diff --git a/.gitattributes b/.gitattributes index ebe93ab374456e8e2dc7920d3527f8a34fa418bc..bbbcfc8547440805f3e1e9883d60ab07a76f67cb 100644 --- a/.gitattributes +++ b/.gitattributes @@ -156,3 +156,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Adapting[[:space:]]In-Domain[[:space:]]Few-Shot[[:space:]]Segmentation[[:space:]]to[[:space:]]New[[:space:]]Domains[[:space:]]without[[:space:]]Source[[:space:]]Domain[[:space:]]Retraining/1f509b59-c5a6-43d8-96e2-83f27bb2df69_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Adapting[[:space:]]Vehicle[[:space:]]Detectors[[:space:]]for[[:space:]]Aerial[[:space:]]Imagery[[:space:]]to[[:space:]]Unseen[[:space:]]Domains[[:space:]]with[[:space:]]Weak[[:space:]]Supervision/8c7d9202-bbbb-43b4-95a8-03f29294f574_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Adaptive[[:space:]]Articulated[[:space:]]Object[[:space:]]Manipulation[[:space:]]On[[:space:]]The[[:space:]]Fly[[:space:]]with[[:space:]]Foundation[[:space:]]Model[[:space:]]Reasoning[[:space:]]and[[:space:]]Part[[:space:]]Grounding/c6de4336-1123-46b0-b3d9-db39393a4712_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Caching[[:space:]]for[[:space:]]Faster[[:space:]]Video[[:space:]]Generation[[:space:]]with[[:space:]]Diffusion[[:space:]]Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Dual[[:space:]]Uncertainty[[:space:]]Optimization_[[:space:]]Boosting[[:space:]]Monocular[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]under[[:space:]]Test-Time[[:space:]]Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Hyper-Graph[[:space:]]Convolution[[:space:]]Network[[:space:]]for[[:space:]]Skeleton-based[[:space:]]Human[[:space:]]Action[[:space:]]Recognition[[:space:]]with[[:space:]]Virtual[[:space:]]Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Learning[[:space:]]of[[:space:]]High-Value[[:space:]]Regions[[:space:]]for[[:space:]]Semi-Supervised[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Prompt[[:space:]]Learning[[:space:]]via[[:space:]]Gaussian[[:space:]]Outlier[[:space:]]Synthesis[[:space:]]for[[:space:]]Out-of-distribution[[:space:]]Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adaptive[[:space:]]Routing[[:space:]]of[[:space:]]Text-to-Image[[:space:]]Generation[[:space:]]Requests[[:space:]]Between[[:space:]]Large[[:space:]]Cloud[[:space:]]Model[[:space:]]and[[:space:]]Light-Weight[[:space:]]Edge[[:space:]]Model/da3df561-f44d-4027-8335-cb684e94422c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AdaptiveAE_[[:space:]]An[[:space:]]Adaptive[[:space:]]Exposure[[:space:]]Strategy[[:space:]]for[[:space:]]HDR[[:space:]]Capturing[[:space:]]in[[:space:]]Dynamic[[:space:]]Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adding[[:space:]]Additional[[:space:]]Control[[:space:]]to[[:space:]]One-Step[[:space:]]Diffusion[[:space:]]with[[:space:]]Joint[[:space:]]Distribution[[:space:]]Matching/c2b3835f-f805-4917-82a0-466f101155c4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Addressing[[:space:]]Representation[[:space:]]Collapse[[:space:]]in[[:space:]]Vector[[:space:]]Quantized[[:space:]]Models[[:space:]]with[[:space:]]One[[:space:]]Linear[[:space:]]Layer/c9ee5a18-482c-48be-ab03-3fe420197718_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Addressing[[:space:]]Text[[:space:]]Embedding[[:space:]]Leakage[[:space:]]in[[:space:]]Diffusion-based[[:space:]]Image[[:space:]]Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AdsQA_[[:space:]]Towards[[:space:]]Advertisement[[:space:]]Video[[:space:]]Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AdvDreamer[[:space:]]Unveils_[[:space:]]Are[[:space:]]Vision-Language[[:space:]]Models[[:space:]]Truly[[:space:]]Ready[[:space:]]for[[:space:]]Real-World[[:space:]]3D[[:space:]]Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Advancing[[:space:]]Text-to-3D[[:space:]]Generation[[:space:]]with[[:space:]]Linearized[[:space:]]Lookahead[[:space:]]Variational[[:space:]]Score[[:space:]]Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Advancing[[:space:]]Textual[[:space:]]Prompt[[:space:]]Learning[[:space:]]with[[:space:]]Anchored[[:space:]]Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Advancing[[:space:]]Visual[[:space:]]Large[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Multi-granular[[:space:]]Versatile[[:space:]]Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Attention[[:space:]]Perturbations[[:space:]]for[[:space:]]Large[[:space:]]Object[[:space:]]Detection[[:space:]]Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Data[[:space:]]Augmentation[[:space:]]for[[:space:]]Single[[:space:]]Domain[[:space:]]Generalization[[:space:]]via[[:space:]]Lyapunov[[:space:]]Exponent-Guided[[:space:]]Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Distribution[[:space:]]Matching[[:space:]]for[[:space:]]Diffusion[[:space:]]Distillation[[:space:]]Towards[[:space:]]Efficient[[:space:]]Image[[:space:]]and[[:space:]]Video[[:space:]]Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Exploitation[[:space:]]of[[:space:]]Data[[:space:]]Diversity[[:space:]]Improves[[:space:]]Visual[[:space:]]Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Purification[[:space:]]via[[:space:]]Super-Resolution[[:space:]]and[[:space:]]Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Reconstruction[[:space:]]Feedback[[:space:]]for[[:space:]]Robust[[:space:]]Fine-grained[[:space:]]Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Robust[[:space:]]Memory-Based[[:space:]]Continual[[:space:]]Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Robustness[[:space:]]of[[:space:]]Discriminative[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]in[[:space:]]Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Adversarial[[:space:]]Training[[:space:]]for[[:space:]]Probabilistic[[:space:]]Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AerialVG_[[:space:]]A[[:space:]]Challenging[[:space:]]Benchmark[[:space:]]for[[:space:]]Aerial[[:space:]]Visual[[:space:]]Grounding[[:space:]]by[[:space:]]Exploring[[:space:]]Positional[[:space:]]Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aether_[[:space:]]Geometric-Aware[[:space:]]Unified[[:space:]]World[[:space:]]Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AffordDexGrasp_[[:space:]]Open-set[[:space:]]Language-guided[[:space:]]Dexterous[[:space:]]Grasp[[:space:]]with[[:space:]]Generalizable-Instructive[[:space:]]Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/After[[:space:]]the[[:space:]]Party_[[:space:]]Navigating[[:space:]]the[[:space:]]Mapping[[:space:]]From[[:space:]]Color[[:space:]]to[[:space:]]Ambient[[:space:]]Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Agreement[[:space:]]aware[[:space:]]and[[:space:]]dissimilarity[[:space:]]oriented[[:space:]]GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AgroBench_[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Benchmark[[:space:]]in[[:space:]]Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AirCache_[[:space:]]Activating[[:space:]]Inter-modal[[:space:]]Relevancy[[:space:]]KV[[:space:]]Cache[[:space:]]Compression[[:space:]]for[[:space:]]Efficient[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Model[[:space:]]Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Align[[:space:]]Your[[:space:]]Rhythm_[[:space:]]Generating[[:space:]]Highly[[:space:]]Aligned[[:space:]]Dance[[:space:]]Poses[[:space:]]with[[:space:]]Gating-Enhanced[[:space:]]Rhythm-Aware[[:space:]]Feature[[:space:]]Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AlignDiff_[[:space:]]Learning[[:space:]]Physically-Grounded[[:space:]]Camera[[:space:]]Alignment[[:space:]]via[[:space:]]Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AlignGuard_[[:space:]]Scalable[[:space:]]Safety[[:space:]]Alignment[[:space:]]for[[:space:]]Text-to-Image[[:space:]]Generation/821c3c22-a165-4d20-a058-8dccd9e55058_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Constraint[[:space:]]Generation[[:space:]]with[[:space:]]Design[[:space:]]Intent[[:space:]]in[[:space:]]Parametric[[:space:]]CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Effective[[:space:]]Tokens[[:space:]]with[[:space:]]Video[[:space:]]Anomaly[[:space:]]in[[:space:]]Large[[:space:]]Language[[:space:]]Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Global[[:space:]]Semantics[[:space:]]and[[:space:]]Local[[:space:]]Textures[[:space:]]in[[:space:]]Generative[[:space:]]Video[[:space:]]Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Information[[:space:]]Capacity[[:space:]]Between[[:space:]]Vision[[:space:]]and[[:space:]]Language[[:space:]]via[[:space:]]Dense-to-Sparse[[:space:]]Feature[[:space:]]Distillation[[:space:]]for[[:space:]]Image-Text[[:space:]]Matching/eee67cc2-5832-464d-b273-17cfe308efbf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Moments[[:space:]]in[[:space:]]Time[[:space:]]using[[:space:]]Video[[:space:]]Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Aligning[[:space:]]Vision[[:space:]]to[[:space:]]Language_[[:space:]]Annotation-Free[[:space:]]Multimodal[[:space:]]Knowledge[[:space:]]Graph[[:space:]]Construction[[:space:]]for[[:space:]]Enhanced[[:space:]]LLMs[[:space:]]Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/All[[:space:]]Parts[[:space:]]Matter_[[:space:]]A[[:space:]]Unified[[:space:]]Mask-Free[[:space:]]Virtual[[:space:]]Try-On[[:space:]]Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/All[[:space:]]in[[:space:]]One_[[:space:]]Visual-Description-Guided[[:space:]]Unified[[:space:]]Point[[:space:]]Cloud[[:space:]]Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AllGCD_[[:space:]]Leveraging[[:space:]]All[[:space:]]Unlabeled[[:space:]]Data[[:space:]]for[[:space:]]Generalized[[:space:]]Category[[:space:]]Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AllTracker_[[:space:]]Efficient[[:space:]]Dense[[:space:]]Point[[:space:]]Tracking[[:space:]]at[[:space:]]High[[:space:]]Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Alleviating[[:space:]]Textual[[:space:]]Reliance[[:space:]]in[[:space:]]Medical[[:space:]]Language-guided[[:space:]]Segmentation[[:space:]]via[[:space:]]Prototype-driven[[:space:]]Semantic[[:space:]]Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Allowing[[:space:]]Oscillation[[:space:]]Quantization_[[:space:]]Overcoming[[:space:]]Solution[[:space:]]Space[[:space:]]Limitation[[:space:]]in[[:space:]]Low[[:space:]]Bit-Width[[:space:]]Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Always[[:space:]]Skip[[:space:]]Attention/569226ba-e747-4584-b317-a54b6e4497d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Amodal[[:space:]]Depth[[:space:]]Anything_[[:space:]]Amodal[[:space:]]Depth[[:space:]]Estimation[[:space:]]in[[:space:]]the[[:space:]]Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Amodal3R_[[:space:]]Amodal[[:space:]]3D[[:space:]]Reconstruction[[:space:]]from[[:space:]]Occluded[[:space:]]2D[[:space:]]Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]Efficient[[:space:]]Hybrid[[:space:]]Vision[[:space:]]Transformer[[:space:]]for[[:space:]]TinyML[[:space:]]Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]Efficient[[:space:]]Post-hoc[[:space:]]Framework[[:space:]]for[[:space:]]Reducing[[:space:]]Task[[:space:]]Discrepancy[[:space:]]of[[:space:]]Text[[:space:]]Encoders[[:space:]]for[[:space:]]Composed[[:space:]]Image[[:space:]]Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]Empirical[[:space:]]Study[[:space:]]of[[:space:]]Autoregressive[[:space:]]Pre-training[[:space:]]from[[:space:]]Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]Information-Theoretic[[:space:]]Regularizer[[:space:]]for[[:space:]]Lossy[[:space:]]Neural[[:space:]]Image[[:space:]]Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]Inversion-based[[:space:]]Measure[[:space:]]of[[:space:]]Memorization[[:space:]]for[[:space:]]Diffusion[[:space:]]Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/An[[:space:]]OpenMind[[:space:]]for[[:space:]]3D[[:space:]]Medical[[:space:]]Vision[[:space:]]Self-supervised[[:space:]]Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Analyzing[[:space:]]Finetuning[[:space:]]Representation[[:space:]]Shift[[:space:]]for[[:space:]]Multimodal[[:space:]]LLMs[[:space:]]Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Anchor[[:space:]]Token[[:space:]]Matching_[[:space:]]Implicit[[:space:]]Structure[[:space:]]Locking[[:space:]]for[[:space:]]Training-free[[:space:]]AR[[:space:]]Image[[:space:]]Editing/02ecc675-dff6-4c74-84e6-883c770f01a4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AnimalClue_[[:space:]]Recognizing[[:space:]]Animals[[:space:]]by[[:space:]]their[[:space:]]Traces/8e6d9b6c-ad00-4a08-84f7-166aef83cedd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Animate[[:space:]]Anyone[[:space:]]2_[[:space:]]High-Fidelity[[:space:]]Character[[:space:]]Image[[:space:]]Animation[[:space:]]with[[:space:]]Environment[[:space:]]Affordance/f25b16a0-97a2-4bcc-93f4-3c50d28debfd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AnimateAnyMesh_[[:space:]]A[[:space:]]Feed-Forward[[:space:]]4D[[:space:]]Foundation[[:space:]]Model[[:space:]]for[[:space:]]Text-Driven[[:space:]]Universal[[:space:]]Mesh[[:space:]]Animation/fc436c83-46ad-42d4-ab61-369656d1b892_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AnimeGamer_[[:space:]]Infinite[[:space:]]Anime[[:space:]]Life[[:space:]]Simulation[[:space:]]with[[:space:]]Next[[:space:]]Game[[:space:]]State[[:space:]]Prediction/842749ed-2ab1-47b2-b183-403e0f55e3a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/AnnofreeOD_[[:space:]]Detecting[[:space:]]All[[:space:]]Classes[[:space:]]at[[:space:]]Low[[:space:]]Frame[[:space:]]Rates[[:space:]]Without[[:space:]]Human[[:space:]]Annotations/c469fc9e-69e6-4a2c-a3c4-94d82008c066_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Anomaly[[:space:]]Detection[[:space:]]of[[:space:]]Integrated[[:space:]]Circuits[[:space:]]Package[[:space:]]Substrates[[:space:]]Using[[:space:]]the[[:space:]]Large[[:space:]]Vision[[:space:]]Model[[:space:]]SAIC_[[:space:]]Dataset[[:space:]]Construction,[[:space:]]Methodology,[[:space:]]and[[:space:]]Application/d944de6f-b792-4861-9144-b4eef3b9b54d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Anti-Tamper[[:space:]]Protection[[:space:]]for[[:space:]]Unauthorized[[:space:]]Individual[[:space:]]Image[[:space:]]Generation/398b8e09-051b-45a4-b77b-d72f2adef0da_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_content_list.json b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44303073f1a3af9df90f89529015a782ee3dd9b2 --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_content_list.json @@ -0,0 +1,1561 @@ +[ + { + "type": "text", + "text": "Adaptive Caching for Faster Video Generation with Diffusion Transformers", + "text_level": 1, + "bbox": [ + 116, + 126, + 880, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kumara Kahatapitiya $^{1,2*}$ Haozhe Liu $^{1}$ Sen He $^{1}$ Ding Liu $^{1}$ Menglin Jia $^{1}$ Chenyang Zhang $^{1}$ Michael S. Ryoo $^{2}$ Tian Xie $^{1}$ Meta AI $^{2}$ Stony Brook University adacache-dit.github.io", + "bbox": [ + 194, + 164, + 785, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs)—despite making significant headway in this context—have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that “not all videos are created equal”: meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to $4.7 \\times$ on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines.", + "bbox": [ + 89, + 300, + 485, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 654, + 220, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [33, 82] have become the standard for generative modeling in recent years, arguably surpassing the quality of VAEs [41, 72], GANs [26, 40] and Auto-Regressive models [9, 10]. This observation holds in a wide-range of applications including image [73, 76], video [6, 81], 3D [52, 69], and audio [37, 44] generation, as well as image [4, 32] and video [70, 99] editing. More recent Diffusion Transformers (DiTs) [60, 67] show better promise in terms of scalability and generalization compared to prior UNet-based diffusion models [73], revealing intriguing horizons in GenAI for the years to come.", + "bbox": [ + 89, + 680, + 482, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the state-of-the-art performance, DiTs can also be computationally expensive both in terms of memory and", + "bbox": [ + 89, + 847, + 482, + 877 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad9bd3d2c074a36ac17b15f1d552409a39d370e2b0812379de409bc9fc3b0ab0.jpg", + "image_caption": [ + "Figure 1. Effectiveness of Adaptive Caching: We show a qualitative comparison of AdaCache (right) applied on top of Open-Sora [120] (left), a baseline video DiT. Here, we consider generating $720\\mathrm{p}$ - 2s video clips, and report VBench [38] quality and average latency (on a single A100 GPU) on the benchmark prompts from Open-Sora [120] gallery. AdaCache generates videos significantly faster (i.e., $4.7\\times$ speedup) with a comparable quality. Also, the number of computed steps varies for each video. Best-viewed with zoom-in. Prompts given in supplementary." + ], + "image_footnote": [], + "bbox": [ + 517, + 267, + 710, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b571e2e99ba297a152cc76f1dad1df0fcb0c21f7f340562909bf67101c1dc0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 266, + 905, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "computational requirements. This becomes especially critical when applied with a large number of input tokens (e.g. high-resolution long video generation). For instance, the reason for models such as Sora [66] not being publicly served is speculated to be the high resource demands and slower inference speeds [54]. To tackle these challenges and reduce the footprint of diffusion models, various research directions have emerged such as latent diffusion [73], step-distillation [79, 111], caching [28, 61, 98], architecture-search [48, 119], token reduction [8, 47] and region-based methods [39, 65]. Fewer techniques transfer readily from UNet-based pipelines to DiTs, whereas others often require novel formulations. Hence, DiT acceleration has been under-explored as of yet.", + "bbox": [ + 511, + 654, + 906, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Moreover, we note that not all videos are created equal. Some videos contain high-frequency textures and signifi", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done at Meta", + "bbox": [ + 109, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15240", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cant motion content, whereas others are much simpler (e.g. with homogeneous textures or static regions). Having a diffusion process tailored specifically for each video generation can be beneficial in terms of realizing the best quality-latency trade-off. This idea has been explored to some extent in region-based methods [4, 39, 65], but not sufficiently in the context of video generation.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above, we introduce Adaptive Caching (AdaCache) for accelerating video diffusion transformers. This approach requires no training and can seamlessly be integrated into a baseline video DiT at inference, as a plug-and-play component. The core idea of our proposal is to cache residual computations within transformer blocks (e.g. attention or MLP outputs) in a certain diffusion step, and reuse them through a number of subsequent steps that is dependent on the generated video. We do this by devising a caching schedule, i.e., deciding when-to-recompute next whenever making a residual computation. This decision is guided by a distance metric that measures the rate-of-change between previously-stored and current representations. If the distance is high we would not cache for an extended period (i.e., #steps), to avoid reusing incompatible representations. We further introduce a Motion Regularization (MoReg) to allocate computations based on the motion content in the video being generated. This is inspired by the observation that high-moving sequences require more diffusion steps to achieve a reasonable quality. Altogether, our pipeline is applied on top of multiple video DiT baselines showing much-faster inference speeds without sacrificing the generation quality (see Fig. 1). Finally, we validate the effectiveness of our contributions and justify our design decisions through ablations and qualitative comparisons.", + "bbox": [ + 91, + 198, + 486, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 588, + 232, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based Video Generation [6, 11, 25, 34, 81] has surpassed the quality and diversity of GAN-based approaches [18, 77, 89, 91, 112], while also being competitive with recent Auto-Regressive models [35, 43, 49, 90, 101, 105]. They have become a standard component in the pipelines for frame interpolation [23, 96], video outpainting [16, 22, 93], image-to-video [6, 27, 103], video-to-video (i.e., video editing or translation) [36, 106, 109], personalization [63, 100], motion customization [104, 116] and compositional generation [51, 107]. The underlying architecture of video diffusion models has evolved from classical UNets [73, 74] with additional spatio-temporal attention layers [7, 17, 25, 29], to fully-fledged transformer-based (i.e., DiT [67]) architectures [24, 56, 62, 114]. In the process, the latency of denoising [55, 82] has also scaled with larger models [24, 68]. This becomes critical especially in applications such as long-video generation [31, 87, 92, 110, 115, 121], while also affecting the growth of commercially-served video models [42, 57, 66, 75].", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficiency of Diffusion models has been actively explored with respect to both training and inference pipelines. Multi-stage training at varying resolutions [12, 13, 24] and high-quality data curation [6, 20, 34, 71] have cut down training costs significantly. In terms of inference acceleration, there exist two main approaches: (1) methods that require re-training such as step-distillation [53, 64, 78, 79], consistency regularization [58, 83], quantization [14, 21, 30, 46, 94], and architecture search/compression [48, 108, 119], or (2) methods that require no re-training such as token reduction [8, 39, 47] and caching [15, 28, 61, 98, 118]. Among these, training-free methods are more-attractive as they can be widely-adopted without any additional costs. This becomes especially relevant for video diffusion models that are both expensive to train and usually very slow at inference. In this paper, we explore a caching-based approach tailored for video DiTs. Different from prior fixed caching schedules in UNet-based [28, 61, 98] and DiT-based [2, 15, 80, 118] pipelines, we introduce a content-dependent (i.e., adaptive) caching scheme to squeeze out the best quality-latency trade-off.", + "bbox": [ + 511, + 90, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Content-adaptive Generation may focus on improving consistency [3-5, 19, 95, 102], quality [1, 85], and/or efficiency [39, 65, 84, 88]. Most region-based methods (e.g. image or video editing) rely on a user-provided mask to ensue consistent generations aligned with context information [4, 102]. Some others automatically detect [85] or retrieve [1] useful information to improve generation quality. Among efficiency-oriented approaches, there exist proposals for selectively-processing a subset of latents [39, 65], switching between diffusion models with varying compute budgets [84], or adaptively-controlling the number of denoising steps [50, 59, 88, 98]. Our proposed AdaCache—which also controls #denoising-steps adaptively—provides better flexibility in the sense that (1) it can selectively-cache any layer or even just a specific module within a layer, and (2) it is tailored to each video generation instead of being fixed for a given architecture, enabling a better-adaptive compute allocation.", + "bbox": [ + 511, + 412, + 908, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Not All Videos Are Created Equal", + "text_level": 1, + "bbox": [ + 511, + 705, + 812, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we motivate the need for a content-dependent denoising process, and show how it can help maximize the quality-latency trade-off. This motivation is based on a couple of interesting observations.", + "bbox": [ + 511, + 732, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we note that each video is unique. Hence, videos have varying levels of complexity. Here, the complexity of a given video can be expressed by the rate-of-change of information across both space and time. Simpler videos may contain more homogeneous regions and/or static content. In contrast, complex videos have more high-frequency details and/or significant motion. The standard video compres", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15241", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67afe584c0b8e55dbf42b2bb79bbf41ac6e06d640e227151de3d3cd63dafaea6.jpg", + "image_caption": [ + "(a) Critical number of denoising-steps varies." + ], + "image_footnote": [], + "bbox": [ + 94, + 109, + 352, + 251 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b634b13384c76986ecce34fac45d5db4eebc648712d51d803402812231cab8e.jpg", + "image_caption": [ + "(b) Changes in subsequent representations are unique.", + "Figure 2. Not all videos are created equal: We show frames from $720\\mathrm{p}$ - 2s video generations based on Open-Sora [120]. (a) We try to break each generation by reducing the number of diffusion steps. Interestingly, not all videos have the same break point. Some sequences are extremely robust (e.g. first column), while others break easily. (b) When we plot the difference between computed representations in subsequent diffusion steps, we see unique variations (Feature distance vs. #steps). If we are to reuse similar representations across denoising steps, it needs to be tailored to each video. (c) There exist different configurations for generating videos at an approximately-fixed latency (e.g. having an arbitrary #denoising-steps, yet only computing a fixed #representations and reusing otherwise). We observe a significant variance in quality in such videos. All these observations suggest the need for a content-dependent denoising process, which is the founding motivation of Adaptive Caching. Best-viewed with zoom-in. Prompts given in supplementary." + ], + "image_footnote": [], + "bbox": [ + 362, + 89, + 630, + 252 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2196f2d689d3b0e3f285f6eac49cfe58f1f8429425cf130762ffa53ad16c6ad1.jpg", + "image_caption": [ + "(c) Different levels of quality at a given latency budget." + ], + "image_footnote": [], + "bbox": [ + 642, + 108, + 905, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion techniques exploit such information to achieve the best possible compression ratios without sacrificing the quality [86, 97]. Motivated by the same, we explore how the compute-cost affects the quality of video generations based on DiTs. We measure this w.r.t. the number of denoising steps, and the observations are shown in Fig. 2 (a). Some video sequences are very robust, and achieve a reasonable quality even at fewer denoising steps. Others break easily when we keep reducing the #steps, but the break-point varies. This suggests that the minimal #steps (or, computations) required to generate a video with a reasonable quality varies, and having a content-dependent denoising schedule can exploit this for speeding-up generation.", + "bbox": [ + 88, + 392, + 483, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we observe how the computed representations (i.e., residual connections in attention or MLP blocks within DiT) change during the denoising process, across different video generations. This may reveal the level of compute redundancy in each video generation, enabling us to reuse representations and improve efficiency. More specifically, we visualize the feature differences between subsequent diffusion steps as histograms given in Fig. 2 (b). Here, we report Feature distance (e.g. L1) vs. #steps. We observe that each histogram is unique. Despite having higher changes in early/latter steps and smaller changes in the middle, the overall distribution and the absolute values vary considerably. A smaller change corresponds to higher redundancy across subsequent computations, and an opportunity for reusing. This motivates the need for a non-uniform compute-schedule not only within the diffusion process of a given video, but also across different videos.", + "bbox": [ + 89, + 593, + 483, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we evaluate the video generation quality at a capped-budget (i.e., fixed computations or latency). We can have multiple generation configurations at an", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approximately-fixed latency, by computing a constant number of representations. For instance, we can cache and reuse representations more-frequently in a setup with more denoising steps, still having the same latency of a process with fewer steps. The observations of a study with either 30 or 100 base denoising steps is shown in Fig. 2 (c). We see that the generation quality varies significantly despite spending a similar cost and having the same underlying pretrained DiT. This motivates us to think about how best to allocate resources at inference, tailored for each video generation.", + "bbox": [ + 511, + 392, + 906, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Adaptive Caching for Faster Video DiTs", + "text_level": 1, + "bbox": [ + 511, + 571, + 872, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Preliminaries: Video Diffusion Transformers", + "text_level": 1, + "bbox": [ + 511, + 601, + 895, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Diffusion Transformers are extended from Latent Diffusion Transformers (DiTs) [67] introduced for image generation. DiTs provide a much-more streamlined, scalable architecture compared to prior UNet-based diffusion models [73], by only having transformer blocks with a homogeneous token resolution (instead of convolutional blocks with up/downsampling). A simplified transformer block (i.e., w/o normalizing or timestep conditioning layers) in a video DiT is shown in Fig. 3 (right)—gray block. It consists of spatial-temporal attention (STA), cross-attention (CA) and linear (MLP) layers. Depending on the implementation, STA may be a single joint spatio-temporal attention layer, or separate spatial and temporal attention layers repeated within alternating blocks. Without loss of generality, let us denote a latent feature at the input/output of such block by $f_{t}^{l}$ and $f_{t}^{l + 1}$ , respectively. Here, $l$ represents the layer index, and $t$ , the diffusion timestep. A simplified flow of computations within each block can be represented as,", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15242", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {t} ^ {l} = \\operatorname {S T A} \\left(f _ {t} ^ {l}\\right); \\quad \\tilde {f} _ {t} ^ {l} = f _ {t} ^ {l} + p _ {t} ^ {l}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 109, + 482, + 128 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {t} ^ {l} = \\mathrm {C A} \\left(\\tilde {f} _ {t} ^ {l}\\right) \\quad ; \\quad \\bar {f} _ {t} ^ {l} = \\tilde {f} _ {t} ^ {l} + q _ {t} ^ {l}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 132, + 482, + 150 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} ^ {l} = \\mathrm {M L P} \\left(\\bar {f} _ {t} ^ {l}\\right); f _ {t} ^ {l + 1} = \\bar {f} _ {t} ^ {l} + r _ {t} ^ {l}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 152, + 482, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $p_t^l$ , $q_t^l$ and $r_t^l$ are residual connections corresponding to each compute-element. Such computations repeat through $L$ layers, generating the noise prediction of each step $t$ , and across a total of $T$ denoising steps. In the current streamlined video DiT architectures with homogeneous token resolutions, each layer or denoising step costs the same.", + "bbox": [ + 89, + 178, + 483, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Adaptive Caching", + "text_level": 1, + "bbox": [ + 89, + 276, + 266, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this subsection, we introduce Adaptive Caching (AdaCache), a training-free mechanism for content-dependent compute allocation in video DiTs. The overview of Adaptive Caching is shown in Fig. 3. Compared to a standard DiT that computes representations for all layers across all diffusion steps, in AdaCache, we decide which layers or steps to compute, adaptively (i.e., dependent on each video being generated). This decision is based on the rate-of-change in the residual connections (e.g. $p_t^l$ , $q_t^l$ or $r_t^l$ ) across diffusion steps, which amount to all significant computations within the DiT. Without loss of generality, let us assume that the residuals in block $l$ in current and immediately-prior diffusion steps $t$ and $t + k$ are already computed. Here, step $t + k$ is identified as 'immediately-prior' to step $t$ since any residuals between these two steps are assumed to be not computed (i.e., cached residuals are reused). We make a decision on the next computation step based on the distance metric $(c_t^l)$ given by,", + "bbox": [ + 89, + 297, + 483, + 570 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {t} ^ {l} = \\operatorname {d i s t} \\left(p _ {t + k} ^ {l}, p _ {t} ^ {l}\\right) = \\left\\| p _ {t} ^ {l} - p _ {t + k} ^ {l} \\right\\| / k. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 575, + 482, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we use L1 distance by default, but other distance metrics can also be applied (e.g. L2, cosine). Once we have the distance metric, we select the next caching rate $(\\tau_t^l)$ based on a pre-defined codebook of basis cache-rates. Here, a 'cache-rate' is defined as the number subsequent steps during which, a previously-computed representation is re-used (i.e., a higher cache-rate gives more compute savings). The codebook is basically a collection of cache-rates defined based on the original denoising schedule (i.e., #steps), coupled with corresponding metric thresholds to select them. Simply put, a higher distance metric will sample a lower cache-rate from the codebook, resulting in more-frequent re-computations. More details about the codebook is available in the supplementary (Sec. A.2).", + "bbox": [ + 89, + 599, + 483, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {t} ^ {l} = \\operatorname {c o d e b o o k} \\left(c _ {t} ^ {l}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 816, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For all denoising steps within $t$ and $t - \\tau$ , we reuse previously-cached representations and only recompute after the current caching schedule (while also estimating the metric, again).", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8fd638c32ce17918dd5cf18a4b185c1ddc6821f457a1c5282bd974e29706e49e.jpg", + "image_caption": [ + "Figure 3. Overview of Adaptive Caching: (Left) During the diffusion process, we choose to cache residual computations within selected DiT blocks. The caching schedule is content-dependent, as we decide when to compute the next representation based on a distance metric $(c_{t})$ . This metric measures the rate-of-change from previously-computed (and, stored) representation to the current one, and can be evaluated per-layer or the DiT as-a-whole. Each computed residual can be cached and reused across multiple steps. (Right) We only cache the residuals (i.e., skip-connections) which amount to the actual computations (e.g. spatial-temporal/cross attention, MLP). The iteratively denoised representation (i.e., $f_{t+k}$ , $f_{t}$ ) always gets updated either with computed or cached residuals." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {t - k} ^ {l} = \\left\\{ \\begin{array}{l l} p _ {t} ^ {l} & \\text {i f} k < \\tau_ {t} ^ {l}; \\\\ p _ {t - k} ^ {l} = \\operatorname {S T A} (f _ {t - k} ^ {l}) & \\text {i f} k = \\tau_ {t} ^ {l}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 414, + 906, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The same applies to other residual computations (e.g. $q_{t - k}^{l}, r_{t - k}^{l}$ ) as well. By design, we can have unique caching schedules for each layer (and, each residual computation). However, we observe that it will make the generations unstable. Therefore, we decide to have a common metric (i.e., $c_t^l = c_t$ ) and hence, a common caching rate (i.e., $\\tau_t^l = \\tau_t$ ) across all DiT layers. For instance, we can consider an averaged metric across all layers, or a metric computed at a certain layer to decide the caching schedule. Meaning, when we recompute residuals in a certain step, we do so for the whole DiT rather than selectively for each layer.", + "bbox": [ + 511, + 462, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, this setup allows us to adaptively-control the compute spent on each video generation, based on framewise information (i.e., no temporal information used as of yet). If the rate-of-change between residuals is high, we will have a smaller caching rate, and otherwise, we have a higher rate. The choice of a lightweight distance metric (e.g. L1) helps us avoid any additional latency overheads.", + "bbox": [ + 511, + 628, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Motion Regularization", + "text_level": 1, + "bbox": [ + 511, + 742, + 723, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further improve Adaptive Caching by making use of video-specific (i.e., temporal) information, we introduce a Motion Regularization (MoReg). This is motivated by the observation that the optimal number of denoising steps varies based on the motion content of each generated video. The core idea here, is to cache less (i.e., recompute more) if a generated video has a high motion content. Simply put, we plan to regularize our caching schedule based on motion. However the problem is that, we need to estimate mo", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15243", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion while the video is still being generated. Therefore, we can not rely on motion estimation algorithms in the pixel space, nor any compute-heavy ones as our focus is on efficiency. As a result, we estimate a noisy latent motion-score $(m_t^l)$ based on residual frame differences. Without loss of generality, let us denote residual latent frames of $p_t^l$ as $\\{p_{t,n}^l\\mid n = 0,\\dots ,N - 1\\}$ where $N$ is the #frames in latent space (generated by the VAE encoder). We estimate the motion-score as,", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t} ^ {l} = \\left\\| p _ {t, i: N} ^ {l} - p _ {t, 0: N - i} ^ {l} \\right\\|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 234, + 482, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $i$ denotes the frame step-size (or, frame-rate), $\\| \\cdot \\|$ , the distance metric (e.g. L1), and $i:j$ , the slice of all frames within the corresponding range. However, since we operate on noisy-latents, we observe that our motion estimate, particularly in early diffusion steps is not reliable. Meaning, it does not provide a reasonable regularization in early steps (i.e., the change in caching schedule does not correlate well with the observed motion in pixel space). To alleviate this, we also compute a motion-gradient $(mg_t^l)$ across diffusion steps, which can act as a reasonable early-predictor of motion that we may observe in latter diffusion steps (that also correlates with the motion in pixel space).", + "bbox": [ + 89, + 265, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm g _ {t} ^ {l} = \\left(m _ {t} ^ {l} - m _ {t + k} ^ {l}\\right) / k. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 460, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the motion-score being noisy, the motion-gradient acts as a better-estimate of trend, as the representations are getting denoised and converging to a noise-free distribution. Finally, we use both motion-score and motion-gradient as a scaling-factor of the distance metric $(c_t^l)$ to regularize our caching schedule.", + "bbox": [ + 89, + 492, + 483, + 583 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {c} _ {t} ^ {l} = c _ {t} ^ {l} \\cdot \\left(m _ {t} ^ {l} + m g _ {t} ^ {l}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 597, + 482, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This means, when we have a higher estimated motion, the distance metric will be increased and a smaller basis cache-rate will be selected from the codebook. As previously discussed, we also enforce a common motion-regularization in all DiT layers by computing a common motion score (i.e., $m_t^l = m_t$ , $mg_t^l = mg_t$ ), ensuring the stability of the denoising process. We can also choose to compute motion at different frame-rates, which we ablate in our experiments. Refer to the supplementary for concrete examples of motion-score and motion-gradient (Fig. A.1).", + "bbox": [ + 89, + 628, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 792, + 223, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 816, + 303, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We select multiple prominent open-source video DiTs as backbone video generation pipelines in our experiments, namely, Open-Sora-v1.2 [120] for Text+Image-to-Video (TI2V) generation, and Open-Sora-Plan-v1.1 [45] and Latte", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[62] for Text-to-Video (T2V) generation. Since we focus on inference-based latency optimizations (i.e., without any re-training), we compare AdaCache against similar methods such as $\\Delta$ -DiT [15], T-GATE [113] and PAB [118]. In our main experiments, we generate $900+$ videos based on standard VBenchmark [38] benchmark prompts, at the corresponding baseline generation settings suggested in [118] for a fair comparison (e.g. $480p$ - 2s with 30-steps in Open-Sora, $512\\times 512$ - 2.7s with 150-steps in Open-Sora-Plan and $512\\times 512$ - 2s with 50-steps in Latte). We report VBenchmark average and reference-based PSNR, SSIM and LPIPS as quality metrics, and also report FLOPs, Latency (s) and Speedup as complexity metrics. Here, Latency is measured on a single A100 GPU. Unless otherwise stated, in our ablations and qualitative results, we experiment on the prompts from Open-Sora benchmark gallery, generating $720p$ - 2s videos with 100-steps, following the recommendation of model authors—which gives better quality video generations (and, higher speedups) compared to the benchmark setting discussed above (i.e., $480p$ - 2s with 30-steps).", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Main results", + "text_level": 1, + "bbox": [ + 513, + 405, + 645, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Table 1, we present a quantitative evaluation of quality and latency on VBench [38] benchmark. We consider three variants of AdaCache: a slow variant, a fast variant with more speedup and the same with motion regularization. We compare with other training-free acceleration methods, showing consistently better speedups with a comparable generation quality. With Open-Sora [120] baseline, AdaCache-slow outperforms others on all quality metrics, while giving a $1.46 \\times$ speedup compared to PAB [118] with $1.20 \\times$ speedup. AdaCache-fast gives the highest acceleration of $2.24 \\times$ with a slight drop in quality. AdaCache-fast (w/ MoReg) shows a clear improvement in quality compared to AdaCache-fast, validating the effectiveness of our regularization and giving a comparable speedup of $2.10 \\times$ . All AdaCache variants outperform even the baseline (w/o any acceleration) on VBench average quality, which aligns better with human preference compared to other reference-based metrics. Similar observations hold with the other baselines as well. With Open-Sora-Plan [45], AdaCache shows the best speedup of $3.70 \\times$ compared to the previous-best $1.45 \\times$ of PAB, and the best quality with a $2.20 \\times$ speedup. With Latte [62], we gain the best speedup of $2.74 \\times$ compared to prior $1.33 \\times$ , and the best overall quality with a $1.59 \\times$ speedup.", + "bbox": [ + 511, + 428, + 906, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "User study: Quantitative metrics on video generation quality can sometimes fall-short in aligning with the perceived visual quality. To better understand the human preference on AdaCache and its comparisons, we conduct a user study in the form of randomized A/B preference tests. Here, we create a questionnaire with 50 multiple-choice questions, each consisting of 3 variants of a single video sequence:", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15244", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c53e9bc2e656c7878443cb4673756141a03924a7f2c13bb9e20f96d3e471a578.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVBatch (%) ↑PSNR ↑LPIPS ↓SSIM ↑FLOPs (T)Latency (s)Speedup
Open-Sora [120]79.22---3230.2454.021.00×
+ Δ-DiT [15]78.2111.910.56920.48113166.47--
+ T-GATE [113]77.6115.500.34950.67602818.4049.111.10×
+ PAB-fast [118]76.9523.580.17430.82202558.2540.231.34×
+ PAB-slow [118]78.5127.040.09250.88472657.7044.931.20×
+ AdaCache-fast79.3924.920.09810.83751331.9724.162.24×
+ AdaCache-fast (w/ MoReg)79.4825.780.08670.85301383.6625.712.10×
+ AdaCache-slow79.6629.970.04560.90852195.5037.011.46×
Open-Sora-Plan [45]80.39---12032.40129.671.00×
+ Δ-DiT [15]77.5513.850.53880.373612027.72--
+ T-GATE [113]80.1518.320.30660.621910663.32113.751.14×
+ PAB-fast [118]71.8115.470.54990.47178551.2689.561.45×
+ PAB-slow [118]80.3018.800.30590.65509276.5798.501.32×
+ AdaCache-fast75.8313.530.54650.43093283.6035.043.70×
+ AdaCache-fast (w/ MoReg)79.3017.690.37450.61473473.6836.773.53×
+ AdaCache-slow80.5022.980.17370.79104983.3058.882.20×
Latte [62]77.40---3439.4732.451.00×
+ Δ-DiT [15]52.008.650.85130.10783437.33--
+ T-GATE [113]75.4219.550.26120.69273059.0229.231.11×
+ PAB-fast [118]73.1317.160.39030.64212576.7724.331.33×
+ PAB-slow [118]76.3219.710.26990.70142767.2226.201.24×
+ AdaCache-fast76.2617.700.35220.66591010.3311.852.74×
+ AdaCache-fast (w/ MoReg)76.4718.160.32220.68321187.3113.202.46×
+ AdaCache-slow77.0722.780.17370.80302023.6520.351.59×
", + "bbox": [ + 184, + 88, + 810, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation of quality and latency: Here, we compare AdaCache with other training-free DiT acceleration methods (e.g. $\\Delta$ -DiT [15], T-GATE [113], PAB [118]) on multiple video baselines (e.g. Open-Sora [120] $480p$ - 2s at 30-steps, Open-Sora-Plan [45] $512\\times 512$ - 2.7s at 150-steps, Latte [62] $512\\times 512$ - 2s at 50-steps). We measure the generation quality with VBenchmark [38], PSNR, LPIPS and SSIM, while reporting complexity with FLOPs, latency and speedup (measured on a single A100 GPU). AdaCache-fast consistently shows the best speedups at a comparable or slightly-lower generation quality. AdaCache-slow gives absolute-best quality while still being faster than prior methods. Our motion-regularization significantly improves the generation quality consistently, with a minimal added-latency.", + "bbox": [ + 89, + 372, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f16ae822e7e91a2d4eca199c514c5bd2d747adfb53b8cfafda3cfcc7235516f5.jpg", + "image_caption": [ + "Figure 4. User study: We collect human preferences, comparing AdaCache with PAB [118] (left) and evaluating our motion regularization (right). AdaCache shows a significantly higher preference-rate over PAB at a comparable latency. Our motion-regularized variant is better-preferred, yet often tied with AdaCache in terms of perceived quality." + ], + "image_footnote": [], + "bbox": [ + 104, + 464, + 269, + 575 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72c24cca44aec9be03284401bd7b301478f68ef9564e4745e4418e5a750295d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 464, + 442, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the baseline, and two efficient generations in a randomized order (comparing either AdaCache vs. PAB [118] at a similar speedup, or AdaCache vs. AdaCache w/ MoReg). We ask the users which efficient variant shows a better quality, and whether it is aligned with (i.e., indistinguishable from) the baseline. We collect a total of 1800 responses from 36 different users, and the results of the study are given in Fig. 4. Between AdaCache and PAB, we see a clear win for our method $(70\\%)$ while being extremely-similar to the baseline more than half the time $(41\\%)$ . Among AdaCache variants, users find these to be often tied $(60\\%)$ in-terms of perceived quality, yet still showing a better preference for motion-regularized variant $(25\\%$ vs. $14\\%)$ . This study validates the effectiveness of Adaptive Caching.", + "bbox": [ + 89, + 689, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7fcaceef2b10348d240fa7dc97d067eaf3a9ec174539fa1fe61af919391d341.jpg", + "image_caption": [ + "Figure 5. Quality-Latency trade-off: We show quality vs. latency curves for different configurations of AdaCache and PAB [118], with Open-Sora [120] 720p - 2s generations. AdaCache outperforms PAB consistently, showing a more-stable performance while reducing latency. This stability is more-prominent in reference-free metric VBenchmark [38] compared to reference-based metrics, validating that AdaCache generations are aligned with human preference even at its fastest speeds, despite not being exactly-aligned with the reference." + ], + "image_footnote": [], + "bbox": [ + 516, + 465, + 903, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 737, + 663, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality-Latency trade-off: In Fig. 5, we compare the quality-latency trade-off of AdaCache with PAB [118]. First, we note that AdaCache enables significantly higher reduction rates (i.e., much smaller absolute latency) compared to PAB. Moreover, across this whole range of latency configurations, AdaCache gives a more-stable performance over PAB, on all quality metrics. Such behavior is especially evident in reference-free metric VBench [38], that aligns better with human preference. Even if we see a drop", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15245", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d1392a61046aae5824c4ca7eba8c3712bd94ce83d6df49a7d5a738e2f0c2143.jpg", + "table_caption": [], + "table_footnote": [ + "(a) AdaCache with Motion Regularization: We show different variants of AdaCache. All versions achieve significant speedups compared to the baseline. AdaCache + MoReg shows a better quality with a slightly-lower speedup." + ], + "table_body": "
MethodVBatch sizeLatencySpeedup
Open-Sora [120]84.16419.601.0×
+ AdaCache83.4089.534.7×
+ AdaCache + MoReg83.5093.504.5×
+ AdaCache + MoReg (w/o grad)83.3689.014.7×
+ AdaCache + MoReg (multi-step)83.4295.654.4×
", + "bbox": [ + 109, + 103, + 532, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/481a7075343115cd24a80ead9e57c9b15dac28575289de0f4671fa59ebe603f0.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Speedups at different resolutions: We compare AdaCache with baselines at different resolutions. It generalizes across resolutions, consistently providing a stable acceleration." + ], + "table_body": "
ResolutionAdaCacheVBenchmarkLatencySpeedup
480p - 2s✘83.68173.841.0×
✓83.1838.524.5×
480p - 4s✘82.77349.901.0×
✓82.1680.164.4×
720p - 2s✘84.16419.601.0×
✓83.4089.534.7×
", + "bbox": [ + 542, + 88, + 893, + 196 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/42bd02f65193b6724cc657a442ba5f65c96aa3907e196f68d096d9644b77ce3a.jpg", + "table_caption": [], + "table_footnote": [ + "(c) Cache metric: Among distance metrics, L1/L2 have similar (and better) performance in contrast to cosine distance." + ], + "table_body": "
DistanceVBenchLatency
L183.4089.53
L283.5092.70
Cosine83.1986.74
", + "bbox": [ + 99, + 247, + 269, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc3f68b7099827592510104a2b630fa60ff3beb814917d97704abd09b735462e.jpg", + "table_caption": [], + "table_footnote": [ + "(d) Cache location: We compute the cache metric at midDiT, resulting in the best quality-latency trade-off." + ], + "table_body": "
LocationVBenchLatency
Start83.3087.55
Mid83.4089.53
End83.4391.20
Multiple83.4190.27
", + "bbox": [ + 279, + 239, + 452, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09e512a8dc4a181d57cdeec573b2f384946b1d915b4e44cdeeade0d2a8c96781.jpg", + "table_caption": [], + "table_footnote": [ + "(e) Cache residual: We consider different residual computations to estimate cache metric. Our default is Temp-attn." + ], + "table_body": "
ResidualVBenchLatency
pt(TA)83.4089.53
pt(SA)83.1989.06
qt(CA)83.2590.70
rt(MLP)83.6299.72
", + "bbox": [ + 462, + 239, + 627, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e287f0fb9b394a35799d4c13ccde5203c58c81b110f19438a8efb8ab6c6f52df.jpg", + "table_caption": [], + "table_footnote": [ + "(f) AdaCache Variants: We achieve a range of speedups (and quality) by controlling the basis cache-rates in AdaCache. Our default setting is AdaCache-fast." + ], + "table_body": "
AdaCacheBasis-ratesVBranchLatency
Fast{12, 10, 8, 6, 4, 3}83.4089.53
Medium{8, 6, 4, 2, 1}83.94143.87
Slow{2, 1}84.12274.30
", + "bbox": [ + 638, + 251, + 905, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in reference-based scores (e.g. PSNR, SSIM) at extreme reduction rates, the qualitative results suggest that the generations are still good (see Fig. 1), despite not being aligned exactly with the reference.", + "bbox": [ + 89, + 406, + 482, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AdaCache with Motion Regularization: We compare AdaCache with different versions of motion regularization in Table 2a. Both vanilla and motion-regularized versions provide significant speedups, $4.7 \\times$ and $4.5 \\times$ respectively, at a comparable quality with baseline Open-Sora [120]. Considering motion-gradient as an early-predictor of motion at latter diffusion steps helps (83.50 vs. 83.36 on VBench). We also estimate motion at different frame-rates by considering a varying step-size in frame differences, which seems to increase the latency without necessarily improving quality. Overall, we consider AdaCache (w/ MoReg) as the configuration with the best quality-latency trade-off. This improvement in quality is more-prominent in qualitative examples in supplementary (Fig. A.2) and the user study (Fig. 4).", + "bbox": [ + 88, + 472, + 482, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acceleration in multi-GPU setups: Aligned with prior work based on Dynamic Sequence Parallelism (DSP) [117], we evaluate how AdaCache performs in video generation across multiple GPUs. This evaluation is relevant in the context of efficiency, as DSP incurs additional latency overheads corresponding to the communication between GPUs, and caching mechanisms can avoid such costs by re-using previous computations. We present the results of this study in Fig. 6 and Table 2. Here, we consider Open-Sora [120] (480p - 2s at 30-steps) and Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) as baselines, and compare against prior-art PAB [118] in terms of latency measurements on A100 GPUs. In Fig. 6, we observe that AdaCache consistently outperforms PAB with better speeds across all settings.", + "bbox": [ + 89, + 689, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97e97b168c1ccc16708bd416f339ce78d11df4cc358a28bc7e7ebc82c88e502c.jpg", + "image_caption": [ + "Figure 6. Acceleration in multi-GPU setups: We evaluate the speedups with varying GPU parallelization, as cached-steps can avoid communication overheads among GPUs. Here, we compare AdaCache with PAB [118], on baselines Open-Sora [120] 480p - 2s generations at 30-steps and Open-Sora-Plan [45] $512 \\times 512 - 2.7$ s generations at 150-steps. AdaCache consistently shows better acceleration over PAB in all settings. Latency measured on A100s." + ], + "image_footnote": [], + "bbox": [ + 529, + 404, + 705, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c9e73ca666ddaa2a3a50d011b634ac4bd56d67716e923dfae106b65cf5d94d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 405, + 885, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speedups at different resolutions: In Table 2b, we compare the trade-offs of AdaCache at various resolutions of video generations, namely, $480\\mathrm{p} - 2\\mathrm{s}$ , $480\\mathrm{p} - 4\\mathrm{s}$ and $720\\mathrm{p} - 2\\mathrm{s}$ , all at 100-steps. AdaCache provides consistent speedups across different resolutions without affecting the quality.", + "bbox": [ + 511, + 638, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cache metric, location and residual: When adaptively deciding the caching schedule, we consider different metrics to compute the rate-of-change between representations, namely, L1/L2 distance or cosine distance. Among these, L1/L2 give an absolute measure which aligns better with the actual change. In contrast, cosine computes a normalized-distance, which is not a good estimate of change (e.g. if the representations differ only by a scale, the distance will be zero, even though we want to have a non-zero value). This observation is verified by the results in Table 2c. Moreover, we consider computing the cache metric at various locations (i.e., layers) in the DiT. Doing so at a single layer (e.g. start,", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15246", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e27b05116571f57b6efd690ef14d7ad54262733746df3acb6b3f30030a325a1e.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison: We show qualitative results on multiple video-DiT baselines including Open-Sora [120] (720p - 2s at 100-steps), Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) and Latte [62] (512×512 - 2s at 50-steps), while comparing against prior training-free inference acceleration method PAB [118]. AdaCache shows a comparable generation quality at much-faster speeds. Best-viewed with zoom-in. Prompts and additional qualitative results (Fig. A.3) are given in supplementary." + ], + "image_footnote": [], + "bbox": [ + 174, + 87, + 467, + 417 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83e6469b73c482a305bea3622859285adee5ba791e51a376c4f4b23989aeb597.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 470, + 88, + 812, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mid, end) is not significantly different from computing an aggregate over multiple-layers (see Table 2d). By default, we compute the cache metric in the mid-layer as a reasonable choice without extra overheads. We choose temporal-attention as the residual for cache metric computation, since it achieves the best trade-off (see Table 2e).", + "bbox": [ + 89, + 489, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AdaCache variants: To achieve a range of speedups (and quality), we consider different basis cache-rates in our AdaCache implementation. For instance, we can have higher-speedup with a slightly-lower quality (e.g. AdaCache-fast), a lower-speedup with a higher-quality (e.g. AdaCacheslow), or balance both (e.g. AdaCache-medium). We can conveniently control this by having corresponding basis cache-rates as shown in Table 2f. By default, we resort to AdaCache-fast which gives the best speedups.", + "bbox": [ + 89, + 584, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 727, + 267, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we present qualitative results on multiple video DiT baselines, including Open-Sora [120], Open-Sora-Plan [45] and Latte [62]. We compare AdaCache against each baseline and prior training-free inference acceleration method for video DiTs, PAB [118]. Here, we consider three different configurations: $720\\mathrm{p}$ - 2s generations at 100-steps for Open-Sora, $512\\times 512$ - 2.7s generations at 150-steps for Open-Sora-Plan, and $512\\times 512$ - 2s generations at 50-steps for Latte, while considering prompts from OpenSora gallery (see supplementary for prompt details). Ada", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cache shows a comparable generation quality, while having much-faster inference pipelines. In fact, it achieves $4.49 \\times$ (vs. $1.26 \\times$ in PAB), $3.53 \\times$ (vs. $1.45 \\times$ in PAB), $2.46 \\times$ (vs. $1.33 \\times$ in PAB) speedups respectively on the three considered baseline video DiTs. In most cases our generations are aligned well with the baseline in the pixel space. Yet this is not a strict requirement, as the denoising process can deviate considerably from that of the baseline, particularly at high caching rates. Still, AdaCache is faithful to the text prompt and is not affected by significant artifacts. Refer supplementary (Fig. A.3) for more qualitative examples.", + "bbox": [ + 511, + 489, + 906, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Adaptive Caching (AdaCache), a plug-and-play component that improves the inference speed of video diffusion transformers without needing any re-training. It caches residual computations, while also devising the caching schedule dependent on each video generation. We further proposed a Motion Regularization (MoReg) to utilize video information and allocate computations based on motion content, improving the quality-latency trade-off. We apply our contributions on multiple open-source video DiTs, showing comparable generation quality at a fraction of latency. We believe AdaCache is widely-applicable with minimal effort, helping democratize high-fidelity long-video generation.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15247", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to thank Ankit Khedia, Arvind Somasundaram, Dustin Johnson, Eugene Vecharynski, Ly Cao, SK Bong, and Yuzi He for the interesting discussions, and Shikun Liu for the support in setting-up the project page. The authors also appreciate the time and effort volunteered by the participants of the user study.", + "bbox": [ + 89, + 113, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 219, + 187, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shady Abu-Hussein, Tom Tirer, and Raja Giryes. Adir: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221, 2022. 2", + "[2] Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Albert Pumarola, Ali Thabet, and Edgar Schonfeld. Flexidit: Your diffusion transformer can easily generate high-quality samples with less compute. arXiv preprint arXiv:2502.20126, 2025. 2", + "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2", + "[4] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 2", + "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2", + "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2", + "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2", + "[8] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599-4603, 2023. 1, 2", + "[9] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1", + "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + ], + "bbox": [ + 99, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2", + "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2", + "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2", + "[14] Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. 2", + "[15] Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. Delta-dit: A training-free acceleration method tailored for diffusion transformers. arXiv preprint arXiv:2406.01125, 2024. 2, 5, 6", + "[16] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024. 2", + "[17] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[18] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint arXiv:1907.06571, 2019. 2", + "[19] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2", + "[20] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2", + "[21] Juncan Deng, Shuaiying Li, Zeyu Wang, Hong Gu, Kedong Xu, and Kejie Huang. Vq4dit: Efficient post-training vector quantization for diffusion transformers. arXiv preprint arXiv:2408.17131, 2024. 2", + "[22] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 2" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15248", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. arXiv preprint arXiv:2403.14611, 2024. 2", + "[24] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 2", + "[25] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2", + "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1", + "[27] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 2", + "[28] Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8352-8361, 2024. 1, 2", + "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2", + "[30] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[31] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streaming2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 2", + "[32] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1", + "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[34] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2", + "[37] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In International Conference on Machine Learning, pages 13916-13932. PMLR, 2023. 1", + "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 5, 6, 7", + "[39] Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. arXiv preprint arXiv:2401.05735, 2024. 1, 2", + "[40] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1", + "[41] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1", + "[42] Inc. Kling AI. Next-generation ai creative studio. URL https://klingai.com/, 2024. 2", + "[43] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2", + "[44] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 1", + "[45] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 5,6,7,8", + "[46] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 2", + "[47] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 1, 2", + "[48] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + ], + "bbox": [ + 522, + 93, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15249", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 2", + "[50] Joseph Liu, Joshua Geddes, Ziyu Guo, Haomiao Jiang, and Mahesh Kumar Nandwana. Smoothcache: A universal inference acceleration technique for diffusion transformers. arXiv preprint arXiv:2411.10510, 2024. 2", + "[51] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2", + "[52] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 1", + "[53] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 1", + "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2", + "[56] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311, 2023. 2", + "[57] Inc. Luma AI. Dream machine. URL https://lumalabs.ai/dream-machine, 2024.2", + "[58] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2", + "[59] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 2", + "[60] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1", + "[61] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 1, 2", + "[62] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 5, 6, 8", + "[63] Yifang Men, Yuan Yao, Miaomiao Cui, and Bo Liefeng. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 2", + "[64] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2", + "[65] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. arXiv preprint arXiv:2404.12382, 2024. 1, 2", + "[66] Inc. OpenAI. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2", + "[67] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 2, 3", + "[68] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[69] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1", + "[70] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023. 1", + "[71] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2", + "[72] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 1", + "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3", + "[74] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2", + "[75] Inc. Runway AI. Gen-3 alpha: A new frontier for video generation. URL https://runwayml.com/research/introducing-gen-3-alpha, 2024.2", + "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15250", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[77] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2", + "[78] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2", + "[79] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2", + "[80] Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425, 2024. 2", + "[81] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 2", + "[82] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2", + "[83] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2", + "[84] Nikita Starodubcev, Dmitry Baranchuk, Artem Fedorov, and Artem Babenko. Your student is better than expected: Adaptive teacher-student collaboration for text-conditional diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275–9285, 2024. 2", + "[85] Maitreya Suin, Nithin Gopalakrishnan Nair, Chun Pong Lau, Vishal M Patel, and Rama Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6343-6352, 2024. 2", + "[86] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3", + "[87] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 2", + "[88] Shengkun Tang, Yaqing Wang, Caiwen Ding, Yi Liang, Yao Li, and Dongkuan Xu. Deediff: Dynamic uncertainty-aware early exiting for accelerating diffusion model generation. arXiv preprint arXiv:2309.17074, 2023. 2", + "[89] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2" + ], + "bbox": [ + 101, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[90] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2", + "[91] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2", + "[92] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 2", + "[93] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 2", + "[94] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 2", + "[95] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2", + "[96] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. arXiv preprint arXiv:2408.15239, 2024. 2", + "[97] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 3", + "[98] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 1, 2", + "[99] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1", + "[100] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 2", + "[101] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu," + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15251", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2", + "[102] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2", + "[103] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2", + "[104] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2", + "[105] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2", + "[106] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2", + "[107] Xingyi Yang and Xinchao Wang. Compositional video generation as flow equalization. arXiv preprint arXiv:2407.06182, 2024. 2", + "[108] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 2", + "[109] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 2", + "[110] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 2", + "[111] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 1", + "[112] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2", + "[113] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747v1, 2024.5, 6", + "[114] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffu" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2", + "[115] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2", + "[116] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2", + "[117] Xuanlei Zhao, Shenggan Cheng, Zangwei Zheng, Zheming Yang, Ziming Liu, and Yang You. Dsp: Dynamic sequence parallelism for multi-dimensional transformers. arXiv preprint arXiv:2403.10266, 2024. 7", + "[118] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 2, 5, 6, 7, 8", + "[119] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 1, 2", + "[120] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3, 5, 6, 7, 8", + "[121] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 530 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "15252", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_model.json b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d182a3a6b2bb04609cb94dc86ecd7c15ead9b3e4 --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_model.json @@ -0,0 +1,2899 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.127, + 0.882, + 0.15 + ], + "angle": 0, + "content": "Adaptive Caching for Faster Video Generation with Diffusion Transformers" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.165, + 0.787, + 0.245 + ], + "angle": 0, + "content": "Kumara Kahatapitiya\\(^{1,2*}\\) Haozhe Liu\\(^{1}\\) Sen He\\(^{1}\\) Ding Liu\\(^{1}\\) Menglin Jia\\(^{1}\\) Chenyang Zhang\\(^{1}\\) Michael S. Ryoo\\(^{2}\\) Tian Xie\\(^{1}\\) Meta AI \\(^{2}\\)Stony Brook University adacache-dit.github.io" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.633 + ], + "angle": 0, + "content": "Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs)—despite making significant headway in this context—have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that “not all videos are created equal”: meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to \\(4.7 \\times\\) on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.222, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.847 + ], + "angle": 0, + "content": "Diffusion models [33, 82] have become the standard for generative modeling in recent years, arguably surpassing the quality of VAEs [41, 72], GANs [26, 40] and Auto-Regressive models [9, 10]. This observation holds in a wide-range of applications including image [73, 76], video [6, 81], 3D [52, 69], and audio [37, 44] generation, as well as image [4, 32] and video [70, 99] editing. More recent Diffusion Transformers (DiTs) [60, 67] show better promise in terms of scalability and generalization compared to prior UNet-based diffusion models [73], revealing intriguing horizons in GenAI for the years to come." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.878 + ], + "angle": 0, + "content": "Despite the state-of-the-art performance, DiTs can also be computationally expensive both in terms of memory and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.268, + 0.712, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.267, + 0.906, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.635 + ], + "angle": 0, + "content": "Figure 1. Effectiveness of Adaptive Caching: We show a qualitative comparison of AdaCache (right) applied on top of Open-Sora [120] (left), a baseline video DiT. Here, we consider generating \\(720\\mathrm{p}\\) - 2s video clips, and report VBench [38] quality and average latency (on a single A100 GPU) on the benchmark prompts from Open-Sora [120] gallery. AdaCache generates videos significantly faster (i.e., \\(4.7\\times\\) speedup) with a comparable quality. Also, the number of computed steps varies for each video. Best-viewed with zoom-in. Prompts given in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.867 + ], + "angle": 0, + "content": "computational requirements. This becomes especially critical when applied with a large number of input tokens (e.g. high-resolution long video generation). For instance, the reason for models such as Sora [66] not being publicly served is speculated to be the high resource demands and slower inference speeds [54]. To tackle these challenges and reduce the footprint of diffusion models, various research directions have emerged such as latent diffusion [73], step-distillation [79, 111], caching [28, 61, 98], architecture-search [48, 119], token reduction [8, 47] and region-based methods [39, 65]. Fewer techniques transfer readily from UNet-based pipelines to DiTs, whereas others often require novel formulations. Hence, DiT acceleration has been under-explored as of yet." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Moreover, we note that not all videos are created equal. Some videos contain high-frequency textures and signifi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.221, + 0.899 + ], + "angle": 0, + "content": "*Work done at Meta" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15240" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "cant motion content, whereas others are much simpler (e.g. with homogeneous textures or static regions). Having a diffusion process tailored specifically for each video generation can be beneficial in terms of realizing the best quality-latency trade-off. This idea has been explored to some extent in region-based methods [4, 39, 65], but not sufficiently in the context of video generation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.199, + 0.487, + 0.576 + ], + "angle": 0, + "content": "Motivated by the above, we introduce Adaptive Caching (AdaCache) for accelerating video diffusion transformers. This approach requires no training and can seamlessly be integrated into a baseline video DiT at inference, as a plug-and-play component. The core idea of our proposal is to cache residual computations within transformer blocks (e.g. attention or MLP outputs) in a certain diffusion step, and reuse them through a number of subsequent steps that is dependent on the generated video. We do this by devising a caching schedule, i.e., deciding when-to-recompute next whenever making a residual computation. This decision is guided by a distance metric that measures the rate-of-change between previously-stored and current representations. If the distance is high we would not cache for an extended period (i.e., #steps), to avoid reusing incompatible representations. We further introduce a Motion Regularization (MoReg) to allocate computations based on the motion content in the video being generated. This is inspired by the observation that high-moving sequences require more diffusion steps to achieve a reasonable quality. Altogether, our pipeline is applied on top of multiple video DiT baselines showing much-faster inference speeds without sacrificing the generation quality (see Fig. 1). Finally, we validate the effectiveness of our contributions and justify our design decisions through ablations and qualitative comparisons." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.233, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion-based Video Generation [6, 11, 25, 34, 81] has surpassed the quality and diversity of GAN-based approaches [18, 77, 89, 91, 112], while also being competitive with recent Auto-Regressive models [35, 43, 49, 90, 101, 105]. They have become a standard component in the pipelines for frame interpolation [23, 96], video outpainting [16, 22, 93], image-to-video [6, 27, 103], video-to-video (i.e., video editing or translation) [36, 106, 109], personalization [63, 100], motion customization [104, 116] and compositional generation [51, 107]. The underlying architecture of video diffusion models has evolved from classical UNets [73, 74] with additional spatio-temporal attention layers [7, 17, 25, 29], to fully-fledged transformer-based (i.e., DiT [67]) architectures [24, 56, 62, 114]. In the process, the latency of denoising [55, 82] has also scaled with larger models [24, 68]. This becomes critical especially in applications such as long-video generation [31, 87, 92, 110, 115, 121], while also affecting the growth of commercially-served video models [42, 57, 66, 75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Efficiency of Diffusion models has been actively explored with respect to both training and inference pipelines. Multi-stage training at varying resolutions [12, 13, 24] and high-quality data curation [6, 20, 34, 71] have cut down training costs significantly. In terms of inference acceleration, there exist two main approaches: (1) methods that require re-training such as step-distillation [53, 64, 78, 79], consistency regularization [58, 83], quantization [14, 21, 30, 46, 94], and architecture search/compression [48, 108, 119], or (2) methods that require no re-training such as token reduction [8, 39, 47] and caching [15, 28, 61, 98, 118]. Among these, training-free methods are more-attractive as they can be widely-adopted without any additional costs. This becomes especially relevant for video diffusion models that are both expensive to train and usually very slow at inference. In this paper, we explore a caching-based approach tailored for video DiTs. Different from prior fixed caching schedules in UNet-based [28, 61, 98] and DiT-based [2, 15, 80, 118] pipelines, we introduce a content-dependent (i.e., adaptive) caching scheme to squeeze out the best quality-latency trade-off." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.909, + 0.686 + ], + "angle": 0, + "content": "Content-adaptive Generation may focus on improving consistency [3-5, 19, 95, 102], quality [1, 85], and/or efficiency [39, 65, 84, 88]. Most region-based methods (e.g. image or video editing) rely on a user-provided mask to ensue consistent generations aligned with context information [4, 102]. Some others automatically detect [85] or retrieve [1] useful information to improve generation quality. Among efficiency-oriented approaches, there exist proposals for selectively-processing a subset of latents [39, 65], switching between diffusion models with varying compute budgets [84], or adaptively-controlling the number of denoising steps [50, 59, 88, 98]. Our proposed AdaCache—which also controls #denoising-steps adaptively—provides better flexibility in the sense that (1) it can selectively-cache any layer or even just a specific module within a layer, and (2) it is tailored to each video generation instead of being fixed for a given architecture, enabling a better-adaptive compute allocation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.813, + 0.723 + ], + "angle": 0, + "content": "3. Not All Videos Are Created Equal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.793 + ], + "angle": 0, + "content": "In this section, we motivate the need for a content-dependent denoising process, and show how it can help maximize the quality-latency trade-off. This motivation is based on a couple of interesting observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "First, we note that each video is unique. Hence, videos have varying levels of complexity. Here, the complexity of a given video can be expressed by the rate-of-change of information across both space and time. Simpler videos may contain more homogeneous regions and/or static content. In contrast, complex videos have more high-frequency details and/or significant motion. The standard video compres" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.111, + 0.353, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.137, + 0.254, + 0.321, + 0.263 + ], + "angle": 0, + "content": "(a) Critical number of denoising-steps varies." + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.09, + 0.631, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.254, + 0.62, + 0.263 + ], + "angle": 0, + "content": "(b) Changes in subsequent representations are unique." + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.109, + 0.906, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.254, + 0.895, + 0.263 + ], + "angle": 0, + "content": "(c) Different levels of quality at a given latency budget." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.268, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Not all videos are created equal: We show frames from \\(720\\mathrm{p}\\) - 2s video generations based on Open-Sora [120]. (a) We try to break each generation by reducing the number of diffusion steps. Interestingly, not all videos have the same break point. Some sequences are extremely robust (e.g. first column), while others break easily. (b) When we plot the difference between computed representations in subsequent diffusion steps, we see unique variations (Feature distance vs. #steps). If we are to reuse similar representations across denoising steps, it needs to be tailored to each video. (c) There exist different configurations for generating videos at an approximately-fixed latency (e.g. having an arbitrary #denoising-steps, yet only computing a fixed #representations and reusing otherwise). We observe a significant variance in quality in such videos. All these observations suggest the need for a content-dependent denoising process, which is the founding motivation of Adaptive Caching. Best-viewed with zoom-in. Prompts given in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.484, + 0.589 + ], + "angle": 0, + "content": "sion techniques exploit such information to achieve the best possible compression ratios without sacrificing the quality [86, 97]. Motivated by the same, we explore how the compute-cost affects the quality of video generations based on DiTs. We measure this w.r.t. the number of denoising steps, and the observations are shown in Fig. 2 (a). Some video sequences are very robust, and achieve a reasonable quality even at fewer denoising steps. Others break easily when we keep reducing the #steps, but the break-point varies. This suggests that the minimal #steps (or, computations) required to generate a video with a reasonable quality varies, and having a content-dependent denoising schedule can exploit this for speeding-up generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Next, we observe how the computed representations (i.e., residual connections in attention or MLP blocks within DiT) change during the denoising process, across different video generations. This may reveal the level of compute redundancy in each video generation, enabling us to reuse representations and improve efficiency. More specifically, we visualize the feature differences between subsequent diffusion steps as histograms given in Fig. 2 (b). Here, we report Feature distance (e.g. L1) vs. #steps. We observe that each histogram is unique. Despite having higher changes in early/latter steps and smaller changes in the middle, the overall distribution and the absolute values vary considerably. A smaller change corresponds to higher redundancy across subsequent computations, and an opportunity for reusing. This motivates the need for a non-uniform compute-schedule not only within the diffusion process of a given video, but also across different videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Finally, we evaluate the video generation quality at a capped-budget (i.e., fixed computations or latency). We can have multiple generation configurations at an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.543 + ], + "angle": 0, + "content": "approximately-fixed latency, by computing a constant number of representations. For instance, we can cache and reuse representations more-frequently in a setup with more denoising steps, still having the same latency of a process with fewer steps. The observations of a study with either 30 or 100 base denoising steps is shown in Fig. 2 (c). We see that the generation quality varies significantly despite spending a similar cost and having the same underlying pretrained DiT. This motivates us to think about how best to allocate resources at inference, tailored for each video generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.874, + 0.59 + ], + "angle": 0, + "content": "4. Adaptive Caching for Faster Video DiTs" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.896, + 0.618 + ], + "angle": 0, + "content": "4.1. Preliminaries: Video Diffusion Transformers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Video Diffusion Transformers are extended from Latent Diffusion Transformers (DiTs) [67] introduced for image generation. DiTs provide a much-more streamlined, scalable architecture compared to prior UNet-based diffusion models [73], by only having transformer blocks with a homogeneous token resolution (instead of convolutional blocks with up/downsampling). A simplified transformer block (i.e., w/o normalizing or timestep conditioning layers) in a video DiT is shown in Fig. 3 (right)—gray block. It consists of spatial-temporal attention (STA), cross-attention (CA) and linear (MLP) layers. Depending on the implementation, STA may be a single joint spatio-temporal attention layer, or separate spatial and temporal attention layers repeated within alternating blocks. Without loss of generality, let us denote a latent feature at the input/output of such block by \\( f_{t}^{l} \\) and \\( f_{t}^{l + 1} \\), respectively. Here, \\( l \\) represents the layer index, and \\( t \\), the diffusion timestep. A simplified flow of computations within each block can be represented as," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15242" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.168, + 0.111, + 0.483, + 0.13 + ], + "angle": 0, + "content": "\\[\np _ {t} ^ {l} = \\operatorname {S T A} \\left(f _ {t} ^ {l}\\right); \\quad \\tilde {f} _ {t} ^ {l} = f _ {t} ^ {l} + p _ {t} ^ {l}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.133, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nq _ {t} ^ {l} = \\mathrm {C A} \\left(\\tilde {f} _ {t} ^ {l}\\right) \\quad ; \\quad \\bar {f} _ {t} ^ {l} = \\tilde {f} _ {t} ^ {l} + q _ {t} ^ {l}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.154, + 0.483, + 0.173 + ], + "angle": 0, + "content": "\\[\nr _ {t} ^ {l} = \\mathrm {M L P} \\left(\\bar {f} _ {t} ^ {l}\\right); f _ {t} ^ {l + 1} = \\bar {f} _ {t} ^ {l} + r _ {t} ^ {l}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.484, + 0.271 + ], + "angle": 0, + "content": "Here \\( p_t^l \\), \\( q_t^l \\) and \\( r_t^l \\) are residual connections corresponding to each compute-element. Such computations repeat through \\( L \\) layers, generating the noise prediction of each step \\( t \\), and across a total of \\( T \\) denoising steps. In the current streamlined video DiT architectures with homogeneous token resolutions, each layer or denoising step costs the same." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.277, + 0.267, + 0.293 + ], + "angle": 0, + "content": "4.2. Adaptive Caching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.571 + ], + "angle": 0, + "content": "In this subsection, we introduce Adaptive Caching (AdaCache), a training-free mechanism for content-dependent compute allocation in video DiTs. The overview of Adaptive Caching is shown in Fig. 3. Compared to a standard DiT that computes representations for all layers across all diffusion steps, in AdaCache, we decide which layers or steps to compute, adaptively (i.e., dependent on each video being generated). This decision is based on the rate-of-change in the residual connections (e.g. \\( p_t^l \\), \\( q_t^l \\) or \\( r_t^l \\)) across diffusion steps, which amount to all significant computations within the DiT. Without loss of generality, let us assume that the residuals in block \\( l \\) in current and immediately-prior diffusion steps \\( t \\) and \\( t + k \\) are already computed. Here, step \\( t + k \\) is identified as 'immediately-prior' to step \\( t \\) since any residuals between these two steps are assumed to be not computed (i.e., cached residuals are reused). We make a decision on the next computation step based on the distance metric \\( (c_t^l) \\) given by," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.576, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\nc _ {t} ^ {l} = \\operatorname {d i s t} \\left(p _ {t + k} ^ {l}, p _ {t} ^ {l}\\right) = \\left\\| p _ {t} ^ {l} - p _ {t + k} ^ {l} \\right\\| / k. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.484, + 0.812 + ], + "angle": 0, + "content": "Here, we use L1 distance by default, but other distance metrics can also be applied (e.g. L2, cosine). Once we have the distance metric, we select the next caching rate \\((\\tau_t^l)\\) based on a pre-defined codebook of basis cache-rates. Here, a 'cache-rate' is defined as the number subsequent steps during which, a previously-computed representation is re-used (i.e., a higher cache-rate gives more compute savings). The codebook is basically a collection of cache-rates defined based on the original denoising schedule (i.e., #steps), coupled with corresponding metric thresholds to select them. Simply put, a higher distance metric will sample a lower cache-rate from the codebook, resulting in more-frequent re-computations. More details about the codebook is available in the supplementary (Sec. A.2)." + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.817, + 0.483, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\tau_ {t} ^ {l} = \\operatorname {c o d e b o o k} \\left(c _ {t} ^ {l}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For all denoising steps within \\( t \\) and \\( t - \\tau \\), we reuse previously-cached representations and only recompute after the current caching schedule (while also estimating the metric, again)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.907, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.39 + ], + "angle": 0, + "content": "Figure 3. Overview of Adaptive Caching: (Left) During the diffusion process, we choose to cache residual computations within selected DiT blocks. The caching schedule is content-dependent, as we decide when to compute the next representation based on a distance metric \\((c_{t})\\). This metric measures the rate-of-change from previously-computed (and, stored) representation to the current one, and can be evaluated per-layer or the DiT as-a-whole. Each computed residual can be cached and reused across multiple steps. (Right) We only cache the residuals (i.e., skip-connections) which amount to the actual computations (e.g. spatial-temporal/cross attention, MLP). The iteratively denoised representation (i.e., \\(f_{t+k}\\), \\(f_{t}\\)) always gets updated either with computed or cached residuals." + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.415, + 0.907, + 0.455 + ], + "angle": 0, + "content": "\\[\np _ {t - k} ^ {l} = \\left\\{ \\begin{array}{l l} p _ {t} ^ {l} & \\text {i f} k < \\tau_ {t} ^ {l}; \\\\ p _ {t - k} ^ {l} = \\operatorname {S T A} (f _ {t - k} ^ {l}) & \\text {i f} k = \\tau_ {t} ^ {l}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.629 + ], + "angle": 0, + "content": "The same applies to other residual computations (e.g. \\( q_{t - k}^{l}, r_{t - k}^{l} \\)) as well. By design, we can have unique caching schedules for each layer (and, each residual computation). However, we observe that it will make the generations unstable. Therefore, we decide to have a common metric (i.e., \\( c_t^l = c_t \\)) and hence, a common caching rate (i.e., \\( \\tau_t^l = \\tau_t \\)) across all DiT layers. For instance, we can consider an averaged metric across all layers, or a metric computed at a certain layer to decide the caching schedule. Meaning, when we recompute residuals in a certain step, we do so for the whole DiT rather than selectively for each layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.906, + 0.735 + ], + "angle": 0, + "content": "Overall, this setup allows us to adaptively-control the compute spent on each video generation, based on framewise information (i.e., no temporal information used as of yet). If the rate-of-change between residuals is high, we will have a smaller caching rate, and otherwise, we have a higher rate. The choice of a lightweight distance metric (e.g. L1) helps us avoid any additional latency overheads." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.725, + 0.759 + ], + "angle": 0, + "content": "4.3. Motion Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To further improve Adaptive Caching by making use of video-specific (i.e., temporal) information, we introduce a Motion Regularization (MoReg). This is motivated by the observation that the optimal number of denoising steps varies based on the motion content of each generated video. The core idea here, is to cache less (i.e., recompute more) if a generated video has a high motion content. Simply put, we plan to regularize our caching schedule based on motion. However the problem is that, we need to estimate mo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15243" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "tion while the video is still being generated. Therefore, we can not rely on motion estimation algorithms in the pixel space, nor any compute-heavy ones as our focus is on efficiency. As a result, we estimate a noisy latent motion-score \\((m_t^l)\\) based on residual frame differences. Without loss of generality, let us denote residual latent frames of \\(p_t^l\\) as \\(\\{p_{t,n}^l\\mid n = 0,\\dots ,N - 1\\}\\) where \\(N\\) is the #frames in latent space (generated by the VAE encoder). We estimate the motion-score as," + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.235, + 0.483, + 0.255 + ], + "angle": 0, + "content": "\\[\nm _ {t} ^ {l} = \\left\\| p _ {t, i: N} ^ {l} - p _ {t, 0: N - i} ^ {l} \\right\\|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Here, \\(i\\) denotes the frame step-size (or, frame-rate), \\(\\| \\cdot \\|\\), the distance metric (e.g. L1), and \\(i:j\\), the slice of all frames within the corresponding range. However, since we operate on noisy-latents, we observe that our motion estimate, particularly in early diffusion steps is not reliable. Meaning, it does not provide a reasonable regularization in early steps (i.e., the change in caching schedule does not correlate well with the observed motion in pixel space). To alleviate this, we also compute a motion-gradient \\((mg_t^l)\\) across diffusion steps, which can act as a reasonable early-predictor of motion that we may observe in latter diffusion steps (that also correlates with the motion in pixel space)." + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.462, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\nm g _ {t} ^ {l} = \\left(m _ {t} ^ {l} - m _ {t + k} ^ {l}\\right) / k. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Despite the motion-score being noisy, the motion-gradient acts as a better-estimate of trend, as the representations are getting denoised and converging to a noise-free distribution. Finally, we use both motion-score and motion-gradient as a scaling-factor of the distance metric \\((c_t^l)\\) to regularize our caching schedule." + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.598, + 0.483, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {c} _ {t} ^ {l} = c _ {t} ^ {l} \\cdot \\left(m _ {t} ^ {l} + m g _ {t} ^ {l}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.78 + ], + "angle": 0, + "content": "This means, when we have a higher estimated motion, the distance metric will be increased and a smaller basis cache-rate will be selected from the codebook. As previously discussed, we also enforce a common motion-regularization in all DiT layers by computing a common motion score (i.e., \\( m_t^l = m_t \\), \\( mg_t^l = mg_t \\)), ensuring the stability of the denoising process. We can also choose to compute motion at different frame-rates, which we ablate in our experiments. Refer to the supplementary for concrete examples of motion-score and motion-gradient (Fig. A.1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.224, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.305, + 0.835 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We select multiple prominent open-source video DiTs as backbone video generation pipelines in our experiments, namely, Open-Sora-v1.2 [120] for Text+Image-to-Video (TI2V) generation, and Open-Sora-Plan-v1.1 [45] and Latte" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.395 + ], + "angle": 0, + "content": "[62] for Text-to-Video (T2V) generation. Since we focus on inference-based latency optimizations (i.e., without any re-training), we compare AdaCache against similar methods such as \\(\\Delta\\)-DiT [15], T-GATE [113] and PAB [118]. In our main experiments, we generate \\(900+\\) videos based on standard VBenchmark [38] benchmark prompts, at the corresponding baseline generation settings suggested in [118] for a fair comparison (e.g. \\(480p\\) - 2s with 30-steps in Open-Sora, \\(512\\times 512\\) - 2.7s with 150-steps in Open-Sora-Plan and \\(512\\times 512\\) - 2s with 50-steps in Latte). We report VBenchmark average and reference-based PSNR, SSIM and LPIPS as quality metrics, and also report FLOPs, Latency (s) and Speedup as complexity metrics. Here, Latency is measured on a single A100 GPU. Unless otherwise stated, in our ablations and qualitative results, we experiment on the prompts from Open-Sora benchmark gallery, generating \\(720p\\) - 2s videos with 100-steps, following the recommendation of model authors—which gives better quality video generations (and, higher speedups) compared to the benchmark setting discussed above (i.e., \\(480p\\) - 2s with 30-steps)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.406, + 0.646, + 0.42 + ], + "angle": 0, + "content": "5.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.791 + ], + "angle": 0, + "content": "In Table 1, we present a quantitative evaluation of quality and latency on VBench [38] benchmark. We consider three variants of AdaCache: a slow variant, a fast variant with more speedup and the same with motion regularization. We compare with other training-free acceleration methods, showing consistently better speedups with a comparable generation quality. With Open-Sora [120] baseline, AdaCache-slow outperforms others on all quality metrics, while giving a \\(1.46 \\times\\) speedup compared to PAB [118] with \\(1.20 \\times\\) speedup. AdaCache-fast gives the highest acceleration of \\(2.24 \\times\\) with a slight drop in quality. AdaCache-fast (w/ MoReg) shows a clear improvement in quality compared to AdaCache-fast, validating the effectiveness of our regularization and giving a comparable speedup of \\(2.10 \\times\\). All AdaCache variants outperform even the baseline (w/o any acceleration) on VBench average quality, which aligns better with human preference compared to other reference-based metrics. Similar observations hold with the other baselines as well. With Open-Sora-Plan [45], AdaCache shows the best speedup of \\(3.70 \\times\\) compared to the previous-best \\(1.45 \\times\\) of PAB, and the best quality with a \\(2.20 \\times\\) speedup. With Latte [62], we gain the best speedup of \\(2.74 \\times\\) compared to prior \\(1.33 \\times\\), and the best overall quality with a \\(1.59 \\times\\) speedup." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "User study: Quantitative metrics on video generation quality can sometimes fall-short in aligning with the perceived visual quality. To better understand the human preference on AdaCache and its comparisons, we conduct a user study in the form of randomized A/B preference tests. Here, we create a questionnaire with 50 multiple-choice questions, each consisting of 3 variants of a single video sequence:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15244" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.186, + 0.089, + 0.812, + 0.37 + ], + "angle": 0, + "content": "
MethodVBatch (%) ↑PSNR ↑LPIPS ↓SSIM ↑FLOPs (T)Latency (s)Speedup
Open-Sora [120]79.22---3230.2454.021.00×
+ Δ-DiT [15]78.2111.910.56920.48113166.47--
+ T-GATE [113]77.6115.500.34950.67602818.4049.111.10×
+ PAB-fast [118]76.9523.580.17430.82202558.2540.231.34×
+ PAB-slow [118]78.5127.040.09250.88472657.7044.931.20×
+ AdaCache-fast79.3924.920.09810.83751331.9724.162.24×
+ AdaCache-fast (w/ MoReg)79.4825.780.08670.85301383.6625.712.10×
+ AdaCache-slow79.6629.970.04560.90852195.5037.011.46×
Open-Sora-Plan [45]80.39---12032.40129.671.00×
+ Δ-DiT [15]77.5513.850.53880.373612027.72--
+ T-GATE [113]80.1518.320.30660.621910663.32113.751.14×
+ PAB-fast [118]71.8115.470.54990.47178551.2689.561.45×
+ PAB-slow [118]80.3018.800.30590.65509276.5798.501.32×
+ AdaCache-fast75.8313.530.54650.43093283.6035.043.70×
+ AdaCache-fast (w/ MoReg)79.3017.690.37450.61473473.6836.773.53×
+ AdaCache-slow80.5022.980.17370.79104983.3058.882.20×
Latte [62]77.40---3439.4732.451.00×
+ Δ-DiT [15]52.008.650.85130.10783437.33--
+ T-GATE [113]75.4219.550.26120.69273059.0229.231.11×
+ PAB-fast [118]73.1317.160.39030.64212576.7724.331.33×
+ PAB-slow [118]76.3219.710.26990.70142767.2226.201.24×
+ AdaCache-fast76.2617.700.35220.66591010.3311.852.74×
+ AdaCache-fast (w/ MoReg)76.4718.160.32220.68321187.3113.202.46×
+ AdaCache-slow77.0722.780.17370.80302023.6520.351.59×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.373, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of quality and latency: Here, we compare AdaCache with other training-free DiT acceleration methods (e.g. \\(\\Delta\\)-DiT [15], T-GATE [113], PAB [118]) on multiple video baselines (e.g. Open-Sora [120] \\(480p\\) - 2s at 30-steps, Open-Sora-Plan [45] \\(512\\times 512\\) - 2.7s at 150-steps, Latte [62] \\(512\\times 512\\) - 2s at 50-steps). We measure the generation quality with VBenchmark [38], PSNR, LPIPS and SSIM, while reporting complexity with FLOPs, latency and speedup (measured on a single A100 GPU). AdaCache-fast consistently shows the best speedups at a comparable or slightly-lower generation quality. AdaCache-slow gives absolute-best quality while still being faster than prior methods. Our motion-regularization significantly improves the generation quality consistently, with a minimal added-latency." + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.465, + 0.271, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.465, + 0.444, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.664 + ], + "angle": 0, + "content": "Figure 4. User study: We collect human preferences, comparing AdaCache with PAB [118] (left) and evaluating our motion regularization (right). AdaCache shows a significantly higher preference-rate over PAB at a comparable latency. Our motion-regularized variant is better-preferred, yet often tied with AdaCache in terms of perceived quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": "the baseline, and two efficient generations in a randomized order (comparing either AdaCache vs. PAB [118] at a similar speedup, or AdaCache vs. AdaCache w/ MoReg). We ask the users which efficient variant shows a better quality, and whether it is aligned with (i.e., indistinguishable from) the baseline. We collect a total of 1800 responses from 36 different users, and the results of the study are given in Fig. 4. Between AdaCache and PAB, we see a clear win for our method \\((70\\%)\\) while being extremely-similar to the baseline more than half the time \\((41\\%)\\). Among AdaCache variants, users find these to be often tied \\((60\\%)\\) in-terms of perceived quality, yet still showing a better preference for motion-regularized variant \\((25\\%\\) vs. \\(14\\%)\\). This study validates the effectiveness of Adaptive Caching." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.467, + 0.905, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Figure 5. Quality-Latency trade-off: We show quality vs. latency curves for different configurations of AdaCache and PAB [118], with Open-Sora [120] 720p - 2s generations. AdaCache outperforms PAB consistently, showing a more-stable performance while reducing latency. This stability is more-prominent in reference-free metric VBenchmark [38] compared to reference-based metrics, validating that AdaCache generations are aligned with human preference even at its fastest speeds, despite not being exactly-aligned with the reference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.665, + 0.755 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Quality-Latency trade-off: In Fig. 5, we compare the quality-latency trade-off of AdaCache with PAB [118]. First, we note that AdaCache enables significantly higher reduction rates (i.e., much smaller absolute latency) compared to PAB. Moreover, across this whole range of latency configurations, AdaCache gives a more-stable performance over PAB, on all quality metrics. Such behavior is especially evident in reference-free metric VBench [38], that aligns better with human preference. Even if we see a drop" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15245" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.104, + 0.533, + 0.197 + ], + "angle": 0, + "content": "
MethodVBatch sizeLatencySpeedup
Open-Sora [120]84.16419.601.0×
+ AdaCache83.4089.534.7×
+ AdaCache + MoReg83.5093.504.5×
+ AdaCache + MoReg (w/o grad)83.3689.014.7×
+ AdaCache + MoReg (multi-step)83.4295.654.4×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.104, + 0.199, + 0.533, + 0.236 + ], + "angle": 0, + "content": "(a) AdaCache with Motion Regularization: We show different variants of AdaCache. All versions achieve significant speedups compared to the baseline. AdaCache + MoReg shows a better quality with a slightly-lower speedup." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.089, + 0.895, + 0.198 + ], + "angle": 0, + "content": "
ResolutionAdaCacheVBenchmarkLatencySpeedup
480p - 2s✘83.68173.841.0×
✓83.1838.524.5×
480p - 4s✘82.77349.901.0×
✓82.1680.164.4×
720p - 2s✘84.16419.601.0×
✓83.4089.534.7×
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.537, + 0.199, + 0.892, + 0.236 + ], + "angle": 0, + "content": "(b) Speedups at different resolutions: We compare AdaCache with baselines at different resolutions. It generalizes across resolutions, consistently providing a stable acceleration." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.248, + 0.271, + 0.307 + ], + "angle": 0, + "content": "
DistanceVBenchLatency
L183.4089.53
L283.5092.70
Cosine83.1986.74
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.092, + 0.309, + 0.272, + 0.357 + ], + "angle": 0, + "content": "(c) Cache metric: Among distance metrics, L1/L2 have similar (and better) performance in contrast to cosine distance." + }, + { + "type": "table", + "bbox": [ + 0.28, + 0.24, + 0.453, + 0.308 + ], + "angle": 0, + "content": "
LocationVBenchLatency
Start83.3087.55
Mid83.4089.53
End83.4391.20
Multiple83.4190.27
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.278, + 0.31, + 0.452, + 0.358 + ], + "angle": 0, + "content": "(d) Cache location: We compute the cache metric at midDiT, resulting in the best quality-latency trade-off." + }, + { + "type": "table", + "bbox": [ + 0.464, + 0.24, + 0.628, + 0.308 + ], + "angle": 0, + "content": "
ResidualVBenchLatency
pt(TA)83.4089.53
pt(SA)83.1989.06
qt(CA)83.2590.70
rt(MLP)83.6299.72
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.457, + 0.31, + 0.627, + 0.358 + ], + "angle": 0, + "content": "(e) Cache residual: We consider different residual computations to estimate cache metric. Our default is Temp-attn." + }, + { + "type": "table", + "bbox": [ + 0.64, + 0.252, + 0.906, + 0.307 + ], + "angle": 0, + "content": "
AdaCacheBasis-ratesVBranchLatency
Fast{12, 10, 8, 6, 4, 3}83.4089.53
Medium{8, 6, 4, 2, 1}83.94143.87
Slow{2, 1}84.12274.30
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.634, + 0.309, + 0.906, + 0.357 + ], + "angle": 0, + "content": "(f) AdaCache Variants: We achieve a range of speedups (and quality) by controlling the basis cache-rates in AdaCache. Our default setting is AdaCache-fast." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.483, + 0.467 + ], + "angle": 0, + "content": "in reference-based scores (e.g. PSNR, SSIM) at extreme reduction rates, the qualitative results suggest that the generations are still good (see Fig. 1), despite not being aligned exactly with the reference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.483, + 0.685 + ], + "angle": 0, + "content": "AdaCache with Motion Regularization: We compare AdaCache with different versions of motion regularization in Table 2a. Both vanilla and motion-regularized versions provide significant speedups, \\(4.7 \\times\\) and \\(4.5 \\times\\) respectively, at a comparable quality with baseline Open-Sora [120]. Considering motion-gradient as an early-predictor of motion at latter diffusion steps helps (83.50 vs. 83.36 on VBench). We also estimate motion at different frame-rates by considering a varying step-size in frame differences, which seems to increase the latency without necessarily improving quality. Overall, we consider AdaCache (w/ MoReg) as the configuration with the best quality-latency trade-off. This improvement in quality is more-prominent in qualitative examples in supplementary (Fig. A.2) and the user study (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Acceleration in multi-GPU setups: Aligned with prior work based on Dynamic Sequence Parallelism (DSP) [117], we evaluate how AdaCache performs in video generation across multiple GPUs. This evaluation is relevant in the context of efficiency, as DSP incurs additional latency overheads corresponding to the communication between GPUs, and caching mechanisms can avoid such costs by re-using previous computations. We present the results of this study in Fig. 6 and Table 2. Here, we consider Open-Sora [120] (480p - 2s at 30-steps) and Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) as baselines, and compare against prior-art PAB [118] in terms of latency measurements on A100 GPUs. In Fig. 6, we observe that AdaCache consistently outperforms PAB with better speeds across all settings." + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.405, + 0.707, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.406, + 0.886, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.63 + ], + "angle": 0, + "content": "Figure 6. Acceleration in multi-GPU setups: We evaluate the speedups with varying GPU parallelization, as cached-steps can avoid communication overheads among GPUs. Here, we compare AdaCache with PAB [118], on baselines Open-Sora [120] 480p - 2s generations at 30-steps and Open-Sora-Plan [45] \\(512 \\times 512 - 2.7\\) s generations at 150-steps. AdaCache consistently shows better acceleration over PAB in all settings. Latency measured on A100s." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Speedups at different resolutions: In Table 2b, we compare the trade-offs of AdaCache at various resolutions of video generations, namely, \\(480\\mathrm{p} - 2\\mathrm{s}\\), \\(480\\mathrm{p} - 4\\mathrm{s}\\) and \\(720\\mathrm{p} - 2\\mathrm{s}\\), all at 100-steps. AdaCache provides consistent speedups across different resolutions without affecting the quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Cache metric, location and residual: When adaptively deciding the caching schedule, we consider different metrics to compute the rate-of-change between representations, namely, L1/L2 distance or cosine distance. Among these, L1/L2 give an absolute measure which aligns better with the actual change. In contrast, cosine computes a normalized-distance, which is not a good estimate of change (e.g. if the representations differ only by a scale, the distance will be zero, even though we want to have a non-zero value). This observation is verified by the results in Table 2c. Moreover, we consider computing the cache metric at various locations (i.e., layers) in the DiT. Doing so at a single layer (e.g. start," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15246" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.088, + 0.468, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.472, + 0.089, + 0.813, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.907, + 0.479 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison: We show qualitative results on multiple video-DiT baselines including Open-Sora [120] (720p - 2s at 100-steps), Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) and Latte [62] (512×512 - 2s at 50-steps), while comparing against prior training-free inference acceleration method PAB [118]. AdaCache shows a comparable generation quality at much-faster speeds. Best-viewed with zoom-in. Prompts and additional qualitative results (Fig. A.3) are given in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.581 + ], + "angle": 0, + "content": "mid, end) is not significantly different from computing an aggregate over multiple-layers (see Table 2d). By default, we compute the cache metric in the mid-layer as a reasonable choice without extra overheads. We choose temporal-attention as the residual for cache metric computation, since it achieves the best trade-off (see Table 2e)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.721 + ], + "angle": 0, + "content": "AdaCache variants: To achieve a range of speedups (and quality), we consider different basis cache-rates in our AdaCache implementation. For instance, we can have higher-speedup with a slightly-lower quality (e.g. AdaCache-fast), a lower-speedup with a higher-quality (e.g. AdaCacheslow), or balance both (e.g. AdaCache-medium). We can conveniently control this by having corresponding basis cache-rates as shown in Table 2f. By default, we resort to AdaCache-fast which gives the best speedups." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.268, + 0.743 + ], + "angle": 0, + "content": "5.4. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Fig. 7, we present qualitative results on multiple video DiT baselines, including Open-Sora [120], Open-Sora-Plan [45] and Latte [62]. We compare AdaCache against each baseline and prior training-free inference acceleration method for video DiTs, PAB [118]. Here, we consider three different configurations: \\(720\\mathrm{p}\\) - 2s generations at 100-steps for Open-Sora, \\(512\\times 512\\) - 2.7s generations at 150-steps for Open-Sora-Plan, and \\(512\\times 512\\) - 2s generations at 50-steps for Latte, while considering prompts from OpenSora gallery (see supplementary for prompt details). Ada" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Cache shows a comparable generation quality, while having much-faster inference pipelines. In fact, it achieves \\(4.49 \\times\\) (vs. \\(1.26 \\times\\) in PAB), \\(3.53 \\times\\) (vs. \\(1.45 \\times\\) in PAB), \\(2.46 \\times\\) (vs. \\(1.33 \\times\\) in PAB) speedups respectively on the three considered baseline video DiTs. In most cases our generations are aligned well with the baseline in the pixel space. Yet this is not a strict requirement, as the denoising process can deviate considerably from that of the baseline, particularly at high caching rates. Still, AdaCache is faithful to the text prompt and is not affected by significant artifacts. Refer supplementary (Fig. A.3) for more qualitative examples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.634, + 0.693 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced Adaptive Caching (AdaCache), a plug-and-play component that improves the inference speed of video diffusion transformers without needing any re-training. It caches residual computations, while also devising the caching schedule dependent on each video generation. We further proposed a Motion Regularization (MoReg) to utilize video information and allocate computations based on motion content, improving the quality-latency trade-off. We apply our contributions on multiple open-source video DiTs, showing comparable generation quality at a fraction of latency. We believe AdaCache is widely-applicable with minimal effort, helping democratize high-fidelity long-video generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15247" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.206 + ], + "angle": 0, + "content": "The authors would like to thank Ankit Khedia, Arvind Somasundaram, Dustin Johnson, Eugene Vecharynski, Ly Cao, SK Bong, and Yuzi He for the interesting discussions, and Shikun Liu for the support in setting-up the project page. The authors also appreciate the time and effort volunteered by the participants of the user study." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.22, + 0.188, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.246, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[1] Shady Abu-Hussein, Tom Tirer, and Raja Giryes. Adir: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.289, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Albert Pumarola, Ali Thabet, and Edgar Schonfeld. Flexidit: Your diffusion transformer can easily generate high-quality samples with less compute. arXiv preprint arXiv:2502.20126, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.375, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.433, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[4] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.476, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.534, + 0.485, + 0.614 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.618, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[8] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599-4603, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[9] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.246, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.175 + ], + "angle": 0, + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.178, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.249, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.32, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[14] Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.377, + 0.906, + 0.445 + ], + "angle": 0, + "content": "[15] Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. Delta-dit: A training-free acceleration method tailored for diffusion transformers. arXiv preprint arXiv:2406.01125, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.448, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[16] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.519, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[17] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.604, + 0.906, + 0.644 + ], + "angle": 0, + "content": "[18] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint arXiv:1907.06571, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.647, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[19] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[20] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.776, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[21] Juncan Deng, Shuaiying Li, Zeyu Wang, Hong Gu, Kedong Xu, and Kejie Huang. Vq4dit: Efficient post-training vector quantization for diffusion transformers. arXiv preprint arXiv:2408.17131, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15248" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. arXiv preprint arXiv:2403.14611, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[24] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[25] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[27] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[28] Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8352-8361, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[30] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[31] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streaming2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[32] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[34] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.193, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[37] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In International Conference on Machine Learning, pages 13916-13932. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.278, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[39] Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. arXiv preprint arXiv:2401.05735, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.433, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[40] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.49, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[41] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.519, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[42] Inc. Kling AI. Next-generation ai creative studio. URL https://klingai.com/, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.548, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[43] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.619, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[44] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.904, + 0.688 + ], + "angle": 0, + "content": "[45] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 5,6,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.69, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[46] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.762, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[47] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15249" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[49] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Joseph Liu, Joshua Geddes, Ziyu Guo, Haomiao Jiang, and Mahesh Kumar Nandwana. Smoothcache: A universal inference acceleration technique for diffusion transformers. arXiv preprint arXiv:2411.10510, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.221, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[51] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[52] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[53] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[56] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[57] Inc. Luma AI. Dream machine. URL https://lumalabs.ai/dream-machine, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[58] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[59] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[60] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[61] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[63] Yifang Men, Yuan Yao, Miaomiao Cui, and Bo Liefeng. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[64] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.247, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[65] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. arXiv preprint arXiv:2404.12382, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.303, + 0.904, + 0.343 + ], + "angle": 0, + "content": "[66] Inc. OpenAI. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[67] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.401, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[68] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[69] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[70] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[71] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[72] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.735, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[74] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[75] Inc. Runway AI. Gen-3 alpha: A new frontier for video generation. URL https://runwayml.com/research/introducing-gen-3-alpha, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15250" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.15, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[77] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[78] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[79] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[80] Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[81] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[82] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[83] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.507, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[84] Nikita Starodubcev, Dmitry Baranchuk, Artem Fedorov, and Artem Babenko. Your student is better than expected: Adaptive teacher-student collaboration for text-conditional diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275–9285, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[85] Maitreya Suin, Nithin Gopalakrishnan Nair, Chun Pong Lau, Vishal M Patel, and Rama Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6343-6352, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[86] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[87] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[88] Shengkun Tang, Yaqing Wang, Caiwen Ding, Yi Liang, Yao Li, and Dongkuan Xu. Deediff: Dynamic uncertainty-aware early exiting for accelerating diffusion model generation. arXiv preprint arXiv:2309.17074, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[89] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[90] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[91] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[92] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.279, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[93] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.35, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[94] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.407, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[95] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.505, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[96] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. arXiv preprint arXiv:2408.15239, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[97] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.633, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[98] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[99] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[100] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[101] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu," + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15251" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[102] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[103] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[104] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[105] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[106] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[107] Xingyi Yang and Xinchao Wang. Compositional video generation as flow equalization. arXiv preprint arXiv:2407.06182, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[108] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[109] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[110] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[111] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[112] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[113] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747v1, 2024.5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[114] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "sion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[115] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[116] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[117] Xuanlei Zhao, Shenggan Cheng, Zangwei Zheng, Zheming Yang, Ziming Liu, and Yang You. Dsp: Dynamic sequence parallelism for multi-dimensional transformers. arXiv preprint arXiv:2403.10266, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[118] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[119] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[120] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[121] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15252" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_origin.pdf b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6d492c9a52b44612d2aa569f07cf148ca301445f --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/6ff754f4-9314-4139-9910-b0b3eedd5a86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eacb3e5729570068613ea50e156e1b5b978bb63f7eaddb9df88d14373e81a0a +size 23941022 diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/full.md b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31bdb281b4deb841204a485a2daf872bc06fcd69 --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/full.md @@ -0,0 +1,361 @@ +# Adaptive Caching for Faster Video Generation with Diffusion Transformers + +Kumara Kahatapitiya $^{1,2*}$ Haozhe Liu $^{1}$ Sen He $^{1}$ Ding Liu $^{1}$ Menglin Jia $^{1}$ Chenyang Zhang $^{1}$ Michael S. Ryoo $^{2}$ Tian Xie $^{1}$ Meta AI $^{2}$ Stony Brook University adacache-dit.github.io + +# Abstract + +Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs)—despite making significant headway in this context—have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that “not all videos are created equal”: meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to $4.7 \times$ on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines. + +# 1. Introduction + +Diffusion models [33, 82] have become the standard for generative modeling in recent years, arguably surpassing the quality of VAEs [41, 72], GANs [26, 40] and Auto-Regressive models [9, 10]. This observation holds in a wide-range of applications including image [73, 76], video [6, 81], 3D [52, 69], and audio [37, 44] generation, as well as image [4, 32] and video [70, 99] editing. More recent Diffusion Transformers (DiTs) [60, 67] show better promise in terms of scalability and generalization compared to prior UNet-based diffusion models [73], revealing intriguing horizons in GenAI for the years to come. + +Despite the state-of-the-art performance, DiTs can also be computationally expensive both in terms of memory and + +![](images/ad9bd3d2c074a36ac17b15f1d552409a39d370e2b0812379de409bc9fc3b0ab0.jpg) +Figure 1. Effectiveness of Adaptive Caching: We show a qualitative comparison of AdaCache (right) applied on top of Open-Sora [120] (left), a baseline video DiT. Here, we consider generating $720\mathrm{p}$ - 2s video clips, and report VBench [38] quality and average latency (on a single A100 GPU) on the benchmark prompts from Open-Sora [120] gallery. AdaCache generates videos significantly faster (i.e., $4.7\times$ speedup) with a comparable quality. Also, the number of computed steps varies for each video. Best-viewed with zoom-in. Prompts given in supplementary. + +![](images/2b571e2e99ba297a152cc76f1dad1df0fcb0c21f7f340562909bf67101c1dc0c.jpg) + +computational requirements. This becomes especially critical when applied with a large number of input tokens (e.g. high-resolution long video generation). For instance, the reason for models such as Sora [66] not being publicly served is speculated to be the high resource demands and slower inference speeds [54]. To tackle these challenges and reduce the footprint of diffusion models, various research directions have emerged such as latent diffusion [73], step-distillation [79, 111], caching [28, 61, 98], architecture-search [48, 119], token reduction [8, 47] and region-based methods [39, 65]. Fewer techniques transfer readily from UNet-based pipelines to DiTs, whereas others often require novel formulations. Hence, DiT acceleration has been under-explored as of yet. + +Moreover, we note that not all videos are created equal. Some videos contain high-frequency textures and signifi + +cant motion content, whereas others are much simpler (e.g. with homogeneous textures or static regions). Having a diffusion process tailored specifically for each video generation can be beneficial in terms of realizing the best quality-latency trade-off. This idea has been explored to some extent in region-based methods [4, 39, 65], but not sufficiently in the context of video generation. + +Motivated by the above, we introduce Adaptive Caching (AdaCache) for accelerating video diffusion transformers. This approach requires no training and can seamlessly be integrated into a baseline video DiT at inference, as a plug-and-play component. The core idea of our proposal is to cache residual computations within transformer blocks (e.g. attention or MLP outputs) in a certain diffusion step, and reuse them through a number of subsequent steps that is dependent on the generated video. We do this by devising a caching schedule, i.e., deciding when-to-recompute next whenever making a residual computation. This decision is guided by a distance metric that measures the rate-of-change between previously-stored and current representations. If the distance is high we would not cache for an extended period (i.e., #steps), to avoid reusing incompatible representations. We further introduce a Motion Regularization (MoReg) to allocate computations based on the motion content in the video being generated. This is inspired by the observation that high-moving sequences require more diffusion steps to achieve a reasonable quality. Altogether, our pipeline is applied on top of multiple video DiT baselines showing much-faster inference speeds without sacrificing the generation quality (see Fig. 1). Finally, we validate the effectiveness of our contributions and justify our design decisions through ablations and qualitative comparisons. + +# 2. Related Work + +Diffusion-based Video Generation [6, 11, 25, 34, 81] has surpassed the quality and diversity of GAN-based approaches [18, 77, 89, 91, 112], while also being competitive with recent Auto-Regressive models [35, 43, 49, 90, 101, 105]. They have become a standard component in the pipelines for frame interpolation [23, 96], video outpainting [16, 22, 93], image-to-video [6, 27, 103], video-to-video (i.e., video editing or translation) [36, 106, 109], personalization [63, 100], motion customization [104, 116] and compositional generation [51, 107]. The underlying architecture of video diffusion models has evolved from classical UNets [73, 74] with additional spatio-temporal attention layers [7, 17, 25, 29], to fully-fledged transformer-based (i.e., DiT [67]) architectures [24, 56, 62, 114]. In the process, the latency of denoising [55, 82] has also scaled with larger models [24, 68]. This becomes critical especially in applications such as long-video generation [31, 87, 92, 110, 115, 121], while also affecting the growth of commercially-served video models [42, 57, 66, 75]. + +Efficiency of Diffusion models has been actively explored with respect to both training and inference pipelines. Multi-stage training at varying resolutions [12, 13, 24] and high-quality data curation [6, 20, 34, 71] have cut down training costs significantly. In terms of inference acceleration, there exist two main approaches: (1) methods that require re-training such as step-distillation [53, 64, 78, 79], consistency regularization [58, 83], quantization [14, 21, 30, 46, 94], and architecture search/compression [48, 108, 119], or (2) methods that require no re-training such as token reduction [8, 39, 47] and caching [15, 28, 61, 98, 118]. Among these, training-free methods are more-attractive as they can be widely-adopted without any additional costs. This becomes especially relevant for video diffusion models that are both expensive to train and usually very slow at inference. In this paper, we explore a caching-based approach tailored for video DiTs. Different from prior fixed caching schedules in UNet-based [28, 61, 98] and DiT-based [2, 15, 80, 118] pipelines, we introduce a content-dependent (i.e., adaptive) caching scheme to squeeze out the best quality-latency trade-off. + +Content-adaptive Generation may focus on improving consistency [3-5, 19, 95, 102], quality [1, 85], and/or efficiency [39, 65, 84, 88]. Most region-based methods (e.g. image or video editing) rely on a user-provided mask to ensue consistent generations aligned with context information [4, 102]. Some others automatically detect [85] or retrieve [1] useful information to improve generation quality. Among efficiency-oriented approaches, there exist proposals for selectively-processing a subset of latents [39, 65], switching between diffusion models with varying compute budgets [84], or adaptively-controlling the number of denoising steps [50, 59, 88, 98]. Our proposed AdaCache—which also controls #denoising-steps adaptively—provides better flexibility in the sense that (1) it can selectively-cache any layer or even just a specific module within a layer, and (2) it is tailored to each video generation instead of being fixed for a given architecture, enabling a better-adaptive compute allocation. + +# 3. Not All Videos Are Created Equal + +In this section, we motivate the need for a content-dependent denoising process, and show how it can help maximize the quality-latency trade-off. This motivation is based on a couple of interesting observations. + +First, we note that each video is unique. Hence, videos have varying levels of complexity. Here, the complexity of a given video can be expressed by the rate-of-change of information across both space and time. Simpler videos may contain more homogeneous regions and/or static content. In contrast, complex videos have more high-frequency details and/or significant motion. The standard video compres + +![](images/67afe584c0b8e55dbf42b2bb79bbf41ac6e06d640e227151de3d3cd63dafaea6.jpg) +(a) Critical number of denoising-steps varies. + +![](images/8b634b13384c76986ecce34fac45d5db4eebc648712d51d803402812231cab8e.jpg) +(b) Changes in subsequent representations are unique. +Figure 2. Not all videos are created equal: We show frames from $720\mathrm{p}$ - 2s video generations based on Open-Sora [120]. (a) We try to break each generation by reducing the number of diffusion steps. Interestingly, not all videos have the same break point. Some sequences are extremely robust (e.g. first column), while others break easily. (b) When we plot the difference between computed representations in subsequent diffusion steps, we see unique variations (Feature distance vs. #steps). If we are to reuse similar representations across denoising steps, it needs to be tailored to each video. (c) There exist different configurations for generating videos at an approximately-fixed latency (e.g. having an arbitrary #denoising-steps, yet only computing a fixed #representations and reusing otherwise). We observe a significant variance in quality in such videos. All these observations suggest the need for a content-dependent denoising process, which is the founding motivation of Adaptive Caching. Best-viewed with zoom-in. Prompts given in supplementary. + +![](images/2196f2d689d3b0e3f285f6eac49cfe58f1f8429425cf130762ffa53ad16c6ad1.jpg) +(c) Different levels of quality at a given latency budget. + +sion techniques exploit such information to achieve the best possible compression ratios without sacrificing the quality [86, 97]. Motivated by the same, we explore how the compute-cost affects the quality of video generations based on DiTs. We measure this w.r.t. the number of denoising steps, and the observations are shown in Fig. 2 (a). Some video sequences are very robust, and achieve a reasonable quality even at fewer denoising steps. Others break easily when we keep reducing the #steps, but the break-point varies. This suggests that the minimal #steps (or, computations) required to generate a video with a reasonable quality varies, and having a content-dependent denoising schedule can exploit this for speeding-up generation. + +Next, we observe how the computed representations (i.e., residual connections in attention or MLP blocks within DiT) change during the denoising process, across different video generations. This may reveal the level of compute redundancy in each video generation, enabling us to reuse representations and improve efficiency. More specifically, we visualize the feature differences between subsequent diffusion steps as histograms given in Fig. 2 (b). Here, we report Feature distance (e.g. L1) vs. #steps. We observe that each histogram is unique. Despite having higher changes in early/latter steps and smaller changes in the middle, the overall distribution and the absolute values vary considerably. A smaller change corresponds to higher redundancy across subsequent computations, and an opportunity for reusing. This motivates the need for a non-uniform compute-schedule not only within the diffusion process of a given video, but also across different videos. + +Finally, we evaluate the video generation quality at a capped-budget (i.e., fixed computations or latency). We can have multiple generation configurations at an + +approximately-fixed latency, by computing a constant number of representations. For instance, we can cache and reuse representations more-frequently in a setup with more denoising steps, still having the same latency of a process with fewer steps. The observations of a study with either 30 or 100 base denoising steps is shown in Fig. 2 (c). We see that the generation quality varies significantly despite spending a similar cost and having the same underlying pretrained DiT. This motivates us to think about how best to allocate resources at inference, tailored for each video generation. + +# 4. Adaptive Caching for Faster Video DiTs + +# 4.1. Preliminaries: Video Diffusion Transformers + +Video Diffusion Transformers are extended from Latent Diffusion Transformers (DiTs) [67] introduced for image generation. DiTs provide a much-more streamlined, scalable architecture compared to prior UNet-based diffusion models [73], by only having transformer blocks with a homogeneous token resolution (instead of convolutional blocks with up/downsampling). A simplified transformer block (i.e., w/o normalizing or timestep conditioning layers) in a video DiT is shown in Fig. 3 (right)—gray block. It consists of spatial-temporal attention (STA), cross-attention (CA) and linear (MLP) layers. Depending on the implementation, STA may be a single joint spatio-temporal attention layer, or separate spatial and temporal attention layers repeated within alternating blocks. Without loss of generality, let us denote a latent feature at the input/output of such block by $f_{t}^{l}$ and $f_{t}^{l + 1}$ , respectively. Here, $l$ represents the layer index, and $t$ , the diffusion timestep. A simplified flow of computations within each block can be represented as, + +$$ +p _ {t} ^ {l} = \operatorname {S T A} \left(f _ {t} ^ {l}\right); \quad \tilde {f} _ {t} ^ {l} = f _ {t} ^ {l} + p _ {t} ^ {l}, \tag {1} +$$ + +$$ +q _ {t} ^ {l} = \mathrm {C A} \left(\tilde {f} _ {t} ^ {l}\right) \quad ; \quad \bar {f} _ {t} ^ {l} = \tilde {f} _ {t} ^ {l} + q _ {t} ^ {l}, \tag {2} +$$ + +$$ +r _ {t} ^ {l} = \mathrm {M L P} \left(\bar {f} _ {t} ^ {l}\right); f _ {t} ^ {l + 1} = \bar {f} _ {t} ^ {l} + r _ {t} ^ {l}. \tag {3} +$$ + +Here $p_t^l$ , $q_t^l$ and $r_t^l$ are residual connections corresponding to each compute-element. Such computations repeat through $L$ layers, generating the noise prediction of each step $t$ , and across a total of $T$ denoising steps. In the current streamlined video DiT architectures with homogeneous token resolutions, each layer or denoising step costs the same. + +# 4.2. Adaptive Caching + +In this subsection, we introduce Adaptive Caching (AdaCache), a training-free mechanism for content-dependent compute allocation in video DiTs. The overview of Adaptive Caching is shown in Fig. 3. Compared to a standard DiT that computes representations for all layers across all diffusion steps, in AdaCache, we decide which layers or steps to compute, adaptively (i.e., dependent on each video being generated). This decision is based on the rate-of-change in the residual connections (e.g. $p_t^l$ , $q_t^l$ or $r_t^l$ ) across diffusion steps, which amount to all significant computations within the DiT. Without loss of generality, let us assume that the residuals in block $l$ in current and immediately-prior diffusion steps $t$ and $t + k$ are already computed. Here, step $t + k$ is identified as 'immediately-prior' to step $t$ since any residuals between these two steps are assumed to be not computed (i.e., cached residuals are reused). We make a decision on the next computation step based on the distance metric $(c_t^l)$ given by, + +$$ +c _ {t} ^ {l} = \operatorname {d i s t} \left(p _ {t + k} ^ {l}, p _ {t} ^ {l}\right) = \left\| p _ {t} ^ {l} - p _ {t + k} ^ {l} \right\| / k. \tag {4} +$$ + +Here, we use L1 distance by default, but other distance metrics can also be applied (e.g. L2, cosine). Once we have the distance metric, we select the next caching rate $(\tau_t^l)$ based on a pre-defined codebook of basis cache-rates. Here, a 'cache-rate' is defined as the number subsequent steps during which, a previously-computed representation is re-used (i.e., a higher cache-rate gives more compute savings). The codebook is basically a collection of cache-rates defined based on the original denoising schedule (i.e., #steps), coupled with corresponding metric thresholds to select them. Simply put, a higher distance metric will sample a lower cache-rate from the codebook, resulting in more-frequent re-computations. More details about the codebook is available in the supplementary (Sec. A.2). + +$$ +\tau_ {t} ^ {l} = \operatorname {c o d e b o o k} \left(c _ {t} ^ {l}\right). \tag {5} +$$ + +For all denoising steps within $t$ and $t - \tau$ , we reuse previously-cached representations and only recompute after the current caching schedule (while also estimating the metric, again). + +![](images/8fd638c32ce17918dd5cf18a4b185c1ddc6821f457a1c5282bd974e29706e49e.jpg) +Figure 3. Overview of Adaptive Caching: (Left) During the diffusion process, we choose to cache residual computations within selected DiT blocks. The caching schedule is content-dependent, as we decide when to compute the next representation based on a distance metric $(c_{t})$ . This metric measures the rate-of-change from previously-computed (and, stored) representation to the current one, and can be evaluated per-layer or the DiT as-a-whole. Each computed residual can be cached and reused across multiple steps. (Right) We only cache the residuals (i.e., skip-connections) which amount to the actual computations (e.g. spatial-temporal/cross attention, MLP). The iteratively denoised representation (i.e., $f_{t+k}$ , $f_{t}$ ) always gets updated either with computed or cached residuals. + +$$ +p _ {t - k} ^ {l} = \left\{ \begin{array}{l l} p _ {t} ^ {l} & \text {i f} k < \tau_ {t} ^ {l}; \\ p _ {t - k} ^ {l} = \operatorname {S T A} (f _ {t - k} ^ {l}) & \text {i f} k = \tau_ {t} ^ {l}. \end{array} \right. \tag {6} +$$ + +The same applies to other residual computations (e.g. $q_{t - k}^{l}, r_{t - k}^{l}$ ) as well. By design, we can have unique caching schedules for each layer (and, each residual computation). However, we observe that it will make the generations unstable. Therefore, we decide to have a common metric (i.e., $c_t^l = c_t$ ) and hence, a common caching rate (i.e., $\tau_t^l = \tau_t$ ) across all DiT layers. For instance, we can consider an averaged metric across all layers, or a metric computed at a certain layer to decide the caching schedule. Meaning, when we recompute residuals in a certain step, we do so for the whole DiT rather than selectively for each layer. + +Overall, this setup allows us to adaptively-control the compute spent on each video generation, based on framewise information (i.e., no temporal information used as of yet). If the rate-of-change between residuals is high, we will have a smaller caching rate, and otherwise, we have a higher rate. The choice of a lightweight distance metric (e.g. L1) helps us avoid any additional latency overheads. + +# 4.3. Motion Regularization + +To further improve Adaptive Caching by making use of video-specific (i.e., temporal) information, we introduce a Motion Regularization (MoReg). This is motivated by the observation that the optimal number of denoising steps varies based on the motion content of each generated video. The core idea here, is to cache less (i.e., recompute more) if a generated video has a high motion content. Simply put, we plan to regularize our caching schedule based on motion. However the problem is that, we need to estimate mo + +tion while the video is still being generated. Therefore, we can not rely on motion estimation algorithms in the pixel space, nor any compute-heavy ones as our focus is on efficiency. As a result, we estimate a noisy latent motion-score $(m_t^l)$ based on residual frame differences. Without loss of generality, let us denote residual latent frames of $p_t^l$ as $\{p_{t,n}^l\mid n = 0,\dots ,N - 1\}$ where $N$ is the #frames in latent space (generated by the VAE encoder). We estimate the motion-score as, + +$$ +m _ {t} ^ {l} = \left\| p _ {t, i: N} ^ {l} - p _ {t, 0: N - i} ^ {l} \right\|. \tag {7} +$$ + +Here, $i$ denotes the frame step-size (or, frame-rate), $\| \cdot \|$ , the distance metric (e.g. L1), and $i:j$ , the slice of all frames within the corresponding range. However, since we operate on noisy-latents, we observe that our motion estimate, particularly in early diffusion steps is not reliable. Meaning, it does not provide a reasonable regularization in early steps (i.e., the change in caching schedule does not correlate well with the observed motion in pixel space). To alleviate this, we also compute a motion-gradient $(mg_t^l)$ across diffusion steps, which can act as a reasonable early-predictor of motion that we may observe in latter diffusion steps (that also correlates with the motion in pixel space). + +$$ +m g _ {t} ^ {l} = \left(m _ {t} ^ {l} - m _ {t + k} ^ {l}\right) / k. \tag {8} +$$ + +Despite the motion-score being noisy, the motion-gradient acts as a better-estimate of trend, as the representations are getting denoised and converging to a noise-free distribution. Finally, we use both motion-score and motion-gradient as a scaling-factor of the distance metric $(c_t^l)$ to regularize our caching schedule. + +$$ +\hat {c} _ {t} ^ {l} = c _ {t} ^ {l} \cdot \left(m _ {t} ^ {l} + m g _ {t} ^ {l}\right). \tag {9} +$$ + +This means, when we have a higher estimated motion, the distance metric will be increased and a smaller basis cache-rate will be selected from the codebook. As previously discussed, we also enforce a common motion-regularization in all DiT layers by computing a common motion score (i.e., $m_t^l = m_t$ , $mg_t^l = mg_t$ ), ensuring the stability of the denoising process. We can also choose to compute motion at different frame-rates, which we ablate in our experiments. Refer to the supplementary for concrete examples of motion-score and motion-gradient (Fig. A.1). + +# 5. Experiments + +# 5.1. Implementation details + +We select multiple prominent open-source video DiTs as backbone video generation pipelines in our experiments, namely, Open-Sora-v1.2 [120] for Text+Image-to-Video (TI2V) generation, and Open-Sora-Plan-v1.1 [45] and Latte + +[62] for Text-to-Video (T2V) generation. Since we focus on inference-based latency optimizations (i.e., without any re-training), we compare AdaCache against similar methods such as $\Delta$ -DiT [15], T-GATE [113] and PAB [118]. In our main experiments, we generate $900+$ videos based on standard VBenchmark [38] benchmark prompts, at the corresponding baseline generation settings suggested in [118] for a fair comparison (e.g. $480p$ - 2s with 30-steps in Open-Sora, $512\times 512$ - 2.7s with 150-steps in Open-Sora-Plan and $512\times 512$ - 2s with 50-steps in Latte). We report VBenchmark average and reference-based PSNR, SSIM and LPIPS as quality metrics, and also report FLOPs, Latency (s) and Speedup as complexity metrics. Here, Latency is measured on a single A100 GPU. Unless otherwise stated, in our ablations and qualitative results, we experiment on the prompts from Open-Sora benchmark gallery, generating $720p$ - 2s videos with 100-steps, following the recommendation of model authors—which gives better quality video generations (and, higher speedups) compared to the benchmark setting discussed above (i.e., $480p$ - 2s with 30-steps). + +# 5.2. Main results + +In Table 1, we present a quantitative evaluation of quality and latency on VBench [38] benchmark. We consider three variants of AdaCache: a slow variant, a fast variant with more speedup and the same with motion regularization. We compare with other training-free acceleration methods, showing consistently better speedups with a comparable generation quality. With Open-Sora [120] baseline, AdaCache-slow outperforms others on all quality metrics, while giving a $1.46 \times$ speedup compared to PAB [118] with $1.20 \times$ speedup. AdaCache-fast gives the highest acceleration of $2.24 \times$ with a slight drop in quality. AdaCache-fast (w/ MoReg) shows a clear improvement in quality compared to AdaCache-fast, validating the effectiveness of our regularization and giving a comparable speedup of $2.10 \times$ . All AdaCache variants outperform even the baseline (w/o any acceleration) on VBench average quality, which aligns better with human preference compared to other reference-based metrics. Similar observations hold with the other baselines as well. With Open-Sora-Plan [45], AdaCache shows the best speedup of $3.70 \times$ compared to the previous-best $1.45 \times$ of PAB, and the best quality with a $2.20 \times$ speedup. With Latte [62], we gain the best speedup of $2.74 \times$ compared to prior $1.33 \times$ , and the best overall quality with a $1.59 \times$ speedup. + +User study: Quantitative metrics on video generation quality can sometimes fall-short in aligning with the perceived visual quality. To better understand the human preference on AdaCache and its comparisons, we conduct a user study in the form of randomized A/B preference tests. Here, we create a questionnaire with 50 multiple-choice questions, each consisting of 3 variants of a single video sequence: + +
MethodVBatch (%) ↑PSNR ↑LPIPS ↓SSIM ↑FLOPs (T)Latency (s)Speedup
Open-Sora [120]79.22---3230.2454.021.00×
+ Δ-DiT [15]78.2111.910.56920.48113166.47--
+ T-GATE [113]77.6115.500.34950.67602818.4049.111.10×
+ PAB-fast [118]76.9523.580.17430.82202558.2540.231.34×
+ PAB-slow [118]78.5127.040.09250.88472657.7044.931.20×
+ AdaCache-fast79.3924.920.09810.83751331.9724.162.24×
+ AdaCache-fast (w/ MoReg)79.4825.780.08670.85301383.6625.712.10×
+ AdaCache-slow79.6629.970.04560.90852195.5037.011.46×
Open-Sora-Plan [45]80.39---12032.40129.671.00×
+ Δ-DiT [15]77.5513.850.53880.373612027.72--
+ T-GATE [113]80.1518.320.30660.621910663.32113.751.14×
+ PAB-fast [118]71.8115.470.54990.47178551.2689.561.45×
+ PAB-slow [118]80.3018.800.30590.65509276.5798.501.32×
+ AdaCache-fast75.8313.530.54650.43093283.6035.043.70×
+ AdaCache-fast (w/ MoReg)79.3017.690.37450.61473473.6836.773.53×
+ AdaCache-slow80.5022.980.17370.79104983.3058.882.20×
Latte [62]77.40---3439.4732.451.00×
+ Δ-DiT [15]52.008.650.85130.10783437.33--
+ T-GATE [113]75.4219.550.26120.69273059.0229.231.11×
+ PAB-fast [118]73.1317.160.39030.64212576.7724.331.33×
+ PAB-slow [118]76.3219.710.26990.70142767.2226.201.24×
+ AdaCache-fast76.2617.700.35220.66591010.3311.852.74×
+ AdaCache-fast (w/ MoReg)76.4718.160.32220.68321187.3113.202.46×
+ AdaCache-slow77.0722.780.17370.80302023.6520.351.59×
+ +Table 1. Quantitative evaluation of quality and latency: Here, we compare AdaCache with other training-free DiT acceleration methods (e.g. $\Delta$ -DiT [15], T-GATE [113], PAB [118]) on multiple video baselines (e.g. Open-Sora [120] $480p$ - 2s at 30-steps, Open-Sora-Plan [45] $512\times 512$ - 2.7s at 150-steps, Latte [62] $512\times 512$ - 2s at 50-steps). We measure the generation quality with VBenchmark [38], PSNR, LPIPS and SSIM, while reporting complexity with FLOPs, latency and speedup (measured on a single A100 GPU). AdaCache-fast consistently shows the best speedups at a comparable or slightly-lower generation quality. AdaCache-slow gives absolute-best quality while still being faster than prior methods. Our motion-regularization significantly improves the generation quality consistently, with a minimal added-latency. + +![](images/f16ae822e7e91a2d4eca199c514c5bd2d747adfb53b8cfafda3cfcc7235516f5.jpg) +Figure 4. User study: We collect human preferences, comparing AdaCache with PAB [118] (left) and evaluating our motion regularization (right). AdaCache shows a significantly higher preference-rate over PAB at a comparable latency. Our motion-regularized variant is better-preferred, yet often tied with AdaCache in terms of perceived quality. + +![](images/72c24cca44aec9be03284401bd7b301478f68ef9564e4745e4418e5a750295d0.jpg) + +the baseline, and two efficient generations in a randomized order (comparing either AdaCache vs. PAB [118] at a similar speedup, or AdaCache vs. AdaCache w/ MoReg). We ask the users which efficient variant shows a better quality, and whether it is aligned with (i.e., indistinguishable from) the baseline. We collect a total of 1800 responses from 36 different users, and the results of the study are given in Fig. 4. Between AdaCache and PAB, we see a clear win for our method $(70\%)$ while being extremely-similar to the baseline more than half the time $(41\%)$ . Among AdaCache variants, users find these to be often tied $(60\%)$ in-terms of perceived quality, yet still showing a better preference for motion-regularized variant $(25\%$ vs. $14\%)$ . This study validates the effectiveness of Adaptive Caching. + +![](images/a7fcaceef2b10348d240fa7dc97d067eaf3a9ec174539fa1fe61af919391d341.jpg) +Figure 5. Quality-Latency trade-off: We show quality vs. latency curves for different configurations of AdaCache and PAB [118], with Open-Sora [120] 720p - 2s generations. AdaCache outperforms PAB consistently, showing a more-stable performance while reducing latency. This stability is more-prominent in reference-free metric VBenchmark [38] compared to reference-based metrics, validating that AdaCache generations are aligned with human preference even at its fastest speeds, despite not being exactly-aligned with the reference. + +# 5.3. Ablation study + +Quality-Latency trade-off: In Fig. 5, we compare the quality-latency trade-off of AdaCache with PAB [118]. First, we note that AdaCache enables significantly higher reduction rates (i.e., much smaller absolute latency) compared to PAB. Moreover, across this whole range of latency configurations, AdaCache gives a more-stable performance over PAB, on all quality metrics. Such behavior is especially evident in reference-free metric VBench [38], that aligns better with human preference. Even if we see a drop + +
MethodVBatch sizeLatencySpeedup
Open-Sora [120]84.16419.601.0×
+ AdaCache83.4089.534.7×
+ AdaCache + MoReg83.5093.504.5×
+ AdaCache + MoReg (w/o grad)83.3689.014.7×
+ AdaCache + MoReg (multi-step)83.4295.654.4×
+ +(a) AdaCache with Motion Regularization: We show different variants of AdaCache. All versions achieve significant speedups compared to the baseline. AdaCache + MoReg shows a better quality with a slightly-lower speedup. + +
ResolutionAdaCacheVBenchmarkLatencySpeedup
480p - 2s✘83.68173.841.0×
✓83.1838.524.5×
480p - 4s✘82.77349.901.0×
✓82.1680.164.4×
720p - 2s✘84.16419.601.0×
✓83.4089.534.7×
+ +(b) Speedups at different resolutions: We compare AdaCache with baselines at different resolutions. It generalizes across resolutions, consistently providing a stable acceleration. + +
DistanceVBenchLatency
L183.4089.53
L283.5092.70
Cosine83.1986.74
+ +(c) Cache metric: Among distance metrics, L1/L2 have similar (and better) performance in contrast to cosine distance. + +
LocationVBenchLatency
Start83.3087.55
Mid83.4089.53
End83.4391.20
Multiple83.4190.27
+ +(d) Cache location: We compute the cache metric at midDiT, resulting in the best quality-latency trade-off. + +
ResidualVBenchLatency
pt(TA)83.4089.53
pt(SA)83.1989.06
qt(CA)83.2590.70
rt(MLP)83.6299.72
+ +(e) Cache residual: We consider different residual computations to estimate cache metric. Our default is Temp-attn. + +
AdaCacheBasis-ratesVBranchLatency
Fast{12, 10, 8, 6, 4, 3}83.4089.53
Medium{8, 6, 4, 2, 1}83.94143.87
Slow{2, 1}84.12274.30
+ +(f) AdaCache Variants: We achieve a range of speedups (and quality) by controlling the basis cache-rates in AdaCache. Our default setting is AdaCache-fast. + +in reference-based scores (e.g. PSNR, SSIM) at extreme reduction rates, the qualitative results suggest that the generations are still good (see Fig. 1), despite not being aligned exactly with the reference. + +AdaCache with Motion Regularization: We compare AdaCache with different versions of motion regularization in Table 2a. Both vanilla and motion-regularized versions provide significant speedups, $4.7 \times$ and $4.5 \times$ respectively, at a comparable quality with baseline Open-Sora [120]. Considering motion-gradient as an early-predictor of motion at latter diffusion steps helps (83.50 vs. 83.36 on VBench). We also estimate motion at different frame-rates by considering a varying step-size in frame differences, which seems to increase the latency without necessarily improving quality. Overall, we consider AdaCache (w/ MoReg) as the configuration with the best quality-latency trade-off. This improvement in quality is more-prominent in qualitative examples in supplementary (Fig. A.2) and the user study (Fig. 4). + +Acceleration in multi-GPU setups: Aligned with prior work based on Dynamic Sequence Parallelism (DSP) [117], we evaluate how AdaCache performs in video generation across multiple GPUs. This evaluation is relevant in the context of efficiency, as DSP incurs additional latency overheads corresponding to the communication between GPUs, and caching mechanisms can avoid such costs by re-using previous computations. We present the results of this study in Fig. 6 and Table 2. Here, we consider Open-Sora [120] (480p - 2s at 30-steps) and Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) as baselines, and compare against prior-art PAB [118] in terms of latency measurements on A100 GPUs. In Fig. 6, we observe that AdaCache consistently outperforms PAB with better speeds across all settings. + +![](images/97e97b168c1ccc16708bd416f339ce78d11df4cc358a28bc7e7ebc82c88e502c.jpg) +Figure 6. Acceleration in multi-GPU setups: We evaluate the speedups with varying GPU parallelization, as cached-steps can avoid communication overheads among GPUs. Here, we compare AdaCache with PAB [118], on baselines Open-Sora [120] 480p - 2s generations at 30-steps and Open-Sora-Plan [45] $512 \times 512 - 2.7$ s generations at 150-steps. AdaCache consistently shows better acceleration over PAB in all settings. Latency measured on A100s. + +![](images/8c9e73ca666ddaa2a3a50d011b634ac4bd56d67716e923dfae106b65cf5d94d5.jpg) + +Speedups at different resolutions: In Table 2b, we compare the trade-offs of AdaCache at various resolutions of video generations, namely, $480\mathrm{p} - 2\mathrm{s}$ , $480\mathrm{p} - 4\mathrm{s}$ and $720\mathrm{p} - 2\mathrm{s}$ , all at 100-steps. AdaCache provides consistent speedups across different resolutions without affecting the quality. + +Cache metric, location and residual: When adaptively deciding the caching schedule, we consider different metrics to compute the rate-of-change between representations, namely, L1/L2 distance or cosine distance. Among these, L1/L2 give an absolute measure which aligns better with the actual change. In contrast, cosine computes a normalized-distance, which is not a good estimate of change (e.g. if the representations differ only by a scale, the distance will be zero, even though we want to have a non-zero value). This observation is verified by the results in Table 2c. Moreover, we consider computing the cache metric at various locations (i.e., layers) in the DiT. Doing so at a single layer (e.g. start, + +![](images/e27b05116571f57b6efd690ef14d7ad54262733746df3acb6b3f30030a325a1e.jpg) +Figure 7. Qualitative comparison: We show qualitative results on multiple video-DiT baselines including Open-Sora [120] (720p - 2s at 100-steps), Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) and Latte [62] (512×512 - 2s at 50-steps), while comparing against prior training-free inference acceleration method PAB [118]. AdaCache shows a comparable generation quality at much-faster speeds. Best-viewed with zoom-in. Prompts and additional qualitative results (Fig. A.3) are given in supplementary. + +![](images/83e6469b73c482a305bea3622859285adee5ba791e51a376c4f4b23989aeb597.jpg) + +mid, end) is not significantly different from computing an aggregate over multiple-layers (see Table 2d). By default, we compute the cache metric in the mid-layer as a reasonable choice without extra overheads. We choose temporal-attention as the residual for cache metric computation, since it achieves the best trade-off (see Table 2e). + +AdaCache variants: To achieve a range of speedups (and quality), we consider different basis cache-rates in our AdaCache implementation. For instance, we can have higher-speedup with a slightly-lower quality (e.g. AdaCache-fast), a lower-speedup with a higher-quality (e.g. AdaCacheslow), or balance both (e.g. AdaCache-medium). We can conveniently control this by having corresponding basis cache-rates as shown in Table 2f. By default, we resort to AdaCache-fast which gives the best speedups. + +# 5.4. Qualitative results + +In Fig. 7, we present qualitative results on multiple video DiT baselines, including Open-Sora [120], Open-Sora-Plan [45] and Latte [62]. We compare AdaCache against each baseline and prior training-free inference acceleration method for video DiTs, PAB [118]. Here, we consider three different configurations: $720\mathrm{p}$ - 2s generations at 100-steps for Open-Sora, $512\times 512$ - 2.7s generations at 150-steps for Open-Sora-Plan, and $512\times 512$ - 2s generations at 50-steps for Latte, while considering prompts from OpenSora gallery (see supplementary for prompt details). Ada + +Cache shows a comparable generation quality, while having much-faster inference pipelines. In fact, it achieves $4.49 \times$ (vs. $1.26 \times$ in PAB), $3.53 \times$ (vs. $1.45 \times$ in PAB), $2.46 \times$ (vs. $1.33 \times$ in PAB) speedups respectively on the three considered baseline video DiTs. In most cases our generations are aligned well with the baseline in the pixel space. Yet this is not a strict requirement, as the denoising process can deviate considerably from that of the baseline, particularly at high caching rates. Still, AdaCache is faithful to the text prompt and is not affected by significant artifacts. Refer supplementary (Fig. A.3) for more qualitative examples. + +# 6. Conclusion + +In this paper, we introduced Adaptive Caching (AdaCache), a plug-and-play component that improves the inference speed of video diffusion transformers without needing any re-training. It caches residual computations, while also devising the caching schedule dependent on each video generation. We further proposed a Motion Regularization (MoReg) to utilize video information and allocate computations based on motion content, improving the quality-latency trade-off. We apply our contributions on multiple open-source video DiTs, showing comparable generation quality at a fraction of latency. We believe AdaCache is widely-applicable with minimal effort, helping democratize high-fidelity long-video generation. + +# Acknowledgements + +The authors would like to thank Ankit Khedia, Arvind Somasundaram, Dustin Johnson, Eugene Vecharynski, Ly Cao, SK Bong, and Yuzi He for the interesting discussions, and Shikun Liu for the support in setting-up the project page. The authors also appreciate the time and effort volunteered by the participants of the user study. + +# References + +[1] Shady Abu-Hussein, Tom Tirer, and Raja Giryes. Adir: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221, 2022. 2 +[2] Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Albert Pumarola, Ali Thabet, and Edgar Schonfeld. Flexidit: Your diffusion transformer can easily generate high-quality samples with less compute. arXiv preprint arXiv:2502.20126, 2025. 2 +[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2 +[4] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 2 +[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2 +[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2 +[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2 +[8] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599-4603, 2023. 1, 2 +[9] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1 +[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1 + +[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2 +[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2 +[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2 +[14] Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. 2 +[15] Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. Delta-dit: A training-free acceleration method tailored for diffusion transformers. arXiv preprint arXiv:2406.01125, 2024. 2, 5, 6 +[16] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024. 2 +[17] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 2 +[18] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint arXiv:1907.06571, 2019. 2 +[19] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2 +[20] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2 +[21] Juncan Deng, Shuaiying Li, Zeyu Wang, Hong Gu, Kedong Xu, and Kejie Huang. Vq4dit: Efficient post-training vector quantization for diffusion transformers. arXiv preprint arXiv:2408.17131, 2024. 2 +[22] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 2 + +[23] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. arXiv preprint arXiv:2403.14611, 2024. 2 +[24] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 2 +[25] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2 +[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1 +[27] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 2 +[28] Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8352-8361, 2024. 1, 2 +[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2 +[30] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[31] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streaming2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 2 +[32] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1 +[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[34] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for + +text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2 +[37] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In International Conference on Machine Learning, pages 13916-13932. PMLR, 2023. 1 +[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 5, 6, 7 +[39] Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. arXiv preprint arXiv:2401.05735, 2024. 1, 2 +[40] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1 +[41] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1 +[42] Inc. Kling AI. Next-generation ai creative studio. URL https://klingai.com/, 2024. 2 +[43] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2 +[44] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 1 +[45] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 5,6,7,8 +[46] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 2 +[47] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 1, 2 +[48] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 + +[49] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 2 +[50] Joseph Liu, Joshua Geddes, Ziyu Guo, Haomiao Jiang, and Mahesh Kumar Nandwana. Smoothcache: A universal inference acceleration technique for diffusion transformers. arXiv preprint arXiv:2411.10510, 2024. 2 +[51] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2 +[52] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 1 +[53] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2 +[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 1 +[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2 +[56] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311, 2023. 2 +[57] Inc. Luma AI. Dream machine. URL https://lumalabs.ai/dream-machine, 2024.2 +[58] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2 +[59] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 2 +[60] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1 +[61] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 1, 2 +[62] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: + +Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 5, 6, 8 +[63] Yifang Men, Yuan Yao, Miaomiao Cui, and Bo Liefeng. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 2 +[64] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2 +[65] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. arXiv preprint arXiv:2404.12382, 2024. 1, 2 +[66] Inc. OpenAI. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2 +[67] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 2, 3 +[68] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[69] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1 +[70] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023. 1 +[71] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2 +[72] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 1 +[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3 +[74] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2 +[75] Inc. Runway AI. Gen-3 alpha: A new frontier for video generation. URL https://runwayml.com/research/introducing-gen-3-alpha, 2024.2 +[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, + +Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[77] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2 +[78] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2 +[79] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2 +[80] Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425, 2024. 2 +[81] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 2 +[82] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2 +[83] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2 +[84] Nikita Starodubcev, Dmitry Baranchuk, Artem Fedorov, and Artem Babenko. Your student is better than expected: Adaptive teacher-student collaboration for text-conditional diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275–9285, 2024. 2 +[85] Maitreya Suin, Nithin Gopalakrishnan Nair, Chun Pong Lau, Vishal M Patel, and Rama Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6343-6352, 2024. 2 +[86] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3 +[87] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 2 +[88] Shengkun Tang, Yaqing Wang, Caiwen Ding, Yi Liang, Yao Li, and Dongkuan Xu. Deediff: Dynamic uncertainty-aware early exiting for accelerating diffusion model generation. arXiv preprint arXiv:2309.17074, 2023. 2 +[89] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2 + +[90] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2 +[91] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2 +[92] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 2 +[93] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 2 +[94] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 2 +[95] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2 +[96] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. arXiv preprint arXiv:2408.15239, 2024. 2 +[97] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 3 +[98] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 1, 2 +[99] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1 +[100] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 2 +[101] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, + +Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2 +[102] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2 +[103] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2 +[104] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2 +[105] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 +[106] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2 +[107] Xingyi Yang and Xinchao Wang. Compositional video generation as flow equalization. arXiv preprint arXiv:2407.06182, 2024. 2 +[108] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 2 +[109] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 2 +[110] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 2 +[111] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 1 +[112] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2 +[113] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747v1, 2024.5, 6 +[114] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffu + +sion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2 +[115] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2 +[116] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2 +[117] Xuanlei Zhao, Shenggan Cheng, Zangwei Zheng, Zheming Yang, Ziming Liu, and Yang You. Dsp: Dynamic sequence parallelism for multi-dimensional transformers. arXiv preprint arXiv:2403.10266, 2024. 7 +[118] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 2, 5, 6, 7, 8 +[119] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 1, 2 +[120] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3, 5, 6, 7, 8 +[121] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 2 \ No newline at end of file diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/images.zip b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2a0c3d9cf83a0ade3eab74bc123d693077148211 --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83624e0b2c56371acf3a4f9a86f0a3f124872ab4065be7f51d522ac20ad4cb50 +size 742755 diff --git a/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/layout.json b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb93f268ff7245ca87e98301b3f334c864ed2cb2 --- /dev/null +++ b/2025/Adaptive Caching for Faster Video Generation with Diffusion Transformers/layout.json @@ -0,0 +1,10688 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 100, + 539, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 100, + 539, + 118 + ], + "spans": [ + { + "bbox": [ + 71, + 100, + 539, + 118 + ], + "type": "text", + "content": "Adaptive Caching for Faster Video Generation with Diffusion Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "spans": [ + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": "Kumara Kahatapitiya" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Haozhe Liu" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Sen He" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Ding Liu" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Menglin Jia" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Chenyang Zhang" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Michael S. Ryoo" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Tian Xie" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": " Meta AI " + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 130, + 481, + 194 + ], + "type": "text", + "content": "Stony Brook University adacache-dit.github.io" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 501 + ], + "type": "text", + "content": "Generating temporally-consistent high-fidelity videos can be computationally expensive, especially over longer temporal spans. More-recent Diffusion Transformers (DiTs)—despite making significant headway in this context—have only heightened such challenges as they rely on larger models and heavier attention mechanisms, resulting in slower inference speeds. In this paper, we introduce a training-free method to accelerate video DiTs, termed Adaptive Caching (AdaCache), which is motivated by the fact that “not all videos are created equal”: meaning, some videos require fewer denoising steps to attain a reasonable quality than others. Building on this, we not only cache computations through the diffusion process, but also devise a caching schedule tailored to each video generation, maximizing the quality-latency trade-off. We further introduce a Motion Regularization (MoReg) scheme to utilize video information within AdaCache, essentially controlling the compute allocation based on motion content. Altogether, our plug-and-play contributions grant significant inference speedups (e.g. up to " + }, + { + "bbox": [ + 55, + 238, + 297, + 501 + ], + "type": "inline_equation", + "content": "4.7 \\times" + }, + { + "bbox": [ + 55, + 238, + 297, + 501 + ], + "type": "text", + "content": " on Open-Sora 720p - 2s video generation) without sacrificing the generation quality, across multiple video DiT baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 670 + ], + "type": "text", + "content": "Diffusion models [33, 82] have become the standard for generative modeling in recent years, arguably surpassing the quality of VAEs [41, 72], GANs [26, 40] and Auto-Regressive models [9, 10]. This observation holds in a wide-range of applications including image [73, 76], video [6, 81], 3D [52, 69], and audio [37, 44] generation, as well as image [4, 32] and video [70, 99] editing. More recent Diffusion Transformers (DiTs) [60, 67] show better promise in terms of scalability and generalization compared to prior UNet-based diffusion models [73], revealing intriguing horizons in GenAI for the years to come." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 695 + ], + "type": "text", + "content": "Despite the state-of-the-art performance, DiTs can also be computationally expensive both in terms of memory and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 212, + 435, + 398 + ], + "blocks": [ + { + "bbox": [ + 317, + 212, + 435, + 398 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 435, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 435, + 398 + ], + "type": "image", + "image_path": "ad9bd3d2c074a36ac17b15f1d552409a39d370e2b0812379de409bc9fc3b0ab0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. Effectiveness of Adaptive Caching: We show a qualitative comparison of AdaCache (right) applied on top of Open-Sora [120] (left), a baseline video DiT. Here, we consider generating " + }, + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "type": "inline_equation", + "content": "720\\mathrm{p}" + }, + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "type": "text", + "content": " - 2s video clips, and report VBench [38] quality and average latency (on a single A100 GPU) on the benchmark prompts from Open-Sora [120] gallery. AdaCache generates videos significantly faster (i.e., " + }, + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "type": "inline_equation", + "content": "4.7\\times" + }, + { + "bbox": [ + 313, + 403, + 555, + 502 + ], + "type": "text", + "content": " speedup) with a comparable quality. Also, the number of computed steps varies for each video. Best-viewed with zoom-in. Prompts given in supplementary." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 211, + 554, + 398 + ], + "blocks": [ + { + "bbox": [ + 437, + 211, + 554, + 398 + ], + "lines": [ + { + "bbox": [ + 437, + 211, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 437, + 211, + 554, + 398 + ], + "type": "image", + "image_path": "2b571e2e99ba297a152cc76f1dad1df0fcb0c21f7f340562909bf67101c1dc0c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 518, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 686 + ], + "type": "text", + "content": "computational requirements. This becomes especially critical when applied with a large number of input tokens (e.g. high-resolution long video generation). For instance, the reason for models such as Sora [66] not being publicly served is speculated to be the high resource demands and slower inference speeds [54]. To tackle these challenges and reduce the footprint of diffusion models, various research directions have emerged such as latent diffusion [73], step-distillation [79, 111], caching [28, 61, 98], architecture-search [48, 119], token reduction [8, 47] and region-based methods [39, 65]. Fewer techniques transfer readily from UNet-based pipelines to DiTs, whereas others often require novel formulations. Hence, DiT acceleration has been under-explored as of yet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Moreover, we note that not all videos are created equal. Some videos contain high-frequency textures and signifi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Work done at Meta" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15240" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "cant motion content, whereas others are much simpler (e.g. with homogeneous textures or static regions). Having a diffusion process tailored specifically for each video generation can be beneficial in terms of realizing the best quality-latency trade-off. This idea has been explored to some extent in region-based methods [4, 39, 65], but not sufficiently in the context of video generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 298, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 298, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 298, + 456 + ], + "type": "text", + "content": "Motivated by the above, we introduce Adaptive Caching (AdaCache) for accelerating video diffusion transformers. This approach requires no training and can seamlessly be integrated into a baseline video DiT at inference, as a plug-and-play component. The core idea of our proposal is to cache residual computations within transformer blocks (e.g. attention or MLP outputs) in a certain diffusion step, and reuse them through a number of subsequent steps that is dependent on the generated video. We do this by devising a caching schedule, i.e., deciding when-to-recompute next whenever making a residual computation. This decision is guided by a distance metric that measures the rate-of-change between previously-stored and current representations. If the distance is high we would not cache for an extended period (i.e., #steps), to avoid reusing incompatible representations. We further introduce a Motion Regularization (MoReg) to allocate computations based on the motion content in the video being generated. This is inspired by the observation that high-moving sequences require more diffusion steps to achieve a reasonable quality. Altogether, our pipeline is applied on top of multiple video DiT baselines showing much-faster inference speeds without sacrificing the generation quality (see Fig. 1). Finally, we validate the effectiveness of our contributions and justify our design decisions through ablations and qualitative comparisons." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 142, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Diffusion-based Video Generation [6, 11, 25, 34, 81] has surpassed the quality and diversity of GAN-based approaches [18, 77, 89, 91, 112], while also being competitive with recent Auto-Regressive models [35, 43, 49, 90, 101, 105]. They have become a standard component in the pipelines for frame interpolation [23, 96], video outpainting [16, 22, 93], image-to-video [6, 27, 103], video-to-video (i.e., video editing or translation) [36, 106, 109], personalization [63, 100], motion customization [104, 116] and compositional generation [51, 107]. The underlying architecture of video diffusion models has evolved from classical UNets [73, 74] with additional spatio-temporal attention layers [7, 17, 25, 29], to fully-fledged transformer-based (i.e., DiT [67]) architectures [24, 56, 62, 114]. In the process, the latency of denoising [55, 82] has also scaled with larger models [24, 68]. This becomes critical especially in applications such as long-video generation [31, 87, 92, 110, 115, 121], while also affecting the growth of commercially-served video models [42, 57, 66, 75]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 324 + ], + "type": "text", + "content": "Efficiency of Diffusion models has been actively explored with respect to both training and inference pipelines. Multi-stage training at varying resolutions [12, 13, 24] and high-quality data curation [6, 20, 34, 71] have cut down training costs significantly. In terms of inference acceleration, there exist two main approaches: (1) methods that require re-training such as step-distillation [53, 64, 78, 79], consistency regularization [58, 83], quantization [14, 21, 30, 46, 94], and architecture search/compression [48, 108, 119], or (2) methods that require no re-training such as token reduction [8, 39, 47] and caching [15, 28, 61, 98, 118]. Among these, training-free methods are more-attractive as they can be widely-adopted without any additional costs. This becomes especially relevant for video diffusion models that are both expensive to train and usually very slow at inference. In this paper, we explore a caching-based approach tailored for video DiTs. Different from prior fixed caching schedules in UNet-based [28, 61, 98] and DiT-based [2, 15, 80, 118] pipelines, we introduce a content-dependent (i.e., adaptive) caching scheme to squeeze out the best quality-latency trade-off." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 327, + 556, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 543 + ], + "type": "text", + "content": "Content-adaptive Generation may focus on improving consistency [3-5, 19, 95, 102], quality [1, 85], and/or efficiency [39, 65, 84, 88]. Most region-based methods (e.g. image or video editing) rely on a user-provided mask to ensue consistent generations aligned with context information [4, 102]. Some others automatically detect [85] or retrieve [1] useful information to improve generation quality. Among efficiency-oriented approaches, there exist proposals for selectively-processing a subset of latents [39, 65], switching between diffusion models with varying compute budgets [84], or adaptively-controlling the number of denoising steps [50, 59, 88, 98]. Our proposed AdaCache—which also controls #denoising-steps adaptively—provides better flexibility in the sense that (1) it can selectively-cache any layer or even just a specific module within a layer, and (2) it is tailored to each video generation instead of being fixed for a given architecture, enabling a better-adaptive compute allocation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 559, + 497, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 497, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 497, + 572 + ], + "type": "text", + "content": "3. Not All Videos Are Created Equal" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": "In this section, we motivate the need for a content-dependent denoising process, and show how it can help maximize the quality-latency trade-off. This motivation is based on a couple of interesting observations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "First, we note that each video is unique. Hence, videos have varying levels of complexity. Here, the complexity of a given video can be expressed by the rate-of-change of information across both space and time. Simpler videos may contain more homogeneous regions and/or static content. In contrast, complex videos have more high-frequency details and/or significant motion. The standard video compres" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15241" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 87, + 216, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 87, + 216, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 87, + 216, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 87, + 216, + 199 + ], + "type": "image", + "image_path": "67afe584c0b8e55dbf42b2bb79bbf41ac6e06d640e227151de3d3cd63dafaea6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 201, + 196, + 208 + ], + "lines": [ + { + "bbox": [ + 83, + 201, + 196, + 208 + ], + "spans": [ + { + "bbox": [ + 83, + 201, + 196, + 208 + ], + "type": "text", + "content": "(a) Critical number of denoising-steps varies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 71, + 386, + 200 + ], + "blocks": [ + { + "bbox": [ + 222, + 71, + 386, + 200 + ], + "lines": [ + { + "bbox": [ + 222, + 71, + 386, + 200 + ], + "spans": [ + { + "bbox": [ + 222, + 71, + 386, + 200 + ], + "type": "image", + "image_path": "8b634b13384c76986ecce34fac45d5db4eebc648712d51d803402812231cab8e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 242, + 201, + 379, + 208 + ], + "lines": [ + { + "bbox": [ + 242, + 201, + 379, + 208 + ], + "spans": [ + { + "bbox": [ + 242, + 201, + 379, + 208 + ], + "type": "text", + "content": "(b) Changes in subsequent representations are unique." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 212, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. Not all videos are created equal: We show frames from " + }, + { + "bbox": [ + 54, + 212, + 555, + 301 + ], + "type": "inline_equation", + "content": "720\\mathrm{p}" + }, + { + "bbox": [ + 54, + 212, + 555, + 301 + ], + "type": "text", + "content": " - 2s video generations based on Open-Sora [120]. (a) We try to break each generation by reducing the number of diffusion steps. Interestingly, not all videos have the same break point. Some sequences are extremely robust (e.g. first column), while others break easily. (b) When we plot the difference between computed representations in subsequent diffusion steps, we see unique variations (Feature distance vs. #steps). If we are to reuse similar representations across denoising steps, it needs to be tailored to each video. (c) There exist different configurations for generating videos at an approximately-fixed latency (e.g. having an arbitrary #denoising-steps, yet only computing a fixed #representations and reusing otherwise). We observe a significant variance in quality in such videos. All these observations suggest the need for a content-dependent denoising process, which is the founding motivation of Adaptive Caching. Best-viewed with zoom-in. Prompts given in supplementary." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 393, + 86, + 554, + 200 + ], + "blocks": [ + { + "bbox": [ + 393, + 86, + 554, + 200 + ], + "lines": [ + { + "bbox": [ + 393, + 86, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 393, + 86, + 554, + 200 + ], + "type": "image", + "image_path": "2196f2d689d3b0e3f285f6eac49cfe58f1f8429425cf130762ffa53ad16c6ad1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 201, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 410, + 201, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 410, + 201, + 547, + 208 + ], + "type": "text", + "content": "(c) Different levels of quality at a given latency budget." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 311, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 296, + 466 + ], + "type": "text", + "content": "sion techniques exploit such information to achieve the best possible compression ratios without sacrificing the quality [86, 97]. Motivated by the same, we explore how the compute-cost affects the quality of video generations based on DiTs. We measure this w.r.t. the number of denoising steps, and the observations are shown in Fig. 2 (a). Some video sequences are very robust, and achieve a reasonable quality even at fewer denoising steps. Others break easily when we keep reducing the #steps, but the break-point varies. This suggests that the minimal #steps (or, computations) required to generate a video with a reasonable quality varies, and having a content-dependent denoising schedule can exploit this for speeding-up generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 470, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 673 + ], + "type": "text", + "content": "Next, we observe how the computed representations (i.e., residual connections in attention or MLP blocks within DiT) change during the denoising process, across different video generations. This may reveal the level of compute redundancy in each video generation, enabling us to reuse representations and improve efficiency. More specifically, we visualize the feature differences between subsequent diffusion steps as histograms given in Fig. 2 (b). Here, we report Feature distance (e.g. L1) vs. #steps. We observe that each histogram is unique. Despite having higher changes in early/latter steps and smaller changes in the middle, the overall distribution and the absolute values vary considerably. A smaller change corresponds to higher redundancy across subsequent computations, and an opportunity for reusing. This motivates the need for a non-uniform compute-schedule not only within the diffusion process of a given video, but also across different videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Finally, we evaluate the video generation quality at a capped-budget (i.e., fixed computations or latency). We can have multiple generation configurations at an" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 311, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 430 + ], + "type": "text", + "content": "approximately-fixed latency, by computing a constant number of representations. For instance, we can cache and reuse representations more-frequently in a setup with more denoising steps, still having the same latency of a process with fewer steps. The observations of a study with either 30 or 100 base denoising steps is shown in Fig. 2 (c). We see that the generation quality varies significantly despite spending a similar cost and having the same underlying pretrained DiT. This motivates us to think about how best to allocate resources at inference, tailored for each video generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 453, + 534, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 534, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 534, + 467 + ], + "type": "text", + "content": "4. Adaptive Caching for Faster Video DiTs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 476, + 548, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 548, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 548, + 489 + ], + "type": "text", + "content": "4.1. Preliminaries: Video Diffusion Transformers" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Video Diffusion Transformers are extended from Latent Diffusion Transformers (DiTs) [67] introduced for image generation. DiTs provide a much-more streamlined, scalable architecture compared to prior UNet-based diffusion models [73], by only having transformer blocks with a homogeneous token resolution (instead of convolutional blocks with up/downsampling). A simplified transformer block (i.e., w/o normalizing or timestep conditioning layers) in a video DiT is shown in Fig. 3 (right)—gray block. It consists of spatial-temporal attention (STA), cross-attention (CA) and linear (MLP) layers. Depending on the implementation, STA may be a single joint spatio-temporal attention layer, or separate spatial and temporal attention layers repeated within alternating blocks. Without loss of generality, let us denote a latent feature at the input/output of such block by " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{t}^{l}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{t}^{l + 1}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", respectively. Here, " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " represents the layer index, and " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", the diffusion timestep. A simplified flow of computations within each block can be represented as," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15242" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 102, + 87, + 295, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 87, + 295, + 102 + ], + "spans": [ + { + "bbox": [ + 102, + 87, + 295, + 102 + ], + "type": "interline_equation", + "content": "p _ {t} ^ {l} = \\operatorname {S T A} \\left(f _ {t} ^ {l}\\right); \\quad \\tilde {f} _ {t} ^ {l} = f _ {t} ^ {l} + p _ {t} ^ {l}, \\tag {1}", + "image_path": "97c8f3f24827c29454d1c996b7b331c777d6976b5b5900797dacdac8edbb8de4.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 105, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 105, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 104, + 105, + 295, + 119 + ], + "type": "interline_equation", + "content": "q _ {t} ^ {l} = \\mathrm {C A} \\left(\\tilde {f} _ {t} ^ {l}\\right) \\quad ; \\quad \\bar {f} _ {t} ^ {l} = \\tilde {f} _ {t} ^ {l} + q _ {t} ^ {l}, \\tag {2}", + "image_path": "684f6ff14f15a005f159cbf753de99cabb27299c536d0b74d48d23bf3670b3f2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 121, + 295, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 121, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 104, + 121, + 295, + 137 + ], + "type": "interline_equation", + "content": "r _ {t} ^ {l} = \\mathrm {M L P} \\left(\\bar {f} _ {t} ^ {l}\\right); f _ {t} ^ {l + 1} = \\bar {f} _ {t} ^ {l} + r _ {t} ^ {l}. \\tag {3}", + "image_path": "02e1ee3b8d8ca867a0a2a5c76a5a9e697848d69bd0feed0448fd7bb3faaee60c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "p_t^l" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "q_t^l" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "r_t^l" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": " are residual connections corresponding to each compute-element. Such computations repeat through " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": " layers, generating the noise prediction of each step " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": ", and across a total of " + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 141, + 296, + 214 + ], + "type": "text", + "content": " denoising steps. In the current streamlined video DiT architectures with homogeneous token resolutions, each layer or denoising step costs the same." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 219, + 163, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 163, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 163, + 232 + ], + "type": "text", + "content": "4.2. Adaptive Caching" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": "In this subsection, we introduce Adaptive Caching (AdaCache), a training-free mechanism for content-dependent compute allocation in video DiTs. The overview of Adaptive Caching is shown in Fig. 3. Compared to a standard DiT that computes representations for all layers across all diffusion steps, in AdaCache, we decide which layers or steps to compute, adaptively (i.e., dependent on each video being generated). This decision is based on the rate-of-change in the residual connections (e.g. " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "p_t^l" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "q_t^l" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "r_t^l" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": ") across diffusion steps, which amount to all significant computations within the DiT. Without loss of generality, let us assume that the residuals in block " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " in current and immediately-prior diffusion steps " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "t + k" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " are already computed. Here, step " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "t + k" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " is identified as 'immediately-prior' to step " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " since any residuals between these two steps are assumed to be not computed (i.e., cached residuals are reused). We make a decision on the next computation step based on the distance metric " + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "inline_equation", + "content": "(c_t^l)" + }, + { + "bbox": [ + 55, + 236, + 296, + 452 + ], + "type": "text", + "content": " given by," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 456, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 456, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 87, + 456, + 295, + 471 + ], + "type": "interline_equation", + "content": "c _ {t} ^ {l} = \\operatorname {d i s t} \\left(p _ {t + k} ^ {l}, p _ {t} ^ {l}\\right) = \\left\\| p _ {t} ^ {l} - p _ {t + k} ^ {l} \\right\\| / k. \\tag {4}", + "image_path": "162a42b187f1e47f97ead03bbba995af4416cd8094c700be53418db270b3c837.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "content": "Here, we use L1 distance by default, but other distance metrics can also be applied (e.g. L2, cosine). Once we have the distance metric, we select the next caching rate " + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "inline_equation", + "content": "(\\tau_t^l)" + }, + { + "bbox": [ + 55, + 475, + 296, + 643 + ], + "type": "text", + "content": " based on a pre-defined codebook of basis cache-rates. Here, a 'cache-rate' is defined as the number subsequent steps during which, a previously-computed representation is re-used (i.e., a higher cache-rate gives more compute savings). The codebook is basically a collection of cache-rates defined based on the original denoising schedule (i.e., #steps), coupled with corresponding metric thresholds to select them. Simply put, a higher distance metric will sample a lower cache-rate from the codebook, resulting in more-frequent re-computations. More details about the codebook is available in the supplementary (Sec. A.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 647, + 295, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 647, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 128, + 647, + 295, + 662 + ], + "type": "interline_equation", + "content": "\\tau_ {t} ^ {l} = \\operatorname {c o d e b o o k} \\left(c _ {t} ^ {l}\\right). \\tag {5}", + "image_path": "f7f95b2697956a54d907c8a5c6cc9767d2e5480355041d10026a7273010f2f6d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "For all denoising steps within " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "t - \\tau" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", we reuse previously-cached representations and only recompute after the current caching schedule (while also estimating the metric, again)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 555, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 167 + ], + "type": "image", + "image_path": "8fd638c32ce17918dd5cf18a4b185c1ddc6821f457a1c5282bd974e29706e49e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": "Figure 3. Overview of Adaptive Caching: (Left) During the diffusion process, we choose to cache residual computations within selected DiT blocks. The caching schedule is content-dependent, as we decide when to compute the next representation based on a distance metric " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "(c_{t})" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ". This metric measures the rate-of-change from previously-computed (and, stored) representation to the current one, and can be evaluated per-layer or the DiT as-a-whole. Each computed residual can be cached and reused across multiple steps. (Right) We only cache the residuals (i.e., skip-connections) which amount to the actual computations (e.g. spatial-temporal/cross attention, MLP). The iteratively denoised representation (i.e., " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{t+k}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 313, + 176, + 555, + 308 + ], + "type": "text", + "content": ") always gets updated either with computed or cached residuals." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 341, + 328, + 555, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 328, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 341, + 328, + 555, + 360 + ], + "type": "interline_equation", + "content": "p _ {t - k} ^ {l} = \\left\\{ \\begin{array}{l l} p _ {t} ^ {l} & \\text {i f} k < \\tau_ {t} ^ {l}; \\\\ p _ {t - k} ^ {l} = \\operatorname {S T A} (f _ {t - k} ^ {l}) & \\text {i f} k = \\tau_ {t} ^ {l}. \\end{array} \\right. \\tag {6}", + "image_path": "827b17cfd9706201709e4a6b4c80bd060e2317b9f12ae88e24b3c69e1243b61f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "content": "The same applies to other residual computations (e.g. " + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "inline_equation", + "content": "q_{t - k}^{l}, r_{t - k}^{l}" + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "content": ") as well. By design, we can have unique caching schedules for each layer (and, each residual computation). However, we observe that it will make the generations unstable. Therefore, we decide to have a common metric (i.e., " + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "inline_equation", + "content": "c_t^l = c_t" + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "content": ") and hence, a common caching rate (i.e., " + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\tau_t^l = \\tau_t" + }, + { + "bbox": [ + 313, + 366, + 555, + 498 + ], + "type": "text", + "content": ") across all DiT layers. For instance, we can consider an averaged metric across all layers, or a metric computed at a certain layer to decide the caching schedule. Meaning, when we recompute residuals in a certain step, we do so for the whole DiT rather than selectively for each layer." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 582 + ], + "type": "text", + "content": "Overall, this setup allows us to adaptively-control the compute spent on each video generation, based on framewise information (i.e., no temporal information used as of yet). If the rate-of-change between residuals is high, we will have a smaller caching rate, and otherwise, we have a higher rate. The choice of a lightweight distance metric (e.g. L1) helps us avoid any additional latency overheads." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "type": "text", + "content": "4.3. Motion Regularization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "To further improve Adaptive Caching by making use of video-specific (i.e., temporal) information, we introduce a Motion Regularization (MoReg). This is motivated by the observation that the optimal number of denoising steps varies based on the motion content of each generated video. The core idea here, is to cache less (i.e., recompute more) if a generated video has a high motion content. Simply put, we plan to regularize our caching schedule based on motion. However the problem is that, we need to estimate mo" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15243" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "tion while the video is still being generated. Therefore, we can not rely on motion estimation algorithms in the pixel space, nor any compute-heavy ones as our focus is on efficiency. As a result, we estimate a noisy latent motion-score " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "(m_t^l)" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " based on residual frame differences. Without loss of generality, let us denote residual latent frames of " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "p_t^l" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "\\{p_{t,n}^l\\mid n = 0,\\dots ,N - 1\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": " is the #frames in latent space (generated by the VAE encoder). We estimate the motion-score as," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 186, + 295, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 186, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 295, + 201 + ], + "type": "interline_equation", + "content": "m _ {t} ^ {l} = \\left\\| p _ {t, i: N} ^ {l} - p _ {t, 0: N - i} ^ {l} \\right\\|. \\tag {7}", + "image_path": "2c0239a48b09dffb7e4f85a517f654c27e91383f91ddc792a516015f1302334a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": " denotes the frame step-size (or, frame-rate), " + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|" + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": ", the distance metric (e.g. L1), and " + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "inline_equation", + "content": "i:j" + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": ", the slice of all frames within the corresponding range. However, since we operate on noisy-latents, we observe that our motion estimate, particularly in early diffusion steps is not reliable. Meaning, it does not provide a reasonable regularization in early steps (i.e., the change in caching schedule does not correlate well with the observed motion in pixel space). To alleviate this, we also compute a motion-gradient " + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "inline_equation", + "content": "(mg_t^l)" + }, + { + "bbox": [ + 55, + 210, + 296, + 354 + ], + "type": "text", + "content": " across diffusion steps, which can act as a reasonable early-predictor of motion that we may observe in latter diffusion steps (that also correlates with the motion in pixel space)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 365, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 365, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 119, + 365, + 295, + 380 + ], + "type": "interline_equation", + "content": "m g _ {t} ^ {l} = \\left(m _ {t} ^ {l} - m _ {t + k} ^ {l}\\right) / k. \\tag {8}", + "image_path": "2a74350d039465c4698d9640873b45c3cf9e5f410e025f871c7d2facc55f6c87.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": "Despite the motion-score being noisy, the motion-gradient acts as a better-estimate of trend, as the representations are getting denoised and converging to a noise-free distribution. Finally, we use both motion-score and motion-gradient as a scaling-factor of the distance metric " + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "inline_equation", + "content": "(c_t^l)" + }, + { + "bbox": [ + 55, + 390, + 296, + 462 + ], + "type": "text", + "content": " to regularize our caching schedule." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 473, + 295, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 473, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 127, + 473, + 295, + 488 + ], + "type": "interline_equation", + "content": "\\hat {c} _ {t} ^ {l} = c _ {t} ^ {l} \\cdot \\left(m _ {t} ^ {l} + m g _ {t} ^ {l}\\right). \\tag {9}", + "image_path": "e0e7373c6ac95887a4eb131273d9522b4c4ba86838c703a0a077399fd4a1fa7b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": "This means, when we have a higher estimated motion, the distance metric will be increased and a smaller basis cache-rate will be selected from the codebook. As previously discussed, we also enforce a common motion-regularization in all DiT layers by computing a common motion score (i.e., " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "m_t^l = m_t" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "inline_equation", + "content": "mg_t^l = mg_t" + }, + { + "bbox": [ + 55, + 498, + 296, + 617 + ], + "type": "text", + "content": "), ensuring the stability of the denoising process. We can also choose to compute motion at different frame-rates, which we ablate in our experiments. Refer to the supplementary for concrete examples of motion-score and motion-gradient (Fig. A.1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 137, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 186, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 186, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 186, + 661 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We select multiple prominent open-source video DiTs as backbone video generation pipelines in our experiments, namely, Open-Sora-v1.2 [120] for Text+Image-to-Video (TI2V) generation, and Open-Sora-Plan-v1.1 [45] and Latte" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "[62] for Text-to-Video (T2V) generation. Since we focus on inference-based latency optimizations (i.e., without any re-training), we compare AdaCache against similar methods such as " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "-DiT [15], T-GATE [113] and PAB [118]. In our main experiments, we generate " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " videos based on standard VBenchmark [38] benchmark prompts, at the corresponding baseline generation settings suggested in [118] for a fair comparison (e.g. " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "480p" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " - 2s with 30-steps in Open-Sora, " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " - 2.7s with 150-steps in Open-Sora-Plan and " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " - 2s with 50-steps in Latte). We report VBenchmark average and reference-based PSNR, SSIM and LPIPS as quality metrics, and also report FLOPs, Latency (s) and Speedup as complexity metrics. Here, Latency is measured on a single A100 GPU. Unless otherwise stated, in our ablations and qualitative results, we experiment on the prompts from Open-Sora benchmark gallery, generating " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "720p" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " - 2s videos with 100-steps, following the recommendation of model authors—which gives better quality video generations (and, higher speedups) compared to the benchmark setting discussed above (i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "480p" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " - 2s with 30-steps)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 321, + 395, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 321, + 395, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 321, + 395, + 332 + ], + "type": "text", + "content": "5.2. Main results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": "In Table 1, we present a quantitative evaluation of quality and latency on VBench [38] benchmark. We consider three variants of AdaCache: a slow variant, a fast variant with more speedup and the same with motion regularization. We compare with other training-free acceleration methods, showing consistently better speedups with a comparable generation quality. With Open-Sora [120] baseline, AdaCache-slow outperforms others on all quality metrics, while giving a " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "1.46 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " speedup compared to PAB [118] with " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "1.20 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " speedup. AdaCache-fast gives the highest acceleration of " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "2.24 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " with a slight drop in quality. AdaCache-fast (w/ MoReg) shows a clear improvement in quality compared to AdaCache-fast, validating the effectiveness of our regularization and giving a comparable speedup of " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "2.10 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": ". All AdaCache variants outperform even the baseline (w/o any acceleration) on VBench average quality, which aligns better with human preference compared to other reference-based metrics. Similar observations hold with the other baselines as well. With Open-Sora-Plan [45], AdaCache shows the best speedup of " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "3.70 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " compared to the previous-best " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "1.45 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " of PAB, and the best quality with a " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "2.20 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " speedup. With Latte [62], we gain the best speedup of " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "2.74 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " compared to prior " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "1.33 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": ", and the best overall quality with a " + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "inline_equation", + "content": "1.59 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 626 + ], + "type": "text", + "content": " speedup." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "User study: Quantitative metrics on video generation quality can sometimes fall-short in aligning with the perceived visual quality. To better understand the human preference on AdaCache and its comparisons, we conduct a user study in the form of randomized A/B preference tests. Here, we create a questionnaire with 50 multiple-choice questions, each consisting of 3 variants of a single video sequence:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15244" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 113, + 70, + 496, + 293 + ], + "blocks": [ + { + "bbox": [ + 113, + 70, + 496, + 293 + ], + "lines": [ + { + "bbox": [ + 113, + 70, + 496, + 293 + ], + "spans": [ + { + "bbox": [ + 113, + 70, + 496, + 293 + ], + "type": "table", + "html": "
MethodVBatch (%) ↑PSNR ↑LPIPS ↓SSIM ↑FLOPs (T)Latency (s)Speedup
Open-Sora [120]79.22---3230.2454.021.00×
+ Δ-DiT [15]78.2111.910.56920.48113166.47--
+ T-GATE [113]77.6115.500.34950.67602818.4049.111.10×
+ PAB-fast [118]76.9523.580.17430.82202558.2540.231.34×
+ PAB-slow [118]78.5127.040.09250.88472657.7044.931.20×
+ AdaCache-fast79.3924.920.09810.83751331.9724.162.24×
+ AdaCache-fast (w/ MoReg)79.4825.780.08670.85301383.6625.712.10×
+ AdaCache-slow79.6629.970.04560.90852195.5037.011.46×
Open-Sora-Plan [45]80.39---12032.40129.671.00×
+ Δ-DiT [15]77.5513.850.53880.373612027.72--
+ T-GATE [113]80.1518.320.30660.621910663.32113.751.14×
+ PAB-fast [118]71.8115.470.54990.47178551.2689.561.45×
+ PAB-slow [118]80.3018.800.30590.65509276.5798.501.32×
+ AdaCache-fast75.8313.530.54650.43093283.6035.043.70×
+ AdaCache-fast (w/ MoReg)79.3017.690.37450.61473473.6836.773.53×
+ AdaCache-slow80.5022.980.17370.79104983.3058.882.20×
Latte [62]77.40---3439.4732.451.00×
+ Δ-DiT [15]52.008.650.85130.10783437.33--
+ T-GATE [113]75.4219.550.26120.69273059.0229.231.11×
+ PAB-fast [118]73.1317.160.39030.64212576.7724.331.33×
+ PAB-slow [118]76.3219.710.26990.70142767.2226.201.24×
+ AdaCache-fast76.2617.700.35220.66591010.3311.852.74×
+ AdaCache-fast (w/ MoReg)76.4718.160.32220.68321187.3113.202.46×
+ AdaCache-slow77.0722.780.17370.80302023.6520.351.59×
", + "image_path": "c53e9bc2e656c7878443cb4673756141a03924a7f2c13bb9e20f96d3e471a578.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of quality and latency: Here, we compare AdaCache with other training-free DiT acceleration methods (e.g. " + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "text", + "content": "-DiT [15], T-GATE [113], PAB [118]) on multiple video baselines (e.g. Open-Sora [120] " + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "inline_equation", + "content": "480p" + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "text", + "content": " - 2s at 30-steps, Open-Sora-Plan [45] " + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "text", + "content": " - 2.7s at 150-steps, Latte [62] " + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 295, + 555, + 361 + ], + "type": "text", + "content": " - 2s at 50-steps). We measure the generation quality with VBenchmark [38], PSNR, LPIPS and SSIM, while reporting complexity with FLOPs, latency and speedup (measured on a single A100 GPU). AdaCache-fast consistently shows the best speedups at a comparable or slightly-lower generation quality. AdaCache-slow gives absolute-best quality while still being faster than prior methods. Our motion-regularization significantly improves the generation quality consistently, with a minimal added-latency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 64, + 368, + 165, + 456 + ], + "blocks": [ + { + "bbox": [ + 64, + 368, + 165, + 456 + ], + "lines": [ + { + "bbox": [ + 64, + 368, + 165, + 456 + ], + "spans": [ + { + "bbox": [ + 64, + 368, + 165, + 456 + ], + "type": "image", + "image_path": "f16ae822e7e91a2d4eca199c514c5bd2d747adfb53b8cfafda3cfcc7235516f5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 459, + 295, + 525 + ], + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 525 + ], + "type": "text", + "content": "Figure 4. User study: We collect human preferences, comparing AdaCache with PAB [118] (left) and evaluating our motion regularization (right). AdaCache shows a significantly higher preference-rate over PAB at a comparable latency. Our motion-regularized variant is better-preferred, yet often tied with AdaCache in terms of perceived quality." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 368, + 271, + 457 + ], + "blocks": [ + { + "bbox": [ + 173, + 368, + 271, + 457 + ], + "lines": [ + { + "bbox": [ + 173, + 368, + 271, + 457 + ], + "spans": [ + { + "bbox": [ + 173, + 368, + 271, + 457 + ], + "type": "image", + "image_path": "72c24cca44aec9be03284401bd7b301478f68ef9564e4745e4418e5a750295d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": "the baseline, and two efficient generations in a randomized order (comparing either AdaCache vs. PAB [118] at a similar speedup, or AdaCache vs. AdaCache w/ MoReg). We ask the users which efficient variant shows a better quality, and whether it is aligned with (i.e., indistinguishable from) the baseline. We collect a total of 1800 responses from 36 different users, and the results of the study are given in Fig. 4. Between AdaCache and PAB, we see a clear win for our method " + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "inline_equation", + "content": "(70\\%)" + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": " while being extremely-similar to the baseline more than half the time " + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "inline_equation", + "content": "(41\\%)" + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": ". Among AdaCache variants, users find these to be often tied " + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "inline_equation", + "content": "(60\\%)" + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": " in-terms of perceived quality, yet still showing a better preference for motion-regularized variant " + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "inline_equation", + "content": "(25\\%" + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "inline_equation", + "content": "14\\%)" + }, + { + "bbox": [ + 55, + 546, + 296, + 714 + ], + "type": "text", + "content": ". This study validates the effectiveness of Adaptive Caching." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 369, + 553, + 464 + ], + "blocks": [ + { + "bbox": [ + 316, + 369, + 553, + 464 + ], + "lines": [ + { + "bbox": [ + 316, + 369, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 553, + 464 + ], + "type": "image", + "image_path": "a7fcaceef2b10348d240fa7dc97d067eaf3a9ec174539fa1fe61af919391d341.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 567 + ], + "type": "text", + "content": "Figure 5. Quality-Latency trade-off: We show quality vs. latency curves for different configurations of AdaCache and PAB [118], with Open-Sora [120] 720p - 2s generations. AdaCache outperforms PAB consistently, showing a more-stable performance while reducing latency. This stability is more-prominent in reference-free metric VBenchmark [38] compared to reference-based metrics, validating that AdaCache generations are aligned with human preference even at its fastest speeds, despite not being exactly-aligned with the reference." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 584, + 406, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 406, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 406, + 597 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Quality-Latency trade-off: In Fig. 5, we compare the quality-latency trade-off of AdaCache with PAB [118]. First, we note that AdaCache enables significantly higher reduction rates (i.e., much smaller absolute latency) compared to PAB. Moreover, across this whole range of latency configurations, AdaCache gives a more-stable performance over PAB, on all quality metrics. Such behavior is especially evident in reference-free metric VBench [38], that aligns better with human preference. Even if we see a drop" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15245" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 82, + 326, + 156 + ], + "blocks": [ + { + "bbox": [ + 67, + 82, + 326, + 156 + ], + "lines": [ + { + "bbox": [ + 67, + 82, + 326, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 82, + 326, + 156 + ], + "type": "table", + "html": "
MethodVBatch sizeLatencySpeedup
Open-Sora [120]84.16419.601.0×
+ AdaCache83.4089.534.7×
+ AdaCache + MoReg83.5093.504.5×
+ AdaCache + MoReg (w/o grad)83.3689.014.7×
+ AdaCache + MoReg (multi-step)83.4295.654.4×
", + "image_path": "8d1392a61046aae5824c4ca7eba8c3712bd94ce83d6df49a7d5a738e2f0c2143.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 157, + 326, + 186 + ], + "lines": [ + { + "bbox": [ + 63, + 157, + 326, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 157, + 326, + 186 + ], + "type": "text", + "content": "(a) AdaCache with Motion Regularization: We show different variants of AdaCache. All versions achieve significant speedups compared to the baseline. AdaCache + MoReg shows a better quality with a slightly-lower speedup." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 332, + 70, + 547, + 156 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 547, + 156 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 547, + 156 + ], + "type": "table", + "html": "
ResolutionAdaCacheVBenchmarkLatencySpeedup
480p - 2s✘83.68173.841.0×
✓83.1838.524.5×
480p - 4s✘82.77349.901.0×
✓82.1680.164.4×
720p - 2s✘84.16419.601.0×
✓83.4089.534.7×
", + "image_path": "481a7075343115cd24a80ead9e57c9b15dac28575289de0f4671fa59ebe603f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 328, + 157, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 328, + 157, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 328, + 157, + 545, + 186 + ], + "type": "text", + "content": "(b) Speedups at different resolutions: We compare AdaCache with baselines at different resolutions. It generalizes across resolutions, consistently providing a stable acceleration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 196, + 165, + 243 + ], + "blocks": [ + { + "bbox": [ + 61, + 196, + 165, + 243 + ], + "lines": [ + { + "bbox": [ + 61, + 196, + 165, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 165, + 243 + ], + "type": "table", + "html": "
DistanceVBenchLatency
L183.4089.53
L283.5092.70
Cosine83.1986.74
", + "image_path": "42bd02f65193b6724cc657a442ba5f65c96aa3907e196f68d096d9644b77ce3a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 244, + 166, + 282 + ], + "lines": [ + { + "bbox": [ + 56, + 244, + 166, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 166, + 282 + ], + "type": "text", + "content": "(c) Cache metric: Among distance metrics, L1/L2 have similar (and better) performance in contrast to cosine distance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 171, + 190, + 277, + 243 + ], + "blocks": [ + { + "bbox": [ + 171, + 190, + 277, + 243 + ], + "lines": [ + { + "bbox": [ + 171, + 190, + 277, + 243 + ], + "spans": [ + { + "bbox": [ + 171, + 190, + 277, + 243 + ], + "type": "table", + "html": "
LocationVBenchLatency
Start83.3087.55
Mid83.4089.53
End83.4391.20
Multiple83.4190.27
", + "image_path": "fc3f68b7099827592510104a2b630fa60ff3beb814917d97704abd09b735462e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 170, + 245, + 276, + 283 + ], + "lines": [ + { + "bbox": [ + 170, + 245, + 276, + 283 + ], + "spans": [ + { + "bbox": [ + 170, + 245, + 276, + 283 + ], + "type": "text", + "content": "(d) Cache location: We compute the cache metric at midDiT, resulting in the best quality-latency trade-off." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 283, + 190, + 384, + 243 + ], + "blocks": [ + { + "bbox": [ + 283, + 190, + 384, + 243 + ], + "lines": [ + { + "bbox": [ + 283, + 190, + 384, + 243 + ], + "spans": [ + { + "bbox": [ + 283, + 190, + 384, + 243 + ], + "type": "table", + "html": "
ResidualVBenchLatency
pt(TA)83.4089.53
pt(SA)83.1989.06
qt(CA)83.2590.70
rt(MLP)83.6299.72
", + "image_path": "09e512a8dc4a181d57cdeec573b2f384946b1d915b4e44cdeeade0d2a8c96781.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 279, + 245, + 383, + 283 + ], + "lines": [ + { + "bbox": [ + 279, + 245, + 383, + 283 + ], + "spans": [ + { + "bbox": [ + 279, + 245, + 383, + 283 + ], + "type": "text", + "content": "(e) Cache residual: We consider different residual computations to estimate cache metric. Our default is Temp-attn." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 391, + 199, + 554, + 243 + ], + "blocks": [ + { + "bbox": [ + 391, + 199, + 554, + 243 + ], + "lines": [ + { + "bbox": [ + 391, + 199, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 391, + 199, + 554, + 243 + ], + "type": "table", + "html": "
AdaCacheBasis-ratesVBranchLatency
Fast{12, 10, 8, 6, 4, 3}83.4089.53
Medium{8, 6, 4, 2, 1}83.94143.87
Slow{2, 1}84.12274.30
", + "image_path": "e287f0fb9b394a35799d4c13ccde5203c58c81b110f19438a8efb8ab6c6f52df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 388, + 244, + 554, + 282 + ], + "lines": [ + { + "bbox": [ + 388, + 244, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 388, + 244, + 554, + 282 + ], + "type": "text", + "content": "(f) AdaCache Variants: We achieve a range of speedups (and quality) by controlling the basis cache-rates in AdaCache. Our default setting is AdaCache-fast." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 322, + 295, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 295, + 369 + ], + "type": "text", + "content": "in reference-based scores (e.g. PSNR, SSIM) at extreme reduction rates, the qualitative results suggest that the generations are still good (see Fig. 1), despite not being aligned exactly with the reference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "content": "AdaCache with Motion Regularization: We compare AdaCache with different versions of motion regularization in Table 2a. Both vanilla and motion-regularized versions provide significant speedups, " + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "inline_equation", + "content": "4.7 \\times" + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "inline_equation", + "content": "4.5 \\times" + }, + { + "bbox": [ + 54, + 374, + 295, + 542 + ], + "type": "text", + "content": " respectively, at a comparable quality with baseline Open-Sora [120]. Considering motion-gradient as an early-predictor of motion at latter diffusion steps helps (83.50 vs. 83.36 on VBench). We also estimate motion at different frame-rates by considering a varying step-size in frame differences, which seems to increase the latency without necessarily improving quality. Overall, we consider AdaCache (w/ MoReg) as the configuration with the best quality-latency trade-off. This improvement in quality is more-prominent in qualitative examples in supplementary (Fig. A.2) and the user study (Fig. 4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 714 + ], + "type": "text", + "content": "Acceleration in multi-GPU setups: Aligned with prior work based on Dynamic Sequence Parallelism (DSP) [117], we evaluate how AdaCache performs in video generation across multiple GPUs. This evaluation is relevant in the context of efficiency, as DSP incurs additional latency overheads corresponding to the communication between GPUs, and caching mechanisms can avoid such costs by re-using previous computations. We present the results of this study in Fig. 6 and Table 2. Here, we consider Open-Sora [120] (480p - 2s at 30-steps) and Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) as baselines, and compare against prior-art PAB [118] in terms of latency measurements on A100 GPUs. In Fig. 6, we observe that AdaCache consistently outperforms PAB with better speeds across all settings." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 324, + 320, + 432, + 417 + ], + "blocks": [ + { + "bbox": [ + 324, + 320, + 432, + 417 + ], + "lines": [ + { + "bbox": [ + 324, + 320, + 432, + 417 + ], + "spans": [ + { + "bbox": [ + 324, + 320, + 432, + 417 + ], + "type": "image", + "image_path": "97e97b168c1ccc16708bd416f339ce78d11df4cc358a28bc7e7ebc82c88e502c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 421, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 498 + ], + "type": "text", + "content": "Figure 6. Acceleration in multi-GPU setups: We evaluate the speedups with varying GPU parallelization, as cached-steps can avoid communication overheads among GPUs. Here, we compare AdaCache with PAB [118], on baselines Open-Sora [120] 480p - 2s generations at 30-steps and Open-Sora-Plan [45] " + }, + { + "bbox": [ + 313, + 421, + 555, + 498 + ], + "type": "inline_equation", + "content": "512 \\times 512 - 2.7" + }, + { + "bbox": [ + 313, + 421, + 555, + 498 + ], + "type": "text", + "content": " s generations at 150-steps. AdaCache consistently shows better acceleration over PAB in all settings. Latency measured on A100s." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 434, + 321, + 542, + 417 + ], + "blocks": [ + { + "bbox": [ + 434, + 321, + 542, + 417 + ], + "lines": [ + { + "bbox": [ + 434, + 321, + 542, + 417 + ], + "spans": [ + { + "bbox": [ + 434, + 321, + 542, + 417 + ], + "type": "image", + "image_path": "8c9e73ca666ddaa2a3a50d011b634ac4bd56d67716e923dfae106b65cf5d94d5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": "Speedups at different resolutions: In Table 2b, we compare the trade-offs of AdaCache at various resolutions of video generations, namely, " + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "inline_equation", + "content": "480\\mathrm{p} - 2\\mathrm{s}" + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "inline_equation", + "content": "480\\mathrm{p} - 4\\mathrm{s}" + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "inline_equation", + "content": "720\\mathrm{p} - 2\\mathrm{s}" + }, + { + "bbox": [ + 313, + 506, + 555, + 567 + ], + "type": "text", + "content": ", all at 100-steps. AdaCache provides consistent speedups across different resolutions without affecting the quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Cache metric, location and residual: When adaptively deciding the caching schedule, we consider different metrics to compute the rate-of-change between representations, namely, L1/L2 distance or cosine distance. Among these, L1/L2 give an absolute measure which aligns better with the actual change. In contrast, cosine computes a normalized-distance, which is not a good estimate of change (e.g. if the representations differ only by a scale, the distance will be zero, even though we want to have a non-zero value). This observation is verified by the results in Table 2c. Moreover, we consider computing the cache metric at various locations (i.e., layers) in the DiT. Doing so at a single layer (e.g. start," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15246" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 69, + 286, + 331 + ], + "blocks": [ + { + "bbox": [ + 107, + 69, + 286, + 331 + ], + "lines": [ + { + "bbox": [ + 107, + 69, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 69, + 286, + 331 + ], + "type": "image", + "image_path": "e27b05116571f57b6efd690ef14d7ad54262733746df3acb6b3f30030a325a1e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 379 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison: We show qualitative results on multiple video-DiT baselines including Open-Sora [120] (720p - 2s at 100-steps), Open-Sora-Plan [45] (512×512 - 2.7s at 150-steps) and Latte [62] (512×512 - 2s at 50-steps), while comparing against prior training-free inference acceleration method PAB [118]. AdaCache shows a comparable generation quality at much-faster speeds. Best-viewed with zoom-in. Prompts and additional qualitative results (Fig. A.3) are given in supplementary." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 288, + 70, + 497, + 331 + ], + "blocks": [ + { + "bbox": [ + 288, + 70, + 497, + 331 + ], + "lines": [ + { + "bbox": [ + 288, + 70, + 497, + 331 + ], + "spans": [ + { + "bbox": [ + 288, + 70, + 497, + 331 + ], + "type": "image", + "image_path": "83e6469b73c482a305bea3622859285adee5ba791e51a376c4f4b23989aeb597.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 460 + ], + "type": "text", + "content": "mid, end) is not significantly different from computing an aggregate over multiple-layers (see Table 2d). By default, we compute the cache metric in the mid-layer as a reasonable choice without extra overheads. We choose temporal-attention as the residual for cache metric computation, since it achieves the best trade-off (see Table 2e)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 571 + ], + "type": "text", + "content": "AdaCache variants: To achieve a range of speedups (and quality), we consider different basis cache-rates in our AdaCache implementation. For instance, we can have higher-speedup with a slightly-lower quality (e.g. AdaCache-fast), a lower-speedup with a higher-quality (e.g. AdaCacheslow), or balance both (e.g. AdaCache-medium). We can conveniently control this by having corresponding basis cache-rates as shown in Table 2f. By default, we resort to AdaCache-fast which gives the best speedups." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 164, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 164, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 164, + 588 + ], + "type": "text", + "content": "5.4. Qualitative results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "In Fig. 7, we present qualitative results on multiple video DiT baselines, including Open-Sora [120], Open-Sora-Plan [45] and Latte [62]. We compare AdaCache against each baseline and prior training-free inference acceleration method for video DiTs, PAB [118]. Here, we consider three different configurations: " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "720\\mathrm{p}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " - 2s generations at 100-steps for Open-Sora, " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " - 2.7s generations at 150-steps for Open-Sora-Plan, and " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " - 2s generations at 50-steps for Latte, while considering prompts from OpenSora gallery (see supplementary for prompt details). Ada" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": "Cache shows a comparable generation quality, while having much-faster inference pipelines. In fact, it achieves " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "4.49 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "1.26 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " in PAB), " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "3.53 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "1.45 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " in PAB), " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "2.46 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "inline_equation", + "content": "1.33 \\times" + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": " in PAB) speedups respectively on the three considered baseline video DiTs. In most cases our generations are aligned well with the baseline in the pixel space. Yet this is not a strict requirement, as the denoising process can deviate considerably from that of the baseline, particularly at high caching rates. Still, AdaCache is faithful to the text prompt and is not affected by significant artifacts. Refer supplementary (Fig. A.3) for more qualitative examples." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced Adaptive Caching (AdaCache), a plug-and-play component that improves the inference speed of video diffusion transformers without needing any re-training. It caches residual computations, while also devising the caching schedule dependent on each video generation. We further proposed a Motion Regularization (MoReg) to utilize video information and allocate computations based on motion content, improving the quality-latency trade-off. We apply our contributions on multiple open-source video DiTs, showing comparable generation quality at a fraction of latency. We believe AdaCache is widely-applicable with minimal effort, helping democratize high-fidelity long-video generation." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15247" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 163 + ], + "type": "text", + "content": "The authors would like to thank Ankit Khedia, Arvind Somasundaram, Dustin Johnson, Eugene Vecharynski, Ly Cao, SK Bong, and Yuzi He for the interesting discussions, and Shikun Liu for the support in setting-up the project page. The authors also appreciate the time and effort volunteered by the participants of the user study." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 194, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 295, + 226 + ], + "type": "text", + "content": "[1] Shady Abu-Hussein, Tom Tirer, and Raja Giryes. Adir: Adaptive diffusion for image reconstruction. arXiv preprint arXiv:2212.03221, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 228, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 228, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 66, + 228, + 296, + 295 + ], + "type": "text", + "content": "[2] Sotiris Anagnostidis, Gregor Bachmann, Yeongmin Kim, Jonas Kohler, Markos Georgopoulos, Artsiom Sanakoyeu, Yuming Du, Albert Pumarola, Ali Thabet, and Edgar Schonfeld. Flexidit: Your diffusion transformer can easily generate high-quality samples with less compute. arXiv preprint arXiv:2502.20126, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 297, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 297, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 66, + 297, + 295, + 341 + ], + "type": "text", + "content": "[3] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18208-18218, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 342, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 342, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 66, + 342, + 295, + 373 + ], + "type": "text", + "content": "[4] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 376, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 376, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 66, + 376, + 296, + 421 + ], + "type": "text", + "content": "[5] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 422, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 422, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 66, + 422, + 296, + 486 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 489, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 489, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 66, + 489, + 296, + 555 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 66, + 557, + 295, + 600 + ], + "type": "text", + "content": "[8] Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599-4603, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 66, + 602, + 295, + 656 + ], + "type": "text", + "content": "[9] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 295, + 713 + ], + "type": "text", + "content": "[10] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 138 + ], + "type": "text", + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 140, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 554, + 195 + ], + "type": "text", + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 197, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 554, + 251 + ], + "type": "text", + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 253, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 253, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 253, + 554, + 297 + ], + "type": "text", + "content": "[14] Lei Chen, Yuan Meng, Chen Tang, Xinzhu Ma, Jingyan Jiang, Xin Wang, Zhi Wang, and Wenwu Zhu. Q-dit: Accurate post-training quantization for diffusion transformers. arXiv preprint arXiv:2406.17343, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 298, + 554, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 298, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 320, + 298, + 554, + 352 + ], + "type": "text", + "content": "[15] Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. Delta-dit: A training-free acceleration method tailored for diffusion transformers. arXiv preprint arXiv:2406.01125, 2024. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 354, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 354, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 354, + 554, + 409 + ], + "type": "text", + "content": "[16] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 411, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 554, + 475 + ], + "type": "text", + "content": "[17] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 478, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 554, + 510 + ], + "type": "text", + "content": "[18] Aidan Clark, Jeff Donahue, and Karen Simonyan. Adversarial video generation on complex datasets. arXiv preprint arXiv:1907.06571, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 554, + 555 + ], + "type": "text", + "content": "[19] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 554, + 612 + ], + "type": "text", + "content": "[20] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 614, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 554, + 656 + ], + "type": "text", + "content": "[21] Juncan Deng, Shuaiying Li, Zeyu Wang, Hong Gu, Kedong Xu, and Kejie Huang. Vq4dit: Efficient post-training vector quantization for diffusion transformers. arXiv preprint arXiv:2408.17131, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 554, + 713 + ], + "type": "text", + "content": "[22] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15248" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[23] Haiwen Feng, Zheng Ding, Zhihao Xia, Simon Niklaus, Victoria Abrevaya, Michael J Black, and Xuaner Zhang. Explorative inbetweening of time and space. arXiv preprint arXiv:2403.14611, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 182 + ], + "type": "text", + "content": "[24] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 239 + ], + "type": "text", + "content": "[25] Rohit Girdhar, Mannat Singh, Andrew Brown, Quentin Duval, Samaneh Azadi, Sai Saketh Rambhatla, Akbar Shah, Xi Yin, Devi Parikh, and Ishan Misra. Emu video: Factorizing text-to-video generation by explicit image conditioning. arXiv preprint arXiv:2311.10709, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[26] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 337 + ], + "type": "text", + "content": "[27] Xun Guo, Mingwu Zheng, Liang Hou, Yuan Gao, Yufan Deng, Chongyang Ma, Weiming Hu, Zhengjun Zha, Haibin Huang, Pengfei Wan, et al. I2v-adapter: A general image-to-video adapter for video diffusion models. arXiv preprint arXiv:2312.16693, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 294, + 403 + ], + "type": "text", + "content": "[28] Amirhossein Habibian, Amir Ghodrati, Noor Fathima, Guillaume Sautiere, Risheek Garrepalli, Fatih Porikli, and Jens Petersen. Clockwork diffusion: Efficient generation with model-step distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8352-8361, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 404, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 447 + ], + "type": "text", + "content": "[29] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 491 + ], + "type": "text", + "content": "[30] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 493, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 294, + 557 + ], + "type": "text", + "content": "[31] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streaming2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "text", + "content": "[32] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 604, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 294, + 635 + ], + "type": "text", + "content": "[33] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 294, + 690 + ], + "type": "text", + "content": "[34] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Image video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[35] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 338, + 74, + 553, + 94 + ], + "type": "text", + "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 150 + ], + "type": "text", + "content": "[36] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 217 + ], + "type": "text", + "content": "[37] Rongjie Huang, Jiawei Huang, Dongchao Yang, Yi Ren, Luping Liu, Mingze Li, Zhenhui Ye, Jinglin Liu, Xiang Yin, and Zhou Zhao. Make-an-audio: Text-to-audio generation with prompt-enhanced diffusion models. In International Conference on Machine Learning, pages 13916-13932. PMLR, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 220, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 553, + 296 + ], + "type": "text", + "content": "[38] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 320, + 297, + 553, + 341 + ], + "type": "text", + "content": "[39] Kumara Kahatapitiya, Adil Karjauv, Davide Abati, Fatih Porikli, Yuki M Asano, and Amirhossein Habibian. Object-centric diffusion for efficient video editing. arXiv preprint arXiv:2401.05735, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 342, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 342, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 321, + 342, + 553, + 387 + ], + "type": "text", + "content": "[40] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 388, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 409 + ], + "type": "text", + "content": "[41] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 411, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 411, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 321, + 411, + 553, + 432 + ], + "type": "text", + "content": "[42] Inc. Kling AI. Next-generation ai creative studio. URL https://klingai.com/, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 434, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 434, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 553, + 488 + ], + "type": "text", + "content": "[43] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 490, + 553, + 521 + ], + "type": "text", + "content": "[44] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 523, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 553, + 544 + ], + "type": "text", + "content": "[45] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 5,6,7,8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 546, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 546, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 546, + 553, + 601 + ], + "type": "text", + "content": "[46] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 603, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 603, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 603, + 553, + 656 + ], + "type": "text", + "content": "[47] Xinui Li, Chao Ma, Xiaokang Yang, and Ming-Hsuan Yang. Vidtome: Video token merging for zero-shot video editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7486-7495, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 713 + ], + "type": "text", + "content": "[48] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15249" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[49] Haozhe Liu, Shikun Liu, Zijian Zhou, Mengmeng Xu, Yanping Xie, Xiao Han, Juan C Pérez, Ding Liu, Kumara Kahatapitiya, Menglin Jia, et al. Mardini: Masked autoregressive diffusion for video generation at scale. arXiv preprint arXiv:2410.20280, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 173 + ], + "type": "text", + "content": "[50] Joseph Liu, Joshua Geddes, Ziyu Guo, Haomiao Jiang, and Mahesh Kumar Nandwana. Smoothcache: A universal inference acceleration technique for diffusion transformers. arXiv preprint arXiv:2411.10510, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 295, + 228 + ], + "type": "text", + "content": "[51] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In European Conference on Computer Vision, pages 423-439. Springer, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 296, + 285 + ], + "type": "text", + "content": "[52] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 294, + 330 + ], + "type": "text", + "content": "[53] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "type": "text", + "content": "[54] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 430 + ], + "type": "text", + "content": "[55] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 475 + ], + "type": "text", + "content": "[56] Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 294, + 499 + ], + "type": "text", + "content": "[57] Inc. Luma AI. Dream machine. URL https://lumalabs.ai/dream-machine, 2024.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "type": "text", + "content": "[58] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 294, + 589 + ], + "type": "text", + "content": "[59] Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K Wong. Fastercache: Training-free video diffusion model acceleration with high quality. arXiv preprint arXiv:2410.19355, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 645 + ], + "type": "text", + "content": "[60] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 690 + ], + "type": "text", + "content": "[61] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15762-15772, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[62] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 95 + ], + "type": "text", + "content": "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 138 + ], + "type": "text", + "content": "[63] Yifang Men, Yuan Yao, Miaomiao Cui, and Bo Liefeng. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 553, + 194 + ], + "type": "text", + "content": "[64] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 195, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 195, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 321, + 195, + 553, + 238 + ], + "type": "text", + "content": "[65] Yotam Nitzan, Zongze Wu, Richard Zhang, Eli Shechtman, Daniel Cohen-Or, Taesung Park, and Michael Gharbi. Lazy diffusion transformer for interactive image editing. arXiv preprint arXiv:2404.12382, 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 553, + 271 + ], + "type": "text", + "content": "[66] Inc. OpenAI. Video generation models as world simulators. 2024. URL https://openai.com/research/video-generation-models-as-world-simulators, 3, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 553, + 315 + ], + "type": "text", + "content": "[67] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 317, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 317, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 321, + 317, + 553, + 369 + ], + "type": "text", + "content": "[68] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 321, + 372, + 553, + 403 + ], + "type": "text", + "content": "[69] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 321, + 405, + 553, + 459 + ], + "type": "text", + "content": "[70] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15932–15942, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 321, + 460, + 553, + 502 + ], + "type": "text", + "content": "[71] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 504, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 553, + 525 + ], + "type": "text", + "content": "[72] Jason Tyler Rolfe. Discrete variational autoencoders. arXiv preprint arXiv:1609.02200, 2016. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 321, + 526, + 553, + 581 + ], + "type": "text", + "content": "[73] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 582, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 582, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 321, + 582, + 553, + 647 + ], + "type": "text", + "content": "[74] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 689 + ], + "type": "text", + "content": "[75] Inc. Runway AI. Gen-3 alpha: A new frontier for video generation. URL https://runwayml.com/research/introducing-gen-3-alpha, 2024.2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 691, + 553, + 713 + ], + "type": "text", + "content": "[76] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15250" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 62, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 117 + ], + "type": "text", + "content": "Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 62, + 118, + 295, + 162 + ], + "type": "text", + "content": "[77] Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value clipping. In Proceedings of the IEEE international conference on computer vision, pages 2830-2839, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 295, + 195 + ], + "type": "text", + "content": "[78] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 230 + ], + "type": "text", + "content": "[79] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 295, + 274 + ], + "type": "text", + "content": "[80] Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. Fora: Fast-forward caching in diffusion transformer acceleration. arXiv preprint arXiv:2407.01425, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 330 + ], + "type": "text", + "content": "[81] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 365 + ], + "type": "text", + "content": "[82] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 367, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 295, + 399 + ], + "type": "text", + "content": "[83] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 401, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 401, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 401, + 295, + 465 + ], + "type": "text", + "content": "[84] Nikita Starodubcev, Dmitry Baranchuk, Artem Fedorov, and Artem Babenko. Your student is better than expected: Adaptive teacher-student collaboration for text-conditional diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275–9285, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 533 + ], + "type": "text", + "content": "[85] Maitreya Suin, Nithin Gopalakrishnan Nair, Chun Pong Lau, Vishal M Patel, and Rama Chellappa. Diffuse and restore: A region-adaptive diffusion model for identity-preserving blind face restoration. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6343-6352, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 578 + ], + "type": "text", + "content": "[86] Gary J Sullivan, Jens-Rainer Ohm, Woo-Jin Han, and Thomas Wiegand. Overview of the high efficiency video coding (hevc) standard. IEEE Transactions on circuits and systems for video technology, 22(12):1649-1668, 2012. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 295, + 612 + ], + "type": "text", + "content": "[87] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 657 + ], + "type": "text", + "content": "[88] Shengkun Tang, Yaqing Wang, Caiwen Ding, Yi Liang, Yao Li, and Dongkuan Xu. Deediff: Dynamic uncertainty-aware early exiting for accelerating diffusion model generation. arXiv preprint arXiv:2309.17074, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 659, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 295, + 712 + ], + "type": "text", + "content": "[89] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526-1535, 2018. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 139 + ], + "type": "text", + "content": "[90] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 321, + 141, + 553, + 173 + ], + "type": "text", + "content": "[91] Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 553, + 217 + ], + "type": "text", + "content": "[92] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-l-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 553, + 274 + ], + "type": "text", + "content": "[93] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. arXiv preprint arXiv:2403.13745, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 277, + 553, + 319 + ], + "type": "text", + "content": "[94] Haoxuan Wang, Yuzhang Shang, Zhihang Yuan, Junyi Wu, and Yan Yan. Quest: Low-bit diffusion model quantization via efficient selective finetuning. arXiv preprint arXiv:2402.03666, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 322, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 322, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 321, + 322, + 553, + 396 + ], + "type": "text", + "content": "[95] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 553, + 453 + ], + "type": "text", + "content": "[96] Xiaojuan Wang, Boyang Zhou, Brian Curless, Ira Kemelmacher-Shlizerman, Aleksander Holynski, and Steven M Seitz. Generative inbetweening: Adapting imageto-video models for keyframe interpolation. arXiv preprint arXiv:2408.15239, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 499 + ], + "type": "text", + "content": "[97] Thomas Wiegand, Gary J Sullivan, Gisle Bjontegaard, and Ajay Luthra. Overview of the h. 264/avc video coding standard. IEEE Transactions on circuits and systems for video technology, 13(7):560-576, 2003. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 501, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 501, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 501, + 553, + 566 + ], + "type": "text", + "content": "[98] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 568, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 633 + ], + "type": "text", + "content": "[99] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[100] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. arXiv preprint arXiv:2408.13239, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[101] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15251" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 116 + ], + "type": "text", + "content": "Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[102] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 218 + ], + "type": "text", + "content": "[103] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[104] Dejia Xu, Weili Nie, Chao Liu, Sifei Liu, Jan Kautz, Zhangyang Wang, and Arash Vahdat. Camco: Camera-controllable 3d-consistent image-to-video generation. arXiv preprint arXiv:2406.02509, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[105] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[106] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 375 + ], + "type": "text", + "content": "[107] Xingyi Yang and Xinchao Wang. Compositional video generation as flow equalization. arXiv preprint arXiv:2407.06182, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 421 + ], + "type": "text", + "content": "[108] Xingyi Yang, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Diffusion probabilistic model made slim. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 22552-22562, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 477 + ], + "type": "text", + "content": "[109] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8466-8476, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[110] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 589 + ], + "type": "text", + "content": "[111] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[112] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[113] Wentian Zhang, Haozhe Liu, Jinheng Xie, Francesco Faccio, Mike Zheng Shou, and Jurgen Schmidhuber. Cross-attention makes inference cumbersome in text-to-image diffusion models. arXiv preprint arXiv:2404.02747v1, 2024.5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[114] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffu" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 420 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "sion transformer for video generation. arXiv preprint arXiv:2407.21705, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[115] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence. arXiv preprint arXiv:2407.16655, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[116] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jiawei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. arXiv preprint arXiv:2310.08465, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 240 + ], + "type": "text", + "content": "[117] Xuanlei Zhao, Shenggan Cheng, Zangwei Zheng, Zheming Yang, Ziming Liu, and Yang You. Dsp: Dynamic sequence parallelism for multi-dimensional transformers. arXiv preprint arXiv:2403.10266, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "text", + "content": "[118] Xuanlei Zhao, Xiaolong Jin, Kai Wang, and Yang You. Real-time video generation with pyramid attention broadcast. arXiv preprint arXiv:2408.12588, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 329 + ], + "type": "text", + "content": "[119] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 375 + ], + "type": "text", + "content": "[120] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[121] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15252" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_content_list.json b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a66b7c6bdae368c230388a2ef3921700936ba7df --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_content_list.json @@ -0,0 +1,1780 @@ +[ + { + "type": "text", + "text": "Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts", + "text_level": 1, + "bbox": [ + 127, + 128, + 870, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zixuan Hu $^{1}$ Dongxiao Li $^{1}$ Xinzhu Ma $^{3}$ Shixiang Tang $^{3}$ Xiaotong Li $^{1}$ Wenhan Yang $^{2}$ Ling-Yu Duan $^{1,2*}$", + "bbox": [ + 184, + 202, + 807, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Peking University, Beijing, China", + "bbox": [ + 285, + 239, + 712, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peng Cheng Laboratory, Shenzhen, China, $^{3}$ The Chinese University of Hong Kong, Hongkong, China.", + "bbox": [ + 156, + 257, + 836, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hzxuan, lingyu}@pku.edu.cn", + "bbox": [ + 372, + 277, + 611, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types. The source code is available at https://github.com/hzcar/DUO.", + "bbox": [ + 86, + 357, + 483, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 823, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D object detection (M3OD) [1, 33, 34] serves as a fundamental perception task, enabling agents to un", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51296b4f1e88d12619b34438f118bb4f0030761c26ff53f3f9edfc3cf01536e5.jpg", + "image_caption": [ + "(a) Overview of Monocular 3D Object Detection" + ], + "image_footnote": [], + "bbox": [ + 527, + 340, + 900, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ed334cae71a4b29534ce7241d1691e88c3a62fb38ecd6712fdab21d5efcd5503.jpg", + "image_caption": [ + "Figure 1. (a) Illustration of M3OD. $(b, c)$ Under test-time shifts, increased semantic and geometric uncertainties lead to a degradation of semantic perception and spatial localization capabilities." + ], + "image_footnote": [], + "bbox": [ + 527, + 428, + 901, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "derstand 3D scenes directly from 2D images, as shown in Fig. 1(a). Due to its low cost and simple hardware configuration, M3OD has attracted widespread attention, leading to the development of numerous detectors [7, 31, 52]. However, when facing adverse weather conditions or sensor failures in real-world deployments, well-trained models often suffer severe performance degradation under data shifts [2, 14, 49]. Therefore, it is crucial to deal with the OOD generalization problem for M3OD.", + "bbox": [ + 509, + 597, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the distribution shifting issue with minimal overhead, Test-Time Adaptation (TTA) has emerged as a critical paradigm, enabling source models to adapt to target distributions via online updates [17, 19, 23]. The dominating strategy in this context involves minimizing prediction entropy, thereby reducing the uncertainty of the model on shifted data [17, 46]. Despite the promising results, existing TTA approaches largely overlook the dual uncertainty inherent in 3D detection—semantic uncertainty (related to class predictions) [42, 50] and geometric uncertainty (related to spatial location) [30, 39], which is a significant dif", + "bbox": [ + 509, + 734, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author.", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7273", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ference compared to conventional 2D tasks.", + "bbox": [ + 89, + 90, + 380, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To investigate these two types of uncertainty under test-time shifts, we analyze detection outcomes for objects subject to common real-world variations. As shown in Fig. 1 (b)&(c), our empirical study reveals that both uncertainties increase markedly with data shifts, creating compounded error accumulation in 3D detection. Moreover, we find that existing uncertainty optimization techniques exhibit significant limitations: 1) Low-score object neglect. Entropy minimization fails to provide effective supervision for challenging objects with low detection scores, resulting in inevitable omissions. 2) Spatial perception collapse. Direct minimization of depth uncertainty can cause model collapse, compromising the perception capacity of spatial attributes.", + "bbox": [ + 88, + 108, + 480, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, to overcome the above limitations, we propose Dual Uncertainty Optimization (DUO), the first TTA framework for joint semantic-geometric uncertainty minimization. Specifically, through the lens of convex optimization theory [3, 45], we present an Legendre-Fenchel structure of the focal loss [25] and reconstruct the semantic uncertainty minimization as a dual optimization problem. Building on this foundation, we apply higher-order approximation analysis to derive a novel Conjugate Focal Loss. This loss breaks the label-dependence barrier in the original objective and introduces a dynamic weighting mechanism for balanced training, effectively reducing the omissions of objects with low scores. In parallel, we introduce a normal field constraint that enforces local consistency of surface normals in regions with high semantic confidence. This spatial coherence clarifies geometric cues, thereby reducing geometric uncertainty. Together, this dual-branch design creates a complementary loop where semantic confident regions bootstrap geometric feature learning, while enhanced spatial perception guides semantic refinement.", + "bbox": [ + 91, + 306, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate the effectiveness and generalizability of our method through experiments on the KITTI dataset [11] with 13 corruption shift types, achieving state-of-the-art results with average improvements of $+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}$ in the Car category. Furthermore, we showcase its superior performance in addressing real-world shift scenarios (daytime $\\leftrightarrow$ night, sunny $\\leftrightarrow$ rainy) of nuScenes dataset [5], yielding an average gain of $+18\\%$ compared to existing methods.", + "bbox": [ + 89, + 611, + 482, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions: 1) To the best of our knowledge, we pioneer dual uncertainty optimization in M3OD by establishing the first TTA framework that jointly minimizes semantic and geometric uncertainties, addressing a critical reliability gap in real-world deployments. 2) Through the lens of convex optimization theory, we derive a novel Conjugate Focal Loss that enables label-agnostic uncertainty weighting and balanced learning for low-score objects. This approach is inherently compatible with the source phase, requiring no additional hyperparameter tuning. 3) We introduce a normal field constraint that enforces the stability of geometric", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "representation with semantic guidance, resolving ambiguous spatial predictions. 4) We analyze and verify that our dual-branch design creates a complementary loop of two types of uncertainty optimization, resulting in significantly improved performance over existing TTA methods.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 172, + 655, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular 3D Object Detection (M3OD) aims to perceive 3D objects from a single 2D image. Existing methods in M3OD can be broadly categorized based on their use of extra data sources, such as CAD models [28], dense depth maps [8, 47], or LiDAR [18, 41]. In this paper, we focus exclusively on approaches that utilize only monocular images, due to their computational efficiency and lower deployment costs. Previous studies, such as MonoDLE [32] and PGD [48], have identified depth estimation as a critical bottleneck in M3OD. To address this, many works leverage multiple geometric cues to integrate diverse depth predictions. For example, MonoFlex [55] integrates depth prediction by combining direct regression with multi-keypoint estimation; MonoGround [40] incorporates the ground plane as prior information; and MonoCD [53] exploits the complementary properties of multi-head estimation. In this paper, we investigate how to enhance the detection performance of such detectors under test-time shifts.", + "bbox": [ + 511, + 195, + 903, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Test-Time Adaptation (TTA) aims to enhance model performance on out-of-distribution samples during inference. Depending on whether the source training process is modified, existing TTA methods can be mainly divided into two groups: Test-Time Training [13, 27, 44] and Fully Test-Time Adaptation [4, 16, 56]. In this paper, we focus on Fully TTA, which adapts models without source data. Prior studies have demonstrated that reducing prediction uncertainty is an effective strategy for improving model generalization in various tasks [10, 21, 26, 51]. These works have developed various strategies to model and optimize uncertainty. For instance, SAR [37] uses entropy as a measure of classification uncertainty along with sharpness-aware optimization. DeYO [22] incorporates entropy with a disentangled factor for uncertainty modeling. ReCAP [17] models regional uncertainty, enforcing implicit data scaling for uncertainty optimization. MonoTTA [24] optimizes positive and negative class uncertainties separately. However, unlike conventional 2D tasks, M3OD inherently exhibits both semantic and geometric uncertainties, which remain largely unexplored and lack an effective training paradigm. In this work, we explore the joint optimization of this dual uncertainty, enhancing the robustness of M3OD models.", + "bbox": [ + 511, + 468, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 829, + 638, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Definition. M3OD aims to predict the 3D and semantic attributes of objects from a single RGB image. Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , the goal is to generate accurate", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7274", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4fd75fea96986fd1e56e5e2a0b6efa144c0b2e6ee797b45f98b7d0d0cd5b93a3.jpg", + "image_caption": [ + "Figure 2. Semantic uncertainty (prediction entropy) and geometric uncertainty (average depth uncertainty) of the M3OD model under domain shifts. (a) shows variations in dual-uncertainty as the shift level increases. (b) records detection scores of predicted objects during inference with the source model. (c) records scores during the entropy minimization process. (d) investigates the depth uncertainty of different heads during the optimization with uncertainty regression loss. Experiments are conducted on the Gauss shift type of KITTI-C." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 294, + 210 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d6f655a59813be432ce02fd130fedc4ddb96960b34278be8b618ed379a7132ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 88, + 500, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cb2fe78abf57b64a5c0ec37fb380299220e70f35dd56280acf667131d80f2017.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 700, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7768d4e328fbde4aa0503cc0c40d6c862df1e126c9327ca51ed07e6a1b4030ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 89, + 903, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D bounding boxes $\\{\\mathcal{B}_i\\}_{i=1}^N$ and semantic labels $\\{\\mathcal{C}_i\\}_{i=1}^N$ for objects present in the scene, where $N$ denotes the number of objects. Each bounding box is typically parameterized as $\\mathcal{B}_i = (\\mathbf{P}_i, \\mathbf{D}_i, \\mathbf{O}_i)$ , where $\\mathbf{P}_i \\in \\mathbb{R}^3$ denotes the 3D center position, $\\mathbf{D}_i \\in \\mathbb{R}^3$ represents the shape dimensions, and $\\mathbf{O}_i \\in [-\\pi, \\pi]$ encodes the orientation. The multi-task nature of M3OD, which involves simultaneous estimation of both geometric and semantic attributes, poses significant challenges for achieving precise and coherent predictions.", + "bbox": [ + 88, + 290, + 480, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-framework. As shown in Fig. 3(a), M3OD models typically employ a backbone network connected with multiple branches to predict various properties, e.g., score heatmaps, depth maps, etc. Since depth estimation is widely recognized as a key bottleneck [32, 48], many approaches employ a multi-head depth estimator to improve prediction accuracy. This estimator comprises a direct regression depth head and multiple geometric depth heads, each providing an individual depth prediction along with an associated uncertainty value. These outputs are then fused via an uncertainty-weighted average to produce the final depth prediction. We utilize the average uncertainty across all heads as our depth uncertainty metric. A detailed explanation is provided in the Appendix. D.2.", + "bbox": [ + 88, + 426, + 482, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TTA Setting. TTA addresses the challenge of distribution shifts by enabling a pre-trained model to adapt to the target distribution during inference, without the need for labeled data. Unlike traditional training on fixed datasets, TTA operates in an online manner, where the model $h_{\\theta}$ with parameter set $\\theta$ produces detection outputs while concurrently updating its parameters based on incoming test data:", + "bbox": [ + 88, + 638, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathcal {B} _ {t} \\right\\}, \\left\\{\\mathcal {C} _ {t} \\right\\} \\leftarrow h _ {\\theta} \\left(\\mathrm {I} _ {t}\\right), \\theta \\leftarrow \\theta - \\nabla \\mathcal {L} _ {t t a} \\left(\\mathrm {I} _ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 750, + 482, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{I}_t$ denotes the incoming test data and $\\mathcal{L}_{tta}$ denotes the loss function used for self-training during adaptation.", + "bbox": [ + 89, + 775, + 482, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Uncovering Dual-Uncertainty under Shifts", + "text_level": 1, + "bbox": [ + 89, + 814, + 470, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since TENT [46] revealed the positive correlation between prediction uncertainty and generalization error under distribution shifts, numerous TTA methods have emerged to optimize uncertainty metrics. While previous works have devel", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "oped diverse measures for conventional 2D tasks [22, 37], the compound semantic and geometric uncertainties in 3D perception remains critically unexplored in the TTA field. In this work, we investigate both types of uncertainty within the context of M3OD to understand their distinct roles.", + "bbox": [ + 511, + 291, + 903, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, to quantify the variation of uncertainty under distribution shifts, we track two metrics: semantic prediction entropy and geometric depth uncertainty (average uncertainty of multi-head depth estimator). As shown in Fig. 2(a), both metrics demonstrate a consistent upward trend as distribution shifts intensify, indicating that more severe shifts lead to higher model uncertainty. Furthermore, we analyze the independent optimization effects of two uncertainties and make the following key observations:", + "bbox": [ + 511, + 367, + 905, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 1: Conventional entropy minimization exacerbates imbalance distribution of detection scores. Different from classification scenarios, object detection suffers from extreme foreground-background imbalance, which hinders the effective optimization of hard positive objects [6, 25]. As shown in Fig. 2(b)&(c), entropy minimization yields a marginal gain for low-score objects while significantly boosting high-score predictions, further exacerbating the imbalance and leading to omissions of low-score objects.", + "bbox": [ + 511, + 503, + 905, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Observation 2: Minimizing depth uncertainty directly causes the model collapse of the multi-head depth estimator. To minimize depth uncertainty, we apply the uncertainty regression loss across multiple depth heads (detailed in the Appendix. D.2). As shown in Fig. 2(d), the regression head which lacks any geometric constraints, exhibits a significantly faster decline of uncertainty compared to keypoint heads. This rapid convergence reduces the multi-head system to a single deterministic head, undermining the model's ability to perform robust spatial understanding.", + "bbox": [ + 511, + 638, + 905, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "5. Methodology", + "text_level": 1, + "bbox": [ + 511, + 799, + 648, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on the above observations, we propose a TTA method. Dual Uncertainty Optimization (DUO) for M3OD, which mainly leverages two novel designs, i.e., conjugate optimization framework and normal consistency constraint, to compatibly optimize semantic and geometric uncertainties.", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7275", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/82b02825d62808bc0a954dfd6d7b61f0b5b229de88224d908cc2d95fe40d88e5.jpg", + "image_caption": [ + "Figure 3. Overview of our DUO method. (a) Meta-framework of the base models, where a backbone network connects to multi-branch predictors for estimating various object properties. (b) Through our theoretical conjugate optimization framework, we derive a conjugate focal loss $\\mathcal{L}_{\\mathrm{CFL}}$ that can adaptively adjust training weights across all classes without the ground-truth labels, while also identifying low-semantic-uncertainty regions for geometric constraints (Sec. 5.1). (c) DUO employs the efficient Sobel operator to convert the depth map to a normal field and design a normal consistency constraint $\\mathcal{L}_{\\mathrm{NCL}}$ to enhance geometric consistency in selected regions (Sec. 5.2)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 390, + 292 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4885a4ae85addbce8880b2ccb215519ee8e39802a24770e6b3c5ec0df8eb0ffa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 88, + 901, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Semantic Uncertainty Conjugate Optimization", + "text_level": 1, + "bbox": [ + 89, + 375, + 482, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the imbalance problem, it is crucial to devote more attention to challenging objects with high uncertainty. A straightforward approach is to employ the focal loss [25] which increases the weight of predictions with low probability. However, focal loss cannot assign an appropriate weight without the ground-truth label and fails to adapt effectively in unsupervised settings. To provide a more flexible weighting solution, we leverage convex optimization theory [3, 45] (a classical analytical tool) to explore the focal loss via a Legendre-Fenchel structure. Building on this foundation, we perform a higher-order approximation analysis to derive a novel loss that can dynamically adjust the weighting without relying on labels, as shown in Fig. 3(b).", + "bbox": [ + 88, + 397, + 482, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Legendre-Fenchel Structure. Given the source model $h_\\theta$ pre-trained with the focal loss, we denote $h \\triangleq h_\\theta(x)$ and its corresponding loss function can be formulated as:", + "bbox": [ + 89, + 594, + 482, + 640 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F L}} (x, y) = - \\alpha (1 - p) ^ {\\gamma} \\cdot y \\log p, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 648, + 482, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p \\triangleq \\text{softmax}(h)$ is the normalized probability over the classes, $x$ is a detected object, and $y$ is the one-hot coding of the ground-truth label. $\\alpha$ and $\\gamma$ are hyperparameters controlling the weighting of uncertain predictions. Motivated by [12], we reformulate it as the following structure:", + "bbox": [ + 89, + 676, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F L}} = f (h) - y ^ {\\top} g (h),\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 761, + 367, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (h) = \\alpha \\log s, g (h) = \\alpha h + \\alpha ((1 - p) ^ {\\gamma} - 1) \\log p, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 777, + 480, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s \\triangleq e^{h_1} + \\ldots + e^{h_c}$ is the sum over the exponential outputs of the model and $c$ is the number of classes. Refer to the Appendix. A for missing proofs of this subsection.", + "bbox": [ + 89, + 809, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Reconstruction. Under this structure, the optimization problem can be regarded as finding an optimal representation $h$ that minimizes the empirical loss. According", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to the Legendre-Fenchel condition [45], the existence of the conjugate function $f^{*}$ is equivalent to the invertibility of the function $g$ . This critical condition is formally established in our work (see Appendix. A), serving as a fundamental theoretical guarantee for our method. Therefore, the minimum value of the objective can be expressed as follows:", + "bbox": [ + 511, + 376, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {h} \\left\\{f (h) - y ^ {\\top} g (h) \\right\\} = \\min _ {z = g (h)} \\left\\{f \\circ g ^ {- 1} (z) - y ^ {\\top} z \\right\\} = f ^ {*} (y). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 470, + 903, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the common assumption that the representation $h$ pre-trained from the large source dataset is already close to a locally optimal solution $h_0$ [9, 35], we can convert the problem into the following relationships:", + "bbox": [ + 511, + 507, + 905, + 569 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf \\circ g ^ {- 1} (z) - y ^ {\\top} z = f ^ {*} (y), \\quad \\nabla_ {z} \\left(f \\circ g ^ {- 1}\\right) = y. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 574, + 903, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conjugate Focal Loss. By applying the chain rule of gradient computation and higher-order approximation, we can obtain the following estimation:", + "bbox": [ + 511, + 599, + 905, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} y _ {0} \\triangleq \\frac {\\nabla_ {h} \\left(f \\circ g ^ {- 1}\\right)}{\\nabla_ {h} z} | _ {z = g (h)} = \\nabla_ {h} g (h) ^ {- 1} \\cdot \\nabla_ {h} f (h) \\tag {6} \\\\ \\approx (I + \\gamma (1 - \\log p) \\cdot p p ^ {\\top} - \\gamma \\log p \\cdot \\operatorname {d i a g} (p))) ^ {- 1} p, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 652, + 903, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{diag}(\\cdot)$ denotes the diagonal matrix and $I$ denotes the identity matrix. This estimation $y_{0}$ ensures that the loss aligns closely with the conjugate function. Ultimately, we substitute it into Equ. 5, yielding an unsupervised approximation of the conjugate function:", + "bbox": [ + 511, + 712, + 905, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C F L}} (x) = - \\alpha (1 - p) ^ {\\gamma} (I + \\gamma (1 - \\log p) \\cdot p ^ {\\top} p \\tag {7} \\\\ - \\gamma \\log p \\cdot \\operatorname {d i a g} (p)) ^ {- 1} p \\log p. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 794, + 903, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast to the focal loss in Equ. 2, our derived Conjugate Focal Loss (CFL) offers a dynamic adjustment of uncertainty weighting without relying on ground-truth labels and it confers the following advantages:", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7276", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Static vs. Dynamic Adjustment. The vanilla focal loss uses a fixed weighting term for the ground-truth class to focus on high-uncertainty predictions. In contrast, our CFL not only incorporates $(1 - p)^{\\gamma}$ to address the imbalance training issue but also dynamically adjusts the weighting across all classes based on the term $(I + \\gamma (1 - \\log p)pp^{\\top} - \\gamma \\log p\\cdot \\mathrm{diag}(p))^{-1}$ , which encodes inter-class prediction relationships.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ground-Truth Independence. While focal loss requires labels to compute the weight for the loss term, CFL operates solely based on the prediction probability, eliminating the need for labeled data during TTA.", + "bbox": [ + 89, + 196, + 480, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compatible Hyperparameters with Source Phase. Our theoretical analysis suggests that hyperparameters $\\alpha, \\gamma$ should remain consistent with their values used in focal loss during the source training. This compatibility provides a practical advantage for TTA scenarios, eliminating the need for extensive hyperparameter tuning. The validity of this consistent setting is empirically verified in Appendix C.", + "bbox": [ + 89, + 257, + 480, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Semantic-Guided Normal Field Constraint", + "text_level": 1, + "bbox": [ + 89, + 368, + 454, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite progress in uncertainty modeling, current methods struggle to handle geometric uncertainty in TTA scenarios due to critical limitations: 1) Model Collapse: Direct optimization of model-predicted uncertainty leads to the degenerated predictor, as discussed in Sec. 4. 2) Substantial Overhead: Geometry-aware methods typically require additional data or offline training for uncertainty estimation, limiting their real-time applicability [15, 39]. To overcome these challenges, we propose an efficient normal field constraint derived from a single image, which enhances the geometric coherence of 3D predictions, thereby reducing the uncertainty stemming from the unstable geometric representation, as shown in Fig. 3(c).", + "bbox": [ + 89, + 388, + 482, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal Field. Given a depth map $D$ , we restore it to the original image resolution using bilinear interpolation, allowing for the alignment with the pixel grid. Then, we compute the spatial gradients of the depth map using efficient Sobel operators [20], which approximate the rate of change in the depth values along horizontal and vertical directions:", + "bbox": [ + 89, + 584, + 482, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla D _ {x} = \\mathbf {S} _ {x} * D, \\quad \\nabla D _ {y} = \\mathbf {S} _ {y} * D, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 685, + 480, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_x$ and $\\mathbf{S}_y$ denote horizontal and vertical Sobel kernels, respectively. These gradient maps capture the variation in depth across neighboring pixels. The surface normal field, which encodes the orientation of the surface at each pixel, is then derived from the gradients as:", + "bbox": [ + 89, + 710, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {N} (u, v) = \\frac {1}{\\sqrt {\\nabla D _ {x} ^ {2} + \\nabla D _ {y} ^ {2} + 1}} \\left[ \\begin{array}{c} - \\nabla D _ {x} \\\\ - \\nabla D _ {y} \\\\ 1 \\end{array} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 795, + 480, + 845 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{N}(u,v)$ denotes the normal orientation at pixel $(u,v)$ . Normal Consistency Loss. To quantify geometric uncertainty, we employ an edge-aware Normal Consistency Loss", + "bbox": [ + 89, + 854, + 491, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(NCL), that encourages smoothness in the local surface by penalizing inconsistencies between neighboring pixels:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {x} (u, v) = \\left\\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u + 1, v) - \\mathbf {N} (u - 1, v) \\right\\| _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 127, + 893, + 146 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi_ {y} (u, v) = \\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u, v + 1) - \\mathbf {N} (u, v - 1) \\| _ {2} ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 148, + 903, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the smoothness terms $\\psi_x(u,v),\\psi_y(u,v)$ enforce horizontal and vertical consistency in the surface normal field. The total normal consistency loss is then given by:", + "bbox": [ + 511, + 179, + 903, + 224 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {N C L}} (u, v) = \\left(\\psi_ {x} (u, v) + \\psi_ {y} (u, v)\\right) \\cdot \\exp (- \\| \\nabla \\mathrm {I} (u, v) \\| _ {2}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 231, + 903, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the edge-aware weighting term $\\exp (-\\| \\nabla \\mathbf{I}(u,v)\\| _2)$ $= \\exp (-\\sqrt{|\\mathbf{S}_x*\\mathbf{I}(u,v)|^2 + |\\mathbf{S}_y*\\mathbf{I}(u,v)|^2})$ preserves discontinuities at boundaries while enforcing smoothness in homogeneous regions. NCL encourages the model to learn a spatially consistent normal field, thereby reducing uncertainty stemming from unstable 3D representations.", + "bbox": [ + 511, + 253, + 903, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Guidance. To enforce geometric-semantic coherence, we generate masks by integrating 2D bounding boxes with semantic predictions, ensuring synchronized focus regions for dual-branch uncertainty optimization. Let $\\{\\mathcal{B}_i\\}_{i=1}^n$ denote detected bounding boxes with scores $\\{s_i\\}_{i=1}^n$ and semantic uncertainties $\\{U_i\\}_{i=1}^n$ derived from the Conjugate Focal Loss ( $\\mathcal{L}_{\\mathrm{CFL}}$ in Eq. 7). To ensure reliable supervision, we select boxes with low semantic uncertainty using an exponentially moving average threshold:", + "bbox": [ + 511, + 344, + 905, + 481 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR = \\{i \\mid U _ {i} \\leq \\bar {U} \\}, \\bar {U} \\leftarrow \\beta \\cdot \\frac {\\sum_ {i = 1} ^ {n} U _ {i}}{n} + (1 - \\beta) \\cdot \\bar {U}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 491, + 903, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta \\in [0,1]$ is a moving average factor (set to 0.1 in default). We then construct the region mask as follows:", + "bbox": [ + 511, + 527, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (u, v) = \\max _ {i \\in R} s _ {i} \\cdot \\mathbb {I} _ {\\text {i n s i d e}} (u, v \\mid \\mathcal {B} _ {i}), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 565, + 903, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{I}_{\\mathrm{inside}}(u,v\\mid \\mathcal{B}_i)$ is an indicator function that returns 1 if pixel $(u,v)$ is inside $\\mathcal{B}_i$ , and 0 otherwise. This semantic-guided mask ensures that only low semantic-uncertainty regions contribute to the normal field constraint, enhancing the reliability of the geometric constraint.", + "bbox": [ + 511, + 595, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Overall Objective", + "text_level": 1, + "bbox": [ + 513, + 675, + 687, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective integrates the conjugate focal loss with the normal consistency constraint into a unified framework, simultaneously addressing semantic and geometric uncertainties. This dual optimization enables a complementary feedback loop: low-uncertainty spatial location enhances the model's ability to perform precise semantic classification, while confident semantic predictions in turn improve spatial understanding. The procedure is as follows:", + "bbox": [ + 511, + 696, + 905, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\Sigma_ {x \\in I} \\mathcal {L} _ {\\mathrm {C F L}} (x) + \\lambda \\Sigma_ {(u, v) \\in I} \\mathcal {M} (u, v) \\cdot \\mathcal {L} _ {\\mathrm {N C L}} (u, v), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 825, + 903, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ iterates over all detected objects and $(u, v)$ spans all pixels in the image I. $\\mathcal{L}_{\\mathrm{CFL}}$ , $\\mathcal{L}_{\\mathrm{NCL}}$ , and $\\mathcal{M}$ are defined in Equ. 7, Equ. 11, and Equ. 13. $\\lambda = 0.7$ by default.", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7277", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a07bb1b7be0223dcbdc4c28b069718a00e4894bcc40447b0299b8db123e0af59.jpg", + "table_caption": [ + "Table 1. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoFlex as the base model. We highlight the best and second results with **bold** and **underline** respectively." + ], + "table_footnote": [], + "table_body": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.000.000.0011.020.005.0712.947.7014.460.002.135.65
• TENTICLR'215.3111.096.402.2225.492.1423.8828.9635.4037.0724.6722.5930.63
• EATAICML'225.4412.124.672.7525.662.4524.7728.9935.6736.9524.3322.4734.10
• DeYOICLR'245.7812.524.523.0126.052.9824.9129.4035.1937.5923.7523.8234.33
• MonoTTAECCV'245.9313.344.053.3528.103.2125.8629.1036.4337.1825.9025.0133.89
• OursThis paper7.3015.409.364.3430.236.8929.0929.7638.3837.7229.3525.8834.97
Pedestrian Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.100.000.004.090.000.511.671.772.090.000.320.95
• TENTICLR'210.712.070.861.3612.700.816.728.3311.8912.797.426.459.78
• EATAICML'220.982.150.761.4412.230.806.778.6711.6013.627.536.4710.33
• DeYOICLR'241.032.250.911.6411.850.806.639.0911.7713.997.596.3910.57
• MonoTTAECCV'241.772.880.341.7812.380.827.039.0212.3113.117.757.1011.08
• OursThis paper1.893.081.541.8613.531.757.689.0912.6613.997.817.2711.27
Cyclist Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.200.000.000.240.002.142.331.724.410.000.000.00
• TENTICLR'210.060.140.040.044.550.936.638.2311.9415.167.721.852.81
• EATAICML'220.050.150.030.024.661.106.737.5813.7714.997.322.032.82
• DeYOICLR'240.060.190.030.034.911.086.486.9113.9414.677.571.792.82
• MonoTTAECCV'240.050.120.010.024.801.256.758.2413.3114.957.552.112.88
• OursThis paper0.110.220.070.106.002.006.898.4113.5815.947.942.132.92
", + "bbox": [ + 132, + 111, + 856, + 483 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ede0e78f33da7e66c03ca6801efb13773ab69451b3b0a9a75ebbb75d6f2a5e5b.jpg", + "table_caption": [ + "Table 2. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoGround as the base model. Due to the space limit, the complete results of three categories are provided in Tab. 7 of Appendix. B." + ], + "table_footnote": [], + "table_body": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoGroundCVPR'220.000.000.000.0011.630.291.956.593.1419.250.004.663.743.94
• TENTICLR'216.8214.818.214.8828.382.6523.9228.0833.0636.7020.2230.6333.2720.90
• EATAICML'227.1215.268.815.0929.082.5224.1828.0333.4336.7821.6130.5033.4221.22
• DeYOICLR'247.3515.729.385.7430.012.9925.0328.5534.3237.3123.4130.9934.1621.92
• MonoTTAECCV'247.8816.7310.355.9731.193.0625.2428.9934.8537.8225.0031.6134.7922.57
• OursThis paper9.7218.8812.747.2433.025.2428.5030.7337.2739.4028.3433.2237.2424.73
", + "bbox": [ + 132, + 527, + 856, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 657, + 223, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 686, + 281, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison, we follow the identical evaluation pipeline with prior work [24], including baseline models, training recipes, and evaluation protocols.", + "bbox": [ + 89, + 714, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on the KITTI [11] and nuScenes [5] datasets. For KITTI, we use the KITTI-C version, which includes 13 distinct data corruption types [14] and five severity levels per type. Results represent the average performance across three difficulty levels, i.e. Easy, Moderate, and Hard. Note that we also provide more results of different severity levels of KITTI in Appendix. B.1. For nuScenes, we adopt the front-view images and construct the Daytime, Night, Sunny, and Rainy scenarios via their scene", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "descriptions following [29]. Based on these scenes, we define 4 real-world adaptation tasks, i.e., Daytime $\\leftrightarrow$ Night and Sunny $\\leftrightarrow$ Rainy. Following the MonoTTA, we transfer the nuScenes dataset into the KITTI format. More details are provided in Appendix. D.3.", + "bbox": [ + 511, + 659, + 906, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared Methods. Based on two representative base models MonoFlex [55] and MonoGround [40], we compare our DUO with several state-of-the-art methods: EATA [36] identifies reliable samples during entropy training. DeYO [22] leverages probability variations under augmentations as an additional cue to enhance entropy optimization. MonoTTA [24] boosts probabilities for high-score classes while applying negative learning to low-score classes.", + "bbox": [ + 511, + 741, + 908, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement our method and other baselines in PyTorch [38]. We employ the Stochastic", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7278", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92d4d99dce17536d135c4261b2f714ad3b09faac07128ddcaec1884fc6375f52.jpg", + "table_caption": [ + "Table 3. Comparison with baselines on Daytime $\\leftrightarrow$ Night and Sunny $\\leftrightarrow$ Rainy of nuScenes dataset, regarding $\\mathrm{AP}_{\\mathrm{3D|R_{40}}}$ ." + ], + "table_footnote": [], + "table_body": "
MethodMonoFlexMonoGround
D → NN → DAvg.D → NN → DAvg.
Source model1.532.752.146.971.094.03
• TENT3.333.453.398.361.665.01
• DeYO4.724.874.7911.011.406.21
• MonoTTA6.923.685.3013.611.297.45
• Ours9.055.417.2315.701.918.81
MethodS → RR → SAvg.S → RR → SAvg.
Source model6.8610.918.898.037.447.73
• TENT8.5311.6110.079.948.719.33
• DeYO9.3312.0410.6810.549.109.82
• MonoTTA9.4712.5511.0110.898.909.90
• Ours11.5413.2112.3812.889.5411.21
", + "bbox": [ + 94, + 107, + 480, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gradient Descent (SGD) optimizer with the same learning rate as MonoTTA, a momentum of 0.9, and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\\lambda$ , $\\alpha$ , $\\gamma$ are assigned default values of 0.7, 4, and 2, respectively.", + "bbox": [ + 89, + 311, + 482, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Protocols. We report experimental results using the Average Precision for 3D bounding boxes, denoted as $\\mathrm{AP}_{3D|R_{40}}$ [43]. Results represent the mean values across three difficulty levels, with Intersection over Union thresholds set to 0.5 for Cars and 0.25 for both Ped. and Cyc.", + "bbox": [ + 89, + 372, + 483, + 448 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 455, + 230, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Corruption Shifts. We first compare our DUO with previous methods on the KITTI-C dataset at severity level 5. Due to space constraints, detailed comparisons with other severity levels are provided in the Appendix. B. The results, reported in Tab. 1 & 2, reveal several observations: 1) Under test-time distribution shifts, pretrained detectors experience significant performance degradation across all categories. 2) Existing TTA methods partially mitigate the adverse effects of distribution shifts in M3OD, but their performance remains suboptimal as they cannot address the dual uncertainty. 3) Our method consistently outperforms compared approaches across all categories and base models. Notably, the proposed DUO achieves the best or comparable performance under 13 types of corruptions, with performance gains of $+2.1$ and $+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}$ in the Car category for two different models. These observations validate the crucial role of our adaptation framework in addressing dual uncertainty in M3OD models and enhancing robustness against distribution shifts.", + "bbox": [ + 89, + 477, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on Real-World Scenario. We further evaluate different methods on four real-world scenarios, as shown in Tab. 3. The experimental results yield the following observations: 1) Under real-world shifts, the pre-trained model still suffers severe performance degradation. 2) For Sunny $\\leftrightarrow$ Rainy adaptation tasks on MonoFlex, existing methods achieve only marginal improvements, whereas our method boosts performance by $+4.7\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}$ . 3) Our DUO brings significant performance improvement on two base models,", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Effects of components in our method. We conduct ablation studies of the conjugate focal loss $\\mathcal{L}_{\\mathrm{CFL}}$ , normal consistency loss $\\mathcal{L}_{\\mathrm{NCL}}$ , and semantic guidance $\\mathcal{M}$ on the KITTI-C validation set. \"Src.\" denotes the source model without adaptation.", + "bbox": [ + 511, + 78, + 903, + 133 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f21d287336aaa078cbbcecf86bde33cceab0b1d23e8398e474300efa9345ee0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Src.\\( \\mathcal{L}_{\\text{CFL}} \\)\\( \\mathcal{L}_{\\text{NCL}} \\)\\( \\mathcal{M} \\)MonoFlexMonoGround
CarPedes.Cyc.Avg.CarPedes.Cyc.Avg.
✓4.540.880.832.083.941.790.522.08
✓20.986.604.3210.6322.898.862.5111.42
✓12.384.632.826.6115.696.401.918.00
✓✓22.176.994.7211.2924.299.272.7812.11
✓✓16.496.234.879.2019.687.982.9410.20
✓✓✓22.977.195.1011.7524.739.623.0212.46
", + "bbox": [ + 516, + 142, + 903, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/560cb5af9fc923d3f9f857e2a68622a8232d8f97d108d6f112b883616fa8dd74.jpg", + "image_caption": [ + "Figure 4. Effects of components in our method. $(a)$ shows the distribution of detection scores during adaptation with conjugate focal loss. $(b)$ records the variation of depth uncertainty in different heads during adaptation with normal field constraints." + ], + "image_footnote": [], + "bbox": [ + 516, + 243, + 709, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a80e21ac4a14bca26c2854d9923aa98753908fe1511cf6606ed338ca526bcf77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 243, + 898, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "maintaining the best performance across all four scenarios, further demonstrating its effectiveness and superiority.", + "bbox": [ + 511, + 444, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 484, + 666, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, without loss of generality, we conduct ablation studies on Monoflex under Gaussian shift in KITTI-C for the sake of brevity. Focusing on two pivotal components of DUO, Conjugate Focal Loss and Normal Field Constraint, we perform extensive experiments to analyze their independent roles and complementary effects, gaining insights into key factors contributing to their effectiveness. Effectiveness of Components. We investigate the impact of individual components by comparing the full method with variations that omit key parts. As shown in Tab. 4, incorporating either the conjugate focal loss or the normal field constraint significantly enhances detection performance, yielding average gains of $+8.9$ and $+7.6\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}$ , respectively. Notably, using the normal consistency loss $(\\mathcal{L}_{\\mathrm{NCL}})$ alone results in unstable, marginal improvements; its effectiveness depends on being combined with $\\mathcal{M}$ , highlighting the necessity of semantic guidance for effective geometric uncertainty reduction. These findings underscore the effectiveness of each component.", + "bbox": [ + 511, + 507, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, a comparison of score distributions in Fig. 4(a) and Fig. 2(c) shows that the conjugate focal loss consistently boosts detection scores, ensuring effective training for challenging low-score objects. Similarly, Fig. 4(b) versus Fig. 2(d) demonstrates that our normal field constraint consistently reduces geometric uncertainty across all heads, preventing the model collapse observed in baseline meth", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7279", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/15faa6deb1adce398932b6e14a93d4cc9a2fe9bb4a0ed1a66ef6defac327c036.jpg", + "image_caption": [ + "(a) Source Model", + "(c) Normal Field Constraint" + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 285, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d031504bb751695f8c862babdca7915fa42a7c436f5497a698c98bd478741c3.jpg", + "image_caption": [ + "(b) Conjugate Optimization", + "(d) Complete DUO" + ], + "image_footnote": [], + "bbox": [ + 295, + 104, + 480, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/330347b61f0aaf811b7912c5b4a22c15c148ee822785e4b924000a1f430f7292.jpg", + "image_caption": [ + "Figure 5. Impacts of different components on dual uncertainty. Colored lines in three-dimensional space show the cumulative average of dual uncertainty during the adaptation process. Green and blue lines in two-dimensional planes record geometric uncertainty and semantic uncertainty, respectively." + ], + "image_footnote": [], + "bbox": [ + 96, + 238, + 284, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/af2a37be2135ef91f1dad218063f98ad50ad6225973849c75edccfc3d570d2dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 238, + 480, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ods. Therefore, these innovations effectively overcome the limitations of existing methods identified in Sec. 4, leading to more robust and reliable detection.", + "bbox": [ + 89, + 460, + 482, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complementary Effects. As shown in Tab. 4, our complete method, which integrates all components, consistently achieves the best detection performance, demonstrating the compatibility of different components. To further investigate the complementary effects of our dual-branch design, we visualize the dual-uncertainty optimization process. Specifically, comparing Fig. 5(a)&(b), we observe that applying the conjugate focal loss significantly reduces semantic uncertainty, while interestingly, geometric uncertainty also shows a modest decline. A similar phenomenon is evident when employing the normal field constraint in Fig. 5(c). This synchronous behavior suggests an intrinsic interdependence between the two types of uncertainty, validating the potential of joint optimization.", + "bbox": [ + 89, + 511, + 482, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, by combining two innovations in our DUO framework, it achieves the fastest and most pronounced decrease in two uncertainties compared to individual components, as shown in Fig. 5(d). These observations further validate that our dual-branch architecture creates a complementary loop for dual-uncertainty optimization, effectively harnessing their complementary effects.", + "bbox": [ + 89, + 728, + 482, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness and Efficiency. To validate the robustness and efficiency of our method, we extend our analysis by examining the sensitivity of key hyper-parameters and comparing running times in Appendix C.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e54163e90545ec7c04503048da046bbc1d4dfad942d61bbe3e0e091266d580c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f967bf6dfef5de4ccfab6c80aac429d164e916bc3273fbed12d4b911beac353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 166, + 903, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15f60dd631e50948e0cadc2ec0e24dbf1d6fb02ce53cf6b56440acb9ccf89e1f.jpg", + "image_caption": [ + "Figure 6. Qualitative examples on KITTI-C. In each row, we provide the front view (left) and the bird's-eye view (right) visualizations. Red represents the ground truth of boxes, while Green represents the predictions. We circle some objects to highlight differences in predictions. The first row corresponds to the source model, the second to MonoTTA, and the third to our method." + ], + "image_footnote": [], + "bbox": [ + 516, + 244, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 513, + 429, + 694, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the qualitative results shown in Fig. 6, our DUO framework produces predictions that are more precisely aligned with the ground-truth annotations. Compared with the latest SOTA method, DUO not only significantly improves the accuracy of 3D location estimation but also reduces missed detections for challenging distant or small-scale objects, which are typically prone to large geometric errors. These visualizations further confirm that our dual uncertainty optimization effectively adapts the source model, achieving precise and reliable 3D perception that closely matches the ground truth.", + "bbox": [ + 511, + 450, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 633, + 782, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a synergistic dual-branch framework designed to address semantic-geometric uncertainty inherent in 3D vision systems under test-time domain shifts. Our approach derives a novel conjugate loss to offer an adaptive, label-free weighting mechanism for balanced training on semantic uncertainty. Meanwhile, it incorporates a normal consistency constraint to reduce uncertainty from inconsistent geometric representation. Our extensive experiments demonstrate that the proposed dual-branch optimization creates a complementary loop, consistently improving performance across diverse domain shifts. In future work, we tend to expand our framework beyond M3OD to cover a broader range of 3D vision tasks. We hope that our work will deepen the understanding of uncertainty-aware model adaptation while providing transferable insights for related fields, such as unsupervised learning.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7280", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), in part by the National Natural Science Foundation of China under Grant 62088102, in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation.", + "bbox": [ + 89, + 114, + 483, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 266, + 187, + 281 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 2019. 1", + "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79:151-175, 2010. 1", + "[3] Dimitri Bertsekas, Angela Nedic, and Asuman Ozdaglar. Convex analysis and optimization. Athena Scientific, 2003. 2, 4", + "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 18", + "[6] Joya Chen, Qi Wu, Dong Liu, and Tong Xu. Foreground-background imbalance problem in deep object detectors: A review. In 2020 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), pages 285-290. IEEE, 2020. 3", + "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 1", + "[8] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops, pages 1000-1001, 2020. 2", + "[9] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 201-208. JMLR Workshop and Conference Proceedings, 2010. 4" + ], + "bbox": [ + 99, + 291, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test-time adaptation for online vision-and-language navigation. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2, 6", + "[12] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204-6218, 2022. 4", + "[13] Gustavo A Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, and Christian Desrosiers. Clust3: Information invariant test-time training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6136-6145, 2023. 2", + "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 6, 18", + "[15] Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In Proceedings of the European Conference on Computer Vision, pages 613-630. Springer, 2022. 5", + "[16] Zixuan Hu, Yichun Hu, and Ling-Yu Duan. Seva: Leveraging single-step ensemble of vicinal augmentations for test-time adaptation. arXiv preprint arXiv:2505.04087, 2025. 2", + "[17] Zixuan Hu, Yichun Hu, Xiaotong Li, Shixiang Tang, and Ling-Yu Duan. Beyond entropy: Region confidence proxy for wild test-time adaptation. arXiv preprint arXiv:2505.20704, 2025. 1, 2", + "[18] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2", + "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1", + "[20] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5", + "[21] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162-14171, 2024. 2", + "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 6, 17" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7281", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1", + "[24] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 6, 17", + "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3, 4", + "[26] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2", + "[27] Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chiman Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218-19227, 2024. 2", + "[28] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2", + "[29] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 18", + "[30] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021.", + "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7941-7952, 2023. 1", + "[32] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 2, 3, 17", + "[33] Xinzhu Ma, Wanli Ouyang, Andrea Simonelli, and Elisa Ricci. 3d object detection from images for autonomous driving: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[34] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda. Power-law escape rate of sgd. In International Conference on Machine Learning, pages 15959-15975. PMLR, 2022. 4", + "[36] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 17", + "[37] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3", + "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 6", + "[39] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3227-3237, 2020. 1, 5", + "[40] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 2, 6, 17", + "[41] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2", + "[42] Prabhu Teja S and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 1", + "[43] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1991-1999, 2019. 7", + "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2", + "[45] Hugo Touchette. Legendre-fenchel transforms in a nutshell. URL http://www.maths.qmul.ac.uk/ht/archive/lfh2.pdf, page 25, 2005. 2, 4", + "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2020. 1, 3, 17", + "[47] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 454-463, 2021. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7282", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 2, 3", + "[49] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 1", + "[50] Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9092-9101, 2021. 1", + "[51] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23231-23241, 2024. 2", + "[52] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerf-like representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 1", + "[53] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 2, 17", + "[54] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 17", + "[55] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 2, 6, 17", + "[56] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION. In The Eleventh International Conference on Learning Representations, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 710 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7283", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_model.json b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f36751602e77e449b7442739a6206c2d8a47ea5 --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.13, + 0.872, + 0.175 + ], + "angle": 0, + "content": "Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.808, + 0.24 + ], + "angle": 0, + "content": "Zixuan Hu\\(^{1}\\) Dongxiao Li\\(^{1}\\) Xinzhu Ma\\(^{3}\\) Shixiang Tang\\(^{3}\\) Xiaotong Li\\(^{1}\\) Wenhan Yang\\(^{2}\\) Ling-Yu Duan\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.241, + 0.713, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Peking University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.258, + 0.838, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Peng Cheng Laboratory, Shenzhen, China, \\(^{3}\\)The Chinese University of Hong Kong, Hongkong, China." + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.278, + 0.612, + 0.293 + ], + "angle": 0, + "content": "{hzxuan, lingyu}@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.358, + 0.485, + 0.797 + ], + "angle": 0, + "content": "Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types. The source code is available at https://github.com/hzcar/DUO." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.824, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Monocular 3D object detection (M3OD) [1, 33, 34] serves as a fundamental perception task, enabling agents to un" + }, + { + "type": "image_caption", + "bbox": [ + 0.587, + 0.328, + 0.832, + 0.341 + ], + "angle": 0, + "content": "(a) Overview of Monocular 3D Object Detection" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.342, + 0.901, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.429, + 0.902, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Figure 1. (a) Illustration of M3OD. \\((b, c)\\) Under test-time shifts, increased semantic and geometric uncertainties lead to a degradation of semantic perception and spatial localization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.598, + 0.907, + 0.734 + ], + "angle": 0, + "content": "derstand 3D scenes directly from 2D images, as shown in Fig. 1(a). Due to its low cost and simple hardware configuration, M3OD has attracted widespread attention, leading to the development of numerous detectors [7, 31, 52]. However, when facing adverse weather conditions or sensor failures in real-world deployments, well-trained models often suffer severe performance degradation under data shifts [2, 14, 49]. Therefore, it is crucial to deal with the OOD generalization problem for M3OD." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To address the distribution shifting issue with minimal overhead, Test-Time Adaptation (TTA) has emerged as a critical paradigm, enabling source models to adapt to target distributions via online updates [17, 19, 23]. The dominating strategy in this context involves minimizing prediction entropy, thereby reducing the uncertainty of the model on shifted data [17, 46]. Despite the promising results, existing TTA approaches largely overlook the dual uncertainty inherent in 3D detection—semantic uncertainty (related to class predictions) [42, 50] and geometric uncertainty (related to spatial location) [30, 39], which is a significant dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7273" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.382, + 0.106 + ], + "angle": 0, + "content": "ference compared to conventional 2D tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.109, + 0.482, + 0.306 + ], + "angle": 0, + "content": "To investigate these two types of uncertainty under test-time shifts, we analyze detection outcomes for objects subject to common real-world variations. As shown in Fig. 1 (b)&(c), our empirical study reveals that both uncertainties increase markedly with data shifts, creating compounded error accumulation in 3D detection. Moreover, we find that existing uncertainty optimization techniques exhibit significant limitations: 1) Low-score object neglect. Entropy minimization fails to provide effective supervision for challenging objects with low detection scores, resulting in inevitable omissions. 2) Spatial perception collapse. Direct minimization of depth uncertainty can cause model collapse, compromising the perception capacity of spatial attributes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.61 + ], + "angle": 0, + "content": "In this work, to overcome the above limitations, we propose Dual Uncertainty Optimization (DUO), the first TTA framework for joint semantic-geometric uncertainty minimization. Specifically, through the lens of convex optimization theory [3, 45], we present an Legendre-Fenchel structure of the focal loss [25] and reconstruct the semantic uncertainty minimization as a dual optimization problem. Building on this foundation, we apply higher-order approximation analysis to derive a novel Conjugate Focal Loss. This loss breaks the label-dependence barrier in the original objective and introduces a dynamic weighting mechanism for balanced training, effectively reducing the omissions of objects with low scores. In parallel, we introduce a normal field constraint that enforces local consistency of surface normals in regions with high semantic confidence. This spatial coherence clarifies geometric cues, thereby reducing geometric uncertainty. Together, this dual-branch design creates a complementary loop where semantic confident regions bootstrap geometric feature learning, while enhanced spatial perception guides semantic refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.734 + ], + "angle": 0, + "content": "We evaluate the effectiveness and generalizability of our method through experiments on the KITTI dataset [11] with 13 corruption shift types, achieving state-of-the-art results with average improvements of \\(+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}\\) in the Car category. Furthermore, we showcase its superior performance in addressing real-world shift scenarios (daytime \\(\\leftrightarrow\\) night, sunny \\(\\leftrightarrow\\) rainy) of nuScenes dataset [5], yielding an average gain of \\(+18\\%\\) compared to existing methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Contributions: 1) To the best of our knowledge, we pioneer dual uncertainty optimization in M3OD by establishing the first TTA framework that jointly minimizes semantic and geometric uncertainties, addressing a critical reliability gap in real-world deployments. 2) Through the lens of convex optimization theory, we derive a novel Conjugate Focal Loss that enables label-agnostic uncertainty weighting and balanced learning for low-score objects. This approach is inherently compatible with the source phase, requiring no additional hyperparameter tuning. 3) We introduce a normal field constraint that enforces the stability of geometric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "representation with semantic guidance, resolving ambiguous spatial predictions. 4) We analyze and verify that our dual-branch design creates a complementary loop of two types of uncertainty optimization, resulting in significantly improved performance over existing TTA methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.174, + 0.656, + 0.19 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.196, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Monocular 3D Object Detection (M3OD) aims to perceive 3D objects from a single 2D image. Existing methods in M3OD can be broadly categorized based on their use of extra data sources, such as CAD models [28], dense depth maps [8, 47], or LiDAR [18, 41]. In this paper, we focus exclusively on approaches that utilize only monocular images, due to their computational efficiency and lower deployment costs. Previous studies, such as MonoDLE [32] and PGD [48], have identified depth estimation as a critical bottleneck in M3OD. To address this, many works leverage multiple geometric cues to integrate diverse depth predictions. For example, MonoFlex [55] integrates depth prediction by combining direct regression with multi-keypoint estimation; MonoGround [40] incorporates the ground plane as prior information; and MonoCD [53] exploits the complementary properties of multi-head estimation. In this paper, we investigate how to enhance the detection performance of such detectors under test-time shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Test-Time Adaptation (TTA) aims to enhance model performance on out-of-distribution samples during inference. Depending on whether the source training process is modified, existing TTA methods can be mainly divided into two groups: Test-Time Training [13, 27, 44] and Fully Test-Time Adaptation [4, 16, 56]. In this paper, we focus on Fully TTA, which adapts models without source data. Prior studies have demonstrated that reducing prediction uncertainty is an effective strategy for improving model generalization in various tasks [10, 21, 26, 51]. These works have developed various strategies to model and optimize uncertainty. For instance, SAR [37] uses entropy as a measure of classification uncertainty along with sharpness-aware optimization. DeYO [22] incorporates entropy with a disentangled factor for uncertainty modeling. ReCAP [17] models regional uncertainty, enforcing implicit data scaling for uncertainty optimization. MonoTTA [24] optimizes positive and negative class uncertainties separately. However, unlike conventional 2D tasks, M3OD inherently exhibits both semantic and geometric uncertainties, which remain largely unexplored and lack an effective training paradigm. In this work, we explore the joint optimization of this dual uncertainty, enhancing the robustness of M3OD models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.64, + 0.847 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Task Definition. M3OD aims to predict the 3D and semantic attributes of objects from a single RGB image. Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), the goal is to generate accurate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7274" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.295, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.501, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.701, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.09, + 0.905, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.22, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Semantic uncertainty (prediction entropy) and geometric uncertainty (average depth uncertainty) of the M3OD model under domain shifts. (a) shows variations in dual-uncertainty as the shift level increases. (b) records detection scores of predicted objects during inference with the source model. (c) records scores during the entropy minimization process. (d) investigates the depth uncertainty of different heads during the optimization with uncertainty regression loss. Experiments are conducted on the Gauss shift type of KITTI-C." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.291, + 0.482, + 0.427 + ], + "angle": 0, + "content": "3D bounding boxes \\(\\{\\mathcal{B}_i\\}_{i=1}^N\\) and semantic labels \\(\\{\\mathcal{C}_i\\}_{i=1}^N\\) for objects present in the scene, where \\(N\\) denotes the number of objects. Each bounding box is typically parameterized as \\(\\mathcal{B}_i = (\\mathbf{P}_i, \\mathbf{D}_i, \\mathbf{O}_i)\\), where \\(\\mathbf{P}_i \\in \\mathbb{R}^3\\) denotes the 3D center position, \\(\\mathbf{D}_i \\in \\mathbb{R}^3\\) represents the shape dimensions, and \\(\\mathbf{O}_i \\in [-\\pi, \\pi]\\) encodes the orientation. The multi-task nature of M3OD, which involves simultaneous estimation of both geometric and semantic attributes, poses significant challenges for achieving precise and coherent predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.427, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Meta-framework. As shown in Fig. 3(a), M3OD models typically employ a backbone network connected with multiple branches to predict various properties, e.g., score heatmaps, depth maps, etc. Since depth estimation is widely recognized as a key bottleneck [32, 48], many approaches employ a multi-head depth estimator to improve prediction accuracy. This estimator comprises a direct regression depth head and multiple geometric depth heads, each providing an individual depth prediction along with an associated uncertainty value. These outputs are then fused via an uncertainty-weighted average to produce the final depth prediction. We utilize the average uncertainty across all heads as our depth uncertainty metric. A detailed explanation is provided in the Appendix. D.2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.64, + 0.483, + 0.746 + ], + "angle": 0, + "content": "TTA Setting. TTA addresses the challenge of distribution shifts by enabling a pre-trained model to adapt to the target distribution during inference, without the need for labeled data. Unlike traditional training on fixed datasets, TTA operates in an online manner, where the model \\( h_{\\theta} \\) with parameter set \\( \\theta \\) produces detection outputs while concurrently updating its parameters based on incoming test data:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.751, + 0.483, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathcal {B} _ {t} \\right\\}, \\left\\{\\mathcal {C} _ {t} \\right\\} \\leftarrow h _ {\\theta} \\left(\\mathrm {I} _ {t}\\right), \\theta \\leftarrow \\theta - \\nabla \\mathcal {L} _ {t t a} \\left(\\mathrm {I} _ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.807 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}_t\\) denotes the incoming test data and \\(\\mathcal{L}_{tta}\\) denotes the loss function used for self-training during adaptation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.815, + 0.472, + 0.833 + ], + "angle": 0, + "content": "4. Uncovering Dual-Uncertainty under Shifts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Since TENT [46] revealed the positive correlation between prediction uncertainty and generalization error under distribution shifts, numerous TTA methods have emerged to optimize uncertainty metrics. While previous works have devel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.905, + 0.367 + ], + "angle": 0, + "content": "oped diverse measures for conventional 2D tasks [22, 37], the compound semantic and geometric uncertainties in 3D perception remains critically unexplored in the TTA field. In this work, we investigate both types of uncertainty within the context of M3OD to understand their distinct roles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.906, + 0.503 + ], + "angle": 0, + "content": "Specifically, to quantify the variation of uncertainty under distribution shifts, we track two metrics: semantic prediction entropy and geometric depth uncertainty (average uncertainty of multi-head depth estimator). As shown in Fig. 2(a), both metrics demonstrate a consistent upward trend as distribution shifts intensify, indicating that more severe shifts lead to higher model uncertainty. Furthermore, we analyze the independent optimization effects of two uncertainties and make the following key observations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.639 + ], + "angle": 0, + "content": "Observation 1: Conventional entropy minimization exacerbates imbalance distribution of detection scores. Different from classification scenarios, object detection suffers from extreme foreground-background imbalance, which hinders the effective optimization of hard positive objects [6, 25]. As shown in Fig. 2(b)&(c), entropy minimization yields a marginal gain for low-score objects while significantly boosting high-score predictions, further exacerbating the imbalance and leading to omissions of low-score objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Observation 2: Minimizing depth uncertainty directly causes the model collapse of the multi-head depth estimator. To minimize depth uncertainty, we apply the uncertainty regression loss across multiple depth heads (detailed in the Appendix. D.2). As shown in Fig. 2(d), the regression head which lacks any geometric constraints, exhibits a significantly faster decline of uncertainty compared to keypoint heads. This rapid convergence reduces the multi-head system to a single deterministic head, undermining the model's ability to perform robust spatial understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.649, + 0.817 + ], + "angle": 0, + "content": "5. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Based on the above observations, we propose a TTA method. Dual Uncertainty Optimization (DUO) for M3OD, which mainly leverages two novel designs, i.e., conjugate optimization framework and normal consistency constraint, to compatibly optimize semantic and geometric uncertainties." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7275" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.391, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.089, + 0.902, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Overview of our DUO method. (a) Meta-framework of the base models, where a backbone network connects to multi-branch predictors for estimating various object properties. (b) Through our theoretical conjugate optimization framework, we derive a conjugate focal loss \\(\\mathcal{L}_{\\mathrm{CFL}}\\) that can adaptively adjust training weights across all classes without the ground-truth labels, while also identifying low-semantic-uncertainty regions for geometric constraints (Sec. 5.1). (c) DUO employs the efficient Sobel operator to convert the depth map to a normal field and design a normal consistency constraint \\(\\mathcal{L}_{\\mathrm{NCL}}\\) to enhance geometric consistency in selected regions (Sec. 5.2)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.376, + 0.483, + 0.392 + ], + "angle": 0, + "content": "5.1. Semantic Uncertainty Conjugate Optimization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.483, + 0.594 + ], + "angle": 0, + "content": "To address the imbalance problem, it is crucial to devote more attention to challenging objects with high uncertainty. A straightforward approach is to employ the focal loss [25] which increases the weight of predictions with low probability. However, focal loss cannot assign an appropriate weight without the ground-truth label and fails to adapt effectively in unsupervised settings. To provide a more flexible weighting solution, we leverage convex optimization theory [3, 45] (a classical analytical tool) to explore the focal loss via a Legendre-Fenchel structure. Building on this foundation, we perform a higher-order approximation analysis to derive a novel loss that can dynamically adjust the weighting without relying on labels, as shown in Fig. 3(b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Legendre-Fenchel Structure. Given the source model \\( h_\\theta \\) pre-trained with the focal loss, we denote \\( h \\triangleq h_\\theta(x) \\) and its corresponding loss function can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.65, + 0.483, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F L}} (x, y) = - \\alpha (1 - p) ^ {\\gamma} \\cdot y \\log p, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.754 + ], + "angle": 0, + "content": "where \\( p \\triangleq \\text{softmax}(h) \\) is the normalized probability over the classes, \\( x \\) is a detected object, and \\( y \\) is the one-hot coding of the ground-truth label. \\( \\alpha \\) and \\( \\gamma \\) are hyperparameters controlling the weighting of uncertain predictions. Motivated by [12], we reformulate it as the following structure:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.762, + 0.369, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F L}} = f (h) - y ^ {\\top} g (h),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.778, + 0.482, + 0.8 + ], + "angle": 0, + "content": "\\[\nf (h) = \\alpha \\log s, g (h) = \\alpha h + \\alpha ((1 - p) ^ {\\gamma} - 1) \\log p, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\( s \\triangleq e^{h_1} + \\ldots + e^{h_c} \\) is the sum over the exponential outputs of the model and \\( c \\) is the number of classes. Refer to the Appendix. A for missing proofs of this subsection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Problem Reconstruction. Under this structure, the optimization problem can be regarded as finding an optimal representation \\( h \\) that minimizes the empirical loss. According" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.907, + 0.468 + ], + "angle": 0, + "content": "to the Legendre-Fenchel condition [45], the existence of the conjugate function \\( f^{*} \\) is equivalent to the invertibility of the function \\( g \\). This critical condition is formally established in our work (see Appendix. A), serving as a fundamental theoretical guarantee for our method. Therefore, the minimum value of the objective can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\min _ {h} \\left\\{f (h) - y ^ {\\top} g (h) \\right\\} = \\min _ {z = g (h)} \\left\\{f \\circ g ^ {- 1} (z) - y ^ {\\top} z \\right\\} = f ^ {*} (y). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.57 + ], + "angle": 0, + "content": "Building on the common assumption that the representation \\( h \\) pre-trained from the large source dataset is already close to a locally optimal solution \\( h_0 \\) [9, 35], we can convert the problem into the following relationships:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.575, + 0.905, + 0.595 + ], + "angle": 0, + "content": "\\[\nf \\circ g ^ {- 1} (z) - y ^ {\\top} z = f ^ {*} (y), \\quad \\nabla_ {z} \\left(f \\circ g ^ {- 1}\\right) = y. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.647 + ], + "angle": 0, + "content": "Conjugate Focal Loss. By applying the chain rule of gradient computation and higher-order approximation, we can obtain the following estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.653, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} y _ {0} \\triangleq \\frac {\\nabla_ {h} \\left(f \\circ g ^ {- 1}\\right)}{\\nabla_ {h} z} | _ {z = g (h)} = \\nabla_ {h} g (h) ^ {- 1} \\cdot \\nabla_ {h} f (h) \\tag {6} \\\\ \\approx (I + \\gamma (1 - \\log p) \\cdot p p ^ {\\top} - \\gamma \\log p \\cdot \\operatorname {d i a g} (p))) ^ {- 1} p, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.906, + 0.789 + ], + "angle": 0, + "content": "where \\(\\mathrm{diag}(\\cdot)\\) denotes the diagonal matrix and \\(I\\) denotes the identity matrix. This estimation \\(y_{0}\\) ensures that the loss aligns closely with the conjugate function. Ultimately, we substitute it into Equ. 5, yielding an unsupervised approximation of the conjugate function:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.795, + 0.905, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C F L}} (x) = - \\alpha (1 - p) ^ {\\gamma} (I + \\gamma (1 - \\log p) \\cdot p ^ {\\top} p \\tag {7} \\\\ - \\gamma \\log p \\cdot \\operatorname {d i a g} (p)) ^ {- 1} p \\log p. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In contrast to the focal loss in Equ. 2, our derived Conjugate Focal Loss (CFL) offers a dynamic adjustment of uncertainty weighting without relying on ground-truth labels and it confers the following advantages:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7276" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Static vs. Dynamic Adjustment. The vanilla focal loss uses a fixed weighting term for the ground-truth class to focus on high-uncertainty predictions. In contrast, our CFL not only incorporates \\((1 - p)^{\\gamma}\\) to address the imbalance training issue but also dynamically adjusts the weighting across all classes based on the term \\((I + \\gamma (1 - \\log p)pp^{\\top} - \\gamma \\log p\\cdot \\mathrm{diag}(p))^{-1}\\), which encodes inter-class prediction relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Ground-Truth Independence. While focal loss requires labels to compute the weight for the loss term, CFL operates solely based on the prediction probability, eliminating the need for labeled data during TTA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.482, + 0.364 + ], + "angle": 0, + "content": "Compatible Hyperparameters with Source Phase. Our theoretical analysis suggests that hyperparameters \\(\\alpha, \\gamma\\) should remain consistent with their values used in focal loss during the source training. This compatibility provides a practical advantage for TTA scenarios, eliminating the need for extensive hyperparameter tuning. The validity of this consistent setting is empirically verified in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.369, + 0.455, + 0.383 + ], + "angle": 0, + "content": "5.2. Semantic-Guided Normal Field Constraint" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.483, + 0.585 + ], + "angle": 0, + "content": "Despite progress in uncertainty modeling, current methods struggle to handle geometric uncertainty in TTA scenarios due to critical limitations: 1) Model Collapse: Direct optimization of model-predicted uncertainty leads to the degenerated predictor, as discussed in Sec. 4. 2) Substantial Overhead: Geometry-aware methods typically require additional data or offline training for uncertainty estimation, limiting their real-time applicability [15, 39]. To overcome these challenges, we propose an efficient normal field constraint derived from a single image, which enhances the geometric coherence of 3D predictions, thereby reducing the uncertainty stemming from the unstable geometric representation, as shown in Fig. 3(c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Normal Field. Given a depth map \\(D\\), we restore it to the original image resolution using bilinear interpolation, allowing for the alignment with the pixel grid. Then, we compute the spatial gradients of the depth map using efficient Sobel operators [20], which approximate the rate of change in the depth values along horizontal and vertical directions:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.686, + 0.482, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\nabla D _ {x} = \\mathbf {S} _ {x} * D, \\quad \\nabla D _ {y} = \\mathbf {S} _ {y} * D, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.788 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_x\\) and \\(\\mathbf{S}_y\\) denote horizontal and vertical Sobel kernels, respectively. These gradient maps capture the variation in depth across neighboring pixels. The surface normal field, which encodes the orientation of the surface at each pixel, is then derived from the gradients as:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.796, + 0.482, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\mathbf {N} (u, v) = \\frac {1}{\\sqrt {\\nabla D _ {x} ^ {2} + \\nabla D _ {y} ^ {2} + 1}} \\left[ \\begin{array}{c} - \\nabla D _ {x} \\\\ - \\nabla D _ {y} \\\\ 1 \\end{array} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.493, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{N}(u,v)\\) denotes the normal orientation at pixel \\((u,v)\\). Normal Consistency Loss. To quantify geometric uncertainty, we employ an edge-aware Normal Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "(NCL), that encourages smoothness in the local surface by penalizing inconsistencies between neighboring pixels:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.128, + 0.895, + 0.147 + ], + "angle": 0, + "content": "\\[\n\\psi_ {x} (u, v) = \\left\\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u + 1, v) - \\mathbf {N} (u - 1, v) \\right\\| _ {2} ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.149, + 0.905, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\psi_ {y} (u, v) = \\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u, v + 1) - \\mathbf {N} (u, v - 1) \\| _ {2} ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.18, + 0.905, + 0.226 + ], + "angle": 0, + "content": "where the smoothness terms \\(\\psi_x(u,v),\\psi_y(u,v)\\) enforce horizontal and vertical consistency in the surface normal field. The total normal consistency loss is then given by:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.232, + 0.905, + 0.248 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {N C L}} (u, v) = \\left(\\psi_ {x} (u, v) + \\psi_ {y} (u, v)\\right) \\cdot \\exp (- \\| \\nabla \\mathrm {I} (u, v) \\| _ {2}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.905, + 0.344 + ], + "angle": 0, + "content": "where the edge-aware weighting term \\(\\exp (-\\| \\nabla \\mathbf{I}(u,v)\\| _2)\\) \\(= \\exp (-\\sqrt{|\\mathbf{S}_x*\\mathbf{I}(u,v)|^2 + |\\mathbf{S}_y*\\mathbf{I}(u,v)|^2})\\) preserves discontinuities at boundaries while enforcing smoothness in homogeneous regions. NCL encourages the model to learn a spatially consistent normal field, thereby reducing uncertainty stemming from unstable 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.345, + 0.906, + 0.482 + ], + "angle": 0, + "content": "Semantic Guidance. To enforce geometric-semantic coherence, we generate masks by integrating 2D bounding boxes with semantic predictions, ensuring synchronized focus regions for dual-branch uncertainty optimization. Let \\(\\{\\mathcal{B}_i\\}_{i=1}^n\\) denote detected bounding boxes with scores \\(\\{s_i\\}_{i=1}^n\\) and semantic uncertainties \\(\\{U_i\\}_{i=1}^n\\) derived from the Conjugate Focal Loss (\\(\\mathcal{L}_{\\mathrm{CFL}}\\) in Eq. 7). To ensure reliable supervision, we select boxes with low semantic uncertainty using an exponentially moving average threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.492, + 0.905, + 0.523 + ], + "angle": 0, + "content": "\\[\nR = \\{i \\mid U _ {i} \\leq \\bar {U} \\}, \\bar {U} \\leftarrow \\beta \\cdot \\frac {\\sum_ {i = 1} ^ {n} U _ {i}}{n} + (1 - \\beta) \\cdot \\bar {U}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.529, + 0.905, + 0.56 + ], + "angle": 0, + "content": "where \\(\\beta \\in [0,1]\\) is a moving average factor (set to 0.1 in default). We then construct the region mask as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.566, + 0.905, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (u, v) = \\max _ {i \\in R} s _ {i} \\cdot \\mathbb {I} _ {\\text {i n s i d e}} (u, v \\mid \\mathcal {B} _ {i}), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.596, + 0.905, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}_{\\mathrm{inside}}(u,v\\mid \\mathcal{B}_i)\\) is an indicator function that returns 1 if pixel \\((u,v)\\) is inside \\(\\mathcal{B}_i\\), and 0 otherwise. This semantic-guided mask ensures that only low semantic-uncertainty regions contribute to the normal field constraint, enhancing the reliability of the geometric constraint." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.676, + 0.688, + 0.692 + ], + "angle": 0, + "content": "5.3. Overall Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.82 + ], + "angle": 0, + "content": "The overall objective integrates the conjugate focal loss with the normal consistency constraint into a unified framework, simultaneously addressing semantic and geometric uncertainties. This dual optimization enables a complementary feedback loop: low-uncertainty spatial location enhances the model's ability to perform precise semantic classification, while confident semantic predictions in turn improve spatial understanding. The procedure is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.827, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\Sigma_ {x \\in I} \\mathcal {L} _ {\\mathrm {C F L}} (x) + \\lambda \\Sigma_ {(u, v) \\in I} \\mathcal {M} (u, v) \\cdot \\mathcal {L} _ {\\mathrm {N C L}} (u, v), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\( x \\) iterates over all detected objects and \\( (u, v) \\) spans all pixels in the image I. \\( \\mathcal{L}_{\\mathrm{CFL}} \\), \\( \\mathcal{L}_{\\mathrm{NCL}} \\), and \\( \\mathcal{M} \\) are defined in Equ. 7, Equ. 11, and Equ. 13. \\( \\lambda = 0.7 \\) by default." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7277" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.077, + 0.908, + 0.107 + ], + "angle": 0, + "content": "Table 1. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoFlex as the base model. We highlight the best and second results with **bold** and **underline** respectively." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.112, + 0.857, + 0.484 + ], + "angle": 0, + "content": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.000.000.0011.020.005.0712.947.7014.460.002.135.65
• TENTICLR'215.3111.096.402.2225.492.1423.8828.9635.4037.0724.6722.5930.63
• EATAICML'225.4412.124.672.7525.662.4524.7728.9935.6736.9524.3322.4734.10
• DeYOICLR'245.7812.524.523.0126.052.9824.9129.4035.1937.5923.7523.8234.33
• MonoTTAECCV'245.9313.344.053.3528.103.2125.8629.1036.4337.1825.9025.0133.89
• OursThis paper7.3015.409.364.3430.236.8929.0929.7638.3837.7229.3525.8834.97
Pedestrian Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.100.000.004.090.000.511.671.772.090.000.320.95
• TENTICLR'210.712.070.861.3612.700.816.728.3311.8912.797.426.459.78
• EATAICML'220.982.150.761.4412.230.806.778.6711.6013.627.536.4710.33
• DeYOICLR'241.032.250.911.6411.850.806.639.0911.7713.997.596.3910.57
• MonoTTAECCV'241.772.880.341.7812.380.827.039.0212.3113.117.757.1011.08
• OursThis paper1.893.081.541.8613.531.757.689.0912.6613.997.817.2711.27
Cyclist Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.200.000.000.240.002.142.331.724.410.000.000.00
• TENTICLR'210.060.140.040.044.550.936.638.2311.9415.167.721.852.81
• EATAICML'220.050.150.030.024.661.106.737.5813.7714.997.322.032.82
• DeYOICLR'240.060.190.030.034.911.086.486.9113.9414.677.571.792.82
• MonoTTAECCV'240.050.120.010.024.801.256.758.2413.3114.957.552.112.88
• OursThis paper0.110.220.070.106.002.006.898.4113.5815.947.942.132.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.492, + 0.908, + 0.521 + ], + "angle": 0, + "content": "Table 2. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoGround as the base model. Due to the space limit, the complete results of three categories are provided in Tab. 7 of Appendix. B." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.528, + 0.857, + 0.655 + ], + "angle": 0, + "content": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoGroundCVPR'220.000.000.000.0011.630.291.956.593.1419.250.004.663.743.94
• TENTICLR'216.8214.818.214.8828.382.6523.9228.0833.0636.7020.2230.6333.2720.90
• EATAICML'227.1215.268.815.0929.082.5224.1828.0333.4336.7821.6130.5033.4221.22
• DeYOICLR'247.3515.729.385.7430.012.9925.0328.5534.3237.3123.4130.9934.1621.92
• MonoTTAECCV'247.8816.7310.355.9731.193.0625.2428.9934.8537.8225.0031.6134.7922.57
• OursThis paper9.7218.8812.747.2433.025.2428.5030.7337.2739.4028.3433.2237.2424.73
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.224, + 0.676 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.282, + 0.706 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.761 + ], + "angle": 0, + "content": "For a fair comparison, we follow the identical evaluation pipeline with prior work [24], including baseline models, training recipes, and evaluation protocols." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on the KITTI [11] and nuScenes [5] datasets. For KITTI, we use the KITTI-C version, which includes 13 distinct data corruption types [14] and five severity levels per type. Results represent the average performance across three difficulty levels, i.e. Easy, Moderate, and Hard. Note that we also provide more results of different severity levels of KITTI in Appendix. B.1. For nuScenes, we adopt the front-view images and construct the Daytime, Night, Sunny, and Rainy scenarios via their scene" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.737 + ], + "angle": 0, + "content": "descriptions following [29]. Based on these scenes, we define 4 real-world adaptation tasks, i.e., Daytime \\(\\leftrightarrow\\) Night and Sunny \\(\\leftrightarrow\\) Rainy. Following the MonoTTA, we transfer the nuScenes dataset into the KITTI format. More details are provided in Appendix. D.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.909, + 0.865 + ], + "angle": 0, + "content": "Compared Methods. Based on two representative base models MonoFlex [55] and MonoGround [40], we compare our DUO with several state-of-the-art methods: EATA [36] identifies reliable samples during entropy training. DeYO [22] leverages probability variations under augmentations as an additional cue to enhance entropy optimization. MonoTTA [24] boosts probabilities for high-score classes while applying negative learning to low-score classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. We implement our method and other baselines in PyTorch [38]. We employ the Stochastic" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7278" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.079, + 0.483, + 0.108 + ], + "angle": 0, + "content": "Table 3. Comparison with baselines on Daytime \\(\\leftrightarrow\\) Night and Sunny \\(\\leftrightarrow\\) Rainy of nuScenes dataset, regarding \\(\\mathrm{AP}_{\\mathrm{3D|R_{40}}}\\)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.482, + 0.287 + ], + "angle": 0, + "content": "
MethodMonoFlexMonoGround
D → NN → DAvg.D → NN → DAvg.
Source model1.532.752.146.971.094.03
• TENT3.333.453.398.361.665.01
• DeYO4.724.874.7911.011.406.21
• MonoTTA6.923.685.3013.611.297.45
• Ours9.055.417.2315.701.918.81
MethodS → RR → SAvg.S → RR → SAvg.
Source model6.8610.918.898.037.447.73
• TENT8.5311.6110.079.948.719.33
• DeYO9.3312.0410.6810.549.109.82
• MonoTTA9.4712.5511.0110.898.909.90
• Ours11.5413.2112.3812.889.5411.21
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.372 + ], + "angle": 0, + "content": "Gradient Descent (SGD) optimizer with the same learning rate as MonoTTA, a momentum of 0.9, and a batch size of 16 for KITTI, 4 for nuScenes. Parameters \\(\\lambda\\), \\(\\alpha\\), \\(\\gamma\\) are assigned default values of 0.7, 4, and 2, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.449 + ], + "angle": 0, + "content": "Evaluation Protocols. We report experimental results using the Average Precision for 3D bounding boxes, denoted as \\(\\mathrm{AP}_{3D|R_{40}}\\) [43]. Results represent the mean values across three difficulty levels, with Intersection over Union thresholds set to 0.5 for Cars and 0.25 for both Ped. and Cyc." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.231, + 0.471 + ], + "angle": 0, + "content": "6.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Evaluation on Corruption Shifts. We first compare our DUO with previous methods on the KITTI-C dataset at severity level 5. Due to space constraints, detailed comparisons with other severity levels are provided in the Appendix. B. The results, reported in Tab. 1 & 2, reveal several observations: 1) Under test-time distribution shifts, pretrained detectors experience significant performance degradation across all categories. 2) Existing TTA methods partially mitigate the adverse effects of distribution shifts in M3OD, but their performance remains suboptimal as they cannot address the dual uncertainty. 3) Our method consistently outperforms compared approaches across all categories and base models. Notably, the proposed DUO achieves the best or comparable performance under 13 types of corruptions, with performance gains of \\(+2.1\\) and \\(+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}\\) in the Car category for two different models. These observations validate the crucial role of our adaptation framework in addressing dual uncertainty in M3OD models and enhancing robustness against distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation on Real-World Scenario. We further evaluate different methods on four real-world scenarios, as shown in Tab. 3. The experimental results yield the following observations: 1) Under real-world shifts, the pre-trained model still suffers severe performance degradation. 2) For Sunny \\(\\leftrightarrow\\) Rainy adaptation tasks on MonoFlex, existing methods achieve only marginal improvements, whereas our method boosts performance by \\(+4.7\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}\\). 3) Our DUO brings significant performance improvement on two base models," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.079, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Table 4. Effects of components in our method. We conduct ablation studies of the conjugate focal loss \\(\\mathcal{L}_{\\mathrm{CFL}}\\), normal consistency loss \\(\\mathcal{L}_{\\mathrm{NCL}}\\), and semantic guidance \\(\\mathcal{M}\\) on the KITTI-C validation set. \"Src.\" denotes the source model without adaptation." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.143, + 0.905, + 0.23 + ], + "angle": 0, + "content": "
Src.\\( \\mathcal{L}_{\\text{CFL}} \\)\\( \\mathcal{L}_{\\text{NCL}} \\)\\( \\mathcal{M} \\)MonoFlexMonoGround
CarPedes.Cyc.Avg.CarPedes.Cyc.Avg.
✓4.540.880.832.083.941.790.522.08
✓20.986.604.3210.6322.898.862.5111.42
✓12.384.632.826.6115.696.401.918.00
✓✓22.176.994.7211.2924.299.272.7812.11
✓✓16.496.234.879.2019.687.982.9410.20
✓✓✓22.977.195.1011.7524.739.623.0212.46
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.244, + 0.71, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.244, + 0.899, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.373, + 0.905, + 0.429 + ], + "angle": 0, + "content": "Figure 4. Effects of components in our method. \\((a)\\) shows the distribution of detection scores during adaptation with conjugate focal loss. \\((b)\\) records the variation of depth uncertainty in different heads during adaptation with normal field constraints." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.445, + 0.905, + 0.476 + ], + "angle": 0, + "content": "maintaining the best performance across all four scenarios, further demonstrating its effectiveness and superiority." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.667, + 0.502 + ], + "angle": 0, + "content": "6.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.794 + ], + "angle": 0, + "content": "In this section, without loss of generality, we conduct ablation studies on Monoflex under Gaussian shift in KITTI-C for the sake of brevity. Focusing on two pivotal components of DUO, Conjugate Focal Loss and Normal Field Constraint, we perform extensive experiments to analyze their independent roles and complementary effects, gaining insights into key factors contributing to their effectiveness. Effectiveness of Components. We investigate the impact of individual components by comparing the full method with variations that omit key parts. As shown in Tab. 4, incorporating either the conjugate focal loss or the normal field constraint significantly enhances detection performance, yielding average gains of \\(+8.9\\) and \\(+7.6\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}\\), respectively. Notably, using the normal consistency loss \\((\\mathcal{L}_{\\mathrm{NCL}})\\) alone results in unstable, marginal improvements; its effectiveness depends on being combined with \\(\\mathcal{M}\\), highlighting the necessity of semantic guidance for effective geometric uncertainty reduction. These findings underscore the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, a comparison of score distributions in Fig. 4(a) and Fig. 2(c) shows that the conjugate focal loss consistently boosts detection scores, ensuring effective training for challenging low-score objects. Similarly, Fig. 4(b) versus Fig. 2(d) demonstrates that our normal field constraint consistently reduces geometric uncertainty across all heads, preventing the model collapse observed in baseline meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7279" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.09, + 0.225, + 0.102 + ], + "angle": 0, + "content": "(a) Source Model" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.105, + 0.287, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.09, + 0.449, + 0.103 + ], + "angle": 0, + "content": "(b) Conjugate Optimization" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.105, + 0.482, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.225, + 0.25, + 0.236 + ], + "angle": 0, + "content": "(c) Normal Field Constraint" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.239, + 0.285, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.225, + 0.423, + 0.236 + ], + "angle": 0, + "content": "(d) Complete DUO" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.239, + 0.482, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.437 + ], + "angle": 0, + "content": "Figure 5. Impacts of different components on dual uncertainty. Colored lines in three-dimensional space show the cumulative average of dual uncertainty during the adaptation process. Green and blue lines in two-dimensional planes record geometric uncertainty and semantic uncertainty, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.506 + ], + "angle": 0, + "content": "ods. Therefore, these innovations effectively overcome the limitations of existing methods identified in Sec. 4, leading to more robust and reliable detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Complementary Effects. As shown in Tab. 4, our complete method, which integrates all components, consistently achieves the best detection performance, demonstrating the compatibility of different components. To further investigate the complementary effects of our dual-branch design, we visualize the dual-uncertainty optimization process. Specifically, comparing Fig. 5(a)&(b), we observe that applying the conjugate focal loss significantly reduces semantic uncertainty, while interestingly, geometric uncertainty also shows a modest decline. A similar phenomenon is evident when employing the normal field constraint in Fig. 5(c). This synchronous behavior suggests an intrinsic interdependence between the two types of uncertainty, validating the potential of joint optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Moreover, by combining two innovations in our DUO framework, it achieves the fastest and most pronounced decrease in two uncertainties compared to individual components, as shown in Fig. 5(d). These observations further validate that our dual-branch architecture creates a complementary loop for dual-uncertainty optimization, effectively harnessing their complementary effects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Robustness and Efficiency. To validate the robustness and efficiency of our method, we extend our analysis by examining the sensitivity of key hyper-parameters and comparing running times in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.245, + 0.905, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Figure 6. Qualitative examples on KITTI-C. In each row, we provide the front view (left) and the bird's-eye view (right) visualizations. Red represents the ground truth of boxes, while Green represents the predictions. We circle some objects to highlight differences in predictions. The first row corresponds to the source model, the second to MonoTTA, and the third to our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.43, + 0.696, + 0.446 + ], + "angle": 0, + "content": "6.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.907, + 0.619 + ], + "angle": 0, + "content": "Based on the qualitative results shown in Fig. 6, our DUO framework produces predictions that are more precisely aligned with the ground-truth annotations. Compared with the latest SOTA method, DUO not only significantly improves the accuracy of 3D location estimation but also reduces missed detections for challenging distant or small-scale objects, which are typically prone to large geometric errors. These visualizations further confirm that our dual uncertainty optimization effectively adapts the source model, achieving precise and reliable 3D perception that closely matches the ground truth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.783, + 0.649 + ], + "angle": 0, + "content": "7. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a synergistic dual-branch framework designed to address semantic-geometric uncertainty inherent in 3D vision systems under test-time domain shifts. Our approach derives a novel conjugate loss to offer an adaptive, label-free weighting mechanism for balanced training on semantic uncertainty. Meanwhile, it incorporates a normal consistency constraint to reduce uncertainty from inconsistent geometric representation. Our extensive experiments demonstrate that the proposed dual-branch optimization creates a complementary loop, consistently improving performance across diverse domain shifts. In future work, we tend to expand our framework beyond M3OD to cover a broader range of 3D vision tasks. We hope that our work will deepen the understanding of uncertainty-aware model adaptation while providing transferable insights for related fields, such as unsupervised learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7280" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.253 + ], + "angle": 0, + "content": "This work was supported by the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), in part by the National Natural Science Foundation of China under Grant 62088102, in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.267, + 0.188, + 0.282 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.292, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[1] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.363, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79:151-175, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[3] Dimitri Bertsekas, Angela Nedic, and Asuman Ozdaglar. Convex analysis and optimization. Athena Scientific, 2003. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.464, + 0.484, + 0.518 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.521, + 0.484, + 0.604 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[6] Joya Chen, Qi Wu, Dong Liu, and Tong Xu. Foreground-background imbalance problem in deep object detectors: A review. In 2020 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), pages 285-290. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[8] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops, pages 1000-1001, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[9] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 201-208. JMLR Workshop and Conference Proceedings, 2010. 4" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.292, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[10] Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test-time adaptation for online vision-and-language navigation. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.152, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.223, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[12] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204-6218, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.281, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[13] Gustavo A Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, and Christian Desrosiers. Clust3: Information invariant test-time training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6136-6145, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.904, + 0.407 + ], + "angle": 0, + "content": "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[15] Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In Proceedings of the European Conference on Computer Vision, pages 613-630. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[16] Zixuan Hu, Yichun Hu, and Ling-Yu Duan. Seva: Leveraging single-step ensemble of vicinal augmentations for test-time adaptation. arXiv preprint arXiv:2505.04087, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.514, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[17] Zixuan Hu, Yichun Hu, Xiaotong Li, Shixiang Tang, and Ling-Yu Duan. Beyond entropy: Region confidence proxy for wild test-time adaptation. arXiv preprint arXiv:2505.20704, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[18] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.644, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[20] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[21] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162-14171, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 6, 17" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7281" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[23] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[24] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[26] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[27] Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chiman Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218-19227, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[28] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[29] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[30] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7941-7952, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[32] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 2, 3, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[33] Xinzhu Ma, Wanli Ouyang, Andrea Simonelli, and Elisa Ricci. 3d object detection from images for autonomous driving: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[34] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[35] Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda. Power-law escape rate of sgd. In International Conference on Machine Learning, pages 15959-15975. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[37] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[39] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3227-3237, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[40] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 2, 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[41] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[42] Prabhu Teja S and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.907, + 0.649 + ], + "angle": 0, + "content": "[43] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1991-1999, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[45] Hugo Touchette. Legendre-fenchel transforms in a nutshell. URL http://www.maths.qmul.ac.uk/ht/archive/lfh2.pdf, page 25, 2005. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2020. 1, 3, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 454-463, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7282" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[48] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[49] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[50] Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9092-9101, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[51] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23231-23241, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.402, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[52] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerf-like representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.472, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[53] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 2, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[54] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[55] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 2, 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[56] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7283" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_origin.pdf b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6b8c712b668c490e09214a489ddce2628416daaf --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/c498df0d-2910-4380-bea1-d1d32a2718e5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:554729f6d7b4d064613904b172cadcab8b33e3eaa50de343d128fc64b44efdae +size 2785108 diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/full.md b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cc2e97dbfda0c13bd8bb918c15643d82e5d4e42b --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/full.md @@ -0,0 +1,363 @@ +# Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts + +Zixuan Hu $^{1}$ Dongxiao Li $^{1}$ Xinzhu Ma $^{3}$ Shixiang Tang $^{3}$ Xiaotong Li $^{1}$ Wenhan Yang $^{2}$ Ling-Yu Duan $^{1,2*}$ + +$^{1}$ School of Computer Science, Peking University, Beijing, China + +$^{2}$ Peng Cheng Laboratory, Shenzhen, China, $^{3}$ The Chinese University of Hong Kong, Hongkong, China. + +{hzxuan, lingyu}@pku.edu.cn + +# Abstract + +Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types. The source code is available at https://github.com/hzcar/DUO. + +# 1. Introduction + +Monocular 3D object detection (M3OD) [1, 33, 34] serves as a fundamental perception task, enabling agents to un + +![](images/51296b4f1e88d12619b34438f118bb4f0030761c26ff53f3f9edfc3cf01536e5.jpg) +(a) Overview of Monocular 3D Object Detection + +![](images/ed334cae71a4b29534ce7241d1691e88c3a62fb38ecd6712fdab21d5efcd5503.jpg) +Figure 1. (a) Illustration of M3OD. $(b, c)$ Under test-time shifts, increased semantic and geometric uncertainties lead to a degradation of semantic perception and spatial localization capabilities. + +derstand 3D scenes directly from 2D images, as shown in Fig. 1(a). Due to its low cost and simple hardware configuration, M3OD has attracted widespread attention, leading to the development of numerous detectors [7, 31, 52]. However, when facing adverse weather conditions or sensor failures in real-world deployments, well-trained models often suffer severe performance degradation under data shifts [2, 14, 49]. Therefore, it is crucial to deal with the OOD generalization problem for M3OD. + +To address the distribution shifting issue with minimal overhead, Test-Time Adaptation (TTA) has emerged as a critical paradigm, enabling source models to adapt to target distributions via online updates [17, 19, 23]. The dominating strategy in this context involves minimizing prediction entropy, thereby reducing the uncertainty of the model on shifted data [17, 46]. Despite the promising results, existing TTA approaches largely overlook the dual uncertainty inherent in 3D detection—semantic uncertainty (related to class predictions) [42, 50] and geometric uncertainty (related to spatial location) [30, 39], which is a significant dif + +ference compared to conventional 2D tasks. + +To investigate these two types of uncertainty under test-time shifts, we analyze detection outcomes for objects subject to common real-world variations. As shown in Fig. 1 (b)&(c), our empirical study reveals that both uncertainties increase markedly with data shifts, creating compounded error accumulation in 3D detection. Moreover, we find that existing uncertainty optimization techniques exhibit significant limitations: 1) Low-score object neglect. Entropy minimization fails to provide effective supervision for challenging objects with low detection scores, resulting in inevitable omissions. 2) Spatial perception collapse. Direct minimization of depth uncertainty can cause model collapse, compromising the perception capacity of spatial attributes. + +In this work, to overcome the above limitations, we propose Dual Uncertainty Optimization (DUO), the first TTA framework for joint semantic-geometric uncertainty minimization. Specifically, through the lens of convex optimization theory [3, 45], we present an Legendre-Fenchel structure of the focal loss [25] and reconstruct the semantic uncertainty minimization as a dual optimization problem. Building on this foundation, we apply higher-order approximation analysis to derive a novel Conjugate Focal Loss. This loss breaks the label-dependence barrier in the original objective and introduces a dynamic weighting mechanism for balanced training, effectively reducing the omissions of objects with low scores. In parallel, we introduce a normal field constraint that enforces local consistency of surface normals in regions with high semantic confidence. This spatial coherence clarifies geometric cues, thereby reducing geometric uncertainty. Together, this dual-branch design creates a complementary loop where semantic confident regions bootstrap geometric feature learning, while enhanced spatial perception guides semantic refinement. + +We evaluate the effectiveness and generalizability of our method through experiments on the KITTI dataset [11] with 13 corruption shift types, achieving state-of-the-art results with average improvements of $+2.2\mathrm{AP}_{3\mathrm{D}|\mathrm{R}_{40}}$ in the Car category. Furthermore, we showcase its superior performance in addressing real-world shift scenarios (daytime $\leftrightarrow$ night, sunny $\leftrightarrow$ rainy) of nuScenes dataset [5], yielding an average gain of $+18\%$ compared to existing methods. + +Contributions: 1) To the best of our knowledge, we pioneer dual uncertainty optimization in M3OD by establishing the first TTA framework that jointly minimizes semantic and geometric uncertainties, addressing a critical reliability gap in real-world deployments. 2) Through the lens of convex optimization theory, we derive a novel Conjugate Focal Loss that enables label-agnostic uncertainty weighting and balanced learning for low-score objects. This approach is inherently compatible with the source phase, requiring no additional hyperparameter tuning. 3) We introduce a normal field constraint that enforces the stability of geometric + +representation with semantic guidance, resolving ambiguous spatial predictions. 4) We analyze and verify that our dual-branch design creates a complementary loop of two types of uncertainty optimization, resulting in significantly improved performance over existing TTA methods. + +# 2. Related Work + +Monocular 3D Object Detection (M3OD) aims to perceive 3D objects from a single 2D image. Existing methods in M3OD can be broadly categorized based on their use of extra data sources, such as CAD models [28], dense depth maps [8, 47], or LiDAR [18, 41]. In this paper, we focus exclusively on approaches that utilize only monocular images, due to their computational efficiency and lower deployment costs. Previous studies, such as MonoDLE [32] and PGD [48], have identified depth estimation as a critical bottleneck in M3OD. To address this, many works leverage multiple geometric cues to integrate diverse depth predictions. For example, MonoFlex [55] integrates depth prediction by combining direct regression with multi-keypoint estimation; MonoGround [40] incorporates the ground plane as prior information; and MonoCD [53] exploits the complementary properties of multi-head estimation. In this paper, we investigate how to enhance the detection performance of such detectors under test-time shifts. + +Test-Time Adaptation (TTA) aims to enhance model performance on out-of-distribution samples during inference. Depending on whether the source training process is modified, existing TTA methods can be mainly divided into two groups: Test-Time Training [13, 27, 44] and Fully Test-Time Adaptation [4, 16, 56]. In this paper, we focus on Fully TTA, which adapts models without source data. Prior studies have demonstrated that reducing prediction uncertainty is an effective strategy for improving model generalization in various tasks [10, 21, 26, 51]. These works have developed various strategies to model and optimize uncertainty. For instance, SAR [37] uses entropy as a measure of classification uncertainty along with sharpness-aware optimization. DeYO [22] incorporates entropy with a disentangled factor for uncertainty modeling. ReCAP [17] models regional uncertainty, enforcing implicit data scaling for uncertainty optimization. MonoTTA [24] optimizes positive and negative class uncertainties separately. However, unlike conventional 2D tasks, M3OD inherently exhibits both semantic and geometric uncertainties, which remain largely unexplored and lack an effective training paradigm. In this work, we explore the joint optimization of this dual uncertainty, enhancing the robustness of M3OD models. + +# 3. Preliminary + +Task Definition. M3OD aims to predict the 3D and semantic attributes of objects from a single RGB image. Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , the goal is to generate accurate + +![](images/4fd75fea96986fd1e56e5e2a0b6efa144c0b2e6ee797b45f98b7d0d0cd5b93a3.jpg) +Figure 2. Semantic uncertainty (prediction entropy) and geometric uncertainty (average depth uncertainty) of the M3OD model under domain shifts. (a) shows variations in dual-uncertainty as the shift level increases. (b) records detection scores of predicted objects during inference with the source model. (c) records scores during the entropy minimization process. (d) investigates the depth uncertainty of different heads during the optimization with uncertainty regression loss. Experiments are conducted on the Gauss shift type of KITTI-C. + +![](images/d6f655a59813be432ce02fd130fedc4ddb96960b34278be8b618ed379a7132ff.jpg) + +![](images/cb2fe78abf57b64a5c0ec37fb380299220e70f35dd56280acf667131d80f2017.jpg) + +![](images/7768d4e328fbde4aa0503cc0c40d6c862df1e126c9327ca51ed07e6a1b4030ff.jpg) + +3D bounding boxes $\{\mathcal{B}_i\}_{i=1}^N$ and semantic labels $\{\mathcal{C}_i\}_{i=1}^N$ for objects present in the scene, where $N$ denotes the number of objects. Each bounding box is typically parameterized as $\mathcal{B}_i = (\mathbf{P}_i, \mathbf{D}_i, \mathbf{O}_i)$ , where $\mathbf{P}_i \in \mathbb{R}^3$ denotes the 3D center position, $\mathbf{D}_i \in \mathbb{R}^3$ represents the shape dimensions, and $\mathbf{O}_i \in [-\pi, \pi]$ encodes the orientation. The multi-task nature of M3OD, which involves simultaneous estimation of both geometric and semantic attributes, poses significant challenges for achieving precise and coherent predictions. + +Meta-framework. As shown in Fig. 3(a), M3OD models typically employ a backbone network connected with multiple branches to predict various properties, e.g., score heatmaps, depth maps, etc. Since depth estimation is widely recognized as a key bottleneck [32, 48], many approaches employ a multi-head depth estimator to improve prediction accuracy. This estimator comprises a direct regression depth head and multiple geometric depth heads, each providing an individual depth prediction along with an associated uncertainty value. These outputs are then fused via an uncertainty-weighted average to produce the final depth prediction. We utilize the average uncertainty across all heads as our depth uncertainty metric. A detailed explanation is provided in the Appendix. D.2. + +TTA Setting. TTA addresses the challenge of distribution shifts by enabling a pre-trained model to adapt to the target distribution during inference, without the need for labeled data. Unlike traditional training on fixed datasets, TTA operates in an online manner, where the model $h_{\theta}$ with parameter set $\theta$ produces detection outputs while concurrently updating its parameters based on incoming test data: + +$$ +\left\{\mathcal {B} _ {t} \right\}, \left\{\mathcal {C} _ {t} \right\} \leftarrow h _ {\theta} \left(\mathrm {I} _ {t}\right), \theta \leftarrow \theta - \nabla \mathcal {L} _ {t t a} \left(\mathrm {I} _ {t}\right), \tag {1} +$$ + +where $\mathbf{I}_t$ denotes the incoming test data and $\mathcal{L}_{tta}$ denotes the loss function used for self-training during adaptation. + +# 4. Uncovering Dual-Uncertainty under Shifts + +Since TENT [46] revealed the positive correlation between prediction uncertainty and generalization error under distribution shifts, numerous TTA methods have emerged to optimize uncertainty metrics. While previous works have devel + +oped diverse measures for conventional 2D tasks [22, 37], the compound semantic and geometric uncertainties in 3D perception remains critically unexplored in the TTA field. In this work, we investigate both types of uncertainty within the context of M3OD to understand their distinct roles. + +Specifically, to quantify the variation of uncertainty under distribution shifts, we track two metrics: semantic prediction entropy and geometric depth uncertainty (average uncertainty of multi-head depth estimator). As shown in Fig. 2(a), both metrics demonstrate a consistent upward trend as distribution shifts intensify, indicating that more severe shifts lead to higher model uncertainty. Furthermore, we analyze the independent optimization effects of two uncertainties and make the following key observations: + +Observation 1: Conventional entropy minimization exacerbates imbalance distribution of detection scores. Different from classification scenarios, object detection suffers from extreme foreground-background imbalance, which hinders the effective optimization of hard positive objects [6, 25]. As shown in Fig. 2(b)&(c), entropy minimization yields a marginal gain for low-score objects while significantly boosting high-score predictions, further exacerbating the imbalance and leading to omissions of low-score objects. + +Observation 2: Minimizing depth uncertainty directly causes the model collapse of the multi-head depth estimator. To minimize depth uncertainty, we apply the uncertainty regression loss across multiple depth heads (detailed in the Appendix. D.2). As shown in Fig. 2(d), the regression head which lacks any geometric constraints, exhibits a significantly faster decline of uncertainty compared to keypoint heads. This rapid convergence reduces the multi-head system to a single deterministic head, undermining the model's ability to perform robust spatial understanding. + +# 5. Methodology + +Based on the above observations, we propose a TTA method. Dual Uncertainty Optimization (DUO) for M3OD, which mainly leverages two novel designs, i.e., conjugate optimization framework and normal consistency constraint, to compatibly optimize semantic and geometric uncertainties. + +![](images/82b02825d62808bc0a954dfd6d7b61f0b5b229de88224d908cc2d95fe40d88e5.jpg) +Figure 3. Overview of our DUO method. (a) Meta-framework of the base models, where a backbone network connects to multi-branch predictors for estimating various object properties. (b) Through our theoretical conjugate optimization framework, we derive a conjugate focal loss $\mathcal{L}_{\mathrm{CFL}}$ that can adaptively adjust training weights across all classes without the ground-truth labels, while also identifying low-semantic-uncertainty regions for geometric constraints (Sec. 5.1). (c) DUO employs the efficient Sobel operator to convert the depth map to a normal field and design a normal consistency constraint $\mathcal{L}_{\mathrm{NCL}}$ to enhance geometric consistency in selected regions (Sec. 5.2). + +![](images/4885a4ae85addbce8880b2ccb215519ee8e39802a24770e6b3c5ec0df8eb0ffa.jpg) + +# 5.1. Semantic Uncertainty Conjugate Optimization + +To address the imbalance problem, it is crucial to devote more attention to challenging objects with high uncertainty. A straightforward approach is to employ the focal loss [25] which increases the weight of predictions with low probability. However, focal loss cannot assign an appropriate weight without the ground-truth label and fails to adapt effectively in unsupervised settings. To provide a more flexible weighting solution, we leverage convex optimization theory [3, 45] (a classical analytical tool) to explore the focal loss via a Legendre-Fenchel structure. Building on this foundation, we perform a higher-order approximation analysis to derive a novel loss that can dynamically adjust the weighting without relying on labels, as shown in Fig. 3(b). + +Legendre-Fenchel Structure. Given the source model $h_\theta$ pre-trained with the focal loss, we denote $h \triangleq h_\theta(x)$ and its corresponding loss function can be formulated as: + +$$ +\mathcal {L} _ {\mathrm {F L}} (x, y) = - \alpha (1 - p) ^ {\gamma} \cdot y \log p, \tag {2} +$$ + +where $p \triangleq \text{softmax}(h)$ is the normalized probability over the classes, $x$ is a detected object, and $y$ is the one-hot coding of the ground-truth label. $\alpha$ and $\gamma$ are hyperparameters controlling the weighting of uncertain predictions. Motivated by [12], we reformulate it as the following structure: + +$$ +\mathcal {L} _ {\mathrm {F L}} = f (h) - y ^ {\top} g (h), +$$ + +$$ +f (h) = \alpha \log s, g (h) = \alpha h + \alpha ((1 - p) ^ {\gamma} - 1) \log p, \tag {3} +$$ + +where $s \triangleq e^{h_1} + \ldots + e^{h_c}$ is the sum over the exponential outputs of the model and $c$ is the number of classes. Refer to the Appendix. A for missing proofs of this subsection. + +Problem Reconstruction. Under this structure, the optimization problem can be regarded as finding an optimal representation $h$ that minimizes the empirical loss. According + +to the Legendre-Fenchel condition [45], the existence of the conjugate function $f^{*}$ is equivalent to the invertibility of the function $g$ . This critical condition is formally established in our work (see Appendix. A), serving as a fundamental theoretical guarantee for our method. Therefore, the minimum value of the objective can be expressed as follows: + +$$ +\min _ {h} \left\{f (h) - y ^ {\top} g (h) \right\} = \min _ {z = g (h)} \left\{f \circ g ^ {- 1} (z) - y ^ {\top} z \right\} = f ^ {*} (y). \tag {4} +$$ + +Building on the common assumption that the representation $h$ pre-trained from the large source dataset is already close to a locally optimal solution $h_0$ [9, 35], we can convert the problem into the following relationships: + +$$ +f \circ g ^ {- 1} (z) - y ^ {\top} z = f ^ {*} (y), \quad \nabla_ {z} \left(f \circ g ^ {- 1}\right) = y. \tag {5} +$$ + +Conjugate Focal Loss. By applying the chain rule of gradient computation and higher-order approximation, we can obtain the following estimation: + +$$ +\begin{array}{l} y _ {0} \triangleq \frac {\nabla_ {h} \left(f \circ g ^ {- 1}\right)}{\nabla_ {h} z} | _ {z = g (h)} = \nabla_ {h} g (h) ^ {- 1} \cdot \nabla_ {h} f (h) \tag {6} \\ \approx (I + \gamma (1 - \log p) \cdot p p ^ {\top} - \gamma \log p \cdot \operatorname {d i a g} (p))) ^ {- 1} p, \\ \end{array} +$$ + +where $\mathrm{diag}(\cdot)$ denotes the diagonal matrix and $I$ denotes the identity matrix. This estimation $y_{0}$ ensures that the loss aligns closely with the conjugate function. Ultimately, we substitute it into Equ. 5, yielding an unsupervised approximation of the conjugate function: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C F L}} (x) = - \alpha (1 - p) ^ {\gamma} (I + \gamma (1 - \log p) \cdot p ^ {\top} p \tag {7} \\ - \gamma \log p \cdot \operatorname {d i a g} (p)) ^ {- 1} p \log p. \\ \end{array} +$$ + +In contrast to the focal loss in Equ. 2, our derived Conjugate Focal Loss (CFL) offers a dynamic adjustment of uncertainty weighting without relying on ground-truth labels and it confers the following advantages: + +Static vs. Dynamic Adjustment. The vanilla focal loss uses a fixed weighting term for the ground-truth class to focus on high-uncertainty predictions. In contrast, our CFL not only incorporates $(1 - p)^{\gamma}$ to address the imbalance training issue but also dynamically adjusts the weighting across all classes based on the term $(I + \gamma (1 - \log p)pp^{\top} - \gamma \log p\cdot \mathrm{diag}(p))^{-1}$ , which encodes inter-class prediction relationships. + +Ground-Truth Independence. While focal loss requires labels to compute the weight for the loss term, CFL operates solely based on the prediction probability, eliminating the need for labeled data during TTA. + +Compatible Hyperparameters with Source Phase. Our theoretical analysis suggests that hyperparameters $\alpha, \gamma$ should remain consistent with their values used in focal loss during the source training. This compatibility provides a practical advantage for TTA scenarios, eliminating the need for extensive hyperparameter tuning. The validity of this consistent setting is empirically verified in Appendix C. + +# 5.2. Semantic-Guided Normal Field Constraint + +Despite progress in uncertainty modeling, current methods struggle to handle geometric uncertainty in TTA scenarios due to critical limitations: 1) Model Collapse: Direct optimization of model-predicted uncertainty leads to the degenerated predictor, as discussed in Sec. 4. 2) Substantial Overhead: Geometry-aware methods typically require additional data or offline training for uncertainty estimation, limiting their real-time applicability [15, 39]. To overcome these challenges, we propose an efficient normal field constraint derived from a single image, which enhances the geometric coherence of 3D predictions, thereby reducing the uncertainty stemming from the unstable geometric representation, as shown in Fig. 3(c). + +Normal Field. Given a depth map $D$ , we restore it to the original image resolution using bilinear interpolation, allowing for the alignment with the pixel grid. Then, we compute the spatial gradients of the depth map using efficient Sobel operators [20], which approximate the rate of change in the depth values along horizontal and vertical directions: + +$$ +\nabla D _ {x} = \mathbf {S} _ {x} * D, \quad \nabla D _ {y} = \mathbf {S} _ {y} * D, \tag {8} +$$ + +where $\mathbf{S}_x$ and $\mathbf{S}_y$ denote horizontal and vertical Sobel kernels, respectively. These gradient maps capture the variation in depth across neighboring pixels. The surface normal field, which encodes the orientation of the surface at each pixel, is then derived from the gradients as: + +$$ +\mathbf {N} (u, v) = \frac {1}{\sqrt {\nabla D _ {x} ^ {2} + \nabla D _ {y} ^ {2} + 1}} \left[ \begin{array}{c} - \nabla D _ {x} \\ - \nabla D _ {y} \\ 1 \end{array} \right], \tag {9} +$$ + +where $\mathbf{N}(u,v)$ denotes the normal orientation at pixel $(u,v)$ . Normal Consistency Loss. To quantify geometric uncertainty, we employ an edge-aware Normal Consistency Loss + +(NCL), that encourages smoothness in the local surface by penalizing inconsistencies between neighboring pixels: + +$$ +\psi_ {x} (u, v) = \left\| 2 \mathbf {N} (u, v) - \mathbf {N} (u + 1, v) - \mathbf {N} (u - 1, v) \right\| _ {2} ^ {2}, +$$ + +$$ +\psi_ {y} (u, v) = \| 2 \mathbf {N} (u, v) - \mathbf {N} (u, v + 1) - \mathbf {N} (u, v - 1) \| _ {2} ^ {2}, \tag {10} +$$ + +where the smoothness terms $\psi_x(u,v),\psi_y(u,v)$ enforce horizontal and vertical consistency in the surface normal field. The total normal consistency loss is then given by: + +$$ +\mathcal {L} _ {\mathrm {N C L}} (u, v) = \left(\psi_ {x} (u, v) + \psi_ {y} (u, v)\right) \cdot \exp (- \| \nabla \mathrm {I} (u, v) \| _ {2}), \tag {11} +$$ + +where the edge-aware weighting term $\exp (-\| \nabla \mathbf{I}(u,v)\| _2)$ $= \exp (-\sqrt{|\mathbf{S}_x*\mathbf{I}(u,v)|^2 + |\mathbf{S}_y*\mathbf{I}(u,v)|^2})$ preserves discontinuities at boundaries while enforcing smoothness in homogeneous regions. NCL encourages the model to learn a spatially consistent normal field, thereby reducing uncertainty stemming from unstable 3D representations. + +Semantic Guidance. To enforce geometric-semantic coherence, we generate masks by integrating 2D bounding boxes with semantic predictions, ensuring synchronized focus regions for dual-branch uncertainty optimization. Let $\{\mathcal{B}_i\}_{i=1}^n$ denote detected bounding boxes with scores $\{s_i\}_{i=1}^n$ and semantic uncertainties $\{U_i\}_{i=1}^n$ derived from the Conjugate Focal Loss ( $\mathcal{L}_{\mathrm{CFL}}$ in Eq. 7). To ensure reliable supervision, we select boxes with low semantic uncertainty using an exponentially moving average threshold: + +$$ +R = \{i \mid U _ {i} \leq \bar {U} \}, \bar {U} \leftarrow \beta \cdot \frac {\sum_ {i = 1} ^ {n} U _ {i}}{n} + (1 - \beta) \cdot \bar {U}, \tag {12} +$$ + +where $\beta \in [0,1]$ is a moving average factor (set to 0.1 in default). We then construct the region mask as follows: + +$$ +\mathcal {M} (u, v) = \max _ {i \in R} s _ {i} \cdot \mathbb {I} _ {\text {i n s i d e}} (u, v \mid \mathcal {B} _ {i}), \tag {13} +$$ + +where $\mathbb{I}_{\mathrm{inside}}(u,v\mid \mathcal{B}_i)$ is an indicator function that returns 1 if pixel $(u,v)$ is inside $\mathcal{B}_i$ , and 0 otherwise. This semantic-guided mask ensures that only low semantic-uncertainty regions contribute to the normal field constraint, enhancing the reliability of the geometric constraint. + +# 5.3. Overall Objective + +The overall objective integrates the conjugate focal loss with the normal consistency constraint into a unified framework, simultaneously addressing semantic and geometric uncertainties. This dual optimization enables a complementary feedback loop: low-uncertainty spatial location enhances the model's ability to perform precise semantic classification, while confident semantic predictions in turn improve spatial understanding. The procedure is as follows: + +$$ +\min _ {\theta} \Sigma_ {x \in I} \mathcal {L} _ {\mathrm {C F L}} (x) + \lambda \Sigma_ {(u, v) \in I} \mathcal {M} (u, v) \cdot \mathcal {L} _ {\mathrm {N C L}} (u, v), \tag {14} +$$ + +where $x$ iterates over all detected objects and $(u, v)$ spans all pixels in the image I. $\mathcal{L}_{\mathrm{CFL}}$ , $\mathcal{L}_{\mathrm{NCL}}$ , and $\mathcal{M}$ are defined in Equ. 7, Equ. 11, and Equ. 13. $\lambda = 0.7$ by default. + +Table 1. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoFlex as the base model. We highlight the best and second results with **bold** and **underline** respectively. + +
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.000.000.0011.020.005.0712.947.7014.460.002.135.65
• TENTICLR'215.3111.096.402.2225.492.1423.8828.9635.4037.0724.6722.5930.63
• EATAICML'225.4412.124.672.7525.662.4524.7728.9935.6736.9524.3322.4734.10
• DeYOICLR'245.7812.524.523.0126.052.9824.9129.4035.1937.5923.7523.8234.33
• MonoTTAECCV'245.9313.344.053.3528.103.2125.8629.1036.4337.1825.9025.0133.89
• OursThis paper7.3015.409.364.3430.236.8929.0929.7638.3837.7229.3525.8834.97
Pedestrian Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.100.000.004.090.000.511.671.772.090.000.320.95
• TENTICLR'210.712.070.861.3612.700.816.728.3311.8912.797.426.459.78
• EATAICML'220.982.150.761.4412.230.806.778.6711.6013.627.536.4710.33
• DeYOICLR'241.032.250.911.6411.850.806.639.0911.7713.997.596.3910.57
• MonoTTAECCV'241.772.880.341.7812.380.827.039.0212.3113.117.757.1011.08
• OursThis paper1.893.081.541.8613.531.757.689.0912.6613.997.817.2711.27
Cyclist Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.200.000.000.240.002.142.331.724.410.000.000.00
• TENTICLR'210.060.140.040.044.550.936.638.2311.9415.167.721.852.81
• EATAICML'220.050.150.030.024.661.106.737.5813.7714.997.322.032.82
• DeYOICLR'240.060.190.030.034.911.086.486.9113.9414.677.571.792.82
• MonoTTAECCV'240.050.120.010.024.801.256.758.2413.3114.957.552.112.88
• OursThis paper0.110.220.070.106.002.006.898.4113.5815.947.942.132.92
+ +Table 2. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoGround as the base model. Due to the space limit, the complete results of three categories are provided in Tab. 7 of Appendix. B. + +
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoGroundCVPR'220.000.000.000.0011.630.291.956.593.1419.250.004.663.743.94
• TENTICLR'216.8214.818.214.8828.382.6523.9228.0833.0636.7020.2230.6333.2720.90
• EATAICML'227.1215.268.815.0929.082.5224.1828.0333.4336.7821.6130.5033.4221.22
• DeYOICLR'247.3515.729.385.7430.012.9925.0328.5534.3237.3123.4130.9934.1621.92
• MonoTTAECCV'247.8816.7310.355.9731.193.0625.2428.9934.8537.8225.0031.6134.7922.57
• OursThis paper9.7218.8812.747.2433.025.2428.5030.7337.2739.4028.3433.2237.2424.73
+ +# 6. Experiments + +# 6.1. Experimental Setup + +For a fair comparison, we follow the identical evaluation pipeline with prior work [24], including baseline models, training recipes, and evaluation protocols. + +Datasets. We conduct experiments on the KITTI [11] and nuScenes [5] datasets. For KITTI, we use the KITTI-C version, which includes 13 distinct data corruption types [14] and five severity levels per type. Results represent the average performance across three difficulty levels, i.e. Easy, Moderate, and Hard. Note that we also provide more results of different severity levels of KITTI in Appendix. B.1. For nuScenes, we adopt the front-view images and construct the Daytime, Night, Sunny, and Rainy scenarios via their scene + +descriptions following [29]. Based on these scenes, we define 4 real-world adaptation tasks, i.e., Daytime $\leftrightarrow$ Night and Sunny $\leftrightarrow$ Rainy. Following the MonoTTA, we transfer the nuScenes dataset into the KITTI format. More details are provided in Appendix. D.3. + +Compared Methods. Based on two representative base models MonoFlex [55] and MonoGround [40], we compare our DUO with several state-of-the-art methods: EATA [36] identifies reliable samples during entropy training. DeYO [22] leverages probability variations under augmentations as an additional cue to enhance entropy optimization. MonoTTA [24] boosts probabilities for high-score classes while applying negative learning to low-score classes. + +Implementation Details. We implement our method and other baselines in PyTorch [38]. We employ the Stochastic + +Table 3. Comparison with baselines on Daytime $\leftrightarrow$ Night and Sunny $\leftrightarrow$ Rainy of nuScenes dataset, regarding $\mathrm{AP}_{\mathrm{3D|R_{40}}}$ . + +
MethodMonoFlexMonoGround
D → NN → DAvg.D → NN → DAvg.
Source model1.532.752.146.971.094.03
• TENT3.333.453.398.361.665.01
• DeYO4.724.874.7911.011.406.21
• MonoTTA6.923.685.3013.611.297.45
• Ours9.055.417.2315.701.918.81
MethodS → RR → SAvg.S → RR → SAvg.
Source model6.8610.918.898.037.447.73
• TENT8.5311.6110.079.948.719.33
• DeYO9.3312.0410.6810.549.109.82
• MonoTTA9.4712.5511.0110.898.909.90
• Ours11.5413.2112.3812.889.5411.21
+ +Gradient Descent (SGD) optimizer with the same learning rate as MonoTTA, a momentum of 0.9, and a batch size of 16 for KITTI, 4 for nuScenes. Parameters $\lambda$ , $\alpha$ , $\gamma$ are assigned default values of 0.7, 4, and 2, respectively. + +Evaluation Protocols. We report experimental results using the Average Precision for 3D bounding boxes, denoted as $\mathrm{AP}_{3D|R_{40}}$ [43]. Results represent the mean values across three difficulty levels, with Intersection over Union thresholds set to 0.5 for Cars and 0.25 for both Ped. and Cyc. + +# 6.2. Main Results + +Evaluation on Corruption Shifts. We first compare our DUO with previous methods on the KITTI-C dataset at severity level 5. Due to space constraints, detailed comparisons with other severity levels are provided in the Appendix. B. The results, reported in Tab. 1 & 2, reveal several observations: 1) Under test-time distribution shifts, pretrained detectors experience significant performance degradation across all categories. 2) Existing TTA methods partially mitigate the adverse effects of distribution shifts in M3OD, but their performance remains suboptimal as they cannot address the dual uncertainty. 3) Our method consistently outperforms compared approaches across all categories and base models. Notably, the proposed DUO achieves the best or comparable performance under 13 types of corruptions, with performance gains of $+2.1$ and $+2.2\mathrm{AP}_{3\mathrm{D}|\mathrm{R}_{40}}$ in the Car category for two different models. These observations validate the crucial role of our adaptation framework in addressing dual uncertainty in M3OD models and enhancing robustness against distribution shifts. + +Evaluation on Real-World Scenario. We further evaluate different methods on four real-world scenarios, as shown in Tab. 3. The experimental results yield the following observations: 1) Under real-world shifts, the pre-trained model still suffers severe performance degradation. 2) For Sunny $\leftrightarrow$ Rainy adaptation tasks on MonoFlex, existing methods achieve only marginal improvements, whereas our method boosts performance by $+4.7\mathrm{AP}_{3\mathrm{D}|\mathrm{R}_{40}}$ . 3) Our DUO brings significant performance improvement on two base models, + +Table 4. Effects of components in our method. We conduct ablation studies of the conjugate focal loss $\mathcal{L}_{\mathrm{CFL}}$ , normal consistency loss $\mathcal{L}_{\mathrm{NCL}}$ , and semantic guidance $\mathcal{M}$ on the KITTI-C validation set. "Src." denotes the source model without adaptation. + +
Src.\( \mathcal{L}_{\text{CFL}} \)\( \mathcal{L}_{\text{NCL}} \)\( \mathcal{M} \)MonoFlexMonoGround
CarPedes.Cyc.Avg.CarPedes.Cyc.Avg.
✓4.540.880.832.083.941.790.522.08
✓20.986.604.3210.6322.898.862.5111.42
✓12.384.632.826.6115.696.401.918.00
✓✓22.176.994.7211.2924.299.272.7812.11
✓✓16.496.234.879.2019.687.982.9410.20
✓✓✓22.977.195.1011.7524.739.623.0212.46
+ +![](images/560cb5af9fc923d3f9f857e2a68622a8232d8f97d108d6f112b883616fa8dd74.jpg) +Figure 4. Effects of components in our method. $(a)$ shows the distribution of detection scores during adaptation with conjugate focal loss. $(b)$ records the variation of depth uncertainty in different heads during adaptation with normal field constraints. + +![](images/a80e21ac4a14bca26c2854d9923aa98753908fe1511cf6606ed338ca526bcf77.jpg) + +maintaining the best performance across all four scenarios, further demonstrating its effectiveness and superiority. + +# 6.3. Ablation Study + +In this section, without loss of generality, we conduct ablation studies on Monoflex under Gaussian shift in KITTI-C for the sake of brevity. Focusing on two pivotal components of DUO, Conjugate Focal Loss and Normal Field Constraint, we perform extensive experiments to analyze their independent roles and complementary effects, gaining insights into key factors contributing to their effectiveness. Effectiveness of Components. We investigate the impact of individual components by comparing the full method with variations that omit key parts. As shown in Tab. 4, incorporating either the conjugate focal loss or the normal field constraint significantly enhances detection performance, yielding average gains of $+8.9$ and $+7.6\mathrm{AP}_{3\mathrm{D}|\mathrm{R}_{40}}$ , respectively. Notably, using the normal consistency loss $(\mathcal{L}_{\mathrm{NCL}})$ alone results in unstable, marginal improvements; its effectiveness depends on being combined with $\mathcal{M}$ , highlighting the necessity of semantic guidance for effective geometric uncertainty reduction. These findings underscore the effectiveness of each component. + +Furthermore, a comparison of score distributions in Fig. 4(a) and Fig. 2(c) shows that the conjugate focal loss consistently boosts detection scores, ensuring effective training for challenging low-score objects. Similarly, Fig. 4(b) versus Fig. 2(d) demonstrates that our normal field constraint consistently reduces geometric uncertainty across all heads, preventing the model collapse observed in baseline meth + +![](images/15faa6deb1adce398932b6e14a93d4cc9a2fe9bb4a0ed1a66ef6defac327c036.jpg) +(a) Source Model +(c) Normal Field Constraint + +![](images/7d031504bb751695f8c862babdca7915fa42a7c436f5497a698c98bd478741c3.jpg) +(b) Conjugate Optimization +(d) Complete DUO + +![](images/330347b61f0aaf811b7912c5b4a22c15c148ee822785e4b924000a1f430f7292.jpg) +Figure 5. Impacts of different components on dual uncertainty. Colored lines in three-dimensional space show the cumulative average of dual uncertainty during the adaptation process. Green and blue lines in two-dimensional planes record geometric uncertainty and semantic uncertainty, respectively. + +![](images/af2a37be2135ef91f1dad218063f98ad50ad6225973849c75edccfc3d570d2dd.jpg) + +ods. Therefore, these innovations effectively overcome the limitations of existing methods identified in Sec. 4, leading to more robust and reliable detection. + +Complementary Effects. As shown in Tab. 4, our complete method, which integrates all components, consistently achieves the best detection performance, demonstrating the compatibility of different components. To further investigate the complementary effects of our dual-branch design, we visualize the dual-uncertainty optimization process. Specifically, comparing Fig. 5(a)&(b), we observe that applying the conjugate focal loss significantly reduces semantic uncertainty, while interestingly, geometric uncertainty also shows a modest decline. A similar phenomenon is evident when employing the normal field constraint in Fig. 5(c). This synchronous behavior suggests an intrinsic interdependence between the two types of uncertainty, validating the potential of joint optimization. + +Moreover, by combining two innovations in our DUO framework, it achieves the fastest and most pronounced decrease in two uncertainties compared to individual components, as shown in Fig. 5(d). These observations further validate that our dual-branch architecture creates a complementary loop for dual-uncertainty optimization, effectively harnessing their complementary effects. + +Robustness and Efficiency. To validate the robustness and efficiency of our method, we extend our analysis by examining the sensitivity of key hyper-parameters and comparing running times in Appendix C. + +![](images/5e54163e90545ec7c04503048da046bbc1d4dfad942d61bbe3e0e091266d580c.jpg) + +![](images/9f967bf6dfef5de4ccfab6c80aac429d164e916bc3273fbed12d4b911beac353.jpg) + +![](images/15f60dd631e50948e0cadc2ec0e24dbf1d6fb02ce53cf6b56440acb9ccf89e1f.jpg) +Figure 6. Qualitative examples on KITTI-C. In each row, we provide the front view (left) and the bird's-eye view (right) visualizations. Red represents the ground truth of boxes, while Green represents the predictions. We circle some objects to highlight differences in predictions. The first row corresponds to the source model, the second to MonoTTA, and the third to our method. + +# 6.4. Qualitative Results + +Based on the qualitative results shown in Fig. 6, our DUO framework produces predictions that are more precisely aligned with the ground-truth annotations. Compared with the latest SOTA method, DUO not only significantly improves the accuracy of 3D location estimation but also reduces missed detections for challenging distant or small-scale objects, which are typically prone to large geometric errors. These visualizations further confirm that our dual uncertainty optimization effectively adapts the source model, achieving precise and reliable 3D perception that closely matches the ground truth. + +# 7. Conclusion and Future Work + +In this paper, we propose a synergistic dual-branch framework designed to address semantic-geometric uncertainty inherent in 3D vision systems under test-time domain shifts. Our approach derives a novel conjugate loss to offer an adaptive, label-free weighting mechanism for balanced training on semantic uncertainty. Meanwhile, it incorporates a normal consistency constraint to reduce uncertainty from inconsistent geometric representation. Our extensive experiments demonstrate that the proposed dual-branch optimization creates a complementary loop, consistently improving performance across diverse domain shifts. In future work, we tend to expand our framework beyond M3OD to cover a broader range of 3D vision tasks. We hope that our work will deepen the understanding of uncertainty-aware model adaptation while providing transferable insights for related fields, such as unsupervised learning. + +# Acknowledgements + +This work was supported by the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), in part by the National Natural Science Foundation of China under Grant 62088102, in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation. + +# References + +[1] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 2019. 1 +[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79:151-175, 2010. 1 +[3] Dimitri Bertsekas, Angela Nedic, and Asuman Ozdaglar. Convex analysis and optimization. Athena Scientific, 2003. 2, 4 +[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 18 +[6] Joya Chen, Qi Wu, Dong Liu, and Tong Xu. Foreground-background imbalance problem in deep object detectors: A review. In 2020 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), pages 285-290. IEEE, 2020. 3 +[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 1 +[8] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops, pages 1000-1001, 2020. 2 +[9] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 201-208. JMLR Workshop and Conference Proceedings, 2010. 4 + +[10] Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test-time adaptation for online vision-and-language navigation. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2, 6 +[12] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204-6218, 2022. 4 +[13] Gustavo A Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, and Christian Desrosiers. Clust3: Information invariant test-time training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6136-6145, 2023. 2 +[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 6, 18 +[15] Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In Proceedings of the European Conference on Computer Vision, pages 613-630. Springer, 2022. 5 +[16] Zixuan Hu, Yichun Hu, and Ling-Yu Duan. Seva: Leveraging single-step ensemble of vicinal augmentations for test-time adaptation. arXiv preprint arXiv:2505.04087, 2025. 2 +[17] Zixuan Hu, Yichun Hu, Xiaotong Li, Shixiang Tang, and Ling-Yu Duan. Beyond entropy: Region confidence proxy for wild test-time adaptation. arXiv preprint arXiv:2505.20704, 2025. 1, 2 +[18] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2 +[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1 +[20] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5 +[21] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162-14171, 2024. 2 +[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 6, 17 + +[23] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1 +[24] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 6, 17 +[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3, 4 +[26] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2 +[27] Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chiman Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218-19227, 2024. 2 +[28] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2 +[29] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 18 +[30] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021. +[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7941-7952, 2023. 1 +[32] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 2, 3, 17 +[33] Xinzhu Ma, Wanli Ouyang, Andrea Simonelli, and Elisa Ricci. 3d object detection from images for autonomous driving: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[34] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 + +[35] Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda. Power-law escape rate of sgd. In International Conference on Machine Learning, pages 15959-15975. PMLR, 2022. 4 +[36] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 17 +[37] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3 +[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 6 +[39] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3227-3237, 2020. 1, 5 +[40] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 2, 6, 17 +[41] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2 +[42] Prabhu Teja S and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 1 +[43] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1991-1999, 2019. 7 +[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2 +[45] Hugo Touchette. Legendre-fenchel transforms in a nutshell. URL http://www.maths.qmul.ac.uk/ht/archive/lfh2.pdf, page 25, 2005. 2, 4 +[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2020. 1, 3, 17 +[47] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 454-463, 2021. 2 + +[48] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 2, 3 +[49] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 1 +[50] Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9092-9101, 2021. 1 +[51] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23231-23241, 2024. 2 +[52] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerf-like representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 1 +[53] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 2, 17 +[54] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 17 +[55] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 2, 6, 17 +[56] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION. In The Eleventh International Conference on Learning Representations, 2023. 2 \ No newline at end of file diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/images.zip b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1364637444e68ca65446bb70f6a03131ad7163f --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f34b02f154714a30bc35492ebb4230d84ad9903ac1416b61ea2e53d560d29157 +size 894698 diff --git a/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/layout.json b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..09ba851d742565eee49bf56a95e6fb851b03f459 --- /dev/null +++ b/2025/Adaptive Dual Uncertainty Optimization_ Boosting Monocular 3D Object Detection under Test-Time Shifts/layout.json @@ -0,0 +1,9252 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 533, + 138 + ], + "type": "text", + "content": "Adaptive Dual Uncertainty Optimization: Boosting Monocular 3D Object Detection under Test-Time Shifts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": "Zixuan Hu" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Dongxiao Li" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Xinzhu Ma" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Shixiang Tang" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Xiaotong Li" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Wenhan Yang" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "text", + "content": " Ling-Yu Duan" + }, + { + "bbox": [ + 113, + 160, + 494, + 190 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 190, + 436, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 190, + 436, + 203 + ], + "spans": [ + { + "bbox": [ + 175, + 190, + 436, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 175, + 190, + 436, + 203 + ], + "type": "text", + "content": "School of Computer Science, Peking University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "spans": [ + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "type": "text", + "content": "Peng Cheng Laboratory, Shenzhen, China, " + }, + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 204, + 512, + 217 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Hongkong, China." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 220, + 374, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 220, + 374, + 232 + ], + "spans": [ + { + "bbox": [ + 228, + 220, + 374, + 232 + ], + "type": "text", + "content": "{hzxuan, lingyu}@pku.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 283, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 296, + 631 + ], + "type": "text", + "content": "Accurate monocular 3D object detection (M3OD) is pivotal for safety-critical applications like autonomous driving, yet its reliability deteriorates significantly under real-world domain shifts caused by environmental or sensor variations. To address these shifts, Test-Time Adaptation (TTA) methods have emerged, enabling models to adapt to target distributions during inference. While prior TTA approaches recognize the positive correlation between low uncertainty and high generalization ability, they fail to address the dual uncertainty inherent to M3OD: semantic uncertainty (ambiguous class predictions) and geometric uncertainty (unstable spatial localization). To bridge this gap, we propose Dual Uncertainty Optimization (DUO), the first TTA framework designed to jointly minimize both uncertainties for robust M3OD. Through a convex optimization lens, we introduce an innovative convex structure of the focal loss and further derive a novel unsupervised version, enabling label-agnostic uncertainty weighting and balanced learning for high-uncertainty objects. In parallel, we design a semantic-aware normal field constraint that preserves geometric coherence in regions with clear semantic cues, reducing uncertainty from the unstable 3D representation. This dual-branch mechanism forms a complementary loop: enhanced spatial perception improves semantic classification, and robust semantic predictions further refine spatial understanding. Extensive experiments demonstrate the superiority of DUO over existing methods across various datasets and domain shift types. The source code is available at https://github.com/hzcar/DUO." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 652, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Monocular 3D object detection (M3OD) [1, 33, 34] serves as a fundamental perception task, enabling agents to un" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 270, + 551, + 335 + ], + "blocks": [ + { + "bbox": [ + 359, + 259, + 509, + 270 + ], + "lines": [ + { + "bbox": [ + 359, + 259, + 509, + 270 + ], + "spans": [ + { + "bbox": [ + 359, + 259, + 509, + 270 + ], + "type": "text", + "content": "(a) Overview of Monocular 3D Object Detection" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 270, + 551, + 335 + ], + "lines": [ + { + "bbox": [ + 323, + 270, + 551, + 335 + ], + "spans": [ + { + "bbox": [ + 323, + 270, + 551, + 335 + ], + "type": "image", + "image_path": "51296b4f1e88d12619b34438f118bb4f0030761c26ff53f3f9edfc3cf01536e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 339, + 552, + 426 + ], + "blocks": [ + { + "bbox": [ + 323, + 339, + 552, + 426 + ], + "lines": [ + { + "bbox": [ + 323, + 339, + 552, + 426 + ], + "spans": [ + { + "bbox": [ + 323, + 339, + 552, + 426 + ], + "type": "image", + "image_path": "ed334cae71a4b29534ce7241d1691e88c3a62fb38ecd6712fdab21d5efcd5503.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 434, + 554, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 468 + ], + "type": "text", + "content": "Figure 1. (a) Illustration of M3OD. " + }, + { + "bbox": [ + 313, + 434, + 554, + 468 + ], + "type": "inline_equation", + "content": "(b, c)" + }, + { + "bbox": [ + 313, + 434, + 554, + 468 + ], + "type": "text", + "content": " Under test-time shifts, increased semantic and geometric uncertainties lead to a degradation of semantic perception and spatial localization capabilities." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 473, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 473, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 312, + 473, + 555, + 581 + ], + "type": "text", + "content": "derstand 3D scenes directly from 2D images, as shown in Fig. 1(a). Due to its low cost and simple hardware configuration, M3OD has attracted widespread attention, leading to the development of numerous detectors [7, 31, 52]. However, when facing adverse weather conditions or sensor failures in real-world deployments, well-trained models often suffer severe performance degradation under data shifts [2, 14, 49]. Therefore, it is crucial to deal with the OOD generalization problem for M3OD." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 714 + ], + "type": "text", + "content": "To address the distribution shifting issue with minimal overhead, Test-Time Adaptation (TTA) has emerged as a critical paradigm, enabling source models to adapt to target distributions via online updates [17, 19, 23]. The dominating strategy in this context involves minimizing prediction entropy, thereby reducing the uncertainty of the model on shifted data [17, 46]. Despite the promising results, existing TTA approaches largely overlook the dual uncertainty inherent in 3D detection—semantic uncertainty (related to class predictions) [42, 50] and geometric uncertainty (related to spatial location) [30, 39], which is a significant dif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": "*Corresponding Author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7273" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 233, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 233, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 233, + 83 + ], + "type": "text", + "content": "ference compared to conventional 2D tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 86, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 86, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 86, + 294, + 242 + ], + "type": "text", + "content": "To investigate these two types of uncertainty under test-time shifts, we analyze detection outcomes for objects subject to common real-world variations. As shown in Fig. 1 (b)&(c), our empirical study reveals that both uncertainties increase markedly with data shifts, creating compounded error accumulation in 3D detection. Moreover, we find that existing uncertainty optimization techniques exhibit significant limitations: 1) Low-score object neglect. Entropy minimization fails to provide effective supervision for challenging objects with low detection scores, resulting in inevitable omissions. 2) Spatial perception collapse. Direct minimization of depth uncertainty can cause model collapse, compromising the perception capacity of spatial attributes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 243, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 483 + ], + "type": "text", + "content": "In this work, to overcome the above limitations, we propose Dual Uncertainty Optimization (DUO), the first TTA framework for joint semantic-geometric uncertainty minimization. Specifically, through the lens of convex optimization theory [3, 45], we present an Legendre-Fenchel structure of the focal loss [25] and reconstruct the semantic uncertainty minimization as a dual optimization problem. Building on this foundation, we apply higher-order approximation analysis to derive a novel Conjugate Focal Loss. This loss breaks the label-dependence barrier in the original objective and introduces a dynamic weighting mechanism for balanced training, effectively reducing the omissions of objects with low scores. In parallel, we introduce a normal field constraint that enforces local consistency of surface normals in regions with high semantic confidence. This spatial coherence clarifies geometric cues, thereby reducing geometric uncertainty. Together, this dual-branch design creates a complementary loop where semantic confident regions bootstrap geometric feature learning, while enhanced spatial perception guides semantic refinement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "content": "We evaluate the effectiveness and generalizability of our method through experiments on the KITTI dataset [11] with 13 corruption shift types, achieving state-of-the-art results with average improvements of " + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "inline_equation", + "content": "+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}" + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "content": " in the Car category. Furthermore, we showcase its superior performance in addressing real-world shift scenarios (daytime " + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "content": " night, sunny " + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "content": " rainy) of nuScenes dataset [5], yielding an average gain of " + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "inline_equation", + "content": "+18\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 581 + ], + "type": "text", + "content": " compared to existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Contributions: 1) To the best of our knowledge, we pioneer dual uncertainty optimization in M3OD by establishing the first TTA framework that jointly minimizes semantic and geometric uncertainties, addressing a critical reliability gap in real-world deployments. 2) Through the lens of convex optimization theory, we derive a novel Conjugate Focal Loss that enables label-agnostic uncertainty weighting and balanced learning for low-score objects. This approach is inherently compatible with the source phase, requiring no additional hyperparameter tuning. 3) We introduce a normal field constraint that enforces the stability of geometric" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "representation with semantic guidance, resolving ambiguous spatial predictions. 4) We analyze and verify that our dual-branch design creates a complementary loop of two types of uncertainty optimization, resulting in significantly improved performance over existing TTA methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 137, + 401, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 401, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 401, + 150 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 155, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 553, + 370 + ], + "type": "text", + "content": "Monocular 3D Object Detection (M3OD) aims to perceive 3D objects from a single 2D image. Existing methods in M3OD can be broadly categorized based on their use of extra data sources, such as CAD models [28], dense depth maps [8, 47], or LiDAR [18, 41]. In this paper, we focus exclusively on approaches that utilize only monocular images, due to their computational efficiency and lower deployment costs. Previous studies, such as MonoDLE [32] and PGD [48], have identified depth estimation as a critical bottleneck in M3OD. To address this, many works leverage multiple geometric cues to integrate diverse depth predictions. For example, MonoFlex [55] integrates depth prediction by combining direct regression with multi-keypoint estimation; MonoGround [40] incorporates the ground plane as prior information; and MonoCD [53] exploits the complementary properties of multi-head estimation. In this paper, we investigate how to enhance the detection performance of such detectors under test-time shifts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "type": "text", + "content": "Test-Time Adaptation (TTA) aims to enhance model performance on out-of-distribution samples during inference. Depending on whether the source training process is modified, existing TTA methods can be mainly divided into two groups: Test-Time Training [13, 27, 44] and Fully Test-Time Adaptation [4, 16, 56]. In this paper, we focus on Fully TTA, which adapts models without source data. Prior studies have demonstrated that reducing prediction uncertainty is an effective strategy for improving model generalization in various tasks [10, 21, 26, 51]. These works have developed various strategies to model and optimize uncertainty. For instance, SAR [37] uses entropy as a measure of classification uncertainty along with sharpness-aware optimization. DeYO [22] incorporates entropy with a disentangled factor for uncertainty modeling. ReCAP [17] models regional uncertainty, enforcing implicit data scaling for uncertainty optimization. MonoTTA [24] optimizes positive and negative class uncertainties separately. However, unlike conventional 2D tasks, M3OD inherently exhibits both semantic and geometric uncertainties, which remain largely unexplored and lack an effective training paradigm. In this work, we explore the joint optimization of this dual uncertainty, enhancing the robustness of M3OD models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 657, + 391, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 391, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 391, + 670 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "Task Definition. M3OD aims to predict the 3D and semantic attributes of objects from a single RGB image. Given an input image " + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": ", the goal is to generate accurate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7274" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 180, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 180, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 180, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 180, + 167 + ], + "type": "image", + "image_path": "4fd75fea96986fd1e56e5e2a0b6efa144c0b2e6ee797b45f98b7d0d0cd5b93a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 174, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 54, + 174, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 54, + 174, + 555, + 220 + ], + "type": "text", + "content": "Figure 2. Semantic uncertainty (prediction entropy) and geometric uncertainty (average depth uncertainty) of the M3OD model under domain shifts. (a) shows variations in dual-uncertainty as the shift level increases. (b) records detection scores of predicted objects during inference with the source model. (c) records scores during the entropy minimization process. (d) investigates the depth uncertainty of different heads during the optimization with uncertainty regression loss. Experiments are conducted on the Gauss shift type of KITTI-C." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 306, + 169 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 306, + 169 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 306, + 169 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 306, + 169 + ], + "type": "image", + "image_path": "d6f655a59813be432ce02fd130fedc4ddb96960b34278be8b618ed379a7132ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 429, + 169 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 429, + 169 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 429, + 169 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 429, + 169 + ], + "type": "image", + "image_path": "cb2fe78abf57b64a5c0ec37fb380299220e70f35dd56280acf667131d80f2017.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 430, + 71, + 553, + 169 + ], + "blocks": [ + { + "bbox": [ + 430, + 71, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 430, + 71, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 553, + 169 + ], + "type": "image", + "image_path": "7768d4e328fbde4aa0503cc0c40d6c862df1e126c9327ca51ed07e6a1b4030ff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": "3D bounding boxes " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " and semantic labels " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " for objects present in the scene, where " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " denotes the number of objects. Each bounding box is typically parameterized as " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i = (\\mathbf{P}_i, \\mathbf{D}_i, \\mathbf{O}_i)" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " denotes the 3D center position, " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " represents the shape dimensions, and " + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i \\in [-\\pi, \\pi]" + }, + { + "bbox": [ + 54, + 230, + 294, + 338 + ], + "type": "text", + "content": " encodes the orientation. The multi-task nature of M3OD, which involves simultaneous estimation of both geometric and semantic attributes, poses significant challenges for achieving precise and coherent predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 338, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 295, + 506 + ], + "type": "text", + "content": "Meta-framework. As shown in Fig. 3(a), M3OD models typically employ a backbone network connected with multiple branches to predict various properties, e.g., score heatmaps, depth maps, etc. Since depth estimation is widely recognized as a key bottleneck [32, 48], many approaches employ a multi-head depth estimator to improve prediction accuracy. This estimator comprises a direct regression depth head and multiple geometric depth heads, each providing an individual depth prediction along with an associated uncertainty value. These outputs are then fused via an uncertainty-weighted average to produce the final depth prediction. We utilize the average uncertainty across all heads as our depth uncertainty metric. A detailed explanation is provided in the Appendix. D.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": "TTA Setting. TTA addresses the challenge of distribution shifts by enabling a pre-trained model to adapt to the target distribution during inference, without the need for labeled data. Unlike traditional training on fixed datasets, TTA operates in an online manner, where the model " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": " with parameter set " + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 506, + 295, + 590 + ], + "type": "text", + "content": " produces detection outputs while concurrently updating its parameters based on incoming test data:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 88, + 594, + 295, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 594, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 88, + 594, + 295, + 609 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathcal {B} _ {t} \\right\\}, \\left\\{\\mathcal {C} _ {t} \\right\\} \\leftarrow h _ {\\theta} \\left(\\mathrm {I} _ {t}\\right), \\theta \\leftarrow \\theta - \\nabla \\mathcal {L} _ {t t a} \\left(\\mathrm {I} _ {t}\\right), \\tag {1}", + "image_path": "4c35e46acdc7d6640cddfc25bc1ec05088fc20d75c69101e732a349a427aa1e4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "text", + "content": " denotes the incoming test data and " + }, + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tta}" + }, + { + "bbox": [ + 55, + 614, + 295, + 639 + ], + "type": "text", + "content": " denotes the loss function used for self-training during adaptation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 645, + 288, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 288, + 659 + ], + "type": "text", + "content": "4. Uncovering Dual-Uncertainty under Shifts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Since TENT [46] revealed the positive correlation between prediction uncertainty and generalization error under distribution shifts, numerous TTA methods have emerged to optimize uncertainty metrics. While previous works have devel" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 231, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 290 + ], + "type": "text", + "content": "oped diverse measures for conventional 2D tasks [22, 37], the compound semantic and geometric uncertainties in 3D perception remains critically unexplored in the TTA field. In this work, we investigate both types of uncertainty within the context of M3OD to understand their distinct roles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 291, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 398 + ], + "type": "text", + "content": "Specifically, to quantify the variation of uncertainty under distribution shifts, we track two metrics: semantic prediction entropy and geometric depth uncertainty (average uncertainty of multi-head depth estimator). As shown in Fig. 2(a), both metrics demonstrate a consistent upward trend as distribution shifts intensify, indicating that more severe shifts lead to higher model uncertainty. Furthermore, we analyze the independent optimization effects of two uncertainties and make the following key observations:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 399, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 506 + ], + "type": "text", + "content": "Observation 1: Conventional entropy minimization exacerbates imbalance distribution of detection scores. Different from classification scenarios, object detection suffers from extreme foreground-background imbalance, which hinders the effective optimization of hard positive objects [6, 25]. As shown in Fig. 2(b)&(c), entropy minimization yields a marginal gain for low-score objects while significantly boosting high-score predictions, further exacerbating the imbalance and leading to omissions of low-score objects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 626 + ], + "type": "text", + "content": "Observation 2: Minimizing depth uncertainty directly causes the model collapse of the multi-head depth estimator. To minimize depth uncertainty, we apply the uncertainty regression loss across multiple depth heads (detailed in the Appendix. D.2). As shown in Fig. 2(d), the regression head which lacks any geometric constraints, exhibits a significantly faster decline of uncertainty compared to keypoint heads. This rapid convergence reduces the multi-head system to a single deterministic head, undermining the model's ability to perform robust spatial understanding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 633, + 397, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 397, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 397, + 647 + ], + "type": "text", + "content": "5. Methodology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Based on the above observations, we propose a TTA method. Dual Uncertainty Optimization (DUO) for M3OD, which mainly leverages two novel designs, i.e., conjugate optimization framework and normal consistency constraint, to compatibly optimize semantic and geometric uncertainties." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7275" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 239, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 239, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 239, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 239, + 232 + ], + "type": "image", + "image_path": "82b02825d62808bc0a954dfd6d7b61f0b5b229de88224d908cc2d95fe40d88e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. Overview of our DUO method. (a) Meta-framework of the base models, where a backbone network connects to multi-branch predictors for estimating various object properties. (b) Through our theoretical conjugate optimization framework, we derive a conjugate focal loss " + }, + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CFL}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "type": "text", + "content": " that can adaptively adjust training weights across all classes without the ground-truth labels, while also identifying low-semantic-uncertainty regions for geometric constraints (Sec. 5.1). (c) DUO employs the efficient Sobel operator to convert the depth map to a normal field and design a normal consistency constraint " + }, + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCL}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 291 + ], + "type": "text", + "content": " to enhance geometric consistency in selected regions (Sec. 5.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 242, + 70, + 552, + 232 + ], + "blocks": [ + { + "bbox": [ + 242, + 70, + 552, + 232 + ], + "lines": [ + { + "bbox": [ + 242, + 70, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 242, + 70, + 552, + 232 + ], + "type": "image", + "image_path": "4885a4ae85addbce8880b2ccb215519ee8e39802a24770e6b3c5ec0df8eb0ffa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 295, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 310 + ], + "type": "text", + "content": "5.1. Semantic Uncertainty Conjugate Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 315, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 295, + 470 + ], + "type": "text", + "content": "To address the imbalance problem, it is crucial to devote more attention to challenging objects with high uncertainty. A straightforward approach is to employ the focal loss [25] which increases the weight of predictions with low probability. However, focal loss cannot assign an appropriate weight without the ground-truth label and fails to adapt effectively in unsupervised settings. To provide a more flexible weighting solution, we leverage convex optimization theory [3, 45] (a classical analytical tool) to explore the focal loss via a Legendre-Fenchel structure. Building on this foundation, we perform a higher-order approximation analysis to derive a novel loss that can dynamically adjust the weighting without relying on labels, as shown in Fig. 3(b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "content": "Legendre-Fenchel Structure. Given the source model " + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "content": " pre-trained with the focal loss, we denote " + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "inline_equation", + "content": "h \\triangleq h_\\theta(x)" + }, + { + "bbox": [ + 55, + 471, + 295, + 507 + ], + "type": "text", + "content": " and its corresponding loss function can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 514, + 295, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 514, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 100, + 514, + 295, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F L}} (x, y) = - \\alpha (1 - p) ^ {\\gamma} \\cdot y \\log p, \\tag {2}", + "image_path": "ab1b97ff16861d6ae028fcd5d6752720081230895e4c7032c4e204d858679233.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "p \\triangleq \\text{softmax}(h)" + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": " is the normalized probability over the classes, " + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": " is a detected object, and " + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": " is the one-hot coding of the ground-truth label. " + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 536, + 296, + 597 + ], + "type": "text", + "content": " are hyperparameters controlling the weighting of uncertain predictions. Motivated by [12], we reformulate it as the following structure:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 603, + 225, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 603, + 225, + 617 + ], + "spans": [ + { + "bbox": [ + 127, + 603, + 225, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F L}} = f (h) - y ^ {\\top} g (h),", + "image_path": "c75c50ae6e41344fac75b1792cdc0e06dedaec74a8034e940766e3aaffbfceda.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 616, + 294, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 616, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 616, + 294, + 633 + ], + "type": "interline_equation", + "content": "f (h) = \\alpha \\log s, g (h) = \\alpha h + \\alpha ((1 - p) ^ {\\gamma} - 1) \\log p, \\tag {3}", + "image_path": "819f4ff573a4f8a374ff30d4470148ded1634c0a753dff7f8ffb53bb282b751f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "s \\triangleq e^{h_1} + \\ldots + e^{h_c}" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is the sum over the exponential outputs of the model and " + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 641, + 295, + 677 + ], + "type": "text", + "content": " is the number of classes. Refer to the Appendix. A for missing proofs of this subsection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Problem Reconstruction. Under this structure, the optimization problem can be regarded as finding an optimal representation " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " that minimizes the empirical loss. According" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "text", + "content": "to the Legendre-Fenchel condition [45], the existence of the conjugate function " + }, + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "text", + "content": " is equivalent to the invertibility of the function " + }, + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 298, + 555, + 370 + ], + "type": "text", + "content": ". This critical condition is formally established in our work (see Appendix. A), serving as a fundamental theoretical guarantee for our method. Therefore, the minimum value of the objective can be expressed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 373, + 553, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 402 + ], + "type": "interline_equation", + "content": "\\min _ {h} \\left\\{f (h) - y ^ {\\top} g (h) \\right\\} = \\min _ {z = g (h)} \\left\\{f \\circ g ^ {- 1} (z) - y ^ {\\top} z \\right\\} = f ^ {*} (y). \\tag {4}", + "image_path": "d004898e0b4fe1ffd8041d5b4b4c2578dd6783dda0f1ae8c7b8c8dca4f88bad4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "text", + "content": "Building on the common assumption that the representation " + }, + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "text", + "content": " pre-trained from the large source dataset is already close to a locally optimal solution " + }, + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 313, + 402, + 554, + 451 + ], + "type": "text", + "content": " [9, 35], we can convert the problem into the following relationships:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 327, + 455, + 553, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 455, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 327, + 455, + 553, + 471 + ], + "type": "interline_equation", + "content": "f \\circ g ^ {- 1} (z) - y ^ {\\top} z = f ^ {*} (y), \\quad \\nabla_ {z} \\left(f \\circ g ^ {- 1}\\right) = y. \\tag {5}", + "image_path": "502d1f53edec2ce7ff9bcbcebb67dbdf8f8b7a21a50eb2c4fa2bfa18b8770d08.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 512 + ], + "type": "text", + "content": "Conjugate Focal Loss. By applying the chain rule of gradient computation and higher-order approximation, we can obtain the following estimation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 517, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 517, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 320, + 517, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} y _ {0} \\triangleq \\frac {\\nabla_ {h} \\left(f \\circ g ^ {- 1}\\right)}{\\nabla_ {h} z} | _ {z = g (h)} = \\nabla_ {h} g (h) ^ {- 1} \\cdot \\nabla_ {h} f (h) \\tag {6} \\\\ \\approx (I + \\gamma (1 - \\log p) \\cdot p p ^ {\\top} - \\gamma \\log p \\cdot \\operatorname {d i a g} (p))) ^ {- 1} p, \\\\ \\end{array}", + "image_path": "1242a5e7512835945cb0bfa5f28c264b44cb96d9349f3abb73219b4450b8acba.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{diag}(\\cdot)" + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "text", + "content": " denotes the diagonal matrix and " + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "text", + "content": " denotes the identity matrix. This estimation " + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "inline_equation", + "content": "y_{0}" + }, + { + "bbox": [ + 313, + 564, + 554, + 624 + ], + "type": "text", + "content": " ensures that the loss aligns closely with the conjugate function. Ultimately, we substitute it into Equ. 5, yielding an unsupervised approximation of the conjugate function:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 330, + 629, + 553, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 629, + 553, + 660 + ], + "spans": [ + { + "bbox": [ + 330, + 629, + 553, + 660 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C F L}} (x) = - \\alpha (1 - p) ^ {\\gamma} (I + \\gamma (1 - \\log p) \\cdot p ^ {\\top} p \\tag {7} \\\\ - \\gamma \\log p \\cdot \\operatorname {d i a g} (p)) ^ {- 1} p \\log p. \\\\ \\end{array}", + "image_path": "e909f3b0166590efb8ce7d2f328608c8a52db9a48c89c57ef919b746c23c4d44.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "In contrast to the focal loss in Equ. 2, our derived Conjugate Focal Loss (CFL) offers a dynamic adjustment of uncertainty weighting without relying on ground-truth labels and it confers the following advantages:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7276" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Static vs. Dynamic Adjustment. The vanilla focal loss uses a fixed weighting term for the ground-truth class to focus on high-uncertainty predictions. In contrast, our CFL not only incorporates " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(1 - p)^{\\gamma}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " to address the imbalance training issue but also dynamically adjusts the weighting across all classes based on the term " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(I + \\gamma (1 - \\log p)pp^{\\top} - \\gamma \\log p\\cdot \\mathrm{diag}(p))^{-1}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", which encodes inter-class prediction relationships." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 204 + ], + "type": "text", + "content": "Ground-Truth Independence. While focal loss requires labels to compute the weight for the loss term, CFL operates solely based on the prediction probability, eliminating the need for labeled data during TTA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 294, + 288 + ], + "type": "text", + "content": "Compatible Hyperparameters with Source Phase. Our theoretical analysis suggests that hyperparameters " + }, + { + "bbox": [ + 55, + 204, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\alpha, \\gamma" + }, + { + "bbox": [ + 55, + 204, + 294, + 288 + ], + "type": "text", + "content": " should remain consistent with their values used in focal loss during the source training. This compatibility provides a practical advantage for TTA scenarios, eliminating the need for extensive hyperparameter tuning. The validity of this consistent setting is empirically verified in Appendix C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 292, + 278, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 278, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 278, + 303 + ], + "type": "text", + "content": "5.2. Semantic-Guided Normal Field Constraint" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 295, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 463 + ], + "type": "text", + "content": "Despite progress in uncertainty modeling, current methods struggle to handle geometric uncertainty in TTA scenarios due to critical limitations: 1) Model Collapse: Direct optimization of model-predicted uncertainty leads to the degenerated predictor, as discussed in Sec. 4. 2) Substantial Overhead: Geometry-aware methods typically require additional data or offline training for uncertainty estimation, limiting their real-time applicability [15, 39]. To overcome these challenges, we propose an efficient normal field constraint derived from a single image, which enhances the geometric coherence of 3D predictions, thereby reducing the uncertainty stemming from the unstable geometric representation, as shown in Fig. 3(c)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "text", + "content": "Normal Field. Given a depth map " + }, + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "text", + "content": ", we restore it to the original image resolution using bilinear interpolation, allowing for the alignment with the pixel grid. Then, we compute the spatial gradients of the depth map using efficient Sobel operators [20], which approximate the rate of change in the depth values along horizontal and vertical directions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 101, + 543, + 294, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 543, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 101, + 543, + 294, + 556 + ], + "type": "interline_equation", + "content": "\\nabla D _ {x} = \\mathbf {S} _ {x} * D, \\quad \\nabla D _ {y} = \\mathbf {S} _ {y} * D, \\tag {8}", + "image_path": "acb8c395c707a61249ef0fa1e789ca887b32363be5916d58ce32b4032080c44e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_x" + }, + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_y" + }, + { + "bbox": [ + 55, + 563, + 295, + 624 + ], + "type": "text", + "content": " denote horizontal and vertical Sobel kernels, respectively. These gradient maps capture the variation in depth across neighboring pixels. The surface normal field, which encodes the orientation of the surface at each pixel, is then derived from the gradients as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 630, + 294, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 630, + 294, + 670 + ], + "spans": [ + { + "bbox": [ + 83, + 630, + 294, + 670 + ], + "type": "interline_equation", + "content": "\\mathbf {N} (u, v) = \\frac {1}{\\sqrt {\\nabla D _ {x} ^ {2} + \\nabla D _ {y} ^ {2} + 1}} \\left[ \\begin{array}{c} - \\nabla D _ {x} \\\\ - \\nabla D _ {y} \\\\ 1 \\end{array} \\right], \\tag {9}", + "image_path": "ae6c2e99e8c44baa91fa8029f11c494286a49921060931485681a4648d5ff056.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{N}(u,v)" + }, + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "text", + "content": " denotes the normal orientation at pixel " + }, + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 677, + 301, + 713 + ], + "type": "text", + "content": ". Normal Consistency Loss. To quantify geometric uncertainty, we employ an edge-aware Normal Consistency Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "(NCL), that encourages smoothness in the local surface by penalizing inconsistencies between neighboring pixels:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 101, + 547, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 101, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 319, + 101, + 547, + 116 + ], + "type": "interline_equation", + "content": "\\psi_ {x} (u, v) = \\left\\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u + 1, v) - \\mathbf {N} (u - 1, v) \\right\\| _ {2} ^ {2},", + "image_path": "de94fc7258761f488cbd5e78ebb6bc55b6302c9e231a779eb2c75b96d106d9ff.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 118, + 553, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 141 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 141 + ], + "type": "interline_equation", + "content": "\\psi_ {y} (u, v) = \\| 2 \\mathbf {N} (u, v) - \\mathbf {N} (u, v + 1) - \\mathbf {N} (u, v - 1) \\| _ {2} ^ {2}, \\tag {10}", + "image_path": "f3bfd7da0cad55cb0579405389f6c1712d496b3b86269833ddf122a74e2e0836.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 142, + 553, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 553, + 178 + ], + "type": "text", + "content": "where the smoothness terms " + }, + { + "bbox": [ + 313, + 142, + 553, + 178 + ], + "type": "inline_equation", + "content": "\\psi_x(u,v),\\psi_y(u,v)" + }, + { + "bbox": [ + 313, + 142, + 553, + 178 + ], + "type": "text", + "content": " enforce horizontal and vertical consistency in the surface normal field. The total normal consistency loss is then given by:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 183, + 553, + 196 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 183, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 318, + 183, + 553, + 196 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {N C L}} (u, v) = \\left(\\psi_ {x} (u, v) + \\psi_ {y} (u, v)\\right) \\cdot \\exp (- \\| \\nabla \\mathrm {I} (u, v) \\| _ {2}), \\tag {11}", + "image_path": "093f3d5e77980c3b4d9a7154bd98d6ee9967cdf8b8717b834a121fbf358674c4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "type": "text", + "content": "where the edge-aware weighting term " + }, + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "type": "inline_equation", + "content": "\\exp (-\\| \\nabla \\mathbf{I}(u,v)\\| _2)" + }, + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "type": "inline_equation", + "content": "= \\exp (-\\sqrt{|\\mathbf{S}_x*\\mathbf{I}(u,v)|^2 + |\\mathbf{S}_y*\\mathbf{I}(u,v)|^2})" + }, + { + "bbox": [ + 313, + 201, + 553, + 272 + ], + "type": "text", + "content": " preserves discontinuities at boundaries while enforcing smoothness in homogeneous regions. NCL encourages the model to learn a spatially consistent normal field, thereby reducing uncertainty stemming from unstable 3D representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "content": "Semantic Guidance. To enforce geometric-semantic coherence, we generate masks by integrating 2D bounding boxes with semantic predictions, ensuring synchronized focus regions for dual-branch uncertainty optimization. Let " + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "content": " denote detected bounding boxes with scores " + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "content": " and semantic uncertainties " + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\{U_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "content": " derived from the Conjugate Focal Loss (" + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CFL}}" + }, + { + "bbox": [ + 313, + 273, + 554, + 381 + ], + "type": "text", + "content": " in Eq. 7). To ensure reliable supervision, we select boxes with low semantic uncertainty using an exponentially moving average threshold:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 389, + 553, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 389, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 389, + 553, + 414 + ], + "type": "interline_equation", + "content": "R = \\{i \\mid U _ {i} \\leq \\bar {U} \\}, \\bar {U} \\leftarrow \\beta \\cdot \\frac {\\sum_ {i = 1} ^ {n} U _ {i}}{n} + (1 - \\beta) \\cdot \\bar {U}, \\tag {12}", + "image_path": "e9416a8d25425625b233f624cfb9970ddaa137d56f4ba42f810c7b819572f91c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 418, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 418, + 553, + 443 + ], + "type": "inline_equation", + "content": "\\beta \\in [0,1]" + }, + { + "bbox": [ + 313, + 418, + 553, + 443 + ], + "type": "text", + "content": " is a moving average factor (set to 0.1 in default). We then construct the region mask as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 355, + 448, + 553, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 448, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 355, + 448, + 553, + 466 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (u, v) = \\max _ {i \\in R} s _ {i} \\cdot \\mathbb {I} _ {\\text {i n s i d e}} (u, v \\mid \\mathcal {B} _ {i}), \\tag {13}", + "image_path": "04f63b250b6644d431108507480370ef2eb6008908be7bf72ccb87a9159b099e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{\\mathrm{inside}}(u,v\\mid \\mathcal{B}_i)" + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "text", + "content": " is an indicator function that returns 1 if pixel " + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "text", + "content": " is inside " + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 472, + 553, + 533 + ], + "type": "text", + "content": ", and 0 otherwise. This semantic-guided mask ensures that only low semantic-uncertainty regions contribute to the normal field constraint, enhancing the reliability of the geometric constraint." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 535, + 421, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 421, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 421, + 548 + ], + "type": "text", + "content": "5.3. Overall Objective" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 649 + ], + "type": "text", + "content": "The overall objective integrates the conjugate focal loss with the normal consistency constraint into a unified framework, simultaneously addressing semantic and geometric uncertainties. This dual optimization enables a complementary feedback loop: low-uncertainty spatial location enhances the model's ability to perform precise semantic classification, while confident semantic predictions in turn improve spatial understanding. The procedure is as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 654, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 654, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 318, + 654, + 553, + 672 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\Sigma_ {x \\in I} \\mathcal {L} _ {\\mathrm {C F L}} (x) + \\lambda \\Sigma_ {(u, v) \\in I} \\mathcal {M} (u, v) \\cdot \\mathcal {L} _ {\\mathrm {N C L}} (u, v), \\tag {14}", + "image_path": "de8152c00c26f78b8b42fcb9734f635d0ea3a0a4f5d66635ee67ac35e8f86392.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " iterates over all detected objects and " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " spans all pixels in the image I. " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CFL}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCL}}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " are defined in Equ. 7, Equ. 11, and Equ. 13. " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.7" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " by default." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7277" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 88, + 524, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "lines": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 60, + 555, + 84 + ], + "type": "text", + "content": "Table 1. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoFlex as the base model. We highlight the best and second results with **bold** and **underline** respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 88, + 524, + 383 + ], + "lines": [ + { + "bbox": [ + 81, + 88, + 524, + 383 + ], + "spans": [ + { + "bbox": [ + 81, + 88, + 524, + 383 + ], + "type": "table", + "html": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.000.000.0011.020.005.0712.947.7014.460.002.135.65
• TENTICLR'215.3111.096.402.2225.492.1423.8828.9635.4037.0724.6722.5930.63
• EATAICML'225.4412.124.672.7525.662.4524.7728.9935.6736.9524.3322.4734.10
• DeYOICLR'245.7812.524.523.0126.052.9824.9129.4035.1937.5923.7523.8234.33
• MonoTTAECCV'245.9313.344.053.3528.103.2125.8629.1036.4337.1825.9025.0133.89
• OursThis paper7.3015.409.364.3430.236.8929.0929.7638.3837.7229.3525.8834.97
Pedestrian Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.100.000.004.090.000.511.671.772.090.000.320.95
• TENTICLR'210.712.070.861.3612.700.816.728.3311.8912.797.426.459.78
• EATAICML'220.982.150.761.4412.230.806.778.6711.6013.627.536.4710.33
• DeYOICLR'241.032.250.911.6411.850.806.639.0911.7713.997.596.3910.57
• MonoTTAECCV'241.772.880.341.7812.380.827.039.0212.3113.117.757.1011.08
• OursThis paper1.893.081.541.8613.531.757.689.0912.6613.997.817.2711.27
Cyclist Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoFlexCVPR'210.000.200.000.000.240.002.142.331.724.410.000.000.00
• TENTICLR'210.060.140.040.044.550.936.638.2311.9415.167.721.852.81
• EATAICML'220.050.150.030.024.661.106.737.5813.7714.997.322.032.82
• DeYOICLR'240.060.190.030.034.911.086.486.9113.9414.677.571.792.82
• MonoTTAECCV'240.050.120.010.024.801.256.758.2413.3114.957.552.112.88
• OursThis paper0.110.220.070.106.002.006.898.4113.5815.947.942.132.92
", + "image_path": "a07bb1b7be0223dcbdc4c28b069718a00e4894bcc40447b0299b8db123e0af59.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 81, + 418, + 524, + 518 + ], + "blocks": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 555, + 412 + ], + "type": "text", + "content": "Table 2. Comparisons with state-of-the-art methods on the KITTI-C validation set (severity level 5) with MonoGround as the base model. Due to the space limit, the complete results of three categories are provided in Tab. 7 of Appendix. B." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 418, + 524, + 518 + ], + "lines": [ + { + "bbox": [ + 81, + 418, + 524, + 518 + ], + "spans": [ + { + "bbox": [ + 81, + 418, + 524, + 518 + ], + "type": "table", + "html": "
Car Category
MethodReferenceNoiseBlurWeatherDigitalAvg
Gauss.ShotImpul.Defoc.GlassMotionSnowFrostFogBrit.Contr.PixelSat.
MonoGroundCVPR'220.000.000.000.0011.630.291.956.593.1419.250.004.663.743.94
• TENTICLR'216.8214.818.214.8828.382.6523.9228.0833.0636.7020.2230.6333.2720.90
• EATAICML'227.1215.268.815.0929.082.5224.1828.0333.4336.7821.6130.5033.4221.22
• DeYOICLR'247.3515.729.385.7430.012.9925.0328.5534.3237.3123.4130.9934.1621.92
• MonoTTAECCV'247.8816.7310.355.9731.193.0625.2428.9934.8537.8225.0031.6134.7922.57
• OursThis paper9.7218.8812.747.2433.025.2428.5030.7337.2739.4028.3433.2237.2424.73
", + "image_path": "ede0e78f33da7e66c03ca6801efb13773ab69451b3b0a9a75ebbb75d6f2a5e5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 137, + 535 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 137, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 137, + 535 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 172, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 172, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 172, + 559 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 602 + ], + "type": "text", + "content": "For a fair comparison, we follow the identical evaluation pipeline with prior work [24], including baseline models, training recipes, and evaluation protocols." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Datasets. We conduct experiments on the KITTI [11] and nuScenes [5] datasets. For KITTI, we use the KITTI-C version, which includes 13 distinct data corruption types [14] and five severity levels per type. Results represent the average performance across three difficulty levels, i.e. Easy, Moderate, and Hard. Note that we also provide more results of different severity levels of KITTI in Appendix. B.1. For nuScenes, we adopt the front-view images and construct the Daytime, Night, Sunny, and Rainy scenarios via their scene" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "text", + "content": "descriptions following [29]. Based on these scenes, we define 4 real-world adaptation tasks, i.e., Daytime " + }, + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "text", + "content": " Night and Sunny " + }, + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 313, + 522, + 555, + 583 + ], + "type": "text", + "content": " Rainy. Following the MonoTTA, we transfer the nuScenes dataset into the KITTI format. More details are provided in Appendix. D.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 587, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 556, + 685 + ], + "type": "text", + "content": "Compared Methods. Based on two representative base models MonoFlex [55] and MonoGround [40], we compare our DUO with several state-of-the-art methods: EATA [36] identifies reliable samples during entropy training. DeYO [22] leverages probability variations under augmentations as an additional cue to enhance entropy optimization. MonoTTA [24] boosts probabilities for high-score classes while applying negative learning to low-score classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Implementation Details. We implement our method and other baselines in PyTorch [38]. We employ the Stochastic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7278" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 294, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "lines": [ + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "text", + "content": "Table 3. Comparison with baselines on Daytime " + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "text", + "content": " Night and Sunny " + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "text", + "content": " Rainy of nuScenes dataset, regarding " + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D|R_{40}}}" + }, + { + "bbox": [ + 55, + 62, + 295, + 85 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 294, + 227 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 294, + 227 + ], + "type": "table", + "html": "
MethodMonoFlexMonoGround
D → NN → DAvg.D → NN → DAvg.
Source model1.532.752.146.971.094.03
• TENT3.333.453.398.361.665.01
• DeYO4.724.874.7911.011.406.21
• MonoTTA6.923.685.3013.611.297.45
• Ours9.055.417.2315.701.918.81
MethodS → RR → SAvg.S → RR → SAvg.
Source model6.8610.918.898.037.447.73
• TENT8.5311.6110.079.948.719.33
• DeYO9.3312.0410.6810.549.109.82
• MonoTTA9.4712.5511.0110.898.909.90
• Ours11.5413.2112.3812.889.5411.21
", + "image_path": "92d4d99dce17536d135c4261b2f714ad3b09faac07128ddcaec1884fc6375f52.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "text", + "content": "Gradient Descent (SGD) optimizer with the same learning rate as MonoTTA, a momentum of 0.9, and a batch size of 16 for KITTI, 4 for nuScenes. Parameters " + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 247, + 295, + 294 + ], + "type": "text", + "content": " are assigned default values of 0.7, 4, and 2, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "text", + "content": "Evaluation Protocols. We report experimental results using the Average Precision for 3D bounding boxes, denoted as " + }, + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3D|R_{40}}" + }, + { + "bbox": [ + 55, + 295, + 296, + 355 + ], + "type": "text", + "content": " [43]. Results represent the mean values across three difficulty levels, with Intersection over Union thresholds set to 0.5 for Cars and 0.25 for both Ped. and Cyc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 141, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 141, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 141, + 373 + ], + "type": "text", + "content": "6.2. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "text", + "content": "Evaluation on Corruption Shifts. We first compare our DUO with previous methods on the KITTI-C dataset at severity level 5. Due to space constraints, detailed comparisons with other severity levels are provided in the Appendix. B. The results, reported in Tab. 1 & 2, reveal several observations: 1) Under test-time distribution shifts, pretrained detectors experience significant performance degradation across all categories. 2) Existing TTA methods partially mitigate the adverse effects of distribution shifts in M3OD, but their performance remains suboptimal as they cannot address the dual uncertainty. 3) Our method consistently outperforms compared approaches across all categories and base models. Notably, the proposed DUO achieves the best or comparable performance under 13 types of corruptions, with performance gains of " + }, + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "inline_equation", + "content": "+2.1" + }, + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "inline_equation", + "content": "+2.2\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}" + }, + { + "bbox": [ + 55, + 378, + 295, + 605 + ], + "type": "text", + "content": " in the Car category for two different models. These observations validate the crucial role of our adaptation framework in addressing dual uncertainty in M3OD models and enhancing robustness against distribution shifts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Evaluation on Real-World Scenario. We further evaluate different methods on four real-world scenarios, as shown in Tab. 3. The experimental results yield the following observations: 1) Under real-world shifts, the pre-trained model still suffers severe performance degradation. 2) For Sunny " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " Rainy adaptation tasks on MonoFlex, existing methods achieve only marginal improvements, whereas our method boosts performance by " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "+4.7\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". 3) Our DUO brings significant performance improvement on two base models," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "text", + "content": "Table 4. Effects of components in our method. We conduct ablation studies of the conjugate focal loss " + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CFL}}" + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "text", + "content": ", normal consistency loss " + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{NCL}}" + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "text", + "content": ", and semantic guidance " + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 62, + 553, + 106 + ], + "type": "text", + "content": " on the KITTI-C validation set. \"Src.\" denotes the source model without adaptation." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 113, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 316, + 113, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 316, + 113, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 113, + 553, + 182 + ], + "type": "table", + "html": "
Src.\\( \\mathcal{L}_{\\text{CFL}} \\)\\( \\mathcal{L}_{\\text{NCL}} \\)\\( \\mathcal{M} \\)MonoFlexMonoGround
CarPedes.Cyc.Avg.CarPedes.Cyc.Avg.
✓4.540.880.832.083.941.790.522.08
✓20.986.604.3210.6322.898.862.5111.42
✓12.384.632.826.6115.696.401.918.00
✓✓22.176.994.7211.2924.299.272.7812.11
✓✓16.496.234.879.2019.687.982.9410.20
✓✓✓22.977.195.1011.7524.739.623.0212.46
", + "image_path": "4f21d287336aaa078cbbcecf86bde33cceab0b1d23e8398e474300efa9345ee0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 193, + 434, + 285 + ], + "blocks": [ + { + "bbox": [ + 316, + 193, + 434, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 193, + 434, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 193, + 434, + 285 + ], + "type": "image", + "image_path": "560cb5af9fc923d3f9f857e2a68622a8232d8f97d108d6f112b883616fa8dd74.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "type": "text", + "content": "Figure 4. Effects of components in our method. " + }, + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "type": "inline_equation", + "content": "(a)" + }, + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "type": "text", + "content": " shows the distribution of detection scores during adaptation with conjugate focal loss. " + }, + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 313, + 295, + 553, + 339 + ], + "type": "text", + "content": " records the variation of depth uncertainty in different heads during adaptation with normal field constraints." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 193, + 550, + 285 + ], + "blocks": [ + { + "bbox": [ + 434, + 193, + 550, + 285 + ], + "lines": [ + { + "bbox": [ + 434, + 193, + 550, + 285 + ], + "spans": [ + { + "bbox": [ + 434, + 193, + 550, + 285 + ], + "type": "image", + "image_path": "a80e21ac4a14bca26c2854d9923aa98753908fe1511cf6606ed338ca526bcf77.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 553, + 376 + ], + "type": "text", + "content": "maintaining the best performance across all four scenarios, further demonstrating its effectiveness and superiority." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 408, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 408, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 408, + 397 + ], + "type": "text", + "content": "6.3. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "content": "In this section, without loss of generality, we conduct ablation studies on Monoflex under Gaussian shift in KITTI-C for the sake of brevity. Focusing on two pivotal components of DUO, Conjugate Focal Loss and Normal Field Constraint, we perform extensive experiments to analyze their independent roles and complementary effects, gaining insights into key factors contributing to their effectiveness. Effectiveness of Components. We investigate the impact of individual components by comparing the full method with variations that omit key parts. As shown in Tab. 4, incorporating either the conjugate focal loss or the normal field constraint significantly enhances detection performance, yielding average gains of " + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "inline_equation", + "content": "+8.9" + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "inline_equation", + "content": "+7.6\\mathrm{AP}_{3\\mathrm{D}|\\mathrm{R}_{40}}" + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "content": ", respectively. Notably, using the normal consistency loss " + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{NCL}})" + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "content": " alone results in unstable, marginal improvements; its effectiveness depends on being combined with " + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 402, + 555, + 628 + ], + "type": "text", + "content": ", highlighting the necessity of semantic guidance for effective geometric uncertainty reduction. These findings underscore the effectiveness of each component." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, a comparison of score distributions in Fig. 4(a) and Fig. 2(c) shows that the conjugate focal loss consistently boosts detection scores, ensuring effective training for challenging low-score objects. Similarly, Fig. 4(b) versus Fig. 2(d) demonstrates that our normal field constraint consistently reduces geometric uncertainty across all heads, preventing the model collapse observed in baseline meth" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7279" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 175, + 176 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 137, + 80 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 137, + 80 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 137, + 80 + ], + "type": "text", + "content": "(a) Source Model" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 175, + 176 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 175, + 176 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 175, + 176 + ], + "type": "image", + "image_path": "15faa6deb1adce398932b6e14a93d4cc9a2fe9bb4a0ed1a66ef6defac327c036.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 178, + 153, + 186 + ], + "lines": [ + { + "bbox": [ + 63, + 178, + 153, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 178, + 153, + 186 + ], + "type": "text", + "content": "(c) Normal Field Constraint" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 181, + 83, + 294, + 176 + ], + "blocks": [ + { + "bbox": [ + 181, + 71, + 274, + 81 + ], + "lines": [ + { + "bbox": [ + 181, + 71, + 274, + 81 + ], + "spans": [ + { + "bbox": [ + 181, + 71, + 274, + 81 + ], + "type": "text", + "content": "(b) Conjugate Optimization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 83, + 294, + 176 + ], + "lines": [ + { + "bbox": [ + 181, + 83, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 181, + 83, + 294, + 176 + ], + "type": "image", + "image_path": "7d031504bb751695f8c862babdca7915fa42a7c436f5497a698c98bd478741c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 178, + 258, + 186 + ], + "lines": [ + { + "bbox": [ + 197, + 178, + 258, + 186 + ], + "spans": [ + { + "bbox": [ + 197, + 178, + 258, + 186 + ], + "type": "text", + "content": "(d) Complete DUO" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 189, + 174, + 283 + ], + "blocks": [ + { + "bbox": [ + 59, + 189, + 174, + 283 + ], + "lines": [ + { + "bbox": [ + 59, + 189, + 174, + 283 + ], + "spans": [ + { + "bbox": [ + 59, + 189, + 174, + 283 + ], + "type": "image", + "image_path": "330347b61f0aaf811b7912c5b4a22c15c148ee822785e4b924000a1f430f7292.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 346 + ], + "type": "text", + "content": "Figure 5. Impacts of different components on dual uncertainty. Colored lines in three-dimensional space show the cumulative average of dual uncertainty during the adaptation process. Green and blue lines in two-dimensional planes record geometric uncertainty and semantic uncertainty, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 181, + 189, + 294, + 283 + ], + "blocks": [ + { + "bbox": [ + 181, + 189, + 294, + 283 + ], + "lines": [ + { + "bbox": [ + 181, + 189, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 181, + 189, + 294, + 283 + ], + "type": "image", + "image_path": "af2a37be2135ef91f1dad218063f98ad50ad6225973849c75edccfc3d570d2dd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "type": "text", + "content": "ods. Therefore, these innovations effectively overcome the limitations of existing methods identified in Sec. 4, leading to more robust and reliable detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 405, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 572 + ], + "type": "text", + "content": "Complementary Effects. As shown in Tab. 4, our complete method, which integrates all components, consistently achieves the best detection performance, demonstrating the compatibility of different components. To further investigate the complementary effects of our dual-branch design, we visualize the dual-uncertainty optimization process. Specifically, comparing Fig. 5(a)&(b), we observe that applying the conjugate focal loss significantly reduces semantic uncertainty, while interestingly, geometric uncertainty also shows a modest decline. A similar phenomenon is evident when employing the normal field constraint in Fig. 5(c). This synchronous behavior suggests an intrinsic interdependence between the two types of uncertainty, validating the potential of joint optimization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 661 + ], + "type": "text", + "content": "Moreover, by combining two innovations in our DUO framework, it achieves the fastest and most pronounced decrease in two uncertainties compared to individual components, as shown in Fig. 5(d). These observations further validate that our dual-branch architecture creates a complementary loop for dual-uncertainty optimization, effectively harnessing their complementary effects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Robustness and Efficiency. To validate the robustness and efficiency of our method, we extend our analysis by examining the sensitivity of key hyper-parameters and comparing running times in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 132 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 132 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 132 + ], + "type": "image", + "image_path": "5e54163e90545ec7c04503048da046bbc1d4dfad942d61bbe3e0e091266d580c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 132, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 316, + 132, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 194 + ], + "type": "image", + "image_path": "9f967bf6dfef5de4ccfab6c80aac429d164e916bc3273fbed12d4b911beac353.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 194, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 316, + 194, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 255 + ], + "type": "image", + "image_path": "15f60dd631e50948e0cadc2ec0e24dbf1d6fb02ce53cf6b56440acb9ccf89e1f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 261, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 327 + ], + "type": "text", + "content": "Figure 6. Qualitative examples on KITTI-C. In each row, we provide the front view (left) and the bird's-eye view (right) visualizations. Red represents the ground truth of boxes, while Green represents the predictions. We circle some objects to highlight differences in predictions. The first row corresponds to the source model, the second to MonoTTA, and the third to our method." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 340, + 425, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 340, + 425, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 340, + 425, + 353 + ], + "type": "text", + "content": "6.4. Qualitative Results" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 357, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 490 + ], + "type": "text", + "content": "Based on the qualitative results shown in Fig. 6, our DUO framework produces predictions that are more precisely aligned with the ground-truth annotations. Compared with the latest SOTA method, DUO not only significantly improves the accuracy of 3D location estimation but also reduces missed detections for challenging distant or small-scale objects, which are typically prone to large geometric errors. These visualizations further confirm that our dual uncertainty optimization effectively adapts the source model, achieving precise and reliable 3D perception that closely matches the ground truth." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 502, + 479, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 479, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 479, + 514 + ], + "type": "text", + "content": "7. Conclusion and Future Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a synergistic dual-branch framework designed to address semantic-geometric uncertainty inherent in 3D vision systems under test-time domain shifts. Our approach derives a novel conjugate loss to offer an adaptive, label-free weighting mechanism for balanced training on semantic uncertainty. Meanwhile, it incorporates a normal consistency constraint to reduce uncertainty from inconsistent geometric representation. Our extensive experiments demonstrate that the proposed dual-branch optimization creates a complementary loop, consistently improving performance across diverse domain shifts. In future work, we tend to expand our framework beyond M3OD to cover a broader range of 3D vision tasks. We hope that our work will deepen the understanding of uncertainty-aware model adaptation while providing transferable insights for related fields, such as unsupervised learning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7280" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 200 + ], + "type": "text", + "content": "This work was supported by the Program of Beijing Municipal Science and Technology Commission Foundation (No.Z241100003524010), in part by the National Natural Science Foundation of China under Grant 62088102, in part by AI Joint Lab of Future Urban Infrastructure sponsored by Fuzhou Chengtou New Infrastructure Group and Boyun Vision Co. Ltd, and in part by the PKU-NTU Joint Research Institute (JRI) sponsored by a donation from the Ng Teng Fong Charitable Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 115, + 223 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 231, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 231, + 296, + 285 + ], + "type": "text", + "content": "[1] Eduardo Arnold, Omar Y Al-Jarrah, Mehrdad Dianati, Saber Fallah, David Oxtoby, and Alex Mouzakitis. A survey on 3d object detection methods for autonomous driving applications. IEEE Transactions on Intelligent Transportation Systems, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 287, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 287, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 287, + 296, + 331 + ], + "type": "text", + "content": "[2] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79:151-175, 2010. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 333, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 296, + 364 + ], + "type": "text", + "content": "[3] Dimitri Bertsekas, Angela Nedic, and Asuman Ozdaglar. Convex analysis and optimization. Athena Scientific, 2003. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 367, + 296, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 367, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 62, + 367, + 296, + 410 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8344-8353, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 412, + 296, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 412, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 412, + 296, + 478 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 18" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 533 + ], + "type": "text", + "content": "[6] Joya Chen, Qi Wu, Dong Liu, and Tong Xu. Foreground-background imbalance problem in deep object detectors: A review. In 2020 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), pages 285-290. IEEE, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[7] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. Monopair: Monocular 3d object detection using pairwise spatial relationships. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12093-12102, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 646 + ], + "type": "text", + "content": "[8] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition workshops, pages 1000-1001, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 712 + ], + "type": "text", + "content": "[9] Dumitru Erhan, Aaron Courville, Yoshua Bengio, and Pascal Vincent. Why does unsupervised pre-training help deep learning? In Proceedings of the thirteenth international conference on artificial intelligence and statistics, pages 201-208. JMLR Workshop and Conference Proceedings, 2010. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[10] Junyu Gao, Xuan Yao, and Changsheng Xu. Fast-slow test-time adaptation for online vision-and-language navigation. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 120, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 553, + 173 + ], + "type": "text", + "content": "[11] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 176, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 176, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 553, + 219 + ], + "type": "text", + "content": "[12] Sachin Goyal, Mingjie Sun, Aditi Raghunathan, and J Zico Kolter. Test time adaptation via conjugate pseudo-labels. Advances in Neural Information Processing Systems, 35:6204-6218, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 222, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 222, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 222, + 553, + 277 + ], + "type": "text", + "content": "[13] Gustavo A Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, and Christian Desrosiers. Clust3: Information invariant test-time training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6136-6145, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 280, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 322 + ], + "type": "text", + "content": "[14] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 1, 6, 18" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "text", + "content": "[15] Julia Hornauer and Vasileios Belagiannis. Gradient-based uncertainty for monocular depth estimation. In Proceedings of the European Conference on Computer Vision, pages 613-630. Springer, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 404 + ], + "type": "text", + "content": "[16] Zixuan Hu, Yichun Hu, and Ling-Yu Duan. Seva: Leveraging single-step ensemble of vicinal augmentations for test-time adaptation. arXiv preprint arXiv:2505.04087, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 407, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 449 + ], + "type": "text", + "content": "[17] Zixuan Hu, Yichun Hu, Xiaotong Li, Shixiang Tang, and Ling-Yu Duan. Beyond entropy: Region confidence proxy for wild test-time adaptation. arXiv preprint arXiv:2505.20704, 2025. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "type": "text", + "content": "[18] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston H Hsu. Monodtr: Monocular 3d object detection with depth-aware transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4012-4021, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 552 + ], + "type": "text", + "content": "[19] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. Advances in Neural Information Processing Systems, 34:2427-2440, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 555, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 598 + ], + "type": "text", + "content": "[20] Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. Design of an image edge detection filter using the sobel operator. IEEE Journal of solid-state circuits, 23(2): 358-367, 1988. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "type": "text", + "content": "[21] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162-14171, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 6, 17" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7281" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[23] Jian Liang, Ran He, and Tieniu Tan. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[24] Hongbin Lin, Yifan Zhang, Shuaicheng Niu, Shuguang Cui, and Zhen Li. Monotta: Fully test-time adaptation for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 96-114. Springer, 2025. 2, 6, 17" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 263 + ], + "type": "text", + "content": "[26] Wei Lin, Muhammad Jehanzeb Mirza, Mateusz Kozinski, Horst Possegger, Hilde Kuehne, and Horst Bischof. Video test-time adaptation for action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22952-22961, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[27] Weihuang Liu, Xi Shen, Haolun Li, Xiuli Bi, Bo Liu, Chiman Pun, and Xiaodong Cun. Depth-aware test-time training for zero-shot video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19218-19227, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "text", + "content": "[28] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "type": "text", + "content": "[29] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 6, 18" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[30] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3111-3121, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[31] Yueru Luo, Chaoda Zheng, Xu Yan, Tang Kun, Chao Zheng, Shuguang Cui, and Zhen Li. Latr: 3d lane detection from monocular images with transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7941-7952, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[32] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4721-4730, 2021. 2, 3, 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[33] Xinzhu Ma, Wanli Ouyang, Andrea Simonelli, and Elisa Ricci. 3d object detection from images for autonomous driving: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Yuexin Ma, Tai Wang, Xuyang Bai, Huitong Yang, Yuenan Hou, Yaming Wang, Yu Qiao, Ruigang Yang, and Xinge Zhu. Vision-centric bev perception: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[35] Takashi Mori, Liu Ziyin, Kangqiao Liu, and Masahito Ueda. Power-law escape rate of sgd. In International Conference on Machine Learning, pages 15959-15975. PMLR, 2022. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[36] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 6, 17" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[37] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 260 + ], + "type": "text", + "content": "[38] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[39] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3227-3237, 2020. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[40] Zequn Qin and Xi Li. Monoground: Detecting monocular 3d objects from the ground. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3793-3802, 2022. 2, 6, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[41] Cody Reading, Ali Harakeh, Julia Chae, and Steven L Waslander. Categorical depth distribution network for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8555-8564, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "text", + "content": "[42] Prabhu Teja S and Francois Fleuret. Uncertainty reduction for model adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9613-9623, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 514 + ], + "type": "text", + "content": "[43] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1991-1999, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[44] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "text", + "content": "[45] Hugo Touchette. Legendre-fenchel transforms in a nutshell. URL http://www.maths.qmul.ac.uk/ht/archive/lfh2.pdf, page 25, 2005. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 647 + ], + "type": "text", + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2020. 1, 3, 17" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[47] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 454-463, 2021. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7282" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 563 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[48] Tai Wang, ZHU Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In Conference on Robot Learning, pages 1475-1485. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "text", + "content": "[49] Yan Wang, Xiangyu Chen, Yurong You, Li Erran Li, Bharath Hariharan, Mark Campbell, Kilian Q Weinberger, and WeiLun Chao. Train in germany, test in the usa: Making 3d object detectors generalize. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11713-11723, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 239 + ], + "type": "text", + "content": "[50] Yuxi Wang, Junran Peng, and ZhaoXiang Zhang. Uncertainty-aware pseudo label refinery for domain adaptive semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9092-9101, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 316 + ], + "type": "text", + "content": "[51] Yanshuo Wang, Ali Cheraghian, Zeeshan Hayden, Jie Hong, Sameera Ramasinghe, Shafin Rahman, David Ahmedt-Aristizabal, Xuesong Li, Lars Petersson, and Mehrtash Harandi. Backpropagation-free network for 3d test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23231-23241, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 318, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 294, + 372 + ], + "type": "text", + "content": "[52] Junkai Xu, Liang Peng, Haoran Cheng, Hao Li, Wei Qian, Ke Li, Wenxiao Wang, and Deng Cai. Mononerd: Nerf-like representations for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6814-6824, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 373, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 294, + 428 + ], + "type": "text", + "content": "[53] Longfei Yan, Pei Yan, Shengzhou Xiong, Xuanyu Xiang, and Yihua Tan. Monocd: Monocular 3d object detection with complementary depths. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10248-10257, 2024. 2, 17" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 294, + 473 + ], + "type": "text", + "content": "[54] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2403-2412, 2018. 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 518 + ], + "type": "text", + "content": "[55] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 2, 6, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[56] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7283" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_content_list.json b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..45eaabea8cbdf897a310655394a06b796fae262a --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_content_list.json @@ -0,0 +1,1723 @@ +[ + { + "type": "text", + "text": "Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections", + "text_level": 1, + "bbox": [ + 173, + 130, + 825, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youwei Zhou $^{1}$ , Tianyang Xu $^{1*}$ , Cong Wu $^{1}$ , Xiaojun Wu $^{1}$ , Josef Kittler $^{2}$ $^{1}$ Jiangnan University, China $^{2}$ University of Surrey, UK", + "bbox": [ + 220, + 202, + 776, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "youwei_zhou@stu.jiangnan.edu.cn", + "bbox": [ + 361, + 241, + 635, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tianyang_xu,Congwu，wu_xiaojun}@jiangnan.edu.cn", + "bbox": [ + 285, + 258, + 710, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "j.kittler@surrey.ac.uk", + "bbox": [ + 397, + 277, + 591, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The shared topology of human skeletons motivated the recent investigation of graph convolutional network (GCN) solutions for action recognition. However, most of the existing GCNs rely on the binary connection of two neighboring vertices (joints) formed by an edge (bone), overlooking the potential of constructing multi-vertex convolution structures. Although some studies have attempted to utilize hyper-graphs to represent the topology, they rely on a fixed construction strategy, which limits their adaptivity in uncovering the intricate latent relationships within the action. In this paper, we address this oversight and explore the merits of an adaptive hyper-graph convolutional network (HyperGCN) to achieve the aggregation of rich semantic information conveyed by skeleton vertices. In particular, our Hyper-GCN adaptively optimises the hyper-graphs during training, revealing the action-driven multi-vertex relations. Besides, virtual connections are often designed to support efficient feature aggregation, implicitly extending the spectrum of dependencies within the skeleton. By injecting virtual connections into hyper-graphs, the semantic clues of diverse action categories can be highlighted. The results of experiments conducted on the NTU-60, NTU-120, and NW-UCLA datasets demonstrate the merits of our Hyper-GCN, compared to the state-of-the-art methods. The code is available at https://github.com/6UOON9/Hyper-GCN.", + "bbox": [ + 89, + 357, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 777, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Skeleton-based human action recognition is a popular research topic in artificial intelligence, with practical applications in video understanding, video surveillance, human-computer interaction, robot vision, VR and AR [16, 21, 28-30, 42, 49, 49]. In general, a skeleton sequence contains", + "bbox": [ + 89, + 803, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77062d86835fd5f1b65731af15262b34468aac1e2504409fc194e5636374c94f.jpg", + "image_caption": [ + "Figure 1. Illustration of our Hyper-GCN. In (a), dotted lines represent the virtual connections. Each coloured part containing multiple joints represents a hyper-graph with hyper-edges. (b) provides an intuitive comparison between normal graph convolution and hyper-graph convolution operations with the same connectivity degree." + ], + "image_footnote": [], + "bbox": [ + 519, + 337, + 901, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a series of 2D or 3D coordinates, which can easily be collected by low-cost depth sensors or obtained by video-based pose estimation algorithms [26, 34, 35, 43]. Compared to RGB and optical flow images, skeleton data, which represents the basic physical structure of a human being, is of lower dimension, conveying human action with higher efficiency. Moreover, it is robust to illumination changes and scene variations. For these reasons, the adoption of this structural data is very popular in skeleton-based action recognition [3, 6, 10, 25, 35, 45, 51].", + "bbox": [ + 511, + 715, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To facilitate skeleton-based action recognition, both Recurrent Neural Networks (RNNs) [10, 25, 35, 45] and Con", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12648", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "volitional Neural Networks (CNNs) [7, 17, 40] have been well explored. However, RNNs themselves cannot depict the intrinsic skeleton topology. The learned filters of CNNs, on the other hand, neglect the spatio-temporal structure of the skeleton. Drawing on these observations, recent studies have focused on how to model the skeleton topology directly. In principle, the physical topology of human joints and bones can be consistently represented by a graph. Accordingly, the graph convolutional network (GCN) is typically introduced to aggregate feature information conveyed by skeleton joints [23, 52, 57]. In terms of data representations, skeleton joints are also one-dimensional sequence data, similar to language. Therefore, with the rise of recent self-attention techniques [9, 44], several Transformer-based action recognition solutions [32, 38, 47, 53] have been released. Transformer considers skeletal joints as tokens and uses attention maps to reflect the topology. These methods have achieved excellent performance but sacrifice much higher GFLOPS and parameters than GCN.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, two neighbouring joints can pass messages through their shared bone. In graph terms, this corresponds to the exchange of information between two vertices along their connecting edge. Besides the physical skeleton topology, attempts have been made to explore the implicit relationships among joints [3, 6, 23, 36], suggesting different variants of skeleton topology. In principle, these approaches assume a binary connection between each connected vertex pair. Mathematically, the constructed topology in the adjacency matrix is represented by a normal graph. However, human actions are jointly defined by several joints. Hence, human actions encompass not only binary relations between vertex pairs but also multi-joint relationships. For instance, the action primitive starting running is manifest in the human raising the left hand while the right leg steps forward, as shown in Figure 1 (a). The binary connections are not sufficient to capture the synergistic interaction of multiple joints. This strongly argues for constructing feature aggregation paths involving multiple vertices.", + "bbox": [ + 91, + 381, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accordingly, we propose to construct a hyper-graph to depict the skeleton topology and take advantage of the outstanding performance of hyper-graph analysis techniques [2, 12, 18, 19]. The hyper-graph topology involves multiple node connections, rather than binary connections of a normal graph. As shown in Figure 1 (a), in a hyper-graph, a hyper-edge can link more than two vertices. This linking has the capacity to represent complex collaborative relations among human joints. As one hyper-edge associates multiple vertices, a single hyper-graph convolution enables aggregating all the features along the hyper-edges.", + "bbox": [ + 89, + 671, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An illustration is provided in Figure 1 (b), where the normal graph and hyper-graph are presented to demonstrate their differences in passing the information conveyed by the vertex features. In the normal graph convolution, after", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2-layer aggregation, the information of the green vertex is spread to 2 other vertices. In contrast, in a hyper-graph convolution, the information of the green vertex spreads to all the vertices, creating an extended receptive field. Theoretically, by modelling a skeleton using a hyper-graph for the aggregation of joint information, improved efficiency can be obtained during learning the action semantics.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the key to achieving enhanced information aggregation by hyper-graph convolution lies in the conformity of the hyper-graph structure to the real topology of human action. Several studies [13, 58] attempt to adopt a hypergraph to model the skeleton topology, mostly based on superficial prior knowledge and an unfounded definition of the structure of hyper-graph. Distinguishing our approach from the existing hyper-graph-based methods, we design an adaptive solution for constructing the hyper-graph with virtual connections. It can construct a Non-uniform hypergraph specific to each individual action type.", + "bbox": [ + 511, + 198, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides the information aggregation issue, the capacity of an action recognition system is also constrained by its input features. In general, the number of skeleton joints is fixed in existing benchmarks, e.g., 25 for NTU-120 [26]. The entire recognition process relies on a successful capture of the interactions among the skeleton joints. In the domain of artistic puppetry, the history of driving actions, such as Shadow Play and Marionette goes back more than 2000 years. This kind of art form provides an inspiration for involving additional 'hyper joints' which can drive or facilitate a better communication between existing joints. The underlying spirit is to alleviate the pressure on the real joints to store and transfer complex semantics. Jointly with the hyper joints, real joints can focus more on storing neighbouring joint features and relegating the task of transferring the global information to the hyper joints. Interestingly, the class token in existing Transformers can also be considered as a hyper-token [9, 44]. As shown in Figure 1, the skeleton is described as a marionette, where the actions are \"controlled\" by connecting hyper joints to the real joints. This suggests that the hyper joints are not only able to capture the representation information of human action, but also reveal the implicit information between physically connected joints as hints for recognition.", + "bbox": [ + 511, + 364, + 903, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By endowing an adaptive non-uniform hyper-graph with hyper joints, virtual connections are created to perform comprehensive hyper-graph convolutions. We construct Hyper-GCN based on the above design principles. Extensive experiments on 3 datasets, NTU RGB+D 60 [35], NTU RGB+D 120 [26], and NW-UCLA [46], are conducted for evaluation. The results validate the merits of our proposed Hyper-GCN. The main contributions are as follows:", + "bbox": [ + 511, + 726, + 903, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- An adaptive non-uniform hyper-graph to represent the hu", + "bbox": [ + 511, + 849, + 903, + 864 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://en.wikipedia.org/wiki/Shadow Playing", + "bbox": [ + 529, + 875, + 862, + 886 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://en.wikipedia.org/wiki/Marionette", + "bbox": [ + 532, + 887, + 852, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12649", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "man skeleton topology. Compared with a fixed hypergraph, the constructed topology is more action-specific, thus boosting the discrimination.", + "bbox": [ + 102, + 90, + 482, + 135 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The injection of virtual hyper joints, enriching the connectivity of the physical joints, from a global semantic perspective.", + "- As the processing architecture, we propose a hyper-graph convolution network (Hyper-GCN). The SOTA performance achieved on three public datasets demonstrates the merits of the Hyper-GCN and virtual connection designs." + ], + "bbox": [ + 91, + 136, + 483, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 258, + 230, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Graph Topology for Action Recognition", + "text_level": 1, + "bbox": [ + 89, + 284, + 431, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A graph can represent a human skeleton, preserving the joint relationships via predefined edges. To aggregate semantics, GCNs [1, 4, 11, 20, 24, 37, 54, 57] and Transformers [8, 32, 38, 53, 55] have been well studied.", + "bbox": [ + 89, + 306, + 482, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For GCN-based approaches, ST-GCN [52] proposes to represent the topology with an adjacency matrix of 3 subsets by modeling the spatio-temporal relevance. Similarly, 2s-AGCN [36], InfoGCN [6], and DS-GCN [51] use a self-attentive mechanism to learn the topology from joint pairs. Besides the intuitive spatio-temporal dimensions, CTR-GCN [3] proposes to learn the topology for channels to refine the skeleton.", + "bbox": [ + 89, + 368, + 483, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For Transformer-based solutions, IIP-Transformer [47] adopt a self-attention mechanism to establish the intra-part and the inter-part joints. IGFormer [31] learns the topology between persons at both semantic and distance levels. STTFormer [33] proposed structure can capture the correlation between joints in consecutive frames.", + "bbox": [ + 89, + 489, + 483, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All the above two paradigms construct the topology from joint pairs, which rely only on the binary relation between two vertices. In their implementation stage, the adjacency matrix or attention map is used to represent the topology. In this case, high-order information among joints is not taken into consideration, neglecting the collaborative power among multiple joints. Though hyper-graph is considered to construct topology using multivariate joint relationships, current [13, 58] solutions manually set the hyper-graphs, which greatly relies on human experience, sacrificing the adaptability of graph learning.", + "bbox": [ + 89, + 580, + 482, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Feature Configurations for Action Recognition", + "text_level": 1, + "bbox": [ + 89, + 756, + 482, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It has been observed that the input features play the most essential role in delivering high recognition accuracy [1, 5, 15, 27, 31, 33, 41, 47] in GCNs and Transformers.", + "bbox": [ + 89, + 779, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing on this, Graph2Net [50] proposes to extract local and global spatial features jointly. CTR-GCN [3] uses multi-scale temporal convolution to extract temporal features. While HD-GCN [23] introduce the hierarchical edge convolution to key edge features. To extend the perception", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "field, STC-Net [22] uses the dilated kernels for graph convolution to capture the features.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, IG-Transformer [31] aggregate the features of two persons as one to mine the semantic information. STST [55] split the coordinate features, semantic features and temporal features into three kinds of tokens. TSTE [53] extracting and merging the spatial and temporal features as input to transformer encoders.", + "bbox": [ + 511, + 121, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After all, existing methods [6, 14, 32, 36, 38, 51, 52] receive the input features from real skeleton joints. However, each skeleton joint acts as an information carrier during forward passing, which is required to deliver both local context and global semantics. Given a fixed representative capacity (number of channels), we believe it is necessary to involve additional virtual joints to balance the pressure of storing local and global information.", + "bbox": [ + 511, + 212, + 906, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 513, + 347, + 620, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 372, + 653, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, the input features in graph convolution are represented by $\\mathbf{F}_{\\mathrm{in}}\\in \\mathbb{R}^{C\\times T\\times V}$ , where $C$ represents the number of channels in the feature maps. Given the topology of a human skeleton, we usually define the graph $\\mathcal{G} = (\\mathcal{V},\\mathcal{E})$ , where $\\mathcal{V}$ represents the set of joints and $\\mathcal{E}$ represents the set of edges between joints. For the set of edges $\\mathcal{E}$ , it is formulated as an adjacent matrix $A\\in \\mathbb{R}^{N\\times N}$ , where $N$ represents the number of human joints. The normalised adjacency matrix is represented by $\\hat{A}\\in \\mathbb{R}^{N\\times N}$ . The normalisation operation is formulated as follows:", + "bbox": [ + 511, + 393, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {A} = \\Lambda^ {- \\frac {1}{2}} A \\Lambda^ {- \\frac {1}{2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 554, + 903, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Lambda$ represents the diagonal matrix stored with the degree of every joint. The entire normal graph convolution process can be formulated as:", + "bbox": [ + 511, + 585, + 905, + 631 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {o u t}} = \\sigma (\\hat {A} \\mathbf {F} _ {\\text {i n}} P), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 643, + 903, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $P \\in \\mathbb{R}^{C \\times C'}$ is the learnable parameters, representing the feature transformation patterns in the feature space. $\\sigma$ denotes the non-linear activation function ReLU.", + "bbox": [ + 511, + 671, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hyper-graph", + "text_level": 1, + "bbox": [ + 513, + 727, + 650, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we use $\\mathcal{G}_{\\mathcal{H}} = (\\mathcal{V}_{\\mathcal{H}},\\mathcal{E}_{\\mathcal{H}},\\mathcal{W}_{\\mathcal{H}})$ to define the spatial hyper-graph [12] with human skeleton. $\\mathcal{V}_{\\mathcal{H}}$ and $\\mathcal{E}_{\\mathcal{H}}$ follow similar definitions in the normal graph, which represent the set of joints and the set of hyper-edges. In addition, we introduce $\\mathcal{W}_{\\mathcal{H}}$ to represent the weights of each hyper-edge. Since a hyper-edge contains multiple joints, the corresponding topology can no longer be simply represented by an adjacency matrix. Therefore, we introduce the incidence matrix $H\\in \\mathbb{R}^{N\\times M}$ to describe the topology of each joint in the hyper-graph. $N$ represents the number of joints and $M$", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12650", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/43bfcada4a3d50f8c88cf05b02a441690a45df1e21f4844af313cbf1e1eee388.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 79, + 715, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db5a874a144d77a1d8dac9f28bd4f697675133051f3bab7ae9a3902818567e32.jpg", + "image_caption": [ + "Figure 2. The framework of our Hyper-GCN. Part (a) represents the Multi-head Hyper-graph Convolution (M-HGC) module. Part (b) represents the process of constructing an adaptive hyper-graph. Part (c) represents the architecture of Hyper-GCN. $\\mathbf{F}_{\\mathbf{P}}$ represents the features of real joints. $\\mathbf{F}_h$ represents the hyper-joints which are learnable parameters. $\\hat{A}$ represents the physical topology." + ], + "image_footnote": [], + "bbox": [ + 96, + 263, + 715, + 381 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2018708943c56b0aae8d52505b46ebc1f9098223d66fdeaea1080aa0cc810b3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 80, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "represents the number of hyper-edges. Given $v \\in \\mathcal{V}_{\\mathcal{H}}$ and $e \\in \\mathcal{E}_{\\mathcal{H}}$ , the values of incidence matrix be determined by:", + "bbox": [ + 89, + 459, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh (e, v) = \\left\\{ \\begin{array}{l l} 1, & v \\in e, \\\\ 0, & v \\notin e \\end{array} \\right.. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 497, + 482, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the normal graph convolution, it also needs to normalise the hyper-graph to modulate the aggregated features. The degree of hyper-graph consists of the degree of joints and the degree of hyper-edges. The degree of joints is represented by the sum of weights of all joints contained in each hyper-edge. Given $v \\in \\mathcal{V}_{\\mathcal{H}}$ , it can be described as follows:", + "bbox": [ + 89, + 546, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd (v) = \\sum_ {e \\in \\mathcal {E} _ {\\mathcal {H}}} w (e) h (v, e), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 648, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use a diagonal matrix $D_{v} \\in \\mathbb{R}^{N \\times N}$ to represent the degree of joints. $W \\in \\mathbb{R}^{M \\times M}$ represents the weight matrix of hyper-edges, which is formulated as a diagonal matrix. The degree of hyper-edges represents the sum of the number of joints contained in each hyper-edge. Given $e \\in \\mathcal{E}_{\\mathcal{H}}$ , it can be described as follows:", + "bbox": [ + 89, + 686, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd (e) = \\sum_ {v \\in \\mathcal {V} _ {\\mathcal {H}}} h (v, e), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 786, + 482, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use a diagonal matrix $D_{e} \\in \\mathbb{R}^{M \\times M}$ to represent the degree of hyper-edges. Therefore, the normalisation of a hyper-graph is formulated as follows:", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {H} = D _ {v} ^ {- 1} H W D _ {e} ^ {- 1} H ^ {T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 882, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d545ccc8059f8da246488210c60d55b923c118918b9210c3bea3f832eee6557a.jpg", + "image_caption": [ + "Figure 3. Instance of the incidence matrix construction by A-NHG with hyper-parameter $K = 3$ . $N$ represents the number of joints. (a) represents the hyper-edge sets are null. (b) represents that measure the distance between 1-st joint and other joints. (c) represents that retrain $K$ -nearest hyper-edges which contain 1-st joint. (d) represents the final state." + ], + "image_footnote": [], + "bbox": [ + 531, + 476, + 897, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{H} \\in \\mathbb{R}^{N \\times N}$ represents the normalised incidence matrix for hyper-graph convolution.", + "bbox": [ + 511, + 806, + 903, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Non-uniform Hyper-graph", + "text_level": 1, + "bbox": [ + 511, + 848, + 828, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Constructing the hyper-graph incidence matrix is critical to understanding human actions. As shown in Figure 2 (b), we", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12651", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "design an Adaptive Non-uniform Hyper-graph (A-NHG) Construction.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the general features $\\mathbf{X} \\in \\mathbb{R}^{N \\times C}$ of spatial skeleton joints, $N$ represent the number of joints, $C$ represent the channels. Our A-NHG construct the hyper-edge with each joint as the center of mass, and a total of $N$ hyper-edges are constructed. A-NHG utilises the Euclidean distance to measure the joint difference.", + "bbox": [ + 89, + 121, + 483, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use one mapping function $\\Phi \\in \\mathbb{R}^{C\\times C_h}$ to embed the features into the subspace for constructing the hypergraph. $C_h$ represents the hidden channels of the mapping subspace. This design enables preserving the original features to reflect their spatial relevance. Based on $X_{H}\\in \\mathbb{R}^{N\\times C_{h}}$ , we define the set of joints $\\nu_{\\mathcal{H}} = \\{v_{1}\\in$ $\\mathbb{R}^{C_h},v_2\\in \\mathbb{R}^{C_h},\\dots,v_N\\in \\mathbb{R}^{C_h}\\}$ . Given $v_{i},v_{j}\\in \\mathcal{V}_{\\mathcal{H}}$ , each element $m_{i,j}$ in the distance matrix $M\\in \\mathbb{R}^{N\\times N}$ can be obtained as follows:", + "bbox": [ + 89, + 215, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i, j} = m _ {j, i} = \\left\\| v _ {i} - v _ {j} \\right\\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 368, + 482, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In principle, we need to transform the distance matrix into a probability incidence matrix, which determines whether the specific joints are contained with the same hyper-edges. In order to guarantee that the model is trainable, we use the softmax operation to assign the probabilities. However, directly using the softmax would inevitably result in each joint belonging to all hyper-edges to some extent.", + "bbox": [ + 89, + 401, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since this outcome is undesirable, we introduce constraints on each joint to limit the number of hyper-edges it can belong to, preventing excessive connections. For one joint, we retain only the $K$ nearest hyper-edges represented by joints in the semantic space for probability estimation. And $K$ is the hyper-parameter of A-NHG.", + "bbox": [ + 89, + 523, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the row vector $\\mathbf{m}_i$ in the distance matrix $M$ , which represents the distances between joint $i$ and the other joints. We select the indices of the minimum $K$ joints from $\\mathbf{m}_i$ and add them to the set $set_i$ . Based on $set_i$ , $h_{i,j}$ in the incidence matrix $H\\in \\mathbb{R}^{N\\times N}$ , can be calculated as follows:", + "bbox": [ + 89, + 616, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i, j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp (- m _ {i , j})}{\\sum_ {k \\in s e t _ {i}} \\exp (- m _ {i , k})}, & j \\in s e t _ {i}, \\\\ 0, & j \\notin s e t _ {i}. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 720, + 482, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to uniform hyper-graph (every hyper-edge contains the same number of joints), our approach limits the maximum number of every joint that can be contained by hyper-edges. Obviously, the number of joints contained in each hyper-edge constructed in this way is non-fixed, allowing the hyper-edges to capture more differentiated associations. The entire process for example with $K = 3$ is illustrated in the Figure 3.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-head Hyper-graph Convolution", + "text_level": 1, + "bbox": [ + 513, + 90, + 839, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accommodate the semantic information reflected by different group channels, we further propose the Multi-head Hyper-graph Convolution (M-HGC), as shown in Figure 2 (a).", + "bbox": [ + 511, + 112, + 905, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use separate branches to independently perform hyper-graph convolution on the topologies represented by the multi-head hyper-graphs, thereby enhancing the computational efficiency of the Hyper-GCN. For instance, the features $\\mathbf{F}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times T \\times V}$ are divided into 8 separate branches along the channel dimension, which are processed by 8 M-HGC parallelly, delivering 8 separate hyper-graphs.", + "bbox": [ + 511, + 172, + 905, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we perform temporal pooling, and then the obtained spatial information is the evidence to determine the optimal hyper-graph construction. This can efficiently decouple the temporal and spatial clues. After performing temporal average-pooling, we can obtain $\\bar{\\mathbf{F}}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times V}$ .", + "bbox": [ + 511, + 279, + 903, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The hyper-graph is depicted by the incidence matrix and weight matrix. So, we use a separate mapping function for each head to embed the features into the subspace for constructing the hyper-graph. This design enables preserving the original features to reflect their spatial relevance. After that, we introduce the A-NHG to obtain the incidence matrix. For the weight matrix, we adopt MLP to measure each hyper-edge. The hyper-graph obtained operation can be formulated as follows:", + "bbox": [ + 511, + 354, + 905, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH = \\uplus_ {k = 1} ^ {8} \\mathrm {A} - \\mathrm {N H G} \\left(\\bar {\\mathbf {F}} _ {\\mathbf {i n}} ^ {k}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 494, + 903, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW = \\sigma_ {2} \\left(\\Psi_ {2} \\left(\\uplus_ {k = 1} ^ {8} \\sigma_ {1} \\left(\\Psi_ {1} ^ {k} \\left(\\overline {{\\mathbf {F}}} _ {\\mathbf {i n}} ^ {k}\\right)\\right)\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 516, + 903, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Psi_{1} \\in \\mathbb{R}^{C_{\\mathrm{in}} / 8 \\times C_{h}}$ , $\\Psi_{2} \\in \\mathbb{R}^{(8 \\times C_{h}) \\times 8}$ are the mapping functions, which are set as learnable parameters. $C_{h}$ represents the hidden channels of the mapping subspace. $k$ represents the channels $[(k - 1) \\times C_{\\mathrm{in}} / 8 + 1, \\dots, k \\times C_{\\mathrm{in}} / 8]$ of $\\overline{\\mathbf{F}}_{\\mathrm{in}}$ . $\\uplus$ represents the channel concatenation. $\\sigma_{1}, \\sigma_{2}$ denote the activation function LeakeyReLU and Tanh to obtain the weight of each hyper-edge, limiting the values with the range of $[-1, 1]$ . Based on $H \\in \\mathbb{R}^{8 \\times N \\times N}$ , $W \\in \\mathbb{R}^{8 \\times N}$ , normalise the hyper-graph by sec.3.2 to obtain the $\\hat{H} \\in \\mathbb{R}^{8 \\times N \\times N}$ .", + "bbox": [ + 511, + 540, + 905, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the hyper-graph, we incorporate the physical topology to emphasise the natural physical relations of a human being's skeleton. To achieve this objective, existing methods [3, 23, 36, 52] divide the physical topology of the human body into 3 subsets. They are represented by $S = \\{s_{id}, s_{cf}, s_{cp}\\}$ , where $s_{id}, s_{cf}$ and $s_{cp}$ denote the identity, centrifugal, and centripetal joint subsets. In our M-HGC, to ensure that the integrated physical topology in each head remains complete, we merge them into a single set. 8 heads are aggregated by channel concatenation after the hyper-graph convolutions, as shown in Figure 2 (a). The operation of M-HGC can be formulated as follows:", + "bbox": [ + 511, + 693, + 905, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {o u t}} = \\uplus_ {k = 1} ^ {8} \\left(\\hat {A} _ {k} + \\alpha \\cdot \\hat {H} _ {k}\\right) \\mathbf {F} _ {\\text {i n}} ^ {k} P _ {k}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 878, + 903, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12652", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{A}$ and $\\hat{H}$ denote the normalized physical adjacency matrix and normalized incidence matrix. $P$ is the learnable weight parameter for feature transform. $\\alpha$ is a learnable parameter for the topology fusion of each head.", + "bbox": [ + 89, + 89, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Virtual Connections", + "text_level": 1, + "bbox": [ + 89, + 160, + 282, + 174 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is worth emphasising that incorporating learnable joints among different samples is essential for enhancing the model capacity, as these learnable joints capture generalised features of human actions. This not only enriches the semantic information but also facilitates easier interaction connections among real joints. Therefore, we introduce the hyper-joints which are to participate in the hyper-graph convolution as shown in Figure 2 (a).", + "bbox": [ + 89, + 181, + 482, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The shape of hyper-joints is consistent with the features of physical joints $\\mathbf{F}_{\\mathbf{p}} \\in \\mathbb{R}^{C \\times T \\times V}$ , which is described as $\\mathbf{F}_{\\mathbf{h}} \\in \\mathbb{R}^{C \\times T \\times V_{h}}$ ( $V_{h}$ represents the number of hyper-joints). To ensure alignment in the temporal dimension, the hyper-joints are shared across each frame. Specifically, the learning of hyper-joints is supervised by the loss function, which guides these hyper-joints to support generalisable driven features embedded within large amounts of data. Furthermore, we set independent hyper-joints at each layer of Hyper-GCN, aiming to harmonise features at different depths. Typically, these hyper-joints are involved in spatial hyper-graph convolution rather than in temporal convolution.", + "bbox": [ + 89, + 303, + 482, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To diversify these hyper-joints, we propose the Divergence Loss for hyper-joints optimisation to mitigate their homogenisation. We adopt a cosine matrix $C \\in \\mathbb{R}^{V_h \\times V_h}$ to measure the differences between hyper-joints, which can be formulated as follows:", + "bbox": [ + 89, + 500, + 482, + 575 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC = \\frac {\\mathbf {F} _ {\\mathrm {h}} \\mathbf {F} _ {\\mathrm {h}} ^ {T}}{| | \\mathbf {F} _ {\\mathrm {h}} | | ^ {2}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 583, + 480, + 618 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, in Divergence Loss, we measure the differences of hyper-joints with the mean of the cosine matrix $C$ in each layer. Since the correlation of hyper-joints themselves cannot be optimised, we manually eliminate this part by minus the $V_{h}$ . The loss calculation can be formulated as:", + "bbox": [ + 89, + 627, + 482, + 703 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} (C) = \\frac {\\sum_ {i = 1} ^ {V _ {h}} \\sum_ {j = 1} ^ {V _ {h}} R e L U \\left(c _ {i , j}\\right) - V _ {h}}{V _ {h} \\left(V _ {h} - 1\\right)}, \\quad c _ {i, j} \\in C, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 712, + 480, + 768 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {C E} + \\frac {1}{L} \\sum_ {l = 1} ^ {L} L _ {h} \\left(C _ {l}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 773, + 482, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V_{h}$ represents the number of hyper-joints we introduced. $C_l$ represents the cosine matrix of the $l$ -th layer. $\\mathcal{L}_{\\mathrm{CE}}$ represents the cross-entropy loss. Additionally, we manually connect the hyper joints to all the physical joints, which optimises the topology from the datasets.", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Entire Architecture", + "text_level": 1, + "bbox": [ + 513, + 90, + 700, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyper-GCN consists of an embedding layer [6], 9 spatial-temporal convolution layers as shown in Figure 2 (c). Each layer consists of the proposed M-HGC and Multi Scale Temporal convolution (MS-TC) [3]. 9 Layers are categorised into 3 stages. In each stage, we introduce the dense connections to integrate deep and shallow features in each stage. Additionally, to further validate the potential of Hyper-GCN, we propose a base version and a large version. The channels in each stage are set to 128, 256, 256 for the base version and 128, 256, 512 for the large version.", + "bbox": [ + 511, + 112, + 906, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 277, + 629, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 301, + 614, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NTU-RGB+D 60 & 120 NTU-RGB+D 60 [35] is a large dataset widely used in skeleton-based human action recognition, which is categorised into 60 classes. NTU-RGB+D 120 [26] is an extension to 120 classes of NTU 60. 4 benchmarks recommended by the official are adopted: (1) NTU60-XSub, (2) NTU60-XView, (3) NTU120-XSub, (4) NTU120-Xset.", + "bbox": [ + 511, + 323, + 903, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Northwestern-UCLA. The Northwestern-UCLA (NWUCLA) dataset [46] contains 1494 video clips, which is categorised into 10 classes. It contains 3 different camera views and is performed by 10 actors.", + "bbox": [ + 511, + 430, + 903, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 500, + 730, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our training and evaluation stages are on a single GPU RTX 3090. In the training phase, Hyper-GCN is optimised by Stochastic Gradient Descent (SGD) with Nesterov momentum set at 0.9 and a weight decay at 0.0004. Our implementation uses label smooth cross-entropy loss with the Divergence Loss we proposed. We set a total of 140 epochs with the start 5 warm-up epochs. The initial learning rate is 0.05, which is reduced to 0.005 at epoch 110 and to 0.0005 at epoch 120.", + "bbox": [ + 511, + 521, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 511, + 667, + 836, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-stream ensemble proposed in [39] have been proven effective by most of the existing state-of-the-art methods. We also use a 4-stream ensemble to evaluate the base version and the large version of Hyper-GCN. The detailed results are reported in Table 1.", + "bbox": [ + 511, + 688, + 903, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the base version of Hyper-GCN comprehensively outperforms all GCN-based and HGCN-based (hyper-graph utilised) SOTA and surpasses the Transformer-based SOTA on the NTU120, with a most lightweight design. Moreover, when scaled to match the parameters of the lightest Transformer-based method, the large version of Hyper-GCN achieves 1-st place in 4 benchmarks and 2-nd place in 1 benchmark. This demonstrates the effectiveness and merits of Hyper-GCN.", + "bbox": [ + 511, + 765, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12653", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1da4c5a2bf7868ad087bb0041fa18f3c5005a9a738d2d6c1e21c17333e4957c2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CategoryMethodsModalitiesParams (M)GFLOPsNTU-RGB+D 60NTU-RGB+D 120NW-UCLA (%)
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)
GCNST-GCN [52]J+B--81.588.370.773.2-
2s-AGCN [36]J+B--88.595.182.584.2-
DC-GCN+ADG [4]J+B+JM+BM4.91.8390.896.686.588.195.3
MS-G3D [27]J+B+JM+BM2.85.2291.596.286.988.4-
MST-GCN [11]J+B+JM+BM12.0-91.596.687.588.8-
CTR-GCN [3]J+B+JM+BM1.51.9792.496.488.990.696.5
EfficientGCN-B4 [41]J+B+JM+BM2.015.2091.795.788.389.1-
InfoGCN [6]J+B+JM+BM1.61.8492.796.989.490.796.6
FR Head [56]J+B+JM+BM2.0-92.896.889.590.996.8
HD-GCN* [23]J+B+J' +B'1.71.7793.097.089.891.296.9
DS-GCN [51]J+B+JM+BM--93.197.589.290.3-
BlockGCN [57]J+B+JM+BM1.31.6393.197.090.391.596.9
TransformerDSTA-Net [38]J+B+JM+BM3.516.1889.595.786.689.0-
IIP-Transformer [47]J+B+JM+BM2.97.2089.595.789.990.9-
SkateFormer [8]J+B+JM+BM2.03.6293.597.889.891.498.3
HGCNHyper-GNN [13]J+B+JM+BM--89.595.7---
Selective-HCN [58]J+B+JM+BM--90.896.6---
DST-HCN [48]J+B+JM+BM3.52.9392.396.888.890.796.6
Ours (B)J+B+JM+BM1.11.6393.397.490.591.797.2
Ours (L)J+B+JM+BM2.32.8893.797.890.992.097.6
", + "bbox": [ + 101, + 87, + 897, + 345 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/16c7aea1e5872b4caac64b6aa9f9c50926a7c837748986f5c7e294076dc4f451.jpg", + "table_caption": [ + "Table 1. Comparison of Hyper-GCN with advanced solutions on NTU-RGB+D 60, NTU-RGB+D 120, and NW-UCLA datasets. For a fair comparison, we use a 4-streams ensemble to evaluate the performance of the mentioned methods. J, B, JB, and JM represent the joint, bone, joint motion, and bone motion. In addition, for HD-GCN [23], we use the results from 4-streams according to the original papers for a fair comparison. The bold font represents the best result and underline font represents the 2-nd best result." + ], + "table_footnote": [], + "table_body": "
NTU RGB+D 120 X-Sub (%)
KUniformNon-uniform
Baseline-84.784.7
w M-HGC386.4 (↑ 1.7)86.0 (↑ 1.3)
586.5 (↑ 1.9)86.2 (↑ 1.5)
786.3 (↑ 1.6)86.5 (↑ 1.8)
986.0 (↑ 1.3)86.7 (↑ 2.2)
1185.9 (↑ 1.2)86.4 (↑ 1.7)
", + "bbox": [ + 114, + 435, + 460, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation of the hyper-parameter $K$ in A-NHG. The Uniform represents K-uniform hyper-graph, in which one hyper-edge contains K joints. The non-uniform represents non-uniform hyper-graph, in which one joint is contained by $K$ hyper-edge.", + "bbox": [ + 89, + 571, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 660, + 243, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments with visualisations to analyse the effectiveness of our design. redAll the ablation study is evaluated with the joint modality of base version.", + "bbox": [ + 89, + 686, + 482, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper-graph Construction: The Non-uniform hypergraphs with hyper-parameter $K$ represent the maximum number of hyper-edges contained in each joint. They reflect the varying degrees of aggregation in the constructed topology. In general, it is common to adopt a uniform hypergraph with hyper-parameter $K$ . Therefore, to explore how to set the hyper-parameter $K$ for the adaptive hyper-graph in M-HGC, we design the ablation experiments to investigate the impact of hyper-parameter $K$ on the model's recognition ability, where both uniform and non-uniform hypergraphs are considered.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8ec9720ed8344ce63630c6b4ae80986a1ae8f38b9813b78118c15cd64f21be9.jpg", + "image_caption": [ + "Figure 4. Visualisation of hyper-graph in selected actions. For ease of observation, the joints contained in the selected hyper-edge are highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 526, + 441, + 906, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are reported in Table 2. In both uniform and non-uniform hyper-graphs, a very large $K$ means that more joints are contained in each hyper-edge. Even though these \"extra\" joints may be optimised for minimal weight, they are likely to act as \"noise\", disrupting the information inter", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12654", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/113ebd7d3b448f5d1863827e00043c445b6b469533c8b14a31c40bf096c5219c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NTU RGB+D 120 X-Sub (%)
w/o Lhw/ Lh
Baseline84.784.7
w/o M-HGCw/ 184.9 (↑ 0.2)84.9 (↑ 0.2)
w/ 384.9 (↑ 0.2)85.2 (↑ 0.5)
w/ 584.785.0 (↑ 0.3)
w M-HGCw/ 186.7 (↑ 2.0)86.7 (↑ 2.0)
w/ 386.6 (↑ 1.8)86.9 (↑ 2.2)
w/ 586.6 (↑ 1.8)86.8 (↑ 2.1)
", + "bbox": [ + 102, + 88, + 470, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablations on hyper joints. w/ $N$ represents the number of hyper joints in each layer. w/ $\\mathcal{L}_h$ represents the training with Divergence Loss. The hyper-parameter $K$ of M-HGC is set to 9.", + "bbox": [ + 89, + 239, + 482, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac8795de8fa7bceae5fa8234961cef74c34fc94ea442eaa851fa91f049d20355.jpg", + "image_caption": [ + "Figure 5. Comparison of the cosine matrices of hyper-joints. Part (a) and (b) represents w/o Divergence Loss and w/ Divergence Loss with 3 hyper-joints. It is calculated by Eqn. (12)." + ], + "image_footnote": [], + "bbox": [ + 99, + 318, + 478, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "action represented by the hyper-edge. As the $K$ decreases, it is difficult for the hyper-edge to represent complex joint combination patterns. Therefore, setting $K = 5$ in uniform and $K = 9$ in Non-uniform achieves the best performance. The non-uniform hyper-graph of A-NHG outperforms the uniform hyper-graph, as the combination of skeleton joints in different actions is diverse.", + "bbox": [ + 89, + 518, + 482, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, we visualise the hyper-graph inferred by Hyper-GCN on selected actions in Figure 4. The hypergraph construction focuses on the joints that are most relevant to the corresponding action category. For \"Kicking something\" as part (d), the left leg joints and the right hand joints are modelled as a hyper-edge. Because in this action, the person's right hand is previously extended forward, and the left leg swings relatively to the right hand. Similarly, in the action \"Stand up\" as shown in part (b), the hand and foot joints serve as fixed joints, while the remaining body joints move relative to them. In addition, \"Taking a selfie\" and \"Drinking water\" both focus on the main joints of movement with the above view as Figure 4 (a), (c).", + "bbox": [ + 89, + 626, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Virtual Connections: We conduct the ablation study to analyse the hyper joints and the Divergence Loss. The results are listed in Table 3. By comparison, the performance is not monotonically increasing with the number of hyper joints. The case of introducing only 3 hyper joints achieves", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bff5685e0e431a8763f1c18c1bb16666f8f38873140dbf55e8c2008dd9bcb0b.jpg", + "image_caption": [ + "Figure 6. t-SNE visualisation of Baseline and Hyper-GCN. Different colours represent different joints. Part (a) and (b) represent the output features in Baseline and Hyper-GCN." + ], + "image_footnote": [], + "bbox": [ + 526, + 97, + 702, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d671b0d9050a0917b58d1dc97e6cb0bf33377a7c33cc04845daa407fe0c2188a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 95, + 890, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the best performance. As our hyper joints are learned from a large amount of data, involving a large number of hyper joints can introduce redundant and ambiguous clues, degrading the performance.", + "bbox": [ + 511, + 281, + 903, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, we visualise the cosine matrices of HyperGCN, as shown in Figure 5. Clearly, in the absence of the Divergence Loss, the homogenisation of the learned hyper-joints is severe. This impedes the ability of hyper-joints to represent the generalised features of human actions. The observation further validates the effectiveness of the Divergence Loss for hyper-joint optimisation, supporting our viewpoint that generalised features with a certain degree of differentiation are needed to participate in hyper-graph convolution.", + "bbox": [ + 511, + 343, + 905, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Hyper-GCN: To further validate that Hyper-GCN increases the efficiency of information interaction, we perform t-SNE projections on the output features in last layer between Baseline and Hyper-GCN, as shown in Figure 6 for comparison. Notice that the semantic information represented by the last Hyper-GCN layer is very convergent. This represents that the semantic information of each joint is adequately conveyed in M-HGC. The joints with similar semantics can prevent sacrificing the information after global average pooling.", + "bbox": [ + 511, + 494, + 905, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 662, + 633, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an Adaptive Hyper-graph Convolutional Network for skeleton-based human action recognition. To exploit the implicit topology of multivariate synergy between joints, we introduce the Adaptive Nonuniform Hyper-graph, the Multi-head Hyper-graph Convolution and virtual connections. We carry out experiments on the dataset NTU-RGB+D 60 & 120, NW-UCLA to validate the effectiveness of Hyper-GCN. The experimental analysis verifies that our design can improve the recognition performance. To the best of our knowledge, Hyper-GCN achieves the SOTA performance on 3 public datasets. The involvement of adaptive non-uniform hypergraph modelling indeed extends existing GCN-based action recognition paradigms.", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12655", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment This work was supported in part by the National Natural Science Foundation of China (62020106012, 62106089, 62332008, 62336004), the Engineering and Physical Sciences Research Council (EPSRC), U.K. (Grants EP/V002856/1, and EP/T022205/1), and the Fundamental Research Funds for the Central Universities (JUSRP202504007).", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 213, + 187, + 228 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tasweer Ahmad, Lianwen Jin, Luojun Lin, and GuoZhi Tang. Skeleton-based action recognition using sparse spatiotemporal gcn with edge effective resistance. Neurocomputing, 423:389-398, 2021. 3", + "[2] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2", + "[3] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, pages 13339-13348, 2021. 1, 2, 3, 5, 6, 7", + "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, page 536-553, Berlin, Heidelberg, 2020. Springer-Verlag. 3, 7", + "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In CVPR, pages 180-189, 2020. 3", + "[6] Hyung-Gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, pages 20154-20164, 2022. 1, 2, 3, 6, 7", + "[7] Guilhem Chéron, Ivan Laptev, and Cordelia Schmid. P-cnn: Pose-based cnn features for action recognition. In ICCV, pages 3218-3226, 2015. 2", + "[8] Jeonghyeok Do and Munchurl Kim. Skateformer: skeletal-temporal transformer for human action recognition. In ECCV, pages 401-420. Springer, 2024. 3, 7", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1", + "[11] Dong Feng, Zhongcheng Wu, Jun Zhang, and Tingting Ren. Multi-scale spatial temporal graph neural network for skeleton-based action recognition. IEEE Access, 9:58256-58265, 2021. 3, 7", + "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. TPAMI, 45(3):3181-3199, 2023. 2, 3" + ], + "bbox": [ + 93, + 238, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Xiaoke Hao, Jie Li, Yingchun Guo, Tao Jiang, and Ming Yu. Hypergraph neural network for skeleton-based action recognition. TIP, 30:2263-2275, 2021. 2, 3, 7", + "[14] Ruijie Hou, Zhihao Wang, Ruimin Ren, Yang Cao, and Zhao Wang. Multi-channel network: Constructing efficient gcn baselines for skeleton-based action recognition. Computers & Graphics, 110:111-117, 2023. 3", + "[15] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Part-level graph convolutional network for skeleton-based action recognition. In AAAI, pages 11045-11052, 2020. 3", + "[16] Cloe Huesser, Simon Schubiger, and Arzu Öltekin. Gesture interaction in virtual reality: A low-cost machine learning system and a qualitative assessment of effectiveness of selected gestures vs. gaze and controller interaction. In International Conference on Human-Computer Interaction, page 151–160, Berlin, Heidelberg, 2021. Springer-Verlag. 1", + "[17] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. TPAMI, 35(1):221-231, 2012. 2", + "[18] Jianwen Jiang, Yuxuan Wei, Yifan Feng, Jingxuan Cao, and Yue Gao. Dynamic hypergraph neural networks. In IJCAI, pages 2635-2641, 2019. 2", + "[19] George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar. Multilevel hypergraph partitioning: Application in vlsi domain. In Proceedings of the 34th annual Design Automation Conference, pages 526-529, 1997. 2", + "[20] Matthew Korban and Xin Li. Ddgen: A dynamic directed graph convolutional network for action recognition. In ECCV, page 761-776, Berlin, Heidelberg, 2020. Springer-Verlag. 3", + "[21] Bokyung Lee, Michael Lee, Pan Zhang, Alexander Tessier, and Azam Khan. Semantic human activity annotation tool using skeletonized surveillance videos. In Proceedings of the 2019 ACM International Symposium on Wearable Computers, pages 312-315, 2019. 1", + "[22] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In ICCV, pages 10221-10230, 2023. 3", + "[23] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In ICCV, pages 10410-10419, 2023. 2, 3, 5, 7", + "[24] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3", + "[25] Jun Liu, Amir Shahroudy, Dong Xu, Alex C. Kot, and Gang Wang. Skeleton-based action recognition using spatiotemporal LSTM network with trust gates. TPAMI, 40(12): 3007-3021, 2018. 1", + "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 42(10):2684-2701, 2020. 1, 2, 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12656", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, pages 140-149, 2020. 3, 7", + "[28] Ashish S. Nikam and Aarti G. Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 Online International Conference on Green Engineering and Technologies, pages 1-5, 2016. 1", + "[29] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory. In International Conference on Artificial Intelligence in Information and Communication, pages 273-278, 2021.", + "[30] Batyrkhan Omarov, Sergazy Narynov, Zhandos Zhumanov, Aidana Gumar, and Mariyam Khassanova. A skeleton-based approach for campus violence detection. Computers, Materials & Continua, 72(1), 2022. 1", + "[31] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In European Conference on Computer Vision, pages 605-622. Springer, 2022. 3", + "[32] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In Pattern recognition. ICPR international workshops and challenges: virtual event, January 10–15, 2021, Proceedings, Part III, pages 694–701. Springer, 2021. 2, 3", + "[33] Helei Qiu, Biao Hou, Bo Ren, and Xiaohua Zhang. Spatiotemporal tuples transformer for skeleton-based action recognition. arXiv preprint arXiv:2201.02849, 2022. 3", + "[34] Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision. Visual Intelligence, 2(1):27, 2024. 1", + "[35] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 1, 2, 6", + "[36] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12018-12027, 2019. 2, 3, 5, 7", + "[37] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In CVPR, pages 7904-7913, 2019. 3", + "[38] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Decoupled spatial-temporal attention network for skeleton-based action-gesture recognition. In Proceedings of the Asian conference on computer vision, 2020. 2, 3, 7", + "[39] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. TIP, 29:9532-9545, 2020. 6", + "[40] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "NIPS, page 568-576, Cambridge, MA, USA, 2014. MIT Press. 2", + "[41] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. TPAMI, 45(2):1474-1488, 2022. 3, 7", + "[42] Ahmed Taha, Hala H Zayed, ME Khalifa, and El-Sayed M El-Horbaty. Skeleton-based human activity recognition for video surveillance. International Journal of Scientific & Engineering Research, 6(1):993-1004, 2015. 1", + "[43] Jianwei Tang, Jieming Wang, and Jian-Fang Hu. Predicting human poses via recurrent attention network. Visual Intelligence, 1(1):18, 2023. 1", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[45] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In ICCV, pages 4041-4049, 2015. 1", + "[46] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning, and recognition. In CVPR, page 2649-2656, USA, 2014. IEEE Computer Society. 2, 6", + "[47] Qingtian Wang, Shuze Shi, Jiabin He, Jianlin Peng, Tingxi Liu, and Renliang Weng. Iip-transformer: Intra-inter-part transformer for skeleton-based action recognition. In BigData, pages 936-945, 2023. 2, 3, 7", + "[48] Shengqin Wang, Yongji Zhang, Hong Qi, Minghao Zhao, and Yu Jiang. Dynamic spatial-temporal hypergraph convolutional network for skeleton-based action recognition. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2147-2152. IEEE, 2023. 7", + "[49] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1", + "[50] Cong Wu, Xiao-Jun Wu, and Josef Kittler. Graph2net: Perceptually-enriched graph learning for skeleton-based action recognition. TCSVT, 32(4):2120-2132, 2022. 3", + "[51] Jianyang Xie, Yanda Meng, Yitian Zhao, Anh Nguyen, Xiaoyun Yang, and Yalin Zheng. Dynamic semantic-based spatial graph convolution network for skeleton-based human action recognition. AAAI, 38(6):6225-6233, 2024. 1, 3, 7", + "[52] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2, 3, 5, 7", + "[53] Hengqi Zhang, Hua Geng, and Geng Yang. Two-stream transformer encoders for skeleton-based action recognition. In International Conference on Computing, Control and Industrial Engineering, pages 272-281. Springer, 2021. 2, 3", + "[54] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In CVPR, pages 1109-1118, 2020. 3", + "[55] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12657", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In CVPR, pages 10608-10617, 2023. 7", + "[57] Yuxuan Zhou, Xudong Yan, Zhi-Qi Cheng, Yan Yan, Qi Dai, and Xian-Sheng Hua. Blockgen: Redefined topology awareness for skeleton-based action recognition. In CVPR, pages 2049-2058, 2024. 2, 3, 7", + "[58] Yiran Zhu, Guangji Huang, Xing Xu, Yanli Ji, and Fumin Shen. Selective hypergraph convolutional networks for skeleton-based action recognition. In ICMR, page 518-526, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 7" + ], + "bbox": [ + 89, + 90, + 482, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12658", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_model.json b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..127ee4348ac0061db71c73c8000083f0e63db50a --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.826, + 0.177 + ], + "angle": 0, + "content": "Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.203, + 0.777, + 0.24 + ], + "angle": 0, + "content": "Youwei Zhou\\(^{1}\\), Tianyang Xu\\(^{1*}\\), Cong Wu\\(^{1}\\), Xiaojun Wu\\(^{1}\\), Josef Kittler\\(^{2}\\) \n\\(^{1}\\)Jiangnan University, China \\(^{2}\\)University of Surrey, UK" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.242, + 0.636, + 0.257 + ], + "angle": 0, + "content": "youwei_zhou@stu.jiangnan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.26, + 0.712, + 0.275 + ], + "angle": 0, + "content": "{tianyang_xu,Congwu，wu_xiaojun}@jiangnan.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.278, + 0.593, + 0.292 + ], + "angle": 0, + "content": "j.kittler@surrey.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.485, + 0.751 + ], + "angle": 0, + "content": "The shared topology of human skeletons motivated the recent investigation of graph convolutional network (GCN) solutions for action recognition. However, most of the existing GCNs rely on the binary connection of two neighboring vertices (joints) formed by an edge (bone), overlooking the potential of constructing multi-vertex convolution structures. Although some studies have attempted to utilize hyper-graphs to represent the topology, they rely on a fixed construction strategy, which limits their adaptivity in uncovering the intricate latent relationships within the action. In this paper, we address this oversight and explore the merits of an adaptive hyper-graph convolutional network (HyperGCN) to achieve the aggregation of rich semantic information conveyed by skeleton vertices. In particular, our Hyper-GCN adaptively optimises the hyper-graphs during training, revealing the action-driven multi-vertex relations. Besides, virtual connections are often designed to support efficient feature aggregation, implicitly extending the spectrum of dependencies within the skeleton. By injecting virtual connections into hyper-graphs, the semantic clues of diverse action categories can be highlighted. The results of experiments conducted on the NTU-60, NTU-120, and NW-UCLA datasets demonstrate the merits of our Hyper-GCN, compared to the state-of-the-art methods. The code is available at https://github.com/6UOON9/Hyper-GCN." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.779, + 0.222, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Skeleton-based human action recognition is a popular research topic in artificial intelligence, with practical applications in video understanding, video surveillance, human-computer interaction, robot vision, VR and AR [16, 21, 28-30, 42, 49, 49]. In general, a skeleton sequence contains" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.338, + 0.903, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Figure 1. Illustration of our Hyper-GCN. In (a), dotted lines represent the virtual connections. Each coloured part containing multiple joints represents a hyper-graph with hyper-edges. (b) provides an intuitive comparison between normal graph convolution and hyper-graph convolution operations with the same connectivity degree." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.907, + 0.868 + ], + "angle": 0, + "content": "a series of 2D or 3D coordinates, which can easily be collected by low-cost depth sensors or obtained by video-based pose estimation algorithms [26, 34, 35, 43]. Compared to RGB and optical flow images, skeleton data, which represents the basic physical structure of a human being, is of lower dimension, conveying human action with higher efficiency. Moreover, it is robust to illumination changes and scene variations. For these reasons, the adoption of this structural data is very popular in skeleton-based action recognition [3, 6, 10, 25, 35, 45, 51]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To facilitate skeleton-based action recognition, both Recurrent Neural Networks (RNNs) [10, 25, 35, 45] and Con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12648" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.379 + ], + "angle": 0, + "content": "volitional Neural Networks (CNNs) [7, 17, 40] have been well explored. However, RNNs themselves cannot depict the intrinsic skeleton topology. The learned filters of CNNs, on the other hand, neglect the spatio-temporal structure of the skeleton. Drawing on these observations, recent studies have focused on how to model the skeleton topology directly. In principle, the physical topology of human joints and bones can be consistently represented by a graph. Accordingly, the graph convolutional network (GCN) is typically introduced to aggregate feature information conveyed by skeleton joints [23, 52, 57]. In terms of data representations, skeleton joints are also one-dimensional sequence data, similar to language. Therefore, with the rise of recent self-attention techniques [9, 44], several Transformer-based action recognition solutions [32, 38, 47, 53] have been released. Transformer considers skeletal joints as tokens and uses attention maps to reflect the topology. These methods have achieved excellent performance but sacrifice much higher GFLOPS and parameters than GCN." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.382, + 0.483, + 0.669 + ], + "angle": 0, + "content": "In general, two neighbouring joints can pass messages through their shared bone. In graph terms, this corresponds to the exchange of information between two vertices along their connecting edge. Besides the physical skeleton topology, attempts have been made to explore the implicit relationships among joints [3, 6, 23, 36], suggesting different variants of skeleton topology. In principle, these approaches assume a binary connection between each connected vertex pair. Mathematically, the constructed topology in the adjacency matrix is represented by a normal graph. However, human actions are jointly defined by several joints. Hence, human actions encompass not only binary relations between vertex pairs but also multi-joint relationships. For instance, the action primitive starting running is manifest in the human raising the left hand while the right leg steps forward, as shown in Figure 1 (a). The binary connections are not sufficient to capture the synergistic interaction of multiple joints. This strongly argues for constructing feature aggregation paths involving multiple vertices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Accordingly, we propose to construct a hyper-graph to depict the skeleton topology and take advantage of the outstanding performance of hyper-graph analysis techniques [2, 12, 18, 19]. The hyper-graph topology involves multiple node connections, rather than binary connections of a normal graph. As shown in Figure 1 (a), in a hyper-graph, a hyper-edge can link more than two vertices. This linking has the capacity to represent complex collaborative relations among human joints. As one hyper-edge associates multiple vertices, a single hyper-graph convolution enables aggregating all the features along the hyper-edges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "An illustration is provided in Figure 1 (b), where the normal graph and hyper-graph are presented to demonstrate their differences in passing the information conveyed by the vertex features. In the normal graph convolution, after" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "2-layer aggregation, the information of the green vertex is spread to 2 other vertices. In contrast, in a hyper-graph convolution, the information of the green vertex spreads to all the vertices, creating an extended receptive field. Theoretically, by modelling a skeleton using a hyper-graph for the aggregation of joint information, improved efficiency can be obtained during learning the action semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.365 + ], + "angle": 0, + "content": "However, the key to achieving enhanced information aggregation by hyper-graph convolution lies in the conformity of the hyper-graph structure to the real topology of human action. Several studies [13, 58] attempt to adopt a hypergraph to model the skeleton topology, mostly based on superficial prior knowledge and an unfounded definition of the structure of hyper-graph. Distinguishing our approach from the existing hyper-graph-based methods, we design an adaptive solution for constructing the hyper-graph with virtual connections. It can construct a Non-uniform hypergraph specific to each individual action type." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.727 + ], + "angle": 0, + "content": "Besides the information aggregation issue, the capacity of an action recognition system is also constrained by its input features. In general, the number of skeleton joints is fixed in existing benchmarks, e.g., 25 for NTU-120 [26]. The entire recognition process relies on a successful capture of the interactions among the skeleton joints. In the domain of artistic puppetry, the history of driving actions, such as Shadow Play and Marionette goes back more than 2000 years. This kind of art form provides an inspiration for involving additional 'hyper joints' which can drive or facilitate a better communication between existing joints. The underlying spirit is to alleviate the pressure on the real joints to store and transfer complex semantics. Jointly with the hyper joints, real joints can focus more on storing neighbouring joint features and relegating the task of transferring the global information to the hyper joints. Interestingly, the class token in existing Transformers can also be considered as a hyper-token [9, 44]. As shown in Figure 1, the skeleton is described as a marionette, where the actions are \"controlled\" by connecting hyper joints to the real joints. This suggests that the hyper joints are not only able to capture the representation information of human action, but also reveal the implicit information between physically connected joints as hints for recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.905, + 0.848 + ], + "angle": 0, + "content": "By endowing an adaptive non-uniform hyper-graph with hyper joints, virtual connections are created to perform comprehensive hyper-graph convolutions. We construct Hyper-GCN based on the above design principles. Extensive experiments on 3 datasets, NTU RGB+D 60 [35], NTU RGB+D 120 [26], and NW-UCLA [46], are conducted for evaluation. The results validate the merits of our proposed Hyper-GCN. The main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.85, + 0.905, + 0.866 + ], + "angle": 0, + "content": "- An adaptive non-uniform hyper-graph to represent the hu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.875, + 0.863, + 0.887 + ], + "angle": 0, + "content": "1https://en.wikipedia.org/wiki/Shadow Playing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.854, + 0.9 + ], + "angle": 0, + "content": "2https://en.wikipedia.org/wiki/Marionette" + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.875, + 0.863, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12649" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.136 + ], + "angle": 0, + "content": "man skeleton topology. Compared with a fixed hypergraph, the constructed topology is more action-specific, thus boosting the discrimination." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.137, + 0.483, + 0.182 + ], + "angle": 0, + "content": "- The injection of virtual hyper joints, enriching the connectivity of the physical joints, from a global semantic perspective." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.183, + 0.484, + 0.243 + ], + "angle": 0, + "content": "- As the processing architecture, we propose a hyper-graph convolution network (Hyper-GCN). The SOTA performance achieved on three public datasets demonstrates the merits of the Hyper-GCN and virtual connection designs." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.137, + 0.484, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.26, + 0.232, + 0.275 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.432, + 0.302 + ], + "angle": 0, + "content": "2.1. Graph Topology for Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.483, + 0.367 + ], + "angle": 0, + "content": "A graph can represent a human skeleton, preserving the joint relationships via predefined edges. To aggregate semantics, GCNs [1, 4, 11, 20, 24, 37, 54, 57] and Transformers [8, 32, 38, 53, 55] have been well studied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.488 + ], + "angle": 0, + "content": "For GCN-based approaches, ST-GCN [52] proposes to represent the topology with an adjacency matrix of 3 subsets by modeling the spatio-temporal relevance. Similarly, 2s-AGCN [36], InfoGCN [6], and DS-GCN [51] use a self-attentive mechanism to learn the topology from joint pairs. Besides the intuitive spatio-temporal dimensions, CTR-GCN [3] proposes to learn the topology for channels to refine the skeleton." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.58 + ], + "angle": 0, + "content": "For Transformer-based solutions, IIP-Transformer [47] adopt a self-attention mechanism to establish the intra-part and the inter-part joints. IGFormer [31] learns the topology between persons at both semantic and distance levels. STTFormer [33] proposed structure can capture the correlation between joints in consecutive frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.747 + ], + "angle": 0, + "content": "All the above two paradigms construct the topology from joint pairs, which rely only on the binary relation between two vertices. In their implementation stage, the adjacency matrix or attention map is used to represent the topology. In this case, high-order information among joints is not taken into consideration, neglecting the collaborative power among multiple joints. Though hyper-graph is considered to construct topology using multivariate joint relationships, current [13, 58] solutions manually set the hyper-graphs, which greatly relies on human experience, sacrificing the adaptability of graph learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.483, + 0.774 + ], + "angle": 0, + "content": "2.2. Feature Configurations for Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.825 + ], + "angle": 0, + "content": "It has been observed that the input features play the most essential role in delivering high recognition accuracy [1, 5, 15, 27, 31, 33, 41, 47] in GCNs and Transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Drawing on this, Graph2Net [50] proposes to extract local and global spatial features jointly. CTR-GCN [3] uses multi-scale temporal convolution to extract temporal features. While HD-GCN [23] introduce the hierarchical edge convolution to key edge features. To extend the perception" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "field, STC-Net [22] uses the dilated kernels for graph convolution to capture the features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Similarly, IG-Transformer [31] aggregate the features of two persons as one to mine the semantic information. STST [55] split the coordinate features, semantic features and temporal features into three kinds of tokens. TSTE [53] extracting and merging the spatial and temporal features as input to transformer encoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.333 + ], + "angle": 0, + "content": "After all, existing methods [6, 14, 32, 36, 38, 51, 52] receive the input features from real skeleton joints. However, each skeleton joint acts as an information carrier during forward passing, which is required to deliver both local context and global semantics. Given a fixed representative capacity (number of channels), we believe it is necessary to involve additional virtual joints to balance the pressure of storing local and global information." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.348, + 0.622, + 0.365 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.373, + 0.655, + 0.387 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.545 + ], + "angle": 0, + "content": "In general, the input features in graph convolution are represented by \\(\\mathbf{F}_{\\mathrm{in}}\\in \\mathbb{R}^{C\\times T\\times V}\\), where \\(C\\) represents the number of channels in the feature maps. Given the topology of a human skeleton, we usually define the graph \\(\\mathcal{G} = (\\mathcal{V},\\mathcal{E})\\), where \\(\\mathcal{V}\\) represents the set of joints and \\(\\mathcal{E}\\) represents the set of edges between joints. For the set of edges \\(\\mathcal{E}\\), it is formulated as an adjacent matrix \\(A\\in \\mathbb{R}^{N\\times N}\\), where \\(N\\) represents the number of human joints. The normalised adjacency matrix is represented by \\(\\hat{A}\\in \\mathbb{R}^{N\\times N}\\). The normalisation operation is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.555, + 0.905, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\hat {A} = \\Lambda^ {- \\frac {1}{2}} A \\Lambda^ {- \\frac {1}{2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.587, + 0.906, + 0.632 + ], + "angle": 0, + "content": "where \\(\\Lambda\\) represents the diagonal matrix stored with the degree of every joint. The entire normal graph convolution process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.644, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {o u t}} = \\sigma (\\hat {A} \\mathbf {F} _ {\\text {i n}} P), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.672, + 0.906, + 0.717 + ], + "angle": 0, + "content": "where \\(P \\in \\mathbb{R}^{C \\times C'}\\) is the learnable parameters, representing the feature transformation patterns in the feature space. \\(\\sigma\\) denotes the non-linear activation function ReLU." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.728, + 0.651, + 0.744 + ], + "angle": 0, + "content": "3.2. Hyper-graph" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Here, we use \\(\\mathcal{G}_{\\mathcal{H}} = (\\mathcal{V}_{\\mathcal{H}},\\mathcal{E}_{\\mathcal{H}},\\mathcal{W}_{\\mathcal{H}})\\) to define the spatial hyper-graph [12] with human skeleton. \\(\\mathcal{V}_{\\mathcal{H}}\\) and \\(\\mathcal{E}_{\\mathcal{H}}\\) follow similar definitions in the normal graph, which represent the set of joints and the set of hyper-edges. In addition, we introduce \\(\\mathcal{W}_{\\mathcal{H}}\\) to represent the weights of each hyper-edge. Since a hyper-edge contains multiple joints, the corresponding topology can no longer be simply represented by an adjacency matrix. Therefore, we introduce the incidence matrix \\(H\\in \\mathbb{R}^{N\\times M}\\) to describe the topology of each joint in the hyper-graph. \\(N\\) represents the number of joints and \\(M\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12650" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.08, + 0.716, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.264, + 0.716, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.081, + 0.891, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. The framework of our Hyper-GCN. Part (a) represents the Multi-head Hyper-graph Convolution (M-HGC) module. Part (b) represents the process of constructing an adaptive hyper-graph. Part (c) represents the architecture of Hyper-GCN. \\(\\mathbf{F}_{\\mathbf{P}}\\) represents the features of real joints. \\(\\mathbf{F}_h\\) represents the hyper-joints which are learnable parameters. \\(\\hat{A}\\) represents the physical topology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.492 + ], + "angle": 0, + "content": "represents the number of hyper-edges. Given \\(v \\in \\mathcal{V}_{\\mathcal{H}}\\) and \\(e \\in \\mathcal{E}_{\\mathcal{H}}\\), the values of incidence matrix be determined by:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.498, + 0.483, + 0.54 + ], + "angle": 0, + "content": "\\[\nh (e, v) = \\left\\{ \\begin{array}{l l} 1, & v \\in e, \\\\ 0, & v \\notin e \\end{array} \\right.. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Similar to the normal graph convolution, it also needs to normalise the hyper-graph to modulate the aggregated features. The degree of hyper-graph consists of the degree of joints and the degree of hyper-edges. The degree of joints is represented by the sum of weights of all joints contained in each hyper-edge. Given \\( v \\in \\mathcal{V}_{\\mathcal{H}} \\), it can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.65, + 0.483, + 0.682 + ], + "angle": 0, + "content": "\\[\nd (v) = \\sum_ {e \\in \\mathcal {E} _ {\\mathcal {H}}} w (e) h (v, e), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.779 + ], + "angle": 0, + "content": "where we use a diagonal matrix \\( D_{v} \\in \\mathbb{R}^{N \\times N} \\) to represent the degree of joints. \\( W \\in \\mathbb{R}^{M \\times M} \\) represents the weight matrix of hyper-edges, which is formulated as a diagonal matrix. The degree of hyper-edges represents the sum of the number of joints contained in each hyper-edge. Given \\( e \\in \\mathcal{E}_{\\mathcal{H}} \\), it can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.787, + 0.483, + 0.82 + ], + "angle": 0, + "content": "\\[\nd (e) = \\sum_ {v \\in \\mathcal {V} _ {\\mathcal {H}}} h (v, e), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.484, + 0.876 + ], + "angle": 0, + "content": "where we use a diagonal matrix \\( D_{e} \\in \\mathbb{R}^{M \\times M} \\) to represent the degree of hyper-edges. Therefore, the normalisation of a hyper-graph is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.883, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {H} = D _ {v} ^ {- 1} H W D _ {e} ^ {- 1} H ^ {T}, \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.477, + 0.898, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.785 + ], + "angle": 0, + "content": "Figure 3. Instance of the incidence matrix construction by A-NHG with hyper-parameter \\( K = 3 \\). \\( N \\) represents the number of joints. (a) represents the hyper-edge sets are null. (b) represents that measure the distance between 1-st joint and other joints. (c) represents that retrain \\( K \\)-nearest hyper-edges which contain 1-st joint. (d) represents the final state." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.905, + 0.84 + ], + "angle": 0, + "content": "where \\(\\hat{H} \\in \\mathbb{R}^{N \\times N}\\) represents the normalised incidence matrix for hyper-graph convolution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.83, + 0.866 + ], + "angle": 0, + "content": "3.3. Adaptive Non-uniform Hyper-graph" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Constructing the hyper-graph incidence matrix is critical to understanding human actions. As shown in Figure 2 (b), we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12651" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "design an Adaptive Non-uniform Hyper-graph (A-NHG) Construction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.214 + ], + "angle": 0, + "content": "Given the general features \\(\\mathbf{X} \\in \\mathbb{R}^{N \\times C}\\) of spatial skeleton joints, \\(N\\) represent the number of joints, \\(C\\) represent the channels. Our A-NHG construct the hyper-edge with each joint as the center of mass, and a total of \\(N\\) hyper-edges are constructed. A-NHG utilises the Euclidean distance to measure the joint difference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.351 + ], + "angle": 0, + "content": "We use one mapping function \\(\\Phi \\in \\mathbb{R}^{C\\times C_h}\\) to embed the features into the subspace for constructing the hypergraph. \\(C_h\\) represents the hidden channels of the mapping subspace. This design enables preserving the original features to reflect their spatial relevance. Based on \\(X_{H}\\in \\mathbb{R}^{N\\times C_{h}}\\), we define the set of joints \\(\\nu_{\\mathcal{H}} = \\{v_{1}\\in\\) \\(\\mathbb{R}^{C_h},v_2\\in \\mathbb{R}^{C_h},\\dots,v_N\\in \\mathbb{R}^{C_h}\\}\\). Given \\(v_{i},v_{j}\\in \\mathcal{V}_{\\mathcal{H}}\\), each element \\(m_{i,j}\\) in the distance matrix \\(M\\in \\mathbb{R}^{N\\times N}\\) can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.369, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nm _ {i, j} = m _ {j, i} = \\left\\| v _ {i} - v _ {j} \\right\\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.52 + ], + "angle": 0, + "content": "In principle, we need to transform the distance matrix into a probability incidence matrix, which determines whether the specific joints are contained with the same hyper-edges. In order to guarantee that the model is trainable, we use the softmax operation to assign the probabilities. However, directly using the softmax would inevitably result in each joint belonging to all hyper-edges to some extent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Since this outcome is undesirable, we introduce constraints on each joint to limit the number of hyper-edges it can belong to, preventing excessive connections. For one joint, we retain only the \\(K\\) nearest hyper-edges represented by joints in the semantic space for probability estimation. And \\(K\\) is the hyper-parameter of A-NHG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.692 + ], + "angle": 0, + "content": "Given the row vector \\(\\mathbf{m}_i\\) in the distance matrix \\(M\\), which represents the distances between joint \\(i\\) and the other joints. We select the indices of the minimum \\(K\\) joints from \\(\\mathbf{m}_i\\) and add them to the set \\(set_i\\). Based on \\(set_i\\), \\(h_{i,j}\\) in the incidence matrix \\(H\\in \\mathbb{R}^{N\\times N}\\), can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.722, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\nh _ {i, j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp (- m _ {i , j})}{\\sum_ {k \\in s e t _ {i}} \\exp (- m _ {i , k})}, & j \\in s e t _ {i}, \\\\ 0, & j \\notin s e t _ {i}. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In contrast to uniform hyper-graph (every hyper-edge contains the same number of joints), our approach limits the maximum number of every joint that can be contained by hyper-edges. Obviously, the number of joints contained in each hyper-edge constructed in this way is non-fixed, allowing the hyper-edges to capture more differentiated associations. The entire process for example with \\( K = 3 \\) is illustrated in the Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.84, + 0.108 + ], + "angle": 0, + "content": "3.4. Multi-head Hyper-graph Convolution" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.906, + 0.173 + ], + "angle": 0, + "content": "To accommodate the semantic information reflected by different group channels, we further propose the Multi-head Hyper-graph Convolution (M-HGC), as shown in Figure 2 (a)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.174, + 0.906, + 0.279 + ], + "angle": 0, + "content": "We use separate branches to independently perform hyper-graph convolution on the topologies represented by the multi-head hyper-graphs, thereby enhancing the computational efficiency of the Hyper-GCN. For instance, the features \\(\\mathbf{F}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times T \\times V}\\) are divided into 8 separate branches along the channel dimension, which are processed by 8 M-HGC parallelly, delivering 8 separate hyper-graphs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.905, + 0.355 + ], + "angle": 0, + "content": "Next, we perform temporal pooling, and then the obtained spatial information is the evidence to determine the optimal hyper-graph construction. This can efficiently decouple the temporal and spatial clues. After performing temporal average-pooling, we can obtain \\(\\bar{\\mathbf{F}}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.906, + 0.489 + ], + "angle": 0, + "content": "The hyper-graph is depicted by the incidence matrix and weight matrix. So, we use a separate mapping function for each head to embed the features into the subspace for constructing the hyper-graph. This design enables preserving the original features to reflect their spatial relevance. After that, we introduce the A-NHG to obtain the incidence matrix. For the weight matrix, we adopt MLP to measure each hyper-edge. The hyper-graph obtained operation can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.496, + 0.905, + 0.514 + ], + "angle": 0, + "content": "\\[\nH = \\uplus_ {k = 1} ^ {8} \\mathrm {A} - \\mathrm {N H G} \\left(\\bar {\\mathbf {F}} _ {\\mathbf {i n}} ^ {k}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.517, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\nW = \\sigma_ {2} \\left(\\Psi_ {2} \\left(\\uplus_ {k = 1} ^ {8} \\sigma_ {1} \\left(\\Psi_ {1} ^ {k} \\left(\\overline {{\\mathbf {F}}} _ {\\mathbf {i n}} ^ {k}\\right)\\right)\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.692 + ], + "angle": 0, + "content": "where \\(\\Psi_{1} \\in \\mathbb{R}^{C_{\\mathrm{in}} / 8 \\times C_{h}}\\), \\(\\Psi_{2} \\in \\mathbb{R}^{(8 \\times C_{h}) \\times 8}\\) are the mapping functions, which are set as learnable parameters. \\(C_{h}\\) represents the hidden channels of the mapping subspace. \\(k\\) represents the channels \\([(k - 1) \\times C_{\\mathrm{in}} / 8 + 1, \\dots, k \\times C_{\\mathrm{in}} / 8]\\) of \\(\\overline{\\mathbf{F}}_{\\mathrm{in}}\\). \\(\\uplus\\) represents the channel concatenation. \\(\\sigma_{1}, \\sigma_{2}\\) denote the activation function LeakeyReLU and Tanh to obtain the weight of each hyper-edge, limiting the values with the range of \\([-1, 1]\\). Based on \\(H \\in \\mathbb{R}^{8 \\times N \\times N}\\), \\(W \\in \\mathbb{R}^{8 \\times N}\\), normalise the hyper-graph by sec.3.2 to obtain the \\(\\hat{H} \\in \\mathbb{R}^{8 \\times N \\times N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.906, + 0.874 + ], + "angle": 0, + "content": "Besides the hyper-graph, we incorporate the physical topology to emphasise the natural physical relations of a human being's skeleton. To achieve this objective, existing methods [3, 23, 36, 52] divide the physical topology of the human body into 3 subsets. They are represented by \\( S = \\{s_{id}, s_{cf}, s_{cp}\\} \\), where \\( s_{id}, s_{cf} \\) and \\( s_{cp} \\) denote the identity, centrifugal, and centripetal joint subsets. In our M-HGC, to ensure that the integrated physical topology in each head remains complete, we merge them into a single set. 8 heads are aggregated by channel concatenation after the hyper-graph convolutions, as shown in Figure 2 (a). The operation of M-HGC can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.88, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {o u t}} = \\uplus_ {k = 1} ^ {8} \\left(\\hat {A} _ {k} + \\alpha \\cdot \\hat {H} _ {k}\\right) \\mathbf {F} _ {\\text {i n}} ^ {k} P _ {k}, \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.152 + ], + "angle": 0, + "content": "where \\(\\hat{A}\\) and \\(\\hat{H}\\) denote the normalized physical adjacency matrix and normalized incidence matrix. \\(P\\) is the learnable weight parameter for feature transform. \\(\\alpha\\) is a learnable parameter for the topology fusion of each head." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.283, + 0.175 + ], + "angle": 0, + "content": "3.5. Virtual Connections" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.304 + ], + "angle": 0, + "content": "It is worth emphasising that incorporating learnable joints among different samples is essential for enhancing the model capacity, as these learnable joints capture generalised features of human actions. This not only enriches the semantic information but also facilitates easier interaction connections among real joints. Therefore, we introduce the hyper-joints which are to participate in the hyper-graph convolution as shown in Figure 2 (a)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.483, + 0.499 + ], + "angle": 0, + "content": "The shape of hyper-joints is consistent with the features of physical joints \\(\\mathbf{F}_{\\mathbf{p}} \\in \\mathbb{R}^{C \\times T \\times V}\\), which is described as \\(\\mathbf{F}_{\\mathbf{h}} \\in \\mathbb{R}^{C \\times T \\times V_{h}}\\) (\\(V_{h}\\) represents the number of hyper-joints). To ensure alignment in the temporal dimension, the hyper-joints are shared across each frame. Specifically, the learning of hyper-joints is supervised by the loss function, which guides these hyper-joints to support generalisable driven features embedded within large amounts of data. Furthermore, we set independent hyper-joints at each layer of Hyper-GCN, aiming to harmonise features at different depths. Typically, these hyper-joints are involved in spatial hyper-graph convolution rather than in temporal convolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.576 + ], + "angle": 0, + "content": "To diversify these hyper-joints, we propose the Divergence Loss for hyper-joints optimisation to mitigate their homogenisation. We adopt a cosine matrix \\(C \\in \\mathbb{R}^{V_h \\times V_h}\\) to measure the differences between hyper-joints, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.584, + 0.482, + 0.619 + ], + "angle": 0, + "content": "\\[\nC = \\frac {\\mathbf {F} _ {\\mathrm {h}} \\mathbf {F} _ {\\mathrm {h}} ^ {T}}{| | \\mathbf {F} _ {\\mathrm {h}} | | ^ {2}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Specifically, in Divergence Loss, we measure the differences of hyper-joints with the mean of the cosine matrix \\(C\\) in each layer. Since the correlation of hyper-joints themselves cannot be optimised, we manually eliminate this part by minus the \\(V_{h}\\). The loss calculation can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.713, + 0.482, + 0.77 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} (C) = \\frac {\\sum_ {i = 1} ^ {V _ {h}} \\sum_ {j = 1} ^ {V _ {h}} R e L U \\left(c _ {i , j}\\right) - V _ {h}}{V _ {h} \\left(V _ {h} - 1\\right)}, \\quad c _ {i, j} \\in C, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.774, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {C E} + \\frac {1}{L} \\sum_ {l = 1} ^ {L} L _ {h} \\left(C _ {l}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\( V_{h} \\) represents the number of hyper-joints we introduced. \\( C_l \\) represents the cosine matrix of the \\( l \\)-th layer. \\( \\mathcal{L}_{\\mathrm{CE}} \\) represents the cross-entropy loss. Additionally, we manually connect the hyper joints to all the physical joints, which optimises the topology from the datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.701, + 0.105 + ], + "angle": 0, + "content": "3.6. Entire Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.265 + ], + "angle": 0, + "content": "Hyper-GCN consists of an embedding layer [6], 9 spatial-temporal convolution layers as shown in Figure 2 (c). Each layer consists of the proposed M-HGC and Multi Scale Temporal convolution (MS-TC) [3]. 9 Layers are categorised into 3 stages. In each stage, we introduce the dense connections to integrate deep and shallow features in each stage. Additionally, to further validate the potential of Hyper-GCN, we propose a base version and a large version. The channels in each stage are set to 128, 256, 256 for the base version and 128, 256, 512 for the large version." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.278, + 0.63, + 0.293 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.303, + 0.615, + 0.317 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.429 + ], + "angle": 0, + "content": "NTU-RGB+D 60 & 120 NTU-RGB+D 60 [35] is a large dataset widely used in skeleton-based human action recognition, which is categorised into 60 classes. NTU-RGB+D 120 [26] is an extension to 120 classes of NTU 60. 4 benchmarks recommended by the official are adopted: (1) NTU60-XSub, (2) NTU60-XView, (3) NTU120-XSub, (4) NTU120-Xset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.905, + 0.492 + ], + "angle": 0, + "content": "Northwestern-UCLA. The Northwestern-UCLA (NWUCLA) dataset [46] contains 1494 video clips, which is categorised into 10 classes. It contains 3 different camera views and is performed by 10 actors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.731, + 0.516 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Our training and evaluation stages are on a single GPU RTX 3090. In the training phase, Hyper-GCN is optimised by Stochastic Gradient Descent (SGD) with Nesterov momentum set at 0.9 and a weight decay at 0.0004. Our implementation uses label smooth cross-entropy loss with the Divergence Loss we proposed. We set a total of 140 epochs with the start 5 warm-up epochs. The initial learning rate is 0.05, which is reduced to 0.005 at epoch 110 and to 0.0005 at epoch 120." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.838, + 0.684 + ], + "angle": 0, + "content": "4.3. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.764 + ], + "angle": 0, + "content": "Multi-stream ensemble proposed in [39] have been proven effective by most of the existing state-of-the-art methods. We also use a 4-stream ensemble to evaluate the base version and the large version of Hyper-GCN. The detailed results are reported in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In summary, the base version of Hyper-GCN comprehensively outperforms all GCN-based and HGCN-based (hyper-graph utilised) SOTA and surpasses the Transformer-based SOTA on the NTU120, with a most lightweight design. Moreover, when scaled to match the parameters of the lightest Transformer-based method, the large version of Hyper-GCN achieves 1-st place in 4 benchmarks and 2-nd place in 1 benchmark. This demonstrates the effectiveness and merits of Hyper-GCN." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12653" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.088, + 0.898, + 0.346 + ], + "angle": 0, + "content": "
CategoryMethodsModalitiesParams (M)GFLOPsNTU-RGB+D 60NTU-RGB+D 120NW-UCLA (%)
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)
GCNST-GCN [52]J+B--81.588.370.773.2-
2s-AGCN [36]J+B--88.595.182.584.2-
DC-GCN+ADG [4]J+B+JM+BM4.91.8390.896.686.588.195.3
MS-G3D [27]J+B+JM+BM2.85.2291.596.286.988.4-
MST-GCN [11]J+B+JM+BM12.0-91.596.687.588.8-
CTR-GCN [3]J+B+JM+BM1.51.9792.496.488.990.696.5
EfficientGCN-B4 [41]J+B+JM+BM2.015.2091.795.788.389.1-
InfoGCN [6]J+B+JM+BM1.61.8492.796.989.490.796.6
FR Head [56]J+B+JM+BM2.0-92.896.889.590.996.8
HD-GCN* [23]J+B+J' +B'1.71.7793.097.089.891.296.9
DS-GCN [51]J+B+JM+BM--93.197.589.290.3-
BlockGCN [57]J+B+JM+BM1.31.6393.197.090.391.596.9
TransformerDSTA-Net [38]J+B+JM+BM3.516.1889.595.786.689.0-
IIP-Transformer [47]J+B+JM+BM2.97.2089.595.789.990.9-
SkateFormer [8]J+B+JM+BM2.03.6293.597.889.891.498.3
HGCNHyper-GNN [13]J+B+JM+BM--89.595.7---
Selective-HCN [58]J+B+JM+BM--90.896.6---
DST-HCN [48]J+B+JM+BM3.52.9392.396.888.890.796.6
Ours (B)J+B+JM+BM1.11.6393.397.490.591.797.2
Ours (L)J+B+JM+BM2.32.8893.797.890.992.097.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.356, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Table 1. Comparison of Hyper-GCN with advanced solutions on NTU-RGB+D 60, NTU-RGB+D 120, and NW-UCLA datasets. For a fair comparison, we use a 4-streams ensemble to evaluate the performance of the mentioned methods. J, B, JB, and JM represent the joint, bone, joint motion, and bone motion. In addition, for HD-GCN [23], we use the results from 4-streams according to the original papers for a fair comparison. The bold font represents the best result and underline font represents the 2-nd best result." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.436, + 0.462, + 0.562 + ], + "angle": 0, + "content": "
NTU RGB+D 120 X-Sub (%)
KUniformNon-uniform
Baseline-84.784.7
w M-HGC386.4 (↑ 1.7)86.0 (↑ 1.3)
586.5 (↑ 1.9)86.2 (↑ 1.5)
786.3 (↑ 1.6)86.5 (↑ 1.8)
986.0 (↑ 1.3)86.7 (↑ 2.2)
1185.9 (↑ 1.2)86.4 (↑ 1.7)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.572, + 0.483, + 0.63 + ], + "angle": 0, + "content": "Table 2. Ablation of the hyper-parameter \\( K \\) in A-NHG. The Uniform represents K-uniform hyper-graph, in which one hyper-edge contains K joints. The non-uniform represents non-uniform hyper-graph, in which one joint is contained by \\( K \\) hyper-edge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.245, + 0.677 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We conduct ablation experiments with visualisations to analyse the effectiveness of our design. redAll the ablation study is evaluated with the joint modality of base version." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Hyper-graph Construction: The Non-uniform hypergraphs with hyper-parameter \\( K \\) represent the maximum number of hyper-edges contained in each joint. They reflect the varying degrees of aggregation in the constructed topology. In general, it is common to adopt a uniform hypergraph with hyper-parameter \\( K \\). Therefore, to explore how to set the hyper-parameter \\( K \\) for the adaptive hyper-graph in M-HGC, we design the ablation experiments to investigate the impact of hyper-parameter \\( K \\) on the model's recognition ability, where both uniform and non-uniform hypergraphs are considered." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.442, + 0.907, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.748, + 0.907, + 0.79 + ], + "angle": 0, + "content": "Figure 4. Visualisation of hyper-graph in selected actions. For ease of observation, the joints contained in the selected hyper-edge are highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The results are reported in Table 2. In both uniform and non-uniform hyper-graphs, a very large \\( K \\) means that more joints are contained in each hyper-edge. Even though these \"extra\" joints may be optimised for minimal weight, they are likely to act as \"noise\", disrupting the information inter" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12654" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.472, + 0.23 + ], + "angle": 0, + "content": "
NTU RGB+D 120 X-Sub (%)
w/o Lhw/ Lh
Baseline84.784.7
w/o M-HGCw/ 184.9 (↑ 0.2)84.9 (↑ 0.2)
w/ 384.9 (↑ 0.2)85.2 (↑ 0.5)
w/ 584.785.0 (↑ 0.3)
w M-HGCw/ 186.7 (↑ 2.0)86.7 (↑ 2.0)
w/ 386.6 (↑ 1.8)86.9 (↑ 2.2)
w/ 586.6 (↑ 1.8)86.8 (↑ 2.1)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.24, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Table 3. Ablations on hyper joints. w/ \\(N\\) represents the number of hyper joints in each layer. w/ \\(\\mathcal{L}_h\\) represents the training with Divergence Loss. The hyper-parameter \\(K\\) of M-HGC is set to 9." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.319, + 0.48, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.446, + 0.483, + 0.489 + ], + "angle": 0, + "content": "Figure 5. Comparison of the cosine matrices of hyper-joints. Part (a) and (b) represents w/o Divergence Loss and w/ Divergence Loss with 3 hyper-joints. It is calculated by Eqn. (12)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.625 + ], + "angle": 0, + "content": "action represented by the hyper-edge. As the \\( K \\) decreases, it is difficult for the hyper-edge to represent complex joint combination patterns. Therefore, setting \\( K = 5 \\) in uniform and \\( K = 9 \\) in Non-uniform achieves the best performance. The non-uniform hyper-graph of A-NHG outperforms the uniform hyper-graph, as the combination of skeleton joints in different actions is diverse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Furthermore, we visualise the hyper-graph inferred by Hyper-GCN on selected actions in Figure 4. The hypergraph construction focuses on the joints that are most relevant to the corresponding action category. For \"Kicking something\" as part (d), the left leg joints and the right hand joints are modelled as a hyper-edge. Because in this action, the person's right hand is previously extended forward, and the left leg swings relatively to the right hand. Similarly, in the action \"Stand up\" as shown in part (b), the hand and foot joints serve as fixed joints, while the remaining body joints move relative to them. In addition, \"Taking a selfie\" and \"Drinking water\" both focus on the main joints of movement with the above view as Figure 4 (a), (c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Virtual Connections: We conduct the ablation study to analyse the hyper joints and the Divergence Loss. The results are listed in Table 3. By comparison, the performance is not monotonically increasing with the number of hyper joints. The case of introducing only 3 hyper joints achieves" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.098, + 0.703, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.097, + 0.892, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.211, + 0.906, + 0.254 + ], + "angle": 0, + "content": "Figure 6. t-SNE visualisation of Baseline and Hyper-GCN. Different colours represent different joints. Part (a) and (b) represent the output features in Baseline and Hyper-GCN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.905, + 0.342 + ], + "angle": 0, + "content": "the best performance. As our hyper joints are learned from a large amount of data, involving a large number of hyper joints can introduce redundant and ambiguous clues, degrading the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.493 + ], + "angle": 0, + "content": "In addition, we visualise the cosine matrices of HyperGCN, as shown in Figure 5. Clearly, in the absence of the Divergence Loss, the homogenisation of the learned hyper-joints is severe. This impedes the ability of hyper-joints to represent the generalised features of human actions. The observation further validates the effectiveness of the Divergence Loss for hyper-joint optimisation, supporting our viewpoint that generalised features with a certain degree of differentiation are needed to participate in hyper-graph convolution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Effectiveness of Hyper-GCN: To further validate that Hyper-GCN increases the efficiency of information interaction, we perform t-SNE projections on the output features in last layer between Baseline and Hyper-GCN, as shown in Figure 6 for comparison. Notice that the semantic information represented by the last Hyper-GCN layer is very convergent. This represents that the semantic information of each joint is adequately conveyed in M-HGC. The joints with similar semantics can prevent sacrificing the information after global average pooling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.663, + 0.634, + 0.679 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an Adaptive Hyper-graph Convolutional Network for skeleton-based human action recognition. To exploit the implicit topology of multivariate synergy between joints, we introduce the Adaptive Nonuniform Hyper-graph, the Multi-head Hyper-graph Convolution and virtual connections. We carry out experiments on the dataset NTU-RGB+D 60 & 120, NW-UCLA to validate the effectiveness of Hyper-GCN. The experimental analysis verifies that our design can improve the recognition performance. To the best of our knowledge, Hyper-GCN achieves the SOTA performance on 3 public datasets. The involvement of adaptive non-uniform hypergraph modelling indeed extends existing GCN-based action recognition paradigms." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12655" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgment This work was supported in part by the National Natural Science Foundation of China (62020106012, 62106089, 62332008, 62336004), the Engineering and Physical Sciences Research Council (EPSRC), U.K. (Grants EP/V002856/1, and EP/T022205/1), and the Fundamental Research Funds for the Central Universities (JUSRP202504007)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.214, + 0.188, + 0.229 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.239, + 0.482, + 0.294 + ], + "angle": 0, + "content": "[1] Tasweer Ahmad, Lianwen Jin, Luojun Lin, and GuoZhi Tang. Skeleton-based action recognition using sparse spatiotemporal gcn with edge effective resistance. Neurocomputing, 423:389-398, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.297, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[2] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.341, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[3] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, pages 13339-13348, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.399, + 0.483, + 0.454 + ], + "angle": 0, + "content": "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, page 536-553, Berlin, Heidelberg, 2020. Springer-Verlag. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In CVPR, pages 180-189, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.514, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[6] Hyung-Gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, pages 20154-20164, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[7] Guilhem Chéron, Ivan Laptev, and Cordelia Schmid. P-cnn: Pose-based cnn features for action recognition. In ICCV, pages 3218-3226, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[8] Jeonghyeok Do and Munchurl Kim. Skateformer: skeletal-temporal transformer for human action recognition. In ECCV, pages 401-420. Springer, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[11] Dong Feng, Zhongcheng Wu, Jun Zhang, and Tingting Ren. Multi-scale spatial temporal graph neural network for skeleton-based action recognition. IEEE Access, 9:58256-58265, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. TPAMI, 45(3):3181-3199, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.239, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[13] Xiaoke Hao, Jie Li, Yingchun Guo, Tao Jiang, and Ming Yu. Hypergraph neural network for skeleton-based action recognition. TIP, 30:2263-2275, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Ruijie Hou, Zhihao Wang, Ruimin Ren, Yang Cao, and Zhao Wang. Multi-channel network: Constructing efficient gcn baselines for skeleton-based action recognition. Computers & Graphics, 110:111-117, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Part-level graph convolutional network for skeleton-based action recognition. In AAAI, pages 11045-11052, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[16] Cloe Huesser, Simon Schubiger, and Arzu Öltekin. Gesture interaction in virtual reality: A low-cost machine learning system and a qualitative assessment of effectiveness of selected gestures vs. gaze and controller interaction. In International Conference on Human-Computer Interaction, page 151–160, Berlin, Heidelberg, 2021. Springer-Verlag. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[17] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. TPAMI, 35(1):221-231, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.382, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[18] Jianwen Jiang, Yuxuan Wei, Yifan Feng, Jingxuan Cao, and Yue Gao. Dynamic hypergraph neural networks. In IJCAI, pages 2635-2641, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[19] George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar. Multilevel hypergraph partitioning: Application in vlsi domain. In Proceedings of the 34th annual Design Automation Conference, pages 526-529, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[20] Matthew Korban and Xin Li. Ddgen: A dynamic directed graph convolutional network for action recognition. In ECCV, page 761-776, Berlin, Heidelberg, 2020. Springer-Verlag. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.907, + 0.612 + ], + "angle": 0, + "content": "[21] Bokyung Lee, Michael Lee, Pan Zhang, Alexander Tessier, and Azam Khan. Semantic human activity annotation tool using skeletonized surveillance videos. In Proceedings of the 2019 ACM International Symposium on Wearable Computers, pages 312-315, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[22] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In ICCV, pages 10221-10230, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[23] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In ICCV, pages 10410-10419, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[24] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Jun Liu, Amir Shahroudy, Dong Xu, Alex C. Kot, and Gang Wang. Skeleton-based action recognition using spatiotemporal LSTM network with trust gates. TPAMI, 40(12): 3007-3021, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 42(10):2684-2701, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12656" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, pages 140-149, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[28] Ashish S. Nikam and Aarti G. Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 Online International Conference on Green Engineering and Technologies, pages 1-5, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.299 + ], + "angle": 0, + "content": "[29] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory. In International Conference on Artificial Intelligence in Information and Communication, pages 273-278, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[30] Batyrkhan Omarov, Sergazy Narynov, Zhandos Zhumanov, Aidana Gumar, and Mariyam Khassanova. A skeleton-based approach for campus violence detection. Computers, Materials & Continua, 72(1), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[31] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In European Conference on Computer Vision, pages 605-622. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[32] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In Pattern recognition. ICPR international workshops and challenges: virtual event, January 10–15, 2021, Proceedings, Part III, pages 694–701. Springer, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[33] Helei Qiu, Biao Hou, Bo Ren, and Xiaohua Zhang. Spatiotemporal tuples transformer for skeleton-based action recognition. arXiv preprint arXiv:2201.02849, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[34] Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision. Visual Intelligence, 2(1):27, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[35] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[36] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12018-12027, 2019. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[37] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In CVPR, pages 7904-7913, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[38] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Decoupled spatial-temporal attention network for skeleton-based action-gesture recognition. In Proceedings of the Asian conference on computer vision, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. TIP, 29:9532-9545, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "NIPS, page 568-576, Cambridge, MA, USA, 2014. MIT Press. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. TPAMI, 45(2):1474-1488, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[42] Ahmed Taha, Hala H Zayed, ME Khalifa, and El-Sayed M El-Horbaty. Skeleton-based human activity recognition for video surveillance. International Journal of Scientific & Engineering Research, 6(1):993-1004, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[43] Jianwei Tang, Jieming Wang, and Jian-Fang Hu. Predicting human poses via recurrent attention network. Visual Intelligence, 1(1):18, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[45] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In ICCV, pages 4041-4049, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[46] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning, and recognition. In CVPR, page 2649-2656, USA, 2014. IEEE Computer Society. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[47] Qingtian Wang, Shuze Shi, Jiabin He, Jianlin Peng, Tingxi Liu, and Renliang Weng. Iip-transformer: Intra-inter-part transformer for skeleton-based action recognition. In BigData, pages 936-945, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[48] Shengqin Wang, Yongji Zhang, Hong Qi, Minghao Zhao, and Yu Jiang. Dynamic spatial-temporal hypergraph convolutional network for skeleton-based action recognition. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2147-2152. IEEE, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[49] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[50] Cong Wu, Xiao-Jun Wu, and Josef Kittler. Graph2net: Perceptually-enriched graph learning for skeleton-based action recognition. TCSVT, 32(4):2120-2132, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[51] Jianyang Xie, Yanda Meng, Yitian Zhao, Anh Nguyen, Xiaoyun Yang, and Yalin Zheng. Dynamic semantic-based spatial graph convolution network for skeleton-based human action recognition. AAAI, 38(6):6225-6233, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[52] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[53] Hengqi Zhang, Hua Geng, and Geng Yang. Two-stream transformer encoders for skeleton-based action recognition. In International Conference on Computing, Control and Industrial Engineering, pages 272-281. Springer, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[54] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In CVPR, pages 1109-1118, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[55] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12657" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[56] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In CVPR, pages 10608-10617, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[57] Yuxuan Zhou, Xudong Yan, Zhi-Qi Cheng, Yan Yan, Qi Dai, and Xian-Sheng Hua. Blockgen: Redefined topology awareness for skeleton-based action recognition. In CVPR, pages 2049-2058, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[58] Yiran Zhu, Guangji Huang, Xing Xu, Yanli Ji, and Fumin Shen. Selective hypergraph convolutional networks for skeleton-based action recognition. In ICMR, page 518-526, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12658" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_origin.pdf b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7b544e6f5571a165e9f8dbd2748a1a923ce44a4f --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/eb5cfddb-33b1-4050-a9eb-d937a7558bf3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a6fd33db5eae373477b99a0dc38ea9307f16f9d1ac0d4d4562d93bad7ed9483 +size 670787 diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/full.md b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4094e7a96908837590b486495a5b6518f8086c20 --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/full.md @@ -0,0 +1,344 @@ +# Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections + +Youwei Zhou $^{1}$ , Tianyang Xu $^{1*}$ , Cong Wu $^{1}$ , Xiaojun Wu $^{1}$ , Josef Kittler $^{2}$ $^{1}$ Jiangnan University, China $^{2}$ University of Surrey, UK + +youwei_zhou@stu.jiangnan.edu.cn + +{tianyang_xu,Congwu，wu_xiaojun}@jiangnan.edu.cn + +j.kittler@surrey.ac.uk + +# Abstract + +The shared topology of human skeletons motivated the recent investigation of graph convolutional network (GCN) solutions for action recognition. However, most of the existing GCNs rely on the binary connection of two neighboring vertices (joints) formed by an edge (bone), overlooking the potential of constructing multi-vertex convolution structures. Although some studies have attempted to utilize hyper-graphs to represent the topology, they rely on a fixed construction strategy, which limits their adaptivity in uncovering the intricate latent relationships within the action. In this paper, we address this oversight and explore the merits of an adaptive hyper-graph convolutional network (HyperGCN) to achieve the aggregation of rich semantic information conveyed by skeleton vertices. In particular, our Hyper-GCN adaptively optimises the hyper-graphs during training, revealing the action-driven multi-vertex relations. Besides, virtual connections are often designed to support efficient feature aggregation, implicitly extending the spectrum of dependencies within the skeleton. By injecting virtual connections into hyper-graphs, the semantic clues of diverse action categories can be highlighted. The results of experiments conducted on the NTU-60, NTU-120, and NW-UCLA datasets demonstrate the merits of our Hyper-GCN, compared to the state-of-the-art methods. The code is available at https://github.com/6UOON9/Hyper-GCN. + +# 1. Introduction + +Skeleton-based human action recognition is a popular research topic in artificial intelligence, with practical applications in video understanding, video surveillance, human-computer interaction, robot vision, VR and AR [16, 21, 28-30, 42, 49, 49]. In general, a skeleton sequence contains + +![](images/77062d86835fd5f1b65731af15262b34468aac1e2504409fc194e5636374c94f.jpg) +Figure 1. Illustration of our Hyper-GCN. In (a), dotted lines represent the virtual connections. Each coloured part containing multiple joints represents a hyper-graph with hyper-edges. (b) provides an intuitive comparison between normal graph convolution and hyper-graph convolution operations with the same connectivity degree. + +a series of 2D or 3D coordinates, which can easily be collected by low-cost depth sensors or obtained by video-based pose estimation algorithms [26, 34, 35, 43]. Compared to RGB and optical flow images, skeleton data, which represents the basic physical structure of a human being, is of lower dimension, conveying human action with higher efficiency. Moreover, it is robust to illumination changes and scene variations. For these reasons, the adoption of this structural data is very popular in skeleton-based action recognition [3, 6, 10, 25, 35, 45, 51]. + +To facilitate skeleton-based action recognition, both Recurrent Neural Networks (RNNs) [10, 25, 35, 45] and Con + +volitional Neural Networks (CNNs) [7, 17, 40] have been well explored. However, RNNs themselves cannot depict the intrinsic skeleton topology. The learned filters of CNNs, on the other hand, neglect the spatio-temporal structure of the skeleton. Drawing on these observations, recent studies have focused on how to model the skeleton topology directly. In principle, the physical topology of human joints and bones can be consistently represented by a graph. Accordingly, the graph convolutional network (GCN) is typically introduced to aggregate feature information conveyed by skeleton joints [23, 52, 57]. In terms of data representations, skeleton joints are also one-dimensional sequence data, similar to language. Therefore, with the rise of recent self-attention techniques [9, 44], several Transformer-based action recognition solutions [32, 38, 47, 53] have been released. Transformer considers skeletal joints as tokens and uses attention maps to reflect the topology. These methods have achieved excellent performance but sacrifice much higher GFLOPS and parameters than GCN. + +In general, two neighbouring joints can pass messages through their shared bone. In graph terms, this corresponds to the exchange of information between two vertices along their connecting edge. Besides the physical skeleton topology, attempts have been made to explore the implicit relationships among joints [3, 6, 23, 36], suggesting different variants of skeleton topology. In principle, these approaches assume a binary connection between each connected vertex pair. Mathematically, the constructed topology in the adjacency matrix is represented by a normal graph. However, human actions are jointly defined by several joints. Hence, human actions encompass not only binary relations between vertex pairs but also multi-joint relationships. For instance, the action primitive starting running is manifest in the human raising the left hand while the right leg steps forward, as shown in Figure 1 (a). The binary connections are not sufficient to capture the synergistic interaction of multiple joints. This strongly argues for constructing feature aggregation paths involving multiple vertices. + +Accordingly, we propose to construct a hyper-graph to depict the skeleton topology and take advantage of the outstanding performance of hyper-graph analysis techniques [2, 12, 18, 19]. The hyper-graph topology involves multiple node connections, rather than binary connections of a normal graph. As shown in Figure 1 (a), in a hyper-graph, a hyper-edge can link more than two vertices. This linking has the capacity to represent complex collaborative relations among human joints. As one hyper-edge associates multiple vertices, a single hyper-graph convolution enables aggregating all the features along the hyper-edges. + +An illustration is provided in Figure 1 (b), where the normal graph and hyper-graph are presented to demonstrate their differences in passing the information conveyed by the vertex features. In the normal graph convolution, after + +2-layer aggregation, the information of the green vertex is spread to 2 other vertices. In contrast, in a hyper-graph convolution, the information of the green vertex spreads to all the vertices, creating an extended receptive field. Theoretically, by modelling a skeleton using a hyper-graph for the aggregation of joint information, improved efficiency can be obtained during learning the action semantics. + +However, the key to achieving enhanced information aggregation by hyper-graph convolution lies in the conformity of the hyper-graph structure to the real topology of human action. Several studies [13, 58] attempt to adopt a hypergraph to model the skeleton topology, mostly based on superficial prior knowledge and an unfounded definition of the structure of hyper-graph. Distinguishing our approach from the existing hyper-graph-based methods, we design an adaptive solution for constructing the hyper-graph with virtual connections. It can construct a Non-uniform hypergraph specific to each individual action type. + +Besides the information aggregation issue, the capacity of an action recognition system is also constrained by its input features. In general, the number of skeleton joints is fixed in existing benchmarks, e.g., 25 for NTU-120 [26]. The entire recognition process relies on a successful capture of the interactions among the skeleton joints. In the domain of artistic puppetry, the history of driving actions, such as Shadow Play and Marionette goes back more than 2000 years. This kind of art form provides an inspiration for involving additional 'hyper joints' which can drive or facilitate a better communication between existing joints. The underlying spirit is to alleviate the pressure on the real joints to store and transfer complex semantics. Jointly with the hyper joints, real joints can focus more on storing neighbouring joint features and relegating the task of transferring the global information to the hyper joints. Interestingly, the class token in existing Transformers can also be considered as a hyper-token [9, 44]. As shown in Figure 1, the skeleton is described as a marionette, where the actions are "controlled" by connecting hyper joints to the real joints. This suggests that the hyper joints are not only able to capture the representation information of human action, but also reveal the implicit information between physically connected joints as hints for recognition. + +By endowing an adaptive non-uniform hyper-graph with hyper joints, virtual connections are created to perform comprehensive hyper-graph convolutions. We construct Hyper-GCN based on the above design principles. Extensive experiments on 3 datasets, NTU RGB+D 60 [35], NTU RGB+D 120 [26], and NW-UCLA [46], are conducted for evaluation. The results validate the merits of our proposed Hyper-GCN. The main contributions are as follows: + +- An adaptive non-uniform hyper-graph to represent the hu + +man skeleton topology. Compared with a fixed hypergraph, the constructed topology is more action-specific, thus boosting the discrimination. + +- The injection of virtual hyper joints, enriching the connectivity of the physical joints, from a global semantic perspective. +- As the processing architecture, we propose a hyper-graph convolution network (Hyper-GCN). The SOTA performance achieved on three public datasets demonstrates the merits of the Hyper-GCN and virtual connection designs. + +# 2. Related Work + +# 2.1. Graph Topology for Action Recognition + +A graph can represent a human skeleton, preserving the joint relationships via predefined edges. To aggregate semantics, GCNs [1, 4, 11, 20, 24, 37, 54, 57] and Transformers [8, 32, 38, 53, 55] have been well studied. + +For GCN-based approaches, ST-GCN [52] proposes to represent the topology with an adjacency matrix of 3 subsets by modeling the spatio-temporal relevance. Similarly, 2s-AGCN [36], InfoGCN [6], and DS-GCN [51] use a self-attentive mechanism to learn the topology from joint pairs. Besides the intuitive spatio-temporal dimensions, CTR-GCN [3] proposes to learn the topology for channels to refine the skeleton. + +For Transformer-based solutions, IIP-Transformer [47] adopt a self-attention mechanism to establish the intra-part and the inter-part joints. IGFormer [31] learns the topology between persons at both semantic and distance levels. STTFormer [33] proposed structure can capture the correlation between joints in consecutive frames. + +All the above two paradigms construct the topology from joint pairs, which rely only on the binary relation between two vertices. In their implementation stage, the adjacency matrix or attention map is used to represent the topology. In this case, high-order information among joints is not taken into consideration, neglecting the collaborative power among multiple joints. Though hyper-graph is considered to construct topology using multivariate joint relationships, current [13, 58] solutions manually set the hyper-graphs, which greatly relies on human experience, sacrificing the adaptability of graph learning. + +# 2.2. Feature Configurations for Action Recognition + +It has been observed that the input features play the most essential role in delivering high recognition accuracy [1, 5, 15, 27, 31, 33, 41, 47] in GCNs and Transformers. + +Drawing on this, Graph2Net [50] proposes to extract local and global spatial features jointly. CTR-GCN [3] uses multi-scale temporal convolution to extract temporal features. While HD-GCN [23] introduce the hierarchical edge convolution to key edge features. To extend the perception + +field, STC-Net [22] uses the dilated kernels for graph convolution to capture the features. + +Similarly, IG-Transformer [31] aggregate the features of two persons as one to mine the semantic information. STST [55] split the coordinate features, semantic features and temporal features into three kinds of tokens. TSTE [53] extracting and merging the spatial and temporal features as input to transformer encoders. + +After all, existing methods [6, 14, 32, 36, 38, 51, 52] receive the input features from real skeleton joints. However, each skeleton joint acts as an information carrier during forward passing, which is required to deliver both local context and global semantics. Given a fixed representative capacity (number of channels), we believe it is necessary to involve additional virtual joints to balance the pressure of storing local and global information. + +# 3. Approach + +# 3.1. Preliminaries + +In general, the input features in graph convolution are represented by $\mathbf{F}_{\mathrm{in}}\in \mathbb{R}^{C\times T\times V}$ , where $C$ represents the number of channels in the feature maps. Given the topology of a human skeleton, we usually define the graph $\mathcal{G} = (\mathcal{V},\mathcal{E})$ , where $\mathcal{V}$ represents the set of joints and $\mathcal{E}$ represents the set of edges between joints. For the set of edges $\mathcal{E}$ , it is formulated as an adjacent matrix $A\in \mathbb{R}^{N\times N}$ , where $N$ represents the number of human joints. The normalised adjacency matrix is represented by $\hat{A}\in \mathbb{R}^{N\times N}$ . The normalisation operation is formulated as follows: + +$$ +\hat {A} = \Lambda^ {- \frac {1}{2}} A \Lambda^ {- \frac {1}{2}}, \tag {1} +$$ + +where $\Lambda$ represents the diagonal matrix stored with the degree of every joint. The entire normal graph convolution process can be formulated as: + +$$ +\mathbf {F} _ {\text {o u t}} = \sigma (\hat {A} \mathbf {F} _ {\text {i n}} P), \tag {2} +$$ + +where $P \in \mathbb{R}^{C \times C'}$ is the learnable parameters, representing the feature transformation patterns in the feature space. $\sigma$ denotes the non-linear activation function ReLU. + +# 3.2. Hyper-graph + +Here, we use $\mathcal{G}_{\mathcal{H}} = (\mathcal{V}_{\mathcal{H}},\mathcal{E}_{\mathcal{H}},\mathcal{W}_{\mathcal{H}})$ to define the spatial hyper-graph [12] with human skeleton. $\mathcal{V}_{\mathcal{H}}$ and $\mathcal{E}_{\mathcal{H}}$ follow similar definitions in the normal graph, which represent the set of joints and the set of hyper-edges. In addition, we introduce $\mathcal{W}_{\mathcal{H}}$ to represent the weights of each hyper-edge. Since a hyper-edge contains multiple joints, the corresponding topology can no longer be simply represented by an adjacency matrix. Therefore, we introduce the incidence matrix $H\in \mathbb{R}^{N\times M}$ to describe the topology of each joint in the hyper-graph. $N$ represents the number of joints and $M$ + +![](images/43bfcada4a3d50f8c88cf05b02a441690a45df1e21f4844af313cbf1e1eee388.jpg) + +![](images/db5a874a144d77a1d8dac9f28bd4f697675133051f3bab7ae9a3902818567e32.jpg) +Figure 2. The framework of our Hyper-GCN. Part (a) represents the Multi-head Hyper-graph Convolution (M-HGC) module. Part (b) represents the process of constructing an adaptive hyper-graph. Part (c) represents the architecture of Hyper-GCN. $\mathbf{F}_{\mathbf{P}}$ represents the features of real joints. $\mathbf{F}_h$ represents the hyper-joints which are learnable parameters. $\hat{A}$ represents the physical topology. + +![](images/2018708943c56b0aae8d52505b46ebc1f9098223d66fdeaea1080aa0cc810b3b.jpg) + +represents the number of hyper-edges. Given $v \in \mathcal{V}_{\mathcal{H}}$ and $e \in \mathcal{E}_{\mathcal{H}}$ , the values of incidence matrix be determined by: + +$$ +h (e, v) = \left\{ \begin{array}{l l} 1, & v \in e, \\ 0, & v \notin e \end{array} \right.. \tag {3} +$$ + +Similar to the normal graph convolution, it also needs to normalise the hyper-graph to modulate the aggregated features. The degree of hyper-graph consists of the degree of joints and the degree of hyper-edges. The degree of joints is represented by the sum of weights of all joints contained in each hyper-edge. Given $v \in \mathcal{V}_{\mathcal{H}}$ , it can be described as follows: + +$$ +d (v) = \sum_ {e \in \mathcal {E} _ {\mathcal {H}}} w (e) h (v, e), \tag {4} +$$ + +where we use a diagonal matrix $D_{v} \in \mathbb{R}^{N \times N}$ to represent the degree of joints. $W \in \mathbb{R}^{M \times M}$ represents the weight matrix of hyper-edges, which is formulated as a diagonal matrix. The degree of hyper-edges represents the sum of the number of joints contained in each hyper-edge. Given $e \in \mathcal{E}_{\mathcal{H}}$ , it can be described as follows: + +$$ +d (e) = \sum_ {v \in \mathcal {V} _ {\mathcal {H}}} h (v, e), \tag {5} +$$ + +where we use a diagonal matrix $D_{e} \in \mathbb{R}^{M \times M}$ to represent the degree of hyper-edges. Therefore, the normalisation of a hyper-graph is formulated as follows: + +$$ +\hat {H} = D _ {v} ^ {- 1} H W D _ {e} ^ {- 1} H ^ {T}, \tag {6} +$$ + +![](images/d545ccc8059f8da246488210c60d55b923c118918b9210c3bea3f832eee6557a.jpg) +Figure 3. Instance of the incidence matrix construction by A-NHG with hyper-parameter $K = 3$ . $N$ represents the number of joints. (a) represents the hyper-edge sets are null. (b) represents that measure the distance between 1-st joint and other joints. (c) represents that retrain $K$ -nearest hyper-edges which contain 1-st joint. (d) represents the final state. + +where $\hat{H} \in \mathbb{R}^{N \times N}$ represents the normalised incidence matrix for hyper-graph convolution. + +# 3.3. Adaptive Non-uniform Hyper-graph + +Constructing the hyper-graph incidence matrix is critical to understanding human actions. As shown in Figure 2 (b), we + +design an Adaptive Non-uniform Hyper-graph (A-NHG) Construction. + +Given the general features $\mathbf{X} \in \mathbb{R}^{N \times C}$ of spatial skeleton joints, $N$ represent the number of joints, $C$ represent the channels. Our A-NHG construct the hyper-edge with each joint as the center of mass, and a total of $N$ hyper-edges are constructed. A-NHG utilises the Euclidean distance to measure the joint difference. + +We use one mapping function $\Phi \in \mathbb{R}^{C\times C_h}$ to embed the features into the subspace for constructing the hypergraph. $C_h$ represents the hidden channels of the mapping subspace. This design enables preserving the original features to reflect their spatial relevance. Based on $X_{H}\in \mathbb{R}^{N\times C_{h}}$ , we define the set of joints $\nu_{\mathcal{H}} = \{v_{1}\in$ $\mathbb{R}^{C_h},v_2\in \mathbb{R}^{C_h},\dots,v_N\in \mathbb{R}^{C_h}\}$ . Given $v_{i},v_{j}\in \mathcal{V}_{\mathcal{H}}$ , each element $m_{i,j}$ in the distance matrix $M\in \mathbb{R}^{N\times N}$ can be obtained as follows: + +$$ +m _ {i, j} = m _ {j, i} = \left\| v _ {i} - v _ {j} \right\| _ {2}. \tag {7} +$$ + +In principle, we need to transform the distance matrix into a probability incidence matrix, which determines whether the specific joints are contained with the same hyper-edges. In order to guarantee that the model is trainable, we use the softmax operation to assign the probabilities. However, directly using the softmax would inevitably result in each joint belonging to all hyper-edges to some extent. + +Since this outcome is undesirable, we introduce constraints on each joint to limit the number of hyper-edges it can belong to, preventing excessive connections. For one joint, we retain only the $K$ nearest hyper-edges represented by joints in the semantic space for probability estimation. And $K$ is the hyper-parameter of A-NHG. + +Given the row vector $\mathbf{m}_i$ in the distance matrix $M$ , which represents the distances between joint $i$ and the other joints. We select the indices of the minimum $K$ joints from $\mathbf{m}_i$ and add them to the set $set_i$ . Based on $set_i$ , $h_{i,j}$ in the incidence matrix $H\in \mathbb{R}^{N\times N}$ , can be calculated as follows: + +$$ +h _ {i, j} = \left\{ \begin{array}{c c} \frac {\exp (- m _ {i , j})}{\sum_ {k \in s e t _ {i}} \exp (- m _ {i , k})}, & j \in s e t _ {i}, \\ 0, & j \notin s e t _ {i}. \end{array} \right. \tag {8} +$$ + +In contrast to uniform hyper-graph (every hyper-edge contains the same number of joints), our approach limits the maximum number of every joint that can be contained by hyper-edges. Obviously, the number of joints contained in each hyper-edge constructed in this way is non-fixed, allowing the hyper-edges to capture more differentiated associations. The entire process for example with $K = 3$ is illustrated in the Figure 3. + +# 3.4. Multi-head Hyper-graph Convolution + +To accommodate the semantic information reflected by different group channels, we further propose the Multi-head Hyper-graph Convolution (M-HGC), as shown in Figure 2 (a). + +We use separate branches to independently perform hyper-graph convolution on the topologies represented by the multi-head hyper-graphs, thereby enhancing the computational efficiency of the Hyper-GCN. For instance, the features $\mathbf{F}_{\mathrm{in}} \in \mathbb{R}^{C_{\mathrm{in}} \times T \times V}$ are divided into 8 separate branches along the channel dimension, which are processed by 8 M-HGC parallelly, delivering 8 separate hyper-graphs. + +Next, we perform temporal pooling, and then the obtained spatial information is the evidence to determine the optimal hyper-graph construction. This can efficiently decouple the temporal and spatial clues. After performing temporal average-pooling, we can obtain $\bar{\mathbf{F}}_{\mathrm{in}} \in \mathbb{R}^{C_{\mathrm{in}} \times V}$ . + +The hyper-graph is depicted by the incidence matrix and weight matrix. So, we use a separate mapping function for each head to embed the features into the subspace for constructing the hyper-graph. This design enables preserving the original features to reflect their spatial relevance. After that, we introduce the A-NHG to obtain the incidence matrix. For the weight matrix, we adopt MLP to measure each hyper-edge. The hyper-graph obtained operation can be formulated as follows: + +$$ +H = \uplus_ {k = 1} ^ {8} \mathrm {A} - \mathrm {N H G} \left(\bar {\mathbf {F}} _ {\mathbf {i n}} ^ {k}\right), \tag {9} +$$ + +$$ +W = \sigma_ {2} \left(\Psi_ {2} \left(\uplus_ {k = 1} ^ {8} \sigma_ {1} \left(\Psi_ {1} ^ {k} \left(\overline {{\mathbf {F}}} _ {\mathbf {i n}} ^ {k}\right)\right)\right)\right), \tag {10} +$$ + +where $\Psi_{1} \in \mathbb{R}^{C_{\mathrm{in}} / 8 \times C_{h}}$ , $\Psi_{2} \in \mathbb{R}^{(8 \times C_{h}) \times 8}$ are the mapping functions, which are set as learnable parameters. $C_{h}$ represents the hidden channels of the mapping subspace. $k$ represents the channels $[(k - 1) \times C_{\mathrm{in}} / 8 + 1, \dots, k \times C_{\mathrm{in}} / 8]$ of $\overline{\mathbf{F}}_{\mathrm{in}}$ . $\uplus$ represents the channel concatenation. $\sigma_{1}, \sigma_{2}$ denote the activation function LeakeyReLU and Tanh to obtain the weight of each hyper-edge, limiting the values with the range of $[-1, 1]$ . Based on $H \in \mathbb{R}^{8 \times N \times N}$ , $W \in \mathbb{R}^{8 \times N}$ , normalise the hyper-graph by sec.3.2 to obtain the $\hat{H} \in \mathbb{R}^{8 \times N \times N}$ . + +Besides the hyper-graph, we incorporate the physical topology to emphasise the natural physical relations of a human being's skeleton. To achieve this objective, existing methods [3, 23, 36, 52] divide the physical topology of the human body into 3 subsets. They are represented by $S = \{s_{id}, s_{cf}, s_{cp}\}$ , where $s_{id}, s_{cf}$ and $s_{cp}$ denote the identity, centrifugal, and centripetal joint subsets. In our M-HGC, to ensure that the integrated physical topology in each head remains complete, we merge them into a single set. 8 heads are aggregated by channel concatenation after the hyper-graph convolutions, as shown in Figure 2 (a). The operation of M-HGC can be formulated as follows: + +$$ +\mathbf {F} _ {\text {o u t}} = \uplus_ {k = 1} ^ {8} \left(\hat {A} _ {k} + \alpha \cdot \hat {H} _ {k}\right) \mathbf {F} _ {\text {i n}} ^ {k} P _ {k}, \tag {11} +$$ + +where $\hat{A}$ and $\hat{H}$ denote the normalized physical adjacency matrix and normalized incidence matrix. $P$ is the learnable weight parameter for feature transform. $\alpha$ is a learnable parameter for the topology fusion of each head. + +# 3.5. Virtual Connections + +It is worth emphasising that incorporating learnable joints among different samples is essential for enhancing the model capacity, as these learnable joints capture generalised features of human actions. This not only enriches the semantic information but also facilitates easier interaction connections among real joints. Therefore, we introduce the hyper-joints which are to participate in the hyper-graph convolution as shown in Figure 2 (a). + +The shape of hyper-joints is consistent with the features of physical joints $\mathbf{F}_{\mathbf{p}} \in \mathbb{R}^{C \times T \times V}$ , which is described as $\mathbf{F}_{\mathbf{h}} \in \mathbb{R}^{C \times T \times V_{h}}$ ( $V_{h}$ represents the number of hyper-joints). To ensure alignment in the temporal dimension, the hyper-joints are shared across each frame. Specifically, the learning of hyper-joints is supervised by the loss function, which guides these hyper-joints to support generalisable driven features embedded within large amounts of data. Furthermore, we set independent hyper-joints at each layer of Hyper-GCN, aiming to harmonise features at different depths. Typically, these hyper-joints are involved in spatial hyper-graph convolution rather than in temporal convolution. + +To diversify these hyper-joints, we propose the Divergence Loss for hyper-joints optimisation to mitigate their homogenisation. We adopt a cosine matrix $C \in \mathbb{R}^{V_h \times V_h}$ to measure the differences between hyper-joints, which can be formulated as follows: + +$$ +C = \frac {\mathbf {F} _ {\mathrm {h}} \mathbf {F} _ {\mathrm {h}} ^ {T}}{| | \mathbf {F} _ {\mathrm {h}} | | ^ {2}}, \tag {12} +$$ + +Specifically, in Divergence Loss, we measure the differences of hyper-joints with the mean of the cosine matrix $C$ in each layer. Since the correlation of hyper-joints themselves cannot be optimised, we manually eliminate this part by minus the $V_{h}$ . The loss calculation can be formulated as: + +$$ +\mathcal {L} _ {h} (C) = \frac {\sum_ {i = 1} ^ {V _ {h}} \sum_ {j = 1} ^ {V _ {h}} R e L U \left(c _ {i , j}\right) - V _ {h}}{V _ {h} \left(V _ {h} - 1\right)}, \quad c _ {i, j} \in C, \tag {13} +$$ + +$$ +\mathcal {L} = \mathcal {L} _ {C E} + \frac {1}{L} \sum_ {l = 1} ^ {L} L _ {h} \left(C _ {l}\right), \tag {14} +$$ + +where $V_{h}$ represents the number of hyper-joints we introduced. $C_l$ represents the cosine matrix of the $l$ -th layer. $\mathcal{L}_{\mathrm{CE}}$ represents the cross-entropy loss. Additionally, we manually connect the hyper joints to all the physical joints, which optimises the topology from the datasets. + +# 3.6. Entire Architecture + +Hyper-GCN consists of an embedding layer [6], 9 spatial-temporal convolution layers as shown in Figure 2 (c). Each layer consists of the proposed M-HGC and Multi Scale Temporal convolution (MS-TC) [3]. 9 Layers are categorised into 3 stages. In each stage, we introduce the dense connections to integrate deep and shallow features in each stage. Additionally, to further validate the potential of Hyper-GCN, we propose a base version and a large version. The channels in each stage are set to 128, 256, 256 for the base version and 128, 256, 512 for the large version. + +# 4. Evaluation + +# 4.1. Datasets + +NTU-RGB+D 60 & 120 NTU-RGB+D 60 [35] is a large dataset widely used in skeleton-based human action recognition, which is categorised into 60 classes. NTU-RGB+D 120 [26] is an extension to 120 classes of NTU 60. 4 benchmarks recommended by the official are adopted: (1) NTU60-XSub, (2) NTU60-XView, (3) NTU120-XSub, (4) NTU120-Xset. + +Northwestern-UCLA. The Northwestern-UCLA (NWUCLA) dataset [46] contains 1494 video clips, which is categorised into 10 classes. It contains 3 different camera views and is performed by 10 actors. + +# 4.2. Implementation Details + +Our training and evaluation stages are on a single GPU RTX 3090. In the training phase, Hyper-GCN is optimised by Stochastic Gradient Descent (SGD) with Nesterov momentum set at 0.9 and a weight decay at 0.0004. Our implementation uses label smooth cross-entropy loss with the Divergence Loss we proposed. We set a total of 140 epochs with the start 5 warm-up epochs. The initial learning rate is 0.05, which is reduced to 0.005 at epoch 110 and to 0.0005 at epoch 120. + +# 4.3. Comparison with the State-of-the-Art + +Multi-stream ensemble proposed in [39] have been proven effective by most of the existing state-of-the-art methods. We also use a 4-stream ensemble to evaluate the base version and the large version of Hyper-GCN. The detailed results are reported in Table 1. + +In summary, the base version of Hyper-GCN comprehensively outperforms all GCN-based and HGCN-based (hyper-graph utilised) SOTA and surpasses the Transformer-based SOTA on the NTU120, with a most lightweight design. Moreover, when scaled to match the parameters of the lightest Transformer-based method, the large version of Hyper-GCN achieves 1-st place in 4 benchmarks and 2-nd place in 1 benchmark. This demonstrates the effectiveness and merits of Hyper-GCN. + +
CategoryMethodsModalitiesParams (M)GFLOPsNTU-RGB+D 60NTU-RGB+D 120NW-UCLA (%)
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)
GCNST-GCN [52]J+B--81.588.370.773.2-
2s-AGCN [36]J+B--88.595.182.584.2-
DC-GCN+ADG [4]J+B+JM+BM4.91.8390.896.686.588.195.3
MS-G3D [27]J+B+JM+BM2.85.2291.596.286.988.4-
MST-GCN [11]J+B+JM+BM12.0-91.596.687.588.8-
CTR-GCN [3]J+B+JM+BM1.51.9792.496.488.990.696.5
EfficientGCN-B4 [41]J+B+JM+BM2.015.2091.795.788.389.1-
InfoGCN [6]J+B+JM+BM1.61.8492.796.989.490.796.6
FR Head [56]J+B+JM+BM2.0-92.896.889.590.996.8
HD-GCN* [23]J+B+J' +B'1.71.7793.097.089.891.296.9
DS-GCN [51]J+B+JM+BM--93.197.589.290.3-
BlockGCN [57]J+B+JM+BM1.31.6393.197.090.391.596.9
TransformerDSTA-Net [38]J+B+JM+BM3.516.1889.595.786.689.0-
IIP-Transformer [47]J+B+JM+BM2.97.2089.595.789.990.9-
SkateFormer [8]J+B+JM+BM2.03.6293.597.889.891.498.3
HGCNHyper-GNN [13]J+B+JM+BM--89.595.7---
Selective-HCN [58]J+B+JM+BM--90.896.6---
DST-HCN [48]J+B+JM+BM3.52.9392.396.888.890.796.6
Ours (B)J+B+JM+BM1.11.6393.397.490.591.797.2
Ours (L)J+B+JM+BM2.32.8893.797.890.992.097.6
+ +Table 1. Comparison of Hyper-GCN with advanced solutions on NTU-RGB+D 60, NTU-RGB+D 120, and NW-UCLA datasets. For a fair comparison, we use a 4-streams ensemble to evaluate the performance of the mentioned methods. J, B, JB, and JM represent the joint, bone, joint motion, and bone motion. In addition, for HD-GCN [23], we use the results from 4-streams according to the original papers for a fair comparison. The bold font represents the best result and underline font represents the 2-nd best result. + +
NTU RGB+D 120 X-Sub (%)
KUniformNon-uniform
Baseline-84.784.7
w M-HGC386.4 (↑ 1.7)86.0 (↑ 1.3)
586.5 (↑ 1.9)86.2 (↑ 1.5)
786.3 (↑ 1.6)86.5 (↑ 1.8)
986.0 (↑ 1.3)86.7 (↑ 2.2)
1185.9 (↑ 1.2)86.4 (↑ 1.7)
+ +Table 2. Ablation of the hyper-parameter $K$ in A-NHG. The Uniform represents K-uniform hyper-graph, in which one hyper-edge contains K joints. The non-uniform represents non-uniform hyper-graph, in which one joint is contained by $K$ hyper-edge. + +# 4.4. Ablation Study + +We conduct ablation experiments with visualisations to analyse the effectiveness of our design. redAll the ablation study is evaluated with the joint modality of base version. + +Hyper-graph Construction: The Non-uniform hypergraphs with hyper-parameter $K$ represent the maximum number of hyper-edges contained in each joint. They reflect the varying degrees of aggregation in the constructed topology. In general, it is common to adopt a uniform hypergraph with hyper-parameter $K$ . Therefore, to explore how to set the hyper-parameter $K$ for the adaptive hyper-graph in M-HGC, we design the ablation experiments to investigate the impact of hyper-parameter $K$ on the model's recognition ability, where both uniform and non-uniform hypergraphs are considered. + +![](images/f8ec9720ed8344ce63630c6b4ae80986a1ae8f38b9813b78118c15cd64f21be9.jpg) +Figure 4. Visualisation of hyper-graph in selected actions. For ease of observation, the joints contained in the selected hyper-edge are highlighted in red. + +The results are reported in Table 2. In both uniform and non-uniform hyper-graphs, a very large $K$ means that more joints are contained in each hyper-edge. Even though these "extra" joints may be optimised for minimal weight, they are likely to act as "noise", disrupting the information inter + +
NTU RGB+D 120 X-Sub (%)
w/o Lhw/ Lh
Baseline84.784.7
w/o M-HGCw/ 184.9 (↑ 0.2)84.9 (↑ 0.2)
w/ 384.9 (↑ 0.2)85.2 (↑ 0.5)
w/ 584.785.0 (↑ 0.3)
w M-HGCw/ 186.7 (↑ 2.0)86.7 (↑ 2.0)
w/ 386.6 (↑ 1.8)86.9 (↑ 2.2)
w/ 586.6 (↑ 1.8)86.8 (↑ 2.1)
+ +Table 3. Ablations on hyper joints. w/ $N$ represents the number of hyper joints in each layer. w/ $\mathcal{L}_h$ represents the training with Divergence Loss. The hyper-parameter $K$ of M-HGC is set to 9. + +![](images/ac8795de8fa7bceae5fa8234961cef74c34fc94ea442eaa851fa91f049d20355.jpg) +Figure 5. Comparison of the cosine matrices of hyper-joints. Part (a) and (b) represents w/o Divergence Loss and w/ Divergence Loss with 3 hyper-joints. It is calculated by Eqn. (12). + +action represented by the hyper-edge. As the $K$ decreases, it is difficult for the hyper-edge to represent complex joint combination patterns. Therefore, setting $K = 5$ in uniform and $K = 9$ in Non-uniform achieves the best performance. The non-uniform hyper-graph of A-NHG outperforms the uniform hyper-graph, as the combination of skeleton joints in different actions is diverse. + +Furthermore, we visualise the hyper-graph inferred by Hyper-GCN on selected actions in Figure 4. The hypergraph construction focuses on the joints that are most relevant to the corresponding action category. For "Kicking something" as part (d), the left leg joints and the right hand joints are modelled as a hyper-edge. Because in this action, the person's right hand is previously extended forward, and the left leg swings relatively to the right hand. Similarly, in the action "Stand up" as shown in part (b), the hand and foot joints serve as fixed joints, while the remaining body joints move relative to them. In addition, "Taking a selfie" and "Drinking water" both focus on the main joints of movement with the above view as Figure 4 (a), (c). + +Virtual Connections: We conduct the ablation study to analyse the hyper joints and the Divergence Loss. The results are listed in Table 3. By comparison, the performance is not monotonically increasing with the number of hyper joints. The case of introducing only 3 hyper joints achieves + +![](images/5bff5685e0e431a8763f1c18c1bb16666f8f38873140dbf55e8c2008dd9bcb0b.jpg) +Figure 6. t-SNE visualisation of Baseline and Hyper-GCN. Different colours represent different joints. Part (a) and (b) represent the output features in Baseline and Hyper-GCN. + +![](images/d671b0d9050a0917b58d1dc97e6cb0bf33377a7c33cc04845daa407fe0c2188a.jpg) + +the best performance. As our hyper joints are learned from a large amount of data, involving a large number of hyper joints can introduce redundant and ambiguous clues, degrading the performance. + +In addition, we visualise the cosine matrices of HyperGCN, as shown in Figure 5. Clearly, in the absence of the Divergence Loss, the homogenisation of the learned hyper-joints is severe. This impedes the ability of hyper-joints to represent the generalised features of human actions. The observation further validates the effectiveness of the Divergence Loss for hyper-joint optimisation, supporting our viewpoint that generalised features with a certain degree of differentiation are needed to participate in hyper-graph convolution. + +Effectiveness of Hyper-GCN: To further validate that Hyper-GCN increases the efficiency of information interaction, we perform t-SNE projections on the output features in last layer between Baseline and Hyper-GCN, as shown in Figure 6 for comparison. Notice that the semantic information represented by the last Hyper-GCN layer is very convergent. This represents that the semantic information of each joint is adequately conveyed in M-HGC. The joints with similar semantics can prevent sacrificing the information after global average pooling. + +# 5. Conclusion + +In this paper, we propose an Adaptive Hyper-graph Convolutional Network for skeleton-based human action recognition. To exploit the implicit topology of multivariate synergy between joints, we introduce the Adaptive Nonuniform Hyper-graph, the Multi-head Hyper-graph Convolution and virtual connections. We carry out experiments on the dataset NTU-RGB+D 60 & 120, NW-UCLA to validate the effectiveness of Hyper-GCN. The experimental analysis verifies that our design can improve the recognition performance. To the best of our knowledge, Hyper-GCN achieves the SOTA performance on 3 public datasets. The involvement of adaptive non-uniform hypergraph modelling indeed extends existing GCN-based action recognition paradigms. + +Acknowledgment This work was supported in part by the National Natural Science Foundation of China (62020106012, 62106089, 62332008, 62336004), the Engineering and Physical Sciences Research Council (EPSRC), U.K. (Grants EP/V002856/1, and EP/T022205/1), and the Fundamental Research Funds for the Central Universities (JUSRP202504007). + +# References + +[1] Tasweer Ahmad, Lianwen Jin, Luojun Lin, and GuoZhi Tang. Skeleton-based action recognition using sparse spatiotemporal gcn with edge effective resistance. Neurocomputing, 423:389-398, 2021. 3 +[2] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2 +[3] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, pages 13339-13348, 2021. 1, 2, 3, 5, 6, 7 +[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, page 536-553, Berlin, Heidelberg, 2020. Springer-Verlag. 3, 7 +[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In CVPR, pages 180-189, 2020. 3 +[6] Hyung-Gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, pages 20154-20164, 2022. 1, 2, 3, 6, 7 +[7] Guilhem Chéron, Ivan Laptev, and Cordelia Schmid. P-cnn: Pose-based cnn features for action recognition. In ICCV, pages 3218-3226, 2015. 2 +[8] Jeonghyeok Do and Munchurl Kim. Skateformer: skeletal-temporal transformer for human action recognition. In ECCV, pages 401-420. Springer, 2024. 3, 7 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1 +[11] Dong Feng, Zhongcheng Wu, Jun Zhang, and Tingting Ren. Multi-scale spatial temporal graph neural network for skeleton-based action recognition. IEEE Access, 9:58256-58265, 2021. 3, 7 +[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. TPAMI, 45(3):3181-3199, 2023. 2, 3 + +[13] Xiaoke Hao, Jie Li, Yingchun Guo, Tao Jiang, and Ming Yu. Hypergraph neural network for skeleton-based action recognition. TIP, 30:2263-2275, 2021. 2, 3, 7 +[14] Ruijie Hou, Zhihao Wang, Ruimin Ren, Yang Cao, and Zhao Wang. Multi-channel network: Constructing efficient gcn baselines for skeleton-based action recognition. Computers & Graphics, 110:111-117, 2023. 3 +[15] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Part-level graph convolutional network for skeleton-based action recognition. In AAAI, pages 11045-11052, 2020. 3 +[16] Cloe Huesser, Simon Schubiger, and Arzu Öltekin. Gesture interaction in virtual reality: A low-cost machine learning system and a qualitative assessment of effectiveness of selected gestures vs. gaze and controller interaction. In International Conference on Human-Computer Interaction, page 151–160, Berlin, Heidelberg, 2021. Springer-Verlag. 1 +[17] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. TPAMI, 35(1):221-231, 2012. 2 +[18] Jianwen Jiang, Yuxuan Wei, Yifan Feng, Jingxuan Cao, and Yue Gao. Dynamic hypergraph neural networks. In IJCAI, pages 2635-2641, 2019. 2 +[19] George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar. Multilevel hypergraph partitioning: Application in vlsi domain. In Proceedings of the 34th annual Design Automation Conference, pages 526-529, 1997. 2 +[20] Matthew Korban and Xin Li. Ddgen: A dynamic directed graph convolutional network for action recognition. In ECCV, page 761-776, Berlin, Heidelberg, 2020. Springer-Verlag. 3 +[21] Bokyung Lee, Michael Lee, Pan Zhang, Alexander Tessier, and Azam Khan. Semantic human activity annotation tool using skeletonized surveillance videos. In Proceedings of the 2019 ACM International Symposium on Wearable Computers, pages 312-315, 2019. 1 +[22] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In ICCV, pages 10221-10230, 2023. 3 +[23] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In ICCV, pages 10410-10419, 2023. 2, 3, 5, 7 +[24] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3 +[25] Jun Liu, Amir Shahroudy, Dong Xu, Alex C. Kot, and Gang Wang. Skeleton-based action recognition using spatiotemporal LSTM network with trust gates. TPAMI, 40(12): 3007-3021, 2018. 1 +[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 42(10):2684-2701, 2020. 1, 2, 6 + +[27] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, pages 140-149, 2020. 3, 7 +[28] Ashish S. Nikam and Aarti G. Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 Online International Conference on Green Engineering and Technologies, pages 1-5, 2016. 1 +[29] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory. In International Conference on Artificial Intelligence in Information and Communication, pages 273-278, 2021. +[30] Batyrkhan Omarov, Sergazy Narynov, Zhandos Zhumanov, Aidana Gumar, and Mariyam Khassanova. A skeleton-based approach for campus violence detection. Computers, Materials & Continua, 72(1), 2022. 1 +[31] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In European Conference on Computer Vision, pages 605-622. Springer, 2022. 3 +[32] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In Pattern recognition. ICPR international workshops and challenges: virtual event, January 10–15, 2021, Proceedings, Part III, pages 694–701. Springer, 2021. 2, 3 +[33] Helei Qiu, Biao Hou, Bo Ren, and Xiaohua Zhang. Spatiotemporal tuples transformer for skeleton-based action recognition. arXiv preprint arXiv:2201.02849, 2022. 3 +[34] Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision. Visual Intelligence, 2(1):27, 2024. 1 +[35] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 1, 2, 6 +[36] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12018-12027, 2019. 2, 3, 5, 7 +[37] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In CVPR, pages 7904-7913, 2019. 3 +[38] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Decoupled spatial-temporal attention network for skeleton-based action-gesture recognition. In Proceedings of the Asian conference on computer vision, 2020. 2, 3, 7 +[39] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. TIP, 29:9532-9545, 2020. 6 +[40] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In + +NIPS, page 568-576, Cambridge, MA, USA, 2014. MIT Press. 2 +[41] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. TPAMI, 45(2):1474-1488, 2022. 3, 7 +[42] Ahmed Taha, Hala H Zayed, ME Khalifa, and El-Sayed M El-Horbaty. Skeleton-based human activity recognition for video surveillance. International Journal of Scientific & Engineering Research, 6(1):993-1004, 2015. 1 +[43] Jianwei Tang, Jieming Wang, and Jian-Fang Hu. Predicting human poses via recurrent attention network. Visual Intelligence, 1(1):18, 2023. 1 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[45] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In ICCV, pages 4041-4049, 2015. 1 +[46] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning, and recognition. In CVPR, page 2649-2656, USA, 2014. IEEE Computer Society. 2, 6 +[47] Qingtian Wang, Shuze Shi, Jiabin He, Jianlin Peng, Tingxi Liu, and Renliang Weng. Iip-transformer: Intra-inter-part transformer for skeleton-based action recognition. In BigData, pages 936-945, 2023. 2, 3, 7 +[48] Shengqin Wang, Yongji Zhang, Hong Qi, Minghao Zhao, and Yu Jiang. Dynamic spatial-temporal hypergraph convolutional network for skeleton-based action recognition. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2147-2152. IEEE, 2023. 7 +[49] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1 +[50] Cong Wu, Xiao-Jun Wu, and Josef Kittler. Graph2net: Perceptually-enriched graph learning for skeleton-based action recognition. TCSVT, 32(4):2120-2132, 2022. 3 +[51] Jianyang Xie, Yanda Meng, Yitian Zhao, Anh Nguyen, Xiaoyun Yang, and Yalin Zheng. Dynamic semantic-based spatial graph convolution network for skeleton-based human action recognition. AAAI, 38(6):6225-6233, 2024. 1, 3, 7 +[52] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2, 3, 5, 7 +[53] Hengqi Zhang, Hua Geng, and Geng Yang. Two-stream transformer encoders for skeleton-based action recognition. In International Conference on Computing, Control and Industrial Engineering, pages 272-281. Springer, 2021. 2, 3 +[54] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In CVPR, pages 1109-1118, 2020. 3 +[55] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3 + +[56] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In CVPR, pages 10608-10617, 2023. 7 +[57] Yuxuan Zhou, Xudong Yan, Zhi-Qi Cheng, Yan Yan, Qi Dai, and Xian-Sheng Hua. Blockgen: Redefined topology awareness for skeleton-based action recognition. In CVPR, pages 2049-2058, 2024. 2, 3, 7 +[58] Yiran Zhu, Guangji Huang, Xing Xu, Yanli Ji, and Fumin Shen. Selective hypergraph convolutional networks for skeleton-based action recognition. In ICMR, page 518-526, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 7 \ No newline at end of file diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/images.zip b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..31898d76a9934354c453901a668bd104fd967c60 --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e48450ac178b17483ef4d4b07735d2271630408efdeafa0193e6f4494c96bef +size 587173 diff --git a/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/layout.json b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7fac0d8498be34bbe1550ec128a0eced7a942b8a --- /dev/null +++ b/2025/Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections/layout.json @@ -0,0 +1,9267 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 505, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 140 + ], + "type": "text", + "content": "Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "spans": [ + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": "Youwei Zhou" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": ", Tianyang Xu" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": ", Cong Wu" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": ", Xiaojun Wu" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": ", Josef Kittler" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": "Jiangnan University, China " + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 160, + 475, + 190 + ], + "type": "text", + "content": "University of Surrey, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 191, + 389, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 191, + 389, + 203 + ], + "spans": [ + { + "bbox": [ + 221, + 191, + 389, + 203 + ], + "type": "text", + "content": "youwei_zhou@stu.jiangnan.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 175, + 205, + 435, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 205, + 435, + 217 + ], + "spans": [ + { + "bbox": [ + 175, + 205, + 435, + 217 + ], + "type": "text", + "content": "{tianyang_xu,Congwu，wu_xiaojun}@jiangnan.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 243, + 220, + 362, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 220, + 362, + 231 + ], + "spans": [ + { + "bbox": [ + 243, + 220, + 362, + 231 + ], + "type": "text", + "content": "j.kittler@surrey.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 283, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 594 + ], + "type": "text", + "content": "The shared topology of human skeletons motivated the recent investigation of graph convolutional network (GCN) solutions for action recognition. However, most of the existing GCNs rely on the binary connection of two neighboring vertices (joints) formed by an edge (bone), overlooking the potential of constructing multi-vertex convolution structures. Although some studies have attempted to utilize hyper-graphs to represent the topology, they rely on a fixed construction strategy, which limits their adaptivity in uncovering the intricate latent relationships within the action. In this paper, we address this oversight and explore the merits of an adaptive hyper-graph convolutional network (HyperGCN) to achieve the aggregation of rich semantic information conveyed by skeleton vertices. In particular, our Hyper-GCN adaptively optimises the hyper-graphs during training, revealing the action-driven multi-vertex relations. Besides, virtual connections are often designed to support efficient feature aggregation, implicitly extending the spectrum of dependencies within the skeleton. By injecting virtual connections into hyper-graphs, the semantic clues of diverse action categories can be highlighted. The results of experiments conducted on the NTU-60, NTU-120, and NW-UCLA datasets demonstrate the merits of our Hyper-GCN, compared to the state-of-the-art methods. The code is available at https://github.com/6UOON9/Hyper-GCN." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 616, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 696 + ], + "type": "text", + "content": "Skeleton-based human action recognition is a popular research topic in artificial intelligence, with practical applications in video understanding, video surveillance, human-computer interaction, robot vision, VR and AR [16, 21, 28-30, 42, 49, 49]. In general, a skeleton sequence contains" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 267, + 552, + 470 + ], + "blocks": [ + { + "bbox": [ + 318, + 267, + 552, + 470 + ], + "lines": [ + { + "bbox": [ + 318, + 267, + 552, + 470 + ], + "spans": [ + { + "bbox": [ + 318, + 267, + 552, + 470 + ], + "type": "image", + "image_path": "77062d86835fd5f1b65731af15262b34468aac1e2504409fc194e5636374c94f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 474, + 555, + 540 + ], + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 540 + ], + "type": "text", + "content": "Figure 1. Illustration of our Hyper-GCN. In (a), dotted lines represent the virtual connections. Each coloured part containing multiple joints represents a hyper-graph with hyper-edges. (b) provides an intuitive comparison between normal graph convolution and hyper-graph convolution operations with the same connectivity degree." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "content": "a series of 2D or 3D coordinates, which can easily be collected by low-cost depth sensors or obtained by video-based pose estimation algorithms [26, 34, 35, 43]. Compared to RGB and optical flow images, skeleton data, which represents the basic physical structure of a human being, is of lower dimension, conveying human action with higher efficiency. Moreover, it is robust to illumination changes and scene variations. For these reasons, the adoption of this structural data is very popular in skeleton-based action recognition [3, 6, 10, 25, 35, 45, 51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To facilitate skeleton-based action recognition, both Recurrent Neural Networks (RNNs) [10, 25, 35, 45] and Con" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12648" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "volitional Neural Networks (CNNs) [7, 17, 40] have been well explored. However, RNNs themselves cannot depict the intrinsic skeleton topology. The learned filters of CNNs, on the other hand, neglect the spatio-temporal structure of the skeleton. Drawing on these observations, recent studies have focused on how to model the skeleton topology directly. In principle, the physical topology of human joints and bones can be consistently represented by a graph. Accordingly, the graph convolutional network (GCN) is typically introduced to aggregate feature information conveyed by skeleton joints [23, 52, 57]. In terms of data representations, skeleton joints are also one-dimensional sequence data, similar to language. Therefore, with the rise of recent self-attention techniques [9, 44], several Transformer-based action recognition solutions [32, 38, 47, 53] have been released. Transformer considers skeletal joints as tokens and uses attention maps to reflect the topology. These methods have achieved excellent performance but sacrifice much higher GFLOPS and parameters than GCN." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 302, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 295, + 529 + ], + "type": "text", + "content": "In general, two neighbouring joints can pass messages through their shared bone. In graph terms, this corresponds to the exchange of information between two vertices along their connecting edge. Besides the physical skeleton topology, attempts have been made to explore the implicit relationships among joints [3, 6, 23, 36], suggesting different variants of skeleton topology. In principle, these approaches assume a binary connection between each connected vertex pair. Mathematically, the constructed topology in the adjacency matrix is represented by a normal graph. However, human actions are jointly defined by several joints. Hence, human actions encompass not only binary relations between vertex pairs but also multi-joint relationships. For instance, the action primitive starting running is manifest in the human raising the left hand while the right leg steps forward, as shown in Figure 1 (a). The binary connections are not sufficient to capture the synergistic interaction of multiple joints. This strongly argues for constructing feature aggregation paths involving multiple vertices." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 532, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 664 + ], + "type": "text", + "content": "Accordingly, we propose to construct a hyper-graph to depict the skeleton topology and take advantage of the outstanding performance of hyper-graph analysis techniques [2, 12, 18, 19]. The hyper-graph topology involves multiple node connections, rather than binary connections of a normal graph. As shown in Figure 1 (a), in a hyper-graph, a hyper-edge can link more than two vertices. This linking has the capacity to represent complex collaborative relations among human joints. As one hyper-edge associates multiple vertices, a single hyper-graph convolution enables aggregating all the features along the hyper-edges." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "An illustration is provided in Figure 1 (b), where the normal graph and hyper-graph are presented to demonstrate their differences in passing the information conveyed by the vertex features. In the normal graph convolution, after" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "2-layer aggregation, the information of the green vertex is spread to 2 other vertices. In contrast, in a hyper-graph convolution, the information of the green vertex spreads to all the vertices, creating an extended receptive field. Theoretically, by modelling a skeleton using a hyper-graph for the aggregation of joint information, improved efficiency can be obtained during learning the action semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 289 + ], + "type": "text", + "content": "However, the key to achieving enhanced information aggregation by hyper-graph convolution lies in the conformity of the hyper-graph structure to the real topology of human action. Several studies [13, 58] attempt to adopt a hypergraph to model the skeleton topology, mostly based on superficial prior knowledge and an unfounded definition of the structure of hyper-graph. Distinguishing our approach from the existing hyper-graph-based methods, we design an adaptive solution for constructing the hyper-graph with virtual connections. It can construct a Non-uniform hypergraph specific to each individual action type." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 553, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 575 + ], + "type": "text", + "content": "Besides the information aggregation issue, the capacity of an action recognition system is also constrained by its input features. In general, the number of skeleton joints is fixed in existing benchmarks, e.g., 25 for NTU-120 [26]. The entire recognition process relies on a successful capture of the interactions among the skeleton joints. In the domain of artistic puppetry, the history of driving actions, such as Shadow Play and Marionette goes back more than 2000 years. This kind of art form provides an inspiration for involving additional 'hyper joints' which can drive or facilitate a better communication between existing joints. The underlying spirit is to alleviate the pressure on the real joints to store and transfer complex semantics. Jointly with the hyper joints, real joints can focus more on storing neighbouring joint features and relegating the task of transferring the global information to the hyper joints. Interestingly, the class token in existing Transformers can also be considered as a hyper-token [9, 44]. As shown in Figure 1, the skeleton is described as a marionette, where the actions are \"controlled\" by connecting hyper joints to the real joints. This suggests that the hyper joints are not only able to capture the representation information of human action, but also reveal the implicit information between physically connected joints as hints for recognition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 575, + 553, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 553, + 671 + ], + "type": "text", + "content": "By endowing an adaptive non-uniform hyper-graph with hyper joints, virtual connections are created to perform comprehensive hyper-graph convolutions. We construct Hyper-GCN based on the above design principles. Extensive experiments on 3 datasets, NTU RGB+D 60 [35], NTU RGB+D 120 [26], and NW-UCLA [46], are conducted for evaluation. The results validate the merits of our proposed Hyper-GCN. The main contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 685 + ], + "type": "text", + "content": "- An adaptive non-uniform hyper-graph to represent the hu" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 528, + 702 + ], + "type": "text", + "content": "1https://en.wikipedia.org/wiki/Shadow Playing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 703, + 522, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 522, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 522, + 712 + ], + "type": "text", + "content": "2https://en.wikipedia.org/wiki/Marionette" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12649" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 107 + ], + "type": "text", + "content": "man skeleton topology. Compared with a fixed hypergraph, the constructed topology is more action-specific, thus boosting the discrimination." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 192 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 108, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 144 + ], + "type": "text", + "content": "- The injection of virtual hyper joints, enriching the connectivity of the physical joints, from a global semantic perspective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 296, + 192 + ], + "type": "text", + "content": "- As the processing architecture, we propose a hyper-graph convolution network (Hyper-GCN). The SOTA performance achieved on three public datasets demonstrates the merits of the Hyper-GCN and virtual connection designs." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 205, + 141, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 141, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 141, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 225, + 264, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 264, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 264, + 239 + ], + "type": "text", + "content": "2.1. Graph Topology for Action Recognition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 243, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 290 + ], + "type": "text", + "content": "A graph can represent a human skeleton, preserving the joint relationships via predefined edges. To aggregate semantics, GCNs [1, 4, 11, 20, 24, 37, 54, 57] and Transformers [8, 32, 38, 53, 55] have been well studied." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 292, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 386 + ], + "type": "text", + "content": "For GCN-based approaches, ST-GCN [52] proposes to represent the topology with an adjacency matrix of 3 subsets by modeling the spatio-temporal relevance. Similarly, 2s-AGCN [36], InfoGCN [6], and DS-GCN [51] use a self-attentive mechanism to learn the topology from joint pairs. Besides the intuitive spatio-temporal dimensions, CTR-GCN [3] proposes to learn the topology for channels to refine the skeleton." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 459 + ], + "type": "text", + "content": "For Transformer-based solutions, IIP-Transformer [47] adopt a self-attention mechanism to establish the intra-part and the inter-part joints. IGFormer [31] learns the topology between persons at both semantic and distance levels. STTFormer [33] proposed structure can capture the correlation between joints in consecutive frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 460, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 591 + ], + "type": "text", + "content": "All the above two paradigms construct the topology from joint pairs, which rely only on the binary relation between two vertices. In their implementation stage, the adjacency matrix or attention map is used to represent the topology. In this case, high-order information among joints is not taken into consideration, neglecting the collaborative power among multiple joints. Though hyper-graph is considered to construct topology using multivariate joint relationships, current [13, 58] solutions manually set the hyper-graphs, which greatly relies on human experience, sacrificing the adaptability of graph learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 295, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 613 + ], + "type": "text", + "content": "2.2. Feature Configurations for Action Recognition" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 653 + ], + "type": "text", + "content": "It has been observed that the input features play the most essential role in delivering high recognition accuracy [1, 5, 15, 27, 31, 33, 41, 47] in GCNs and Transformers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Drawing on this, Graph2Net [50] proposes to extract local and global spatial features jointly. CTR-GCN [3] uses multi-scale temporal convolution to extract temporal features. While HD-GCN [23] introduce the hierarchical edge convolution to key edge features. To extend the perception" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "field, STC-Net [22] uses the dilated kernels for graph convolution to capture the features." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 167 + ], + "type": "text", + "content": "Similarly, IG-Transformer [31] aggregate the features of two persons as one to mine the semantic information. STST [55] split the coordinate features, semantic features and temporal features into three kinds of tokens. TSTE [53] extracting and merging the spatial and temporal features as input to transformer encoders." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 168, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 263 + ], + "type": "text", + "content": "After all, existing methods [6, 14, 32, 36, 38, 51, 52] receive the input features from real skeleton joints. However, each skeleton joint acts as an information carrier during forward passing, which is required to deliver both local context and global semantics. Given a fixed representative capacity (number of channels), we believe it is necessary to involve additional virtual joints to balance the pressure of storing local and global information." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 275, + 380, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 275, + 380, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 275, + 380, + 289 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 295, + 400, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 295, + 400, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 295, + 400, + 306 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": "In general, the input features in graph convolution are represented by " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{in}}\\in \\mathbb{R}^{C\\times T\\times V}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": " represents the number of channels in the feature maps. Given the topology of a human skeleton, we usually define the graph " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = (\\mathcal{V},\\mathcal{E})" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": " represents the set of joints and " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": " represents the set of edges between joints. For the set of edges " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": ", it is formulated as an adjacent matrix " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": " represents the number of human joints. The normalised adjacency matrix is represented by " + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\hat{A}\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": ". The normalisation operation is formulated as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 396, + 439, + 553, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 439, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 396, + 439, + 553, + 454 + ], + "type": "interline_equation", + "content": "\\hat {A} = \\Lambda^ {- \\frac {1}{2}} A \\Lambda^ {- \\frac {1}{2}}, \\tag {1}", + "image_path": "f0be19b30777cae76015bdb15bf826899470e764ea1bf53d5b6e48e6e34052fe.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "inline_equation", + "content": "\\Lambda" + }, + { + "bbox": [ + 313, + 464, + 554, + 500 + ], + "type": "text", + "content": " represents the diagonal matrix stored with the degree of every joint. The entire normal graph convolution process can be formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 392, + 510, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 510, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 392, + 510, + 553, + 523 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {o u t}} = \\sigma (\\hat {A} \\mathbf {F} _ {\\text {i n}} P), \\tag {2}", + "image_path": "47ddd7d0ee85804db8cff0174a3ba329583e7690bb72f2110302a8e108ed9eae.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{C \\times C'}" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " is the learnable parameters, representing the feature transformation patterns in the feature space. " + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 532, + 554, + 567 + ], + "type": "text", + "content": " denotes the non-linear activation function ReLU." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 576, + 398, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 576, + 398, + 589 + ], + "spans": [ + { + "bbox": [ + 314, + 576, + 398, + 589 + ], + "type": "text", + "content": "3.2. Hyper-graph" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Here, we use " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\mathcal{H}} = (\\mathcal{V}_{\\mathcal{H}},\\mathcal{E}_{\\mathcal{H}},\\mathcal{W}_{\\mathcal{H}})" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " to define the spatial hyper-graph [12] with human skeleton. " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " follow similar definitions in the normal graph, which represent the set of joints and the set of hyper-edges. In addition, we introduce " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\mathcal{H}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " to represent the weights of each hyper-edge. Since a hyper-edge contains multiple joints, the corresponding topology can no longer be simply represented by an adjacency matrix. Therefore, we introduce the incidence matrix " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "H\\in \\mathbb{R}^{N\\times M}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " to describe the topology of each joint in the hyper-graph. " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " represents the number of joints and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12650" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 63, + 438, + 208 + ], + "blocks": [ + { + "bbox": [ + 59, + 63, + 438, + 208 + ], + "lines": [ + { + "bbox": [ + 59, + 63, + 438, + 208 + ], + "spans": [ + { + "bbox": [ + 59, + 63, + 438, + 208 + ], + "type": "image", + "image_path": "43bfcada4a3d50f8c88cf05b02a441690a45df1e21f4844af313cbf1e1eee388.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 438, + 302 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 438, + 302 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 438, + 302 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 438, + 302 + ], + "type": "image", + "image_path": "db5a874a144d77a1d8dac9f28bd4f697675133051f3bab7ae9a3902818567e32.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. The framework of our Hyper-GCN. Part (a) represents the Multi-head Hyper-graph Convolution (M-HGC) module. Part (b) represents the process of constructing an adaptive hyper-graph. Part (c) represents the architecture of Hyper-GCN. " + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "text", + "content": " represents the features of real joints. " + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_h" + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "text", + "content": " represents the hyper-joints which are learnable parameters. " + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\hat{A}" + }, + { + "bbox": [ + 55, + 308, + 555, + 344 + ], + "type": "text", + "content": " represents the physical topology." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 442, + 64, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 442, + 64, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 442, + 64, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 442, + 64, + 545, + 304 + ], + "type": "image", + "image_path": "2018708943c56b0aae8d52505b46ebc1f9098223d66fdeaea1080aa0cc810b3b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": "represents the number of hyper-edges. Given " + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "inline_equation", + "content": "e \\in \\mathcal{E}_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 364, + 295, + 389 + ], + "type": "text", + "content": ", the values of incidence matrix be determined by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 394, + 295, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 394, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 124, + 394, + 295, + 427 + ], + "type": "interline_equation", + "content": "h (e, v) = \\left\\{ \\begin{array}{l l} 1, & v \\in e, \\\\ 0, & v \\notin e \\end{array} \\right.. \\tag {3}", + "image_path": "9f81b8c6c5bdfb96dafaea7bff7faf44dba377c167ff424fc69e2b6751617272.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 433, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 514 + ], + "type": "text", + "content": "Similar to the normal graph convolution, it also needs to normalise the hyper-graph to modulate the aggregated features. The degree of hyper-graph consists of the degree of joints and the degree of hyper-edges. The degree of joints is represented by the sum of weights of all joints contained in each hyper-edge. Given " + }, + { + "bbox": [ + 55, + 433, + 296, + 514 + ], + "type": "inline_equation", + "content": "v \\in \\mathcal{V}_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 433, + 296, + 514 + ], + "type": "text", + "content": ", it can be described as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 514, + 295, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 514, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 120, + 514, + 295, + 540 + ], + "type": "interline_equation", + "content": "d (v) = \\sum_ {e \\in \\mathcal {E} _ {\\mathcal {H}}} w (e) h (v, e), \\tag {4}", + "image_path": "55e0713c51afe5739b75108593f42c737ce2827d2e6e5e29f85416dad6a8262a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": "where we use a diagonal matrix " + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "inline_equation", + "content": "D_{v} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": " to represent the degree of joints. " + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{M \\times M}" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": " represents the weight matrix of hyper-edges, which is formulated as a diagonal matrix. The degree of hyper-edges represents the sum of the number of joints contained in each hyper-edge. Given " + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "inline_equation", + "content": "e \\in \\mathcal{E}_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 544, + 296, + 616 + ], + "type": "text", + "content": ", it can be described as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 130, + 623, + 295, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 623, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 130, + 623, + 295, + 649 + ], + "type": "interline_equation", + "content": "d (e) = \\sum_ {v \\in \\mathcal {V} _ {\\mathcal {H}}} h (v, e), \\tag {5}", + "image_path": "273860514d58aca3dfb29600406e542b12581a8abbba4e3a3088f8f8288acf92.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "where we use a diagonal matrix " + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "inline_equation", + "content": "D_{e} \\in \\mathbb{R}^{M \\times M}" + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": " to represent the degree of hyper-edges. Therefore, the normalisation of a hyper-graph is formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 699, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 699, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 124, + 699, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {H} = D _ {v} ^ {- 1} H W D _ {e} ^ {- 1} H ^ {T}, \\tag {6}", + "image_path": "1e15bbbf104569dc98ac0b1e9b7bf25a43fe7855f806417c83d4b0fa71f68870.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 325, + 377, + 549, + 545 + ], + "blocks": [ + { + "bbox": [ + 325, + 377, + 549, + 545 + ], + "lines": [ + { + "bbox": [ + 325, + 377, + 549, + 545 + ], + "spans": [ + { + "bbox": [ + 325, + 377, + 549, + 545 + ], + "type": "image", + "image_path": "d545ccc8059f8da246488210c60d55b923c118918b9210c3bea3f832eee6557a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "text", + "content": "Figure 3. Instance of the incidence matrix construction by A-NHG with hyper-parameter " + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "text", + "content": " represents the number of joints. (a) represents the hyper-edge sets are null. (b) represents that measure the distance between 1-st joint and other joints. (c) represents that retrain " + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 555, + 555, + 621 + ], + "type": "text", + "content": "-nearest hyper-edges which contain 1-st joint. (d) represents the final state." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "inline_equation", + "content": "\\hat{H} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 639, + 553, + 665 + ], + "type": "text", + "content": " represents the normalised incidence matrix for hyper-graph convolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 507, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 507, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 507, + 685 + ], + "type": "text", + "content": "3.3. Adaptive Non-uniform Hyper-graph" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Constructing the hyper-graph incidence matrix is critical to understanding human actions. As shown in Figure 2 (b), we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12651" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "design an Adaptive Non-uniform Hyper-graph (A-NHG) Construction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "content": "Given the general features " + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "content": " of spatial skeleton joints, " + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "content": " represent the number of joints, " + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "content": " represent the channels. Our A-NHG construct the hyper-edge with each joint as the center of mass, and a total of " + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 96, + 296, + 169 + ], + "type": "text", + "content": " hyper-edges are constructed. A-NHG utilises the Euclidean distance to measure the joint difference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": "We use one mapping function " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{C\\times C_h}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": " to embed the features into the subspace for constructing the hypergraph. " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": " represents the hidden channels of the mapping subspace. This design enables preserving the original features to reflect their spatial relevance. Based on " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "X_{H}\\in \\mathbb{R}^{N\\times C_{h}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": ", we define the set of joints " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\nu_{\\mathcal{H}} = \\{v_{1}\\in" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C_h},v_2\\in \\mathbb{R}^{C_h},\\dots,v_N\\in \\mathbb{R}^{C_h}\\}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "v_{i},v_{j}\\in \\mathcal{V}_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": ", each element " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "m_{i,j}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": " in the distance matrix " + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 171, + 296, + 277 + ], + "type": "text", + "content": " can be obtained as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 292, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 292, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 118, + 292, + 295, + 306 + ], + "type": "interline_equation", + "content": "m _ {i, j} = m _ {j, i} = \\left\\| v _ {i} - v _ {j} \\right\\| _ {2}. \\tag {7}", + "image_path": "902fbd64ffaf06f3bc895c0cc01edd5faffbe3ea07300b081e407587e2c29750.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 318, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 411 + ], + "type": "text", + "content": "In principle, we need to transform the distance matrix into a probability incidence matrix, which determines whether the specific joints are contained with the same hyper-edges. In order to guarantee that the model is trainable, we use the softmax operation to assign the probabilities. However, directly using the softmax would inevitably result in each joint belonging to all hyper-edges to some extent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": "Since this outcome is undesirable, we introduce constraints on each joint to limit the number of hyper-edges it can belong to, preventing excessive connections. For one joint, we retain only the " + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": " nearest hyper-edges represented by joints in the semantic space for probability estimation. And " + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 415, + 296, + 486 + ], + "type": "text", + "content": " is the hyper-parameter of A-NHG." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": "Given the row vector " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " in the distance matrix " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": ", which represents the distances between joint " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " and the other joints. We select the indices of the minimum " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " joints from " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " and add them to the set " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "set_i" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "set_i" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "h_{i,j}" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": " in the incidence matrix " + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "inline_equation", + "content": "H\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 55, + 488, + 296, + 548 + ], + "type": "text", + "content": ", can be calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 94, + 571, + 295, + 604 + ], + "type": "interline_equation", + "content": "h _ {i, j} = \\left\\{ \\begin{array}{c c} \\frac {\\exp (- m _ {i , j})}{\\sum_ {k \\in s e t _ {i}} \\exp (- m _ {i , k})}, & j \\in s e t _ {i}, \\\\ 0, & j \\notin s e t _ {i}. \\end{array} \\right. \\tag {8}", + "image_path": "80b3f3d3112f19213d7882206e984c9ace89dc75bad4c5d7260c3d4ebe770544.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "In contrast to uniform hyper-graph (every hyper-edge contains the same number of joints), our approach limits the maximum number of every joint that can be contained by hyper-edges. Obviously, the number of joints contained in each hyper-edge constructed in this way is non-fixed, allowing the hyper-edges to capture more differentiated associations. The entire process for example with " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " is illustrated in the Figure 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 514, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 514, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 514, + 85 + ], + "type": "text", + "content": "3.4. Multi-head Hyper-graph Convolution" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 137 + ], + "type": "text", + "content": "To accommodate the semantic information reflected by different group channels, we further propose the Multi-head Hyper-graph Convolution (M-HGC), as shown in Figure 2 (a)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 137, + 554, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 554, + 220 + ], + "type": "text", + "content": "We use separate branches to independently perform hyper-graph convolution on the topologies represented by the multi-head hyper-graphs, thereby enhancing the computational efficiency of the Hyper-GCN. For instance, the features " + }, + { + "bbox": [ + 313, + 137, + 554, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times T \\times V}" + }, + { + "bbox": [ + 313, + 137, + 554, + 220 + ], + "type": "text", + "content": " are divided into 8 separate branches along the channel dimension, which are processed by 8 M-HGC parallelly, delivering 8 separate hyper-graphs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 221, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 281 + ], + "type": "text", + "content": "Next, we perform temporal pooling, and then the obtained spatial information is the evidence to determine the optimal hyper-graph construction. This can efficiently decouple the temporal and spatial clues. After performing temporal average-pooling, we can obtain " + }, + { + "bbox": [ + 313, + 221, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{F}}_{\\mathrm{in}} \\in \\mathbb{R}^{C_{\\mathrm{in}} \\times V}" + }, + { + "bbox": [ + 313, + 221, + 553, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 387 + ], + "type": "text", + "content": "The hyper-graph is depicted by the incidence matrix and weight matrix. So, we use a separate mapping function for each head to embed the features into the subspace for constructing the hyper-graph. This design enables preserving the original features to reflect their spatial relevance. After that, we introduce the A-NHG to obtain the incidence matrix. For the weight matrix, we adopt MLP to measure each hyper-edge. The hyper-graph obtained operation can be formulated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 392, + 553, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 392, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 365, + 392, + 553, + 407 + ], + "type": "interline_equation", + "content": "H = \\uplus_ {k = 1} ^ {8} \\mathrm {A} - \\mathrm {N H G} \\left(\\bar {\\mathbf {F}} _ {\\mathbf {i n}} ^ {k}\\right), \\tag {9}", + "image_path": "c314da9df1115a729cac5ad4d61a49153a012d24c381ba87a961215ff6beba9c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 409, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 409, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 364, + 409, + 553, + 423 + ], + "type": "interline_equation", + "content": "W = \\sigma_ {2} \\left(\\Psi_ {2} \\left(\\uplus_ {k = 1} ^ {8} \\sigma_ {1} \\left(\\Psi_ {1} ^ {k} \\left(\\overline {{\\mathbf {F}}} _ {\\mathbf {i n}} ^ {k}\\right)\\right)\\right)\\right), \\tag {10}", + "image_path": "dacc78ec96c8b22a51d0bd273d63ad5a77880d05dd11a4bc65fe4ae2c9f1966c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\Psi_{1} \\in \\mathbb{R}^{C_{\\mathrm{in}} / 8 \\times C_{h}}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\Psi_{2} \\in \\mathbb{R}^{(8 \\times C_{h}) \\times 8}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " are the mapping functions, which are set as learnable parameters. " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "C_{h}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " represents the hidden channels of the mapping subspace. " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " represents the channels " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "[(k - 1) \\times C_{\\mathrm{in}} / 8 + 1, \\dots, k \\times C_{\\mathrm{in}} / 8]" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{F}}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\uplus" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " represents the channel concatenation. " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\sigma_{1}, \\sigma_{2}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": " denote the activation function LeakeyReLU and Tanh to obtain the weight of each hyper-edge, limiting the values with the range of " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": ". Based on " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "H \\in \\mathbb{R}^{8 \\times N \\times N}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{8 \\times N}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": ", normalise the hyper-graph by sec.3.2 to obtain the " + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\hat{H} \\in \\mathbb{R}^{8 \\times N \\times N}" + }, + { + "bbox": [ + 313, + 428, + 554, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "text", + "content": "Besides the hyper-graph, we incorporate the physical topology to emphasise the natural physical relations of a human being's skeleton. To achieve this objective, existing methods [3, 23, 36, 52] divide the physical topology of the human body into 3 subsets. They are represented by " + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "inline_equation", + "content": "S = \\{s_{id}, s_{cf}, s_{cp}\\}" + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "inline_equation", + "content": "s_{id}, s_{cf}" + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "inline_equation", + "content": "s_{cp}" + }, + { + "bbox": [ + 313, + 549, + 554, + 692 + ], + "type": "text", + "content": " denote the identity, centrifugal, and centripetal joint subsets. In our M-HGC, to ensure that the integrated physical topology in each head remains complete, we merge them into a single set. 8 heads are aggregated by channel concatenation after the hyper-graph convolutions, as shown in Figure 2 (a). The operation of M-HGC can be formulated as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 359, + 696, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 696, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 359, + 696, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {o u t}} = \\uplus_ {k = 1} ^ {8} \\left(\\hat {A} _ {k} + \\alpha \\cdot \\hat {H} _ {k}\\right) \\mathbf {F} _ {\\text {i n}} ^ {k} P _ {k}, \\tag {11}", + "image_path": "a29082b8a30787ccb54a6485201bca31434d3915eb3b9debf884f5c1b7085282.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12652" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\hat{A}" + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\hat{H}" + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "content": " denote the normalized physical adjacency matrix and normalized incidence matrix. " + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "content": " is the learnable weight parameter for feature transform. " + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 71, + 294, + 120 + ], + "type": "text", + "content": " is a learnable parameter for the topology fusion of each head." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 173, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 173, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 173, + 138 + ], + "type": "text", + "content": "3.5. Virtual Connections" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 240 + ], + "type": "text", + "content": "It is worth emphasising that incorporating learnable joints among different samples is essential for enhancing the model capacity, as these learnable joints capture generalised features of human actions. This not only enriches the semantic information but also facilitates easier interaction connections among real joints. Therefore, we introduce the hyper-joints which are to participate in the hyper-graph convolution as shown in Figure 2 (a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "text", + "content": "The shape of hyper-joints is consistent with the features of physical joints " + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{p}} \\in \\mathbb{R}^{C \\times T \\times V}" + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "text", + "content": ", which is described as " + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{h}} \\in \\mathbb{R}^{C \\times T \\times V_{h}}" + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 55, + 240, + 295, + 395 + ], + "type": "text", + "content": " represents the number of hyper-joints). To ensure alignment in the temporal dimension, the hyper-joints are shared across each frame. Specifically, the learning of hyper-joints is supervised by the loss function, which guides these hyper-joints to support generalisable driven features embedded within large amounts of data. Furthermore, we set independent hyper-joints at each layer of Hyper-GCN, aiming to harmonise features at different depths. Typically, these hyper-joints are involved in spatial hyper-graph convolution rather than in temporal convolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": "To diversify these hyper-joints, we propose the Divergence Loss for hyper-joints optimisation to mitigate their homogenisation. We adopt a cosine matrix " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{V_h \\times V_h}" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": " to measure the differences between hyper-joints, which can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 462, + 294, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 462, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 146, + 462, + 294, + 490 + ], + "type": "interline_equation", + "content": "C = \\frac {\\mathbf {F} _ {\\mathrm {h}} \\mathbf {F} _ {\\mathrm {h}} ^ {T}}{| | \\mathbf {F} _ {\\mathrm {h}} | | ^ {2}}, \\tag {12}", + "image_path": "6c7b7a822cdb791c65df0f01b297b14665e16882e16e0b9d4ccb3ad35c94cea8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "content": "Specifically, in Divergence Loss, we measure the differences of hyper-joints with the mean of the cosine matrix " + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "content": " in each layer. Since the correlation of hyper-joints themselves cannot be optimised, we manually eliminate this part by minus the " + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 55, + 497, + 295, + 557 + ], + "type": "text", + "content": ". The loss calculation can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 564, + 294, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 564, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 65, + 564, + 294, + 609 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} (C) = \\frac {\\sum_ {i = 1} ^ {V _ {h}} \\sum_ {j = 1} ^ {V _ {h}} R e L U \\left(c _ {i , j}\\right) - V _ {h}}{V _ {h} \\left(V _ {h} - 1\\right)}, \\quad c _ {i, j} \\in C, \\tag {13}", + "image_path": "37cac8de0dc96bb5f2a4841894be805ce0290e5fb8d3a5afe750d05e0185c2a0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 613, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 613, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 85, + 613, + 295, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {C E} + \\frac {1}{L} \\sum_ {l = 1} ^ {L} L _ {h} \\left(C _ {l}\\right), \\tag {14}", + "image_path": "a82a566c02bb31f05928656f5ae58d819be64c5bef8f0c987887db5fc9033b2b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "V_{h}" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " represents the number of hyper-joints we introduced. " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " represents the cosine matrix of the " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "-th layer. " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " represents the cross-entropy loss. Additionally, we manually connect the hyper joints to all the physical joints, which optimises the topology from the datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 429, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 429, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 429, + 83 + ], + "type": "text", + "content": "3.6. Entire Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 209 + ], + "type": "text", + "content": "Hyper-GCN consists of an embedding layer [6], 9 spatial-temporal convolution layers as shown in Figure 2 (c). Each layer consists of the proposed M-HGC and Multi Scale Temporal convolution (MS-TC) [3]. 9 Layers are categorised into 3 stages. In each stage, we introduce the dense connections to integrate deep and shallow features in each stage. Additionally, to further validate the potential of Hyper-GCN, we propose a base version and a large version. The channels in each stage are set to 128, 256, 256 for the base version and 128, 256, 512 for the large version." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 220, + 385, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 385, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 385, + 232 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 239, + 376, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 376, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 376, + 251 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 339 + ], + "type": "text", + "content": "NTU-RGB+D 60 & 120 NTU-RGB+D 60 [35] is a large dataset widely used in skeleton-based human action recognition, which is categorised into 60 classes. NTU-RGB+D 120 [26] is an extension to 120 classes of NTU 60. 4 benchmarks recommended by the official are adopted: (1) NTU60-XSub, (2) NTU60-XView, (3) NTU120-XSub, (4) NTU120-Xset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 389 + ], + "type": "text", + "content": "Northwestern-UCLA. The Northwestern-UCLA (NWUCLA) dataset [46] contains 1494 video clips, which is categorised into 10 classes. It contains 3 different camera views and is performed by 10 actors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 396, + 447, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 447, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 447, + 408 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "content": "Our training and evaluation stages are on a single GPU RTX 3090. In the training phase, Hyper-GCN is optimised by Stochastic Gradient Descent (SGD) with Nesterov momentum set at 0.9 and a weight decay at 0.0004. Our implementation uses label smooth cross-entropy loss with the Divergence Loss we proposed. We set a total of 140 epochs with the start 5 warm-up epochs. The initial learning rate is 0.05, which is reduced to 0.005 at epoch 110 and to 0.0005 at epoch 120." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 529, + 512, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 512, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 512, + 541 + ], + "type": "text", + "content": "4.3. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 605 + ], + "type": "text", + "content": "Multi-stream ensemble proposed in [39] have been proven effective by most of the existing state-of-the-art methods. We also use a 4-stream ensemble to evaluate the base version and the large version of Hyper-GCN. The detailed results are reported in Table 1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "type": "text", + "content": "In summary, the base version of Hyper-GCN comprehensively outperforms all GCN-based and HGCN-based (hyper-graph utilised) SOTA and surpasses the Transformer-based SOTA on the NTU120, with a most lightweight design. Moreover, when scaled to match the parameters of the lightest Transformer-based method, the large version of Hyper-GCN achieves 1-st place in 4 benchmarks and 2-nd place in 1 benchmark. This demonstrates the effectiveness and merits of Hyper-GCN." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12653" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 69, + 549, + 274 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 549, + 274 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 549, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 549, + 274 + ], + "type": "table", + "html": "
CategoryMethodsModalitiesParams (M)GFLOPsNTU-RGB+D 60NTU-RGB+D 120NW-UCLA (%)
X-Sub (%)X-View (%)X-Sub (%)X-Set (%)
GCNST-GCN [52]J+B--81.588.370.773.2-
2s-AGCN [36]J+B--88.595.182.584.2-
DC-GCN+ADG [4]J+B+JM+BM4.91.8390.896.686.588.195.3
MS-G3D [27]J+B+JM+BM2.85.2291.596.286.988.4-
MST-GCN [11]J+B+JM+BM12.0-91.596.687.588.8-
CTR-GCN [3]J+B+JM+BM1.51.9792.496.488.990.696.5
EfficientGCN-B4 [41]J+B+JM+BM2.015.2091.795.788.389.1-
InfoGCN [6]J+B+JM+BM1.61.8492.796.989.490.796.6
FR Head [56]J+B+JM+BM2.0-92.896.889.590.996.8
HD-GCN* [23]J+B+J' +B'1.71.7793.097.089.891.296.9
DS-GCN [51]J+B+JM+BM--93.197.589.290.3-
BlockGCN [57]J+B+JM+BM1.31.6393.197.090.391.596.9
TransformerDSTA-Net [38]J+B+JM+BM3.516.1889.595.786.689.0-
IIP-Transformer [47]J+B+JM+BM2.97.2089.595.789.990.9-
SkateFormer [8]J+B+JM+BM2.03.6293.597.889.891.498.3
HGCNHyper-GNN [13]J+B+JM+BM--89.595.7---
Selective-HCN [58]J+B+JM+BM--90.896.6---
DST-HCN [48]J+B+JM+BM3.52.9392.396.888.890.796.6
Ours (B)J+B+JM+BM1.11.6393.397.490.591.797.2
Ours (L)J+B+JM+BM2.32.8893.797.890.992.097.6
", + "image_path": "1da4c5a2bf7868ad087bb0041fa18f3c5005a9a738d2d6c1e21c17333e4957c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 345, + 282, + 445 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 327 + ], + "type": "text", + "content": "Table 1. Comparison of Hyper-GCN with advanced solutions on NTU-RGB+D 60, NTU-RGB+D 120, and NW-UCLA datasets. For a fair comparison, we use a 4-streams ensemble to evaluate the performance of the mentioned methods. J, B, JB, and JM represent the joint, bone, joint motion, and bone motion. In addition, for HD-GCN [23], we use the results from 4-streams according to the original papers for a fair comparison. The bold font represents the best result and underline font represents the 2-nd best result." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 345, + 282, + 445 + ], + "lines": [ + { + "bbox": [ + 70, + 345, + 282, + 445 + ], + "spans": [ + { + "bbox": [ + 70, + 345, + 282, + 445 + ], + "type": "table", + "html": "
NTU RGB+D 120 X-Sub (%)
KUniformNon-uniform
Baseline-84.784.7
w M-HGC386.4 (↑ 1.7)86.0 (↑ 1.3)
586.5 (↑ 1.9)86.2 (↑ 1.5)
786.3 (↑ 1.6)86.5 (↑ 1.8)
986.0 (↑ 1.3)86.7 (↑ 2.2)
1185.9 (↑ 1.2)86.4 (↑ 1.7)
", + "image_path": "16c7aea1e5872b4caac64b6aa9f9c50926a7c837748986f5c7e294076dc4f451.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "type": "text", + "content": "Table 2. Ablation of the hyper-parameter " + }, + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "type": "text", + "content": " in A-NHG. The Uniform represents K-uniform hyper-graph, in which one hyper-edge contains K joints. The non-uniform represents non-uniform hyper-graph, in which one joint is contained by " + }, + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 453, + 295, + 498 + ], + "type": "text", + "content": " hyper-edge." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 523, + 149, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 149, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 149, + 536 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 579 + ], + "type": "text", + "content": "We conduct ablation experiments with visualisations to analyse the effectiveness of our design. redAll the ablation study is evaluated with the joint modality of base version." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Hyper-graph Construction: The Non-uniform hypergraphs with hyper-parameter " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " represent the maximum number of hyper-edges contained in each joint. They reflect the varying degrees of aggregation in the constructed topology. In general, it is common to adopt a uniform hypergraph with hyper-parameter " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ". Therefore, to explore how to set the hyper-parameter " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for the adaptive hyper-graph in M-HGC, we design the ablation experiments to investigate the impact of hyper-parameter " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " on the model's recognition ability, where both uniform and non-uniform hypergraphs are considered." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 350, + 555, + 585 + ], + "blocks": [ + { + "bbox": [ + 322, + 350, + 555, + 585 + ], + "lines": [ + { + "bbox": [ + 322, + 350, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 322, + 350, + 555, + 585 + ], + "type": "image", + "image_path": "f8ec9720ed8344ce63630c6b4ae80986a1ae8f38b9813b78118c15cd64f21be9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 592, + 555, + 625 + ], + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 625 + ], + "type": "text", + "content": "Figure 4. Visualisation of hyper-graph in selected actions. For ease of observation, the joints contained in the selected hyper-edge are highlighted in red." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "The results are reported in Table 2. In both uniform and non-uniform hyper-graphs, a very large " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " means that more joints are contained in each hyper-edge. Even though these \"extra\" joints may be optimised for minimal weight, they are likely to act as \"noise\", disrupting the information inter" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12654" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 288, + 182 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 182 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 182 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 182 + ], + "type": "table", + "html": "
NTU RGB+D 120 X-Sub (%)
w/o Lhw/ Lh
Baseline84.784.7
w/o M-HGCw/ 184.9 (↑ 0.2)84.9 (↑ 0.2)
w/ 384.9 (↑ 0.2)85.2 (↑ 0.5)
w/ 584.785.0 (↑ 0.3)
w M-HGCw/ 186.7 (↑ 2.0)86.7 (↑ 2.0)
w/ 386.6 (↑ 1.8)86.9 (↑ 2.2)
w/ 586.6 (↑ 1.8)86.8 (↑ 2.1)
", + "image_path": "113ebd7d3b448f5d1863827e00043c445b6b469533c8b14a31c40bf096c5219c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "text", + "content": "Table 3. Ablations on hyper joints. w/ " + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "text", + "content": " represents the number of hyper joints in each layer. w/ " + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_h" + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "text", + "content": " represents the training with Divergence Loss. The hyper-parameter " + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 190, + 295, + 224 + ], + "type": "text", + "content": " of M-HGC is set to 9." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 61, + 252, + 293, + 343 + ], + "blocks": [ + { + "bbox": [ + 61, + 252, + 293, + 343 + ], + "lines": [ + { + "bbox": [ + 61, + 252, + 293, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 252, + 293, + 343 + ], + "type": "image", + "image_path": "ac8795de8fa7bceae5fa8234961cef74c34fc94ea442eaa851fa91f049d20355.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 353, + 295, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 353, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 295, + 387 + ], + "type": "text", + "content": "Figure 5. Comparison of the cosine matrices of hyper-joints. Part (a) and (b) represents w/o Divergence Loss and w/ Divergence Loss with 3 hyper-joints. It is calculated by Eqn. (12)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": "action represented by the hyper-edge. As the " + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": " decreases, it is difficult for the hyper-edge to represent complex joint combination patterns. Therefore, setting " + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "inline_equation", + "content": "K = 5" + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": " in uniform and " + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "inline_equation", + "content": "K = 9" + }, + { + "bbox": [ + 55, + 411, + 295, + 495 + ], + "type": "text", + "content": " in Non-uniform achieves the best performance. The non-uniform hyper-graph of A-NHG outperforms the uniform hyper-graph, as the combination of skeleton joints in different actions is diverse." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 496, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 652 + ], + "type": "text", + "content": "Furthermore, we visualise the hyper-graph inferred by Hyper-GCN on selected actions in Figure 4. The hypergraph construction focuses on the joints that are most relevant to the corresponding action category. For \"Kicking something\" as part (d), the left leg joints and the right hand joints are modelled as a hyper-edge. Because in this action, the person's right hand is previously extended forward, and the left leg swings relatively to the right hand. Similarly, in the action \"Stand up\" as shown in part (b), the hand and foot joints serve as fixed joints, while the remaining body joints move relative to them. In addition, \"Taking a selfie\" and \"Drinking water\" both focus on the main joints of movement with the above view as Figure 4 (a), (c)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "Virtual Connections: We conduct the ablation study to analyse the hyper joints and the Divergence Loss. The results are listed in Table 3. By comparison, the performance is not monotonically increasing with the number of hyper joints. The case of introducing only 3 hyper joints achieves" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 77, + 430, + 158 + ], + "blocks": [ + { + "bbox": [ + 322, + 77, + 430, + 158 + ], + "lines": [ + { + "bbox": [ + 322, + 77, + 430, + 158 + ], + "spans": [ + { + "bbox": [ + 322, + 77, + 430, + 158 + ], + "type": "image", + "image_path": "5bff5685e0e431a8763f1c18c1bb16666f8f38873140dbf55e8c2008dd9bcb0b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 167, + 554, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 201 + ], + "type": "text", + "content": "Figure 6. t-SNE visualisation of Baseline and Hyper-GCN. Different colours represent different joints. Part (a) and (b) represent the output features in Baseline and Hyper-GCN." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 76, + 545, + 159 + ], + "blocks": [ + { + "bbox": [ + 434, + 76, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 434, + 76, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 434, + 76, + 545, + 159 + ], + "type": "image", + "image_path": "d671b0d9050a0917b58d1dc97e6cb0bf33377a7c33cc04845daa407fe0c2188a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 223, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 270 + ], + "type": "text", + "content": "the best performance. As our hyper joints are learned from a large amount of data, involving a large number of hyper joints can introduce redundant and ambiguous clues, degrading the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 272, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 390 + ], + "type": "text", + "content": "In addition, we visualise the cosine matrices of HyperGCN, as shown in Figure 5. Clearly, in the absence of the Divergence Loss, the homogenisation of the learned hyper-joints is severe. This impedes the ability of hyper-joints to represent the generalised features of human actions. The observation further validates the effectiveness of the Divergence Loss for hyper-joint optimisation, supporting our viewpoint that generalised features with a certain degree of differentiation are needed to participate in hyper-graph convolution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 513 + ], + "type": "text", + "content": "Effectiveness of Hyper-GCN: To further validate that Hyper-GCN increases the efficiency of information interaction, we perform t-SNE projections on the output features in last layer between Baseline and Hyper-GCN, as shown in Figure 6 for comparison. Notice that the semantic information represented by the last Hyper-GCN layer is very convergent. This represents that the semantic information of each joint is adequately conveyed in M-HGC. The joints with similar semantics can prevent sacrificing the information after global average pooling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 388, + 537 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "In this paper, we propose an Adaptive Hyper-graph Convolutional Network for skeleton-based human action recognition. To exploit the implicit topology of multivariate synergy between joints, we introduce the Adaptive Nonuniform Hyper-graph, the Multi-head Hyper-graph Convolution and virtual connections. We carry out experiments on the dataset NTU-RGB+D 60 & 120, NW-UCLA to validate the effectiveness of Hyper-GCN. The experimental analysis verifies that our design can improve the recognition performance. To the best of our knowledge, Hyper-GCN achieves the SOTA performance on 3 public datasets. The involvement of adaptive non-uniform hypergraph modelling indeed extends existing GCN-based action recognition paradigms." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12655" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgment This work was supported in part by the National Natural Science Foundation of China (62020106012, 62106089, 62332008, 62336004), the Engineering and Physical Sciences Research Council (EPSRC), U.K. (Grants EP/V002856/1, and EP/T022205/1), and the Fundamental Research Funds for the Central Universities (JUSRP202504007)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 169, + 115, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 169, + 115, + 181 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 115, + 181 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 189, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 189, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 189, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 294, + 232 + ], + "type": "text", + "content": "[1] Tasweer Ahmad, Lianwen Jin, Luojun Lin, and GuoZhi Tang. Skeleton-based action recognition using sparse spatiotemporal gcn with edge effective resistance. Neurocomputing, 423:389-398, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 235, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 294, + 267 + ], + "type": "text", + "content": "[2] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention. Pattern Recognition, 110:107637, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 313 + ], + "type": "text", + "content": "[3] Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In ICCV, pages 13339-13348, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 295, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 359 + ], + "type": "text", + "content": "[4] Ke Cheng, Yifan Zhang, Congqi Cao, Lei Shi, Jian Cheng, and Hanqing Lu. Decoupling gcn with dropgraph module for skeleton-based action recognition. In ECCV, page 536-553, Berlin, Heidelberg, 2020. Springer-Verlag. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 404 + ], + "type": "text", + "content": "[5] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In CVPR, pages 180-189, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 407, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 295, + 460 + ], + "type": "text", + "content": "[6] Hyung-Gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. Infogcn: Representation learning for human skeleton-based action recognition. In CVPR, pages 20154-20164, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 463, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 294, + 496 + ], + "type": "text", + "content": "[7] Guilhem Chéron, Ivan Laptev, and Cordelia Schmid. P-cnn: Pose-based cnn features for action recognition. In ICCV, pages 3218-3226, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 294, + 531 + ], + "type": "text", + "content": "[8] Jeonghyeok Do and Munchurl Kim. Skateformer: skeletal-temporal transformer for human action recognition. In ECCV, pages 401-420. Springer, 2024. 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 294, + 597 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 633 + ], + "type": "text", + "content": "[10] Yong Du, Wei Wang, and Liang Wang. Hierarchical recurrent neural network for skeleton based action recognition. In CVPR, pages 1110-1118, 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[11] Dong Feng, Zhongcheng Wu, Jun Zhang, and Tingting Ren. Multi-scale spatial temporal graph neural network for skeleton-based action recognition. IEEE Access, 9:58256-58265, 2021. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[12] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. TPAMI, 45(3):3181-3199, 2023. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Xiaoke Hao, Jie Li, Yingchun Guo, Tao Jiang, and Ming Yu. Hypergraph neural network for skeleton-based action recognition. TIP, 30:2263-2275, 2021. 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[14] Ruijie Hou, Zhihao Wang, Ruimin Ren, Yang Cao, and Zhao Wang. Multi-channel network: Constructing efficient gcn baselines for skeleton-based action recognition. Computers & Graphics, 110:111-117, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "text", + "content": "[15] Linjiang Huang, Yan Huang, Wanli Ouyang, and Liang Wang. Part-level graph convolutional network for skeleton-based action recognition. In AAAI, pages 11045-11052, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 199, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 266 + ], + "type": "text", + "content": "[16] Cloe Huesser, Simon Schubiger, and Arzu Öltekin. Gesture interaction in virtual reality: A low-cost machine learning system and a qualitative assessment of effectiveness of selected gestures vs. gaze and controller interaction. In International Conference on Human-Computer Interaction, page 151–160, Berlin, Heidelberg, 2021. Springer-Verlag. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "type": "text", + "content": "[17] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. TPAMI, 35(1):221-231, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 302, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 302, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 302, + 553, + 335 + ], + "type": "text", + "content": "[18] Jianwen Jiang, Yuxuan Wei, Yifan Feng, Jingxuan Cao, and Yue Gao. Dynamic hypergraph neural networks. In IJCAI, pages 2635-2641, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 337, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 381 + ], + "type": "text", + "content": "[19] George Karypis, Rajat Aggarwal, Vipin Kumar, and Shashi Shekhar. Multilevel hypergraph partitioning: Application in vlsi domain. In Proceedings of the 34th annual Design Automation Conference, pages 526-529, 1997. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 427 + ], + "type": "text", + "content": "[20] Matthew Korban and Xin Li. Ddgen: A dynamic directed graph convolutional network for action recognition. In ECCV, page 761-776, Berlin, Heidelberg, 2020. Springer-Verlag. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 429, + 555, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 555, + 484 + ], + "type": "text", + "content": "[21] Bokyung Lee, Michael Lee, Pan Zhang, Alexander Tessier, and Azam Khan. Semantic human activity annotation tool using skeletonized surveillance videos. In Proceedings of the 2019 ACM International Symposium on Wearable Computers, pages 312-315, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "type": "text", + "content": "[22] Jungho Lee, Minhyeok Lee, Suhwan Cho, Sungmin Woo, Sungjun Jang, and Sangyoun Lee. Leveraging spatiotemporal dependency for skeleton-based action recognition. In ICCV, pages 10221-10230, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 575 + ], + "type": "text", + "content": "[23] Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. In ICCV, pages 10410-10419, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 621 + ], + "type": "text", + "content": "[24] Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian. Actional-structural graph convolutional networks for skeleton-based action recognition. In CVPR, pages 3595–3603, 2019. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[25] Jun Liu, Amir Shahroudy, Dong Xu, Alex C. Kot, and Gang Wang. Skeleton-based action recognition using spatiotemporal LSTM network with trust gates. TPAMI, 40(12): 3007-3021, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C. Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. TPAMI, 42(10):2684-2701, 2020. 1, 2, 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12656" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang. Disentangling and unifying graph convolutions for skeleton-based action recognition. In CVPR, pages 140-149, 2020. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[28] Ashish S. Nikam and Aarti G. Ambekar. Sign language recognition using image based hand gesture recognition techniques. In 2016 Online International Conference on Green Engineering and Technologies, pages 1-5, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 236 + ], + "type": "text", + "content": "[29] Cosmas Ifeanyi Nwakanma, Fabliha Bushra Islam, Mareska Pratiwi Maharani, Dong-Seong Kim, and JaeMin Lee. Iot-based vibration sensor data collection and emergency detection classification using long short term memory. In International Conference on Artificial Intelligence in Information and Communication, pages 273-278, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[30] Batyrkhan Omarov, Sergazy Narynov, Zhandos Zhumanov, Aidana Gumar, and Mariyam Khassanova. A skeleton-based approach for campus violence detection. Computers, Materials & Continua, 72(1), 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 335 + ], + "type": "text", + "content": "[31] Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. In European Conference on Computer Vision, pages 605-622. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 402 + ], + "type": "text", + "content": "[32] Chiara Plizzari, Marco Cannici, and Matteo Matteucci. Spatial temporal transformer network for skeleton-based action recognition. In Pattern recognition. ICPR international workshops and challenges: virtual event, January 10–15, 2021, Proceedings, Part III, pages 694–701. Springer, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 436 + ], + "type": "text", + "content": "[33] Helei Qiu, Biao Hou, Bo Ren, and Xiaohua Zhang. Spatiotemporal tuples transformer for skeleton-based action recognition. arXiv preprint arXiv:2201.02849, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[34] Megani Rajendran, Chek Tien Tan, Indriyati Atmosukarto, Aik Beng Ng, and Simon See. Review on synergizing the metaverse and ai-driven synthetic data: enhancing virtual realms and activity recognition in computer vision. Visual Intelligence, 2(1):27, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[35] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In CVPR, pages 1010-1019, 2016. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 568 + ], + "type": "text", + "content": "[36] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In CVPR, pages 12018-12027, 2019. 2, 3, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[37] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In CVPR, pages 7904-7913, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 646 + ], + "type": "text", + "content": "[38] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Decoupled spatial-temporal attention network for skeleton-based action-gesture recognition. In Proceedings of the Asian conference on computer vision, 2020. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks. TIP, 29:9532-9545, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[40] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "NIPS, page 568-576, Cambridge, MA, USA, 2014. MIT Press. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[41] Yi-Fan Song, Zhang Zhang, Caifeng Shan, and Liang Wang. Constructing stronger and faster baselines for skeleton-based action recognition. TPAMI, 45(2):1474-1488, 2022. 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[42] Ahmed Taha, Hala H Zayed, ME Khalifa, and El-Sayed M El-Horbaty. Skeleton-based human activity recognition for video surveillance. International Journal of Scientific & Engineering Research, 6(1):993-1004, 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "text", + "content": "[43] Jianwei Tang, Jieming Wang, and Jian-Fang Hu. Predicting human poses via recurrent attention network. Visual Intelligence, 1(1):18, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[45] Vivek Veeriah, Naifan Zhuang, and Guo-Jun Qi. Differential recurrent neural networks for action recognition. In ICCV, pages 4041-4049, 2015. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[46] Jiang Wang, Xiaohan Nie, Yin Xia, Ying Wu, and Song-Chun Zhu. Cross-view action modeling, learning, and recognition. In CVPR, page 2649-2656, USA, 2014. IEEE Computer Society. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 370 + ], + "type": "text", + "content": "[47] Qingtian Wang, Shuze Shi, Jiabin He, Jianlin Peng, Tingxi Liu, and Renliang Weng. Iip-transformer: Intra-inter-part transformer for skeleton-based action recognition. In BigData, pages 936-945, 2023. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "text", + "content": "[48] Shengqin Wang, Yongji Zhang, Hong Qi, Minghao Zhao, and Yu Jiang. Dynamic spatial-temporal hypergraph convolutional network for skeleton-based action recognition. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 2147-2152. IEEE, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "text", + "content": "[49] Yizhe Wang, Congqi Cao, and Yanning Zhang. Visualsemantic network: a visual and semantic enhanced model for gesture recognition. Visual Intelligence, 1(1):25, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "text", + "content": "[50] Cong Wu, Xiao-Jun Wu, and Josef Kittler. Graph2net: Perceptually-enriched graph learning for skeleton-based action recognition. TCSVT, 32(4):2120-2132, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 536 + ], + "type": "text", + "content": "[51] Jianyang Xie, Yanda Meng, Yitian Zhao, Anh Nguyen, Xiaoyun Yang, and Yalin Zheng. Dynamic semantic-based spatial graph convolution network for skeleton-based human action recognition. AAAI, 38(6):6225-6233, 2024. 1, 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 568 + ], + "type": "text", + "content": "[52] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In AAAI, 2018. 2, 3, 5, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[53] Hengqi Zhang, Hua Geng, and Geng Yang. Two-stream transformer encoders for skeleton-based action recognition. In International Conference on Computing, Control and Industrial Engineering, pages 272-281. Springer, 2021. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "text", + "content": "[54] Pengfei Zhang, Cuiling Lan, Wenjun Zeng, Junliang Xing, Jianru Xue, and Nanning Zheng. Semantics-guided neural networks for efficient skeleton-based human action recognition. In CVPR, pages 1109-1118, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "text", + "content": "[55] Yuhan Zhang, Bo Wu, Wen Li, Lixin Duan, and Chuang Gan. Stst: Spatial-temporal specialized transformer for skeleton-based action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3229-3237, 2021. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12657" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 206 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "[56] Huanyu Zhou, Qingjie Liu, and Yunhong Wang. Learning discriminative representations for skeleton based action recognition. In CVPR, pages 10608-10617, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[57] Yuxuan Zhou, Xudong Yan, Zhi-Qi Cheng, Yan Yan, Qi Dai, and Xian-Sheng Hua. Blockgen: Redefined topology awareness for skeleton-based action recognition. In CVPR, pages 2049-2058, 2024. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 206 + ], + "type": "text", + "content": "[58] Yiran Zhu, Guangji Huang, Xing Xu, Yanli Ji, and Fumin Shen. Selective hypergraph convolutional networks for skeleton-based action recognition. In ICMR, page 518-526, New York, NY, USA, 2022. Association for Computing Machinery. 2, 3, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12658" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_content_list.json b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f44aa9f79280a034b97ed3202be9c1ec61bd8b4c --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_content_list.json @@ -0,0 +1,1758 @@ +[ + { + "type": "text", + "text": "Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tao Lei $^{1*}$ , Ziyao Yang $^{1}$ , Xingwu Wang $^{2}$ , Yi Wang $^{1}$ , Xuan Wang $^{3}$ , Feiman Sun $^{4}$ , Asoke K. Nandi $^{5}$ $^{1}$ Shaanxi University of Science and Technology, China \n $^{2}$ Beijing University of Posts and Telecommunications, China \n $^{3}$ Northwestern Polytechnical University, China \n $^{4}$ Southwest Jiaotong University, China $^{5}$ Brunel University of London, UK", + "bbox": [ + 112, + 202, + 883, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "*Corresponding author:leitao@sust.edu.cn", + "bbox": [ + 318, + 294, + 673, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing semi-supervised learning methods typically mitigate the impact of unreliable predictions by suppressing low-confidence regions. However, these methods fail to explore which regions hold higher learning value and how to design adaptive learning strategies for these regions. To address these issues, we propose a novel adaptive learning of high-value regions (ALHVR) framework. By exploiting the diversity of predictions from dual-branch networks, the prediction regions are classified into three groups: reliable stable region, reliable unstable region, and unreliable stable region. For high-value regions (reliable unstable region and unreliable stable region), different training strategies are designed. Specifically, for reliable unstable region, we propose a confidence-guided cross-prototype consistency learning (CG-CPCL) module, which enforces prototype consistency constraints in the feature space. By leveraging confidence information, the high-confidence predictions from one network selectively supervise the low-confidence predictions from the other, thus helping the model learn inter-class discrimination more stably. Additionally, for unreliable stable region, we design a dynamic teacher competition teaching (DTCT) module, which dynamically selects the most reliable pixels as teachers by evaluating the unperturbed predictions from both networks. These selected pixels are then used to supervise perturbed predictions, thereby enhancing the model's learning capability in unreliable region. Experimental results show that our method outperforms state-of-the-art approaches on three public datasets. Code is available at https://github.com/ziziyao/ALHVR.", + "bbox": [ + 86, + 376, + 483, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical images contain various types of tissues with complex anatomical structures, leading to ambiguous bound", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e75b38b68f71efdd50fad78b6529c4231dd794f41870f3c05bf3803bf93c3b36.jpg", + "image_caption": [ + "Figure 1. Training on the Brats dataset using the CPS framework, where the predictions of a given sample are classified based on a confidence threshold. Region 1 represents the reliable stable region, Region 2 represents the reliable unstable region, and Region 3 represents the unreliable stable region." + ], + "image_footnote": [], + "bbox": [ + 544, + 352, + 877, + 436 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ddf6369b6081855551deb4b8fb328102d1a4efe38f924a9d537656a351ff9a2e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetRegion 1Region 2Region 3Dice
ACDC✓✓✓85.34
✓87.13
✓88.46
✓88.20
✓✓87.41
✓✓82.51
✓✓88.58
", + "bbox": [ + 596, + 515, + 826, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. Experimental results of training on the ACDC dataset with $10\\%$ labeled data using the CPS framework under different masked region settings.", + "bbox": [ + 511, + 609, + 906, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "aries between target regions and surrounding tissues. This issue increases the segmentation difficulty for the model at boundary regions. Some semi-supervised methods [23, 34] estimate confidence based on predictions from a single network to distinguish between high-confidence and low-confidence pixels. These methods guide the model to focus on high-confidence predictions by discarding low-confidence pixels. However, predictions from a single network are biased, leading to inaccurate pixel classification. Therefore, some approaches evaluate pixel reliability based on dual-network predictions. For instance, MCF [27] treats inconsistent predictions between networks as errors and attempts to correct them, but it fails to handle cases where both predictions are incorrect. MLRP [24] employs a pixelwise comparison strategy to select relatively reliable predic", + "bbox": [ + 511, + 672, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21450", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions from two networks as pseudo-labels. However, mutual learning between low-confidence predictions may introduce biases, leading to suboptimal learning. These methods fail to identify or utilize high-value regions, particularly those with high segmentation difficulty. Furthermore, they lack adaptive learning strategies tailored to distinct regions, limiting the model's ability to learn in crucial segmentation regions, such as tissue boundaries.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to resolve the aforementioned issues, this paper combines the diversity of dual-network predictions with confidence thresholds to achieve more precise pixel classification. Taking CPS-based methods [6] as an example, we analyze this problem. When two networks predict the same image, the confidence of each pixel relative to a predefined threshold can be categorized into three types: (1) both networks have confidence scores above the threshold, (2) one network has a confidence score above the threshold while the other is below it, and (3) both networks have confidence scores below the threshold. Based on confidence levels and prediction consistency, we define these three regions as the reliable stable region, the reliable unstable region, and the unreliable stable region, respectively. As shown in Fig. 1, different regions are represented by different colors. To further explore the most valuable learning regions, we conducted experiments on the ACDC dataset using the CPS framework. Specifically, we applied masks to occlude different regions, training the model only on unmasked areas. As shown in Tab. 1, the highest Dice score is achieved when pixels in the reliable stable region are masked out during training process. This indicates that the reliable unstable region and the unreliable stable region are high-value regions.", + "bbox": [ + 91, + 214, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the lack of adaptive learning strategies for regions of varying value in existing methods, this paper proposes a semi-supervised medical image segmentation method based on adaptive learning for high-value regions (ALHVR). By analyzing the confidence distributions of model predictions, high-value regions are identified, and region-specific training strategies are designed to improve segmentation performance. Specifically, we propose the confidence-guided cross-prototype consistency learning (CG-CPCL) module and the dynamic teacher competition teaching (DTCT) module, which are designed to train the reliable unstable region and the unreliable stable region, respectively. The main contributions of this paper are as follows:", + "bbox": [ + 89, + 563, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) We conduct an investigation into the high-value regions in semi-supervised medical image segmentation, categorizing them into reliable unstable region and unreliable stable region. Furthermore, we design adaptive learning strategies to enhance segmentation performance in these areas.", + "(2) We propose CG-CPCL to address the challenge where unstable predictions within a region pre" + ], + "bbox": [ + 89, + 777, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vent the model from accurately recognizing key region features. CG-CPCL enhances cross-consistency between features and prototypes by constructing class prototypes while incorporating confidence constraints to ensure that low-confidence predictions learn exclusively from high-confidence predictions. This approach effectively guides the model to learn more reliable feature representations in reliable unstable region, thereby improving its ability to identify key regions.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(3) We design DTCT to tackle the issue where unreliable predictions within a region lead to the accumulation of erroneous signals, causing prediction bias. DTCT employs pixel-level reliability assessment to dynamically and competitively select more reliable pixels from unperturbed predictions as teacher predictions to supervise perturbed predictions. This method provides the model with more valuable and reliable learning signals, effectively reducing prediction bias.", + "(4) The proposed framework integrates both learning modules and is evaluated on three public datasets: ACDC, AbdomenCT-1K, and Brats. Experimental results demonstrate that our framework outperforms current state-of-the-art semi-supervised learning methods." + ], + "bbox": [ + 511, + 228, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 457, + 653, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semi-Supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 511, + 483, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenge in semi-supervised image segmentation is to effectively exploit unlabeled data to improve model performance. To address this, researchers have proposed various approaches, among which consistency regularization and pseudo-labeling methods have become mainstream directions.", + "bbox": [ + 511, + 506, + 903, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, consistency regularization methods are based on the smoothness assumption, which states that a model's predictions should remain similar under different levels of perturbations, including data-level, model-level, and feature-level perturbations. At the data level, common perturbations involve transformations such as rotation and color jittering. To generate more diverse samples, copy-paste techniques are popular, such as CutMix [35] and ClassMix [20]. However, existing copy-paste methods overlook the distribution mismatch between labeled and unlabeled data. To tackle this issue, BCP [2] introduces bidirectional copy-paste between labeled and unlabeled data to generate new samples. At the model level, perturbations are typically introduced by modifying model parameters or structures. For example, MT [25] employs an exponential moving average (EMA) to update the teacher model's parameters with those of the student model, CPS [6] initializes two networks with different parameters and trains them independently. Additionally, MCF [27] and CC-Net [10] introduce variations in encoder or decoder structures, ensuring that different net", + "bbox": [ + 511, + 599, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21451", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work architectures produce consistent predictions. At the feature level, CCT [21] injects noise or drops activations to enforce consistency between decoders under perturbations. PS-MT [14] uses T-VAT from the teacher model for more accurate student perturbations. Since these consistency regularization methods rely on individual perturbations, ABD [8] dynamically adjusts image displacement, improving the model's adaptability to complex data variations and robustness.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, the pseudo-labeling methods are based on the clustering assumption. It expands the dataset by generating pseudo-labels for unlabeled data using a pre-trained model. To improve the quality of pseudo-labels, $\\mathrm{ST} + +$ [32] prioritizes reliable unlabeled images based on prediction stability and performs selective retraining. DMT [9] overcomes the limitation of self-training by identifying incorrect pseudo-labels through prediction differences between multiple models. Additionally, some methods combine pseudolabeling with consistency regularization. For example, FixMatch [23] and RCPS [37] encourage the model to produce consistent predictions between strongly augmented data and pseudo-labels generated from weakly augmented data. While these methods enhance model performance through data diversification or improved pseudo-labels' quality, they overlook the varying complexity of predictions across different image regions, and fail to learn more discriminative features from high-value regions.", + "bbox": [ + 91, + 228, + 486, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Uncertainty-Guided Semi-Supervised Medical Image Segmentation", + "text_level": 1, + "bbox": [ + 89, + 513, + 483, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In semi-supervised medical image segmentation, uncertainty analysis is a key strategy for effectively exploiting pseudo-labels. Existing methods typically estimate uncertainty using Monte Carlo Dropout [11]. Additionally, some approaches employ confidence scores or entropy measures for uncertainty guidance. Specifically, to enable the model to learn more reliable targets, UA-MT [34] utilizes Monte Carlo Dropout for uncertainty estimation and trains the model only with low-uncertainty pixels. URPC [16] refines uncertainty by computing the KL divergence of predictions, allowing the model to progressively learn from multi-scale reliable predictions. To further improve pseudo-label quality, UniMatch [33] evaluates the confidence of model predictions using the maximum class probability and retains only high-confidence pixels from weakly augmented data as pseudo-labels. MLRP [24] assesses the reliability of different subnet predictions via entropy and uses more reliable predictions as pseudo-labels to supervise the other subnetwork's training. To fully exploit pixel information with varying confidence levels, UFC [13] distinguishes high-performing and under-performing pixels based on confidence differences between the student and teacher models. Both DC-Net [5] and DD-Net [26] decouple predictions us", + "bbox": [ + 91, + 553, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing confidence thresholds and optimize different pixels with distinct loss functions, effectively improving model performance. However, although these methods assess and train pixel reliability using various strategies, they rely solely on simple consistency losses (e.g. cross-entropy loss and mean squared error loss) and fail to design adaptive training strategies based on the characteristics of the data.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 209, + 604, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of the Framework", + "text_level": 1, + "bbox": [ + 511, + 234, + 767, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In semi-supervised image segmentation tasks, the training dataset typically consists of a small amount of labeled data and a large amount of unlabeled data. Mathematically, the training dataset is defined as $D = \\{D_L, D_U\\}$ , where $D_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N$ represents the labeled dataset containing $N$ samples, and $D_U = \\{x_i^U\\}_{i=1}^M$ represents the unlabeled dataset containing $M$ samples, with $N \\ll M$ . We define the input image as $x_i \\in \\mathbb{R}^{H \\times W \\times Z}$ , where $H$ , $W$ , and $Z$ denote the height, width, and depth of the image, respectively. The ground truth corresponding to the labeled data is defined as $y_i \\in \\{0, 1, \\dots, C-1\\}^{H \\times W \\times Z}$ , where $C$ represents the number of classes.", + "bbox": [ + 511, + 256, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The training process of ALHVR is shown in Fig. 2, which consists of two parallel networks $f_{A}(\\cdot)$ and $f_{B}(\\cdot)$ with the same structure but independently updating parameters. Each network's decoder takes the encoded feature $F_{1}$ and its perturbed version $F_{2}$ as inputs to generate predictions from different perspectives. It is worth noting that feature perturbations are generated by randomly adding noise or dropout.", + "bbox": [ + 511, + 436, + 905, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The original and perturbed features from both networks are fed into their corresponding decoders, $f_{A}^{d}(\\cdot)$ and $f_{B}^{d}(\\cdot)$ , to generate the final prediction outputs:", + "bbox": [ + 511, + 558, + 905, + 603 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {a 1} = f _ {A} ^ {d} \\left(F _ {a 1}\\right), \\hat {y} _ {a 2} = f _ {A} ^ {d} \\left(F _ {a 2}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 606, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {b 1} = f _ {B} ^ {d} \\left(F _ {b 1}\\right), \\hat {y} _ {b 2} = f _ {B} ^ {d} \\left(F _ {b 2}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 630, + 905, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the decoder outputs include the prediction results for both labeled and unlabeled data, $\\hat{y} = \\hat{y}^L\\cup \\hat{y}^U$", + "bbox": [ + 511, + 654, + 905, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For labeled data, a combination of cross-entropy loss and dice loss is used as the supervised learning loss function:", + "bbox": [ + 511, + 684, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} ^ {a} = L _ {c e} \\left(\\hat {y} _ {a 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {a 1}, y\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 717, + 906, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {s}} ^ {b} = L _ {c e} \\left(\\hat {y} _ {b 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {b 1}, y\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 742, + 903, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For unlabeled data, an adaptive training strategy is proposed based on high-value regions. Specifically, for reliable unstable region, the CG-CPCL module is designed to introduce a cross-prototype consistency loss. Additionally, to further enhance prototype consistency and reduce intraclass variance, the cosine similarity loss between prototypes and features is incorporated as a soft loss term to optimize intra-class feature clustering. The specific definition is as follows:", + "bbox": [ + 511, + 763, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21452", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/584537d36a8a929a644cff7bd016a2712a23a439ec0ac716968a93f621f67b7f.jpg", + "image_caption": [], + "image_footnote": [ + "random perturbation", + "C cosine similarity" + ], + "bbox": [ + 186, + 97, + 808, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58f9d35eee1e1d2f88f4e3a52c74fbc666c823b964642a832cecee5bcef1a2b6.jpg", + "image_caption": [ + "Figure 2. The process of our proposed ALHVR framework. This method first classifies the perturbation predictions of the network output into value regions, then applies an adaptive training strategy to high-value regions, where CG-CPCL is used to train Region 2 and DTCT is used to train Region 3. The supervised learning process and the pseudo-label conversion process are omitted in the figure." + ], + "image_footnote": [ + "$\\text{M}$ multiply pixel by pixel protetype generation", + "1 region1: reliable stable region", + "2 region2: reliable unstable region", + "3 region3:unreliable stable region" + ], + "bbox": [ + 442, + 306, + 803, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p c} ^ {a} = L _ {c e} ^ {a} \\left(\\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\hat {y} _ {b 1} ^ {p}\\right) + L _ {\\cos} ^ {a} \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 503, + 482, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p c} ^ {b} = L _ {c e} ^ {b} \\left(\\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\hat {y} _ {a 1} ^ {p}\\right) + L _ {\\cos} ^ {b} \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 527, + 482, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{a2}^{d}$ and $F_{b2}^{d}$ are perturbed features from the penultimate layer of the decoder, $p_{a1}^{c}$ and $p_{b1}^{c}$ denotes the c-th class prototype computed from the original features of the decoder, $\\hat{y}_{a1}^{p}$ and $\\hat{y}_{b1}^{p}$ denote the pseudo-labels generated from $\\hat{y}_{a1}$ and $\\hat{y}_{b1}$ , respectively. $L_{ce}(\\cdot)$ represents the cross-entropy loss, and $L_{\\cos}(\\cdot)$ denotes the cosine similarity loss.", + "bbox": [ + 89, + 551, + 482, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For unreliable stable region, the mean squared error loss is introduced through DTCT. To further constrain the model's probability distribution and reduce uncertainty in these unreliable regions, we incorporate the entropy of the model's predictions as a regularization term, thereby enhancing the model's robustness in such regions. This is formally defined as:", + "bbox": [ + 89, + 643, + 482, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m h} ^ {a} = L _ {m s e} ^ {a} \\left(\\hat {y} _ {a 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {a 1} \\left(\\hat {y} _ {a 1}\\right) + H _ {a 2} \\left(\\hat {y} _ {a 2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 750, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m h} ^ {b} = L _ {m s e} ^ {b} \\left(\\hat {y} _ {b 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {b 1} \\left(\\hat {y} _ {b 1}\\right) + H _ {b 2} \\left(\\hat {y} _ {b 2}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 770, + 482, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{mse}(\\cdot)$ represents the mean squared error loss, and $H(\\cdot)$ denotes the entropy.", + "bbox": [ + 91, + 792, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unsupervised loss is defined as:", + "bbox": [ + 109, + 823, + 349, + 837 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {u} ^ {a} = L _ {p c} ^ {a} + L _ {m h} ^ {a}, L _ {u} ^ {b} = L _ {p c} ^ {b} + L _ {m h} ^ {b}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 837, + 480, + 856 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss function for each network is:", + "bbox": [ + 109, + 862, + 395, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a} = L _ {s} ^ {a} + \\lambda L _ {u} ^ {a}, L _ {b} = L _ {s} ^ {b} + \\lambda L _ {u} ^ {b}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a time-dependent Gaussian warming up function [12], used to balance the supervised and unsupervised losses.", + "bbox": [ + 511, + 505, + 906, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Valuable Region Classification Method", + "text_level": 1, + "bbox": [ + 511, + 556, + 846, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing semi-supervised learning methods typically use fixed thresholds to separate high-confidence and low-confidence pixels. However, this binary approach provides a coarse distinction of pixel prediction difficulty, neglecting the potential information in low-confidence pixels. To address these limitations, an adaptive threshold is employed to refine pixel classification. Specifically, we directly derive the adaptive threshold from the confidence maps produced by two networks, ensuring better adaptation to current data variations. The confidence maps are generated using perturbed predictions: $\\text{con}_{a2} = \\max(\\hat{y}_{a2}), \\text{con}_{b2} = \\max(\\hat{y}_{b2})$ . By computing the pixel-wise average of the two confidence maps and sorting the values in ascending order, we obtain the sorted average confidence vector $\\text{con}$ .", + "bbox": [ + 511, + 579, + 906, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the lowest $K$ of the sorted confidence values are selected, and their mean is computed as the adaptive threshold:", + "bbox": [ + 511, + 791, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\operatorname {c o n} (i), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 825, + 903, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K = \\beta \\cdot (H\\times W\\times Z)$ , and $\\beta$ is a hyperparameter within the range $(0,1)$ , controlling the proportion of", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21453", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7e8b48bd560b512da09e8e191307169ae2ac82ae377f4c2fcdadab9e450e7cee.jpg", + "image_caption": [ + "Figure 3. The structure of the DTCT module. By comparing the confidence of pixels in the Region 3 from the unperturbed predictions, the more reliable prediction is selected as the teacher. A probability sharpening function is then applied to adjust the probability distribution, generating more reliable pseudo-labels to supervise the perturbed predictions." + ], + "image_footnote": [], + "bbox": [ + 171, + 103, + 823, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "selected pixels.", + "bbox": [ + 89, + 330, + 194, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the confidence threshold, we classify predicted pixels from the two networks interactively. By comparing the confidence values of corresponding pixels with the threshold, the pixels are categorized into three regions including reliable stable region $\\Omega_{1}$ , reliable unstable region $\\Omega_{2}$ and the unreliable stable region $\\Omega_{3}$ (as described in Chapter 1). These regions can be defined as:", + "bbox": [ + 89, + 345, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {1} = \\left\\{c o n _ {a 2} \\geq \\gamma \\& c o n _ {b 2} \\geq \\gamma \\right\\}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 462, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {2} = \\left\\{\\left(c o n _ {a 2} > \\gamma \\& c o n _ {b 2} < \\gamma\\right) \\mid \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 488, + 482, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left(c o n _ {a 2} < \\gamma \\& c o n _ {b 2} > \\gamma\\right) \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 510, + 410, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {3} = \\left\\{c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma \\right\\}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 530, + 482, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The reliable stable region is easily learnable by both networks, while the reliable unstable region reflects areas with unclear predictions, and the unreliable stable region represents difficult-to-learn areas. Thus, focusing on the reliable unstable region and unreliable stable region, rather than the reliable stable region, can better enhance the model's learning ability and provide higher learning value, as confirmed by the experimental results in Tab. 1.", + "bbox": [ + 89, + 551, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Confidence-Guided Cross-Prototype Consistency Learning", + "text_level": 1, + "bbox": [ + 89, + 681, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to complex semantic information in reliable unstable region and differences in model learning capabilities, different models produce divergent predictions for the same region. To address this, we propose the CG-CPCL module, which introduces prototypes as category center references, providing a global and stable reference point. Similar to UPCoL [15], we approximate the probability distribution of voxels by calculating cosine similarity between features and prototypes, introducing global constraints to prevent overreliance on local features. Additionally, we use confidence-based conditional guidance to ensure the model learns only from reliable targets.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first generate class prototypes for the unlabeled data within the current mini-batch. We upsample the features from the second-to-last layer of the decoder using trilinear interpolation to match the size of the pseudolabels. Then, we use the pseudo-labels to select the corresponding class features, and generate class prototypes by calculating the mean of these features. This process is described as follows:", + "bbox": [ + 511, + 330, + 906, + 449 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {a 1} ^ {c} = \\frac {\\sum F _ {a 1} ^ {d} \\cdot \\mathbb {I} (\\hat {y} _ {a 1} ^ {p} = c)}{\\sum (\\hat {y} _ {a 1} ^ {p} = c)}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 450, + 905, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{a1}^{d}$ represents the original features from the decoder, and $\\mathbb{I}(\\cdot)$ denotes the indicator function. For simplicity, the summation process omits the dimensional details of the features and pseudo-labels. Use the same method to calculate the $p_{b1}^{c}$ . After obtaining the class prototypes, we approximate the model's prediction probability by calculating the cosine similarity between the features and the prototypes:", + "bbox": [ + 511, + 493, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} _ {a} = \\cos \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\operatorname {s i m} _ {b} = \\cos \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 607, + 903, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos(a, b) = \\frac{a \\cdot b}{\\|a\\|_2 \\cdot \\|b\\|_2}$ . To ensure that low-confidence predictions in the $\\Omega_2$ region only learn from reliable targets, we need to perform pixel selection on the confidence maps predicted by each network. This generates a position mask that contains the interactive information from both networks, which will be used as a guiding signal for the model's learning:", + "bbox": [ + 511, + 633, + 905, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM a s k _ {a} = \\operatorname {c o n} _ {a 2} < \\gamma \\& \\operatorname {c o n} _ {b 2} > \\gamma , \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 744, + 903, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM a s k _ {b} = \\operatorname {c o n} _ {a 2} > \\gamma \\& \\operatorname {c o n} _ {b 2} < \\gamma . \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 768, + 903, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the position mask can be used to calculate the cross-prototype consistency loss, and the formula is as follows:", + "bbox": [ + 511, + 789, + 905, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {c e} ^ {a} = L _ {c e} \\left(\\operatorname {s i m} _ {a}, \\hat {y} _ {b 1} ^ {p}\\right) \\odot \\operatorname {M a s k} _ {a}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 834, + 903, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ represents the pixel-wise multiplication operation. Calculate $L_{ce}^{b}$ in the same way. Moreover, to further enhance the model's ability to learn class features, the cosine", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21454", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "similarity loss between the features and prototypes is also calculated:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\cos} ^ {a} = (1 - \\operatorname {s i m} _ {a}) \\odot \\operatorname {M a s k} _ {a}. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 126, + 482, + 142 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We calculate $L_{\\mathrm{cos}}^b$ similarly. Finally, the total loss for each model training the reliable unstable region is a linear combination of $L_{ce}$ and $L_{\\mathrm{cos}}$ , as shown in Eqs. (5) and (6).", + "bbox": [ + 89, + 152, + 483, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Dynamic Teacher Competition Teaching", + "text_level": 1, + "bbox": [ + 89, + 207, + 436, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The unreliable stable region is an area where both models struggle to make accurate predictions and typically exhibits high uncertainty. The module improves the adaptability of the model to complex regions by reducing the influence of noise and uncertainty in the training process. The structure of the DTCT module is shown in Fig. 3.", + "bbox": [ + 89, + 229, + 483, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To generate the new teacher prediction, we perform pixel-wise confidence comparison based on the confidence maps of predictions without feature perturbation. We retain the pixels with higher confidence as the new prediction while discarding the pixels with lower confidence. Specifically, we first generate the confidence maps of the two predictions without feature perturbation: $\\text{con}_{a1} = \\max(\\hat{y}_{a1}), \\text{con}_{b1} = \\max(\\hat{y}_{b1})$ . Then, we compare the generated confidence maps and select the prediction probabilities corresponding to the pixels with higher confidence as the new teacher prediction. Next, we apply a probability sharpening function [29] to convert this into a soft pseudolabel. This process can be defined as:", + "bbox": [ + 89, + 320, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} = \\left\\{ \\begin{array}{l} \\hat {y} _ {a 1}, c o n _ {a 1} \\geq c o n _ {b 1} \\\\ \\hat {y} _ {b 1}, c o n _ {a 1} < c o n _ {b 1} \\end{array} , \\right. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 520, + 482, + 554 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t} ^ {\\mathrm {s p}} = \\frac {\\hat {y} _ {t} ^ {1 / T}}{\\hat {y} _ {t} ^ {1 / T} + (1 - \\hat {y} _ {t}) ^ {1 / T}}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 559, + 482, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $T$ is a hyperparameter that controls the output distribution of the model. Since the DTCT module is only used to train the pixels in the unreliable stable region, a location mask needs to be generated first to determine the region location of $\\Omega_3$ :", + "bbox": [ + 89, + 603, + 483, + 679 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM a s k = c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma . \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 685, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Then, the newly generated reliable pseudo-labels are used to supervise the perturbed version of the predictions, as shown in the following formula:", + "bbox": [ + 89, + 710, + 483, + 756 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {m s e} ^ {a} = L _ {m s e} \\left(\\hat {y} _ {a 2}, \\hat {y} _ {t} ^ {s p}\\right) \\odot M a s k. \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 761, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We can calculate $L_{mse}^{b}$ in the same way. To enhance the model's ability to perceive uncertainty in the unreliable stable region, guide the model to focus on regions with higher uncertainty, and effectively reduce the uncertainty in these regions, we quantify uncertainty using entropy and add it as a regularization term to the loss function:", + "bbox": [ + 89, + 787, + 483, + 878 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH _ {a 1} = H \\left(\\hat {y} _ {a 1}\\right) \\odot \\text {M a s k}, H _ {a 2} = H \\left(\\hat {y} _ {a 2}\\right) \\odot \\text {M a s k}, \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 883, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a9afa876456c526020e31c2787d5e73156cc94f71161b219f06e76ed7491cf20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net7(10%)081.5970.748.072.35
U-Net14(20%)084.8375.337.612.09
U-Net70(All)091.1084.044.901.14
MT[25][NIPS'2017]83.6573.1513.453.61
CPS[6][CVPR'2021]85.3475.508.782.37
MC-Net[28][MICCAI'2021]86.0776.5811.483.37
MCF[27][CVPR'2023]85.1875.2210.732.78
AC-MT[31][MIA'2023]86.3676.869.642.59
UG-MCL[36][AIIM'2023]7(10%)63(90%)86.8477.607.771.94
EVIL[7][CBM'2024]87.2578.185.311.35
AAU[1][MIA'2024]86.4676.998.002.06
MLRP[24][MIA'2024]87.1078.164.911.30
ours90.56[3,31]83.212.570.78
MT[25][NIPS'2017]86.5577.496.862.15
CPS[6][CVPR'2021]87.0378.136.662.12
MC-Net[28][MICCAI'2021]86.5877.6815.994.93
MCF[27][CVPR'2023]87.3278.436.532.00
AC-MT[31][MIA'2023]87.9979.347.852.14
UG-MCL[36][AIIM'2023]14(20%)56(80%)88.4079.869.122.44
EVIL[7][CBM'2024]88.3479.906.461.68
AAU[1][MIA'2024]87.7178.928.512.14
MLRP[24][MIA'2024]88.0079.545.411.58
ours91.09[2,69]84.052.710.73
", + "bbox": [ + 519, + 88, + 901, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparisons with state-of-the-art semi-supervised segmentation methods on the ACDC dataset.", + "bbox": [ + 511, + 316, + 903, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $H(\\hat{y}) = -\\hat{y}\\log \\hat{y}$ . We can use the same method to calculate the $H_{b1}$ and $H_{b2}$ . Finally, the total loss for training the unreliable stable region in each model is a linear combination of $L_{mse}$ and $H$ , as shown in Eqs. (7) and (8).", + "bbox": [ + 511, + 366, + 905, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 441, + 648, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 467, + 800, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The ACDC dataset [4] consists of cardiac dynamic MRI scans and corresponding ground truth from 100 patients, categorized into three classes: right ventricle, left ventricle, and myocardium. Following the experimental setup in CTCT [17], we used scans from 70 patients for training, 10 patients for validation, and 20 patients for testing.", + "bbox": [ + 511, + 489, + 905, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The AbdomenCT-1K dataset [18] contains over 1,000 (1K) CT scans from 12 medical centers, categorized into four classes: liver, kidneys, spleen, and pancreas. Following the experimental setup in BFFC [38], we selected 30 scans for training and 11 scans for testing.", + "bbox": [ + 511, + 580, + 905, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Brats dataset [3] consists of preoperative MRI scans from 335 glioma patients. Following the experimental setup in CPCL [30], we used 250 samples for training, 25 samples for validation, and 60 for testing.", + "bbox": [ + 511, + 656, + 905, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We evaluate the method using four metrics categorized into region-sensitive and edge-sensitive types. Region-sensitive metrics include Dice and Jaccard, while edge-sensitive metrics include 95HD and ASD.", + "bbox": [ + 511, + 715, + 905, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 787, + 730, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments were implemented using the PyTorch 1.7 on NVIDIA GeForce RTX 3090 GPU. The model is trained using the SGD optimizer with a weight decay factor of 0.0001, a momentum of 0.9 and an initial learning rate of 0.01. For segmentation on the ACDC and AbdomenCT-1K datasets, we use U-Net [22] as the backbone, with a", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21455", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/57cbfd2b8d4f591c120d9119272c7e84e40bb9b6b87e916b58c911e66fa3dd6b.jpg", + "image_caption": [ + "Figure 4. Visualization of some semi-supervised methods. The first two rows show the segmentation results on the ACDC dataset trained with $10\\%$ labeled data. The last two rows show the segmentation results on the AbdomenCT-1K dataset trained with $5\\%$ labeled data." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 475, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "batch size of 16 and $30\\mathrm{k}$ training iterations. For Brats, we use V-Net [19] as the backbone, with a batch size of 4 and $15\\mathrm{k}$ training iterations. In the total loss function of the network, the Gaussian warming up function is defined as $\\lambda (t) = 0.1\\times e^{-5(1 - t / t_{\\mathrm{max}})^2}$ , where $\\mathfrak{t}$ denotes the current number of iterations and $t_\\mathrm{max}$ represents the maximum number of iterations. In the probability sharpening function, the hyperparameter $T$ is usually empirically set to 0.1. For calculating the confidence threshold, the hyperparameter $\\beta$ values were configured as 0.8, 0.7, and 0.7 in the ACDC, AbdomenCT-1K, and Brats datasets, respectively (set by Ablation Studies).", + "bbox": [ + 88, + 422, + 482, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 618, + 459, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Firstly, we evaluate the proposed method on the ACDC dataset, comparing it with MT [25], CPS [6], MC-Net [28], MCF [27], AC-MT [31], UG-MCL [36], EVIL [7], AAU [1], and MLRP [24]. Additionally, we use U-Net trained with $10\\%$ , $20\\%$ , and $100\\%$ of labeled data as the baseline and upper bound for reference. As shown in Tab. 2, our method achieves significant improvements across all four evaluation metrics compared to the baseline. Specifically, with $10\\%$ labeled data, the Dice score increases from $81.59\\%$ to $90.56\\%$ . With $20\\%$ labeled data, it improves from $84.83\\%$ to $91.09\\%$ , approaching the upper bound of $91.10\\%$ achieved with full supervision. Compared to state-of-the-art methods, ALHVR consistently outperforms others. With $10\\%$ labeled data, it surpasses the second-best method (EVIL) by $3.31\\%$ in Dice. When using $20\\%$ labeled data, it outperforms UG-MCL by $2.69\\%$ in Dice.", + "bbox": [ + 91, + 642, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the generalization capability of the", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac74a443a1e566f936981a034bd384b47b26201fa77df67d330cb867ac751d25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net1(5%)067.3957.1446.8711.37
U-Net3(10%)075.6666.3322.446.89
U-Net30(All)086.9379.2711.552.45
MT[25][NIPS'2017]71.8261.1942.039.23
CPS[6][CVPR'2021]78.1468.3226.256.68
MC-Net[28][MICCAI'2021]80.9771.4332.987.48
MCF[27][CVPR'2023]80.7971.3026.137.81
AC-MT[31][MIA'2023]76.1865.9936.128.89
UG-MCL[36][AIIM'2023]1(5%)29(95%)74.0663.5826.937.06
EVIL[7][CBM'2024]80.6871.4723.726.33
AAU[1][MIA'2024]80.6571.0332.689.04
MLRP[24][MIA'2024]80.3070.6720.945.43
ours84.7576.3521.155.59
MT[25][NIPS'2017]78.7569.1025.356.88
CPS[6][CVPR'2021]82.7174.1117.304.90
MC-Net[28][MICCAI'2021]82.9474.1319.354.59
MCF[27][CVPR'2023]83.3474.6316.974.72
AC-MT[31][MIA'2023]82.3272.9726.527.54
UG-MCL[36][AIIM'2023]3(10%)27(90%)81.4372.5620.295.67
EVIL[7][CBM'2024]84.2475.7518.145.25
AAU[1][MIA'2024]81.4271.8229.698.25
MLRP[24][MIA'2024]84.5575.9913.824.05
ours85.6877.3521.966.22
", + "bbox": [ + 519, + 88, + 901, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparisons with state-of-the-art semi-supervised segmentation methods on the AbdomenCT-1K dataset.", + "bbox": [ + 511, + 316, + 903, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proposed method across different datasets, we evaluate its performance on the five-class AbdomenCT-1K dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 3, with $5\\%$ labeled data, ALHVR improves the Dice score by $17.36\\%$ compared to the baseline. When increasing the labeled data to $20\\%$ , the Dice score improves by $10.02\\%$ . Compared to state-of-the-art methods, ALHVR consistently achieves the best performance in segmenting four organs. Specifically, with $5\\%$ and $10\\%$ labeled data, it surpasses the second-best methods (MC-Net and MLRP) by $3.78\\%$ and $1.13\\%$ in Dice, respectively. Furthermore, as illustrated in Fig. 4, our method provides more precise segmentation of fine structures and boundary regions across different organs.", + "bbox": [ + 511, + 369, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to organ segmentation tasks, we evaluated our method on the Brats dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 4, with only $10\\%$ labeled data, ALHVR improves the Dice score by $5.91\\%$ compared to the baseline. Furthermore, it outperforms state-of-the-art methods, achieving a $1.18\\%$ higher Dice score than the second-best method (MLRP).", + "bbox": [ + 511, + 580, + 905, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 698, + 679, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies to validate the effectiveness of each part and analyze the hyperparameters of the proposed method. The experiments were conducted on the ACDC (labeled ratio: $10\\%$ ), AbdomenCT-1K (labeled ratio: $5\\%$ ), and Brats (labeled ratio: $10\\%$ ) datasets.", + "bbox": [ + 511, + 720, + 905, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Impact of Different Modules", + "text_level": 1, + "bbox": [ + 511, + 805, + 758, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, we examine the effectiveness of the major components of ALHVR, CG-CPCL, and DTCT, on different datasets. When CG-CPCL training (learning only reliable unstable region) was used, the performance of all datasets was significantly improved compared with the baseline", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21456", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/51923632ee0abd4b0cf8932699df08d2b6f3efe5d86681473099d88904ffade4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
V-Net25(10%)079.2368.4015.834.34
V-Net250(All)086.4977.238.681.98
MT[25][NIPS'2017]80.4969.9911.713.46
CPS[6][CVPR'2021]82.7072.3012.993.46
MC-Net[28][MICCAI'2021]82.5172.9814.573.78
MCF[27][CVPR'2023]81.9671.7613.963.99
AC-MT[31][MIA'2023]25(10%)225(90%)82.0971.6515.914.81
UG-MCL[36][AHIM'2023]82.8572.6911.122.28
EVIL[7][CBM'2024]82.6072.4611.442.76
AAU[1][MIA'2024]82.4372.3313.834.18
MLRP[24][MIA'2024]83.9674.0015.164.21
ours85.14↑1.1875.757.792.00
", + "bbox": [ + 98, + 88, + 477, + 213 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/202ff8e39da6ba3303e7c4b165c1afef9122e4e0a45554914dc3726b418ae08e.jpg", + "table_caption": [ + "Table 4. Comparisons with state-of-the-art semi-supervised segmentation methods on the Brats dataset." + ], + "table_footnote": [], + "table_body": "
DatasetsCG-CPCLDTCTDice↑Jaccard↑95HD↓ASD↓
ACDC✓86.6177.147.302.15
89.3281.313.451.02
✓89.1980.973.260.94
✓✓90.5683.212.570.78
AbdomenCT-1K✓79.1969.2129.777.56
83.9875.919.592.81
✓83.7975.2521.395.28
✓✓84.7576.3521.155.59
Brats✓83.2073.7311.913.01
84.6274.818.922.36
✓84.8575.229.962.61
✓✓85.1475.757.792.00
", + "bbox": [ + 122, + 251, + 450, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method, especially on the AbdomenCT-1K dataset, where the Dice index improved by $4.79\\%$ . In the ACDC dataset, when cross pseudo supervision similar to CPS was applied (see Tab. 1 for experimental results), CG-CPCL increased by $0.86\\%$ . The reliable unstable region usually has feature differences within the class, which is prone to misjudgment. Therefore, the class prototype is constructed, and cross-prototype consistency is used to enhance the model's discrimination. In addition, the confidence-based conditional guidance mechanism effectively reduces the performance difference between models.", + "bbox": [ + 91, + 462, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similarly, with DTCT training (learning only the unreliable stable region), the performance of all datasets improved significantly, with the Brats dataset improving by $1.65\\%$ . On the ACDC dataset, CG-CPCL improved by $0.99\\%$ when cross pseudo supervision similar to CPS was used. The unreliable stable region has high uncertainty, and a single model is prone to bias. To this end, we improved the accuracy of the model in this area by integrating predictions from multiple models, competitively generating teacher predictions, and selecting the most credible predictions as pseudo-labels.", + "bbox": [ + 91, + 628, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When CG-CPCL and DTCT are combined, the model can achieve optimal segmentation performance by using differentiated learning strategies to optimize both reliable unstable and unreliable stable regions. Compared with the baseline model, the Dice indexes of multiple datasets improved by $3.95\\%$ , $5.56\\%$ , and $1.94\\%$ , respectively. On the ACDC dataset, ALHVR increased the Dice indicator from", + "bbox": [ + 91, + 795, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/66057fde1b1c7200840f4b8661b9e5575849ccb6db8651719657ae64a61dfa6f.jpg", + "table_caption": [ + "Table 5. Ablation results of different components of ALHVR. We conducted experiments to test the impact of CG-CPCL and DTCT on model performance." + ], + "table_footnote": [], + "table_body": "
β0.50.60.70.80.9
ACDC89.8889.9490.2690.5689.85
AbdomenCT-1K84.1684.3984.7584.4983.71
Brats84.7185.0285.1484.8484.50
", + "bbox": [ + 578, + 89, + 841, + 140 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies on the hyperparameter $\\beta$ with ${10}\\% ,5\\%$ and ${10}\\%$ labeled data.", + "bbox": [ + 514, + 145, + 903, + 171 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "88.58% to 90.56%. Since these two types of regions have different distributions and prediction characteristics, a unified training strategy is difficult to account for both. Therefore, our method optimizes the two types of regions with targeted strategies, effectively improving the overall performance of the model and its generalization ability under complex scenarios.", + "bbox": [ + 514, + 199, + 903, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Hyperparameter Value Analysis", + "text_level": 1, + "bbox": [ + 516, + 319, + 781, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigate the impact of confidence thresholds calculated based on different pixel proportions on model training. As shown in Tab. 6, the performance of dynamic thresholds on multiple datasets is evaluated using Dice coefficients. The results indicate that the optimal proportions are 0.8 for the ACDC dataset and 0.7 for the AbdomenCT-1K and Brats datasets. When the proportion is set lower, the threshold decreases, causing the model to focus more on high-confidence pixels, which may reduce its ability to handle difficult samples and degrade the quality of pseudolabels. Conversely, when the threshold increases, the model may overlook subtle features, affecting detail prediction. Therefore, it is necessary to balance the pixel proportion to optimize the model's attention on the reliable unstable region and the unreliable stable region, thereby improving performance and robustness.", + "bbox": [ + 514, + 339, + 903, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 601, + 630, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel semi-supervised learning framework, ALHVR, which explores high-value learning regions and designs adaptive training strategies. Specifically, CG-CPCL is used to train the reliable unstable region, while DTCT is employed for the unreliable stable region. To enhance stability in the reliable unstable region, CG-CPCL enables low-confidence predictions to learn from high-confidence predictions through confidence-guided cross-pseudo supervision and prototype consistency constraints, improving intra-class consistency and inter-class discriminability. To enhance reliability in the unreliable stable region, DTCT dynamically selects more reliable pixels as teachers using a dynamic teacher selection mechanism and competitive teaching strategy, reducing noise interference while strengthening the model's learning capability. Experimental results on three public datasets demonstrate that ALHVR achieves better performance than existing state-of-the-art methods.", + "bbox": [ + 514, + 628, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21457", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62271296, 62201334), the Shaanxi Province Innovation Capacity Support Program (Grant No. 2025RS-CXTD-012), and the Scientific Research Program funded by the Shaanxi Provincial Education Commission (Grant Nos. 23JP022, 23JP014).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 237 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sukesh Adiga, Jose Dolz, and Herve Lombaert. Anatomically-aware uncertainty for semi-supervised image segmentation. Medical Image Analysis, 91:103011, 2024. 6, 7, 8", + "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11514-11524, 2023. 2", + "[3] Spyridon (Spyros) Bakas. Brats miccai brain tumor dataset, 2020. https://dx.doi.org/10.21227/hdtd-5j88.6", + "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6", + "[5] Faquan Chen, Jingjing Fei, Yaqi Chen, and Chenxi Huang. Decoupled consistency for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 551-561. Springer, 2023. 3", + "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 2, 6, 7, 8", + "[7] Yingyu Chen, Ziyuan Yang, Chenyu Shen, Zhiwen Wang, Zhongzhou Zhang, Yang Qin, Xin Wei, Jingfeng Lu, Yan Liu, and Yi Zhang. Evidence-based uncertainty-aware semi-supervised medical image segmentation. Computers in Biology and Medicine, 170:108004, 2024. 6, 7, 8", + "[8] Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4070-4080, 2024. 3", + "[9] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, 130:108777, 2022. 3", + "[10] Hejun Huang, Zuguo Chen, Chaoyang Chen, Ming Lu, and Ying Zou. Complementary consistency semi-supervised learning for 3d left atrial image segmentation. Computers in Biology and Medicine, 165:107368, 2023. 2" + ], + "bbox": [ + 93, + 247, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3", + "[12] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 4", + "[13] Tao Lei, Yi Wang, Xingwu Wang, Xuan Wang, Bin Hu, and Asoke K Nandi. Unified feature consistency of underperforming pixels and valid regions for semi-supervised medical image segmentation. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 3", + "[14] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 3", + "[15] Wenjing Lu, Jiahao Lei, Peng Qiu, Rui Sheng, Jinhua Zhou, Xinwu Lu, and Yang Yang. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 662-672. Springer, 2023. 5", + "[16] Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, and Shaoting Zhang. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 318-329. Springer, 2021. 3", + "[17] Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In International conference on medical imaging with deep learning, pages 820-833. PMLR, 2022. 6", + "[18] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695-6714, 2021. 6", + "[19] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 7", + "[20] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369-1378, 2021. 2", + "[21] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12674-12684, 2020. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21458", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6", + "[23] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3", + "[24] Jiawei Su, Zhiming Luo, Sheng Lian, Dazhen Lin, and Shaozi Li. Mutual learning with reliable pseudo label for semi-supervised medical image segmentation. Medical Image Analysis, 94:103111, 2024. 1, 3, 6, 7, 8", + "[25] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 6, 7, 8", + "[26] Bing Wang, Taifeng Huang, Shuo Yang, Ying Yang, Junhai Zhai, and Xin Zhang. Dual-decoder data decoupling training for semi-supervised medical image segmentation. Biomedical Signal Processing and Control, 100:106984, 2025. 3", + "[27] Yongchao Wang, Bin Xiao, Xiuli Bi, Weisheng Li, and Xinbo Gao. Mcf: Mutual correction framework for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15651-15660, 2023. 1, 2, 6, 7, 8", + "[28] Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, and Lei Zhang. Semi-supervised left atrium segmentation with mutual consistency training. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part II 24, pages 297-306. Springer, 2021. 6, 7, 8", + "[29] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 6", + "[30] Zhe Xu, Yixin Wang, Donghuan Lu, Lequan Yu, Jiangpeng Yan, Jie Luo, Kai Ma, Yefeng Zheng, and Raymond Kaiyu Tong. All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 26(7):3174-3184, 2022. 6", + "[31] Zhe Xu, Yixin Wang, Donghuan Lu, Xiangde Luo, Jiangpeng Yan, Yefeng Zheng, and Raymond Kai-yu Tong. Ambiguity-selective consistency regularization for meantacher semi-supervised medical image segmentation. Medical Image Analysis, 88:102880, 2023. 6, 7, 8", + "[32] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4268-4277, 2022. 3", + "[33] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in" + ], + "bbox": [ + 91, + 92, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7236-7246, 2023. 3", + "[34] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical image computing and computer assisted intervention-MICCAI 2019: 22nd international conference, Shenzhen, China, October 13–17, 2019, proceedings, part II 22, pages 605–613. Springer, 2019. 1, 3", + "[35] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2", + "[36] Yichi Zhang, Rushi Jiao, Qingcheng Liao, Dongyang Li, and Jicong Zhang. Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation. Artificial Intelligence in Medicine, 138:102476, 2023. 6, 7, 8", + "[37] Xiangyu Zhao, Zengxin Qi, Sheng Wang, Qian Wang, Xuehai Wu, Ying Mao, and Lichi Zhang. Rcps: Rectified contrastive pseudo supervision for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 28(1):251-261, 2023. 3", + "[38] Zhongda Zhao, Haiyan Wang, Tao Lei, Xuan Wang, Xiaohong Shen, and Haiyang Yao. Balanced feature fusion collaborative training for semi-supervised medical image segmentation. Pattern Recognition, 157:110856, 2025. 6" + ], + "bbox": [ + 516, + 92, + 906, + 486 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21459", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_model.json b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4fc69f108298f837ab81c54c0235f7c1eb92f74e --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.183, + 0.131, + 0.819, + 0.177 + ], + "angle": 0, + "content": "Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.203, + 0.884, + 0.293 + ], + "angle": 0, + "content": "Tao Lei\\(^{1*}\\), Ziyao Yang\\(^{1}\\), Xingwu Wang\\(^{2}\\), Yi Wang\\(^{1}\\), Xuan Wang\\(^{3}\\), Feiman Sun\\(^{4}\\), Asoke K. Nandi\\(^{5}\\) \n\\(^{1}\\)Shaanxi University of Science and Technology, China \n\\(^{2}\\)Beijing University of Posts and Telecommunications, China \n\\(^{3}\\)Northwestern Polytechnical University, China \n\\(^{4}\\)Southwest Jiaotong University, China \\(^{5}\\)Brunel University of London, UK" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.295, + 0.674, + 0.31 + ], + "angle": 0, + "content": "*Corresponding author:leitao@sust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.377, + 0.485, + 0.817 + ], + "angle": 0, + "content": "Existing semi-supervised learning methods typically mitigate the impact of unreliable predictions by suppressing low-confidence regions. However, these methods fail to explore which regions hold higher learning value and how to design adaptive learning strategies for these regions. To address these issues, we propose a novel adaptive learning of high-value regions (ALHVR) framework. By exploiting the diversity of predictions from dual-branch networks, the prediction regions are classified into three groups: reliable stable region, reliable unstable region, and unreliable stable region. For high-value regions (reliable unstable region and unreliable stable region), different training strategies are designed. Specifically, for reliable unstable region, we propose a confidence-guided cross-prototype consistency learning (CG-CPCL) module, which enforces prototype consistency constraints in the feature space. By leveraging confidence information, the high-confidence predictions from one network selectively supervise the low-confidence predictions from the other, thus helping the model learn inter-class discrimination more stably. Additionally, for unreliable stable region, we design a dynamic teacher competition teaching (DTCT) module, which dynamically selects the most reliable pixels as teachers by evaluating the unperturbed predictions from both networks. These selected pixels are then used to supervise perturbed predictions, thereby enhancing the model's learning capability in unreliable region. Experimental results show that our method outperforms state-of-the-art approaches on three public datasets. Code is available at https://github.com/ziziyao/ALHVR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Medical images contain various types of tissues with complex anatomical structures, leading to ambiguous bound" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.353, + 0.878, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.442, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Figure 1. Training on the Brats dataset using the CPS framework, where the predictions of a given sample are classified based on a confidence threshold. Region 1 represents the reliable stable region, Region 2 represents the reliable unstable region, and Region 3 represents the unreliable stable region." + }, + { + "type": "table", + "bbox": [ + 0.597, + 0.516, + 0.828, + 0.606 + ], + "angle": 0, + "content": "
DatasetRegion 1Region 2Region 3Dice
ACDC✓✓✓85.34
✓87.13
✓88.46
✓88.20
✓✓87.41
✓✓82.51
✓✓88.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Table 1. Experimental results of training on the ACDC dataset with \\(10\\%\\) labeled data using the CPS framework under different masked region settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.903 + ], + "angle": 0, + "content": "aries between target regions and surrounding tissues. This issue increases the segmentation difficulty for the model at boundary regions. Some semi-supervised methods [23, 34] estimate confidence based on predictions from a single network to distinguish between high-confidence and low-confidence pixels. These methods guide the model to focus on high-confidence predictions by discarding low-confidence pixels. However, predictions from a single network are biased, leading to inaccurate pixel classification. Therefore, some approaches evaluate pixel reliability based on dual-network predictions. For instance, MCF [27] treats inconsistent predictions between networks as errors and attempts to correct them, but it fails to handle cases where both predictions are incorrect. MLRP [24] employs a pixelwise comparison strategy to select relatively reliable predic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21450" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "tions from two networks as pseudo-labels. However, mutual learning between low-confidence predictions may introduce biases, leading to suboptimal learning. These methods fail to identify or utilize high-value regions, particularly those with high segmentation difficulty. Furthermore, they lack adaptive learning strategies tailored to distinct regions, limiting the model's ability to learn in crucial segmentation regions, such as tissue boundaries." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.215, + 0.485, + 0.562 + ], + "angle": 0, + "content": "In order to resolve the aforementioned issues, this paper combines the diversity of dual-network predictions with confidence thresholds to achieve more precise pixel classification. Taking CPS-based methods [6] as an example, we analyze this problem. When two networks predict the same image, the confidence of each pixel relative to a predefined threshold can be categorized into three types: (1) both networks have confidence scores above the threshold, (2) one network has a confidence score above the threshold while the other is below it, and (3) both networks have confidence scores below the threshold. Based on confidence levels and prediction consistency, we define these three regions as the reliable stable region, the reliable unstable region, and the unreliable stable region, respectively. As shown in Fig. 1, different regions are represented by different colors. To further explore the most valuable learning regions, we conducted experiments on the ACDC dataset using the CPS framework. Specifically, we applied masks to occlude different regions, training the model only on unmasked areas. As shown in Tab. 1, the highest Dice score is achieved when pixels in the reliable stable region are masked out during training process. This indicates that the reliable unstable region and the unreliable stable region are high-value regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.775 + ], + "angle": 0, + "content": "To address the lack of adaptive learning strategies for regions of varying value in existing methods, this paper proposes a semi-supervised medical image segmentation method based on adaptive learning for high-value regions (ALHVR). By analyzing the confidence distributions of model predictions, high-value regions are identified, and region-specific training strategies are designed to improve segmentation performance. Specifically, we propose the confidence-guided cross-prototype consistency learning (CG-CPCL) module and the dynamic teacher competition teaching (DTCT) module, which are designed to train the reliable unstable region and the unreliable stable region, respectively. The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.868 + ], + "angle": 0, + "content": "(1) We conduct an investigation into the high-value regions in semi-supervised medical image segmentation, categorizing them into reliable unstable region and unreliable stable region. Furthermore, we design adaptive learning strategies to enhance segmentation performance in these areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "(2) We propose CG-CPCL to address the challenge where unstable predictions within a region pre" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "vent the model from accurately recognizing key region features. CG-CPCL enhances cross-consistency between features and prototypes by constructing class prototypes while incorporating confidence constraints to ensure that low-confidence predictions learn exclusively from high-confidence predictions. This approach effectively guides the model to learn more reliable feature representations in reliable unstable region, thereby improving its ability to identify key regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.363 + ], + "angle": 0, + "content": "(3) We design DTCT to tackle the issue where unreliable predictions within a region lead to the accumulation of erroneous signals, causing prediction bias. DTCT employs pixel-level reliability assessment to dynamically and competitively select more reliable pixels from unperturbed predictions as teacher predictions to supervise perturbed predictions. This method provides the model with more valuable and reliable learning signals, effectively reducing prediction bias." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.442 + ], + "angle": 0, + "content": "(4) The proposed framework integrates both learning modules and is evaluated on three public datasets: ACDC, AbdomenCT-1K, and Brats. Experimental results demonstrate that our framework outperforms current state-of-the-art semi-supervised learning methods." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.229, + 0.905, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.458, + 0.655, + 0.473 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.905, + 0.501 + ], + "angle": 0, + "content": "2.1. Semi-Supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.597 + ], + "angle": 0, + "content": "The key challenge in semi-supervised image segmentation is to effectively exploit unlabeled data to improve model performance. To address this, researchers have proposed various approaches, among which consistency regularization and pseudo-labeling methods have become mainstream directions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Firstly, consistency regularization methods are based on the smoothness assumption, which states that a model's predictions should remain similar under different levels of perturbations, including data-level, model-level, and feature-level perturbations. At the data level, common perturbations involve transformations such as rotation and color jittering. To generate more diverse samples, copy-paste techniques are popular, such as CutMix [35] and ClassMix [20]. However, existing copy-paste methods overlook the distribution mismatch between labeled and unlabeled data. To tackle this issue, BCP [2] introduces bidirectional copy-paste between labeled and unlabeled data to generate new samples. At the model level, perturbations are typically introduced by modifying model parameters or structures. For example, MT [25] employs an exponential moving average (EMA) to update the teacher model's parameters with those of the student model, CPS [6] initializes two networks with different parameters and trains them independently. Additionally, MCF [27] and CC-Net [10] introduce variations in encoder or decoder structures, ensuring that different net" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21451" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "work architectures produce consistent predictions. At the feature level, CCT [21] injects noise or drops activations to enforce consistency between decoders under perturbations. PS-MT [14] uses T-VAT from the teacher model for more accurate student perturbations. Since these consistency regularization methods rely on individual perturbations, ABD [8] dynamically adjusts image displacement, improving the model's adaptability to complex data variations and robustness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.229, + 0.487, + 0.502 + ], + "angle": 0, + "content": "Secondly, the pseudo-labeling methods are based on the clustering assumption. It expands the dataset by generating pseudo-labels for unlabeled data using a pre-trained model. To improve the quality of pseudo-labels, \\(\\mathrm{ST} + +\\) [32] prioritizes reliable unlabeled images based on prediction stability and performs selective retraining. DMT [9] overcomes the limitation of self-training by identifying incorrect pseudo-labels through prediction differences between multiple models. Additionally, some methods combine pseudolabeling with consistency regularization. For example, FixMatch [23] and RCPS [37] encourage the model to produce consistent predictions between strongly augmented data and pseudo-labels generated from weakly augmented data. While these methods enhance model performance through data diversification or improved pseudo-labels' quality, they overlook the varying complexity of predictions across different image regions, and fail to learn more discriminative features from high-value regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.515, + 0.484, + 0.547 + ], + "angle": 0, + "content": "2.2. Uncertainty-Guided Semi-Supervised Medical Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In semi-supervised medical image segmentation, uncertainty analysis is a key strategy for effectively exploiting pseudo-labels. Existing methods typically estimate uncertainty using Monte Carlo Dropout [11]. Additionally, some approaches employ confidence scores or entropy measures for uncertainty guidance. Specifically, to enable the model to learn more reliable targets, UA-MT [34] utilizes Monte Carlo Dropout for uncertainty estimation and trains the model only with low-uncertainty pixels. URPC [16] refines uncertainty by computing the KL divergence of predictions, allowing the model to progressively learn from multi-scale reliable predictions. To further improve pseudo-label quality, UniMatch [33] evaluates the confidence of model predictions using the maximum class probability and retains only high-confidence pixels from weakly augmented data as pseudo-labels. MLRP [24] assesses the reliability of different subnet predictions via entropy and uses more reliable predictions as pseudo-labels to supervise the other subnetwork's training. To fully exploit pixel information with varying confidence levels, UFC [13] distinguishes high-performing and under-performing pixels based on confidence differences between the student and teacher models. Both DC-Net [5] and DD-Net [26] decouple predictions us" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.198 + ], + "angle": 0, + "content": "ing confidence thresholds and optimize different pixels with distinct loss functions, effectively improving model performance. However, although these methods assess and train pixel reliability using various strategies, they rely solely on simple consistency losses (e.g. cross-entropy loss and mean squared error loss) and fail to design adaptive training strategies based on the characteristics of the data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.21, + 0.605, + 0.226 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.235, + 0.768, + 0.249 + ], + "angle": 0, + "content": "3.1. Overview of the Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.437 + ], + "angle": 0, + "content": "In semi-supervised image segmentation tasks, the training dataset typically consists of a small amount of labeled data and a large amount of unlabeled data. Mathematically, the training dataset is defined as \\( D = \\{D_L, D_U\\} \\), where \\( D_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N \\) represents the labeled dataset containing \\( N \\) samples, and \\( D_U = \\{x_i^U\\}_{i=1}^M \\) represents the unlabeled dataset containing \\( M \\) samples, with \\( N \\ll M \\). We define the input image as \\( x_i \\in \\mathbb{R}^{H \\times W \\times Z} \\), where \\( H \\), \\( W \\), and \\( Z \\) denote the height, width, and depth of the image, respectively. The ground truth corresponding to the labeled data is defined as \\( y_i \\in \\{0, 1, \\dots, C-1\\}^{H \\times W \\times Z} \\), where \\( C \\) represents the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.906, + 0.558 + ], + "angle": 0, + "content": "The training process of ALHVR is shown in Fig. 2, which consists of two parallel networks \\( f_{A}(\\cdot) \\) and \\( f_{B}(\\cdot) \\) with the same structure but independently updating parameters. Each network's decoder takes the encoded feature \\( F_{1} \\) and its perturbed version \\( F_{2} \\) as inputs to generate predictions from different perspectives. It is worth noting that feature perturbations are generated by randomly adding noise or dropout." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.906, + 0.604 + ], + "angle": 0, + "content": "The original and perturbed features from both networks are fed into their corresponding decoders, \\( f_{A}^{d}(\\cdot) \\) and \\( f_{B}^{d}(\\cdot) \\), to generate the final prediction outputs:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.607, + 0.907, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {a 1} = f _ {A} ^ {d} \\left(F _ {a 1}\\right), \\hat {y} _ {a 2} = f _ {A} ^ {d} \\left(F _ {a 2}\\right), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.631, + 0.906, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {b 1} = f _ {B} ^ {d} \\left(F _ {b 1}\\right), \\hat {y} _ {b 2} = f _ {B} ^ {d} \\left(F _ {b 2}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.684 + ], + "angle": 0, + "content": "where the decoder outputs include the prediction results for both labeled and unlabeled data, \\(\\hat{y} = \\hat{y}^L\\cup \\hat{y}^U\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.906, + 0.715 + ], + "angle": 0, + "content": "For labeled data, a combination of cross-entropy loss and dice loss is used as the supervised learning loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.718, + 0.907, + 0.735 + ], + "angle": 0, + "content": "\\[\nL _ {s} ^ {a} = L _ {c e} \\left(\\hat {y} _ {a 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {a 1}, y\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.743, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {s}} ^ {b} = L _ {c e} \\left(\\hat {y} _ {b 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {b 1}, y\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.9 + ], + "angle": 0, + "content": "For unlabeled data, an adaptive training strategy is proposed based on high-value regions. Specifically, for reliable unstable region, the CG-CPCL module is designed to introduce a cross-prototype consistency loss. Additionally, to further enhance prototype consistency and reduce intraclass variance, the cosine similarity loss between prototypes and features is incorporated as a soft loss term to optimize intra-class feature clustering. The specific definition is as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21452" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.098, + 0.81, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.194, + 0.312, + 0.306, + 0.332 + ], + "angle": 0, + "content": "random perturbation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.323, + 0.316, + 0.416, + 0.331 + ], + "angle": 0, + "content": "C cosine similarity" + }, + { + "type": "image_footnote", + "bbox": [ + 0.195, + 0.337, + 0.437, + 0.356 + ], + "angle": 0, + "content": "\\(\\text{M}\\) multiply pixel by pixel protetype generation" + }, + { + "type": "image_footnote", + "bbox": [ + 0.232, + 0.368, + 0.382, + 0.381 + ], + "angle": 0, + "content": "1 region1: reliable stable region" + }, + { + "type": "image_footnote", + "bbox": [ + 0.231, + 0.389, + 0.391, + 0.402 + ], + "angle": 0, + "content": "2 region2: reliable unstable region" + }, + { + "type": "image_footnote", + "bbox": [ + 0.232, + 0.411, + 0.392, + 0.425 + ], + "angle": 0, + "content": "3 region3:unreliable stable region" + }, + { + "type": "image", + "bbox": [ + 0.443, + 0.307, + 0.805, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Figure 2. The process of our proposed ALHVR framework. This method first classifies the perturbation predictions of the network output into value regions, then applies an adaptive training strategy to high-value regions, where CG-CPCL is used to train Region 2 and DTCT is used to train Region 3. The supervised learning process and the pseudo-label conversion process are omitted in the figure." + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.504, + 0.483, + 0.522 + ], + "angle": 0, + "content": "\\[\nL _ {p c} ^ {a} = L _ {c e} ^ {a} \\left(\\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\hat {y} _ {b 1} ^ {p}\\right) + L _ {\\cos} ^ {a} \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.529, + 0.483, + 0.548 + ], + "angle": 0, + "content": "\\[\nL _ {p c} ^ {b} = L _ {c e} ^ {b} \\left(\\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\hat {y} _ {a 1} ^ {p}\\right) + L _ {\\cos} ^ {b} \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.483, + 0.643 + ], + "angle": 0, + "content": "where \\( F_{a2}^{d} \\) and \\( F_{b2}^{d} \\) are perturbed features from the penultimate layer of the decoder, \\( p_{a1}^{c} \\) and \\( p_{b1}^{c} \\) denotes the c-th class prototype computed from the original features of the decoder, \\( \\hat{y}_{a1}^{p} \\) and \\( \\hat{y}_{b1}^{p} \\) denote the pseudo-labels generated from \\( \\hat{y}_{a1} \\) and \\( \\hat{y}_{b1} \\), respectively. \\( L_{ce}(\\cdot) \\) represents the cross-entropy loss, and \\( L_{\\cos}(\\cdot) \\) denotes the cosine similarity loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.748 + ], + "angle": 0, + "content": "For unreliable stable region, the mean squared error loss is introduced through DTCT. To further constrain the model's probability distribution and reduce uncertainty in these unreliable regions, we incorporate the entropy of the model's predictions as a regularization term, thereby enhancing the model's robustness in such regions. This is formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.75, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\nL _ {m h} ^ {a} = L _ {m s e} ^ {a} \\left(\\hat {y} _ {a 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {a 1} \\left(\\hat {y} _ {a 1}\\right) + H _ {a 2} \\left(\\hat {y} _ {a 2}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.771, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\nL _ {m h} ^ {b} = L _ {m s e} ^ {b} \\left(\\hat {y} _ {b 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {b 1} \\left(\\hat {y} _ {b 1}\\right) + H _ {b 2} \\left(\\hat {y} _ {b 2}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.793, + 0.483, + 0.823 + ], + "angle": 0, + "content": "where \\(L_{mse}(\\cdot)\\) represents the mean squared error loss, and \\(H(\\cdot)\\) denotes the entropy." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.351, + 0.838 + ], + "angle": 0, + "content": "The unsupervised loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.838, + 0.482, + 0.857 + ], + "angle": 0, + "content": "\\[\nL _ {u} ^ {a} = L _ {p c} ^ {a} + L _ {m h} ^ {a}, L _ {u} ^ {b} = L _ {p c} ^ {b} + L _ {m h} ^ {b}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.863, + 0.396, + 0.877 + ], + "angle": 0, + "content": "The total loss function for each network is:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nL _ {a} = L _ {s} ^ {a} + \\lambda L _ {u} ^ {a}, L _ {b} = L _ {s} ^ {b} + \\lambda L _ {u} ^ {b}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.907, + 0.549 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a time-dependent Gaussian warming up function [12], used to balance the supervised and unsupervised losses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.847, + 0.573 + ], + "angle": 0, + "content": "3.2. Valuable Region Classification Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.791 + ], + "angle": 0, + "content": "Existing semi-supervised learning methods typically use fixed thresholds to separate high-confidence and low-confidence pixels. However, this binary approach provides a coarse distinction of pixel prediction difficulty, neglecting the potential information in low-confidence pixels. To address these limitations, an adaptive threshold is employed to refine pixel classification. Specifically, we directly derive the adaptive threshold from the confidence maps produced by two networks, ensuring better adaptation to current data variations. The confidence maps are generated using perturbed predictions: \\( \\text{con}_{a2} = \\max(\\hat{y}_{a2}), \\text{con}_{b2} = \\max(\\hat{y}_{b2}) \\). By computing the pixel-wise average of the two confidence maps and sorting the values in ascending order, we obtain the sorted average confidence vector \\( \\text{con} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.832 + ], + "angle": 0, + "content": "Finally, the lowest \\( K \\) of the sorted confidence values are selected, and their mean is computed as the adaptive threshold:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.827, + 0.905, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\operatorname {c o n} (i), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(K = \\beta \\cdot (H\\times W\\times Z)\\) , and \\(\\beta\\) is a hyperparameter within the range \\((0,1)\\) , controlling the proportion of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.104, + 0.825, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.908, + 0.31 + ], + "angle": 0, + "content": "Figure 3. The structure of the DTCT module. By comparing the confidence of pixels in the Region 3 from the unperturbed predictions, the more reliable prediction is selected as the teacher. A probability sharpening function is then applied to adjust the probability distribution, generating more reliable pseudo-labels to supervise the perturbed predictions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.331, + 0.195, + 0.345 + ], + "angle": 0, + "content": "selected pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.452 + ], + "angle": 0, + "content": "Using the confidence threshold, we classify predicted pixels from the two networks interactively. By comparing the confidence values of corresponding pixels with the threshold, the pixels are categorized into three regions including reliable stable region \\(\\Omega_{1}\\), reliable unstable region \\(\\Omega_{2}\\) and the unreliable stable region \\(\\Omega_{3}\\) (as described in Chapter 1). These regions can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.463, + 0.483, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {1} = \\left\\{c o n _ {a 2} \\geq \\gamma \\& c o n _ {b 2} \\geq \\gamma \\right\\}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.489, + 0.483, + 0.514 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {2} = \\left\\{\\left(c o n _ {a 2} > \\gamma \\& c o n _ {b 2} < \\gamma\\right) \\mid \\right. \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.511, + 0.411, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\left. \\left(c o n _ {a 2} < \\gamma \\& c o n _ {b 2} > \\gamma\\right) \\right\\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.531, + 0.483, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {3} = \\left\\{c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma \\right\\}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.675 + ], + "angle": 0, + "content": "The reliable stable region is easily learnable by both networks, while the reliable unstable region reflects areas with unclear predictions, and the unreliable stable region represents difficult-to-learn areas. Thus, focusing on the reliable unstable region and unreliable stable region, rather than the reliable stable region, can better enhance the model's learning ability and provide higher learning value, as confirmed by the experimental results in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.484, + 0.716 + ], + "angle": 0, + "content": "3.3. Confidence-Guided Cross-Prototype Consistency Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Due to complex semantic information in reliable unstable region and differences in model learning capabilities, different models produce divergent predictions for the same region. To address this, we propose the CG-CPCL module, which introduces prototypes as category center references, providing a global and stable reference point. Similar to UPCoL [15], we approximate the probability distribution of voxels by calculating cosine similarity between features and prototypes, introducing global constraints to prevent overreliance on local features. Additionally, we use confidence-based conditional guidance to ensure the model learns only from reliable targets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.45 + ], + "angle": 0, + "content": "Specifically, we first generate class prototypes for the unlabeled data within the current mini-batch. We upsample the features from the second-to-last layer of the decoder using trilinear interpolation to match the size of the pseudolabels. Then, we use the pseudo-labels to select the corresponding class features, and generate class prototypes by calculating the mean of these features. This process is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.452, + 0.906, + 0.488 + ], + "angle": 0, + "content": "\\[\np _ {a 1} ^ {c} = \\frac {\\sum F _ {a 1} ^ {d} \\cdot \\mathbb {I} (\\hat {y} _ {a 1} ^ {p} = c)}{\\sum (\\hat {y} _ {a 1} ^ {p} = c)}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.601 + ], + "angle": 0, + "content": "where \\(F_{a1}^{d}\\) represents the original features from the decoder, and \\(\\mathbb{I}(\\cdot)\\) denotes the indicator function. For simplicity, the summation process omits the dimensional details of the features and pseudo-labels. Use the same method to calculate the \\(p_{b1}^{c}\\). After obtaining the class prototypes, we approximate the model's prediction probability by calculating the cosine similarity between the features and the prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.608, + 0.905, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} _ {a} = \\cos \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\operatorname {s i m} _ {b} = \\cos \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.741 + ], + "angle": 0, + "content": "where \\( \\cos(a, b) = \\frac{a \\cdot b}{\\|a\\|_2 \\cdot \\|b\\|_2} \\). To ensure that low-confidence predictions in the \\( \\Omega_2 \\) region only learn from reliable targets, we need to perform pixel selection on the confidence maps predicted by each network. This generates a position mask that contains the interactive information from both networks, which will be used as a guiding signal for the model's learning:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.745, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\nM a s k _ {a} = \\operatorname {c o n} _ {a 2} < \\gamma \\& \\operatorname {c o n} _ {b 2} > \\gamma , \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.77, + 0.905, + 0.785 + ], + "angle": 0, + "content": "\\[\nM a s k _ {b} = \\operatorname {c o n} _ {a 2} > \\gamma \\& \\operatorname {c o n} _ {b 2} < \\gamma . \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.906, + 0.833 + ], + "angle": 0, + "content": "Then, the position mask can be used to calculate the cross-prototype consistency loss, and the formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.835, + 0.905, + 0.852 + ], + "angle": 0, + "content": "\\[\nL _ {c e} ^ {a} = L _ {c e} \\left(\\operatorname {s i m} _ {a}, \\hat {y} _ {b 1} ^ {p}\\right) \\odot \\operatorname {M a s k} _ {a}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents the pixel-wise multiplication operation. Calculate \\(L_{ce}^{b}\\) in the same way. Moreover, to further enhance the model's ability to learn class features, the cosine" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21454" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "similarity loss between the features and prototypes is also calculated:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.127, + 0.483, + 0.143 + ], + "angle": 0, + "content": "\\[\nL _ {\\cos} ^ {a} = (1 - \\operatorname {s i m} _ {a}) \\odot \\operatorname {M a s k} _ {a}. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.201 + ], + "angle": 0, + "content": "We calculate \\( L_{\\mathrm{cos}}^b \\) similarly. Finally, the total loss for each model training the reliable unstable region is a linear combination of \\( L_{ce} \\) and \\( L_{\\mathrm{cos}} \\), as shown in Eqs. (5) and (6)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.437, + 0.225 + ], + "angle": 0, + "content": "3.4. Dynamic Teacher Competition Teaching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.484, + 0.32 + ], + "angle": 0, + "content": "The unreliable stable region is an area where both models struggle to make accurate predictions and typically exhibits high uncertainty. The module improves the adaptability of the model to complex regions by reducing the influence of noise and uncertainty in the training process. The structure of the DTCT module is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.517 + ], + "angle": 0, + "content": "To generate the new teacher prediction, we perform pixel-wise confidence comparison based on the confidence maps of predictions without feature perturbation. We retain the pixels with higher confidence as the new prediction while discarding the pixels with lower confidence. Specifically, we first generate the confidence maps of the two predictions without feature perturbation: \\( \\text{con}_{a1} = \\max(\\hat{y}_{a1}), \\text{con}_{b1} = \\max(\\hat{y}_{b1}) \\). Then, we compare the generated confidence maps and select the prediction probabilities corresponding to the pixels with higher confidence as the new teacher prediction. Next, we apply a probability sharpening function [29] to convert this into a soft pseudolabel. This process can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.521, + 0.483, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} = \\left\\{ \\begin{array}{l} \\hat {y} _ {a 1}, c o n _ {a 1} \\geq c o n _ {b 1} \\\\ \\hat {y} _ {b 1}, c o n _ {a 1} < c o n _ {b 1} \\end{array} , \\right. \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.56, + 0.483, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t} ^ {\\mathrm {s p}} = \\frac {\\hat {y} _ {t} ^ {1 / T}}{\\hat {y} _ {t} ^ {1 / T} + (1 - \\hat {y} _ {t}) ^ {1 / T}}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.68 + ], + "angle": 0, + "content": "where \\( T \\) is a hyperparameter that controls the output distribution of the model. Since the DTCT module is only used to train the pixels in the unreliable stable region, a location mask needs to be generated first to determine the region location of \\( \\Omega_3 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.686, + 0.483, + 0.701 + ], + "angle": 0, + "content": "\\[\nM a s k = c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma . \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.757 + ], + "angle": 0, + "content": "Then, the newly generated reliable pseudo-labels are used to supervise the perturbed version of the predictions, as shown in the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.762, + 0.483, + 0.78 + ], + "angle": 0, + "content": "\\[\nL _ {m s e} ^ {a} = L _ {m s e} \\left(\\hat {y} _ {a 2}, \\hat {y} _ {t} ^ {s p}\\right) \\odot M a s k. \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.879 + ], + "angle": 0, + "content": "We can calculate \\( L_{mse}^{b} \\) in the same way. To enhance the model's ability to perceive uncertainty in the unreliable stable region, guide the model to focus on regions with higher uncertainty, and effectively reduce the uncertainty in these regions, we quantify uncertainty using entropy and add it as a regularization term to the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nH _ {a 1} = H \\left(\\hat {y} _ {a 1}\\right) \\odot \\text {M a s k}, H _ {a 2} = H \\left(\\hat {y} _ {a 2}\\right) \\odot \\text {M a s k}, \\tag {25}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.903, + 0.313 + ], + "angle": 0, + "content": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net7(10%)081.5970.748.072.35
U-Net14(20%)084.8375.337.612.09
U-Net70(All)091.1084.044.901.14
MT[25][NIPS'2017]83.6573.1513.453.61
CPS[6][CVPR'2021]85.3475.508.782.37
MC-Net[28][MICCAI'2021]86.0776.5811.483.37
MCF[27][CVPR'2023]85.1875.2210.732.78
AC-MT[31][MIA'2023]86.3676.869.642.59
UG-MCL[36][AIIM'2023]7(10%)63(90%)86.8477.607.771.94
EVIL[7][CBM'2024]87.2578.185.311.35
AAU[1][MIA'2024]86.4676.998.002.06
MLRP[24][MIA'2024]87.1078.164.911.30
ours90.56[3,31]83.212.570.78
MT[25][NIPS'2017]86.5577.496.862.15
CPS[6][CVPR'2021]87.0378.136.662.12
MC-Net[28][MICCAI'2021]86.5877.6815.994.93
MCF[27][CVPR'2023]87.3278.436.532.00
AC-MT[31][MIA'2023]87.9979.347.852.14
UG-MCL[36][AIIM'2023]14(20%)56(80%)88.4079.869.122.44
EVIL[7][CBM'2024]88.3479.906.461.68
AAU[1][MIA'2024]87.7178.928.512.14
MLRP[24][MIA'2024]88.0079.545.411.58
ours91.09[2,69]84.052.710.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.317, + 0.905, + 0.345 + ], + "angle": 0, + "content": "Table 2. Comparisons with state-of-the-art semi-supervised segmentation methods on the ACDC dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.429 + ], + "angle": 0, + "content": "where \\(H(\\hat{y}) = -\\hat{y}\\log \\hat{y}\\). We can use the same method to calculate the \\(H_{b1}\\) and \\(H_{b2}\\). Finally, the total loss for training the unreliable stable region in each model is a linear combination of \\(L_{mse}\\) and \\(H\\), as shown in Eqs. (7) and (8)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.443, + 0.649, + 0.46 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.802, + 0.483 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.581 + ], + "angle": 0, + "content": "The ACDC dataset [4] consists of cardiac dynamic MRI scans and corresponding ground truth from 100 patients, categorized into three classes: right ventricle, left ventricle, and myocardium. Following the experimental setup in CTCT [17], we used scans from 70 patients for training, 10 patients for validation, and 20 patients for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.656 + ], + "angle": 0, + "content": "The AbdomenCT-1K dataset [18] contains over 1,000 (1K) CT scans from 12 medical centers, categorized into four classes: liver, kidneys, spleen, and pancreas. Following the experimental setup in BFFC [38], we selected 30 scans for training and 11 scans for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.717 + ], + "angle": 0, + "content": "The Brats dataset [3] consists of preoperative MRI scans from 335 glioma patients. Following the experimental setup in CPCL [30], we used 250 samples for training, 25 samples for validation, and 60 for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Metrics. We evaluate the method using four metrics categorized into region-sensitive and edge-sensitive types. Region-sensitive metrics include Dice and Jaccard, while edge-sensitive metrics include 95HD and ASD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.731, + 0.804 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "All experiments were implemented using the PyTorch 1.7 on NVIDIA GeForce RTX 3090 GPU. The model is trained using the SGD optimizer with a weight decay factor of 0.0001, a momentum of 0.9 and an initial learning rate of 0.01. For segmentation on the ACDC and AbdomenCT-1K datasets, we use U-Net [22] as the backbone, with a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.476, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.485, + 0.402 + ], + "angle": 0, + "content": "Figure 4. Visualization of some semi-supervised methods. The first two rows show the segmentation results on the ACDC dataset trained with \\(10\\%\\) labeled data. The last two rows show the segmentation results on the AbdomenCT-1K dataset trained with \\(5\\%\\) labeled data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.423, + 0.483, + 0.606 + ], + "angle": 0, + "content": "batch size of 16 and \\(30\\mathrm{k}\\) training iterations. For Brats, we use V-Net [19] as the backbone, with a batch size of 4 and \\(15\\mathrm{k}\\) training iterations. In the total loss function of the network, the Gaussian warming up function is defined as \\(\\lambda (t) = 0.1\\times e^{-5(1 - t / t_{\\mathrm{max}})^2}\\), where \\(\\mathfrak{t}\\) denotes the current number of iterations and \\(t_\\mathrm{max}\\) represents the maximum number of iterations. In the probability sharpening function, the hyperparameter \\(T\\) is usually empirically set to 0.1. For calculating the confidence threshold, the hyperparameter \\(\\beta\\) values were configured as 0.8, 0.7, and 0.7 in the ACDC, AbdomenCT-1K, and Brats datasets, respectively (set by Ablation Studies)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.619, + 0.46, + 0.635 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.643, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Firstly, we evaluate the proposed method on the ACDC dataset, comparing it with MT [25], CPS [6], MC-Net [28], MCF [27], AC-MT [31], UG-MCL [36], EVIL [7], AAU [1], and MLRP [24]. Additionally, we use U-Net trained with \\(10\\%\\), \\(20\\%\\), and \\(100\\%\\) of labeled data as the baseline and upper bound for reference. As shown in Tab. 2, our method achieves significant improvements across all four evaluation metrics compared to the baseline. Specifically, with \\(10\\%\\) labeled data, the Dice score increases from \\(81.59\\%\\) to \\(90.56\\%\\). With \\(20\\%\\) labeled data, it improves from \\(84.83\\%\\) to \\(91.09\\%\\), approaching the upper bound of \\(91.10\\%\\) achieved with full supervision. Compared to state-of-the-art methods, ALHVR consistently outperforms others. With \\(10\\%\\) labeled data, it surpasses the second-best method (EVIL) by \\(3.31\\%\\) in Dice. When using \\(20\\%\\) labeled data, it outperforms UG-MCL by \\(2.69\\%\\) in Dice." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To further validate the generalization capability of the" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.903, + 0.315 + ], + "angle": 0, + "content": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net1(5%)067.3957.1446.8711.37
U-Net3(10%)075.6666.3322.446.89
U-Net30(All)086.9379.2711.552.45
MT[25][NIPS'2017]71.8261.1942.039.23
CPS[6][CVPR'2021]78.1468.3226.256.68
MC-Net[28][MICCAI'2021]80.9771.4332.987.48
MCF[27][CVPR'2023]80.7971.3026.137.81
AC-MT[31][MIA'2023]76.1865.9936.128.89
UG-MCL[36][AIIM'2023]1(5%)29(95%)74.0663.5826.937.06
EVIL[7][CBM'2024]80.6871.4723.726.33
AAU[1][MIA'2024]80.6571.0332.689.04
MLRP[24][MIA'2024]80.3070.6720.945.43
ours84.7576.3521.155.59
MT[25][NIPS'2017]78.7569.1025.356.88
CPS[6][CVPR'2021]82.7174.1117.304.90
MC-Net[28][MICCAI'2021]82.9474.1319.354.59
MCF[27][CVPR'2023]83.3474.6316.974.72
AC-MT[31][MIA'2023]82.3272.9726.527.54
UG-MCL[36][AIIM'2023]3(10%)27(90%)81.4372.5620.295.67
EVIL[7][CBM'2024]84.2475.7518.145.25
AAU[1][MIA'2024]81.4271.8229.698.25
MLRP[24][MIA'2024]84.5575.9913.824.05
ours85.6877.3521.966.22
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.346 + ], + "angle": 0, + "content": "Table 3. Comparisons with state-of-the-art semi-supervised segmentation methods on the AbdomenCT-1K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.581 + ], + "angle": 0, + "content": "proposed method across different datasets, we evaluate its performance on the five-class AbdomenCT-1K dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 3, with \\(5\\%\\) labeled data, ALHVR improves the Dice score by \\(17.36\\%\\) compared to the baseline. When increasing the labeled data to \\(20\\%\\), the Dice score improves by \\(10.02\\%\\). Compared to state-of-the-art methods, ALHVR consistently achieves the best performance in segmenting four organs. Specifically, with \\(5\\%\\) and \\(10\\%\\) labeled data, it surpasses the second-best methods (MC-Net and MLRP) by \\(3.78\\%\\) and \\(1.13\\%\\) in Dice, respectively. Furthermore, as illustrated in Fig. 4, our method provides more precise segmentation of fine structures and boundary regions across different organs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.687 + ], + "angle": 0, + "content": "In addition to organ segmentation tasks, we evaluated our method on the Brats dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 4, with only \\(10\\%\\) labeled data, ALHVR improves the Dice score by \\(5.91\\%\\) compared to the baseline. Furthermore, it outperforms state-of-the-art methods, achieving a \\(1.18\\%\\) higher Dice score than the second-best method (MLRP)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.68, + 0.714 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.721, + 0.906, + 0.797 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies to validate the effectiveness of each part and analyze the hyperparameters of the proposed method. The experiments were conducted on the ACDC (labeled ratio: \\(10\\%\\)), AbdomenCT-1K (labeled ratio: \\(5\\%\\)), and Brats (labeled ratio: \\(10\\%\\)) datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.759, + 0.821 + ], + "angle": 0, + "content": "4.4.1. Impact of Different Modules" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In Tab. 5, we examine the effectiveness of the major components of ALHVR, CG-CPCL, and DTCT, on different datasets. When CG-CPCL training (learning only reliable unstable region) was used, the performance of all datasets was significantly improved compared with the baseline" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21456" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.478, + 0.214 + ], + "angle": 0, + "content": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
V-Net25(10%)079.2368.4015.834.34
V-Net250(All)086.4977.238.681.98
MT[25][NIPS'2017]80.4969.9911.713.46
CPS[6][CVPR'2021]82.7072.3012.993.46
MC-Net[28][MICCAI'2021]82.5172.9814.573.78
MCF[27][CVPR'2023]81.9671.7613.963.99
AC-MT[31][MIA'2023]25(10%)225(90%)82.0971.6515.914.81
UG-MCL[36][AHIM'2023]82.8572.6911.122.28
EVIL[7][CBM'2024]82.6072.4611.442.76
AAU[1][MIA'2024]82.4372.3313.834.18
MLRP[24][MIA'2024]83.9674.0015.164.21
ours85.14↑1.1875.757.792.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.246 + ], + "angle": 0, + "content": "Table 4. Comparisons with state-of-the-art semi-supervised segmentation methods on the Brats dataset." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.252, + 0.452, + 0.393 + ], + "angle": 0, + "content": "
DatasetsCG-CPCLDTCTDice↑Jaccard↑95HD↓ASD↓
ACDC✓86.6177.147.302.15
89.3281.313.451.02
✓89.1980.973.260.94
✓✓90.5683.212.570.78
AbdomenCT-1K✓79.1969.2129.777.56
83.9875.919.592.81
✓83.7975.2521.395.28
✓✓84.7576.3521.155.59
Brats✓83.2073.7311.913.01
84.6274.818.922.36
✓84.8575.229.962.61
✓✓85.1475.757.792.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.44 + ], + "angle": 0, + "content": "Table 5. Ablation results of different components of ALHVR. We conducted experiments to test the impact of CG-CPCL and DTCT on model performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.627 + ], + "angle": 0, + "content": "method, especially on the AbdomenCT-1K dataset, where the Dice index improved by \\(4.79\\%\\). In the ACDC dataset, when cross pseudo supervision similar to CPS was applied (see Tab. 1 for experimental results), CG-CPCL increased by \\(0.86\\%\\). The reliable unstable region usually has feature differences within the class, which is prone to misjudgment. Therefore, the class prototype is constructed, and cross-prototype consistency is used to enhance the model's discrimination. In addition, the confidence-based conditional guidance mechanism effectively reduces the performance difference between models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.63, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Similarly, with DTCT training (learning only the unreliable stable region), the performance of all datasets improved significantly, with the Brats dataset improving by \\(1.65\\%\\). On the ACDC dataset, CG-CPCL improved by \\(0.99\\%\\) when cross pseudo supervision similar to CPS was used. The unreliable stable region has high uncertainty, and a single model is prone to bias. To this end, we improved the accuracy of the model in this area by integrating predictions from multiple models, competitively generating teacher predictions, and selecting the most credible predictions as pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.796, + 0.483, + 0.9 + ], + "angle": 0, + "content": "When CG-CPCL and DTCT are combined, the model can achieve optimal segmentation performance by using differentiated learning strategies to optimize both reliable unstable and unreliable stable regions. Compared with the baseline model, the Dice indexes of multiple datasets improved by \\(3.95\\%\\), \\(5.56\\%\\), and \\(1.94\\%\\), respectively. On the ACDC dataset, ALHVR increased the Dice indicator from" + }, + { + "type": "table", + "bbox": [ + 0.579, + 0.09, + 0.842, + 0.141 + ], + "angle": 0, + "content": "
β0.50.60.70.80.9
ACDC89.8889.9490.2690.5689.85
AbdomenCT-1K84.1684.3984.7584.4983.71
Brats84.7185.0285.1484.8484.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.146, + 0.905, + 0.172 + ], + "angle": 0, + "content": "Table 6. Ablation studies on the hyperparameter \\( \\beta \\) with \\( {10}\\% ,5\\% \\) and \\( {10}\\% \\) labeled data." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.2, + 0.905, + 0.305 + ], + "angle": 0, + "content": "88.58% to 90.56%. Since these two types of regions have different distributions and prediction characteristics, a unified training strategy is difficult to account for both. Therefore, our method optimizes the two types of regions with targeted strategies, effectively improving the overall performance of the model and its generalization ability under complex scenarios." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.32, + 0.782, + 0.336 + ], + "angle": 0, + "content": "4.4.2. Hyperparameter Value Analysis" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.34, + 0.905, + 0.581 + ], + "angle": 0, + "content": "We investigate the impact of confidence thresholds calculated based on different pixel proportions on model training. As shown in Tab. 6, the performance of dynamic thresholds on multiple datasets is evaluated using Dice coefficients. The results indicate that the optimal proportions are 0.8 for the ACDC dataset and 0.7 for the AbdomenCT-1K and Brats datasets. When the proportion is set lower, the threshold decreases, causing the model to focus more on high-confidence pixels, which may reduce its ability to handle difficult samples and degrade the quality of pseudolabels. Conversely, when the threshold increases, the model may overlook subtle features, affecting detail prediction. Therefore, it is necessary to balance the pixel proportion to optimize the model's attention on the reliable unstable region and the unreliable stable region, thereby improving performance and robustness." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.602, + 0.631, + 0.617 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.629, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this paper, we propose a novel semi-supervised learning framework, ALHVR, which explores high-value learning regions and designs adaptive training strategies. Specifically, CG-CPCL is used to train the reliable unstable region, while DTCT is employed for the unreliable stable region. To enhance stability in the reliable unstable region, CG-CPCL enables low-confidence predictions to learn from high-confidence predictions through confidence-guided cross-pseudo supervision and prototype consistency constraints, improving intra-class consistency and inter-class discriminability. To enhance reliability in the unreliable stable region, DTCT dynamically selects more reliable pixels as teachers using a dynamic teacher selection mechanism and competitive teaching strategy, reducing noise interference while strengthening the model's learning capability. Experimental results on three public datasets demonstrate that ALHVR achieves better performance than existing state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21457" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62271296, 62201334), the Shaanxi Province Innovation Capacity Support Program (Grant No. 2025RS-CXTD-012), and the Scientific Research Program funded by the Shaanxi Provincial Education Commission (Grant Nos. 23JP022, 23JP014)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.238 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.248, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Sukesh Adiga, Jose Dolz, and Herve Lombaert. Anatomically-aware uncertainty for semi-supervised image segmentation. Medical Image Analysis, 91:103011, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.305, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11514-11524, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.376, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[3] Spyridon (Spyros) Bakas. Brats miccai brain tumor dataset, 2020. https://dx.doi.org/10.21227/hdtd-5j88.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.484, + 0.501 + ], + "angle": 0, + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[5] Faquan Chen, Jingjing Fei, Yaqi Chen, and Chenxi Huang. Decoupled consistency for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 551-561. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[7] Yingyu Chen, Ziyuan Yang, Chenyu Shen, Zhiwen Wang, Zhongzhou Zhang, Yang Qin, Xin Wei, Jingfeng Lu, Yan Liu, and Yi Zhang. Evidence-based uncertainty-aware semi-supervised medical image segmentation. Computers in Biology and Medicine, 170:108004, 2024. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[8] Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4070-4080, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[9] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, 130:108777, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Hejun Huang, Zuguo Chen, Chaoyang Chen, Ming Lu, and Ying Zou. Complementary consistency semi-supervised learning for 3d left atrial image segmentation. Computers in Biology and Medicine, 165:107368, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.248, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[11] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Tao Lei, Yi Wang, Xingwu Wang, Xuan Wang, Bin Hu, and Asoke K Nandi. Unified feature consistency of underperforming pixels and valid regions for semi-supervised medical image segmentation. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[14] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[15] Wenjing Lu, Jiahao Lei, Peng Qiu, Rui Sheng, Jinhua Zhou, Xinwu Lu, and Yang Yang. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 662-672. Springer, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[16] Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, and Shaoting Zhang. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 318-329. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[17] Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In International conference on medical imaging with deep learning, pages 820-833. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[18] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695-6714, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[19] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[20] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369-1378, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12674-12684, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21458" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[23] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[24] Jiawei Su, Zhiming Luo, Sheng Lian, Dazhen Lin, and Shaozi Li. Mutual learning with reliable pseudo label for semi-supervised medical image segmentation. Medical Image Analysis, 94:103111, 2024. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[25] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[26] Bing Wang, Taifeng Huang, Shuo Yang, Ying Yang, Junhai Zhai, and Xin Zhang. Dual-decoder data decoupling training for semi-supervised medical image segmentation. Biomedical Signal Processing and Control, 100:106984, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[27] Yongchao Wang, Bin Xiao, Xiuli Bi, Weisheng Li, and Xinbo Gao. Mcf: Mutual correction framework for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15651-15660, 2023. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[28] Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, and Lei Zhang. Semi-supervised left atrium segmentation with mutual consistency training. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part II 24, pages 297-306. Springer, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[29] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[30] Zhe Xu, Yixin Wang, Donghuan Lu, Lequan Yu, Jiangpeng Yan, Jie Luo, Kai Ma, Yefeng Zheng, and Raymond Kaiyu Tong. All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 26(7):3174-3184, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[31] Zhe Xu, Yixin Wang, Donghuan Lu, Xiangde Luo, Jiangpeng Yan, Yefeng Zheng, and Raymond Kai-yu Tong. Ambiguity-selective consistency regularization for meantacher semi-supervised medical image segmentation. Medical Image Analysis, 88:102880, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[32] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4268-4277, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[33] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7236-7246, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[34] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical image computing and computer assisted intervention-MICCAI 2019: 22nd international conference, Shenzhen, China, October 13–17, 2019, proceedings, part II 22, pages 605–613. Springer, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.303 + ], + "angle": 0, + "content": "[35] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[36] Yichi Zhang, Rushi Jiao, Qingcheng Liao, Dongyang Li, and Jicong Zhang. Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation. Artificial Intelligence in Medicine, 138:102476, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[37] Xiangyu Zhao, Zengxin Qi, Sheng Wang, Qian Wang, Xuehai Wu, Ying Mao, and Lichi Zhang. Rcps: Rectified contrastive pseudo supervision for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 28(1):251-261, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[38] Zhongda Zhao, Haiyan Wang, Tao Lei, Xuan Wang, Xiaohong Shen, and Haiyang Yao. Balanced feature fusion collaborative training for semi-supervised medical image segmentation. Pattern Recognition, 157:110856, 2025. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21459" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_origin.pdf b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..571c21fcaf8521cd0a7813108db7456406024a5a --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/6962b9f6-7f28-4f4a-b794-c5ccf66928cc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb451b366987b5a4d09ffc83eee67aa9a3335d46a5051f52a1e49e8cee0b547c +size 1883755 diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/full.md b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4925f4f74a1eb38da68259cb799c9450e3a529a8 --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/full.md @@ -0,0 +1,373 @@ +# Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation + +Tao Lei $^{1*}$ , Ziyao Yang $^{1}$ , Xingwu Wang $^{2}$ , Yi Wang $^{1}$ , Xuan Wang $^{3}$ , Feiman Sun $^{4}$ , Asoke K. Nandi $^{5}$ $^{1}$ Shaanxi University of Science and Technology, China + $^{2}$ Beijing University of Posts and Telecommunications, China + $^{3}$ Northwestern Polytechnical University, China + $^{4}$ Southwest Jiaotong University, China $^{5}$ Brunel University of London, UK + +*Corresponding author:leitao@sust.edu.cn + +# Abstract + +Existing semi-supervised learning methods typically mitigate the impact of unreliable predictions by suppressing low-confidence regions. However, these methods fail to explore which regions hold higher learning value and how to design adaptive learning strategies for these regions. To address these issues, we propose a novel adaptive learning of high-value regions (ALHVR) framework. By exploiting the diversity of predictions from dual-branch networks, the prediction regions are classified into three groups: reliable stable region, reliable unstable region, and unreliable stable region. For high-value regions (reliable unstable region and unreliable stable region), different training strategies are designed. Specifically, for reliable unstable region, we propose a confidence-guided cross-prototype consistency learning (CG-CPCL) module, which enforces prototype consistency constraints in the feature space. By leveraging confidence information, the high-confidence predictions from one network selectively supervise the low-confidence predictions from the other, thus helping the model learn inter-class discrimination more stably. Additionally, for unreliable stable region, we design a dynamic teacher competition teaching (DTCT) module, which dynamically selects the most reliable pixels as teachers by evaluating the unperturbed predictions from both networks. These selected pixels are then used to supervise perturbed predictions, thereby enhancing the model's learning capability in unreliable region. Experimental results show that our method outperforms state-of-the-art approaches on three public datasets. Code is available at https://github.com/ziziyao/ALHVR. + +# 1. Introduction + +Medical images contain various types of tissues with complex anatomical structures, leading to ambiguous bound + +![](images/e75b38b68f71efdd50fad78b6529c4231dd794f41870f3c05bf3803bf93c3b36.jpg) +Figure 1. Training on the Brats dataset using the CPS framework, where the predictions of a given sample are classified based on a confidence threshold. Region 1 represents the reliable stable region, Region 2 represents the reliable unstable region, and Region 3 represents the unreliable stable region. + +
DatasetRegion 1Region 2Region 3Dice
ACDC✓✓✓85.34
✓87.13
✓88.46
✓88.20
✓✓87.41
✓✓82.51
✓✓88.58
+ +Table 1. Experimental results of training on the ACDC dataset with $10\%$ labeled data using the CPS framework under different masked region settings. + +aries between target regions and surrounding tissues. This issue increases the segmentation difficulty for the model at boundary regions. Some semi-supervised methods [23, 34] estimate confidence based on predictions from a single network to distinguish between high-confidence and low-confidence pixels. These methods guide the model to focus on high-confidence predictions by discarding low-confidence pixels. However, predictions from a single network are biased, leading to inaccurate pixel classification. Therefore, some approaches evaluate pixel reliability based on dual-network predictions. For instance, MCF [27] treats inconsistent predictions between networks as errors and attempts to correct them, but it fails to handle cases where both predictions are incorrect. MLRP [24] employs a pixelwise comparison strategy to select relatively reliable predic + +tions from two networks as pseudo-labels. However, mutual learning between low-confidence predictions may introduce biases, leading to suboptimal learning. These methods fail to identify or utilize high-value regions, particularly those with high segmentation difficulty. Furthermore, they lack adaptive learning strategies tailored to distinct regions, limiting the model's ability to learn in crucial segmentation regions, such as tissue boundaries. + +In order to resolve the aforementioned issues, this paper combines the diversity of dual-network predictions with confidence thresholds to achieve more precise pixel classification. Taking CPS-based methods [6] as an example, we analyze this problem. When two networks predict the same image, the confidence of each pixel relative to a predefined threshold can be categorized into three types: (1) both networks have confidence scores above the threshold, (2) one network has a confidence score above the threshold while the other is below it, and (3) both networks have confidence scores below the threshold. Based on confidence levels and prediction consistency, we define these three regions as the reliable stable region, the reliable unstable region, and the unreliable stable region, respectively. As shown in Fig. 1, different regions are represented by different colors. To further explore the most valuable learning regions, we conducted experiments on the ACDC dataset using the CPS framework. Specifically, we applied masks to occlude different regions, training the model only on unmasked areas. As shown in Tab. 1, the highest Dice score is achieved when pixels in the reliable stable region are masked out during training process. This indicates that the reliable unstable region and the unreliable stable region are high-value regions. + +To address the lack of adaptive learning strategies for regions of varying value in existing methods, this paper proposes a semi-supervised medical image segmentation method based on adaptive learning for high-value regions (ALHVR). By analyzing the confidence distributions of model predictions, high-value regions are identified, and region-specific training strategies are designed to improve segmentation performance. Specifically, we propose the confidence-guided cross-prototype consistency learning (CG-CPCL) module and the dynamic teacher competition teaching (DTCT) module, which are designed to train the reliable unstable region and the unreliable stable region, respectively. The main contributions of this paper are as follows: + +(1) We conduct an investigation into the high-value regions in semi-supervised medical image segmentation, categorizing them into reliable unstable region and unreliable stable region. Furthermore, we design adaptive learning strategies to enhance segmentation performance in these areas. +(2) We propose CG-CPCL to address the challenge where unstable predictions within a region pre + +vent the model from accurately recognizing key region features. CG-CPCL enhances cross-consistency between features and prototypes by constructing class prototypes while incorporating confidence constraints to ensure that low-confidence predictions learn exclusively from high-confidence predictions. This approach effectively guides the model to learn more reliable feature representations in reliable unstable region, thereby improving its ability to identify key regions. + +(3) We design DTCT to tackle the issue where unreliable predictions within a region lead to the accumulation of erroneous signals, causing prediction bias. DTCT employs pixel-level reliability assessment to dynamically and competitively select more reliable pixels from unperturbed predictions as teacher predictions to supervise perturbed predictions. This method provides the model with more valuable and reliable learning signals, effectively reducing prediction bias. +(4) The proposed framework integrates both learning modules and is evaluated on three public datasets: ACDC, AbdomenCT-1K, and Brats. Experimental results demonstrate that our framework outperforms current state-of-the-art semi-supervised learning methods. + +# 2. Related Work + +# 2.1. Semi-Supervised Medical Image Segmentation + +The key challenge in semi-supervised image segmentation is to effectively exploit unlabeled data to improve model performance. To address this, researchers have proposed various approaches, among which consistency regularization and pseudo-labeling methods have become mainstream directions. + +Firstly, consistency regularization methods are based on the smoothness assumption, which states that a model's predictions should remain similar under different levels of perturbations, including data-level, model-level, and feature-level perturbations. At the data level, common perturbations involve transformations such as rotation and color jittering. To generate more diverse samples, copy-paste techniques are popular, such as CutMix [35] and ClassMix [20]. However, existing copy-paste methods overlook the distribution mismatch between labeled and unlabeled data. To tackle this issue, BCP [2] introduces bidirectional copy-paste between labeled and unlabeled data to generate new samples. At the model level, perturbations are typically introduced by modifying model parameters or structures. For example, MT [25] employs an exponential moving average (EMA) to update the teacher model's parameters with those of the student model, CPS [6] initializes two networks with different parameters and trains them independently. Additionally, MCF [27] and CC-Net [10] introduce variations in encoder or decoder structures, ensuring that different net + +work architectures produce consistent predictions. At the feature level, CCT [21] injects noise or drops activations to enforce consistency between decoders under perturbations. PS-MT [14] uses T-VAT from the teacher model for more accurate student perturbations. Since these consistency regularization methods rely on individual perturbations, ABD [8] dynamically adjusts image displacement, improving the model's adaptability to complex data variations and robustness. + +Secondly, the pseudo-labeling methods are based on the clustering assumption. It expands the dataset by generating pseudo-labels for unlabeled data using a pre-trained model. To improve the quality of pseudo-labels, $\mathrm{ST} + +$ [32] prioritizes reliable unlabeled images based on prediction stability and performs selective retraining. DMT [9] overcomes the limitation of self-training by identifying incorrect pseudo-labels through prediction differences between multiple models. Additionally, some methods combine pseudolabeling with consistency regularization. For example, FixMatch [23] and RCPS [37] encourage the model to produce consistent predictions between strongly augmented data and pseudo-labels generated from weakly augmented data. While these methods enhance model performance through data diversification or improved pseudo-labels' quality, they overlook the varying complexity of predictions across different image regions, and fail to learn more discriminative features from high-value regions. + +# 2.2. Uncertainty-Guided Semi-Supervised Medical Image Segmentation + +In semi-supervised medical image segmentation, uncertainty analysis is a key strategy for effectively exploiting pseudo-labels. Existing methods typically estimate uncertainty using Monte Carlo Dropout [11]. Additionally, some approaches employ confidence scores or entropy measures for uncertainty guidance. Specifically, to enable the model to learn more reliable targets, UA-MT [34] utilizes Monte Carlo Dropout for uncertainty estimation and trains the model only with low-uncertainty pixels. URPC [16] refines uncertainty by computing the KL divergence of predictions, allowing the model to progressively learn from multi-scale reliable predictions. To further improve pseudo-label quality, UniMatch [33] evaluates the confidence of model predictions using the maximum class probability and retains only high-confidence pixels from weakly augmented data as pseudo-labels. MLRP [24] assesses the reliability of different subnet predictions via entropy and uses more reliable predictions as pseudo-labels to supervise the other subnetwork's training. To fully exploit pixel information with varying confidence levels, UFC [13] distinguishes high-performing and under-performing pixels based on confidence differences between the student and teacher models. Both DC-Net [5] and DD-Net [26] decouple predictions us + +ing confidence thresholds and optimize different pixels with distinct loss functions, effectively improving model performance. However, although these methods assess and train pixel reliability using various strategies, they rely solely on simple consistency losses (e.g. cross-entropy loss and mean squared error loss) and fail to design adaptive training strategies based on the characteristics of the data. + +# 3. Method + +# 3.1. Overview of the Framework + +In semi-supervised image segmentation tasks, the training dataset typically consists of a small amount of labeled data and a large amount of unlabeled data. Mathematically, the training dataset is defined as $D = \{D_L, D_U\}$ , where $D_L = \{(x_i^L, y_i^L)\}_{i=1}^N$ represents the labeled dataset containing $N$ samples, and $D_U = \{x_i^U\}_{i=1}^M$ represents the unlabeled dataset containing $M$ samples, with $N \ll M$ . We define the input image as $x_i \in \mathbb{R}^{H \times W \times Z}$ , where $H$ , $W$ , and $Z$ denote the height, width, and depth of the image, respectively. The ground truth corresponding to the labeled data is defined as $y_i \in \{0, 1, \dots, C-1\}^{H \times W \times Z}$ , where $C$ represents the number of classes. + +The training process of ALHVR is shown in Fig. 2, which consists of two parallel networks $f_{A}(\cdot)$ and $f_{B}(\cdot)$ with the same structure but independently updating parameters. Each network's decoder takes the encoded feature $F_{1}$ and its perturbed version $F_{2}$ as inputs to generate predictions from different perspectives. It is worth noting that feature perturbations are generated by randomly adding noise or dropout. + +The original and perturbed features from both networks are fed into their corresponding decoders, $f_{A}^{d}(\cdot)$ and $f_{B}^{d}(\cdot)$ , to generate the final prediction outputs: + +$$ +\hat {y} _ {a 1} = f _ {A} ^ {d} \left(F _ {a 1}\right), \hat {y} _ {a 2} = f _ {A} ^ {d} \left(F _ {a 2}\right), \tag {1} +$$ + +$$ +\hat {y} _ {b 1} = f _ {B} ^ {d} \left(F _ {b 1}\right), \hat {y} _ {b 2} = f _ {B} ^ {d} \left(F _ {b 2}\right), \tag {2} +$$ + +where the decoder outputs include the prediction results for both labeled and unlabeled data, $\hat{y} = \hat{y}^L\cup \hat{y}^U$ + +For labeled data, a combination of cross-entropy loss and dice loss is used as the supervised learning loss function: + +$$ +L _ {s} ^ {a} = L _ {c e} \left(\hat {y} _ {a 1}, y\right) + L _ {d i c e} \left(\hat {y} _ {a 1}, y\right), \tag {3} +$$ + +$$ +L _ {\mathrm {s}} ^ {b} = L _ {c e} \left(\hat {y} _ {b 1}, y\right) + L _ {d i c e} \left(\hat {y} _ {b 1}, y\right). \tag {4} +$$ + +For unlabeled data, an adaptive training strategy is proposed based on high-value regions. Specifically, for reliable unstable region, the CG-CPCL module is designed to introduce a cross-prototype consistency loss. Additionally, to further enhance prototype consistency and reduce intraclass variance, the cosine similarity loss between prototypes and features is incorporated as a soft loss term to optimize intra-class feature clustering. The specific definition is as follows: + +![](images/584537d36a8a929a644cff7bd016a2712a23a439ec0ac716968a93f621f67b7f.jpg) +random perturbation +C cosine similarity + +Figure 2. The process of our proposed ALHVR framework. This method first classifies the perturbation predictions of the network output into value regions, then applies an adaptive training strategy to high-value regions, where CG-CPCL is used to train Region 2 and DTCT is used to train Region 3. The supervised learning process and the pseudo-label conversion process are omitted in the figure. +![](images/58f9d35eee1e1d2f88f4e3a52c74fbc666c823b964642a832cecee5bcef1a2b6.jpg) +$\text{M}$ multiply pixel by pixel protetype generation +1 region1: reliable stable region +2 region2: reliable unstable region +3 region3:unreliable stable region + +$$ +L _ {p c} ^ {a} = L _ {c e} ^ {a} \left(\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\right), \hat {y} _ {b 1} ^ {p}\right) + L _ {\cos} ^ {a} \left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\right), \tag {5} +$$ + +$$ +L _ {p c} ^ {b} = L _ {c e} ^ {b} \left(\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\right), \hat {y} _ {a 1} ^ {p}\right) + L _ {\cos} ^ {b} \left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\right), \tag {6} +$$ + +where $F_{a2}^{d}$ and $F_{b2}^{d}$ are perturbed features from the penultimate layer of the decoder, $p_{a1}^{c}$ and $p_{b1}^{c}$ denotes the c-th class prototype computed from the original features of the decoder, $\hat{y}_{a1}^{p}$ and $\hat{y}_{b1}^{p}$ denote the pseudo-labels generated from $\hat{y}_{a1}$ and $\hat{y}_{b1}$ , respectively. $L_{ce}(\cdot)$ represents the cross-entropy loss, and $L_{\cos}(\cdot)$ denotes the cosine similarity loss. + +For unreliable stable region, the mean squared error loss is introduced through DTCT. To further constrain the model's probability distribution and reduce uncertainty in these unreliable regions, we incorporate the entropy of the model's predictions as a regularization term, thereby enhancing the model's robustness in such regions. This is formally defined as: + +$$ +L _ {m h} ^ {a} = L _ {m s e} ^ {a} \left(\hat {y} _ {a 2}, \left(\hat {y} _ {a 1}, \hat {y} _ {b 1}\right)\right) + H _ {a 1} \left(\hat {y} _ {a 1}\right) + H _ {a 2} \left(\hat {y} _ {a 2}\right), \tag {7} +$$ + +$$ +L _ {m h} ^ {b} = L _ {m s e} ^ {b} \left(\hat {y} _ {b 2}, \left(\hat {y} _ {a 1}, \hat {y} _ {b 1}\right)\right) + H _ {b 1} \left(\hat {y} _ {b 1}\right) + H _ {b 2} \left(\hat {y} _ {b 2}\right), \tag {8} +$$ + +where $L_{mse}(\cdot)$ represents the mean squared error loss, and $H(\cdot)$ denotes the entropy. + +The unsupervised loss is defined as: + +$$ +L _ {u} ^ {a} = L _ {p c} ^ {a} + L _ {m h} ^ {a}, L _ {u} ^ {b} = L _ {p c} ^ {b} + L _ {m h} ^ {b}. \tag {9} +$$ + +The total loss function for each network is: + +$$ +L _ {a} = L _ {s} ^ {a} + \lambda L _ {u} ^ {a}, L _ {b} = L _ {s} ^ {b} + \lambda L _ {u} ^ {b}, \tag {10} +$$ + +where $\lambda$ is a time-dependent Gaussian warming up function [12], used to balance the supervised and unsupervised losses. + +# 3.2. Valuable Region Classification Method + +Existing semi-supervised learning methods typically use fixed thresholds to separate high-confidence and low-confidence pixels. However, this binary approach provides a coarse distinction of pixel prediction difficulty, neglecting the potential information in low-confidence pixels. To address these limitations, an adaptive threshold is employed to refine pixel classification. Specifically, we directly derive the adaptive threshold from the confidence maps produced by two networks, ensuring better adaptation to current data variations. The confidence maps are generated using perturbed predictions: $\text{con}_{a2} = \max(\hat{y}_{a2}), \text{con}_{b2} = \max(\hat{y}_{b2})$ . By computing the pixel-wise average of the two confidence maps and sorting the values in ascending order, we obtain the sorted average confidence vector $\text{con}$ . + +Finally, the lowest $K$ of the sorted confidence values are selected, and their mean is computed as the adaptive threshold: + +$$ +\gamma = \frac {1}{K} \sum_ {i = 1} ^ {K} \operatorname {c o n} (i), \tag {11} +$$ + +where $K = \beta \cdot (H\times W\times Z)$ , and $\beta$ is a hyperparameter within the range $(0,1)$ , controlling the proportion of + +![](images/7e8b48bd560b512da09e8e191307169ae2ac82ae377f4c2fcdadab9e450e7cee.jpg) +Figure 3. The structure of the DTCT module. By comparing the confidence of pixels in the Region 3 from the unperturbed predictions, the more reliable prediction is selected as the teacher. A probability sharpening function is then applied to adjust the probability distribution, generating more reliable pseudo-labels to supervise the perturbed predictions. + +selected pixels. + +Using the confidence threshold, we classify predicted pixels from the two networks interactively. By comparing the confidence values of corresponding pixels with the threshold, the pixels are categorized into three regions including reliable stable region $\Omega_{1}$ , reliable unstable region $\Omega_{2}$ and the unreliable stable region $\Omega_{3}$ (as described in Chapter 1). These regions can be defined as: + +$$ +\Omega_ {1} = \left\{c o n _ {a 2} \geq \gamma \& c o n _ {b 2} \geq \gamma \right\}, \tag {12} +$$ + +$$ +\Omega_ {2} = \left\{\left(c o n _ {a 2} > \gamma \& c o n _ {b 2} < \gamma\right) \mid \right. \tag {13} +$$ + +$$ +\left. \left(c o n _ {a 2} < \gamma \& c o n _ {b 2} > \gamma\right) \right\}, +$$ + +$$ +\Omega_ {3} = \left\{c o n _ {a 2} < \gamma \& c o n _ {b 2} < \gamma \right\}. \tag {14} +$$ + +The reliable stable region is easily learnable by both networks, while the reliable unstable region reflects areas with unclear predictions, and the unreliable stable region represents difficult-to-learn areas. Thus, focusing on the reliable unstable region and unreliable stable region, rather than the reliable stable region, can better enhance the model's learning ability and provide higher learning value, as confirmed by the experimental results in Tab. 1. + +# 3.3. Confidence-Guided Cross-Prototype Consistency Learning + +Due to complex semantic information in reliable unstable region and differences in model learning capabilities, different models produce divergent predictions for the same region. To address this, we propose the CG-CPCL module, which introduces prototypes as category center references, providing a global and stable reference point. Similar to UPCoL [15], we approximate the probability distribution of voxels by calculating cosine similarity between features and prototypes, introducing global constraints to prevent overreliance on local features. Additionally, we use confidence-based conditional guidance to ensure the model learns only from reliable targets. + +Specifically, we first generate class prototypes for the unlabeled data within the current mini-batch. We upsample the features from the second-to-last layer of the decoder using trilinear interpolation to match the size of the pseudolabels. Then, we use the pseudo-labels to select the corresponding class features, and generate class prototypes by calculating the mean of these features. This process is described as follows: + +$$ +p _ {a 1} ^ {c} = \frac {\sum F _ {a 1} ^ {d} \cdot \mathbb {I} (\hat {y} _ {a 1} ^ {p} = c)}{\sum (\hat {y} _ {a 1} ^ {p} = c)}, \tag {15} +$$ + +where $F_{a1}^{d}$ represents the original features from the decoder, and $\mathbb{I}(\cdot)$ denotes the indicator function. For simplicity, the summation process omits the dimensional details of the features and pseudo-labels. Use the same method to calculate the $p_{b1}^{c}$ . After obtaining the class prototypes, we approximate the model's prediction probability by calculating the cosine similarity between the features and the prototypes: + +$$ +\operatorname {s i m} _ {a} = \cos \left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\right), \operatorname {s i m} _ {b} = \cos \left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\right), \tag {16} +$$ + +where $\cos(a, b) = \frac{a \cdot b}{\|a\|_2 \cdot \|b\|_2}$ . To ensure that low-confidence predictions in the $\Omega_2$ region only learn from reliable targets, we need to perform pixel selection on the confidence maps predicted by each network. This generates a position mask that contains the interactive information from both networks, which will be used as a guiding signal for the model's learning: + +$$ +M a s k _ {a} = \operatorname {c o n} _ {a 2} < \gamma \& \operatorname {c o n} _ {b 2} > \gamma , \tag {17} +$$ + +$$ +M a s k _ {b} = \operatorname {c o n} _ {a 2} > \gamma \& \operatorname {c o n} _ {b 2} < \gamma . \tag {18} +$$ + +Then, the position mask can be used to calculate the cross-prototype consistency loss, and the formula is as follows: + +$$ +L _ {c e} ^ {a} = L _ {c e} \left(\operatorname {s i m} _ {a}, \hat {y} _ {b 1} ^ {p}\right) \odot \operatorname {M a s k} _ {a}, \tag {19} +$$ + +where $\odot$ represents the pixel-wise multiplication operation. Calculate $L_{ce}^{b}$ in the same way. Moreover, to further enhance the model's ability to learn class features, the cosine + +similarity loss between the features and prototypes is also calculated: + +$$ +L _ {\cos} ^ {a} = (1 - \operatorname {s i m} _ {a}) \odot \operatorname {M a s k} _ {a}. \tag {20} +$$ + +We calculate $L_{\mathrm{cos}}^b$ similarly. Finally, the total loss for each model training the reliable unstable region is a linear combination of $L_{ce}$ and $L_{\mathrm{cos}}$ , as shown in Eqs. (5) and (6). + +# 3.4. Dynamic Teacher Competition Teaching + +The unreliable stable region is an area where both models struggle to make accurate predictions and typically exhibits high uncertainty. The module improves the adaptability of the model to complex regions by reducing the influence of noise and uncertainty in the training process. The structure of the DTCT module is shown in Fig. 3. + +To generate the new teacher prediction, we perform pixel-wise confidence comparison based on the confidence maps of predictions without feature perturbation. We retain the pixels with higher confidence as the new prediction while discarding the pixels with lower confidence. Specifically, we first generate the confidence maps of the two predictions without feature perturbation: $\text{con}_{a1} = \max(\hat{y}_{a1}), \text{con}_{b1} = \max(\hat{y}_{b1})$ . Then, we compare the generated confidence maps and select the prediction probabilities corresponding to the pixels with higher confidence as the new teacher prediction. Next, we apply a probability sharpening function [29] to convert this into a soft pseudolabel. This process can be defined as: + +$$ +\hat {y} _ {t} = \left\{ \begin{array}{l} \hat {y} _ {a 1}, c o n _ {a 1} \geq c o n _ {b 1} \\ \hat {y} _ {b 1}, c o n _ {a 1} < c o n _ {b 1} \end{array} , \right. \tag {21} +$$ + +$$ +\hat {y} _ {t} ^ {\mathrm {s p}} = \frac {\hat {y} _ {t} ^ {1 / T}}{\hat {y} _ {t} ^ {1 / T} + (1 - \hat {y} _ {t}) ^ {1 / T}}, \tag {22} +$$ + +where $T$ is a hyperparameter that controls the output distribution of the model. Since the DTCT module is only used to train the pixels in the unreliable stable region, a location mask needs to be generated first to determine the region location of $\Omega_3$ : + +$$ +M a s k = c o n _ {a 2} < \gamma \& c o n _ {b 2} < \gamma . \tag {23} +$$ + +Then, the newly generated reliable pseudo-labels are used to supervise the perturbed version of the predictions, as shown in the following formula: + +$$ +L _ {m s e} ^ {a} = L _ {m s e} \left(\hat {y} _ {a 2}, \hat {y} _ {t} ^ {s p}\right) \odot M a s k. \tag {24} +$$ + +We can calculate $L_{mse}^{b}$ in the same way. To enhance the model's ability to perceive uncertainty in the unreliable stable region, guide the model to focus on regions with higher uncertainty, and effectively reduce the uncertainty in these regions, we quantify uncertainty using entropy and add it as a regularization term to the loss function: + +$$ +H _ {a 1} = H \left(\hat {y} _ {a 1}\right) \odot \text {M a s k}, H _ {a 2} = H \left(\hat {y} _ {a 2}\right) \odot \text {M a s k}, \tag {25} +$$ + +
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net7(10%)081.5970.748.072.35
U-Net14(20%)084.8375.337.612.09
U-Net70(All)091.1084.044.901.14
MT[25][NIPS'2017]83.6573.1513.453.61
CPS[6][CVPR'2021]85.3475.508.782.37
MC-Net[28][MICCAI'2021]86.0776.5811.483.37
MCF[27][CVPR'2023]85.1875.2210.732.78
AC-MT[31][MIA'2023]86.3676.869.642.59
UG-MCL[36][AIIM'2023]7(10%)63(90%)86.8477.607.771.94
EVIL[7][CBM'2024]87.2578.185.311.35
AAU[1][MIA'2024]86.4676.998.002.06
MLRP[24][MIA'2024]87.1078.164.911.30
ours90.56[3,31]83.212.570.78
MT[25][NIPS'2017]86.5577.496.862.15
CPS[6][CVPR'2021]87.0378.136.662.12
MC-Net[28][MICCAI'2021]86.5877.6815.994.93
MCF[27][CVPR'2023]87.3278.436.532.00
AC-MT[31][MIA'2023]87.9979.347.852.14
UG-MCL[36][AIIM'2023]14(20%)56(80%)88.4079.869.122.44
EVIL[7][CBM'2024]88.3479.906.461.68
AAU[1][MIA'2024]87.7178.928.512.14
MLRP[24][MIA'2024]88.0079.545.411.58
ours91.09[2,69]84.052.710.73
+ +Table 2. Comparisons with state-of-the-art semi-supervised segmentation methods on the ACDC dataset. + +where $H(\hat{y}) = -\hat{y}\log \hat{y}$ . We can use the same method to calculate the $H_{b1}$ and $H_{b2}$ . Finally, the total loss for training the unreliable stable region in each model is a linear combination of $L_{mse}$ and $H$ , as shown in Eqs. (7) and (8). + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +The ACDC dataset [4] consists of cardiac dynamic MRI scans and corresponding ground truth from 100 patients, categorized into three classes: right ventricle, left ventricle, and myocardium. Following the experimental setup in CTCT [17], we used scans from 70 patients for training, 10 patients for validation, and 20 patients for testing. + +The AbdomenCT-1K dataset [18] contains over 1,000 (1K) CT scans from 12 medical centers, categorized into four classes: liver, kidneys, spleen, and pancreas. Following the experimental setup in BFFC [38], we selected 30 scans for training and 11 scans for testing. + +The Brats dataset [3] consists of preoperative MRI scans from 335 glioma patients. Following the experimental setup in CPCL [30], we used 250 samples for training, 25 samples for validation, and 60 for testing. + +Metrics. We evaluate the method using four metrics categorized into region-sensitive and edge-sensitive types. Region-sensitive metrics include Dice and Jaccard, while edge-sensitive metrics include 95HD and ASD. + +# 4.2. Implementation Details + +All experiments were implemented using the PyTorch 1.7 on NVIDIA GeForce RTX 3090 GPU. The model is trained using the SGD optimizer with a weight decay factor of 0.0001, a momentum of 0.9 and an initial learning rate of 0.01. For segmentation on the ACDC and AbdomenCT-1K datasets, we use U-Net [22] as the backbone, with a + +![](images/57cbfd2b8d4f591c120d9119272c7e84e40bb9b6b87e916b58c911e66fa3dd6b.jpg) +Figure 4. Visualization of some semi-supervised methods. The first two rows show the segmentation results on the ACDC dataset trained with $10\%$ labeled data. The last two rows show the segmentation results on the AbdomenCT-1K dataset trained with $5\%$ labeled data. + +batch size of 16 and $30\mathrm{k}$ training iterations. For Brats, we use V-Net [19] as the backbone, with a batch size of 4 and $15\mathrm{k}$ training iterations. In the total loss function of the network, the Gaussian warming up function is defined as $\lambda (t) = 0.1\times e^{-5(1 - t / t_{\mathrm{max}})^2}$ , where $\mathfrak{t}$ denotes the current number of iterations and $t_\mathrm{max}$ represents the maximum number of iterations. In the probability sharpening function, the hyperparameter $T$ is usually empirically set to 0.1. For calculating the confidence threshold, the hyperparameter $\beta$ values were configured as 0.8, 0.7, and 0.7 in the ACDC, AbdomenCT-1K, and Brats datasets, respectively (set by Ablation Studies). + +# 4.3. Comparison with State-of-the-Art Methods + +Firstly, we evaluate the proposed method on the ACDC dataset, comparing it with MT [25], CPS [6], MC-Net [28], MCF [27], AC-MT [31], UG-MCL [36], EVIL [7], AAU [1], and MLRP [24]. Additionally, we use U-Net trained with $10\%$ , $20\%$ , and $100\%$ of labeled data as the baseline and upper bound for reference. As shown in Tab. 2, our method achieves significant improvements across all four evaluation metrics compared to the baseline. Specifically, with $10\%$ labeled data, the Dice score increases from $81.59\%$ to $90.56\%$ . With $20\%$ labeled data, it improves from $84.83\%$ to $91.09\%$ , approaching the upper bound of $91.10\%$ achieved with full supervision. Compared to state-of-the-art methods, ALHVR consistently outperforms others. With $10\%$ labeled data, it surpasses the second-best method (EVIL) by $3.31\%$ in Dice. When using $20\%$ labeled data, it outperforms UG-MCL by $2.69\%$ in Dice. + +To further validate the generalization capability of the + +
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net1(5%)067.3957.1446.8711.37
U-Net3(10%)075.6666.3322.446.89
U-Net30(All)086.9379.2711.552.45
MT[25][NIPS'2017]71.8261.1942.039.23
CPS[6][CVPR'2021]78.1468.3226.256.68
MC-Net[28][MICCAI'2021]80.9771.4332.987.48
MCF[27][CVPR'2023]80.7971.3026.137.81
AC-MT[31][MIA'2023]76.1865.9936.128.89
UG-MCL[36][AIIM'2023]1(5%)29(95%)74.0663.5826.937.06
EVIL[7][CBM'2024]80.6871.4723.726.33
AAU[1][MIA'2024]80.6571.0332.689.04
MLRP[24][MIA'2024]80.3070.6720.945.43
ours84.7576.3521.155.59
MT[25][NIPS'2017]78.7569.1025.356.88
CPS[6][CVPR'2021]82.7174.1117.304.90
MC-Net[28][MICCAI'2021]82.9474.1319.354.59
MCF[27][CVPR'2023]83.3474.6316.974.72
AC-MT[31][MIA'2023]82.3272.9726.527.54
UG-MCL[36][AIIM'2023]3(10%)27(90%)81.4372.5620.295.67
EVIL[7][CBM'2024]84.2475.7518.145.25
AAU[1][MIA'2024]81.4271.8229.698.25
MLRP[24][MIA'2024]84.5575.9913.824.05
ours85.6877.3521.966.22
+ +Table 3. Comparisons with state-of-the-art semi-supervised segmentation methods on the AbdomenCT-1K dataset. + +proposed method across different datasets, we evaluate its performance on the five-class AbdomenCT-1K dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 3, with $5\%$ labeled data, ALHVR improves the Dice score by $17.36\%$ compared to the baseline. When increasing the labeled data to $20\%$ , the Dice score improves by $10.02\%$ . Compared to state-of-the-art methods, ALHVR consistently achieves the best performance in segmenting four organs. Specifically, with $5\%$ and $10\%$ labeled data, it surpasses the second-best methods (MC-Net and MLRP) by $3.78\%$ and $1.13\%$ in Dice, respectively. Furthermore, as illustrated in Fig. 4, our method provides more precise segmentation of fine structures and boundary regions across different organs. + +In addition to organ segmentation tasks, we evaluated our method on the Brats dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 4, with only $10\%$ labeled data, ALHVR improves the Dice score by $5.91\%$ compared to the baseline. Furthermore, it outperforms state-of-the-art methods, achieving a $1.18\%$ higher Dice score than the second-best method (MLRP). + +# 4.4. Ablation Studies + +In this section, we conduct ablation studies to validate the effectiveness of each part and analyze the hyperparameters of the proposed method. The experiments were conducted on the ACDC (labeled ratio: $10\%$ ), AbdomenCT-1K (labeled ratio: $5\%$ ), and Brats (labeled ratio: $10\%$ ) datasets. + +# 4.4.1. Impact of Different Modules + +In Tab. 5, we examine the effectiveness of the major components of ALHVR, CG-CPCL, and DTCT, on different datasets. When CG-CPCL training (learning only reliable unstable region) was used, the performance of all datasets was significantly improved compared with the baseline + +
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
V-Net25(10%)079.2368.4015.834.34
V-Net250(All)086.4977.238.681.98
MT[25][NIPS'2017]80.4969.9911.713.46
CPS[6][CVPR'2021]82.7072.3012.993.46
MC-Net[28][MICCAI'2021]82.5172.9814.573.78
MCF[27][CVPR'2023]81.9671.7613.963.99
AC-MT[31][MIA'2023]25(10%)225(90%)82.0971.6515.914.81
UG-MCL[36][AHIM'2023]82.8572.6911.122.28
EVIL[7][CBM'2024]82.6072.4611.442.76
AAU[1][MIA'2024]82.4372.3313.834.18
MLRP[24][MIA'2024]83.9674.0015.164.21
ours85.14↑1.1875.757.792.00
+ +Table 4. Comparisons with state-of-the-art semi-supervised segmentation methods on the Brats dataset. + +
DatasetsCG-CPCLDTCTDice↑Jaccard↑95HD↓ASD↓
ACDC✓86.6177.147.302.15
89.3281.313.451.02
✓89.1980.973.260.94
✓✓90.5683.212.570.78
AbdomenCT-1K✓79.1969.2129.777.56
83.9875.919.592.81
✓83.7975.2521.395.28
✓✓84.7576.3521.155.59
Brats✓83.2073.7311.913.01
84.6274.818.922.36
✓84.8575.229.962.61
✓✓85.1475.757.792.00
+ +method, especially on the AbdomenCT-1K dataset, where the Dice index improved by $4.79\%$ . In the ACDC dataset, when cross pseudo supervision similar to CPS was applied (see Tab. 1 for experimental results), CG-CPCL increased by $0.86\%$ . The reliable unstable region usually has feature differences within the class, which is prone to misjudgment. Therefore, the class prototype is constructed, and cross-prototype consistency is used to enhance the model's discrimination. In addition, the confidence-based conditional guidance mechanism effectively reduces the performance difference between models. + +Similarly, with DTCT training (learning only the unreliable stable region), the performance of all datasets improved significantly, with the Brats dataset improving by $1.65\%$ . On the ACDC dataset, CG-CPCL improved by $0.99\%$ when cross pseudo supervision similar to CPS was used. The unreliable stable region has high uncertainty, and a single model is prone to bias. To this end, we improved the accuracy of the model in this area by integrating predictions from multiple models, competitively generating teacher predictions, and selecting the most credible predictions as pseudo-labels. + +When CG-CPCL and DTCT are combined, the model can achieve optimal segmentation performance by using differentiated learning strategies to optimize both reliable unstable and unreliable stable regions. Compared with the baseline model, the Dice indexes of multiple datasets improved by $3.95\%$ , $5.56\%$ , and $1.94\%$ , respectively. On the ACDC dataset, ALHVR increased the Dice indicator from + +Table 5. Ablation results of different components of ALHVR. We conducted experiments to test the impact of CG-CPCL and DTCT on model performance. + +
β0.50.60.70.80.9
ACDC89.8889.9490.2690.5689.85
AbdomenCT-1K84.1684.3984.7584.4983.71
Brats84.7185.0285.1484.8484.50
+ +Table 6. Ablation studies on the hyperparameter $\beta$ with ${10}\% ,5\%$ and ${10}\%$ labeled data. + +88.58% to 90.56%. Since these two types of regions have different distributions and prediction characteristics, a unified training strategy is difficult to account for both. Therefore, our method optimizes the two types of regions with targeted strategies, effectively improving the overall performance of the model and its generalization ability under complex scenarios. + +# 4.4.2. Hyperparameter Value Analysis + +We investigate the impact of confidence thresholds calculated based on different pixel proportions on model training. As shown in Tab. 6, the performance of dynamic thresholds on multiple datasets is evaluated using Dice coefficients. The results indicate that the optimal proportions are 0.8 for the ACDC dataset and 0.7 for the AbdomenCT-1K and Brats datasets. When the proportion is set lower, the threshold decreases, causing the model to focus more on high-confidence pixels, which may reduce its ability to handle difficult samples and degrade the quality of pseudolabels. Conversely, when the threshold increases, the model may overlook subtle features, affecting detail prediction. Therefore, it is necessary to balance the pixel proportion to optimize the model's attention on the reliable unstable region and the unreliable stable region, thereby improving performance and robustness. + +# 5. Conclusion + +In this paper, we propose a novel semi-supervised learning framework, ALHVR, which explores high-value learning regions and designs adaptive training strategies. Specifically, CG-CPCL is used to train the reliable unstable region, while DTCT is employed for the unreliable stable region. To enhance stability in the reliable unstable region, CG-CPCL enables low-confidence predictions to learn from high-confidence predictions through confidence-guided cross-pseudo supervision and prototype consistency constraints, improving intra-class consistency and inter-class discriminability. To enhance reliability in the unreliable stable region, DTCT dynamically selects more reliable pixels as teachers using a dynamic teacher selection mechanism and competitive teaching strategy, reducing noise interference while strengthening the model's learning capability. Experimental results on three public datasets demonstrate that ALHVR achieves better performance than existing state-of-the-art methods. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (Grant Nos. 62271296, 62201334), the Shaanxi Province Innovation Capacity Support Program (Grant No. 2025RS-CXTD-012), and the Scientific Research Program funded by the Shaanxi Provincial Education Commission (Grant Nos. 23JP022, 23JP014). + +# References + +[1] Sukesh Adiga, Jose Dolz, and Herve Lombaert. Anatomically-aware uncertainty for semi-supervised image segmentation. Medical Image Analysis, 91:103011, 2024. 6, 7, 8 +[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11514-11524, 2023. 2 +[3] Spyridon (Spyros) Bakas. Brats miccai brain tumor dataset, 2020. https://dx.doi.org/10.21227/hdtd-5j88.6 +[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6 +[5] Faquan Chen, Jingjing Fei, Yaqi Chen, and Chenxi Huang. Decoupled consistency for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 551-561. Springer, 2023. 3 +[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 2, 6, 7, 8 +[7] Yingyu Chen, Ziyuan Yang, Chenyu Shen, Zhiwen Wang, Zhongzhou Zhang, Yang Qin, Xin Wei, Jingfeng Lu, Yan Liu, and Yi Zhang. Evidence-based uncertainty-aware semi-supervised medical image segmentation. Computers in Biology and Medicine, 170:108004, 2024. 6, 7, 8 +[8] Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4070-4080, 2024. 3 +[9] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, 130:108777, 2022. 3 +[10] Hejun Huang, Zuguo Chen, Chaoyang Chen, Ming Lu, and Ying Zou. Complementary consistency semi-supervised learning for 3d left atrial image segmentation. Computers in Biology and Medicine, 165:107368, 2023. 2 + +[11] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3 +[12] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 4 +[13] Tao Lei, Yi Wang, Xingwu Wang, Xuan Wang, Bin Hu, and Asoke K Nandi. Unified feature consistency of underperforming pixels and valid regions for semi-supervised medical image segmentation. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 3 +[14] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 3 +[15] Wenjing Lu, Jiahao Lei, Peng Qiu, Rui Sheng, Jinhua Zhou, Xinwu Lu, and Yang Yang. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 662-672. Springer, 2023. 5 +[16] Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, and Shaoting Zhang. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 318-329. Springer, 2021. 3 +[17] Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In International conference on medical imaging with deep learning, pages 820-833. PMLR, 2022. 6 +[18] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695-6714, 2021. 6 +[19] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 7 +[20] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369-1378, 2021. 2 +[21] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12674-12684, 2020. 3 + +[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6 +[23] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3 +[24] Jiawei Su, Zhiming Luo, Sheng Lian, Dazhen Lin, and Shaozi Li. Mutual learning with reliable pseudo label for semi-supervised medical image segmentation. Medical Image Analysis, 94:103111, 2024. 1, 3, 6, 7, 8 +[25] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 6, 7, 8 +[26] Bing Wang, Taifeng Huang, Shuo Yang, Ying Yang, Junhai Zhai, and Xin Zhang. Dual-decoder data decoupling training for semi-supervised medical image segmentation. Biomedical Signal Processing and Control, 100:106984, 2025. 3 +[27] Yongchao Wang, Bin Xiao, Xiuli Bi, Weisheng Li, and Xinbo Gao. Mcf: Mutual correction framework for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15651-15660, 2023. 1, 2, 6, 7, 8 +[28] Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, and Lei Zhang. Semi-supervised left atrium segmentation with mutual consistency training. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part II 24, pages 297-306. Springer, 2021. 6, 7, 8 +[29] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 6 +[30] Zhe Xu, Yixin Wang, Donghuan Lu, Lequan Yu, Jiangpeng Yan, Jie Luo, Kai Ma, Yefeng Zheng, and Raymond Kaiyu Tong. All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 26(7):3174-3184, 2022. 6 +[31] Zhe Xu, Yixin Wang, Donghuan Lu, Xiangde Luo, Jiangpeng Yan, Yefeng Zheng, and Raymond Kai-yu Tong. Ambiguity-selective consistency regularization for meantacher semi-supervised medical image segmentation. Medical Image Analysis, 88:102880, 2023. 6, 7, 8 +[32] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4268-4277, 2022. 3 +[33] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in + +semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7236-7246, 2023. 3 +[34] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical image computing and computer assisted intervention-MICCAI 2019: 22nd international conference, Shenzhen, China, October 13–17, 2019, proceedings, part II 22, pages 605–613. Springer, 2019. 1, 3 +[35] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2 +[36] Yichi Zhang, Rushi Jiao, Qingcheng Liao, Dongyang Li, and Jicong Zhang. Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation. Artificial Intelligence in Medicine, 138:102476, 2023. 6, 7, 8 +[37] Xiangyu Zhao, Zengxin Qi, Sheng Wang, Qian Wang, Xuehai Wu, Ying Mao, and Lichi Zhang. Rcps: Rectified contrastive pseudo supervision for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 28(1):251-261, 2023. 3 +[38] Zhongda Zhao, Haiyan Wang, Tao Lei, Xuan Wang, Xiaohong Shen, and Haiyang Yao. Balanced feature fusion collaborative training for semi-supervised medical image segmentation. Pattern Recognition, 157:110856, 2025. 6 \ No newline at end of file diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/images.zip b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9d3c65ae60af492d80a14d81e512794b0d643bb4 --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c670afd8ef525a996d51b98aad068d16323189978e9f3cc15871979e50f39ca +size 565211 diff --git a/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/layout.json b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ed2a1df5d9f8e50c5c9715f9ce501a2e13d0d66 --- /dev/null +++ b/2025/Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation/layout.json @@ -0,0 +1,9209 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "Adaptive Learning of High-Value Regions for Semi-Supervised Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Tao Lei" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Ziyao Yang" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Xingwu Wang" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Yi Wang" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Xuan Wang" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Feiman Sun" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": ", Asoke K. Nandi" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Shaanxi University of Science and Technology, China \n" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications, China \n" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Northwestern Polytechnical University, China \n" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Southwest Jiaotong University, China " + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 69, + 160, + 541, + 232 + ], + "type": "text", + "content": "Brunel University of London, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 233, + 412, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 233, + 412, + 245 + ], + "spans": [ + { + "bbox": [ + 195, + 233, + 412, + 245 + ], + "type": "text", + "content": "*Corresponding author:leitao@sust.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 296, + 647 + ], + "type": "text", + "content": "Existing semi-supervised learning methods typically mitigate the impact of unreliable predictions by suppressing low-confidence regions. However, these methods fail to explore which regions hold higher learning value and how to design adaptive learning strategies for these regions. To address these issues, we propose a novel adaptive learning of high-value regions (ALHVR) framework. By exploiting the diversity of predictions from dual-branch networks, the prediction regions are classified into three groups: reliable stable region, reliable unstable region, and unreliable stable region. For high-value regions (reliable unstable region and unreliable stable region), different training strategies are designed. Specifically, for reliable unstable region, we propose a confidence-guided cross-prototype consistency learning (CG-CPCL) module, which enforces prototype consistency constraints in the feature space. By leveraging confidence information, the high-confidence predictions from one network selectively supervise the low-confidence predictions from the other, thus helping the model learn inter-class discrimination more stably. Additionally, for unreliable stable region, we design a dynamic teacher competition teaching (DTCT) module, which dynamically selects the most reliable pixels as teachers by evaluating the unperturbed predictions from both networks. These selected pixels are then used to supervise perturbed predictions, thereby enhancing the model's learning capability in unreliable region. Experimental results show that our method outperforms state-of-the-art approaches on three public datasets. Code is available at https://github.com/ziziyao/ALHVR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Medical images contain various types of tissues with complex anatomical structures, leading to ambiguous bound" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 279, + 537, + 346 + ], + "blocks": [ + { + "bbox": [ + 333, + 279, + 537, + 346 + ], + "lines": [ + { + "bbox": [ + 333, + 279, + 537, + 346 + ], + "spans": [ + { + "bbox": [ + 333, + 279, + 537, + 346 + ], + "type": "image", + "image_path": "e75b38b68f71efdd50fad78b6529c4231dd794f41870f3c05bf3803bf93c3b36.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "type": "text", + "content": "Figure 1. Training on the Brats dataset using the CPS framework, where the predictions of a given sample are classified based on a confidence threshold. Region 1 represents the reliable stable region, Region 2 represents the reliable unstable region, and Region 3 represents the unreliable stable region." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 365, + 408, + 506, + 479 + ], + "blocks": [ + { + "bbox": [ + 365, + 408, + 506, + 479 + ], + "lines": [ + { + "bbox": [ + 365, + 408, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 365, + 408, + 506, + 479 + ], + "type": "table", + "html": "
DatasetRegion 1Region 2Region 3Dice
ACDC✓✓✓85.34
✓87.13
✓88.46
✓88.20
✓✓87.41
✓✓82.51
✓✓88.58
", + "image_path": "ddf6369b6081855551deb4b8fb328102d1a4efe38f924a9d537656a351ff9a2e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 517 + ], + "type": "text", + "content": "Table 1. Experimental results of training on the ACDC dataset with " + }, + { + "bbox": [ + 313, + 483, + 555, + 517 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 483, + 555, + 517 + ], + "type": "text", + "content": " labeled data using the CPS framework under different masked region settings." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 715 + ], + "type": "text", + "content": "aries between target regions and surrounding tissues. This issue increases the segmentation difficulty for the model at boundary regions. Some semi-supervised methods [23, 34] estimate confidence based on predictions from a single network to distinguish between high-confidence and low-confidence pixels. These methods guide the model to focus on high-confidence predictions by discarding low-confidence pixels. However, predictions from a single network are biased, leading to inaccurate pixel classification. Therefore, some approaches evaluate pixel reliability based on dual-network predictions. For instance, MCF [27] treats inconsistent predictions between networks as errors and attempts to correct them, but it fails to handle cases where both predictions are incorrect. MLRP [24] employs a pixelwise comparison strategy to select relatively reliable predic" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21450" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "tions from two networks as pseudo-labels. However, mutual learning between low-confidence predictions may introduce biases, leading to suboptimal learning. These methods fail to identify or utilize high-value regions, particularly those with high segmentation difficulty. Furthermore, they lack adaptive learning strategies tailored to distinct regions, limiting the model's ability to learn in crucial segmentation regions, such as tissue boundaries." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 170, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 296, + 445 + ], + "type": "text", + "content": "In order to resolve the aforementioned issues, this paper combines the diversity of dual-network predictions with confidence thresholds to achieve more precise pixel classification. Taking CPS-based methods [6] as an example, we analyze this problem. When two networks predict the same image, the confidence of each pixel relative to a predefined threshold can be categorized into three types: (1) both networks have confidence scores above the threshold, (2) one network has a confidence score above the threshold while the other is below it, and (3) both networks have confidence scores below the threshold. Based on confidence levels and prediction consistency, we define these three regions as the reliable stable region, the reliable unstable region, and the unreliable stable region, respectively. As shown in Fig. 1, different regions are represented by different colors. To further explore the most valuable learning regions, we conducted experiments on the ACDC dataset using the CPS framework. Specifically, we applied masks to occlude different regions, training the model only on unmasked areas. As shown in Tab. 1, the highest Dice score is achieved when pixels in the reliable stable region are masked out during training process. This indicates that the reliable unstable region and the unreliable stable region are high-value regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "content": "To address the lack of adaptive learning strategies for regions of varying value in existing methods, this paper proposes a semi-supervised medical image segmentation method based on adaptive learning for high-value regions (ALHVR). By analyzing the confidence distributions of model predictions, high-value regions are identified, and region-specific training strategies are designed to improve segmentation performance. Specifically, we propose the confidence-guided cross-prototype consistency learning (CG-CPCL) module and the dynamic teacher competition teaching (DTCT) module, which are designed to train the reliable unstable region and the unreliable stable region, respectively. The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 616, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 616, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 687 + ], + "type": "text", + "content": "(1) We conduct an investigation into the high-value regions in semi-supervised medical image segmentation, categorizing them into reliable unstable region and unreliable stable region. Furthermore, we design adaptive learning strategies to enhance segmentation performance in these areas." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "(2) We propose CG-CPCL to address the challenge where unstable predictions within a region pre" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "vent the model from accurately recognizing key region features. CG-CPCL enhances cross-consistency between features and prototypes by constructing class prototypes while incorporating confidence constraints to ensure that low-confidence predictions learn exclusively from high-confidence predictions. This approach effectively guides the model to learn more reliable feature representations in reliable unstable region, thereby improving its ability to identify key regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 181, + 553, + 350 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 553, + 287 + ], + "type": "text", + "content": "(3) We design DTCT to tackle the issue where unreliable predictions within a region lead to the accumulation of erroneous signals, causing prediction bias. DTCT employs pixel-level reliability assessment to dynamically and competitively select more reliable pixels from unperturbed predictions as teacher predictions to supervise perturbed predictions. This method provides the model with more valuable and reliable learning signals, effectively reducing prediction bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 350 + ], + "type": "text", + "content": "(4) The proposed framework integrates both learning modules and is evaluated on three public datasets: ACDC, AbdomenCT-1K, and Brats. Experimental results demonstrate that our framework outperforms current state-of-the-art semi-supervised learning methods." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 383, + 553, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 553, + 396 + ], + "type": "text", + "content": "2.1. Semi-Supervised Medical Image Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 401, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 472 + ], + "type": "text", + "content": "The key challenge in semi-supervised image segmentation is to effectively exploit unlabeled data to improve model performance. To address this, researchers have proposed various approaches, among which consistency regularization and pseudo-labeling methods have become mainstream directions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "content": "Firstly, consistency regularization methods are based on the smoothness assumption, which states that a model's predictions should remain similar under different levels of perturbations, including data-level, model-level, and feature-level perturbations. At the data level, common perturbations involve transformations such as rotation and color jittering. To generate more diverse samples, copy-paste techniques are popular, such as CutMix [35] and ClassMix [20]. However, existing copy-paste methods overlook the distribution mismatch between labeled and unlabeled data. To tackle this issue, BCP [2] introduces bidirectional copy-paste between labeled and unlabeled data to generate new samples. At the model level, perturbations are typically introduced by modifying model parameters or structures. For example, MT [25] employs an exponential moving average (EMA) to update the teacher model's parameters with those of the student model, CPS [6] initializes two networks with different parameters and trains them independently. Additionally, MCF [27] and CC-Net [10] introduce variations in encoder or decoder structures, ensuring that different net" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21451" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "work architectures produce consistent predictions. At the feature level, CCT [21] injects noise or drops activations to enforce consistency between decoders under perturbations. PS-MT [14] uses T-VAT from the teacher model for more accurate student perturbations. Since these consistency regularization methods rely on individual perturbations, ABD [8] dynamically adjusts image displacement, improving the model's adaptability to complex data variations and robustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 181, + 298, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 181, + 298, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 181, + 298, + 397 + ], + "type": "text", + "content": "Secondly, the pseudo-labeling methods are based on the clustering assumption. It expands the dataset by generating pseudo-labels for unlabeled data using a pre-trained model. To improve the quality of pseudo-labels, " + }, + { + "bbox": [ + 56, + 181, + 298, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{ST} + +" + }, + { + "bbox": [ + 56, + 181, + 298, + 397 + ], + "type": "text", + "content": " [32] prioritizes reliable unlabeled images based on prediction stability and performs selective retraining. DMT [9] overcomes the limitation of self-training by identifying incorrect pseudo-labels through prediction differences between multiple models. Additionally, some methods combine pseudolabeling with consistency regularization. For example, FixMatch [23] and RCPS [37] encourage the model to produce consistent predictions between strongly augmented data and pseudo-labels generated from weakly augmented data. While these methods enhance model performance through data diversification or improved pseudo-labels' quality, they overlook the varying complexity of predictions across different image regions, and fail to learn more discriminative features from high-value regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "content": "2.2. Uncertainty-Guided Semi-Supervised Medical Image Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "In semi-supervised medical image segmentation, uncertainty analysis is a key strategy for effectively exploiting pseudo-labels. Existing methods typically estimate uncertainty using Monte Carlo Dropout [11]. Additionally, some approaches employ confidence scores or entropy measures for uncertainty guidance. Specifically, to enable the model to learn more reliable targets, UA-MT [34] utilizes Monte Carlo Dropout for uncertainty estimation and trains the model only with low-uncertainty pixels. URPC [16] refines uncertainty by computing the KL divergence of predictions, allowing the model to progressively learn from multi-scale reliable predictions. To further improve pseudo-label quality, UniMatch [33] evaluates the confidence of model predictions using the maximum class probability and retains only high-confidence pixels from weakly augmented data as pseudo-labels. MLRP [24] assesses the reliability of different subnet predictions via entropy and uses more reliable predictions as pseudo-labels to supervise the other subnetwork's training. To fully exploit pixel information with varying confidence levels, UFC [13] distinguishes high-performing and under-performing pixels based on confidence differences between the student and teacher models. Both DC-Net [5] and DD-Net [26] decouple predictions us" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "ing confidence thresholds and optimize different pixels with distinct loss functions, effectively improving model performance. However, although these methods assess and train pixel reliability using various strategies, they rely solely on simple consistency losses (e.g. cross-entropy loss and mean squared error loss) and fail to design adaptive training strategies based on the characteristics of the data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 166, + 370, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 166, + 370, + 178 + ], + "spans": [ + { + "bbox": [ + 314, + 166, + 370, + 178 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 186, + 470, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 470, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 470, + 197 + ], + "type": "text", + "content": "3.1. Overview of the Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": "In semi-supervised image segmentation tasks, the training dataset typically consists of a small amount of labeled data and a large amount of unlabeled data. Mathematically, the training dataset is defined as " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "D = \\{D_L, D_U\\}" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "D_L = \\{(x_i^L, y_i^L)\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " represents the labeled dataset containing " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " samples, and " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "D_U = \\{x_i^U\\}_{i=1}^M" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " represents the unlabeled dataset containing " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " samples, with " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "N \\ll M" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ". We define the input image as " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathbb{R}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " denote the height, width, and depth of the image, respectively. The ground truth corresponding to the labeled data is defined as " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "y_i \\in \\{0, 1, \\dots, C-1\\}^{H \\times W \\times Z}" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 203, + 555, + 346 + ], + "type": "text", + "content": " represents the number of classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "content": "The training process of ALHVR is shown in Fig. 2, which consists of two parallel networks " + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "inline_equation", + "content": "f_{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "inline_equation", + "content": "f_{B}(\\cdot)" + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "content": " with the same structure but independently updating parameters. Each network's decoder takes the encoded feature " + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "content": " and its perturbed version " + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 313, + 346, + 554, + 441 + ], + "type": "text", + "content": " as inputs to generate predictions from different perspectives. It is worth noting that feature perturbations are generated by randomly adding noise or dropout." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "text", + "content": "The original and perturbed features from both networks are fed into their corresponding decoders, " + }, + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "inline_equation", + "content": "f_{A}^{d}(\\cdot)" + }, + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "inline_equation", + "content": "f_{B}^{d}(\\cdot)" + }, + { + "bbox": [ + 313, + 442, + 554, + 478 + ], + "type": "text", + "content": ", to generate the final prediction outputs:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 367, + 480, + 555, + 495 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {a 1} = f _ {A} ^ {d} \\left(F _ {a 1}\\right), \\hat {y} _ {a 2} = f _ {A} ^ {d} \\left(F _ {a 2}\\right), \\tag {1}", + "image_path": "34ac535b8fdc74208414d13764302e890ca7c13b528063b382582e3cca834e41.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 499, + 554, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 499, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 369, + 499, + 554, + 514 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {b 1} = f _ {B} ^ {d} \\left(F _ {b 1}\\right), \\hat {y} _ {b 2} = f _ {B} ^ {d} \\left(F _ {b 2}\\right), \\tag {2}", + "image_path": "4ae7fb7fc749b215d293e2b916fc4bf218dacb715764ea8c93c7a57f63a57a95.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "type": "text", + "content": "where the decoder outputs include the prediction results for both labeled and unlabeled data, " + }, + { + "bbox": [ + 313, + 518, + 554, + 541 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\hat{y}^L\\cup \\hat{y}^U" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 566 + ], + "type": "text", + "content": "For labeled data, a combination of cross-entropy loss and dice loss is used as the supervised learning loss function:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 568, + 555, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 568, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 364, + 568, + 555, + 582 + ], + "type": "interline_equation", + "content": "L _ {s} ^ {a} = L _ {c e} \\left(\\hat {y} _ {a 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {a 1}, y\\right), \\tag {3}", + "image_path": "dea11f9b61193256e2945b163ab451215624ba563e653c988ef1f460e47fe1f6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 588, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 588, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 365, + 588, + 553, + 602 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {s}} ^ {b} = L _ {c e} \\left(\\hat {y} _ {b 1}, y\\right) + L _ {d i c e} \\left(\\hat {y} _ {b 1}, y\\right). \\tag {4}", + "image_path": "2031d2a7c6cddac205bf2638ef259ee20107f335a5f500acfd1365c864012886.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": "For unlabeled data, an adaptive training strategy is proposed based on high-value regions. Specifically, for reliable unstable region, the CG-CPCL module is designed to introduce a cross-prototype consistency loss. Additionally, to further enhance prototype consistency and reduce intraclass variance, the cosine similarity loss between prototypes and features is incorporated as a soft loss term to optimize intra-class feature clustering. The specific definition is as follows:" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21452" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 77, + 495, + 241 + ], + "blocks": [ + { + "bbox": [ + 114, + 77, + 495, + 241 + ], + "lines": [ + { + "bbox": [ + 114, + 77, + 495, + 241 + ], + "spans": [ + { + "bbox": [ + 114, + 77, + 495, + 241 + ], + "type": "image", + "image_path": "584537d36a8a929a644cff7bd016a2712a23a439ec0ac716968a93f621f67b7f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 247, + 187, + 262 + ], + "lines": [ + { + "bbox": [ + 118, + 247, + 187, + 262 + ], + "spans": [ + { + "bbox": [ + 118, + 247, + 187, + 262 + ], + "type": "text", + "content": "random perturbation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 197, + 250, + 254, + 262 + ], + "lines": [ + { + "bbox": [ + 197, + 250, + 254, + 262 + ], + "spans": [ + { + "bbox": [ + 197, + 250, + 254, + 262 + ], + "type": "text", + "content": "C cosine similarity" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 271, + 243, + 492, + 339 + ], + "blocks": [ + { + "bbox": [ + 119, + 266, + 267, + 281 + ], + "lines": [ + { + "bbox": [ + 119, + 266, + 267, + 281 + ], + "spans": [ + { + "bbox": [ + 119, + 266, + 267, + 281 + ], + "type": "inline_equation", + "content": "\\text{M}" + }, + { + "bbox": [ + 119, + 266, + 267, + 281 + ], + "type": "text", + "content": " multiply pixel by pixel protetype generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 141, + 291, + 233, + 301 + ], + "lines": [ + { + "bbox": [ + 141, + 291, + 233, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 233, + 301 + ], + "type": "text", + "content": "1 region1: reliable stable region" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 141, + 308, + 239, + 318 + ], + "lines": [ + { + "bbox": [ + 141, + 308, + 239, + 318 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 239, + 318 + ], + "type": "text", + "content": "2 region2: reliable unstable region" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 141, + 325, + 239, + 336 + ], + "lines": [ + { + "bbox": [ + 141, + 325, + 239, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 239, + 336 + ], + "type": "text", + "content": "3 region3:unreliable stable region" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 271, + 243, + 492, + 339 + ], + "lines": [ + { + "bbox": [ + 271, + 243, + 492, + 339 + ], + "spans": [ + { + "bbox": [ + 271, + 243, + 492, + 339 + ], + "type": "image", + "image_path": "58f9d35eee1e1d2f88f4e3a52c74fbc666c823b964642a832cecee5bcef1a2b6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 384 + ], + "type": "text", + "content": "Figure 2. The process of our proposed ALHVR framework. This method first classifies the perturbation predictions of the network output into value regions, then applies an adaptive training strategy to high-value regions, where CG-CPCL is used to train Region 2 and DTCT is used to train Region 3. The supervised learning process and the pseudo-label conversion process are omitted in the figure." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 399, + 295, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 399, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 83, + 399, + 295, + 413 + ], + "type": "interline_equation", + "content": "L _ {p c} ^ {a} = L _ {c e} ^ {a} \\left(\\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\hat {y} _ {b 1} ^ {p}\\right) + L _ {\\cos} ^ {a} \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\tag {5}", + "image_path": "14074b672c7e713d68c9cb009829f2554f1ef2af13d1ad421133fb7846913f75.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 83, + 418, + 295, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 418, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 83, + 418, + 295, + 434 + ], + "type": "interline_equation", + "content": "L _ {p c} ^ {b} = L _ {c e} ^ {b} \\left(\\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\hat {y} _ {a 1} ^ {p}\\right) + L _ {\\cos} ^ {b} \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {6}", + "image_path": "fe9fafac04767766bf6ede508d23e05d9e93d306b062d6d0e690ba61d6304ae7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "F_{a2}^{d}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "F_{b2}^{d}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " are perturbed features from the penultimate layer of the decoder, " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "p_{a1}^{c}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "p_{b1}^{c}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " denotes the c-th class prototype computed from the original features of the decoder, " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\hat{y}_{a1}^{p}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\hat{y}_{b1}^{p}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " denote the pseudo-labels generated from " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\hat{y}_{a1}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\hat{y}_{b1}" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "L_{ce}(\\cdot)" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " represents the cross-entropy loss, and " + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "inline_equation", + "content": "L_{\\cos}(\\cdot)" + }, + { + "bbox": [ + 55, + 437, + 295, + 509 + ], + "type": "text", + "content": " denotes the cosine similarity loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 510, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 592 + ], + "type": "text", + "content": "For unreliable stable region, the mean squared error loss is introduced through DTCT. To further constrain the model's probability distribution and reduce uncertainty in these unreliable regions, we incorporate the entropy of the model's predictions as a regularization term, thereby enhancing the model's robustness in such regions. This is formally defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 594, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 594, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 61, + 594, + 295, + 607 + ], + "type": "interline_equation", + "content": "L _ {m h} ^ {a} = L _ {m s e} ^ {a} \\left(\\hat {y} _ {a 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {a 1} \\left(\\hat {y} _ {a 1}\\right) + H _ {a 2} \\left(\\hat {y} _ {a 2}\\right), \\tag {7}", + "image_path": "501e7d344d60144da60483e0d6c5eef8128382f3778a034eaae58671f08ba21b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 59, + 610, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 610, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 59, + 610, + 295, + 625 + ], + "type": "interline_equation", + "content": "L _ {m h} ^ {b} = L _ {m s e} ^ {b} \\left(\\hat {y} _ {b 2}, \\left(\\hat {y} _ {a 1}, \\hat {y} _ {b 1}\\right)\\right) + H _ {b 1} \\left(\\hat {y} _ {b 1}\\right) + H _ {b 2} \\left(\\hat {y} _ {b 2}\\right), \\tag {8}", + "image_path": "6f85b0070edda3881b80d867e9ad95d736b53a450b2facb63e95d945399350ff.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "inline_equation", + "content": "L_{mse}(\\cdot)" + }, + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "text", + "content": " represents the mean squared error loss, and " + }, + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 56, + 628, + 295, + 651 + ], + "type": "text", + "content": " denotes the entropy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 652, + 214, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 214, + 663 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 214, + 663 + ], + "type": "text", + "content": "The unsupervised loss is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 99, + 663, + 294, + 678 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 663, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 99, + 663, + 294, + 678 + ], + "type": "interline_equation", + "content": "L _ {u} ^ {a} = L _ {p c} ^ {a} + L _ {m h} ^ {a}, L _ {u} ^ {b} = L _ {p c} ^ {b} + L _ {m h} ^ {b}. \\tag {9}", + "image_path": "de6a281d8076f23a7ce925c5f44e48dee1db32c99d3e5b5c4c665abe2dba7d21.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 683, + 242, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 242, + 694 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 242, + 694 + ], + "type": "text", + "content": "The total loss function for each network is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 104, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "L _ {a} = L _ {s} ^ {a} + \\lambda L _ {u} ^ {a}, L _ {b} = L _ {s} ^ {b} + \\lambda L _ {u} ^ {b}, \\tag {10}", + "image_path": "d85bc8667692a8d9aff76b823e91e5b850e11991aaff142a103579c8d79a681c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "type": "text", + "content": " is a time-dependent Gaussian warming up function [12], used to balance the supervised and unsupervised losses." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 441, + 518, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 518, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 518, + 453 + ], + "type": "text", + "content": "3.2. Valuable Region Classification Method" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "text", + "content": "Existing semi-supervised learning methods typically use fixed thresholds to separate high-confidence and low-confidence pixels. However, this binary approach provides a coarse distinction of pixel prediction difficulty, neglecting the potential information in low-confidence pixels. To address these limitations, an adaptive threshold is employed to refine pixel classification. Specifically, we directly derive the adaptive threshold from the confidence maps produced by two networks, ensuring better adaptation to current data variations. The confidence maps are generated using perturbed predictions: " + }, + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\text{con}_{a2} = \\max(\\hat{y}_{a2}), \\text{con}_{b2} = \\max(\\hat{y}_{b2})" + }, + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "text", + "content": ". By computing the pixel-wise average of the two confidence maps and sorting the values in ascending order, we obtain the sorted average confidence vector " + }, + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "inline_equation", + "content": "\\text{con}" + }, + { + "bbox": [ + 313, + 459, + 555, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 627, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 658 + ], + "type": "text", + "content": "Finally, the lowest " + }, + { + "bbox": [ + 313, + 627, + 554, + 658 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 627, + 554, + 658 + ], + "type": "text", + "content": " of the sorted confidence values are selected, and their mean is computed as the adaptive threshold:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 394, + 654, + 553, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 654, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 394, + 654, + 553, + 687 + ], + "type": "interline_equation", + "content": "\\gamma = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\operatorname {c o n} (i), \\tag {11}", + "image_path": "580ff066556bed216bfeec3fde925ed7c2cfd33b746d0c99a31cbffd543b1102.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "K = \\beta \\cdot (H\\times W\\times Z)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is a hyperparameter within the range " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " , controlling the proportion of" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21453" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 82, + 504, + 205 + ], + "blocks": [ + { + "bbox": [ + 105, + 82, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 504, + 205 + ], + "type": "image", + "image_path": "7e8b48bd560b512da09e8e191307169ae2ac82ae377f4c2fcdadab9e450e7cee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 555, + 245 + ], + "type": "text", + "content": "Figure 3. The structure of the DTCT module. By comparing the confidence of pixels in the Region 3 from the unperturbed predictions, the more reliable prediction is selected as the teacher. A probability sharpening function is then applied to adjust the probability distribution, generating more reliable pseudo-labels to supervise the perturbed predictions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 262, + 119, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 119, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 119, + 273 + ], + "type": "text", + "content": "selected pixels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "text", + "content": "Using the confidence threshold, we classify predicted pixels from the two networks interactively. By comparing the confidence values of corresponding pixels with the threshold, the pixels are categorized into three regions including reliable stable region " + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\Omega_{1}" + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "text", + "content": ", reliable unstable region " + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\Omega_{2}" + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "text", + "content": " and the unreliable stable region " + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\Omega_{3}" + }, + { + "bbox": [ + 55, + 274, + 296, + 357 + ], + "type": "text", + "content": " (as described in Chapter 1). These regions can be defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 366, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 366, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 102, + 366, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\Omega_ {1} = \\left\\{c o n _ {a 2} \\geq \\gamma \\& c o n _ {b 2} \\geq \\gamma \\right\\}, \\tag {12}", + "image_path": "c1b5de30994a2f999b8d9f923bb68328b63cf3eeb14638eead2ebfeeaf943d7c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 387, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 387, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 99, + 387, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\Omega_ {2} = \\left\\{\\left(c o n _ {a 2} > \\gamma \\& c o n _ {b 2} < \\gamma\\right) \\mid \\right. \\tag {13}", + "image_path": "db6da90f98684c56c3f89183cd95bfe71722a925541ff6223f9be5dbe90bdad1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 404, + 251, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 404, + 251, + 415 + ], + "spans": [ + { + "bbox": [ + 131, + 404, + 251, + 415 + ], + "type": "interline_equation", + "content": "\\left. \\left(c o n _ {a 2} < \\gamma \\& c o n _ {b 2} > \\gamma\\right) \\right\\},", + "image_path": "b9a514e16c32024f1f97587b017a4628bc370f99e8607711286ca2c35ae76955.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 420, + 295, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 420, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 102, + 420, + 295, + 433 + ], + "type": "interline_equation", + "content": "\\Omega_ {3} = \\left\\{c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma \\right\\}. \\tag {14}", + "image_path": "51a50a75a83a9abd6eddf48052b011e2326d3e1dd7cb083ac3ec39bd36aa83ee.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 534 + ], + "type": "text", + "content": "The reliable stable region is easily learnable by both networks, while the reliable unstable region reflects areas with unclear predictions, and the unreliable stable region represents difficult-to-learn areas. Thus, focusing on the reliable unstable region and unreliable stable region, rather than the reliable stable region, can better enhance the model's learning ability and provide higher learning value, as confirmed by the experimental results in Tab. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 540, + 296, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 567 + ], + "type": "text", + "content": "3.3. Confidence-Guided Cross-Prototype Consistency Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Due to complex semantic information in reliable unstable region and differences in model learning capabilities, different models produce divergent predictions for the same region. To address this, we propose the CG-CPCL module, which introduces prototypes as category center references, providing a global and stable reference point. Similar to UPCoL [15], we approximate the probability distribution of voxels by calculating cosine similarity between features and prototypes, introducing global constraints to prevent overreliance on local features. Additionally, we use confidence-based conditional guidance to ensure the model learns only from reliable targets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 356 + ], + "type": "text", + "content": "Specifically, we first generate class prototypes for the unlabeled data within the current mini-batch. We upsample the features from the second-to-last layer of the decoder using trilinear interpolation to match the size of the pseudolabels. Then, we use the pseudo-labels to select the corresponding class features, and generate class prototypes by calculating the mean of these features. This process is described as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 357, + 554, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 357, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 376, + 357, + 554, + 386 + ], + "type": "interline_equation", + "content": "p _ {a 1} ^ {c} = \\frac {\\sum F _ {a 1} ^ {d} \\cdot \\mathbb {I} (\\hat {y} _ {a 1} ^ {p} = c)}{\\sum (\\hat {y} _ {a 1} ^ {p} = c)}, \\tag {15}", + "image_path": "9cf75f611e0e9d9d7a9c14d20a504f1e09f196711549c7b6c81324980130783d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "inline_equation", + "content": "F_{a1}^{d}" + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "text", + "content": " represents the original features from the decoder, and " + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "text", + "content": " denotes the indicator function. For simplicity, the summation process omits the dimensional details of the features and pseudo-labels. Use the same method to calculate the " + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "inline_equation", + "content": "p_{b1}^{c}" + }, + { + "bbox": [ + 313, + 391, + 554, + 475 + ], + "type": "text", + "content": ". After obtaining the class prototypes, we approximate the model's prediction probability by calculating the cosine similarity between the features and the prototypes:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 481, + 553, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 481, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 331, + 481, + 553, + 496 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} _ {a} = \\cos \\left(F _ {a 2} ^ {d}, p _ {b 1} ^ {c}\\right), \\operatorname {s i m} _ {b} = \\cos \\left(F _ {b 2} ^ {d}, p _ {a 1} ^ {c}\\right), \\tag {16}", + "image_path": "52322b2f9c4b1b143292500a4ee3b96b241a65dca2f2324228e61b1e9723d719.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\cos(a, b) = \\frac{a \\cdot b}{\\|a\\|_2 \\cdot \\|b\\|_2}" + }, + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "text", + "content": ". To ensure that low-confidence predictions in the " + }, + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "inline_equation", + "content": "\\Omega_2" + }, + { + "bbox": [ + 313, + 502, + 554, + 586 + ], + "type": "text", + "content": " region only learn from reliable targets, we need to perform pixel selection on the confidence maps predicted by each network. This generates a position mask that contains the interactive information from both networks, which will be used as a guiding signal for the model's learning:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 358, + 590, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 590, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 358, + 590, + 553, + 602 + ], + "type": "interline_equation", + "content": "M a s k _ {a} = \\operatorname {c o n} _ {a 2} < \\gamma \\& \\operatorname {c o n} _ {b 2} > \\gamma , \\tag {17}", + "image_path": "b73d94c512d29376e3d75c7989b646129e2f46fbe0f4583718b0ebf24f20b2ff.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 358, + 609, + 553, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 609, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 358, + 609, + 553, + 621 + ], + "type": "interline_equation", + "content": "M a s k _ {b} = \\operatorname {c o n} _ {a 2} > \\gamma \\& \\operatorname {c o n} _ {b 2} < \\gamma . \\tag {18}", + "image_path": "f138c3bdd7d22fbb003a7319d7c426d281673a914c6e45cdb2e6fd729d46a9b8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 625, + 554, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 554, + 659 + ], + "type": "text", + "content": "Then, the position mask can be used to calculate the cross-prototype consistency loss, and the formula is as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 661, + 553, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 661, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 365, + 661, + 553, + 674 + ], + "type": "interline_equation", + "content": "L _ {c e} ^ {a} = L _ {c e} \\left(\\operatorname {s i m} _ {a}, \\hat {y} _ {b 1} ^ {p}\\right) \\odot \\operatorname {M a s k} _ {a}, \\tag {19}", + "image_path": "0e215642be57bcc12c2e47ce6bb546cc252edff1aeb7a4ab4bd5eda03af6607b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " represents the pixel-wise multiplication operation. Calculate " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "L_{ce}^{b}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " in the same way. Moreover, to further enhance the model's ability to learn class features, the cosine" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21454" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "similarity loss between the features and prototypes is also calculated:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 112, + 100, + 295, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 100, + 295, + 113 + ], + "spans": [ + { + "bbox": [ + 112, + 100, + 295, + 113 + ], + "type": "interline_equation", + "content": "L _ {\\cos} ^ {a} = (1 - \\operatorname {s i m} _ {a}) \\odot \\operatorname {M a s k} _ {a}. \\tag {20}", + "image_path": "9e55612170ac2ee6cba759f9f4dc476327e0411769f5ee06eed14ef0746633b0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "text", + "content": "We calculate " + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cos}}^b" + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "text", + "content": " similarly. Finally, the total loss for each model training the reliable unstable region is a linear combination of " + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "inline_equation", + "content": "L_{ce}" + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{cos}}" + }, + { + "bbox": [ + 55, + 121, + 296, + 159 + ], + "type": "text", + "content": ", as shown in Eqs. (5) and (6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 164, + 267, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 267, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 267, + 178 + ], + "type": "text", + "content": "3.4. Dynamic Teacher Competition Teaching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 253 + ], + "type": "text", + "content": "The unreliable stable region is an area where both models struggle to make accurate predictions and typically exhibits high uncertainty. The module improves the adaptability of the model to complex regions by reducing the influence of noise and uncertainty in the training process. The structure of the DTCT module is shown in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 254, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 409 + ], + "type": "text", + "content": "To generate the new teacher prediction, we perform pixel-wise confidence comparison based on the confidence maps of predictions without feature perturbation. We retain the pixels with higher confidence as the new prediction while discarding the pixels with lower confidence. Specifically, we first generate the confidence maps of the two predictions without feature perturbation: " + }, + { + "bbox": [ + 55, + 254, + 296, + 409 + ], + "type": "inline_equation", + "content": "\\text{con}_{a1} = \\max(\\hat{y}_{a1}), \\text{con}_{b1} = \\max(\\hat{y}_{b1})" + }, + { + "bbox": [ + 55, + 254, + 296, + 409 + ], + "type": "text", + "content": ". Then, we compare the generated confidence maps and select the prediction probabilities corresponding to the pixels with higher confidence as the new teacher prediction. Next, we apply a probability sharpening function [29] to convert this into a soft pseudolabel. This process can be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 412, + 295, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 412, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 295, + 439 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} = \\left\\{ \\begin{array}{l} \\hat {y} _ {a 1}, c o n _ {a 1} \\geq c o n _ {b 1} \\\\ \\hat {y} _ {b 1}, c o n _ {a 1} < c o n _ {b 1} \\end{array} , \\right. \\tag {21}", + "image_path": "1d3804ef99c2f7bd64ce4cfedf6308ee872589289cdf4f9070f63abab225ff18.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 443, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 443, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 119, + 443, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t} ^ {\\mathrm {s p}} = \\frac {\\hat {y} _ {t} ^ {1 / T}}{\\hat {y} _ {t} ^ {1 / T} + (1 - \\hat {y} _ {t}) ^ {1 / T}}, \\tag {22}", + "image_path": "3d8acfc72a4296f10e8765b753d757f5ebb38d694abfdc9ab5e2b00e1ea4ca33.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "text", + "content": " is a hyperparameter that controls the output distribution of the model. Since the DTCT module is only used to train the pixels in the unreliable stable region, a location mask needs to be generated first to determine the region location of " + }, + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "inline_equation", + "content": "\\Omega_3" + }, + { + "bbox": [ + 55, + 478, + 296, + 538 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 101, + 543, + 295, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 543, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 101, + 543, + 295, + 555 + ], + "type": "interline_equation", + "content": "M a s k = c o n _ {a 2} < \\gamma \\& c o n _ {b 2} < \\gamma . \\tag {23}", + "image_path": "57dbbea4b0a23899f9ad6358e0b7b37ccb24862fce043f5260f52c40a7225f10.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": "Then, the newly generated reliable pseudo-labels are used to supervise the perturbed version of the predictions, as shown in the following formula:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 603, + 295, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 603, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 295, + 617 + ], + "type": "interline_equation", + "content": "L _ {m s e} ^ {a} = L _ {m s e} \\left(\\hat {y} _ {a 2}, \\hat {y} _ {t} ^ {s p}\\right) \\odot M a s k. \\tag {24}", + "image_path": "238da6f6bf4f62ebfbad2cd82c6bf7cb3e10e7dfc2028b32be13640a3919f611.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": "We can calculate " + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "inline_equation", + "content": "L_{mse}^{b}" + }, + { + "bbox": [ + 55, + 624, + 296, + 696 + ], + "type": "text", + "content": " in the same way. To enhance the model's ability to perceive uncertainty in the unreliable stable region, guide the model to focus on regions with higher uncertainty, and effectively reduce the uncertainty in these regions, we quantify uncertainty using entropy and add it as a regularization term to the loss function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 63, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "H _ {a 1} = H \\left(\\hat {y} _ {a 1}\\right) \\odot \\text {M a s k}, H _ {a 2} = H \\left(\\hat {y} _ {a 2}\\right) \\odot \\text {M a s k}, \\tag {25}", + "image_path": "12da46ce277bdb3f84a08d1cc27ea419638ad5779eb61035d8e98e8dacd94875.jpg" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 552, + 247 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "type": "table", + "html": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net7(10%)081.5970.748.072.35
U-Net14(20%)084.8375.337.612.09
U-Net70(All)091.1084.044.901.14
MT[25][NIPS'2017]83.6573.1513.453.61
CPS[6][CVPR'2021]85.3475.508.782.37
MC-Net[28][MICCAI'2021]86.0776.5811.483.37
MCF[27][CVPR'2023]85.1875.2210.732.78
AC-MT[31][MIA'2023]86.3676.869.642.59
UG-MCL[36][AIIM'2023]7(10%)63(90%)86.8477.607.771.94
EVIL[7][CBM'2024]87.2578.185.311.35
AAU[1][MIA'2024]86.4676.998.002.06
MLRP[24][MIA'2024]87.1078.164.911.30
ours90.56[3,31]83.212.570.78
MT[25][NIPS'2017]86.5577.496.862.15
CPS[6][CVPR'2021]87.0378.136.662.12
MC-Net[28][MICCAI'2021]86.5877.6815.994.93
MCF[27][CVPR'2023]87.3278.436.532.00
AC-MT[31][MIA'2023]87.9979.347.852.14
UG-MCL[36][AIIM'2023]14(20%)56(80%)88.4079.869.122.44
EVIL[7][CBM'2024]88.3479.906.461.68
AAU[1][MIA'2024]87.7178.928.512.14
MLRP[24][MIA'2024]88.0079.545.411.58
ours91.09[2,69]84.052.710.73
", + "image_path": "a9afa876456c526020e31c2787d5e73156cc94f71161b219f06e76ed7491cf20.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 273 + ], + "type": "text", + "content": "Table 2. Comparisons with state-of-the-art semi-supervised segmentation methods on the ACDC dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "inline_equation", + "content": "H(\\hat{y}) = -\\hat{y}\\log \\hat{y}" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": ". We can use the same method to calculate the " + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "inline_equation", + "content": "H_{b1}" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "inline_equation", + "content": "H_{b2}" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": ". Finally, the total loss for training the unreliable stable region in each model is a linear combination of " + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "inline_equation", + "content": "L_{mse}" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 290, + 554, + 339 + ], + "type": "text", + "content": ", as shown in Eqs. (7) and (8)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 350, + 397, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 397, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 397, + 364 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 370, + 490, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 490, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 490, + 382 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 460 + ], + "type": "text", + "content": "The ACDC dataset [4] consists of cardiac dynamic MRI scans and corresponding ground truth from 100 patients, categorized into three classes: right ventricle, left ventricle, and myocardium. Following the experimental setup in CTCT [17], we used scans from 70 patients for training, 10 patients for validation, and 20 patients for testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 519 + ], + "type": "text", + "content": "The AbdomenCT-1K dataset [18] contains over 1,000 (1K) CT scans from 12 medical centers, categorized into four classes: liver, kidneys, spleen, and pancreas. Following the experimental setup in BFFC [38], we selected 30 scans for training and 11 scans for testing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "content": "The Brats dataset [3] consists of preoperative MRI scans from 335 glioma patients. Following the experimental setup in CPCL [30], we used 250 samples for training, 25 samples for validation, and 60 for testing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "content": "Metrics. We evaluate the method using four metrics categorized into region-sensitive and edge-sensitive types. Region-sensitive metrics include Dice and Jaccard, while edge-sensitive metrics include 95HD and ASD." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 447, + 636 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "All experiments were implemented using the PyTorch 1.7 on NVIDIA GeForce RTX 3090 GPU. The model is trained using the SGD optimizer with a weight decay factor of 0.0001, a momentum of 0.9 and an initial learning rate of 0.01. For segmentation on the ACDC and AbdomenCT-1K datasets, we use U-Net [22] as the backbone, with a" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21455" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 291, + 257 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 291, + 257 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 291, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 291, + 257 + ], + "type": "image", + "image_path": "57cbfd2b8d4f591c120d9119272c7e84e40bb9b6b87e916b58c911e66fa3dd6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "type": "text", + "content": "Figure 4. Visualization of some semi-supervised methods. The first two rows show the segmentation results on the ACDC dataset trained with " + }, + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "type": "text", + "content": " labeled data. The last two rows show the segmentation results on the AbdomenCT-1K dataset trained with " + }, + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 262, + 296, + 318 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": "batch size of 16 and " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "30\\mathrm{k}" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " training iterations. For Brats, we use V-Net [19] as the backbone, with a batch size of 4 and " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "15\\mathrm{k}" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " training iterations. In the total loss function of the network, the Gaussian warming up function is defined as " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\lambda (t) = 0.1\\times e^{-5(1 - t / t_{\\mathrm{max}})^2}" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathfrak{t}" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " denotes the current number of iterations and " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "t_\\mathrm{max}" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " represents the maximum number of iterations. In the probability sharpening function, the hyperparameter " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " is usually empirically set to 0.1. For calculating the confidence threshold, the hyperparameter " + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 335, + 295, + 479 + ], + "type": "text", + "content": " values were configured as 0.8, 0.7, and 0.7 in the ACDC, AbdomenCT-1K, and Brats datasets, respectively (set by Ablation Studies)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 490, + 281, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 281, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 281, + 502 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": "Firstly, we evaluate the proposed method on the ACDC dataset, comparing it with MT [25], CPS [6], MC-Net [28], MCF [27], AC-MT [31], UG-MCL [36], EVIL [7], AAU [1], and MLRP [24]. Additionally, we use U-Net trained with " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " of labeled data as the baseline and upper bound for reference. As shown in Tab. 2, our method achieves significant improvements across all four evaluation metrics compared to the baseline. Specifically, with " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " labeled data, the Dice score increases from " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "81.59\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "90.56\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " labeled data, it improves from " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "84.83\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "91.09\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": ", approaching the upper bound of " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "91.10\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " achieved with full supervision. Compared to state-of-the-art methods, ALHVR consistently outperforms others. With " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " labeled data, it surpasses the second-best method (EVIL) by " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " in Dice. When using " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " labeled data, it outperforms UG-MCL by " + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "inline_equation", + "content": "2.69\\%" + }, + { + "bbox": [ + 56, + 509, + 295, + 700 + ], + "type": "text", + "content": " in Dice." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "To further validate the generalization capability of the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 552, + 249 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 552, + 249 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 552, + 249 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 552, + 249 + ], + "type": "table", + "html": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
U-Net1(5%)067.3957.1446.8711.37
U-Net3(10%)075.6666.3322.446.89
U-Net30(All)086.9379.2711.552.45
MT[25][NIPS'2017]71.8261.1942.039.23
CPS[6][CVPR'2021]78.1468.3226.256.68
MC-Net[28][MICCAI'2021]80.9771.4332.987.48
MCF[27][CVPR'2023]80.7971.3026.137.81
AC-MT[31][MIA'2023]76.1865.9936.128.89
UG-MCL[36][AIIM'2023]1(5%)29(95%)74.0663.5826.937.06
EVIL[7][CBM'2024]80.6871.4723.726.33
AAU[1][MIA'2024]80.6571.0332.689.04
MLRP[24][MIA'2024]80.3070.6720.945.43
ours84.7576.3521.155.59
MT[25][NIPS'2017]78.7569.1025.356.88
CPS[6][CVPR'2021]82.7174.1117.304.90
MC-Net[28][MICCAI'2021]82.9474.1319.354.59
MCF[27][CVPR'2023]83.3474.6316.974.72
AC-MT[31][MIA'2023]82.3272.9726.527.54
UG-MCL[36][AIIM'2023]3(10%)27(90%)81.4372.5620.295.67
EVIL[7][CBM'2024]84.2475.7518.145.25
AAU[1][MIA'2024]81.4271.8229.698.25
MLRP[24][MIA'2024]84.5575.9913.824.05
ours85.6877.3521.966.22
", + "image_path": "ac74a443a1e566f936981a034bd384b47b26201fa77df67d330cb867ac751d25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "type": "text", + "content": "Table 3. Comparisons with state-of-the-art semi-supervised segmentation methods on the AbdomenCT-1K dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": "proposed method across different datasets, we evaluate its performance on the five-class AbdomenCT-1K dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 3, with " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " labeled data, ALHVR improves the Dice score by " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "17.36\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " compared to the baseline. When increasing the labeled data to " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": ", the Dice score improves by " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "10.02\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": ". Compared to state-of-the-art methods, ALHVR consistently achieves the best performance in segmenting four organs. Specifically, with " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " labeled data, it surpasses the second-best methods (MC-Net and MLRP) by " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "3.78\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "inline_equation", + "content": "1.13\\%" + }, + { + "bbox": [ + 313, + 293, + 555, + 460 + ], + "type": "text", + "content": " in Dice, respectively. Furthermore, as illustrated in Fig. 4, our method provides more precise segmentation of fine structures and boundary regions across different organs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": "In addition to organ segmentation tasks, we evaluated our method on the Brats dataset using the same comparison methods as in the ACDC experiment. As shown in Tab. 4, with only " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " labeled data, ALHVR improves the Dice score by " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "5.91\\%" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " compared to the baseline. Furthermore, it outperforms state-of-the-art methods, achieving a " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " higher Dice score than the second-best method (MLRP)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 553, + 416, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 416, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 416, + 565 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": "In this section, we conduct ablation studies to validate the effectiveness of each part and analyze the hyperparameters of the proposed method. The experiments were conducted on the ACDC (labeled ratio: " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": "), AbdomenCT-1K (labeled ratio: " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": "), and Brats (labeled ratio: " + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 571, + 554, + 631 + ], + "type": "text", + "content": ") datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 638, + 464, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 464, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 464, + 650 + ], + "type": "text", + "content": "4.4.1. Impact of Different Modules" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "In Tab. 5, we examine the effectiveness of the major components of ALHVR, CG-CPCL, and DTCT, on different datasets. When CG-CPCL training (learning only reliable unstable region) was used, the performance of all datasets was significantly improved compared with the baseline" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21456" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 292, + 169 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 292, + 169 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 292, + 169 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 292, + 169 + ], + "type": "table", + "html": "
MethodLabeledUnlabeledDice↑Jaccard↑95HD↓ASD↓
V-Net25(10%)079.2368.4015.834.34
V-Net250(All)086.4977.238.681.98
MT[25][NIPS'2017]80.4969.9911.713.46
CPS[6][CVPR'2021]82.7072.3012.993.46
MC-Net[28][MICCAI'2021]82.5172.9814.573.78
MCF[27][CVPR'2023]81.9671.7613.963.99
AC-MT[31][MIA'2023]25(10%)225(90%)82.0971.6515.914.81
UG-MCL[36][AHIM'2023]82.8572.6911.122.28
EVIL[7][CBM'2024]82.6072.4611.442.76
AAU[1][MIA'2024]82.4372.3313.834.18
MLRP[24][MIA'2024]83.9674.0015.164.21
ours85.14↑1.1875.757.792.00
", + "image_path": "51923632ee0abd4b0cf8932699df08d2b6f3efe5d86681473099d88904ffade4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 75, + 199, + 276, + 311 + ], + "blocks": [ + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 194 + ], + "type": "text", + "content": "Table 4. Comparisons with state-of-the-art semi-supervised segmentation methods on the Brats dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 199, + 276, + 311 + ], + "lines": [ + { + "bbox": [ + 75, + 199, + 276, + 311 + ], + "spans": [ + { + "bbox": [ + 75, + 199, + 276, + 311 + ], + "type": "table", + "html": "
DatasetsCG-CPCLDTCTDice↑Jaccard↑95HD↓ASD↓
ACDC✓86.6177.147.302.15
89.3281.313.451.02
✓89.1980.973.260.94
✓✓90.5683.212.570.78
AbdomenCT-1K✓79.1969.2129.777.56
83.9875.919.592.81
✓83.7975.2521.395.28
✓✓84.7576.3521.155.59
Brats✓83.2073.7311.913.01
84.6274.818.922.36
✓84.8575.229.962.61
✓✓85.1475.757.792.00
", + "image_path": "202ff8e39da6ba3303e7c4b165c1afef9122e4e0a45554914dc3726b418ae08e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "text", + "content": "method, especially on the AbdomenCT-1K dataset, where the Dice index improved by " + }, + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "4.79\\%" + }, + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "text", + "content": ". In the ACDC dataset, when cross pseudo supervision similar to CPS was applied (see Tab. 1 for experimental results), CG-CPCL increased by " + }, + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 56, + 366, + 295, + 496 + ], + "type": "text", + "content": ". The reliable unstable region usually has feature differences within the class, which is prone to misjudgment. Therefore, the class prototype is constructed, and cross-prototype consistency is used to enhance the model's discrimination. In addition, the confidence-based conditional guidance mechanism effectively reduces the performance difference between models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "text", + "content": "Similarly, with DTCT training (learning only the unreliable stable region), the performance of all datasets improved significantly, with the Brats dataset improving by " + }, + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "1.65\\%" + }, + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "text", + "content": ". On the ACDC dataset, CG-CPCL improved by " + }, + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "0.99\\%" + }, + { + "bbox": [ + 56, + 498, + 295, + 628 + ], + "type": "text", + "content": " when cross pseudo supervision similar to CPS was used. The unreliable stable region has high uncertainty, and a single model is prone to bias. To this end, we improved the accuracy of the model in this area by integrating predictions from multiple models, competitively generating teacher predictions, and selecting the most credible predictions as pseudo-labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "text", + "content": "When CG-CPCL and DTCT are combined, the model can achieve optimal segmentation performance by using differentiated learning strategies to optimize both reliable unstable and unreliable stable regions. Compared with the baseline model, the Dice indexes of multiple datasets improved by " + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "inline_equation", + "content": "5.56\\%" + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "inline_equation", + "content": "1.94\\%" + }, + { + "bbox": [ + 56, + 630, + 295, + 712 + ], + "type": "text", + "content": ", respectively. On the ACDC dataset, ALHVR increased the Dice indicator from" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 354, + 71, + 515, + 111 + ], + "blocks": [ + { + "bbox": [ + 56, + 315, + 294, + 348 + ], + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 348 + ], + "type": "text", + "content": "Table 5. Ablation results of different components of ALHVR. We conducted experiments to test the impact of CG-CPCL and DTCT on model performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 71, + 515, + 111 + ], + "lines": [ + { + "bbox": [ + 354, + 71, + 515, + 111 + ], + "spans": [ + { + "bbox": [ + 354, + 71, + 515, + 111 + ], + "type": "table", + "html": "
β0.50.60.70.80.9
ACDC89.8889.9490.2690.5689.85
AbdomenCT-1K84.1684.3984.7584.4983.71
Brats84.7185.0285.1484.8484.50
", + "image_path": "66057fde1b1c7200840f4b8661b9e5575849ccb6db8651719657ae64a61dfa6f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "lines": [ + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "text", + "content": "Table 6. Ablation studies on the hyperparameter " + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "inline_equation", + "content": "{10}\\% ,5\\%" + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "inline_equation", + "content": "{10}\\%" + }, + { + "bbox": [ + 315, + 115, + 553, + 136 + ], + "type": "text", + "content": " labeled data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 315, + 158, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 158, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 315, + 158, + 553, + 241 + ], + "type": "text", + "content": "88.58% to 90.56%. Since these two types of regions have different distributions and prediction characteristics, a unified training strategy is difficult to account for both. Therefore, our method optimizes the two types of regions with targeted strategies, effectively improving the overall performance of the model and its generalization ability under complex scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 253, + 478, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 478, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 478, + 266 + ], + "type": "text", + "content": "4.4.2. Hyperparameter Value Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 269, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 553, + 460 + ], + "type": "text", + "content": "We investigate the impact of confidence thresholds calculated based on different pixel proportions on model training. As shown in Tab. 6, the performance of dynamic thresholds on multiple datasets is evaluated using Dice coefficients. The results indicate that the optimal proportions are 0.8 for the ACDC dataset and 0.7 for the AbdomenCT-1K and Brats datasets. When the proportion is set lower, the threshold decreases, causing the model to focus more on high-confidence pixels, which may reduce its ability to handle difficult samples and degrade the quality of pseudolabels. Conversely, when the threshold increases, the model may overlook subtle features, affecting detail prediction. Therefore, it is necessary to balance the pixel proportion to optimize the model's attention on the reliable unstable region and the unreliable stable region, thereby improving performance and robustness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 476, + 386, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 386, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 386, + 488 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 498, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 498, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 498, + 553, + 712 + ], + "type": "text", + "content": "In this paper, we propose a novel semi-supervised learning framework, ALHVR, which explores high-value learning regions and designs adaptive training strategies. Specifically, CG-CPCL is used to train the reliable unstable region, while DTCT is employed for the unreliable stable region. To enhance stability in the reliable unstable region, CG-CPCL enables low-confidence predictions to learn from high-confidence predictions through confidence-guided cross-pseudo supervision and prototype consistency constraints, improving intra-class consistency and inter-class discriminability. To enhance reliability in the unreliable stable region, DTCT dynamically selects more reliable pixels as teachers using a dynamic teacher selection mechanism and competitive teaching strategy, reducing noise interference while strengthening the model's learning capability. Experimental results on three public datasets demonstrate that ALHVR achieves better performance than existing state-of-the-art methods." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21457" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62271296, 62201334), the Shaanxi Province Innovation Capacity Support Program (Grant No. 2025RS-CXTD-012), and the Scientific Research Program funded by the Shaanxi Provincial Education Commission (Grant Nos. 23JP022, 23JP014)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 188 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 196, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 196, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 196, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 296, + 239 + ], + "type": "text", + "content": "[1] Sukesh Adiga, Jose Dolz, and Herve Lombaert. Anatomically-aware uncertainty for semi-supervised image segmentation. Medical Image Analysis, 91:103011, 2024. 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 241, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 241, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 241, + 296, + 295 + ], + "type": "text", + "content": "[2] Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11514-11524, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 294, + 319 + ], + "type": "text", + "content": "[3] Spyridon (Spyros) Bakas. Brats miccai brain tumor dataset, 2020. https://dx.doi.org/10.21227/hdtd-5j88.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 321, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 296, + 396 + ], + "type": "text", + "content": "[4] Olivier Bernard, Alain Lalande, Clement Zotti, Frederick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging, 37 (11):2514-2525, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 453 + ], + "type": "text", + "content": "[5] Faquan Chen, Jingjing Fei, Yaqi Chen, and Chenxi Huang. Decoupled consistency for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 551-561. Springer, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 294, + 510 + ], + "type": "text", + "content": "[6] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2613-2622, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 566 + ], + "type": "text", + "content": "[7] Yingyu Chen, Ziyuan Yang, Chenyu Shen, Zhiwen Wang, Zhongzhou Zhang, Yang Qin, Xin Wei, Jingfeng Lu, Yan Liu, and Yi Zhang. Evidence-based uncertainty-aware semi-supervised medical image segmentation. Computers in Biology and Medicine, 170:108004, 2024. 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 622 + ], + "type": "text", + "content": "[8] Hanyang Chi, Jian Pang, Bingfeng Zhang, and Weifeng Liu. Adaptive bidirectional displacement for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4070-4080, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 667 + ], + "type": "text", + "content": "[9] Zhengyang Feng, Qianyu Zhou, Qiqi Gu, Xin Tan, Guangliang Cheng, Xuequan Lu, Jianping Shi, and Lizhuang Ma. Dmt: Dynamic mutual training for semi-supervised learning. Pattern Recognition, 130:108777, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[10] Hejun Huang, Zuguo Chen, Chaoyang Chen, Ming Lu, and Ying Zou. Complementary consistency semi-supervised learning for 3d left atrial image segmentation. Computers in Biology and Medicine, 165:107368, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[11] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 139 + ], + "type": "text", + "content": "[12] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 196 + ], + "type": "text", + "content": "[13] Tao Lei, Yi Wang, Xingwu Wang, Xuan Wang, Bin Hu, and Asoke K Nandi. Unified feature consistency of underperforming pixels and valid regions for semi-supervised medical image segmentation. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 262 + ], + "type": "text", + "content": "[14] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4258-4267, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 331 + ], + "type": "text", + "content": "[15] Wenjing Lu, Jiahao Lei, Peng Qiu, Rui Sheng, Jinhua Zhou, Xinwu Lu, and Yang Yang. Upcol: Uncertainty-informed prototype consistency learning for semi-supervised medical image segmentation. In International conference on medical image computing and computer-assisted intervention, pages 662-672. Springer, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 333, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 553, + 431 + ], + "type": "text", + "content": "[16] Xiangde Luo, Wenjun Liao, Jieneng Chen, Tao Song, Yinan Chen, Shichuan Zhang, Nianyong Chen, Guotai Wang, and Shaoting Zhang. Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency. In Medical Image Computing and Computer Assisted Intervention-MICCAI 2021: 24th International Conference, Strasbourg, France, September 27-October 1, 2021, Proceedings, Part II 24, pages 318-329. Springer, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[17] Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmentation via cross teaching between cnn and transformer. In International conference on medical imaging with deep learning, pages 820-833. PMLR, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[18] Jun Ma, Yao Zhang, Song Gu, Cheng Zhu, Cheng Ge, Yichi Zhang, Xingle An, Congcong Wang, Qiyuan Wang, Xin Liu, et al. Abdomenct-1k: Is abdominal organ segmentation a solved problem? IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6695-6714, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "text", + "content": "[19] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "text", + "content": "[20] Viktor Olsson, Wilhelm Tranheden, Juliano Pinto, and Lennart Svensson. Classmix: Segmentation-based data augmentation for semi-supervised learning. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1369-1378, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[21] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12674-12684, 2020. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21458" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[22] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[23] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "text", + "content": "[24] Jiawei Su, Zhiming Luo, Sheng Lian, Dazhen Lin, and Shaozi Li. Mutual learning with reliable pseudo label for semi-supervised medical image segmentation. Medical Image Analysis, 94:103111, 2024. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 294 + ], + "type": "text", + "content": "[25] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "text", + "content": "[26] Bing Wang, Taifeng Huang, Shuo Yang, Ying Yang, Junhai Zhai, and Xin Zhang. Dual-decoder data decoupling training for semi-supervised medical image segmentation. Biomedical Signal Processing and Control, 100:106984, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 393 + ], + "type": "text", + "content": "[27] Yongchao Wang, Bin Xiao, Xiuli Bi, Weisheng Li, and Xinbo Gao. Mcf: Mutual correction framework for semi-supervised medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15651-15660, 2023. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 470 + ], + "type": "text", + "content": "[28] Yicheng Wu, Minfeng Xu, Zongyuan Ge, Jianfei Cai, and Lei Zhang. Semi-supervised left atrium segmentation with mutual consistency training. In Medical image computing and computer assisted intervention-MICCAI 2021: 24th international conference, Strasbourg, France, September 27-October 1, 2021, proceedings, part II 24, pages 297-306. Springer, 2021. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[29] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. Advances in neural information processing systems, 33:6256-6268, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 580 + ], + "type": "text", + "content": "[30] Zhe Xu, Yixin Wang, Donghuan Lu, Lequan Yu, Jiangpeng Yan, Jie Luo, Kai Ma, Yefeng Zheng, and Raymond Kaiyu Tong. All-around real label supervision: Cyclic prototype consistency learning for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 26(7):3174-3184, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[31] Zhe Xu, Yixin Wang, Donghuan Lu, Xiangde Luo, Jiangpeng Yan, Yefeng Zheng, and Raymond Kai-yu Tong. Ambiguity-selective consistency regularization for meantacher semi-supervised medical image segmentation. Medical Image Analysis, 88:102880, 2023. 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "text", + "content": "[32] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-training work better for semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4268-4277, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[33] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 385 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "semi-supervised semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7236-7246, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 184 + ], + "type": "text", + "content": "[34] Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In Medical image computing and computer assisted intervention-MICCAI 2019: 22nd international conference, Shenzhen, China, October 13–17, 2019, proceedings, part II 22, pages 605–613. Springer, 2019. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 239 + ], + "type": "text", + "content": "[35] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6023-6032, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 240, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 284 + ], + "type": "text", + "content": "[36] Yichi Zhang, Rushi Jiao, Qingcheng Liao, Dongyang Li, and Jicong Zhang. Uncertainty-guided mutual consistency learning for semi-supervised medical image segmentation. Artificial Intelligence in Medicine, 138:102476, 2023. 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 340 + ], + "type": "text", + "content": "[37] Xiangyu Zhao, Zengxin Qi, Sheng Wang, Qian Wang, Xuehai Wu, Ying Mao, and Lichi Zhang. Rcps: Rectified contrastive pseudo supervision for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 28(1):251-261, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 341, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 554, + 385 + ], + "type": "text", + "content": "[38] Zhongda Zhao, Haiyan Wang, Tao Lei, Xuan Wang, Xiaohong Shen, and Haiyang Yao. Balanced feature fusion collaborative training for semi-supervised medical image segmentation. Pattern Recognition, 157:110856, 2025. 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21459" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_content_list.json b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a01a45d1d6bca1c1bd873c7a3d775920c20c230f --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": "Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongkang Zhang $^{1,2}$ Dongyu She $^{2}$ Zhong Zhou $^{1,2,*}$", + "bbox": [ + 267, + 202, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, China $^{2}$ Zhongguancun Laboratory, Beijing, China", + "bbox": [ + 129, + 220, + 867, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zyk，zz}@buaa.edu.cn，shedy@zgclab.edu.cn", + "bbox": [ + 310, + 258, + 681, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Out-of-distribution (OOD) detection aims to distinguish whether detected objects belong to known categories or not. Existing methods extract OOD samples from Indistribution (ID) data to regularize the model's decision boundaries. However, the decision boundaries are not adequately regularized because the model does not have sufficient knowledge about the distribution of OOD data. To address the above issue, we propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis (APL-GOS) for OOD detection. Specifically, we leverage the Vision-Language Model (VLM) to initialize learnable ID prompts by sampling standardized results from pre-defined Q&A pairs. Region-level prompts are synthesised in low-likelihood regions of class-conditional gaussian distributions. These prompts are then utilized to initialize learnable OOD prompts and optimized with adaptive prompt learning. Also, OOD pseudo-samples are synthesised via gaussian outlier synthesis. The aforementioned methodology regularizes the model to learn more compact decision boundaries for ID and OOD categories. Extensive experiments show that APLGOS achieves state-of-the-art performance with less ID data on four mainstream datasets.", + "bbox": [ + 88, + 340, + 485, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has made significant progress in recent years. It encompasses a multitude of research domains, including object detection [22, 40, 52], autonomous driving [41, 51] and image generation [20, 42]. Various existing deep learning methods rely on large-scale datasets to regularize the model, enabling it to learn sufficient data distribution and supervision signals of the training data. In real-world scenarios, where the number of unknown categories is significantly greater than that in the training dataset, the model lacks knowledge about the distribution of unknown data in", + "bbox": [ + 89, + 727, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8e0bda1e5603869afc5b7406559ead6efef0d0508171576b86f37136ca38d19.jpg", + "image_caption": [ + "Figure 1. Quantitative comparisons with state-of-the-art OOD detection methods in terms of FPR95, AUROC and mAP metrics. Note that larger points denote higher mAP, and the numerical values are also given next to each point. Our APLGOS provides remarkable performance boost on all the metrics." + ], + "image_footnote": [], + "bbox": [ + 516, + 308, + 905, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "practical applications and struggles to learn compact decision boundaries that effectively distinguish between known and unknown categories. During the testing phase, unknown categories are likely to result in erroneous predictions accompanied by a high confidence score. This leads to severe safety risks in critical safety domains such as autonomous driving.", + "bbox": [ + 511, + 613, + 906, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "OD detection [5, 21, 24, 33] is a research hotspot in recent years, which aims to enable the detectors to accurately distinguish not only seen categories, but also unseen categories during training. The detectors need to learn compact decision boundaries during training, ensuring low uncertainty for ID categories while maintaining high uncertainty away from them. To achieve this, existing OOD detection methods [12, 13, 34-36] provide sufficient supervision of OOD data for model training by extracting OOD pseudo-samples from ID data, helping the model better distinguish between known and unknown categories. However, due to the unpredictable quality of OOD pseudo", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3235", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "samples extracted from the ID data and the requirement for a large volume of ID data, the detector is not adequately regularized to learn compact decision boundaries for both ID and OOD categories. Therefore, synthesis-based methods [6, 18] have been proposed to generate OOD pseudo-samples. They synthesize out-of-distribution RGB images directly or virtual outliers in lower-dimensional hidden space, which to some extent mitigate the limitations of extracting pseudo-samples from ID data. In recent years, Vision-Language Models (VLMs) [30, 31, 48], owing to their powerful pre-trained knowledge and representation capabilities, have achieved considerable success and been applied across numerous fields.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose APLGOS, a synthesis-based vision-language method that leverages the powerful learning and representation capabilities of VLMs to assist in synthesizing virtual outliers using ID data. APLGOS mainly consists of Prompt Learning Module (PLM) and Text-Image Alignment Module (TAM). PLM employs two distinct strategies to generate ID prompts and OOD pseudo-prompts, respectively, to assist in regularizing the model. For ID data, we first provide a predefined Q&A pair and templates with location and category names, e.g., “Q: What is in the region with coordinates $, , , $ ? A: That’s a $$ ”, guiding the detector to incorporate location coordinates for more fine-grained observation. We guide ChatGPT through multiple rounds of standardization for the aforementioned prompts to generate a set of statements for the model to sample during training. The statements sampled from this set are then directly used to initialize the learnable ID prompts. In order to ensure that the generated OOD pseudo-samples better fit the distribution of OOD data, PLM generates OOD prompts through adaptive prompt learning via Gaussian Outlier Synthesis, where it samples virtual OOD prompts in the low-likelihood region of the class-conditional Gaussian distribution of ID prompts in high-dimensional hidden space. TAM calculates similarity scores for images and prompts and combines contrastive learning to align multimodal data, thereby regularizing model's decision boundaries.", + "bbox": [ + 91, + 291, + 483, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are:", + "bbox": [ + 109, + 717, + 460, + 731 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a vision-language OOD detection model namely APLGOS. Through adaptive prompt learning, APLGOS generates adaptive region-level prompts for ID and OOD images. Based on contrastive learning, APLGOS calculates similarity for images and prompts to ensure model learn compact decision boundaries.", + "- ID prompts, OOD prompts and OOD images are all virtual. ChatGPT standardizes pre-defined Q&A pairs with templates and instructions. Then we sample them to initialize learnable ID prompts. We synthesise virtual OOD prompts and OOD images in low-likelihood regions of" + ], + "bbox": [ + 91, + 734, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "class-conditional Gaussian distribution.", + "bbox": [ + 527, + 90, + 789, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experiments on mainstream datasets show that APLGOS achieves state-of-the-art performance in terms of FPR95, AUROC, AUPR and mAP metrics. Compared to the baseline method [6], when using Berkley DeepDrive-100k as ID dataset and OpenImages as OOD dataset, our method reduces FPR95 by $7.76\\%$ .", + "bbox": [ + 513, + 106, + 906, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 217, + 653, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Out-of-distribution Detection", + "text_level": 1, + "bbox": [ + 513, + 244, + 774, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD detection [14, 26, 36, 39] aims to learn a compact decision boundary on training data that allows model to detect not only the categories with low uncertainty, that have been seen in training phase, but also the unseen categories with high uncertainty. Since in physical world, the number of unseen categories for the model is much bigger than seen categories, using large-scale dataset to regularize the model [13, 27] is difficult to fully cover all unseen categories of physical world. Liang et al. [23] use temperature scaling and add small perturbations to the input to separate the softmax score distributions between ID and OOD images. Based on energy theory [17], the work [26] replace traditional softmax score with energy score to distinguish ID and OOD images. Recently, outlier based methods are proposed, which utilize outliers exposure [29, 47] or generate virtual outliers in pixel [9, 18] space or hidden feature space [6] to regularize the model. Nevertheless, they are inefficient and the quality of the synthesised virtual outliers is worrying. With the emergence of vision-language models, vision-language model-based methods are proposed to address open-vocabulary problems [28, 37, 39]. To the best of our knowledge, no prior work has explored the use of prompt learning in OOD detection task.", + "bbox": [ + 511, + 270, + 906, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Prompt Learning", + "text_level": 1, + "bbox": [ + 513, + 633, + 684, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt learning is to view pre-trained language models, such as BERT [4], GPT [1, 2] and BLOOM [16] as knowledge bases, and use them to provide text prompts to optimize the performance of downstream tasks. In contrast to hand-designed prompts, the goal of prompt learning is to adaptively provide accurate prompts for downstream tasks. Zhou et al. [50] propose CoOp, which models a prompt's context words with learnable vectors while keeping pre-trained parameters fixed. To prevent CoOp from overfitting base classes, Zhou et al. [49] introduce CoCoOp, which uses conditional context optimization to generate an input-conditional token for each image, but this approach introduces high computational costs. At the same time, due to the effectiveness of prompt learning, there are various methods incorporating it with computer vision tasks [3, 8, 45, 46].", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3236", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bbe25e5b8f3e014da3260cc976908de2bfce9293fac31efebbfba0d13ed9f37e.jpg", + "image_caption": [ + "Figure 2. The proposed APLGOS network architecture. Prompt learning module is responsible for using ChatGPT to standardize Q&A pairs with guidance introduction and templates, then it generates a statements set. The module samples prompts from the statements set to initialize the learnable ID prompts, and synthesises virtual OOD prompts in low-likelihood regions of class-conditional gaussian distributions. The Text-Image Alignment Module computes similarity scores to align text and image embeddings in the hidden space." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 445, + 225, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis for OOD Detection. As shown in Figure 2, APLGOS mainly consists of two modules, i.e. PLM and TAM. PLM leverages ChatGPT to standardize pre-defined Q&A pairs using guidance instructions and predefined templates, generating a set of statements. During training, PLM samples statements from this set to initialize the learnable prompts. For ID categories, APLGOS directly employs the initialized prompts as input to the text encoder, whereas for OOD categories, it synthesizes virtual OOD prompts and images within the low-likelihood region of the class-conditional Gaussian distribution of ID classes in the hidden space. Notably, only ID images are sourced from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual and synthesized. This approach enables the model to enhance the quality of pseudo-samples with less ID data while better capturing the distribution of OOD data. Additionally, through contrastive learning, TAM computes similarity scores to align images and prompts within the high-dimensional hidden space.", + "bbox": [ + 88, + 474, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For clarity, we omit the batchsize of data in the following description and consider a single batch as an example. The input to APLGOS consists of two modalities: detected region images $[\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_b]$ extracted from a raw RGB image $\\mathbf{X}\\in \\mathbb{R}^{C\\times H\\times W}$ , and text prompts $\\mathbf{T}\\in \\mathbb{R}^{b\\times l}$ . Here, $C$ , $H$ , and $W$ denote the number of channels, height, and width of the image, respectively. $b$ represents the number of detected region images from a single raw RGB image.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$l$ indicates the length of the text prompts. The text input is given as $\\mathbf{T} = [\\mathbf{T}_1, \\mathbf{T}_2, \\dots, \\mathbf{T}_b]$ . token in the sampled prompts has been replaced with the corresponding labels.", + "bbox": [ + 511, + 446, + 906, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Prompt Learning Module", + "text_level": 1, + "bbox": [ + 511, + 500, + 748, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ID Prompts. To enhance the model's representation ability and more effectively regularize its decision boundaries, we generate a set of statements for the Prompt Learning Module to sample from, rather than using a single invariant statement to initialize the learnable ID prompts. Specifically, we first predefine a Q&A pair, such as “Q: What is in the region with coordinates $, , , $ ? A: That's a $$ .” We then input this Q&A pair into ChatGPT for standardization. During this process, we provide predefined templates and guiding instructions to ensure that ChatGPT standardizes the Q&A pair accordingly. The standardization process is illustrated below with an example prompt:", + "bbox": [ + 511, + 522, + 906, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega_ {0} = g \\left(\\mathrm {Q} ^ {\\mathrm {A}} + \\mathrm {M} + \\mathrm {G} _ {0}\\right), \\quad \\Omega_ {i} = g \\left(\\Omega_ {i - 1} + \\mathrm {G} _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 739, + 906, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega_{i}$ denotes generated prompt result in $i_{th}$ round, $\\mathrm{Q^A}$ denotes Q&A pair, M denotes predefined template, $\\mathrm{G}_i$ denotes guidance instruction for $i_{th}$ standardizing round and $g$ is ChatGPT's standardizing operation. We collect the statements from these $t$ rounds to obtain statements set $\\Omega_t$ . These statements are then used for sampling during the initialization of learnable ID prompts.", + "bbox": [ + 511, + 763, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce no extra character sets and vocabularies, and the generated prompts are represented in natural lan", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3237", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "guage. The learnable prompts follow the paradigm e.g. $\\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle < V_1 > < V_2 > \\dots < V_m >$ $\\langle CLS\\rangle$ , which is initialized by sampled prompt. $\\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle$ are learnable location tokens, which implicitly introduce location information into the prompts. $\\langle V_1\\rangle < \\langle V_2\\rangle \\dots < V_m\\rangle$ are learnable description tokens, and $m$ is its length. $\\langle CLS\\rangle$ is class token.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {T}} = f _ {\\theta} \\left(h \\left(r \\left(g \\left(\\Omega_ {t - 1} + \\mathrm {G} _ {t}\\right)\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 212, + 482, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{T}} = [\\hat{\\mathbf{T}}_1,\\hat{\\mathbf{T}}_2,\\dots ,\\hat{\\mathbf{T}}_b],\\hat{\\mathbf{T}}_i\\in \\mathbb{R}^{l'},t$ is rounds of standardizing operations, $l^{\\prime}$ is length of prompt embedding. Here, for ease of understanding, we use one $\\hat{\\mathbf{T}}_i$ as an example to describe the subsequent operations, and standardize $\\hat{\\mathbf{T}}_i$ as $\\hat{\\mathbf{T}},\\hat{\\mathbf{T}}\\in \\mathbb{R}^{l'}$ . $f_{\\theta}$ is transformer-based text encoder, $h$ is tokenizer, $r$ is replacement function for $\\langle \\mathrm{CLS}\\rangle$ token. We replace $\\langle \\mathrm{CLS}\\rangle$ with the category label of the object in the current region (i.e., the corresponding ID class label).", + "bbox": [ + 89, + 239, + 483, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OOD Prompts. In the hidden space, distinct decision boundaries should be established between ID and OOD prompts. In the OOD detection task, we refine the decision boundaries as much as possible. By incorporating prompt learning, we synthesize region-level OOD pseudoprompts using Gaussian outlier synthesis. Specifically, the Prompt Learning Module synthesizes virtual OOD prompts in the low-likelihood regions of class-conditional Gaussian distributions in hidden space. This allows the Text-Image Alignment Module to perceive the distribution difference between ID and OOD categories in hidden space and align images and prompts through contrastive learning. Provided that the quantity of data is large enough, we assume the ID prompts embedding from the text encoder form a class-conditional multivariate Gaussian distribution:", + "bbox": [ + 89, + 388, + 483, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (\\hat {\\mathbf {T}} | y = i) = \\mathcal {N} (\\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 631, + 482, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ is the parameter of text encoder $f_{\\theta},y$ is ground truth label, $\\hat{\\mu}_i$ is empirical gaussian mean of $i_{th}$ in-distribution category prompts embedding, and $i\\in \\{1,2,\\dots,K\\}$ , $K$ represents the number of in-distribution classes, $\\mathcal{N}(\\hat{\\mu}_i,\\hat{\\sigma}) = \\frac{1}{\\sqrt{2\\pi\\hat{\\sigma}}} e^{-\\frac{(\\hat{\\Gamma} - \\hat{\\mu}_i)^2}{2\\hat{\\sigma}^2}}$ , $\\hat{\\sigma}$ denotes the tied covariance matrix.", + "bbox": [ + 89, + 657, + 482, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we calculate the empirical gaussian mean of $i_{th}$ ID category prompts embedding as follows:", + "bbox": [ + 89, + 743, + 482, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mu} _ {i} = \\frac {1}{| \\mathcal {Q} _ {T} |} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\hat {\\mathbf {T}} _ {i, j}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 786, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $|\\mathcal{Q}_T|$ denotes the length of the prompts queue $\\mathcal{Q}_T$ used to buffer ID prompts, and $\\mathcal{Q}_T\\in \\mathbb{R}^{K\\times |\\mathcal{Q}_T|}$ .", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we calculate the tied covariance matrix of ID prompts embedding as follows:", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} = \\frac {1}{K | \\mathcal {Q} _ {T} |} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) ^ {T} + \\beta \\mathbf {E}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 114, + 903, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is learnable matrix initialized by randomly gaussian noise, $\\mathbf{E}$ is unit matrix, $\\alpha, \\beta$ are hyper-parameters, $\\hat{\\sigma}$ is tied covariance matrix, and $\\hat{\\sigma} = [\\hat{\\sigma}_1, \\hat{\\sigma}_2, \\dots, \\hat{\\sigma}_K]^T$ .", + "bbox": [ + 511, + 172, + 903, + 218 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After computing the empirical Gaussian mean $\\hat{\\mu}$ and the tied covariance matrix $\\hat{\\sigma}$ , the Prompt Learning Module samples virtual OOD prompts from the low-likelihood regions of the class-conditional Gaussian distributions in hidden space, based on the estimated multivariate distributions. Then, it selects the top-k prompts with the lowest probability from this $\\epsilon$ -likelihood region:", + "bbox": [ + 511, + 218, + 906, + 323 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {V} _ {i} = \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 338, + 903, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Psi$ is class-conditional gaussian distribution probability density and satisfies the following relation:", + "bbox": [ + 511, + 359, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Psi (\\hat {\\mathbf {T}}, \\hat {\\boldsymbol {\\mu}} _ {1}, \\hat {\\boldsymbol {\\mu}} _ {2}, \\dots , \\hat {\\boldsymbol {\\mu}} _ {K}, \\hat {\\boldsymbol {\\sigma}}) = \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {1}, \\hat {\\sigma}) \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {2}, \\hat {\\sigma}) \\dots \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {K}, \\hat {\\sigma}), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 414, + 903, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each $\\Psi (\\hat{\\mathbf{T}},\\hat{\\mu}_i,\\hat{\\sigma})$ , its expansion can be formulated as:", + "bbox": [ + 511, + 467, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi \\left(\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}\\right) = \\left\\{v _ {i} \\mid \\frac {1}{\\sqrt {2 \\pi^ {\\frac {l ^ {\\prime}}{2}}} \\left| \\hat {\\sigma} \\right| ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right) ^ {T} \\hat {\\sigma} ^ {-} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right)} < \\epsilon \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 505, + 906, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{i} \\sim \\mathcal{N}(\\hat{\\mu}_{i},\\hat{\\sigma})$ denotes sampled virtual prompt using $i_{th}$ ID category prompts, $i = \\{1,2,\\dots,K\\}$ , and “-” denotes matrix inverse operation. The final synthesised OOD prompts are denoted as $\\hat{\\mathbf{T}}^{\\dagger}$ .", + "bbox": [ + 511, + 558, + 905, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. OOD Virtual Images Synthesis", + "text_level": 1, + "bbox": [ + 511, + 631, + 785, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing methods [12, 13, 34-36] directly extract OOD pseudo-samples from ID data. However, the extracted pseudo-samples are unable to fit the distribution of OOD data adequately. We also use synthesis method to get OOD data. The principle of synthesizing OOD image is similar to Eq. 3 to Eq. 8. Compared with synthesizing OOD prompts, the input for calculating empirical Gaussian mean and tied covariance is ID image embedding instead of ID prompts embedding. We define the final synthesised virtual images using current ID image embedding queue $\\mathcal{Q}_I$ as $\\hat{\\mathbf{X}}^{\\dagger}$ .", + "bbox": [ + 511, + 652, + 906, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Text-Image Alignment Module", + "text_level": 1, + "bbox": [ + 511, + 816, + 785, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first encode ID and OOD images and prompts to generate their embeddings. Then, the similarity score between prompts embedding and image embeddings is computed as follows:", + "bbox": [ + 511, + 839, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3238", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\frac {\\left| \\left| \\hat {\\mathbf {X}} \\right| \\right| _ {p} \\left(\\left| \\left| \\hat {\\mathbf {T}} \\right| \\right| _ {p}\\right) ^ {T}}{e ^ {\\omega}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 102, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}$ is the embedding of detected region images in the second training phase, and $\\hat{\\mathbf{X}} = [\\hat{\\mathbf{X}}_1,\\hat{\\mathbf{X}}_2,\\dots ,\\hat{\\mathbf{X}}_b],\\hat{\\mathbf{X}}_i$ is one detected region image embedding, $\\hat{\\mathbf{X}}_i\\in \\mathbb{R}^{l'}$ . In the third training phase, the input is embedding of synthesised virtual image $\\hat{\\mathbf{X}}^{\\dagger}$ instead of $\\hat{\\mathbf{X}},\\omega$ is hyper-parameters for scaling. $\\mathbf{S}$ is similarity score. The prompts embedding in Eq. 9 is ID prompts embedding $\\hat{\\mathbf{T}}$ in the second phase and synthesised OOD prompts embedding $\\hat{\\mathbf{T}}^{\\dagger}$ in the third phase, $||\\cdot ||_{p}$ is normalization, in addition, $||\\hat{\\mathbf{X}}_i||_p = \\hat{\\mathbf{X}}_i / \\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{X}}_{i,j}|^2}$ and $||\\hat{\\mathbf{T}}_i^\\dagger ||_p = \\hat{\\mathbf{T}}_i^\\dagger /\\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{T}}_{i,j}^\\dagger|^2}$", + "bbox": [ + 89, + 142, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 332, + 235, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alignment loss $\\mathcal{L}_{\\text {align }}$ constrains the contrastive learning process during alignment, receiving ID or OOD data at different training phases. The similarity score between prompts embedding and image embeddings is used to calculate the alignment loss:", + "bbox": [ + 89, + 354, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} (\\mathbf {S}, y) = - \\sum_ {i = 1} ^ {K ^ {\\prime}} t _ {i} \\log \\left(\\mathcal {R} _ {i} (\\mathbf {S})\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 443, + 482, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_i$ represents the category label of the object contained in the currently detected region. $\\mathcal{R}_i$ represents the standardized prediction score. We treat all OOD categories as a single category, i.e., \"background\". During the training phase, if the ID dataset contains a total of $K$ classes, each detected region image is required to calculate similarity scores with $(K + 1)$ text prompts, i.e., $K' = K + 1$ .", + "bbox": [ + 89, + 494, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous methods typically generate simple prompts that lack location information, such as \"a photo of $a < CLS >$ \" [49, 50], or provide brief prompts with relative location information for the entire image [43]. We argue that these prompts lack the fine granularity needed for the model to learn essential location information in vision-language-based detection tasks. $\\mathcal{L}_{loc}$ is designed to implicitly incorporate location information, enabling the generation of fine-grained prompts for detected image regions.", + "bbox": [ + 89, + 614, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l o c} = \\frac {\\lambda}{\\Phi (\\mathbf {B} _ {g})} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\sqrt {\\mathbf {B} _ {g}} _ {i} - \\sqrt {u \\left(\\mathbf {B} _ {r}\\right) _ {i}}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 775, + 482, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_g$ represents the ground truth coordinates of detected image region, $\\mathbf{B}_r$ represents the regression results of coordinates, and $\\mathbf{B}_g\\in \\mathbb{R}^{b\\times 4}$ , $\\mathbf{B}_r\\in \\mathbb{R}^{b\\times 4}$ , $z = 4$ , $u$ represents calculating absolute values, $\\Phi$ represents calculating the dimension of vector, $\\lambda$ is hyper-parameter.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After incorporating the classification loss $\\mathcal{L}_{cls}$ and the location loss $\\mathcal{L}_{loc}$ , the total loss can be expressed as:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\xi_ {1} \\left[ \\gamma_ {1} \\tau \\mathcal {L} _ {\\text {a l i g n}} ^ {i d} + \\gamma_ {2} (1 - \\tau) \\mathcal {L} _ {\\text {a l i g n}} ^ {o o d} \\right] \\\\ + \\gamma_ {3} \\xi_ {2} \\left[ \\kappa \\mathcal {L} _ {l o c} ^ {i d} + (1 - \\kappa) \\mathcal {L} _ {l o c} ^ {o o d} \\right] \\tag {12} \\\\ + \\gamma_ {4} \\xi_ {3} \\mathcal {L} _ {c l s} + \\gamma_ {5} \\xi_ {4} \\mathcal {L} _ {r e g} + \\overline {{\\mathscr {W}}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 132, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}$ are the hyper-parameters, $\\xi ,\\tau ,\\kappa$ determine the loss functions used in the current training phase and $\\xi_{i} = \\{0,1\\}$ $\\tau = \\{0,1\\}$ . In order to better regularize the model, in the actual implementation of $\\mathcal{L}$ we also add the regularization term $\\mathcal{W}$ and $\\mathcal{W}_i = [\\Delta^{2}_{(\\mathcal{F}(\\mathbf{O}_1),\\mathcal{B}_1)_i} + \\Delta^{2}_{(\\mathcal{G}(\\mathbf{O}_2),\\mathcal{B}_2)_i}],\\overline{\\mathcal{W}} = 1 / N\\sum_{i = 1}^{N}\\mathcal{W}_i,\\Delta^{2}_{(a,b)}$ represents $(a - b)^2$ $\\mathcal{F},\\mathcal{G}$ represent regression blocks, $\\mathbf{O}_i$ represents regularization matrix, $\\mathcal{B}_i$ represents bias matrix of regression block, $i = \\{1,2\\}$", + "bbox": [ + 511, + 202, + 906, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 354, + 645, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 378, + 614, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We verify our proposed APLGOS on four commonly used datasets: PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017 and OpenImages. The PASCAL VOC [7] dataset contains 9963 images in 20 categories, split into 5011 training and 4952 test images, with a resolution of $500 \\times 375$ ( $375 \\times 500$ ). The BDD-100k [44] dataset consists of 100,000 high-resolution driving scenarios with detailed road object annotations. The MS-COCO2017 [25] dataset includes 328,000 images across 91 categories and 2.5 million instance tags, with 82 categories having more than 5000 tags. OpenImages V4 [15] contains 9.2 million images across 500 categories, commonly used for classification, object detection, and visual relationship detection. The above four datasets comprehensively evaluate our proposed method from different aspects and perspectives.", + "bbox": [ + 511, + 400, + 906, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 637, + 730, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a transformer-based text encoder in the Prompt Learning Module, we employ ResNet50 [10] and RegNetX4.0 [32] as backbones, respectively. We use ChatGPT-3.5 to standardize Q&A pairs. The ratio of ID data used for training to synthesised OOD data is approximately 1:1. We use PASCAL VOC and Berkeley DeepDrive-100K as ID datasets, and evaluate on two OOD datasets containing subsets randomly sampled from MS-COCO2017 and OpenImages, respectively. To ensure the fairness of the test, we manually exclude categories in the OOD dataset that overlap with those in the ID dataset before evaluating on the OOD dataset. We set $B = 16$ and train APLGOS on PASCAL VOC for 18,000 iterations, and set $B = 8$ to train on Berkeley DeepDrive-100k for 90,000 iterations. We set the learning rate $lr = 0.01$ . The length of prompt embedding and length of image embedding $l' = 1024$ . We use 1000", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3239", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2447cca546421005cb3391856f55013a187b6a3e8c4d4f09afd3378d7ed651c0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ID DatasetMethodFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
PASCAL VOCMSP [11]70.99 / 73.1383.45 / 81.91-48.7
ODIN [23]59.82 / 63.1482.20 / 82.59-48.7
Mahalanobis [19]96.46 / 96.2759.25 / 57.42-48.7
Energy score [26]56.89 / 58.6983.69 / 82.98-48.7
Gram matrices [33]62.75 / 67.4279.88 / 77.62-48.7
Generalized ODIN [14]59.57 / 70.2883.12 / 79.23-48.1
CSI [36]59.91 / 57.4181.83 / 82.95-48.1
GAN-synthesis [18]60.93 / 59.9783.67 / 82.67-48.5
VOS-ResNet50* [6]48.28 / 52.1487.65 / 85.398.76 / 96.9847.8
VOS-RegX4.0* [6]50.53 / 50.2788.10 / 87.0898.92 / 97.8049.1
APLGOS (ResNet50)47.16 / 49.6687.89 / 85.9198.80 / 97.5448.8
APLGOS (RegNetX4.0)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
Berkeley DeepDrive-100kMSP [11]80.94 / 79.0475.87 / 77.38-31.2
ODIN [23]62.85 / 58.9274.44 / 76.61-31.2
Mahalanobis [19]57.66 / 60.1684.92 / 86.88-31.2
Energy score [26]60.06 / 54.9777.48 / 79.60-31.2
Gram matrices [33]60.93 / 77.5574.93 / 59.38-31.2
Generalized ODIN [14]57.27 / 50.1785.22 / 87.18-31.8
CSI [36]47.10 / 37.0684.09 / 87.99-30.6
GAN-synthesis [18]57.03 / 50.6178.82 / 81.25-31.4
VOS-ResNet50* [6]46.97 / 31.2584.97 / 89.8299.67 / 99.8635.7
VOS-RegX4.0* [6]42.82 / 27.5586.36 / 92.1199.76 / 99.9337.0
Dynamic Prototypes [38]45.72 / 35.0585.14 / 88.92-31.5
APLGOS (ResNet50)41.10 / 23.3087.36 / 92.8799.73 / 99.8935.8
APLGOS (RegNetX4.0)39.48 / 19.7987.47 / 93.5999.79 / 99.9437.6
", + "bbox": [ + 107, + 88, + 893, + 444 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6678457f62cd16546957ce4424139e9f357f6315210f54ecb252a5258acb067.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art methods on mainstream datasets. Here we use PASCAL VOC and Berkeley DeepDrive-100k as ID datasets, MS-COCO2017 and OpenImages as OOD datasets, respectively. “-” denotes that the data is not available." + ], + "table_footnote": [], + "table_body": "
StrategyFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
(a)VOS-RegNetX4.0* [6]50.53 / 50.2788.10 / 87.0898.82 / 97.8049.1
(b)[6] + <CLS>50.12 / 49.5088.56 / 86.8398.91 / 97.7948.2
(c)[6] + “a region of a” + <CLS>51.31 / 50.9688.20 / 86.7398.98 / 97.8548.7
(d)[6] + RP + <CLS>49.50 / 49.4088.49 / 86.7398.82 / 97.7748.9
(e)[6] + <LOC> + “a region of a” + <CLS>49.56 / 47.6088.23 / 87.0798.89 / 97.8749.1
(f)[6] + <LOC> + RP + <CLS> (Ours)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
", + "bbox": [ + 107, + 486, + 893, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Ablation studies for prompt strategies. “+” denotes the combination of strategies. “RP” represents sampled prompts from statements set, which is standardized by ChatGPT using Q&A pair and guidance instructions. (b) denotes the simplest prompt strategy, i.e., only providing the ground-truth label for the ID data, (for synthesised OOD image, we define its label as “background”). (c) denotes the original prompt strategy of CLIP [31]. (d) denotes that we replace the prompts in CLIP [31] with the statements by ChatGPT standardizing the Q&A pairs. (e) denotes adding location tokens $<\\mathrm{LOC}>$ to (c). (f) represents the prompts of our proposed APLGOS.", + "bbox": [ + 88, + 613, + 906, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "samples to estimate the class-conditional Gaussian distribution of ID image embeddings and 10000 samples for ID prompts embedding (i.e., $|\\mathcal{Q}_I| = 1000$ , $|\\mathcal{Q}_T| = 10000$ ). The total length $l$ of the standardized Q&A pair does not exceed 77. In the experimental tables, “*” denotes results from local replication based on open-source code. “↓” and “↑” indicate lower/greater is better, respectively.", + "bbox": [ + 88, + 696, + 486, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with The State-of-the-Art", + "text_level": 1, + "bbox": [ + 89, + 816, + 423, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the results of our proposed framework with different image encoder backbones (ResNet50 and RegNetX4.0) on PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017, and OpenImages datasets, as shown in", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The best results for the same dataset and the same backbone settings are shown in bold. For the same evaluation metric on the same dataset, the best results are underlined. When using Transformer-based text encoder and ResNet50-based image encoder, APLGOS achieves an FPR95 of $47.16\\%$ and an mAP of $48.8\\%$ on PASCAL VOC (ID) with MS-COCO2017 as the OOD dataset. When OpenImages is used as the OOD dataset, FPR95 increases to $49.66\\%$ . Compared to the state-of-the-art OOD detection model [6], APLGOS reduces FPR95 by $1.12\\%$ and $2.48\\%$ on MS-COCO2017 and OpenImages, respectively. With Transformer-based text encoder and RegNetX4.0-based image encoder, FPR95 decreases to $45.96\\%$ on MS-", + "bbox": [ + 511, + 696, + 908, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3240", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ee46864c4ae3f6bff9b72b80cd1398287898e58cda985355b911e8a7d54f141.jpg", + "image_caption": [ + "Figure 3. Ablation on number of sampled OOD prompts $\\mathcal{K}$ . The horizontal coordinate is the number of sampled ood prompts $\\mathcal{K} (\\times 10^3)$ , while the vertical coordinates are, from left to right, FPR95, AUROC, AUPR, and mAP, respectively. Red line and Blue line represent using MS-COCO2017 and OpenImages as OOD datasets, respectively. Pink line represents using PASCAL VOC as ID dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 269, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73c6ad0c18438ff92f9cc9d14ff5d7f846cf62942ac99a9b166ee318d2b2a349.jpg", + "image_caption": [ + "Number of sampled out-of-distribution prompts $(\\times 1000)$" + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 450, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e42fb880866069d188e5d5f77a6441d8294ddfd0716dda7e0e78d78e11b84c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 89, + 638, + 247 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cfaa103478de814d41924f66e90ef2fa4c6892b36f8fa401613060180cdb0528.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 90, + 901, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34af21c4a891e98f864dd5209d8977676c1911159887d680ca81570e37542c9c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
αFPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
051.63 / 50.8887.86 / 87.2449.2
0.551.90 / 51.4887.55 / 87.0248.9
1.045.96 / 47.1089.19 / 88.4949.4
1.555.88 / 53.3386.29 / 86.7548.9
2.055.92 / 49.5486.75 / 88.0048.9
", + "bbox": [ + 120, + 325, + 455, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "COCO2017 and $47.1\\%$ on OpenImages, while the mAP on PASCAL VOC improves to $49.4\\%$ . This setup further reduces FPR95 by $4.57\\%$ and $3.17\\%$ on MS-COCO2017 and OpenImages, respectively, compared to [6]. For Berkeley DeepDrive-100k (ID), using ResNet50-based image encoder and Transformer-based text encoder, APLGOS achieves an FPR95 of $41.10\\%$ on MS-COCO2017 and $23.30\\%$ on OpenImages, with an mAP of $35.8\\%$ . When using RegNetX4.0-based image encoder instead, FPR95 further decreases to $39.48\\%$ on MS-COCO2017 and $19.79\\%$ on OpenImages, while mAP improves to $37.6\\%$ .", + "bbox": [ + 89, + 515, + 483, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 695, + 254, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prompt strategies. To further validate the effectiveness of our prompt strategies, we conduct extensive ablation experiments on APLGOS's prompt strategies, and the results are shown in Table 2. Sampling from the statements set brings greater performance gains than simply initializing learnable prompts with \"a region of a\" ((c) vs (d)). Moreover, adding location tokens to prompts significantly improves performance, as it refines the scope of the prompts ((c) vs (e)). Compared to other prompt strategies, our APLGOS prompt strategy (f) integrates the advantages of the aforementioned strategies and achieves the best performance.", + "bbox": [ + 89, + 717, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Sampled OOD Prompts. APLGOS synthe", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f57427ca8ae30f04ab2cfa03145e495f53b7c5fbf9824314af718a823b1d6abd.jpg", + "table_caption": [ + "Table 3. The Ablation Experiments on The Strength of Random Gaussian Noise $\\varepsilon$ . $\\alpha$ represents the strength of added gaussian noise. The value of $\\alpha$ increases gradually from 0 to 2.0, and we take the value at 0.5 intervals." + ], + "table_footnote": [], + "table_body": "
Γ1FPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
1:450.11 / 58.3887.71 / 85.6749.1
1:349.40 / 55.1287.91 / 86.3849.2
1:247.98 / 54.4988.40 / 85.9449.2
1:145.96 / 47.1089.19 / 88.4949.4
2:148.25 / 50.2088.30 / 87.7649.2
3:150.95 / 53.9486.81 / 84.7047.5
4:150.20 / 51.5686.70 / 84.8947.3
", + "bbox": [ + 542, + 325, + 877, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. The ablation experiments on the ratio $\\Gamma_{1}$ of ID and OOD data used during training. Our default parameters and results are shown in bold. Parameters and results of baseline [6] are shown with a dark base color.", + "bbox": [ + 511, + 478, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sises virtual prompts for OOD categories and for each ID category, APLGOS samples $\\kappa$ virtual OOD prompts in low-likelihood regions of ID class-conditional gaussian distributions in high-dimensional hidden space. We conduct ablation experiments on $\\kappa$ , the results of its effect on performance are shown in Figure 3. When $\\kappa$ is too small, it may fail to adequately cover the region outside the ID categories' decision boundaries in the hidden space. On the other hand, when $\\kappa$ is too large, the excessive randomness in the sampled OOD prompts makes it difficult to effectively regularize the decision boundaries with the limited model parameters. Therefore, we set $\\kappa = 10000$ as the default value.", + "bbox": [ + 511, + 547, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Strength of Random Gaussian Noise $\\varepsilon$ . To enhance the size and diversity of the OOD prompts embedding sampling space and prevent the model from overly relying on the ID category distribution, we introduce a learnable matrix initialized with random Gaussian noise $\\varepsilon$ during the OOD prompt sampling stage (Eq. 5). We conduct ablation experiments on its strength $\\alpha$ , and the results are shown in Table 3. A small value of $\\alpha$ makes the sampling space of OOD prompts embedding too narrow, while a large value of $\\alpha$ results in an overly large sampling space. Only by appropriately expanding the sampling space of OOD prompts", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3241", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d14e61dc7c29443c7937d3b6931d8983489bba33a013ccb1803fbaf9b71dddfe.jpg", + "image_caption": [ + "Figure 4. Detection results on ID dataset. Here we use Berkeley DeepDrive-100k dataset as ID dataset. We use RegNetX4.0 and Transformer as backbone. The first row is the detection results of baseline [6]. The second row is the detection result of our APLGOS. Our APLGOS rarely misclassifies the ID class as OOD class, and there is almost no missed detection." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 883, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2eb827ad38b660063abcfa24ab55d7a413aa719962bb5b9fb9030164b2a1b70.jpg", + "image_caption": [ + "Figure 5. Detection results on OOD datasets. Here we use Berkeley DeepDrive-100k dataset as ID dataset, MS-COCO2017 and OpenImages as OOD datasets. The first row is the detection results of baseline [6]. The second row is the detection results of our APLGOS. Compared to the baseline, APLGOS rarely misses detections and hardly produces overlapping boxes for the same object." + ], + "image_footnote": [], + "bbox": [ + 112, + 276, + 883, + 404 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/daf57df2e22470715408fde8cad652fbaf0191c0d40a902dbf2f45661e00ec71.jpg", + "image_caption": [ + "Figure 6. Detection results in Real World. Here we use Berkeley DeepDrive-100k dataset as in-distribution dataset. Pictures we take ourselves with our phone as out-of-distribution dataset." + ], + "image_footnote": [], + "bbox": [ + 102, + 474, + 472, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "embedding can the model's ability to fit the OOD distribution be effectively enhanced.", + "bbox": [ + 89, + 703, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ratio of ID and OOD Data Used During Training. To verify that APLGOS can achieve better performance with less ID data, we conduct ablation experiments on the amount of ID data used during training, and the results are shown in Table 4. By default, APLGOS adopts a ratio $\\Gamma_{1}$ of 1:1 for ID and OOD data during training, whereas the baseline [6] uses a ratio of 2:1. However, in this case, the performance of APLGOS decreases instead.", + "bbox": [ + 89, + 734, + 482, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Detection Results. To better evaluate the performance of APLGOS, we visualize its detection results on ID datasets, OOD datasets, and real-world scenarios.", + "bbox": [ + 89, + 854, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results are presented in Figures 4, 5 and 6. The images in real-world scenarios are captured using an iPhone 14 Pro Max. The results demonstrate that APLGOS outperforms the baseline method in detecting ID and OOD categories. Moreover, real-world visualizations further demonstrate its strong generalization.", + "bbox": [ + 511, + 477, + 906, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 585, + 633, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a vision-language method, Adaptive Prompt Learning via Gaussian Outlier Synthesis (APLGOS) for Out-of-distribution Detection. Through prompt learning approach, APLGOS provides adaptive region-level prompts with location information for ID / OOD images. We use ChatGPT to standardize pre-defined Q&A pairs and generate a statements set. During training, only ID images are from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual. We sample statements from the statements set to initialize learnable ID prompts. We samples virtual OOD prompts and OOD images in the low-likelihood region of the class-conditional gaussian distribution in high-dimensional hidden space. The similarity score between prompts and images is utilized to calculate contrastive learning loss in high-dimensional hidden space, which guarantees the quality of virtual outliers as well as better regularization of the model. Through comprehensive experimental evaluations, we demonstrated the effectiveness of the proposed APLGOS.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3242", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This paper is supported by Zhongguancun Laboratory and the National Natural Science Foundation of China (Grant No. 62272018).", + "bbox": [ + 89, + 114, + 473, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 186 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877-1901, 2020. 2", + "[3] L. Chen, G. Wang, L. Yuan, K. Wang, K. Deng, and P. Torr. Nerf-vpt: Learning novel view representations with neural radiance fields via view prompt tuning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1156-1164, 2024. 2", + "[4] J. Devlin, M. Chang, K. Lee, and K. Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[5] X. Du, Z Wang, M. Cai, and S Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2022. 1", + "[6] X. Du, Z. Wang, M. Cai, and Y. Li. Vos: Learning what you don't know by virtual outlier synthesis. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7, 8", + "[7] M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 5", + "[8] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context learning for 3d point cloud understanding. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[9] M. Grcic, P. Bevandić, and S. Šegvic. Dense open-set recognition with synthetic outliers generated by real nvp. arXiv preprint arXiv:2011.11094, 2020. 2", + "[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5", + "[11] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR), 2016. 6", + "[12] D Hendrycks, M. Mazeika, and T. Dietterich. Deep anomaly detection with outlier exposure. International Conference on Learning Representations (ICLR), 2018. 1, 4", + "[13] D. Hendrycks, S. Basart, M. Mazeika, A. Zou, J. Kwon, M. Mostajabi, J. Steinhardt, and D. Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 4" + ], + "bbox": [ + 93, + 196, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Y. Hsu, Y. Shen, H. Jin, and Z. Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10951-10960, 2020. 2, 6", + "[15] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV), 128 (7):1956-1981, 2020. 5", + "[16] T. Le Scao, A. Fan, C. Akiki, E. Pavlick, S. Ilic, D. Hesslow, R. Castagné, A. Luccioni, F. Yvon, M. Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. 2023. 2", + "[17] Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2", + "[18] K. Lee, H. Lee, K. Lee, and J. Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. International Conference on Learning Representations (ICLR), 2017. 2, 6", + "[19] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems (NeurIPS), 31, 2018. 6", + "[20] D. Li, J. Li, and S. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1", + "[21] J. Li, P. Chen, Z. He, S. Yu, S. Liu, and J. Jia. Rethinking out-of-distribution (OOD) detection: Masked image modeling is all you need. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11578-11589, 2023. 1", + "[22] K. Li, Q. Geng, M. Wan, X. Cao, and Z. Zhou. Context and spatial feature calibration for real-time semantic segmentation. IEEE Transactions on Image Processing (TIP), 32: 5465-5477, 2023. 1", + "[23] S. Liang, Y. Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. International Conference on Learning Representations (ICLR), 2017. 2, 6", + "[24] J. Liao, X. Xu, M. Nguyen, A. Goodge, and C. Foo. Coftad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing (TIP), 2024. 1", + "[25] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 5", + "[26] W. Liu, X. Wang, J. Owens, and Y. Li. Energy-based out-of-distribution detection. Advances in neural information processing systems (NeurIPS), 33:21464-21475, 2020. 2, 6", + "[27] X. Liu, Y. Lochman, and C. Zach. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23946-23955, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3243", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems (NeurIPS), 35:35087-35102, 2022. 2", + "[29] S. Mohseni, M. Pitale, J. Yadawa, and Z. Wang. Self-supervised learning for generalizable out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 5216-5223, 2020. 2", + "[30] S. Ning, L. Qiu, Y. Liu, and X. He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2", + "[31] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763, 2021. 2, 6", + "[32] I. Radosavovic, R. Kosaraju, R. Girshick, K. He, and P. Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10428-10436, 2020. 5", + "[33] C. Sastry and S. Oore. Detecting out-of-distribution examples with gram matrices. In International Conference on Machine Learning (ICML), pages 8491-8501, 2020. 1, 6", + "[34] B. Su, H. Zhang, and Z. Zhou. Hsic-based moving weight averaging for few-shot open-set object detection. In Proceedings of the 31st ACM International Conference on Multimedia (ACMM), pages 5358-5369, 2023. 1, 4", + "[35] B. Su, H. Zhang, J. Li, and Z. Zhou. Toward generalized few-shot open-set object detection. IEEE Transactions on Image Processing (TIP), 33:1389-1402, 2024.", + "[36] J. Tack, S. Mo, J. Jeong, and J. Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems (NeurIPS), 33:11839-11852, 2020. 1, 2, 4, 6", + "[37] H. Wang, Y. Li, H. Yao, and X. Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1802-1812, 2023. 2", + "[38] A. Wu, C. Deng, and W. Liu. Unsupervised out-of-distribution object detection via pca-driven dynamic prototype enhancement. IEEE Transactions on Image Processing (TIP), 2024. 6", + "[39] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7031-7040, 2023. 2", + "[40] C. Xie, Z. Zhang, Y. Wu, F. Zhu, R. Zhao, and S. Liang. Described object detection: Liberating object detection with flexible expressions. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1", + "[41] X. Xiong, Y. Liu, T. Yuan, Y. Wang, Y. Wang, and H. Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17535-17544, 2023. 1" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1", + "[43] Z. Yang, Z. Gan, J. Wang, X. Hu, F. Ahmed, Z. Liu, Y. Lu, and L. Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision (ECCV), pages 521-539. Springer, 2022. 5", + "[44] F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2636-2645, 2020. 5", + "[45] Y. Zhai, Y. Zeng, Z. Huang, Z. Qin, X. Jin, and D. Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6979–6987, 2024. 2", + "[46] C. Zhang, X. Chen, S. Chai, C. Wu, D. Lagun, T. Beeler, and F. De la Torre. Iti-gen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3969-3980, 2023. 2", + "[47] J. Zhang, N. Inkawich, Y. Chen, and H. Li. Fine-grained out-of-distribution detection with mixup outlier exposure. arXiv preprint arXiv:2106.03917, 2(5), 2021. 2", + "[48] J. Zhang, J. Huang, S. Jin, and S. Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2", + "[49] K. Zhou, J. Yang, C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16816-16825, 2022. 2, 5", + "[50] K. Zhou, J. Yang, C. Loy, and Z. Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 2, 5", + "[51] Z. Zhu, Y. Zhang, H. Chen, Y. Dong, S. Zhao, W. Ding, J. Zhong, and S. Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21600-21610, 2023. 1", + "[52] O. Zohar, K. Wang, and S. Yeung. Prob: Probabilistic objectness for open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11444-11453, 2023. 1" + ], + "bbox": [ + 514, + 90, + 906, + 742 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3244", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_model.json b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e087ec286711420e8a2396ac46b6410c546e593 --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.188, + 0.131, + 0.812, + 0.175 + ], + "angle": 0, + "content": "Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.203, + 0.735, + 0.222 + ], + "angle": 0, + "content": "Yongkang Zhang\\(^{1,2}\\) Dongyu She\\(^{2}\\) Zhong Zhou\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.222, + 0.868, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, China \\(^{2}\\)Zhongguancun Laboratory, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.259, + 0.682, + 0.275 + ], + "angle": 0, + "content": "{zyk，zz}@buaa.edu.cn，shedy@zgclab.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.675 + ], + "angle": 0, + "content": "Out-of-distribution (OOD) detection aims to distinguish whether detected objects belong to known categories or not. Existing methods extract OOD samples from Indistribution (ID) data to regularize the model's decision boundaries. However, the decision boundaries are not adequately regularized because the model does not have sufficient knowledge about the distribution of OOD data. To address the above issue, we propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis (APL-GOS) for OOD detection. Specifically, we leverage the Vision-Language Model (VLM) to initialize learnable ID prompts by sampling standardized results from pre-defined Q&A pairs. Region-level prompts are synthesised in low-likelihood regions of class-conditional gaussian distributions. These prompts are then utilized to initialize learnable OOD prompts and optimized with adaptive prompt learning. Also, OOD pseudo-samples are synthesised via gaussian outlier synthesis. The aforementioned methodology regularizes the model to learn more compact decision boundaries for ID and OOD categories. Extensive experiments show that APLGOS achieves state-of-the-art performance with less ID data on four mainstream datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Deep learning has made significant progress in recent years. It encompasses a multitude of research domains, including object detection [22, 40, 52], autonomous driving [41, 51] and image generation [20, 42]. Various existing deep learning methods rely on large-scale datasets to regularize the model, enabling it to learn sufficient data distribution and supervision signals of the training data. In real-world scenarios, where the number of unknown categories is significantly greater than that in the training dataset, the model lacks knowledge about the distribution of unknown data in" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.309, + 0.906, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.608 + ], + "angle": 0, + "content": "Figure 1. Quantitative comparisons with state-of-the-art OOD detection methods in terms of FPR95, AUROC and mAP metrics. Note that larger points denote higher mAP, and the numerical values are also given next to each point. Our APLGOS provides remarkable performance boost on all the metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.72 + ], + "angle": 0, + "content": "practical applications and struggles to learn compact decision boundaries that effectively distinguish between known and unknown categories. During the testing phase, unknown categories are likely to result in erroneous predictions accompanied by a high confidence score. This leads to severe safety risks in critical safety domains such as autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "OD detection [5, 21, 24, 33] is a research hotspot in recent years, which aims to enable the detectors to accurately distinguish not only seen categories, but also unseen categories during training. The detectors need to learn compact decision boundaries during training, ensuring low uncertainty for ID categories while maintaining high uncertainty away from them. To achieve this, existing OOD detection methods [12, 13, 34-36] provide sufficient supervision of OOD data for model training by extracting OOD pseudo-samples from ID data, helping the model better distinguish between known and unknown categories. However, due to the unpredictable quality of OOD pseudo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3235" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "samples extracted from the ID data and the requirement for a large volume of ID data, the detector is not adequately regularized to learn compact decision boundaries for both ID and OOD categories. Therefore, synthesis-based methods [6, 18] have been proposed to generate OOD pseudo-samples. They synthesize out-of-distribution RGB images directly or virtual outliers in lower-dimensional hidden space, which to some extent mitigate the limitations of extracting pseudo-samples from ID data. In recent years, Vision-Language Models (VLMs) [30, 31, 48], owing to their powerful pre-trained knowledge and representation capabilities, have achieved considerable success and been applied across numerous fields." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.713 + ], + "angle": 0, + "content": "In this paper, we propose APLGOS, a synthesis-based vision-language method that leverages the powerful learning and representation capabilities of VLMs to assist in synthesizing virtual outliers using ID data. APLGOS mainly consists of Prompt Learning Module (PLM) and Text-Image Alignment Module (TAM). PLM employs two distinct strategies to generate ID prompts and OOD pseudo-prompts, respectively, to assist in regularizing the model. For ID data, we first provide a predefined Q&A pair and templates with location and category names, e.g., “Q: What is in the region with coordinates \\(, , , \\)? A: That’s a \\(\\)”, guiding the detector to incorporate location coordinates for more fine-grained observation. We guide ChatGPT through multiple rounds of standardization for the aforementioned prompts to generate a set of statements for the model to sample during training. The statements sampled from this set are then directly used to initialize the learnable ID prompts. In order to ensure that the generated OOD pseudo-samples better fit the distribution of OOD data, PLM generates OOD prompts through adaptive prompt learning via Gaussian Outlier Synthesis, where it samples virtual OOD prompts in the low-likelihood region of the class-conditional Gaussian distribution of ID prompts in high-dimensional hidden space. TAM calculates similarity scores for images and prompts and combines contrastive learning to align multimodal data, thereby regularizing model's decision boundaries." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.718, + 0.462, + 0.732 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.735, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- We propose a vision-language OOD detection model namely APLGOS. Through adaptive prompt learning, APLGOS generates adaptive region-level prompts for ID and OOD images. Based on contrastive learning, APLGOS calculates similarity for images and prompts to ensure model learn compact decision boundaries." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- ID prompts, OOD prompts and OOD images are all virtual. ChatGPT standardizes pre-defined Q&A pairs with templates and instructions. Then we sample them to initialize learnable ID prompts. We synthesise virtual OOD prompts and OOD images in low-likelihood regions of" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.79, + 0.106 + ], + "angle": 0, + "content": "class-conditional Gaussian distribution." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.907, + 0.198 + ], + "angle": 0, + "content": "- Extensive experiments on mainstream datasets show that APLGOS achieves state-of-the-art performance in terms of FPR95, AUROC, AUPR and mAP metrics. Compared to the baseline method [6], when using Berkley DeepDrive-100k as ID dataset and OpenImages as OOD dataset, our method reduces FPR95 by \\(7.76\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.218, + 0.655, + 0.234 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.246, + 0.776, + 0.261 + ], + "angle": 0, + "content": "2.1. Out-of-distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.271, + 0.908, + 0.619 + ], + "angle": 0, + "content": "OOD detection [14, 26, 36, 39] aims to learn a compact decision boundary on training data that allows model to detect not only the categories with low uncertainty, that have been seen in training phase, but also the unseen categories with high uncertainty. Since in physical world, the number of unseen categories for the model is much bigger than seen categories, using large-scale dataset to regularize the model [13, 27] is difficult to fully cover all unseen categories of physical world. Liang et al. [23] use temperature scaling and add small perturbations to the input to separate the softmax score distributions between ID and OOD images. Based on energy theory [17], the work [26] replace traditional softmax score with energy score to distinguish ID and OOD images. Recently, outlier based methods are proposed, which utilize outliers exposure [29, 47] or generate virtual outliers in pixel [9, 18] space or hidden feature space [6] to regularize the model. Nevertheless, they are inefficient and the quality of the synthesised virtual outliers is worrying. With the emergence of vision-language models, vision-language model-based methods are proposed to address open-vocabulary problems [28, 37, 39]. To the best of our knowledge, no prior work has explored the use of prompt learning in OOD detection task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.635, + 0.686, + 0.651 + ], + "angle": 0, + "content": "2.2. Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Prompt learning is to view pre-trained language models, such as BERT [4], GPT [1, 2] and BLOOM [16] as knowledge bases, and use them to provide text prompts to optimize the performance of downstream tasks. In contrast to hand-designed prompts, the goal of prompt learning is to adaptively provide accurate prompts for downstream tasks. Zhou et al. [50] propose CoOp, which models a prompt's context words with learnable vectors while keeping pre-trained parameters fixed. To prevent CoOp from overfitting base classes, Zhou et al. [49] introduce CoCoOp, which uses conditional context optimization to generate an input-conditional token for each image, but this approach introduces high computational costs. At the same time, due to the effectiveness of prompt learning, there are various methods incorporating it with computer vision tasks [3, 8, 45, 46]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3236" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.904, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.376, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. The proposed APLGOS network architecture. Prompt learning module is responsible for using ChatGPT to standardize Q&A pairs with guidance introduction and templates, then it generates a statements set. The module samples prompts from the statements set to initialize the learnable ID prompts, and synthesises virtual OOD prompts in low-likelihood regions of class-conditional gaussian distributions. The Text-Image Alignment Module computes similarity scores to align text and image embeddings in the hidden space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.446, + 0.226, + 0.464 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis for OOD Detection. As shown in Figure 2, APLGOS mainly consists of two modules, i.e. PLM and TAM. PLM leverages ChatGPT to standardize pre-defined Q&A pairs using guidance instructions and predefined templates, generating a set of statements. During training, PLM samples statements from this set to initialize the learnable prompts. For ID categories, APLGOS directly employs the initialized prompts as input to the text encoder, whereas for OOD categories, it synthesizes virtual OOD prompts and images within the low-likelihood region of the class-conditional Gaussian distribution of ID classes in the hidden space. Notably, only ID images are sourced from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual and synthesized. This approach enables the model to enhance the quality of pseudo-samples with less ID data while better capturing the distribution of OOD data. Additionally, through contrastive learning, TAM computes similarity scores to align images and prompts within the high-dimensional hidden space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For clarity, we omit the batchsize of data in the following description and consider a single batch as an example. The input to APLGOS consists of two modalities: detected region images \\([\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_b]\\) extracted from a raw RGB image \\(\\mathbf{X}\\in \\mathbb{R}^{C\\times H\\times W}\\), and text prompts \\(\\mathbf{T}\\in \\mathbb{R}^{b\\times l}\\). Here, \\(C\\), \\(H\\), and \\(W\\) denote the number of channels, height, and width of the image, respectively. \\(b\\) represents the number of detected region images from a single raw RGB image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.495 + ], + "angle": 0, + "content": "\\(l\\) indicates the length of the text prompts. The text input is given as \\(\\mathbf{T} = [\\mathbf{T}_1, \\mathbf{T}_2, \\dots, \\mathbf{T}_b]\\). token in the sampled prompts has been replaced with the corresponding labels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.749, + 0.517 + ], + "angle": 0, + "content": "3.1. Prompt Learning Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.721 + ], + "angle": 0, + "content": "ID Prompts. To enhance the model's representation ability and more effectively regularize its decision boundaries, we generate a set of statements for the Prompt Learning Module to sample from, rather than using a single invariant statement to initialize the learnable ID prompts. Specifically, we first predefine a Q&A pair, such as “Q: What is in the region with coordinates \\(, , , \\)? A: That's a \\(\\).” We then input this Q&A pair into ChatGPT for standardization. During this process, we provide predefined templates and guiding instructions to ensure that ChatGPT standardizes the Q&A pair accordingly. The standardization process is illustrated below with an example prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.741, + 0.907, + 0.759 + ], + "angle": 0, + "content": "\\[\n\\Omega_ {0} = g \\left(\\mathrm {Q} ^ {\\mathrm {A}} + \\mathrm {M} + \\mathrm {G} _ {0}\\right), \\quad \\Omega_ {i} = g \\left(\\Omega_ {i - 1} + \\mathrm {G} _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\(\\Omega_{i}\\) denotes generated prompt result in \\(i_{th}\\) round, \\(\\mathrm{Q^A}\\) denotes Q&A pair, M denotes predefined template, \\(\\mathrm{G}_i\\) denotes guidance instruction for \\(i_{th}\\) standardizing round and \\(g\\) is ChatGPT's standardizing operation. We collect the statements from these \\(t\\) rounds to obtain statements set \\(\\Omega_t\\). These statements are then used for sampling during the initialization of learnable ID prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We introduce no extra character sets and vocabularies, and the generated prompts are represented in natural lan" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3237" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "guage. The learnable prompts follow the paradigm e.g. \\( \\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle < V_1 > < V_2 > \\dots < V_m > \\) \\( \\langle CLS\\rangle \\), which is initialized by sampled prompt. \\( \\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle \\) are learnable location tokens, which implicitly introduce location information into the prompts. \\( \\langle V_1\\rangle < \\langle V_2\\rangle \\dots < V_m\\rangle \\) are learnable description tokens, and \\( m \\) is its length. \\( \\langle CLS\\rangle \\) is class token." + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.213, + 0.483, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {T}} = f _ {\\theta} \\left(h \\left(r \\left(g \\left(\\Omega_ {t - 1} + \\mathrm {G} _ {t}\\right)\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.485, + 0.368 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{T}} = [\\hat{\\mathbf{T}}_1,\\hat{\\mathbf{T}}_2,\\dots ,\\hat{\\mathbf{T}}_b],\\hat{\\mathbf{T}}_i\\in \\mathbb{R}^{l'},t\\) is rounds of standardizing operations, \\(l^{\\prime}\\) is length of prompt embedding. Here, for ease of understanding, we use one \\(\\hat{\\mathbf{T}}_i\\) as an example to describe the subsequent operations, and standardize \\(\\hat{\\mathbf{T}}_i\\) as \\(\\hat{\\mathbf{T}},\\hat{\\mathbf{T}}\\in \\mathbb{R}^{l'}\\) . \\(f_{\\theta}\\) is transformer-based text encoder, \\(h\\) is tokenizer, \\(r\\) is replacement function for \\(\\langle \\mathrm{CLS}\\rangle\\) token. We replace \\(\\langle \\mathrm{CLS}\\rangle\\) with the category label of the object in the current region (i.e., the corresponding ID class label)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.616 + ], + "angle": 0, + "content": "OOD Prompts. In the hidden space, distinct decision boundaries should be established between ID and OOD prompts. In the OOD detection task, we refine the decision boundaries as much as possible. By incorporating prompt learning, we synthesize region-level OOD pseudoprompts using Gaussian outlier synthesis. Specifically, the Prompt Learning Module synthesizes virtual OOD prompts in the low-likelihood regions of class-conditional Gaussian distributions in hidden space. This allows the Text-Image Alignment Module to perceive the distribution difference between ID and OOD categories in hidden space and align images and prompts through contrastive learning. Provided that the quantity of data is large enough, we assume the ID prompts embedding from the text encoder form a class-conditional multivariate Gaussian distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.632, + 0.483, + 0.65 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (\\hat {\\mathbf {T}} | y = i) = \\mathcal {N} (\\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.744 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the parameter of text encoder \\(f_{\\theta},y\\) is ground truth label, \\(\\hat{\\mu}_i\\) is empirical gaussian mean of \\(i_{th}\\) in-distribution category prompts embedding, and \\(i\\in \\{1,2,\\dots,K\\}\\), \\(K\\) represents the number of in-distribution classes, \\(\\mathcal{N}(\\hat{\\mu}_i,\\hat{\\sigma}) = \\frac{1}{\\sqrt{2\\pi\\hat{\\sigma}}} e^{-\\frac{(\\hat{\\Gamma} - \\hat{\\mu}_i)^2}{2\\hat{\\sigma}^2}}\\), \\(\\hat{\\sigma}\\) denotes the tied covariance matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.773 + ], + "angle": 0, + "content": "First, we calculate the empirical gaussian mean of \\( i_{th} \\) ID category prompts embedding as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.787, + 0.483, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mu} _ {i} = \\frac {1}{| \\mathcal {Q} _ {T} |} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\hat {\\mathbf {T}} _ {i, j}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(|\\mathcal{Q}_T|\\) denotes the length of the prompts queue \\(\\mathcal{Q}_T\\) used to buffer ID prompts, and \\(\\mathcal{Q}_T\\in \\mathbb{R}^{K\\times |\\mathcal{Q}_T|}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then we calculate the tied covariance matrix of ID prompts embedding as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.116, + 0.905, + 0.173 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} = \\frac {1}{K | \\mathcal {Q} _ {T} |} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) ^ {T} + \\beta \\mathbf {E}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.173, + 0.905, + 0.219 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is learnable matrix initialized by randomly gaussian noise, \\(\\mathbf{E}\\) is unit matrix, \\(\\alpha, \\beta\\) are hyper-parameters, \\(\\hat{\\sigma}\\) is tied covariance matrix, and \\(\\hat{\\sigma} = [\\hat{\\sigma}_1, \\hat{\\sigma}_2, \\dots, \\hat{\\sigma}_K]^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.907, + 0.324 + ], + "angle": 0, + "content": "After computing the empirical Gaussian mean \\(\\hat{\\mu}\\) and the tied covariance matrix \\(\\hat{\\sigma}\\), the Prompt Learning Module samples virtual OOD prompts from the low-likelihood regions of the class-conditional Gaussian distributions in hidden space, based on the estimated multivariate distributions. Then, it selects the top-k prompts with the lowest probability from this \\(\\epsilon\\)-likelihood region:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.339, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {V} _ {i} = \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.905, + 0.392 + ], + "angle": 0, + "content": "where \\(\\Psi\\) is class-conditional gaussian distribution probability density and satisfies the following relation:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.415, + 0.905, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Psi (\\hat {\\mathbf {T}}, \\hat {\\boldsymbol {\\mu}} _ {1}, \\hat {\\boldsymbol {\\mu}} _ {2}, \\dots , \\hat {\\boldsymbol {\\mu}} _ {K}, \\hat {\\boldsymbol {\\sigma}}) = \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {1}, \\hat {\\sigma}) \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {2}, \\hat {\\sigma}) \\dots \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {K}, \\hat {\\sigma}), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.468, + 0.905, + 0.497 + ], + "angle": 0, + "content": "For each \\(\\Psi (\\hat{\\mathbf{T}},\\hat{\\mu}_i,\\hat{\\sigma})\\) , its expansion can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.506, + 0.907, + 0.559 + ], + "angle": 0, + "content": "\\[\n\\Psi \\left(\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}\\right) = \\left\\{v _ {i} \\mid \\frac {1}{\\sqrt {2 \\pi^ {\\frac {l ^ {\\prime}}{2}}} \\left| \\hat {\\sigma} \\right| ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right) ^ {T} \\hat {\\sigma} ^ {-} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right)} < \\epsilon \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.559, + 0.906, + 0.624 + ], + "angle": 0, + "content": "where \\( v_{i} \\sim \\mathcal{N}(\\hat{\\mu}_{i},\\hat{\\sigma}) \\) denotes sampled virtual prompt using \\( i_{th} \\) ID category prompts, \\( i = \\{1,2,\\dots,K\\} \\), and “-” denotes matrix inverse operation. The final synthesised OOD prompts are denoted as \\( \\hat{\\mathbf{T}}^{\\dagger} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.632, + 0.787, + 0.649 + ], + "angle": 0, + "content": "3.2. OOD Virtual Images Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Existing methods [12, 13, 34-36] directly extract OOD pseudo-samples from ID data. However, the extracted pseudo-samples are unable to fit the distribution of OOD data adequately. We also use synthesis method to get OOD data. The principle of synthesizing OOD image is similar to Eq. 3 to Eq. 8. Compared with synthesizing OOD prompts, the input for calculating empirical Gaussian mean and tied covariance is ID image embedding instead of ID prompts embedding. We define the final synthesised virtual images using current ID image embedding queue \\(\\mathcal{Q}_I\\) as \\(\\hat{\\mathbf{X}}^{\\dagger}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.787, + 0.835 + ], + "angle": 0, + "content": "3.3. Text-Image Alignment Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.9 + ], + "angle": 0, + "content": "We first encode ID and OOD images and prompts to generate their embeddings. Then, the similarity score between prompts embedding and image embeddings is computed as follows:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3238" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.216, + 0.103, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\frac {\\left| \\left| \\hat {\\mathbf {X}} \\right| \\right| _ {p} \\left(\\left| \\left| \\hat {\\mathbf {T}} \\right| \\right| _ {p}\\right) ^ {T}}{e ^ {\\omega}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.143, + 0.484, + 0.326 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}\\) is the embedding of detected region images in the second training phase, and \\(\\hat{\\mathbf{X}} = [\\hat{\\mathbf{X}}_1,\\hat{\\mathbf{X}}_2,\\dots ,\\hat{\\mathbf{X}}_b],\\hat{\\mathbf{X}}_i\\) is one detected region image embedding, \\(\\hat{\\mathbf{X}}_i\\in \\mathbb{R}^{l'}\\) . In the third training phase, the input is embedding of synthesised virtual image \\(\\hat{\\mathbf{X}}^{\\dagger}\\) instead of \\(\\hat{\\mathbf{X}},\\omega\\) is hyper-parameters for scaling. \\(\\mathbf{S}\\) is similarity score. The prompts embedding in Eq. 9 is ID prompts embedding \\(\\hat{\\mathbf{T}}\\) in the second phase and synthesised OOD prompts embedding \\(\\hat{\\mathbf{T}}^{\\dagger}\\) in the third phase, \\(||\\cdot ||_{p}\\) is normalization, in addition, \\(||\\hat{\\mathbf{X}}_i||_p = \\hat{\\mathbf{X}}_i / \\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{X}}_{i,j}|^2}\\) and \\(||\\hat{\\mathbf{T}}_i^\\dagger ||_p = \\hat{\\mathbf{T}}_i^\\dagger /\\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{T}}_{i,j}^\\dagger|^2}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.236, + 0.348 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Alignment loss \\(\\mathcal{L}_{\\text {align }}\\) constrains the contrastive learning process during alignment, receiving ID or OOD data at different training phases. The similarity score between prompts embedding and image embeddings is used to calculate the alignment loss:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.444, + 0.483, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} (\\mathbf {S}, y) = - \\sum_ {i = 1} ^ {K ^ {\\prime}} t _ {i} \\log \\left(\\mathcal {R} _ {i} (\\mathbf {S})\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.601 + ], + "angle": 0, + "content": "where \\( t_i \\) represents the category label of the object contained in the currently detected region. \\( \\mathcal{R}_i \\) represents the standardized prediction score. We treat all OOD categories as a single category, i.e., \"background\". During the training phase, if the ID dataset contains a total of \\( K \\) classes, each detected region image is required to calculate similarity scores with \\( (K + 1) \\) text prompts, i.e., \\( K' = K + 1 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Previous methods typically generate simple prompts that lack location information, such as \"a photo of \\( a < CLS > \\)\" [49, 50], or provide brief prompts with relative location information for the entire image [43]. We argue that these prompts lack the fine granularity needed for the model to learn essential location information in vision-language-based detection tasks. \\( \\mathcal{L}_{loc} \\) is designed to implicitly incorporate location information, enabling the generation of fine-grained prompts for detected image regions." + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.776, + 0.483, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l o c} = \\frac {\\lambda}{\\Phi (\\mathbf {B} _ {g})} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\sqrt {\\mathbf {B} _ {g}} _ {i} - \\sqrt {u \\left(\\mathbf {B} _ {r}\\right) _ {i}}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_g\\) represents the ground truth coordinates of detected image region, \\(\\mathbf{B}_r\\) represents the regression results of coordinates, and \\(\\mathbf{B}_g\\in \\mathbb{R}^{b\\times 4}\\), \\(\\mathbf{B}_r\\in \\mathbb{R}^{b\\times 4}\\), \\(z = 4\\), \\(u\\) represents calculating absolute values, \\(\\Phi\\) represents calculating the dimension of vector, \\(\\lambda\\) is hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "After incorporating the classification loss \\(\\mathcal{L}_{cls}\\) and the location loss \\(\\mathcal{L}_{loc}\\), the total loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.133, + 0.905, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\xi_ {1} \\left[ \\gamma_ {1} \\tau \\mathcal {L} _ {\\text {a l i g n}} ^ {i d} + \\gamma_ {2} (1 - \\tau) \\mathcal {L} _ {\\text {a l i g n}} ^ {o o d} \\right] \\\\ + \\gamma_ {3} \\xi_ {2} \\left[ \\kappa \\mathcal {L} _ {l o c} ^ {i d} + (1 - \\kappa) \\mathcal {L} _ {l o c} ^ {o o d} \\right] \\tag {12} \\\\ + \\gamma_ {4} \\xi_ {3} \\mathcal {L} _ {c l s} + \\gamma_ {5} \\xi_ {4} \\mathcal {L} _ {r e g} + \\overline {{\\mathscr {W}}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Note that \\(\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}\\) are the hyper-parameters, \\(\\xi ,\\tau ,\\kappa\\) determine the loss functions used in the current training phase and \\(\\xi_{i} = \\{0,1\\}\\) \\(\\tau = \\{0,1\\}\\) . In order to better regularize the model, in the actual implementation of \\(\\mathcal{L}\\) we also add the regularization term \\(\\mathcal{W}\\) and \\(\\mathcal{W}_i = [\\Delta^{2}_{(\\mathcal{F}(\\mathbf{O}_1),\\mathcal{B}_1)_i} + \\Delta^{2}_{(\\mathcal{G}(\\mathbf{O}_2),\\mathcal{B}_2)_i}],\\overline{\\mathcal{W}} = 1 / N\\sum_{i = 1}^{N}\\mathcal{W}_i,\\Delta^{2}_{(a,b)}\\) represents \\((a - b)^2\\) \\(\\mathcal{F},\\mathcal{G}\\) represent regression blocks, \\(\\mathbf{O}_i\\) represents regularization matrix, \\(\\mathcal{B}_i\\) represents bias matrix of regression block, \\(i = \\{1,2\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.646, + 0.372 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.615, + 0.394 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.629 + ], + "angle": 0, + "content": "We verify our proposed APLGOS on four commonly used datasets: PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017 and OpenImages. The PASCAL VOC [7] dataset contains 9963 images in 20 categories, split into 5011 training and 4952 test images, with a resolution of \\(500 \\times 375\\) (\\(375 \\times 500\\)). The BDD-100k [44] dataset consists of 100,000 high-resolution driving scenarios with detailed road object annotations. The MS-COCO2017 [25] dataset includes 328,000 images across 91 categories and 2.5 million instance tags, with 82 categories having more than 5000 tags. OpenImages V4 [15] contains 9.2 million images across 500 categories, commonly used for classification, object detection, and visual relationship detection. The above four datasets comprehensively evaluate our proposed method from different aspects and perspectives." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.731, + 0.654 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We use a transformer-based text encoder in the Prompt Learning Module, we employ ResNet50 [10] and RegNetX4.0 [32] as backbones, respectively. We use ChatGPT-3.5 to standardize Q&A pairs. The ratio of ID data used for training to synthesised OOD data is approximately 1:1. We use PASCAL VOC and Berkeley DeepDrive-100K as ID datasets, and evaluate on two OOD datasets containing subsets randomly sampled from MS-COCO2017 and OpenImages, respectively. To ensure the fairness of the test, we manually exclude categories in the OOD dataset that overlap with those in the ID dataset before evaluating on the OOD dataset. We set \\( B = 16 \\) and train APLGOS on PASCAL VOC for 18,000 iterations, and set \\( B = 8 \\) to train on Berkeley DeepDrive-100k for 90,000 iterations. We set the learning rate \\( lr = 0.01 \\). The length of prompt embedding and length of image embedding \\( l' = 1024 \\). We use 1000" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3239" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.108, + 0.089, + 0.895, + 0.445 + ], + "angle": 0, + "content": "
ID DatasetMethodFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
PASCAL VOCMSP [11]70.99 / 73.1383.45 / 81.91-48.7
ODIN [23]59.82 / 63.1482.20 / 82.59-48.7
Mahalanobis [19]96.46 / 96.2759.25 / 57.42-48.7
Energy score [26]56.89 / 58.6983.69 / 82.98-48.7
Gram matrices [33]62.75 / 67.4279.88 / 77.62-48.7
Generalized ODIN [14]59.57 / 70.2883.12 / 79.23-48.1
CSI [36]59.91 / 57.4181.83 / 82.95-48.1
GAN-synthesis [18]60.93 / 59.9783.67 / 82.67-48.5
VOS-ResNet50* [6]48.28 / 52.1487.65 / 85.398.76 / 96.9847.8
VOS-RegX4.0* [6]50.53 / 50.2788.10 / 87.0898.92 / 97.8049.1
APLGOS (ResNet50)47.16 / 49.6687.89 / 85.9198.80 / 97.5448.8
APLGOS (RegNetX4.0)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
Berkeley DeepDrive-100kMSP [11]80.94 / 79.0475.87 / 77.38-31.2
ODIN [23]62.85 / 58.9274.44 / 76.61-31.2
Mahalanobis [19]57.66 / 60.1684.92 / 86.88-31.2
Energy score [26]60.06 / 54.9777.48 / 79.60-31.2
Gram matrices [33]60.93 / 77.5574.93 / 59.38-31.2
Generalized ODIN [14]57.27 / 50.1785.22 / 87.18-31.8
CSI [36]47.10 / 37.0684.09 / 87.99-30.6
GAN-synthesis [18]57.03 / 50.6178.82 / 81.25-31.4
VOS-ResNet50* [6]46.97 / 31.2584.97 / 89.8299.67 / 99.8635.7
VOS-RegX4.0* [6]42.82 / 27.5586.36 / 92.1199.76 / 99.9337.0
Dynamic Prototypes [38]45.72 / 35.0585.14 / 88.92-31.5
APLGOS (ResNet50)41.10 / 23.3087.36 / 92.8799.73 / 99.8935.8
APLGOS (RegNetX4.0)39.48 / 19.7987.47 / 93.5999.79 / 99.9437.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.452, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods on mainstream datasets. Here we use PASCAL VOC and Berkeley DeepDrive-100k as ID datasets, MS-COCO2017 and OpenImages as OOD datasets, respectively. “-” denotes that the data is not available." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.487, + 0.895, + 0.607 + ], + "angle": 0, + "content": "
StrategyFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
(a)VOS-RegNetX4.0* [6]50.53 / 50.2788.10 / 87.0898.82 / 97.8049.1
(b)[6] + <CLS>50.12 / 49.5088.56 / 86.8398.91 / 97.7948.2
(c)[6] + “a region of a” + <CLS>51.31 / 50.9688.20 / 86.7398.98 / 97.8548.7
(d)[6] + RP + <CLS>49.50 / 49.4088.49 / 86.7398.82 / 97.7748.9
(e)[6] + <LOC> + “a region of a” + <CLS>49.56 / 47.6088.23 / 87.0798.89 / 97.8749.1
(f)[6] + <LOC> + RP + <CLS> (Ours)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.614, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Table 2. Ablation studies for prompt strategies. “+” denotes the combination of strategies. “RP” represents sampled prompts from statements set, which is standardized by ChatGPT using Q&A pair and guidance instructions. (b) denotes the simplest prompt strategy, i.e., only providing the ground-truth label for the ID data, (for synthesised OOD image, we define its label as “background”). (c) denotes the original prompt strategy of CLIP [31]. (d) denotes that we replace the prompts in CLIP [31] with the statements by ChatGPT standardizing the Q&A pairs. (e) denotes adding location tokens \\(<\\mathrm{LOC}>\\) to (c). (f) represents the prompts of our proposed APLGOS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.697, + 0.487, + 0.804 + ], + "angle": 0, + "content": "samples to estimate the class-conditional Gaussian distribution of ID image embeddings and 10000 samples for ID prompts embedding (i.e., \\( |\\mathcal{Q}_I| = 1000 \\), \\( |\\mathcal{Q}_T| = 10000 \\)). The total length \\( l \\) of the standardized Q&A pair does not exceed 77. In the experimental tables, “*” denotes results from local replication based on open-source code. “↓” and “↑” indicate lower/greater is better, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.817, + 0.424, + 0.833 + ], + "angle": 0, + "content": "4.3. Comparison with The State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We report the results of our proposed framework with different image encoder backbones (ResNet50 and RegNetX4.0) on PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017, and OpenImages datasets, as shown in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.895 + ], + "angle": 0, + "content": "Table 1. The best results for the same dataset and the same backbone settings are shown in bold. For the same evaluation metric on the same dataset, the best results are underlined. When using Transformer-based text encoder and ResNet50-based image encoder, APLGOS achieves an FPR95 of \\(47.16\\%\\) and an mAP of \\(48.8\\%\\) on PASCAL VOC (ID) with MS-COCO2017 as the OOD dataset. When OpenImages is used as the OOD dataset, FPR95 increases to \\(49.66\\%\\). Compared to the state-of-the-art OOD detection model [6], APLGOS reduces FPR95 by \\(1.12\\%\\) and \\(2.48\\%\\) on MS-COCO2017 and OpenImages, respectively. With Transformer-based text encoder and RegNetX4.0-based image encoder, FPR95 decreases to \\(45.96\\%\\) on MS-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3240" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.27, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.452, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.09, + 0.639, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.091, + 0.902, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.25, + 0.625, + 0.261 + ], + "angle": 0, + "content": "Number of sampled out-of-distribution prompts \\((\\times 1000)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Figure 3. Ablation on number of sampled OOD prompts \\(\\mathcal{K}\\). The horizontal coordinate is the number of sampled ood prompts \\(\\mathcal{K} (\\times 10^3)\\), while the vertical coordinates are, from left to right, FPR95, AUROC, AUPR, and mAP, respectively. Red line and Blue line represent using MS-COCO2017 and OpenImages as OOD datasets, respectively. Pink line represents using PASCAL VOC as ID dataset." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.327, + 0.456, + 0.442 + ], + "angle": 0, + "content": "
αFPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
051.63 / 50.8887.86 / 87.2449.2
0.551.90 / 51.4887.55 / 87.0248.9
1.045.96 / 47.1089.19 / 88.4949.4
1.555.88 / 53.3386.29 / 86.7548.9
2.055.92 / 49.5486.75 / 88.0048.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.509 + ], + "angle": 0, + "content": "Table 3. The Ablation Experiments on The Strength of Random Gaussian Noise \\(\\varepsilon\\). \\(\\alpha\\) represents the strength of added gaussian noise. The value of \\(\\alpha\\) increases gradually from 0 to 2.0, and we take the value at 0.5 intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.683 + ], + "angle": 0, + "content": "COCO2017 and \\(47.1\\%\\) on OpenImages, while the mAP on PASCAL VOC improves to \\(49.4\\%\\). This setup further reduces FPR95 by \\(4.57\\%\\) and \\(3.17\\%\\) on MS-COCO2017 and OpenImages, respectively, compared to [6]. For Berkeley DeepDrive-100k (ID), using ResNet50-based image encoder and Transformer-based text encoder, APLGOS achieves an FPR95 of \\(41.10\\%\\) on MS-COCO2017 and \\(23.30\\%\\) on OpenImages, with an mAP of \\(35.8\\%\\). When using RegNetX4.0-based image encoder instead, FPR95 further decreases to \\(39.48\\%\\) on MS-COCO2017 and \\(19.79\\%\\) on OpenImages, while mAP improves to \\(37.6\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.255, + 0.71 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Prompt strategies. To further validate the effectiveness of our prompt strategies, we conduct extensive ablation experiments on APLGOS's prompt strategies, and the results are shown in Table 2. Sampling from the statements set brings greater performance gains than simply initializing learnable prompts with \"a region of a\" ((c) vs (d)). Moreover, adding location tokens to prompts significantly improves performance, as it refines the scope of the prompts ((c) vs (e)). Compared to other prompt strategies, our APLGOS prompt strategy (f) integrates the advantages of the aforementioned strategies and achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Number of Sampled OOD Prompts. APLGOS synthe" + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.327, + 0.878, + 0.469 + ], + "angle": 0, + "content": "
Γ1FPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
1:450.11 / 58.3887.71 / 85.6749.1
1:349.40 / 55.1287.91 / 86.3849.2
1:247.98 / 54.4988.40 / 85.9449.2
1:145.96 / 47.1089.19 / 88.4949.4
2:148.25 / 50.2088.30 / 87.7649.2
3:150.95 / 53.9486.81 / 84.7047.5
4:150.20 / 51.5686.70 / 84.8947.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.479, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Table 4. The ablation experiments on the ratio \\( \\Gamma_{1} \\) of ID and OOD data used during training. Our default parameters and results are shown in bold. Parameters and results of baseline [6] are shown with a dark base color." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.73 + ], + "angle": 0, + "content": "sises virtual prompts for OOD categories and for each ID category, APLGOS samples \\(\\kappa\\) virtual OOD prompts in low-likelihood regions of ID class-conditional gaussian distributions in high-dimensional hidden space. We conduct ablation experiments on \\(\\kappa\\), the results of its effect on performance are shown in Figure 3. When \\(\\kappa\\) is too small, it may fail to adequately cover the region outside the ID categories' decision boundaries in the hidden space. On the other hand, when \\(\\kappa\\) is too large, the excessive randomness in the sampled OOD prompts makes it difficult to effectively regularize the decision boundaries with the limited model parameters. Therefore, we set \\(\\kappa = 10000\\) as the default value." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Strength of Random Gaussian Noise \\(\\varepsilon\\). To enhance the size and diversity of the OOD prompts embedding sampling space and prevent the model from overly relying on the ID category distribution, we introduce a learnable matrix initialized with random Gaussian noise \\(\\varepsilon\\) during the OOD prompt sampling stage (Eq. 5). We conduct ablation experiments on its strength \\(\\alpha\\), and the results are shown in Table 3. A small value of \\(\\alpha\\) makes the sampling space of OOD prompts embedding too narrow, while a large value of \\(\\alpha\\) results in an overly large sampling space. Only by appropriately expanding the sampling space of OOD prompts" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "3241" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.885, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.908, + 0.271 + ], + "angle": 0, + "content": "Figure 4. Detection results on ID dataset. Here we use Berkeley DeepDrive-100k dataset as ID dataset. We use RegNetX4.0 and Transformer as backbone. The first row is the detection results of baseline [6]. The second row is the detection result of our APLGOS. Our APLGOS rarely misclassifies the ID class as OOD class, and there is almost no missed detection." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.277, + 0.885, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.417, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 5. Detection results on OOD datasets. Here we use Berkeley DeepDrive-100k dataset as ID dataset, MS-COCO2017 and OpenImages as OOD datasets. The first row is the detection results of baseline [6]. The second row is the detection results of our APLGOS. Compared to the baseline, APLGOS rarely misses detections and hardly produces overlapping boxes for the same object." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.476, + 0.473, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.69 + ], + "angle": 0, + "content": "Figure 6. Detection results in Real World. Here we use Berkeley DeepDrive-100k dataset as in-distribution dataset. Pictures we take ourselves with our phone as out-of-distribution dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.734 + ], + "angle": 0, + "content": "embedding can the model's ability to fit the OOD distribution be effectively enhanced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Ratio of ID and OOD Data Used During Training. To verify that APLGOS can achieve better performance with less ID data, we conduct ablation experiments on the amount of ID data used during training, and the results are shown in Table 4. By default, APLGOS adopts a ratio \\(\\Gamma_{1}\\) of 1:1 for ID and OOD data during training, whereas the baseline [6] uses a ratio of 2:1. However, in this case, the performance of APLGOS decreases instead." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Visualization of Detection Results. To better evaluate the performance of APLGOS, we visualize its detection results on ID datasets, OOD datasets, and real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.569 + ], + "angle": 0, + "content": "The results are presented in Figures 4, 5 and 6. The images in real-world scenarios are captured using an iPhone 14 Pro Max. The results demonstrate that APLGOS outperforms the baseline method in detecting ID and OOD categories. Moreover, real-world visualizations further demonstrate its strong generalization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.587, + 0.634, + 0.603 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a vision-language method, Adaptive Prompt Learning via Gaussian Outlier Synthesis (APLGOS) for Out-of-distribution Detection. Through prompt learning approach, APLGOS provides adaptive region-level prompts with location information for ID / OOD images. We use ChatGPT to standardize pre-defined Q&A pairs and generate a statements set. During training, only ID images are from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual. We sample statements from the statements set to initialize learnable ID prompts. We samples virtual OOD prompts and OOD images in the low-likelihood region of the class-conditional gaussian distribution in high-dimensional hidden space. The similarity score between prompts and images is utilized to calculate contrastive learning loss in high-dimensional hidden space, which guarantees the quality of virtual outliers as well as better regularization of the model. Through comprehensive experimental evaluations, we demonstrated the effectiveness of the proposed APLGOS." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3242" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.475, + 0.16 + ], + "angle": 0, + "content": "This paper is supported by Zhongguancun Laboratory and the National Natural Science Foundation of China (Grant No. 62272018)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.188 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.198, + 0.484, + 0.252 + ], + "angle": 0, + "content": "[1] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.255, + 0.484, + 0.323 + ], + "angle": 0, + "content": "[2] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.325, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[3] L. Chen, G. Wang, L. Yuan, K. Wang, K. Deng, and P. Torr. Nerf-vpt: Learning novel view representations with neural radiance fields via view prompt tuning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1156-1164, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.395, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[4] J. Devlin, M. Chang, K. Lee, and K. Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[5] X. Du, Z Wang, M. Cai, and S Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.533 + ], + "angle": 0, + "content": "[6] X. Du, Z. Wang, M. Cai, and Y. Li. Vos: Learning what you don't know by virtual outlier synthesis. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.589 + ], + "angle": 0, + "content": "[7] M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[8] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context learning for 3d point cloud understanding. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[9] M. Grcic, P. Bevandić, and S. Šegvic. Dense open-set recognition with synthetic outliers generated by real nvp. arXiv preprint arXiv:2011.11094, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[12] D Hendrycks, M. Mazeika, and T. Dietterich. Deep anomaly detection with outlier exposure. International Conference on Learning Representations (ICLR), 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] D. Hendrycks, S. Basart, M. Mazeika, A. Zou, J. Kwon, M. Mostajabi, J. Steinhardt, and D. Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.198, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Y. Hsu, Y. Shen, H. Jin, and Z. Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10951-10960, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV), 128 (7):1956-1981, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] T. Le Scao, A. Fan, C. Akiki, E. Pavlick, S. Ilic, D. Hesslow, R. Castagné, A. Luccioni, F. Yvon, M. Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[17] Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[18] K. Lee, H. Lee, K. Lee, and J. Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. International Conference on Learning Representations (ICLR), 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[19] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems (NeurIPS), 31, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[20] D. Li, J. Li, and S. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[21] J. Li, P. Chen, Z. He, S. Yu, S. Liu, and J. Jia. Rethinking out-of-distribution (OOD) detection: Masked image modeling is all you need. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11578-11589, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[22] K. Li, Q. Geng, M. Wan, X. Cao, and Z. Zhou. Context and spatial feature calibration for real-time semantic segmentation. IEEE Transactions on Image Processing (TIP), 32: 5465-5477, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.7 + ], + "angle": 0, + "content": "[23] S. Liang, Y. Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. International Conference on Learning Representations (ICLR), 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[24] J. Liao, X. Xu, M. Nguyen, A. Goodge, and C. Foo. Coftad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing (TIP), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[26] W. Liu, X. Wang, J. Owens, and Y. Li. Energy-based out-of-distribution detection. Advances in neural information processing systems (NeurIPS), 33:21464-21475, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] X. Liu, Y. Lochman, and C. Zach. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23946-23955, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3243" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems (NeurIPS), 35:35087-35102, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[29] S. Mohseni, M. Pitale, J. Yadawa, and Z. Wang. Self-supervised learning for generalizable out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 5216-5223, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.485, + 0.288 + ], + "angle": 0, + "content": "[30] S. Ning, L. Qiu, Y. Liu, and X. He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.485, + 0.357 + ], + "angle": 0, + "content": "[31] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.485, + 0.413 + ], + "angle": 0, + "content": "[32] I. Radosavovic, R. Kosaraju, R. Girshick, K. He, and P. Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10428-10436, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.485, + 0.455 + ], + "angle": 0, + "content": "[33] C. Sastry and S. Oore. Detecting out-of-distribution examples with gram matrices. In International Conference on Machine Learning (ICML), pages 8491-8501, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.485, + 0.511 + ], + "angle": 0, + "content": "[34] B. Su, H. Zhang, and Z. Zhou. Hsic-based moving weight averaging for few-shot open-set object detection. In Proceedings of the 31st ACM International Conference on Multimedia (ACMM), pages 5358-5369, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.485, + 0.553 + ], + "angle": 0, + "content": "[35] B. Su, H. Zhang, J. Li, and Z. Zhou. Toward generalized few-shot open-set object detection. IEEE Transactions on Image Processing (TIP), 33:1389-1402, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.485, + 0.608 + ], + "angle": 0, + "content": "[36] J. Tack, S. Mo, J. Jeong, and J. Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems (NeurIPS), 33:11839-11852, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.485, + 0.664 + ], + "angle": 0, + "content": "[37] H. Wang, Y. Li, H. Yao, and X. Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1802-1812, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.485, + 0.719 + ], + "angle": 0, + "content": "[38] A. Wu, C. Deng, and W. Liu. Unsupervised out-of-distribution object detection via pca-driven dynamic prototype enhancement. IEEE Transactions on Image Processing (TIP), 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.485, + 0.789 + ], + "angle": 0, + "content": "[39] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7031-7040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.485, + 0.845 + ], + "angle": 0, + "content": "[40] C. Xie, Z. Zhang, Y. Wu, F. Zhu, R. Zhao, and S. Liang. Described object detection: Liberating object detection with flexible expressions. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[41] X. Xiong, Y. Liu, T. Yuan, Y. Wang, Y. Wang, and H. Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17535-17544, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.149 + ], + "angle": 0, + "content": "[42] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[43] Z. Yang, Z. Gan, J. Wang, X. Hu, F. Ahmed, Z. Liu, Y. Lu, and L. Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision (ECCV), pages 521-539. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.218, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[44] F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2636-2645, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.29, + 0.908, + 0.359 + ], + "angle": 0, + "content": "[45] Y. Zhai, Y. Zeng, Z. Huang, Z. Qin, X. Jin, and D. Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6979–6987, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.36, + 0.908, + 0.416 + ], + "angle": 0, + "content": "[46] C. Zhang, X. Chen, S. Chai, C. Wu, D. Lagun, T. Beeler, and F. De la Torre. Iti-gen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3969-3980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.416, + 0.908, + 0.46 + ], + "angle": 0, + "content": "[47] J. Zhang, N. Inkawich, Y. Chen, and H. Li. Fine-grained out-of-distribution detection with mixup outlier exposure. arXiv preprint arXiv:2106.03917, 2(5), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.46, + 0.908, + 0.503 + ], + "angle": 0, + "content": "[48] J. Zhang, J. Huang, S. Jin, and S. Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.503, + 0.908, + 0.559 + ], + "angle": 0, + "content": "[49] K. Zhou, J. Yang, C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16816-16825, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.559, + 0.908, + 0.601 + ], + "angle": 0, + "content": "[50] K. Zhou, J. Yang, C. Loy, and Z. Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.601, + 0.908, + 0.686 + ], + "angle": 0, + "content": "[51] Z. Zhu, Y. Zhang, H. Chen, Y. Dong, S. Zhao, W. Ding, J. Zhong, and S. Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21600-21610, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.686, + 0.908, + 0.743 + ], + "angle": 0, + "content": "[52] O. Zohar, K. Wang, and S. Yeung. Prob: Probabilistic objectness for open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11444-11453, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3244" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_origin.pdf b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1967f32263288d5b30547ec748f031ace16ae56 --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/64d4c9d4-3703-4d8c-8396-0f791cf0fe59_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:649f59978d41dc3b163d964919a8c7aa7c54caa32087d378043103ebe59256db +size 4079291 diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/full.md b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b44b5f0afe72035f4dd298721091efbe05caf0e3 --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/full.md @@ -0,0 +1,301 @@ +# Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection + +Yongkang Zhang $^{1,2}$ Dongyu She $^{2}$ Zhong Zhou $^{1,2,*}$ + +$^{1}$ State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, China $^{2}$ Zhongguancun Laboratory, Beijing, China + +{zyk，zz}@buaa.edu.cn，shedy@zgclab.edu.cn + +# Abstract + +Out-of-distribution (OOD) detection aims to distinguish whether detected objects belong to known categories or not. Existing methods extract OOD samples from Indistribution (ID) data to regularize the model's decision boundaries. However, the decision boundaries are not adequately regularized because the model does not have sufficient knowledge about the distribution of OOD data. To address the above issue, we propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis (APL-GOS) for OOD detection. Specifically, we leverage the Vision-Language Model (VLM) to initialize learnable ID prompts by sampling standardized results from pre-defined Q&A pairs. Region-level prompts are synthesised in low-likelihood regions of class-conditional gaussian distributions. These prompts are then utilized to initialize learnable OOD prompts and optimized with adaptive prompt learning. Also, OOD pseudo-samples are synthesised via gaussian outlier synthesis. The aforementioned methodology regularizes the model to learn more compact decision boundaries for ID and OOD categories. Extensive experiments show that APLGOS achieves state-of-the-art performance with less ID data on four mainstream datasets. + +# 1. Introduction + +Deep learning has made significant progress in recent years. It encompasses a multitude of research domains, including object detection [22, 40, 52], autonomous driving [41, 51] and image generation [20, 42]. Various existing deep learning methods rely on large-scale datasets to regularize the model, enabling it to learn sufficient data distribution and supervision signals of the training data. In real-world scenarios, where the number of unknown categories is significantly greater than that in the training dataset, the model lacks knowledge about the distribution of unknown data in + +![](images/a8e0bda1e5603869afc5b7406559ead6efef0d0508171576b86f37136ca38d19.jpg) +Figure 1. Quantitative comparisons with state-of-the-art OOD detection methods in terms of FPR95, AUROC and mAP metrics. Note that larger points denote higher mAP, and the numerical values are also given next to each point. Our APLGOS provides remarkable performance boost on all the metrics. + +practical applications and struggles to learn compact decision boundaries that effectively distinguish between known and unknown categories. During the testing phase, unknown categories are likely to result in erroneous predictions accompanied by a high confidence score. This leads to severe safety risks in critical safety domains such as autonomous driving. + +OD detection [5, 21, 24, 33] is a research hotspot in recent years, which aims to enable the detectors to accurately distinguish not only seen categories, but also unseen categories during training. The detectors need to learn compact decision boundaries during training, ensuring low uncertainty for ID categories while maintaining high uncertainty away from them. To achieve this, existing OOD detection methods [12, 13, 34-36] provide sufficient supervision of OOD data for model training by extracting OOD pseudo-samples from ID data, helping the model better distinguish between known and unknown categories. However, due to the unpredictable quality of OOD pseudo + +samples extracted from the ID data and the requirement for a large volume of ID data, the detector is not adequately regularized to learn compact decision boundaries for both ID and OOD categories. Therefore, synthesis-based methods [6, 18] have been proposed to generate OOD pseudo-samples. They synthesize out-of-distribution RGB images directly or virtual outliers in lower-dimensional hidden space, which to some extent mitigate the limitations of extracting pseudo-samples from ID data. In recent years, Vision-Language Models (VLMs) [30, 31, 48], owing to their powerful pre-trained knowledge and representation capabilities, have achieved considerable success and been applied across numerous fields. + +In this paper, we propose APLGOS, a synthesis-based vision-language method that leverages the powerful learning and representation capabilities of VLMs to assist in synthesizing virtual outliers using ID data. APLGOS mainly consists of Prompt Learning Module (PLM) and Text-Image Alignment Module (TAM). PLM employs two distinct strategies to generate ID prompts and OOD pseudo-prompts, respectively, to assist in regularizing the model. For ID data, we first provide a predefined Q&A pair and templates with location and category names, e.g., “Q: What is in the region with coordinates $, , , $ ? A: That’s a $$ ”, guiding the detector to incorporate location coordinates for more fine-grained observation. We guide ChatGPT through multiple rounds of standardization for the aforementioned prompts to generate a set of statements for the model to sample during training. The statements sampled from this set are then directly used to initialize the learnable ID prompts. In order to ensure that the generated OOD pseudo-samples better fit the distribution of OOD data, PLM generates OOD prompts through adaptive prompt learning via Gaussian Outlier Synthesis, where it samples virtual OOD prompts in the low-likelihood region of the class-conditional Gaussian distribution of ID prompts in high-dimensional hidden space. TAM calculates similarity scores for images and prompts and combines contrastive learning to align multimodal data, thereby regularizing model's decision boundaries. + +In summary, the main contributions of this paper are: + +- We propose a vision-language OOD detection model namely APLGOS. Through adaptive prompt learning, APLGOS generates adaptive region-level prompts for ID and OOD images. Based on contrastive learning, APLGOS calculates similarity for images and prompts to ensure model learn compact decision boundaries. +- ID prompts, OOD prompts and OOD images are all virtual. ChatGPT standardizes pre-defined Q&A pairs with templates and instructions. Then we sample them to initialize learnable ID prompts. We synthesise virtual OOD prompts and OOD images in low-likelihood regions of + +class-conditional Gaussian distribution. + +- Extensive experiments on mainstream datasets show that APLGOS achieves state-of-the-art performance in terms of FPR95, AUROC, AUPR and mAP metrics. Compared to the baseline method [6], when using Berkley DeepDrive-100k as ID dataset and OpenImages as OOD dataset, our method reduces FPR95 by $7.76\%$ . + +# 2. Related Work + +# 2.1. Out-of-distribution Detection + +OOD detection [14, 26, 36, 39] aims to learn a compact decision boundary on training data that allows model to detect not only the categories with low uncertainty, that have been seen in training phase, but also the unseen categories with high uncertainty. Since in physical world, the number of unseen categories for the model is much bigger than seen categories, using large-scale dataset to regularize the model [13, 27] is difficult to fully cover all unseen categories of physical world. Liang et al. [23] use temperature scaling and add small perturbations to the input to separate the softmax score distributions between ID and OOD images. Based on energy theory [17], the work [26] replace traditional softmax score with energy score to distinguish ID and OOD images. Recently, outlier based methods are proposed, which utilize outliers exposure [29, 47] or generate virtual outliers in pixel [9, 18] space or hidden feature space [6] to regularize the model. Nevertheless, they are inefficient and the quality of the synthesised virtual outliers is worrying. With the emergence of vision-language models, vision-language model-based methods are proposed to address open-vocabulary problems [28, 37, 39]. To the best of our knowledge, no prior work has explored the use of prompt learning in OOD detection task. + +# 2.2. Prompt Learning + +Prompt learning is to view pre-trained language models, such as BERT [4], GPT [1, 2] and BLOOM [16] as knowledge bases, and use them to provide text prompts to optimize the performance of downstream tasks. In contrast to hand-designed prompts, the goal of prompt learning is to adaptively provide accurate prompts for downstream tasks. Zhou et al. [50] propose CoOp, which models a prompt's context words with learnable vectors while keeping pre-trained parameters fixed. To prevent CoOp from overfitting base classes, Zhou et al. [49] introduce CoCoOp, which uses conditional context optimization to generate an input-conditional token for each image, but this approach introduces high computational costs. At the same time, due to the effectiveness of prompt learning, there are various methods incorporating it with computer vision tasks [3, 8, 45, 46]. + +![](images/bbe25e5b8f3e014da3260cc976908de2bfce9293fac31efebbfba0d13ed9f37e.jpg) +Figure 2. The proposed APLGOS network architecture. Prompt learning module is responsible for using ChatGPT to standardize Q&A pairs with guidance introduction and templates, then it generates a statements set. The module samples prompts from the statements set to initialize the learnable ID prompts, and synthesises virtual OOD prompts in low-likelihood regions of class-conditional gaussian distributions. The Text-Image Alignment Module computes similarity scores to align text and image embeddings in the hidden space. + +# 3. Methodology + +We propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis for OOD Detection. As shown in Figure 2, APLGOS mainly consists of two modules, i.e. PLM and TAM. PLM leverages ChatGPT to standardize pre-defined Q&A pairs using guidance instructions and predefined templates, generating a set of statements. During training, PLM samples statements from this set to initialize the learnable prompts. For ID categories, APLGOS directly employs the initialized prompts as input to the text encoder, whereas for OOD categories, it synthesizes virtual OOD prompts and images within the low-likelihood region of the class-conditional Gaussian distribution of ID classes in the hidden space. Notably, only ID images are sourced from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual and synthesized. This approach enables the model to enhance the quality of pseudo-samples with less ID data while better capturing the distribution of OOD data. Additionally, through contrastive learning, TAM computes similarity scores to align images and prompts within the high-dimensional hidden space. + +For clarity, we omit the batchsize of data in the following description and consider a single batch as an example. The input to APLGOS consists of two modalities: detected region images $[\mathbf{X}_1,\mathbf{X}_2,\dots,\mathbf{X}_b]$ extracted from a raw RGB image $\mathbf{X}\in \mathbb{R}^{C\times H\times W}$ , and text prompts $\mathbf{T}\in \mathbb{R}^{b\times l}$ . Here, $C$ , $H$ , and $W$ denote the number of channels, height, and width of the image, respectively. $b$ represents the number of detected region images from a single raw RGB image. + +$l$ indicates the length of the text prompts. The text input is given as $\mathbf{T} = [\mathbf{T}_1, \mathbf{T}_2, \dots, \mathbf{T}_b]$ . token in the sampled prompts has been replaced with the corresponding labels. + +# 3.1. Prompt Learning Module + +ID Prompts. To enhance the model's representation ability and more effectively regularize its decision boundaries, we generate a set of statements for the Prompt Learning Module to sample from, rather than using a single invariant statement to initialize the learnable ID prompts. Specifically, we first predefine a Q&A pair, such as “Q: What is in the region with coordinates $, , , $ ? A: That's a $$ .” We then input this Q&A pair into ChatGPT for standardization. During this process, we provide predefined templates and guiding instructions to ensure that ChatGPT standardizes the Q&A pair accordingly. The standardization process is illustrated below with an example prompt: + +$$ +\Omega_ {0} = g \left(\mathrm {Q} ^ {\mathrm {A}} + \mathrm {M} + \mathrm {G} _ {0}\right), \quad \Omega_ {i} = g \left(\Omega_ {i - 1} + \mathrm {G} _ {i}\right), \tag {1} +$$ + +where $\Omega_{i}$ denotes generated prompt result in $i_{th}$ round, $\mathrm{Q^A}$ denotes Q&A pair, M denotes predefined template, $\mathrm{G}_i$ denotes guidance instruction for $i_{th}$ standardizing round and $g$ is ChatGPT's standardizing operation. We collect the statements from these $t$ rounds to obtain statements set $\Omega_t$ . These statements are then used for sampling during the initialization of learnable ID prompts. + +We introduce no extra character sets and vocabularies, and the generated prompts are represented in natural lan + +guage. The learnable prompts follow the paradigm e.g. $\langle loc_1\rangle < \langle loc_2\rangle < \langle loc_3\rangle < \langle loc_4\rangle < V_1 > < V_2 > \dots < V_m >$ $\langle CLS\rangle$ , which is initialized by sampled prompt. $\langle loc_1\rangle < \langle loc_2\rangle < \langle loc_3\rangle < \langle loc_4\rangle$ are learnable location tokens, which implicitly introduce location information into the prompts. $\langle V_1\rangle < \langle V_2\rangle \dots < V_m\rangle$ are learnable description tokens, and $m$ is its length. $\langle CLS\rangle$ is class token. + +$$ +\hat {\mathbf {T}} = f _ {\theta} \left(h \left(r \left(g \left(\Omega_ {t - 1} + \mathrm {G} _ {t}\right)\right)\right)\right), \tag {2} +$$ + +where $\hat{\mathbf{T}} = [\hat{\mathbf{T}}_1,\hat{\mathbf{T}}_2,\dots ,\hat{\mathbf{T}}_b],\hat{\mathbf{T}}_i\in \mathbb{R}^{l'},t$ is rounds of standardizing operations, $l^{\prime}$ is length of prompt embedding. Here, for ease of understanding, we use one $\hat{\mathbf{T}}_i$ as an example to describe the subsequent operations, and standardize $\hat{\mathbf{T}}_i$ as $\hat{\mathbf{T}},\hat{\mathbf{T}}\in \mathbb{R}^{l'}$ . $f_{\theta}$ is transformer-based text encoder, $h$ is tokenizer, $r$ is replacement function for $\langle \mathrm{CLS}\rangle$ token. We replace $\langle \mathrm{CLS}\rangle$ with the category label of the object in the current region (i.e., the corresponding ID class label). + +OOD Prompts. In the hidden space, distinct decision boundaries should be established between ID and OOD prompts. In the OOD detection task, we refine the decision boundaries as much as possible. By incorporating prompt learning, we synthesize region-level OOD pseudoprompts using Gaussian outlier synthesis. Specifically, the Prompt Learning Module synthesizes virtual OOD prompts in the low-likelihood regions of class-conditional Gaussian distributions in hidden space. This allows the Text-Image Alignment Module to perceive the distribution difference between ID and OOD categories in hidden space and align images and prompts through contrastive learning. Provided that the quantity of data is large enough, we assume the ID prompts embedding from the text encoder form a class-conditional multivariate Gaussian distribution: + +$$ +p _ {\theta} (\hat {\mathbf {T}} | y = i) = \mathcal {N} (\hat {\mu} _ {i}, \hat {\sigma}), \tag {3} +$$ + +where $\theta$ is the parameter of text encoder $f_{\theta},y$ is ground truth label, $\hat{\mu}_i$ is empirical gaussian mean of $i_{th}$ in-distribution category prompts embedding, and $i\in \{1,2,\dots,K\}$ , $K$ represents the number of in-distribution classes, $\mathcal{N}(\hat{\mu}_i,\hat{\sigma}) = \frac{1}{\sqrt{2\pi\hat{\sigma}}} e^{-\frac{(\hat{\Gamma} - \hat{\mu}_i)^2}{2\hat{\sigma}^2}}$ , $\hat{\sigma}$ denotes the tied covariance matrix. + +First, we calculate the empirical gaussian mean of $i_{th}$ ID category prompts embedding as follows: + +$$ +\hat {\mu} _ {i} = \frac {1}{| \mathcal {Q} _ {T} |} \sum_ {j = 1} ^ {| \mathcal {Q} _ {T} |} \hat {\mathbf {T}} _ {i, j}, \tag {4} +$$ + +where $|\mathcal{Q}_T|$ denotes the length of the prompts queue $\mathcal{Q}_T$ used to buffer ID prompts, and $\mathcal{Q}_T\in \mathbb{R}^{K\times |\mathcal{Q}_T|}$ . + +Then we calculate the tied covariance matrix of ID prompts embedding as follows: + +$$ +\hat {\sigma} = \frac {1}{K | \mathcal {Q} _ {T} |} \sum_ {i = 1} ^ {K} \sum_ {j = 1} ^ {| \mathcal {Q} _ {T} |} \left(\hat {\mathbf {T}} _ {i, j} + \alpha \varepsilon - \hat {\mu} _ {i}\right) \left(\hat {\mathbf {T}} _ {i, j} + \alpha \varepsilon - \hat {\mu} _ {i}\right) ^ {T} + \beta \mathbf {E}, \tag {5} +$$ + +where $\varepsilon$ is learnable matrix initialized by randomly gaussian noise, $\mathbf{E}$ is unit matrix, $\alpha, \beta$ are hyper-parameters, $\hat{\sigma}$ is tied covariance matrix, and $\hat{\sigma} = [\hat{\sigma}_1, \hat{\sigma}_2, \dots, \hat{\sigma}_K]^T$ . + +After computing the empirical Gaussian mean $\hat{\mu}$ and the tied covariance matrix $\hat{\sigma}$ , the Prompt Learning Module samples virtual OOD prompts from the low-likelihood regions of the class-conditional Gaussian distributions in hidden space, based on the estimated multivariate distributions. Then, it selects the top-k prompts with the lowest probability from this $\epsilon$ -likelihood region: + +$$ +\mathcal {V} _ {i} = \Psi (\hat {\mathbf {T}}, \hat {\mu} _ {i}, \hat {\sigma}), \tag {6} +$$ + +where $\Psi$ is class-conditional gaussian distribution probability density and satisfies the following relation: + +$$ +\begin{array}{l} \Psi (\hat {\mathbf {T}}, \hat {\boldsymbol {\mu}} _ {1}, \hat {\boldsymbol {\mu}} _ {2}, \dots , \hat {\boldsymbol {\mu}} _ {K}, \hat {\boldsymbol {\sigma}}) = \\ \Psi (\hat {\mathbf {T}}, \hat {\mu} _ {1}, \hat {\sigma}) \Psi (\hat {\mathbf {T}}, \hat {\mu} _ {2}, \hat {\sigma}) \dots \Psi (\hat {\mathbf {T}}, \hat {\mu} _ {K}, \hat {\sigma}), \tag {7} \\ \end{array} +$$ + +For each $\Psi (\hat{\mathbf{T}},\hat{\mu}_i,\hat{\sigma})$ , its expansion can be formulated as: + +$$ +\Psi \left(\hat {\mathbf {T}}, \hat {\mu} _ {i}, \hat {\sigma}\right) = \left\{v _ {i} \mid \frac {1}{\sqrt {2 \pi^ {\frac {l ^ {\prime}}{2}}} \left| \hat {\sigma} \right| ^ {\frac {1}{2}}} e ^ {- \frac {1}{2} \left(v _ {i} - \hat {\mu} _ {i}\right) ^ {T} \hat {\sigma} ^ {-} \left(v _ {i} - \hat {\mu} _ {i}\right)} < \epsilon \right\}, \tag {8} +$$ + +where $v_{i} \sim \mathcal{N}(\hat{\mu}_{i},\hat{\sigma})$ denotes sampled virtual prompt using $i_{th}$ ID category prompts, $i = \{1,2,\dots,K\}$ , and “-” denotes matrix inverse operation. The final synthesised OOD prompts are denoted as $\hat{\mathbf{T}}^{\dagger}$ . + +# 3.2. OOD Virtual Images Synthesis + +Existing methods [12, 13, 34-36] directly extract OOD pseudo-samples from ID data. However, the extracted pseudo-samples are unable to fit the distribution of OOD data adequately. We also use synthesis method to get OOD data. The principle of synthesizing OOD image is similar to Eq. 3 to Eq. 8. Compared with synthesizing OOD prompts, the input for calculating empirical Gaussian mean and tied covariance is ID image embedding instead of ID prompts embedding. We define the final synthesised virtual images using current ID image embedding queue $\mathcal{Q}_I$ as $\hat{\mathbf{X}}^{\dagger}$ . + +# 3.3. Text-Image Alignment Module + +We first encode ID and OOD images and prompts to generate their embeddings. Then, the similarity score between prompts embedding and image embeddings is computed as follows: + +$$ +\mathbf {S} = \frac {\left| \left| \hat {\mathbf {X}} \right| \right| _ {p} \left(\left| \left| \hat {\mathbf {T}} \right| \right| _ {p}\right) ^ {T}}{e ^ {\omega}}, \tag {9} +$$ + +where $\hat{\mathbf{X}}$ is the embedding of detected region images in the second training phase, and $\hat{\mathbf{X}} = [\hat{\mathbf{X}}_1,\hat{\mathbf{X}}_2,\dots ,\hat{\mathbf{X}}_b],\hat{\mathbf{X}}_i$ is one detected region image embedding, $\hat{\mathbf{X}}_i\in \mathbb{R}^{l'}$ . In the third training phase, the input is embedding of synthesised virtual image $\hat{\mathbf{X}}^{\dagger}$ instead of $\hat{\mathbf{X}},\omega$ is hyper-parameters for scaling. $\mathbf{S}$ is similarity score. The prompts embedding in Eq. 9 is ID prompts embedding $\hat{\mathbf{T}}$ in the second phase and synthesised OOD prompts embedding $\hat{\mathbf{T}}^{\dagger}$ in the third phase, $||\cdot ||_{p}$ is normalization, in addition, $||\hat{\mathbf{X}}_i||_p = \hat{\mathbf{X}}_i / \sqrt{\sum_{j = 1}^{l'}|\hat{\mathbf{X}}_{i,j}|^2}$ and $||\hat{\mathbf{T}}_i^\dagger ||_p = \hat{\mathbf{T}}_i^\dagger /\sqrt{\sum_{j = 1}^{l'}|\hat{\mathbf{T}}_{i,j}^\dagger|^2}$ + +# 3.4. Loss Function + +Alignment loss $\mathcal{L}_{\text {align }}$ constrains the contrastive learning process during alignment, receiving ID or OOD data at different training phases. The similarity score between prompts embedding and image embeddings is used to calculate the alignment loss: + +$$ +\mathcal {L} _ {\text {a l i g n}} (\mathbf {S}, y) = - \sum_ {i = 1} ^ {K ^ {\prime}} t _ {i} \log \left(\mathcal {R} _ {i} (\mathbf {S})\right), \tag {10} +$$ + +where $t_i$ represents the category label of the object contained in the currently detected region. $\mathcal{R}_i$ represents the standardized prediction score. We treat all OOD categories as a single category, i.e., "background". During the training phase, if the ID dataset contains a total of $K$ classes, each detected region image is required to calculate similarity scores with $(K + 1)$ text prompts, i.e., $K' = K + 1$ . + +Previous methods typically generate simple prompts that lack location information, such as "a photo of $a < CLS >$ " [49, 50], or provide brief prompts with relative location information for the entire image [43]. We argue that these prompts lack the fine granularity needed for the model to learn essential location information in vision-language-based detection tasks. $\mathcal{L}_{loc}$ is designed to implicitly incorporate location information, enabling the generation of fine-grained prompts for detected image regions. + +$$ +\mathcal {L} _ {l o c} = \frac {\lambda}{\Phi (\mathbf {B} _ {g})} \left[ \sum_ {i = 1} ^ {z} \left(\sqrt {\mathbf {B} _ {g}} _ {i} - \sqrt {u \left(\mathbf {B} _ {r}\right) _ {i}}\right) ^ {2} \right] ^ {\frac {1}{2}}, \tag {11} +$$ + +where $\mathbf{B}_g$ represents the ground truth coordinates of detected image region, $\mathbf{B}_r$ represents the regression results of coordinates, and $\mathbf{B}_g\in \mathbb{R}^{b\times 4}$ , $\mathbf{B}_r\in \mathbb{R}^{b\times 4}$ , $z = 4$ , $u$ represents calculating absolute values, $\Phi$ represents calculating the dimension of vector, $\lambda$ is hyper-parameter. + +After incorporating the classification loss $\mathcal{L}_{cls}$ and the location loss $\mathcal{L}_{loc}$ , the total loss can be expressed as: + +$$ +\begin{array}{l} \mathcal {L} = \xi_ {1} \left[ \gamma_ {1} \tau \mathcal {L} _ {\text {a l i g n}} ^ {i d} + \gamma_ {2} (1 - \tau) \mathcal {L} _ {\text {a l i g n}} ^ {o o d} \right] \\ + \gamma_ {3} \xi_ {2} \left[ \kappa \mathcal {L} _ {l o c} ^ {i d} + (1 - \kappa) \mathcal {L} _ {l o c} ^ {o o d} \right] \tag {12} \\ + \gamma_ {4} \xi_ {3} \mathcal {L} _ {c l s} + \gamma_ {5} \xi_ {4} \mathcal {L} _ {r e g} + \overline {{\mathscr {W}}}. \\ \end{array} +$$ + +Note that $\gamma_{1},\gamma_{2},\gamma_{3},\gamma_{4},\gamma_{5}$ are the hyper-parameters, $\xi ,\tau ,\kappa$ determine the loss functions used in the current training phase and $\xi_{i} = \{0,1\}$ $\tau = \{0,1\}$ . In order to better regularize the model, in the actual implementation of $\mathcal{L}$ we also add the regularization term $\mathcal{W}$ and $\mathcal{W}_i = [\Delta^{2}_{(\mathcal{F}(\mathbf{O}_1),\mathcal{B}_1)_i} + \Delta^{2}_{(\mathcal{G}(\mathbf{O}_2),\mathcal{B}_2)_i}],\overline{\mathcal{W}} = 1 / N\sum_{i = 1}^{N}\mathcal{W}_i,\Delta^{2}_{(a,b)}$ represents $(a - b)^2$ $\mathcal{F},\mathcal{G}$ represent regression blocks, $\mathbf{O}_i$ represents regularization matrix, $\mathcal{B}_i$ represents bias matrix of regression block, $i = \{1,2\}$ + +# 4. Experiments + +# 4.1. Datasets + +We verify our proposed APLGOS on four commonly used datasets: PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017 and OpenImages. The PASCAL VOC [7] dataset contains 9963 images in 20 categories, split into 5011 training and 4952 test images, with a resolution of $500 \times 375$ ( $375 \times 500$ ). The BDD-100k [44] dataset consists of 100,000 high-resolution driving scenarios with detailed road object annotations. The MS-COCO2017 [25] dataset includes 328,000 images across 91 categories and 2.5 million instance tags, with 82 categories having more than 5000 tags. OpenImages V4 [15] contains 9.2 million images across 500 categories, commonly used for classification, object detection, and visual relationship detection. The above four datasets comprehensively evaluate our proposed method from different aspects and perspectives. + +# 4.2. Implementation Details + +We use a transformer-based text encoder in the Prompt Learning Module, we employ ResNet50 [10] and RegNetX4.0 [32] as backbones, respectively. We use ChatGPT-3.5 to standardize Q&A pairs. The ratio of ID data used for training to synthesised OOD data is approximately 1:1. We use PASCAL VOC and Berkeley DeepDrive-100K as ID datasets, and evaluate on two OOD datasets containing subsets randomly sampled from MS-COCO2017 and OpenImages, respectively. To ensure the fairness of the test, we manually exclude categories in the OOD dataset that overlap with those in the ID dataset before evaluating on the OOD dataset. We set $B = 16$ and train APLGOS on PASCAL VOC for 18,000 iterations, and set $B = 8$ to train on Berkeley DeepDrive-100k for 90,000 iterations. We set the learning rate $lr = 0.01$ . The length of prompt embedding and length of image embedding $l' = 1024$ . We use 1000 + +
ID DatasetMethodFPR95 ↓AUROC ↑ +OOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
PASCAL VOCMSP [11]70.99 / 73.1383.45 / 81.91-48.7
ODIN [23]59.82 / 63.1482.20 / 82.59-48.7
Mahalanobis [19]96.46 / 96.2759.25 / 57.42-48.7
Energy score [26]56.89 / 58.6983.69 / 82.98-48.7
Gram matrices [33]62.75 / 67.4279.88 / 77.62-48.7
Generalized ODIN [14]59.57 / 70.2883.12 / 79.23-48.1
CSI [36]59.91 / 57.4181.83 / 82.95-48.1
GAN-synthesis [18]60.93 / 59.9783.67 / 82.67-48.5
VOS-ResNet50* [6]48.28 / 52.1487.65 / 85.398.76 / 96.9847.8
VOS-RegX4.0* [6]50.53 / 50.2788.10 / 87.0898.92 / 97.8049.1
APLGOS (ResNet50)47.16 / 49.6687.89 / 85.9198.80 / 97.5448.8
APLGOS (RegNetX4.0)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
Berkeley DeepDrive-100kMSP [11]80.94 / 79.0475.87 / 77.38-31.2
ODIN [23]62.85 / 58.9274.44 / 76.61-31.2
Mahalanobis [19]57.66 / 60.1684.92 / 86.88-31.2
Energy score [26]60.06 / 54.9777.48 / 79.60-31.2
Gram matrices [33]60.93 / 77.5574.93 / 59.38-31.2
Generalized ODIN [14]57.27 / 50.1785.22 / 87.18-31.8
CSI [36]47.10 / 37.0684.09 / 87.99-30.6
GAN-synthesis [18]57.03 / 50.6178.82 / 81.25-31.4
VOS-ResNet50* [6]46.97 / 31.2584.97 / 89.8299.67 / 99.8635.7
VOS-RegX4.0* [6]42.82 / 27.5586.36 / 92.1199.76 / 99.9337.0
Dynamic Prototypes [38]45.72 / 35.0585.14 / 88.92-31.5
APLGOS (ResNet50)41.10 / 23.3087.36 / 92.8799.73 / 99.8935.8
APLGOS (RegNetX4.0)39.48 / 19.7987.47 / 93.5999.79 / 99.9437.6
+ +Table 1. Comparison with the state-of-the-art methods on mainstream datasets. Here we use PASCAL VOC and Berkeley DeepDrive-100k as ID datasets, MS-COCO2017 and OpenImages as OOD datasets, respectively. “-” denotes that the data is not available. + +
StrategyFPR95 ↓AUROC ↑ +OOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
(a)VOS-RegNetX4.0* [6]50.53 / 50.2788.10 / 87.0898.82 / 97.8049.1
(b)[6] + <CLS>50.12 / 49.5088.56 / 86.8398.91 / 97.7948.2
(c)[6] + “a region of a” + <CLS>51.31 / 50.9688.20 / 86.7398.98 / 97.8548.7
(d)[6] + RP + <CLS>49.50 / 49.4088.49 / 86.7398.82 / 97.7748.9
(e)[6] + <LOC> + “a region of a” + <CLS>49.56 / 47.6088.23 / 87.0798.89 / 97.8749.1
(f)[6] + <LOC> + RP + <CLS> (Ours)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
+ +Table 2. Ablation studies for prompt strategies. “+” denotes the combination of strategies. “RP” represents sampled prompts from statements set, which is standardized by ChatGPT using Q&A pair and guidance instructions. (b) denotes the simplest prompt strategy, i.e., only providing the ground-truth label for the ID data, (for synthesised OOD image, we define its label as “background”). (c) denotes the original prompt strategy of CLIP [31]. (d) denotes that we replace the prompts in CLIP [31] with the statements by ChatGPT standardizing the Q&A pairs. (e) denotes adding location tokens $<\mathrm{LOC}>$ to (c). (f) represents the prompts of our proposed APLGOS. + +samples to estimate the class-conditional Gaussian distribution of ID image embeddings and 10000 samples for ID prompts embedding (i.e., $|\mathcal{Q}_I| = 1000$ , $|\mathcal{Q}_T| = 10000$ ). The total length $l$ of the standardized Q&A pair does not exceed 77. In the experimental tables, “*” denotes results from local replication based on open-source code. “↓” and “↑” indicate lower/greater is better, respectively. + +# 4.3. Comparison with The State-of-the-Art + +We report the results of our proposed framework with different image encoder backbones (ResNet50 and RegNetX4.0) on PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017, and OpenImages datasets, as shown in + +Table 1. The best results for the same dataset and the same backbone settings are shown in bold. For the same evaluation metric on the same dataset, the best results are underlined. When using Transformer-based text encoder and ResNet50-based image encoder, APLGOS achieves an FPR95 of $47.16\%$ and an mAP of $48.8\%$ on PASCAL VOC (ID) with MS-COCO2017 as the OOD dataset. When OpenImages is used as the OOD dataset, FPR95 increases to $49.66\%$ . Compared to the state-of-the-art OOD detection model [6], APLGOS reduces FPR95 by $1.12\%$ and $2.48\%$ on MS-COCO2017 and OpenImages, respectively. With Transformer-based text encoder and RegNetX4.0-based image encoder, FPR95 decreases to $45.96\%$ on MS- + +![](images/0ee46864c4ae3f6bff9b72b80cd1398287898e58cda985355b911e8a7d54f141.jpg) +Figure 3. Ablation on number of sampled OOD prompts $\mathcal{K}$ . The horizontal coordinate is the number of sampled ood prompts $\mathcal{K} (\times 10^3)$ , while the vertical coordinates are, from left to right, FPR95, AUROC, AUPR, and mAP, respectively. Red line and Blue line represent using MS-COCO2017 and OpenImages as OOD datasets, respectively. Pink line represents using PASCAL VOC as ID dataset. + +![](images/73c6ad0c18438ff92f9cc9d14ff5d7f846cf62942ac99a9b166ee318d2b2a349.jpg) +Number of sampled out-of-distribution prompts $(\times 1000)$ + +![](images/1e42fb880866069d188e5d5f77a6441d8294ddfd0716dda7e0e78d78e11b84c7.jpg) + +![](images/cfaa103478de814d41924f66e90ef2fa4c6892b36f8fa401613060180cdb0528.jpg) + +
αFPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
051.63 / 50.8887.86 / 87.2449.2
0.551.90 / 51.4887.55 / 87.0248.9
1.045.96 / 47.1089.19 / 88.4949.4
1.555.88 / 53.3386.29 / 86.7548.9
2.055.92 / 49.5486.75 / 88.0048.9
+ +COCO2017 and $47.1\%$ on OpenImages, while the mAP on PASCAL VOC improves to $49.4\%$ . This setup further reduces FPR95 by $4.57\%$ and $3.17\%$ on MS-COCO2017 and OpenImages, respectively, compared to [6]. For Berkeley DeepDrive-100k (ID), using ResNet50-based image encoder and Transformer-based text encoder, APLGOS achieves an FPR95 of $41.10\%$ on MS-COCO2017 and $23.30\%$ on OpenImages, with an mAP of $35.8\%$ . When using RegNetX4.0-based image encoder instead, FPR95 further decreases to $39.48\%$ on MS-COCO2017 and $19.79\%$ on OpenImages, while mAP improves to $37.6\%$ . + +# 4.4. Ablation Studies + +Prompt strategies. To further validate the effectiveness of our prompt strategies, we conduct extensive ablation experiments on APLGOS's prompt strategies, and the results are shown in Table 2. Sampling from the statements set brings greater performance gains than simply initializing learnable prompts with "a region of a" ((c) vs (d)). Moreover, adding location tokens to prompts significantly improves performance, as it refines the scope of the prompts ((c) vs (e)). Compared to other prompt strategies, our APLGOS prompt strategy (f) integrates the advantages of the aforementioned strategies and achieves the best performance. + +Number of Sampled OOD Prompts. APLGOS synthe + +Table 3. The Ablation Experiments on The Strength of Random Gaussian Noise $\varepsilon$ . $\alpha$ represents the strength of added gaussian noise. The value of $\alpha$ increases gradually from 0 to 2.0, and we take the value at 0.5 intervals. + +
Γ1FPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
1:450.11 / 58.3887.71 / 85.6749.1
1:349.40 / 55.1287.91 / 86.3849.2
1:247.98 / 54.4988.40 / 85.9449.2
1:145.96 / 47.1089.19 / 88.4949.4
2:148.25 / 50.2088.30 / 87.7649.2
3:150.95 / 53.9486.81 / 84.7047.5
4:150.20 / 51.5686.70 / 84.8947.3
+ +Table 4. The ablation experiments on the ratio $\Gamma_{1}$ of ID and OOD data used during training. Our default parameters and results are shown in bold. Parameters and results of baseline [6] are shown with a dark base color. + +sises virtual prompts for OOD categories and for each ID category, APLGOS samples $\kappa$ virtual OOD prompts in low-likelihood regions of ID class-conditional gaussian distributions in high-dimensional hidden space. We conduct ablation experiments on $\kappa$ , the results of its effect on performance are shown in Figure 3. When $\kappa$ is too small, it may fail to adequately cover the region outside the ID categories' decision boundaries in the hidden space. On the other hand, when $\kappa$ is too large, the excessive randomness in the sampled OOD prompts makes it difficult to effectively regularize the decision boundaries with the limited model parameters. Therefore, we set $\kappa = 10000$ as the default value. + +Strength of Random Gaussian Noise $\varepsilon$ . To enhance the size and diversity of the OOD prompts embedding sampling space and prevent the model from overly relying on the ID category distribution, we introduce a learnable matrix initialized with random Gaussian noise $\varepsilon$ during the OOD prompt sampling stage (Eq. 5). We conduct ablation experiments on its strength $\alpha$ , and the results are shown in Table 3. A small value of $\alpha$ makes the sampling space of OOD prompts embedding too narrow, while a large value of $\alpha$ results in an overly large sampling space. Only by appropriately expanding the sampling space of OOD prompts + +![](images/d14e61dc7c29443c7937d3b6931d8983489bba33a013ccb1803fbaf9b71dddfe.jpg) +Figure 4. Detection results on ID dataset. Here we use Berkeley DeepDrive-100k dataset as ID dataset. We use RegNetX4.0 and Transformer as backbone. The first row is the detection results of baseline [6]. The second row is the detection result of our APLGOS. Our APLGOS rarely misclassifies the ID class as OOD class, and there is almost no missed detection. + +![](images/d2eb827ad38b660063abcfa24ab55d7a413aa719962bb5b9fb9030164b2a1b70.jpg) +Figure 5. Detection results on OOD datasets. Here we use Berkeley DeepDrive-100k dataset as ID dataset, MS-COCO2017 and OpenImages as OOD datasets. The first row is the detection results of baseline [6]. The second row is the detection results of our APLGOS. Compared to the baseline, APLGOS rarely misses detections and hardly produces overlapping boxes for the same object. + +![](images/daf57df2e22470715408fde8cad652fbaf0191c0d40a902dbf2f45661e00ec71.jpg) +Figure 6. Detection results in Real World. Here we use Berkeley DeepDrive-100k dataset as in-distribution dataset. Pictures we take ourselves with our phone as out-of-distribution dataset. + +embedding can the model's ability to fit the OOD distribution be effectively enhanced. + +Ratio of ID and OOD Data Used During Training. To verify that APLGOS can achieve better performance with less ID data, we conduct ablation experiments on the amount of ID data used during training, and the results are shown in Table 4. By default, APLGOS adopts a ratio $\Gamma_{1}$ of 1:1 for ID and OOD data during training, whereas the baseline [6] uses a ratio of 2:1. However, in this case, the performance of APLGOS decreases instead. + +Visualization of Detection Results. To better evaluate the performance of APLGOS, we visualize its detection results on ID datasets, OOD datasets, and real-world scenarios. + +The results are presented in Figures 4, 5 and 6. The images in real-world scenarios are captured using an iPhone 14 Pro Max. The results demonstrate that APLGOS outperforms the baseline method in detecting ID and OOD categories. Moreover, real-world visualizations further demonstrate its strong generalization. + +# 5. Conclusion + +In this paper, we propose a vision-language method, Adaptive Prompt Learning via Gaussian Outlier Synthesis (APLGOS) for Out-of-distribution Detection. Through prompt learning approach, APLGOS provides adaptive region-level prompts with location information for ID / OOD images. We use ChatGPT to standardize pre-defined Q&A pairs and generate a statements set. During training, only ID images are from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual. We sample statements from the statements set to initialize learnable ID prompts. We samples virtual OOD prompts and OOD images in the low-likelihood region of the class-conditional gaussian distribution in high-dimensional hidden space. The similarity score between prompts and images is utilized to calculate contrastive learning loss in high-dimensional hidden space, which guarantees the quality of virtual outliers as well as better regularization of the model. Through comprehensive experimental evaluations, we demonstrated the effectiveness of the proposed APLGOS. + +# Acknowledgements + +This paper is supported by Zhongguancun Laboratory and the National Natural Science Foundation of China (Grant No. 62272018). + +# References + +[1] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877-1901, 2020. 2 +[3] L. Chen, G. Wang, L. Yuan, K. Wang, K. Deng, and P. Torr. Nerf-vpt: Learning novel view representations with neural radiance fields via view prompt tuning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1156-1164, 2024. 2 +[4] J. Devlin, M. Chang, K. Lee, and K. Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[5] X. Du, Z Wang, M. Cai, and S Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2022. 1 +[6] X. Du, Z. Wang, M. Cai, and Y. Li. Vos: Learning what you don't know by virtual outlier synthesis. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7, 8 +[7] M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 5 +[8] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context learning for 3d point cloud understanding. Advances in Neural Information Processing Systems, 36, 2024. 2 +[9] M. Grcic, P. Bevandić, and S. Šegvic. Dense open-set recognition with synthetic outliers generated by real nvp. arXiv preprint arXiv:2011.11094, 2020. 2 +[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5 +[11] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR), 2016. 6 +[12] D Hendrycks, M. Mazeika, and T. Dietterich. Deep anomaly detection with outlier exposure. International Conference on Learning Representations (ICLR), 2018. 1, 4 +[13] D. Hendrycks, S. Basart, M. Mazeika, A. Zou, J. Kwon, M. Mostajabi, J. Steinhardt, and D. Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 4 + +[14] Y. Hsu, Y. Shen, H. Jin, and Z. Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10951-10960, 2020. 2, 6 +[15] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV), 128 (7):1956-1981, 2020. 5 +[16] T. Le Scao, A. Fan, C. Akiki, E. Pavlick, S. Ilic, D. Hesslow, R. Castagné, A. Luccioni, F. Yvon, M. Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. 2023. 2 +[17] Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2 +[18] K. Lee, H. Lee, K. Lee, and J. Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. International Conference on Learning Representations (ICLR), 2017. 2, 6 +[19] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems (NeurIPS), 31, 2018. 6 +[20] D. Li, J. Li, and S. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1 +[21] J. Li, P. Chen, Z. He, S. Yu, S. Liu, and J. Jia. Rethinking out-of-distribution (OOD) detection: Masked image modeling is all you need. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11578-11589, 2023. 1 +[22] K. Li, Q. Geng, M. Wan, X. Cao, and Z. Zhou. Context and spatial feature calibration for real-time semantic segmentation. IEEE Transactions on Image Processing (TIP), 32: 5465-5477, 2023. 1 +[23] S. Liang, Y. Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. International Conference on Learning Representations (ICLR), 2017. 2, 6 +[24] J. Liao, X. Xu, M. Nguyen, A. Goodge, and C. Foo. Coftad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing (TIP), 2024. 1 +[25] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 5 +[26] W. Liu, X. Wang, J. Owens, and Y. Li. Energy-based out-of-distribution detection. Advances in neural information processing systems (NeurIPS), 33:21464-21475, 2020. 2, 6 +[27] X. Liu, Y. Lochman, and C. Zach. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23946-23955, 2023. 2 + +[28] Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems (NeurIPS), 35:35087-35102, 2022. 2 +[29] S. Mohseni, M. Pitale, J. Yadawa, and Z. Wang. Self-supervised learning for generalizable out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 5216-5223, 2020. 2 +[30] S. Ning, L. Qiu, Y. Liu, and X. He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2 +[31] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763, 2021. 2, 6 +[32] I. Radosavovic, R. Kosaraju, R. Girshick, K. He, and P. Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10428-10436, 2020. 5 +[33] C. Sastry and S. Oore. Detecting out-of-distribution examples with gram matrices. In International Conference on Machine Learning (ICML), pages 8491-8501, 2020. 1, 6 +[34] B. Su, H. Zhang, and Z. Zhou. Hsic-based moving weight averaging for few-shot open-set object detection. In Proceedings of the 31st ACM International Conference on Multimedia (ACMM), pages 5358-5369, 2023. 1, 4 +[35] B. Su, H. Zhang, J. Li, and Z. Zhou. Toward generalized few-shot open-set object detection. IEEE Transactions on Image Processing (TIP), 33:1389-1402, 2024. +[36] J. Tack, S. Mo, J. Jeong, and J. Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems (NeurIPS), 33:11839-11852, 2020. 1, 2, 4, 6 +[37] H. Wang, Y. Li, H. Yao, and X. Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1802-1812, 2023. 2 +[38] A. Wu, C. Deng, and W. Liu. Unsupervised out-of-distribution object detection via pca-driven dynamic prototype enhancement. IEEE Transactions on Image Processing (TIP), 2024. 6 +[39] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7031-7040, 2023. 2 +[40] C. Xie, Z. Zhang, Y. Wu, F. Zhu, R. Zhao, and S. Liang. Described object detection: Liberating object detection with flexible expressions. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1 +[41] X. Xiong, Y. Liu, T. Yuan, Y. Wang, Y. Wang, and H. Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17535-17544, 2023. 1 + +[42] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1 +[43] Z. Yang, Z. Gan, J. Wang, X. Hu, F. Ahmed, Z. Liu, Y. Lu, and L. Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision (ECCV), pages 521-539. Springer, 2022. 5 +[44] F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2636-2645, 2020. 5 +[45] Y. Zhai, Y. Zeng, Z. Huang, Z. Qin, X. Jin, and D. Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6979–6987, 2024. 2 +[46] C. Zhang, X. Chen, S. Chai, C. Wu, D. Lagun, T. Beeler, and F. De la Torre. Iti-gen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3969-3980, 2023. 2 +[47] J. Zhang, N. Inkawich, Y. Chen, and H. Li. Fine-grained out-of-distribution detection with mixup outlier exposure. arXiv preprint arXiv:2106.03917, 2(5), 2021. 2 +[48] J. Zhang, J. Huang, S. Jin, and S. Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2 +[49] K. Zhou, J. Yang, C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16816-16825, 2022. 2, 5 +[50] K. Zhou, J. Yang, C. Loy, and Z. Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 2, 5 +[51] Z. Zhu, Y. Zhang, H. Chen, Y. Dong, S. Zhao, W. Ding, J. Zhong, and S. Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21600-21610, 2023. 1 +[52] O. Zohar, K. Wang, and S. Yeung. Prob: Probabilistic objectness for open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11444-11453, 2023. 1 \ No newline at end of file diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/images.zip b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..357181eeafff7525e5f1f491c742439d0c18867f --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8df233ff96aa477991b0021cb80e667bef13d71d4c52dd0a84e16aecd9ae011 +size 819622 diff --git a/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/layout.json b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab79c904226b34bdc19d45560d3730021e4605cc --- /dev/null +++ b/2025/Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection/layout.json @@ -0,0 +1,9295 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "text", + "content": "Adaptive Prompt Learning via Gaussian Outlier Synthesis for Out-of-distribution Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "text", + "content": "Yongkang Zhang" + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "text", + "content": " Dongyu She" + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "text", + "content": " Zhong Zhou" + }, + { + "bbox": [ + 164, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "spans": [ + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, China " + }, + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 175, + 531, + 204 + ], + "type": "text", + "content": "Zhongguancun Laboratory, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 190, + 205, + 417, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 205, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 190, + 205, + 417, + 217 + ], + "type": "text", + "content": "{zyk，zz}@buaa.edu.cn，shedy@zgclab.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 534 + ], + "type": "text", + "content": "Out-of-distribution (OOD) detection aims to distinguish whether detected objects belong to known categories or not. Existing methods extract OOD samples from Indistribution (ID) data to regularize the model's decision boundaries. However, the decision boundaries are not adequately regularized because the model does not have sufficient knowledge about the distribution of OOD data. To address the above issue, we propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis (APL-GOS) for OOD detection. Specifically, we leverage the Vision-Language Model (VLM) to initialize learnable ID prompts by sampling standardized results from pre-defined Q&A pairs. Region-level prompts are synthesised in low-likelihood regions of class-conditional gaussian distributions. These prompts are then utilized to initialize learnable OOD prompts and optimized with adaptive prompt learning. Also, OOD pseudo-samples are synthesised via gaussian outlier synthesis. The aforementioned methodology regularizes the model to learn more compact decision boundaries for ID and OOD categories. Extensive experiments show that APLGOS achieves state-of-the-art performance with less ID data on four mainstream datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 696 + ], + "type": "text", + "content": "Deep learning has made significant progress in recent years. It encompasses a multitude of research domains, including object detection [22, 40, 52], autonomous driving [41, 51] and image generation [20, 42]. Various existing deep learning methods rely on large-scale datasets to regularize the model, enabling it to learn sufficient data distribution and supervision signals of the training data. In real-world scenarios, where the number of unknown categories is significantly greater than that in the training dataset, the model lacks knowledge about the distribution of unknown data in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 244, + 554, + 417 + ], + "blocks": [ + { + "bbox": [ + 316, + 244, + 554, + 417 + ], + "lines": [ + { + "bbox": [ + 316, + 244, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 244, + 554, + 417 + ], + "type": "image", + "image_path": "a8e0bda1e5603869afc5b7406559ead6efef0d0508171576b86f37136ca38d19.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 481 + ], + "type": "text", + "content": "Figure 1. Quantitative comparisons with state-of-the-art OOD detection methods in terms of FPR95, AUROC and mAP metrics. Note that larger points denote higher mAP, and the numerical values are also given next to each point. Our APLGOS provides remarkable performance boost on all the metrics." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 570 + ], + "type": "text", + "content": "practical applications and struggles to learn compact decision boundaries that effectively distinguish between known and unknown categories. During the testing phase, unknown categories are likely to result in erroneous predictions accompanied by a high confidence score. This leads to severe safety risks in critical safety domains such as autonomous driving." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "OD detection [5, 21, 24, 33] is a research hotspot in recent years, which aims to enable the detectors to accurately distinguish not only seen categories, but also unseen categories during training. The detectors need to learn compact decision boundaries during training, ensuring low uncertainty for ID categories while maintaining high uncertainty away from them. To achieve this, existing OOD detection methods [12, 13, 34-36] provide sufficient supervision of OOD data for model training by extracting OOD pseudo-samples from ID data, helping the model better distinguish between known and unknown categories. However, due to the unpredictable quality of OOD pseudo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3235" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "samples extracted from the ID data and the requirement for a large volume of ID data, the detector is not adequately regularized to learn compact decision boundaries for both ID and OOD categories. Therefore, synthesis-based methods [6, 18] have been proposed to generate OOD pseudo-samples. They synthesize out-of-distribution RGB images directly or virtual outliers in lower-dimensional hidden space, which to some extent mitigate the limitations of extracting pseudo-samples from ID data. In recent years, Vision-Language Models (VLMs) [30, 31, 48], owing to their powerful pre-trained knowledge and representation capabilities, have achieved considerable success and been applied across numerous fields." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "text", + "content": "In this paper, we propose APLGOS, a synthesis-based vision-language method that leverages the powerful learning and representation capabilities of VLMs to assist in synthesizing virtual outliers using ID data. APLGOS mainly consists of Prompt Learning Module (PLM) and Text-Image Alignment Module (TAM). PLM employs two distinct strategies to generate ID prompts and OOD pseudo-prompts, respectively, to assist in regularizing the model. For ID data, we first provide a predefined Q&A pair and templates with location and category names, e.g., “Q: What is in the region with coordinates " + }, + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "inline_equation", + "content": ", , , " + }, + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "text", + "content": "? A: That’s a " + }, + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 56, + 231, + 296, + 564 + ], + "type": "text", + "content": "”, guiding the detector to incorporate location coordinates for more fine-grained observation. We guide ChatGPT through multiple rounds of standardization for the aforementioned prompts to generate a set of statements for the model to sample during training. The statements sampled from this set are then directly used to initialize the learnable ID prompts. In order to ensure that the generated OOD pseudo-samples better fit the distribution of OOD data, PLM generates OOD prompts through adaptive prompt learning via Gaussian Outlier Synthesis, where it samples virtual OOD prompts in the low-likelihood region of the class-conditional Gaussian distribution of ID prompts in high-dimensional hidden space. TAM calculates similarity scores for images and prompts and combines contrastive learning to align multimodal data, thereby regularizing model's decision boundaries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 568, + 282, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 568, + 282, + 579 + ], + "spans": [ + { + "bbox": [ + 67, + 568, + 282, + 579 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 582, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 582, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 653 + ], + "type": "text", + "content": "- We propose a vision-language OOD detection model namely APLGOS. Through adaptive prompt learning, APLGOS generates adaptive region-level prompts for ID and OOD images. Based on contrastive learning, APLGOS calculates similarity for images and prompts to ensure model learn compact decision boundaries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 296, + 714 + ], + "type": "text", + "content": "- ID prompts, OOD prompts and OOD images are all virtual. ChatGPT standardizes pre-defined Q&A pairs with templates and instructions. Then we sample them to initialize learnable ID prompts. We synthesise virtual OOD prompts and OOD images in low-likelihood regions of" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 483, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 483, + 83 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 483, + 83 + ], + "type": "text", + "content": "class-conditional Gaussian distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "text", + "content": "- Extensive experiments on mainstream datasets show that APLGOS achieves state-of-the-art performance in terms of FPR95, AUROC, AUPR and mAP metrics. Compared to the baseline method [6], when using Berkley DeepDrive-100k as ID dataset and OpenImages as OOD dataset, our method reduces FPR95 by " + }, + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "inline_equation", + "content": "7.76\\%" + }, + { + "bbox": [ + 314, + 84, + 555, + 156 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 172, + 400, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 172, + 400, + 185 + ], + "spans": [ + { + "bbox": [ + 314, + 172, + 400, + 185 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 194, + 474, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 194, + 474, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 474, + 206 + ], + "type": "text", + "content": "2.1. Out-of-distribution Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 214, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 490 + ], + "type": "text", + "content": "OOD detection [14, 26, 36, 39] aims to learn a compact decision boundary on training data that allows model to detect not only the categories with low uncertainty, that have been seen in training phase, but also the unseen categories with high uncertainty. Since in physical world, the number of unseen categories for the model is much bigger than seen categories, using large-scale dataset to regularize the model [13, 27] is difficult to fully cover all unseen categories of physical world. Liang et al. [23] use temperature scaling and add small perturbations to the input to separate the softmax score distributions between ID and OOD images. Based on energy theory [17], the work [26] replace traditional softmax score with energy score to distinguish ID and OOD images. Recently, outlier based methods are proposed, which utilize outliers exposure [29, 47] or generate virtual outliers in pixel [9, 18] space or hidden feature space [6] to regularize the model. Nevertheless, they are inefficient and the quality of the synthesised virtual outliers is worrying. With the emergence of vision-language models, vision-language model-based methods are proposed to address open-vocabulary problems [28, 37, 39]. To the best of our knowledge, no prior work has explored the use of prompt learning in OOD detection task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 502, + 419, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 419, + 515 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 419, + 515 + ], + "type": "text", + "content": "2.2. Prompt Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Prompt learning is to view pre-trained language models, such as BERT [4], GPT [1, 2] and BLOOM [16] as knowledge bases, and use them to provide text prompts to optimize the performance of downstream tasks. In contrast to hand-designed prompts, the goal of prompt learning is to adaptively provide accurate prompts for downstream tasks. Zhou et al. [50] propose CoOp, which models a prompt's context words with learnable vectors while keeping pre-trained parameters fixed. To prevent CoOp from overfitting base classes, Zhou et al. [49] introduce CoCoOp, which uses conditional context optimization to generate an input-conditional token for each image, but this approach introduces high computational costs. At the same time, due to the effectiveness of prompt learning, there are various methods incorporating it with computer vision tasks [3, 8, 45, 46]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3236" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 289 + ], + "type": "image", + "image_path": "bbe25e5b8f3e014da3260cc976908de2bfce9293fac31efebbfba0d13ed9f37e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. The proposed APLGOS network architecture. Prompt learning module is responsible for using ChatGPT to standardize Q&A pairs with guidance introduction and templates, then it generates a statements set. The module samples prompts from the statements set to initialize the learnable ID prompts, and synthesises virtual OOD prompts in low-likelihood regions of class-conditional gaussian distributions. The Text-Image Alignment Module computes similarity scores to align text and image embeddings in the hidden space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 353, + 138, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 138, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 138, + 367 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 376, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 616 + ], + "type": "text", + "content": "We propose an Adaptive Prompt Learning framework via Gaussian Outlier Synthesis for OOD Detection. As shown in Figure 2, APLGOS mainly consists of two modules, i.e. PLM and TAM. PLM leverages ChatGPT to standardize pre-defined Q&A pairs using guidance instructions and predefined templates, generating a set of statements. During training, PLM samples statements from this set to initialize the learnable prompts. For ID categories, APLGOS directly employs the initialized prompts as input to the text encoder, whereas for OOD categories, it synthesizes virtual OOD prompts and images within the low-likelihood region of the class-conditional Gaussian distribution of ID classes in the hidden space. Notably, only ID images are sourced from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual and synthesized. This approach enables the model to enhance the quality of pseudo-samples with less ID data while better capturing the distribution of OOD data. Additionally, through contrastive learning, TAM computes similarity scores to align images and prompts within the high-dimensional hidden space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "For clarity, we omit the batchsize of data in the following description and consider a single batch as an example. The input to APLGOS consists of two modalities: detected region images " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "[\\mathbf{X}_1,\\mathbf{X}_2,\\dots,\\mathbf{X}_b]" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " extracted from a raw RGB image " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", and text prompts " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\in \\mathbb{R}^{b\\times l}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " denote the number of channels, height, and width of the image, respectively. " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " represents the number of detected region images from a single raw RGB image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "type": "text", + "content": " indicates the length of the text prompts. The text input is given as " + }, + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = [\\mathbf{T}_1, \\mathbf{T}_2, \\dots, \\mathbf{T}_b]" + }, + { + "bbox": [ + 313, + 354, + 555, + 392 + ], + "type": "text", + "content": ". token in the sampled prompts has been replaced with the corresponding labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 396, + 458, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 458, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 458, + 409 + ], + "type": "text", + "content": "3.1. Prompt Learning Module" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "text", + "content": "ID Prompts. To enhance the model's representation ability and more effectively regularize its decision boundaries, we generate a set of statements for the Prompt Learning Module to sample from, rather than using a single invariant statement to initialize the learnable ID prompts. Specifically, we first predefine a Q&A pair, such as “Q: What is in the region with coordinates " + }, + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "inline_equation", + "content": ", , , " + }, + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "text", + "content": "? A: That's a " + }, + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 414, + 555, + 571 + ], + "type": "text", + "content": ".” We then input this Q&A pair into ChatGPT for standardization. During this process, we provide predefined templates and guiding instructions to ensure that ChatGPT standardizes the Q&A pair accordingly. The standardization process is illustrated below with an example prompt:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 330, + 586, + 555, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 586, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 330, + 586, + 555, + 601 + ], + "type": "interline_equation", + "content": "\\Omega_ {0} = g \\left(\\mathrm {Q} ^ {\\mathrm {A}} + \\mathrm {M} + \\mathrm {G} _ {0}\\right), \\quad \\Omega_ {i} = g \\left(\\Omega_ {i - 1} + \\mathrm {G} _ {i}\\right), \\tag {1}", + "image_path": "0e4bdbcdbd055ebd2665048a06ac92e7f9ba129916df572a8001875496799c9d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\Omega_{i}" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " denotes generated prompt result in " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " round, " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{Q^A}" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " denotes Q&A pair, M denotes predefined template, " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{G}_i" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " denotes guidance instruction for " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " standardizing round and " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " is ChatGPT's standardizing operation. We collect the statements from these " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": " rounds to obtain statements set " + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\Omega_t" + }, + { + "bbox": [ + 313, + 605, + 553, + 689 + ], + "type": "text", + "content": ". These statements are then used for sampling during the initialization of learnable ID prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We introduce no extra character sets and vocabularies, and the generated prompts are represented in natural lan" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3237" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "guage. The learnable prompts follow the paradigm e.g. " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle < V_1 > < V_2 > \\dots < V_m >" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle CLS\\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", which is initialized by sampled prompt. " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle loc_1\\rangle < \\langle loc_2\\rangle < \\langle loc_3\\rangle < \\langle loc_4\\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are learnable location tokens, which implicitly introduce location information into the prompts. " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle V_1\\rangle < \\langle V_2\\rangle \\dots < V_m\\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are learnable description tokens, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is its length. " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle CLS\\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is class token." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 168, + 295, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 168, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 111, + 168, + 295, + 183 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {T}} = f _ {\\theta} \\left(h \\left(r \\left(g \\left(\\Omega_ {t - 1} + \\mathrm {G} _ {t}\\right)\\right)\\right)\\right), \\tag {2}", + "image_path": "dddfade8a161046558a2eb3e21c924db42eb0205c4b5d2e442b82fdd7a5e3a6f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}} = [\\hat{\\mathbf{T}}_1,\\hat{\\mathbf{T}}_2,\\dots ,\\hat{\\mathbf{T}}_b],\\hat{\\mathbf{T}}_i\\in \\mathbb{R}^{l'},t" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " is rounds of standardizing operations, " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "l^{\\prime}" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " is length of prompt embedding. Here, for ease of understanding, we use one " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}_i" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " as an example to describe the subsequent operations, and standardize " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}_i" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}},\\hat{\\mathbf{T}}\\in \\mathbb{R}^{l'}" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " is transformer-based text encoder, " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " is tokenizer, " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " is replacement function for " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{CLS}\\rangle" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " token. We replace " + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\langle \\mathrm{CLS}\\rangle" + }, + { + "bbox": [ + 55, + 190, + 296, + 291 + ], + "type": "text", + "content": " with the category label of the object in the current region (i.e., the corresponding ID class label)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 487 + ], + "type": "text", + "content": "OOD Prompts. In the hidden space, distinct decision boundaries should be established between ID and OOD prompts. In the OOD detection task, we refine the decision boundaries as much as possible. By incorporating prompt learning, we synthesize region-level OOD pseudoprompts using Gaussian outlier synthesis. Specifically, the Prompt Learning Module synthesizes virtual OOD prompts in the low-likelihood regions of class-conditional Gaussian distributions in hidden space. This allows the Text-Image Alignment Module to perceive the distribution difference between ID and OOD categories in hidden space and align images and prompts through contrastive learning. Provided that the quantity of data is large enough, we assume the ID prompts embedding from the text encoder form a class-conditional multivariate Gaussian distribution:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 500, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 500, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 122, + 500, + 295, + 514 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (\\hat {\\mathbf {T}} | y = i) = \\mathcal {N} (\\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {3}", + "image_path": "ade71e8337beada10b64783a2ce64b393c746f1730185e2e8ea04e7f2e4bc1d3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " is the parameter of text encoder " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "f_{\\theta},y" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " is ground truth label, " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}_i" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " is empirical gaussian mean of " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " in-distribution category prompts embedding, and " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " represents the number of in-distribution classes, " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\hat{\\mu}_i,\\hat{\\sigma}) = \\frac{1}{\\sqrt{2\\pi\\hat{\\sigma}}} e^{-\\frac{(\\hat{\\Gamma} - \\hat{\\mu}_i)^2}{2\\hat{\\sigma}^2}}" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 55, + 521, + 295, + 589 + ], + "type": "text", + "content": " denotes the tied covariance matrix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": "First, we calculate the empirical gaussian mean of " + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 589, + 295, + 612 + ], + "type": "text", + "content": " ID category prompts embedding as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 623, + 295, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 623, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 131, + 623, + 295, + 658 + ], + "type": "interline_equation", + "content": "\\hat {\\mu} _ {i} = \\frac {1}{| \\mathcal {Q} _ {T} |} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\hat {\\mathbf {T}} _ {i, j}, \\tag {4}", + "image_path": "ecc143bbb3e6683541cffda2e127dc21074426b13056c7407895e60ba1f6bc8b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_T|" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " denotes the length of the prompts queue " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_T" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " used to buffer ID prompts, and " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_T\\in \\mathbb{R}^{K\\times |\\mathcal{Q}_T|}" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Then we calculate the tied covariance matrix of ID prompts embedding as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 553, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 137 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} = \\frac {1}{K | \\mathcal {Q} _ {T} |} \\sum_ {i = 1} ^ {K} \\sum_ {j = 1} ^ {| \\mathcal {Q} _ {T} |} \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) \\left(\\hat {\\mathbf {T}} _ {i, j} + \\alpha \\varepsilon - \\hat {\\mu} _ {i}\\right) ^ {T} + \\beta \\mathbf {E}, \\tag {5}", + "image_path": "9ec9050130ea0ce598834900b1f3709d7786f614064fff8c042f4226d3a9e543.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": " is learnable matrix initialized by randomly gaussian noise, " + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": " is unit matrix, " + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta" + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": " are hyper-parameters, " + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": " is tied covariance matrix, and " + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma} = [\\hat{\\sigma}_1, \\hat{\\sigma}_2, \\dots, \\hat{\\sigma}_K]^T" + }, + { + "bbox": [ + 313, + 137, + 553, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "text", + "content": "After computing the empirical Gaussian mean " + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\hat{\\mu}" + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "text", + "content": " and the tied covariance matrix " + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "text", + "content": ", the Prompt Learning Module samples virtual OOD prompts from the low-likelihood regions of the class-conditional Gaussian distributions in hidden space, based on the estimated multivariate distributions. Then, it selects the top-k prompts with the lowest probability from this " + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 173, + 555, + 256 + ], + "type": "text", + "content": "-likelihood region:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 396, + 268, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 268, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 396, + 268, + 553, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {V} _ {i} = \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}), \\tag {6}", + "image_path": "f00641e6165f0bfb45faf989eeed6bcb4e788184485a81a32a0545c8a76e64b7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 285, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 285, + 553, + 310 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 285, + 553, + 310 + ], + "type": "text", + "content": " is class-conditional gaussian distribution probability density and satisfies the following relation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 328, + 553, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 328, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 324, + 328, + 553, + 370 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Psi (\\hat {\\mathbf {T}}, \\hat {\\boldsymbol {\\mu}} _ {1}, \\hat {\\boldsymbol {\\mu}} _ {2}, \\dots , \\hat {\\boldsymbol {\\mu}} _ {K}, \\hat {\\boldsymbol {\\sigma}}) = \\\\ \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {1}, \\hat {\\sigma}) \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {2}, \\hat {\\sigma}) \\dots \\Psi (\\hat {\\mathbf {T}}, \\hat {\\mu} _ {K}, \\hat {\\sigma}), \\tag {7} \\\\ \\end{array}", + "image_path": "77af5af05502f0e2ba93873eab66c8a4e94980906743257bb0e81200693144c1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 370, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 553, + 393 + ], + "type": "text", + "content": "For each " + }, + { + "bbox": [ + 313, + 370, + 553, + 393 + ], + "type": "inline_equation", + "content": "\\Psi (\\hat{\\mathbf{T}},\\hat{\\mu}_i,\\hat{\\sigma})" + }, + { + "bbox": [ + 313, + 370, + 553, + 393 + ], + "type": "text", + "content": " , its expansion can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 400, + 555, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 400, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 314, + 400, + 555, + 442 + ], + "type": "interline_equation", + "content": "\\Psi \\left(\\hat {\\mathbf {T}}, \\hat {\\mu} _ {i}, \\hat {\\sigma}\\right) = \\left\\{v _ {i} \\mid \\frac {1}{\\sqrt {2 \\pi^ {\\frac {l ^ {\\prime}}{2}}} \\left| \\hat {\\sigma} \\right| ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right) ^ {T} \\hat {\\sigma} ^ {-} \\left(v _ {i} - \\hat {\\mu} _ {i}\\right)} < \\epsilon \\right\\}, \\tag {8}", + "image_path": "41fcee3901a183bef428d2c7e11db818cc4d6fc60cac5d745eb8b4400accbca8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "inline_equation", + "content": "v_{i} \\sim \\mathcal{N}(\\hat{\\mu}_{i},\\hat{\\sigma})" + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "content": " denotes sampled virtual prompt using " + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "content": " ID category prompts, " + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "inline_equation", + "content": "i = \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "content": ", and “-” denotes matrix inverse operation. The final synthesised OOD prompts are denoted as " + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}^{\\dagger}" + }, + { + "bbox": [ + 313, + 442, + 554, + 494 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 500, + 481, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 481, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 481, + 514 + ], + "type": "text", + "content": "3.2. OOD Virtual Images Synthesis" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "text", + "content": "Existing methods [12, 13, 34-36] directly extract OOD pseudo-samples from ID data. However, the extracted pseudo-samples are unable to fit the distribution of OOD data adequately. We also use synthesis method to get OOD data. The principle of synthesizing OOD image is similar to Eq. 3 to Eq. 8. Compared with synthesizing OOD prompts, the input for calculating empirical Gaussian mean and tied covariance is ID image embedding instead of ID prompts embedding. We define the final synthesised virtual images using current ID image embedding queue " + }, + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}_I" + }, + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\dagger}" + }, + { + "bbox": [ + 313, + 517, + 555, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 647, + 481, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 481, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 481, + 661 + ], + "type": "text", + "content": "3.3. Text-Image Alignment Module" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 712 + ], + "type": "text", + "content": "We first encode ID and OOD images and prompts to generate their embeddings. Then, the similarity score between prompts embedding and image embeddings is computed as follows:" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3238" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 132, + 81, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 81, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 132, + 81, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\frac {\\left| \\left| \\hat {\\mathbf {X}} \\right| \\right| _ {p} \\left(\\left| \\left| \\hat {\\mathbf {T}} \\right| \\right| _ {p}\\right) ^ {T}}{e ^ {\\omega}}, \\tag {9}", + "image_path": "441664764a82dfc5190d6d303a23a37ef509753c0bbfc06dac2e017e34ada5f4.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " is the embedding of detected region images in the second training phase, and " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = [\\hat{\\mathbf{X}}_1,\\hat{\\mathbf{X}}_2,\\dots ,\\hat{\\mathbf{X}}_b],\\hat{\\mathbf{X}}_i" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " is one detected region image embedding, " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_i\\in \\mathbb{R}^{l'}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " . In the third training phase, the input is embedding of synthesised virtual image " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}^{\\dagger}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}},\\omega" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " is hyper-parameters for scaling. " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " is similarity score. The prompts embedding in Eq. 9 is ID prompts embedding " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " in the second phase and synthesised OOD prompts embedding " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{T}}^{\\dagger}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " in the third phase, " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "||\\cdot ||_{p}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " is normalization, in addition, " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "||\\hat{\\mathbf{X}}_i||_p = \\hat{\\mathbf{X}}_i / \\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{X}}_{i,j}|^2}" + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 113, + 296, + 258 + ], + "type": "inline_equation", + "content": "||\\hat{\\mathbf{T}}_i^\\dagger ||_p = \\hat{\\mathbf{T}}_i^\\dagger /\\sqrt{\\sum_{j = 1}^{l'}|\\hat{\\mathbf{T}}_{i,j}^\\dagger|^2}" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 144, + 275 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": "Alignment loss " + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {align }}" + }, + { + "bbox": [ + 55, + 281, + 296, + 342 + ], + "type": "text", + "content": " constrains the contrastive learning process during alignment, receiving ID or OOD data at different training phases. The similarity score between prompts embedding and image embeddings is used to calculate the alignment loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 351, + 295, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 351, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 102, + 351, + 295, + 386 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} (\\mathbf {S}, y) = - \\sum_ {i = 1} ^ {K ^ {\\prime}} t _ {i} \\log \\left(\\mathcal {R} _ {i} (\\mathbf {S})\\right), \\tag {10}", + "image_path": "7a4c7ef631dff7afe247eb929c1852189c5baab094570685b4cb53a041854bfc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " represents the category label of the object contained in the currently detected region. " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_i" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " represents the standardized prediction score. We treat all OOD categories as a single category, i.e., \"background\". During the training phase, if the ID dataset contains a total of " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " classes, each detected region image is required to calculate similarity scores with " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "(K + 1)" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": " text prompts, i.e., " + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "inline_equation", + "content": "K' = K + 1" + }, + { + "bbox": [ + 55, + 392, + 296, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "text", + "content": "Previous methods typically generate simple prompts that lack location information, such as \"a photo of " + }, + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "inline_equation", + "content": "a < CLS >" + }, + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "text", + "content": "\" [49, 50], or provide brief prompts with relative location information for the entire image [43]. We argue that these prompts lack the fine granularity needed for the model to learn essential location information in vision-language-based detection tasks. " + }, + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{loc}" + }, + { + "bbox": [ + 55, + 487, + 296, + 594 + ], + "type": "text", + "content": " is designed to implicitly incorporate location information, enabling the generation of fine-grained prompts for detected image regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 614, + 295, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 614, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 78, + 614, + 295, + 644 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l o c} = \\frac {\\lambda}{\\Phi (\\mathbf {B} _ {g})} \\left[ \\sum_ {i = 1} ^ {z} \\left(\\sqrt {\\mathbf {B} _ {g}} _ {i} - \\sqrt {u \\left(\\mathbf {B} _ {r}\\right) _ {i}}\\right) ^ {2} \\right] ^ {\\frac {1}{2}}, \\tag {11}", + "image_path": "e4102b3ce2be0112284377cbbfcded574ae0d91ab6170e5228ac2aadb0361714.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_g" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents the ground truth coordinates of detected image region, " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_r" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents the regression results of coordinates, and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_g\\in \\mathbb{R}^{b\\times 4}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_r\\in \\mathbb{R}^{b\\times 4}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "z = 4" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents calculating absolute values, " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " represents calculating the dimension of vector, " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is hyper-parameter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "After incorporating the classification loss " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and the location loss " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{loc}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", the total loss can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 356, + 105, + 553, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 105, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 356, + 105, + 553, + 152 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\xi_ {1} \\left[ \\gamma_ {1} \\tau \\mathcal {L} _ {\\text {a l i g n}} ^ {i d} + \\gamma_ {2} (1 - \\tau) \\mathcal {L} _ {\\text {a l i g n}} ^ {o o d} \\right] \\\\ + \\gamma_ {3} \\xi_ {2} \\left[ \\kappa \\mathcal {L} _ {l o c} ^ {i d} + (1 - \\kappa) \\mathcal {L} _ {l o c} ^ {o o d} \\right] \\tag {12} \\\\ + \\gamma_ {4} \\xi_ {3} \\mathcal {L} _ {c l s} + \\gamma_ {5} \\xi_ {4} \\mathcal {L} _ {r e g} + \\overline {{\\mathscr {W}}}. \\\\ \\end{array}", + "image_path": "63c72531532b19fe023b662458125a0a7271f70f14d7e28e3e6e200572ed9054.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\gamma_{1},\\gamma_{2},\\gamma_{3},\\gamma_{4},\\gamma_{5}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " are the hyper-parameters, " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\xi ,\\tau ,\\kappa" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " determine the loss functions used in the current training phase and " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\xi_{i} = \\{0,1\\}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\tau = \\{0,1\\}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " . In order to better regularize the model, in the actual implementation of " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " we also add the regularization term " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_i = [\\Delta^{2}_{(\\mathcal{F}(\\mathbf{O}_1),\\mathcal{B}_1)_i} + \\Delta^{2}_{(\\mathcal{G}(\\mathbf{O}_2),\\mathcal{B}_2)_i}],\\overline{\\mathcal{W}} = 1 / N\\sum_{i = 1}^{N}\\mathcal{W}_i,\\Delta^{2}_{(a,b)}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "(a - b)^2" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{F},\\mathcal{G}" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " represent regression blocks, " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_i" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " represents regularization matrix, " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "text", + "content": " represents bias matrix of regression block, " + }, + { + "bbox": [ + 313, + 160, + 555, + 270 + ], + "type": "inline_equation", + "content": "i = \\{1,2\\}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 395, + 294 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 300, + 376, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 376, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 376, + 312 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "text", + "content": "We verify our proposed APLGOS on four commonly used datasets: PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017 and OpenImages. The PASCAL VOC [7] dataset contains 9963 images in 20 categories, split into 5011 training and 4952 test images, with a resolution of " + }, + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "inline_equation", + "content": "500 \\times 375" + }, + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "inline_equation", + "content": "375 \\times 500" + }, + { + "bbox": [ + 313, + 317, + 555, + 498 + ], + "type": "text", + "content": "). The BDD-100k [44] dataset consists of 100,000 high-resolution driving scenarios with detailed road object annotations. The MS-COCO2017 [25] dataset includes 328,000 images across 91 categories and 2.5 million instance tags, with 82 categories having more than 5000 tags. OpenImages V4 [15] contains 9.2 million images across 500 categories, commonly used for classification, object detection, and visual relationship detection. The above four datasets comprehensively evaluate our proposed method from different aspects and perspectives." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 505, + 447, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 447, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 447, + 517 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "We use a transformer-based text encoder in the Prompt Learning Module, we employ ResNet50 [10] and RegNetX4.0 [32] as backbones, respectively. We use ChatGPT-3.5 to standardize Q&A pairs. The ratio of ID data used for training to synthesised OOD data is approximately 1:1. We use PASCAL VOC and Berkeley DeepDrive-100K as ID datasets, and evaluate on two OOD datasets containing subsets randomly sampled from MS-COCO2017 and OpenImages, respectively. To ensure the fairness of the test, we manually exclude categories in the OOD dataset that overlap with those in the ID dataset before evaluating on the OOD dataset. We set " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "B = 16" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " and train APLGOS on PASCAL VOC for 18,000 iterations, and set " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "B = 8" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " to train on Berkeley DeepDrive-100k for 90,000 iterations. We set the learning rate " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "lr = 0.01" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". The length of prompt embedding and length of image embedding " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "l' = 1024" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". We use 1000" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3239" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 547, + 352 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 547, + 352 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 547, + 352 + ], + "type": "table", + "html": "
ID DatasetMethodFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
PASCAL VOCMSP [11]70.99 / 73.1383.45 / 81.91-48.7
ODIN [23]59.82 / 63.1482.20 / 82.59-48.7
Mahalanobis [19]96.46 / 96.2759.25 / 57.42-48.7
Energy score [26]56.89 / 58.6983.69 / 82.98-48.7
Gram matrices [33]62.75 / 67.4279.88 / 77.62-48.7
Generalized ODIN [14]59.57 / 70.2883.12 / 79.23-48.1
CSI [36]59.91 / 57.4181.83 / 82.95-48.1
GAN-synthesis [18]60.93 / 59.9783.67 / 82.67-48.5
VOS-ResNet50* [6]48.28 / 52.1487.65 / 85.398.76 / 96.9847.8
VOS-RegX4.0* [6]50.53 / 50.2788.10 / 87.0898.92 / 97.8049.1
APLGOS (ResNet50)47.16 / 49.6687.89 / 85.9198.80 / 97.5448.8
APLGOS (RegNetX4.0)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
Berkeley DeepDrive-100kMSP [11]80.94 / 79.0475.87 / 77.38-31.2
ODIN [23]62.85 / 58.9274.44 / 76.61-31.2
Mahalanobis [19]57.66 / 60.1684.92 / 86.88-31.2
Energy score [26]60.06 / 54.9777.48 / 79.60-31.2
Gram matrices [33]60.93 / 77.5574.93 / 59.38-31.2
Generalized ODIN [14]57.27 / 50.1785.22 / 87.18-31.8
CSI [36]47.10 / 37.0684.09 / 87.99-30.6
GAN-synthesis [18]57.03 / 50.6178.82 / 81.25-31.4
VOS-ResNet50* [6]46.97 / 31.2584.97 / 89.8299.67 / 99.8635.7
VOS-RegX4.0* [6]42.82 / 27.5586.36 / 92.1199.76 / 99.9337.0
Dynamic Prototypes [38]45.72 / 35.0585.14 / 88.92-31.5
APLGOS (ResNet50)41.10 / 23.3087.36 / 92.8799.73 / 99.8935.8
APLGOS (RegNetX4.0)39.48 / 19.7987.47 / 93.5999.79 / 99.9437.6
", + "image_path": "2447cca546421005cb3391856f55013a187b6a3e8c4d4f09afd3378d7ed651c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 66, + 385, + 547, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 357, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 381 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods on mainstream datasets. Here we use PASCAL VOC and Berkeley DeepDrive-100k as ID datasets, MS-COCO2017 and OpenImages as OOD datasets, respectively. “-” denotes that the data is not available." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 385, + 547, + 480 + ], + "lines": [ + { + "bbox": [ + 66, + 385, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 66, + 385, + 547, + 480 + ], + "type": "table", + "html": "
StrategyFPR95 ↓AUROC ↑\nOOD: MS-COCO2017 / OpenImagesmAP (ID) ↑
(a)VOS-RegNetX4.0* [6]50.53 / 50.2788.10 / 87.0898.82 / 97.8049.1
(b)[6] + <CLS>50.12 / 49.5088.56 / 86.8398.91 / 97.7948.2
(c)[6] + “a region of a” + <CLS>51.31 / 50.9688.20 / 86.7398.98 / 97.8548.7
(d)[6] + RP + <CLS>49.50 / 49.4088.49 / 86.7398.82 / 97.7748.9
(e)[6] + <LOC> + “a region of a” + <CLS>49.56 / 47.6088.23 / 87.0798.89 / 97.8749.1
(f)[6] + <LOC> + RP + <CLS> (Ours)45.96 / 47.1089.19 / 88.4999.00 / 98.3049.4
", + "image_path": "e6678457f62cd16546957ce4424139e9f357f6315210f54ecb252a5258acb067.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 486, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 54, + 486, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 555, + 544 + ], + "type": "text", + "content": "Table 2. Ablation studies for prompt strategies. “+” denotes the combination of strategies. “RP” represents sampled prompts from statements set, which is standardized by ChatGPT using Q&A pair and guidance instructions. (b) denotes the simplest prompt strategy, i.e., only providing the ground-truth label for the ID data, (for synthesised OOD image, we define its label as “background”). (c) denotes the original prompt strategy of CLIP [31]. (d) denotes that we replace the prompts in CLIP [31] with the statements by ChatGPT standardizing the Q&A pairs. (e) denotes adding location tokens " + }, + { + "bbox": [ + 54, + 486, + 555, + 544 + ], + "type": "inline_equation", + "content": "<\\mathrm{LOC}>" + }, + { + "bbox": [ + 54, + 486, + 555, + 544 + ], + "type": "text", + "content": " to (c). (f) represents the prompts of our proposed APLGOS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "spans": [ + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "text", + "content": "samples to estimate the class-conditional Gaussian distribution of ID image embeddings and 10000 samples for ID prompts embedding (i.e., " + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_I| = 1000" + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "inline_equation", + "content": "|\\mathcal{Q}_T| = 10000" + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "text", + "content": "). The total length " + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 552, + 298, + 636 + ], + "type": "text", + "content": " of the standardized Q&A pair does not exceed 77. In the experimental tables, “*” denotes results from local replication based on open-source code. “↓” and “↑” indicate lower/greater is better, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 259, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 259, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 259, + 659 + ], + "type": "text", + "content": "4.3. Comparison with The State-of-the-Art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "We report the results of our proposed framework with different image encoder backbones (ResNet50 and RegNetX4.0) on PASCAL VOC, Berkeley DeepDrive-100k, MS-COCO2017, and OpenImages datasets, as shown in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": "Table 1. The best results for the same dataset and the same backbone settings are shown in bold. For the same evaluation metric on the same dataset, the best results are underlined. When using Transformer-based text encoder and ResNet50-based image encoder, APLGOS achieves an FPR95 of " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "47.16\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": " and an mAP of " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "48.8\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": " on PASCAL VOC (ID) with MS-COCO2017 as the OOD dataset. When OpenImages is used as the OOD dataset, FPR95 increases to " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "49.66\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": ". Compared to the state-of-the-art OOD detection model [6], APLGOS reduces FPR95 by " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "1.12\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "2.48\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": " on MS-COCO2017 and OpenImages, respectively. With Transformer-based text encoder and RegNetX4.0-based image encoder, FPR95 decreases to " + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "inline_equation", + "content": "45.96\\%" + }, + { + "bbox": [ + 313, + 552, + 556, + 708 + ], + "type": "text", + "content": " on MS-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3240" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 165, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 165, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 165, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 165, + 198 + ], + "type": "image", + "image_path": "0ee46864c4ae3f6bff9b72b80cd1398287898e58cda985355b911e8a7d54f141.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "type": "text", + "content": "Figure 3. Ablation on number of sampled OOD prompts " + }, + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "type": "text", + "content": ". The horizontal coordinate is the number of sampled ood prompts " + }, + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{K} (\\times 10^3)" + }, + { + "bbox": [ + 55, + 215, + 555, + 250 + ], + "type": "text", + "content": ", while the vertical coordinates are, from left to right, FPR95, AUROC, AUPR, and mAP, respectively. Red line and Blue line represent using MS-COCO2017 and OpenImages as OOD datasets, respectively. Pink line represents using PASCAL VOC as ID dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 276, + 196 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 276, + 196 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 276, + 196 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 276, + 196 + ], + "type": "image", + "image_path": "73c6ad0c18438ff92f9cc9d14ff5d7f846cf62942ac99a9b166ee318d2b2a349.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 198, + 382, + 206 + ], + "lines": [ + { + "bbox": [ + 228, + 198, + 382, + 206 + ], + "spans": [ + { + "bbox": [ + 228, + 198, + 382, + 206 + ], + "type": "text", + "content": "Number of sampled out-of-distribution prompts " + }, + { + "bbox": [ + 228, + 198, + 382, + 206 + ], + "type": "inline_equation", + "content": "(\\times 1000)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 282, + 71, + 391, + 196 + ], + "blocks": [ + { + "bbox": [ + 282, + 71, + 391, + 196 + ], + "lines": [ + { + "bbox": [ + 282, + 71, + 391, + 196 + ], + "spans": [ + { + "bbox": [ + 282, + 71, + 391, + 196 + ], + "type": "image", + "image_path": "1e42fb880866069d188e5d5f77a6441d8294ddfd0716dda7e0e78d78e11b84c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 396, + 72, + 552, + 196 + ], + "blocks": [ + { + "bbox": [ + 396, + 72, + 552, + 196 + ], + "lines": [ + { + "bbox": [ + 396, + 72, + 552, + 196 + ], + "spans": [ + { + "bbox": [ + 396, + 72, + 552, + 196 + ], + "type": "image", + "image_path": "cfaa103478de814d41924f66e90ef2fa4c6892b36f8fa401613060180cdb0528.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 74, + 258, + 279, + 350 + ], + "blocks": [ + { + "bbox": [ + 74, + 258, + 279, + 350 + ], + "lines": [ + { + "bbox": [ + 74, + 258, + 279, + 350 + ], + "spans": [ + { + "bbox": [ + 74, + 258, + 279, + 350 + ], + "type": "table", + "html": "
αFPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
051.63 / 50.8887.86 / 87.2449.2
0.551.90 / 51.4887.55 / 87.0248.9
1.045.96 / 47.1089.19 / 88.4949.4
1.555.88 / 53.3386.29 / 86.7548.9
2.055.92 / 49.5486.75 / 88.0048.9
", + "image_path": "34af21c4a891e98f864dd5209d8977676c1911159887d680ca81570e37542c9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": "COCO2017 and " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "47.1\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on OpenImages, while the mAP on PASCAL VOC improves to " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "49.4\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": ". This setup further reduces FPR95 by " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "4.57\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on MS-COCO2017 and OpenImages, respectively, compared to [6]. For Berkeley DeepDrive-100k (ID), using ResNet50-based image encoder and Transformer-based text encoder, APLGOS achieves an FPR95 of " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "41.10\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on MS-COCO2017 and " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "23.30\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on OpenImages, with an mAP of " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "35.8\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": ". When using RegNetX4.0-based image encoder instead, FPR95 further decreases to " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "39.48\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on MS-COCO2017 and " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "19.79\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": " on OpenImages, while mAP improves to " + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "inline_equation", + "content": "37.6\\%" + }, + { + "bbox": [ + 55, + 408, + 296, + 540 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 156, + 562 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 568, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 700 + ], + "type": "text", + "content": "Prompt strategies. To further validate the effectiveness of our prompt strategies, we conduct extensive ablation experiments on APLGOS's prompt strategies, and the results are shown in Table 2. Sampling from the statements set brings greater performance gains than simply initializing learnable prompts with \"a region of a\" ((c) vs (d)). Moreover, adding location tokens to prompts significantly improves performance, as it refines the scope of the prompts ((c) vs (e)). Compared to other prompt strategies, our APLGOS prompt strategy (f) integrates the advantages of the aforementioned strategies and achieves the best performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Number of Sampled OOD Prompts. APLGOS synthe" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 332, + 258, + 537, + 371 + ], + "blocks": [ + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "text", + "content": "Table 3. The Ablation Experiments on The Strength of Random Gaussian Noise " + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "text", + "content": " represents the strength of added gaussian noise. The value of " + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 358, + 296, + 403 + ], + "type": "text", + "content": " increases gradually from 0 to 2.0, and we take the value at 0.5 intervals." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 258, + 537, + 371 + ], + "lines": [ + { + "bbox": [ + 332, + 258, + 537, + 371 + ], + "spans": [ + { + "bbox": [ + 332, + 258, + 537, + 371 + ], + "type": "table", + "html": "
Γ1FPR95 ↓AUROC ↑mAP (ID) ↑
OOD: MS-COCO2017 / OpenImages
1:450.11 / 58.3887.71 / 85.6749.1
1:349.40 / 55.1287.91 / 86.3849.2
1:247.98 / 54.4988.40 / 85.9449.2
1:145.96 / 47.1089.19 / 88.4949.4
2:148.25 / 50.2088.30 / 87.7649.2
3:150.95 / 53.9486.81 / 84.7047.5
4:150.20 / 51.5686.70 / 84.8947.3
", + "image_path": "f57427ca8ae30f04ab2cfa03145e495f53b7c5fbf9824314af718a823b1d6abd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 422 + ], + "type": "text", + "content": "Table 4. The ablation experiments on the ratio " + }, + { + "bbox": [ + 313, + 379, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\Gamma_{1}" + }, + { + "bbox": [ + 313, + 379, + 555, + 422 + ], + "type": "text", + "content": " of ID and OOD data used during training. Our default parameters and results are shown in bold. Parameters and results of baseline [6] are shown with a dark base color." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": "sises virtual prompts for OOD categories and for each ID category, APLGOS samples " + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": " virtual OOD prompts in low-likelihood regions of ID class-conditional gaussian distributions in high-dimensional hidden space. We conduct ablation experiments on " + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": ", the results of its effect on performance are shown in Figure 3. When " + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": " is too small, it may fail to adequately cover the region outside the ID categories' decision boundaries in the hidden space. On the other hand, when " + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": " is too large, the excessive randomness in the sampled OOD prompts makes it difficult to effectively regularize the decision boundaries with the limited model parameters. Therefore, we set " + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "inline_equation", + "content": "\\kappa = 10000" + }, + { + "bbox": [ + 313, + 434, + 555, + 578 + ], + "type": "text", + "content": " as the default value." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Strength of Random Gaussian Noise " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ". To enhance the size and diversity of the OOD prompts embedding sampling space and prevent the model from overly relying on the ID category distribution, we introduce a learnable matrix initialized with random Gaussian noise " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " during the OOD prompt sampling stage (Eq. 5). We conduct ablation experiments on its strength " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": ", and the results are shown in Table 3. A small value of " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " makes the sampling space of OOD prompts embedding too narrow, while a large value of " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " results in an overly large sampling space. Only by appropriately expanding the sampling space of OOD prompts" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3241" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 541, + 171 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 171 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 171 + ], + "type": "image", + "image_path": "d14e61dc7c29443c7937d3b6931d8983489bba33a013ccb1803fbaf9b71dddfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 555, + 214 + ], + "type": "text", + "content": "Figure 4. Detection results on ID dataset. Here we use Berkeley DeepDrive-100k dataset as ID dataset. We use RegNetX4.0 and Transformer as backbone. The first row is the detection results of baseline [6]. The second row is the detection result of our APLGOS. Our APLGOS rarely misclassifies the ID class as OOD class, and there is almost no missed detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 219, + 541, + 320 + ], + "blocks": [ + { + "bbox": [ + 69, + 219, + 541, + 320 + ], + "lines": [ + { + "bbox": [ + 69, + 219, + 541, + 320 + ], + "spans": [ + { + "bbox": [ + 69, + 219, + 541, + 320 + ], + "type": "image", + "image_path": "d2eb827ad38b660063abcfa24ab55d7a413aa719962bb5b9fb9030164b2a1b70.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "type": "text", + "content": "Figure 5. Detection results on OOD datasets. Here we use Berkeley DeepDrive-100k dataset as ID dataset, MS-COCO2017 and OpenImages as OOD datasets. The first row is the detection results of baseline [6]. The second row is the detection results of our APLGOS. Compared to the baseline, APLGOS rarely misses detections and hardly produces overlapping boxes for the same object." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 376, + 289, + 504 + ], + "blocks": [ + { + "bbox": [ + 63, + 376, + 289, + 504 + ], + "lines": [ + { + "bbox": [ + 63, + 376, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 63, + 376, + 289, + 504 + ], + "type": "image", + "image_path": "daf57df2e22470715408fde8cad652fbaf0191c0d40a902dbf2f45661e00ec71.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 513, + 295, + 546 + ], + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 546 + ], + "type": "text", + "content": "Figure 6. Detection results in Real World. Here we use Berkeley DeepDrive-100k dataset as in-distribution dataset. Pictures we take ourselves with our phone as out-of-distribution dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 581 + ], + "type": "text", + "content": "embedding can the model's ability to fit the OOD distribution be effectively enhanced." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": "Ratio of ID and OOD Data Used During Training. To verify that APLGOS can achieve better performance with less ID data, we conduct ablation experiments on the amount of ID data used during training, and the results are shown in Table 4. By default, APLGOS adopts a ratio " + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\Gamma_{1}" + }, + { + "bbox": [ + 55, + 582, + 295, + 676 + ], + "type": "text", + "content": " of 1:1 for ID and OOD data during training, whereas the baseline [6] uses a ratio of 2:1. However, in this case, the performance of APLGOS decreases instead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "content": "Visualization of Detection Results. To better evaluate the performance of APLGOS, we visualize its detection results on ID datasets, OOD datasets, and real-world scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 450 + ], + "type": "text", + "content": "The results are presented in Figures 4, 5 and 6. The images in real-world scenarios are captured using an iPhone 14 Pro Max. The results demonstrate that APLGOS outperforms the baseline method in detecting ID and OOD categories. Moreover, real-world visualizations further demonstrate its strong generalization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 464, + 388, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 464, + 388, + 477 + ], + "spans": [ + { + "bbox": [ + 314, + 464, + 388, + 477 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a vision-language method, Adaptive Prompt Learning via Gaussian Outlier Synthesis (APLGOS) for Out-of-distribution Detection. Through prompt learning approach, APLGOS provides adaptive region-level prompts with location information for ID / OOD images. We use ChatGPT to standardize pre-defined Q&A pairs and generate a statements set. During training, only ID images are from the dataset, while ID prompts, OOD prompts, and OOD images are all virtual. We sample statements from the statements set to initialize learnable ID prompts. We samples virtual OOD prompts and OOD images in the low-likelihood region of the class-conditional gaussian distribution in high-dimensional hidden space. The similarity score between prompts and images is utilized to calculate contrastive learning loss in high-dimensional hidden space, which guarantees the quality of virtual outliers as well as better regularization of the model. Through comprehensive experimental evaluations, we demonstrated the effectiveness of the proposed APLGOS." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3242" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 290, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 290, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 290, + 126 + ], + "type": "text", + "content": "This paper is supported by Zhongguancun Laboratory and the National Natural Science Foundation of China (Grant No. 62272018)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 148 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 156, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 156, + 296, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 156, + 296, + 199 + ], + "spans": [ + { + "bbox": [ + 61, + 156, + 296, + 199 + ], + "type": "text", + "content": "[1] J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 201, + 296, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 296, + 255 + ], + "type": "text", + "content": "[2] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 257, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 257, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 62, + 257, + 296, + 311 + ], + "type": "text", + "content": "[3] L. Chen, G. Wang, L. Yuan, K. Wang, K. Deng, and P. Torr. Nerf-vpt: Learning novel view representations with neural radiance fields via view prompt tuning. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1156-1164, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 312, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 295, + 345 + ], + "type": "text", + "content": "[4] J. Devlin, M. Chang, K. Lee, and K. Toutanova. Bert: Pretraining of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 378 + ], + "type": "text", + "content": "[5] X. Du, Z Wang, M. Cai, and S Li. Towards unknown-aware learning with virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 422 + ], + "type": "text", + "content": "[6] X. Du, Z. Wang, M. Cai, and Y. Li. Vos: Learning what you don't know by virtual outlier synthesis. International Conference on Learning Representations (ICLR), 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 466 + ], + "type": "text", + "content": "[7] M. Everingham, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The Pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV), 88:303-338, 2010. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 512 + ], + "type": "text", + "content": "[8] Zhongbin Fang, Xiangtai Li, Xia Li, Joachim M Buhmann, Chen Change Loy, and Mengyuan Liu. Explore in-context learning for 3d point cloud understanding. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 545 + ], + "type": "text", + "content": "[9] M. Grcic, P. Bevandić, and S. Šegvic. Dense open-set recognition with synthetic outliers generated by real nvp. arXiv preprint arXiv:2011.11094, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "text", + "content": "[11] D. Hendrycks and K. Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. International Conference on Learning Representations (ICLR), 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "text", + "content": "[12] D Hendrycks, M. Mazeika, and T. Dietterich. Deep anomaly detection with outlier exposure. International Conference on Learning Representations (ICLR), 2018. 1, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] D. Hendrycks, S. Basart, M. Mazeika, A. Zou, J. Kwon, M. Mostajabi, J. Steinhardt, and D. Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[14] Y. Hsu, Y. Shen, H. Jin, and Z. Kira. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10951-10960, 2020. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "text", + "content": "[15] A. Kuznetsova, H. Rom, N. Alldrin, J. Uijlings, I. Krasin, J. Pont-Tuset, S. Kamali, S. Popov, M. Malloci, A. Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV), 128 (7):1956-1981, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[16] T. Le Scao, A. Fan, C. Akiki, E. Pavlick, S. Ilic, D. Hesslow, R. Castagné, A. Luccioni, F. Yvon, M. Galle, et al. Bloom: A 176b-parameter open-access multilingual language model. 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[17] Y. LeCun, S. Chopra, R. Hadsell, M. Ranzato, and F. Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[18] K. Lee, H. Lee, K. Lee, and J. Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. International Conference on Learning Representations (ICLR), 2017. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "type": "text", + "content": "[19] K. Lee, K. Lee, H. Lee, and J. Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems (NeurIPS), 31, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[20] D. Li, J. Li, and S. Hoi. Blip-diffusion: Pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[21] J. Li, P. Chen, Z. He, S. Yu, S. Liu, and J. Jia. Rethinking out-of-distribution (OOD) detection: Masked image modeling is all you need. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11578-11589, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 509 + ], + "type": "text", + "content": "[22] K. Li, Q. Geng, M. Wan, X. Cao, and Z. Zhou. Context and spatial feature calibration for real-time semantic segmentation. IEEE Transactions on Image Processing (TIP), 32: 5465-5477, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 511, + 553, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 554 + ], + "type": "text", + "content": "[23] S. Liang, Y. Li, and R. Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. International Conference on Learning Representations (ICLR), 2017. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 590 + ], + "type": "text", + "content": "[24] J. Liao, X. Xu, M. Nguyen, A. Goodge, and C. Foo. Coftad: Contrastive fine-tuning for few-shot anomaly detection. IEEE Transactions on Image Processing (TIP), 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 634 + ], + "type": "text", + "content": "[25] T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollar, and C. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[26] W. Liu, X. Wang, J. Owens, and Y. Li. Energy-based out-of-distribution detection. Advances in neural information processing systems (NeurIPS), 33:21464-21475, 2020. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] X. Liu, Y. Lochman, and C. Zach. Gen: Pushing the limits of softmax-based out-of-distribution detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23946-23955, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3243" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[28] Yifei Ming, Ziyang Cai, Jiuxiang Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. Advances in Neural Information Processing Systems (NeurIPS), 35:35087-35102, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 172 + ], + "type": "text", + "content": "[29] S. Mohseni, M. Pitale, J. Yadawa, and Z. Wang. Self-supervised learning for generalizable out-of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 5216-5223, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 228 + ], + "type": "text", + "content": "[30] S. Ning, L. Qiu, Y. Liu, and X. He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "text", + "content": "[31] A. Radford, J. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763, 2021. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 327 + ], + "type": "text", + "content": "[32] I. Radosavovic, R. Kosaraju, R. Girshick, K. He, and P. Dollar. Designing network design spaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10428-10436, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 360 + ], + "type": "text", + "content": "[33] C. Sastry and S. Oore. Detecting out-of-distribution examples with gram matrices. In International Conference on Machine Learning (ICML), pages 8491-8501, 2020. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 404 + ], + "type": "text", + "content": "[34] B. Su, H. Zhang, and Z. Zhou. Hsic-based moving weight averaging for few-shot open-set object detection. In Proceedings of the 31st ACM International Conference on Multimedia (ACMM), pages 5358-5369, 2023. 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "text", + "content": "[35] B. Su, H. Zhang, J. Li, and Z. Zhou. Toward generalized few-shot open-set object detection. IEEE Transactions on Image Processing (TIP), 33:1389-1402, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 481 + ], + "type": "text", + "content": "[36] J. Tack, S. Mo, J. Jeong, and J. Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. Advances in neural information processing systems (NeurIPS), 33:11839-11852, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 296, + 525 + ], + "type": "text", + "content": "[37] H. Wang, Y. Li, H. Yao, and X. Li. Clipn for zero-shot ood detection: Teaching clip to say no. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1802-1812, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "text", + "content": "[38] A. Wu, C. Deng, and W. Liu. Unsupervised out-of-distribution object detection via pca-driven dynamic prototype enhancement. IEEE Transactions on Image Processing (TIP), 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "text", + "content": "[39] X. Wu, F. Zhu, R. Zhao, and H. Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7031-7040, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[40] C. Xie, Z. Zhang, Y. Wu, F. Zhu, R. Zhao, and S. Liang. Described object detection: Liberating object detection with flexible expressions. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 714 + ], + "type": "text", + "content": "[41] X. Xiong, Y. Liu, T. Yuan, Y. Wang, Y. Wang, and H. Zhao. Neural map prior for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17535-17544, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 588 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 118 + ], + "type": "text", + "content": "[42] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems (NeurIPS), 36, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 555, + 172 + ], + "type": "text", + "content": "[43] Z. Yang, Z. Gan, J. Wang, X. Hu, F. Ahmed, Z. Liu, Y. Lu, and L. Wang. Unitab: Unifying text and box outputs for grounded vision-language modeling. In European Conference on Computer Vision (ECCV), pages 521-539. Springer, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 172, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 172, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 172, + 555, + 229 + ], + "type": "text", + "content": "[44] F. Yu, H. Chen, X. Wang, W. Xian, Y. Chen, F. Liu, V. Madhavan, and T. Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2636-2645, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 229, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 229, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 315, + 229, + 555, + 284 + ], + "type": "text", + "content": "[45] Y. Zhai, Y. Zeng, Z. Huang, Z. Qin, X. Jin, and D. Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 6979–6987, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 285, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 555, + 329 + ], + "type": "text", + "content": "[46] C. Zhang, X. Chen, S. Chai, C. Wu, D. Lagun, T. Beeler, and F. De la Torre. Iti-gen: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3969-3980, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 329, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 329, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 329, + 555, + 364 + ], + "type": "text", + "content": "[47] J. Zhang, N. Inkawich, Y. Chen, and H. Li. Fine-grained out-of-distribution detection with mixup outlier exposure. arXiv preprint arXiv:2106.03917, 2(5), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 364, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 364, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 315, + 364, + 555, + 398 + ], + "type": "text", + "content": "[48] J. Zhang, J. Huang, S. Jin, and S. Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 398, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 398, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 315, + 398, + 555, + 442 + ], + "type": "text", + "content": "[49] K. Zhou, J. Yang, C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 16816-16825, 2022. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 442, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 442, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 315, + 442, + 555, + 475 + ], + "type": "text", + "content": "[50] K. Zhou, J. Yang, C. Loy, and Z. Liu. Learning to prompt for vision-language models. International Journal of Computer Vision (IJCV), 130(9):2337-2348, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 475, + 555, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 555, + 543 + ], + "type": "text", + "content": "[51] Z. Zhu, Y. Zhang, H. Chen, Y. Dong, S. Zhao, W. Ding, J. Zhong, and S. Zheng. Understanding the robustness of 3d object detection with bird's-eye-view representations in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21600-21610, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 543, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 543, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 315, + 543, + 555, + 588 + ], + "type": "text", + "content": "[52] O. Zohar, K. Wang, and S. Yeung. Prob: Probabilistic objectness for open world object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11444-11453, 2023. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3244" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_content_list.json b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d711539d1baec2af29e437b329f30939beca9a96 --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_content_list.json @@ -0,0 +1,1886 @@ +[ + { + "type": "text", + "text": "Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zewei Xin, Qinya Li*, Chaoyue Niu*, Fan Wu, Guihai Chen \nShanghai Jiao Tong University, China", + "bbox": [ + 267, + 203, + 730, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xin_zeweil12138，qinyali，rvince}@sjtu.edu.cn，{fwu，gchen}@cs.sjtu.edu.cn", + "bbox": [ + 169, + 241, + 821, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation.", + "bbox": [ + 89, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 715, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, text-to-image (T2I) models like Imagen [3], Stable Diffusion [28], and DALL-E [27] have achieved significant success in generating diverse, high-quality images given user prompts. However, the impressive generation quality comes with large model and high cost. For example, Stable Diffusion 3.5 [1] has 8 billion parameters. Such a large model necessitates reliance on cloud servers, leading to high serving cost. As shown in Tab. 1, it is particularly costly in commercial scenarios with millions of requests.", + "bbox": [ + 89, + 739, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/ffacc91aeca21aae7aeeb76f5d2f76911aa0888db9f5ff7d12638571a42c6c29.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Text-to-Image Model#ParamPricing ($/M)
Stable Diffusion 1.6 [28]0.86 B9 K
Stable Diffusion XL [24]2.6 B9 K
Stable Diffusion 3 Medium [9]2 B35 K
Stable Diffusion 3 [9]8 B65 K
Stable Diffusion 3.5 [1]8 B65 K
", + "bbox": [ + 532, + 290, + 887, + 375 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. For the Stable Diffusion models, newer versions come with more parameters and higher API costs.", + "bbox": [ + 511, + 386, + 906, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The high cost of cloud-based image generation drives the trend of deploying T2I models on edge devices. Techniques such as quantization [16, 34, 41], structural pruning [4, 17], and reducing denoising steps [20, 22, 30] are used to minimize model sizes and speed up inference, showing that edge deployment is feasible. Compared to cloud models, edge models leverage users' local computing devices to provide instant services anytime and anywhere without cloud serving cost and communication overhead. However, the lightweight edge T2I models often have lower generation quality compared to the cloud large models.", + "bbox": [ + 511, + 444, + 906, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although cloud models often offer superior quality, not all user prompts need these large models. For some prompts, light-weight edge models can produce comparable or even better results. To ensure high-image quality generation while limiting requests to cloud models to reduce cost, it is necessary to introduce a routing mechanism such that suitable models are selected for different user prompts based on their complexity. Intuitively, as shown in Fig. 1, a routing framework for T2I models routes only challenging prompts to the cloud, while handling easy ones on the edge.", + "bbox": [ + 511, + 612, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous routing methods are primarily designed for large language models (LLMs) [8, 11, 21, 23], presenting significant challenges when applied to T2I models. One major challenge lies in comparing the quality of generated images. Unlike text, image quality does not have unified metrics and is subject to many influencing factors, including color, clarity, and the unique distortions typical of generated images [15]. This makes it hard to compare image quality to guide routing decisions. Furthermore, the output space of", + "bbox": [ + 511, + 765, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Q. Li and C. Niu are the corresponding authors.", + "bbox": [ + 107, + 886, + 367, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19482", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c991700e86973aca2b80858dd5d63f89dc9bdfc6f13659db4ef332c05c4f8ba6.jpg", + "image_caption": [ + "Figure 1. Routing for text-to-image generation requests." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "T2I models is substantially larger than the input text space, and a single prompt can correspond to numerous images, hindering the prediction of image quality before generation.", + "bbox": [ + 88, + 272, + 482, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose an edge-cloud T2I routing framework RouteT2I, deploying a routing model and a small T2I model on the edge, and a large T2I model on the cloud. Our goal is to optimize the overall quality of generated images at a limited cost, by effectively routing user prompts to the edge or the cloud. To evaluate the quality gap between generated images, we first propose a multi-metric quality measure, including metrics for both real photos and those unique to generated images. Each metric assesses image quality by measuring the similarity between the image and a pair of positive and negative prompts that describe the quality metric. We then define Pareto relative superiority (PRS) to quantify quality differences between images. RouteT2I predicts PRS based on user prompts, inspired by the cross-attention operation between text and image during generation. In particular, inspired by mixture-of-experts (MoE) selective focus on experts or tokens in LLMs [2, 32, 43], we design a novel dual-gate token selection MoE, where a user prompt is treated as a sequence of tokens, with experts aligned to quality metrics. A token selection gate allows experts to proactively choose tokens that markedly impact these metrics, thereby focusing on key tokens. A dual-gate MoE uses both positive and negative gates to evaluate dominant effects when tokens have opposing effects simultaneously. Furthermore, multiple heads are introduced to predict multiple quality metrics. Based on the multi-metric quality prediction, RouteT2I determines quality differences between images generated on the edge over those in the cloud. Only user prompts that show substantial quality improvement when generated in the cloud are routed there, while others remain at the edge. Our key contributions are as follows:", + "bbox": [ + 91, + 319, + 483, + 804 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We, for the first time, study how to route text-to-image generation requests between cloud and edge models for low cost and high quality.", + "- Specified to multi-metric quality measures of generated images, we design a routing model architecture with dual gate token selection MoE, combined with a routing strat" + ], + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "egy for user prompts.", + "bbox": [ + 526, + 92, + 671, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We evaluate the proposed design using 18 image generation model pairs on the public COCO 2014 dataset. Key results include: (1) by routing $50\\%$ of requests to edge and cloud models, respectively, the overall quality improvement over the edge model reaches $83.97\\%$ of that achieved by fully using the cloud model; and (2) to achieve $50\\%$ of the quality improvement attained by fully using the cloud model, the number of requests routed to the cloud model is reduced by $70.24\\%$ compared to a random routing policy.", + "bbox": [ + 511, + 106, + 906, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 271, + 655, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model routing selects suitable models for user requests to ensure quality while reducing high cost of invoking powerful models. Depending on whether routing decision occurs before or after weak model execution, model routing can be categorized into predictive and non-predictive routing.", + "bbox": [ + 511, + 296, + 906, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-predictive routing takes the output of a weak model to decide whether to route to the next, powerful model. LAECIPS [12] evaluated the confidence of predictions made by the edge semantic segmentation model to decide if cloud assistance from a large visual model is needed. For classification models, Hybrid [13] used a routing model to partition the data domain between edge and cloud models, routing cases with incorrect edge outputs to the cloud. For cascading LLM, LLM Cascade [40] routed based on the consistency of responses from weaker LLMs, while FrugalGPT [6] considered current quality and past costs to determine routing. Tabi [36] not only routed based on the confidence of responses from weaker LLMs but also enhanced output quality by aggregating historic outputs.", + "bbox": [ + 511, + 372, + 908, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Different from non-predictive routing, predictive routing selects model before running the weak model on resource-constrained edge device. Previous studies focused on LLM, with the difference in routing strategies. RouteLLM [23] predicted the outcomes of quality comparisons between LLM and routed based on prediction confidence. Hybrid LLM [8] relaxed comparisons, allowing the weak model to succeed if the quality gap is within a threshold, thus saving costs with some quality compromise. Recent research [33] advances the confidence routing strategy by identifying out-of-distribution data through confidence thresholds. ZOOTER [21] introduced a new strategy by predicting the normalized quality of candidate model outputs, making routing decision based on relative quality, with routing models distilled from existing quality scoring models.", + "bbox": [ + 511, + 584, + 908, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous routing work were mainly for LLMs, whereas the predictive routing for T2I in our work remains unexplored. A key distinction lies in the image outputs of T2I, which necessitate predicting quality variations in a larger output image space from a smaller input text space (Sec. B in Appendix). On one hand, unlike many LLM tasks with", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19483", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb7a08002e54d00522e1bc917989117d998fc99d2bbab8e3455c240629b7235a.jpg", + "image_caption": [ + "SD2.1", + "SD3", + "There is a dog lying on the ground." + ], + "image_footnote": [], + "bbox": [ + 96, + 99, + 285, + 172 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b212a8d7beae69e53ccac2f15934339672dc4efaca4c612c69f65075983dd916.jpg", + "image_caption": [ + "SD2.1" + ], + "image_footnote": [], + "bbox": [ + 289, + 101, + 383, + 171 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/667356aa2dd35d9944466490e4ae2f60245c52c2e87407c7f44aa43256f4996b.jpg", + "image_caption": [ + "SD3" + ], + "image_footnote": [], + "bbox": [ + 383, + 101, + 475, + 171 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/41a45d196d09eb111a1fa5f3f18ba31527feb677211034be1a5a05abf83889fe.jpg", + "image_caption": [ + "A cat sitting motionless on a television stand behind a mug in the shape of a skeleton head." + ], + "image_footnote": [], + "bbox": [ + 96, + 181, + 284, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/90303fed76d78ba2c4ff90cbeb63402d60fa842b59a34a10fdd87f43978f3f81.jpg", + "image_caption": [ + "A sandwich cut in half on a white plate with a side of salad and a fork and knife wrapped in a napkin." + ], + "image_footnote": [], + "bbox": [ + 287, + 181, + 383, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/296310e1fda8a1a11f508c534489c61459e0b2e27d3aba5943e772e421f0e188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 181, + 475, + 253 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e4355162532ad6fd1e9204ff5e6b557f33ad69cf0d4b421d1b5e0e17168d80c.jpg", + "image_caption": [ + "Figure 2. SD2.1 tends to match or outperform SD3 on simple prompts, while more likely falling short on complex ones. Entities to be generated are **bolded**.", + "(a)", + "Figure 3. Impact of the noun count in text prompts on ImageReward score [39] gap and win rate between SD3 and SD2.1." + ], + "image_footnote": [], + "bbox": [ + 94, + 349, + 284, + 460 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/156f812081bfb20dd78360833c4f508218850ef1e087c7e65701f52075e19c99.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 287, + 352, + 475, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "definitive answers, image quality lacks fixed standards and is highly subjective. Previous single-objective LLM routing struggles to adapt to such ambiguous quality criteria. On the other hand, prior work lacked dedicated routing model designs, making it challenging to predict quality relationships within the vast output image space. This underscores the necessity of effective routing model design.", + "bbox": [ + 89, + 542, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 662, + 295, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In text-to-image generation scenario, the cloud normally hosts a large T2I model, denoted as $\\mathcal{M}_c$ , while a resource-constraint edge device, such as a smartphone, often deploys a light-weight T2I model, denoted as $\\mathcal{M}_e$ . For example, Google has released Imagen [3, 29] for cloud deployment and MobileDiffusion [42] for edge deployment. From the perspective of a user, the large model generally offers superior generation quality, but incurs edge-cloud communication cost and cloud serving fee, which together are denoted as $F_c$ . In contrast, the edge model can leverage the user's local device, thereby eliminating these additional costs.", + "bbox": [ + 89, + 689, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, user requests involve not only complex ones but also simple ones. One intuitive factor influencing request complexity is the number of entities that need to", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be generated in images, i.e. the number of nouns in text prompts. As shown in Fig. 2, 3, as noun count increases, the large model is more likely to outperform the small model in quality, with quality differences becoming more noticeable. However, while large models generally excel in overall quality, small models may perform better on specific prompts. This likelihood increases as the prompt becomes simpler. In such cases, the purely large model serving not only fails to improve performance but also incurs high cost. Therefore, it is necessary to design a routing mechanism to select the most suitable model based on the prompt's characteristics. Intuitively, the routing mechanism should direct simpler prompts to the cost-effective edge model and more complex prompts to the high-quality cloud model, thus maximizing generation quality while minimizing cost.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, a T2I routing framework $R: \\mathcal{X} \\to \\{0,1\\}$ assigns user prompts $\\mathcal{X}$ to the cloud large model $\\mathcal{M}_c$ as 1 and to the edge light-weight model $\\mathcal{M}_e$ as 0. The generated images after routing can be expressed as $\\mathcal{I}_r = R(\\mathcal{X})\\mathcal{I}_c + (1 - R(\\mathcal{X}))\\mathcal{I}_e$ , where $\\mathcal{I}_c$ and $\\mathcal{I}_e$ denote the images generated using $\\mathcal{M}_c$ and $\\mathcal{M}_e$ , respectively. Given a quality scoring function $\\mathcal{Q}(\\cdot)$ of the generated images, the cost budget $\\tau_{fee}$ of cloud model serving and edge-cloud communication, and the response latency constraint $\\tau_{time}$ , the optimization objective is maximizing the quality under the budget and latency constraints, formulated as", + "bbox": [ + 511, + 316, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot F _ {c} \\leq \\tau_ {f e e} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} - D _ {c} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} - D _ {c} \\end{array} \\tag {1} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot D _ {\\mathcal {M} _ {c}} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} \\cdot D _ {\\mathcal {M} _ {e}} \\\\ + D _ {R} \\leq \\tau_ {\\text {t i m e}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 492, + 903, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{\\mathcal{M}_c}$ , $D_{\\mathcal{M}_e}$ , and $D_R$ denote the latency of cloud large model serving plus edge-cloud communication, the latency of the edge light-weight model serving, and the latency of the routing model execution. Considering the constraints are linear, the optimization objective can be approximated as imposing an upper bound $\\rho_r$ on the routing rate to the cloud large model as", + "bbox": [ + 511, + 574, + 905, + 680 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\leq \\rho_ {r}. \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 690, + 903, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the lack of a universal standard for evaluating the quality of generated images, a multi-dimensional quality scoring function $\\mathcal{Q}:\\mathcal{I}\\to [0,1]^N$ is required for a more comprehensive comparison. Thus, different from existing routing for typical classification and text generation tasks, the T2I routing framework's optimization objective is a new multi-objective problem. In addition, to predict quality changes in the larger image output space from the smaller text input space, we conduct a detailed analysis of how different prompt tokens influence the generated images and their multi-dimensional quality metrics, by examining", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19484", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the interaction between text and images during the T2I generation process. In what follow, we define the multi-metric quality measures for routing objectives in Sec. 4 and introduce the T2I routing framework in Sec. 5.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Multi-Metric Image Generation Quality in Routing Objectives", + "text_level": 1, + "bbox": [ + 89, + 162, + 482, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The connection between text and images allows image attributes to be described through corresponding prompts. Thus, a metric can be measured by the relative relationship between an image and the positive/negative text pair describing that metric [18, 35]. The contrastive quality metric for image $I$ is defined as", + "bbox": [ + 89, + 205, + 483, + 296 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq (I, m) = \\sigma \\left(C L I P \\left(I, m ^ {+}\\right) - C L I P \\left(I, m ^ {-}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 303, + 482, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m = (m^{+}, m^{-})$ denotes positive and negative pair, and $\\sigma$ denotes the sigmoid function that transforms output values in the range from 0 to 1. According to this definition, if an image is more related to the positive prompt, and is less related to the negative prompt, the contrastive quality metric is higher. Compared to quality measure with only the positive prompt, the contrastive method evaluates whether positive or negative quality is more dominant in the image, leading to a more robust and reliable evaluation.", + "bbox": [ + 89, + 325, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the ambiguity and complexity of image quality, we introduce a multi-dimensional metric to comprehensively evaluate generated image quality, providing noise resistance and stability for quality-prediction-based routing. Just as real photo quality depends on factors like definition and color, our multi-metric combines traditional measures of real photos with unique aspects of generated images, such as realism and object integrity. Thus, we extend the single-metric evaluation in Eq. (3) to an $N$ -dimensional quality assessment of an image $I$ , expressed as", + "bbox": [ + 89, + 463, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} (I) = \\left[ q (I, m _ {i}) \\mid i = 1, 2, \\dots , N \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 619, + 480, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m_{i}$ is an instantiation of $m$ in Eq. (3). We adopt 10 common metrics inspired by prior works [14, 37, 38], with details in Appendix Sec. C. For example, for the definition metric, we use the positive prompt \"High definition photo\" and the negative prompt \"Low definition photo\"; and for the object integrity, we take the positive prompt \"Object completion\" and the negative prompt \"Object twisting\".", + "bbox": [ + 89, + 643, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the multi-objective optimization setting for routing with multi-metric quality, the initial goal is to find a Pareto optimal generated image that excels in all the metrics. However, this is challenging in practice due to the potential nonexistence of such an image generated from the cloud or edge model. We thus relax the constraint of Pareto optimality by allowing suboptimal performance in some metrics if the image significantly outperforms in others when selecting higher-quality images. We define Pareto relative superiority (PRS) to quantify the quality advantage of the images", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1642754cac28ecc02eed28cddaaba7c38d98d82204342cc6613be67deb64e388.jpg", + "image_caption": [ + "Figure 4. Overview of RouteT2I. RouteT2I assembles a pair of off-the-shelf edge and cloud text-to-image models and evaluates them using multi-metric image quality. RouteT2I utilizes Pareto relative superiority between qualities as supervision to train the routing model. Then, to balance quality and cost, the routing strategy determines the most suitable model for each user prompt, choosing between the edge or cloud model." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{I}_e$ generated by the edge model over the images $\\mathcal{I}_c$ generated by the cloud model. In particular, we first normalize the quality distance for the metric $i$ between an edge-generated image $I_e \\in \\mathcal{I}_e$ and a cloud-generated image $I_c \\in \\mathcal{I}_c$ as", + "bbox": [ + 511, + 392, + 905, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} (I _ {e}, I _ {c}) = \\sigma \\left(\\frac {q (I _ {e} , m _ {i}) - q (I _ {c} , m _ {i})}{\\Gamma | \\mu_ {i} (\\mathcal {I} _ {e}) - \\mu_ {i} (\\mathcal {I} _ {c}) |}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 464, + 905, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_i(\\cdot)$ is the average quality of the set of cloud or edge generated images, while the sigmoid function $\\sigma$ and the temperature parameter $\\Gamma$ are used to modulate the data distribution, effectively distinguishing similar qualities and preventing centralization. We then define the overall quality gap as a weighted sum of distances across all $N$ metrics:", + "bbox": [ + 511, + 508, + 905, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP R S \\left(I _ {e}, I _ {c}\\right) = \\sum_ {i = 1} ^ {N} w _ {i} D _ {i} \\left(I _ {e}, I _ {c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 611, + 903, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ denotes the weight of metric $i$ , and $\\sum_{i} w_{i} = 1$ . We can evaluate the relative quality of $I_{e}$ compared to $I_{c}$ by examining how much Pareto relative superiority deviates from 0.5. This evaluation helps to route a user prompt to the model that generates an image with higher overall quality.", + "bbox": [ + 511, + 664, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Design of RouteT2I", + "text_level": 1, + "bbox": [ + 511, + 753, + 714, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve the optimization objective in Eq. (2), we propose a text-to-image model routing framework RouteT2I. As shown in Fig. 4, RouteT2I comprises a routing model that predicts the multi-metric generation quality given user prompts and a routing strategy that selects cloud or edge model. In particular, (1) the routing model predicts the quality relationship between images generated on the edge and cloud, namely, Pareto relative superiority. Inspired by the", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19485", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fba23697dd652034e0744e5e4567e6470325d210cae48cc56f74a5576d3b4970.jpg", + "image_caption": [ + "(a) Token selection gate." + ], + "image_footnote": [], + "bbox": [ + 147, + 89, + 428, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4fcdf9af33b1f26cc5f64f2650bc2e4bc46c0b98bb5747a72b9d7503897021c7.jpg", + "image_caption": [ + "(b) Dual-gate MoE.", + "Figure 5. Architecture of a dual-gate MoE that uses the token selection gate as the gate network. First, the gate selects the most relevant tokens for each expert. Then, experts project the tokens selected by positive and negative gates into their respective spaces and evaluate them using score matrices. Finally, the evaluations of the tokens are gathered and compared to form new tokens." + ], + "image_footnote": [], + "bbox": [ + 464, + 88, + 846, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "role of prompts in generation, where prompt as sequences of tokens interact with images through cross-attention to determine the image content, we treat prompt as a set of tokens in the routing model and select key tokens with influential contextual information for each quality metric. Then, the dominant influence of these tokens is accessed by comparing their positive and negative effects; and (2) the routing strategy aims to maximize quality under cost constraints. The key idea is to use Pareto relative superiority to describe the quality disparity of images generated by the cloud and the edge models. Only prompts that show a notable quality gap when generated in the cloud are routed there, while the rest remain on the cost-effective edge.", + "bbox": [ + 88, + 294, + 485, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Routing Model Architecture Design with Dual-Gate Token Selection MoE", + "text_level": 1, + "bbox": [ + 89, + 500, + 483, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To predict Pareto relative superiority between the images generated by the cloud model and the edge model from user prompts, we design a dual-gate token selection mixture-of-experts (MoE), as shown in Fig. 5, and introduce it into Transformer network to replace linear layers.", + "bbox": [ + 89, + 537, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model treats the user prompt as a sequence of tokens, whose influences on the image quality differ due to their weights in cross-attention during generation. To focus on key tokens for each metric, we design a token selection gate, where experts align with quality metrics and actively choose the most relevant tokens to simulate metric attention to different image attributes. Only tokens that significantly impact the quality metrics corresponding to the experts are selected, while those with minimal influences are disregarded, reducing potential interference.", + "bbox": [ + 88, + 614, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Regarding the design of a dual-gate MoE, both positive and negative gates are introduced to distinguish the opposing aspects of token influences on image quality. Similar to images that have both superior and inferior regions when evaluated by metrics, these tokens are assessed by the corresponding experts for their positive and negative influences to each metric. Additionally, we introduce a comparison between two gates to identify which influence of each token predominates overall. We further improve the structure of", + "bbox": [ + 88, + 763, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "expert to handle both positive and negative influences on a single metric simultaneously.", + "bbox": [ + 511, + 294, + 906, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To support multi-metric quality prediction, our design incorporates multiple prediction heads within the model architecture. Each head outputs the prediction of a specific quality metric. Such multi-head design not only enhances noise resistance, thereby increasing robustness, but also facilitates a comprehensive evaluation of quality.", + "bbox": [ + 511, + 324, + 906, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1. Token Selection Gate", + "text_level": 1, + "bbox": [ + 511, + 420, + 707, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the token selection gate, the dot product between the token representations $T \\in \\mathbb{R}^{n \\times d}$ and the expert embeddings $E \\in \\mathbb{R}^{k \\times d}$ is computed as an affinity matrix:", + "bbox": [ + 511, + 438, + 906, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {S o f t m a x} \\left(T \\cdot E ^ {T}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 488, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ denotes the number of tokens in user prompt, $k$ denotes the number of experts, $d$ denotes the hidden dimension, and the Softmax function is applied along the expert axis. $A[t,i]$ in the affinity matrix captures the correlation between token $t$ and expert $i$ . Then, top- $K$ tokens with the highest relevance for each expert are selected", + "bbox": [ + 511, + 512, + 905, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = \\operatorname {T o p} - \\mathrm {K} (A), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 609, + 903, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M \\in \\mathbb{R}^{n \\times k}$ is the binary mask matrix such that $M[t, i] = 1$ denotes token $t$ is selected by expert $i$ , and $M[t, i] = 0$ denotes not selected. Experts tend to have similar weights because the metrics they are responsible for are weighted similarly. This ensures that the tokens selected by these experts possess significant global importance. This practice minimizes interference from tokens that are insignificant across all metrics. The influences of key tokens are further evidenced by their frequent selection through the top- $K$ function and their corresponding scores within the affinity matrix. Moreover, by allocating an equal number of $K$ tokens to each expert, workloads remain balanced, and bias in quality dimensions is avoided, ultimately resulting in more comprehensive and reliable quality prediction. Overall, the token selection gate is summarized as", + "bbox": [ + 511, + 632, + 906, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA, M = G (T, E), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 880, + 903, + 897 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19486", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G$ denotes the gate network.", + "bbox": [ + 89, + 90, + 321, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2. Dual-Gate MoE", + "text_level": 1, + "bbox": [ + 89, + 112, + 246, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both a positive gate $G^{+}$ and a negative gate $G^{-}$ are introduced for different experts to select tokens that have positive and negative impacts on their corresponding quality metrics, while capturing the weights of their influences. The token selection results are denoted as", + "bbox": [ + 89, + 131, + 483, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {o}, M ^ {o} = G ^ {o} \\left(T, E ^ {o}\\right), \\text {w h e r e} o \\in \\{+, - \\}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 215, + 480, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The affinity matrix $A^o$ and the binary mask matrix $M^o$ are captured by the gate $G^o$ for the positive or negative aspect using the corresponding expert embedding $E^o$ . The tokens selected by two gates are then evaluated by the corresponding positive or negative experts.", + "bbox": [ + 89, + 241, + 483, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To eliminate the redundancy of having separate experts for evaluating both positive and negative aspects of the same quality metric, the structure of experts is revised to allow an expert to assess both aspects simultaneously. This revision acknowledges that both positive and negative evaluations are inherently part of the same quality metric. Specifically, the original linear layer $W_{i} \\in \\mathbb{R}^{d \\times h}$ of expert on metric $i$ is split into two projection matrices $P_{i}^{+}, P_{i}^{-} \\in \\mathbb{R}^{d \\times l}$ and one score matrix $S_{i} \\in \\mathbb{R}^{l \\times h}$ . Two projection matrices are specific to positive and negative aspects, projecting tokens into the corresponding representations in an $l$ -dimensional low-dimensional space. Since $l \\ll h, d$ , the parameter size and the computation cost are reduced from $O(hd)$ to $O(l(h + d))$ , allowing an expert to efficiently focus on both positive and negative aspects of a metric with a low cost. The shared score matrix then evaluates the impact of these representations on metric $i$ . Thus, the impact of $t$ -th token $T[t] \\in \\mathbb{R}^d$ on quality metric $i$ for positive or negative aspect $o$ is described as", + "bbox": [ + 89, + 316, + 483, + 603 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {o} [ t ] = T [ t ] \\cdot P _ {i} ^ {o} \\cdot S _ {i}, \\text {w h e r e} o \\in \\{+, - \\}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 613, + 480, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By separating the projection matrices and score matrix, we can evaluate both positive and negative impacts on a metric within the same expert. Then, the overall positive or negative impact of the $t$ -th token on image quality can be obtained by a weighted sum of its impacts for corresponding aspects across all metrics", + "bbox": [ + 89, + 638, + 483, + 729 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {o} [ t ] = \\sum_ {i = 0} ^ {k} \\lambda_ {t, i} ^ {o} \\cdot T _ {i} ^ {o} [ t ], \\text {w h e r e} o \\in \\{+, - \\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 739, + 480, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $\\lambda_{t,i}^{o} = \\frac{\\mathbb{1}\\{M^{o}[t,i] = 1\\}A^{o}[t,i]}{\\sum_{j = 0}^{k}\\mathbb{1}\\{M^{o}[t,j] = 1\\}A^{o}[t,j]}$ denotes the normalized affinity of the $i$ -th expert for the $t$ -th token. By contrasting the impact of a token from both positive and negative aspects, we can assess the token's predominant influence on the image quality as", + "bbox": [ + 89, + 787, + 483, + 871 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {T} [ t ] = \\sigma \\left(T ^ {+} [ t ] - T ^ {-} [ t ]\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 878, + 480, + 897 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The introduction of contrasting helps diminish ambiguity in predictions, especially when tokens exhibit influences on multiple metrics from both positive and negative aspects.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Routing Strategy Design", + "text_level": 1, + "bbox": [ + 511, + 143, + 738, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the Pareto relative superiority predicted by routing model, we propose a routing strategy to effectively route user prompts to suitable models to achieve the trade-off between quality and costs. With a preset routing rate, we can determine the proportion of user prompts sent to the cloud, focusing on those where the cloud model significantly outperforms the edge model in quality. Since Pareto relative superiority represents this quality gap, we can efficiently filter prompts by setting a threshold $\\alpha$ on it. In particular, the prompts with a Pareto relative superiority above the threshold are better handled by the edge model for cost efficiency, while those below the threshold achieve superior quality when processed in the cloud. Thus, the optimization objective in Eq. (2) can be expressed as:", + "bbox": [ + 511, + 166, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\alpha \\leq 1 / 2} \\mathbb {P} \\left\\{P R S \\left(I _ {e}, I _ {c}\\right) < \\alpha \\mid I _ {e}, I _ {c} \\in \\mathcal {I} _ {e}, \\mathcal {I} _ {c} \\right\\} \\leq \\rho_ {r}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 386, + 906, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since Pareto relative superiority being greater than or less than $1/2$ indicates a relative advantage or disadvantage, respectively, we set a upper bound $1/2$ on $\\alpha$ to prevent user prompts with images generated at the edge of better quality from being routed to the cloud.", + "bbox": [ + 511, + 417, + 905, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 511, + 506, + 638, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 531, + 702, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We take COCO2014 [19], a comprehensive dataset for object detection, segmentation, and captioning tasks. We select its captions as user prompts $\\mathcal{X}$ . Given these prompts, we generate images $\\mathcal{I}_e, \\mathcal{I}_c$ using different open-source text-to-image models. Similarly, we use the LAION dataset [31], with results shown in Appendix Sec. F.", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We take T2I models with varying performances and sizes for cloud and edge usage, include Stable Diffusion 1.5 (SD1.5) [28], Stable Diffusion 2.1 (SD2.1) [28], Stable Diffusion XL (SDXL) [24], Stable Diffusion XL-Refiner (XL-Refiner) [24], Stable Diffusion 3 (SD3) [9], PixArt $\\alpha$ [5], and Infinity [10]. These models use various text encoders, such as CLIP [25], OpenCLIP [7], and T5 [26], and are based on architectures like diffusion, diffusion Transformer and autoregressive. Details are in Appendix Sec. A.", + "bbox": [ + 511, + 643, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines and Settings. We first introduce random routing as a baseline, where prompts are randomly assigned to candidate models, given a specified routing rate to the cloud. Additionally, although there does not exist any previous work on text-to-image generation routing, for a comprehensive comparison, we also reproduce several representative routing methods for LLM by adapting them to our scenarios, including RouteLLM [23] with BERT classifier or", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19487", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/092065c5714afe1210b9e9a7a282a8f9b041d7a5ea8c71bbb1c381ab9157b869.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RouterImage Quality MetricsΔP(%)
DefinitionDetailClaritySharpnessHarmonyRealismColorConsistencyLayoutIntegrity
Edge Model0.62510.66850.60760.65370.59490.55750.46800.50880.48600.4690-
Cloud Model0.63370.68470.63460.67030.59300.58680.51340.51990.53450.4972-
Random0.62940.67660.62110.66200.59390.57210.49070.51440.51020.483140.00
RouteLLM-BERT [23]0.63470.67920.63050.66510.59600.57880.49820.51600.51670.486671.51
RouteLLM-MF [23]0.63640.68140.62990.66600.59520.57760.49700.51640.51490.485069.90
Hybrid LLM [8]0.63270.67840.63060.66770.59640.57870.50080.51610.51910.486473.49
ZOOTER [21]0.63500.67960.63150.66720.59660.57880.50040.51660.51790.485477.95
RouteT2I (Ours)0.63500.67860.63180.66790.59750.58040.50100.51670.51890.486583.97
", + "bbox": [ + 94, + 88, + 903, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "matrix factorization, Hybrid LLM [8], and ZOOTER [21]. Specifically, we keep their text input processing and quality threshold design, but replace their optimization goal with the average image quality across metrics as they only handle single-objective optimization.", + "bbox": [ + 88, + 266, + 480, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unless specified in their paper, the routing models are based on Transformers and take the same hyperparameters as in our method. All routers are implemented in PyTorch 2.3.1 and trained using Adam optimizer with a learning rate of 2e-5, a weight decay of 0, a batch size of 16, on a NVIDIA 4090D for about 10 epochs.", + "bbox": [ + 89, + 342, + 483, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. For cost efficiency, we introduce routing rate, namely, the proportion of user prompts to the cloud model:", + "bbox": [ + 89, + 431, + 483, + 463 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np = \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 487, + 480, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Smaller routing rate indicates better cost efficiency.", + "bbox": [ + 89, + 513, + 429, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For image generation quality, the absolute performance of routing is constrained by the gap between edge and cloud T2I models. Thus, we introduce a set of relative performance metrics for evaluation. We first fix a routing rate $p$ and introduce the win rate of the selected models over the cloud model after routing:", + "bbox": [ + 89, + 529, + 483, + 619 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(\\mathcal {I} _ {r, p}\\right) = \\mathbb {P} \\left\\{\\mathcal {Q} \\left(I _ {r}\\right) \\geq \\mathcal {Q} \\left(I _ {c}\\right) \\mid I _ {r}, I _ {c} \\in \\mathcal {I} _ {r, p}, \\mathcal {I} _ {c} \\right\\}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 628, + 480, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{r,p}$ denotes the generated images after routing at a routing rate $p$ . To intuitively show the improvement in win rate, we introduce the normalized win rate improvement, which is the increase in win rate achieved by a routing method compared to the baseline, divided by the increase achieved by the optimal oracle routing. A higher value indicates that the routing method is closer to the oracle routing:", + "bbox": [ + 89, + 656, + 483, + 763 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\bar {w} (p) = \\frac {w \\left(\\mathcal {I} _ {r , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}{w \\left(\\mathcal {I} _ {o , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 782, + 483, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{I}_{b,p}$ and $\\mathcal{I}_{o,p}$ denote the generated images using the random baseline and oracle, respectively. To quantify the improvement of image quality metrics, we introduce relative performance improvement, which measures the proportion of the performance gain achieved through routing", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7a6d158b858116da566d7d33dfc4677aa4b9ad7fd05565206ea1aca3d6dfc02b.jpg", + "table_caption": [ + "Table 2. The multi-dimensional quality of images generated by text-to-image models, with the router selecting between edge and cloud models for each prompt at a $50\\%$ routing rate. The higher the metrics, the better." + ], + "table_footnote": [], + "table_body": "
Router\\p40%50%60%70%80%
RouteLLM-BERT [23]24.2922.4519.0717.6220.59
RouteLLM-MF [23]25.6523.2919.9216.6720.59
Hybrid LLM [8]23.7719.7514.9213.6216.09
ZOOTER [21]26.7721.9717.9716.9721.04
RouteT2I (Ours)30.6025.8120.3218.0221.94
", + "bbox": [ + 537, + 263, + 883, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Normalized win rate improvements $\\Delta \\overline{w}$ (\\%) at different routing rates, i.e. the ratio of improvement to oracle improvement.", + "bbox": [ + 511, + 351, + 905, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to the edge model, relative to the total possible improvement brought by fully using the cloud model:", + "bbox": [ + 511, + 404, + 905, + 434 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta P = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\mu_ {i} \\left(\\mathcal {I} _ {r}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right)}{\\left| \\mu_ {i} \\left(\\mathcal {I} _ {c}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right) \\right|}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 441, + 903, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This metric effectively quantifies routing effectiveness while accounting for the original quality gap.", + "bbox": [ + 511, + 489, + 905, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For cost-quality balance, we introduce the cost saving to measure the reduction of the routing rate to the cloud model, compared to the random baseline at a given relative performance improvement $\\Delta P$ :", + "bbox": [ + 511, + 521, + 905, + 580 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma (\\Delta P) = \\frac {p _ {b} (\\Delta P) - p _ {r} (\\Delta P)}{p _ {b} (\\Delta P)}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 587, + 903, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $p_b(\\Delta P)$ and $p_r(\\Delta P)$ represent the routing rates of baseline and router at a given $\\Delta P$ , respectively.", + "bbox": [ + 511, + 628, + 906, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 667, + 653, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present the routing performance of our RouteT2I and baselines using SD3 as the cloud model and SD2.1 as the edge model. Human evaluation, cost analysis, visual results, and more T2I pair results are provided in Appendix.", + "bbox": [ + 511, + 689, + 905, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Win Rates. Tab. 3 shows our RouteT2I demonstrates a significant improvements on normalized win rate improvements across all routing rates. Specifically, at a $40\\%$ routing rate, our improvement reaches $30.60\\%$ , exceeding the baselines by at least $3.83\\%$ and demonstrating strong consistency with the oracle.", + "bbox": [ + 511, + 750, + 905, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Quality. Tab. 2 shows RouteT2I outperforms all baselines in 6 of 10 metrics. Our multi-metric optimization balances all metrics simultaneously, while others focus on just a few, often at the cost of others. We", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19488", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5c8ca62473dac4f3f24a8621da42127503a220f825d59dd7fa1a8be75a44aecf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ΔP Router40%50%60%70%80%
RouteLLM-BERT [23]56.1551.3946.9242.7040.21
RouteLLM-MF [23]48.8649.9048.2044.8941.50
Hybrid LLM [8]62.0658.8553.6349.9233.38
ZOOTER [21]69.2865.7660.8157.3549.64
RouteT2I (Ours)71.8170.2466.6160.0153.53
", + "bbox": [ + 114, + 88, + 460, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69c67ba2cf1dbda5d25956402a4ade59774f323ac9a92217ee3331020bd77084.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 122, + 232, + 285, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/228111b9e8cc0d6ab2e87be145202cdd29122e322b9cc3680b153e94bbf2492a.jpg", + "image_caption": [ + "(b)", + "Figure 6. Visualization of (a) Normalized win rate improvement $\\Delta \\overline{w}$ and (b) relative performance improvement $\\Delta P$ at varying routing rates $p$ to the cloud." + ], + "image_footnote": [], + "bbox": [ + 289, + 232, + 452, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "also achieve the highest relative performance improvement, reaching $83.97\\%$ of cloud model's gain over edge model, surpassing others by at least $6\\%$ . Experiments in Appendix Sec. H show our method may lead to performance gains beyond those of the cloud model itself and is effective for T2I models with various text encoders and architectures.", + "bbox": [ + 89, + 416, + 483, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cost Efficiency. Tab. 4 presents the cost savings at given relative performance improvement targets. At a $40\\%$ relative performance improvement, RouteT2I reduce cloud service needs by up to $71.81\\%$ of the calls required by random routing and by at least $5.80\\%$ compared to other baselines. In other cases, RouteT2I still reduces over $50\\%$ costs, demonstrating high cost efficiency.", + "bbox": [ + 89, + 507, + 483, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. As shown in Fig. 6, RouteT2I performs well across routing rates, demonstrating its ability to predict quality gaps between T2I models and route effectively.", + "bbox": [ + 89, + 613, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual examples are shown in Appendix Sec. G.", + "bbox": [ + 109, + 657, + 429, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. Routing Performance on More T2I Model Pairs", + "text_level": 1, + "bbox": [ + 89, + 681, + 482, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To test our method across T2I model pairs, Tab. 5 shows RouteT2I's performance with various cloud and edge models. Results reveal significant improvements, especially for pairs with large quality gaps, such as SD3 on cloud and others on edge, where RouteT2I achieves over $25\\%$ of the oracle routing's improvement at a $40\\%$ routing rate. Even with similar model pairs like XL-Refiner and SDXL, it reaches $10\\%$ of the oracle's improvement, demonstrating its ability to discern subtle differences between T2I models.", + "bbox": [ + 89, + 704, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Experiment", + "text_level": 1, + "bbox": [ + 89, + 848, + 290, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 6 presents ablation results for RouteT2I's multi-metric quality optimization, token selection gate, and dual", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e57877800500c575387534514b260a4a44ea1440ba8f7fd416da6933e947f574.jpg", + "table_caption": [ + "Table 4. Cost saving $\\gamma$ (\\%) under a given relative performance improvements $(\\Delta P)$ target. $\\gamma$ is the ratio of savings to the routing rate required by the random baseline." + ], + "table_footnote": [], + "table_body": "
Cloud-Edgep40%50%60%70%80%
SD3XL-Refiner27.4222.3918.0516.8520.89
SD3SDXL29.7826.8122.3516.9821.30
SD3SD1.526.2420.8816.1914.0419.22
SD2.1XL-Refiner21.6218.3120.4625.3028.66
SD2.1SDXL19.4616.8717.2720.4223.69
SD2.1SD1.510.779.2810.1212.2215.76
XL-RefinerSDXL13.5710.159.7411.8112.59
XL-RefinerSD1.519.6516.9818.8722.7925.57
SDXLSD1.518.3817.0121.5724.7532.49
SD3PixArt-α32.2627.5924.7430.7338.97
PixArt-αSD2.125.8221.2622.5528.1933.76
PixArt-αSDXL16.7113.5712.3716.3522.14
PixArt-αXL-Refiner22.5119.1418.1823.7228.44
PixArt-αSD1.522.3820.5219.9325.4032.43
SD3Infinity22.1919.9717.7719.8125.39
SD2.1Infinity21.9819.8622.6028.4536.18
PixArt-αInfinity27.6923.4824.1230.8938.91
", + "bbox": [ + 517, + 89, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/406f07290d81eb9c80e0d6a285c6d7d30e678f078d0598396f04a08ccd6b6898.jpg", + "table_caption": [ + "Table 5. Normalized win rate improvement $\\Delta \\overline{w} (\\%)$ over oracle routing of our RouteT2I with different cloud-edge model pairs." + ], + "table_footnote": [], + "table_body": "
Router\\p40%50%60%70%80%
w/o Multi-Metric27.3722.8118.8716.4719.92
w/o Token Selection27.8223.0518.2716.7719.24
w/o Dual-Gate27.2222.0919.2717.3721.62
RouteT2I30.6025.8120.3218.0221.94
", + "bbox": [ + 537, + 340, + 883, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation experiment of RouteT2I, showing normalized win rate improvement $\\Delta \\overline{w}$ (%) on given routing rates.", + "bbox": [ + 511, + 421, + 906, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gate MoE. The multi-metric quality optimization allows routers to comprehensively assess the image quality gap, significantly enhancing overall performance. Removing this results in a performance decline across all routing rates, particularly at $40\\%$ routing rate, where the drop is approximately $3.23\\%$ . The token selection gate prioritizes critical tokens, ensuring routing focuses on key quality factors while reducing interference from irrelevant elements. Its removal notably impacts high routing rates, causing a performance decrease of around $2\\%$ at $80\\%$ routing rate. The dual-gate MoE effectively distinguishes subtle differences between cloud and edge model outputs, proving most effective at medium routing rates, especially near $50\\%$ .", + "bbox": [ + 511, + 479, + 906, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 691, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed a new text-to-image model routing design RouteT2I between edge and cloud. RouteT2I adopts the routing model with a dual-gate token selection MoE to predict how user prompts affect multiple image quality metrics by identifying and evaluating key tokens with contrastive method. RouteT2I further introduces a routing strategy that exploits the predicted advantage of the edge model over the cloud model to determine which side for image generation. Extensive evaluation has demonstrated that RouteT2I can significantly enhance generation quality at a specified routing rate and meanwhile reduce cost at a given quality target.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19489", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledge", + "text_level": 1, + "bbox": [ + 91, + 90, + 209, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Key R&D Program of China (No. 2022ZD0119100), China NSF grant No. 62202297, No. 62025204, No. 62202296, No. 62272293, No. 62441236, and No. U24A20326, SJTU-Huawei Research Program, and Tencent WeChat Research Program. The opinions, findings, conclusions, and recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agencies or the government.", + "bbox": [ + 89, + 114, + 485, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] stabilityai/stable-diffusion-3.5-large · Hugging Face — huggingface.co. https://huggingface.co/stabilityai/stable-diffusion-3.5-large.1", + "[2] Szymon Antoniak, Michal Krutul, Maciej Pioro, Jakub Krajewski, Jan Ludziejewski, Kamil Ciebiera, Krystian Król, Tomasz Odrzygóźdź, Marek Cygan, and Sebastian Jaszcur. Mixture of tokens: Continuous moe through cross-example aggregation. Advances in Neural Information Processing Systems, 37:103873-103896, 2025. 2", + "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3", + "[4] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 821–830, 2024. 1", + "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6", + "[6] Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023. 2", + "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2818-2829, 2023. 6", + "[8] Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618, 2024. 1, 2, 7, 8", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 6" + ], + "bbox": [ + 99, + 287, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 6", + "[11] Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. *Routerbench: A benchmark for multi-llm routing system.* arXiv preprint arXiv:2403.12031, 2024. 1", + "[12] Shijing Hu, Ruijun Deng, Xin Du, Zhihui Lu, Qiang Duan, Yi He, Shih-Chia Huang, and Jie Wu. Laecips: Large vision model assisted adaptive edge-cloud collaboration for iot-based perception system. arXiv preprint arXiv:2404.10498, 2024. 2", + "[13] Anil Kag and Igor Fedorov. Efficient edge inference by selective query. In International Conference on Learning Representations, 2023. 2", + "[14] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36:69981-70011, 2023. 4", + "[15] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[16] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1", + "[17] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[18] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 4", + "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[20] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 1", + "[21] Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692, 2023. 1, 2, 7, 8", + "[22] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19490", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1", + "[23] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data. arXiv preprint arXiv:2406.18665, 2024. 1, 2, 6, 7, 8", + "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 6", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6", + "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6", + "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3", + "[30] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 1", + "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6", + "[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2", + "[33] Tal Shnitzer, Anthony Ou, Mfrian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets. arXiv preprint arXiv:2309.15789, 2023. 2", + "[34] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for dif" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "fusion models. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[35] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 4", + "[36] Yiding Wang, Kai Chen, Haisheng Tan, and Kun Guo. Tabi: An efficient multi-level inference system for large language models. In Proceedings of the Eighteenth European Conference on Computer Systems, pages 233-248, 2023. 2", + "[37] Haoning Wu, Xiele Wu, Chunyi Li, Zicheng Zhang, Chaofeng Chen, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. T2i-scorer: Quantitative evaluation on text-to-image generation via fine-tuned large multi-modal models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3676-3685, 2024. 4", + "[38] Xun Wu, Shaohan Huang, Guolong Wang, Jing Xiong, and Furu Wei. Multimodal large language models make text-to-image generative models align better. Advances in Neural Information Processing Systems, 37:81287-81323, 2024. 4", + "[39] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3", + "[40] Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. Large language model cascades with mixture of thoughts representations for cost-efficient reasoning. arXiv preprint arXiv:2310.03094, 2023. 2", + "[41] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. arXiv preprint arXiv:2405.17873, 2024. 1", + "[42] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3", + "[43] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19491", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_model.json b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f10dbe49d24e11a4320db2ac71cc11bf37ce20b8 --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.204, + 0.731, + 0.24 + ], + "angle": 0, + "content": "Zewei Xin, Qinya Li*, Chaoyue Niu*, Fan Wu, Guihai Chen \nShanghai Jiao Tong University, China" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.242, + 0.822, + 0.258 + ], + "angle": 0, + "content": "{xin_zeweil12138，qinyali，rvince}@sjtu.edu.cn，{fwu，gchen}@cs.sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.716, + 0.222, + 0.731 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Nowadays, text-to-image (T2I) models like Imagen [3], Stable Diffusion [28], and DALL-E [27] have achieved significant success in generating diverse, high-quality images given user prompts. However, the impressive generation quality comes with large model and high cost. For example, Stable Diffusion 3.5 [1] has 8 billion parameters. Such a large model necessitates reliance on cloud servers, leading to high serving cost. As shown in Tab. 1, it is particularly costly in commercial scenarios with millions of requests." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.291, + 0.888, + 0.376 + ], + "angle": 0, + "content": "
Text-to-Image Model#ParamPricing ($/M)
Stable Diffusion 1.6 [28]0.86 B9 K
Stable Diffusion XL [24]2.6 B9 K
Stable Diffusion 3 Medium [9]2 B35 K
Stable Diffusion 3 [9]8 B65 K
Stable Diffusion 3.5 [1]8 B65 K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.387, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Table 1. For the Stable Diffusion models, newer versions come with more parameters and higher API costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.612 + ], + "angle": 0, + "content": "The high cost of cloud-based image generation drives the trend of deploying T2I models on edge devices. Techniques such as quantization [16, 34, 41], structural pruning [4, 17], and reducing denoising steps [20, 22, 30] are used to minimize model sizes and speed up inference, showing that edge deployment is feasible. Compared to cloud models, edge models leverage users' local computing devices to provide instant services anytime and anywhere without cloud serving cost and communication overhead. However, the lightweight edge T2I models often have lower generation quality compared to the cloud large models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Although cloud models often offer superior quality, not all user prompts need these large models. For some prompts, light-weight edge models can produce comparable or even better results. To ensure high-image quality generation while limiting requests to cloud models to reduce cost, it is necessary to introduce a routing mechanism such that suitable models are selected for different user prompts based on their complexity. Intuitively, as shown in Fig. 1, a routing framework for T2I models routes only challenging prompts to the cloud, while handling easy ones on the edge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Previous routing methods are primarily designed for large language models (LLMs) [8, 11, 21, 23], presenting significant challenges when applied to T2I models. One major challenge lies in comparing the quality of generated images. Unlike text, image quality does not have unified metrics and is subject to many influencing factors, including color, clarity, and the unique distortions typical of generated images [15]. This makes it hard to compare image quality to guide routing decisions. Furthermore, the output space of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.368, + 0.901 + ], + "angle": 0, + "content": "*Q. Li and C. Niu are the corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.228, + 0.454, + 0.243 + ], + "angle": 0, + "content": "Figure 1. Routing for text-to-image generation requests." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.273, + 0.483, + 0.319 + ], + "angle": 0, + "content": "T2I models is substantially larger than the input text space, and a single prompt can correspond to numerous images, hindering the prediction of image quality before generation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.32, + 0.484, + 0.805 + ], + "angle": 0, + "content": "To address these challenges, we propose an edge-cloud T2I routing framework RouteT2I, deploying a routing model and a small T2I model on the edge, and a large T2I model on the cloud. Our goal is to optimize the overall quality of generated images at a limited cost, by effectively routing user prompts to the edge or the cloud. To evaluate the quality gap between generated images, we first propose a multi-metric quality measure, including metrics for both real photos and those unique to generated images. Each metric assesses image quality by measuring the similarity between the image and a pair of positive and negative prompts that describe the quality metric. We then define Pareto relative superiority (PRS) to quantify quality differences between images. RouteT2I predicts PRS based on user prompts, inspired by the cross-attention operation between text and image during generation. In particular, inspired by mixture-of-experts (MoE) selective focus on experts or tokens in LLMs [2, 32, 43], we design a novel dual-gate token selection MoE, where a user prompt is treated as a sequence of tokens, with experts aligned to quality metrics. A token selection gate allows experts to proactively choose tokens that markedly impact these metrics, thereby focusing on key tokens. A dual-gate MoE uses both positive and negative gates to evaluate dominant effects when tokens have opposing effects simultaneously. Furthermore, multiple heads are introduced to predict multiple quality metrics. Based on the multi-metric quality prediction, RouteT2I determines quality differences between images generated on the edge over those in the cloud. Only user prompts that show substantial quality improvement when generated in the cloud are routed there, while others remain at the edge. Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We, for the first time, study how to route text-to-image generation requests between cloud and edge models for low cost and high quality." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Specified to multi-metric quality measures of generated images, we design a routing model architecture with dual gate token selection MoE, combined with a routing strat" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.093, + 0.672, + 0.107 + ], + "angle": 0, + "content": "egy for user prompts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.907, + 0.259 + ], + "angle": 0, + "content": "- We evaluate the proposed design using 18 image generation model pairs on the public COCO 2014 dataset. Key results include: (1) by routing \\(50\\%\\) of requests to edge and cloud models, respectively, the overall quality improvement over the edge model reaches \\(83.97\\%\\) of that achieved by fully using the cloud model; and (2) to achieve \\(50\\%\\) of the quality improvement attained by fully using the cloud model, the number of requests routed to the cloud model is reduced by \\(70.24\\%\\) compared to a random routing policy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.656, + 0.287 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Model routing selects suitable models for user requests to ensure quality while reducing high cost of invoking powerful models. Depending on whether routing decision occurs before or after weak model execution, model routing can be categorized into predictive and non-predictive routing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.909, + 0.584 + ], + "angle": 0, + "content": "Non-predictive routing takes the output of a weak model to decide whether to route to the next, powerful model. LAECIPS [12] evaluated the confidence of predictions made by the edge semantic segmentation model to decide if cloud assistance from a large visual model is needed. For classification models, Hybrid [13] used a routing model to partition the data domain between edge and cloud models, routing cases with incorrect edge outputs to the cloud. For cascading LLM, LLM Cascade [40] routed based on the consistency of responses from weaker LLMs, while FrugalGPT [6] considered current quality and past costs to determine routing. Tabi [36] not only routed based on the confidence of responses from weaker LLMs but also enhanced output quality by aggregating historic outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.811 + ], + "angle": 0, + "content": "Different from non-predictive routing, predictive routing selects model before running the weak model on resource-constrained edge device. Previous studies focused on LLM, with the difference in routing strategies. RouteLLM [23] predicted the outcomes of quality comparisons between LLM and routed based on prediction confidence. Hybrid LLM [8] relaxed comparisons, allowing the weak model to succeed if the quality gap is within a threshold, thus saving costs with some quality compromise. Recent research [33] advances the confidence routing strategy by identifying out-of-distribution data through confidence thresholds. ZOOTER [21] introduced a new strategy by predicting the normalized quality of candidate model outputs, making routing decision based on relative quality, with routing models distilled from existing quality scoring models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Previous routing work were mainly for LLMs, whereas the predictive routing for T2I in our work remains unexplored. A key distinction lies in the image outputs of T2I, which necessitate predicting quality variations in a larger output image space from a smaller input text space (Sec. B in Appendix). On one hand, unlike many LLM tasks with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19483" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.092, + 0.162, + 0.101 + ], + "angle": 0, + "content": "SD2.1" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.092, + 0.244, + 0.101 + ], + "angle": 0, + "content": "SD3" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.101, + 0.287, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.092, + 0.35, + 0.101 + ], + "angle": 0, + "content": "SD2.1" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.102, + 0.385, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.092, + 0.436, + 0.101 + ], + "angle": 0, + "content": "SD3" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.102, + 0.477, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.173, + 0.266, + 0.182 + ], + "angle": 0, + "content": "There is a dog lying on the ground." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.182, + 0.285, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.255, + 0.276, + 0.283 + ], + "angle": 0, + "content": "A cat sitting motionless on a television stand behind a mug in the shape of a skeleton head." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.182, + 0.385, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.182, + 0.477, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.255, + 0.471, + 0.283 + ], + "angle": 0, + "content": "A sandwich cut in half on a white plate with a side of salad and a fork and knife wrapped in a napkin." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.337 + ], + "angle": 0, + "content": "Figure 2. SD2.1 tends to match or outperform SD3 on simple prompts, while more likely falling short on complex ones. Entities to be generated are **bolded**." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.351, + 0.285, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.183, + 0.467, + 0.2, + 0.477 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.353, + 0.477, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.466, + 0.394, + 0.477 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Figure 3. Impact of the noun count in text prompts on ImageReward score [39] gap and win rate between SD3 and SD2.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.651 + ], + "angle": 0, + "content": "definitive answers, image quality lacks fixed standards and is highly subjective. Previous single-objective LLM routing struggles to adapt to such ambiguous quality criteria. On the other hand, prior work lacked dedicated routing model designs, making it challenging to predict quality relationships within the vast output image space. This underscores the necessity of effective routing model design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.297, + 0.68 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.856 + ], + "angle": 0, + "content": "In text-to-image generation scenario, the cloud normally hosts a large T2I model, denoted as \\(\\mathcal{M}_c\\), while a resource-constraint edge device, such as a smartphone, often deploys a light-weight T2I model, denoted as \\(\\mathcal{M}_e\\). For example, Google has released Imagen [3, 29] for cloud deployment and MobileDiffusion [42] for edge deployment. From the perspective of a user, the large model generally offers superior generation quality, but incurs edge-cloud communication cost and cloud serving fee, which together are denoted as \\(F_c\\). In contrast, the edge model can leverage the user's local device, thereby eliminating these additional costs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In practice, user requests involve not only complex ones but also simple ones. One intuitive factor influencing request complexity is the number of entities that need to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "be generated in images, i.e. the number of nouns in text prompts. As shown in Fig. 2, 3, as noun count increases, the large model is more likely to outperform the small model in quality, with quality differences becoming more noticeable. However, while large models generally excel in overall quality, small models may perform better on specific prompts. This likelihood increases as the prompt becomes simpler. In such cases, the purely large model serving not only fails to improve performance but also incurs high cost. Therefore, it is necessary to design a routing mechanism to select the most suitable model based on the prompt's characteristics. Intuitively, the routing mechanism should direct simpler prompts to the cost-effective edge model and more complex prompts to the high-quality cloud model, thus maximizing generation quality while minimizing cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Formally, a T2I routing framework \\( R: \\mathcal{X} \\to \\{0,1\\} \\) assigns user prompts \\( \\mathcal{X} \\) to the cloud large model \\( \\mathcal{M}_c \\) as 1 and to the edge light-weight model \\( \\mathcal{M}_e \\) as 0. The generated images after routing can be expressed as \\( \\mathcal{I}_r = R(\\mathcal{X})\\mathcal{I}_c + (1 - R(\\mathcal{X}))\\mathcal{I}_e \\), where \\( \\mathcal{I}_c \\) and \\( \\mathcal{I}_e \\) denote the images generated using \\( \\mathcal{M}_c \\) and \\( \\mathcal{M}_e \\), respectively. Given a quality scoring function \\( \\mathcal{Q}(\\cdot) \\) of the generated images, the cost budget \\( \\tau_{fee} \\) of cloud model serving and edge-cloud communication, and the response latency constraint \\( \\tau_{time} \\), the optimization objective is maximizing the quality under the budget and latency constraints, formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.493, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot F _ {c} \\leq \\tau_ {f e e} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} - D _ {c} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} - D _ {c} \\end{array} \\tag {1} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot D _ {\\mathcal {M} _ {c}} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} \\cdot D _ {\\mathcal {M} _ {e}} \\\\ + D _ {R} \\leq \\tau_ {\\text {t i m e}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.681 + ], + "angle": 0, + "content": "where \\(D_{\\mathcal{M}_c}\\), \\(D_{\\mathcal{M}_e}\\), and \\(D_R\\) denote the latency of cloud large model serving plus edge-cloud communication, the latency of the edge light-weight model serving, and the latency of the routing model execution. Considering the constraints are linear, the optimization objective can be approximated as imposing an upper bound \\(\\rho_r\\) on the routing rate to the cloud large model as" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.691, + 0.905, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\leq \\rho_ {r}. \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Due to the lack of a universal standard for evaluating the quality of generated images, a multi-dimensional quality scoring function \\(\\mathcal{Q}:\\mathcal{I}\\to [0,1]^N\\) is required for a more comprehensive comparison. Thus, different from existing routing for typical classification and text generation tasks, the T2I routing framework's optimization objective is a new multi-objective problem. In addition, to predict quality changes in the larger image output space from the smaller text input space, we conduct a detailed analysis of how different prompt tokens influence the generated images and their multi-dimensional quality metrics, by examining" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19484" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "the interaction between text and images during the T2I generation process. In what follow, we define the multi-metric quality measures for routing objectives in Sec. 4 and introduce the T2I routing framework in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.163, + 0.483, + 0.199 + ], + "angle": 0, + "content": "4. Multi-Metric Image Generation Quality in Routing Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.297 + ], + "angle": 0, + "content": "The connection between text and images allows image attributes to be described through corresponding prompts. Thus, a metric can be measured by the relative relationship between an image and the positive/negative text pair describing that metric [18, 35]. The contrastive quality metric for image \\(I\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.304, + 0.483, + 0.321 + ], + "angle": 0, + "content": "\\[\nq (I, m) = \\sigma \\left(C L I P \\left(I, m ^ {+}\\right) - C L I P \\left(I, m ^ {-}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.463 + ], + "angle": 0, + "content": "where \\( m = (m^{+}, m^{-}) \\) denotes positive and negative pair, and \\( \\sigma \\) denotes the sigmoid function that transforms output values in the range from 0 to 1. According to this definition, if an image is more related to the positive prompt, and is less related to the negative prompt, the contrastive quality metric is higher. Compared to quality measure with only the positive prompt, the contrastive method evaluates whether positive or negative quality is more dominant in the image, leading to a more robust and reliable evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Considering the ambiguity and complexity of image quality, we introduce a multi-dimensional metric to comprehensively evaluate generated image quality, providing noise resistance and stability for quality-prediction-based routing. Just as real photo quality depends on factors like definition and color, our multi-metric combines traditional measures of real photos with unique aspects of generated images, such as realism and object integrity. Thus, we extend the single-metric evaluation in Eq. (3) to an \\(N\\)-dimensional quality assessment of an image \\(I\\), expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.621, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} (I) = \\left[ q (I, m _ {i}) \\mid i = 1, 2, \\dots , N \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.75 + ], + "angle": 0, + "content": "where \\( m_{i} \\) is an instantiation of \\( m \\) in Eq. (3). We adopt 10 common metrics inspired by prior works [14, 37, 38], with details in Appendix Sec. C. For example, for the definition metric, we use the positive prompt \"High definition photo\" and the negative prompt \"Low definition photo\"; and for the object integrity, we take the positive prompt \"Object completion\" and the negative prompt \"Object twisting\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the multi-objective optimization setting for routing with multi-metric quality, the initial goal is to find a Pareto optimal generated image that excels in all the metrics. However, this is challenging in practice due to the potential nonexistence of such an image generated from the cloud or edge model. We thus relax the constraint of Pareto optimality by allowing suboptimal performance in some metrics if the image significantly outperforms in others when selecting higher-quality images. We define Pareto relative superiority (PRS) to quantify the quality advantage of the images" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.905, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.269, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 4. Overview of RouteT2I. RouteT2I assembles a pair of off-the-shelf edge and cloud text-to-image models and evaluates them using multi-metric image quality. RouteT2I utilizes Pareto relative superiority between qualities as supervision to train the routing model. Then, to balance quality and cost, the routing strategy determines the most suitable model for each user prompt, choosing between the edge or cloud model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.906, + 0.455 + ], + "angle": 0, + "content": "\\(\\mathcal{I}_e\\) generated by the edge model over the images \\(\\mathcal{I}_c\\) generated by the cloud model. In particular, we first normalize the quality distance for the metric \\(i\\) between an edge-generated image \\(I_e \\in \\mathcal{I}_e\\) and a cloud-generated image \\(I_c \\in \\mathcal{I}_c\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.465, + 0.906, + 0.5 + ], + "angle": 0, + "content": "\\[\nD _ {i} (I _ {e}, I _ {c}) = \\sigma \\left(\\frac {q (I _ {e} , m _ {i}) - q (I _ {c} , m _ {i})}{\\Gamma | \\mu_ {i} (\\mathcal {I} _ {e}) - \\mu_ {i} (\\mathcal {I} _ {c}) |}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mu_i(\\cdot)\\) is the average quality of the set of cloud or edge generated images, while the sigmoid function \\(\\sigma\\) and the temperature parameter \\(\\Gamma\\) are used to modulate the data distribution, effectively distinguishing similar qualities and preventing centralization. We then define the overall quality gap as a weighted sum of distances across all \\(N\\) metrics:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.612, + 0.905, + 0.655 + ], + "angle": 0, + "content": "\\[\nP R S \\left(I _ {e}, I _ {c}\\right) = \\sum_ {i = 1} ^ {N} w _ {i} D _ {i} \\left(I _ {e}, I _ {c}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.743 + ], + "angle": 0, + "content": "where \\( w_{i} \\) denotes the weight of metric \\( i \\), and \\( \\sum_{i} w_{i} = 1 \\). We can evaluate the relative quality of \\( I_{e} \\) compared to \\( I_{c} \\) by examining how much Pareto relative superiority deviates from 0.5. This evaluation helps to route a user prompt to the model that generates an image with higher overall quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.715, + 0.772 + ], + "angle": 0, + "content": "5. Design of RouteT2I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To achieve the optimization objective in Eq. (2), we propose a text-to-image model routing framework RouteT2I. As shown in Fig. 4, RouteT2I comprises a routing model that predicts the multi-metric generation quality given user prompts and a routing strategy that selects cloud or edge model. In particular, (1) the routing model predicts the quality relationship between images generated on the edge and cloud, namely, Pareto relative superiority. Inspired by the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19485" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.148, + 0.09, + 0.429, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.201, + 0.352, + 0.214 + ], + "angle": 0, + "content": "(a) Token selection gate." + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.847, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.201, + 0.708, + 0.214 + ], + "angle": 0, + "content": "(b) Dual-gate MoE." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.226, + 0.908, + 0.269 + ], + "angle": 0, + "content": "Figure 5. Architecture of a dual-gate MoE that uses the token selection gate as the gate network. First, the gate selects the most relevant tokens for each expert. Then, experts project the tokens selected by positive and negative gates into their respective spaces and evaluate them using score matrices. Finally, the evaluations of the tokens are gathered and compared to form new tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.295, + 0.486, + 0.492 + ], + "angle": 0, + "content": "role of prompts in generation, where prompt as sequences of tokens interact with images through cross-attention to determine the image content, we treat prompt as a set of tokens in the routing model and select key tokens with influential contextual information for each quality metric. Then, the dominant influence of these tokens is accessed by comparing their positive and negative effects; and (2) the routing strategy aims to maximize quality under cost constraints. The key idea is to use Pareto relative superiority to describe the quality disparity of images generated by the cloud and the edge models. Only prompts that show a notable quality gap when generated in the cloud are routed there, while the rest remain on the cost-effective edge." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.531 + ], + "angle": 0, + "content": "5.1. Routing Model Architecture Design with Dual-Gate Token Selection MoE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.614 + ], + "angle": 0, + "content": "To predict Pareto relative superiority between the images generated by the cloud model and the edge model from user prompts, we design a dual-gate token selection mixture-of-experts (MoE), as shown in Fig. 5, and introduce it into Transformer network to replace linear layers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Our model treats the user prompt as a sequence of tokens, whose influences on the image quality differ due to their weights in cross-attention during generation. To focus on key tokens for each metric, we design a token selection gate, where experts align with quality metrics and actively choose the most relevant tokens to simulate metric attention to different image attributes. Only tokens that significantly impact the quality metrics corresponding to the experts are selected, while those with minimal influences are disregarded, reducing potential interference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Regarding the design of a dual-gate MoE, both positive and negative gates are introduced to distinguish the opposing aspects of token influences on image quality. Similar to images that have both superior and inferior regions when evaluated by metrics, these tokens are assessed by the corresponding experts for their positive and negative influences to each metric. Additionally, we introduce a comparison between two gates to identify which influence of each token predominates overall. We further improve the structure of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.324 + ], + "angle": 0, + "content": "expert to handle both positive and negative influences on a single metric simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.416 + ], + "angle": 0, + "content": "To support multi-metric quality prediction, our design incorporates multiple prediction heads within the model architecture. Each head outputs the prediction of a specific quality metric. Such multi-head design not only enhances noise resistance, thereby increasing robustness, but also facilitates a comprehensive evaluation of quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.708, + 0.434 + ], + "angle": 0, + "content": "5.1.1. Token Selection Gate" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.485 + ], + "angle": 0, + "content": "In the token selection gate, the dot product between the token representations \\( T \\in \\mathbb{R}^{n \\times d} \\) and the expert embeddings \\( E \\in \\mathbb{R}^{k \\times d} \\) is computed as an affinity matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.489, + 0.905, + 0.508 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {S o f t m a x} \\left(T \\cdot E ^ {T}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.906, + 0.604 + ], + "angle": 0, + "content": "where \\( n \\) denotes the number of tokens in user prompt, \\( k \\) denotes the number of experts, \\( d \\) denotes the hidden dimension, and the Softmax function is applied along the expert axis. \\( A[t,i] \\) in the affinity matrix captures the correlation between token \\( t \\) and expert \\( i \\). Then, top- \\( K \\) tokens with the highest relevance for each expert are selected" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.61, + 0.905, + 0.627 + ], + "angle": 0, + "content": "\\[\nM = \\operatorname {T o p} - \\mathrm {K} (A), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.86 + ], + "angle": 0, + "content": "where \\( M \\in \\mathbb{R}^{n \\times k} \\) is the binary mask matrix such that \\( M[t, i] = 1 \\) denotes token \\( t \\) is selected by expert \\( i \\), and \\( M[t, i] = 0 \\) denotes not selected. Experts tend to have similar weights because the metrics they are responsible for are weighted similarly. This ensures that the tokens selected by these experts possess significant global importance. This practice minimizes interference from tokens that are insignificant across all metrics. The influences of key tokens are further evidenced by their frequent selection through the top- \\( K \\) function and their corresponding scores within the affinity matrix. Moreover, by allocating an equal number of \\( K \\) tokens to each expert, workloads remain balanced, and bias in quality dimensions is avoided, ultimately resulting in more comprehensive and reliable quality prediction. Overall, the token selection gate is summarized as" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.881, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\nA, M = G (T, E), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19486" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.323, + 0.106 + ], + "angle": 0, + "content": "where \\(G\\) denotes the gate network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.113, + 0.248, + 0.127 + ], + "angle": 0, + "content": "5.1.2. Dual-Gate MoE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.132, + 0.484, + 0.207 + ], + "angle": 0, + "content": "Both a positive gate \\( G^{+} \\) and a negative gate \\( G^{-} \\) are introduced for different experts to select tokens that have positive and negative impacts on their corresponding quality metrics, while capturing the weights of their influences. The token selection results are denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.216, + 0.482, + 0.233 + ], + "angle": 0, + "content": "\\[\nA ^ {o}, M ^ {o} = G ^ {o} \\left(T, E ^ {o}\\right), \\text {w h e r e} o \\in \\{+, - \\}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.484, + 0.317 + ], + "angle": 0, + "content": "The affinity matrix \\( A^o \\) and the binary mask matrix \\( M^o \\) are captured by the gate \\( G^o \\) for the positive or negative aspect using the corresponding expert embedding \\( E^o \\). The tokens selected by two gates are then evaluated by the corresponding positive or negative experts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.604 + ], + "angle": 0, + "content": "To eliminate the redundancy of having separate experts for evaluating both positive and negative aspects of the same quality metric, the structure of experts is revised to allow an expert to assess both aspects simultaneously. This revision acknowledges that both positive and negative evaluations are inherently part of the same quality metric. Specifically, the original linear layer \\( W_{i} \\in \\mathbb{R}^{d \\times h} \\) of expert on metric \\( i \\) is split into two projection matrices \\( P_{i}^{+}, P_{i}^{-} \\in \\mathbb{R}^{d \\times l} \\) and one score matrix \\( S_{i} \\in \\mathbb{R}^{l \\times h} \\). Two projection matrices are specific to positive and negative aspects, projecting tokens into the corresponding representations in an \\( l \\)-dimensional low-dimensional space. Since \\( l \\ll h, d \\), the parameter size and the computation cost are reduced from \\( O(hd) \\) to \\( O(l(h + d)) \\), allowing an expert to efficiently focus on both positive and negative aspects of a metric with a low cost. The shared score matrix then evaluates the impact of these representations on metric \\( i \\). Thus, the impact of \\( t \\)-th token \\( T[t] \\in \\mathbb{R}^d \\) on quality metric \\( i \\) for positive or negative aspect \\( o \\) is described as" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.614, + 0.482, + 0.631 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {o} [ t ] = T [ t ] \\cdot P _ {i} ^ {o} \\cdot S _ {i}, \\text {w h e r e} o \\in \\{+, - \\}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.73 + ], + "angle": 0, + "content": "By separating the projection matrices and score matrix, we can evaluate both positive and negative impacts on a metric within the same expert. Then, the overall positive or negative impact of the \\( t \\)-th token on image quality can be obtained by a weighted sum of its impacts for corresponding aspects across all metrics" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.74, + 0.482, + 0.78 + ], + "angle": 0, + "content": "\\[\nT ^ {o} [ t ] = \\sum_ {i = 0} ^ {k} \\lambda_ {t, i} ^ {o} \\cdot T _ {i} ^ {o} [ t ], \\text {w h e r e} o \\in \\{+, - \\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.872 + ], + "angle": 0, + "content": "Here, \\(\\lambda_{t,i}^{o} = \\frac{\\mathbb{1}\\{M^{o}[t,i] = 1\\}A^{o}[t,i]}{\\sum_{j = 0}^{k}\\mathbb{1}\\{M^{o}[t,j] = 1\\}A^{o}[t,j]}\\) denotes the normalized affinity of the \\(i\\)-th expert for the \\(t\\)-th token. By contrasting the impact of a token from both positive and negative aspects, we can assess the token's predominant influence on the image quality as" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.88, + 0.482, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\hat {T} [ t ] = \\sigma \\left(T ^ {+} [ t ] - T ^ {-} [ t ]\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "The introduction of contrasting helps diminish ambiguity in predictions, especially when tokens exhibit influences on multiple metrics from both positive and negative aspects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.145, + 0.74, + 0.162 + ], + "angle": 0, + "content": "5.2. Routing Strategy Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Given the Pareto relative superiority predicted by routing model, we propose a routing strategy to effectively route user prompts to suitable models to achieve the trade-off between quality and costs. With a preset routing rate, we can determine the proportion of user prompts sent to the cloud, focusing on those where the cloud model significantly outperforms the edge model in quality. Since Pareto relative superiority represents this quality gap, we can efficiently filter prompts by setting a threshold \\(\\alpha\\) on it. In particular, the prompts with a Pareto relative superiority above the threshold are better handled by the edge model for cost efficiency, while those below the threshold achieve superior quality when processed in the cloud. Thus, the optimization objective in Eq. (2) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.387, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\alpha \\leq 1 / 2} \\mathbb {P} \\left\\{P R S \\left(I _ {e}, I _ {c}\\right) < \\alpha \\mid I _ {e}, I _ {c} \\in \\mathcal {I} _ {e}, \\mathcal {I} _ {c} \\right\\} \\leq \\rho_ {r}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.419, + 0.906, + 0.495 + ], + "angle": 0, + "content": "Since Pareto relative superiority being greater than or less than \\(1/2\\) indicates a relative advantage or disadvantage, respectively, we set a upper bound \\(1/2\\) on \\(\\alpha\\) to prevent user prompts with images generated at the edge of better quality from being routed to the cloud." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.64, + 0.525 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.704, + 0.548 + ], + "angle": 0, + "content": "6.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Dataset. We take COCO2014 [19], a comprehensive dataset for object detection, segmentation, and captioning tasks. We select its captions as user prompts \\(\\mathcal{X}\\). Given these prompts, we generate images \\(\\mathcal{I}_e, \\mathcal{I}_c\\) using different open-source text-to-image models. Similarly, we use the LAION dataset [31], with results shown in Appendix Sec. F." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Models. We take T2I models with varying performances and sizes for cloud and edge usage, include Stable Diffusion 1.5 (SD1.5) [28], Stable Diffusion 2.1 (SD2.1) [28], Stable Diffusion XL (SDXL) [24], Stable Diffusion XL-Refiner (XL-Refiner) [24], Stable Diffusion 3 (SD3) [9], PixArt\\(\\alpha\\) [5], and Infinity [10]. These models use various text encoders, such as CLIP [25], OpenCLIP [7], and T5 [26], and are based on architectures like diffusion, diffusion Transformer and autoregressive. Details are in Appendix Sec. A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Baselines and Settings. We first introduce random routing as a baseline, where prompts are randomly assigned to candidate models, given a specified routing rate to the cloud. Additionally, although there does not exist any previous work on text-to-image generation routing, for a comprehensive comparison, we also reproduce several representative routing methods for LLM by adapting them to our scenarios, including RouteLLM [23] with BERT classifier or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19487" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.201 + ], + "angle": 0, + "content": "
RouterImage Quality MetricsΔP(%)
DefinitionDetailClaritySharpnessHarmonyRealismColorConsistencyLayoutIntegrity
Edge Model0.62510.66850.60760.65370.59490.55750.46800.50880.48600.4690-
Cloud Model0.63370.68470.63460.67030.59300.58680.51340.51990.53450.4972-
Random0.62940.67660.62110.66200.59390.57210.49070.51440.51020.483140.00
RouteLLM-BERT [23]0.63470.67920.63050.66510.59600.57880.49820.51600.51670.486671.51
RouteLLM-MF [23]0.63640.68140.62990.66600.59520.57760.49700.51640.51490.485069.90
Hybrid LLM [8]0.63270.67840.63060.66770.59640.57870.50080.51610.51910.486473.49
ZOOTER [21]0.63500.67960.63150.66720.59660.57880.50040.51660.51790.485477.95
RouteT2I (Ours)0.63500.67860.63180.66790.59750.58040.50100.51670.51890.486583.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.212, + 0.908, + 0.242 + ], + "angle": 0, + "content": "Table 2. The multi-dimensional quality of images generated by text-to-image models, with the router selecting between edge and cloud models for each prompt at a \\(50\\%\\) routing rate. The higher the metrics, the better." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.482, + 0.342 + ], + "angle": 0, + "content": "matrix factorization, Hybrid LLM [8], and ZOOTER [21]. Specifically, we keep their text input processing and quality threshold design, but replace their optimization goal with the average image quality across metrics as they only handle single-objective optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Unless specified in their paper, the routing models are based on Transformers and take the same hyperparameters as in our method. All routers are implemented in PyTorch 2.3.1 and trained using Adam optimizer with a learning rate of 2e-5, a weight decay of 0, a batch size of 16, on a NVIDIA 4090D for about 10 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.464 + ], + "angle": 0, + "content": "Metrics. For cost efficiency, we introduce routing rate, namely, the proportion of user prompts to the cloud model:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.488, + 0.482, + 0.505 + ], + "angle": 0, + "content": "\\[\np = \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.431, + 0.529 + ], + "angle": 0, + "content": "Smaller routing rate indicates better cost efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.621 + ], + "angle": 0, + "content": "For image generation quality, the absolute performance of routing is constrained by the gap between edge and cloud T2I models. Thus, we introduce a set of relative performance metrics for evaluation. We first fix a routing rate \\( p \\) and introduce the win rate of the selected models over the cloud model after routing:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.63, + 0.482, + 0.649 + ], + "angle": 0, + "content": "\\[\nw \\left(\\mathcal {I} _ {r, p}\\right) = \\mathbb {P} \\left\\{\\mathcal {Q} \\left(I _ {r}\\right) \\geq \\mathcal {Q} \\left(I _ {c}\\right) \\mid I _ {r}, I _ {c} \\in \\mathcal {I} _ {r, p}, \\mathcal {I} _ {c} \\right\\}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{r,p}\\) denotes the generated images after routing at a routing rate \\(p\\). To intuitively show the improvement in win rate, we introduce the normalized win rate improvement, which is the increase in win rate achieved by a routing method compared to the baseline, divided by the increase achieved by the optimal oracle routing. A higher value indicates that the routing method is closer to the oracle routing:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.784, + 0.484, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\Delta \\bar {w} (p) = \\frac {w \\left(\\mathcal {I} _ {r , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}{w \\left(\\mathcal {I} _ {o , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}_{b,p}\\) and \\(\\mathcal{I}_{o,p}\\) denote the generated images using the random baseline and oracle, respectively. To quantify the improvement of image quality metrics, we introduce relative performance improvement, which measures the proportion of the performance gain achieved through routing" + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.265, + 0.885, + 0.342 + ], + "angle": 0, + "content": "
Router\\p40%50%60%70%80%
RouteLLM-BERT [23]24.2922.4519.0717.6220.59
RouteLLM-MF [23]25.6523.2919.9216.6720.59
Hybrid LLM [8]23.7719.7514.9213.6216.09
ZOOTER [21]26.7721.9717.9716.9721.04
RouteT2I (Ours)30.6025.8120.3218.0221.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.352, + 0.906, + 0.381 + ], + "angle": 0, + "content": "Table 3. Normalized win rate improvements \\( \\Delta \\overline{w} \\) (\\%) at different routing rates, i.e. the ratio of improvement to oracle improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.435 + ], + "angle": 0, + "content": "compared to the edge model, relative to the total possible improvement brought by fully using the cloud model:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.443, + 0.905, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\Delta P = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\mu_ {i} \\left(\\mathcal {I} _ {r}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right)}{\\left| \\mu_ {i} \\left(\\mathcal {I} _ {c}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right) \\right|}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.491, + 0.906, + 0.521 + ], + "angle": 0, + "content": "This metric effectively quantifies routing effectiveness while accounting for the original quality gap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.581 + ], + "angle": 0, + "content": "For cost-quality balance, we introduce the cost saving to measure the reduction of the routing rate to the cloud model, compared to the random baseline at a given relative performance improvement \\(\\Delta P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.588, + 0.905, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\gamma (\\Delta P) = \\frac {p _ {b} (\\Delta P) - p _ {r} (\\Delta P)}{p _ {b} (\\Delta P)}, \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.908, + 0.661 + ], + "angle": 0, + "content": "where \\(p_b(\\Delta P)\\) and \\(p_r(\\Delta P)\\) represent the routing rates of baseline and router at a given \\(\\Delta P\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.654, + 0.683 + ], + "angle": 0, + "content": "6.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.75 + ], + "angle": 0, + "content": "We present the routing performance of our RouteT2I and baselines using SD3 as the cloud model and SD2.1 as the edge model. Human evaluation, cost analysis, visual results, and more T2I pair results are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Win Rates. Tab. 3 shows our RouteT2I demonstrates a significant improvements on normalized win rate improvements across all routing rates. Specifically, at a \\(40\\%\\) routing rate, our improvement reaches \\(30.60\\%\\), exceeding the baselines by at least \\(3.83\\%\\) and demonstrating strong consistency with the oracle." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Image Quality. Tab. 2 shows RouteT2I outperforms all baselines in 6 of 10 metrics. Our multi-metric optimization balances all metrics simultaneously, while others focus on just a few, often at the cost of others. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19488" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.461, + 0.168 + ], + "angle": 0, + "content": "
ΔP Router40%50%60%70%80%
RouteLLM-BERT [23]56.1551.3946.9242.7040.21
RouteLLM-MF [23]48.8649.9048.2044.8941.50
Hybrid LLM [8]62.0658.8553.6349.9233.38
ZOOTER [21]69.2865.7660.8157.3549.64
RouteT2I (Ours)71.8170.2466.6160.0153.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.22 + ], + "angle": 0, + "content": "Table 4. Cost saving \\(\\gamma\\) (\\%) under a given relative performance improvements \\((\\Delta P)\\) target. \\(\\gamma\\) is the ratio of savings to the routing rate required by the random baseline." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.233, + 0.287, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.327, + 0.214, + 0.338 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.233, + 0.454, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.327, + 0.38, + 0.338 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Figure 6. Visualization of (a) Normalized win rate improvement \\(\\Delta \\overline{w}\\) and (b) relative performance improvement \\(\\Delta P\\) at varying routing rates \\(p\\) to the cloud." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.507 + ], + "angle": 0, + "content": "also achieve the highest relative performance improvement, reaching \\(83.97\\%\\) of cloud model's gain over edge model, surpassing others by at least \\(6\\%\\). Experiments in Appendix Sec. H show our method may lead to performance gains beyond those of the cloud model itself and is effective for T2I models with various text encoders and architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Cost Efficiency. Tab. 4 presents the cost savings at given relative performance improvement targets. At a \\(40\\%\\) relative performance improvement, RouteT2I reduce cloud service needs by up to \\(71.81\\%\\) of the calls required by random routing and by at least \\(5.80\\%\\) compared to other baselines. In other cases, RouteT2I still reduces over \\(50\\%\\) costs, demonstrating high cost efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Visualization. As shown in Fig. 6, RouteT2I performs well across routing rates, demonstrating its ability to predict quality gaps between T2I models and route effectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.659, + 0.431, + 0.675 + ], + "angle": 0, + "content": "Visual examples are shown in Appendix Sec. G." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.699 + ], + "angle": 0, + "content": "6.3. Routing Performance on More T2I Model Pairs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.841 + ], + "angle": 0, + "content": "To test our method across T2I model pairs, Tab. 5 shows RouteT2I's performance with various cloud and edge models. Results reveal significant improvements, especially for pairs with large quality gaps, such as SD3 on cloud and others on edge, where RouteT2I achieves over \\(25\\%\\) of the oracle routing's improvement at a \\(40\\%\\) routing rate. Even with similar model pairs like XL-Refiner and SDXL, it reaches \\(10\\%\\) of the oracle's improvement, demonstrating its ability to discern subtle differences between T2I models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.291, + 0.866 + ], + "angle": 0, + "content": "6.4. Ablation Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Tab. 6 presents ablation results for RouteT2I's multi-metric quality optimization, token selection gate, and dual" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.288 + ], + "angle": 0, + "content": "
Cloud-Edgep40%50%60%70%80%
SD3XL-Refiner27.4222.3918.0516.8520.89
SD3SDXL29.7826.8122.3516.9821.30
SD3SD1.526.2420.8816.1914.0419.22
SD2.1XL-Refiner21.6218.3120.4625.3028.66
SD2.1SDXL19.4616.8717.2720.4223.69
SD2.1SD1.510.779.2810.1212.2215.76
XL-RefinerSDXL13.5710.159.7411.8112.59
XL-RefinerSD1.519.6516.9818.8722.7925.57
SDXLSD1.518.3817.0121.5724.7532.49
SD3PixArt-α32.2627.5924.7430.7338.97
PixArt-αSD2.125.8221.2622.5528.1933.76
PixArt-αSDXL16.7113.5712.3716.3522.14
PixArt-αXL-Refiner22.5119.1418.1823.7228.44
PixArt-αSD1.522.3820.5219.9325.4032.43
SD3Infinity22.1919.9717.7719.8125.39
SD2.1Infinity21.9819.8622.6028.4536.18
PixArt-αInfinity27.6923.4824.1230.8938.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.297, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Table 5. Normalized win rate improvement \\( \\Delta \\overline{w} (\\%) \\) over oracle routing of our RouteT2I with different cloud-edge model pairs." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.342, + 0.885, + 0.412 + ], + "angle": 0, + "content": "
Router\\p40%50%60%70%80%
w/o Multi-Metric27.3722.8118.8716.4719.92
w/o Token Selection27.8223.0518.2716.7719.24
w/o Dual-Gate27.2222.0919.2717.3721.62
RouteT2I30.6025.8120.3218.0221.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.422, + 0.907, + 0.452 + ], + "angle": 0, + "content": "Table 6. Ablation experiment of RouteT2I, showing normalized win rate improvement \\( \\Delta \\overline{w} \\) (%) on given routing rates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.907, + 0.678 + ], + "angle": 0, + "content": "gate MoE. The multi-metric quality optimization allows routers to comprehensively assess the image quality gap, significantly enhancing overall performance. Removing this results in a performance decline across all routing rates, particularly at \\(40\\%\\) routing rate, where the drop is approximately \\(3.23\\%\\). The token selection gate prioritizes critical tokens, ensuring routing focuses on key quality factors while reducing interference from irrelevant elements. Its removal notably impacts high routing rates, causing a performance decrease of around \\(2\\%\\) at \\(80\\%\\) routing rate. The dual-gate MoE effectively distinguishes subtle differences between cloud and edge model outputs, proving most effective at medium routing rates, especially near \\(50\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.634, + 0.71 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we have proposed a new text-to-image model routing design RouteT2I between edge and cloud. RouteT2I adopts the routing model with a dual-gate token selection MoE to predict how user prompts affect multiple image quality metrics by identifying and evaluating key tokens with contrastive method. RouteT2I further introduces a routing strategy that exploits the predicted advantage of the edge model over the cloud model to determine which side for image generation. Extensive evaluation has demonstrated that RouteT2I can significantly enhance generation quality at a specified routing rate and meanwhile reduce cost at a given quality target." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19489" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.21, + 0.108 + ], + "angle": 0, + "content": "Acknowledge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.252 + ], + "angle": 0, + "content": "This work was supported in part by National Key R&D Program of China (No. 2022ZD0119100), China NSF grant No. 62202297, No. 62025204, No. 62202296, No. 62272293, No. 62441236, and No. U24A20326, SJTU-Huawei Research Program, and Tencent WeChat Research Program. The opinions, findings, conclusions, and recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agencies or the government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.264, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[1] stabilityai/stable-diffusion-3.5-large · Hugging Face — huggingface.co. https://huggingface.co/stabilityai/stable-diffusion-3.5-large.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.331, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[2] Szymon Antoniak, Michal Krutul, Maciej Pioro, Jakub Krajewski, Jan Ludziejewski, Kamil Ciebiera, Krystian Król, Tomasz Odrzygóźdź, Marek Cygan, and Sebastian Jaszcur. Mixture of tokens: Continuous moe through cross-example aggregation. Advances in Neural Information Processing Systems, 37:103873-103896, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.471, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[4] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 821–830, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[6] Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2818-2829, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[8] Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[10] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[11] Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. *Routerbench: A benchmark for multi-llm routing system.* arXiv preprint arXiv:2403.12031, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.285 + ], + "angle": 0, + "content": "[12] Shijing Hu, Ruijun Deng, Xin Du, Zhihui Lu, Qiang Duan, Yi He, Shih-Chia Huang, and Jie Wu. Laecips: Large vision model assisted adaptive edge-cloud collaboration for iot-based perception system. arXiv preprint arXiv:2404.10498, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[13] Anil Kag and Igor Fedorov. Efficient edge inference by selective query. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[14] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36:69981-70011, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[15] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[16] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[17] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[18] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[20] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[21] Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692, 2023. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19490" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.119 + ], + "angle": 0, + "content": "resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[23] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data. arXiv preprint arXiv:2406.18665, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[30] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[33] Tal Shnitzer, Anthony Ou, Mfrian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets. arXiv preprint arXiv:2309.15789, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for dif" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "fusion models. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[35] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[36] Yiding Wang, Kai Chen, Haisheng Tan, and Kun Guo. Tabi: An efficient multi-level inference system for large language models. In Proceedings of the Eighteenth European Conference on Computer Systems, pages 233-248, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[37] Haoning Wu, Xiele Wu, Chunyi Li, Zicheng Zhang, Chaofeng Chen, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. T2i-scorer: Quantitative evaluation on text-to-image generation via fine-tuned large multi-modal models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3676-3685, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[38] Xun Wu, Shaohan Huang, Guolong Wang, Jing Xiong, and Furu Wei. Multimodal large language models make text-to-image generative models align better. Advances in Neural Information Processing Systems, 37:81287-81323, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[39] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[40] Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. Large language model cascades with mixture of thoughts representations for cost-efficient reasoning. arXiv preprint arXiv:2310.03094, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[41] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. arXiv preprint arXiv:2405.17873, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[42] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[43] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19491" + } + ] +] \ No newline at end of file diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_origin.pdf b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..18423ca07e2dfdc770598bc89ca995e6ee35dfd6 --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/da3df561-f44d-4027-8335-cb684e94422c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14dee4b2a08dffdbdca6a84646916f8ec7fadae22bfc6bacf200d631c9038c14 +size 3162320 diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/full.md b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4a93eb9e58d91c77f3904560d148bbab8025d2fd --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/full.md @@ -0,0 +1,389 @@ +# Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model + +Zewei Xin, Qinya Li*, Chaoyue Niu*, Fan Wu, Guihai Chen +Shanghai Jiao Tong University, China + +{xin_zeweil12138，qinyali，rvince}@sjtu.edu.cn，{fwu，gchen}@cs.sjtu.edu.cn + +# Abstract + +Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation. + +# 1. Introduction + +Nowadays, text-to-image (T2I) models like Imagen [3], Stable Diffusion [28], and DALL-E [27] have achieved significant success in generating diverse, high-quality images given user prompts. However, the impressive generation quality comes with large model and high cost. For example, Stable Diffusion 3.5 [1] has 8 billion parameters. Such a large model necessitates reliance on cloud servers, leading to high serving cost. As shown in Tab. 1, it is particularly costly in commercial scenarios with millions of requests. + +
Text-to-Image Model#ParamPricing ($/M)
Stable Diffusion 1.6 [28]0.86 B9 K
Stable Diffusion XL [24]2.6 B9 K
Stable Diffusion 3 Medium [9]2 B35 K
Stable Diffusion 3 [9]8 B65 K
Stable Diffusion 3.5 [1]8 B65 K
+ +Table 1. For the Stable Diffusion models, newer versions come with more parameters and higher API costs. + +The high cost of cloud-based image generation drives the trend of deploying T2I models on edge devices. Techniques such as quantization [16, 34, 41], structural pruning [4, 17], and reducing denoising steps [20, 22, 30] are used to minimize model sizes and speed up inference, showing that edge deployment is feasible. Compared to cloud models, edge models leverage users' local computing devices to provide instant services anytime and anywhere without cloud serving cost and communication overhead. However, the lightweight edge T2I models often have lower generation quality compared to the cloud large models. + +Although cloud models often offer superior quality, not all user prompts need these large models. For some prompts, light-weight edge models can produce comparable or even better results. To ensure high-image quality generation while limiting requests to cloud models to reduce cost, it is necessary to introduce a routing mechanism such that suitable models are selected for different user prompts based on their complexity. Intuitively, as shown in Fig. 1, a routing framework for T2I models routes only challenging prompts to the cloud, while handling easy ones on the edge. + +Previous routing methods are primarily designed for large language models (LLMs) [8, 11, 21, 23], presenting significant challenges when applied to T2I models. One major challenge lies in comparing the quality of generated images. Unlike text, image quality does not have unified metrics and is subject to many influencing factors, including color, clarity, and the unique distortions typical of generated images [15]. This makes it hard to compare image quality to guide routing decisions. Furthermore, the output space of + +![](images/c991700e86973aca2b80858dd5d63f89dc9bdfc6f13659db4ef332c05c4f8ba6.jpg) +Figure 1. Routing for text-to-image generation requests. + +T2I models is substantially larger than the input text space, and a single prompt can correspond to numerous images, hindering the prediction of image quality before generation. + +To address these challenges, we propose an edge-cloud T2I routing framework RouteT2I, deploying a routing model and a small T2I model on the edge, and a large T2I model on the cloud. Our goal is to optimize the overall quality of generated images at a limited cost, by effectively routing user prompts to the edge or the cloud. To evaluate the quality gap between generated images, we first propose a multi-metric quality measure, including metrics for both real photos and those unique to generated images. Each metric assesses image quality by measuring the similarity between the image and a pair of positive and negative prompts that describe the quality metric. We then define Pareto relative superiority (PRS) to quantify quality differences between images. RouteT2I predicts PRS based on user prompts, inspired by the cross-attention operation between text and image during generation. In particular, inspired by mixture-of-experts (MoE) selective focus on experts or tokens in LLMs [2, 32, 43], we design a novel dual-gate token selection MoE, where a user prompt is treated as a sequence of tokens, with experts aligned to quality metrics. A token selection gate allows experts to proactively choose tokens that markedly impact these metrics, thereby focusing on key tokens. A dual-gate MoE uses both positive and negative gates to evaluate dominant effects when tokens have opposing effects simultaneously. Furthermore, multiple heads are introduced to predict multiple quality metrics. Based on the multi-metric quality prediction, RouteT2I determines quality differences between images generated on the edge over those in the cloud. Only user prompts that show substantial quality improvement when generated in the cloud are routed there, while others remain at the edge. Our key contributions are as follows: + +- We, for the first time, study how to route text-to-image generation requests between cloud and edge models for low cost and high quality. +- Specified to multi-metric quality measures of generated images, we design a routing model architecture with dual gate token selection MoE, combined with a routing strat + +egy for user prompts. + +- We evaluate the proposed design using 18 image generation model pairs on the public COCO 2014 dataset. Key results include: (1) by routing $50\%$ of requests to edge and cloud models, respectively, the overall quality improvement over the edge model reaches $83.97\%$ of that achieved by fully using the cloud model; and (2) to achieve $50\%$ of the quality improvement attained by fully using the cloud model, the number of requests routed to the cloud model is reduced by $70.24\%$ compared to a random routing policy. + +# 2. Related Work + +Model routing selects suitable models for user requests to ensure quality while reducing high cost of invoking powerful models. Depending on whether routing decision occurs before or after weak model execution, model routing can be categorized into predictive and non-predictive routing. + +Non-predictive routing takes the output of a weak model to decide whether to route to the next, powerful model. LAECIPS [12] evaluated the confidence of predictions made by the edge semantic segmentation model to decide if cloud assistance from a large visual model is needed. For classification models, Hybrid [13] used a routing model to partition the data domain between edge and cloud models, routing cases with incorrect edge outputs to the cloud. For cascading LLM, LLM Cascade [40] routed based on the consistency of responses from weaker LLMs, while FrugalGPT [6] considered current quality and past costs to determine routing. Tabi [36] not only routed based on the confidence of responses from weaker LLMs but also enhanced output quality by aggregating historic outputs. + +Different from non-predictive routing, predictive routing selects model before running the weak model on resource-constrained edge device. Previous studies focused on LLM, with the difference in routing strategies. RouteLLM [23] predicted the outcomes of quality comparisons between LLM and routed based on prediction confidence. Hybrid LLM [8] relaxed comparisons, allowing the weak model to succeed if the quality gap is within a threshold, thus saving costs with some quality compromise. Recent research [33] advances the confidence routing strategy by identifying out-of-distribution data through confidence thresholds. ZOOTER [21] introduced a new strategy by predicting the normalized quality of candidate model outputs, making routing decision based on relative quality, with routing models distilled from existing quality scoring models. + +Previous routing work were mainly for LLMs, whereas the predictive routing for T2I in our work remains unexplored. A key distinction lies in the image outputs of T2I, which necessitate predicting quality variations in a larger output image space from a smaller input text space (Sec. B in Appendix). On one hand, unlike many LLM tasks with + +![](images/bb7a08002e54d00522e1bc917989117d998fc99d2bbab8e3455c240629b7235a.jpg) +SD2.1 +SD3 +There is a dog lying on the ground. + +![](images/b212a8d7beae69e53ccac2f15934339672dc4efaca4c612c69f65075983dd916.jpg) +SD2.1 + +![](images/667356aa2dd35d9944466490e4ae2f60245c52c2e87407c7f44aa43256f4996b.jpg) +SD3 + +![](images/41a45d196d09eb111a1fa5f3f18ba31527feb677211034be1a5a05abf83889fe.jpg) +A cat sitting motionless on a television stand behind a mug in the shape of a skeleton head. + +![](images/90303fed76d78ba2c4ff90cbeb63402d60fa842b59a34a10fdd87f43978f3f81.jpg) +A sandwich cut in half on a white plate with a side of salad and a fork and knife wrapped in a napkin. + +![](images/296310e1fda8a1a11f508c534489c61459e0b2e27d3aba5943e772e421f0e188.jpg) + +![](images/2e4355162532ad6fd1e9204ff5e6b557f33ad69cf0d4b421d1b5e0e17168d80c.jpg) +Figure 2. SD2.1 tends to match or outperform SD3 on simple prompts, while more likely falling short on complex ones. Entities to be generated are **bolded**. +(a) +Figure 3. Impact of the noun count in text prompts on ImageReward score [39] gap and win rate between SD3 and SD2.1. + +![](images/156f812081bfb20dd78360833c4f508218850ef1e087c7e65701f52075e19c99.jpg) +(b) + +definitive answers, image quality lacks fixed standards and is highly subjective. Previous single-objective LLM routing struggles to adapt to such ambiguous quality criteria. On the other hand, prior work lacked dedicated routing model designs, making it challenging to predict quality relationships within the vast output image space. This underscores the necessity of effective routing model design. + +# 3. Problem Formulation + +In text-to-image generation scenario, the cloud normally hosts a large T2I model, denoted as $\mathcal{M}_c$ , while a resource-constraint edge device, such as a smartphone, often deploys a light-weight T2I model, denoted as $\mathcal{M}_e$ . For example, Google has released Imagen [3, 29] for cloud deployment and MobileDiffusion [42] for edge deployment. From the perspective of a user, the large model generally offers superior generation quality, but incurs edge-cloud communication cost and cloud serving fee, which together are denoted as $F_c$ . In contrast, the edge model can leverage the user's local device, thereby eliminating these additional costs. + +In practice, user requests involve not only complex ones but also simple ones. One intuitive factor influencing request complexity is the number of entities that need to + +be generated in images, i.e. the number of nouns in text prompts. As shown in Fig. 2, 3, as noun count increases, the large model is more likely to outperform the small model in quality, with quality differences becoming more noticeable. However, while large models generally excel in overall quality, small models may perform better on specific prompts. This likelihood increases as the prompt becomes simpler. In such cases, the purely large model serving not only fails to improve performance but also incurs high cost. Therefore, it is necessary to design a routing mechanism to select the most suitable model based on the prompt's characteristics. Intuitively, the routing mechanism should direct simpler prompts to the cost-effective edge model and more complex prompts to the high-quality cloud model, thus maximizing generation quality while minimizing cost. + +Formally, a T2I routing framework $R: \mathcal{X} \to \{0,1\}$ assigns user prompts $\mathcal{X}$ to the cloud large model $\mathcal{M}_c$ as 1 and to the edge light-weight model $\mathcal{M}_e$ as 0. The generated images after routing can be expressed as $\mathcal{I}_r = R(\mathcal{X})\mathcal{I}_c + (1 - R(\mathcal{X}))\mathcal{I}_e$ , where $\mathcal{I}_c$ and $\mathcal{I}_e$ denote the images generated using $\mathcal{M}_c$ and $\mathcal{M}_e$ , respectively. Given a quality scoring function $\mathcal{Q}(\cdot)$ of the generated images, the cost budget $\tau_{fee}$ of cloud model serving and edge-cloud communication, and the response latency constraint $\tau_{time}$ , the optimization objective is maximizing the quality under the budget and latency constraints, formulated as + +$$ +\begin{array}{l} \max R (\mathcal {X}) \mathcal {Q} \left(\mathcal {I} _ {c}\right) + (1 - R (\mathcal {X})) \mathcal {Q} \left(\mathcal {I} _ {e}\right) \\ \begin{array}{l} \text {s . t .} \mathbb {P} \{R (\mathcal {X}) = 1 \} \cdot F _ {c} \leq \tau_ {f e e} \\ \mathbb {P} \{R (\mathcal {X}) = 1 \} - D _ {c} + \mathbb {P} \{R (\mathcal {X}) = 0 \} - D _ {c} \end{array} \tag {1} \\ \mathbb {P} \{R (\mathcal {X}) = 1 \} \cdot D _ {\mathcal {M} _ {c}} + \mathbb {P} \{R (\mathcal {X}) = 0 \} \cdot D _ {\mathcal {M} _ {e}} \\ + D _ {R} \leq \tau_ {\text {t i m e}}, \\ \end{array} +$$ + +where $D_{\mathcal{M}_c}$ , $D_{\mathcal{M}_e}$ , and $D_R$ denote the latency of cloud large model serving plus edge-cloud communication, the latency of the edge light-weight model serving, and the latency of the routing model execution. Considering the constraints are linear, the optimization objective can be approximated as imposing an upper bound $\rho_r$ on the routing rate to the cloud large model as + +$$ +\begin{array}{l} \max R (\mathcal {X}) \mathcal {Q} \left(\mathcal {I} _ {c}\right) + (1 - R (\mathcal {X})) \mathcal {Q} \left(\mathcal {I} _ {e}\right) \\ \text {s . t .} \mathbb {P} \{R (\mathcal {X}) = 1 \} \leq \rho_ {r}. \tag {2} \\ \end{array} +$$ + +Due to the lack of a universal standard for evaluating the quality of generated images, a multi-dimensional quality scoring function $\mathcal{Q}:\mathcal{I}\to [0,1]^N$ is required for a more comprehensive comparison. Thus, different from existing routing for typical classification and text generation tasks, the T2I routing framework's optimization objective is a new multi-objective problem. In addition, to predict quality changes in the larger image output space from the smaller text input space, we conduct a detailed analysis of how different prompt tokens influence the generated images and their multi-dimensional quality metrics, by examining + +the interaction between text and images during the T2I generation process. In what follow, we define the multi-metric quality measures for routing objectives in Sec. 4 and introduce the T2I routing framework in Sec. 5. + +# 4. Multi-Metric Image Generation Quality in Routing Objectives + +The connection between text and images allows image attributes to be described through corresponding prompts. Thus, a metric can be measured by the relative relationship between an image and the positive/negative text pair describing that metric [18, 35]. The contrastive quality metric for image $I$ is defined as + +$$ +q (I, m) = \sigma \left(C L I P \left(I, m ^ {+}\right) - C L I P \left(I, m ^ {-}\right)\right), \tag {3} +$$ + +where $m = (m^{+}, m^{-})$ denotes positive and negative pair, and $\sigma$ denotes the sigmoid function that transforms output values in the range from 0 to 1. According to this definition, if an image is more related to the positive prompt, and is less related to the negative prompt, the contrastive quality metric is higher. Compared to quality measure with only the positive prompt, the contrastive method evaluates whether positive or negative quality is more dominant in the image, leading to a more robust and reliable evaluation. + +Considering the ambiguity and complexity of image quality, we introduce a multi-dimensional metric to comprehensively evaluate generated image quality, providing noise resistance and stability for quality-prediction-based routing. Just as real photo quality depends on factors like definition and color, our multi-metric combines traditional measures of real photos with unique aspects of generated images, such as realism and object integrity. Thus, we extend the single-metric evaluation in Eq. (3) to an $N$ -dimensional quality assessment of an image $I$ , expressed as + +$$ +\mathcal {Q} (I) = \left[ q (I, m _ {i}) \mid i = 1, 2, \dots , N \right], \tag {4} +$$ + +where $m_{i}$ is an instantiation of $m$ in Eq. (3). We adopt 10 common metrics inspired by prior works [14, 37, 38], with details in Appendix Sec. C. For example, for the definition metric, we use the positive prompt "High definition photo" and the negative prompt "Low definition photo"; and for the object integrity, we take the positive prompt "Object completion" and the negative prompt "Object twisting". + +In the multi-objective optimization setting for routing with multi-metric quality, the initial goal is to find a Pareto optimal generated image that excels in all the metrics. However, this is challenging in practice due to the potential nonexistence of such an image generated from the cloud or edge model. We thus relax the constraint of Pareto optimality by allowing suboptimal performance in some metrics if the image significantly outperforms in others when selecting higher-quality images. We define Pareto relative superiority (PRS) to quantify the quality advantage of the images + +![](images/1642754cac28ecc02eed28cddaaba7c38d98d82204342cc6613be67deb64e388.jpg) +Figure 4. Overview of RouteT2I. RouteT2I assembles a pair of off-the-shelf edge and cloud text-to-image models and evaluates them using multi-metric image quality. RouteT2I utilizes Pareto relative superiority between qualities as supervision to train the routing model. Then, to balance quality and cost, the routing strategy determines the most suitable model for each user prompt, choosing between the edge or cloud model. + +$\mathcal{I}_e$ generated by the edge model over the images $\mathcal{I}_c$ generated by the cloud model. In particular, we first normalize the quality distance for the metric $i$ between an edge-generated image $I_e \in \mathcal{I}_e$ and a cloud-generated image $I_c \in \mathcal{I}_c$ as + +$$ +D _ {i} (I _ {e}, I _ {c}) = \sigma \left(\frac {q (I _ {e} , m _ {i}) - q (I _ {c} , m _ {i})}{\Gamma | \mu_ {i} (\mathcal {I} _ {e}) - \mu_ {i} (\mathcal {I} _ {c}) |}\right), \tag {5} +$$ + +where $\mu_i(\cdot)$ is the average quality of the set of cloud or edge generated images, while the sigmoid function $\sigma$ and the temperature parameter $\Gamma$ are used to modulate the data distribution, effectively distinguishing similar qualities and preventing centralization. We then define the overall quality gap as a weighted sum of distances across all $N$ metrics: + +$$ +P R S \left(I _ {e}, I _ {c}\right) = \sum_ {i = 1} ^ {N} w _ {i} D _ {i} \left(I _ {e}, I _ {c}\right), \tag {6} +$$ + +where $w_{i}$ denotes the weight of metric $i$ , and $\sum_{i} w_{i} = 1$ . We can evaluate the relative quality of $I_{e}$ compared to $I_{c}$ by examining how much Pareto relative superiority deviates from 0.5. This evaluation helps to route a user prompt to the model that generates an image with higher overall quality. + +# 5. Design of RouteT2I + +To achieve the optimization objective in Eq. (2), we propose a text-to-image model routing framework RouteT2I. As shown in Fig. 4, RouteT2I comprises a routing model that predicts the multi-metric generation quality given user prompts and a routing strategy that selects cloud or edge model. In particular, (1) the routing model predicts the quality relationship between images generated on the edge and cloud, namely, Pareto relative superiority. Inspired by the + +![](images/fba23697dd652034e0744e5e4567e6470325d210cae48cc56f74a5576d3b4970.jpg) +(a) Token selection gate. + +![](images/4fcdf9af33b1f26cc5f64f2650bc2e4bc46c0b98bb5747a72b9d7503897021c7.jpg) +(b) Dual-gate MoE. +Figure 5. Architecture of a dual-gate MoE that uses the token selection gate as the gate network. First, the gate selects the most relevant tokens for each expert. Then, experts project the tokens selected by positive and negative gates into their respective spaces and evaluate them using score matrices. Finally, the evaluations of the tokens are gathered and compared to form new tokens. + +role of prompts in generation, where prompt as sequences of tokens interact with images through cross-attention to determine the image content, we treat prompt as a set of tokens in the routing model and select key tokens with influential contextual information for each quality metric. Then, the dominant influence of these tokens is accessed by comparing their positive and negative effects; and (2) the routing strategy aims to maximize quality under cost constraints. The key idea is to use Pareto relative superiority to describe the quality disparity of images generated by the cloud and the edge models. Only prompts that show a notable quality gap when generated in the cloud are routed there, while the rest remain on the cost-effective edge. + +# 5.1. Routing Model Architecture Design with Dual-Gate Token Selection MoE + +To predict Pareto relative superiority between the images generated by the cloud model and the edge model from user prompts, we design a dual-gate token selection mixture-of-experts (MoE), as shown in Fig. 5, and introduce it into Transformer network to replace linear layers. + +Our model treats the user prompt as a sequence of tokens, whose influences on the image quality differ due to their weights in cross-attention during generation. To focus on key tokens for each metric, we design a token selection gate, where experts align with quality metrics and actively choose the most relevant tokens to simulate metric attention to different image attributes. Only tokens that significantly impact the quality metrics corresponding to the experts are selected, while those with minimal influences are disregarded, reducing potential interference. + +Regarding the design of a dual-gate MoE, both positive and negative gates are introduced to distinguish the opposing aspects of token influences on image quality. Similar to images that have both superior and inferior regions when evaluated by metrics, these tokens are assessed by the corresponding experts for their positive and negative influences to each metric. Additionally, we introduce a comparison between two gates to identify which influence of each token predominates overall. We further improve the structure of + +expert to handle both positive and negative influences on a single metric simultaneously. + +To support multi-metric quality prediction, our design incorporates multiple prediction heads within the model architecture. Each head outputs the prediction of a specific quality metric. Such multi-head design not only enhances noise resistance, thereby increasing robustness, but also facilitates a comprehensive evaluation of quality. + +# 5.1.1. Token Selection Gate + +In the token selection gate, the dot product between the token representations $T \in \mathbb{R}^{n \times d}$ and the expert embeddings $E \in \mathbb{R}^{k \times d}$ is computed as an affinity matrix: + +$$ +A = \operatorname {S o f t m a x} \left(T \cdot E ^ {T}\right), \tag {7} +$$ + +where $n$ denotes the number of tokens in user prompt, $k$ denotes the number of experts, $d$ denotes the hidden dimension, and the Softmax function is applied along the expert axis. $A[t,i]$ in the affinity matrix captures the correlation between token $t$ and expert $i$ . Then, top- $K$ tokens with the highest relevance for each expert are selected + +$$ +M = \operatorname {T o p} - \mathrm {K} (A), \tag {8} +$$ + +where $M \in \mathbb{R}^{n \times k}$ is the binary mask matrix such that $M[t, i] = 1$ denotes token $t$ is selected by expert $i$ , and $M[t, i] = 0$ denotes not selected. Experts tend to have similar weights because the metrics they are responsible for are weighted similarly. This ensures that the tokens selected by these experts possess significant global importance. This practice minimizes interference from tokens that are insignificant across all metrics. The influences of key tokens are further evidenced by their frequent selection through the top- $K$ function and their corresponding scores within the affinity matrix. Moreover, by allocating an equal number of $K$ tokens to each expert, workloads remain balanced, and bias in quality dimensions is avoided, ultimately resulting in more comprehensive and reliable quality prediction. Overall, the token selection gate is summarized as + +$$ +A, M = G (T, E), \tag {9} +$$ + +where $G$ denotes the gate network. + +# 5.1.2. Dual-Gate MoE + +Both a positive gate $G^{+}$ and a negative gate $G^{-}$ are introduced for different experts to select tokens that have positive and negative impacts on their corresponding quality metrics, while capturing the weights of their influences. The token selection results are denoted as + +$$ +A ^ {o}, M ^ {o} = G ^ {o} \left(T, E ^ {o}\right), \text {w h e r e} o \in \{+, - \}. \tag {10} +$$ + +The affinity matrix $A^o$ and the binary mask matrix $M^o$ are captured by the gate $G^o$ for the positive or negative aspect using the corresponding expert embedding $E^o$ . The tokens selected by two gates are then evaluated by the corresponding positive or negative experts. + +To eliminate the redundancy of having separate experts for evaluating both positive and negative aspects of the same quality metric, the structure of experts is revised to allow an expert to assess both aspects simultaneously. This revision acknowledges that both positive and negative evaluations are inherently part of the same quality metric. Specifically, the original linear layer $W_{i} \in \mathbb{R}^{d \times h}$ of expert on metric $i$ is split into two projection matrices $P_{i}^{+}, P_{i}^{-} \in \mathbb{R}^{d \times l}$ and one score matrix $S_{i} \in \mathbb{R}^{l \times h}$ . Two projection matrices are specific to positive and negative aspects, projecting tokens into the corresponding representations in an $l$ -dimensional low-dimensional space. Since $l \ll h, d$ , the parameter size and the computation cost are reduced from $O(hd)$ to $O(l(h + d))$ , allowing an expert to efficiently focus on both positive and negative aspects of a metric with a low cost. The shared score matrix then evaluates the impact of these representations on metric $i$ . Thus, the impact of $t$ -th token $T[t] \in \mathbb{R}^d$ on quality metric $i$ for positive or negative aspect $o$ is described as + +$$ +T _ {i} ^ {o} [ t ] = T [ t ] \cdot P _ {i} ^ {o} \cdot S _ {i}, \text {w h e r e} o \in \{+, - \}. \tag {11} +$$ + +By separating the projection matrices and score matrix, we can evaluate both positive and negative impacts on a metric within the same expert. Then, the overall positive or negative impact of the $t$ -th token on image quality can be obtained by a weighted sum of its impacts for corresponding aspects across all metrics + +$$ +T ^ {o} [ t ] = \sum_ {i = 0} ^ {k} \lambda_ {t, i} ^ {o} \cdot T _ {i} ^ {o} [ t ], \text {w h e r e} o \in \{+, - \}. \tag {12} +$$ + +Here, $\lambda_{t,i}^{o} = \frac{\mathbb{1}\{M^{o}[t,i] = 1\}A^{o}[t,i]}{\sum_{j = 0}^{k}\mathbb{1}\{M^{o}[t,j] = 1\}A^{o}[t,j]}$ denotes the normalized affinity of the $i$ -th expert for the $t$ -th token. By contrasting the impact of a token from both positive and negative aspects, we can assess the token's predominant influence on the image quality as + +$$ +\hat {T} [ t ] = \sigma \left(T ^ {+} [ t ] - T ^ {-} [ t ]\right). \tag {13} +$$ + +The introduction of contrasting helps diminish ambiguity in predictions, especially when tokens exhibit influences on multiple metrics from both positive and negative aspects. + +# 5.2. Routing Strategy Design + +Given the Pareto relative superiority predicted by routing model, we propose a routing strategy to effectively route user prompts to suitable models to achieve the trade-off between quality and costs. With a preset routing rate, we can determine the proportion of user prompts sent to the cloud, focusing on those where the cloud model significantly outperforms the edge model in quality. Since Pareto relative superiority represents this quality gap, we can efficiently filter prompts by setting a threshold $\alpha$ on it. In particular, the prompts with a Pareto relative superiority above the threshold are better handled by the edge model for cost efficiency, while those below the threshold achieve superior quality when processed in the cloud. Thus, the optimization objective in Eq. (2) can be expressed as: + +$$ +\max _ {\alpha \leq 1 / 2} \mathbb {P} \left\{P R S \left(I _ {e}, I _ {c}\right) < \alpha \mid I _ {e}, I _ {c} \in \mathcal {I} _ {e}, \mathcal {I} _ {c} \right\} \leq \rho_ {r}. \tag {14} +$$ + +Since Pareto relative superiority being greater than or less than $1/2$ indicates a relative advantage or disadvantage, respectively, we set a upper bound $1/2$ on $\alpha$ to prevent user prompts with images generated at the edge of better quality from being routed to the cloud. + +# 6. Experiment + +# 6.1. Experimental Setup + +Dataset. We take COCO2014 [19], a comprehensive dataset for object detection, segmentation, and captioning tasks. We select its captions as user prompts $\mathcal{X}$ . Given these prompts, we generate images $\mathcal{I}_e, \mathcal{I}_c$ using different open-source text-to-image models. Similarly, we use the LAION dataset [31], with results shown in Appendix Sec. F. + +Models. We take T2I models with varying performances and sizes for cloud and edge usage, include Stable Diffusion 1.5 (SD1.5) [28], Stable Diffusion 2.1 (SD2.1) [28], Stable Diffusion XL (SDXL) [24], Stable Diffusion XL-Refiner (XL-Refiner) [24], Stable Diffusion 3 (SD3) [9], PixArt $\alpha$ [5], and Infinity [10]. These models use various text encoders, such as CLIP [25], OpenCLIP [7], and T5 [26], and are based on architectures like diffusion, diffusion Transformer and autoregressive. Details are in Appendix Sec. A. + +Baselines and Settings. We first introduce random routing as a baseline, where prompts are randomly assigned to candidate models, given a specified routing rate to the cloud. Additionally, although there does not exist any previous work on text-to-image generation routing, for a comprehensive comparison, we also reproduce several representative routing methods for LLM by adapting them to our scenarios, including RouteLLM [23] with BERT classifier or + +
RouterImage Quality MetricsΔP(%)
DefinitionDetailClaritySharpnessHarmonyRealismColorConsistencyLayoutIntegrity
Edge Model0.62510.66850.60760.65370.59490.55750.46800.50880.48600.4690-
Cloud Model0.63370.68470.63460.67030.59300.58680.51340.51990.53450.4972-
Random0.62940.67660.62110.66200.59390.57210.49070.51440.51020.483140.00
RouteLLM-BERT [23]0.63470.67920.63050.66510.59600.57880.49820.51600.51670.486671.51
RouteLLM-MF [23]0.63640.68140.62990.66600.59520.57760.49700.51640.51490.485069.90
Hybrid LLM [8]0.63270.67840.63060.66770.59640.57870.50080.51610.51910.486473.49
ZOOTER [21]0.63500.67960.63150.66720.59660.57880.50040.51660.51790.485477.95
RouteT2I (Ours)0.63500.67860.63180.66790.59750.58040.50100.51670.51890.486583.97
+ +matrix factorization, Hybrid LLM [8], and ZOOTER [21]. Specifically, we keep their text input processing and quality threshold design, but replace their optimization goal with the average image quality across metrics as they only handle single-objective optimization. + +Unless specified in their paper, the routing models are based on Transformers and take the same hyperparameters as in our method. All routers are implemented in PyTorch 2.3.1 and trained using Adam optimizer with a learning rate of 2e-5, a weight decay of 0, a batch size of 16, on a NVIDIA 4090D for about 10 epochs. + +Metrics. For cost efficiency, we introduce routing rate, namely, the proportion of user prompts to the cloud model: + +$$ +p = \mathbb {P} \{R (\mathcal {X}) = 1 \}. \tag {15} +$$ + +Smaller routing rate indicates better cost efficiency. + +For image generation quality, the absolute performance of routing is constrained by the gap between edge and cloud T2I models. Thus, we introduce a set of relative performance metrics for evaluation. We first fix a routing rate $p$ and introduce the win rate of the selected models over the cloud model after routing: + +$$ +w \left(\mathcal {I} _ {r, p}\right) = \mathbb {P} \left\{\mathcal {Q} \left(I _ {r}\right) \geq \mathcal {Q} \left(I _ {c}\right) \mid I _ {r}, I _ {c} \in \mathcal {I} _ {r, p}, \mathcal {I} _ {c} \right\}, \tag {16} +$$ + +where $\mathcal{I}_{r,p}$ denotes the generated images after routing at a routing rate $p$ . To intuitively show the improvement in win rate, we introduce the normalized win rate improvement, which is the increase in win rate achieved by a routing method compared to the baseline, divided by the increase achieved by the optimal oracle routing. A higher value indicates that the routing method is closer to the oracle routing: + +$$ +\Delta \bar {w} (p) = \frac {w \left(\mathcal {I} _ {r , p}\right) - w \left(\mathcal {I} _ {b , p}\right)}{w \left(\mathcal {I} _ {o , p}\right) - w \left(\mathcal {I} _ {b , p}\right)}, \tag {17} +$$ + +where $\mathcal{I}_{b,p}$ and $\mathcal{I}_{o,p}$ denote the generated images using the random baseline and oracle, respectively. To quantify the improvement of image quality metrics, we introduce relative performance improvement, which measures the proportion of the performance gain achieved through routing + +Table 2. The multi-dimensional quality of images generated by text-to-image models, with the router selecting between edge and cloud models for each prompt at a $50\%$ routing rate. The higher the metrics, the better. + +
Router\p40%50%60%70%80%
RouteLLM-BERT [23]24.2922.4519.0717.6220.59
RouteLLM-MF [23]25.6523.2919.9216.6720.59
Hybrid LLM [8]23.7719.7514.9213.6216.09
ZOOTER [21]26.7721.9717.9716.9721.04
RouteT2I (Ours)30.6025.8120.3218.0221.94
+ +Table 3. Normalized win rate improvements $\Delta \overline{w}$ (\%) at different routing rates, i.e. the ratio of improvement to oracle improvement. + +compared to the edge model, relative to the total possible improvement brought by fully using the cloud model: + +$$ +\Delta P = \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {\mu_ {i} \left(\mathcal {I} _ {r}\right) - \mu_ {i} \left(\mathcal {I} _ {e}\right)}{\left| \mu_ {i} \left(\mathcal {I} _ {c}\right) - \mu_ {i} \left(\mathcal {I} _ {e}\right) \right|}. \tag {18} +$$ + +This metric effectively quantifies routing effectiveness while accounting for the original quality gap. + +For cost-quality balance, we introduce the cost saving to measure the reduction of the routing rate to the cloud model, compared to the random baseline at a given relative performance improvement $\Delta P$ : + +$$ +\gamma (\Delta P) = \frac {p _ {b} (\Delta P) - p _ {r} (\Delta P)}{p _ {b} (\Delta P)}, \tag {19} +$$ + +where $p_b(\Delta P)$ and $p_r(\Delta P)$ represent the routing rates of baseline and router at a given $\Delta P$ , respectively. + +# 6.2. Main Results + +We present the routing performance of our RouteT2I and baselines using SD3 as the cloud model and SD2.1 as the edge model. Human evaluation, cost analysis, visual results, and more T2I pair results are provided in Appendix. + +Win Rates. Tab. 3 shows our RouteT2I demonstrates a significant improvements on normalized win rate improvements across all routing rates. Specifically, at a $40\%$ routing rate, our improvement reaches $30.60\%$ , exceeding the baselines by at least $3.83\%$ and demonstrating strong consistency with the oracle. + +Image Quality. Tab. 2 shows RouteT2I outperforms all baselines in 6 of 10 metrics. Our multi-metric optimization balances all metrics simultaneously, while others focus on just a few, often at the cost of others. We + +
ΔP Router40%50%60%70%80%
RouteLLM-BERT [23]56.1551.3946.9242.7040.21
RouteLLM-MF [23]48.8649.9048.2044.8941.50
Hybrid LLM [8]62.0658.8553.6349.9233.38
ZOOTER [21]69.2865.7660.8157.3549.64
RouteT2I (Ours)71.8170.2466.6160.0153.53
+ +![](images/69c67ba2cf1dbda5d25956402a4ade59774f323ac9a92217ee3331020bd77084.jpg) +(a) + +![](images/228111b9e8cc0d6ab2e87be145202cdd29122e322b9cc3680b153e94bbf2492a.jpg) +(b) +Figure 6. Visualization of (a) Normalized win rate improvement $\Delta \overline{w}$ and (b) relative performance improvement $\Delta P$ at varying routing rates $p$ to the cloud. + +also achieve the highest relative performance improvement, reaching $83.97\%$ of cloud model's gain over edge model, surpassing others by at least $6\%$ . Experiments in Appendix Sec. H show our method may lead to performance gains beyond those of the cloud model itself and is effective for T2I models with various text encoders and architectures. + +Cost Efficiency. Tab. 4 presents the cost savings at given relative performance improvement targets. At a $40\%$ relative performance improvement, RouteT2I reduce cloud service needs by up to $71.81\%$ of the calls required by random routing and by at least $5.80\%$ compared to other baselines. In other cases, RouteT2I still reduces over $50\%$ costs, demonstrating high cost efficiency. + +Visualization. As shown in Fig. 6, RouteT2I performs well across routing rates, demonstrating its ability to predict quality gaps between T2I models and route effectively. + +Visual examples are shown in Appendix Sec. G. + +# 6.3. Routing Performance on More T2I Model Pairs + +To test our method across T2I model pairs, Tab. 5 shows RouteT2I's performance with various cloud and edge models. Results reveal significant improvements, especially for pairs with large quality gaps, such as SD3 on cloud and others on edge, where RouteT2I achieves over $25\%$ of the oracle routing's improvement at a $40\%$ routing rate. Even with similar model pairs like XL-Refiner and SDXL, it reaches $10\%$ of the oracle's improvement, demonstrating its ability to discern subtle differences between T2I models. + +# 6.4. Ablation Experiment + +Tab. 6 presents ablation results for RouteT2I's multi-metric quality optimization, token selection gate, and dual + +Table 4. Cost saving $\gamma$ (\%) under a given relative performance improvements $(\Delta P)$ target. $\gamma$ is the ratio of savings to the routing rate required by the random baseline. + +
Cloud-Edgep40%50%60%70%80%
SD3XL-Refiner27.4222.3918.0516.8520.89
SD3SDXL29.7826.8122.3516.9821.30
SD3SD1.526.2420.8816.1914.0419.22
SD2.1XL-Refiner21.6218.3120.4625.3028.66
SD2.1SDXL19.4616.8717.2720.4223.69
SD2.1SD1.510.779.2810.1212.2215.76
XL-RefinerSDXL13.5710.159.7411.8112.59
XL-RefinerSD1.519.6516.9818.8722.7925.57
SDXLSD1.518.3817.0121.5724.7532.49
SD3PixArt-α32.2627.5924.7430.7338.97
PixArt-αSD2.125.8221.2622.5528.1933.76
PixArt-αSDXL16.7113.5712.3716.3522.14
PixArt-αXL-Refiner22.5119.1418.1823.7228.44
PixArt-αSD1.522.3820.5219.9325.4032.43
SD3Infinity22.1919.9717.7719.8125.39
SD2.1Infinity21.9819.8622.6028.4536.18
PixArt-αInfinity27.6923.4824.1230.8938.91
+ +Table 5. Normalized win rate improvement $\Delta \overline{w} (\%)$ over oracle routing of our RouteT2I with different cloud-edge model pairs. + +
Router\p40%50%60%70%80%
w/o Multi-Metric27.3722.8118.8716.4719.92
w/o Token Selection27.8223.0518.2716.7719.24
w/o Dual-Gate27.2222.0919.2717.3721.62
RouteT2I30.6025.8120.3218.0221.94
+ +Table 6. Ablation experiment of RouteT2I, showing normalized win rate improvement $\Delta \overline{w}$ (%) on given routing rates. + +gate MoE. The multi-metric quality optimization allows routers to comprehensively assess the image quality gap, significantly enhancing overall performance. Removing this results in a performance decline across all routing rates, particularly at $40\%$ routing rate, where the drop is approximately $3.23\%$ . The token selection gate prioritizes critical tokens, ensuring routing focuses on key quality factors while reducing interference from irrelevant elements. Its removal notably impacts high routing rates, causing a performance decrease of around $2\%$ at $80\%$ routing rate. The dual-gate MoE effectively distinguishes subtle differences between cloud and edge model outputs, proving most effective at medium routing rates, especially near $50\%$ . + +# 7. Conclusion + +In this work, we have proposed a new text-to-image model routing design RouteT2I between edge and cloud. RouteT2I adopts the routing model with a dual-gate token selection MoE to predict how user prompts affect multiple image quality metrics by identifying and evaluating key tokens with contrastive method. RouteT2I further introduces a routing strategy that exploits the predicted advantage of the edge model over the cloud model to determine which side for image generation. Extensive evaluation has demonstrated that RouteT2I can significantly enhance generation quality at a specified routing rate and meanwhile reduce cost at a given quality target. + +# Acknowledge + +This work was supported in part by National Key R&D Program of China (No. 2022ZD0119100), China NSF grant No. 62202297, No. 62025204, No. 62202296, No. 62272293, No. 62441236, and No. U24A20326, SJTU-Huawei Research Program, and Tencent WeChat Research Program. The opinions, findings, conclusions, and recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agencies or the government. + +# References + +[1] stabilityai/stable-diffusion-3.5-large · Hugging Face — huggingface.co. https://huggingface.co/stabilityai/stable-diffusion-3.5-large.1 +[2] Szymon Antoniak, Michal Krutul, Maciej Pioro, Jakub Krajewski, Jan Ludziejewski, Kamil Ciebiera, Krystian Król, Tomasz Odrzygóźdź, Marek Cygan, and Sebastian Jaszcur. Mixture of tokens: Continuous moe through cross-example aggregation. Advances in Neural Information Processing Systems, 37:103873-103896, 2025. 2 +[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3 +[4] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 821–830, 2024. 1 +[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6 +[6] Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023. 2 +[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2818-2829, 2023. 6 +[8] Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618, 2024. 1, 2, 7, 8 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 6 + +[10] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 6 +[11] Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. *Routerbench: A benchmark for multi-llm routing system.* arXiv preprint arXiv:2403.12031, 2024. 1 +[12] Shijing Hu, Ruijun Deng, Xin Du, Zhihui Lu, Qiang Duan, Yi He, Shih-Chia Huang, and Jie Wu. Laecips: Large vision model assisted adaptive edge-cloud collaboration for iot-based perception system. arXiv preprint arXiv:2404.10498, 2024. 2 +[13] Anil Kag and Igor Fedorov. Efficient edge inference by selective query. In International Conference on Learning Representations, 2023. 2 +[14] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36:69981-70011, 2023. 4 +[15] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 1 +[16] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1 +[17] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1 +[18] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 4 +[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[20] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 1 +[21] Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692, 2023. 1, 2, 7, 8 +[22] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high + +resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1 +[23] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data. arXiv preprint arXiv:2406.18665, 2024. 1, 2, 6, 7, 8 +[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 6 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6 +[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6 +[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3 +[30] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 1 +[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6 +[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2 +[33] Tal Shnitzer, Anthony Ou, Mfrian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets. arXiv preprint arXiv:2309.15789, 2023. 2 +[34] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for dif + +fusion models. Advances in Neural Information Processing Systems, 36, 2024. 1 +[35] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 4 +[36] Yiding Wang, Kai Chen, Haisheng Tan, and Kun Guo. Tabi: An efficient multi-level inference system for large language models. In Proceedings of the Eighteenth European Conference on Computer Systems, pages 233-248, 2023. 2 +[37] Haoning Wu, Xiele Wu, Chunyi Li, Zicheng Zhang, Chaofeng Chen, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. T2i-scorer: Quantitative evaluation on text-to-image generation via fine-tuned large multi-modal models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3676-3685, 2024. 4 +[38] Xun Wu, Shaohan Huang, Guolong Wang, Jing Xiong, and Furu Wei. Multimodal large language models make text-to-image generative models align better. Advances in Neural Information Processing Systems, 37:81287-81323, 2024. 4 +[39] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3 +[40] Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. Large language model cascades with mixture of thoughts representations for cost-efficient reasoning. arXiv preprint arXiv:2310.03094, 2023. 2 +[41] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. arXiv preprint arXiv:2405.17873, 2024. 1 +[42] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3 +[43] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2 \ No newline at end of file diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/images.zip b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1de2a4ebf9533d327fc50bbc249d55ecef41da86 --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:458d6162caa237ee1e37475d4b2c4fe015a46ccf8452799ad2f041d9551d4ca5 +size 578510 diff --git a/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/layout.json b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79f749e83764ea8e2d09a17cbf84ae9b287aec49 --- /dev/null +++ b/2025/Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model/layout.json @@ -0,0 +1,10228 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "text", + "content": "Adaptive Routing of Text-to-Image Generation Requests Between Large Cloud Model and Light-Weight Edge Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 161, + 447, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 161, + 447, + 190 + ], + "spans": [ + { + "bbox": [ + 164, + 161, + 447, + 190 + ], + "type": "text", + "content": "Zewei Xin, Qinya Li*, Chaoyue Niu*, Fan Wu, Guihai Chen \nShanghai Jiao Tong University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 191, + 503, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 191, + 503, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 503, + 204 + ], + "type": "text", + "content": "{xin_zeweil12138，qinyali，rvince}@sjtu.edu.cn，{fwu，gchen}@cs.sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 544 + ], + "type": "text", + "content": "Large text-to-image models demonstrate impressive generation capabilities; however, their substantial size necessitates expensive cloud servers for deployment. Conversely, light-weight models can be deployed on edge devices at lower cost but often with inferior generation quality for complex user prompts. To strike a balance between performance and cost, we propose a routing framework, called RouteT2I, which dynamically selects either the large cloud model or the light-weight edge model for each user prompt. Since generated image quality is challenging to measure and compare directly, RouteT2I establishes multi-dimensional quality metrics, particularly, by evaluating the similarity between the generated images and both positive and negative texts that describe each specific quality metric. RouteT2I then predicts the expected quality of the generated images by identifying key tokens in the prompt and comparing their impact on the quality. RouteT2I further introduces the Pareto relative superiority to compare the multi-metric quality of the generated images. Based on this comparison and predefined cost constraints, RouteT2I allocates prompts to either the edge or the cloud. Evaluation reveals that RouteT2I significantly reduces the number of requesting large cloud model while maintaining high-quality image generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "content": "Nowadays, text-to-image (T2I) models like Imagen [3], Stable Diffusion [28], and DALL-E [27] have achieved significant success in generating diverse, high-quality images given user prompts. However, the impressive generation quality comes with large model and high cost. For example, Stable Diffusion 3.5 [1] has 8 billion parameters. Such a large model necessitates reliance on cloud servers, leading to high serving cost. As shown in Tab. 1, it is particularly costly in commercial scenarios with millions of requests." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 326, + 230, + 543, + 297 + ], + "blocks": [ + { + "bbox": [ + 326, + 230, + 543, + 297 + ], + "lines": [ + { + "bbox": [ + 326, + 230, + 543, + 297 + ], + "spans": [ + { + "bbox": [ + 326, + 230, + 543, + 297 + ], + "type": "table", + "html": "
Text-to-Image Model#ParamPricing ($/M)
Stable Diffusion 1.6 [28]0.86 B9 K
Stable Diffusion XL [24]2.6 B9 K
Stable Diffusion 3 Medium [9]2 B35 K
Stable Diffusion 3 [9]8 B65 K
Stable Diffusion 3.5 [1]8 B65 K
", + "image_path": "ffacc91aeca21aae7aeeb76f5d2f76911aa0888db9f5ff7d12638571a42c6c29.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 328 + ], + "type": "text", + "content": "Table 1. For the Stable Diffusion models, newer versions come with more parameters and higher API costs." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 352, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 484 + ], + "type": "text", + "content": "The high cost of cloud-based image generation drives the trend of deploying T2I models on edge devices. Techniques such as quantization [16, 34, 41], structural pruning [4, 17], and reducing denoising steps [20, 22, 30] are used to minimize model sizes and speed up inference, showing that edge deployment is feasible. Compared to cloud models, edge models leverage users' local computing devices to provide instant services anytime and anywhere without cloud serving cost and communication overhead. However, the lightweight edge T2I models often have lower generation quality compared to the cloud large models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "type": "text", + "content": "Although cloud models often offer superior quality, not all user prompts need these large models. For some prompts, light-weight edge models can produce comparable or even better results. To ensure high-image quality generation while limiting requests to cloud models to reduce cost, it is necessary to introduce a routing mechanism such that suitable models are selected for different user prompts based on their complexity. Intuitively, as shown in Fig. 1, a routing framework for T2I models routes only challenging prompts to the cloud, while handling easy ones on the edge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 557, + 715 + ], + "type": "text", + "content": "Previous routing methods are primarily designed for large language models (LLMs) [8, 11, 21, 23], presenting significant challenges when applied to T2I models. One major challenge lies in comparing the quality of generated images. Unlike text, image quality does not have unified metrics and is subject to many influencing factors, including color, clarity, and the unique distortions typical of generated images [15]. This makes it hard to compare image quality to guide routing decisions. Furthermore, the output space of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "text", + "content": "*Q. Li and C. Niu are the corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19482" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 171 + ], + "type": "image", + "image_path": "c991700e86973aca2b80858dd5d63f89dc9bdfc6f13659db4ef332c05c4f8ba6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 180, + 277, + 192 + ], + "lines": [ + { + "bbox": [ + 72, + 180, + 277, + 192 + ], + "spans": [ + { + "bbox": [ + 72, + 180, + 277, + 192 + ], + "type": "text", + "content": "Figure 1. Routing for text-to-image generation requests." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 216, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 216, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 295, + 252 + ], + "type": "text", + "content": "T2I models is substantially larger than the input text space, and a single prompt can correspond to numerous images, hindering the prediction of image quality before generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 637 + ], + "type": "text", + "content": "To address these challenges, we propose an edge-cloud T2I routing framework RouteT2I, deploying a routing model and a small T2I model on the edge, and a large T2I model on the cloud. Our goal is to optimize the overall quality of generated images at a limited cost, by effectively routing user prompts to the edge or the cloud. To evaluate the quality gap between generated images, we first propose a multi-metric quality measure, including metrics for both real photos and those unique to generated images. Each metric assesses image quality by measuring the similarity between the image and a pair of positive and negative prompts that describe the quality metric. We then define Pareto relative superiority (PRS) to quantify quality differences between images. RouteT2I predicts PRS based on user prompts, inspired by the cross-attention operation between text and image during generation. In particular, inspired by mixture-of-experts (MoE) selective focus on experts or tokens in LLMs [2, 32, 43], we design a novel dual-gate token selection MoE, where a user prompt is treated as a sequence of tokens, with experts aligned to quality metrics. A token selection gate allows experts to proactively choose tokens that markedly impact these metrics, thereby focusing on key tokens. A dual-gate MoE uses both positive and negative gates to evaluate dominant effects when tokens have opposing effects simultaneously. Furthermore, multiple heads are introduced to predict multiple quality metrics. Based on the multi-metric quality prediction, RouteT2I determines quality differences between images generated on the edge over those in the cloud. Only user prompts that show substantial quality improvement when generated in the cloud are routed there, while others remain at the edge. Our key contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "content": "- We, for the first time, study how to route text-to-image generation requests between cloud and edge models for low cost and high quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- Specified to multi-metric quality measures of generated images, we design a routing model architecture with dual gate token selection MoE, combined with a routing strat" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 73, + 411, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 73, + 411, + 84 + ], + "spans": [ + { + "bbox": [ + 322, + 73, + 411, + 84 + ], + "type": "text", + "content": "egy for user prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "content": "- We evaluate the proposed design using 18 image generation model pairs on the public COCO 2014 dataset. Key results include: (1) by routing " + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "content": " of requests to edge and cloud models, respectively, the overall quality improvement over the edge model reaches " + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "inline_equation", + "content": "83.97\\%" + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "content": " of that achieved by fully using the cloud model; and (2) to achieve " + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "content": " of the quality improvement attained by fully using the cloud model, the number of requests routed to the cloud model is reduced by " + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "inline_equation", + "content": "70.24\\%" + }, + { + "bbox": [ + 313, + 84, + 555, + 205 + ], + "type": "text", + "content": " compared to a random routing policy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 215, + 401, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 401, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 401, + 227 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 295 + ], + "type": "text", + "content": "Model routing selects suitable models for user requests to ensure quality while reducing high cost of invoking powerful models. Depending on whether routing decision occurs before or after weak model execution, model routing can be categorized into predictive and non-predictive routing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 295, + 556, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 462 + ], + "type": "text", + "content": "Non-predictive routing takes the output of a weak model to decide whether to route to the next, powerful model. LAECIPS [12] evaluated the confidence of predictions made by the edge semantic segmentation model to decide if cloud assistance from a large visual model is needed. For classification models, Hybrid [13] used a routing model to partition the data domain between edge and cloud models, routing cases with incorrect edge outputs to the cloud. For cascading LLM, LLM Cascade [40] routed based on the consistency of responses from weaker LLMs, while FrugalGPT [6] considered current quality and past costs to determine routing. Tabi [36] not only routed based on the confidence of responses from weaker LLMs but also enhanced output quality by aggregating historic outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 556, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 556, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 556, + 642 + ], + "type": "text", + "content": "Different from non-predictive routing, predictive routing selects model before running the weak model on resource-constrained edge device. Previous studies focused on LLM, with the difference in routing strategies. RouteLLM [23] predicted the outcomes of quality comparisons between LLM and routed based on prediction confidence. Hybrid LLM [8] relaxed comparisons, allowing the weak model to succeed if the quality gap is within a threshold, thus saving costs with some quality compromise. Recent research [33] advances the confidence routing strategy by identifying out-of-distribution data through confidence thresholds. ZOOTER [21] introduced a new strategy by predicting the normalized quality of candidate model outputs, making routing decision based on relative quality, with routing models distilled from existing quality scoring models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "Previous routing work were mainly for LLMs, whereas the predictive routing for T2I in our work remains unexplored. A key distinction lies in the image outputs of T2I, which necessitate predicting quality variations in a larger output image space from a smaller input text space (Sec. B in Appendix). On one hand, unlike many LLM tasks with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19483" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 79, + 175, + 137 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 99, + 79 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 99, + 79 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 99, + 79 + ], + "type": "text", + "content": "SD2.1" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 72, + 149, + 79 + ], + "lines": [ + { + "bbox": [ + 137, + 72, + 149, + 79 + ], + "spans": [ + { + "bbox": [ + 137, + 72, + 149, + 79 + ], + "type": "text", + "content": "SD3" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 79, + 175, + 137 + ], + "lines": [ + { + "bbox": [ + 59, + 79, + 175, + 137 + ], + "spans": [ + { + "bbox": [ + 59, + 79, + 175, + 137 + ], + "type": "image", + "image_path": "bb7a08002e54d00522e1bc917989117d998fc99d2bbab8e3455c240629b7235a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 137, + 162, + 144 + ], + "lines": [ + { + "bbox": [ + 71, + 137, + 162, + 144 + ], + "spans": [ + { + "bbox": [ + 71, + 137, + 162, + 144 + ], + "type": "text", + "content": "There is a dog lying on the ground." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 80, + 235, + 136 + ], + "blocks": [ + { + "bbox": [ + 195, + 72, + 214, + 79 + ], + "lines": [ + { + "bbox": [ + 195, + 72, + 214, + 79 + ], + "spans": [ + { + "bbox": [ + 195, + 72, + 214, + 79 + ], + "type": "text", + "content": "SD2.1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 80, + 235, + 136 + ], + "lines": [ + { + "bbox": [ + 177, + 80, + 235, + 136 + ], + "spans": [ + { + "bbox": [ + 177, + 80, + 235, + 136 + ], + "type": "image", + "image_path": "b212a8d7beae69e53ccac2f15934339672dc4efaca4c612c69f65075983dd916.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 235, + 80, + 291, + 136 + ], + "blocks": [ + { + "bbox": [ + 253, + 72, + 266, + 79 + ], + "lines": [ + { + "bbox": [ + 253, + 72, + 266, + 79 + ], + "spans": [ + { + "bbox": [ + 253, + 72, + 266, + 79 + ], + "type": "text", + "content": "SD3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 80, + 291, + 136 + ], + "lines": [ + { + "bbox": [ + 235, + 80, + 291, + 136 + ], + "spans": [ + { + "bbox": [ + 235, + 80, + 291, + 136 + ], + "type": "image", + "image_path": "667356aa2dd35d9944466490e4ae2f60245c52c2e87407c7f44aa43256f4996b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 144, + 174, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 144, + 174, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 144, + 174, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 144, + 174, + 201 + ], + "type": "image", + "image_path": "41a45d196d09eb111a1fa5f3f18ba31527feb677211034be1a5a05abf83889fe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 201, + 168, + 224 + ], + "lines": [ + { + "bbox": [ + 66, + 201, + 168, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 201, + 168, + 224 + ], + "type": "text", + "content": "A cat sitting motionless on a television stand behind a mug in the shape of a skeleton head." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 144, + 235, + 201 + ], + "blocks": [ + { + "bbox": [ + 176, + 144, + 235, + 201 + ], + "lines": [ + { + "bbox": [ + 176, + 144, + 235, + 201 + ], + "spans": [ + { + "bbox": [ + 176, + 144, + 235, + 201 + ], + "type": "image", + "image_path": "90303fed76d78ba2c4ff90cbeb63402d60fa842b59a34a10fdd87f43978f3f81.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 201, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 201, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 201, + 288, + 224 + ], + "type": "text", + "content": "A sandwich cut in half on a white plate with a side of salad and a fork and knife wrapped in a napkin." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 144, + 291, + 201 + ], + "blocks": [ + { + "bbox": [ + 235, + 144, + 291, + 201 + ], + "lines": [ + { + "bbox": [ + 235, + 144, + 291, + 201 + ], + "spans": [ + { + "bbox": [ + 235, + 144, + 291, + 201 + ], + "type": "image", + "image_path": "296310e1fda8a1a11f508c534489c61459e0b2e27d3aba5943e772e421f0e188.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 277, + 174, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 233, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 266 + ], + "type": "text", + "content": "Figure 2. SD2.1 tends to match or outperform SD3 on simple prompts, while more likely falling short on complex ones. Entities to be generated are **bolded**." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 277, + 174, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 277, + 174, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 277, + 174, + 365 + ], + "type": "image", + "image_path": "2e4355162532ad6fd1e9204ff5e6b557f33ad69cf0d4b421d1b5e0e17168d80c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 369, + 122, + 377 + ], + "lines": [ + { + "bbox": [ + 111, + 369, + 122, + 377 + ], + "spans": [ + { + "bbox": [ + 111, + 369, + 122, + 377 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 388, + 295, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 410 + ], + "type": "text", + "content": "Figure 3. Impact of the noun count in text prompts on ImageReward score [39] gap and win rate between SD3 and SD2.1." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 279, + 291, + 365 + ], + "blocks": [ + { + "bbox": [ + 176, + 279, + 291, + 365 + ], + "lines": [ + { + "bbox": [ + 176, + 279, + 291, + 365 + ], + "spans": [ + { + "bbox": [ + 176, + 279, + 291, + 365 + ], + "type": "image", + "image_path": "156f812081bfb20dd78360833c4f508218850ef1e087c7e65701f52075e19c99.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 369, + 241, + 377 + ], + "lines": [ + { + "bbox": [ + 229, + 369, + 241, + 377 + ], + "spans": [ + { + "bbox": [ + 229, + 369, + 241, + 377 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 515 + ], + "type": "text", + "content": "definitive answers, image quality lacks fixed standards and is highly subjective. Previous single-objective LLM routing struggles to adapt to such ambiguous quality criteria. On the other hand, prior work lacked dedicated routing model designs, making it challenging to predict quality relationships within the vast output image space. This underscores the necessity of effective routing model design." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 525, + 181, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 181, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 181, + 538 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": "In text-to-image generation scenario, the cloud normally hosts a large T2I model, denoted as " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": ", while a resource-constraint edge device, such as a smartphone, often deploys a light-weight T2I model, denoted as " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_e" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": ". For example, Google has released Imagen [3, 29] for cloud deployment and MobileDiffusion [42] for edge deployment. From the perspective of a user, the large model generally offers superior generation quality, but incurs edge-cloud communication cost and cloud serving fee, which together are denoted as " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "F_c" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": ". In contrast, the edge model can leverage the user's local device, thereby eliminating these additional costs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In practice, user requests involve not only complex ones but also simple ones. One intuitive factor influencing request complexity is the number of entities that need to" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "be generated in images, i.e. the number of nouns in text prompts. As shown in Fig. 2, 3, as noun count increases, the large model is more likely to outperform the small model in quality, with quality differences becoming more noticeable. However, while large models generally excel in overall quality, small models may perform better on specific prompts. This likelihood increases as the prompt becomes simpler. In such cases, the purely large model serving not only fails to improve performance but also incurs high cost. Therefore, it is necessary to design a routing mechanism to select the most suitable model based on the prompt's characteristics. Intuitively, the routing mechanism should direct simpler prompts to the cost-effective edge model and more complex prompts to the high-quality cloud model, thus maximizing generation quality while minimizing cost." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": "Formally, a T2I routing framework " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "R: \\mathcal{X} \\to \\{0,1\\}" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " assigns user prompts " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " to the cloud large model " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " as 1 and to the edge light-weight model " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_e" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " as 0. The generated images after routing can be expressed as " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_r = R(\\mathcal{X})\\mathcal{I}_c + (1 - R(\\mathcal{X}))\\mathcal{I}_e" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_c" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_e" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " denote the images generated using " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_e" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": ", respectively. Given a quality scoring function " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " of the generated images, the cost budget " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\tau_{fee}" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": " of cloud model serving and edge-cloud communication, and the response latency constraint " + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\tau_{time}" + }, + { + "bbox": [ + 313, + 251, + 555, + 384 + ], + "type": "text", + "content": ", the optimization objective is maximizing the quality under the budget and latency constraints, formulated as" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 322, + 390, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 390, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 322, + 390, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\begin{array}{l} \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot F _ {c} \\leq \\tau_ {f e e} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} - D _ {c} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} - D _ {c} \\end{array} \\tag {1} \\\\ \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\cdot D _ {\\mathcal {M} _ {c}} + \\mathbb {P} \\{R (\\mathcal {X}) = 0 \\} \\cdot D _ {\\mathcal {M} _ {e}} \\\\ + D _ {R} \\leq \\tau_ {\\text {t i m e}}, \\\\ \\end{array}", + "image_path": "2a4a545ce1e2447df0408de4ed039d9e6f03ecbefcc3c6fe5604ae77c631fadf.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{M}_c}" + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "inline_equation", + "content": "D_{\\mathcal{M}_e}" + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "inline_equation", + "content": "D_R" + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "content": " denote the latency of cloud large model serving plus edge-cloud communication, the latency of the edge light-weight model serving, and the latency of the routing model execution. Considering the constraints are linear, the optimization objective can be approximated as imposing an upper bound " + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "inline_equation", + "content": "\\rho_r" + }, + { + "bbox": [ + 313, + 455, + 554, + 539 + ], + "type": "text", + "content": " on the routing rate to the cloud large model as" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 353, + 547, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 547, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 353, + 547, + 553, + 574 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max R (\\mathcal {X}) \\mathcal {Q} \\left(\\mathcal {I} _ {c}\\right) + (1 - R (\\mathcal {X})) \\mathcal {Q} \\left(\\mathcal {I} _ {e}\\right) \\\\ \\text {s . t .} \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\} \\leq \\rho_ {r}. \\tag {2} \\\\ \\end{array}", + "image_path": "d92c267d6d0b332752de9f9995847ce8264a993222b1d40d1852de047f02e63c.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Due to the lack of a universal standard for evaluating the quality of generated images, a multi-dimensional quality scoring function " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}:\\mathcal{I}\\to [0,1]^N" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " is required for a more comprehensive comparison. Thus, different from existing routing for typical classification and text generation tasks, the T2I routing framework's optimization objective is a new multi-objective problem. In addition, to predict quality changes in the larger image output space from the smaller text input space, we conduct a detailed analysis of how different prompt tokens influence the generated images and their multi-dimensional quality metrics, by examining" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19484" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "the interaction between text and images during the T2I generation process. In what follow, we define the multi-metric quality measures for routing objectives in Sec. 4 and introduce the T2I routing framework in Sec. 5." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 295, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 157 + ], + "type": "text", + "content": "4. Multi-Metric Image Generation Quality in Routing Objectives" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 235 + ], + "type": "text", + "content": "The connection between text and images allows image attributes to be described through corresponding prompts. Thus, a metric can be measured by the relative relationship between an image and the positive/negative text pair describing that metric [18, 35]. The contrastive quality metric for image " + }, + { + "bbox": [ + 55, + 163, + 296, + 235 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 163, + 296, + 235 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 240, + 295, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 240, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 70, + 240, + 295, + 254 + ], + "type": "interline_equation", + "content": "q (I, m) = \\sigma \\left(C L I P \\left(I, m ^ {+}\\right) - C L I P \\left(I, m ^ {-}\\right)\\right), \\tag {3}", + "image_path": "096adc5a0327b01225bcd5da3d4e223a5fba4f94ffbabc612be89d0391b14f99.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "inline_equation", + "content": "m = (m^{+}, m^{-})" + }, + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "text", + "content": " denotes positive and negative pair, and " + }, + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 258, + 296, + 366 + ], + "type": "text", + "content": " denotes the sigmoid function that transforms output values in the range from 0 to 1. According to this definition, if an image is more related to the positive prompt, and is less related to the negative prompt, the contrastive quality metric is higher. Compared to quality measure with only the positive prompt, the contrastive method evaluates whether positive or negative quality is more dominant in the image, leading to a more robust and reliable evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "content": "Considering the ambiguity and complexity of image quality, we introduce a multi-dimensional metric to comprehensively evaluate generated image quality, providing noise resistance and stability for quality-prediction-based routing. Just as real photo quality depends on factors like definition and color, our multi-metric combines traditional measures of real photos with unique aspects of generated images, such as realism and object integrity. Thus, we extend the single-metric evaluation in Eq. (3) to an " + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "content": "-dimensional quality assessment of an image " + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "content": ", expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 491, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 491, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 103, + 491, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} (I) = \\left[ q (I, m _ {i}) \\mid i = 1, 2, \\dots , N \\right], \\tag {4}", + "image_path": "70167a1c90cd91e17ce0e19297927c16c6eb120d4eb8a814017ae4be4bc3c46d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " is an instantiation of " + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 510, + 296, + 594 + ], + "type": "text", + "content": " in Eq. (3). We adopt 10 common metrics inspired by prior works [14, 37, 38], with details in Appendix Sec. C. For example, for the definition metric, we use the positive prompt \"High definition photo\" and the negative prompt \"Low definition photo\"; and for the object integrity, we take the positive prompt \"Object completion\" and the negative prompt \"Object twisting\"." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "In the multi-objective optimization setting for routing with multi-metric quality, the initial goal is to find a Pareto optimal generated image that excels in all the metrics. However, this is challenging in practice due to the potential nonexistence of such an image generated from the cloud or edge model. We thus relax the constraint of Pareto optimality by allowing suboptimal performance in some metrics if the image significantly outperforms in others when selecting higher-quality images. We define Pareto relative superiority (PRS) to quantify the quality advantage of the images" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 553, + 205 + ], + "type": "image", + "image_path": "1642754cac28ecc02eed28cddaaba7c38d98d82204342cc6613be67deb64e388.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 213, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 289 + ], + "type": "text", + "content": "Figure 4. Overview of RouteT2I. RouteT2I assembles a pair of off-the-shelf edge and cloud text-to-image models and evaluates them using multi-metric image quality. RouteT2I utilizes Pareto relative superiority between qualities as supervision to train the routing model. Then, to balance quality and cost, the routing strategy determines the most suitable model for each user prompt, choosing between the edge or cloud model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_e" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": " generated by the edge model over the images " + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_c" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": " generated by the cloud model. In particular, we first normalize the quality distance for the metric " + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": " between an edge-generated image " + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "I_e \\in \\mathcal{I}_e" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": " and a cloud-generated image " + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "inline_equation", + "content": "I_c \\in \\mathcal{I}_c" + }, + { + "bbox": [ + 313, + 311, + 554, + 360 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 368, + 554, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 368, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 347, + 368, + 554, + 396 + ], + "type": "interline_equation", + "content": "D _ {i} (I _ {e}, I _ {c}) = \\sigma \\left(\\frac {q (I _ {e} , m _ {i}) - q (I _ {c} , m _ {i})}{\\Gamma | \\mu_ {i} (\\mathcal {I} _ {e}) - \\mu_ {i} (\\mathcal {I} _ {c}) |}\\right), \\tag {5}", + "image_path": "8b69d10551a517a7546fbf70d7a5767706e4d29f495db20f3be89749ea5fe0f6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\mu_i(\\cdot)" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": " is the average quality of the set of cloud or edge generated images, while the sigmoid function " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": " and the temperature parameter " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": " are used to modulate the data distribution, effectively distinguishing similar qualities and preventing centralization. We then define the overall quality gap as a weighted sum of distances across all " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": " metrics:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 484, + 553, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 484, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 365, + 484, + 553, + 518 + ], + "type": "interline_equation", + "content": "P R S \\left(I _ {e}, I _ {c}\\right) = \\sum_ {i = 1} ^ {N} w _ {i} D _ {i} \\left(I _ {e}, I _ {c}\\right), \\tag {6}", + "image_path": "15a9ca23aa58129653922d653da2cccc1c9c2e30789fa6148ca48d75490b5df1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": " denotes the weight of metric " + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "inline_equation", + "content": "\\sum_{i} w_{i} = 1" + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": ". We can evaluate the relative quality of " + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "inline_equation", + "content": "I_{e}" + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "inline_equation", + "content": "I_{c}" + }, + { + "bbox": [ + 313, + 526, + 554, + 588 + ], + "type": "text", + "content": " by examining how much Pareto relative superiority deviates from 0.5. This evaluation helps to route a user prompt to the model that generates an image with higher overall quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 437, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 437, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 437, + 611 + ], + "type": "text", + "content": "5. Design of RouteT2I" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "To achieve the optimization objective in Eq. (2), we propose a text-to-image model routing framework RouteT2I. As shown in Fig. 4, RouteT2I comprises a routing model that predicts the multi-metric generation quality given user prompts and a routing strategy that selects cloud or edge model. In particular, (1) the routing model predicts the quality relationship between images generated on the edge and cloud, namely, Pareto relative superiority. Inspired by the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19485" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 90, + 71, + 262, + 155 + ], + "blocks": [ + { + "bbox": [ + 90, + 71, + 262, + 155 + ], + "lines": [ + { + "bbox": [ + 90, + 71, + 262, + 155 + ], + "spans": [ + { + "bbox": [ + 90, + 71, + 262, + 155 + ], + "type": "image", + "image_path": "fba23697dd652034e0744e5e4567e6470325d210cae48cc56f74a5576d3b4970.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 159, + 215, + 169 + ], + "lines": [ + { + "bbox": [ + 135, + 159, + 215, + 169 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 215, + 169 + ], + "type": "text", + "content": "(a) Token selection gate." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 518, + 157 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 518, + 157 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 518, + 157 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 518, + 157 + ], + "type": "image", + "image_path": "4fcdf9af33b1f26cc5f64f2650bc2e4bc46c0b98bb5747a72b9d7503897021c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 159, + 433, + 169 + ], + "lines": [ + { + "bbox": [ + 368, + 159, + 433, + 169 + ], + "spans": [ + { + "bbox": [ + 368, + 159, + 433, + 169 + ], + "type": "text", + "content": "(b) Dual-gate MoE." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 178, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 54, + 178, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 54, + 178, + 555, + 213 + ], + "type": "text", + "content": "Figure 5. Architecture of a dual-gate MoE that uses the token selection gate as the gate network. First, the gate selects the most relevant tokens for each expert. Then, experts project the tokens selected by positive and negative gates into their respective spaces and evaluate them using score matrices. Finally, the evaluations of the tokens are gathered and compared to form new tokens." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 233, + 297, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 297, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 297, + 389 + ], + "type": "text", + "content": "role of prompts in generation, where prompt as sequences of tokens interact with images through cross-attention to determine the image content, we treat prompt as a set of tokens in the routing model and select key tokens with influential contextual information for each quality metric. Then, the dominant influence of these tokens is accessed by comparing their positive and negative effects; and (2) the routing strategy aims to maximize quality under cost constraints. The key idea is to use Pareto relative superiority to describe the quality disparity of images generated by the cloud and the edge models. Only prompts that show a notable quality gap when generated in the cloud are routed there, while the rest remain on the cost-effective edge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 420 + ], + "type": "text", + "content": "5.1. Routing Model Architecture Design with Dual-Gate Token Selection MoE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "To predict Pareto relative superiority between the images generated by the cloud model and the edge model from user prompts, we design a dual-gate token selection mixture-of-experts (MoE), as shown in Fig. 5, and introduce it into Transformer network to replace linear layers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "type": "text", + "content": "Our model treats the user prompt as a sequence of tokens, whose influences on the image quality differ due to their weights in cross-attention during generation. To focus on key tokens for each metric, we design a token selection gate, where experts align with quality metrics and actively choose the most relevant tokens to simulate metric attention to different image attributes. Only tokens that significantly impact the quality metrics corresponding to the experts are selected, while those with minimal influences are disregarded, reducing potential interference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 296, + 714 + ], + "type": "text", + "content": "Regarding the design of a dual-gate MoE, both positive and negative gates are introduced to distinguish the opposing aspects of token influences on image quality. Similar to images that have both superior and inferior regions when evaluated by metrics, these tokens are assessed by the corresponding experts for their positive and negative influences to each metric. Additionally, we introduce a comparison between two gates to identify which influence of each token predominates overall. We further improve the structure of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 256 + ], + "type": "text", + "content": "expert to handle both positive and negative influences on a single metric simultaneously." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 257, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 329 + ], + "type": "text", + "content": "To support multi-metric quality prediction, our design incorporates multiple prediction heads within the model architecture. Each head outputs the prediction of a specific quality metric. Such multi-head design not only enhances noise resistance, thereby increasing robustness, but also facilitates a comprehensive evaluation of quality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 433, + 343 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 433, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 433, + 343 + ], + "type": "text", + "content": "5.1.1. Token Selection Gate" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "text", + "content": "In the token selection gate, the dot product between the token representations " + }, + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "text", + "content": " and the expert embeddings " + }, + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{k \\times d}" + }, + { + "bbox": [ + 313, + 347, + 555, + 384 + ], + "type": "text", + "content": " is computed as an affinity matrix:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 387, + 553, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 387, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 386, + 387, + 553, + 402 + ], + "type": "interline_equation", + "content": "A = \\operatorname {S o f t m a x} \\left(T \\cdot E ^ {T}\\right), \\tag {7}", + "image_path": "12e8f5f366bb2438e632c3f515cc1a495613fe5227d770ecb81c5f6d471ef614.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " denotes the number of tokens in user prompt, " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " denotes the number of experts, " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " denotes the hidden dimension, and the Softmax function is applied along the expert axis. " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "A[t,i]" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " in the affinity matrix captures the correlation between token " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " and expert " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": ". Then, top- " + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 406, + 554, + 478 + ], + "type": "text", + "content": " tokens with the highest relevance for each expert are selected" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 483, + 553, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 483, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 399, + 483, + 553, + 496 + ], + "type": "interline_equation", + "content": "M = \\operatorname {T o p} - \\mathrm {K} (A), \\tag {8}", + "image_path": "32be201609eeaa9ad301b532e7f783cf75912e37f69ccd7ecf4cc6080ead27a3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{n \\times k}" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " is the binary mask matrix such that " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "M[t, i] = 1" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " denotes token " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " is selected by expert " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "M[t, i] = 0" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " denotes not selected. Experts tend to have similar weights because the metrics they are responsible for are weighted similarly. This ensures that the tokens selected by these experts possess significant global importance. This practice minimizes interference from tokens that are insignificant across all metrics. The influences of key tokens are further evidenced by their frequent selection through the top- " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " function and their corresponding scores within the affinity matrix. Moreover, by allocating an equal number of " + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 501, + 555, + 681 + ], + "type": "text", + "content": " tokens to each expert, workloads remain balanced, and bias in quality dimensions is avoided, ultimately resulting in more comprehensive and reliable quality prediction. Overall, the token selection gate is summarized as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 395, + 697, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 697, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 395, + 697, + 553, + 711 + ], + "type": "interline_equation", + "content": "A, M = G (T, E), \\tag {9}", + "image_path": "a7b61dfba89f99205f7e852a0066209071b4b784976a518c10ef3f0b540bde8c.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19486" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "text", + "content": " denotes the gate network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 151, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 151, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 151, + 100 + ], + "type": "text", + "content": "5.1.2. Dual-Gate MoE" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "text", + "content": "Both a positive gate " + }, + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "inline_equation", + "content": "G^{+}" + }, + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "text", + "content": " and a negative gate " + }, + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "inline_equation", + "content": "G^{-}" + }, + { + "bbox": [ + 55, + 104, + 296, + 163 + ], + "type": "text", + "content": " are introduced for different experts to select tokens that have positive and negative impacts on their corresponding quality metrics, while capturing the weights of their influences. The token selection results are denoted as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 171, + 294, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 171, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 80, + 171, + 294, + 184 + ], + "type": "interline_equation", + "content": "A ^ {o}, M ^ {o} = G ^ {o} \\left(T, E ^ {o}\\right), \\text {w h e r e} o \\in \\{+, - \\}. \\tag {10}", + "image_path": "003007744ca12755257600b889b7e471a1a0c29c6af9a2c654a023586b818e9d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": "The affinity matrix " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "A^o" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": " and the binary mask matrix " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "M^o" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": " are captured by the gate " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "G^o" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": " for the positive or negative aspect using the corresponding expert embedding " + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "inline_equation", + "content": "E^o" + }, + { + "bbox": [ + 55, + 191, + 296, + 251 + ], + "type": "text", + "content": ". The tokens selected by two gates are then evaluated by the corresponding positive or negative experts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": "To eliminate the redundancy of having separate experts for evaluating both positive and negative aspects of the same quality metric, the structure of experts is revised to allow an expert to assess both aspects simultaneously. This revision acknowledges that both positive and negative evaluations are inherently part of the same quality metric. Specifically, the original linear layer " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "W_{i} \\in \\mathbb{R}^{d \\times h}" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " of expert on metric " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " is split into two projection matrices " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "P_{i}^{+}, P_{i}^{-} \\in \\mathbb{R}^{d \\times l}" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " and one score matrix " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "S_{i} \\in \\mathbb{R}^{l \\times h}" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": ". Two projection matrices are specific to positive and negative aspects, projecting tokens into the corresponding representations in an " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": "-dimensional low-dimensional space. Since " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "l \\ll h, d" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": ", the parameter size and the computation cost are reduced from " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "O(hd)" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "O(l(h + d))" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": ", allowing an expert to efficiently focus on both positive and negative aspects of a metric with a low cost. The shared score matrix then evaluates the impact of these representations on metric " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": ". Thus, the impact of " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": "-th token " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "T[t] \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " on quality metric " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " for positive or negative aspect " + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 251, + 296, + 478 + ], + "type": "text", + "content": " is described as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 486, + 294, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 486, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 88, + 486, + 294, + 499 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {o} [ t ] = T [ t ] \\cdot P _ {i} ^ {o} \\cdot S _ {i}, \\text {w h e r e} o \\in \\{+, - \\}. \\tag {11}", + "image_path": "7fb22f26b3657898155edeb76185a138bfea420f5985cc6a0ee60499df394bee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 578 + ], + "type": "text", + "content": "By separating the projection matrices and score matrix, we can evaluate both positive and negative impacts on a metric within the same expert. Then, the overall positive or negative impact of the " + }, + { + "bbox": [ + 55, + 506, + 296, + 578 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 506, + 296, + 578 + ], + "type": "text", + "content": "-th token on image quality can be obtained by a weighted sum of its impacts for corresponding aspects across all metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 586, + 294, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 586, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 77, + 586, + 294, + 617 + ], + "type": "interline_equation", + "content": "T ^ {o} [ t ] = \\sum_ {i = 0} ^ {k} \\lambda_ {t, i} ^ {o} \\cdot T _ {i} ^ {o} [ t ], \\text {w h e r e} o \\in \\{+, - \\}. \\tag {12}", + "image_path": "a391119cfd322d46a666a387303b3f9cdb68812ddb8e200a3196ef7ac88013dd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "inline_equation", + "content": "\\lambda_{t,i}^{o} = \\frac{\\mathbb{1}\\{M^{o}[t,i] = 1\\}A^{o}[t,i]}{\\sum_{j = 0}^{k}\\mathbb{1}\\{M^{o}[t,j] = 1\\}A^{o}[t,j]}" + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "text", + "content": " denotes the normalized affinity of the " + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "text", + "content": "-th expert for the " + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 624, + 296, + 690 + ], + "type": "text", + "content": "-th token. By contrasting the impact of a token from both positive and negative aspects, we can assess the token's predominant influence on the image quality as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 121, + 696, + 294, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 696, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 121, + 696, + 294, + 711 + ], + "type": "interline_equation", + "content": "\\hat {T} [ t ] = \\sigma \\left(T ^ {+} [ t ] - T ^ {-} [ t ]\\right). \\tag {13}", + "image_path": "041df0f3aa06e05adfb6cbc6dbd6322c41b5cf3b1328ae9db3ec100708814f04.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "The introduction of contrasting helps diminish ambiguity in predictions, especially when tokens exhibit influences on multiple metrics from both positive and negative aspects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 114, + 452, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 452, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 452, + 128 + ], + "type": "text", + "content": "5.2. Routing Strategy Design" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 132, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 300 + ], + "type": "text", + "content": "Given the Pareto relative superiority predicted by routing model, we propose a routing strategy to effectively route user prompts to suitable models to achieve the trade-off between quality and costs. With a preset routing rate, we can determine the proportion of user prompts sent to the cloud, focusing on those where the cloud model significantly outperforms the edge model in quality. Since Pareto relative superiority represents this quality gap, we can efficiently filter prompts by setting a threshold " + }, + { + "bbox": [ + 313, + 132, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 132, + 555, + 300 + ], + "type": "text", + "content": " on it. In particular, the prompts with a Pareto relative superiority above the threshold are better handled by the edge model for cost efficiency, while those below the threshold achieve superior quality when processed in the cloud. Thus, the optimization objective in Eq. (2) can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 306, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 306, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 306, + 555, + 326 + ], + "type": "interline_equation", + "content": "\\max _ {\\alpha \\leq 1 / 2} \\mathbb {P} \\left\\{P R S \\left(I _ {e}, I _ {c}\\right) < \\alpha \\mid I _ {e}, I _ {c} \\in \\mathcal {I} _ {e}, \\mathcal {I} _ {c} \\right\\} \\leq \\rho_ {r}. \\tag {14}", + "image_path": "f3ad603e2cce9c2db59e1ff3390634846c1592aa99895d29dd2f4743647617e0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "text", + "content": "Since Pareto relative superiority being greater than or less than " + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "text", + "content": " indicates a relative advantage or disadvantage, respectively, we set a upper bound " + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 331, + 554, + 392 + ], + "type": "text", + "content": " to prevent user prompts with images generated at the edge of better quality from being routed to the cloud." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 391, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 391, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 391, + 415 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "type": "text", + "content": "6.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Dataset. We take COCO2014 [19], a comprehensive dataset for object detection, segmentation, and captioning tasks. We select its captions as user prompts " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": ". Given these prompts, we generate images " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_e, \\mathcal{I}_c" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " using different open-source text-to-image models. Similarly, we use the LAION dataset [31], with results shown in Appendix Sec. F." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": "Models. We take T2I models with varying performances and sizes for cloud and edge usage, include Stable Diffusion 1.5 (SD1.5) [28], Stable Diffusion 2.1 (SD2.1) [28], Stable Diffusion XL (SDXL) [24], Stable Diffusion XL-Refiner (XL-Refiner) [24], Stable Diffusion 3 (SD3) [9], PixArt" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 510, + 555, + 617 + ], + "type": "text", + "content": " [5], and Infinity [10]. These models use various text encoders, such as CLIP [25], OpenCLIP [7], and T5 [26], and are based on architectures like diffusion, diffusion Transformer and autoregressive. Details are in Appendix Sec. A." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "Baselines and Settings. We first introduce random routing as a baseline, where prompts are randomly assigned to candidate models, given a specified routing rate to the cloud. Additionally, although there does not exist any previous work on text-to-image generation routing, for a comprehensive comparison, we also reproduce several representative routing methods for LLM by adapting them to our scenarios, including RouteLLM [23] with BERT classifier or" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19487" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 159 + ], + "type": "table", + "html": "
RouterImage Quality MetricsΔP(%)
DefinitionDetailClaritySharpnessHarmonyRealismColorConsistencyLayoutIntegrity
Edge Model0.62510.66850.60760.65370.59490.55750.46800.50880.48600.4690-
Cloud Model0.63370.68470.63460.67030.59300.58680.51340.51990.53450.4972-
Random0.62940.67660.62110.66200.59390.57210.49070.51440.51020.483140.00
RouteLLM-BERT [23]0.63470.67920.63050.66510.59600.57880.49820.51600.51670.486671.51
RouteLLM-MF [23]0.63640.68140.62990.66600.59520.57760.49700.51640.51490.485069.90
Hybrid LLM [8]0.63270.67840.63060.66770.59640.57870.50080.51610.51910.486473.49
ZOOTER [21]0.63500.67960.63150.66720.59660.57880.50040.51660.51790.485477.95
RouteT2I (Ours)0.63500.67860.63180.66790.59750.58040.50100.51670.51890.486583.97
", + "image_path": "092065c5714afe1210b9e9a7a282a8f9b041d7a5ea8c71bbb1c381ab9157b869.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 211, + 294, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 294, + 270 + ], + "type": "text", + "content": "matrix factorization, Hybrid LLM [8], and ZOOTER [21]. Specifically, we keep their text input processing and quality threshold design, but replace their optimization goal with the average image quality across metrics as they only handle single-objective optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 342 + ], + "type": "text", + "content": "Unless specified in their paper, the routing models are based on Transformers and take the same hyperparameters as in our method. All routers are implemented in PyTorch 2.3.1 and trained using Adam optimizer with a learning rate of 2e-5, a weight decay of 0, a batch size of 16, on a NVIDIA 4090D for about 10 epochs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 367 + ], + "type": "text", + "content": "Metrics. For cost efficiency, we introduce routing rate, namely, the proportion of user prompts to the cloud model:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 386, + 294, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 386, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 132, + 386, + 294, + 399 + ], + "type": "interline_equation", + "content": "p = \\mathbb {P} \\{R (\\mathcal {X}) = 1 \\}. \\tag {15}", + "image_path": "705165e6092c6e17c35dd94ce52b1dc2c1940b5780d0e5b8e6962718dd31ae76.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 263, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 263, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 263, + 418 + ], + "type": "text", + "content": "Smaller routing rate indicates better cost efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": "For image generation quality, the absolute performance of routing is constrained by the gap between edge and cloud T2I models. Thus, we introduce a set of relative performance metrics for evaluation. We first fix a routing rate " + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": " and introduce the win rate of the selected models over the cloud model after routing:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 498, + 294, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 498, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 66, + 498, + 294, + 514 + ], + "type": "interline_equation", + "content": "w \\left(\\mathcal {I} _ {r, p}\\right) = \\mathbb {P} \\left\\{\\mathcal {Q} \\left(I _ {r}\\right) \\geq \\mathcal {Q} \\left(I _ {c}\\right) \\mid I _ {r}, I _ {c} \\in \\mathcal {I} _ {r, p}, \\mathcal {I} _ {c} \\right\\}, \\tag {16}", + "image_path": "71d3cb0d27f5d52a07ced7406fe983670feb1223bf6c4fb4e2e956fa56d4623a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{r,p}" + }, + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "text", + "content": " denotes the generated images after routing at a routing rate " + }, + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 520, + 296, + 605 + ], + "type": "text", + "content": ". To intuitively show the improvement in win rate, we introduce the normalized win rate improvement, which is the increase in win rate achieved by a routing method compared to the baseline, divided by the increase achieved by the optimal oracle routing. A higher value indicates that the routing method is closer to the oracle routing:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 620, + 296, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 620, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 113, + 620, + 296, + 648 + ], + "type": "interline_equation", + "content": "\\Delta \\bar {w} (p) = \\frac {w \\left(\\mathcal {I} _ {r , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}{w \\left(\\mathcal {I} _ {o , p}\\right) - w \\left(\\mathcal {I} _ {b , p}\\right)}, \\tag {17}", + "image_path": "4e75fe3ab6f680995f96bf8a999dcd0054b2beca2c0ff19d2013e4e8ae6a6df9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{b,p}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{o,p}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " denote the generated images using the random baseline and oracle, respectively. To quantify the improvement of image quality metrics, we introduce relative performance improvement, which measures the proportion of the performance gain achieved through routing" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 329, + 209, + 541, + 270 + ], + "blocks": [ + { + "bbox": [ + 55, + 167, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 555, + 191 + ], + "type": "text", + "content": "Table 2. The multi-dimensional quality of images generated by text-to-image models, with the router selecting between edge and cloud models for each prompt at a " + }, + { + "bbox": [ + 55, + 167, + 555, + 191 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 167, + 555, + 191 + ], + "type": "text", + "content": " routing rate. The higher the metrics, the better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 209, + 541, + 270 + ], + "lines": [ + { + "bbox": [ + 329, + 209, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 329, + 209, + 541, + 270 + ], + "type": "table", + "html": "
Router\\p40%50%60%70%80%
RouteLLM-BERT [23]24.2922.4519.0717.6220.59
RouteLLM-MF [23]25.6523.2919.9216.6720.59
Hybrid LLM [8]23.7719.7514.9213.6216.09
ZOOTER [21]26.7721.9717.9716.9721.04
RouteT2I (Ours)30.6025.8120.3218.0221.94
", + "image_path": "7a6d158b858116da566d7d33dfc4677aa4b9ad7fd05565206ea1aca3d6dfc02b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 278, + 554, + 301 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 301 + ], + "type": "text", + "content": "Table 3. Normalized win rate improvements " + }, + { + "bbox": [ + 313, + 278, + 554, + 301 + ], + "type": "inline_equation", + "content": "\\Delta \\overline{w}" + }, + { + "bbox": [ + 313, + 278, + 554, + 301 + ], + "type": "text", + "content": " (\\%) at different routing rates, i.e. the ratio of improvement to oracle improvement." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 320, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 344 + ], + "type": "text", + "content": "compared to the edge model, relative to the total possible improvement brought by fully using the cloud model:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 365, + 350, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 350, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 365, + 350, + 553, + 384 + ], + "type": "interline_equation", + "content": "\\Delta P = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {\\mu_ {i} \\left(\\mathcal {I} _ {r}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right)}{\\left| \\mu_ {i} \\left(\\mathcal {I} _ {c}\\right) - \\mu_ {i} \\left(\\mathcal {I} _ {e}\\right) \\right|}. \\tag {18}", + "image_path": "986b15f7b5f2d64173c3fa1b7a872fe9e8734f4ce1c7114ad45886e8214adce6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 388, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 412 + ], + "type": "text", + "content": "This metric effectively quantifies routing effectiveness while accounting for the original quality gap." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "content": "For cost-quality balance, we introduce the cost saving to measure the reduction of the routing rate to the cloud model, compared to the random baseline at a given relative performance improvement " + }, + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 465, + 553, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 465, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 369, + 465, + 553, + 493 + ], + "type": "interline_equation", + "content": "\\gamma (\\Delta P) = \\frac {p _ {b} (\\Delta P) - p _ {r} (\\Delta P)}{p _ {b} (\\Delta P)}, \\tag {19}", + "image_path": "b33c7d95266c673c12a7af2380e00de1901c07836b6f8812d17e37fc86a78364.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "inline_equation", + "content": "p_b(\\Delta P)" + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "inline_equation", + "content": "p_r(\\Delta P)" + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "text", + "content": " represent the routing rates of baseline and router at a given " + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 313, + 498, + 555, + 523 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 529, + 400, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 400, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 400, + 540 + ], + "type": "text", + "content": "6.2. Main Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 546, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 594 + ], + "type": "text", + "content": "We present the routing performance of our RouteT2I and baselines using SD3 as the cloud model and SD2.1 as the edge model. Human evaluation, cost analysis, visual results, and more T2I pair results are provided in Appendix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": "Win Rates. Tab. 3 shows our RouteT2I demonstrates a significant improvements on normalized win rate improvements across all routing rates. Specifically, at a " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " routing rate, our improvement reaches " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "30.60\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": ", exceeding the baselines by at least " + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "inline_equation", + "content": "3.83\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 665 + ], + "type": "text", + "content": " and demonstrating strong consistency with the oracle." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Image Quality. Tab. 2 shows RouteT2I outperforms all baselines in 6 of 10 metrics. Our multi-metric optimization balances all metrics simultaneously, while others focus on just a few, often at the cost of others. We" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19488" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "type": "table", + "html": "
ΔP Router40%50%60%70%80%
RouteLLM-BERT [23]56.1551.3946.9242.7040.21
RouteLLM-MF [23]48.8649.9048.2044.8941.50
Hybrid LLM [8]62.0658.8553.6349.9233.38
ZOOTER [21]69.2865.7660.8157.3549.64
RouteT2I (Ours)71.8170.2466.6160.0153.53
", + "image_path": "5c8ca62473dac4f3f24a8621da42127503a220f825d59dd7fa1a8be75a44aecf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 184, + 175, + 256 + ], + "blocks": [ + { + "bbox": [ + 75, + 184, + 175, + 256 + ], + "lines": [ + { + "bbox": [ + 75, + 184, + 175, + 256 + ], + "spans": [ + { + "bbox": [ + 75, + 184, + 175, + 256 + ], + "type": "image", + "image_path": "69c67ba2cf1dbda5d25956402a4ade59774f323ac9a92217ee3331020bd77084.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 258, + 130, + 267 + ], + "lines": [ + { + "bbox": [ + 120, + 258, + 130, + 267 + ], + "spans": [ + { + "bbox": [ + 120, + 258, + 130, + 267 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 184, + 277, + 255 + ], + "blocks": [ + { + "bbox": [ + 177, + 184, + 277, + 255 + ], + "lines": [ + { + "bbox": [ + 177, + 184, + 277, + 255 + ], + "spans": [ + { + "bbox": [ + 177, + 184, + 277, + 255 + ], + "type": "image", + "image_path": "228111b9e8cc0d6ab2e87be145202cdd29122e322b9cc3680b153e94bbf2492a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 258, + 232, + 267 + ], + "lines": [ + { + "bbox": [ + 220, + 258, + 232, + 267 + ], + "spans": [ + { + "bbox": [ + 220, + 258, + 232, + 267 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "text", + "content": "Figure 6. Visualization of (a) Normalized win rate improvement " + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\Delta \\overline{w}" + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "text", + "content": " and (b) relative performance improvement " + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\Delta P" + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "text", + "content": " at varying routing rates " + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 277, + 296, + 310 + ], + "type": "text", + "content": " to the cloud." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": "also achieve the highest relative performance improvement, reaching " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "83.97\\%" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": " of cloud model's gain over edge model, surpassing others by at least " + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": ". Experiments in Appendix Sec. H show our method may lead to performance gains beyond those of the cloud model itself and is effective for T2I models with various text encoders and architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": "Cost Efficiency. Tab. 4 presents the cost savings at given relative performance improvement targets. At a " + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": " relative performance improvement, RouteT2I reduce cloud service needs by up to " + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "inline_equation", + "content": "71.81\\%" + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": " of the calls required by random routing and by at least " + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "inline_equation", + "content": "5.80\\%" + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": " compared to other baselines. In other cases, RouteT2I still reduces over " + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 402, + 296, + 485 + ], + "type": "text", + "content": " costs, demonstrating high cost efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 486, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 521 + ], + "type": "text", + "content": "Visualization. As shown in Fig. 6, RouteT2I performs well across routing rates, demonstrating its ability to predict quality gaps between T2I models and route effectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 521, + 263, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 521, + 263, + 534 + ], + "spans": [ + { + "bbox": [ + 67, + 521, + 263, + 534 + ], + "type": "text", + "content": "Visual examples are shown in Appendix Sec. G." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 540, + 295, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 553 + ], + "type": "text", + "content": "6.3. Routing Performance on More T2I Model Pairs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "content": "To test our method across T2I model pairs, Tab. 5 shows RouteT2I's performance with various cloud and edge models. Results reveal significant improvements, especially for pairs with large quality gaps, such as SD3 on cloud and others on edge, where RouteT2I achieves over " + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "content": " of the oracle routing's improvement at a " + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "content": " routing rate. Even with similar model pairs like XL-Refiner and SDXL, it reaches " + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 666 + ], + "type": "text", + "content": " of the oracle's improvement, demonstrating its ability to discern subtle differences between T2I models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 178, + 685 + ], + "type": "text", + "content": "6.4. Ablation Experiment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Tab. 6 presents ablation results for RouteT2I's multi-metric quality optimization, token selection gate, and dual" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "text", + "content": "Table 4. Cost saving " + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "text", + "content": " (\\%) under a given relative performance improvements " + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "inline_equation", + "content": "(\\Delta P)" + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "text", + "content": " target. " + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 140, + 295, + 174 + ], + "type": "text", + "content": " is the ratio of savings to the routing rate required by the random baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 228 + ], + "type": "table", + "html": "
Cloud-Edgep40%50%60%70%80%
SD3XL-Refiner27.4222.3918.0516.8520.89
SD3SDXL29.7826.8122.3516.9821.30
SD3SD1.526.2420.8816.1914.0419.22
SD2.1XL-Refiner21.6218.3120.4625.3028.66
SD2.1SDXL19.4616.8717.2720.4223.69
SD2.1SD1.510.779.2810.1212.2215.76
XL-RefinerSDXL13.5710.159.7411.8112.59
XL-RefinerSD1.519.6516.9818.8722.7925.57
SDXLSD1.518.3817.0121.5724.7532.49
SD3PixArt-α32.2627.5924.7430.7338.97
PixArt-αSD2.125.8221.2622.5528.1933.76
PixArt-αSDXL16.7113.5712.3716.3522.14
PixArt-αXL-Refiner22.5119.1418.1823.7228.44
PixArt-αSD1.522.3820.5219.9325.4032.43
SD3Infinity22.1919.9717.7719.8125.39
SD2.1Infinity21.9819.8622.6028.4536.18
PixArt-αInfinity27.6923.4824.1230.8938.91
", + "image_path": "e57877800500c575387534514b260a4a44ea1440ba8f7fd416da6933e947f574.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 329, + 270, + 541, + 326 + ], + "blocks": [ + { + "bbox": [ + 313, + 235, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 258 + ], + "type": "text", + "content": "Table 5. Normalized win rate improvement " + }, + { + "bbox": [ + 313, + 235, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\Delta \\overline{w} (\\%)" + }, + { + "bbox": [ + 313, + 235, + 555, + 258 + ], + "type": "text", + "content": " over oracle routing of our RouteT2I with different cloud-edge model pairs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 270, + 541, + 326 + ], + "lines": [ + { + "bbox": [ + 329, + 270, + 541, + 326 + ], + "spans": [ + { + "bbox": [ + 329, + 270, + 541, + 326 + ], + "type": "table", + "html": "
Router\\p40%50%60%70%80%
w/o Multi-Metric27.3722.8118.8716.4719.92
w/o Token Selection27.8223.0518.2716.7719.24
w/o Dual-Gate27.2222.0919.2717.3721.62
RouteT2I30.6025.8120.3218.0221.94
", + "image_path": "406f07290d81eb9c80e0d6a285c6d7d30e678f078d0598396f04a08ccd6b6898.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": "Table 6. Ablation experiment of RouteT2I, showing normalized win rate improvement " + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\Delta \\overline{w}" + }, + { + "bbox": [ + 313, + 334, + 555, + 357 + ], + "type": "text", + "content": " (%) on given routing rates." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": "gate MoE. The multi-metric quality optimization allows routers to comprehensively assess the image quality gap, significantly enhancing overall performance. Removing this results in a performance decline across all routing rates, particularly at " + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": " routing rate, where the drop is approximately " + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "inline_equation", + "content": "3.23\\%" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": ". The token selection gate prioritizes critical tokens, ensuring routing focuses on key quality factors while reducing interference from irrelevant elements. Its removal notably impacts high routing rates, causing a performance decrease of around " + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": " routing rate. The dual-gate MoE effectively distinguishes subtle differences between cloud and edge model outputs, proving most effective at medium routing rates, especially near " + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 380, + 555, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 548, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 388, + 562 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "In this work, we have proposed a new text-to-image model routing design RouteT2I between edge and cloud. RouteT2I adopts the routing model with a dual-gate token selection MoE to predict how user prompts affect multiple image quality metrics by identifying and evaluating key tokens with contrastive method. RouteT2I further introduces a routing strategy that exploits the predicted advantage of the edge model over the cloud model to determine which side for image generation. Extensive evaluation has demonstrated that RouteT2I can significantly enhance generation quality at a specified routing rate and meanwhile reduce cost at a given quality target." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19489" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 128, + 85 + ], + "type": "text", + "content": "Acknowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 199 + ], + "type": "text", + "content": "This work was supported in part by National Key R&D Program of China (No. 2022ZD0119100), China NSF grant No. 62202297, No. 62025204, No. 62202296, No. 62272293, No. 62441236, and No. U24A20326, SJTU-Huawei Research Program, and Tencent WeChat Research Program. The opinions, findings, conclusions, and recommendations expressed in this paper are those of the authors and do not necessarily reflect the views of the funding agencies or the government." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 228, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "type": "text", + "content": "[1] stabilityai/stable-diffusion-3.5-large · Hugging Face — huggingface.co. https://huggingface.co/stabilityai/stable-diffusion-3.5-large.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 262, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 296, + 327 + ], + "type": "text", + "content": "[2] Szymon Antoniak, Michal Krutul, Maciej Pioro, Jakub Krajewski, Jan Ludziejewski, Kamil Ciebiera, Krystian Król, Tomasz Odrzygóźdź, Marek Cygan, and Sebastian Jaszcur. Mixture of tokens: Continuous moe through cross-example aggregation. Advances in Neural Information Processing Systems, 37:103873-103896, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 296, + 371 + ], + "type": "text", + "content": "[3] Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, et al. Imagen 3. arXiv preprint arXiv:2408.07009, 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "type": "text", + "content": "[4] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. Ld-pruner: Efficient pruning of latent diffusion models using task-agnostic insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 821–830, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "type": "text", + "content": "[5] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 296, + 525 + ], + "type": "text", + "content": "[6] Lingjiao Chen, Matei Zaharia, and James Zou. Frugalgpt: How to use large language models while reducing cost and improving performance. arXiv preprint arXiv:2305.05176, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 526, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 592 + ], + "type": "text", + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2818-2829, 2023. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 593, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 647 + ], + "type": "text", + "content": "[8] Dujian Ding, Ankur Mallick, Chi Wang, Robert Sim, Subhabrata Mukherjee, Victor Ruhle, Laks VS Lakshmanan, and Ahmed Hassan Awadallah. Hybrid llm: Cost-efficient and quality-aware query routing. arXiv preprint arXiv:2404.14618, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 713 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[10] Jian Han, Jinlai Liu, Yi Jiang, Bin Yan, Yuqi Zhang, Zehuan Yuan, Bingyue Peng, and Xiaobing Liu. Infinity: Scaling bitwise autoregressive modeling for high-resolution image synthesis. arXiv preprint arXiv:2412.04431, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 171 + ], + "type": "text", + "content": "[11] Qitian Jason Hu, Jacob Bieker, Xiuyu Li, Nan Jiang, Benjamin Keigwin, Gaurav Ranganath, Kurt Keutzer, and Shriyash Kaustubh Upadhyay. *Routerbench: A benchmark for multi-llm routing system.* arXiv preprint arXiv:2403.12031, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 225 + ], + "type": "text", + "content": "[12] Shijing Hu, Ruijun Deng, Xin Du, Zhihui Lu, Qiang Duan, Yi He, Shih-Chia Huang, and Jie Wu. Laecips: Large vision model assisted adaptive edge-cloud collaboration for iot-based perception system. arXiv preprint arXiv:2404.10498, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 259 + ], + "type": "text", + "content": "[13] Anil Kag and Igor Fedorov. Efficient edge inference by selective query. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "type": "text", + "content": "[14] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36:69981-70011, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "text", + "content": "[15] Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Hannah Teufel, Marco Bellagente, et al. Holistic evaluation of text-to-image models. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "text", + "content": "[16] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 481 + ], + "type": "text", + "content": "[17] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "text", + "content": "[18] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "text", + "content": "[19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 647 + ], + "type": "text", + "content": "[20] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[21] Keming Lu, Hongyi Yuan, Runji Lin, Junyang Lin, Zheng Yuan, Chang Zhou, and Jingren Zhou. Routing to the expert: Efficient reward-guided ensemble of large language models. arXiv preprint arXiv:2311.08692, 2023. 1, 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[22] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19490" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "text", + "content": "resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[23] Isaac Ong, Amjad Almahairi, Vincent Wu, Wei-Lin Chiang, Tianhao Wu, Joseph E Gonzalez, M Waleed Kadous, and Ion Stoica. Routellm: Learning to route llms with preference data. arXiv preprint arXiv:2406.18665, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[24] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[26] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 487 + ], + "type": "text", + "content": "[29] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "text", + "content": "[30] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 600 + ], + "type": "text", + "content": "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 645 + ], + "type": "text", + "content": "[32] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[33] Tal Shnitzer, Anthony Ou, Mfrian Silva, Kate Soule, Yuekai Sun, Justin Solomon, Neil Thompson, and Mikhail Yurochkin. Large language model routing with benchmark datasets. arXiv preprint arXiv:2309.15789, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[34] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for dif" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 542 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "fusion models. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[35] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555-2563, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[36] Yiding Wang, Kai Chen, Haisheng Tan, and Kun Guo. Tabi: An efficient multi-level inference system for large language models. In Proceedings of the Eighteenth European Conference on Computer Systems, pages 233-248, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 251 + ], + "type": "text", + "content": "[37] Haoning Wu, Xiele Wu, Chunyi Li, Zicheng Zhang, Chaofeng Chen, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. T2i-scorer: Quantitative evaluation on text-to-image generation via fine-tuned large multi-modal models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3676-3685, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[38] Xun Wu, Shaohan Huang, Guolong Wang, Jing Xiong, and Furu Wei. Multimodal large language models make text-to-image generative models align better. Advances in Neural Information Processing Systems, 37:81287-81323, 2024. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[39] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[40] Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. Large language model cascades with mixture of thoughts representations for cost-efficient reasoning. arXiv preprint arXiv:2310.03094, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "text", + "content": "[41] Tianchen Zhao, Xuefei Ning, Tongcheng Fang, Enshu Liu, Guyue Huang, Zinan Lin, Shengen Yan, Guohao Dai, and Yu Wang. Mixdq: Memory-efficient few-step text-to-image diffusion models with metric-decoupled mixed precision quantization. arXiv preprint arXiv:2405.17873, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 487 + ], + "type": "text", + "content": "[42] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 542 + ], + "type": "text", + "content": "[43] Yanqi Zhou, Tao Lei, Hanxiao Liu, Nan Du, Yanping Huang, Vincent Zhao, Andrew M Dai, Quoc V Le, James Laudon, et al. Mixture-of-experts with expert choice routing. Advances in Neural Information Processing Systems, 35:7103-7114, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19491" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_content_list.json b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d2f3670ed9859708aaa3df7cddcb2b27633b8c72 --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_content_list.json @@ -0,0 +1,1887 @@ +[ + { + "type": "text", + "text": "AdaptiveAE: An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyi Xu $^{1,3,4*}$ Fan Zhang $^{1}$ Boxin Shi $^{3,4\\dagger}$ Tianfan Xue $^{2,1\\dagger}$ Yujin Wang $^{1\\dagger}$ $^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 176, + 200, + 821, + 239 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{3}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University", + "$^{4}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 169, + 241, + 825, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "photon@stu.pku.edu.cn, zhangfan@pjlab.org.cn, shiboxin@pu.edu.cn", + "bbox": [ + 209, + 277, + 785, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tfxue@ie.cuhk.edu.hk, wangyujin@pjlab.org.cn", + "bbox": [ + 300, + 295, + 692, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57f3c8d85b7a63c3a2fbc6de4a21d338865288976d58cc1ae4ab9ff40d79a3e2.jpg", + "image_caption": [ + "Ours-stage I" + ], + "image_footnote": [], + "bbox": [ + 93, + 337, + 274, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7320d55c71ddaab53b4bbc6ceb1939e3dd086d50357da90a8152ae7a1df51e33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 431, + 130, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2886d4a1d6d4777176cda51e17973d40068c6e830cd0b441092666ed009a261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 434, + 163, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56ec4291011bcb79a4ca3b1f902221cc986364ec40153fea96aa4141f92991f8.jpg", + "image_caption": [ + "Pourreza et al." + ], + "image_footnote": [], + "bbox": [ + 94, + 459, + 274, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bdae9d0326c6336d7e0abab175dad8fcd1c359cc7ae8d0a033b13992973278a.jpg", + "image_caption": [ + "Figure 1. AdaptiveAE takes camera preview images as input and automatically predicts the ISO and shutter speed for each LDR captures for exposure fusion through a 3-stage sequential refinement procedure to achieve an optimal balance between noise level and motion-related problems for high quality HDR capturing in dynamic scenes with deep reinforcement learning. AdaptiveAE achieves PSNR 39.7 on HDR dataset [26], while baseline methods [6, 21, 32] that either only predicts shutter speed or do not consider motion can only achieve PSNR below 37.6 and has evident motion blur and ghosting artifacts in HDR results." + ], + "image_footnote": [], + "bbox": [ + 102, + 551, + 132, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9bc6a12b1ee7ce7ba9b63901fca253196c0b7345bc72b85851d39b5ea9e4ad93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 135, + 551, + 169, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24ac4af5b61e35b75f58a8413d4b3b3ac2b009d82fca3a44ae6ab797abf83485.jpg", + "image_caption": [ + "Ours-stage II" + ], + "image_footnote": [], + "bbox": [ + 307, + 338, + 488, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9ca708596d774d0fee0494d5a772d970d6737b486135e1e9a5f1953f0623985.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 433, + 346, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f1e9ded7fc2e2f587f1ac5a0029376077738f1913e8a3787b025b3a8217bcd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 434, + 385, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5786f0da83e045f29d5458f464df7b65a4a6475864afedcd792a41098065600.jpg", + "image_caption": [ + "Wang et al." + ], + "image_footnote": [], + "bbox": [ + 307, + 459, + 488, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea6d2f0a457b0594deb06d35feea7700f2991fd802355eb09cd642cde415849e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 551, + 346, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eeb940db43cbc39522bc5f7706bedbbd72fc8ba16daef7f505747fa6233c53a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 553, + 385, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1683a3d83a049a09236d36091e3e3df2a9ba1a3e4cfe032df20d396250c7954a.jpg", + "image_caption": [ + "Ours-stage III" + ], + "image_footnote": [], + "bbox": [ + 521, + 338, + 702, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/667677b13404260285ccecab2d683a89ff51a2737f805fdbb8f411bcd326418a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 433, + 560, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57fd6fba537d893b39e3c90ba0fa00d528f88ac04e54e2afc7fc81f0ce5bf920.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 435, + 599, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20b2781409bd4a9565ee62119418fffe646f4b2d365a478af1fffae848ef2e8c.jpg", + "image_caption": [ + "Hasinoff et al." + ], + "image_footnote": [], + "bbox": [ + 521, + 459, + 702, + 551 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d11aebd9323ae48813cbc7cd114dc47a85b2ac92049c785ae2782ddba9d400b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 551, + 560, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb291a2e952040657bcb9d209cbe247a66a0846af8700c5f0af7d498013b017c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 566, + 553, + 599, + 575 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/424effc145ab0444c1bb6a282fe56b3522b5122336f6055248b0ad4fc7ef21ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 340, + 903, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f2c83c59ef5ff751ce19bf48dad459f0ce71b3086d490ddaebc6d9e9ace42c8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 483, + 754, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a37337102f97533ed0e0cfa3318249836f2fcb80b9c7c84e6a5bb43f7ea5bd3.jpg", + "image_caption": [ + "Deep Learning Based" + ], + "image_footnote": [], + "bbox": [ + 759, + 483, + 787, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c4db17d59b56dd50e25d451caaaf17be993c3382138924ca0dc9b09542a434c.jpg", + "image_caption": [ + "Shutter Speed" + ], + "image_footnote": [], + "bbox": [ + 723, + 515, + 751, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa947fc2e4b90cd3dd695852f772b1a5e764057e22e59b5a82bea74be07c6ae4.jpg", + "image_caption": [ + "ISO" + ], + "image_footnote": [], + "bbox": [ + 844, + 515, + 872, + 536 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e266a634297ed36372e3b0bc8a42db9165bae7d04c60382110604934bb274d38.jpg", + "image_caption": [ + "Adjustable / Unadjustable" + ], + "image_footnote": [], + "bbox": [ + 723, + 546, + 756, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 670, + 326, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mainstream high dynamic range imaging techniques typically rely on fusing multiple images captured with different exposure setups (shutter speed and ISO). A good balance between shutter speed and ISO is crucial for achieving high-quality HDR, as high ISO values introduce significant noise, while long shutter speeds can lead to noticeable motion blur. However, existing methods often overlook the complex interaction between shutter speed and ISO and fail to account for motion blur effects in dynamic scenes.", + "bbox": [ + 89, + 702, + 482, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose AdaptiveAE, a reinforcement", + "bbox": [ + 109, + 838, + 482, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning-based method that optimizes the selection of shutter speed and ISO combinations to maximize HDR reconstruction quality in dynamic environments. AdaptiveAE integrates an image synthesis pipeline that incorporates motion blur and noise simulation into our training procedure, leveraging semantic information and exposure histograms. It can adaptively select optimal ISO and shutter speed sequences based on a user-defined exposure time budget, and find a better exposure schedule than traditional solutions. Experimental results across multiple datasets demonstrate that it achieves the state-of-the-art performance.", + "bbox": [ + 511, + 671, + 906, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 845, + 645, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-dynamic-range (HDR) imaging plays a pivotal role in computational photography. To capture an HDR scene,", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during Tianyi Xu's internship at Shanghai AI Laboratory.", + "bbox": [ + 89, + 862, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 116, + 887, + 246, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25176", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "due to hardware limits, a single capture can only cover a Low Dynamic Range (LDR), and HDR fusion techniques are proposed to combine multiple LDR images with varying exposures to cover a wide dynamic range [7, 10, 14, 15, 17, 19, 23, 35, 36, 40]. The typical way to vary exposure is to change either shutter speed or ISO. This is a challenging process, as a longer shutter speed may increase the signal-to-noise ratio but introduce unpleasant motion blur, a large ISO may increase brightness but also magnify noise, and significant exposure differences can cover a wider dynamic range but also increase the risk of misalignment. Therefore, the choice of exposure values (EVs) for each capture is critical in this process to ensure high-quality HDR results.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Still, very limited work discusses how to choose the optimal exposure levels, particularly in dynamic scenes. Prior works on exposure scheduling mainly focus on static scenes and ignores potential motion blur. Learning-based techniques overlook the intricate interplay between ISO and shutter speed, resulting in suboptimal image quality under varying conditions [6, 21, 32]. Additionally, many existing methods treat ghosting and motion blur as separate, computationally intensive post-processing tasks, which is unsuitable for real-time applications [11, 25, 27, 29, 31, 38, 39].", + "bbox": [ + 89, + 289, + 480, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose AdaptiveAE, an efficient exposure control algorithm designed for HDR capturing, which addresses both motion blur and noise during image acquisition. Given the previously captured image, our method optimizes the exposure bracketing strategy for the subsequent capture, based on illumination information and semantic data from previous frames. Unlike previous approaches that treat motion blur as a separate post-processing task, we aim to address it during the capture process.", + "bbox": [ + 89, + 443, + 480, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Designing both efficient and adaptive exposure control is non-trivial, and we resort to reinforcement learning [18] to solve this challenge. Our solution mimics an experienced photographer. At each iteration, the policy network takes the previously captured LDR images as input, together with extracted semantic and illumination information. Given this information, the policy network learns to determine the optimal exposure setup for the following captures, which maximizes the additional information provided by this frame while also reducing the risk of misalignment and motion blur in a dynamic scene. Once the next frame is captured, this newly captured image will be used as input for the subsequent refinement iteration. As shown in Fig. 1 right, the final quality (PSNR) of fusion increases as more images are captured.", + "bbox": [ + 89, + 580, + 480, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed approach offers several advantages over traditional exposure controls. First, AdaptiveAE controls both exposure time and ISO and also adapts to different scenes. As a result, it has a much higher upper bound compared to either a fixed exposure schedule or an adaptive control algorithm that only changes exposure time. As shown", + "bbox": [ + 89, + 810, + 480, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Fig. 1, our method iteratively achieves an optimal balance between noise and blur, resulting in less noise, reduced motion blur, and superior image quality compared to other baseline methods. Second, our method can handle both static and dynamic scenes. In static scenes, it achieves performance comparable to state-of-the-art techniques, and in dynamic ones, it produces visually compelling HDR images with minimal motion blur and ghosting artifacts. Third, our method can automatically choose the best number of frames for HDR imaging. Unlike traditional HDR approaches that often use a fixed number of frames, such as three, our approach provides flexibility to determine whether capturing three or more frames is optimal for certain scenes, balancing image quality and time budget.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach enables the concurrent analysis of blur and noise in AE prediction, thereby enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately.", + "bbox": [ + 511, + 304, + 903, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Additionally, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and underscore the critical importance of incorporating blur synthesis into our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II, demonstrates superior noise control and effectively reduces motion blur, outperforming baselines in the visual quality of the fused HDR images.", + "bbox": [ + 511, + 428, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach allows for concurrent analysis of blur and noise in AE prediction, enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately.", + "bbox": [ + 511, + 626, + 903, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Also, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and highlight the critical importance of incorporating blur synthesis within our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II,", + "bbox": [ + 511, + 750, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25177", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "demonstrates superior noise control and effectively reduces motion blur, outperforming compared baselines in the visual quality of the fused HDR images.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 143, + 228, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Strategy for exposure bracketing. Determining the optimal set of exposures for multiple-exposure dynamic range imaging is a well-established problem. Most digital cameras allow users to set the compensation ratio for exposure bracketing, while mobile cameras typically impose fixed ratios during automatic exposure bracketing. Heuristic strategies based on the histogram are proposed by [5, 21, 30], to balance single-to-noise ratio (SNR) and saturation. The work by [6] firstly formulates this challenge as a constrained optimization problem in the linear RGB domain, addressing the scenario that follows multiple exposure fusion and precedes tone mapping and denoising. This concise formulation facilitates the use of a straightforward numerical solver. Further extensions of this formulation consider the alignment of multiple input images to account for handshake [24]. In the context of structured light 3D reconstruction, various exposures are likewise treated as an optimization problem [4]. Exposure influences not only the noise but also, to some extent, the tone of the image in the standard RGB domain. Evaluating the final image quality complicates the problem further, as subsequent tone mapping or retouching can significantly alter image quality. Therefore, a neural network is utilized to estimate exposures and fuse multiple exposed images to achieve optimal fidelity in the gamma-corrected domain [12]. Additionally, reinforcement learning is employed by [37] and [32] to assess the rewards on comprehensive image quality after more sophisticated tone adjustment.", + "bbox": [ + 89, + 167, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenge in dynamic scenes. In static scenes, increasing under-exposures can reduce saturation, and a longer shutter speed improves signal-to-noise ratio in dark areas. However, in dynamic scenes, excessive exposure may cause ghosting artifacts, and a prolonged shutter speed can lead to motion blur. Consequently, heuristic exposure bracketing is usually limited to two or three EV settings [5, 30]. The method by [6] imposes an upper limit on total shutter speed, while [24] addresses handshake motion through image registration but scarcely tackles local object motion. Most approaches do not fully address motion in dynamic scenes, leaving motion blur and ghosting artifacts to be addressed through post-processing.", + "bbox": [ + 89, + 590, + 483, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 800, + 181, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conventional weighted linear combination methods, such as exposure fusion, provide straightforward SNR estimation but fail in dynamic scenes where moving objects create unquantifiable motion blur and ghosting artifacts due to misalignment. Our approach addresses these challenges dur", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/421b9851af850fb7339a2f6eb81a66eed6ccda7a3b310a3c64c6411c95561657.jpg", + "image_caption": [ + "(a) Our overall image synthesis pipeline." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 174 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d5227d29cef88f714dd43c985a785ac119f2bd0cbd7814b458e7ea00f089ca0.jpg", + "image_caption": [ + "(b) Our blur synthesis pipeline." + ], + "image_footnote": [], + "bbox": [ + 516, + 191, + 903, + 275 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a616eb802b232413d45415954c7e9bc8ec5cc90dc0eb01e5ea4715d8a2852c0c.jpg", + "image_caption": [ + "(c) Our noise synthesis pipeline.", + "Figure 2. Our blur-aware data synthesis pipeline." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 903, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing capturing rather than in post-processing, as the latter is shown to yield suboptimal results. We predict exposure-related risks—such as motion blur, ghosting, noise, and saturation—based on a limited number of previously captured frames. Additionally, we employ a sequential strategy for exposure and ISO parameter determination, rather than simultaneously predicting settings for all three LDR images, which reflects the iterative nature of auto-exposure in mobile cameras that enables adaptation to significant brightness transitions.", + "bbox": [ + 511, + 444, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.Blur-aware data synthesis pipeline", + "text_level": 1, + "bbox": [ + 511, + 604, + 815, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To simulate capturing in real environments, we designed an image synthesis pipeline to generate realistic motion blur and noise in LDR images from HDR videos in the training dataset, for use in training. Typical exposure settings involve adjusting the exposure value (EV), which is calculated as:", + "bbox": [ + 511, + 626, + 905, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E V} = \\log_ {2} \\left(\\frac {F ^ {2}}{T} \\times \\frac {1 0 0}{\\mathrm {I S O}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 712, + 903, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F$ denotes the aperture's f-number, ISO represents the ISO sensitivity and $T$ is the exposure time in seconds.", + "bbox": [ + 511, + 750, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to recent methods [6, 21, 32] concerning high dynamic range capture, we assume that aperture and focus are held constant to prevent changes in defocus. This leaves just two camera settings to manipulate: (1) Shutter speed, which controls the amount of light to collect, and (2) ISO, which determines the sensor gain.", + "bbox": [ + 511, + 780, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our pipeline synthesizes motion blur and noise for the ground truth static HDR image based on a specified ISO", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25178", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9900aad84df9c2809d531a94055d7112d27cbb6801c39951e924b0b80d54dc8c.jpg", + "image_caption": [ + "Figure 3. The training scheme of AdaptiveAE. States are defined as the three LDRs synthesized using predicted ISOs and shutter speeds. Starting from $s_0$ where the three LDRs has EV $\\{-2,0, + 2\\}$ with arbitrary EV 0 baseline, ISOs and shutter speeds, the agent sequentially predicts, customizes or inherits capturing parameters (i.e. ISO and shutter speed) for the next stage and synthesize the corresponding LDR using our image synthesis pipeline. Unlike training, the LDRs will be captured rather than synthesized during inference." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 898, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and shutter speed. As depicted in Fig. 2, it uses two consecutive static HDR images as input. Motion blur is first synthesized according to shutter speed to produce a blurred HDR image in linear space. Next, noise is added based on shutter speed and ISO to create an LDR image, reflecting our exposure choice. Note that motion blur should be applied before adding noise, as it influences the raw input by affecting the number and pattern of captured photons during photography.", + "bbox": [ + 88, + 367, + 480, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthesizing blur. On-the-shelf training datasets consist of consecutive HDR ground truth of a scene with motion. Now we explain how we simulate motion blur to a frame of HDR $f_{i}^{L}$ , as shown in Fig. 2b, which is the $i$ -th HDR ground truth frame in the dataset scene, and the superscript $L$ indicates it is in linear space. We first use $\\mu$ -law tone-mapping with $\\mu = 5000$ to transfer $f_{i}^{L}$ and $f_{i+1}^{L}$ from HDR space to LDR space where the image interpolation algorithm we applied is trained upon, receiving $f_{i}^{\\mathcal{T}}$ and $f_{i+1}^{\\mathcal{T}}$ , where the superscript $\\mathcal{T}$ denotes they are in LDR space. Then we use RIFE [7] to interpolate the them to 256 frames and get the sequence of images $\\{f_{i}^{\\mathcal{T}}, s_{1}^{\\mathcal{T}}, s_{2}^{\\mathcal{T}}, \\dots, s_{254}^{\\mathcal{T}}, f_{i+1}^{\\mathcal{T}}\\}$ . Then for the selected shutter speed $T_{j}$ for the $j$ -th LDR $l_{j}^{\\mathcal{T}}$ to take, the blurred HDR $b_{j}^{L}$ is simulated as:", + "bbox": [ + 89, + 503, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb _ {j} ^ {L} = \\mathbf {i T M O} \\left(\\frac {f _ {i} ^ {\\mathcal {T}} + \\sum_ {m = 1} ^ {m _ {j}} s _ {m} ^ {\\mathcal {T}}}{m _ {j}}\\right), m _ {j} = \\left\\lceil \\frac {2 5 6 T _ {j}}{\\Delta \\tau} \\right\\rceil , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 729, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta \\tau$ denotes the time elapsed from $f_{i}^{L}$ is taken to $f_{i + 1}^{L}$ is not yet taken and iTMO indicates the inverse $\\mu$ -law tonemapping function with $\\mu = 5000$ .", + "bbox": [ + 89, + 777, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthesizing noise. We adopt the noise model mentioned in [6], in which noise is modeled as a zero-mean variable, coming from three independent sources, including photon noise, which represents the Poisson distribution of photon arrivals and depends linearly upon the number of recorded", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "electrons, $\\Phi T$ , readout noise, which comes from sensor readout, and analog-to-digital conversion(ADC) noise, which comes from the combined effect of amplifier and quantization. Hence, for pixels below the saturation level:", + "bbox": [ + 511, + 367, + 906, + 428 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV a r (n) = \\frac {\\Phi T \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\frac {\\sigma_ {\\text {r e a d}} ^ {2} \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\sigma_ {\\mathrm {A D C}} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 436, + 906, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Phi$ is the radiance level, $T$ is the shutter speed, $U$ is a camera-dependent variable.", + "bbox": [ + 511, + 474, + 905, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following our noise model, as shown in Fig. 2c, we can synthesize the corresponding noise with selected ISO and shutter speed to the blurred HDR $b_{j}^{L}$ to get the LDR image $l_{j}^{\\mathcal{T}}$ . For details, please refer to our supplementary material.", + "bbox": [ + 511, + 506, + 905, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Problem formulation of AdaptiveAE", + "text_level": 1, + "bbox": [ + 511, + 575, + 830, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a scene, the goal of AdaptiveAE, which has access to three initial preview LDR images $\\{p_j^T\\}_{1,2,3}$ (i.e. underexposed, mid-exposed, overexposed) before capturing, is to find an optimal exposure setup (i.e. ISO and shutter speed) for LDR capturing, with which the fused HDR output will result in pleasing visual performance.", + "bbox": [ + 511, + 598, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method formulates exposure bracketing as a Markov Decision Process [22], solved via deep reinforcement learning to refine exposure parameters (ISO, shutter speed) sequentially. As illustrated in Fig. 4, the process starts from three LDRs at a default $\\{-2_p,0_p, + 2_p\\}$ EV spacing relative to an arbitrary reference (subscript $p$ ). The refinement then proceeds in stages, as shown in Fig. 3.", + "bbox": [ + 511, + 689, + 905, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, the agent predicts optimal parameters for the midexposed frame, establishing a new 0-EV reference (subscript $x$ ). The side frames are then customized—their parameters are procedurally set to achieve a symmetric $\\{-2_x,0_x, + 2_x\\}$ EV bracket. Next, the agent refines the underexposed frame to an EV of $-y_{x}$ . The mid-exposed frame $(0_{x})$ is inherited (its parameters are reused), while", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25179", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b2f7eae99e8282802337d2f8763386e293aecbd6a5f4edae5c134f6a55f4a1e8.jpg", + "image_caption": [ + "Figure 4. Training pipeline of our method. The ISO and shutter speed prediction process is conceptualized as a Markov Decision Process, where a CNN-based policy network predicts the ISO and shutter speed of the next exposure sets. Concurrently, a CNN-based value network estimates the state value. We leverage our blur-aware image synthesis pipeline to synthesize the predicted LDRs and employ DeepHDR [7] to fuse the predicted LDR images, generating our HDR result and calculating the reward for the current policy. The entire system is optimized using the A3C (Asynchronous Advantage Actor-Critic) method [18]." + ], + "image_footnote": [], + "bbox": [ + 130, + 85, + 861, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the overexposed frame is customized to $+y_{x}$ to maintain symmetry, yielding $\\{-y_{x}, 0_{x}, + y_{x}\\}$ . Finally, the agent predicts the overexposed frame's EV as $+z_{x}$ , creating a potentially asymmetric set $\\{-y_{x}, 0_{x}, + z_{x}\\}$ . This sequential prediction can be extended, as our adopted fusion method [7] handles more than three LDRs. A video providing further details on this process, along with an example of extended exposure bracketing, is available in the supplementary materials.", + "bbox": [ + 88, + 332, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimization objectives", + "text_level": 1, + "bbox": [ + 89, + 477, + 308, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let us denote the problem as $P = (S, A)$ , where S is a state space and A is an action space. Specifically, in our task, $S$ is the space of the exposure setups (i.e., ISO and shutter speed) for an LDR set that typically contains three LDRs (i.e. under-exposed, mid-exposed, and overexposed), while $A$ is the set of all possible ISO and shutter speed combinations, which is discrete. During training, at stage $s_j = \\{(\\mathrm{ISO}_{j1}, T_{j1}), (\\mathrm{ISO}_{j2}, T_{j2}), (\\mathrm{ISO}_{j3}, T_{j3}), \\}$ , we first find the corresponding HDR ground truth image pairs $(f_i^L, f_{i+1}^L)_{1,2,3}$ and generate the corresponding LDRs $\\{l_j^T\\}_{1,2,3}$ through our image synthesis pipeline in Fig. 2. The footnote 1, 2, 3 denotes that the same operation is done for the under-, mid-, and over-exposed LDRs. Taking $\\{l_j^T\\}_{1,2,3}$ as input, the agent predicts an action $a_j = (\\mathrm{ISO}_j, T_j)$ , which is expanded to an exposure setup for three LDRs by customizing EV or inheriting from state $s_{j-1}$ , mapping state $s_j$ to state $s_{j+1}$ . Adding a sequence of $M$ LDRs to exposure bracketing corresponds to a trajectory $\\tau$ of states and actions:", + "bbox": [ + 91, + 500, + 483, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\left(s _ {0}, a _ {0}, \\dots , s _ {M - 1}, a _ {M - 1}, s _ {M}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 799, + 480, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_M$ is the stopping state. Our goal is to find a policy that maximizes the accumulated reward during the decision-making process. In this paper, the reward function with the $j$ -th action (i.e., corresponding to deciding the exposure setting for the $j$ -th LDR for exposure fusion) is thus written as:", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr \\left(s _ {j}, a _ {j}\\right) = \\mathcal {R} \\left(s _ {j + 1}\\right) - \\mathcal {R} \\left(s _ {j}\\right) - \\mathcal {P} (j), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 348, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{j + 1} = p(s_j,a_j)$ , and $\\mathcal{R}$ denotes our reward design and $\\mathcal{P}$ denotes the $L_{\\mathrm{step}}$ penalty, detailed in 3.4.", + "bbox": [ + 511, + 372, + 903, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As depicted in Fig. 4, our model comprises a policy network and a value network, both of which utilize a CNN-based architecture. The policy network predicts the optimal ISO and shutter speed for the subsequent exposure, outputting a distribution of action probabilities $\\pi(s,\\theta)$ for an input image $s$ . Concurrently, the value network $V^{\\pi}(s,\\omega)$ estimates the corresponding state value. These networks, with combined parameters $\\psi = (\\theta,\\omega)$ , are trained by maximizing our objective $J(\\theta)_{\\psi}$ to learn the optimal policy $\\pi(s)$ . Specifically, to train the policy network and the value network, we apply the A3C (Asynchronous Advantage Actor-Critic) method [18], where the actor is represented by the policy network and the critic is the value network. Network details are in the supplementary materials.", + "bbox": [ + 511, + 402, + 906, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Reward", + "text_level": 1, + "bbox": [ + 511, + 625, + 611, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When designing the reward function for our system, we considered four key factors: (1) similarity between the fused HDR and the ground truth HDR; (2) quality of important regions in the fused HDR; (3) quality of moving regions in the fused HDR; and (4) a penalty for overly long LDR stacks. Thus, our reward function is:", + "bbox": [ + 511, + 647, + 905, + 737 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = - \\left(P _ {\\text {c o n s t r u c t i o n}} + P _ {\\text {p r i o r i t y}} + P _ {\\text {g h o s t}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 750, + 903, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider $P_{\\text {construction }}$ , the L2 loss between our fused HDR and the ground truth, as our major reward component, which is affected by noise and saturation. Note that through the entire sequential decision process, the middle-exposed frame is used as a reference for HDR fusion. Conforming to Eq. (3), during training, noise is synthesized according to the irradiance of ground truth HDR, and during inference, noise is estimated with the irradiance of noisy signals.", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25180", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$P_{\\mathrm{priority}}$ represents an L2 loss within the areas in the image masked by an importance mask, which is generated by a saliency predictor [20]. This ensures that the highest quality is maintained in the most significant areas, thereby enhancing the overall visual fidelity where it matters most.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$P_{\\mathrm{ghost}}$ is also an L2 loss within a masked area, denoting areas with large motions and thus having a higher risk of motion blur- or ghosting-caused HDR quality degradation. The mask is computed by calculating the optical flow using RAFT [28] between the HDR ground truth of the middle-exposed frame (i.e., the reference frame for fusion) and the corresponding HDR $f_{i}^{L}$ and selecting the pixels where the mode of the flow vector exceeds a constant threshold $K$ . Normalizing the largest optical flow vector, we empirically set $K$ to 0.2. $P_{\\mathrm{ghost}}$ guides the agent to deal more carefully with regions that are prone to artifacts caused by motion and is helpful for high-quality HDR capturing, as is verified by the result of our ablation studies in Tab. 2.", + "bbox": [ + 88, + 167, + 482, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{P}(j)$ is a penalty designed to penalize excessively long exposure brackets. Capturing an excessive number of shots, such as 10, even with random exposure settings, can lead to nearly perfect HDR fusion results, but it is time-consuming. Typically, three shots [2, 6, 21] are sufficient to achieve high-quality outcomes. To this end, we incorporate a penalty for taking more than three shots, as follows:", + "bbox": [ + 89, + 362, + 482, + 468 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} (j) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} j \\leq H \\\\ \\alpha (j - H) ^ {2} & \\text {i f} j > H \\end{array} , \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 479, + 480, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha$ is a positive coefficient and $H$ is set to 3.", + "bbox": [ + 89, + 529, + 419, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this manner, the autonomous agent optimizes exposure parameters by predicting relatively fast shutter speeds for LDR images, particularly for the middle-exposed reference frame, thereby minimizing motion blur while avoiding excessive ISO values that would introduce noise-related degradation. When confronted with potential ghosting artifacts—which emerge from information deficiency in LDR images due to concurrent saturation and motion—the agent adaptively selects EVs that minimize both underexposure and saturation, resulting in a significant reduction of ghosting artifacts in the final reconstruction.", + "bbox": [ + 88, + 545, + 482, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 724, + 220, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment details. We use Real-HDRV [26] for training and tested our performance on Real-HDRV [26] and DeepHDRVideo [3]. For HDR fusion, we adopt DeepHDR [33] to generate the HDR image based on the selected exposure bracketing. Additionally, we set the number of LDR frames to 3 for all methods involved. For methods that do not account for changes in ISO values, we set the ISO to 200 as a standard value for most cases, which is rationalized in our supplementary materials. Since RIFE [7] is relatively time-consuming, the image interpolation and blur synthesis step", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is performed before training. We apply random flipping, rotation, and cropping with $512 \\times 512$ pixels for data augmentation. Our training dataset consists of a total of 770 scenes, including 440 dynamic and 330 static scenes.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We directly evaluate the performance of the fused HDR results. Similar to previous HDR fusion methods [7, 14, 19], we employ PSNR- $\\mu$ , SSIM- $\\mu$ , PU-PSNR, PU-SSIM, and HDR-VDP-2 [16] as evaluation metrics. PSNR- $\\mu$ and SSIM- $\\mu$ denote PSNR and SSIM of the fused HDR after $\\mu$ -law tone-mapping with $\\mu = 5000$ . PU-PSNR and PU-SSIM are computed after perceptually uniform encoding [1]. When computing the HDR-VDP-2 [16], the diagonal display size is 30 inches.", + "bbox": [ + 511, + 152, + 906, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 511, + 301, + 607, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Real-HDRV dataset. We compared our trained agent's performance with several state-of-the-art HDR exposure bracketing methods, including Pourreza-Shahri et al. [21], Hasinoff et al. [6], and Wang et al. [32]. The first two are non-deep-learning methods that do not consider motion: Pourreza-Shahri et al. use K-means clustering to adjust shutter speed based on image brightness, while Hasinoff et al. mathematically optimize ISO and shutter speed for the best worst-case SNR. Wang et al. utilize reinforcement learning to predict shutter speed for maximizing PSNR, but also ignore motion.", + "bbox": [ + 511, + 324, + 906, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method achieves state-of-the-art performance on the HDRV-Test dataset, as shown in Tab. 1. By treating ISO as a variable, it provides flexibility in handling extremely dark scenes. Additionally, it excels in dynamic scenes due to a blur synthesis model and carefully designed rewards. Thus, our model effectively balances noise reduction and motion artifact minimization, delivering high-quality HDR results. Visualization results in Fig. 5 show that deep learning-based exposure fusion models affect EV selection differently. In Hasinoff et al.'s setup, three LDRs are equally weighted, but dynamic scenes require a reference image, emphasizing the middle-exposed LDR's quality. This often shifts its EV toward under-exposure to reduce motion blur. Our experiments confirm that if the middle-exposed image is blurry—a common problem in other methods—the fused HDR will also be blurred.", + "bbox": [ + 511, + 491, + 905, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference time. Our method takes less than $250\\mathrm{ms}$ for each HDR image, which is acceptable. With average exposure time $n$ ( $\\leq 30ms$ ) and prediction time $m$ ( $\\leq 10ms$ ), the total execution time is $6n + 3m$ ( $\\leq 250ms$ ). This can be further lowered to $6n$ ( $\\leq 200ms$ ) adopting concurrency. Note that DeepHDR is only required during training for calculating rewards, and on-camera inference can be performed with a small agent, taking only $3.5ms$ for each frame without optimization. Besides, our method can be further accelerated by on-camera processors when applied to DSLRs. More discussions are in our supplementary materials.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25181", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d3de9365d8e7bccc28e520fe57a679c3167bff0481545594d08799c11e01c78c.jpg", + "image_caption": [ + "LDR Images", + "Figure 5. Qualitative comparisons with other auto-exposure methods on HDRV dataset [26]. Left: Predicted LDRs with varying ISO and shutter speed settings and synthesized using our image synthesis pipeline. Middle: Fused HDR image using DeepHDR [7] and tonemapined using Photomatix. Right: Zoom-in results for tested methods." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 187, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff5409d764a92d23a08ab76c626a512db868d36c4263fde7651fe6f9fedffb98.jpg", + "image_caption": [ + "Our Tonemapped HDR Image" + ], + "image_footnote": [], + "bbox": [ + 187, + 90, + 464, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/489655b8709e24eaf98b598deef205f42c532ce5f3c062feee6528dfc86214e7.jpg", + "image_caption": [ + "Wang et al. Pourreza et al. Hasinoff et al.", + "Ours", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 464, + 90, + 910, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4216b509773b183ce3fa1ad6f25ccb14aa0826772344021ca2a071b6f7bf962d.jpg", + "table_caption": [ + "Table 1. Comparison of Different Methods. We utilized one preview image and three preview images, respectively, for the compared methods and our method to predict exposure settings for three LDRs. We leveraged DeepHDR [7] for exposure fusion and used the mentioned metrics to evaluate the quality of the fused HDR. Bold: The best." + ], + "table_footnote": [], + "table_body": "
MethodsHDRV [26]DeepHDRVideo [3]
PSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIMPSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIM
Pourreza et al. [21]33.640.861754.5530.610.867935.570.878055.6731.590.8791
Hasinoff et al. [6]37.590.905257.0232.870.898038.470.915758.6534.450.9132
Wang et al. [32]36.460.890256.0932.680.893337.950.901957.3933.270.9008
Ours39.700.940859.2034.670.946539.810.937158.9036.190.9338
", + "bbox": [ + 91, + 512, + 903, + 606 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/038b226dd5a15e9b15194a0e866e3525d6f4ac371205fcad5b44942eb116127e.jpg", + "table_caption": [ + "Table 2. Ablation study of AdaptiveAE on HDRV [26] dataset. Base denotes our model trained with only the step penalty and construction reward. Bold: The best." + ], + "table_footnote": [], + "table_body": "
ModelPSNR-μSSIM-μPU-PSNRPU-SSIM
Base38.210.922732.680.9198
Base+Ppriority38.570.926133.020.9239
Base+Ppriority+Pghost39.700.940834.670.9465
", + "bbox": [ + 106, + 664, + 460, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gap to the best-achievable. For each scene in the test set of Real-HDRV, we iteratively search for the best set of predictions by Gaussian sampling around our initial prediction (50 times per exposure parameter per frame, with a deviation of $20\\%$ of the mean for both ISO and shutter speed). Statistics in Tab. 4 show that our method approaches the locally optimal result while being efficient.", + "bbox": [ + 89, + 733, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross datasets test. To test the generalization ability of our model, we also evaluated the performance of our trained agent on DeepHDRVideo [3], as shown in the right of Tab. 1. Our agent exhibits good generalization abilities.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f262e18bc78981af742204067dd8dd8386c5be96d524ceffc8e32ec1d81a8bf1.jpg", + "table_caption": [ + "Table 3. Comparison of different exposure fusion methods. We utilize the four auto-exposure methods to predict exposure settings and employ our image synthesis pipeline to create three LDRs. Then, we apply different exposure fusion methods to fuse them and compare the HDR quality on the HDRV dataset [26]. HDRTransformer [14] is pretrained on HDRV [26] dataset. P: PSNR- $\\mu$ S: SSIM- $\\mu$ . Pou: Pourreza et al. [21], Has: Hasinoff et al. [6], W: Wang et al. [32]. HDR-Trans: HDR-Transformer. Bold: The best." + ], + "table_footnote": [], + "table_body": "
ModelDeepHDRHDR-GANHDR-Trans
P-μS-μP-μS-μP-μS-μ
Pou33.640.861735.710.889235.840.8824
Has37.590.925238.580.926339.110.9372
W36.460.900237.950.916938.890.9210
Ours39.700.940840.730.937641.370.9478
", + "bbox": [ + 514, + 733, + 913, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the DeepHDRVideo dataset only provides ground truth HDR for the middle image in the sequence, we synthesize HDR for the other images using DeepHDR [33].", + "bbox": [ + 511, + 838, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross HDR fusion methods test. In our training scheme,", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25182", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abbd4ca540e4e79a1b0ae6ff6636e6682c1523b10f16d76cec4be8557c7d5e6b.jpg", + "table_caption": [ + "Table 4. Performance comparison investigating the gap between our method and the optimum on the HDRV-test dataset." + ], + "table_footnote": [], + "table_body": "
OursWorstAverageBest
PSNR-μ39.7025.7632.4139.93
SSIM-μ0.94080.77380.86090.9412
", + "bbox": [ + 114, + 127, + 460, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e2a8baf7457b8550cb3c40fa02c2876435cb8eb97581dfd60240a5bc4e405e5.jpg", + "image_caption": [ + "Wang et al.", + "Figure 6. Results of our real capture data. The subject performs steady and repetitive movements, taking shots according to the exposure settings predicted by various methods." + ], + "image_footnote": [], + "bbox": [ + 98, + 200, + 223, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88ea1c1e3281f4498568455941fb3e88f7022e81d5092aef7f902a778a642e35.jpg", + "image_caption": [ + "Hasinoff et al." + ], + "image_footnote": [], + "bbox": [ + 223, + 200, + 346, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e3290c7c7e4426deb4314c3ff183fdb66d0c9015fceacff00c9a0891c2ed8bf.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 346, + 200, + 470, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we used DeepHDR for exposure fusion. For further testing, we adopted different exposure methods, including HDR-GAN [19] and HDR-Transformer [14] (pre-trained on the HDRV-dataset [26]), as post-processing exposure fusion methods. All tests are conducted on the HDRV-Test dataset. As shown in Tab. 3, without considering motion before LDR capturing, however powerful an exposure fusion method (i.e. HDR-GAN and HDR-Transformer) fails to yield a satisfactory result. Notably, when stronger fusion models are used, the performance gap between our model and traditional models, which do not account for motion, increases. This is because, when blur and ghosting risk are mitigated, post-processing models can effectively manage the remaining challenges. In contrast, ignoring motion makes these challenges more difficult for post-processing models, leading to potential failure.", + "bbox": [ + 88, + 412, + 482, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results for real capture. We used a SONY Alpha 7C-II to evaluate our model on real-world scenes. Subjects performed steady, repetitive movements while ISO and shutter speed were manually set for each capture, with the aperture fixed at f/2.8. The camera and subject positions remained unchanged during multi-exposure captures. As shown in Fig. 6, our method provides noise control comparable to other baselines while effectively mitigating motion blur, which can impair the visual quality of the fused HDR image; for more results, see our supplementary materials.", + "bbox": [ + 89, + 655, + 482, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 816, + 251, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of penalty item. We conduct ablation studies to validate the effectiveness of our reward design. All quantitative evaluations are conducted on the HDRV-Test dataset. We train our networks using only the construction", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c6298add16b1a6e25f9232f8c0a6775c18860039f33cab00fc887f63509e7e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 589, + 135 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8476f314e935179366fe64aeb1230113e43f4749748421b5d6db7187ddb9b9eb.jpg", + "image_caption": [ + "Under-exposed" + ], + "image_footnote": [], + "bbox": [ + 517, + 135, + 589, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c97e927027381d99e123cb7e88b6ebad07e02e3b20507659125b7546a55a5eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 90, + 665, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cadbdc4c20577aae27ce1dd7cc0d1f84670f00b2d51c33269d4fb84ac7320e0d.jpg", + "image_caption": [ + "Middle-exposed" + ], + "image_footnote": [], + "bbox": [ + 591, + 133, + 665, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/388664262c000dcf514fbaa58a3a7bf53d35ef51626fe922f26614cd52a12fad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 90, + 738, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4f9563ac7c766b35eb5e362f1d6561a514e6b3010ca5966c0c2950c0da47e24.jpg", + "image_caption": [ + "Over-exposed" + ], + "image_footnote": [], + "bbox": [ + 666, + 133, + 738, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc7c33839d6b7b2edd687c4a7c39c108f56aa8aa86578565109d16e9bf7ba057.jpg", + "image_caption": [ + "Figure 7. Left: Zoom-in details of the moving regions and our predicted ISO and shutter speed (seconds) for each LDR. Up: LDRs for scenes with an average motion level of around 30 pixels, Down: around 60 pixels. Right: Comparisons with baseline methods across different motion magnitude ranges. Tested on Real-HDRV [26] dataset." + ], + "image_footnote": [], + "bbox": [ + 741, + 90, + 910, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reward and step penalty as the base model and test the effectiveness of $P_{\\mathrm{ghost}}$ and $P_{\\mathrm{priority}}$ . As shown in Tab. 2, these two reward elements enhance our agent's performance by directing the model to focus on semantically important regions and moving areas. The former typically corresponds to the scene's focal point (e.g. a person's face), while the latter involves regions at high risk of motion blur and ghosting artifacts. Qualitative evidence is provided in our supplementary materials.", + "bbox": [ + 511, + 284, + 906, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to dynamic scenes. We offer a pair of examples from the Real-HDRV [26] dataset with around 30 and 60 pixels of regional movement, respectively. As shown in Fig. 7: Left, our trained agent tends to predict faster shutter speeds for scenes with stronger movement, validating the responsiveness of our agent in dynamic scenes. We assess our method's performance against previous auto-exposure methods [6, 32] at varying motion magnitudes (Fig. 7: Right). To create the evaluation dataset for robustness to different motion levels, we follow HDRFlow [34], using RAFT [28] to process dynamic scenes from DRV and obtain optical flow maps. We then manually crop these images with reasonable flow predictions, dividing them into $128 \\times 128$ blocks, and calculate the average motion magnitude for each block. Finally, we evaluate the PSNR of blocks corresponding to different motion magnitudes. As shown in Fig. 7: Right, our AdaptiveAE demonstrates greater robustness than other methods as motion magnitude increases.", + "bbox": [ + 511, + 421, + 908, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 708, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce AdaptiveAE, which optimizes HDR exposure in dynamic settings using deep reinforcement learning, treating exposure bracketing as a Markov Decision Process. It autonomously adjusts ISO and shutter speed for a pretrained exposure fusion algorithm. Reward systems focus on moving and key regions, minimizing sequence lengths. Experiments show AdaptiveAE outperforms state-of-the-art methods in dynamic scenes while matching top models in static ones, allowing flexible HDR capture. Our analysis of noise models and exposure settings offers insights for future research, with plans to include adjustable apertures.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25183", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 257, + 187, + 273 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Maryam Azimi, Vedad Hulusic, Philippe Hanhart, and Touradj Ebrahimi. PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR. In Proceedings of the Picture Coding Symposium (PCS), pages 1-5, 2021. 6", + "[2] Neil Barakat, A. Nicholas Hone, and Thomas E. Darcie. Minimal-bracketing sets for high-dynamic-range image capture. IEEE Transactions on Image Processing, 17(10):1864-1875, 2008. 6, 1", + "[3] Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, and Lei Zhang. HDR video reconstruction: A coarse-to-fine network and a real-world benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2502-2511, 2021. 2, 6, 7, 1", + "[4] Wenyuan Chen, Xingjian Liu, Shanghai Ru, and Yu Sun. Automated exposures selection for high dynamic range structured-light 3-D scanning. IEEE Transactions on Industrial Electronics, 70(7):7428-7437, 2023. 3", + "[5] Natasha Gelfand, Andrew Adams, Sung Hee Park, and Kari Pulli. Multi-exposure imaging on mobile devices. In Proceedings of the 18th ACM International Conference on Multimedia, pages 823-826, 2010. 3", + "[6] Samuel W Hasinoff, Frédo Durand, and William T Freeman. Noise-optimal capture for high dynamic range photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 553-560, 2010. 1, 2, 3, 4, 6, 7, 8", + "[7] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 6, 7, 1, 3", + "[8] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (TOG), 36(4):144:1-144:12, 2017. 2", + "[9] Nima Khademi Kalantari, Eli Shechtman, Connelly Barnes, Soheil Darabi, Dan B. Goldman, and Pradeep Sen. Patch-based high dynamic range video. ACM Transactions on Graphics (TOG), 32(6):202:1-202:10, 2013. 2", + "[10] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. SAFNet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of European Conference on Computer Vision (ECCV), 2024. 2" + ], + "bbox": [ + 93, + 282, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Acknowledgement. This work was supported by the National Key R&D Program of China No. 2022ZD0160201, Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No. 62136001, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources.", + "[11] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 2", + "[12] Jieyu Li, Ruiwen Zhen, and Robert L. Stevenson. A lightweight exposure bracketing strategy for HDR imaging without access to camera RAW. *Electronic Imaging*, 2023 (15):116-1-116-6, 2023. 3", + "[13] Shuaizheng Liu, Xindong Zhang, Lingchen Sun, Zhetong Liang, Hui Zeng, and Lei Zhang. Joint HDR denoising and fusion: A real-world mobile HDR image dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13966-13975, 2023. 2", + "[14] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 6, 7, 8", + "[15] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2020. 2", + "[16] Rafal Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM Transactions on Graphics (TOG), 30(4):1-14, 2011. 6", + "[17] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion: A simple and practical alternative to high dynamic range photography. Computer Graphics Forum, 28(1):161-171, 2009. 2", + "[18] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. arXiv preprint:1602.01783, 2016. 2, 5", + "[19] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson W. H. Lau. HDR-GAN: HDR image reconstruction from multi-exposed LDR images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 2, 6, 8", + "[20] Junting Pan, Cristian Canton Ferrer, Kevin McGuinness, Noel E. O'Connor, Jordi Torres, Elisa Sayrol, and Xavier Giro-i Nieto. SalGAN: Visual saliency prediction with generative adversarial networks. arXiv preprint:1701.01081, 2017. 6", + "[21] Reza Pourreza-Shahri and Nasser Kehtarnavaz. Exposure bracketing via automatic exposure selection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 320-323, 2015. 1, 2, 3, 6, 7", + "[22] Martin L. Puterman. Markov decision processes. In *Handbooks in Operations Research and Management Science*, chapter 8, pages 331-434. North-Holland, Amsterdam, 1990. 4", + "[23] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. TransMEF: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2126-2134, 2022. 2" + ], + "bbox": [ + 93, + 90, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25184", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Kalpana Seshadrinathan, Sung Hee Park, and Oscar Nestares. Noise and dynamic range optimal computational imaging. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 2785-2788, 2012. 3", + "[25] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5572-5581, 2019. 2", + "[26] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 1, 2, 6, 7, 8, 3", + "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2", + "[28] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 6, 8", + "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. BANet: A blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 2, 1, 3", + "[30] Peter van Beek. Improved image selection for stack-based HDR imaging. *Electronic Imaging*, 2019(4):581-1-581-6, 2019. 3", + "[31] Channarayapatna Shivaram Vijay, Chandramouli Paramanand, Ambasamudram Narayanan Rajagopalan, and Rama Chellappa. Non-uniform deblurring in HDR image reconstruction. IEEE Transactions on Image Processing, 22 (10):3739-3750, 2013. 2", + "[32] Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, and Jimmy S. Ren. Learning a reinforced agent for flexible exposure bracketing selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1820-1828, 2020. 1, 2, 3, 6, 7, 8", + "[33] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 6, 7", + "[34] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. HDRFlow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 8", + "[35] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2", + "[36] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Runsheng Yu, Wenyu Liu, Yasen Zhang, Zhi Qu, Deli Zhao, and Bo Zhang. DeepExposure: Learning to expose photos with asynchronously reinforced adversarial learning. In Advances in Neural Information Processing Systems 31, pages 8377-8386, 2018. 3", + "[38] Lu Yuan, Jian Sun, Long Quan, and Heung-Yeung Shum. Image deblurring with blurred/noisy image pairs. ACM Transactions on Graphics (TOG), 26(3):1:1-1:10, 2007. 2", + "[39] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Björn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2", + "[40] Zhilu Zhang, Haoyu Wang, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised high dynamic range imaging with multi-exposure images in dynamic scenes. arXiv preprint:2310.01840, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25185", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_model.json b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1b0a5626d1ac2637fbf84373a3289d557a37cf08 --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_model.json @@ -0,0 +1,2541 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.833, + 0.177 + ], + "angle": 0, + "content": "AdaptiveAE: An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.202, + 0.822, + 0.24 + ], + "angle": 0, + "content": "Tianyi Xu\\(^{1,3,4*}\\) Fan Zhang\\(^{1}\\) Boxin Shi\\(^{3,4\\dagger}\\) Tianfan Xue\\(^{2,1\\dagger}\\) Yujin Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)The Chinese University of Hong Kong" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.242, + 0.826, + 0.256 + ], + "angle": 0, + "content": "\\(^{3}\\)State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.259, + 0.825, + 0.275 + ], + "angle": 0, + "content": "\\(^{4}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.171, + 0.242, + 0.826, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.278, + 0.786, + 0.292 + ], + "angle": 0, + "content": "photon@stu.pku.edu.cn, zhangfan@pjlab.org.cn, shiboxin@pu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.296, + 0.693, + 0.31 + ], + "angle": 0, + "content": "tfxue@ie.cuhk.edu.hk, wangyujin@pjlab.org.cn" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.338, + 0.276, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.433, + 0.131, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.435, + 0.165, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.172, + 0.439, + 0.256, + 0.453 + ], + "angle": 0, + "content": "Ours-stage I" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.46, + 0.276, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.553, + 0.133, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.553, + 0.171, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.556, + 0.272, + 0.569 + ], + "angle": 0, + "content": "Pourreza et al." + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.339, + 0.49, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.434, + 0.348, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.435, + 0.386, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.438, + 0.478, + 0.452 + ], + "angle": 0, + "content": "Ours-stage II" + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.46, + 0.49, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.553, + 0.348, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.554, + 0.386, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.556, + 0.466, + 0.57 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.339, + 0.704, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.434, + 0.562, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.436, + 0.6, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.439, + 0.696, + 0.453 + ], + "angle": 0, + "content": "Ours-stage III" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.46, + 0.703, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.553, + 0.562, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.554, + 0.6, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.557, + 0.696, + 0.57 + ], + "angle": 0, + "content": "Hasinoff et al." + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.342, + 0.904, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.484, + 0.755, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.484, + 0.788, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.491, + 0.901, + 0.503 + ], + "angle": 0, + "content": "Deep Learning Based" + }, + { + "type": "image", + "bbox": [ + 0.724, + 0.516, + 0.753, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.519, + 0.829, + 0.531 + ], + "angle": 0, + "content": "Shutter Speed" + }, + { + "type": "image", + "bbox": [ + 0.845, + 0.516, + 0.874, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.883, + 0.522, + 0.904, + 0.532 + ], + "angle": 0, + "content": "ISO" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.547, + 0.758, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.548, + 0.899, + 0.56 + ], + "angle": 0, + "content": "Adjustable / Unadjustable" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.588, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Figure 1. AdaptiveAE takes camera preview images as input and automatically predicts the ISO and shutter speed for each LDR captures for exposure fusion through a 3-stage sequential refinement procedure to achieve an optimal balance between noise level and motion-related problems for high quality HDR capturing in dynamic scenes with deep reinforcement learning. AdaptiveAE achieves PSNR 39.7 on HDR dataset [26], while baseline methods [6, 21, 32] that either only predicts shutter speed or do not consider motion can only achieve PSNR below 37.6 and has evident motion blur and ghosting artifacts in HDR results." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.671, + 0.327, + 0.687 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Mainstream high dynamic range imaging techniques typically rely on fusing multiple images captured with different exposure setups (shutter speed and ISO). A good balance between shutter speed and ISO is crucial for achieving high-quality HDR, as high ISO values introduce significant noise, while long shutter speeds can lead to noticeable motion blur. However, existing methods often overlook the complex interaction between shutter speed and ISO and fail to account for motion blur effects in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.839, + 0.483, + 0.854 + ], + "angle": 0, + "content": "In this work, we propose AdaptiveAE, a reinforcement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.838 + ], + "angle": 0, + "content": "learning-based method that optimizes the selection of shutter speed and ISO combinations to maximize HDR reconstruction quality in dynamic environments. AdaptiveAE integrates an image synthesis pipeline that incorporates motion blur and noise simulation into our training procedure, leveraging semantic information and exposure histograms. It can adaptively select optimal ISO and shutter speed sequences based on a user-defined exposure time budget, and find a better exposure schedule than traditional solutions. Experimental results across multiple datasets demonstrate that it achieves the state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.846, + 0.646, + 0.861 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "High-dynamic-range (HDR) imaging plays a pivotal role in computational photography. To capture an HDR scene," + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.888 + ], + "angle": 0, + "content": "*This work was done during Tianyi Xu's internship at Shanghai AI Laboratory." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.247, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "due to hardware limits, a single capture can only cover a Low Dynamic Range (LDR), and HDR fusion techniques are proposed to combine multiple LDR images with varying exposures to cover a wide dynamic range [7, 10, 14, 15, 17, 19, 23, 35, 36, 40]. The typical way to vary exposure is to change either shutter speed or ISO. This is a challenging process, as a longer shutter speed may increase the signal-to-noise ratio but introduce unpleasant motion blur, a large ISO may increase brightness but also magnify noise, and significant exposure differences can cover a wider dynamic range but also increase the risk of misalignment. Therefore, the choice of exposure values (EVs) for each capture is critical in this process to ensure high-quality HDR results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.482, + 0.442 + ], + "angle": 0, + "content": "Still, very limited work discusses how to choose the optimal exposure levels, particularly in dynamic scenes. Prior works on exposure scheduling mainly focus on static scenes and ignores potential motion blur. Learning-based techniques overlook the intricate interplay between ISO and shutter speed, resulting in suboptimal image quality under varying conditions [6, 21, 32]. Additionally, many existing methods treat ghosting and motion blur as separate, computationally intensive post-processing tasks, which is unsuitable for real-time applications [11, 25, 27, 29, 31, 38, 39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.482, + 0.581 + ], + "angle": 0, + "content": "In this work, we propose AdaptiveAE, an efficient exposure control algorithm designed for HDR capturing, which addresses both motion blur and noise during image acquisition. Given the previously captured image, our method optimizes the exposure bracketing strategy for the subsequent capture, based on illumination information and semantic data from previous frames. Unlike previous approaches that treat motion blur as a separate post-processing task, we aim to address it during the capture process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.482, + 0.809 + ], + "angle": 0, + "content": "Designing both efficient and adaptive exposure control is non-trivial, and we resort to reinforcement learning [18] to solve this challenge. Our solution mimics an experienced photographer. At each iteration, the policy network takes the previously captured LDR images as input, together with extracted semantic and illumination information. Given this information, the policy network learns to determine the optimal exposure setup for the following captures, which maximizes the additional information provided by this frame while also reducing the risk of misalignment and motion blur in a dynamic scene. Once the next frame is captured, this newly captured image will be used as input for the subsequent refinement iteration. As shown in Fig. 1 right, the final quality (PSNR) of fusion increases as more images are captured." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.903 + ], + "angle": 0, + "content": "Our proposed approach offers several advantages over traditional exposure controls. First, AdaptiveAE controls both exposure time and ISO and also adapts to different scenes. As a result, it has a much higher upper bound compared to either a fixed exposure schedule or an adaptive control algorithm that only changes exposure time. As shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "in Fig. 1, our method iteratively achieves an optimal balance between noise and blur, resulting in less noise, reduced motion blur, and superior image quality compared to other baseline methods. Second, our method can handle both static and dynamic scenes. In static scenes, it achieves performance comparable to state-of-the-art techniques, and in dynamic ones, it produces visually compelling HDR images with minimal motion blur and ghosting artifacts. Third, our method can automatically choose the best number of frames for HDR imaging. Unlike traditional HDR approaches that often use a fixed number of frames, such as three, our approach provides flexibility to determine whether capturing three or more frames is optimal for certain scenes, balancing image quality and time budget." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.905, + 0.427 + ], + "angle": 0, + "content": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach enables the concurrent analysis of blur and noise in AE prediction, thereby enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.626 + ], + "angle": 0, + "content": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Additionally, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and underscore the critical importance of incorporating blur synthesis into our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II, demonstrates superior noise control and effectively reduces motion blur, outperforming baselines in the visual quality of the fused HDR images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.748 + ], + "angle": 0, + "content": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach allows for concurrent analysis of blur and noise in AE prediction, enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.903 + ], + "angle": 0, + "content": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Also, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and highlight the critical importance of incorporating blur synthesis within our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "demonstrates superior noise control and effectively reduces motion blur, outperforming compared baselines in the visual quality of the fused HDR images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.144, + 0.229, + 0.159 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.484, + 0.591 + ], + "angle": 0, + "content": "Strategy for exposure bracketing. Determining the optimal set of exposures for multiple-exposure dynamic range imaging is a well-established problem. Most digital cameras allow users to set the compensation ratio for exposure bracketing, while mobile cameras typically impose fixed ratios during automatic exposure bracketing. Heuristic strategies based on the histogram are proposed by [5, 21, 30], to balance single-to-noise ratio (SNR) and saturation. The work by [6] firstly formulates this challenge as a constrained optimization problem in the linear RGB domain, addressing the scenario that follows multiple exposure fusion and precedes tone mapping and denoising. This concise formulation facilitates the use of a straightforward numerical solver. Further extensions of this formulation consider the alignment of multiple input images to account for handshake [24]. In the context of structured light 3D reconstruction, various exposures are likewise treated as an optimization problem [4]. Exposure influences not only the noise but also, to some extent, the tone of the image in the standard RGB domain. Evaluating the final image quality complicates the problem further, as subsequent tone mapping or retouching can significantly alter image quality. Therefore, a neural network is utilized to estimate exposures and fuse multiple exposed images to achieve optimal fidelity in the gamma-corrected domain [12]. Additionally, reinforcement learning is employed by [37] and [32] to assess the rewards on comprehensive image quality after more sophisticated tone adjustment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.789 + ], + "angle": 0, + "content": "Challenge in dynamic scenes. In static scenes, increasing under-exposures can reduce saturation, and a longer shutter speed improves signal-to-noise ratio in dark areas. However, in dynamic scenes, excessive exposure may cause ghosting artifacts, and a prolonged shutter speed can lead to motion blur. Consequently, heuristic exposure bracketing is usually limited to two or three EV settings [5, 30]. The method by [6] imposes an upper limit on total shutter speed, while [24] addresses handshake motion through image registration but scarcely tackles local object motion. Most approaches do not fully address motion in dynamic scenes, leaving motion blur and ghosting artifacts to be addressed through post-processing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.182, + 0.815 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Conventional weighted linear combination methods, such as exposure fusion, provide straightforward SNR estimation but fail in dynamic scenes where moving objects create unquantifiable motion blur and ghosting artifacts due to misalignment. Our approach addresses these challenges dur" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.178, + 0.817, + 0.191 + ], + "angle": 0, + "content": "(a) Our overall image synthesis pipeline." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.279, + 0.793, + 0.292 + ], + "angle": 0, + "content": "(b) Our blur synthesis pipeline." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.382, + 0.796, + 0.394 + ], + "angle": 0, + "content": "(c) Our noise synthesis pipeline." + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.406, + 0.855, + 0.42 + ], + "angle": 0, + "content": "Figure 2. Our blur-aware data synthesis pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.595 + ], + "angle": 0, + "content": "ing capturing rather than in post-processing, as the latter is shown to yield suboptimal results. We predict exposure-related risks—such as motion blur, ghosting, noise, and saturation—based on a limited number of previously captured frames. Additionally, we employ a sequential strategy for exposure and ISO parameter determination, rather than simultaneously predicting settings for all three LDR images, which reflects the iterative nature of auto-exposure in mobile cameras that enables adaptation to significant brightness transitions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.816, + 0.621 + ], + "angle": 0, + "content": "3.1.Blur-aware data synthesis pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.715 + ], + "angle": 0, + "content": "To simulate capturing in real environments, we designed an image synthesis pipeline to generate realistic motion blur and noise in LDR images from HDR videos in the training dataset, for use in training. Typical exposure settings involve adjusting the exposure value (EV), which is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.713, + 0.905, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E V} = \\log_ {2} \\left(\\frac {F ^ {2}}{T} \\times \\frac {1 0 0}{\\mathrm {I S O}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.78 + ], + "angle": 0, + "content": "where \\(F\\) denotes the aperture's f-number, ISO represents the ISO sensitivity and \\(T\\) is the exposure time in seconds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Similar to recent methods [6, 21, 32] concerning high dynamic range capture, we assume that aperture and focus are held constant to prevent changes in defocus. This leaves just two camera settings to manipulate: (1) Shutter speed, which controls the amount of light to collect, and (2) ISO, which determines the sensor gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our pipeline synthesizes motion blur and noise for the ground truth static HDR image based on a specified ISO" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.9, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.304, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 3. The training scheme of AdaptiveAE. States are defined as the three LDRs synthesized using predicted ISOs and shutter speeds. Starting from \\( s_0 \\) where the three LDRs has EV \\( \\{-2,0, + 2\\} \\) with arbitrary EV 0 baseline, ISOs and shutter speeds, the agent sequentially predicts, customizes or inherits capturing parameters (i.e. ISO and shutter speed) for the next stage and synthesize the corresponding LDR using our image synthesis pipeline. Unlike training, the LDRs will be captured rather than synthesized during inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.368, + 0.482, + 0.502 + ], + "angle": 0, + "content": "and shutter speed. As depicted in Fig. 2, it uses two consecutive static HDR images as input. Motion blur is first synthesized according to shutter speed to produce a blurred HDR image in linear space. Next, noise is added based on shutter speed and ISO to create an LDR image, reflecting our exposure choice. Note that motion blur should be applied before adding noise, as it influences the raw input by affecting the number and pattern of captured photons during photography." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Synthesizing blur. On-the-shelf training datasets consist of consecutive HDR ground truth of a scene with motion. Now we explain how we simulate motion blur to a frame of HDR \\( f_{i}^{L} \\), as shown in Fig. 2b, which is the \\( i \\)-th HDR ground truth frame in the dataset scene, and the superscript \\( L \\) indicates it is in linear space. We first use \\( \\mu \\)-law tone-mapping with \\( \\mu = 5000 \\) to transfer \\( f_{i}^{L} \\) and \\( f_{i+1}^{L} \\) from HDR space to LDR space where the image interpolation algorithm we applied is trained upon, receiving \\( f_{i}^{\\mathcal{T}} \\) and \\( f_{i+1}^{\\mathcal{T}} \\), where the superscript \\( \\mathcal{T} \\) denotes they are in LDR space. Then we use RIFE [7] to interpolate the them to 256 frames and get the sequence of images \\( \\{f_{i}^{\\mathcal{T}}, s_{1}^{\\mathcal{T}}, s_{2}^{\\mathcal{T}}, \\dots, s_{254}^{\\mathcal{T}}, f_{i+1}^{\\mathcal{T}}\\} \\). Then for the selected shutter speed \\( T_{j} \\) for the \\( j \\)-th LDR \\( l_{j}^{\\mathcal{T}} \\) to take, the blurred HDR \\( b_{j}^{L} \\) is simulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.731, + 0.483, + 0.767 + ], + "angle": 0, + "content": "\\[\nb _ {j} ^ {L} = \\mathbf {i T M O} \\left(\\frac {f _ {i} ^ {\\mathcal {T}} + \\sum_ {m = 1} ^ {m _ {j}} s _ {m} ^ {\\mathcal {T}}}{m _ {j}}\\right), m _ {j} = \\left\\lceil \\frac {2 5 6 T _ {j}}{\\Delta \\tau} \\right\\rceil , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\(\\Delta \\tau\\) denotes the time elapsed from \\(f_{i}^{L}\\) is taken to \\(f_{i + 1}^{L}\\) is not yet taken and iTMO indicates the inverse \\(\\mu\\)-law tonemapping function with \\(\\mu = 5000\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Synthesizing noise. We adopt the noise model mentioned in [6], in which noise is modeled as a zero-mean variable, coming from three independent sources, including photon noise, which represents the Poisson distribution of photon arrivals and depends linearly upon the number of recorded" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.429 + ], + "angle": 0, + "content": "electrons, \\(\\Phi T\\), readout noise, which comes from sensor readout, and analog-to-digital conversion(ADC) noise, which comes from the combined effect of amplifier and quantization. Hence, for pixels below the saturation level:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.437, + 0.907, + 0.469 + ], + "angle": 0, + "content": "\\[\nV a r (n) = \\frac {\\Phi T \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\frac {\\sigma_ {\\text {r e a d}} ^ {2} \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\sigma_ {\\mathrm {A D C}} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.506 + ], + "angle": 0, + "content": "where \\(\\Phi\\) is the radiance level, \\(T\\) is the shutter speed, \\(U\\) is a camera-dependent variable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.571 + ], + "angle": 0, + "content": "Following our noise model, as shown in Fig. 2c, we can synthesize the corresponding noise with selected ISO and shutter speed to the blurred HDR \\( b_{j}^{L} \\) to get the LDR image \\( l_{j}^{\\mathcal{T}} \\). For details, please refer to our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.831, + 0.594 + ], + "angle": 0, + "content": "3.2. Problem formulation of AdaptiveAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Given a scene, the goal of AdaptiveAE, which has access to three initial preview LDR images \\(\\{p_j^T\\}_{1,2,3}\\) (i.e. underexposed, mid-exposed, overexposed) before capturing, is to find an optimal exposure setup (i.e. ISO and shutter speed) for LDR capturing, with which the fused HDR output will result in pleasing visual performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.796 + ], + "angle": 0, + "content": "Our method formulates exposure bracketing as a Markov Decision Process [22], solved via deep reinforcement learning to refine exposure parameters (ISO, shutter speed) sequentially. As illustrated in Fig. 4, the process starts from three LDRs at a default \\(\\{-2_p,0_p, + 2_p\\}\\) EV spacing relative to an arbitrary reference (subscript \\(p\\)). The refinement then proceeds in stages, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "First, the agent predicts optimal parameters for the midexposed frame, establishing a new 0-EV reference (subscript \\(x\\)). The side frames are then customized—their parameters are procedurally set to achieve a symmetric \\(\\{-2_x,0_x, + 2_x\\}\\) EV bracket. Next, the agent refines the underexposed frame to an EV of \\(-y_{x}\\). The mid-exposed frame \\((0_{x})\\) is inherited (its parameters are reused), while" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.087, + 0.862, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Training pipeline of our method. The ISO and shutter speed prediction process is conceptualized as a Markov Decision Process, where a CNN-based policy network predicts the ISO and shutter speed of the next exposure sets. Concurrently, a CNN-based value network estimates the state value. We leverage our blur-aware image synthesis pipeline to synthesize the predicted LDRs and employ DeepHDR [7] to fuse the predicted LDR images, generating our HDR result and calculating the reward for the current policy. The entire system is optimized using the A3C (Asynchronous Advantage Actor-Critic) method [18]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.484, + 0.468 + ], + "angle": 0, + "content": "the overexposed frame is customized to \\( +y_{x} \\) to maintain symmetry, yielding \\( \\{-y_{x}, 0_{x}, + y_{x}\\} \\). Finally, the agent predicts the overexposed frame's EV as \\( +z_{x} \\), creating a potentially asymmetric set \\( \\{-y_{x}, 0_{x}, + z_{x}\\} \\). This sequential prediction can be extended, as our adopted fusion method [7] handles more than three LDRs. A video providing further details on this process, along with an example of extended exposure bracketing, is available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.31, + 0.494 + ], + "angle": 0, + "content": "3.3. Optimization objectives" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.5, + 0.484, + 0.787 + ], + "angle": 0, + "content": "Let us denote the problem as \\( P = (S, A) \\), where S is a state space and A is an action space. Specifically, in our task, \\( S \\) is the space of the exposure setups (i.e., ISO and shutter speed) for an LDR set that typically contains three LDRs (i.e. under-exposed, mid-exposed, and overexposed), while \\( A \\) is the set of all possible ISO and shutter speed combinations, which is discrete. During training, at stage \\( s_j = \\{(\\mathrm{ISO}_{j1}, T_{j1}), (\\mathrm{ISO}_{j2}, T_{j2}), (\\mathrm{ISO}_{j3}, T_{j3}), \\} \\), we first find the corresponding HDR ground truth image pairs \\( (f_i^L, f_{i+1}^L)_{1,2,3} \\) and generate the corresponding LDRs \\( \\{l_j^T\\}_{1,2,3} \\) through our image synthesis pipeline in Fig. 2. The footnote 1, 2, 3 denotes that the same operation is done for the under-, mid-, and over-exposed LDRs. Taking \\( \\{l_j^T\\}_{1,2,3} \\) as input, the agent predicts an action \\( a_j = (\\mathrm{ISO}_j, T_j) \\), which is expanded to an exposure setup for three LDRs by customizing EV or inheriting from state \\( s_{j-1} \\), mapping state \\( s_j \\) to state \\( s_{j+1} \\). Adding a sequence of \\( M \\) LDRs to exposure bracketing corresponds to a trajectory \\( \\tau \\) of states and actions:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.8, + 0.482, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\tau = \\left(s _ {0}, a _ {0}, \\dots , s _ {M - 1}, a _ {M - 1}, s _ {M}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( s_M \\) is the stopping state. Our goal is to find a policy that maximizes the accumulated reward during the decision-making process. In this paper, the reward function with the \\( j \\)-th action (i.e., corresponding to deciding the exposure setting for the \\( j \\)-th LDR for exposure fusion) is thus written as:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.349, + 0.905, + 0.366 + ], + "angle": 0, + "content": "\\[\nr \\left(s _ {j}, a _ {j}\\right) = \\mathcal {R} \\left(s _ {j + 1}\\right) - \\mathcal {R} \\left(s _ {j}\\right) - \\mathcal {P} (j), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.403 + ], + "angle": 0, + "content": "where \\( s_{j + 1} = p(s_j,a_j) \\), and \\( \\mathcal{R} \\) denotes our reward design and \\( \\mathcal{P} \\) denotes the \\( L_{\\mathrm{step}} \\) penalty, detailed in 3.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.616 + ], + "angle": 0, + "content": "As depicted in Fig. 4, our model comprises a policy network and a value network, both of which utilize a CNN-based architecture. The policy network predicts the optimal ISO and shutter speed for the subsequent exposure, outputting a distribution of action probabilities \\(\\pi(s,\\theta)\\) for an input image \\(s\\). Concurrently, the value network \\(V^{\\pi}(s,\\omega)\\) estimates the corresponding state value. These networks, with combined parameters \\(\\psi = (\\theta,\\omega)\\), are trained by maximizing our objective \\(J(\\theta)_{\\psi}\\) to learn the optimal policy \\(\\pi(s)\\). Specifically, to train the policy network and the value network, we apply the A3C (Asynchronous Advantage Actor-Critic) method [18], where the actor is represented by the policy network and the critic is the value network. Network details are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.612, + 0.64 + ], + "angle": 0, + "content": "3.4. Reward" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.906, + 0.738 + ], + "angle": 0, + "content": "When designing the reward function for our system, we considered four key factors: (1) similarity between the fused HDR and the ground truth HDR; (2) quality of important regions in the fused HDR; (3) quality of moving regions in the fused HDR; and (4) a penalty for overly long LDR stacks. Thus, our reward function is:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.751, + 0.905, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = - \\left(P _ {\\text {c o n s t r u c t i o n}} + P _ {\\text {p r i o r i t y}} + P _ {\\text {g h o s t}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We consider \\( P_{\\text {construction }} \\), the L2 loss between our fused HDR and the ground truth, as our major reward component, which is affected by noise and saturation. Note that through the entire sequential decision process, the middle-exposed frame is used as a reference for HDR fusion. Conforming to Eq. (3), during training, noise is synthesized according to the irradiance of ground truth HDR, and during inference, noise is estimated with the irradiance of noisy signals." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25180" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "\\(P_{\\mathrm{priority}}\\) represents an L2 loss within the areas in the image masked by an importance mask, which is generated by a saliency predictor [20]. This ensures that the highest quality is maintained in the most significant areas, thereby enhancing the overall visual fidelity where it matters most." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.168, + 0.483, + 0.362 + ], + "angle": 0, + "content": "\\(P_{\\mathrm{ghost}}\\) is also an L2 loss within a masked area, denoting areas with large motions and thus having a higher risk of motion blur- or ghosting-caused HDR quality degradation. The mask is computed by calculating the optical flow using RAFT [28] between the HDR ground truth of the middle-exposed frame (i.e., the reference frame for fusion) and the corresponding HDR \\(f_{i}^{L}\\) and selecting the pixels where the mode of the flow vector exceeds a constant threshold \\(K\\). Normalizing the largest optical flow vector, we empirically set \\(K\\) to 0.2. \\(P_{\\mathrm{ghost}}\\) guides the agent to deal more carefully with regions that are prone to artifacts caused by motion and is helpful for high-quality HDR capturing, as is verified by the result of our ablation studies in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.469 + ], + "angle": 0, + "content": "\\(\\mathcal{P}(j)\\) is a penalty designed to penalize excessively long exposure brackets. Capturing an excessive number of shots, such as 10, even with random exposure settings, can lead to nearly perfect HDR fusion results, but it is time-consuming. Typically, three shots [2, 6, 21] are sufficient to achieve high-quality outcomes. To this end, we incorporate a penalty for taking more than three shots, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.48, + 0.482, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} (j) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} j \\leq H \\\\ \\alpha (j - H) ^ {2} & \\text {i f} j > H \\end{array} , \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.421, + 0.544 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a positive coefficient and \\(H\\) is set to 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.546, + 0.483, + 0.711 + ], + "angle": 0, + "content": "In this manner, the autonomous agent optimizes exposure parameters by predicting relatively fast shutter speeds for LDR images, particularly for the middle-exposed reference frame, thereby minimizing motion blur while avoiding excessive ISO values that would introduce noise-related degradation. When confronted with potential ghosting artifacts—which emerge from information deficiency in LDR images due to concurrent saturation and motion—the agent adaptively selects EVs that minimize both underexposure and saturation, resulting in a significant reduction of ghosting artifacts in the final reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.222, + 0.743 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Experiment details. We use Real-HDRV [26] for training and tested our performance on Real-HDRV [26] and DeepHDRVideo [3]. For HDR fusion, we adopt DeepHDR [33] to generate the HDR image based on the selected exposure bracketing. Additionally, we set the number of LDR frames to 3 for all methods involved. For methods that do not account for changes in ISO values, we set the ISO to 200 as a standard value for most cases, which is rationalized in our supplementary materials. Since RIFE [7] is relatively time-consuming, the image interpolation and blur synthesis step" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "is performed before training. We apply random flipping, rotation, and cropping with \\(512 \\times 512\\) pixels for data augmentation. Our training dataset consists of a total of 770 scenes, including 440 dynamic and 330 static scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Evaluation metrics. We directly evaluate the performance of the fused HDR results. Similar to previous HDR fusion methods [7, 14, 19], we employ PSNR-\\(\\mu\\), SSIM-\\(\\mu\\), PU-PSNR, PU-SSIM, and HDR-VDP-2 [16] as evaluation metrics. PSNR-\\(\\mu\\) and SSIM-\\(\\mu\\) denote PSNR and SSIM of the fused HDR after \\(\\mu\\)-law tone-mapping with \\(\\mu = 5000\\). PU-PSNR and PU-SSIM are computed after perceptually uniform encoding [1]. When computing the HDR-VDP-2 [16], the diagonal display size is 30 inches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.302, + 0.608, + 0.316 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.49 + ], + "angle": 0, + "content": "Results on Real-HDRV dataset. We compared our trained agent's performance with several state-of-the-art HDR exposure bracketing methods, including Pourreza-Shahri et al. [21], Hasinoff et al. [6], and Wang et al. [32]. The first two are non-deep-learning methods that do not consider motion: Pourreza-Shahri et al. use K-means clustering to adjust shutter speed based on image brightness, while Hasinoff et al. mathematically optimize ISO and shutter speed for the best worst-case SNR. Wang et al. utilize reinforcement learning to predict shutter speed for maximizing PSNR, but also ignore motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.734 + ], + "angle": 0, + "content": "Our method achieves state-of-the-art performance on the HDRV-Test dataset, as shown in Tab. 1. By treating ISO as a variable, it provides flexibility in handling extremely dark scenes. Additionally, it excels in dynamic scenes due to a blur synthesis model and carefully designed rewards. Thus, our model effectively balances noise reduction and motion artifact minimization, delivering high-quality HDR results. Visualization results in Fig. 5 show that deep learning-based exposure fusion models affect EV selection differently. In Hasinoff et al.'s setup, three LDRs are equally weighted, but dynamic scenes require a reference image, emphasizing the middle-exposed LDR's quality. This often shifts its EV toward under-exposure to reduce motion blur. Our experiments confirm that if the middle-exposed image is blurry—a common problem in other methods—the fused HDR will also be blurred." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Inference time. Our method takes less than \\(250\\mathrm{ms}\\) for each HDR image, which is acceptable. With average exposure time \\(n\\) (\\(\\leq 30ms\\)) and prediction time \\(m\\) (\\(\\leq 10ms\\)), the total execution time is \\(6n + 3m\\) (\\(\\leq 250ms\\)). This can be further lowered to \\(6n\\) (\\(\\leq 200ms\\)) adopting concurrency. Note that DeepHDR is only required during training for calculating rewards, and on-camera inference can be performed with a small agent, taking only \\(3.5ms\\) for each frame without optimization. Besides, our method can be further accelerated by on-camera processors when applied to DSLRs. More discussions are in our supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.188, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.38, + 0.183, + 0.395 + ], + "angle": 0, + "content": "LDR Images" + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.091, + 0.465, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.38, + 0.424, + 0.395 + ], + "angle": 0, + "content": "Our Tonemapped HDR Image" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.091, + 0.911, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.38, + 0.763, + 0.393 + ], + "angle": 0, + "content": "Wang et al. Pourreza et al. Hasinoff et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.38, + 0.79, + 0.391 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.38, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons with other auto-exposure methods on HDRV dataset [26]. Left: Predicted LDRs with varying ISO and shutter speed settings and synthesized using our image synthesis pipeline. Middle: Fused HDR image using DeepHDR [7] and tonemapined using Photomatix. Right: Zoom-in results for tested methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.466, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Table 1. Comparison of Different Methods. We utilized one preview image and three preview images, respectively, for the compared methods and our method to predict exposure settings for three LDRs. We leveraged DeepHDR [7] for exposure fusion and used the mentioned metrics to evaluate the quality of the fused HDR. Bold: The best." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.513, + 0.905, + 0.607 + ], + "angle": 0, + "content": "
MethodsHDRV [26]DeepHDRVideo [3]
PSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIMPSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIM
Pourreza et al. [21]33.640.861754.5530.610.867935.570.878055.6731.590.8791
Hasinoff et al. [6]37.590.905257.0232.870.898038.470.915758.6534.450.9132
Wang et al. [32]36.460.890256.0932.680.893337.950.901957.3933.270.9008
Ours39.700.940859.2034.670.946539.810.937158.9036.190.9338
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Table 2. Ablation study of AdaptiveAE on HDRV [26] dataset. Base denotes our model trained with only the step penalty and construction reward. Bold: The best." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.665, + 0.462, + 0.726 + ], + "angle": 0, + "content": "
ModelPSNR-μSSIM-μPU-PSNRPU-SSIM
Base38.210.922732.680.9198
Base+Ppriority38.570.926133.020.9239
Base+Ppriority+Pghost39.700.940834.670.9465
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Gap to the best-achievable. For each scene in the test set of Real-HDRV, we iteratively search for the best set of predictions by Gaussian sampling around our initial prediction (50 times per exposure parameter per frame, with a deviation of \\(20\\%\\) of the mean for both ISO and shutter speed). Statistics in Tab. 4 show that our method approaches the locally optimal result while being efficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Cross datasets test. To test the generalization ability of our model, we also evaluated the performance of our trained agent on DeepHDRVideo [3], as shown in the right of Tab. 1. Our agent exhibits good generalization abilities." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.612, + 0.907, + 0.725 + ], + "angle": 0, + "content": "Table 3. Comparison of different exposure fusion methods. We utilize the four auto-exposure methods to predict exposure settings and employ our image synthesis pipeline to create three LDRs. Then, we apply different exposure fusion methods to fuse them and compare the HDR quality on the HDRV dataset [26]. HDRTransformer [14] is pretrained on HDRV [26] dataset. P: PSNR- \\(\\mu\\) S: SSIM- \\(\\mu\\) . Pou: Pourreza et al. [21], Has: Hasinoff et al. [6], W: Wang et al. [32]. HDR-Trans: HDR-Transformer. Bold: The best." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.734, + 0.915, + 0.823 + ], + "angle": 0, + "content": "
ModelDeepHDRHDR-GANHDR-Trans
P-μS-μP-μS-μP-μS-μ
Pou33.640.861735.710.889235.840.8824
Has37.590.925238.580.926339.110.9372
W36.460.900237.950.916938.890.9210
Ours39.700.940840.730.937641.370.9478
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Since the DeepHDRVideo dataset only provides ground truth HDR for the middle image in the sequence, we synthesize HDR for the other images using DeepHDR [33]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Cross HDR fusion methods test. In our training scheme," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25182" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 4. Performance comparison investigating the gap between our method and the optimum on the HDRV-test dataset." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.128, + 0.462, + 0.179 + ], + "angle": 0, + "content": "
OursWorstAverageBest
PSNR-μ39.7025.7632.4139.93
SSIM-μ0.94080.77380.86090.9412
" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.202, + 0.225, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.331, + 0.191, + 0.342 + ], + "angle": 0, + "content": "Wang et al." + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.202, + 0.347, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.331, + 0.325, + 0.342 + ], + "angle": 0, + "content": "Hasinoff et al." + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.202, + 0.471, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.331, + 0.427, + 0.341 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Figure 6. Results of our real capture data. The subject performs steady and repetitive movements, taking shots according to the exposure settings predicted by various methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.414, + 0.483, + 0.655 + ], + "angle": 0, + "content": "we used DeepHDR for exposure fusion. For further testing, we adopted different exposure methods, including HDR-GAN [19] and HDR-Transformer [14] (pre-trained on the HDRV-dataset [26]), as post-processing exposure fusion methods. All tests are conducted on the HDRV-Test dataset. As shown in Tab. 3, without considering motion before LDR capturing, however powerful an exposure fusion method (i.e. HDR-GAN and HDR-Transformer) fails to yield a satisfactory result. Notably, when stronger fusion models are used, the performance gap between our model and traditional models, which do not account for motion, increases. This is because, when blur and ghosting risk are mitigated, post-processing models can effectively manage the remaining challenges. In contrast, ignoring motion makes these challenges more difficult for post-processing models, leading to potential failure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Results for real capture. We used a SONY Alpha 7C-II to evaluate our model on real-world scenes. Subjects performed steady, repetitive movements while ISO and shutter speed were manually set for each capture, with the aperture fixed at f/2.8. The camera and subject positions remained unchanged during multi-exposure captures. As shown in Fig. 6, our method provides noise control comparable to other baselines while effectively mitigating motion blur, which can impair the visual quality of the fused HDR image; for more results, see our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.253, + 0.833 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of penalty item. We conduct ablation studies to validate the effectiveness of our reward design. All quantitative evaluations are conducted on the HDRV-Test dataset. We train our networks using only the construction" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.591, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.136, + 0.591, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.179, + 0.59, + 0.19 + ], + "angle": 0, + "content": "Under-exposed" + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.092, + 0.666, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.135, + 0.666, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.179, + 0.666, + 0.19 + ], + "angle": 0, + "content": "Middle-exposed" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.092, + 0.74, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.135, + 0.74, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.179, + 0.737, + 0.19 + ], + "angle": 0, + "content": "Over-exposed" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.091, + 0.911, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.196, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Figure 7. Left: Zoom-in details of the moving regions and our predicted ISO and shutter speed (seconds) for each LDR. Up: LDRs for scenes with an average motion level of around 30 pixels, Down: around 60 pixels. Right: Comparisons with baseline methods across different motion magnitude ranges. Tested on Real-HDRV [26] dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.42 + ], + "angle": 0, + "content": "reward and step penalty as the base model and test the effectiveness of \\( P_{\\mathrm{ghost}} \\) and \\( P_{\\mathrm{priority}} \\). As shown in Tab. 2, these two reward elements enhance our agent's performance by directing the model to focus on semantically important regions and moving areas. The former typically corresponds to the scene's focal point (e.g. a person's face), while the latter involves regions at high risk of motion blur and ghosting artifacts. Qualitative evidence is provided in our supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.909, + 0.694 + ], + "angle": 0, + "content": "Robustness to dynamic scenes. We offer a pair of examples from the Real-HDRV [26] dataset with around 30 and 60 pixels of regional movement, respectively. As shown in Fig. 7: Left, our trained agent tends to predict faster shutter speeds for scenes with stronger movement, validating the responsiveness of our agent in dynamic scenes. We assess our method's performance against previous auto-exposure methods [6, 32] at varying motion magnitudes (Fig. 7: Right). To create the evaluation dataset for robustness to different motion levels, we follow HDRFlow [34], using RAFT [28] to process dynamic scenes from DRV and obtain optical flow maps. We then manually crop these images with reasonable flow predictions, dividing them into \\(128 \\times 128\\) blocks, and calculate the average motion magnitude for each block. Finally, we evaluate the PSNR of blocks corresponding to different motion magnitudes. As shown in Fig. 7: Right, our AdaptiveAE demonstrates greater robustness than other methods as motion magnitude increases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.709, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce AdaptiveAE, which optimizes HDR exposure in dynamic settings using deep reinforcement learning, treating exposure bracketing as a Markov Decision Process. It autonomously adjusts ISO and shutter speed for a pretrained exposure fusion algorithm. Reward systems focus on moving and key regions, minimizing sequence lengths. Experiments show AdaptiveAE outperforms state-of-the-art methods in dynamic scenes while matching top models in static ones, allowing flexible HDR capture. Our analysis of noise models and exposure settings offers insights for future research, with plans to include adjustable apertures." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25183" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.234 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Key R&D Program of China No. 2022ZD0160201, Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No. 62136001, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.258, + 0.188, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[1] Maryam Azimi, Vedad Hulusic, Philippe Hanhart, and Touradj Ebrahimi. PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR. In Proceedings of the Picture Coding Symposium (PCS), pages 1-5, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[2] Neil Barakat, A. Nicholas Hone, and Thomas E. Darcie. Minimal-bracketing sets for high-dynamic-range image capture. IEEE Transactions on Image Processing, 17(10):1864-1875, 2008. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.491 + ], + "angle": 0, + "content": "[3] Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, and Lei Zhang. HDR video reconstruction: A coarse-to-fine network and a real-world benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2502-2511, 2021. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[4] Wenyuan Chen, Xingjian Liu, Shanghai Ru, and Yu Sun. Automated exposures selection for high dynamic range structured-light 3-D scanning. IEEE Transactions on Industrial Electronics, 70(7):7428-7437, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[5] Natasha Gelfand, Andrew Adams, Sung Hee Park, and Kari Pulli. Multi-exposure imaging on mobile devices. In Proceedings of the 18th ACM International Conference on Multimedia, pages 823-826, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[6] Samuel W Hasinoff, Frédo Durand, and William T Freeman. Noise-optimal capture for high dynamic range photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 553-560, 2010. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[7] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 6, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[8] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (TOG), 36(4):144:1-144:12, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[9] Nima Khademi Kalantari, Eli Shechtman, Connelly Barnes, Soheil Darabi, Dan B. Goldman, and Pradeep Sen. Patch-based high dynamic range video. ACM Transactions on Graphics (TOG), 32(6):202:1-202:10, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. SAFNet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.284, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[12] Jieyu Li, Ruiwen Zhen, and Robert L. Stevenson. A lightweight exposure bracketing strategy for HDR imaging without access to camera RAW. *Electronic Imaging*, 2023 (15):116-1-116-6, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[13] Shuaizheng Liu, Xindong Zhang, Lingchen Sun, Zhetong Liang, Hui Zeng, and Lei Zhang. Joint HDR denoising and fusion: A real-world mobile HDR image dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13966-13975, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[14] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[15] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[16] Rafal Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM Transactions on Graphics (TOG), 30(4):1-14, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[17] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion: A simple and practical alternative to high dynamic range photography. Computer Graphics Forum, 28(1):161-171, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[18] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. arXiv preprint:1602.01783, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[19] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson W. H. Lau. HDR-GAN: HDR image reconstruction from multi-exposed LDR images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[20] Junting Pan, Cristian Canton Ferrer, Kevin McGuinness, Noel E. O'Connor, Jordi Torres, Elisa Sayrol, and Xavier Giro-i Nieto. SalGAN: Visual saliency prediction with generative adversarial networks. arXiv preprint:1701.01081, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.776 + ], + "angle": 0, + "content": "[21] Reza Pourreza-Shahri and Nasser Kehtarnavaz. Exposure bracketing via automatic exposure selection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 320-323, 2015. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[22] Martin L. Puterman. Markov decision processes. In *Handbooks in Operations Research and Management Science*, chapter 8, pages 331-434. North-Holland, Amsterdam, 1990. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. TransMEF: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2126-2134, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[24] Kalpana Seshadrinathan, Sung Hee Park, and Oscar Nestares. Noise and dynamic range optimal computational imaging. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 2785-2788, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[25] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5572-5581, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[26] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 1, 2, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[28] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. BANet: A blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 2, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[30] Peter van Beek. Improved image selection for stack-based HDR imaging. *Electronic Imaging*, 2019(4):581-1-581-6, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[31] Channarayapatna Shivaram Vijay, Chandramouli Paramanand, Ambasamudram Narayanan Rajagopalan, and Rama Chellappa. Non-uniform deblurring in HDR image reconstruction. IEEE Transactions on Image Processing, 22 (10):3739-3750, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[32] Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, and Jimmy S. Ren. Learning a reinforced agent for flexible exposure bracketing selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1820-1828, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[33] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[34] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. HDRFlow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[35] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[37] Runsheng Yu, Wenyu Liu, Yasen Zhang, Zhi Qu, Deli Zhao, and Bo Zhang. DeepExposure: Learning to expose photos with asynchronously reinforced adversarial learning. In Advances in Neural Information Processing Systems 31, pages 8377-8386, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[38] Lu Yuan, Jian Sun, Long Quan, and Heung-Yeung Shum. Image deblurring with blurred/noisy image pairs. ACM Transactions on Graphics (TOG), 26(3):1:1-1:10, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[39] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Björn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[40] Zhilu Zhang, Haoyu Wang, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised high dynamic range imaging with multi-exposure images in dynamic scenes. arXiv preprint:2310.01840, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25185" + } + ] +] \ No newline at end of file diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_origin.pdf b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dae6a21ca0100641ea0a1a9f0279870f1c838c7e --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/facb16c1-8dd7-4d43-a13f-75a64353a2a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3faaa3b5b6e57d4ad21e6e255076939504f10f518a196fd3ea9fef0f2ab2d3e1 +size 9730385 diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/full.md b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8bc876d0e30a1a97821b6c1a20d8bdaa0fe0e4b3 --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/full.md @@ -0,0 +1,362 @@ +# AdaptiveAE: An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes + +Tianyi Xu $^{1,3,4*}$ Fan Zhang $^{1}$ Boxin Shi $^{3,4\dagger}$ Tianfan Xue $^{2,1\dagger}$ Yujin Wang $^{1\dagger}$ $^{1}$ Shanghai AI Laboratory $^{2}$ The Chinese University of Hong Kong + +$^{3}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University +$^{4}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +photon@stu.pku.edu.cn, zhangfan@pjlab.org.cn, shiboxin@pu.edu.cn + +tfxue@ie.cuhk.edu.hk, wangyujin@pjlab.org.cn + +![](images/57f3c8d85b7a63c3a2fbc6de4a21d338865288976d58cc1ae4ab9ff40d79a3e2.jpg) +Ours-stage I + +![](images/7320d55c71ddaab53b4bbc6ceb1939e3dd086d50357da90a8152ae7a1df51e33.jpg) + +![](images/d2886d4a1d6d4777176cda51e17973d40068c6e830cd0b441092666ed009a261.jpg) + +![](images/56ec4291011bcb79a4ca3b1f902221cc986364ec40153fea96aa4141f92991f8.jpg) +Pourreza et al. + +![](images/2bdae9d0326c6336d7e0abab175dad8fcd1c359cc7ae8d0a033b13992973278a.jpg) +Figure 1. AdaptiveAE takes camera preview images as input and automatically predicts the ISO and shutter speed for each LDR captures for exposure fusion through a 3-stage sequential refinement procedure to achieve an optimal balance between noise level and motion-related problems for high quality HDR capturing in dynamic scenes with deep reinforcement learning. AdaptiveAE achieves PSNR 39.7 on HDR dataset [26], while baseline methods [6, 21, 32] that either only predicts shutter speed or do not consider motion can only achieve PSNR below 37.6 and has evident motion blur and ghosting artifacts in HDR results. + +![](images/9bc6a12b1ee7ce7ba9b63901fca253196c0b7345bc72b85851d39b5ea9e4ad93.jpg) + +![](images/24ac4af5b61e35b75f58a8413d4b3b3ac2b009d82fca3a44ae6ab797abf83485.jpg) +Ours-stage II + +![](images/f9ca708596d774d0fee0494d5a772d970d6737b486135e1e9a5f1953f0623985.jpg) + +![](images/4f1e9ded7fc2e2f587f1ac5a0029376077738f1913e8a3787b025b3a8217bcd0.jpg) + +![](images/d5786f0da83e045f29d5458f464df7b65a4a6475864afedcd792a41098065600.jpg) +Wang et al. + +![](images/ea6d2f0a457b0594deb06d35feea7700f2991fd802355eb09cd642cde415849e.jpg) + +![](images/eeb940db43cbc39522bc5f7706bedbbd72fc8ba16daef7f505747fa6233c53a6.jpg) + +![](images/1683a3d83a049a09236d36091e3e3df2a9ba1a3e4cfe032df20d396250c7954a.jpg) +Ours-stage III + +![](images/667677b13404260285ccecab2d683a89ff51a2737f805fdbb8f411bcd326418a.jpg) + +![](images/57fd6fba537d893b39e3c90ba0fa00d528f88ac04e54e2afc7fc81f0ce5bf920.jpg) + +![](images/20b2781409bd4a9565ee62119418fffe646f4b2d365a478af1fffae848ef2e8c.jpg) +Hasinoff et al. + +![](images/d11aebd9323ae48813cbc7cd114dc47a85b2ac92049c785ae2782ddba9d400b4.jpg) + +![](images/fb291a2e952040657bcb9d209cbe247a66a0846af8700c5f0af7d498013b017c.jpg) + +![](images/424effc145ab0444c1bb6a282fe56b3522b5122336f6055248b0ad4fc7ef21ad.jpg) + +![](images/f2c83c59ef5ff751ce19bf48dad459f0ce71b3086d490ddaebc6d9e9ace42c8b.jpg) + +![](images/1a37337102f97533ed0e0cfa3318249836f2fcb80b9c7c84e6a5bb43f7ea5bd3.jpg) +Deep Learning Based + +![](images/9c4db17d59b56dd50e25d451caaaf17be993c3382138924ca0dc9b09542a434c.jpg) +Shutter Speed + +![](images/fa947fc2e4b90cd3dd695852f772b1a5e764057e22e59b5a82bea74be07c6ae4.jpg) +ISO + +![](images/e266a634297ed36372e3b0bc8a42db9165bae7d04c60382110604934bb274d38.jpg) +Adjustable / Unadjustable + +# Abstract + +Mainstream high dynamic range imaging techniques typically rely on fusing multiple images captured with different exposure setups (shutter speed and ISO). A good balance between shutter speed and ISO is crucial for achieving high-quality HDR, as high ISO values introduce significant noise, while long shutter speeds can lead to noticeable motion blur. However, existing methods often overlook the complex interaction between shutter speed and ISO and fail to account for motion blur effects in dynamic scenes. + +In this work, we propose AdaptiveAE, a reinforcement + +learning-based method that optimizes the selection of shutter speed and ISO combinations to maximize HDR reconstruction quality in dynamic environments. AdaptiveAE integrates an image synthesis pipeline that incorporates motion blur and noise simulation into our training procedure, leveraging semantic information and exposure histograms. It can adaptively select optimal ISO and shutter speed sequences based on a user-defined exposure time budget, and find a better exposure schedule than traditional solutions. Experimental results across multiple datasets demonstrate that it achieves the state-of-the-art performance. + +# 1. Introduction + +High-dynamic-range (HDR) imaging plays a pivotal role in computational photography. To capture an HDR scene, + +due to hardware limits, a single capture can only cover a Low Dynamic Range (LDR), and HDR fusion techniques are proposed to combine multiple LDR images with varying exposures to cover a wide dynamic range [7, 10, 14, 15, 17, 19, 23, 35, 36, 40]. The typical way to vary exposure is to change either shutter speed or ISO. This is a challenging process, as a longer shutter speed may increase the signal-to-noise ratio but introduce unpleasant motion blur, a large ISO may increase brightness but also magnify noise, and significant exposure differences can cover a wider dynamic range but also increase the risk of misalignment. Therefore, the choice of exposure values (EVs) for each capture is critical in this process to ensure high-quality HDR results. + +Still, very limited work discusses how to choose the optimal exposure levels, particularly in dynamic scenes. Prior works on exposure scheduling mainly focus on static scenes and ignores potential motion blur. Learning-based techniques overlook the intricate interplay between ISO and shutter speed, resulting in suboptimal image quality under varying conditions [6, 21, 32]. Additionally, many existing methods treat ghosting and motion blur as separate, computationally intensive post-processing tasks, which is unsuitable for real-time applications [11, 25, 27, 29, 31, 38, 39]. + +In this work, we propose AdaptiveAE, an efficient exposure control algorithm designed for HDR capturing, which addresses both motion blur and noise during image acquisition. Given the previously captured image, our method optimizes the exposure bracketing strategy for the subsequent capture, based on illumination information and semantic data from previous frames. Unlike previous approaches that treat motion blur as a separate post-processing task, we aim to address it during the capture process. + +Designing both efficient and adaptive exposure control is non-trivial, and we resort to reinforcement learning [18] to solve this challenge. Our solution mimics an experienced photographer. At each iteration, the policy network takes the previously captured LDR images as input, together with extracted semantic and illumination information. Given this information, the policy network learns to determine the optimal exposure setup for the following captures, which maximizes the additional information provided by this frame while also reducing the risk of misalignment and motion blur in a dynamic scene. Once the next frame is captured, this newly captured image will be used as input for the subsequent refinement iteration. As shown in Fig. 1 right, the final quality (PSNR) of fusion increases as more images are captured. + +Our proposed approach offers several advantages over traditional exposure controls. First, AdaptiveAE controls both exposure time and ISO and also adapts to different scenes. As a result, it has a much higher upper bound compared to either a fixed exposure schedule or an adaptive control algorithm that only changes exposure time. As shown + +in Fig. 1, our method iteratively achieves an optimal balance between noise and blur, resulting in less noise, reduced motion blur, and superior image quality compared to other baseline methods. Second, our method can handle both static and dynamic scenes. In static scenes, it achieves performance comparable to state-of-the-art techniques, and in dynamic ones, it produces visually compelling HDR images with minimal motion blur and ghosting artifacts. Third, our method can automatically choose the best number of frames for HDR imaging. Unlike traditional HDR approaches that often use a fixed number of frames, such as three, our approach provides flexibility to determine whether capturing three or more frames is optimal for certain scenes, balancing image quality and time budget. + +Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach enables the concurrent analysis of blur and noise in AE prediction, thereby enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately. + +We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Additionally, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and underscore the critical importance of incorporating blur synthesis into our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II, demonstrates superior noise control and effectively reduces motion blur, outperforming baselines in the visual quality of the fused HDR images. + +Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach allows for concurrent analysis of blur and noise in AE prediction, enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately. + +We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Also, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and highlight the critical importance of incorporating blur synthesis within our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II, + +demonstrates superior noise control and effectively reduces motion blur, outperforming compared baselines in the visual quality of the fused HDR images. + +# 2. Related work + +Strategy for exposure bracketing. Determining the optimal set of exposures for multiple-exposure dynamic range imaging is a well-established problem. Most digital cameras allow users to set the compensation ratio for exposure bracketing, while mobile cameras typically impose fixed ratios during automatic exposure bracketing. Heuristic strategies based on the histogram are proposed by [5, 21, 30], to balance single-to-noise ratio (SNR) and saturation. The work by [6] firstly formulates this challenge as a constrained optimization problem in the linear RGB domain, addressing the scenario that follows multiple exposure fusion and precedes tone mapping and denoising. This concise formulation facilitates the use of a straightforward numerical solver. Further extensions of this formulation consider the alignment of multiple input images to account for handshake [24]. In the context of structured light 3D reconstruction, various exposures are likewise treated as an optimization problem [4]. Exposure influences not only the noise but also, to some extent, the tone of the image in the standard RGB domain. Evaluating the final image quality complicates the problem further, as subsequent tone mapping or retouching can significantly alter image quality. Therefore, a neural network is utilized to estimate exposures and fuse multiple exposed images to achieve optimal fidelity in the gamma-corrected domain [12]. Additionally, reinforcement learning is employed by [37] and [32] to assess the rewards on comprehensive image quality after more sophisticated tone adjustment. + +Challenge in dynamic scenes. In static scenes, increasing under-exposures can reduce saturation, and a longer shutter speed improves signal-to-noise ratio in dark areas. However, in dynamic scenes, excessive exposure may cause ghosting artifacts, and a prolonged shutter speed can lead to motion blur. Consequently, heuristic exposure bracketing is usually limited to two or three EV settings [5, 30]. The method by [6] imposes an upper limit on total shutter speed, while [24] addresses handshake motion through image registration but scarcely tackles local object motion. Most approaches do not fully address motion in dynamic scenes, leaving motion blur and ghosting artifacts to be addressed through post-processing. + +# 3. Method + +Conventional weighted linear combination methods, such as exposure fusion, provide straightforward SNR estimation but fail in dynamic scenes where moving objects create unquantifiable motion blur and ghosting artifacts due to misalignment. Our approach addresses these challenges dur + +![](images/421b9851af850fb7339a2f6eb81a66eed6ccda7a3b310a3c64c6411c95561657.jpg) +(a) Our overall image synthesis pipeline. + +![](images/4d5227d29cef88f714dd43c985a785ac119f2bd0cbd7814b458e7ea00f089ca0.jpg) +(b) Our blur synthesis pipeline. + +![](images/a616eb802b232413d45415954c7e9bc8ec5cc90dc0eb01e5ea4715d8a2852c0c.jpg) +(c) Our noise synthesis pipeline. +Figure 2. Our blur-aware data synthesis pipeline. + +ing capturing rather than in post-processing, as the latter is shown to yield suboptimal results. We predict exposure-related risks—such as motion blur, ghosting, noise, and saturation—based on a limited number of previously captured frames. Additionally, we employ a sequential strategy for exposure and ISO parameter determination, rather than simultaneously predicting settings for all three LDR images, which reflects the iterative nature of auto-exposure in mobile cameras that enables adaptation to significant brightness transitions. + +# 3.1.Blur-aware data synthesis pipeline + +To simulate capturing in real environments, we designed an image synthesis pipeline to generate realistic motion blur and noise in LDR images from HDR videos in the training dataset, for use in training. Typical exposure settings involve adjusting the exposure value (EV), which is calculated as: + +$$ +\mathrm {E V} = \log_ {2} \left(\frac {F ^ {2}}{T} \times \frac {1 0 0}{\mathrm {I S O}}\right), \tag {1} +$$ + +where $F$ denotes the aperture's f-number, ISO represents the ISO sensitivity and $T$ is the exposure time in seconds. + +Similar to recent methods [6, 21, 32] concerning high dynamic range capture, we assume that aperture and focus are held constant to prevent changes in defocus. This leaves just two camera settings to manipulate: (1) Shutter speed, which controls the amount of light to collect, and (2) ISO, which determines the sensor gain. + +Our pipeline synthesizes motion blur and noise for the ground truth static HDR image based on a specified ISO + +![](images/9900aad84df9c2809d531a94055d7112d27cbb6801c39951e924b0b80d54dc8c.jpg) +Figure 3. The training scheme of AdaptiveAE. States are defined as the three LDRs synthesized using predicted ISOs and shutter speeds. Starting from $s_0$ where the three LDRs has EV $\{-2,0, + 2\}$ with arbitrary EV 0 baseline, ISOs and shutter speeds, the agent sequentially predicts, customizes or inherits capturing parameters (i.e. ISO and shutter speed) for the next stage and synthesize the corresponding LDR using our image synthesis pipeline. Unlike training, the LDRs will be captured rather than synthesized during inference. + +and shutter speed. As depicted in Fig. 2, it uses two consecutive static HDR images as input. Motion blur is first synthesized according to shutter speed to produce a blurred HDR image in linear space. Next, noise is added based on shutter speed and ISO to create an LDR image, reflecting our exposure choice. Note that motion blur should be applied before adding noise, as it influences the raw input by affecting the number and pattern of captured photons during photography. + +Synthesizing blur. On-the-shelf training datasets consist of consecutive HDR ground truth of a scene with motion. Now we explain how we simulate motion blur to a frame of HDR $f_{i}^{L}$ , as shown in Fig. 2b, which is the $i$ -th HDR ground truth frame in the dataset scene, and the superscript $L$ indicates it is in linear space. We first use $\mu$ -law tone-mapping with $\mu = 5000$ to transfer $f_{i}^{L}$ and $f_{i+1}^{L}$ from HDR space to LDR space where the image interpolation algorithm we applied is trained upon, receiving $f_{i}^{\mathcal{T}}$ and $f_{i+1}^{\mathcal{T}}$ , where the superscript $\mathcal{T}$ denotes they are in LDR space. Then we use RIFE [7] to interpolate the them to 256 frames and get the sequence of images $\{f_{i}^{\mathcal{T}}, s_{1}^{\mathcal{T}}, s_{2}^{\mathcal{T}}, \dots, s_{254}^{\mathcal{T}}, f_{i+1}^{\mathcal{T}}\}$ . Then for the selected shutter speed $T_{j}$ for the $j$ -th LDR $l_{j}^{\mathcal{T}}$ to take, the blurred HDR $b_{j}^{L}$ is simulated as: + +$$ +b _ {j} ^ {L} = \mathbf {i T M O} \left(\frac {f _ {i} ^ {\mathcal {T}} + \sum_ {m = 1} ^ {m _ {j}} s _ {m} ^ {\mathcal {T}}}{m _ {j}}\right), m _ {j} = \left\lceil \frac {2 5 6 T _ {j}}{\Delta \tau} \right\rceil , \tag {2} +$$ + +where $\Delta \tau$ denotes the time elapsed from $f_{i}^{L}$ is taken to $f_{i + 1}^{L}$ is not yet taken and iTMO indicates the inverse $\mu$ -law tonemapping function with $\mu = 5000$ . + +Synthesizing noise. We adopt the noise model mentioned in [6], in which noise is modeled as a zero-mean variable, coming from three independent sources, including photon noise, which represents the Poisson distribution of photon arrivals and depends linearly upon the number of recorded + +electrons, $\Phi T$ , readout noise, which comes from sensor readout, and analog-to-digital conversion(ADC) noise, which comes from the combined effect of amplifier and quantization. Hence, for pixels below the saturation level: + +$$ +V a r (n) = \frac {\Phi T \times \mathrm {I S O} ^ {2}}{U ^ {2}} + \frac {\sigma_ {\text {r e a d}} ^ {2} \times \mathrm {I S O} ^ {2}}{U ^ {2}} + \sigma_ {\mathrm {A D C}} ^ {2}, \tag {3} +$$ + +where $\Phi$ is the radiance level, $T$ is the shutter speed, $U$ is a camera-dependent variable. + +Following our noise model, as shown in Fig. 2c, we can synthesize the corresponding noise with selected ISO and shutter speed to the blurred HDR $b_{j}^{L}$ to get the LDR image $l_{j}^{\mathcal{T}}$ . For details, please refer to our supplementary material. + +# 3.2. Problem formulation of AdaptiveAE + +Given a scene, the goal of AdaptiveAE, which has access to three initial preview LDR images $\{p_j^T\}_{1,2,3}$ (i.e. underexposed, mid-exposed, overexposed) before capturing, is to find an optimal exposure setup (i.e. ISO and shutter speed) for LDR capturing, with which the fused HDR output will result in pleasing visual performance. + +Our method formulates exposure bracketing as a Markov Decision Process [22], solved via deep reinforcement learning to refine exposure parameters (ISO, shutter speed) sequentially. As illustrated in Fig. 4, the process starts from three LDRs at a default $\{-2_p,0_p, + 2_p\}$ EV spacing relative to an arbitrary reference (subscript $p$ ). The refinement then proceeds in stages, as shown in Fig. 3. + +First, the agent predicts optimal parameters for the midexposed frame, establishing a new 0-EV reference (subscript $x$ ). The side frames are then customized—their parameters are procedurally set to achieve a symmetric $\{-2_x,0_x, + 2_x\}$ EV bracket. Next, the agent refines the underexposed frame to an EV of $-y_{x}$ . The mid-exposed frame $(0_{x})$ is inherited (its parameters are reused), while + +![](images/b2f7eae99e8282802337d2f8763386e293aecbd6a5f4edae5c134f6a55f4a1e8.jpg) +Figure 4. Training pipeline of our method. The ISO and shutter speed prediction process is conceptualized as a Markov Decision Process, where a CNN-based policy network predicts the ISO and shutter speed of the next exposure sets. Concurrently, a CNN-based value network estimates the state value. We leverage our blur-aware image synthesis pipeline to synthesize the predicted LDRs and employ DeepHDR [7] to fuse the predicted LDR images, generating our HDR result and calculating the reward for the current policy. The entire system is optimized using the A3C (Asynchronous Advantage Actor-Critic) method [18]. + +the overexposed frame is customized to $+y_{x}$ to maintain symmetry, yielding $\{-y_{x}, 0_{x}, + y_{x}\}$ . Finally, the agent predicts the overexposed frame's EV as $+z_{x}$ , creating a potentially asymmetric set $\{-y_{x}, 0_{x}, + z_{x}\}$ . This sequential prediction can be extended, as our adopted fusion method [7] handles more than three LDRs. A video providing further details on this process, along with an example of extended exposure bracketing, is available in the supplementary materials. + +# 3.3. Optimization objectives + +Let us denote the problem as $P = (S, A)$ , where S is a state space and A is an action space. Specifically, in our task, $S$ is the space of the exposure setups (i.e., ISO and shutter speed) for an LDR set that typically contains three LDRs (i.e. under-exposed, mid-exposed, and overexposed), while $A$ is the set of all possible ISO and shutter speed combinations, which is discrete. During training, at stage $s_j = \{(\mathrm{ISO}_{j1}, T_{j1}), (\mathrm{ISO}_{j2}, T_{j2}), (\mathrm{ISO}_{j3}, T_{j3}), \}$ , we first find the corresponding HDR ground truth image pairs $(f_i^L, f_{i+1}^L)_{1,2,3}$ and generate the corresponding LDRs $\{l_j^T\}_{1,2,3}$ through our image synthesis pipeline in Fig. 2. The footnote 1, 2, 3 denotes that the same operation is done for the under-, mid-, and over-exposed LDRs. Taking $\{l_j^T\}_{1,2,3}$ as input, the agent predicts an action $a_j = (\mathrm{ISO}_j, T_j)$ , which is expanded to an exposure setup for three LDRs by customizing EV or inheriting from state $s_{j-1}$ , mapping state $s_j$ to state $s_{j+1}$ . Adding a sequence of $M$ LDRs to exposure bracketing corresponds to a trajectory $\tau$ of states and actions: + +$$ +\tau = \left(s _ {0}, a _ {0}, \dots , s _ {M - 1}, a _ {M - 1}, s _ {M}\right), \tag {4} +$$ + +where $s_M$ is the stopping state. Our goal is to find a policy that maximizes the accumulated reward during the decision-making process. In this paper, the reward function with the $j$ -th action (i.e., corresponding to deciding the exposure setting for the $j$ -th LDR for exposure fusion) is thus written as: + +$$ +r \left(s _ {j}, a _ {j}\right) = \mathcal {R} \left(s _ {j + 1}\right) - \mathcal {R} \left(s _ {j}\right) - \mathcal {P} (j), \tag {5} +$$ + +where $s_{j + 1} = p(s_j,a_j)$ , and $\mathcal{R}$ denotes our reward design and $\mathcal{P}$ denotes the $L_{\mathrm{step}}$ penalty, detailed in 3.4. + +As depicted in Fig. 4, our model comprises a policy network and a value network, both of which utilize a CNN-based architecture. The policy network predicts the optimal ISO and shutter speed for the subsequent exposure, outputting a distribution of action probabilities $\pi(s,\theta)$ for an input image $s$ . Concurrently, the value network $V^{\pi}(s,\omega)$ estimates the corresponding state value. These networks, with combined parameters $\psi = (\theta,\omega)$ , are trained by maximizing our objective $J(\theta)_{\psi}$ to learn the optimal policy $\pi(s)$ . Specifically, to train the policy network and the value network, we apply the A3C (Asynchronous Advantage Actor-Critic) method [18], where the actor is represented by the policy network and the critic is the value network. Network details are in the supplementary materials. + +# 3.4. Reward + +When designing the reward function for our system, we considered four key factors: (1) similarity between the fused HDR and the ground truth HDR; (2) quality of important regions in the fused HDR; (3) quality of moving regions in the fused HDR; and (4) a penalty for overly long LDR stacks. Thus, our reward function is: + +$$ +\mathcal {R} = - \left(P _ {\text {c o n s t r u c t i o n}} + P _ {\text {p r i o r i t y}} + P _ {\text {g h o s t}}\right). \tag {6} +$$ + +We consider $P_{\text {construction }}$ , the L2 loss between our fused HDR and the ground truth, as our major reward component, which is affected by noise and saturation. Note that through the entire sequential decision process, the middle-exposed frame is used as a reference for HDR fusion. Conforming to Eq. (3), during training, noise is synthesized according to the irradiance of ground truth HDR, and during inference, noise is estimated with the irradiance of noisy signals. + +$P_{\mathrm{priority}}$ represents an L2 loss within the areas in the image masked by an importance mask, which is generated by a saliency predictor [20]. This ensures that the highest quality is maintained in the most significant areas, thereby enhancing the overall visual fidelity where it matters most. + +$P_{\mathrm{ghost}}$ is also an L2 loss within a masked area, denoting areas with large motions and thus having a higher risk of motion blur- or ghosting-caused HDR quality degradation. The mask is computed by calculating the optical flow using RAFT [28] between the HDR ground truth of the middle-exposed frame (i.e., the reference frame for fusion) and the corresponding HDR $f_{i}^{L}$ and selecting the pixels where the mode of the flow vector exceeds a constant threshold $K$ . Normalizing the largest optical flow vector, we empirically set $K$ to 0.2. $P_{\mathrm{ghost}}$ guides the agent to deal more carefully with regions that are prone to artifacts caused by motion and is helpful for high-quality HDR capturing, as is verified by the result of our ablation studies in Tab. 2. + +$\mathcal{P}(j)$ is a penalty designed to penalize excessively long exposure brackets. Capturing an excessive number of shots, such as 10, even with random exposure settings, can lead to nearly perfect HDR fusion results, but it is time-consuming. Typically, three shots [2, 6, 21] are sufficient to achieve high-quality outcomes. To this end, we incorporate a penalty for taking more than three shots, as follows: + +$$ +\mathcal {P} (j) = \left\{ \begin{array}{l l} 0 & \text {i f} j \leq H \\ \alpha (j - H) ^ {2} & \text {i f} j > H \end{array} , \right. \tag {7} +$$ + +where $\alpha$ is a positive coefficient and $H$ is set to 3. + +In this manner, the autonomous agent optimizes exposure parameters by predicting relatively fast shutter speeds for LDR images, particularly for the middle-exposed reference frame, thereby minimizing motion blur while avoiding excessive ISO values that would introduce noise-related degradation. When confronted with potential ghosting artifacts—which emerge from information deficiency in LDR images due to concurrent saturation and motion—the agent adaptively selects EVs that minimize both underexposure and saturation, resulting in a significant reduction of ghosting artifacts in the final reconstruction. + +# 4. Experiments + +Experiment details. We use Real-HDRV [26] for training and tested our performance on Real-HDRV [26] and DeepHDRVideo [3]. For HDR fusion, we adopt DeepHDR [33] to generate the HDR image based on the selected exposure bracketing. Additionally, we set the number of LDR frames to 3 for all methods involved. For methods that do not account for changes in ISO values, we set the ISO to 200 as a standard value for most cases, which is rationalized in our supplementary materials. Since RIFE [7] is relatively time-consuming, the image interpolation and blur synthesis step + +is performed before training. We apply random flipping, rotation, and cropping with $512 \times 512$ pixels for data augmentation. Our training dataset consists of a total of 770 scenes, including 440 dynamic and 330 static scenes. + +Evaluation metrics. We directly evaluate the performance of the fused HDR results. Similar to previous HDR fusion methods [7, 14, 19], we employ PSNR- $\mu$ , SSIM- $\mu$ , PU-PSNR, PU-SSIM, and HDR-VDP-2 [16] as evaluation metrics. PSNR- $\mu$ and SSIM- $\mu$ denote PSNR and SSIM of the fused HDR after $\mu$ -law tone-mapping with $\mu = 5000$ . PU-PSNR and PU-SSIM are computed after perceptually uniform encoding [1]. When computing the HDR-VDP-2 [16], the diagonal display size is 30 inches. + +# 4.1. Results + +Results on Real-HDRV dataset. We compared our trained agent's performance with several state-of-the-art HDR exposure bracketing methods, including Pourreza-Shahri et al. [21], Hasinoff et al. [6], and Wang et al. [32]. The first two are non-deep-learning methods that do not consider motion: Pourreza-Shahri et al. use K-means clustering to adjust shutter speed based on image brightness, while Hasinoff et al. mathematically optimize ISO and shutter speed for the best worst-case SNR. Wang et al. utilize reinforcement learning to predict shutter speed for maximizing PSNR, but also ignore motion. + +Our method achieves state-of-the-art performance on the HDRV-Test dataset, as shown in Tab. 1. By treating ISO as a variable, it provides flexibility in handling extremely dark scenes. Additionally, it excels in dynamic scenes due to a blur synthesis model and carefully designed rewards. Thus, our model effectively balances noise reduction and motion artifact minimization, delivering high-quality HDR results. Visualization results in Fig. 5 show that deep learning-based exposure fusion models affect EV selection differently. In Hasinoff et al.'s setup, three LDRs are equally weighted, but dynamic scenes require a reference image, emphasizing the middle-exposed LDR's quality. This often shifts its EV toward under-exposure to reduce motion blur. Our experiments confirm that if the middle-exposed image is blurry—a common problem in other methods—the fused HDR will also be blurred. + +Inference time. Our method takes less than $250\mathrm{ms}$ for each HDR image, which is acceptable. With average exposure time $n$ ( $\leq 30ms$ ) and prediction time $m$ ( $\leq 10ms$ ), the total execution time is $6n + 3m$ ( $\leq 250ms$ ). This can be further lowered to $6n$ ( $\leq 200ms$ ) adopting concurrency. Note that DeepHDR is only required during training for calculating rewards, and on-camera inference can be performed with a small agent, taking only $3.5ms$ for each frame without optimization. Besides, our method can be further accelerated by on-camera processors when applied to DSLRs. More discussions are in our supplementary materials. + +![](images/d3de9365d8e7bccc28e520fe57a679c3167bff0481545594d08799c11e01c78c.jpg) +LDR Images +Figure 5. Qualitative comparisons with other auto-exposure methods on HDRV dataset [26]. Left: Predicted LDRs with varying ISO and shutter speed settings and synthesized using our image synthesis pipeline. Middle: Fused HDR image using DeepHDR [7] and tonemapined using Photomatix. Right: Zoom-in results for tested methods. + +![](images/ff5409d764a92d23a08ab76c626a512db868d36c4263fde7651fe6f9fedffb98.jpg) +Our Tonemapped HDR Image + +![](images/489655b8709e24eaf98b598deef205f42c532ce5f3c062feee6528dfc86214e7.jpg) +Wang et al. Pourreza et al. Hasinoff et al. +Ours +Ground Truth + +Table 1. Comparison of Different Methods. We utilized one preview image and three preview images, respectively, for the compared methods and our method to predict exposure settings for three LDRs. We leveraged DeepHDR [7] for exposure fusion and used the mentioned metrics to evaluate the quality of the fused HDR. Bold: The best. + +
MethodsHDRV [26]DeepHDRVideo [3]
PSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIMPSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIM
Pourreza et al. [21]33.640.861754.5530.610.867935.570.878055.6731.590.8791
Hasinoff et al. [6]37.590.905257.0232.870.898038.470.915758.6534.450.9132
Wang et al. [32]36.460.890256.0932.680.893337.950.901957.3933.270.9008
Ours39.700.940859.2034.670.946539.810.937158.9036.190.9338
+ +Table 2. Ablation study of AdaptiveAE on HDRV [26] dataset. Base denotes our model trained with only the step penalty and construction reward. Bold: The best. + +
ModelPSNR-μSSIM-μPU-PSNRPU-SSIM
Base38.210.922732.680.9198
Base+Ppriority38.570.926133.020.9239
Base+Ppriority+Pghost39.700.940834.670.9465
+ +Gap to the best-achievable. For each scene in the test set of Real-HDRV, we iteratively search for the best set of predictions by Gaussian sampling around our initial prediction (50 times per exposure parameter per frame, with a deviation of $20\%$ of the mean for both ISO and shutter speed). Statistics in Tab. 4 show that our method approaches the locally optimal result while being efficient. + +Cross datasets test. To test the generalization ability of our model, we also evaluated the performance of our trained agent on DeepHDRVideo [3], as shown in the right of Tab. 1. Our agent exhibits good generalization abilities. + +Table 3. Comparison of different exposure fusion methods. We utilize the four auto-exposure methods to predict exposure settings and employ our image synthesis pipeline to create three LDRs. Then, we apply different exposure fusion methods to fuse them and compare the HDR quality on the HDRV dataset [26]. HDRTransformer [14] is pretrained on HDRV [26] dataset. P: PSNR- $\mu$ S: SSIM- $\mu$ . Pou: Pourreza et al. [21], Has: Hasinoff et al. [6], W: Wang et al. [32]. HDR-Trans: HDR-Transformer. Bold: The best. + +
ModelDeepHDRHDR-GANHDR-Trans
P-μS-μP-μS-μP-μS-μ
Pou33.640.861735.710.889235.840.8824
Has37.590.925238.580.926339.110.9372
W36.460.900237.950.916938.890.9210
Ours39.700.940840.730.937641.370.9478
+ +Since the DeepHDRVideo dataset only provides ground truth HDR for the middle image in the sequence, we synthesize HDR for the other images using DeepHDR [33]. + +Cross HDR fusion methods test. In our training scheme, + +Table 4. Performance comparison investigating the gap between our method and the optimum on the HDRV-test dataset. + +
OursWorstAverageBest
PSNR-μ39.7025.7632.4139.93
SSIM-μ0.94080.77380.86090.9412
+ +![](images/7e2a8baf7457b8550cb3c40fa02c2876435cb8eb97581dfd60240a5bc4e405e5.jpg) +Wang et al. +Figure 6. Results of our real capture data. The subject performs steady and repetitive movements, taking shots according to the exposure settings predicted by various methods. + +![](images/88ea1c1e3281f4498568455941fb3e88f7022e81d5092aef7f902a778a642e35.jpg) +Hasinoff et al. + +![](images/2e3290c7c7e4426deb4314c3ff183fdb66d0c9015fceacff00c9a0891c2ed8bf.jpg) +Ours + +we used DeepHDR for exposure fusion. For further testing, we adopted different exposure methods, including HDR-GAN [19] and HDR-Transformer [14] (pre-trained on the HDRV-dataset [26]), as post-processing exposure fusion methods. All tests are conducted on the HDRV-Test dataset. As shown in Tab. 3, without considering motion before LDR capturing, however powerful an exposure fusion method (i.e. HDR-GAN and HDR-Transformer) fails to yield a satisfactory result. Notably, when stronger fusion models are used, the performance gap between our model and traditional models, which do not account for motion, increases. This is because, when blur and ghosting risk are mitigated, post-processing models can effectively manage the remaining challenges. In contrast, ignoring motion makes these challenges more difficult for post-processing models, leading to potential failure. + +Results for real capture. We used a SONY Alpha 7C-II to evaluate our model on real-world scenes. Subjects performed steady, repetitive movements while ISO and shutter speed were manually set for each capture, with the aperture fixed at f/2.8. The camera and subject positions remained unchanged during multi-exposure captures. As shown in Fig. 6, our method provides noise control comparable to other baselines while effectively mitigating motion blur, which can impair the visual quality of the fused HDR image; for more results, see our supplementary materials. + +# 4.2. Ablation studies + +Effectiveness of penalty item. We conduct ablation studies to validate the effectiveness of our reward design. All quantitative evaluations are conducted on the HDRV-Test dataset. We train our networks using only the construction + +![](images/2c6298add16b1a6e25f9232f8c0a6775c18860039f33cab00fc887f63509e7e9.jpg) + +![](images/8476f314e935179366fe64aeb1230113e43f4749748421b5d6db7187ddb9b9eb.jpg) +Under-exposed + +![](images/0c97e927027381d99e123cb7e88b6ebad07e02e3b20507659125b7546a55a5eb.jpg) + +![](images/cadbdc4c20577aae27ce1dd7cc0d1f84670f00b2d51c33269d4fb84ac7320e0d.jpg) +Middle-exposed + +![](images/388664262c000dcf514fbaa58a3a7bf53d35ef51626fe922f26614cd52a12fad.jpg) + +![](images/b4f9563ac7c766b35eb5e362f1d6561a514e6b3010ca5966c0c2950c0da47e24.jpg) +Over-exposed + +![](images/dc7c33839d6b7b2edd687c4a7c39c108f56aa8aa86578565109d16e9bf7ba057.jpg) +Figure 7. Left: Zoom-in details of the moving regions and our predicted ISO and shutter speed (seconds) for each LDR. Up: LDRs for scenes with an average motion level of around 30 pixels, Down: around 60 pixels. Right: Comparisons with baseline methods across different motion magnitude ranges. Tested on Real-HDRV [26] dataset. + +reward and step penalty as the base model and test the effectiveness of $P_{\mathrm{ghost}}$ and $P_{\mathrm{priority}}$ . As shown in Tab. 2, these two reward elements enhance our agent's performance by directing the model to focus on semantically important regions and moving areas. The former typically corresponds to the scene's focal point (e.g. a person's face), while the latter involves regions at high risk of motion blur and ghosting artifacts. Qualitative evidence is provided in our supplementary materials. + +Robustness to dynamic scenes. We offer a pair of examples from the Real-HDRV [26] dataset with around 30 and 60 pixels of regional movement, respectively. As shown in Fig. 7: Left, our trained agent tends to predict faster shutter speeds for scenes with stronger movement, validating the responsiveness of our agent in dynamic scenes. We assess our method's performance against previous auto-exposure methods [6, 32] at varying motion magnitudes (Fig. 7: Right). To create the evaluation dataset for robustness to different motion levels, we follow HDRFlow [34], using RAFT [28] to process dynamic scenes from DRV and obtain optical flow maps. We then manually crop these images with reasonable flow predictions, dividing them into $128 \times 128$ blocks, and calculate the average motion magnitude for each block. Finally, we evaluate the PSNR of blocks corresponding to different motion magnitudes. As shown in Fig. 7: Right, our AdaptiveAE demonstrates greater robustness than other methods as motion magnitude increases. + +# 5. Conclusion + +We introduce AdaptiveAE, which optimizes HDR exposure in dynamic settings using deep reinforcement learning, treating exposure bracketing as a Markov Decision Process. It autonomously adjusts ISO and shutter speed for a pretrained exposure fusion algorithm. Reward systems focus on moving and key regions, minimizing sequence lengths. Experiments show AdaptiveAE outperforms state-of-the-art methods in dynamic scenes while matching top models in static ones, allowing flexible HDR capture. Our analysis of noise models and exposure settings offers insights for future research, with plans to include adjustable apertures. + +# References + +[1] Maryam Azimi, Vedad Hulusic, Philippe Hanhart, and Touradj Ebrahimi. PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR. In Proceedings of the Picture Coding Symposium (PCS), pages 1-5, 2021. 6 +[2] Neil Barakat, A. Nicholas Hone, and Thomas E. Darcie. Minimal-bracketing sets for high-dynamic-range image capture. IEEE Transactions on Image Processing, 17(10):1864-1875, 2008. 6, 1 +[3] Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, and Lei Zhang. HDR video reconstruction: A coarse-to-fine network and a real-world benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2502-2511, 2021. 2, 6, 7, 1 +[4] Wenyuan Chen, Xingjian Liu, Shanghai Ru, and Yu Sun. Automated exposures selection for high dynamic range structured-light 3-D scanning. IEEE Transactions on Industrial Electronics, 70(7):7428-7437, 2023. 3 +[5] Natasha Gelfand, Andrew Adams, Sung Hee Park, and Kari Pulli. Multi-exposure imaging on mobile devices. In Proceedings of the 18th ACM International Conference on Multimedia, pages 823-826, 2010. 3 +[6] Samuel W Hasinoff, Frédo Durand, and William T Freeman. Noise-optimal capture for high dynamic range photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 553-560, 2010. 1, 2, 3, 4, 6, 7, 8 +[7] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 6, 7, 1, 3 +[8] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (TOG), 36(4):144:1-144:12, 2017. 2 +[9] Nima Khademi Kalantari, Eli Shechtman, Connelly Barnes, Soheil Darabi, Dan B. Goldman, and Pradeep Sen. Patch-based high dynamic range video. ACM Transactions on Graphics (TOG), 32(6):202:1-202:10, 2013. 2 +[10] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. SAFNet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of European Conference on Computer Vision (ECCV), 2024. 2 + +Acknowledgement. This work was supported by the National Key R&D Program of China No. 2022ZD0160201, Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No. 62136001, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources. +[11] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 2 +[12] Jieyu Li, Ruiwen Zhen, and Robert L. Stevenson. A lightweight exposure bracketing strategy for HDR imaging without access to camera RAW. *Electronic Imaging*, 2023 (15):116-1-116-6, 2023. 3 +[13] Shuaizheng Liu, Xindong Zhang, Lingchen Sun, Zhetong Liang, Hui Zeng, and Lei Zhang. Joint HDR denoising and fusion: A real-world mobile HDR image dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13966-13975, 2023. 2 +[14] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 6, 7, 8 +[15] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2020. 2 +[16] Rafal Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM Transactions on Graphics (TOG), 30(4):1-14, 2011. 6 +[17] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion: A simple and practical alternative to high dynamic range photography. Computer Graphics Forum, 28(1):161-171, 2009. 2 +[18] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. arXiv preprint:1602.01783, 2016. 2, 5 +[19] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson W. H. Lau. HDR-GAN: HDR image reconstruction from multi-exposed LDR images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 2, 6, 8 +[20] Junting Pan, Cristian Canton Ferrer, Kevin McGuinness, Noel E. O'Connor, Jordi Torres, Elisa Sayrol, and Xavier Giro-i Nieto. SalGAN: Visual saliency prediction with generative adversarial networks. arXiv preprint:1701.01081, 2017. 6 +[21] Reza Pourreza-Shahri and Nasser Kehtarnavaz. Exposure bracketing via automatic exposure selection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 320-323, 2015. 1, 2, 3, 6, 7 +[22] Martin L. Puterman. Markov decision processes. In *Handbooks in Operations Research and Management Science*, chapter 8, pages 331-434. North-Holland, Amsterdam, 1990. 4 +[23] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. TransMEF: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2126-2134, 2022. 2 + +[24] Kalpana Seshadrinathan, Sung Hee Park, and Oscar Nestares. Noise and dynamic range optimal computational imaging. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 2785-2788, 2012. 3 +[25] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5572-5581, 2019. 2 +[26] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 1, 2, 6, 7, 8, 3 +[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2 +[28] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 6, 8 +[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. BANet: A blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 2, 1, 3 +[30] Peter van Beek. Improved image selection for stack-based HDR imaging. *Electronic Imaging*, 2019(4):581-1-581-6, 2019. 3 +[31] Channarayapatna Shivaram Vijay, Chandramouli Paramanand, Ambasamudram Narayanan Rajagopalan, and Rama Chellappa. Non-uniform deblurring in HDR image reconstruction. IEEE Transactions on Image Processing, 22 (10):3739-3750, 2013. 2 +[32] Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, and Jimmy S. Ren. Learning a reinforced agent for flexible exposure bracketing selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1820-1828, 2020. 1, 2, 3, 6, 7, 8 +[33] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 6, 7 +[34] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. HDRFlow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 8 +[35] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2 +[36] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2 + +[37] Runsheng Yu, Wenyu Liu, Yasen Zhang, Zhi Qu, Deli Zhao, and Bo Zhang. DeepExposure: Learning to expose photos with asynchronously reinforced adversarial learning. In Advances in Neural Information Processing Systems 31, pages 8377-8386, 2018. 3 +[38] Lu Yuan, Jian Sun, Long Quan, and Heung-Yeung Shum. Image deblurring with blurred/noisy image pairs. ACM Transactions on Graphics (TOG), 26(3):1:1-1:10, 2007. 2 +[39] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Björn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2 +[40] Zhilu Zhang, Haoyu Wang, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised high dynamic range imaging with multi-exposure images in dynamic scenes. arXiv preprint:2310.01840, 2023. 2 \ No newline at end of file diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/images.zip b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..519395858ea3a96738a3563e0b1ec4027a61ea27 --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4df01caaa24b03c50cdb750c1016cfc1fec20f700cfcfe53bb6296aecf4de04f +size 661262 diff --git a/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/layout.json b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9a5210ae75fabc9ed7d558f3322eaffc72f9019a --- /dev/null +++ b/2025/AdaptiveAE_ An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes/layout.json @@ -0,0 +1,10228 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "text", + "content": "AdaptiveAE: An Adaptive Exposure Strategy for HDR Capturing in Dynamic Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": "Tianyi Xu" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1,3,4*}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": " Fan Zhang" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{3,4\\dagger}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": " Tianfan Xue" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{2,1\\dagger}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": " Yujin Wang" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 108, + 159, + 503, + 190 + ], + "type": "text", + "content": "The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 191, + 505, + 217 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 104, + 191, + 505, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 191, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 505, + 202 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 191, + 505, + 202 + ], + "type": "text", + "content": "State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 205, + 504, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 205, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 104, + 205, + 504, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 104, + 205, + 504, + 217 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 128, + 220, + 481, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 220, + 481, + 231 + ], + "spans": [ + { + "bbox": [ + 128, + 220, + 481, + 231 + ], + "type": "text", + "content": "photon@stu.pku.edu.cn, zhangfan@pjlab.org.cn, shiboxin@pu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 184, + 234, + 424, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 234, + 424, + 245 + ], + "spans": [ + { + "bbox": [ + 184, + 234, + 424, + 245 + ], + "type": "text", + "content": "tfxue@ie.cuhk.edu.hk, wangyujin@pjlab.org.cn" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 267, + 168, + 342 + ], + "blocks": [ + { + "bbox": [ + 57, + 267, + 168, + 342 + ], + "lines": [ + { + "bbox": [ + 57, + 267, + 168, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 267, + 168, + 342 + ], + "type": "image", + "image_path": "57f3c8d85b7a63c3a2fbc6de4a21d338865288976d58cc1ae4ab9ff40d79a3e2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 347, + 156, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 347, + 156, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 156, + 358 + ], + "type": "text", + "content": "Ours-stage I" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 62, + 342, + 80, + 363 + ], + "blocks": [ + { + "bbox": [ + 62, + 342, + 80, + 363 + ], + "lines": [ + { + "bbox": [ + 62, + 342, + 80, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 80, + 363 + ], + "type": "image", + "image_path": "7320d55c71ddaab53b4bbc6ceb1939e3dd086d50357da90a8152ae7a1df51e33.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 82, + 344, + 100, + 364 + ], + "blocks": [ + { + "bbox": [ + 82, + 344, + 100, + 364 + ], + "lines": [ + { + "bbox": [ + 82, + 344, + 100, + 364 + ], + "spans": [ + { + "bbox": [ + 82, + 344, + 100, + 364 + ], + "type": "image", + "image_path": "d2886d4a1d6d4777176cda51e17973d40068c6e830cd0b441092666ed009a261.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 58, + 364, + 168, + 437 + ], + "blocks": [ + { + "bbox": [ + 58, + 364, + 168, + 437 + ], + "lines": [ + { + "bbox": [ + 58, + 364, + 168, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 364, + 168, + 437 + ], + "type": "image", + "image_path": "56ec4291011bcb79a4ca3b1f902221cc986364ec40153fea96aa4141f92991f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 440, + 166, + 450 + ], + "lines": [ + { + "bbox": [ + 107, + 440, + 166, + 450 + ], + "spans": [ + { + "bbox": [ + 107, + 440, + 166, + 450 + ], + "type": "text", + "content": "Pourreza et al." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 63, + 437, + 81, + 456 + ], + "blocks": [ + { + "bbox": [ + 63, + 437, + 81, + 456 + ], + "lines": [ + { + "bbox": [ + 63, + 437, + 81, + 456 + ], + "spans": [ + { + "bbox": [ + 63, + 437, + 81, + 456 + ], + "type": "image", + "image_path": "2bdae9d0326c6336d7e0abab175dad8fcd1c359cc7ae8d0a033b13992973278a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 465, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 55, + 465, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 555, + 521 + ], + "type": "text", + "content": "Figure 1. AdaptiveAE takes camera preview images as input and automatically predicts the ISO and shutter speed for each LDR captures for exposure fusion through a 3-stage sequential refinement procedure to achieve an optimal balance between noise level and motion-related problems for high quality HDR capturing in dynamic scenes with deep reinforcement learning. AdaptiveAE achieves PSNR 39.7 on HDR dataset [26], while baseline methods [6, 21, 32] that either only predicts shutter speed or do not consider motion can only achieve PSNR below 37.6 and has evident motion blur and ghosting artifacts in HDR results." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 83, + 437, + 104, + 456 + ], + "blocks": [ + { + "bbox": [ + 83, + 437, + 104, + 456 + ], + "lines": [ + { + "bbox": [ + 83, + 437, + 104, + 456 + ], + "spans": [ + { + "bbox": [ + 83, + 437, + 104, + 456 + ], + "type": "image", + "image_path": "9bc6a12b1ee7ce7ba9b63901fca253196c0b7345bc72b85851d39b5ea9e4ad93.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 188, + 268, + 299, + 342 + ], + "blocks": [ + { + "bbox": [ + 188, + 268, + 299, + 342 + ], + "lines": [ + { + "bbox": [ + 188, + 268, + 299, + 342 + ], + "spans": [ + { + "bbox": [ + 188, + 268, + 299, + 342 + ], + "type": "image", + "image_path": "24ac4af5b61e35b75f58a8413d4b3b3ac2b009d82fca3a44ae6ab797abf83485.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 346, + 292, + 357 + ], + "lines": [ + { + "bbox": [ + 239, + 346, + 292, + 357 + ], + "spans": [ + { + "bbox": [ + 239, + 346, + 292, + 357 + ], + "type": "text", + "content": "Ours-stage II" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 195, + 343, + 212, + 363 + ], + "blocks": [ + { + "bbox": [ + 195, + 343, + 212, + 363 + ], + "lines": [ + { + "bbox": [ + 195, + 343, + 212, + 363 + ], + "spans": [ + { + "bbox": [ + 195, + 343, + 212, + 363 + ], + "type": "image", + "image_path": "f9ca708596d774d0fee0494d5a772d970d6737b486135e1e9a5f1953f0623985.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 216, + 344, + 236, + 363 + ], + "blocks": [ + { + "bbox": [ + 216, + 344, + 236, + 363 + ], + "lines": [ + { + "bbox": [ + 216, + 344, + 236, + 363 + ], + "spans": [ + { + "bbox": [ + 216, + 344, + 236, + 363 + ], + "type": "image", + "image_path": "4f1e9ded7fc2e2f587f1ac5a0029376077738f1913e8a3787b025b3a8217bcd0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 188, + 364, + 299, + 437 + ], + "blocks": [ + { + "bbox": [ + 188, + 364, + 299, + 437 + ], + "lines": [ + { + "bbox": [ + 188, + 364, + 299, + 437 + ], + "spans": [ + { + "bbox": [ + 188, + 364, + 299, + 437 + ], + "type": "image", + "image_path": "d5786f0da83e045f29d5458f464df7b65a4a6475864afedcd792a41098065600.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 440, + 285, + 451 + ], + "lines": [ + { + "bbox": [ + 238, + 440, + 285, + 451 + ], + "spans": [ + { + "bbox": [ + 238, + 440, + 285, + 451 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 194, + 437, + 212, + 456 + ], + "blocks": [ + { + "bbox": [ + 194, + 437, + 212, + 456 + ], + "lines": [ + { + "bbox": [ + 194, + 437, + 212, + 456 + ], + "spans": [ + { + "bbox": [ + 194, + 437, + 212, + 456 + ], + "type": "image", + "image_path": "ea6d2f0a457b0594deb06d35feea7700f2991fd802355eb09cd642cde415849e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 215, + 438, + 236, + 456 + ], + "blocks": [ + { + "bbox": [ + 215, + 438, + 236, + 456 + ], + "lines": [ + { + "bbox": [ + 215, + 438, + 236, + 456 + ], + "spans": [ + { + "bbox": [ + 215, + 438, + 236, + 456 + ], + "type": "image", + "image_path": "eeb940db43cbc39522bc5f7706bedbbd72fc8ba16daef7f505747fa6233c53a6.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 319, + 268, + 430, + 342 + ], + "blocks": [ + { + "bbox": [ + 319, + 268, + 430, + 342 + ], + "lines": [ + { + "bbox": [ + 319, + 268, + 430, + 342 + ], + "spans": [ + { + "bbox": [ + 319, + 268, + 430, + 342 + ], + "type": "image", + "image_path": "1683a3d83a049a09236d36091e3e3df2a9ba1a3e4cfe032df20d396250c7954a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 347, + 425, + 358 + ], + "lines": [ + { + "bbox": [ + 370, + 347, + 425, + 358 + ], + "spans": [ + { + "bbox": [ + 370, + 347, + 425, + 358 + ], + "type": "text", + "content": "Ours-stage III" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 326, + 343, + 343, + 363 + ], + "blocks": [ + { + "bbox": [ + 326, + 343, + 343, + 363 + ], + "lines": [ + { + "bbox": [ + 326, + 343, + 343, + 363 + ], + "spans": [ + { + "bbox": [ + 326, + 343, + 343, + 363 + ], + "type": "image", + "image_path": "667677b13404260285ccecab2d683a89ff51a2737f805fdbb8f411bcd326418a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 347, + 345, + 367, + 363 + ], + "blocks": [ + { + "bbox": [ + 347, + 345, + 367, + 363 + ], + "lines": [ + { + "bbox": [ + 347, + 345, + 367, + 363 + ], + "spans": [ + { + "bbox": [ + 347, + 345, + 367, + 363 + ], + "type": "image", + "image_path": "57fd6fba537d893b39e3c90ba0fa00d528f88ac04e54e2afc7fc81f0ce5bf920.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 319, + 364, + 430, + 437 + ], + "blocks": [ + { + "bbox": [ + 319, + 364, + 430, + 437 + ], + "lines": [ + { + "bbox": [ + 319, + 364, + 430, + 437 + ], + "spans": [ + { + "bbox": [ + 319, + 364, + 430, + 437 + ], + "type": "image", + "image_path": "20b2781409bd4a9565ee62119418fffe646f4b2d365a478af1fffae848ef2e8c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 441, + 425, + 451 + ], + "lines": [ + { + "bbox": [ + 369, + 441, + 425, + 451 + ], + "spans": [ + { + "bbox": [ + 369, + 441, + 425, + 451 + ], + "type": "text", + "content": "Hasinoff et al." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 325, + 437, + 343, + 456 + ], + "blocks": [ + { + "bbox": [ + 325, + 437, + 343, + 456 + ], + "lines": [ + { + "bbox": [ + 325, + 437, + 343, + 456 + ], + "spans": [ + { + "bbox": [ + 325, + 437, + 343, + 456 + ], + "type": "image", + "image_path": "d11aebd9323ae48813cbc7cd114dc47a85b2ac92049c785ae2782ddba9d400b4.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 347, + 438, + 367, + 456 + ], + "blocks": [ + { + "bbox": [ + 347, + 438, + 367, + 456 + ], + "lines": [ + { + "bbox": [ + 347, + 438, + 367, + 456 + ], + "spans": [ + { + "bbox": [ + 347, + 438, + 367, + 456 + ], + "type": "image", + "image_path": "fb291a2e952040657bcb9d209cbe247a66a0846af8700c5f0af7d498013b017c.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 433, + 270, + 553, + 373 + ], + "blocks": [ + { + "bbox": [ + 433, + 270, + 553, + 373 + ], + "lines": [ + { + "bbox": [ + 433, + 270, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 433, + 270, + 553, + 373 + ], + "type": "image", + "image_path": "424effc145ab0444c1bb6a282fe56b3522b5122336f6055248b0ad4fc7ef21ad.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 442, + 383, + 462, + 402 + ], + "blocks": [ + { + "bbox": [ + 442, + 383, + 462, + 402 + ], + "lines": [ + { + "bbox": [ + 442, + 383, + 462, + 402 + ], + "spans": [ + { + "bbox": [ + 442, + 383, + 462, + 402 + ], + "type": "image", + "image_path": "f2c83c59ef5ff751ce19bf48dad459f0ce71b3086d490ddaebc6d9e9ace42c8b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 465, + 383, + 482, + 402 + ], + "blocks": [ + { + "bbox": [ + 465, + 383, + 482, + 402 + ], + "lines": [ + { + "bbox": [ + 465, + 383, + 482, + 402 + ], + "spans": [ + { + "bbox": [ + 465, + 383, + 482, + 402 + ], + "type": "image", + "image_path": "1a37337102f97533ed0e0cfa3318249836f2fcb80b9c7c84e6a5bb43f7ea5bd3.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 388, + 551, + 398 + ], + "lines": [ + { + "bbox": [ + 483, + 388, + 551, + 398 + ], + "spans": [ + { + "bbox": [ + 483, + 388, + 551, + 398 + ], + "type": "text", + "content": "Deep Learning Based" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 443, + 408, + 460, + 426 + ], + "blocks": [ + { + "bbox": [ + 443, + 408, + 460, + 426 + ], + "lines": [ + { + "bbox": [ + 443, + 408, + 460, + 426 + ], + "spans": [ + { + "bbox": [ + 443, + 408, + 460, + 426 + ], + "type": "image", + "image_path": "9c4db17d59b56dd50e25d451caaaf17be993c3382138924ca0dc9b09542a434c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 411, + 507, + 420 + ], + "lines": [ + { + "bbox": [ + 463, + 411, + 507, + 420 + ], + "spans": [ + { + "bbox": [ + 463, + 411, + 507, + 420 + ], + "type": "text", + "content": "Shutter Speed" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 517, + 408, + 534, + 425 + ], + "blocks": [ + { + "bbox": [ + 517, + 408, + 534, + 425 + ], + "lines": [ + { + "bbox": [ + 517, + 408, + 534, + 425 + ], + "spans": [ + { + "bbox": [ + 517, + 408, + 534, + 425 + ], + "type": "image", + "image_path": "fa947fc2e4b90cd3dd695852f772b1a5e764057e22e59b5a82bea74be07c6ae4.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 540, + 413, + 553, + 421 + ], + "lines": [ + { + "bbox": [ + 540, + 413, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 540, + 413, + 553, + 421 + ], + "type": "text", + "content": "ISO" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 443, + 433, + 463, + 444 + ], + "blocks": [ + { + "bbox": [ + 443, + 433, + 463, + 444 + ], + "lines": [ + { + "bbox": [ + 443, + 433, + 463, + 444 + ], + "spans": [ + { + "bbox": [ + 443, + 433, + 463, + 444 + ], + "type": "image", + "image_path": "e266a634297ed36372e3b0bc8a42db9165bae7d04c60382110604934bb274d38.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 434, + 550, + 443 + ], + "lines": [ + { + "bbox": [ + 470, + 434, + 550, + 443 + ], + "spans": [ + { + "bbox": [ + 470, + 434, + 550, + 443 + ], + "type": "text", + "content": "Adjustable / Unadjustable" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "bbox": [ + 152, + 531, + 200, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 531, + 200, + 544 + ], + "spans": [ + { + "bbox": [ + 152, + 531, + 200, + 544 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 556, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 663 + ], + "type": "text", + "content": "Mainstream high dynamic range imaging techniques typically rely on fusing multiple images captured with different exposure setups (shutter speed and ISO). A good balance between shutter speed and ISO is crucial for achieving high-quality HDR, as high ISO values introduce significant noise, while long shutter speeds can lead to noticeable motion blur. However, existing methods often overlook the complex interaction between shutter speed and ISO and fail to account for motion blur effects in dynamic scenes." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 67, + 664, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 295, + 676 + ], + "type": "text", + "content": "In this work, we propose AdaptiveAE, a reinforcement" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 532, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 663 + ], + "type": "text", + "content": "learning-based method that optimizes the selection of shutter speed and ISO combinations to maximize HDR reconstruction quality in dynamic environments. AdaptiveAE integrates an image synthesis pipeline that incorporates motion blur and noise simulation into our training procedure, leveraging semantic information and exposure histograms. It can adaptively select optimal ISO and shutter speed sequences based on a user-defined exposure time budget, and find a better exposure schedule than traditional solutions. Experimental results across multiple datasets demonstrate that it achieves the state-of-the-art performance." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 314, + 670, + 395, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 670, + 395, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 670, + 395, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "High-dynamic-range (HDR) imaging plays a pivotal role in computational photography. To capture an HDR scene," + } + ] + } + ], + "index": 49 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 703 + ], + "type": "text", + "content": "*This work was done during Tianyi Xu's internship at Shanghai AI Laboratory." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 71, + 703, + 151, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 151, + 713 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 151, + 713 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25176" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "due to hardware limits, a single capture can only cover a Low Dynamic Range (LDR), and HDR fusion techniques are proposed to combine multiple LDR images with varying exposures to cover a wide dynamic range [7, 10, 14, 15, 17, 19, 23, 35, 36, 40]. The typical way to vary exposure is to change either shutter speed or ISO. This is a challenging process, as a longer shutter speed may increase the signal-to-noise ratio but introduce unpleasant motion blur, a large ISO may increase brightness but also magnify noise, and significant exposure differences can cover a wider dynamic range but also increase the risk of misalignment. Therefore, the choice of exposure values (EVs) for each capture is critical in this process to ensure high-quality HDR results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 294, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 294, + 350 + ], + "type": "text", + "content": "Still, very limited work discusses how to choose the optimal exposure levels, particularly in dynamic scenes. Prior works on exposure scheduling mainly focus on static scenes and ignores potential motion blur. Learning-based techniques overlook the intricate interplay between ISO and shutter speed, resulting in suboptimal image quality under varying conditions [6, 21, 32]. Additionally, many existing methods treat ghosting and motion blur as separate, computationally intensive post-processing tasks, which is unsuitable for real-time applications [11, 25, 27, 29, 31, 38, 39]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 351, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 294, + 460 + ], + "type": "text", + "content": "In this work, we propose AdaptiveAE, an efficient exposure control algorithm designed for HDR capturing, which addresses both motion blur and noise during image acquisition. Given the previously captured image, our method optimizes the exposure bracketing strategy for the subsequent capture, based on illumination information and semantic data from previous frames. Unlike previous approaches that treat motion blur as a separate post-processing task, we aim to address it during the capture process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 294, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 640 + ], + "type": "text", + "content": "Designing both efficient and adaptive exposure control is non-trivial, and we resort to reinforcement learning [18] to solve this challenge. Our solution mimics an experienced photographer. At each iteration, the policy network takes the previously captured LDR images as input, together with extracted semantic and illumination information. Given this information, the policy network learns to determine the optimal exposure setup for the following captures, which maximizes the additional information provided by this frame while also reducing the risk of misalignment and motion blur in a dynamic scene. Once the next frame is captured, this newly captured image will be used as input for the subsequent refinement iteration. As shown in Fig. 1 right, the final quality (PSNR) of fusion increases as more images are captured." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 294, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 715 + ], + "type": "text", + "content": "Our proposed approach offers several advantages over traditional exposure controls. First, AdaptiveAE controls both exposure time and ISO and also adapts to different scenes. As a result, it has a much higher upper bound compared to either a fixed exposure schedule or an adaptive control algorithm that only changes exposure time. As shown" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "in Fig. 1, our method iteratively achieves an optimal balance between noise and blur, resulting in less noise, reduced motion blur, and superior image quality compared to other baseline methods. Second, our method can handle both static and dynamic scenes. In static scenes, it achieves performance comparable to state-of-the-art techniques, and in dynamic ones, it produces visually compelling HDR images with minimal motion blur and ghosting artifacts. Third, our method can automatically choose the best number of frames for HDR imaging. Unlike traditional HDR approaches that often use a fixed number of frames, such as three, our approach provides flexibility to determine whether capturing three or more frames is optimal for certain scenes, balancing image quality and time budget." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 241, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 338 + ], + "type": "text", + "content": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach enables the concurrent analysis of blur and noise in AE prediction, thereby enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 339, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 495 + ], + "type": "text", + "content": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Additionally, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and underscore the critical importance of incorporating blur synthesis into our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II, demonstrates superior noise control and effectively reduces motion blur, outperforming baselines in the visual quality of the fused HDR images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 496, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 592 + ], + "type": "text", + "content": "Current datasets [3, 8, 9, 13, 26] are inadequate for studying auto-exposure (AE) with simultaneous noise and motion blur considerations. To bridge this gap, we introduce a blur-aware data synthesis pipeline. This novel approach allows for concurrent analysis of blur and noise in AE prediction, enhancing HDR image quality. Our method uniquely integrates these factors, departing from traditional practices that address them separately." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "type": "text", + "content": "We evaluate AdaptiveAE on established benchmarks, including the DeepHDR Video dataset [3] and the HDRV dataset [26], employing various downstream exposure fusion techniques. Our results demonstrate state-of-the-art performance compared to existing auto-exposure methods. Also, comprehensive ablation studies and targeted experiments focusing on motion blur confirm the efficacy of our approach and highlight the critical importance of incorporating blur synthesis within our pipeline. Our method, tested on real-world scenes using a SONY Alpha 7C-II," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25177" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "demonstrates superior noise control and effectively reduces motion blur, outperforming compared baselines in the visual quality of the fused HDR images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 140, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 140, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 140, + 125 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 468 + ], + "type": "text", + "content": "Strategy for exposure bracketing. Determining the optimal set of exposures for multiple-exposure dynamic range imaging is a well-established problem. Most digital cameras allow users to set the compensation ratio for exposure bracketing, while mobile cameras typically impose fixed ratios during automatic exposure bracketing. Heuristic strategies based on the histogram are proposed by [5, 21, 30], to balance single-to-noise ratio (SNR) and saturation. The work by [6] firstly formulates this challenge as a constrained optimization problem in the linear RGB domain, addressing the scenario that follows multiple exposure fusion and precedes tone mapping and denoising. This concise formulation facilitates the use of a straightforward numerical solver. Further extensions of this formulation consider the alignment of multiple input images to account for handshake [24]. In the context of structured light 3D reconstruction, various exposures are likewise treated as an optimization problem [4]. Exposure influences not only the noise but also, to some extent, the tone of the image in the standard RGB domain. Evaluating the final image quality complicates the problem further, as subsequent tone mapping or retouching can significantly alter image quality. Therefore, a neural network is utilized to estimate exposures and fuse multiple exposed images to achieve optimal fidelity in the gamma-corrected domain [12]. Additionally, reinforcement learning is employed by [37] and [32] to assess the rewards on comprehensive image quality after more sophisticated tone adjustment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 624 + ], + "type": "text", + "content": "Challenge in dynamic scenes. In static scenes, increasing under-exposures can reduce saturation, and a longer shutter speed improves signal-to-noise ratio in dark areas. However, in dynamic scenes, excessive exposure may cause ghosting artifacts, and a prolonged shutter speed can lead to motion blur. Consequently, heuristic exposure bracketing is usually limited to two or three EV settings [5, 30]. The method by [6] imposes an upper limit on total shutter speed, while [24] addresses handshake motion through image registration but scarcely tackles local object motion. Most approaches do not fully address motion in dynamic scenes, leaving motion blur and ghosting artifacts to be addressed through post-processing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 634, + 111, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 111, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 111, + 645 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Conventional weighted linear combination methods, such as exposure fusion, provide straightforward SNR estimation but fail in dynamic scenes where moving objects create unquantifiable motion blur and ghosting artifacts due to misalignment. Our approach addresses these challenges dur" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 138 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 138 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 138 + ], + "type": "image", + "image_path": "421b9851af850fb7339a2f6eb81a66eed6ccda7a3b310a3c64c6411c95561657.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 140, + 500, + 151 + ], + "lines": [ + { + "bbox": [ + 369, + 140, + 500, + 151 + ], + "spans": [ + { + "bbox": [ + 369, + 140, + 500, + 151 + ], + "type": "text", + "content": "(a) Our overall image synthesis pipeline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 152, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 316, + 152, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 218 + ], + "type": "image", + "image_path": "4d5227d29cef88f714dd43c985a785ac119f2bd0cbd7814b458e7ea00f089ca0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 220, + 485, + 231 + ], + "lines": [ + { + "bbox": [ + 384, + 220, + 485, + 231 + ], + "spans": [ + { + "bbox": [ + 384, + 220, + 485, + 231 + ], + "type": "text", + "content": "(b) Our blur synthesis pipeline." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 299 + ], + "type": "image", + "image_path": "a616eb802b232413d45415954c7e9bc8ec5cc90dc0eb01e5ea4715d8a2852c0c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 302, + 487, + 312 + ], + "lines": [ + { + "bbox": [ + 382, + 302, + 487, + 312 + ], + "spans": [ + { + "bbox": [ + 382, + 302, + 487, + 312 + ], + "type": "text", + "content": "(c) Our noise synthesis pipeline." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 344, + 321, + 523, + 332 + ], + "lines": [ + { + "bbox": [ + 344, + 321, + 523, + 332 + ], + "spans": [ + { + "bbox": [ + 344, + 321, + 523, + 332 + ], + "type": "text", + "content": "Figure 2. Our blur-aware data synthesis pipeline." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 352, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 471 + ], + "type": "text", + "content": "ing capturing rather than in post-processing, as the latter is shown to yield suboptimal results. We predict exposure-related risks—such as motion blur, ghosting, noise, and saturation—based on a limited number of previously captured frames. Additionally, we employ a sequential strategy for exposure and ISO parameter determination, rather than simultaneously predicting settings for all three LDR images, which reflects the iterative nature of auto-exposure in mobile cameras that enables adaptation to significant brightness transitions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 479, + 499, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 499, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 499, + 491 + ], + "type": "text", + "content": "3.1.Blur-aware data synthesis pipeline" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 496, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 566 + ], + "type": "text", + "content": "To simulate capturing in real environments, we designed an image synthesis pipeline to generate realistic motion blur and noise in LDR images from HDR videos in the training dataset, for use in training. Typical exposure settings involve adjusting the exposure value (EV), which is calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 376, + 564, + 553, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 564, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 376, + 564, + 553, + 592 + ], + "type": "interline_equation", + "content": "\\mathrm {E V} = \\log_ {2} \\left(\\frac {F ^ {2}}{T} \\times \\frac {1 0 0}{\\mathrm {I S O}}\\right), \\tag {1}", + "image_path": "ee8a825c1ad9560c146b180a12a65a5a746e64ef5e90dcf083cc8589a75a1eff.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "content": " denotes the aperture's f-number, ISO represents the ISO sensitivity and " + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 594, + 553, + 617 + ], + "type": "text", + "content": " is the exposure time in seconds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 689 + ], + "type": "text", + "content": "Similar to recent methods [6, 21, 32] concerning high dynamic range capture, we assume that aperture and focus are held constant to prevent changes in defocus. This leaves just two camera settings to manipulate: (1) Shutter speed, which controls the amount of light to collect, and (2) ISO, which determines the sensor gain." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Our pipeline synthesizes motion blur and noise for the ground truth static HDR image based on a specified ISO" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25178" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 550, + 236 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 550, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 550, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 550, + 236 + ], + "type": "image", + "image_path": "9900aad84df9c2809d531a94055d7112d27cbb6801c39951e924b0b80d54dc8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. The training scheme of AdaptiveAE. States are defined as the three LDRs synthesized using predicted ISOs and shutter speeds. Starting from " + }, + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "type": "text", + "content": " where the three LDRs has EV " + }, + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\{-2,0, + 2\\}" + }, + { + "bbox": [ + 55, + 240, + 555, + 285 + ], + "type": "text", + "content": " with arbitrary EV 0 baseline, ISOs and shutter speeds, the agent sequentially predicts, customizes or inherits capturing parameters (i.e. ISO and shutter speed) for the next stage and synthesize the corresponding LDR using our image synthesis pipeline. Unlike training, the LDRs will be captured rather than synthesized during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 291, + 294, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 291, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 294, + 397 + ], + "type": "text", + "content": "and shutter speed. As depicted in Fig. 2, it uses two consecutive static HDR images as input. Motion blur is first synthesized according to shutter speed to produce a blurred HDR image in linear space. Next, noise is added based on shutter speed and ISO to create an LDR image, reflecting our exposure choice. Note that motion blur should be applied before adding noise, as it influences the raw input by affecting the number and pattern of captured photons during photography." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": "Synthesizing blur. On-the-shelf training datasets consist of consecutive HDR ground truth of a scene with motion. Now we explain how we simulate motion blur to a frame of HDR " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "f_{i}^{L}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": ", as shown in Fig. 2b, which is the " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": "-th HDR ground truth frame in the dataset scene, and the superscript " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " indicates it is in linear space. We first use " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": "-law tone-mapping with " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mu = 5000" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " to transfer " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "f_{i}^{L}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "f_{i+1}^{L}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " from HDR space to LDR space where the image interpolation algorithm we applied is trained upon, receiving " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "f_{i}^{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "f_{i+1}^{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": ", where the superscript " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " denotes they are in LDR space. Then we use RIFE [7] to interpolate the them to 256 frames and get the sequence of images " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\{f_{i}^{\\mathcal{T}}, s_{1}^{\\mathcal{T}}, s_{2}^{\\mathcal{T}}, \\dots, s_{254}^{\\mathcal{T}}, f_{i+1}^{\\mathcal{T}}\\}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": ". Then for the selected shutter speed " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "T_{j}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": "-th LDR " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "l_{j}^{\\mathcal{T}}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " to take, the blurred HDR " + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "inline_equation", + "content": "b_{j}^{L}" + }, + { + "bbox": [ + 55, + 399, + 296, + 570 + ], + "type": "text", + "content": " is simulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 578, + 295, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 578, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 64, + 578, + 295, + 607 + ], + "type": "interline_equation", + "content": "b _ {j} ^ {L} = \\mathbf {i T M O} \\left(\\frac {f _ {i} ^ {\\mathcal {T}} + \\sum_ {m = 1} ^ {m _ {j}} s _ {m} ^ {\\mathcal {T}}}{m _ {j}}\\right), m _ {j} = \\left\\lceil \\frac {2 5 6 T _ {j}}{\\Delta \\tau} \\right\\rceil , \\tag {2}", + "image_path": "62b53e4f019c7586b0fb054ef9115dae332e20a16597c784c8f2a0ecfc400775.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\Delta \\tau" + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": " denotes the time elapsed from " + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "inline_equation", + "content": "f_{i}^{L}" + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": " is taken to " + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "inline_equation", + "content": "f_{i + 1}^{L}" + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": " is not yet taken and iTMO indicates the inverse " + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "-law tonemapping function with " + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu = 5000" + }, + { + "bbox": [ + 55, + 616, + 295, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "Synthesizing noise. We adopt the noise model mentioned in [6], in which noise is modeled as a zero-mean variable, coming from three independent sources, including photon noise, which represents the Poisson distribution of photon arrivals and depends linearly upon the number of recorded" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 291, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 339 + ], + "type": "text", + "content": "electrons, " + }, + { + "bbox": [ + 313, + 291, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\Phi T" + }, + { + "bbox": [ + 313, + 291, + 555, + 339 + ], + "type": "text", + "content": ", readout noise, which comes from sensor readout, and analog-to-digital conversion(ADC) noise, which comes from the combined effect of amplifier and quantization. Hence, for pixels below the saturation level:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 328, + 346, + 555, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 346, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 328, + 346, + 555, + 371 + ], + "type": "interline_equation", + "content": "V a r (n) = \\frac {\\Phi T \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\frac {\\sigma_ {\\text {r e a d}} ^ {2} \\times \\mathrm {I S O} ^ {2}}{U ^ {2}} + \\sigma_ {\\mathrm {A D C}} ^ {2}, \\tag {3}", + "image_path": "ad51bbade6230343150a5718ce7c9769ff007cc063934aaeac6c1b42712bc5ce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "text", + "content": " is the radiance level, " + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "text", + "content": " is the shutter speed, " + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 376, + 554, + 400 + ], + "type": "text", + "content": " is a camera-dependent variable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "text", + "content": "Following our noise model, as shown in Fig. 2c, we can synthesize the corresponding noise with selected ISO and shutter speed to the blurred HDR " + }, + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "inline_equation", + "content": "b_{j}^{L}" + }, + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "text", + "content": " to get the LDR image " + }, + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "inline_equation", + "content": "l_{j}^{\\mathcal{T}}" + }, + { + "bbox": [ + 313, + 401, + 554, + 452 + ], + "type": "text", + "content": ". For details, please refer to our supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 508, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 508, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 508, + 470 + ], + "type": "text", + "content": "3.2. Problem formulation of AdaptiveAE" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": "Given a scene, the goal of AdaptiveAE, which has access to three initial preview LDR images " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\{p_j^T\\}_{1,2,3}" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " (i.e. underexposed, mid-exposed, overexposed) before capturing, is to find an optimal exposure setup (i.e. ISO and shutter speed) for LDR capturing, with which the fused HDR output will result in pleasing visual performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "text", + "content": "Our method formulates exposure bracketing as a Markov Decision Process [22], solved via deep reinforcement learning to refine exposure parameters (ISO, shutter speed) sequentially. As illustrated in Fig. 4, the process starts from three LDRs at a default " + }, + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "inline_equation", + "content": "\\{-2_p,0_p, + 2_p\\}" + }, + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "text", + "content": " EV spacing relative to an arbitrary reference (subscript " + }, + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 546, + 554, + 630 + ], + "type": "text", + "content": "). The refinement then proceeds in stages, as shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "First, the agent predicts optimal parameters for the midexposed frame, establishing a new 0-EV reference (subscript " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "). The side frames are then customized—their parameters are procedurally set to achieve a symmetric " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{-2_x,0_x, + 2_x\\}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " EV bracket. Next, the agent refines the underexposed frame to an EV of " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "-y_{x}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". The mid-exposed frame " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "(0_{x})" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " is inherited (its parameters are reused), while" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25179" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 527, + 198 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 527, + 198 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 527, + 198 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 527, + 198 + ], + "type": "image", + "image_path": "b2f7eae99e8282802337d2f8763386e293aecbd6a5f4edae5c134f6a55f4a1e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 202, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Training pipeline of our method. The ISO and shutter speed prediction process is conceptualized as a Markov Decision Process, where a CNN-based policy network predicts the ISO and shutter speed of the next exposure sets. Concurrently, a CNN-based value network estimates the state value. We leverage our blur-aware image synthesis pipeline to synthesize the predicted LDRs and employ DeepHDR [7] to fuse the predicted LDR images, generating our HDR result and calculating the reward for the current policy. The entire system is optimized using the A3C (Asynchronous Advantage Actor-Critic) method [18]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "content": "the overexposed frame is customized to " + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "inline_equation", + "content": "+y_{x}" + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "content": " to maintain symmetry, yielding " + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\{-y_{x}, 0_{x}, + y_{x}\\}" + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "content": ". Finally, the agent predicts the overexposed frame's EV as " + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "inline_equation", + "content": "+z_{x}" + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "content": ", creating a potentially asymmetric set " + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\{-y_{x}, 0_{x}, + z_{x}\\}" + }, + { + "bbox": [ + 54, + 263, + 296, + 370 + ], + "type": "text", + "content": ". This sequential prediction can be extended, as our adopted fusion method [7] handles more than three LDRs. A video providing further details on this process, along with an example of extended exposure bracketing, is available in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 378, + 189, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 189, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 189, + 391 + ], + "type": "text", + "content": "3.3. Optimization objectives" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": "Let us denote the problem as " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "P = (S, A)" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": ", where S is a state space and A is an action space. Specifically, in our task, " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " is the space of the exposure setups (i.e., ISO and shutter speed) for an LDR set that typically contains three LDRs (i.e. under-exposed, mid-exposed, and overexposed), while " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " is the set of all possible ISO and shutter speed combinations, which is discrete. During training, at stage " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "s_j = \\{(\\mathrm{ISO}_{j1}, T_{j1}), (\\mathrm{ISO}_{j2}, T_{j2}), (\\mathrm{ISO}_{j3}, T_{j3}), \\}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": ", we first find the corresponding HDR ground truth image pairs " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "(f_i^L, f_{i+1}^L)_{1,2,3}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " and generate the corresponding LDRs " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\{l_j^T\\}_{1,2,3}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " through our image synthesis pipeline in Fig. 2. The footnote 1, 2, 3 denotes that the same operation is done for the under-, mid-, and over-exposed LDRs. Taking " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\{l_j^T\\}_{1,2,3}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " as input, the agent predicts an action " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "a_j = (\\mathrm{ISO}_j, T_j)" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": ", which is expanded to an exposure setup for three LDRs by customizing EV or inheriting from state " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "s_{j-1}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": ", mapping state " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " to state " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "s_{j+1}" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": ". Adding a sequence of " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " LDRs to exposure bracketing corresponds to a trajectory " + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 396, + 296, + 623 + ], + "type": "text", + "content": " of states and actions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 633, + 294, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 633, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 100, + 633, + 294, + 644 + ], + "type": "interline_equation", + "content": "\\tau = \\left(s _ {0}, a _ {0}, \\dots , s _ {M - 1}, a _ {M - 1}, s _ {M}\\right), \\tag {4}", + "image_path": "90e09697a7eaf597ae7d301bcdd79da0a450adbbd64dd31570c6b255cf9dc156.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "s_M" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is the stopping state. Our goal is to find a policy that maximizes the accumulated reward during the decision-making process. In this paper, the reward function with the " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "-th action (i.e., corresponding to deciding the exposure setting for the " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "-th LDR for exposure fusion) is thus written as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 276, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 276, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 353, + 276, + 553, + 289 + ], + "type": "interline_equation", + "content": "r \\left(s _ {j}, a _ {j}\\right) = \\mathcal {R} \\left(s _ {j + 1}\\right) - \\mathcal {R} \\left(s _ {j}\\right) - \\mathcal {P} (j), \\tag {5}", + "image_path": "9dcd9d8b10bc3e16092d0e576ed397f4cf1b66fe67b56daaad025f46123afde8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "inline_equation", + "content": "s_{j + 1} = p(s_j,a_j)" + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "content": " denotes our reward design and " + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{step}}" + }, + { + "bbox": [ + 313, + 295, + 553, + 319 + ], + "type": "text", + "content": " penalty, detailed in 3.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": "As depicted in Fig. 4, our model comprises a policy network and a value network, both of which utilize a CNN-based architecture. The policy network predicts the optimal ISO and shutter speed for the subsequent exposure, outputting a distribution of action probabilities " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\pi(s,\\theta)" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": " for an input image " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": ". Concurrently, the value network " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "V^{\\pi}(s,\\omega)" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": " estimates the corresponding state value. These networks, with combined parameters " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\psi = (\\theta,\\omega)" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": ", are trained by maximizing our objective " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "J(\\theta)_{\\psi}" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": " to learn the optimal policy " + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\pi(s)" + }, + { + "bbox": [ + 313, + 319, + 555, + 487 + ], + "type": "text", + "content": ". Specifically, to train the policy network and the value network, we apply the A3C (Asynchronous Advantage Actor-Critic) method [18], where the actor is represented by the policy network and the critic is the value network. Network details are in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 374, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 374, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 374, + 506 + ], + "type": "text", + "content": "3.4. Reward" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 554, + 584 + ], + "type": "text", + "content": "When designing the reward function for our system, we considered four key factors: (1) similarity between the fused HDR and the ground truth HDR; (2) quality of important regions in the fused HDR; (3) quality of moving regions in the fused HDR; and (4) a penalty for overly long LDR stacks. Thus, our reward function is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 594, + 553, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 594, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 354, + 594, + 553, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = - \\left(P _ {\\text {c o n s t r u c t i o n}} + P _ {\\text {p r i o r i t y}} + P _ {\\text {g h o s t}}\\right). \\tag {6}", + "image_path": "a090e29d6907f772e73c966d6072d28c22bb2b5f1cf06e74b3e3fb54c3a9b682.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "We consider " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "P_{\\text {construction }}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", the L2 loss between our fused HDR and the ground truth, as our major reward component, which is affected by noise and saturation. Note that through the entire sequential decision process, the middle-exposed frame is used as a reference for HDR fusion. Conforming to Eq. (3), during training, noise is synthesized according to the irradiance of ground truth HDR, and during inference, noise is estimated with the irradiance of noisy signals." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25180" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{priority}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": " represents an L2 loss within the areas in the image masked by an importance mask, which is generated by a saliency predictor [20]. This ensures that the highest quality is maintained in the most significant areas, thereby enhancing the overall visual fidelity where it matters most." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ghost}}" + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "content": " is also an L2 loss within a masked area, denoting areas with large motions and thus having a higher risk of motion blur- or ghosting-caused HDR quality degradation. The mask is computed by calculating the optical flow using RAFT [28] between the HDR ground truth of the middle-exposed frame (i.e., the reference frame for fusion) and the corresponding HDR " + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "inline_equation", + "content": "f_{i}^{L}" + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "content": " and selecting the pixels where the mode of the flow vector exceeds a constant threshold " + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "content": ". Normalizing the largest optical flow vector, we empirically set " + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "content": " to 0.2. " + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ghost}}" + }, + { + "bbox": [ + 54, + 133, + 295, + 286 + ], + "type": "text", + "content": " guides the agent to deal more carefully with regions that are prone to artifacts caused by motion and is helpful for high-quality HDR capturing, as is verified by the result of our ablation studies in Tab. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 287, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(j)" + }, + { + "bbox": [ + 55, + 287, + 295, + 371 + ], + "type": "text", + "content": " is a penalty designed to penalize excessively long exposure brackets. Capturing an excessive number of shots, such as 10, even with random exposure settings, can lead to nearly perfect HDR fusion results, but it is time-consuming. Typically, three shots [2, 6, 21] are sufficient to achieve high-quality outcomes. To this end, we incorporate a penalty for taking more than three shots, as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 380, + 294, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 294, + 412 + ], + "type": "interline_equation", + "content": "\\mathcal {P} (j) = \\left\\{ \\begin{array}{l l} 0 & \\text {i f} j \\leq H \\\\ \\alpha (j - H) ^ {2} & \\text {i f} j > H \\end{array} , \\right. \\tag {7}", + "image_path": "a0ba3e79d10d773cc589bb61f02663a9171d78ff31234fa449898555dccb8962.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "text", + "content": " is a positive coefficient and " + }, + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 419, + 257, + 430 + ], + "type": "text", + "content": " is set to 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 432, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 432, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 432, + 295, + 563 + ], + "type": "text", + "content": "In this manner, the autonomous agent optimizes exposure parameters by predicting relatively fast shutter speeds for LDR images, particularly for the middle-exposed reference frame, thereby minimizing motion blur while avoiding excessive ISO values that would introduce noise-related degradation. When confronted with potential ghosting artifacts—which emerge from information deficiency in LDR images due to concurrent saturation and motion—the agent adaptively selects EVs that minimize both underexposure and saturation, resulting in a significant reduction of ghosting artifacts in the final reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 135, + 588 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Experiment details. We use Real-HDRV [26] for training and tested our performance on Real-HDRV [26] and DeepHDRVideo [3]. For HDR fusion, we adopt DeepHDR [33] to generate the HDR image based on the selected exposure bracketing. Additionally, we set the number of LDR frames to 3 for all methods involved. For methods that do not account for changes in ISO values, we set the ISO to 200 as a standard value for most cases, which is rationalized in our supplementary materials. Since RIFE [7] is relatively time-consuming, the image interpolation and blur synthesis step" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "is performed before training. We apply random flipping, rotation, and cropping with " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " pixels for data augmentation. Our training dataset consists of a total of 770 scenes, including 440 dynamic and 330 static scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": "Evaluation metrics. We directly evaluate the performance of the fused HDR results. Similar to previous HDR fusion methods [7, 14, 19], we employ PSNR-" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": ", SSIM-" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": ", PU-PSNR, PU-SSIM, and HDR-VDP-2 [16] as evaluation metrics. PSNR-" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": " and SSIM-" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": " denote PSNR and SSIM of the fused HDR after " + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": "-law tone-mapping with " + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\mu = 5000" + }, + { + "bbox": [ + 313, + 121, + 555, + 229 + ], + "type": "text", + "content": ". PU-PSNR and PU-SSIM are computed after perceptually uniform encoding [1]. When computing the HDR-VDP-2 [16], the diagonal display size is 30 inches." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 372, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 372, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 372, + 250 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 257, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 388 + ], + "type": "text", + "content": "Results on Real-HDRV dataset. We compared our trained agent's performance with several state-of-the-art HDR exposure bracketing methods, including Pourreza-Shahri et al. [21], Hasinoff et al. [6], and Wang et al. [32]. The first two are non-deep-learning methods that do not consider motion: Pourreza-Shahri et al. use K-means clustering to adjust shutter speed based on image brightness, while Hasinoff et al. mathematically optimize ISO and shutter speed for the best worst-case SNR. Wang et al. utilize reinforcement learning to predict shutter speed for maximizing PSNR, but also ignore motion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 389, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 581 + ], + "type": "text", + "content": "Our method achieves state-of-the-art performance on the HDRV-Test dataset, as shown in Tab. 1. By treating ISO as a variable, it provides flexibility in handling extremely dark scenes. Additionally, it excels in dynamic scenes due to a blur synthesis model and carefully designed rewards. Thus, our model effectively balances noise reduction and motion artifact minimization, delivering high-quality HDR results. Visualization results in Fig. 5 show that deep learning-based exposure fusion models affect EV selection differently. In Hasinoff et al.'s setup, three LDRs are equally weighted, but dynamic scenes require a reference image, emphasizing the middle-exposed LDR's quality. This often shifts its EV toward under-exposure to reduce motion blur. Our experiments confirm that if the middle-exposed image is blurry—a common problem in other methods—the fused HDR will also be blurred." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Inference time. Our method takes less than " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "250\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " for each HDR image, which is acceptable. With average exposure time " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\leq 30ms" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ") and prediction time " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\leq 10ms" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "), the total execution time is " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "6n + 3m" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\leq 250ms" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "). This can be further lowered to " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "6n" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\leq 200ms" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ") adopting concurrency. Note that DeepHDR is only required during training for calculating rewards, and on-camera inference can be performed with a small agent, taking only " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "3.5ms" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " for each frame without optimization. Besides, our method can be further accelerated by on-camera processors when applied to DSLRs. More discussions are in our supplementary materials." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25181" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 115, + 297 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 115, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 115, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 115, + 297 + ], + "type": "image", + "image_path": "d3de9365d8e7bccc28e520fe57a679c3167bff0481545594d08799c11e01c78c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 300, + 111, + 312 + ], + "lines": [ + { + "bbox": [ + 63, + 300, + 111, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 300, + 111, + 312 + ], + "type": "text", + "content": "LDR Images" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 361 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons with other auto-exposure methods on HDRV dataset [26]. Left: Predicted LDRs with varying ISO and shutter speed settings and synthesized using our image synthesis pipeline. Middle: Fused HDR image using DeepHDR [7] and tonemapined using Photomatix. Right: Zoom-in results for tested methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 72, + 284, + 297 + ], + "blocks": [ + { + "bbox": [ + 115, + 72, + 284, + 297 + ], + "lines": [ + { + "bbox": [ + 115, + 72, + 284, + 297 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 284, + 297 + ], + "type": "image", + "image_path": "ff5409d764a92d23a08ab76c626a512db868d36c4263fde7651fe6f9fedffb98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 300, + 259, + 312 + ], + "lines": [ + { + "bbox": [ + 143, + 300, + 259, + 312 + ], + "spans": [ + { + "bbox": [ + 143, + 300, + 259, + 312 + ], + "type": "text", + "content": "Our Tonemapped HDR Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 284, + 72, + 557, + 297 + ], + "blocks": [ + { + "bbox": [ + 284, + 72, + 557, + 297 + ], + "lines": [ + { + "bbox": [ + 284, + 72, + 557, + 297 + ], + "spans": [ + { + "bbox": [ + 284, + 72, + 557, + 297 + ], + "type": "image", + "image_path": "489655b8709e24eaf98b598deef205f42c532ce5f3c062feee6528dfc86214e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 300, + 466, + 311 + ], + "lines": [ + { + "bbox": [ + 290, + 300, + 466, + 311 + ], + "spans": [ + { + "bbox": [ + 290, + 300, + 466, + 311 + ], + "type": "text", + "content": "Wang et al. Pourreza et al. Hasinoff et al." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 300, + 483, + 309 + ], + "lines": [ + { + "bbox": [ + 463, + 300, + 483, + 309 + ], + "spans": [ + { + "bbox": [ + 463, + 300, + 483, + 309 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 300, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 503, + 300, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 503, + 300, + 555, + 310 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 406, + 553, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 403 + ], + "type": "text", + "content": "Table 1. Comparison of Different Methods. We utilized one preview image and three preview images, respectively, for the compared methods and our method to predict exposure settings for three LDRs. We leveraged DeepHDR [7] for exposure fusion and used the mentioned metrics to evaluate the quality of the fused HDR. Bold: The best." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 406, + 553, + 480 + ], + "lines": [ + { + "bbox": [ + 56, + 406, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 406, + 553, + 480 + ], + "type": "table", + "html": "
MethodsHDRV [26]DeepHDRVideo [3]
PSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIMPSNR-μSSIM-μHDR-VDP-2PU-PSNRPU-SSIM
Pourreza et al. [21]33.640.861754.5530.610.867935.570.878055.6731.590.8791
Hasinoff et al. [6]37.590.905257.0232.870.898038.470.915758.6534.450.9132
Wang et al. [32]36.460.890256.0932.680.893337.950.901957.3933.270.9008
Ours39.700.940859.2034.670.946539.810.937158.9036.190.9338
", + "image_path": "4216b509773b183ce3fa1ad6f25ccb14aa0826772344021ca2a071b6f7bf962d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 65, + 526, + 282, + 574 + ], + "blocks": [ + { + "bbox": [ + 55, + 484, + 295, + 518 + ], + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 518 + ], + "type": "text", + "content": "Table 2. Ablation study of AdaptiveAE on HDRV [26] dataset. Base denotes our model trained with only the step penalty and construction reward. Bold: The best." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 526, + 282, + 574 + ], + "lines": [ + { + "bbox": [ + 65, + 526, + 282, + 574 + ], + "spans": [ + { + "bbox": [ + 65, + 526, + 282, + 574 + ], + "type": "table", + "html": "
ModelPSNR-μSSIM-μPU-PSNRPU-SSIM
Base38.210.922732.680.9198
Base+Ppriority38.570.926133.020.9239
Base+Ppriority+Pghost39.700.940834.670.9465
", + "image_path": "038b226dd5a15e9b15194a0e866e3525d6f4ac371205fcad5b44942eb116127e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": "Gap to the best-achievable. For each scene in the test set of Real-HDRV, we iteratively search for the best set of predictions by Gaussian sampling around our initial prediction (50 times per exposure parameter per frame, with a deviation of " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " of the mean for both ISO and shutter speed). Statistics in Tab. 4 show that our method approaches the locally optimal result while being efficient." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Cross datasets test. To test the generalization ability of our model, we also evaluated the performance of our trained agent on DeepHDRVideo [3], as shown in the right of Tab. 1. Our agent exhibits good generalization abilities." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 581, + 559, + 651 + ], + "blocks": [ + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "type": "text", + "content": "Table 3. Comparison of different exposure fusion methods. We utilize the four auto-exposure methods to predict exposure settings and employ our image synthesis pipeline to create three LDRs. Then, we apply different exposure fusion methods to fuse them and compare the HDR quality on the HDRV dataset [26]. HDRTransformer [14] is pretrained on HDRV [26] dataset. P: PSNR- " + }, + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "type": "text", + "content": " S: SSIM- " + }, + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 484, + 555, + 574 + ], + "type": "text", + "content": " . Pou: Pourreza et al. [21], Has: Hasinoff et al. [6], W: Wang et al. [32]. HDR-Trans: HDR-Transformer. Bold: The best." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 581, + 559, + 651 + ], + "lines": [ + { + "bbox": [ + 315, + 581, + 559, + 651 + ], + "spans": [ + { + "bbox": [ + 315, + 581, + 559, + 651 + ], + "type": "table", + "html": "
ModelDeepHDRHDR-GANHDR-Trans
P-μS-μP-μS-μP-μS-μ
Pou33.640.861735.710.889235.840.8824
Has37.590.925238.580.926339.110.9372
W36.460.900237.950.916938.890.9210
Ours39.700.940840.730.937641.370.9478
", + "image_path": "f262e18bc78981af742204067dd8dd8386c5be96d524ceffc8e32ec1d81a8bf1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 664, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 555, + 700 + ], + "type": "text", + "content": "Since the DeepHDRVideo dataset only provides ground truth HDR for the middle image in the sequence, we synthesize HDR for the other images using DeepHDR [33]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Cross HDR fusion methods test. In our training scheme," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25182" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 101, + 282, + 141 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "type": "text", + "content": "Table 4. Performance comparison investigating the gap between our method and the optimum on the HDRV-test dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 101, + 282, + 141 + ], + "lines": [ + { + "bbox": [ + 70, + 101, + 282, + 141 + ], + "spans": [ + { + "bbox": [ + 70, + 101, + 282, + 141 + ], + "type": "table", + "html": "
OursWorstAverageBest
PSNR-μ39.7025.7632.4139.93
SSIM-μ0.94080.77380.86090.9412
", + "image_path": "abbd4ca540e4e79a1b0ae6ff6636e6682c1523b10f16d76cec4be8557c7d5e6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 159, + 137, + 260 + ], + "blocks": [ + { + "bbox": [ + 60, + 159, + 137, + 260 + ], + "lines": [ + { + "bbox": [ + 60, + 159, + 137, + 260 + ], + "spans": [ + { + "bbox": [ + 60, + 159, + 137, + 260 + ], + "type": "image", + "image_path": "7e2a8baf7457b8550cb3c40fa02c2876435cb8eb97581dfd60240a5bc4e405e5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 262, + 116, + 270 + ], + "lines": [ + { + "bbox": [ + 80, + 262, + 116, + 270 + ], + "spans": [ + { + "bbox": [ + 80, + 262, + 116, + 270 + ], + "type": "text", + "content": "Wang et al." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 310 + ], + "type": "text", + "content": "Figure 6. Results of our real capture data. The subject performs steady and repetitive movements, taking shots according to the exposure settings predicted by various methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 159, + 212, + 260 + ], + "blocks": [ + { + "bbox": [ + 137, + 159, + 212, + 260 + ], + "lines": [ + { + "bbox": [ + 137, + 159, + 212, + 260 + ], + "spans": [ + { + "bbox": [ + 137, + 159, + 212, + 260 + ], + "type": "image", + "image_path": "88ea1c1e3281f4498568455941fb3e88f7022e81d5092aef7f902a778a642e35.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 262, + 198, + 270 + ], + "lines": [ + { + "bbox": [ + 153, + 262, + 198, + 270 + ], + "spans": [ + { + "bbox": [ + 153, + 262, + 198, + 270 + ], + "type": "text", + "content": "Hasinoff et al." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 212, + 159, + 288, + 260 + ], + "blocks": [ + { + "bbox": [ + 212, + 159, + 288, + 260 + ], + "lines": [ + { + "bbox": [ + 212, + 159, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 212, + 159, + 288, + 260 + ], + "type": "image", + "image_path": "2e3290c7c7e4426deb4314c3ff183fdb66d0c9015fceacff00c9a0891c2ed8bf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 262, + 261, + 270 + ], + "lines": [ + { + "bbox": [ + 244, + 262, + 261, + 270 + ], + "spans": [ + { + "bbox": [ + 244, + 262, + 261, + 270 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 327, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 295, + 518 + ], + "type": "text", + "content": "we used DeepHDR for exposure fusion. For further testing, we adopted different exposure methods, including HDR-GAN [19] and HDR-Transformer [14] (pre-trained on the HDRV-dataset [26]), as post-processing exposure fusion methods. All tests are conducted on the HDRV-Test dataset. As shown in Tab. 3, without considering motion before LDR capturing, however powerful an exposure fusion method (i.e. HDR-GAN and HDR-Transformer) fails to yield a satisfactory result. Notably, when stronger fusion models are used, the performance gap between our model and traditional models, which do not account for motion, increases. This is because, when blur and ghosting risk are mitigated, post-processing models can effectively manage the remaining challenges. In contrast, ignoring motion makes these challenges more difficult for post-processing models, leading to potential failure." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 639 + ], + "type": "text", + "content": "Results for real capture. We used a SONY Alpha 7C-II to evaluate our model on real-world scenes. Subjects performed steady, repetitive movements while ISO and shutter speed were manually set for each capture, with the aperture fixed at f/2.8. The camera and subject positions remained unchanged during multi-exposure captures. As shown in Fig. 6, our method provides noise control comparable to other baselines while effectively mitigating motion blur, which can impair the visual quality of the fused HDR image; for more results, see our supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 154, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 154, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 154, + 659 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of penalty item. We conduct ablation studies to validate the effectiveness of our reward design. All quantitative evaluations are conducted on the HDRV-Test dataset. We train our networks using only the construction" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 361, + 107 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 361, + 107 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 361, + 107 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 361, + 107 + ], + "type": "image", + "image_path": "2c6298add16b1a6e25f9232f8c0a6775c18860039f33cab00fc887f63509e7e9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 107, + 361, + 141 + ], + "blocks": [ + { + "bbox": [ + 317, + 107, + 361, + 141 + ], + "lines": [ + { + "bbox": [ + 317, + 107, + 361, + 141 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 361, + 141 + ], + "type": "image", + "image_path": "8476f314e935179366fe64aeb1230113e43f4749748421b5d6db7187ddb9b9eb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 141, + 361, + 150 + ], + "lines": [ + { + "bbox": [ + 318, + 141, + 361, + 150 + ], + "spans": [ + { + "bbox": [ + 318, + 141, + 361, + 150 + ], + "type": "text", + "content": "Under-exposed" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 362, + 72, + 407, + 106 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 407, + 106 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 407, + 106 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 407, + 106 + ], + "type": "image", + "image_path": "0c97e927027381d99e123cb7e88b6ebad07e02e3b20507659125b7546a55a5eb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 362, + 106, + 407, + 141 + ], + "blocks": [ + { + "bbox": [ + 362, + 106, + 407, + 141 + ], + "lines": [ + { + "bbox": [ + 362, + 106, + 407, + 141 + ], + "spans": [ + { + "bbox": [ + 362, + 106, + 407, + 141 + ], + "type": "image", + "image_path": "cadbdc4c20577aae27ce1dd7cc0d1f84670f00b2d51c33269d4fb84ac7320e0d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 141, + 407, + 150 + ], + "lines": [ + { + "bbox": [ + 363, + 141, + 407, + 150 + ], + "spans": [ + { + "bbox": [ + 363, + 141, + 407, + 150 + ], + "type": "text", + "content": "Middle-exposed" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 452, + 106 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 452, + 106 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 452, + 106 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 452, + 106 + ], + "type": "image", + "image_path": "388664262c000dcf514fbaa58a3a7bf53d35ef51626fe922f26614cd52a12fad.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 408, + 106, + 452, + 141 + ], + "blocks": [ + { + "bbox": [ + 408, + 106, + 452, + 141 + ], + "lines": [ + { + "bbox": [ + 408, + 106, + 452, + 141 + ], + "spans": [ + { + "bbox": [ + 408, + 106, + 452, + 141 + ], + "type": "image", + "image_path": "b4f9563ac7c766b35eb5e362f1d6561a514e6b3010ca5966c0c2950c0da47e24.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 141, + 451, + 150 + ], + "lines": [ + { + "bbox": [ + 410, + 141, + 451, + 150 + ], + "spans": [ + { + "bbox": [ + 410, + 141, + 451, + 150 + ], + "type": "text", + "content": "Over-exposed" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 454, + 72, + 557, + 152 + ], + "blocks": [ + { + "bbox": [ + 454, + 72, + 557, + 152 + ], + "lines": [ + { + "bbox": [ + 454, + 72, + 557, + 152 + ], + "spans": [ + { + "bbox": [ + 454, + 72, + 557, + 152 + ], + "type": "image", + "image_path": "dc7c33839d6b7b2edd687c4a7c39c108f56aa8aa86578565109d16e9bf7ba057.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 155, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 555, + 220 + ], + "type": "text", + "content": "Figure 7. Left: Zoom-in details of the moving regions and our predicted ISO and shutter speed (seconds) for each LDR. Up: LDRs for scenes with an average motion level of around 30 pixels, Down: around 60 pixels. Right: Comparisons with baseline methods across different motion magnitude ranges. Tested on Real-HDRV [26] dataset." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "text", + "content": "reward and step penalty as the base model and test the effectiveness of " + }, + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ghost}}" + }, + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{priority}}" + }, + { + "bbox": [ + 313, + 225, + 555, + 332 + ], + "type": "text", + "content": ". As shown in Tab. 2, these two reward elements enhance our agent's performance by directing the model to focus on semantically important regions and moving areas. The former typically corresponds to the scene's focal point (e.g. a person's face), while the latter involves regions at high risk of motion blur and ghosting artifacts. Qualitative evidence is provided in our supplementary materials." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 334, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 556, + 549 + ], + "type": "text", + "content": "Robustness to dynamic scenes. We offer a pair of examples from the Real-HDRV [26] dataset with around 30 and 60 pixels of regional movement, respectively. As shown in Fig. 7: Left, our trained agent tends to predict faster shutter speeds for scenes with stronger movement, validating the responsiveness of our agent in dynamic scenes. We assess our method's performance against previous auto-exposure methods [6, 32] at varying motion magnitudes (Fig. 7: Right). To create the evaluation dataset for robustness to different motion levels, we follow HDRFlow [34], using RAFT [28] to process dynamic scenes from DRV and obtain optical flow maps. We then manually crop these images with reasonable flow predictions, dividing them into " + }, + { + "bbox": [ + 313, + 334, + 556, + 549 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 334, + 556, + 549 + ], + "type": "text", + "content": " blocks, and calculate the average motion magnitude for each block. Finally, we evaluate the PSNR of blocks corresponding to different motion magnitudes. As shown in Fig. 7: Right, our AdaptiveAE demonstrates greater robustness than other methods as motion magnitude increases." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 561, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 561, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 561, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We introduce AdaptiveAE, which optimizes HDR exposure in dynamic settings using deep reinforcement learning, treating exposure bracketing as a Markov Decision Process. It autonomously adjusts ISO and shutter speed for a pretrained exposure fusion algorithm. Reward systems focus on moving and key regions, minimizing sequence lengths. Experiments show AdaptiveAE outperforms state-of-the-art methods in dynamic scenes while matching top models in static ones, allowing flexible HDR capture. Our analysis of noise models and exposure settings offers insights for future research, with plans to include adjustable apertures." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25183" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 224, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 294, + 278 + ], + "type": "text", + "content": "[1] Maryam Azimi, Vedad Hulusic, Philippe Hanhart, and Touradj Ebrahimi. PU21: A novel perceptually uniform encoding for adapting existing quality metrics for HDR. In Proceedings of the Picture Coding Symposium (PCS), pages 1-5, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "type": "text", + "content": "[2] Neil Barakat, A. Nicholas Hone, and Thomas E. Darcie. Minimal-bracketing sets for high-dynamic-range image capture. IEEE Transactions on Image Processing, 17(10):1864-1875, 2008. 6, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 324, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 388 + ], + "type": "text", + "content": "[3] Guanying Chen, Chaofeng Chen, Shi Guo, Zhetong Liang, Kwan-Yee K. Wong, and Lei Zhang. HDR video reconstruction: A coarse-to-fine network and a real-world benchmark dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2502-2511, 2021. 2, 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 294, + 434 + ], + "type": "text", + "content": "[4] Wenyuan Chen, Xingjian Liu, Shanghai Ru, and Yu Sun. Automated exposures selection for high dynamic range structured-light 3-D scanning. IEEE Transactions on Industrial Electronics, 70(7):7428-7437, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[5] Natasha Gelfand, Andrew Adams, Sung Hee Park, and Kari Pulli. Multi-exposure imaging on mobile devices. In Proceedings of the 18th ACM International Conference on Multimedia, pages 823-826, 2010. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 480, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 533 + ], + "type": "text", + "content": "[6] Samuel W Hasinoff, Frédo Durand, and William T Freeman. Noise-optimal capture for high dynamic range photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 553-560, 2010. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 590 + ], + "type": "text", + "content": "[7] Zhewei Huang, Tianyuan Zhang, Wen Heng, Boxin Shi, and Shuchang Zhou. Real-time intermediate flow estimation for video frame interpolation. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 4, 5, 6, 7, 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 591, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 294, + 624 + ], + "type": "text", + "content": "[8] Nima Khademi Kalantari and Ravi Ramamoorthi. Deep high dynamic range imaging of dynamic scenes. ACM Transactions on Graphics (TOG), 36(4):144:1-144:12, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 668 + ], + "type": "text", + "content": "[9] Nima Khademi Kalantari, Eli Shechtman, Connelly Barnes, Soheil Darabi, Dan B. Goldman, and Pradeep Sen. Patch-based high dynamic range video. ACM Transactions on Graphics (TOG), 32(6):202:1-202:10, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[10] Lingtong Kong, Bo Li, Yike Xiong, Hao Zhang, Hong Gu, and Jinwei Chen. SAFNet: Selective alignment fusion network for efficient HDR imaging. In Proceedings of European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 185 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Key R&D Program of China No. 2022ZD0160201, Shanghai Artificial Intelligence Laboratory, National Natural Science Foundation of China (Grant No. 62136001, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. DeblurGAN-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8878-8887, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 171 + ], + "type": "text", + "content": "[12] Jieyu Li, Ruiwen Zhen, and Robert L. Stevenson. A lightweight exposure bracketing strategy for HDR imaging without access to camera RAW. *Electronic Imaging*, 2023 (15):116-1-116-6, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 228 + ], + "type": "text", + "content": "[13] Shuaizheng Liu, Xindong Zhang, Lingchen Sun, Zhetong Liang, Hui Zeng, and Lei Zhang. Joint HDR denoising and fusion: A real-world mobile HDR image dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13966-13975, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "text", + "content": "[14] Zhen Liu, Yinglong Wang, Bing Zeng, and Shuaicheng Liu. Ghost-free high dynamic range imaging with context-aware transformer. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "text", + "content": "[15] Kede Ma, Zhengfang Duanmu, Hanwei Zhu, Yuming Fang, and Zhou Wang. Deep guided learning for fast multi-exposure image fusion. IEEE Transactions on Image Processing, 29:2808-2819, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "text", + "content": "[16] Rafal Mantiuk, Kil Joong Kim, Allan G. Rempel, and Wolfgang Heidrich. HDR-VDP-2: A calibrated visual metric for visibility and quality predictions in all luminance conditions. ACM Transactions on Graphics (TOG), 30(4):1-14, 2011. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 404 + ], + "type": "text", + "content": "[17] Tom Mertens, Jan Kautz, and Frank Van Reeth. Exposure fusion: A simple and practical alternative to high dynamic range photography. Computer Graphics Forum, 28(1):161-171, 2009. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "type": "text", + "content": "[18] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. arXiv preprint:1602.01783, 2016. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 514 + ], + "type": "text", + "content": "[19] Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, and Rynson W. H. Lau. HDR-GAN: HDR image reconstruction from multi-exposed LDR images with large motions. IEEE Transactions on Image Processing, 30:3885-3896, 2021. 2, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[20] Junting Pan, Cristian Canton Ferrer, Kevin McGuinness, Noel E. O'Connor, Jordi Torres, Elisa Sayrol, and Xavier Giro-i Nieto. SalGAN: Visual saliency prediction with generative adversarial networks. arXiv preprint:1701.01081, 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 570, + 553, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 614 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 614 + ], + "type": "text", + "content": "[21] Reza Pourreza-Shahri and Nasser Kehtarnavaz. Exposure bracketing via automatic exposure selection. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 320-323, 2015. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[22] Martin L. Puterman. Markov decision processes. In *Handbooks in Operations Research and Management Science*, chapter 8, pages 331-434. North-Holland, Amsterdam, 1990. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[23] Linhao Qu, Shaolei Liu, Manning Wang, and Zhijian Song. TransMEF: A transformer-based multi-exposure image fusion framework using self-supervised multi-task learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2126-2134, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25184" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[24] Kalpana Seshadrinathan, Sung Hee Park, and Oscar Nestares. Noise and dynamic range optimal computational imaging. In Proceedings of the IEEE International Conference on Image Processing (ICIP), pages 2785-2788, 2012. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 182 + ], + "type": "text", + "content": "[25] Ziyi Shen, Wenguan Wang, Xiankai Lu, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5572-5581, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 248 + ], + "type": "text", + "content": "[26] Yong Shu, Liquan Shen, Xiangyu Hu, Mengyao Li, and Zihao Zhou. Towards real-world HDR video reconstruction: A large-scale benchmark dataset and a two-stage alignment network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2879-2888, 2024. 1, 2, 6, 7, 8, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[27] Xin Tao, Hongyun Gao, Xiaoyong Shen, Jue Wang, and Jiaya Jia. Scale-recurrent network for deep image deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8174-8182, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[28] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In Proceedings of European Conference on Computer Vision (ECCV), 2020. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[29] Fu-Jen Tsai, Yan-Tsung Peng, Chung-Chi Tsai, Yen-Yu Lin, and Chia-Wen Lin. BANet: A blur-aware attention network for dynamic scene deblurring. IEEE Transactions on Image Processing, 31:6789-6799, 2022. 2, 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 403 + ], + "type": "text", + "content": "[30] Peter van Beek. Improved image selection for stack-based HDR imaging. *Electronic Imaging*, 2019(4):581-1-581-6, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[31] Channarayapatna Shivaram Vijay, Chandramouli Paramanand, Ambasamudram Narayanan Rajagopalan, and Rama Chellappa. Non-uniform deblurring in HDR image reconstruction. IEEE Transactions on Image Processing, 22 (10):3739-3750, 2013. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[32] Zhouxia Wang, Jiawei Zhang, Mude Lin, Jiong Wang, Ping Luo, and Jimmy S. Ren. Learning a reinforced agent for flexible exposure bracketing selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1820-1828, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "text", + "content": "[33] Shangzhe Wu, Jiarui Xu, Yu-Wing Tai, and Chi-Keung Tang. Deep high dynamic range imaging with large foreground motions. In Proceedings of European Conference on Computer Vision (ECCV), 2018. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[34] Gangwei Xu, Yujin Wang, Jinwei Gu, Tianfan Xue, and Xin Yang. HDRFlow: Real-time HDR video reconstruction with large motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24851-24860, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[35] Qingsen Yan, Dong Gong, Qinfeng Shi, Anton van den Hengel, Chunhua Shen, Ian Reid, and Yanning Zhang. Attention-guided network for ghost-free high dynamic range imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1751-1760, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[36] Qingsen Yan, Lei Zhang, Yu Liu, Yu Zhu, Jinqiu Sun, Qinfeng Shi, and Yanning Zhang. Deep HDR imaging via a non-local network. IEEE Transactions on Image Processing, 29: 4308-4322, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 262 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[37] Runsheng Yu, Wenyu Liu, Yasen Zhang, Zhi Qu, Deli Zhao, and Bo Zhang. DeepExposure: Learning to expose photos with asynchronously reinforced adversarial learning. In Advances in Neural Information Processing Systems 31, pages 8377-8386, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[38] Lu Yuan, Jian Sun, Long Quan, and Heung-Yeung Shum. Image deblurring with blurred/noisy image pairs. ACM Transactions on Graphics (TOG), 26(3):1:1-1:10, 2007. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Kaihao Zhang, Wenhan Luo, Yiran Zhong, Lin Ma, Björn Stenger, Wei Liu, and Hongdong Li. Deblurring by realistic blurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2737-2746, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[40] Zhilu Zhang, Haoyu Wang, Shuai Liu, Xiaotao Wang, Lei Lei, and Wangmeng Zuo. Self-supervised high dynamic range imaging with multi-exposure images in dynamic scenes. arXiv preprint:2310.01840, 2023. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25185" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_content_list.json b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6dbb99f14496c6c69a122fab771de519b4fd9f80 --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_content_list.json @@ -0,0 +1,1706 @@ +[ + { + "type": "text", + "text": "Adding Additional Control to One-Step Diffusion with Joint Distribution Matching", + "text_level": 1, + "bbox": [ + 230, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yihong Luo", + "bbox": [ + 279, + 202, + 392, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiacheng Sun", + "bbox": [ + 282, + 220, + 401, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianyang Hu", + "bbox": [ + 426, + 203, + 544, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenguo Li", + "bbox": [ + 436, + 220, + 544, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Song", + "bbox": [ + 576, + 203, + 683, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jing Tang $^{3,1*}$", + "bbox": [ + 578, + 220, + 687, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 HKUST", + "bbox": [ + 250, + 243, + 333, + 261 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Huawei Noah's Ark Lab", + "bbox": [ + 359, + 244, + 568, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ HKUST(GZ)", + "bbox": [ + 596, + 244, + 718, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2dcb62db6757db3a3ecb22438db114f064f7959f3f1f2993f02a376ff24a787e.jpg", + "image_caption": [ + "Figure 1. Visual comparison of different strategies of adding controls. The compared baselines include 1) the diffusion with integrated standard ControlNet (denoted as ControlNet), and 2) the integration of pre-trained standard ControlNet with Diff-Instruct's pre-trained one-step generator (denoted as $\\mathrm{DI} + \\mathrm{ControlNet}$ ). Notably, our method not only maintains computational efficiency but also surpasses the visual quality achieved by the standard ControlNet approach. While the standard ControlNet approach relies heavily on high Classifier-Free Guidance (CFG) to achieve high-quality generation, this dependency might introduce unwanted artifacts in the final samples." + ], + "image_footnote": [], + "bbox": [ + 93, + 314, + 589, + 536 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ee40dbef056b33d2061172ff63216b0fc5e6fe090ed76474b1aca56a7796ff8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 321, + 906, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 625, + 326, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While diffusion distillation has enabled one-step generation through methods like Diff-Instruct [17] and Variational Score Distillation [34], adapting distilled models to emerging new controls – such as novel structural constraints or latest user preferences – remains challenging. Conventional approaches typically require modifying the base diffusion model and redistilling it – a process that is both computationally intensive and time-consuming. To address these challenges, we introduce Joint Distribution Matching (JDM), a novel approach that minimizes the reverse KL divergence between image-condition joint distributions. By deriving a tractable upper bound, JDM decouples fidelity learning from condition learning. This asymmetric distillation scheme enables our one-step student to handle con", + "bbox": [ + 88, + 657, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trols unknown to the teacher model and facilitates improved classifier-free guidance (CFG) usage and seamless integration of human feedback learning (HFL). Experimental results demonstrate that JDM surpasses baseline methods such as multi-step ControlNet [45] by mere one-step in most cases, while achieving state-of-the-art performance in one-step text-to-image synthesis through improved usage of CFG or HFL integration.", + "bbox": [ + 511, + 625, + 908, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 760, + 645, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [9, 31] have significantly advanced generative modeling, particularly in text-to-image synthesis, by producing high-quality and diverse images in a controllable manner [45]. However, their practical deployment is often hindered by the inefficiency of the sampling process, which usually takes tens to hundreds of Network Function Evaluations (NFEs). Thanks to the recent progress in diffusion distillation, the sampling efficiency has been", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Jing Tang, jingtang@ust.hk.", + "bbox": [ + 107, + 886, + 392, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4009", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "greatly enhanced, and photo-realistic images can be generated with as few as 1 NFE [17, 18, 42]. With huge computational savings in model serving, diffusion distillation of the base DM has become the standard procedure.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As artificial intelligence-generated content (AIGC) applications continue to evolve, new scenarios demand models to adapt to novel conditions and controls. These conditions encompass structural constraints, semantic guidelines, and external factors such as user preferences and additional sensory inputs. The conventional approach to integrating such controls into diffusion models entails modifying the base model and subsequently performing diffusion distillation for one-step student [44] — a process that is both computationally expensive and time-intensive. A more efficient alternative would be to extend the distillation pipeline to accommodate new controls directly, bypassing the need for extensive retraining.", + "bbox": [ + 89, + 154, + 482, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing work on this important matter is scarce. On one hand, learning additional control for one-step students remains largely unexplored. Doing so for the base DM typically relies on pre-trained ControlNet models optimized through denoising score matching (DSM) to obtain new controllability [36, 45]. However, extending ControlNet for one-step generation incurs significant limitations, such as degraded fine-grained control and suboptimal sample quality, underscoring the need for new learning paradigms that better integrate control mechanisms into one-step generators. On the other hand, incorporating additional control during diffusion distillation is also challenging. Current diffusion distillation methodologies for one-step generation predominantly focus on distilling a student model that replicates the capabilities of the teacher diffusion model [17, 42, 44], without investigating how to extend the student's abilities beyond those of the teacher. This limitation is particularly relevant when adding novel controls that the original diffusion model was not designed to handle.", + "bbox": [ + 89, + 352, + 482, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel approach termed JDM, that minimizes the reverse Kullback-Leibler (KL) divergence between the image-condition joint distributions. We derive a tractable upper bound for this divergence, which effectively decouples fidelity learning from condition learning. The asymmetrical nature of our objective enables us to obtain a one-step student that can handle controls unknown to the teacher diffusion model. Moreover, this decoupling mechanism not only facilitates improved usage of classifier-free guidance (CFG), but also enables the seamless integration of human feedback learning (HFL) into the training process. Consequently, our method enhances both the controllability and quality of generated images, providing a more flexible and efficient framework for one-step diffusion generation.", + "bbox": [ + 89, + 642, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate the superiority of our proposed JDM. For controllable generation tasks, our", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "one-step approach achieves better performance than multi-step controllable DM (50 NFE), with lower FID scores in average (14.58 vs 15.21) and better controllability measured by consistency scores (improves $24\\%$ in average). Besides, in text-to-image generation, by incorporating either human feedback learning or improved usage of CFG, our method establishes new state-of-the-art (SOTA) performance among one-step approaches. Specifically, our variant with better CFG achieves CLIP scores of 33.97, clearly outperforming the multi-step DM's 33.03.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 511, + 250, + 640, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models (DMs). DMs [9, 31] establish a forward diffusion process that progressively introduces Gaussian noise into the data over $T$ steps: $q(\\mathbf{x}_t|\\mathbf{x}) \\triangleq \\mathcal{N}(\\mathbf{x}_t; \\alpha_t\\mathbf{x}, \\sigma_t^2\\mathbf{I})$ , where $\\alpha_t$ and $\\sigma_t$ are hyper-parameters that control the diffused schedule. The diffused samples can be directly calculated as $\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon$ , with $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The diffusion network $\\epsilon_{\\phi}$ is then trained to perform denoising by minimizing: $\\mathbb{E}_{\\mathbf{x},\\epsilon,t}||\\epsilon_{\\phi}(\\mathbf{x}_t,t) - \\epsilon||_2^2$ . Once trained, the score of the diffused samples $\\mathbf{x}_t$ can be estimated using:", + "bbox": [ + 511, + 272, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\approx \\nabla_ {\\mathbf {x} _ {t}} \\log p _ {\\phi} (\\mathbf {x} _ {t}) = s _ {\\phi} (\\mathbf {x} _ {t}, t) = - \\frac {\\epsilon_ {\\phi} (\\mathbf {x} _ {t} , t)}{\\sigma_ {t}}, (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 420, + 906, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the estimated score, sampling from DMs can be achieved by solving the corresponding diffusion stochastic differential equations (SDEs) or probability flow ordinary differential equations (PF-ODEs) with multiple steps.", + "bbox": [ + 511, + 450, + 903, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ControlNet. Among other methods [1, 21], ControlNet [45] is the most promising method in adding additional Control to DMs. Given a pretrained diffusion model $\\epsilon_{\\phi}$ , a ControlNet parameterized by $\\beta$ can be trained by minimizing the denoising loss $L(\\beta)$ for injecting additional controls, where $L(\\beta) = \\mathbb{E}_{\\mathbf{x},\\epsilon ,t}||\\epsilon -\\epsilon_{\\phi ,\\beta}(\\mathbf{x}_t,c)||_2^2$ .", + "bbox": [ + 511, + 512, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion distillation. Despite the active progress in training-free accelerated sampling of DMs [12, 19, 30, 39, 46], diffusion distillation is dispensable for satisfactory few-step generation. Diffusion distillation typically follows two main appealing approaches: : 1) Trajectory distillation [13, 13, 20, 27, 32, 33, 40], which attempts to replicate the teacher model's ODE trajectories on instances level. These methods face challenges in difficult instance-level matching; 2) Distribution matching via score distillation [17, 42, 48], which aims to replicate the teacher model on distribution level using distribution divergence metrics.", + "bbox": [ + 511, + 604, + 905, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Score Distillation. Diff-Instruct (DI) [17] and Variational Score Distillation (VSD) [34] train a conditional one-step student by minimizing the reverse integral KL divergence:", + "bbox": [ + 511, + 770, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) \\| p \\left(\\mathbf {x} _ {t} \\mid c\\right)\\right), \\lambda_ {t} > 0 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 827, + 903, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $p_{\\theta}(\\mathbf{x}_t|c) \\triangleq \\int p_{\\theta}(\\mathbf{x}|c) q(\\mathbf{x}_t|\\mathbf{x}) d\\mathbf{x}$ is the distribution of diffused sample. A student trained in this way can effectively replicate the teacher's capabilities while enabling", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "one-step generation. However, training a conditional student requires a two-step process: first, a conditional teacher must be trained, followed by distillation to transfer its knowledge to the student. This sequential approach renders the introduction of new control guidance inefficient.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additional Controls for One-step Diffusion. Distilling one-step generator by score distillation has been well explored [17, 42, 48], however, how to distill one-step generator with additional controls has not been well explored. CCM [36] explores integrating consistency training with ControlNet, showing reasonable performance with four steps, while our work can surpass the standard ControlNet with mere one step in most cases. The success of previous work [17, 42, 43] in score distillation relies on initializing one-step students with the teacher. SDXS [44] explored learning one-step generator with control via score distillation, however, their teacher and fake score are required to have a ControlNet that supports the injected condition. In contrast, we minimize a tractable upper bound of joint KL divergence. This approach enables an asymmetric formulation between the teacher and student, where our student is partially initialized by the teacher, supplemented with an additional ControlNet. Our work demonstrate that it is possible to train a one-step student through score distillation, allowing it to handle conditions unknown to teacher.", + "bbox": [ + 89, + 167, + 482, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Preference Alignment For One-Step Diffusion. Since our framework supports universal guidance, we also explore integrating the reward model as additional guidance into training one-step generators for aligning with human preference. Although there are many works that try to align diffusion models with human preferences [2, 4-6, 10, 22, 23, 41], we note that how to align a one-step generator with human preference is not well explored. The existing few works [16, 25] directly maximize the reward of generated images, which can lead to obvious artifacts (see Fig. 6). In contrast, our method integrates preference learning with modeling $\\log p(c|\\mathbf{x}_t)$ , suffering less from artifacts and providing better visual quality. Besides, our frameworks decouple the condition and fidelity learning. This provides flexibility in applying CFG by an additional teacher model, showing better performance in distilling a one-step generator.", + "bbox": [ + 89, + 472, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 741, + 225, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setup. Consider a pre-trained DM with a multilevel score network $s_{\\phi}(\\mathbf{x}_t,t) = \\nabla_{\\mathbf{x}_t}\\log p_{\\phi}(\\mathbf{x}_t,t)\\approx \\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t)$ , where $p_t(\\mathbf{x}_t)$ represents the marginal diffusedistributions at time $t$ . We assume that this pre-trained model provides a high-quality approximation of the data distribution, such that $p(\\mathbf{x}_0)\\approx p_d$ . Additionally, we want to implement a new control $c$ given by a conditional discriminative model $\\log p(c|\\mathbf{x})$ . Our objective is to train a one-step generator that incorporates additional control $c$ through", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a127f32689b52919de94e84d10547aa0232b4aa40ea2ca582f7ad9082862a0e1.jpg", + "image_caption": [ + "Figure 2. The framework description of our proposed JDM." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 906, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "marginal diffusion. In essence, we aim to develop an algorithm that enables a student model to acquire new capabilities beyond those of the teacher diffusion model.", + "bbox": [ + 511, + 267, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to directly inject new controls in learning one-step student, we propose minimizing the joint reverse KL divergence between $p_{\\theta}(\\mathbf{x}_t,c) \\triangleq p_{\\theta}(\\mathbf{x}_t|c)p(c)$ and $p(\\mathbf{x},c) \\triangleq p(c|\\mathbf{x}_t)p_{\\phi}(\\mathbf{x}_t)$ , i.e.,", + "bbox": [ + 511, + 313, + 903, + 375 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\operatorname {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) | | p \\left(\\mathbf {x} _ {t}, c\\right)\\right) \\\\ = \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c) \\mid \\mid p (c \\mid \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t})\\right) \\tag {3} \\\\ = \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} [ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}, c) ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 382, + 903, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_t > 0$ , $p(c)$ is a known fixed distribution. Notably, the joint KL divergence exhibits asymmetry between the target and student distributions. The target joint distribution factorizes into the marginal distribution $p_{\\phi}(\\mathbf{x})$ and the conditional distribution $p(c|\\mathbf{x}_t)$ , with the latter being accessible through a discriminative model. This formulation enables us to distill a one-step conditional generative student that handles condition unknown to teacher generative model.", + "bbox": [ + 511, + 441, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However minimizing the KL divergence in Eq. (3) still requires access to the gradient of the conditional student. Luckily, we can access its tractable upper bound as follows.", + "bbox": [ + 511, + 561, + 905, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lemma 3.1 Suppose the condition $c$ is discrete, a upper bound of Eq. (3) can be computed by:", + "bbox": [ + 511, + 616, + 905, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} (p _ {\\theta} (\\mathbf {x} _ {t}, c) | | p (\\mathbf {x} _ {t}, c)) \\\\ \\leq \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\left[ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}) \\right] \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 654, + 903, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "See proof in the Appendix A. The gradient of the upper bound for learning conditional generator $p_{\\theta}(\\mathbf{x}|c)$ can be computed as follows:", + "bbox": [ + 511, + 695, + 905, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log p (c | \\mathbf {x} _ {t})} _ {\\text {c o n d i t i o n l e a r i n g}} \\\\ + \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})}} _ {\\text {f i d e l i t y l e a n i n g}} ] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}. \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 748, + 903, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can approximates $\\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t)$ by a score model $s_{\\psi}(\\mathbf{x}_t,t)$ , which is readily learnable through initialization via $s_{\\phi}(\\mathbf{x}_t,t)$ . Following GAN's tradition [7], we call $s_{\\phi}$ as real score and $s_{\\psi}$ as fake score.", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4011", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This upper bound naturally decomposes into two distinct learning components: conditional alignment, and generation fidelity. This decomposition contrasts with VSD and diff-instruct frameworks, where the fake score serves multiple purposes in both conditional learning and fidelity learning, potentially compromising effectiveness. Our approach reduces the burden on the fake score while eliminating the requirement for teachers to understand training conditions.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning Fake Score. We employ an auxiliary diffusion model $s_{\\psi}$ to model $\\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t)$ . The fake score can be efficiently learned through denoising:", + "bbox": [ + 89, + 212, + 483, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | \\epsilon_ {\\psi} (\\mathbf {x} _ {t}, t) - \\epsilon | | _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 267, + 482, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon$ and $p(\\epsilon)$ is the standard Gaussian distribution. After trained, we have $\\nabla_{\\mathbf{x}_t}\\log p_{\\theta ,t}(\\mathbf{x}_t)\\approx s_\\psi (\\mathbf{x}_t,t) = -\\frac{\\epsilon_\\psi(\\mathbf{x}_t,t)}{\\sigma_t}$ .", + "bbox": [ + 89, + 296, + 483, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modeling the $\\log p(c|\\mathbf{x}_t)$ . Unfortunately, conditional alignment density is defined in terms of clean samples in most cases, and rarely in terms of noisy samples. Hence, we need to find a way to approximate it. The $\\log p(c|\\mathbf{x}_t)$ can be modeled as follows:", + "bbox": [ + 89, + 345, + 483, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (c | \\mathbf {x} _ {t}) \\triangleq \\int_ {c} p (c | \\mathbf {x} _ {t}, \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\tag {7} \\\\ = \\int p (c | \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 426, + 482, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we substitute $p(c|\\mathbf{x}_t,\\mathbf{x})$ with $p(c|\\mathbf{x})$ , since the condition $c$ is fully relied on $\\mathbf{x}$ . The remaining challenge is how to model $p(\\mathbf{x}|\\mathbf{x}_t)$ . We propose to parameterize $p(\\mathbf{x}|\\mathbf{x}_t)$ using an implicit generator. An easy way can be directly parameterizing it with the fake score. However, the distribution is defined over clean samples, and using a fake score can not estimate the distribution accurately. Hence, we propose parameterizing it using a consistency model, which can be efficiently trained upon the fake score through LoRA finetuning. We note that modeling $p(c|\\mathbf{x}_t)$ via consistency models by itself is interesting, how to integrate this technique with training-free controllable generation will be promising future work, however, it is beyond the scope of this work.", + "bbox": [ + 89, + 501, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning the $p(\\mathbf{x}|\\mathbf{x}_t)$ . The $p(\\mathbf{x}|\\mathbf{x}_t)$ is modeled by a consistency model, we suggest learning the model by inserting a LoRA over the fake score for efficiency. Specifically, the consistency model [33] can be efficiently trained through:", + "bbox": [ + 89, + 696, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\beta} \\mathbb {E} _ {k, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k}}, t _ {k}) - \\mathrm {s g} (f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k - 1}}, t _ {k - 1})) | | _ {2} ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 766, + 473, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\psi , \\beta} (\\mathbf {x} _ {t}, t) \\triangleq \\left\\{ \\begin{array}{l l} \\mathbf {x} _ {t}, & \\text {i f} t = 0 \\\\ \\frac {\\mathbf {x} _ {t} - \\sigma_ {t} \\epsilon_ {\\psi , \\beta} (\\mathbf {x} _ {t} , t)}{\\alpha_ {t}}, & \\text {i f} t > 0 \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 792, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_k > t_{k-1}$ , $\\beta$ denotes the parameters of a lightweight LoRA, and $\\mathrm{sg}(\\cdot)$ denotes the stop-gradient operator.", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark. Our learning framework supports universal guidance, as demonstrated through the following instances.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Learning Better Aligned One-Step Generator", + "text_level": 1, + "bbox": [ + 511, + 90, + 897, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human Feedback Integration. We demonstrate that Human-Feedback Learning (HFL) can be seamlessly integrated into this framework. Specifically, we introduce \"human-preferred images\" as a single conditioning factor. Since we are dealing with only one condition, it is unnecessary to inject it into the desired generator. The learning gradient can then be expressed as:", + "bbox": [ + 511, + 112, + 905, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c), t} \\lambda_ {t} \\left[ \\nabla_ {\\mathbf {x} _ {t}} r \\left(\\mathbf {x} _ {t}\\right) \\right. \\\\ \\left. + \\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})} \\right] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 241, + 903, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)$ models the gradient of log probability of an image being human-preferred, formally defined as $\\nabla_{\\mathbf{x}_t}\\log p(\\text{\"Human-preferred images\"}|\\mathbf{x}_t)\\triangleq \\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)$ .", + "bbox": [ + 511, + 297, + 905, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoupled CFG. Our framework separates the learning of condition and fidelity into two distinct components. When the conditional probability $p(c|\\mathbf{x}_t)$ represents text-image alignment, its gradient can be computed using Classifier-Free Guidance (CFG). While previous approaches [42] have incorporated CFG during distillation, however, their CFG is coupled with the real score. In contrast, our framework explicitly leverages CFG for conditional learning. This key difference allows us to compute CFG using a diffusion model different from the one used for real score computation, enabling the use of more sophisticated diffusion models to guide text-image alignment.", + "bbox": [ + 511, + 345, + 906, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning One-Step Generator with Additional Control", + "text_level": 1, + "bbox": [ + 511, + 535, + 905, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controllable Generation. To incorporate additional control capabilities similar to ControlNet [45], we parameterize the student model as a diffusion model with an associated ControlNet. The generator's learning objective becomes:", + "bbox": [ + 511, + 573, + 905, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , \\beta} - \\mathbb {E} _ {p _ {\\theta , \\beta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\log p (c | \\mathbf {x} _ {t}) + \\log \\frac {p _ {\\phi , t} (\\mathbf {x} _ {t})}{p _ {\\psi , t} (\\mathbf {x} _ {t})} ], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 645, + 905, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ denotes the additional parameters for ControlNet. This formulation enables asymmetric capability development - the student model can learn conditional generation tasks beyond the teacher's abilities.", + "bbox": [ + 511, + 688, + 905, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To the best of our knowledge, we are the first to explore learning additional control one-step generator by score distillation in an asymmetric form. To highlight the effectiveness of our proposed formulation, we discuss two potential competitive variants for learning one-step generators with additional control conditions. 1) Train ControlNet via Denoising: Given a pre-trained diffusion model $\\epsilon_{\\phi}(\\mathbf{x}_t,t)$ , a ControlNet parameterized by $\\beta$ can be trained by denoising score matching loss. Hence a naive idea is training the ControlNet for one-step generation using denoising loss too.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4012", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cabc5a9d8872a46cdd005a89d572c505ff6b30f4f3ef9e5d72ec877c728dce76.jpg", + "image_caption": [ + "Control Signal", + "Implicit Training", + "Denoising Training", + "JDM(Ours)", + "Figure 3. The qualitative comparison of the proposed method and potential baselines in one-step controllable generation." + ], + "image_footnote": [], + "bbox": [ + 94, + 114, + 480, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Empirically, we find this approach can transfer the control well but generate blurry images (see Tab. 2 and Fig. 3); 2) Implicitly Train ControlNet via Score Distillation Training ControlNet for Diffusion Models is straightforward by directly using the original diffusion loss. Hence, a natural idea is to follow the training of ControlNet for DMs and directly inject the condition via ControlNet with the original VSD loss. This approach shares a high-level idea with training ControlNet for diffusion models: maintaining the original training loss while only adding ControlNet to the generator for condition injection. However, we empirically find this approach can generate high-quality images but ignores the control signal (see Tab. 2 and Fig. 3).", + "bbox": [ + 89, + 314, + 483, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Shared One-Step Generator Between Different Additional Control", + "text_level": 1, + "bbox": [ + 89, + 518, + 482, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While Eq. (10) establishes the joint training of One-Step Generator $G_{\\theta}$ and ControlNet $\\phi$ , this approach necessitates retraining the entire One-Step Generator for each new control condition. This requirement leads to inefficient use of computational resources and storage space, which limits practical applications.", + "bbox": [ + 89, + 551, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A simple solution would be to first train a one-step generator without control using diff-instruct [17], then train ControlNet to incorporate additional control into $G_{\\theta}$ . However, previous work has shown that one-step generators trained in this manner often suffer from mode collapse, making it difficult to accommodate additional control signals. One potential approach is to use teacher diffusion to generate millions of noise-image pairs, then add ODE regression loss for a one-step generator [42]. However, this method is extremely computationally inefficient and impractical for real-world applications.", + "bbox": [ + 89, + 643, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel two-phase warm-up training strategy: 1) Initial Phase: Joint training of $G_{\\theta}$ and ControlNet for a primary condition; 2) Extension Phase: Fixing $G_{\\theta}$ while training only ControlNet for subsequent conditions. This approach leverages the joint KL divergence to regularize $G_{\\theta}$ during initial training,", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resulting in better accommodating the condition and avoiding mode collapse. The resulting well-trained $G_{\\theta}$ can then effectively incorporate other forms of control.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 145, + 645, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Controllable Generation", + "text_level": 1, + "bbox": [ + 511, + 167, + 736, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment Setting. All the models are trained in an internally collected dataset. We use Stable Diffusion 1.5 [26] as the frozen teacher model and initialization for the fake score model and student. The ControlNet is initialized in the way introduced in its original paper [45]. We utilize four conditions for evaluating our method in one-step controllable generation as follows: Canny [3], Hed [37], Depth map, and lower resolution samples. Details for these conditions can be found in the Appendix B.", + "bbox": [ + 511, + 189, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric. We select the FID [8] to measure the image quality and consistency to measure the controllability. In particular, the FID is calculated between generated images by SD without additional control and generated images based on additional conditions. The consistency is defined by Consistecny $= ||h(\\mathbf{x}) - c||_{1}$ , where $c$ is the condition, and $h(\\cdot)$ is the function used for obtaining condition. We also report the number of function evaluations (NFE) required for generating an image for comparing efficiency.", + "bbox": [ + 511, + 325, + 903, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. We conduct comprehensive evaluations comparing our proposed approach against two baseline methods: (1) the standard diffusion model (DM) with ControlNet and (2) pre-trained one-step generator by diff-instruct integrated with DM's ControlNet. The quantitative results are summarized in Tab. 1, where we evaluate both image quality (FID) and condition consistency across multiple control tasks. Our results reveal several key findings: 1) The proposed method achieves a significant reduction in the number of function evaluations (NFEs) from 50 to 1, while maintaining or improving performance metrics. Specifically, our method demonstrates superior FID scores and consistency measures across various conditioning tasks, indicating both better image quality and more precise condition adherence. 2) Our shared one-step generator architecture exhibits substantial improvements over the naive integration of a pre-trained one-step generator, validating the effectiveness of our unified training strategy. This is evidenced by consistent performance gains across all evaluated metrics and tasks. 3) While it is technically feasible to directly combine DM's ControlNet with a pre-trained one-step generator, this approach yields substantially inferior results. This observation underscores the importance of our proposed strategy tailored for training one-step generators with additional control. Overall, these results demonstrate that our method successfully achieves better trade-off between efficiency and sample quality in controlled image generation, achieving state-of-the-art performance with significantly reduced computational overhead", + "bbox": [ + 511, + 463, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4013", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3908136c05d7541d41ae9979e2fb548c9a870e6c3abd35498523fe8accc6e94d.jpg", + "table_caption": [ + "Table 1. Comparison of machine metrics of different methods across tasks." + ], + "table_footnote": [], + "table_body": "
MethodNFE↓FID↓Canny Consistency↓FID↓HED Consistency↓FID↓Depth Consistency↓8× Super ResolutionFID↓Avg Consistency↓
FID↓Consistency↓
ControlNet5014.480.11319.210.10115.250.09311.930.06515.210.093
DI + ControlNet122.740.14128.040.11322.490.09715.570.12622.210.120
JDM (Ours)113.070.10216.510.04516.240.08112.480.05314.580.071
", + "bbox": [ + 94, + 114, + 903, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Control Signal", + "bbox": [ + 117, + 207, + 171, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ControlNet, 50 NFE", + "bbox": [ + 200, + 207, + 281, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DI+Control Net, 1 NFE", + "bbox": [ + 300, + 207, + 380, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "JDM(Ours), 1 NFE", + "bbox": [ + 395, + 207, + 486, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cca29ca67671377cd9f6af027a1d20526e5a37db60c44322f972b6722edb9bbf.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons on controllable generation across different control signals against competing methods." + ], + "image_footnote": [], + "bbox": [ + 94, + 238, + 488, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Control Signal", + "bbox": [ + 531, + 207, + 584, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ControlNet, 50 NFE", + "bbox": [ + 614, + 207, + 696, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DI+Control Net,1 NFE", + "bbox": [ + 712, + 207, + 795, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "JDM(Ours), 1 NFE", + "bbox": [ + 808, + 207, + 898, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/229a600e95d6d33214bdb82006387db45f65bf693a08eeac54791e75721d2fc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 238, + 903, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. We present the qualitative comparison in Fig. 4. We can observe that DM's ControlNet provides high-level control for one-step generators. However, this approach produces lower-quality images. Our customized one-step generator training with additional control can generate much higher-quality images. This validates the effectiveness of our proposed method and indicates that we can teach students conditions unknown to teachers by minimizing the upper bound of joint KL divergence.", + "bbox": [ + 89, + 595, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to Potential Baselines", + "text_level": 1, + "bbox": [ + 89, + 742, + 390, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comprehensive studies to validate our design choices by comparing with two potential baseline approaches as discussed in Sec. 3.2: 1) training ControlNet via denoising loss; 2) implicitly training ControlNet via score distillation. The quantitative results are shown in Tab. 1. The quantitative results are shown in Fig. 3.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Details. We introduce the training of two potential baselines as follows:", + "bbox": [ + 89, + 854, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Train ControlNet via Denoising: Given a pre-trained", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "one-step diffusion model $G_{\\theta}$ , the ControlNet is trained by denoising:", + "bbox": [ + 524, + 595, + 903, + 627 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL (\\beta) = \\mathbb {E} _ {\\mathbf {x}, \\epsilon} | | G _ {\\{\\phi , \\beta \\}} (\\mathbf {x} _ {T}, c) - \\mathbf {x} | | _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 633, + 903, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "* where $T$ is the terminal step in the one-step generator.", + "bbox": [ + 526, + 660, + 895, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Implicitly Train ControlNet via score distillation: The one-step student with additional control is directly trained by the original VSD loss as following:", + "bbox": [ + 513, + 675, + 905, + 720 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left(\\int_ {c} q _ {\\theta , \\beta} (\\mathbf {x} | c) p (c) d c \\mid \\mid p _ {\\phi} (\\mathbf {x})\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 728, + 903, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The condition injection is implicitly trained via reverse KL divergence.", + "bbox": [ + 524, + 763, + 903, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing to Train ControlNet via Denoising. The baseline of training ControlNet through DSM loss achieves reasonable consistency scores (0.109 for Canny and 0.094 for Depth) but suffers from significantly degraded image quality, as reflected by the worse FID scores (28.83 for Canny and 35.14 for Depth). This reveals that this approach tends to generate blurry images while maintaining decent", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4014", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ef4fbc53e9191a79d473c22101dfde5bb1667032e72c26478744a2415423389.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 109, + 85, + 883, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "control. The degradation in image quality can be attributed to the mismatch between the denoising objective and the reverse KL divergence, as the model is forced to learn denoising behavior that may not be optimal for direct generation.", + "bbox": [ + 89, + 550, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing to Implicitly Train ControlNet via score distillation. The implicit training baseline through score distillation shows better FID scores (22.87 for Canny and 22.34 for Depth) compared to the DSM approach, indicating its capability to generate higher quality images. However, it demonstrates worse consistency scores (0.151 for Canny and 0.150 for Depth), suggesting that the control signals are not effectively incorporated. This indicates that simply maintaining the original VSD loss while adding ControlNet may lead to the model ignoring the control conditions.", + "bbox": [ + 89, + 611, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Other Application in Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 89, + 772, + 482, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. We employs multiple metrics to assess different aspects of generation quality: Aesthetic Score (AeS) [29] evaluates the image quality; CLIP Score (CS) measures the text-to-image alignment; Human Preference Score (HPS) v2.1 [35] correlates strongly with the human preference, capturing both image-text alignment and aesthetic quality.", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/043c45106854410112a15c6fec88e4137f3ffab8ea5e35360ced2ae5f695f0aa.jpg", + "table_caption": [ + "Table 2. Comparison of machine metrics of different methods across tasks." + ], + "table_footnote": [], + "table_body": "
MethodNFE↓FID↓Canny Consistency↓FID↓Depth Consistency↓
JDM (Ours)113.070.10216.240.081
w/ warm-up shared UNet114.350.12216.710.093
w/o modeling p(c|xt)114.890.11318.560.085
Implicit Training122.870.15122.340.150
DI w/ Denoising128.830.10935.140.094
", + "bbox": [ + 516, + 574, + 903, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline Models. We conduct our experiments on SD-v1.5. We perform human feedback learning by ImageReward [38]. To ensure fairness in evaluation, we do not report this metric. We perform decoupled CFG by using SD-v2.1 in distilling SD-v1.5. We mainly compare our model against the open-source state-of-the-art (SOTA) models, e.g., LCM [14], Hyper-SD [25], and DMD2 [43].", + "bbox": [ + 511, + 699, + 906, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. Table 3 presents a comprehensive comparison of our method against existing approaches. Our one-step generator, enhanced with SD-v2.1-based CFG and human feedback learning, demonstrates superior performance across all evaluation metrics. Notably, our method significantly outperforms the direct baselines such as Diff-", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4015", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e3c59cc25ec5134fde73ecf2a180bc5910ae3040a17d236724bac823844c5822.jpg", + "table_caption": [ + "Table 3. Comparison of machine metrics on text-to-image generation across state-of-the-art methods. HFL denotes human feedback learning which might hack the machine metrics. We highlight the best and second best among distillation methods." + ], + "table_footnote": [], + "table_body": "
ModelBackboneHFLStepsHPS↑Aes↑CS↑
AnimationConcept-ArtPaintingPhotoAverage
Base ModelSD-v1.5No2526.2924.8524.8726.0125.505.4933.03
Base ModelSD-v2.1No2527.8227.1427.1728.1727.585.6633.46
SD Turbo [28]SD 2.1No128.3026.9226.4325.5426.805.3132.21
InstaFlow [11]SD-v1.5No123.1723.0422.7322.9722.985.2531.97
TCD [47]SD-v1.5No423.1421.1121.0823.6222.245.4329.07
LCM-dreamshaper [15]SD-v1.5No426.5126.4025.9624.3225.805.9431.55
PeRFlow [40]SD-v1.5No422.7922.1721.2823.5022.435.3530.77
DMD2 [43]SD-v1.5No124.1722.6822.9724.3023.535.8230.92
Diff Instruct [17]SD-v1.5No127.3226.1526.4125.5026.355.7132.08
Hyper-SD [25]SD-v1.5Yes128.6528.1628.4126.9028.015.6430.87
JDM w/ HFL (Ours)SD-v1.5Yes130.1629.1730.1428.3529.465.8933.75
JDM w/ better CFG (Ours)SD-v1.5No130.5629.4630.3828.5929.755.9033.97
", + "bbox": [ + 94, + 119, + 903, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1807028b1a6b7bc2baa42845ececf527a4ead7481c30891a3e972ac48612324f.jpg", + "image_caption": [ + "Figure 6. We compare JDM with the variant directly using $\\log p(c|\\mathbf{x})$ . It is clear that the variant suffers from artifacts." + ], + "image_footnote": [], + "bbox": [ + 93, + 347, + 480, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instruct [17] and DMD2 [43]. An intriguing observation is that the variant utilizing better CFG achieves even better metrics compared to the HFL variant. This unexpected finding suggests a promising research direction: the potential benefits of leveraging multiple teacher diffusion models for student model distillation, which merits further investigation in future work.", + "bbox": [ + 88, + 551, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. Since using HFL may lead to the hack of machine metrics, we further conducted qualitative comparisons. We qualitatively compared our w/ HFL and our w/ better CFG with the most competitive baseline, as shown in Fig. 5. Our method demonstrates significantly better visual quality and text-image alignment. Notably, while Hyper-SD also employs HFL, their HFL is performed separately, whereas ours is conducted collaboratively with one-step learning, allowing fake scores to participate and eliminate artifacts caused by reward maximization. As shown in Fig. 5, Hyper-SD exhibits noticeable artifacts, likely due to their flawed HFL approach.", + "bbox": [ + 88, + 657, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Additional Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 331, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Warm-up Shared UNet The proposed two-phase warm-up training strategy is crucial for learning a", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shared one-step UNet among different controls. Here, we compare our strategy to naively using pre-trained DI as shared UNet. The results are shown in Tab. 2. Without the warm-up strategy, the performance will degrade severely, especially regarding the consistency metric. This indicates our warm-up strategy can help in learning a One-step UNet that can handle multiple conditions well.", + "bbox": [ + 511, + 351, + 906, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Modeling $\\log p(c|\\mathbf{x}_t)$ Modeling $\\log p(c|\\mathbf{x}_t)$ not only makes the formulation more elegant but also ensures that the information from the gradients of condition learning and fidelity pertains to the same diffused $\\mathbf{x}_t$ . If we were to simply substitute $\\log p(c|\\mathbf{x}_t)$ with $\\log p(c|\\mathbf{x}_0)$ , the gradient information would relate separately to $\\mathbf{x}_t$ and $\\mathbf{x}_0$ , resulting in less stable learning. As demonstrated in Tab. 2, omitting the modeling of $\\log p(c|\\mathbf{x}_t)$ leads to a significant drop in performance, particularly concerning the FID. This highlights the importance of modeling $\\log p(c|\\mathbf{x}_t)$ for achieving high-quality controllable one-step generation. Furthermore, modeling $\\log p(c|\\mathbf{x}_t)$ is more robust than using $\\log p(c|\\mathbf{x}_0)$ . In HFL learning scenarios, directly modeling $\\log p(c|\\mathbf{x}_0)$ can result in noticeable artifacts caused by reward hacking, as illustrated in Fig. 6.", + "bbox": [ + 511, + 479, + 908, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose JDM for adding new control unknown to teacher DMs into one-step student. Our method minimizes the upper bound of reverse KL divergence between image-condition joint distributions. This approach decouples fidelity and condition learning, allowing the one-step student to handle controls unknown to the teacher. Extensive experiments show that JDM outperforms multi-step controllable DMs by one-step, while achieving SOTA performance in one-step text-to-image synthesis by the integration of decoupled CFG or human feedback learning.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4016", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jing Tang's work is partially supported by National Key R&D Program of China under Grant No. 2024YFA1012700 and No. 2023YFF0725100, by the National Natural Science Foundation of China (NSFC) under Grant No. 62402410 and No. U22B2060, by Guangdong Provincial Project (No. 2023QN10X025), by Guangdong Basic and Applied Basic Research Foundation under Grant No. 2023A1515110131, by Guangzhou Municipal Science and Technology Bureau under Grant No. 2024A04J4454, by Guangzhou Municipal Education Bureau (No. 2024312263), by Nansha District Project (No. 2023ZD022), and by Guangzhou Industrial Information and Intelligent Key Laboratory Project (No. 2024A03J0628) and Guangzhou Municipal Key Laboratory of Financial Technology Cutting-Edge Research (No. 2024A03J0630).", + "bbox": [ + 89, + 114, + 485, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 358, + 187, + 375 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Roni Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3", + "[3] John Canny. A computational approach to edge detection. PAMI, 1986. 5, 1", + "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3", + "[5] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023.", + "[6] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3", + "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + ], + "bbox": [ + 99, + 383, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3", + "[11] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 8", + "[12] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2", + "[13] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2", + "[14] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 7", + "[15] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 8", + "[16] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. Transactions on Machine Learning Research, 2025. 3", + "[17] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3, 5, 8", + "[18] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans, 2024. 2", + "[19] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 2", + "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14297-14306, 2023. 2", + "[21] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2", + "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3", + "[23] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4017", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1", + "[25] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis, 2024. 3, 7, 8", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5", + "[27] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2", + "[28] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation, 2023. 8", + "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7", + "[30] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2", + "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2", + "[32] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2, 4", + "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2", + "[35] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7", + "[36] Jie Xiao, Kai Zhu, Han Zhang, Zhiheng Liu, Yujun Shen, Zhantao Yang, Ruili Feng, Yu Liu, Xueyang Fu, and Zheng-Jun Zha. CCM: Real-time controllable visual content creation using text-to-image consistency models. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3", + "[37] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, 2015. 5", + "[38] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[39] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffusion sampling with optimized time steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292-8301, 2024. 2", + "[40] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 2, 8", + "[41] Zilyu Ye, Zhiyang Chen, Tiancheng Li, Zemin Huang, Wei-jian Luo, and Guo-Jun Qi. Schedule on the fly: Diffusion time prediction for faster and better image generation. arXiv preprint arXiv:2412.01243, 2024. 3", + "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3, 4, 5", + "[43] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis, 2024. 3, 7, 8", + "[44] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 2, 3", + "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 5", + "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36:49842-49869, 2023. 2", + "[47] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation, 2024. 8", + "[48] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In International Conference on Machine Learning, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 700 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4018", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_model.json b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..411030c663fea5e19a18e9884fa1fa52b73a3ea1 --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "Adding Additional Control to One-Step Diffusion with Joint Distribution Matching" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.203, + 0.393, + 0.221 + ], + "angle": 0, + "content": "Yihong Luo" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.222, + 0.403, + 0.24 + ], + "angle": 0, + "content": "Jiacheng Sun" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.204, + 0.545, + 0.222 + ], + "angle": 0, + "content": "Tianyang Hu" + }, + { + "type": "text", + "bbox": [ + 0.437, + 0.222, + 0.545, + 0.24 + ], + "angle": 0, + "content": "Zhenguo Li" + }, + { + "type": "text", + "bbox": [ + 0.578, + 0.204, + 0.684, + 0.222 + ], + "angle": 0, + "content": "Yifan Song" + }, + { + "type": "text", + "bbox": [ + 0.579, + 0.222, + 0.689, + 0.24 + ], + "angle": 0, + "content": "Jing Tang \\(^{3,1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.244, + 0.334, + 0.262 + ], + "angle": 0, + "content": "1 HKUST" + }, + { + "type": "text", + "bbox": [ + 0.361, + 0.245, + 0.57, + 0.263 + ], + "angle": 0, + "content": "\\(^{2}\\) Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.598, + 0.245, + 0.72, + 0.264 + ], + "angle": 0, + "content": "\\(^{3}\\) HKUST(GZ)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.315, + 0.59, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.322, + 0.907, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.547, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of different strategies of adding controls. The compared baselines include 1) the diffusion with integrated standard ControlNet (denoted as ControlNet), and 2) the integration of pre-trained standard ControlNet with Diff-Instruct's pre-trained one-step generator (denoted as \\(\\mathrm{DI} + \\mathrm{ControlNet}\\)). Notably, our method not only maintains computational efficiency but also surpasses the visual quality achieved by the standard ControlNet approach. While the standard ControlNet approach relies heavily on high Classifier-Free Guidance (CFG) to achieve high-quality generation, this dependency might introduce unwanted artifacts in the final samples." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.625, + 0.327, + 0.641 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.486, + 0.871 + ], + "angle": 0, + "content": "While diffusion distillation has enabled one-step generation through methods like Diff-Instruct [17] and Variational Score Distillation [34], adapting distilled models to emerging new controls – such as novel structural constraints or latest user preferences – remains challenging. Conventional approaches typically require modifying the base diffusion model and redistilling it – a process that is both computationally intensive and time-consuming. To address these challenges, we introduce Joint Distribution Matching (JDM), a novel approach that minimizes the reverse KL divergence between image-condition joint distributions. By deriving a tractable upper bound, JDM decouples fidelity learning from condition learning. This asymmetric distillation scheme enables our one-step student to handle con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.749 + ], + "angle": 0, + "content": "trols unknown to the teacher model and facilitates improved classifier-free guidance (CFG) usage and seamless integration of human feedback learning (HFL). Experimental results demonstrate that JDM surpasses baseline methods such as multi-step ControlNet [45] by mere one-step in most cases, while achieving state-of-the-art performance in one-step text-to-image synthesis through improved usage of CFG or HFL integration." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.761, + 0.646, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion models (DMs) [9, 31] have significantly advanced generative modeling, particularly in text-to-image synthesis, by producing high-quality and diverse images in a controllable manner [45]. However, their practical deployment is often hindered by the inefficiency of the sampling process, which usually takes tens to hundreds of Network Function Evaluations (NFEs). Thanks to the recent progress in diffusion distillation, the sampling efficiency has been" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.393, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author: Jing Tang, jingtang@ust.hk." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "greatly enhanced, and photo-realistic images can be generated with as few as 1 NFE [17, 18, 42]. With huge computational savings in model serving, diffusion distillation of the base DM has become the standard procedure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.351 + ], + "angle": 0, + "content": "As artificial intelligence-generated content (AIGC) applications continue to evolve, new scenarios demand models to adapt to novel conditions and controls. These conditions encompass structural constraints, semantic guidelines, and external factors such as user preferences and additional sensory inputs. The conventional approach to integrating such controls into diffusion models entails modifying the base model and subsequently performing diffusion distillation for one-step student [44] — a process that is both computationally expensive and time-intensive. A more efficient alternative would be to extend the distillation pipeline to accommodate new controls directly, bypassing the need for extensive retraining." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Existing work on this important matter is scarce. On one hand, learning additional control for one-step students remains largely unexplored. Doing so for the base DM typically relies on pre-trained ControlNet models optimized through denoising score matching (DSM) to obtain new controllability [36, 45]. However, extending ControlNet for one-step generation incurs significant limitations, such as degraded fine-grained control and suboptimal sample quality, underscoring the need for new learning paradigms that better integrate control mechanisms into one-step generators. On the other hand, incorporating additional control during diffusion distillation is also challenging. Current diffusion distillation methodologies for one-step generation predominantly focus on distilling a student model that replicates the capabilities of the teacher diffusion model [17, 42, 44], without investigating how to extend the student's abilities beyond those of the teacher. This limitation is particularly relevant when adding novel controls that the original diffusion model was not designed to handle." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.868 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel approach termed JDM, that minimizes the reverse Kullback-Leibler (KL) divergence between the image-condition joint distributions. We derive a tractable upper bound for this divergence, which effectively decouples fidelity learning from condition learning. The asymmetrical nature of our objective enables us to obtain a one-step student that can handle controls unknown to the teacher diffusion model. Moreover, this decoupling mechanism not only facilitates improved usage of classifier-free guidance (CFG), but also enables the seamless integration of human feedback learning (HFL) into the training process. Consequently, our method enhances both the controllability and quality of generated images, providing a more flexible and efficient framework for one-step diffusion generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Extensive experiments demonstrate the superiority of our proposed JDM. For controllable generation tasks, our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "one-step approach achieves better performance than multi-step controllable DM (50 NFE), with lower FID scores in average (14.58 vs 15.21) and better controllability measured by consistency scores (improves \\(24\\%\\) in average). Besides, in text-to-image generation, by incorporating either human feedback learning or improved usage of CFG, our method establishes new state-of-the-art (SOTA) performance among one-step approaches. Specifically, our variant with better CFG achieves CLIP scores of 33.97, clearly outperforming the multi-step DM's 33.03." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.642, + 0.268 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Diffusion models (DMs). DMs [9, 31] establish a forward diffusion process that progressively introduces Gaussian noise into the data over \\(T\\) steps: \\(q(\\mathbf{x}_t|\\mathbf{x}) \\triangleq \\mathcal{N}(\\mathbf{x}_t; \\alpha_t\\mathbf{x}, \\sigma_t^2\\mathbf{I})\\), where \\(\\alpha_t\\) and \\(\\sigma_t\\) are hyper-parameters that control the diffused schedule. The diffused samples can be directly calculated as \\(\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon\\), with \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The diffusion network \\(\\epsilon_{\\phi}\\) is then trained to perform denoising by minimizing: \\(\\mathbb{E}_{\\mathbf{x},\\epsilon,t}||\\epsilon_{\\phi}(\\mathbf{x}_t,t) - \\epsilon||_2^2\\). Once trained, the score of the diffused samples \\(\\mathbf{x}_t\\) can be estimated using:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.421, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\approx \\nabla_ {\\mathbf {x} _ {t}} \\log p _ {\\phi} (\\mathbf {x} _ {t}) = s _ {\\phi} (\\mathbf {x} _ {t}, t) = - \\frac {\\epsilon_ {\\phi} (\\mathbf {x} _ {t} , t)}{\\sigma_ {t}}, (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.905, + 0.512 + ], + "angle": 0, + "content": "With the estimated score, sampling from DMs can be achieved by solving the corresponding diffusion stochastic differential equations (SDEs) or probability flow ordinary differential equations (PF-ODEs) with multiple steps." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.513, + 0.905, + 0.604 + ], + "angle": 0, + "content": "ControlNet. Among other methods [1, 21], ControlNet [45] is the most promising method in adding additional Control to DMs. Given a pretrained diffusion model \\(\\epsilon_{\\phi}\\), a ControlNet parameterized by \\(\\beta\\) can be trained by minimizing the denoising loss \\(L(\\beta)\\) for injecting additional controls, where \\(L(\\beta) = \\mathbb{E}_{\\mathbf{x},\\epsilon ,t}||\\epsilon -\\epsilon_{\\phi ,\\beta}(\\mathbf{x}_t,c)||_2^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.77 + ], + "angle": 0, + "content": "Diffusion distillation. Despite the active progress in training-free accelerated sampling of DMs [12, 19, 30, 39, 46], diffusion distillation is dispensable for satisfactory few-step generation. Diffusion distillation typically follows two main appealing approaches: : 1) Trajectory distillation [13, 13, 20, 27, 32, 33, 40], which attempts to replicate the teacher model's ODE trajectories on instances level. These methods face challenges in difficult instance-level matching; 2) Distribution matching via score distillation [17, 42, 48], which aims to replicate the teacher model on distribution level using distribution divergence metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Score Distillation. Diff-Instruct (DI) [17] and Variational Score Distillation (VSD) [34] train a conditional one-step student by minimizing the reverse integral KL divergence:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.828, + 0.905, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) \\| p \\left(\\mathbf {x} _ {t} \\mid c\\right)\\right), \\lambda_ {t} > 0 \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(p_{\\theta}(\\mathbf{x}_t|c) \\triangleq \\int p_{\\theta}(\\mathbf{x}|c) q(\\mathbf{x}_t|\\mathbf{x}) d\\mathbf{x}\\) is the distribution of diffused sample. A student trained in this way can effectively replicate the teacher's capabilities while enabling" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "one-step generation. However, training a conditional student requires a two-step process: first, a conditional teacher must be trained, followed by distillation to transfer its knowledge to the student. This sequential approach renders the introduction of new control guidance inefficient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Additional Controls for One-step Diffusion. Distilling one-step generator by score distillation has been well explored [17, 42, 48], however, how to distill one-step generator with additional controls has not been well explored. CCM [36] explores integrating consistency training with ControlNet, showing reasonable performance with four steps, while our work can surpass the standard ControlNet with mere one step in most cases. The success of previous work [17, 42, 43] in score distillation relies on initializing one-step students with the teacher. SDXS [44] explored learning one-step generator with control via score distillation, however, their teacher and fake score are required to have a ControlNet that supports the injected condition. In contrast, we minimize a tractable upper bound of joint KL divergence. This approach enables an asymmetric formulation between the teacher and student, where our student is partially initialized by the teacher, supplemented with an additional ControlNet. Our work demonstrate that it is possible to train a one-step student through score distillation, allowing it to handle conditions unknown to teacher." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Human Preference Alignment For One-Step Diffusion. Since our framework supports universal guidance, we also explore integrating the reward model as additional guidance into training one-step generators for aligning with human preference. Although there are many works that try to align diffusion models with human preferences [2, 4-6, 10, 22, 23, 41], we note that how to align a one-step generator with human preference is not well explored. The existing few works [16, 25] directly maximize the reward of generated images, which can lead to obvious artifacts (see Fig. 6). In contrast, our method integrates preference learning with modeling \\(\\log p(c|\\mathbf{x}_t)\\), suffering less from artifacts and providing better visual quality. Besides, our frameworks decouple the condition and fidelity learning. This provides flexibility in applying CFG by an additional teacher model, showing better performance in distilling a one-step generator." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.226, + 0.759 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Problem Setup. Consider a pre-trained DM with a multilevel score network \\( s_{\\phi}(\\mathbf{x}_t,t) = \\nabla_{\\mathbf{x}_t}\\log p_{\\phi}(\\mathbf{x}_t,t)\\approx \\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t) \\), where \\( p_t(\\mathbf{x}_t) \\) represents the marginal diffusedistributions at time \\( t \\). We assume that this pre-trained model provides a high-quality approximation of the data distribution, such that \\( p(\\mathbf{x}_0)\\approx p_d \\). Additionally, we want to implement a new control \\( c \\) given by a conditional discriminative model \\( \\log p(c|\\mathbf{x}) \\). Our objective is to train a one-step generator that incorporates additional control \\( c \\) through" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.907, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.246, + 0.888, + 0.261 + ], + "angle": 0, + "content": "Figure 2. The framework description of our proposed JDM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.313 + ], + "angle": 0, + "content": "marginal diffusion. In essence, we aim to develop an algorithm that enables a student model to acquire new capabilities beyond those of the teacher diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.905, + 0.375 + ], + "angle": 0, + "content": "In order to directly inject new controls in learning one-step student, we propose minimizing the joint reverse KL divergence between \\( p_{\\theta}(\\mathbf{x}_t,c) \\triangleq p_{\\theta}(\\mathbf{x}_t|c)p(c) \\) and \\( p(\\mathbf{x},c) \\triangleq p(c|\\mathbf{x}_t)p_{\\phi}(\\mathbf{x}_t) \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.383, + 0.905, + 0.435 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\operatorname {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) | | p \\left(\\mathbf {x} _ {t}, c\\right)\\right) \\\\ = \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c) \\mid \\mid p (c \\mid \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t})\\right) \\tag {3} \\\\ = \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} [ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}, c) ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.906, + 0.563 + ], + "angle": 0, + "content": "where \\(\\lambda_t > 0\\), \\(p(c)\\) is a known fixed distribution. Notably, the joint KL divergence exhibits asymmetry between the target and student distributions. The target joint distribution factorizes into the marginal distribution \\(p_{\\phi}(\\mathbf{x})\\) and the conditional distribution \\(p(c|\\mathbf{x}_t)\\), with the latter being accessible through a discriminative model. This formulation enables us to distill a one-step conditional generative student that handles condition unknown to teacher generative model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.906, + 0.608 + ], + "angle": 0, + "content": "However minimizing the KL divergence in Eq. (3) still requires access to the gradient of the conditional student. Luckily, we can access its tractable upper bound as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Lemma 3.1 Suppose the condition \\( c \\) is discrete, a upper bound of Eq. (3) can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.655, + 0.905, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} (p _ {\\theta} (\\mathbf {x} _ {t}, c) | | p (\\mathbf {x} _ {t}, c)) \\\\ \\leq \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\left[ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}) \\right] \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.906, + 0.741 + ], + "angle": 0, + "content": "See proof in the Appendix A. The gradient of the upper bound for learning conditional generator \\( p_{\\theta}(\\mathbf{x}|c) \\) can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.749, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log p (c | \\mathbf {x} _ {t})} _ {\\text {c o n d i t i o n l e a r i n g}} \\\\ + \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})}} _ {\\text {f i d e l i t y l e a n i n g}} ] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}. \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We can approximates \\(\\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t)\\) by a score model \\(s_{\\psi}(\\mathbf{x}_t,t)\\), which is readily learnable through initialization via \\(s_{\\phi}(\\mathbf{x}_t,t)\\). Following GAN's tradition [7], we call \\(s_{\\phi}\\) as real score and \\(s_{\\psi}\\) as fake score." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4011" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.212 + ], + "angle": 0, + "content": "This upper bound naturally decomposes into two distinct learning components: conditional alignment, and generation fidelity. This decomposition contrasts with VSD and diff-instruct frameworks, where the fake score serves multiple purposes in both conditional learning and fidelity learning, potentially compromising effectiveness. Our approach reduces the burden on the fake score while eliminating the requirement for teachers to understand training conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.259 + ], + "angle": 0, + "content": "Learning Fake Score. We employ an auxiliary diffusion model \\( s_{\\psi} \\) to model \\( \\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t) \\). The fake score can be efficiently learned through denoising:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.268, + 0.483, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | \\epsilon_ {\\psi} (\\mathbf {x} _ {t}, t) - \\epsilon | | _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.346 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon\\) and \\(p(\\epsilon)\\) is the standard Gaussian distribution. After trained, we have \\(\\nabla_{\\mathbf{x}_t}\\log p_{\\theta ,t}(\\mathbf{x}_t)\\approx s_\\psi (\\mathbf{x}_t,t) = -\\frac{\\epsilon_\\psi(\\mathbf{x}_t,t)}{\\sigma_t}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.42 + ], + "angle": 0, + "content": "Modeling the \\(\\log p(c|\\mathbf{x}_t)\\). Unfortunately, conditional alignment density is defined in terms of clean samples in most cases, and rarely in terms of noisy samples. Hence, we need to find a way to approximate it. The \\(\\log p(c|\\mathbf{x}_t)\\) can be modeled as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.428, + 0.483, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (c | \\mathbf {x} _ {t}) \\triangleq \\int_ {c} p (c | \\mathbf {x} _ {t}, \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\tag {7} \\\\ = \\int p (c | \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.698 + ], + "angle": 0, + "content": "where we substitute \\( p(c|\\mathbf{x}_t,\\mathbf{x}) \\) with \\( p(c|\\mathbf{x}) \\), since the condition \\( c \\) is fully relied on \\( \\mathbf{x} \\). The remaining challenge is how to model \\( p(\\mathbf{x}|\\mathbf{x}_t) \\). We propose to parameterize \\( p(\\mathbf{x}|\\mathbf{x}_t) \\) using an implicit generator. An easy way can be directly parameterizing it with the fake score. However, the distribution is defined over clean samples, and using a fake score can not estimate the distribution accurately. Hence, we propose parameterizing it using a consistency model, which can be efficiently trained upon the fake score through LoRA finetuning. We note that modeling \\( p(c|\\mathbf{x}_t) \\) via consistency models by itself is interesting, how to integrate this technique with training-free controllable generation will be promising future work, however, it is beyond the scope of this work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Learning the \\( p(\\mathbf{x}|\\mathbf{x}_t) \\). The \\( p(\\mathbf{x}|\\mathbf{x}_t) \\) is modeled by a consistency model, we suggest learning the model by inserting a LoRA over the fake score for efficiency. Specifically, the consistency model [33] can be efficiently trained through:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.767, + 0.474, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\beta} \\mathbb {E} _ {k, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k}}, t _ {k}) - \\mathrm {s g} (f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k - 1}}, t _ {k - 1})) | | _ {2} ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.793, + 0.484, + 0.84 + ], + "angle": 0, + "content": "\\[\nf _ {\\psi , \\beta} (\\mathbf {x} _ {t}, t) \\triangleq \\left\\{ \\begin{array}{l l} \\mathbf {x} _ {t}, & \\text {i f} t = 0 \\\\ \\frac {\\mathbf {x} _ {t} - \\sigma_ {t} \\epsilon_ {\\psi , \\beta} (\\mathbf {x} _ {t} , t)}{\\alpha_ {t}}, & \\text {i f} t > 0 \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "where \\(t_k > t_{k-1}\\), \\(\\beta\\) denotes the parameters of a lightweight LoRA, and \\(\\mathrm{sg}(\\cdot)\\) denotes the stop-gradient operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Remark. Our learning framework supports universal guidance, as demonstrated through the following instances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.898, + 0.108 + ], + "angle": 0, + "content": "3.1. Learning Better Aligned One-Step Generator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.906, + 0.22 + ], + "angle": 0, + "content": "Human Feedback Integration. We demonstrate that Human-Feedback Learning (HFL) can be seamlessly integrated into this framework. Specifically, we introduce \"human-preferred images\" as a single conditioning factor. Since we are dealing with only one condition, it is unnecessary to inject it into the desired generator. The learning gradient can then be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.242, + 0.905, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c), t} \\lambda_ {t} \\left[ \\nabla_ {\\mathbf {x} _ {t}} r \\left(\\mathbf {x} _ {t}\\right) \\right. \\\\ \\left. + \\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})} \\right] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.906, + 0.346 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)\\) models the gradient of log probability of an image being human-preferred, formally defined as \\(\\nabla_{\\mathbf{x}_t}\\log p(\\text{\"Human-preferred images\"}|\\mathbf{x}_t)\\triangleq \\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.527 + ], + "angle": 0, + "content": "Decoupled CFG. Our framework separates the learning of condition and fidelity into two distinct components. When the conditional probability \\( p(c|\\mathbf{x}_t) \\) represents text-image alignment, its gradient can be computed using Classifier-Free Guidance (CFG). While previous approaches [42] have incorporated CFG during distillation, however, their CFG is coupled with the real score. In contrast, our framework explicitly leverages CFG for conditional learning. This key difference allows us to compute CFG using a diffusion model different from the one used for real score computation, enabling the use of more sophisticated diffusion models to guide text-image alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.906, + 0.567 + ], + "angle": 0, + "content": "3.2. Learning One-Step Generator with Additional Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.635 + ], + "angle": 0, + "content": "Controllable Generation. To incorporate additional control capabilities similar to ControlNet [45], we parameterize the student model as a diffusion model with an associated ControlNet. The generator's learning objective becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.646, + 0.906, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , \\beta} - \\mathbb {E} _ {p _ {\\theta , \\beta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\log p (c | \\mathbf {x} _ {t}) + \\log \\frac {p _ {\\phi , t} (\\mathbf {x} _ {t})}{p _ {\\psi , t} (\\mathbf {x} _ {t})} ], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.749 + ], + "angle": 0, + "content": "where \\(\\beta\\) denotes the additional parameters for ControlNet. This formulation enables asymmetric capability development - the student model can learn conditional generation tasks beyond the teacher's abilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To the best of our knowledge, we are the first to explore learning additional control one-step generator by score distillation in an asymmetric form. To highlight the effectiveness of our proposed formulation, we discuss two potential competitive variants for learning one-step generators with additional control conditions. 1) Train ControlNet via Denoising: Given a pre-trained diffusion model \\(\\epsilon_{\\phi}(\\mathbf{x}_t,t)\\), a ControlNet parameterized by \\(\\beta\\) can be trained by denoising score matching loss. Hence a naive idea is training the ControlNet for one-step generation using denoising loss too." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4012" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.09, + 0.166, + 0.114 + ], + "angle": 0, + "content": "Control Signal" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.09, + 0.264, + 0.114 + ], + "angle": 0, + "content": "Implicit Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.09, + 0.365, + 0.114 + ], + "angle": 0, + "content": "Denoising Training" + }, + { + "type": "image_caption", + "bbox": [ + 0.396, + 0.096, + 0.466, + 0.11 + ], + "angle": 0, + "content": "JDM(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.116, + 0.481, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.483, + 0.303 + ], + "angle": 0, + "content": "Figure 3. The qualitative comparison of the proposed method and potential baselines in one-step controllable generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Empirically, we find this approach can transfer the control well but generate blurry images (see Tab. 2 and Fig. 3); 2) Implicitly Train ControlNet via Score Distillation Training ControlNet for Diffusion Models is straightforward by directly using the original diffusion loss. Hence, a natural idea is to follow the training of ControlNet for DMs and directly inject the condition via ControlNet with the original VSD loss. This approach shares a high-level idea with training ControlNet for diffusion models: maintaining the original training loss while only adding ControlNet to the generator for condition injection. However, we empirically find this approach can generate high-quality images but ignores the control signal (see Tab. 2 and Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.483, + 0.549 + ], + "angle": 0, + "content": "3.2.1. Shared One-Step Generator Between Different Additional Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.645 + ], + "angle": 0, + "content": "While Eq. (10) establishes the joint training of One-Step Generator \\( G_{\\theta} \\) and ControlNet \\( \\phi \\), this approach necessitates retraining the entire One-Step Generator for each new control condition. This requirement leads to inefficient use of computational resources and storage space, which limits practical applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.81 + ], + "angle": 0, + "content": "A simple solution would be to first train a one-step generator without control using diff-instruct [17], then train ControlNet to incorporate additional control into \\( G_{\\theta} \\). However, previous work has shown that one-step generators trained in this manner often suffer from mode collapse, making it difficult to accommodate additional control signals. One potential approach is to use teacher diffusion to generate millions of noise-image pairs, then add ODE regression loss for a one-step generator [42]. However, this method is extremely computationally inefficient and impractical for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel two-phase warm-up training strategy: 1) Initial Phase: Joint training of \\( G_{\\theta} \\) and ControlNet for a primary condition; 2) Extension Phase: Fixing \\( G_{\\theta} \\) while training only ControlNet for subsequent conditions. This approach leverages the joint KL divergence to regularize \\( G_{\\theta} \\) during initial training," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "resulting in better accommodating the condition and avoiding mode collapse. The resulting well-trained \\( G_{\\theta} \\) can then effectively incorporate other forms of control." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.146, + 0.646, + 0.162 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.168, + 0.738, + 0.182 + ], + "angle": 0, + "content": "4.1. Controllable Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.905, + 0.326 + ], + "angle": 0, + "content": "Experiment Setting. All the models are trained in an internally collected dataset. We use Stable Diffusion 1.5 [26] as the frozen teacher model and initialization for the fake score model and student. The ControlNet is initialized in the way introduced in its original paper [45]. We utilize four conditions for evaluating our method in one-step controllable generation as follows: Canny [3], Hed [37], Depth map, and lower resolution samples. Details for these conditions can be found in the Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.463 + ], + "angle": 0, + "content": "Evaluation Metric. We select the FID [8] to measure the image quality and consistency to measure the controllability. In particular, the FID is calculated between generated images by SD without additional control and generated images based on additional conditions. The consistency is defined by Consistecny \\( = ||h(\\mathbf{x}) - c||_{1} \\), where \\( c \\) is the condition, and \\( h(\\cdot) \\) is the function used for obtaining condition. We also report the number of function evaluations (NFE) required for generating an image for comparing efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We conduct comprehensive evaluations comparing our proposed approach against two baseline methods: (1) the standard diffusion model (DM) with ControlNet and (2) pre-trained one-step generator by diff-instruct integrated with DM's ControlNet. The quantitative results are summarized in Tab. 1, where we evaluate both image quality (FID) and condition consistency across multiple control tasks. Our results reveal several key findings: 1) The proposed method achieves a significant reduction in the number of function evaluations (NFEs) from 50 to 1, while maintaining or improving performance metrics. Specifically, our method demonstrates superior FID scores and consistency measures across various conditioning tasks, indicating both better image quality and more precise condition adherence. 2) Our shared one-step generator architecture exhibits substantial improvements over the naive integration of a pre-trained one-step generator, validating the effectiveness of our unified training strategy. This is evidenced by consistent performance gains across all evaluated metrics and tasks. 3) While it is technically feasible to directly combine DM's ControlNet with a pre-trained one-step generator, this approach yields substantially inferior results. This observation underscores the importance of our proposed strategy tailored for training one-step generators with additional control. Overall, these results demonstrate that our method successfully achieves better trade-off between efficiency and sample quality in controlled image generation, achieving state-of-the-art performance with significantly reduced computational overhead" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4013" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.277, + 0.09, + 0.722, + 0.103 + ], + "angle": 0, + "content": "Table 1. Comparison of machine metrics of different methods across tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.904, + 0.196 + ], + "angle": 0, + "content": "
MethodNFE↓FID↓Canny Consistency↓FID↓HED Consistency↓FID↓Depth Consistency↓8× Super ResolutionFID↓Avg Consistency↓
FID↓Consistency↓
ControlNet5014.480.11319.210.10115.250.09311.930.06515.210.093
DI + ControlNet122.740.14128.040.11322.490.09715.570.12622.210.120
JDM (Ours)113.070.10216.510.04516.240.08112.480.05314.580.071
" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.208, + 0.173, + 0.237 + ], + "angle": 0, + "content": "Control Signal" + }, + { + "type": "text", + "bbox": [ + 0.201, + 0.208, + 0.282, + 0.236 + ], + "angle": 0, + "content": "ControlNet, 50 NFE" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.208, + 0.382, + 0.236 + ], + "angle": 0, + "content": "DI+Control Net, 1 NFE" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.208, + 0.487, + 0.236 + ], + "angle": 0, + "content": "JDM(Ours), 1 NFE" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.239, + 0.49, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.208, + 0.585, + 0.236 + ], + "angle": 0, + "content": "Control Signal" + }, + { + "type": "text", + "bbox": [ + 0.616, + 0.208, + 0.697, + 0.236 + ], + "angle": 0, + "content": "ControlNet, 50 NFE" + }, + { + "type": "text", + "bbox": [ + 0.713, + 0.208, + 0.796, + 0.236 + ], + "angle": 0, + "content": "DI+Control Net,1 NFE" + }, + { + "type": "text", + "bbox": [ + 0.81, + 0.208, + 0.899, + 0.236 + ], + "angle": 0, + "content": "JDM(Ours), 1 NFE" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.239, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.556, + 0.855, + 0.571 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on controllable generation across different control signals against competing methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Qualitative Comparison. We present the qualitative comparison in Fig. 4. We can observe that DM's ControlNet provides high-level control for one-step generators. However, this approach produces lower-quality images. Our customized one-step generator training with additional control can generate much higher-quality images. This validates the effectiveness of our proposed method and indicates that we can teach students conditions unknown to teachers by minimizing the upper bound of joint KL divergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.391, + 0.759 + ], + "angle": 0, + "content": "4.2. Comparison to Potential Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.855 + ], + "angle": 0, + "content": "We conduct comprehensive studies to validate our design choices by comparing with two potential baseline approaches as discussed in Sec. 3.2: 1) training ControlNet via denoising loss; 2) implicitly training ControlNet via score distillation. The quantitative results are shown in Tab. 1. The quantitative results are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Baseline Details. We introduce the training of two potential baselines as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Train ControlNet via Denoising: Given a pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.597, + 0.905, + 0.628 + ], + "angle": 0, + "content": "one-step diffusion model \\( G_{\\theta} \\), the ControlNet is trained by denoising:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.635, + 0.905, + 0.653 + ], + "angle": 0, + "content": "\\[\nL (\\beta) = \\mathbb {E} _ {\\mathbf {x}, \\epsilon} | | G _ {\\{\\phi , \\beta \\}} (\\mathbf {x} _ {T}, c) - \\mathbf {x} | | _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.661, + 0.897, + 0.675 + ], + "angle": 0, + "content": "* where \\( T \\) is the terminal step in the one-step generator." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.676, + 0.906, + 0.721 + ], + "angle": 0, + "content": "- Implicitly Train ControlNet via score distillation: The one-step student with additional control is directly trained by the original VSD loss as following:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.729, + 0.905, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left(\\int_ {c} q _ {\\theta , \\beta} (\\mathbf {x} | c) p (c) d c \\mid \\mid p _ {\\phi} (\\mathbf {x})\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.765, + 0.905, + 0.795 + ], + "angle": 0, + "content": "The condition injection is implicitly trained via reverse KL divergence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparing to Train ControlNet via Denoising. The baseline of training ControlNet through DSM loss achieves reasonable consistency scores (0.109 for Canny and 0.094 for Depth) but suffers from significantly degraded image quality, as reflected by the worse FID scores (28.83 for Canny and 35.14 for Depth). This reveals that this approach tends to generate blurry images while maintaining decent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4014" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.885, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.611 + ], + "angle": 0, + "content": "control. The degradation in image quality can be attributed to the mismatch between the denoising objective and the reverse KL divergence, as the model is forced to learn denoising behavior that may not be optimal for direct generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Comparing to Implicitly Train ControlNet via score distillation. The implicit training baseline through score distillation shows better FID scores (22.87 for Canny and 22.34 for Depth) compared to the DSM approach, indicating its capability to generate higher quality images. However, it demonstrates worse consistency scores (0.151 for Canny and 0.150 for Depth), suggesting that the control signals are not effectively incorporated. This indicates that simply maintaining the original VSD loss while adding ControlNet may lead to the model ignoring the control conditions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.79 + ], + "angle": 0, + "content": "4.3. Other Application in Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation. We employs multiple metrics to assess different aspects of generation quality: Aesthetic Score (AeS) [29] evaluates the image quality; CLIP Score (CS) measures the text-to-image alignment; Human Preference Score (HPS) v2.1 [35] correlates strongly with the human preference, capturing both image-text alignment and aesthetic quality." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.549, + 0.905, + 0.574 + ], + "angle": 0, + "content": "Table 2. Comparison of machine metrics of different methods across tasks." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.668 + ], + "angle": 0, + "content": "
MethodNFE↓FID↓Canny Consistency↓FID↓Depth Consistency↓
JDM (Ours)113.070.10216.240.081
w/ warm-up shared UNet114.350.12216.710.093
w/o modeling p(c|xt)114.890.11318.560.085
Implicit Training122.870.15122.340.150
DI w/ Denoising128.830.10935.140.094
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Baseline Models. We conduct our experiments on SD-v1.5. We perform human feedback learning by ImageReward [38]. To ensure fairness in evaluation, we do not report this metric. We perform decoupled CFG by using SD-v2.1 in distilling SD-v1.5. We mainly compare our model against the open-source state-of-the-art (SOTA) models, e.g., LCM [14], Hyper-SD [25], and DMD2 [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative Results. Table 3 presents a comprehensive comparison of our method against existing approaches. Our one-step generator, enhanced with SD-v2.1-based CFG and human feedback learning, demonstrates superior performance across all evaluation metrics. Notably, our method significantly outperforms the direct baselines such as Diff-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4015" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 3. Comparison of machine metrics on text-to-image generation across state-of-the-art methods. HFL denotes human feedback learning which might hack the machine metrics. We highlight the best and second best among distillation methods." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.904, + 0.328 + ], + "angle": 0, + "content": "
ModelBackboneHFLStepsHPS↑Aes↑CS↑
AnimationConcept-ArtPaintingPhotoAverage
Base ModelSD-v1.5No2526.2924.8524.8726.0125.505.4933.03
Base ModelSD-v2.1No2527.8227.1427.1728.1727.585.6633.46
SD Turbo [28]SD 2.1No128.3026.9226.4325.5426.805.3132.21
InstaFlow [11]SD-v1.5No123.1723.0422.7322.9722.985.2531.97
TCD [47]SD-v1.5No423.1421.1121.0823.6222.245.4329.07
LCM-dreamshaper [15]SD-v1.5No426.5126.4025.9624.3225.805.9431.55
PeRFlow [40]SD-v1.5No422.7922.1721.2823.5022.435.3530.77
DMD2 [43]SD-v1.5No124.1722.6822.9724.3023.535.8230.92
Diff Instruct [17]SD-v1.5No127.3226.1526.4125.5026.355.7132.08
Hyper-SD [25]SD-v1.5Yes128.6528.1628.4126.9028.015.6430.87
JDM w/ HFL (Ours)SD-v1.5Yes130.1629.1730.1428.3529.465.8933.75
JDM w/ better CFG (Ours)SD-v1.5No130.5629.4630.3828.5929.755.9033.97
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.348, + 0.482, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.53 + ], + "angle": 0, + "content": "Figure 6. We compare JDM with the variant directly using \\(\\log p(c|\\mathbf{x})\\). It is clear that the variant suffers from artifacts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Instruct [17] and DMD2 [43]. An intriguing observation is that the variant utilizing better CFG achieves even better metrics compared to the HFL variant. This unexpected finding suggests a promising research direction: the potential benefits of leveraging multiple teacher diffusion models for student model distillation, which merits further investigation in future work." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Qualitative Comparison. Since using HFL may lead to the hack of machine metrics, we further conducted qualitative comparisons. We qualitatively compared our w/ HFL and our w/ better CFG with the most competitive baseline, as shown in Fig. 5. Our method demonstrates significantly better visual quality and text-image alignment. Notably, while Hyper-SD also employs HFL, their HFL is performed separately, whereas ours is conducted collaboratively with one-step learning, allowing fake scores to participate and eliminate artifacts caused by reward maximization. As shown in Fig. 5, Hyper-SD exhibits noticeable artifacts, likely due to their flawed HFL approach." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.332, + 0.866 + ], + "angle": 0, + "content": "4.4. Additional Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Effect of Warm-up Shared UNet The proposed two-phase warm-up training strategy is crucial for learning a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.459 + ], + "angle": 0, + "content": "shared one-step UNet among different controls. Here, we compare our strategy to naively using pre-trained DI as shared UNet. The results are shown in Tab. 2. Without the warm-up strategy, the performance will degrade severely, especially regarding the consistency metric. This indicates our warm-up strategy can help in learning a One-step UNet that can handle multiple conditions well." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.909, + 0.709 + ], + "angle": 0, + "content": "Effect of Modeling \\(\\log p(c|\\mathbf{x}_t)\\) Modeling \\(\\log p(c|\\mathbf{x}_t)\\) not only makes the formulation more elegant but also ensures that the information from the gradients of condition learning and fidelity pertains to the same diffused \\(\\mathbf{x}_t\\). If we were to simply substitute \\(\\log p(c|\\mathbf{x}_t)\\) with \\(\\log p(c|\\mathbf{x}_0)\\), the gradient information would relate separately to \\(\\mathbf{x}_t\\) and \\(\\mathbf{x}_0\\), resulting in less stable learning. As demonstrated in Tab. 2, omitting the modeling of \\(\\log p(c|\\mathbf{x}_t)\\) leads to a significant drop in performance, particularly concerning the FID. This highlights the importance of modeling \\(\\log p(c|\\mathbf{x}_t)\\) for achieving high-quality controllable one-step generation. Furthermore, modeling \\(\\log p(c|\\mathbf{x}_t)\\) is more robust than using \\(\\log p(c|\\mathbf{x}_0)\\). In HFL learning scenarios, directly modeling \\(\\log p(c|\\mathbf{x}_0)\\) can result in noticeable artifacts caused by reward hacking, as illustrated in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose JDM for adding new control unknown to teacher DMs into one-step student. Our method minimizes the upper bound of reverse KL divergence between image-condition joint distributions. This approach decouples fidelity and condition learning, allowing the one-step student to handle controls unknown to the teacher. Extensive experiments show that JDM outperforms multi-step controllable DMs by one-step, while achieving SOTA performance in one-step text-to-image synthesis by the integration of decoupled CFG or human feedback learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4016" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.345 + ], + "angle": 0, + "content": "Jing Tang's work is partially supported by National Key R&D Program of China under Grant No. 2024YFA1012700 and No. 2023YFF0725100, by the National Natural Science Foundation of China (NSFC) under Grant No. 62402410 and No. U22B2060, by Guangdong Provincial Project (No. 2023QN10X025), by Guangdong Basic and Applied Basic Research Foundation under Grant No. 2023A1515110131, by Guangzhou Municipal Science and Technology Bureau under Grant No. 2024A04J4454, by Guangzhou Municipal Education Bureau (No. 2024312263), by Nansha District Project (No. 2023ZD022), and by Guangzhou Industrial Information and Intelligent Key Laboratory Project (No. 2024A03J0628) and Guangzhou Municipal Key Laboratory of Financial Technology Cutting-Edge Research (No. 2024A03J0630)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.359, + 0.188, + 0.375 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.384, + 0.483, + 0.452 + ], + "angle": 0, + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Roni Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.456, + 0.484, + 0.498 + ], + "angle": 0, + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[3] John Canny. A computational approach to edge detection. PAMI, 1986. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[5] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[6] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.384, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[10] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[11] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[12] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[13] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[14] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[15] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[16] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. Transactions on Machine Learning Research, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[17] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[18] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[19] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.906, + 0.717 + ], + "angle": 0, + "content": "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14297-14306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[21] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4017" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[24] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[25] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[28] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[30] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[32] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[35] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[36] Jie Xiao, Kai Zhu, Han Zhang, Zhiheng Liu, Yujun Shen, Zhantao Yang, Ruili Feng, Yu Liu, Xueyang Fu, and Zheng-Jun Zha. CCM: Real-time controllable visual content creation using text-to-image consistency models. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[39] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffusion sampling with optimized time steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292-8301, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[40] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[41] Zilyu Ye, Zhiyang Chen, Tiancheng Li, Zemin Huang, Wei-jian Luo, and Guo-Jun Qi. Schedule on the fly: Diffusion time prediction for faster and better image generation. arXiv preprint arXiv:2412.01243, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[43] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[44] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36:49842-49869, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[47] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[48] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In International Conference on Machine Learning, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4018" + } + ] +] \ No newline at end of file diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_origin.pdf b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..386c53d723dedd2258faf6a560486df567243a4c --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/c2b3835f-f805-4917-82a0-466f101155c4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61de1292c536e6adf4753a7f24c74dbafa081b6673418a4f9022b36adfa7557f +size 5421302 diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/full.md b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e04c5892dee6c77f7d344d13501646a45d97b6b --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/full.md @@ -0,0 +1,345 @@ +# Adding Additional Control to One-Step Diffusion with Joint Distribution Matching + +Yihong Luo + +Jiacheng Sun + +Tianyang Hu + +Zhenguo Li + +Yifan Song + +Jing Tang $^{3,1*}$ + +1 HKUST + +$^{2}$ Huawei Noah's Ark Lab + +$^{3}$ HKUST(GZ) + +![](images/2dcb62db6757db3a3ecb22438db114f064f7959f3f1f2993f02a376ff24a787e.jpg) +Figure 1. Visual comparison of different strategies of adding controls. The compared baselines include 1) the diffusion with integrated standard ControlNet (denoted as ControlNet), and 2) the integration of pre-trained standard ControlNet with Diff-Instruct's pre-trained one-step generator (denoted as $\mathrm{DI} + \mathrm{ControlNet}$ ). Notably, our method not only maintains computational efficiency but also surpasses the visual quality achieved by the standard ControlNet approach. While the standard ControlNet approach relies heavily on high Classifier-Free Guidance (CFG) to achieve high-quality generation, this dependency might introduce unwanted artifacts in the final samples. + +![](images/6ee40dbef056b33d2061172ff63216b0fc5e6fe090ed76474b1aca56a7796ff8.jpg) + +# Abstract + +While diffusion distillation has enabled one-step generation through methods like Diff-Instruct [17] and Variational Score Distillation [34], adapting distilled models to emerging new controls – such as novel structural constraints or latest user preferences – remains challenging. Conventional approaches typically require modifying the base diffusion model and redistilling it – a process that is both computationally intensive and time-consuming. To address these challenges, we introduce Joint Distribution Matching (JDM), a novel approach that minimizes the reverse KL divergence between image-condition joint distributions. By deriving a tractable upper bound, JDM decouples fidelity learning from condition learning. This asymmetric distillation scheme enables our one-step student to handle con + +trols unknown to the teacher model and facilitates improved classifier-free guidance (CFG) usage and seamless integration of human feedback learning (HFL). Experimental results demonstrate that JDM surpasses baseline methods such as multi-step ControlNet [45] by mere one-step in most cases, while achieving state-of-the-art performance in one-step text-to-image synthesis through improved usage of CFG or HFL integration. + +# 1. Introduction + +Diffusion models (DMs) [9, 31] have significantly advanced generative modeling, particularly in text-to-image synthesis, by producing high-quality and diverse images in a controllable manner [45]. However, their practical deployment is often hindered by the inefficiency of the sampling process, which usually takes tens to hundreds of Network Function Evaluations (NFEs). Thanks to the recent progress in diffusion distillation, the sampling efficiency has been + +greatly enhanced, and photo-realistic images can be generated with as few as 1 NFE [17, 18, 42]. With huge computational savings in model serving, diffusion distillation of the base DM has become the standard procedure. + +As artificial intelligence-generated content (AIGC) applications continue to evolve, new scenarios demand models to adapt to novel conditions and controls. These conditions encompass structural constraints, semantic guidelines, and external factors such as user preferences and additional sensory inputs. The conventional approach to integrating such controls into diffusion models entails modifying the base model and subsequently performing diffusion distillation for one-step student [44] — a process that is both computationally expensive and time-intensive. A more efficient alternative would be to extend the distillation pipeline to accommodate new controls directly, bypassing the need for extensive retraining. + +Existing work on this important matter is scarce. On one hand, learning additional control for one-step students remains largely unexplored. Doing so for the base DM typically relies on pre-trained ControlNet models optimized through denoising score matching (DSM) to obtain new controllability [36, 45]. However, extending ControlNet for one-step generation incurs significant limitations, such as degraded fine-grained control and suboptimal sample quality, underscoring the need for new learning paradigms that better integrate control mechanisms into one-step generators. On the other hand, incorporating additional control during diffusion distillation is also challenging. Current diffusion distillation methodologies for one-step generation predominantly focus on distilling a student model that replicates the capabilities of the teacher diffusion model [17, 42, 44], without investigating how to extend the student's abilities beyond those of the teacher. This limitation is particularly relevant when adding novel controls that the original diffusion model was not designed to handle. + +To address these challenges, we propose a novel approach termed JDM, that minimizes the reverse Kullback-Leibler (KL) divergence between the image-condition joint distributions. We derive a tractable upper bound for this divergence, which effectively decouples fidelity learning from condition learning. The asymmetrical nature of our objective enables us to obtain a one-step student that can handle controls unknown to the teacher diffusion model. Moreover, this decoupling mechanism not only facilitates improved usage of classifier-free guidance (CFG), but also enables the seamless integration of human feedback learning (HFL) into the training process. Consequently, our method enhances both the controllability and quality of generated images, providing a more flexible and efficient framework for one-step diffusion generation. + +Extensive experiments demonstrate the superiority of our proposed JDM. For controllable generation tasks, our + +one-step approach achieves better performance than multi-step controllable DM (50 NFE), with lower FID scores in average (14.58 vs 15.21) and better controllability measured by consistency scores (improves $24\%$ in average). Besides, in text-to-image generation, by incorporating either human feedback learning or improved usage of CFG, our method establishes new state-of-the-art (SOTA) performance among one-step approaches. Specifically, our variant with better CFG achieves CLIP scores of 33.97, clearly outperforming the multi-step DM's 33.03. + +# 2. Background + +Diffusion models (DMs). DMs [9, 31] establish a forward diffusion process that progressively introduces Gaussian noise into the data over $T$ steps: $q(\mathbf{x}_t|\mathbf{x}) \triangleq \mathcal{N}(\mathbf{x}_t; \alpha_t\mathbf{x}, \sigma_t^2\mathbf{I})$ , where $\alpha_t$ and $\sigma_t$ are hyper-parameters that control the diffused schedule. The diffused samples can be directly calculated as $\mathbf{x}_t = \alpha_t\mathbf{x} + \sigma_t\epsilon$ , with $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The diffusion network $\epsilon_{\phi}$ is then trained to perform denoising by minimizing: $\mathbb{E}_{\mathbf{x},\epsilon,t}||\epsilon_{\phi}(\mathbf{x}_t,t) - \epsilon||_2^2$ . Once trained, the score of the diffused samples $\mathbf{x}_t$ can be estimated using: + +$$ +\nabla_ {\mathbf {x} _ {t}} \log p (\mathbf {x} _ {t}) \approx \nabla_ {\mathbf {x} _ {t}} \log p _ {\phi} (\mathbf {x} _ {t}) = s _ {\phi} (\mathbf {x} _ {t}, t) = - \frac {\epsilon_ {\phi} (\mathbf {x} _ {t} , t)}{\sigma_ {t}}, (1) +$$ + +With the estimated score, sampling from DMs can be achieved by solving the corresponding diffusion stochastic differential equations (SDEs) or probability flow ordinary differential equations (PF-ODEs) with multiple steps. + +ControlNet. Among other methods [1, 21], ControlNet [45] is the most promising method in adding additional Control to DMs. Given a pretrained diffusion model $\epsilon_{\phi}$ , a ControlNet parameterized by $\beta$ can be trained by minimizing the denoising loss $L(\beta)$ for injecting additional controls, where $L(\beta) = \mathbb{E}_{\mathbf{x},\epsilon ,t}||\epsilon -\epsilon_{\phi ,\beta}(\mathbf{x}_t,c)||_2^2$ . + +Diffusion distillation. Despite the active progress in training-free accelerated sampling of DMs [12, 19, 30, 39, 46], diffusion distillation is dispensable for satisfactory few-step generation. Diffusion distillation typically follows two main appealing approaches: : 1) Trajectory distillation [13, 13, 20, 27, 32, 33, 40], which attempts to replicate the teacher model's ODE trajectories on instances level. These methods face challenges in difficult instance-level matching; 2) Distribution matching via score distillation [17, 42, 48], which aims to replicate the teacher model on distribution level using distribution divergence metrics. + +Score Distillation. Diff-Instruct (DI) [17] and Variational Score Distillation (VSD) [34] train a conditional one-step student by minimizing the reverse integral KL divergence: + +$$ +\mathbb {E} _ {t} \lambda_ {t} \mathrm {K L} \left(p _ {\theta} \left(\mathbf {x} _ {t} \mid c\right) \| p \left(\mathbf {x} _ {t} \mid c\right)\right), \lambda_ {t} > 0 \tag {2} +$$ + +where $p_{\theta}(\mathbf{x}_t|c) \triangleq \int p_{\theta}(\mathbf{x}|c) q(\mathbf{x}_t|\mathbf{x}) d\mathbf{x}$ is the distribution of diffused sample. A student trained in this way can effectively replicate the teacher's capabilities while enabling + +one-step generation. However, training a conditional student requires a two-step process: first, a conditional teacher must be trained, followed by distillation to transfer its knowledge to the student. This sequential approach renders the introduction of new control guidance inefficient. + +Additional Controls for One-step Diffusion. Distilling one-step generator by score distillation has been well explored [17, 42, 48], however, how to distill one-step generator with additional controls has not been well explored. CCM [36] explores integrating consistency training with ControlNet, showing reasonable performance with four steps, while our work can surpass the standard ControlNet with mere one step in most cases. The success of previous work [17, 42, 43] in score distillation relies on initializing one-step students with the teacher. SDXS [44] explored learning one-step generator with control via score distillation, however, their teacher and fake score are required to have a ControlNet that supports the injected condition. In contrast, we minimize a tractable upper bound of joint KL divergence. This approach enables an asymmetric formulation between the teacher and student, where our student is partially initialized by the teacher, supplemented with an additional ControlNet. Our work demonstrate that it is possible to train a one-step student through score distillation, allowing it to handle conditions unknown to teacher. + +Human Preference Alignment For One-Step Diffusion. Since our framework supports universal guidance, we also explore integrating the reward model as additional guidance into training one-step generators for aligning with human preference. Although there are many works that try to align diffusion models with human preferences [2, 4-6, 10, 22, 23, 41], we note that how to align a one-step generator with human preference is not well explored. The existing few works [16, 25] directly maximize the reward of generated images, which can lead to obvious artifacts (see Fig. 6). In contrast, our method integrates preference learning with modeling $\log p(c|\mathbf{x}_t)$ , suffering less from artifacts and providing better visual quality. Besides, our frameworks decouple the condition and fidelity learning. This provides flexibility in applying CFG by an additional teacher model, showing better performance in distilling a one-step generator. + +# 3. Methodology + +Problem Setup. Consider a pre-trained DM with a multilevel score network $s_{\phi}(\mathbf{x}_t,t) = \nabla_{\mathbf{x}_t}\log p_{\phi}(\mathbf{x}_t,t)\approx \nabla_{\mathbf{x}_t}\log p_t(\mathbf{x}_t)$ , where $p_t(\mathbf{x}_t)$ represents the marginal diffusedistributions at time $t$ . We assume that this pre-trained model provides a high-quality approximation of the data distribution, such that $p(\mathbf{x}_0)\approx p_d$ . Additionally, we want to implement a new control $c$ given by a conditional discriminative model $\log p(c|\mathbf{x})$ . Our objective is to train a one-step generator that incorporates additional control $c$ through + +![](images/a127f32689b52919de94e84d10547aa0232b4aa40ea2ca582f7ad9082862a0e1.jpg) +Figure 2. The framework description of our proposed JDM. + +marginal diffusion. In essence, we aim to develop an algorithm that enables a student model to acquire new capabilities beyond those of the teacher diffusion model. + +In order to directly inject new controls in learning one-step student, we propose minimizing the joint reverse KL divergence between $p_{\theta}(\mathbf{x}_t,c) \triangleq p_{\theta}(\mathbf{x}_t|c)p(c)$ and $p(\mathbf{x},c) \triangleq p(c|\mathbf{x}_t)p_{\phi}(\mathbf{x}_t)$ , i.e., + +$$ +\begin{array}{l} \mathbb {E} _ {t} \lambda_ {t} \operatorname {K L} \left(p _ {\theta} \left(\mathbf {x} _ {t}, c\right) | | p \left(\mathbf {x} _ {t}, c\right)\right) \\ = \mathbb {E} _ {t} \lambda_ {t} \mathrm {K L} \left(p _ {\theta} \left(\mathbf {x} _ {t} \mid c\right) p (c) \mid \mid p (c \mid \mathbf {x} _ {t}) p _ {\phi} (\mathbf {x} _ {t})\right) \tag {3} \\ = \lambda_ {t} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {t} | c) p (c), t} [ - \log p (c | \mathbf {x} _ {t}) p _ {\phi} (\mathbf {x} _ {t}) + \log p _ {\theta} (\mathbf {x} _ {t}, c) ], \\ \end{array} +$$ + +where $\lambda_t > 0$ , $p(c)$ is a known fixed distribution. Notably, the joint KL divergence exhibits asymmetry between the target and student distributions. The target joint distribution factorizes into the marginal distribution $p_{\phi}(\mathbf{x})$ and the conditional distribution $p(c|\mathbf{x}_t)$ , with the latter being accessible through a discriminative model. This formulation enables us to distill a one-step conditional generative student that handles condition unknown to teacher generative model. + +However minimizing the KL divergence in Eq. (3) still requires access to the gradient of the conditional student. Luckily, we can access its tractable upper bound as follows. + +Lemma 3.1 Suppose the condition $c$ is discrete, a upper bound of Eq. (3) can be computed by: + +$$ +\begin{array}{l} \mathbb {E} _ {t} \lambda_ {t} \mathrm {K L} (p _ {\theta} (\mathbf {x} _ {t}, c) | | p (\mathbf {x} _ {t}, c)) \\ \leq \lambda_ {t} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {t} | c) p (c), t} \left[ - \log p (c | \mathbf {x} _ {t}) p _ {\phi} (\mathbf {x} _ {t}) + \log p _ {\theta} (\mathbf {x} _ {t}) \right] \tag {4} \\ \end{array} +$$ + +See proof in the Appendix A. The gradient of the upper bound for learning conditional generator $p_{\theta}(\mathbf{x}|c)$ can be computed as follows: + +$$ +\begin{array}{l} \operatorname {G r a d} (\theta) = - \alpha_ {t} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {t} | c) p (c), t} \lambda_ {t} [ \underbrace {\nabla_ {\mathbf {x} _ {t}} \log p (c | \mathbf {x} _ {t})} _ {\text {c o n d i t i o n l e a r i n g}} \\ + \underbrace {\nabla_ {\mathbf {x} _ {t}} \log \frac {p _ {\phi} (\mathbf {x} _ {t})}{p _ {\theta} (\mathbf {x} _ {t})}} _ {\text {f i d e l i t y l e a n i n g}} ] \frac {\partial \mathbf {x}}{\partial \theta}. \tag {5} \\ \end{array} +$$ + +We can approximates $\nabla_{\mathbf{x}_t}\log p_\theta (\mathbf{x}_t)$ by a score model $s_{\psi}(\mathbf{x}_t,t)$ , which is readily learnable through initialization via $s_{\phi}(\mathbf{x}_t,t)$ . Following GAN's tradition [7], we call $s_{\phi}$ as real score and $s_{\psi}$ as fake score. + +This upper bound naturally decomposes into two distinct learning components: conditional alignment, and generation fidelity. This decomposition contrasts with VSD and diff-instruct frameworks, where the fake score serves multiple purposes in both conditional learning and fidelity learning, potentially compromising effectiveness. Our approach reduces the burden on the fake score while eliminating the requirement for teachers to understand training conditions. + +Learning Fake Score. We employ an auxiliary diffusion model $s_{\psi}$ to model $\nabla_{\mathbf{x}_t}\log p_\theta (\mathbf{x}_t)$ . The fake score can be efficiently learned through denoising: + +$$ +\mathbb {E} _ {t, p (\epsilon), p _ {\theta} (\mathbf {x})} | | \epsilon_ {\psi} (\mathbf {x} _ {t}, t) - \epsilon | | _ {2} ^ {2}, \tag {6} +$$ + +where $\mathbf{x}_t = \alpha_t\mathbf{x} + \sigma_t\epsilon$ and $p(\epsilon)$ is the standard Gaussian distribution. After trained, we have $\nabla_{\mathbf{x}_t}\log p_{\theta ,t}(\mathbf{x}_t)\approx s_\psi (\mathbf{x}_t,t) = -\frac{\epsilon_\psi(\mathbf{x}_t,t)}{\sigma_t}$ . + +Modeling the $\log p(c|\mathbf{x}_t)$ . Unfortunately, conditional alignment density is defined in terms of clean samples in most cases, and rarely in terms of noisy samples. Hence, we need to find a way to approximate it. The $\log p(c|\mathbf{x}_t)$ can be modeled as follows: + +$$ +\begin{array}{l} p (c | \mathbf {x} _ {t}) \triangleq \int_ {c} p (c | \mathbf {x} _ {t}, \mathbf {x}) p (\mathbf {x} | \mathbf {x} _ {t}) d \mathbf {x} \tag {7} \\ = \int p (c | \mathbf {x}) p (\mathbf {x} | \mathbf {x} _ {t}) d \mathbf {x} \\ \end{array} +$$ + +where we substitute $p(c|\mathbf{x}_t,\mathbf{x})$ with $p(c|\mathbf{x})$ , since the condition $c$ is fully relied on $\mathbf{x}$ . The remaining challenge is how to model $p(\mathbf{x}|\mathbf{x}_t)$ . We propose to parameterize $p(\mathbf{x}|\mathbf{x}_t)$ using an implicit generator. An easy way can be directly parameterizing it with the fake score. However, the distribution is defined over clean samples, and using a fake score can not estimate the distribution accurately. Hence, we propose parameterizing it using a consistency model, which can be efficiently trained upon the fake score through LoRA finetuning. We note that modeling $p(c|\mathbf{x}_t)$ via consistency models by itself is interesting, how to integrate this technique with training-free controllable generation will be promising future work, however, it is beyond the scope of this work. + +Learning the $p(\mathbf{x}|\mathbf{x}_t)$ . The $p(\mathbf{x}|\mathbf{x}_t)$ is modeled by a consistency model, we suggest learning the model by inserting a LoRA over the fake score for efficiency. Specifically, the consistency model [33] can be efficiently trained through: + +$$ +\min _ {\beta} \mathbb {E} _ {k, p (\epsilon), p _ {\theta} (\mathbf {x})} | | f _ {\psi , \beta} (\mathbf {x} _ {t _ {k}}, t _ {k}) - \mathrm {s g} (f _ {\psi , \beta} (\mathbf {x} _ {t _ {k - 1}}, t _ {k - 1})) | | _ {2} ^ {2}, +$$ + +$$ +f _ {\psi , \beta} (\mathbf {x} _ {t}, t) \triangleq \left\{ \begin{array}{l l} \mathbf {x} _ {t}, & \text {i f} t = 0 \\ \frac {\mathbf {x} _ {t} - \sigma_ {t} \epsilon_ {\psi , \beta} (\mathbf {x} _ {t} , t)}{\alpha_ {t}}, & \text {i f} t > 0 \end{array} \right. \tag {8} +$$ + +where $t_k > t_{k-1}$ , $\beta$ denotes the parameters of a lightweight LoRA, and $\mathrm{sg}(\cdot)$ denotes the stop-gradient operator. + +Remark. Our learning framework supports universal guidance, as demonstrated through the following instances. + +# 3.1. Learning Better Aligned One-Step Generator + +Human Feedback Integration. We demonstrate that Human-Feedback Learning (HFL) can be seamlessly integrated into this framework. Specifically, we introduce "human-preferred images" as a single conditioning factor. Since we are dealing with only one condition, it is unnecessary to inject it into the desired generator. The learning gradient can then be expressed as: + +$$ +\begin{array}{l} \operatorname {G r a d} (\theta) = - \alpha_ {t} \mathbb {E} _ {p _ {\theta} \left(\mathbf {x} _ {t} \mid c\right) p (c), t} \lambda_ {t} \left[ \nabla_ {\mathbf {x} _ {t}} r \left(\mathbf {x} _ {t}\right) \right. \\ \left. + \nabla_ {\mathbf {x} _ {t}} \log \frac {p _ {\phi} (\mathbf {x} _ {t})}{p _ {\theta} (\mathbf {x} _ {t})} \right] \frac {\partial \mathbf {x}}{\partial \theta}, \tag {9} \\ \end{array} +$$ + +where $\nabla_{\mathbf{x}_t}r(\mathbf{x}_t)$ models the gradient of log probability of an image being human-preferred, formally defined as $\nabla_{\mathbf{x}_t}\log p(\text{"Human-preferred images"}|\mathbf{x}_t)\triangleq \nabla_{\mathbf{x}_t}r(\mathbf{x}_t)$ . + +Decoupled CFG. Our framework separates the learning of condition and fidelity into two distinct components. When the conditional probability $p(c|\mathbf{x}_t)$ represents text-image alignment, its gradient can be computed using Classifier-Free Guidance (CFG). While previous approaches [42] have incorporated CFG during distillation, however, their CFG is coupled with the real score. In contrast, our framework explicitly leverages CFG for conditional learning. This key difference allows us to compute CFG using a diffusion model different from the one used for real score computation, enabling the use of more sophisticated diffusion models to guide text-image alignment. + +# 3.2. Learning One-Step Generator with Additional Control + +Controllable Generation. To incorporate additional control capabilities similar to ControlNet [45], we parameterize the student model as a diffusion model with an associated ControlNet. The generator's learning objective becomes: + +$$ +\min _ {\theta , \beta} - \mathbb {E} _ {p _ {\theta , \beta} (\mathbf {x} _ {t} | c) p (c), t} \lambda_ {t} [ \log p (c | \mathbf {x} _ {t}) + \log \frac {p _ {\phi , t} (\mathbf {x} _ {t})}{p _ {\psi , t} (\mathbf {x} _ {t})} ], \tag {10} +$$ + +where $\beta$ denotes the additional parameters for ControlNet. This formulation enables asymmetric capability development - the student model can learn conditional generation tasks beyond the teacher's abilities. + +To the best of our knowledge, we are the first to explore learning additional control one-step generator by score distillation in an asymmetric form. To highlight the effectiveness of our proposed formulation, we discuss two potential competitive variants for learning one-step generators with additional control conditions. 1) Train ControlNet via Denoising: Given a pre-trained diffusion model $\epsilon_{\phi}(\mathbf{x}_t,t)$ , a ControlNet parameterized by $\beta$ can be trained by denoising score matching loss. Hence a naive idea is training the ControlNet for one-step generation using denoising loss too. + +![](images/cabc5a9d8872a46cdd005a89d572c505ff6b30f4f3ef9e5d72ec877c728dce76.jpg) +Control Signal +Implicit Training +Denoising Training +JDM(Ours) +Figure 3. The qualitative comparison of the proposed method and potential baselines in one-step controllable generation. + +Empirically, we find this approach can transfer the control well but generate blurry images (see Tab. 2 and Fig. 3); 2) Implicitly Train ControlNet via Score Distillation Training ControlNet for Diffusion Models is straightforward by directly using the original diffusion loss. Hence, a natural idea is to follow the training of ControlNet for DMs and directly inject the condition via ControlNet with the original VSD loss. This approach shares a high-level idea with training ControlNet for diffusion models: maintaining the original training loss while only adding ControlNet to the generator for condition injection. However, we empirically find this approach can generate high-quality images but ignores the control signal (see Tab. 2 and Fig. 3). + +# 3.2.1. Shared One-Step Generator Between Different Additional Control + +While Eq. (10) establishes the joint training of One-Step Generator $G_{\theta}$ and ControlNet $\phi$ , this approach necessitates retraining the entire One-Step Generator for each new control condition. This requirement leads to inefficient use of computational resources and storage space, which limits practical applications. + +A simple solution would be to first train a one-step generator without control using diff-instruct [17], then train ControlNet to incorporate additional control into $G_{\theta}$ . However, previous work has shown that one-step generators trained in this manner often suffer from mode collapse, making it difficult to accommodate additional control signals. One potential approach is to use teacher diffusion to generate millions of noise-image pairs, then add ODE regression loss for a one-step generator [42]. However, this method is extremely computationally inefficient and impractical for real-world applications. + +To address these limitations, we propose a novel two-phase warm-up training strategy: 1) Initial Phase: Joint training of $G_{\theta}$ and ControlNet for a primary condition; 2) Extension Phase: Fixing $G_{\theta}$ while training only ControlNet for subsequent conditions. This approach leverages the joint KL divergence to regularize $G_{\theta}$ during initial training, + +resulting in better accommodating the condition and avoiding mode collapse. The resulting well-trained $G_{\theta}$ can then effectively incorporate other forms of control. + +# 4. Experiments + +# 4.1. Controllable Generation + +Experiment Setting. All the models are trained in an internally collected dataset. We use Stable Diffusion 1.5 [26] as the frozen teacher model and initialization for the fake score model and student. The ControlNet is initialized in the way introduced in its original paper [45]. We utilize four conditions for evaluating our method in one-step controllable generation as follows: Canny [3], Hed [37], Depth map, and lower resolution samples. Details for these conditions can be found in the Appendix B. + +Evaluation Metric. We select the FID [8] to measure the image quality and consistency to measure the controllability. In particular, the FID is calculated between generated images by SD without additional control and generated images based on additional conditions. The consistency is defined by Consistecny $= ||h(\mathbf{x}) - c||_{1}$ , where $c$ is the condition, and $h(\cdot)$ is the function used for obtaining condition. We also report the number of function evaluations (NFE) required for generating an image for comparing efficiency. + +Quantitative Results. We conduct comprehensive evaluations comparing our proposed approach against two baseline methods: (1) the standard diffusion model (DM) with ControlNet and (2) pre-trained one-step generator by diff-instruct integrated with DM's ControlNet. The quantitative results are summarized in Tab. 1, where we evaluate both image quality (FID) and condition consistency across multiple control tasks. Our results reveal several key findings: 1) The proposed method achieves a significant reduction in the number of function evaluations (NFEs) from 50 to 1, while maintaining or improving performance metrics. Specifically, our method demonstrates superior FID scores and consistency measures across various conditioning tasks, indicating both better image quality and more precise condition adherence. 2) Our shared one-step generator architecture exhibits substantial improvements over the naive integration of a pre-trained one-step generator, validating the effectiveness of our unified training strategy. This is evidenced by consistent performance gains across all evaluated metrics and tasks. 3) While it is technically feasible to directly combine DM's ControlNet with a pre-trained one-step generator, this approach yields substantially inferior results. This observation underscores the importance of our proposed strategy tailored for training one-step generators with additional control. Overall, these results demonstrate that our method successfully achieves better trade-off between efficiency and sample quality in controlled image generation, achieving state-of-the-art performance with significantly reduced computational overhead + +Table 1. Comparison of machine metrics of different methods across tasks. + +
MethodNFE↓FID↓Canny Consistency↓FID↓HED Consistency↓FID↓Depth Consistency↓8× Super ResolutionFID↓Avg Consistency↓
FID↓Consistency↓
ControlNet5014.480.11319.210.10115.250.09311.930.06515.210.093
DI + ControlNet122.740.14128.040.11322.490.09715.570.12622.210.120
JDM (Ours)113.070.10216.510.04516.240.08112.480.05314.580.071
+ +Control Signal + +ControlNet, 50 NFE + +DI+Control Net, 1 NFE + +JDM(Ours), 1 NFE + +![](images/cca29ca67671377cd9f6af027a1d20526e5a37db60c44322f972b6722edb9bbf.jpg) +Figure 4. Qualitative comparisons on controllable generation across different control signals against competing methods. + +Control Signal + +ControlNet, 50 NFE + +DI+Control Net,1 NFE + +JDM(Ours), 1 NFE + +![](images/229a600e95d6d33214bdb82006387db45f65bf693a08eeac54791e75721d2fc1.jpg) + +Qualitative Comparison. We present the qualitative comparison in Fig. 4. We can observe that DM's ControlNet provides high-level control for one-step generators. However, this approach produces lower-quality images. Our customized one-step generator training with additional control can generate much higher-quality images. This validates the effectiveness of our proposed method and indicates that we can teach students conditions unknown to teachers by minimizing the upper bound of joint KL divergence. + +# 4.2. Comparison to Potential Baselines + +We conduct comprehensive studies to validate our design choices by comparing with two potential baseline approaches as discussed in Sec. 3.2: 1) training ControlNet via denoising loss; 2) implicitly training ControlNet via score distillation. The quantitative results are shown in Tab. 1. The quantitative results are shown in Fig. 3. + +Baseline Details. We introduce the training of two potential baselines as follows: + +- Train ControlNet via Denoising: Given a pre-trained + +one-step diffusion model $G_{\theta}$ , the ControlNet is trained by denoising: + +$$ +L (\beta) = \mathbb {E} _ {\mathbf {x}, \epsilon} | | G _ {\{\phi , \beta \}} (\mathbf {x} _ {T}, c) - \mathbf {x} | | _ {2} ^ {2}, \tag {11} +$$ + +* where $T$ is the terminal step in the one-step generator. + +- Implicitly Train ControlNet via score distillation: The one-step student with additional control is directly trained by the original VSD loss as following: + +$$ +\operatorname {K L} \left(\int_ {c} q _ {\theta , \beta} (\mathbf {x} | c) p (c) d c \mid \mid p _ {\phi} (\mathbf {x})\right). \tag {12} +$$ + +The condition injection is implicitly trained via reverse KL divergence. + +Comparing to Train ControlNet via Denoising. The baseline of training ControlNet through DSM loss achieves reasonable consistency scores (0.109 for Canny and 0.094 for Depth) but suffers from significantly degraded image quality, as reflected by the worse FID scores (28.83 for Canny and 35.14 for Depth). This reveals that this approach tends to generate blurry images while maintaining decent + +![](images/9ef4fbc53e9191a79d473c22101dfde5bb1667032e72c26478744a2415423389.jpg) + +control. The degradation in image quality can be attributed to the mismatch between the denoising objective and the reverse KL divergence, as the model is forced to learn denoising behavior that may not be optimal for direct generation. + +Comparing to Implicitly Train ControlNet via score distillation. The implicit training baseline through score distillation shows better FID scores (22.87 for Canny and 22.34 for Depth) compared to the DSM approach, indicating its capability to generate higher quality images. However, it demonstrates worse consistency scores (0.151 for Canny and 0.150 for Depth), suggesting that the control signals are not effectively incorporated. This indicates that simply maintaining the original VSD loss while adding ControlNet may lead to the model ignoring the control conditions. + +# 4.3. Other Application in Text-to-Image Generation + +Evaluation. We employs multiple metrics to assess different aspects of generation quality: Aesthetic Score (AeS) [29] evaluates the image quality; CLIP Score (CS) measures the text-to-image alignment; Human Preference Score (HPS) v2.1 [35] correlates strongly with the human preference, capturing both image-text alignment and aesthetic quality. + +Table 2. Comparison of machine metrics of different methods across tasks. + +
MethodNFE↓FID↓Canny Consistency↓FID↓Depth Consistency↓
JDM (Ours)113.070.10216.240.081
w/ warm-up shared UNet114.350.12216.710.093
w/o modeling p(c|xt)114.890.11318.560.085
Implicit Training122.870.15122.340.150
DI w/ Denoising128.830.10935.140.094
+ +Baseline Models. We conduct our experiments on SD-v1.5. We perform human feedback learning by ImageReward [38]. To ensure fairness in evaluation, we do not report this metric. We perform decoupled CFG by using SD-v2.1 in distilling SD-v1.5. We mainly compare our model against the open-source state-of-the-art (SOTA) models, e.g., LCM [14], Hyper-SD [25], and DMD2 [43]. + +Quantitative Results. Table 3 presents a comprehensive comparison of our method against existing approaches. Our one-step generator, enhanced with SD-v2.1-based CFG and human feedback learning, demonstrates superior performance across all evaluation metrics. Notably, our method significantly outperforms the direct baselines such as Diff- + +Table 3. Comparison of machine metrics on text-to-image generation across state-of-the-art methods. HFL denotes human feedback learning which might hack the machine metrics. We highlight the best and second best among distillation methods. + +
ModelBackboneHFLStepsHPS↑Aes↑CS↑
AnimationConcept-ArtPaintingPhotoAverage
Base ModelSD-v1.5No2526.2924.8524.8726.0125.505.4933.03
Base ModelSD-v2.1No2527.8227.1427.1728.1727.585.6633.46
SD Turbo [28]SD 2.1No128.3026.9226.4325.5426.805.3132.21
InstaFlow [11]SD-v1.5No123.1723.0422.7322.9722.985.2531.97
TCD [47]SD-v1.5No423.1421.1121.0823.6222.245.4329.07
LCM-dreamshaper [15]SD-v1.5No426.5126.4025.9624.3225.805.9431.55
PeRFlow [40]SD-v1.5No422.7922.1721.2823.5022.435.3530.77
DMD2 [43]SD-v1.5No124.1722.6822.9724.3023.535.8230.92
Diff Instruct [17]SD-v1.5No127.3226.1526.4125.5026.355.7132.08
Hyper-SD [25]SD-v1.5Yes128.6528.1628.4126.9028.015.6430.87
JDM w/ HFL (Ours)SD-v1.5Yes130.1629.1730.1428.3529.465.8933.75
JDM w/ better CFG (Ours)SD-v1.5No130.5629.4630.3828.5929.755.9033.97
+ +![](images/1807028b1a6b7bc2baa42845ececf527a4ead7481c30891a3e972ac48612324f.jpg) +Figure 6. We compare JDM with the variant directly using $\log p(c|\mathbf{x})$ . It is clear that the variant suffers from artifacts. + +Instruct [17] and DMD2 [43]. An intriguing observation is that the variant utilizing better CFG achieves even better metrics compared to the HFL variant. This unexpected finding suggests a promising research direction: the potential benefits of leveraging multiple teacher diffusion models for student model distillation, which merits further investigation in future work. + +Qualitative Comparison. Since using HFL may lead to the hack of machine metrics, we further conducted qualitative comparisons. We qualitatively compared our w/ HFL and our w/ better CFG with the most competitive baseline, as shown in Fig. 5. Our method demonstrates significantly better visual quality and text-image alignment. Notably, while Hyper-SD also employs HFL, their HFL is performed separately, whereas ours is conducted collaboratively with one-step learning, allowing fake scores to participate and eliminate artifacts caused by reward maximization. As shown in Fig. 5, Hyper-SD exhibits noticeable artifacts, likely due to their flawed HFL approach. + +# 4.4. Additional Ablation Study + +Effect of Warm-up Shared UNet The proposed two-phase warm-up training strategy is crucial for learning a + +shared one-step UNet among different controls. Here, we compare our strategy to naively using pre-trained DI as shared UNet. The results are shown in Tab. 2. Without the warm-up strategy, the performance will degrade severely, especially regarding the consistency metric. This indicates our warm-up strategy can help in learning a One-step UNet that can handle multiple conditions well. + +Effect of Modeling $\log p(c|\mathbf{x}_t)$ Modeling $\log p(c|\mathbf{x}_t)$ not only makes the formulation more elegant but also ensures that the information from the gradients of condition learning and fidelity pertains to the same diffused $\mathbf{x}_t$ . If we were to simply substitute $\log p(c|\mathbf{x}_t)$ with $\log p(c|\mathbf{x}_0)$ , the gradient information would relate separately to $\mathbf{x}_t$ and $\mathbf{x}_0$ , resulting in less stable learning. As demonstrated in Tab. 2, omitting the modeling of $\log p(c|\mathbf{x}_t)$ leads to a significant drop in performance, particularly concerning the FID. This highlights the importance of modeling $\log p(c|\mathbf{x}_t)$ for achieving high-quality controllable one-step generation. Furthermore, modeling $\log p(c|\mathbf{x}_t)$ is more robust than using $\log p(c|\mathbf{x}_0)$ . In HFL learning scenarios, directly modeling $\log p(c|\mathbf{x}_0)$ can result in noticeable artifacts caused by reward hacking, as illustrated in Fig. 6. + +# 5. Conclusion + +In this work, we propose JDM for adding new control unknown to teacher DMs into one-step student. Our method minimizes the upper bound of reverse KL divergence between image-condition joint distributions. This approach decouples fidelity and condition learning, allowing the one-step student to handle controls unknown to the teacher. Extensive experiments show that JDM outperforms multi-step controllable DMs by one-step, while achieving SOTA performance in one-step text-to-image synthesis by the integration of decoupled CFG or human feedback learning. + +# Acknowledgments + +Jing Tang's work is partially supported by National Key R&D Program of China under Grant No. 2024YFA1012700 and No. 2023YFF0725100, by the National Natural Science Foundation of China (NSFC) under Grant No. 62402410 and No. U22B2060, by Guangdong Provincial Project (No. 2023QN10X025), by Guangdong Basic and Applied Basic Research Foundation under Grant No. 2023A1515110131, by Guangzhou Municipal Science and Technology Bureau under Grant No. 2024A04J4454, by Guangzhou Municipal Education Bureau (No. 2024312263), by Nansha District Project (No. 2023ZD022), and by Guangzhou Industrial Information and Intelligent Key Laboratory Project (No. 2024A03J0628) and Guangzhou Municipal Key Laboratory of Financial Technology Cutting-Edge Research (No. 2024A03J0630). + +# References + +[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Roni Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3 +[3] John Canny. A computational approach to edge detection. PAMI, 1986. 5, 1 +[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3 +[5] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023. +[6] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3 +[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 +[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 + +[10] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3 +[11] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 8 +[12] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2 +[13] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2 +[14] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 7 +[15] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 8 +[16] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. Transactions on Machine Learning Research, 2025. 3 +[17] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3, 5, 8 +[18] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans, 2024. 2 +[19] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 2 +[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14297-14306, 2023. 2 +[21] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2 +[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3 +[23] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3 + +[24] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1 +[25] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis, 2024. 3, 7, 8 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5 +[27] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2 +[28] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation, 2023. 8 +[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7 +[30] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2 +[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2 +[32] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2, 4 +[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2 +[35] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7 +[36] Jie Xiao, Kai Zhu, Han Zhang, Zhiheng Liu, Yujun Shen, Zhantao Yang, Ruili Feng, Yu Liu, Xueyang Fu, and Zheng-Jun Zha. CCM: Real-time controllable visual content creation using text-to-image consistency models. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3 +[37] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, 2015. 5 +[38] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere- + +ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7 +[39] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffusion sampling with optimized time steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292-8301, 2024. 2 +[40] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 2, 8 +[41] Zilyu Ye, Zhiyang Chen, Tiancheng Li, Zemin Huang, Wei-jian Luo, and Guo-Jun Qi. Schedule on the fly: Diffusion time prediction for faster and better image generation. arXiv preprint arXiv:2412.01243, 2024. 3 +[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3, 4, 5 +[43] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis, 2024. 3, 7, 8 +[44] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 2, 3 +[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 5 +[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36:49842-49869, 2023. 2 +[47] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation, 2024. 8 +[48] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In International Conference on Machine Learning, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/images.zip b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a98f8bd7e55300c226b24446e26b737f2467f756 --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10306d7e11da893ec2ca68a16ebee82f63d70047d62d86baa49515849bf4b011 +size 887527 diff --git a/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/layout.json b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5115965486c613f74e855b0513c76d3066a6420c --- /dev/null +++ b/2025/Adding Additional Control to One-Step Diffusion with Joint Distribution Matching/layout.json @@ -0,0 +1,8690 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "type": "text", + "content": "Adding Additional Control to One-Step Diffusion with Joint Distribution Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 160, + 240, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 160, + 240, + 175 + ], + "spans": [ + { + "bbox": [ + 171, + 160, + 240, + 175 + ], + "type": "text", + "content": "Yihong Luo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 175, + 246, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 175, + 246, + 190 + ], + "spans": [ + { + "bbox": [ + 173, + 175, + 246, + 190 + ], + "type": "text", + "content": "Jiacheng Sun" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 261, + 161, + 333, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 161, + 333, + 175 + ], + "spans": [ + { + "bbox": [ + 261, + 161, + 333, + 175 + ], + "type": "text", + "content": "Tianyang Hu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 267, + 175, + 333, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 267, + 175, + 333, + 190 + ], + "spans": [ + { + "bbox": [ + 267, + 175, + 333, + 190 + ], + "type": "text", + "content": "Zhenguo Li" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 353, + 161, + 418, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 161, + 418, + 175 + ], + "spans": [ + { + "bbox": [ + 353, + 161, + 418, + 175 + ], + "type": "text", + "content": "Yifan Song" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 175, + 421, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 175, + 421, + 190 + ], + "spans": [ + { + "bbox": [ + 354, + 175, + 421, + 190 + ], + "type": "text", + "content": "Jing Tang " + }, + { + "bbox": [ + 354, + 175, + 421, + 190 + ], + "type": "inline_equation", + "content": "^{3,1*}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 153, + 193, + 204, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 193, + 204, + 207 + ], + "spans": [ + { + "bbox": [ + 153, + 193, + 204, + 207 + ], + "type": "text", + "content": "1 HKUST" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 220, + 194, + 348, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 220, + 194, + 348, + 208 + ], + "spans": [ + { + "bbox": [ + 220, + 194, + 348, + 208 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 220, + 194, + 348, + 208 + ], + "type": "text", + "content": " Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 194, + 440, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 194, + 440, + 209 + ], + "spans": [ + { + "bbox": [ + 365, + 194, + 440, + 209 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 365, + 194, + 440, + 209 + ], + "type": "text", + "content": " HKUST(GZ)" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 57, + 249, + 361, + 425 + ], + "blocks": [ + { + "bbox": [ + 57, + 249, + 361, + 425 + ], + "lines": [ + { + "bbox": [ + 57, + 249, + 361, + 425 + ], + "spans": [ + { + "bbox": [ + 57, + 249, + 361, + 425 + ], + "type": "image", + "image_path": "2dcb62db6757db3a3ecb22438db114f064f7959f3f1f2993f02a376ff24a787e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 433, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 433, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. Visual comparison of different strategies of adding controls. The compared baselines include 1) the diffusion with integrated standard ControlNet (denoted as ControlNet), and 2) the integration of pre-trained standard ControlNet with Diff-Instruct's pre-trained one-step generator (denoted as " + }, + { + "bbox": [ + 54, + 433, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathrm{DI} + \\mathrm{ControlNet}" + }, + { + "bbox": [ + 54, + 433, + 555, + 489 + ], + "type": "text", + "content": "). Notably, our method not only maintains computational efficiency but also surpasses the visual quality achieved by the standard ControlNet approach. While the standard ControlNet approach relies heavily on high Classifier-Free Guidance (CFG) to achieve high-quality generation, this dependency might introduce unwanted artifacts in the final samples." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 370, + 255, + 555, + 422 + ], + "blocks": [ + { + "bbox": [ + 370, + 255, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 370, + 255, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 370, + 255, + 555, + 422 + ], + "type": "image", + "image_path": "6ee40dbef056b33d2061172ff63216b0fc5e6fe090ed76474b1aca56a7796ff8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 495, + 200, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 495, + 200, + 507 + ], + "spans": [ + { + "bbox": [ + 151, + 495, + 200, + 507 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 521, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 297, + 689 + ], + "type": "text", + "content": "While diffusion distillation has enabled one-step generation through methods like Diff-Instruct [17] and Variational Score Distillation [34], adapting distilled models to emerging new controls – such as novel structural constraints or latest user preferences – remains challenging. Conventional approaches typically require modifying the base diffusion model and redistilling it – a process that is both computationally intensive and time-consuming. To address these challenges, we introduce Joint Distribution Matching (JDM), a novel approach that minimizes the reverse KL divergence between image-condition joint distributions. By deriving a tractable upper bound, JDM decouples fidelity learning from condition learning. This asymmetric distillation scheme enables our one-step student to handle con" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 495, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 593 + ], + "type": "text", + "content": "trols unknown to the teacher model and facilitates improved classifier-free guidance (CFG) usage and seamless integration of human feedback learning (HFL). Experimental results demonstrate that JDM surpasses baseline methods such as multi-step ControlNet [45] by mere one-step in most cases, while achieving state-of-the-art performance in one-step text-to-image synthesis through improved usage of CFG or HFL integration." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 602, + 395, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 602, + 395, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 602, + 395, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Diffusion models (DMs) [9, 31] have significantly advanced generative modeling, particularly in text-to-image synthesis, by producing high-quality and diverse images in a controllable manner [45]. However, their practical deployment is often hindered by the inefficiency of the sampling process, which usually takes tens to hundreds of Network Function Evaluations (NFEs). Thanks to the recent progress in diffusion distillation, the sampling efficiency has been" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 240, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 240, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 240, + 713 + ], + "type": "text", + "content": "*Corresponding Author: Jing Tang, jingtang@ust.hk." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4009" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "greatly enhanced, and photo-realistic images can be generated with as few as 1 NFE [17, 18, 42]. With huge computational savings in model serving, diffusion distillation of the base DM has become the standard procedure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 277 + ], + "type": "text", + "content": "As artificial intelligence-generated content (AIGC) applications continue to evolve, new scenarios demand models to adapt to novel conditions and controls. These conditions encompass structural constraints, semantic guidelines, and external factors such as user preferences and additional sensory inputs. The conventional approach to integrating such controls into diffusion models entails modifying the base model and subsequently performing diffusion distillation for one-step student [44] — a process that is both computationally expensive and time-intensive. A more efficient alternative would be to extend the distillation pipeline to accommodate new controls directly, bypassing the need for extensive retraining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 506 + ], + "type": "text", + "content": "Existing work on this important matter is scarce. On one hand, learning additional control for one-step students remains largely unexplored. Doing so for the base DM typically relies on pre-trained ControlNet models optimized through denoising score matching (DSM) to obtain new controllability [36, 45]. However, extending ControlNet for one-step generation incurs significant limitations, such as degraded fine-grained control and suboptimal sample quality, underscoring the need for new learning paradigms that better integrate control mechanisms into one-step generators. On the other hand, incorporating additional control during diffusion distillation is also challenging. Current diffusion distillation methodologies for one-step generation predominantly focus on distilling a student model that replicates the capabilities of the teacher diffusion model [17, 42, 44], without investigating how to extend the student's abilities beyond those of the teacher. This limitation is particularly relevant when adding novel controls that the original diffusion model was not designed to handle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 687 + ], + "type": "text", + "content": "To address these challenges, we propose a novel approach termed JDM, that minimizes the reverse Kullback-Leibler (KL) divergence between the image-condition joint distributions. We derive a tractable upper bound for this divergence, which effectively decouples fidelity learning from condition learning. The asymmetrical nature of our objective enables us to obtain a one-step student that can handle controls unknown to the teacher diffusion model. Moreover, this decoupling mechanism not only facilitates improved usage of classifier-free guidance (CFG), but also enables the seamless integration of human feedback learning (HFL) into the training process. Consequently, our method enhances both the controllability and quality of generated images, providing a more flexible and efficient framework for one-step diffusion generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Extensive experiments demonstrate the superiority of our proposed JDM. For controllable generation tasks, our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "one-step approach achieves better performance than multi-step controllable DM (50 NFE), with lower FID scores in average (14.58 vs 15.21) and better controllability measured by consistency scores (improves " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " in average). Besides, in text-to-image generation, by incorporating either human feedback learning or improved usage of CFG, our method establishes new state-of-the-art (SOTA) performance among one-step approaches. Specifically, our variant with better CFG achieves CLIP scores of 33.97, clearly outperforming the multi-step DM's 33.03." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 392, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 392, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 392, + 212 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": "Diffusion models (DMs). DMs [9, 31] establish a forward diffusion process that progressively introduces Gaussian noise into the data over " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " steps: " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t|\\mathbf{x}) \\triangleq \\mathcal{N}(\\mathbf{x}_t; \\alpha_t\\mathbf{x}, \\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\alpha_t" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " are hyper-parameters that control the diffused schedule. The diffused samples can be directly calculated as " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ". The diffusion network " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " is then trained to perform denoising by minimizing: " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{x},\\epsilon,t}||\\epsilon_{\\phi}(\\mathbf{x}_t,t) - \\epsilon||_2^2" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": ". Once trained, the score of the diffused samples " + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 216, + 555, + 324 + ], + "type": "text", + "content": " can be estimated using:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 333, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 333, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 318, + 333, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\mathbf {x} _ {t}} \\log p (\\mathbf {x} _ {t}) \\approx \\nabla_ {\\mathbf {x} _ {t}} \\log p _ {\\phi} (\\mathbf {x} _ {t}) = s _ {\\phi} (\\mathbf {x} _ {t}, t) = - \\frac {\\epsilon_ {\\phi} (\\mathbf {x} _ {t} , t)}{\\sigma_ {t}}, (1)", + "image_path": "5060408177ec67a0ae0f0ed74489bd4111d232ce7a43695e17ba1a6d9e36de63.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 553, + 405 + ], + "type": "text", + "content": "With the estimated score, sampling from DMs can be achieved by solving the corresponding diffusion stochastic differential equations (SDEs) or probability flow ordinary differential equations (PF-ODEs) with multiple steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "content": "ControlNet. Among other methods [1, 21], ControlNet [45] is the most promising method in adding additional Control to DMs. Given a pretrained diffusion model " + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}" + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "content": ", a ControlNet parameterized by " + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "content": " can be trained by minimizing the denoising loss " + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "inline_equation", + "content": "L(\\beta)" + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "content": " for injecting additional controls, where " + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "inline_equation", + "content": "L(\\beta) = \\mathbb{E}_{\\mathbf{x},\\epsilon ,t}||\\epsilon -\\epsilon_{\\phi ,\\beta}(\\mathbf{x}_t,c)||_2^2" + }, + { + "bbox": [ + 313, + 406, + 553, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 479, + 554, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 609 + ], + "type": "text", + "content": "Diffusion distillation. Despite the active progress in training-free accelerated sampling of DMs [12, 19, 30, 39, 46], diffusion distillation is dispensable for satisfactory few-step generation. Diffusion distillation typically follows two main appealing approaches: : 1) Trajectory distillation [13, 13, 20, 27, 32, 33, 40], which attempts to replicate the teacher model's ODE trajectories on instances level. These methods face challenges in difficult instance-level matching; 2) Distribution matching via score distillation [17, 42, 48], which aims to replicate the teacher model on distribution level using distribution divergence metrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 610, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 553, + 647 + ], + "type": "text", + "content": "Score Distillation. Diff-Instruct (DI) [17] and Variational Score Distillation (VSD) [34] train a conditional one-step student by minimizing the reverse integral KL divergence:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 655, + 553, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 655, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 365, + 655, + 553, + 668 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) \\| p \\left(\\mathbf {x} _ {t} \\mid c\\right)\\right), \\lambda_ {t} > 0 \\tag {2}", + "image_path": "3e406f4d8dea89fd777adb407e3d5476024ffe3e9f3f0130e390ac5e00ee041e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_t|c) \\triangleq \\int p_{\\theta}(\\mathbf{x}|c) q(\\mathbf{x}_t|\\mathbf{x}) d\\mathbf{x}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " is the distribution of diffused sample. A student trained in this way can effectively replicate the teacher's capabilities while enabling" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4010" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "one-step generation. However, training a conditional student requires a two-step process: first, a conditional teacher must be trained, followed by distillation to transfer its knowledge to the student. This sequential approach renders the introduction of new control guidance inefficient." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 373 + ], + "type": "text", + "content": "Additional Controls for One-step Diffusion. Distilling one-step generator by score distillation has been well explored [17, 42, 48], however, how to distill one-step generator with additional controls has not been well explored. CCM [36] explores integrating consistency training with ControlNet, showing reasonable performance with four steps, while our work can surpass the standard ControlNet with mere one step in most cases. The success of previous work [17, 42, 43] in score distillation relies on initializing one-step students with the teacher. SDXS [44] explored learning one-step generator with control via score distillation, however, their teacher and fake score are required to have a ControlNet that supports the injected condition. In contrast, we minimize a tractable upper bound of joint KL divergence. This approach enables an asymmetric formulation between the teacher and student, where our student is partially initialized by the teacher, supplemented with an additional ControlNet. Our work demonstrate that it is possible to train a one-step student through score distillation, allowing it to handle conditions unknown to teacher." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 374, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 579 + ], + "type": "text", + "content": "Human Preference Alignment For One-Step Diffusion. Since our framework supports universal guidance, we also explore integrating the reward model as additional guidance into training one-step generators for aligning with human preference. Although there are many works that try to align diffusion models with human preferences [2, 4-6, 10, 22, 23, 41], we note that how to align a one-step generator with human preference is not well explored. The existing few works [16, 25] directly maximize the reward of generated images, which can lead to obvious artifacts (see Fig. 6). In contrast, our method integrates preference learning with modeling " + }, + { + "bbox": [ + 55, + 374, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 374, + 295, + 579 + ], + "type": "text", + "content": ", suffering less from artifacts and providing better visual quality. Besides, our frameworks decouple the condition and fidelity learning. This provides flexibility in applying CFG by an additional teacher model, showing better performance in distilling a one-step generator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 138, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 138, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 138, + 601 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Problem Setup. Consider a pre-trained DM with a multilevel score network " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "s_{\\phi}(\\mathbf{x}_t,t) = \\nabla_{\\mathbf{x}_t}\\log p_{\\phi}(\\mathbf{x}_t,t)\\approx \\nabla_{\\mathbf{x}_t}\\log p_t(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "p_t(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " represents the marginal diffusedistributions at time " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". We assume that this pre-trained model provides a high-quality approximation of the data distribution, such that " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)\\approx p_d" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". Additionally, we want to implement a new control " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " given by a conditional discriminative model " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". Our objective is to train a one-step generator that incorporates additional control " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " through" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 555, + 190 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 555, + 190 + ], + "type": "image", + "image_path": "a127f32689b52919de94e84d10547aa0232b4aa40ea2ca582f7ad9082862a0e1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 194, + 543, + 206 + ], + "lines": [ + { + "bbox": [ + 324, + 194, + 543, + 206 + ], + "spans": [ + { + "bbox": [ + 324, + 194, + 543, + 206 + ], + "type": "text", + "content": "Figure 2. The framework description of our proposed JDM." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 247 + ], + "type": "text", + "content": "marginal diffusion. In essence, we aim to develop an algorithm that enables a student model to acquire new capabilities beyond those of the teacher diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "text", + "content": "In order to directly inject new controls in learning one-step student, we propose minimizing the joint reverse KL divergence between " + }, + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_t,c) \\triangleq p_{\\theta}(\\mathbf{x}_t|c)p(c)" + }, + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x},c) \\triangleq p(c|\\mathbf{x}_t)p_{\\phi}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 248, + 553, + 297 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 303, + 553, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 303, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 321, + 303, + 553, + 344 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\operatorname {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t}, c\\right) | | p \\left(\\mathbf {x} _ {t}, c\\right)\\right) \\\\ = \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} \\left(p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c) \\mid \\mid p (c \\mid \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t})\\right) \\tag {3} \\\\ = \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} [ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}, c) ], \\\\ \\end{array}", + "image_path": "fdf4057bdde5d123835d8a718d4e90212fb523cc33730d8913ad2c2b68413fae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\lambda_t > 0" + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "inline_equation", + "content": "p(c)" + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "content": " is a known fixed distribution. Notably, the joint KL divergence exhibits asymmetry between the target and student distributions. The target joint distribution factorizes into the marginal distribution " + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "inline_equation", + "content": "p_{\\phi}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "content": " and the conditional distribution " + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 350, + 554, + 445 + ], + "type": "text", + "content": ", with the latter being accessible through a discriminative model. This formulation enables us to distill a one-step conditional generative student that handles condition unknown to teacher generative model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 481 + ], + "type": "text", + "content": "However minimizing the KL divergence in Eq. (3) still requires access to the gradient of the conditional student. Luckily, we can access its tractable upper bound as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 488, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 513 + ], + "type": "text", + "content": "Lemma 3.1 Suppose the condition " + }, + { + "bbox": [ + 313, + 488, + 554, + 513 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 488, + 554, + 513 + ], + "type": "text", + "content": " is discrete, a upper bound of Eq. (3) can be computed by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 518, + 553, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 518, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 328, + 518, + 553, + 545 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbb {E} _ {t} \\lambda_ {t} \\mathrm {K L} (p _ {\\theta} (\\mathbf {x} _ {t}, c) | | p (\\mathbf {x} _ {t}, c)) \\\\ \\leq \\lambda_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\left[ - \\log p (c | \\mathbf {x} _ {t}) p _ {\\phi} (\\mathbf {x} _ {t}) + \\log p _ {\\theta} (\\mathbf {x} _ {t}) \\right] \\tag {4} \\\\ \\end{array}", + "image_path": "27d97b0c4d11a21a6aecebf4dd003a04333310a5b3095e1fd146382b1fa2df4c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 554, + 586 + ], + "type": "text", + "content": "See proof in the Appendix A. The gradient of the upper bound for learning conditional generator " + }, + { + "bbox": [ + 313, + 551, + 554, + 586 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}|c)" + }, + { + "bbox": [ + 313, + 551, + 554, + 586 + ], + "type": "text", + "content": " can be computed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 593, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 593, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 320, + 593, + 553, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log p (c | \\mathbf {x} _ {t})} _ {\\text {c o n d i t i o n l e a r i n g}} \\\\ + \\underbrace {\\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})}} _ {\\text {f i d e l i t y l e a n i n g}} ] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}. \\tag {5} \\\\ \\end{array}", + "image_path": "674ddd3614d0ae31141d8bbf9baf61d2ba1b44374fb9406935d04d2c31c0d9af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "We can approximates " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " by a score model " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "s_{\\psi}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", which is readily learnable through initialization via " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "s_{\\phi}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ". Following GAN's tradition [7], we call " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "s_{\\phi}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " as real score and " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "s_{\\psi}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " as fake score." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4011" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "This upper bound naturally decomposes into two distinct learning components: conditional alignment, and generation fidelity. This decomposition contrasts with VSD and diff-instruct frameworks, where the fake score serves multiple purposes in both conditional learning and fidelity learning, potentially compromising effectiveness. Our approach reduces the burden on the fake score while eliminating the requirement for teachers to understand training conditions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "text", + "content": "Learning Fake Score. We employ an auxiliary diffusion model " + }, + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "inline_equation", + "content": "s_{\\psi}" + }, + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "text", + "content": " to model " + }, + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p_\\theta (\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 168, + 296, + 205 + ], + "type": "text", + "content": ". The fake score can be efficiently learned through denoising:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 212, + 295, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 212, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 295, + 228 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | \\epsilon_ {\\psi} (\\mathbf {x} _ {t}, t) - \\epsilon | | _ {2} ^ {2}, \\tag {6}", + "image_path": "52428da9f895f5a057ca09b56a05d964e946a94a1d1baf72c12f003a3c014c7f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t = \\alpha_t\\mathbf{x} + \\sigma_t\\epsilon" + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "inline_equation", + "content": "p(\\epsilon)" + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "text", + "content": " is the standard Gaussian distribution. After trained, we have " + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p_{\\theta ,t}(\\mathbf{x}_t)\\approx s_\\psi (\\mathbf{x}_t,t) = -\\frac{\\epsilon_\\psi(\\mathbf{x}_t,t)}{\\sigma_t}" + }, + { + "bbox": [ + 55, + 235, + 296, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "text", + "content": "Modeling the " + }, + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "text", + "content": ". Unfortunately, conditional alignment density is defined in terms of clean samples in most cases, and rarely in terms of noisy samples. Hence, we need to find a way to approximate it. The " + }, + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 332 + ], + "type": "text", + "content": " can be modeled as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 338, + 295, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 338, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 338, + 295, + 389 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (c | \\mathbf {x} _ {t}) \\triangleq \\int_ {c} p (c | \\mathbf {x} _ {t}, \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\tag {7} \\\\ = \\int p (c | \\mathbf {x}) p (\\mathbf {x} | \\mathbf {x} _ {t}) d \\mathbf {x} \\\\ \\end{array}", + "image_path": "780e961c036c2a2ae42cd2936e598cb097017d15830196e8cfbcbe69604d6deb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": "where we substitute " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x}_t,\\mathbf{x})" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": ", since the condition " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": " is fully relied on " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": ". The remaining challenge is how to model " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": ". We propose to parameterize " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": " using an implicit generator. An easy way can be directly parameterizing it with the fake score. However, the distribution is defined over clean samples, and using a fake score can not estimate the distribution accurately. Hence, we propose parameterizing it using a consistency model, which can be efficiently trained upon the fake score through LoRA finetuning. We note that modeling " + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 397, + 296, + 552 + ], + "type": "text", + "content": " via consistency models by itself is interesting, how to integrate this technique with training-free controllable generation will be promising future work, however, it is beyond the scope of this work." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": "Learning the " + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 552, + 296, + 601 + ], + "type": "text", + "content": " is modeled by a consistency model, we suggest learning the model by inserting a LoRA over the fake score for efficiency. Specifically, the consistency model [33] can be efficiently trained through:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 607, + 290, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 607, + 290, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 607, + 290, + 624 + ], + "type": "interline_equation", + "content": "\\min _ {\\beta} \\mathbb {E} _ {k, p (\\epsilon), p _ {\\theta} (\\mathbf {x})} | | f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k}}, t _ {k}) - \\mathrm {s g} (f _ {\\psi , \\beta} (\\mathbf {x} _ {t _ {k - 1}}, t _ {k - 1})) | | _ {2} ^ {2},", + "image_path": "f272e382933967cfba8b112e0a990159526d0ab4b0b54eaf8ca6f618166f472f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 628, + 296, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 628, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 62, + 628, + 296, + 665 + ], + "type": "interline_equation", + "content": "f _ {\\psi , \\beta} (\\mathbf {x} _ {t}, t) \\triangleq \\left\\{ \\begin{array}{l l} \\mathbf {x} _ {t}, & \\text {i f} t = 0 \\\\ \\frac {\\mathbf {x} _ {t} - \\sigma_ {t} \\epsilon_ {\\psi , \\beta} (\\mathbf {x} _ {t} , t)}{\\alpha_ {t}}, & \\text {i f} t > 0 \\end{array} \\right. \\tag {8}", + "image_path": "588b09dd71eb96334dbbc0cb5754d4b19bd752e35038975edc945b11dbc44148.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "t_k > t_{k-1}" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " denotes the parameters of a lightweight LoRA, and " + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(\\cdot)" + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": " denotes the stop-gradient operator." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Remark. Our learning framework supports universal guidance, as demonstrated through the following instances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 549, + 85 + ], + "type": "text", + "content": "3.1. Learning Better Aligned One-Step Generator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 89, + 554, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 174 + ], + "type": "text", + "content": "Human Feedback Integration. We demonstrate that Human-Feedback Learning (HFL) can be seamlessly integrated into this framework. Specifically, we introduce \"human-preferred images\" as a single conditioning factor. Since we are dealing with only one condition, it is unnecessary to inject it into the desired generator. The learning gradient can then be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 191, + 553, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 191, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 321, + 191, + 553, + 229 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {G r a d} (\\theta) = - \\alpha_ {t} \\mathbb {E} _ {p _ {\\theta} \\left(\\mathbf {x} _ {t} \\mid c\\right) p (c), t} \\lambda_ {t} \\left[ \\nabla_ {\\mathbf {x} _ {t}} r \\left(\\mathbf {x} _ {t}\\right) \\right. \\\\ \\left. + \\nabla_ {\\mathbf {x} _ {t}} \\log \\frac {p _ {\\phi} (\\mathbf {x} _ {t})}{p _ {\\theta} (\\mathbf {x} _ {t})} \\right] \\frac {\\partial \\mathbf {x}}{\\partial \\theta}, \\tag {9} \\\\ \\end{array}", + "image_path": "fdcf78eb309e33e7f7b8339cb686dee191d38229eae42b0f1ba9da9d2c1c7df4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "text", + "content": " models the gradient of log probability of an image being human-preferred, formally defined as " + }, + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t}\\log p(\\text{\"Human-preferred images\"}|\\mathbf{x}_t)\\triangleq \\nabla_{\\mathbf{x}_t}r(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 236, + 554, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "text", + "content": "Decoupled CFG. Our framework separates the learning of condition and fidelity into two distinct components. When the conditional probability " + }, + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "inline_equation", + "content": "p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "text", + "content": " represents text-image alignment, its gradient can be computed using Classifier-Free Guidance (CFG). While previous approaches [42] have incorporated CFG during distillation, however, their CFG is coupled with the real score. In contrast, our framework explicitly leverages CFG for conditional learning. This key difference allows us to compute CFG using a diffusion model different from the one used for real score computation, enabling the use of more sophisticated diffusion models to guide text-image alignment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 424, + 554, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 449 + ], + "type": "text", + "content": "3.2. Learning One-Step Generator with Additional Control" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 454, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 502 + ], + "type": "text", + "content": "Controllable Generation. To incorporate additional control capabilities similar to ControlNet [45], we parameterize the student model as a diffusion model with an associated ControlNet. The generator's learning objective becomes:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 511, + 554, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 511, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 318, + 511, + 554, + 538 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , \\beta} - \\mathbb {E} _ {p _ {\\theta , \\beta} (\\mathbf {x} _ {t} | c) p (c), t} \\lambda_ {t} [ \\log p (c | \\mathbf {x} _ {t}) + \\log \\frac {p _ {\\phi , t} (\\mathbf {x} _ {t})}{p _ {\\psi , t} (\\mathbf {x} _ {t})} ], \\tag {10}", + "image_path": "e2b3681ed7bed3653b49439c978142686320d782d5635c030f517f374e0cedf2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "content": " denotes the additional parameters for ControlNet. This formulation enables asymmetric capability development - the student model can learn conditional generation tasks beyond the teacher's abilities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "To the best of our knowledge, we are the first to explore learning additional control one-step generator by score distillation in an asymmetric form. To highlight the effectiveness of our proposed formulation, we discuss two potential competitive variants for learning one-step generators with additional control conditions. 1) Train ControlNet via Denoising: Given a pre-trained diffusion model " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": ", a ControlNet parameterized by " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " can be trained by denoising score matching loss. Hence a naive idea is training the ControlNet for one-step generation using denoising loss too." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4012" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 91, + 294, + 210 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 101, + 90 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 101, + 90 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 101, + 90 + ], + "type": "text", + "content": "Control Signal" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 131, + 71, + 161, + 90 + ], + "lines": [ + { + "bbox": [ + 131, + 71, + 161, + 90 + ], + "spans": [ + { + "bbox": [ + 131, + 71, + 161, + 90 + ], + "type": "text", + "content": "Implicit Training" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 71, + 223, + 90 + ], + "lines": [ + { + "bbox": [ + 187, + 71, + 223, + 90 + ], + "spans": [ + { + "bbox": [ + 187, + 71, + 223, + 90 + ], + "type": "text", + "content": "Denoising Training" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 242, + 76, + 285, + 87 + ], + "lines": [ + { + "bbox": [ + 242, + 76, + 285, + 87 + ], + "spans": [ + { + "bbox": [ + 242, + 76, + 285, + 87 + ], + "type": "text", + "content": "JDM(Ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 91, + 294, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 294, + 210 + ], + "type": "image", + "image_path": "cabc5a9d8872a46cdd005a89d572c505ff6b30f4f3ef9e5d72ec877c728dce76.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "type": "text", + "content": "Figure 3. The qualitative comparison of the proposed method and potential baselines in one-step controllable generation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 249, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 406 + ], + "type": "text", + "content": "Empirically, we find this approach can transfer the control well but generate blurry images (see Tab. 2 and Fig. 3); 2) Implicitly Train ControlNet via Score Distillation Training ControlNet for Diffusion Models is straightforward by directly using the original diffusion loss. Hence, a natural idea is to follow the training of ControlNet for DMs and directly inject the condition via ControlNet with the original VSD loss. This approach shares a high-level idea with training ControlNet for diffusion models: maintaining the original training loss while only adding ControlNet to the generator for condition injection. However, we empirically find this approach can generate high-quality images but ignores the control signal (see Tab. 2 and Fig. 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "type": "text", + "content": "3.2.1. Shared One-Step Generator Between Different Additional Control" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": "While Eq. (10) establishes the joint training of One-Step Generator " + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": " and ControlNet " + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 437, + 295, + 510 + ], + "type": "text", + "content": ", this approach necessitates retraining the entire One-Step Generator for each new control condition. This requirement leads to inefficient use of computational resources and storage space, which limits practical applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "text", + "content": "A simple solution would be to first train a one-step generator without control using diff-instruct [17], then train ControlNet to incorporate additional control into " + }, + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "text", + "content": ". However, previous work has shown that one-step generators trained in this manner often suffer from mode collapse, making it difficult to accommodate additional control signals. One potential approach is to use teacher diffusion to generate millions of noise-image pairs, then add ODE regression loss for a one-step generator [42]. However, this method is extremely computationally inefficient and impractical for real-world applications." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "To address these limitations, we propose a novel two-phase warm-up training strategy: 1) Initial Phase: Joint training of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " and ControlNet for a primary condition; 2) Extension Phase: Fixing " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " while training only ControlNet for subsequent conditions. This approach leverages the joint KL divergence to regularize " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " during initial training," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "resulting in better accommodating the condition and avoiding mode collapse. The resulting well-trained " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " can then effectively incorporate other forms of control." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 115, + 395, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 115, + 395, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 115, + 395, + 128 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 133, + 451, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 451, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 451, + 144 + ], + "type": "text", + "content": "4.1. Controllable Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 150, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 553, + 258 + ], + "type": "text", + "content": "Experiment Setting. All the models are trained in an internally collected dataset. We use Stable Diffusion 1.5 [26] as the frozen teacher model and initialization for the fake score model and student. The ControlNet is initialized in the way introduced in its original paper [45]. We utilize four conditions for evaluating our method in one-step controllable generation as follows: Canny [3], Hed [37], Depth map, and lower resolution samples. Details for these conditions can be found in the Appendix B." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "text", + "content": "Evaluation Metric. We select the FID [8] to measure the image quality and consistency to measure the controllability. In particular, the FID is calculated between generated images by SD without additional control and generated images based on additional conditions. The consistency is defined by Consistecny " + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "inline_equation", + "content": "= ||h(\\mathbf{x}) - c||_{1}" + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "text", + "content": " is the condition, and " + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "inline_equation", + "content": "h(\\cdot)" + }, + { + "bbox": [ + 313, + 258, + 553, + 366 + ], + "type": "text", + "content": " is the function used for obtaining condition. We also report the number of function evaluations (NFE) required for generating an image for comparing efficiency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 367, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 713 + ], + "type": "text", + "content": "Quantitative Results. We conduct comprehensive evaluations comparing our proposed approach against two baseline methods: (1) the standard diffusion model (DM) with ControlNet and (2) pre-trained one-step generator by diff-instruct integrated with DM's ControlNet. The quantitative results are summarized in Tab. 1, where we evaluate both image quality (FID) and condition consistency across multiple control tasks. Our results reveal several key findings: 1) The proposed method achieves a significant reduction in the number of function evaluations (NFEs) from 50 to 1, while maintaining or improving performance metrics. Specifically, our method demonstrates superior FID scores and consistency measures across various conditioning tasks, indicating both better image quality and more precise condition adherence. 2) Our shared one-step generator architecture exhibits substantial improvements over the naive integration of a pre-trained one-step generator, validating the effectiveness of our unified training strategy. This is evidenced by consistent performance gains across all evaluated metrics and tasks. 3) While it is technically feasible to directly combine DM's ControlNet with a pre-trained one-step generator, this approach yields substantially inferior results. This observation underscores the importance of our proposed strategy tailored for training one-step generators with additional control. Overall, these results demonstrate that our method successfully achieves better trade-off between efficiency and sample quality in controlled image generation, achieving state-of-the-art performance with significantly reduced computational overhead" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4013" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 169, + 71, + 441, + 81 + ], + "lines": [ + { + "bbox": [ + 169, + 71, + 441, + 81 + ], + "spans": [ + { + "bbox": [ + 169, + 71, + 441, + 81 + ], + "type": "text", + "content": "Table 1. Comparison of machine metrics of different methods across tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 553, + 155 + ], + "type": "table", + "html": "
MethodNFE↓FID↓Canny Consistency↓FID↓HED Consistency↓FID↓Depth Consistency↓8× Super ResolutionFID↓Avg Consistency↓
FID↓Consistency↓
ControlNet5014.480.11319.210.10115.250.09311.930.06515.210.093
DI + ControlNet122.740.14128.040.11322.490.09715.570.12622.210.120
JDM (Ours)113.070.10216.510.04516.240.08112.480.05314.580.071
", + "image_path": "3908136c05d7541d41ae9979e2fb548c9a870e6c3abd35498523fe8accc6e94d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 164, + 105, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 164, + 105, + 187 + ], + "spans": [ + { + "bbox": [ + 72, + 164, + 105, + 187 + ], + "type": "text", + "content": "Control Signal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 164, + 172, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 164, + 172, + 186 + ], + "spans": [ + { + "bbox": [ + 123, + 164, + 172, + 186 + ], + "type": "text", + "content": "ControlNet, 50 NFE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 164, + 233, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 164, + 233, + 186 + ], + "spans": [ + { + "bbox": [ + 184, + 164, + 233, + 186 + ], + "type": "text", + "content": "DI+Control Net, 1 NFE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 164, + 298, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 164, + 298, + 186 + ], + "spans": [ + { + "bbox": [ + 242, + 164, + 298, + 186 + ], + "type": "text", + "content": "JDM(Ours), 1 NFE" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 189, + 299, + 430 + ], + "blocks": [ + { + "bbox": [ + 58, + 189, + 299, + 430 + ], + "lines": [ + { + "bbox": [ + 58, + 189, + 299, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 189, + 299, + 430 + ], + "type": "image", + "image_path": "cca29ca67671377cd9f6af027a1d20526e5a37db60c44322f972b6722edb9bbf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 440, + 523, + 452 + ], + "lines": [ + { + "bbox": [ + 86, + 440, + 523, + 452 + ], + "spans": [ + { + "bbox": [ + 86, + 440, + 523, + 452 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on controllable generation across different control signals against competing methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 164, + 358, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 164, + 358, + 186 + ], + "spans": [ + { + "bbox": [ + 325, + 164, + 358, + 186 + ], + "type": "text", + "content": "Control Signal" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 376, + 164, + 426, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 164, + 426, + 186 + ], + "spans": [ + { + "bbox": [ + 376, + 164, + 426, + 186 + ], + "type": "text", + "content": "ControlNet, 50 NFE" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 436, + 164, + 487, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 436, + 164, + 487, + 186 + ], + "spans": [ + { + "bbox": [ + 436, + 164, + 487, + 186 + ], + "type": "text", + "content": "DI+Control Net,1 NFE" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 495, + 164, + 550, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 495, + 164, + 550, + 186 + ], + "spans": [ + { + "bbox": [ + 495, + 164, + 550, + 186 + ], + "type": "text", + "content": "JDM(Ours), 1 NFE" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 312, + 189, + 553, + 430 + ], + "blocks": [ + { + "bbox": [ + 312, + 189, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 312, + 189, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 189, + 553, + 430 + ], + "type": "image", + "image_path": "229a600e95d6d33214bdb82006387db45f65bf693a08eeac54791e75721d2fc1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 472, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 581 + ], + "type": "text", + "content": "Qualitative Comparison. We present the qualitative comparison in Fig. 4. We can observe that DM's ControlNet provides high-level control for one-step generators. However, this approach produces lower-quality images. Our customized one-step generator training with additional control can generate much higher-quality images. This validates the effectiveness of our proposed method and indicates that we can teach students conditions unknown to teachers by minimizing the upper bound of joint KL divergence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 588, + 239, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 239, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 239, + 601 + ], + "type": "text", + "content": "4.2. Comparison to Potential Baselines" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "We conduct comprehensive studies to validate our design choices by comparing with two potential baseline approaches as discussed in Sec. 3.2: 1) training ControlNet via denoising loss; 2) implicitly training ControlNet via score distillation. The quantitative results are shown in Tab. 1. The quantitative results are shown in Fig. 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": "Baseline Details. We introduce the training of two potential baselines as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- Train ControlNet via Denoising: Given a pre-trained" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 472, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 472, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 321, + 472, + 553, + 497 + ], + "type": "text", + "content": "one-step diffusion model " + }, + { + "bbox": [ + 321, + 472, + 553, + 497 + ], + "type": "inline_equation", + "content": "G_{\\theta}" + }, + { + "bbox": [ + 321, + 472, + 553, + 497 + ], + "type": "text", + "content": ", the ControlNet is trained by denoising:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 369, + 502, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 502, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 369, + 502, + 553, + 517 + ], + "type": "interline_equation", + "content": "L (\\beta) = \\mathbb {E} _ {\\mathbf {x}, \\epsilon} | | G _ {\\{\\phi , \\beta \\}} (\\mathbf {x} _ {T}, c) - \\mathbf {x} | | _ {2} ^ {2}, \\tag {11}", + "image_path": "730d5815f453209aed75c82811ca4dc760cacfa9b19ebe35561a7236f765f7f6.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 322, + 523, + 548, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 523, + 548, + 534 + ], + "spans": [ + { + "bbox": [ + 322, + 523, + 548, + 534 + ], + "type": "text", + "content": "* where " + }, + { + "bbox": [ + 322, + 523, + 548, + 534 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 322, + 523, + 548, + 534 + ], + "type": "text", + "content": " is the terminal step in the one-step generator." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 535, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 535, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 314, + 535, + 554, + 571 + ], + "type": "text", + "content": "- Implicitly Train ControlNet via score distillation: The one-step student with additional control is directly trained by the original VSD loss as following:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 376, + 577, + 553, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 577, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 376, + 577, + 553, + 600 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left(\\int_ {c} q _ {\\theta , \\beta} (\\mathbf {x} | c) p (c) d c \\mid \\mid p _ {\\phi} (\\mathbf {x})\\right). \\tag {12}", + "image_path": "525ecc76bea339d1e0b9b8c65d431792aea018b9a19b5329a51ada2a39617dcf.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 321, + 605, + 553, + 629 + ], + "type": "text", + "content": "The condition injection is implicitly trained via reverse KL divergence." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Comparing to Train ControlNet via Denoising. The baseline of training ControlNet through DSM loss achieves reasonable consistency scores (0.109 for Canny and 0.094 for Depth) but suffers from significantly degraded image quality, as reflected by the worse FID scores (28.83 for Canny and 35.14 for Depth). This reveals that this approach tends to generate blurry images while maintaining decent" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4014" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 541, + 429 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 541, + 429 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 541, + 429 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 541, + 429 + ], + "type": "image", + "image_path": "9ef4fbc53e9191a79d473c22101dfde5bb1667032e72c26478744a2415423389.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 483 + ], + "type": "text", + "content": "control. The degradation in image quality can be attributed to the mismatch between the denoising objective and the reverse KL divergence, as the model is forced to learn denoising behavior that may not be optimal for direct generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 604 + ], + "type": "text", + "content": "Comparing to Implicitly Train ControlNet via score distillation. The implicit training baseline through score distillation shows better FID scores (22.87 for Canny and 22.34 for Depth) compared to the DSM approach, indicating its capability to generate higher quality images. However, it demonstrates worse consistency scores (0.151 for Canny and 0.150 for Depth), suggesting that the control signals are not effectively incorporated. This indicates that simply maintaining the original VSD loss while adding ControlNet may lead to the model ignoring the control conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 612, + 295, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 625 + ], + "type": "text", + "content": "4.3. Other Application in Text-to-Image Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Evaluation. We employs multiple metrics to assess different aspects of generation quality: Aesthetic Score (AeS) [29] evaluates the image quality; CLIP Score (CS) measures the text-to-image alignment; Human Preference Score (HPS) v2.1 [35] correlates strongly with the human preference, capturing both image-text alignment and aesthetic quality." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 455, + 553, + 529 + ], + "blocks": [ + { + "bbox": [ + 313, + 434, + 553, + 454 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 454 + ], + "type": "text", + "content": "Table 2. Comparison of machine metrics of different methods across tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 455, + 553, + 529 + ], + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 529 + ], + "type": "table", + "html": "
MethodNFE↓FID↓Canny Consistency↓FID↓Depth Consistency↓
JDM (Ours)113.070.10216.240.081
w/ warm-up shared UNet114.350.12216.710.093
w/o modeling p(c|xt)114.890.11318.560.085
Implicit Training122.870.15122.340.150
DI w/ Denoising128.830.10935.140.094
", + "image_path": "043c45106854410112a15c6fec88e4137f3ffab8ea5e35360ced2ae5f695f0aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 639 + ], + "type": "text", + "content": "Baseline Models. We conduct our experiments on SD-v1.5. We perform human feedback learning by ImageReward [38]. To ensure fairness in evaluation, we do not report this metric. We perform decoupled CFG by using SD-v2.1 in distilling SD-v1.5. We mainly compare our model against the open-source state-of-the-art (SOTA) models, e.g., LCM [14], Hyper-SD [25], and DMD2 [43]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Quantitative Results. Table 3 presents a comprehensive comparison of our method against existing approaches. Our one-step generator, enhanced with SD-v2.1-based CFG and human feedback learning, demonstrates superior performance across all evaluation metrics. Notably, our method significantly outperforms the direct baselines such as Diff-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4015" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Comparison of machine metrics on text-to-image generation across state-of-the-art methods. HFL denotes human feedback learning which might hack the machine metrics. We highlight the best and second best among distillation methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 553, + 259 + ], + "type": "table", + "html": "
ModelBackboneHFLStepsHPS↑Aes↑CS↑
AnimationConcept-ArtPaintingPhotoAverage
Base ModelSD-v1.5No2526.2924.8524.8726.0125.505.4933.03
Base ModelSD-v2.1No2527.8227.1427.1728.1727.585.6633.46
SD Turbo [28]SD 2.1No128.3026.9226.4325.5426.805.3132.21
InstaFlow [11]SD-v1.5No123.1723.0422.7322.9722.985.2531.97
TCD [47]SD-v1.5No423.1421.1121.0823.6222.245.4329.07
LCM-dreamshaper [15]SD-v1.5No426.5126.4025.9624.3225.805.9431.55
PeRFlow [40]SD-v1.5No422.7922.1721.2823.5022.435.3530.77
DMD2 [43]SD-v1.5No124.1722.6822.9724.3023.535.8230.92
Diff Instruct [17]SD-v1.5No127.3226.1526.4125.5026.355.7132.08
Hyper-SD [25]SD-v1.5Yes128.6528.1628.4126.9028.015.6430.87
JDM w/ HFL (Ours)SD-v1.5Yes130.1629.1730.1428.3529.465.8933.75
JDM w/ better CFG (Ours)SD-v1.5No130.5629.4630.3828.5929.755.9033.97
", + "image_path": "e3c59cc25ec5134fde73ecf2a180bc5910ae3040a17d236724bac823844c5822.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 275, + 294, + 390 + ], + "blocks": [ + { + "bbox": [ + 57, + 275, + 294, + 390 + ], + "lines": [ + { + "bbox": [ + 57, + 275, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 57, + 275, + 294, + 390 + ], + "type": "image", + "image_path": "1807028b1a6b7bc2baa42845ececf527a4ead7481c30891a3e972ac48612324f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "content": "Figure 6. We compare JDM with the variant directly using " + }, + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 396, + 295, + 419 + ], + "type": "text", + "content": ". It is clear that the variant suffers from artifacts." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 521 + ], + "type": "text", + "content": "Instruct [17] and DMD2 [43]. An intriguing observation is that the variant utilizing better CFG achieves even better metrics compared to the HFL variant. This unexpected finding suggests a promising research direction: the potential benefits of leveraging multiple teacher diffusion models for student model distillation, which merits further investigation in future work." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 666 + ], + "type": "text", + "content": "Qualitative Comparison. Since using HFL may lead to the hack of machine metrics, we further conducted qualitative comparisons. We qualitatively compared our w/ HFL and our w/ better CFG with the most competitive baseline, as shown in Fig. 5. Our method demonstrates significantly better visual quality and text-image alignment. Notably, while Hyper-SD also employs HFL, their HFL is performed separately, whereas ours is conducted collaboratively with one-step learning, allowing fake scores to participate and eliminate artifacts caused by reward maximization. As shown in Fig. 5, Hyper-SD exhibits noticeable artifacts, likely due to their flawed HFL approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 203, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 203, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 203, + 685 + ], + "type": "text", + "content": "4.4. Additional Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Effect of Warm-up Shared UNet The proposed two-phase warm-up training strategy is crucial for learning a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 363 + ], + "type": "text", + "content": "shared one-step UNet among different controls. Here, we compare our strategy to naively using pre-trained DI as shared UNet. The results are shown in Tab. 2. Without the warm-up strategy, the performance will degrade severely, especially regarding the consistency metric. This indicates our warm-up strategy can help in learning a One-step UNet that can handle multiple conditions well." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": "Effect of Modeling " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " Modeling " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " not only makes the formulation more elegant but also ensures that the information from the gradients of condition learning and fidelity pertains to the same diffused " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": ". If we were to simply substitute " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": ", the gradient information would relate separately to " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": ", resulting in less stable learning. As demonstrated in Tab. 2, omitting the modeling of " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " leads to a significant drop in performance, particularly concerning the FID. This highlights the importance of modeling " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " for achieving high-quality controllable one-step generation. Furthermore, modeling " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " is more robust than using " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": ". In HFL learning scenarios, directly modeling " + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\log p(c|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 380, + 556, + 561 + ], + "type": "text", + "content": " can result in noticeable artifacts caused by reward hacking, as illustrated in Fig. 6." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this work, we propose JDM for adding new control unknown to teacher DMs into one-step student. Our method minimizes the upper bound of reverse KL divergence between image-condition joint distributions. This approach decouples fidelity and condition learning, allowing the one-step student to handle controls unknown to the teacher. Extensive experiments show that JDM outperforms multi-step controllable DMs by one-step, while achieving SOTA performance in one-step text-to-image synthesis by the integration of decoupled CFG or human feedback learning." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4016" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 273 + ], + "type": "text", + "content": "Jing Tang's work is partially supported by National Key R&D Program of China under Grant No. 2024YFA1012700 and No. 2023YFF0725100, by the National Natural Science Foundation of China (NSFC) under Grant No. 62402410 and No. U22B2060, by Guangdong Provincial Project (No. 2023QN10X025), by Guangdong Basic and Applied Basic Research Foundation under Grant No. 2023A1515110131, by Guangzhou Municipal Science and Technology Bureau under Grant No. 2024A04J4454, by Guangzhou Municipal Education Bureau (No. 2024312263), by Nansha District Project (No. 2023ZD022), and by Guangzhou Industrial Information and Intelligent Key Laboratory Project (No. 2024A03J0628) and Guangzhou Municipal Key Laboratory of Financial Technology Cutting-Edge Research (No. 2024A03J0630)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 284, + 115, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 115, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 115, + 297 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 304, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 304, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 304, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 295, + 357 + ], + "type": "text", + "content": "[1] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Roni Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 361, + 296, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 394 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 394 + ], + "type": "text", + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 396, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 294, + 418 + ], + "type": "text", + "content": "[3] John Canny. A computational approach to edge detection. PAMI, 1986. 5, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 419, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 453 + ], + "type": "text", + "content": "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "text", + "content": "[5] Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, et al. Emu: Enhancing image generation models using photogenic needles in a haystack. arXiv preprint arXiv:2309.15807, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "type": "text", + "content": "[6] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 578, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 622 + ], + "type": "text", + "content": "[7] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 678 + ], + "type": "text", + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 296, + 714 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[10] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 173 + ], + "type": "text", + "content": "[11] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. arXiv preprint arXiv:2309.06380, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 217 + ], + "type": "text", + "content": "[12] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[13] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[14] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 554, + 342 + ], + "type": "text", + "content": "[15] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[16] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. Transactions on Machine Learning Research, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 378, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 554, + 432 + ], + "type": "text", + "content": "[17] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 466 + ], + "type": "text", + "content": "[18] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 554, + 511 + ], + "type": "text", + "content": "[19] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. arXiv preprint arXiv:2402.15170, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 513, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 554, + 567 + ], + "type": "text", + "content": "[20] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14297-14306, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 554, + 612 + ], + "type": "text", + "content": "[21] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "text", + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[23] Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4017" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[24] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[25] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis, 2024. 3, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[27] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 285 + ], + "type": "text", + "content": "[28] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 352 + ], + "type": "text", + "content": "[29] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 397 + ], + "type": "text", + "content": "[30] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "text", + "content": "[32] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "type": "text", + "content": "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "text", + "content": "[34] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. arXiv preprint arXiv:2305.16213, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "text", + "content": "[35] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 667 + ], + "type": "text", + "content": "[36] Jie Xiao, Kai Zhu, Han Zhang, Zhiheng Liu, Yujun Shen, Zhantao Yang, Ruili Feng, Yu Liu, Xueyang Fu, and Zheng-Jun Zha. CCM: Real-time controllable visual content creation using text-to-image consistency models. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "text", + "content": "[37] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In ICCV, 2015. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagere-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 555 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "ward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[39] Shuchen Xue, Zhaoqiang Liu, Fei Chen, Shifeng Zhang, Tianyang Hu, Enze Xie, and Zhenguo Li. Accelerating diffusion sampling with optimized time steps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8292-8301, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[40] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. arXiv preprint arXiv:2405.07510, 2024. 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[41] Zilyu Ye, Zhiyang Chen, Tiancheng Li, Zemin Huang, Wei-jian Luo, and Guo-Jun Qi. Schedule on the fly: Diffusion time prediction for faster and better image generation. arXiv preprint arXiv:2412.01243, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[42] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[43] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis, 2024. 3, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[44] Xuanwu Yin Yuda Song, Zehao Sun. Sdxs: Real-time one-step latent diffusion models with image conditions. arxiv, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[45] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[46] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. Advances in Neural Information Processing Systems, 36:49842-49869, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 498 + ], + "type": "text", + "content": "[47] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[48] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In International Conference on Machine Learning, 2024. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4018" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_content_list.json b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5545a6802422a3afdf284232f707388f51fc500a --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "Addressing Representation Collapse in Vector Quantized Models with One Linear Layer", + "text_level": 1, + "bbox": [ + 147, + 128, + 849, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongxin Zhu $^{1,2}$ , Bocheng Li $^{1,2}$ , Yifei Xin $^{3}$ , Zhihua Xia $^{4}$ , Linli Xu $^{1,2*}$", + "bbox": [ + 228, + 202, + 772, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 307, + 220, + 691, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ State Key Laboratory of Cognitive Intelligence, $^{3}$ Peking University, $^{4}$ Jinan University", + "bbox": [ + 156, + 238, + 838, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zyx2016@mail.ustc.edu.cn,bcli@mail.ustc.edu.cn", + "bbox": [ + 227, + 257, + 769, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xinyifei@stu.pku.edu.cn, xiazhihua@jnu.edu.cn, linlixu@ustc.edu.cn", + "bbox": [ + 117, + 273, + 872, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vector Quantization (VQ) is essential for discretizing continuous representations in unsupervised learning but suffers from representation collapse, causing low codebook utilization and limiting scalability. Existing solutions often rely on complex optimizations or reduce latent dimensionality, which compromises model capacity and fails to fully solve the problem. We identify the root cause as disjoint codebook optimization, where only a few code vectors are updated via gradient descent. To fix this, we propose SimpleVQ, which reparameterizes code vectors through a learnable linear transformation layer over a latent basis, optimizing the entire linear space rather than nearest individual code vectors. Although the multiplication of two linear matrices is equivalent to applying a single linear layer, this simple approach effectively prevents collapse. Extensive experiments on image and audio tasks demonstrate that SimVQ improves codebook usage, is easy to implement, and generalizes well across modalities and architectures.", + "bbox": [ + 88, + 357, + 485, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 657, + 220, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, vector quantization (VQ) [19, 29] has emerged as a foundational technique in unsupervised representation learning [2, 5] and latent generative models [4, 21, 31, 33, 34, 44]. By converting continuous representations into discrete codes, VQ models can effectively identify the inherent structure of data and enable various discrete modeling methods on continuous data, from high-quality image generation [9] to audio synthesis [6]. The recent success of Large Language Models (LLMs) [1] has highlighted the effectiveness of next-token prediction as a powerful and versatile training objective. Consequently, VQ models are taken as the direct method to transform data from various modalities [25, 28, 37] or scientific domains [10] to discrete", + "bbox": [ + 89, + 681, + 482, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequences for next token prediction training. However, attempts to integrate VQ models as multimodal tokenizers to leverage the scaling laws of LLMs face significant challenges because of the difficulty of expanding the codebook. For example, the Chameleon model [28] constrains its codebook size to $8k$ , which is significantly trailing behind the vocabulary size of LLMs (e.g., LLaMA3's vocabulary size is $128k$ [8]).", + "bbox": [ + 511, + 328, + 905, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a broad agreement that increasing vocabulary size can consistently improve the performance of LLMs [27]. However, recent studies [42] indicate that traditional VQ models often fail to utilize the additional parameters introduced by codebook expansion, leaving most codes inactive during training. The contradiction between codebook expansion and low codebook utilization in VQ models is known as the representation collapse problem [22], where increasing the codebook size fails to improve the performance. To address these discrepancies, we conduct a theoretical analysis of the optimization procedure of VQ models and identify that the disjoint optimization of the codebook is the root cause of representation collapse. As illustrated in Fig. 1(a), the core mechanism of VQ models involves a nearest-neighbor replacement strategy, where the encoder's output features are replaced by the nearest vector in the codebook to serve as input to the decoder. The indices of the nearest vector are taken as the discrete representation of the data. This nearest-selection operator results in only a subset of codes being updated through gradient descent, while the remaining codes remain unchanged.", + "bbox": [ + 511, + 455, + 906, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some methods mitigate representation collapse by hand-designing complex optimization strategies, such as stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] through sophisticated training strategies. Recently, some approaches [16, 33, 35] propose to reduce the dimension of the latent space to a very small scale (e.g., 8 v.s. 128) to alleviate the curse of dimensionality, thereby improving the overlap between the encoder's fea", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22968", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e32b10883c72bc623c65cca52363719c50009416e5845bc98d83fcb0158b8bd.jpg", + "image_caption": [ + "Figure 1. Comparison of Vanilla VQ and SimVQ. (a): (left) Disjoint optimization in Vanilla VQ. Only the nearest codes are updated, resulting in a high percentage of \"dead\" codes that are not updated. (b): (right) Joint optimization in SimVQ. The entire codebook is updated with a latent basis, ensuring all codes remain active." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 504, + 233 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9eaa319bcd3756109ed568cab93cd588d8097f25b904638324a06aa7447b3d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 85, + 887, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tures and the codebook. However, while these methods enhance codebook utilization, they do so at the cost of model capacity, leading to worse performance compared to vanilla VQ models when the codebook size is small or representation collapse is not severe. Another approach, VQGAN-LC [42], initializes the codebook with features extracted from the pre-trained CLIP model [17] to create a well-structured latent space that better matches the distribution of the encoder output. Nevertheless, the latent space defined by an external pre-trained model limits the model's ability to generalize to diverse datasets and reaches a performance plateau as the codebook size increases. These limitations highlight the need for a more effective method to improve codebook utilization without compromising model capacity or relying on external models.", + "bbox": [ + 88, + 324, + 485, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We critically assess prevalent methodologies and reveal that optimizing the latent space rather than individual code vectors is key to preventing representation collapse. Building on this insight, we introduce a simple yet effective method, termed SimVQ, to directly update the latent space spanned by the codebook by linear transforming the code vectors via a learnable latent basis. Specifically, the vectors in the codebook are reparameterized as a linear combination of the basis in the learnable linear layer $W$ :", + "bbox": [ + 88, + 551, + 483, + 686 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} \\in \\mathbb {R} ^ {K \\times d} \\Rightarrow \\boldsymbol {C W} \\text {w i t h} \\boldsymbol {W} \\in \\mathbb {R} ^ {d \\times d}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 693, + 482, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $K$ denotes the codebook size and $d$ represents the dimension of latent space. This reparameterization with linear transformation disentangles the optimization of the codebook into two components: the coefficient matrix $\\mathbf{C}$ and the basis of linear space $\\mathbf{W}$ respectively. As illustrated in Fig. 1(b), by optimizing the basis matrix $\\mathbf{W}$ , the latent space spanned by $\\mathbf{CW}$ is rotated and stretched to match encoder's output feature. The entire codebook is updated jointly to prevent the representation collapse problem. The simplicity of the proposed method makes it highly portable and easily adaptable for improving VQ-based models across a wide range of domains, requiring only one linear layer.", + "bbox": [ + 88, + 719, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions to vector quantized models are as follows:", + "bbox": [ + 511, + 325, + 903, + 353 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We theoretically analyze the representation collapse problem in VQ models and reveal that optimizing the latent space spanned by the codebook, rather than individual code vectors, is crucial to addressing this issue.", + "- We propose a novel method, SimVQ, which reparameterizes the codebook vectors in VQ models via a linear transformation with a learnable latent basis. This simple yet effective approach is highly adaptable and easy to implement, making it broadly applicable across various machine learning contexts.", + "- We conduct an extensive evaluation of SimVQ across diverse modalities, including image and audio with different model architectures. The results show that SimVQ not only effectively addresses the representation collapse problem by achieving near-complete codebook utilization regardless of the codebook size, but also establishes new state-of-the-art performance. Furthermore, when scaling up the codebook size, SimVQ consistently delivers improved results." + ], + "bbox": [ + 511, + 356, + 906, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 661, + 656, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VQ-VAE [29] is the pioneering work to encode data into discrete representations. Building on these developments, VQGAN [9] combines VQ-VAE with adversarial networks to improve the perceptual quality of generated samples and establish a fundamental quantization protocol in latent generative models [21, 28, 34]. Many traditional VQ works focus on training a better discrete representation rather than codebook utilization to improve reconstruction performance. For example, RVQ [15] and MoVQ [41] enhance the reconstruction details with multichannel quantization. However, these methods suffer from a critical issue of representation collapse, as they struggle to scale the codebook size beyond $10\\mathrm{k}$ entries, limiting their scalability. In response to this challenge, several approaches have been pro", + "bbox": [ + 509, + 688, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22969", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "posed recently. DALLE [18] employs the gumbel-softmax trick [12] and stochastic sampling strategies to activate most codes during training. However, during inference, only a small subset of codes is utilized for quantization [38]. Some methods design complex optimization strategies to improve codebook utilization, such as stochastic quantization [26], distribution penalty [30, 32] and codebook reset [40]. Huh et al. [11] proposes rescaling the vectors in the codebook during training to match the distributions in the latent space. VQGAN-FC [33] introduces a method to map latent vectors into a lower-dimensional space followed by $l_{2}$ normalization to alleviate representation collapse. FSQ [16] extends this idea by projecting representations into a reduced-dimensional space, where they are quantized into a small set of fixed values. LFQ [35], a variant of FSQ, uses binary values for quantized representations, thereby simplifying the encoding process. While these methods improve the codebook utilization, they do so at the cost of model capacity by significantly reducing the dimensionality of latent space (often to as low as 8), leading to worse performance compared to vanilla VQ models when the codebook size is small and representation collapse is not severe. Additionally, VQGAN-LC [42] proposes to initialize the codebook using features extracted from the pre-trained CLIP model to avoid representation collapse. However, the reliance on the pre-trained model limits the VQ model's ability to generalize to diverse datasets and results in a performance plateau as the codebook size increases. In contrast, our method, SimVQ, effectively addresses the representation collapse problem with a simple linear layer, without sacrificing model capacity or relying on external pre-trained models.", + "bbox": [ + 89, + 90, + 483, + 573 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Representation Collapse in VQ Models", + "text_level": 1, + "bbox": [ + 89, + 594, + 441, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 621, + 232, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A vector quantized model is typically a reconstructive encoder-decoder architecture that includes a vector quantization layer to convert continuous representations into discrete codes. For simplicity, we represent an image with a single random variable $x$ . Formally, the encoder $f_{\\theta}$ maps the input image into a latent space, producing a continuous representation $z_{e} = f_{\\theta}(x) \\in \\mathbb{R}^{d}$ . This representation is then quantized using a learnable codebook $C = [q_{1}, \\ldots, q_{K}] \\in \\mathbb{R}^{K \\times d}$ , where $q_{i}$ is a codebook vector. We define $\\delta_{k} \\in \\{0, 1\\}^{1 \\times K}$ as a characteristic (one-hot) vector where only the $k$ -th element is 1, such that $q_{k} = \\delta_{k} C \\in \\mathbb{R}^{1 \\times d}$ . The quantization layer selects the nearest codebook vector $q_{k}$ by minimizing the Euclidean distance between $z_{e}$ and the codebook entries [29]:", + "bbox": [ + 89, + 646, + 483, + 858 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk = \\arg \\min _ {j} \\| z _ {e} - q _ {j} \\| _ {2} ^ {2} = \\arg \\min _ {j} \\| z _ {e} - \\delta_ {j} C \\| _ {2} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 872, + 483, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The selected vector $q_{k}$ is then passed to the decoder $g_{\\phi}$ to reconstruct the input image.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable gradient propagation through the non-differentiable characteristic vector $\\delta_{k}$ , the straight-through estimator (STE) [3] is applied. During the backward pass, the gradient of $z_{q} = \\delta_{k}C$ is copied to $z_{e}$ as follows,", + "bbox": [ + 511, + 121, + 903, + 183 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {q} = \\mathrm {s g} (\\delta_ {k} \\pmb {C} - z _ {e}) + z _ {e}, \\Rightarrow \\frac {\\partial z _ {q}}{\\partial z _ {e}} = 1 (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 191, + 903, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathfrak{sg}$ is the stop gradient operator, ensuring the gradient for $\\delta_k C$ is discarded during the backward pass.", + "bbox": [ + 511, + 233, + 903, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The learning objective is the combination of a reconstruction loss and commitment loss that ensures that the encoder commits to an embedding and the encoder's output does not drift:", + "bbox": [ + 511, + 263, + 906, + 323 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\log p (x | z _ {q}) + \\| \\operatorname {s g} \\left(\\delta_ {k} \\boldsymbol {C}\\right) - z _ {e} \\| _ {2} ^ {2} + \\beta \\| \\delta_ {k} \\boldsymbol {C} - \\operatorname {s g} \\left(z _ {e}\\right) \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 334, + 903, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\log p(x|z_q)$ is typically the mean squared error (MSE) loss $\\| x - g_{\\phi}(z_q)\\| _2^2$ for image and audio data.", + "bbox": [ + 511, + 366, + 903, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Disjoint Optimization of Codebook", + "text_level": 1, + "bbox": [ + 511, + 406, + 818, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In VQ models, only the nearest code is selected and updated via gradient descent. Ideally, all codebook entries should be updated and utilized for decoding. However, experimental evidence shows that only a small fraction of the codebook gets updated and utilized, leading to what is known as the representation collapse problem [22]. To investigate the root cause of this issue, we provide a theoretical analysis of the optimization dynamics in VQ models.", + "bbox": [ + 511, + 428, + 905, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the use of the straight-through estimator (STE) for gradient propagation, the codebook $C$ can only be updated through the gradient of the commitment loss, which is defined as:", + "bbox": [ + 511, + 547, + 905, + 607 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C}) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\right\\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 619, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The codebook $\\pmb{C}$ is updated according to the following equation, where $\\eta$ is the learning rate:", + "bbox": [ + 511, + 648, + 903, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {C} ^ {(t + 1)} = \\boldsymbol {C} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C} ^ {(t)})}{\\partial \\boldsymbol {C} ^ {(t)}} \\right] (6) \\\\ = \\boldsymbol {C} ^ {(t)} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} ^ {(t)} \\right] + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} z _ {e} \\right] (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 689, + 903, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta_k^T\\delta_k$ is the Kronecker delta matrix, defined as:", + "bbox": [ + 511, + 762, + 867, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\delta_ {k} ^ {T} \\delta_ {k}\\right) _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j = k, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 790, + 903, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All vectors in $C$ will be updated and utilized if and only if the expectation $\\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right]$ converges to the identity matrix. Unlike variational autoencoders (VAEs) [14], which enforce a Gaussian distribution on the latent space via a", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22970", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "KL-divergence penalty, VQ models optimize $z_{e}$ towards the selected codebook vectors $\\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_kC\\right]$ . At the same time, the selected codebook vectors are optimized towards the distribution of $z_{e}$ , resulting in the same selected subset of vectors moving closer to $z_{e}$ , somewhat akin to a cocoon effect. However, this disjoint optimization of the codebook leads to part of the codebook, specifically $(I - \\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right])C$ , remaining un-updated and underutilized once the optimization process begins. This phenomenon occurs because the optimization focuses only on a subset of codebook vectors, leaving other vectors stagnant.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This analysis reveals the fundamental cause of representation collapse in VQ models: the disjoint optimization process that updates only a subset of codebook vectors. This insight forms the basis for our proposed solution, SimVQ, which aims to address this issue by optimizing the entire latent space spanned by the codebook, rather than individual code vectors.", + "bbox": [ + 89, + 258, + 483, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Addressing Collapse with Latent Linear Transformation", + "text_level": 1, + "bbox": [ + 89, + 378, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Reparameterize Codes with Latent Basis", + "text_level": 1, + "bbox": [ + 89, + 422, + 439, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\pmb{W} = \\{\\pmb{w}_1, \\dots, \\pmb{w}_n\\}$ be a basis of a linear space. Any vector $\\pmb{v}$ in the space can be uniquely expressed as a linear combination of the basis vectors with coefficients $c_1, \\dots, c_n \\in \\mathbb{R}$ :", + "bbox": [ + 89, + 445, + 483, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} = c _ {1} \\boldsymbol {w} _ {1} + \\dots + c _ {n} \\boldsymbol {w} _ {n} = \\boldsymbol {c} \\boldsymbol {W}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 521, + 480, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the equivalence between $\\pmb{v}$ and $c\\pmb{W}$ in the linear space, we can reparameterize each vector in the codebook of VQ models with a new basis matrix $\\pmb{W} \\in \\mathbb{R}^{d \\times d}$ . Specifically, the codebook $\\pmb{C} = \\{c_1, \\dots, c_K\\}$ can be reparameterized as:", + "bbox": [ + 89, + 549, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {\\boldsymbol {c}} _ {1} \\boldsymbol {W}, \\dots , \\hat {\\boldsymbol {c}} _ {N} \\boldsymbol {W} \\} = \\hat {\\boldsymbol {C}} \\boldsymbol {W} \\in \\mathbb {R} ^ {K \\times d}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 637, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This reparameterization introduces two components: the basis matrix $\\mathbf{W}$ and the coefficient matrix $\\hat{\\mathbf{C}}$ . In the following, we will discuss the optimization of both the basis matrix $\\mathbf{W}$ and the coefficient matrix $\\hat{\\mathbf{C}}$ . For simplicity, we slightly abuse $\\mathbf{C}$ and $\\hat{\\mathbf{C}}$ below.", + "bbox": [ + 89, + 669, + 483, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Asymmetric Optimization Dynamics", + "text_level": 1, + "bbox": [ + 89, + 757, + 408, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While it is commonly accepted that multiplying two linear matrices is equivalent to a single linear layer, we argue that the disjoint optimization problem of the codebook in VQ models can be addressed by linear transformation. In vanilla VQ models, only the codebook $C$ is responsible for minimizing commitment loss, leading to the disjoint optimization problem where only the selected codes will be updated.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/acaa8fcbfd39e70edaae65d82bff2a10fc759155be00813d480232790e21b499.jpg", + "image_caption": [ + "Figure 2. (a): (left) The optimization trajectory of the objective $\\| \\pmb{x} - \\pmb{q}\\|_2^2$ , which is the same as vanilla VQ. Only a small fraction of points are updated while others remain inactive. (b): (right) The optimization trajectory of the objective $\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2$ with $\\pmb{q}$ frozen, which is the same as SimVQ. All the points are updated towards targets $x$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 709, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/26aa03cbd2994a9c5092415892590343fd3a37c6be1fc0685e8a96b9a619c387.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 901, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ab17f67fcb4d5ed7446cec316c3649888bf536cb9d657eeef1534741bb777ca8.jpg", + "image_caption": [ + "Figure 3. (a): (left) The optimization trajectory of the optimization objective: $\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2$ with both $\\pmb{q}$ and $\\pmb{w}$ unfrozen. (b): (right) The Frobenius norm of the projection matrix $\\pmb{w}$ and loss curves. The loss quickly converges to 0 with $\\pmb{w}$ almost unchanged." + ], + "image_footnote": [], + "bbox": [ + 517, + 299, + 702, + 396 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cbaae70c1eeb78bd94e2902795523e31e2fa602a5508e663652b2ab6c1c7321f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 299, + 901, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In contrast, when the codebook is reparameterized as $CW$ , both the basis $W$ and the coefficient matrix $C$ contribute to minimizing the commitment loss. The gradients $\\frac{\\partial\\mathcal{L}}{\\partial W}$ and $\\frac{\\partial\\mathcal{L}}{\\partial C}$ can simultaneously reduce the loss. As a result, the optimization of the reparameterized codebook can be divided into three scenarios:", + "bbox": [ + 511, + 489, + 905, + 577 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Updating $C$ with $W$ frozen: Only the selected codes adapt to the latent distribution of $z_{e}$ , as depicted on Fig. 1(a). The vanilla VQ is a special case with $W = I$ .", + "- Updating $W$ with $C$ frozen: The entire codebook $CW$ adjusts to the latent distribution of $z_{e}$ . The basis matrix $W$ rotates and stretches the space as shown in Fig. 1(b).", + "- Updating both $C$ and $W$ : The selected subset of codes moves towards $z_{e}$ while the space spanned by $W$ undergoes simultaneous rotation and stretching." + ], + "bbox": [ + 513, + 579, + 903, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To highlight the difference in optimization between $C$ and $CW$ , we conduct a toy experiment in a two-dimensional setting and visualize the optimization process in Fig. 2 and Fig. 3.", + "bbox": [ + 511, + 715, + 905, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1. Toy Examples", + "text_level": 1, + "bbox": [ + 511, + 782, + 658, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We randomly sample two target points $x$ from Gaussian distribution as follows:", + "bbox": [ + 511, + 801, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {1} \\sim \\mathcal {N} (\\binom {2} {2}, \\binom {1 0} {0 1}), \\quad \\boldsymbol {x} _ {2} \\sim \\mathcal {N} (\\binom {- 2} {- 2}, \\binom {1 0} {0 1}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 839, + 903, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we initialize 10 learnable vectors $\\mathbf{q}$ from a Gaussian", + "bbox": [ + 513, + 885, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22971", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training Procedure for SimVQ", + "text_level": 1, + "bbox": [ + 91, + 90, + 387, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Encoder $f_{\\theta}$ , Decoder $g_{\\phi}$ , Codebook $C \\in \\mathbb{R}^{K \\times d}$ , Linear projector matrix $\\boldsymbol{W}_{\\psi}$ , commitment weight $\\beta$ .", + "bbox": [ + 106, + 109, + 482, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: Model parameters $\\theta, \\phi, \\psi$ and Codebook $C$ . Initialize Codebook with Gaussian distribution and freeze the parameter of Codebook $C$ ;", + "bbox": [ + 106, + 143, + 483, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "repeat", + "text_level": 1, + "bbox": [ + 107, + 189, + 156, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Draw $x\\sim p_{data}(\\pmb {x})$", + "bbox": [ + 125, + 203, + 261, + 218 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {e} = f _ {\\theta} (x);\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 219, + 205, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$/^{*}$ Replace $q_{j}$ in vanilla VQ with proposed $q_{j}\\pmb{W}_{\\psi}$ .", + "bbox": [ + 125, + 233, + 452, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Nearest code search: $k = \\arg \\min_{j} \\| z_{e} - q_{j}W_{\\psi} \\|_{2}^{2}$ , where $q_{j} \\in C$ ;", + "bbox": [ + 125, + 248, + 480, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Straight Through Estimation: $z_{q} = \\mathrm{sg}(q_{k}\\mathbf{W}_{\\psi} - z_{e}) + z_{e}$", + "bbox": [ + 125, + 277, + 480, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = g _ {\\phi} (z _ {q});\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 309, + 207, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Minimize $\\mathcal{L}(\\theta ,\\phi ,\\psi) = \\mathrm{MSE}(x,\\hat{x})$ + $\\beta \\| z_{e}$ - $\\mathrm{sg}(q_kW_\\psi)\\| _2^2 +\\| \\mathrm{sg}(z_e) - q_kW_\\psi \\| _2^2;$", + "bbox": [ + 125, + 324, + 480, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "until converged", + "bbox": [ + 107, + 354, + 215, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distribution:", + "bbox": [ + 89, + 400, + 174, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {q} _ {i} \\right\\} _ {i = 1} ^ {1 0} \\sim \\mathcal {N} \\left(\\binom {0} {0}, \\binom {1} {0}, \\binom {0} {1}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 422, + 480, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training with gradient descent, we introduce perturbation noise $\\mathcal{N}(0,0.01)$ to the targets. In Fig. 2(a), the optimization objective is similar to vanilla VQ: $\\| \\pmb {x} - \\pmb {q}\\| _2^2$ . Only the nearest points $q_{4}$ and $q_{10}$ are updated. In contrast, in Fig. 2(b), the optimization objective $\\| \\pmb {x} - \\pmb {q}\\pmb {w}\\| _2^2$ is similar to SimVQ with the points reparameterized by a learnable latent basis $\\pmb{w}$ and $\\pmb{q}$ frozen, resulting in the entire codebook $\\{\\pmb {q}\\}_{i = 1}^{10}$ being jointly updated.", + "bbox": [ + 89, + 468, + 483, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. The simultaneous optimization of the latent basis $\\mathbf{w}$ and the coefficient matrix $\\mathbf{q}$ may lead to the collapse.", + "bbox": [ + 89, + 599, + 483, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide an example in Fig. 3(a) where the optimization objective is $\\| \\pmb{x} - \\pmb{q}\\pmb{w} \\|_2^2$ with $\\pmb{q}$ unfrozen this time. In the training process, only the nearest point $\\pmb{q}_1$ and point $\\pmb{q}_{10}$ move towards the target point, while other points remain almost unchanged. We also visualize the loss curve in Fig. 3(b). The optimization objective with both $\\pmb{q}$ and $\\pmb{w}$ unfrozen converges quickly, where the norm of basis $\\pmb{w}$ is much smaller than the objective with $\\pmb{q}$ frozen. This indicates that the disjoint optimization of the codebook persists: $\\pmb{q}$ can directly commit to the loss and dominate the optimization process, with $\\pmb{w}$ being ignored, leading to the collapse quickly.", + "bbox": [ + 89, + 641, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Joint Optimization of the Codebook", + "bbox": [ + 89, + 832, + 401, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose SimVQ by simply using a learnable basis $W \\in \\mathbb{R}^{d \\times d}$ to reparameterize the codebook such that the codebook is transformed into $CW$ . The pseudo-code for this", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "approach is provided in Algorithm 1. During training, we optimize only the latent basis matrix $\\mathbf{W}$ , while keeping the coefficient matrix $\\mathbf{C}$ frozen. The commitment loss for SimVQ is defined as:", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e}, q _ {k}\\right) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\boldsymbol {W} \\right\\| _ {2} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 161, + 903, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The vanilla VQ model is a special case of SimVQ, where the linear basis matrix $\\mathbf{W}$ is fixed as the identity matrix $\\mathbf{I}$ . The update for $\\mathbf{W}$ with learning rate $\\eta$ is:", + "bbox": [ + 511, + 190, + 905, + 236 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {W} ^ {(t + 1)} = \\boldsymbol {W} ^ {(t)} - \\eta \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e} , \\boldsymbol {q} _ {k}\\right)}{\\partial \\boldsymbol {W} ^ {(t)}} \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 244, + 903, + 276 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n= \\left(\\boldsymbol {I} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} \\right]\\right) \\boldsymbol {W} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} z _ {e} \\right]. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 277, + 903, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The term $\\mathbb{E}\\left[\\boldsymbol{C}^T\\delta_k^T\\delta_k\\boldsymbol{C}\\right]$ represents the expectation of the quadratic form, and simplifies to $\\mathbb{E}[\\pmb{q}_k^T\\pmb{q}_k]$ . Since the codes are randomly sampled from a Gaussian distribution, we have:", + "bbox": [ + 511, + 321, + 905, + 381 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {q} _ {k} \\right] = \\boldsymbol {I}, \\text {w h e r e} \\boldsymbol {q} \\sim \\mathcal {N} (0, 1), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 382, + 903, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which ensures that all elements of $\\mathbf{W}$ are updated. As training progresses, the latent basis $\\mathbf{W}$ converges to:", + "bbox": [ + 511, + 406, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow \\infty} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {q} _ {k} ^ {T} z _ {e} \\right] \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 446, + 903, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, in the limit:", + "bbox": [ + 513, + 479, + 635, + 493 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {t \\rightarrow \\infty} \\boldsymbol {q} _ {k} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} [ \\boldsymbol {q} _ {k} \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {e} ] = \\mathbb {E} [ \\boldsymbol {e} ] \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 503, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At convergence, the product $q_{k}W$ equals the nearest feature.", + "bbox": [ + 511, + 537, + 903, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Efficiency Analysis", + "bbox": [ + 513, + 575, + 694, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SimVQ demonstrates greater efficiency than vanilla VQ due to its asymmetric training strategy, wherein the codebook $C$ remains static and only the linear projection $W$ is optimized. This approach results in a significant reduction in memory usage during the gradient backpropagation process. In vanilla VQ, the memory cost for codebook optimization is $O(Kd)$ , where $K$ is the number of vectors in the codebook, and $d$ is the dimension of each vector. In our experiments, $K = 65,536$ is much larger than $d = 128$ . As the vocabulary size increases, the memory required for backpropagation grows proportionally, significantly impacting resource consumption. In contrast, SimVQ's memory cost for backpropagation is only $O(d^2)$ because the codebook $C$ is fixed, and only the linear layer $W$ is updated. This results in a constant memory requirement in backpropagation, independent of the vocabulary size. The $d \\times d$ scaling becomes particularly advantageous as $K$ increases in practical applications. This structural design minimizes the computational overhead and improves training efficiency, especially when dealing with large vocabularies.", + "bbox": [ + 511, + 598, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22972", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d2fcd8f4be6e6f4e5e9f18836fa7e11d59eb3dd579768e613730e9718c561a6f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLatent dimCodebook sizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN [9]12865,5361.4%3.740.1722.2070.6
VQGAN-EMA [19]12865,5364.5%3.230.1522.8972.3
VQGAN-FC [33]12865,5361.4%5.330.1821.4568.8
VQGAN-FC [33]865,536100.0%2.630.1323.7977.5
FSQ†[16]664,000100.0%2.800.1323.6375.8
LFQ [35]1665,536100.0%2.880.1323.6077.2
VQGAN-LC-CLIP+ [42]76865,536100.0%2.400.1323.9877.3
SimVQ (ours)12865,536100.0%2.240.1224.1578.4
SimVQ (ours)128262,144100.0%1.990.1124.6880.3
", + "bbox": [ + 142, + 88, + 854, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Reconstruction performance on ImageNet-1k with a resolution of ${128} \\times {128}$ . All models are trained using images downsampled into ${16} \\times {16}$ tokens. † Results are reproduced using the codebook size of $\\left\\lbrack {8,8,8,5,5,5}\\right\\rbrack$ to approximately match ${65},{536}. +$ Following VQGAN-LC,we extract CLIP features with the codebook frozen. The codebook utilization is calculated as the fraction of the codes that are activated at least once when encoding the validation set.", + "bbox": [ + 89, + 273, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 354, + 222, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the efficacy and versatility of the proposed SimVQ, we conduct experiments across both image and audio modalities. Subsequently, we analyze the learned linear layer to investigate the latent basis. The experimental configurations are listed in Appendix 8.1.", + "bbox": [ + 89, + 380, + 482, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Vision Modality", + "text_level": 1, + "bbox": [ + 89, + 467, + 250, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1. Baselines", + "text_level": 1, + "bbox": [ + 89, + 489, + 202, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our baseline selection focuses on methods that enhance codebook utilization through architectural designs. We include VQGAN-FC [33], FSQ [16], LFQ [35] and VQGAN-LC-CLIP [42] as our primary baselines, as they represent the current state-of-the-art in improving reconstruction performance through codebook architecture innovations. Traditional VQ variants such as RVQ [15] and MoVQ [41] address fundamentally different technical challenges - they focus on training better discrete representations, rather than enhancing codebook utilization to improve reconstruction performance like our work. Another important research direction explores optimization-based solutions, where methods like stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] tackle codebook utilization through sophisticated training strategies. Given our focus on architectural innovations in codebook design, we evaluate SimVQ against methods that share this technical foundation for a meaningful assessment of our contributions.", + "bbox": [ + 89, + 508, + 482, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 806, + 302, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To rigorously evaluate the proposed SimVQ, we reproduce all the VQ models listed in Tab. 1 using the same architecture of VQGAN [9] with the quantization layer different only. Among the baselines, for VQGAN-FC [33], we follow the original setting to reduce the dimension of the", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "latent space to 8 followed by $l_{2}$ normalization to improve codebook utilization. For FSQ [16], we adopt a codebook size of $[8,8,8,5,5,5]$ as recommended, to approximately match the default codebook size. For VQGAN-LC [42], we follow them and leverage an external pre-trained CLIP model to extract features of the training dataset in advance for a well-defined latent space. All models are trained on the ImageNet [7] dataset for 50 epochs with a batch size of 256. Input images are processed at a resolution of $128 \\times 128$ pixels and downsampled by a factor of 8, yielding a feature map of $16 \\times 16 \\times 128$ , where 128 is the dimension of the latent space. We set the default codebook size to a large number of $2^{16} = 65536$ rather than the traditional number 8192 to highlight the representation collapse problem. Performance is evaluated using rFID, LPIPS, PSNR, and SSIM metrics on the ImageNet validation set.", + "bbox": [ + 511, + 356, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.3. Main Results", + "text_level": 1, + "bbox": [ + 511, + 609, + 653, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the reconstruction performance of various VQ models on image data. We make three key observations: 1) Traditional VQGAN models utilize only a very small subset of the codebook, with a utilization rate of just $1.4\\%$ . Although VQGAN-EMA is proposed to improve codebook utilization, especially when the codebook size scales up to $65k$ , it still suffers from severe representation collapse. 2) Recently proposed methods, such as LFQ, FSQ, and VQGAN-FC, effectively improve codebook utilization to $100\\%$ . However, these methods require reducing the latent space to a very low dimension. For example, applying VQGAN-FC to the standard latent dimension of 128 results in severe representation collapse and degraded reconstruction performance. Additionally, these models face limitations in model capacity due to the low-dimensional latent space. While they achieve full codebook utilization, their reconstruction quality on rFID score lags significantly behind SimVQ. 3) VQGAN-LC-CLIP leverages an exter", + "bbox": [ + 511, + 628, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22973", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9be7bdf5e71d460d31e29223d0350c2bdb60a1c47ec2b3ce5a7bfdf09a704831.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBandwidthUtil↑UTMOS↑PESQ↑STOI↑V/UV F1↑
GT--4.06/3.48---
EnCodec [6]3.0kbps-2.31/2.092.05/2.050.90/0.880.92/0.89
Vocos [24]3.0kbps-3.53/3.062.40/2.190.92/0.900.94/0.91
SpeechTokenizer [39]3.0kbps-3.56/3.021.93/1.740.88/0.840.93/0.89
WavTokenizer [13]0.9kbps100/100%3.74/3.43*2.01/2.26*0.89/0.89*0.92/0.92*
SimVQ (ours)0.9kbps100.0/100.0%4.00/3.512.33/2.080.91/0.880.94/0.91
WavTokenizer [13]0.975kbps68/-%4.02*/-2.39*/-0.92*/-0.94*/-
WavTokenizer [13]1.05kbps27/-%4.00*/-2.36*/-0.81*/-0.94*/-
SimVQ (ours)0.975kbps99.4/99.4%4.03/3.522.42/2.150.92/0.880.94/0.92
SimVQ (ours)1.2kbps99.4/99.0%4.03/3.522.54/2.260.93/0.900.94/0.92
SimVQ (ours)1.35kbps95.6/94.7%4.03/3.532.61/2.310.93/0.900.95/0.93
", + "bbox": [ + 143, + 88, + 857, + 296 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/91167f61c8c19f7c0c766a91a082ba6ad1c736e4e5f5ba99230064ebf0d03df9.jpg", + "table_caption": [ + "Table 2. Reconstruction performance on LibriTTS test-clean/test-other dataset. * WavTokenizer is trained with a window size of 3 seconds. The bandwidth of 0.9kbps, 0.975kbps, 1.2kbps, 1.35kbps means the codebook size of 4096, 8192, 65536, 262144 respectively." + ], + "table_footnote": [], + "table_body": "
MethodCodebook SizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN-LC-CLIP† [42]50,00099.9%2.750.1323.858.4
VQGAN-LC-CLIP† [42]100,00099.9%2.620.1223.858.9
VQGAN-LC-CLIP† [42]200,00099.8%2.660.1223.959.2
SimVQ1,024100.0%3.670.1622.3470.8
SimVQ8,192100.0%2.980.1423.2374.7
SimVQ65,536100.0%2.240.1224.1578.4
SimVQ262,144100.0%1.990.1124.6880.3
", + "bbox": [ + 93, + 357, + 480, + 446 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc5a6459ba67b87be6675f4db95627fb6edf24f95d25769183437e2548600549.jpg", + "table_caption": [ + "Table 3. Ablation study on the effect of various codebook sizes on ImageNet at a resolution of $128 \\times 128$ . $\\dagger$ We directly copy the reported results of VQGAN-LC from the original paper on ImageNet $256 \\times 256$ resolution." + ], + "table_footnote": [], + "table_body": "
InitializationTrainableUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
GaussianYes100.0%2.310.1224.0477.2
UniformNo100.0%2.310.1224.1578.4
GaussianNo100.0%2.240.1224.1578.4
", + "bbox": [ + 94, + 523, + 480, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study of codebook optimization strategy.", + "bbox": [ + 111, + 593, + 460, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nal pre-trained CLIP model to provide a well-defined latent space. However, VQGAN-LC relies on CLIP features pretrained on much larger datasets than ImageNet, which introduces generalization issues and a lower performance ceiling (degradation issue in Tab. 3). In contrast, SimVQ can be applied to a wide range of data types and achieves superior performance (rFID 2.40 vs. 2.24) without the limitations imposed by a pre-trained feature extraction model.", + "bbox": [ + 89, + 632, + 482, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 760, + 243, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the Codebook Size In Tab. 3, we explore the impact of different codebook sizes, ranging from $1k$ to $262k$ , which is typically the level of LLM's vocabulary size. SimVQ consistently improves performance as the codebook size increases. For instance, the rFID score decreases to 1.99, and SSIM surpasses 80.0. In contrast, while VQGAN-LC-CLIP can keep high codebook utilization as increasing codebook size, it encounters performance degradation, with the rFID", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "score worsening from 2.62 to 2.66 when the codebook size is increased from 100,000 to 200,000.", + "bbox": [ + 511, + 359, + 906, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the Codebook Optimization Strategy We investigate codebook initialization and the training of the linear layer in Tab. 4. Our findings are as follows: 1) The codebook is robust to different initialization strategies, yielding similar results with both Gaussian and uniform initialization. 2) When the codebook is updated during training, SimVQ continues to address the representation collapse issue, though with a slight degradation in performance.", + "bbox": [ + 511, + 411, + 906, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Audio Modality", + "text_level": 1, + "bbox": [ + 511, + 542, + 669, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1. Baselines and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 565, + 821, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the LibriTTS dataset [36] for audio-based VQ model training. The baselines such as Encodec [6], VOCos [24], and SpeechTokenizer [39] are based on residual vector quantization method. Our SimVQ model adopts the same architecture as WavTokenizer [13] with the only modification being the replacement of their EMA codebook with our one linear layer reparameterization method. We train SimVQ on LibriTTS-580h for 50 epochs with a batch size of 64. Note that WavTokenizer is trained with a 3-second window size for optimal performance, we train SimVQ using a 1-second window to accelerate training. For objective evaluation of the reconstructed audio, we follow VOCos [24] and employ metrics such as UTMOS [23], PESQ [20], STOI, and the F1 score for voiced/unvoiced classification (V/UV F1). UTMOS is particularly valuable as it produces scores highly correlated with human evaluations.", + "bbox": [ + 511, + 584, + 906, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 835, + 653, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 presents the reconstruction performance of various VQ models on audio data. Baseline models using residual vector quantization perform significantly worse than", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22974", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c4d5bb5c26c628314a7b0a9b5f415c277866b4fcfb803dbcbed7b2a897f0476.jpg", + "image_caption": [ + "Figure 4. (a):(left) The rank of latent basis matrix $\\mathbf{W}$ over training epochs. (b):(right) The Frobenius norm of latent basis matrix $\\mathbf{W}$ over training epochs." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 480, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SimVQ, even when utilizing much larger bandwidths. Despite using the same architecture as WAVTokenizer, our model, which replaces the quantization layer with SimVQ, achieves superior performance with a 1-second window size and maintains nearly $100\\%$ codebook utilization when scaling up to a size of 262,144. The consistent performance of the SimVQ model across both image and audio data demonstrates that SimVQ is a general method for addressing the representation collapse problem in VQ models and can be effectively applied across multiple modalities.", + "bbox": [ + 88, + 287, + 482, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 89, + 450, + 192, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 4(a), we plot the rank of the latent basis matrix over training epochs. Notably, SimVQ demonstrates the ability to adaptively adjust the rank of the latent space. Specifically, when the codebook size increases from $65k$ to $262k$ , the rank of the latent basis matrix decreases more rapidly and converges to a lower value. This observation suggests that a larger codebook can effectively alleviate the pressure on the latent space dimensionality, allowing the model to learn to represent data more efficiently. Additionally, despite the rank decreasing to a lower-rank space, SimVQ maintains $100\\%$ codebook utilization, highlighting its superiority over VQGAN-FC, which struggles when increasing the latent dimension from 8 to 128. We also calculate the Frobenius norm of the latent basis matrix, as shown in Fig. 4. The norm of a codebook size of $262k$ is slightly larger than for $65k$ , indicating that a larger codebook can span a broader area in the linear space. For a comprehensive evaluation, we also provide the reconstruction loss curve on the ImageNet validation dataset in Appendix 8.2. The results consistently show that SimVQ achieves improved performance, further validating the effectiveness of our approach.", + "bbox": [ + 91, + 474, + 482, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 801, + 302, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the distribution of encoder features and codebook embeddings in Fig. 5 and the frequency in Appendix 8.3. Compared to vanilla VQ models that most codebook vectors are not unused, SimVQ can update the whole codebook and distribute the features evenly in the whole space.", + "bbox": [ + 89, + 824, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da97486972720edf2e8b89b0011a519ffb75d5ce39f845826b6685e876e4a1f4.jpg", + "image_caption": [ + "Figure 5. Visualization of the divergence between the encoder features and codebook embeddings on a random subset of ImageNet validation dataset. The left figure is vanilla VQ model and the right one is SimVQ." + ], + "image_footnote": [], + "bbox": [ + 557, + 89, + 866, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We qualitatively compare the reconstruction quality of both images and audio in Appendix 8.4. SimVQ can preserve more details with an enlarged codebook size, such as \"eyes\" and \"text\", which are challenging for vanilla VQ models.", + "bbox": [ + 511, + 311, + 905, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 511, + 388, + 627, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VQ Performance and Generative Models Many generative models [21, 25, 28, 34] utilize VQ models as a \"tokenizer\" to obtain discrete tokens. While it is intuitive that VQ reconstruction quality should impact generation performance, recent studies [35, 43] have revealed that the relationship is more nuanced: improved VQ reconstruction metrics do not necessarily translate to better generative outcomes. This complex relationship suggests that evaluating VQ models primarily through downstream generation tasks may not provide the most insightful assessment of their fundamental properties. Therefore, we focus on addressing the critical issue of representation collapse in VQ models, aiming to advance our understanding of their core representation learning mechanisms.", + "bbox": [ + 511, + 414, + 906, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 641, + 633, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore the representation collapse problem in VQ models. We conduct a theoretical analysis of the optimization process in VQ models and propose a simple yet effective method, SimVQ, to address this issue. Our method addresses the representation collapse by jointly optimizing the latent space through a linear transformation with one linear layer. Experimental results demonstrate that SimVQ outperforms previous approaches on both image and audio datasets, highlighting its broad applicability across diverse machine learning tasks.", + "bbox": [ + 511, + 667, + 906, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 833, + 671, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the National Natural Science Foundation of China (Grant No.62276245).", + "bbox": [ + 511, + 857, + 905, + 887 + ], + "page_idx": 7 + }, + { + "type": "header", + "text": "Rank of W (Codebook Size = 65536)", + "bbox": [ + 509, + 58, + 611, + 65 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22975", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 1", + "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3", + "[4] Zalan Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523-2533, 2023. 1", + "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Maria Elisabeth Bechtle, Feryal Behbahani, Stephanie C.Y. Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando De Freitas, Satinder Singh, and Tim Roktaschel. Genie: Generative interactive environments. In Proceedings of the 41st International Conference on Machine Learning, pages 4603-4623. PMLR, 2024. 1", + "[6] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. Transactions on Machine Learning Research, 2023. Featured Certification, Reproducibility Certification. 1, 7", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6", + "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1", + "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 1, 2, 6", + "[10] Zhangyang Gao, Cheng Tan, Jue Wang, Yufei Huang, Lirong Wu, and Stan Z Li. Foldtoken: Learning protein language via vector quantization and beyond. arXiv preprint arXiv:2403.09673, 2024. 1", + "[11] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized net" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works. In Proceedings of the 40th International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3", + "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 3", + "[13] Shengpeng Ji, Ziyue Jiang, Xize Cheng, Yifu Chen, Minghui Fang, Jialong Zuo, Qian Yang, Ruiqi Li, Zhang Zhang, Xiaoda Yang, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532, 2024. 7", + "[14] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 3", + "[15] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2, 6", + "[16] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6", + "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[18] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3", + "[19] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 6", + "[20] A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra. Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and CODECs. In 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), pages 749-752 vol.2, 2001. 7", + "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 8", + "[22] Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018. 1, 3", + "[23] Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. ArXiv, abs/2204.02152, 2022. 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22976", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Hubert Siuzdak. Vocos: Closing the gap between timedomain and fourier-based neural vocoders for high-quality audio synthesis. In The Twelfth International Conference on Learning Representations, 2024. 7", + "[25] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 8", + "[26] Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi, Toshiyuki Kumakura, and Yuki Mitsu-fuji. SQ-VAE: Variational Bayes on discrete representation with self-annealed stochastic quantization. In Proceedings of the 39th International Conference on Machine Learning, pages 20987-21012. PMLR, 2022. 1, 3, 6", + "[27] Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, and Ngai Wong. Scaling laws with vocabulary: Larger models deserve larger vocabularies. arXiv preprint arXiv:2407.13623, 2024. 1", + "[28] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 8", + "[29] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 2, 3", + "[30] Tung-Long Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quantized wasserstein auto-encoder. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 1, 3, 6", + "[31] Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023. 1", + "[32] Pan Xiao, Peijie Qiu, and Aristeidis Sotiras. Scvae: Sparse coding-based variational autoencoder. ArXiv, abs/2303.16666, 2023. 1, 3, 6", + "[33] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In International Conference on Learning Representations, 2022. 1, 3, 6", + "[34] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 8", + "[35] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 8" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. Libritts: A corpus derived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882, 2019. 7", + "[37] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757-15773, Singapore, 2023. Association for Computational Linguistics. 1", + "[38] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18467-18476, 2023. 3", + "[39] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech language models. In The Twelfth International Conference on Learning Representations, 2024. 7", + "[40] Chuanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798-22807, 2023. 1, 3, 6", + "[41] Chuanxia Zheng, Guoxian Song, Tat-Jen Cham, Jianfei Cai, Dinh Q. Phung, and Linjie Luo. High-quality pluralistic image completion via code shared vqgan. ArXiv, abs/2204.01931, 2022. 2, 6", + "[42] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of $99\\%$ . ArXiv, abs/2406.11837, 2024. 1, 2, 3, 6, 7", + "[43] Yongxin Zhu, Bocheng Li, Hang Zhang, Xin Li, Linli Xu, and Lidong Bing. Stabilize the latent space for image autoregressive modeling: A unified perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8", + "[44] Yongxin Zhu, Dan Su, Liqiang He, Linli Xu, and Dong Yu. Generative pre-trained speech language model with efficient hierarchical transformer. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1764-1775, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + ], + "bbox": [ + 516, + 92, + 903, + 695 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22977", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_model.json b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4600b7a6703ad280566d8f36a699c3e6fc4a1d6a --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.13, + 0.851, + 0.176 + ], + "angle": 0, + "content": "Addressing Representation Collapse in Vector Quantized Models with One Linear Layer" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.203, + 0.774, + 0.221 + ], + "angle": 0, + "content": "Yongxin Zhu\\(^{1,2}\\), Bocheng Li\\(^{1,2}\\), Yifei Xin\\(^{3}\\), Zhihua Xia\\(^{4}\\), Linli Xu\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.222, + 0.692, + 0.239 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.239, + 0.839, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)State Key Laboratory of Cognitive Intelligence, \\(^{3}\\)Peking University, \\(^{4}\\)Jinan University" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.258, + 0.77, + 0.274 + ], + "angle": 0, + "content": "zyx2016@mail.ustc.edu.cn,bcli@mail.ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.275, + 0.873, + 0.293 + ], + "angle": 0, + "content": "xinyifei@stu.pku.edu.cn, xiazhihua@jnu.edu.cn, linlixu@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.486, + 0.631 + ], + "angle": 0, + "content": "Vector Quantization (VQ) is essential for discretizing continuous representations in unsupervised learning but suffers from representation collapse, causing low codebook utilization and limiting scalability. Existing solutions often rely on complex optimizations or reduce latent dimensionality, which compromises model capacity and fails to fully solve the problem. We identify the root cause as disjoint codebook optimization, where only a few code vectors are updated via gradient descent. To fix this, we propose SimpleVQ, which reparameterizes code vectors through a learnable linear transformation layer over a latent basis, optimizing the entire linear space rather than nearest individual code vectors. Although the multiplication of two linear matrices is equivalent to applying a single linear layer, this simple approach effectively prevents collapse. Extensive experiments on image and audio tasks demonstrate that SimVQ improves codebook usage, is easy to implement, and generalizes well across modalities and architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.222, + 0.674 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.881 + ], + "angle": 0, + "content": "In recent years, vector quantization (VQ) [19, 29] has emerged as a foundational technique in unsupervised representation learning [2, 5] and latent generative models [4, 21, 31, 33, 34, 44]. By converting continuous representations into discrete codes, VQ models can effectively identify the inherent structure of data and enable various discrete modeling methods on continuous data, from high-quality image generation [9] to audio synthesis [6]. The recent success of Large Language Models (LLMs) [1] has highlighted the effectiveness of next-token prediction as a powerful and versatile training objective. Consequently, VQ models are taken as the direct method to transform data from various modalities [25, 28, 37] or scientific domains [10] to discrete" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.906, + 0.45 + ], + "angle": 0, + "content": "sequences for next token prediction training. However, attempts to integrate VQ models as multimodal tokenizers to leverage the scaling laws of LLMs face significant challenges because of the difficulty of expanding the codebook. For example, the Chameleon model [28] constrains its codebook size to \\( 8k \\), which is significantly trailing behind the vocabulary size of LLMs (e.g., LLaMA3's vocabulary size is \\( 128k \\) [8])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.908, + 0.775 + ], + "angle": 0, + "content": "There is a broad agreement that increasing vocabulary size can consistently improve the performance of LLMs [27]. However, recent studies [42] indicate that traditional VQ models often fail to utilize the additional parameters introduced by codebook expansion, leaving most codes inactive during training. The contradiction between codebook expansion and low codebook utilization in VQ models is known as the representation collapse problem [22], where increasing the codebook size fails to improve the performance. To address these discrepancies, we conduct a theoretical analysis of the optimization procedure of VQ models and identify that the disjoint optimization of the codebook is the root cause of representation collapse. As illustrated in Fig. 1(a), the core mechanism of VQ models involves a nearest-neighbor replacement strategy, where the encoder's output features are replaced by the nearest vector in the codebook to serve as input to the decoder. The indices of the nearest vector are taken as the discrete representation of the data. This nearest-selection operator results in only a subset of codes being updated through gradient descent, while the remaining codes remain unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Some methods mitigate representation collapse by hand-designing complex optimization strategies, such as stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] through sophisticated training strategies. Recently, some approaches [16, 33, 35] propose to reduce the dimension of the latent space to a very small scale (e.g., 8 v.s. 128) to alleviate the curse of dimensionality, thereby improving the overlap between the encoder's fea" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22968" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.505, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.087, + 0.888, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.256, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 1. Comparison of Vanilla VQ and SimVQ. (a): (left) Disjoint optimization in Vanilla VQ. Only the nearest codes are updated, resulting in a high percentage of \"dead\" codes that are not updated. (b): (right) Joint optimization in SimVQ. The entire codebook is updated with a latent basis, ensuring all codes remain active." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.55 + ], + "angle": 0, + "content": "tures and the codebook. However, while these methods enhance codebook utilization, they do so at the cost of model capacity, leading to worse performance compared to vanilla VQ models when the codebook size is small or representation collapse is not severe. Another approach, VQGAN-LC [42], initializes the codebook with features extracted from the pre-trained CLIP model [17] to create a well-structured latent space that better matches the distribution of the encoder output. Nevertheless, the latent space defined by an external pre-trained model limits the model's ability to generalize to diverse datasets and reaches a performance plateau as the codebook size increases. These limitations highlight the need for a more effective method to improve codebook utilization without compromising model capacity or relying on external models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.485, + 0.688 + ], + "angle": 0, + "content": "We critically assess prevalent methodologies and reveal that optimizing the latent space rather than individual code vectors is key to preventing representation collapse. Building on this insight, we introduce a simple yet effective method, termed SimVQ, to directly update the latent space spanned by the codebook by linear transforming the code vectors via a learnable latent basis. Specifically, the vectors in the codebook are reparameterized as a linear combination of the basis in the learnable linear layer \\( W \\):" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.694, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} \\in \\mathbb {R} ^ {K \\times d} \\Rightarrow \\boldsymbol {C W} \\text {w i t h} \\boldsymbol {W} \\in \\mathbb {R} ^ {d \\times d}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "where \\(K\\) denotes the codebook size and \\(d\\) represents the dimension of latent space. This reparameterization with linear transformation disentangles the optimization of the codebook into two components: the coefficient matrix \\(\\mathbf{C}\\) and the basis of linear space \\(\\mathbf{W}\\) respectively. As illustrated in Fig. 1(b), by optimizing the basis matrix \\(\\mathbf{W}\\), the latent space spanned by \\(\\mathbf{CW}\\) is rotated and stretched to match encoder's output feature. The entire codebook is updated jointly to prevent the representation collapse problem. The simplicity of the proposed method makes it highly portable and easily adaptable for improving VQ-based models across a wide range of domains, requiring only one linear layer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.905, + 0.354 + ], + "angle": 0, + "content": "In summary, our contributions to vector quantized models are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.417 + ], + "angle": 0, + "content": "- We theoretically analyze the representation collapse problem in VQ models and reveal that optimizing the latent space spanned by the codebook, rather than individual code vectors, is crucial to addressing this issue." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.906, + 0.508 + ], + "angle": 0, + "content": "- We propose a novel method, SimVQ, which reparameterizes the codebook vectors in VQ models via a linear transformation with a learnable latent basis. This simple yet effective approach is highly adaptable and easy to implement, making it broadly applicable across various machine learning contexts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.907, + 0.645 + ], + "angle": 0, + "content": "- We conduct an extensive evaluation of SimVQ across diverse modalities, including image and audio with different model architectures. The results show that SimVQ not only effectively addresses the representation collapse problem by achieving near-complete codebook utilization regardless of the codebook size, but also establishes new state-of-the-art performance. Furthermore, when scaling up the codebook size, SimVQ consistently delivers improved results." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.657, + 0.679 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "VQ-VAE [29] is the pioneering work to encode data into discrete representations. Building on these developments, VQGAN [9] combines VQ-VAE with adversarial networks to improve the perceptual quality of generated samples and establish a fundamental quantization protocol in latent generative models [21, 28, 34]. Many traditional VQ works focus on training a better discrete representation rather than codebook utilization to improve reconstruction performance. For example, RVQ [15] and MoVQ [41] enhance the reconstruction details with multichannel quantization. However, these methods suffer from a critical issue of representation collapse, as they struggle to scale the codebook size beyond \\(10\\mathrm{k}\\) entries, limiting their scalability. In response to this challenge, several approaches have been pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22969" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.574 + ], + "angle": 0, + "content": "posed recently. DALLE [18] employs the gumbel-softmax trick [12] and stochastic sampling strategies to activate most codes during training. However, during inference, only a small subset of codes is utilized for quantization [38]. Some methods design complex optimization strategies to improve codebook utilization, such as stochastic quantization [26], distribution penalty [30, 32] and codebook reset [40]. Huh et al. [11] proposes rescaling the vectors in the codebook during training to match the distributions in the latent space. VQGAN-FC [33] introduces a method to map latent vectors into a lower-dimensional space followed by \\( l_{2} \\) normalization to alleviate representation collapse. FSQ [16] extends this idea by projecting representations into a reduced-dimensional space, where they are quantized into a small set of fixed values. LFQ [35], a variant of FSQ, uses binary values for quantized representations, thereby simplifying the encoding process. While these methods improve the codebook utilization, they do so at the cost of model capacity by significantly reducing the dimensionality of latent space (often to as low as 8), leading to worse performance compared to vanilla VQ models when the codebook size is small and representation collapse is not severe. Additionally, VQGAN-LC [42] proposes to initialize the codebook using features extracted from the pre-trained CLIP model to avoid representation collapse. However, the reliance on the pre-trained model limits the VQ model's ability to generalize to diverse datasets and results in a performance plateau as the codebook size increases. In contrast, our method, SimVQ, effectively addresses the representation collapse problem with a simple linear layer, without sacrificing model capacity or relying on external pre-trained models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.442, + 0.613 + ], + "angle": 0, + "content": "3. Representation Collapse in VQ Models" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.622, + 0.233, + 0.637 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.859 + ], + "angle": 0, + "content": "A vector quantized model is typically a reconstructive encoder-decoder architecture that includes a vector quantization layer to convert continuous representations into discrete codes. For simplicity, we represent an image with a single random variable \\( x \\). Formally, the encoder \\( f_{\\theta} \\) maps the input image into a latent space, producing a continuous representation \\( z_{e} = f_{\\theta}(x) \\in \\mathbb{R}^{d} \\). This representation is then quantized using a learnable codebook \\( C = [q_{1}, \\ldots, q_{K}] \\in \\mathbb{R}^{K \\times d} \\), where \\( q_{i} \\) is a codebook vector. We define \\( \\delta_{k} \\in \\{0, 1\\}^{1 \\times K} \\) as a characteristic (one-hot) vector where only the \\( k \\)-th element is 1, such that \\( q_{k} = \\delta_{k} C \\in \\mathbb{R}^{1 \\times d} \\). The quantization layer selects the nearest codebook vector \\( q_{k} \\) by minimizing the Euclidean distance between \\( z_{e} \\) and the codebook entries [29]:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.873, + 0.484, + 0.898 + ], + "angle": 0, + "content": "\\[\nk = \\arg \\min _ {j} \\| z _ {e} - q _ {j} \\| _ {2} ^ {2} = \\arg \\min _ {j} \\| z _ {e} - \\delta_ {j} C \\| _ {2} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "The selected vector \\(q_{k}\\) is then passed to the decoder \\(g_{\\phi}\\) to reconstruct the input image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.184 + ], + "angle": 0, + "content": "To enable gradient propagation through the non-differentiable characteristic vector \\(\\delta_{k}\\), the straight-through estimator (STE) [3] is applied. During the backward pass, the gradient of \\(z_{q} = \\delta_{k}C\\) is copied to \\(z_{e}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.193, + 0.905, + 0.226 + ], + "angle": 0, + "content": "\\[\nz _ {q} = \\mathrm {s g} (\\delta_ {k} \\pmb {C} - z _ {e}) + z _ {e}, \\Rightarrow \\frac {\\partial z _ {q}}{\\partial z _ {e}} = 1 (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.264 + ], + "angle": 0, + "content": "where \\(\\mathfrak{sg}\\) is the stop gradient operator, ensuring the gradient for \\(\\delta_k C\\) is discarded during the backward pass." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.324 + ], + "angle": 0, + "content": "The learning objective is the combination of a reconstruction loss and commitment loss that ensures that the encoder commits to an embedding and the encoder's output does not drift:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\log p (x | z _ {q}) + \\| \\operatorname {s g} \\left(\\delta_ {k} \\boldsymbol {C}\\right) - z _ {e} \\| _ {2} ^ {2} + \\beta \\| \\delta_ {k} \\boldsymbol {C} - \\operatorname {s g} \\left(z _ {e}\\right) \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.905, + 0.399 + ], + "angle": 0, + "content": "where \\(\\log p(x|z_q)\\) is typically the mean squared error (MSE) loss \\(\\| x - g_{\\phi}(z_q)\\| _2^2\\) for image and audio data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.82, + 0.423 + ], + "angle": 0, + "content": "3.2. Disjoint Optimization of Codebook" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.548 + ], + "angle": 0, + "content": "In VQ models, only the nearest code is selected and updated via gradient descent. Ideally, all codebook entries should be updated and utilized for decoding. However, experimental evidence shows that only a small fraction of the codebook gets updated and utilized, leading to what is known as the representation collapse problem [22]. To investigate the root cause of this issue, we provide a theoretical analysis of the optimization dynamics in VQ models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.608 + ], + "angle": 0, + "content": "Due to the use of the straight-through estimator (STE) for gradient propagation, the codebook \\( C \\) can only be updated through the gradient of the commitment loss, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.621, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C}) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\right\\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.905, + 0.68 + ], + "angle": 0, + "content": "The codebook \\( \\pmb{C} \\) is updated according to the following equation, where \\( \\eta \\) is the learning rate:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.69, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {C} ^ {(t + 1)} = \\boldsymbol {C} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C} ^ {(t)})}{\\partial \\boldsymbol {C} ^ {(t)}} \\right] (6) \\\\ = \\boldsymbol {C} ^ {(t)} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} ^ {(t)} \\right] + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} z _ {e} \\right] (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.763, + 0.868, + 0.78 + ], + "angle": 0, + "content": "where \\(\\delta_k^T\\delta_k\\) is the Kronecker delta matrix, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.791, + 0.905, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\left(\\delta_ {k} ^ {T} \\delta_ {k}\\right) _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j = k, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "All vectors in \\(C\\) will be updated and utilized if and only if the expectation \\(\\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right]\\) converges to the identity matrix. Unlike variational autoencoders (VAEs) [14], which enforce a Gaussian distribution on the latent space via a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22970" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "KL-divergence penalty, VQ models optimize \\( z_{e} \\) towards the selected codebook vectors \\( \\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_kC\\right] \\). At the same time, the selected codebook vectors are optimized towards the distribution of \\( z_{e} \\), resulting in the same selected subset of vectors moving closer to \\( z_{e} \\), somewhat akin to a cocoon effect. However, this disjoint optimization of the codebook leads to part of the codebook, specifically \\( (I - \\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right])C \\), remaining un-updated and underutilized once the optimization process begins. This phenomenon occurs because the optimization focuses only on a subset of codebook vectors, leaving other vectors stagnant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.364 + ], + "angle": 0, + "content": "This analysis reveals the fundamental cause of representation collapse in VQ models: the disjoint optimization process that updates only a subset of codebook vectors. This insight forms the basis for our proposed solution, SimVQ, which aims to address this issue by optimizing the entire latent space spanned by the codebook, rather than individual code vectors." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.414 + ], + "angle": 0, + "content": "4. Addressing Collapse with Latent Linear Transformation" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.423, + 0.441, + 0.44 + ], + "angle": 0, + "content": "4.1. Reparameterize Codes with Latent Basis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.507 + ], + "angle": 0, + "content": "Let \\( \\pmb{W} = \\{\\pmb{w}_1, \\dots, \\pmb{w}_n\\} \\) be a basis of a linear space. Any vector \\( \\pmb{v} \\) in the space can be uniquely expressed as a linear combination of the basis vectors with coefficients \\( c_1, \\dots, c_n \\in \\mathbb{R} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.522, + 0.482, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} = c _ {1} \\boldsymbol {w} _ {1} + \\dots + c _ {n} \\boldsymbol {w} _ {n} = \\boldsymbol {c} \\boldsymbol {W}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Given the equivalence between \\(\\pmb{v}\\) and \\(c\\pmb{W}\\) in the linear space, we can reparameterize each vector in the codebook of VQ models with a new basis matrix \\(\\pmb{W} \\in \\mathbb{R}^{d \\times d}\\). Specifically, the codebook \\(\\pmb{C} = \\{c_1, \\dots, c_K\\}\\) can be reparameterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.638, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {\\boldsymbol {c}} _ {1} \\boldsymbol {W}, \\dots , \\hat {\\boldsymbol {c}} _ {N} \\boldsymbol {W} \\} = \\hat {\\boldsymbol {C}} \\boldsymbol {W} \\in \\mathbb {R} ^ {K \\times d}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.746 + ], + "angle": 0, + "content": "This reparameterization introduces two components: the basis matrix \\( \\mathbf{W} \\) and the coefficient matrix \\( \\hat{\\mathbf{C}} \\). In the following, we will discuss the optimization of both the basis matrix \\( \\mathbf{W} \\) and the coefficient matrix \\( \\hat{\\mathbf{C}} \\). For simplicity, we slightly abuse \\( \\mathbf{C} \\) and \\( \\hat{\\mathbf{C}} \\) below." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.409, + 0.774 + ], + "angle": 0, + "content": "4.2. Asymmetric Optimization Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "While it is commonly accepted that multiplying two linear matrices is equivalent to a single linear layer, we argue that the disjoint optimization problem of the codebook in VQ models can be addressed by linear transformation. In vanilla VQ models, only the codebook \\( C \\) is responsible for minimizing commitment loss, leading to the disjoint optimization problem where only the selected codes will be updated." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.71, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.091, + 0.903, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.202, + 0.906, + 0.285 + ], + "angle": 0, + "content": "Figure 2. (a): (left) The optimization trajectory of the objective \\(\\| \\pmb{x} - \\pmb{q}\\|_2^2\\), which is the same as vanilla VQ. Only a small fraction of points are updated while others remain inactive. (b): (right) The optimization trajectory of the objective \\(\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2\\) with \\(\\pmb{q}\\) frozen, which is the same as SimVQ. All the points are updated towards targets \\(x\\)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.3, + 0.703, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.3, + 0.903, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.407, + 0.906, + 0.464 + ], + "angle": 0, + "content": "Figure 3. (a): (left) The optimization trajectory of the optimization objective: \\(\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2\\) with both \\(\\pmb{q}\\) and \\(\\pmb{w}\\) unfrozen. (b): (right) The Frobenius norm of the projection matrix \\(\\pmb{w}\\) and loss curves. The loss quickly converges to 0 with \\(\\pmb{w}\\) almost unchanged." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.49, + 0.906, + 0.578 + ], + "angle": 0, + "content": "In contrast, when the codebook is reparameterized as \\( CW \\), both the basis \\( W \\) and the coefficient matrix \\( C \\) contribute to minimizing the commitment loss. The gradients \\( \\frac{\\partial\\mathcal{L}}{\\partial W} \\) and \\( \\frac{\\partial\\mathcal{L}}{\\partial C} \\) can simultaneously reduce the loss. As a result, the optimization of the reparameterized codebook can be divided into three scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.58, + 0.905, + 0.625 + ], + "angle": 0, + "content": "- Updating \\( C \\) with \\( W \\) frozen: Only the selected codes adapt to the latent distribution of \\( z_{e} \\), as depicted on Fig. 1(a). The vanilla VQ is a special case with \\( W = I \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.625, + 0.905, + 0.67 + ], + "angle": 0, + "content": "- Updating \\(W\\) with \\(C\\) frozen: The entire codebook \\(CW\\) adjusts to the latent distribution of \\(z_{e}\\). The basis matrix \\(W\\) rotates and stretches the space as shown in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.671, + 0.905, + 0.716 + ], + "angle": 0, + "content": "- Updating both \\( C \\) and \\( W \\): The selected subset of codes moves towards \\( z_{e} \\) while the space spanned by \\( W \\) undergoes simultaneous rotation and stretching." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.58, + 0.905, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.906, + 0.777 + ], + "angle": 0, + "content": "To highlight the difference in optimization between \\( C \\) and \\( CW \\), we conduct a toy experiment in a two-dimensional setting and visualize the optimization process in Fig. 2 and Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.659, + 0.799 + ], + "angle": 0, + "content": "4.2.1. Toy Examples" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.802, + 0.906, + 0.831 + ], + "angle": 0, + "content": "We randomly sample two target points \\( x \\) from Gaussian distribution as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.84, + 0.905, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {1} \\sim \\mathcal {N} (\\binom {2} {2}, \\binom {1 0} {0 1}), \\quad \\boldsymbol {x} _ {2} \\sim \\mathcal {N} (\\binom {- 2} {- 2}, \\binom {1 0} {0 1}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Then we initialize 10 learnable vectors \\( \\mathbf{q} \\) from a Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22971" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.388, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 Training Procedure for SimVQ" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.111, + 0.483, + 0.143 + ], + "angle": 0, + "content": "Input: Encoder \\(f_{\\theta}\\), Decoder \\(g_{\\phi}\\), Codebook \\(C \\in \\mathbb{R}^{K \\times d}\\), Linear projector matrix \\(\\boldsymbol{W}_{\\psi}\\), commitment weight \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.144, + 0.485, + 0.187 + ], + "angle": 0, + "content": "Output: Model parameters \\(\\theta, \\phi, \\psi\\) and Codebook \\(C\\). Initialize Codebook with Gaussian distribution and freeze the parameter of Codebook \\(C\\);" + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.19, + 0.157, + 0.203 + ], + "angle": 0, + "content": "repeat" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.204, + 0.262, + 0.219 + ], + "angle": 0, + "content": "Draw \\(x\\sim p_{data}(\\pmb {x})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.22, + 0.207, + 0.234 + ], + "angle": 0, + "content": "\\[\nz _ {e} = f _ {\\theta} (x);\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.234, + 0.454, + 0.249 + ], + "angle": 0, + "content": "\\(/^{*}\\) Replace \\(q_{j}\\) in vanilla VQ with proposed \\(q_{j}\\pmb{W}_{\\psi}\\)." + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.249, + 0.482, + 0.279 + ], + "angle": 0, + "content": "Nearest code search: \\( k = \\arg \\min_{j} \\| z_{e} - q_{j}W_{\\psi} \\|_{2}^{2} \\), where \\( q_{j} \\in C \\);" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.279, + 0.482, + 0.308 + ], + "angle": 0, + "content": "Straight Through Estimation: \\(z_{q} = \\mathrm{sg}(q_{k}\\mathbf{W}_{\\psi} - z_{e}) + z_{e}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.31, + 0.208, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = g _ {\\phi} (z _ {q});\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.325, + 0.482, + 0.355 + ], + "angle": 0, + "content": "Minimize \\(\\mathcal{L}(\\theta ,\\phi ,\\psi) = \\mathrm{MSE}(x,\\hat{x})\\) + \\(\\beta \\| z_{e}\\) - \\(\\mathrm{sg}(q_kW_\\psi)\\| _2^2 +\\| \\mathrm{sg}(z_e) - q_kW_\\psi \\| _2^2;\\)" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.356, + 0.216, + 0.369 + ], + "angle": 0, + "content": "until converged" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.401, + 0.175, + 0.414 + ], + "angle": 0, + "content": "distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.424, + 0.482, + 0.459 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {q} _ {i} \\right\\} _ {i = 1} ^ {1 0} \\sim \\mathcal {N} \\left(\\binom {0} {0}, \\binom {1} {0}, \\binom {0} {1}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.591 + ], + "angle": 0, + "content": "During training with gradient descent, we introduce perturbation noise \\(\\mathcal{N}(0,0.01)\\) to the targets. In Fig. 2(a), the optimization objective is similar to vanilla VQ: \\(\\| \\pmb {x} - \\pmb {q}\\| _2^2\\). Only the nearest points \\(q_{4}\\) and \\(q_{10}\\) are updated. In contrast, in Fig. 2(b), the optimization objective \\(\\| \\pmb {x} - \\pmb {q}\\pmb {w}\\| _2^2\\) is similar to SimVQ with the points reparameterized by a learnable latent basis \\(\\pmb{w}\\) and \\(\\pmb{q}\\) frozen, resulting in the entire codebook \\(\\{\\pmb {q}\\}_{i = 1}^{10}\\) being jointly updated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.632 + ], + "angle": 0, + "content": "Remark 1. The simultaneous optimization of the latent basis \\(\\mathbf{w}\\) and the coefficient matrix \\(\\mathbf{q}\\) may lead to the collapse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.824 + ], + "angle": 0, + "content": "We provide an example in Fig. 3(a) where the optimization objective is \\( \\| \\pmb{x} - \\pmb{q}\\pmb{w} \\|_2^2 \\) with \\( \\pmb{q} \\) unfrozen this time. In the training process, only the nearest point \\( \\pmb{q}_1 \\) and point \\( \\pmb{q}_{10} \\) move towards the target point, while other points remain almost unchanged. We also visualize the loss curve in Fig. 3(b). The optimization objective with both \\( \\pmb{q} \\) and \\( \\pmb{w} \\) unfrozen converges quickly, where the norm of basis \\( \\pmb{w} \\) is much smaller than the objective with \\( \\pmb{q} \\) frozen. This indicates that the disjoint optimization of the codebook persists: \\( \\pmb{q} \\) can directly commit to the loss and dominate the optimization process, with \\( \\pmb{w} \\) being ignored, leading to the collapse quickly." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.833, + 0.403, + 0.849 + ], + "angle": 0, + "content": "4.3. Joint Optimization of the Codebook" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We propose SimVQ by simply using a learnable basis \\(W \\in \\mathbb{R}^{d \\times d}\\) to reparameterize the codebook such that the codebook is transformed into \\(CW\\). The pseudo-code for this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "approach is provided in Algorithm 1. During training, we optimize only the latent basis matrix \\( \\mathbf{W} \\), while keeping the coefficient matrix \\( \\mathbf{C} \\) frozen. The commitment loss for SimVQ is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.162, + 0.905, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e}, q _ {k}\\right) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\boldsymbol {W} \\right\\| _ {2} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.191, + 0.906, + 0.237 + ], + "angle": 0, + "content": "The vanilla VQ model is a special case of SimVQ, where the linear basis matrix \\(\\mathbf{W}\\) is fixed as the identity matrix \\(\\mathbf{I}\\). The update for \\(\\mathbf{W}\\) with learning rate \\(\\eta\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.246, + 0.905, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {W} ^ {(t + 1)} = \\boldsymbol {W} ^ {(t)} - \\eta \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e} , \\boldsymbol {q} _ {k}\\right)}{\\partial \\boldsymbol {W} ^ {(t)}} \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.278, + 0.905, + 0.311 + ], + "angle": 0, + "content": "\\[\n= \\left(\\boldsymbol {I} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} \\right]\\right) \\boldsymbol {W} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} z _ {e} \\right]. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.906, + 0.382 + ], + "angle": 0, + "content": "The term \\(\\mathbb{E}\\left[\\boldsymbol{C}^T\\delta_k^T\\delta_k\\boldsymbol{C}\\right]\\) represents the expectation of the quadratic form, and simplifies to \\(\\mathbb{E}[\\pmb{q}_k^T\\pmb{q}_k]\\). Since the codes are randomly sampled from a Gaussian distribution, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.383, + 0.905, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {q} _ {k} \\right] = \\boldsymbol {I}, \\text {w h e r e} \\boldsymbol {q} \\sim \\mathcal {N} (0, 1), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.407, + 0.906, + 0.438 + ], + "angle": 0, + "content": "which ensures that all elements of \\(\\mathbf{W}\\) are updated. As training progresses, the latent basis \\(\\mathbf{W}\\) converges to:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.447, + 0.905, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow \\infty} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {q} _ {k} ^ {T} z _ {e} \\right] \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.48, + 0.636, + 0.494 + ], + "angle": 0, + "content": "Thus, in the limit:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.505, + 0.905, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\lim _ {t \\rightarrow \\infty} \\boldsymbol {q} _ {k} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} [ \\boldsymbol {q} _ {k} \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {e} ] = \\mathbb {E} [ \\boldsymbol {e} ] \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.538, + 0.905, + 0.566 + ], + "angle": 0, + "content": "At convergence, the product \\( q_{k}W \\) equals the nearest feature." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.577, + 0.696, + 0.593 + ], + "angle": 0, + "content": "4.4. Efficiency Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.902 + ], + "angle": 0, + "content": "SimVQ demonstrates greater efficiency than vanilla VQ due to its asymmetric training strategy, wherein the codebook \\( C \\) remains static and only the linear projection \\( W \\) is optimized. This approach results in a significant reduction in memory usage during the gradient backpropagation process. In vanilla VQ, the memory cost for codebook optimization is \\( O(Kd) \\), where \\( K \\) is the number of vectors in the codebook, and \\( d \\) is the dimension of each vector. In our experiments, \\( K = 65,536 \\) is much larger than \\( d = 128 \\). As the vocabulary size increases, the memory required for backpropagation grows proportionally, significantly impacting resource consumption. In contrast, SimVQ's memory cost for backpropagation is only \\( O(d^2) \\) because the codebook \\( C \\) is fixed, and only the linear layer \\( W \\) is updated. This results in a constant memory requirement in backpropagation, independent of the vocabulary size. The \\( d \\times d \\) scaling becomes particularly advantageous as \\( K \\) increases in practical applications. This structural design minimizes the computational overhead and improves training efficiency, especially when dealing with large vocabularies." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22972" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.855, + 0.264 + ], + "angle": 0, + "content": "
MethodLatent dimCodebook sizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN [9]12865,5361.4%3.740.1722.2070.6
VQGAN-EMA [19]12865,5364.5%3.230.1522.8972.3
VQGAN-FC [33]12865,5361.4%5.330.1821.4568.8
VQGAN-FC [33]865,536100.0%2.630.1323.7977.5
FSQ†[16]664,000100.0%2.800.1323.6375.8
LFQ [35]1665,536100.0%2.880.1323.6077.2
VQGAN-LC-CLIP+ [42]76865,536100.0%2.400.1323.9877.3
SimVQ (ours)12865,536100.0%2.240.1224.1578.4
SimVQ (ours)128262,144100.0%1.990.1124.6880.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.275, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Table 1. Reconstruction performance on ImageNet-1k with a resolution of \\( {128} \\times {128} \\) . All models are trained using images downsampled into \\( {16} \\times {16} \\) tokens. † Results are reproduced using the codebook size of \\( \\left\\lbrack {8,8,8,5,5,5}\\right\\rbrack \\) to approximately match \\( {65},{536}. + \\) Following VQGAN-LC,we extract CLIP features with the codebook frozen. The codebook utilization is calculated as the fraction of the codes that are activated at least once when encoding the validation set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.356, + 0.223, + 0.373 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.458 + ], + "angle": 0, + "content": "To assess the efficacy and versatility of the proposed SimVQ, we conduct experiments across both image and audio modalities. Subsequently, we analyze the learned linear layer to investigate the latent basis. The experimental configurations are listed in Appendix 8.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.25, + 0.484 + ], + "angle": 0, + "content": "5.1. Vision Modality" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.203, + 0.505 + ], + "angle": 0, + "content": "5.1.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Our baseline selection focuses on methods that enhance codebook utilization through architectural designs. We include VQGAN-FC [33], FSQ [16], LFQ [35] and VQGAN-LC-CLIP [42] as our primary baselines, as they represent the current state-of-the-art in improving reconstruction performance through codebook architecture innovations. Traditional VQ variants such as RVQ [15] and MoVQ [41] address fundamentally different technical challenges - they focus on training better discrete representations, rather than enhancing codebook utilization to improve reconstruction performance like our work. Another important research direction explores optimization-based solutions, where methods like stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] tackle codebook utilization through sophisticated training strategies. Given our focus on architectural innovations in codebook design, we evaluate SimVQ against methods that share this technical foundation for a meaningful assessment of our contributions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.303, + 0.822 + ], + "angle": 0, + "content": "5.1.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To rigorously evaluate the proposed SimVQ, we reproduce all the VQ models listed in Tab. 1 using the same architecture of VQGAN [9] with the quantization layer different only. Among the baselines, for VQGAN-FC [33], we follow the original setting to reduce the dimension of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.6 + ], + "angle": 0, + "content": "latent space to 8 followed by \\( l_{2} \\) normalization to improve codebook utilization. For FSQ [16], we adopt a codebook size of \\( [8,8,8,5,5,5] \\) as recommended, to approximately match the default codebook size. For VQGAN-LC [42], we follow them and leverage an external pre-trained CLIP model to extract features of the training dataset in advance for a well-defined latent space. All models are trained on the ImageNet [7] dataset for 50 epochs with a batch size of 256. Input images are processed at a resolution of \\( 128 \\times 128 \\) pixels and downsampled by a factor of 8, yielding a feature map of \\( 16 \\times 16 \\times 128 \\), where 128 is the dimension of the latent space. We set the default codebook size to a large number of \\( 2^{16} = 65536 \\) rather than the traditional number 8192 to highlight the representation collapse problem. Performance is evaluated using rFID, LPIPS, PSNR, and SSIM metrics on the ImageNet validation set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.61, + 0.655, + 0.624 + ], + "angle": 0, + "content": "5.1.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 presents the reconstruction performance of various VQ models on image data. We make three key observations: 1) Traditional VQGAN models utilize only a very small subset of the codebook, with a utilization rate of just \\(1.4\\%\\). Although VQGAN-EMA is proposed to improve codebook utilization, especially when the codebook size scales up to \\(65k\\), it still suffers from severe representation collapse. 2) Recently proposed methods, such as LFQ, FSQ, and VQGAN-FC, effectively improve codebook utilization to \\(100\\%\\). However, these methods require reducing the latent space to a very low dimension. For example, applying VQGAN-FC to the standard latent dimension of 128 results in severe representation collapse and degraded reconstruction performance. Additionally, these models face limitations in model capacity due to the low-dimensional latent space. While they achieve full codebook utilization, their reconstruction quality on rFID score lags significantly behind SimVQ. 3) VQGAN-LC-CLIP leverages an exter" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22973" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.858, + 0.297 + ], + "angle": 0, + "content": "
MethodBandwidthUtil↑UTMOS↑PESQ↑STOI↑V/UV F1↑
GT--4.06/3.48---
EnCodec [6]3.0kbps-2.31/2.092.05/2.050.90/0.880.92/0.89
Vocos [24]3.0kbps-3.53/3.062.40/2.190.92/0.900.94/0.91
SpeechTokenizer [39]3.0kbps-3.56/3.021.93/1.740.88/0.840.93/0.89
WavTokenizer [13]0.9kbps100/100%3.74/3.43*2.01/2.26*0.89/0.89*0.92/0.92*
SimVQ (ours)0.9kbps100.0/100.0%4.00/3.512.33/2.080.91/0.880.94/0.91
WavTokenizer [13]0.975kbps68/-%4.02*/-2.39*/-0.92*/-0.94*/-
WavTokenizer [13]1.05kbps27/-%4.00*/-2.36*/-0.81*/-0.94*/-
SimVQ (ours)0.975kbps99.4/99.4%4.03/3.522.42/2.150.92/0.880.94/0.92
SimVQ (ours)1.2kbps99.4/99.0%4.03/3.522.54/2.260.93/0.900.94/0.92
SimVQ (ours)1.35kbps95.6/94.7%4.03/3.532.61/2.310.93/0.900.95/0.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.306, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Table 2. Reconstruction performance on LibriTTS test-clean/test-other dataset. * WavTokenizer is trained with a window size of 3 seconds. The bandwidth of 0.9kbps, 0.975kbps, 1.2kbps, 1.35kbps means the codebook size of 4096, 8192, 65536, 262144 respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.358, + 0.481, + 0.447 + ], + "angle": 0, + "content": "
MethodCodebook SizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN-LC-CLIP† [42]50,00099.9%2.750.1323.858.4
VQGAN-LC-CLIP† [42]100,00099.9%2.620.1223.858.9
VQGAN-LC-CLIP† [42]200,00099.8%2.660.1223.959.2
SimVQ1,024100.0%3.670.1622.3470.8
SimVQ8,192100.0%2.980.1423.2374.7
SimVQ65,536100.0%2.240.1224.1578.4
SimVQ262,144100.0%1.990.1124.6880.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Table 3. Ablation study on the effect of various codebook sizes on ImageNet at a resolution of \\(128 \\times 128\\). \\(\\dagger\\) We directly copy the reported results of VQGAN-LC from the original paper on ImageNet \\(256 \\times 256\\) resolution." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.525, + 0.481, + 0.584 + ], + "angle": 0, + "content": "
InitializationTrainableUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
GaussianYes100.0%2.310.1224.0477.2
UniformNo100.0%2.310.1224.1578.4
GaussianNo100.0%2.240.1224.1578.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.594, + 0.462, + 0.609 + ], + "angle": 0, + "content": "Table 4. Ablation study of codebook optimization strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.756 + ], + "angle": 0, + "content": "nal pre-trained CLIP model to provide a well-defined latent space. However, VQGAN-LC relies on CLIP features pretrained on much larger datasets than ImageNet, which introduces generalization issues and a lower performance ceiling (degradation issue in Tab. 3). In contrast, SimVQ can be applied to a wide range of data types and achieves superior performance (rFID 2.40 vs. 2.24) without the limitations imposed by a pre-trained feature extraction model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.761, + 0.245, + 0.777 + ], + "angle": 0, + "content": "5.1.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "On the Codebook Size In Tab. 3, we explore the impact of different codebook sizes, ranging from \\(1k\\) to \\(262k\\), which is typically the level of LLM's vocabulary size. SimVQ consistently improves performance as the codebook size increases. For instance, the rFID score decreases to 1.99, and SSIM surpasses 80.0. In contrast, while VQGAN-LC-CLIP can keep high codebook utilization as increasing codebook size, it encounters performance degradation, with the rFID" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.907, + 0.393 + ], + "angle": 0, + "content": "score worsening from 2.62 to 2.66 when the codebook size is increased from 100,000 to 200,000." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.534 + ], + "angle": 0, + "content": "On the Codebook Optimization Strategy We investigate codebook initialization and the training of the linear layer in Tab. 4. Our findings are as follows: 1) The codebook is robust to different initialization strategies, yielding similar results with both Gaussian and uniform initialization. 2) When the codebook is updated during training, SimVQ continues to address the representation collapse issue, though with a slight degradation in performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.544, + 0.671, + 0.56 + ], + "angle": 0, + "content": "5.2. Audio Modality" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.822, + 0.581 + ], + "angle": 0, + "content": "5.2.1. Baselines and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.829 + ], + "angle": 0, + "content": "We use the LibriTTS dataset [36] for audio-based VQ model training. The baselines such as Encodec [6], VOCos [24], and SpeechTokenizer [39] are based on residual vector quantization method. Our SimVQ model adopts the same architecture as WavTokenizer [13] with the only modification being the replacement of their EMA codebook with our one linear layer reparameterization method. We train SimVQ on LibriTTS-580h for 50 epochs with a batch size of 64. Note that WavTokenizer is trained with a 3-second window size for optimal performance, we train SimVQ using a 1-second window to accelerate training. For objective evaluation of the reconstructed audio, we follow VOCos [24] and employ metrics such as UTMOS [23], PESQ [20], STOI, and the F1 score for voiced/unvoiced classification (V/UV F1). UTMOS is particularly valuable as it produces scores highly correlated with human evaluations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.655, + 0.85 + ], + "angle": 0, + "content": "5.2.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 presents the reconstruction performance of various VQ models on audio data. Baseline models using residual vector quantization perform significantly worse than" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22974" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.51, + 0.059, + 0.612, + 0.066 + ], + "angle": 0, + "content": "Rank of W (Codebook Size = 65536)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.482, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Figure 4. (a):(left) The rank of latent basis matrix \\(\\mathbf{W}\\) over training epochs. (b):(right) The Frobenius norm of latent basis matrix \\(\\mathbf{W}\\) over training epochs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.289, + 0.483, + 0.442 + ], + "angle": 0, + "content": "SimVQ, even when utilizing much larger bandwidths. Despite using the same architecture as WAVTokenizer, our model, which replaces the quantization layer with SimVQ, achieves superior performance with a 1-second window size and maintains nearly \\(100\\%\\) codebook utilization when scaling up to a size of 262,144. The consistent performance of the SimVQ model across both image and audio data demonstrates that SimVQ is a general method for addressing the representation collapse problem in VQ models and can be effectively applied across multiple modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.452, + 0.194, + 0.468 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.792 + ], + "angle": 0, + "content": "In Fig. 4(a), we plot the rank of the latent basis matrix over training epochs. Notably, SimVQ demonstrates the ability to adaptively adjust the rank of the latent space. Specifically, when the codebook size increases from \\(65k\\) to \\(262k\\), the rank of the latent basis matrix decreases more rapidly and converges to a lower value. This observation suggests that a larger codebook can effectively alleviate the pressure on the latent space dimensionality, allowing the model to learn to represent data more efficiently. Additionally, despite the rank decreasing to a lower-rank space, SimVQ maintains \\(100\\%\\) codebook utilization, highlighting its superiority over VQGAN-FC, which struggles when increasing the latent dimension from 8 to 128. We also calculate the Frobenius norm of the latent basis matrix, as shown in Fig. 4. The norm of a codebook size of \\(262k\\) is slightly larger than for \\(65k\\), indicating that a larger codebook can span a broader area in the linear space. For a comprehensive evaluation, we also provide the reconstruction loss curve on the ImageNet validation dataset in Appendix 8.2. The results consistently show that SimVQ achieves improved performance, further validating the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.303, + 0.819 + ], + "angle": 0, + "content": "5.4. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We visualize the distribution of encoder features and codebook embeddings in Fig. 5 and the frequency in Appendix 8.3. Compared to vanilla VQ models that most codebook vectors are not unused, SimVQ can update the whole codebook and distribute the features evenly in the whole space." + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.09, + 0.867, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.23, + 0.906, + 0.286 + ], + "angle": 0, + "content": "Figure 5. Visualization of the divergence between the encoder features and codebook embeddings on a random subset of ImageNet validation dataset. The left figure is vanilla VQ model and the right one is SimVQ." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.906, + 0.376 + ], + "angle": 0, + "content": "We qualitatively compare the reconstruction quality of both images and audio in Appendix 8.4. SimVQ can preserve more details with an enlarged codebook size, such as \"eyes\" and \"text\", which are challenging for vanilla VQ models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.628, + 0.405 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.627 + ], + "angle": 0, + "content": "VQ Performance and Generative Models Many generative models [21, 25, 28, 34] utilize VQ models as a \"tokenizer\" to obtain discrete tokens. While it is intuitive that VQ reconstruction quality should impact generation performance, recent studies [35, 43] have revealed that the relationship is more nuanced: improved VQ reconstruction metrics do not necessarily translate to better generative outcomes. This complex relationship suggests that evaluating VQ models primarily through downstream generation tasks may not provide the most insightful assessment of their fundamental properties. Therefore, we focus on addressing the critical issue of representation collapse in VQ models, aiming to advance our understanding of their core representation learning mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.642, + 0.634, + 0.658 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.819 + ], + "angle": 0, + "content": "In this paper, we explore the representation collapse problem in VQ models. We conduct a theoretical analysis of the optimization process in VQ models and propose a simple yet effective method, SimVQ, to address this issue. Our method addresses the representation collapse by jointly optimizing the latent space through a linear transformation with one linear layer. Experimental results demonstrate that SimVQ outperforms previous approaches on both image and audio datasets, highlighting its broad applicability across diverse machine learning tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.673, + 0.851 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.906, + 0.888 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China (Grant No.62276245)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22975" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[4] Zalan Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523-2533, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.402, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Maria Elisabeth Bechtle, Feryal Behbahani, Stephanie C.Y. Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando De Freitas, Satinder Singh, and Tim Roktaschel. Genie: Generative interactive environments. In Proceedings of the 41st International Conference on Machine Learning, pages 4603-4623. PMLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[6] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. Transactions on Machine Learning Research, 2023. Featured Certification, Reproducibility Certification. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Zhangyang Gao, Cheng Tan, Jue Wang, Yufei Huang, Lirong Wu, and Stan Z Li. Foldtoken: Learning protein language via vector quantization and beyond. arXiv preprint arXiv:2403.09673, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized net" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "works. In Proceedings of the 40th International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[13] Shengpeng Ji, Ziyue Jiang, Xize Cheng, Yifu Chen, Minghui Fang, Jialong Zuo, Qian Yang, Ruiqi Li, Zhang Zhang, Xiaoda Yang, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.904, + 0.291 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[15] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[16] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.422, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[18] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[19] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[20] A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra. Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and CODECs. In 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), pages 749-752 vol.2, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[22] Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. ArXiv, abs/2204.02152, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22976" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Hubert Siuzdak. Vocos: Closing the gap between timedomain and fourier-based neural vocoders for high-quality audio synthesis. In The Twelfth International Conference on Learning Representations, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[25] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[26] Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi, Toshiyuki Kumakura, and Yuki Mitsu-fuji. SQ-VAE: Variational Bayes on discrete representation with self-annealed stochastic quantization. In Proceedings of the 39th International Conference on Machine Learning, pages 20987-21012. PMLR, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[27] Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, and Ngai Wong. Scaling laws with vocabulary: Larger models deserve larger vocabularies. arXiv preprint arXiv:2407.13623, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[28] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[29] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Tung-Long Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quantized wasserstein auto-encoder. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[31] Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[32] Pan Xiao, Peijie Qiu, and Aristeidis Sotiras. Scvae: Sparse coding-based variational autoencoder. ArXiv, abs/2303.16666, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.638, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[33] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In International Conference on Learning Representations, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[34] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[36] Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. Libritts: A corpus derived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[37] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757-15773, Singapore, 2023. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[38] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18467-18476, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[39] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech language models. In The Twelfth International Conference on Learning Representations, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.429 + ], + "angle": 0, + "content": "[40] Chuanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798-22807, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[41] Chuanxia Zheng, Guoxian Song, Tat-Jen Cham, Jianfei Cai, Dinh Q. Phung, and Linjie Luo. High-quality pluralistic image completion via code shared vqgan. ArXiv, abs/2204.01931, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[42] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of \\(99\\%\\). ArXiv, abs/2406.11837, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.531, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[43] Yongxin Zhu, Bocheng Li, Hang Zhang, Xin Li, Linli Xu, and Lidong Bing. Stabilize the latent space for image autoregressive modeling: A unified perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[44] Yongxin Zhu, Dan Su, Liqiang He, Linli Xu, and Dong Yu. Generative pre-trained speech language model with efficient hierarchical transformer. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1764-1775, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22977" + } + ] +] \ No newline at end of file diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_origin.pdf b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1573ba594681d98fed0a29578a976d90230ac7b9 --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/c9ee5a18-482c-48be-ab03-3fe420197718_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7700d144094d893f48403454f31587fe3b4ad8d77cb4cc7e22fe674ab1187a9 +size 4610944 diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/full.md b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae097b5e388ae197608119dfbff03ae3e6137e4e --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/full.md @@ -0,0 +1,379 @@ +# Addressing Representation Collapse in Vector Quantized Models with One Linear Layer + +Yongxin Zhu $^{1,2}$ , Bocheng Li $^{1,2}$ , Yifei Xin $^{3}$ , Zhihua Xia $^{4}$ , Linli Xu $^{1,2*}$ + +1University of Science and Technology of China + +$^{2}$ State Key Laboratory of Cognitive Intelligence, $^{3}$ Peking University, $^{4}$ Jinan University + +zyx2016@mail.ustc.edu.cn,bcli@mail.ustc.edu.cn + +xinyifei@stu.pku.edu.cn, xiazhihua@jnu.edu.cn, linlixu@ustc.edu.cn + +# Abstract + +Vector Quantization (VQ) is essential for discretizing continuous representations in unsupervised learning but suffers from representation collapse, causing low codebook utilization and limiting scalability. Existing solutions often rely on complex optimizations or reduce latent dimensionality, which compromises model capacity and fails to fully solve the problem. We identify the root cause as disjoint codebook optimization, where only a few code vectors are updated via gradient descent. To fix this, we propose SimpleVQ, which reparameterizes code vectors through a learnable linear transformation layer over a latent basis, optimizing the entire linear space rather than nearest individual code vectors. Although the multiplication of two linear matrices is equivalent to applying a single linear layer, this simple approach effectively prevents collapse. Extensive experiments on image and audio tasks demonstrate that SimVQ improves codebook usage, is easy to implement, and generalizes well across modalities and architectures. + +# 1. Introduction + +In recent years, vector quantization (VQ) [19, 29] has emerged as a foundational technique in unsupervised representation learning [2, 5] and latent generative models [4, 21, 31, 33, 34, 44]. By converting continuous representations into discrete codes, VQ models can effectively identify the inherent structure of data and enable various discrete modeling methods on continuous data, from high-quality image generation [9] to audio synthesis [6]. The recent success of Large Language Models (LLMs) [1] has highlighted the effectiveness of next-token prediction as a powerful and versatile training objective. Consequently, VQ models are taken as the direct method to transform data from various modalities [25, 28, 37] or scientific domains [10] to discrete + +sequences for next token prediction training. However, attempts to integrate VQ models as multimodal tokenizers to leverage the scaling laws of LLMs face significant challenges because of the difficulty of expanding the codebook. For example, the Chameleon model [28] constrains its codebook size to $8k$ , which is significantly trailing behind the vocabulary size of LLMs (e.g., LLaMA3's vocabulary size is $128k$ [8]). + +There is a broad agreement that increasing vocabulary size can consistently improve the performance of LLMs [27]. However, recent studies [42] indicate that traditional VQ models often fail to utilize the additional parameters introduced by codebook expansion, leaving most codes inactive during training. The contradiction between codebook expansion and low codebook utilization in VQ models is known as the representation collapse problem [22], where increasing the codebook size fails to improve the performance. To address these discrepancies, we conduct a theoretical analysis of the optimization procedure of VQ models and identify that the disjoint optimization of the codebook is the root cause of representation collapse. As illustrated in Fig. 1(a), the core mechanism of VQ models involves a nearest-neighbor replacement strategy, where the encoder's output features are replaced by the nearest vector in the codebook to serve as input to the decoder. The indices of the nearest vector are taken as the discrete representation of the data. This nearest-selection operator results in only a subset of codes being updated through gradient descent, while the remaining codes remain unchanged. + +Some methods mitigate representation collapse by hand-designing complex optimization strategies, such as stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] through sophisticated training strategies. Recently, some approaches [16, 33, 35] propose to reduce the dimension of the latent space to a very small scale (e.g., 8 v.s. 128) to alleviate the curse of dimensionality, thereby improving the overlap between the encoder's fea + +![](images/0e32b10883c72bc623c65cca52363719c50009416e5845bc98d83fcb0158b8bd.jpg) +Figure 1. Comparison of Vanilla VQ and SimVQ. (a): (left) Disjoint optimization in Vanilla VQ. Only the nearest codes are updated, resulting in a high percentage of "dead" codes that are not updated. (b): (right) Joint optimization in SimVQ. The entire codebook is updated with a latent basis, ensuring all codes remain active. + +![](images/e9eaa319bcd3756109ed568cab93cd588d8097f25b904638324a06aa7447b3d0.jpg) + +tures and the codebook. However, while these methods enhance codebook utilization, they do so at the cost of model capacity, leading to worse performance compared to vanilla VQ models when the codebook size is small or representation collapse is not severe. Another approach, VQGAN-LC [42], initializes the codebook with features extracted from the pre-trained CLIP model [17] to create a well-structured latent space that better matches the distribution of the encoder output. Nevertheless, the latent space defined by an external pre-trained model limits the model's ability to generalize to diverse datasets and reaches a performance plateau as the codebook size increases. These limitations highlight the need for a more effective method to improve codebook utilization without compromising model capacity or relying on external models. + +We critically assess prevalent methodologies and reveal that optimizing the latent space rather than individual code vectors is key to preventing representation collapse. Building on this insight, we introduce a simple yet effective method, termed SimVQ, to directly update the latent space spanned by the codebook by linear transforming the code vectors via a learnable latent basis. Specifically, the vectors in the codebook are reparameterized as a linear combination of the basis in the learnable linear layer $W$ : + +$$ +\boldsymbol {C} \in \mathbb {R} ^ {K \times d} \Rightarrow \boldsymbol {C W} \text {w i t h} \boldsymbol {W} \in \mathbb {R} ^ {d \times d}, \tag {1} +$$ + +where $K$ denotes the codebook size and $d$ represents the dimension of latent space. This reparameterization with linear transformation disentangles the optimization of the codebook into two components: the coefficient matrix $\mathbf{C}$ and the basis of linear space $\mathbf{W}$ respectively. As illustrated in Fig. 1(b), by optimizing the basis matrix $\mathbf{W}$ , the latent space spanned by $\mathbf{CW}$ is rotated and stretched to match encoder's output feature. The entire codebook is updated jointly to prevent the representation collapse problem. The simplicity of the proposed method makes it highly portable and easily adaptable for improving VQ-based models across a wide range of domains, requiring only one linear layer. + +In summary, our contributions to vector quantized models are as follows: + +- We theoretically analyze the representation collapse problem in VQ models and reveal that optimizing the latent space spanned by the codebook, rather than individual code vectors, is crucial to addressing this issue. +- We propose a novel method, SimVQ, which reparameterizes the codebook vectors in VQ models via a linear transformation with a learnable latent basis. This simple yet effective approach is highly adaptable and easy to implement, making it broadly applicable across various machine learning contexts. +- We conduct an extensive evaluation of SimVQ across diverse modalities, including image and audio with different model architectures. The results show that SimVQ not only effectively addresses the representation collapse problem by achieving near-complete codebook utilization regardless of the codebook size, but also establishes new state-of-the-art performance. Furthermore, when scaling up the codebook size, SimVQ consistently delivers improved results. + +# 2. Related Work + +VQ-VAE [29] is the pioneering work to encode data into discrete representations. Building on these developments, VQGAN [9] combines VQ-VAE with adversarial networks to improve the perceptual quality of generated samples and establish a fundamental quantization protocol in latent generative models [21, 28, 34]. Many traditional VQ works focus on training a better discrete representation rather than codebook utilization to improve reconstruction performance. For example, RVQ [15] and MoVQ [41] enhance the reconstruction details with multichannel quantization. However, these methods suffer from a critical issue of representation collapse, as they struggle to scale the codebook size beyond $10\mathrm{k}$ entries, limiting their scalability. In response to this challenge, several approaches have been pro + +posed recently. DALLE [18] employs the gumbel-softmax trick [12] and stochastic sampling strategies to activate most codes during training. However, during inference, only a small subset of codes is utilized for quantization [38]. Some methods design complex optimization strategies to improve codebook utilization, such as stochastic quantization [26], distribution penalty [30, 32] and codebook reset [40]. Huh et al. [11] proposes rescaling the vectors in the codebook during training to match the distributions in the latent space. VQGAN-FC [33] introduces a method to map latent vectors into a lower-dimensional space followed by $l_{2}$ normalization to alleviate representation collapse. FSQ [16] extends this idea by projecting representations into a reduced-dimensional space, where they are quantized into a small set of fixed values. LFQ [35], a variant of FSQ, uses binary values for quantized representations, thereby simplifying the encoding process. While these methods improve the codebook utilization, they do so at the cost of model capacity by significantly reducing the dimensionality of latent space (often to as low as 8), leading to worse performance compared to vanilla VQ models when the codebook size is small and representation collapse is not severe. Additionally, VQGAN-LC [42] proposes to initialize the codebook using features extracted from the pre-trained CLIP model to avoid representation collapse. However, the reliance on the pre-trained model limits the VQ model's ability to generalize to diverse datasets and results in a performance plateau as the codebook size increases. In contrast, our method, SimVQ, effectively addresses the representation collapse problem with a simple linear layer, without sacrificing model capacity or relying on external pre-trained models. + +# 3. Representation Collapse in VQ Models + +# 3.1. Preliminaries + +A vector quantized model is typically a reconstructive encoder-decoder architecture that includes a vector quantization layer to convert continuous representations into discrete codes. For simplicity, we represent an image with a single random variable $x$ . Formally, the encoder $f_{\theta}$ maps the input image into a latent space, producing a continuous representation $z_{e} = f_{\theta}(x) \in \mathbb{R}^{d}$ . This representation is then quantized using a learnable codebook $C = [q_{1}, \ldots, q_{K}] \in \mathbb{R}^{K \times d}$ , where $q_{i}$ is a codebook vector. We define $\delta_{k} \in \{0, 1\}^{1 \times K}$ as a characteristic (one-hot) vector where only the $k$ -th element is 1, such that $q_{k} = \delta_{k} C \in \mathbb{R}^{1 \times d}$ . The quantization layer selects the nearest codebook vector $q_{k}$ by minimizing the Euclidean distance between $z_{e}$ and the codebook entries [29]: + +$$ +k = \arg \min _ {j} \| z _ {e} - q _ {j} \| _ {2} ^ {2} = \arg \min _ {j} \| z _ {e} - \delta_ {j} C \| _ {2} ^ {2}. \tag {2} +$$ + +The selected vector $q_{k}$ is then passed to the decoder $g_{\phi}$ to reconstruct the input image. + +To enable gradient propagation through the non-differentiable characteristic vector $\delta_{k}$ , the straight-through estimator (STE) [3] is applied. During the backward pass, the gradient of $z_{q} = \delta_{k}C$ is copied to $z_{e}$ as follows, + +$$ +z _ {q} = \mathrm {s g} (\delta_ {k} \pmb {C} - z _ {e}) + z _ {e}, \Rightarrow \frac {\partial z _ {q}}{\partial z _ {e}} = 1 (3) +$$ + +where $\mathfrak{sg}$ is the stop gradient operator, ensuring the gradient for $\delta_k C$ is discarded during the backward pass. + +The learning objective is the combination of a reconstruction loss and commitment loss that ensures that the encoder commits to an embedding and the encoder's output does not drift: + +$$ +\mathcal {L} = \log p (x | z _ {q}) + \| \operatorname {s g} \left(\delta_ {k} \boldsymbol {C}\right) - z _ {e} \| _ {2} ^ {2} + \beta \| \delta_ {k} \boldsymbol {C} - \operatorname {s g} \left(z _ {e}\right) \| _ {2} ^ {2}, \tag {4} +$$ + +where $\log p(x|z_q)$ is typically the mean squared error (MSE) loss $\| x - g_{\phi}(z_q)\| _2^2$ for image and audio data. + +# 3.2. Disjoint Optimization of Codebook + +In VQ models, only the nearest code is selected and updated via gradient descent. Ideally, all codebook entries should be updated and utilized for decoding. However, experimental evidence shows that only a small fraction of the codebook gets updated and utilized, leading to what is known as the representation collapse problem [22]. To investigate the root cause of this issue, we provide a theoretical analysis of the optimization dynamics in VQ models. + +Due to the use of the straight-through estimator (STE) for gradient propagation, the codebook $C$ can only be updated through the gradient of the commitment loss, which is defined as: + +$$ +\mathcal {L} _ {\text {c o m m i t}} (\boldsymbol {C}) = \left\| z _ {e} - \delta_ {k} \boldsymbol {C} \right\| _ {2} ^ {2}. \tag {5} +$$ + +The codebook $\pmb{C}$ is updated according to the following equation, where $\eta$ is the learning rate: + +$$ +\begin{array}{l} \boldsymbol {C} ^ {(t + 1)} = \boldsymbol {C} ^ {(t)} + \eta \mathbb {E} _ {z _ {e}} \left[ \frac {\partial \mathcal {L} _ {\text {c o m m i t}} (\boldsymbol {C} ^ {(t)})}{\partial \boldsymbol {C} ^ {(t)}} \right] (6) \\ = \boldsymbol {C} ^ {(t)} - \eta \mathbb {E} _ {z _ {e}} \left[ \delta_ {k} ^ {T} \delta_ {k} \boldsymbol {C} ^ {(t)} \right] + \eta \mathbb {E} _ {z _ {e}} \left[ \delta_ {k} ^ {T} z _ {e} \right] (7) \\ \end{array} +$$ + +where $\delta_k^T\delta_k$ is the Kronecker delta matrix, defined as: + +$$ +\left(\delta_ {k} ^ {T} \delta_ {k}\right) _ {i j} = \left\{ \begin{array}{l l} 1 & \text {i f} i = j = k, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {8} +$$ + +All vectors in $C$ will be updated and utilized if and only if the expectation $\mathbb{E}_{z_e}\left[\delta_k^T\delta_k\right]$ converges to the identity matrix. Unlike variational autoencoders (VAEs) [14], which enforce a Gaussian distribution on the latent space via a + +KL-divergence penalty, VQ models optimize $z_{e}$ towards the selected codebook vectors $\mathbb{E}_{z_e}\left[\delta_k^T\delta_kC\right]$ . At the same time, the selected codebook vectors are optimized towards the distribution of $z_{e}$ , resulting in the same selected subset of vectors moving closer to $z_{e}$ , somewhat akin to a cocoon effect. However, this disjoint optimization of the codebook leads to part of the codebook, specifically $(I - \mathbb{E}_{z_e}\left[\delta_k^T\delta_k\right])C$ , remaining un-updated and underutilized once the optimization process begins. This phenomenon occurs because the optimization focuses only on a subset of codebook vectors, leaving other vectors stagnant. + +This analysis reveals the fundamental cause of representation collapse in VQ models: the disjoint optimization process that updates only a subset of codebook vectors. This insight forms the basis for our proposed solution, SimVQ, which aims to address this issue by optimizing the entire latent space spanned by the codebook, rather than individual code vectors. + +# 4. Addressing Collapse with Latent Linear Transformation + +# 4.1. Reparameterize Codes with Latent Basis + +Let $\pmb{W} = \{\pmb{w}_1, \dots, \pmb{w}_n\}$ be a basis of a linear space. Any vector $\pmb{v}$ in the space can be uniquely expressed as a linear combination of the basis vectors with coefficients $c_1, \dots, c_n \in \mathbb{R}$ : + +$$ +\boldsymbol {v} = c _ {1} \boldsymbol {w} _ {1} + \dots + c _ {n} \boldsymbol {w} _ {n} = \boldsymbol {c} \boldsymbol {W}. \tag {9} +$$ + +Given the equivalence between $\pmb{v}$ and $c\pmb{W}$ in the linear space, we can reparameterize each vector in the codebook of VQ models with a new basis matrix $\pmb{W} \in \mathbb{R}^{d \times d}$ . Specifically, the codebook $\pmb{C} = \{c_1, \dots, c_K\}$ can be reparameterized as: + +$$ +\{\hat {\boldsymbol {c}} _ {1} \boldsymbol {W}, \dots , \hat {\boldsymbol {c}} _ {N} \boldsymbol {W} \} = \hat {\boldsymbol {C}} \boldsymbol {W} \in \mathbb {R} ^ {K \times d}. \tag {10} +$$ + +This reparameterization introduces two components: the basis matrix $\mathbf{W}$ and the coefficient matrix $\hat{\mathbf{C}}$ . In the following, we will discuss the optimization of both the basis matrix $\mathbf{W}$ and the coefficient matrix $\hat{\mathbf{C}}$ . For simplicity, we slightly abuse $\mathbf{C}$ and $\hat{\mathbf{C}}$ below. + +# 4.2. Asymmetric Optimization Dynamics + +While it is commonly accepted that multiplying two linear matrices is equivalent to a single linear layer, we argue that the disjoint optimization problem of the codebook in VQ models can be addressed by linear transformation. In vanilla VQ models, only the codebook $C$ is responsible for minimizing commitment loss, leading to the disjoint optimization problem where only the selected codes will be updated. + +![](images/acaa8fcbfd39e70edaae65d82bff2a10fc759155be00813d480232790e21b499.jpg) +Figure 2. (a): (left) The optimization trajectory of the objective $\| \pmb{x} - \pmb{q}\|_2^2$ , which is the same as vanilla VQ. Only a small fraction of points are updated while others remain inactive. (b): (right) The optimization trajectory of the objective $\| \pmb{x} - \pmb{q}\pmb{w}\|_2^2$ with $\pmb{q}$ frozen, which is the same as SimVQ. All the points are updated towards targets $x$ . + +![](images/26aa03cbd2994a9c5092415892590343fd3a37c6be1fc0685e8a96b9a619c387.jpg) + +![](images/ab17f67fcb4d5ed7446cec316c3649888bf536cb9d657eeef1534741bb777ca8.jpg) +Figure 3. (a): (left) The optimization trajectory of the optimization objective: $\| \pmb{x} - \pmb{q}\pmb{w}\|_2^2$ with both $\pmb{q}$ and $\pmb{w}$ unfrozen. (b): (right) The Frobenius norm of the projection matrix $\pmb{w}$ and loss curves. The loss quickly converges to 0 with $\pmb{w}$ almost unchanged. + +![](images/cbaae70c1eeb78bd94e2902795523e31e2fa602a5508e663652b2ab6c1c7321f.jpg) + +In contrast, when the codebook is reparameterized as $CW$ , both the basis $W$ and the coefficient matrix $C$ contribute to minimizing the commitment loss. The gradients $\frac{\partial\mathcal{L}}{\partial W}$ and $\frac{\partial\mathcal{L}}{\partial C}$ can simultaneously reduce the loss. As a result, the optimization of the reparameterized codebook can be divided into three scenarios: + +- Updating $C$ with $W$ frozen: Only the selected codes adapt to the latent distribution of $z_{e}$ , as depicted on Fig. 1(a). The vanilla VQ is a special case with $W = I$ . +- Updating $W$ with $C$ frozen: The entire codebook $CW$ adjusts to the latent distribution of $z_{e}$ . The basis matrix $W$ rotates and stretches the space as shown in Fig. 1(b). +- Updating both $C$ and $W$ : The selected subset of codes moves towards $z_{e}$ while the space spanned by $W$ undergoes simultaneous rotation and stretching. + +To highlight the difference in optimization between $C$ and $CW$ , we conduct a toy experiment in a two-dimensional setting and visualize the optimization process in Fig. 2 and Fig. 3. + +# 4.2.1. Toy Examples + +We randomly sample two target points $x$ from Gaussian distribution as follows: + +$$ +\boldsymbol {x} _ {1} \sim \mathcal {N} (\binom {2} {2}, \binom {1 0} {0 1}), \quad \boldsymbol {x} _ {2} \sim \mathcal {N} (\binom {- 2} {- 2}, \binom {1 0} {0 1}). \tag {11} +$$ + +Then we initialize 10 learnable vectors $\mathbf{q}$ from a Gaussian + +# Algorithm 1 Training Procedure for SimVQ + +Input: Encoder $f_{\theta}$ , Decoder $g_{\phi}$ , Codebook $C \in \mathbb{R}^{K \times d}$ , Linear projector matrix $\boldsymbol{W}_{\psi}$ , commitment weight $\beta$ . + +Output: Model parameters $\theta, \phi, \psi$ and Codebook $C$ . Initialize Codebook with Gaussian distribution and freeze the parameter of Codebook $C$ ; + +# repeat + +Draw $x\sim p_{data}(\pmb {x})$ + +$$ +z _ {e} = f _ {\theta} (x); +$$ + +$/^{*}$ Replace $q_{j}$ in vanilla VQ with proposed $q_{j}\pmb{W}_{\psi}$ . + +Nearest code search: $k = \arg \min_{j} \| z_{e} - q_{j}W_{\psi} \|_{2}^{2}$ , where $q_{j} \in C$ ; + +Straight Through Estimation: $z_{q} = \mathrm{sg}(q_{k}\mathbf{W}_{\psi} - z_{e}) + z_{e}$ + +$$ +\hat {x} = g _ {\phi} (z _ {q}); +$$ + +Minimize $\mathcal{L}(\theta ,\phi ,\psi) = \mathrm{MSE}(x,\hat{x})$ + $\beta \| z_{e}$ - $\mathrm{sg}(q_kW_\psi)\| _2^2 +\| \mathrm{sg}(z_e) - q_kW_\psi \| _2^2;$ + +until converged + +distribution: + +$$ +\left\{\boldsymbol {q} _ {i} \right\} _ {i = 1} ^ {1 0} \sim \mathcal {N} \left(\binom {0} {0}, \binom {1} {0}, \binom {0} {1}\right), \tag {12} +$$ + +During training with gradient descent, we introduce perturbation noise $\mathcal{N}(0,0.01)$ to the targets. In Fig. 2(a), the optimization objective is similar to vanilla VQ: $\| \pmb {x} - \pmb {q}\| _2^2$ . Only the nearest points $q_{4}$ and $q_{10}$ are updated. In contrast, in Fig. 2(b), the optimization objective $\| \pmb {x} - \pmb {q}\pmb {w}\| _2^2$ is similar to SimVQ with the points reparameterized by a learnable latent basis $\pmb{w}$ and $\pmb{q}$ frozen, resulting in the entire codebook $\{\pmb {q}\}_{i = 1}^{10}$ being jointly updated. + +Remark 1. The simultaneous optimization of the latent basis $\mathbf{w}$ and the coefficient matrix $\mathbf{q}$ may lead to the collapse. + +We provide an example in Fig. 3(a) where the optimization objective is $\| \pmb{x} - \pmb{q}\pmb{w} \|_2^2$ with $\pmb{q}$ unfrozen this time. In the training process, only the nearest point $\pmb{q}_1$ and point $\pmb{q}_{10}$ move towards the target point, while other points remain almost unchanged. We also visualize the loss curve in Fig. 3(b). The optimization objective with both $\pmb{q}$ and $\pmb{w}$ unfrozen converges quickly, where the norm of basis $\pmb{w}$ is much smaller than the objective with $\pmb{q}$ frozen. This indicates that the disjoint optimization of the codebook persists: $\pmb{q}$ can directly commit to the loss and dominate the optimization process, with $\pmb{w}$ being ignored, leading to the collapse quickly. + +4.3. Joint Optimization of the Codebook + +We propose SimVQ by simply using a learnable basis $W \in \mathbb{R}^{d \times d}$ to reparameterize the codebook such that the codebook is transformed into $CW$ . The pseudo-code for this + +approach is provided in Algorithm 1. During training, we optimize only the latent basis matrix $\mathbf{W}$ , while keeping the coefficient matrix $\mathbf{C}$ frozen. The commitment loss for SimVQ is defined as: + +$$ +\mathcal {L} _ {\text {c o m m i t}} \left(z _ {e}, q _ {k}\right) = \left\| z _ {e} - \delta_ {k} \boldsymbol {C} \boldsymbol {W} \right\| _ {2} ^ {2}. \tag {13} +$$ + +The vanilla VQ model is a special case of SimVQ, where the linear basis matrix $\mathbf{W}$ is fixed as the identity matrix $\mathbf{I}$ . The update for $\mathbf{W}$ with learning rate $\eta$ is: + +$$ +\boldsymbol {W} ^ {(t + 1)} = \boldsymbol {W} ^ {(t)} - \eta \frac {\partial \mathcal {L} _ {\text {c o m m i t}} \left(z _ {e} , \boldsymbol {q} _ {k}\right)}{\partial \boldsymbol {W} ^ {(t)}} \tag {14} +$$ + +$$ += \left(\boldsymbol {I} - \eta \mathbb {E} _ {z _ {e}} \left[ \boldsymbol {C} ^ {T} \delta_ {k} ^ {T} \delta_ {k} \boldsymbol {C} \right]\right) \boldsymbol {W} ^ {(t)} + \eta \mathbb {E} _ {z _ {e}} \left[ \boldsymbol {C} ^ {T} \delta_ {k} ^ {T} z _ {e} \right]. \tag {15} +$$ + +The term $\mathbb{E}\left[\boldsymbol{C}^T\delta_k^T\delta_k\boldsymbol{C}\right]$ represents the expectation of the quadratic form, and simplifies to $\mathbb{E}[\pmb{q}_k^T\pmb{q}_k]$ . Since the codes are randomly sampled from a Gaussian distribution, we have: + +$$ +\mathbb {E} \left[ \boldsymbol {q} _ {k} ^ {T} \boldsymbol {q} _ {k} \right] = \boldsymbol {I}, \text {w h e r e} \boldsymbol {q} \sim \mathcal {N} (0, 1), \tag {16} +$$ + +which ensures that all elements of $\mathbf{W}$ are updated. As training progresses, the latent basis $\mathbf{W}$ converges to: + +$$ +\lim _ {t \rightarrow \infty} \boldsymbol {W} ^ {(t)} = \mathbb {E} _ {z _ {e}} \left[ \boldsymbol {q} _ {k} ^ {T} z _ {e} \right] \tag {17} +$$ + +Thus, in the limit: + +$$ +\lim _ {t \rightarrow \infty} \boldsymbol {q} _ {k} \boldsymbol {W} ^ {(t)} = \mathbb {E} [ \boldsymbol {q} _ {k} \boldsymbol {q} _ {k} ^ {T} \boldsymbol {e} ] = \mathbb {E} [ \boldsymbol {e} ] \tag {18} +$$ + +At convergence, the product $q_{k}W$ equals the nearest feature. + +4.4. Efficiency Analysis + +SimVQ demonstrates greater efficiency than vanilla VQ due to its asymmetric training strategy, wherein the codebook $C$ remains static and only the linear projection $W$ is optimized. This approach results in a significant reduction in memory usage during the gradient backpropagation process. In vanilla VQ, the memory cost for codebook optimization is $O(Kd)$ , where $K$ is the number of vectors in the codebook, and $d$ is the dimension of each vector. In our experiments, $K = 65,536$ is much larger than $d = 128$ . As the vocabulary size increases, the memory required for backpropagation grows proportionally, significantly impacting resource consumption. In contrast, SimVQ's memory cost for backpropagation is only $O(d^2)$ because the codebook $C$ is fixed, and only the linear layer $W$ is updated. This results in a constant memory requirement in backpropagation, independent of the vocabulary size. The $d \times d$ scaling becomes particularly advantageous as $K$ increases in practical applications. This structural design minimizes the computational overhead and improves training efficiency, especially when dealing with large vocabularies. + +
MethodLatent dimCodebook sizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN [9]12865,5361.4%3.740.1722.2070.6
VQGAN-EMA [19]12865,5364.5%3.230.1522.8972.3
VQGAN-FC [33]12865,5361.4%5.330.1821.4568.8
VQGAN-FC [33]865,536100.0%2.630.1323.7977.5
FSQ†[16]664,000100.0%2.800.1323.6375.8
LFQ [35]1665,536100.0%2.880.1323.6077.2
VQGAN-LC-CLIP+ [42]76865,536100.0%2.400.1323.9877.3
SimVQ (ours)12865,536100.0%2.240.1224.1578.4
SimVQ (ours)128262,144100.0%1.990.1124.6880.3
+ +Table 1. Reconstruction performance on ImageNet-1k with a resolution of ${128} \times {128}$ . All models are trained using images downsampled into ${16} \times {16}$ tokens. † Results are reproduced using the codebook size of $\left\lbrack {8,8,8,5,5,5}\right\rbrack$ to approximately match ${65},{536}. +$ Following VQGAN-LC,we extract CLIP features with the codebook frozen. The codebook utilization is calculated as the fraction of the codes that are activated at least once when encoding the validation set. + +# 5. Experiments + +To assess the efficacy and versatility of the proposed SimVQ, we conduct experiments across both image and audio modalities. Subsequently, we analyze the learned linear layer to investigate the latent basis. The experimental configurations are listed in Appendix 8.1. + +# 5.1. Vision Modality + +# 5.1.1. Baselines + +Our baseline selection focuses on methods that enhance codebook utilization through architectural designs. We include VQGAN-FC [33], FSQ [16], LFQ [35] and VQGAN-LC-CLIP [42] as our primary baselines, as they represent the current state-of-the-art in improving reconstruction performance through codebook architecture innovations. Traditional VQ variants such as RVQ [15] and MoVQ [41] address fundamentally different technical challenges - they focus on training better discrete representations, rather than enhancing codebook utilization to improve reconstruction performance like our work. Another important research direction explores optimization-based solutions, where methods like stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] tackle codebook utilization through sophisticated training strategies. Given our focus on architectural innovations in codebook design, we evaluate SimVQ against methods that share this technical foundation for a meaningful assessment of our contributions. + +# 5.1.2. Implementation Details + +To rigorously evaluate the proposed SimVQ, we reproduce all the VQ models listed in Tab. 1 using the same architecture of VQGAN [9] with the quantization layer different only. Among the baselines, for VQGAN-FC [33], we follow the original setting to reduce the dimension of the + +latent space to 8 followed by $l_{2}$ normalization to improve codebook utilization. For FSQ [16], we adopt a codebook size of $[8,8,8,5,5,5]$ as recommended, to approximately match the default codebook size. For VQGAN-LC [42], we follow them and leverage an external pre-trained CLIP model to extract features of the training dataset in advance for a well-defined latent space. All models are trained on the ImageNet [7] dataset for 50 epochs with a batch size of 256. Input images are processed at a resolution of $128 \times 128$ pixels and downsampled by a factor of 8, yielding a feature map of $16 \times 16 \times 128$ , where 128 is the dimension of the latent space. We set the default codebook size to a large number of $2^{16} = 65536$ rather than the traditional number 8192 to highlight the representation collapse problem. Performance is evaluated using rFID, LPIPS, PSNR, and SSIM metrics on the ImageNet validation set. + +# 5.1.3. Main Results + +Tab. 1 presents the reconstruction performance of various VQ models on image data. We make three key observations: 1) Traditional VQGAN models utilize only a very small subset of the codebook, with a utilization rate of just $1.4\%$ . Although VQGAN-EMA is proposed to improve codebook utilization, especially when the codebook size scales up to $65k$ , it still suffers from severe representation collapse. 2) Recently proposed methods, such as LFQ, FSQ, and VQGAN-FC, effectively improve codebook utilization to $100\%$ . However, these methods require reducing the latent space to a very low dimension. For example, applying VQGAN-FC to the standard latent dimension of 128 results in severe representation collapse and degraded reconstruction performance. Additionally, these models face limitations in model capacity due to the low-dimensional latent space. While they achieve full codebook utilization, their reconstruction quality on rFID score lags significantly behind SimVQ. 3) VQGAN-LC-CLIP leverages an exter + +
MethodBandwidthUtil↑UTMOS↑PESQ↑STOI↑V/UV F1↑
GT--4.06/3.48---
EnCodec [6]3.0kbps-2.31/2.092.05/2.050.90/0.880.92/0.89
Vocos [24]3.0kbps-3.53/3.062.40/2.190.92/0.900.94/0.91
SpeechTokenizer [39]3.0kbps-3.56/3.021.93/1.740.88/0.840.93/0.89
WavTokenizer [13]0.9kbps100/100%3.74/3.43*2.01/2.26*0.89/0.89*0.92/0.92*
SimVQ (ours)0.9kbps100.0/100.0%4.00/3.512.33/2.080.91/0.880.94/0.91
WavTokenizer [13]0.975kbps68/-%4.02*/-2.39*/-0.92*/-0.94*/-
WavTokenizer [13]1.05kbps27/-%4.00*/-2.36*/-0.81*/-0.94*/-
SimVQ (ours)0.975kbps99.4/99.4%4.03/3.522.42/2.150.92/0.880.94/0.92
SimVQ (ours)1.2kbps99.4/99.0%4.03/3.522.54/2.260.93/0.900.94/0.92
SimVQ (ours)1.35kbps95.6/94.7%4.03/3.532.61/2.310.93/0.900.95/0.93
+ +Table 2. Reconstruction performance on LibriTTS test-clean/test-other dataset. * WavTokenizer is trained with a window size of 3 seconds. The bandwidth of 0.9kbps, 0.975kbps, 1.2kbps, 1.35kbps means the codebook size of 4096, 8192, 65536, 262144 respectively. + +
MethodCodebook SizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN-LC-CLIP† [42]50,00099.9%2.750.1323.858.4
VQGAN-LC-CLIP† [42]100,00099.9%2.620.1223.858.9
VQGAN-LC-CLIP† [42]200,00099.8%2.660.1223.959.2
SimVQ1,024100.0%3.670.1622.3470.8
SimVQ8,192100.0%2.980.1423.2374.7
SimVQ65,536100.0%2.240.1224.1578.4
SimVQ262,144100.0%1.990.1124.6880.3
+ +Table 3. Ablation study on the effect of various codebook sizes on ImageNet at a resolution of $128 \times 128$ . $\dagger$ We directly copy the reported results of VQGAN-LC from the original paper on ImageNet $256 \times 256$ resolution. + +
InitializationTrainableUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
GaussianYes100.0%2.310.1224.0477.2
UniformNo100.0%2.310.1224.1578.4
GaussianNo100.0%2.240.1224.1578.4
+ +Table 4. Ablation study of codebook optimization strategy. + +nal pre-trained CLIP model to provide a well-defined latent space. However, VQGAN-LC relies on CLIP features pretrained on much larger datasets than ImageNet, which introduces generalization issues and a lower performance ceiling (degradation issue in Tab. 3). In contrast, SimVQ can be applied to a wide range of data types and achieves superior performance (rFID 2.40 vs. 2.24) without the limitations imposed by a pre-trained feature extraction model. + +# 5.1.4. Ablation Study + +On the Codebook Size In Tab. 3, we explore the impact of different codebook sizes, ranging from $1k$ to $262k$ , which is typically the level of LLM's vocabulary size. SimVQ consistently improves performance as the codebook size increases. For instance, the rFID score decreases to 1.99, and SSIM surpasses 80.0. In contrast, while VQGAN-LC-CLIP can keep high codebook utilization as increasing codebook size, it encounters performance degradation, with the rFID + +score worsening from 2.62 to 2.66 when the codebook size is increased from 100,000 to 200,000. + +On the Codebook Optimization Strategy We investigate codebook initialization and the training of the linear layer in Tab. 4. Our findings are as follows: 1) The codebook is robust to different initialization strategies, yielding similar results with both Gaussian and uniform initialization. 2) When the codebook is updated during training, SimVQ continues to address the representation collapse issue, though with a slight degradation in performance. + +# 5.2. Audio Modality + +# 5.2.1. Baselines and Implementation Details + +We use the LibriTTS dataset [36] for audio-based VQ model training. The baselines such as Encodec [6], VOCos [24], and SpeechTokenizer [39] are based on residual vector quantization method. Our SimVQ model adopts the same architecture as WavTokenizer [13] with the only modification being the replacement of their EMA codebook with our one linear layer reparameterization method. We train SimVQ on LibriTTS-580h for 50 epochs with a batch size of 64. Note that WavTokenizer is trained with a 3-second window size for optimal performance, we train SimVQ using a 1-second window to accelerate training. For objective evaluation of the reconstructed audio, we follow VOCos [24] and employ metrics such as UTMOS [23], PESQ [20], STOI, and the F1 score for voiced/unvoiced classification (V/UV F1). UTMOS is particularly valuable as it produces scores highly correlated with human evaluations. + +# 5.2.2. Main Results + +Tab. 2 presents the reconstruction performance of various VQ models on audio data. Baseline models using residual vector quantization perform significantly worse than + +![](images/9c4d5bb5c26c628314a7b0a9b5f415c277866b4fcfb803dbcbed7b2a897f0476.jpg) +Figure 4. (a):(left) The rank of latent basis matrix $\mathbf{W}$ over training epochs. (b):(right) The Frobenius norm of latent basis matrix $\mathbf{W}$ over training epochs. + +SimVQ, even when utilizing much larger bandwidths. Despite using the same architecture as WAVTokenizer, our model, which replaces the quantization layer with SimVQ, achieves superior performance with a 1-second window size and maintains nearly $100\%$ codebook utilization when scaling up to a size of 262,144. The consistent performance of the SimVQ model across both image and audio data demonstrates that SimVQ is a general method for addressing the representation collapse problem in VQ models and can be effectively applied across multiple modalities. + +# 5.3. Analysis + +In Fig. 4(a), we plot the rank of the latent basis matrix over training epochs. Notably, SimVQ demonstrates the ability to adaptively adjust the rank of the latent space. Specifically, when the codebook size increases from $65k$ to $262k$ , the rank of the latent basis matrix decreases more rapidly and converges to a lower value. This observation suggests that a larger codebook can effectively alleviate the pressure on the latent space dimensionality, allowing the model to learn to represent data more efficiently. Additionally, despite the rank decreasing to a lower-rank space, SimVQ maintains $100\%$ codebook utilization, highlighting its superiority over VQGAN-FC, which struggles when increasing the latent dimension from 8 to 128. We also calculate the Frobenius norm of the latent basis matrix, as shown in Fig. 4. The norm of a codebook size of $262k$ is slightly larger than for $65k$ , indicating that a larger codebook can span a broader area in the linear space. For a comprehensive evaluation, we also provide the reconstruction loss curve on the ImageNet validation dataset in Appendix 8.2. The results consistently show that SimVQ achieves improved performance, further validating the effectiveness of our approach. + +# 5.4. Qualitative Evaluation + +We visualize the distribution of encoder features and codebook embeddings in Fig. 5 and the frequency in Appendix 8.3. Compared to vanilla VQ models that most codebook vectors are not unused, SimVQ can update the whole codebook and distribute the features evenly in the whole space. + +![](images/da97486972720edf2e8b89b0011a519ffb75d5ce39f845826b6685e876e4a1f4.jpg) +Figure 5. Visualization of the divergence between the encoder features and codebook embeddings on a random subset of ImageNet validation dataset. The left figure is vanilla VQ model and the right one is SimVQ. + +We qualitatively compare the reconstruction quality of both images and audio in Appendix 8.4. SimVQ can preserve more details with an enlarged codebook size, such as "eyes" and "text", which are challenging for vanilla VQ models. + +# 6. Discussion + +VQ Performance and Generative Models Many generative models [21, 25, 28, 34] utilize VQ models as a "tokenizer" to obtain discrete tokens. While it is intuitive that VQ reconstruction quality should impact generation performance, recent studies [35, 43] have revealed that the relationship is more nuanced: improved VQ reconstruction metrics do not necessarily translate to better generative outcomes. This complex relationship suggests that evaluating VQ models primarily through downstream generation tasks may not provide the most insightful assessment of their fundamental properties. Therefore, we focus on addressing the critical issue of representation collapse in VQ models, aiming to advance our understanding of their core representation learning mechanisms. + +# 7. Conclusion + +In this paper, we explore the representation collapse problem in VQ models. We conduct a theoretical analysis of the optimization process in VQ models and propose a simple yet effective method, SimVQ, to address this issue. Our method addresses the representation collapse by jointly optimizing the latent space through a linear transformation with one linear layer. Experimental results demonstrate that SimVQ outperforms previous approaches on both image and audio datasets, highlighting its broad applicability across diverse machine learning tasks. + +# Acknowledgement + +This research was supported by the National Natural Science Foundation of China (Grant No.62276245). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 1 +[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3 +[4] Zalan Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523-2533, 2023. 1 +[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Maria Elisabeth Bechtle, Feryal Behbahani, Stephanie C.Y. Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando De Freitas, Satinder Singh, and Tim Roktaschel. Genie: Generative interactive environments. In Proceedings of the 41st International Conference on Machine Learning, pages 4603-4623. PMLR, 2024. 1 +[6] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. Transactions on Machine Learning Research, 2023. Featured Certification, Reproducibility Certification. 1, 7 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6 +[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1 +[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 1, 2, 6 +[10] Zhangyang Gao, Cheng Tan, Jue Wang, Yufei Huang, Lirong Wu, and Stan Z Li. Foldtoken: Learning protein language via vector quantization and beyond. arXiv preprint arXiv:2403.09673, 2024. 1 +[11] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized net + +works. In Proceedings of the 40th International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3 +[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 3 +[13] Shengpeng Ji, Ziyue Jiang, Xize Cheng, Yifu Chen, Minghui Fang, Jialong Zuo, Qian Yang, Ruiqi Li, Zhang Zhang, Xiaoda Yang, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532, 2024. 7 +[14] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 3 +[15] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2, 6 +[16] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6 +[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[18] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3 +[19] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 6 +[20] A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra. Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and CODECs. In 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), pages 749-752 vol.2, 2001. 7 +[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 8 +[22] Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018. 1, 3 +[23] Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. ArXiv, abs/2204.02152, 2022. 7 + +[24] Hubert Siuzdak. Vocos: Closing the gap between timedomain and fourier-based neural vocoders for high-quality audio synthesis. In The Twelfth International Conference on Learning Representations, 2024. 7 +[25] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 8 +[26] Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi, Toshiyuki Kumakura, and Yuki Mitsu-fuji. SQ-VAE: Variational Bayes on discrete representation with self-annealed stochastic quantization. In Proceedings of the 39th International Conference on Machine Learning, pages 20987-21012. PMLR, 2022. 1, 3, 6 +[27] Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, and Ngai Wong. Scaling laws with vocabulary: Larger models deserve larger vocabularies. arXiv preprint arXiv:2407.13623, 2024. 1 +[28] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 8 +[29] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 2, 3 +[30] Tung-Long Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quantized wasserstein auto-encoder. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 1, 3, 6 +[31] Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023. 1 +[32] Pan Xiao, Peijie Qiu, and Aristeidis Sotiras. Scvae: Sparse coding-based variational autoencoder. ArXiv, abs/2303.16666, 2023. 1, 3, 6 +[33] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In International Conference on Learning Representations, 2022. 1, 3, 6 +[34] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 8 +[35] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 8 + +[36] Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. Libritts: A corpus derived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882, 2019. 7 +[37] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757-15773, Singapore, 2023. Association for Computational Linguistics. 1 +[38] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18467-18476, 2023. 3 +[39] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech language models. In The Twelfth International Conference on Learning Representations, 2024. 7 +[40] Chuanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798-22807, 2023. 1, 3, 6 +[41] Chuanxia Zheng, Guoxian Song, Tat-Jen Cham, Jianfei Cai, Dinh Q. Phung, and Linjie Luo. High-quality pluralistic image completion via code shared vqgan. ArXiv, abs/2204.01931, 2022. 2, 6 +[42] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of $99\%$ . ArXiv, abs/2406.11837, 2024. 1, 2, 3, 6, 7 +[43] Yongxin Zhu, Bocheng Li, Hang Zhang, Xin Li, Linli Xu, and Lidong Bing. Stabilize the latent space for image autoregressive modeling: A unified perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8 +[44] Yongxin Zhu, Dan Su, Liqiang He, Linli Xu, and Dong Yu. Generative pre-trained speech language model with efficient hierarchical transformer. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1764-1775, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1 \ No newline at end of file diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/images.zip b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6303106752174b43ed9b3a8bb8deb66f7947e4b6 --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f626fd5ce97ced46c99a537140b17401beb49aceb1b536c93734a685e3b2e49b +size 461917 diff --git a/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/layout.json b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab48141c5e5ff606ae283810d166697ac40a0fba --- /dev/null +++ b/2025/Addressing Representation Collapse in Vector Quantized Models with One Linear Layer/layout.json @@ -0,0 +1,10897 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 102, + 520, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 102, + 520, + 139 + ], + "spans": [ + { + "bbox": [ + 90, + 102, + 520, + 139 + ], + "type": "text", + "content": "Addressing Representation Collapse in Vector Quantized Models with One Linear Layer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "spans": [ + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "content": "Yongxin Zhu" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "content": ", Bocheng Li" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "content": ", Yifei Xin" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "content": ", Zhihua Xia" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "text", + "content": ", Linli Xu" + }, + { + "bbox": [ + 140, + 160, + 473, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 175, + 423, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 175, + 423, + 189 + ], + "spans": [ + { + "bbox": [ + 188, + 175, + 423, + 189 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "spans": [ + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "text", + "content": "State Key Laboratory of Cognitive Intelligence, " + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "text", + "content": "Peking University, " + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 189, + 513, + 203 + ], + "type": "text", + "content": "Jinan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 204, + 471, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 204, + 471, + 217 + ], + "spans": [ + { + "bbox": [ + 139, + 204, + 471, + 217 + ], + "type": "text", + "content": "zyx2016@mail.ustc.edu.cn,bcli@mail.ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 217, + 534, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 217, + 534, + 232 + ], + "spans": [ + { + "bbox": [ + 72, + 217, + 534, + 232 + ], + "type": "text", + "content": "xinyifei@stu.pku.edu.cn, xiazhihua@jnu.edu.cn, linlixu@ustc.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 283, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 297, + 499 + ], + "type": "text", + "content": "Vector Quantization (VQ) is essential for discretizing continuous representations in unsupervised learning but suffers from representation collapse, causing low codebook utilization and limiting scalability. Existing solutions often rely on complex optimizations or reduce latent dimensionality, which compromises model capacity and fails to fully solve the problem. We identify the root cause as disjoint codebook optimization, where only a few code vectors are updated via gradient descent. To fix this, we propose SimpleVQ, which reparameterizes code vectors through a learnable linear transformation layer over a latent basis, optimizing the entire linear space rather than nearest individual code vectors. Although the multiplication of two linear matrices is equivalent to applying a single linear layer, this simple approach effectively prevents collapse. Extensive experiments on image and audio tasks demonstrate that SimVQ improves codebook usage, is easy to implement, and generalizes well across modalities and architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 135, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 135, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 135, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 540, + 295, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 697 + ], + "type": "text", + "content": "In recent years, vector quantization (VQ) [19, 29] has emerged as a foundational technique in unsupervised representation learning [2, 5] and latent generative models [4, 21, 31, 33, 34, 44]. By converting continuous representations into discrete codes, VQ models can effectively identify the inherent structure of data and enable various discrete modeling methods on continuous data, from high-quality image generation [9] to audio synthesis [6]. The recent success of Large Language Models (LLMs) [1] has highlighted the effectiveness of next-token prediction as a powerful and versatile training objective. Consequently, VQ models are taken as the direct method to transform data from various modalities [25, 28, 37] or scientific domains [10] to discrete" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "text", + "content": "sequences for next token prediction training. However, attempts to integrate VQ models as multimodal tokenizers to leverage the scaling laws of LLMs face significant challenges because of the difficulty of expanding the codebook. For example, the Chameleon model [28] constrains its codebook size to " + }, + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "inline_equation", + "content": "8k" + }, + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "text", + "content": ", which is significantly trailing behind the vocabulary size of LLMs (e.g., LLaMA3's vocabulary size is " + }, + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "inline_equation", + "content": "128k" + }, + { + "bbox": [ + 313, + 260, + 554, + 356 + ], + "type": "text", + "content": " [8])." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 361, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 613 + ], + "type": "text", + "content": "There is a broad agreement that increasing vocabulary size can consistently improve the performance of LLMs [27]. However, recent studies [42] indicate that traditional VQ models often fail to utilize the additional parameters introduced by codebook expansion, leaving most codes inactive during training. The contradiction between codebook expansion and low codebook utilization in VQ models is known as the representation collapse problem [22], where increasing the codebook size fails to improve the performance. To address these discrepancies, we conduct a theoretical analysis of the optimization procedure of VQ models and identify that the disjoint optimization of the codebook is the root cause of representation collapse. As illustrated in Fig. 1(a), the core mechanism of VQ models involves a nearest-neighbor replacement strategy, where the encoder's output features are replaced by the nearest vector in the codebook to serve as input to the decoder. The indices of the nearest vector are taken as the discrete representation of the data. This nearest-selection operator results in only a subset of codes being updated through gradient descent, while the remaining codes remain unchanged." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Some methods mitigate representation collapse by hand-designing complex optimization strategies, such as stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] through sophisticated training strategies. Recently, some approaches [16, 33, 35] propose to reduce the dimension of the latent space to a very small scale (e.g., 8 v.s. 128) to alleviate the curse of dimensionality, thereby improving the overlap between the encoder's fea" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22968" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 309, + 185 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 309, + 185 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 309, + 185 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 309, + 185 + ], + "type": "image", + "image_path": "0e32b10883c72bc623c65cca52363719c50009416e5845bc98d83fcb0158b8bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 555, + 237 + ], + "type": "text", + "content": "Figure 1. Comparison of Vanilla VQ and SimVQ. (a): (left) Disjoint optimization in Vanilla VQ. Only the nearest codes are updated, resulting in a high percentage of \"dead\" codes that are not updated. (b): (right) Joint optimization in SimVQ. The entire codebook is updated with a latent basis, ensuring all codes remain active." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 312, + 68, + 543, + 185 + ], + "blocks": [ + { + "bbox": [ + 312, + 68, + 543, + 185 + ], + "lines": [ + { + "bbox": [ + 312, + 68, + 543, + 185 + ], + "spans": [ + { + "bbox": [ + 312, + 68, + 543, + 185 + ], + "type": "image", + "image_path": "e9eaa319bcd3756109ed568cab93cd588d8097f25b904638324a06aa7447b3d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 257, + 297, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 435 + ], + "type": "text", + "content": "tures and the codebook. However, while these methods enhance codebook utilization, they do so at the cost of model capacity, leading to worse performance compared to vanilla VQ models when the codebook size is small or representation collapse is not severe. Another approach, VQGAN-LC [42], initializes the codebook with features extracted from the pre-trained CLIP model [17] to create a well-structured latent space that better matches the distribution of the encoder output. Nevertheless, the latent space defined by an external pre-trained model limits the model's ability to generalize to diverse datasets and reaches a performance plateau as the codebook size increases. These limitations highlight the need for a more effective method to improve codebook utilization without compromising model capacity or relying on external models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "text", + "content": "We critically assess prevalent methodologies and reveal that optimizing the latent space rather than individual code vectors is key to preventing representation collapse. Building on this insight, we introduce a simple yet effective method, termed SimVQ, to directly update the latent space spanned by the codebook by linear transforming the code vectors via a learnable latent basis. Specifically, the vectors in the codebook are reparameterized as a linear combination of the basis in the learnable linear layer " + }, + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 437, + 296, + 544 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 549, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 549, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 96, + 549, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} \\in \\mathbb {R} ^ {K \\times d} \\Rightarrow \\boldsymbol {C W} \\text {w i t h} \\boldsymbol {W} \\in \\mathbb {R} ^ {d \\times d}, \\tag {1}", + "image_path": "ada3aa0b4bd0aa79380024b257a8d7d842ce571e5360ac2c771e0a2e3ddedfdc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " denotes the codebook size and " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " represents the dimension of latent space. This reparameterization with linear transformation disentangles the optimization of the codebook into two components: the coefficient matrix " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " and the basis of linear space " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " respectively. As illustrated in Fig. 1(b), by optimizing the basis matrix " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": ", the latent space spanned by " + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{CW}" + }, + { + "bbox": [ + 54, + 570, + 297, + 714 + ], + "type": "text", + "content": " is rotated and stretched to match encoder's output feature. The entire codebook is updated jointly to prevent the representation collapse problem. The simplicity of the proposed method makes it highly portable and easily adaptable for improving VQ-based models across a wide range of domains, requiring only one linear layer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 553, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 280 + ], + "type": "text", + "content": "In summary, our contributions to vector quantized models are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 282, + 555, + 510 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 282, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 330 + ], + "type": "text", + "content": "- We theoretically analyze the representation collapse problem in VQ models and reveal that optimizing the latent space spanned by the codebook, rather than individual code vectors, is crucial to addressing this issue." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 402 + ], + "type": "text", + "content": "- We propose a novel method, SimVQ, which reparameterizes the codebook vectors in VQ models via a linear transformation with a learnable latent basis. This simple yet effective approach is highly adaptable and easy to implement, making it broadly applicable across various machine learning contexts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 510 + ], + "type": "text", + "content": "- We conduct an extensive evaluation of SimVQ across diverse modalities, including image and audio with different model architectures. The results show that SimVQ not only effectively addresses the representation collapse problem by achieving near-complete codebook utilization regardless of the codebook size, but also establishes new state-of-the-art performance. Furthermore, when scaling up the codebook size, SimVQ consistently delivers improved results." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 524, + 402, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 402, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 402, + 537 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "text", + "content": "VQ-VAE [29] is the pioneering work to encode data into discrete representations. Building on these developments, VQGAN [9] combines VQ-VAE with adversarial networks to improve the perceptual quality of generated samples and establish a fundamental quantization protocol in latent generative models [21, 28, 34]. Many traditional VQ works focus on training a better discrete representation rather than codebook utilization to improve reconstruction performance. For example, RVQ [15] and MoVQ [41] enhance the reconstruction details with multichannel quantization. However, these methods suffer from a critical issue of representation collapse, as they struggle to scale the codebook size beyond " + }, + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 312, + 545, + 556, + 715 + ], + "type": "text", + "content": " entries, limiting their scalability. In response to this challenge, several approaches have been pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22969" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 454 + ], + "type": "text", + "content": "posed recently. DALLE [18] employs the gumbel-softmax trick [12] and stochastic sampling strategies to activate most codes during training. However, during inference, only a small subset of codes is utilized for quantization [38]. Some methods design complex optimization strategies to improve codebook utilization, such as stochastic quantization [26], distribution penalty [30, 32] and codebook reset [40]. Huh et al. [11] proposes rescaling the vectors in the codebook during training to match the distributions in the latent space. VQGAN-FC [33] introduces a method to map latent vectors into a lower-dimensional space followed by " + }, + { + "bbox": [ + 55, + 72, + 296, + 454 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 454 + ], + "type": "text", + "content": " normalization to alleviate representation collapse. FSQ [16] extends this idea by projecting representations into a reduced-dimensional space, where they are quantized into a small set of fixed values. LFQ [35], a variant of FSQ, uses binary values for quantized representations, thereby simplifying the encoding process. While these methods improve the codebook utilization, they do so at the cost of model capacity by significantly reducing the dimensionality of latent space (often to as low as 8), leading to worse performance compared to vanilla VQ models when the codebook size is small and representation collapse is not severe. Additionally, VQGAN-LC [42] proposes to initialize the codebook using features extracted from the pre-trained CLIP model to avoid representation collapse. However, the reliance on the pre-trained model limits the VQ model's ability to generalize to diverse datasets and results in a performance plateau as the codebook size increases. In contrast, our method, SimVQ, effectively addresses the representation collapse problem with a simple linear layer, without sacrificing model capacity or relying on external pre-trained models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 471, + 270, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 270, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 270, + 485 + ], + "type": "text", + "content": "3. Representation Collapse in VQ Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 492, + 142, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 142, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 142, + 504 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": "A vector quantized model is typically a reconstructive encoder-decoder architecture that includes a vector quantization layer to convert continuous representations into discrete codes. For simplicity, we represent an image with a single random variable " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": ". Formally, the encoder " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": " maps the input image into a latent space, producing a continuous representation " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "z_{e} = f_{\\theta}(x) \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": ". This representation is then quantized using a learnable codebook " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "C = [q_{1}, \\ldots, q_{K}] \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": " is a codebook vector. We define " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "\\delta_{k} \\in \\{0, 1\\}^{1 \\times K}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": " as a characteristic (one-hot) vector where only the " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": "-th element is 1, such that " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "q_{k} = \\delta_{k} C \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": ". The quantization layer selects the nearest codebook vector " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "q_{k}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": " by minimizing the Euclidean distance between " + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 55, + 512, + 296, + 680 + ], + "type": "text", + "content": " and the codebook entries [29]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 691, + 296, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 691, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 67, + 691, + 296, + 711 + ], + "type": "interline_equation", + "content": "k = \\arg \\min _ {j} \\| z _ {e} - q _ {j} \\| _ {2} ^ {2} = \\arg \\min _ {j} \\| z _ {e} - \\delta_ {j} C \\| _ {2} ^ {2}. \\tag {2}", + "image_path": "43b0cdb96c355febae01442bb1f210b9b9936bbfa4fed43a28c9559ab911fdbb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "The selected vector " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "q_{k}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is then passed to the decoder " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " to reconstruct the input image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "text", + "content": "To enable gradient propagation through the non-differentiable characteristic vector " + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "text", + "content": ", the straight-through estimator (STE) [3] is applied. During the backward pass, the gradient of " + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "inline_equation", + "content": "z_{q} = \\delta_{k}C" + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "text", + "content": " is copied to " + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 313, + 96, + 553, + 145 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 351, + 152, + 553, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 152, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 351, + 152, + 553, + 178 + ], + "type": "interline_equation", + "content": "z _ {q} = \\mathrm {s g} (\\delta_ {k} \\pmb {C} - z _ {e}) + z _ {e}, \\Rightarrow \\frac {\\partial z _ {q}}{\\partial z _ {e}} = 1 (3)", + "image_path": "f5a724fca635e2ad0c4c1d23d1f9640980743aeab22be29a0480dc2c31377e7f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\mathfrak{sg}" + }, + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "text", + "content": " is the stop gradient operator, ensuring the gradient for " + }, + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "inline_equation", + "content": "\\delta_k C" + }, + { + "bbox": [ + 313, + 185, + 553, + 209 + ], + "type": "text", + "content": " is discarded during the backward pass." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 209, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 256 + ], + "type": "text", + "content": "The learning objective is the combination of a reconstruction loss and commitment loss that ensures that the encoder commits to an embedding and the encoder's output does not drift:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 265, + 553, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 553, + 290 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\log p (x | z _ {q}) + \\| \\operatorname {s g} \\left(\\delta_ {k} \\boldsymbol {C}\\right) - z _ {e} \\| _ {2} ^ {2} + \\beta \\| \\delta_ {k} \\boldsymbol {C} - \\operatorname {s g} \\left(z _ {e}\\right) \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "0f3a2dc1aa1c973ae06bacb5fc5f31d8f4afe2ab283721f3f88daa9b5e7d0fbc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "inline_equation", + "content": "\\log p(x|z_q)" + }, + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "text", + "content": " is typically the mean squared error (MSE) loss " + }, + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "inline_equation", + "content": "\\| x - g_{\\phi}(z_q)\\| _2^2" + }, + { + "bbox": [ + 313, + 290, + 553, + 316 + ], + "type": "text", + "content": " for image and audio data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 322, + 501, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 501, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 501, + 335 + ], + "type": "text", + "content": "3.2. Disjoint Optimization of Codebook" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 434 + ], + "type": "text", + "content": "In VQ models, only the nearest code is selected and updated via gradient descent. Ideally, all codebook entries should be updated and utilized for decoding. However, experimental evidence shows that only a small fraction of the codebook gets updated and utilized, leading to what is known as the representation collapse problem [22]. To investigate the root cause of this issue, we provide a theoretical analysis of the optimization dynamics in VQ models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "content": "Due to the use of the straight-through estimator (STE) for gradient propagation, the codebook " + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 434, + 554, + 481 + ], + "type": "text", + "content": " can only be updated through the gradient of the commitment loss, which is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 372, + 491, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 491, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 372, + 491, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C}) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\right\\| _ {2} ^ {2}. \\tag {5}", + "image_path": "59588341841faeb268c7ff2e57828cf840bc31fbf73a15c5791e513614529b4f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "text", + "content": "The codebook " + }, + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "text", + "content": " is updated according to the following equation, where " + }, + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 514, + 553, + 538 + ], + "type": "text", + "content": " is the learning rate:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 546, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 546, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 344, + 546, + 553, + 596 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {C} ^ {(t + 1)} = \\boldsymbol {C} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} (\\boldsymbol {C} ^ {(t)})}{\\partial \\boldsymbol {C} ^ {(t)}} \\right] (6) \\\\ = \\boldsymbol {C} ^ {(t)} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} ^ {(t)} \\right] + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\delta_ {k} ^ {T} z _ {e} \\right] (7) \\\\ \\end{array}", + "image_path": "fa8d6b021959fc0a73ed78a34b7456482cd9069850654202c584ed26ce85d309.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 604, + 531, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 531, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 531, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 604, + 531, + 617 + ], + "type": "inline_equation", + "content": "\\delta_k^T\\delta_k" + }, + { + "bbox": [ + 313, + 604, + 531, + 617 + ], + "type": "text", + "content": " is the Kronecker delta matrix, defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 370, + 626, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 626, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 370, + 626, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\left(\\delta_ {k} ^ {T} \\delta_ {k}\\right) _ {i j} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} i = j = k, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {8}", + "image_path": "08e7bb2db009e4bc096af892d3bc76c199e9a0a5638e485cbf6ba9f74c1792b7.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "All vectors in " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " will be updated and utilized if and only if the expectation " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right]" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " converges to the identity matrix. Unlike variational autoencoders (VAEs) [14], which enforce a Gaussian distribution on the latent space via a" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22970" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "KL-divergence penalty, VQ models optimize " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " towards the selected codebook vectors " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_kC\\right]" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". At the same time, the selected codebook vectors are optimized towards the distribution of " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", resulting in the same selected subset of vectors moving closer to " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", somewhat akin to a cocoon effect. However, this disjoint optimization of the codebook leads to part of the codebook, specifically " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "(I - \\mathbb{E}_{z_e}\\left[\\delta_k^T\\delta_k\\right])C" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", remaining un-updated and underutilized once the optimization process begins. This phenomenon occurs because the optimization focuses only on a subset of codebook vectors, leaving other vectors stagnant." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 288 + ], + "type": "text", + "content": "This analysis reveals the fundamental cause of representation collapse in VQ models: the disjoint optimization process that updates only a subset of codebook vectors. This insight forms the basis for our proposed solution, SimVQ, which aims to address this issue by optimizing the entire latent space spanned by the codebook, rather than individual code vectors." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 327 + ], + "type": "text", + "content": "4. Addressing Collapse with Latent Linear Transformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 269, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 269, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 269, + 348 + ], + "type": "text", + "content": "4.1. Reparameterize Codes with Latent Basis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\pmb{W} = \\{\\pmb{w}_1, \\dots, \\pmb{w}_n\\}" + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": " be a basis of a linear space. Any vector " + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": " in the space can be uniquely expressed as a linear combination of the basis vectors with coefficients " + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "inline_equation", + "content": "c_1, \\dots, c_n \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 353, + 296, + 401 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 413, + 294, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 413, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 294, + 425 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} = c _ {1} \\boldsymbol {w} _ {1} + \\dots + c _ {n} \\boldsymbol {w} _ {n} = \\boldsymbol {c} \\boldsymbol {W}. \\tag {9}", + "image_path": "8304014f622b7d224b1c3c625b48d0abe12fdc49bcfaed985044610dfe040edf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": "Given the equivalence between " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "c\\pmb{W}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": " in the linear space, we can reparameterize each vector in the codebook of VQ models with a new basis matrix " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\pmb{W} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": ". Specifically, the codebook " + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\pmb{C} = \\{c_1, \\dots, c_K\\}" + }, + { + "bbox": [ + 55, + 435, + 296, + 495 + ], + "type": "text", + "content": " can be reparameterized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 505, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 505, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 97, + 505, + 295, + 520 + ], + "type": "interline_equation", + "content": "\\{\\hat {\\boldsymbol {c}} _ {1} \\boldsymbol {W}, \\dots , \\hat {\\boldsymbol {c}} _ {N} \\boldsymbol {W} \\} = \\hat {\\boldsymbol {C}} \\boldsymbol {W} \\in \\mathbb {R} ^ {K \\times d}. \\tag {10}", + "image_path": "174ce729b212318a0f3ba0e710013e64e878cdff9c2b4871090e8ab5a3b803e2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": "This reparameterization introduces two components: the basis matrix " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " and the coefficient matrix " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": ". In the following, we will discuss the optimization of both the basis matrix " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " and the coefficient matrix " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": ". For simplicity, we slightly abuse " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 590 + ], + "type": "text", + "content": " below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 600, + 250, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 250, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 250, + 613 + ], + "type": "text", + "content": "4.2. Asymmetric Optimization Dynamics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "While it is commonly accepted that multiplying two linear matrices is equivalent to a single linear layer, we argue that the disjoint optimization problem of the codebook in VQ models can be addressed by linear transformation. In vanilla VQ models, only the codebook " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " is responsible for minimizing commitment loss, leading to the disjoint optimization problem where only the selected codes will be updated." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 434, + 151 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 434, + 151 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 434, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 434, + 151 + ], + "type": "image", + "image_path": "acaa8fcbfd39e70edaae65d82bff2a10fc759155be00813d480232790e21b499.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "text", + "content": "Figure 2. (a): (left) The optimization trajectory of the objective " + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "inline_equation", + "content": "\\| \\pmb{x} - \\pmb{q}\\|_2^2" + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "text", + "content": ", which is the same as vanilla VQ. Only a small fraction of points are updated while others remain inactive. (b): (right) The optimization trajectory of the objective " + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "inline_equation", + "content": "\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2" + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "text", + "content": " frozen, which is the same as SimVQ. All the points are updated towards targets " + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 159, + 554, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 552, + 151 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 552, + 151 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 552, + 151 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 552, + 151 + ], + "type": "image", + "image_path": "26aa03cbd2994a9c5092415892590343fd3a37c6be1fc0685e8a96b9a619c387.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 237, + 430, + 314 + ], + "blocks": [ + { + "bbox": [ + 317, + 237, + 430, + 314 + ], + "lines": [ + { + "bbox": [ + 317, + 237, + 430, + 314 + ], + "spans": [ + { + "bbox": [ + 317, + 237, + 430, + 314 + ], + "type": "image", + "image_path": "ab17f67fcb4d5ed7446cec316c3649888bf536cb9d657eeef1534741bb777ca8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": "Figure 3. (a): (left) The optimization trajectory of the optimization objective: " + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\| \\pmb{x} - \\pmb{q}\\pmb{w}\\|_2^2" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": " with both " + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": " unfrozen. (b): (right) The Frobenius norm of the projection matrix " + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": " and loss curves. The loss quickly converges to 0 with " + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 322, + 554, + 367 + ], + "type": "text", + "content": " almost unchanged." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 431, + 237, + 552, + 314 + ], + "blocks": [ + { + "bbox": [ + 431, + 237, + 552, + 314 + ], + "lines": [ + { + "bbox": [ + 431, + 237, + 552, + 314 + ], + "spans": [ + { + "bbox": [ + 431, + 237, + 552, + 314 + ], + "type": "image", + "image_path": "cbaae70c1eeb78bd94e2902795523e31e2fa602a5508e663652b2ab6c1c7321f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": "In contrast, when the codebook is reparameterized as " + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "inline_equation", + "content": "CW" + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": ", both the basis " + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": " and the coefficient matrix " + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": " contribute to minimizing the commitment loss. The gradients " + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial W}" + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\frac{\\partial\\mathcal{L}}{\\partial C}" + }, + { + "bbox": [ + 313, + 388, + 554, + 457 + ], + "type": "text", + "content": " can simultaneously reduce the loss. As a result, the optimization of the reparameterized codebook can be divided into three scenarios:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 459, + 553, + 567 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": "- Updating " + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": " frozen: Only the selected codes adapt to the latent distribution of " + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": ", as depicted on Fig. 1(a). The vanilla VQ is a special case with " + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "inline_equation", + "content": "W = I" + }, + { + "bbox": [ + 314, + 459, + 553, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": "- Updating " + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": " frozen: The entire codebook " + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "inline_equation", + "content": "CW" + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": " adjusts to the latent distribution of " + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": ". The basis matrix " + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 314, + 495, + 553, + 530 + ], + "type": "text", + "content": " rotates and stretches the space as shown in Fig. 1(b)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "content": "- Updating both " + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "content": ": The selected subset of codes moves towards " + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "inline_equation", + "content": "z_{e}" + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "content": " while the space spanned by " + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 314, + 531, + 553, + 567 + ], + "type": "text", + "content": " undergoes simultaneous rotation and stretching." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "content": "To highlight the difference in optimization between " + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "inline_equation", + "content": "CW" + }, + { + "bbox": [ + 313, + 567, + 554, + 615 + ], + "type": "text", + "content": ", we conduct a toy experiment in a two-dimensional setting and visualize the optimization process in Fig. 2 and Fig. 3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 620, + 403, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 403, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 403, + 632 + ], + "type": "text", + "content": "4.2.1. Toy Examples" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": "We randomly sample two target points " + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": " from Gaussian distribution as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 665, + 553, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 665, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 665, + 553, + 700 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {1} \\sim \\mathcal {N} (\\binom {2} {2}, \\binom {1 0} {0 1}), \\quad \\boldsymbol {x} _ {2} \\sim \\mathcal {N} (\\binom {- 2} {- 2}, \\binom {1 0} {0 1}). \\tag {11}", + "image_path": "f2237269b87afccfcbaa2da931d0531560d69dfa792e9b17f95ab689ea8c1281.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "content": "Then we initialize 10 learnable vectors " + }, + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 314, + 701, + 553, + 713 + ], + "type": "text", + "content": " from a Gaussian" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22971" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 237, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 237, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 237, + 83 + ], + "type": "text", + "content": "Algorithm 1 Training Procedure for SimVQ" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "spans": [ + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": "Input: Encoder " + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": ", Decoder " + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": ", Codebook " + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": ", Linear projector matrix " + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W}_{\\psi}" + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": ", commitment weight " + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 65, + 87, + 295, + 113 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "text", + "content": "Output: Model parameters " + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "inline_equation", + "content": "\\theta, \\phi, \\psi" + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "text", + "content": " and Codebook " + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "text", + "content": ". Initialize Codebook with Gaussian distribution and freeze the parameter of Codebook " + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 65, + 114, + 296, + 148 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 150, + 96, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 150, + 96, + 160 + ], + "spans": [ + { + "bbox": [ + 66, + 150, + 96, + 160 + ], + "type": "text", + "content": "repeat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 161, + 160, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 161, + 160, + 173 + ], + "spans": [ + { + "bbox": [ + 77, + 161, + 160, + 173 + ], + "type": "text", + "content": "Draw " + }, + { + "bbox": [ + 77, + 161, + 160, + 173 + ], + "type": "inline_equation", + "content": "x\\sim p_{data}(\\pmb {x})" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 174, + 126, + 185 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 174, + 126, + 185 + ], + "spans": [ + { + "bbox": [ + 77, + 174, + 126, + 185 + ], + "type": "interline_equation", + "content": "z _ {e} = f _ {\\theta} (x);", + "image_path": "eb9e8481244dd023c60bea770c70a38346c3a84d44476f499fa501c41df67c17.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "spans": [ + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "inline_equation", + "content": "/^{*}" + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "text", + "content": " Replace " + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "text", + "content": " in vanilla VQ with proposed " + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "inline_equation", + "content": "q_{j}\\pmb{W}_{\\psi}" + }, + { + "bbox": [ + 77, + 185, + 277, + 197 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "text", + "content": "Nearest code search: " + }, + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "inline_equation", + "content": "k = \\arg \\min_{j} \\| z_{e} - q_{j}W_{\\psi} \\|_{2}^{2}" + }, + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "inline_equation", + "content": "q_{j} \\in C" + }, + { + "bbox": [ + 77, + 197, + 294, + 220 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 77, + 220, + 294, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 220, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 77, + 220, + 294, + 243 + ], + "type": "text", + "content": "Straight Through Estimation: " + }, + { + "bbox": [ + 77, + 220, + 294, + 243 + ], + "type": "inline_equation", + "content": "z_{q} = \\mathrm{sg}(q_{k}\\mathbf{W}_{\\psi} - z_{e}) + z_{e}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 245, + 127, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 245, + 127, + 257 + ], + "spans": [ + { + "bbox": [ + 77, + 245, + 127, + 257 + ], + "type": "interline_equation", + "content": "\\hat {x} = g _ {\\phi} (z _ {q});", + "image_path": "66f277f0113c7dbe2f54d1315a49502aef5150189efbe24d2c62ec150c13573e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "text", + "content": "Minimize " + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta ,\\phi ,\\psi) = \\mathrm{MSE}(x,\\hat{x})" + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "text", + "content": " + " + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "inline_equation", + "content": "\\beta \\| z_{e}" + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "text", + "content": " - " + }, + { + "bbox": [ + 77, + 257, + 294, + 281 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}(q_kW_\\psi)\\| _2^2 +\\| \\mathrm{sg}(z_e) - q_kW_\\psi \\| _2^2;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 281, + 132, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 132, + 292 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 132, + 292 + ], + "type": "text", + "content": "until converged" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 317, + 107, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 107, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 107, + 327 + ], + "type": "text", + "content": "distribution:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 111, + 335, + 294, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 335, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 111, + 335, + 294, + 363 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {q} _ {i} \\right\\} _ {i = 1} ^ {1 0} \\sim \\mathcal {N} \\left(\\binom {0} {0}, \\binom {1} {0}, \\binom {0} {1}\\right), \\tag {12}", + "image_path": "8418d979280398cb93613632f79b81e6c5f82dfa3a54b22ace0b8d10a502e341.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": "During training with gradient descent, we introduce perturbation noise " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,0.01)" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " to the targets. In Fig. 2(a), the optimization objective is similar to vanilla VQ: " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\| \\pmb {x} - \\pmb {q}\\| _2^2" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": ". Only the nearest points " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "q_{4}" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "q_{10}" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " are updated. In contrast, in Fig. 2(b), the optimization objective " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\| \\pmb {x} - \\pmb {q}\\pmb {w}\\| _2^2" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " is similar to SimVQ with the points reparameterized by a learnable latent basis " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " frozen, resulting in the entire codebook " + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\{\\pmb {q}\\}_{i = 1}^{10}" + }, + { + "bbox": [ + 55, + 371, + 296, + 468 + ], + "type": "text", + "content": " being jointly updated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "text", + "content": "Remark 1. The simultaneous optimization of the latent basis " + }, + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "text", + "content": " and the coefficient matrix " + }, + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 475, + 296, + 500 + ], + "type": "text", + "content": " may lead to the collapse." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": "We provide an example in Fig. 3(a) where the optimization objective is " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\| \\pmb{x} - \\pmb{q}\\pmb{w} \\|_2^2" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " unfrozen this time. In the training process, only the nearest point " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}_1" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " and point " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{10}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " move towards the target point, while other points remain almost unchanged. We also visualize the loss curve in Fig. 3(b). The optimization objective with both " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " unfrozen converges quickly, where the norm of basis " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " is much smaller than the objective with " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " frozen. This indicates that the disjoint optimization of the codebook persists: " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " can directly commit to the loss and dominate the optimization process, with " + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 55, + 508, + 296, + 652 + ], + "type": "text", + "content": " being ignored, leading to the collapse quickly." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 659, + 246, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 246, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 246, + 672 + ], + "type": "text", + "content": "4.3. Joint Optimization of the Codebook" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We propose SimVQ by simply using a learnable basis " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " to reparameterize the codebook such that the codebook is transformed into " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "CW" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". The pseudo-code for this" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "approach is provided in Algorithm 1. During training, we optimize only the latent basis matrix " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": ", while keeping the coefficient matrix " + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": " frozen. The commitment loss for SimVQ is defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 359, + 128, + 553, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 128, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 359, + 128, + 553, + 143 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e}, q _ {k}\\right) = \\left\\| z _ {e} - \\delta_ {k} \\boldsymbol {C} \\boldsymbol {W} \\right\\| _ {2} ^ {2}. \\tag {13}", + "image_path": "79a297543fd670a72dfd6c2008c6b39e40186d5317712d03329b99148de76d60.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "content": "The vanilla VQ model is a special case of SimVQ, where the linear basis matrix " + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "content": " is fixed as the identity matrix " + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "content": ". The update for " + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "content": " with learning rate " + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 151, + 554, + 187 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 194, + 553, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 194, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 324, + 194, + 553, + 219 + ], + "type": "interline_equation", + "content": "\\boldsymbol {W} ^ {(t + 1)} = \\boldsymbol {W} ^ {(t)} - \\eta \\frac {\\partial \\mathcal {L} _ {\\text {c o m m i t}} \\left(z _ {e} , \\boldsymbol {q} _ {k}\\right)}{\\partial \\boldsymbol {W} ^ {(t)}} \\tag {14}", + "image_path": "0f7db4fbe9c196df38226b0542437c91012e79f9c619da7125bd41d1b50518f7.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 327, + 220, + 553, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 220, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 327, + 220, + 553, + 246 + ], + "type": "interline_equation", + "content": "= \\left(\\boldsymbol {I} - \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} \\delta_ {k} \\boldsymbol {C} \\right]\\right) \\boldsymbol {W} ^ {(t)} + \\eta \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {C} ^ {T} \\delta_ {k} ^ {T} z _ {e} \\right]. \\tag {15}", + "image_path": "60d45ef1b80b1361298190fefa4591dff77a5cd1035c84f9604441fdb9043e2c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "text", + "content": "The term " + }, + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "inline_equation", + "content": "\\mathbb{E}\\left[\\boldsymbol{C}^T\\delta_k^T\\delta_k\\boldsymbol{C}\\right]" + }, + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "text", + "content": " represents the expectation of the quadratic form, and simplifies to " + }, + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\pmb{q}_k^T\\pmb{q}_k]" + }, + { + "bbox": [ + 313, + 255, + 554, + 302 + ], + "type": "text", + "content": ". Since the codes are randomly sampled from a Gaussian distribution, we have:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 361, + 303, + 553, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 303, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 361, + 303, + 553, + 319 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {q} _ {k} \\right] = \\boldsymbol {I}, \\text {w h e r e} \\boldsymbol {q} \\sim \\mathcal {N} (0, 1), \\tag {16}", + "image_path": "bba50db26f2a4db64807f61eae1795ab51db83cbddf7297583efd0c4ebc49b85.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": "which ensures that all elements of " + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": " are updated. As training progresses, the latent basis " + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": " converges to:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 381, + 354, + 553, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 354, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 381, + 354, + 553, + 372 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow \\infty} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} _ {z _ {e}} \\left[ \\boldsymbol {q} _ {k} ^ {T} z _ {e} \\right] \\tag {17}", + "image_path": "75caa752d195702837f1977c333f82b3b5361e25f2f23da2031dfe8a178d0495.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 380, + 389, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 389, + 391 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 389, + 391 + ], + "type": "text", + "content": "Thus, in the limit:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 361, + 399, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 399, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 361, + 399, + 553, + 418 + ], + "type": "interline_equation", + "content": "\\lim _ {t \\rightarrow \\infty} \\boldsymbol {q} _ {k} \\boldsymbol {W} ^ {(t)} = \\mathbb {E} [ \\boldsymbol {q} _ {k} \\boldsymbol {q} _ {k} ^ {T} \\boldsymbol {e} ] = \\mathbb {E} [ \\boldsymbol {e} ] \\tag {18}", + "image_path": "17dbbd42cd39676807aaa2035201667224a766facb2b386cc7815b5191775029.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 426, + 553, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 448 + ], + "type": "text", + "content": "At convergence, the product " + }, + { + "bbox": [ + 313, + 426, + 553, + 448 + ], + "type": "inline_equation", + "content": "q_{k}W" + }, + { + "bbox": [ + 313, + 426, + 553, + 448 + ], + "type": "text", + "content": " equals the nearest feature." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 456, + 425, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 425, + 469 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 425, + 469 + ], + "type": "text", + "content": "4.4. Efficiency Analysis" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": "SimVQ demonstrates greater efficiency than vanilla VQ due to its asymmetric training strategy, wherein the codebook " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " remains static and only the linear projection " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is optimized. This approach results in a significant reduction in memory usage during the gradient backpropagation process. In vanilla VQ, the memory cost for codebook optimization is " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "O(Kd)" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is the number of vectors in the codebook, and " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is the dimension of each vector. In our experiments, " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "K = 65,536" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is much larger than " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "d = 128" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": ". As the vocabulary size increases, the memory required for backpropagation grows proportionally, significantly impacting resource consumption. In contrast, SimVQ's memory cost for backpropagation is only " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "O(d^2)" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " because the codebook " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is fixed, and only the linear layer " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " is updated. This results in a constant memory requirement in backpropagation, independent of the vocabulary size. The " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "d \\times d" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " scaling becomes particularly advantageous as " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " increases in practical applications. This structural design minimizes the computational overhead and improves training efficiency, especially when dealing with large vocabularies." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22972" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 523, + 209 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 523, + 209 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 523, + 209 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 523, + 209 + ], + "type": "table", + "html": "
MethodLatent dimCodebook sizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN [9]12865,5361.4%3.740.1722.2070.6
VQGAN-EMA [19]12865,5364.5%3.230.1522.8972.3
VQGAN-FC [33]12865,5361.4%5.330.1821.4568.8
VQGAN-FC [33]865,536100.0%2.630.1323.7977.5
FSQ†[16]664,000100.0%2.800.1323.6375.8
LFQ [35]1665,536100.0%2.880.1323.6077.2
VQGAN-LC-CLIP+ [42]76865,536100.0%2.400.1323.9877.3
SimVQ (ours)12865,536100.0%2.240.1224.1578.4
SimVQ (ours)128262,144100.0%1.990.1124.6880.3
", + "image_path": "d2fcd8f4be6e6f4e5e9f18836fa7e11d59eb3dd579768e613730e9718c561a6f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": "Table 1. Reconstruction performance on ImageNet-1k with a resolution of " + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "inline_equation", + "content": "{128} \\times {128}" + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": " . All models are trained using images downsampled into " + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "inline_equation", + "content": "{16} \\times {16}" + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": " tokens. † Results are reproduced using the codebook size of " + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\left\\lbrack {8,8,8,5,5,5}\\right\\rbrack" + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": " to approximately match " + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "inline_equation", + "content": "{65},{536}. +" + }, + { + "bbox": [ + 55, + 217, + 555, + 262 + ], + "type": "text", + "content": " Following VQGAN-LC,we extract CLIP features with the codebook frozen. The codebook utilization is calculated as the fraction of the codes that are activated at least once when encoding the validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 281, + 136, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 136, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 136, + 295 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 301, + 295, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 362 + ], + "type": "text", + "content": "To assess the efficacy and versatility of the proposed SimVQ, we conduct experiments across both image and audio modalities. Subsequently, we analyze the learned linear layer to investigate the latent basis. The experimental configurations are listed in Appendix 8.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 153, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 153, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 153, + 383 + ], + "type": "text", + "content": "5.1. Vision Modality" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 388, + 124, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 124, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 124, + 399 + ], + "type": "text", + "content": "5.1.1. Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 631 + ], + "type": "text", + "content": "Our baseline selection focuses on methods that enhance codebook utilization through architectural designs. We include VQGAN-FC [33], FSQ [16], LFQ [35] and VQGAN-LC-CLIP [42] as our primary baselines, as they represent the current state-of-the-art in improving reconstruction performance through codebook architecture innovations. Traditional VQ variants such as RVQ [15] and MoVQ [41] address fundamentally different technical challenges - they focus on training better discrete representations, rather than enhancing codebook utilization to improve reconstruction performance like our work. Another important research direction explores optimization-based solutions, where methods like stochastic quantization [26], distribution penalty [30, 32], and codebook reset [40] tackle codebook utilization through sophisticated training strategies. Given our focus on architectural innovations in codebook design, we evaluate SimVQ against methods that share this technical foundation for a meaningful assessment of our contributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 639, + 185, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 185, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 185, + 651 + ], + "type": "text", + "content": "5.1.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "To rigorously evaluate the proposed SimVQ, we reproduce all the VQ models listed in Tab. 1 using the same architecture of VQGAN [9] with the quantization layer different only. Among the baselines, for VQGAN-FC [33], we follow the original setting to reduce the dimension of the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": "latent space to 8 followed by " + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": " normalization to improve codebook utilization. For FSQ [16], we adopt a codebook size of " + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "inline_equation", + "content": "[8,8,8,5,5,5]" + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": " as recommended, to approximately match the default codebook size. For VQGAN-LC [42], we follow them and leverage an external pre-trained CLIP model to extract features of the training dataset in advance for a well-defined latent space. All models are trained on the ImageNet [7] dataset for 50 epochs with a batch size of 256. Input images are processed at a resolution of " + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": " pixels and downsampled by a factor of 8, yielding a feature map of " + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 128" + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": ", where 128 is the dimension of the latent space. We set the default codebook size to a large number of " + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "inline_equation", + "content": "2^{16} = 65536" + }, + { + "bbox": [ + 313, + 282, + 555, + 475 + ], + "type": "text", + "content": " rather than the traditional number 8192 to highlight the representation collapse problem. Performance is evaluated using rFID, LPIPS, PSNR, and SSIM metrics on the ImageNet validation set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 483, + 400, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 400, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 400, + 494 + ], + "type": "text", + "content": "5.1.3. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "content": "Tab. 1 presents the reconstruction performance of various VQ models on image data. We make three key observations: 1) Traditional VQGAN models utilize only a very small subset of the codebook, with a utilization rate of just " + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "content": ". Although VQGAN-EMA is proposed to improve codebook utilization, especially when the codebook size scales up to " + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "inline_equation", + "content": "65k" + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "content": ", it still suffers from severe representation collapse. 2) Recently proposed methods, such as LFQ, FSQ, and VQGAN-FC, effectively improve codebook utilization to " + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "content": ". However, these methods require reducing the latent space to a very low dimension. For example, applying VQGAN-FC to the standard latent dimension of 128 results in severe representation collapse and degraded reconstruction performance. Additionally, these models face limitations in model capacity due to the low-dimensional latent space. While they achieve full codebook utilization, their reconstruction quality on rFID score lags significantly behind SimVQ. 3) VQGAN-LC-CLIP leverages an exter" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22973" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 525, + 235 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 525, + 235 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 525, + 235 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 525, + 235 + ], + "type": "table", + "html": "
MethodBandwidthUtil↑UTMOS↑PESQ↑STOI↑V/UV F1↑
GT--4.06/3.48---
EnCodec [6]3.0kbps-2.31/2.092.05/2.050.90/0.880.92/0.89
Vocos [24]3.0kbps-3.53/3.062.40/2.190.92/0.900.94/0.91
SpeechTokenizer [39]3.0kbps-3.56/3.021.93/1.740.88/0.840.93/0.89
WavTokenizer [13]0.9kbps100/100%3.74/3.43*2.01/2.26*0.89/0.89*0.92/0.92*
SimVQ (ours)0.9kbps100.0/100.0%4.00/3.512.33/2.080.91/0.880.94/0.91
WavTokenizer [13]0.975kbps68/-%4.02*/-2.39*/-0.92*/-0.94*/-
WavTokenizer [13]1.05kbps27/-%4.00*/-2.36*/-0.81*/-0.94*/-
SimVQ (ours)0.975kbps99.4/99.4%4.03/3.522.42/2.150.92/0.880.94/0.92
SimVQ (ours)1.2kbps99.4/99.0%4.03/3.522.54/2.260.93/0.900.94/0.92
SimVQ (ours)1.35kbps95.6/94.7%4.03/3.532.61/2.310.93/0.900.95/0.93
", + "image_path": "9be7bdf5e71d460d31e29223d0350c2bdb60a1c47ec2b3ce5a7bfdf09a704831.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 283, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 55, + 242, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 553, + 266 + ], + "type": "text", + "content": "Table 2. Reconstruction performance on LibriTTS test-clean/test-other dataset. * WavTokenizer is trained with a window size of 3 seconds. The bandwidth of 0.9kbps, 0.975kbps, 1.2kbps, 1.35kbps means the codebook size of 4096, 8192, 65536, 262144 respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 283, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 57, + 283, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 57, + 283, + 294, + 354 + ], + "type": "table", + "html": "
MethodCodebook SizeUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
VQGAN-LC-CLIP† [42]50,00099.9%2.750.1323.858.4
VQGAN-LC-CLIP† [42]100,00099.9%2.620.1223.858.9
VQGAN-LC-CLIP† [42]200,00099.8%2.660.1223.959.2
SimVQ1,024100.0%3.670.1622.3470.8
SimVQ8,192100.0%2.980.1423.2374.7
SimVQ65,536100.0%2.240.1224.1578.4
SimVQ262,144100.0%1.990.1124.6880.3
", + "image_path": "91167f61c8c19f7c0c766a91a082ba6ad1c736e4e5f5ba99230064ebf0d03df9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 415, + 294, + 462 + ], + "blocks": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": "Table 3. Ablation study on the effect of various codebook sizes on ImageNet at a resolution of " + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": " We directly copy the reported results of VQGAN-LC from the original paper on ImageNet " + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 361, + 295, + 407 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 415, + 294, + 462 + ], + "lines": [ + { + "bbox": [ + 58, + 415, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 415, + 294, + 462 + ], + "type": "table", + "html": "
InitializationTrainableUtil↑rFID↓LPIPS↓PSNR↑SSIM↑
GaussianYes100.0%2.310.1224.0477.2
UniformNo100.0%2.310.1224.1578.4
GaussianNo100.0%2.240.1224.1578.4
", + "image_path": "cc5a6459ba67b87be6675f4db95627fb6edf24f95d25769183437e2548600549.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 470, + 282, + 482 + ], + "lines": [ + { + "bbox": [ + 68, + 470, + 282, + 482 + ], + "spans": [ + { + "bbox": [ + 68, + 470, + 282, + 482 + ], + "type": "text", + "content": "Table 4. Ablation study of codebook optimization strategy." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 501, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 598 + ], + "type": "text", + "content": "nal pre-trained CLIP model to provide a well-defined latent space. However, VQGAN-LC relies on CLIP features pretrained on much larger datasets than ImageNet, which introduces generalization issues and a lower performance ceiling (degradation issue in Tab. 3). In contrast, SimVQ can be applied to a wide range of data types and achieves superior performance (rFID 2.40 vs. 2.24) without the limitations imposed by a pre-trained feature extraction model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 149, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 149, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 149, + 615 + ], + "type": "text", + "content": "5.1.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "On the Codebook Size In Tab. 3, we explore the impact of different codebook sizes, ranging from " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "262k" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", which is typically the level of LLM's vocabulary size. SimVQ consistently improves performance as the codebook size increases. For instance, the rFID score decreases to 1.99, and SSIM surpasses 80.0. In contrast, while VQGAN-LC-CLIP can keep high codebook utilization as increasing codebook size, it encounters performance degradation, with the rFID" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 311 + ], + "type": "text", + "content": "score worsening from 2.62 to 2.66 when the codebook size is increased from 100,000 to 200,000." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 326, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 422 + ], + "type": "text", + "content": "On the Codebook Optimization Strategy We investigate codebook initialization and the training of the linear layer in Tab. 4. Our findings are as follows: 1) The codebook is robust to different initialization strategies, yielding similar results with both Gaussian and uniform initialization. 2) When the codebook is updated during training, SimVQ continues to address the representation collapse issue, though with a slight degradation in performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 410, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 410, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 410, + 443 + ], + "type": "text", + "content": "5.2. Audio Modality" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 448, + 503, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 503, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 503, + 460 + ], + "type": "text", + "content": "5.2.1. Baselines and Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 463, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 656 + ], + "type": "text", + "content": "We use the LibriTTS dataset [36] for audio-based VQ model training. The baselines such as Encodec [6], VOCos [24], and SpeechTokenizer [39] are based on residual vector quantization method. Our SimVQ model adopts the same architecture as WavTokenizer [13] with the only modification being the replacement of their EMA codebook with our one linear layer reparameterization method. We train SimVQ on LibriTTS-580h for 50 epochs with a batch size of 64. Note that WavTokenizer is trained with a 3-second window size for optimal performance, we train SimVQ using a 1-second window to accelerate training. For objective evaluation of the reconstructed audio, we follow VOCos [24] and employ metrics such as UTMOS [23], PESQ [20], STOI, and the F1 score for voiced/unvoiced classification (V/UV F1). UTMOS is particularly valuable as it produces scores highly correlated with human evaluations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 662, + 400, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 400, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 400, + 673 + ], + "type": "text", + "content": "5.2.2. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Tab. 2 presents the reconstruction performance of various VQ models on audio data. Baseline models using residual vector quantization perform significantly worse than" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22974" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 294, + 166 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 294, + 166 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 294, + 166 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 294, + 166 + ], + "type": "image", + "image_path": "9c4d5bb5c26c628314a7b0a9b5f415c277866b4fcfb803dbcbed7b2a897f0476.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "type": "text", + "content": "Figure 4. (a):(left) The rank of latent basis matrix " + }, + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "type": "text", + "content": " over training epochs. (b):(right) The Frobenius norm of latent basis matrix " + }, + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 173, + 295, + 207 + ], + "type": "text", + "content": " over training epochs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 228, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 295, + 350 + ], + "type": "text", + "content": "SimVQ, even when utilizing much larger bandwidths. Despite using the same architecture as WAVTokenizer, our model, which replaces the quantization layer with SimVQ, achieves superior performance with a 1-second window size and maintains nearly " + }, + { + "bbox": [ + 54, + 228, + 295, + 350 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 228, + 295, + 350 + ], + "type": "text", + "content": " codebook utilization when scaling up to a size of 262,144. The consistent performance of the SimVQ model across both image and audio data demonstrates that SimVQ is a general method for addressing the representation collapse problem in VQ models and can be effectively applied across multiple modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 118, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 118, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 118, + 370 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": "In Fig. 4(a), we plot the rank of the latent basis matrix over training epochs. Notably, SimVQ demonstrates the ability to adaptively adjust the rank of the latent space. Specifically, when the codebook size increases from " + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "inline_equation", + "content": "65k" + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "inline_equation", + "content": "262k" + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": ", the rank of the latent basis matrix decreases more rapidly and converges to a lower value. This observation suggests that a larger codebook can effectively alleviate the pressure on the latent space dimensionality, allowing the model to learn to represent data more efficiently. Additionally, despite the rank decreasing to a lower-rank space, SimVQ maintains " + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": " codebook utilization, highlighting its superiority over VQGAN-FC, which struggles when increasing the latent dimension from 8 to 128. We also calculate the Frobenius norm of the latent basis matrix, as shown in Fig. 4. The norm of a codebook size of " + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "inline_equation", + "content": "262k" + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": " is slightly larger than for " + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "inline_equation", + "content": "65k" + }, + { + "bbox": [ + 56, + 376, + 295, + 627 + ], + "type": "text", + "content": ", indicating that a larger codebook can span a broader area in the linear space. For a comprehensive evaluation, we also provide the reconstruction loss curve on the ImageNet validation dataset in Appendix 8.2. The results consistently show that SimVQ achieves improved performance, further validating the effectiveness of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 185, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 185, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 185, + 648 + ], + "type": "text", + "content": "5.4. Qualitative Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 715 + ], + "type": "text", + "content": "We visualize the distribution of encoder features and codebook embeddings in Fig. 5 and the frequency in Appendix 8.3. Compared to vanilla VQ models that most codebook vectors are not unused, SimVQ can update the whole codebook and distribute the features evenly in the whole space." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 71, + 530, + 171 + ], + "blocks": [ + { + "bbox": [ + 341, + 71, + 530, + 171 + ], + "lines": [ + { + "bbox": [ + 341, + 71, + 530, + 171 + ], + "spans": [ + { + "bbox": [ + 341, + 71, + 530, + 171 + ], + "type": "image", + "image_path": "da97486972720edf2e8b89b0011a519ffb75d5ce39f845826b6685e876e4a1f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 554, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 226 + ], + "type": "text", + "content": "Figure 5. Visualization of the divergence between the encoder features and codebook embeddings on a random subset of ImageNet validation dataset. The left figure is vanilla VQ model and the right one is SimVQ." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 247, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 297 + ], + "type": "text", + "content": "We qualitatively compare the reconstruction quality of both images and audio in Appendix 8.4. SimVQ can preserve more details with an enlarged codebook size, such as \"eyes\" and \"text\", which are challenging for vanilla VQ models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 384, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 384, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 384, + 320 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 328, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 496 + ], + "type": "text", + "content": "VQ Performance and Generative Models Many generative models [21, 25, 28, 34] utilize VQ models as a \"tokenizer\" to obtain discrete tokens. While it is intuitive that VQ reconstruction quality should impact generation performance, recent studies [35, 43] have revealed that the relationship is more nuanced: improved VQ reconstruction metrics do not necessarily translate to better generative outcomes. This complex relationship suggests that evaluating VQ models primarily through downstream generation tasks may not provide the most insightful assessment of their fundamental properties. Therefore, we focus on addressing the critical issue of representation collapse in VQ models, aiming to advance our understanding of their core representation learning mechanisms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 388, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 388, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 388, + 521 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 648 + ], + "type": "text", + "content": "In this paper, we explore the representation collapse problem in VQ models. We conduct a theoretical analysis of the optimization process in VQ models and propose a simple yet effective method, SimVQ, to address this issue. Our method addresses the representation collapse by jointly optimizing the latent space through a linear transformation with one linear layer. Experimental results demonstrate that SimVQ outperforms previous approaches on both image and audio datasets, highlighting its broad applicability across diverse machine learning tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 679, + 554, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 554, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 554, + 703 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China (Grant No.62276245)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 312, + 46, + 374, + 52 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 46, + 374, + 52 + ], + "spans": [ + { + "bbox": [ + 312, + 46, + 374, + 52 + ], + "type": "text", + "content": "Rank of W (Codebook Size = 65536)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22975" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Advances in Neural Information Processing Systems, pages 12449-12460. Curran Associates, Inc., 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 205, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 295, + 248 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 316 + ], + "type": "text", + "content": "[4] Zalan Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, and Neil Zeghidour. Audiolm: A language modeling approach to audio generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 31:2523-2533, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 318, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 296, + 428 + ], + "type": "text", + "content": "[5] Jake Bruce, Michael D Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Maria Elisabeth Bechtle, Feryal Behbahani, Stephanie C.Y. Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando De Freitas, Satinder Singh, and Tim Roktaschel. Genie: Generative interactive environments. In Proceedings of the 41st International Conference on Machine Learning, pages 4603-4623. PMLR, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 474 + ], + "type": "text", + "content": "[6] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. Transactions on Machine Learning Research, 2023. Featured Certification, Reproducibility Certification. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 520 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 296, + 575 + ], + "type": "text", + "content": "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 578, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 632 + ], + "type": "text", + "content": "[9] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 1, 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[10] Zhangyang Gao, Cheng Tan, Jue Wang, Yufei Huang, Lirong Wu, and Stan Z Li. Foldtoken: Learning protein language via vector quantization and beyond. arXiv preprint arXiv:2403.09673, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "text", + "content": "[11] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized net" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "works. In Proceedings of the 40th International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "text", + "content": "[12] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In 5th International Conference on Learning Representations, ICLR 2017, Toulouse, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[13] Shengpeng Ji, Ziyue Jiang, Xize Cheng, Yifu Chen, Minghui Fang, Jialong Zuo, Qian Yang, Ruiqi Li, Zhang Zhang, Xiaoda Yang, et al. Wavtokenizer: an efficient acoustic discrete codec tokenizer for audio language modeling. arXiv preprint arXiv:2408.16532, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 230 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 286 + ], + "type": "text", + "content": "[15] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "type": "text", + "content": "[16] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 334, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 334, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 334, + 553, + 409 + ], + "type": "text", + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[18] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "type": "text", + "content": "[19] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 578 + ], + "type": "text", + "content": "[20] A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra. Perceptual evaluation of speech quality (pesq)-a new method for speech quality assessment of telephone networks and CODECs. In 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), pages 749-752 vol.2, 2001. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 579, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 555, + 634 + ], + "type": "text", + "content": "[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 669 + ], + "type": "text", + "content": "[22] Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. ArXiv, abs/2204.02152, 2022. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22976" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Hubert Siuzdak. Vocos: Closing the gap between timedomain and fourier-based neural vocoders for high-quality audio synthesis. In The Twelfth International Conference on Learning Representations, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[25] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 238 + ], + "type": "text", + "content": "[26] Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi, Toshiyuki Kumakura, and Yuki Mitsu-fuji. SQ-VAE: Variational Bayes on discrete representation with self-annealed stochastic quantization. In Proceedings of the 39th International Conference on Machine Learning, pages 20987-21012. PMLR, 2022. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[27] Chaofan Tao, Qian Liu, Longxu Dou, Niklas Muennighoff, Zhongwei Wan, Ping Luo, Min Lin, and Ngai Wong. Scaling laws with vocabulary: Larger models deserve larger vocabularies. arXiv preprint arXiv:2407.13623, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "text", + "content": "[28] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 358 + ], + "type": "text", + "content": "[29] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "text", + "content": "[30] Tung-Long Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quantized wasserstein auto-encoder. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[31] Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[32] Pan Xiao, Peijie Qiu, and Aristeidis Sotiras. Scvae: Sparse coding-based variational autoencoder. ArXiv, abs/2303.16666, 2023. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 505, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 294, + 559 + ], + "type": "text", + "content": "[33] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In International Conference on Learning Representations, 2022. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 635 + ], + "type": "text", + "content": "[34] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 1, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 712 + ], + "type": "text", + "content": "[35] Lijun Yu, Jose Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, 2024. 1, 3, 6, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 551 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[36] Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J Weiss, Ye Jia, Zhifeng Chen, and Yonghui Wu. Libritts: A corpus derived from librispeech for text-to-speech. arXiv preprint arXiv:1904.02882, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 194 + ], + "type": "text", + "content": "[37] Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. SpeechGPT: Empowering large language models with intrinsic cross-modal conversational abilities. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757-15773, Singapore, 2023. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "type": "text", + "content": "[38] Jiahui Zhang, Fangneng Zhan, Christian Theobalt, and Shijian Lu. Regularized vector quantization for tokenized image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18467-18476, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 295 + ], + "type": "text", + "content": "[39] Xin Zhang, Dong Zhang, Shimin Li, Yaqian Zhou, and Xipeng Qiu. Speechtokenizer: Unified speech tokenizer for speech language models. In The Twelfth International Conference on Learning Representations, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 339 + ], + "type": "text", + "content": "[40] Chuanxia Zheng and Andrea Vedaldi. Online clustered codebook. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22798-22807, 2023. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "type": "text", + "content": "[41] Chuanxia Zheng, Guoxian Song, Tat-Jen Cham, Jianfei Cai, Dinh Q. Phung, and Linjie Luo. High-quality pluralistic image completion via code shared vqgan. ArXiv, abs/2204.01931, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 386, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 418 + ], + "type": "text", + "content": "[42] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of " + }, + { + "bbox": [ + 316, + 386, + 553, + 418 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 316, + 386, + 553, + 418 + ], + "type": "text", + "content": ". ArXiv, abs/2406.11837, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 420, + 553, + 474 + ], + "type": "text", + "content": "[43] Yongxin Zhu, Bocheng Li, Hang Zhang, Xin Li, Linli Xu, and Lidong Bing. Stabilize the latent space for image autoregressive modeling: A unified perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 551 + ], + "type": "text", + "content": "[44] Yongxin Zhu, Dan Su, Liqiang He, Linli Xu, and Dong Yu. Generative pre-trained speech language model with efficient hierarchical transformer. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1764-1775, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22977" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_content_list.json b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c6286843f833c9a9c7a528abb31fa1b1f178cb1 --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_content_list.json @@ -0,0 +1,2001 @@ +[ + { + "type": "text", + "text": "Addressing Text Embedding Leakage in Diffusion-based Image Editing", + "text_level": 1, + "bbox": [ + 138, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunung Mun $^{1*}$", + "bbox": [ + 181, + 178, + 303, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinhwan $\\mathbf{Nam}^{1*}$", + "bbox": [ + 341, + 179, + 468, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sunghyun Cho $^{1,2}$", + "bbox": [ + 509, + 179, + 648, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jungseul Ok $^{1,2\\dagger}$", + "bbox": [ + 687, + 178, + 813, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graduate School of AI, POSTECH $^{1}$ , Dept. of CSE, POSTECH $^{2}$", + "bbox": [ + 225, + 215, + 769, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mtablo, njh18, s.cho, jungseul}@postech.ac.kr", + "bbox": [ + 292, + 236, + 699, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60910e3dd9c8b845fd2853f1d6be7d6beb876e46911842a404c22415a9a2e636.jpg", + "image_caption": [ + "Figure 1. Qualitative comparison of tuning-free image-editing methods. Left to right: (1) source image, (2) visualization-only binary masks of the source objects and corresponding edit prompts (source $\\rightarrow$ target), and (3-7) results from our ALE and four baselines. The masks are not provided to any method; they are shown solely to indicate which region is supposed to change. Rows 1-2 illustrate single-object edits where baselines spill changes outside the intended area or distort geometry, while ALE keeps the background intact. Rows 3-4 demonstrate multi-object edits: baselines often entangle attributes, whereas ALE preserves each attribute in its designated region, yielding leakage-free results." + ], + "image_footnote": [], + "bbox": [ + 98, + 262, + 330, + 618 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8cd935d78ddff3f8895afa43d9e3de16c940a124ec485e48ea0fdfaf2d4f9cb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 262, + 450, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a149c77d6299793243285d290ba617087ad05412ab2b6e126f4377899589fa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 262, + 560, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2c843289f0eaad14bfc97a4c5e0933db27f8501b01b6fffb5bdc61b7c3d65d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 262, + 669, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee3987fe2eb806b14470c90f54376a9cc304a9884c665414278f767d053ddc73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 262, + 782, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b773ecdf6874bf78957ae9c2cb420262a064e8d7df7eaa69705fa86e67838f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 262, + 893, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b47de48d6739de8d20e99e33af746be8941ec253d16db254c1a5349976c0b3cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 348, + 450, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9bfdfcc28bd1f0625359194a210cc2ff88188b3278ca7e4b5ac52429527b7e70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 348, + 560, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/867b57c2ff9a84de45f2c68da0ec168bcca64b06d6444832852ccd7feee4db92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 348, + 669, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93b07de86845280ccab728c6a0325f0c487e975a936b17bf260fd2be071b79e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 348, + 782, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4eae2ac0aef7db039a611d58d886a98996013035796023eff86ca091ff1bee30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 348, + 893, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/95679ccc00ef3eda2c673e2acb736ebc619e982ac035d1eea352b30ebb20d027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 434, + 450, + 517 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49dd9e33c114f83a12cd5dbe62eed3d567ef2babd7fdd5b8a91253ba579b42a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 434, + 560, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75c671510375350284f108486addf98c25695c14344de8b687db6e2b44e52b46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 434, + 669, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37fe325af5d7ab827fb5119650b5aa736f29792f36fd332546dc1b075ecb1bc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 434, + 782, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b691138771d81c44e7947cda28f60500b07624f8298cde0e0850b8a2979773a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 434, + 893, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ba60764e9f199613546b9e4f364fcb734f72614041a5e60ee43a691cf6e853a.jpg", + "image_caption": [ + "ALE (Ours)" + ], + "image_footnote": [], + "bbox": [ + 338, + 518, + 450, + 604 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48f236a30788ca0a3474a36c8f8fc8637baeea1d6039a6cd542c707b31d88bad.jpg", + "image_caption": [ + "InfEdit" + ], + "image_footnote": [], + "bbox": [ + 450, + 518, + 560, + 604 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ff41c806ef77281748a17115b132da54b5a0c882f4b1a0064c6be3b92d75a4a.jpg", + "image_caption": [ + "MasaCtrl" + ], + "image_footnote": [], + "bbox": [ + 560, + 518, + 669, + 604 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00c1800cb5264654e45150856a706a7e8ef8c99fa11335f3bd2823f401284963.jpg", + "image_caption": [ + "FPE" + ], + "image_footnote": [], + "bbox": [ + 669, + 518, + 782, + 604 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef91215f5457b213be8af7fba68321759507a4479726d547772f0ff27e9b10a8.jpg", + "image_caption": [ + "P2P" + ], + "image_footnote": [], + "bbox": [ + 782, + 518, + 893, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 719, + 326, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based image editing, powered by generative diffusion models, lets users modify images through natural-language prompts and has dramatically simplified traditional workflows. Despite these advances, current methods still suffer from a critical problem: attribute leakage, where edits meant for specific objects unintentionally affect unrelated regions or other target objects. Our analy", + "bbox": [ + 86, + 753, + 482, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sis reveals the root cause as the semantic entanglement inherent in End-of-Sequence (EOS) embeddings generated by autoregressive text encoders, which indiscriminately aggregate attributes across prompts. To address this issue, we introduce Attribute-Leakage-free Editing (ALE), a framework that tackles attribute leakage at its source. ALE combines Object-Restricted Embeddings (ORE) to disentangle text embeddings, Region-Guided Blending for Cross-Attention Masking (RGB-CAM) for spatially precise attention, and Background Blending (BB) to preserve non-edited content. To quantitatively evaluate attribute leakage across various", + "bbox": [ + 511, + 720, + 906, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution; † Correspondence to jungseul@postech.ac.kr", + "bbox": [ + 112, + 875, + 470, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://mtablo.github.io/ALE_Edit_page/", + "bbox": [ + 112, + 887, + 401, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16451", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "editing methods, we propose the Attribute-Leakage Evaluation Benchmark (ALE-Bench), featuring comprehensive editing scenarios and new metrics. Extensive experiments show that ALE reduces attribute leakage by large margins, thereby enabling accurate, multi-object, text-driven image editing while faithfully preserving non-target content.", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 210, + 222, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-based image editing, where users modify existing images via natural language prompts, has emerged as a powerful alternative to traditional manual editing. Conventional editing workflows typically demand significant manual effort and domain expertise [1, 16, 25, 33]. In contrast, recent advancements leveraging generative diffusion models have substantially simplified the editing process [6, 12, 15, 20]. These models enable users to perform high-quality edits through intuitive textual prompts, making image editing more accessible and flexible.", + "bbox": [ + 89, + 234, + 482, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advances, existing text-based editing methods frequently suffer from a critical limitation: attribute leakage, where edits intended for specific objects inadvertently affect unrelated regions within the image. Attribute leakage can be categorized into two distinct types: Target-External Leakage (TEL), where attributes of a target object unintentionally affect non-target regions, and Target-Internal Leakage (TIL), where attributes intended for one target object inadvertently influence another target object within the same editing prompt. To mitigate this, recent studies attempt to spatially constrain editing effects by manipulating cross-attention maps [3, 29]. However, as illustrated in Figure 1, even state-of-the-art methods [3, 13, 29, 34] continue to exhibit significant TEL in single-object editing scenarios, and both TEL and substantial TIL in more complex, multi-object editing scenarios.", + "bbox": [ + 89, + 387, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our analysis highlights that attribute leakage fundamentally stems from overlooked issues of the entanglement of text embeddings, specifically those associated with the End-of-Sequence (EOS) tokens, building upon recent study [14]. Most text-based editing pipelines employ autoregressive text encoders such as CLIP [22], which append EOS tokens to text prompts until reaching a fixed length (e.g., 77 tokens) to generate embeddings. Consequently, EOS embeddings inherently aggregate information from all tokens within the prompt and attend indiscriminately across image regions via cross-attention layers, exacerbating both TEL and TIL. While recent studies [14] attempt to alleviate leakage stemming from EOS embedding entanglement, our analysis demonstrates that these approaches are insufficient to fully eliminate attribute leakage (see Section 2).", + "bbox": [ + 89, + 628, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel framework, Attribute-Leakage-Free Editing (ALE), primarily consisting of three complementary components: Object-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Restricted Embeddings (ORE), Region-Guided Blending for Cross-Attention Masking (RGB-CAM), and Background Blending (BB). ORE assigns distinct, semantically isolated embeddings to each object in the prompt, explicitly avoiding embedding entanglement. RGB-CAM enhances spatial precision in cross-attention maps by leveraging segmentation masks, restricting attention solely to intended regions. BB preserves the structural integrity of non-edited regions by integrating latents from the source image.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to systematically quantify attribute leakage, we introduce the Attribute-Leakage Evaluation Benchmark (ALE-Bench), a specialized benchmark designed explicitly for leakage evaluation in multi-object editing. Existing benchmarks predominantly focus on single-object scenarios, lacking metrics for evaluating attribute leakage comprehensively [17, 27, 31]. ALE-Bench covers a diverse range of editing scenarios, including multi-object editing in various edit types. We also propose two novel evaluation metrics: Target-External Leakage Score (TELS) and Target-Internal Leakage Score (TILS), explicitly quantifying TEL and TIL, respectively.", + "bbox": [ + 511, + 227, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are:", + "bbox": [ + 532, + 407, + 764, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Identifying the previously overlooked role of EOS embeddings as a fundamental cause of attribute leakage in text-based image editing (Section 2).", + "2. Proposing the novel Attribute-Leakage-Free Editing (ALE) framework, which specifically addresses leakage induced by EOS embeddings (Section 3).", + "3. Introducing a comprehensive benchmark, ALE-Bench, along with novel metrics (TELS, TILS) designed explicitly to quantify attribute leakage in multi-object editing scenarios (Section 4)." + ], + "bbox": [ + 513, + 422, + 903, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Attribute Leakage Problem and Analysis", + "text_level": 1, + "bbox": [ + 511, + 587, + 879, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we discuss the attribute leakage problem in multi-object text-based image editing and highlight limitations of existing methods. Specifically, Section 2.1 formally introduces multi-object text-based image editing and briefly describes dual-branch frameworks commonly used for such tasks. Section 2.2 defines attribute leakage and categorizes it into two types: TEL and TIL. Finally, Section 2.3 analyzes embedding entanglement, particularly from EOS tokens, as the primary cause of attribute leakage and explains why existing methods fail to adequately address this issue.", + "bbox": [ + 511, + 612, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multi-Object Text-based Image Editing", + "text_level": 1, + "bbox": [ + 511, + 772, + 852, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-based image editing modifies specific regions of a source image according to textual prompts describing desired changes. Formally, given a source image $x^{\\mathrm{src}}$ and a textual prompt pair $(y^{\\mathrm{src}}, y^{\\mathrm{tgt}})$ where $y^{\\mathrm{src}}$ specifies the source objects to be edited (i.e., objects to be replaced or modified), and $y^{\\mathrm{tgt}}$ specifies the target objects (i.e., new or modified objects to appear)—the goal is to generate an edited", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16452", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e44203661143ee2aa69180641c8be3891f8f29bdf686a09f4cdecc1e2c62fcb3.jpg", + "image_caption": [ + "(a) Source Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 222, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/689a93332ea52064fb0056f231127df8c8c546e7ad64dadd7ab3777616c78598.jpg", + "image_caption": [ + "(b) Leakage Example" + ], + "image_footnote": [], + "bbox": [ + 225, + 88, + 349, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c3a833f91689b324bdf29c3247fd13df24b15473787f059cd29cddaf8d8a8d6f.jpg", + "image_caption": [ + "(c) Reference Result", + "Figure 2. Illustration of attribute leakage in image editing. (a) The source image and editing prompts. (b) An editing result exhibiting attribute leakage. Regions highlighted in green indicate target-external leakage, where the editing spills into non-target (e.g., the green bell pepper turns into a golden pumpkin). Regions in blue show target-internal leakage, where the attributes of \"a red pumpkin\" undesirably influence the appearance of the \"golden apple\" region. (c) A reference image showing the desired editing result without attribute leakage." + ], + "image_footnote": [], + "bbox": [ + 352, + 88, + 477, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image $x^{\\mathrm{tgt}}$ . Ideally, $x^{\\mathrm{tgt}}$ reflects modifications described by $y^{\\mathrm{tgt}}$ exclusively within regions indicated by $y^{\\mathrm{src}}$ , leaving all other areas unchanged. When edits involve multiple objects simultaneously, the task is classified as multi-object editing.", + "bbox": [ + 89, + 349, + 482, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Practically, multi-object editing prompts can be decomposed into individual object-level prompt pairs \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^K\\), typically via language models or noun- chunk parsers [2, 11]. Our research specifically focuses on scenarios involving up to \\(K = 3\\) objects. For example, consider the editing scenario in Figure 2a. Given the prompt pair \\((y^{\\mathrm{src}}, y^{\\mathrm{tgt}}) = (\\text{\"a yellow bell pepper and a red bell pepper\", \"a red pumpkin and a golden apple\")}, it can be decomposed into object-level prompts \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^2 = [\"\\text{\"a yellow bell pepper\", \"a red pumpkin\", \"a red bell pepper\", \"a golden apple\"}]\\). The resulting image \\(x^{\\mathrm{tgt}}\\) should contain a red pumpkin and a golden apple, precisely aligned to corresponding regions.", + "bbox": [ + 89, + 412, + 482, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To perform such precise edits, diffusion-based editing methods encode textual prompts into embeddings that guide the editing process through cross-attention layers. These layers spatially align text-described attributes with corresponding regions in the image. Achieving accurate alignment, however, requires effectively preserving the original spatial structure while synthesizing novel visual attributes. To this end, recent methods utilize dual-branch editing frameworks, which have become popular due to their ability to simultaneously retain the original image layout and introduce new content [3, 19, 29].", + "bbox": [ + 88, + 611, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual-branch frameworks operate through two parallel pathways—a source branch and a target branch—using the same pretrained text-to-image diffusion model. Specifically, the source branch reconstructs the original image $x^{\\mathrm{src}}$ guided by the source textual prompt $y^{\\mathrm{src}}$ , thereby capturing structural and spatial information inherent in $x^{\\mathrm{src}}$ . The target branch, in parallel, synthesizes new visual attributes guided by the target textual prompt $y^{\\mathrm{tgt}}$ . Structural consis", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tency is maintained by injecting intermediate self-attention layer components—such as queries and keys—computed from the source branch into the corresponding self-attention layers of the target branch. Despite their strengths, dual-branch frameworks face fundamental challenges, particularly attribute leakage, caused by embedding entanglement involving EOS tokens, as analyzed in subsequent sections.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Attribute Leakage", + "text_level": 1, + "bbox": [ + 511, + 208, + 691, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribute leakage is a critical challenge in multi-object text-based image editing, characterized by unintended propagation of attributes from target objects to unrelated regions or other target objects. Formally, given a source image $x^{\\mathrm{src}}$ and object-level prompt pairs $[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K}$ , attribute leakage occurs when modifying an object from its original description $y_i^{\\mathrm{src}}$ to a new target description $y_i^{\\mathrm{tgt}}$ unintentionally impacts regions or objects not specified by the editing prompt. Attribute leakage can be categorized into two distinct types:", + "bbox": [ + 511, + 229, + 903, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Target-External Leakage (TEL): This occurs when editing an object specified by $(y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})$ unintentionally affects regions not described by the prompt pair (i.e., non-target regions). For example, as highlighted by the green region in Figure 2b, editing $y_{2}^{\\mathrm{src}} =$ \"a red bell pepper\" to $y_{2}^{\\mathrm{tgt}} =$ \"a golden apple\" inadvertently transforms an unrelated \"green bell pepper\" into a golden object.", + "bbox": [ + 513, + 367, + 903, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Target-Internal Leakage (TIL): This occurs when editing an object specified by $(y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})$ unintentionally affects another target object specified by a different prompt pair $(y_{j}^{\\mathrm{src}}, y_{j}^{\\mathrm{tgt}})$ , where $i \\neq j$ . For instance, as illustrated by the blue region in Figure 2b, editing $y_{1}^{\\mathrm{src}} = \\text{\"a yellow bell pepper\"}$ to $y_{1}^{\\mathrm{tgt}} = \\text{\"a red pumpkin\"}$ inadvertently impacts another target object described by $y_{2}^{\\mathrm{tgt}} = \\text{\"a golden apple\"}$ , causing it to appear as a mixture of red and golden pumpkin-like attributes.", + "bbox": [ + 513, + 472, + 905, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Effectively mitigating both TEL and TIL is essential to achieving precise, user-intended edits, as exemplified by the desired reference result in Figure 2c.", + "bbox": [ + 511, + 609, + 903, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Causes of Attribute Leakage", + "text_level": 1, + "bbox": [ + 511, + 666, + 767, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Attribute leakage primarily arises from embedding entanglement, which occurs during prompt encoding. Widely used text encoders, such as CLIP, encode tokens autoregressively, causing embeddings of later tokens to unintentionally accumulate mixed semantics from preceding attributes or objects. To mitigate this entanglement, some approaches introduce object-wise embeddings by parsing prompts into distinct noun-phrase spans and encoding each segment independently [10]. However, these methods only address entanglement among original tokens in the prompt, failing to adequately resolve entanglement involving EOS tokens.", + "bbox": [ + 511, + 688, + 905, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since CLIP pads prompts to a fixed length using EOS tokens, the EOS embeddings inevitably aggregate semantic information from multiple attributes and objects. For", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16453", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32c0e86db5a4ebbcff02a591ee601d2882c0599e4ad51960bca4acfad3d81952.jpg", + "image_caption": [ + "(a) Source image, edit prompt, and reference result image" + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 477, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e5011d5123124ed16453e35a6ef9654c1e4fa7e1336c8756f2e450cab919918.jpg", + "image_caption": [ + "\"a red diamond\"" + ], + "image_footnote": [], + "bbox": [ + 98, + 218, + 189, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/190c145d79dbe5dca7ebeef84eedb62209c51550f37e7f7783a537a06b09c67a.jpg", + "image_caption": [ + "\"agolden apple\"" + ], + "image_footnote": [], + "bbox": [ + 192, + 218, + 287, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e4a610d22d19eac4a021b572f9a494a26127dd6e13ef021b17e49fe4a8e10ddf.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 290, + 218, + 380, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/862b80c60bb589a05585c7286671791eb694f42b4af5b6e4c68187a425204928.jpg", + "image_caption": [ + "Edited Result" + ], + "image_footnote": [], + "bbox": [ + 383, + 218, + 475, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7293e174c80bdaa4ce9f8195a27448c7a7fca34c9fd726384962f36d64d9d8ec.jpg", + "image_caption": [ + "(b) Cross-attention maps and an edited result of ETS", + "\"a red diamond\"", + "(c) Cross-attention maps and an edited result of ETS + Masking", + "Figure 3. Illustration of attribute leakage from EOS embeddings and misaligned cross-attention. (b) ETS with object-wise embeddings shows spatial misalignment: embeddings of one object (e.g., \"a red diamond\") can influence unrelated regions. (c) Adding cross-attention masking correctly localizes object-wise embeddings but fails to suppress leakage from EOS embeddings (e.g., diamond-like decoration on the apple). Cross-attention maps are averaged across timesteps and summed over tokens (or padded EOS tokens)." + ], + "image_footnote": [], + "bbox": [ + 98, + 330, + 191, + 402 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19e9e0f25fefb322b6384149154d6ddf045a54693d41da150ed6273d00f33455.jpg", + "image_caption": [ + "\"a golden apple\"" + ], + "image_footnote": [], + "bbox": [ + 192, + 330, + 287, + 402 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0780115aa17c43cd02c821ada78ca615c66dbe38534c79a95d6dbc9e5c05281.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 290, + 330, + 380, + 402 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7dff0d0171ad472d7ad8b4434567e05be59d2e8406819c854dab1d70e677993e.jpg", + "image_caption": [ + "Edited Result" + ], + "image_footnote": [], + "bbox": [ + 383, + 330, + 475, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instance, when encoding the prompt \"a red diamond and a golden apple\", the EOS embeddings inherently encapsulate combined semantics from all attributes and objects (e.g., \"red\", \"diamond\", \"golden\", and \"apple\"). To address this issue, End-Token-Substitution (ETS) [14] replaces attribute-rich EOS embeddings with attribute-free embeddings obtained from prompts without descriptors (e.g., \"a diamond and an apple\"). Nevertheless, ETS remains insufficient even when combined with object-wise embeddings, as the simplified EOS embeddings still aggregate semantic information across multiple objects, thereby continuing to propagate attribute leakage (see Figure 3).", + "bbox": [ + 89, + 582, + 482, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Embedding entanglement further exacerbates spatial inaccuracies in dual-branch text-based image editing frameworks. Since dual-branch methods inject structural information from the source image into the target editing branch, entangled embeddings can produce misaligned cross-attention maps, causing embeddings to incorrectly attend to visually similar but semantically incorrect regions during editing (Figure 3b). This visual-semantic confusion significantly exacerbates spatial inaccuracies, intensifying", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93944ffe58acff755f597afe7531939e0543abe600c9656ec9b003002004a1dc.jpg", + "image_caption": [ + "(a) Source image", + "Figure 4. Comparison of different EOS embedding modification strategies in a single editing scenario, where \"yellow bell pepper\" is edited to \"diamond\" and \"red bell pepper\" to \"moon\". (b) Uses the original EOS embeddings. (c) Replaces EOS embeddings with zero vectors. (d) Replaces EOS embeddings with those obtained from an empty prompt \"." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 612, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1ebd3c6b680ff5cadc1301f594ff61d3dff9667e00a26d4b18f92be9cf45c91b.jpg", + "image_caption": [ + "(b) Original EOS" + ], + "image_footnote": [], + "bbox": [ + 614, + 88, + 709, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d182492bf06aa5a5a96b20ad99124cb1c8702b34f37ebb8daef509779bedc82e.jpg", + "image_caption": [ + "(c) EOS $\\leftarrow$ 0" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 805, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e32a98110f1ec9ac7d624b454c8b659c636d1da60c56f4c276994a1442466bca.jpg", + "image_caption": [ + "(d) EOS $\\leftarrow$" + ], + "image_footnote": [], + "bbox": [ + 807, + 88, + 901, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attribute leakage. Although existing methods attempt to mitigate this issue through refined cross-attention alignment or explicit masking [3, 29], these strategies fail to address the fundamental problem: EOS embeddings inherently lack spatial specificity, as they integrate semantic content from the entire prompt. Therefore, restricting the spatial attention of EOS embeddings to specific regions is inherently ineffective, further compounding attribute leakage even when employing combined strategies (Figure 3c).", + "bbox": [ + 511, + 281, + 905, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One naive alternative could involve removing semantic content entirely from EOS embeddings by substituting them with zero vectors or embeddings derived from empty prompts. However, as demonstrated empirically in Figure 4 and detailed in Appendix E, this simplistic solution severely degrades visual quality and editing accuracy. These observations imply that diffusion models intrinsically depend on EOS embeddings containing semantics to achieve high-quality image editing outcomes. Therefore, resolving attribute leakage effectively requires a dedicated strategy that carefully mitigates the unintended influence of entangled EOS embeddings without completely eliminating or overly simplifying their semantic content. We propose such a targeted approach in the next section.", + "bbox": [ + 511, + 417, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Attribute-Leakage-Free Editing (ALE)", + "text_level": 1, + "bbox": [ + 511, + 643, + 862, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a source image $x^{\\mathrm{src}}$ and a list of $K$ object-level prompt pairs $[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K}$ , our goal is to generate an edited image $x^{\\mathrm{tgt}}$ that (1) replaces every $y_i^{\\mathrm{src}}$ with $y_i^{\\mathrm{tgt}}$ inside its designated region, (2) preserves all non-target content, and (3) avoids both TIL and TEL.", + "bbox": [ + 511, + 669, + 905, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed method, ALE, is built on the dual-branch editing framework with the Denoising Diffusion Consistent Model (DDCM) virtual inversion scheme: a source branch reconstructs $x^{\\mathrm{src}}$ from latent $\\{z_{\\tau}^{\\mathrm{src}}\\}_{\\tau = T}^{0}$ under the prompt $y_{\\mathrm{base}}^{\\mathrm{src}} =$ \" and \".join $\\left(\\left[y_i^{\\mathrm{src}}\\right]\\right)$ , while a target branch denoises latent $\\{z_{\\tau}^{\\mathrm{tgt}}\\}_{\\tau = T}^{0}$ toward the edited image guided by $y_{\\mathrm{base}}^{\\mathrm{tgt}} =$ \" and \".join $\\left(\\left[y_i^{\\mathrm{tgt}}\\right]\\right)$ . On top of this backbone we introduce three key components: ORE, RGB-CAM, and BB. Figure 5 visualizes an overall pipeline, and Algorithm 1 lists the complete procedure.", + "bbox": [ + 511, + 747, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16454", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c5e06dc87c55347db03588732e8ca5f77ffea4767f5b53fcb6092e35c13ec485.jpg", + "image_caption": [ + "Figure 5. Overview of ALE. The framework consists of two branches: the upper branch (source branch) processes the source latent $z_{\\tau}^{\\mathrm{src}}$ , and the lower branch (target branch) processes the target latent $z_{\\tau}^{\\mathrm{tgt}}$ at each timestep $\\tau$ . ORE produces multiple semantically-isolated embeddings for each target object to reduce interference between unrelated objects. RGB-CAM refines cross-attention activations using segmentation masks, aligning the attention of each embedding to spatial regions corresponding to its target object. BB merges the source latent for background regions and the target latent for edited regions. Dashed lines indicate omitted components (e.g., the encoder, the decoder, and the segmentation model) for simplicity." + ], + "image_footnote": [], + "bbox": [ + 148, + 88, + 861, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. DDCM and Dual-Branch Editing Framework", + "text_level": 1, + "bbox": [ + 89, + 345, + 478, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Virtual inversion via DDCM We adopt the Denoising Diffusion Consistent Model (DDCM) [29], which chooses the special variance schedule so that any noisy latent $z_{\\tau}$ retains a closed-form link to its clean latent $z_0$ at every timestep $\\tau \\in [T, \\dots, 0]$ . This property enables a virtual inversion that avoids the costly DDIM / null-text inversion. Because this form matches the multi-step consistency sampler of Latent Consistency Models [8], we can edit images in as few as 4-20 steps without explicit inversion steps.", + "bbox": [ + 89, + 369, + 483, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dual-branch framework Starting from an initial noise, the source branch follows DDCM updates with $y_{\\mathrm{base}}^{\\mathrm{src}}$ , while the target branch starts with $z_{\\tau}^{\\mathrm{tgt}} = z_{\\tau}^{\\mathrm{src}}$ and calculates the next latent $z_{\\tau - 1}^{\\mathrm{tgt}}$ at each step using the U-Net noise $\\varepsilon_{\\theta}(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}})$ and the consistency correction term. To preserve structure of $x^{\\mathrm{src}}$ , we copy the query-key tensors $(Q, K)_{\\ell, \\tau}^{\\mathrm{src}}$ from every self-attention layer $\\ell$ of the source branch into the corresponding layer of the target branch according to self-attention injection schedule $S = \\{\\tau | T_S \\leq \\tau \\leq T\\}$ , where $T_S \\in [0, T]$ controls the length of the self-attention injection schedule. Thus, a shorter schedule $(T_s \\approx T)$ touches only early denoising steps and enables stronger edits, whereas a longer schedule $(T_s \\approx 0)$ enforces stricter structural preservation.", + "bbox": [ + 89, + 511, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Object-Restricted Embeddings (ORE)", + "text_level": 1, + "bbox": [ + 89, + 736, + 419, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address leakages in prompt embedding level, ORE encodes each object prompt $y_{i}^{\\mathrm{tgt}}$ in isolation, yielding a set of token-embedding matrices", + "bbox": [ + 89, + 758, + 483, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {i} ^ {\\prime} = \\left[ e _ {\\text {B O S}}, \\underbrace {e _ {\\text {t o k e n} _ {1}} , \\cdots} _ {\\text {f r o m t o k e n s i n y} _ {i} ^ {\\mathrm {t g t}}}, \\underbrace {e _ {\\text {E O S}} , \\cdots} _ {\\text {f r o m p a d d e d E O S t o k e n s}} \\right] \\in \\mathbb {R} ^ {L \\times d}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 816, + 472, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ is the padded prompt length and $d$ the embedding dimension. For example, in Figure 3 case, $E_1'$ is obtained", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as $[e_{\\mathrm{BOS}}, e_{\\mathrm{a}}, e_{\\mathrm{red}}, e_{\\mathrm{diamond}}, e_{\\mathrm{EOS}}, \\ldots]$ . And we construct a base embedding $E_{\\mathrm{base}}'$ , by encoding $y_{\\mathrm{base}}^{\\mathrm{tgt}}$ and splicing $E_i'[y_i^{\\mathrm{tgt}}]$ back into their original spans, to calculate base value tensor in RGB-CAM. Because no token embedding in $\\{E_i'\\}$ can influence another object's span and EOS embeddings in $E_i'$ only contain semantics of $y_i^{\\mathrm{tgt}}$ , subsequent cross-attention receives semantically disentangled embeddings, thereby preventing leakage at its source.", + "bbox": [ + 511, + 345, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Region-Guided Blending for Cross-Attention Masking (RGB-CAM)", + "text_level": 1, + "bbox": [ + 511, + 476, + 906, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Standard cross-attention layers in diffusion U-Net accept a single value tensor $V$ and thus cannot exploit multiple OREs. RGB-CAM replaces the vanilla cross-attention output with a spatially blended tensor", + "bbox": [ + 511, + 513, + 906, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\sum_ {i = 1} ^ {K} (M \\odot m _ {i}) V _ {i} + (M \\odot m _ {\\text {b a c k}}) V _ {\\text {b a s e}},\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 583, + 867, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M =$ attention\\_map(Q, K) is the base cross attention map, $V_{i} = W_{v}(E_{i}^{\\prime})$ , $K = W_{k}(\\text{Encoder}_{\\text{text}}(y_{\\text{base}}^{\\text{tgt}}))$ , $V_{\\text{base}} = W_{v}(E_{\\text{base}}^{\\prime})$ , and $\\{m_i\\}$ , $m_{\\text{back}}$ are object and background segmentation masks from Grounded-SAM [24]. Since masks are not pixel-perfect, we apply a slight dilation. The masked tensors $(M \\odot m_i)V_i$ localize each ORE to its designated region, eliminating target-internal leakage, while the background term preserves areas outside all masks. Note that only when ORE and RGB-CAM operate in tandem does ALE produce leakage-free results.", + "bbox": [ + 511, + 633, + 906, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Background Blending (BB)", + "text_level": 1, + "bbox": [ + 511, + 792, + 759, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Even with perfect cross-attention, backgrounds remain weakly constrained because $\\{y_i^{\\mathrm{tgt}}\\}$ mention only target objects. At every timestep $\\tau$ we blend the source latent by the background mask as a final step:", + "bbox": [ + 511, + 815, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {\\tau} ^ {\\mathrm {l g t}} = m _ {\\text {b a c k}} \\odot z _ {\\tau} ^ {\\mathrm {s r c}} + (1 - m _ {\\text {b a c k}}) \\odot z _ {\\tau} ^ {\\mathrm {t g t}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 883, + 852, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16455", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Attribute-Leakage-Free Editing" + ], + "code_body": "Input: Source image $x^{\\mathrm{src}}$ , object-level prompt pairs $\\left[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})\\right]_{i=1}^{K}$ , self-attention injection schedule $\\mathcal{S} = \\{\\tau \\mid 0 \\leq \\tau \\leq T_S\\}$", + "bbox": [ + 109, + 112, + 450, + 159 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Output: Edited image $x^{\\mathrm{tgt}}$", + "bbox": [ + 112, + 160, + 287, + 175 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Pre-processing" + ], + "code_body": "Form concatenated base prompts $y_{\\mathrm{base}}^{\\mathrm{src}}, y_{\\mathrm{base}}^{\\mathrm{tgt}}$ \nEncode object-restricted embeddings $\\{E_i'\\}_{i=1}^K$ and the base embedding $E_{\\mathrm{base}}'$ (ORE); \nObtain object masks $\\{m_i\\}_{i=1}^K$ and background mask $m_{\\mathrm{back}}$ with Grounded-SAM; \nInitialization \nSample initial noise $z_T^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)$ ; \nSet $z_T^{\\mathrm{tgt}} \\gets z_T^{\\mathrm{src}}$ ; \nfor $\\tau = T$ to 1 do \n// Source branch \nPredict noise $\\hat{\\varepsilon}_{\\tau}^{\\mathrm{src}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{src}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{src}}\\right)$ ; \nUpdate $z_{\\tau-1}^{\\mathrm{src}}$ with DDCM sampling; \n// Target branch \nif $\\tau \\in S$ then copy self-attention $Q, K$ tensors from the source branch; \nPredict noise $\\hat{\\varepsilon}_{\\tau}^{\\mathrm{tgt}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}}\\right)$ ; RGB-CAM[\\{E_i', m_i\\}, E'_base, $m_{\\mathrm{back}}$ ] \nUpdate $z_{\\tau-1}^{\\mathrm{tgt}}$ with DDCM sampling; \n// Background blending (BB) \n $z_{\\tau-1}^{\\mathrm{tgt}} \\gets m_{\\mathrm{back}} \\odot z_{\\tau-1}^{\\mathrm{src}} + (1 - m_{\\mathrm{back}}) \\odot z_{\\tau-1}^{\\mathrm{tgt}}$ ; \nend \n $x^{\\mathrm{tgt}} \\gets \\text{Decoder}(z_0^{\\mathrm{tgt}})$ ; \nreturn $x^{tgt}$", + "bbox": [ + 109, + 196, + 472, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BB guarantees preservation of non-edited regions, suppressing TEL without expensive threshold tuning required by prior local-blending heuristics [13].", + "bbox": [ + 89, + 606, + 482, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 667, + 223, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 694, + 266, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ALE-Bench construction Prior benchmarks for text-guided image editing [7, 17] focus on the visual quality of the edited result itself and neglect attribute leakage. Although they measure background preservation metrics that are similar to TELS, they overlook TIL. This makes it difficult to analyze how well a method achieves precise editing. To fill this gap, we introduce Attribute-Leakage-Evaluation Benchmark (ALE-Bench), a dedicated testbed for attribute leakage evaluation. ALE-Bench systematically varies the number of objects to be edited and the editing type, grouped into five categories: (1) color, (2) object, (3) material, (4) color + object, and (5) object + material. Each source im", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age is paired with multiple prompts for each edit type, enabling fine-grained analysis across diverse scenarios. Figure 6 shows examples of ALE-Bench. Full construction details and dataset statistics are provided in Appendix B.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics We evaluate image-editing performance using the following metrics:", + "bbox": [ + 511, + 154, + 903, + 185 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Structure Distance [26] quantifies how well the edited image $x^{\\mathrm{tgt}}$ preserves the spatial layout of $x^{\\mathrm{src}}$ . A lower score indicates better structural consistency.", + "- Editing Performance is measured by the cosine similarity between the CLIP embeddings of $x^{\\mathrm{tgt}}$ and prompt $y^{\\mathrm{tgt}}$ . Higher similarity reflects more faithful edits.", + "- Background Preservation is assessed on the non-edited regions using PSNR, SSIM [28], LPIPS [32], and MSE between $x^{\\mathrm{tgt}}$ and $x^{\\mathrm{src}}$ . Higher PSNR/SSIM and lower LPIPS/MSE signify better preservation." + ], + "bbox": [ + 513, + 185, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we introduce new metrics:", + "bbox": [ + 531, + 337, + 799, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Target-Internal Leakage Score (TILS) measures unintended modifications inside other target-object regions as follows:", + "bbox": [ + 513, + 353, + 903, + 396 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T I L S} = \\frac {1}{K (K - 1)} \\sum_ {i \\neq j} ^ {K} \\operatorname {C L I P} \\left(x _ {\\mathrm {t g t}} \\odot m _ {j}, y _ {i} ^ {\\mathrm {t g t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 407, + 872, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where CLIP represents the CLIP similarity score, $K$ is the number of objects to be edited, $x^{\\mathrm{tgt}}$ is the edited image, $m_j$ is the $j$ -th object mask, and $y_i^{\\mathrm{tgt}}$ is the target prompt for $i$ -th object. A lower TILS implies that, as the user intended, the target objects did not affect each other.", + "bbox": [ + 524, + 462, + 903, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Target-External Leakage Score (TELS) measures unintended changes in the background (non-edited regions) as follows:", + "bbox": [ + 513, + 537, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T E L S} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathrm {C L I P} \\left(x ^ {\\mathrm {t g t}} \\odot \\left(\\mathbf {1} - \\bigcup_ {j = 1} ^ {K} m _ {j}\\right), y _ {i} ^ {\\mathrm {t g t}}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 592, + 895, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The mean CLIP scores between the background and each target prompt are computed for multiple object edits. A lower TELS indicates minimal TEL, which ensures that the outside of the targets remains unchanged.", + "bbox": [ + 524, + 652, + 906, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines For comparison, we selected tuning-free image editing methods including Prompt-to-Prompt (P2P) [13], MasaCtrl [3], Free-Prompt-Editing (FPE) [34], and InfEdit [29] (see Appendix A, C for further details).", + "bbox": [ + 511, + 715, + 903, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 787, + 651, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ALE outperforms existing methods in both mitigating attribute leakage and producing high-quality edits, as seen in Table 1 and Figure 7. In particular, ALE achieves the lowest TELS and TILS, reflecting its ability to precisely apply attributes solely to the designated target regions. Across different numbers of editing objects (Table 2) and different", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16456", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1854bd306a3086f1f0f6767ba2ddcccc916104e2308284167bfc554c6894bf8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Source ImageObject MasksColorObjectMaterialColor + ObjectObject + Material
yoga mat → pink-colored yoga matman → alienfootwear → footwear made of amberfootwear → orange-colored sandalsman → woman made of bronze
cup → white-colored cup, steam → green-colored steambook → cushion, cup → tumblerbook → book made of chrome, cup → cup made of papersteam → gray-colored smoke, cup → blue-colored jarbook → journal made of gold, cup → mug made of amethyst
wolf → pink-colored wolf, moon → purple-colored moon, mountain → gray-colored mountainwolf → eagle, moon → soccer ball, mountain → glacierwolf → wolf made of velvet, moon → moon made of steel, mountain → mountain made of leatherwolf → crimson-colored dragon, moon → navy-colored soccer ball, mountain → brown-colored gardenwolf → dog made of linen, moon → balloon made of paper, mountain → glacier made of amber
", + "bbox": [ + 93, + 89, + 906, + 344 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fa94a571a6825a3b3fbf8a6a430a25f809d4df85c0ec0ed4ae089569023a24b.jpg", + "table_caption": [ + "Figure 6. Examples of ALE-Bench. Source images are shown with object binary masks. Edit types are color, object, material modifications, and their combinations. Rows depict 1-, 2-, and 3-object edits, respectively." + ], + "table_footnote": [], + "table_body": "
Source ImageColorObjectMaterialColor + ObjectObject + Material
tree→ blue-colored treepine treetree made of diamondyellow-colored Christmas treepine tree made of bamboo
couch→ purple-colored couchbenchcouch made of leatherkhaki-colored deskdesk made of chrome
floor→ khaki-colored floorlawnfloor made of turquoiseyellow-colored tiletile made of jade
", + "bbox": [ + 91, + 381, + 903, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7. Qualitative examples of ALE applied to three source objects (tree, couch, and floor) under different editing types. The bottom rows show the corresponding prompts describing each transformation.", + "bbox": [ + 89, + 566, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "editing types (Table 3), ALE demonstrates its robust performance. Figure 7 illustrates qualitative examples of ALE for various editing types. These results indicate that our method can effectively address both TEL and TIL. Detailed results are in Appendix D. Furthermore, the quantitative and qualitative results on PIE-Bench are in Appendix E.", + "bbox": [ + 89, + 609, + 483, + 702 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8aeb21f30d8130acb860a8f2a3596564895f0e53be614c0a181014a073ff0378.jpg", + "image_caption": [ + "(a) Source image" + ], + "image_footnote": [], + "bbox": [ + 93, + 723, + 187, + 797 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5a8e8c1c91f26e2029738dd7b93dc757d1fdc87b653999d5378a3e9fd50f60a.jpg", + "image_caption": [ + "(b) BB" + ], + "image_footnote": [], + "bbox": [ + 189, + 723, + 285, + 797 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a95e448b9c30e8e105ed2cb2414026fc3630be3f72b61368822421f1b772b55.jpg", + "image_caption": [ + "(c) $\\mathrm{ORE} + \\mathrm{RGB}$", + "Figure 8. Qualitative ablation results for the editing: yellow bell pepper $\\rightarrow$ diamond, red bell pepper $\\rightarrow$ moon. (b) BB resolves TEL but not TIL. (c) ORE + RGB-CAM reduces TIL but not TEL. (d) ALE (BB + ORE + RGB-CAM) shows no TIL and TEL." + ], + "image_footnote": [], + "bbox": [ + 285, + 723, + 380, + 797 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14f3931906e9b8449240c78b1047e456bbc0995370a0de8a5c9f87bab07bdb1e.jpg", + "image_caption": [ + "(d) ALE" + ], + "image_footnote": [], + "bbox": [ + 383, + 723, + 478, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 609, + 679, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present ablation results on BB, ORE, and RGB-CAM in Figure 8. Each module plays a complementary role in preventing attribute leakage. Using only BB (Figure 8b) helps preserve background regions and suppresses TEL, but fails to prevent TIL. On the other hand, $\\mathrm{ORE} + \\mathrm{RGB}$ -CAM (Figure 8c) reduces TIL by disentangling text embeddings and aligned cross-attention. However, without BB, it cannot preserve the original background, and TEL remains. Only when all three modules—ORE, RGB-CAM, and BB—are used together (Figure 8d), ALE achieves precise and leakage-free editing. Qualitative results and more examples are provided in Appendix E.", + "bbox": [ + 511, + 632, + 906, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 513, + 828, + 635, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While ALE-Bench provides a focused framework for evaluating attribute leakage, and ALE achieves strong performance on rigid attribute edits, both are currently lim", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16457", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/22a9acea6edced92cf72efb101418b43b7e1c6991c1a3aecabe735b6a226aa17.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
P2P21.5217.260.151420.6711.150.44950.08790.5589
MasaCtrl20.1816.740.092920.0114.990.29300.04180.7346
FPE21.0717.380.116421.8912.820.39030.06560.6052
InfEdit19.5916.690.048421.7816.740.20340.03400.7709
ALE16.0315.280.016722.2030.040.03610.00140.9228
", + "bbox": [ + 91, + 88, + 903, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a546bc681b2a127d94cc00a14335a2a232f2d0b31de27de3955f3a111155c65f.jpg", + "table_caption": [ + "Table 1. Comparison of editing performances on ALE-Bench. ALE demonstrates the lowest attribute leakage, highest structure preservation, and superior editing performance, indicating a more precise and controlled editing." + ], + "table_footnote": [], + "table_body": "
# of Editing ObjectsTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
116.41-0.0087622.6230.010.04050.001670.9049
216.0015.420.0164822.0630.060.03600.001460.9235
315.8915.360.0246022.1930.010.03230.001540.9426
", + "bbox": [ + 91, + 252, + 903, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/952084cb4b842e1894adf37255429ef3e0b73b9c563896f6092ec4c148a76de8.jpg", + "table_caption": [ + "Table 2. Performance of ALE on ALE-Bench based on the number of objects edited. Our method maintains low attribute leakage and strong background preservation even as the number of editing objects increases." + ], + "table_footnote": [], + "table_body": "
Editing TypeTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
Color17.6316.210.0089023.1232.970.02880.000790.9309
Material17.1515.960.0117922.9430.630.03390.001200.9248
Object15.8616.250.0197421.8229.030.03860.001820.9218
Color+Object15.3014.010.0230622.1528.600.04070.002050.9206
Object+Material14.5514.510.0195621.4228.880.03930.001910.9201
", + "bbox": [ + 91, + 385, + 903, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Performance of ALE on ALE-Bench across various editing types (color, object, material, and combinations). The results show consistent low attribute leakage and high editing performance.", + "bbox": [ + 89, + 501, + 906, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ited to local and relatively simple transformations—such as changes in color, object identity, or material. They do not support or evaluate non-rigid transformations like style transfer, pose changes, or adding/deleting objects, where defining and detecting attribute leakage becomes ambiguous. This focus reflects our aim to establish a clear and measurable foundation before addressing more complex editing scenarios. Although the benchmark offers 3,000 detailed editing scenarios across 20 carefully curated images, the small image set may limit how well results generalize to models trained on larger or more diverse datasets. Future work could expand both the editing model and benchmark to support richer transformations and broader datasets. More limitations are discussed in Appendix F.", + "bbox": [ + 88, + 556, + 482, + 770 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 782, + 209, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we addressed the issue of attribute leakage in diffusion-based image editing, focusing on two types of leakage: Target-External Leakage (TEL)—unintended edits in non-target regions—and Target-Internal Leakage (TIL)—interference among attributes of different targets. To mitigate these problems, we introduced Attribute", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Leakage-Free Editing (ALE), a tuning-free framework that combines three key components: Object-Restricted Embeddings (ORE), which localize attribute semantics to each target object embedding; Region-Guided Blending for Cross-Attention Masking (RGB-CAM), which constrains cross-attention with segmentation masks to avoid unintended inter-object attribute mixing; and Background Blending (BB), which preserves the source image in backgrounds.", + "bbox": [ + 511, + 556, + 906, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also presented ALE-Bench, a dedicated benchmark for rigorously evaluating attribute leakage across diverse multi-object editing scenarios. ALE-Bench introduces new quantitative metrics—TELS and TILS—that effectively quantify unintended modifications, providing comprehensive measures to assess editing fidelity and consistency.", + "bbox": [ + 511, + 678, + 908, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our extensive experimental validation demonstrated that ALE significantly outperforms existing tuning-free editing methods, achieving state-of-the-art performance by effectively minimizing attribute leakage while maintaining high editing quality and structural consistency. By effectively addressing attribute leakage with minimal computational overhead, ALE enhances the reliability and precision of multi-object image editing tasks.", + "bbox": [ + 511, + 768, + 910, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16458", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) and IITP-ITRC (Information Technology Research Center) grant funded by the Korea government (MSIT) (No.RS-2019-II191906, Artificial Intelligence Graduate School Program (POSTECH); No.RS-2021-II212068, Artificial Intelligence Innovation Hub; IITP-2025-00437866; RS-2024-00509258, Global AI Frontier Lab) and partly supported by Seoul R&BD Program through the Seoul Business Agency (SBA) funded by The Seoul Metropolitan Government (SP240008).", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 294, + 187, + 310 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3, 4, 6, 11", + "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1-10, 2023. 11", + "[5] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7430-7440, 2023. 11", + "[6] Adham Elarabawy, Harish Kamath, and Samuel Denton. Direct inversion: Optimization-free text-driven real image editing with diffusion models. arXiv preprint arXiv:2211.07825, 2022. 2", + "[7] Chakrabarty et al. Lomoe: Localized multi-object editing via multi-diffusion. In Proc. ACM Multimedia, 2024. 6, 12", + "[8] Luo et al. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv, 2023. 5", + "[9] Yu et al. Zero-shot referring image segmentation with global-local context features. In Proc. CVPR, 2023. 11", + "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 3, 11", + "[11] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan," + ], + "bbox": [ + 93, + 319, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3", + "[12] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 2", + "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6, 11", + "[14] Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Shahbaz Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, and Yaxing Wang. Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems, 37:137646-137672, 2024. 2, 4, 11, 13, 14", + "[15] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[17] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6, 13", + "[18] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 11", + "[19] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817–7826, 2024. 3, 11", + "[20] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2", + "[21] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 11", + "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[23] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16459", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "through attention map alignment. Advances in Neural Information Processing Systems, 36:3536-3559, 2023. 11", + "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kun-chang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 5", + "[25] Patsorn Sangkloy, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Scribbler: Controlling deep image synthesis with sketch and color. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5400-5409, 2017. 2", + "[26] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 6", + "[27] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2", + "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[29] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 2, 3, 4, 5, 6, 11", + "[30] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 11", + "[31] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 2", + "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[33] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In Computer vision-ECCV 2016: 14th European conference, Amsterdam, the netherlands, October 11-14, 2016, proceedings, part v 14, pages 597-613. Springer, 2016. 2", + "[34] Siyu Zou, Jiji Tang, Yiyi Zhou, Jing He, Chaoyi Zhao, Rongsheng Zhang, Zhipeng Hu, and Xiaoshuai Sun. Towards efficient diffusion-based image editing with instant attention masks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7864-7872, 2024. 2, 6" + ], + "bbox": [ + 91, + 92, + 482, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16460", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_model.json b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23433fd068b81a386c1b6c841e58b2f33adfe3f5 --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_model.json @@ -0,0 +1,2662 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "Addressing Text Embedding Leakage in Diffusion-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.179, + 0.304, + 0.2 + ], + "angle": 0, + "content": "Sunung Mun\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.18, + 0.47, + 0.198 + ], + "angle": 0, + "content": "Jinhwan \\(\\mathbf{Nam}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.18, + 0.65, + 0.199 + ], + "angle": 0, + "content": "Sunghyun Cho\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.688, + 0.179, + 0.815, + 0.199 + ], + "angle": 0, + "content": "Jungseul Ok\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.216, + 0.771, + 0.235 + ], + "angle": 0, + "content": "Graduate School of AI, POSTECH\\(^{1}\\), Dept. of CSE, POSTECH\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.237, + 0.7, + 0.253 + ], + "angle": 0, + "content": "{mtablo, njh18, s.cho, jungseul}@postech.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.263, + 0.331, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.263, + 0.451, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.263, + 0.561, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.263, + 0.671, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.263, + 0.783, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.263, + 0.895, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.349, + 0.451, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.349, + 0.561, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.349, + 0.671, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.349, + 0.783, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.349, + 0.894, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.435, + 0.451, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.435, + 0.561, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.435, + 0.671, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.435, + 0.783, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.435, + 0.894, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.519, + 0.451, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.519, + 0.561, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.519, + 0.671, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.519, + 0.783, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.519, + 0.894, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.607, + 0.424, + 0.617 + ], + "angle": 0, + "content": "ALE (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.607, + 0.524, + 0.617 + ], + "angle": 0, + "content": "InfEdit" + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.607, + 0.639, + 0.617 + ], + "angle": 0, + "content": "MasaCtrl" + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.607, + 0.739, + 0.617 + ], + "angle": 0, + "content": "FPE" + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.607, + 0.849, + 0.617 + ], + "angle": 0, + "content": "P2P" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.624, + 0.907, + 0.707 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparison of tuning-free image-editing methods. Left to right: (1) source image, (2) visualization-only binary masks of the source objects and corresponding edit prompts (source \\(\\rightarrow\\) target), and (3-7) results from our ALE and four baselines. The masks are not provided to any method; they are shown solely to indicate which region is supposed to change. Rows 1-2 illustrate single-object edits where baselines spill changes outside the intended area or distort geometry, while ALE keeps the background intact. Rows 3-4 demonstrate multi-object edits: baselines often entangle attributes, whereas ALE preserves each attribute in its designated region, yielding leakage-free results." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.72, + 0.327, + 0.736 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.755, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Text-based image editing, powered by generative diffusion models, lets users modify images through natural-language prompts and has dramatically simplified traditional workflows. Despite these advances, current methods still suffer from a critical problem: attribute leakage, where edits meant for specific objects unintentionally affect unrelated regions or other target objects. Our analy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.722, + 0.907, + 0.889 + ], + "angle": 0, + "content": "sis reveals the root cause as the semantic entanglement inherent in End-of-Sequence (EOS) embeddings generated by autoregressive text encoders, which indiscriminately aggregate attributes across prompts. To address this issue, we introduce Attribute-Leakage-free Editing (ALE), a framework that tackles attribute leakage at its source. ALE combines Object-Restricted Embeddings (ORE) to disentangle text embeddings, Region-Guided Blending for Cross-Attention Masking (RGB-CAM) for spatially precise attention, and Background Blending (BB) to preserve non-edited content. To quantitatively evaluate attribute leakage across various" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.471, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution; † Correspondence to jungseul@postech.ac.kr" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.403, + 0.901 + ], + "angle": 0, + "content": "Project page: https://mtablo.github.io/ALE_Edit_page/" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16451" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "editing methods, we propose the Attribute-Leakage Evaluation Benchmark (ALE-Bench), featuring comprehensive editing scenarios and new metrics. Extensive experiments show that ALE reduces attribute leakage by large margins, thereby enabling accurate, multi-object, text-driven image editing while faithfully preserving non-target content." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.223, + 0.227 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.483, + 0.386 + ], + "angle": 0, + "content": "Text-based image editing, where users modify existing images via natural language prompts, has emerged as a powerful alternative to traditional manual editing. Conventional editing workflows typically demand significant manual effort and domain expertise [1, 16, 25, 33]. In contrast, recent advancements leveraging generative diffusion models have substantially simplified the editing process [6, 12, 15, 20]. These models enable users to perform high-quality edits through intuitive textual prompts, making image editing more accessible and flexible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Despite these advances, existing text-based editing methods frequently suffer from a critical limitation: attribute leakage, where edits intended for specific objects inadvertently affect unrelated regions within the image. Attribute leakage can be categorized into two distinct types: Target-External Leakage (TEL), where attributes of a target object unintentionally affect non-target regions, and Target-Internal Leakage (TIL), where attributes intended for one target object inadvertently influence another target object within the same editing prompt. To mitigate this, recent studies attempt to spatially constrain editing effects by manipulating cross-attention maps [3, 29]. However, as illustrated in Figure 1, even state-of-the-art methods [3, 13, 29, 34] continue to exhibit significant TEL in single-object editing scenarios, and both TEL and substantial TIL in more complex, multi-object editing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Our analysis highlights that attribute leakage fundamentally stems from overlooked issues of the entanglement of text embeddings, specifically those associated with the End-of-Sequence (EOS) tokens, building upon recent study [14]. Most text-based editing pipelines employ autoregressive text encoders such as CLIP [22], which append EOS tokens to text prompts until reaching a fixed length (e.g., 77 tokens) to generate embeddings. Consequently, EOS embeddings inherently aggregate information from all tokens within the prompt and attend indiscriminately across image regions via cross-attention layers, exacerbating both TEL and TIL. While recent studies [14] attempt to alleviate leakage stemming from EOS embedding entanglement, our analysis demonstrates that these approaches are insufficient to fully eliminate attribute leakage (see Section 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel framework, Attribute-Leakage-Free Editing (ALE), primarily consisting of three complementary components: Object-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Restricted Embeddings (ORE), Region-Guided Blending for Cross-Attention Masking (RGB-CAM), and Background Blending (BB). ORE assigns distinct, semantically isolated embeddings to each object in the prompt, explicitly avoiding embedding entanglement. RGB-CAM enhances spatial precision in cross-attention maps by leveraging segmentation masks, restricting attention solely to intended regions. BB preserves the structural integrity of non-edited regions by integrating latents from the source image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.408 + ], + "angle": 0, + "content": "Furthermore, to systematically quantify attribute leakage, we introduce the Attribute-Leakage Evaluation Benchmark (ALE-Bench), a specialized benchmark designed explicitly for leakage evaluation in multi-object editing. Existing benchmarks predominantly focus on single-object scenarios, lacking metrics for evaluating attribute leakage comprehensively [17, 27, 31]. ALE-Bench covers a diverse range of editing scenarios, including multi-object editing in various edit types. We also propose two novel evaluation metrics: Target-External Leakage Score (TELS) and Target-Internal Leakage Score (TILS), explicitly quantifying TEL and TIL, respectively." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.409, + 0.765, + 0.423 + ], + "angle": 0, + "content": "In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.424, + 0.905, + 0.468 + ], + "angle": 0, + "content": "1. Identifying the previously overlooked role of EOS embeddings as a fundamental cause of attribute leakage in text-based image editing (Section 2)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.469, + 0.905, + 0.514 + ], + "angle": 0, + "content": "2. Proposing the novel Attribute-Leakage-Free Editing (ALE) framework, which specifically addresses leakage induced by EOS embeddings (Section 3)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.905, + 0.574 + ], + "angle": 0, + "content": "3. Introducing a comprehensive benchmark, ALE-Bench, along with novel metrics (TELS, TILS) designed explicitly to quantify attribute leakage in multi-object editing scenarios (Section 4)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.424, + 0.905, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.88, + 0.605 + ], + "angle": 0, + "content": "2. Attribute Leakage Problem and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.765 + ], + "angle": 0, + "content": "In this section, we discuss the attribute leakage problem in multi-object text-based image editing and highlight limitations of existing methods. Specifically, Section 2.1 formally introduces multi-object text-based image editing and briefly describes dual-branch frameworks commonly used for such tasks. Section 2.2 defines attribute leakage and categorizes it into two types: TEL and TIL. Finally, Section 2.3 analyzes embedding entanglement, particularly from EOS tokens, as the primary cause of attribute leakage and explains why existing methods fail to adequately address this issue." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.854, + 0.79 + ], + "angle": 0, + "content": "2.1. Multi-Object Text-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Text-based image editing modifies specific regions of a source image according to textual prompts describing desired changes. Formally, given a source image \\( x^{\\mathrm{src}} \\) and a textual prompt pair \\( (y^{\\mathrm{src}}, y^{\\mathrm{tgt}}) \\) where \\( y^{\\mathrm{src}} \\) specifies the source objects to be edited (i.e., objects to be replaced or modified), and \\( y^{\\mathrm{tgt}} \\) specifies the target objects (i.e., new or modified objects to appear)—the goal is to generate an edited" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16452" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.223, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.19, + 0.207, + 0.202 + ], + "angle": 0, + "content": "(a) Source Image" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.089, + 0.351, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.23, + 0.19, + 0.345, + 0.202 + ], + "angle": 0, + "content": "(b) Leakage Example" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.089, + 0.478, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.19, + 0.471, + 0.201 + ], + "angle": 0, + "content": "(c) Reference Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.483, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Illustration of attribute leakage in image editing. (a) The source image and editing prompts. (b) An editing result exhibiting attribute leakage. Regions highlighted in green indicate target-external leakage, where the editing spills into non-target (e.g., the green bell pepper turns into a golden pumpkin). Regions in blue show target-internal leakage, where the attributes of \"a red pumpkin\" undesirably influence the appearance of the \"golden apple\" region. (c) A reference image showing the desired editing result without attribute leakage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.412 + ], + "angle": 0, + "content": "image \\( x^{\\mathrm{tgt}} \\). Ideally, \\( x^{\\mathrm{tgt}} \\) reflects modifications described by \\( y^{\\mathrm{tgt}} \\) exclusively within regions indicated by \\( y^{\\mathrm{src}} \\), leaving all other areas unchanged. When edits involve multiple objects simultaneously, the task is classified as multi-object editing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Practically, multi-object editing prompts can be decomposed into individual object-level prompt pairs \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^K\\), typically via language models or noun- chunk parsers [2, 11]. Our research specifically focuses on scenarios involving up to \\(K = 3\\) objects. For example, consider the editing scenario in Figure 2a. Given the prompt pair \\((y^{\\mathrm{src}}, y^{\\mathrm{tgt}}) = (\\text{\"a yellow bell pepper and a red bell pepper\", \"a red pumpkin and a golden apple\")}, it can be decomposed into object-level prompts \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^2 = [\"\\text{\"a yellow bell pepper\", \"a red pumpkin\", \"a red bell pepper\", \"a golden apple\"}]\\). The resulting image \\(x^{\\mathrm{tgt}}\\) should contain a red pumpkin and a golden apple, precisely aligned to corresponding regions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To perform such precise edits, diffusion-based editing methods encode textual prompts into embeddings that guide the editing process through cross-attention layers. These layers spatially align text-described attributes with corresponding regions in the image. Achieving accurate alignment, however, requires effectively preserving the original spatial structure while synthesizing novel visual attributes. To this end, recent methods utilize dual-branch editing frameworks, which have become popular due to their ability to simultaneously retain the original image layout and introduce new content [3, 19, 29]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Dual-branch frameworks operate through two parallel pathways—a source branch and a target branch—using the same pretrained text-to-image diffusion model. Specifically, the source branch reconstructs the original image \\( x^{\\mathrm{src}} \\) guided by the source textual prompt \\( y^{\\mathrm{src}} \\), thereby capturing structural and spatial information inherent in \\( x^{\\mathrm{src}} \\). The target branch, in parallel, synthesizes new visual attributes guided by the target textual prompt \\( y^{\\mathrm{tgt}} \\). Structural consis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "tency is maintained by injecting intermediate self-attention layer components—such as queries and keys—computed from the source branch into the corresponding self-attention layers of the target branch. Despite their strengths, dual-branch frameworks face fundamental challenges, particularly attribute leakage, caused by embedding entanglement involving EOS tokens, as analyzed in subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.209, + 0.692, + 0.224 + ], + "angle": 0, + "content": "2.2. Attribute Leakage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.367 + ], + "angle": 0, + "content": "Attribute leakage is a critical challenge in multi-object text-based image editing, characterized by unintended propagation of attributes from target objects to unrelated regions or other target objects. Formally, given a source image \\( x^{\\mathrm{src}} \\) and object-level prompt pairs \\( [(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K} \\), attribute leakage occurs when modifying an object from its original description \\( y_i^{\\mathrm{src}} \\) to a new target description \\( y_i^{\\mathrm{tgt}} \\) unintentionally impacts regions or objects not specified by the editing prompt. Attribute leakage can be categorized into two distinct types:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.368, + 0.905, + 0.473 + ], + "angle": 0, + "content": "- Target-External Leakage (TEL): This occurs when editing an object specified by \\((y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})\\) unintentionally affects regions not described by the prompt pair (i.e., non-target regions). For example, as highlighted by the green region in Figure 2b, editing \\(y_{2}^{\\mathrm{src}} =\\) \"a red bell pepper\" to \\(y_{2}^{\\mathrm{tgt}} =\\) \"a golden apple\" inadvertently transforms an unrelated \"green bell pepper\" into a golden object." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.473, + 0.906, + 0.609 + ], + "angle": 0, + "content": "- Target-Internal Leakage (TIL): This occurs when editing an object specified by \\((y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})\\) unintentionally affects another target object specified by a different prompt pair \\((y_{j}^{\\mathrm{src}}, y_{j}^{\\mathrm{tgt}})\\), where \\(i \\neq j\\). For instance, as illustrated by the blue region in Figure 2b, editing \\(y_{1}^{\\mathrm{src}} = \\text{\"a yellow bell pepper\"}\\) to \\(y_{1}^{\\mathrm{tgt}} = \\text{\"a red pumpkin\"}\\) inadvertently impacts another target object described by \\(y_{2}^{\\mathrm{tgt}} = \\text{\"a golden apple\"}\\), causing it to appear as a mixture of red and golden pumpkin-like attributes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.61, + 0.905, + 0.655 + ], + "angle": 0, + "content": "Effectively mitigating both TEL and TIL is essential to achieving precise, user-intended edits, as exemplified by the desired reference result in Figure 2c." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.769, + 0.683 + ], + "angle": 0, + "content": "2.3. Causes of Attribute Leakage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Attribute leakage primarily arises from embedding entanglement, which occurs during prompt encoding. Widely used text encoders, such as CLIP, encode tokens autoregressively, causing embeddings of later tokens to unintentionally accumulate mixed semantics from preceding attributes or objects. To mitigate this entanglement, some approaches introduce object-wise embeddings by parsing prompts into distinct noun-phrase spans and encoding each segment independently [10]. However, these methods only address entanglement among original tokens in the prompt, failing to adequately resolve entanglement involving EOS tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Since CLIP pads prompts to a fixed length using EOS tokens, the EOS embeddings inevitably aggregate semantic information from multiple attributes and objects. For" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16453" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.478, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.199, + 0.44, + 0.211 + ], + "angle": 0, + "content": "(a) Source image, edit prompt, and reference result image" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.219, + 0.191, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.292, + 0.184, + 0.302 + ], + "angle": 0, + "content": "\"a red diamond\"" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.219, + 0.288, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.293, + 0.28, + 0.303 + ], + "angle": 0, + "content": "\"agolden apple\"" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.219, + 0.381, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.293, + 0.353, + 0.302 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.219, + 0.477, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.293, + 0.462, + 0.302 + ], + "angle": 0, + "content": "Edited Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.311, + 0.427, + 0.323 + ], + "angle": 0, + "content": "(b) Cross-attention maps and an edited result of ETS" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.331, + 0.192, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.404, + 0.184, + 0.414 + ], + "angle": 0, + "content": "\"a red diamond\"" + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.331, + 0.288, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.404, + 0.28, + 0.415 + ], + "angle": 0, + "content": "\"a golden apple\"" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.331, + 0.381, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.404, + 0.353, + 0.413 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.331, + 0.477, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.404, + 0.462, + 0.413 + ], + "angle": 0, + "content": "Edited Result" + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.423, + 0.456, + 0.435 + ], + "angle": 0, + "content": "(c) Cross-attention maps and an edited result of ETS + Masking" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Figure 3. Illustration of attribute leakage from EOS embeddings and misaligned cross-attention. (b) ETS with object-wise embeddings shows spatial misalignment: embeddings of one object (e.g., \"a red diamond\") can influence unrelated regions. (c) Adding cross-attention masking correctly localizes object-wise embeddings but fails to suppress leakage from EOS embeddings (e.g., diamond-like decoration on the apple). Cross-attention maps are averaged across timesteps and summed over tokens (or padded EOS tokens)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.764 + ], + "angle": 0, + "content": "instance, when encoding the prompt \"a red diamond and a golden apple\", the EOS embeddings inherently encapsulate combined semantics from all attributes and objects (e.g., \"red\", \"diamond\", \"golden\", and \"apple\"). To address this issue, End-Token-Substitution (ETS) [14] replaces attribute-rich EOS embeddings with attribute-free embeddings obtained from prompts without descriptors (e.g., \"a diamond and an apple\"). Nevertheless, ETS remains insufficient even when combined with object-wise embeddings, as the simplified EOS embeddings still aggregate semantic information across multiple objects, thereby continuing to propagate attribute leakage (see Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Embedding entanglement further exacerbates spatial inaccuracies in dual-branch text-based image editing frameworks. Since dual-branch methods inject structural information from the source image into the target editing branch, entangled embeddings can produce misaligned cross-attention maps, causing embeddings to incorrectly attend to visually similar but semantically incorrect regions during editing (Figure 3b). This visual-semantic confusion significantly exacerbates spatial inaccuracies, intensifying" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.614, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.165, + 0.612, + 0.177 + ], + "angle": 0, + "content": "(a) Source image" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.089, + 0.71, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.165, + 0.708, + 0.177 + ], + "angle": 0, + "content": "(b) Original EOS" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.806, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.165, + 0.795, + 0.176 + ], + "angle": 0, + "content": "(c) EOS \\(\\leftarrow\\) 0" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.089, + 0.903, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.165, + 0.896, + 0.176 + ], + "angle": 0, + "content": "(d) EOS \\(\\leftarrow\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.189, + 0.906, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Comparison of different EOS embedding modification strategies in a single editing scenario, where \"yellow bell pepper\" is edited to \"diamond\" and \"red bell pepper\" to \"moon\". (b) Uses the original EOS embeddings. (c) Replaces EOS embeddings with zero vectors. (d) Replaces EOS embeddings with those obtained from an empty prompt \"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.906, + 0.418 + ], + "angle": 0, + "content": "attribute leakage. Although existing methods attempt to mitigate this issue through refined cross-attention alignment or explicit masking [3, 29], these strategies fail to address the fundamental problem: EOS embeddings inherently lack spatial specificity, as they integrate semantic content from the entire prompt. Therefore, restricting the spatial attention of EOS embeddings to specific regions is inherently ineffective, further compounding attribute leakage even when employing combined strategies (Figure 3c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.906, + 0.63 + ], + "angle": 0, + "content": "One naive alternative could involve removing semantic content entirely from EOS embeddings by substituting them with zero vectors or embeddings derived from empty prompts. However, as demonstrated empirically in Figure 4 and detailed in Appendix E, this simplistic solution severely degrades visual quality and editing accuracy. These observations imply that diffusion models intrinsically depend on EOS embeddings containing semantics to achieve high-quality image editing outcomes. Therefore, resolving attribute leakage effectively requires a dedicated strategy that carefully mitigates the unintended influence of entangled EOS embeddings without completely eliminating or overly simplifying their semantic content. We propose such a targeted approach in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.644, + 0.863, + 0.663 + ], + "angle": 0, + "content": "3. Attribute-Leakage-Free Editing (ALE)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.747 + ], + "angle": 0, + "content": "Given a source image \\( x^{\\mathrm{src}} \\) and a list of \\( K \\) object-level prompt pairs \\( [(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K} \\), our goal is to generate an edited image \\( x^{\\mathrm{tgt}} \\) that (1) replaces every \\( y_i^{\\mathrm{src}} \\) with \\( y_i^{\\mathrm{tgt}} \\) inside its designated region, (2) preserves all non-target content, and (3) avoids both TIL and TEL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our proposed method, ALE, is built on the dual-branch editing framework with the Denoising Diffusion Consistent Model (DDCM) virtual inversion scheme: a source branch reconstructs \\( x^{\\mathrm{src}} \\) from latent \\( \\{z_{\\tau}^{\\mathrm{src}}\\}_{\\tau = T}^{0} \\) under the prompt \\( y_{\\mathrm{base}}^{\\mathrm{src}} = \\) \" and \".join \\( \\left(\\left[y_i^{\\mathrm{src}}\\right]\\right) \\), while a target branch denoises latent \\( \\{z_{\\tau}^{\\mathrm{tgt}}\\}_{\\tau = T}^{0} \\) toward the edited image guided by \\( y_{\\mathrm{base}}^{\\mathrm{tgt}} = \\) \" and \".join \\( \\left(\\left[y_i^{\\mathrm{tgt}}\\right]\\right) \\). On top of this backbone we introduce three key components: ORE, RGB-CAM, and BB. Figure 5 visualizes an overall pipeline, and Algorithm 1 lists the complete procedure." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16454" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.15, + 0.089, + 0.862, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.254, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 5. Overview of ALE. The framework consists of two branches: the upper branch (source branch) processes the source latent \\( z_{\\tau}^{\\mathrm{src}} \\), and the lower branch (target branch) processes the target latent \\( z_{\\tau}^{\\mathrm{tgt}} \\) at each timestep \\( \\tau \\). ORE produces multiple semantically-isolated embeddings for each target object to reduce interference between unrelated objects. RGB-CAM refines cross-attention activations using segmentation masks, aligning the attention of each embedding to spatial regions corresponding to its target object. BB merges the source latent for background regions and the target latent for edited regions. Dashed lines indicate omitted components (e.g., the encoder, the decoder, and the segmentation model) for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.346, + 0.479, + 0.363 + ], + "angle": 0, + "content": "3.1. DDCM and Dual-Branch Editing Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Virtual inversion via DDCM We adopt the Denoising Diffusion Consistent Model (DDCM) [29], which chooses the special variance schedule so that any noisy latent \\( z_{\\tau} \\) retains a closed-form link to its clean latent \\( z_0 \\) at every timestep \\( \\tau \\in [T, \\dots, 0] \\). This property enables a virtual inversion that avoids the costly DDIM / null-text inversion. Because this form matches the multi-step consistency sampler of Latent Consistency Models [8], we can edit images in as few as 4-20 steps without explicit inversion steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Dual-branch framework Starting from an initial noise, the source branch follows DDCM updates with \\( y_{\\mathrm{base}}^{\\mathrm{src}} \\), while the target branch starts with \\( z_{\\tau}^{\\mathrm{tgt}} = z_{\\tau}^{\\mathrm{src}} \\) and calculates the next latent \\( z_{\\tau - 1}^{\\mathrm{tgt}} \\) at each step using the U-Net noise \\( \\varepsilon_{\\theta}(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}}) \\) and the consistency correction term. To preserve structure of \\( x^{\\mathrm{src}} \\), we copy the query-key tensors \\( (Q, K)_{\\ell, \\tau}^{\\mathrm{src}} \\) from every self-attention layer \\( \\ell \\) of the source branch into the corresponding layer of the target branch according to self-attention injection schedule \\( S = \\{\\tau | T_S \\leq \\tau \\leq T\\} \\), where \\( T_S \\in [0, T] \\) controls the length of the self-attention injection schedule. Thus, a shorter schedule \\( (T_s \\approx T) \\) touches only early denoising steps and enables stronger edits, whereas a longer schedule \\( (T_s \\approx 0) \\) enforces stricter structural preservation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.737, + 0.421, + 0.754 + ], + "angle": 0, + "content": "3.2. Object-Restricted Embeddings (ORE)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.485, + 0.806 + ], + "angle": 0, + "content": "To address leakages in prompt embedding level, ORE encodes each object prompt \\( y_{i}^{\\mathrm{tgt}} \\) in isolation, yielding a set of token-embedding matrices" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.817, + 0.473, + 0.859 + ], + "angle": 0, + "content": "\\[\nE _ {i} ^ {\\prime} = \\left[ e _ {\\text {B O S}}, \\underbrace {e _ {\\text {t o k e n} _ {1}} , \\cdots} _ {\\text {f r o m t o k e n s i n y} _ {i} ^ {\\mathrm {t g t}}}, \\underbrace {e _ {\\text {E O S}} , \\cdots} _ {\\text {f r o m p a d d e d E O S t o k e n s}} \\right] \\in \\mathbb {R} ^ {L \\times d}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(L\\) is the padded prompt length and \\(d\\) the embedding dimension. For example, in Figure 3 case, \\(E_1'\\) is obtained" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.472 + ], + "angle": 0, + "content": "as \\([e_{\\mathrm{BOS}}, e_{\\mathrm{a}}, e_{\\mathrm{red}}, e_{\\mathrm{diamond}}, e_{\\mathrm{EOS}}, \\ldots]\\). And we construct a base embedding \\(E_{\\mathrm{base}}'\\), by encoding \\(y_{\\mathrm{base}}^{\\mathrm{tgt}}\\) and splicing \\(E_i'[y_i^{\\mathrm{tgt}}]\\) back into their original spans, to calculate base value tensor in RGB-CAM. Because no token embedding in \\(\\{E_i'\\}\\) can influence another object's span and EOS embeddings in \\(E_i'\\) only contain semantics of \\(y_i^{\\mathrm{tgt}}\\), subsequent cross-attention receives semantically disentangled embeddings, thereby preventing leakage at its source." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.509 + ], + "angle": 0, + "content": "3.3. Region-Guided Blending for Cross-Attention Masking (RGB-CAM)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Standard cross-attention layers in diffusion U-Net accept a single value tensor \\( V \\) and thus cannot exploit multiple OREs. RGB-CAM replaces the vanilla cross-attention output with a spatially blended tensor" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.584, + 0.869, + 0.626 + ], + "angle": 0, + "content": "\\[\nA = \\sum_ {i = 1} ^ {K} (M \\odot m _ {i}) V _ {i} + (M \\odot m _ {\\text {b a c k}}) V _ {\\text {b a s e}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.786 + ], + "angle": 0, + "content": "where \\( M = \\) attention\\_map(Q, K) is the base cross attention map, \\( V_{i} = W_{v}(E_{i}^{\\prime}) \\), \\( K = W_{k}(\\text{Encoder}_{\\text{text}}(y_{\\text{base}}^{\\text{tgt}})) \\), \\( V_{\\text{base}} = W_{v}(E_{\\text{base}}^{\\prime}) \\), and \\( \\{m_i\\} \\), \\( m_{\\text{back}} \\) are object and background segmentation masks from Grounded-SAM [24]. Since masks are not pixel-perfect, we apply a slight dilation. The masked tensors \\( (M \\odot m_i)V_i \\) localize each ORE to its designated region, eliminating target-internal leakage, while the background term preserves areas outside all masks. Note that only when ORE and RGB-CAM operate in tandem does ALE produce leakage-free results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.793, + 0.76, + 0.81 + ], + "angle": 0, + "content": "3.4. Background Blending (BB)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.816, + 0.906, + 0.876 + ], + "angle": 0, + "content": "Even with perfect cross-attention, backgrounds remain weakly constrained because \\(\\{y_i^{\\mathrm{tgt}}\\}\\) mention only target objects. At every timestep \\(\\tau\\) we blend the source latent by the background mask as a final step:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.885, + 0.853, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {\\tau} ^ {\\mathrm {l g t}} = m _ {\\text {b a c k}} \\odot z _ {\\tau} ^ {\\mathrm {s r c}} + (1 - m _ {\\text {b a c k}}) \\odot z _ {\\tau} ^ {\\mathrm {t g t}}.\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16455" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.103, + 0.096, + 0.406, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: Attribute-Leakage-Free Editing" + }, + { + "type": "algorithm", + "bbox": [ + 0.111, + 0.113, + 0.452, + 0.16 + ], + "angle": 0, + "content": "Input: Source image \\(x^{\\mathrm{src}}\\), object-level prompt pairs \\(\\left[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})\\right]_{i=1}^{K}\\), self-attention injection schedule \\(\\mathcal{S} = \\{\\tau \\mid 0 \\leq \\tau \\leq T_S\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.161, + 0.288, + 0.176 + ], + "angle": 0, + "content": "Output: Edited image \\(x^{\\mathrm{tgt}}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.113, + 0.184, + 0.219, + 0.198 + ], + "angle": 0, + "content": "Pre-processing" + }, + { + "type": "algorithm", + "bbox": [ + 0.111, + 0.198, + 0.473, + 0.576 + ], + "angle": 0, + "content": "Form concatenated base prompts \\(y_{\\mathrm{base}}^{\\mathrm{src}}, y_{\\mathrm{base}}^{\\mathrm{tgt}}\\) \nEncode object-restricted embeddings \\(\\{E_i'\\}_{i=1}^K\\) and the base embedding \\(E_{\\mathrm{base}}'\\) (ORE); \nObtain object masks \\(\\{m_i\\}_{i=1}^K\\) and background mask \\(m_{\\mathrm{back}}\\) with Grounded-SAM; \nInitialization \nSample initial noise \\(z_T^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)\\); \nSet \\(z_T^{\\mathrm{tgt}} \\gets z_T^{\\mathrm{src}}\\); \nfor \\(\\tau = T\\) to 1 do \n// Source branch \nPredict noise \\(\\hat{\\varepsilon}_{\\tau}^{\\mathrm{src}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{src}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{src}}\\right)\\); \nUpdate \\(z_{\\tau-1}^{\\mathrm{src}}\\) with DDCM sampling; \n// Target branch \nif \\(\\tau \\in S\\) then copy self-attention \\(Q, K\\) tensors from the source branch; \nPredict noise \\(\\hat{\\varepsilon}_{\\tau}^{\\mathrm{tgt}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}}\\right)\\); RGB-CAM[\\{E_i', m_i\\}, E'_base, \\(m_{\\mathrm{back}}\\)] \nUpdate \\(z_{\\tau-1}^{\\mathrm{tgt}}\\) with DDCM sampling; \n// Background blending (BB) \n\\(z_{\\tau-1}^{\\mathrm{tgt}} \\gets m_{\\mathrm{back}} \\odot z_{\\tau-1}^{\\mathrm{src}} + (1 - m_{\\mathrm{back}}) \\odot z_{\\tau-1}^{\\mathrm{tgt}}\\); \nend \n\\(x^{\\mathrm{tgt}} \\gets \\text{Decoder}(z_0^{\\mathrm{tgt}})\\); \nreturn \\(x^{tgt}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.655 + ], + "angle": 0, + "content": "BB guarantees preservation of non-edited regions, suppressing TEL without expensive threshold tuning required by prior local-blending heuristics [13]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.224, + 0.687 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.267, + 0.71 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "ALE-Bench construction Prior benchmarks for text-guided image editing [7, 17] focus on the visual quality of the edited result itself and neglect attribute leakage. Although they measure background preservation metrics that are similar to TELS, they overlook TIL. This makes it difficult to analyze how well a method achieves precise editing. To fill this gap, we introduce Attribute-Leakage-Evaluation Benchmark (ALE-Bench), a dedicated testbed for attribute leakage evaluation. ALE-Bench systematically varies the number of objects to be edited and the editing type, grouped into five categories: (1) color, (2) object, (3) material, (4) color + object, and (5) object + material. Each source im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "age is paired with multiple prompts for each edit type, enabling fine-grained analysis across diverse scenarios. Figure 6 shows examples of ALE-Bench. Full construction details and dataset statistics are provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.155, + 0.905, + 0.186 + ], + "angle": 0, + "content": "Evaluation Metrics We evaluate image-editing performance using the following metrics:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.186, + 0.905, + 0.231 + ], + "angle": 0, + "content": "- Structure Distance [26] quantifies how well the edited image \\( x^{\\mathrm{tgt}} \\) preserves the spatial layout of \\( x^{\\mathrm{src}} \\). A lower score indicates better structural consistency." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.232, + 0.905, + 0.276 + ], + "angle": 0, + "content": "- Editing Performance is measured by the cosine similarity between the CLIP embeddings of \\( x^{\\mathrm{tgt}} \\) and prompt \\( y^{\\mathrm{tgt}} \\). Higher similarity reflects more faithful edits." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.277, + 0.905, + 0.337 + ], + "angle": 0, + "content": "- Background Preservation is assessed on the non-edited regions using PSNR, SSIM [28], LPIPS [32], and MSE between \\( x^{\\mathrm{tgt}} \\) and \\( x^{\\mathrm{src}} \\). Higher PSNR/SSIM and lower LPIPS/MSE signify better preservation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.186, + 0.905, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.338, + 0.8, + 0.351 + ], + "angle": 0, + "content": "Furthermore, we introduce new metrics:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.354, + 0.905, + 0.397 + ], + "angle": 0, + "content": "- Target-Internal Leakage Score (TILS) measures unintended modifications inside other target-object regions as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.408, + 0.874, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T I L S} = \\frac {1}{K (K - 1)} \\sum_ {i \\neq j} ^ {K} \\operatorname {C L I P} \\left(x _ {\\mathrm {t g t}} \\odot m _ {j}, y _ {i} ^ {\\mathrm {t g t}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.463, + 0.905, + 0.538 + ], + "angle": 0, + "content": "where CLIP represents the CLIP similarity score, \\(K\\) is the number of objects to be edited, \\(x^{\\mathrm{tgt}}\\) is the edited image, \\(m_j\\) is the \\(j\\)-th object mask, and \\(y_i^{\\mathrm{tgt}}\\) is the target prompt for \\(i\\)-th object. A lower TILS implies that, as the user intended, the target objects did not affect each other." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.539, + 0.905, + 0.582 + ], + "angle": 0, + "content": "- Target-External Leakage Score (TELS) measures unintended changes in the background (non-edited regions) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.593, + 0.897, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T E L S} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathrm {C L I P} \\left(x ^ {\\mathrm {t g t}} \\odot \\left(\\mathbf {1} - \\bigcup_ {j = 1} ^ {K} m _ {j}\\right), y _ {i} ^ {\\mathrm {t g t}}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.653, + 0.907, + 0.714 + ], + "angle": 0, + "content": "The mean CLIP scores between the background and each target prompt are computed for multiple object edits. A lower TELS indicates minimal TEL, which ensures that the outside of the targets remains unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.905, + 0.777 + ], + "angle": 0, + "content": "Baselines For comparison, we selected tuning-free image editing methods including Prompt-to-Prompt (P2P) [13], MasaCtrl [3], Free-Prompt-Editing (FPE) [34], and InfEdit [29] (see Appendix A, C for further details)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.652, + 0.803 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ALE outperforms existing methods in both mitigating attribute leakage and producing high-quality edits, as seen in Table 1 and Figure 7. In particular, ALE achieves the lowest TELS and TILS, reflecting its ability to precisely apply attributes solely to the designated target regions. Across different numbers of editing objects (Table 2) and different" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16456" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.907, + 0.345 + ], + "angle": 0, + "content": "
Source ImageObject MasksColorObjectMaterialColor + ObjectObject + Material
yoga mat → pink-colored yoga matman → alienfootwear → footwear made of amberfootwear → orange-colored sandalsman → woman made of bronze
cup → white-colored cup, steam → green-colored steambook → cushion, cup → tumblerbook → book made of chrome, cup → cup made of papersteam → gray-colored smoke, cup → blue-colored jarbook → journal made of gold, cup → mug made of amethyst
wolf → pink-colored wolf, moon → purple-colored moon, mountain → gray-colored mountainwolf → eagle, moon → soccer ball, mountain → glacierwolf → wolf made of velvet, moon → moon made of steel, mountain → mountain made of leatherwolf → crimson-colored dragon, moon → navy-colored soccer ball, mountain → brown-colored gardenwolf → dog made of linen, moon → balloon made of paper, mountain → glacier made of amber
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.348, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 6. Examples of ALE-Bench. Source images are shown with object binary masks. Edit types are color, object, material modifications, and their combinations. Rows depict 1-, 2-, and 3-object edits, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.382, + 0.905, + 0.562 + ], + "angle": 0, + "content": "
Source ImageColorObjectMaterialColor + ObjectObject + Material
tree→ blue-colored treepine treetree made of diamondyellow-colored Christmas treepine tree made of bamboo
couch→ purple-colored couchbenchcouch made of leatherkhaki-colored deskdesk made of chrome
floor→ khaki-colored floorlawnfloor made of turquoiseyellow-colored tiletile made of jade
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.567, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 7. Qualitative examples of ALE applied to three source objects (tree, couch, and floor) under different editing types. The bottom rows show the corresponding prompts describing each transformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.703 + ], + "angle": 0, + "content": "editing types (Table 3), ALE demonstrates its robust performance. Figure 7 illustrates qualitative examples of ALE for various editing types. These results indicate that our method can effectively address both TEL and TIL. Detailed results are in Appendix D. Furthermore, the quantitative and qualitative results on PIE-Bench are in Appendix E." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.724, + 0.189, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.8, + 0.187, + 0.813 + ], + "angle": 0, + "content": "(a) Source image" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.724, + 0.286, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.8, + 0.256, + 0.813 + ], + "angle": 0, + "content": "(b) BB" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.724, + 0.382, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.8, + 0.378, + 0.813 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{ORE} + \\mathrm{RGB}\\)" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.724, + 0.479, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.8, + 0.455, + 0.813 + ], + "angle": 0, + "content": "(d) ALE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.821, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Figure 8. Qualitative ablation results for the editing: yellow bell pepper \\(\\rightarrow\\) diamond, red bell pepper \\(\\rightarrow\\) moon. (b) BB resolves TEL but not TIL. (c) ORE + RGB-CAM reduces TIL but not TEL. (d) ALE (BB + ORE + RGB-CAM) shows no TIL and TEL." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.61, + 0.68, + 0.625 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.815 + ], + "angle": 0, + "content": "We present ablation results on BB, ORE, and RGB-CAM in Figure 8. Each module plays a complementary role in preventing attribute leakage. Using only BB (Figure 8b) helps preserve background regions and suppresses TEL, but fails to prevent TIL. On the other hand, \\(\\mathrm{ORE} + \\mathrm{RGB}\\)-CAM (Figure 8c) reduces TIL by disentangling text embeddings and aligned cross-attention. However, without BB, it cannot preserve the original background, and TEL remains. Only when all three modules—ORE, RGB-CAM, and BB—are used together (Figure 8d), ALE achieves precise and leakage-free editing. Qualitative results and more examples are provided in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.829, + 0.637, + 0.845 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "While ALE-Bench provides a focused framework for evaluating attribute leakage, and ALE achieves strong performance on rigid attribute edits, both are currently lim" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16457" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.204 + ], + "angle": 0, + "content": "
MethodTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
P2P21.5217.260.151420.6711.150.44950.08790.5589
MasaCtrl20.1816.740.092920.0114.990.29300.04180.7346
FPE21.0717.380.116421.8912.820.39030.06560.6052
InfEdit19.5916.690.048421.7816.740.20340.03400.7709
ALE16.0315.280.016722.2030.040.03610.00140.9228
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.212, + 0.905, + 0.242 + ], + "angle": 0, + "content": "Table 1. Comparison of editing performances on ALE-Bench. ALE demonstrates the lowest attribute leakage, highest structure preservation, and superior editing performance, indicating a more precise and controlled editing." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.253, + 0.905, + 0.337 + ], + "angle": 0, + "content": "
# of Editing ObjectsTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
116.41-0.0087622.6230.010.04050.001670.9049
216.0015.420.0164822.0630.060.03600.001460.9235
315.8915.360.0246022.1930.010.03230.001540.9426
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Table 2. Performance of ALE on ALE-Bench based on the number of objects edited. Our method maintains low attribute leakage and strong background preservation even as the number of editing objects increases." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.386, + 0.905, + 0.496 + ], + "angle": 0, + "content": "
Editing TypeTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
Color17.6316.210.0089023.1232.970.02880.000790.9309
Material17.1515.960.0117922.9430.630.03390.001200.9248
Object15.8616.250.0197421.8229.030.03860.001820.9218
Color+Object15.3014.010.0230622.1528.600.04070.002050.9206
Object+Material14.5514.510.0195621.4228.880.03930.001910.9201
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.502, + 0.908, + 0.532 + ], + "angle": 0, + "content": "Table 3. Performance of ALE on ALE-Bench across various editing types (color, object, material, and combinations). The results show consistent low attribute leakage and high editing performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.557, + 0.483, + 0.771 + ], + "angle": 0, + "content": "ited to local and relatively simple transformations—such as changes in color, object identity, or material. They do not support or evaluate non-rigid transformations like style transfer, pose changes, or adding/deleting objects, where defining and detecting attribute leakage becomes ambiguous. This focus reflects our aim to establish a clear and measurable foundation before addressing more complex editing scenarios. Although the benchmark offers 3,000 detailed editing scenarios across 20 carefully curated images, the small image set may limit how well results generalize to models trained on larger or more diverse datasets. Future work could expand both the editing model and benchmark to support richer transformations and broader datasets. More limitations are discussed in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.784, + 0.21, + 0.801 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this paper, we addressed the issue of attribute leakage in diffusion-based image editing, focusing on two types of leakage: Target-External Leakage (TEL)—unintended edits in non-target regions—and Target-Internal Leakage (TIL)—interference among attributes of different targets. To mitigate these problems, we introduced Attribute" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Leakage-Free Editing (ALE), a tuning-free framework that combines three key components: Object-Restricted Embeddings (ORE), which localize attribute semantics to each target object embedding; Region-Guided Blending for Cross-Attention Masking (RGB-CAM), which constrains cross-attention with segmentation masks to avoid unintended inter-object attribute mixing; and Background Blending (BB), which preserves the source image in backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.679, + 0.909, + 0.77 + ], + "angle": 0, + "content": "We also presented ALE-Bench, a dedicated benchmark for rigorously evaluating attribute leakage across diverse multi-object editing scenarios. ALE-Bench introduces new quantitative metrics—TELS and TILS—that effectively quantify unintended modifications, providing comprehensive measures to assess editing fidelity and consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.911, + 0.892 + ], + "angle": 0, + "content": "Our extensive experimental validation demonstrated that ALE significantly outperforms existing tuning-free editing methods, achieving state-of-the-art performance by effectively minimizing attribute leakage while maintaining high editing quality and structural consistency. By effectively addressing attribute leakage with minimal computational overhead, ALE enhances the reliability and precision of multi-object image editing tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16458" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) and IITP-ITRC (Information Technology Research Center) grant funded by the Korea government (MSIT) (No.RS-2019-II191906, Artificial Intelligence Graduate School Program (POSTECH); No.RS-2021-II212068, Artificial Intelligence Innovation Hub; IITP-2025-00437866; RS-2024-00509258, Global AI Frontier Lab) and partly supported by Seoul R&BD Program through the Seoul Business Agency (SBA) funded by The Seoul Metropolitan Government (SP240008)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.295, + 0.188, + 0.311 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.32, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3, 4, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1-10, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[5] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7430-7440, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[6] Adham Elarabawy, Harish Kamath, and Samuel Denton. Direct inversion: Optimization-free text-driven real image editing with diffusion models. arXiv preprint arXiv:2211.07825, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[7] Chakrabarty et al. Lomoe: Localized multi-object editing via multi-diffusion. In Proc. ACM Multimedia, 2024. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[8] Luo et al. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[9] Yu et al. Zero-shot referring image segmentation with global-local context features. In Proc. CVPR, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.099, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.32, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.119 + ], + "angle": 0, + "content": "et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[12] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[14] Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Shahbaz Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, and Yaxing Wang. Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems, 37:137646-137672, 2024. 2, 4, 11, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[15] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[17] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[19] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817–7826, 2024. 3, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.706 + ], + "angle": 0, + "content": "[20] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.908, + 0.775 + ], + "angle": 0, + "content": "[21] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16459" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "through attention map alignment. Advances in Neural Information Processing Systems, 36:3536-3559, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kun-chang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[25] Patsorn Sangkloy, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Scribbler: Controlling deep image synthesis with sketch and color. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5400-5409, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[26] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[27] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.445, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[29] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 2, 3, 4, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[30] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.669 + ], + "angle": 0, + "content": "[31] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[33] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In Computer vision-ECCV 2016: 14th European conference, Amsterdam, the netherlands, October 11-14, 2016, proceedings, part v 14, pages 597-613. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.826, + 0.482, + 0.895 + ], + "angle": 0, + "content": "[34] Siyu Zou, Jiji Tang, Yiyi Zhou, Jing He, Chaoyi Zhao, Rongsheng Zhang, Zhipeng Hu, and Xiaoshuai Sun. Towards efficient diffusion-based image editing with instant attention masks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7864-7872, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.956 + ], + "angle": 0, + "content": "16460" + } + ] +] \ No newline at end of file diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_origin.pdf b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71c1d7e63e643aac45e3c30afbacd072ff3b3691 --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/7ef26365-88b2-4b63-b82e-fc5e555360c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76542607b628eaa85fecbb8bc4b3ea476f2b0eb6568b1eafc1ffa2221f6563fb +size 6147210 diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/full.md b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c5e099fe442667b5fa21eeb6cca053001205830a --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/full.md @@ -0,0 +1,395 @@ +# Addressing Text Embedding Leakage in Diffusion-based Image Editing + +Sunung Mun $^{1*}$ + +Jinhwan $\mathbf{Nam}^{1*}$ + +Sunghyun Cho $^{1,2}$ + +Jungseul Ok $^{1,2\dagger}$ + +Graduate School of AI, POSTECH $^{1}$ , Dept. of CSE, POSTECH $^{2}$ + +{mtablo, njh18, s.cho, jungseul}@postech.ac.kr + +![](images/60910e3dd9c8b845fd2853f1d6be7d6beb876e46911842a404c22415a9a2e636.jpg) +Figure 1. Qualitative comparison of tuning-free image-editing methods. Left to right: (1) source image, (2) visualization-only binary masks of the source objects and corresponding edit prompts (source $\rightarrow$ target), and (3-7) results from our ALE and four baselines. The masks are not provided to any method; they are shown solely to indicate which region is supposed to change. Rows 1-2 illustrate single-object edits where baselines spill changes outside the intended area or distort geometry, while ALE keeps the background intact. Rows 3-4 demonstrate multi-object edits: baselines often entangle attributes, whereas ALE preserves each attribute in its designated region, yielding leakage-free results. + +![](images/8cd935d78ddff3f8895afa43d9e3de16c940a124ec485e48ea0fdfaf2d4f9cb0.jpg) + +![](images/9a149c77d6299793243285d290ba617087ad05412ab2b6e126f4377899589fa5.jpg) + +![](images/a2c843289f0eaad14bfc97a4c5e0933db27f8501b01b6fffb5bdc61b7c3d65d3.jpg) + +![](images/ee3987fe2eb806b14470c90f54376a9cc304a9884c665414278f767d053ddc73.jpg) + +![](images/9b773ecdf6874bf78957ae9c2cb420262a064e8d7df7eaa69705fa86e67838f1.jpg) + +![](images/b47de48d6739de8d20e99e33af746be8941ec253d16db254c1a5349976c0b3cc.jpg) + +![](images/9bfdfcc28bd1f0625359194a210cc2ff88188b3278ca7e4b5ac52429527b7e70.jpg) + +![](images/867b57c2ff9a84de45f2c68da0ec168bcca64b06d6444832852ccd7feee4db92.jpg) + +![](images/93b07de86845280ccab728c6a0325f0c487e975a936b17bf260fd2be071b79e8.jpg) + +![](images/4eae2ac0aef7db039a611d58d886a98996013035796023eff86ca091ff1bee30.jpg) + +![](images/95679ccc00ef3eda2c673e2acb736ebc619e982ac035d1eea352b30ebb20d027.jpg) + +![](images/49dd9e33c114f83a12cd5dbe62eed3d567ef2babd7fdd5b8a91253ba579b42a3.jpg) + +![](images/75c671510375350284f108486addf98c25695c14344de8b687db6e2b44e52b46.jpg) + +![](images/37fe325af5d7ab827fb5119650b5aa736f29792f36fd332546dc1b075ecb1bc1.jpg) + +![](images/b691138771d81c44e7947cda28f60500b07624f8298cde0e0850b8a2979773a0.jpg) + +![](images/0ba60764e9f199613546b9e4f364fcb734f72614041a5e60ee43a691cf6e853a.jpg) +ALE (Ours) + +![](images/48f236a30788ca0a3474a36c8f8fc8637baeea1d6039a6cd542c707b31d88bad.jpg) +InfEdit + +![](images/3ff41c806ef77281748a17115b132da54b5a0c882f4b1a0064c6be3b92d75a4a.jpg) +MasaCtrl + +![](images/00c1800cb5264654e45150856a706a7e8ef8c99fa11335f3bd2823f401284963.jpg) +FPE + +![](images/ef91215f5457b213be8af7fba68321759507a4479726d547772f0ff27e9b10a8.jpg) +P2P + +# Abstract + +Text-based image editing, powered by generative diffusion models, lets users modify images through natural-language prompts and has dramatically simplified traditional workflows. Despite these advances, current methods still suffer from a critical problem: attribute leakage, where edits meant for specific objects unintentionally affect unrelated regions or other target objects. Our analy + +sis reveals the root cause as the semantic entanglement inherent in End-of-Sequence (EOS) embeddings generated by autoregressive text encoders, which indiscriminately aggregate attributes across prompts. To address this issue, we introduce Attribute-Leakage-free Editing (ALE), a framework that tackles attribute leakage at its source. ALE combines Object-Restricted Embeddings (ORE) to disentangle text embeddings, Region-Guided Blending for Cross-Attention Masking (RGB-CAM) for spatially precise attention, and Background Blending (BB) to preserve non-edited content. To quantitatively evaluate attribute leakage across various + +editing methods, we propose the Attribute-Leakage Evaluation Benchmark (ALE-Bench), featuring comprehensive editing scenarios and new metrics. Extensive experiments show that ALE reduces attribute leakage by large margins, thereby enabling accurate, multi-object, text-driven image editing while faithfully preserving non-target content. + +# 1. Introduction + +Text-based image editing, where users modify existing images via natural language prompts, has emerged as a powerful alternative to traditional manual editing. Conventional editing workflows typically demand significant manual effort and domain expertise [1, 16, 25, 33]. In contrast, recent advancements leveraging generative diffusion models have substantially simplified the editing process [6, 12, 15, 20]. These models enable users to perform high-quality edits through intuitive textual prompts, making image editing more accessible and flexible. + +Despite these advances, existing text-based editing methods frequently suffer from a critical limitation: attribute leakage, where edits intended for specific objects inadvertently affect unrelated regions within the image. Attribute leakage can be categorized into two distinct types: Target-External Leakage (TEL), where attributes of a target object unintentionally affect non-target regions, and Target-Internal Leakage (TIL), where attributes intended for one target object inadvertently influence another target object within the same editing prompt. To mitigate this, recent studies attempt to spatially constrain editing effects by manipulating cross-attention maps [3, 29]. However, as illustrated in Figure 1, even state-of-the-art methods [3, 13, 29, 34] continue to exhibit significant TEL in single-object editing scenarios, and both TEL and substantial TIL in more complex, multi-object editing scenarios. + +Our analysis highlights that attribute leakage fundamentally stems from overlooked issues of the entanglement of text embeddings, specifically those associated with the End-of-Sequence (EOS) tokens, building upon recent study [14]. Most text-based editing pipelines employ autoregressive text encoders such as CLIP [22], which append EOS tokens to text prompts until reaching a fixed length (e.g., 77 tokens) to generate embeddings. Consequently, EOS embeddings inherently aggregate information from all tokens within the prompt and attend indiscriminately across image regions via cross-attention layers, exacerbating both TEL and TIL. While recent studies [14] attempt to alleviate leakage stemming from EOS embedding entanglement, our analysis demonstrates that these approaches are insufficient to fully eliminate attribute leakage (see Section 2). + +To address these limitations, we propose a novel framework, Attribute-Leakage-Free Editing (ALE), primarily consisting of three complementary components: Object- + +Restricted Embeddings (ORE), Region-Guided Blending for Cross-Attention Masking (RGB-CAM), and Background Blending (BB). ORE assigns distinct, semantically isolated embeddings to each object in the prompt, explicitly avoiding embedding entanglement. RGB-CAM enhances spatial precision in cross-attention maps by leveraging segmentation masks, restricting attention solely to intended regions. BB preserves the structural integrity of non-edited regions by integrating latents from the source image. + +Furthermore, to systematically quantify attribute leakage, we introduce the Attribute-Leakage Evaluation Benchmark (ALE-Bench), a specialized benchmark designed explicitly for leakage evaluation in multi-object editing. Existing benchmarks predominantly focus on single-object scenarios, lacking metrics for evaluating attribute leakage comprehensively [17, 27, 31]. ALE-Bench covers a diverse range of editing scenarios, including multi-object editing in various edit types. We also propose two novel evaluation metrics: Target-External Leakage Score (TELS) and Target-Internal Leakage Score (TILS), explicitly quantifying TEL and TIL, respectively. + +In summary, our contributions are: + +1. Identifying the previously overlooked role of EOS embeddings as a fundamental cause of attribute leakage in text-based image editing (Section 2). +2. Proposing the novel Attribute-Leakage-Free Editing (ALE) framework, which specifically addresses leakage induced by EOS embeddings (Section 3). +3. Introducing a comprehensive benchmark, ALE-Bench, along with novel metrics (TELS, TILS) designed explicitly to quantify attribute leakage in multi-object editing scenarios (Section 4). + +# 2. Attribute Leakage Problem and Analysis + +In this section, we discuss the attribute leakage problem in multi-object text-based image editing and highlight limitations of existing methods. Specifically, Section 2.1 formally introduces multi-object text-based image editing and briefly describes dual-branch frameworks commonly used for such tasks. Section 2.2 defines attribute leakage and categorizes it into two types: TEL and TIL. Finally, Section 2.3 analyzes embedding entanglement, particularly from EOS tokens, as the primary cause of attribute leakage and explains why existing methods fail to adequately address this issue. + +# 2.1. Multi-Object Text-based Image Editing + +Text-based image editing modifies specific regions of a source image according to textual prompts describing desired changes. Formally, given a source image $x^{\mathrm{src}}$ and a textual prompt pair $(y^{\mathrm{src}}, y^{\mathrm{tgt}})$ where $y^{\mathrm{src}}$ specifies the source objects to be edited (i.e., objects to be replaced or modified), and $y^{\mathrm{tgt}}$ specifies the target objects (i.e., new or modified objects to appear)—the goal is to generate an edited + +![](images/e44203661143ee2aa69180641c8be3891f8f29bdf686a09f4cdecc1e2c62fcb3.jpg) +(a) Source Image + +![](images/689a93332ea52064fb0056f231127df8c8c546e7ad64dadd7ab3777616c78598.jpg) +(b) Leakage Example + +![](images/c3a833f91689b324bdf29c3247fd13df24b15473787f059cd29cddaf8d8a8d6f.jpg) +(c) Reference Result +Figure 2. Illustration of attribute leakage in image editing. (a) The source image and editing prompts. (b) An editing result exhibiting attribute leakage. Regions highlighted in green indicate target-external leakage, where the editing spills into non-target (e.g., the green bell pepper turns into a golden pumpkin). Regions in blue show target-internal leakage, where the attributes of "a red pumpkin" undesirably influence the appearance of the "golden apple" region. (c) A reference image showing the desired editing result without attribute leakage. + +image $x^{\mathrm{tgt}}$ . Ideally, $x^{\mathrm{tgt}}$ reflects modifications described by $y^{\mathrm{tgt}}$ exclusively within regions indicated by $y^{\mathrm{src}}$ , leaving all other areas unchanged. When edits involve multiple objects simultaneously, the task is classified as multi-object editing. + +Practically, multi-object editing prompts can be decomposed into individual object-level prompt pairs \([(y_i^{\mathrm{src}}, y_i^{\mathrm{tgt}})]_{i=1}^K\), typically via language models or noun- chunk parsers [2, 11]. Our research specifically focuses on scenarios involving up to \(K = 3\) objects. For example, consider the editing scenario in Figure 2a. Given the prompt pair \((y^{\mathrm{src}}, y^{\mathrm{tgt}}) = (\text{"a yellow bell pepper and a red bell pepper", "a red pumpkin and a golden apple")}, it can be decomposed into object-level prompts \([(y_i^{\mathrm{src}}, y_i^{\mathrm{tgt}})]_{i=1}^2 = ["\text{"a yellow bell pepper", "a red pumpkin", "a red bell pepper", "a golden apple"}]\). The resulting image \(x^{\mathrm{tgt}}\) should contain a red pumpkin and a golden apple, precisely aligned to corresponding regions. + +To perform such precise edits, diffusion-based editing methods encode textual prompts into embeddings that guide the editing process through cross-attention layers. These layers spatially align text-described attributes with corresponding regions in the image. Achieving accurate alignment, however, requires effectively preserving the original spatial structure while synthesizing novel visual attributes. To this end, recent methods utilize dual-branch editing frameworks, which have become popular due to their ability to simultaneously retain the original image layout and introduce new content [3, 19, 29]. + +Dual-branch frameworks operate through two parallel pathways—a source branch and a target branch—using the same pretrained text-to-image diffusion model. Specifically, the source branch reconstructs the original image $x^{\mathrm{src}}$ guided by the source textual prompt $y^{\mathrm{src}}$ , thereby capturing structural and spatial information inherent in $x^{\mathrm{src}}$ . The target branch, in parallel, synthesizes new visual attributes guided by the target textual prompt $y^{\mathrm{tgt}}$ . Structural consis + +tency is maintained by injecting intermediate self-attention layer components—such as queries and keys—computed from the source branch into the corresponding self-attention layers of the target branch. Despite their strengths, dual-branch frameworks face fundamental challenges, particularly attribute leakage, caused by embedding entanglement involving EOS tokens, as analyzed in subsequent sections. + +# 2.2. Attribute Leakage + +Attribute leakage is a critical challenge in multi-object text-based image editing, characterized by unintended propagation of attributes from target objects to unrelated regions or other target objects. Formally, given a source image $x^{\mathrm{src}}$ and object-level prompt pairs $[(y_i^{\mathrm{src}}, y_i^{\mathrm{tgt}})]_{i=1}^{K}$ , attribute leakage occurs when modifying an object from its original description $y_i^{\mathrm{src}}$ to a new target description $y_i^{\mathrm{tgt}}$ unintentionally impacts regions or objects not specified by the editing prompt. Attribute leakage can be categorized into two distinct types: + +- Target-External Leakage (TEL): This occurs when editing an object specified by $(y_{i}^{\mathrm{src}}, y_{i}^{\mathrm{tgt}})$ unintentionally affects regions not described by the prompt pair (i.e., non-target regions). For example, as highlighted by the green region in Figure 2b, editing $y_{2}^{\mathrm{src}} =$ "a red bell pepper" to $y_{2}^{\mathrm{tgt}} =$ "a golden apple" inadvertently transforms an unrelated "green bell pepper" into a golden object. + +- Target-Internal Leakage (TIL): This occurs when editing an object specified by $(y_{i}^{\mathrm{src}}, y_{i}^{\mathrm{tgt}})$ unintentionally affects another target object specified by a different prompt pair $(y_{j}^{\mathrm{src}}, y_{j}^{\mathrm{tgt}})$ , where $i \neq j$ . For instance, as illustrated by the blue region in Figure 2b, editing $y_{1}^{\mathrm{src}} = \text{"a yellow bell pepper"}$ to $y_{1}^{\mathrm{tgt}} = \text{"a red pumpkin"}$ inadvertently impacts another target object described by $y_{2}^{\mathrm{tgt}} = \text{"a golden apple"}$ , causing it to appear as a mixture of red and golden pumpkin-like attributes. + +Effectively mitigating both TEL and TIL is essential to achieving precise, user-intended edits, as exemplified by the desired reference result in Figure 2c. + +# 2.3. Causes of Attribute Leakage + +Attribute leakage primarily arises from embedding entanglement, which occurs during prompt encoding. Widely used text encoders, such as CLIP, encode tokens autoregressively, causing embeddings of later tokens to unintentionally accumulate mixed semantics from preceding attributes or objects. To mitigate this entanglement, some approaches introduce object-wise embeddings by parsing prompts into distinct noun-phrase spans and encoding each segment independently [10]. However, these methods only address entanglement among original tokens in the prompt, failing to adequately resolve entanglement involving EOS tokens. + +Since CLIP pads prompts to a fixed length using EOS tokens, the EOS embeddings inevitably aggregate semantic information from multiple attributes and objects. For + +![](images/32c0e86db5a4ebbcff02a591ee601d2882c0599e4ad51960bca4acfad3d81952.jpg) +(a) Source image, edit prompt, and reference result image + +![](images/1e5011d5123124ed16453e35a6ef9654c1e4fa7e1336c8756f2e450cab919918.jpg) +"a red diamond" + +![](images/190c145d79dbe5dca7ebeef84eedb62209c51550f37e7f7783a537a06b09c67a.jpg) +"agolden apple" + +![](images/e4a610d22d19eac4a021b572f9a494a26127dd6e13ef021b17e49fe4a8e10ddf.jpg) + + +![](images/862b80c60bb589a05585c7286671791eb694f42b4af5b6e4c68187a425204928.jpg) +Edited Result + +![](images/7293e174c80bdaa4ce9f8195a27448c7a7fca34c9fd726384962f36d64d9d8ec.jpg) +(b) Cross-attention maps and an edited result of ETS +"a red diamond" +(c) Cross-attention maps and an edited result of ETS + Masking +Figure 3. Illustration of attribute leakage from EOS embeddings and misaligned cross-attention. (b) ETS with object-wise embeddings shows spatial misalignment: embeddings of one object (e.g., "a red diamond") can influence unrelated regions. (c) Adding cross-attention masking correctly localizes object-wise embeddings but fails to suppress leakage from EOS embeddings (e.g., diamond-like decoration on the apple). Cross-attention maps are averaged across timesteps and summed over tokens (or padded EOS tokens). + +![](images/19e9e0f25fefb322b6384149154d6ddf045a54693d41da150ed6273d00f33455.jpg) +"a golden apple" + +![](images/f0780115aa17c43cd02c821ada78ca615c66dbe38534c79a95d6dbc9e5c05281.jpg) + + +![](images/7dff0d0171ad472d7ad8b4434567e05be59d2e8406819c854dab1d70e677993e.jpg) +Edited Result + +instance, when encoding the prompt "a red diamond and a golden apple", the EOS embeddings inherently encapsulate combined semantics from all attributes and objects (e.g., "red", "diamond", "golden", and "apple"). To address this issue, End-Token-Substitution (ETS) [14] replaces attribute-rich EOS embeddings with attribute-free embeddings obtained from prompts without descriptors (e.g., "a diamond and an apple"). Nevertheless, ETS remains insufficient even when combined with object-wise embeddings, as the simplified EOS embeddings still aggregate semantic information across multiple objects, thereby continuing to propagate attribute leakage (see Figure 3). + +Embedding entanglement further exacerbates spatial inaccuracies in dual-branch text-based image editing frameworks. Since dual-branch methods inject structural information from the source image into the target editing branch, entangled embeddings can produce misaligned cross-attention maps, causing embeddings to incorrectly attend to visually similar but semantically incorrect regions during editing (Figure 3b). This visual-semantic confusion significantly exacerbates spatial inaccuracies, intensifying + +![](images/93944ffe58acff755f597afe7531939e0543abe600c9656ec9b003002004a1dc.jpg) +(a) Source image +Figure 4. Comparison of different EOS embedding modification strategies in a single editing scenario, where "yellow bell pepper" is edited to "diamond" and "red bell pepper" to "moon". (b) Uses the original EOS embeddings. (c) Replaces EOS embeddings with zero vectors. (d) Replaces EOS embeddings with those obtained from an empty prompt ". + +![](images/1ebd3c6b680ff5cadc1301f594ff61d3dff9667e00a26d4b18f92be9cf45c91b.jpg) +(b) Original EOS + +![](images/d182492bf06aa5a5a96b20ad99124cb1c8702b34f37ebb8daef509779bedc82e.jpg) +(c) EOS $\leftarrow$ 0 + +![](images/e32a98110f1ec9ac7d624b454c8b659c636d1da60c56f4c276994a1442466bca.jpg) +(d) EOS $\leftarrow$ + +attribute leakage. Although existing methods attempt to mitigate this issue through refined cross-attention alignment or explicit masking [3, 29], these strategies fail to address the fundamental problem: EOS embeddings inherently lack spatial specificity, as they integrate semantic content from the entire prompt. Therefore, restricting the spatial attention of EOS embeddings to specific regions is inherently ineffective, further compounding attribute leakage even when employing combined strategies (Figure 3c). + +One naive alternative could involve removing semantic content entirely from EOS embeddings by substituting them with zero vectors or embeddings derived from empty prompts. However, as demonstrated empirically in Figure 4 and detailed in Appendix E, this simplistic solution severely degrades visual quality and editing accuracy. These observations imply that diffusion models intrinsically depend on EOS embeddings containing semantics to achieve high-quality image editing outcomes. Therefore, resolving attribute leakage effectively requires a dedicated strategy that carefully mitigates the unintended influence of entangled EOS embeddings without completely eliminating or overly simplifying their semantic content. We propose such a targeted approach in the next section. + +# 3. Attribute-Leakage-Free Editing (ALE) + +Given a source image $x^{\mathrm{src}}$ and a list of $K$ object-level prompt pairs $[(y_i^{\mathrm{src}}, y_i^{\mathrm{tgt}})]_{i=1}^{K}$ , our goal is to generate an edited image $x^{\mathrm{tgt}}$ that (1) replaces every $y_i^{\mathrm{src}}$ with $y_i^{\mathrm{tgt}}$ inside its designated region, (2) preserves all non-target content, and (3) avoids both TIL and TEL. + +Our proposed method, ALE, is built on the dual-branch editing framework with the Denoising Diffusion Consistent Model (DDCM) virtual inversion scheme: a source branch reconstructs $x^{\mathrm{src}}$ from latent $\{z_{\tau}^{\mathrm{src}}\}_{\tau = T}^{0}$ under the prompt $y_{\mathrm{base}}^{\mathrm{src}} =$ " and ".join $\left(\left[y_i^{\mathrm{src}}\right]\right)$ , while a target branch denoises latent $\{z_{\tau}^{\mathrm{tgt}}\}_{\tau = T}^{0}$ toward the edited image guided by $y_{\mathrm{base}}^{\mathrm{tgt}} =$ " and ".join $\left(\left[y_i^{\mathrm{tgt}}\right]\right)$ . On top of this backbone we introduce three key components: ORE, RGB-CAM, and BB. Figure 5 visualizes an overall pipeline, and Algorithm 1 lists the complete procedure. + +![](images/c5e06dc87c55347db03588732e8ca5f77ffea4767f5b53fcb6092e35c13ec485.jpg) +Figure 5. Overview of ALE. The framework consists of two branches: the upper branch (source branch) processes the source latent $z_{\tau}^{\mathrm{src}}$ , and the lower branch (target branch) processes the target latent $z_{\tau}^{\mathrm{tgt}}$ at each timestep $\tau$ . ORE produces multiple semantically-isolated embeddings for each target object to reduce interference between unrelated objects. RGB-CAM refines cross-attention activations using segmentation masks, aligning the attention of each embedding to spatial regions corresponding to its target object. BB merges the source latent for background regions and the target latent for edited regions. Dashed lines indicate omitted components (e.g., the encoder, the decoder, and the segmentation model) for simplicity. + +# 3.1. DDCM and Dual-Branch Editing Framework + +Virtual inversion via DDCM We adopt the Denoising Diffusion Consistent Model (DDCM) [29], which chooses the special variance schedule so that any noisy latent $z_{\tau}$ retains a closed-form link to its clean latent $z_0$ at every timestep $\tau \in [T, \dots, 0]$ . This property enables a virtual inversion that avoids the costly DDIM / null-text inversion. Because this form matches the multi-step consistency sampler of Latent Consistency Models [8], we can edit images in as few as 4-20 steps without explicit inversion steps. + +Dual-branch framework Starting from an initial noise, the source branch follows DDCM updates with $y_{\mathrm{base}}^{\mathrm{src}}$ , while the target branch starts with $z_{\tau}^{\mathrm{tgt}} = z_{\tau}^{\mathrm{src}}$ and calculates the next latent $z_{\tau - 1}^{\mathrm{tgt}}$ at each step using the U-Net noise $\varepsilon_{\theta}(z_{\tau}^{\mathrm{tgt}}, \tau, y_{\mathrm{base}}^{\mathrm{tgt}})$ and the consistency correction term. To preserve structure of $x^{\mathrm{src}}$ , we copy the query-key tensors $(Q, K)_{\ell, \tau}^{\mathrm{src}}$ from every self-attention layer $\ell$ of the source branch into the corresponding layer of the target branch according to self-attention injection schedule $S = \{\tau | T_S \leq \tau \leq T\}$ , where $T_S \in [0, T]$ controls the length of the self-attention injection schedule. Thus, a shorter schedule $(T_s \approx T)$ touches only early denoising steps and enables stronger edits, whereas a longer schedule $(T_s \approx 0)$ enforces stricter structural preservation. + +# 3.2. Object-Restricted Embeddings (ORE) + +To address leakages in prompt embedding level, ORE encodes each object prompt $y_{i}^{\mathrm{tgt}}$ in isolation, yielding a set of token-embedding matrices + +$$ +E _ {i} ^ {\prime} = \left[ e _ {\text {B O S}}, \underbrace {e _ {\text {t o k e n} _ {1}} , \cdots} _ {\text {f r o m t o k e n s i n y} _ {i} ^ {\mathrm {t g t}}}, \underbrace {e _ {\text {E O S}} , \cdots} _ {\text {f r o m p a d d e d E O S t o k e n s}} \right] \in \mathbb {R} ^ {L \times d} +$$ + +where $L$ is the padded prompt length and $d$ the embedding dimension. For example, in Figure 3 case, $E_1'$ is obtained + +as $[e_{\mathrm{BOS}}, e_{\mathrm{a}}, e_{\mathrm{red}}, e_{\mathrm{diamond}}, e_{\mathrm{EOS}}, \ldots]$ . And we construct a base embedding $E_{\mathrm{base}}'$ , by encoding $y_{\mathrm{base}}^{\mathrm{tgt}}$ and splicing $E_i'[y_i^{\mathrm{tgt}}]$ back into their original spans, to calculate base value tensor in RGB-CAM. Because no token embedding in $\{E_i'\}$ can influence another object's span and EOS embeddings in $E_i'$ only contain semantics of $y_i^{\mathrm{tgt}}$ , subsequent cross-attention receives semantically disentangled embeddings, thereby preventing leakage at its source. + +# 3.3. Region-Guided Blending for Cross-Attention Masking (RGB-CAM) + +Standard cross-attention layers in diffusion U-Net accept a single value tensor $V$ and thus cannot exploit multiple OREs. RGB-CAM replaces the vanilla cross-attention output with a spatially blended tensor + +$$ +A = \sum_ {i = 1} ^ {K} (M \odot m _ {i}) V _ {i} + (M \odot m _ {\text {b a c k}}) V _ {\text {b a s e}}, +$$ + +where $M =$ attention\_map(Q, K) is the base cross attention map, $V_{i} = W_{v}(E_{i}^{\prime})$ , $K = W_{k}(\text{Encoder}_{\text{text}}(y_{\text{base}}^{\text{tgt}}))$ , $V_{\text{base}} = W_{v}(E_{\text{base}}^{\prime})$ , and $\{m_i\}$ , $m_{\text{back}}$ are object and background segmentation masks from Grounded-SAM [24]. Since masks are not pixel-perfect, we apply a slight dilation. The masked tensors $(M \odot m_i)V_i$ localize each ORE to its designated region, eliminating target-internal leakage, while the background term preserves areas outside all masks. Note that only when ORE and RGB-CAM operate in tandem does ALE produce leakage-free results. + +# 3.4. Background Blending (BB) + +Even with perfect cross-attention, backgrounds remain weakly constrained because $\{y_i^{\mathrm{tgt}}\}$ mention only target objects. At every timestep $\tau$ we blend the source latent by the background mask as a final step: + +$$ +\tilde {z} _ {\tau} ^ {\mathrm {l g t}} = m _ {\text {b a c k}} \odot z _ {\tau} ^ {\mathrm {s r c}} + (1 - m _ {\text {b a c k}}) \odot z _ {\tau} ^ {\mathrm {t g t}}. +$$ + +Algorithm 1: Attribute-Leakage-Free Editing +Input: Source image $x^{\mathrm{src}}$ , object-level prompt pairs $\left[(y_i^{\mathrm{src}}, y_i^{\mathrm{tgt}})\right]_{i=1}^{K}$ , self-attention injection schedule $\mathcal{S} = \{\tau \mid 0 \leq \tau \leq T_S\}$ + +Output: Edited image $x^{\mathrm{tgt}}$ + +Pre-processing +Form concatenated base prompts $y_{\mathrm{base}}^{\mathrm{src}}, y_{\mathrm{base}}^{\mathrm{tgt}}$ +Encode object-restricted embeddings $\{E_i'\}_{i=1}^K$ and the base embedding $E_{\mathrm{base}}'$ (ORE); +Obtain object masks $\{m_i\}_{i=1}^K$ and background mask $m_{\mathrm{back}}$ with Grounded-SAM; +Initialization +Sample initial noise $z_T^{\mathrm{src}} \sim \mathcal{N}(0, I)$ ; +Set $z_T^{\mathrm{tgt}} \gets z_T^{\mathrm{src}}$ ; +for $\tau = T$ to 1 do +// Source branch +Predict noise $\hat{\varepsilon}_{\tau}^{\mathrm{src}} \gets \varepsilon_{\theta}\left(z_{\tau}^{\mathrm{src}}, \tau, y_{\mathrm{base}}^{\mathrm{src}}\right)$ ; +Update $z_{\tau-1}^{\mathrm{src}}$ with DDCM sampling; +// Target branch +if $\tau \in S$ then copy self-attention $Q, K$ tensors from the source branch; +Predict noise $\hat{\varepsilon}_{\tau}^{\mathrm{tgt}} \gets \varepsilon_{\theta}\left(z_{\tau}^{\mathrm{tgt}}, \tau, y_{\mathrm{base}}^{\mathrm{tgt}}\right)$ ; RGB-CAM[\{E_i', m_i\}, E'_base, $m_{\mathrm{back}}$ ] +Update $z_{\tau-1}^{\mathrm{tgt}}$ with DDCM sampling; +// Background blending (BB) + $z_{\tau-1}^{\mathrm{tgt}} \gets m_{\mathrm{back}} \odot z_{\tau-1}^{\mathrm{src}} + (1 - m_{\mathrm{back}}) \odot z_{\tau-1}^{\mathrm{tgt}}$ ; +end + $x^{\mathrm{tgt}} \gets \text{Decoder}(z_0^{\mathrm{tgt}})$ ; +return $x^{tgt}$ + +BB guarantees preservation of non-edited regions, suppressing TEL without expensive threshold tuning required by prior local-blending heuristics [13]. + +# 4. Experiments + +# 4.1. Experiment Setup + +ALE-Bench construction Prior benchmarks for text-guided image editing [7, 17] focus on the visual quality of the edited result itself and neglect attribute leakage. Although they measure background preservation metrics that are similar to TELS, they overlook TIL. This makes it difficult to analyze how well a method achieves precise editing. To fill this gap, we introduce Attribute-Leakage-Evaluation Benchmark (ALE-Bench), a dedicated testbed for attribute leakage evaluation. ALE-Bench systematically varies the number of objects to be edited and the editing type, grouped into five categories: (1) color, (2) object, (3) material, (4) color + object, and (5) object + material. Each source im + +age is paired with multiple prompts for each edit type, enabling fine-grained analysis across diverse scenarios. Figure 6 shows examples of ALE-Bench. Full construction details and dataset statistics are provided in Appendix B. + +Evaluation Metrics We evaluate image-editing performance using the following metrics: + +- Structure Distance [26] quantifies how well the edited image $x^{\mathrm{tgt}}$ preserves the spatial layout of $x^{\mathrm{src}}$ . A lower score indicates better structural consistency. +- Editing Performance is measured by the cosine similarity between the CLIP embeddings of $x^{\mathrm{tgt}}$ and prompt $y^{\mathrm{tgt}}$ . Higher similarity reflects more faithful edits. +- Background Preservation is assessed on the non-edited regions using PSNR, SSIM [28], LPIPS [32], and MSE between $x^{\mathrm{tgt}}$ and $x^{\mathrm{src}}$ . Higher PSNR/SSIM and lower LPIPS/MSE signify better preservation. + +Furthermore, we introduce new metrics: + +- Target-Internal Leakage Score (TILS) measures unintended modifications inside other target-object regions as follows: + +$$ +\mathrm {T I L S} = \frac {1}{K (K - 1)} \sum_ {i \neq j} ^ {K} \operatorname {C L I P} \left(x _ {\mathrm {t g t}} \odot m _ {j}, y _ {i} ^ {\mathrm {t g t}}\right), +$$ + +where CLIP represents the CLIP similarity score, $K$ is the number of objects to be edited, $x^{\mathrm{tgt}}$ is the edited image, $m_j$ is the $j$ -th object mask, and $y_i^{\mathrm{tgt}}$ is the target prompt for $i$ -th object. A lower TILS implies that, as the user intended, the target objects did not affect each other. + +- Target-External Leakage Score (TELS) measures unintended changes in the background (non-edited regions) as follows: + +$$ +\mathrm {T E L S} = \frac {1}{K} \sum_ {i = 1} ^ {K} \mathrm {C L I P} \left(x ^ {\mathrm {t g t}} \odot \left(\mathbf {1} - \bigcup_ {j = 1} ^ {K} m _ {j}\right), y _ {i} ^ {\mathrm {t g t}}\right). +$$ + +The mean CLIP scores between the background and each target prompt are computed for multiple object edits. A lower TELS indicates minimal TEL, which ensures that the outside of the targets remains unchanged. + +Baselines For comparison, we selected tuning-free image editing methods including Prompt-to-Prompt (P2P) [13], MasaCtrl [3], Free-Prompt-Editing (FPE) [34], and InfEdit [29] (see Appendix A, C for further details). + +# 4.2. Main Results + +ALE outperforms existing methods in both mitigating attribute leakage and producing high-quality edits, as seen in Table 1 and Figure 7. In particular, ALE achieves the lowest TELS and TILS, reflecting its ability to precisely apply attributes solely to the designated target regions. Across different numbers of editing objects (Table 2) and different + +
Source ImageObject MasksColorObjectMaterialColor + ObjectObject + Material
yoga mat → pink-colored yoga matman → alienfootwear → footwear made of amberfootwear → orange-colored sandalsman → woman made of bronze
cup → white-colored cup, steam → green-colored steambook → cushion, cup → tumblerbook → book made of chrome, cup → cup made of papersteam → gray-colored smoke, cup → blue-colored jarbook → journal made of gold, cup → mug made of amethyst
wolf → pink-colored wolf, moon → purple-colored moon, mountain → gray-colored mountainwolf → eagle, moon → soccer ball, mountain → glacierwolf → wolf made of velvet, moon → moon made of steel, mountain → mountain made of leatherwolf → crimson-colored dragon, moon → navy-colored soccer ball, mountain → brown-colored gardenwolf → dog made of linen, moon → balloon made of paper, mountain → glacier made of amber
+ +Figure 6. Examples of ALE-Bench. Source images are shown with object binary masks. Edit types are color, object, material modifications, and their combinations. Rows depict 1-, 2-, and 3-object edits, respectively. + +
Source ImageColorObjectMaterialColor + ObjectObject + Material
tree→ blue-colored treepine treetree made of diamondyellow-colored Christmas treepine tree made of bamboo
couch→ purple-colored couchbenchcouch made of leatherkhaki-colored deskdesk made of chrome
floor→ khaki-colored floorlawnfloor made of turquoiseyellow-colored tiletile made of jade
+ +Figure 7. Qualitative examples of ALE applied to three source objects (tree, couch, and floor) under different editing types. The bottom rows show the corresponding prompts describing each transformation. + +editing types (Table 3), ALE demonstrates its robust performance. Figure 7 illustrates qualitative examples of ALE for various editing types. These results indicate that our method can effectively address both TEL and TIL. Detailed results are in Appendix D. Furthermore, the quantitative and qualitative results on PIE-Bench are in Appendix E. + +![](images/8aeb21f30d8130acb860a8f2a3596564895f0e53be614c0a181014a073ff0378.jpg) +(a) Source image + +![](images/a5a8e8c1c91f26e2029738dd7b93dc757d1fdc87b653999d5378a3e9fd50f60a.jpg) +(b) BB + +![](images/8a95e448b9c30e8e105ed2cb2414026fc3630be3f72b61368822421f1b772b55.jpg) +(c) $\mathrm{ORE} + \mathrm{RGB}$ +Figure 8. Qualitative ablation results for the editing: yellow bell pepper $\rightarrow$ diamond, red bell pepper $\rightarrow$ moon. (b) BB resolves TEL but not TIL. (c) ORE + RGB-CAM reduces TIL but not TEL. (d) ALE (BB + ORE + RGB-CAM) shows no TIL and TEL. + +![](images/14f3931906e9b8449240c78b1047e456bbc0995370a0de8a5c9f87bab07bdb1e.jpg) +(d) ALE + +# 4.3. Ablation Studies + +We present ablation results on BB, ORE, and RGB-CAM in Figure 8. Each module plays a complementary role in preventing attribute leakage. Using only BB (Figure 8b) helps preserve background regions and suppresses TEL, but fails to prevent TIL. On the other hand, $\mathrm{ORE} + \mathrm{RGB}$ -CAM (Figure 8c) reduces TIL by disentangling text embeddings and aligned cross-attention. However, without BB, it cannot preserve the original background, and TEL remains. Only when all three modules—ORE, RGB-CAM, and BB—are used together (Figure 8d), ALE achieves precise and leakage-free editing. Qualitative results and more examples are provided in Appendix E. + +# 5. Limitations + +While ALE-Bench provides a focused framework for evaluating attribute leakage, and ALE achieves strong performance on rigid attribute edits, both are currently lim + +
MethodTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
P2P21.5217.260.151420.6711.150.44950.08790.5589
MasaCtrl20.1816.740.092920.0114.990.29300.04180.7346
FPE21.0717.380.116421.8912.820.39030.06560.6052
InfEdit19.5916.690.048421.7816.740.20340.03400.7709
ALE16.0315.280.016722.2030.040.03610.00140.9228
+ +Table 1. Comparison of editing performances on ALE-Bench. ALE demonstrates the lowest attribute leakage, highest structure preservation, and superior editing performance, indicating a more precise and controlled editing. + +
# of Editing ObjectsTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
116.41-0.0087622.6230.010.04050.001670.9049
216.0015.420.0164822.0630.060.03600.001460.9235
315.8915.360.0246022.1930.010.03230.001540.9426
+ +Table 2. Performance of ALE on ALE-Bench based on the number of objects edited. Our method maintains low attribute leakage and strong background preservation even as the number of editing objects increases. + +
Editing TypeTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
Color17.6316.210.0089023.1232.970.02880.000790.9309
Material17.1515.960.0117922.9430.630.03390.001200.9248
Object15.8616.250.0197421.8229.030.03860.001820.9218
Color+Object15.3014.010.0230622.1528.600.04070.002050.9206
Object+Material14.5514.510.0195621.4228.880.03930.001910.9201
+ +Table 3. Performance of ALE on ALE-Bench across various editing types (color, object, material, and combinations). The results show consistent low attribute leakage and high editing performance. + +ited to local and relatively simple transformations—such as changes in color, object identity, or material. They do not support or evaluate non-rigid transformations like style transfer, pose changes, or adding/deleting objects, where defining and detecting attribute leakage becomes ambiguous. This focus reflects our aim to establish a clear and measurable foundation before addressing more complex editing scenarios. Although the benchmark offers 3,000 detailed editing scenarios across 20 carefully curated images, the small image set may limit how well results generalize to models trained on larger or more diverse datasets. Future work could expand both the editing model and benchmark to support richer transformations and broader datasets. More limitations are discussed in Appendix F. + +# 6. Conclusion + +In this paper, we addressed the issue of attribute leakage in diffusion-based image editing, focusing on two types of leakage: Target-External Leakage (TEL)—unintended edits in non-target regions—and Target-Internal Leakage (TIL)—interference among attributes of different targets. To mitigate these problems, we introduced Attribute + +Leakage-Free Editing (ALE), a tuning-free framework that combines three key components: Object-Restricted Embeddings (ORE), which localize attribute semantics to each target object embedding; Region-Guided Blending for Cross-Attention Masking (RGB-CAM), which constrains cross-attention with segmentation masks to avoid unintended inter-object attribute mixing; and Background Blending (BB), which preserves the source image in backgrounds. + +We also presented ALE-Bench, a dedicated benchmark for rigorously evaluating attribute leakage across diverse multi-object editing scenarios. ALE-Bench introduces new quantitative metrics—TELS and TILS—that effectively quantify unintended modifications, providing comprehensive measures to assess editing fidelity and consistency. + +Our extensive experimental validation demonstrated that ALE significantly outperforms existing tuning-free editing methods, achieving state-of-the-art performance by effectively minimizing attribute leakage while maintaining high editing quality and structural consistency. By effectively addressing attribute leakage with minimal computational overhead, ALE enhances the reliability and precision of multi-object image editing tasks. + +# Acknowledgements + +This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) and IITP-ITRC (Information Technology Research Center) grant funded by the Korea government (MSIT) (No.RS-2019-II191906, Artificial Intelligence Graduate School Program (POSTECH); No.RS-2021-II212068, Artificial Intelligence Innovation Hub; IITP-2025-00437866; RS-2024-00509258, Global AI Frontier Lab) and partly supported by Seoul R&BD Program through the Seoul Business Agency (SBA) funded by The Seoul Metropolitan Government (SP240008). + +# References + +[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3, 4, 6, 11 +[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1-10, 2023. 11 +[5] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7430-7440, 2023. 11 +[6] Adham Elarabawy, Harish Kamath, and Samuel Denton. Direct inversion: Optimization-free text-driven real image editing with diffusion models. arXiv preprint arXiv:2211.07825, 2022. 2 +[7] Chakrabarty et al. Lomoe: Localized multi-object editing via multi-diffusion. In Proc. ACM Multimedia, 2024. 6, 12 +[8] Luo et al. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv, 2023. 5 +[9] Yu et al. Zero-shot referring image segmentation with global-local context features. In Proc. CVPR, 2023. 11 +[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 3, 11 +[11] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, + +et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3 +[12] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 2 +[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6, 11 +[14] Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Shahbaz Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, and Yaxing Wang. Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems, 37:137646-137672, 2024. 2, 4, 11, 13, 14 +[15] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[17] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6, 13 +[18] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 11 +[19] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817–7826, 2024. 3, 11 +[20] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2 +[21] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 11 +[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[23] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence + +through attention map alignment. Advances in Neural Information Processing Systems, 36:3536-3559, 2023. 11 +[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kun-chang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 5 +[25] Patsorn Sangkloy, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Scribbler: Controlling deep image synthesis with sketch and color. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5400-5409, 2017. 2 +[26] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 6 +[27] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2 +[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[29] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 2, 3, 4, 5, 6, 11 +[30] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 11 +[31] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 2 +[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[33] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In Computer vision-ECCV 2016: 14th European conference, Amsterdam, the netherlands, October 11-14, 2016, proceedings, part v 14, pages 597-613. Springer, 2016. 2 +[34] Siyu Zou, Jiji Tang, Yiyi Zhou, Jing He, Chaoyi Zhao, Rongsheng Zhang, Zhipeng Hu, and Xiaoshuai Sun. Towards efficient diffusion-based image editing with instant attention masks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7864-7872, 2024. 2, 6 \ No newline at end of file diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/images.zip b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f0f920a4cac17e15ba979833df259f04f20ac422 --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dfb8dcd865d290407856a28ea0e1d9eaab068e62e87c4a94db930a265e110a7 +size 739551 diff --git a/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/layout.json b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c9ec105426d4d417c1ed8e601282912f6f1dd6f0 --- /dev/null +++ b/2025/Addressing Text Embedding Leakage in Diffusion-based Image Editing/layout.json @@ -0,0 +1,10868 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "text", + "content": "Addressing Text Embedding Leakage in Diffusion-based Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 186, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 186, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 186, + 158 + ], + "type": "text", + "content": "Sunung Mun" + }, + { + "bbox": [ + 111, + 141, + 186, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 142, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 142, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 209, + 142, + 287, + 156 + ], + "type": "text", + "content": "Jinhwan " + }, + { + "bbox": [ + 209, + 142, + 287, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{Nam}^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 142, + 397, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 142, + 397, + 157 + ], + "spans": [ + { + "bbox": [ + 312, + 142, + 397, + 157 + ], + "type": "text", + "content": "Sunghyun Cho" + }, + { + "bbox": [ + 312, + 142, + 397, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 421, + 141, + 498, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 421, + 141, + 498, + 157 + ], + "spans": [ + { + "bbox": [ + 421, + 141, + 498, + 157 + ], + "type": "text", + "content": "Jungseul Ok" + }, + { + "bbox": [ + 421, + 141, + 498, + 157 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "text", + "content": "Graduate School of AI, POSTECH" + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "text", + "content": ", Dept. of CSE, POSTECH" + }, + { + "bbox": [ + 138, + 171, + 471, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 179, + 187, + 428, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 187, + 428, + 200 + ], + "spans": [ + { + "bbox": [ + 179, + 187, + 428, + 200 + ], + "type": "text", + "content": "{mtablo, njh18, s.cho, jungseul}@postech.ac.kr" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 208, + 202, + 490 + ], + "blocks": [ + { + "bbox": [ + 60, + 208, + 202, + 490 + ], + "lines": [ + { + "bbox": [ + 60, + 208, + 202, + 490 + ], + "spans": [ + { + "bbox": [ + 60, + 208, + 202, + 490 + ], + "type": "image", + "image_path": "60910e3dd9c8b845fd2853f1d6be7d6beb876e46911842a404c22415a9a2e636.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 494, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 555, + 559 + ], + "type": "text", + "content": "Figure 1. Qualitative comparison of tuning-free image-editing methods. Left to right: (1) source image, (2) visualization-only binary masks of the source objects and corresponding edit prompts (source " + }, + { + "bbox": [ + 55, + 494, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 494, + 555, + 559 + ], + "type": "text", + "content": " target), and (3-7) results from our ALE and four baselines. The masks are not provided to any method; they are shown solely to indicate which region is supposed to change. Rows 1-2 illustrate single-object edits where baselines spill changes outside the intended area or distort geometry, while ALE keeps the background intact. Rows 3-4 demonstrate multi-object edits: baselines often entangle attributes, whereas ALE preserves each attribute in its designated region, yielding leakage-free results." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 206, + 208, + 276, + 275 + ], + "blocks": [ + { + "bbox": [ + 206, + 208, + 276, + 275 + ], + "lines": [ + { + "bbox": [ + 206, + 208, + 276, + 275 + ], + "spans": [ + { + "bbox": [ + 206, + 208, + 276, + 275 + ], + "type": "image", + "image_path": "8cd935d78ddff3f8895afa43d9e3de16c940a124ec485e48ea0fdfaf2d4f9cb0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 276, + 208, + 343, + 275 + ], + "blocks": [ + { + "bbox": [ + 276, + 208, + 343, + 275 + ], + "lines": [ + { + "bbox": [ + 276, + 208, + 343, + 275 + ], + "spans": [ + { + "bbox": [ + 276, + 208, + 343, + 275 + ], + "type": "image", + "image_path": "9a149c77d6299793243285d290ba617087ad05412ab2b6e126f4377899589fa5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 343, + 208, + 410, + 275 + ], + "blocks": [ + { + "bbox": [ + 343, + 208, + 410, + 275 + ], + "lines": [ + { + "bbox": [ + 343, + 208, + 410, + 275 + ], + "spans": [ + { + "bbox": [ + 343, + 208, + 410, + 275 + ], + "type": "image", + "image_path": "a2c843289f0eaad14bfc97a4c5e0933db27f8501b01b6fffb5bdc61b7c3d65d3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 208, + 479, + 275 + ], + "blocks": [ + { + "bbox": [ + 410, + 208, + 479, + 275 + ], + "lines": [ + { + "bbox": [ + 410, + 208, + 479, + 275 + ], + "spans": [ + { + "bbox": [ + 410, + 208, + 479, + 275 + ], + "type": "image", + "image_path": "ee3987fe2eb806b14470c90f54376a9cc304a9884c665414278f767d053ddc73.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 479, + 208, + 547, + 275 + ], + "blocks": [ + { + "bbox": [ + 479, + 208, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 479, + 208, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 479, + 208, + 547, + 275 + ], + "type": "image", + "image_path": "9b773ecdf6874bf78957ae9c2cb420262a064e8d7df7eaa69705fa86e67838f1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 207, + 276, + 276, + 343 + ], + "blocks": [ + { + "bbox": [ + 207, + 276, + 276, + 343 + ], + "lines": [ + { + "bbox": [ + 207, + 276, + 276, + 343 + ], + "spans": [ + { + "bbox": [ + 207, + 276, + 276, + 343 + ], + "type": "image", + "image_path": "b47de48d6739de8d20e99e33af746be8941ec253d16db254c1a5349976c0b3cc.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 276, + 276, + 343, + 343 + ], + "blocks": [ + { + "bbox": [ + 276, + 276, + 343, + 343 + ], + "lines": [ + { + "bbox": [ + 276, + 276, + 343, + 343 + ], + "spans": [ + { + "bbox": [ + 276, + 276, + 343, + 343 + ], + "type": "image", + "image_path": "9bfdfcc28bd1f0625359194a210cc2ff88188b3278ca7e4b5ac52429527b7e70.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 343, + 276, + 410, + 344 + ], + "blocks": [ + { + "bbox": [ + 343, + 276, + 410, + 344 + ], + "lines": [ + { + "bbox": [ + 343, + 276, + 410, + 344 + ], + "spans": [ + { + "bbox": [ + 343, + 276, + 410, + 344 + ], + "type": "image", + "image_path": "867b57c2ff9a84de45f2c68da0ec168bcca64b06d6444832852ccd7feee4db92.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 276, + 479, + 344 + ], + "blocks": [ + { + "bbox": [ + 410, + 276, + 479, + 344 + ], + "lines": [ + { + "bbox": [ + 410, + 276, + 479, + 344 + ], + "spans": [ + { + "bbox": [ + 410, + 276, + 479, + 344 + ], + "type": "image", + "image_path": "93b07de86845280ccab728c6a0325f0c487e975a936b17bf260fd2be071b79e8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 479, + 276, + 547, + 344 + ], + "blocks": [ + { + "bbox": [ + 479, + 276, + 547, + 344 + ], + "lines": [ + { + "bbox": [ + 479, + 276, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 479, + 276, + 547, + 344 + ], + "type": "image", + "image_path": "4eae2ac0aef7db039a611d58d886a98996013035796023eff86ca091ff1bee30.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 207, + 344, + 276, + 410 + ], + "blocks": [ + { + "bbox": [ + 207, + 344, + 276, + 410 + ], + "lines": [ + { + "bbox": [ + 207, + 344, + 276, + 410 + ], + "spans": [ + { + "bbox": [ + 207, + 344, + 276, + 410 + ], + "type": "image", + "image_path": "95679ccc00ef3eda2c673e2acb736ebc619e982ac035d1eea352b30ebb20d027.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 276, + 344, + 343, + 411 + ], + "blocks": [ + { + "bbox": [ + 276, + 344, + 343, + 411 + ], + "lines": [ + { + "bbox": [ + 276, + 344, + 343, + 411 + ], + "spans": [ + { + "bbox": [ + 276, + 344, + 343, + 411 + ], + "type": "image", + "image_path": "49dd9e33c114f83a12cd5dbe62eed3d567ef2babd7fdd5b8a91253ba579b42a3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 343, + 344, + 410, + 411 + ], + "blocks": [ + { + "bbox": [ + 343, + 344, + 410, + 411 + ], + "lines": [ + { + "bbox": [ + 343, + 344, + 410, + 411 + ], + "spans": [ + { + "bbox": [ + 343, + 344, + 410, + 411 + ], + "type": "image", + "image_path": "75c671510375350284f108486addf98c25695c14344de8b687db6e2b44e52b46.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 410, + 344, + 479, + 411 + ], + "blocks": [ + { + "bbox": [ + 410, + 344, + 479, + 411 + ], + "lines": [ + { + "bbox": [ + 410, + 344, + 479, + 411 + ], + "spans": [ + { + "bbox": [ + 410, + 344, + 479, + 411 + ], + "type": "image", + "image_path": "37fe325af5d7ab827fb5119650b5aa736f29792f36fd332546dc1b075ecb1bc1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 479, + 344, + 547, + 411 + ], + "blocks": [ + { + "bbox": [ + 479, + 344, + 547, + 411 + ], + "lines": [ + { + "bbox": [ + 479, + 344, + 547, + 411 + ], + "spans": [ + { + "bbox": [ + 479, + 344, + 547, + 411 + ], + "type": "image", + "image_path": "b691138771d81c44e7947cda28f60500b07624f8298cde0e0850b8a2979773a0.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 207, + 411, + 276, + 479 + ], + "blocks": [ + { + "bbox": [ + 207, + 411, + 276, + 479 + ], + "lines": [ + { + "bbox": [ + 207, + 411, + 276, + 479 + ], + "spans": [ + { + "bbox": [ + 207, + 411, + 276, + 479 + ], + "type": "image", + "image_path": "0ba60764e9f199613546b9e4f364fcb734f72614041a5e60ee43a691cf6e853a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 480, + 259, + 488 + ], + "lines": [ + { + "bbox": [ + 224, + 480, + 259, + 488 + ], + "spans": [ + { + "bbox": [ + 224, + 480, + 259, + 488 + ], + "type": "text", + "content": "ALE (Ours)" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 276, + 411, + 343, + 479 + ], + "blocks": [ + { + "bbox": [ + 276, + 411, + 343, + 479 + ], + "lines": [ + { + "bbox": [ + 276, + 411, + 343, + 479 + ], + "spans": [ + { + "bbox": [ + 276, + 411, + 343, + 479 + ], + "type": "image", + "image_path": "48f236a30788ca0a3474a36c8f8fc8637baeea1d6039a6cd542c707b31d88bad.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 480, + 320, + 488 + ], + "lines": [ + { + "bbox": [ + 299, + 480, + 320, + 488 + ], + "spans": [ + { + "bbox": [ + 299, + 480, + 320, + 488 + ], + "type": "text", + "content": "InfEdit" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 343, + 411, + 410, + 479 + ], + "blocks": [ + { + "bbox": [ + 343, + 411, + 410, + 479 + ], + "lines": [ + { + "bbox": [ + 343, + 411, + 410, + 479 + ], + "spans": [ + { + "bbox": [ + 343, + 411, + 410, + 479 + ], + "type": "image", + "image_path": "3ff41c806ef77281748a17115b132da54b5a0c882f4b1a0064c6be3b92d75a4a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 480, + 391, + 488 + ], + "lines": [ + { + "bbox": [ + 364, + 480, + 391, + 488 + ], + "spans": [ + { + "bbox": [ + 364, + 480, + 391, + 488 + ], + "type": "text", + "content": "MasaCtrl" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 410, + 411, + 479, + 479 + ], + "blocks": [ + { + "bbox": [ + 410, + 411, + 479, + 479 + ], + "lines": [ + { + "bbox": [ + 410, + 411, + 479, + 479 + ], + "spans": [ + { + "bbox": [ + 410, + 411, + 479, + 479 + ], + "type": "image", + "image_path": "00c1800cb5264654e45150856a706a7e8ef8c99fa11335f3bd2823f401284963.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 480, + 452, + 488 + ], + "lines": [ + { + "bbox": [ + 438, + 480, + 452, + 488 + ], + "spans": [ + { + "bbox": [ + 438, + 480, + 452, + 488 + ], + "type": "text", + "content": "FPE" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 479, + 411, + 547, + 479 + ], + "blocks": [ + { + "bbox": [ + 479, + 411, + 547, + 479 + ], + "lines": [ + { + "bbox": [ + 479, + 411, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 479, + 411, + 547, + 479 + ], + "type": "image", + "image_path": "ef91215f5457b213be8af7fba68321759507a4479726d547772f0ff27e9b10a8.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 480, + 519, + 488 + ], + "lines": [ + { + "bbox": [ + 506, + 480, + 519, + 488 + ], + "spans": [ + { + "bbox": [ + 506, + 480, + 519, + 488 + ], + "type": "text", + "content": "P2P" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 151, + 570, + 200, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 570, + 200, + 582 + ], + "spans": [ + { + "bbox": [ + 151, + 570, + 200, + 582 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 53, + 597, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 597, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 597, + 295, + 682 + ], + "type": "text", + "content": "Text-based image editing, powered by generative diffusion models, lets users modify images through natural-language prompts and has dramatically simplified traditional workflows. Despite these advances, current methods still suffer from a critical problem: attribute leakage, where edits meant for specific objects unintentionally affect unrelated regions or other target objects. Our analy" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 571, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 704 + ], + "type": "text", + "content": "sis reveals the root cause as the semantic entanglement inherent in End-of-Sequence (EOS) embeddings generated by autoregressive text encoders, which indiscriminately aggregate attributes across prompts. To address this issue, we introduce Attribute-Leakage-free Editing (ALE), a framework that tackles attribute leakage at its source. ALE combines Object-Restricted Embeddings (ORE) to disentangle text embeddings, Region-Guided Blending for Cross-Attention Masking (RGB-CAM) for spatially precise attention, and Background Blending (BB) to preserve non-edited content. To quantitatively evaluate attribute leakage across various" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 288, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 288, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 288, + 703 + ], + "type": "text", + "content": "* Equal contribution; † Correspondence to jungseul@postech.ac.kr" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 69, + 703, + 246, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 246, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 246, + 713 + ], + "type": "text", + "content": "Project page: https://mtablo.github.io/ALE_Edit_page/" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16451" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "editing methods, we propose the Attribute-Leakage Evaluation Benchmark (ALE-Bench), featuring comprehensive editing scenarios and new metrics. Extensive experiments show that ALE reduces attribute leakage by large margins, thereby enabling accurate, multi-object, text-driven image editing while faithfully preserving non-target content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 295, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 305 + ], + "type": "text", + "content": "Text-based image editing, where users modify existing images via natural language prompts, has emerged as a powerful alternative to traditional manual editing. Conventional editing workflows typically demand significant manual effort and domain expertise [1, 16, 25, 33]. In contrast, recent advancements leveraging generative diffusion models have substantially simplified the editing process [6, 12, 15, 20]. These models enable users to perform high-quality edits through intuitive textual prompts, making image editing more accessible and flexible." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 498 + ], + "type": "text", + "content": "Despite these advances, existing text-based editing methods frequently suffer from a critical limitation: attribute leakage, where edits intended for specific objects inadvertently affect unrelated regions within the image. Attribute leakage can be categorized into two distinct types: Target-External Leakage (TEL), where attributes of a target object unintentionally affect non-target regions, and Target-Internal Leakage (TIL), where attributes intended for one target object inadvertently influence another target object within the same editing prompt. To mitigate this, recent studies attempt to spatially constrain editing effects by manipulating cross-attention maps [3, 29]. However, as illustrated in Figure 1, even state-of-the-art methods [3, 13, 29, 34] continue to exhibit significant TEL in single-object editing scenarios, and both TEL and substantial TIL in more complex, multi-object editing scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 677 + ], + "type": "text", + "content": "Our analysis highlights that attribute leakage fundamentally stems from overlooked issues of the entanglement of text embeddings, specifically those associated with the End-of-Sequence (EOS) tokens, building upon recent study [14]. Most text-based editing pipelines employ autoregressive text encoders such as CLIP [22], which append EOS tokens to text prompts until reaching a fixed length (e.g., 77 tokens) to generate embeddings. Consequently, EOS embeddings inherently aggregate information from all tokens within the prompt and attend indiscriminately across image regions via cross-attention layers, exacerbating both TEL and TIL. While recent studies [14] attempt to alleviate leakage stemming from EOS embedding entanglement, our analysis demonstrates that these approaches are insufficient to fully eliminate attribute leakage (see Section 2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To address these limitations, we propose a novel framework, Attribute-Leakage-Free Editing (ALE), primarily consisting of three complementary components: Object-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "Restricted Embeddings (ORE), Region-Guided Blending for Cross-Attention Masking (RGB-CAM), and Background Blending (BB). ORE assigns distinct, semantically isolated embeddings to each object in the prompt, explicitly avoiding embedding entanglement. RGB-CAM enhances spatial precision in cross-attention maps by leveraging segmentation masks, restricting attention solely to intended regions. BB preserves the structural integrity of non-edited regions by integrating latents from the source image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 323 + ], + "type": "text", + "content": "Furthermore, to systematically quantify attribute leakage, we introduce the Attribute-Leakage Evaluation Benchmark (ALE-Bench), a specialized benchmark designed explicitly for leakage evaluation in multi-object editing. Existing benchmarks predominantly focus on single-object scenarios, lacking metrics for evaluating attribute leakage comprehensively [17, 27, 31]. ALE-Bench covers a diverse range of editing scenarios, including multi-object editing in various edit types. We also propose two novel evaluation metrics: Target-External Leakage Score (TELS) and Target-Internal Leakage Score (TILS), explicitly quantifying TEL and TIL, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 323, + 468, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 323, + 468, + 335 + ], + "spans": [ + { + "bbox": [ + 326, + 323, + 468, + 335 + ], + "type": "text", + "content": "In summary, our contributions are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 335, + 553, + 454 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 315, + 335, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 335, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 335, + 553, + 370 + ], + "type": "text", + "content": "1. Identifying the previously overlooked role of EOS embeddings as a fundamental cause of attribute leakage in text-based image editing (Section 2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 371, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 553, + 407 + ], + "type": "text", + "content": "2. Proposing the novel Attribute-Leakage-Free Editing (ALE) framework, which specifically addresses leakage induced by EOS embeddings (Section 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 553, + 454 + ], + "type": "text", + "content": "3. Introducing a comprehensive benchmark, ALE-Bench, along with novel metrics (TELS, TILS) designed explicitly to quantify attribute leakage in multi-object editing scenarios (Section 4)." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 465, + 538, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 538, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 538, + 479 + ], + "type": "text", + "content": "2. Attribute Leakage Problem and Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 485, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 605 + ], + "type": "text", + "content": "In this section, we discuss the attribute leakage problem in multi-object text-based image editing and highlight limitations of existing methods. Specifically, Section 2.1 formally introduces multi-object text-based image editing and briefly describes dual-branch frameworks commonly used for such tasks. Section 2.2 defines attribute leakage and categorizes it into two types: TEL and TIL. Finally, Section 2.3 analyzes embedding entanglement, particularly from EOS tokens, as the primary cause of attribute leakage and explains why existing methods fail to adequately address this issue." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 612, + 522, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 522, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 522, + 625 + ], + "type": "text", + "content": "2.1. Multi-Object Text-based Image Editing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Text-based image editing modifies specific regions of a source image according to textual prompts describing desired changes. Formally, given a source image " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " and a textual prompt pair " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "(y^{\\mathrm{src}}, y^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " specifies the source objects to be edited (i.e., objects to be replaced or modified), and " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " specifies the target objects (i.e., new or modified objects to appear)—the goal is to generate an edited" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16452" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 136, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "type": "image", + "image_path": "e44203661143ee2aa69180641c8be3891f8f29bdf686a09f4cdecc1e2c62fcb3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 150, + 126, + 159 + ], + "lines": [ + { + "bbox": [ + 69, + 150, + 126, + 159 + ], + "spans": [ + { + "bbox": [ + 69, + 150, + 126, + 159 + ], + "type": "text", + "content": "(a) Source Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 70, + 214, + 148 + ], + "blocks": [ + { + "bbox": [ + 138, + 70, + 214, + 148 + ], + "lines": [ + { + "bbox": [ + 138, + 70, + 214, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 70, + 214, + 148 + ], + "type": "image", + "image_path": "689a93332ea52064fb0056f231127df8c8c546e7ad64dadd7ab3777616c78598.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 150, + 211, + 159 + ], + "lines": [ + { + "bbox": [ + 140, + 150, + 211, + 159 + ], + "spans": [ + { + "bbox": [ + 140, + 150, + 211, + 159 + ], + "type": "text", + "content": "(b) Leakage Example" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 292, + 148 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 292, + 148 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 292, + 148 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 292, + 148 + ], + "type": "image", + "image_path": "c3a833f91689b324bdf29c3247fd13df24b15473787f059cd29cddaf8d8a8d6f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 150, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 220, + 150, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 220, + 150, + 288, + 159 + ], + "type": "text", + "content": "(c) Reference Result" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 169, + 295, + 268 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 295, + 268 + ], + "type": "text", + "content": "Figure 2. Illustration of attribute leakage in image editing. (a) The source image and editing prompts. (b) An editing result exhibiting attribute leakage. Regions highlighted in green indicate target-external leakage, where the editing spills into non-target (e.g., the green bell pepper turns into a golden pumpkin). Regions in blue show target-internal leakage, where the attributes of \"a red pumpkin\" undesirably influence the appearance of the \"golden apple\" region. (c) A reference image showing the desired editing result without attribute leakage." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "content": "image " + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "content": ". Ideally, " + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "content": " reflects modifications described by " + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "content": " exclusively within regions indicated by " + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 277, + 295, + 326 + ], + "type": "text", + "content": ", leaving all other areas unchanged. When edits involve multiple objects simultaneously, the task is classified as multi-object editing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 327, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 483 + ], + "type": "text", + "content": "Practically, multi-object editing prompts can be decomposed into individual object-level prompt pairs \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^K\\), typically via language models or noun- chunk parsers [2, 11]. Our research specifically focuses on scenarios involving up to \\(K = 3\\) objects. For example, consider the editing scenario in Figure 2a. Given the prompt pair \\((y^{\\mathrm{src}}, y^{\\mathrm{tgt}}) = (\\text{\"a yellow bell pepper and a red bell pepper\", \"a red pumpkin and a golden apple\")}, it can be decomposed into object-level prompts \\([(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^2 = [\"\\text{\"a yellow bell pepper\", \"a red pumpkin\", \"a red bell pepper\", \"a golden apple\"}]\\). The resulting image \\(x^{\\mathrm{tgt}}\\) should contain a red pumpkin and a golden apple, precisely aligned to corresponding regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 295, + 616 + ], + "type": "text", + "content": "To perform such precise edits, diffusion-based editing methods encode textual prompts into embeddings that guide the editing process through cross-attention layers. These layers spatially align text-described attributes with corresponding regions in the image. Achieving accurate alignment, however, requires effectively preserving the original spatial structure while synthesizing novel visual attributes. To this end, recent methods utilize dual-branch editing frameworks, which have become popular due to their ability to simultaneously retain the original image layout and introduce new content [3, 19, 29]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Dual-branch frameworks operate through two parallel pathways—a source branch and a target branch—using the same pretrained text-to-image diffusion model. Specifically, the source branch reconstructs the original image " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " guided by the source textual prompt " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", thereby capturing structural and spatial information inherent in " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". The target branch, in parallel, synthesizes new visual attributes guided by the target textual prompt " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Structural consis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "tency is maintained by injecting intermediate self-attention layer components—such as queries and keys—computed from the source branch into the corresponding self-attention layers of the target branch. Despite their strengths, dual-branch frameworks face fundamental challenges, particularly attribute leakage, caused by embedding entanglement involving EOS tokens, as analyzed in subsequent sections." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 165, + 423, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 423, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 423, + 177 + ], + "type": "text", + "content": "2.2. Attribute Leakage" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": "Attribute leakage is a critical challenge in multi-object text-based image editing, characterized by unintended propagation of attributes from target objects to unrelated regions or other target objects. Formally, given a source image " + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": " and object-level prompt pairs " + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "inline_equation", + "content": "[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K}" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": ", attribute leakage occurs when modifying an object from its original description " + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": " to a new target description " + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": " unintentionally impacts regions or objects not specified by the editing prompt. Attribute leakage can be categorized into two distinct types:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "text", + "content": "- Target-External Leakage (TEL): This occurs when editing an object specified by " + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "inline_equation", + "content": "(y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "text", + "content": " unintentionally affects regions not described by the prompt pair (i.e., non-target regions). For example, as highlighted by the green region in Figure 2b, editing " + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "inline_equation", + "content": "y_{2}^{\\mathrm{src}} =" + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "text", + "content": " \"a red bell pepper\" to " + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "inline_equation", + "content": "y_{2}^{\\mathrm{tgt}} =" + }, + { + "bbox": [ + 314, + 291, + 553, + 374 + ], + "type": "text", + "content": " \"a golden apple\" inadvertently transforms an unrelated \"green bell pepper\" into a golden object." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": "- Target-Internal Leakage (TIL): This occurs when editing an object specified by " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "(y_{i}^{\\mathrm{src}}, y_{i}^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": " unintentionally affects another target object specified by a different prompt pair " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "(y_{j}^{\\mathrm{src}}, y_{j}^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": ". For instance, as illustrated by the blue region in Figure 2b, editing " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "y_{1}^{\\mathrm{src}} = \\text{\"a yellow bell pepper\"}" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "y_{1}^{\\mathrm{tgt}} = \\text{\"a red pumpkin\"}" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": " inadvertently impacts another target object described by " + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "inline_equation", + "content": "y_{2}^{\\mathrm{tgt}} = \\text{\"a golden apple\"}" + }, + { + "bbox": [ + 314, + 374, + 554, + 482 + ], + "type": "text", + "content": ", causing it to appear as a mixture of red and golden pumpkin-like attributes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 483, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 518 + ], + "type": "text", + "content": "Effectively mitigating both TEL and TIL is essential to achieving precise, user-intended edits, as exemplified by the desired reference result in Figure 2c." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 528, + 470, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 470, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 470, + 540 + ], + "type": "text", + "content": "2.3. Causes of Attribute Leakage" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 676 + ], + "type": "text", + "content": "Attribute leakage primarily arises from embedding entanglement, which occurs during prompt encoding. Widely used text encoders, such as CLIP, encode tokens autoregressively, causing embeddings of later tokens to unintentionally accumulate mixed semantics from preceding attributes or objects. To mitigate this entanglement, some approaches introduce object-wise embeddings by parsing prompts into distinct noun-phrase spans and encoding each segment independently [10]. However, these methods only address entanglement among original tokens in the prompt, failing to adequately resolve entanglement involving EOS tokens." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Since CLIP pads prompts to a fixed length using EOS tokens, the EOS embeddings inevitably aggregate semantic information from multiple attributes and objects. For" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16453" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 292, + 155 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 292, + 155 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 292, + 155 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 292, + 155 + ], + "type": "image", + "image_path": "32c0e86db5a4ebbcff02a591ee601d2882c0599e4ad51960bca4acfad3d81952.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 157, + 269, + 167 + ], + "lines": [ + { + "bbox": [ + 83, + 157, + 269, + 167 + ], + "spans": [ + { + "bbox": [ + 83, + 157, + 269, + 167 + ], + "type": "text", + "content": "(a) Source image, edit prompt, and reference result image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 173, + 116, + 230 + ], + "blocks": [ + { + "bbox": [ + 60, + 173, + 116, + 230 + ], + "lines": [ + { + "bbox": [ + 60, + 173, + 116, + 230 + ], + "spans": [ + { + "bbox": [ + 60, + 173, + 116, + 230 + ], + "type": "image", + "image_path": "1e5011d5123124ed16453e35a6ef9654c1e4fa7e1336c8756f2e450cab919918.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 231, + 112, + 239 + ], + "lines": [ + { + "bbox": [ + 65, + 231, + 112, + 239 + ], + "spans": [ + { + "bbox": [ + 65, + 231, + 112, + 239 + ], + "type": "text", + "content": "\"a red diamond\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 173, + 176, + 230 + ], + "blocks": [ + { + "bbox": [ + 118, + 173, + 176, + 230 + ], + "lines": [ + { + "bbox": [ + 118, + 173, + 176, + 230 + ], + "spans": [ + { + "bbox": [ + 118, + 173, + 176, + 230 + ], + "type": "image", + "image_path": "190c145d79dbe5dca7ebeef84eedb62209c51550f37e7f7783a537a06b09c67a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 232, + 171, + 239 + ], + "lines": [ + { + "bbox": [ + 123, + 232, + 171, + 239 + ], + "spans": [ + { + "bbox": [ + 123, + 232, + 171, + 239 + ], + "type": "text", + "content": "\"agolden apple\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 173, + 233, + 230 + ], + "blocks": [ + { + "bbox": [ + 178, + 173, + 233, + 230 + ], + "lines": [ + { + "bbox": [ + 178, + 173, + 233, + 230 + ], + "spans": [ + { + "bbox": [ + 178, + 173, + 233, + 230 + ], + "type": "image", + "image_path": "e4a610d22d19eac4a021b572f9a494a26127dd6e13ef021b17e49fe4a8e10ddf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 232, + 216, + 239 + ], + "lines": [ + { + "bbox": [ + 194, + 232, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 194, + 232, + 216, + 239 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 173, + 291, + 230 + ], + "blocks": [ + { + "bbox": [ + 235, + 173, + 291, + 230 + ], + "lines": [ + { + "bbox": [ + 235, + 173, + 291, + 230 + ], + "spans": [ + { + "bbox": [ + 235, + 173, + 291, + 230 + ], + "type": "image", + "image_path": "862b80c60bb589a05585c7286671791eb694f42b4af5b6e4c68187a425204928.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 232, + 282, + 239 + ], + "lines": [ + { + "bbox": [ + 243, + 232, + 282, + 239 + ], + "spans": [ + { + "bbox": [ + 243, + 232, + 282, + 239 + ], + "type": "text", + "content": "Edited Result" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 60, + 262, + 117, + 319 + ], + "blocks": [ + { + "bbox": [ + 90, + 246, + 261, + 255 + ], + "lines": [ + { + "bbox": [ + 90, + 246, + 261, + 255 + ], + "spans": [ + { + "bbox": [ + 90, + 246, + 261, + 255 + ], + "type": "text", + "content": "(b) Cross-attention maps and an edited result of ETS" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 262, + 117, + 319 + ], + "lines": [ + { + "bbox": [ + 60, + 262, + 117, + 319 + ], + "spans": [ + { + "bbox": [ + 60, + 262, + 117, + 319 + ], + "type": "image", + "image_path": "7293e174c80bdaa4ce9f8195a27448c7a7fca34c9fd726384962f36d64d9d8ec.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 319, + 112, + 327 + ], + "lines": [ + { + "bbox": [ + 64, + 319, + 112, + 327 + ], + "spans": [ + { + "bbox": [ + 64, + 319, + 112, + 327 + ], + "type": "text", + "content": "\"a red diamond\"" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 335, + 279, + 344 + ], + "lines": [ + { + "bbox": [ + 72, + 335, + 279, + 344 + ], + "spans": [ + { + "bbox": [ + 72, + 335, + 279, + 344 + ], + "type": "text", + "content": "(c) Cross-attention maps and an edited result of ETS + Masking" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 354, + 296, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 452 + ], + "type": "text", + "content": "Figure 3. Illustration of attribute leakage from EOS embeddings and misaligned cross-attention. (b) ETS with object-wise embeddings shows spatial misalignment: embeddings of one object (e.g., \"a red diamond\") can influence unrelated regions. (c) Adding cross-attention masking correctly localizes object-wise embeddings but fails to suppress leakage from EOS embeddings (e.g., diamond-like decoration on the apple). Cross-attention maps are averaged across timesteps and summed over tokens (or padded EOS tokens)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 118, + 262, + 176, + 319 + ], + "blocks": [ + { + "bbox": [ + 118, + 262, + 176, + 319 + ], + "lines": [ + { + "bbox": [ + 118, + 262, + 176, + 319 + ], + "spans": [ + { + "bbox": [ + 118, + 262, + 176, + 319 + ], + "type": "image", + "image_path": "19e9e0f25fefb322b6384149154d6ddf045a54693d41da150ed6273d00f33455.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 319, + 171, + 328 + ], + "lines": [ + { + "bbox": [ + 123, + 319, + 171, + 328 + ], + "spans": [ + { + "bbox": [ + 123, + 319, + 171, + 328 + ], + "type": "text", + "content": "\"a golden apple\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 178, + 262, + 233, + 319 + ], + "blocks": [ + { + "bbox": [ + 178, + 262, + 233, + 319 + ], + "lines": [ + { + "bbox": [ + 178, + 262, + 233, + 319 + ], + "spans": [ + { + "bbox": [ + 178, + 262, + 233, + 319 + ], + "type": "image", + "image_path": "f0780115aa17c43cd02c821ada78ca615c66dbe38534c79a95d6dbc9e5c05281.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 319, + 216, + 327 + ], + "lines": [ + { + "bbox": [ + 194, + 319, + 216, + 327 + ], + "spans": [ + { + "bbox": [ + 194, + 319, + 216, + 327 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 235, + 262, + 291, + 319 + ], + "blocks": [ + { + "bbox": [ + 235, + 262, + 291, + 319 + ], + "lines": [ + { + "bbox": [ + 235, + 262, + 291, + 319 + ], + "spans": [ + { + "bbox": [ + 235, + 262, + 291, + 319 + ], + "type": "image", + "image_path": "7dff0d0171ad472d7ad8b4434567e05be59d2e8406819c854dab1d70e677993e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 319, + 282, + 327 + ], + "lines": [ + { + "bbox": [ + 243, + 319, + 282, + 327 + ], + "spans": [ + { + "bbox": [ + 243, + 319, + 282, + 327 + ], + "type": "text", + "content": "Edited Result" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": "instance, when encoding the prompt \"a red diamond and a golden apple\", the EOS embeddings inherently encapsulate combined semantics from all attributes and objects (e.g., \"red\", \"diamond\", \"golden\", and \"apple\"). To address this issue, End-Token-Substitution (ETS) [14] replaces attribute-rich EOS embeddings with attribute-free embeddings obtained from prompts without descriptors (e.g., \"a diamond and an apple\"). Nevertheless, ETS remains insufficient even when combined with object-wise embeddings, as the simplified EOS embeddings still aggregate semantic information across multiple objects, thereby continuing to propagate attribute leakage (see Figure 3)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Embedding entanglement further exacerbates spatial inaccuracies in dual-branch text-based image editing frameworks. Since dual-branch methods inject structural information from the source image into the target editing branch, entangled embeddings can produce misaligned cross-attention maps, causing embeddings to incorrectly attend to visually similar but semantically incorrect regions during editing (Figure 3b). This visual-semantic confusion significantly exacerbates spatial inaccuracies, intensifying" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 375, + 129 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 375, + 129 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 375, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 375, + 129 + ], + "type": "image", + "image_path": "93944ffe58acff755f597afe7531939e0543abe600c9656ec9b003002004a1dc.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 130, + 374, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 130, + 374, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 130, + 374, + 140 + ], + "type": "text", + "content": "(a) Source image" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 149, + 554, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 554, + 216 + ], + "type": "text", + "content": "Figure 4. Comparison of different EOS embedding modification strategies in a single editing scenario, where \"yellow bell pepper\" is edited to \"diamond\" and \"red bell pepper\" to \"moon\". (b) Uses the original EOS embeddings. (c) Replaces EOS embeddings with zero vectors. (d) Replaces EOS embeddings with those obtained from an empty prompt \"." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 434, + 129 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 434, + 129 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 434, + 129 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 434, + 129 + ], + "type": "image", + "image_path": "1ebd3c6b680ff5cadc1301f594ff61d3dff9667e00a26d4b18f92be9cf45c91b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 130, + 433, + 140 + ], + "lines": [ + { + "bbox": [ + 378, + 130, + 433, + 140 + ], + "spans": [ + { + "bbox": [ + 378, + 130, + 433, + 140 + ], + "type": "text", + "content": "(b) Original EOS" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 493, + 129 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 493, + 129 + ], + "type": "image", + "image_path": "d182492bf06aa5a5a96b20ad99124cb1c8702b34f37ebb8daef509779bedc82e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 130, + 486, + 139 + ], + "lines": [ + { + "bbox": [ + 443, + 130, + 486, + 139 + ], + "spans": [ + { + "bbox": [ + 443, + 130, + 486, + 139 + ], + "type": "text", + "content": "(c) EOS " + }, + { + "bbox": [ + 443, + 130, + 486, + 139 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 443, + 130, + 486, + 139 + ], + "type": "text", + "content": " 0" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 494, + 70, + 552, + 129 + ], + "blocks": [ + { + "bbox": [ + 494, + 70, + 552, + 129 + ], + "lines": [ + { + "bbox": [ + 494, + 70, + 552, + 129 + ], + "spans": [ + { + "bbox": [ + 494, + 70, + 552, + 129 + ], + "type": "image", + "image_path": "e32a98110f1ec9ac7d624b454c8b659c636d1da60c56f4c276994a1442466bca.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 130, + 548, + 139 + ], + "lines": [ + { + "bbox": [ + 500, + 130, + 548, + 139 + ], + "spans": [ + { + "bbox": [ + 500, + 130, + 548, + 139 + ], + "type": "text", + "content": "(d) EOS " + }, + { + "bbox": [ + 500, + 130, + 548, + 139 + ], + "type": "inline_equation", + "content": "\\leftarrow" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 223, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 554, + 331 + ], + "type": "text", + "content": "attribute leakage. Although existing methods attempt to mitigate this issue through refined cross-attention alignment or explicit masking [3, 29], these strategies fail to address the fundamental problem: EOS embeddings inherently lack spatial specificity, as they integrate semantic content from the entire prompt. Therefore, restricting the spatial attention of EOS embeddings to specific regions is inherently ineffective, further compounding attribute leakage even when employing combined strategies (Figure 3c)." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "type": "text", + "content": "One naive alternative could involve removing semantic content entirely from EOS embeddings by substituting them with zero vectors or embeddings derived from empty prompts. However, as demonstrated empirically in Figure 4 and detailed in Appendix E, this simplistic solution severely degrades visual quality and editing accuracy. These observations imply that diffusion models intrinsically depend on EOS embeddings containing semantics to achieve high-quality image editing outcomes. Therefore, resolving attribute leakage effectively requires a dedicated strategy that carefully mitigates the unintended influence of entangled EOS embeddings without completely eliminating or overly simplifying their semantic content. We propose such a targeted approach in the next section." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 510, + 528, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 528, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 528, + 525 + ], + "type": "text", + "content": "3. Attribute-Leakage-Free Editing (ALE)" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": "Given a source image " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": " and a list of " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": " object-level prompt pairs " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})]_{i=1}^{K}" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": ", our goal is to generate an edited image " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": " that (1) replaces every " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 591 + ], + "type": "text", + "content": " inside its designated region, (2) preserves all non-target content, and (3) avoids both TIL and TEL." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": "Our proposed method, ALE, is built on the dual-branch editing framework with the Denoising Diffusion Consistent Model (DDCM) virtual inversion scheme: a source branch reconstructs " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": " from latent " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{z_{\\tau}^{\\mathrm{src}}\\}_{\\tau = T}^{0}" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": " under the prompt " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{base}}^{\\mathrm{src}} =" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": " \" and \".join " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\left(\\left[y_i^{\\mathrm{src}}\\right]\\right)" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": ", while a target branch denoises latent " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\{z_{\\tau}^{\\mathrm{tgt}}\\}_{\\tau = T}^{0}" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": " toward the edited image guided by " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{base}}^{\\mathrm{tgt}} =" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": " \" and \".join " + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\left(\\left[y_i^{\\mathrm{tgt}}\\right]\\right)" + }, + { + "bbox": [ + 313, + 592, + 554, + 713 + ], + "type": "text", + "content": ". On top of this backbone we introduce three key components: ORE, RGB-CAM, and BB. Figure 5 visualizes an overall pipeline, and Algorithm 1 lists the complete procedure." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16454" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 70, + 527, + 194 + ], + "blocks": [ + { + "bbox": [ + 91, + 70, + 527, + 194 + ], + "lines": [ + { + "bbox": [ + 91, + 70, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 91, + 70, + 527, + 194 + ], + "type": "image", + "image_path": "c5e06dc87c55347db03588732e8ca5f77ffea4767f5b53fcb6092e35c13ec485.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": "Figure 5. Overview of ALE. The framework consists of two branches: the upper branch (source branch) processes the source latent " + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "inline_equation", + "content": "z_{\\tau}^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": ", and the lower branch (target branch) processes the target latent " + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "inline_equation", + "content": "z_{\\tau}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": " at each timestep " + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 201, + 555, + 269 + ], + "type": "text", + "content": ". ORE produces multiple semantically-isolated embeddings for each target object to reduce interference between unrelated objects. RGB-CAM refines cross-attention activations using segmentation masks, aligning the attention of each embedding to spatial regions corresponding to its target object. BB merges the source latent for background regions and the target latent for edited regions. Dashed lines indicate omitted components (e.g., the encoder, the decoder, and the segmentation model) for simplicity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 274, + 293, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 293, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 293, + 287 + ], + "type": "text", + "content": "3.1. DDCM and Dual-Branch Editing Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "text", + "content": "Virtual inversion via DDCM We adopt the Denoising Diffusion Consistent Model (DDCM) [29], which chooses the special variance schedule so that any noisy latent " + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "inline_equation", + "content": "z_{\\tau}" + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "text", + "content": " retains a closed-form link to its clean latent " + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "text", + "content": " at every timestep " + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\tau \\in [T, \\dots, 0]" + }, + { + "bbox": [ + 55, + 293, + 296, + 402 + ], + "type": "text", + "content": ". This property enables a virtual inversion that avoids the costly DDIM / null-text inversion. Because this form matches the multi-step consistency sampler of Latent Consistency Models [8], we can edit images in as few as 4-20 steps without explicit inversion steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": "Dual-branch framework Starting from an initial noise, the source branch follows DDCM updates with " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{base}}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": ", while the target branch starts with " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "z_{\\tau}^{\\mathrm{tgt}} = z_{\\tau}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " and calculates the next latent " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "z_{\\tau - 1}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " at each step using the U-Net noise " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\theta}(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " and the consistency correction term. To preserve structure of " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": ", we copy the query-key tensors " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "(Q, K)_{\\ell, \\tau}^{\\mathrm{src}}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " from every self-attention layer " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " of the source branch into the corresponding layer of the target branch according to self-attention injection schedule " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "S = \\{\\tau | T_S \\leq \\tau \\leq T\\}" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "T_S \\in [0, T]" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " controls the length of the self-attention injection schedule. Thus, a shorter schedule " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "(T_s \\approx T)" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " touches only early denoising steps and enables stronger edits, whereas a longer schedule " + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "inline_equation", + "content": "(T_s \\approx 0)" + }, + { + "bbox": [ + 55, + 405, + 296, + 574 + ], + "type": "text", + "content": " enforces stricter structural preservation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 583, + 257, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 257, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 257, + 597 + ], + "type": "text", + "content": "3.2. Object-Restricted Embeddings (ORE)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "text", + "content": "To address leakages in prompt embedding level, ORE encodes each object prompt " + }, + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "inline_equation", + "content": "y_{i}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "text", + "content": " in isolation, yielding a set of token-embedding matrices" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 647, + 289, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 647, + 289, + 680 + ], + "spans": [ + { + "bbox": [ + 63, + 647, + 289, + 680 + ], + "type": "interline_equation", + "content": "E _ {i} ^ {\\prime} = \\left[ e _ {\\text {B O S}}, \\underbrace {e _ {\\text {t o k e n} _ {1}} , \\cdots} _ {\\text {f r o m t o k e n s i n y} _ {i} ^ {\\mathrm {t g t}}}, \\underbrace {e _ {\\text {E O S}} , \\cdots} _ {\\text {f r o m p a d d e d E O S t o k e n s}} \\right] \\in \\mathbb {R} ^ {L \\times d}", + "image_path": "76bd73cd9543e0343aebfda7924981bbc41008c9d2a4177249aeb1e56056e182.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is the padded prompt length and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " the embedding dimension. For example, in Figure 3 case, " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_1'" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is obtained" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "[e_{\\mathrm{BOS}}, e_{\\mathrm{a}}, e_{\\mathrm{red}}, e_{\\mathrm{diamond}}, e_{\\mathrm{EOS}}, \\ldots]" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": ". And we construct a base embedding " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{base}}'" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": ", by encoding " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{base}}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": " and splicing " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "E_i'[y_i^{\\mathrm{tgt}}]" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": " back into their original spans, to calculate base value tensor in RGB-CAM. Because no token embedding in " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\{E_i'\\}" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": " can influence another object's span and EOS embeddings in " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "E_i'" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": " only contain semantics of " + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 274, + 555, + 373 + ], + "type": "text", + "content": ", subsequent cross-attention receives semantically disentangled embeddings, thereby preventing leakage at its source." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 403 + ], + "type": "text", + "content": "3.3. Region-Guided Blending for Cross-Attention Masking (RGB-CAM)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 456 + ], + "type": "text", + "content": "Standard cross-attention layers in diffusion U-Net accept a single value tensor " + }, + { + "bbox": [ + 313, + 407, + 555, + 456 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 407, + 555, + 456 + ], + "type": "text", + "content": " and thus cannot exploit multiple OREs. RGB-CAM replaces the vanilla cross-attention output with a spatially blended tensor" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 462, + 531, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 462, + 531, + 495 + ], + "spans": [ + { + "bbox": [ + 336, + 462, + 531, + 495 + ], + "type": "interline_equation", + "content": "A = \\sum_ {i = 1} ^ {K} (M \\odot m _ {i}) V _ {i} + (M \\odot m _ {\\text {b a c k}}) V _ {\\text {b a s e}},", + "image_path": "903804d1073e97412dc1c757e4e8876d0cbac65261d0bc33af89470fb5bee6f3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "M =" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": " attention\\_map(Q, K) is the base cross attention map, " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "V_{i} = W_{v}(E_{i}^{\\prime})" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "K = W_{k}(\\text{Encoder}_{\\text{text}}(y_{\\text{base}}^{\\text{tgt}}))" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "V_{\\text{base}} = W_{v}(E_{\\text{base}}^{\\prime})" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "\\{m_i\\}" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "m_{\\text{back}}" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": " are object and background segmentation masks from Grounded-SAM [24]. Since masks are not pixel-perfect, we apply a slight dilation. The masked tensors " + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "inline_equation", + "content": "(M \\odot m_i)V_i" + }, + { + "bbox": [ + 313, + 502, + 555, + 622 + ], + "type": "text", + "content": " localize each ORE to its designated region, eliminating target-internal leakage, while the background term preserves areas outside all masks. Note that only when ORE and RGB-CAM operate in tandem does ALE produce leakage-free results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 465, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 465, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 465, + 641 + ], + "type": "text", + "content": "3.4. Background Blending (BB)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "content": "Even with perfect cross-attention, backgrounds remain weakly constrained because " + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\{y_i^{\\mathrm{tgt}}\\}" + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "content": " mention only target objects. At every timestep " + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 646, + 554, + 693 + ], + "type": "text", + "content": " we blend the source latent by the background mask as a final step:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 700, + 522, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 700, + 522, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 700, + 522, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {\\tau} ^ {\\mathrm {l g t}} = m _ {\\text {b a c k}} \\odot z _ {\\tau} ^ {\\mathrm {s r c}} + (1 - m _ {\\text {b a c k}}) \\odot z _ {\\tau} ^ {\\mathrm {t g t}}.", + "image_path": "4e7613459ba3a5ac1b16f2803e7b96757fad7e0985595a062c311270c769deca.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16455" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 67, + 89, + 276, + 126 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 248, + 87 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 248, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 248, + 87 + ], + "type": "text", + "content": "Algorithm 1: Attribute-Leakage-Free Editing" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "lines": [ + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "text", + "content": "Input: Source image " + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "text", + "content": ", object-level prompt pairs " + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "inline_equation", + "content": "\\left[(y_i^{\\mathrm{src}}, y_i^{\\mathrm{tgt}})\\right]_{i=1}^{K}" + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "text", + "content": ", self-attention injection schedule " + }, + { + "bbox": [ + 67, + 89, + 276, + 126 + ], + "type": "inline_equation", + "content": "\\mathcal{S} = \\{\\tau \\mid 0 \\leq \\tau \\leq T_S\\}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 69, + 127, + 176, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 127, + 176, + 139 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 176, + 139 + ], + "type": "text", + "content": "Output: Edited image " + }, + { + "bbox": [ + 69, + 127, + 176, + 139 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + } + ] + } + ], + "index": 2 + }, + { + "type": "code", + "bbox": [ + 67, + 156, + 289, + 456 + ], + "blocks": [ + { + "bbox": [ + 69, + 145, + 134, + 156 + ], + "lines": [ + { + "bbox": [ + 69, + 145, + 134, + 156 + ], + "spans": [ + { + "bbox": [ + 69, + 145, + 134, + 156 + ], + "type": "text", + "content": "Pre-processing" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "lines": [ + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "Form concatenated base prompts " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{base}}^{\\mathrm{src}}, y_{\\mathrm{base}}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " \nEncode object-restricted embeddings " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\{E_i'\\}_{i=1}^K" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " and the base embedding " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{base}}'" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " (ORE); \nObtain object masks " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\{m_i\\}_{i=1}^K" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " and background mask " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{back}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " with Grounded-SAM; \nInitialization \nSample initial noise " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "z_T^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; \nSet " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "z_T^{\\mathrm{tgt}} \\gets z_T^{\\mathrm{src}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; \nfor " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\tau = T" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " to 1 do \n// Source branch \nPredict noise " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\varepsilon}_{\\tau}^{\\mathrm{src}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{src}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{src}}\\right)" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; \nUpdate " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "z_{\\tau-1}^{\\mathrm{src}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " with DDCM sampling; \n// Target branch \nif " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\tau \\in S" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " then copy self-attention " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " tensors from the source branch; \nPredict noise " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\varepsilon}_{\\tau}^{\\mathrm{tgt}} \\gets \\varepsilon_{\\theta}\\left(z_{\\tau}^{\\mathrm{tgt}}, \\tau, y_{\\mathrm{base}}^{\\mathrm{tgt}}\\right)" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; RGB-CAM[\\{E_i', m_i\\}, E'_base, " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{back}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "] \nUpdate " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "z_{\\tau-1}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": " with DDCM sampling; \n// Background blending (BB) \n" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "z_{\\tau-1}^{\\mathrm{tgt}} \\gets m_{\\mathrm{back}} \\odot z_{\\tau-1}^{\\mathrm{src}} + (1 - m_{\\mathrm{back}}) \\odot z_{\\tau-1}^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; \nend \n" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}} \\gets \\text{Decoder}(z_0^{\\mathrm{tgt}})" + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "text", + "content": "; \nreturn " + }, + { + "bbox": [ + 67, + 156, + 289, + 456 + ], + "type": "inline_equation", + "content": "x^{tgt}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 480, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 518 + ], + "type": "text", + "content": "BB guarantees preservation of non-edited regions, suppressing TEL without expensive threshold tuning required by prior local-blending heuristics [13]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 137, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 137, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 137, + 544 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 550, + 163, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 163, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 163, + 562 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "ALE-Bench construction Prior benchmarks for text-guided image editing [7, 17] focus on the visual quality of the edited result itself and neglect attribute leakage. Although they measure background preservation metrics that are similar to TELS, they overlook TIL. This makes it difficult to analyze how well a method achieves precise editing. To fill this gap, we introduce Attribute-Leakage-Evaluation Benchmark (ALE-Bench), a dedicated testbed for attribute leakage evaluation. ALE-Bench systematically varies the number of objects to be edited and the editing type, grouped into five categories: (1) color, (2) object, (3) material, (4) color + object, and (5) object + material. Each source im" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "age is paired with multiple prompts for each edit type, enabling fine-grained analysis across diverse scenarios. Figure 6 shows examples of ALE-Bench. Full construction details and dataset statistics are provided in Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 122, + 553, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 147 + ], + "type": "text", + "content": "Evaluation Metrics We evaluate image-editing performance using the following metrics:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 147, + 553, + 266 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "text", + "content": "- Structure Distance [26] quantifies how well the edited image " + }, + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "text", + "content": " preserves the spatial layout of " + }, + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 314, + 147, + 553, + 182 + ], + "type": "text", + "content": ". A lower score indicates better structural consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "text", + "content": "- Editing Performance is measured by the cosine similarity between the CLIP embeddings of " + }, + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "text", + "content": " and prompt " + }, + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "inline_equation", + "content": "y^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 314, + 183, + 553, + 218 + ], + "type": "text", + "content": ". Higher similarity reflects more faithful edits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "content": "- Background Preservation is assessed on the non-edited regions using PSNR, SSIM [28], LPIPS [32], and MSE between " + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{src}}" + }, + { + "bbox": [ + 314, + 219, + 553, + 266 + ], + "type": "text", + "content": ". Higher PSNR/SSIM and lower LPIPS/MSE signify better preservation." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 267, + 489, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 267, + 489, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 267, + 489, + 277 + ], + "type": "text", + "content": "Furthermore, we introduce new metrics:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 280, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 280, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 314, + 280, + 553, + 314 + ], + "type": "text", + "content": "- Target-Internal Leakage Score (TILS) measures unintended modifications inside other target-object regions as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 340, + 323, + 534, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 323, + 534, + 357 + ], + "spans": [ + { + "bbox": [ + 340, + 323, + 534, + 357 + ], + "type": "interline_equation", + "content": "\\mathrm {T I L S} = \\frac {1}{K (K - 1)} \\sum_ {i \\neq j} ^ {K} \\operatorname {C L I P} \\left(x _ {\\mathrm {t g t}} \\odot m _ {j}, y _ {i} ^ {\\mathrm {t g t}}\\right),", + "image_path": "9af930266f923d5bd2fc2d88d2410a9bdd1e02156cfdfa83f791cbc8dbd027c4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": "where CLIP represents the CLIP similarity score, " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": " is the number of objects to be edited, " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": " is the edited image, " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": "-th object mask, and " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "y_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": " is the target prompt for " + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 321, + 366, + 553, + 426 + ], + "type": "text", + "content": "-th object. A lower TILS implies that, as the user intended, the target objects did not affect each other." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 426, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 553, + 460 + ], + "type": "text", + "content": "- Target-External Leakage Score (TELS) measures unintended changes in the background (non-edited regions) as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 469, + 548, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 469, + 548, + 509 + ], + "spans": [ + { + "bbox": [ + 326, + 469, + 548, + 509 + ], + "type": "interline_equation", + "content": "\\mathrm {T E L S} = \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\mathrm {C L I P} \\left(x ^ {\\mathrm {t g t}} \\odot \\left(\\mathbf {1} - \\bigcup_ {j = 1} ^ {K} m _ {j}\\right), y _ {i} ^ {\\mathrm {t g t}}\\right).", + "image_path": "9d8598a3eb66c579902e53323272b641fe1fe7c55bf2b1eac5dcb4151c601bfc.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 517, + 555, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 517, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 321, + 517, + 555, + 565 + ], + "type": "text", + "content": "The mean CLIP scores between the background and each target prompt are computed for multiple object edits. A lower TELS indicates minimal TEL, which ensures that the outside of the targets remains unchanged." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 567, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 615 + ], + "type": "text", + "content": "Baselines For comparison, we selected tuning-free image editing methods including Prompt-to-Prompt (P2P) [13], MasaCtrl [3], Free-Prompt-Editing (FPE) [34], and InfEdit [29] (see Appendix A, C for further details)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 399, + 635 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "ALE outperforms existing methods in both mitigating attribute leakage and producing high-quality edits, as seen in Table 1 and Figure 7. In particular, ALE achieves the lowest TELS and TILS, reflecting its ability to precisely apply attributes solely to the designated target regions. Across different numbers of editing objects (Table 2) and different" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16456" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 555, + 273 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 555, + 273 + ], + "type": "table", + "html": "
Source ImageObject MasksColorObjectMaterialColor + ObjectObject + Material
yoga mat → pink-colored yoga matman → alienfootwear → footwear made of amberfootwear → orange-colored sandalsman → woman made of bronze
cup → white-colored cup, steam → green-colored steambook → cushion, cup → tumblerbook → book made of chrome, cup → cup made of papersteam → gray-colored smoke, cup → blue-colored jarbook → journal made of gold, cup → mug made of amethyst
wolf → pink-colored wolf, moon → purple-colored moon, mountain → gray-colored mountainwolf → eagle, moon → soccer ball, mountain → glacierwolf → wolf made of velvet, moon → moon made of steel, mountain → mountain made of leatherwolf → crimson-colored dragon, moon → navy-colored soccer ball, mountain → brown-colored gardenwolf → dog made of linen, moon → balloon made of paper, mountain → glacier made of amber
", + "image_path": "c1854bd306a3086f1f0f6767ba2ddcccc916104e2308284167bfc554c6894bf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 302, + 553, + 445 + ], + "blocks": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 299 + ], + "type": "text", + "content": "Figure 6. Examples of ALE-Bench. Source images are shown with object binary masks. Edit types are color, object, material modifications, and their combinations. Rows depict 1-, 2-, and 3-object edits, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 302, + 553, + 445 + ], + "lines": [ + { + "bbox": [ + 56, + 302, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 553, + 445 + ], + "type": "table", + "html": "
Source ImageColorObjectMaterialColor + ObjectObject + Material
tree→ blue-colored treepine treetree made of diamondyellow-colored Christmas treepine tree made of bamboo
couch→ purple-colored couchbenchcouch made of leatherkhaki-colored deskdesk made of chrome
floor→ khaki-colored floorlawnfloor made of turquoiseyellow-colored tiletile made of jade
", + "image_path": "4fa94a571a6825a3b3fbf8a6a430a25f809d4df85c0ec0ed4ae089569023a24b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "text", + "content": "Figure 7. Qualitative examples of ALE applied to three source objects (tree, couch, and floor) under different editing types. The bottom rows show the corresponding prompts describing each transformation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 483, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 556 + ], + "type": "text", + "content": "editing types (Table 3), ALE demonstrates its robust performance. Figure 7 illustrates qualitative examples of ALE for various editing types. These results indicate that our method can effectively address both TEL and TIL. Detailed results are in Appendix D. Furthermore, the quantitative and qualitative results on PIE-Bench are in Appendix E." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 573, + 115, + 632 + ], + "blocks": [ + { + "bbox": [ + 57, + 573, + 115, + 632 + ], + "lines": [ + { + "bbox": [ + 57, + 573, + 115, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 573, + 115, + 632 + ], + "type": "image", + "image_path": "8aeb21f30d8130acb860a8f2a3596564895f0e53be614c0a181014a073ff0378.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 633, + 114, + 643 + ], + "lines": [ + { + "bbox": [ + 57, + 633, + 114, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 633, + 114, + 643 + ], + "type": "text", + "content": "(a) Source image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 116, + 573, + 175, + 632 + ], + "blocks": [ + { + "bbox": [ + 116, + 573, + 175, + 632 + ], + "lines": [ + { + "bbox": [ + 116, + 573, + 175, + 632 + ], + "spans": [ + { + "bbox": [ + 116, + 573, + 175, + 632 + ], + "type": "image", + "image_path": "a5a8e8c1c91f26e2029738dd7b93dc757d1fdc87b653999d5378a3e9fd50f60a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 633, + 156, + 643 + ], + "lines": [ + { + "bbox": [ + 133, + 633, + 156, + 643 + ], + "spans": [ + { + "bbox": [ + 133, + 633, + 156, + 643 + ], + "type": "text", + "content": "(b) BB" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 175, + 573, + 233, + 632 + ], + "blocks": [ + { + "bbox": [ + 175, + 573, + 233, + 632 + ], + "lines": [ + { + "bbox": [ + 175, + 573, + 233, + 632 + ], + "spans": [ + { + "bbox": [ + 175, + 573, + 233, + 632 + ], + "type": "image", + "image_path": "8a95e448b9c30e8e105ed2cb2414026fc3630be3f72b61368822421f1b772b55.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 633, + 231, + 643 + ], + "lines": [ + { + "bbox": [ + 178, + 633, + 231, + 643 + ], + "spans": [ + { + "bbox": [ + 178, + 633, + 231, + 643 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 178, + 633, + 231, + 643 + ], + "type": "inline_equation", + "content": "\\mathrm{ORE} + \\mathrm{RGB}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "type": "text", + "content": "Figure 8. Qualitative ablation results for the editing: yellow bell pepper " + }, + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "type": "text", + "content": " diamond, red bell pepper " + }, + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 650, + 296, + 694 + ], + "type": "text", + "content": " moon. (b) BB resolves TEL but not TIL. (c) ORE + RGB-CAM reduces TIL but not TEL. (d) ALE (BB + ORE + RGB-CAM) shows no TIL and TEL." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 573, + 293, + 632 + ], + "blocks": [ + { + "bbox": [ + 235, + 573, + 293, + 632 + ], + "lines": [ + { + "bbox": [ + 235, + 573, + 293, + 632 + ], + "spans": [ + { + "bbox": [ + 235, + 573, + 293, + 632 + ], + "type": "image", + "image_path": "14f3931906e9b8449240c78b1047e456bbc0995370a0de8a5c9f87bab07bdb1e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 633, + 278, + 643 + ], + "lines": [ + { + "bbox": [ + 249, + 633, + 278, + 643 + ], + "spans": [ + { + "bbox": [ + 249, + 633, + 278, + 643 + ], + "type": "text", + "content": "(d) ALE" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 483, + 416, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 416, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 416, + 495 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 501, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 645 + ], + "type": "text", + "content": "We present ablation results on BB, ORE, and RGB-CAM in Figure 8. Each module plays a complementary role in preventing attribute leakage. Using only BB (Figure 8b) helps preserve background regions and suppresses TEL, but fails to prevent TIL. On the other hand, " + }, + { + "bbox": [ + 313, + 501, + 555, + 645 + ], + "type": "inline_equation", + "content": "\\mathrm{ORE} + \\mathrm{RGB}" + }, + { + "bbox": [ + 313, + 501, + 555, + 645 + ], + "type": "text", + "content": "-CAM (Figure 8c) reduces TIL by disentangling text embeddings and aligned cross-attention. However, without BB, it cannot preserve the original background, and TEL remains. Only when all three modules—ORE, RGB-CAM, and BB—are used together (Figure 8d), ALE achieves precise and leakage-free editing. Qualitative results and more examples are provided in Appendix E." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 656, + 389, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 656, + 389, + 669 + ], + "spans": [ + { + "bbox": [ + 314, + 656, + 389, + 669 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "While ALE-Bench provides a focused framework for evaluating attribute leakage, and ALE achieves strong performance on rigid attribute edits, both are currently lim" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16457" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 161 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 161 + ], + "type": "table", + "html": "
MethodTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
P2P21.5217.260.151420.6711.150.44950.08790.5589
MasaCtrl20.1816.740.092920.0114.990.29300.04180.7346
FPE21.0717.380.116421.8912.820.39030.06560.6052
InfEdit19.5916.690.048421.7816.740.20340.03400.7709
ALE16.0315.280.016722.2030.040.03610.00140.9228
", + "image_path": "22a9acea6edced92cf72efb101418b43b7e1c6991c1a3aecabe735b6a226aa17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 200, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 167, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 553, + 191 + ], + "type": "text", + "content": "Table 1. Comparison of editing performances on ALE-Bench. ALE demonstrates the lowest attribute leakage, highest structure preservation, and superior editing performance, indicating a more precise and controlled editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 200, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 553, + 266 + ], + "type": "table", + "html": "
# of Editing ObjectsTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
116.41-0.0087622.6230.010.04050.001670.9049
216.0015.420.0164822.0630.060.03600.001460.9235
315.8915.360.0246022.1930.010.03230.001540.9426
", + "image_path": "a546bc681b2a127d94cc00a14335a2a232f2d0b31de27de3955f3a111155c65f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 305, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 272, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 297 + ], + "type": "text", + "content": "Table 2. Performance of ALE on ALE-Bench based on the number of objects edited. Our method maintains low attribute leakage and strong background preservation even as the number of editing objects increases." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 305, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 305, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 553, + 392 + ], + "type": "table", + "html": "
Editing TypeTELS ↓TILS ↓Structure Distance ↓Editing Performance ↑Background Preservation
PSNR ↑LPIPS ↓MSE ↓SSIM ↑
Color17.6316.210.0089023.1232.970.02880.000790.9309
Material17.1515.960.0117922.9430.630.03390.001200.9248
Object15.8616.250.0197421.8229.030.03860.001820.9218
Color+Object15.3014.010.0230622.1528.600.04070.002050.9206
Object+Material14.5514.510.0195621.4228.880.03930.001910.9201
", + "image_path": "952084cb4b842e1894adf37255429ef3e0b73b9c563896f6092ec4c148a76de8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 555, + 421 + ], + "type": "text", + "content": "Table 3. Performance of ALE on ALE-Bench across various editing types (color, object, material, and combinations). The results show consistent low attribute leakage and high editing performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 441, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 610 + ], + "type": "text", + "content": "ited to local and relatively simple transformations—such as changes in color, object identity, or material. They do not support or evaluate non-rigid transformations like style transfer, pose changes, or adding/deleting objects, where defining and detecting attribute leakage becomes ambiguous. This focus reflects our aim to establish a clear and measurable foundation before addressing more complex editing scenarios. Although the benchmark offers 3,000 detailed editing scenarios across 20 carefully curated images, the small image set may limit how well results generalize to models trained on larger or more diverse datasets. Future work could expand both the editing model and benchmark to support richer transformations and broader datasets. More limitations are discussed in Appendix F." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 620, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 128, + 634 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we addressed the issue of attribute leakage in diffusion-based image editing, focusing on two types of leakage: Target-External Leakage (TEL)—unintended edits in non-target regions—and Target-Internal Leakage (TIL)—interference among attributes of different targets. To mitigate these problems, we introduced Attribute" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 537 + ], + "type": "text", + "content": "Leakage-Free Editing (ALE), a tuning-free framework that combines three key components: Object-Restricted Embeddings (ORE), which localize attribute semantics to each target object embedding; Region-Guided Blending for Cross-Attention Masking (RGB-CAM), which constrains cross-attention with segmentation masks to avoid unintended inter-object attribute mixing; and Background Blending (BB), which preserves the source image in backgrounds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 537, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 556, + 609 + ], + "type": "text", + "content": "We also presented ALE-Bench, a dedicated benchmark for rigorously evaluating attribute leakage across diverse multi-object editing scenarios. ALE-Bench introduces new quantitative metrics—TELS and TILS—that effectively quantify unintended modifications, providing comprehensive measures to assess editing fidelity and consistency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 609, + 557, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 557, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 557, + 706 + ], + "type": "text", + "content": "Our extensive experimental validation demonstrated that ALE significantly outperforms existing tuning-free editing methods, achieving state-of-the-art performance by effectively minimizing attribute leakage while maintaining high editing quality and structural consistency. By effectively addressing attribute leakage with minimal computational overhead, ALE enhances the reliability and precision of multi-object image editing tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16458" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) and IITP-ITRC (Information Technology Research Center) grant funded by the Korea government (MSIT) (No.RS-2019-II191906, Artificial Intelligence Graduate School Program (POSTECH); No.RS-2021-II212068, Artificial Intelligence Innovation Hub; IITP-2025-00437866; RS-2024-00509258, Global AI Frontier Lab) and partly supported by Seoul R&BD Program through the Seoul Business Agency (SBA) funded by The Seoul Metropolitan Government (SP240008)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 115, + 246 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 253, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 253, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 253, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 253, + 294, + 286 + ], + "type": "text", + "content": "[1] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 288, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 342 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 343, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 408 + ], + "type": "text", + "content": "[3] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22560-22570, 2023. 2, 3, 4, 6, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 295, + 453 + ], + "type": "text", + "content": "[4] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 42(4):1-10, 2023. 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 455, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 509 + ], + "type": "text", + "content": "[5] Wenkai Dong, Song Xue, Xiaoyue Duan, and Shumin Han. Prompt tuning inversion for text-driven image editing using diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7430-7440, 2023. 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 511, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 294, + 553 + ], + "type": "text", + "content": "[6] Adham Elarabawy, Harish Kamath, and Samuel Denton. Direct inversion: Optimization-free text-driven real image editing with diffusion models. arXiv preprint arXiv:2211.07825, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 294, + 577 + ], + "type": "text", + "content": "[7] Chakrabarty et al. Lomoe: Localized multi-object editing via multi-diffusion. In Proc. ACM Multimedia, 2024. 6, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 578, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 294, + 600 + ], + "type": "text", + "content": "[8] Luo et al. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 623 + ], + "type": "text", + "content": "[9] Yu et al. Zero-shot referring image segmentation with global-local context features. In Proc. CVPR, 2023. 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 60, + 624, + 295, + 679 + ], + "type": "text", + "content": "[10] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 3, 11" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[11] Aaron Grattafori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 94 + ], + "type": "text", + "content": "et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 148 + ], + "type": "text", + "content": "[12] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6986-6996, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "type": "text", + "content": "[13] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 6, 11" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 259 + ], + "type": "text", + "content": "[14] Taihang Hu, Linxuan Li, Joost van de Weijer, Hongcheng Gao, Fahad Shahbaz Khan, Jian Yang, Ming-Ming Cheng, Kai Wang, and Yaxing Wang. Token merging for training-free semantic binding in text-to-image synthesis. Advances in Neural Information Processing Systems, 37:137646-137672, 2024. 2, 4, 11, 13, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 316 + ], + "type": "text", + "content": "[15] Mingzhen Huang, Jialing Cai, Shan Jia, Vishnu Suresh Lokhande, and Siwei Lyu. Paralleledits: Efficient multi-aspect text-driven image editing with attention grouping. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 369 + ], + "type": "text", + "content": "[16] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 414 + ], + "type": "text", + "content": "[17] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Direct inversion: Boosting diffusion-based editing with 3 lines of code. arXiv preprint arXiv:2310.01506, 2023. 2, 6, 13" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "type": "text", + "content": "[18] Senmao Li, Joost Van De Weijer, Taihang Hu, Fahad Shahbaz Khan, Qibin Hou, Yaxing Wang, and Jian Yang. Stylediffusion: Prompt-embedding inversion for text-based editing. arXiv preprint arXiv:2303.15649, 2023. 11" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "text", + "content": "[19] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7817–7826, 2024. 3, 11" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 559 + ], + "type": "text", + "content": "[20] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "text", + "content": "[21] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 679 + ], + "type": "text", + "content": "[22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[23] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16459" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 708 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "through attention map alignment. Advances in Neural Information Processing Systems, 36:3536-3559, 2023. 11" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[24] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kun-chang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[25] Patsorn Sangkloy, Jingwan Lu, Chen Fang, Fisher Yu, and James Hays. Scribbler: Controlling deep image synthesis with sketch and color. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5400-5409, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 271 + ], + "type": "text", + "content": "[26] Narek Tumanyan, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Splicing vit features for semantic appearance transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10748-10757, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 294, + 350 + ], + "type": "text", + "content": "[27] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J Fleet, Radu Soricut, et al. Imagen editor and editbench: Advancing and evaluating text-guided image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18359-18369, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 294, + 396 + ], + "type": "text", + "content": "[28] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "type": "text", + "content": "[29] Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language. arXiv preprint arXiv:2312.04965, 2023. 2, 3, 4, 5, 6, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "text", + "content": "[30] Fei Yang, Shiqi Yang, Muhammad Atif Butt, Joost van de Weijer, et al. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. Advances in Neural Information Processing Systems, 36:26291–26303, 2023. 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 529 + ], + "type": "text", + "content": "[31] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. Magicbrush: A manually annotated dataset for instruction-guided image editing. In Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 586 + ], + "type": "text", + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 653 + ], + "type": "text", + "content": "[33] Jun-Yan Zhu, Philipp Krahenbuhl, Eli Shechtman, and Alexei A Efros. Generative visual manipulation on the natural image manifold. In Computer vision-ECCV 2016: 14th European conference, Amsterdam, the netherlands, October 11-14, 2016, proceedings, part v 14, pages 597-613. Springer, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 654, + 294, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 294, + 708 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 294, + 708 + ], + "type": "text", + "content": "[34] Siyu Zou, Jiji Tang, Yiyi Zhou, Jing He, Chaoyi Zhao, Rongsheng Zhang, Zhipeng Hu, and Xiaoshuai Sun. Towards efficient diffusion-based image editing with instant attention masks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7864-7872, 2024. 2, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16460" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_content_list.json b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a78dbe04f1bb0a882e18d26840a430f31c8ed2dd --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "AdsQA: Towards Advertisement Video Understanding", + "text_level": 1, + "bbox": [ + 222, + 130, + 774, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinwei Long $^{1}$ Kai Tian $^{1*}$ Peng Xu $^{1\\boxtimes}$ Guoli Jia $^{1}$ Jingxuan Li $^{2}$ Sa Yang $^{3}$ Yihua Shao $^{4}$ Kaiyan Zhang $^{1}$ Che Jiang $^{1}$ Hao Xu $^{5}$ Yang Liu $^{2}$ Jiaheng Ma $^{2}$ Bowen Zhou $^{1,6\\boxtimes}$ $^{1}$ Tsinghua University, $^{2}$ Independent Researcher, $^{3}$ Peking University, $^{4}$ CASIA $^{5}$ Harvard University, $^{6}$ Shanghai Artificial Intelligence Lab \n{longxw22, tk23}@ mails.tsinghua.edu.cn; {peng_xu, zhoubowen}@tsinghua.edu.cn \nProject page: https://github.com/TsinghuaC3I/AdsQA \nMARS2 workshop & challenge: https://lens4mllms.github.io/mars2-workshop-iccv2025/", + "bbox": [ + 158, + 178, + 834, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 356, + 326, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 toptier LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin.", + "bbox": [ + 89, + 388, + 483, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A recent milestone flagged by OpenAI o1 [25] and DeepSeek-R1 [21] has been in the spotlight, and already opened up an era of large reasoning models. Meanwhile, going beyond general domains, the specialized do", + "bbox": [ + 89, + 818, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mains [13, 93] are increasingly studied to boost these general-purpose LLMs to evolve continuously. The specialized domains contribute significantly in optimizing and evaluating generalist models towards specialized reasoning like domain experts [18, 52]. Recently, a clear phenomenon [20, 84] has been observed that LLMs are already good at reasoning a kind of step-wise problems represented by math and programming. As has been widely discussed [20, 84], the step-wise problems based on explicit theorems and programming syntax are compatible with the chain-style reasoning approach ``if A, then B'. To further extend the diversity of specialized reasoning for LLMs, it would be better to exploit other novel domains where the domain-specific reasoning is unmanageable for \"if A, then B\".", + "bbox": [ + 511, + 358, + 906, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, we, for the first time, propose to use advertisement videos to probe the reasoning boundary of LLMs in perceiving implicit multimodal reasoning. The domain-unique features of clue-rich and information-dense ad videos can be summarized as key words implicit, nonphysical, mental, heuristic, etc. Different from user-uploaded videos on social media, ads are typically meticulously crafted by commercial or non-commercial organizations, making them entertaining, creative, aesthetically appealing, and capable of offering enjoyment and attracting viewer engagement.", + "bbox": [ + 511, + 575, + 908, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on these domain-specific advantages of ad videos, we introduce AdsQA, a comprehensive and carefully curated VideoQA benchmark derived from 1,544 ad videos containing 10,962 clips, spanning a total of 22.7 hours of video content. The AdsQA introduces five tasks, each requiring different types and levels of reasoning: (1) Visual Concept Understanding: Identify and analyze visual elements in ads. (2) Emotion Recognition: Detect emotions and infer their roles in ads. (3) Theme and Core Message Extraction: Summarize the central theme and key messages", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Kai Tian has equal contribution with Xinwei Long.", + "bbox": [ + 107, + 886, + 385, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23396", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a2d7424407d699f5bae0578200c1caa3d9b56e9f5e6530d8f365cfe61d06597.jpg", + "image_caption": [ + "Figure 1. Overview of AdsQA benchmark. Subfits A - B: statistics & diversity, and examples." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 359, + 440 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0f358a0cf58c7e5d6515b1f8e8307bfe04f1fbbdf5ac2a886a1e1846d70e9d69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 369, + 97, + 630, + 439 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6c02812ced8ce6bd3c1969ab7c8519b559f914be6dcd90aebe080c99e0eea425.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 98, + 897, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of the ad. (4) Persuasion Strategy Mining: Analyze the strategies used to persuade the audience. (5) Potential Audience Modeling: Identify and characterize the target audience. The five tasks are formalized as the open-ended QA format. During the construction process, we propose an innovative Role-Played Multi-Agent Annotation framework, which simulates the role of advertising experts to generate specialized data and significantly reduces the workload of human annotators. To ensure data quality, we conduct multiple rounds of automated data cleaning and manual data revision. Moreover, based on our AdsQA, we benchmark 14 well-known LLMs, including the GPT-4 level models, recently-released reasoning-based models, etc.", + "bbox": [ + 88, + 494, + 483, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Humans process visual stimuli from ads and naturally form mental impressions, rather than relying on rigid logical reasoning templates, e.g., \"if A, then B\". Inspired by this heuristic process, we propose ReAd-R, a DeepseekR1 styled Reinforced Ad Reasoner, which learns to gather effective visual stimuli from ad videos, then reflects on the questions, and finally answers the given questions in a human-like manner. Specifically, ReAd-R utilizes a reward function to evaluate the model's responses and adjusts its parameters based on correctness, facilitating learning through trial and error. As a result, ReAd-R enhances reasoning abilities through outcome-reward-driven optimization, eliminating the need for costly step-wise su", + "bbox": [ + 88, + 703, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pervision or chain-of-thought (COT) training data. Experiments demonstrate that our model brings reasoning ability an obvious gain in comprehending implicit logic of ad videos. In contrast, other reasoning-based methods obtain limited improvement due to fixed COT templates and suboptimal process reward models.", + "bbox": [ + 511, + 494, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To our knowledge, our main contribution can be stated as: (1) AdsQA is the first video QA benchmark for advertisement domain, which is also the first ad benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current mainstream 'if A, then B' LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. Moreover, it advances ad video understanding beyond physical-content dominated shallow perception towards deeper cognitive reasoning.", + "bbox": [ + 511, + 598, + 906, + 750 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, which enhances the specialized reasoning capability to understand implicit logic in ad videos like the human thinking process from perception to cognition. ReAd-R also can be regarded as one of the earliest attempts evolving R1-technique to vision research.", + "(3) We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning (e.g., variants of GPT4, LLaVA, and Qwen) by a clear margin." + ], + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23397", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Question Answering Benchmarks. Video question answering (Video QA) [7, 48, 55, 58, 68] aims to answer user questions based on a given video, requiring a detailed understanding of both the spatio-temporal information and the relationships between objects and events [5, 12, 66, 74]. The Video QA community has developed a variety of benchmarks. From the perspective of video content, most benchmarks focus on human-centric videos, such as NextQA [65], ActivityQA [77], MVBench [36], FunQA [67], and VideoMME [16], which are derived from movies [19, 59], TV shows [32], social media [67], etc. Additionally, there are benchmarks centered on object-centric videos, such as EgoSchema [47], which are sourced from instructional and operational videos [38, 92]. Despite their advancements, these benchmarks still have several limitations. MovieQA [59] overly relies on dialogue understanding, which restricts the in-depth comprehension of visual content. Although TGIF-QA [26] requires some complex reasoning, the GIFs used are typically no longer than 3 seconds. ActivityQA [77] and Next-QA [65] have introduced open-ended QA tasks, but the annotated answers are often overly simplistic, usually consisting of just a few words. To sum up, current video QA benchmarks face limitations in video diversity and the setups of QA pairs.", + "bbox": [ + 93, + 114, + 480, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video QA Methods and Video-LLMs. Early video QA methods employed graph structures [29, 56, 90] or transformers [17, 33, 37] to model the spatiotemporal relationships in videos and capture interactions between humans and objects. Recently, Video-LLMs [6, 45, 46] have demonstrated impressive capabilities in video understanding. Representative Video-LLMs include the VideoLLaMA [9, 47] series, LLaVA series [34, 39, 41, 70, 86], InternLM series [62, 85], and Qwen series [4, 60]. Video-LLMs take both the question and the video as input and directly generate the answer, implicitly performing the reasoning and thinking process in an efficient manner. Moreover, Some methods attempt to explicitly model the reasoning process through step-by-step thinking or multi-round debates. For example, some studies [15, 22, 57, 63, 81] leverages chain-of-thought to simulate the thinking process, while VideoAgent [14, 61, 73] improves reasoning capabilities through agent collaboration.", + "bbox": [ + 93, + 478, + 480, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Advertisement Understanding. Internet companies generate substantial profits by automatically distributing advertisements to target users, making the understanding of ad content highly important [28, 79]. The Pit dataset [24], one of the earliest efforts in automatic ad understanding, formalized this task as a visual question-answering (VQA) problem [43]. Subsequent research has built upon the Pit dataset or further developed it [53], or utilized their own private datasets [72]. These studies have mainly focused on a single aspect of advertisement understanding, such as persua", + "bbox": [ + 89, + 750, + 480, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sive strategies in image ads [31, 76], image ad search [89], intent understanding [27], and visual metaphor comprehension [3, 53, 71, 82, 83]. Although the Pit dataset primarily focuses on image advertisements, it also provided a subset of video advertisements they collected. However, the Pit dataset suffers from issues such as data inaccessibility, lack of diversity, and limited Q&A formats. Therefore, there is currently no comprehensive Video QA benchmark available in the advertising domain.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement Learning for Reasoning. Recent research efforts [25, 30, 44, 94] have attempted to improve the reasoning capabilities of LLMs through reinforcement learning. Recently, DeepSeek-R1 [21] achieved significant improvements in reasoning based on reinforcement learning, eliminating the need for intermediate reasoning signals. S1 [49] demonstrated that even with only a few hundred data points, the model can effectively perform reasoning in specialized tasks. However, current research on RL-based reasoning primarily focuses on math and code problems, with few works [42] attempting to apply it to multimodal domains, particularly open-ended video QA tasks.", + "bbox": [ + 511, + 229, + 903, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The AdsQA Benchmark", + "text_level": 1, + "bbox": [ + 511, + 430, + 736, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition", + "text_level": 1, + "bbox": [ + 511, + 455, + 666, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To comprehensively evaluate the model's ability to understand ad videos, we divided the questions into five types, with each one targeting a specific angle of ad analysis.", + "bbox": [ + 511, + 479, + 903, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Concept Understanding (VU) task evaluates the model's ability to comprehend specific visual concepts, such as characters, objects, scenes, slogans, and other details, as well as their interrelationships.", + "bbox": [ + 511, + 527, + 903, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Emotion Recognition (ER) assesses if the models understand what emotion the ad evokes, and how ads establish connections with audiences through emotions.", + "bbox": [ + 511, + 590, + 903, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theme and Core Message Extraction (TE) drives the models to extract the underlying message or central idea that the ad explores. This task requires deep reasoning to gather visual information from the full video.", + "bbox": [ + 511, + 637, + 903, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Persuasion Strategy Mining (PS) task evaluates the model's ability to uncover the strategies used to convey core messages and persuade audiences, such as humor, exaggeration, and visual rhetoric. The task includes questions, such as: how an ad conveys its central message, why the ad is appealing, and what strategies the ad employs.", + "bbox": [ + 511, + 700, + 903, + 791 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Potential Audience Modeling (AM) probes the model's performance in identifying potential audience groups and their profiles. It sets key questions, such as who the ad targets and what the characteristics of the audience are. This task directly reflects the ad's influence and value.", + "bbox": [ + 511, + 792, + 903, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These five tasks are defined as the open-ended video QA and each question is annotated with a ground-truth answer.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23398", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dataset Construction Pipeline", + "text_level": 1, + "bbox": [ + 89, + 90, + 357, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AdsQA benchmark construction pipeline is in three stages: Pre-Processing. To serve the ad video understanding, we consider the creativity, aesthetic quality, and availability of videos during our collection process. We do not use any private data; instead, we collect creative ad videos that adhere to the Creative Commons License from a creative community platform [1]. This platform offers publicly accessible, high-quality ad videos uploaded by creators. Specifically, we first crawl videos along with their metadata. The metadata is typically written by the uploader (i.e., usually the creator of the ad) and includes information such as the theme, content, and key creative elements of the ad. It can be considered the \"ground-truth\" information of the ad video and serves as an important reference for our benchmark construction. Then, human experts are asked to carefully review the completeness and accuracy of the metadata and manually remove incomplete samples. Additionally, we conduct automated filtering based on video duration, aesthetic score, and content to ensure video quality and exclude non-ethical and negative content.", + "bbox": [ + 89, + 112, + 480, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering efficiency and accuracy, we employ PySceneDetect to segment ad videos into fine-grained video clips to avoid missing detailed information in subsequent steps. For each clip $C_i$ , we use Video-LLMs to generate its description $Desc_i$ . To preserve key visual elements, we sample $n$ keyframes $\\{F_i^0, \\dots, F_i^{N_{max}}\\}$ from each clip based on frame similarity calculated using SSIM [64], where $n$ is determined by the duration of each clip. To retain speech information, we extract speech information $Asr_i$ for each clip using the Whisper model, and translate them into English using GPT-4 [23]. Therefore, an ad video with $N$ clips can be represented as a modality-interleaved sequence, as Eq. 1,", + "bbox": [ + 89, + 415, + 482, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = \\left\\{M, \\left\\{F _ {i} ^ {0}, \\dots , F _ {i} ^ {n _ {i}}, D e s c _ {i}, A s r _ {i} \\right\\} _ {i = 1} ^ {N} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 618, + 480, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ denotes the meta-information of the video. The modality-interleaved sequence will be used for automated annotation generation.", + "bbox": [ + 89, + 643, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Role-Played Multi-Agent Annotation. To balance data quality and cost, ensure data diversity, and avoid template-driven outputs, we propose a Role-Played Multi-Agent Annotation framework to automatically generate video QA pairs. Inspired by previous research [78], we found that agents can develop specialized capabilities in the ad domain by configuring specific skill sets, such as marketing, visual design, and consumer psychology. Therefore, we can leverage AI agents to act as ad experts, analyzing diverse facets of ad videos and designing more specialized and challenging questions. We ask human experts to create profiles for advertising expert skills, enabling AI agents to adaptively select profiles (or autonomously generate new profiles) to accomplish role-playing tasks. Additionally, human experts", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "provide examples of QA pairs as in-context demonstration to guide the agents in generating appropriate QA pairs.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Role-Played Multi-Agent Annotation is formulated into a three-stage pipeline. (1) In the initial stage, a master agent is recruited to generate a preliminary QA annotation for the given modality-interleaved sequence $V$ . (2) In the iterative stage, the master agent checks the quality of the current QA annotations and decides whether to recruit a specialized expert agent. If necessary, the master agent selects a profile for the expert agent and instantiates it accordingly. The expert agent then generates new QA annotations leveraging its specialized expertise. Finally, the master agent revises the original annotations based on the expert agent's provided annotations and assesses whether to terminate the iteration or continue to recruit a new expert agent. (3) After terminating the iteration, the master agent synthesizes outputs from previous iterations to produce the final QA annotation.", + "bbox": [ + 511, + 122, + 903, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Automated Data Cleaning. After automated annotation through multi-agent collaboration, we employ the IXC-2.5-Reward model [80] to automatically verify the correctness of the answers. Given the video, question, and meta information, the reward model scores the answer. QA pairs with scores below a certain threshold are flagged and carefully reviewed by annotators.", + "bbox": [ + 511, + 364, + 903, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Manual Check and Quality Control", + "text_level": 1, + "bbox": [ + 511, + 481, + 823, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the automatic annotation generation, we conduct a careful manual check and revision on all Q&A pairs, focusing on the question suitability, annotation correctness, video content, and task difficulty. To ensure the quality of Q&A pairs, we first remove questions that are homogeneous, unrelated to the ads' theme, or cannot be answered based on the ad video (e.g., those requiring external knowledge). Then, we incorporate the meta information as ground truth to verify the accuracy of the reference answers. We eliminated pairs containing inaccurate, ambiguous, or biased answers. After two rounds of rigorous selection, only $37\\%$ of the QA pairs remain. Additionally, we manually revise the wording of some QA pairs to improve their clarity and accuracy. Any modifications to a sample are reviewed by other annotators to ensure consistency and reliability.", + "bbox": [ + 511, + 503, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Dataset Statistics", + "text_level": 1, + "bbox": [ + 511, + 742, + 681, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Videos. Our AdsQA Benchmark includes a total of 1,544 unique advertisements, comprising 10,962 video clips for evaluation. As shown in Fig. 2a, these ad videos span 9 primary domains. The video durations range from 15 to 120 seconds, with an average length of 52.9 seconds, amounting to a total of 22.7 hours of content. Each ad video is accompanied by metadata, including a title, tags, automatic speech recognition (ASR) results, and descriptions. This metadata is utilized for Q&A generation, as described in Sec. 3.1.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23399", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9c3eb2e16091b492d57aebf28575e025d78491d315cb07edc4afcecefae0c0a3.jpg", + "image_caption": [ + "(a) Alluvial diagram of task and domain distribution." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 364, + 388 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34870af13fa888f42083da9d1df9b525871bc63d5f9f75ad7b8cd32135046149.jpg", + "image_caption": [ + "(b) Distribution of video release years." + ], + "image_footnote": [], + "bbox": [ + 372, + 103, + 633, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06177250c26462f1fdd9a3fcae91f072a445bd4ccf434c79515d1c45d42fb4d9.jpg", + "image_caption": [ + "(c) Distribution of video release region." + ], + "image_footnote": [], + "bbox": [ + 656, + 102, + 836, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd712e94ee7232e8a170b984b209dd73b019e3aaa29da2e203d084256af2ff6a.jpg", + "image_caption": [ + "(d) The distribution of the video duration in the dataset." + ], + "image_footnote": [], + "bbox": [ + 377, + 263, + 609, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f31804cdf242274beb7906f6fc7320d21de1b1197405bf284b36e8a948429c7b.jpg", + "image_caption": [ + "(e) Number of Q&A entries for different tasks." + ], + "image_footnote": [], + "bbox": [ + 622, + 262, + 859, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most of the ad videos are in English and primarily originate from North America and Europe.", + "bbox": [ + 89, + 460, + 482, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q&A Pairs. Our AdsQA Benchmark includes 7,859 QA pairs (among which $29.2\\%$ of the questions can be categorized into two types, resulting in a total of 10,154 QA pairs for evaluation). Among these: Visual Concept Understanding accounts for $21.4\\%$ of the QA pairs, Theme and Message Extraction accounts for $28.5\\%$ , Persuasion Strategy Mining accounts for $25.3\\%$ , Potential Audience Modeling accounts for $15.1\\%$ , and Emotion Recognition accounts for $9.8\\%$ . Typical persuasion strategies include metaphor, symbolism, humor, expert opinion, and others. Some questions may fit into two categories, such as \"What is the role of a particular visual concept in relation to the theme?\" On average, each question consists of 17 words and each answer contains approximately 12 words.", + "bbox": [ + 89, + 492, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 722, + 225, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, we propose ReAd-R, a DeepSeekR1 styled Reinforced Ad Reasoner, to simulate human heuristic thinking and learn from trial and error. The model takes an ad video and a question as input. The policy model generates a set of reasoning processes, including thoughts and answers, based on the given input. Each reasoning process is evaluated using a reward function to compute its reward value. After calculating the reward values for all outputs, each reasoning process is assessed and used to update the policy model. Additionally, ReAd-R employs KL di", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43d0446add253e528766c953740086ad6a20ae79ce139ce2224461df6420dab0.jpg", + "image_caption": [ + "Figure 2. Statistics of Our AdsQA Bench.", + "Reasoning: \"think>Loop's noise-reduction products, as inferred from the context, are intended to help individuals in noisy environments, such as subway rides, music parties, and bedrooms where noise can disturb others...nPeople who are concerned with noise reduction in their daily lives—such as commuters, festival-goers, night owls, or anyone sensitive to loud environments—would benefit the most from using Loop's noise-reducing earwear.", + "Figure 3. Framework of ReAd-R. Given a question and video, the policy model generates multiple responses. The reward model evaluates and scores them, and the rewards are used to update the policy model via policy gradient optimization." + ], + "image_footnote": [], + "bbox": [ + 527, + 458, + 890, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "vergence to control the difference between the policy model and the reference model, ensuring stable training.", + "bbox": [ + 511, + 776, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Preparation. Inspired by previous work [21, 49], RL-based reasoners can improve reasoning capabilities with limited high-quality data. We additionally crawled a total of 1,053 ad videos along with their metadata and automatically generated annotations using the same method. We utilized metadata tags collected during data crawling (e.g., topic,", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23400", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "domain) to ensure diversity, adopted Video-LLM evaluation results to assess difficulty, and performed manual quality assessment based on the same criteria as in Sec. 3.3. Finally, we selected 100 videos and 500 question-answer pairs. We designed a prompt format to guide the model to output its reasoning process before providing the final answer.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reward Modeling. The reward model guides the model to find the optimal direction through trial and error. Ideally, ReAd-R's responses should (1) include as many elements of the standard answer as possible and (2) avoid containing content not mentioned in the ad video (e.g., hallucinations). To achieve this, we propose a rule-guided LLM evaluator as the reward model to assess the quality of each response. We introduce two sets of rules: inclusion rules and exclusion rules. The inclusion rule requires that the generated answer incorporate as many elements of the standard answer as possible. If satisfied, the reward value is 1.0; otherwise, it is 0. The exclusion rule states that if the generated content includes elements not mentioned in the standard answer and these elements cannot be inferred from the meta-information, they should be judged as incorrect. To avoid sparse reward values in the early stages, we relaxed the inclusion rule. If the generated response partially includes factual elements, it is assigned a reward value of 0.5. Additionally, the format reward is used to enforce the model's predictions to adhere to the required format of and , as $R(\\cdot) = R_{\\text{ans}}(\\cdot) + R_{\\text{format}}(\\cdot)$ .", + "bbox": [ + 89, + 181, + 482, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reinforced Fine-tuning. Following DeepSeek-R1, we employ the GRPO algorithm instead of PPO to optimize our model. ReAd-R first generates $n$ distinct responses $O = \\{o_1, o_2, \\dots, o_n\\}$ from the current policy model $\\pi_{\\theta_{old}}$ . Then, the reward model $R(\\cdot)$ evaluates these responses $O$ to obtain their reward value as $\\{r_1, r_2, \\dots, r_n\\}$ , and the advantage $A_i$ is defined as", + "bbox": [ + 89, + 500, + 482, + 604 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} = \\frac {r _ {i} - \\operatorname {m e a n} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 614, + 480, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$A_{i}$ denotes relative quality of the $i^{th}$ answer when compared to the group of rewards $\\{r_1, r_2, \\ldots, r_n\\}$ . This strategy encourages the model to generate better answers with higher reward values within the group. We use GRPO algorithm [20] to optimize our policy model based on $A_{i}$ .", + "bbox": [ + 89, + 657, + 482, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 748, + 223, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 772, + 297, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluated Models. We select 14 top-tier LLMs as strong competitors covering diverse categories: (1) GPT-4 Level Baselines: including GPT-4o [23], GPT-4V [2] and Qwen2.5-VL-72B [4]. (2) Open-sourced Video LLMs: We present at least one representative model released in the past six months from each series of Video-LLMs, such as Qwen2.5-VL-7B [4], LLaVA-Onevision-7B [34], and oth", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers. (3) Reasoning Based Methods: We re-implemented several reasoning-based methods and applied them to the AdsQA task. These methods include the Video Chain-of-Thought (VOT) [15], the Role-Played Agent (EvoAgent) [78], and the Monte Carlo Tree Search (MCTS) algorithms [84]. Detailed descriptions and implementation methods are provided in the appendix.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. All five tasks of AdsQA are evaluated as open-ended QA, allowing for the free-text evaluation methods. Traditional free-text metrics [40, 51] are sensitive to lexical variations, which may introduce bias when evaluating different LLMs. Recent studies [8, 10, 19, 67] have shown promising results in using large generalist models to evaluate generated text. Therefore, we follow their framework and employ GPT-4o to assist in evaluating fretext similarity. Specifically, we use the same inclusion and exclusion rules mentioned in Sec. 4 to guide the scoring of text generated by the large model. We provided each sample with a ground-truth answer and meta-information; each meta-information includes relevant information such as the creative elements, content, storyline, and themes of the advertisement video. We propose both relaxed and strict scoring modes. Our strict accuracy Acc_strict is defined as:", + "bbox": [ + 511, + 196, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A c c} _ {\\text {s t r i c t}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 448, + 903, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}\\{\\cdot\\}$ is the indicator function that returns 1 if the model's response $x_{i}$ semantically matches all elements of the standard answer $y_{i}$ , and 0 otherwise.", + "bbox": [ + 511, + 496, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the relaxed mode, the relaxed accuracy $Acc_{\\mathrm{relaxed}}$ is defined as:", + "bbox": [ + 511, + 542, + 905, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A c c} _ {\\text {r e l a x e d}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\} + \\lambda \\cdot \\mathbb {1} \\left\\{x _ {i} \\approx y _ {i} \\right\\}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 580, + 903, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{1}\\{x_i\\approx y_i\\}$ returns 1 if the response $x_{i}$ partially addresses the elements of $y_{i}$ , and 0 otherwise. It is observed $\\lambda = 0.5$ works well in experiments. Our prompt template for model-based evaluation are in the appendix.", + "bbox": [ + 511, + 627, + 905, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our ReAd-R is model-agnostic. In our main experiments, we utilize Qwen2-VL-7B as the base model. We froze the parameters of the visual backbone and only fine-tuned the parameters of the language model. We conducted the experiments with a batch size of 4 and a learning rate of 1e-6 on a server with eight A100 GPUs, with a training duration of 12 hours. Additional details and hyperparameters are provided in the appendix.", + "bbox": [ + 511, + 688, + 905, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results and Observations", + "text_level": 1, + "bbox": [ + 511, + 818, + 745, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although GPT-4 achieves over $85\\%$ accuracy on other datasets such as Next-QA, its strict accuracy on our dataset is only $29.4\\%$ , and its relaxed accuracy is just $56.6\\%$ . This shows that our benchmark exceeds the capabilities of some", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23401", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/82a46d8678949b682964e7115aee82579d2b54d022ca73cc7921f7b8b8db381d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelStrict AccuracyRelaxed Accuracy
VUERTEPSAMOverallVUERTEPSAMOverall
Human Results*55.139.457.157.545.851.374.565.975.776.264.671.4
Commercial Large Multimodal Model
GPT-4o [23]24.926.532.632.331.029.450.157.862.455.254.856.6
Gemini-2.5-Pro* [11]36.044.040.429.333.335.959.269.064.054.462.460.7
Open-sourced Video-LLMs
VideoLLaMA2-7B [9]4.567.757.286.067.386.4821.629.028.422.627.225.2
InternLM-XComp.2.5-7B [85]11.210.716.311.912.112.634.735.941.534.137.336.5
LLaVA-OneVision-7B [34]11.811.616.213.715.114.035.839.543.236.841.739.1
LLaVA-Video-7B [87]14.014.418.715.217.316.137.841.645.138.143.641.0
MiniCPM-o_2.6-7B [75]13.015.317.314.918.916.338.143.445.739.145.042.5
Qwen2-VL-7B [60]13.715.420.216.020.017.236.841.346.037.744.641.0
Qwen2.5-VL-7B [4]20.223.224.620.524.323.045.850.451.845.350.948.9
Qwen2.5-VL-72B [4]26.730.234.829.834.731.051.857.459.553.859.255.8
Reasoning-based Models
VOT (Qwen2-VL-7B) [15]14.316.419.512.622.017.036.543.342.535.245.140.2
EvolAgent (Qwen2-VL-7B) [78]15.016.121.616.520.418.338.643.446.440.445.342.6
MCTSr (Qwen2-VL-7B) [84]14.810.219.316.116.417.140.641.346.541.142.643.4
Qwen2-VL-7B (SFT) [60]10.512.517.913.815.914.134.842.142.734.941.438.8
ReAd-R (Qwen2-VL-7B) (Ours)15.819.321.718.518.818.642.346.450.043.447.844.9
ReAd-R (Qwen2.5-VL-7B) (Ours)20.427.927.222.125.525.046.256.254.648.252.651.5
", + "bbox": [ + 94, + 88, + 903, + 467 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Experimental Results.", + "bbox": [ + 403, + 476, + 591, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "multi-modal large models. We observe that the state-of-the-art multi-modal models show strong visual perception, often generating partial answers by describing video segments. However, fully understanding the underlying meaning of ads is difficult for all baseline models. This requires not only identifying specific visual elements but also interpreting their implicit logic.", + "bbox": [ + 88, + 516, + 480, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different tasks vary in difficulty, but their performance is comparable. Though tasks focus on different aspects, these aspects are closely linked. For example, ads may use emotions or specific visuals to convey themes. Most models perform best on Theme and Message Extraction and Audience Modeling tasks. This is because ad videos aim to deliver clear messages to specific audiences, making it relatively easier to identify themes and target users. Even without fully understanding creative elements or marketing strategies, models can infer these from characters, scenes, or slogans in the ad. The Persuasion Strategy Mining task is more challenging. It requires models to explain how and why ads use certain designs or elements. These questions are often not directly expressed in the video and may even require interpreting counterfactual or unexpected visual information, which increases the difficulty of this task. Models also perform worse on the Visual Concept Understanding (VU) task. While other video QA benchmarks [77] focus on", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "surface-level questions like \"What color are the gloves?\", our VU task requires understanding global ad information, such as \"Which scene represents the ad's theme?\" Answering this question requires understanding the theme, identifying related scenes, and describing them briefly.", + "bbox": [ + 511, + 516, + 903, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the AdsQA task, reasoning models that excel in code and math tasks achieved only marginal results, even with ten times the computational cost. This shows that AdsQA reasoning differs significantly from tasks like math and code. Math and code rely on structured \"if A, then B\" reasoning, while ads require associative reasoning, connecting concrete visuals to abstract concepts. Prior work [91] also suggests chain-of-thought reasoning can hinder abstract reasoning performance. Though MCSTr improved results by $2.4\\%$ over Qwen2-VL, its search process is inefficient and highly dependent on the reward model. In some cases, Qwen2-VL answers correctly, but MCSTr fails due to incorrect search directions from the reward model. Moreover, training an effective reward model is difficult, especially in the data-scarce ad domain.", + "bbox": [ + 511, + 595, + 905, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method, ReAd-R, achieves a $3.9\\%$ improvement using only 500 video-question-answer triplets, outperforming Qwen2-VL and other reasoning models. During inference, our model reasons about video content and questions by generating free-text explanations, eliminating the need for", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23402", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/424e9fa799f80ed841a3d57493321b04081976f50e8193b3fae688c5d74a05cd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Strict Acc.Relaxed Acc.
ReAd-R (Our Model)18.644.9
w. Uncurated Data13.236.2
w. Strict Reward15.141.7
w.o. Prompt Constraints16.842.9
Qwen2-VL (SFT)14.138.8
w. Uncurated Data16.040.9
VOT (Chain length: 3)17.040.2
w. (Chain length: 5)13.636.2
", + "bbox": [ + 114, + 88, + 460, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study on ReAd-R, Qwen2-VL SFT, and VOT.", + "bbox": [ + 93, + 255, + 478, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fixed chain-of-thought (COT) templates (e.g., VOT), additional reward functions (e.g., MCTSr), or iterative search processes (e.g., MCTSr and EvolAgent). Interestingly, we observe that supervised fine-tuning (SFT) on the same data caused a sharp performance drop of $2.2\\%$ for Qwen2-VL. Using limited but high-quality data does not improve the performance of SFT models. SFT may overfit to the features of the training data, leading to poor generalization on diverse test sets. In contrast, our ReAd-R model improves reasoning capabilities through RL on limited data, achieving generalization across diverse advertisement videos.", + "bbox": [ + 89, + 295, + 483, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 470, + 243, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation studies for our ReAd-R model in Tab 2. We first show results of RL fine-tuning on 1,053 uncurated videos and 5,159 QA pairs. Our model performs worse than Qwen2-VL. This is due to the GRPO algorithm's sensitivity to data quality. Early in fine-tuning, low-quality data can mislead optimization, causing suboptimal results. Next, we test a strict reward rule (only fully correct answers receive a $R_{ans}$ reward of 1; otherwise, the reward is 0). Our method does not achieve improvement through RL fine-tuning. During training, we observe that the strict reward rule makes it hard for the model to earn rewards, limiting gains. We also find ReAd-R is sensitive to prompt templates. Properly constraining the model's thinking and answers (e.g., avoiding repetitive content in and tags) stabilizes training and improves performance.", + "bbox": [ + 89, + 492, + 483, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also conduct ablation studies on our implemented baselines (e.g., VOT, Qwen2-VL SFT). We present the results of Qwen2-VL fine-tuned using uncurated data. Interestingly, the model fine-tuned on uncurated data improves by $2.1\\%$ over limited-data fine-tuning but still lags behind the non-SFT model. This shows SFT does not enhance reasoning on ad data. Additionally, we experiment with VOT using a COT length of 5 (as set in its original paper) and observe a significant performance drop of $4.0\\%$ . The model does not benefit from step-by-step reasoning; instead, accumulated errors degrade its performance. This highlights the", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2f70e5ef148338af5e39fba0cddb1e3f378f09232e1682a79af06df15e52509.jpg", + "image_caption": [ + "Figure 4. Case Study. Due to the limited page space, we omit the model's thinking process and show only final answer." + ], + "image_footnote": [], + "bbox": [ + 526, + 88, + 893, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "difference between abstract reasoning in the ad domain and chain-style logical reasoning.", + "bbox": [ + 511, + 383, + 906, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Case Study", + "text_level": 1, + "bbox": [ + 511, + 431, + 635, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 4, Qwen2.5-VL generates vague and overly general answers (scored 0 by GPT-4). The COT baseline produces better answers (scored 0.5 by GPT-4) by gathering information step-by-step through a reasoning process, but it still misses some important clues in the video. Our ReAd-R learns to perform video-grounded reasoning through reinforcement fine-tuning, deriving the final answer by summarizing its reasoning process. The reinforced thinking-answer mechanism helps ReAd-R generate more complete and precise answers.", + "bbox": [ + 511, + 455, + 908, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 628, + 633, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We contibute AdsQA, the first advertisement benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current \"if A, then B\" LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, no need of costly supervision, which enhances the specialized reasoning capability to understand implicit logic in ad videos. We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities (e.g., variants of GPT4, LLaVA, and Qwen).", + "bbox": [ + 511, + 656, + 906, + 852 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported by the National Key Research and Development Program of China (2022ZD0160603), and NSFC No.62306162.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23403", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://www.adsofttheworld.com/.4", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6, 15", + "[3] Arjun R Akula, Brendan Driscoll, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T Freeman, et al. Metaclue: Towards comprehensive visual metaphors research. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23201-23211, 2023. 3", + "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-v1 technical report. arXiv preprint arXiv:2502.13923, 2025. 3, 6, 7, 15, 16", + "[5] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, Yao Dou, Jaden Park, Jianfeng Gao, Yong Jae Lee, and Jianwei Yang. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. CoRR, abs/2410.10818, 2024. 3, 16", + "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3", + "[7] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 3", + "[8] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 6", + "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. CoRR, abs/2406.07476, 2024. 3, 7, 15", + "[10] Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023. 6", + "[11] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. 7, 15", + "[12] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3", + "[13] Ning Ding, Shang Qu, Linhai Xie, Yifei Li, Zaoqu Liu, Kaiyan Zhang, Yibai Xiong, Yuxin Zuo, Zhangren Chen, Ermo Hua, et al. Automating exploratory pro" + ], + "bbox": [ + 93, + 114, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "teomics research via language models. arXiv preprint arXiv:2411.03743, 2024. 1", + "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 3", + "[15] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 3, 6, 7", + "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 3", + "[17] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14773-14783, 2023. 3", + "[18] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 23923-23932, 2025. 1", + "[19] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (SFD): A benchmark for story-level video understanding. CoRR, abs/2406.10221, 2024. 3, 6, 16", + "[20] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1, 6", + "[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 5", + "[22] Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding. arXiv preprint arXiv:2502.06428, 2025. 3", + "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4, 6, 7", + "[24] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In CVPR, pages 1100-1110. IEEE Computer Society, 2017. 3, 16", + "[25] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23404", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 1, 3", + "[26] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 3", + "[27] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12986-12996, 2021. 3", + "[28] Zhiwei Jia, Pradyumna Narayana, Arjun R Akula, Garima Pruthi, Hao Su, Sugato Basu, and Varun Jampani. Kafa: Rethinking image ad understanding with knowledge-augmented feature adaptation of vision-language models. arXiv preprint arXiv:2305.18373, 2023. 3", + "[29] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 3", + "[30] Fangkai Jiao, Geyang Guo, Xingxing Zhang, Nancy F Chen, Shafiq Joty, and Furu Wei. Preference optimization for reasoning with pseudo feedback. arXiv preprint arXiv:2411.16345, 2024. 3", + "[31] Yaman Kumar, Rajat Jha, Arunim Gupta, Milan Aggarwal, Aditya Garg, Tushar Malyan, Ayush Bhardwaj, Rajiv Ratn Shah, Balaji Krishnamurthy, and Changyou Chen. Persuasion strategies in advertisements. In Proceedings of the AAAI conference on artificial intelligence, pages 57-66, 2023. 3", + "[32] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379. Association for Computational Linguistics, 2018. 3, 16", + "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021. 3", + "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 6, 7, 16", + "[35] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Lou, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark. In CVPR, pages 22195-22206. IEEE, 2024. 16", + "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3", + "[37] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8658-8665, 2019. 3", + "[38] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In European Conference on Computer Vision, pages 200-216. Springer, 2024. 3", + "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 3", + "[40] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6", + "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3", + "[42] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025. 3", + "[43] Xinwei Long, Zhiyuan Ma, Ermo Hua, Kaiyan Zhang, Biqing Qi, and Bowen Zhou. Retrieval-augmented visual question answering via built-in autoregressive search engines. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 24723-24731, 2025. 3", + "[44] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967, 2024. 3", + "[45] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. 2023. 3", + "[46] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding. arxiv, 2024. 3", + "[47] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS, 2023. 3, 16", + "[48] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 3", + "[49] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. 3, 5", + "[50] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. CoRR, abs/2311.16103, 2023. 16", + "[51] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23405", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Biqing Qi, Kaiyan Zhang, Kai Tian, Haoxiang Li, ZhangRen Chen, Sihang Zeng, Ermo Hua, Hu Jinfang, and Bowen Zhou. Large language models as biomedical hypothesis generators: a comprehensive evaluation. arXiv preprint arXiv:2407.08940, 2024. 1", + "[53] Abisek Rajakumar Kalarani, Pushpak Bhattacharyya, and Sumit Shekhar. Seeing the unseen: Visual metaphor captioning for videos. arXiv e-prints, pages arXiv-2406, 2024. 3", + "[54] Ramon Sanabria, Ozan Caglayan, Shruti Palaskar, Desmond Elliott, Loic Barrault, Lucia Specia, and Florian Metze. How2: a large-scale dataset for multimodal language understanding. arXiv preprint arXiv:1811.00347, 2018. 16", + "[55] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 3", + "[56] Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection. arXiv preprint arXiv:2504.13092, 2025. 3", + "[57] Yudi Shi, Shangzhe Di, Qirui Chen, and Weidi Xie. Unlocking video-llm via agent-of-thoughts distillation. arXiv preprint arXiv:2412.01694, 2024. 3", + "[58] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 3", + "[59] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640. IEEE Computer Society, 2016. 3", + "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 15", + "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 3", + "[62] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3", + "[63] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool. arXiv preprint arXiv:2407.05355, 2024. 3", + "[64] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 9777-9786. Computer Vision Foundation / IEEE, 2021. 3", + "[66] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3", + "[67] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2024. 3, 6", + "[68] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025.3", + "[69] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645–1653, 2017. 16", + "[70] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3", + "[71] Yanzhi Xu, Yueying Hua, Shichen Li, and Zhongqing Wang. Exploring chain-of-thought for multi-modal metaphor detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 91-101, 2024. 3", + "[72] Dingyi Yang, Chunru Zhan, Ziheng Wang, Biao Wang, Tiezheng Ge, Bo Zheng, and Qin Jin. Synchronized video storytelling: Generating video narrations with structured storyline. arXiv preprint arXiv:2405.14040, 2024. 3", + "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 3", + "[74] Ruilin Yao, Bo Zhang, Jirui Huang, Xinwei Long, Yifang Zhang, Tianyu Zou, Yufei Wu, Shichao Su, Yifan Xu, Wenxi Zeng, et al. Lens: Multi-level evaluation of multimodal reasoning with large language models. arXiv preprint arXiv:2505.15616, 2025.3", + "[75] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 7, 15", + "[76] Keren Ye, Narges Honarvar Nazari, James Hahn, Zaeem Hussain, Mingda Zhang, and Adriana Kovashka. Interpreting the rhetoric of visual advertisements. IEEE transactions on pattern analysis and machine intelligence, 43(4):1308-1323, 2019. 3" + ], + "bbox": [ + 516, + 90, + 905, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23406", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134. AAAI Press, 2019. 3, 7, 15, 16", + "[78] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms. arXiv preprint arXiv:2406.14228, 2024. 4, 6, 7", + "[79] Majid Zahmati, Seyed Morteza Azimzadeh, Mohammad Saber Sotoodeh, and Amid Asgari. An eye-tracking study on how the popularity and gender of the endorsers affected the audience's attention on the advertisement. Electronic Commerce Research, 23(3):1665-1676, 2023. 3", + "[80] Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, et al. Internlm-xcomposer2. 5-reward: A simple yet effective multi-modal reward model. arXiv preprint arXiv:2501.12368, 2025. 4", + "[81] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresight-drive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025.3", + "[82] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 3", + "[83] Dongyu Zhang, Jingwei Yu, Senyuan Jin, Liang Yang, and Hongfei Lin. Multicmet: A novel chinese benchmark for understanding multimodal metaphor. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6141-6154, 2023. 3", + "[84] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b. arXiv preprint arXiv:2406.07394, 2024. 1, 6, 7", + "[85] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 3, 7, 15", + "[86] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3", + "[87] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7", + "[88] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 15", + "[89] Kang Zhao, Xinyu Zhao, Zhipeng Jin, Yi Yang, Wen Tao, Cong Han, Shuanglong Li, and Lin Liu. Enhancing baidu" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multimodal advertisement with chinese text-to-image generation via bilingual alignment and caption synthesis. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2855-2859, 2024. 3", + "[90] Zhou Zhao, Zhu Zhang, Shuwen Xiao, Zhou Yu, Jun Yu, Deng Cai, Fei Wu, and Yueting Zhuang. Open-ended long-form video question answering via adaptive hierarchical reinforced networks. In IJCAI, page 8, 2018. 3", + "[91] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 7", + "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3", + "[93] Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding. arXiv preprint arXiv:2501.18362, 2025. 1", + "[94] Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al. Trl: Test-time reinforcement learning. arXiv preprint arXiv:2504.16084, 2025.3" + ], + "bbox": [ + 516, + 92, + 903, + 484 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_model.json b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72b90752ca7a40d0ef8fff117ac441568607924d --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_model.json @@ -0,0 +1,2545 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.776, + 0.154 + ], + "angle": 0, + "content": "AdsQA: Towards Advertisement Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.179, + 0.836, + 0.322 + ], + "angle": 0, + "content": "Xinwei Long\\(^{1}\\) Kai Tian\\(^{1*}\\) Peng Xu\\(^{1\\boxtimes}\\) Guoli Jia\\(^{1}\\) Jingxuan Li\\(^{2}\\) Sa Yang\\(^{3}\\) Yihua Shao\\(^{4}\\) Kaiyan Zhang\\(^{1}\\) Che Jiang\\(^{1}\\) Hao Xu\\(^{5}\\) Yang Liu\\(^{2}\\) Jiaheng Ma\\(^{2}\\) Bowen Zhou\\(^{1,6\\boxtimes}\\) \n\\(^{1}\\) Tsinghua University, \\(^{2}\\) Independent Researcher, \\(^{3}\\) Peking University, \\(^{4}\\) CASIA \\(^{5}\\) Harvard University, \\(^{6}\\) Shanghai Artificial Intelligence Lab \n{longxw22, tk23}@ mails.tsinghua.edu.cn; {peng_xu, zhoubowen}@tsinghua.edu.cn \nProject page: https://github.com/TsinghuaC3I/AdsQA \nMARS2 workshop & challenge: https://lens4mllms.github.io/mars2-workshop-iccv2025/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.357, + 0.327, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.766 + ], + "angle": 0, + "content": "Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 toptier LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.794, + 0.222, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.484, + 0.88 + ], + "angle": 0, + "content": "A recent milestone flagged by OpenAI o1 [25] and DeepSeek-R1 [21] has been in the spotlight, and already opened up an era of large reasoning models. Meanwhile, going beyond general domains, the specialized do" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.908, + 0.57 + ], + "angle": 0, + "content": "mains [13, 93] are increasingly studied to boost these general-purpose LLMs to evolve continuously. The specialized domains contribute significantly in optimizing and evaluating generalist models towards specialized reasoning like domain experts [18, 52]. Recently, a clear phenomenon [20, 84] has been observed that LLMs are already good at reasoning a kind of step-wise problems represented by math and programming. As has been widely discussed [20, 84], the step-wise problems based on explicit theorems and programming syntax are compatible with the chain-style reasoning approach ``if A, then B'. To further extend the diversity of specialized reasoning for LLMs, it would be better to exploit other novel domains where the domain-specific reasoning is unmanageable for \"if A, then B\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.744 + ], + "angle": 0, + "content": "Therefore, we, for the first time, propose to use advertisement videos to probe the reasoning boundary of LLMs in perceiving implicit multimodal reasoning. The domain-unique features of clue-rich and information-dense ad videos can be summarized as key words implicit, nonphysical, mental, heuristic, etc. Different from user-uploaded videos on social media, ads are typically meticulously crafted by commercial or non-commercial organizations, making them entertaining, creative, aesthetically appealing, and capable of offering enjoyment and attracting viewer engagement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Based on these domain-specific advantages of ad videos, we introduce AdsQA, a comprehensive and carefully curated VideoQA benchmark derived from 1,544 ad videos containing 10,962 clips, spanning a total of 22.7 hours of video content. The AdsQA introduces five tasks, each requiring different types and levels of reasoning: (1) Visual Concept Understanding: Identify and analyze visual elements in ads. (2) Emotion Recognition: Detect emotions and infer their roles in ads. (3) Theme and Core Message Extraction: Summarize the central theme and key messages" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.386, + 0.901 + ], + "angle": 0, + "content": "*Kai Tian has equal contribution with Xinwei Long." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.094, + 0.361, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.098, + 0.631, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.099, + 0.898, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.454, + 0.778, + 0.469 + ], + "angle": 0, + "content": "Figure 1. Overview of AdsQA benchmark. Subfits A - B: statistics & diversity, and examples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.495, + 0.484, + 0.692 + ], + "angle": 0, + "content": "of the ad. (4) Persuasion Strategy Mining: Analyze the strategies used to persuade the audience. (5) Potential Audience Modeling: Identify and characterize the target audience. The five tasks are formalized as the open-ended QA format. During the construction process, we propose an innovative Role-Played Multi-Agent Annotation framework, which simulates the role of advertising experts to generate specialized data and significantly reduces the workload of human annotators. To ensure data quality, we conduct multiple rounds of automated data cleaning and manual data revision. Moreover, based on our AdsQA, we benchmark 14 well-known LLMs, including the GPT-4 level models, recently-released reasoning-based models, etc." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Humans process visual stimuli from ads and naturally form mental impressions, rather than relying on rigid logical reasoning templates, e.g., \"if A, then B\". Inspired by this heuristic process, we propose ReAd-R, a DeepseekR1 styled Reinforced Ad Reasoner, which learns to gather effective visual stimuli from ad videos, then reflects on the questions, and finally answers the given questions in a human-like manner. Specifically, ReAd-R utilizes a reward function to evaluate the model's responses and adjusts its parameters based on correctness, facilitating learning through trial and error. As a result, ReAd-R enhances reasoning abilities through outcome-reward-driven optimization, eliminating the need for costly step-wise su" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.587 + ], + "angle": 0, + "content": "pervision or chain-of-thought (COT) training data. Experiments demonstrate that our model brings reasoning ability an obvious gain in comprehending implicit logic of ad videos. In contrast, other reasoning-based methods obtain limited improvement due to fixed COT templates and suboptimal process reward models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.75 + ], + "angle": 0, + "content": "To our knowledge, our main contribution can be stated as: (1) AdsQA is the first video QA benchmark for advertisement domain, which is also the first ad benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current mainstream 'if A, then B' LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. Moreover, it advances ad video understanding beyond physical-content dominated shallow perception towards deeper cognitive reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.84 + ], + "angle": 0, + "content": "(2) We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, which enhances the specialized reasoning capability to understand implicit logic in ad videos like the human thinking process from perception to cognition. ReAd-R also can be regarded as one of the earliest attempts evolving R1-technique to vision research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "(3) We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning (e.g., variants of GPT4, LLaVA, and Qwen) by a clear margin." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23397" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.477 + ], + "angle": 0, + "content": "Video Question Answering Benchmarks. Video question answering (Video QA) [7, 48, 55, 58, 68] aims to answer user questions based on a given video, requiring a detailed understanding of both the spatio-temporal information and the relationships between objects and events [5, 12, 66, 74]. The Video QA community has developed a variety of benchmarks. From the perspective of video content, most benchmarks focus on human-centric videos, such as NextQA [65], ActivityQA [77], MVBench [36], FunQA [67], and VideoMME [16], which are derived from movies [19, 59], TV shows [32], social media [67], etc. Additionally, there are benchmarks centered on object-centric videos, such as EgoSchema [47], which are sourced from instructional and operational videos [38, 92]. Despite their advancements, these benchmarks still have several limitations. MovieQA [59] overly relies on dialogue understanding, which restricts the in-depth comprehension of visual content. Although TGIF-QA [26] requires some complex reasoning, the GIFs used are typically no longer than 3 seconds. ActivityQA [77] and Next-QA [65] have introduced open-ended QA tasks, but the annotated answers are often overly simplistic, usually consisting of just a few words. To sum up, current video QA benchmarks face limitations in video diversity and the setups of QA pairs." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.479, + 0.482, + 0.749 + ], + "angle": 0, + "content": "Video QA Methods and Video-LLMs. Early video QA methods employed graph structures [29, 56, 90] or transformers [17, 33, 37] to model the spatiotemporal relationships in videos and capture interactions between humans and objects. Recently, Video-LLMs [6, 45, 46] have demonstrated impressive capabilities in video understanding. Representative Video-LLMs include the VideoLLaMA [9, 47] series, LLaVA series [34, 39, 41, 70, 86], InternLM series [62, 85], and Qwen series [4, 60]. Video-LLMs take both the question and the video as input and directly generate the answer, implicitly performing the reasoning and thinking process in an efficient manner. Moreover, Some methods attempt to explicitly model the reasoning process through step-by-step thinking or multi-round debates. For example, some studies [15, 22, 57, 63, 81] leverages chain-of-thought to simulate the thinking process, while VideoAgent [14, 61, 73] improves reasoning capabilities through agent collaboration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Advertisement Understanding. Internet companies generate substantial profits by automatically distributing advertisements to target users, making the understanding of ad content highly important [28, 79]. The Pit dataset [24], one of the earliest efforts in automatic ad understanding, formalized this task as a visual question-answering (VQA) problem [43]. Subsequent research has built upon the Pit dataset or further developed it [53], or utilized their own private datasets [72]. These studies have mainly focused on a single aspect of advertisement understanding, such as persua" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "sive strategies in image ads [31, 76], image ad search [89], intent understanding [27], and visual metaphor comprehension [3, 53, 71, 82, 83]. Although the Pit dataset primarily focuses on image advertisements, it also provided a subset of video advertisements they collected. However, the Pit dataset suffers from issues such as data inaccessibility, lack of diversity, and limited Q&A formats. Therefore, there is currently no comprehensive Video QA benchmark available in the advertising domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.905, + 0.411 + ], + "angle": 0, + "content": "Reinforcement Learning for Reasoning. Recent research efforts [25, 30, 44, 94] have attempted to improve the reasoning capabilities of LLMs through reinforcement learning. Recently, DeepSeek-R1 [21] achieved significant improvements in reasoning based on reinforcement learning, eliminating the need for intermediate reasoning signals. S1 [49] demonstrated that even with only a few hundred data points, the model can effectively perform reasoning in specialized tasks. However, current research on RL-based reasoning primarily focuses on math and code problems, with few works [42] attempting to apply it to multimodal domains, particularly open-ended video QA tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.431, + 0.738, + 0.447 + ], + "angle": 0, + "content": "3. The AdsQA Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.667, + 0.471 + ], + "angle": 0, + "content": "3.1. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.905, + 0.527 + ], + "angle": 0, + "content": "To comprehensively evaluate the model's ability to understand ad videos, we divided the questions into five types, with each one targeting a specific angle of ad analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.905, + 0.589 + ], + "angle": 0, + "content": "Visual Concept Understanding (VU) task evaluates the model's ability to comprehend specific visual concepts, such as characters, objects, scenes, slogans, and other details, as well as their interrelationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.905, + 0.637 + ], + "angle": 0, + "content": "Emotion Recognition (ER) assesses if the models understand what emotion the ad evokes, and how ads establish connections with audiences through emotions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.905, + 0.699 + ], + "angle": 0, + "content": "Theme and Core Message Extraction (TE) drives the models to extract the underlying message or central idea that the ad explores. This task requires deep reasoning to gather visual information from the full video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.905, + 0.792 + ], + "angle": 0, + "content": "Persuasion Strategy Mining (PS) task evaluates the model's ability to uncover the strategies used to convey core messages and persuade audiences, such as humor, exaggeration, and visual rhetoric. The task includes questions, such as: how an ad conveys its central message, why the ad is appealing, and what strategies the ad employs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Potential Audience Modeling (AM) probes the model's performance in identifying potential audience groups and their profiles. It sets key questions, such as who the ad targets and what the characteristics of the audience are. This task directly reflects the ad's influence and value." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "These five tasks are defined as the open-ended video QA and each question is annotated with a ground-truth answer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23398" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.358, + 0.108 + ], + "angle": 0, + "content": "3.2. Dataset Construction Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.482, + 0.415 + ], + "angle": 0, + "content": "AdsQA benchmark construction pipeline is in three stages: Pre-Processing. To serve the ad video understanding, we consider the creativity, aesthetic quality, and availability of videos during our collection process. We do not use any private data; instead, we collect creative ad videos that adhere to the Creative Commons License from a creative community platform [1]. This platform offers publicly accessible, high-quality ad videos uploaded by creators. Specifically, we first crawl videos along with their metadata. The metadata is typically written by the uploader (i.e., usually the creator of the ad) and includes information such as the theme, content, and key creative elements of the ad. It can be considered the \"ground-truth\" information of the ad video and serves as an important reference for our benchmark construction. Then, human experts are asked to carefully review the completeness and accuracy of the metadata and manually remove incomplete samples. Additionally, we conduct automated filtering based on video duration, aesthetic score, and content to ensure video quality and exclude non-ethical and negative content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Considering efficiency and accuracy, we employ PySceneDetect to segment ad videos into fine-grained video clips to avoid missing detailed information in subsequent steps. For each clip \\( C_i \\), we use Video-LLMs to generate its description \\( Desc_i \\). To preserve key visual elements, we sample \\( n \\) keyframes \\( \\{F_i^0, \\dots, F_i^{N_{max}}\\} \\) from each clip based on frame similarity calculated using SSIM [64], where \\( n \\) is determined by the duration of each clip. To retain speech information, we extract speech information \\( Asr_i \\) for each clip using the Whisper model, and translate them into English using GPT-4 [23]. Therefore, an ad video with \\( N \\) clips can be represented as a modality-interleaved sequence, as Eq. 1," + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.619, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\nV = \\left\\{M, \\left\\{F _ {i} ^ {0}, \\dots , F _ {i} ^ {n _ {i}}, D e s c _ {i}, A s r _ {i} \\right\\} _ {i = 1} ^ {N} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\( M \\) denotes the meta-information of the video. The modality-interleaved sequence will be used for automated annotation generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Role-Played Multi-Agent Annotation. To balance data quality and cost, ensure data diversity, and avoid template-driven outputs, we propose a Role-Played Multi-Agent Annotation framework to automatically generate video QA pairs. Inspired by previous research [78], we found that agents can develop specialized capabilities in the ad domain by configuring specific skill sets, such as marketing, visual design, and consumer psychology. Therefore, we can leverage AI agents to act as ad experts, analyzing diverse facets of ad videos and designing more specialized and challenging questions. We ask human experts to create profiles for advertising expert skills, enabling AI agents to adaptively select profiles (or autonomously generate new profiles) to accomplish role-playing tasks. Additionally, human experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "provide examples of QA pairs as in-context demonstration to guide the agents in generating appropriate QA pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.905, + 0.363 + ], + "angle": 0, + "content": "The Role-Played Multi-Agent Annotation is formulated into a three-stage pipeline. (1) In the initial stage, a master agent is recruited to generate a preliminary QA annotation for the given modality-interleaved sequence \\( V \\). (2) In the iterative stage, the master agent checks the quality of the current QA annotations and decides whether to recruit a specialized expert agent. If necessary, the master agent selects a profile for the expert agent and instantiates it accordingly. The expert agent then generates new QA annotations leveraging its specialized expertise. Finally, the master agent revises the original annotations based on the expert agent's provided annotations and assesses whether to terminate the iteration or continue to recruit a new expert agent. (3) After terminating the iteration, the master agent synthesizes outputs from previous iterations to produce the final QA annotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.471 + ], + "angle": 0, + "content": "Automated Data Cleaning. After automated annotation through multi-agent collaboration, we employ the IXC-2.5-Reward model [80] to automatically verify the correctness of the answers. Given the video, question, and meta information, the reward model scores the answer. QA pairs with scores below a certain threshold are flagged and carefully reviewed by annotators." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.824, + 0.499 + ], + "angle": 0, + "content": "3.3. Manual Check and Quality Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Following the automatic annotation generation, we conduct a careful manual check and revision on all Q&A pairs, focusing on the question suitability, annotation correctness, video content, and task difficulty. To ensure the quality of Q&A pairs, we first remove questions that are homogeneous, unrelated to the ads' theme, or cannot be answered based on the ad video (e.g., those requiring external knowledge). Then, we incorporate the meta information as ground truth to verify the accuracy of the reference answers. We eliminated pairs containing inaccurate, ambiguous, or biased answers. After two rounds of rigorous selection, only \\(37\\%\\) of the QA pairs remain. Additionally, we manually revise the wording of some QA pairs to improve their clarity and accuracy. Any modifications to a sample are reviewed by other annotators to ensure consistency and reliability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.682, + 0.758 + ], + "angle": 0, + "content": "3.4. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Videos. Our AdsQA Benchmark includes a total of 1,544 unique advertisements, comprising 10,962 video clips for evaluation. As shown in Fig. 2a, these ad videos span 9 primary domains. The video durations range from 15 to 120 seconds, with an average length of 52.9 seconds, amounting to a total of 22.7 hours of content. Each ad video is accompanied by metadata, including a title, tags, automatic speech recognition (ASR) results, and descriptions. This metadata is utilized for Q&A generation, as described in Sec. 3.1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "23399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.365, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.394, + 0.356, + 0.404 + ], + "angle": 0, + "content": "(a) Alluvial diagram of task and domain distribution." + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.104, + 0.634, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.25, + 0.587, + 0.261 + ], + "angle": 0, + "content": "(b) Distribution of video release years." + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.103, + 0.837, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.25, + 0.822, + 0.261 + ], + "angle": 0, + "content": "(c) Distribution of video release region." + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.265, + 0.611, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.393, + 0.608, + 0.404 + ], + "angle": 0, + "content": "(d) The distribution of the video duration in the dataset." + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.263, + 0.86, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.393, + 0.839, + 0.404 + ], + "angle": 0, + "content": "(e) Number of Q&A entries for different tasks." + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.42, + 0.622, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Statistics of Our AdsQA Bench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Most of the ad videos are in English and primarily originate from North America and Europe." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.706 + ], + "angle": 0, + "content": "Q&A Pairs. Our AdsQA Benchmark includes 7,859 QA pairs (among which \\(29.2\\%\\) of the questions can be categorized into two types, resulting in a total of 10,154 QA pairs for evaluation). Among these: Visual Concept Understanding accounts for \\(21.4\\%\\) of the QA pairs, Theme and Message Extraction accounts for \\(28.5\\%\\), Persuasion Strategy Mining accounts for \\(25.3\\%\\), Potential Audience Modeling accounts for \\(15.1\\%\\), and Emotion Recognition accounts for \\(9.8\\%\\). Typical persuasion strategies include metaphor, symbolism, humor, expert opinion, and others. Some questions may fit into two categories, such as \"What is the role of a particular visual concept in relation to the theme?\" On average, each question consists of 17 words and each answer contains approximately 12 words." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.226, + 0.74 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, we propose ReAd-R, a DeepSeekR1 styled Reinforced Ad Reasoner, to simulate human heuristic thinking and learn from trial and error. The model takes an ad video and a question as input. The policy model generates a set of reasoning processes, including thoughts and answers, based on the given input. Each reasoning process is evaluated using a reward function to compute its reward value. After calculating the reward values for all outputs, each reasoning process is assessed and used to update the policy model. Additionally, ReAd-R employs KL di" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.459, + 0.891, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.622, + 0.89, + 0.675 + ], + "angle": 0, + "content": "Reasoning: \"think>Loop's noise-reduction products, as inferred from the context, are intended to help individuals in noisy environments, such as subway rides, music parties, and bedrooms where noise can disturb others...nPeople who are concerned with noise reduction in their daily lives—such as commuters, festival-goers, night owls, or anyone sensitive to loud environments—would benefit the most from using Loop's noise-reducing earwear." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Figure 3. Framework of ReAd-R. Given a question and video, the policy model generates multiple responses. The reward model evaluates and scores them, and the rewards are used to update the policy model via policy gradient optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.907, + 0.809 + ], + "angle": 0, + "content": "vergence to control the difference between the policy model and the reference model, ensuring stable training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Data Preparation. Inspired by previous work [21, 49], RL-based reasoners can improve reasoning capabilities with limited high-quality data. We additionally crawled a total of 1,053 ad videos along with their metadata and automatically generated annotations using the same method. We utilized metadata tags collected during data crawling (e.g., topic," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "domain) to ensure diversity, adopted Video-LLM evaluation results to assess difficulty, and performed manual quality assessment based on the same criteria as in Sec. 3.3. Finally, we selected 100 videos and 500 question-answer pairs. We designed a prompt format to guide the model to output its reasoning process before providing the final answer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Reward Modeling. The reward model guides the model to find the optimal direction through trial and error. Ideally, ReAd-R's responses should (1) include as many elements of the standard answer as possible and (2) avoid containing content not mentioned in the ad video (e.g., hallucinations). To achieve this, we propose a rule-guided LLM evaluator as the reward model to assess the quality of each response. We introduce two sets of rules: inclusion rules and exclusion rules. The inclusion rule requires that the generated answer incorporate as many elements of the standard answer as possible. If satisfied, the reward value is 1.0; otherwise, it is 0. The exclusion rule states that if the generated content includes elements not mentioned in the standard answer and these elements cannot be inferred from the meta-information, they should be judged as incorrect. To avoid sparse reward values in the early stages, we relaxed the inclusion rule. If the generated response partially includes factual elements, it is assigned a reward value of 0.5. Additionally, the format reward is used to enforce the model's predictions to adhere to the required format of and , as \\( R(\\cdot) = R_{\\text{ans}}(\\cdot) + R_{\\text{format}}(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.605 + ], + "angle": 0, + "content": "Reinforced Fine-tuning. Following DeepSeek-R1, we employ the GRPO algorithm instead of PPO to optimize our model. ReAd-R first generates \\( n \\) distinct responses \\( O = \\{o_1, o_2, \\dots, o_n\\} \\) from the current policy model \\( \\pi_{\\theta_{old}} \\). Then, the reward model \\( R(\\cdot) \\) evaluates these responses \\( O \\) to obtain their reward value as \\( \\{r_1, r_2, \\dots, r_n\\} \\), and the advantage \\( A_i \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.615, + 0.482, + 0.648 + ], + "angle": 0, + "content": "\\[\nA _ {i} = \\frac {r _ {i} - \\operatorname {m e a n} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\(A_{i}\\) denotes relative quality of the \\(i^{th}\\) answer when compared to the group of rewards \\(\\{r_1, r_2, \\ldots, r_n\\}\\). This strategy encourages the model to generate better answers with higher reward values within the group. We use GRPO algorithm [20] to optimize our policy model based on \\(A_{i}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.224, + 0.765 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.298, + 0.79 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluated Models. We select 14 top-tier LLMs as strong competitors covering diverse categories: (1) GPT-4 Level Baselines: including GPT-4o [23], GPT-4V [2] and Qwen2.5-VL-72B [4]. (2) Open-sourced Video LLMs: We present at least one representative model released in the past six months from each series of Video-LLMs, such as Qwen2.5-VL-7B [4], LLaVA-Onevision-7B [34], and oth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ers. (3) Reasoning Based Methods: We re-implemented several reasoning-based methods and applied them to the AdsQA task. These methods include the Video Chain-of-Thought (VOT) [15], the Role-Played Agent (EvoAgent) [78], and the Monte Carlo Tree Search (MCTS) algorithms [84]. Detailed descriptions and implementation methods are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Evaluation Metric. All five tasks of AdsQA are evaluated as open-ended QA, allowing for the free-text evaluation methods. Traditional free-text metrics [40, 51] are sensitive to lexical variations, which may introduce bias when evaluating different LLMs. Recent studies [8, 10, 19, 67] have shown promising results in using large generalist models to evaluate generated text. Therefore, we follow their framework and employ GPT-4o to assist in evaluating fretext similarity. Specifically, we use the same inclusion and exclusion rules mentioned in Sec. 4 to guide the scoring of text generated by the large model. We provided each sample with a ground-truth answer and meta-information; each meta-information includes relevant information such as the creative elements, content, storyline, and themes of the advertisement video. We propose both relaxed and strict scoring modes. Our strict accuracy Acc_strict is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.449, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A c c} _ {\\text {s t r i c t}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.497, + 0.907, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\{\\cdot\\}\\) is the indicator function that returns 1 if the model's response \\(x_{i}\\) semantically matches all elements of the standard answer \\(y_{i}\\), and 0 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.543, + 0.906, + 0.571 + ], + "angle": 0, + "content": "In the relaxed mode, the relaxed accuracy \\( Acc_{\\mathrm{relaxed}} \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.581, + 0.905, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A c c} _ {\\text {r e l a x e d}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\} + \\lambda \\cdot \\mathbb {1} \\left\\{x _ {i} \\approx y _ {i} \\right\\}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.628, + 0.906, + 0.689 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\{x_i\\approx y_i\\}\\) returns 1 if the response \\(x_{i}\\) partially addresses the elements of \\(y_{i}\\), and 0 otherwise. It is observed \\(\\lambda = 0.5\\) works well in experiments. Our prompt template for model-based evaluation are in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Implementation Details. Our ReAd-R is model-agnostic. In our main experiments, we utilize Qwen2-VL-7B as the base model. We froze the parameters of the visual backbone and only fine-tuned the parameters of the language model. We conducted the experiments with a batch size of 4 and a learning rate of 1e-6 on a server with eight A100 GPUs, with a training duration of 12 hours. Additional details and hyperparameters are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.746, + 0.833 + ], + "angle": 0, + "content": "5.2. Results and Observations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Although GPT-4 achieves over \\(85\\%\\) accuracy on other datasets such as Next-QA, its strict accuracy on our dataset is only \\(29.4\\%\\), and its relaxed accuracy is just \\(56.6\\%\\). This shows that our benchmark exceeds the capabilities of some" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23401" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.468 + ], + "angle": 0, + "content": "
ModelStrict AccuracyRelaxed Accuracy
VUERTEPSAMOverallVUERTEPSAMOverall
Human Results*55.139.457.157.545.851.374.565.975.776.264.671.4
Commercial Large Multimodal Model
GPT-4o [23]24.926.532.632.331.029.450.157.862.455.254.856.6
Gemini-2.5-Pro* [11]36.044.040.429.333.335.959.269.064.054.462.460.7
Open-sourced Video-LLMs
VideoLLaMA2-7B [9]4.567.757.286.067.386.4821.629.028.422.627.225.2
InternLM-XComp.2.5-7B [85]11.210.716.311.912.112.634.735.941.534.137.336.5
LLaVA-OneVision-7B [34]11.811.616.213.715.114.035.839.543.236.841.739.1
LLaVA-Video-7B [87]14.014.418.715.217.316.137.841.645.138.143.641.0
MiniCPM-o_2.6-7B [75]13.015.317.314.918.916.338.143.445.739.145.042.5
Qwen2-VL-7B [60]13.715.420.216.020.017.236.841.346.037.744.641.0
Qwen2.5-VL-7B [4]20.223.224.620.524.323.045.850.451.845.350.948.9
Qwen2.5-VL-72B [4]26.730.234.829.834.731.051.857.459.553.859.255.8
Reasoning-based Models
VOT (Qwen2-VL-7B) [15]14.316.419.512.622.017.036.543.342.535.245.140.2
EvolAgent (Qwen2-VL-7B) [78]15.016.121.616.520.418.338.643.446.440.445.342.6
MCTSr (Qwen2-VL-7B) [84]14.810.219.316.116.417.140.641.346.541.142.643.4
Qwen2-VL-7B (SFT) [60]10.512.517.913.815.914.134.842.142.734.941.438.8
ReAd-R (Qwen2-VL-7B) (Ours)15.819.321.718.518.818.642.346.450.043.447.844.9
ReAd-R (Qwen2.5-VL-7B) (Ours)20.427.927.222.125.525.046.256.254.648.252.651.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.405, + 0.477, + 0.593, + 0.492 + ], + "angle": 0, + "content": "Table 1. Experimental Results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.482, + 0.625 + ], + "angle": 0, + "content": "multi-modal large models. We observe that the state-of-the-art multi-modal models show strong visual perception, often generating partial answers by describing video segments. However, fully understanding the underlying meaning of ads is difficult for all baseline models. This requires not only identifying specific visual elements but also interpreting their implicit logic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Different tasks vary in difficulty, but their performance is comparable. Though tasks focus on different aspects, these aspects are closely linked. For example, ads may use emotions or specific visuals to convey themes. Most models perform best on Theme and Message Extraction and Audience Modeling tasks. This is because ad videos aim to deliver clear messages to specific audiences, making it relatively easier to identify themes and target users. Even without fully understanding creative elements or marketing strategies, models can infer these from characters, scenes, or slogans in the ad. The Persuasion Strategy Mining task is more challenging. It requires models to explain how and why ads use certain designs or elements. These questions are often not directly expressed in the video and may even require interpreting counterfactual or unexpected visual information, which increases the difficulty of this task. Models also perform worse on the Visual Concept Understanding (VU) task. While other video QA benchmarks [77] focus on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.905, + 0.594 + ], + "angle": 0, + "content": "surface-level questions like \"What color are the gloves?\", our VU task requires understanding global ad information, such as \"Which scene represents the ad's theme?\" Answering this question requires understanding the theme, identifying related scenes, and describing them briefly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.822 + ], + "angle": 0, + "content": "On the AdsQA task, reasoning models that excel in code and math tasks achieved only marginal results, even with ten times the computational cost. This shows that AdsQA reasoning differs significantly from tasks like math and code. Math and code rely on structured \"if A, then B\" reasoning, while ads require associative reasoning, connecting concrete visuals to abstract concepts. Prior work [91] also suggests chain-of-thought reasoning can hinder abstract reasoning performance. Though MCSTr improved results by \\(2.4\\%\\) over Qwen2-VL, its search process is inefficient and highly dependent on the reward model. In some cases, Qwen2-VL answers correctly, but MCSTr fails due to incorrect search directions from the reward model. Moreover, training an effective reward model is difficult, especially in the data-scarce ad domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our method, ReAd-R, achieves a \\(3.9\\%\\) improvement using only 500 video-question-answer triplets, outperforming Qwen2-VL and other reasoning models. During inference, our model reasons about video content and questions by generating free-text explanations, eliminating the need for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23402" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.246 + ], + "angle": 0, + "content": "
Strict Acc.Relaxed Acc.
ReAd-R (Our Model)18.644.9
w. Uncurated Data13.236.2
w. Strict Reward15.141.7
w.o. Prompt Constraints16.842.9
Qwen2-VL (SFT)14.138.8
w. Uncurated Data16.040.9
VOT (Chain length: 3)17.040.2
w. (Chain length: 5)13.636.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.256, + 0.48, + 0.27 + ], + "angle": 0, + "content": "Table 2. Ablation study on ReAd-R, Qwen2-VL SFT, and VOT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.465 + ], + "angle": 0, + "content": "fixed chain-of-thought (COT) templates (e.g., VOT), additional reward functions (e.g., MCTSr), or iterative search processes (e.g., MCTSr and EvolAgent). Interestingly, we observe that supervised fine-tuning (SFT) on the same data caused a sharp performance drop of \\(2.2\\%\\) for Qwen2-VL. Using limited but high-quality data does not improve the performance of SFT models. SFT may overfit to the features of the training data, leading to poor generalization on diverse test sets. In contrast, our ReAd-R model improves reasoning capabilities through RL on limited data, achieving generalization across diverse advertisement videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.471, + 0.245, + 0.488 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.734 + ], + "angle": 0, + "content": "We conduct ablation studies for our ReAd-R model in Tab 2. We first show results of RL fine-tuning on 1,053 uncurated videos and 5,159 QA pairs. Our model performs worse than Qwen2-VL. This is due to the GRPO algorithm's sensitivity to data quality. Early in fine-tuning, low-quality data can mislead optimization, causing suboptimal results. Next, we test a strict reward rule (only fully correct answers receive a \\( R_{ans} \\) reward of 1; otherwise, the reward is 0). Our method does not achieve improvement through RL fine-tuning. During training, we observe that the strict reward rule makes it hard for the model to earn rewards, limiting gains. We also find ReAd-R is sensitive to prompt templates. Properly constraining the model's thinking and answers (e.g., avoiding repetitive content in and tags) stabilizes training and improves performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We also conduct ablation studies on our implemented baselines (e.g., VOT, Qwen2-VL SFT). We present the results of Qwen2-VL fine-tuned using uncurated data. Interestingly, the model fine-tuned on uncurated data improves by \\(2.1\\%\\) over limited-data fine-tuning but still lags behind the non-SFT model. This shows SFT does not enhance reasoning on ad data. Additionally, we experiment with VOT using a COT length of 5 (as set in its original paper) and observe a significant performance drop of \\(4.0\\%\\). The model does not benefit from step-by-step reasoning; instead, accumulated errors degrade its performance. This highlights the" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.089, + 0.895, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.325, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 4. Case Study. Due to the limited page space, we omit the model's thinking process and show only final answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.416 + ], + "angle": 0, + "content": "difference between abstract reasoning in the ad domain and chain-style logical reasoning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.432, + 0.637, + 0.449 + ], + "angle": 0, + "content": "5.4. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.609 + ], + "angle": 0, + "content": "As shown in Fig. 4, Qwen2.5-VL generates vague and overly general answers (scored 0 by GPT-4). The COT baseline produces better answers (scored 0.5 by GPT-4) by gathering information step-by-step through a reasoning process, but it still misses some important clues in the video. Our ReAd-R learns to perform video-grounded reasoning through reinforcement fine-tuning, deriving the final answer by summarizing its reasoning process. The reinforced thinking-answer mechanism helps ReAd-R generate more complete and precise answers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.629, + 0.634, + 0.644 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.853 + ], + "angle": 0, + "content": "We contibute AdsQA, the first advertisement benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current \"if A, then B\" LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, no need of costly supervision, which enhances the specialized reasoning capability to understand implicit logic in ad videos. We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities (e.g., variants of GPT4, LLaVA, and Qwen)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported by the National Key Research and Development Program of China (2022ZD0160603), and NSFC No.62306162." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23403" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.41, + 0.128 + ], + "angle": 0, + "content": "[1] https://www.adsofttheworld.com/.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.131, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.203, + 0.483, + 0.298 + ], + "angle": 0, + "content": "[3] Arjun R Akula, Brendan Driscoll, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T Freeman, et al. Metaclue: Towards comprehensive visual metaphors research. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23201-23211, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-v1 technical report. arXiv preprint arXiv:2502.13923, 2025. 3, 6, 7, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, Yao Dou, Jaden Park, Jianfeng Gao, Yong Jae Lee, and Jianwei Yang. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. CoRR, abs/2410.10818, 2024. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.542 + ], + "angle": 0, + "content": "[7] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.6 + ], + "angle": 0, + "content": "[8] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. CoRR, abs/2406.07476, 2024. 3, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[10] Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Ning Ding, Shang Qu, Linhai Xie, Yifei Li, Zaoqu Liu, Kaiyan Zhang, Yibai Xiong, Yuxin Zuo, Zhangren Chen, Ermo Hua, et al. Automating exploratory pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "teomics research via language models. arXiv preprint arXiv:2411.03743, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[17] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14773-14783, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[18] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 23923-23932, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[19] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (SFD): A benchmark for story-level video understanding. CoRR, abs/2406.10221, 2024. 3, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[20] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[22] Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding. arXiv preprint arXiv:2502.06428, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[24] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In CVPR, pages 1100-1110. IEEE Computer Society, 2017. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12986-12996, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[28] Zhiwei Jia, Pradyumna Narayana, Arjun R Akula, Garima Pruthi, Hao Su, Sugato Basu, and Varun Jampani. Kafa: Rethinking image ad understanding with knowledge-augmented feature adaptation of vision-language models. arXiv preprint arXiv:2305.18373, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[29] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Fangkai Jiao, Geyang Guo, Xingxing Zhang, Nancy F Chen, Shafiq Joty, and Furu Wei. Preference optimization for reasoning with pseudo feedback. arXiv preprint arXiv:2411.16345, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[31] Yaman Kumar, Rajat Jha, Arunim Gupta, Milan Aggarwal, Aditya Garg, Tushar Malyan, Ayush Bhardwaj, Rajiv Ratn Shah, Balaji Krishnamurthy, and Changyou Chen. Persuasion strategies in advertisements. In Proceedings of the AAAI conference on artificial intelligence, pages 57-66, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[32] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379. Association for Computational Linguistics, 2018. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 6, 7, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[35] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Lou, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark. In CVPR, pages 22195-22206. IEEE, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[37] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8658-8665, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[38] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In European Conference on Computer Vision, pages 200-216. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[40] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[42] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[43] Xinwei Long, Zhiyuan Ma, Ermo Hua, Kaiyan Zhang, Biqing Qi, and Bowen Zhou. Retrieval-augmented visual question answering via built-in autoregressive search engines. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 24723-24731, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[44] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[45] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[46] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding. arxiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[47] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS, 2023. 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[48] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[49] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[50] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. CoRR, abs/2311.16103, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[51] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23405" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[52] Biqing Qi, Kaiyan Zhang, Kai Tian, Haoxiang Li, ZhangRen Chen, Sihang Zeng, Ermo Hua, Hu Jinfang, and Bowen Zhou. Large language models as biomedical hypothesis generators: a comprehensive evaluation. arXiv preprint arXiv:2407.08940, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[53] Abisek Rajakumar Kalarani, Pushpak Bhattacharyya, and Sumit Shekhar. Seeing the unseen: Visual metaphor captioning for videos. arXiv e-prints, pages arXiv-2406, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[54] Ramon Sanabria, Ozan Caglayan, Shruti Palaskar, Desmond Elliott, Loic Barrault, Lucia Specia, and Florian Metze. How2: a large-scale dataset for multimodal language understanding. arXiv preprint arXiv:1811.00347, 2018. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[55] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[56] Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection. arXiv preprint arXiv:2504.13092, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[57] Yudi Shi, Shangzhe Di, Qirui Chen, and Weidi Xie. Unlocking video-llm via agent-of-thoughts distillation. arXiv preprint arXiv:2412.01694, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[58] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[59] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640. IEEE Computer Society, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[62] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[63] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool. arXiv preprint arXiv:2407.05355, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[64] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[65] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 9777-9786. Computer Vision Foundation / IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[66] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[67] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.387 + ], + "angle": 0, + "content": "[68] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[69] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645–1653, 2017. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[70] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.588 + ], + "angle": 0, + "content": "[71] Yanzhi Xu, Yueying Hua, Shichen Li, and Zhongqing Wang. Exploring chain-of-thought for multi-modal metaphor detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 91-101, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[72] Dingyi Yang, Chunru Zhan, Ziheng Wang, Biao Wang, Tiezheng Ge, Bo Zheng, and Qin Jin. Synchronized video storytelling: Generating video narrations with structured storyline. arXiv preprint arXiv:2405.14040, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.771 + ], + "angle": 0, + "content": "[74] Ruilin Yao, Bo Zhang, Jirui Huang, Xinwei Long, Yifang Zhang, Tianyu Zou, Yufei Wu, Shichao Su, Yifan Xu, Wenxi Zeng, et al. Lens: Multi-level evaluation of multimodal reasoning with large language models. arXiv preprint arXiv:2505.15616, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[75] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[76] Keren Ye, Narges Honarvar Nazari, James Hahn, Zaeem Hussain, Mingda Zhang, and Adriana Kovashka. Interpreting the rhetoric of visual advertisements. IEEE transactions on pattern analysis and machine intelligence, 43(4):1308-1323, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[77] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134. AAAI Press, 2019. 3, 7, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[78] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms. arXiv preprint arXiv:2406.14228, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[79] Majid Zahmati, Seyed Morteza Azimzadeh, Mohammad Saber Sotoodeh, and Amid Asgari. An eye-tracking study on how the popularity and gender of the endorsers affected the audience's attention on the advertisement. Electronic Commerce Research, 23(3):1665-1676, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[80] Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, et al. Internlm-xcomposer2. 5-reward: A simple yet effective multi-modal reward model. arXiv preprint arXiv:2501.12368, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[81] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresight-drive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[82] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[83] Dongyu Zhang, Jingwei Yu, Senyuan Jin, Liang Yang, and Hongfei Lin. Multicmet: A novel chinese benchmark for understanding multimodal metaphor. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6141-6154, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[84] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b. arXiv preprint arXiv:2406.07394, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[85] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 3, 7, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[86] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[87] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[88] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[89] Kang Zhao, Xinyu Zhao, Zhipeng Jin, Yi Yang, Wen Tao, Cong Han, Shuanglong Li, and Lin Liu. Enhancing baidu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "multimodal advertisement with chinese text-to-image generation via bilingual alignment and caption synthesis. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2855-2859, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[90] Zhou Zhao, Zhu Zhang, Shuwen Xiao, Zhou Yu, Jun Yu, Deng Cai, Fei Wu, and Yueting Zhuang. Open-ended long-form video question answering via adaptive hierarchical reinforced networks. In IJCAI, page 8, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[91] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[93] Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding. arXiv preprint arXiv:2501.18362, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[94] Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al. Trl: Test-time reinforcement learning. arXiv preprint arXiv:2504.16084, 2025.3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23407" + } + ] +] \ No newline at end of file diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_origin.pdf b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17c75da3b8871ce15c5c6195fbdbeecd4d63c8a1 --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/6375f078-3f0b-4521-b432-64d23e64bff4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58c9f6a72139246168e83f4e9c0e334939e96c257c320f5aea3d73fb7c6f2eef +size 5207974 diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/full.md b/2025/AdsQA_ Towards Advertisement Video Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eed1cab9edabc5b2a6a48578ec4379d1dc4526b8 --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/full.md @@ -0,0 +1,319 @@ +# AdsQA: Towards Advertisement Video Understanding + +Xinwei Long $^{1}$ Kai Tian $^{1*}$ Peng Xu $^{1\boxtimes}$ Guoli Jia $^{1}$ Jingxuan Li $^{2}$ Sa Yang $^{3}$ Yihua Shao $^{4}$ Kaiyan Zhang $^{1}$ Che Jiang $^{1}$ Hao Xu $^{5}$ Yang Liu $^{2}$ Jiaheng Ma $^{2}$ Bowen Zhou $^{1,6\boxtimes}$ $^{1}$ Tsinghua University, $^{2}$ Independent Researcher, $^{3}$ Peking University, $^{4}$ CASIA $^{5}$ Harvard University, $^{6}$ Shanghai Artificial Intelligence Lab +{longxw22, tk23}@ mails.tsinghua.edu.cn; {peng_xu, zhoubowen}@tsinghua.edu.cn +Project page: https://github.com/TsinghuaC3I/AdsQA +MARS2 workshop & challenge: https://lens4mllms.github.io/mars2-workshop-iccv2025/ + +# Abstract + +Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 toptier LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin. + +# 1. Introduction + +A recent milestone flagged by OpenAI o1 [25] and DeepSeek-R1 [21] has been in the spotlight, and already opened up an era of large reasoning models. Meanwhile, going beyond general domains, the specialized do + +mains [13, 93] are increasingly studied to boost these general-purpose LLMs to evolve continuously. The specialized domains contribute significantly in optimizing and evaluating generalist models towards specialized reasoning like domain experts [18, 52]. Recently, a clear phenomenon [20, 84] has been observed that LLMs are already good at reasoning a kind of step-wise problems represented by math and programming. As has been widely discussed [20, 84], the step-wise problems based on explicit theorems and programming syntax are compatible with the chain-style reasoning approach ``if A, then B'. To further extend the diversity of specialized reasoning for LLMs, it would be better to exploit other novel domains where the domain-specific reasoning is unmanageable for "if A, then B". + +Therefore, we, for the first time, propose to use advertisement videos to probe the reasoning boundary of LLMs in perceiving implicit multimodal reasoning. The domain-unique features of clue-rich and information-dense ad videos can be summarized as key words implicit, nonphysical, mental, heuristic, etc. Different from user-uploaded videos on social media, ads are typically meticulously crafted by commercial or non-commercial organizations, making them entertaining, creative, aesthetically appealing, and capable of offering enjoyment and attracting viewer engagement. + +Based on these domain-specific advantages of ad videos, we introduce AdsQA, a comprehensive and carefully curated VideoQA benchmark derived from 1,544 ad videos containing 10,962 clips, spanning a total of 22.7 hours of video content. The AdsQA introduces five tasks, each requiring different types and levels of reasoning: (1) Visual Concept Understanding: Identify and analyze visual elements in ads. (2) Emotion Recognition: Detect emotions and infer their roles in ads. (3) Theme and Core Message Extraction: Summarize the central theme and key messages + +![](images/2a2d7424407d699f5bae0578200c1caa3d9b56e9f5e6530d8f365cfe61d06597.jpg) +Figure 1. Overview of AdsQA benchmark. Subfits A - B: statistics & diversity, and examples. + +![](images/0f358a0cf58c7e5d6515b1f8e8307bfe04f1fbbdf5ac2a886a1e1846d70e9d69.jpg) + +![](images/6c02812ced8ce6bd3c1969ab7c8519b559f914be6dcd90aebe080c99e0eea425.jpg) + +of the ad. (4) Persuasion Strategy Mining: Analyze the strategies used to persuade the audience. (5) Potential Audience Modeling: Identify and characterize the target audience. The five tasks are formalized as the open-ended QA format. During the construction process, we propose an innovative Role-Played Multi-Agent Annotation framework, which simulates the role of advertising experts to generate specialized data and significantly reduces the workload of human annotators. To ensure data quality, we conduct multiple rounds of automated data cleaning and manual data revision. Moreover, based on our AdsQA, we benchmark 14 well-known LLMs, including the GPT-4 level models, recently-released reasoning-based models, etc. + +Humans process visual stimuli from ads and naturally form mental impressions, rather than relying on rigid logical reasoning templates, e.g., "if A, then B". Inspired by this heuristic process, we propose ReAd-R, a DeepseekR1 styled Reinforced Ad Reasoner, which learns to gather effective visual stimuli from ad videos, then reflects on the questions, and finally answers the given questions in a human-like manner. Specifically, ReAd-R utilizes a reward function to evaluate the model's responses and adjusts its parameters based on correctness, facilitating learning through trial and error. As a result, ReAd-R enhances reasoning abilities through outcome-reward-driven optimization, eliminating the need for costly step-wise su + +pervision or chain-of-thought (COT) training data. Experiments demonstrate that our model brings reasoning ability an obvious gain in comprehending implicit logic of ad videos. In contrast, other reasoning-based methods obtain limited improvement due to fixed COT templates and suboptimal process reward models. + +To our knowledge, our main contribution can be stated as: (1) AdsQA is the first video QA benchmark for advertisement domain, which is also the first ad benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current mainstream 'if A, then B' LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. Moreover, it advances ad video understanding beyond physical-content dominated shallow perception towards deeper cognitive reasoning. + +(2) We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, which enhances the specialized reasoning capability to understand implicit logic in ad videos like the human thinking process from perception to cognition. ReAd-R also can be regarded as one of the earliest attempts evolving R1-technique to vision research. +(3) We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning (e.g., variants of GPT4, LLaVA, and Qwen) by a clear margin. + +# 2. Related Work + +Video Question Answering Benchmarks. Video question answering (Video QA) [7, 48, 55, 58, 68] aims to answer user questions based on a given video, requiring a detailed understanding of both the spatio-temporal information and the relationships between objects and events [5, 12, 66, 74]. The Video QA community has developed a variety of benchmarks. From the perspective of video content, most benchmarks focus on human-centric videos, such as NextQA [65], ActivityQA [77], MVBench [36], FunQA [67], and VideoMME [16], which are derived from movies [19, 59], TV shows [32], social media [67], etc. Additionally, there are benchmarks centered on object-centric videos, such as EgoSchema [47], which are sourced from instructional and operational videos [38, 92]. Despite their advancements, these benchmarks still have several limitations. MovieQA [59] overly relies on dialogue understanding, which restricts the in-depth comprehension of visual content. Although TGIF-QA [26] requires some complex reasoning, the GIFs used are typically no longer than 3 seconds. ActivityQA [77] and Next-QA [65] have introduced open-ended QA tasks, but the annotated answers are often overly simplistic, usually consisting of just a few words. To sum up, current video QA benchmarks face limitations in video diversity and the setups of QA pairs. + +Video QA Methods and Video-LLMs. Early video QA methods employed graph structures [29, 56, 90] or transformers [17, 33, 37] to model the spatiotemporal relationships in videos and capture interactions between humans and objects. Recently, Video-LLMs [6, 45, 46] have demonstrated impressive capabilities in video understanding. Representative Video-LLMs include the VideoLLaMA [9, 47] series, LLaVA series [34, 39, 41, 70, 86], InternLM series [62, 85], and Qwen series [4, 60]. Video-LLMs take both the question and the video as input and directly generate the answer, implicitly performing the reasoning and thinking process in an efficient manner. Moreover, Some methods attempt to explicitly model the reasoning process through step-by-step thinking or multi-round debates. For example, some studies [15, 22, 57, 63, 81] leverages chain-of-thought to simulate the thinking process, while VideoAgent [14, 61, 73] improves reasoning capabilities through agent collaboration. + +Advertisement Understanding. Internet companies generate substantial profits by automatically distributing advertisements to target users, making the understanding of ad content highly important [28, 79]. The Pit dataset [24], one of the earliest efforts in automatic ad understanding, formalized this task as a visual question-answering (VQA) problem [43]. Subsequent research has built upon the Pit dataset or further developed it [53], or utilized their own private datasets [72]. These studies have mainly focused on a single aspect of advertisement understanding, such as persua + +sive strategies in image ads [31, 76], image ad search [89], intent understanding [27], and visual metaphor comprehension [3, 53, 71, 82, 83]. Although the Pit dataset primarily focuses on image advertisements, it also provided a subset of video advertisements they collected. However, the Pit dataset suffers from issues such as data inaccessibility, lack of diversity, and limited Q&A formats. Therefore, there is currently no comprehensive Video QA benchmark available in the advertising domain. + +Reinforcement Learning for Reasoning. Recent research efforts [25, 30, 44, 94] have attempted to improve the reasoning capabilities of LLMs through reinforcement learning. Recently, DeepSeek-R1 [21] achieved significant improvements in reasoning based on reinforcement learning, eliminating the need for intermediate reasoning signals. S1 [49] demonstrated that even with only a few hundred data points, the model can effectively perform reasoning in specialized tasks. However, current research on RL-based reasoning primarily focuses on math and code problems, with few works [42] attempting to apply it to multimodal domains, particularly open-ended video QA tasks. + +# 3. The AdsQA Benchmark + +# 3.1. Task Definition + +To comprehensively evaluate the model's ability to understand ad videos, we divided the questions into five types, with each one targeting a specific angle of ad analysis. + +Visual Concept Understanding (VU) task evaluates the model's ability to comprehend specific visual concepts, such as characters, objects, scenes, slogans, and other details, as well as their interrelationships. + +Emotion Recognition (ER) assesses if the models understand what emotion the ad evokes, and how ads establish connections with audiences through emotions. + +Theme and Core Message Extraction (TE) drives the models to extract the underlying message or central idea that the ad explores. This task requires deep reasoning to gather visual information from the full video. + +Persuasion Strategy Mining (PS) task evaluates the model's ability to uncover the strategies used to convey core messages and persuade audiences, such as humor, exaggeration, and visual rhetoric. The task includes questions, such as: how an ad conveys its central message, why the ad is appealing, and what strategies the ad employs. + +Potential Audience Modeling (AM) probes the model's performance in identifying potential audience groups and their profiles. It sets key questions, such as who the ad targets and what the characteristics of the audience are. This task directly reflects the ad's influence and value. + +These five tasks are defined as the open-ended video QA and each question is annotated with a ground-truth answer. + +# 3.2. Dataset Construction Pipeline + +AdsQA benchmark construction pipeline is in three stages: Pre-Processing. To serve the ad video understanding, we consider the creativity, aesthetic quality, and availability of videos during our collection process. We do not use any private data; instead, we collect creative ad videos that adhere to the Creative Commons License from a creative community platform [1]. This platform offers publicly accessible, high-quality ad videos uploaded by creators. Specifically, we first crawl videos along with their metadata. The metadata is typically written by the uploader (i.e., usually the creator of the ad) and includes information such as the theme, content, and key creative elements of the ad. It can be considered the "ground-truth" information of the ad video and serves as an important reference for our benchmark construction. Then, human experts are asked to carefully review the completeness and accuracy of the metadata and manually remove incomplete samples. Additionally, we conduct automated filtering based on video duration, aesthetic score, and content to ensure video quality and exclude non-ethical and negative content. + +Considering efficiency and accuracy, we employ PySceneDetect to segment ad videos into fine-grained video clips to avoid missing detailed information in subsequent steps. For each clip $C_i$ , we use Video-LLMs to generate its description $Desc_i$ . To preserve key visual elements, we sample $n$ keyframes $\{F_i^0, \dots, F_i^{N_{max}}\}$ from each clip based on frame similarity calculated using SSIM [64], where $n$ is determined by the duration of each clip. To retain speech information, we extract speech information $Asr_i$ for each clip using the Whisper model, and translate them into English using GPT-4 [23]. Therefore, an ad video with $N$ clips can be represented as a modality-interleaved sequence, as Eq. 1, + +$$ +V = \left\{M, \left\{F _ {i} ^ {0}, \dots , F _ {i} ^ {n _ {i}}, D e s c _ {i}, A s r _ {i} \right\} _ {i = 1} ^ {N} \right\}, \tag {1} +$$ + +where $M$ denotes the meta-information of the video. The modality-interleaved sequence will be used for automated annotation generation. + +Role-Played Multi-Agent Annotation. To balance data quality and cost, ensure data diversity, and avoid template-driven outputs, we propose a Role-Played Multi-Agent Annotation framework to automatically generate video QA pairs. Inspired by previous research [78], we found that agents can develop specialized capabilities in the ad domain by configuring specific skill sets, such as marketing, visual design, and consumer psychology. Therefore, we can leverage AI agents to act as ad experts, analyzing diverse facets of ad videos and designing more specialized and challenging questions. We ask human experts to create profiles for advertising expert skills, enabling AI agents to adaptively select profiles (or autonomously generate new profiles) to accomplish role-playing tasks. Additionally, human experts + +provide examples of QA pairs as in-context demonstration to guide the agents in generating appropriate QA pairs. + +The Role-Played Multi-Agent Annotation is formulated into a three-stage pipeline. (1) In the initial stage, a master agent is recruited to generate a preliminary QA annotation for the given modality-interleaved sequence $V$ . (2) In the iterative stage, the master agent checks the quality of the current QA annotations and decides whether to recruit a specialized expert agent. If necessary, the master agent selects a profile for the expert agent and instantiates it accordingly. The expert agent then generates new QA annotations leveraging its specialized expertise. Finally, the master agent revises the original annotations based on the expert agent's provided annotations and assesses whether to terminate the iteration or continue to recruit a new expert agent. (3) After terminating the iteration, the master agent synthesizes outputs from previous iterations to produce the final QA annotation. + +Automated Data Cleaning. After automated annotation through multi-agent collaboration, we employ the IXC-2.5-Reward model [80] to automatically verify the correctness of the answers. Given the video, question, and meta information, the reward model scores the answer. QA pairs with scores below a certain threshold are flagged and carefully reviewed by annotators. + +# 3.3. Manual Check and Quality Control + +Following the automatic annotation generation, we conduct a careful manual check and revision on all Q&A pairs, focusing on the question suitability, annotation correctness, video content, and task difficulty. To ensure the quality of Q&A pairs, we first remove questions that are homogeneous, unrelated to the ads' theme, or cannot be answered based on the ad video (e.g., those requiring external knowledge). Then, we incorporate the meta information as ground truth to verify the accuracy of the reference answers. We eliminated pairs containing inaccurate, ambiguous, or biased answers. After two rounds of rigorous selection, only $37\%$ of the QA pairs remain. Additionally, we manually revise the wording of some QA pairs to improve their clarity and accuracy. Any modifications to a sample are reviewed by other annotators to ensure consistency and reliability. + +# 3.4. Dataset Statistics + +Videos. Our AdsQA Benchmark includes a total of 1,544 unique advertisements, comprising 10,962 video clips for evaluation. As shown in Fig. 2a, these ad videos span 9 primary domains. The video durations range from 15 to 120 seconds, with an average length of 52.9 seconds, amounting to a total of 22.7 hours of content. Each ad video is accompanied by metadata, including a title, tags, automatic speech recognition (ASR) results, and descriptions. This metadata is utilized for Q&A generation, as described in Sec. 3.1. + +![](images/9c3eb2e16091b492d57aebf28575e025d78491d315cb07edc4afcecefae0c0a3.jpg) +(a) Alluvial diagram of task and domain distribution. + +![](images/34870af13fa888f42083da9d1df9b525871bc63d5f9f75ad7b8cd32135046149.jpg) +(b) Distribution of video release years. + +![](images/06177250c26462f1fdd9a3fcae91f072a445bd4ccf434c79515d1c45d42fb4d9.jpg) +(c) Distribution of video release region. + +![](images/fd712e94ee7232e8a170b984b209dd73b019e3aaa29da2e203d084256af2ff6a.jpg) +(d) The distribution of the video duration in the dataset. + +![](images/f31804cdf242274beb7906f6fc7320d21de1b1197405bf284b36e8a948429c7b.jpg) +(e) Number of Q&A entries for different tasks. + +Most of the ad videos are in English and primarily originate from North America and Europe. + +Q&A Pairs. Our AdsQA Benchmark includes 7,859 QA pairs (among which $29.2\%$ of the questions can be categorized into two types, resulting in a total of 10,154 QA pairs for evaluation). Among these: Visual Concept Understanding accounts for $21.4\%$ of the QA pairs, Theme and Message Extraction accounts for $28.5\%$ , Persuasion Strategy Mining accounts for $25.3\%$ , Potential Audience Modeling accounts for $15.1\%$ , and Emotion Recognition accounts for $9.8\%$ . Typical persuasion strategies include metaphor, symbolism, humor, expert opinion, and others. Some questions may fit into two categories, such as "What is the role of a particular visual concept in relation to the theme?" On average, each question consists of 17 words and each answer contains approximately 12 words. + +# 4. Methodology + +As illustrated in Fig. 3, we propose ReAd-R, a DeepSeekR1 styled Reinforced Ad Reasoner, to simulate human heuristic thinking and learn from trial and error. The model takes an ad video and a question as input. The policy model generates a set of reasoning processes, including thoughts and answers, based on the given input. Each reasoning process is evaluated using a reward function to compute its reward value. After calculating the reward values for all outputs, each reasoning process is assessed and used to update the policy model. Additionally, ReAd-R employs KL di + +![](images/43d0446add253e528766c953740086ad6a20ae79ce139ce2224461df6420dab0.jpg) +Figure 2. Statistics of Our AdsQA Bench. +Reasoning: "think>Loop's noise-reduction products, as inferred from the context, are intended to help individuals in noisy environments, such as subway rides, music parties, and bedrooms where noise can disturb others...nPeople who are concerned with noise reduction in their daily lives—such as commuters, festival-goers, night owls, or anyone sensitive to loud environments—would benefit the most from using Loop's noise-reducing earwear. +Figure 3. Framework of ReAd-R. Given a question and video, the policy model generates multiple responses. The reward model evaluates and scores them, and the rewards are used to update the policy model via policy gradient optimization. + +vergence to control the difference between the policy model and the reference model, ensuring stable training. + +Data Preparation. Inspired by previous work [21, 49], RL-based reasoners can improve reasoning capabilities with limited high-quality data. We additionally crawled a total of 1,053 ad videos along with their metadata and automatically generated annotations using the same method. We utilized metadata tags collected during data crawling (e.g., topic, + +domain) to ensure diversity, adopted Video-LLM evaluation results to assess difficulty, and performed manual quality assessment based on the same criteria as in Sec. 3.3. Finally, we selected 100 videos and 500 question-answer pairs. We designed a prompt format to guide the model to output its reasoning process before providing the final answer. + +Reward Modeling. The reward model guides the model to find the optimal direction through trial and error. Ideally, ReAd-R's responses should (1) include as many elements of the standard answer as possible and (2) avoid containing content not mentioned in the ad video (e.g., hallucinations). To achieve this, we propose a rule-guided LLM evaluator as the reward model to assess the quality of each response. We introduce two sets of rules: inclusion rules and exclusion rules. The inclusion rule requires that the generated answer incorporate as many elements of the standard answer as possible. If satisfied, the reward value is 1.0; otherwise, it is 0. The exclusion rule states that if the generated content includes elements not mentioned in the standard answer and these elements cannot be inferred from the meta-information, they should be judged as incorrect. To avoid sparse reward values in the early stages, we relaxed the inclusion rule. If the generated response partially includes factual elements, it is assigned a reward value of 0.5. Additionally, the format reward is used to enforce the model's predictions to adhere to the required format of and , as $R(\cdot) = R_{\text{ans}}(\cdot) + R_{\text{format}}(\cdot)$ . + +Reinforced Fine-tuning. Following DeepSeek-R1, we employ the GRPO algorithm instead of PPO to optimize our model. ReAd-R first generates $n$ distinct responses $O = \{o_1, o_2, \dots, o_n\}$ from the current policy model $\pi_{\theta_{old}}$ . Then, the reward model $R(\cdot)$ evaluates these responses $O$ to obtain their reward value as $\{r_1, r_2, \dots, r_n\}$ , and the advantage $A_i$ is defined as + +$$ +A _ {i} = \frac {r _ {i} - \operatorname {m e a n} \left(\left\{r _ {1} , r _ {2} , \dots , r _ {n} \right\}\right)}{\operatorname {s t d} \left(\left\{r _ {1} , r _ {2} , \dots , r _ {n} \right\}\right)}. \tag {2} +$$ + +$A_{i}$ denotes relative quality of the $i^{th}$ answer when compared to the group of rewards $\{r_1, r_2, \ldots, r_n\}$ . This strategy encourages the model to generate better answers with higher reward values within the group. We use GRPO algorithm [20] to optimize our policy model based on $A_{i}$ . + +# 5. Experiments + +# 5.1. Experimental Settings + +Evaluated Models. We select 14 top-tier LLMs as strong competitors covering diverse categories: (1) GPT-4 Level Baselines: including GPT-4o [23], GPT-4V [2] and Qwen2.5-VL-72B [4]. (2) Open-sourced Video LLMs: We present at least one representative model released in the past six months from each series of Video-LLMs, such as Qwen2.5-VL-7B [4], LLaVA-Onevision-7B [34], and oth + +ers. (3) Reasoning Based Methods: We re-implemented several reasoning-based methods and applied them to the AdsQA task. These methods include the Video Chain-of-Thought (VOT) [15], the Role-Played Agent (EvoAgent) [78], and the Monte Carlo Tree Search (MCTS) algorithms [84]. Detailed descriptions and implementation methods are provided in the appendix. + +Evaluation Metric. All five tasks of AdsQA are evaluated as open-ended QA, allowing for the free-text evaluation methods. Traditional free-text metrics [40, 51] are sensitive to lexical variations, which may introduce bias when evaluating different LLMs. Recent studies [8, 10, 19, 67] have shown promising results in using large generalist models to evaluate generated text. Therefore, we follow their framework and employ GPT-4o to assist in evaluating fretext similarity. Specifically, we use the same inclusion and exclusion rules mentioned in Sec. 4 to guide the scoring of text generated by the large model. We provided each sample with a ground-truth answer and meta-information; each meta-information includes relevant information such as the creative elements, content, storyline, and themes of the advertisement video. We propose both relaxed and strict scoring modes. Our strict accuracy Acc_strict is defined as: + +$$ +\operatorname {A c c} _ {\text {s t r i c t}} = \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbb {1} \left\{x _ {i} = y _ {i} \right\}, \tag {3} +$$ + +where $\mathbb{1}\{\cdot\}$ is the indicator function that returns 1 if the model's response $x_{i}$ semantically matches all elements of the standard answer $y_{i}$ , and 0 otherwise. + +In the relaxed mode, the relaxed accuracy $Acc_{\mathrm{relaxed}}$ is defined as: + +$$ +\operatorname {A c c} _ {\text {r e l a x e d}} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\mathbb {1} \left\{x _ {i} = y _ {i} \right\} + \lambda \cdot \mathbb {1} \left\{x _ {i} \approx y _ {i} \right\}\right), \tag {4} +$$ + +where $\mathbb{1}\{x_i\approx y_i\}$ returns 1 if the response $x_{i}$ partially addresses the elements of $y_{i}$ , and 0 otherwise. It is observed $\lambda = 0.5$ works well in experiments. Our prompt template for model-based evaluation are in the appendix. + +Implementation Details. Our ReAd-R is model-agnostic. In our main experiments, we utilize Qwen2-VL-7B as the base model. We froze the parameters of the visual backbone and only fine-tuned the parameters of the language model. We conducted the experiments with a batch size of 4 and a learning rate of 1e-6 on a server with eight A100 GPUs, with a training duration of 12 hours. Additional details and hyperparameters are provided in the appendix. + +# 5.2. Results and Observations + +Although GPT-4 achieves over $85\%$ accuracy on other datasets such as Next-QA, its strict accuracy on our dataset is only $29.4\%$ , and its relaxed accuracy is just $56.6\%$ . This shows that our benchmark exceeds the capabilities of some + +
ModelStrict AccuracyRelaxed Accuracy
VUERTEPSAMOverallVUERTEPSAMOverall
Human Results*55.139.457.157.545.851.374.565.975.776.264.671.4
Commercial Large Multimodal Model
GPT-4o [23]24.926.532.632.331.029.450.157.862.455.254.856.6
Gemini-2.5-Pro* [11]36.044.040.429.333.335.959.269.064.054.462.460.7
Open-sourced Video-LLMs
VideoLLaMA2-7B [9]4.567.757.286.067.386.4821.629.028.422.627.225.2
InternLM-XComp.2.5-7B [85]11.210.716.311.912.112.634.735.941.534.137.336.5
LLaVA-OneVision-7B [34]11.811.616.213.715.114.035.839.543.236.841.739.1
LLaVA-Video-7B [87]14.014.418.715.217.316.137.841.645.138.143.641.0
MiniCPM-o_2.6-7B [75]13.015.317.314.918.916.338.143.445.739.145.042.5
Qwen2-VL-7B [60]13.715.420.216.020.017.236.841.346.037.744.641.0
Qwen2.5-VL-7B [4]20.223.224.620.524.323.045.850.451.845.350.948.9
Qwen2.5-VL-72B [4]26.730.234.829.834.731.051.857.459.553.859.255.8
Reasoning-based Models
VOT (Qwen2-VL-7B) [15]14.316.419.512.622.017.036.543.342.535.245.140.2
EvolAgent (Qwen2-VL-7B) [78]15.016.121.616.520.418.338.643.446.440.445.342.6
MCTSr (Qwen2-VL-7B) [84]14.810.219.316.116.417.140.641.346.541.142.643.4
Qwen2-VL-7B (SFT) [60]10.512.517.913.815.914.134.842.142.734.941.438.8
ReAd-R (Qwen2-VL-7B) (Ours)15.819.321.718.518.818.642.346.450.043.447.844.9
ReAd-R (Qwen2.5-VL-7B) (Ours)20.427.927.222.125.525.046.256.254.648.252.651.5
+ +Table 1. Experimental Results. + +multi-modal large models. We observe that the state-of-the-art multi-modal models show strong visual perception, often generating partial answers by describing video segments. However, fully understanding the underlying meaning of ads is difficult for all baseline models. This requires not only identifying specific visual elements but also interpreting their implicit logic. + +Different tasks vary in difficulty, but their performance is comparable. Though tasks focus on different aspects, these aspects are closely linked. For example, ads may use emotions or specific visuals to convey themes. Most models perform best on Theme and Message Extraction and Audience Modeling tasks. This is because ad videos aim to deliver clear messages to specific audiences, making it relatively easier to identify themes and target users. Even without fully understanding creative elements or marketing strategies, models can infer these from characters, scenes, or slogans in the ad. The Persuasion Strategy Mining task is more challenging. It requires models to explain how and why ads use certain designs or elements. These questions are often not directly expressed in the video and may even require interpreting counterfactual or unexpected visual information, which increases the difficulty of this task. Models also perform worse on the Visual Concept Understanding (VU) task. While other video QA benchmarks [77] focus on + +surface-level questions like "What color are the gloves?", our VU task requires understanding global ad information, such as "Which scene represents the ad's theme?" Answering this question requires understanding the theme, identifying related scenes, and describing them briefly. + +On the AdsQA task, reasoning models that excel in code and math tasks achieved only marginal results, even with ten times the computational cost. This shows that AdsQA reasoning differs significantly from tasks like math and code. Math and code rely on structured "if A, then B" reasoning, while ads require associative reasoning, connecting concrete visuals to abstract concepts. Prior work [91] also suggests chain-of-thought reasoning can hinder abstract reasoning performance. Though MCSTr improved results by $2.4\%$ over Qwen2-VL, its search process is inefficient and highly dependent on the reward model. In some cases, Qwen2-VL answers correctly, but MCSTr fails due to incorrect search directions from the reward model. Moreover, training an effective reward model is difficult, especially in the data-scarce ad domain. + +Our method, ReAd-R, achieves a $3.9\%$ improvement using only 500 video-question-answer triplets, outperforming Qwen2-VL and other reasoning models. During inference, our model reasons about video content and questions by generating free-text explanations, eliminating the need for + +
Strict Acc.Relaxed Acc.
ReAd-R (Our Model)18.644.9
w. Uncurated Data13.236.2
w. Strict Reward15.141.7
w.o. Prompt Constraints16.842.9
Qwen2-VL (SFT)14.138.8
w. Uncurated Data16.040.9
VOT (Chain length: 3)17.040.2
w. (Chain length: 5)13.636.2
+ +Table 2. Ablation study on ReAd-R, Qwen2-VL SFT, and VOT. + +fixed chain-of-thought (COT) templates (e.g., VOT), additional reward functions (e.g., MCTSr), or iterative search processes (e.g., MCTSr and EvolAgent). Interestingly, we observe that supervised fine-tuning (SFT) on the same data caused a sharp performance drop of $2.2\%$ for Qwen2-VL. Using limited but high-quality data does not improve the performance of SFT models. SFT may overfit to the features of the training data, leading to poor generalization on diverse test sets. In contrast, our ReAd-R model improves reasoning capabilities through RL on limited data, achieving generalization across diverse advertisement videos. + +# 5.3. Ablation Study + +We conduct ablation studies for our ReAd-R model in Tab 2. We first show results of RL fine-tuning on 1,053 uncurated videos and 5,159 QA pairs. Our model performs worse than Qwen2-VL. This is due to the GRPO algorithm's sensitivity to data quality. Early in fine-tuning, low-quality data can mislead optimization, causing suboptimal results. Next, we test a strict reward rule (only fully correct answers receive a $R_{ans}$ reward of 1; otherwise, the reward is 0). Our method does not achieve improvement through RL fine-tuning. During training, we observe that the strict reward rule makes it hard for the model to earn rewards, limiting gains. We also find ReAd-R is sensitive to prompt templates. Properly constraining the model's thinking and answers (e.g., avoiding repetitive content in and tags) stabilizes training and improves performance. + +We also conduct ablation studies on our implemented baselines (e.g., VOT, Qwen2-VL SFT). We present the results of Qwen2-VL fine-tuned using uncurated data. Interestingly, the model fine-tuned on uncurated data improves by $2.1\%$ over limited-data fine-tuning but still lags behind the non-SFT model. This shows SFT does not enhance reasoning on ad data. Additionally, we experiment with VOT using a COT length of 5 (as set in its original paper) and observe a significant performance drop of $4.0\%$ . The model does not benefit from step-by-step reasoning; instead, accumulated errors degrade its performance. This highlights the + +![](images/d2f70e5ef148338af5e39fba0cddb1e3f378f09232e1682a79af06df15e52509.jpg) +Figure 4. Case Study. Due to the limited page space, we omit the model's thinking process and show only final answer. + +difference between abstract reasoning in the ad domain and chain-style logical reasoning. + +# 5.4. Case Study + +As shown in Fig. 4, Qwen2.5-VL generates vague and overly general answers (scored 0 by GPT-4). The COT baseline produces better answers (scored 0.5 by GPT-4) by gathering information step-by-step through a reasoning process, but it still misses some important clues in the video. Our ReAd-R learns to perform video-grounded reasoning through reinforcement fine-tuning, deriving the final answer by summarizing its reasoning process. The reinforced thinking-answer mechanism helps ReAd-R generate more complete and precise answers. + +# 6. Conclusion + +We contibute AdsQA, the first advertisement benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current "if A, then B" LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, no need of costly supervision, which enhances the specialized reasoning capability to understand implicit logic in ad videos. We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities (e.g., variants of GPT4, LLaVA, and Qwen). + +Acknowledgments. This work was supported by the National Key Research and Development Program of China (2022ZD0160603), and NSFC No.62306162. + +# References + +[1] https://www.adsofttheworld.com/.4 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6, 15 +[3] Arjun R Akula, Brendan Driscoll, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T Freeman, et al. Metaclue: Towards comprehensive visual metaphors research. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23201-23211, 2023. 3 +[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-v1 technical report. arXiv preprint arXiv:2502.13923, 2025. 3, 6, 7, 15, 16 +[5] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, Yao Dou, Jaden Park, Jianfeng Gao, Yong Jae Lee, and Jianwei Yang. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. CoRR, abs/2410.10818, 2024. 3, 16 +[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3 +[7] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 3 +[8] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 6 +[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. CoRR, abs/2406.07476, 2024. 3, 7, 15 +[10] Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023. 6 +[11] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. 7, 15 +[12] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3 +[13] Ning Ding, Shang Qu, Linhai Xie, Yifei Li, Zaoqu Liu, Kaiyan Zhang, Yibai Xiong, Yuxin Zuo, Zhangren Chen, Ermo Hua, et al. Automating exploratory pro + +teomics research via language models. arXiv preprint arXiv:2411.03743, 2024. 1 +[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 3 +[15] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 3, 6, 7 +[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 3 +[17] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14773-14783, 2023. 3 +[18] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 23923-23932, 2025. 1 +[19] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (SFD): A benchmark for story-level video understanding. CoRR, abs/2406.10221, 2024. 3, 6, 16 +[20] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1, 6 +[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 5 +[22] Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding. arXiv preprint arXiv:2502.06428, 2025. 3 +[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4, 6, 7 +[24] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In CVPR, pages 1100-1110. IEEE Computer Society, 2017. 3, 16 +[25] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander + +Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 1, 3 +[26] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 3 +[27] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12986-12996, 2021. 3 +[28] Zhiwei Jia, Pradyumna Narayana, Arjun R Akula, Garima Pruthi, Hao Su, Sugato Basu, and Varun Jampani. Kafa: Rethinking image ad understanding with knowledge-augmented feature adaptation of vision-language models. arXiv preprint arXiv:2305.18373, 2023. 3 +[29] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 3 +[30] Fangkai Jiao, Geyang Guo, Xingxing Zhang, Nancy F Chen, Shafiq Joty, and Furu Wei. Preference optimization for reasoning with pseudo feedback. arXiv preprint arXiv:2411.16345, 2024. 3 +[31] Yaman Kumar, Rajat Jha, Arunim Gupta, Milan Aggarwal, Aditya Garg, Tushar Malyan, Ayush Bhardwaj, Rajiv Ratn Shah, Balaji Krishnamurthy, and Changyou Chen. Persuasion strategies in advertisements. In Proceedings of the AAAI conference on artificial intelligence, pages 57-66, 2023. 3 +[32] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379. Association for Computational Linguistics, 2018. 3, 16 +[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021. 3 +[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 6, 7, 16 +[35] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Lou, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark. In CVPR, pages 22195-22206. IEEE, 2024. 16 +[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3 +[37] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video + +question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8658-8665, 2019. 3 +[38] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In European Conference on Computer Vision, pages 200-216. Springer, 2024. 3 +[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 3 +[40] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3 +[42] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025. 3 +[43] Xinwei Long, Zhiyuan Ma, Ermo Hua, Kaiyan Zhang, Biqing Qi, and Bowen Zhou. Retrieval-augmented visual question answering via built-in autoregressive search engines. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 24723-24731, 2025. 3 +[44] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967, 2024. 3 +[45] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. 2023. 3 +[46] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding. arxiv, 2024. 3 +[47] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS, 2023. 3, 16 +[48] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 3 +[49] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. 3, 5 +[50] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. CoRR, abs/2311.16103, 2023. 16 +[51] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6 + +[52] Biqing Qi, Kaiyan Zhang, Kai Tian, Haoxiang Li, ZhangRen Chen, Sihang Zeng, Ermo Hua, Hu Jinfang, and Bowen Zhou. Large language models as biomedical hypothesis generators: a comprehensive evaluation. arXiv preprint arXiv:2407.08940, 2024. 1 +[53] Abisek Rajakumar Kalarani, Pushpak Bhattacharyya, and Sumit Shekhar. Seeing the unseen: Visual metaphor captioning for videos. arXiv e-prints, pages arXiv-2406, 2024. 3 +[54] Ramon Sanabria, Ozan Caglayan, Shruti Palaskar, Desmond Elliott, Loic Barrault, Lucia Specia, and Florian Metze. How2: a large-scale dataset for multimodal language understanding. arXiv preprint arXiv:1811.00347, 2018. 16 +[55] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 3 +[56] Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection. arXiv preprint arXiv:2504.13092, 2025. 3 +[57] Yudi Shi, Shangzhe Di, Qirui Chen, and Weidi Xie. Unlocking video-llm via agent-of-thoughts distillation. arXiv preprint arXiv:2412.01694, 2024. 3 +[58] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 3 +[59] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640. IEEE Computer Society, 2016. 3 +[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 15 +[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 3 +[62] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3 +[63] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool. arXiv preprint arXiv:2407.05355, 2024. 3 +[64] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4 + +[65] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 9777-9786. Computer Vision Foundation / IEEE, 2021. 3 +[66] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3 +[67] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2024. 3, 6 +[68] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025.3 +[69] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645–1653, 2017. 16 +[70] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3 +[71] Yanzhi Xu, Yueying Hua, Shichen Li, and Zhongqing Wang. Exploring chain-of-thought for multi-modal metaphor detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 91-101, 2024. 3 +[72] Dingyi Yang, Chunru Zhan, Ziheng Wang, Biao Wang, Tiezheng Ge, Bo Zheng, and Qin Jin. Synchronized video storytelling: Generating video narrations with structured storyline. arXiv preprint arXiv:2405.14040, 2024. 3 +[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 3 +[74] Ruilin Yao, Bo Zhang, Jirui Huang, Xinwei Long, Yifang Zhang, Tianyu Zou, Yufei Wu, Shichao Su, Yifan Xu, Wenxi Zeng, et al. Lens: Multi-level evaluation of multimodal reasoning with large language models. arXiv preprint arXiv:2505.15616, 2025.3 +[75] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 7, 15 +[76] Keren Ye, Narges Honarvar Nazari, James Hahn, Zaeem Hussain, Mingda Zhang, and Adriana Kovashka. Interpreting the rhetoric of visual advertisements. IEEE transactions on pattern analysis and machine intelligence, 43(4):1308-1323, 2019. 3 + +[77] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134. AAAI Press, 2019. 3, 7, 15, 16 +[78] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms. arXiv preprint arXiv:2406.14228, 2024. 4, 6, 7 +[79] Majid Zahmati, Seyed Morteza Azimzadeh, Mohammad Saber Sotoodeh, and Amid Asgari. An eye-tracking study on how the popularity and gender of the endorsers affected the audience's attention on the advertisement. Electronic Commerce Research, 23(3):1665-1676, 2023. 3 +[80] Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, et al. Internlm-xcomposer2. 5-reward: A simple yet effective multi-modal reward model. arXiv preprint arXiv:2501.12368, 2025. 4 +[81] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresight-drive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025.3 +[82] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 3 +[83] Dongyu Zhang, Jingwei Yu, Senyuan Jin, Liang Yang, and Hongfei Lin. Multicmet: A novel chinese benchmark for understanding multimodal metaphor. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6141-6154, 2023. 3 +[84] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b. arXiv preprint arXiv:2406.07394, 2024. 1, 6, 7 +[85] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 3, 7, 15 +[86] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3 +[87] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7 +[88] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 15 +[89] Kang Zhao, Xinyu Zhao, Zhipeng Jin, Yi Yang, Wen Tao, Cong Han, Shuanglong Li, and Lin Liu. Enhancing baidu + +multimodal advertisement with chinese text-to-image generation via bilingual alignment and caption synthesis. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2855-2859, 2024. 3 +[90] Zhou Zhao, Zhu Zhang, Shuwen Xiao, Zhou Yu, Jun Yu, Deng Cai, Fei Wu, and Yueting Zhuang. Open-ended long-form video question answering via adaptive hierarchical reinforced networks. In IJCAI, page 8, 2018. 3 +[91] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 7 +[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3 +[93] Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding. arXiv preprint arXiv:2501.18362, 2025. 1 +[94] Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al. Trl: Test-time reinforcement learning. arXiv preprint arXiv:2504.16084, 2025.3 \ No newline at end of file diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/images.zip b/2025/AdsQA_ Towards Advertisement Video Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6061e28492dfaa33ebc0acdb506f7878e8638d2c --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c765624eb89aff95f5a38541efd2f4329d286e0a95856048676a3980241213fc +size 628922 diff --git a/2025/AdsQA_ Towards Advertisement Video Understanding/layout.json b/2025/AdsQA_ Towards Advertisement Video Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d6a91a024e6f72e68822410e041defb20b0e92c6 --- /dev/null +++ b/2025/AdsQA_ Towards Advertisement Video Understanding/layout.json @@ -0,0 +1,8934 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 474, + 121 + ], + "type": "text", + "content": "AdsQA: Towards Advertisement Video Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "spans": [ + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": "Xinwei Long" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Kai Tian" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Peng Xu" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Guoli Jia" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Jingxuan Li" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Sa Yang" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Yihua Shao" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Kaiyan Zhang" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Che Jiang" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Hao Xu" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Yang Liu" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Jiaheng Ma" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Bowen Zhou" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1,6\\boxtimes}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Tsinghua University, " + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Independent Researcher, " + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Peking University, " + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " CASIA " + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Harvard University, " + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 97, + 141, + 511, + 255 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Lab \n{longxw22, tk23}@ mails.tsinghua.edu.cn; {peng_xu, zhoubowen}@tsinghua.edu.cn \nProject page: https://github.com/TsinghuaC3I/AdsQA \nMARS2 workshop & challenge: https://lens4mllms.github.io/mars2-workshop-iccv2025/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 308, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 606 + ], + "type": "text", + "content": "Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 toptier LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "content": "A recent milestone flagged by OpenAI o1 [25] and DeepSeek-R1 [21] has been in the spotlight, and already opened up an era of large reasoning models. Meanwhile, going beyond general domains, the specialized do" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 284, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 451 + ], + "type": "text", + "content": "mains [13, 93] are increasingly studied to boost these general-purpose LLMs to evolve continuously. The specialized domains contribute significantly in optimizing and evaluating generalist models towards specialized reasoning like domain experts [18, 52]. Recently, a clear phenomenon [20, 84] has been observed that LLMs are already good at reasoning a kind of step-wise problems represented by math and programming. As has been widely discussed [20, 84], the step-wise problems based on explicit theorems and programming syntax are compatible with the chain-style reasoning approach ``if A, then B'. To further extend the diversity of specialized reasoning for LLMs, it would be better to exploit other novel domains where the domain-specific reasoning is unmanageable for \"if A, then B\"." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 589 + ], + "type": "text", + "content": "Therefore, we, for the first time, propose to use advertisement videos to probe the reasoning boundary of LLMs in perceiving implicit multimodal reasoning. The domain-unique features of clue-rich and information-dense ad videos can be summarized as key words implicit, nonphysical, mental, heuristic, etc. Different from user-uploaded videos on social media, ads are typically meticulously crafted by commercial or non-commercial organizations, making them entertaining, creative, aesthetically appealing, and capable of offering enjoyment and attracting viewer engagement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Based on these domain-specific advantages of ad videos, we introduce AdsQA, a comprehensive and carefully curated VideoQA benchmark derived from 1,544 ad videos containing 10,962 clips, spanning a total of 22.7 hours of video content. The AdsQA introduces five tasks, each requiring different types and levels of reasoning: (1) Visual Concept Understanding: Identify and analyze visual elements in ads. (2) Emotion Recognition: Detect emotions and infer their roles in ads. (3) Theme and Core Message Extraction: Summarize the central theme and key messages" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "type": "text", + "content": "*Kai Tian has equal contribution with Xinwei Long." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23396" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 220, + 349 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 220, + 349 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 220, + 349 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 220, + 349 + ], + "type": "image", + "image_path": "2a2d7424407d699f5bae0578200c1caa3d9b56e9f5e6530d8f365cfe61d06597.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 359, + 476, + 371 + ], + "lines": [ + { + "bbox": [ + 133, + 359, + 476, + 371 + ], + "spans": [ + { + "bbox": [ + 133, + 359, + 476, + 371 + ], + "type": "text", + "content": "Figure 1. Overview of AdsQA benchmark. Subfits A - B: statistics & diversity, and examples." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 226, + 77, + 386, + 348 + ], + "blocks": [ + { + "bbox": [ + 226, + 77, + 386, + 348 + ], + "lines": [ + { + "bbox": [ + 226, + 77, + 386, + 348 + ], + "spans": [ + { + "bbox": [ + 226, + 77, + 386, + 348 + ], + "type": "image", + "image_path": "0f358a0cf58c7e5d6515b1f8e8307bfe04f1fbbdf5ac2a886a1e1846d70e9d69.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 389, + 78, + 549, + 348 + ], + "blocks": [ + { + "bbox": [ + 389, + 78, + 549, + 348 + ], + "lines": [ + { + "bbox": [ + 389, + 78, + 549, + 348 + ], + "spans": [ + { + "bbox": [ + 389, + 78, + 549, + 348 + ], + "type": "image", + "image_path": "6c02812ced8ce6bd3c1969ab7c8519b559f914be6dcd90aebe080c99e0eea425.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 392, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 392, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 392, + 296, + 548 + ], + "type": "text", + "content": "of the ad. (4) Persuasion Strategy Mining: Analyze the strategies used to persuade the audience. (5) Potential Audience Modeling: Identify and characterize the target audience. The five tasks are formalized as the open-ended QA format. During the construction process, we propose an innovative Role-Played Multi-Agent Annotation framework, which simulates the role of advertising experts to generate specialized data and significantly reduces the workload of human annotators. To ensure data quality, we conduct multiple rounds of automated data cleaning and manual data revision. Moreover, based on our AdsQA, we benchmark 14 well-known LLMs, including the GPT-4 level models, recently-released reasoning-based models, etc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 557, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 296, + 715 + ], + "type": "text", + "content": "Humans process visual stimuli from ads and naturally form mental impressions, rather than relying on rigid logical reasoning templates, e.g., \"if A, then B\". Inspired by this heuristic process, we propose ReAd-R, a DeepseekR1 styled Reinforced Ad Reasoner, which learns to gather effective visual stimuli from ad videos, then reflects on the questions, and finally answers the given questions in a human-like manner. Specifically, ReAd-R utilizes a reward function to evaluate the model's responses and adjusts its parameters based on correctness, facilitating learning through trial and error. As a result, ReAd-R enhances reasoning abilities through outcome-reward-driven optimization, eliminating the need for costly step-wise su" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 464 + ], + "type": "text", + "content": "pervision or chain-of-thought (COT) training data. Experiments demonstrate that our model brings reasoning ability an obvious gain in comprehending implicit logic of ad videos. In contrast, other reasoning-based methods obtain limited improvement due to fixed COT templates and suboptimal process reward models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "content": "To our knowledge, our main contribution can be stated as: (1) AdsQA is the first video QA benchmark for advertisement domain, which is also the first ad benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current mainstream 'if A, then B' LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. Moreover, it advances ad video understanding beyond physical-content dominated shallow perception towards deeper cognitive reasoning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 665 + ], + "type": "text", + "content": "(2) We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, which enhances the specialized reasoning capability to understand implicit logic in ad videos like the human thinking process from perception to cognition. ReAd-R also can be regarded as one of the earliest attempts evolving R1-technique to vision research." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "(3) We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning (e.g., variants of GPT4, LLaVA, and Qwen) by a clear margin." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23397" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 294, + 377 + ], + "type": "text", + "content": "Video Question Answering Benchmarks. Video question answering (Video QA) [7, 48, 55, 58, 68] aims to answer user questions based on a given video, requiring a detailed understanding of both the spatio-temporal information and the relationships between objects and events [5, 12, 66, 74]. The Video QA community has developed a variety of benchmarks. From the perspective of video content, most benchmarks focus on human-centric videos, such as NextQA [65], ActivityQA [77], MVBench [36], FunQA [67], and VideoMME [16], which are derived from movies [19, 59], TV shows [32], social media [67], etc. Additionally, there are benchmarks centered on object-centric videos, such as EgoSchema [47], which are sourced from instructional and operational videos [38, 92]. Despite their advancements, these benchmarks still have several limitations. MovieQA [59] overly relies on dialogue understanding, which restricts the in-depth comprehension of visual content. Although TGIF-QA [26] requires some complex reasoning, the GIFs used are typically no longer than 3 seconds. ActivityQA [77] and Next-QA [65] have introduced open-ended QA tasks, but the annotated answers are often overly simplistic, usually consisting of just a few words. To sum up, current video QA benchmarks face limitations in video diversity and the setups of QA pairs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 379, + 294, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 379, + 294, + 593 + ], + "spans": [ + { + "bbox": [ + 57, + 379, + 294, + 593 + ], + "type": "text", + "content": "Video QA Methods and Video-LLMs. Early video QA methods employed graph structures [29, 56, 90] or transformers [17, 33, 37] to model the spatiotemporal relationships in videos and capture interactions between humans and objects. Recently, Video-LLMs [6, 45, 46] have demonstrated impressive capabilities in video understanding. Representative Video-LLMs include the VideoLLaMA [9, 47] series, LLaVA series [34, 39, 41, 70, 86], InternLM series [62, 85], and Qwen series [4, 60]. Video-LLMs take both the question and the video as input and directly generate the answer, implicitly performing the reasoning and thinking process in an efficient manner. Moreover, Some methods attempt to explicitly model the reasoning process through step-by-step thinking or multi-round debates. For example, some studies [15, 22, 57, 63, 81] leverages chain-of-thought to simulate the thinking process, while VideoAgent [14, 61, 73] improves reasoning capabilities through agent collaboration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 714 + ], + "type": "text", + "content": "Advertisement Understanding. Internet companies generate substantial profits by automatically distributing advertisements to target users, making the understanding of ad content highly important [28, 79]. The Pit dataset [24], one of the earliest efforts in automatic ad understanding, formalized this task as a visual question-answering (VQA) problem [43]. Subsequent research has built upon the Pit dataset or further developed it [53], or utilized their own private datasets [72]. These studies have mainly focused on a single aspect of advertisement understanding, such as persua" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "sive strategies in image ads [31, 76], image ad search [89], intent understanding [27], and visual metaphor comprehension [3, 53, 71, 82, 83]. Although the Pit dataset primarily focuses on image advertisements, it also provided a subset of video advertisements they collected. However, the Pit dataset suffers from issues such as data inaccessibility, lack of diversity, and limited Q&A formats. Therefore, there is currently no comprehensive Video QA benchmark available in the advertising domain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 182, + 553, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 325 + ], + "type": "text", + "content": "Reinforcement Learning for Reasoning. Recent research efforts [25, 30, 44, 94] have attempted to improve the reasoning capabilities of LLMs through reinforcement learning. Recently, DeepSeek-R1 [21] achieved significant improvements in reasoning based on reinforcement learning, eliminating the need for intermediate reasoning signals. S1 [49] demonstrated that even with only a few hundred data points, the model can effectively perform reasoning in specialized tasks. However, current research on RL-based reasoning primarily focuses on math and code problems, with few works [42] attempting to apply it to multimodal domains, particularly open-ended video QA tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 341, + 451, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 451, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 451, + 354 + ], + "type": "text", + "content": "3. The AdsQA Benchmark" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 408, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 408, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 408, + 373 + ], + "type": "text", + "content": "3.1. Task Definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 417 + ], + "type": "text", + "content": "To comprehensively evaluate the model's ability to understand ad videos, we divided the questions into five types, with each one targeting a specific angle of ad analysis." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 418, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 466 + ], + "type": "text", + "content": "Visual Concept Understanding (VU) task evaluates the model's ability to comprehend specific visual concepts, such as characters, objects, scenes, slogans, and other details, as well as their interrelationships." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 553, + 504 + ], + "type": "text", + "content": "Emotion Recognition (ER) assesses if the models understand what emotion the ad evokes, and how ads establish connections with audiences through emotions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 505, + 553, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 553, + 553 + ], + "type": "text", + "content": "Theme and Core Message Extraction (TE) drives the models to extract the underlying message or central idea that the ad explores. This task requires deep reasoning to gather visual information from the full video." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 555, + 553, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 553, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 553, + 627 + ], + "type": "text", + "content": "Persuasion Strategy Mining (PS) task evaluates the model's ability to uncover the strategies used to convey core messages and persuade audiences, such as humor, exaggeration, and visual rhetoric. The task includes questions, such as: how an ad conveys its central message, why the ad is appealing, and what strategies the ad employs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 628, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 687 + ], + "type": "text", + "content": "Potential Audience Modeling (AM) probes the model's performance in identifying potential audience groups and their profiles. It sets key questions, such as who the ad targets and what the characteristics of the audience are. This task directly reflects the ad's influence and value." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "These five tasks are defined as the open-ended video QA and each question is annotated with a ground-truth answer." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23398" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 219, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 219, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 219, + 85 + ], + "type": "text", + "content": "3.2. Dataset Construction Pipeline" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 294, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 294, + 328 + ], + "type": "text", + "content": "AdsQA benchmark construction pipeline is in three stages: Pre-Processing. To serve the ad video understanding, we consider the creativity, aesthetic quality, and availability of videos during our collection process. We do not use any private data; instead, we collect creative ad videos that adhere to the Creative Commons License from a creative community platform [1]. This platform offers publicly accessible, high-quality ad videos uploaded by creators. Specifically, we first crawl videos along with their metadata. The metadata is typically written by the uploader (i.e., usually the creator of the ad) and includes information such as the theme, content, and key creative elements of the ad. It can be considered the \"ground-truth\" information of the ad video and serves as an important reference for our benchmark construction. Then, human experts are asked to carefully review the completeness and accuracy of the metadata and manually remove incomplete samples. Additionally, we conduct automated filtering based on video duration, aesthetic score, and content to ensure video quality and exclude non-ethical and negative content." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": "Considering efficiency and accuracy, we employ PySceneDetect to segment ad videos into fine-grained video clips to avoid missing detailed information in subsequent steps. For each clip " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": ", we use Video-LLMs to generate its description " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "Desc_i" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": ". To preserve key visual elements, we sample " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": " keyframes " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "\\{F_i^0, \\dots, F_i^{N_{max}}\\}" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": " from each clip based on frame similarity calculated using SSIM [64], where " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": " is determined by the duration of each clip. To retain speech information, we extract speech information " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "Asr_i" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": " for each clip using the Whisper model, and translate them into English using GPT-4 [23]. Therefore, an ad video with " + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 329, + 295, + 484 + ], + "type": "text", + "content": " clips can be represented as a modality-interleaved sequence, as Eq. 1," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 490, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 490, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 87, + 490, + 294, + 506 + ], + "type": "interline_equation", + "content": "V = \\left\\{M, \\left\\{F _ {i} ^ {0}, \\dots , F _ {i} ^ {n _ {i}}, D e s c _ {i}, A s r _ {i} \\right\\} _ {i = 1} ^ {N} \\right\\}, \\tag {1}", + "image_path": "12de4a71ff1692d1275e9506049f4c1fc52f9bdef58341019fc265e84735c3c9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": " denotes the meta-information of the video. The modality-interleaved sequence will be used for automated annotation generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Role-Played Multi-Agent Annotation. To balance data quality and cost, ensure data diversity, and avoid template-driven outputs, we propose a Role-Played Multi-Agent Annotation framework to automatically generate video QA pairs. Inspired by previous research [78], we found that agents can develop specialized capabilities in the ad domain by configuring specific skill sets, such as marketing, visual design, and consumer psychology. Therefore, we can leverage AI agents to act as ad experts, analyzing diverse facets of ad videos and designing more specialized and challenging questions. We ask human experts to create profiles for advertising expert skills, enabling AI agents to adaptively select profiles (or autonomously generate new profiles) to accomplish role-playing tasks. Additionally, human experts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "provide examples of QA pairs as in-context demonstration to guide the agents in generating appropriate QA pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 97, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 287 + ], + "type": "text", + "content": "The Role-Played Multi-Agent Annotation is formulated into a three-stage pipeline. (1) In the initial stage, a master agent is recruited to generate a preliminary QA annotation for the given modality-interleaved sequence " + }, + { + "bbox": [ + 313, + 97, + 553, + 287 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 97, + 553, + 287 + ], + "type": "text", + "content": ". (2) In the iterative stage, the master agent checks the quality of the current QA annotations and decides whether to recruit a specialized expert agent. If necessary, the master agent selects a profile for the expert agent and instantiates it accordingly. The expert agent then generates new QA annotations leveraging its specialized expertise. Finally, the master agent revises the original annotations based on the expert agent's provided annotations and assesses whether to terminate the iteration or continue to recruit a new expert agent. (3) After terminating the iteration, the master agent synthesizes outputs from previous iterations to produce the final QA annotation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 553, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 373 + ], + "type": "text", + "content": "Automated Data Cleaning. After automated annotation through multi-agent collaboration, we employ the IXC-2.5-Reward model [80] to automatically verify the correctness of the answers. Given the video, question, and meta information, the reward model scores the answer. QA pairs with scores below a certain threshold are flagged and carefully reviewed by annotators." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 504, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 504, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 504, + 395 + ], + "type": "text", + "content": "3.3. Manual Check and Quality Control" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 579 + ], + "type": "text", + "content": "Following the automatic annotation generation, we conduct a careful manual check and revision on all Q&A pairs, focusing on the question suitability, annotation correctness, video content, and task difficulty. To ensure the quality of Q&A pairs, we first remove questions that are homogeneous, unrelated to the ads' theme, or cannot be answered based on the ad video (e.g., those requiring external knowledge). Then, we incorporate the meta information as ground truth to verify the accuracy of the reference answers. We eliminated pairs containing inaccurate, ambiguous, or biased answers. After two rounds of rigorous selection, only " + }, + { + "bbox": [ + 313, + 399, + 555, + 579 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 313, + 399, + 555, + 579 + ], + "type": "text", + "content": " of the QA pairs remain. Additionally, we manually revise the wording of some QA pairs to improve their clarity and accuracy. Any modifications to a sample are reviewed by other annotators to ensure consistency and reliability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 417, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 417, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 417, + 600 + ], + "type": "text", + "content": "3.4. Dataset Statistics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Videos. Our AdsQA Benchmark includes a total of 1,544 unique advertisements, comprising 10,962 video clips for evaluation. As shown in Fig. 2a, these ad videos span 9 primary domains. The video durations range from 15 to 120 seconds, with an average length of 52.9 seconds, amounting to a total of 22.7 hours of content. Each ad video is accompanied by metadata, including a title, tags, automatic speech recognition (ASR) results, and descriptions. This metadata is utilized for Q&A generation, as described in Sec. 3.1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23399" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 223, + 308 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 223, + 308 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 223, + 308 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 223, + 308 + ], + "type": "image", + "image_path": "9c3eb2e16091b492d57aebf28575e025d78491d315cb07edc4afcecefae0c0a3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 312, + 217, + 319 + ], + "lines": [ + { + "bbox": [ + 86, + 312, + 217, + 319 + ], + "spans": [ + { + "bbox": [ + 86, + 312, + 217, + 319 + ], + "type": "text", + "content": "(a) Alluvial diagram of task and domain distribution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 82, + 388, + 194 + ], + "blocks": [ + { + "bbox": [ + 228, + 82, + 388, + 194 + ], + "lines": [ + { + "bbox": [ + 228, + 82, + 388, + 194 + ], + "spans": [ + { + "bbox": [ + 228, + 82, + 388, + 194 + ], + "type": "image", + "image_path": "34870af13fa888f42083da9d1df9b525871bc63d5f9f75ad7b8cd32135046149.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 198, + 359, + 206 + ], + "lines": [ + { + "bbox": [ + 263, + 198, + 359, + 206 + ], + "spans": [ + { + "bbox": [ + 263, + 198, + 359, + 206 + ], + "type": "text", + "content": "(b) Distribution of video release years." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 81, + 512, + 194 + ], + "blocks": [ + { + "bbox": [ + 402, + 81, + 512, + 194 + ], + "lines": [ + { + "bbox": [ + 402, + 81, + 512, + 194 + ], + "spans": [ + { + "bbox": [ + 402, + 81, + 512, + 194 + ], + "type": "image", + "image_path": "06177250c26462f1fdd9a3fcae91f072a445bd4ccf434c79515d1c45d42fb4d9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 198, + 503, + 206 + ], + "lines": [ + { + "bbox": [ + 405, + 198, + 503, + 206 + ], + "spans": [ + { + "bbox": [ + 405, + 198, + 503, + 206 + ], + "type": "text", + "content": "(c) Distribution of video release region." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 231, + 209, + 373, + 306 + ], + "blocks": [ + { + "bbox": [ + 231, + 209, + 373, + 306 + ], + "lines": [ + { + "bbox": [ + 231, + 209, + 373, + 306 + ], + "spans": [ + { + "bbox": [ + 231, + 209, + 373, + 306 + ], + "type": "image", + "image_path": "fd712e94ee7232e8a170b984b209dd73b019e3aaa29da2e203d084256af2ff6a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 235, + 311, + 372, + 319 + ], + "lines": [ + { + "bbox": [ + 235, + 311, + 372, + 319 + ], + "spans": [ + { + "bbox": [ + 235, + 311, + 372, + 319 + ], + "type": "text", + "content": "(d) The distribution of the video duration in the dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 381, + 208, + 526, + 306 + ], + "blocks": [ + { + "bbox": [ + 381, + 208, + 526, + 306 + ], + "lines": [ + { + "bbox": [ + 381, + 208, + 526, + 306 + ], + "spans": [ + { + "bbox": [ + 381, + 208, + 526, + 306 + ], + "type": "image", + "image_path": "f31804cdf242274beb7906f6fc7320d21de1b1197405bf284b36e8a948429c7b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 311, + 513, + 319 + ], + "lines": [ + { + "bbox": [ + 397, + 311, + 513, + 319 + ], + "spans": [ + { + "bbox": [ + 397, + 311, + 513, + 319 + ], + "type": "text", + "content": "(e) Number of Q&A entries for different tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "content": "Most of the ad videos are in English and primarily originate from North America and Europe." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": "Q&A Pairs. Our AdsQA Benchmark includes 7,859 QA pairs (among which " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "29.2\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": " of the questions can be categorized into two types, resulting in a total of 10,154 QA pairs for evaluation). Among these: Visual Concept Understanding accounts for " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "21.4\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": " of the QA pairs, Theme and Message Extraction accounts for " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": ", Persuasion Strategy Mining accounts for " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "25.3\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": ", Potential Audience Modeling accounts for " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": ", and Emotion Recognition accounts for " + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 55, + 390, + 296, + 559 + ], + "type": "text", + "content": ". Typical persuasion strategies include metaphor, symbolism, humor, expert opinion, and others. Some questions may fit into two categories, such as \"What is the role of a particular visual concept in relation to the theme?\" On average, each question consists of 17 words and each answer contains approximately 12 words." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 572, + 138, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 138, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 138, + 586 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 3, we propose ReAd-R, a DeepSeekR1 styled Reinforced Ad Reasoner, to simulate human heuristic thinking and learn from trial and error. The model takes an ad video and a question as input. The policy model generates a set of reasoning processes, including thoughts and answers, based on the given input. Each reasoning process is evaluated using a reward function to compute its reward value. After calculating the reward values for all outputs, each reasoning process is assessed and used to update the policy model. Additionally, ReAd-R employs KL di" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 323, + 363, + 545, + 492 + ], + "blocks": [ + { + "bbox": [ + 228, + 332, + 380, + 344 + ], + "lines": [ + { + "bbox": [ + 228, + 332, + 380, + 344 + ], + "spans": [ + { + "bbox": [ + 228, + 332, + 380, + 344 + ], + "type": "text", + "content": "Figure 2. Statistics of Our AdsQA Bench." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 363, + 545, + 492 + ], + "lines": [ + { + "bbox": [ + 323, + 363, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 323, + 363, + 545, + 492 + ], + "type": "image", + "image_path": "43d0446add253e528766c953740086ad6a20ae79ce139ce2224461df6420dab0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 492, + 544, + 534 + ], + "lines": [ + { + "bbox": [ + 325, + 492, + 544, + 534 + ], + "spans": [ + { + "bbox": [ + 325, + 492, + 544, + 534 + ], + "type": "text", + "content": "Reasoning: \"think>Loop's noise-reduction products, as inferred from the context, are intended to help individuals in noisy environments, such as subway rides, music parties, and bedrooms where noise can disturb others...nPeople who are concerned with noise reduction in their daily lives—such as commuters, festival-goers, night owls, or anyone sensitive to loud environments—would benefit the most from using Loop's noise-reducing earwear." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 590 + ], + "type": "text", + "content": "Figure 3. Framework of ReAd-R. Given a question and video, the policy model generates multiple responses. The reward model evaluates and scores them, and the rewards are used to update the policy model via policy gradient optimization." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 640 + ], + "type": "text", + "content": "vergence to control the difference between the policy model and the reference model, ensuring stable training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Data Preparation. Inspired by previous work [21, 49], RL-based reasoners can improve reasoning capabilities with limited high-quality data. We additionally crawled a total of 1,053 ad videos along with their metadata and automatically generated annotations using the same method. We utilized metadata tags collected during data crawling (e.g., topic," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23400" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "domain) to ensure diversity, adopted Video-LLM evaluation results to assess difficulty, and performed manual quality assessment based on the same criteria as in Sec. 3.3. Finally, we selected 100 videos and 500 question-answer pairs. We designed a prompt format to guide the model to output its reasoning process before providing the final answer." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 396 + ], + "type": "text", + "content": "Reward Modeling. The reward model guides the model to find the optimal direction through trial and error. Ideally, ReAd-R's responses should (1) include as many elements of the standard answer as possible and (2) avoid containing content not mentioned in the ad video (e.g., hallucinations). To achieve this, we propose a rule-guided LLM evaluator as the reward model to assess the quality of each response. We introduce two sets of rules: inclusion rules and exclusion rules. The inclusion rule requires that the generated answer incorporate as many elements of the standard answer as possible. If satisfied, the reward value is 1.0; otherwise, it is 0. The exclusion rule states that if the generated content includes elements not mentioned in the standard answer and these elements cannot be inferred from the meta-information, they should be judged as incorrect. To avoid sparse reward values in the early stages, we relaxed the inclusion rule. If the generated response partially includes factual elements, it is assigned a reward value of 0.5. Additionally, the format reward is used to enforce the model's predictions to adhere to the required format of and , as " + }, + { + "bbox": [ + 55, + 144, + 295, + 396 + ], + "type": "inline_equation", + "content": "R(\\cdot) = R_{\\text{ans}}(\\cdot) + R_{\\text{format}}(\\cdot)" + }, + { + "bbox": [ + 55, + 144, + 295, + 396 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": "Reinforced Fine-tuning. Following DeepSeek-R1, we employ the GRPO algorithm instead of PPO to optimize our model. ReAd-R first generates " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " distinct responses " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "O = \\{o_1, o_2, \\dots, o_n\\}" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " from the current policy model " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_{old}}" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": ". Then, the reward model " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "R(\\cdot)" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " evaluates these responses " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " to obtain their reward value as " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\{r_1, r_2, \\dots, r_n\\}" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": ", and the advantage " + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 55, + 396, + 295, + 479 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 487, + 294, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 487, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 294, + 513 + ], + "type": "interline_equation", + "content": "A _ {i} = \\frac {r _ {i} - \\operatorname {m e a n} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {1} , r _ {2} , \\dots , r _ {n} \\right\\}\\right)}. \\tag {2}", + "image_path": "f69c1dcf1f1ba00caaeb25043b693b621b0c15593860bcd922d82d8bf126c20f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "text", + "content": " denotes relative quality of the " + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "text", + "content": " answer when compared to the group of rewards " + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\{r_1, r_2, \\ldots, r_n\\}" + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "text", + "content": ". This strategy encourages the model to generate better answers with higher reward values within the group. We use GRPO algorithm [20] to optimize our policy model based on " + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 55, + 521, + 295, + 583 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 137, + 605 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 182, + 625 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Evaluated Models. We select 14 top-tier LLMs as strong competitors covering diverse categories: (1) GPT-4 Level Baselines: including GPT-4o [23], GPT-4V [2] and Qwen2.5-VL-72B [4]. (2) Open-sourced Video LLMs: We present at least one representative model released in the past six months from each series of Video-LLMs, such as Qwen2.5-VL-7B [4], LLaVA-Onevision-7B [34], and oth" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ers. (3) Reasoning Based Methods: We re-implemented several reasoning-based methods and applied them to the AdsQA task. These methods include the Video Chain-of-Thought (VOT) [15], the Role-Played Agent (EvoAgent) [78], and the Monte Carlo Tree Search (MCTS) algorithms [84]. Detailed descriptions and implementation methods are provided in the appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 156, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 348 + ], + "type": "text", + "content": "Evaluation Metric. All five tasks of AdsQA are evaluated as open-ended QA, allowing for the free-text evaluation methods. Traditional free-text metrics [40, 51] are sensitive to lexical variations, which may introduce bias when evaluating different LLMs. Recent studies [8, 10, 19, 67] have shown promising results in using large generalist models to evaluate generated text. Therefore, we follow their framework and employ GPT-4o to assist in evaluating fretext similarity. Specifically, we use the same inclusion and exclusion rules mentioned in Sec. 4 to guide the scoring of text generated by the large model. We provided each sample with a ground-truth answer and meta-information; each meta-information includes relevant information such as the creative elements, content, storyline, and themes of the advertisement video. We propose both relaxed and strict scoring modes. Our strict accuracy Acc_strict is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 355, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 355, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 372, + 355, + 553, + 386 + ], + "type": "interline_equation", + "content": "\\operatorname {A c c} _ {\\text {s t r i c t}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\}, \\tag {3}", + "image_path": "e4a3a16d9696283714acdacda16198ba7e25a3ca817e5453bff28dde6428b9ef.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "inline_equation", + "content": "\\mathbb{1}\\{\\cdot\\}" + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "content": " is the indicator function that returns 1 if the model's response " + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "content": " semantically matches all elements of the standard answer " + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 393, + 555, + 430 + ], + "type": "text", + "content": ", and 0 otherwise." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 430, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 452 + ], + "type": "text", + "content": "In the relaxed mode, the relaxed accuracy " + }, + { + "bbox": [ + 313, + 430, + 554, + 452 + ], + "type": "inline_equation", + "content": "Acc_{\\mathrm{relaxed}}" + }, + { + "bbox": [ + 313, + 430, + 554, + 452 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 460, + 553, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 324, + 460, + 553, + 491 + ], + "type": "interline_equation", + "content": "\\operatorname {A c c} _ {\\text {r e l a x e d}} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\mathbb {1} \\left\\{x _ {i} = y _ {i} \\right\\} + \\lambda \\cdot \\mathbb {1} \\left\\{x _ {i} \\approx y _ {i} \\right\\}\\right), \\tag {4}", + "image_path": "9a9847a847a37940d74526512a169f530f85dc239c8497f8924191874ee98c36.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbb{1}\\{x_i\\approx y_i\\}" + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "content": " returns 1 if the response " + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "content": " partially addresses the elements of " + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "content": ", and 0 otherwise. It is observed " + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\lambda = 0.5" + }, + { + "bbox": [ + 313, + 497, + 554, + 545 + ], + "type": "text", + "content": " works well in experiments. Our prompt template for model-based evaluation are in the appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 641 + ], + "type": "text", + "content": "Implementation Details. Our ReAd-R is model-agnostic. In our main experiments, we utilize Qwen2-VL-7B as the base model. We froze the parameters of the visual backbone and only fine-tuned the parameters of the language model. We conducted the experiments with a batch size of 4 and a learning rate of 1e-6 on a server with eight A100 GPUs, with a training duration of 12 hours. Additional details and hyperparameters are provided in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 648, + 456, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 456, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 456, + 659 + ], + "type": "text", + "content": "5.2. Results and Observations" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Although GPT-4 achieves over " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " accuracy on other datasets such as Next-QA, its strict accuracy on our dataset is only " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "29.4\\%" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", and its relaxed accuracy is just " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "56.6\\%" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ". This shows that our benchmark exceeds the capabilities of some" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23401" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 370 + ], + "type": "table", + "html": "
ModelStrict AccuracyRelaxed Accuracy
VUERTEPSAMOverallVUERTEPSAMOverall
Human Results*55.139.457.157.545.851.374.565.975.776.264.671.4
Commercial Large Multimodal Model
GPT-4o [23]24.926.532.632.331.029.450.157.862.455.254.856.6
Gemini-2.5-Pro* [11]36.044.040.429.333.335.959.269.064.054.462.460.7
Open-sourced Video-LLMs
VideoLLaMA2-7B [9]4.567.757.286.067.386.4821.629.028.422.627.225.2
InternLM-XComp.2.5-7B [85]11.210.716.311.912.112.634.735.941.534.137.336.5
LLaVA-OneVision-7B [34]11.811.616.213.715.114.035.839.543.236.841.739.1
LLaVA-Video-7B [87]14.014.418.715.217.316.137.841.645.138.143.641.0
MiniCPM-o_2.6-7B [75]13.015.317.314.918.916.338.143.445.739.145.042.5
Qwen2-VL-7B [60]13.715.420.216.020.017.236.841.346.037.744.641.0
Qwen2.5-VL-7B [4]20.223.224.620.524.323.045.850.451.845.350.948.9
Qwen2.5-VL-72B [4]26.730.234.829.834.731.051.857.459.553.859.255.8
Reasoning-based Models
VOT (Qwen2-VL-7B) [15]14.316.419.512.622.017.036.543.342.535.245.140.2
EvolAgent (Qwen2-VL-7B) [78]15.016.121.616.520.418.338.643.446.440.445.342.6
MCTSr (Qwen2-VL-7B) [84]14.810.219.316.116.417.140.641.346.541.142.643.4
Qwen2-VL-7B (SFT) [60]10.512.517.913.815.914.134.842.142.734.941.438.8
ReAd-R (Qwen2-VL-7B) (Ours)15.819.321.718.518.818.642.346.450.043.447.844.9
ReAd-R (Qwen2.5-VL-7B) (Ours)20.427.927.222.125.525.046.256.254.648.252.651.5
", + "image_path": "82a46d8678949b682964e7115aee82579d2b54d022ca73cc7921f7b8b8db381d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 247, + 377, + 362, + 389 + ], + "lines": [ + { + "bbox": [ + 247, + 377, + 362, + 389 + ], + "spans": [ + { + "bbox": [ + 247, + 377, + 362, + 389 + ], + "type": "text", + "content": "Table 1. Experimental Results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 409, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 294, + 495 + ], + "type": "text", + "content": "multi-modal large models. We observe that the state-of-the-art multi-modal models show strong visual perception, often generating partial answers by describing video segments. However, fully understanding the underlying meaning of ads is difficult for all baseline models. This requires not only identifying specific visual elements but also interpreting their implicit logic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "Different tasks vary in difficulty, but their performance is comparable. Though tasks focus on different aspects, these aspects are closely linked. For example, ads may use emotions or specific visuals to convey themes. Most models perform best on Theme and Message Extraction and Audience Modeling tasks. This is because ad videos aim to deliver clear messages to specific audiences, making it relatively easier to identify themes and target users. Even without fully understanding creative elements or marketing strategies, models can infer these from characters, scenes, or slogans in the ad. The Persuasion Strategy Mining task is more challenging. It requires models to explain how and why ads use certain designs or elements. These questions are often not directly expressed in the video and may even require interpreting counterfactual or unexpected visual information, which increases the difficulty of this task. Models also perform worse on the Visual Concept Understanding (VU) task. While other video QA benchmarks [77] focus on" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 409, + 553, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 553, + 470 + ], + "type": "text", + "content": "surface-level questions like \"What color are the gloves?\", our VU task requires understanding global ad information, such as \"Which scene represents the ad's theme?\" Answering this question requires understanding the theme, identifying related scenes, and describing them briefly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 472, + 554, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 651 + ], + "type": "text", + "content": "On the AdsQA task, reasoning models that excel in code and math tasks achieved only marginal results, even with ten times the computational cost. This shows that AdsQA reasoning differs significantly from tasks like math and code. Math and code rely on structured \"if A, then B\" reasoning, while ads require associative reasoning, connecting concrete visuals to abstract concepts. Prior work [91] also suggests chain-of-thought reasoning can hinder abstract reasoning performance. Though MCSTr improved results by " + }, + { + "bbox": [ + 313, + 472, + 554, + 651 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 472, + 554, + 651 + ], + "type": "text", + "content": " over Qwen2-VL, its search process is inefficient and highly dependent on the reward model. In some cases, Qwen2-VL answers correctly, but MCSTr fails due to incorrect search directions from the reward model. Moreover, training an effective reward model is difficult, especially in the data-scarce ad domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Our method, ReAd-R, achieves a " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " improvement using only 500 video-question-answer triplets, outperforming Qwen2-VL and other reasoning models. During inference, our model reasons about video content and questions by generating free-text explanations, eliminating the need for" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23402" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 194 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 194 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 194 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 194 + ], + "type": "table", + "html": "
Strict Acc.Relaxed Acc.
ReAd-R (Our Model)18.644.9
w. Uncurated Data13.236.2
w. Strict Reward15.141.7
w.o. Prompt Constraints16.842.9
Qwen2-VL (SFT)14.138.8
w. Uncurated Data16.040.9
VOT (Chain length: 3)17.040.2
w. (Chain length: 5)13.636.2
", + "image_path": "424e9fa799f80ed841a3d57493321b04081976f50e8193b3fae688c5d74a05cd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 202, + 293, + 213 + ], + "lines": [ + { + "bbox": [ + 57, + 202, + 293, + 213 + ], + "spans": [ + { + "bbox": [ + 57, + 202, + 293, + 213 + ], + "type": "text", + "content": "Table 2. Ablation study on ReAd-R, Qwen2-VL SFT, and VOT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "text", + "content": "fixed chain-of-thought (COT) templates (e.g., VOT), additional reward functions (e.g., MCTSr), or iterative search processes (e.g., MCTSr and EvolAgent). Interestingly, we observe that supervised fine-tuning (SFT) on the same data caused a sharp performance drop of " + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 55, + 234, + 296, + 368 + ], + "type": "text", + "content": " for Qwen2-VL. Using limited but high-quality data does not improve the performance of SFT models. SFT may overfit to the features of the training data, leading to poor generalization on diverse test sets. In contrast, our ReAd-R model improves reasoning capabilities through RL on limited data, achieving generalization across diverse advertisement videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 149, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 149, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 149, + 386 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 581 + ], + "type": "text", + "content": "We conduct ablation studies for our ReAd-R model in Tab 2. We first show results of RL fine-tuning on 1,053 uncurated videos and 5,159 QA pairs. Our model performs worse than Qwen2-VL. This is due to the GRPO algorithm's sensitivity to data quality. Early in fine-tuning, low-quality data can mislead optimization, causing suboptimal results. Next, we test a strict reward rule (only fully correct answers receive a " + }, + { + "bbox": [ + 55, + 390, + 296, + 581 + ], + "type": "inline_equation", + "content": "R_{ans}" + }, + { + "bbox": [ + 55, + 390, + 296, + 581 + ], + "type": "text", + "content": " reward of 1; otherwise, the reward is 0). Our method does not achieve improvement through RL fine-tuning. During training, we observe that the strict reward rule makes it hard for the model to earn rewards, limiting gains. We also find ReAd-R is sensitive to prompt templates. Properly constraining the model's thinking and answers (e.g., avoiding repetitive content in and tags) stabilizes training and improves performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "We also conduct ablation studies on our implemented baselines (e.g., VOT, Qwen2-VL SFT). We present the results of Qwen2-VL fine-tuned using uncurated data. Interestingly, the model fine-tuned on uncurated data improves by " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " over limited-data fine-tuning but still lags behind the non-SFT model. This shows SFT does not enhance reasoning on ad data. Additionally, we experiment with VOT using a COT length of 5 (as set in its original paper) and observe a significant performance drop of " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "4.0\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". The model does not benefit from step-by-step reasoning; instead, accumulated errors degrade its performance. This highlights the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 322, + 70, + 547, + 248 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 547, + 248 + ], + "type": "image", + "image_path": "d2f70e5ef148338af5e39fba0cddb1e3f378f09232e1682a79af06df15e52509.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 281 + ], + "type": "text", + "content": "Figure 4. Case Study. Due to the limited page space, we omit the model's thinking process and show only final answer." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "text", + "content": "difference between abstract reasoning in the ad domain and chain-style logical reasoning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 389, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 389, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 389, + 355 + ], + "type": "text", + "content": "5.4. Case Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 482 + ], + "type": "text", + "content": "As shown in Fig. 4, Qwen2.5-VL generates vague and overly general answers (scored 0 by GPT-4). The COT baseline produces better answers (scored 0.5 by GPT-4) by gathering information step-by-step through a reasoning process, but it still misses some important clues in the video. Our ReAd-R learns to perform video-grounded reasoning through reinforcement fine-tuning, deriving the final answer by summarizing its reasoning process. The reinforced thinking-answer mechanism helps ReAd-R generate more complete and precise answers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 388, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 388, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 388, + 510 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 520, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 675 + ], + "type": "text", + "content": "We contibute AdsQA, the first advertisement benchmark for LLMs, with domain-unique features implicit, non-physical, mental, heuristic, etc. It presents a new challenge to the current \"if A, then B\" LLM thinking approach, hence further extend the domain specialization reasoning scenarios for LLMs. We propose ReAd-R, a DeepSeek-R1 styled RL reasoning model, no need of costly supervision, which enhances the specialized reasoning capability to understand implicit logic in ad videos. We benchmark 14 flagship LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities (e.g., variants of GPT4, LLaVA, and Qwen)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgments. This work was supported by the National Key Research and Development Program of China (2022ZD0160603), and NSFC No.62306162." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23403" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 250, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 250, + 101 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 250, + 101 + ], + "type": "text", + "content": "[1] https://www.adsofttheworld.com/.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 103, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 295, + 158 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6, 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 160, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 160, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 160, + 295, + 236 + ], + "type": "text", + "content": "[3] Arjun R Akula, Brendan Driscoll, Pradyumna Narayana, Soravit Changpinyo, Zhiwei Jia, Suyash Damle, Garima Pruthi, Sugato Basu, Leonidas Guibas, William T Freeman, et al. Metaclue: Towards comprehensive visual metaphors research. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23201-23211, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-v1 technical report. arXiv preprint arXiv:2502.13923, 2025. 3, 6, 7, 15, 16" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "type": "text", + "content": "[5] Mu Cai, Reuben Tan, Jianrui Zhang, Bocheng Zou, Kai Zhang, Feng Yao, Fangrui Zhu, Jing Gu, Yiwu Zhong, Yuzhang Shang, Yao Dou, Jaden Park, Jianfeng Gao, Yong Jae Lee, and Jianwei Yang. Temporalbench: Benchmarking fine-grained temporal understanding for multimodal video models. CoRR, abs/2410.10818, 2024. 3, 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 395 + ], + "type": "text", + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 429 + ], + "type": "text", + "content": "[7] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 475 + ], + "type": "text", + "content": "[8] Xiuyuan Chen, Yuan Lin, Yuchen Zhang, and Weiran Huang. Autoeval-video: An automatic benchmark for assessing large vision language models in open-ended video question answering. arXiv preprint arXiv:2311.14906, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "text", + "content": "[9] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and Lidong Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in videolms. CoRR, abs/2406.07476, 2024. 3, 7, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "text", + "content": "[10] Cheng-Han Chiang and Hung-yi Lee. Can large language models be an alternative to human evaluations? arXiv preprint arXiv:2305.01937, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 568, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 295, + 633 + ], + "type": "text", + "content": "[11] Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261, 2025. 7, 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[12] Shangzhe Di and Weidi Xie. Grounded question-answering in long egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12934–12943, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[13] Ning Ding, Shang Qu, Linhai Xie, Yifei Li, Zaoqu Liu, Kaiyan Zhang, Yibai Xiong, Yuxin Zuo, Zhangren Chen, Ermo Hua, et al. Automating exploratory pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "text", + "content": "teomics research via language models. arXiv preprint arXiv:2411.03743, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "text", + "content": "[14] Yue Fan, Xiaojian Ma, Rujie Wu, Yuntao Du, Jiaqi Li, Zhi Gao, and Qing Li. Videoagent: A memory-augmented multimodal agent for video understanding. In European Conference on Computer Vision, pages 75-92. Springer, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[15] Hao Fei, Shengqiong Wu, Wei Ji, Hanwang Zhang, Meishan Zhang, Mong-Li Lee, and Wynne Hsu. Video-of-thought: Step-by-step video reasoning from perception to cognition. arXiv preprint arXiv:2501.03230, 2024. 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 270 + ], + "type": "text", + "content": "[16] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 272, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 336 + ], + "type": "text", + "content": "[17] Difei Gao, Luowei Zhou, Lei Ji, Linchao Zhu, Yi Yang, and Mike Zheng Shou. Mist: Multi-modal iterative spatial-temporal transformer for long-form video question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14773-14783, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 338, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 402 + ], + "type": "text", + "content": "[18] Weibo Gao, Qi Liu, Linan Yue, Fangzhou Yao, Rui Lv, Zheng Zhang, Hao Wang, and Zhenya Huang. Agent4edu: Generating learner response data by generative agents for intelligent education systems. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 23923-23932, 2025. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 447 + ], + "type": "text", + "content": "[19] Ridouane Ghermi, Xi Wang, Vicky Kalogeiton, and Ivan Laptev. Short film dataset (SFD): A benchmark for story-level video understanding. CoRR, abs/2406.10221, 2024. 3, 6, 16" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 502 + ], + "type": "text", + "content": "[20] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Yu Wu, YK Li, et al. Deepseek-coder: When the large language model meets programming-the rise of code intelligence. arXiv preprint arXiv:2401.14196, 2024. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "text", + "content": "[22] Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding. arXiv preprint arXiv:2502.06428, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 691 + ], + "type": "text", + "content": "[24] Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic understanding of image and video advertisements. In CVPR, pages 1100-1110. IEEE Computer Society, 2017. 3, 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 693, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 553, + 713 + ], + "type": "text", + "content": "[25] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23404" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Yunseok Jang, Yale Song, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Tgif-qa: Toward spatio-temporal reasoning in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2758–2766, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[27] Menglin Jia, Zuxuan Wu, Austin Reiter, Claire Cardie, Serge Belongie, and Ser-Nam Lim. Intentonomy: a dataset and study towards human intent understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12986-12996, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[28] Zhiwei Jia, Pradyumna Narayana, Arjun R Akula, Garima Pruthi, Hao Su, Sugato Basu, and Varun Jampani. Kafa: Rethinking image ad understanding with knowledge-augmented feature adaptation of vision-language models. arXiv preprint arXiv:2305.18373, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[29] Pin Jiang and Yahong Han. Reasoning with heterogeneous graph alignment for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11109-11116, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[30] Fangkai Jiao, Geyang Guo, Xingxing Zhang, Nancy F Chen, Shafiq Joty, and Furu Wei. Preference optimization for reasoning with pseudo feedback. arXiv preprint arXiv:2411.16345, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "text", + "content": "[31] Yaman Kumar, Rajat Jha, Arunim Gupta, Milan Aggarwal, Aditya Garg, Tushar Malyan, Ayush Bhardwaj, Rajiv Ratn Shah, Balaji Krishnamurthy, and Changyou Chen. Persuasion strategies in advertisements. In Proceedings of the AAAI conference on artificial intelligence, pages 57-66, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 455 + ], + "type": "text", + "content": "[32] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L. Berg. TVQA: localized, compositional video question answering. In EMNLP, pages 1369-1379. Association for Computational Linguistics, 2018. 3, 16" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 511 + ], + "type": "text", + "content": "[33] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7331-7341, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 555 + ], + "type": "text", + "content": "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 6, 7, 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 611 + ], + "type": "text", + "content": "[35] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Lou, Limin Wang, and Yu Qiao. Mvbench: A comprehensive multimodal video understanding benchmark. In CVPR, pages 22195-22206. IEEE, 2024. 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[37] Xiangpeng Li, Jingkuan Song, Lianli Gao, Xianglong Liu, Wenbing Huang, Xiangnan He, and Chuang Gan. Beyond rnns: Positional self-attention with co-attention for video" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "question answering. In Proceedings of the AAAI conference on artificial intelligence, pages 8658-8665, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[38] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhang, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. In European Conference on Computer Vision, pages 200-216. Springer, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[39] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. CoRR, abs/2311.10122, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 216 + ], + "type": "text", + "content": "[40] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[42] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[43] Xinwei Long, Zhiyuan Ma, Ermo Hua, Kaiyan Zhang, Biqing Qi, and Bowen Zhou. Retrieval-augmented visual question answering via built-in autoregressive search engines. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 24723-24731, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[44] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning. arXiv preprint arXiv:2401.08967, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 384, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 424 + ], + "type": "text", + "content": "[45] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[46] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Videogpt+: Integrating image and video encoders for enhanced video understanding. arxiv, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[47] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. In NeurIPS, 2023. 3, 16" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "text", + "content": "[48] Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13235-13245, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 612 + ], + "type": "text", + "content": "[49] Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[50] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models. CoRR, abs/2311.16103, 2023. 16" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[51] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23405" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[52] Biqing Qi, Kaiyan Zhang, Kai Tian, Haoxiang Li, ZhangRen Chen, Sihang Zeng, Ermo Hua, Hu Jinfang, and Bowen Zhou. Large language models as biomedical hypothesis generators: a comprehensive evaluation. arXiv preprint arXiv:2407.08940, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[53] Abisek Rajakumar Kalarani, Pushpak Bhattacharyya, and Sumit Shekhar. Seeing the unseen: Visual metaphor captioning for videos. arXiv e-prints, pages arXiv-2406, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[54] Ramon Sanabria, Ozan Caglayan, Shruti Palaskar, Desmond Elliott, Loic Barrault, Lucia Specia, and Florian Metze. How2: a large-scale dataset for multimodal language understanding. arXiv preprint arXiv:1811.00347, 2018. 16" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "text", + "content": "[55] Paul Hongsuck Seo, Arsha Nagrani, and Cordelia Schmid. Look before you speak: Visually contextualized utterances. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16877-16887, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 329 + ], + "type": "text", + "content": "[56] Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection. arXiv preprint arXiv:2504.13092, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 365 + ], + "type": "text", + "content": "[57] Yudi Shi, Shangzhe Di, Qirui Chen, and Weidi Xie. Unlocking video-llm via agent-of-thoughts distillation. arXiv preprint arXiv:2412.01694, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[58] Enxin Song, Wenhao Chai, Tian Ye, Jenq-Neng Hwang, Xi Li, and Gaoang Wang. Moviechat+: Question-aware sparse memory for long video question answering. arXiv preprint arXiv:2404.17176, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[59] Makarand Tapaswi, Yukun Zhu, Rainer Stiefelhagen, Antonio Torralba, Raquel Urtasun, and Sanja Fidler. Movieqa: Understanding stories in movies through question-answering. In CVPR, pages 4631–4640. IEEE Computer Society, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[60] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[61] Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video understanding with large language model as agent. In European Conference on Computer Vision, pages 58-76. Springer, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[62] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[63] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool. arXiv preprint arXiv:2407.05355, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[64] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "text", + "content": "[65] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 9777-9786. Computer Vision Foundation / IEEE, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "text", + "content": "[66] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13204-13214, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 554, + 250 + ], + "type": "text", + "content": "[67] Binzhu Xie, Sicheng Zhang, Zitang Zhou, Bo Li, Yuanhan Zhang, Jack Hessel, Jingkang Yang, and Ziwei Liu. Funqa: Towards surprising video comprehension. In European Conference on Computer Vision, pages 39-57. Springer, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 306 + ], + "type": "text", + "content": "[68] Haomiao Xiong, Zongxin Yang, Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Jiawen Zhu, and Huchuan Lu. Streaming video understanding and multi-round interaction with memory-enhanced knowledge. arXiv preprint arXiv:2501.13468, 2025.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 364 + ], + "type": "text", + "content": "[69] Dejing Xu, Zhou Zhao, Jun Xiao, Fei Wu, Hanwang Zhang, Xiangnan He, and Yueting Zhuang. Video question answering via gradually refined attention over appearance and motion. In Proceedings of the 25th ACM international conference on Multimedia, pages 1645–1653, 2017. 16" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 408 + ], + "type": "text", + "content": "[70] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 554, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 465 + ], + "type": "text", + "content": "[71] Yanzhi Xu, Yueying Hua, Shichen Li, and Zhongqing Wang. Exploring chain-of-thought for multi-modal metaphor detection. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 91-101, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 554, + 510 + ], + "type": "text", + "content": "[72] Dingyi Yang, Chunru Zhan, Ziheng Wang, Biao Wang, Tiezheng Ge, Bo Zheng, and Qin Jin. Synchronized video storytelling: Generating video narrations with structured storyline. arXiv preprint arXiv:2405.14040, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 555 + ], + "type": "text", + "content": "[73] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). arXiv preprint arXiv:2401.08392, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 610 + ], + "type": "text", + "content": "[74] Ruilin Yao, Bo Zhang, Jirui Huang, Xinwei Long, Yifang Zhang, Tianyu Zou, Yufei Wu, Shichao Su, Yifan Xu, Wenxi Zeng, et al. Lens: Multi-level evaluation of multimodal reasoning with large language models. arXiv preprint arXiv:2505.15616, 2025.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "text", + "content": "[75] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 7, 15" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[76] Keren Ye, Narges Honarvar Nazari, James Hahn, Zaeem Hussain, Mingda Zhang, and Adriana Kovashka. Interpreting the rhetoric of visual advertisements. IEEE transactions on pattern analysis and machine intelligence, 43(4):1308-1323, 2019. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23406" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[77] Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. Activitynet-qa: A dataset for understanding complex web videos via question answering. In AAAI, pages 9127–9134. AAAI Press, 2019. 3, 7, 15, 16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[78] Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Dongsheng Li, and Deqing Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms. arXiv preprint arXiv:2406.14228, 2024. 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "text", + "content": "[79] Majid Zahmati, Seyed Morteza Azimzadeh, Mohammad Saber Sotoodeh, and Amid Asgari. An eye-tracking study on how the popularity and gender of the endorsers affected the audience's attention on the advertisement. Electronic Commerce Research, 23(3):1665-1676, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[80] Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, et al. Internlm-xcomposer2. 5-reward: A simple yet effective multi-modal reward model. arXiv preprint arXiv:2501.12368, 2025. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 324 + ], + "type": "text", + "content": "[81] Shuang Zeng, Xinyuan Chang, Mengwei Xie, Xinran Liu, Yifan Bai, Zheng Pan, Mu Xu, and Xing Wei. Futuresight-drive: Thinking visually with spatio-temporal cot for autonomous driving. arXiv preprint arXiv:2505.17685, 2025.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 402 + ], + "type": "text", + "content": "[82] Dongyu Zhang, Minghao Zhang, Heting Zhang, Liang Yang, and Hongfei Lin. Multimet: A multimodal dataset for metaphor understanding. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 3214-3225, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[83] Dongyu Zhang, Jingwei Yu, Senyuan Jin, Liang Yang, and Hongfei Lin. Multicmet: A novel chinese benchmark for understanding multimodal metaphor. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 6141-6154, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 512 + ], + "type": "text", + "content": "[84] Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b. arXiv preprint arXiv:2406.07394, 2024. 1, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 578 + ], + "type": "text", + "content": "[85] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 3, 7, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "type": "text", + "content": "[86] Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llavanext: A strong zero-shot video understanding model, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 657 + ], + "type": "text", + "content": "[87] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[88] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data, 2024. 15" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[89] Kang Zhao, Xinyu Zhao, Zhipeng Jin, Yi Yang, Wen Tao, Cong Han, Shuanglong Li, and Lin Liu. Enhancing baidu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 127 + ], + "type": "text", + "content": "multimodal advertisement with chinese text-to-image generation via bilingual alignment and caption synthesis. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2855-2859, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[90] Zhou Zhao, Zhu Zhang, Shuwen Xiao, Zhou Yu, Jun Yu, Deng Cai, Fei Wu, and Yueting Zhuang. Open-ended long-form video question answering via adaptive hierarchical reinforced networks. In IJCAI, page 8, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[91] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "text", + "content": "[92] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 338 + ], + "type": "text", + "content": "[93] Yuxin Zuo, Shang Qu, Yifei Li, Zhangren Chen, Xuekai Zhu, Ermo Hua, Kaiyan Zhang, Ning Ding, and Bowen Zhou. Medxpertqa: Benchmarking expert-level medical reasoning and understanding. arXiv preprint arXiv:2501.18362, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 553, + 384 + ], + "type": "text", + "content": "[94] Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al. Trl: Test-time reinforcement learning. arXiv preprint arXiv:2504.16084, 2025.3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23407" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_content_list.json b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56012d841cc003f727820555263dc44713f845fb --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_content_list.json @@ -0,0 +1,1826 @@ +[ + { + "type": "text", + "text": "AdvDreamer Unveils:", + "text_level": 1, + "bbox": [ + 387, + 131, + 609, + 148 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Are Vision-Language Models Truly Ready for Real-World 3D Variations?", + "text_level": 1, + "bbox": [ + 127, + 152, + 869, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shouwei Ruan $^{1\\dagger}$ , Hanqing Liu $^{1\\dagger}$ , Yao Huang $^{1\\dagger}$ , Xiaqi Wang $^{1}$ , Caixin Kang $^{1}$ , Hang Su $^{2}$ , Yinpeng Dong $^{3}$ , Xingxing Wei $^{1*}$", + "bbox": [ + 192, + 202, + 803, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Artificial Intelligence,", + "bbox": [ + 362, + 238, + 635, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University", + "bbox": [ + 161, + 256, + 834, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Dept. of Comp. Sci. and Tech., Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab, BNRist Center, Tsinghua University", + "bbox": [ + 166, + 273, + 828, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 College of AI, Tsinghua University", + "bbox": [ + 352, + 309, + 643, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shouweiruan, hqliu, y_huang, xxwei}@buaa.edu.cn, suhangss@tsinghua.edu.cn, dongyinpeng@mail.tsinghua.edu.cn", + "bbox": [ + 127, + 330, + 866, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(A) Adversarial 3D Variations Samples form Single-view Natural Images —— (C) Background Variations for Adv-3DT samples", + "bbox": [ + 153, + 349, + 841, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d4983076e76f69f2c01aefff157f73c37a9ec199c7865d471d4722f1c55056f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 158, + 368, + 274, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91513af6c0274a700ac71aea6cfc52a39e93dfec9cb07450b16365e25fe6be56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 368, + 467, + 440 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e744f7a72a21a6fb0987cd19d08e4cb0f2bafc9faf47370d3e067e3742e20343.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 368, + 633, + 439 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(B) Physical Reproduction of Adversarial 3D Variations Samples", + "bbox": [ + 153, + 443, + 388, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c916d5b785362732141f6114aeb74e746360d29d2dcb1d6b5768dd8871de4679.jpg", + "image_caption": [ + "Figure 1. The 3D Variation Vulnerabilities in VLMs. (A) The proposed AdvDreamer allows the capture of adversarial 3D transformation (Adv-3DT) samples from single natural images, threatening various VLMs. (B) These Adv-3DT samples can be successfully reproduced in the physical world and (C) retain their aggressiveness under background variations. (D) We also explain the GPT-Score metric used in the experiments. In each VLM response, we highlight the correct, significant error, and less precise parts using different colors." + ], + "image_footnote": [], + "bbox": [ + 133, + 453, + 264, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e1357629bf1be364d02f6f37868cce6b703a631fcb3d42a2a4a261caf59cc593.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 453, + 383, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7bd5121d7cfb2b332e5d7ea86ac4f2282435d98bb4d151e2bbf754144ee322b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 455, + 506, + 558 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17f84b046911aa55a6bb4c55168ef72fc3f589bb2e69bf4776c697faedc92b96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 455, + 635, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac1996e7d1a94690ce288b9d8e76acf48d3d5edae43fffdfe99953a22c0c4068.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 352, + 864, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cef88438797ec505262a305d763b93ec92d5512fd0eeec51beeaa8a6059bb26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 487, + 864, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 648, + 326, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Language Models (VLMs) have exhibited remarkable generalization capabilities, yet their robustness in dynamic real-world scenarios remains largely unexplored. To systematically evaluate VLMs' robustness to real-world 3D variations, we propose AdvDreamer, the first framework capable of generating physically reproducible Adversarial 3D Transformation (Adv-3DT) samples from single-view observations. In AdvDreamer, we integrate three key innovations: Firstly, to characterize real-world 3D variations with limited prior knowledge precisely, we design a zero-shot Monocular Pose Manipulation pipeline built upon generative 3D priors. Secondly, to ensure the visual qual", + "bbox": [ + 89, + 681, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ity of worst-case Adv-3DT samples, we propose a Naturalness Reward Model that provides continuous naturalness regularization during adversarial optimization, effectively preventing convergence to hallucinated or unnatural elements. Thirdly, to enable systematic evaluation across diverse VLM architectures and visual-language tasks, we introduce the Inverse Semantic Probability loss as the adversarial optimization objective, which solely operates in the fundamental visual-textual alignment space. Based on the captured Adv-3DT samples with high aggressiveness and transferability, we establish MM3DTBench, the first VQA benchmark dataset tailored to evaluate VLM robustness under challenging 3D variations. Extensive evaluations of representative VLMs with varying architectures reveal that real-world 3D variations can pose severe threats to model performance across various tasks.", + "bbox": [ + 511, + 650, + 908, + 892 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal contribution.", + "bbox": [ + 114, + 875, + 227, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 114, + 888, + 245, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65f7a3ecc01c3954051e3bff120793d5a59b93dd7a949d9ec21e82e1c2cd882f.jpg", + "image_caption": [ + "Figure 2. Comparison of paradigms to characterize Adversarial 3D Transformation (Adv-3DT) samples." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 285, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a35f2f9f3b56dc9fc20c47dcb736c0b822dcff76a70a07edbfccaf520881b5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 90, + 477, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e7f87b135794b238f091b2754fb9dbcb9580202b93b25ef6009a58b04e479628.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 90, + 666, + 191 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/013f93479f7280a02673e7bc8ccc94ae40796dd9cb8d75e30d2599568c88ff33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 90, + 861, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 223, + 222, + 238 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Vision-Language Models (VLMs) [7, 13, 35, 36] have demonstrated remarkable capabilities in bridging visual perception and natural language understanding. Through large-scale image-text pre-training followed by fine-grained instruction tuning [31, 36], these models effectively tackle a wide range of visual-centric tasks, including instruction-based image recognition [47, 55, 65], visual question answering [3, 31, 35, 36], and visual reasoning [9, 10]. Due to their strong generalizability and impressive zero-shot capability, VLMs are increasingly deployed in various safety-critical applications, particularly in autonomous driving [56, 63] and robotic systems [25, 64].", + "bbox": [ + 89, + 250, + 482, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite extensive studies validating VLMs' resilience against various distribution shifts (e.g., style variations [2, 39, 40, 68], image corruptions [2], and adversarial perturbations [16, 38, 67, 70]), they primarily focus on 2D perturbations in the digital domain, overlooking a critical challenge in real-world deployment: the 3D variations. As VLMs are increasingly integrated into dynamic scenarios, their ability to handle 3D variation becomes crucial, raising a fundamental question: Have current VLMs attained sufficient robustness to the distribution shifts arising from ubiquitous 3D variations in the real world?", + "bbox": [ + 89, + 431, + 482, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To thoroughly explore this problem and offer comprehensive answers, we identify three fundamental challenges:", + "bbox": [ + 89, + 598, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) How to accurately characterize real-world 3D variations with limited prior knowledge? Existing methods exploring the 3D variation robustness heavily rely on explicit 3D structures [4, 21] or representations derived from dense multi-view observations [4, 21, 42]. However, in most practical scenarios, the available priors are severely limited (e.g., single-view observation only). This constraint necessitates developing a method that minimizes dependency on extensive scene priors. To meet this, we introduce a zero-shot Monocular Pose-Manipulation (MPM) pipeline based on generative representations. Specifically, MPM leverages the rich 3D priors embedded in pre-trained Large-Reconstructive Models (LRM) [23, 24] and Image Re-Composition Models [12, 66], enabling fine-grained 3D adjustments with only single-view images and specified 3D variation parameters. While this generative-based paradigm establishes a foundation for capturing worst-case 3D variations through optimizing 3D variation parameters, it still", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "exhibits inherent limitations that may introduce undesirable image corruptions, thus presenting our second challenge:", + "bbox": [ + 511, + 224, + 903, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) How to capture the worst-case adversarial 3D transformation (Adv-3DT) samples while ensuring that the performance degradation is inherently caused by the 3D variation? Prior works utilizing synthetic 3D assets [4, 21, 42] frequently produce samples with noticeable texture discrepancies compared to real-world images. These limitations in samples' visual fidelity and contextual coherence highlight a critical concern: observed performance degradation under Adv-3DT samples may stem from undesirable corruption rather than the 3D variation itself. To address this, we incorporate naturalness reward constraints within the adversarial optimization process. Specifically, we treat rigid 3D variations as an adversarial attack and propose the Naturalness Reward Model (NRM) that preserves image quality throughout optimization. NRM leverages visual context features provided by the DINOv2 [46] backbone to predict a naturalness reward that jointly reflects the visual fidelity and physical plausibility of the samples, effectively regularizing the optimization trajectory and preventing convergence toward unnatural or hallucinated elements.", + "bbox": [ + 509, + 258, + 906, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(3) How to ensure effective generalization of Adv-3DT samples across diverse VLM tasks and architectures? Existing studies [4, 18, 50] typically rely on task-specific objectives with end-to-end optimization, substantially limiting their applicability across diverse VLM architectures and a wide range of visual language tasks. To transcend this limitation, we formulate the adversarial objective in a task-and architecture-agnostic manner. Specifically, we introduce the Inverse Semantic Probability (ISP) loss, which operates solely through the foundational component of the VLM—the visual encoder, aiming to minimize the image-text matching probability between Adv-3DT samples and their corresponding ground-truth textual descriptions. Extensive experiments demonstrate that Adv-3DT samples optimized under the ISP loss exhibit remarkable adversarial transferability across various downstream tasks and models. Building on this, we further introduce the MM3DTBench, the first VQA benchmark dataset specifically designed to evaluate the 3D variation robustness of VLMs in real-world scenarios, comprising highly challenging Adv-3DT samples with carefully crafted candidate answers.", + "bbox": [ + 511, + 564, + 908, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating the aforementioned innovations, this paper", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7895", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proposes AdvDreamer, a novel framework that enables the characterization of Adv-3DT samples from monocular observations, with broad adaptability across various visual-language tasks and VLM architectures. Our extensive digital and physical experiments reveal significant vulnerabilities in contemporary VLMs when confronted with challenging 3D variations. In visual recognition tasks, foundational models like OpenCLIP [26] and BLIP-2 [31] suffer a performance degradation of $65\\% \\sim 80\\%$ on Adv-3DT samples. More concerning still, in open-ended tasks including image captioning and visual question answering, even the most advanced commercial models like GPT-4o [45] exhibit performance declines approaching $50\\%$ . Our findings highlight the urgent need to enhance VLM robustness against 3D variations and set new requirements for deploying VLMs in complex, dynamic real-world applications.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 349, + 233, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 375, + 316, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classic Vision-Language Models (VLMs) [30, 32, 41, 47] achieve cross-modal alignment through diverse pre-training objectives on large-scale image-text pairs [33, 53]. Recent advances in Large Language Models (LLMs) [59, 60] have inspired new VLM architectures that integrate LLM capabilities via MLP projection layers [35, 36] or specialized modules like Q-Former [31]. Through techniques like instruction tuning [36], current VLMs [13, 35, 36] demonstrate enhanced performance across vision-centric tasks, including visual grounding, question answering, and image captioning, etc. These models also enable intelligent systems for task planning and decision-making [25, 64].", + "bbox": [ + 89, + 396, + 483, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite their impressive capabilities [37, 47, 67], concerns remain about the robustness and generalization of VLMs, especially in safety-critical environments. Recent studies [52, 58] highlight that VLMs still exhibit deficiencies in processing visual details, such as orientation, quantity, and colour, which may be attributed to their visual representations. Moreover, VLMs remain susceptible to $L_{p}$ -norm adversarial perturbations [16, 43], with some efforts aiming to enhance their robustness through adversarial finetuning techniques [43]. In contrast to previous works, our study specifically investigates VLM robustness against 3D variations, addressing a crucial gap in understanding their reliability for dynamic real-world applications.", + "bbox": [ + 89, + 579, + 483, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Evaluation for 3D Variation Robustness", + "text_level": 1, + "bbox": [ + 89, + 787, + 431, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness to 3D variations remains a longstanding challenge in computer vision. Evaluating this property often involves capturing worst-case 3D transformations, i.e. the adversarial poses or viewpoints, to create challenging samples for assessment. Prior research can be categorized into three paradigms (see Fig.2): (A) Human-guided methods, like", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ObjectNet [8] and OOD-CV [69], manually collect samples with diverse poses and viewpoints, but they lack systematic coverage of worst-case examples and are costly. (B) Simulator-based [4, 21, 42] methods optimize worst-case 3D parameters via differentiable rendering of synthetic objects but require manually designed 3D assets and produce samples with limited realism. (C) Multi-view-based methods, like ViewFool [18] and GMVFool [50, 51], use NeRF-based [44] representations to optimize adversarial viewpoints, but they require dense observations and struggle with background integration. In contrast, the proposed AdvDreamer employs a (D) Generative-based approach, utilizing robust 3D priors from generative models. Given a single natural image, AdvDreamer generate realistic Adv-3DT samples in a zero-shot manner, providing a more efficient solution for evaluating 3D variation robustness.", + "bbox": [ + 511, + 90, + 906, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 344, + 648, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overview of AdvDreamer is shown in Fig. 3. It consists $\\bullet$ Zero-shot Monocular Pose Manipulation pipeline, leveraging generative 3D representations to obtain specified 3D variation samples from single-view images; $\\bullet$ Naturalness Reward Model, providing naturalness regularization during adversarial optimization; and $\\bullet$ Inverse Semantic Probability loss, ensuring the Adv-3DT samples effectively threaten VLMs and generalize across different architectures and tasks. The following sections will firstly formalize the optimization problem of AdvDreamer (Sec.3.1), then sequentially describe the key components (Sec.3.2), and finally outline the full optimization algorithm (Sec.3.3).", + "bbox": [ + 511, + 369, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 559, + 718, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parametrization of Real-World 3D Variations. In this paper, we primarily focus on rigid 3D transformations of foreground objects, i.e., the rotation, translation, and scaling, since they reflect the most typical 3D changes observed in real-world environments. Formally, we define a 6-dimensional vector $\\Theta = \\{\\alpha, \\beta, \\gamma, \\Delta_x, \\Delta_y, s\\}$ to uniquely parameterize any arbitrary transformation, where $\\{\\alpha, \\beta, \\gamma\\}$ denote the Tait-Bryan angles (yaw, pitch, roll) in the $Z-Y-X$ sequence, $\\{\\Delta_x, \\Delta_y\\}$ represent the translation along the $x$ and $y$ axes in the $xy$ -plane, and $s$ is the uniform scaling factor. To prevent the loss of critical visual cues and ensure human recognizability, we follow previous work [18] to constrain $\\Theta$ within a bounded range $[\\Theta \\min, \\Theta \\max]$ .", + "bbox": [ + 511, + 580, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimization Problem. The objective of AdvDreamer is to find an optimal distribution $p^*(\\Theta)$ , ensuring that Adv-3DT sampled from this distribution are both aggressive against VLMs and maintaining visual quality, which can be abstractly formulated as the following optimization problem:", + "bbox": [ + 511, + 777, + 908, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np ^ {*} (\\Theta) = \\arg \\max _ {p (\\Theta)} \\mathbb {E} _ {\\Theta \\sim p (\\Theta)} \\left[ \\mathcal {L} _ {I S P} \\left(X ^ {\\prime}\\right) + \\mathcal {L} _ {N a t} \\left(X ^ {\\prime}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 861, + 903, + 887 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e X ^ {\\prime} = \\mathcal {T} (\\Theta , X),\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 887, + 777, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7896", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/225896af9b58da7824ec8152570a7f4886dc366d21724adb3066679c1bfaca94.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed AdvDreamer framework. To capture worst-case 3D variations in the real world, we treat rigid 3D transformations as adversarial attacks, optimizing the distribution of transformation parameters through a query-based approach. In AdvDreamer, we introduce the Monocular Pose Manipulation pipeline to perform specified 3D transformations on single-view images and guide the optimization process using the proposed Naturalness Reward Model and Inverse Semantic Probability loss." + ], + "image_footnote": [], + "bbox": [ + 142, + 90, + 851, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{T}(\\Theta, X)$ is a transformation function to produce a novel image with the specified transformation $\\Theta$ applied to the original image $X \\in \\mathbb{R}^3$ . To implement $\\mathcal{T}(\\cdot)$ , we design the MPM pipeline, which will be detailed in Sec.3.2.1. The $\\mathcal{L}_{Nat}$ and $\\mathcal{L}_{ISP}$ measure the naturalness and adversarial efficacy of the transformed samples, respectively. These will be further elaborated in Sec.3.2.2 and Sec.3.2.3.", + "bbox": [ + 89, + 366, + 483, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimizing over the distribution $p(\\Theta)$ rather than a deterministic $\\Theta$ offers two key advantages. 1) As highlighted in [18, 50], learning the underlying distribution enables comprehensive exploration of the 3D variation space, facilitating the identification of the model's vulnerable regions. 2) Given the stochastic nature of the generation, optimizing over a continuous distribution can mitigate the impact of appearance inconsistency, as appearance variations induced by the generation are unlikely to consistently deceive the model without optimizing the adversarial 3D variations [6].", + "bbox": [ + 89, + 472, + 483, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. AdvDreamer Framework", + "text_level": 1, + "bbox": [ + 89, + 633, + 323, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Zero-shot Monocular Pose Manipulation", + "text_level": 1, + "bbox": [ + 89, + 654, + 419, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To implement the transformation function $\\mathcal{T}$ , previous approaches typically rely on explicit 3D structures [4, 21, 42] or neural radiance representations [18, 50]. However, these approaches cannot accommodate real-world scenarios with limited prior knowledge. To address this, we design a simple yet effective zero-shot monocular pose manipulation (MPM) pipeline that consists of the following process:", + "bbox": [ + 89, + 674, + 482, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1 Foreground-Background Pairs Preparation. MPM first decompose the input image $X$ into foreground-background pairs $\\{X_{f},X_{b}\\}$ . Specifically, we leverage Grounded-SAM [48] to semantically segment the major instance as foreground $X_{f}$ , followed by diffusion-based inpainting [15, 49] to obtain a complete background $X_{b}$ . This process is denoted as $\\{X_f,X_b\\} = \\mathcal{F}(X)$ . Additionally, we support directly generating synthetic $\\{X_f,X_b\\}$ pairs", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through Stable-Diffusion [49] to enhance sample diversity.", + "bbox": [ + 511, + 366, + 901, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Theta$ Object-level Pose Manipulation. For each optimization iteration, we transform the foreground $X_{f}$ into a batch of $X_{f}^{\\prime}$ under $\\Theta$ sampled from current $p(\\Theta)$ . In this stage, MPM leverages the robust priors provided by a pre-trained Large Reconstruction Model (e.g., TripoSR [57]) to construct single-view-based 3D representations and apply sampled transformations. This process is formulated as $X_{f}^{\\prime} = \\mathcal{R}_{\\mathbf{w}_{0}}(X_{f},\\Theta)$ , where $\\Theta \\sim p(\\Theta)$ and $\\mathcal{R}_{\\mathbf{w}}(\\cdot)$ denotes neural rendering function with weights $\\mathbf{w}_0$ .", + "bbox": [ + 511, + 382, + 903, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\Theta$ Re-Composition. We then compose the transformed foreground $X_{f}^{\\prime}$ with background $X_{b}$ through a pre-trained diffusion-based composition model [12, 66], to handles shadows and occlusions. This process is defined as $X^{\\prime} = \\mathcal{C}_{\\mathbf{w}_1}(X_f',X_b)$ , where $\\mathbf{w}_1$ denotes diffusion model's weights.", + "bbox": [ + 511, + 518, + 903, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining the above stages, the transformation process by MPM can be formally represented as follows:", + "bbox": [ + 511, + 594, + 903, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r l} X ^ {\\prime} = \\mathcal {T} (\\boldsymbol {\\Theta}, X) & = \\mathcal {C} _ {\\mathbf {w} _ {1}} \\left(\\mathcal {R} _ {\\mathbf {w} _ {0}} \\left(X _ {f}, \\boldsymbol {\\Theta}\\right), X _ {b}\\right), \\\\ \\text {w h e r e} \\{X _ {f}, X _ {b} \\} & = \\mathcal {F} (X). \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 636, + 903, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Naturalness Reward Model", + "text_level": 1, + "bbox": [ + 511, + 684, + 751, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the MPM pipeline utilizing generative priors to achieve realistic 3D variation samples, we observe that directly applying adversarial optimization frequently leads to unnatural artifacts, such as shape distortions and physically implausible cases in Fig. 5. These issues stem from two factors: generalization limitations of existing generative models and the inherent high semantic bias in low-quality samples. To address this challenge, we propose a novel Naturalness Reward Model (NRM) that continuously regularizes $p(\\Theta)$ during optimization based on sample naturalness, preventing convergence to low-quality pseudo-optimal regions.", + "bbox": [ + 511, + 704, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig.4, the NRM employs DINOv2[46] as its backbone, chosen for several reasons. Firstly, DINO", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fbf19ad0829fae426c631676671a2ca58d284ef7c89a9d396eecf8e158cc5fbe.jpg", + "image_caption": [ + "Figure 4. Architecture of NRM and computational cost and accuracy of naturalness assessment across Human, GPT, and NRM." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 470, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "captures finer visual contextual features compared to traditional pre-trained vision encoders [58], aiding better convergence during training. Secondly, unlike vision-language aligned encoders such as CLIP, DINO's visual features are language-agnostic, making them more faithfully reflect image quality rather than semantic consistency. We use DINOv2 to extract image tokens from the input Adv-3DT samples, followed by dual-stream prediction heads to estimate visual fidelity and physical plausibility scores:", + "bbox": [ + 89, + 210, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) = f _ {R} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right) = f _ {P} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 359, + 483, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{R}(\\cdot)$ and $f_{P}(\\cdot)$ are two-layer MLPs. To train the NRM, we construct a large-scale naturalness preference dataset with fine-grained scoring. We generate numerous samples from the MPM pipeline on ImageNet and synthetic images, scoring each on a 5-point scale for visual fidelity and physical plausibility. These scores are initially annotated automatically by GPT-4o with Chain-of-Thought prompts and verified by volunteers for correction. This process resulted in 120K high-quality training samples.", + "bbox": [ + 89, + 391, + 483, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The architecture and training parameters of NRM are detailed in Appendix B. Our experimental results demonstrate that the naturalness feedback provided by NRM highly aligns with human evaluation. In terms of prediction efficiency and stability, it outperforms methods that directly employ GPT-4o for naturalness assessment (see Fig. 4 and Appendix C.2). Based on the NRM's predictions, we define the naturalness regularization loss for the optimization as:", + "bbox": [ + 89, + 527, + 483, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) = - \\frac {1}{2} \\left(\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) + \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right)\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 659, + 482, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Inverse Semantic Probability Loss", + "text_level": 1, + "bbox": [ + 89, + 699, + 375, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous studies typically employ task-specific optimization objectives with end-to-end manner to obtain worst-case Adv-3DT samples. However, directly applying this approach to VLMs with substantial parameters and diverse architectures spanning various tasks is challenging. To address this, we design the Inverse Semantic Probability (ISP) loss, which minimizes the probability that VLMs assign to the correct semantic attributes of Adv-3DT samples.", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This design have two key considerations. 1) It is based solely on the visual and text encoder $\\mathcal{E}_v$ , $\\mathcal{E}_t$ , which are fundamental components across modern VLMs, ensuring architecture-agnostic applicability. 2) By operating in the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "image-text alignment space rather than model-specific layers or task-specific heads, it enables task generalization.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a generated sample $X'$ and a semantic label set $\\mathcal{Y} = \\{y_i\\}_{i=1}^N$ containing the ground-truth label $y_t$ , we first compute a semantic similarity vector $\\mathbf{S}$ :", + "bbox": [ + 511, + 121, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} = \\left\\{s _ {i} \\right\\} _ {i = 1} ^ {N}, \\text {w h e r e} s _ {i} = \\cos \\left[ \\mathcal {E} _ {v} \\left(X ^ {\\prime}\\right), \\mathcal {E} _ {t} \\left(\\mathcal {P} \\left(y _ {i}\\right)\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 172, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot)$ computes cosine similarity and $\\mathcal{P}(\\cdot)$ follows the template \"a photo of a $\\{y_i\\}$ \" in [47]. The vector $\\mathbf{S}$ captures the representation similarities between $X'$ and each semantic label. We then derive the conditional semantic probability through softmax transformation:", + "bbox": [ + 511, + 198, + 905, + 275 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(y _ {i} \\mid X ^ {\\prime}\\right) = \\exp \\left(s _ {i}\\right) / \\sum_ {j = 1} ^ {N} \\exp \\left(s _ {j}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 280, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the $\\mathcal{L}_{\\mathrm{ISP}}$ is defined as the negative log-likelihood of the ground-truth's semantic probability:", + "bbox": [ + 511, + 306, + 905, + 338 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) = - \\log p \\left(y _ {t} \\mid X ^ {\\prime}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 344, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, $\\mathcal{L}_{\\mathrm{ISP}}$ quantifies the likelihood of $X^{\\prime}$ being misclassified by comparing its similarity to the ground-truth and irrelevant categories. Maximizing it directs the $p(\\Theta)$ towards regions that effectively deceive the model.", + "bbox": [ + 511, + 369, + 905, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Query-Based Adversarial Optimization", + "text_level": 1, + "bbox": [ + 511, + 436, + 852, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve Problem (1), we parameterize the adversarial distribution $p(\\Theta)$ as a Multivariate Gaussian form:", + "bbox": [ + 511, + 459, + 903, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Theta} = \\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, \\text {w h e r e} \\mathbf {z} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 498, + 903, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{A} = (\\Theta_{\\mathrm{min}} - \\Theta_{\\mathrm{max}}) / 2$ , $\\mathbf{A} = (\\Theta_{\\mathrm{min}} + \\Theta_{\\mathrm{max}}) / 2$ , $\\mathbf{z}$ follows a Gaussian Distribution with mean $\\pmb{\\mu} \\in \\mathbb{R}^6$ and covariance $\\pmb{\\Sigma} \\in \\mathbb{R}^{6 \\times 6}$ . Thus, we rewrite (1) in a specific form:", + "bbox": [ + 511, + 522, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}} {\\arg \\max } \\mathbb {E} _ {\\boldsymbol {\\mathbf {z}} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma})} \\left[ \\mathcal {L} _ {\\text {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) + \\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 575, + 903, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e X ^ {\\prime} = \\mathcal {T} (\\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, X).\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 603, + 834, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Solving this problem typically involves computing the gradient of the loss w.r.t. $\\pmb{\\mu}$ , $\\pmb{\\Sigma}$ and updating them via gradient ascent. However, the forward steps involve multiple components, e.g., the denoise process, introducing uncertainty in the gradient propagation path, making gradient-based optimization challenging [54]. Therefore, we adopt the Covariance Matrix Adaptation Evolution Strategy [19, 22], an efficient query-based black-box optimizer. At each iteration $t$ , we: 1) Sample $K$ candidates $\\{\\mathbf{z}_i^{t + 1}\\}_{i = 1}^K$ from $\\mathcal{N}(\\boldsymbol {\\mu}^t,\\boldsymbol {\\Sigma}^t)$ . 2) Generate corresponding samples $\\{(X')_{i}^{t + 1}\\}_{i = 1}^{K}$ . 3) Update distribution parameters $\\pmb{\\mu}^t,\\pmb{\\Sigma}^t$ as follows:", + "bbox": [ + 511, + 625, + 905, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {t + 1} = \\sum_ {i = 1} ^ {k} w _ {i} \\cdot \\mathbf {z} _ {(i: k)} ^ {t + 1}, \\quad \\text {w h e r e} \\sum_ {i = 1} ^ {k} w _ {i} = 1, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 792, + 903, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {t + 1} = \\left(1 - \\eta_ {1} - \\eta_ {\\mu}\\right) \\cdot \\boldsymbol {\\Sigma} ^ {t} + \\eta_ {1} \\cdot p _ {\\Sigma} ^ {t + 1} \\left(p _ {\\Sigma} ^ {t + 1}\\right) ^ {T} +\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 839, + 859, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {\\mu} \\cdot \\sum_ {i = 1} ^ {h} w _ {i} \\cdot \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) ^ {T}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 861, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2ce3e60904567f2362d7d27536c611b2c097c3d8ec000624b2c43656df8f47eb.jpg", + "table_caption": [ + "Table 1. Zero-shot Classification Accuracy $(\\%)$ and degradation $(\\downarrow \\%)$ Under Adv-3DT Samples Generated from ImageNet [17]." + ], + "table_footnote": [], + "table_body": "
Target Models#ParamsMethods
CleanRandomp*(Θ)Θ*
OpenCLIP ViT-B/16 [26]149M98.062.6 (↓35)54.0 (↓44)18.0 (↓80)
OpenCLIP ViT-L/14 [26]428M94.461.7 (↓33)50.9 (↓44)15.3 (↓79)
OpenCLIP ViT-G/14 [26]2.5B96.463.5 (↓33)53.5 (↓43)18.7 (↓78)
BLIP ViT-B/16 [30]583M83.056.0 (↓27)51.3 (↓32)17.3 (↓66)
BLIP-2 ViT-L/14 [31]3.4B84.057.0 (↓27)52.2 (↓32)18.7 (↓65)
BLIP-2 ViT-G/14 [31]4.1B81.055.7 (↓25)49.1 (↓32)15.7 (↓65)
", + "bbox": [ + 93, + 121, + 485, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "here, $\\mathbf{z}_{(i:k)}$ denotes samples ranked by losses: we first select top-10 samples based on $\\mathcal{L}_{\\mathrm{ISP}}$ , then choose the top-k $(k = 5)$ with highest $\\mathcal{L}_{\\mathrm{Nat}}$ for updates. $w_{i} = 1 / k$ are importance weights, $\\eta_{1}$ and $\\eta_{\\mu}$ are learning rates, $p_{\\Sigma}$ tracks the evolution path, and $\\sigma^t$ is the step size updated via Cumulative Step-size Adaptation (CSA) [22]. Due to space constraints, complete derivations of Eq. (10), (11) and hyperparameter settings are provided in the Appendix.A.1. We also provide the pseudocode for optimization process in Appendix A.2.", + "bbox": [ + 89, + 224, + 483, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 369, + 225, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 395, + 281, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tasks & Datasets. We evaluate VLMs' robustness to 3D variations using AdvDreamer on three representative vision-language tasks. (1) For zero-shot image classification, we utilize the ImageNet test set [17]. (2) For image captioning, we employ the test splits from COCO Caption [11] and NoCaps [1] datasets. (3) For Visual Question Answering (VQA), we use the VQAv2 [20] test set. For each dataset, we select 300 clean samples spanning 30 categories with clear and well-defined object instances. During the optimization process, we employ ImageNet-1K categories as the semantic label set $\\mathcal{Y}$ for the classification task and use COCO-80 categories for the captioning and VQA tasks.", + "bbox": [ + 89, + 417, + 483, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Victim Models. For zero-shot classification, we target vision-language foundation models: OpenCLIP [14, 26], BLIPs [30, 31], and additional models for transfer attacks. For other tasks, we select mainstream open-source VLMs (LLaVa [34, 36], MiniGPT-4 [72], etc.) as well as representative commercial VLMs (GPT-4o and GPT-4o-mini).", + "bbox": [ + 89, + 598, + 483, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Methods & Baselines. We use two variants with AdvDreamer: (1) $p^*(\\Theta)$ : represents the average results over 10 Adv-3DT samples randomly sampled from the optimal adversarial distribution, while (2) $\\Theta^*$ : indicates the evaluation results using the Adv-3DT sample corresponding to the distribution center. Additionally, we establish two potential baselines following [18, 50]: the original natural images without transformation (Clean) and the average results over 10 images with random 3D transformations (Random). Metrics. For zero-shot classification, we follow the CLIP-Benchmark [27] protocol, reporting Top-1 accuracy and its degradation relative to clean samples. For captioning, in addition to conventional metrics (CIDEr, BLEU), we adopt the LLM-as-a-judge approach [36, 67, 71] to address the open", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8fac97422681d8f344213e62437351241009df581fc688a61d5822a6cbf831ca.jpg", + "table_caption": [ + "Table 2. Cross-model transferability of Adv-3DT samples, utilizing OpenCLIP and BLIP as surrogate models." + ], + "table_footnote": [], + "table_body": "
Target ModelCleanOpenCLIP [26]BLIP [30]
p*(Θ)Θ*p*(Θ)Θ*
ALBEF [29]65.037.2 (↓28)27.7 (↓37)40.2 (↓25)25.7 (↓39)
OpenAI CLIP ViT-B/16 [47]85.348.7 (↓37)29.7 (↓56)52.8 (↓32)31.0 (↓54)
OpenCLIP ViT-B/16 [26]97.754.0 (↓44)18.0 (↓80)59.3 (↓38)34.7 (↓63)
Meta-CLIP ViT-B/16 [62]91.049.3 (↓42)30.0 (↓61)53.8 (↓37)29.7 (↓61)
BLIP ViT-B/16 [30]82.748.4 (↓34)30.7 (↓52)51.3 (↓31)17.3 (↓65)
BLIP-2 ViT-L/14 [31]84.050.0 (↓34)32.0 (↓52)52.3 (↓32)27.7 (↓56)
SigLIP ViT-B/16 [65]95.355.3 (↓40)34.0 (↓61)59.3 (↓36)34.7 (↓61)
", + "bbox": [ + 511, + 121, + 901, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ended nature of VLM responses, aligning evaluations more closely with human preferences. We use GPT-4 to assess three aspects of the response on a 10-point scale: semantic accuracy, tone confidence, and overall coherence, with the scores summed to yield the GPT-Score (see Fig.1 (D)). For the VQA task, we introduce GPT-Acc, which leverages LLM-based judgment to evaluate response correctness. The judgment prompts are detailed in AppendixC.", + "bbox": [ + 509, + 224, + 906, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Research Questions (RQs) and Findings", + "text_level": 1, + "bbox": [ + 511, + 358, + 856, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RQ1: Do current VLMs demonstrate sufficient robustness to 3D variations? Take-away: \"Emphatically not. Our evaluation reveals the severe vulnerability of current VLMs to worst-case 3D variations across multiple vision-language tasks.\" As shown in Tab. 1, in zero-shot classification, Adv-3DT samples generated by AdvDreamer significantly compromise VLMs' performance, causing accuracy drops of up to $80\\%$ . This substantially outperforms the Random baseline, validating AdvDreamer's effectiveness in capturing worst-case transformations. Notably, samples under $\\Theta^{*}$ induce more severe performance degradation than those under the adversarial distribution $p^{*}(\\Theta)$ , indicating the existence of extremely vulnerable points.", + "bbox": [ + 511, + 383, + 906, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This fundamental vulnerability extends beyond classification to more complex open-ended tasks. As illustrated in Tab.3, for image captioning, Adv-3DT samples under both $p^* (\\Theta)$ and $\\Theta^{*}$ effectively degrade model performance across all evaluation metrics, with GPT-Score decreasing by nearly $50\\%$ across various VLMs. This significant degradation reveals a critical limitation in current VLMs' 3D understanding and reasoning capabilities. For VQA tasks, GPT4o, despite achieving the highest accuracy $(75.4\\%)$ on clean samples—exhibits a substantial $25\\%$ accuracy drop when under Adv-3DT attacks. Fig. 1 (A) provides visualizations of representative Adv-3DT samples generated by AdvDreamer, with more examples presented in Appendix.D.", + "bbox": [ + 511, + 580, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RQ2: To what extent do different VLMs share common 3D variation vulnerability regions? Take-away: \"Substantially. Our analysis reveals that Adv-3DT samples demonstrate strong cross-model transferability across various VLMs, indicating shared vulnerability regions in the 3D variation space.\" As demonstrated in Tab. 2, we investigate transferability by utilizing OpenCLIP and BLIP as surrogate models to optimize $p^*(\\Theta)$ , then evaluate six", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7899", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/112e95e535dc5a705d3bf28ae44ab8aef4ae74475c5c3df19a5a48f31e085a7f.jpg", + "table_caption": [ + "Table 3. Performance Degradation (↓) of VLMs on Image Captioning and VQA Tasks. For closed-source VLMs (GPT-4o/4o-mini), we performed transfer-based attacks, using OpenAI-CLIP (ViT-L@336px) as surrogate encoder to optimize $p^*(\\Theta)$ . For other VLMs, $p^*(\\Theta)$ are optimized w.r.t. respective image encoders (e.g., Eva-CLIP and OpenAI CLIP, etc.). Metrics: C-CIDEr, B-BLUE@4." + ], + "table_footnote": [], + "table_body": "
ModelMethodCOCO Caption [11]NoCaps [1]VQA-V2 [20]
CB@4GPT-ScoreCB@4GPT-ScoreGPT-Acc
BLIP-2 [30]Clean133.436.024.798.144.124.453.4
Random91.8 (↓42)23.5 (↓13)19.5 (↓5)60.2 (↓38)28.6 (↓16)20.9 (↓4)34.0 (↓19)
p*(Θ)83.8 (↓50)23.3 (↓13)18.9 (↓6)55.7 (↓42)26.9 (↓17)17.1 (↓7)31.7 (↓22)
Θ*72.2 (↓61)21.5 (↓15)17.0 (↓8)40.0 (↓58)22.4 (↓22)15.0 (↓9)28.0 (↓25)
LLaVa-1.5 [36]Clean126.130.425.0118.750.624.768.6
Random93.4 (↓33)25.6 (↓5)20.0 (↓5)75.3 (↓43)35.7 (↓15)21.7 (↓3)56.3 (↓12)
p*(Θ)85.8 (↓40)22.9 (↓8)18.9 (↓6)60.0 (↓59)29.6 (↓21)17.4 (↓7)54.3 (↓14)
Θ*73.2 (↓53)20.5 (↓10)16.5 (↓9)45.8 (↓73)25.5 (↓25)15.6 (↓9)53.8 (↓15)
LLaVa-1.6 [35]Clean41.66.922.930.510.822.671.2
Random25.1 (↓17)6.0 (↓1)17.5 (↓5)18.3 (↓12)7.1 (↓4)21.0 (↓2)54.3 (↓17)
p*(Θ)21.9 (↓20)5.3 (↓2)16.9 (↓6)13.6 (↓17)6.5 (↓4)15.7 (↓7)53.0 (↓18)
Θ*16.7 (↓25)4.6 (↓2)15.4 (↓8)10.5 (↓20)4.3 (↓7)13.5 (↓9)46.9 (↓24)
Otter [28]Clean127.034.924.076.429.722.553.9
Random71.6 (↓55)22.5 (↓12)21.0 (↓3)55.4 (↓21)23.7 (↓6)21.1 (↓1)49.7 (↓4)
p*(Θ)67.9 (↓59)17.8 (↓17)17.2 (↓7)45.4 (↓31)19.2 (↓11)16.2 (↓6)44.9 (↓9)
Θ*58.9 (↓68)17.5 (↓17)15.8 (↓8)36.0 (↓40)15.7 (↓14)14.6 (↓8)42.0 (↓12)
MiniGPT-4 [72]Clean78.523.424.060.828.524.751.3
Random58.2 (↓20)17.9 (↓6)19.7 (↓4)38.0 (↓23)19.8 (↓9)21.9 (↓3)43.7 (↓8)
p*(Θ)52.3 (↓26)16.3 (↓7)18.1 (↓6)32.9 (↓28)18.4 (↓10)16.7 (↓8)41.0 (↓10)
Θ*42.2 (↓36)12.9 (↓11)16.2 (↓8)26.6 (↓34)14.9 (↓14)15.0 (↓10)34.3 (↓17)
GPT-4o-mini [45]Clean23.73.023.315.72.923.068.6
Random14.1 (↓10)1.8 (↓1)16.9 (↓6)10.8 (↓5)0.9 (↓2)20.0 (↓3)51.7 (↓17)
p*(Θ)12.5 (↓11)1.2 (↓2)18.0 (↓5)7.8 (↓8)0.6 (↓2)17.2 (↓6)43.4 (↓25)
Θ*9.6 (↓14)1.6 (↓1)16.3 (↓7)7.3 (↓8)0.4 (↓3)15.5 (↓8)42.7 (↓26)
GPT-4o [45]Clean51.29.725.040.711.724.975.4
Random36.5 (↓15)5.1 (↓5)19.1 (↓6)24.7 (↓16)9.5 (↓2)20.9 (↓4)56.3 (↓19)
p*(Θ)26.4 (↓25)4.6 (↓5)17.7 (↓7)19.1 (↓22)6.5 (↓5)17.0 (↓8)50.7 (↓25)
Θ*20.3 (↓31)4.7 (↓5)16.1 (↓9)15.8 (↓25)5.4 (↓6)14.6 (↓10)50.3 (↓25)
", + "bbox": [ + 145, + 138, + 854, + 422 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45ea2226032cb952d532b1d73e6bbe929d56394b1a5d58b6447670aecacb80f2.jpg", + "table_caption": [ + "Table 4. Top-1 Accuracy and Naturalness Score of Adv-3DT samples generated by AdvDreamer w/ and w/o NRM. (under $p^{*}(\\Theta)$ )" + ], + "table_footnote": [], + "table_body": "
MethodsTop-1 Acc.ScoreRScoreP
AdvDreamer w/o. NRM48.6% (↓49.4)1.601.39
AdvDreamer w/. NRM54.0% (↓44.0)2.522.57
", + "bbox": [ + 107, + 465, + 465, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5. Visualization of Adv-3DT samples w/ and w/o NRM.", + "bbox": [ + 96, + 619, + 475, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "target VLMs with diverse architectures and training objectives. The transfer attacks maintain strong effectiveness, with performance degradation comparable to direct attacks. Notably, $p^*(\\Theta)$ exhibits smaller transfer gaps than $\\Theta^*$ , suggesting that learning the underlying distribution can better capture generalizable vulnerability regions. This transferability extends consistently to captioning and VQA tasks (Tab. 3). When employing OpenAI-CLIP as the surrogate encoder, Adv-3DT samples generated by AdvDreamer effectively compromise commercial systems like GPT-4o. We attribute this universal transferability to a fundamental limitation: current image-text pretraining datasets exhibit significant 3D biases and fail to encompass diverse certain special 3D transformations, leading to universal performance deficiencies under these transformation regions.", + "bbox": [ + 89, + 641, + 482, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RQ3: Does the performance degradation under Adv-3DT samples generated by AdvDreamer fundamentally", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "arise from 3D variations? Take-away: \"Yes, we successfully reproduce these Adv-3DT samples in physical environments and demonstrate that degradation is inherently caused by the 3D variation rather than undesirable image quality problems.\" With the naturalness feedback provided by NRM, AdvDreamer effectively prevents unwanted degradation of samples during the optimization process. To verify this, we evaluate Adv-3DT samples generated by AdvDreamer with and without NRM under ImageNet. Tab. 4 presents the accuracy under OpenCLIP (ViT-B/16), alongside GPT-4o-assessed visual realism $(Score_R)$ and physical plausibility $(Score_P)$ . The results indicate that NRM significantly improves samples' realism and physical feasibility, with a marginal trade-off in aggressiveness. Furthermore, Fig. 5 presents a qualitative comparison, illustrating the effectiveness of NRM's naturality feedback in improving the image quality and mitigating the distribution discrepancy compared to natural image distributions. In addition, we validate the alignment between NRM predictions and human evaluations, as detailed in Appendix C.2.", + "bbox": [ + 511, + 436, + 906, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we conduct physical-world experiments on 12 common objects from traffic and household environments to validate this. Our experimental procedure consisted of three main steps: 1) capturing images of objects in their natural state, 2) using AdvDreamer with OpenCLIP as the surrogate model to generate Adv-3DT samples, and 3) physically reproducing the adversarial 3D transformations captured by AdvDreamer through video recording. This process yielded 3,807 frames of physical Adv-3DT samples. We evaluate effectiveness through zero-shot classifi", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7900", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ff6655d3561980f8a81695232f81e1255ecf3c5b12e0fbb3b26bca2964e78de.jpg", + "table_caption": [ + "Table 5. Quantitative results of physical experiments." + ], + "table_footnote": [], + "table_body": "
SampleOpenCLIPLLaVa-1.5GPT-4o
Acc.Conf.Acc.1Acc.2Acc.1Acc.2
Natural (Physical)100.00.89283.310091.7100.0
Adv-3DT (AdvDreamer)0.0 (↓100)0.00325.0 (↓58)25.0 (↓75)25.0 (↓67)41.7 (↓58)
Adv-3DT (Physical)51.3 (↓49)0.41333.6 (↓50)45.1 (↓55)56.8 (↓35)64.8 (↓35)
", + "bbox": [ + 178, + 109, + 816, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f9a6bf5ad1d1e92827a403764db9ef89035254f643abf19fcf555156ba945ea5.jpg", + "image_caption": [ + "Figure 6. Attack success rate of Adv-3DT samples w.r.t. (A) optimization steps and (B) scaling ratios of the optimal distribution." + ], + "image_footnote": [], + "bbox": [ + 102, + 194, + 284, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/823610e126a0141de068c3a144c0f16c5153d0b618c8811fc9b075a9a60c749a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 193, + 464, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cation and VQA tasks. For VQA, we designed two evaluation protocols: multiple-choice classification (Acc.1) and binary verification (Acc.2). As shown in Table 5, physically reproduced Adv-3DT samples significantly degraded VLM performance: the accuracy of zero-shot classification dropped to $51.3\\%$ , while VQA performance decreased to $33.6\\%$ (Acc.1) and $45.1\\%$ (Acc.2). These results confirm that AdvDreamer successfully captures physically realizable 3D variations that represent real world corner cases. However, an aggressiveness gap between digital and physical attacks is observed, primarily attributed to data distribution differences and camera parameter fluctuations. Fig. 1 (B) presents representative examples from physical experiments, with more detailed results provided in Appendix D.", + "bbox": [ + 88, + 345, + 482, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies and Additional Analysis", + "text_level": 1, + "bbox": [ + 89, + 571, + 444, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Convergence Analysis & Computational Efficiency. We analyze the attack success rates (ASR) of $p^*(\\Theta)$ against OpenCLIP (ViT-B/16) and BLIP (ViT-B/16) across different optimization iterations. As shown in Fig. 6, AdvDreamer's effectiveness consistently improves with more iterations and converges after 15 steps. Thus, we set the iteration step default to 15 throughout our experiments to balance effectiveness and efficiency. The computational cost is discussed in Appendix C.4 by detailing the overhead of each component. For each clean sample, AdvDreamer requires an average of 0.28 GPU hours with a single RTX 3090.", + "bbox": [ + 89, + 595, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does Optimal Distribution $p^*(\\Theta)$ Effectively Characterize the Worst-case 3D Variations? We conduct analysis on ImageNet using five random natural images. For each image, we generate 100 Adv-3DT samples by scaling the variance $\\Sigma$ of their $p^*(\\Theta)$ with different ratios. Fig. 6(B) shows that ASR consistently decreases as the sampling range expands beyond the optimal distribution. This inverse relationship confirms that AdvDreamer successfully identifies the most vulnerable regions in the 3D variation space.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc00b9da4be52c7ea4a64396ae900b607f1ac28a22d41b60ea2eecf73e39f57.jpg", + "image_caption": [ + "Figure 7. Benchmarking result of VLMs' 3D variation robustness." + ], + "image_footnote": [], + "bbox": [ + 517, + 191, + 893, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. MM3DTBench", + "text_level": 1, + "bbox": [ + 511, + 304, + 666, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, We introduce MM3DTBench, it comprises 215 challenging Adv-3DT samples, consisting of samples generated by AdvDreamer and their physical replications. For each sample, we provide question templates with candidate answers and precise orientation annotations for the transformed objects. We employ two evaluation tasks: multiple-choice (Choice) and free-form description (Free answer), with accuracy computed through text-only GPT-4 judgment. More Details are provided in the Appendix.E. Our evaluation of 13 representative VLMs reveals robustness gaps. As shown in Fig. 7, even top-performing models (GPT-4o [45], CogVLM [61], and Intern-VL [13]) achieve limited success, while most models struggle with accuracy below $50\\%$ . Notably, Claude-3 [5], despite its recognized excellence, only achieves an accuracy of $30.0\\%$ . This performance correlates with models' general out-of-distribution robustness measured by MultiTrust [67], suggesting a fundamental limitation of current VLMs. We encourage the adoption of MM3DTBench in VLM development to advance security evaluation in dynamic scenarios.", + "bbox": [ + 511, + 325, + 906, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 640, + 633, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed AdvDreamer, a novel framework for characterizing real-world adversarial 3D transformation samples from single-view observations. Our comprehensive evaluation revealed critical robustness gaps in existing VLMs, highlighting the urgent need to enhance VLMs' 3D variation perception and understanding capabilities in safety-critical applications. Through MM3DTBench, we established the first benchmark for worst-case 3D variations, aiming to facilitate the development of more robust vision-language systems for real-world deployment.", + "bbox": [ + 511, + 666, + 906, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 830, + 671, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Fundamental Research Funds for the Central Universities, and was also supported by NSFC Projects (U2341228, 62276149).", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7901", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 6, 7", + "[2] Haider Al-Tahan, Quentin Garrido, Randall Balestriero, Diane Bouchacourt, Caner Hazirbas, and Mark Ibrahim. Unibench: Visual reasoning requires rethinking vision-language beyond scaling. arXiv preprint arXiv:2408.04810, 2024. 2", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[4] Michael A Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4845-4854, 2019. 2, 3, 4", + "[5] Anthropic. claude-3. https://www.anthropic.com/claude, 2024.8", + "[6] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 4", + "[7] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2", + "[8] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 3", + "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2", + "[10] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 7", + "[12] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level im" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2, 4", + "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 8", + "[14] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6", + "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 4", + "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2, 3", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[18] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. In Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6", + "[19] Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1487-1495, 2017. 5", + "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6, 7", + "[21] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. In ECCV, pages 22-38. Springer, 2020. 2, 3, 4", + "[22] Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772, 2016. 5, 6", + "[23] Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2", + "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7902", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3", + "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. openclip, 2021. 3, 6", + "[27] LAION-AI. Clip-benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2024.6", + "[28] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 7", + "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 6", + "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3, 6, 7", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2, 3, 6", + "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3", + "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 7", + "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 7", + "[37] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2024. 3", + "[38] Jiahuan Long, Tingsong Jiang, Wen Yao, Shuai Jia, Weijia Zhang, Weien Zhou, Chao Ma, and Xiaoqian Chen. Papmot: Exploring adversarial patch attack against multiple ob" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ject tracking. In European Conference on Computer Vision, pages 128-144. Springer, 2024. 2", + "[39] Jiahuan Long, Zhengqin Xu, Tingsong Jiang, Wen Yao, Shuai Jia, Chao Ma, and Xiaoqian Chen. Robust sam: On the adversarial robustness of vision foundation models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5775-5783, 2025. 2", + "[40] Jiahuan Long, Wen Yao, Tingsong Jiang, and Chao Ma. Cdupatch: Color-driven universal adversarial patch attack for dual-modal visible-infrared detectors. arXiv preprint arXiv:2504.10888, 2025. 2", + "[41] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3", + "[42] Spandan Madan, Timothy Henry, Jamell Dozier, Helen Ho, Nishchal Bhandari, Tomotake Sasaki, Frédo Durand, Hanspeter Pfister, and Xavier Boix. When and how convolutional neural networks generalize to out-of-distribution category-viewpoint combinations. Nature Machine Intelligence, 4(2):146-153, 2022. 2, 3, 4", + "[43] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 3", + "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3", + "[45] openai. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.3,7,8", + "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6", + "[48] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4", + "[50] Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, and Xingxing Wei. Towards viewpoint-invariant visual recognition via adversarial training. In ICCV, pages 1-10, 2023. 2, 3, 4, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7903", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Shouwei Ruan, Yinpeng Dong, Hang Su, Jianteng Peng, Ning Chen, and Xingxing Wei. Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528, 2023. 3", + "[52] Shouwei Ruan, Yinpeng Dong, Hanqing Liu, Yao Huang, Hang Su, and Xingxing Wei. Omniview-tuning: Boosting viewpoint invariance of vision-language pre-training models. arXiv preprint arXiv:2404.12139, 2024. 3", + "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3", + "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[55] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2", + "[56] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2", + "[57] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 4", + "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 3, 5", + "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3", + "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3", + "[61] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 8", + "[62] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In The Twelfth International Conference on Learning Representations, 2024. 6", + "[63] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3", + "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 6", + "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 2, 4", + "[67] Yichi Zhang, Yao Huang, Yitong Sun, Chang Liu, Zhe Zhao, Zhengwei Fang, Yifan Wang, Huanran Chen, Xiao Yang, Xingxing Wei, Hang Su, Yinpeng Dong, and Jun Zhu. Benchmarking trustworthiness of multimodal large language models: A comprehensive study, 2024. 2, 3, 6, 8", + "[68] Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. Proceedings of the European conference on computer vision (ECCV), 2024. 2", + "[69] Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, and Adam Kortylewski. Ood-cv: A benchmark for robustness to out-of-distribution shifts of individual nuisances in natural images. In European conference on computer vision, pages 163-180. Springer, 2022. 3", + "[70] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2", + "[71] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6", + "[72] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 696 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_model.json b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c7881944747b2ba052b24998db0ebf4888833dc --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.388, + 0.132, + 0.61, + 0.15 + ], + "angle": 0, + "content": "AdvDreamer Unveils:" + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.154, + 0.87, + 0.175 + ], + "angle": 0, + "content": "Are Vision-Language Models Truly Ready for Real-World 3D Variations?" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.203, + 0.804, + 0.24 + ], + "angle": 0, + "content": "Shouwei Ruan\\(^{1\\dagger}\\), Hanqing Liu\\(^{1\\dagger}\\), Yao Huang\\(^{1\\dagger}\\), Xiaqi Wang\\(^{1}\\), Caixin Kang\\(^{1}\\), Hang Su\\(^{2}\\), Yinpeng Dong\\(^{3}\\), Xingxing Wei\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.239, + 0.636, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Artificial Intelligence," + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.257, + 0.835, + 0.275 + ], + "angle": 0, + "content": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.275, + 0.829, + 0.31 + ], + "angle": 0, + "content": "\\(^{2}\\) Dept. of Comp. Sci. and Tech., Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab, BNRist Center, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.31, + 0.645, + 0.328 + ], + "angle": 0, + "content": "3 College of AI, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.332, + 0.867, + 0.344 + ], + "angle": 0, + "content": "{shouweiruan, hqliu, y_huang, xxwei}@buaa.edu.cn, suhangss@tsinghua.edu.cn, dongyinpeng@mail.tsinghua.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.351, + 0.842, + 0.362 + ], + "angle": 0, + "content": "(A) Adversarial 3D Variations Samples form Single-view Natural Images —— (C) Background Variations for Adv-3DT samples" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.369, + 0.276, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.369, + 0.468, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.369, + 0.634, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.444, + 0.39, + 0.453 + ], + "angle": 0, + "content": "(B) Physical Reproduction of Adversarial 3D Variations Samples" + }, + { + "type": "image", + "bbox": [ + 0.135, + 0.454, + 0.266, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.454, + 0.385, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.397, + 0.457, + 0.507, + 0.559 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.457, + 0.636, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.353, + 0.866, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.488, + 0.865, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.581, + 0.907, + 0.637 + ], + "angle": 0, + "content": "Figure 1. The 3D Variation Vulnerabilities in VLMs. (A) The proposed AdvDreamer allows the capture of adversarial 3D transformation (Adv-3DT) samples from single natural images, threatening various VLMs. (B) These Adv-3DT samples can be successfully reproduced in the physical world and (C) retain their aggressiveness under background variations. (D) We also explain the GPT-Score metric used in the experiments. In each VLM response, we highlight the correct, significant error, and less precise parts using different colors." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.65, + 0.327, + 0.665 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Vision Language Models (VLMs) have exhibited remarkable generalization capabilities, yet their robustness in dynamic real-world scenarios remains largely unexplored. To systematically evaluate VLMs' robustness to real-world 3D variations, we propose AdvDreamer, the first framework capable of generating physically reproducible Adversarial 3D Transformation (Adv-3DT) samples from single-view observations. In AdvDreamer, we integrate three key innovations: Firstly, to characterize real-world 3D variations with limited prior knowledge precisely, we design a zero-shot Monocular Pose Manipulation pipeline built upon generative 3D priors. Secondly, to ensure the visual qual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.909, + 0.893 + ], + "angle": 0, + "content": "ity of worst-case Adv-3DT samples, we propose a Naturalness Reward Model that provides continuous naturalness regularization during adversarial optimization, effectively preventing convergence to hallucinated or unnatural elements. Thirdly, to enable systematic evaluation across diverse VLM architectures and visual-language tasks, we introduce the Inverse Semantic Probability loss as the adversarial optimization objective, which solely operates in the fundamental visual-textual alignment space. Based on the captured Adv-3DT samples with high aggressiveness and transferability, we establish MM3DTBench, the first VQA benchmark dataset tailored to evaluate VLM robustness under challenging 3D variations. Extensive evaluations of representative VLMs with varying architectures reveal that real-world 3D variations can pose severe threats to model performance across various tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.876, + 0.228, + 0.888 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.889, + 0.246, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.876, + 0.246, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7894" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.286, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.092, + 0.478, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.092, + 0.667, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.672, + 0.092, + 0.862, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.202, + 0.806, + 0.217 + ], + "angle": 0, + "content": "Figure 2. Comparison of paradigms to characterize Adversarial 3D Transformation (Adv-3DT) samples." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.224, + 0.223, + 0.239 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.431 + ], + "angle": 0, + "content": "Recently, Vision-Language Models (VLMs) [7, 13, 35, 36] have demonstrated remarkable capabilities in bridging visual perception and natural language understanding. Through large-scale image-text pre-training followed by fine-grained instruction tuning [31, 36], these models effectively tackle a wide range of visual-centric tasks, including instruction-based image recognition [47, 55, 65], visual question answering [3, 31, 35, 36], and visual reasoning [9, 10]. Due to their strong generalizability and impressive zero-shot capability, VLMs are increasingly deployed in various safety-critical applications, particularly in autonomous driving [56, 63] and robotic systems [25, 64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Despite extensive studies validating VLMs' resilience against various distribution shifts (e.g., style variations [2, 39, 40, 68], image corruptions [2], and adversarial perturbations [16, 38, 67, 70]), they primarily focus on 2D perturbations in the digital domain, overlooking a critical challenge in real-world deployment: the 3D variations. As VLMs are increasingly integrated into dynamic scenarios, their ability to handle 3D variation becomes crucial, raising a fundamental question: Have current VLMs attained sufficient robustness to the distribution shifts arising from ubiquitous 3D variations in the real world?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.629 + ], + "angle": 0, + "content": "To thoroughly explore this problem and offer comprehensive answers, we identify three fundamental challenges:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(1) How to accurately characterize real-world 3D variations with limited prior knowledge? Existing methods exploring the 3D variation robustness heavily rely on explicit 3D structures [4, 21] or representations derived from dense multi-view observations [4, 21, 42]. However, in most practical scenarios, the available priors are severely limited (e.g., single-view observation only). This constraint necessitates developing a method that minimizes dependency on extensive scene priors. To meet this, we introduce a zero-shot Monocular Pose-Manipulation (MPM) pipeline based on generative representations. Specifically, MPM leverages the rich 3D priors embedded in pre-trained Large-Reconstructive Models (LRM) [23, 24] and Image Re-Composition Models [12, 66], enabling fine-grained 3D adjustments with only single-view images and specified 3D variation parameters. While this generative-based paradigm establishes a foundation for capturing worst-case 3D variations through optimizing 3D variation parameters, it still" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.905, + 0.256 + ], + "angle": 0, + "content": "exhibits inherent limitations that may introduce undesirable image corruptions, thus presenting our second challenge:" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.259, + 0.907, + 0.56 + ], + "angle": 0, + "content": "(2) How to capture the worst-case adversarial 3D transformation (Adv-3DT) samples while ensuring that the performance degradation is inherently caused by the 3D variation? Prior works utilizing synthetic 3D assets [4, 21, 42] frequently produce samples with noticeable texture discrepancies compared to real-world images. These limitations in samples' visual fidelity and contextual coherence highlight a critical concern: observed performance degradation under Adv-3DT samples may stem from undesirable corruption rather than the 3D variation itself. To address this, we incorporate naturalness reward constraints within the adversarial optimization process. Specifically, we treat rigid 3D variations as an adversarial attack and propose the Naturalness Reward Model (NRM) that preserves image quality throughout optimization. NRM leverages visual context features provided by the DINOv2 [46] backbone to predict a naturalness reward that jointly reflects the visual fidelity and physical plausibility of the samples, effectively regularizing the optimization trajectory and preventing convergence toward unnatural or hallucinated elements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.882 + ], + "angle": 0, + "content": "(3) How to ensure effective generalization of Adv-3DT samples across diverse VLM tasks and architectures? Existing studies [4, 18, 50] typically rely on task-specific objectives with end-to-end optimization, substantially limiting their applicability across diverse VLM architectures and a wide range of visual language tasks. To transcend this limitation, we formulate the adversarial objective in a task-and architecture-agnostic manner. Specifically, we introduce the Inverse Semantic Probability (ISP) loss, which operates solely through the foundational component of the VLM—the visual encoder, aiming to minimize the image-text matching probability between Adv-3DT samples and their corresponding ground-truth textual descriptions. Extensive experiments demonstrate that Adv-3DT samples optimized under the ISP loss exhibit remarkable adversarial transferability across various downstream tasks and models. Building on this, we further introduce the MM3DTBench, the first VQA benchmark dataset specifically designed to evaluate the 3D variation robustness of VLMs in real-world scenarios, comprising highly challenging Adv-3DT samples with carefully crafted candidate answers." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Integrating the aforementioned innovations, this paper" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7895" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "proposes AdvDreamer, a novel framework that enables the characterization of Adv-3DT samples from monocular observations, with broad adaptability across various visual-language tasks and VLM architectures. Our extensive digital and physical experiments reveal significant vulnerabilities in contemporary VLMs when confronted with challenging 3D variations. In visual recognition tasks, foundational models like OpenCLIP [26] and BLIP-2 [31] suffer a performance degradation of \\(65\\% \\sim 80\\%\\) on Adv-3DT samples. More concerning still, in open-ended tasks including image captioning and visual question answering, even the most advanced commercial models like GPT-4o [45] exhibit performance declines approaching \\(50\\%\\). Our findings highlight the urgent need to enhance VLM robustness against 3D variations and set new requirements for deploying VLMs in complex, dynamic real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.234, + 0.365 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.375, + 0.318, + 0.391 + ], + "angle": 0, + "content": "2.1. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.579 + ], + "angle": 0, + "content": "Classic Vision-Language Models (VLMs) [30, 32, 41, 47] achieve cross-modal alignment through diverse pre-training objectives on large-scale image-text pairs [33, 53]. Recent advances in Large Language Models (LLMs) [59, 60] have inspired new VLM architectures that integrate LLM capabilities via MLP projection layers [35, 36] or specialized modules like Q-Former [31]. Through techniques like instruction tuning [36], current VLMs [13, 35, 36] demonstrate enhanced performance across vision-centric tasks, including visual grounding, question answering, and image captioning, etc. These models also enable intelligent systems for task planning and decision-making [25, 64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Despite their impressive capabilities [37, 47, 67], concerns remain about the robustness and generalization of VLMs, especially in safety-critical environments. Recent studies [52, 58] highlight that VLMs still exhibit deficiencies in processing visual details, such as orientation, quantity, and colour, which may be attributed to their visual representations. Moreover, VLMs remain susceptible to \\( L_{p} \\)-norm adversarial perturbations [16, 43], with some efforts aiming to enhance their robustness through adversarial finetuning techniques [43]. In contrast to previous works, our study specifically investigates VLM robustness against 3D variations, addressing a crucial gap in understanding their reliability for dynamic real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.433, + 0.803 + ], + "angle": 0, + "content": "2.2. Evaluation for 3D Variation Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Robustness to 3D variations remains a longstanding challenge in computer vision. Evaluating this property often involves capturing worst-case 3D transformations, i.e. the adversarial poses or viewpoints, to create challenging samples for assessment. Prior research can be categorized into three paradigms (see Fig.2): (A) Human-guided methods, like" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.334 + ], + "angle": 0, + "content": "ObjectNet [8] and OOD-CV [69], manually collect samples with diverse poses and viewpoints, but they lack systematic coverage of worst-case examples and are costly. (B) Simulator-based [4, 21, 42] methods optimize worst-case 3D parameters via differentiable rendering of synthetic objects but require manually designed 3D assets and produce samples with limited realism. (C) Multi-view-based methods, like ViewFool [18] and GMVFool [50, 51], use NeRF-based [44] representations to optimize adversarial viewpoints, but they require dense observations and struggle with background integration. In contrast, the proposed AdvDreamer employs a (D) Generative-based approach, utilizing robust 3D priors from generative models. Given a single natural image, AdvDreamer generate realistic Adv-3DT samples in a zero-shot manner, providing a more efficient solution for evaluating 3D variation robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.649, + 0.363 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.553 + ], + "angle": 0, + "content": "The overview of AdvDreamer is shown in Fig. 3. It consists \\(\\bullet\\) Zero-shot Monocular Pose Manipulation pipeline, leveraging generative 3D representations to obtain specified 3D variation samples from single-view images; \\(\\bullet\\) Naturalness Reward Model, providing naturalness regularization during adversarial optimization; and \\(\\bullet\\) Inverse Semantic Probability loss, ensuring the Adv-3DT samples effectively threaten VLMs and generalize across different architectures and tasks. The following sections will firstly formalize the optimization problem of AdvDreamer (Sec.3.1), then sequentially describe the key components (Sec.3.2), and finally outline the full optimization algorithm (Sec.3.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.719, + 0.575 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Parametrization of Real-World 3D Variations. In this paper, we primarily focus on rigid 3D transformations of foreground objects, i.e., the rotation, translation, and scaling, since they reflect the most typical 3D changes observed in real-world environments. Formally, we define a 6-dimensional vector \\(\\Theta = \\{\\alpha, \\beta, \\gamma, \\Delta_x, \\Delta_y, s\\}\\) to uniquely parameterize any arbitrary transformation, where \\(\\{\\alpha, \\beta, \\gamma\\}\\) denote the Tait-Bryan angles (yaw, pitch, roll) in the \\(Z-Y-X\\) sequence, \\(\\{\\Delta_x, \\Delta_y\\}\\) represent the translation along the \\(x\\) and \\(y\\) axes in the \\(xy\\)-plane, and \\(s\\) is the uniform scaling factor. To prevent the loss of critical visual cues and ensure human recognizability, we follow previous work [18] to constrain \\(\\Theta\\) within a bounded range \\([\\Theta \\min, \\Theta \\max]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.909, + 0.855 + ], + "angle": 0, + "content": "Optimization Problem. The objective of AdvDreamer is to find an optimal distribution \\( p^*(\\Theta) \\), ensuring that Adv-3DT sampled from this distribution are both aggressive against VLMs and maintaining visual quality, which can be abstractly formulated as the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.862, + 0.905, + 0.888 + ], + "angle": 0, + "content": "\\[\np ^ {*} (\\Theta) = \\arg \\max _ {p (\\Theta)} \\mathbb {E} _ {\\Theta \\sim p (\\Theta)} \\left[ \\mathcal {L} _ {I S P} \\left(X ^ {\\prime}\\right) + \\mathcal {L} _ {N a t} \\left(X ^ {\\prime}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.888, + 0.779, + 0.904 + ], + "angle": 0, + "content": "\\[\nw h e r e X ^ {\\prime} = \\mathcal {T} (\\Theta , X),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7896" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.143, + 0.092, + 0.852, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.299, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed AdvDreamer framework. To capture worst-case 3D variations in the real world, we treat rigid 3D transformations as adversarial attacks, optimizing the distribution of transformation parameters through a query-based approach. In AdvDreamer, we introduce the Monocular Pose Manipulation pipeline to perform specified 3D transformations on single-view images and guide the optimization process using the proposed Naturalness Reward Model and Inverse Semantic Probability loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.472 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}(\\Theta, X)\\) is a transformation function to produce a novel image with the specified transformation \\(\\Theta\\) applied to the original image \\(X \\in \\mathbb{R}^3\\). To implement \\(\\mathcal{T}(\\cdot)\\), we design the MPM pipeline, which will be detailed in Sec.3.2.1. The \\(\\mathcal{L}_{Nat}\\) and \\(\\mathcal{L}_{ISP}\\) measure the naturalness and adversarial efficacy of the transformed samples, respectively. These will be further elaborated in Sec.3.2.2 and Sec.3.2.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.624 + ], + "angle": 0, + "content": "Optimizing over the distribution \\( p(\\Theta) \\) rather than a deterministic \\( \\Theta \\) offers two key advantages. 1) As highlighted in [18, 50], learning the underlying distribution enables comprehensive exploration of the 3D variation space, facilitating the identification of the model's vulnerable regions. 2) Given the stochastic nature of the generation, optimizing over a continuous distribution can mitigate the impact of appearance inconsistency, as appearance variations induced by the generation are unlikely to consistently deceive the model without optimizing the adversarial 3D variations [6]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.325, + 0.648 + ], + "angle": 0, + "content": "3.2. AdvDreamer Framework" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.421, + 0.671 + ], + "angle": 0, + "content": "3.2.1. Zero-shot Monocular Pose Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.78 + ], + "angle": 0, + "content": "To implement the transformation function \\(\\mathcal{T}\\), previous approaches typically rely on explicit 3D structures [4, 21, 42] or neural radiance representations [18, 50]. However, these approaches cannot accommodate real-world scenarios with limited prior knowledge. To address this, we design a simple yet effective zero-shot monocular pose manipulation (MPM) pipeline that consists of the following process:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "1 Foreground-Background Pairs Preparation. MPM first decompose the input image \\(X\\) into foreground-background pairs \\(\\{X_{f},X_{b}\\}\\). Specifically, we leverage Grounded-SAM [48] to semantically segment the major instance as foreground \\(X_{f}\\), followed by diffusion-based inpainting [15, 49] to obtain a complete background \\(X_{b}\\). This process is denoted as \\(\\{X_f,X_b\\} = \\mathcal{F}(X)\\). Additionally, we support directly generating synthetic \\(\\{X_f,X_b\\}\\) pairs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.903, + 0.383 + ], + "angle": 0, + "content": "through Stable-Diffusion [49] to enhance sample diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.519 + ], + "angle": 0, + "content": "\\(\\Theta\\) Object-level Pose Manipulation. For each optimization iteration, we transform the foreground \\(X_{f}\\) into a batch of \\(X_{f}^{\\prime}\\) under \\(\\Theta\\) sampled from current \\(p(\\Theta)\\). In this stage, MPM leverages the robust priors provided by a pre-trained Large Reconstruction Model (e.g., TripoSR [57]) to construct single-view-based 3D representations and apply sampled transformations. This process is formulated as \\(X_{f}^{\\prime} = \\mathcal{R}_{\\mathbf{w}_{0}}(X_{f},\\Theta)\\), where \\(\\Theta \\sim p(\\Theta)\\) and \\(\\mathcal{R}_{\\mathbf{w}}(\\cdot)\\) denotes neural rendering function with weights \\(\\mathbf{w}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.905, + 0.595 + ], + "angle": 0, + "content": "\\(\\Theta\\) Re-Composition. We then compose the transformed foreground \\(X_{f}^{\\prime}\\) with background \\(X_{b}\\) through a pre-trained diffusion-based composition model [12, 66], to handles shadows and occlusions. This process is defined as \\(X^{\\prime} = \\mathcal{C}_{\\mathbf{w}_1}(X_f',X_b)\\), where \\(\\mathbf{w}_1\\) denotes diffusion model's weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.625 + ], + "angle": 0, + "content": "Combining the above stages, the transformation process by MPM can be formally represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.637, + 0.905, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r l} X ^ {\\prime} = \\mathcal {T} (\\boldsymbol {\\Theta}, X) & = \\mathcal {C} _ {\\mathbf {w} _ {1}} \\left(\\mathcal {R} _ {\\mathbf {w} _ {0}} \\left(X _ {f}, \\boldsymbol {\\Theta}\\right), X _ {b}\\right), \\\\ \\text {w h e r e} \\{X _ {f}, X _ {b} \\} & = \\mathcal {F} (X). \\end{array} \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.753, + 0.699 + ], + "angle": 0, + "content": "3.2.2. Naturalness Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Despite the MPM pipeline utilizing generative priors to achieve realistic 3D variation samples, we observe that directly applying adversarial optimization frequently leads to unnatural artifacts, such as shape distortions and physically implausible cases in Fig. 5. These issues stem from two factors: generalization limitations of existing generative models and the inherent high semantic bias in low-quality samples. To address this challenge, we propose a novel Naturalness Reward Model (NRM) that continuously regularizes \\( p(\\Theta) \\) during optimization based on sample naturalness, preventing convergence to low-quality pseudo-optimal regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig.4, the NRM employs DINOv2[46] as its backbone, chosen for several reasons. Firstly, DINO" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.472, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Figure 4. Architecture of NRM and computational cost and accuracy of naturalness assessment across Human, GPT, and NRM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.349 + ], + "angle": 0, + "content": "captures finer visual contextual features compared to traditional pre-trained vision encoders [58], aiding better convergence during training. Secondly, unlike vision-language aligned encoders such as CLIP, DINO's visual features are language-agnostic, making them more faithfully reflect image quality rather than semantic consistency. We use DINOv2 to extract image tokens from the input Adv-3DT samples, followed by dual-stream prediction heads to estimate visual fidelity and physical plausibility scores:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.391 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) = f _ {R} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right) = f _ {P} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.527 + ], + "angle": 0, + "content": "where \\( f_{R}(\\cdot) \\) and \\( f_{P}(\\cdot) \\) are two-layer MLPs. To train the NRM, we construct a large-scale naturalness preference dataset with fine-grained scoring. We generate numerous samples from the MPM pipeline on ImageNet and synthetic images, scoring each on a 5-point scale for visual fidelity and physical plausibility. These scores are initially annotated automatically by GPT-4o with Chain-of-Thought prompts and verified by volunteers for correction. This process resulted in 120K high-quality training samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.65 + ], + "angle": 0, + "content": "The architecture and training parameters of NRM are detailed in Appendix B. Our experimental results demonstrate that the naturalness feedback provided by NRM highly aligns with human evaluation. In terms of prediction efficiency and stability, it outperforms methods that directly employ GPT-4o for naturalness assessment (see Fig. 4 and Appendix C.2). Based on the NRM's predictions, we define the naturalness regularization loss for the optimization as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.66, + 0.483, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) = - \\frac {1}{2} \\left(\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) + \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right)\\right) \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.377, + 0.715 + ], + "angle": 0, + "content": "3.2.3. Inverse Semantic Probability Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Previous studies typically employ task-specific optimization objectives with end-to-end manner to obtain worst-case Adv-3DT samples. However, directly applying this approach to VLMs with substantial parameters and diverse architectures spanning various tasks is challenging. To address this, we design the Inverse Semantic Probability (ISP) loss, which minimizes the probability that VLMs assign to the correct semantic attributes of Adv-3DT samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This design have two key considerations. 1) It is based solely on the visual and text encoder \\(\\mathcal{E}_v\\), \\(\\mathcal{E}_t\\), which are fundamental components across modern VLMs, ensuring architecture-agnostic applicability. 2) By operating in the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "image-text alignment space rather than model-specific layers or task-specific heads, it enables task generalization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.168 + ], + "angle": 0, + "content": "Given a generated sample \\( X' \\) and a semantic label set \\( \\mathcal{Y} = \\{y_i\\}_{i=1}^N \\) containing the ground-truth label \\( y_t \\), we first compute a semantic similarity vector \\( \\mathbf{S} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.174, + 0.905, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} = \\left\\{s _ {i} \\right\\} _ {i = 1} ^ {N}, \\text {w h e r e} s _ {i} = \\cos \\left[ \\mathcal {E} _ {v} \\left(X ^ {\\prime}\\right), \\mathcal {E} _ {t} \\left(\\mathcal {P} \\left(y _ {i}\\right)\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.906, + 0.276 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot)\\) computes cosine similarity and \\(\\mathcal{P}(\\cdot)\\) follows the template \"a photo of a \\(\\{y_i\\}\\)\" in [47]. The vector \\(\\mathbf{S}\\) captures the representation similarities between \\(X'\\) and each semantic label. We then derive the conditional semantic probability through softmax transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.281, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\np \\left(y _ {i} \\mid X ^ {\\prime}\\right) = \\exp \\left(s _ {i}\\right) / \\sum_ {j = 1} ^ {N} \\exp \\left(s _ {j}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.308, + 0.906, + 0.339 + ], + "angle": 0, + "content": "Then, the \\(\\mathcal{L}_{\\mathrm{ISP}}\\) is defined as the negative log-likelihood of the ground-truth's semantic probability:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.345, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) = - \\log p \\left(y _ {t} \\mid X ^ {\\prime}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.906, + 0.431 + ], + "angle": 0, + "content": "Intuitively, \\(\\mathcal{L}_{\\mathrm{ISP}}\\) quantifies the likelihood of \\(X^{\\prime}\\) being misclassified by comparing its similarity to the ground-truth and irrelevant categories. Maximizing it directs the \\(p(\\Theta)\\) towards regions that effectively deceive the model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.854, + 0.455 + ], + "angle": 0, + "content": "3.3. Query-Based Adversarial Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.905, + 0.492 + ], + "angle": 0, + "content": "To solve Problem (1), we parameterize the adversarial distribution \\( p(\\Theta) \\) as a Multivariate Gaussian form:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.499, + 0.905, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Theta} = \\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, \\text {w h e r e} \\mathbf {z} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.906, + 0.569 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} = (\\Theta_{\\mathrm{min}} - \\Theta_{\\mathrm{max}}) / 2\\), \\(\\mathbf{A} = (\\Theta_{\\mathrm{min}} + \\Theta_{\\mathrm{max}}) / 2\\), \\(\\mathbf{z}\\) follows a Gaussian Distribution with mean \\(\\pmb{\\mu} \\in \\mathbb{R}^6\\) and covariance \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{6 \\times 6}\\). Thus, we rewrite (1) in a specific form:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.576, + 0.905, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}} {\\arg \\max } \\mathbb {E} _ {\\boldsymbol {\\mathbf {z}} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma})} \\left[ \\mathcal {L} _ {\\text {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) + \\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) \\right], \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.604, + 0.836, + 0.62 + ], + "angle": 0, + "content": "\\[\nw h e r e X ^ {\\prime} = \\mathcal {T} (\\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, X).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Solving this problem typically involves computing the gradient of the loss w.r.t. \\(\\pmb{\\mu}\\), \\(\\pmb{\\Sigma}\\) and updating them via gradient ascent. However, the forward steps involve multiple components, e.g., the denoise process, introducing uncertainty in the gradient propagation path, making gradient-based optimization challenging [54]. Therefore, we adopt the Covariance Matrix Adaptation Evolution Strategy [19, 22], an efficient query-based black-box optimizer. At each iteration \\(t\\), we: 1) Sample \\(K\\) candidates \\(\\{\\mathbf{z}_i^{t + 1}\\}_{i = 1}^K\\) from \\(\\mathcal{N}(\\boldsymbol {\\mu}^t,\\boldsymbol {\\Sigma}^t)\\). 2) Generate corresponding samples \\(\\{(X')_{i}^{t + 1}\\}_{i = 1}^{K}\\). 3) Update distribution parameters \\(\\pmb{\\mu}^t,\\pmb{\\Sigma}^t\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.794, + 0.905, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {t + 1} = \\sum_ {i = 1} ^ {k} w _ {i} \\cdot \\mathbf {z} _ {(i: k)} ^ {t + 1}, \\quad \\text {w h e r e} \\sum_ {i = 1} ^ {k} w _ {i} = 1, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.84, + 0.86, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {t + 1} = \\left(1 - \\eta_ {1} - \\eta_ {\\mu}\\right) \\cdot \\boldsymbol {\\Sigma} ^ {t} + \\eta_ {1} \\cdot p _ {\\Sigma} ^ {t + 1} \\left(p _ {\\Sigma} ^ {t + 1}\\right) ^ {T} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.862, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\eta_ {\\mu} \\cdot \\sum_ {i = 1} ^ {h} w _ {i} \\cdot \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) ^ {T}, \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7898" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 1. Zero-shot Classification Accuracy \\((\\%)\\) and degradation \\((\\downarrow \\%)\\) Under Adv-3DT Samples Generated from ImageNet [17]." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.122, + 0.486, + 0.219 + ], + "angle": 0, + "content": "
Target Models#ParamsMethods
CleanRandomp*(Θ)Θ*
OpenCLIP ViT-B/16 [26]149M98.062.6 (↓35)54.0 (↓44)18.0 (↓80)
OpenCLIP ViT-L/14 [26]428M94.461.7 (↓33)50.9 (↓44)15.3 (↓79)
OpenCLIP ViT-G/14 [26]2.5B96.463.5 (↓33)53.5 (↓43)18.7 (↓78)
BLIP ViT-B/16 [30]583M83.056.0 (↓27)51.3 (↓32)17.3 (↓66)
BLIP-2 ViT-L/14 [31]3.4B84.057.0 (↓27)52.2 (↓32)18.7 (↓65)
BLIP-2 ViT-G/14 [31]4.1B81.055.7 (↓25)49.1 (↓32)15.7 (↓65)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.485, + 0.362 + ], + "angle": 0, + "content": "here, \\(\\mathbf{z}_{(i:k)}\\) denotes samples ranked by losses: we first select top-10 samples based on \\(\\mathcal{L}_{\\mathrm{ISP}}\\), then choose the top-k \\((k = 5)\\) with highest \\(\\mathcal{L}_{\\mathrm{Nat}}\\) for updates. \\(w_{i} = 1 / k\\) are importance weights, \\(\\eta_{1}\\) and \\(\\eta_{\\mu}\\) are learning rates, \\(p_{\\Sigma}\\) tracks the evolution path, and \\(\\sigma^t\\) is the step size updated via Cumulative Step-size Adaptation (CSA) [22]. Due to space constraints, complete derivations of Eq. (10), (11) and hyperparameter settings are provided in the Appendix.A.1. We also provide the pseudocode for optimization process in Appendix A.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.371, + 0.226, + 0.389 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.396, + 0.282, + 0.412 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.599 + ], + "angle": 0, + "content": "Tasks & Datasets. We evaluate VLMs' robustness to 3D variations using AdvDreamer on three representative vision-language tasks. (1) For zero-shot image classification, we utilize the ImageNet test set [17]. (2) For image captioning, we employ the test splits from COCO Caption [11] and NoCaps [1] datasets. (3) For Visual Question Answering (VQA), we use the VQAv2 [20] test set. For each dataset, we select 300 clean samples spanning 30 categories with clear and well-defined object instances. During the optimization process, we employ ImageNet-1K categories as the semantic label set \\(\\mathcal{Y}\\) for the classification task and use COCO-80 categories for the captioning and VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Victim Models. For zero-shot classification, we target vision-language foundation models: OpenCLIP [14, 26], BLIPs [30, 31], and additional models for transfer attacks. For other tasks, we select mainstream open-source VLMs (LLaVa [34, 36], MiniGPT-4 [72], etc.) as well as representative commercial VLMs (GPT-4o and GPT-4o-mini)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Methods & Baselines. We use two variants with AdvDreamer: (1) \\( p^*(\\Theta) \\): represents the average results over 10 Adv-3DT samples randomly sampled from the optimal adversarial distribution, while (2) \\( \\Theta^* \\): indicates the evaluation results using the Adv-3DT sample corresponding to the distribution center. Additionally, we establish two potential baselines following [18, 50]: the original natural images without transformation (Clean) and the average results over 10 images with random 3D transformations (Random). Metrics. For zero-shot classification, we follow the CLIP-Benchmark [27] protocol, reporting Top-1 accuracy and its degradation relative to clean samples. For captioning, in addition to conventional metrics (CIDEr, BLEU), we adopt the LLM-as-a-judge approach [36, 67, 71] to address the open" + }, + { + "type": "table_caption", + "bbox": [ + 0.507, + 0.089, + 0.902, + 0.118 + ], + "angle": 0, + "content": "Table 2. Cross-model transferability of Adv-3DT samples, utilizing OpenCLIP and BLIP as surrogate models." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.122, + 0.902, + 0.218 + ], + "angle": 0, + "content": "
Target ModelCleanOpenCLIP [26]BLIP [30]
p*(Θ)Θ*p*(Θ)Θ*
ALBEF [29]65.037.2 (↓28)27.7 (↓37)40.2 (↓25)25.7 (↓39)
OpenAI CLIP ViT-B/16 [47]85.348.7 (↓37)29.7 (↓56)52.8 (↓32)31.0 (↓54)
OpenCLIP ViT-B/16 [26]97.754.0 (↓44)18.0 (↓80)59.3 (↓38)34.7 (↓63)
Meta-CLIP ViT-B/16 [62]91.049.3 (↓42)30.0 (↓61)53.8 (↓37)29.7 (↓61)
BLIP ViT-B/16 [30]82.748.4 (↓34)30.7 (↓52)51.3 (↓31)17.3 (↓65)
BLIP-2 ViT-L/14 [31]84.050.0 (↓34)32.0 (↓52)52.3 (↓32)27.7 (↓56)
SigLIP ViT-B/16 [65]95.355.3 (↓40)34.0 (↓61)59.3 (↓36)34.7 (↓61)
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.225, + 0.907, + 0.348 + ], + "angle": 0, + "content": "ended nature of VLM responses, aligning evaluations more closely with human preferences. We use GPT-4 to assess three aspects of the response on a 10-point scale: semantic accuracy, tone confidence, and overall coherence, with the scores summed to yield the GPT-Score (see Fig.1 (D)). For the VQA task, we introduce GPT-Acc, which leverages LLM-based judgment to evaluate response correctness. The judgment prompts are detailed in AppendixC." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.857, + 0.376 + ], + "angle": 0, + "content": "4.2. Research Questions (RQs) and Findings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.907, + 0.58 + ], + "angle": 0, + "content": "RQ1: Do current VLMs demonstrate sufficient robustness to 3D variations? Take-away: \"Emphatically not. Our evaluation reveals the severe vulnerability of current VLMs to worst-case 3D variations across multiple vision-language tasks.\" As shown in Tab. 1, in zero-shot classification, Adv-3DT samples generated by AdvDreamer significantly compromise VLMs' performance, causing accuracy drops of up to \\(80\\%\\). This substantially outperforms the Random baseline, validating AdvDreamer's effectiveness in capturing worst-case transformations. Notably, samples under \\(\\Theta^{*}\\) induce more severe performance degradation than those under the adversarial distribution \\(p^{*}(\\Theta)\\), indicating the existence of extremely vulnerable points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.779 + ], + "angle": 0, + "content": "This fundamental vulnerability extends beyond classification to more complex open-ended tasks. As illustrated in Tab.3, for image captioning, Adv-3DT samples under both \\( p^* (\\Theta) \\) and \\( \\Theta^{*} \\) effectively degrade model performance across all evaluation metrics, with GPT-Score decreasing by nearly \\( 50\\% \\) across various VLMs. This significant degradation reveals a critical limitation in current VLMs' 3D understanding and reasoning capabilities. For VQA tasks, GPT4o, despite achieving the highest accuracy \\( (75.4\\%) \\) on clean samples—exhibits a substantial \\( 25\\% \\) accuracy drop when under Adv-3DT attacks. Fig. 1 (A) provides visualizations of representative Adv-3DT samples generated by AdvDreamer, with more examples presented in Appendix.D." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "RQ2: To what extent do different VLMs share common 3D variation vulnerability regions? Take-away: \"Substantially. Our analysis reveals that Adv-3DT samples demonstrate strong cross-model transferability across various VLMs, indicating shared vulnerability regions in the 3D variation space.\" As demonstrated in Tab. 2, we investigate transferability by utilizing OpenCLIP and BLIP as surrogate models to optimize \\( p^*(\\Theta) \\), then evaluate six" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7899" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 3. Performance Degradation (↓) of VLMs on Image Captioning and VQA Tasks. For closed-source VLMs (GPT-4o/4o-mini), we performed transfer-based attacks, using OpenAI-CLIP (ViT-L@336px) as surrogate encoder to optimize \\( p^*(\\Theta) \\). For other VLMs, \\( p^*(\\Theta) \\) are optimized w.r.t. respective image encoders (e.g., Eva-CLIP and OpenAI CLIP, etc.). Metrics: C-CIDEr, B-BLUE@4." + }, + { + "type": "table", + "bbox": [ + 0.146, + 0.139, + 0.855, + 0.424 + ], + "angle": 0, + "content": "
ModelMethodCOCO Caption [11]NoCaps [1]VQA-V2 [20]
CB@4GPT-ScoreCB@4GPT-ScoreGPT-Acc
BLIP-2 [30]Clean133.436.024.798.144.124.453.4
Random91.8 (↓42)23.5 (↓13)19.5 (↓5)60.2 (↓38)28.6 (↓16)20.9 (↓4)34.0 (↓19)
p*(Θ)83.8 (↓50)23.3 (↓13)18.9 (↓6)55.7 (↓42)26.9 (↓17)17.1 (↓7)31.7 (↓22)
Θ*72.2 (↓61)21.5 (↓15)17.0 (↓8)40.0 (↓58)22.4 (↓22)15.0 (↓9)28.0 (↓25)
LLaVa-1.5 [36]Clean126.130.425.0118.750.624.768.6
Random93.4 (↓33)25.6 (↓5)20.0 (↓5)75.3 (↓43)35.7 (↓15)21.7 (↓3)56.3 (↓12)
p*(Θ)85.8 (↓40)22.9 (↓8)18.9 (↓6)60.0 (↓59)29.6 (↓21)17.4 (↓7)54.3 (↓14)
Θ*73.2 (↓53)20.5 (↓10)16.5 (↓9)45.8 (↓73)25.5 (↓25)15.6 (↓9)53.8 (↓15)
LLaVa-1.6 [35]Clean41.66.922.930.510.822.671.2
Random25.1 (↓17)6.0 (↓1)17.5 (↓5)18.3 (↓12)7.1 (↓4)21.0 (↓2)54.3 (↓17)
p*(Θ)21.9 (↓20)5.3 (↓2)16.9 (↓6)13.6 (↓17)6.5 (↓4)15.7 (↓7)53.0 (↓18)
Θ*16.7 (↓25)4.6 (↓2)15.4 (↓8)10.5 (↓20)4.3 (↓7)13.5 (↓9)46.9 (↓24)
Otter [28]Clean127.034.924.076.429.722.553.9
Random71.6 (↓55)22.5 (↓12)21.0 (↓3)55.4 (↓21)23.7 (↓6)21.1 (↓1)49.7 (↓4)
p*(Θ)67.9 (↓59)17.8 (↓17)17.2 (↓7)45.4 (↓31)19.2 (↓11)16.2 (↓6)44.9 (↓9)
Θ*58.9 (↓68)17.5 (↓17)15.8 (↓8)36.0 (↓40)15.7 (↓14)14.6 (↓8)42.0 (↓12)
MiniGPT-4 [72]Clean78.523.424.060.828.524.751.3
Random58.2 (↓20)17.9 (↓6)19.7 (↓4)38.0 (↓23)19.8 (↓9)21.9 (↓3)43.7 (↓8)
p*(Θ)52.3 (↓26)16.3 (↓7)18.1 (↓6)32.9 (↓28)18.4 (↓10)16.7 (↓8)41.0 (↓10)
Θ*42.2 (↓36)12.9 (↓11)16.2 (↓8)26.6 (↓34)14.9 (↓14)15.0 (↓10)34.3 (↓17)
GPT-4o-mini [45]Clean23.73.023.315.72.923.068.6
Random14.1 (↓10)1.8 (↓1)16.9 (↓6)10.8 (↓5)0.9 (↓2)20.0 (↓3)51.7 (↓17)
p*(Θ)12.5 (↓11)1.2 (↓2)18.0 (↓5)7.8 (↓8)0.6 (↓2)17.2 (↓6)43.4 (↓25)
Θ*9.6 (↓14)1.6 (↓1)16.3 (↓7)7.3 (↓8)0.4 (↓3)15.5 (↓8)42.7 (↓26)
GPT-4o [45]Clean51.29.725.040.711.724.975.4
Random36.5 (↓15)5.1 (↓5)19.1 (↓6)24.7 (↓16)9.5 (↓2)20.9 (↓4)56.3 (↓19)
p*(Θ)26.4 (↓25)4.6 (↓5)17.7 (↓7)19.1 (↓22)6.5 (↓5)17.0 (↓8)50.7 (↓25)
Θ*20.3 (↓31)4.7 (↓5)16.1 (↓9)15.8 (↓25)5.4 (↓6)14.6 (↓10)50.3 (↓25)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.435, + 0.483, + 0.464 + ], + "angle": 0, + "content": "Table 4. Top-1 Accuracy and Naturalness Score of Adv-3DT samples generated by AdvDreamer w/ and w/o NRM. (under \\( p^{*}(\\Theta) \\))" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.467, + 0.466, + 0.615 + ], + "angle": 0, + "content": "
MethodsTop-1 Acc.ScoreRScoreP
AdvDreamer w/o. NRM48.6% (↓49.4)1.601.39
AdvDreamer w/. NRM54.0% (↓44.0)2.522.57
" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.62, + 0.476, + 0.635 + ], + "angle": 0, + "content": "Figure 5. Visualization of Adv-3DT samples w/ and w/o NRM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.868 + ], + "angle": 0, + "content": "target VLMs with diverse architectures and training objectives. The transfer attacks maintain strong effectiveness, with performance degradation comparable to direct attacks. Notably, \\( p^*(\\Theta) \\) exhibits smaller transfer gaps than \\( \\Theta^* \\), suggesting that learning the underlying distribution can better capture generalizable vulnerability regions. This transferability extends consistently to captioning and VQA tasks (Tab. 3). When employing OpenAI-CLIP as the surrogate encoder, Adv-3DT samples generated by AdvDreamer effectively compromise commercial systems like GPT-4o. We attribute this universal transferability to a fundamental limitation: current image-text pretraining datasets exhibit significant 3D biases and fail to encompass diverse certain special 3D transformations, leading to universal performance deficiencies under these transformation regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "RQ3: Does the performance degradation under Adv-3DT samples generated by AdvDreamer fundamentally" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.74 + ], + "angle": 0, + "content": "arise from 3D variations? Take-away: \"Yes, we successfully reproduce these Adv-3DT samples in physical environments and demonstrate that degradation is inherently caused by the 3D variation rather than undesirable image quality problems.\" With the naturalness feedback provided by NRM, AdvDreamer effectively prevents unwanted degradation of samples during the optimization process. To verify this, we evaluate Adv-3DT samples generated by AdvDreamer with and without NRM under ImageNet. Tab. 4 presents the accuracy under OpenCLIP (ViT-B/16), alongside GPT-4o-assessed visual realism \\((Score_R)\\) and physical plausibility \\((Score_P)\\). The results indicate that NRM significantly improves samples' realism and physical feasibility, with a marginal trade-off in aggressiveness. Furthermore, Fig. 5 presents a qualitative comparison, illustrating the effectiveness of NRM's naturality feedback in improving the image quality and mitigating the distribution discrepancy compared to natural image distributions. In addition, we validate the alignment between NRM predictions and human evaluations, as detailed in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, we conduct physical-world experiments on 12 common objects from traffic and household environments to validate this. Our experimental procedure consisted of three main steps: 1) capturing images of objects in their natural state, 2) using AdvDreamer with OpenCLIP as the surrogate model to generate Adv-3DT samples, and 3) physically reproducing the adversarial 3D transformations captured by AdvDreamer through video recording. This process yielded 3,807 frames of physical Adv-3DT samples. We evaluate effectiveness through zero-shot classifi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7900" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.34, + 0.09, + 0.658, + 0.105 + ], + "angle": 0, + "content": "Table 5. Quantitative results of physical experiments." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.111, + 0.817, + 0.178 + ], + "angle": 0, + "content": "
SampleOpenCLIPLLaVa-1.5GPT-4o
Acc.Conf.Acc.1Acc.2Acc.1Acc.2
Natural (Physical)100.00.89283.310091.7100.0
Adv-3DT (AdvDreamer)0.0 (↓100)0.00325.0 (↓58)25.0 (↓75)25.0 (↓67)41.7 (↓58)
Adv-3DT (Physical)51.3 (↓49)0.41333.6 (↓50)45.1 (↓55)56.8 (↓35)64.8 (↓35)
" + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.195, + 0.285, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.194, + 0.465, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.298, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Figure 6. Attack success rate of Adv-3DT samples w.r.t. (A) optimization steps and (B) scaling ratios of the optimal distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.346, + 0.483, + 0.558 + ], + "angle": 0, + "content": "cation and VQA tasks. For VQA, we designed two evaluation protocols: multiple-choice classification (Acc.1) and binary verification (Acc.2). As shown in Table 5, physically reproduced Adv-3DT samples significantly degraded VLM performance: the accuracy of zero-shot classification dropped to \\(51.3\\%\\), while VQA performance decreased to \\(33.6\\%\\) (Acc.1) and \\(45.1\\%\\) (Acc.2). These results confirm that AdvDreamer successfully captures physically realizable 3D variations that represent real world corner cases. However, an aggressiveness gap between digital and physical attacks is observed, primarily attributed to data distribution differences and camera parameter fluctuations. Fig. 1 (B) presents representative examples from physical experiments, with more detailed results provided in Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.573, + 0.445, + 0.589 + ], + "angle": 0, + "content": "4.3. Ablation Studies and Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Convergence Analysis & Computational Efficiency. We analyze the attack success rates (ASR) of \\( p^*(\\Theta) \\) against OpenCLIP (ViT-B/16) and BLIP (ViT-B/16) across different optimization iterations. As shown in Fig. 6, AdvDreamer's effectiveness consistently improves with more iterations and converges after 15 steps. Thus, we set the iteration step default to 15 throughout our experiments to balance effectiveness and efficiency. The computational cost is discussed in Appendix C.4 by detailing the overhead of each component. For each clean sample, AdvDreamer requires an average of 0.28 GPU hours with a single RTX 3090." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Does Optimal Distribution \\( p^*(\\Theta) \\) Effectively Characterize the Worst-case 3D Variations? We conduct analysis on ImageNet using five random natural images. For each image, we generate 100 Adv-3DT samples by scaling the variance \\( \\Sigma \\) of their \\( p^*(\\Theta) \\) with different ratios. Fig. 6(B) shows that ASR consistently decreases as the sampling range expands beyond the optimal distribution. This inverse relationship confirms that AdvDreamer successfully identifies the most vulnerable regions in the 3D variation space." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.193, + 0.895, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Figure 7. Benchmarking result of VLMs' 3D variation robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.305, + 0.667, + 0.319 + ], + "angle": 0, + "content": "4.4. MM3DTBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.908, + 0.629 + ], + "angle": 0, + "content": "Finally, We introduce MM3DTBench, it comprises 215 challenging Adv-3DT samples, consisting of samples generated by AdvDreamer and their physical replications. For each sample, we provide question templates with candidate answers and precise orientation annotations for the transformed objects. We employ two evaluation tasks: multiple-choice (Choice) and free-form description (Free answer), with accuracy computed through text-only GPT-4 judgment. More Details are provided in the Appendix.E. Our evaluation of 13 representative VLMs reveals robustness gaps. As shown in Fig. 7, even top-performing models (GPT-4o [45], CogVLM [61], and Intern-VL [13]) achieve limited success, while most models struggle with accuracy below \\(50\\%\\). Notably, Claude-3 [5], despite its recognized excellence, only achieves an accuracy of \\(30.0\\%\\). This performance correlates with models' general out-of-distribution robustness measured by MultiTrust [67], suggesting a fundamental limitation of current VLMs. We encourage the adoption of MM3DTBench in VLM development to advance security evaluation in dynamic scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.634, + 0.657 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.907, + 0.818 + ], + "angle": 0, + "content": "We proposed AdvDreamer, a novel framework for characterizing real-world adversarial 3D transformation samples from single-view observations. Our comprehensive evaluation revealed critical robustness gaps in existing VLMs, highlighting the urgent need to enhance VLMs' 3D variation perception and understanding capabilities in safety-critical applications. Through MM3DTBench, we established the first benchmark for worst-case 3D variations, aiming to facilitate the development of more robust vision-language systems for real-world deployment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.673, + 0.848 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the Fundamental Research Funds for the Central Universities, and was also supported by NSFC Projects (U2341228, 62276149)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "7901" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Haider Al-Tahan, Quentin Garrido, Randall Balestriero, Diane Bouchacourt, Caner Hazirbas, and Mark Ibrahim. Unibench: Visual reasoning requires rethinking vision-language beyond scaling. arXiv preprint arXiv:2408.04810, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.483, + 0.351 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[4] Michael A Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4845-4854, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[5] Anthropic. claude-3. https://www.anthropic.com/claude, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.52 + ], + "angle": 0, + "content": "[6] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.578 + ], + "angle": 0, + "content": "[7] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[8] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[10] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level im" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.134 + ], + "angle": 0, + "content": "age customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[14] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[18] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. In Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[19] Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1487-1495, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.631, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[21] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. In ECCV, pages 22-38. Springer, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.904, + 0.772 + ], + "angle": 0, + "content": "[22] Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7902" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. openclip, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[27] LAION-AI. Clip-benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[28] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[37] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Jiahuan Long, Tingsong Jiang, Wen Yao, Shuai Jia, Weijia Zhang, Weien Zhou, Chao Ma, and Xiaoqian Chen. Papmot: Exploring adversarial patch attack against multiple ob" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ject tracking. In European Conference on Computer Vision, pages 128-144. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[39] Jiahuan Long, Zhengqin Xu, Tingsong Jiang, Wen Yao, Shuai Jia, Chao Ma, and Xiaoqian Chen. Robust sam: On the adversarial robustness of vision foundation models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5775-5783, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[40] Jiahuan Long, Wen Yao, Tingsong Jiang, and Chao Ma. Cdupatch: Color-driven universal adversarial patch attack for dual-modal visible-infrared detectors. arXiv preprint arXiv:2504.10888, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[41] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.312, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[42] Spandan Madan, Timothy Henry, Jamell Dozier, Helen Ho, Nishchal Bhandari, Tomotake Sasaki, Frédo Durand, Hanspeter Pfister, and Xavier Boix. When and how convolutional neural networks generalize to out-of-distribution category-viewpoint combinations. Nature Machine Intelligence, 4(2):146-153, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[43] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[45] openai. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.3,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[48] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[50] Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, and Xingxing Wei. Towards viewpoint-invariant visual recognition via adversarial training. In ICCV, pages 1-10, 2023. 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7903" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[51] Shouwei Ruan, Yinpeng Dong, Hang Su, Jianteng Peng, Ning Chen, and Xingxing Wei. Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[52] Shouwei Ruan, Yinpeng Dong, Hanqing Liu, Yao Huang, Hang Su, and Xingxing Wei. Omniview-tuning: Boosting viewpoint invariance of vision-language pre-training models. arXiv preprint arXiv:2404.12139, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[55] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[56] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[57] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[61] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[62] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In The Twelfth International Conference on Learning Representations, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[63] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[64] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[67] Yichi Zhang, Yao Huang, Yitong Sun, Chang Liu, Zhe Zhao, Zhengwei Fang, Yifan Wang, Huanran Chen, Xiao Yang, Xingxing Wei, Hang Su, Yinpeng Dong, and Jun Zhu. Benchmarking trustworthiness of multimodal large language models: A comprehensive study, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.904, + 0.415 + ], + "angle": 0, + "content": "[68] Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. Proceedings of the European conference on computer vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.5 + ], + "angle": 0, + "content": "[69] Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, and Adam Kortylewski. Ood-cv: A benchmark for robustness to out-of-distribution shifts of individual nuisances in natural images. In European conference on computer vision, pages 163-180. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.502, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[70] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[71] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[72] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7904" + } + ] +] \ No newline at end of file diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_origin.pdf b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..299018ea9de5fc7f2e43807493dc768217346929 --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/74239f2c-88ab-4f0c-965d-c8f65d948118_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e50d1258f8b624c53d214f213c40c30ffbfe691ddc1195f793832cb2a640b7d +size 4271206 diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/full.md b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d554eff0d7ffeb53d4c99dab42e2ec1c87bd34eb --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/full.md @@ -0,0 +1,378 @@ +# AdvDreamer Unveils: + +# Are Vision-Language Models Truly Ready for Real-World 3D Variations? + +Shouwei Ruan $^{1\dagger}$ , Hanqing Liu $^{1\dagger}$ , Yao Huang $^{1\dagger}$ , Xiaqi Wang $^{1}$ , Caixin Kang $^{1}$ , Hang Su $^{2}$ , Yinpeng Dong $^{3}$ , Xingxing Wei $^{1*}$ + +$^{1}$ Institute of Artificial Intelligence, + +State Key Laboratory of Virtual Reality Technology and Systems, Beihang University + +$^{2}$ Dept. of Comp. Sci. and Tech., Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab, BNRist Center, Tsinghua University + +3 College of AI, Tsinghua University + +{shouweiruan, hqliu, y_huang, xxwei}@buaa.edu.cn, suhangss@tsinghua.edu.cn, dongyinpeng@mail.tsinghua.edu.cn + +(A) Adversarial 3D Variations Samples form Single-view Natural Images —— (C) Background Variations for Adv-3DT samples + +![](images/2d4983076e76f69f2c01aefff157f73c37a9ec199c7865d471d4722f1c55056f.jpg) + +![](images/91513af6c0274a700ac71aea6cfc52a39e93dfec9cb07450b16365e25fe6be56.jpg) + +![](images/e744f7a72a21a6fb0987cd19d08e4cb0f2bafc9faf47370d3e067e3742e20343.jpg) + +(B) Physical Reproduction of Adversarial 3D Variations Samples + +![](images/c916d5b785362732141f6114aeb74e746360d29d2dcb1d6b5768dd8871de4679.jpg) +Figure 1. The 3D Variation Vulnerabilities in VLMs. (A) The proposed AdvDreamer allows the capture of adversarial 3D transformation (Adv-3DT) samples from single natural images, threatening various VLMs. (B) These Adv-3DT samples can be successfully reproduced in the physical world and (C) retain their aggressiveness under background variations. (D) We also explain the GPT-Score metric used in the experiments. In each VLM response, we highlight the correct, significant error, and less precise parts using different colors. + +![](images/e1357629bf1be364d02f6f37868cce6b703a631fcb3d42a2a4a261caf59cc593.jpg) + +![](images/7bd5121d7cfb2b332e5d7ea86ac4f2282435d98bb4d151e2bbf754144ee322b7.jpg) + +![](images/17f84b046911aa55a6bb4c55168ef72fc3f589bb2e69bf4776c697faedc92b96.jpg) + +![](images/ac1996e7d1a94690ce288b9d8e76acf48d3d5edae43fffdfe99953a22c0c4068.jpg) + +![](images/3cef88438797ec505262a305d763b93ec92d5512fd0eeec51beeaa8a6059bb26.jpg) + +# Abstract + +Vision Language Models (VLMs) have exhibited remarkable generalization capabilities, yet their robustness in dynamic real-world scenarios remains largely unexplored. To systematically evaluate VLMs' robustness to real-world 3D variations, we propose AdvDreamer, the first framework capable of generating physically reproducible Adversarial 3D Transformation (Adv-3DT) samples from single-view observations. In AdvDreamer, we integrate three key innovations: Firstly, to characterize real-world 3D variations with limited prior knowledge precisely, we design a zero-shot Monocular Pose Manipulation pipeline built upon generative 3D priors. Secondly, to ensure the visual qual + +ity of worst-case Adv-3DT samples, we propose a Naturalness Reward Model that provides continuous naturalness regularization during adversarial optimization, effectively preventing convergence to hallucinated or unnatural elements. Thirdly, to enable systematic evaluation across diverse VLM architectures and visual-language tasks, we introduce the Inverse Semantic Probability loss as the adversarial optimization objective, which solely operates in the fundamental visual-textual alignment space. Based on the captured Adv-3DT samples with high aggressiveness and transferability, we establish MM3DTBench, the first VQA benchmark dataset tailored to evaluate VLM robustness under challenging 3D variations. Extensive evaluations of representative VLMs with varying architectures reveal that real-world 3D variations can pose severe threats to model performance across various tasks. + +![](images/65f7a3ecc01c3954051e3bff120793d5a59b93dd7a949d9ec21e82e1c2cd882f.jpg) +Figure 2. Comparison of paradigms to characterize Adversarial 3D Transformation (Adv-3DT) samples. + +![](images/5a35f2f9f3b56dc9fc20c47dcb736c0b822dcff76a70a07edbfccaf520881b5a.jpg) + +![](images/e7f87b135794b238f091b2754fb9dbcb9580202b93b25ef6009a58b04e479628.jpg) + +![](images/013f93479f7280a02673e7bc8ccc94ae40796dd9cb8d75e30d2599568c88ff33.jpg) + +# 1. Introduction + +Recently, Vision-Language Models (VLMs) [7, 13, 35, 36] have demonstrated remarkable capabilities in bridging visual perception and natural language understanding. Through large-scale image-text pre-training followed by fine-grained instruction tuning [31, 36], these models effectively tackle a wide range of visual-centric tasks, including instruction-based image recognition [47, 55, 65], visual question answering [3, 31, 35, 36], and visual reasoning [9, 10]. Due to their strong generalizability and impressive zero-shot capability, VLMs are increasingly deployed in various safety-critical applications, particularly in autonomous driving [56, 63] and robotic systems [25, 64]. + +Despite extensive studies validating VLMs' resilience against various distribution shifts (e.g., style variations [2, 39, 40, 68], image corruptions [2], and adversarial perturbations [16, 38, 67, 70]), they primarily focus on 2D perturbations in the digital domain, overlooking a critical challenge in real-world deployment: the 3D variations. As VLMs are increasingly integrated into dynamic scenarios, their ability to handle 3D variation becomes crucial, raising a fundamental question: Have current VLMs attained sufficient robustness to the distribution shifts arising from ubiquitous 3D variations in the real world? + +To thoroughly explore this problem and offer comprehensive answers, we identify three fundamental challenges: + +(1) How to accurately characterize real-world 3D variations with limited prior knowledge? Existing methods exploring the 3D variation robustness heavily rely on explicit 3D structures [4, 21] or representations derived from dense multi-view observations [4, 21, 42]. However, in most practical scenarios, the available priors are severely limited (e.g., single-view observation only). This constraint necessitates developing a method that minimizes dependency on extensive scene priors. To meet this, we introduce a zero-shot Monocular Pose-Manipulation (MPM) pipeline based on generative representations. Specifically, MPM leverages the rich 3D priors embedded in pre-trained Large-Reconstructive Models (LRM) [23, 24] and Image Re-Composition Models [12, 66], enabling fine-grained 3D adjustments with only single-view images and specified 3D variation parameters. While this generative-based paradigm establishes a foundation for capturing worst-case 3D variations through optimizing 3D variation parameters, it still + +exhibits inherent limitations that may introduce undesirable image corruptions, thus presenting our second challenge: + +(2) How to capture the worst-case adversarial 3D transformation (Adv-3DT) samples while ensuring that the performance degradation is inherently caused by the 3D variation? Prior works utilizing synthetic 3D assets [4, 21, 42] frequently produce samples with noticeable texture discrepancies compared to real-world images. These limitations in samples' visual fidelity and contextual coherence highlight a critical concern: observed performance degradation under Adv-3DT samples may stem from undesirable corruption rather than the 3D variation itself. To address this, we incorporate naturalness reward constraints within the adversarial optimization process. Specifically, we treat rigid 3D variations as an adversarial attack and propose the Naturalness Reward Model (NRM) that preserves image quality throughout optimization. NRM leverages visual context features provided by the DINOv2 [46] backbone to predict a naturalness reward that jointly reflects the visual fidelity and physical plausibility of the samples, effectively regularizing the optimization trajectory and preventing convergence toward unnatural or hallucinated elements. + +(3) How to ensure effective generalization of Adv-3DT samples across diverse VLM tasks and architectures? Existing studies [4, 18, 50] typically rely on task-specific objectives with end-to-end optimization, substantially limiting their applicability across diverse VLM architectures and a wide range of visual language tasks. To transcend this limitation, we formulate the adversarial objective in a task-and architecture-agnostic manner. Specifically, we introduce the Inverse Semantic Probability (ISP) loss, which operates solely through the foundational component of the VLM—the visual encoder, aiming to minimize the image-text matching probability between Adv-3DT samples and their corresponding ground-truth textual descriptions. Extensive experiments demonstrate that Adv-3DT samples optimized under the ISP loss exhibit remarkable adversarial transferability across various downstream tasks and models. Building on this, we further introduce the MM3DTBench, the first VQA benchmark dataset specifically designed to evaluate the 3D variation robustness of VLMs in real-world scenarios, comprising highly challenging Adv-3DT samples with carefully crafted candidate answers. + +Integrating the aforementioned innovations, this paper + +proposes AdvDreamer, a novel framework that enables the characterization of Adv-3DT samples from monocular observations, with broad adaptability across various visual-language tasks and VLM architectures. Our extensive digital and physical experiments reveal significant vulnerabilities in contemporary VLMs when confronted with challenging 3D variations. In visual recognition tasks, foundational models like OpenCLIP [26] and BLIP-2 [31] suffer a performance degradation of $65\% \sim 80\%$ on Adv-3DT samples. More concerning still, in open-ended tasks including image captioning and visual question answering, even the most advanced commercial models like GPT-4o [45] exhibit performance declines approaching $50\%$ . Our findings highlight the urgent need to enhance VLM robustness against 3D variations and set new requirements for deploying VLMs in complex, dynamic real-world applications. + +# 2. Related Work + +# 2.1. Vision-Language Models + +Classic Vision-Language Models (VLMs) [30, 32, 41, 47] achieve cross-modal alignment through diverse pre-training objectives on large-scale image-text pairs [33, 53]. Recent advances in Large Language Models (LLMs) [59, 60] have inspired new VLM architectures that integrate LLM capabilities via MLP projection layers [35, 36] or specialized modules like Q-Former [31]. Through techniques like instruction tuning [36], current VLMs [13, 35, 36] demonstrate enhanced performance across vision-centric tasks, including visual grounding, question answering, and image captioning, etc. These models also enable intelligent systems for task planning and decision-making [25, 64]. + +Despite their impressive capabilities [37, 47, 67], concerns remain about the robustness and generalization of VLMs, especially in safety-critical environments. Recent studies [52, 58] highlight that VLMs still exhibit deficiencies in processing visual details, such as orientation, quantity, and colour, which may be attributed to their visual representations. Moreover, VLMs remain susceptible to $L_{p}$ -norm adversarial perturbations [16, 43], with some efforts aiming to enhance their robustness through adversarial finetuning techniques [43]. In contrast to previous works, our study specifically investigates VLM robustness against 3D variations, addressing a crucial gap in understanding their reliability for dynamic real-world applications. + +# 2.2. Evaluation for 3D Variation Robustness + +Robustness to 3D variations remains a longstanding challenge in computer vision. Evaluating this property often involves capturing worst-case 3D transformations, i.e. the adversarial poses or viewpoints, to create challenging samples for assessment. Prior research can be categorized into three paradigms (see Fig.2): (A) Human-guided methods, like + +ObjectNet [8] and OOD-CV [69], manually collect samples with diverse poses and viewpoints, but they lack systematic coverage of worst-case examples and are costly. (B) Simulator-based [4, 21, 42] methods optimize worst-case 3D parameters via differentiable rendering of synthetic objects but require manually designed 3D assets and produce samples with limited realism. (C) Multi-view-based methods, like ViewFool [18] and GMVFool [50, 51], use NeRF-based [44] representations to optimize adversarial viewpoints, but they require dense observations and struggle with background integration. In contrast, the proposed AdvDreamer employs a (D) Generative-based approach, utilizing robust 3D priors from generative models. Given a single natural image, AdvDreamer generate realistic Adv-3DT samples in a zero-shot manner, providing a more efficient solution for evaluating 3D variation robustness. + +# 3. Methodology + +The overview of AdvDreamer is shown in Fig. 3. It consists $\bullet$ Zero-shot Monocular Pose Manipulation pipeline, leveraging generative 3D representations to obtain specified 3D variation samples from single-view images; $\bullet$ Naturalness Reward Model, providing naturalness regularization during adversarial optimization; and $\bullet$ Inverse Semantic Probability loss, ensuring the Adv-3DT samples effectively threaten VLMs and generalize across different architectures and tasks. The following sections will firstly formalize the optimization problem of AdvDreamer (Sec.3.1), then sequentially describe the key components (Sec.3.2), and finally outline the full optimization algorithm (Sec.3.3). + +# 3.1. Problem Formulation + +Parametrization of Real-World 3D Variations. In this paper, we primarily focus on rigid 3D transformations of foreground objects, i.e., the rotation, translation, and scaling, since they reflect the most typical 3D changes observed in real-world environments. Formally, we define a 6-dimensional vector $\Theta = \{\alpha, \beta, \gamma, \Delta_x, \Delta_y, s\}$ to uniquely parameterize any arbitrary transformation, where $\{\alpha, \beta, \gamma\}$ denote the Tait-Bryan angles (yaw, pitch, roll) in the $Z-Y-X$ sequence, $\{\Delta_x, \Delta_y\}$ represent the translation along the $x$ and $y$ axes in the $xy$ -plane, and $s$ is the uniform scaling factor. To prevent the loss of critical visual cues and ensure human recognizability, we follow previous work [18] to constrain $\Theta$ within a bounded range $[\Theta \min, \Theta \max]$ . + +Optimization Problem. The objective of AdvDreamer is to find an optimal distribution $p^*(\Theta)$ , ensuring that Adv-3DT sampled from this distribution are both aggressive against VLMs and maintaining visual quality, which can be abstractly formulated as the following optimization problem: + +$$ +p ^ {*} (\Theta) = \arg \max _ {p (\Theta)} \mathbb {E} _ {\Theta \sim p (\Theta)} \left[ \mathcal {L} _ {I S P} \left(X ^ {\prime}\right) + \mathcal {L} _ {N a t} \left(X ^ {\prime}\right) \right], \tag {1} +$$ + +$$ +w h e r e X ^ {\prime} = \mathcal {T} (\Theta , X), +$$ + +![](images/225896af9b58da7824ec8152570a7f4886dc366d21724adb3066679c1bfaca94.jpg) +Figure 3. Overview of the proposed AdvDreamer framework. To capture worst-case 3D variations in the real world, we treat rigid 3D transformations as adversarial attacks, optimizing the distribution of transformation parameters through a query-based approach. In AdvDreamer, we introduce the Monocular Pose Manipulation pipeline to perform specified 3D transformations on single-view images and guide the optimization process using the proposed Naturalness Reward Model and Inverse Semantic Probability loss. + +where $\mathcal{T}(\Theta, X)$ is a transformation function to produce a novel image with the specified transformation $\Theta$ applied to the original image $X \in \mathbb{R}^3$ . To implement $\mathcal{T}(\cdot)$ , we design the MPM pipeline, which will be detailed in Sec.3.2.1. The $\mathcal{L}_{Nat}$ and $\mathcal{L}_{ISP}$ measure the naturalness and adversarial efficacy of the transformed samples, respectively. These will be further elaborated in Sec.3.2.2 and Sec.3.2.3. + +Optimizing over the distribution $p(\Theta)$ rather than a deterministic $\Theta$ offers two key advantages. 1) As highlighted in [18, 50], learning the underlying distribution enables comprehensive exploration of the 3D variation space, facilitating the identification of the model's vulnerable regions. 2) Given the stochastic nature of the generation, optimizing over a continuous distribution can mitigate the impact of appearance inconsistency, as appearance variations induced by the generation are unlikely to consistently deceive the model without optimizing the adversarial 3D variations [6]. + +# 3.2. AdvDreamer Framework + +# 3.2.1. Zero-shot Monocular Pose Manipulation + +To implement the transformation function $\mathcal{T}$ , previous approaches typically rely on explicit 3D structures [4, 21, 42] or neural radiance representations [18, 50]. However, these approaches cannot accommodate real-world scenarios with limited prior knowledge. To address this, we design a simple yet effective zero-shot monocular pose manipulation (MPM) pipeline that consists of the following process: + +1 Foreground-Background Pairs Preparation. MPM first decompose the input image $X$ into foreground-background pairs $\{X_{f},X_{b}\}$ . Specifically, we leverage Grounded-SAM [48] to semantically segment the major instance as foreground $X_{f}$ , followed by diffusion-based inpainting [15, 49] to obtain a complete background $X_{b}$ . This process is denoted as $\{X_f,X_b\} = \mathcal{F}(X)$ . Additionally, we support directly generating synthetic $\{X_f,X_b\}$ pairs + +through Stable-Diffusion [49] to enhance sample diversity. + +$\Theta$ Object-level Pose Manipulation. For each optimization iteration, we transform the foreground $X_{f}$ into a batch of $X_{f}^{\prime}$ under $\Theta$ sampled from current $p(\Theta)$ . In this stage, MPM leverages the robust priors provided by a pre-trained Large Reconstruction Model (e.g., TripoSR [57]) to construct single-view-based 3D representations and apply sampled transformations. This process is formulated as $X_{f}^{\prime} = \mathcal{R}_{\mathbf{w}_{0}}(X_{f},\Theta)$ , where $\Theta \sim p(\Theta)$ and $\mathcal{R}_{\mathbf{w}}(\cdot)$ denotes neural rendering function with weights $\mathbf{w}_0$ . + +$\Theta$ Re-Composition. We then compose the transformed foreground $X_{f}^{\prime}$ with background $X_{b}$ through a pre-trained diffusion-based composition model [12, 66], to handles shadows and occlusions. This process is defined as $X^{\prime} = \mathcal{C}_{\mathbf{w}_1}(X_f',X_b)$ , where $\mathbf{w}_1$ denotes diffusion model's weights. + +Combining the above stages, the transformation process by MPM can be formally represented as follows: + +$$ +\begin{array}{r l} X ^ {\prime} = \mathcal {T} (\boldsymbol {\Theta}, X) & = \mathcal {C} _ {\mathbf {w} _ {1}} \left(\mathcal {R} _ {\mathbf {w} _ {0}} \left(X _ {f}, \boldsymbol {\Theta}\right), X _ {b}\right), \\ \text {w h e r e} \{X _ {f}, X _ {b} \} & = \mathcal {F} (X). \end{array} \tag {2} +$$ + +# 3.2.2. Naturalness Reward Model + +Despite the MPM pipeline utilizing generative priors to achieve realistic 3D variation samples, we observe that directly applying adversarial optimization frequently leads to unnatural artifacts, such as shape distortions and physically implausible cases in Fig. 5. These issues stem from two factors: generalization limitations of existing generative models and the inherent high semantic bias in low-quality samples. To address this challenge, we propose a novel Naturalness Reward Model (NRM) that continuously regularizes $p(\Theta)$ during optimization based on sample naturalness, preventing convergence to low-quality pseudo-optimal regions. + +As shown in Fig.4, the NRM employs DINOv2[46] as its backbone, chosen for several reasons. Firstly, DINO + +![](images/fbf19ad0829fae426c631676671a2ca58d284ef7c89a9d396eecf8e158cc5fbe.jpg) +Figure 4. Architecture of NRM and computational cost and accuracy of naturalness assessment across Human, GPT, and NRM. + +captures finer visual contextual features compared to traditional pre-trained vision encoders [58], aiding better convergence during training. Secondly, unlike vision-language aligned encoders such as CLIP, DINO's visual features are language-agnostic, making them more faithfully reflect image quality rather than semantic consistency. We use DINOv2 to extract image tokens from the input Adv-3DT samples, followed by dual-stream prediction heads to estimate visual fidelity and physical plausibility scores: + +$$ +\operatorname {S c o r e} _ {R} \left(X ^ {\prime}\right) = f _ {R} \left(\mathcal {E} _ {D} \left(X ^ {\prime}\right)\right), \operatorname {S c o r e} _ {P} \left(X ^ {\prime}\right) = f _ {P} \left(\mathcal {E} _ {D} \left(X ^ {\prime}\right)\right), \tag {3} +$$ + +where $f_{R}(\cdot)$ and $f_{P}(\cdot)$ are two-layer MLPs. To train the NRM, we construct a large-scale naturalness preference dataset with fine-grained scoring. We generate numerous samples from the MPM pipeline on ImageNet and synthetic images, scoring each on a 5-point scale for visual fidelity and physical plausibility. These scores are initially annotated automatically by GPT-4o with Chain-of-Thought prompts and verified by volunteers for correction. This process resulted in 120K high-quality training samples. + +The architecture and training parameters of NRM are detailed in Appendix B. Our experimental results demonstrate that the naturalness feedback provided by NRM highly aligns with human evaluation. In terms of prediction efficiency and stability, it outperforms methods that directly employ GPT-4o for naturalness assessment (see Fig. 4 and Appendix C.2). Based on the NRM's predictions, we define the naturalness regularization loss for the optimization as: + +$$ +\mathcal {L} _ {\text {N a t}} \left(X ^ {\prime}\right) = - \frac {1}{2} \left(\operatorname {S c o r e} _ {R} \left(X ^ {\prime}\right) + \operatorname {S c o r e} _ {P} \left(X ^ {\prime}\right)\right) \tag {4} +$$ + +# 3.2.3. Inverse Semantic Probability Loss + +Previous studies typically employ task-specific optimization objectives with end-to-end manner to obtain worst-case Adv-3DT samples. However, directly applying this approach to VLMs with substantial parameters and diverse architectures spanning various tasks is challenging. To address this, we design the Inverse Semantic Probability (ISP) loss, which minimizes the probability that VLMs assign to the correct semantic attributes of Adv-3DT samples. + +This design have two key considerations. 1) It is based solely on the visual and text encoder $\mathcal{E}_v$ , $\mathcal{E}_t$ , which are fundamental components across modern VLMs, ensuring architecture-agnostic applicability. 2) By operating in the + +image-text alignment space rather than model-specific layers or task-specific heads, it enables task generalization. + +Given a generated sample $X'$ and a semantic label set $\mathcal{Y} = \{y_i\}_{i=1}^N$ containing the ground-truth label $y_t$ , we first compute a semantic similarity vector $\mathbf{S}$ : + +$$ +\mathbf {S} = \left\{s _ {i} \right\} _ {i = 1} ^ {N}, \text {w h e r e} s _ {i} = \cos \left[ \mathcal {E} _ {v} \left(X ^ {\prime}\right), \mathcal {E} _ {t} \left(\mathcal {P} \left(y _ {i}\right)\right) \right], \tag {5} +$$ + +where $\cos (\cdot)$ computes cosine similarity and $\mathcal{P}(\cdot)$ follows the template "a photo of a $\{y_i\}$ " in [47]. The vector $\mathbf{S}$ captures the representation similarities between $X'$ and each semantic label. We then derive the conditional semantic probability through softmax transformation: + +$$ +p \left(y _ {i} \mid X ^ {\prime}\right) = \exp \left(s _ {i}\right) / \sum_ {j = 1} ^ {N} \exp \left(s _ {j}\right). \tag {6} +$$ + +Then, the $\mathcal{L}_{\mathrm{ISP}}$ is defined as the negative log-likelihood of the ground-truth's semantic probability: + +$$ +\mathcal {L} _ {\mathrm {I S P}} \left(X ^ {\prime}, \mathcal {Y}\right) = - \log p \left(y _ {t} \mid X ^ {\prime}\right). \tag {7} +$$ + +Intuitively, $\mathcal{L}_{\mathrm{ISP}}$ quantifies the likelihood of $X^{\prime}$ being misclassified by comparing its similarity to the ground-truth and irrelevant categories. Maximizing it directs the $p(\Theta)$ towards regions that effectively deceive the model. + +# 3.3. Query-Based Adversarial Optimization + +To solve Problem (1), we parameterize the adversarial distribution $p(\Theta)$ as a Multivariate Gaussian form: + +$$ +\boldsymbol {\Theta} = \mathbf {A} \cdot \tanh (\mathbf {z}) + \mathbf {B}, \text {w h e r e} \mathbf {z} \sim \mathcal {N} (\boldsymbol {\mu}, \boldsymbol {\Sigma}), \tag {8} +$$ + +where $\mathbf{A} = (\Theta_{\mathrm{min}} - \Theta_{\mathrm{max}}) / 2$ , $\mathbf{A} = (\Theta_{\mathrm{min}} + \Theta_{\mathrm{max}}) / 2$ , $\mathbf{z}$ follows a Gaussian Distribution with mean $\pmb{\mu} \in \mathbb{R}^6$ and covariance $\pmb{\Sigma} \in \mathbb{R}^{6 \times 6}$ . Thus, we rewrite (1) in a specific form: + +$$ +\underset {\boldsymbol {\mu}, \boldsymbol {\Sigma}} {\arg \max } \mathbb {E} _ {\boldsymbol {\mathbf {z}} \sim \mathcal {N} (\boldsymbol {\mu}, \boldsymbol {\Sigma})} \left[ \mathcal {L} _ {\text {I S P}} \left(X ^ {\prime}, \mathcal {Y}\right) + \mathcal {L} _ {\text {N a t}} \left(X ^ {\prime}\right) \right], \tag {9} +$$ + +$$ +w h e r e X ^ {\prime} = \mathcal {T} (\mathbf {A} \cdot \tanh (\mathbf {z}) + \mathbf {B}, X). +$$ + +Solving this problem typically involves computing the gradient of the loss w.r.t. $\pmb{\mu}$ , $\pmb{\Sigma}$ and updating them via gradient ascent. However, the forward steps involve multiple components, e.g., the denoise process, introducing uncertainty in the gradient propagation path, making gradient-based optimization challenging [54]. Therefore, we adopt the Covariance Matrix Adaptation Evolution Strategy [19, 22], an efficient query-based black-box optimizer. At each iteration $t$ , we: 1) Sample $K$ candidates $\{\mathbf{z}_i^{t + 1}\}_{i = 1}^K$ from $\mathcal{N}(\boldsymbol {\mu}^t,\boldsymbol {\Sigma}^t)$ . 2) Generate corresponding samples $\{(X')_{i}^{t + 1}\}_{i = 1}^{K}$ . 3) Update distribution parameters $\pmb{\mu}^t,\pmb{\Sigma}^t$ as follows: + +$$ +\boldsymbol {\mu} ^ {t + 1} = \sum_ {i = 1} ^ {k} w _ {i} \cdot \mathbf {z} _ {(i: k)} ^ {t + 1}, \quad \text {w h e r e} \sum_ {i = 1} ^ {k} w _ {i} = 1, \tag {10} +$$ + +$$ +\boldsymbol {\Sigma} ^ {t + 1} = \left(1 - \eta_ {1} - \eta_ {\mu}\right) \cdot \boldsymbol {\Sigma} ^ {t} + \eta_ {1} \cdot p _ {\Sigma} ^ {t + 1} \left(p _ {\Sigma} ^ {t + 1}\right) ^ {T} + +$$ + +$$ +\eta_ {\mu} \cdot \sum_ {i = 1} ^ {h} w _ {i} \cdot \left(\frac {\mathbf {z} _ {(i : k)} ^ {t + 1} - \boldsymbol {\mu} ^ {t}}{\boldsymbol {\sigma} ^ {t}}\right) \left(\frac {\mathbf {z} _ {(i : k)} ^ {t + 1} - \boldsymbol {\mu} ^ {t}}{\boldsymbol {\sigma} ^ {t}}\right) ^ {T}, \tag {11} +$$ + +Table 1. Zero-shot Classification Accuracy $(\%)$ and degradation $(\downarrow \%)$ Under Adv-3DT Samples Generated from ImageNet [17]. + +
Target Models#ParamsMethods
CleanRandomp*(Θ)Θ*
OpenCLIP ViT-B/16 [26]149M98.062.6 (↓35)54.0 (↓44)18.0 (↓80)
OpenCLIP ViT-L/14 [26]428M94.461.7 (↓33)50.9 (↓44)15.3 (↓79)
OpenCLIP ViT-G/14 [26]2.5B96.463.5 (↓33)53.5 (↓43)18.7 (↓78)
BLIP ViT-B/16 [30]583M83.056.0 (↓27)51.3 (↓32)17.3 (↓66)
BLIP-2 ViT-L/14 [31]3.4B84.057.0 (↓27)52.2 (↓32)18.7 (↓65)
BLIP-2 ViT-G/14 [31]4.1B81.055.7 (↓25)49.1 (↓32)15.7 (↓65)
+ +here, $\mathbf{z}_{(i:k)}$ denotes samples ranked by losses: we first select top-10 samples based on $\mathcal{L}_{\mathrm{ISP}}$ , then choose the top-k $(k = 5)$ with highest $\mathcal{L}_{\mathrm{Nat}}$ for updates. $w_{i} = 1 / k$ are importance weights, $\eta_{1}$ and $\eta_{\mu}$ are learning rates, $p_{\Sigma}$ tracks the evolution path, and $\sigma^t$ is the step size updated via Cumulative Step-size Adaptation (CSA) [22]. Due to space constraints, complete derivations of Eq. (10), (11) and hyperparameter settings are provided in the Appendix.A.1. We also provide the pseudocode for optimization process in Appendix A.2. + +# 4. Experiments + +# 4.1. Experimental Setup + +Tasks & Datasets. We evaluate VLMs' robustness to 3D variations using AdvDreamer on three representative vision-language tasks. (1) For zero-shot image classification, we utilize the ImageNet test set [17]. (2) For image captioning, we employ the test splits from COCO Caption [11] and NoCaps [1] datasets. (3) For Visual Question Answering (VQA), we use the VQAv2 [20] test set. For each dataset, we select 300 clean samples spanning 30 categories with clear and well-defined object instances. During the optimization process, we employ ImageNet-1K categories as the semantic label set $\mathcal{Y}$ for the classification task and use COCO-80 categories for the captioning and VQA tasks. + +Victim Models. For zero-shot classification, we target vision-language foundation models: OpenCLIP [14, 26], BLIPs [30, 31], and additional models for transfer attacks. For other tasks, we select mainstream open-source VLMs (LLaVa [34, 36], MiniGPT-4 [72], etc.) as well as representative commercial VLMs (GPT-4o and GPT-4o-mini). + +Methods & Baselines. We use two variants with AdvDreamer: (1) $p^*(\Theta)$ : represents the average results over 10 Adv-3DT samples randomly sampled from the optimal adversarial distribution, while (2) $\Theta^*$ : indicates the evaluation results using the Adv-3DT sample corresponding to the distribution center. Additionally, we establish two potential baselines following [18, 50]: the original natural images without transformation (Clean) and the average results over 10 images with random 3D transformations (Random). Metrics. For zero-shot classification, we follow the CLIP-Benchmark [27] protocol, reporting Top-1 accuracy and its degradation relative to clean samples. For captioning, in addition to conventional metrics (CIDEr, BLEU), we adopt the LLM-as-a-judge approach [36, 67, 71] to address the open + +Table 2. Cross-model transferability of Adv-3DT samples, utilizing OpenCLIP and BLIP as surrogate models. + +
Target ModelCleanOpenCLIP [26]BLIP [30]
p*(Θ)Θ*p*(Θ)Θ*
ALBEF [29]65.037.2 (↓28)27.7 (↓37)40.2 (↓25)25.7 (↓39)
OpenAI CLIP ViT-B/16 [47]85.348.7 (↓37)29.7 (↓56)52.8 (↓32)31.0 (↓54)
OpenCLIP ViT-B/16 [26]97.754.0 (↓44)18.0 (↓80)59.3 (↓38)34.7 (↓63)
Meta-CLIP ViT-B/16 [62]91.049.3 (↓42)30.0 (↓61)53.8 (↓37)29.7 (↓61)
BLIP ViT-B/16 [30]82.748.4 (↓34)30.7 (↓52)51.3 (↓31)17.3 (↓65)
BLIP-2 ViT-L/14 [31]84.050.0 (↓34)32.0 (↓52)52.3 (↓32)27.7 (↓56)
SigLIP ViT-B/16 [65]95.355.3 (↓40)34.0 (↓61)59.3 (↓36)34.7 (↓61)
+ +ended nature of VLM responses, aligning evaluations more closely with human preferences. We use GPT-4 to assess three aspects of the response on a 10-point scale: semantic accuracy, tone confidence, and overall coherence, with the scores summed to yield the GPT-Score (see Fig.1 (D)). For the VQA task, we introduce GPT-Acc, which leverages LLM-based judgment to evaluate response correctness. The judgment prompts are detailed in AppendixC. + +# 4.2. Research Questions (RQs) and Findings + +RQ1: Do current VLMs demonstrate sufficient robustness to 3D variations? Take-away: "Emphatically not. Our evaluation reveals the severe vulnerability of current VLMs to worst-case 3D variations across multiple vision-language tasks." As shown in Tab. 1, in zero-shot classification, Adv-3DT samples generated by AdvDreamer significantly compromise VLMs' performance, causing accuracy drops of up to $80\%$ . This substantially outperforms the Random baseline, validating AdvDreamer's effectiveness in capturing worst-case transformations. Notably, samples under $\Theta^{*}$ induce more severe performance degradation than those under the adversarial distribution $p^{*}(\Theta)$ , indicating the existence of extremely vulnerable points. + +This fundamental vulnerability extends beyond classification to more complex open-ended tasks. As illustrated in Tab.3, for image captioning, Adv-3DT samples under both $p^* (\Theta)$ and $\Theta^{*}$ effectively degrade model performance across all evaluation metrics, with GPT-Score decreasing by nearly $50\%$ across various VLMs. This significant degradation reveals a critical limitation in current VLMs' 3D understanding and reasoning capabilities. For VQA tasks, GPT4o, despite achieving the highest accuracy $(75.4\%)$ on clean samples—exhibits a substantial $25\%$ accuracy drop when under Adv-3DT attacks. Fig. 1 (A) provides visualizations of representative Adv-3DT samples generated by AdvDreamer, with more examples presented in Appendix.D. + +RQ2: To what extent do different VLMs share common 3D variation vulnerability regions? Take-away: "Substantially. Our analysis reveals that Adv-3DT samples demonstrate strong cross-model transferability across various VLMs, indicating shared vulnerability regions in the 3D variation space." As demonstrated in Tab. 2, we investigate transferability by utilizing OpenCLIP and BLIP as surrogate models to optimize $p^*(\Theta)$ , then evaluate six + +Table 3. Performance Degradation (↓) of VLMs on Image Captioning and VQA Tasks. For closed-source VLMs (GPT-4o/4o-mini), we performed transfer-based attacks, using OpenAI-CLIP (ViT-L@336px) as surrogate encoder to optimize $p^*(\Theta)$ . For other VLMs, $p^*(\Theta)$ are optimized w.r.t. respective image encoders (e.g., Eva-CLIP and OpenAI CLIP, etc.). Metrics: C-CIDEr, B-BLUE@4. + +
ModelMethodCOCO Caption [11]NoCaps [1]VQA-V2 [20]
CB@4GPT-ScoreCB@4GPT-ScoreGPT-Acc
BLIP-2 [30]Clean133.436.024.798.144.124.453.4
Random91.8 (↓42)23.5 (↓13)19.5 (↓5)60.2 (↓38)28.6 (↓16)20.9 (↓4)34.0 (↓19)
p*(Θ)83.8 (↓50)23.3 (↓13)18.9 (↓6)55.7 (↓42)26.9 (↓17)17.1 (↓7)31.7 (↓22)
Θ*72.2 (↓61)21.5 (↓15)17.0 (↓8)40.0 (↓58)22.4 (↓22)15.0 (↓9)28.0 (↓25)
LLaVa-1.5 [36]Clean126.130.425.0118.750.624.768.6
Random93.4 (↓33)25.6 (↓5)20.0 (↓5)75.3 (↓43)35.7 (↓15)21.7 (↓3)56.3 (↓12)
p*(Θ)85.8 (↓40)22.9 (↓8)18.9 (↓6)60.0 (↓59)29.6 (↓21)17.4 (↓7)54.3 (↓14)
Θ*73.2 (↓53)20.5 (↓10)16.5 (↓9)45.8 (↓73)25.5 (↓25)15.6 (↓9)53.8 (↓15)
LLaVa-1.6 [35]Clean41.66.922.930.510.822.671.2
Random25.1 (↓17)6.0 (↓1)17.5 (↓5)18.3 (↓12)7.1 (↓4)21.0 (↓2)54.3 (↓17)
p*(Θ)21.9 (↓20)5.3 (↓2)16.9 (↓6)13.6 (↓17)6.5 (↓4)15.7 (↓7)53.0 (↓18)
Θ*16.7 (↓25)4.6 (↓2)15.4 (↓8)10.5 (↓20)4.3 (↓7)13.5 (↓9)46.9 (↓24)
Otter [28]Clean127.034.924.076.429.722.553.9
Random71.6 (↓55)22.5 (↓12)21.0 (↓3)55.4 (↓21)23.7 (↓6)21.1 (↓1)49.7 (↓4)
p*(Θ)67.9 (↓59)17.8 (↓17)17.2 (↓7)45.4 (↓31)19.2 (↓11)16.2 (↓6)44.9 (↓9)
Θ*58.9 (↓68)17.5 (↓17)15.8 (↓8)36.0 (↓40)15.7 (↓14)14.6 (↓8)42.0 (↓12)
MiniGPT-4 [72]Clean78.523.424.060.828.524.751.3
Random58.2 (↓20)17.9 (↓6)19.7 (↓4)38.0 (↓23)19.8 (↓9)21.9 (↓3)43.7 (↓8)
p*(Θ)52.3 (↓26)16.3 (↓7)18.1 (↓6)32.9 (↓28)18.4 (↓10)16.7 (↓8)41.0 (↓10)
Θ*42.2 (↓36)12.9 (↓11)16.2 (↓8)26.6 (↓34)14.9 (↓14)15.0 (↓10)34.3 (↓17)
GPT-4o-mini [45]Clean23.73.023.315.72.923.068.6
Random14.1 (↓10)1.8 (↓1)16.9 (↓6)10.8 (↓5)0.9 (↓2)20.0 (↓3)51.7 (↓17)
p*(Θ)12.5 (↓11)1.2 (↓2)18.0 (↓5)7.8 (↓8)0.6 (↓2)17.2 (↓6)43.4 (↓25)
Θ*9.6 (↓14)1.6 (↓1)16.3 (↓7)7.3 (↓8)0.4 (↓3)15.5 (↓8)42.7 (↓26)
GPT-4o [45]Clean51.29.725.040.711.724.975.4
Random36.5 (↓15)5.1 (↓5)19.1 (↓6)24.7 (↓16)9.5 (↓2)20.9 (↓4)56.3 (↓19)
p*(Θ)26.4 (↓25)4.6 (↓5)17.7 (↓7)19.1 (↓22)6.5 (↓5)17.0 (↓8)50.7 (↓25)
Θ*20.3 (↓31)4.7 (↓5)16.1 (↓9)15.8 (↓25)5.4 (↓6)14.6 (↓10)50.3 (↓25)
+ +Table 4. Top-1 Accuracy and Naturalness Score of Adv-3DT samples generated by AdvDreamer w/ and w/o NRM. (under $p^{*}(\Theta)$ ) + +
MethodsTop-1 Acc.ScoreRScoreP
AdvDreamer w/o. NRM48.6% (↓49.4)1.601.39
AdvDreamer w/. NRM54.0% (↓44.0)2.522.57
+ +Figure 5. Visualization of Adv-3DT samples w/ and w/o NRM. + +target VLMs with diverse architectures and training objectives. The transfer attacks maintain strong effectiveness, with performance degradation comparable to direct attacks. Notably, $p^*(\Theta)$ exhibits smaller transfer gaps than $\Theta^*$ , suggesting that learning the underlying distribution can better capture generalizable vulnerability regions. This transferability extends consistently to captioning and VQA tasks (Tab. 3). When employing OpenAI-CLIP as the surrogate encoder, Adv-3DT samples generated by AdvDreamer effectively compromise commercial systems like GPT-4o. We attribute this universal transferability to a fundamental limitation: current image-text pretraining datasets exhibit significant 3D biases and fail to encompass diverse certain special 3D transformations, leading to universal performance deficiencies under these transformation regions. + +RQ3: Does the performance degradation under Adv-3DT samples generated by AdvDreamer fundamentally + +arise from 3D variations? Take-away: "Yes, we successfully reproduce these Adv-3DT samples in physical environments and demonstrate that degradation is inherently caused by the 3D variation rather than undesirable image quality problems." With the naturalness feedback provided by NRM, AdvDreamer effectively prevents unwanted degradation of samples during the optimization process. To verify this, we evaluate Adv-3DT samples generated by AdvDreamer with and without NRM under ImageNet. Tab. 4 presents the accuracy under OpenCLIP (ViT-B/16), alongside GPT-4o-assessed visual realism $(Score_R)$ and physical plausibility $(Score_P)$ . The results indicate that NRM significantly improves samples' realism and physical feasibility, with a marginal trade-off in aggressiveness. Furthermore, Fig. 5 presents a qualitative comparison, illustrating the effectiveness of NRM's naturality feedback in improving the image quality and mitigating the distribution discrepancy compared to natural image distributions. In addition, we validate the alignment between NRM predictions and human evaluations, as detailed in Appendix C.2. + +Furthermore, we conduct physical-world experiments on 12 common objects from traffic and household environments to validate this. Our experimental procedure consisted of three main steps: 1) capturing images of objects in their natural state, 2) using AdvDreamer with OpenCLIP as the surrogate model to generate Adv-3DT samples, and 3) physically reproducing the adversarial 3D transformations captured by AdvDreamer through video recording. This process yielded 3,807 frames of physical Adv-3DT samples. We evaluate effectiveness through zero-shot classifi + +Table 5. Quantitative results of physical experiments. + +
SampleOpenCLIPLLaVa-1.5GPT-4o
Acc.Conf.Acc.1Acc.2Acc.1Acc.2
Natural (Physical)100.00.89283.310091.7100.0
Adv-3DT (AdvDreamer)0.0 (↓100)0.00325.0 (↓58)25.0 (↓75)25.0 (↓67)41.7 (↓58)
Adv-3DT (Physical)51.3 (↓49)0.41333.6 (↓50)45.1 (↓55)56.8 (↓35)64.8 (↓35)
+ +![](images/f9a6bf5ad1d1e92827a403764db9ef89035254f643abf19fcf555156ba945ea5.jpg) +Figure 6. Attack success rate of Adv-3DT samples w.r.t. (A) optimization steps and (B) scaling ratios of the optimal distribution. + +![](images/823610e126a0141de068c3a144c0f16c5153d0b618c8811fc9b075a9a60c749a.jpg) + +cation and VQA tasks. For VQA, we designed two evaluation protocols: multiple-choice classification (Acc.1) and binary verification (Acc.2). As shown in Table 5, physically reproduced Adv-3DT samples significantly degraded VLM performance: the accuracy of zero-shot classification dropped to $51.3\%$ , while VQA performance decreased to $33.6\%$ (Acc.1) and $45.1\%$ (Acc.2). These results confirm that AdvDreamer successfully captures physically realizable 3D variations that represent real world corner cases. However, an aggressiveness gap between digital and physical attacks is observed, primarily attributed to data distribution differences and camera parameter fluctuations. Fig. 1 (B) presents representative examples from physical experiments, with more detailed results provided in Appendix D. + +# 4.3. Ablation Studies and Additional Analysis + +Convergence Analysis & Computational Efficiency. We analyze the attack success rates (ASR) of $p^*(\Theta)$ against OpenCLIP (ViT-B/16) and BLIP (ViT-B/16) across different optimization iterations. As shown in Fig. 6, AdvDreamer's effectiveness consistently improves with more iterations and converges after 15 steps. Thus, we set the iteration step default to 15 throughout our experiments to balance effectiveness and efficiency. The computational cost is discussed in Appendix C.4 by detailing the overhead of each component. For each clean sample, AdvDreamer requires an average of 0.28 GPU hours with a single RTX 3090. + +Does Optimal Distribution $p^*(\Theta)$ Effectively Characterize the Worst-case 3D Variations? We conduct analysis on ImageNet using five random natural images. For each image, we generate 100 Adv-3DT samples by scaling the variance $\Sigma$ of their $p^*(\Theta)$ with different ratios. Fig. 6(B) shows that ASR consistently decreases as the sampling range expands beyond the optimal distribution. This inverse relationship confirms that AdvDreamer successfully identifies the most vulnerable regions in the 3D variation space. + +![](images/1fc00b9da4be52c7ea4a64396ae900b607f1ac28a22d41b60ea2eecf73e39f57.jpg) +Figure 7. Benchmarking result of VLMs' 3D variation robustness. + +# 4.4. MM3DTBench + +Finally, We introduce MM3DTBench, it comprises 215 challenging Adv-3DT samples, consisting of samples generated by AdvDreamer and their physical replications. For each sample, we provide question templates with candidate answers and precise orientation annotations for the transformed objects. We employ two evaluation tasks: multiple-choice (Choice) and free-form description (Free answer), with accuracy computed through text-only GPT-4 judgment. More Details are provided in the Appendix.E. Our evaluation of 13 representative VLMs reveals robustness gaps. As shown in Fig. 7, even top-performing models (GPT-4o [45], CogVLM [61], and Intern-VL [13]) achieve limited success, while most models struggle with accuracy below $50\%$ . Notably, Claude-3 [5], despite its recognized excellence, only achieves an accuracy of $30.0\%$ . This performance correlates with models' general out-of-distribution robustness measured by MultiTrust [67], suggesting a fundamental limitation of current VLMs. We encourage the adoption of MM3DTBench in VLM development to advance security evaluation in dynamic scenarios. + +# 5. Conclusion + +We proposed AdvDreamer, a novel framework for characterizing real-world adversarial 3D transformation samples from single-view observations. Our comprehensive evaluation revealed critical robustness gaps in existing VLMs, highlighting the urgent need to enhance VLMs' 3D variation perception and understanding capabilities in safety-critical applications. Through MM3DTBench, we established the first benchmark for worst-case 3D variations, aiming to facilitate the development of more robust vision-language systems for real-world deployment. + +# Acknowledgement + +This work was supported by the Fundamental Research Funds for the Central Universities, and was also supported by NSFC Projects (U2341228, 62276149). + +# References + +[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 6, 7 +[2] Haider Al-Tahan, Quentin Garrido, Randall Balestriero, Diane Bouchacourt, Caner Hazirbas, and Mark Ibrahim. Unibench: Visual reasoning requires rethinking vision-language beyond scaling. arXiv preprint arXiv:2408.04810, 2024. 2 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[4] Michael A Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4845-4854, 2019. 2, 3, 4 +[5] Anthropic. claude-3. https://www.anthropic.com/claude, 2024.8 +[6] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 4 +[7] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2 +[8] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 3 +[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2 +[10] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. Advances in Neural Information Processing Systems, 36, 2024. 2 +[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 7 +[12] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level im + +age customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2, 4 +[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 8 +[14] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6 +[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 4 +[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2, 3 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[18] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. In Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6 +[19] Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1487-1495, 2017. 5 +[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6, 7 +[21] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. In ECCV, pages 22-38. Springer, 2020. 2, 3, 4 +[22] Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772, 2016. 5, 6 +[23] Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2 +[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value + +maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3 +[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. openclip, 2021. 3, 6 +[27] LAION-AI. Clip-benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2024.6 +[28] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 7 +[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 6 +[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3, 6, 7 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2, 3, 6 +[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 7 +[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 7 +[37] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2024. 3 +[38] Jiahuan Long, Tingsong Jiang, Wen Yao, Shuai Jia, Weijia Zhang, Weien Zhou, Chao Ma, and Xiaoqian Chen. Papmot: Exploring adversarial patch attack against multiple ob + +ject tracking. In European Conference on Computer Vision, pages 128-144. Springer, 2024. 2 +[39] Jiahuan Long, Zhengqin Xu, Tingsong Jiang, Wen Yao, Shuai Jia, Chao Ma, and Xiaoqian Chen. Robust sam: On the adversarial robustness of vision foundation models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5775-5783, 2025. 2 +[40] Jiahuan Long, Wen Yao, Tingsong Jiang, and Chao Ma. Cdupatch: Color-driven universal adversarial patch attack for dual-modal visible-infrared detectors. arXiv preprint arXiv:2504.10888, 2025. 2 +[41] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3 +[42] Spandan Madan, Timothy Henry, Jamell Dozier, Helen Ho, Nishchal Bhandari, Tomotake Sasaki, Frédo Durand, Hanspeter Pfister, and Xavier Boix. When and how convolutional neural networks generalize to out-of-distribution category-viewpoint combinations. Nature Machine Intelligence, 4(2):146-153, 2022. 2, 3, 4 +[43] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 3 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3 +[45] openai. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.3,7,8 +[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6 +[48] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4 +[50] Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, and Xingxing Wei. Towards viewpoint-invariant visual recognition via adversarial training. In ICCV, pages 1-10, 2023. 2, 3, 4, 6 + +[51] Shouwei Ruan, Yinpeng Dong, Hang Su, Jianteng Peng, Ning Chen, and Xingxing Wei. Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528, 2023. 3 +[52] Shouwei Ruan, Yinpeng Dong, Hanqing Liu, Yao Huang, Hang Su, and Xingxing Wei. Omniview-tuning: Boosting viewpoint invariance of vision-language pre-training models. arXiv preprint arXiv:2404.12139, 2024. 3 +[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3 +[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[55] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2 +[56] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2 +[57] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 4 +[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 3, 5 +[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[61] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 8 +[62] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In The Twelfth International Conference on Learning Representations, 2024. 6 +[63] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2 + +[64] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3 +[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 6 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 2, 4 +[67] Yichi Zhang, Yao Huang, Yitong Sun, Chang Liu, Zhe Zhao, Zhengwei Fang, Yifan Wang, Huanran Chen, Xiao Yang, Xingxing Wei, Hang Su, Yinpeng Dong, and Jun Zhu. Benchmarking trustworthiness of multimodal large language models: A comprehensive study, 2024. 2, 3, 6, 8 +[68] Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. Proceedings of the European conference on computer vision (ECCV), 2024. 2 +[69] Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, and Adam Kortylewski. Ood-cv: A benchmark for robustness to out-of-distribution shifts of individual nuisances in natural images. In European conference on computer vision, pages 163-180. Springer, 2022. 3 +[70] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2 +[71] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6 +[72] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6, 7 \ No newline at end of file diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/images.zip b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..94ea71594bbd32b4cbcfee4e2419c398c8e9adc1 --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:218f3e7c39ff0c4b97ad01db79796a54185128a3c6dd4188fba30736374b5eba +size 765269 diff --git a/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/layout.json b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9bd86d07300361f84e49839d069460501322c2b7 --- /dev/null +++ b/2025/AdvDreamer Unveils_ Are Vision-Language Models Truly Ready for Real-World 3D Variations_/layout.json @@ -0,0 +1,10530 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 237, + 104, + 373, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 104, + 373, + 118 + ], + "spans": [ + { + "bbox": [ + 237, + 104, + 373, + 118 + ], + "type": "text", + "content": "AdvDreamer Unveils:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 121, + 532, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 121, + 532, + 138 + ], + "spans": [ + { + "bbox": [ + 78, + 121, + 532, + 138 + ], + "type": "text", + "content": "Are Vision-Language Models Truly Ready for Real-World 3D Variations?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": "Shouwei Ruan" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Hanqing Liu" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Yao Huang" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Xiaqi Wang" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Caixin Kang" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Hang Su" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Yinpeng Dong" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "text", + "content": ", Xingxing Wei" + }, + { + "bbox": [ + 118, + 160, + 492, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 222, + 189, + 389, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 189, + 389, + 203 + ], + "spans": [ + { + "bbox": [ + 222, + 189, + 389, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 222, + 189, + 389, + 203 + ], + "type": "text", + "content": "Institute of Artificial Intelligence," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 203, + 511, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 203, + 511, + 217 + ], + "spans": [ + { + "bbox": [ + 99, + 203, + 511, + 217 + ], + "type": "text", + "content": "State Key Laboratory of Virtual Reality Technology and Systems, Beihang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 217, + 507, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 217, + 507, + 245 + ], + "spans": [ + { + "bbox": [ + 102, + 217, + 507, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 217, + 507, + 245 + ], + "type": "text", + "content": " Dept. of Comp. Sci. and Tech., Institute for AI, Tsinghua-Bosch Joint ML Center, THBI Lab, BNRist Center, Tsinghua University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 245, + 394, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 245, + 394, + 259 + ], + "spans": [ + { + "bbox": [ + 216, + 245, + 394, + 259 + ], + "type": "text", + "content": "3 College of AI, Tsinghua University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 262, + 530, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 262, + 530, + 272 + ], + "spans": [ + { + "bbox": [ + 78, + 262, + 530, + 272 + ], + "type": "text", + "content": "{shouweiruan, hqliu, y_huang, xxwei}@buaa.edu.cn, suhangss@tsinghua.edu.cn, dongyinpeng@mail.tsinghua.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 277, + 515, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 277, + 515, + 286 + ], + "spans": [ + { + "bbox": [ + 94, + 277, + 515, + 286 + ], + "type": "text", + "content": "(A) Adversarial 3D Variations Samples form Single-view Natural Images —— (C) Background Variations for Adv-3DT samples" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 97, + 292, + 168, + 346 + ], + "blocks": [ + { + "bbox": [ + 97, + 292, + 168, + 346 + ], + "lines": [ + { + "bbox": [ + 97, + 292, + 168, + 346 + ], + "spans": [ + { + "bbox": [ + 97, + 292, + 168, + 346 + ], + "type": "image", + "image_path": "2d4983076e76f69f2c01aefff157f73c37a9ec199c7865d471d4722f1c55056f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 179, + 292, + 286, + 349 + ], + "blocks": [ + { + "bbox": [ + 179, + 292, + 286, + 349 + ], + "lines": [ + { + "bbox": [ + 179, + 292, + 286, + 349 + ], + "spans": [ + { + "bbox": [ + 179, + 292, + 286, + 349 + ], + "type": "image", + "image_path": "91513af6c0274a700ac71aea6cfc52a39e93dfec9cb07450b16365e25fe6be56.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 293, + 292, + 388, + 348 + ], + "blocks": [ + { + "bbox": [ + 293, + 292, + 388, + 348 + ], + "lines": [ + { + "bbox": [ + 293, + 292, + 388, + 348 + ], + "spans": [ + { + "bbox": [ + 293, + 292, + 388, + 348 + ], + "type": "image", + "image_path": "e744f7a72a21a6fb0987cd19d08e4cb0f2bafc9faf47370d3e067e3742e20343.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 94, + 351, + 238, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 351, + 238, + 358 + ], + "spans": [ + { + "bbox": [ + 94, + 351, + 238, + 358 + ], + "type": "text", + "content": "(B) Physical Reproduction of Adversarial 3D Variations Samples" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 82, + 359, + 162, + 442 + ], + "blocks": [ + { + "bbox": [ + 82, + 359, + 162, + 442 + ], + "lines": [ + { + "bbox": [ + 82, + 359, + 162, + 442 + ], + "spans": [ + { + "bbox": [ + 82, + 359, + 162, + 442 + ], + "type": "image", + "image_path": "c916d5b785362732141f6114aeb74e746360d29d2dcb1d6b5768dd8871de4679.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 555, + 504 + ], + "type": "text", + "content": "Figure 1. The 3D Variation Vulnerabilities in VLMs. (A) The proposed AdvDreamer allows the capture of adversarial 3D transformation (Adv-3DT) samples from single natural images, threatening various VLMs. (B) These Adv-3DT samples can be successfully reproduced in the physical world and (C) retain their aggressiveness under background variations. (D) We also explain the GPT-Score metric used in the experiments. In each VLM response, we highlight the correct, significant error, and less precise parts using different colors." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 165, + 359, + 235, + 442 + ], + "blocks": [ + { + "bbox": [ + 165, + 359, + 235, + 442 + ], + "lines": [ + { + "bbox": [ + 165, + 359, + 235, + 442 + ], + "spans": [ + { + "bbox": [ + 165, + 359, + 235, + 442 + ], + "type": "image", + "image_path": "e1357629bf1be364d02f6f37868cce6b703a631fcb3d42a2a4a261caf59cc593.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 242, + 361, + 310, + 442 + ], + "blocks": [ + { + "bbox": [ + 242, + 361, + 310, + 442 + ], + "lines": [ + { + "bbox": [ + 242, + 361, + 310, + 442 + ], + "spans": [ + { + "bbox": [ + 242, + 361, + 310, + 442 + ], + "type": "image", + "image_path": "7bd5121d7cfb2b332e5d7ea86ac4f2282435d98bb4d151e2bbf754144ee322b7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 361, + 389, + 444 + ], + "blocks": [ + { + "bbox": [ + 317, + 361, + 389, + 444 + ], + "lines": [ + { + "bbox": [ + 317, + 361, + 389, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 389, + 444 + ], + "type": "image", + "image_path": "17f84b046911aa55a6bb4c55168ef72fc3f589bb2e69bf4776c697faedc92b96.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 279, + 529, + 384 + ], + "blocks": [ + { + "bbox": [ + 395, + 279, + 529, + 384 + ], + "lines": [ + { + "bbox": [ + 395, + 279, + 529, + 384 + ], + "spans": [ + { + "bbox": [ + 395, + 279, + 529, + 384 + ], + "type": "image", + "image_path": "ac1996e7d1a94690ce288b9d8e76acf48d3d5edae43fffdfe99953a22c0c4068.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 396, + 386, + 529, + 445 + ], + "blocks": [ + { + "bbox": [ + 396, + 386, + 529, + 445 + ], + "lines": [ + { + "bbox": [ + 396, + 386, + 529, + 445 + ], + "spans": [ + { + "bbox": [ + 396, + 386, + 529, + 445 + ], + "type": "image", + "image_path": "3cef88438797ec505262a305d763b93ec92d5512fd0eeec51beeaa8a6059bb26.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "spans": [ + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 684 + ], + "type": "text", + "content": "Vision Language Models (VLMs) have exhibited remarkable generalization capabilities, yet their robustness in dynamic real-world scenarios remains largely unexplored. To systematically evaluate VLMs' robustness to real-world 3D variations, we propose AdvDreamer, the first framework capable of generating physically reproducible Adversarial 3D Transformation (Adv-3DT) samples from single-view observations. In AdvDreamer, we integrate three key innovations: Firstly, to characterize real-world 3D variations with limited prior knowledge precisely, we design a zero-shot Monocular Pose Manipulation pipeline built upon generative 3D priors. Secondly, to ensure the visual qual" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 515, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 556, + 707 + ], + "type": "text", + "content": "ity of worst-case Adv-3DT samples, we propose a Naturalness Reward Model that provides continuous naturalness regularization during adversarial optimization, effectively preventing convergence to hallucinated or unnatural elements. Thirdly, to enable systematic evaluation across diverse VLM architectures and visual-language tasks, we introduce the Inverse Semantic Probability loss as the adversarial optimization objective, which solely operates in the fundamental visual-textual alignment space. Based on the captured Adv-3DT samples with high aggressiveness and transferability, we establish MM3DTBench, the first VQA benchmark dataset tailored to evaluate VLM robustness under challenging 3D variations. Extensive evaluations of representative VLMs with varying architectures reveal that real-world 3D variations can pose severe threats to model performance across various tasks." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 693, + 139, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 139, + 703 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 139, + 703 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 70, + 693, + 139, + 703 + ], + "type": "text", + "content": " Equal contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 70, + 704, + 150, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 704, + 150, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 704, + 150, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7894" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 175, + 152 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 175, + 152 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 175, + 152 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 175, + 152 + ], + "type": "image", + "image_path": "65f7a3ecc01c3954051e3bff120793d5a59b93dd7a949d9ec21e82e1c2cd882f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 159, + 493, + 171 + ], + "lines": [ + { + "bbox": [ + 115, + 159, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 159, + 493, + 171 + ], + "type": "text", + "content": "Figure 2. Comparison of paradigms to characterize Adversarial 3D Transformation (Adv-3DT) samples." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 292, + 152 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 292, + 152 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 292, + 152 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 292, + 152 + ], + "type": "image", + "image_path": "5a35f2f9f3b56dc9fc20c47dcb736c0b822dcff76a70a07edbfccaf520881b5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 408, + 152 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 408, + 152 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 408, + 152 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 408, + 152 + ], + "type": "image", + "image_path": "e7f87b135794b238f091b2754fb9dbcb9580202b93b25ef6009a58b04e479628.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 411, + 72, + 527, + 155 + ], + "blocks": [ + { + "bbox": [ + 411, + 72, + 527, + 155 + ], + "lines": [ + { + "bbox": [ + 411, + 72, + 527, + 155 + ], + "spans": [ + { + "bbox": [ + 411, + 72, + 527, + 155 + ], + "type": "image", + "image_path": "013f93479f7280a02673e7bc8ccc94ae40796dd9cb8d75e30d2599568c88ff33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 177, + 136, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 136, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 136, + 189 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 198, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 341 + ], + "type": "text", + "content": "Recently, Vision-Language Models (VLMs) [7, 13, 35, 36] have demonstrated remarkable capabilities in bridging visual perception and natural language understanding. Through large-scale image-text pre-training followed by fine-grained instruction tuning [31, 36], these models effectively tackle a wide range of visual-centric tasks, including instruction-based image recognition [47, 55, 65], visual question answering [3, 31, 35, 36], and visual reasoning [9, 10]. Due to their strong generalizability and impressive zero-shot capability, VLMs are increasingly deployed in various safety-critical applications, particularly in autonomous driving [56, 63] and robotic systems [25, 64]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 472 + ], + "type": "text", + "content": "Despite extensive studies validating VLMs' resilience against various distribution shifts (e.g., style variations [2, 39, 40, 68], image corruptions [2], and adversarial perturbations [16, 38, 67, 70]), they primarily focus on 2D perturbations in the digital domain, overlooking a critical challenge in real-world deployment: the 3D variations. As VLMs are increasingly integrated into dynamic scenarios, their ability to handle 3D variation becomes crucial, raising a fundamental question: Have current VLMs attained sufficient robustness to the distribution shifts arising from ubiquitous 3D variations in the real world?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 498 + ], + "type": "text", + "content": "To thoroughly explore this problem and offer comprehensive answers, we identify three fundamental challenges:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "(1) How to accurately characterize real-world 3D variations with limited prior knowledge? Existing methods exploring the 3D variation robustness heavily rely on explicit 3D structures [4, 21] or representations derived from dense multi-view observations [4, 21, 42]. However, in most practical scenarios, the available priors are severely limited (e.g., single-view observation only). This constraint necessitates developing a method that minimizes dependency on extensive scene priors. To meet this, we introduce a zero-shot Monocular Pose-Manipulation (MPM) pipeline based on generative representations. Specifically, MPM leverages the rich 3D priors embedded in pre-trained Large-Reconstructive Models (LRM) [23, 24] and Image Re-Composition Models [12, 66], enabling fine-grained 3D adjustments with only single-view images and specified 3D variation parameters. While this generative-based paradigm establishes a foundation for capturing worst-case 3D variations through optimizing 3D variation parameters, it still" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 178, + 553, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 202 + ], + "type": "text", + "content": "exhibits inherent limitations that may introduce undesirable image corruptions, thus presenting our second challenge:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 205, + 555, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 205, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 205, + 555, + 443 + ], + "type": "text", + "content": "(2) How to capture the worst-case adversarial 3D transformation (Adv-3DT) samples while ensuring that the performance degradation is inherently caused by the 3D variation? Prior works utilizing synthetic 3D assets [4, 21, 42] frequently produce samples with noticeable texture discrepancies compared to real-world images. These limitations in samples' visual fidelity and contextual coherence highlight a critical concern: observed performance degradation under Adv-3DT samples may stem from undesirable corruption rather than the 3D variation itself. To address this, we incorporate naturalness reward constraints within the adversarial optimization process. Specifically, we treat rigid 3D variations as an adversarial attack and propose the Naturalness Reward Model (NRM) that preserves image quality throughout optimization. NRM leverages visual context features provided by the DINOv2 [46] backbone to predict a naturalness reward that jointly reflects the visual fidelity and physical plausibility of the samples, effectively regularizing the optimization trajectory and preventing convergence toward unnatural or hallucinated elements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 698 + ], + "type": "text", + "content": "(3) How to ensure effective generalization of Adv-3DT samples across diverse VLM tasks and architectures? Existing studies [4, 18, 50] typically rely on task-specific objectives with end-to-end optimization, substantially limiting their applicability across diverse VLM architectures and a wide range of visual language tasks. To transcend this limitation, we formulate the adversarial objective in a task-and architecture-agnostic manner. Specifically, we introduce the Inverse Semantic Probability (ISP) loss, which operates solely through the foundational component of the VLM—the visual encoder, aiming to minimize the image-text matching probability between Adv-3DT samples and their corresponding ground-truth textual descriptions. Extensive experiments demonstrate that Adv-3DT samples optimized under the ISP loss exhibit remarkable adversarial transferability across various downstream tasks and models. Building on this, we further introduce the MM3DTBench, the first VQA benchmark dataset specifically designed to evaluate the 3D variation robustness of VLMs in real-world scenarios, comprising highly challenging Adv-3DT samples with carefully crafted candidate answers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "Integrating the aforementioned innovations, this paper" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7895" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "proposes AdvDreamer, a novel framework that enables the characterization of Adv-3DT samples from monocular observations, with broad adaptability across various visual-language tasks and VLM architectures. Our extensive digital and physical experiments reveal significant vulnerabilities in contemporary VLMs when confronted with challenging 3D variations. In visual recognition tasks, foundational models like OpenCLIP [26] and BLIP-2 [31] suffer a performance degradation of " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "65\\% \\sim 80\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " on Adv-3DT samples. More concerning still, in open-ended tasks including image captioning and visual question answering, even the most advanced commercial models like GPT-4o [45] exhibit performance declines approaching " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": ". Our findings highlight the urgent need to enhance VLM robustness against 3D variations and set new requirements for deploying VLMs in complex, dynamic real-world applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 143, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 143, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 143, + 289 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 194, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 194, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 194, + 309 + ], + "type": "text", + "content": "2.1. Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 314, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 458 + ], + "type": "text", + "content": "Classic Vision-Language Models (VLMs) [30, 32, 41, 47] achieve cross-modal alignment through diverse pre-training objectives on large-scale image-text pairs [33, 53]. Recent advances in Large Language Models (LLMs) [59, 60] have inspired new VLM architectures that integrate LLM capabilities via MLP projection layers [35, 36] or specialized modules like Q-Former [31]. Through techniques like instruction tuning [36], current VLMs [13, 35, 36] demonstrate enhanced performance across vision-centric tasks, including visual grounding, question answering, and image captioning, etc. These models also enable intelligent systems for task planning and decision-making [25, 64]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 459, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 615 + ], + "type": "text", + "content": "Despite their impressive capabilities [37, 47, 67], concerns remain about the robustness and generalization of VLMs, especially in safety-critical environments. Recent studies [52, 58] highlight that VLMs still exhibit deficiencies in processing visual details, such as orientation, quantity, and colour, which may be attributed to their visual representations. Moreover, VLMs remain susceptible to " + }, + { + "bbox": [ + 55, + 459, + 296, + 615 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 55, + 459, + 296, + 615 + ], + "type": "text", + "content": "-norm adversarial perturbations [16, 43], with some efforts aiming to enhance their robustness through adversarial finetuning techniques [43]. In contrast to previous works, our study specifically investigates VLM robustness against 3D variations, addressing a crucial gap in understanding their reliability for dynamic real-world applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 264, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 264, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 264, + 635 + ], + "type": "text", + "content": "2.2. Evaluation for 3D Variation Robustness" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Robustness to 3D variations remains a longstanding challenge in computer vision. Evaluating this property often involves capturing worst-case 3D transformations, i.e. the adversarial poses or viewpoints, to create challenging samples for assessment. Prior research can be categorized into three paradigms (see Fig.2): (A) Human-guided methods, like" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 264 + ], + "type": "text", + "content": "ObjectNet [8] and OOD-CV [69], manually collect samples with diverse poses and viewpoints, but they lack systematic coverage of worst-case examples and are costly. (B) Simulator-based [4, 21, 42] methods optimize worst-case 3D parameters via differentiable rendering of synthetic objects but require manually designed 3D assets and produce samples with limited realism. (C) Multi-view-based methods, like ViewFool [18] and GMVFool [50, 51], use NeRF-based [44] representations to optimize adversarial viewpoints, but they require dense observations and struggle with background integration. In contrast, the proposed AdvDreamer employs a (D) Generative-based approach, utilizing robust 3D priors from generative models. Given a single natural image, AdvDreamer generate realistic Adv-3DT samples in a zero-shot manner, providing a more efficient solution for evaluating 3D variation robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 273, + 397, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 397, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 397, + 287 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "text", + "content": "The overview of AdvDreamer is shown in Fig. 3. It consists " + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "text", + "content": " Zero-shot Monocular Pose Manipulation pipeline, leveraging generative 3D representations to obtain specified 3D variation samples from single-view images; " + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "text", + "content": " Naturalness Reward Model, providing naturalness regularization during adversarial optimization; and " + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 313, + 293, + 555, + 437 + ], + "type": "text", + "content": " Inverse Semantic Probability loss, ensuring the Adv-3DT samples effectively threaten VLMs and generalize across different architectures and tasks. The following sections will firstly formalize the optimization problem of AdvDreamer (Sec.3.1), then sequentially describe the key components (Sec.3.2), and finally outline the full optimization algorithm (Sec.3.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 443, + 440, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 440, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 440, + 455 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": "Parametrization of Real-World 3D Variations. In this paper, we primarily focus on rigid 3D transformations of foreground objects, i.e., the rotation, translation, and scaling, since they reflect the most typical 3D changes observed in real-world environments. Formally, we define a 6-dimensional vector " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\Theta = \\{\\alpha, \\beta, \\gamma, \\Delta_x, \\Delta_y, s\\}" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " to uniquely parameterize any arbitrary transformation, where " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\{\\alpha, \\beta, \\gamma\\}" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " denote the Tait-Bryan angles (yaw, pitch, roll) in the " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "Z-Y-X" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " sequence, " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\{\\Delta_x, \\Delta_y\\}" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " represent the translation along the " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " axes in the " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "xy" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": "-plane, and " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " is the uniform scaling factor. To prevent the loss of critical visual cues and ensure human recognizability, we follow previous work [18] to constrain " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": " within a bounded range " + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "inline_equation", + "content": "[\\Theta \\min, \\Theta \\max]" + }, + { + "bbox": [ + 313, + 460, + 555, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 616, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 677 + ], + "type": "text", + "content": "Optimization Problem. The objective of AdvDreamer is to find an optimal distribution " + }, + { + "bbox": [ + 313, + 616, + 556, + 677 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 313, + 616, + 556, + 677 + ], + "type": "text", + "content": ", ensuring that Adv-3DT sampled from this distribution are both aggressive against VLMs and maintaining visual quality, which can be abstractly formulated as the following optimization problem:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 682, + 553, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 682, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 325, + 682, + 553, + 703 + ], + "type": "interline_equation", + "content": "p ^ {*} (\\Theta) = \\arg \\max _ {p (\\Theta)} \\mathbb {E} _ {\\Theta \\sim p (\\Theta)} \\left[ \\mathcal {L} _ {I S P} \\left(X ^ {\\prime}\\right) + \\mathcal {L} _ {N a t} \\left(X ^ {\\prime}\\right) \\right], \\tag {1}", + "image_path": "26c6059cc3805f597b9a2bad377f4141456913d4ac91b3d4a736d5a871d311d1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 703, + 476, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 703, + 476, + 715 + ], + "spans": [ + { + "bbox": [ + 382, + 703, + 476, + 715 + ], + "type": "interline_equation", + "content": "w h e r e X ^ {\\prime} = \\mathcal {T} (\\Theta , X),", + "image_path": "69a0a538296accbeb82a8ae12b5613c4e73010700483ec657be76435502c54d3.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7896" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 72, + 521, + 232 + ], + "blocks": [ + { + "bbox": [ + 87, + 72, + 521, + 232 + ], + "lines": [ + { + "bbox": [ + 87, + 72, + 521, + 232 + ], + "spans": [ + { + "bbox": [ + 87, + 72, + 521, + 232 + ], + "type": "image", + "image_path": "225896af9b58da7824ec8152570a7f4886dc366d21724adb3066679c1bfaca94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 281 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed AdvDreamer framework. To capture worst-case 3D variations in the real world, we treat rigid 3D transformations as adversarial attacks, optimizing the distribution of transformation parameters through a query-based approach. In AdvDreamer, we introduce the Monocular Pose Manipulation pipeline to perform specified 3D transformations on single-view images and guide the optimization process using the proposed Naturalness Reward Model and Inverse Semantic Probability loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\Theta, X)" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": " is a transformation function to produce a novel image with the specified transformation " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": " applied to the original image " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": ". To implement " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": ", we design the MPM pipeline, which will be detailed in Sec.3.2.1. The " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Nat}" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ISP}" + }, + { + "bbox": [ + 55, + 290, + 296, + 373 + ], + "type": "text", + "content": " measure the naturalness and adversarial efficacy of the transformed samples, respectively. These will be further elaborated in Sec.3.2.2 and Sec.3.2.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "text", + "content": "Optimizing over the distribution " + }, + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "inline_equation", + "content": "p(\\Theta)" + }, + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "text", + "content": " rather than a deterministic " + }, + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 374, + 296, + 494 + ], + "type": "text", + "content": " offers two key advantages. 1) As highlighted in [18, 50], learning the underlying distribution enables comprehensive exploration of the 3D variation space, facilitating the identification of the model's vulnerable regions. 2) Given the stochastic nature of the generation, optimizing over a continuous distribution can mitigate the impact of appearance inconsistency, as appearance variations induced by the generation are unlikely to consistently deceive the model without optimizing the adversarial 3D variations [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 198, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 198, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 198, + 513 + ], + "type": "text", + "content": "3.2. AdvDreamer Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 518, + 257, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 257, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 257, + 531 + ], + "type": "text", + "content": "3.2.1. Zero-shot Monocular Pose Manipulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "content": "To implement the transformation function " + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 534, + 295, + 617 + ], + "type": "text", + "content": ", previous approaches typically rely on explicit 3D structures [4, 21, 42] or neural radiance representations [18, 50]. However, these approaches cannot accommodate real-world scenarios with limited prior knowledge. To address this, we design a simple yet effective zero-shot monocular pose manipulation (MPM) pipeline that consists of the following process:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "1 Foreground-Background Pairs Preparation. MPM first decompose the input image " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " into foreground-background pairs " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{X_{f},X_{b}\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ". Specifically, we leverage Grounded-SAM [48] to semantically segment the major instance as foreground " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{f}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ", followed by diffusion-based inpainting [15, 49] to obtain a complete background " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{b}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ". This process is denoted as " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{X_f,X_b\\} = \\mathcal{F}(X)" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": ". Additionally, we support directly generating synthetic " + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{X_f,X_b\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": " pairs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 290, + 552, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 552, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 552, + 303 + ], + "type": "text", + "content": "through Stable-Diffusion [49] to enhance sample diversity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " Object-level Pose Manipulation. For each optimization iteration, we transform the foreground " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "X_{f}" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " into a batch of " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "X_{f}^{\\prime}" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " sampled from current " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "p(\\Theta)" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": ". In this stage, MPM leverages the robust priors provided by a pre-trained Large Reconstruction Model (e.g., TripoSR [57]) to construct single-view-based 3D representations and apply sampled transformations. This process is formulated as " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "X_{f}^{\\prime} = \\mathcal{R}_{\\mathbf{w}_{0}}(X_{f},\\Theta)" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\Theta \\sim p(\\Theta)" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathbf{w}}(\\cdot)" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": " denotes neural rendering function with weights " + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_0" + }, + { + "bbox": [ + 313, + 303, + 553, + 411 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "content": " Re-Composition. We then compose the transformed foreground " + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "inline_equation", + "content": "X_{f}^{\\prime}" + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "content": " with background " + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "inline_equation", + "content": "X_{b}" + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "content": " through a pre-trained diffusion-based composition model [12, 66], to handles shadows and occlusions. This process is defined as " + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "inline_equation", + "content": "X^{\\prime} = \\mathcal{C}_{\\mathbf{w}_1}(X_f',X_b)" + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_1" + }, + { + "bbox": [ + 313, + 411, + 553, + 471 + ], + "type": "text", + "content": " denotes diffusion model's weights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "content": "Combining the above stages, the transformation process by MPM can be formally represented as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 504, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 504, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 347, + 504, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\begin{array}{r l} X ^ {\\prime} = \\mathcal {T} (\\boldsymbol {\\Theta}, X) & = \\mathcal {C} _ {\\mathbf {w} _ {1}} \\left(\\mathcal {R} _ {\\mathbf {w} _ {0}} \\left(X _ {f}, \\boldsymbol {\\Theta}\\right), X _ {b}\\right), \\\\ \\text {w h e r e} \\{X _ {f}, X _ {b} \\} & = \\mathcal {F} (X). \\end{array} \\tag {2}", + "image_path": "fa638702eef575c9dc03fda96127c3cc7985f73c25884c7fc2fa854295982ad5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 460, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 460, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 460, + 553 + ], + "type": "text", + "content": "3.2.2. Naturalness Reward Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": "Despite the MPM pipeline utilizing generative priors to achieve realistic 3D variation samples, we observe that directly applying adversarial optimization frequently leads to unnatural artifacts, such as shape distortions and physically implausible cases in Fig. 5. These issues stem from two factors: generalization limitations of existing generative models and the inherent high semantic bias in low-quality samples. To address this challenge, we propose a novel Naturalness Reward Model (NRM) that continuously regularizes " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "p(\\Theta)" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": " during optimization based on sample naturalness, preventing convergence to low-quality pseudo-optimal regions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "As shown in Fig.4, the NRM employs DINOv2[46] as its backbone, chosen for several reasons. Firstly, DINO" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7897" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 288, + 131 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 288, + 131 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 288, + 131 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 288, + 131 + ], + "type": "image", + "image_path": "fbf19ad0829fae426c631676671a2ca58d284ef7c89a9d396eecf8e158cc5fbe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 160 + ], + "type": "text", + "content": "Figure 4. Architecture of NRM and computational cost and accuracy of naturalness assessment across Human, GPT, and NRM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 276 + ], + "type": "text", + "content": "captures finer visual contextual features compared to traditional pre-trained vision encoders [58], aiding better convergence during training. Secondly, unlike vision-language aligned encoders such as CLIP, DINO's visual features are language-agnostic, making them more faithfully reflect image quality rather than semantic consistency. We use DINOv2 to extract image tokens from the input Adv-3DT samples, followed by dual-stream prediction heads to estimate visual fidelity and physical plausibility scores:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 309 + ], + "type": "interline_equation", + "content": "\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) = f _ {R} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right) = f _ {P} \\left(\\mathcal {E} _ {D} \\left(X ^ {\\prime}\\right)\\right), \\tag {3}", + "image_path": "1235b0db729d5034321300ec1dead6ebe933e3f020f9dbf98daf80ce72215696.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "inline_equation", + "content": "f_{R}(\\cdot)" + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "inline_equation", + "content": "f_{P}(\\cdot)" + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "content": " are two-layer MLPs. To train the NRM, we construct a large-scale naturalness preference dataset with fine-grained scoring. We generate numerous samples from the MPM pipeline on ImageNet and synthetic images, scoring each on a 5-point scale for visual fidelity and physical plausibility. These scores are initially annotated automatically by GPT-4o with Chain-of-Thought prompts and verified by volunteers for correction. This process resulted in 120K high-quality training samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 514 + ], + "type": "text", + "content": "The architecture and training parameters of NRM are detailed in Appendix B. Our experimental results demonstrate that the naturalness feedback provided by NRM highly aligns with human evaluation. In terms of prediction efficiency and stability, it outperforms methods that directly employ GPT-4o for naturalness assessment (see Fig. 4 and Appendix C.2). Based on the NRM's predictions, we define the naturalness regularization loss for the optimization as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 522, + 295, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 522, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 80, + 522, + 295, + 546 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) = - \\frac {1}{2} \\left(\\operatorname {S c o r e} _ {R} \\left(X ^ {\\prime}\\right) + \\operatorname {S c o r e} _ {P} \\left(X ^ {\\prime}\\right)\\right) \\tag {4}", + "image_path": "57c44aca59a98ee9f9c0f2a696a8b927177b3900f83fdbb084a2120e23892d8a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 230, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 230, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 230, + 566 + ], + "type": "text", + "content": "3.2.3. Inverse Semantic Probability Loss" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Previous studies typically employ task-specific optimization objectives with end-to-end manner to obtain worst-case Adv-3DT samples. However, directly applying this approach to VLMs with substantial parameters and diverse architectures spanning various tasks is challenging. To address this, we design the Inverse Semantic Probability (ISP) loss, which minimizes the probability that VLMs assign to the correct semantic attributes of Adv-3DT samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "This design have two key considerations. 1) It is based solely on the visual and text encoder " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_v" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", which are fundamental components across modern VLMs, ensuring architecture-agnostic applicability. 2) By operating in the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "image-text alignment space rather than model-specific layers or task-specific heads, it enables task generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": "Given a generated sample " + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": " and a semantic label set " + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{y_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": " containing the ground-truth label " + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "inline_equation", + "content": "y_t" + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": ", we first compute a semantic similarity vector " + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 96, + 553, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 322, + 137, + 553, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 137, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 322, + 137, + 553, + 152 + ], + "type": "interline_equation", + "content": "\\mathbf {S} = \\left\\{s _ {i} \\right\\} _ {i = 1} ^ {N}, \\text {w h e r e} s _ {i} = \\cos \\left[ \\mathcal {E} _ {v} \\left(X ^ {\\prime}\\right), \\mathcal {E} _ {t} \\left(\\mathcal {P} \\left(y _ {i}\\right)\\right) \\right], \\tag {5}", + "image_path": "c4fd0b3fd1111f1675608b96ef930fa099d064f1949be8ffbb95635df839fc2f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot)" + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": " computes cosine similarity and " + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": " follows the template \"a photo of a " + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "inline_equation", + "content": "\\{y_i\\}" + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": "\" in [47]. The vector " + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": " captures the representation similarities between " + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "inline_equation", + "content": "X'" + }, + { + "bbox": [ + 313, + 157, + 554, + 218 + ], + "type": "text", + "content": " and each semantic label. We then derive the conditional semantic probability through softmax transformation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 222, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 222, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 358, + 222, + 553, + 239 + ], + "type": "interline_equation", + "content": "p \\left(y _ {i} \\mid X ^ {\\prime}\\right) = \\exp \\left(s _ {i}\\right) / \\sum_ {j = 1} ^ {N} \\exp \\left(s _ {j}\\right). \\tag {6}", + "image_path": "66c23fff467735905996f39eed40b4a2e5056e73a8f4fdd36dfba8980df7216a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 243, + 554, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 554, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 554, + 268 + ], + "type": "text", + "content": "Then, the " + }, + { + "bbox": [ + 313, + 243, + 554, + 268 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ISP}}" + }, + { + "bbox": [ + 313, + 243, + 554, + 268 + ], + "type": "text", + "content": " is defined as the negative log-likelihood of the ground-truth's semantic probability:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 273, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 273, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 370, + 273, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) = - \\log p \\left(y _ {t} \\mid X ^ {\\prime}\\right). \\tag {7}", + "image_path": "06ef48e35c9014b6137fb05aae4a194554f2eccd8bfa856107f683ed8e1b7676.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ISP}}" + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "text", + "content": " quantifies the likelihood of " + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "inline_equation", + "content": "X^{\\prime}" + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "text", + "content": " being misclassified by comparing its similarity to the ground-truth and irrelevant categories. Maximizing it directs the " + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "inline_equation", + "content": "p(\\Theta)" + }, + { + "bbox": [ + 313, + 293, + 554, + 341 + ], + "type": "text", + "content": " towards regions that effectively deceive the model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 346, + 522, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 522, + 360 + ], + "type": "text", + "content": "3.3. Query-Based Adversarial Optimization" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 364, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 553, + 389 + ], + "type": "text", + "content": "To solve Problem (1), we parameterize the adversarial distribution " + }, + { + "bbox": [ + 313, + 364, + 553, + 389 + ], + "type": "inline_equation", + "content": "p(\\Theta)" + }, + { + "bbox": [ + 313, + 364, + 553, + 389 + ], + "type": "text", + "content": " as a Multivariate Gaussian form:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 340, + 395, + 553, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 395, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 340, + 395, + 553, + 408 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Theta} = \\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, \\text {w h e r e} \\mathbf {z} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}), \\tag {8}", + "image_path": "170926bbbc3c6c5979cf0a614ed7c17332693f8fef0998a8ca371e17d1425dbb.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = (\\Theta_{\\mathrm{min}} - \\Theta_{\\mathrm{max}}) / 2" + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{A} = (\\Theta_{\\mathrm{min}} + \\Theta_{\\mathrm{max}}) / 2" + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": " follows a Gaussian Distribution with mean " + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{6 \\times 6}" + }, + { + "bbox": [ + 313, + 414, + 554, + 450 + ], + "type": "text", + "content": ". Thus, we rewrite (1) in a specific form:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 338, + 456, + 553, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 456, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 338, + 456, + 553, + 478 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma}} {\\arg \\max } \\mathbb {E} _ {\\boldsymbol {\\mathbf {z}} \\sim \\mathcal {N} (\\boldsymbol {\\mu}, \\boldsymbol {\\Sigma})} \\left[ \\mathcal {L} _ {\\text {I S P}} \\left(X ^ {\\prime}, \\mathcal {Y}\\right) + \\mathcal {L} _ {\\text {N a t}} \\left(X ^ {\\prime}\\right) \\right], \\tag {9}", + "image_path": "685330fe67fa63c578fd639890c4e2913a83320fdb41709bbe90a7412d24a09a.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 357, + 478, + 511, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 478, + 511, + 491 + ], + "spans": [ + { + "bbox": [ + 357, + 478, + 511, + 491 + ], + "type": "interline_equation", + "content": "w h e r e X ^ {\\prime} = \\mathcal {T} (\\mathbf {A} \\cdot \\tanh (\\mathbf {z}) + \\mathbf {B}, X).", + "image_path": "3665c26657ef002327103cde7538acc07589e24522b84b189cf68b050dd067f2.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": "Solving this problem typically involves computing the gradient of the loss w.r.t. " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": " and updating them via gradient ascent. However, the forward steps involve multiple components, e.g., the denoise process, introducing uncertainty in the gradient propagation path, making gradient-based optimization challenging [54]. Therefore, we adopt the Covariance Matrix Adaptation Evolution Strategy [19, 22], an efficient query-based black-box optimizer. At each iteration " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": ", we: 1) Sample " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": " candidates " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_i^{t + 1}\\}_{i = 1}^K" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\boldsymbol {\\mu}^t,\\boldsymbol {\\Sigma}^t)" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": ". 2) Generate corresponding samples " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\{(X')_{i}^{t + 1}\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": ". 3) Update distribution parameters " + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^t,\\pmb{\\Sigma}^t" + }, + { + "bbox": [ + 313, + 495, + 554, + 628 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 336, + 628, + 553, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 628, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 336, + 628, + 553, + 661 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {t + 1} = \\sum_ {i = 1} ^ {k} w _ {i} \\cdot \\mathbf {z} _ {(i: k)} ^ {t + 1}, \\quad \\text {w h e r e} \\sum_ {i = 1} ^ {k} w _ {i} = 1, \\tag {10}", + "image_path": "f4e53921326e8308ac711d7e73b31558a9c4f89ebe411f76240d84ef9f633ee1.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 324, + 665, + 526, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 665, + 526, + 681 + ], + "spans": [ + { + "bbox": [ + 324, + 665, + 526, + 681 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {t + 1} = \\left(1 - \\eta_ {1} - \\eta_ {\\mu}\\right) \\cdot \\boldsymbol {\\Sigma} ^ {t} + \\eta_ {1} \\cdot p _ {\\Sigma} ^ {t + 1} \\left(p _ {\\Sigma} ^ {t + 1}\\right) ^ {T} +", + "image_path": "628bf28dd52152d7b181f16dfa820294f1bb0402e25f40bbe60b73e462e52436.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 339, + 682, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 682, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 682, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\eta_ {\\mu} \\cdot \\sum_ {i = 1} ^ {h} w _ {i} \\cdot \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) \\left(\\frac {\\mathbf {z} _ {(i : k)} ^ {t + 1} - \\boldsymbol {\\mu} ^ {t}}{\\boldsymbol {\\sigma} ^ {t}}\\right) ^ {T}, \\tag {11}", + "image_path": "7a180b63f037ff2e7dd8a163482771778642e2814150b99e8a655801534305c7.jpg" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7898" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 96, + 297, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 1. Zero-shot Classification Accuracy " + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": " and degradation " + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "inline_equation", + "content": "(\\downarrow \\%)" + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": " Under Adv-3DT Samples Generated from ImageNet [17]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 96, + 297, + 173 + ], + "lines": [ + { + "bbox": [ + 57, + 96, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 297, + 173 + ], + "type": "table", + "html": "
Target Models#ParamsMethods
CleanRandomp*(Θ)Θ*
OpenCLIP ViT-B/16 [26]149M98.062.6 (↓35)54.0 (↓44)18.0 (↓80)
OpenCLIP ViT-L/14 [26]428M94.461.7 (↓33)50.9 (↓44)15.3 (↓79)
OpenCLIP ViT-G/14 [26]2.5B96.463.5 (↓33)53.5 (↓43)18.7 (↓78)
BLIP ViT-B/16 [30]583M83.056.0 (↓27)51.3 (↓32)17.3 (↓66)
BLIP-2 ViT-L/14 [31]3.4B84.057.0 (↓27)52.2 (↓32)18.7 (↓65)
BLIP-2 ViT-G/14 [31]4.1B81.055.7 (↓25)49.1 (↓32)15.7 (↓65)
", + "image_path": "2ce3e60904567f2362d7d27536c611b2c097c3d8ec000624b2c43656df8f47eb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": "here, " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{(i:k)}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " denotes samples ranked by losses: we first select top-10 samples based on " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ISP}}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": ", then choose the top-k " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "(k = 5)" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " with highest " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Nat}}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " for updates. " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "w_{i} = 1 / k" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " are importance weights, " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\eta_{1}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\eta_{\\mu}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " are learning rates, " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "p_{\\Sigma}" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " tracks the evolution path, and " + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\sigma^t" + }, + { + "bbox": [ + 55, + 178, + 296, + 286 + ], + "type": "text", + "content": " is the step size updated via Cumulative Step-size Adaptation (CSA) [22]. Due to space constraints, complete derivations of Eq. (10), (11) and hyperparameter settings are provided in the Appendix.A.1. We also provide the pseudocode for optimization process in Appendix A.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 138, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 138, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 138, + 308 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 313, + 172, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 172, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 172, + 326 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "content": "Tasks & Datasets. We evaluate VLMs' robustness to 3D variations using AdvDreamer on three representative vision-language tasks. (1) For zero-shot image classification, we utilize the ImageNet test set [17]. (2) For image captioning, we employ the test splits from COCO Caption [11] and NoCaps [1] datasets. (3) For Visual Question Answering (VQA), we use the VQAv2 [20] test set. For each dataset, we select 300 clean samples spanning 30 categories with clear and well-defined object instances. During the optimization process, we employ ImageNet-1K categories as the semantic label set " + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 55, + 331, + 296, + 474 + ], + "type": "text", + "content": " for the classification task and use COCO-80 categories for the captioning and VQA tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 545 + ], + "type": "text", + "content": "Victim Models. For zero-shot classification, we target vision-language foundation models: OpenCLIP [14, 26], BLIPs [30, 31], and additional models for transfer attacks. For other tasks, we select mainstream open-source VLMs (LLaVa [34, 36], MiniGPT-4 [72], etc.) as well as representative commercial VLMs (GPT-4o and GPT-4o-mini)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Methods & Baselines. We use two variants with AdvDreamer: (1) " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ": represents the average results over 10 Adv-3DT samples randomly sampled from the optimal adversarial distribution, while (2) " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta^*" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ": indicates the evaluation results using the Adv-3DT sample corresponding to the distribution center. Additionally, we establish two potential baselines following [18, 50]: the original natural images without transformation (Clean) and the average results over 10 images with random 3D transformations (Random). Metrics. For zero-shot classification, we follow the CLIP-Benchmark [27] protocol, reporting Top-1 accuracy and its degradation relative to clean samples. For captioning, in addition to conventional metrics (CIDEr, BLEU), we adopt the LLM-as-a-judge approach [36, 67, 71] to address the open" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 313, + 96, + 552, + 172 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 552, + 93 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 552, + 93 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 552, + 93 + ], + "type": "text", + "content": "Table 2. Cross-model transferability of Adv-3DT samples, utilizing OpenCLIP and BLIP as surrogate models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 96, + 552, + 172 + ], + "lines": [ + { + "bbox": [ + 313, + 96, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 552, + 172 + ], + "type": "table", + "html": "
Target ModelCleanOpenCLIP [26]BLIP [30]
p*(Θ)Θ*p*(Θ)Θ*
ALBEF [29]65.037.2 (↓28)27.7 (↓37)40.2 (↓25)25.7 (↓39)
OpenAI CLIP ViT-B/16 [47]85.348.7 (↓37)29.7 (↓56)52.8 (↓32)31.0 (↓54)
OpenCLIP ViT-B/16 [26]97.754.0 (↓44)18.0 (↓80)59.3 (↓38)34.7 (↓63)
Meta-CLIP ViT-B/16 [62]91.049.3 (↓42)30.0 (↓61)53.8 (↓37)29.7 (↓61)
BLIP ViT-B/16 [30]82.748.4 (↓34)30.7 (↓52)51.3 (↓31)17.3 (↓65)
BLIP-2 ViT-L/14 [31]84.050.0 (↓34)32.0 (↓52)52.3 (↓32)27.7 (↓56)
SigLIP ViT-B/16 [65]95.355.3 (↓40)34.0 (↓61)59.3 (↓36)34.7 (↓61)
", + "image_path": "8fac97422681d8f344213e62437351241009df581fc688a61d5822a6cbf831ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 178, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 178, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 312, + 178, + 555, + 275 + ], + "type": "text", + "content": "ended nature of VLM responses, aligning evaluations more closely with human preferences. We use GPT-4 to assess three aspects of the response on a 10-point scale: semantic accuracy, tone confidence, and overall coherence, with the scores summed to yield the GPT-Score (see Fig.1 (D)). For the VQA task, we introduce GPT-Acc, which leverages LLM-based judgment to evaluate response correctness. The judgment prompts are detailed in AppendixC." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 284, + 524, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 524, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 524, + 297 + ], + "type": "text", + "content": "4.2. Research Questions (RQs) and Findings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "text", + "content": "RQ1: Do current VLMs demonstrate sufficient robustness to 3D variations? Take-away: \"Emphatically not. Our evaluation reveals the severe vulnerability of current VLMs to worst-case 3D variations across multiple vision-language tasks.\" As shown in Tab. 1, in zero-shot classification, Adv-3DT samples generated by AdvDreamer significantly compromise VLMs' performance, causing accuracy drops of up to " + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "text", + "content": ". This substantially outperforms the Random baseline, validating AdvDreamer's effectiveness in capturing worst-case transformations. Notably, samples under " + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "inline_equation", + "content": "\\Theta^{*}" + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "text", + "content": " induce more severe performance degradation than those under the adversarial distribution " + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "inline_equation", + "content": "p^{*}(\\Theta)" + }, + { + "bbox": [ + 313, + 304, + 555, + 459 + ], + "type": "text", + "content": ", indicating the existence of extremely vulnerable points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": "This fundamental vulnerability extends beyond classification to more complex open-ended tasks. As illustrated in Tab.3, for image captioning, Adv-3DT samples under both " + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "inline_equation", + "content": "p^* (\\Theta)" + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\Theta^{*}" + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": " effectively degrade model performance across all evaluation metrics, with GPT-Score decreasing by nearly " + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": " across various VLMs. This significant degradation reveals a critical limitation in current VLMs' 3D understanding and reasoning capabilities. For VQA tasks, GPT4o, despite achieving the highest accuracy " + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "inline_equation", + "content": "(75.4\\%)" + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": " on clean samples—exhibits a substantial " + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 460, + 556, + 616 + ], + "type": "text", + "content": " accuracy drop when under Adv-3DT attacks. Fig. 1 (A) provides visualizations of representative Adv-3DT samples generated by AdvDreamer, with more examples presented in Appendix.D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "RQ2: To what extent do different VLMs share common 3D variation vulnerability regions? Take-away: \"Substantially. Our analysis reveals that Adv-3DT samples demonstrate strong cross-model transferability across various VLMs, indicating shared vulnerability regions in the 3D variation space.\" As demonstrated in Tab. 2, we investigate transferability by utilizing OpenCLIP and BLIP as surrogate models to optimize " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": ", then evaluate six" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7899" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 110, + 523, + 335 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 3. Performance Degradation (↓) of VLMs on Image Captioning and VQA Tasks. For closed-source VLMs (GPT-4o/4o-mini), we performed transfer-based attacks, using OpenAI-CLIP (ViT-L@336px) as surrogate encoder to optimize " + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": ". For other VLMs, " + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": " are optimized w.r.t. respective image encoders (e.g., Eva-CLIP and OpenAI CLIP, etc.). Metrics: C-CIDEr, B-BLUE@4." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 110, + 523, + 335 + ], + "lines": [ + { + "bbox": [ + 89, + 110, + 523, + 335 + ], + "spans": [ + { + "bbox": [ + 89, + 110, + 523, + 335 + ], + "type": "table", + "html": "
ModelMethodCOCO Caption [11]NoCaps [1]VQA-V2 [20]
CB@4GPT-ScoreCB@4GPT-ScoreGPT-Acc
BLIP-2 [30]Clean133.436.024.798.144.124.453.4
Random91.8 (↓42)23.5 (↓13)19.5 (↓5)60.2 (↓38)28.6 (↓16)20.9 (↓4)34.0 (↓19)
p*(Θ)83.8 (↓50)23.3 (↓13)18.9 (↓6)55.7 (↓42)26.9 (↓17)17.1 (↓7)31.7 (↓22)
Θ*72.2 (↓61)21.5 (↓15)17.0 (↓8)40.0 (↓58)22.4 (↓22)15.0 (↓9)28.0 (↓25)
LLaVa-1.5 [36]Clean126.130.425.0118.750.624.768.6
Random93.4 (↓33)25.6 (↓5)20.0 (↓5)75.3 (↓43)35.7 (↓15)21.7 (↓3)56.3 (↓12)
p*(Θ)85.8 (↓40)22.9 (↓8)18.9 (↓6)60.0 (↓59)29.6 (↓21)17.4 (↓7)54.3 (↓14)
Θ*73.2 (↓53)20.5 (↓10)16.5 (↓9)45.8 (↓73)25.5 (↓25)15.6 (↓9)53.8 (↓15)
LLaVa-1.6 [35]Clean41.66.922.930.510.822.671.2
Random25.1 (↓17)6.0 (↓1)17.5 (↓5)18.3 (↓12)7.1 (↓4)21.0 (↓2)54.3 (↓17)
p*(Θ)21.9 (↓20)5.3 (↓2)16.9 (↓6)13.6 (↓17)6.5 (↓4)15.7 (↓7)53.0 (↓18)
Θ*16.7 (↓25)4.6 (↓2)15.4 (↓8)10.5 (↓20)4.3 (↓7)13.5 (↓9)46.9 (↓24)
Otter [28]Clean127.034.924.076.429.722.553.9
Random71.6 (↓55)22.5 (↓12)21.0 (↓3)55.4 (↓21)23.7 (↓6)21.1 (↓1)49.7 (↓4)
p*(Θ)67.9 (↓59)17.8 (↓17)17.2 (↓7)45.4 (↓31)19.2 (↓11)16.2 (↓6)44.9 (↓9)
Θ*58.9 (↓68)17.5 (↓17)15.8 (↓8)36.0 (↓40)15.7 (↓14)14.6 (↓8)42.0 (↓12)
MiniGPT-4 [72]Clean78.523.424.060.828.524.751.3
Random58.2 (↓20)17.9 (↓6)19.7 (↓4)38.0 (↓23)19.8 (↓9)21.9 (↓3)43.7 (↓8)
p*(Θ)52.3 (↓26)16.3 (↓7)18.1 (↓6)32.9 (↓28)18.4 (↓10)16.7 (↓8)41.0 (↓10)
Θ*42.2 (↓36)12.9 (↓11)16.2 (↓8)26.6 (↓34)14.9 (↓14)15.0 (↓10)34.3 (↓17)
GPT-4o-mini [45]Clean23.73.023.315.72.923.068.6
Random14.1 (↓10)1.8 (↓1)16.9 (↓6)10.8 (↓5)0.9 (↓2)20.0 (↓3)51.7 (↓17)
p*(Θ)12.5 (↓11)1.2 (↓2)18.0 (↓5)7.8 (↓8)0.6 (↓2)17.2 (↓6)43.4 (↓25)
Θ*9.6 (↓14)1.6 (↓1)16.3 (↓7)7.3 (↓8)0.4 (↓3)15.5 (↓8)42.7 (↓26)
GPT-4o [45]Clean51.29.725.040.711.724.975.4
Random36.5 (↓15)5.1 (↓5)19.1 (↓6)24.7 (↓16)9.5 (↓2)20.9 (↓4)56.3 (↓19)
p*(Θ)26.4 (↓25)4.6 (↓5)17.7 (↓7)19.1 (↓22)6.5 (↓5)17.0 (↓8)50.7 (↓25)
Θ*20.3 (↓31)4.7 (↓5)16.1 (↓9)15.8 (↓25)5.4 (↓6)14.6 (↓10)50.3 (↓25)
", + "image_path": "112e95e535dc5a705d3bf28ae44ab8aef4ae74475c5c3df19a5a48f31e085a7f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 66, + 369, + 285, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "type": "text", + "content": "Table 4. Top-1 Accuracy and Naturalness Score of Adv-3DT samples generated by AdvDreamer w/ and w/o NRM. (under " + }, + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "type": "inline_equation", + "content": "p^{*}(\\Theta)" + }, + { + "bbox": [ + 55, + 344, + 295, + 367 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 369, + 285, + 487 + ], + "lines": [ + { + "bbox": [ + 66, + 369, + 285, + 487 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 285, + 487 + ], + "type": "table", + "html": "
MethodsTop-1 Acc.ScoreRScoreP
AdvDreamer w/o. NRM48.6% (↓49.4)1.601.39
AdvDreamer w/. NRM54.0% (↓44.0)2.522.57
", + "image_path": "45ea2226032cb952d532b1d73e6bbe929d56394b1a5d58b6447670aecacb80f2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 491, + 291, + 502 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 491, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 59, + 491, + 291, + 502 + ], + "type": "text", + "content": "Figure 5. Visualization of Adv-3DT samples w/ and w/o NRM." + } + ] + } + ], + "index": 4, + "type": "text" + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "content": "target VLMs with diverse architectures and training objectives. The transfer attacks maintain strong effectiveness, with performance degradation comparable to direct attacks. Notably, " + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "content": " exhibits smaller transfer gaps than " + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\Theta^*" + }, + { + "bbox": [ + 55, + 508, + 295, + 687 + ], + "type": "text", + "content": ", suggesting that learning the underlying distribution can better capture generalizable vulnerability regions. This transferability extends consistently to captioning and VQA tasks (Tab. 3). When employing OpenAI-CLIP as the surrogate encoder, Adv-3DT samples generated by AdvDreamer effectively compromise commercial systems like GPT-4o. We attribute this universal transferability to a fundamental limitation: current image-text pretraining datasets exhibit significant 3D biases and fail to encompass diverse certain special 3D transformations, leading to universal performance deficiencies under these transformation regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "RQ3: Does the performance degradation under Adv-3DT samples generated by AdvDreamer fundamentally" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "text", + "content": "arise from 3D variations? Take-away: \"Yes, we successfully reproduce these Adv-3DT samples in physical environments and demonstrate that degradation is inherently caused by the 3D variation rather than undesirable image quality problems.\" With the naturalness feedback provided by NRM, AdvDreamer effectively prevents unwanted degradation of samples during the optimization process. To verify this, we evaluate Adv-3DT samples generated by AdvDreamer with and without NRM under ImageNet. Tab. 4 presents the accuracy under OpenCLIP (ViT-B/16), alongside GPT-4o-assessed visual realism " + }, + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "inline_equation", + "content": "(Score_R)" + }, + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "text", + "content": " and physical plausibility " + }, + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "inline_equation", + "content": "(Score_P)" + }, + { + "bbox": [ + 313, + 346, + 555, + 586 + ], + "type": "text", + "content": ". The results indicate that NRM significantly improves samples' realism and physical feasibility, with a marginal trade-off in aggressiveness. Furthermore, Fig. 5 presents a qualitative comparison, illustrating the effectiveness of NRM's naturality feedback in improving the image quality and mitigating the distribution discrepancy compared to natural image distributions. In addition, we validate the alignment between NRM predictions and human evaluations, as detailed in Appendix C.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, we conduct physical-world experiments on 12 common objects from traffic and household environments to validate this. Our experimental procedure consisted of three main steps: 1) capturing images of objects in their natural state, 2) using AdvDreamer with OpenCLIP as the surrogate model to generate Adv-3DT samples, and 3) physically reproducing the adversarial 3D transformations captured by AdvDreamer through video recording. This process yielded 3,807 frames of physical Adv-3DT samples. We evaluate effectiveness through zero-shot classifi" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7900" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 87, + 500, + 140 + ], + "blocks": [ + { + "bbox": [ + 208, + 71, + 402, + 83 + ], + "lines": [ + { + "bbox": [ + 208, + 71, + 402, + 83 + ], + "spans": [ + { + "bbox": [ + 208, + 71, + 402, + 83 + ], + "type": "text", + "content": "Table 5. Quantitative results of physical experiments." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 87, + 500, + 140 + ], + "lines": [ + { + "bbox": [ + 109, + 87, + 500, + 140 + ], + "spans": [ + { + "bbox": [ + 109, + 87, + 500, + 140 + ], + "type": "table", + "html": "
SampleOpenCLIPLLaVa-1.5GPT-4o
Acc.Conf.Acc.1Acc.2Acc.1Acc.2
Natural (Physical)100.00.89283.310091.7100.0
Adv-3DT (AdvDreamer)0.0 (↓100)0.00325.0 (↓58)25.0 (↓75)25.0 (↓67)41.7 (↓58)
Adv-3DT (Physical)51.3 (↓49)0.41333.6 (↓50)45.1 (↓55)56.8 (↓35)64.8 (↓35)
", + "image_path": "2ff6655d3561980f8a81695232f81e1255ecf3c5b12e0fbb3b26bca2964e78de.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 154, + 174, + 229 + ], + "blocks": [ + { + "bbox": [ + 63, + 154, + 174, + 229 + ], + "lines": [ + { + "bbox": [ + 63, + 154, + 174, + 229 + ], + "spans": [ + { + "bbox": [ + 63, + 154, + 174, + 229 + ], + "type": "image", + "image_path": "f9a6bf5ad1d1e92827a403764db9ef89035254f643abf19fcf555156ba945ea5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "type": "text", + "content": "Figure 6. Attack success rate of Adv-3DT samples w.r.t. (A) optimization steps and (B) scaling ratios of the optimal distribution." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 171, + 153, + 284, + 228 + ], + "blocks": [ + { + "bbox": [ + 171, + 153, + 284, + 228 + ], + "lines": [ + { + "bbox": [ + 171, + 153, + 284, + 228 + ], + "spans": [ + { + "bbox": [ + 171, + 153, + 284, + 228 + ], + "type": "image", + "image_path": "823610e126a0141de068c3a144c0f16c5153d0b618c8811fc9b075a9a60c749a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "text", + "content": "cation and VQA tasks. For VQA, we designed two evaluation protocols: multiple-choice classification (Acc.1) and binary verification (Acc.2). As shown in Table 5, physically reproduced Adv-3DT samples significantly degraded VLM performance: the accuracy of zero-shot classification dropped to " + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "inline_equation", + "content": "51.3\\%" + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "text", + "content": ", while VQA performance decreased to " + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "inline_equation", + "content": "33.6\\%" + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "text", + "content": " (Acc.1) and " + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "inline_equation", + "content": "45.1\\%" + }, + { + "bbox": [ + 54, + 274, + 295, + 441 + ], + "type": "text", + "content": " (Acc.2). These results confirm that AdvDreamer successfully captures physically realizable 3D variations that represent real world corner cases. However, an aggressiveness gap between digital and physical attacks is observed, primarily attributed to data distribution differences and camera parameter fluctuations. Fig. 1 (B) presents representative examples from physical experiments, with more detailed results provided in Appendix D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 272, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 272, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 272, + 466 + ], + "type": "text", + "content": "4.3. Ablation Studies and Additional Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": "Convergence Analysis & Computational Efficiency. We analyze the attack success rates (ASR) of " + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 472, + 295, + 604 + ], + "type": "text", + "content": " against OpenCLIP (ViT-B/16) and BLIP (ViT-B/16) across different optimization iterations. As shown in Fig. 6, AdvDreamer's effectiveness consistently improves with more iterations and converges after 15 steps. Thus, we set the iteration step default to 15 throughout our experiments to balance effectiveness and efficiency. The computational cost is discussed in Appendix C.4 by detailing the overhead of each component. For each clean sample, AdvDreamer requires an average of 0.28 GPU hours with a single RTX 3090." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Does Optimal Distribution " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " Effectively Characterize the Worst-case 3D Variations? We conduct analysis on ImageNet using five random natural images. For each image, we generate 100 Adv-3DT samples by scaling the variance " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " of their " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "p^*(\\Theta)" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " with different ratios. Fig. 6(B) shows that ASR consistently decreases as the sampling range expands beyond the optimal distribution. This inverse relationship confirms that AdvDreamer successfully identifies the most vulnerable regions in the 3D variation space." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 152, + 547, + 219 + ], + "blocks": [ + { + "bbox": [ + 317, + 152, + 547, + 219 + ], + "lines": [ + { + "bbox": [ + 317, + 152, + 547, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 547, + 219 + ], + "type": "image", + "image_path": "1fc00b9da4be52c7ea4a64396ae900b607f1ac28a22d41b60ea2eecf73e39f57.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 222, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 233 + ], + "type": "text", + "content": "Figure 7. Benchmarking result of VLMs' 3D variation robustness." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 408, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 408, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 408, + 252 + ], + "type": "text", + "content": "4.4. MM3DTBench" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": "Finally, We introduce MM3DTBench, it comprises 215 challenging Adv-3DT samples, consisting of samples generated by AdvDreamer and their physical replications. For each sample, we provide question templates with candidate answers and precise orientation annotations for the transformed objects. We employ two evaluation tasks: multiple-choice (Choice) and free-form description (Free answer), with accuracy computed through text-only GPT-4 judgment. More Details are provided in the Appendix.E. Our evaluation of 13 representative VLMs reveals robustness gaps. As shown in Fig. 7, even top-performing models (GPT-4o [45], CogVLM [61], and Intern-VL [13]) achieve limited success, while most models struggle with accuracy below " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": ". Notably, Claude-3 [5], despite its recognized excellence, only achieves an accuracy of " + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "inline_equation", + "content": "30.0\\%" + }, + { + "bbox": [ + 313, + 258, + 555, + 498 + ], + "type": "text", + "content": ". This performance correlates with models' general out-of-distribution robustness measured by MultiTrust [67], suggesting a fundamental limitation of current VLMs. We encourage the adoption of MM3DTBench in VLM development to advance security evaluation in dynamic scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 555, + 647 + ], + "type": "text", + "content": "We proposed AdvDreamer, a novel framework for characterizing real-world adversarial 3D transformation samples from single-view observations. Our comprehensive evaluation revealed critical robustness gaps in existing VLMs, highlighting the urgent need to enhance VLMs' 3D variation perception and understanding capabilities in safety-critical applications. Through MM3DTBench, we established the first benchmark for worst-case 3D variations, aiming to facilitate the development of more robust vision-language systems for real-world deployment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "This work was supported by the Fundamental Research Funds for the Central Universities, and was also supported by NSFC Projects (U2341228, 62276149)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "7901" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 212 + ], + "type": "text", + "content": "[2] Haider Al-Tahan, Quentin Garrido, Randall Balestriero, Diane Bouchacourt, Caner Hazirbas, and Mark Ibrahim. Unibench: Visual reasoning requires rethinking vision-language beyond scaling. arXiv preprint arXiv:2408.04810, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 295, + 277 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 345 + ], + "type": "text", + "content": "[4] Michael A Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4845-4854, 2019. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 346, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 367 + ], + "type": "text", + "content": "[5] Anthropic. claude-3. https://www.anthropic.com/claude, 2024.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 411 + ], + "type": "text", + "content": "[6] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International conference on machine learning, pages 284-293. PMLR, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 457 + ], + "type": "text", + "content": "[7] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 523 + ], + "type": "text", + "content": "[8] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 590 + ], + "type": "text", + "content": "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "text", + "content": "[10] Liangyu Chen, Bo Li, Sheng Shen, Jingkang Yang, Chunyuan Li, Kurt Keutzer, Trevor Darrell, and Ziwei Liu. Large language models are visual reasoning coordinators. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[11] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level im" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 106 + ], + "type": "text", + "content": "age customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 173 + ], + "type": "text", + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 239 + ], + "type": "text", + "content": "[14] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 553, + 285 + ], + "type": "text", + "content": "[15] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. In ICLR 2023 (Eleventh International Conference on Learning Representations), 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 341 + ], + "type": "text", + "content": "[16] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 553, + 386 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 553, + 441 + ], + "type": "text", + "content": "[18] Yinpeng Dong, Shouwei Ruan, Hang Su, Caixin Kang, Xingxing Wei, and Jun Zhu. Viewfool: Evaluating the robustness of visual recognition to adversarial viewpoints. In Neural Information Processing Systems (NeurIPS), 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 498 + ], + "type": "text", + "content": "[19] Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1487-1495, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 499, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 499, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 499, + 553, + 553 + ], + "type": "text", + "content": "[20] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "text", + "content": "[21] Abdullah Hamdi and Bernard Ghanem. Towards analyzing semantic robustness of deep neural networks. In ECCV, pages 22-38. Springer, 2020. 2, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 611 + ], + "type": "text", + "content": "[22] Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772, 2016. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "text", + "content": "[23] Zexin He and Tengfei Wang. Openlrm: Open-source large reconstruction models. https://github.com/3DTopia/OpenLRM, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7902" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. openclip, 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "text", + "content": "[27] LAION-AI. Clip-benchmark. https://github.com/LAION-AI/CLIP_benchmark, 2024.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[28] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Fanyi Pu, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Mimicit: Multi-modal in-context instruction tuning. arXiv preprint arXiv:2306.05425, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 330 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[32] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 499 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 587 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR, and world knowledge, 2024. 2, 3, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 678 + ], + "type": "text", + "content": "[37] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[38] Jiahuan Long, Tingsong Jiang, Wen Yao, Shuai Jia, Weijia Zhang, Weien Zhou, Chao Ma, and Xiaoqian Chen. Papmot: Exploring adversarial patch attack against multiple ob" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ject tracking. In European Conference on Computer Vision, pages 128-144. Springer, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "text", + "content": "[39] Jiahuan Long, Zhengqin Xu, Tingsong Jiang, Wen Yao, Shuai Jia, Chao Ma, and Xiaoqian Chen. Robust sam: On the adversarial robustness of vision foundation models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5775-5783, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 154, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 197 + ], + "type": "text", + "content": "[40] Jiahuan Long, Wen Yao, Tingsong Jiang, and Chao Ma. Cdupatch: Color-driven universal adversarial patch attack for dual-modal visible-infrared detectors. arXiv preprint arXiv:2504.10888, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "text", + "content": "[41] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 247, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 247, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 247, + 553, + 312 + ], + "type": "text", + "content": "[42] Spandan Madan, Timothy Henry, Jamell Dozier, Helen Ho, Nishchal Bhandari, Tomotake Sasaki, Frédo Durand, Hanspeter Pfister, and Xavier Boix. When and how convolutional neural networks generalize to out-of-distribution category-viewpoint combinations. Nature Machine Intelligence, 4(2):146-153, 2022. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "text", + "content": "[43] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "text", + "content": "[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 407, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 427 + ], + "type": "text", + "content": "[45] openai. Gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024.3,7,8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "text", + "content": "[46] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 552 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "text", + "content": "[48] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv preprint arXiv:2401.14159, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[50] Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, and Xingxing Wei. Towards viewpoint-invariant visual recognition via adversarial training. In ICCV, pages 1-10, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7903" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[51] Shouwei Ruan, Yinpeng Dong, Hang Su, Jianteng Peng, Ning Chen, and Xingxing Wei. Improving viewpoint robustness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[52] Shouwei Ruan, Yinpeng Dong, Hanqing Liu, Yao Huang, Hang Su, and Xingxing Wei. Omniview-tuning: Boosting viewpoint invariance of vision-language pre-training models. arXiv preprint arXiv:2404.12139, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 229 + ], + "type": "text", + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[55] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "text", + "content": "[56] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[57] Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, , Adam Letts, Yangguang Li, Ding Liang, Christian Laforte, Varun Jampani, and Yan-Pei Cao. Triposr: Fast 3d object reconstruction from a single image. arXiv preprint arXiv:2403.02151, 2024. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 443 + ], + "type": "text", + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[60] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 601 + ], + "type": "text", + "content": "[61] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 656 + ], + "type": "text", + "content": "[62] Hu Xu, Saining Xie, Xiaqing Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. In The Twelfth International Conference on Learning Representations, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[63] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 552 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[64] Naoki Yokoyama, Sehoon Ha, Dhruv Batra, Jiuguang Wang, and Bernadette Bucher. Vlfm: Vision-language frontier maps for zero-shot semantic navigation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 42-48. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[65] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Scaling in-the-wild training for diffusion-based illumination harmonization and editing by imposing consistent light transport. In The Thirteenth International Conference on Learning Representations, 2025. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 284 + ], + "type": "text", + "content": "[67] Yichi Zhang, Yao Huang, Yitong Sun, Chang Liu, Zhe Zhao, Zhengwei Fang, Yifan Wang, Huanran Chen, Xiao Yang, Xingxing Wei, Hang Su, Yinpeng Dong, and Jun Zhu. Benchmarking trustworthiness of multimodal large language models: A comprehensive study, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 328 + ], + "type": "text", + "content": "[68] Yabin Zhang, Wenjie Zhu, Chenhang He, and Lei Zhang. Lapt: Label-driven automated prompt tuning for ood detection with vision-language models. Proceedings of the European conference on computer vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 396 + ], + "type": "text", + "content": "[69] Bingchen Zhao, Shaozuo Yu, Wufei Ma, Mingxin Yu, Shenxiao Mei, Angtian Wang, Ju He, Alan Yuille, and Adam Kortylewski. Ood-cv: A benchmark for robustness to out-of-distribution shifts of individual nuisances in natural images. In European conference on computer vision, pages 163-180. Springer, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 553, + 452 + ], + "type": "text", + "content": "[70] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Chongxuan Li, Ngai-Man Cheung, and Min Lin. On evaluating adversarial robustness of large vision-language models. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "text", + "content": "[71] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[72] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 6, 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7904" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_content_list.json b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5e777f4f4a48e3e987ceb03406c9c7a20f2e8415 --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_content_list.json @@ -0,0 +1,2261 @@ +[ + { + "type": "text", + "text": "Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation", + "text_level": 1, + "bbox": [ + 233, + 130, + 763, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu Lei $^{1}$ Bingde Liu $^{1}$ Qingsong Xie $^{2}$ Haonan Lu $^{2}$ Zhijie Deng $^{1*}$ $^{1}$ Shanghai Jiao Tong University OPPO AI Center \n{tony-lei, bingde, zhijied}@sjtu.edu.cn {xieqingsong1, luhaonan}@oppo.com", + "bbox": [ + 158, + 202, + 834, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-3D generation based on score distillation of pretrained 2D diffusion models has gained increasing interest, with variational score distillation (VSD) as a remarkable example. VSD proves that vanilla score distillation can be improved by introducing an extra score-based model, which characterizes the distribution of images rendered from 3D models, to correct the distillation gradient. Despite the theoretical foundations, VSD, in practice, is likely to suffer from slow and sometimes ill-posed convergence. In this paper, we perform an in-depth investigation of the interplay between the introduced score model and the 3D model, and find that there exists a mismatching problem between LoRA and 3D distributions in practical implementation. We can simply adjust their optimization order to improve the generation quality. By doing so, the score model looks ahead to the current 3D state and hence yields more reasonable corrections. Nevertheless, naive lookahead VSD may suffer from unstable training in practice due to the potential over-fitting. To address this, we propose to use a linearized variant of the model for score distillation, giving rise to the Linearized Lookahead Variational Score Distillation ( $L^2$ -VSD). $L^2$ -VSD can be realized efficiently with forward-mode autodiff functionalities of existing deep learning libraries. Extensive experiments validate the efficacy of $L^2$ -VSD, revealing its clear superiority over prior score distillation-based methods. We also show that our method can be seamlessly incorporated into any other VSD-based text-to-3D framework.", + "bbox": [ + 91, + 323, + 483, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 775, + 220, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D content creation is important for a variety of applications, such as interactive gaming [1, 32], cinematic arts [3], AR/VR [4, 16], and building simulated environments for training agents in robotics [26]. However, it is still challenging and expensive to create high-quality 3D assets as", + "bbox": [ + 89, + 801, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "it requires a high level of expertise. Therefore, automating this process with generative models has become an important problem [12], while remaining non-trivial due to the scarcity of data and complexity of 3D representations.", + "bbox": [ + 511, + 292, + 906, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Score distillation has emerged as an attractive way for 3D generation given textual condition [2, 17, 21, 28, 29]. It leverages pretrained 2D diffusion models [9, 22] to define priors to guide the evolution of 3D content without reliance on annotations. Score Distillation Sampling (SDS) [21] is a seminal work in this line, but it is widely criticized that its generations suffer from the over-smoothing issue. Variational Score Distillation (VSD) [29] remediates this by introducing an extra model that captures the score of the images rendered from the 3D model to correct the distillation gradient. However, VSD often requires a lengthy optimization through 3 stages: NeRF generation, geometry refinement, and texture refinement. The outcomes obtained in the initial stage are often blurry, prone to collapsing, and not directly applicable [30]. Though existing works begin to understand and improve SDS [13, 28, 34], there are great but fewer efforts dedicated to improving the more promising VSD [18, 31].", + "bbox": [ + 511, + 354, + 908, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To identify the root of VSD's drawback, we conduct a comprehensive analysis of the interaction between the introduced score model and the 3D model, revealing that adjusting their optimization order can sometimes lead to a considerable enhancement in generation quality. This adjustment allows the score model to look ahead to the current 3D state, resulting in more accurate and sensible corrections for the distillation gradient. Yet, naive lookahead VSD can encounter unstable training due to the risk of the score model overfitting the single 3D particle and sampled camera view.", + "bbox": [ + 511, + 627, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, we formally compare the correction gradients before and after looking ahead and identify two major differences—a linear first-order term and a high-order one. Upon closer examination, we observe that the former accommodates subtle semantic information, whereas the latter contains non-trivial high-frequency noises. Given these findings, we propose to use only the first-order term for correction, yielding Linearized Looka", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19567", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d3b0fc460214293e8ba038790c5abc4bb9cca111bc6107f6bb0624b8873a71f.jpg", + "image_caption": [ + "SDS" + ], + "image_footnote": [], + "bbox": [ + 96, + 104, + 245, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9c257d2bd731145e8f678f672749af66cdf9a823db8599f65355d4987c3836b2.jpg", + "image_caption": [ + "ESD" + ], + "image_footnote": [], + "bbox": [ + 259, + 104, + 408, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b7a1ae5eb4f969194dd17ebda8ab3b49d85c0c1bc4e538608c41a9ed7a98baef.jpg", + "image_caption": [ + "VSD" + ], + "image_footnote": [], + "bbox": [ + 449, + 117, + 558, + 170 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7bdd383f924c808728d508a0a327640088326bb1677f9dce145370bd2686670.jpg", + "image_caption": [ + "VSD-HiFA" + ], + "image_footnote": [], + "bbox": [ + 584, + 104, + 732, + 180 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3fa276734451d37c268a5e4060133dcb32582ead2a61833fb560d860697db2c8.jpg", + "image_caption": [ + "$L^{2}\\mathrm{VSD}$" + ], + "image_footnote": [], + "bbox": [ + 748, + 104, + 893, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94b1de41d1b4198ade1a568fa248a561e45a696fbded2f940383fe496d73e305.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 203, + 243, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30d2c53ba9303ea2ec18ee6edf1a1669737bdad2d64da669396e7045fd6b708d.jpg", + "image_caption": [ + "\"A DSLR photo of a fox holding a videogame controller\"" + ], + "image_footnote": [], + "bbox": [ + 261, + 203, + 408, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e6a91a5ec9de5b9b9102fdf14fed87701df8885ac2e09f86e0f6bcfefef00c19.jpg", + "image_caption": [ + "\"A hamburger\"" + ], + "image_footnote": [], + "bbox": [ + 424, + 203, + 571, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1f63c0f875be4d0c5ce2c4ca3c9d3ff21bec8ea9a66e1b0a2cc465487d6e3511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 202, + 733, + 279 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1aa008b842e50efdcd111ccd4fedc516e1fdc4ea0080240cc8cd03b85428b6a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 202, + 893, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f079dd78bc6a074ef105de3d9bdba45aef52d6a02ec9b93a1ce64f009f74e8d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 301, + 241, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/16aa64529646b3a4aa3eca8db2f4f2a9ccef431c93a33ab6a63b61f97c7691c8.jpg", + "image_caption": [ + "\"A humanoid robot lying on the couch while on a laptop\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 301, + 406, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9544160126ef40a311aa50af5a5ca1466822f2cfc4655bbcd52a2a62770247f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 301, + 570, + 377 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eba0e145701bd819cfc5d19d00cfe4531f91560f9689ddeb36b9ae2b3b5faa58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 303, + 730, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fcada5bb2068ed860bfefe1adfd040490d64d8a11379bec59afff9f9be9e3f56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 303, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ba5a94cc9db707b643889fcd293ebd7091260ff2fcd96f37ed2bc190fee0384a.jpg", + "image_caption": [ + "Figure 1. Qualitative comparison with high resolution of 256. $L^2$ -VSD can generate meticulously detailed and realistic 3D assets from easy to very complex prompts and scenes. $L^2$ -VSD tends to yield more complete structures than HiFA." + ], + "image_footnote": [], + "bbox": [ + 98, + 398, + 243, + 474 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/97df17223b69a826bbb461f5daf914b8d85383b1e69e63df0e3dd88b42f490da.jpg", + "image_caption": [ + "\"A zoomed out DSLR photo of Sydney opera house, aerial view\"" + ], + "image_footnote": [], + "bbox": [ + 264, + 398, + 410, + 474 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9baf8e025a77ef1e889062b180035f24386c3f71d22daa6a854f5100907ab0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 398, + 573, + 474 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/63efc9f36f977474e9efc9ae69012835a08a54025bd3cc76d8f1069a107d0941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 398, + 733, + 474 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f8d964175c6c116b3d5417f5203aba0d50ce28a19af22acc6b4364b78487881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 400, + 897, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "head Variational Score Distillation ( $L^2$ -VSD), to reliably and consistently boost the generation quality of VSD. $L^2$ -VSD is both easy to implement and computationally efficient—the added linear term can be computed by only one additional forward process of the score model under the scope of forward-mode automatic differentiation, which is supported in many deep learning libraries [14, 20].", + "bbox": [ + 89, + 536, + 482, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through extensive experiments, we demonstrate the significant superiority of the proposed $L^2$ -VSD in improving 3D generation quality compared to competing baseline methods, as shown in Fig. 1. $L^2$ -VSD can even produce realistic results with low resolution directly in the first stage. Moreover, we empirically show that $L^2$ -VSD can be seamlessly integrated into other VSD-based 3D generation pipelines and combined with other parallel techniques for VSD, e.g., Entropy Score Distillation(ESD) [28], which mitigates the Janus problem, for further improvement. We summarize our technical contributions as follows:", + "bbox": [ + 89, + 643, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For VSD, a fundamental method in text-to-3D generation, we carefully identify the gaps between its theory and implementation and analyze the potential impact the gaps may bring us, providing a direction for possible improvements.", + "- We propose $L^2$ -VSD, an easy-to-implement and compu" + ], + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/706c2a086a603e2d319cc023547327e7afed13a0b2123ccf8b935b9f30bfc12f.jpg", + "image_caption": [ + "Figure 2. Comparison of convergence between VSD and L-VSD with an illustrative 2D Gaussian example. In this toy example we consider $x \\in \\mathbb{R}^2$ from a single Gaussian distribution. We optimize $q(x)$ towards ground truth distribution $p(x)$ and use $r(x)$ to function as LoRA, which is used to estimate $q(x)$ . This example validates the existence of a mismatching issue in VSD, and we leave the details in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 519, + 536, + 705, + 638 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d9cfbf43f5fd6dbb0084bf9f572e453363e268b31c2b0b3770b0bbe967d7233.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 536, + 901, + 638 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "tationally efficient variant of VSD, which mitigates the mismatching problem to some extent. We demonstrate its significant improvement over baselines and are comparable to other state-of-the-art.", + "- We demonstrate that our method can be seamlessly combined with other VSD-based improving techniques without much effort." + ], + "bbox": [ + 513, + 800, + 903, + 905 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19568", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 91, + 89, + 220, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 113, + 259, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models (DMs) [9, 23, 24] are defined with a forward diffusion process on data $x \\in \\mathbb{R}^d$ with a Gaussian transition kernel. The conditional distribution at some timestep $t \\in [0, T]$ usually satisfies", + "bbox": [ + 89, + 135, + 483, + 198 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 207, + 482, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_0 \\coloneqq x$ , and $\\alpha_{t}$ and $\\sigma_{t}$ are pre-defined noise schedules. DMs learn a reverse diffusion process, specified by a parameterized distribution $p(x_{t - 1}|x_t)\\coloneqq \\mathcal{N}(x_{t - 1};\\mu_\\psi (x_t,t),\\sigma_t^2\\mathbf{I})$ with $\\mu_{\\psi}$ as a neural network (NN), to enable the sampling of generations from Gaussian noise. The training objective of $\\mu_{\\psi}$ is the variational lower bound of the log data likelihood. In practice, $\\mu_{\\psi}$ is reparameterized as a denoising network $\\epsilon_{\\psi}$ and the training loss can be further simplified as a Mean Squared Error (MSE) form [9, 15]:", + "bbox": [ + 89, + 234, + 483, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D i f f} := \\mathbb {E} _ {x, t, \\epsilon} [ \\omega (t) | | \\epsilon_ {\\psi} (\\alpha_ {t} x + \\sigma_ {t} \\epsilon , t) - \\epsilon | | _ {2} ^ {2} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 396, + 482, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ follows the data distribution, $t$ is uniformly drawn from $[0,T]$ , $\\epsilon$ is a standard Gaussian noise, and $\\omega (t)$ is a time-dependent coefficient. $\\epsilon_{\\psi}$ also inherently connects to score matching [24, 27].", + "bbox": [ + 89, + 422, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance (CFG) [8]. We can augment the model $\\epsilon_{\\psi}$ with an extra input, the condition $y$ , to characterize the corresponding conditional distribution, leaving $\\epsilon_{\\psi}(x_t,t,\\emptyset)$ account for the original unconditional one. Then, we can resort to CFG to further boost the quality of conditional generation. Typically, CFG leverages the following term in the sampling process:", + "bbox": [ + 89, + 484, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\epsilon} _ {\\psi} \\left(x _ {t}, t, y\\right) := (1 + s) \\epsilon_ {\\psi} \\left(x _ {t}, t, y\\right) - s \\epsilon_ {\\psi} \\left(x _ {t}, t, \\emptyset\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 601, + 482, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s > 0$ refers to a guidance scale.", + "bbox": [ + 89, + 628, + 349, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-3D Generation with Score Distillation", + "text_level": 1, + "bbox": [ + 89, + 651, + 475, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D generation aims to identify the parameters $\\theta \\in \\mathbb{R}^N$ of a 3D model given a text condition $y$ . Neural radiance field (NeRF) [19] is a typical 3D representation based on neural networks. In particular, NeRF renders a new view of the scene with the input of a sequence of images as known views. Additionally, textured mesh can be applied to represent the geometry of a 3D object with triangle meshes and textures with color on the mesh surface.", + "bbox": [ + 89, + 672, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denote $g(\\theta, c)$ as the differential rendering function projecting the 3D scene to a 2D image given a camera angle $c$ . Score distillation approaches for text-to-3D generations demand the image sample $g(\\theta, c)$ to respect the prior specified by a text-to-2D diffusion model $\\epsilon_{\\text{pretrain}}(\\cdot, \\cdot, y)$ pretrained on vast real text-image pairs, based on which the optimization goal is constructed [13, 21, 25, 28, 29, 33, 34].", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation Sampling (SDS) [21] updates the 3D model using view-dependent prompt $y^{c}$ :", + "bbox": [ + 513, + 90, + 906, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\theta) := \\mathbb {E} _ {t, \\epsilon , c} [ \\omega (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y ^ {c}\\right) - \\epsilon\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 131, + 913, + 172 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c$ is a randomly sampled camera angle and $x_{t} \\coloneqq \\alpha_{t} g(\\theta, c) + \\sigma_{t} \\epsilon$ . The gradient is a simplification of that of the denoising objective w.r.t. $\\theta$ [21]. Intuitively, it encourages the rendered images to move toward the high-probability regions of the pretrained model, thus a good 3D model emerges. SDS is a seminal work in the line of text-to-3D generation, but it is sensitive to the CFG scale $s$ [8]. A small $s$ often results in over-smooth outcomes, whereas a large $s$ leads to over-saturation.", + "bbox": [ + 511, + 174, + 906, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Score Distillation (VSD) [29] addresses the issue of SDS with a thorough theoretical analysis and proposes a new algorithm. In particular, apart from the pretrained model $\\epsilon_{\\text{pretrain}}(x_t, t, y)$ for capturing the data distribution, VSD introduces a tunable model $\\epsilon_{\\phi}(x_t, t, c, y)$ , often instantiated as a LoRA [10] adaptation of $\\epsilon_{\\text{pretrain}}$ , to account for the distribution of images rendered from all possible 3D models given condition $y$ and camera angle $c$ . We refer the introduced model as the LoRA model hereinafter. VSD proves the LoRA model can reliably correct the original distillation gradient.", + "bbox": [ + 511, + 309, + 905, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VSD, in practice, usually considers only a single 3D particle $\\theta$ and performs an iterative optimization of $\\theta$ and $\\phi$ until convergence. More specifically, let $\\theta_{i}$ and $\\phi_{i}$ denote the parameters at $i$ -th training iteration. VSD updates $\\theta_{i}$ with the following gradient:", + "bbox": [ + 511, + 476, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {V S D} (\\theta_ {i}) := \\mathbb {E} _ {t, \\epsilon , c}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 561, + 674, + 578 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ w (t) \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 580, + 905, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{t} = \\alpha_{t}g(\\theta_{i},c) + \\sigma_{t}\\epsilon$ . To avoid more than once differentiable rendering of the 3D model for efficiency, VSD updates $\\phi_{i}$ still with $g(\\theta_i,c)$ while using a different noisy state $x_{t^{\\prime}} = \\alpha_{t^{\\prime}}g(\\theta_{i},c) + \\sigma_{t^{\\prime}}\\epsilon^{\\prime}$ . The learning objective is the denoising loss defined in Eq. (2), whose gradient is:", + "bbox": [ + 511, + 627, + 905, + 703 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\phi_ {i}} \\mathcal {L} _ {V S D} (\\phi_ {i}) := \\tag {6} \\\\ \\mathbb {E} _ {t ^ {\\prime}, \\epsilon^ {\\prime}, c} \\left[ \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 713, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J_{\\phi_i}(x_{t'}, t', c, y) \\coloneqq \\frac{\\partial \\epsilon_\\phi(x_{t'}, t', c, y)}{\\partial \\phi}|_{\\phi = \\phi_i}$ and we omit some time-dependent scaling factor.", + "bbox": [ + 511, + 760, + 905, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Apart from the compromise to maintain one 3D particle for efficiency, the practical algorithm of VSD exhibits several significant gaps from the theory: 1) one-step update cannot guarantee $\\phi$ to converge in each iteration, and 2) the updates to $\\theta_{i}$ are computed given the LoRA model $\\phi_{i}$ , which characterizes the distribution associated with the previous 3D state $\\theta_{i - 1}$ instead of $\\theta_{i}$ . We hypothesize these", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19569", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9cb847eda96a244646352f071dea1534af88495da9eca23321e0e163aae99273.jpg", + "image_caption": [ + "(a) VSD: From left to right, the results correspond to setting $\\gamma$ to 1, 2, 5, and 10. We can observe that the quality of the model is not fundamentally improved." + ], + "image_footnote": [], + "bbox": [ + 91, + 75, + 493, + 154 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d2e006c6201119b322950e3a20ab0e64c286780b8e224401eab014d5c8ddad18.jpg", + "image_caption": [ + "(b) L-VSD: From left to right, the first three results correspond to setting $\\gamma$ to 1, 2, and 5; the last corresponds to scaling the learning rate by 0.1 with $\\gamma = 1$ .", + "Figure 3. Qualitative Examples of training LoRA model for multiple steps per optimization iteration for VSD and L-VSD." + ], + "image_footnote": [], + "bbox": [ + 500, + 75, + 901, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "are probably the root of the unstable performance and sometimes corrupted outcomes of VSD, and perform an in-depth investigation regarding them below.", + "bbox": [ + 89, + 220, + 485, + 267 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Diagnose the Issues of VSD", + "text_level": 1, + "bbox": [ + 89, + 281, + 344, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess whether the identified gaps contribute to the issues of VSD, we conduct two sets of pivot experiments in this section. As introduced in Sec 2.2, we analyze the impact of convergence of LoRA in Sec 3.1; then, we correct the optimization order to see the consequences in Sec 3.2; lastly, we combine these two factors to check if they are interfered in Sec 3.3. We provide an illustrative 2D Gaussian example for better understanding and to evaluate the effectiveness of lookahead in Sec 3.2. We base the implementation on the open-source framework threestudio [5].", + "bbox": [ + 89, + 306, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Better Convergence? Maybe No.", + "text_level": 1, + "bbox": [ + 89, + 469, + 377, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 1 The one-step update in vanilla VSD cannot guarantee the LoRA model to converge well, thus possibly harming the effectiveness of score distillation. We assume updating the LoRA model for $\\gamma (\\gamma >1)$ steps in each iteration could alleviate the pathology.", + "bbox": [ + 89, + 491, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experiments Setting We evaluate the effects of various $\\gamma$ , including 1, 2, 5, and 10. In each step, we optimize the LoRA model with different noisy images under different views. Unless otherwise specified, we take NeRF as the default 3D representation and use a simple prompt, \"a delicious hamburger\", in the study. We keep all other hyperparameters the same as the vanilla VSD. Usually, the whole VSD process contains multiple stages, where in the first stage a NeRF is constructed and then the geometry and texture are refined, respectively. We directly report the first-stage learning outcomes because they establish the foundation for the following parts.", + "bbox": [ + 89, + 568, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Findings 1 We present the training loss of the LoRA model in Fig. 4a to indicate the convergence and the final generations in Fig. 3a to reflect if the issues still exist. As shown, although the loss curves for various $\\gamma$ share a periodic rise and fall, the loss of a larger $\\gamma$ (e.g., 5 or 10) is floating in a relatively smaller range, and $\\gamma = 5$ roughly makes LoRA model converge. However, as shown in Fig. 3a, it's hard to tell if the shape becomes relatively more reasonable, and the overall quality of the 3D model does not witness a continual improvement as $\\gamma$ rises. Thereby, we conclude", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that improving the convergence of the LoRA model on the original VSD is not sufficient, thus not being the key factor.", + "bbox": [ + 511, + 220, + 903, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Lookahead? Maybe Yes!", + "text_level": 1, + "bbox": [ + 511, + 258, + 741, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assumption 2 As shown in Eq. (5), the updates to $\\theta_{i}$ are computed given $\\phi_{i}$ , which characterizes the distribution associated with $\\theta_{i-1}$ instead of $\\theta_{i}$ . This is inconsistent with Theorem 2 of [29], where the LoRA model should first adapt to the current 3D model (i.e., $\\theta_{i}$ ) to serve as a reliable score estimator for the corresponding distribution. Fixing such a mismatch may address the issues of VSD.", + "bbox": [ + 511, + 281, + 903, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experiments Setting We update $\\phi_{i}$ first to obtain $\\phi_{i + 1}$ , based on which $\\theta_{i}$ is updated. We name such a modification Lookahead-VSD (L-VSD) because it makes the LoRA model look ahead for one step compared to the original VSD. All other parameters are kept unchanged.", + "bbox": [ + 511, + 387, + 903, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Findings 2 We show the result in Fig. 3b. Intuitively, the rendered image has clearer edges compared to those in Fig. 3a. Besides, inspecting the optimization process, we find L-VSD acquires the geometries and textures for the 3D model more quickly than VSD, and the loss of the 3D model in L-VSD with $\\gamma = 1$ floats in a similar level to VSD with $\\gamma = 10$ . These results validate the necessity for the LoRA model to look ahead during the optimization of VSD. However, we also observe that the 3D model can easily suffer from being over-saturated as optimization continues, which means the 3D models can not converge with normal shapes and colors in L-VSD.", + "bbox": [ + 511, + 462, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Illustrative Example It seems weird and self-contradictory with only the unsatisfactory results above. Here we provide an illustrative 2D Gaussian example in Fig. 2, to show the existence of the mismatching problem, the effectiveness of lookahead, and point out the possible reason that harms the performance of L-VSD. In this experiment, we assume that $\\theta = x \\in \\mathbb{R}^2$ , and preset a Gaussian distribution as ground truth, which should be taken as the pretrained distribution. We randomly initialize the parameterized distribution $q(x)$ , and use another Gaussian distribution $r(x)$ to approximate $q$ , which can be interpreted as LoRA in VSD. In each iteration, we randomly sample $x_{\\text{sample}}$ from $q(x)$ , which is similar to differentiable rendering in VSD, and then use the mean and variance of a Gaussian distribution to calculate the optimization direction. The learning trajectories are illustrated in Fig. 2. The results confirm the mismatching", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19570", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f0c3328783ce29ca73eef3f66de0d6a5e3b4e362f5d509869f9d163a32f3a80a.jpg", + "image_caption": [ + "(a) Loss of the LoRA model during training given various $\\gamma$ . When increasing $\\gamma$ , the loss is relatively lower, while also showing periodic changes." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 356, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7c1d655b475f5f7715decd053e1f75b8007d361c517c069b15d7578b757a5418.jpg", + "image_caption": [ + "(b) The norm of the LoRA prediction $||\\epsilon_{\\phi}||$ varies w.r.t. training iteration at various $\\gamma$ in L-VSD." + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 629, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4597b38211e1ce0fc92b4221a81e57c1040859c65aa44005203dce8ef4bdd9e0.jpg", + "image_caption": [ + "(c) The norm of various scores during the training of L-VSD.", + "Figure 4. Curve Plots of Pivot Experimentsin Sec. 3." + ], + "image_footnote": [], + "bbox": [ + 640, + 88, + 893, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "problem of VSD, which hinders the distribution matching process. Correcting the optimization order can lead to better results. But, if we overfit the $r(x)$ on the samples $x_{\\text{sample}}$ , the results cannot converge normally towards the region of $p(x)$ when the timestep is small, which usually leads to color saturation in text-to-3D generation [11, 25]. This evidence supports our finding with the L-VSD above. More illustrative examples and complete runnable code can be found in Appendix A and F.", + "bbox": [ + 88, + 263, + 483, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Lookahead $\\times$ Convergence? Definitely No.", + "text_level": 1, + "bbox": [ + 89, + 409, + 454, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the above studies, we learn that fitting the LoRA model to the 3D model first is essential for VSD, while enhancing the convergence of the LoRA model is also beneficial to some degree. Here we conduct additional experiments to find out the consequences of combining these two factors.", + "bbox": [ + 89, + 431, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the setting of L-VSD in Sec. 3.2, we increase the LoRA training steps $\\gamma$ to 2 and 5. We also perform a study where the learning rate of LoRA is scaled to 1/10 of the original one to indicate under-convergence. Fig. 3b exhibits the results. As shown, increasing $\\gamma$ in L-VSD exacerbates the phenomenon of over-saturation and makes the geometry and texture easier to collapse, while decreasing the learning rate of the LoRA model somehow improves the quality of the generation.", + "bbox": [ + 89, + 522, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These deviate from our expectations. To chase a deeper understanding, we examine the output of the LoRA by inspecting its norm $||\\epsilon_{\\phi}||$ , and plot its variation during training in Fig. 4b. We observe that the norm rapidly drops to zero for L-VSD with $\\gamma = 5$ , which implies ill-posed convergence. The underlying reason is probably that we maintain only one single 3D particle training of LoRA for efficiency, thus the distribution to fit by the LoRA model is biased. The pathology is more obvious for L-VSD when optimizing LoRA more intensively. Besides, we note that the trick of reducing the learning rate, despite being effective in this case, is unstable when handling harder prompts. Please refer to Appendix C.1 for more failure cases.", + "bbox": [ + 89, + 659, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conclusion Although lookahead is important for generating 3D models with clear outlines and realistic texture, the risk is the possibility of the LoRA model over-fitting on the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "isolated particle, which sometimes exhibits flaws in geometry and texture after each optimization step. Given these, this work tries to figure out a way to interpolate between the original VSD with a relatively stable formation process and L-VSD with higher-quality outcomes.", + "bbox": [ + 511, + 263, + 905, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 345, + 650, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Sec 4.1, we conduct a rigorous comparison between VSD and L-VSD and perform thorough studies to gain an understanding of their difference. Please also refer to Appendix A for an illustrative overview of the training pipelines of VSD and L-VSD to understand their difference better. We then derive the novel Linearized Lookahead VSD ( $L^2$ -VSD) in Sec. 4.2.", + "bbox": [ + 511, + 372, + 906, + 476 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Compare VSD with L-VSD", + "text_level": 1, + "bbox": [ + 511, + 479, + 761, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first lay out the details of the update rules of L-VSD. Concretely, L-VSD first updates $\\phi_{i}$ with", + "bbox": [ + 511, + 502, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {i + 1} = \\phi_ {i} - 2 \\eta \\Delta_ {\\phi_ {i}}, \\Delta_ {\\phi_ {i}} := \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 541, + 919, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ denotes the learning rate of the LoRA model. Then, L-VSD updates $\\theta_{i}$ with the updated LoRA model $\\epsilon_{\\phi_{i + 1}}$ :", + "bbox": [ + 511, + 571, + 905, + 603 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {L - V S D} (\\theta_ {i}) =\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 612, + 640, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, \\epsilon , c} \\left[ \\right. w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\left. \\right| _ {\\theta = \\theta_ {i}} \\left. \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 630, + 924, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can decompose $\\epsilon_{\\phi_{i+1}}(x_t, t, c, y)$ by Taylor series to understand the gap between the update rule of VSD (Eq. (5)) and that of L-VSD (Eq. (8)) for $\\theta_i$ :", + "bbox": [ + 511, + 672, + 906, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right) \\\\ = \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + (\\phi_ {i + 1} - \\phi_ {i}) J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y) + \\dots \\\\ = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\underbrace {\\left(- 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} \\left(x _ {t} , t , c , y\\right)\\right)} _ {\\Delta \\epsilon_ {\\text {f i r s t}}} + \\underbrace {\\mathcal {O} \\left(\\Delta_ {\\phi_ {i}} ^ {2}\\right)} _ {\\Delta \\epsilon_ {\\text {h i g h}}} \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 729, + 903, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Namely, both the first-order term $\\Delta \\epsilon_{\\text{first}}$ and the high-order one $\\Delta \\epsilon_{\\text{high}}$ may contribute to the fast formation of geometry and texture in L-VSD.", + "bbox": [ + 511, + 825, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To disentangle their effects, we opt to plot how their norms vary during the training procedure. In particular,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19571", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cc272166c354be7731f981c11052e445d5740da343e949c55baf4af373548562.jpg", + "image_caption": [ + "Figure 5. Visualization of $\\Delta \\epsilon_{\\text{first}}$ and $\\Delta \\epsilon_{\\text{high}}$ . The left column in each group represents the decoded first-order term, while the right column represents the decoded high-order term. Prompt for each group: (left) \"a delicious hamburger\"; (mid) \"an astronaut riding a horse\";(right) \"an Iron man\"." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 217, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21ae0f28761a55beb72dc0cf348490a4d07f0a2091138e29946b2f2630936b08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 88, + 349, + 137 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33bb2d6b2689d6fc1060b8c53d187910d94a90da6541d49451cd046b121b3411.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 480, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we set $\\eta$ to 0.01. We leave how to estimate $\\Delta \\epsilon_{\\text{first}}$ to the next subsection and compute $\\Delta \\epsilon_{\\text{high}}$ by $\\epsilon_{\\phi_{i+1}}(x_t, t, c, y) - \\epsilon_{\\phi_i}(x_t, t, c, y) - \\Delta \\epsilon_{\\text{first}}$ . As illustrated in Fig. 4c, $||\\Delta \\epsilon_{\\text{high}}||$ is significantly larger than $||\\Delta \\epsilon_{\\text{first}}||$ . It is even larger than the norm of the entire LoRA model $||\\epsilon_{\\phi}||$ , which means that $\\epsilon_{\\phi_{i+1}}$ is probably dominated by $\\Delta \\epsilon_{\\text{high}}$ . Moreover, the norm of $\\Delta \\epsilon_{\\text{high}}$ varies in a considerable range, which indicates much greater randomness than the first-order term.", + "bbox": [ + 89, + 232, + 483, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain a more intuitive understanding of the two terms, we also pass them through the decoder of the pretrained DM [22] to obtain visualizations of them, presented in Fig. 5. As shown, the visualization of $\\Delta \\epsilon_{\\text{first}}$ indicates an object shape corresponding to the prompt, while $\\Delta \\epsilon_{\\text{high}}$ is more random, and we cannot witness any semantic information within it.", + "bbox": [ + 89, + 359, + 483, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on all the above observations and inferences, one question arises: whether $\\Delta \\epsilon_{\\text{first}}$ is the essential component for the appealing generations, and keeping it can improve score distillation further. We answer this in the next subsection.", + "bbox": [ + 89, + 470, + 483, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Linearized Lookahead VSD", + "text_level": 1, + "bbox": [ + 89, + 561, + 341, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)\\coloneqq \\epsilon_{\\phi_i}(x_t,t,c,y) + \\Delta \\epsilon_{first}$ . In fact, it corresponds to performing one-step SGD with $(x_{t'},t')$ under the denoising loss (Eq. (2)) to update the following linear noise-prediction model:", + "bbox": [ + 89, + 589, + 483, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {\\phi} ^ {\\operatorname {l i n}} \\left(x _ {t}, t, c, y\\right) = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\left(\\phi - \\phi_ {i}\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 674, + 483, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the low complexity of the linear model, the risk of overfitting to the current training data is low, which properly addresses the problem of the original L-VSD. With these understandings, our method boils down to using only the linearized lookahead correction $\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)$ for estimating the score function of noisy rendered images of $\\theta_{i}$ . i.e.,", + "bbox": [ + 89, + 712, + 482, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta_ {i}} \\mathcal {L} ^ {*} (\\theta_ {i}) = \\\\ \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right] \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 839, + 504, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 517, + 92, + 563, + 104 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) \\right. \\\\ \\left. + 2 \\eta \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) \\left[ J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right) \\right]\\right) \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 112, + 911, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Viewing $J_{\\phi_i}(x_{t'}, t', c, y) J_{\\phi_i}^T(x_t, t, c, y) \\in \\mathbb{R}^{d \\times d}$ as a preconditioning matrix, the above update rule involves two score contrast terms—one corresponds to the original VSD objective and the other accounts for a linearized lookahead correction for practical iterative optimization.", + "bbox": [ + 511, + 223, + 903, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we clarify the term $\\Delta \\epsilon_{\\text{first}} := -2\\eta \\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)$ is friendly to estimate. Concretely, $\\Delta_{\\phi_i}$ is exactly the gradient to update the LoRA model, which a backward pass can calculate. Then, the vector-Jacobian product $\\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)$ can be realized by a forward pass of the score model using forward-mode automatic differentiation, which is equipped in many deep learning frameworks [20]. As a result, $L^2$ -VSD only needs a single additional forward pass of the LoRA model per iteration compared to VSD, which is much more efficient than updating the LoRA multiple times. More importantly, we can even use only the entities associated with the last layer of the LoRA model for estimating the vector-Jacobian product to achieve a trade-off between quality and efficiency.", + "bbox": [ + 511, + 300, + 905, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 532, + 643, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 551, + 611, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the efficacy of our proposed Linearized Lookahead VSD method on text-guided 3D generation. Our baseline approaches include SDS, VSD, ESD (Entropy Score Distillation) [28], and VSD-based HiFA [37], which include representative state-of-the-art methods. For a fair comparison, all experiments are benchmarked under the open-source framework threestudio [5]. To clearly demonstrate the superior performance brought by our method, we compare both the results produced with high resolution and low resolution. Please refer to Appendix B for more implementation details.", + "bbox": [ + 511, + 573, + 905, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 511, + 748, + 735, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our generation results with SDS, VSD, ESD, L-VSD, and HiFA in high resolution of 256 and low resolution of 64. We present some representative results without bias in Fig. 1 and in Fig. 6, and we refer readers to Appendix C for results of L-VSD and more cases, as L-VSD tends to fail in generating complete objects. It's noteworthy that we consider our method as a baseline similar to SDS and VSD. ESD and HiFA are state-of-the-art based on VSD, which try to improve certain properties of generation", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19572", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55fd417d3d16ab3269df0a8398d30f316780898fec4361d8b6309ba0e0a54bee.jpg", + "image_caption": [ + "SDS" + ], + "image_footnote": [], + "bbox": [ + 94, + 107, + 245, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63d43fe71cd2459f50e540c7ffc2777f5a1e43b804d33375b83a5f9c6a52580c.jpg", + "image_caption": [ + "ESD" + ], + "image_footnote": [], + "bbox": [ + 259, + 107, + 410, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fa0961d93774f1a90ec730b828213f2960e6543d0438597b017a4636115c4dfb.jpg", + "image_caption": [ + "VSD" + ], + "image_footnote": [], + "bbox": [ + 426, + 107, + 578, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17c25c039b0e36b16c801eec5585c2696b0bc4e4d78a6c48880768dc9f9e5816.jpg", + "image_caption": [ + "VSD-HiFA" + ], + "image_footnote": [], + "bbox": [ + 594, + 107, + 733, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cc3c42068195db89cc9569473fae7ba60152aa2977443bfd948cbe415b0fa22.jpg", + "image_caption": [ + "$L^2\\mathrm{VSD}$" + ], + "image_footnote": [], + "bbox": [ + 750, + 107, + 890, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e2070fd0ac2cda253b6388b9a2f41ea14444c0b8a50047c3aff221888508dd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 195, + 245, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/235f7830a7bb6a1723f42245a22bddf4a9c71614147a4704774b61d6e9b6aaaf.jpg", + "image_caption": [ + "\"A red fire hydrant spraying water.\"" + ], + "image_footnote": [], + "bbox": [ + 259, + 195, + 410, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/918f334c19bce008fb44b64a19b5c6594e2fa8eb30cc4468dc6d36f28cb43a99.jpg", + "image_caption": [ + "\"A ficus planted in a pot.\"" + ], + "image_footnote": [], + "bbox": [ + 424, + 196, + 576, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/51e59be5137a136b4ece0082a022b5187cce15f638158940e8647a90622d5833.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 593, + 196, + 733, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04dfbcad533d34278c80e48aeac7eab9d3a0c88268e4a45298d3d9821a5169d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 196, + 900, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8a9038b5cdb62eb857914788e19dab1bb7c4823c2d5f3dd36cd8dac39dce143.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 285, + 243, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6eb2a64417515fc669eb9069e01d91b102593487cca9fc1a51acf28c51c013f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 285, + 408, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45f383171f9d9bdfabbe7ed427c19c84932e7cf8ecfe565d83b69bb7f54127cd.jpg", + "image_caption": [ + "\"A completely destroyed car\"" + ], + "image_footnote": [], + "bbox": [ + 426, + 285, + 576, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc33bed8073459d458ada39c0d5bad3362d5f44d7a562ff41181cf2363cfd404.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 285, + 733, + 356 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3ad41f68c382fbb6267758ce1051fe47730a41302cfe0796417446bc0bf6dc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 285, + 900, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1150da8a6c1dd7af86565e6e661902d96ab5a9b18ec2ad336ad020d54c159cad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 375, + 243, + 445 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b01c2c0c0c7a744dc5d8cb7e98c3576597e1d70118c135f2eb5650945ff15c1.jpg", + "image_caption": [ + "\"A DSLR photo of a cracked egg with the yolk spilling out on a wooden table\"", + "Figure 6. More Qualitative comparison with high resolution." + ], + "image_footnote": [], + "bbox": [ + 261, + 375, + 408, + 444 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/831a340c3076052af2f04274cd77382d5257c69e55be25781439d93b1321606d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 375, + 576, + 444 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/208b8c978a65a339ec1fbef20e14939d6e43e66b10985e79293184d320ec16bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 375, + 733, + 444 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c514e95edba7d1ceb1fed7edbc93f28687c335dca06ee38a8c2b4c20bf6f6ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 373, + 900, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "quality by introducing novel techniques. We compare with them in order to position our performance more rigorously. We can find that our method can generate high-quality 3D assets with a more realistic appearance than baselines. As for VSD-HiFA, there exist more floating around the objects while ours do not. It can generate better textures, which is due to their explicit loss design on the image space. We demonstrate in Sec. 5.5 that our methods can be seamlessly combined with other techniques like ESD and HiFA. Compared with VSD and its variant ESD, our results have significantly more realistic appearances and reasonable geometry. Moreover, our method has better convergence, which means our results are more robust to optimization and do not suffer from generating floaters in space. It's clearly shown that with our linearized lookahead correction term, even only trained with low-resolution rendering settings, we can generate 3D scenes with realistic and detailed appearances. In conclusion, our method surpasses the other baseline methods significantly with linearized lookahead and achieves comparable results to state-of-the-art.", + "bbox": [ + 89, + 482, + 482, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.Quantitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 787, + 321, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow previous work [21, 34] to quantitatively evaluate the generation quality using the angle of CLIP similarity [6] and Frechet Inception Distance [7] and list the results in Table 1. Specifically, the CLIP similarity measures the cosine similarity between the rendered image embeddings of the generated 3D object and the text embeddings,", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and we calculate the angle. The FID measures the distance between the image distribution by randomly rendering 3D representation and the text-conditioned image distribution from the pretrained diffusion model. We use DreamFusion gallery prompts to ensure that there is no man-made bias in the evaluation. Please refer to Appendix B for more metric computation details.", + "bbox": [ + 511, + 483, + 906, + 587 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8fbcdfb889793e1812fd20ad6b80c4321438b25ee237572755cdf651277aa8e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SDSESDVSDL-VSDHiFAL2-VSD
Averaged CLIP sim (↓)0.3050.3160.3240.3370.3130.285
Averaged FID (↓)372.35315.15301.54496292.88284.06
", + "bbox": [ + 526, + 589, + 893, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparison. $(\\downarrow)$ means the lower the better. We measure the scenes across 20 prompts, which are randomly sampled from DreamFusion's gallery and include scenes shown in Fig. 1, Fig. 6, and figures in appendix. The quantitative results align with the qualitative results.", + "bbox": [ + 511, + 643, + 906, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 723, + 666, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4.1. Ablation on $\\eta$", + "text_level": 1, + "bbox": [ + 511, + 744, + 653, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study on the scale of $\\eta$ , as shown in Fig. 7a, where only $\\eta$ changes while other parameters are unchanged. We still use the simple prompt \"a delicious hamburger\". It's demonstrated that our method is robust to the scale of $\\eta$ . It's worth noting that even if the norm of the first-order term $\\Delta \\epsilon_{\\text{first}}$ is only around 1e-2 when we set $\\eta$ to 1e-3, the results still improve a lot. So, even a minor correction for each iteration can lead to incredible improvement in such a long-term optimization. Ablation on", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c9371a5608726abdfe3a7a03be835b82a300107748d8bc71b2f014d219fc8ea7.jpg", + "image_caption": [ + "(a) Ablation on the $\\eta$ in $\\Delta \\epsilon_{\\text{first}}$ . From left to right, the results correspond to setting $\\eta$ to 1e-3, 1e-2, 0.1, and 1 respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 486, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f1cd55eeda82783fcb47691870c8cf0fa4e81e7a2a4df58c087dfbbe83564ee.jpg", + "image_caption": [ + "(b) Ablation on last-layer approximation. From left to right, the results correspond to settings in (a), but with last-layer approximation." + ], + "image_footnote": [], + "bbox": [ + 508, + 88, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71d2142c30b18b0bc42dc8c6327b338f6a5ab92f7d498369555ed7bb3241b152.jpg", + "image_caption": [ + "Figure 7. Qualitative results of ablation studies.", + "Figure 8. Examples of combining ESD with $L^2$ -VSD. We can observe that in the case of \"sunflower\", by combining our method, we obtain a reasonable sunflower rather than a \"sunball\". In the case of \"bulldozer\", with a linearized lookahead, ESD can generate additional elements like bricks." + ], + "image_footnote": [], + "bbox": [ + 91, + 224, + 485, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "high-order term correction can be found in Appendix C.4.", + "bbox": [ + 89, + 377, + 472, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2. Ablation on Last-layer Approximation", + "text_level": 1, + "bbox": [ + 89, + 396, + 405, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As stated above, we can use only the entities associated with the last layer of the LoRA model to further reduce additional time costs. Admittedly, this approach may result in performance loss. As the correction term is calculated with the Jacobian matrices of the LoRA model, if we only use the last-layer gradients to approximate, we actually ignore the variable updates within other layers, thus losing accuracy. We present the corresponding results in Fig. 7b. Although the outcomes tend to generate floaters, the realism of results still gets largely improved. Also, we believe the floaters can be eliminated in the following geometry refinement stage. We compare the computation efficiency with the baseline VSD in Appendix B.3. We can save a significant computation time cost by only using this approximation.", + "bbox": [ + 88, + 415, + 483, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Connection with Other Methods", + "text_level": 1, + "bbox": [ + 89, + 635, + 375, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ESD [28] is dedicated to solving the Janus problem based on VSD. It maximizes the entropy of different views of the generated results, encouraging diversity across views. The gradient of ESD is theoretically equivalent to adopting the CFG [8] trick upon VSD. Our method can also be incorporated into ESD. We name this as $L^2$ -ESD. We present the corresponding results in Fig. 8. The prompts we use are \"a sunflower on a flowerpot\" and \"a bulldozer made out of toy bricks\" respectively. To avoid unnecessary trials, we simply set $\\lambda$ to 0.5, which is recommended in ESD. As we can observe, compared with ESD, the 3D results are more photo-realistic with distinct shapes.", + "bbox": [ + 89, + 657, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HiFA [37] is a state-of-the-art approach for high-quality 3D generation in a single-stage training based on basic methods. It distills denoising scores from pretrained models in both the image and latent spaces and proposes several", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09602a92e0edf871dbea8998e7fb656540e17549e30bb3e352c7bb68d3b9f3e4.jpg", + "image_caption": [ + "Figure 9. Examples of combining HiFA with $L^2$ -VSD. The colors are more realistic and the appearance of our method's results aligns better with the prompts." + ], + "image_footnote": [], + "bbox": [ + 514, + 224, + 903, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "techniques to improve NeRF generation, which are orthogonal to our method. We demonstrate our method's compatibility by combining it with HiFA, named as $L^2$ -HiFA. We present the results in Fig. 9. The prompts we use are \"an elephant skull\" and \"Pumpkin head zombie, skinny, highly detailed, photorealistic\" respectively.", + "bbox": [ + 511, + 351, + 906, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, it's believed that $L^2$ -VSD can be combined with other parallel techniques in the future.", + "bbox": [ + 511, + 443, + 905, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion, Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 479, + 861, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we dive deep into the theory of VSD, having a comprehensive understanding of the inner process, and identify two potential directions for improvement. In terms of the algorithm formulation, we propose Linearized Lookahead Variational Score Distillation( $L^2$ -VSD), a novel framework based on VSD that achieves state-of-the-art results on text-to-3D generation. The linearized lookahead term enables us to benefit from both better convergence and lookahead for the next iteration. More importantly, our method can be incorporated into any VSD-based framework in the future.", + "bbox": [ + 511, + 506, + 906, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and broader impact. Firstly, although $L^2$ -VSD achieves remarkable improvement on text-to-3D results, the generation process still takes hours of time, which is a common issue for score distillation-based methods. We believe orthogonal improvements on distillation accelerating [35] can mitigate this problem. Secondly, while we find the first-order term shows a regular pattern and some similarity may exist between our method and SiD [36], which is briefly discussed in appendix section E, we have not yet been able to formulate a distribution-based objective that guides this optimization. Investigating the underlying reasons is of significant interest in the future. Lastly, like other generative models, our method may be utilized to generate fake and malicious content, which needs more attention and caution.", + "bbox": [ + 511, + 674, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19574", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by NSF of China (Nos. 92470118, 62306176), Natural Science Foundation of Shanghai (No. 23ZR1428700), CCF-ALIMAMA TECH Kangaroo Fund (NO. CCF-ALIMAMA OF 2025010), and CCF-Zhipu Large Model Innovation Fund (No. CCF-Zhipu202412).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Singh, and Tim Rocktäschel. Genie: Generative interactive environments, 2024. 1", + "[2] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 1", + "[3] Matthew Conlen, Jeffrey Heer, Hillary Mushkin, and Scott Davidoff. Cinematic techniques in narrative visualization, 2023. 1", + "[4] Chris Creed, Maadh Al-Kalbani, Arthur Theil, Sayan Sarcar, and Ian Williams. Inclusive ar/vr: accessibility barriers for immersive technologies. Universal Access in the Information Society, 23(1):59-73, 2023. 1", + "[5] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.4.6", + "[6] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 7", + "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 8", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 3", + "[10] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 3", + "[11] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023.5", + "[12] Chenhan Jiang. A survey on text-to-3d contents generation in the wild, 2024. 1" + ], + "bbox": [ + 93, + 227, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation, 2023. 1, 3", + "[14] Nikhil Ketkar, Jojo Moolayil, Nikhil Ketkar, and Jojo Moolayil. Automatic differentiation in deep learning. Deep Learning with Python: Learn Best Practices of Deep Learning Models with PyTorch, pages 133-145, 2021. 2", + "[15] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. 3", + "[16] Sixu Li, Chaojian Li, Wenbo Zhu, Boyang, Yu, Yang, Zhao, Cheng Wan, Haoran You, Huihong Shi, Yingyan, and Lin. Instant-3d: Instant neural radiance field training towards ondevice ar/vr 3d reconstruction, 2024. 1", + "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 1", + "[18] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1–19. Springer, 2025. 1", + "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3", + "[20] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 2, 6", + "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 7", + "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6", + "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3", + "[24] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 3", + "[25] Boshi Tang, Jianan Wang, Zhiyong Wu, and Lei Zhang. Stable score distillation for high-quality 3d generation, 2024. 3, 5", + "[26] SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy," + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Martin Engelcke, Ryan Faulkner, Frankie Garcia, Charles Gbadamosi, Zhitao Gong, Lucy Gonzales, Kshitij Gupta, Karol Gregor, Arne Olav Hallingstad, Tim Harley, Sam Haves, Felix Hill, Ed Hirst, Drew A. Hudson, Jony Hudson, Steph Hughes-Fitt, Danilo J. Rezende, Mimi Jasarevic, Laura Kampis, Rosemary Ke, Thomas Keck, Junkyung Kim, Oscar Knagg, Kavya Kopparapu, Andrew Lampinen, Shane Legg, Alexander Lerchner, Marjorie Limont, Yulan Liu, Maria Loks-Thompson, Joseph Marino, Kathryn Martin Cussons, Loic Matthey, Siobhan Mcloughlin, Piermaria Mendolicchio, Hamza Merzic, Anna Mitenkova, Alexandre Moufarek, Valeria Oliveira, Yanko Oliveira, Hannah Openshaw, Renke Pan, Aneesh Pappu, Alex Platonov, Ollie Purkiss, David Reichert, John Reid, Pierre Harvey Richemond, Tyson Roberts, Giles Ruscoe, Jaume Sanchez Elias, Tasha Sandars, Daniel P. Sawyer, Tim Scholtes, Guy Simmons, Daniel Slater, Hubert Soyer, Heiko Strathmann, Peter Stys, Allison C. Tam, Denis Teptyashin, Tayfun Terzi, Davide Vercelli, Bojan Vujatovic, Marcus Wainwright, Jane X. Wang, Zhengdong Wang, Daan Wierstra, Duncan Williams, Nathaniel Wong, Sarah York, and Nick Young. Scaling instructable agents across many simulated worlds, 2024. 1", + "[27] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3", + "[28] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming mode collapse in score distillation for text-to-3d generation, 2024. 1, 2, 3, 6, 8", + "[29] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 1, 3, 4", + "[30] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan, 2023. 1", + "[31] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8131-8141, 2024. 1", + "[32] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video, 2024. 1", + "[33] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting, 2023. 3", + "[34] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation, 2023. 1, 3, 7", + "[35] Linqi Zhou, Andy Shih, Chenlin Meng, and Stefano Ermon. Dreampropeller: Supercharge text-to-3d generation" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with parallel sampling. arXiv preprint arXiv:2311.17082, 2023.8", + "[36] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. arXiv preprint arXiv:2404.04057, 2024.8", + "[37] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 6, 8" + ], + "bbox": [ + 516, + 92, + 903, + 233 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19576", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_model.json b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4c56781ac514287c2f9fd07b11b90de4e873923 --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_model.json @@ -0,0 +1,2860 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.764, + 0.175 + ], + "angle": 0, + "content": "Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.836, + 0.258 + ], + "angle": 0, + "content": "Yu Lei\\(^{1}\\) Bingde Liu\\(^{1}\\) Qingsong Xie\\(^{2}\\) Haonan Lu\\(^{2}\\) Zhijie Deng\\(^{1*}\\) \n\\(^{1}\\) Shanghai Jiao Tong University OPPO AI Center \n{tony-lei, bingde, zhijied}@sjtu.edu.cn {xieqingsong1, luhaonan}@oppo.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.485, + 0.747 + ], + "angle": 0, + "content": "Text-to-3D generation based on score distillation of pretrained 2D diffusion models has gained increasing interest, with variational score distillation (VSD) as a remarkable example. VSD proves that vanilla score distillation can be improved by introducing an extra score-based model, which characterizes the distribution of images rendered from 3D models, to correct the distillation gradient. Despite the theoretical foundations, VSD, in practice, is likely to suffer from slow and sometimes ill-posed convergence. In this paper, we perform an in-depth investigation of the interplay between the introduced score model and the 3D model, and find that there exists a mismatching problem between LoRA and 3D distributions in practical implementation. We can simply adjust their optimization order to improve the generation quality. By doing so, the score model looks ahead to the current 3D state and hence yields more reasonable corrections. Nevertheless, naive lookahead VSD may suffer from unstable training in practice due to the potential over-fitting. To address this, we propose to use a linearized variant of the model for score distillation, giving rise to the Linearized Lookahead Variational Score Distillation (\\( L^2 \\)-VSD). \\( L^2 \\)-VSD can be realized efficiently with forward-mode autodiff functionalities of existing deep learning libraries. Extensive experiments validate the efficacy of \\( L^2 \\)-VSD, revealing its clear superiority over prior score distillation-based methods. We also show that our method can be seamlessly incorporated into any other VSD-based text-to-3D framework." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.776, + 0.222, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.483, + 0.879 + ], + "angle": 0, + "content": "3D content creation is important for a variety of applications, such as interactive gaming [1, 32], cinematic arts [3], AR/VR [4, 16], and building simulated environments for training agents in robotics [26]. However, it is still challenging and expensive to create high-quality 3D assets as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.354 + ], + "angle": 0, + "content": "it requires a high level of expertise. Therefore, automating this process with generative models has become an important problem [12], while remaining non-trivial due to the scarcity of data and complexity of 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.909, + 0.627 + ], + "angle": 0, + "content": "Score distillation has emerged as an attractive way for 3D generation given textual condition [2, 17, 21, 28, 29]. It leverages pretrained 2D diffusion models [9, 22] to define priors to guide the evolution of 3D content without reliance on annotations. Score Distillation Sampling (SDS) [21] is a seminal work in this line, but it is widely criticized that its generations suffer from the over-smoothing issue. Variational Score Distillation (VSD) [29] remediates this by introducing an extra model that captures the score of the images rendered from the 3D model to correct the distillation gradient. However, VSD often requires a lengthy optimization through 3 stages: NeRF generation, geometry refinement, and texture refinement. The outcomes obtained in the initial stage are often blurry, prone to collapsing, and not directly applicable [30]. Though existing works begin to understand and improve SDS [13, 28, 34], there are great but fewer efforts dedicated to improving the more promising VSD [18, 31]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.909, + 0.78 + ], + "angle": 0, + "content": "To identify the root of VSD's drawback, we conduct a comprehensive analysis of the interaction between the introduced score model and the 3D model, revealing that adjusting their optimization order can sometimes lead to a considerable enhancement in generation quality. This adjustment allows the score model to look ahead to the current 3D state, resulting in more accurate and sensible corrections for the distillation gradient. Yet, naive lookahead VSD can encounter unstable training due to the risk of the score model overfitting the single 3D particle and sampled camera view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this issue, we formally compare the correction gradients before and after looking ahead and identify two major differences—a linear first-order term and a high-order one. Upon closer examination, we observe that the former accommodates subtle semantic information, whereas the latter contains non-trivial high-frequency noises. Given these findings, we propose to use only the first-order term for correction, yielding Linearized Looka" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19567" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.093, + 0.19, + 0.104 + ], + "angle": 0, + "content": "SDS" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.105, + 0.246, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.093, + 0.35, + 0.103 + ], + "angle": 0, + "content": "ESD" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.106, + 0.409, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.093, + 0.514, + 0.104 + ], + "angle": 0, + "content": "VSD" + }, + { + "type": "image", + "bbox": [ + 0.45, + 0.118, + 0.56, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.093, + 0.694, + 0.104 + ], + "angle": 0, + "content": "VSD-HiFA" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.105, + 0.733, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.093, + 0.843, + 0.104 + ], + "angle": 0, + "content": "\\(L^{2}\\mathrm{VSD}\\)" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.106, + 0.895, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.186, + 0.545, + 0.199 + ], + "angle": 0, + "content": "\"A hamburger\"" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.204, + 0.245, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.204, + 0.41, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.204, + 0.572, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.203, + 0.734, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.203, + 0.895, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.284, + 0.66, + 0.298 + ], + "angle": 0, + "content": "\"A DSLR photo of a fox holding a videogame controller\"" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.303, + 0.242, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.303, + 0.407, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.303, + 0.571, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.304, + 0.732, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.304, + 0.895, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.382, + 0.657, + 0.395 + ], + "angle": 0, + "content": "\"A humanoid robot lying on the couch while on a laptop\"" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.4, + 0.245, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.265, + 0.4, + 0.411, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.4, + 0.574, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.4, + 0.734, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.401, + 0.898, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.48, + 0.682, + 0.493 + ], + "angle": 0, + "content": "\"A zoomed out DSLR photo of Sydney opera house, aerial view\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.906, + 0.523 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparison with high resolution of 256. \\(L^2\\)-VSD can generate meticulously detailed and realistic 3D assets from easy to very complex prompts and scenes. \\(L^2\\)-VSD tends to yield more complete structures than HiFA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.643 + ], + "angle": 0, + "content": "head Variational Score Distillation (\\(L^2\\)-VSD), to reliably and consistently boost the generation quality of VSD. \\(L^2\\)-VSD is both easy to implement and computationally efficient—the added linear term can be computed by only one additional forward process of the score model under the scope of forward-mode automatic differentiation, which is supported in many deep learning libraries [14, 20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Through extensive experiments, we demonstrate the significant superiority of the proposed \\( L^2 \\)-VSD in improving 3D generation quality compared to competing baseline methods, as shown in Fig. 1. \\( L^2 \\)-VSD can even produce realistic results with low resolution directly in the first stage. Moreover, we empirically show that \\( L^2 \\)-VSD can be seamlessly integrated into other VSD-based 3D generation pipelines and combined with other parallel techniques for VSD, e.g., Entropy Score Distillation(ESD) [28], which mitigates the Janus problem, for further improvement. We summarize our technical contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- For VSD, a fundamental method in text-to-3D generation, we carefully identify the gaps between its theory and implementation and analyze the potential impact the gaps may bring us, providing a direction for possible improvements." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose \\(L^2\\)-VSD, an easy-to-implement and compu" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.537, + 0.706, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.537, + 0.902, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Figure 2. Comparison of convergence between VSD and L-VSD with an illustrative 2D Gaussian example. In this toy example we consider \\( x \\in \\mathbb{R}^2 \\) from a single Gaussian distribution. We optimize \\( q(x) \\) towards ground truth distribution \\( p(x) \\) and use \\( r(x) \\) to function as LoRA, which is used to estimate \\( q(x) \\). This example validates the existence of a mismatching issue in VSD, and we leave the details in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.801, + 0.905, + 0.861 + ], + "angle": 0, + "content": "tationally efficient variant of VSD, which mitigates the mismatching problem to some extent. We demonstrate its significant improvement over baselines and are comparable to other state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.862, + 0.905, + 0.906 + ], + "angle": 0, + "content": "- We demonstrate that our method can be seamlessly combined with other VSD-based improving techniques without much effort." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.801, + 0.905, + 0.906 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19568" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.221, + 0.108 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.114, + 0.26, + 0.13 + ], + "angle": 0, + "content": "2.1. Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.136, + 0.485, + 0.199 + ], + "angle": 0, + "content": "Diffusion models (DMs) [9, 23, 24] are defined with a forward diffusion process on data \\( x \\in \\mathbb{R}^d \\) with a Gaussian transition kernel. The conditional distribution at some timestep \\( t \\in [0, T] \\) usually satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.208, + 0.483, + 0.226 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.484, + 0.388 + ], + "angle": 0, + "content": "where \\(x_0 \\coloneqq x\\), and \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are pre-defined noise schedules. DMs learn a reverse diffusion process, specified by a parameterized distribution \\(p(x_{t - 1}|x_t)\\coloneqq \\mathcal{N}(x_{t - 1};\\mu_\\psi (x_t,t),\\sigma_t^2\\mathbf{I})\\) with \\(\\mu_{\\psi}\\) as a neural network (NN), to enable the sampling of generations from Gaussian noise. The training objective of \\(\\mu_{\\psi}\\) is the variational lower bound of the log data likelihood. In practice, \\(\\mu_{\\psi}\\) is reparameterized as a denoising network \\(\\epsilon_{\\psi}\\) and the training loss can be further simplified as a Mean Squared Error (MSE) form [9, 15]:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.397, + 0.483, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D i f f} := \\mathbb {E} _ {x, t, \\epsilon} [ \\omega (t) | | \\epsilon_ {\\psi} (\\alpha_ {t} x + \\sigma_ {t} \\epsilon , t) - \\epsilon | | _ {2} ^ {2} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.485 + ], + "angle": 0, + "content": "where \\( x \\) follows the data distribution, \\( t \\) is uniformly drawn from \\( [0,T] \\), \\( \\epsilon \\) is a standard Gaussian noise, and \\( \\omega (t) \\) is a time-dependent coefficient. \\( \\epsilon_{\\psi} \\) also inherently connects to score matching [24, 27]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Classifier-free guidance (CFG) [8]. We can augment the model \\(\\epsilon_{\\psi}\\) with an extra input, the condition \\(y\\), to characterize the corresponding conditional distribution, leaving \\(\\epsilon_{\\psi}(x_t,t,\\emptyset)\\) account for the original unconditional one. Then, we can resort to CFG to further boost the quality of conditional generation. Typically, CFG leverages the following term in the sampling process:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.602, + 0.483, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\hat {\\epsilon} _ {\\psi} \\left(x _ {t}, t, y\\right) := (1 + s) \\epsilon_ {\\psi} \\left(x _ {t}, t, y\\right) - s \\epsilon_ {\\psi} \\left(x _ {t}, t, \\emptyset\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.629, + 0.351, + 0.644 + ], + "angle": 0, + "content": "where \\( s > 0 \\) refers to a guidance scale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.476, + 0.669 + ], + "angle": 0, + "content": "2.2. Text-to-3D Generation with Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Text-to-3D generation aims to identify the parameters \\(\\theta \\in \\mathbb{R}^N\\) of a 3D model given a text condition \\(y\\). Neural radiance field (NeRF) [19] is a typical 3D representation based on neural networks. In particular, NeRF renders a new view of the scene with the input of a sequence of images as known views. Additionally, textured mesh can be applied to represent the geometry of a 3D object with triangle meshes and textures with color on the mesh surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Denote \\( g(\\theta, c) \\) as the differential rendering function projecting the 3D scene to a 2D image given a camera angle \\( c \\). Score distillation approaches for text-to-3D generations demand the image sample \\( g(\\theta, c) \\) to respect the prior specified by a text-to-2D diffusion model \\( \\epsilon_{\\text{pretrain}}(\\cdot, \\cdot, y) \\) pretrained on vast real text-image pairs, based on which the optimization goal is constructed [13, 21, 25, 28, 29, 33, 34]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "Score Distillation Sampling (SDS) [21] updates the 3D model using view-dependent prompt \\( y^{c} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.132, + 0.915, + 0.174 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\theta) := \\mathbb {E} _ {t, \\epsilon , c} [ \\omega (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y ^ {c}\\right) - \\epsilon\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.175, + 0.907, + 0.309 + ], + "angle": 0, + "content": "where \\( c \\) is a randomly sampled camera angle and \\( x_{t} \\coloneqq \\alpha_{t} g(\\theta, c) + \\sigma_{t} \\epsilon \\). The gradient is a simplification of that of the denoising objective w.r.t. \\( \\theta \\) [21]. Intuitively, it encourages the rendered images to move toward the high-probability regions of the pretrained model, thus a good 3D model emerges. SDS is a seminal work in the line of text-to-3D generation, but it is sensitive to the CFG scale \\( s \\) [8]. A small \\( s \\) often results in over-smooth outcomes, whereas a large \\( s \\) leads to over-saturation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.906, + 0.476 + ], + "angle": 0, + "content": "Variational Score Distillation (VSD) [29] addresses the issue of SDS with a thorough theoretical analysis and proposes a new algorithm. In particular, apart from the pretrained model \\(\\epsilon_{\\text{pretrain}}(x_t, t, y)\\) for capturing the data distribution, VSD introduces a tunable model \\(\\epsilon_{\\phi}(x_t, t, c, y)\\), often instantiated as a LoRA [10] adaptation of \\(\\epsilon_{\\text{pretrain}}\\), to account for the distribution of images rendered from all possible 3D models given condition \\(y\\) and camera angle \\(c\\). We refer the introduced model as the LoRA model hereinafter. VSD proves the LoRA model can reliably correct the original distillation gradient." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.552 + ], + "angle": 0, + "content": "VSD, in practice, usually considers only a single 3D particle \\(\\theta\\) and performs an iterative optimization of \\(\\theta\\) and \\(\\phi\\) until convergence. More specifically, let \\(\\theta_{i}\\) and \\(\\phi_{i}\\) denote the parameters at \\(i\\)-th training iteration. VSD updates \\(\\theta_{i}\\) with the following gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.563, + 0.676, + 0.579 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {V S D} (\\theta_ {i}) := \\mathbb {E} _ {t, \\epsilon , c}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.582, + 0.906, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\left[ w (t) \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.704 + ], + "angle": 0, + "content": "where \\( x_{t} = \\alpha_{t}g(\\theta_{i},c) + \\sigma_{t}\\epsilon \\). To avoid more than once differentiable rendering of the 3D model for efficiency, VSD updates \\( \\phi_{i} \\) still with \\( g(\\theta_i,c) \\) while using a different noisy state \\( x_{t^{\\prime}} = \\alpha_{t^{\\prime}}g(\\theta_{i},c) + \\sigma_{t^{\\prime}}\\epsilon^{\\prime} \\). The learning objective is the denoising loss defined in Eq. (2), whose gradient is:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.714, + 0.906, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\phi_ {i}} \\mathcal {L} _ {V S D} (\\phi_ {i}) := \\tag {6} \\\\ \\mathbb {E} _ {t ^ {\\prime}, \\epsilon^ {\\prime}, c} \\left[ \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.906, + 0.796 + ], + "angle": 0, + "content": "where \\(J_{\\phi_i}(x_{t'}, t', c, y) \\coloneqq \\frac{\\partial \\epsilon_\\phi(x_{t'}, t', c, y)}{\\partial \\phi}|_{\\phi = \\phi_i}\\) and we omit some time-dependent scaling factor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Apart from the compromise to maintain one 3D particle for efficiency, the practical algorithm of VSD exhibits several significant gaps from the theory: 1) one-step update cannot guarantee \\(\\phi\\) to converge in each iteration, and 2) the updates to \\(\\theta_{i}\\) are computed given the LoRA model \\(\\phi_{i}\\), which characterizes the distribution associated with the previous 3D state \\(\\theta_{i - 1}\\) instead of \\(\\theta_{i}\\). We hypothesize these" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.077, + 0.495, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.157, + 0.493, + 0.194 + ], + "angle": 0, + "content": "(a) VSD: From left to right, the results correspond to setting \\(\\gamma\\) to 1, 2, 5, and 10. We can observe that the quality of the model is not fundamentally improved." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.076, + 0.903, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.157, + 0.902, + 0.194 + ], + "angle": 0, + "content": "(b) L-VSD: From left to right, the first three results correspond to setting \\(\\gamma\\) to 1, 2, and 5; the last corresponds to scaling the learning rate by 0.1 with \\(\\gamma = 1\\)." + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.199, + 0.88, + 0.214 + ], + "angle": 0, + "content": "Figure 3. Qualitative Examples of training LoRA model for multiple steps per optimization iteration for VSD and L-VSD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.486, + 0.268 + ], + "angle": 0, + "content": "are probably the root of the unstable performance and sometimes corrupted outcomes of VSD, and perform an in-depth investigation regarding them below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.345, + 0.3 + ], + "angle": 0, + "content": "3. Diagnose the Issues of VSD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.46 + ], + "angle": 0, + "content": "To assess whether the identified gaps contribute to the issues of VSD, we conduct two sets of pivot experiments in this section. As introduced in Sec 2.2, we analyze the impact of convergence of LoRA in Sec 3.1; then, we correct the optimization order to see the consequences in Sec 3.2; lastly, we combine these two factors to check if they are interfered in Sec 3.3. We provide an illustrative 2D Gaussian example for better understanding and to evaluate the effectiveness of lookahead in Sec 3.2. We base the implementation on the open-source framework threestudio [5]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.379, + 0.485 + ], + "angle": 0, + "content": "3.1. Better Convergence? Maybe No." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Assumption 1 The one-step update in vanilla VSD cannot guarantee the LoRA model to converge well, thus possibly harming the effectiveness of score distillation. We assume updating the LoRA model for \\(\\gamma (\\gamma >1)\\) steps in each iteration could alleviate the pathology." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Experiments Setting We evaluate the effects of various \\(\\gamma\\), including 1, 2, 5, and 10. In each step, we optimize the LoRA model with different noisy images under different views. Unless otherwise specified, we take NeRF as the default 3D representation and use a simple prompt, \"a delicious hamburger\", in the study. We keep all other hyperparameters the same as the vanilla VSD. Usually, the whole VSD process contains multiple stages, where in the first stage a NeRF is constructed and then the geometry and texture are refined, respectively. We directly report the first-stage learning outcomes because they establish the foundation for the following parts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Findings 1 We present the training loss of the LoRA model in Fig. 4a to indicate the convergence and the final generations in Fig. 3a to reflect if the issues still exist. As shown, although the loss curves for various \\(\\gamma\\) share a periodic rise and fall, the loss of a larger \\(\\gamma\\) (e.g., 5 or 10) is floating in a relatively smaller range, and \\(\\gamma = 5\\) roughly makes LoRA model converge. However, as shown in Fig. 3a, it's hard to tell if the shape becomes relatively more reasonable, and the overall quality of the 3D model does not witness a continual improvement as \\(\\gamma\\) rises. Thereby, we conclude" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.905, + 0.253 + ], + "angle": 0, + "content": "that improving the convergence of the LoRA model on the original VSD is not sufficient, thus not being the key factor." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.26, + 0.742, + 0.275 + ], + "angle": 0, + "content": "3.2. Lookahead? Maybe Yes!" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.905, + 0.387 + ], + "angle": 0, + "content": "Assumption 2 As shown in Eq. (5), the updates to \\(\\theta_{i}\\) are computed given \\(\\phi_{i}\\), which characterizes the distribution associated with \\(\\theta_{i-1}\\) instead of \\(\\theta_{i}\\). This is inconsistent with Theorem 2 of [29], where the LoRA model should first adapt to the current 3D model (i.e., \\(\\theta_{i}\\)) to serve as a reliable score estimator for the corresponding distribution. Fixing such a mismatch may address the issues of VSD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.463 + ], + "angle": 0, + "content": "Experiments Setting We update \\(\\phi_{i}\\) first to obtain \\(\\phi_{i + 1}\\), based on which \\(\\theta_{i}\\) is updated. We name such a modification Lookahead-VSD (L-VSD) because it makes the LoRA model look ahead for one step compared to the original VSD. All other parameters are kept unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Findings 2 We show the result in Fig. 3b. Intuitively, the rendered image has clearer edges compared to those in Fig. 3a. Besides, inspecting the optimization process, we find L-VSD acquires the geometries and textures for the 3D model more quickly than VSD, and the loss of the 3D model in L-VSD with \\(\\gamma = 1\\) floats in a similar level to VSD with \\(\\gamma = 10\\). These results validate the necessity for the LoRA model to look ahead during the optimization of VSD. However, we also observe that the 3D model can easily suffer from being over-saturated as optimization continues, which means the 3D models can not converge with normal shapes and colors in L-VSD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Illustrative Example It seems weird and self-contradictory with only the unsatisfactory results above. Here we provide an illustrative 2D Gaussian example in Fig. 2, to show the existence of the mismatching problem, the effectiveness of lookahead, and point out the possible reason that harms the performance of L-VSD. In this experiment, we assume that \\(\\theta = x \\in \\mathbb{R}^2\\), and preset a Gaussian distribution as ground truth, which should be taken as the pretrained distribution. We randomly initialize the parameterized distribution \\(q(x)\\), and use another Gaussian distribution \\(r(x)\\) to approximate \\(q\\), which can be interpreted as LoRA in VSD. In each iteration, we randomly sample \\(x_{\\text{sample}}\\) from \\(q(x)\\), which is similar to differentiable rendering in VSD, and then use the mean and variance of a Gaussian distribution to calculate the optimization direction. The learning trajectories are illustrated in Fig. 2. The results confirm the mismatching" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19570" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.357, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.201, + 0.357, + 0.249 + ], + "angle": 0, + "content": "(a) Loss of the LoRA model during training given various \\(\\gamma\\). When increasing \\(\\gamma\\), the loss is relatively lower, while also showing periodic changes." + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.63, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.2, + 0.63, + 0.237 + ], + "angle": 0, + "content": "(b) The norm of the LoRA prediction \\(||\\epsilon_{\\phi}||\\) varies w.r.t. training iteration at various \\(\\gamma\\) in L-VSD." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.089, + 0.895, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.2, + 0.902, + 0.226 + ], + "angle": 0, + "content": "(c) The norm of various scores during the training of L-VSD." + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.248, + 0.663, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Curve Plots of Pivot Experimentsin Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.265, + 0.485, + 0.401 + ], + "angle": 0, + "content": "problem of VSD, which hinders the distribution matching process. Correcting the optimization order can lead to better results. But, if we overfit the \\( r(x) \\) on the samples \\( x_{\\text{sample}} \\), the results cannot converge normally towards the region of \\( p(x) \\) when the timestep is small, which usually leads to color saturation in text-to-3D generation [11, 25]. This evidence supports our finding with the L-VSD above. More illustrative examples and complete runnable code can be found in Appendix A and F." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.41, + 0.455, + 0.427 + ], + "angle": 0, + "content": "3.3. Lookahead \\(\\times\\) Convergence? Definitely No." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.522 + ], + "angle": 0, + "content": "From the above studies, we learn that fitting the LoRA model to the 3D model first is essential for VSD, while enhancing the convergence of the LoRA model is also beneficial to some degree. Here we conduct additional experiments to find out the consequences of combining these two factors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Based on the setting of L-VSD in Sec. 3.2, we increase the LoRA training steps \\(\\gamma\\) to 2 and 5. We also perform a study where the learning rate of LoRA is scaled to 1/10 of the original one to indicate under-convergence. Fig. 3b exhibits the results. As shown, increasing \\(\\gamma\\) in L-VSD exacerbates the phenomenon of over-saturation and makes the geometry and texture easier to collapse, while decreasing the learning rate of the LoRA model somehow improves the quality of the generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.855 + ], + "angle": 0, + "content": "These deviate from our expectations. To chase a deeper understanding, we examine the output of the LoRA by inspecting its norm \\(||\\epsilon_{\\phi}||\\), and plot its variation during training in Fig. 4b. We observe that the norm rapidly drops to zero for L-VSD with \\(\\gamma = 5\\), which implies ill-posed convergence. The underlying reason is probably that we maintain only one single 3D particle training of LoRA for efficiency, thus the distribution to fit by the LoRA model is biased. The pathology is more obvious for L-VSD when optimizing LoRA more intensively. Besides, we note that the trick of reducing the learning rate, despite being effective in this case, is unstable when handling harder prompts. Please refer to Appendix C.1 for more failure cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Conclusion Although lookahead is important for generating 3D models with clear outlines and realistic texture, the risk is the possibility of the LoRA model over-fitting on the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.906, + 0.341 + ], + "angle": 0, + "content": "isolated particle, which sometimes exhibits flaws in geometry and texture after each optimization step. Given these, this work tries to figure out a way to interpolate between the original VSD with a relatively stable formation process and L-VSD with higher-quality outcomes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.651, + 0.365 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.477 + ], + "angle": 0, + "content": "In Sec 4.1, we conduct a rigorous comparison between VSD and L-VSD and perform thorough studies to gain an understanding of their difference. Please also refer to Appendix A for an illustrative overview of the training pipelines of VSD and L-VSD to understand their difference better. We then derive the novel Linearized Lookahead VSD (\\(L^2\\)-VSD) in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.762, + 0.497 + ], + "angle": 0, + "content": "4.1. Compare VSD with L-VSD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.534 + ], + "angle": 0, + "content": "We first lay out the details of the update rules of L-VSD. Concretely, L-VSD first updates \\(\\phi_{i}\\) with" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.542, + 0.921, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\phi_ {i + 1} = \\phi_ {i} - 2 \\eta \\Delta_ {\\phi_ {i}}, \\Delta_ {\\phi_ {i}} := \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.906, + 0.604 + ], + "angle": 0, + "content": "where \\(\\eta\\) denotes the learning rate of the LoRA model. Then, L-VSD updates \\(\\theta_{i}\\) with the updated LoRA model \\(\\epsilon_{\\phi_{i + 1}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.613, + 0.641, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {L - V S D} (\\theta_ {i}) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.631, + 0.926, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, \\epsilon , c} \\left[ \\right. w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\left. \\right| _ {\\theta = \\theta_ {i}} \\left. \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.719 + ], + "angle": 0, + "content": "We can decompose \\(\\epsilon_{\\phi_{i+1}}(x_t, t, c, y)\\) by Taylor series to understand the gap between the update rule of VSD (Eq. (5)) and that of L-VSD (Eq. (8)) for \\(\\theta_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.73, + 0.905, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right) \\\\ = \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + (\\phi_ {i + 1} - \\phi_ {i}) J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y) + \\dots \\\\ = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\underbrace {\\left(- 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} \\left(x _ {t} , t , c , y\\right)\\right)} _ {\\Delta \\epsilon_ {\\text {f i r s t}}} + \\underbrace {\\mathcal {O} \\left(\\Delta_ {\\phi_ {i}} ^ {2}\\right)} _ {\\Delta \\epsilon_ {\\text {h i g h}}} \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Namely, both the first-order term \\(\\Delta \\epsilon_{\\text{first}}\\) and the high-order one \\(\\Delta \\epsilon_{\\text{high}}\\) may contribute to the fast formation of geometry and texture in L-VSD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To disentangle their effects, we opt to plot how their norms vary during the training procedure. In particular," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.218, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.089, + 0.35, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.089, + 0.482, + 0.138 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "Figure 5. Visualization of \\(\\Delta \\epsilon_{\\text{first}}\\) and \\(\\Delta \\epsilon_{\\text{high}}\\). The left column in each group represents the decoded first-order term, while the right column represents the decoded high-order term. Prompt for each group: (left) \"a delicious hamburger\"; (mid) \"an astronaut riding a horse\";(right) \"an Iron man\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.233, + 0.484, + 0.354 + ], + "angle": 0, + "content": "we set \\(\\eta\\) to 0.01. We leave how to estimate \\(\\Delta \\epsilon_{\\text{first}}\\) to the next subsection and compute \\(\\Delta \\epsilon_{\\text{high}}\\) by \\(\\epsilon_{\\phi_{i+1}}(x_t, t, c, y) - \\epsilon_{\\phi_i}(x_t, t, c, y) - \\Delta \\epsilon_{\\text{first}}\\). As illustrated in Fig. 4c, \\(||\\Delta \\epsilon_{\\text{high}}||\\) is significantly larger than \\(||\\Delta \\epsilon_{\\text{first}}||\\). It is even larger than the norm of the entire LoRA model \\(||\\epsilon_{\\phi}||\\), which means that \\(\\epsilon_{\\phi_{i+1}}\\) is probably dominated by \\(\\Delta \\epsilon_{\\text{high}}\\). Moreover, the norm of \\(\\Delta \\epsilon_{\\text{high}}\\) varies in a considerable range, which indicates much greater randomness than the first-order term." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.465 + ], + "angle": 0, + "content": "To obtain a more intuitive understanding of the two terms, we also pass them through the decoder of the pretrained DM [22] to obtain visualizations of them, presented in Fig. 5. As shown, the visualization of \\(\\Delta \\epsilon_{\\text{first}}\\) indicates an object shape corresponding to the prompt, while \\(\\Delta \\epsilon_{\\text{high}}\\) is more random, and we cannot witness any semantic information within it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Based on all the above observations and inferences, one question arises: whether \\(\\Delta \\epsilon_{\\text{first}}\\) is the essential component for the appealing generations, and keeping it can improve score distillation further. We answer this in the next subsection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.342, + 0.578 + ], + "angle": 0, + "content": "4.2. Linearized Lookahead VSD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Let \\(\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)\\coloneqq \\epsilon_{\\phi_i}(x_t,t,c,y) + \\Delta \\epsilon_{first}\\). In fact, it corresponds to performing one-step SGD with \\((x_{t'},t')\\) under the denoising loss (Eq. (2)) to update the following linear noise-prediction model:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.675, + 0.484, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {\\phi} ^ {\\operatorname {l i n}} \\left(x _ {t}, t, c, y\\right) = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\left(\\phi - \\phi_ {i}\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.483, + 0.813 + ], + "angle": 0, + "content": "Due to the low complexity of the linear model, the risk of overfitting to the current training data is low, which properly addresses the problem of the original L-VSD. With these understandings, our method boils down to using only the linearized lookahead correction \\(\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)\\) for estimating the score function of noisy rendered images of \\(\\theta_{i}\\). i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.84, + 0.506, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta_ {i}} \\mathcal {L} ^ {*} (\\theta_ {i}) = \\\\ \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right] \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.093, + 0.565, + 0.105 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.113, + 0.913, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) \\right. \\\\ \\left. + 2 \\eta \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) \\left[ J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right) \\right]\\right) \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.299 + ], + "angle": 0, + "content": "Viewing \\( J_{\\phi_i}(x_{t'}, t', c, y) J_{\\phi_i}^T(x_t, t, c, y) \\in \\mathbb{R}^{d \\times d} \\) as a preconditioning matrix, the above update rule involves two score contrast terms—one corresponds to the original VSD objective and the other accounts for a linearized lookahead correction for practical iterative optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.526 + ], + "angle": 0, + "content": "Moreover, we clarify the term \\(\\Delta \\epsilon_{\\text{first}} := -2\\eta \\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)\\) is friendly to estimate. Concretely, \\(\\Delta_{\\phi_i}\\) is exactly the gradient to update the LoRA model, which a backward pass can calculate. Then, the vector-Jacobian product \\(\\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)\\) can be realized by a forward pass of the score model using forward-mode automatic differentiation, which is equipped in many deep learning frameworks [20]. As a result, \\(L^2\\)-VSD only needs a single additional forward pass of the LoRA model per iteration compared to VSD, which is much more efficient than updating the LoRA multiple times. More importantly, we can even use only the entities associated with the last layer of the LoRA model for estimating the vector-Jacobian product to achieve a trade-off between quality and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.534, + 0.645, + 0.551 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.612, + 0.568 + ], + "angle": 0, + "content": "5.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.74 + ], + "angle": 0, + "content": "In this section, we evaluate the efficacy of our proposed Linearized Lookahead VSD method on text-guided 3D generation. Our baseline approaches include SDS, VSD, ESD (Entropy Score Distillation) [28], and VSD-based HiFA [37], which include representative state-of-the-art methods. For a fair comparison, all experiments are benchmarked under the open-source framework threestudio [5]. To clearly demonstrate the superior performance brought by our method, we compare both the results produced with high resolution and low resolution. Please refer to Appendix B for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.736, + 0.765 + ], + "angle": 0, + "content": "5.2. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We compare our generation results with SDS, VSD, ESD, L-VSD, and HiFA in high resolution of 256 and low resolution of 64. We present some representative results without bias in Fig. 1 and in Fig. 6, and we refer readers to Appendix C for results of L-VSD and more cases, as L-VSD tends to fail in generating complete objects. It's noteworthy that we consider our method as a baseline similar to SDS and VSD. ESD and HiFA are state-of-the-art based on VSD, which try to improve certain properties of generation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19572" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.092, + 0.19, + 0.102 + ], + "angle": 0, + "content": "SDS" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.108, + 0.246, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.092, + 0.353, + 0.102 + ], + "angle": 0, + "content": "ESD" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.108, + 0.411, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.486, + 0.092, + 0.516, + 0.102 + ], + "angle": 0, + "content": "VSD" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.108, + 0.579, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.093, + 0.693, + 0.103 + ], + "angle": 0, + "content": "VSD-HiFA" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.108, + 0.735, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.093, + 0.844, + 0.103 + ], + "angle": 0, + "content": "\\(L^2\\mathrm{VSD}\\)" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.108, + 0.892, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.184, + 0.605, + 0.196 + ], + "angle": 0, + "content": "\"A red fire hydrant spraying water.\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.196, + 0.246, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.196, + 0.411, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.197, + 0.578, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.594, + 0.197, + 0.735, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.197, + 0.901, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.272, + 0.577, + 0.284 + ], + "angle": 0, + "content": "\"A ficus planted in a pot.\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.286, + 0.245, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.286, + 0.41, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.286, + 0.577, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.286, + 0.735, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.286, + 0.901, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.359, + 0.593, + 0.371 + ], + "angle": 0, + "content": "\"A completely destroyed car\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.375, + 0.245, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.375, + 0.41, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.375, + 0.577, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.375, + 0.735, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.374, + 0.901, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.448, + 0.728, + 0.461 + ], + "angle": 0, + "content": "\"A DSLR photo of a cracked egg with the yolk spilling out on a wooden table\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.468, + 0.687, + 0.481 + ], + "angle": 0, + "content": "Figure 6. More Qualitative comparison with high resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.483, + 0.785 + ], + "angle": 0, + "content": "quality by introducing novel techniques. We compare with them in order to position our performance more rigorously. We can find that our method can generate high-quality 3D assets with a more realistic appearance than baselines. As for VSD-HiFA, there exist more floating around the objects while ours do not. It can generate better textures, which is due to their explicit loss design on the image space. We demonstrate in Sec. 5.5 that our methods can be seamlessly combined with other techniques like ESD and HiFA. Compared with VSD and its variant ESD, our results have significantly more realistic appearances and reasonable geometry. Moreover, our method has better convergence, which means our results are more robust to optimization and do not suffer from generating floaters in space. It's clearly shown that with our linearized lookahead correction term, even only trained with low-resolution rendering settings, we can generate 3D scenes with realistic and detailed appearances. In conclusion, our method surpasses the other baseline methods significantly with linearized lookahead and achieves comparable results to state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.323, + 0.805 + ], + "angle": 0, + "content": "5.3.Quantitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We follow previous work [21, 34] to quantitatively evaluate the generation quality using the angle of CLIP similarity [6] and Frechet Inception Distance [7] and list the results in Table 1. Specifically, the CLIP similarity measures the cosine similarity between the rendered image embeddings of the generated 3D object and the text embeddings," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.588 + ], + "angle": 0, + "content": "and we calculate the angle. The FID measures the distance between the image distribution by randomly rendering 3D representation and the text-conditioned image distribution from the pretrained diffusion model. We use DreamFusion gallery prompts to ensure that there is no man-made bias in the evaluation. Please refer to Appendix B for more metric computation details." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.59, + 0.895, + 0.633 + ], + "angle": 0, + "content": "
SDSESDVSDL-VSDHiFAL2-VSD
Averaged CLIP sim (↓)0.3050.3160.3240.3370.3130.285
Averaged FID (↓)372.35315.15301.54496292.88284.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.714 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparison. \\((\\downarrow)\\) means the lower the better. We measure the scenes across 20 prompts, which are randomly sampled from DreamFusion's gallery and include scenes shown in Fig. 1, Fig. 6, and figures in appendix. The quantitative results align with the qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.667, + 0.74 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.655, + 0.761 + ], + "angle": 0, + "content": "5.4.1. Ablation on \\(\\eta\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study on the scale of \\(\\eta\\), as shown in Fig. 7a, where only \\(\\eta\\) changes while other parameters are unchanged. We still use the simple prompt \"a delicious hamburger\". It's demonstrated that our method is robust to the scale of \\(\\eta\\). It's worth noting that even if the norm of the first-order term \\(\\Delta \\epsilon_{\\text{first}}\\) is only around 1e-2 when we set \\(\\eta\\) to 1e-3, the results still improve a lot. So, even a minor correction for each iteration can lead to incredible improvement in such a long-term optimization. Ablation on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19573" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.488, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.169, + 0.486, + 0.194 + ], + "angle": 0, + "content": "(a) Ablation on the \\(\\eta\\) in \\(\\Delta \\epsilon_{\\text{first}}\\). From left to right, the results correspond to setting \\(\\eta\\) to 1e-3, 1e-2, 0.1, and 1 respectively." + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.089, + 0.905, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.169, + 0.902, + 0.194 + ], + "angle": 0, + "content": "(b) Ablation on last-layer approximation. From left to right, the results correspond to settings in (a), but with last-layer approximation." + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.199, + 0.651, + 0.213 + ], + "angle": 0, + "content": "Figure 7. Qualitative results of ablation studies." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.225, + 0.486, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Figure 8. Examples of combining ESD with \\( L^2 \\)-VSD. We can observe that in the case of \"sunflower\", by combining our method, we obtain a reasonable sunflower rather than a \"sunball\". In the case of \"bulldozer\", with a linearized lookahead, ESD can generate additional elements like bricks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.473, + 0.394 + ], + "angle": 0, + "content": "high-order term correction can be found in Appendix C.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.397, + 0.406, + 0.412 + ], + "angle": 0, + "content": "5.4.2. Ablation on Last-layer Approximation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.485, + 0.629 + ], + "angle": 0, + "content": "As stated above, we can use only the entities associated with the last layer of the LoRA model to further reduce additional time costs. Admittedly, this approach may result in performance loss. As the correction term is calculated with the Jacobian matrices of the LoRA model, if we only use the last-layer gradients to approximate, we actually ignore the variable updates within other layers, thus losing accuracy. We present the corresponding results in Fig. 7b. Although the outcomes tend to generate floaters, the realism of results still gets largely improved. Also, we believe the floaters can be eliminated in the following geometry refinement stage. We compare the computation efficiency with the baseline VSD in Appendix B.3. We can save a significant computation time cost by only using this approximation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.377, + 0.65 + ], + "angle": 0, + "content": "5.5. Connection with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.839 + ], + "angle": 0, + "content": "ESD [28] is dedicated to solving the Janus problem based on VSD. It maximizes the entropy of different views of the generated results, encouraging diversity across views. The gradient of ESD is theoretically equivalent to adopting the CFG [8] trick upon VSD. Our method can also be incorporated into ESD. We name this as \\( L^2 \\)-ESD. We present the corresponding results in Fig. 8. The prompts we use are \"a sunflower on a flowerpot\" and \"a bulldozer made out of toy bricks\" respectively. To avoid unnecessary trials, we simply set \\( \\lambda \\) to 0.5, which is recommended in ESD. As we can observe, compared with ESD, the 3D results are more photo-realistic with distinct shapes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "HiFA [37] is a state-of-the-art approach for high-quality 3D generation in a single-stage training based on basic methods. It distills denoising scores from pretrained models in both the image and latent spaces and proposes several" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.225, + 0.905, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.336 + ], + "angle": 0, + "content": "Figure 9. Examples of combining HiFA with \\( L^2 \\)-VSD. The colors are more realistic and the appearance of our method's results aligns better with the prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.442 + ], + "angle": 0, + "content": "techniques to improve NeRF generation, which are orthogonal to our method. We demonstrate our method's compatibility by combining it with HiFA, named as \\( L^2 \\)-HiFA. We present the results in Fig. 9. The prompts we use are \"an elephant skull\" and \"Pumpkin head zombie, skinny, highly detailed, photorealistic\" respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.474 + ], + "angle": 0, + "content": "In conclusion, it's believed that \\( L^2 \\)-VSD can be combined with other parallel techniques in the future." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.481, + 0.862, + 0.497 + ], + "angle": 0, + "content": "6. Discussion, Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.672 + ], + "angle": 0, + "content": "In this paper, we dive deep into the theory of VSD, having a comprehensive understanding of the inner process, and identify two potential directions for improvement. In terms of the algorithm formulation, we propose Linearized Lookahead Variational Score Distillation(\\(L^2\\)-VSD), a novel framework based on VSD that achieves state-of-the-art results on text-to-3D generation. The linearized lookahead term enables us to benefit from both better convergence and lookahead for the next iteration. More importantly, our method can be incorporated into any VSD-based framework in the future." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitations and broader impact. Firstly, although \\( L^2 \\)-VSD achieves remarkable improvement on text-to-3D results, the generation process still takes hours of time, which is a common issue for score distillation-based methods. We believe orthogonal improvements on distillation accelerating [35] can mitigate this problem. Secondly, while we find the first-order term shows a regular pattern and some similarity may exist between our method and SiD [36], which is briefly discussed in appendix section E, we have not yet been able to formulate a distribution-based objective that guides this optimization. Investigating the underlying reasons is of significant interest in the future. Lastly, like other generative models, our method may be utilized to generate fake and malicious content, which needs more attention and caution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19574" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was supported by NSF of China (Nos. 92470118, 62306176), Natural Science Foundation of Shanghai (No. 23ZR1428700), CCF-ALIMAMA TECH Kangaroo Fund (NO. CCF-ALIMAMA OF 2025010), and CCF-Zhipu Large Model Innovation Fund (No. CCF-Zhipu202412)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.339 + ], + "angle": 0, + "content": "[1] Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Singh, and Tim Rocktäschel. Genie: Generative interactive environments, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.484, + 0.382 + ], + "angle": 0, + "content": "[2] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[3] Matthew Conlen, Jeffrey Heer, Hillary Mushkin, and Scott Davidoff. Cinematic techniques in narrative visualization, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.479 + ], + "angle": 0, + "content": "[4] Chris Creed, Maadh Al-Kalbani, Arthur Theil, Sayan Sarcar, and Ian Williams. Inclusive ar/vr: accessibility barriers for immersive technologies. Universal Access in the Information Society, 23(1):59-73, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[5] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.4.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[6] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[10] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[11] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Chenhan Jiang. A survey on text-to-3d contents generation in the wild, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[13] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[14] Nikhil Ketkar, Jojo Moolayil, Nikhil Ketkar, and Jojo Moolayil. Automatic differentiation in deep learning. Deep Learning with Python: Learn Best Practices of Deep Learning Models with PyTorch, pages 133-145, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Sixu Li, Chaojian Li, Wenbo Zhu, Boyang, Yu, Yang, Zhao, Cheng Wan, Haoran You, Huihong Shi, Yingyan, and Lin. Instant-3d: Instant neural radiance field training towards ondevice ar/vr 3d reconstruction, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.265, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1–19. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[20] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.906, + 0.632 + ], + "angle": 0, + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[24] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[25] Boshi Tang, Jianan Wang, Zhiyong Wu, and Lei Zhang. Stable score distillation for high-quality 3d generation, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Martin Engelcke, Ryan Faulkner, Frankie Garcia, Charles Gbadamosi, Zhitao Gong, Lucy Gonzales, Kshitij Gupta, Karol Gregor, Arne Olav Hallingstad, Tim Harley, Sam Haves, Felix Hill, Ed Hirst, Drew A. Hudson, Jony Hudson, Steph Hughes-Fitt, Danilo J. Rezende, Mimi Jasarevic, Laura Kampis, Rosemary Ke, Thomas Keck, Junkyung Kim, Oscar Knagg, Kavya Kopparapu, Andrew Lampinen, Shane Legg, Alexander Lerchner, Marjorie Limont, Yulan Liu, Maria Loks-Thompson, Joseph Marino, Kathryn Martin Cussons, Loic Matthey, Siobhan Mcloughlin, Piermaria Mendolicchio, Hamza Merzic, Anna Mitenkova, Alexandre Moufarek, Valeria Oliveira, Yanko Oliveira, Hannah Openshaw, Renke Pan, Aneesh Pappu, Alex Platonov, Ollie Purkiss, David Reichert, John Reid, Pierre Harvey Richemond, Tyson Roberts, Giles Ruscoe, Jaume Sanchez Elias, Tasha Sandars, Daniel P. Sawyer, Tim Scholtes, Guy Simmons, Daniel Slater, Hubert Soyer, Heiko Strathmann, Peter Stys, Allison C. Tam, Denis Teptyashin, Tayfun Terzi, Davide Vercelli, Bojan Vujatovic, Marcus Wainwright, Jane X. Wang, Zhengdong Wang, Daan Wierstra, Duncan Williams, Nathaniel Wong, Sarah York, and Nick Young. Scaling instructable agents across many simulated worlds, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.454 + ], + "angle": 0, + "content": "[27] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[28] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming mode collapse in score distillation for text-to-3d generation, 2024. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[29] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[30] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[31] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8131-8141, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[32] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[33] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.829, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[34] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Linqi Zhou, Andy Shih, Chenlin Meng, and Stefano Ermon. Dreampropeller: Supercharge text-to-3d generation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "with parallel sampling. arXiv preprint arXiv:2311.17082, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[36] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. arXiv preprint arXiv:2404.04057, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[37] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19576" + } + ] +] \ No newline at end of file diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_origin.pdf b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..398bf846aa53c8fffb30e058cc2869fc39d5a09a --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/93194ba5-cad9-47e5-bf50-5b562a6b61ac_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3de15f86e5032b0a6c0ea7da59e18b6c463b56165ad5a561034aff1ce532ee06 +size 27923828 diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/full.md b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b95de54970e1d5efbacb5341a22aa7f75644b0d --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/full.md @@ -0,0 +1,435 @@ +# Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation + +Yu Lei $^{1}$ Bingde Liu $^{1}$ Qingsong Xie $^{2}$ Haonan Lu $^{2}$ Zhijie Deng $^{1*}$ $^{1}$ Shanghai Jiao Tong University OPPO AI Center +{tony-lei, bingde, zhijied}@sjtu.edu.cn {xieqingsong1, luhaonan}@oppo.com + +# Abstract + +Text-to-3D generation based on score distillation of pretrained 2D diffusion models has gained increasing interest, with variational score distillation (VSD) as a remarkable example. VSD proves that vanilla score distillation can be improved by introducing an extra score-based model, which characterizes the distribution of images rendered from 3D models, to correct the distillation gradient. Despite the theoretical foundations, VSD, in practice, is likely to suffer from slow and sometimes ill-posed convergence. In this paper, we perform an in-depth investigation of the interplay between the introduced score model and the 3D model, and find that there exists a mismatching problem between LoRA and 3D distributions in practical implementation. We can simply adjust their optimization order to improve the generation quality. By doing so, the score model looks ahead to the current 3D state and hence yields more reasonable corrections. Nevertheless, naive lookahead VSD may suffer from unstable training in practice due to the potential over-fitting. To address this, we propose to use a linearized variant of the model for score distillation, giving rise to the Linearized Lookahead Variational Score Distillation ( $L^2$ -VSD). $L^2$ -VSD can be realized efficiently with forward-mode autodiff functionalities of existing deep learning libraries. Extensive experiments validate the efficacy of $L^2$ -VSD, revealing its clear superiority over prior score distillation-based methods. We also show that our method can be seamlessly incorporated into any other VSD-based text-to-3D framework. + +# 1. Introduction + +3D content creation is important for a variety of applications, such as interactive gaming [1, 32], cinematic arts [3], AR/VR [4, 16], and building simulated environments for training agents in robotics [26]. However, it is still challenging and expensive to create high-quality 3D assets as + +it requires a high level of expertise. Therefore, automating this process with generative models has become an important problem [12], while remaining non-trivial due to the scarcity of data and complexity of 3D representations. + +Score distillation has emerged as an attractive way for 3D generation given textual condition [2, 17, 21, 28, 29]. It leverages pretrained 2D diffusion models [9, 22] to define priors to guide the evolution of 3D content without reliance on annotations. Score Distillation Sampling (SDS) [21] is a seminal work in this line, but it is widely criticized that its generations suffer from the over-smoothing issue. Variational Score Distillation (VSD) [29] remediates this by introducing an extra model that captures the score of the images rendered from the 3D model to correct the distillation gradient. However, VSD often requires a lengthy optimization through 3 stages: NeRF generation, geometry refinement, and texture refinement. The outcomes obtained in the initial stage are often blurry, prone to collapsing, and not directly applicable [30]. Though existing works begin to understand and improve SDS [13, 28, 34], there are great but fewer efforts dedicated to improving the more promising VSD [18, 31]. + +To identify the root of VSD's drawback, we conduct a comprehensive analysis of the interaction between the introduced score model and the 3D model, revealing that adjusting their optimization order can sometimes lead to a considerable enhancement in generation quality. This adjustment allows the score model to look ahead to the current 3D state, resulting in more accurate and sensible corrections for the distillation gradient. Yet, naive lookahead VSD can encounter unstable training due to the risk of the score model overfitting the single 3D particle and sampled camera view. + +To address this issue, we formally compare the correction gradients before and after looking ahead and identify two major differences—a linear first-order term and a high-order one. Upon closer examination, we observe that the former accommodates subtle semantic information, whereas the latter contains non-trivial high-frequency noises. Given these findings, we propose to use only the first-order term for correction, yielding Linearized Looka + +![](images/7d3b0fc460214293e8ba038790c5abc4bb9cca111bc6107f6bb0624b8873a71f.jpg) +SDS + +![](images/9c257d2bd731145e8f678f672749af66cdf9a823db8599f65355d4987c3836b2.jpg) +ESD + +![](images/b7a1ae5eb4f969194dd17ebda8ab3b49d85c0c1bc4e538608c41a9ed7a98baef.jpg) +VSD + +![](images/d7bdd383f924c808728d508a0a327640088326bb1677f9dce145370bd2686670.jpg) +VSD-HiFA + +![](images/3fa276734451d37c268a5e4060133dcb32582ead2a61833fb560d860697db2c8.jpg) +$L^{2}\mathrm{VSD}$ + +![](images/94b1de41d1b4198ade1a568fa248a561e45a696fbded2f940383fe496d73e305.jpg) + +![](images/30d2c53ba9303ea2ec18ee6edf1a1669737bdad2d64da669396e7045fd6b708d.jpg) +"A DSLR photo of a fox holding a videogame controller" + +![](images/e6a91a5ec9de5b9b9102fdf14fed87701df8885ac2e09f86e0f6bcfefef00c19.jpg) +"A hamburger" + +![](images/1f63c0f875be4d0c5ce2c4ca3c9d3ff21bec8ea9a66e1b0a2cc465487d6e3511.jpg) + +![](images/1aa008b842e50efdcd111ccd4fedc516e1fdc4ea0080240cc8cd03b85428b6a5.jpg) + +![](images/f079dd78bc6a074ef105de3d9bdba45aef52d6a02ec9b93a1ce64f009f74e8d4.jpg) + +![](images/16aa64529646b3a4aa3eca8db2f4f2a9ccef431c93a33ab6a63b61f97c7691c8.jpg) +"A humanoid robot lying on the couch while on a laptop" + +![](images/9544160126ef40a311aa50af5a5ca1466822f2cfc4655bbcd52a2a62770247f8.jpg) + +![](images/eba0e145701bd819cfc5d19d00cfe4531f91560f9689ddeb36b9ae2b3b5faa58.jpg) + +![](images/fcada5bb2068ed860bfefe1adfd040490d64d8a11379bec59afff9f9be9e3f56.jpg) + +![](images/ba5a94cc9db707b643889fcd293ebd7091260ff2fcd96f37ed2bc190fee0384a.jpg) +Figure 1. Qualitative comparison with high resolution of 256. $L^2$ -VSD can generate meticulously detailed and realistic 3D assets from easy to very complex prompts and scenes. $L^2$ -VSD tends to yield more complete structures than HiFA. + +![](images/97df17223b69a826bbb461f5daf914b8d85383b1e69e63df0e3dd88b42f490da.jpg) +"A zoomed out DSLR photo of Sydney opera house, aerial view" + +![](images/a9baf8e025a77ef1e889062b180035f24386c3f71d22daa6a854f5100907ab0e.jpg) + +![](images/63efc9f36f977474e9efc9ae69012835a08a54025bd3cc76d8f1069a107d0941.jpg) + +![](images/2f8d964175c6c116b3d5417f5203aba0d50ce28a19af22acc6b4364b78487881.jpg) + +head Variational Score Distillation ( $L^2$ -VSD), to reliably and consistently boost the generation quality of VSD. $L^2$ -VSD is both easy to implement and computationally efficient—the added linear term can be computed by only one additional forward process of the score model under the scope of forward-mode automatic differentiation, which is supported in many deep learning libraries [14, 20]. + +Through extensive experiments, we demonstrate the significant superiority of the proposed $L^2$ -VSD in improving 3D generation quality compared to competing baseline methods, as shown in Fig. 1. $L^2$ -VSD can even produce realistic results with low resolution directly in the first stage. Moreover, we empirically show that $L^2$ -VSD can be seamlessly integrated into other VSD-based 3D generation pipelines and combined with other parallel techniques for VSD, e.g., Entropy Score Distillation(ESD) [28], which mitigates the Janus problem, for further improvement. We summarize our technical contributions as follows: + +- For VSD, a fundamental method in text-to-3D generation, we carefully identify the gaps between its theory and implementation and analyze the potential impact the gaps may bring us, providing a direction for possible improvements. +- We propose $L^2$ -VSD, an easy-to-implement and compu + +![](images/706c2a086a603e2d319cc023547327e7afed13a0b2123ccf8b935b9f30bfc12f.jpg) +Figure 2. Comparison of convergence between VSD and L-VSD with an illustrative 2D Gaussian example. In this toy example we consider $x \in \mathbb{R}^2$ from a single Gaussian distribution. We optimize $q(x)$ towards ground truth distribution $p(x)$ and use $r(x)$ to function as LoRA, which is used to estimate $q(x)$ . This example validates the existence of a mismatching issue in VSD, and we leave the details in Sec. 3.2. + +![](images/4d9cfbf43f5fd6dbb0084bf9f572e453363e268b31c2b0b3770b0bbe967d7233.jpg) + +tationally efficient variant of VSD, which mitigates the mismatching problem to some extent. We demonstrate its significant improvement over baselines and are comparable to other state-of-the-art. +- We demonstrate that our method can be seamlessly combined with other VSD-based improving techniques without much effort. + +# 2. Background + +# 2.1. Diffusion Models + +Diffusion models (DMs) [9, 23, 24] are defined with a forward diffusion process on data $x \in \mathbb{R}^d$ with a Gaussian transition kernel. The conditional distribution at some timestep $t \in [0, T]$ usually satisfies + +$$ +q \left(x _ {t} \mid x _ {0}\right) = \mathcal {N} \left(x _ {t}; \alpha_ {t} x _ {0}, \sigma_ {t} ^ {2} \mathbf {I}\right) \tag {1} +$$ + +where $x_0 \coloneqq x$ , and $\alpha_{t}$ and $\sigma_{t}$ are pre-defined noise schedules. DMs learn a reverse diffusion process, specified by a parameterized distribution $p(x_{t - 1}|x_t)\coloneqq \mathcal{N}(x_{t - 1};\mu_\psi (x_t,t),\sigma_t^2\mathbf{I})$ with $\mu_{\psi}$ as a neural network (NN), to enable the sampling of generations from Gaussian noise. The training objective of $\mu_{\psi}$ is the variational lower bound of the log data likelihood. In practice, $\mu_{\psi}$ is reparameterized as a denoising network $\epsilon_{\psi}$ and the training loss can be further simplified as a Mean Squared Error (MSE) form [9, 15]: + +$$ +\mathcal {L} _ {D i f f} := \mathbb {E} _ {x, t, \epsilon} [ \omega (t) | | \epsilon_ {\psi} (\alpha_ {t} x + \sigma_ {t} \epsilon , t) - \epsilon | | _ {2} ^ {2} ], \tag {2} +$$ + +where $x$ follows the data distribution, $t$ is uniformly drawn from $[0,T]$ , $\epsilon$ is a standard Gaussian noise, and $\omega (t)$ is a time-dependent coefficient. $\epsilon_{\psi}$ also inherently connects to score matching [24, 27]. + +Classifier-free guidance (CFG) [8]. We can augment the model $\epsilon_{\psi}$ with an extra input, the condition $y$ , to characterize the corresponding conditional distribution, leaving $\epsilon_{\psi}(x_t,t,\emptyset)$ account for the original unconditional one. Then, we can resort to CFG to further boost the quality of conditional generation. Typically, CFG leverages the following term in the sampling process: + +$$ +\hat {\epsilon} _ {\psi} \left(x _ {t}, t, y\right) := (1 + s) \epsilon_ {\psi} \left(x _ {t}, t, y\right) - s \epsilon_ {\psi} \left(x _ {t}, t, \emptyset\right) \tag {3} +$$ + +where $s > 0$ refers to a guidance scale. + +# 2.2. Text-to-3D Generation with Score Distillation + +Text-to-3D generation aims to identify the parameters $\theta \in \mathbb{R}^N$ of a 3D model given a text condition $y$ . Neural radiance field (NeRF) [19] is a typical 3D representation based on neural networks. In particular, NeRF renders a new view of the scene with the input of a sequence of images as known views. Additionally, textured mesh can be applied to represent the geometry of a 3D object with triangle meshes and textures with color on the mesh surface. + +Denote $g(\theta, c)$ as the differential rendering function projecting the 3D scene to a 2D image given a camera angle $c$ . Score distillation approaches for text-to-3D generations demand the image sample $g(\theta, c)$ to respect the prior specified by a text-to-2D diffusion model $\epsilon_{\text{pretrain}}(\cdot, \cdot, y)$ pretrained on vast real text-image pairs, based on which the optimization goal is constructed [13, 21, 25, 28, 29, 33, 34]. + +Score Distillation Sampling (SDS) [21] updates the 3D model using view-dependent prompt $y^{c}$ : + +$$ +\nabla_ {\theta} \mathcal {L} _ {S D S} (\theta) := \mathbb {E} _ {t, \epsilon , c} [ \omega (t) \left(\epsilon_ {\text {p r e t r a i n}} \left(x _ {t}, t, y ^ {c}\right) - \epsilon\right) \frac {\partial g (\theta , c)}{\partial \theta} ], \tag {4} +$$ + +where $c$ is a randomly sampled camera angle and $x_{t} \coloneqq \alpha_{t} g(\theta, c) + \sigma_{t} \epsilon$ . The gradient is a simplification of that of the denoising objective w.r.t. $\theta$ [21]. Intuitively, it encourages the rendered images to move toward the high-probability regions of the pretrained model, thus a good 3D model emerges. SDS is a seminal work in the line of text-to-3D generation, but it is sensitive to the CFG scale $s$ [8]. A small $s$ often results in over-smooth outcomes, whereas a large $s$ leads to over-saturation. + +Variational Score Distillation (VSD) [29] addresses the issue of SDS with a thorough theoretical analysis and proposes a new algorithm. In particular, apart from the pretrained model $\epsilon_{\text{pretrain}}(x_t, t, y)$ for capturing the data distribution, VSD introduces a tunable model $\epsilon_{\phi}(x_t, t, c, y)$ , often instantiated as a LoRA [10] adaptation of $\epsilon_{\text{pretrain}}$ , to account for the distribution of images rendered from all possible 3D models given condition $y$ and camera angle $c$ . We refer the introduced model as the LoRA model hereinafter. VSD proves the LoRA model can reliably correct the original distillation gradient. + +VSD, in practice, usually considers only a single 3D particle $\theta$ and performs an iterative optimization of $\theta$ and $\phi$ until convergence. More specifically, let $\theta_{i}$ and $\phi_{i}$ denote the parameters at $i$ -th training iteration. VSD updates $\theta_{i}$ with the following gradient: + +$$ +\nabla_ {\theta_ {i}} \mathcal {L} _ {V S D} (\theta_ {i}) := \mathbb {E} _ {t, \epsilon , c} +$$ + +$$ +\left[ w (t) \left(\epsilon_ {p r e t r a i n} \left(x _ {t}, t, y\right) - \epsilon_ {\phi_ {i}} \left(x _ {t}, t, c, y\right)\right) \frac {\partial g (\theta , c)}{\partial \theta} \Big | _ {\theta = \theta_ {i}} \right], \tag {5} +$$ + +where $x_{t} = \alpha_{t}g(\theta_{i},c) + \sigma_{t}\epsilon$ . To avoid more than once differentiable rendering of the 3D model for efficiency, VSD updates $\phi_{i}$ still with $g(\theta_i,c)$ while using a different noisy state $x_{t^{\prime}} = \alpha_{t^{\prime}}g(\theta_{i},c) + \sigma_{t^{\prime}}\epsilon^{\prime}$ . The learning objective is the denoising loss defined in Eq. (2), whose gradient is: + +$$ +\begin{array}{l} \nabla_ {\phi_ {i}} \mathcal {L} _ {V S D} (\phi_ {i}) := \tag {6} \\ \mathbb {E} _ {t ^ {\prime}, \epsilon^ {\prime}, c} \left[ \left(\epsilon_ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) - \epsilon^ {\prime}\right) J _ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) \right]. \\ \end{array} +$$ + +where $J_{\phi_i}(x_{t'}, t', c, y) \coloneqq \frac{\partial \epsilon_\phi(x_{t'}, t', c, y)}{\partial \phi}|_{\phi = \phi_i}$ and we omit some time-dependent scaling factor. + +Apart from the compromise to maintain one 3D particle for efficiency, the practical algorithm of VSD exhibits several significant gaps from the theory: 1) one-step update cannot guarantee $\phi$ to converge in each iteration, and 2) the updates to $\theta_{i}$ are computed given the LoRA model $\phi_{i}$ , which characterizes the distribution associated with the previous 3D state $\theta_{i - 1}$ instead of $\theta_{i}$ . We hypothesize these + +![](images/9cb847eda96a244646352f071dea1534af88495da9eca23321e0e163aae99273.jpg) +(a) VSD: From left to right, the results correspond to setting $\gamma$ to 1, 2, 5, and 10. We can observe that the quality of the model is not fundamentally improved. + +![](images/d2e006c6201119b322950e3a20ab0e64c286780b8e224401eab014d5c8ddad18.jpg) +(b) L-VSD: From left to right, the first three results correspond to setting $\gamma$ to 1, 2, and 5; the last corresponds to scaling the learning rate by 0.1 with $\gamma = 1$ . +Figure 3. Qualitative Examples of training LoRA model for multiple steps per optimization iteration for VSD and L-VSD. + +are probably the root of the unstable performance and sometimes corrupted outcomes of VSD, and perform an in-depth investigation regarding them below. + +# 3. Diagnose the Issues of VSD + +To assess whether the identified gaps contribute to the issues of VSD, we conduct two sets of pivot experiments in this section. As introduced in Sec 2.2, we analyze the impact of convergence of LoRA in Sec 3.1; then, we correct the optimization order to see the consequences in Sec 3.2; lastly, we combine these two factors to check if they are interfered in Sec 3.3. We provide an illustrative 2D Gaussian example for better understanding and to evaluate the effectiveness of lookahead in Sec 3.2. We base the implementation on the open-source framework threestudio [5]. + +# 3.1. Better Convergence? Maybe No. + +Assumption 1 The one-step update in vanilla VSD cannot guarantee the LoRA model to converge well, thus possibly harming the effectiveness of score distillation. We assume updating the LoRA model for $\gamma (\gamma >1)$ steps in each iteration could alleviate the pathology. + +Experiments Setting We evaluate the effects of various $\gamma$ , including 1, 2, 5, and 10. In each step, we optimize the LoRA model with different noisy images under different views. Unless otherwise specified, we take NeRF as the default 3D representation and use a simple prompt, "a delicious hamburger", in the study. We keep all other hyperparameters the same as the vanilla VSD. Usually, the whole VSD process contains multiple stages, where in the first stage a NeRF is constructed and then the geometry and texture are refined, respectively. We directly report the first-stage learning outcomes because they establish the foundation for the following parts. + +Findings 1 We present the training loss of the LoRA model in Fig. 4a to indicate the convergence and the final generations in Fig. 3a to reflect if the issues still exist. As shown, although the loss curves for various $\gamma$ share a periodic rise and fall, the loss of a larger $\gamma$ (e.g., 5 or 10) is floating in a relatively smaller range, and $\gamma = 5$ roughly makes LoRA model converge. However, as shown in Fig. 3a, it's hard to tell if the shape becomes relatively more reasonable, and the overall quality of the 3D model does not witness a continual improvement as $\gamma$ rises. Thereby, we conclude + +that improving the convergence of the LoRA model on the original VSD is not sufficient, thus not being the key factor. + +# 3.2. Lookahead? Maybe Yes! + +Assumption 2 As shown in Eq. (5), the updates to $\theta_{i}$ are computed given $\phi_{i}$ , which characterizes the distribution associated with $\theta_{i-1}$ instead of $\theta_{i}$ . This is inconsistent with Theorem 2 of [29], where the LoRA model should first adapt to the current 3D model (i.e., $\theta_{i}$ ) to serve as a reliable score estimator for the corresponding distribution. Fixing such a mismatch may address the issues of VSD. + +Experiments Setting We update $\phi_{i}$ first to obtain $\phi_{i + 1}$ , based on which $\theta_{i}$ is updated. We name such a modification Lookahead-VSD (L-VSD) because it makes the LoRA model look ahead for one step compared to the original VSD. All other parameters are kept unchanged. + +Findings 2 We show the result in Fig. 3b. Intuitively, the rendered image has clearer edges compared to those in Fig. 3a. Besides, inspecting the optimization process, we find L-VSD acquires the geometries and textures for the 3D model more quickly than VSD, and the loss of the 3D model in L-VSD with $\gamma = 1$ floats in a similar level to VSD with $\gamma = 10$ . These results validate the necessity for the LoRA model to look ahead during the optimization of VSD. However, we also observe that the 3D model can easily suffer from being over-saturated as optimization continues, which means the 3D models can not converge with normal shapes and colors in L-VSD. + +Illustrative Example It seems weird and self-contradictory with only the unsatisfactory results above. Here we provide an illustrative 2D Gaussian example in Fig. 2, to show the existence of the mismatching problem, the effectiveness of lookahead, and point out the possible reason that harms the performance of L-VSD. In this experiment, we assume that $\theta = x \in \mathbb{R}^2$ , and preset a Gaussian distribution as ground truth, which should be taken as the pretrained distribution. We randomly initialize the parameterized distribution $q(x)$ , and use another Gaussian distribution $r(x)$ to approximate $q$ , which can be interpreted as LoRA in VSD. In each iteration, we randomly sample $x_{\text{sample}}$ from $q(x)$ , which is similar to differentiable rendering in VSD, and then use the mean and variance of a Gaussian distribution to calculate the optimization direction. The learning trajectories are illustrated in Fig. 2. The results confirm the mismatching + +![](images/f0c3328783ce29ca73eef3f66de0d6a5e3b4e362f5d509869f9d163a32f3a80a.jpg) +(a) Loss of the LoRA model during training given various $\gamma$ . When increasing $\gamma$ , the loss is relatively lower, while also showing periodic changes. + +![](images/7c1d655b475f5f7715decd053e1f75b8007d361c517c069b15d7578b757a5418.jpg) +(b) The norm of the LoRA prediction $||\epsilon_{\phi}||$ varies w.r.t. training iteration at various $\gamma$ in L-VSD. + +![](images/4597b38211e1ce0fc92b4221a81e57c1040859c65aa44005203dce8ef4bdd9e0.jpg) +(c) The norm of various scores during the training of L-VSD. +Figure 4. Curve Plots of Pivot Experimentsin Sec. 3. + +problem of VSD, which hinders the distribution matching process. Correcting the optimization order can lead to better results. But, if we overfit the $r(x)$ on the samples $x_{\text{sample}}$ , the results cannot converge normally towards the region of $p(x)$ when the timestep is small, which usually leads to color saturation in text-to-3D generation [11, 25]. This evidence supports our finding with the L-VSD above. More illustrative examples and complete runnable code can be found in Appendix A and F. + +# 3.3. Lookahead $\times$ Convergence? Definitely No. + +From the above studies, we learn that fitting the LoRA model to the 3D model first is essential for VSD, while enhancing the convergence of the LoRA model is also beneficial to some degree. Here we conduct additional experiments to find out the consequences of combining these two factors. + +Based on the setting of L-VSD in Sec. 3.2, we increase the LoRA training steps $\gamma$ to 2 and 5. We also perform a study where the learning rate of LoRA is scaled to 1/10 of the original one to indicate under-convergence. Fig. 3b exhibits the results. As shown, increasing $\gamma$ in L-VSD exacerbates the phenomenon of over-saturation and makes the geometry and texture easier to collapse, while decreasing the learning rate of the LoRA model somehow improves the quality of the generation. + +These deviate from our expectations. To chase a deeper understanding, we examine the output of the LoRA by inspecting its norm $||\epsilon_{\phi}||$ , and plot its variation during training in Fig. 4b. We observe that the norm rapidly drops to zero for L-VSD with $\gamma = 5$ , which implies ill-posed convergence. The underlying reason is probably that we maintain only one single 3D particle training of LoRA for efficiency, thus the distribution to fit by the LoRA model is biased. The pathology is more obvious for L-VSD when optimizing LoRA more intensively. Besides, we note that the trick of reducing the learning rate, despite being effective in this case, is unstable when handling harder prompts. Please refer to Appendix C.1 for more failure cases. + +Conclusion Although lookahead is important for generating 3D models with clear outlines and realistic texture, the risk is the possibility of the LoRA model over-fitting on the + +isolated particle, which sometimes exhibits flaws in geometry and texture after each optimization step. Given these, this work tries to figure out a way to interpolate between the original VSD with a relatively stable formation process and L-VSD with higher-quality outcomes. + +# 4. Methodology + +In Sec 4.1, we conduct a rigorous comparison between VSD and L-VSD and perform thorough studies to gain an understanding of their difference. Please also refer to Appendix A for an illustrative overview of the training pipelines of VSD and L-VSD to understand their difference better. We then derive the novel Linearized Lookahead VSD ( $L^2$ -VSD) in Sec. 4.2. + +# 4.1. Compare VSD with L-VSD + +We first lay out the details of the update rules of L-VSD. Concretely, L-VSD first updates $\phi_{i}$ with + +$$ +\phi_ {i + 1} = \phi_ {i} - 2 \eta \Delta_ {\phi_ {i}}, \Delta_ {\phi_ {i}} := \left(\epsilon_ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) - \epsilon^ {\prime}\right) J _ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) \tag {7} +$$ + +where $\eta$ denotes the learning rate of the LoRA model. Then, L-VSD updates $\theta_{i}$ with the updated LoRA model $\epsilon_{\phi_{i + 1}}$ : + +$$ +\nabla_ {\theta_ {i}} \mathcal {L} _ {L - V S D} (\theta_ {i}) = +$$ + +$$ +\mathbb {E} _ {t, \epsilon , c} \left[ \right. w (t) \left(\epsilon_ {\text {p r e t r a i n}} \left(x _ {t}, t, y\right) - \epsilon_ {\phi_ {i + 1}} \left(x _ {t}, t, c, y\right)\right) \frac {\partial g (\theta , c)}{\partial \theta} \left. \right| _ {\theta = \theta_ {i}} \left. \right]. \tag {8} +$$ + +We can decompose $\epsilon_{\phi_{i+1}}(x_t, t, c, y)$ by Taylor series to understand the gap between the update rule of VSD (Eq. (5)) and that of L-VSD (Eq. (8)) for $\theta_i$ : + +$$ +\begin{array}{l} \epsilon_ {\phi_ {i + 1}} \left(x _ {t}, t, c, y\right) \\ = \epsilon_ {\phi_ {i}} (x _ {t}, t, c, y) + (\phi_ {i + 1} - \phi_ {i}) J _ {\phi_ {i}} ^ {T} (x _ {t}, t, c, y) + \dots \\ = \epsilon_ {\phi_ {i}} \left(x _ {t}, t, c, y\right) + \underbrace {\left(- 2 \eta \Delta_ {\phi_ {i}} J _ {\phi_ {i}} ^ {T} \left(x _ {t} , t , c , y\right)\right)} _ {\Delta \epsilon_ {\text {f i r s t}}} + \underbrace {\mathcal {O} \left(\Delta_ {\phi_ {i}} ^ {2}\right)} _ {\Delta \epsilon_ {\text {h i g h}}} \tag {9} \\ \end{array} +$$ + +Namely, both the first-order term $\Delta \epsilon_{\text{first}}$ and the high-order one $\Delta \epsilon_{\text{high}}$ may contribute to the fast formation of geometry and texture in L-VSD. + +To disentangle their effects, we opt to plot how their norms vary during the training procedure. In particular, + +![](images/cc272166c354be7731f981c11052e445d5740da343e949c55baf4af373548562.jpg) +Figure 5. Visualization of $\Delta \epsilon_{\text{first}}$ and $\Delta \epsilon_{\text{high}}$ . The left column in each group represents the decoded first-order term, while the right column represents the decoded high-order term. Prompt for each group: (left) "a delicious hamburger"; (mid) "an astronaut riding a horse";(right) "an Iron man". + +![](images/21ae0f28761a55beb72dc0cf348490a4d07f0a2091138e29946b2f2630936b08.jpg) + +![](images/33bb2d6b2689d6fc1060b8c53d187910d94a90da6541d49451cd046b121b3411.jpg) + +we set $\eta$ to 0.01. We leave how to estimate $\Delta \epsilon_{\text{first}}$ to the next subsection and compute $\Delta \epsilon_{\text{high}}$ by $\epsilon_{\phi_{i+1}}(x_t, t, c, y) - \epsilon_{\phi_i}(x_t, t, c, y) - \Delta \epsilon_{\text{first}}$ . As illustrated in Fig. 4c, $||\Delta \epsilon_{\text{high}}||$ is significantly larger than $||\Delta \epsilon_{\text{first}}||$ . It is even larger than the norm of the entire LoRA model $||\epsilon_{\phi}||$ , which means that $\epsilon_{\phi_{i+1}}$ is probably dominated by $\Delta \epsilon_{\text{high}}$ . Moreover, the norm of $\Delta \epsilon_{\text{high}}$ varies in a considerable range, which indicates much greater randomness than the first-order term. + +To obtain a more intuitive understanding of the two terms, we also pass them through the decoder of the pretrained DM [22] to obtain visualizations of them, presented in Fig. 5. As shown, the visualization of $\Delta \epsilon_{\text{first}}$ indicates an object shape corresponding to the prompt, while $\Delta \epsilon_{\text{high}}$ is more random, and we cannot witness any semantic information within it. + +Based on all the above observations and inferences, one question arises: whether $\Delta \epsilon_{\text{first}}$ is the essential component for the appealing generations, and keeping it can improve score distillation further. We answer this in the next subsection. + +# 4.2. Linearized Lookahead VSD + +Let $\epsilon_{\phi_{i + 1}}^{\mathrm{lin}}(x_t,t,c,y)\coloneqq \epsilon_{\phi_i}(x_t,t,c,y) + \Delta \epsilon_{first}$ . In fact, it corresponds to performing one-step SGD with $(x_{t'},t')$ under the denoising loss (Eq. (2)) to update the following linear noise-prediction model: + +$$ +\epsilon_ {\phi} ^ {\operatorname {l i n}} \left(x _ {t}, t, c, y\right) = \epsilon_ {\phi_ {i}} \left(x _ {t}, t, c, y\right) + \left(\phi - \phi_ {i}\right) J _ {\phi_ {i}} ^ {T} \left(x _ {t}, t, c, y\right). \tag {10} +$$ + +Due to the low complexity of the linear model, the risk of overfitting to the current training data is low, which properly addresses the problem of the original L-VSD. With these understandings, our method boils down to using only the linearized lookahead correction $\epsilon_{\phi_{i + 1}}^{\mathrm{lin}}(x_t,t,c,y)$ for estimating the score function of noisy rendered images of $\theta_{i}$ . i.e., + +$$ +\begin{array}{l} \nabla_ {\theta_ {i}} \mathcal {L} ^ {*} (\theta_ {i}) = \\ \mathbb {E} _ {t, \epsilon , c} \left[ w (t) \left(\epsilon_ {\text {p r e t r a i n}} \left(x _ {t}, t, y\right) - \epsilon_ {\phi_ {i + 1}} ^ {\lim } \left(x _ {t}, t, c, y\right)\right) \frac {\partial g (\theta , c)}{\partial \theta} \Big | _ {\theta = \theta_ {i}} \right] \tag {11} \\ \end{array} +$$ + +where + +$$ +\begin{array}{l} \left(\epsilon_ {p r e t r a i n} \left(x _ {t}, t, y\right) - \epsilon_ {\phi_ {i + 1}} ^ {\lim } \left(x _ {t}, t, c, y\right)\right) \\ = \left(\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \epsilon_ {\phi_ {i}} (x _ {t}, t, c, y) + 2 \eta \Delta_ {\phi_ {i}} J _ {\phi_ {i}} ^ {T} (x _ {t}, t, c, y)\right) \\ = \left(\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \epsilon_ {\phi_ {i}} (x _ {t}, t, c, y) \right. \\ \left. + 2 \eta \left(\epsilon_ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) - \epsilon^ {\prime}\right) \left[ J _ {\phi_ {i}} \left(x _ {t ^ {\prime}}, t ^ {\prime}, c, y\right) J _ {\phi_ {i}} ^ {T} \left(x _ {t}, t, c, y\right) \right]\right) \tag {12} \\ \end{array} +$$ + +Viewing $J_{\phi_i}(x_{t'}, t', c, y) J_{\phi_i}^T(x_t, t, c, y) \in \mathbb{R}^{d \times d}$ as a preconditioning matrix, the above update rule involves two score contrast terms—one corresponds to the original VSD objective and the other accounts for a linearized lookahead correction for practical iterative optimization. + +Moreover, we clarify the term $\Delta \epsilon_{\text{first}} := -2\eta \Delta_{\phi_i} J_{\phi_i}^T(x_t, t, c, y)$ is friendly to estimate. Concretely, $\Delta_{\phi_i}$ is exactly the gradient to update the LoRA model, which a backward pass can calculate. Then, the vector-Jacobian product $\Delta_{\phi_i} J_{\phi_i}^T(x_t, t, c, y)$ can be realized by a forward pass of the score model using forward-mode automatic differentiation, which is equipped in many deep learning frameworks [20]. As a result, $L^2$ -VSD only needs a single additional forward pass of the LoRA model per iteration compared to VSD, which is much more efficient than updating the LoRA multiple times. More importantly, we can even use only the entities associated with the last layer of the LoRA model for estimating the vector-Jacobian product to achieve a trade-off between quality and efficiency. + +# 5. Experiments + +# 5.1. Settings + +In this section, we evaluate the efficacy of our proposed Linearized Lookahead VSD method on text-guided 3D generation. Our baseline approaches include SDS, VSD, ESD (Entropy Score Distillation) [28], and VSD-based HiFA [37], which include representative state-of-the-art methods. For a fair comparison, all experiments are benchmarked under the open-source framework threestudio [5]. To clearly demonstrate the superior performance brought by our method, we compare both the results produced with high resolution and low resolution. Please refer to Appendix B for more implementation details. + +# 5.2. Qualitative Comparison + +We compare our generation results with SDS, VSD, ESD, L-VSD, and HiFA in high resolution of 256 and low resolution of 64. We present some representative results without bias in Fig. 1 and in Fig. 6, and we refer readers to Appendix C for results of L-VSD and more cases, as L-VSD tends to fail in generating complete objects. It's noteworthy that we consider our method as a baseline similar to SDS and VSD. ESD and HiFA are state-of-the-art based on VSD, which try to improve certain properties of generation + +![](images/55fd417d3d16ab3269df0a8398d30f316780898fec4361d8b6309ba0e0a54bee.jpg) +SDS + +![](images/63d43fe71cd2459f50e540c7ffc2777f5a1e43b804d33375b83a5f9c6a52580c.jpg) +ESD + +![](images/fa0961d93774f1a90ec730b828213f2960e6543d0438597b017a4636115c4dfb.jpg) +VSD + +![](images/17c25c039b0e36b16c801eec5585c2696b0bc4e4d78a6c48880768dc9f9e5816.jpg) +VSD-HiFA + +![](images/3cc3c42068195db89cc9569473fae7ba60152aa2977443bfd948cbe415b0fa22.jpg) +$L^2\mathrm{VSD}$ + +![](images/0e2070fd0ac2cda253b6388b9a2f41ea14444c0b8a50047c3aff221888508dd3.jpg) + +![](images/235f7830a7bb6a1723f42245a22bddf4a9c71614147a4704774b61d6e9b6aaaf.jpg) +"A red fire hydrant spraying water." + +![](images/918f334c19bce008fb44b64a19b5c6594e2fa8eb30cc4468dc6d36f28cb43a99.jpg) +"A ficus planted in a pot." + +![](images/51e59be5137a136b4ece0082a022b5187cce15f638158940e8647a90622d5833.jpg) + +![](images/04dfbcad533d34278c80e48aeac7eab9d3a0c88268e4a45298d3d9821a5169d7.jpg) + +![](images/f8a9038b5cdb62eb857914788e19dab1bb7c4823c2d5f3dd36cd8dac39dce143.jpg) + +![](images/e6eb2a64417515fc669eb9069e01d91b102593487cca9fc1a51acf28c51c013f.jpg) + +![](images/45f383171f9d9bdfabbe7ed427c19c84932e7cf8ecfe565d83b69bb7f54127cd.jpg) +"A completely destroyed car" + +![](images/cc33bed8073459d458ada39c0d5bad3362d5f44d7a562ff41181cf2363cfd404.jpg) + +![](images/c3ad41f68c382fbb6267758ce1051fe47730a41302cfe0796417446bc0bf6dc5.jpg) + +![](images/1150da8a6c1dd7af86565e6e661902d96ab5a9b18ec2ad336ad020d54c159cad.jpg) + +![](images/8b01c2c0c0c7a744dc5d8cb7e98c3576597e1d70118c135f2eb5650945ff15c1.jpg) +"A DSLR photo of a cracked egg with the yolk spilling out on a wooden table" +Figure 6. More Qualitative comparison with high resolution. + +![](images/831a340c3076052af2f04274cd77382d5257c69e55be25781439d93b1321606d.jpg) + +![](images/208b8c978a65a339ec1fbef20e14939d6e43e66b10985e79293184d320ec16bf.jpg) + +![](images/2c514e95edba7d1ceb1fed7edbc93f28687c335dca06ee38a8c2b4c20bf6f6ed.jpg) + +quality by introducing novel techniques. We compare with them in order to position our performance more rigorously. We can find that our method can generate high-quality 3D assets with a more realistic appearance than baselines. As for VSD-HiFA, there exist more floating around the objects while ours do not. It can generate better textures, which is due to their explicit loss design on the image space. We demonstrate in Sec. 5.5 that our methods can be seamlessly combined with other techniques like ESD and HiFA. Compared with VSD and its variant ESD, our results have significantly more realistic appearances and reasonable geometry. Moreover, our method has better convergence, which means our results are more robust to optimization and do not suffer from generating floaters in space. It's clearly shown that with our linearized lookahead correction term, even only trained with low-resolution rendering settings, we can generate 3D scenes with realistic and detailed appearances. In conclusion, our method surpasses the other baseline methods significantly with linearized lookahead and achieves comparable results to state-of-the-art. + +# 5.3.Quantitative Comparison + +We follow previous work [21, 34] to quantitatively evaluate the generation quality using the angle of CLIP similarity [6] and Frechet Inception Distance [7] and list the results in Table 1. Specifically, the CLIP similarity measures the cosine similarity between the rendered image embeddings of the generated 3D object and the text embeddings, + +and we calculate the angle. The FID measures the distance between the image distribution by randomly rendering 3D representation and the text-conditioned image distribution from the pretrained diffusion model. We use DreamFusion gallery prompts to ensure that there is no man-made bias in the evaluation. Please refer to Appendix B for more metric computation details. + +
SDSESDVSDL-VSDHiFAL2-VSD
Averaged CLIP sim (↓)0.3050.3160.3240.3370.3130.285
Averaged FID (↓)372.35315.15301.54496292.88284.06
+ +Table 1. Quantitative Comparison. $(\downarrow)$ means the lower the better. We measure the scenes across 20 prompts, which are randomly sampled from DreamFusion's gallery and include scenes shown in Fig. 1, Fig. 6, and figures in appendix. The quantitative results align with the qualitative results. + +# 5.4. Ablation Study + +# 5.4.1. Ablation on $\eta$ + +We conduct an ablation study on the scale of $\eta$ , as shown in Fig. 7a, where only $\eta$ changes while other parameters are unchanged. We still use the simple prompt "a delicious hamburger". It's demonstrated that our method is robust to the scale of $\eta$ . It's worth noting that even if the norm of the first-order term $\Delta \epsilon_{\text{first}}$ is only around 1e-2 when we set $\eta$ to 1e-3, the results still improve a lot. So, even a minor correction for each iteration can lead to incredible improvement in such a long-term optimization. Ablation on + +![](images/c9371a5608726abdfe3a7a03be835b82a300107748d8bc71b2f014d219fc8ea7.jpg) +(a) Ablation on the $\eta$ in $\Delta \epsilon_{\text{first}}$ . From left to right, the results correspond to setting $\eta$ to 1e-3, 1e-2, 0.1, and 1 respectively. + +![](images/7f1cd55eeda82783fcb47691870c8cf0fa4e81e7a2a4df58c087dfbbe83564ee.jpg) +(b) Ablation on last-layer approximation. From left to right, the results correspond to settings in (a), but with last-layer approximation. + +![](images/71d2142c30b18b0bc42dc8c6327b338f6a5ab92f7d498369555ed7bb3241b152.jpg) +Figure 7. Qualitative results of ablation studies. +Figure 8. Examples of combining ESD with $L^2$ -VSD. We can observe that in the case of "sunflower", by combining our method, we obtain a reasonable sunflower rather than a "sunball". In the case of "bulldozer", with a linearized lookahead, ESD can generate additional elements like bricks. + +high-order term correction can be found in Appendix C.4. + +# 5.4.2. Ablation on Last-layer Approximation + +As stated above, we can use only the entities associated with the last layer of the LoRA model to further reduce additional time costs. Admittedly, this approach may result in performance loss. As the correction term is calculated with the Jacobian matrices of the LoRA model, if we only use the last-layer gradients to approximate, we actually ignore the variable updates within other layers, thus losing accuracy. We present the corresponding results in Fig. 7b. Although the outcomes tend to generate floaters, the realism of results still gets largely improved. Also, we believe the floaters can be eliminated in the following geometry refinement stage. We compare the computation efficiency with the baseline VSD in Appendix B.3. We can save a significant computation time cost by only using this approximation. + +# 5.5. Connection with Other Methods + +ESD [28] is dedicated to solving the Janus problem based on VSD. It maximizes the entropy of different views of the generated results, encouraging diversity across views. The gradient of ESD is theoretically equivalent to adopting the CFG [8] trick upon VSD. Our method can also be incorporated into ESD. We name this as $L^2$ -ESD. We present the corresponding results in Fig. 8. The prompts we use are "a sunflower on a flowerpot" and "a bulldozer made out of toy bricks" respectively. To avoid unnecessary trials, we simply set $\lambda$ to 0.5, which is recommended in ESD. As we can observe, compared with ESD, the 3D results are more photo-realistic with distinct shapes. + +HiFA [37] is a state-of-the-art approach for high-quality 3D generation in a single-stage training based on basic methods. It distills denoising scores from pretrained models in both the image and latent spaces and proposes several + +![](images/09602a92e0edf871dbea8998e7fb656540e17549e30bb3e352c7bb68d3b9f3e4.jpg) +Figure 9. Examples of combining HiFA with $L^2$ -VSD. The colors are more realistic and the appearance of our method's results aligns better with the prompts. + +techniques to improve NeRF generation, which are orthogonal to our method. We demonstrate our method's compatibility by combining it with HiFA, named as $L^2$ -HiFA. We present the results in Fig. 9. The prompts we use are "an elephant skull" and "Pumpkin head zombie, skinny, highly detailed, photorealistic" respectively. + +In conclusion, it's believed that $L^2$ -VSD can be combined with other parallel techniques in the future. + +# 6. Discussion, Conclusion and Limitation + +In this paper, we dive deep into the theory of VSD, having a comprehensive understanding of the inner process, and identify two potential directions for improvement. In terms of the algorithm formulation, we propose Linearized Lookahead Variational Score Distillation( $L^2$ -VSD), a novel framework based on VSD that achieves state-of-the-art results on text-to-3D generation. The linearized lookahead term enables us to benefit from both better convergence and lookahead for the next iteration. More importantly, our method can be incorporated into any VSD-based framework in the future. + +Limitations and broader impact. Firstly, although $L^2$ -VSD achieves remarkable improvement on text-to-3D results, the generation process still takes hours of time, which is a common issue for score distillation-based methods. We believe orthogonal improvements on distillation accelerating [35] can mitigate this problem. Secondly, while we find the first-order term shows a regular pattern and some similarity may exist between our method and SiD [36], which is briefly discussed in appendix section E, we have not yet been able to formulate a distribution-based objective that guides this optimization. Investigating the underlying reasons is of significant interest in the future. Lastly, like other generative models, our method may be utilized to generate fake and malicious content, which needs more attention and caution. + +# Acknowledgements + +This work was supported by NSF of China (Nos. 92470118, 62306176), Natural Science Foundation of Shanghai (No. 23ZR1428700), CCF-ALIMAMA TECH Kangaroo Fund (NO. CCF-ALIMAMA OF 2025010), and CCF-Zhipu Large Model Innovation Fund (No. CCF-Zhipu202412). + +# References + +[1] Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Singh, and Tim Rocktäschel. Genie: Generative interactive environments, 2024. 1 +[2] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 1 +[3] Matthew Conlen, Jeffrey Heer, Hillary Mushkin, and Scott Davidoff. Cinematic techniques in narrative visualization, 2023. 1 +[4] Chris Creed, Maadh Al-Kalbani, Arthur Theil, Sayan Sarcar, and Ian Williams. Inclusive ar/vr: accessibility barriers for immersive technologies. Universal Access in the Information Society, 23(1):59-73, 2023. 1 +[5] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.4.6 +[6] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 7 +[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 8 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 3 +[10] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 3 +[11] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023.5 +[12] Chenhan Jiang. A survey on text-to-3d contents generation in the wild, 2024. 1 + +[13] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation, 2023. 1, 3 +[14] Nikhil Ketkar, Jojo Moolayil, Nikhil Ketkar, and Jojo Moolayil. Automatic differentiation in deep learning. Deep Learning with Python: Learn Best Practices of Deep Learning Models with PyTorch, pages 133-145, 2021. 2 +[15] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. 3 +[16] Sixu Li, Chaojian Li, Wenbo Zhu, Boyang, Yu, Yang, Zhao, Cheng Wan, Haoran You, Huihong Shi, Yingyan, and Lin. Instant-3d: Instant neural radiance field training towards ondevice ar/vr 3d reconstruction, 2024. 1 +[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 1 +[18] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1–19. Springer, 2025. 1 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3 +[20] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 2, 6 +[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 7 +[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6 +[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3 +[24] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 3 +[25] Boshi Tang, Jianan Wang, Zhiyong Wu, and Lei Zhang. Stable score distillation for high-quality 3d generation, 2024. 3, 5 +[26] SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy, + +Martin Engelcke, Ryan Faulkner, Frankie Garcia, Charles Gbadamosi, Zhitao Gong, Lucy Gonzales, Kshitij Gupta, Karol Gregor, Arne Olav Hallingstad, Tim Harley, Sam Haves, Felix Hill, Ed Hirst, Drew A. Hudson, Jony Hudson, Steph Hughes-Fitt, Danilo J. Rezende, Mimi Jasarevic, Laura Kampis, Rosemary Ke, Thomas Keck, Junkyung Kim, Oscar Knagg, Kavya Kopparapu, Andrew Lampinen, Shane Legg, Alexander Lerchner, Marjorie Limont, Yulan Liu, Maria Loks-Thompson, Joseph Marino, Kathryn Martin Cussons, Loic Matthey, Siobhan Mcloughlin, Piermaria Mendolicchio, Hamza Merzic, Anna Mitenkova, Alexandre Moufarek, Valeria Oliveira, Yanko Oliveira, Hannah Openshaw, Renke Pan, Aneesh Pappu, Alex Platonov, Ollie Purkiss, David Reichert, John Reid, Pierre Harvey Richemond, Tyson Roberts, Giles Ruscoe, Jaume Sanchez Elias, Tasha Sandars, Daniel P. Sawyer, Tim Scholtes, Guy Simmons, Daniel Slater, Hubert Soyer, Heiko Strathmann, Peter Stys, Allison C. Tam, Denis Teptyashin, Tayfun Terzi, Davide Vercelli, Bojan Vujatovic, Marcus Wainwright, Jane X. Wang, Zhengdong Wang, Daan Wierstra, Duncan Williams, Nathaniel Wong, Sarah York, and Nick Young. Scaling instructable agents across many simulated worlds, 2024. 1 +[27] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3 +[28] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming mode collapse in score distillation for text-to-3d generation, 2024. 1, 2, 3, 6, 8 +[29] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 1, 3, 4 +[30] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan, 2023. 1 +[31] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8131-8141, 2024. 1 +[32] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video, 2024. 1 +[33] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting, 2023. 3 +[34] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation, 2023. 1, 3, 7 +[35] Linqi Zhou, Andy Shih, Chenlin Meng, and Stefano Ermon. Dreampropeller: Supercharge text-to-3d generation + +with parallel sampling. arXiv preprint arXiv:2311.17082, 2023.8 +[36] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. arXiv preprint arXiv:2404.04057, 2024.8 +[37] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 6, 8 \ No newline at end of file diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/images.zip b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..68f45307f5d08e1c22407734c75f7ba63833fc0f --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08df9c2e38de7ca22201a2582091816320795ad016b8507edf6ee953a8808c49 +size 618514 diff --git a/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/layout.json b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6c492b72dd1f42fc168f92e383eca848acde8eb6 --- /dev/null +++ b/2025/Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation/layout.json @@ -0,0 +1,12471 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 467, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 467, + 138 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 467, + 138 + ], + "type": "text", + "content": "Advancing Text-to-3D Generation with Linearized Lookahead Variational Score Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": "Yu Lei" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": " Bingde Liu" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": " Qingsong Xie" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": " Haonan Lu" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": " Zhijie Deng" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 511, + 204 + ], + "type": "text", + "content": " Shanghai Jiao Tong University OPPO AI Center \n{tony-lei, bingde, zhijied}@sjtu.edu.cn {xieqingsong1, luhaonan}@oppo.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "text", + "content": "Text-to-3D generation based on score distillation of pretrained 2D diffusion models has gained increasing interest, with variational score distillation (VSD) as a remarkable example. VSD proves that vanilla score distillation can be improved by introducing an extra score-based model, which characterizes the distribution of images rendered from 3D models, to correct the distillation gradient. Despite the theoretical foundations, VSD, in practice, is likely to suffer from slow and sometimes ill-posed convergence. In this paper, we perform an in-depth investigation of the interplay between the introduced score model and the 3D model, and find that there exists a mismatching problem between LoRA and 3D distributions in practical implementation. We can simply adjust their optimization order to improve the generation quality. By doing so, the score model looks ahead to the current 3D state and hence yields more reasonable corrections. Nevertheless, naive lookahead VSD may suffer from unstable training in practice due to the potential over-fitting. To address this, we propose to use a linearized variant of the model for score distillation, giving rise to the Linearized Lookahead Variational Score Distillation (" + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "text", + "content": "-VSD). " + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "text", + "content": "-VSD can be realized efficiently with forward-mode autodiff functionalities of existing deep learning libraries. Extensive experiments validate the efficacy of " + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 56, + 256, + 296, + 591 + ], + "type": "text", + "content": "-VSD, revealing its clear superiority over prior score distillation-based methods. We also show that our method can be seamlessly incorporated into any other VSD-based text-to-3D framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 614, + 135, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 135, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 135, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 696 + ], + "type": "text", + "content": "3D content creation is important for a variety of applications, such as interactive gaming [1, 32], cinematic arts [3], AR/VR [4, 16], and building simulated environments for training agents in robotics [26]. However, it is still challenging and expensive to create high-quality 3D assets as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 280 + ], + "type": "text", + "content": "it requires a high level of expertise. Therefore, automating this process with generative models has become an important problem [12], while remaining non-trivial due to the scarcity of data and complexity of 3D representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 281, + 556, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 496 + ], + "type": "text", + "content": "Score distillation has emerged as an attractive way for 3D generation given textual condition [2, 17, 21, 28, 29]. It leverages pretrained 2D diffusion models [9, 22] to define priors to guide the evolution of 3D content without reliance on annotations. Score Distillation Sampling (SDS) [21] is a seminal work in this line, but it is widely criticized that its generations suffer from the over-smoothing issue. Variational Score Distillation (VSD) [29] remediates this by introducing an extra model that captures the score of the images rendered from the 3D model to correct the distillation gradient. However, VSD often requires a lengthy optimization through 3 stages: NeRF generation, geometry refinement, and texture refinement. The outcomes obtained in the initial stage are often blurry, prone to collapsing, and not directly applicable [30]. Though existing works begin to understand and improve SDS [13, 28, 34], there are great but fewer efforts dedicated to improving the more promising VSD [18, 31]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 556, + 617 + ], + "type": "text", + "content": "To identify the root of VSD's drawback, we conduct a comprehensive analysis of the interaction between the introduced score model and the 3D model, revealing that adjusting their optimization order can sometimes lead to a considerable enhancement in generation quality. This adjustment allows the score model to look ahead to the current 3D state, resulting in more accurate and sensible corrections for the distillation gradient. Yet, naive lookahead VSD can encounter unstable training due to the risk of the score model overfitting the single 3D particle and sampled camera view." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "To address this issue, we formally compare the correction gradients before and after looking ahead and identify two major differences—a linear first-order term and a high-order one. Upon closer examination, we observe that the former accommodates subtle semantic information, whereas the latter contains non-trivial high-frequency noises. Given these findings, we propose to use only the first-order term for correction, yielding Linearized Looka" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19567" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 83, + 150, + 144 + ], + "blocks": [ + { + "bbox": [ + 97, + 73, + 116, + 82 + ], + "lines": [ + { + "bbox": [ + 97, + 73, + 116, + 82 + ], + "spans": [ + { + "bbox": [ + 97, + 73, + 116, + 82 + ], + "type": "text", + "content": "SDS" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 83, + 150, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 83, + 150, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 83, + 150, + 144 + ], + "type": "image", + "image_path": "7d3b0fc460214293e8ba038790c5abc4bb9cca111bc6107f6bb0624b8873a71f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 159, + 83, + 250, + 144 + ], + "blocks": [ + { + "bbox": [ + 196, + 73, + 214, + 81 + ], + "lines": [ + { + "bbox": [ + 196, + 73, + 214, + 81 + ], + "spans": [ + { + "bbox": [ + 196, + 73, + 214, + 81 + ], + "type": "text", + "content": "ESD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 83, + 250, + 144 + ], + "lines": [ + { + "bbox": [ + 159, + 83, + 250, + 144 + ], + "spans": [ + { + "bbox": [ + 159, + 83, + 250, + 144 + ], + "type": "image", + "image_path": "9c257d2bd731145e8f678f672749af66cdf9a823db8599f65355d4987c3836b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 275, + 93, + 342, + 135 + ], + "blocks": [ + { + "bbox": [ + 295, + 73, + 314, + 82 + ], + "lines": [ + { + "bbox": [ + 295, + 73, + 314, + 82 + ], + "spans": [ + { + "bbox": [ + 295, + 73, + 314, + 82 + ], + "type": "text", + "content": "VSD" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 275, + 93, + 342, + 135 + ], + "lines": [ + { + "bbox": [ + 275, + 93, + 342, + 135 + ], + "spans": [ + { + "bbox": [ + 275, + 93, + 342, + 135 + ], + "type": "image", + "image_path": "b7a1ae5eb4f969194dd17ebda8ab3b49d85c0c1bc4e538608c41a9ed7a98baef.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 358, + 83, + 448, + 143 + ], + "blocks": [ + { + "bbox": [ + 384, + 73, + 424, + 82 + ], + "lines": [ + { + "bbox": [ + 384, + 73, + 424, + 82 + ], + "spans": [ + { + "bbox": [ + 384, + 73, + 424, + 82 + ], + "type": "text", + "content": "VSD-HiFA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 83, + 448, + 143 + ], + "lines": [ + { + "bbox": [ + 358, + 83, + 448, + 143 + ], + "spans": [ + { + "bbox": [ + 358, + 83, + 448, + 143 + ], + "type": "image", + "image_path": "d7bdd383f924c808728d508a0a327640088326bb1677f9dce145370bd2686670.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 458, + 83, + 547, + 144 + ], + "blocks": [ + { + "bbox": [ + 488, + 73, + 515, + 82 + ], + "lines": [ + { + "bbox": [ + 488, + 73, + 515, + 82 + ], + "spans": [ + { + "bbox": [ + 488, + 73, + 515, + 82 + ], + "type": "inline_equation", + "content": "L^{2}\\mathrm{VSD}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 458, + 83, + 547, + 144 + ], + "lines": [ + { + "bbox": [ + 458, + 83, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 458, + 83, + 547, + 144 + ], + "type": "image", + "image_path": "3fa276734451d37c268a5e4060133dcb32582ead2a61833fb560d860697db2c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 161, + 149, + 221 + ], + "blocks": [ + { + "bbox": [ + 59, + 161, + 149, + 221 + ], + "lines": [ + { + "bbox": [ + 59, + 161, + 149, + 221 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 149, + 221 + ], + "type": "image", + "image_path": "94b1de41d1b4198ade1a568fa248a561e45a696fbded2f940383fe496d73e305.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 160, + 161, + 250, + 221 + ], + "blocks": [ + { + "bbox": [ + 160, + 161, + 250, + 221 + ], + "lines": [ + { + "bbox": [ + 160, + 161, + 250, + 221 + ], + "spans": [ + { + "bbox": [ + 160, + 161, + 250, + 221 + ], + "type": "image", + "image_path": "30d2c53ba9303ea2ec18ee6edf1a1669737bdad2d64da669396e7045fd6b708d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 224, + 403, + 236 + ], + "lines": [ + { + "bbox": [ + 204, + 224, + 403, + 236 + ], + "spans": [ + { + "bbox": [ + 204, + 224, + 403, + 236 + ], + "type": "text", + "content": "\"A DSLR photo of a fox holding a videogame controller\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 260, + 161, + 350, + 221 + ], + "blocks": [ + { + "bbox": [ + 278, + 147, + 333, + 157 + ], + "lines": [ + { + "bbox": [ + 278, + 147, + 333, + 157 + ], + "spans": [ + { + "bbox": [ + 278, + 147, + 333, + 157 + ], + "type": "text", + "content": "\"A hamburger\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 161, + 350, + 221 + ], + "lines": [ + { + "bbox": [ + 260, + 161, + 350, + 221 + ], + "spans": [ + { + "bbox": [ + 260, + 161, + 350, + 221 + ], + "type": "image", + "image_path": "e6a91a5ec9de5b9b9102fdf14fed87701df8885ac2e09f86e0f6bcfefef00c19.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 359, + 160, + 449, + 221 + ], + "blocks": [ + { + "bbox": [ + 359, + 160, + 449, + 221 + ], + "lines": [ + { + "bbox": [ + 359, + 160, + 449, + 221 + ], + "spans": [ + { + "bbox": [ + 359, + 160, + 449, + 221 + ], + "type": "image", + "image_path": "1f63c0f875be4d0c5ce2c4ca3c9d3ff21bec8ea9a66e1b0a2cc465487d6e3511.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 459, + 160, + 547, + 220 + ], + "blocks": [ + { + "bbox": [ + 459, + 160, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 459, + 160, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 459, + 160, + 547, + 220 + ], + "type": "image", + "image_path": "1aa008b842e50efdcd111ccd4fedc516e1fdc4ea0080240cc8cd03b85428b6a5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 239, + 148, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 239, + 148, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 239, + 148, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 239, + 148, + 300 + ], + "type": "image", + "image_path": "f079dd78bc6a074ef105de3d9bdba45aef52d6a02ec9b93a1ce64f009f74e8d4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 159, + 239, + 249, + 300 + ], + "blocks": [ + { + "bbox": [ + 159, + 239, + 249, + 300 + ], + "lines": [ + { + "bbox": [ + 159, + 239, + 249, + 300 + ], + "spans": [ + { + "bbox": [ + 159, + 239, + 249, + 300 + ], + "type": "image", + "image_path": "16aa64529646b3a4aa3eca8db2f4f2a9ccef431c93a33ab6a63b61f97c7691c8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 302, + 402, + 312 + ], + "lines": [ + { + "bbox": [ + 203, + 302, + 402, + 312 + ], + "spans": [ + { + "bbox": [ + 203, + 302, + 402, + 312 + ], + "type": "text", + "content": "\"A humanoid robot lying on the couch while on a laptop\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 260, + 239, + 349, + 299 + ], + "blocks": [ + { + "bbox": [ + 260, + 239, + 349, + 299 + ], + "lines": [ + { + "bbox": [ + 260, + 239, + 349, + 299 + ], + "spans": [ + { + "bbox": [ + 260, + 239, + 349, + 299 + ], + "type": "image", + "image_path": "9544160126ef40a311aa50af5a5ca1466822f2cfc4655bbcd52a2a62770247f8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 359, + 240, + 447, + 300 + ], + "blocks": [ + { + "bbox": [ + 359, + 240, + 447, + 300 + ], + "lines": [ + { + "bbox": [ + 359, + 240, + 447, + 300 + ], + "spans": [ + { + "bbox": [ + 359, + 240, + 447, + 300 + ], + "type": "image", + "image_path": "eba0e145701bd819cfc5d19d00cfe4531f91560f9689ddeb36b9ae2b3b5faa58.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 458, + 240, + 547, + 300 + ], + "blocks": [ + { + "bbox": [ + 458, + 240, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 458, + 240, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 458, + 240, + 547, + 300 + ], + "type": "image", + "image_path": "fcada5bb2068ed860bfefe1adfd040490d64d8a11379bec59afff9f9be9e3f56.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 60, + 316, + 149, + 376 + ], + "blocks": [ + { + "bbox": [ + 60, + 316, + 149, + 376 + ], + "lines": [ + { + "bbox": [ + 60, + 316, + 149, + 376 + ], + "spans": [ + { + "bbox": [ + 60, + 316, + 149, + 376 + ], + "type": "image", + "image_path": "ba5a94cc9db707b643889fcd293ebd7091260ff2fcd96f37ed2bc190fee0384a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "type": "text", + "content": "Figure 1. Qualitative comparison with high resolution of 256. " + }, + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "type": "text", + "content": "-VSD can generate meticulously detailed and realistic 3D assets from easy to very complex prompts and scenes. " + }, + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 392, + 554, + 414 + ], + "type": "text", + "content": "-VSD tends to yield more complete structures than HiFA." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 162, + 316, + 251, + 376 + ], + "blocks": [ + { + "bbox": [ + 162, + 316, + 251, + 376 + ], + "lines": [ + { + "bbox": [ + 162, + 316, + 251, + 376 + ], + "spans": [ + { + "bbox": [ + 162, + 316, + 251, + 376 + ], + "type": "image", + "image_path": "97df17223b69a826bbb461f5daf914b8d85383b1e69e63df0e3dd88b42f490da.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 380, + 417, + 390 + ], + "lines": [ + { + "bbox": [ + 191, + 380, + 417, + 390 + ], + "spans": [ + { + "bbox": [ + 191, + 380, + 417, + 390 + ], + "type": "text", + "content": "\"A zoomed out DSLR photo of Sydney opera house, aerial view\"" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 262, + 316, + 351, + 376 + ], + "blocks": [ + { + "bbox": [ + 262, + 316, + 351, + 376 + ], + "lines": [ + { + "bbox": [ + 262, + 316, + 351, + 376 + ], + "spans": [ + { + "bbox": [ + 262, + 316, + 351, + 376 + ], + "type": "image", + "image_path": "a9baf8e025a77ef1e889062b180035f24386c3f71d22daa6a854f5100907ab0e.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 359, + 316, + 449, + 376 + ], + "blocks": [ + { + "bbox": [ + 359, + 316, + 449, + 376 + ], + "lines": [ + { + "bbox": [ + 359, + 316, + 449, + 376 + ], + "spans": [ + { + "bbox": [ + 359, + 316, + 449, + 376 + ], + "type": "image", + "image_path": "63efc9f36f977474e9efc9ae69012835a08a54025bd3cc76d8f1069a107d0941.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 460, + 317, + 549, + 376 + ], + "blocks": [ + { + "bbox": [ + 460, + 317, + 549, + 376 + ], + "lines": [ + { + "bbox": [ + 460, + 317, + 549, + 376 + ], + "spans": [ + { + "bbox": [ + 460, + 317, + 549, + 376 + ], + "type": "image", + "image_path": "2f8d964175c6c116b3d5417f5203aba0d50ce28a19af22acc6b4364b78487881.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "text", + "content": "head Variational Score Distillation (" + }, + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "text", + "content": "-VSD), to reliably and consistently boost the generation quality of VSD. " + }, + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 425, + 295, + 509 + ], + "type": "text", + "content": "-VSD is both easy to implement and computationally efficient—the added linear term can be computed by only one additional forward process of the score model under the scope of forward-mode automatic differentiation, which is supported in many deep learning libraries [14, 20]." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "text", + "content": "Through extensive experiments, we demonstrate the significant superiority of the proposed " + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "text", + "content": "-VSD in improving 3D generation quality compared to competing baseline methods, as shown in Fig. 1. " + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "text", + "content": "-VSD can even produce realistic results with low resolution directly in the first stage. Moreover, we empirically show that " + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 510, + 295, + 640 + ], + "type": "text", + "content": "-VSD can be seamlessly integrated into other VSD-based 3D generation pipelines and combined with other parallel techniques for VSD, e.g., Entropy Score Distillation(ESD) [28], which mitigates the Janus problem, for further improvement. We summarize our technical contributions as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "- For VSD, a fundamental method in text-to-3D generation, we carefully identify the gaps between its theory and implementation and analyze the potential impact the gaps may bring us, providing a direction for possible improvements." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- We propose " + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "-VSD, an easy-to-implement and compu" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 318, + 425, + 432, + 506 + ], + "blocks": [ + { + "bbox": [ + 318, + 425, + 432, + 506 + ], + "lines": [ + { + "bbox": [ + 318, + 425, + 432, + 506 + ], + "spans": [ + { + "bbox": [ + 318, + 425, + 432, + 506 + ], + "type": "image", + "image_path": "706c2a086a603e2d319cc023547327e7afed13a0b2123ccf8b935b9f30bfc12f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": "Figure 2. Comparison of convergence between VSD and L-VSD with an illustrative 2D Gaussian example. In this toy example we consider " + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": " from a single Gaussian distribution. We optimize " + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": " towards ground truth distribution " + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "inline_equation", + "content": "r(x)" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": " to function as LoRA, which is used to estimate " + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 529, + 555, + 605 + ], + "type": "text", + "content": ". This example validates the existence of a mismatching issue in VSD, and we leave the details in Sec. 3.2." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 438, + 425, + 552, + 506 + ], + "blocks": [ + { + "bbox": [ + 438, + 425, + 552, + 506 + ], + "lines": [ + { + "bbox": [ + 438, + 425, + 552, + 506 + ], + "spans": [ + { + "bbox": [ + 438, + 425, + 552, + 506 + ], + "type": "image", + "image_path": "4d9cfbf43f5fd6dbb0084bf9f572e453363e268b31c2b0b3770b0bbe967d7233.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 634, + 553, + 717 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 322, + 634, + 553, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 634, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 322, + 634, + 553, + 681 + ], + "type": "text", + "content": "tationally efficient variant of VSD, which mitigates the mismatching problem to some extent. We demonstrate its significant improvement over baselines and are comparable to other state-of-the-art." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 314, + 682, + 553, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 682, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 314, + 682, + 553, + 717 + ], + "type": "text", + "content": "- We demonstrate that our method can be seamlessly combined with other VSD-based improving techniques without much effort." + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19568" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 135, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 135, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 135, + 85 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 159, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 159, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 159, + 102 + ], + "type": "text", + "content": "2.1. Diffusion Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "text", + "content": "Diffusion models (DMs) [9, 23, 24] are defined with a forward diffusion process on data " + }, + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "text", + "content": " with a Gaussian transition kernel. The conditional distribution at some timestep " + }, + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 55, + 107, + 296, + 157 + ], + "type": "text", + "content": " usually satisfies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 164, + 295, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 164, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 295, + 178 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {0}\\right) = \\mathcal {N} \\left(x _ {t}; \\alpha_ {t} x _ {0}, \\sigma_ {t} ^ {2} \\mathbf {I}\\right) \\tag {1}", + "image_path": "4d51c93524239ccad778908af364aa057a081cc51cdd826d6c1f9fab1dcfce8c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "x_0 \\coloneqq x" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " are pre-defined noise schedules. DMs learn a reverse diffusion process, specified by a parameterized distribution " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "p(x_{t - 1}|x_t)\\coloneqq \\mathcal{N}(x_{t - 1};\\mu_\\psi (x_t,t),\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mu_{\\psi}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " as a neural network (NN), to enable the sampling of generations from Gaussian noise. The training objective of " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mu_{\\psi}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " is the variational lower bound of the log data likelihood. In practice, " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mu_{\\psi}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " is reparameterized as a denoising network " + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 55, + 186, + 296, + 307 + ], + "type": "text", + "content": " and the training loss can be further simplified as a Mean Squared Error (MSE) form [9, 15]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 314, + 295, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 314, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 73, + 314, + 295, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D i f f} := \\mathbb {E} _ {x, t, \\epsilon} [ \\omega (t) | | \\epsilon_ {\\psi} (\\alpha_ {t} x + \\sigma_ {t} \\epsilon , t) - \\epsilon | | _ {2} ^ {2} ], \\tag {2}", + "image_path": "a82a61c56e9675b2341f929c960afb74b294127590d9edcf1f0945dfc7f9f4c9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " follows the data distribution, " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " is uniformly drawn from " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "[0,T]" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " is a standard Gaussian noise, and " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\omega (t)" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " is a time-dependent coefficient. " + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 55, + 335, + 296, + 384 + ], + "type": "text", + "content": " also inherently connects to score matching [24, 27]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "content": "Classifier-free guidance (CFG) [8]. We can augment the model " + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}" + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "content": " with an extra input, the condition " + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "content": ", to characterize the corresponding conditional distribution, leaving " + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(x_t,t,\\emptyset)" + }, + { + "bbox": [ + 55, + 384, + 296, + 468 + ], + "type": "text", + "content": " account for the original unconditional one. Then, we can resort to CFG to further boost the quality of conditional generation. Typically, CFG leverages the following term in the sampling process:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 476, + 295, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 476, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 69, + 476, + 295, + 490 + ], + "type": "interline_equation", + "content": "\\hat {\\epsilon} _ {\\psi} \\left(x _ {t}, t, y\\right) := (1 + s) \\epsilon_ {\\psi} \\left(x _ {t}, t, y\\right) - s \\epsilon_ {\\psi} \\left(x _ {t}, t, \\emptyset\\right) \\tag {3}", + "image_path": "24728d591139377239893598913dc9d1fee3f4e28ced205f7d5c99e6c7d026f5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 214, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 214, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 214, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 214, + 510 + ], + "type": "inline_equation", + "content": "s > 0" + }, + { + "bbox": [ + 55, + 498, + 214, + 510 + ], + "type": "text", + "content": " refers to a guidance scale." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 291, + 529 + ], + "type": "text", + "content": "2.2. Text-to-3D Generation with Score Distillation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "content": "Text-to-3D generation aims to identify the parameters " + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\theta \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "content": " of a 3D model given a text condition " + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 533, + 296, + 629 + ], + "type": "text", + "content": ". Neural radiance field (NeRF) [19] is a typical 3D representation based on neural networks. In particular, NeRF renders a new view of the scene with the input of a sequence of images as known views. Additionally, textured mesh can be applied to represent the geometry of a 3D object with triangle meshes and textures with color on the mesh surface." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Denote " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "g(\\theta, c)" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " as the differential rendering function projecting the 3D scene to a 2D image given a camera angle " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Score distillation approaches for text-to-3D generations demand the image sample " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "g(\\theta, c)" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " to respect the prior specified by a text-to-2D diffusion model " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{pretrain}}(\\cdot, \\cdot, y)" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " pretrained on vast real text-image pairs, based on which the optimization goal is constructed [13, 21, 25, 28, 29, 33, 34]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 97 + ], + "type": "text", + "content": "Score Distillation Sampling (SDS) [21] updates the 3D model using view-dependent prompt " + }, + { + "bbox": [ + 314, + 72, + 555, + 97 + ], + "type": "inline_equation", + "content": "y^{c}" + }, + { + "bbox": [ + 314, + 72, + 555, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 104, + 559, + 137 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 104, + 559, + 137 + ], + "spans": [ + { + "bbox": [ + 314, + 104, + 559, + 137 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {S D S} (\\theta) := \\mathbb {E} _ {t, \\epsilon , c} [ \\omega (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y ^ {c}\\right) - \\epsilon\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} ], \\tag {4}", + "image_path": "94968a3d7723a3b5de92e3f883a9c4f843409cbeab6d07b9e7b88f8a24bdabd0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": " is a randomly sampled camera angle and " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "x_{t} \\coloneqq \\alpha_{t} g(\\theta, c) + \\sigma_{t} \\epsilon" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": ". The gradient is a simplification of that of the denoising objective w.r.t. " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": " [21]. Intuitively, it encourages the rendered images to move toward the high-probability regions of the pretrained model, thus a good 3D model emerges. SDS is a seminal work in the line of text-to-3D generation, but it is sensitive to the CFG scale " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": " [8]. A small " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": " often results in over-smooth outcomes, whereas a large " + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 138, + 555, + 244 + ], + "type": "text", + "content": " leads to over-saturation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": "Variational Score Distillation (VSD) [29] addresses the issue of SDS with a thorough theoretical analysis and proposes a new algorithm. In particular, apart from the pretrained model " + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{pretrain}}(x_t, t, y)" + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": " for capturing the data distribution, VSD introduces a tunable model " + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(x_t, t, c, y)" + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": ", often instantiated as a LoRA [10] adaptation of " + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\text{pretrain}}" + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": ", to account for the distribution of images rendered from all possible 3D models given condition " + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": " and camera angle " + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 245, + 554, + 376 + ], + "type": "text", + "content": ". We refer the introduced model as the LoRA model hereinafter. VSD proves the LoRA model can reliably correct the original distillation gradient." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": "VSD, in practice, usually considers only a single 3D particle " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " and performs an iterative optimization of " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " until convergence. More specifically, let " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " denote the parameters at " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": "-th training iteration. VSD updates " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": " with the following gradient:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 445, + 413, + 458 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 413, + 458 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 413, + 458 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {V S D} (\\theta_ {i}) := \\mathbb {E} _ {t, \\epsilon , c}", + "image_path": "075cf78c2cf99a6de7ae12103c1b35a2b1680a3284b93448ce3a369f8f4b9b2e.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 460, + 554, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 460, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 318, + 460, + 554, + 496 + ], + "type": "interline_equation", + "content": "\\left[ w (t) \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right], \\tag {5}", + "image_path": "46cb7b73c533c30a6950986a7cc8f6c9c3687697210befd2221ec2eee3b55b0f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "inline_equation", + "content": "x_{t} = \\alpha_{t}g(\\theta_{i},c) + \\sigma_{t}\\epsilon" + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": ". To avoid more than once differentiable rendering of the 3D model for efficiency, VSD updates " + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": " still with " + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "inline_equation", + "content": "g(\\theta_i,c)" + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": " while using a different noisy state " + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "inline_equation", + "content": "x_{t^{\\prime}} = \\alpha_{t^{\\prime}}g(\\theta_{i},c) + \\sigma_{t^{\\prime}}\\epsilon^{\\prime}" + }, + { + "bbox": [ + 313, + 497, + 554, + 557 + ], + "type": "text", + "content": ". The learning objective is the denoising loss defined in Eq. (2), whose gradient is:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 565, + 554, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 565, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 323, + 565, + 554, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\phi_ {i}} \\mathcal {L} _ {V S D} (\\phi_ {i}) := \\tag {6} \\\\ \\mathbb {E} _ {t ^ {\\prime}, \\epsilon^ {\\prime}, c} \\left[ \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\right]. \\\\ \\end{array}", + "image_path": "84d2e7a14d5634d5517477b424484124f0e76361ffbba58542c74401f4dfc189.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 602, + 554, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 554, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 554, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 602, + 554, + 630 + ], + "type": "inline_equation", + "content": "J_{\\phi_i}(x_{t'}, t', c, y) \\coloneqq \\frac{\\partial \\epsilon_\\phi(x_{t'}, t', c, y)}{\\partial \\phi}|_{\\phi = \\phi_i}" + }, + { + "bbox": [ + 313, + 602, + 554, + 630 + ], + "type": "text", + "content": " and we omit some time-dependent scaling factor." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Apart from the compromise to maintain one 3D particle for efficiency, the practical algorithm of VSD exhibits several significant gaps from the theory: 1) one-step update cannot guarantee " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " to converge in each iteration, and 2) the updates to " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " are computed given the LoRA model " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ", which characterizes the distribution associated with the previous 3D state " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i - 1}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ". We hypothesize these" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19569" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 60, + 302, + 122 + ], + "blocks": [ + { + "bbox": [ + 56, + 60, + 302, + 122 + ], + "lines": [ + { + "bbox": [ + 56, + 60, + 302, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 60, + 302, + 122 + ], + "type": "image", + "image_path": "9cb847eda96a244646352f071dea1534af88495da9eca23321e0e163aae99273.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 124, + 301, + 153 + ], + "lines": [ + { + "bbox": [ + 55, + 124, + 301, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 301, + 153 + ], + "type": "text", + "content": "(a) VSD: From left to right, the results correspond to setting " + }, + { + "bbox": [ + 55, + 124, + 301, + 153 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 124, + 301, + 153 + ], + "type": "text", + "content": " to 1, 2, 5, and 10. We can observe that the quality of the model is not fundamentally improved." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 60, + 552, + 123 + ], + "blocks": [ + { + "bbox": [ + 306, + 60, + 552, + 123 + ], + "lines": [ + { + "bbox": [ + 306, + 60, + 552, + 123 + ], + "spans": [ + { + "bbox": [ + 306, + 60, + 552, + 123 + ], + "type": "image", + "image_path": "d2e006c6201119b322950e3a20ab0e64c286780b8e224401eab014d5c8ddad18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "lines": [ + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "type": "text", + "content": "(b) L-VSD: From left to right, the first three results correspond to setting " + }, + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "type": "text", + "content": " to 1, 2, and 5; the last corresponds to scaling the learning rate by 0.1 with " + }, + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 306, + 124, + 552, + 153 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 157, + 538, + 169 + ], + "lines": [ + { + "bbox": [ + 70, + 157, + 538, + 169 + ], + "spans": [ + { + "bbox": [ + 70, + 157, + 538, + 169 + ], + "type": "text", + "content": "Figure 3. Qualitative Examples of training LoRA model for multiple steps per optimization iteration for VSD and L-VSD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 175, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 297, + 212 + ], + "type": "text", + "content": "are probably the root of the unstable performance and sometimes corrupted outcomes of VSD, and perform an in-depth investigation regarding them below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 223, + 211, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 211, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 211, + 237 + ], + "type": "text", + "content": "3. Diagnose the Issues of VSD" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 243, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 364 + ], + "type": "text", + "content": "To assess whether the identified gaps contribute to the issues of VSD, we conduct two sets of pivot experiments in this section. As introduced in Sec 2.2, we analyze the impact of convergence of LoRA in Sec 3.1; then, we correct the optimization order to see the consequences in Sec 3.2; lastly, we combine these two factors to check if they are interfered in Sec 3.3. We provide an illustrative 2D Gaussian example for better understanding and to evaluate the effectiveness of lookahead in Sec 3.2. We base the implementation on the open-source framework threestudio [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 372, + 231, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 231, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 231, + 384 + ], + "type": "text", + "content": "3.1. Better Convergence? Maybe No." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "text", + "content": "Assumption 1 The one-step update in vanilla VSD cannot guarantee the LoRA model to converge well, thus possibly harming the effectiveness of score distillation. We assume updating the LoRA model for " + }, + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\gamma (\\gamma >1)" + }, + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "text", + "content": " steps in each iteration could alleviate the pathology." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "content": "Experiments Setting We evaluate the effects of various " + }, + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 450, + 296, + 594 + ], + "type": "text", + "content": ", including 1, 2, 5, and 10. In each step, we optimize the LoRA model with different noisy images under different views. Unless otherwise specified, we take NeRF as the default 3D representation and use a simple prompt, \"a delicious hamburger\", in the study. We keep all other hyperparameters the same as the vanilla VSD. Usually, the whole VSD process contains multiple stages, where in the first stage a NeRF is constructed and then the geometry and texture are refined, respectively. We directly report the first-stage learning outcomes because they establish the foundation for the following parts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Findings 1 We present the training loss of the LoRA model in Fig. 4a to indicate the convergence and the final generations in Fig. 3a to reflect if the issues still exist. As shown, although the loss curves for various " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " share a periodic rise and fall, the loss of a larger " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " (e.g., 5 or 10) is floating in a relatively smaller range, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " roughly makes LoRA model converge. However, as shown in Fig. 3a, it's hard to tell if the shape becomes relatively more reasonable, and the overall quality of the 3D model does not witness a continual improvement as " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " rises. Thereby, we conclude" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 175, + 553, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 200 + ], + "type": "text", + "content": "that improving the convergence of the LoRA model on the original VSD is not sufficient, thus not being the key factor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 205, + 454, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 454, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 454, + 217 + ], + "type": "text", + "content": "3.2. Lookahead? Maybe Yes!" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": "Assumption 2 As shown in Eq. (5), the updates to " + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": " are computed given " + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": ", which characterizes the distribution associated with " + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\theta_{i-1}" + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": ". This is inconsistent with Theorem 2 of [29], where the LoRA model should first adapt to the current 3D model (i.e., " + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 223, + 553, + 306 + ], + "type": "text", + "content": ") to serve as a reliable score estimator for the corresponding distribution. Fixing such a mismatch may address the issues of VSD." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": "Experiments Setting We update " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": " first to obtain " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "\\phi_{i + 1}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": ", based on which " + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 307, + 553, + 366 + ], + "type": "text", + "content": " is updated. We name such a modification Lookahead-VSD (L-VSD) because it makes the LoRA model look ahead for one step compared to the original VSD. All other parameters are kept unchanged." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "content": "Findings 2 We show the result in Fig. 3b. Intuitively, the rendered image has clearer edges compared to those in Fig. 3a. Besides, inspecting the optimization process, we find L-VSD acquires the geometries and textures for the 3D model more quickly than VSD, and the loss of the 3D model in L-VSD with " + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\gamma = 1" + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "content": " floats in a similar level to VSD with " + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\gamma = 10" + }, + { + "bbox": [ + 313, + 366, + 554, + 510 + ], + "type": "text", + "content": ". These results validate the necessity for the LoRA model to look ahead during the optimization of VSD. However, we also observe that the 3D model can easily suffer from being over-saturated as optimization continues, which means the 3D models can not converge with normal shapes and colors in L-VSD." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "Illustrative Example It seems weird and self-contradictory with only the unsatisfactory results above. Here we provide an illustrative 2D Gaussian example in Fig. 2, to show the existence of the mismatching problem, the effectiveness of lookahead, and point out the possible reason that harms the performance of L-VSD. In this experiment, we assume that " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta = x \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", and preset a Gaussian distribution as ground truth, which should be taken as the pretrained distribution. We randomly initialize the parameterized distribution " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", and use another Gaussian distribution " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "r(x)" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " to approximate " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", which can be interpreted as LoRA in VSD. In each iteration, we randomly sample " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{\\text{sample}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": ", which is similar to differentiable rendering in VSD, and then use the mean and variance of a Gaussian distribution to calculate the optimization direction. The learning trajectories are illustrated in Fig. 2. The results confirm the mismatching" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19570" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 218, + 156 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 218, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 218, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 218, + 156 + ], + "type": "image", + "image_path": "f0c3328783ce29ca73eef3f66de0d6a5e3b4e362f5d509869f9d163a32f3a80a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "type": "text", + "content": "(a) Loss of the LoRA model during training given various " + }, + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "type": "text", + "content": ". When increasing " + }, + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 159, + 218, + 197 + ], + "type": "text", + "content": ", the loss is relatively lower, while also showing periodic changes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 385, + 156 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 385, + 156 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 385, + 156 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 385, + 156 + ], + "type": "image", + "image_path": "7c1d655b475f5f7715decd053e1f75b8007d361c517c069b15d7578b757a5418.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "lines": [ + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "spans": [ + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "type": "text", + "content": "(b) The norm of the LoRA prediction " + }, + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "type": "inline_equation", + "content": "||\\epsilon_{\\phi}||" + }, + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "type": "text", + "content": " varies w.r.t. training iteration at various " + }, + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 222, + 158, + 385, + 187 + ], + "type": "text", + "content": " in L-VSD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 70, + 547, + 156 + ], + "blocks": [ + { + "bbox": [ + 392, + 70, + 547, + 156 + ], + "lines": [ + { + "bbox": [ + 392, + 70, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 392, + 70, + 547, + 156 + ], + "type": "image", + "image_path": "4597b38211e1ce0fc92b4221a81e57c1040859c65aa44005203dce8ef4bdd9e0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 158, + 552, + 178 + ], + "lines": [ + { + "bbox": [ + 390, + 158, + 552, + 178 + ], + "spans": [ + { + "bbox": [ + 390, + 158, + 552, + 178 + ], + "type": "text", + "content": "(c) The norm of various scores during the training of L-VSD." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 204, + 196, + 405, + 208 + ], + "lines": [ + { + "bbox": [ + 204, + 196, + 405, + 208 + ], + "spans": [ + { + "bbox": [ + 204, + 196, + 405, + 208 + ], + "type": "text", + "content": "Figure 4. Curve Plots of Pivot Experimentsin Sec. 3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "text", + "content": "problem of VSD, which hinders the distribution matching process. Correcting the optimization order can lead to better results. But, if we overfit the " + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "inline_equation", + "content": "r(x)" + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "text", + "content": " on the samples " + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "inline_equation", + "content": "x_{\\text{sample}}" + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "text", + "content": ", the results cannot converge normally towards the region of " + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 54, + 209, + 296, + 317 + ], + "type": "text", + "content": " when the timestep is small, which usually leads to color saturation in text-to-3D generation [11, 25]. This evidence supports our finding with the L-VSD above. More illustrative examples and complete runnable code can be found in Appendix A and F." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 324, + 278, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 278, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 278, + 338 + ], + "type": "text", + "content": "3.3. Lookahead " + }, + { + "bbox": [ + 55, + 324, + 278, + 338 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 324, + 278, + 338 + ], + "type": "text", + "content": " Convergence? Definitely No." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 342, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 413 + ], + "type": "text", + "content": "From the above studies, we learn that fitting the LoRA model to the 3D model first is essential for VSD, while enhancing the convergence of the LoRA model is also beneficial to some degree. Here we conduct additional experiments to find out the consequences of combining these two factors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "text", + "content": "Based on the setting of L-VSD in Sec. 3.2, we increase the LoRA training steps " + }, + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "text", + "content": " to 2 and 5. We also perform a study where the learning rate of LoRA is scaled to 1/10 of the original one to indicate under-convergence. Fig. 3b exhibits the results. As shown, increasing " + }, + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 414, + 296, + 521 + ], + "type": "text", + "content": " in L-VSD exacerbates the phenomenon of over-saturation and makes the geometry and texture easier to collapse, while decreasing the learning rate of the LoRA model somehow improves the quality of the generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "content": "These deviate from our expectations. To chase a deeper understanding, we examine the output of the LoRA by inspecting its norm " + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "inline_equation", + "content": "||\\epsilon_{\\phi}||" + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "content": ", and plot its variation during training in Fig. 4b. We observe that the norm rapidly drops to zero for L-VSD with " + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "content": ", which implies ill-posed convergence. The underlying reason is probably that we maintain only one single 3D particle training of LoRA for efficiency, thus the distribution to fit by the LoRA model is biased. The pathology is more obvious for L-VSD when optimizing LoRA more intensively. Besides, we note that the trick of reducing the learning rate, despite being effective in this case, is unstable when handling harder prompts. Please refer to Appendix C.1 for more failure cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Conclusion Although lookahead is important for generating 3D models with clear outlines and realistic texture, the risk is the possibility of the LoRA model over-fitting on the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": "isolated particle, which sometimes exhibits flaws in geometry and texture after each optimization step. Given these, this work tries to figure out a way to interpolate between the original VSD with a relatively stable formation process and L-VSD with higher-quality outcomes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 274, + 398, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 398, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 398, + 289 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 295, + 555, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 377 + ], + "type": "text", + "content": "In Sec 4.1, we conduct a rigorous comparison between VSD and L-VSD and perform thorough studies to gain an understanding of their difference. Please also refer to Appendix A for an illustrative overview of the training pipelines of VSD and L-VSD to understand their difference better. We then derive the novel Linearized Lookahead VSD (" + }, + { + "bbox": [ + 313, + 295, + 555, + 377 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 295, + 555, + 377 + ], + "type": "text", + "content": "-VSD) in Sec. 4.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 380, + 466, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 466, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 466, + 393 + ], + "type": "text", + "content": "4.1. Compare VSD with L-VSD" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "content": "We first lay out the details of the update rules of L-VSD. Concretely, L-VSD first updates " + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 313, + 398, + 554, + 422 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 429, + 563, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 563, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 563, + 452 + ], + "type": "interline_equation", + "content": "\\phi_ {i + 1} = \\phi_ {i} - 2 \\eta \\Delta_ {\\phi_ {i}}, \\Delta_ {\\phi_ {i}} := \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) \\tag {7}", + "image_path": "68d33296105bb6de015377fa685904e8a6cb8226e085983b853abc3bb4233476.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "text", + "content": " denotes the learning rate of the LoRA model. Then, L-VSD updates " + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "text", + "content": " with the updated LoRA model " + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i + 1}}" + }, + { + "bbox": [ + 313, + 453, + 554, + 478 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 485, + 392, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 392, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 392, + 498 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta_ {i}} \\mathcal {L} _ {L - V S D} (\\theta_ {i}) =", + "image_path": "9d85d5eef15eab5897679de5f961bf0fa60793e61e0b4f1362d24dbf5d60e4a4.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 499, + 566, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 566, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 566, + 533 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, \\epsilon , c} \\left[ \\right. w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\left. \\right| _ {\\theta = \\theta_ {i}} \\left. \\right]. \\tag {8}", + "image_path": "2514fd0a53dfc36a2a24db5cb0764634bd1cb2230d1d8f7173a18516ba95af53.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "text", + "content": "We can decompose " + }, + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i+1}}(x_t, t, c, y)" + }, + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "text", + "content": " by Taylor series to understand the gap between the update rule of VSD (Eq. (5)) and that of L-VSD (Eq. (8)) for " + }, + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "inline_equation", + "content": "\\theta_i" + }, + { + "bbox": [ + 313, + 533, + 555, + 569 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 318, + 578, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 578, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 318, + 578, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\epsilon_ {\\phi_ {i + 1}} \\left(x _ {t}, t, c, y\\right) \\\\ = \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + (\\phi_ {i + 1} - \\phi_ {i}) J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y) + \\dots \\\\ = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\underbrace {\\left(- 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} \\left(x _ {t} , t , c , y\\right)\\right)} _ {\\Delta \\epsilon_ {\\text {f i r s t}}} + \\underbrace {\\mathcal {O} \\left(\\Delta_ {\\phi_ {i}} ^ {2}\\right)} _ {\\Delta \\epsilon_ {\\text {h i g h}}} \\tag {9} \\\\ \\end{array}", + "image_path": "ced83e9cd1eea9f75bca90ef1d856dda4ef4c1d9a4ff2f33490f75e660a90012.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": "Namely, both the first-order term " + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": " and the high-order one " + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 689 + ], + "type": "text", + "content": " may contribute to the fast formation of geometry and texture in L-VSD." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To disentangle their effects, we opt to plot how their norms vary during the training procedure. In particular," + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19571" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 133, + 109 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 133, + 109 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 133, + 109 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 133, + 109 + ], + "type": "image", + "image_path": "cc272166c354be7731f981c11052e445d5740da343e949c55baf4af373548562.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "text", + "content": "Figure 5. Visualization of " + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 55, + 118, + 295, + 173 + ], + "type": "text", + "content": ". The left column in each group represents the decoded first-order term, while the right column represents the decoded high-order term. Prompt for each group: (left) \"a delicious hamburger\"; (mid) \"an astronaut riding a horse\";(right) \"an Iron man\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 70, + 214, + 109 + ], + "blocks": [ + { + "bbox": [ + 138, + 70, + 214, + 109 + ], + "lines": [ + { + "bbox": [ + 138, + 70, + 214, + 109 + ], + "spans": [ + { + "bbox": [ + 138, + 70, + 214, + 109 + ], + "type": "image", + "image_path": "21ae0f28761a55beb72dc0cf348490a4d07f0a2091138e29946b2f2630936b08.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 294, + 109 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 294, + 109 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 294, + 109 + ], + "type": "image", + "image_path": "33bb2d6b2689d6fc1060b8c53d187910d94a90da6541d49451cd046b121b3411.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": "we set " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " to 0.01. We leave how to estimate " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " to the next subsection and compute " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i+1}}(x_t, t, c, y) - \\epsilon_{\\phi_i}(x_t, t, c, y) - \\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": ". As illustrated in Fig. 4c, " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "||\\Delta \\epsilon_{\\text{high}}||" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " is significantly larger than " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "||\\Delta \\epsilon_{\\text{first}}||" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": ". It is even larger than the norm of the entire LoRA model " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "||\\epsilon_{\\phi}||" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": ", which means that " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i+1}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " is probably dominated by " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": ". Moreover, the norm of " + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 55, + 184, + 296, + 280 + ], + "type": "text", + "content": " varies in a considerable range, which indicates much greater randomness than the first-order term." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "text", + "content": "To obtain a more intuitive understanding of the two terms, we also pass them through the decoder of the pretrained DM [22] to obtain visualizations of them, presented in Fig. 5. As shown, the visualization of " + }, + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "text", + "content": " indicates an object shape corresponding to the prompt, while " + }, + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{high}}" + }, + { + "bbox": [ + 55, + 285, + 296, + 368 + ], + "type": "text", + "content": " is more random, and we cannot witness any semantic information within it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": "Based on all the above observations and inferences, one question arises: whether " + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 373, + 296, + 433 + ], + "type": "text", + "content": " is the essential component for the appealing generations, and keeping it can improve score distillation further. We answer this in the next subsection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 209, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 209, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 209, + 457 + ], + "type": "text", + "content": "4.2. Linearized Lookahead VSD" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)\\coloneqq \\epsilon_{\\phi_i}(x_t,t,c,y) + \\Delta \\epsilon_{first}" + }, + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "text", + "content": ". In fact, it corresponds to performing one-step SGD with " + }, + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "inline_equation", + "content": "(x_{t'},t')" + }, + { + "bbox": [ + 55, + 467, + 296, + 517 + ], + "type": "text", + "content": " under the denoising loss (Eq. (2)) to update the following linear noise-prediction model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 534, + 296, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 534, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 58, + 534, + 296, + 561 + ], + "type": "interline_equation", + "content": "\\epsilon_ {\\phi} ^ {\\operatorname {l i n}} \\left(x _ {t}, t, c, y\\right) = \\epsilon_ {\\phi_ {i}} \\left(x _ {t}, t, c, y\\right) + \\left(\\phi - \\phi_ {i}\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right). \\tag {10}", + "image_path": "23b0c31ba0d8c17087fc00f4e40dc8e2c727d86ce6e40b01d1494d14db4764dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "text", + "content": "Due to the low complexity of the linear model, the risk of overfitting to the current training data is low, which properly addresses the problem of the original L-VSD. With these understandings, our method boils down to using only the linearized lookahead correction " + }, + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi_{i + 1}}^{\\mathrm{lin}}(x_t,t,c,y)" + }, + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "text", + "content": " for estimating the score function of noisy rendered images of " + }, + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 564, + 295, + 643 + ], + "type": "text", + "content": ". i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 309, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 309, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 309, + 714 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta_ {i}} \\mathcal {L} ^ {*} (\\theta_ {i}) = \\\\ \\mathbb {E} _ {t, \\epsilon , c} \\left[ w (t) \\left(\\epsilon_ {\\text {p r e t r a i n}} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\frac {\\partial g (\\theta , c)}{\\partial \\theta} \\Big | _ {\\theta = \\theta_ {i}} \\right] \\tag {11} \\\\ \\end{array}", + "image_path": "fea60d5e72b5eeeeb2eaf4c9619453c057f0f03a514fc00fb9d6836c979ef36a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 73, + 345, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 345, + 83 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 345, + 83 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 89, + 558, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 89, + 558, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 89, + 558, + 177 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left(\\epsilon_ {p r e t r a i n} \\left(x _ {t}, t, y\\right) - \\epsilon_ {\\phi_ {i + 1}} ^ {\\lim } \\left(x _ {t}, t, c, y\\right)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) + 2 \\eta \\Delta_ {\\phi_ {i}} J _ {\\phi_ {i}} ^ {T} (x _ {t}, t, c, y)\\right) \\\\ = \\left(\\epsilon_ {p r e t r a i n} (x _ {t}, t, y) - \\epsilon_ {\\phi_ {i}} (x _ {t}, t, c, y) \\right. \\\\ \\left. + 2 \\eta \\left(\\epsilon_ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) - \\epsilon^ {\\prime}\\right) \\left[ J _ {\\phi_ {i}} \\left(x _ {t ^ {\\prime}}, t ^ {\\prime}, c, y\\right) J _ {\\phi_ {i}} ^ {T} \\left(x _ {t}, t, c, y\\right) \\right]\\right) \\tag {12} \\\\ \\end{array}", + "image_path": "77566ad3db4be41dfbe309b7e4f6398502894be1b03a35fcfaf7fa258c91562a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 177, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 236 + ], + "type": "text", + "content": "Viewing " + }, + { + "bbox": [ + 313, + 177, + 553, + 236 + ], + "type": "inline_equation", + "content": "J_{\\phi_i}(x_{t'}, t', c, y) J_{\\phi_i}^T(x_t, t, c, y) \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 177, + 553, + 236 + ], + "type": "text", + "content": " as a preconditioning matrix, the above update rule involves two score contrast terms—one corresponds to the original VSD objective and the other accounts for a linearized lookahead correction for practical iterative optimization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "content": "Moreover, we clarify the term " + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}} := -2\\eta \\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)" + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "content": " is friendly to estimate. Concretely, " + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\Delta_{\\phi_i}" + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "content": " is exactly the gradient to update the LoRA model, which a backward pass can calculate. Then, the vector-Jacobian product " + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "inline_equation", + "content": "\\Delta_{\\phi_i} J_{\\phi_i}^T(x_t, t, c, y)" + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "content": " can be realized by a forward pass of the score model using forward-mode automatic differentiation, which is equipped in many deep learning frameworks [20]. As a result, " + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 238, + 554, + 416 + ], + "type": "text", + "content": "-VSD only needs a single additional forward pass of the LoRA model per iteration compared to VSD, which is much more efficient than updating the LoRA multiple times. More importantly, we can even use only the entities associated with the last layer of the LoRA model for estimating the vector-Jacobian product to achieve a trade-off between quality and efficiency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 394, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 394, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 394, + 436 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 437, + 374, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 374, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 374, + 449 + ], + "type": "text", + "content": "5.1. Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 454, + 554, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 586 + ], + "type": "text", + "content": "In this section, we evaluate the efficacy of our proposed Linearized Lookahead VSD method on text-guided 3D generation. Our baseline approaches include SDS, VSD, ESD (Entropy Score Distillation) [28], and VSD-based HiFA [37], which include representative state-of-the-art methods. For a fair comparison, all experiments are benchmarked under the open-source framework threestudio [5]. To clearly demonstrate the superior performance brought by our method, we compare both the results produced with high resolution and low resolution. Please refer to Appendix B for more implementation details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 593, + 450, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 450, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 450, + 605 + ], + "type": "text", + "content": "5.2. Qualitative Comparison" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "We compare our generation results with SDS, VSD, ESD, L-VSD, and HiFA in high resolution of 256 and low resolution of 64. We present some representative results without bias in Fig. 1 and in Fig. 6, and we refer readers to Appendix C for results of L-VSD and more cases, as L-VSD tends to fail in generating complete objects. It's noteworthy that we consider our method as a baseline similar to SDS and VSD. ESD and HiFA are state-of-the-art based on VSD, which try to improve certain properties of generation" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19572" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 85, + 150, + 141 + ], + "blocks": [ + { + "bbox": [ + 99, + 72, + 116, + 80 + ], + "lines": [ + { + "bbox": [ + 99, + 72, + 116, + 80 + ], + "spans": [ + { + "bbox": [ + 99, + 72, + 116, + 80 + ], + "type": "text", + "content": "SDS" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 85, + 150, + 141 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 150, + 141 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 150, + 141 + ], + "type": "image", + "image_path": "55fd417d3d16ab3269df0a8398d30f316780898fec4361d8b6309ba0e0a54bee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 159, + 85, + 251, + 141 + ], + "blocks": [ + { + "bbox": [ + 198, + 72, + 216, + 80 + ], + "lines": [ + { + "bbox": [ + 198, + 72, + 216, + 80 + ], + "spans": [ + { + "bbox": [ + 198, + 72, + 216, + 80 + ], + "type": "text", + "content": "ESD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 85, + 251, + 141 + ], + "lines": [ + { + "bbox": [ + 159, + 85, + 251, + 141 + ], + "spans": [ + { + "bbox": [ + 159, + 85, + 251, + 141 + ], + "type": "image", + "image_path": "63d43fe71cd2459f50e540c7ffc2777f5a1e43b804d33375b83a5f9c6a52580c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 261, + 85, + 354, + 141 + ], + "blocks": [ + { + "bbox": [ + 297, + 72, + 315, + 80 + ], + "lines": [ + { + "bbox": [ + 297, + 72, + 315, + 80 + ], + "spans": [ + { + "bbox": [ + 297, + 72, + 315, + 80 + ], + "type": "text", + "content": "VSD" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 85, + 354, + 141 + ], + "lines": [ + { + "bbox": [ + 261, + 85, + 354, + 141 + ], + "spans": [ + { + "bbox": [ + 261, + 85, + 354, + 141 + ], + "type": "image", + "image_path": "fa0961d93774f1a90ec730b828213f2960e6543d0438597b017a4636115c4dfb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 364, + 85, + 449, + 141 + ], + "blocks": [ + { + "bbox": [ + 383, + 73, + 424, + 81 + ], + "lines": [ + { + "bbox": [ + 383, + 73, + 424, + 81 + ], + "spans": [ + { + "bbox": [ + 383, + 73, + 424, + 81 + ], + "type": "text", + "content": "VSD-HiFA" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 364, + 85, + 449, + 141 + ], + "lines": [ + { + "bbox": [ + 364, + 85, + 449, + 141 + ], + "spans": [ + { + "bbox": [ + 364, + 85, + 449, + 141 + ], + "type": "image", + "image_path": "17c25c039b0e36b16c801eec5585c2696b0bc4e4d78a6c48880768dc9f9e5816.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 459, + 85, + 545, + 142 + ], + "blocks": [ + { + "bbox": [ + 489, + 73, + 516, + 81 + ], + "lines": [ + { + "bbox": [ + 489, + 73, + 516, + 81 + ], + "spans": [ + { + "bbox": [ + 489, + 73, + 516, + 81 + ], + "type": "inline_equation", + "content": "L^2\\mathrm{VSD}" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 85, + 545, + 142 + ], + "lines": [ + { + "bbox": [ + 459, + 85, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 459, + 85, + 545, + 142 + ], + "type": "image", + "image_path": "3cc3c42068195db89cc9569473fae7ba60152aa2977443bfd948cbe415b0fa22.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 155, + 150, + 212 + ], + "blocks": [ + { + "bbox": [ + 59, + 155, + 150, + 212 + ], + "lines": [ + { + "bbox": [ + 59, + 155, + 150, + 212 + ], + "spans": [ + { + "bbox": [ + 59, + 155, + 150, + 212 + ], + "type": "image", + "image_path": "0e2070fd0ac2cda253b6388b9a2f41ea14444c0b8a50047c3aff221888508dd3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 159, + 155, + 251, + 212 + ], + "blocks": [ + { + "bbox": [ + 244, + 145, + 370, + 155 + ], + "lines": [ + { + "bbox": [ + 244, + 145, + 370, + 155 + ], + "spans": [ + { + "bbox": [ + 244, + 145, + 370, + 155 + ], + "type": "text", + "content": "\"A red fire hydrant spraying water.\"" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 155, + 251, + 212 + ], + "lines": [ + { + "bbox": [ + 159, + 155, + 251, + 212 + ], + "spans": [ + { + "bbox": [ + 159, + 155, + 251, + 212 + ], + "type": "image", + "image_path": "235f7830a7bb6a1723f42245a22bddf4a9c71614147a4704774b61d6e9b6aaaf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 260, + 156, + 353, + 213 + ], + "blocks": [ + { + "bbox": [ + 260, + 156, + 353, + 213 + ], + "lines": [ + { + "bbox": [ + 260, + 156, + 353, + 213 + ], + "spans": [ + { + "bbox": [ + 260, + 156, + 353, + 213 + ], + "type": "image", + "image_path": "918f334c19bce008fb44b64a19b5c6594e2fa8eb30cc4468dc6d36f28cb43a99.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 215, + 353, + 224 + ], + "lines": [ + { + "bbox": [ + 260, + 215, + 353, + 224 + ], + "spans": [ + { + "bbox": [ + 260, + 215, + 353, + 224 + ], + "type": "text", + "content": "\"A ficus planted in a pot.\"" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 363, + 156, + 449, + 212 + ], + "blocks": [ + { + "bbox": [ + 363, + 156, + 449, + 212 + ], + "lines": [ + { + "bbox": [ + 363, + 156, + 449, + 212 + ], + "spans": [ + { + "bbox": [ + 363, + 156, + 449, + 212 + ], + "type": "image", + "image_path": "51e59be5137a136b4ece0082a022b5187cce15f638158940e8647a90622d5833.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 460, + 156, + 551, + 212 + ], + "blocks": [ + { + "bbox": [ + 460, + 156, + 551, + 212 + ], + "lines": [ + { + "bbox": [ + 460, + 156, + 551, + 212 + ], + "spans": [ + { + "bbox": [ + 460, + 156, + 551, + 212 + ], + "type": "image", + "image_path": "04dfbcad533d34278c80e48aeac7eab9d3a0c88268e4a45298d3d9821a5169d7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 226, + 149, + 282 + ], + "blocks": [ + { + "bbox": [ + 59, + 226, + 149, + 282 + ], + "lines": [ + { + "bbox": [ + 59, + 226, + 149, + 282 + ], + "spans": [ + { + "bbox": [ + 59, + 226, + 149, + 282 + ], + "type": "image", + "image_path": "f8a9038b5cdb62eb857914788e19dab1bb7c4823c2d5f3dd36cd8dac39dce143.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 160, + 226, + 250, + 281 + ], + "blocks": [ + { + "bbox": [ + 160, + 226, + 250, + 281 + ], + "lines": [ + { + "bbox": [ + 160, + 226, + 250, + 281 + ], + "spans": [ + { + "bbox": [ + 160, + 226, + 250, + 281 + ], + "type": "image", + "image_path": "e6eb2a64417515fc669eb9069e01d91b102593487cca9fc1a51acf28c51c013f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 261, + 226, + 353, + 282 + ], + "blocks": [ + { + "bbox": [ + 261, + 226, + 353, + 282 + ], + "lines": [ + { + "bbox": [ + 261, + 226, + 353, + 282 + ], + "spans": [ + { + "bbox": [ + 261, + 226, + 353, + 282 + ], + "type": "image", + "image_path": "45f383171f9d9bdfabbe7ed427c19c84932e7cf8ecfe565d83b69bb7f54127cd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 284, + 362, + 293 + ], + "lines": [ + { + "bbox": [ + 258, + 284, + 362, + 293 + ], + "spans": [ + { + "bbox": [ + 258, + 284, + 362, + 293 + ], + "type": "text", + "content": "\"A completely destroyed car\"" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 364, + 226, + 449, + 282 + ], + "blocks": [ + { + "bbox": [ + 364, + 226, + 449, + 282 + ], + "lines": [ + { + "bbox": [ + 364, + 226, + 449, + 282 + ], + "spans": [ + { + "bbox": [ + 364, + 226, + 449, + 282 + ], + "type": "image", + "image_path": "cc33bed8073459d458ada39c0d5bad3362d5f44d7a562ff41181cf2363cfd404.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 460, + 226, + 551, + 281 + ], + "blocks": [ + { + "bbox": [ + 460, + 226, + 551, + 281 + ], + "lines": [ + { + "bbox": [ + 460, + 226, + 551, + 281 + ], + "spans": [ + { + "bbox": [ + 460, + 226, + 551, + 281 + ], + "type": "image", + "image_path": "c3ad41f68c382fbb6267758ce1051fe47730a41302cfe0796417446bc0bf6dc5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 59, + 297, + 149, + 353 + ], + "blocks": [ + { + "bbox": [ + 59, + 297, + 149, + 353 + ], + "lines": [ + { + "bbox": [ + 59, + 297, + 149, + 353 + ], + "spans": [ + { + "bbox": [ + 59, + 297, + 149, + 353 + ], + "type": "image", + "image_path": "1150da8a6c1dd7af86565e6e661902d96ab5a9b18ec2ad336ad020d54c159cad.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 160, + 297, + 250, + 352 + ], + "blocks": [ + { + "bbox": [ + 160, + 297, + 250, + 352 + ], + "lines": [ + { + "bbox": [ + 160, + 297, + 250, + 352 + ], + "spans": [ + { + "bbox": [ + 160, + 297, + 250, + 352 + ], + "type": "image", + "image_path": "8b01c2c0c0c7a744dc5d8cb7e98c3576597e1d70118c135f2eb5650945ff15c1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 354, + 445, + 365 + ], + "lines": [ + { + "bbox": [ + 167, + 354, + 445, + 365 + ], + "spans": [ + { + "bbox": [ + 167, + 354, + 445, + 365 + ], + "type": "text", + "content": "\"A DSLR photo of a cracked egg with the yolk spilling out on a wooden table\"" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 370, + 420, + 380 + ], + "lines": [ + { + "bbox": [ + 188, + 370, + 420, + 380 + ], + "spans": [ + { + "bbox": [ + 188, + 370, + 420, + 380 + ], + "type": "text", + "content": "Figure 6. More Qualitative comparison with high resolution." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 261, + 297, + 353, + 352 + ], + "blocks": [ + { + "bbox": [ + 261, + 297, + 353, + 352 + ], + "lines": [ + { + "bbox": [ + 261, + 297, + 353, + 352 + ], + "spans": [ + { + "bbox": [ + 261, + 297, + 353, + 352 + ], + "type": "image", + "image_path": "831a340c3076052af2f04274cd77382d5257c69e55be25781439d93b1321606d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 364, + 297, + 449, + 352 + ], + "blocks": [ + { + "bbox": [ + 364, + 297, + 449, + 352 + ], + "lines": [ + { + "bbox": [ + 364, + 297, + 449, + 352 + ], + "spans": [ + { + "bbox": [ + 364, + 297, + 449, + 352 + ], + "type": "image", + "image_path": "208b8c978a65a339ec1fbef20e14939d6e43e66b10985e79293184d320ec16bf.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 460, + 296, + 551, + 352 + ], + "blocks": [ + { + "bbox": [ + 460, + 296, + 551, + 352 + ], + "lines": [ + { + "bbox": [ + 460, + 296, + 551, + 352 + ], + "spans": [ + { + "bbox": [ + 460, + 296, + 551, + 352 + ], + "type": "image", + "image_path": "2c514e95edba7d1ceb1fed7edbc93f28687c335dca06ee38a8c2b4c20bf6f6ed.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 382, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 295, + 621 + ], + "type": "text", + "content": "quality by introducing novel techniques. We compare with them in order to position our performance more rigorously. We can find that our method can generate high-quality 3D assets with a more realistic appearance than baselines. As for VSD-HiFA, there exist more floating around the objects while ours do not. It can generate better textures, which is due to their explicit loss design on the image space. We demonstrate in Sec. 5.5 that our methods can be seamlessly combined with other techniques like ESD and HiFA. Compared with VSD and its variant ESD, our results have significantly more realistic appearances and reasonable geometry. Moreover, our method has better convergence, which means our results are more robust to optimization and do not suffer from generating floaters in space. It's clearly shown that with our linearized lookahead correction term, even only trained with low-resolution rendering settings, we can generate 3D scenes with realistic and detailed appearances. In conclusion, our method surpasses the other baseline methods significantly with linearized lookahead and achieves comparable results to state-of-the-art." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 197, + 637 + ], + "type": "text", + "content": "5.3.Quantitative Comparison" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We follow previous work [21, 34] to quantitatively evaluate the generation quality using the angle of CLIP similarity [6] and Frechet Inception Distance [7] and list the results in Table 1. Specifically, the CLIP similarity measures the cosine similarity between the rendered image embeddings of the generated 3D object and the text embeddings," + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 383, + 555, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 465 + ], + "type": "text", + "content": "and we calculate the angle. The FID measures the distance between the image distribution by randomly rendering 3D representation and the text-conditioned image distribution from the pretrained diffusion model. We use DreamFusion gallery prompts to ensure that there is no man-made bias in the evaluation. Please refer to Appendix B for more metric computation details." + } + ] + } + ], + "index": 33 + }, + { + "type": "table", + "bbox": [ + 322, + 467, + 547, + 501 + ], + "blocks": [ + { + "bbox": [ + 322, + 467, + 547, + 501 + ], + "lines": [ + { + "bbox": [ + 322, + 467, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 322, + 467, + 547, + 501 + ], + "type": "table", + "html": "
SDSESDVSDL-VSDHiFAL2-VSD
Averaged CLIP sim (↓)0.3050.3160.3240.3370.3130.285
Averaged FID (↓)372.35315.15301.54496292.88284.06
", + "image_path": "8fbcdfb889793e1812fd20ad6b80c4321438b25ee237572755cdf651277aa8e0.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "type": "text", + "content": "Table 1. Quantitative Comparison. " + }, + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "type": "text", + "content": " means the lower the better. We measure the scenes across 20 prompts, which are randomly sampled from DreamFusion's gallery and include scenes shown in Fig. 1, Fig. 6, and figures in appendix. The quantitative results align with the qualitative results." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 408, + 586 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 590, + 400, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 400, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 400, + 602 + ], + "type": "text", + "content": "5.4.1. Ablation on " + }, + { + "bbox": [ + 313, + 590, + 400, + 602 + ], + "type": "inline_equation", + "content": "\\eta" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "We conduct an ablation study on the scale of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", as shown in Fig. 7a, where only " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " changes while other parameters are unchanged. We still use the simple prompt \"a delicious hamburger\". It's demonstrated that our method is robust to the scale of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". It's worth noting that even if the norm of the first-order term " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " is only around 1e-2 when we set " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " to 1e-3, the results still improve a lot. So, even a minor correction for each iteration can lead to incredible improvement in such a long-term optimization. Ablation on" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19573" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 298, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 133 + ], + "type": "image", + "image_path": "c9371a5608726abdfe3a7a03be835b82a300107748d8bc71b2f014d219fc8ea7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "text", + "content": "(a) Ablation on the " + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "inline_equation", + "content": "\\Delta \\epsilon_{\\text{first}}" + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "text", + "content": ". From left to right, the results correspond to setting " + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 133, + 297, + 153 + ], + "type": "text", + "content": " to 1e-3, 1e-2, 0.1, and 1 respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 311, + 70, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 311, + 70, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 311, + 70, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 311, + 70, + 553, + 133 + ], + "type": "image", + "image_path": "7f1cd55eeda82783fcb47691870c8cf0fa4e81e7a2a4df58c087dfbbe83564ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 133, + 552, + 153 + ], + "lines": [ + { + "bbox": [ + 310, + 133, + 552, + 153 + ], + "spans": [ + { + "bbox": [ + 310, + 133, + 552, + 153 + ], + "type": "text", + "content": "(b) Ablation on last-layer approximation. From left to right, the results correspond to settings in (a), but with last-layer approximation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 178, + 297, + 224 + ], + "blocks": [ + { + "bbox": [ + 212, + 157, + 398, + 168 + ], + "lines": [ + { + "bbox": [ + 212, + 157, + 398, + 168 + ], + "spans": [ + { + "bbox": [ + 212, + 157, + 398, + 168 + ], + "type": "text", + "content": "Figure 7. Qualitative results of ablation studies." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 178, + 297, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 178, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 297, + 224 + ], + "type": "image", + "image_path": "71d2142c30b18b0bc42dc8c6327b338f6a5ab92f7d498369555ed7bb3241b152.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 296, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 286 + ], + "type": "text", + "content": "Figure 8. Examples of combining ESD with " + }, + { + "bbox": [ + 55, + 232, + 296, + 286 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 232, + 296, + 286 + ], + "type": "text", + "content": "-VSD. We can observe that in the case of \"sunflower\", by combining our method, we obtain a reasonable sunflower rather than a \"sunball\". In the case of \"bulldozer\", with a linearized lookahead, ESD can generate additional elements like bricks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 299, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 289, + 312 + ], + "type": "text", + "content": "high-order term correction can be found in Appendix C.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 314, + 248, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 248, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 248, + 326 + ], + "type": "text", + "content": "5.4.2. Ablation on Last-layer Approximation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 329, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 296, + 498 + ], + "type": "text", + "content": "As stated above, we can use only the entities associated with the last layer of the LoRA model to further reduce additional time costs. Admittedly, this approach may result in performance loss. As the correction term is calculated with the Jacobian matrices of the LoRA model, if we only use the last-layer gradients to approximate, we actually ignore the variable updates within other layers, thus losing accuracy. We present the corresponding results in Fig. 7b. Although the outcomes tend to generate floaters, the realism of results still gets largely improved. Also, we believe the floaters can be eliminated in the following geometry refinement stage. We compare the computation efficiency with the baseline VSD in Appendix B.3. We can save a significant computation time cost by only using this approximation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 503, + 230, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 230, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 230, + 514 + ], + "type": "text", + "content": "5.5. Connection with Other Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "content": "ESD [28] is dedicated to solving the Janus problem based on VSD. It maximizes the entropy of different views of the generated results, encouraging diversity across views. The gradient of ESD is theoretically equivalent to adopting the CFG [8] trick upon VSD. Our method can also be incorporated into ESD. We name this as " + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "content": "-ESD. We present the corresponding results in Fig. 8. The prompts we use are \"a sunflower on a flowerpot\" and \"a bulldozer made out of toy bricks\" respectively. To avoid unnecessary trials, we simply set " + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 521, + 295, + 664 + ], + "type": "text", + "content": " to 0.5, which is recommended in ESD. As we can observe, compared with ESD, the 3D results are more photo-realistic with distinct shapes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "HiFA [37] is a state-of-the-art approach for high-quality 3D generation in a single-stage training based on basic methods. It distills denoising scores from pretrained models in both the image and latent spaces and proposes several" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 178, + 553, + 222 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 553, + 222 + ], + "type": "image", + "image_path": "09602a92e0edf871dbea8998e7fb656540e17549e30bb3e352c7bb68d3b9f3e4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 232, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 266 + ], + "type": "text", + "content": "Figure 9. Examples of combining HiFA with " + }, + { + "bbox": [ + 313, + 232, + 555, + 266 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 232, + 555, + 266 + ], + "type": "text", + "content": "-VSD. The colors are more realistic and the appearance of our method's results aligns better with the prompts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 350 + ], + "type": "text", + "content": "techniques to improve NeRF generation, which are orthogonal to our method. We demonstrate our method's compatibility by combining it with HiFA, named as " + }, + { + "bbox": [ + 313, + 278, + 555, + 350 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 278, + 555, + 350 + ], + "type": "text", + "content": "-HiFA. We present the results in Fig. 9. The prompts we use are \"an elephant skull\" and \"Pumpkin head zombie, skinny, highly detailed, photorealistic\" respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 351, + 554, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 375 + ], + "type": "text", + "content": "In conclusion, it's believed that " + }, + { + "bbox": [ + 313, + 351, + 554, + 375 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 351, + 554, + 375 + ], + "type": "text", + "content": "-VSD can be combined with other parallel techniques in the future." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 380, + 527, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 527, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 527, + 393 + ], + "type": "text", + "content": "6. Discussion, Conclusion and Limitation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "content": "In this paper, we dive deep into the theory of VSD, having a comprehensive understanding of the inner process, and identify two potential directions for improvement. In terms of the algorithm formulation, we propose Linearized Lookahead Variational Score Distillation(" + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 401, + 555, + 532 + ], + "type": "text", + "content": "-VSD), a novel framework based on VSD that achieves state-of-the-art results on text-to-3D generation. The linearized lookahead term enables us to benefit from both better convergence and lookahead for the next iteration. More importantly, our method can be incorporated into any VSD-based framework in the future." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "text", + "content": "Limitations and broader impact. Firstly, although " + }, + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "inline_equation", + "content": "L^2" + }, + { + "bbox": [ + 313, + 534, + 556, + 712 + ], + "type": "text", + "content": "-VSD achieves remarkable improvement on text-to-3D results, the generation process still takes hours of time, which is a common issue for score distillation-based methods. We believe orthogonal improvements on distillation accelerating [35] can mitigate this problem. Secondly, while we find the first-order term shows a regular pattern and some similarity may exist between our method and SiD [36], which is briefly discussed in appendix section E, we have not yet been able to formulate a distribution-based objective that guides this optimization. Investigating the underlying reasons is of significant interest in the future. Lastly, like other generative models, our method may be utilized to generate fake and malicious content, which needs more attention and caution." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19574" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported by NSF of China (Nos. 92470118, 62306176), Natural Science Foundation of Shanghai (No. 23ZR1428700), CCF-ALIMAMA TECH Kangaroo Fund (NO. CCF-ALIMAMA OF 2025010), and CCF-Zhipu Large Model Innovation Fund (No. CCF-Zhipu202412)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 268 + ], + "type": "text", + "content": "[1] Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang, Konrad Zolna, Jeff Clune, Nando de Freitas, Satinder Singh, and Tim Rocktäschel. Genie: Generative interactive environments, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 296, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 302 + ], + "type": "text", + "content": "[2] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 334 + ], + "type": "text", + "content": "[3] Matthew Conlen, Jeffrey Heer, Hillary Mushkin, and Scott Davidoff. Cinematic techniques in narrative visualization, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 296, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 379 + ], + "type": "text", + "content": "[4] Chris Creed, Maadh Al-Kalbani, Arthur Theil, Sayan Sarcar, and Ian Williams. Inclusive ar/vr: accessibility barriers for immersive technologies. Universal Access in the Information Society, 23(1):59-73, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 445 + ], + "type": "text", + "content": "[5] Yuan-Chen Guo, Ying-Tian Liu, Ruizhi Shao, Christian Laforte, Vikram Voleti, Guan Luo, Chia-Hao Chen, Zixin Zou, Chen Wang, Yan-Pei Cao, and Song-Hai Zhang. threestudio: A unified framework for 3d content generation. https://github.com/threestudio-project/threestudio, 2023.4.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 479 + ], + "type": "text", + "content": "[6] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: a reference-free evaluation metric for image captioning. In EMNLP, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "text", + "content": "[7] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 557 + ], + "type": "text", + "content": "[8] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 601 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 296, + 646 + ], + "type": "text", + "content": "[10] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[11] Yukun Huang, Jianan Wang, Yukai Shi, Xianbiao Qi, Zheng-Jun Zha, and Lei Zhang. Dreamtime: An improved optimization strategy for text-to-3d content creation. arXiv preprint arXiv:2306.12422, 2023.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 712 + ], + "type": "text", + "content": "[12] Chenhan Jiang. A survey on text-to-3d contents generation in the wild, 2024. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[13] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free score distillation, 2023. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 554, + 140 + ], + "type": "text", + "content": "[14] Nikhil Ketkar, Jojo Moolayil, Nikhil Ketkar, and Jojo Moolayil. Automatic differentiation in deep learning. Deep Learning with Python: Learn Best Practices of Deep Learning Models with PyTorch, pages 133-145, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "type": "text", + "content": "[15] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 207 + ], + "type": "text", + "content": "[16] Sixu Li, Chaojian Li, Wenbo Zhu, Boyang, Yu, Yang, Zhao, Cheng Wan, Haoran You, Huihong Shi, Yingyan, and Lin. Instant-3d: Instant neural radiance field training towards ondevice ar/vr 3d reconstruction, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "text", + "content": "[17] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 553, + 308 + ], + "type": "text", + "content": "[18] Zhiyuan Ma, Yuxiang Wei, Yabin Zhang, Xiangyu Zhu, Zhen Lei, and Lei Zhang. Scaledreamer: Scalable text-to-3d synthesis with asynchronous score distillation. In European Conference on Computer Vision, pages 1–19. Springer, 2025. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "text", + "content": "[20] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 401, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 444 + ], + "type": "text", + "content": "[21] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In The Eleventh International Conference on Learning Representations, 2022. 1, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 500 + ], + "type": "text", + "content": "[22] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 502, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 554, + 555 + ], + "type": "text", + "content": "[23] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 554, + 612 + ], + "type": "text", + "content": "[24] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 644 + ], + "type": "text", + "content": "[25] Boshi Tang, Jianan Wang, Zhiyong Wu, and Lei Zhang. Stable score distillation for high-quality 3d generation, 2024. 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[26] SIMA Team, Maria Abi Raad, Arun Ahuja, Catarina Barros, Frederic Besse, Andrew Bolt, Adrian Bolton, Bethanie Brownfield, Gavin Buttimore, Max Cant, Sarah Chakera, Stephanie C. Y. Chan, Jeff Clune, Adrian Collister, Vikki Copeman, Alex Cullum, Ishita Dasgupta, Dario de Cesare, Julia Di Trapani, Yani Donchev, Emma Dunleavy," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19575" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 324 + ], + "type": "text", + "content": "Martin Engelcke, Ryan Faulkner, Frankie Garcia, Charles Gbadamosi, Zhitao Gong, Lucy Gonzales, Kshitij Gupta, Karol Gregor, Arne Olav Hallingstad, Tim Harley, Sam Haves, Felix Hill, Ed Hirst, Drew A. Hudson, Jony Hudson, Steph Hughes-Fitt, Danilo J. Rezende, Mimi Jasarevic, Laura Kampis, Rosemary Ke, Thomas Keck, Junkyung Kim, Oscar Knagg, Kavya Kopparapu, Andrew Lampinen, Shane Legg, Alexander Lerchner, Marjorie Limont, Yulan Liu, Maria Loks-Thompson, Joseph Marino, Kathryn Martin Cussons, Loic Matthey, Siobhan Mcloughlin, Piermaria Mendolicchio, Hamza Merzic, Anna Mitenkova, Alexandre Moufarek, Valeria Oliveira, Yanko Oliveira, Hannah Openshaw, Renke Pan, Aneesh Pappu, Alex Platonov, Ollie Purkiss, David Reichert, John Reid, Pierre Harvey Richemond, Tyson Roberts, Giles Ruscoe, Jaume Sanchez Elias, Tasha Sandars, Daniel P. Sawyer, Tim Scholtes, Guy Simmons, Daniel Slater, Hubert Soyer, Heiko Strathmann, Peter Stys, Allison C. Tam, Denis Teptyashin, Tayfun Terzi, Davide Vercelli, Bojan Vujatovic, Marcus Wainwright, Jane X. Wang, Zhengdong Wang, Daan Wierstra, Duncan Williams, Nathaniel Wong, Sarah York, and Nick Young. Scaling instructable agents across many simulated worlds, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 327, + 295, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 359 + ], + "type": "text", + "content": "[27] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 415 + ], + "type": "text", + "content": "[28] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming mode collapse in score distillation for text-to-3d generation, 2024. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "text", + "content": "[29] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 1, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 496 + ], + "type": "text", + "content": "[30] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 552 + ], + "type": "text", + "content": "[31] Min Wei, Jingkai Zhou, Junyao Sun, and Xuesong Zhang. Adversarial score distillation: When score distillation meets gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8131-8141, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 597 + ], + "type": "text", + "content": "[32] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and browser-compatible environment from a single video, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 600, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 654 + ], + "type": "text", + "content": "[33] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to optimize denoising scores for 3d generation: A unified and improved diffusion prior on nerf and 3d gaussian splatting, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 656, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 294, + 689 + ], + "type": "text", + "content": "[34] Xin Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3d with classifier score distillation, 2023. 1, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Linqi Zhou, Andy Shih, Chenlin Meng, and Stefano Ermon. Dreampropeller: Supercharge text-to-3d generation" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 185 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "with parallel sampling. arXiv preprint arXiv:2311.17082, 2023.8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[36] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. arXiv preprint arXiv:2404.04057, 2024.8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[37] Junzhe Zhu, Peiye Zhuang, and Sanmi Koyejo. Hifa: High-fidelity text-to-3d generation with advanced diffusion guidance. arXiv preprint arXiv:2305.18766, 2023. 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19576" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_content_list.json b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7d59a1a425390a2b40f8a7cd957f647ec5b9a83e --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_content_list.json @@ -0,0 +1,1581 @@ +[ + { + "type": "text", + "text": "Advancing Textual Prompt Learning with Anchored Attributes", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheng Li $^{1}$ , Yibing Song $^{2}$ , Ming-Ming Cheng $^{1}$ , Xiang Li $^{1*}$ , Jian Yang $^{1*}$ $^{1}$ PCA Lab, VCIP, College of Computer Science, Nankai University, $^{2}$ DAMO Academy, Alibaba Group", + "bbox": [ + 225, + 179, + 776, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhengli97@mail.nankai.edu.cn, yibingsong.csv@gmail.com {cmm, xiang.li.implus, csjyang}@nankai.edu.cn", + "bbox": [ + 263, + 236, + 733, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Textual-based prompt learning methods primarily employ multiple learnable soft prompts and hard class tokens in a cascading manner as text inputs, aiming to align image and text (category) spaces for downstream tasks. However, current training is restricted to aligning images with predefined known categories and cannot be associated with unknown categories. In this work, we propose utilizing universal attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an Attribute-anchored Textual Prompt learning method for vision-language models, named ATPrompt. This approach expands the learning space of soft prompts from the original one-dimensional category level into the multi-dimensional attribute level by incorporating multiple attribute tokens into the learnable soft prompts. Through this modification, we transform the text prompt from a category-centric form to an attribute-category hybrid form. Additionally, we introduce a straightforward differentiable attribute search method to identify representative and suitable attributes for downstream tasks. As an easy-to-use plug-in technique, ATPrompt can seamlessly replace the existing basic prompt format in textual-based methods, providing general improvements at a negligible computational cost. Extensive experiments across 11 datasets validate the effectiveness of our method. Code is publicly available at https://github.com/zhengli97/ATPrompt.", + "bbox": [ + 88, + 337, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) [10, 20, 21, 28, 29, 32, 34, 35], such as CLIP [3, 32] and ALIGN [10], have demonstrated exceptional performance in recent years. These models are trained with a contrastive loss to establish alignment between image and text (category) space. Inspired by the success of NLP [18, 19], prompt learning [52] has", + "bbox": [ + 89, + 784, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/404ad711120a7070a0481b67a504644d00bcebcafb9099f47527b71126fd0ab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 306, + 851, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc204e01d93682cd3d39978bd1666f2a400475d7802824a1f0afd311d5c2603e.jpg", + "image_caption": [ + "Figure 1. Comparison of image and text (category) alignment processes through learnable prompts. (a) Current prompt learning methods align images with predefined categories but fail to establish accurate associations with unknown categories. (b) ATPrompt leverages universal attributes as an intermediary to create more accurate alignments between images and unknown categories." + ], + "image_footnote": [], + "bbox": [ + 563, + 460, + 849, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emerged as a parameter-efficient tool to adapt powerful VLMs to downstream tasks. Models with a few learnable soft prompt tokens [11, 47, 51] can achieve performance parity with, or even outperform, fully fine-tuned ones [11]. Depending on how the soft prompt tokens are applied, existing methods can be broadly classified into textual-based [13, 14, 23, 40, 51, 52] and visual-based approaches [1, 2, 11, 16, 46]. Among these, the textual-based method is the most fundamental and straightforward, comprising the majority.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In typical image classification tasks, current text-based methods [13, 14, 51, 52] predominantly employ the traditional approach of concatenating learnable soft prompts with hard class tokens to replace handcrafted text prompts (e.g., \"a photo of a {class}\") as inputs to the encoder. Although this text prompt demonstrates strong performance, it restricts image alignment during training to predefined known categories only, thereby preventing accurate associations with unknown categories, as shown in Fig. 1(a). Intuitively, when confronted with an unfamiliar category, humans often associate it with additional attributes (e.g., color, shape or texture) to increase comprehensibility and clarity, rather than merely stating the object's name. For instance, one might describe a cheetah as: \"The cheetah is a cat-like animal with a small head, short yellow hair, and black spots.\" or refer to an apple as \"That red spherical fruit with orange stripes is an apple.\" instead of using a general description such as \"This is a cheetah.\" or \"That fruit is an apple.\" Attributes can serve as bridges that connect unknown categories to our known knowledge.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these observations, we propose a novel approach that leverages attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an attribute-anchored textual prompt learning method for VLMs, named ATPrompt. This method extends the learning space of soft prompts from the original one-dimensional category level to a multidimensional attribute level by integrating multiple fixed universal attribute tokens into the learnable soft prompts. Guided by these anchored attributes, soft tokens acquire not only category-specific but also attribute-related general representations during training. This results in improved alignment between images and unknown categories compared to the original method, as shown in Fig. 1(b). Additionally, based on the depth at which soft prompts are applied, we propose two versions of ATPrompt: shallow and deep, respectively, ensuring compatibility with existing methods of varying depths [13, 14, 52]. To finalize these attributes, we present a simple and effective differentiable attribute search method that learns to identify suitable attributes from a candidate pool constructed by LLMs. The search operation only needs to be performed once per task, and once completed, the selected attributes can be used by ATPrompt for model training.", + "bbox": [ + 91, + 396, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As an easy-to-use plug-in technique, ATPrompt can seamlessly substitute existing forms used in textual-based prompt learning methods, yielding general improvements with negligible additional computational overhead.", + "bbox": [ + 89, + 760, + 480, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 823, + 434, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose an efficient attribute-anchored textual prompt learning method that expands the learning space of soft prompts from a one-dimensional class level to the multidimensional attribute level.", + "bbox": [ + 91, + 839, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce an effective differentiable search method to select appropriate attributes for downstream tasks.", + "- Both shallow and deep versions of ATPrompt are introduced to ensure compatibility with existing prompt learning methods of varying depths.", + "- Extensive experiments demonstrate that ATPrompt can be seamlessly integrated into existing textual-based methods, resulting in consistently improved performance with negligible computational costs." + ], + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 239, + 653, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Learning for VLMs. Inspired by recent advancements in NLP [18, 19], prompt learning [13, 14, 23, 33, 48, 51-53] has garnered significant interest among vision researchers aiming to apply these techniques to VLMs [10, 32, 42], such as CLIP. CoOp [52] is the pioneering text-based approach that introduced the concept of using a combination of soft textual tokens and a hard class token as input. Subsequent studies [13, 14, 17, 23, 41, 44, 51] have predominantly followed this textual prompt format. However, this form constrains the soft prompts to align with images within a one-dimensional, predefined category space, limiting their applicability to unknown categories. Therefore, training based on the current text form will be more likely to overfit to known categories, diminishing their zero-shot generalization capability for unknown categories. To address this limitation, several methods have been proposed [12, 14, 23, 44, 53]. For instance, Kg-CoOp [44] uses hand-crafted hard prompts to regularize learnable soft prompts during training. PromptSRC [14] utilizes CLIP's [32] original features to regularize the learning of soft prompts for image and text branches. PromptKD [23] utilizes a pre-trained strong teacher model to guide the learning of a student model [8, 22, 24, 43] with learnable prompts. Despite these advancements, none of the above methods address the inherent limitations of the format itself. In this work, we introduce the attribute-anchored textual prompt format for VLMs, which proposes to utilize attributes as a bridge to build more accurate associations between images and unknown categories.", + "bbox": [ + 511, + 266, + 903, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attributes for VLMs. In practice, categories typically encompass multiple attributes. When individuals encounter an unfamiliar category, they often describe it using additional attributes to enhance the clarity of their communication, rather than merely stating its name. Inspired by this observation, numerous studies [4, 30, 37, 38, 45] have begun to leverage attributes to support their objectives. VCD [30] was the pioneering work to propose the use of LLMs to decompose class names into multiple in-class attributes (e.g., beak and tail for birds) for classification. AAPL [15] introduces a meta-network to extract visual attribute features based on encoded image features, facilitating image-text alignment. TAP [5] presents a structured \"Tree of At", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1d40b41b880b211d87c229cda2c4c252092476e1cbfc4054c41aceab0af5c0d.jpg", + "image_caption": [ + "Figure 2. Architectural comparison among existing methods. (a) Vanilla CLIP employs a hand-crafted text template as input to the text encoder. (b) Classical prompt learning proposes a new text form that concatenates multiple learnable soft tokens with class tokens. (c) ArGue [37] employs multiple in-class attributes mined by LLMs as supplementary information. These attributes are utilized to construct distinct text groups, which serve as learning targets, thereby regularizing the learning of soft tokens. The final prediction is achieved by assembling all groups. (d) Our ATPrompt treats universal attributes as learning components and anchors them into existing soft prompt templates. Though this operation, we expand the learning space of soft tokens to multi-dimensional attribute levels and facilitate the alignment of images with unknown class texts." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 264, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/597a08010caa7fab85fbd011e3d4b0549890acd44cd4ecfe046c3dea2f263c3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 92, + 452, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f1c2500e0697a38ef1eca8915369c9e894a2ce207cc27b37fd466af3bf5db66a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 90, + 661, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/540938326c885781b22f84838f8676c2e0ff0eacc8034305263230f84f9a2dbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 90, + 893, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "attributes\" approach that leverages attribute-specific knowledge graphs to enhance VLMs. ArGue [37] leverages large language models to mine multiple in-class attributes and integrates them into soft prompts and fixed templates to create multiple text groups. It employs the original text features generated by the fixed template to regularize the learning of soft tokens, as illustrated in Fig. 2(c). Most prior studies have focused on utilizing in-class attributes to enhance model performance by providing supplementary attribute information. However, when dealing with an unknown class, reacquiring the attributes of the new class becomes necessary—a process that is both complex and costly.", + "bbox": [ + 88, + 333, + 480, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we believe that universal (inter-class) attributes are more efficient and robust than the in-class attributes used in previous works. Instead of taking attributes as the learning objective, we treat it as a learning component and propose to anchor universal attributes into soft prompt templates, transforming the existing class-centric form [52] into a hybrid attribute-class learnable textual prompt form. Our approach can be seamlessly integrated into existing textual-based methods, enhancing their performance without incurring additional computational costs.", + "bbox": [ + 88, + 515, + 482, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 679, + 181, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt learning [13, 14, 51-53] aims to enhance the generalization ability of pre-trained VLMs like CLIP on downstream tasks by training inserted learnable soft tokens. Existing textual-based methods all follow the classic prompt paradigm, concatenating soft prompt tokens and hard class tokens as the input to the text encoder, as shown in Fig. 2(b). In this paper, we propose a simple and effective textual prompt learning method, named ATPrompt, which anchors multiple fixed universal attribute tokens into the original soft prompts, as shown in Fig. 2(d). Guided by these attributes, soft prompts can learn not only category-specific but also attribute-related general representations through training. When encountering unknown categories, these", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "learned attribute-related tokens can provide additional information to promote better image-text alignment. Furthermore, to identify these universal attributes, we present an automated pipeline that encompasses sequential steps. Initially, we employ LLMs to synthesize the attribute pool for the current downstream category. Subsequently, we propose a differentiable attribute search method designed to identify attributes within the pool that are most suitable for our attribute-anchored prompt forms. For every task, this search operation is conducted only once. Once the attributes are finalized, they are integrated into our ATPrompt for specialized model fine-tuning.", + "bbox": [ + 511, + 333, + 906, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 527, + 643, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Models. Existing VLMs, such as CLIP [32], have demonstrated remarkable zero-shot generalization performance after training with 400 million image-text pairs. The primary objective of these models is to learn the alignment between image and text modalities produced by each encoder [6, 27, 49]. Given a labeled image classification dataset $\\mathrm{D} = \\{(x,c)\\}$ which includes $N$ class labels $C = \\{c_i\\}_{i = 1}^N$ , CLIP makes predictions by calculating the cosine similarity between image features and the text features of each class. Specifically, for each input image $x$ , it undergoes feature extraction via the image encoder $h_I(x)$ and obtains a feature vector $u = h_I(x)$ . Simultaneously, for each class, a series of textual descriptions $t$ are generated using the hand-crafted template. Then, these text descriptions are fed into the text encoder $h_T(x)$ to obtain text features $w = h_{T}(t)$ . Finally, the output probability for image $x$ classified to $c$ is calculated as follows:", + "bbox": [ + 511, + 550, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (c | x) = \\frac {\\exp \\left(\\cos \\left(u , w _ {c}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\cos \\left(u , w _ {i}\\right) / \\tau\\right)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 821, + 903, + 858 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter and $\\cos (\\cdot ,\\cdot)$ denotes cosine similarity.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f77496042a509cb5255fb771dda64d1fcce2ecffd22ac7cf5c3499613787eff.jpg", + "image_caption": [ + "(a) Shallow Version" + ], + "image_footnote": [], + "bbox": [ + 153, + 89, + 444, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46cb0444479a79004dbd45c76924ac713536483e63109a0b0f71e2f0fbf63e3f.jpg", + "image_caption": [ + "(b) Deep Version", + "Figure 3. An illustration of the computation process for shallow and deep versions. Take two attributes [A] and [B] as examples. (a) The shallow version concatenates hard attribute tokens, soft prompt tokens, and class tokens and inputs them into the encoder for calculation. (b) The deep version uses the same input but discards the class-related soft prompt tokens after calculating the self-attention and introduces them again before the next layer. These two forms can be compatible with existing methods of varying prompt depths, including input-level ones like CoOp [52], CoCoOp [51] and depth-level ones like MaPLe [13]." + ], + "image_footnote": [], + "bbox": [ + 477, + 93, + 705, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/24b9b8627f22e44be57067d982e180ebfec30d25f3bdc9419e909b7f4ccf3eff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 125, + 843, + 231 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Learning for VLMs. Instead of manually designed hard prompts for image-text alignment, which is inaccurate and inflexible, recent prompt learning works [13, 14, 44, 51, 52] like CoOp propose to learn appropriate soft textual prompts for downstream tasks. Concretely, $M$ learnable soft tokens $[T_i]_{i=1}^M$ are concatenated with the hard class token [CLS] as the input of the text encoder, as shown in Fig. 2(b). Its form is shown as follows:", + "bbox": [ + 89, + 340, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {T} = \\left[ T _ {1} \\right] \\left[ T _ {2} \\right] \\dots \\left[ T _ {M} \\right] [ \\text {C L S} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 470, + 480, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ represents the length of soft tokens. For simplicity, we omit the prefix and suffix tokens in the input.", + "bbox": [ + 89, + 494, + 482, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to embedding soft tokens at the input level, existing studies [11, 13, 14, 23] have also explored introducing them at deeper layers. This is achieved by adding soft tokens within the Transformer blocks and subsequently removing them after the self-attention computations. For the $i$ -th block, this process can be described as follows:", + "bbox": [ + 89, + 526, + 483, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 625, + 482, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{i}$ represents the $i$ -th transformer block and $\\mathrm{T}_i$ denotes the set of learnable soft tokens, defined as $\\mathrm{T}_i = \\{[T_1]_i,\\dots,[T_M]_i\\}$ .", + "bbox": [ + 89, + 648, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning soft prompts with universal attributes", + "text_level": 1, + "bbox": [ + 89, + 703, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach introduces two variants, distinguished by the number of layers at which soft tokens are applied: a shallow version and a deep version, as shown in Fig. 3.", + "bbox": [ + 89, + 724, + 482, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shallow Version. We begin by introducing the shallow version, in which hard attribute tokens are anchored solely at the input level, as illustrated in Fig. 3(a). Consider two universal attributes, A and B. According to Eqn. (2), the shallow-level text prompt $P_T$ provided to the text encoder can be expressed as follows:", + "bbox": [ + 89, + 771, + 483, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {T} = \\left[ T _ {a _ {1}} \\right] \\dots \\left[ T _ {a _ {m}} \\right] \\left[ A \\right] \\left[ T _ {b _ {1}} \\right] \\dots \\left[ T _ {b _ {m}} \\right] \\left[ B \\right] \\left[ T _ {1} \\right] \\dots \\left[ T _ {M} \\right] \\left[ C L S \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{m}$ and $b_{m}$ are hyperparameters specifying the length of soft tokens for attributes A and B. In our method, we set these parameters to be the same by default.", + "bbox": [ + 511, + 340, + 906, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While this example places the class token at the end, we also evaluate configurations where it is positioned at the front or in the middle. As shown in Tab. 5, the end position yields the best performance and is therefore adopted as our default configuration.", + "bbox": [ + 511, + 387, + 905, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep Version. In this version, learnable soft tokens are introduced at the input of the deep layers. Previous works, such as VPT and MaPLe, discard all soft tokens and subsequently reintroduce them after the block. When this operation is applied to attribute-related words, a gap will emerge between the introduced excessive low-level tokens and the existing high-level tokens, thereby weakening the feature continuity across layers. In this study, our approach selectively discards and then re-adds only class-related soft tokens in the input, specifically $[T_1],\\dots,[T_M]$ , as shown in Fig. 3(b). Based on Eqn. (3), the deep version of ATPrompt can be rewritten as follows:", + "bbox": [ + 511, + 463, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathrm {F} _ {1}, \\rightarrow , \\mathrm {C L S} _ {1} ] = L _ {1} ([ \\mathrm {T} _ {a _ {0}}, \\mathrm {A}, \\mathrm {T} _ {b _ {0}}, \\mathrm {B}, \\mathrm {T} _ {0}, \\mathrm {C L S} _ {0} ]), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 648, + 903, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathrm {F} _ {i, -}, \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {F} _ {i - 1}, \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 670, + 903, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ni = 2, 3, \\dots , M.\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 691, + 746, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{F}_i$ represents the features computed by the $i$ -th Transformer layer. We demonstrate the effectiveness of this operation in Tab. 6.", + "bbox": [ + 511, + 709, + 905, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. Let $\\theta$ represent the weight of the total soft tokens and let $v$ denote the selected fixed attribute tokens. Training is conducted on a labeled dataset $\\mathrm{D} = \\{(x,c)\\}$ , with the objective of minimizing the cross-entropy loss between predicted values and ground truth labels. This process can be formulated as follows:", + "bbox": [ + 511, + 753, + 906, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} L _ {t r a i n} = \\min _ {\\theta} \\sum_ {x \\in D} \\operatorname {C E} (f (x; v, \\theta), c). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 848, + 903, + 880 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f(\\cdot)$ represents the function of the CLIP model.", + "bbox": [ + 511, + 885, + 870, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87de5756e56376e383ee44596442002f12bec98cf0bd6d9b0208e909ac934773.jpg", + "image_caption": [ + "Figure 4. An overview of our attribute search pipeline. (a) We first query the LLM iteratively to obtain multiple independent attributes. These are subsequently aggregated to form a pool of candidate combinations, which serves as the input for the search process. (b) The forward computation for each candidate combination is represented by a distinct colored path. To identify the optimal attributes, we employ an alternating optimization algorithm that co-optimizes the soft tokens and a corresponding path weight vector. Upon completion of training, the combination associated with the highest-weighted path is selected as the final output." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Attribute search", + "text_level": 1, + "bbox": [ + 89, + 436, + 253, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selecting the attributes involves two key considerations: their content and their quantity. While directly querying a LLM is a straightforward approach, it has notable drawbacks. This method cannot determine the optimal number of attributes for a specific downstream dataset, and querying by category name alone can introduce semantic bias. To address these issues, we propose an automated pipeline that selects the appropriate content and quantity of attributes for the current downstream task, as illustrated in Fig. 4.", + "bbox": [ + 88, + 460, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute Pool. Inspired by CoT [39, 50], we divide the entire process into multiple steps to enhance the reasoning ability of LLMs. First, we prompt the LLM to generate descriptive sentences for each known category, thereby enriching category-related information. Using these descriptions as context, we then prompt the LLM to summarize a set of independent attribute bases that are common across these categories. An attribute pool is then formed by creating all possible combinations of these bases, as shown in Fig. 4(a). For $N$ attribute bases, this result in a total of $w = C_N^1 + C_N^2 + \\ldots + C_N^N$ candidates in the pool, which constitutes our search space. Note that we do not consider the order of attributes, as permutations generally do not introduce significant semantic bias or affect the final performance, a claim we validate in our experiments.", + "bbox": [ + 89, + 597, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute Searching. Inspired by DARTS [26], we introduce a differentiable attribute search method that learns to find representative attributes $v$ from the search space $\\mathcal{V}$ , as shown in Fig. 4(b). To make the search space continuous, we relax the discrete attribute selection into a softmax-", + "bbox": [ + 88, + 824, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weighted sum over all $w$ possible candidates:", + "bbox": [ + 511, + 438, + 813, + 453 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x, v; \\alpha , \\theta) = \\sum_ {i \\in \\mathcal {V}} \\frac {\\exp (\\alpha_ {i})}{\\sum_ {i ^ {\\prime} \\in \\mathcal {V}} \\exp (\\alpha_ {i ^ {\\prime}})} f (x, v _ {i}; \\theta). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 464, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ represents the weight for attribute combination $v_{i}$ . The task of attribute search is thus reduced to learning the weight vector $\\alpha$ for the candidate pool.", + "bbox": [ + 511, + 512, + 903, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After relaxation, our goal is to jointly learn the attribute weight $\\alpha$ and soft prompt tokens $\\theta$ . Following standard practice [26, 31, 54], we optimize the weights $\\alpha$ by minimizing the validation loss $L_{val}$ , while the soft tokens are learned by minimizing the training loss $L_{train}$ . We employ an alternating algorithm [9, 22] to solve this bi-level optimization problem, alternating between these two subproblems:", + "bbox": [ + 511, + 556, + 905, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\alpha} = \\arg \\min _ {\\alpha} L _ {v a l} (f (x, v; \\alpha , \\hat {\\theta}), c), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 691, + 906, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\theta} = \\arg \\min _ {\\theta} L _ {t r a i n} (f (x, v; \\hat {\\alpha}, \\theta), c). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 720, + 903, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{train}$ and $L_{val}$ both use the cross-entropy loss function. After the search, the attribute combination with the highest weight $(\\alpha_{i})$ is selected.", + "bbox": [ + 511, + 750, + 903, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cost Analysis. Unlike traditional Neural Architecture Search (NAS) methods [7, 25, 36], which search computationally expensive network-level parameters, our approach focuses on a lightweight token-level search space. This design makes our method significantly more efficient than previous approaches. In practice, our search converges in approximately 5 epochs, often requiring less than 5 minutes", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2e0d9f3ca23c08cad645adb7b4ba05e64709c4977a6e0994169a44181a00e5c6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAverageImageNetCaltech101OxfordPots
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (UCV 22)82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp (CVPR 22)80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
MaPLc (CVPR 23)82.2875.1478.5576.6670.5473.4797.7494.3696.0295.4397.7696.58
PromptSRC (ICCV 23)84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
ArGue (CVPR 24)83.6978.0780.7876.9272.0674.4198.4395.2096.7995.3697.9596.64
DePT (CVPR 24)83.6671.8277.2977.1370.1073.4598.3394.3396.2994.7097.6396.14
CoPrompt (ICLR 24)84.0077.2380.4877.6771.2774.3398.2794.9096.5595.6798.1096.87
PromptKD (CVPR 24)86.9680.7383.7380.8374.6677.6298.9196.6597.7796.3098.0197.15
CoOp + ATPrompt82.6868.0474.65 (+2.99)76.2770.6073.3397.9593.6395.7494.7796.5995.67
CoCoOp + ATPrompt81.6974.5477.95 (+2.12)76.4370.5073.3597.9695.2796.6095.4697.8996.66
MaPLc + ATPrompt82.9875.7679.21 (+0.66)76.9470.7273.7098.3295.0996.6895.6297.6396.61
DePT + ATPrompt83.8073.7578.45 (+1.16)77.3270.6573.8398.4894.6096.5094.6597.9996.29
PromptKD + ATPrompt87.0581.8284.35 (+0.62)80.9074.8377.7598.9096.5297.7096.9298.2797.59
", + "bbox": [ + 155, + 88, + 841, + 282 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/225817b9a93d90998862d2e1c7069390a3523530d42fe02503432fcc3c476a6f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (JCV 22)78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp (CVPR 22)70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
MaPLe (CVPR 23)72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
PromptSRC (ICCV 23)78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
ArGue (CVPR 24)75.6473.3874.4998.3475.4185.3692.3391.9692.1440.4638.0339.21
DePT (CVPR 24)79.6772.4075.8698.2072.0083.0890.4391.3390.8842.5322.5329.46
CoPrompt (ICLR 24)76.9774.4075.6697.2776.6085.7190.7392.0791.4040.2039.3339.76
PromptKD (CVPR 24)82.8083.3783.1399.4282.6290.2492.4393.6893.0549.1241.8145.17
CoOp + ATPrompt77.4366.5571.5897.4467.5279.7788.7487.4488.0940.3827.2232.52
CoCoOp + ATPrompt74.5073.4773.9896.5273.5983.5190.5991.7491.1637.3033.1535.10
MaPLe + ATPrompt75.3973.8474.6197.8275.0784.9590.6592.0091.3237.6136.1536.87
DePT + ATPrompt79.2973.4776.2798.2073.6984.2090.4291.6991.0543.1933.2337.56
PromptKD + ATPrompt82.5184.0383.2699.1582.0389.7892.4893.8693.2249.6342.3545.70
", + "bbox": [ + 166, + 287, + 828, + 481 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35b6822f6e6a4221919295fc645e77f9894154ef62ffaad7d49cfaaa77d5273d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSUN397DTDEuroSATUCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (IUCV 22)80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp (CVPR 22)79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
MaPLe (CVPR 23)80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
PromptSRC (ICCV 23)82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
ArGue (CVPR 24)81.5280.7481.1381.6066.5573.3194.4388.2491.2385.5679.2982.31
DePT (CVPR 24)82.3775.0778.5583.2056.1367.0488.2766.2775.7085.4372.1778.24
CoPrompt (ICLR 24)82.6380.0381.3083.1364.7372.7994.6078.5785.8486.9079.5783.07
PromptKD (CVPR 24)83.6981.5482.6085.8471.3777.9497.5482.0889.1489.7182.2786.10
CoOp + ATPrompt80.8468.6474.2480.8345.4958.2290.3459.7971.9684.4964.9673.45
CoCoOp + ATPrompt80.5076.8678.6478.6356.8966.0287.9574.1580.4682.7476.4079.44
MaPLe + ATPrompt80.9878.1579.5480.5058.2867.6194.8477.5985.3584.0878.8881.40
DePT + ATPrompt82.4276.4879.3482.6456.7767.3089.6069.5078.2885.6073.1578.89
PromptKD + ATPrompt83.8781.3582.5986.9272.3478.9697.0592.0794.4989.2982.4485.73
", + "bbox": [ + 168, + 486, + 828, + 681 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69bc9ee7fe9d6ad3bc44c32d673523ed019ee6508426f6d794e9624e9b88a337.jpg", + "table_caption": [ + "Table 1. Base-to-novel generalization experiments of five baselines with and without our ATPrompt on 11 datasets. Our method achieves consistent average performance improvement over different baselines." + ], + "table_footnote": [], + "table_body": "
MethodSourceTarget DatasetAverage
Image NetCaltech 101Oxford PetsStanford CarsFlowers 102Food101FGVC AircraftSUN397DTDEuro SATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
+ATPrompt71.6793.9690.6565.0170.4085.8620.9765.7743.4446.5969.9265.26 (+1.38)
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
+ATPrompt71.2793.7990.6265.9071.1786.0323.2266.6344.4448.7070.7166.59 (+0.85)
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
+ATPrompt70.6994.0491.0366.0671.9986.3324.4267.0545.2148.6369.1566.75 (+0.45)
", + "bbox": [ + 145, + 722, + 848, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Cross-dataset generalization experiments of three baselines with and without our ATPrompt on 11 datasets. ATPrompt achieves consistent average performance improvement on target datasets.", + "bbox": [ + 89, + 852, + 906, + 882 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3623", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c659414e05650acb23dae8983b6f6b35400b87560abc82ed5c01704f0817a311.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAttribute BasesSearched Results
ImageNetcolor, size, shape, habitat, behavior(color, shape)
Caltech101shape, color, material, function, size(shape,size)
OxfordPetsloyalty, affection, energy, playfulness, intelligence(playfulness, energy)
StanfordCarsdesign, engine, performance, luxury, color(luxury)
Flowers102color, flower, habitat, growth, season(color, habitat, growth)
Food101flavor, texture, origin, ingredients, preparation(flavor, preparation)
", + "bbox": [ + 96, + 88, + 478, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on a single A800 GPU for some datasets. In addition, search efficiency can be improved by curating a smaller set of base attributes, which narrows the search space.", + "bbox": [ + 89, + 338, + 482, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 398, + 223, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 422, + 191, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Base-to-Novel Generalization. Following [14, 23, 51, 52], we split the dataset into base and novel classes. The model is trained on the base class training set and evaluated on the test set. The attributes applied in this experiment are searched based on the base class. For datasets without a dedicated validation set, like ImageNet, we split the 16-shot labeled data, using half for training and the other half for attribute search.", + "bbox": [ + 89, + 444, + 482, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-dataset Experiments. Consistent with previous works [14, 51, 52], we first train a model on the ImageNet-1K source dataset and then evaluate its generalization performance on several out-of-distribution datasets. The attributes used are obtained from the source dataset.", + "bbox": [ + 89, + 566, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attribute Search. We select five independent attributes as the basis in the attribute pool. This results in 31 candidate attribute combinations for the search process. We use ChatGPT-4o for attribute queries. In Tab. 3 reports a subset of the queried attribute bases and the final combinations selected by our search.", + "bbox": [ + 89, + 642, + 482, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We evaluate the model performance on 15 datasets. We report base and novel class accuracy and their Harmonic Mean (HM) averaged over 3 runs. The details of each dataset are attached in the Appendix.", + "bbox": [ + 89, + 732, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Base-to-Novel Generalization", + "text_level": 1, + "bbox": [ + 89, + 803, + 352, + 818 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As demonstrated in Tab. 1, we evaluate the base-to-novel generalization performance of five baseline methods, both with and without integrating ATPrompt, across 11 diverse recognition datasets. Notably, ATPrompt consistently improves the average performance of all baseline methods.", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reasons for Limited Improvement in Some Conditions. (1) Learnable text prompts are the central component in earlier works, and optimizing them through ATPrompt can obviously improve the performance. (2) Recent studies have expanded beyond learnable text prompts by introducing additional learnable modules. Since our work only involves the optimization of the learnable text prompt part, the improvement have become less obvious.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Cross-dataset Evaluation", + "text_level": 1, + "bbox": [ + 511, + 224, + 743, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows the cross-dataset generalization results for three baseline methods. Our method demonstrates superior performance, with improvements of $1.38\\%$ , $0.85\\%$ and $0.45\\%$ for CoOp, CoCoOp and MaPLe, respectively.", + "bbox": [ + 511, + 247, + 906, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 320, + 728, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 shows the domain generalization results for three baseline methods. The results show that our method improves CoOp, CoCoOp and MaPLe methods by $0.90\\%$ , $0.49\\%$ and $0.33\\%$ respectively.", + "bbox": [ + 511, + 343, + 905, + 404 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/67d825f9ea92e00cb5d83fd843d495a5e2aa29b9d01de307be9b92e3b11eae2c.jpg", + "table_caption": [ + "Table 3. Part of the results obtained after differentiable attribute search. Please refer to the appendix for the complete results." + ], + "table_footnote": [], + "table_body": "
MethodSourceTarget DatasetAverage
ImageNet-V2-S-A-R
CoOp71.5164.2047.9949.7175.2159.28
+ATPrompt71.6764.4349.1350.9176.2460.18 (+0.90)
CoCoOp71.0264.0748.7550.6376.1859.91
+ATPrompt71.2764.6649.1551.4476.3360.40 (+0.49)
MaPLe70.7264.0749.1550.9076.9860.27
+ATPrompt70.6964.4049.1051.7777.1160.60 (+0.33)
", + "bbox": [ + 521, + 417, + 898, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Domain generalization experiments of three baselines with and without our ATPrompt on four datasets. The integration of ATPrompt resulted in better generalization performance.", + "bbox": [ + 511, + 546, + 906, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 611, + 681, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By default, the experiments are conducted on the ImageNet. To minimize the influence of other components, we mainly adopt CoOp as the baseline method. Two attributes (color and shape) are used in our ATPrompt.", + "bbox": [ + 511, + 633, + 905, + 696 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d195a143a3ea8c18ad49ef0f1be6805308cb3ced856a3581842236793d708d6.jpg", + "image_caption": [ + "Figure 5. Illustration of varying soft token length on ImageNet. Increased tokens can lead to overfitting to the base class and weaken generalization to the novel class." + ], + "image_footnote": [], + "bbox": [ + 589, + 712, + 826, + 857 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft Prompt Length. In Fig. 5, we examine the optimal soft token length for both attribute and class tokens. By varying the length from 1 to 10, we observe that longer prompts dilute the guiding influence of the attribute tokens, thereby reducing generalization to novel classes.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class Token Position. The relative positioning of attribute tokens and class tokens in our method requires careful consideration. In Tab. 5, we examine configurations where the class token is positioned in the middle or on either side of two attribute tokens. The results demonstrate that optimal performance is achieved when the class token is placed at the end, which aligns with the findings of CoOp.", + "bbox": [ + 89, + 167, + 480, + 272 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ff2e23003089086000920d60a7859c2184728f6d84ee44bdb7e93fdf63c3a33f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PositionBaseNovelHM
Front76.1270.5073.20
Middle76.1370.2973.09
End76.2770.6073.33
", + "bbox": [ + 179, + 282, + 393, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Operation of Deep Version. In ATPrompt-Deep, we exclusively drop class soft tokens while retaining both hard and soft attribute tokens after they pass through the block. In this part, we compare the performance of partial drop (i.e., removing attribute soft tokens while retaining hard tokens) and full drop (i.e., removing both attribute soft and hard tokens) operations, as illustrated in Tab. 6.", + "bbox": [ + 89, + 398, + 480, + 503 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6d9d88adef2551af2a4c77466166cd871c1684e672c5c1440df6fd7dae5669f3.jpg", + "table_caption": [ + "Table 5. Comparison of different class token positions on ImageNet. The end position works best." + ], + "table_footnote": [], + "table_body": "
Operation of Attribute TokenBaseNovelHM
Retain all hard and soft tokens76.9470.7273.70
Partial drop and re-add76.8770.4473.51
Full drop and re-add76.8370.1073.31
", + "bbox": [ + 124, + 516, + 450, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results show that maintaining the attributes of both hard and soft tokens during the forward process results in optimal performance. Conversely, dropping and reintroducing either hard or soft attribute tokens harms performance, likely because it disrupts the continuity of attribute representations across layers and complicates optimization.", + "bbox": [ + 89, + 643, + 480, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Attribute Order. In this study, we do not specifically focus on the order of attributes because varying the sequence usually does not result in semantic deviations in reality. Tab. 7 quantitatively assesses the impact of attribute order on prompt learning performance. From this table, we observe that despite variations in order, similar results are consistently produced, and the performance fluctuations across different orders remain within a reasonable range.", + "bbox": [ + 89, + 734, + 480, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison to Other Attributes. In Tab. 8, we explore the effectiveness of attributes derived through alternative methods, specifically by manually selecting class-irrelevant", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f24709d6501914fc69ef83e9642fd0d4be9a6d69c932a562f3df22357f49224.jpg", + "table_caption": [ + "Table 6. Comparison of operations on deep soft and hard attribute tokens based on MaPLe+ATPrompt. Preserving hard and soft attribute tokens in deep layers performs better than other operations." + ], + "table_footnote": [], + "table_body": "
AttributesBaseNovelHM
(shape, color)76.3270.3973.24
(color, shape)76.2770.6073.33
(size, habitat)76.4470.2373.20
(habitat, size)76.4670.1673.14
", + "bbox": [ + 596, + 88, + 823, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and common attributes. It shows that manually selected irrelevant attributes exhibit comparable performance during training; however, they perform poorly when applied to new categories. This suggests that incorrect attribute tokens cause the soft tokens to develop biased representations, thereby diminishing their zero-shot generalization ability.", + "bbox": [ + 511, + 238, + 903, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/95ea42de93842aa5533cd72c5c5f78ba3b8844031cbeb076e73a69565eaf9fe1.jpg", + "table_caption": [ + "Table 7. Comparison of different orders on ImageNet. The order of attributes does not significantly affect the model, and performance fluctuations are within a reasonable range." + ], + "table_footnote": [], + "table_body": "
TypeAttributesBaseNovelHM
Common(shape, size)82.8367.1374.16
(color, texture)82.7367.5674.38
Irrelevant( plane, engines)82.8166.2273.59
(football, sport)82.7767.1474.14
Searched-82.6868.0474.65
", + "bbox": [ + 555, + 340, + 864, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Comparison of average performance for various attribute configurations on 11 datasets. The attributes obtained by our method achieve the best performance.", + "bbox": [ + 511, + 448, + 906, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 512, + 632, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce ATPrompt, an attribute-anchored textual prompt learning method that uses universal attributes as a bridge to improve generalization from seen to unseen categories. Our approach expands the learning space of soft prompts from a one-dimensional, category-centric structure to a multi-dimensional attribute space by anchoring fixed attribute tokens with the prompt. To ensure the selection of optimal attributes, we propose an automated pipeline designed to identify the most suitable candidates for any given downstream task. ATPrompt is designed with both shallow and deep architectural variants, rendering it broadly compatible with existing prompt learning methods. Extensive experiments validate the effectiveness of our approach. We believe this work offers a new direction for research into the fundamental structure of learnable prompts in prompt learning area.", + "bbox": [ + 511, + 537, + 903, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works. This work is a preliminary study of the basic prompt form, which has not been able to achieve comparable performance to regularization-based methods when working alone. Furthermore, the current selection of attribute anchors relies on manual experimentation, and automatically discovering the optimal anchor positions through a learning-based approach remains a promising direction for future research.", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62361166670 and U24A20330. This research was also supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206134), the Fundamental Research Funds for the Central Universities 070-63253222 and the Tianjin Key Laboratory of Visual Computing and Intelligent Perception (VCIP). Computation is supported by the Supercomputing Center of Nankai University (NKSC).", + "bbox": [ + 89, + 114, + 483, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 282, + 187, + 299 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 1", + "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 35:25005-25017, 2022. 1", + "[3] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1", + "[4] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, pages 23518-23527, 2023. 2", + "[5] Tong Ding, Wanhua Li, Zhongqi Miao, and Hanspeter Pfister. Tree of attributes prompt learning for vision-language models. arXiv preprint arXiv:2410.11201, 2024. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[7] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. JMLR, 20(55):1-21, 2019. 5", + "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[9] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. NeurIPS, 33:5632-5643, 2020. 5", + "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2", + "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 1, 4" + ], + "bbox": [ + 93, + 309, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 2", + "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 4", + "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 7", + "[15] Gahyeon Kim, Sohee Kim, and Seokju Lee. Aapl: Adding attributes to prompt learning for vision-language models. In CVPR Workshop, pages 1572-1582, 2024. 2", + "[16] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 1", + "[17] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 2", + "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2", + "[19] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 1, 2", + "[20] Yunheng Li, Yuxuan Li, Quansheng Zeng, Wenhai Wang, Qibin Hou, and Ming-Ming Cheng. Unbiased region-language alignment for open-vocabulary dense prediction. arXiv preprint arXiv:2412.06244, 2024. 1", + "[21] Yunheng Li, Zhong-Yu Li, Quan-Sheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-CLIP: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243-28258. PMLR, 2024. 1", + "[22] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 5", + "[23] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. In CVPR, pages 26617-26626, 2024. 1, 2, 4, 7", + "[24] Zheng Li, Xiang Li, Lingfeng Yang, Renjie Song, Jian Yang, and Zhigeng Pan. Dual teachers for self-knowledge distillation. Pattern Recognition, 151:110422, 2024. 2", + "[25] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 5", + "[26] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 5", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3626", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3", + "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 1", + "[29] Hao Ma, Ming Li, Jingyuan Yang, Or Patashnik, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Clip-flow: Decoding images encoded in clip space. CVMJ, 10(6):1157-1168, 2024. 1", + "[30] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2", + "[31] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 5", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3", + "[33] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 2", + "[34] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, pages 13019-13029, 2024. 1", + "[35] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1", + "[36] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 5", + "[37] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3", + "[38] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757, 2024. 2", + "[39] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5", + "[40] Ge Wu, Xin Zhang, Zheng Li, Zhaowei Chen, Jiajun Liang, Jian Yang, and Xiang Li. Cascade prompt learning for vision-language model adaptation. In ECCV, 2024. 1", + "[41] Shihan Wu, Ji Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves. In CVPR, pages 14723-14732, 2025. 2", + "[42] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, and Yongjun Xu. Clip-kd: An empirical study of distilling clip models. arXiv preprint arXiv:2307.12732, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 2", + "[44] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 2, 4", + "[45] Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, and Da Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In AAAI, pages 6979–6987, 2024. 2", + "[46] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In CVPR, pages 12924-12933, 2024. 1", + "[47] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, et al. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. 1", + "[48] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 2", + "[49] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025.3", + "[50] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 5", + "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 4, 7", + "[52] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 7", + "[53] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, pages 15659-15669, 2023. 2, 3", + "[54] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 5" + ], + "bbox": [ + 516, + 90, + 905, + 756 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_model.json b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ad62145aca5500665b32dac81faf47668a56757b --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.181, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "Advancing Textual Prompt Learning with Anchored Attributes" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.777, + 0.235 + ], + "angle": 0, + "content": "Zheng Li\\(^{1}\\), Yibing Song\\(^{2}\\), Ming-Ming Cheng\\(^{1}\\), Xiang Li\\(^{1*}\\), Jian Yang\\(^{1*}\\) \n\\(^{1}\\) PCA Lab, VCIP, College of Computer Science, Nankai University, \\(^{2}\\) DAMO Academy, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.237, + 0.735, + 0.27 + ], + "angle": 0, + "content": "zhengli97@mail.nankai.edu.cn, yibingsong.csv@gmail.com {cmm, xiang.li.implus, csjyang}@nankai.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Textual-based prompt learning methods primarily employ multiple learnable soft prompts and hard class tokens in a cascading manner as text inputs, aiming to align image and text (category) spaces for downstream tasks. However, current training is restricted to aligning images with predefined known categories and cannot be associated with unknown categories. In this work, we propose utilizing universal attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an Attribute-anchored Textual Prompt learning method for vision-language models, named ATPrompt. This approach expands the learning space of soft prompts from the original one-dimensional category level into the multi-dimensional attribute level by incorporating multiple attribute tokens into the learnable soft prompts. Through this modification, we transform the text prompt from a category-centric form to an attribute-category hybrid form. Additionally, we introduce a straightforward differentiable attribute search method to identify representative and suitable attributes for downstream tasks. As an easy-to-use plug-in technique, ATPrompt can seamlessly replace the existing basic prompt format in textual-based methods, providing general improvements at a negligible computational cost. Extensive experiments across 11 datasets validate the effectiveness of our method. Code is publicly available at https://github.com/zhengli97/ATPrompt." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.76, + 0.222, + 0.775 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) [10, 20, 21, 28, 29, 32, 34, 35], such as CLIP [3, 32] and ALIGN [10], have demonstrated exceptional performance in recent years. These models are trained with a contrastive loss to establish alignment between image and text (category) space. Inspired by the success of NLP [18, 19], prompt learning [52] has" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.308, + 0.852, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.461, + 0.851, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.652, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Figure 1. Comparison of image and text (category) alignment processes through learnable prompts. (a) Current prompt learning methods align images with predefined categories but fail to establish accurate associations with unknown categories. (b) ATPrompt leverages universal attributes as an intermediary to create more accurate alignments between images and unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "emerged as a parameter-efficient tool to adapt powerful VLMs to downstream tasks. Models with a few learnable soft prompt tokens [11, 47, 51] can achieve performance parity with, or even outperform, fully fine-tuned ones [11]. Depending on how the soft prompt tokens are applied, existing methods can be broadly classified into textual-based [13, 14, 23, 40, 51, 52] and visual-based approaches [1, 2, 11, 16, 46]. Among these, the textual-based method is the most fundamental and straightforward, comprising the majority." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3618" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "In typical image classification tasks, current text-based methods [13, 14, 51, 52] predominantly employ the traditional approach of concatenating learnable soft prompts with hard class tokens to replace handcrafted text prompts (e.g., \"a photo of a {class}\") as inputs to the encoder. Although this text prompt demonstrates strong performance, it restricts image alignment during training to predefined known categories only, thereby preventing accurate associations with unknown categories, as shown in Fig. 1(a). Intuitively, when confronted with an unfamiliar category, humans often associate it with additional attributes (e.g., color, shape or texture) to increase comprehensibility and clarity, rather than merely stating the object's name. For instance, one might describe a cheetah as: \"The cheetah is a cat-like animal with a small head, short yellow hair, and black spots.\" or refer to an apple as \"That red spherical fruit with orange stripes is an apple.\" instead of using a general description such as \"This is a cheetah.\" or \"That fruit is an apple.\" Attributes can serve as bridges that connect unknown categories to our known knowledge." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.397, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Building on these observations, we propose a novel approach that leverages attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an attribute-anchored textual prompt learning method for VLMs, named ATPrompt. This method extends the learning space of soft prompts from the original one-dimensional category level to a multidimensional attribute level by integrating multiple fixed universal attribute tokens into the learnable soft prompts. Guided by these anchored attributes, soft tokens acquire not only category-specific but also attribute-related general representations during training. This results in improved alignment between images and unknown categories compared to the original method, as shown in Fig. 1(b). Additionally, based on the depth at which soft prompts are applied, we propose two versions of ATPrompt: shallow and deep, respectively, ensuring compatibility with existing methods of varying depths [13, 14, 52]. To finalize these attributes, we present a simple and effective differentiable attribute search method that learns to identify suitable attributes from a candidate pool constructed by LLMs. The search operation only needs to be performed once per task, and once completed, the selected attributes can be used by ATPrompt for model training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.482, + 0.821 + ], + "angle": 0, + "content": "As an easy-to-use plug-in technique, ATPrompt can seamlessly substitute existing forms used in textual-based prompt learning methods, yielding general improvements with negligible additional computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.435, + 0.837 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- We propose an efficient attribute-anchored textual prompt learning method that expands the learning space of soft prompts from a one-dimensional class level to the multidimensional attribute level." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "- We introduce an effective differentiable search method to select appropriate attributes for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.167 + ], + "angle": 0, + "content": "- Both shallow and deep versions of ATPrompt are introduced to ensure compatibility with existing prompt learning methods of varying depths." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that ATPrompt can be seamlessly integrated into existing textual-based methods, resulting in consistently improved performance with negligible computational costs." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.241, + 0.655, + 0.258 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.705 + ], + "angle": 0, + "content": "Prompt Learning for VLMs. Inspired by recent advancements in NLP [18, 19], prompt learning [13, 14, 23, 33, 48, 51-53] has garnered significant interest among vision researchers aiming to apply these techniques to VLMs [10, 32, 42], such as CLIP. CoOp [52] is the pioneering text-based approach that introduced the concept of using a combination of soft textual tokens and a hard class token as input. Subsequent studies [13, 14, 17, 23, 41, 44, 51] have predominantly followed this textual prompt format. However, this form constrains the soft prompts to align with images within a one-dimensional, predefined category space, limiting their applicability to unknown categories. Therefore, training based on the current text form will be more likely to overfit to known categories, diminishing their zero-shot generalization capability for unknown categories. To address this limitation, several methods have been proposed [12, 14, 23, 44, 53]. For instance, Kg-CoOp [44] uses hand-crafted hard prompts to regularize learnable soft prompts during training. PromptSRC [14] utilizes CLIP's [32] original features to regularize the learning of soft prompts for image and text branches. PromptKD [23] utilizes a pre-trained strong teacher model to guide the learning of a student model [8, 22, 24, 43] with learnable prompts. Despite these advancements, none of the above methods address the inherent limitations of the format itself. In this work, we introduce the attribute-anchored textual prompt format for VLMs, which proposes to utilize attributes as a bridge to build more accurate associations between images and unknown categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Attributes for VLMs. In practice, categories typically encompass multiple attributes. When individuals encounter an unfamiliar category, they often describe it using additional attributes to enhance the clarity of their communication, rather than merely stating its name. Inspired by this observation, numerous studies [4, 30, 37, 38, 45] have begun to leverage attributes to support their objectives. VCD [30] was the pioneering work to propose the use of LLMs to decompose class names into multiple in-class attributes (e.g., beak and tail for birds) for classification. AAPL [15] introduces a meta-network to extract visual attribute features based on encoded image features, facilitating image-text alignment. TAP [5] presents a structured \"Tree of At" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.266, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.093, + 0.454, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.092, + 0.663, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.092, + 0.895, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.223, + 0.907, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Architectural comparison among existing methods. (a) Vanilla CLIP employs a hand-crafted text template as input to the text encoder. (b) Classical prompt learning proposes a new text form that concatenates multiple learnable soft tokens with class tokens. (c) ArGue [37] employs multiple in-class attributes mined by LLMs as supplementary information. These attributes are utilized to construct distinct text groups, which serve as learning targets, thereby regularizing the learning of soft tokens. The final prediction is achieved by assembling all groups. (d) Our ATPrompt treats universal attributes as learning components and anchors them into existing soft prompt templates. Though this operation, we expand the learning space of soft tokens to multi-dimensional attribute levels and facilitate the alignment of images with unknown class texts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.482, + 0.516 + ], + "angle": 0, + "content": "attributes\" approach that leverages attribute-specific knowledge graphs to enhance VLMs. ArGue [37] leverages large language models to mine multiple in-class attributes and integrates them into soft prompts and fixed templates to create multiple text groups. It employs the original text features generated by the fixed template to regularize the learning of soft tokens, as illustrated in Fig. 2(c). Most prior studies have focused on utilizing in-class attributes to enhance model performance by providing supplementary attribute information. However, when dealing with an unknown class, reacquiring the attributes of the new class becomes necessary—a process that is both complex and costly." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.668 + ], + "angle": 0, + "content": "In this work, we believe that universal (inter-class) attributes are more efficient and robust than the in-class attributes used in previous works. Instead of taking attributes as the learning objective, we treat it as a learning component and propose to anchor universal attributes into soft prompt templates, transforming the existing class-centric form [52] into a hybrid attribute-class learnable textual prompt form. Our approach can be seamlessly integrated into existing textual-based methods, enhancing their performance without incurring additional computational costs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.182, + 0.695 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Prompt learning [13, 14, 51-53] aims to enhance the generalization ability of pre-trained VLMs like CLIP on downstream tasks by training inserted learnable soft tokens. Existing textual-based methods all follow the classic prompt paradigm, concatenating soft prompt tokens and hard class tokens as the input to the text encoder, as shown in Fig. 2(b). In this paper, we propose a simple and effective textual prompt learning method, named ATPrompt, which anchors multiple fixed universal attribute tokens into the original soft prompts, as shown in Fig. 2(d). Guided by these attributes, soft prompts can learn not only category-specific but also attribute-related general representations through training. When encountering unknown categories, these" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.516 + ], + "angle": 0, + "content": "learned attribute-related tokens can provide additional information to promote better image-text alignment. Furthermore, to identify these universal attributes, we present an automated pipeline that encompasses sequential steps. Initially, we employ LLMs to synthesize the attribute pool for the current downstream category. Subsequently, we propose a differentiable attribute search method designed to identify attributes within the pool that are most suitable for our attribute-anchored prompt forms. For every task, this search operation is conducted only once. Once the attributes are finalized, they are integrated into our ATPrompt for specialized model fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.644, + 0.545 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Vision-Language Models. Existing VLMs, such as CLIP [32], have demonstrated remarkable zero-shot generalization performance after training with 400 million image-text pairs. The primary objective of these models is to learn the alignment between image and text modalities produced by each encoder [6, 27, 49]. Given a labeled image classification dataset \\(\\mathrm{D} = \\{(x,c)\\}\\) which includes \\(N\\) class labels \\(C = \\{c_i\\}_{i = 1}^N\\), CLIP makes predictions by calculating the cosine similarity between image features and the text features of each class. Specifically, for each input image \\(x\\), it undergoes feature extraction via the image encoder \\(h_I(x)\\) and obtains a feature vector \\(u = h_I(x)\\). Simultaneously, for each class, a series of textual descriptions \\(t\\) are generated using the hand-crafted template. Then, these text descriptions are fed into the text encoder \\(h_T(x)\\) to obtain text features \\(w = h_{T}(t)\\). Finally, the output probability for image \\(x\\) classified to \\(c\\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.822, + 0.905, + 0.859 + ], + "angle": 0, + "content": "\\[\np (c | x) = \\frac {\\exp \\left(\\cos \\left(u , w _ {c}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\cos \\left(u , w _ {i}\\right) / \\tau\\right)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter and \\(\\cos (\\cdot ,\\cdot)\\) denotes cosine similarity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3620" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.155, + 0.09, + 0.446, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.287, + 0.236, + 0.398, + 0.248 + ], + "angle": 0, + "content": "(a) Shallow Version" + }, + { + "type": "image", + "bbox": [ + 0.478, + 0.094, + 0.707, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.236, + 0.652, + 0.248 + ], + "angle": 0, + "content": "(b) Deep Version" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.126, + 0.844, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 3. An illustration of the computation process for shallow and deep versions. Take two attributes [A] and [B] as examples. (a) The shallow version concatenates hard attribute tokens, soft prompt tokens, and class tokens and inputs them into the encoder for calculation. (b) The deep version uses the same input but discards the class-related soft prompt tokens after calculating the self-attention and introduces them again before the next layer. These two forms can be compatible with existing methods of varying prompt depths, including input-level ones like CoOp [52], CoCoOp [51] and depth-level ones like MaPLe [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.463 + ], + "angle": 0, + "content": "Prompt Learning for VLMs. Instead of manually designed hard prompts for image-text alignment, which is inaccurate and inflexible, recent prompt learning works [13, 14, 44, 51, 52] like CoOp propose to learn appropriate soft textual prompts for downstream tasks. Concretely, \\(M\\) learnable soft tokens \\([T_i]_{i=1}^M\\) are concatenated with the hard class token [CLS] as the input of the text encoder, as shown in Fig. 2(b). Its form is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.471, + 0.482, + 0.489 + ], + "angle": 0, + "content": "\\[\nP _ {T} = \\left[ T _ {1} \\right] \\left[ T _ {2} \\right] \\dots \\left[ T _ {M} \\right] [ \\text {C L S} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.526 + ], + "angle": 0, + "content": "where \\( M \\) represents the length of soft tokens. For simplicity, we omit the prefix and suffix tokens in the input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.618 + ], + "angle": 0, + "content": "In addition to embedding soft tokens at the input level, existing studies [11, 13, 14, 23] have also explored introducing them at deeper layers. This is achieved by adding soft tokens within the Transformer blocks and subsequently removing them after the self-attention computations. For the \\(i\\)-th block, this process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.626, + 0.483, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.65, + 0.484, + 0.698 + ], + "angle": 0, + "content": "where \\(L_{i}\\) represents the \\(i\\)-th transformer block and \\(\\mathrm{T}_i\\) denotes the set of learnable soft tokens, defined as \\(\\mathrm{T}_i = \\{[T_1]_i,\\dots,[T_M]_i\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.704, + 0.484, + 0.721 + ], + "angle": 0, + "content": "3.2. Learning soft prompts with universal attributes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Our approach introduces two variants, distinguished by the number of layers at which soft tokens are applied: a shallow version and a deep version, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Shallow Version. We begin by introducing the shallow version, in which hard attribute tokens are anchored solely at the input level, as illustrated in Fig. 3(a). Consider two universal attributes, A and B. According to Eqn. (2), the shallow-level text prompt \\( P_T \\) provided to the text encoder can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nP _ {T} = \\left[ T _ {a _ {1}} \\right] \\dots \\left[ T _ {a _ {m}} \\right] \\left[ A \\right] \\left[ T _ {b _ {1}} \\right] \\dots \\left[ T _ {b _ {m}} \\right] \\left[ B \\right] \\left[ T _ {1} \\right] \\dots \\left[ T _ {M} \\right] \\left[ C L S \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.387 + ], + "angle": 0, + "content": "where \\( a_{m} \\) and \\( b_{m} \\) are hyperparameters specifying the length of soft tokens for attributes A and B. In our method, we set these parameters to be the same by default." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.463 + ], + "angle": 0, + "content": "While this example places the class token at the end, we also evaluate configurations where it is positioned at the front or in the middle. As shown in Tab. 5, the end position yields the best performance and is therefore adopted as our default configuration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Deep Version. In this version, learnable soft tokens are introduced at the input of the deep layers. Previous works, such as VPT and MaPLe, discard all soft tokens and subsequently reintroduce them after the block. When this operation is applied to attribute-related words, a gap will emerge between the introduced excessive low-level tokens and the existing high-level tokens, thereby weakening the feature continuity across layers. In this study, our approach selectively discards and then re-adds only class-related soft tokens in the input, specifically \\([T_1],\\dots,[T_M]\\), as shown in Fig. 3(b). Based on Eqn. (3), the deep version of ATPrompt can be rewritten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.65, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n[ \\mathrm {F} _ {1}, \\rightarrow , \\mathrm {C L S} _ {1} ] = L _ {1} ([ \\mathrm {T} _ {a _ {0}}, \\mathrm {A}, \\mathrm {T} _ {b _ {0}}, \\mathrm {B}, \\mathrm {T} _ {0}, \\mathrm {C L S} _ {0} ]), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.671, + 0.905, + 0.696 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathrm {F} _ {i, -}, \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {F} _ {i - 1}, \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.692, + 0.748, + 0.706 + ], + "angle": 0, + "content": "\\[\ni = 2, 3, \\dots , M.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.906, + 0.754 + ], + "angle": 0, + "content": "where \\(\\mathrm{F}_i\\) represents the features computed by the \\(i\\)-th Transformer layer. We demonstrate the effectiveness of this operation in Tab. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.907, + 0.844 + ], + "angle": 0, + "content": "Training. Let \\(\\theta\\) represent the weight of the total soft tokens and let \\(v\\) denote the selected fixed attribute tokens. Training is conducted on a labeled dataset \\(\\mathrm{D} = \\{(x,c)\\}\\), with the objective of minimizing the cross-entropy loss between predicted values and ground truth labels. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.849, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} L _ {t r a i n} = \\min _ {\\theta} \\sum_ {x \\in D} \\operatorname {C E} (f (x; v, \\theta), c). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.871, + 0.902 + ], + "angle": 0, + "content": "where \\( f(\\cdot) \\) represents the function of the CLIP model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3621" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.352, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 4. An overview of our attribute search pipeline. (a) We first query the LLM iteratively to obtain multiple independent attributes. These are subsequently aggregated to form a pool of candidate combinations, which serves as the input for the search process. (b) The forward computation for each candidate combination is represented by a distinct colored path. To identify the optimal attributes, we employ an alternating optimization algorithm that co-optimizes the soft tokens and a corresponding path weight vector. Upon completion of training, the combination associated with the highest-weighted path is selected as the final output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.254, + 0.453 + ], + "angle": 0, + "content": "3.3. Attribute search" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.596 + ], + "angle": 0, + "content": "Selecting the attributes involves two key considerations: their content and their quantity. While directly querying a LLM is a straightforward approach, it has notable drawbacks. This method cannot determine the optimal number of attributes for a specific downstream dataset, and querying by category name alone can introduce semantic bias. To address these issues, we propose an automated pipeline that selects the appropriate content and quantity of attributes for the current downstream task, as illustrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Attribute Pool. Inspired by CoT [39, 50], we divide the entire process into multiple steps to enhance the reasoning ability of LLMs. First, we prompt the LLM to generate descriptive sentences for each known category, thereby enriching category-related information. Using these descriptions as context, we then prompt the LLM to summarize a set of independent attribute bases that are common across these categories. An attribute pool is then formed by creating all possible combinations of these bases, as shown in Fig. 4(a). For \\( N \\) attribute bases, this result in a total of \\( w = C_N^1 + C_N^2 + \\ldots + C_N^N \\) candidates in the pool, which constitutes our search space. Note that we do not consider the order of attributes, as permutations generally do not introduce significant semantic bias or affect the final performance, a claim we validate in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Attribute Searching. Inspired by DARTS [26], we introduce a differentiable attribute search method that learns to find representative attributes \\( v \\) from the search space \\( \\mathcal{V} \\), as shown in Fig. 4(b). To make the search space continuous, we relax the discrete attribute selection into a softmax-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.814, + 0.454 + ], + "angle": 0, + "content": "weighted sum over all \\( w \\) possible candidates:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.465, + 0.907, + 0.502 + ], + "angle": 0, + "content": "\\[\nf (x, v; \\alpha , \\theta) = \\sum_ {i \\in \\mathcal {V}} \\frac {\\exp (\\alpha_ {i})}{\\sum_ {i ^ {\\prime} \\in \\mathcal {V}} \\exp (\\alpha_ {i ^ {\\prime}})} f (x, v _ {i}; \\theta). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.905, + 0.558 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) represents the weight for attribute combination \\(v_{i}\\). The task of attribute search is thus reduced to learning the weight vector \\(\\alpha\\) for the candidate pool." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.678 + ], + "angle": 0, + "content": "After relaxation, our goal is to jointly learn the attribute weight \\(\\alpha\\) and soft prompt tokens \\(\\theta\\). Following standard practice [26, 31, 54], we optimize the weights \\(\\alpha\\) by minimizing the validation loss \\(L_{val}\\), while the soft tokens are learned by minimizing the training loss \\(L_{train}\\). We employ an alternating algorithm [9, 22] to solve this bi-level optimization problem, alternating between these two subproblems:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.692, + 0.907, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\hat {\\alpha} = \\arg \\min _ {\\alpha} L _ {v a l} (f (x, v; \\alpha , \\hat {\\theta}), c), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.721, + 0.905, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\hat {\\theta} = \\arg \\min _ {\\theta} L _ {t r a i n} (f (x, v; \\hat {\\alpha}, \\theta), c). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.795 + ], + "angle": 0, + "content": "where \\( L_{train} \\) and \\( L_{val} \\) both use the cross-entropy loss function. After the search, the attribute combination with the highest weight \\( (\\alpha_{i}) \\) is selected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Cost Analysis. Unlike traditional Neural Architecture Search (NAS) methods [7, 25, 36], which search computationally expensive network-level parameters, our approach focuses on a lightweight token-level search space. This design makes our method significantly more efficient than previous approaches. In practice, our search converges in approximately 5 epochs, often requiring less than 5 minutes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3622" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.842, + 0.284 + ], + "angle": 0, + "content": "
MethodAverageImageNetCaltech101OxfordPots
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (UCV 22)82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp (CVPR 22)80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
MaPLc (CVPR 23)82.2875.1478.5576.6670.5473.4797.7494.3696.0295.4397.7696.58
PromptSRC (ICCV 23)84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
ArGue (CVPR 24)83.6978.0780.7876.9272.0674.4198.4395.2096.7995.3697.9596.64
DePT (CVPR 24)83.6671.8277.2977.1370.1073.4598.3394.3396.2994.7097.6396.14
CoPrompt (ICLR 24)84.0077.2380.4877.6771.2774.3398.2794.9096.5595.6798.1096.87
PromptKD (CVPR 24)86.9680.7383.7380.8374.6677.6298.9196.6597.7796.3098.0197.15
CoOp + ATPrompt82.6868.0474.65 (+2.99)76.2770.6073.3397.9593.6395.7494.7796.5995.67
CoCoOp + ATPrompt81.6974.5477.95 (+2.12)76.4370.5073.3597.9695.2796.6095.4697.8996.66
MaPLc + ATPrompt82.9875.7679.21 (+0.66)76.9470.7273.7098.3295.0996.6895.6297.6396.61
DePT + ATPrompt83.8073.7578.45 (+1.16)77.3270.6573.8398.4894.6096.5094.6597.9996.29
PromptKD + ATPrompt87.0581.8284.35 (+0.62)80.9074.8377.7598.9096.5297.7096.9298.2797.59
" + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.288, + 0.83, + 0.482 + ], + "angle": 0, + "content": "
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (JCV 22)78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp (CVPR 22)70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
MaPLe (CVPR 23)72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
PromptSRC (ICCV 23)78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
ArGue (CVPR 24)75.6473.3874.4998.3475.4185.3692.3391.9692.1440.4638.0339.21
DePT (CVPR 24)79.6772.4075.8698.2072.0083.0890.4391.3390.8842.5322.5329.46
CoPrompt (ICLR 24)76.9774.4075.6697.2776.6085.7190.7392.0791.4040.2039.3339.76
PromptKD (CVPR 24)82.8083.3783.1399.4282.6290.2492.4393.6893.0549.1241.8145.17
CoOp + ATPrompt77.4366.5571.5897.4467.5279.7788.7487.4488.0940.3827.2232.52
CoCoOp + ATPrompt74.5073.4773.9896.5273.5983.5190.5991.7491.1637.3033.1535.10
MaPLe + ATPrompt75.3973.8474.6197.8275.0784.9590.6592.0091.3237.6136.1536.87
DePT + ATPrompt79.2973.4776.2798.2073.6984.2090.4291.6991.0543.1933.2337.56
PromptKD + ATPrompt82.5184.0383.2699.1582.0389.7892.4893.8693.2249.6342.3545.70
" + }, + { + "type": "table", + "bbox": [ + 0.169, + 0.487, + 0.829, + 0.682 + ], + "angle": 0, + "content": "
MethodSUN397DTDEuroSATUCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (IUCV 22)80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp (CVPR 22)79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
MaPLe (CVPR 23)80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
PromptSRC (ICCV 23)82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
ArGue (CVPR 24)81.5280.7481.1381.6066.5573.3194.4388.2491.2385.5679.2982.31
DePT (CVPR 24)82.3775.0778.5583.2056.1367.0488.2766.2775.7085.4372.1778.24
CoPrompt (ICLR 24)82.6380.0381.3083.1364.7372.7994.6078.5785.8486.9079.5783.07
PromptKD (CVPR 24)83.6981.5482.6085.8471.3777.9497.5482.0889.1489.7182.2786.10
CoOp + ATPrompt80.8468.6474.2480.8345.4958.2290.3459.7971.9684.4964.9673.45
CoCoOp + ATPrompt80.5076.8678.6478.6356.8966.0287.9574.1580.4682.7476.4079.44
MaPLe + ATPrompt80.9878.1579.5480.5058.2867.6194.8477.5985.3584.0878.8881.40
DePT + ATPrompt82.4276.4879.3482.6456.7767.3089.6069.5078.2885.6073.1578.89
PromptKD + ATPrompt83.8781.3582.5986.9272.3478.9697.0592.0794.4989.2982.4485.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.689, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Table 1. Base-to-novel generalization experiments of five baselines with and without our ATPrompt on 11 datasets. Our method achieves consistent average performance improvement over different baselines." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.723, + 0.849, + 0.848 + ], + "angle": 0, + "content": "
MethodSourceTarget DatasetAverage
Image NetCaltech 101Oxford PetsStanford CarsFlowers 102Food101FGVC AircraftSUN397DTDEuro SATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
+ATPrompt71.6793.9690.6565.0170.4085.8620.9765.7743.4446.5969.9265.26 (+1.38)
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
+ATPrompt71.2793.7990.6265.9071.1786.0323.2266.6344.4448.7070.7166.59 (+0.85)
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
+ATPrompt70.6994.0491.0366.0671.9986.3324.4267.0545.2148.6369.1566.75 (+0.45)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.853, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Table 2. Cross-dataset generalization experiments of three baselines with and without our ATPrompt on 11 datasets. ATPrompt achieves consistent average performance improvement on target datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3623" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.479, + 0.287 + ], + "angle": 0, + "content": "
DatasetAttribute BasesSearched Results
ImageNetcolor, size, shape, habitat, behavior(color, shape)
Caltech101shape, color, material, function, size(shape,size)
OxfordPetsloyalty, affection, energy, playfulness, intelligence(playfulness, energy)
StanfordCarsdesign, engine, performance, luxury, color(luxury)
Flowers102color, flower, habitat, growth, season(color, habitat, growth)
Food101flavor, texture, origin, ingredients, preparation(flavor, preparation)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Table 3. Part of the results obtained after differentiable attribute search. Please refer to the appendix for the complete results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.483, + 0.386 + ], + "angle": 0, + "content": "on a single A800 GPU for some datasets. In addition, search efficiency can be improved by curating a smaller set of base attributes, which narrows the search space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.224, + 0.416 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.192, + 0.44 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Base-to-Novel Generalization. Following [14, 23, 51, 52], we split the dataset into base and novel classes. The model is trained on the base class training set and evaluated on the test set. The attributes applied in this experiment are searched based on the base class. For datasets without a dedicated validation set, like ImageNet, we split the 16-shot labeled data, using half for training and the other half for attribute search." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Cross-dataset Experiments. Consistent with previous works [14, 51, 52], we first train a model on the ImageNet-1K source dataset and then evaluate its generalization performance on several out-of-distribution datasets. The attributes used are obtained from the source dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.733 + ], + "angle": 0, + "content": "Attribute Search. We select five independent attributes as the basis in the attribute pool. This results in 31 candidate attribute combinations for the search process. We use ChatGPT-4o for attribute queries. In Tab. 3 reports a subset of the queried attribute bases and the final combinations selected by our search." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Implementation Details. We evaluate the model performance on 15 datasets. We report base and novel class accuracy and their Harmonic Mean (HM) averaged over 3 runs. The details of each dataset are attached in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.353, + 0.819 + ], + "angle": 0, + "content": "4.2. Base-to-Novel Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "As demonstrated in Tab. 1, we evaluate the base-to-novel generalization performance of five baseline methods, both with and without integrating ATPrompt, across 11 diverse recognition datasets. Notably, ATPrompt consistently improves the average performance of all baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "Reasons for Limited Improvement in Some Conditions. (1) Learnable text prompts are the central component in earlier works, and optimizing them through ATPrompt can obviously improve the performance. (2) Recent studies have expanded beyond learnable text prompts by introducing additional learnable modules. Since our work only involves the optimization of the learnable text prompt part, the improvement have become less obvious." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.744, + 0.24 + ], + "angle": 0, + "content": "4.3. Cross-dataset Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Tab. 2 shows the cross-dataset generalization results for three baseline methods. Our method demonstrates superior performance, with improvements of \\(1.38\\%\\), \\(0.85\\%\\) and \\(0.45\\%\\) for CoOp, CoCoOp and MaPLe, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.73, + 0.336 + ], + "angle": 0, + "content": "4.4. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.405 + ], + "angle": 0, + "content": "Tab. 4 shows the domain generalization results for three baseline methods. The results show that our method improves CoOp, CoCoOp and MaPLe methods by \\(0.90\\%\\), \\(0.49\\%\\) and \\(0.33\\%\\) respectively." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.418, + 0.9, + 0.537 + ], + "angle": 0, + "content": "
MethodSourceTarget DatasetAverage
ImageNet-V2-S-A-R
CoOp71.5164.2047.9949.7175.2159.28
+ATPrompt71.6764.4349.1350.9176.2460.18 (+0.90)
CoCoOp71.0264.0748.7550.6376.1859.91
+ATPrompt71.2764.6649.1551.4476.3360.40 (+0.49)
MaPLe70.7264.0749.1550.9076.9860.27
+ATPrompt70.6964.4049.1051.7777.1160.60 (+0.33)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.547, + 0.907, + 0.589 + ], + "angle": 0, + "content": "Table 4. Domain generalization experiments of three baselines with and without our ATPrompt on four datasets. The integration of ATPrompt resulted in better generalization performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.682, + 0.629 + ], + "angle": 0, + "content": "4.5. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.906, + 0.697 + ], + "angle": 0, + "content": "By default, the experiments are conducted on the ImageNet. To minimize the influence of other components, we mainly adopt CoOp as the baseline method. Two attributes (color and shape) are used in our ATPrompt." + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.713, + 0.828, + 0.858 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.862, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Figure 5. Illustration of varying soft token length on ImageNet. Increased tokens can lead to overfitting to the base class and weaken generalization to the novel class." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "Soft Prompt Length. In Fig. 5, we examine the optimal soft token length for both attribute and class tokens. By varying the length from 1 to 10, we observe that longer prompts dilute the guiding influence of the attribute tokens, thereby reducing generalization to novel classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Class Token Position. The relative positioning of attribute tokens and class tokens in our method requires careful consideration. In Tab. 5, we examine configurations where the class token is positioned in the middle or on either side of two attribute tokens. The results demonstrate that optimal performance is achieved when the class token is placed at the end, which aligns with the findings of CoOp." + }, + { + "type": "table", + "bbox": [ + 0.18, + 0.284, + 0.395, + 0.352 + ], + "angle": 0, + "content": "
PositionBaseNovelHM
Front76.1270.5073.20
Middle76.1370.2973.09
End76.2770.6073.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.391 + ], + "angle": 0, + "content": "Table 5. Comparison of different class token positions on ImageNet. The end position works best." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.482, + 0.505 + ], + "angle": 0, + "content": "Prompt Operation of Deep Version. In ATPrompt-Deep, we exclusively drop class soft tokens while retaining both hard and soft attribute tokens after they pass through the block. In this part, we compare the performance of partial drop (i.e., removing attribute soft tokens while retaining hard tokens) and full drop (i.e., removing both attribute soft and hard tokens) operations, as illustrated in Tab. 6." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.517, + 0.452, + 0.582 + ], + "angle": 0, + "content": "
Operation of Attribute TokenBaseNovelHM
Retain all hard and soft tokens76.9470.7273.70
Partial drop and re-add76.8770.4473.51
Full drop and re-add76.8370.1073.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Table 6. Comparison of operations on deep soft and hard attribute tokens based on MaPLe+ATPrompt. Preserving hard and soft attribute tokens in deep layers performs better than other operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.482, + 0.735 + ], + "angle": 0, + "content": "The results show that maintaining the attributes of both hard and soft tokens during the forward process results in optimal performance. Conversely, dropping and reintroducing either hard or soft attribute tokens harms performance, likely because it disrupts the continuity of attribute representations across layers and complicates optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.482, + 0.855 + ], + "angle": 0, + "content": "Attribute Order. In this study, we do not specifically focus on the order of attributes because varying the sequence usually does not result in semantic deviations in reality. Tab. 7 quantitatively assesses the impact of attribute order on prompt learning performance. From this table, we observe that despite variations in order, similar results are consistently produced, and the performance fluctuations across different orders remain within a reasonable range." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Comparison to Other Attributes. In Tab. 8, we explore the effectiveness of attributes derived through alternative methods, specifically by manually selecting class-irrelevant" + }, + { + "type": "table", + "bbox": [ + 0.598, + 0.089, + 0.825, + 0.168 + ], + "angle": 0, + "content": "
AttributesBaseNovelHM
(shape, color)76.3270.3973.24
(color, shape)76.2770.6073.33
(size, habitat)76.4470.2373.20
(habitat, size)76.4670.1673.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.178, + 0.905, + 0.22 + ], + "angle": 0, + "content": "Table 7. Comparison of different orders on ImageNet. The order of attributes does not significantly affect the model, and performance fluctuations are within a reasonable range." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.905, + 0.331 + ], + "angle": 0, + "content": "and common attributes. It shows that manually selected irrelevant attributes exhibit comparable performance during training; however, they perform poorly when applied to new categories. This suggests that incorrect attribute tokens cause the soft tokens to develop biased representations, thereby diminishing their zero-shot generalization ability." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.341, + 0.865, + 0.438 + ], + "angle": 0, + "content": "
TypeAttributesBaseNovelHM
Common(shape, size)82.8367.1374.16
(color, texture)82.7367.5674.38
Irrelevant( plane, engines)82.8166.2273.59
(football, sport)82.7767.1474.14
Searched-82.6868.0474.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.449, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Table 8. Comparison of average performance for various attribute configurations on 11 datasets. The attributes obtained by our method achieve the best performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.633, + 0.53 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.905, + 0.78 + ], + "angle": 0, + "content": "In this work, we introduce ATPrompt, an attribute-anchored textual prompt learning method that uses universal attributes as a bridge to improve generalization from seen to unseen categories. Our approach expands the learning space of soft prompts from a one-dimensional, category-centric structure to a multi-dimensional attribute space by anchoring fixed attribute tokens with the prompt. To ensure the selection of optimal attributes, we propose an automated pipeline designed to identify the most suitable candidates for any given downstream task. ATPrompt is designed with both shallow and deep architectural variants, rendering it broadly compatible with existing prompt learning methods. Extensive experiments validate the effectiveness of our approach. We believe this work offers a new direction for research into the fundamental structure of learnable prompts in prompt learning area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Limitations and future works. This work is a preliminary study of the basic prompt form, which has not been able to achieve comparable performance to regularization-based methods when working alone. Furthermore, the current selection of attribute anchors relies on manual experimentation, and automatically discovering the optimal anchor positions through a learning-based approach remains a promising direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3625" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.268 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62361166670 and U24A20330. This research was also supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206134), the Fundamental Research Funds for the Central Universities 070-63253222 and the Tianjin Key Laboratory of Visual Computing and Intelligent Perception (VCIP). Computation is supported by the Supercomputing Center of Nankai University (NKSC)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.284, + 0.188, + 0.3 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.31, + 0.484, + 0.351 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.354, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 35:25005-25017, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.398, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[3] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[4] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, pages 23518-23527, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[5] Tong Ding, Wanhua Li, Zhongqi Miao, and Hanspeter Pfister. Tree of attributes prompt learning for vision-language models. arXiv preprint arXiv:2410.11201, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.484, + 0.683 + ], + "angle": 0, + "content": "[7] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. JMLR, 20(55):1-21, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.484, + 0.727 + ], + "angle": 0, + "content": "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[9] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. NeurIPS, 33:5632-5643, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.31, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Gahyeon Kim, Sohee Kim, and Seokju Lee. Aapl: Adding attributes to prompt learning for vision-language models. In CVPR Workshop, pages 1572-1582, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[19] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[20] Yunheng Li, Yuxuan Li, Quansheng Zeng, Wenhai Wang, Qibin Hou, and Ming-Ming Cheng. Unbiased region-language alignment for open-vocabulary dense prediction. arXiv preprint arXiv:2412.06244, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[21] Yunheng Li, Zhong-Yu Li, Quan-Sheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-CLIP: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243-28258. PMLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[22] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[23] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. In CVPR, pages 26617-26626, 2024. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[24] Zheng Li, Xiang Li, Lingfeng Yang, Renjie Song, Jian Yang, and Zhigeng Pan. Dual teachers for self-knowledge distillation. Pattern Recognition, 151:110422, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3626" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.166, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[29] Hao Ma, Ming Li, Jingyuan Yang, Or Patashnik, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Clip-flow: Decoding images encoded in clip space. CVMJ, 10(6):1157-1168, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.483, + 0.264 + ], + "angle": 0, + "content": "[30] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.483, + 0.309 + ], + "angle": 0, + "content": "[31] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[33] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[34] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, pages 13019-13029, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[35] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[36] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[37] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.483, + 0.684 + ], + "angle": 0, + "content": "[38] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[39] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[40] Ge Wu, Xin Zhang, Zheng Li, Zhaowei Chen, Jiajun Liang, Jian Yang, and Xiang Li. Cascade prompt learning for vision-language model adaptation. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[41] Shihan Wu, Ji Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves. In CVPR, pages 14723-14732, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[42] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, and Yongjun Xu. Clip-kd: An empirical study of distilling clip models. arXiv preprint arXiv:2307.12732, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[43] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[44] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[45] Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, and Da Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In AAAI, pages 6979–6987, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[46] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In CVPR, pages 12924-12933, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.373 + ], + "angle": 0, + "content": "[47] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, et al. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.444 + ], + "angle": 0, + "content": "[48] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.906, + 0.514 + ], + "angle": 0, + "content": "[49] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.906, + 0.584 + ], + "angle": 0, + "content": "[50] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.906, + 0.628 + ], + "angle": 0, + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.67 + ], + "angle": 0, + "content": "[52] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.906, + 0.714 + ], + "angle": 0, + "content": "[53] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, pages 15659-15669, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[54] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3627" + } + ] +] \ No newline at end of file diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_origin.pdf b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c2bc8dea8aea512260bccd0dbb8e2f393eb34540 --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/4e7100e1-8488-4372-a2bb-73e12989371e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96d2a9d5961e3c0800e94f0abc02e8e0c7bfe79031d07b5de26b42ff1135961e +size 1382090 diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/full.md b/2025/Advancing Textual Prompt Learning with Anchored Attributes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a81115b9338eacf5af0782c48a00456f39a396e7 --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/full.md @@ -0,0 +1,321 @@ +# Advancing Textual Prompt Learning with Anchored Attributes + +Zheng Li $^{1}$ , Yibing Song $^{2}$ , Ming-Ming Cheng $^{1}$ , Xiang Li $^{1*}$ , Jian Yang $^{1*}$ $^{1}$ PCA Lab, VCIP, College of Computer Science, Nankai University, $^{2}$ DAMO Academy, Alibaba Group + +zhengli97@mail.nankai.edu.cn, yibingsong.csv@gmail.com {cmm, xiang.li.implus, csjyang}@nankai.edu.cn + +# Abstract + +Textual-based prompt learning methods primarily employ multiple learnable soft prompts and hard class tokens in a cascading manner as text inputs, aiming to align image and text (category) spaces for downstream tasks. However, current training is restricted to aligning images with predefined known categories and cannot be associated with unknown categories. In this work, we propose utilizing universal attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an Attribute-anchored Textual Prompt learning method for vision-language models, named ATPrompt. This approach expands the learning space of soft prompts from the original one-dimensional category level into the multi-dimensional attribute level by incorporating multiple attribute tokens into the learnable soft prompts. Through this modification, we transform the text prompt from a category-centric form to an attribute-category hybrid form. Additionally, we introduce a straightforward differentiable attribute search method to identify representative and suitable attributes for downstream tasks. As an easy-to-use plug-in technique, ATPrompt can seamlessly replace the existing basic prompt format in textual-based methods, providing general improvements at a negligible computational cost. Extensive experiments across 11 datasets validate the effectiveness of our method. Code is publicly available at https://github.com/zhengli97/ATPrompt. + +# 1. Introduction + +Vision-Language Models (VLMs) [10, 20, 21, 28, 29, 32, 34, 35], such as CLIP [3, 32] and ALIGN [10], have demonstrated exceptional performance in recent years. These models are trained with a contrastive loss to establish alignment between image and text (category) space. Inspired by the success of NLP [18, 19], prompt learning [52] has + +![](images/404ad711120a7070a0481b67a504644d00bcebcafb9099f47527b71126fd0ab6.jpg) + +![](images/dc204e01d93682cd3d39978bd1666f2a400475d7802824a1f0afd311d5c2603e.jpg) +Figure 1. Comparison of image and text (category) alignment processes through learnable prompts. (a) Current prompt learning methods align images with predefined categories but fail to establish accurate associations with unknown categories. (b) ATPrompt leverages universal attributes as an intermediary to create more accurate alignments between images and unknown categories. + +emerged as a parameter-efficient tool to adapt powerful VLMs to downstream tasks. Models with a few learnable soft prompt tokens [11, 47, 51] can achieve performance parity with, or even outperform, fully fine-tuned ones [11]. Depending on how the soft prompt tokens are applied, existing methods can be broadly classified into textual-based [13, 14, 23, 40, 51, 52] and visual-based approaches [1, 2, 11, 16, 46]. Among these, the textual-based method is the most fundamental and straightforward, comprising the majority. + +In typical image classification tasks, current text-based methods [13, 14, 51, 52] predominantly employ the traditional approach of concatenating learnable soft prompts with hard class tokens to replace handcrafted text prompts (e.g., "a photo of a {class}") as inputs to the encoder. Although this text prompt demonstrates strong performance, it restricts image alignment during training to predefined known categories only, thereby preventing accurate associations with unknown categories, as shown in Fig. 1(a). Intuitively, when confronted with an unfamiliar category, humans often associate it with additional attributes (e.g., color, shape or texture) to increase comprehensibility and clarity, rather than merely stating the object's name. For instance, one might describe a cheetah as: "The cheetah is a cat-like animal with a small head, short yellow hair, and black spots." or refer to an apple as "That red spherical fruit with orange stripes is an apple." instead of using a general description such as "This is a cheetah." or "That fruit is an apple." Attributes can serve as bridges that connect unknown categories to our known knowledge. + +Building on these observations, we propose a novel approach that leverages attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an attribute-anchored textual prompt learning method for VLMs, named ATPrompt. This method extends the learning space of soft prompts from the original one-dimensional category level to a multidimensional attribute level by integrating multiple fixed universal attribute tokens into the learnable soft prompts. Guided by these anchored attributes, soft tokens acquire not only category-specific but also attribute-related general representations during training. This results in improved alignment between images and unknown categories compared to the original method, as shown in Fig. 1(b). Additionally, based on the depth at which soft prompts are applied, we propose two versions of ATPrompt: shallow and deep, respectively, ensuring compatibility with existing methods of varying depths [13, 14, 52]. To finalize these attributes, we present a simple and effective differentiable attribute search method that learns to identify suitable attributes from a candidate pool constructed by LLMs. The search operation only needs to be performed once per task, and once completed, the selected attributes can be used by ATPrompt for model training. + +As an easy-to-use plug-in technique, ATPrompt can seamlessly substitute existing forms used in textual-based prompt learning methods, yielding general improvements with negligible additional computational overhead. + +Our contributions can be summarized as follows: + +- We propose an efficient attribute-anchored textual prompt learning method that expands the learning space of soft prompts from a one-dimensional class level to the multidimensional attribute level. + +- We introduce an effective differentiable search method to select appropriate attributes for downstream tasks. +- Both shallow and deep versions of ATPrompt are introduced to ensure compatibility with existing prompt learning methods of varying depths. +- Extensive experiments demonstrate that ATPrompt can be seamlessly integrated into existing textual-based methods, resulting in consistently improved performance with negligible computational costs. + +# 2. Related Work + +Prompt Learning for VLMs. Inspired by recent advancements in NLP [18, 19], prompt learning [13, 14, 23, 33, 48, 51-53] has garnered significant interest among vision researchers aiming to apply these techniques to VLMs [10, 32, 42], such as CLIP. CoOp [52] is the pioneering text-based approach that introduced the concept of using a combination of soft textual tokens and a hard class token as input. Subsequent studies [13, 14, 17, 23, 41, 44, 51] have predominantly followed this textual prompt format. However, this form constrains the soft prompts to align with images within a one-dimensional, predefined category space, limiting their applicability to unknown categories. Therefore, training based on the current text form will be more likely to overfit to known categories, diminishing their zero-shot generalization capability for unknown categories. To address this limitation, several methods have been proposed [12, 14, 23, 44, 53]. For instance, Kg-CoOp [44] uses hand-crafted hard prompts to regularize learnable soft prompts during training. PromptSRC [14] utilizes CLIP's [32] original features to regularize the learning of soft prompts for image and text branches. PromptKD [23] utilizes a pre-trained strong teacher model to guide the learning of a student model [8, 22, 24, 43] with learnable prompts. Despite these advancements, none of the above methods address the inherent limitations of the format itself. In this work, we introduce the attribute-anchored textual prompt format for VLMs, which proposes to utilize attributes as a bridge to build more accurate associations between images and unknown categories. + +Attributes for VLMs. In practice, categories typically encompass multiple attributes. When individuals encounter an unfamiliar category, they often describe it using additional attributes to enhance the clarity of their communication, rather than merely stating its name. Inspired by this observation, numerous studies [4, 30, 37, 38, 45] have begun to leverage attributes to support their objectives. VCD [30] was the pioneering work to propose the use of LLMs to decompose class names into multiple in-class attributes (e.g., beak and tail for birds) for classification. AAPL [15] introduces a meta-network to extract visual attribute features based on encoded image features, facilitating image-text alignment. TAP [5] presents a structured "Tree of At + +![](images/b1d40b41b880b211d87c229cda2c4c252092476e1cbfc4054c41aceab0af5c0d.jpg) +Figure 2. Architectural comparison among existing methods. (a) Vanilla CLIP employs a hand-crafted text template as input to the text encoder. (b) Classical prompt learning proposes a new text form that concatenates multiple learnable soft tokens with class tokens. (c) ArGue [37] employs multiple in-class attributes mined by LLMs as supplementary information. These attributes are utilized to construct distinct text groups, which serve as learning targets, thereby regularizing the learning of soft tokens. The final prediction is achieved by assembling all groups. (d) Our ATPrompt treats universal attributes as learning components and anchors them into existing soft prompt templates. Though this operation, we expand the learning space of soft tokens to multi-dimensional attribute levels and facilitate the alignment of images with unknown class texts. + +![](images/597a08010caa7fab85fbd011e3d4b0549890acd44cd4ecfe046c3dea2f263c3b.jpg) + +![](images/f1c2500e0697a38ef1eca8915369c9e894a2ce207cc27b37fd466af3bf5db66a.jpg) + +![](images/540938326c885781b22f84838f8676c2e0ff0eacc8034305263230f84f9a2dbb.jpg) + +attributes" approach that leverages attribute-specific knowledge graphs to enhance VLMs. ArGue [37] leverages large language models to mine multiple in-class attributes and integrates them into soft prompts and fixed templates to create multiple text groups. It employs the original text features generated by the fixed template to regularize the learning of soft tokens, as illustrated in Fig. 2(c). Most prior studies have focused on utilizing in-class attributes to enhance model performance by providing supplementary attribute information. However, when dealing with an unknown class, reacquiring the attributes of the new class becomes necessary—a process that is both complex and costly. + +In this work, we believe that universal (inter-class) attributes are more efficient and robust than the in-class attributes used in previous works. Instead of taking attributes as the learning objective, we treat it as a learning component and propose to anchor universal attributes into soft prompt templates, transforming the existing class-centric form [52] into a hybrid attribute-class learnable textual prompt form. Our approach can be seamlessly integrated into existing textual-based methods, enhancing their performance without incurring additional computational costs. + +# 3. Method + +Prompt learning [13, 14, 51-53] aims to enhance the generalization ability of pre-trained VLMs like CLIP on downstream tasks by training inserted learnable soft tokens. Existing textual-based methods all follow the classic prompt paradigm, concatenating soft prompt tokens and hard class tokens as the input to the text encoder, as shown in Fig. 2(b). In this paper, we propose a simple and effective textual prompt learning method, named ATPrompt, which anchors multiple fixed universal attribute tokens into the original soft prompts, as shown in Fig. 2(d). Guided by these attributes, soft prompts can learn not only category-specific but also attribute-related general representations through training. When encountering unknown categories, these + +learned attribute-related tokens can provide additional information to promote better image-text alignment. Furthermore, to identify these universal attributes, we present an automated pipeline that encompasses sequential steps. Initially, we employ LLMs to synthesize the attribute pool for the current downstream category. Subsequently, we propose a differentiable attribute search method designed to identify attributes within the pool that are most suitable for our attribute-anchored prompt forms. For every task, this search operation is conducted only once. Once the attributes are finalized, they are integrated into our ATPrompt for specialized model fine-tuning. + +# 3.1. Preliminary + +Vision-Language Models. Existing VLMs, such as CLIP [32], have demonstrated remarkable zero-shot generalization performance after training with 400 million image-text pairs. The primary objective of these models is to learn the alignment between image and text modalities produced by each encoder [6, 27, 49]. Given a labeled image classification dataset $\mathrm{D} = \{(x,c)\}$ which includes $N$ class labels $C = \{c_i\}_{i = 1}^N$ , CLIP makes predictions by calculating the cosine similarity between image features and the text features of each class. Specifically, for each input image $x$ , it undergoes feature extraction via the image encoder $h_I(x)$ and obtains a feature vector $u = h_I(x)$ . Simultaneously, for each class, a series of textual descriptions $t$ are generated using the hand-crafted template. Then, these text descriptions are fed into the text encoder $h_T(x)$ to obtain text features $w = h_{T}(t)$ . Finally, the output probability for image $x$ classified to $c$ is calculated as follows: + +$$ +p (c | x) = \frac {\exp \left(\cos \left(u , w _ {c}\right) / \tau\right)}{\sum_ {i = 1} ^ {N} \exp \left(\cos \left(u , w _ {i}\right) / \tau\right)}. \tag {1} +$$ + +where $\tau$ is the temperature parameter and $\cos (\cdot ,\cdot)$ denotes cosine similarity. + +![](images/2f77496042a509cb5255fb771dda64d1fcce2ecffd22ac7cf5c3499613787eff.jpg) +(a) Shallow Version + +![](images/46cb0444479a79004dbd45c76924ac713536483e63109a0b0f71e2f0fbf63e3f.jpg) +(b) Deep Version +Figure 3. An illustration of the computation process for shallow and deep versions. Take two attributes [A] and [B] as examples. (a) The shallow version concatenates hard attribute tokens, soft prompt tokens, and class tokens and inputs them into the encoder for calculation. (b) The deep version uses the same input but discards the class-related soft prompt tokens after calculating the self-attention and introduces them again before the next layer. These two forms can be compatible with existing methods of varying prompt depths, including input-level ones like CoOp [52], CoCoOp [51] and depth-level ones like MaPLe [13]. + +![](images/24b9b8627f22e44be57067d982e180ebfec30d25f3bdc9419e909b7f4ccf3eff.jpg) + +Prompt Learning for VLMs. Instead of manually designed hard prompts for image-text alignment, which is inaccurate and inflexible, recent prompt learning works [13, 14, 44, 51, 52] like CoOp propose to learn appropriate soft textual prompts for downstream tasks. Concretely, $M$ learnable soft tokens $[T_i]_{i=1}^M$ are concatenated with the hard class token [CLS] as the input of the text encoder, as shown in Fig. 2(b). Its form is shown as follows: + +$$ +P _ {T} = \left[ T _ {1} \right] \left[ T _ {2} \right] \dots \left[ T _ {M} \right] [ \text {C L S} ], \tag {2} +$$ + +where $M$ represents the length of soft tokens. For simplicity, we omit the prefix and suffix tokens in the input. + +In addition to embedding soft tokens at the input level, existing studies [11, 13, 14, 23] have also explored introducing them at deeper layers. This is achieved by adding soft tokens within the Transformer blocks and subsequently removing them after the self-attention computations. For the $i$ -th block, this process can be described as follows: + +$$ +\left[ \mathrm {C L S} _ {i} \right] = L _ {i} \left(\left[ \mathrm {T} _ {i - 1}, \mathrm {C L S} _ {i - 1} \right]\right). \tag {3} +$$ + +where $L_{i}$ represents the $i$ -th transformer block and $\mathrm{T}_i$ denotes the set of learnable soft tokens, defined as $\mathrm{T}_i = \{[T_1]_i,\dots,[T_M]_i\}$ . + +# 3.2. Learning soft prompts with universal attributes + +Our approach introduces two variants, distinguished by the number of layers at which soft tokens are applied: a shallow version and a deep version, as shown in Fig. 3. + +Shallow Version. We begin by introducing the shallow version, in which hard attribute tokens are anchored solely at the input level, as illustrated in Fig. 3(a). Consider two universal attributes, A and B. According to Eqn. (2), the shallow-level text prompt $P_T$ provided to the text encoder can be expressed as follows: + +$$ +P _ {T} = \left[ T _ {a _ {1}} \right] \dots \left[ T _ {a _ {m}} \right] \left[ A \right] \left[ T _ {b _ {1}} \right] \dots \left[ T _ {b _ {m}} \right] \left[ B \right] \left[ T _ {1} \right] \dots \left[ T _ {M} \right] \left[ C L S \right]. \tag {4} +$$ + +where $a_{m}$ and $b_{m}$ are hyperparameters specifying the length of soft tokens for attributes A and B. In our method, we set these parameters to be the same by default. + +While this example places the class token at the end, we also evaluate configurations where it is positioned at the front or in the middle. As shown in Tab. 5, the end position yields the best performance and is therefore adopted as our default configuration. + +Deep Version. In this version, learnable soft tokens are introduced at the input of the deep layers. Previous works, such as VPT and MaPLe, discard all soft tokens and subsequently reintroduce them after the block. When this operation is applied to attribute-related words, a gap will emerge between the introduced excessive low-level tokens and the existing high-level tokens, thereby weakening the feature continuity across layers. In this study, our approach selectively discards and then re-adds only class-related soft tokens in the input, specifically $[T_1],\dots,[T_M]$ , as shown in Fig. 3(b). Based on Eqn. (3), the deep version of ATPrompt can be rewritten as follows: + +$$ +[ \mathrm {F} _ {1}, \rightarrow , \mathrm {C L S} _ {1} ] = L _ {1} ([ \mathrm {T} _ {a _ {0}}, \mathrm {A}, \mathrm {T} _ {b _ {0}}, \mathrm {B}, \mathrm {T} _ {0}, \mathrm {C L S} _ {0} ]), \tag {5} +$$ + +$$ +\left[ \mathrm {F} _ {i, -}, \mathrm {C L S} _ {i} \right] = L _ {i} \left(\left[ \mathrm {F} _ {i - 1}, \mathrm {T} _ {i - 1}, \mathrm {C L S} _ {i - 1} \right]\right). \tag {6} +$$ + +$$ +i = 2, 3, \dots , M. +$$ + +where $\mathrm{F}_i$ represents the features computed by the $i$ -th Transformer layer. We demonstrate the effectiveness of this operation in Tab. 6. + +Training. Let $\theta$ represent the weight of the total soft tokens and let $v$ denote the selected fixed attribute tokens. Training is conducted on a labeled dataset $\mathrm{D} = \{(x,c)\}$ , with the objective of minimizing the cross-entropy loss between predicted values and ground truth labels. This process can be formulated as follows: + +$$ +\min _ {\theta} L _ {t r a i n} = \min _ {\theta} \sum_ {x \in D} \operatorname {C E} (f (x; v, \theta), c). \tag {7} +$$ + +where $f(\cdot)$ represents the function of the CLIP model. + +![](images/87de5756e56376e383ee44596442002f12bec98cf0bd6d9b0208e909ac934773.jpg) +Figure 4. An overview of our attribute search pipeline. (a) We first query the LLM iteratively to obtain multiple independent attributes. These are subsequently aggregated to form a pool of candidate combinations, which serves as the input for the search process. (b) The forward computation for each candidate combination is represented by a distinct colored path. To identify the optimal attributes, we employ an alternating optimization algorithm that co-optimizes the soft tokens and a corresponding path weight vector. Upon completion of training, the combination associated with the highest-weighted path is selected as the final output. + +# 3.3. Attribute search + +Selecting the attributes involves two key considerations: their content and their quantity. While directly querying a LLM is a straightforward approach, it has notable drawbacks. This method cannot determine the optimal number of attributes for a specific downstream dataset, and querying by category name alone can introduce semantic bias. To address these issues, we propose an automated pipeline that selects the appropriate content and quantity of attributes for the current downstream task, as illustrated in Fig. 4. + +Attribute Pool. Inspired by CoT [39, 50], we divide the entire process into multiple steps to enhance the reasoning ability of LLMs. First, we prompt the LLM to generate descriptive sentences for each known category, thereby enriching category-related information. Using these descriptions as context, we then prompt the LLM to summarize a set of independent attribute bases that are common across these categories. An attribute pool is then formed by creating all possible combinations of these bases, as shown in Fig. 4(a). For $N$ attribute bases, this result in a total of $w = C_N^1 + C_N^2 + \ldots + C_N^N$ candidates in the pool, which constitutes our search space. Note that we do not consider the order of attributes, as permutations generally do not introduce significant semantic bias or affect the final performance, a claim we validate in our experiments. + +Attribute Searching. Inspired by DARTS [26], we introduce a differentiable attribute search method that learns to find representative attributes $v$ from the search space $\mathcal{V}$ , as shown in Fig. 4(b). To make the search space continuous, we relax the discrete attribute selection into a softmax- + +weighted sum over all $w$ possible candidates: + +$$ +f (x, v; \alpha , \theta) = \sum_ {i \in \mathcal {V}} \frac {\exp (\alpha_ {i})}{\sum_ {i ^ {\prime} \in \mathcal {V}} \exp (\alpha_ {i ^ {\prime}})} f (x, v _ {i}; \theta). \tag {8} +$$ + +where $\alpha_{i}$ represents the weight for attribute combination $v_{i}$ . The task of attribute search is thus reduced to learning the weight vector $\alpha$ for the candidate pool. + +After relaxation, our goal is to jointly learn the attribute weight $\alpha$ and soft prompt tokens $\theta$ . Following standard practice [26, 31, 54], we optimize the weights $\alpha$ by minimizing the validation loss $L_{val}$ , while the soft tokens are learned by minimizing the training loss $L_{train}$ . We employ an alternating algorithm [9, 22] to solve this bi-level optimization problem, alternating between these two subproblems: + +$$ +\hat {\alpha} = \arg \min _ {\alpha} L _ {v a l} (f (x, v; \alpha , \hat {\theta}), c), \tag {9} +$$ + +$$ +\hat {\theta} = \arg \min _ {\theta} L _ {t r a i n} (f (x, v; \hat {\alpha}, \theta), c). \tag {10} +$$ + +where $L_{train}$ and $L_{val}$ both use the cross-entropy loss function. After the search, the attribute combination with the highest weight $(\alpha_{i})$ is selected. + +Cost Analysis. Unlike traditional Neural Architecture Search (NAS) methods [7, 25, 36], which search computationally expensive network-level parameters, our approach focuses on a lightweight token-level search space. This design makes our method significantly more efficient than previous approaches. In practice, our search converges in approximately 5 epochs, often requiring less than 5 minutes + +
MethodAverageImageNetCaltech101OxfordPots
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (UCV 22)82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp (CVPR 22)80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
MaPLc (CVPR 23)82.2875.1478.5576.6670.5473.4797.7494.3696.0295.4397.7696.58
PromptSRC (ICCV 23)84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
ArGue (CVPR 24)83.6978.0780.7876.9272.0674.4198.4395.2096.7995.3697.9596.64
DePT (CVPR 24)83.6671.8277.2977.1370.1073.4598.3394.3396.2994.7097.6396.14
CoPrompt (ICLR 24)84.0077.2380.4877.6771.2774.3398.2794.9096.5595.6798.1096.87
PromptKD (CVPR 24)86.9680.7383.7380.8374.6677.6298.9196.6597.7796.3098.0197.15
CoOp + ATPrompt82.6868.0474.65 (+2.99)76.2770.6073.3397.9593.6395.7494.7796.5995.67
CoCoOp + ATPrompt81.6974.5477.95 (+2.12)76.4370.5073.3597.9695.2796.6095.4697.8996.66
MaPLc + ATPrompt82.9875.7679.21 (+0.66)76.9470.7273.7098.3295.0996.6895.6297.6396.61
DePT + ATPrompt83.8073.7578.45 (+1.16)77.3270.6573.8398.4894.6096.5094.6597.9996.29
PromptKD + ATPrompt87.0581.8284.35 (+0.62)80.9074.8377.7598.9096.5297.7096.9298.2797.59
+ +
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (JCV 22)78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp (CVPR 22)70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
MaPLe (CVPR 23)72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
PromptSRC (ICCV 23)78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
ArGue (CVPR 24)75.6473.3874.4998.3475.4185.3692.3391.9692.1440.4638.0339.21
DePT (CVPR 24)79.6772.4075.8698.2072.0083.0890.4391.3390.8842.5322.5329.46
CoPrompt (ICLR 24)76.9774.4075.6697.2776.6085.7190.7392.0791.4040.2039.3339.76
PromptKD (CVPR 24)82.8083.3783.1399.4282.6290.2492.4393.6893.0549.1241.8145.17
CoOp + ATPrompt77.4366.5571.5897.4467.5279.7788.7487.4488.0940.3827.2232.52
CoCoOp + ATPrompt74.5073.4773.9896.5273.5983.5190.5991.7491.1637.3033.1535.10
MaPLe + ATPrompt75.3973.8474.6197.8275.0784.9590.6592.0091.3237.6136.1536.87
DePT + ATPrompt79.2973.4776.2798.2073.6984.2090.4291.6991.0543.1933.2337.56
PromptKD + ATPrompt82.5184.0383.2699.1582.0389.7892.4893.8693.2249.6342.3545.70
+ +
MethodSUN397DTDEuroSATUCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (IUCV 22)80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp (CVPR 22)79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
MaPLe (CVPR 23)80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
PromptSRC (ICCV 23)82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
ArGue (CVPR 24)81.5280.7481.1381.6066.5573.3194.4388.2491.2385.5679.2982.31
DePT (CVPR 24)82.3775.0778.5583.2056.1367.0488.2766.2775.7085.4372.1778.24
CoPrompt (ICLR 24)82.6380.0381.3083.1364.7372.7994.6078.5785.8486.9079.5783.07
PromptKD (CVPR 24)83.6981.5482.6085.8471.3777.9497.5482.0889.1489.7182.2786.10
CoOp + ATPrompt80.8468.6474.2480.8345.4958.2290.3459.7971.9684.4964.9673.45
CoCoOp + ATPrompt80.5076.8678.6478.6356.8966.0287.9574.1580.4682.7476.4079.44
MaPLe + ATPrompt80.9878.1579.5480.5058.2867.6194.8477.5985.3584.0878.8881.40
DePT + ATPrompt82.4276.4879.3482.6456.7767.3089.6069.5078.2885.6073.1578.89
PromptKD + ATPrompt83.8781.3582.5986.9272.3478.9697.0592.0794.4989.2982.4485.73
+ +Table 1. Base-to-novel generalization experiments of five baselines with and without our ATPrompt on 11 datasets. Our method achieves consistent average performance improvement over different baselines. + +
MethodSourceTarget DatasetAverage
Image NetCaltech 101Oxford PetsStanford CarsFlowers 102Food101FGVC AircraftSUN397DTDEuro SATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
+ATPrompt71.6793.9690.6565.0170.4085.8620.9765.7743.4446.5969.9265.26 (+1.38)
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
+ATPrompt71.2793.7990.6265.9071.1786.0323.2266.6344.4448.7070.7166.59 (+0.85)
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
+ATPrompt70.6994.0491.0366.0671.9986.3324.4267.0545.2148.6369.1566.75 (+0.45)
+ +Table 2. Cross-dataset generalization experiments of three baselines with and without our ATPrompt on 11 datasets. ATPrompt achieves consistent average performance improvement on target datasets. + +
DatasetAttribute BasesSearched Results
ImageNetcolor, size, shape, habitat, behavior(color, shape)
Caltech101shape, color, material, function, size(shape,size)
OxfordPetsloyalty, affection, energy, playfulness, intelligence(playfulness, energy)
StanfordCarsdesign, engine, performance, luxury, color(luxury)
Flowers102color, flower, habitat, growth, season(color, habitat, growth)
Food101flavor, texture, origin, ingredients, preparation(flavor, preparation)
+ +on a single A800 GPU for some datasets. In addition, search efficiency can be improved by curating a smaller set of base attributes, which narrows the search space. + +# 4. Experiments + +# 4.1. Settings + +Base-to-Novel Generalization. Following [14, 23, 51, 52], we split the dataset into base and novel classes. The model is trained on the base class training set and evaluated on the test set. The attributes applied in this experiment are searched based on the base class. For datasets without a dedicated validation set, like ImageNet, we split the 16-shot labeled data, using half for training and the other half for attribute search. + +Cross-dataset Experiments. Consistent with previous works [14, 51, 52], we first train a model on the ImageNet-1K source dataset and then evaluate its generalization performance on several out-of-distribution datasets. The attributes used are obtained from the source dataset. + +Attribute Search. We select five independent attributes as the basis in the attribute pool. This results in 31 candidate attribute combinations for the search process. We use ChatGPT-4o for attribute queries. In Tab. 3 reports a subset of the queried attribute bases and the final combinations selected by our search. + +Implementation Details. We evaluate the model performance on 15 datasets. We report base and novel class accuracy and their Harmonic Mean (HM) averaged over 3 runs. The details of each dataset are attached in the Appendix. + +# 4.2. Base-to-Novel Generalization + +As demonstrated in Tab. 1, we evaluate the base-to-novel generalization performance of five baseline methods, both with and without integrating ATPrompt, across 11 diverse recognition datasets. Notably, ATPrompt consistently improves the average performance of all baseline methods. + +Reasons for Limited Improvement in Some Conditions. (1) Learnable text prompts are the central component in earlier works, and optimizing them through ATPrompt can obviously improve the performance. (2) Recent studies have expanded beyond learnable text prompts by introducing additional learnable modules. Since our work only involves the optimization of the learnable text prompt part, the improvement have become less obvious. + +# 4.3. Cross-dataset Evaluation + +Tab. 2 shows the cross-dataset generalization results for three baseline methods. Our method demonstrates superior performance, with improvements of $1.38\%$ , $0.85\%$ and $0.45\%$ for CoOp, CoCoOp and MaPLe, respectively. + +# 4.4. Domain Generalization + +Tab. 4 shows the domain generalization results for three baseline methods. The results show that our method improves CoOp, CoCoOp and MaPLe methods by $0.90\%$ , $0.49\%$ and $0.33\%$ respectively. + +Table 3. Part of the results obtained after differentiable attribute search. Please refer to the appendix for the complete results. + +
MethodSourceTarget DatasetAverage
ImageNet-V2-S-A-R
CoOp71.5164.2047.9949.7175.2159.28
+ATPrompt71.6764.4349.1350.9176.2460.18 (+0.90)
CoCoOp71.0264.0748.7550.6376.1859.91
+ATPrompt71.2764.6649.1551.4476.3360.40 (+0.49)
MaPLe70.7264.0749.1550.9076.9860.27
+ATPrompt70.6964.4049.1051.7777.1160.60 (+0.33)
+ +Table 4. Domain generalization experiments of three baselines with and without our ATPrompt on four datasets. The integration of ATPrompt resulted in better generalization performance. + +# 4.5. Further Analysis + +By default, the experiments are conducted on the ImageNet. To minimize the influence of other components, we mainly adopt CoOp as the baseline method. Two attributes (color and shape) are used in our ATPrompt. + +![](images/7d195a143a3ea8c18ad49ef0f1be6805308cb3ced856a3581842236793d708d6.jpg) +Figure 5. Illustration of varying soft token length on ImageNet. Increased tokens can lead to overfitting to the base class and weaken generalization to the novel class. + +Soft Prompt Length. In Fig. 5, we examine the optimal soft token length for both attribute and class tokens. By varying the length from 1 to 10, we observe that longer prompts dilute the guiding influence of the attribute tokens, thereby reducing generalization to novel classes. + +Class Token Position. The relative positioning of attribute tokens and class tokens in our method requires careful consideration. In Tab. 5, we examine configurations where the class token is positioned in the middle or on either side of two attribute tokens. The results demonstrate that optimal performance is achieved when the class token is placed at the end, which aligns with the findings of CoOp. + +
PositionBaseNovelHM
Front76.1270.5073.20
Middle76.1370.2973.09
End76.2770.6073.33
+ +Prompt Operation of Deep Version. In ATPrompt-Deep, we exclusively drop class soft tokens while retaining both hard and soft attribute tokens after they pass through the block. In this part, we compare the performance of partial drop (i.e., removing attribute soft tokens while retaining hard tokens) and full drop (i.e., removing both attribute soft and hard tokens) operations, as illustrated in Tab. 6. + +Table 5. Comparison of different class token positions on ImageNet. The end position works best. + +
Operation of Attribute TokenBaseNovelHM
Retain all hard and soft tokens76.9470.7273.70
Partial drop and re-add76.8770.4473.51
Full drop and re-add76.8370.1073.31
+ +The results show that maintaining the attributes of both hard and soft tokens during the forward process results in optimal performance. Conversely, dropping and reintroducing either hard or soft attribute tokens harms performance, likely because it disrupts the continuity of attribute representations across layers and complicates optimization. + +Attribute Order. In this study, we do not specifically focus on the order of attributes because varying the sequence usually does not result in semantic deviations in reality. Tab. 7 quantitatively assesses the impact of attribute order on prompt learning performance. From this table, we observe that despite variations in order, similar results are consistently produced, and the performance fluctuations across different orders remain within a reasonable range. + +Comparison to Other Attributes. In Tab. 8, we explore the effectiveness of attributes derived through alternative methods, specifically by manually selecting class-irrelevant + +Table 6. Comparison of operations on deep soft and hard attribute tokens based on MaPLe+ATPrompt. Preserving hard and soft attribute tokens in deep layers performs better than other operations. + +
AttributesBaseNovelHM
(shape, color)76.3270.3973.24
(color, shape)76.2770.6073.33
(size, habitat)76.4470.2373.20
(habitat, size)76.4670.1673.14
+ +and common attributes. It shows that manually selected irrelevant attributes exhibit comparable performance during training; however, they perform poorly when applied to new categories. This suggests that incorrect attribute tokens cause the soft tokens to develop biased representations, thereby diminishing their zero-shot generalization ability. + +Table 7. Comparison of different orders on ImageNet. The order of attributes does not significantly affect the model, and performance fluctuations are within a reasonable range. + +
TypeAttributesBaseNovelHM
Common(shape, size)82.8367.1374.16
(color, texture)82.7367.5674.38
Irrelevant( plane, engines)82.8166.2273.59
(football, sport)82.7767.1474.14
Searched-82.6868.0474.65
+ +Table 8. Comparison of average performance for various attribute configurations on 11 datasets. The attributes obtained by our method achieve the best performance. + +# 5. Conclusion + +In this work, we introduce ATPrompt, an attribute-anchored textual prompt learning method that uses universal attributes as a bridge to improve generalization from seen to unseen categories. Our approach expands the learning space of soft prompts from a one-dimensional, category-centric structure to a multi-dimensional attribute space by anchoring fixed attribute tokens with the prompt. To ensure the selection of optimal attributes, we propose an automated pipeline designed to identify the most suitable candidates for any given downstream task. ATPrompt is designed with both shallow and deep architectural variants, rendering it broadly compatible with existing prompt learning methods. Extensive experiments validate the effectiveness of our approach. We believe this work offers a new direction for research into the fundamental structure of learnable prompts in prompt learning area. + +Limitations and future works. This work is a preliminary study of the basic prompt form, which has not been able to achieve comparable performance to regularization-based methods when working alone. Furthermore, the current selection of attribute anchors relies on manual experimentation, and automatically discovering the optimal anchor positions through a learning-based approach remains a promising direction for future research. + +# 6. Acknowledgments + +This work was supported by the National Natural Science Foundation of China under Grant Nos. 62361166670 and U24A20330. This research was also supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206134), the Fundamental Research Funds for the Central Universities 070-63253222 and the Tianjin Key Laboratory of Visual Computing and Intelligent Perception (VCIP). Computation is supported by the Supercomputing Center of Nankai University (NKSC). + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 1 +[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 35:25005-25017, 2022. 1 +[3] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1 +[4] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, pages 23518-23527, 2023. 2 +[5] Tong Ding, Wanhua Li, Zhongqi Miao, and Hanspeter Pfister. Tree of attributes prompt learning for vision-language models. arXiv preprint arXiv:2410.11201, 2024. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[7] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. JMLR, 20(55):1-21, 2019. 5 +[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[9] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. NeurIPS, 33:5632-5643, 2020. 5 +[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2 +[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 1, 4 + +[12] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 2 +[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 4 +[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 7 +[15] Gahyeon Kim, Sohee Kim, and Seokju Lee. Aapl: Adding attributes to prompt learning for vision-language models. In CVPR Workshop, pages 1572-1582, 2024. 2 +[16] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 1 +[17] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 2 +[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2 +[19] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 1, 2 +[20] Yunheng Li, Yuxuan Li, Quansheng Zeng, Wenhai Wang, Qibin Hou, and Ming-Ming Cheng. Unbiased region-language alignment for open-vocabulary dense prediction. arXiv preprint arXiv:2412.06244, 2024. 1 +[21] Yunheng Li, Zhong-Yu Li, Quan-Sheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-CLIP: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243-28258. PMLR, 2024. 1 +[22] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 5 +[23] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. In CVPR, pages 26617-26626, 2024. 1, 2, 4, 7 +[24] Zheng Li, Xiang Li, Lingfeng Yang, Renjie Song, Jian Yang, and Zhigeng Pan. Dual teachers for self-knowledge distillation. Pattern Recognition, 151:110422, 2024. 2 +[25] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 5 +[26] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 5 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3 +[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 1 +[29] Hao Ma, Ming Li, Jingyuan Yang, Or Patashnik, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Clip-flow: Decoding images encoded in clip space. CVMJ, 10(6):1157-1168, 2024. 1 +[30] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2 +[31] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 5 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3 +[33] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 2 +[34] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, pages 13019-13029, 2024. 1 +[35] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1 +[36] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 5 +[37] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3 +[38] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757, 2024. 2 +[39] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5 +[40] Ge Wu, Xin Zhang, Zheng Li, Zhaowei Chen, Jiajun Liang, Jian Yang, and Xiang Li. Cascade prompt learning for vision-language model adaptation. In ECCV, 2024. 1 +[41] Shihan Wu, Ji Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves. In CVPR, pages 14723-14732, 2025. 2 +[42] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, and Yongjun Xu. Clip-kd: An empirical study of distilling clip models. arXiv preprint arXiv:2307.12732, 2023. 2 + +[43] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 2 +[44] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 2, 4 +[45] Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, and Da Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In AAAI, pages 6979–6987, 2024. 2 +[46] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In CVPR, pages 12924-12933, 2024. 1 +[47] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, et al. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. 1 +[48] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 2 +[49] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025.3 +[50] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 5 +[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 4, 7 +[52] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 7 +[53] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, pages 15659-15669, 2023. 2, 3 +[54] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 5 \ No newline at end of file diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/images.zip b/2025/Advancing Textual Prompt Learning with Anchored Attributes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..966057c859b1e550f75326c45eb8efad4560bb21 --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6852f8830f37b9a598bbdeea583f24f5b4200f29ade1ce13eed1a9612ec626ed +size 876857 diff --git a/2025/Advancing Textual Prompt Learning with Anchored Attributes/layout.json b/2025/Advancing Textual Prompt Learning with Anchored Attributes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad1f9fdcd028be6ce97bb812c66e72b9796290d2 --- /dev/null +++ b/2025/Advancing Textual Prompt Learning with Anchored Attributes/layout.json @@ -0,0 +1,8018 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "text", + "content": "Advancing Textual Prompt Learning with Anchored Attributes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": "Zheng Li" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": ", Yibing Song" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": ", Ming-Ming Cheng" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": ", Xiang Li" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": " PCA Lab, VCIP, College of Computer Science, Nankai University, " + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 475, + 186 + ], + "type": "text", + "content": " DAMO Academy, Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 161, + 187, + 449, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 187, + 449, + 213 + ], + "spans": [ + { + "bbox": [ + 161, + 187, + 449, + 213 + ], + "type": "text", + "content": "zhengli97@mail.nankai.edu.cn, yibingsong.csv@gmail.com {cmm, xiang.li.implus, csjyang}@nankai.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 267, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 296, + 579 + ], + "type": "text", + "content": "Textual-based prompt learning methods primarily employ multiple learnable soft prompts and hard class tokens in a cascading manner as text inputs, aiming to align image and text (category) spaces for downstream tasks. However, current training is restricted to aligning images with predefined known categories and cannot be associated with unknown categories. In this work, we propose utilizing universal attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an Attribute-anchored Textual Prompt learning method for vision-language models, named ATPrompt. This approach expands the learning space of soft prompts from the original one-dimensional category level into the multi-dimensional attribute level by incorporating multiple attribute tokens into the learnable soft prompts. Through this modification, we transform the text prompt from a category-centric form to an attribute-category hybrid form. Additionally, we introduce a straightforward differentiable attribute search method to identify representative and suitable attributes for downstream tasks. As an easy-to-use plug-in technique, ATPrompt can seamlessly replace the existing basic prompt format in textual-based methods, providing general improvements at a negligible computational cost. Extensive experiments across 11 datasets validate the effectiveness of our method. Code is publicly available at https://github.com/zhengli97/ATPrompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 694 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) [10, 20, 21, 28, 29, 32, 34, 35], such as CLIP [3, 32] and ALIGN [10], have demonstrated exceptional performance in recent years. These models are trained with a contrastive loss to establish alignment between image and text (category) space. Inspired by the success of NLP [18, 19], prompt learning [52] has" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 340, + 243, + 521, + 361 + ], + "blocks": [ + { + "bbox": [ + 340, + 243, + 521, + 361 + ], + "lines": [ + { + "bbox": [ + 340, + 243, + 521, + 361 + ], + "spans": [ + { + "bbox": [ + 340, + 243, + 521, + 361 + ], + "type": "image", + "image_path": "404ad711120a7070a0481b67a504644d00bcebcafb9099f47527b71126fd0ab6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 365, + 520, + 509 + ], + "blocks": [ + { + "bbox": [ + 345, + 365, + 520, + 509 + ], + "lines": [ + { + "bbox": [ + 345, + 365, + 520, + 509 + ], + "spans": [ + { + "bbox": [ + 345, + 365, + 520, + 509 + ], + "type": "image", + "image_path": "dc204e01d93682cd3d39978bd1666f2a400475d7802824a1f0afd311d5c2603e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 516, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 583 + ], + "type": "text", + "content": "Figure 1. Comparison of image and text (category) alignment processes through learnable prompts. (a) Current prompt learning methods align images with predefined categories but fail to establish accurate associations with unknown categories. (b) ATPrompt leverages universal attributes as an intermediary to create more accurate alignments between images and unknown categories." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "emerged as a parameter-efficient tool to adapt powerful VLMs to downstream tasks. Models with a few learnable soft prompt tokens [11, 47, 51] can achieve performance parity with, or even outperform, fully fine-tuned ones [11]. Depending on how the soft prompt tokens are applied, existing methods can be broadly classified into textual-based [13, 14, 23, 40, 51, 52] and visual-based approaches [1, 2, 11, 16, 46]. Among these, the textual-based method is the most fundamental and straightforward, comprising the majority." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3618" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "In typical image classification tasks, current text-based methods [13, 14, 51, 52] predominantly employ the traditional approach of concatenating learnable soft prompts with hard class tokens to replace handcrafted text prompts (e.g., \"a photo of a {class}\") as inputs to the encoder. Although this text prompt demonstrates strong performance, it restricts image alignment during training to predefined known categories only, thereby preventing accurate associations with unknown categories, as shown in Fig. 1(a). Intuitively, when confronted with an unfamiliar category, humans often associate it with additional attributes (e.g., color, shape or texture) to increase comprehensibility and clarity, rather than merely stating the object's name. For instance, one might describe a cheetah as: \"The cheetah is a cat-like animal with a small head, short yellow hair, and black spots.\" or refer to an apple as \"That red spherical fruit with orange stripes is an apple.\" instead of using a general description such as \"This is a cheetah.\" or \"That fruit is an apple.\" Attributes can serve as bridges that connect unknown categories to our known knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 314, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 295, + 601 + ], + "type": "text", + "content": "Building on these observations, we propose a novel approach that leverages attributes as a bridge to enhance the alignment between images and unknown categories. Specifically, we introduce an attribute-anchored textual prompt learning method for VLMs, named ATPrompt. This method extends the learning space of soft prompts from the original one-dimensional category level to a multidimensional attribute level by integrating multiple fixed universal attribute tokens into the learnable soft prompts. Guided by these anchored attributes, soft tokens acquire not only category-specific but also attribute-related general representations during training. This results in improved alignment between images and unknown categories compared to the original method, as shown in Fig. 1(b). Additionally, based on the depth at which soft prompts are applied, we propose two versions of ATPrompt: shallow and deep, respectively, ensuring compatibility with existing methods of varying depths [13, 14, 52]. To finalize these attributes, we present a simple and effective differentiable attribute search method that learns to identify suitable attributes from a candidate pool constructed by LLMs. The search operation only needs to be performed once per task, and once completed, the selected attributes can be used by ATPrompt for model training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 602, + 294, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 294, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 294, + 650 + ], + "type": "text", + "content": "As an easy-to-use plug-in technique, ATPrompt can seamlessly substitute existing forms used in textual-based prompt learning methods, yielding general improvements with negligible additional computational overhead." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 712 + ], + "type": "text", + "content": "- We propose an efficient attribute-anchored textual prompt learning method that expands the learning space of soft prompts from a one-dimensional class level to the multidimensional attribute level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 95 + ], + "type": "text", + "content": "- We introduce an effective differentiable search method to select appropriate attributes for downstream tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 96, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 132 + ], + "type": "text", + "content": "- Both shallow and deep versions of ATPrompt are introduced to ensure compatibility with existing prompt learning methods of varying depths." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 179 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that ATPrompt can be seamlessly integrated into existing textual-based methods, resulting in consistently improved performance with negligible computational costs." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 190, + 400, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 400, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 400, + 204 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 211, + 553, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 558 + ], + "type": "text", + "content": "Prompt Learning for VLMs. Inspired by recent advancements in NLP [18, 19], prompt learning [13, 14, 23, 33, 48, 51-53] has garnered significant interest among vision researchers aiming to apply these techniques to VLMs [10, 32, 42], such as CLIP. CoOp [52] is the pioneering text-based approach that introduced the concept of using a combination of soft textual tokens and a hard class token as input. Subsequent studies [13, 14, 17, 23, 41, 44, 51] have predominantly followed this textual prompt format. However, this form constrains the soft prompts to align with images within a one-dimensional, predefined category space, limiting their applicability to unknown categories. Therefore, training based on the current text form will be more likely to overfit to known categories, diminishing their zero-shot generalization capability for unknown categories. To address this limitation, several methods have been proposed [12, 14, 23, 44, 53]. For instance, Kg-CoOp [44] uses hand-crafted hard prompts to regularize learnable soft prompts during training. PromptSRC [14] utilizes CLIP's [32] original features to regularize the learning of soft prompts for image and text branches. PromptKD [23] utilizes a pre-trained strong teacher model to guide the learning of a student model [8, 22, 24, 43] with learnable prompts. Despite these advancements, none of the above methods address the inherent limitations of the format itself. In this work, we introduce the attribute-anchored textual prompt format for VLMs, which proposes to utilize attributes as a bridge to build more accurate associations between images and unknown categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Attributes for VLMs. In practice, categories typically encompass multiple attributes. When individuals encounter an unfamiliar category, they often describe it using additional attributes to enhance the clarity of their communication, rather than merely stating its name. Inspired by this observation, numerous studies [4, 30, 37, 38, 45] have begun to leverage attributes to support their objectives. VCD [30] was the pioneering work to propose the use of LLMs to decompose class names into multiple in-class attributes (e.g., beak and tail for birds) for classification. AAPL [15] introduces a meta-network to extract visual attribute features based on encoded image features, facilitating image-text alignment. TAP [5] presents a structured \"Tree of At" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3619" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 162, + 171 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 162, + 171 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 162, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 162, + 171 + ], + "type": "image", + "image_path": "b1d40b41b880b211d87c229cda2c4c252092476e1cbfc4054c41aceab0af5c0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 176, + 555, + 254 + ], + "lines": [ + { + "bbox": [ + 54, + 176, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 54, + 176, + 555, + 254 + ], + "type": "text", + "content": "Figure 2. Architectural comparison among existing methods. (a) Vanilla CLIP employs a hand-crafted text template as input to the text encoder. (b) Classical prompt learning proposes a new text form that concatenates multiple learnable soft tokens with class tokens. (c) ArGue [37] employs multiple in-class attributes mined by LLMs as supplementary information. These attributes are utilized to construct distinct text groups, which serve as learning targets, thereby regularizing the learning of soft tokens. The final prediction is achieved by assembling all groups. (d) Our ATPrompt treats universal attributes as learning components and anchors them into existing soft prompt templates. Though this operation, we expand the learning space of soft tokens to multi-dimensional attribute levels and facilitate the alignment of images with unknown class texts." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 73, + 277, + 171 + ], + "blocks": [ + { + "bbox": [ + 170, + 73, + 277, + 171 + ], + "lines": [ + { + "bbox": [ + 170, + 73, + 277, + 171 + ], + "spans": [ + { + "bbox": [ + 170, + 73, + 277, + 171 + ], + "type": "image", + "image_path": "597a08010caa7fab85fbd011e3d4b0549890acd44cd4ecfe046c3dea2f263c3b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 280, + 72, + 405, + 171 + ], + "blocks": [ + { + "bbox": [ + 280, + 72, + 405, + 171 + ], + "lines": [ + { + "bbox": [ + 280, + 72, + 405, + 171 + ], + "spans": [ + { + "bbox": [ + 280, + 72, + 405, + 171 + ], + "type": "image", + "image_path": "f1c2500e0697a38ef1eca8915369c9e894a2ce207cc27b37fd466af3bf5db66a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 415, + 72, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 415, + 72, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 415, + 72, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 415, + 72, + 547, + 171 + ], + "type": "image", + "image_path": "540938326c885781b22f84838f8676c2e0ff0eacc8034305263230f84f9a2dbb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 264, + 294, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 294, + 408 + ], + "type": "text", + "content": "attributes\" approach that leverages attribute-specific knowledge graphs to enhance VLMs. ArGue [37] leverages large language models to mine multiple in-class attributes and integrates them into soft prompts and fixed templates to create multiple text groups. It employs the original text features generated by the fixed template to regularize the learning of soft tokens, as illustrated in Fig. 2(c). Most prior studies have focused on utilizing in-class attributes to enhance model performance by providing supplementary attribute information. However, when dealing with an unknown class, reacquiring the attributes of the new class becomes necessary—a process that is both complex and costly." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 408, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 529 + ], + "type": "text", + "content": "In this work, we believe that universal (inter-class) attributes are more efficient and robust than the in-class attributes used in previous works. Instead of taking attributes as the learning objective, we treat it as a learning component and propose to anchor universal attributes into soft prompt templates, transforming the existing class-centric form [52] into a hybrid attribute-class learnable textual prompt form. Our approach can be seamlessly integrated into existing textual-based methods, enhancing their performance without incurring additional computational costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 111, + 550 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Prompt learning [13, 14, 51-53] aims to enhance the generalization ability of pre-trained VLMs like CLIP on downstream tasks by training inserted learnable soft tokens. Existing textual-based methods all follow the classic prompt paradigm, concatenating soft prompt tokens and hard class tokens as the input to the text encoder, as shown in Fig. 2(b). In this paper, we propose a simple and effective textual prompt learning method, named ATPrompt, which anchors multiple fixed universal attribute tokens into the original soft prompts, as shown in Fig. 2(d). Guided by these attributes, soft prompts can learn not only category-specific but also attribute-related general representations through training. When encountering unknown categories, these" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 408 + ], + "type": "text", + "content": "learned attribute-related tokens can provide additional information to promote better image-text alignment. Furthermore, to identify these universal attributes, we present an automated pipeline that encompasses sequential steps. Initially, we employ LLMs to synthesize the attribute pool for the current downstream category. Subsequently, we propose a differentiable attribute search method designed to identify attributes within the pool that are most suitable for our attribute-anchored prompt forms. For every task, this search operation is conducted only once. Once the attributes are finalized, they are integrated into our ATPrompt for specialized model fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 418, + 394, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 394, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 394, + 431 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": "Vision-Language Models. Existing VLMs, such as CLIP [32], have demonstrated remarkable zero-shot generalization performance after training with 400 million image-text pairs. The primary objective of these models is to learn the alignment between image and text modalities produced by each encoder [6, 27, 49]. Given a labeled image classification dataset " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathrm{D} = \\{(x,c)\\}" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " which includes " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " class labels " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "C = \\{c_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": ", CLIP makes predictions by calculating the cosine similarity between image features and the text features of each class. Specifically, for each input image " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": ", it undergoes feature extraction via the image encoder " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "h_I(x)" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " and obtains a feature vector " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "u = h_I(x)" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": ". Simultaneously, for each class, a series of textual descriptions " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " are generated using the hand-crafted template. Then, these text descriptions are fed into the text encoder " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "h_T(x)" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " to obtain text features " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "w = h_{T}(t)" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": ". Finally, the output probability for image " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " classified to " + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 436, + 555, + 641 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 651, + 553, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 651, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 359, + 651, + 553, + 680 + ], + "type": "interline_equation", + "content": "p (c | x) = \\frac {\\exp \\left(\\cos \\left(u , w _ {c}\\right) / \\tau\\right)}{\\sum_ {i = 1} ^ {N} \\exp \\left(\\cos \\left(u , w _ {i}\\right) / \\tau\\right)}. \\tag {1}", + "image_path": "d29233c139a4596b5cbdb389d1e92e0007b64e5dff192cde2682bdda194c51bc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the temperature parameter and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " denotes cosine similarity." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3620" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 71, + 272, + 183 + ], + "blocks": [ + { + "bbox": [ + 94, + 71, + 272, + 183 + ], + "lines": [ + { + "bbox": [ + 94, + 71, + 272, + 183 + ], + "spans": [ + { + "bbox": [ + 94, + 71, + 272, + 183 + ], + "type": "image", + "image_path": "2f77496042a509cb5255fb771dda64d1fcce2ecffd22ac7cf5c3499613787eff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 186, + 243, + 196 + ], + "lines": [ + { + "bbox": [ + 175, + 186, + 243, + 196 + ], + "spans": [ + { + "bbox": [ + 175, + 186, + 243, + 196 + ], + "type": "text", + "content": "(a) Shallow Version" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 292, + 74, + 432, + 183 + ], + "blocks": [ + { + "bbox": [ + 292, + 74, + 432, + 183 + ], + "lines": [ + { + "bbox": [ + 292, + 74, + 432, + 183 + ], + "spans": [ + { + "bbox": [ + 292, + 74, + 432, + 183 + ], + "type": "image", + "image_path": "46cb0444479a79004dbd45c76924ac713536483e63109a0b0f71e2f0fbf63e3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 186, + 399, + 196 + ], + "lines": [ + { + "bbox": [ + 340, + 186, + 399, + 196 + ], + "spans": [ + { + "bbox": [ + 340, + 186, + 399, + 196 + ], + "type": "text", + "content": "(b) Deep Version" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 202, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 258 + ], + "type": "text", + "content": "Figure 3. An illustration of the computation process for shallow and deep versions. Take two attributes [A] and [B] as examples. (a) The shallow version concatenates hard attribute tokens, soft prompt tokens, and class tokens and inputs them into the encoder for calculation. (b) The deep version uses the same input but discards the class-related soft prompt tokens after calculating the self-attention and introduces them again before the next layer. These two forms can be compatible with existing methods of varying prompt depths, including input-level ones like CoOp [52], CoCoOp [51] and depth-level ones like MaPLe [13]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 444, + 99, + 516, + 183 + ], + "blocks": [ + { + "bbox": [ + 444, + 99, + 516, + 183 + ], + "lines": [ + { + "bbox": [ + 444, + 99, + 516, + 183 + ], + "spans": [ + { + "bbox": [ + 444, + 99, + 516, + 183 + ], + "type": "image", + "image_path": "24b9b8627f22e44be57067d982e180ebfec30d25f3bdc9419e909b7f4ccf3eff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "text", + "content": "Prompt Learning for VLMs. Instead of manually designed hard prompts for image-text alignment, which is inaccurate and inflexible, recent prompt learning works [13, 14, 44, 51, 52] like CoOp propose to learn appropriate soft textual prompts for downstream tasks. Concretely, " + }, + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "text", + "content": " learnable soft tokens " + }, + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "inline_equation", + "content": "[T_i]_{i=1}^M" + }, + { + "bbox": [ + 55, + 270, + 296, + 366 + ], + "type": "text", + "content": " are concatenated with the hard class token [CLS] as the input of the text encoder, as shown in Fig. 2(b). Its form is shown as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 373, + 294, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 373, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 116, + 373, + 294, + 387 + ], + "type": "interline_equation", + "content": "P _ {T} = \\left[ T _ {1} \\right] \\left[ T _ {2} \\right] \\dots \\left[ T _ {M} \\right] [ \\text {C L S} ], \\tag {2}", + "image_path": "654f63b402accdf4fd597d51ac0cde430e587cc8a0da1601e7d3846c1c39d612.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 392, + 295, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 392, + 295, + 416 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 392, + 295, + 416 + ], + "type": "text", + "content": " represents the length of soft tokens. For simplicity, we omit the prefix and suffix tokens in the input." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "content": "In addition to embedding soft tokens at the input level, existing studies [11, 13, 14, 23] have also explored introducing them at deeper layers. This is achieved by adding soft tokens within the Transformer blocks and subsequently removing them after the self-attention computations. For the " + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 417, + 296, + 489 + ], + "type": "text", + "content": "-th block, this process can be described as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 495, + 295, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 495, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 111, + 495, + 295, + 509 + ], + "type": "interline_equation", + "content": "\\left[ \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {3}", + "image_path": "fc795c8e30d5b3acc2da41a403e2a9849e5acd92f6f6127932e8898ad95372fd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "content": "-th transformer block and " + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_i" + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "content": " denotes the set of learnable soft tokens, defined as " + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_i = \\{[T_1]_i,\\dots,[T_M]_i\\}" + }, + { + "bbox": [ + 55, + 514, + 296, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 557, + 296, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 571 + ], + "type": "text", + "content": "3.2. Learning soft prompts with universal attributes" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 574, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 610 + ], + "type": "text", + "content": "Our approach introduces two variants, distinguished by the number of layers at which soft tokens are applied: a shallow version and a deep version, as shown in Fig. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 611, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 682 + ], + "type": "text", + "content": "Shallow Version. We begin by introducing the shallow version, in which hard attribute tokens are anchored solely at the input level, as illustrated in Fig. 3(a). Consider two universal attributes, A and B. According to Eqn. (2), the shallow-level text prompt " + }, + { + "bbox": [ + 55, + 611, + 296, + 682 + ], + "type": "inline_equation", + "content": "P_T" + }, + { + "bbox": [ + 55, + 611, + 296, + 682 + ], + "type": "text", + "content": " provided to the text encoder can be expressed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 689, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 61, + 689, + 295, + 714 + ], + "type": "interline_equation", + "content": "P _ {T} = \\left[ T _ {a _ {1}} \\right] \\dots \\left[ T _ {a _ {m}} \\right] \\left[ A \\right] \\left[ T _ {b _ {1}} \\right] \\dots \\left[ T _ {b _ {m}} \\right] \\left[ B \\right] \\left[ T _ {1} \\right] \\dots \\left[ T _ {M} \\right] \\left[ C L S \\right]. \\tag {4}", + "image_path": "3671d7fa97b1dcfdaf0f17ddf24d0869267f55d162089acb22cfb1686950627d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "inline_equation", + "content": "a_{m}" + }, + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "inline_equation", + "content": "b_{m}" + }, + { + "bbox": [ + 313, + 270, + 555, + 306 + ], + "type": "text", + "content": " are hyperparameters specifying the length of soft tokens for attributes A and B. In our method, we set these parameters to be the same by default." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "type": "text", + "content": "While this example places the class token at the end, we also evaluate configurations where it is positioned at the front or in the middle. As shown in Tab. 5, the end position yields the best performance and is therefore adopted as our default configuration." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "text", + "content": "Deep Version. In this version, learnable soft tokens are introduced at the input of the deep layers. Previous works, such as VPT and MaPLe, discard all soft tokens and subsequently reintroduce them after the block. When this operation is applied to attribute-related words, a gap will emerge between the introduced excessive low-level tokens and the existing high-level tokens, thereby weakening the feature continuity across layers. In this study, our approach selectively discards and then re-adds only class-related soft tokens in the input, specifically " + }, + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "inline_equation", + "content": "[T_1],\\dots,[T_M]" + }, + { + "bbox": [ + 313, + 367, + 555, + 510 + ], + "type": "text", + "content": ", as shown in Fig. 3(b). Based on Eqn. (3), the deep version of ATPrompt can be rewritten as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 514, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 514, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 329, + 514, + 553, + 529 + ], + "type": "interline_equation", + "content": "[ \\mathrm {F} _ {1}, \\rightarrow , \\mathrm {C L S} _ {1} ] = L _ {1} ([ \\mathrm {T} _ {a _ {0}}, \\mathrm {A}, \\mathrm {T} _ {b _ {0}}, \\mathrm {B}, \\mathrm {T} _ {0}, \\mathrm {C L S} _ {0} ]), \\tag {5}", + "image_path": "a5b736bf839ddfbe0adf25faabfa85ed635c226e528146387b1413ca8bfb5f2d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 531, + 553, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 531, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 347, + 531, + 553, + 551 + ], + "type": "interline_equation", + "content": "\\left[ \\mathrm {F} _ {i, -}, \\mathrm {C L S} _ {i} \\right] = L _ {i} \\left(\\left[ \\mathrm {F} _ {i - 1}, \\mathrm {T} _ {i - 1}, \\mathrm {C L S} _ {i - 1} \\right]\\right). \\tag {6}", + "image_path": "f66ab69bdd709d1aaa9f5fe5c716ca7891453bd39442142b7c981a9ac230bb8f.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 394, + 548, + 457, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 548, + 457, + 559 + ], + "spans": [ + { + "bbox": [ + 394, + 548, + 457, + 559 + ], + "type": "interline_equation", + "content": "i = 2, 3, \\dots , M.", + "image_path": "46bb18c672e7d277217d9ee1a480cc05c073e8b1905376e3dad43e70024eea01.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_i" + }, + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "text", + "content": " represents the features computed by the " + }, + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 562, + 554, + 597 + ], + "type": "text", + "content": "-th Transformer layer. We demonstrate the effectiveness of this operation in Tab. 6." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "text", + "content": "Training. Let " + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "text", + "content": " represent the weight of the total soft tokens and let " + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "text", + "content": " denote the selected fixed attribute tokens. Training is conducted on a labeled dataset " + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "inline_equation", + "content": "\\mathrm{D} = \\{(x,c)\\}" + }, + { + "bbox": [ + 313, + 597, + 555, + 668 + ], + "type": "text", + "content": ", with the objective of minimizing the cross-entropy loss between predicted values and ground truth labels. This process can be formulated as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 348, + 672, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 672, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 348, + 672, + 553, + 697 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} L _ {t r a i n} = \\min _ {\\theta} \\sum_ {x \\in D} \\operatorname {C E} (f (x; v, \\theta), c). \\tag {7}", + "image_path": "7ef73974f48f277b0c3c5507bc0f0b65eb31f44026e5269f7fe7a55fc1fe3363.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 701, + 533, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 533, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 533, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 533, + 714 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 701, + 533, + 714 + ], + "type": "text", + "content": " represents the function of the CLIP model." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3621" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 277 + ], + "type": "image", + "image_path": "87de5756e56376e383ee44596442002f12bec98cf0bd6d9b0208e909ac934773.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 278, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 278, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 555, + 335 + ], + "type": "text", + "content": "Figure 4. An overview of our attribute search pipeline. (a) We first query the LLM iteratively to obtain multiple independent attributes. These are subsequently aggregated to form a pool of candidate combinations, which serves as the input for the search process. (b) The forward computation for each candidate combination is represented by a distinct colored path. To identify the optimal attributes, we employ an alternating optimization algorithm that co-optimizes the soft tokens and a corresponding path weight vector. Upon completion of training, the combination associated with the highest-weighted path is selected as the final output." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 155, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 155, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 155, + 358 + ], + "type": "text", + "content": "3.3. Attribute search" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 365, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 472 + ], + "type": "text", + "content": "Selecting the attributes involves two key considerations: their content and their quantity. While directly querying a LLM is a straightforward approach, it has notable drawbacks. This method cannot determine the optimal number of attributes for a specific downstream dataset, and querying by category name alone can introduce semantic bias. To address these issues, we propose an automated pipeline that selects the appropriate content and quantity of attributes for the current downstream task, as illustrated in Fig. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "text", + "content": "Attribute Pool. Inspired by CoT [39, 50], we divide the entire process into multiple steps to enhance the reasoning ability of LLMs. First, we prompt the LLM to generate descriptive sentences for each known category, thereby enriching category-related information. Using these descriptions as context, we then prompt the LLM to summarize a set of independent attribute bases that are common across these categories. An attribute pool is then formed by creating all possible combinations of these bases, as shown in Fig. 4(a). For " + }, + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "text", + "content": " attribute bases, this result in a total of " + }, + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "inline_equation", + "content": "w = C_N^1 + C_N^2 + \\ldots + C_N^N" + }, + { + "bbox": [ + 55, + 473, + 295, + 652 + ], + "type": "text", + "content": " candidates in the pool, which constitutes our search space. Note that we do not consider the order of attributes, as permutations generally do not introduce significant semantic bias or affect the final performance, a claim we validate in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "text", + "content": "Attribute Searching. Inspired by DARTS [26], we introduce a differentiable attribute search method that learns to find representative attributes " + }, + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "text", + "content": " from the search space " + }, + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 653, + 295, + 712 + ], + "type": "text", + "content": ", as shown in Fig. 4(b). To make the search space continuous, we relax the discrete attribute selection into a softmax-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 347, + 498, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 498, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 498, + 359 + ], + "type": "text", + "content": "weighted sum over all " + }, + { + "bbox": [ + 313, + 347, + 498, + 359 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 347, + 498, + 359 + ], + "type": "text", + "content": " possible candidates:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 337, + 368, + 555, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 368, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 337, + 368, + 555, + 397 + ], + "type": "interline_equation", + "content": "f (x, v; \\alpha , \\theta) = \\sum_ {i \\in \\mathcal {V}} \\frac {\\exp (\\alpha_ {i})}{\\sum_ {i ^ {\\prime} \\in \\mathcal {V}} \\exp (\\alpha_ {i ^ {\\prime}})} f (x, v _ {i}; \\theta). \\tag {8}", + "image_path": "ebdc59364ceb7a3d47fcb06743ab09e0c1577d38b3cf2ddcd71d6fa129f675b2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "content": " represents the weight for attribute combination " + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "content": ". The task of attribute search is thus reduced to learning the weight vector " + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 406, + 553, + 441 + ], + "type": "text", + "content": " for the candidate pool." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": "After relaxation, our goal is to jointly learn the attribute weight " + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": " and soft prompt tokens " + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": ". Following standard practice [26, 31, 54], we optimize the weights " + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": " by minimizing the validation loss " + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "inline_equation", + "content": "L_{val}" + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": ", while the soft tokens are learned by minimizing the training loss " + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "inline_equation", + "content": "L_{train}" + }, + { + "bbox": [ + 313, + 441, + 554, + 536 + ], + "type": "text", + "content": ". We employ an alternating algorithm [9, 22] to solve this bi-level optimization problem, alternating between these two subproblems:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 361, + 548, + 555, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 548, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 361, + 548, + 555, + 566 + ], + "type": "interline_equation", + "content": "\\hat {\\alpha} = \\arg \\min _ {\\alpha} L _ {v a l} (f (x, v; \\alpha , \\hat {\\theta}), c), \\tag {9}", + "image_path": "cce17c19686ee11cf4bb2f25c6ad18a8d52c77228d1db1a1d30875d35ece6a59.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 571, + 553, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 571, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 358, + 571, + 553, + 589 + ], + "type": "interline_equation", + "content": "\\hat {\\theta} = \\arg \\min _ {\\theta} L _ {t r a i n} (f (x, v; \\hat {\\alpha}, \\theta), c). \\tag {10}", + "image_path": "96b0635de427c037b2c25ce40ed2b264263164cbe3aa285614ed0c3306200d72.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "inline_equation", + "content": "L_{train}" + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "inline_equation", + "content": "L_{val}" + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "content": " both use the cross-entropy loss function. After the search, the attribute combination with the highest weight " + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "inline_equation", + "content": "(\\alpha_{i})" + }, + { + "bbox": [ + 313, + 594, + 553, + 629 + ], + "type": "text", + "content": " is selected." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Cost Analysis. Unlike traditional Neural Architecture Search (NAS) methods [7, 25, 36], which search computationally expensive network-level parameters, our approach focuses on a lightweight token-level search space. This design makes our method significantly more efficient than previous approaches. In practice, our search converges in approximately 5 epochs, often requiring less than 5 minutes" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3622" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 515, + 224 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 515, + 224 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 515, + 224 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 515, + 224 + ], + "type": "table", + "html": "
MethodAverageImageNetCaltech101OxfordPots
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (UCV 22)82.6963.2271.6676.4767.8871.9298.0089.8193.7393.6795.2994.47
CoCoOp (CVPR 22)80.4771.6975.8375.9870.4373.1097.9693.8195.8495.2097.6996.43
MaPLc (CVPR 23)82.2875.1478.5576.6670.5473.4797.7494.3696.0295.4397.7696.58
PromptSRC (ICCV 23)84.2676.1079.9777.6070.7374.0198.1094.0396.0295.3397.3096.30
ArGue (CVPR 24)83.6978.0780.7876.9272.0674.4198.4395.2096.7995.3697.9596.64
DePT (CVPR 24)83.6671.8277.2977.1370.1073.4598.3394.3396.2994.7097.6396.14
CoPrompt (ICLR 24)84.0077.2380.4877.6771.2774.3398.2794.9096.5595.6798.1096.87
PromptKD (CVPR 24)86.9680.7383.7380.8374.6677.6298.9196.6597.7796.3098.0197.15
CoOp + ATPrompt82.6868.0474.65 (+2.99)76.2770.6073.3397.9593.6395.7494.7796.5995.67
CoCoOp + ATPrompt81.6974.5477.95 (+2.12)76.4370.5073.3597.9695.2796.6095.4697.8996.66
MaPLc + ATPrompt82.9875.7679.21 (+0.66)76.9470.7273.7098.3295.0996.6895.6297.6396.61
DePT + ATPrompt83.8073.7578.45 (+1.16)77.3270.6573.8398.4894.6096.5094.6597.9996.29
PromptKD + ATPrompt87.0581.8284.35 (+0.62)80.9074.8377.7598.9096.5297.7096.9298.2797.59
", + "image_path": "2e0d9f3ca23c08cad645adb7b4ba05e64709c4977a6e0994169a44181a00e5c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 102, + 228, + 507, + 381 + ], + "blocks": [ + { + "bbox": [ + 102, + 228, + 507, + 381 + ], + "lines": [ + { + "bbox": [ + 102, + 228, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 102, + 228, + 507, + 381 + ], + "type": "table", + "html": "
MethodStanfordCarsFlowers102Food101FGVCAircraft
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (JCV 22)78.1260.4068.1397.6059.6774.0688.3382.2685.1940.4422.3028.75
CoCoOp (CVPR 22)70.4973.5972.0194.8771.7581.7190.7091.2990.9933.4123.7127.74
MaPLe (CVPR 23)72.9474.0073.4795.9272.4682.5690.7192.0591.3837.4435.6136.50
PromptSRC (ICCV 23)78.2774.9776.5898.0776.5085.9590.6791.5391.1042.7337.8740.15
ArGue (CVPR 24)75.6473.3874.4998.3475.4185.3692.3391.9692.1440.4638.0339.21
DePT (CVPR 24)79.6772.4075.8698.2072.0083.0890.4391.3390.8842.5322.5329.46
CoPrompt (ICLR 24)76.9774.4075.6697.2776.6085.7190.7392.0791.4040.2039.3339.76
PromptKD (CVPR 24)82.8083.3783.1399.4282.6290.2492.4393.6893.0549.1241.8145.17
CoOp + ATPrompt77.4366.5571.5897.4467.5279.7788.7487.4488.0940.3827.2232.52
CoCoOp + ATPrompt74.5073.4773.9896.5273.5983.5190.5991.7491.1637.3033.1535.10
MaPLe + ATPrompt75.3973.8474.6197.8275.0784.9590.6592.0091.3237.6136.1536.87
DePT + ATPrompt79.2973.4776.2798.2073.6984.2090.4291.6991.0543.1933.2337.56
PromptKD + ATPrompt82.5184.0383.2699.1582.0389.7892.4893.8693.2249.6342.3545.70
", + "image_path": "225817b9a93d90998862d2e1c7069390a3523530d42fe02503432fcc3c476a6f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 103, + 385, + 507, + 540 + ], + "blocks": [ + { + "bbox": [ + 103, + 385, + 507, + 540 + ], + "lines": [ + { + "bbox": [ + 103, + 385, + 507, + 540 + ], + "spans": [ + { + "bbox": [ + 103, + 385, + 507, + 540 + ], + "type": "table", + "html": "
MethodSUN397DTDEuroSATUCF101
BaseNovelHMBaseNovelHMBaseNovelHMBaseNovelHM
CoOp (IUCV 22)80.6065.8972.5179.4441.1854.2492.1954.7468.6984.6956.0567.46
CoCoOp (CVPR 22)79.7476.8678.2777.0156.0064.8587.4960.0471.2182.3373.4577.64
MaPLe (CVPR 23)80.8278.7079.7580.3659.1868.1694.0773.2382.3583.0078.6680.77
PromptSRC (ICCV 23)82.6778.4780.5283.3762.9771.7592.9073.9082.3287.1078.8082.74
ArGue (CVPR 24)81.5280.7481.1381.6066.5573.3194.4388.2491.2385.5679.2982.31
DePT (CVPR 24)82.3775.0778.5583.2056.1367.0488.2766.2775.7085.4372.1778.24
CoPrompt (ICLR 24)82.6380.0381.3083.1364.7372.7994.6078.5785.8486.9079.5783.07
PromptKD (CVPR 24)83.6981.5482.6085.8471.3777.9497.5482.0889.1489.7182.2786.10
CoOp + ATPrompt80.8468.6474.2480.8345.4958.2290.3459.7971.9684.4964.9673.45
CoCoOp + ATPrompt80.5076.8678.6478.6356.8966.0287.9574.1580.4682.7476.4079.44
MaPLe + ATPrompt80.9878.1579.5480.5058.2867.6194.8477.5985.3584.0878.8881.40
DePT + ATPrompt82.4276.4879.3482.6456.7767.3089.6069.5078.2885.6073.1578.89
PromptKD + ATPrompt83.8781.3582.5986.9272.3478.9697.0592.0794.4989.2982.4485.73
", + "image_path": "35b6822f6e6a4221919295fc645e77f9894154ef62ffaad7d49cfaaa77d5273d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 89, + 572, + 519, + 671 + ], + "blocks": [ + { + "bbox": [ + 55, + 545, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 545, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 555, + 568 + ], + "type": "text", + "content": "Table 1. Base-to-novel generalization experiments of five baselines with and without our ATPrompt on 11 datasets. Our method achieves consistent average performance improvement over different baselines." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 572, + 519, + 671 + ], + "lines": [ + { + "bbox": [ + 89, + 572, + 519, + 671 + ], + "spans": [ + { + "bbox": [ + 89, + 572, + 519, + 671 + ], + "type": "table", + "html": "
MethodSourceTarget DatasetAverage
Image NetCaltech 101Oxford PetsStanford CarsFlowers 102Food101FGVC AircraftSUN397DTDEuro SATUCF101
CoOp71.5193.7089.1464.5168.7185.3018.4764.1541.9246.3966.5563.88
+ATPrompt71.6793.9690.6565.0170.4085.8620.9765.7743.4446.5969.9265.26 (+1.38)
CoCoOp71.0294.4390.1465.3271.8886.0622.9467.3645.7345.3768.2165.74
+ATPrompt71.2793.7990.6265.9071.1786.0323.2266.6344.4448.7070.7166.59 (+0.85)
MaPLe70.7293.5390.4965.5772.2386.2024.7467.0146.4948.0668.6966.30
+ATPrompt70.6994.0491.0366.0671.9986.3324.4267.0545.2148.6369.1566.75 (+0.45)
", + "image_path": "69bc9ee7fe9d6ad3bc44c32d673523ed019ee6508426f6d794e9624e9b88a337.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 675, + 555, + 699 + ], + "lines": [ + { + "bbox": [ + 55, + 675, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 555, + 699 + ], + "type": "text", + "content": "Table 2. Cross-dataset generalization experiments of three baselines with and without our ATPrompt on 11 datasets. ATPrompt achieves consistent average performance improvement on target datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3623" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 227 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 227 + ], + "type": "table", + "html": "
DatasetAttribute BasesSearched Results
ImageNetcolor, size, shape, habitat, behavior(color, shape)
Caltech101shape, color, material, function, size(shape,size)
OxfordPetsloyalty, affection, energy, playfulness, intelligence(playfulness, energy)
StanfordCarsdesign, engine, performance, luxury, color(luxury)
Flowers102color, flower, habitat, growth, season(color, habitat, growth)
Food101flavor, texture, origin, ingredients, preparation(flavor, preparation)
", + "image_path": "c659414e05650acb23dae8983b6f6b35400b87560abc82ed5c01704f0817a311.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 268, + 295, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 305 + ], + "type": "text", + "content": "on a single A800 GPU for some datasets. In addition, search efficiency can be improved by curating a smaller set of base attributes, which narrows the search space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 137, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 137, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 137, + 329 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 117, + 348 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 352, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 448 + ], + "type": "text", + "content": "Base-to-Novel Generalization. Following [14, 23, 51, 52], we split the dataset into base and novel classes. The model is trained on the base class training set and evaluated on the test set. The attributes applied in this experiment are searched based on the base class. For datasets without a dedicated validation set, like ImageNet, we split the 16-shot labeled data, using half for training and the other half for attribute search." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 508 + ], + "type": "text", + "content": "Cross-dataset Experiments. Consistent with previous works [14, 51, 52], we first train a model on the ImageNet-1K source dataset and then evaluate its generalization performance on several out-of-distribution datasets. The attributes used are obtained from the source dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "type": "text", + "content": "Attribute Search. We select five independent attributes as the basis in the attribute pool. This results in 31 candidate attribute combinations for the search process. We use ChatGPT-4o for attribute queries. In Tab. 3 reports a subset of the queried attribute bases and the final combinations selected by our search." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 580, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 629 + ], + "type": "text", + "content": "Implementation Details. We evaluate the model performance on 15 datasets. We report base and novel class accuracy and their Harmonic Mean (HM) averaged over 3 runs. The details of each dataset are attached in the Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 216, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 216, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 216, + 648 + ], + "type": "text", + "content": "4.2. Base-to-Novel Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "As demonstrated in Tab. 1, we evaluate the base-to-novel generalization performance of five baseline methods, both with and without integrating ATPrompt, across 11 diverse recognition datasets. Notably, ATPrompt consistently improves the average performance of all baseline methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "Reasons for Limited Improvement in Some Conditions. (1) Learnable text prompts are the central component in earlier works, and optimizing them through ATPrompt can obviously improve the performance. (2) Recent studies have expanded beyond learnable text prompts by introducing additional learnable modules. Since our work only involves the optimization of the learnable text prompt part, the improvement have become less obvious." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 178, + 455, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 455, + 190 + ], + "type": "text", + "content": "4.3. Cross-dataset Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "text", + "content": "Tab. 2 shows the cross-dataset generalization results for three baseline methods. Our method demonstrates superior performance, with improvements of " + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "inline_equation", + "content": "1.38\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "inline_equation", + "content": "0.85\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "inline_equation", + "content": "0.45\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 245 + ], + "type": "text", + "content": " for CoOp, CoCoOp and MaPLe, respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 254, + 446, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 446, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 446, + 266 + ], + "type": "text", + "content": "4.4. Domain Generalization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "text", + "content": "Tab. 4 shows the domain generalization results for three baseline methods. The results show that our method improves CoOp, CoCoOp and MaPLe methods by " + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "inline_equation", + "content": "0.90\\%" + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "inline_equation", + "content": "0.33\\%" + }, + { + "bbox": [ + 313, + 272, + 554, + 320 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 319, + 331, + 550, + 425 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 258 + ], + "type": "text", + "content": "Table 3. Part of the results obtained after differentiable attribute search. Please refer to the appendix for the complete results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 331, + 550, + 425 + ], + "lines": [ + { + "bbox": [ + 319, + 331, + 550, + 425 + ], + "spans": [ + { + "bbox": [ + 319, + 331, + 550, + 425 + ], + "type": "table", + "html": "
MethodSourceTarget DatasetAverage
ImageNet-V2-S-A-R
CoOp71.5164.2047.9949.7175.2159.28
+ATPrompt71.6764.4349.1350.9176.2460.18 (+0.90)
CoCoOp71.0264.0748.7550.6376.1859.91
+ATPrompt71.2764.6649.1551.4476.3360.40 (+0.49)
MaPLe70.7264.0749.1550.9076.9860.27
+ATPrompt70.6964.4049.1051.7777.1160.60 (+0.33)
", + "image_path": "67d825f9ea92e00cb5d83fd843d495a5e2aa29b9d01de307be9b92e3b11eae2c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 433, + 555, + 466 + ], + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 466 + ], + "type": "text", + "content": "Table 4. Domain generalization experiments of three baselines with and without our ATPrompt on four datasets. The integration of ATPrompt resulted in better generalization performance." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 484, + 417, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 417, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 417, + 498 + ], + "type": "text", + "content": "4.5. Further Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 502, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 552 + ], + "type": "text", + "content": "By default, the experiments are conducted on the ImageNet. To minimize the influence of other components, we mainly adopt CoOp as the baseline method. Two attributes (color and shape) are used in our ATPrompt." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 361, + 564, + 506, + 679 + ], + "blocks": [ + { + "bbox": [ + 361, + 564, + 506, + 679 + ], + "lines": [ + { + "bbox": [ + 361, + 564, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 361, + 564, + 506, + 679 + ], + "type": "image", + "image_path": "7d195a143a3ea8c18ad49ef0f1be6805308cb3ced856a3581842236793d708d6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 682, + 555, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 682, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 682, + 555, + 715 + ], + "type": "text", + "content": "Figure 5. Illustration of varying soft token length on ImageNet. Increased tokens can lead to overfitting to the base class and weaken generalization to the novel class." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3624" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "Soft Prompt Length. In Fig. 5, we examine the optimal soft token length for both attribute and class tokens. By varying the length from 1 to 10, we observe that longer prompts dilute the guiding influence of the attribute tokens, thereby reducing generalization to novel classes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 216 + ], + "type": "text", + "content": "Class Token Position. The relative positioning of attribute tokens and class tokens in our method requires careful consideration. In Tab. 5, we examine configurations where the class token is positioned in the middle or on either side of two attribute tokens. The results demonstrate that optimal performance is achieved when the class token is placed at the end, which aligns with the findings of CoOp." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 110, + 224, + 241, + 278 + ], + "blocks": [ + { + "bbox": [ + 110, + 224, + 241, + 278 + ], + "lines": [ + { + "bbox": [ + 110, + 224, + 241, + 278 + ], + "spans": [ + { + "bbox": [ + 110, + 224, + 241, + 278 + ], + "type": "table", + "html": "
PositionBaseNovelHM
Front76.1270.5073.20
Middle76.1370.2973.09
End76.2770.6073.33
", + "image_path": "ff2e23003089086000920d60a7859c2184728f6d84ee44bdb7e93fdf63c3a33f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 399 + ], + "type": "text", + "content": "Prompt Operation of Deep Version. In ATPrompt-Deep, we exclusively drop class soft tokens while retaining both hard and soft attribute tokens after they pass through the block. In this part, we compare the performance of partial drop (i.e., removing attribute soft tokens while retaining hard tokens) and full drop (i.e., removing both attribute soft and hard tokens) operations, as illustrated in Tab. 6." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 76, + 409, + 276, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 286, + 295, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 309 + ], + "type": "text", + "content": "Table 5. Comparison of different class token positions on ImageNet. The end position works best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 409, + 276, + 460 + ], + "lines": [ + { + "bbox": [ + 76, + 409, + 276, + 460 + ], + "spans": [ + { + "bbox": [ + 76, + 409, + 276, + 460 + ], + "type": "table", + "html": "
Operation of Attribute TokenBaseNovelHM
Retain all hard and soft tokens76.9470.7273.70
Partial drop and re-add76.8770.4473.51
Full drop and re-add76.8370.1073.31
", + "image_path": "6d9d88adef2551af2a4c77466166cd871c1684e672c5c1440df6fd7dae5669f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 294, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 294, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 294, + 582 + ], + "type": "text", + "content": "The results show that maintaining the attributes of both hard and soft tokens during the forward process results in optimal performance. Conversely, dropping and reintroducing either hard or soft attribute tokens harms performance, likely because it disrupts the continuity of attribute representations across layers and complicates optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 294, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 294, + 677 + ], + "type": "text", + "content": "Attribute Order. In this study, we do not specifically focus on the order of attributes because varying the sequence usually does not result in semantic deviations in reality. Tab. 7 quantitatively assesses the impact of attribute order on prompt learning performance. From this table, we observe that despite variations in order, similar results are consistently produced, and the performance fluctuations across different orders remain within a reasonable range." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Comparison to Other Attributes. In Tab. 8, we explore the effectiveness of attributes derived through alternative methods, specifically by manually selecting class-irrelevant" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 365, + 70, + 504, + 133 + ], + "blocks": [ + { + "bbox": [ + 55, + 469, + 295, + 503 + ], + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 503 + ], + "type": "text", + "content": "Table 6. Comparison of operations on deep soft and hard attribute tokens based on MaPLe+ATPrompt. Preserving hard and soft attribute tokens in deep layers performs better than other operations." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 70, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 365, + 70, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 365, + 70, + 504, + 133 + ], + "type": "table", + "html": "
AttributesBaseNovelHM
(shape, color)76.3270.3973.24
(color, shape)76.2770.6073.33
(size, habitat)76.4470.2373.20
(habitat, size)76.4670.1673.14
", + "image_path": "5f24709d6501914fc69ef83e9642fd0d4be9a6d69c932a562f3df22357f49224.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 189, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 553, + 262 + ], + "type": "text", + "content": "and common attributes. It shows that manually selected irrelevant attributes exhibit comparable performance during training; however, they perform poorly when applied to new categories. This suggests that incorrect attribute tokens cause the soft tokens to develop biased representations, thereby diminishing their zero-shot generalization ability." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 340, + 270, + 529, + 346 + ], + "blocks": [ + { + "bbox": [ + 313, + 140, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 174 + ], + "type": "text", + "content": "Table 7. Comparison of different orders on ImageNet. The order of attributes does not significantly affect the model, and performance fluctuations are within a reasonable range." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 270, + 529, + 346 + ], + "lines": [ + { + "bbox": [ + 340, + 270, + 529, + 346 + ], + "spans": [ + { + "bbox": [ + 340, + 270, + 529, + 346 + ], + "type": "table", + "html": "
TypeAttributesBaseNovelHM
Common(shape, size)82.8367.1374.16
(color, texture)82.7367.5674.38
Irrelevant( plane, engines)82.8166.2273.59
(football, sport)82.7767.1474.14
Searched-82.6868.0474.65
", + "image_path": "95ea42de93842aa5533cd72c5c5f78ba3b8844031cbeb076e73a69565eaf9fe1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 355, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 389 + ], + "type": "text", + "content": "Table 8. Comparison of average performance for various attribute configurations on 11 datasets. The attributes obtained by our method achieve the best performance." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 406, + 387, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 387, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 387, + 419 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 617 + ], + "type": "text", + "content": "In this work, we introduce ATPrompt, an attribute-anchored textual prompt learning method that uses universal attributes as a bridge to improve generalization from seen to unseen categories. Our approach expands the learning space of soft prompts from a one-dimensional, category-centric structure to a multi-dimensional attribute space by anchoring fixed attribute tokens with the prompt. To ensure the selection of optimal attributes, we propose an automated pipeline designed to identify the most suitable candidates for any given downstream task. ATPrompt is designed with both shallow and deep architectural variants, rendering it broadly compatible with existing prompt learning methods. Extensive experiments validate the effectiveness of our approach. We believe this work offers a new direction for research into the fundamental structure of learnable prompts in prompt learning area." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Limitations and future works. This work is a preliminary study of the basic prompt form, which has not been able to achieve comparable performance to regularization-based methods when working alone. Furthermore, the current selection of attribute anchors relies on manual experimentation, and automatically discovering the optimal anchor positions through a learning-based approach remains a promising direction for future research." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3625" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 212 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant Nos. 62361166670 and U24A20330. This research was also supported by the Young Scientists Fund of the National Natural Science Foundation of China (Grant No.62206134), the Fundamental Research Funds for the Central Universities 070-63253222 and the Tianjin Key Laboratory of Visual Computing and Intelligent Perception (VCIP). Computation is supported by the Supercomputing Center of Nankai University (NKSC)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 224, + 115, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 115, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 115, + 237 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 245, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 296, + 277 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 312 + ], + "type": "text", + "content": "[2] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros. Visual prompting via image inpainting. NeurIPS, 35:25005-25017, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 315, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 315, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 61, + 315, + 296, + 358 + ], + "type": "text", + "content": "[3] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "text", + "content": "[4] Keyan Chen, Xiaolong Jiang, Yao Hu, Xu Tang, Yan Gao, Jianqi Chen, and Weidi Xie. Ovarnet: Towards open-vocabulary object attribute recognition. In CVPR, pages 23518-23527, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 438 + ], + "type": "text", + "content": "[5] Tong Ding, Wanhua Li, Zhongqi Miao, and Hanspeter Pfister. Tree of attributes prompt learning for vision-language models. arXiv preprint arXiv:2410.11201, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 506 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 509, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 296, + 540 + ], + "type": "text", + "content": "[7] Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. JMLR, 20(55):1-21, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 543, + 296, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 296, + 575 + ], + "type": "text", + "content": "[8] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 610 + ], + "type": "text", + "content": "[9] Yan Huang, Shang Li, Liang Wang, Tieniu Tan, et al. Unfolding the alternating optimization for blind super resolution. NeurIPS, 33:5632-5643, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "text", + "content": "[10] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[11] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 1, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Baoshuo Kan, Teng Wang, Wenpeng Lu, Xiantong Zhen, Weili Guan, and Feng Zheng. Knowledge-aware prompt tuning for generalizable vision-language models. In ICCV, pages 15670-15680, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[13] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In CVPR, pages 19113-19122, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 217 + ], + "type": "text", + "content": "[14] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In ICCV, pages 15190-15200, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[15] Gahyeon Kim, Sohee Kim, and Seokju Lee. Aapl: Adding attributes to prompt learning for vision-language models. In CVPR Workshop, pages 1572-1582, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 285 + ], + "type": "text", + "content": "[16] Nilakshan Kunananthaseelan, Jing Zhang, and Mehrtash Harandi. Lavip: Language-grounded visual prompting. In AAAI, pages 2840-2848, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[17] Dongjun Lee, Seokwon Song, Jihee Suh, Joonmyeong Choi, Sanghyeok Lee, and Hyunwoo J Kim. Read-only prompt optimization for vision-language few-shot learning. In ICCV, pages 1401-1411, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "text", + "content": "[18] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[19] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[20] Yunheng Li, Yuxuan Li, Quansheng Zeng, Wenhai Wang, Qibin Hou, and Ming-Ming Cheng. Unbiased region-language alignment for open-vocabulary dense prediction. arXiv preprint arXiv:2412.06244, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 487 + ], + "type": "text", + "content": "[21] Yunheng Li, Zhong-Yu Li, Quan-Sheng Zeng, Qibin Hou, and Ming-Ming Cheng. Cascade-CLIP: Cascaded vision-language embeddings alignment for zero-shot semantic segmentation. In ICML, pages 28243-28258. PMLR, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 532 + ], + "type": "text", + "content": "[22] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 577 + ], + "type": "text", + "content": "[23] Zheng Li, Xiang Li, Xinyi Fu, Xin Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. In CVPR, pages 26617-26626, 2024. 1, 2, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 579, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 611 + ], + "type": "text", + "content": "[24] Zheng Li, Xiang Li, Lingfeng Yang, Renjie Song, Jian Yang, and Zhigeng Pan. Dual teachers for self-knowledge distillation. Pattern Recognition, 151:110422, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 656 + ], + "type": "text", + "content": "[25] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In ECCV, pages 19–34, 2018. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[26] Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3626" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[28] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. NeurIPS, 32, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 131, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 295, + 174 + ], + "type": "text", + "content": "[29] Hao Ma, Ming Li, Jingyuan Yang, Or Patashnik, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Clip-flow: Decoding images encoded in clip space. CVMJ, 10(6):1157-1168, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 295, + 209 + ], + "type": "text", + "content": "[30] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 295, + 244 + ], + "type": "text", + "content": "[31] Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameters sharing. In ICML, pages 4095-4104. PMLR, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 246, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 295, + 300 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 295, + 335 + ], + "type": "text", + "content": "[33] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 380 + ], + "type": "text", + "content": "[34] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, pages 13019-13029, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 415 + ], + "type": "text", + "content": "[35] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 295, + 460 + ], + "type": "text", + "content": "[36] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V Le. Mnasnet: Platform-aware neural architecture search for mobile. In CVPR, pages 2820-2828, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 463, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 295, + 496 + ], + "type": "text", + "content": "[37] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argument: Attribute-guided prompt tuning for vision-language models. In CVPR, pages 28578-28587, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 498, + 295, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 541 + ], + "type": "text", + "content": "[38] Yubin Wang, Xinyang Jiang, De Cheng, Dongsheng Li, and Cairong Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749-5757, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "type": "text", + "content": "[39] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS, 35:24824-24837, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 295, + 622 + ], + "type": "text", + "content": "[40] Ge Wu, Xin Zhang, Zheng Li, Zhaowei Chen, Jiajun Liang, Jian Yang, and Xiang Li. Cascade prompt learning for vision-language model adaptation. In ECCV, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[41] Shihan Wu, Ji Zhang, Pengpeng Zeng, Lianli Gao, Jingkuan Song, and Heng Tao Shen. Skip tuning: Pre-trained vision-language models are effective and efficient adapters themselves. In CVPR, pages 14723-14732, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[42] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, and Yongjun Xu. Clip-kd: An empirical study of distilling clip models. arXiv preprint arXiv:2307.12732, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 599 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[43] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 162 + ], + "type": "text", + "content": "[44] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In CVPR, pages 6757-6767, 2023. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 206 + ], + "type": "text", + "content": "[45] Yajing Zhai, Yawen Zeng, Zhiyong Huang, Zheng Qin, Xin Jin, and Da Cao. Multi-prompts learning with cross-modal alignment for attribute-based person re-identification. In AAAI, pages 6979–6987, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 240 + ], + "type": "text", + "content": "[46] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In CVPR, pages 12924-12933, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 554, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 295 + ], + "type": "text", + "content": "[47] Xuying Zhang, Yutong Liu, Yangguang Li, Renrui Zhang, Yufei Liu, Kai Wang, Wanli Ouyang, Zhiwei Xiong, Peng Gao, Qibin Hou, et al. Tar3d: Creating high-quality 3d assets via next-part prediction. arXiv preprint arXiv:2412.16919, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 554, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 351 + ], + "type": "text", + "content": "[48] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 353, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 554, + 407 + ], + "type": "text", + "content": "[49] Xuying Zhang, Yupeng Zhou, Kai Wang, Yikai Wang, Zhen Li, Shaohui Jiao, Daquan Zhou, Qibin Hou, and Ming-Ming Cheng. Ar-1-to-3: Single image to consistent 3d object generation via next-view prediction. arXiv preprint arXiv:2503.12929, 2025.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "type": "text", + "content": "[50] Denny Zhou, Nathanael Scharli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 464, + 554, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 497 + ], + "type": "text", + "content": "[51] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 498, + 554, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 530 + ], + "type": "text", + "content": "[52] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "type": "text", + "content": "[53] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In ICCV, pages 15659-15669, 2023. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 599 + ], + "type": "text", + "content": "[54] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697-8710, 2018. 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3627" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_content_list.json b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5d6ea1b2880ff3106ff9915f2961a155182557a4 --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_content_list.json @@ -0,0 +1,1505 @@ +[ + { + "type": "text", + "text": "Advancing Visual Large Language Model for Multi-granular Versatile Perception", + "text_level": 1, + "bbox": [ + 143, + 128, + 852, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wentao Xiang $^{1*}$ , Haoxian Tan $^{2*}$ , Yujie Zhong $^{2\\dagger}$ , Cong Wei $^{1}$ , Dengjie Li $^{2}$ , and Yujiu Yang $^{1\\dagger}$ $^{1}$ Tsinghua University \n $^{2}$ Meituan Inc.", + "bbox": [ + 137, + 202, + 857, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thu_xiangwentao@163.com, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn", + "bbox": [ + 163, + 258, + 826, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perception is a fundamental task in the field of computer vision, encompassing a diverse set of subtasks that can be systematically categorized into four distinct groups based on two dimensions: prediction type and instruction type. Notably, existing researches often focus solely on a limited subset of these potential combinations, which constrains their applicability and versatility across various contexts. In response to this challenge, we present MVP-LM, a Multi-granular and Versatile Perception framework incorporating Visual Large Language Model. Our framework is designed to integrate both word-based and sentence-based perception tasks alongside box and mask predictions within a single architecture. MVP-LM features an innovative multi-granularity decoder in conjunction with a CoT-inspired dataset unification strategy, enabling seamless supervised fine-tuning across a wide spectrum of tasks, including but not limited to panoptic segmentation, detection, grounding, and referring expression segmentation. Furthermore, we introduce a query enhancement strategy aimed at harnessing the decoding and generative capabilities inherent in VLLMs. Extensive experiments conducted across a range of benchmarks in both word-based and sentence-based perception tasks substantiate the efficacy of our framework. The code will be available here.", + "bbox": [ + 89, + 339, + 483, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 731, + 220, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perception constitutes a fundamental task within the field of computer vision, necessitating that models accurately identify and locate objects within images or videos in accordance with specified instructions. These perception tasks can be categorized along two dimensions: prediction type (box vs. mask) and instruction type (word-based vs. sentence-based). This classification yields four different groups.", + "bbox": [ + 89, + 756, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/69b304799a32a07b9bc17a257e4adaed2a29d581c2eb0cb30a379bf4e6960d70.jpg", + "table_caption": [ + "Table 1. The comparison of capabilities of different methods. Current works can only address a subset of the combinations, while our MVP-LM can cover all tasks." + ], + "table_footnote": [], + "table_body": "
MethodOutput TypeInstruction Type
BoxMaskWordSentence
Specialists
OWL[52], PromptDet[14], OV-DINO[66]✓✓
X-Decoder[90], OpenSeeD[85], Mask DINO[36]✓✓✓
MDETR[29], GLIP[38], Grounding DINO[46]✓✓✓
VLT[10], LAVT[79], PolyFormer[45]✓✓
VLLM-based Generalists
QwenVL[1], InternVL[6], DeepseekVL[51]✓✓
LISA[32], PixelLM[58], Glamm[57]✓✓
OMG-LLaVA[87], PSALM[88], HyperSeg[71]✓✓✓
MVP-LM (Ours)✓✓✓✓
", + "bbox": [ + 516, + 349, + 901, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The distinction between coordinate-level prediction and pixel-level prediction has been extensively explored in the literature. Given that mask annotations can be converted to box annotations without loss of information, the concurrent optimization of box predictions and mask predictions within the same dataset has emerged as a prominent training paradigm. This approach has demonstrated mutual benefits for both prediction types [23, 36]. In contrast to the costly annotations required for segmentation tasks, box annotations are significantly easier to obtain, resulting in a considerably larger dataset. Consequently, another line of research focuses on enhancing segmentation performance by leveraging additional box-annotated data. Numerous methods [4, 23, 34, 35, 39, 64, 85] have been developed to achieve this, ranging from the utilization of pre-trained models to natively annotate box data, to the design of specialized segmentation loss functions tailored for box annotations.", + "bbox": [ + 509, + 518, + 906, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Word-based perception employs individual words to denote the categories of targets, which often results in a degree of vagueness and ambiguity. A category name can correspond to none or multiple objects within an image. In contrast, sentence-based perception utilizes complete sentences to describe the objects of interest, resulting in greater precision. In most cases, only one object within the im", + "bbox": [ + 509, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 114, + 886, + 372, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22153", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age satisfies the specified conditions. A common approach in the literature [29, 38, 46] involves the deployment of a two-stream framework, wherein visual queries extracted by any vision backbone are utilized to decode object locations and compute similarity with the text features generated by a language encoder to identify the objects. However, many of these methods overlook the underlying semantic concepts inherent in the descriptions, resulting in suboptimal outcomes, particularly for sentence-based perception.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, Large Language Models (LLMs) have demonstrated a remarkable ability to comprehend user queries and generate appropriate responses, positioning them as a viable option for sentence-based perception [1, 5, 6, 33, 51, 57, 59]. Additionally, LLMs possess the capability to extract language features through forwarding, which aligns well with word-based perception. Therefore, LLMs reveal substantial potential for unifying both perceptions.", + "bbox": [ + 89, + 228, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The representative of sentence-based location tasks, Referring Expression Comprehension (REC), has emerged as the standard benchmark for Visual Large Language Models (VLLMs), as box coordinates can be expressed as a sequence of numbers in text form, rendering them compatible with traditional VLLMs. However, many existing models [1, 6, 51] encounter difficulties in achieving pixel-level predictions. Recent advances [5, 33, 57, 59] have successfully output masks by incorporating an additional mask decoder. Nevertheless, these approaches have not adequately addressed the challenges associated with word-based perception.", + "bbox": [ + 89, + 351, + 482, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, other studies [87, 88] have demonstrated comparable performance on standard word-based perception benchmarks by employing the VLLM as a versatile decoder. Nonetheless, these models focus solely on pixel-level prediction and overlook box prediction.", + "bbox": [ + 89, + 532, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Tab. 1, although joint training has been examined for certain combinations of tasks, the collaborative effects of joint training across all four groups have not been thoroughly explored in the existing literature.", + "bbox": [ + 89, + 609, + 482, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we introduce a Multi-granular and Versatile Perception framework incorporating VLLM, termed MVPLM, which is capable of detecting and segmenting targeted objects in images according to various types of user instructions within a single model. By employing a Chain-of-Thought (CoT)-inspired data curation [72], we transform multiple box- and mask-annotated datasets into a unified supervised fine-tuning (SFT) dataset format, thereby accommodating a diverse set of tasks, including panoptic segmentation, object detection, visual grounding, and referring expression segmentation. Furthermore, to leverage the decoding and generative capabilities of LLMs, we enhance the queries for object identification and localization by incorporating features derived from LLM-generated sequences.", + "bbox": [ + 89, + 672, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are outlined as follows:", + "bbox": [ + 109, + 885, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce MVP-LM, a framework that harnesses both the decoding and generative capabilities of VLLM to perform both word-based and sentence-based perception tasks across varying granularities.", + "- We have developed a CoT-inspired data curation that unifies datasets from diverse tasks into a single SFT dataset, thereby encouraging the model to adopt a \"thinking-then-perceiving\" paradigm.", + "- MVP-LM demonstrates competitive performance across various benchmarks in both word-based and sentence-based perception, thereby validating the potential of our framework. Notably, our MVP-LM achieves remarkable results on both closed-set and open-set tasks compared with other specialists and VLLMs." + ], + "bbox": [ + 513, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 316, + 653, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Perception in computer vision can be categorized into four groups along two dimensions: prediction type and instruction type. Below, we review both the related work within each of these four groups and CoT, and highlight the gaps MVP-LM aims to fill.", + "bbox": [ + 511, + 342, + 903, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Word-based Preception", + "text_level": 1, + "bbox": [ + 513, + 425, + 730, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Detection and Grounding. Word-based location tasks encompass object detection, grounding, and instance detection, focusing on identifying and localizing objects via bounding boxes with categorical labels. Conventional frameworks [13-15, 17, 18, 31, 37, 73, 83] utilize CLIP-like encoders to compute visual-linguistic similarities, enabling zero-shot capabilities. Significant advances include GLIP [38], which unifies phrase grounding with object detection, and OV-DINO [66], which implements an end-to-end training pipeline for enhanced detection performance and generalization capability.", + "bbox": [ + 511, + 446, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Segmentation. Word-based segmentation tasks, encompassing panoptic, semantic, and instance segmentation, demand precise pixel-level predictions. However, generating such masks from textual inputs remains underexplored, largely due to the scarcity of fine-grained annotated data. Existing methods [2, 22, 40, 65, 75, 82, 85, 90] primarily align word embeddings with visual features, resulting in coarse segmentation masks. Nevertheless, these approaches [85, 90] often demonstrate limited textual understanding and lack the requisite fine-grained control for high-quality mask generation, which ultimately restricts their performance in both closed-set and open-set tasks.", + "bbox": [ + 511, + 614, + 903, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Sentence-based Preception", + "text_level": 1, + "bbox": [ + 513, + 803, + 754, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Expression Comprehension. Sentence-based location tasks, which require a deeper multi-modality understanding, have become a standard benchmark for VLLMs. Specifically, Referring Expression Comprehension (REC) is often the preferred task, as box coordinates", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22154", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b4a92503a5159d1dde21a5c44042b32c28964e78512e74880dc0bcbf71808ee7.jpg", + "image_caption": [ + "Figure 1. Overview of MVP-LM. MVP-LM implements perception by integrating a multi-granularity decoder into the existing VLLM framework. We utilize a unified prompt template to construct the input sequence for the LLM across different tasks. The base query is derived from the summary token of the generated response. Concurrently, we extract the instruction embeddings from the input sequence (denoted by the same color) and select the corresponding residuals from the multi-scale visual features based on these embeddings. After aggregating the base query vector with the residuals, we decode the bounding box and segmentation mask by inputting the resulting queries into the multi-granularity decoder." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 861, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "can be represented as a text sequence of numbers, making them compatible with conventional VLLMs [1, 6, 41, 47, 48, 51, 60, 67, 77, 84]. However, many existing models struggle to achieve pixel-level predictions, and even more so, to facilitate joint training with sentence-based mask perception.", + "bbox": [ + 88, + 383, + 480, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Referring Expression Segmentation. Sentence-based segmentation tasks demand a fine-grained understanding of descriptive sentences to accurately generate segmentation masks for target objects. Traditional approaches [3, 9, 24-27, 56] have made strides in fusing visual and textual features to improve segmentation performance, but often fall short in fully leveraging the richness of sentence descriptions, resulting in suboptimal performance in both closed-set and open-set scenarios. Recent efforts [5, 33, 43, 57, 59] have successfully implemented Referring Expression Segmentation (RES) by incorporating an additional mask decoder. However, these approaches fall short in addressing word-based perception.", + "bbox": [ + 89, + 481, + 483, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Chain-of-Thought", + "text_level": 1, + "bbox": [ + 89, + 705, + 266, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Chain-of-Thought prompting [72] was initially introduced during inference to enhance the reasoning capabilities of LLMs. Subsequently, CoT has evolved into a prevalent paradigm in the post-training phase, wherein models first undergo supervised fine-tuning on CoT datasets, followed by reinforcement learning [62]. A similar trend has also been observed in VLLMs, which have demonstrated promising performance in multi-modal understanding and reasoning tasks. Various meticulously designed rewards [16, 21, 50, 68, 78, 86], ranging from format to geometry, have been proposed to handle diverse scenarios.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Identified Gaps and Our Contribution", + "text_level": 1, + "bbox": [ + 511, + 383, + 841, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While significant progress has been made in addressing some of the combinations [6, 33, 51, 59, 88], a comprehensive solution effectively handling all four groups remains elusive. MVP-LM bridges this gap by offering a solution capable of simultaneous perception across all four combinations, marking a significant advancement in the field, addressing the current limitations, and setting a new benchmark for multi-dimensional visual perception.", + "bbox": [ + 511, + 407, + 906, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 547, + 648, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides an in-depth description of our proposed methodology. We begin by detailing the architecture of our MVP-LM in Sec. 3.1, followed by an exposition of our strategies for dataset processing in Sec. 3.2.", + "bbox": [ + 511, + 574, + 906, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. MVP-LM Architecture Design", + "text_level": 1, + "bbox": [ + 511, + 648, + 782, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "System Overview. As illustrated in Fig. 1, our architecture comprises four essential components: an image encoder (Swin-Base Transformer [49]), a connector module, an LLM (Phi-1.5 [20]), and a multi-granularity decoder. The image encoder extracts rich visual features from input images, serving as the foundation for subsequent processing. The connector module facilitates seamless integration between visual and textual modalities by aligning features from the image encoder with the token embeddings for the LLM. The multi-granularity decoder is responsible for producing both bounding boxes and segmentation masks.", + "bbox": [ + 511, + 672, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Query Generation. The dynamic generation of visual queries is pivotal for the model's ability to handle diverse perception tasks effectively. As depicted in Fig. 2, each visual query is composed of two components: a", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22155", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dc119ff177d4b859279c9495fc6e2fc84ff92a1ea65c270c53de66f99ae09716.jpg", + "image_caption": [ + "Figure 2. Dynamic query selection. MVP-LM expands the embedding (hidden states) of the summary token into $N$ base query vectors. Concurrently, multi-scale visual features are flattened, and their similarity with the instruction embedding from the input sequence is computed. The top $N$ similar features are then selected and integrated with the corresponding $N$ base query vectors, resulting in $N$ final queries tailored for multi-granularity perception tasks. This dynamic selection mechanism ensures that the most relevant visual features are incorporated into each query, enhancing the precision and adaptability of the decoder." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 459, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "context-aware base query vector and language-guided residuals.", + "bbox": [ + 89, + 356, + 482, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Initially, MVP-LM leverages the VLLM to generate a specialized summary token that encapsulates the essential information from the input sequences. The hidden states corresponding to this summary token are projected into the base query vectors using an MLP, providing a robust foundation for query formulation.", + "bbox": [ + 89, + 386, + 483, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Simultaneously, we compute the similarity between multi-scale visual features and the text embedding derived from the input instructions (i.e., category names for word-based perception and referring expressions for sentence-based perception). To facilitate this comparison, an MLP aligns the dimensionality of the text embeddings with that of the multi-scale visual features. The top $N$ most similar visual features are then identified, ensuring that the queries are informed by the most relevant visual information. These selected features are integrated with the corresponding base query vectors, resulting in $N$ refined final queries that encapsulate both contextual and visual clues. This dynamic query generation process enables MVP-LM to adaptively focus on relevant regions of the image, thereby enhancing the model's ability to generate accurate bounding boxes and segmentation masks.", + "bbox": [ + 89, + 478, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-granularity Decoder. The multi-granularity decoder is the cornerstone of our architecture, facilitating the simultaneous output of bounding boxes and segmentation masks. Our decoder is inspired by the design principles of OpenSeeD [85], which emphasizes the profound integration of multi-scale features and comprehensive query processing. Initially, a set of content queries and their corresponding reference points are generated through a query selection mechanism as seen in Fig. 2. These queries are then iteratively cross-attended with multi-scale visual features via deformable attention layers, enabling the decoder to capture both global contextual information and detailed", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "local features. The outputs from each MSDeform layer are subsequently processed by three shared heads—a cross-modal similarity computation head, a box head, and a mask head—to produce predictions. The unified output streamlines the training process and allows the model to learn shared representations that benefit both detection and segmentation tasks.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Objectives. To fully exploit the capabilities of MVP-LM, we employ a joint training approach that simultaneously optimizes multiple tasks across diverse datasets. The training objective is formulated as a composite loss function comprising several components:", + "bbox": [ + 511, + 210, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {l l m} + \\lambda_ {w o r d} \\mathcal {L} _ {w o r d} + \\lambda_ {s e n t} \\mathcal {L} _ {s e n t} \\\\ + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {b o x} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 297, + 841, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{llm}$ represents the standard loss for next-token prediction, ensuring that the LLM accurately generates coherent and contextually appropriate summaries. The term $\\mathcal{L}_{word}$ corresponds to the Cross-Entropy (CE) loss for word-based perception conditions, facilitating precise classification based on single-word instructions. Conversely, $\\mathcal{L}_{sent}$ denotes the Binary Cross Entropy (BCE) loss for sentence-based perception conditions, enabling the model to handle more complex, descriptive instructions effectively. The mask loss, $\\mathcal{L}_{mask}$ , is composed of a pixel-level BCE loss and a Dice loss, which together promote accurate and smooth segmentation masks by penalizing discrepancies between predicted and ground-truth masks at both the pixel and region levels. The box loss, $\\mathcal{L}_{box}$ , includes the $l1$ loss to minimize the absolute differences between predicted and ground-truth bounding box coordinates, as well as the Generalized Intersection over Union (GIoU) loss to enhance the alignment and overlap between predicted and actual boxes. Formally, these loss components are defined as:", + "bbox": [ + 511, + 344, + 906, + 645 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\\\ \\mathcal {L} _ {b o x} = \\lambda_ {l 1} \\mathcal {L} _ {l 1} + \\lambda_ {g i o u} \\mathcal {L} _ {g i o u}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 657, + 820, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the methodology of OpenSeeD, we employ bipartite matching to match the model's predictions with their corresponding ground truth annotations during training. Predictions from each layer are assigned to their corresponding annotations to calculate losses through Hungarian matching. Moreover, a denoising strategy is incorporated, similar to that used in MaskDINO, to stabilize optimization and accelerate convergence. However, during inference, only the final layer's output is utilized.", + "bbox": [ + 511, + 703, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Curation", + "text_level": 1, + "bbox": [ + 511, + 848, + 661, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CoT-inspired Dataset Unification. We unified heterogeneous datasets into an SFT dataset specially tailored for", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22156", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the training of VLLM. For each sample in the perception datasets, we construct a question-answer pair based on the original annotations. Each training pair consists of three key components: a task description, an instruction, and a response that embodies a \"thinking-then-perceiving\" process.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task description specifies the particular task associated with the data. In the case of word-based perception, the description may be articulated as follows: \"Please identify all objects according to the given phrase list. This is all the candidate phrases.\" Conversely, for sentence-based perception, the description is framed as: \"Please identify the target according to the following instruction.\"", + "bbox": [ + 89, + 183, + 480, + 289 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The instruction varies according to the specific tasks involved. For word-based perception, we concatenate the provided word lists into a single sentence, using a comma as a separator. For instance, given three categories: apple, banana, and orange, the resultant sentence would be \"apple, banana, orange.\" After passing this sentence through VLLM, we select the corresponding output embeddings and apply average pooling to obtain the text embedding for each word, while for sentence-based perception, the referring expression directly serves as the final instruction, requiring no additional processing. The text embedding is derived simply by applying pooling over the entire expression.", + "bbox": [ + 89, + 291, + 480, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The response format is standardized across various tasks. We adopt a \"thinking-then-perceiving\" strategy by preponding the input image caption to the summary, thereby encouraging the VLLM to effectively capture image details. Specifically, the response is structured as follows: \"[image caption]. The perception result is $<\\text{PER}>$ \", where \"<\\text{PER}>\" denotes a special token representing the summary. The hidden states associated with this summary token are subsequently utilized to derive the visual queries.", + "bbox": [ + 89, + 474, + 480, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-caption Auto-labeling. To fully leverage the generative prowess of VLLMs, we design our system to generate comprehensive image captions in conjunction with summary tokens. This generation approach enriches the contextual information available to the model, facilitating a deeper understanding of the visual content. However, a significant challenge arises as the majority of our training datasets lack image captions. To address this, we integrate open-source VLLMs to autonomously generate descriptive captions for each image prior to the training phase.", + "bbox": [ + 89, + 611, + 480, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we employ a diverse set of VLLMs, including VILA-3B, VILA-13B, InternVL2-8B, and InternVL2-26B, to produce multiple unique captions for the same image. The utilization of distinct VLLMs not only enhances the accuracy of the generated captions but also introduces substantial diversity, mitigating the risk of model overfitting to a single captioning style or vocabulary. By incorporating multi-caption auto-labeling, our approach ensures a more comprehensive and resilient training paradigm, effec", + "bbox": [ + 89, + 763, + 480, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e8256d8e90baff093c2564635e1f696d2d470b827208a99bfbbe44a3ca38dcb0.jpg", + "table_caption": [ + "Table 2. Hyperparameters for both training stages. \"CP\", \"RC\", \"O\", and \"GG\" denote COCO-Panoptic, RefCOCO, Objects365, and GoldG, respectively." + ], + "table_footnote": [], + "table_body": "
ParametersStage1Stage2
Training ComponentsConnectorConnector + LLM + Multi-granularity Decoder
OptimizerAdamW
Training Rate2 × 10-34 × 10-5
Batch Size12864
Number of Steps465080000
Learning Rate ScheduleCosine Decay
Weight Decay0.00.05
Warmup Ratio0.03
Training DataCC3MCP(33.3%) RC(33.3%) O(16.7%) GG(16.7%)
LossLLLMLLLM + 2 · Lword/sent + 5 · LL1 + 2 · LGIoU + 5 · LBCE + 5 · LDICE
Image Size1024 × 1024
Image ProcessingResize longer to 1024 and pad shorter to 1024
", + "bbox": [ + 516, + 128, + 903, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tively bridging the gap between visual inputs and linguistic descriptions.", + "bbox": [ + 511, + 308, + 903, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data Refinement. During the data generation process, the occurrence of irregular or erroneous data is inevitable. Consequently, we implement a straightforward data-cleaning procedure for the dataset. First, we utilize the length and information entropy of the generated captions to filter out meaningless results. Second, we remove sentences containing suggestive keywords such as \"indicate\", \"might\", \"may\", and \"imply\", as well as other similar terms, to mitigate the hallucinations that are inherent in large language models. These two mechanisms ensure the quality of the generated captions.", + "bbox": [ + 511, + 344, + 905, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Randomness in Training Pairs. The variability of training pairs is essential for preventing the model from learning spurious correlations and for enhancing its generalization capabilities. In our framework, such diversity is achieved by randomly selecting one of four generated captions for each training response. However, fixed category name ordering in word-based datasets can still introduce spurious causal relationships, leading to overfitting and impaired performance on unseen data.", + "bbox": [ + 511, + 516, + 905, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To counteract these issues, our mitigation strategy is twofold. Firstly, we introduce randomness by shuffling the order of the word list in each iteration. This disrupts fixed positional biases, preventing the model from relying on word order for predictions. Secondly, word lists are dynamically constructed by randomly incorporating a subset of negative category names (categories absent from the image). This selective inclusion diversifies training prompts, proving particularly beneficial for datasets with extensive vocabularies. By carefully balancing positive and negative categories within VLLM input sequence length constraints, the model's ability to discern relevant from irrelevant objects is enhanced. This approach fosters a more adaptive learning process, significantly reducing overfitting and ensuring robust performance across both closed-set and open-set detection scenarios.", + "bbox": [ + 511, + 659, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22157", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d49f4138d048c396083526bdde60e8869859d5ad4bcd964fdcf080907a5c0d8c.jpg", + "table_caption": [ + "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation and open-set segmentation tasks. Our model MVP-LM achieves remarkable performance among previous VLLM-based methods." + ], + "table_footnote": [], + "table_body": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
SpecialistsMask2former [7]Swin-B55.165.1-----
OneFormer [28]Swin-L57.967.4-----
OpenSeeD [85]Swin-L59.568.619.723.441.4--
SEEM [91]DaViT-B56.166.3-----
MaskCLIP [11]ViT-L30.947.615.123.7-45.9-
SimBaseline [76]ViT-B---20.5-47.788.4
DeOP [22]ResNet-101c---22.9-48.891.7
DaTaSeg [19]ViTDet-B52.862.712.318.328.051.1-
VLLM-based GeneralistsOMG-LLaVA [87]ConvNeXt-L53.8------
PSALM [32]Swin-B55.966.613.718.228.848.581.3
MVP-LMSwin-B56.166.819.420.535.344.185.7
", + "bbox": [ + 147, + 116, + 846, + 294 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b78b6cf1cc6c9010bc3b6f28b56a7a77c67f380cd7303426ae5ae4c6abfac1a.jpg", + "table_caption": [ + "Table 4. Comparison with the state-of-the-art models on the closed-set Referring Segmentation benchmarks (RefCOCO series) and more challenging Generalized Referring Expression Segmentation benchmark gRefCOCO[43]. Gray numbers denote the method using gRefCOCO for training." + ], + "table_footnote": [], + "table_body": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SpecialistsVLT [10]67.570.565.256.361.050.155.057.752.562.250.5
CRIS [69]70.573.266.162.368.153.759.960.455.363.851.0
LAVT [79]72.775.868.862.168.455.161.262.157.665.355.0
PolyFormer-B [45]74.876.671.167.672.959.367.869.1---
VLLM-based GeneralistsLISA-7B [32]74.176.571.162.467.456.566.468.538.752.644.8
PixelLM-7B [58]73.076.568.266.371.758.369.370.5---
GSVA-7B [74]76.477.472.864.567.758.671.172.061.769.260.3
LaSagnA-7B [70]76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [87]78.080.374.169.173.163.072.972.9---
PSALM [88]83.684.781.672.975.570.173.874.442.052.450.6
MVP-LM83.685.182.573.976.865.475.175.637.650.449.9
", + "bbox": [ + 138, + 349, + 857, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 550, + 223, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We train MVP-LM in a multi-dataset and multitask manner like [19, 88]. For generic perception, we use COCO Panoptic Segmentation [42](with 118K samples in total) and O365 [61](with 1.7M samples in total). We use RefCOCO series [54, 80](with 126K samples in total) for referring segmentation tasks while utilizing the GoldG [30](with 770K samples in total) dataset for grounding. For evaluation, we report the main results on conventional COCO image segmentation tasks (panoptic/semantic) and RefCOCO series. Besides, we conduct evaluation on more challenging perception tasks like open-set segmentation and detection [8, 12, 53, 89].", + "bbox": [ + 89, + 578, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We report results in widely used evaluation metrics: panoptic quality (PQ) and mean intersection-over-union (mIoU) for panoptic and semantic segmentation, respectively. Additionally, we use cumulative Intersection-over-Union (cIoU) for referring segmentation evaluation consistent with previous studies.", + "bbox": [ + 89, + 762, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The model architecture integrates Phi1.5 [20] as the language model and Swin-B [49] as visual encoder, with our multi-granularity decoder building", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on OpenSeeD [85]. For most tasks, we initialize 100 queries for subsequent mask decoding, and for panoptic segmentation, we introduce an additional 100 learnable queries specifically for stuff categories. Our training proceeds in two phases following LLaVA [44]: 1) visual-language alignment trains exclusively the connector on CC3M [63] while freezes other components; 2) subsequent fine-tuning stage optimizes the whole model except the vision encoder. Detailed hyperparameters and training configurations are listed in Tab. 2. Training spans 80k iterations on the aggregated datasets and 9k for ablation studies, which utilize only the COCO Panoptic and RefCOCO datasets.", + "bbox": [ + 511, + 551, + 906, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main results", + "text_level": 1, + "bbox": [ + 511, + 742, + 648, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Word-based Perception Results. We present the performance of MVP-LM on the COCO-Panoptic [42] and open-set segmentation [8, 12, 53, 89] tasks as detailed in Tab. 3. MVP-LM demonstrates competitive performance when compared to both specialized models and VLLM-based methods in closed-set tasks. Notably, it exhibits exceptional performance in the realm of open-set segmentation within the context of VLLM. Specifically, when evaluated on the open-set ADE20K dataset [89], our", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22158", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/af10024bab862a3a9c57f07ac4ffc081cb4860a2e438146429d38c6b519b8b84.jpg", + "table_caption": [ + "Table 5. Results on referring expression comprehension tasks. DS-VL2 is the abbreviation for DeepSeek-VL2." + ], + "table_footnote": [], + "table_body": "
MethodParamsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra13B87.891.181.882.987.874.482.683.1
MiniGPTv27B88.791.785.380.085.174.584.484.7
Octopus7B89.092.683.483.589.476.084.486.2
MVP-LM1.3B93.594.591.684.988.879.386.787.4
DS-VL2200B+95.196.795.191.294.987.492.892.9
", + "bbox": [ + 93, + 116, + 480, + 202 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MVP-LM demonstrates substantial improvements over the VLLM-based PSALM [88] and achieves gains of 5.7 and 2.3 in PQ and mIoU metrics, respectively. Additionally, it surpasses PSALM by 6.5 and 4.4 on the mIoU metric for the open-set datasets Cityscapes [8] and PascalVOC-20 [12], respectively.", + "bbox": [ + 89, + 220, + 482, + 310 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sentence-based Perception Results. We evaluate MVP-LM against state-of-the-art methods on the referring expression segmentation (RES) benchmarks, RefCOCO+/g[55, 81] as well as the more challenging generalized referring expression segmentation benchmark, gRefCOCO[43] in Tab. 4. We also include RefCOCO series for referring expression comprehension (REC) in Tab. 5. In the domain of RES, MVP-LM achieves state-of-the-art performance on the RefCOCO and RefCOCOg datasets across all subsets. In the RefCOCO+ dataset, MVP-LM attains scores of 73.9 and 76.8 on val and testA, respectively. Moreover, our MVP-LM, only 1.3B parameters, outperforms listed 7B and 13B models, and achieves metrics that are second only to DeepSeek-VL2 with over 200B parameters in REC. These results highlight the robustness and generalization capabilities of MVP-LM when compared to previous methods.", + "bbox": [ + 89, + 311, + 482, + 553 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 561, + 241, + 578 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This section presents an ablation study investigating critical architectural decisions, specifically: multi-dataset joint learning, performance across diverse response configurations, the impact of query quantity, and the effect of query selection.", + "bbox": [ + 89, + 583, + 482, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint Training. The progressive incorporation of training data sources is analyzed in Tab. 6. Initial training on COCO and RefCOCO datasets yields a MVP-LM mIoU of 77.6 on the RefCOCO validation set. Subsequent inclusion of O365 significantly increases cIoU to 81.8, with further enhancement to 83.6 cIoU upon integrating the Grounding dataset.", + "bbox": [ + 89, + 657, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For COCO evaluation, sequential integration of O365 yields a PQ of 55.8 and mIoU of 65.9. Subsequent integration of the Grounding dataset marginally improves PQ to 56.1 and mIoU to 66.8.", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Crucially, sentence-based perception is enhanced by this multi-dataset approach by 6.0 points while word-based perception remains consistent. These findings empirically validate the effectiveness of our unified multi-task training strategy in coordinating multi-modal representations and significantly enhancing cross-task generalization.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63a29edc1634511024a4bf67593c7bb9cc08f220a65f29286174723766eca910.jpg", + "table_caption": [ + "Table 6. Ablation of the datasets involved in the joint training. We use the abbreviations C, R, O, and G to represent COCO, Ref-COCO, O365, and Grounding, respectively. Both experiments are trained for 56k iterations." + ], + "table_footnote": [], + "table_body": "
DatasetRefCOCO val cIoUCOCO
PQmIoU
C, R77.656.466.3
C, R, O81.855.865.9
C, R, O, G83.656.166.8
", + "bbox": [ + 593, + 143, + 823, + 212 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/05bc8ed1c55b0ae871b3c2169719067fa1b5ada91523c0b2133179eb0c210af9.jpg", + "table_caption": [ + "Table 7. Ablation of different response settings. 'None' denotes the basic response setting containing the summary token only. 'Existing Obj' means generating existing objects' names or descriptions in the image before summary tokens. 'Caption' indicates the regression of the image caption prior to summary tokens. $^\\dagger$ denotes the direct use of similar visual features as the final queries without LLM-generated base query vectors." + ], + "table_footnote": [], + "table_body": "
Response settingRefCOCO val cIoUCOCO
PQmIoU
None75.655.365.7
Existing Obj.75.654.965.6
Caption75.755.666.2
Caption†75.955.066.1
", + "bbox": [ + 576, + 328, + 841, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Response Settings. We provide a comprehensive ablation study on various response settings in Tab. 7. The baseline response, \"The perception result is $$ \", remains consistent across all inputs, achieving 75.6 cIoU on the RefCOCO validation set and 55.3 PQ/65.7 mIoU on COCO. When the LLM is instructed to output existing category names or object descriptions from the input image prior to the baseline response, the model maintains its performance on RefCOCO (75.6 cIoU) but exhibits a decline in COCO metrics, resulting in 54.9 PQ and 65.6 mIoU. Conversely, our final configuration, which directs the LLM to generate the image caption before delivering the baseline response, results in simultaneous improvements across all metrics (75.7 cIoU, 55.6 PQ, and 66.2 mIoU).", + "bbox": [ + 511, + 429, + 906, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we evaluate the performance of utilizing the most similar visual features as the final queries. Compared to the approach of combining these features with the base query vectors generated by the LLM, this configuration yields a modest improvement in the RefCOCO validation set (+0.2). However, it is accompanied by a noticeable decline in COCO PQ, resulting in a decrease of 0.6.", + "bbox": [ + 511, + 642, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The incorporation of descriptive captions in the responses facilitates the mutual enhancement of both sentence-level (RefCOCO) and word-level (COCO) perception tasks. This observation verifies our hypothesis that structured semantic prompting significantly improves visual-language understanding.", + "bbox": [ + 511, + 750, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of Queries. As presented in Tab. 8, the response to variations in the number of queries differs between the RefCOCO and COCO datasets. Specifically, performance on the RefCOCO validation set deteriorates with an increasing", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22159", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e9b7b06f213aabed6e6096e71a8b5fec4acfa25638b4039e03241eddd5bb914.jpg", + "image_caption": [ + "COCO", + "Figure 3. Visualization of the segmentation results on COCO, ADE20K and gRefCOCO datasets. We plot the predicted mask with a confidence larger than 0.5 for a clear visualization. The text above each image in the gRefCOCO denotes the corresponding reference text." + ], + "image_footnote": [], + "bbox": [ + 145, + 90, + 848, + 385 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7cf3d7fc19d49319883a821070af5fea32ae12cff39c09df76b7689598cf5594.jpg", + "table_caption": [ + "Table 8. Ablation of the number of queries. 100 queries is the default setting, considering the balance between performance and computational costs." + ], + "table_footnote": [], + "table_body": "
# queriesRefCOCO val cIoUCOCO
PQmIoU
3076.353.165.2
10075.755.666.2
30072.454.965.5
", + "bbox": [ + 173, + 489, + 398, + 559 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/600b64eef6d7ee204b3a8394fab4d6dc2cc4f6614d0545b08ce8219ac8831e02.jpg", + "table_caption": [ + "Table 9. Ablation of whether applying query selection." + ], + "table_footnote": [], + "table_body": "
Apply Query SelectionRefCOCO val cIoUCOCO
PQmIoU
✓75.755.666.2
78.354.465.1
", + "bbox": [ + 138, + 593, + 431, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "number of queries, decreasing from 76.3 cIoU to 72.4 cIoU. We hypothesize that, due to the presence of only one target in the RefCOCO images, an increase in the number of candidates may complicate the model's task. In contrast, for the COCO dataset, the use of 100 queries yields optimal performance among the three settings evaluated. Additionally, it is important to note that an increased number of queries incurs additional computational costs. Consequently, we adopt 100 queries as our default setting to achieve an optimal balance between performance and efficiency.", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Query Selection. As conveyed in Tab. 9, applying separate learnable query embeddings instead of query selection improves RefCOCO val results (from 75.7 to 78.3) but lowers COCO-Panoptic results (from 55.6 to 54.4). It reflects that RefCOCO's simpler targets suit fixed queries, and that COCO's complexity needs more flexible query selection.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 438, + 694, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present visualizations for the COCO, ADE20K, and gRefCOCO datasets in Fig. 3. The results demonstrate that MVP-LM effectively generates precise masks and is proficient in identifying a comprehensive range of objects.", + "bbox": [ + 511, + 460, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 534, + 633, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we categorize perception tasks into four groups based on two dimensions: prediction type and instruction type. We observe that current works typically address only a subset of the possible combinations. To address this limitation, we propose a unified Visual Large Language Model framework, MVP-LM, encompassing all these tasks. By incorporating a multi-granularity decoder and a CoT-inspired dataset unification strategy, our model enables joint training across multiple datasets and adopts a \"thinking-then-perceiving\" paradigm. Furthermore, we design a query enhancement strategy to exploit the decoding and generative capabilities of VLLM. Extensive experiments demonstrate the potential across various benchmarks. Considering the improvement brought by our CoT-inspired data curation, we would investigate the application of R1-like training in the realm of perception in our future work.", + "bbox": [ + 511, + 559, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 815, + 692, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903), the Shenzhen Science and Technology Program (JSGG20220831093004008).", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22160", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3", + "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4974-4983, 2019. 1", + "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185-24198, 2024. 1, 2, 3", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 6", + "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 7", + "[9] Henghui Ding, Scott Cohen, Brian Price, and Xudong Jiang. Phraseclick: toward achieving flexible interactive segmentation by phrase and click. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 417-435. Springer, 2020. 3", + "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 6", + "[11] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. 2023. 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6, 7", + "[13] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R. Scott, and Weilin Huang. TOOD: Task-aligned One-stage Object Detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490-3499, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 2", + "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In Computer Vision - ECCV 2022, pages 701-717, Cham, 2022. Springer Nature Switzerland. 1", + "[15] Chengjian Feng, Yujie Zhong, Zequn Jie, Weidi Xie, and Lin Ma. Instagen: Enhancing object detection by training on synthetic dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14121-14130, 2024. 2", + "[16] Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms, 2025. 3", + "[17] Shenghao Fu, Junkai Yan, Qize Yang, Xihan Wei, Xiaohua Xie, and Wei-Shi Zheng. Frozen-detr: Enhancing detr with image understanding from frozen foundation models. In NeurIPS, 2024. 2", + "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2021. 2", + "[19] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[20] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 3, 6", + "[21] Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: An Imm perceiving any aspect ratio and high-resolution images. In Computer Vision - ECCV 2024, pages 390-406, Cham, 2025. Springer Nature Switzerland. 3", + "[22] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1086-1096, 2023. 2, 6", + "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1", + "[24] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22161", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3", + "[25] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation, 2025.", + "[26] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020.", + "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 3", + "[28] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 6", + "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 1, 2", + "[30] Gaoussou Youssouf Kebe, Padraig Higgins, Patrick Jenkins, Kasra Darvish, Rishabh Sachdeva, Ryan Barron, John Winder, Donald Engel, Edward Raff, Francis Ferraro, and Cynthia Matuszek. A spoken language dataset of descriptions for speech-based grounded language learning. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6", + "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2", + "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 1, 6", + "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3", + "[34] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 1", + "[35] Shiyi Lan, Xitong Yang, Zhiding Yu, Zuxuan Wu, Jose M Alvarez, and Anima Anandkumar. Vision transformers are good mask auto-labelers. arXiv preprint arXiv:2301.03992, 2023. 1", + "[36] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 1", + "[37] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2", + "[38] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 1, 2", + "[39] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In Computer VisionECCV 2022:17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 1-18. Springer, 2022. 1", + "[40] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE transactions on pattern analysis and machine intelligence, 45(4):4552-4568, 2022. 2", + "[41] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3", + "[42] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6", + "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 3, 6, 7", + "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6", + "[45] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 1, 6", + "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2", + "[47] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Li Su, and Qingming Huang. Knowledge-guided pairwise reconstruction network for weakly supervised referring expression" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22162", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "grounding. In Proceedings of the 27th ACM International Conference on Multimedia, pages 539-547, 2019. 3", + "[48] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 3", + "[49] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6", + "[50] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning, 2025. 3", + "[51] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 3", + "[52] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983-73007, 2023. 1", + "[53] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6", + "[54] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6", + "[55] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7", + "[56] Hai Nguyen-Truong, E-Ro Nguyen, Tuan-Anh Vu, Minh-Triet Tran, Binh-Son Hua, and Sai-Kit Yeung. Vision-aware text features in referring image segmentation: From object understanding to context understanding, 2024. 3", + "[57] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 1, 2, 3", + "[58] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 1, 6", + "[59] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2, 3", + "[60] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3", + "[61] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6", + "[62] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. 3", + "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6", + "[64] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1", + "[65] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 2", + "[66] Hao Wang, Pengzhen Ren, Zequn Jie, Xiao Dong, Chengjian Feng, Yinlong Qian, Lin Ma, Dongmei Jiang, Yaowei Wang, Xiangyuan Lan, et al. Ov-dino: Unified open-vocabulary detection with language-aware selective fusion. arXiv preprint arXiv:2407.07844, 2024. 1, 2", + "[67] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 3", + "[68] Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, Xiangnan Fang, Zewen He, Zhenbo Luo, Wenxuan Wang, Junqi Lin, Jian Luan, and Qin Jin. Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding. arXiv e-prints, art. arXiv:2503.13377, 2025. 3", + "[69] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6", + "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22163", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Jie Hu, Dengjie Li, Zheng Zhao, and Yujiu Yang. Hyperseg: Hybrid segmentation assistant with fine-grained visual perceiver. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8931-8941, 2025. 1", + "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3", + "[73] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2", + "[74] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 6", + "[75] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019. 2", + "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 6", + "[77] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3", + "[78] Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, Bo Zhang, and Wei Chen. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025. 3", + "[79] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 6", + "[80] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6", + "[81] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7", + "[82] Chang-Bin Zhang, Jinhong Ni, Yujie Zhong, and Kai Han. v-clr: View-consistent learning for open-world instance segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 20307-20317, 2025. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Chang-Bin Zhang, Yujie Zhong, and Kai Han. Mr. detr: Instructive multi-route training for detection transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 9933-9943, 2025. 2", + "[84] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4158-4166, 2018. 3", + "[85] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1, 2, 4, 6", + "[86] Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via stepwise group relative policy optimization, 2025. 3", + "[87] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 1, 2, 6", + "[88] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2025. 1, 2, 3, 6, 7", + "[89] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6", + "[90] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 1, 2", + "[91] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 656 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22164", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_model.json b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_model.json new file mode 100644 index 0000000000000000000000000000000000000000..378486fa385573d10c078e758dcaae9339354999 --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_model.json @@ -0,0 +1,2556 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.145, + 0.13, + 0.853, + 0.176 + ], + "angle": 0, + "content": "Advancing Visual Large Language Model for Multi-granular Versatile Perception" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.203, + 0.858, + 0.256 + ], + "angle": 0, + "content": "Wentao Xiang\\(^{1*}\\), Haoxian Tan\\(^{2*}\\), Yujie Zhong\\(^{2\\dagger}\\), Cong Wei\\(^{1}\\), Dengjie Li\\(^{2}\\), and Yujiu Yang\\(^{1\\dagger}\\) \n\\(^{1}\\) Tsinghua University \n\\(^{2}\\) Meituan Inc." + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.26, + 0.828, + 0.275 + ], + "angle": 0, + "content": "thu_xiangwentao@163.com, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.485, + 0.705 + ], + "angle": 0, + "content": "Perception is a fundamental task in the field of computer vision, encompassing a diverse set of subtasks that can be systematically categorized into four distinct groups based on two dimensions: prediction type and instruction type. Notably, existing researches often focus solely on a limited subset of these potential combinations, which constrains their applicability and versatility across various contexts. In response to this challenge, we present MVP-LM, a Multi-granular and Versatile Perception framework incorporating Visual Large Language Model. Our framework is designed to integrate both word-based and sentence-based perception tasks alongside box and mask predictions within a single architecture. MVP-LM features an innovative multi-granularity decoder in conjunction with a CoT-inspired dataset unification strategy, enabling seamless supervised fine-tuning across a wide spectrum of tasks, including but not limited to panoptic segmentation, detection, grounding, and referring expression segmentation. Furthermore, we introduce a query enhancement strategy aimed at harnessing the decoding and generative capabilities inherent in VLLMs. Extensive experiments conducted across a range of benchmarks in both word-based and sentence-based perception tasks substantiate the efficacy of our framework. The code will be available here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.222, + 0.748 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Perception constitutes a fundamental task within the field of computer vision, necessitating that models accurately identify and locate objects within images or videos in accordance with specified instructions. These perception tasks can be categorized along two dimensions: prediction type (box vs. mask) and instruction type (word-based vs. sentence-based). This classification yields four different groups." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.303, + 0.906, + 0.344 + ], + "angle": 0, + "content": "Table 1. The comparison of capabilities of different methods. Current works can only address a subset of the combinations, while our MVP-LM can cover all tasks." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.35, + 0.902, + 0.504 + ], + "angle": 0, + "content": "
MethodOutput TypeInstruction Type
BoxMaskWordSentence
Specialists
OWL[52], PromptDet[14], OV-DINO[66]✓✓
X-Decoder[90], OpenSeeD[85], Mask DINO[36]✓✓✓
MDETR[29], GLIP[38], Grounding DINO[46]✓✓✓
VLT[10], LAVT[79], PolyFormer[45]✓✓
VLLM-based Generalists
QwenVL[1], InternVL[6], DeepseekVL[51]✓✓
LISA[32], PixelLM[58], Glamm[57]✓✓
OMG-LLaVA[87], PSALM[88], HyperSeg[71]✓✓✓
MVP-LM (Ours)✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.519, + 0.907, + 0.791 + ], + "angle": 0, + "content": "The distinction between coordinate-level prediction and pixel-level prediction has been extensively explored in the literature. Given that mask annotations can be converted to box annotations without loss of information, the concurrent optimization of box predictions and mask predictions within the same dataset has emerged as a prominent training paradigm. This approach has demonstrated mutual benefits for both prediction types [23, 36]. In contrast to the costly annotations required for segmentation tasks, box annotations are significantly easier to obtain, resulting in a considerably larger dataset. Consequently, another line of research focuses on enhancing segmentation performance by leveraging additional box-annotated data. Numerous methods [4, 23, 34, 35, 39, 64, 85] have been developed to achieve this, ranging from the utilization of pre-trained models to natively annotate box data, to the design of specialized segmentation loss functions tailored for box annotations." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Word-based perception employs individual words to denote the categories of targets, which often results in a degree of vagueness and ambiguity. A category name can correspond to none or multiple objects within an image. In contrast, sentence-based perception utilizes complete sentences to describe the objects of interest, resulting in greater precision. In most cases, only one object within the im" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.374, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22153" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "age satisfies the specified conditions. A common approach in the literature [29, 38, 46] involves the deployment of a two-stream framework, wherein visual queries extracted by any vision backbone are utilized to decode object locations and compute similarity with the text features generated by a language encoder to identify the objects. However, many of these methods overlook the underlying semantic concepts inherent in the descriptions, resulting in suboptimal outcomes, particularly for sentence-based perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Recently, Large Language Models (LLMs) have demonstrated a remarkable ability to comprehend user queries and generate appropriate responses, positioning them as a viable option for sentence-based perception [1, 5, 6, 33, 51, 57, 59]. Additionally, LLMs possess the capability to extract language features through forwarding, which aligns well with word-based perception. Therefore, LLMs reveal substantial potential for unifying both perceptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.532 + ], + "angle": 0, + "content": "The representative of sentence-based location tasks, Referring Expression Comprehension (REC), has emerged as the standard benchmark for Visual Large Language Models (VLLMs), as box coordinates can be expressed as a sequence of numbers in text form, rendering them compatible with traditional VLLMs. However, many existing models [1, 6, 51] encounter difficulties in achieving pixel-level predictions. Recent advances [5, 33, 57, 59] have successfully output masks by incorporating an additional mask decoder. Nevertheless, these approaches have not adequately addressed the challenges associated with word-based perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Alternatively, other studies [87, 88] have demonstrated comparable performance on standard word-based perception benchmarks by employing the VLLM as a versatile decoder. Nonetheless, these models focus solely on pixel-level prediction and overlook box prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.672 + ], + "angle": 0, + "content": "As illustrated in Tab. 1, although joint training has been examined for certain combinations of tasks, the collaborative effects of joint training across all four groups have not been thoroughly explored in the existing literature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.884 + ], + "angle": 0, + "content": "To this end, we introduce a Multi-granular and Versatile Perception framework incorporating VLLM, termed MVPLM, which is capable of detecting and segmenting targeted objects in images according to various types of user instructions within a single model. By employing a Chain-of-Thought (CoT)-inspired data curation [72], we transform multiple box- and mask-annotated datasets into a unified supervised fine-tuning (SFT) dataset format, thereby accommodating a diverse set of tasks, including panoptic segmentation, object detection, visual grounding, and referring expression segmentation. Furthermore, to leverage the decoding and generative capabilities of LLMs, we enhance the queries for object identification and localization by incorporating features derived from LLM-generated sequences." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In summary, our contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "- We introduce MVP-LM, a framework that harnesses both the decoding and generative capabilities of VLLM to perform both word-based and sentence-based perception tasks across varying granularities." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We have developed a CoT-inspired data curation that unifies datasets from diverse tasks into a single SFT dataset, thereby encouraging the model to adopt a \"thinking-then-perceiving\" paradigm." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.213, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- MVP-LM demonstrates competitive performance across various benchmarks in both word-based and sentence-based perception, thereby validating the potential of our framework. Notably, our MVP-LM achieves remarkable results on both closed-set and open-set tasks compared with other specialists and VLLMs." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.317, + 0.655, + 0.334 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.905, + 0.418 + ], + "angle": 0, + "content": "Perception in computer vision can be categorized into four groups along two dimensions: prediction type and instruction type. Below, we review both the related work within each of these four groups and CoT, and highlight the gaps MVP-LM aims to fill." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.426, + 0.731, + 0.443 + ], + "angle": 0, + "content": "2.1. Word-based Preception" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Detection and Grounding. Word-based location tasks encompass object detection, grounding, and instance detection, focusing on identifying and localizing objects via bounding boxes with categorical labels. Conventional frameworks [13-15, 17, 18, 31, 37, 73, 83] utilize CLIP-like encoders to compute visual-linguistic similarities, enabling zero-shot capabilities. Significant advances include GLIP [38], which unifies phrase grounding with object detection, and OV-DINO [66], which implements an end-to-end training pipeline for enhanced detection performance and generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.615, + 0.905, + 0.796 + ], + "angle": 0, + "content": "Segmentation. Word-based segmentation tasks, encompassing panoptic, semantic, and instance segmentation, demand precise pixel-level predictions. However, generating such masks from textual inputs remains underexplored, largely due to the scarcity of fine-grained annotated data. Existing methods [2, 22, 40, 65, 75, 82, 85, 90] primarily align word embeddings with visual features, resulting in coarse segmentation masks. Nevertheless, these approaches [85, 90] often demonstrate limited textual understanding and lack the requisite fine-grained control for high-quality mask generation, which ultimately restricts their performance in both closed-set and open-set tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.804, + 0.755, + 0.82 + ], + "angle": 0, + "content": "2.2. Sentence-based Preception" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Referring Expression Comprehension. Sentence-based location tasks, which require a deeper multi-modality understanding, have become a standard benchmark for VLLMs. Specifically, Referring Expression Comprehension (REC) is often the preferred task, as box coordinates" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22154" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.862, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.292, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 1. Overview of MVP-LM. MVP-LM implements perception by integrating a multi-granularity decoder into the existing VLLM framework. We utilize a unified prompt template to construct the input sequence for the LLM across different tasks. The base query is derived from the summary token of the generated response. Concurrently, we extract the instruction embeddings from the input sequence (denoted by the same color) and select the corresponding residuals from the multi-scale visual features based on these embeddings. After aggregating the base query vector with the residuals, we decode the bounding box and segmentation mask by inputting the resulting queries into the multi-granularity decoder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.482, + 0.476 + ], + "angle": 0, + "content": "can be represented as a text sequence of numbers, making them compatible with conventional VLLMs [1, 6, 41, 47, 48, 51, 60, 67, 77, 84]. However, many existing models struggle to achieve pixel-level predictions, and even more so, to facilitate joint training with sentence-based mask perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Referring Expression Segmentation. Sentence-based segmentation tasks demand a fine-grained understanding of descriptive sentences to accurately generate segmentation masks for target objects. Traditional approaches [3, 9, 24-27, 56] have made strides in fusing visual and textual features to improve segmentation performance, but often fall short in fully leveraging the richness of sentence descriptions, resulting in suboptimal performance in both closed-set and open-set scenarios. Recent efforts [5, 33, 43, 57, 59] have successfully implemented Referring Expression Segmentation (RES) by incorporating an additional mask decoder. However, these approaches fall short in addressing word-based perception." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.267, + 0.723 + ], + "angle": 0, + "content": "2.3. Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Chain-of-Thought prompting [72] was initially introduced during inference to enhance the reasoning capabilities of LLMs. Subsequently, CoT has evolved into a prevalent paradigm in the post-training phase, wherein models first undergo supervised fine-tuning on CoT datasets, followed by reinforcement learning [62]. A similar trend has also been observed in VLLMs, which have demonstrated promising performance in multi-modal understanding and reasoning tasks. Various meticulously designed rewards [16, 21, 50, 68, 78, 86], ranging from format to geometry, have been proposed to handle diverse scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.843, + 0.401 + ], + "angle": 0, + "content": "2.4. Identified Gaps and Our Contribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.53 + ], + "angle": 0, + "content": "While significant progress has been made in addressing some of the combinations [6, 33, 51, 59, 88], a comprehensive solution effectively handling all four groups remains elusive. MVP-LM bridges this gap by offering a solution capable of simultaneous perception across all four combinations, marking a significant advancement in the field, addressing the current limitations, and setting a new benchmark for multi-dimensional visual perception." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.649, + 0.566 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.636 + ], + "angle": 0, + "content": "This section provides an in-depth description of our proposed methodology. We begin by detailing the architecture of our MVP-LM in Sec. 3.1, followed by an exposition of our strategies for dataset processing in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.783, + 0.666 + ], + "angle": 0, + "content": "3.1. MVP-LM Architecture Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.84 + ], + "angle": 0, + "content": "System Overview. As illustrated in Fig. 1, our architecture comprises four essential components: an image encoder (Swin-Base Transformer [49]), a connector module, an LLM (Phi-1.5 [20]), and a multi-granularity decoder. The image encoder extracts rich visual features from input images, serving as the foundation for subsequent processing. The connector module facilitates seamless integration between visual and textual modalities by aligning features from the image encoder with the token embeddings for the LLM. The multi-granularity decoder is responsible for producing both bounding boxes and segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dynamic Query Generation. The dynamic generation of visual queries is pivotal for the model's ability to handle diverse perception tasks effectively. As depicted in Fig. 2, each visual query is composed of two components: a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.46, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Dynamic query selection. MVP-LM expands the embedding (hidden states) of the summary token into \\(N\\) base query vectors. Concurrently, multi-scale visual features are flattened, and their similarity with the instruction embedding from the input sequence is computed. The top \\(N\\) similar features are then selected and integrated with the corresponding \\(N\\) base query vectors, resulting in \\(N\\) final queries tailored for multi-granularity perception tasks. This dynamic selection mechanism ensures that the most relevant visual features are incorporated into each query, enhancing the precision and adaptability of the decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.483, + 0.387 + ], + "angle": 0, + "content": "context-aware base query vector and language-guided residuals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.478 + ], + "angle": 0, + "content": "Initially, MVP-LM leverages the VLLM to generate a specialized summary token that encapsulates the essential information from the input sequences. The hidden states corresponding to this summary token are projected into the base query vectors using an MLP, providing a robust foundation for query formulation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Simultaneously, we compute the similarity between multi-scale visual features and the text embedding derived from the input instructions (i.e., category names for word-based perception and referring expressions for sentence-based perception). To facilitate this comparison, an MLP aligns the dimensionality of the text embeddings with that of the multi-scale visual features. The top \\( N \\) most similar visual features are then identified, ensuring that the queries are informed by the most relevant visual information. These selected features are integrated with the corresponding base query vectors, resulting in \\( N \\) refined final queries that encapsulate both contextual and visual clues. This dynamic query generation process enables MVP-LM to adaptively focus on relevant regions of the image, thereby enhancing the model's ability to generate accurate bounding boxes and segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multi-granularity Decoder. The multi-granularity decoder is the cornerstone of our architecture, facilitating the simultaneous output of bounding boxes and segmentation masks. Our decoder is inspired by the design principles of OpenSeeD [85], which emphasizes the profound integration of multi-scale features and comprehensive query processing. Initially, a set of content queries and their corresponding reference points are generated through a query selection mechanism as seen in Fig. 2. These queries are then iteratively cross-attended with multi-scale visual features via deformable attention layers, enabling the decoder to capture both global contextual information and detailed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "local features. The outputs from each MSDeform layer are subsequently processed by three shared heads—a cross-modal similarity computation head, a box head, and a mask head—to produce predictions. The unified output streamlines the training process and allows the model to learn shared representations that benefit both detection and segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.908, + 0.289 + ], + "angle": 0, + "content": "Training Objectives. To fully exploit the capabilities of MVP-LM, we employ a joint training approach that simultaneously optimizes multiple tasks across diverse datasets. The training objective is formulated as a composite loss function comprising several components:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.299, + 0.843, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {l l m} + \\lambda_ {w o r d} \\mathcal {L} _ {w o r d} + \\lambda_ {s e n t} \\mathcal {L} _ {s e n t} \\\\ + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {b o x} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{llm}\\) represents the standard loss for next-token prediction, ensuring that the LLM accurately generates coherent and contextually appropriate summaries. The term \\(\\mathcal{L}_{word}\\) corresponds to the Cross-Entropy (CE) loss for word-based perception conditions, facilitating precise classification based on single-word instructions. Conversely, \\(\\mathcal{L}_{sent}\\) denotes the Binary Cross Entropy (BCE) loss for sentence-based perception conditions, enabling the model to handle more complex, descriptive instructions effectively. The mask loss, \\(\\mathcal{L}_{mask}\\), is composed of a pixel-level BCE loss and a Dice loss, which together promote accurate and smooth segmentation masks by penalizing discrepancies between predicted and ground-truth masks at both the pixel and region levels. The box loss, \\(\\mathcal{L}_{box}\\), includes the \\(l1\\) loss to minimize the absolute differences between predicted and ground-truth bounding box coordinates, as well as the Generalized Intersection over Union (GIoU) loss to enhance the alignment and overlap between predicted and actual boxes. Formally, these loss components are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.659, + 0.821, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\\\ \\mathcal {L} _ {b o x} = \\lambda_ {l 1} \\mathcal {L} _ {l 1} + \\lambda_ {g i o u} \\mathcal {L} _ {g i o u}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Following the methodology of OpenSeeD, we employ bipartite matching to match the model's predictions with their corresponding ground truth annotations during training. Predictions from each layer are assigned to their corresponding annotations to calculate losses through Hungarian matching. Moreover, a denoising strategy is incorporated, similar to that used in MaskDINO, to stabilize optimization and accelerate convergence. However, during inference, only the final layer's output is utilized." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.663, + 0.864 + ], + "angle": 0, + "content": "3.2. Data Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "CoT-inspired Dataset Unification. We unified heterogeneous datasets into an SFT dataset specially tailored for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "the training of VLLM. For each sample in the perception datasets, we construct a question-answer pair based on the original annotations. Each training pair consists of three key components: a task description, an instruction, and a response that embodies a \"thinking-then-perceiving\" process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.482, + 0.29 + ], + "angle": 0, + "content": "The task description specifies the particular task associated with the data. In the case of word-based perception, the description may be articulated as follows: \"Please identify all objects according to the given phrase list. This is all the candidate phrases.\" Conversely, for sentence-based perception, the description is framed as: \"Please identify the target according to the following instruction.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.482, + 0.473 + ], + "angle": 0, + "content": "The instruction varies according to the specific tasks involved. For word-based perception, we concatenate the provided word lists into a single sentence, using a comma as a separator. For instance, given three categories: apple, banana, and orange, the resultant sentence would be \"apple, banana, orange.\" After passing this sentence through VLLM, we select the corresponding output embeddings and apply average pooling to obtain the text embedding for each word, while for sentence-based perception, the referring expression directly serves as the final instruction, requiring no additional processing. The text embedding is derived simply by applying pooling over the entire expression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.482, + 0.611 + ], + "angle": 0, + "content": "The response format is standardized across various tasks. We adopt a \"thinking-then-perceiving\" strategy by preponding the input image caption to the summary, thereby encouraging the VLLM to effectively capture image details. Specifically, the response is structured as follows: \"[image caption]. The perception result is \\(<\\text{PER}>\\)\", where \"<\\text{PER}>\" denotes a special token representing the summary. The hidden states associated with this summary token are subsequently utilized to derive the visual queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.482, + 0.764 + ], + "angle": 0, + "content": "Multi-caption Auto-labeling. To fully leverage the generative prowess of VLLMs, we design our system to generate comprehensive image captions in conjunction with summary tokens. This generation approach enriches the contextual information available to the model, facilitating a deeper understanding of the visual content. However, a significant challenge arises as the majority of our training datasets lack image captions. To address this, we integrate open-source VLLMs to autonomously generate descriptive captions for each image prior to the training phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Specifically, we employ a diverse set of VLLMs, including VILA-3B, VILA-13B, InternVL2-8B, and InternVL2-26B, to produce multiple unique captions for the same image. The utilization of distinct VLLMs not only enhances the accuracy of the generated captions but also introduces substantial diversity, mitigating the risk of model overfitting to a single captioning style or vocabulary. By incorporating multi-caption auto-labeling, our approach ensures a more comprehensive and resilient training paradigm, effec" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.083, + 0.905, + 0.126 + ], + "angle": 0, + "content": "Table 2. Hyperparameters for both training stages. \"CP\", \"RC\", \"O\", and \"GG\" denote COCO-Panoptic, RefCOCO, Objects365, and GoldG, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.13, + 0.905, + 0.292 + ], + "angle": 0, + "content": "
ParametersStage1Stage2
Training ComponentsConnectorConnector + LLM + Multi-granularity Decoder
OptimizerAdamW
Training Rate2 × 10-34 × 10-5
Batch Size12864
Number of Steps465080000
Learning Rate ScheduleCosine Decay
Weight Decay0.00.05
Warmup Ratio0.03
Training DataCC3MCP(33.3%) RC(33.3%) O(16.7%) GG(16.7%)
LossLLLMLLLM + 2 · Lword/sent + 5 · LL1 + 2 · LGIoU + 5 · LBCE + 5 · LDICE
Image Size1024 × 1024
Image ProcessingResize longer to 1024 and pad shorter to 1024
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.905, + 0.339 + ], + "angle": 0, + "content": "tively bridging the gap between visual inputs and linguistic descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.906, + 0.512 + ], + "angle": 0, + "content": "Data Refinement. During the data generation process, the occurrence of irregular or erroneous data is inevitable. Consequently, we implement a straightforward data-cleaning procedure for the dataset. First, we utilize the length and information entropy of the generated captions to filter out meaningless results. Second, we remove sentences containing suggestive keywords such as \"indicate\", \"might\", \"may\", and \"imply\", as well as other similar terms, to mitigate the hallucinations that are inherent in large language models. These two mechanisms ensure the quality of the generated captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.653 + ], + "angle": 0, + "content": "Randomness in Training Pairs. The variability of training pairs is essential for preventing the model from learning spurious correlations and for enhancing its generalization capabilities. In our framework, such diversity is achieved by randomly selecting one of four generated captions for each training response. However, fixed category name ordering in word-based datasets can still introduce spurious causal relationships, leading to overfitting and impaired performance on unseen data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To counteract these issues, our mitigation strategy is twofold. Firstly, we introduce randomness by shuffling the order of the word list in each iteration. This disrupts fixed positional biases, preventing the model from relying on word order for predictions. Secondly, word lists are dynamically constructed by randomly incorporating a subset of negative category names (categories absent from the image). This selective inclusion diversifies training prompts, proving particularly beneficial for datasets with extensive vocabularies. By carefully balancing positive and negative categories within VLLM input sequence length constraints, the model's ability to discern relevant from irrelevant objects is enhanced. This approach fosters a more adaptive learning process, significantly reducing overfitting and ensuring robust performance across both closed-set and open-set detection scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22157" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.083, + 0.907, + 0.111 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation and open-set segmentation tasks. Our model MVP-LM achieves remarkable performance among previous VLLM-based methods." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.117, + 0.847, + 0.295 + ], + "angle": 0, + "content": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
SpecialistsMask2former [7]Swin-B55.165.1-----
OneFormer [28]Swin-L57.967.4-----
OpenSeeD [85]Swin-L59.568.619.723.441.4--
SEEM [91]DaViT-B56.166.3-----
MaskCLIP [11]ViT-L30.947.615.123.7-45.9-
SimBaseline [76]ViT-B---20.5-47.788.4
DeOP [22]ResNet-101c---22.9-48.891.7
DaTaSeg [19]ViTDet-B52.862.712.318.328.051.1-
VLLM-based GeneralistsOMG-LLaVA [87]ConvNeXt-L53.8------
PSALM [32]Swin-B55.966.613.718.228.848.581.3
MVP-LMSwin-B56.166.819.420.535.344.185.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.302, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Table 4. Comparison with the state-of-the-art models on the closed-set Referring Segmentation benchmarks (RefCOCO series) and more challenging Generalized Referring Expression Segmentation benchmark gRefCOCO[43]. Gray numbers denote the method using gRefCOCO for training." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.35, + 0.859, + 0.535 + ], + "angle": 0, + "content": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SpecialistsVLT [10]67.570.565.256.361.050.155.057.752.562.250.5
CRIS [69]70.573.266.162.368.153.759.960.455.363.851.0
LAVT [79]72.775.868.862.168.455.161.262.157.665.355.0
PolyFormer-B [45]74.876.671.167.672.959.367.869.1---
VLLM-based GeneralistsLISA-7B [32]74.176.571.162.467.456.566.468.538.752.644.8
PixelLM-7B [58]73.076.568.266.371.758.369.370.5---
GSVA-7B [74]76.477.472.864.567.758.671.172.061.769.260.3
LaSagnA-7B [70]76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [87]78.080.374.169.173.163.072.972.9---
PSALM [88]83.684.781.672.975.570.173.874.442.052.450.6
MVP-LM83.685.182.573.976.865.475.175.637.650.449.9
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.225, + 0.569 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Datasets. We train MVP-LM in a multi-dataset and multitask manner like [19, 88]. For generic perception, we use COCO Panoptic Segmentation [42](with 118K samples in total) and O365 [61](with 1.7M samples in total). We use RefCOCO series [54, 80](with 126K samples in total) for referring segmentation tasks while utilizing the GoldG [30](with 770K samples in total) dataset for grounding. For evaluation, we report the main results on conventional COCO image segmentation tasks (panoptic/semantic) and RefCOCO series. Besides, we conduct evaluation on more challenging perception tasks like open-set segmentation and detection [8, 12, 53, 89]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Evaluation Metrics. We report results in widely used evaluation metrics: panoptic quality (PQ) and mean intersection-over-union (mIoU) for panoptic and semantic segmentation, respectively. Additionally, we use cumulative Intersection-over-Union (cIoU) for referring segmentation evaluation consistent with previous studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. The model architecture integrates Phi1.5 [20] as the language model and Swin-B [49] as visual encoder, with our multi-granularity decoder building" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.735 + ], + "angle": 0, + "content": "on OpenSeeD [85]. For most tasks, we initialize 100 queries for subsequent mask decoding, and for panoptic segmentation, we introduce an additional 100 learnable queries specifically for stuff categories. Our training proceeds in two phases following LLaVA [44]: 1) visual-language alignment trains exclusively the connector on CC3M [63] while freezes other components; 2) subsequent fine-tuning stage optimizes the whole model except the vision encoder. Detailed hyperparameters and training configurations are listed in Tab. 2. Training spans 80k iterations on the aggregated datasets and 9k for ablation studies, which utilize only the COCO Panoptic and RefCOCO datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.649, + 0.758 + ], + "angle": 0, + "content": "4.1. Main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Word-based Perception Results. We present the performance of MVP-LM on the COCO-Panoptic [42] and open-set segmentation [8, 12, 53, 89] tasks as detailed in Tab. 3. MVP-LM demonstrates competitive performance when compared to both specialized models and VLLM-based methods in closed-set tasks. Notably, it exhibits exceptional performance in the realm of open-set segmentation within the context of VLLM. Specifically, when evaluated on the open-set ADE20K dataset [89], our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22158" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.083, + 0.483, + 0.111 + ], + "angle": 0, + "content": "Table 5. Results on referring expression comprehension tasks. DS-VL2 is the abbreviation for DeepSeek-VL2." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.117, + 0.482, + 0.203 + ], + "angle": 0, + "content": "
MethodParamsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra13B87.891.181.882.987.874.482.683.1
MiniGPTv27B88.791.785.380.085.174.584.484.7
Octopus7B89.092.683.483.589.476.084.486.2
MVP-LM1.3B93.594.591.684.988.879.386.787.4
DS-VL2200B+95.196.795.191.294.987.492.892.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.483, + 0.311 + ], + "angle": 0, + "content": "MVP-LM demonstrates substantial improvements over the VLLM-based PSALM [88] and achieves gains of 5.7 and 2.3 in PQ and mIoU metrics, respectively. Additionally, it surpasses PSALM by 6.5 and 4.4 on the mIoU metric for the open-set datasets Cityscapes [8] and PascalVOC-20 [12], respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.554 + ], + "angle": 0, + "content": "Sentence-based Perception Results. We evaluate MVP-LM against state-of-the-art methods on the referring expression segmentation (RES) benchmarks, RefCOCO+/g[55, 81] as well as the more challenging generalized referring expression segmentation benchmark, gRefCOCO[43] in Tab. 4. We also include RefCOCO series for referring expression comprehension (REC) in Tab. 5. In the domain of RES, MVP-LM achieves state-of-the-art performance on the RefCOCO and RefCOCOg datasets across all subsets. In the RefCOCO+ dataset, MVP-LM attains scores of 73.9 and 76.8 on val and testA, respectively. Moreover, our MVP-LM, only 1.3B parameters, outperforms listed 7B and 13B models, and achieves metrics that are second only to DeepSeek-VL2 with over 200B parameters in REC. These results highlight the robustness and generalization capabilities of MVP-LM when compared to previous methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.242, + 0.579 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.658 + ], + "angle": 0, + "content": "This section presents an ablation study investigating critical architectural decisions, specifically: multi-dataset joint learning, performance across diverse response configurations, the impact of query quantity, and the effect of query selection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Joint Training. The progressive incorporation of training data sources is analyzed in Tab. 6. Initial training on COCO and RefCOCO datasets yields a MVP-LM mIoU of 77.6 on the RefCOCO validation set. Subsequent inclusion of O365 significantly increases cIoU to 81.8, with further enhancement to 83.6 cIoU upon integrating the Grounding dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "For COCO evaluation, sequential integration of O365 yields a PQ of 55.8 and mIoU of 65.9. Subsequent integration of the Grounding dataset marginally improves PQ to 56.1 and mIoU to 66.8." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Crucially, sentence-based perception is enhanced by this multi-dataset approach by 6.0 points while word-based perception remains consistent. These findings empirically validate the effectiveness of our unified multi-task training strategy in coordinating multi-modal representations and significantly enhancing cross-task generalization." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.083, + 0.905, + 0.138 + ], + "angle": 0, + "content": "Table 6. Ablation of the datasets involved in the joint training. We use the abbreviations C, R, O, and G to represent COCO, Ref-COCO, O365, and Grounding, respectively. Both experiments are trained for 56k iterations." + }, + { + "type": "table", + "bbox": [ + 0.594, + 0.144, + 0.825, + 0.213 + ], + "angle": 0, + "content": "
DatasetRefCOCO val cIoUCOCO
PQmIoU
C, R77.656.466.3
C, R, O81.855.865.9
C, R, O, G83.656.166.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.221, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Table 7. Ablation of different response settings. 'None' denotes the basic response setting containing the summary token only. 'Existing Obj' means generating existing objects' names or descriptions in the image before summary tokens. 'Caption' indicates the regression of the image caption prior to summary tokens. \\(^\\dagger\\) denotes the direct use of similar visual features as the final queries without LLM-generated base query vectors." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.329, + 0.842, + 0.412 + ], + "angle": 0, + "content": "
Response settingRefCOCO val cIoUCOCO
PQmIoU
None75.655.365.7
Existing Obj.75.654.965.6
Caption75.755.666.2
Caption†75.955.066.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.642 + ], + "angle": 0, + "content": "Response Settings. We provide a comprehensive ablation study on various response settings in Tab. 7. The baseline response, \"The perception result is \\(\\)\", remains consistent across all inputs, achieving 75.6 cIoU on the RefCOCO validation set and 55.3 PQ/65.7 mIoU on COCO. When the LLM is instructed to output existing category names or object descriptions from the input image prior to the baseline response, the model maintains its performance on RefCOCO (75.6 cIoU) but exhibits a decline in COCO metrics, resulting in 54.9 PQ and 65.6 mIoU. Conversely, our final configuration, which directs the LLM to generate the image caption before delivering the baseline response, results in simultaneous improvements across all metrics (75.7 cIoU, 55.6 PQ, and 66.2 mIoU)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Additionally, we evaluate the performance of utilizing the most similar visual features as the final queries. Compared to the approach of combining these features with the base query vectors generated by the LLM, this configuration yields a modest improvement in the RefCOCO validation set (+0.2). However, it is accompanied by a noticeable decline in COCO PQ, resulting in a decrease of 0.6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.84 + ], + "angle": 0, + "content": "The incorporation of descriptive captions in the responses facilitates the mutual enhancement of both sentence-level (RefCOCO) and word-level (COCO) perception tasks. This observation verifies our hypothesis that structured semantic prompting significantly improves visual-language understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Number of Queries. As presented in Tab. 8, the response to variations in the number of queries differs between the RefCOCO and COCO datasets. Specifically, performance on the RefCOCO validation set deteriorates with an increasing" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22159" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.078, + 0.514, + 0.089 + ], + "angle": 0, + "content": "COCO" + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.091, + 0.849, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.395, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 3. Visualization of the segmentation results on COCO, ADE20K and gRefCOCO datasets. We plot the predicted mask with a confidence larger than 0.5 for a clear visualization. The text above each image in the gRefCOCO denotes the corresponding reference text." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.437, + 0.484, + 0.479 + ], + "angle": 0, + "content": "Table 8. Ablation of the number of queries. 100 queries is the default setting, considering the balance between performance and computational costs." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.491, + 0.399, + 0.56 + ], + "angle": 0, + "content": "
# queriesRefCOCO val cIoUCOCO
PQmIoU
3076.353.165.2
10075.755.666.2
30072.454.965.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.569, + 0.449, + 0.584 + ], + "angle": 0, + "content": "Table 9. Ablation of whether applying query selection." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.594, + 0.433, + 0.652 + ], + "angle": 0, + "content": "
Apply Query SelectionRefCOCO val cIoUCOCO
PQmIoU
✓75.755.666.2
78.354.465.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "number of queries, decreasing from 76.3 cIoU to 72.4 cIoU. We hypothesize that, due to the presence of only one target in the RefCOCO images, an increase in the number of candidates may complicate the model's task. In contrast, for the COCO dataset, the use of 100 queries yields optimal performance among the three settings evaluated. Additionally, it is important to note that an increased number of queries incurs additional computational costs. Consequently, we adopt 100 queries as our default setting to achieve an optimal balance between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Query Selection. As conveyed in Tab. 9, applying separate learnable query embeddings instead of query selection improves RefCOCO val results (from 75.7 to 78.3) but lowers COCO-Panoptic results (from 55.6 to 54.4). It reflects that RefCOCO's simpler targets suit fixed queries, and that COCO's complexity needs more flexible query selection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.696, + 0.455 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.522 + ], + "angle": 0, + "content": "We present visualizations for the COCO, ADE20K, and gRefCOCO datasets in Fig. 3. The results demonstrate that MVP-LM effectively generates precise masks and is proficient in identifying a comprehensive range of objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.634, + 0.55 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.907, + 0.803 + ], + "angle": 0, + "content": "In this study, we categorize perception tasks into four groups based on two dimensions: prediction type and instruction type. We observe that current works typically address only a subset of the possible combinations. To address this limitation, we propose a unified Visual Large Language Model framework, MVP-LM, encompassing all these tasks. By incorporating a multi-granularity decoder and a CoT-inspired dataset unification strategy, our model enables joint training across multiple datasets and adopts a \"thinking-then-perceiving\" paradigm. Furthermore, we design a query enhancement strategy to exploit the decoding and generative capabilities of VLLM. Extensive experiments demonstrate the potential across various benchmarks. Considering the improvement brought by our CoT-inspired data curation, we would investigate the application of R1-like training in the realm of perception in our future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.693, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903), the Shenzhen Science and Technology Program (JSGG20220831093004008)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22160" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4974-4983, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185-24198, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[9] Henghui Ding, Scott Cohen, Brian Price, and Xudong Jiang. Phraseclick: toward achieving flexible interactive segmentation by phrase and click. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 417-435. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.788, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[13] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R. Scott, and Weilin Huang. TOOD: Task-aligned One-stage Object Detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490-3499, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In Computer Vision - ECCV 2022, pages 701-717, Cham, 2022. Springer Nature Switzerland. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[15] Chengjian Feng, Yujie Zhong, Zequn Jie, Weidi Xie, and Lin Ma. Instagen: Enhancing object detection by training on synthetic dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14121-14130, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[16] Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[17] Shenghao Fu, Junkai Yan, Qize Yang, Xihan Wei, Xiaohua Xie, and Wei-Shi Zheng. Frozen-detr: Enhancing detr with image understanding from frozen foundation models. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[19] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[20] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[21] Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: An Imm perceiving any aspect ratio and high-resolution images. In Computer Vision - ECCV 2024, pages 390-406, Cham, 2025. Springer Nature Switzerland. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[22] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1086-1096, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22161" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[25] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.244 + ], + "angle": 0, + "content": "[26] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.247, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[28] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.387, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[30] Gaoussou Youssouf Kebe, Padraig Higgins, Patrick Jenkins, Kasra Darvish, Rishabh Sachdeva, Ryan Barron, John Winder, Donald Engel, Edward Raff, Francis Ferraro, and Cynthia Matuszek. A spoken language dataset of descriptions for speech-based grounded language learning. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.567, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.61, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[34] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[35] Shiyi Lan, Xitong Yang, Zhiding Yu, Zuxuan Wu, Jose M Alvarez, and Anima Anandkumar. Vision transformers are good mask auto-labelers. arXiv preprint arXiv:2301.03992, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[37] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[38] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[39] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In Computer VisionECCV 2022:17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 1-18. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.904, + 0.446 + ], + "angle": 0, + "content": "[40] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE transactions on pattern analysis and machine intelligence, 45(4):4552-4568, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[41] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[45] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Li Su, and Qingming Huang. Knowledge-guided pairwise reconstruction network for weakly supervised referring expression" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22162" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "grounding. In Proceedings of the 27th ACM International Conference on Multimedia, pages 539-547, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[48] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[49] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[50] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[51] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[52] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983-73007, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[53] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[54] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[55] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[56] Hai Nguyen-Truong, E-Ro Nguyen, Tuan-Anh Vu, Minh-Triet Tran, Binh-Son Hua, and Sai-Kit Yeung. Vision-aware text features in referring image segmentation: From object understanding to context understanding, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[57] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[58] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[59] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[60] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[61] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[62] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[64] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[65] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[66] Hao Wang, Pengzhen Ren, Zequn Jie, Xiao Dong, Chengjian Feng, Yinlong Qian, Lin Ma, Dongmei Jiang, Yaowei Wang, Xiangyuan Lan, et al. Ov-dino: Unified open-vocabulary detection with language-aware selective fusion. arXiv preprint arXiv:2407.07844, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[67] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[68] Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, Xiangnan Fang, Zewen He, Zhenbo Luo, Wenxuan Wang, Junqi Lin, Jian Luan, and Qin Jin. Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding. arXiv e-prints, art. arXiv:2503.13377, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[69] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22163" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[71] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Jie Hu, Dengjie Li, Zheng Zhao, and Yujiu Yang. Hyperseg: Hybrid segmentation assistant with fine-grained visual perceiver. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8931-8941, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[73] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[74] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[75] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[77] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[78] Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, Bo Zhang, and Wei Chen. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[79] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[80] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[81] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[82] Chang-Bin Zhang, Jinhong Ni, Yujie Zhong, and Kai Han. v-clr: View-consistent learning for open-world instance segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 20307-20317, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[83] Chang-Bin Zhang, Yujie Zhong, and Kai Han. Mr. detr: Instructive multi-route training for detection transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 9933-9943, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[84] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4158-4166, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[85] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[86] Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via stepwise group relative policy optimization, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[87] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[88] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[89] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[90] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[91] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22164" + } + ] +] \ No newline at end of file diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_origin.pdf b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1ad83e9431732ed48efecf45938dc9d2925523d1 --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/2e0e9703-37e8-4376-9e70-3d183c7f0c36_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b418b25476eaabfb2362dbc8abf4590fce8ab57c97ec66e63976e848980ae00 +size 10130596 diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/full.md b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b5dd796992e59dab764f3cf9fcaff4622dd1b2b2 --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/full.md @@ -0,0 +1,323 @@ +# Advancing Visual Large Language Model for Multi-granular Versatile Perception + +Wentao Xiang $^{1*}$ , Haoxian Tan $^{2*}$ , Yujie Zhong $^{2\dagger}$ , Cong Wei $^{1}$ , Dengjie Li $^{2}$ , and Yujiu Yang $^{1\dagger}$ $^{1}$ Tsinghua University + $^{2}$ Meituan Inc. + +thu_xiangwentao@163.com, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn + +# Abstract + +Perception is a fundamental task in the field of computer vision, encompassing a diverse set of subtasks that can be systematically categorized into four distinct groups based on two dimensions: prediction type and instruction type. Notably, existing researches often focus solely on a limited subset of these potential combinations, which constrains their applicability and versatility across various contexts. In response to this challenge, we present MVP-LM, a Multi-granular and Versatile Perception framework incorporating Visual Large Language Model. Our framework is designed to integrate both word-based and sentence-based perception tasks alongside box and mask predictions within a single architecture. MVP-LM features an innovative multi-granularity decoder in conjunction with a CoT-inspired dataset unification strategy, enabling seamless supervised fine-tuning across a wide spectrum of tasks, including but not limited to panoptic segmentation, detection, grounding, and referring expression segmentation. Furthermore, we introduce a query enhancement strategy aimed at harnessing the decoding and generative capabilities inherent in VLLMs. Extensive experiments conducted across a range of benchmarks in both word-based and sentence-based perception tasks substantiate the efficacy of our framework. The code will be available here. + +# 1. Introduction + +Perception constitutes a fundamental task within the field of computer vision, necessitating that models accurately identify and locate objects within images or videos in accordance with specified instructions. These perception tasks can be categorized along two dimensions: prediction type (box vs. mask) and instruction type (word-based vs. sentence-based). This classification yields four different groups. + +Table 1. The comparison of capabilities of different methods. Current works can only address a subset of the combinations, while our MVP-LM can cover all tasks. + +
MethodOutput TypeInstruction Type
BoxMaskWordSentence
Specialists
OWL[52], PromptDet[14], OV-DINO[66]✓✓
X-Decoder[90], OpenSeeD[85], Mask DINO[36]✓✓✓
MDETR[29], GLIP[38], Grounding DINO[46]✓✓✓
VLT[10], LAVT[79], PolyFormer[45]✓✓
VLLM-based Generalists
QwenVL[1], InternVL[6], DeepseekVL[51]✓✓
LISA[32], PixelLM[58], Glamm[57]✓✓
OMG-LLaVA[87], PSALM[88], HyperSeg[71]✓✓✓
MVP-LM (Ours)✓✓✓✓
+ +The distinction between coordinate-level prediction and pixel-level prediction has been extensively explored in the literature. Given that mask annotations can be converted to box annotations without loss of information, the concurrent optimization of box predictions and mask predictions within the same dataset has emerged as a prominent training paradigm. This approach has demonstrated mutual benefits for both prediction types [23, 36]. In contrast to the costly annotations required for segmentation tasks, box annotations are significantly easier to obtain, resulting in a considerably larger dataset. Consequently, another line of research focuses on enhancing segmentation performance by leveraging additional box-annotated data. Numerous methods [4, 23, 34, 35, 39, 64, 85] have been developed to achieve this, ranging from the utilization of pre-trained models to natively annotate box data, to the design of specialized segmentation loss functions tailored for box annotations. + +Word-based perception employs individual words to denote the categories of targets, which often results in a degree of vagueness and ambiguity. A category name can correspond to none or multiple objects within an image. In contrast, sentence-based perception utilizes complete sentences to describe the objects of interest, resulting in greater precision. In most cases, only one object within the im + +age satisfies the specified conditions. A common approach in the literature [29, 38, 46] involves the deployment of a two-stream framework, wherein visual queries extracted by any vision backbone are utilized to decode object locations and compute similarity with the text features generated by a language encoder to identify the objects. However, many of these methods overlook the underlying semantic concepts inherent in the descriptions, resulting in suboptimal outcomes, particularly for sentence-based perception. + +Recently, Large Language Models (LLMs) have demonstrated a remarkable ability to comprehend user queries and generate appropriate responses, positioning them as a viable option for sentence-based perception [1, 5, 6, 33, 51, 57, 59]. Additionally, LLMs possess the capability to extract language features through forwarding, which aligns well with word-based perception. Therefore, LLMs reveal substantial potential for unifying both perceptions. + +The representative of sentence-based location tasks, Referring Expression Comprehension (REC), has emerged as the standard benchmark for Visual Large Language Models (VLLMs), as box coordinates can be expressed as a sequence of numbers in text form, rendering them compatible with traditional VLLMs. However, many existing models [1, 6, 51] encounter difficulties in achieving pixel-level predictions. Recent advances [5, 33, 57, 59] have successfully output masks by incorporating an additional mask decoder. Nevertheless, these approaches have not adequately addressed the challenges associated with word-based perception. + +Alternatively, other studies [87, 88] have demonstrated comparable performance on standard word-based perception benchmarks by employing the VLLM as a versatile decoder. Nonetheless, these models focus solely on pixel-level prediction and overlook box prediction. + +As illustrated in Tab. 1, although joint training has been examined for certain combinations of tasks, the collaborative effects of joint training across all four groups have not been thoroughly explored in the existing literature. + +To this end, we introduce a Multi-granular and Versatile Perception framework incorporating VLLM, termed MVPLM, which is capable of detecting and segmenting targeted objects in images according to various types of user instructions within a single model. By employing a Chain-of-Thought (CoT)-inspired data curation [72], we transform multiple box- and mask-annotated datasets into a unified supervised fine-tuning (SFT) dataset format, thereby accommodating a diverse set of tasks, including panoptic segmentation, object detection, visual grounding, and referring expression segmentation. Furthermore, to leverage the decoding and generative capabilities of LLMs, we enhance the queries for object identification and localization by incorporating features derived from LLM-generated sequences. + +In summary, our contributions are outlined as follows: + +- We introduce MVP-LM, a framework that harnesses both the decoding and generative capabilities of VLLM to perform both word-based and sentence-based perception tasks across varying granularities. +- We have developed a CoT-inspired data curation that unifies datasets from diverse tasks into a single SFT dataset, thereby encouraging the model to adopt a "thinking-then-perceiving" paradigm. +- MVP-LM demonstrates competitive performance across various benchmarks in both word-based and sentence-based perception, thereby validating the potential of our framework. Notably, our MVP-LM achieves remarkable results on both closed-set and open-set tasks compared with other specialists and VLLMs. + +# 2. Related Work + +Perception in computer vision can be categorized into four groups along two dimensions: prediction type and instruction type. Below, we review both the related work within each of these four groups and CoT, and highlight the gaps MVP-LM aims to fill. + +# 2.1. Word-based Preception + +Detection and Grounding. Word-based location tasks encompass object detection, grounding, and instance detection, focusing on identifying and localizing objects via bounding boxes with categorical labels. Conventional frameworks [13-15, 17, 18, 31, 37, 73, 83] utilize CLIP-like encoders to compute visual-linguistic similarities, enabling zero-shot capabilities. Significant advances include GLIP [38], which unifies phrase grounding with object detection, and OV-DINO [66], which implements an end-to-end training pipeline for enhanced detection performance and generalization capability. + +Segmentation. Word-based segmentation tasks, encompassing panoptic, semantic, and instance segmentation, demand precise pixel-level predictions. However, generating such masks from textual inputs remains underexplored, largely due to the scarcity of fine-grained annotated data. Existing methods [2, 22, 40, 65, 75, 82, 85, 90] primarily align word embeddings with visual features, resulting in coarse segmentation masks. Nevertheless, these approaches [85, 90] often demonstrate limited textual understanding and lack the requisite fine-grained control for high-quality mask generation, which ultimately restricts their performance in both closed-set and open-set tasks. + +# 2.2. Sentence-based Preception + +Referring Expression Comprehension. Sentence-based location tasks, which require a deeper multi-modality understanding, have become a standard benchmark for VLLMs. Specifically, Referring Expression Comprehension (REC) is often the preferred task, as box coordinates + +![](images/b4a92503a5159d1dde21a5c44042b32c28964e78512e74880dc0bcbf71808ee7.jpg) +Figure 1. Overview of MVP-LM. MVP-LM implements perception by integrating a multi-granularity decoder into the existing VLLM framework. We utilize a unified prompt template to construct the input sequence for the LLM across different tasks. The base query is derived from the summary token of the generated response. Concurrently, we extract the instruction embeddings from the input sequence (denoted by the same color) and select the corresponding residuals from the multi-scale visual features based on these embeddings. After aggregating the base query vector with the residuals, we decode the bounding box and segmentation mask by inputting the resulting queries into the multi-granularity decoder. + +can be represented as a text sequence of numbers, making them compatible with conventional VLLMs [1, 6, 41, 47, 48, 51, 60, 67, 77, 84]. However, many existing models struggle to achieve pixel-level predictions, and even more so, to facilitate joint training with sentence-based mask perception. + +Referring Expression Segmentation. Sentence-based segmentation tasks demand a fine-grained understanding of descriptive sentences to accurately generate segmentation masks for target objects. Traditional approaches [3, 9, 24-27, 56] have made strides in fusing visual and textual features to improve segmentation performance, but often fall short in fully leveraging the richness of sentence descriptions, resulting in suboptimal performance in both closed-set and open-set scenarios. Recent efforts [5, 33, 43, 57, 59] have successfully implemented Referring Expression Segmentation (RES) by incorporating an additional mask decoder. However, these approaches fall short in addressing word-based perception. + +# 2.3. Chain-of-Thought + +Chain-of-Thought prompting [72] was initially introduced during inference to enhance the reasoning capabilities of LLMs. Subsequently, CoT has evolved into a prevalent paradigm in the post-training phase, wherein models first undergo supervised fine-tuning on CoT datasets, followed by reinforcement learning [62]. A similar trend has also been observed in VLLMs, which have demonstrated promising performance in multi-modal understanding and reasoning tasks. Various meticulously designed rewards [16, 21, 50, 68, 78, 86], ranging from format to geometry, have been proposed to handle diverse scenarios. + +# 2.4. Identified Gaps and Our Contribution + +While significant progress has been made in addressing some of the combinations [6, 33, 51, 59, 88], a comprehensive solution effectively handling all four groups remains elusive. MVP-LM bridges this gap by offering a solution capable of simultaneous perception across all four combinations, marking a significant advancement in the field, addressing the current limitations, and setting a new benchmark for multi-dimensional visual perception. + +# 3. Methodology + +This section provides an in-depth description of our proposed methodology. We begin by detailing the architecture of our MVP-LM in Sec. 3.1, followed by an exposition of our strategies for dataset processing in Sec. 3.2. + +# 3.1. MVP-LM Architecture Design + +System Overview. As illustrated in Fig. 1, our architecture comprises four essential components: an image encoder (Swin-Base Transformer [49]), a connector module, an LLM (Phi-1.5 [20]), and a multi-granularity decoder. The image encoder extracts rich visual features from input images, serving as the foundation for subsequent processing. The connector module facilitates seamless integration between visual and textual modalities by aligning features from the image encoder with the token embeddings for the LLM. The multi-granularity decoder is responsible for producing both bounding boxes and segmentation masks. + +Dynamic Query Generation. The dynamic generation of visual queries is pivotal for the model's ability to handle diverse perception tasks effectively. As depicted in Fig. 2, each visual query is composed of two components: a + +![](images/dc119ff177d4b859279c9495fc6e2fc84ff92a1ea65c270c53de66f99ae09716.jpg) +Figure 2. Dynamic query selection. MVP-LM expands the embedding (hidden states) of the summary token into $N$ base query vectors. Concurrently, multi-scale visual features are flattened, and their similarity with the instruction embedding from the input sequence is computed. The top $N$ similar features are then selected and integrated with the corresponding $N$ base query vectors, resulting in $N$ final queries tailored for multi-granularity perception tasks. This dynamic selection mechanism ensures that the most relevant visual features are incorporated into each query, enhancing the precision and adaptability of the decoder. + +context-aware base query vector and language-guided residuals. + +Initially, MVP-LM leverages the VLLM to generate a specialized summary token that encapsulates the essential information from the input sequences. The hidden states corresponding to this summary token are projected into the base query vectors using an MLP, providing a robust foundation for query formulation. + +Simultaneously, we compute the similarity between multi-scale visual features and the text embedding derived from the input instructions (i.e., category names for word-based perception and referring expressions for sentence-based perception). To facilitate this comparison, an MLP aligns the dimensionality of the text embeddings with that of the multi-scale visual features. The top $N$ most similar visual features are then identified, ensuring that the queries are informed by the most relevant visual information. These selected features are integrated with the corresponding base query vectors, resulting in $N$ refined final queries that encapsulate both contextual and visual clues. This dynamic query generation process enables MVP-LM to adaptively focus on relevant regions of the image, thereby enhancing the model's ability to generate accurate bounding boxes and segmentation masks. + +Multi-granularity Decoder. The multi-granularity decoder is the cornerstone of our architecture, facilitating the simultaneous output of bounding boxes and segmentation masks. Our decoder is inspired by the design principles of OpenSeeD [85], which emphasizes the profound integration of multi-scale features and comprehensive query processing. Initially, a set of content queries and their corresponding reference points are generated through a query selection mechanism as seen in Fig. 2. These queries are then iteratively cross-attended with multi-scale visual features via deformable attention layers, enabling the decoder to capture both global contextual information and detailed + +local features. The outputs from each MSDeform layer are subsequently processed by three shared heads—a cross-modal similarity computation head, a box head, and a mask head—to produce predictions. The unified output streamlines the training process and allows the model to learn shared representations that benefit both detection and segmentation tasks. + +Training Objectives. To fully exploit the capabilities of MVP-LM, we employ a joint training approach that simultaneously optimizes multiple tasks across diverse datasets. The training objective is formulated as a composite loss function comprising several components: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {l l m} + \lambda_ {w o r d} \mathcal {L} _ {w o r d} + \lambda_ {s e n t} \mathcal {L} _ {s e n t} \\ + \mathcal {L} _ {m a s k} + \mathcal {L} _ {b o x} \\ \end{array} +$$ + +Here, $\mathcal{L}_{llm}$ represents the standard loss for next-token prediction, ensuring that the LLM accurately generates coherent and contextually appropriate summaries. The term $\mathcal{L}_{word}$ corresponds to the Cross-Entropy (CE) loss for word-based perception conditions, facilitating precise classification based on single-word instructions. Conversely, $\mathcal{L}_{sent}$ denotes the Binary Cross Entropy (BCE) loss for sentence-based perception conditions, enabling the model to handle more complex, descriptive instructions effectively. The mask loss, $\mathcal{L}_{mask}$ , is composed of a pixel-level BCE loss and a Dice loss, which together promote accurate and smooth segmentation masks by penalizing discrepancies between predicted and ground-truth masks at both the pixel and region levels. The box loss, $\mathcal{L}_{box}$ , includes the $l1$ loss to minimize the absolute differences between predicted and ground-truth bounding box coordinates, as well as the Generalized Intersection over Union (GIoU) loss to enhance the alignment and overlap between predicted and actual boxes. Formally, these loss components are defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {m a s k}} = \lambda_ {b c e} \mathcal {L} _ {b c e} + \lambda_ {d i c e} \mathcal {L} _ {d i c e}, \\ \mathcal {L} _ {b o x} = \lambda_ {l 1} \mathcal {L} _ {l 1} + \lambda_ {g i o u} \mathcal {L} _ {g i o u}, \\ \end{array} +$$ + +Following the methodology of OpenSeeD, we employ bipartite matching to match the model's predictions with their corresponding ground truth annotations during training. Predictions from each layer are assigned to their corresponding annotations to calculate losses through Hungarian matching. Moreover, a denoising strategy is incorporated, similar to that used in MaskDINO, to stabilize optimization and accelerate convergence. However, during inference, only the final layer's output is utilized. + +# 3.2. Data Curation + +CoT-inspired Dataset Unification. We unified heterogeneous datasets into an SFT dataset specially tailored for + +the training of VLLM. For each sample in the perception datasets, we construct a question-answer pair based on the original annotations. Each training pair consists of three key components: a task description, an instruction, and a response that embodies a "thinking-then-perceiving" process. + +The task description specifies the particular task associated with the data. In the case of word-based perception, the description may be articulated as follows: "Please identify all objects according to the given phrase list. This is all the candidate phrases." Conversely, for sentence-based perception, the description is framed as: "Please identify the target according to the following instruction." + +The instruction varies according to the specific tasks involved. For word-based perception, we concatenate the provided word lists into a single sentence, using a comma as a separator. For instance, given three categories: apple, banana, and orange, the resultant sentence would be "apple, banana, orange." After passing this sentence through VLLM, we select the corresponding output embeddings and apply average pooling to obtain the text embedding for each word, while for sentence-based perception, the referring expression directly serves as the final instruction, requiring no additional processing. The text embedding is derived simply by applying pooling over the entire expression. + +The response format is standardized across various tasks. We adopt a "thinking-then-perceiving" strategy by preponding the input image caption to the summary, thereby encouraging the VLLM to effectively capture image details. Specifically, the response is structured as follows: "[image caption]. The perception result is $<\text{PER}>$ ", where "<\text{PER}>" denotes a special token representing the summary. The hidden states associated with this summary token are subsequently utilized to derive the visual queries. + +Multi-caption Auto-labeling. To fully leverage the generative prowess of VLLMs, we design our system to generate comprehensive image captions in conjunction with summary tokens. This generation approach enriches the contextual information available to the model, facilitating a deeper understanding of the visual content. However, a significant challenge arises as the majority of our training datasets lack image captions. To address this, we integrate open-source VLLMs to autonomously generate descriptive captions for each image prior to the training phase. + +Specifically, we employ a diverse set of VLLMs, including VILA-3B, VILA-13B, InternVL2-8B, and InternVL2-26B, to produce multiple unique captions for the same image. The utilization of distinct VLLMs not only enhances the accuracy of the generated captions but also introduces substantial diversity, mitigating the risk of model overfitting to a single captioning style or vocabulary. By incorporating multi-caption auto-labeling, our approach ensures a more comprehensive and resilient training paradigm, effec + +Table 2. Hyperparameters for both training stages. "CP", "RC", "O", and "GG" denote COCO-Panoptic, RefCOCO, Objects365, and GoldG, respectively. + +
ParametersStage1Stage2
Training ComponentsConnectorConnector + LLM + Multi-granularity Decoder
OptimizerAdamW
Training Rate2 × 10-34 × 10-5
Batch Size12864
Number of Steps465080000
Learning Rate ScheduleCosine Decay
Weight Decay0.00.05
Warmup Ratio0.03
Training DataCC3MCP(33.3%) RC(33.3%) O(16.7%) GG(16.7%)
LossLLLMLLLM + 2 · Lword/sent + 5 · LL1 + 2 · LGIoU + 5 · LBCE + 5 · LDICE
Image Size1024 × 1024
Image ProcessingResize longer to 1024 and pad shorter to 1024
+ +tively bridging the gap between visual inputs and linguistic descriptions. + +Data Refinement. During the data generation process, the occurrence of irregular or erroneous data is inevitable. Consequently, we implement a straightforward data-cleaning procedure for the dataset. First, we utilize the length and information entropy of the generated captions to filter out meaningless results. Second, we remove sentences containing suggestive keywords such as "indicate", "might", "may", and "imply", as well as other similar terms, to mitigate the hallucinations that are inherent in large language models. These two mechanisms ensure the quality of the generated captions. + +Randomness in Training Pairs. The variability of training pairs is essential for preventing the model from learning spurious correlations and for enhancing its generalization capabilities. In our framework, such diversity is achieved by randomly selecting one of four generated captions for each training response. However, fixed category name ordering in word-based datasets can still introduce spurious causal relationships, leading to overfitting and impaired performance on unseen data. + +To counteract these issues, our mitigation strategy is twofold. Firstly, we introduce randomness by shuffling the order of the word list in each iteration. This disrupts fixed positional biases, preventing the model from relying on word order for predictions. Secondly, word lists are dynamically constructed by randomly incorporating a subset of negative category names (categories absent from the image). This selective inclusion diversifies training prompts, proving particularly beneficial for datasets with extensive vocabularies. By carefully balancing positive and negative categories within VLLM input sequence length constraints, the model's ability to discern relevant from irrelevant objects is enhanced. This approach fosters a more adaptive learning process, significantly reducing overfitting and ensuring robust performance across both closed-set and open-set detection scenarios. + +Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation and open-set segmentation tasks. Our model MVP-LM achieves remarkable performance among previous VLLM-based methods. + +
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
SpecialistsMask2former [7]Swin-B55.165.1-----
OneFormer [28]Swin-L57.967.4-----
OpenSeeD [85]Swin-L59.568.619.723.441.4--
SEEM [91]DaViT-B56.166.3-----
MaskCLIP [11]ViT-L30.947.615.123.7-45.9-
SimBaseline [76]ViT-B---20.5-47.788.4
DeOP [22]ResNet-101c---22.9-48.891.7
DaTaSeg [19]ViTDet-B52.862.712.318.328.051.1-
VLLM-based GeneralistsOMG-LLaVA [87]ConvNeXt-L53.8------
PSALM [32]Swin-B55.966.613.718.228.848.581.3
MVP-LMSwin-B56.166.819.420.535.344.185.7
+ +Table 4. Comparison with the state-of-the-art models on the closed-set Referring Segmentation benchmarks (RefCOCO series) and more challenging Generalized Referring Expression Segmentation benchmark gRefCOCO[43]. Gray numbers denote the method using gRefCOCO for training. + +
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SpecialistsVLT [10]67.570.565.256.361.050.155.057.752.562.250.5
CRIS [69]70.573.266.162.368.153.759.960.455.363.851.0
LAVT [79]72.775.868.862.168.455.161.262.157.665.355.0
PolyFormer-B [45]74.876.671.167.672.959.367.869.1---
VLLM-based GeneralistsLISA-7B [32]74.176.571.162.467.456.566.468.538.752.644.8
PixelLM-7B [58]73.076.568.266.371.758.369.370.5---
GSVA-7B [74]76.477.472.864.567.758.671.172.061.769.260.3
LaSagnA-7B [70]76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [87]78.080.374.169.173.163.072.972.9---
PSALM [88]83.684.781.672.975.570.173.874.442.052.450.6
MVP-LM83.685.182.573.976.865.475.175.637.650.449.9
+ +# 4. Experiments + +Datasets. We train MVP-LM in a multi-dataset and multitask manner like [19, 88]. For generic perception, we use COCO Panoptic Segmentation [42](with 118K samples in total) and O365 [61](with 1.7M samples in total). We use RefCOCO series [54, 80](with 126K samples in total) for referring segmentation tasks while utilizing the GoldG [30](with 770K samples in total) dataset for grounding. For evaluation, we report the main results on conventional COCO image segmentation tasks (panoptic/semantic) and RefCOCO series. Besides, we conduct evaluation on more challenging perception tasks like open-set segmentation and detection [8, 12, 53, 89]. + +Evaluation Metrics. We report results in widely used evaluation metrics: panoptic quality (PQ) and mean intersection-over-union (mIoU) for panoptic and semantic segmentation, respectively. Additionally, we use cumulative Intersection-over-Union (cIoU) for referring segmentation evaluation consistent with previous studies. + +Implementation Details. The model architecture integrates Phi1.5 [20] as the language model and Swin-B [49] as visual encoder, with our multi-granularity decoder building + +on OpenSeeD [85]. For most tasks, we initialize 100 queries for subsequent mask decoding, and for panoptic segmentation, we introduce an additional 100 learnable queries specifically for stuff categories. Our training proceeds in two phases following LLaVA [44]: 1) visual-language alignment trains exclusively the connector on CC3M [63] while freezes other components; 2) subsequent fine-tuning stage optimizes the whole model except the vision encoder. Detailed hyperparameters and training configurations are listed in Tab. 2. Training spans 80k iterations on the aggregated datasets and 9k for ablation studies, which utilize only the COCO Panoptic and RefCOCO datasets. + +# 4.1. Main results + +Word-based Perception Results. We present the performance of MVP-LM on the COCO-Panoptic [42] and open-set segmentation [8, 12, 53, 89] tasks as detailed in Tab. 3. MVP-LM demonstrates competitive performance when compared to both specialized models and VLLM-based methods in closed-set tasks. Notably, it exhibits exceptional performance in the realm of open-set segmentation within the context of VLLM. Specifically, when evaluated on the open-set ADE20K dataset [89], our + +Table 5. Results on referring expression comprehension tasks. DS-VL2 is the abbreviation for DeepSeek-VL2. + +
MethodParamsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra13B87.891.181.882.987.874.482.683.1
MiniGPTv27B88.791.785.380.085.174.584.484.7
Octopus7B89.092.683.483.589.476.084.486.2
MVP-LM1.3B93.594.591.684.988.879.386.787.4
DS-VL2200B+95.196.795.191.294.987.492.892.9
+ +MVP-LM demonstrates substantial improvements over the VLLM-based PSALM [88] and achieves gains of 5.7 and 2.3 in PQ and mIoU metrics, respectively. Additionally, it surpasses PSALM by 6.5 and 4.4 on the mIoU metric for the open-set datasets Cityscapes [8] and PascalVOC-20 [12], respectively. + +Sentence-based Perception Results. We evaluate MVP-LM against state-of-the-art methods on the referring expression segmentation (RES) benchmarks, RefCOCO+/g[55, 81] as well as the more challenging generalized referring expression segmentation benchmark, gRefCOCO[43] in Tab. 4. We also include RefCOCO series for referring expression comprehension (REC) in Tab. 5. In the domain of RES, MVP-LM achieves state-of-the-art performance on the RefCOCO and RefCOCOg datasets across all subsets. In the RefCOCO+ dataset, MVP-LM attains scores of 73.9 and 76.8 on val and testA, respectively. Moreover, our MVP-LM, only 1.3B parameters, outperforms listed 7B and 13B models, and achieves metrics that are second only to DeepSeek-VL2 with over 200B parameters in REC. These results highlight the robustness and generalization capabilities of MVP-LM when compared to previous methods. + +# 4.2. Ablation study + +This section presents an ablation study investigating critical architectural decisions, specifically: multi-dataset joint learning, performance across diverse response configurations, the impact of query quantity, and the effect of query selection. + +Joint Training. The progressive incorporation of training data sources is analyzed in Tab. 6. Initial training on COCO and RefCOCO datasets yields a MVP-LM mIoU of 77.6 on the RefCOCO validation set. Subsequent inclusion of O365 significantly increases cIoU to 81.8, with further enhancement to 83.6 cIoU upon integrating the Grounding dataset. + +For COCO evaluation, sequential integration of O365 yields a PQ of 55.8 and mIoU of 65.9. Subsequent integration of the Grounding dataset marginally improves PQ to 56.1 and mIoU to 66.8. + +Crucially, sentence-based perception is enhanced by this multi-dataset approach by 6.0 points while word-based perception remains consistent. These findings empirically validate the effectiveness of our unified multi-task training strategy in coordinating multi-modal representations and significantly enhancing cross-task generalization. + +Table 6. Ablation of the datasets involved in the joint training. We use the abbreviations C, R, O, and G to represent COCO, Ref-COCO, O365, and Grounding, respectively. Both experiments are trained for 56k iterations. + +
DatasetRefCOCO val cIoUCOCO
PQmIoU
C, R77.656.466.3
C, R, O81.855.865.9
C, R, O, G83.656.166.8
+ +Table 7. Ablation of different response settings. 'None' denotes the basic response setting containing the summary token only. 'Existing Obj' means generating existing objects' names or descriptions in the image before summary tokens. 'Caption' indicates the regression of the image caption prior to summary tokens. $^\dagger$ denotes the direct use of similar visual features as the final queries without LLM-generated base query vectors. + +
Response settingRefCOCO val cIoUCOCO
PQmIoU
None75.655.365.7
Existing Obj.75.654.965.6
Caption75.755.666.2
Caption†75.955.066.1
+ +Response Settings. We provide a comprehensive ablation study on various response settings in Tab. 7. The baseline response, "The perception result is $$ ", remains consistent across all inputs, achieving 75.6 cIoU on the RefCOCO validation set and 55.3 PQ/65.7 mIoU on COCO. When the LLM is instructed to output existing category names or object descriptions from the input image prior to the baseline response, the model maintains its performance on RefCOCO (75.6 cIoU) but exhibits a decline in COCO metrics, resulting in 54.9 PQ and 65.6 mIoU. Conversely, our final configuration, which directs the LLM to generate the image caption before delivering the baseline response, results in simultaneous improvements across all metrics (75.7 cIoU, 55.6 PQ, and 66.2 mIoU). + +Additionally, we evaluate the performance of utilizing the most similar visual features as the final queries. Compared to the approach of combining these features with the base query vectors generated by the LLM, this configuration yields a modest improvement in the RefCOCO validation set (+0.2). However, it is accompanied by a noticeable decline in COCO PQ, resulting in a decrease of 0.6. + +The incorporation of descriptive captions in the responses facilitates the mutual enhancement of both sentence-level (RefCOCO) and word-level (COCO) perception tasks. This observation verifies our hypothesis that structured semantic prompting significantly improves visual-language understanding. + +Number of Queries. As presented in Tab. 8, the response to variations in the number of queries differs between the RefCOCO and COCO datasets. Specifically, performance on the RefCOCO validation set deteriorates with an increasing + +![](images/3e9b7b06f213aabed6e6096e71a8b5fec4acfa25638b4039e03241eddd5bb914.jpg) +COCO +Figure 3. Visualization of the segmentation results on COCO, ADE20K and gRefCOCO datasets. We plot the predicted mask with a confidence larger than 0.5 for a clear visualization. The text above each image in the gRefCOCO denotes the corresponding reference text. + +Table 8. Ablation of the number of queries. 100 queries is the default setting, considering the balance between performance and computational costs. + +
# queriesRefCOCO val cIoUCOCO
PQmIoU
3076.353.165.2
10075.755.666.2
30072.454.965.5
+ +Table 9. Ablation of whether applying query selection. + +
Apply Query SelectionRefCOCO val cIoUCOCO
PQmIoU
✓75.755.666.2
78.354.465.1
+ +number of queries, decreasing from 76.3 cIoU to 72.4 cIoU. We hypothesize that, due to the presence of only one target in the RefCOCO images, an increase in the number of candidates may complicate the model's task. In contrast, for the COCO dataset, the use of 100 queries yields optimal performance among the three settings evaluated. Additionally, it is important to note that an increased number of queries incurs additional computational costs. Consequently, we adopt 100 queries as our default setting to achieve an optimal balance between performance and efficiency. + +Query Selection. As conveyed in Tab. 9, applying separate learnable query embeddings instead of query selection improves RefCOCO val results (from 75.7 to 78.3) but lowers COCO-Panoptic results (from 55.6 to 54.4). It reflects that RefCOCO's simpler targets suit fixed queries, and that COCO's complexity needs more flexible query selection. + +# 4.3. Qualitative Results + +We present visualizations for the COCO, ADE20K, and gRefCOCO datasets in Fig. 3. The results demonstrate that MVP-LM effectively generates precise masks and is proficient in identifying a comprehensive range of objects. + +# 5. Conclusion + +In this study, we categorize perception tasks into four groups based on two dimensions: prediction type and instruction type. We observe that current works typically address only a subset of the possible combinations. To address this limitation, we propose a unified Visual Large Language Model framework, MVP-LM, encompassing all these tasks. By incorporating a multi-granularity decoder and a CoT-inspired dataset unification strategy, our model enables joint training across multiple datasets and adopts a "thinking-then-perceiving" paradigm. Furthermore, we design a query enhancement strategy to exploit the decoding and generative capabilities of VLLM. Extensive experiments demonstrate the potential across various benchmarks. Considering the improvement brought by our CoT-inspired data curation, we would investigate the application of R1-like training in the realm of perception in our future work. + +# 6. Acknowledgments + +This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903), the Shenzhen Science and Technology Program (JSGG20220831093004008). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3 +[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4974-4983, 2019. 1 +[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185-24198, 2024. 1, 2, 3 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 6 +[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 7 +[9] Henghui Ding, Scott Cohen, Brian Price, and Xudong Jiang. Phraseclick: toward achieving flexible interactive segmentation by phrase and click. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 417-435. Springer, 2020. 3 +[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 6 +[11] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. 2023. 6 + +[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6, 7 +[13] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R. Scott, and Weilin Huang. TOOD: Task-aligned One-stage Object Detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490-3499, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 2 +[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In Computer Vision - ECCV 2022, pages 701-717, Cham, 2022. Springer Nature Switzerland. 1 +[15] Chengjian Feng, Yujie Zhong, Zequn Jie, Weidi Xie, and Lin Ma. Instagen: Enhancing object detection by training on synthetic dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14121-14130, 2024. 2 +[16] Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms, 2025. 3 +[17] Shenghao Fu, Junkai Yan, Qize Yang, Xihan Wei, Xiaohua Xie, and Wei-Shi Zheng. Frozen-detr: Enhancing detr with image understanding from frozen foundation models. In NeurIPS, 2024. 2 +[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2021. 2 +[19] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 6 +[20] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 3, 6 +[21] Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: An Imm perceiving any aspect ratio and high-resolution images. In Computer Vision - ECCV 2024, pages 390-406, Cham, 2025. Springer Nature Switzerland. 3 +[22] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1086-1096, 2023. 2, 6 +[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1 +[24] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3 +[25] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation, 2025. +[26] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020. +[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 3 +[28] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 6 +[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 1, 2 +[30] Gaoussou Youssouf Kebe, Padraig Higgins, Patrick Jenkins, Kasra Darvish, Rishabh Sachdeva, Ryan Barron, John Winder, Donald Engel, Edward Raff, Francis Ferraro, and Cynthia Matuszek. A spoken language dataset of descriptions for speech-based grounded language learning. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6 +[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2 +[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 1, 6 +[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3 +[34] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 1 +[35] Shiyi Lan, Xitong Yang, Zhiding Yu, Zuxuan Wu, Jose M Alvarez, and Anima Anandkumar. Vision transformers are good mask auto-labelers. arXiv preprint arXiv:2301.03992, 2023. 1 +[36] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards + +a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 1 +[37] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2 +[38] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 1, 2 +[39] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In Computer VisionECCV 2022:17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 1-18. Springer, 2022. 1 +[40] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE transactions on pattern analysis and machine intelligence, 45(4):4552-4568, 2022. 2 +[41] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3 +[42] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6 +[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 3, 6, 7 +[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6 +[45] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 1, 6 +[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2 +[47] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Li Su, and Qingming Huang. Knowledge-guided pairwise reconstruction network for weakly supervised referring expression + +grounding. In Proceedings of the 27th ACM International Conference on Multimedia, pages 539-547, 2019. 3 +[48] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 3 +[49] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6 +[50] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning, 2025. 3 +[51] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 3 +[52] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983-73007, 2023. 1 +[53] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6 +[54] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6 +[55] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7 +[56] Hai Nguyen-Truong, E-Ro Nguyen, Tuan-Anh Vu, Minh-Triet Tran, Binh-Son Hua, and Sai-Kit Yeung. Vision-aware text features in referring image segmentation: From object understanding to context understanding, 2024. 3 +[57] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 1, 2, 3 +[58] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 1, 6 +[59] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2, 3 +[60] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3 +[61] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6 +[62] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. 3 +[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6 +[64] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1 +[65] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 2 +[66] Hao Wang, Pengzhen Ren, Zequn Jie, Xiao Dong, Chengjian Feng, Yinlong Qian, Lin Ma, Dongmei Jiang, Yaowei Wang, Xiangyuan Lan, et al. Ov-dino: Unified open-vocabulary detection with language-aware selective fusion. arXiv preprint arXiv:2407.07844, 2024. 1, 2 +[67] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 3 +[68] Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, Xiangnan Fang, Zewen He, Zhenbo Luo, Wenxuan Wang, Junqi Lin, Jian Luan, and Qin Jin. Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding. arXiv e-prints, art. arXiv:2503.13377, 2025. 3 +[69] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6 +[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 6 + +[71] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Jie Hu, Dengjie Li, Zheng Zhao, and Yujiu Yang. Hyperseg: Hybrid segmentation assistant with fine-grained visual perceiver. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8931-8941, 2025. 1 +[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3 +[73] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2 +[74] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 6 +[75] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019. 2 +[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 6 +[77] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3 +[78] Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, Bo Zhang, and Wei Chen. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025. 3 +[79] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 6 +[80] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6 +[81] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7 +[82] Chang-Bin Zhang, Jinhong Ni, Yujie Zhong, and Kai Han. v-clr: View-consistent learning for open-world instance segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 20307-20317, 2025. 2 + +[83] Chang-Bin Zhang, Yujie Zhong, and Kai Han. Mr. detr: Instructive multi-route training for detection transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 9933-9943, 2025. 2 +[84] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4158-4166, 2018. 3 +[85] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1, 2, 4, 6 +[86] Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via stepwise group relative policy optimization, 2025. 3 +[87] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 1, 2, 6 +[88] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2025. 1, 2, 3, 6, 7 +[89] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6 +[90] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 1, 2 +[91] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 6 \ No newline at end of file diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/images.zip b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4ee765573da8bd2c07e0b81c314535a9040000d7 --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1dab78c8ba0b28158db2ca10cb9b32684b6fcb85e647e61d011a89721b93db9 +size 506470 diff --git a/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/layout.json b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9dc6144bc96635bd191fcd1c78676bb9045b678e --- /dev/null +++ b/2025/Advancing Visual Large Language Model for Multi-granular Versatile Perception/layout.json @@ -0,0 +1,8162 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 102, + 522, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 102, + 522, + 139 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 522, + 139 + ], + "type": "text", + "content": "Advancing Visual Large Language Model for Multi-granular Versatile Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "spans": [ + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": "Wentao Xiang" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": ", Haoxian Tan" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": ", Yujie Zhong" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": ", Cong Wei" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": ", Dengjie Li" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": ", and Yujiu Yang" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": " Tsinghua University \n" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 84, + 160, + 525, + 202 + ], + "type": "text", + "content": " Meituan Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 205, + 506, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 205, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 205, + 506, + 217 + ], + "type": "text", + "content": "thu_xiangwentao@163.com, jaszhong@hotmail.com, yang.yujiu@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 269, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 558 + ], + "type": "text", + "content": "Perception is a fundamental task in the field of computer vision, encompassing a diverse set of subtasks that can be systematically categorized into four distinct groups based on two dimensions: prediction type and instruction type. Notably, existing researches often focus solely on a limited subset of these potential combinations, which constrains their applicability and versatility across various contexts. In response to this challenge, we present MVP-LM, a Multi-granular and Versatile Perception framework incorporating Visual Large Language Model. Our framework is designed to integrate both word-based and sentence-based perception tasks alongside box and mask predictions within a single architecture. MVP-LM features an innovative multi-granularity decoder in conjunction with a CoT-inspired dataset unification strategy, enabling seamless supervised fine-tuning across a wide spectrum of tasks, including but not limited to panoptic segmentation, detection, grounding, and referring expression segmentation. Furthermore, we introduce a query enhancement strategy aimed at harnessing the decoding and generative capabilities inherent in VLLMs. Extensive experiments conducted across a range of benchmarks in both word-based and sentence-based perception tasks substantiate the efficacy of our framework. The code will be available here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 579, + 135, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 135, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 135, + 592 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 696 + ], + "type": "text", + "content": "Perception constitutes a fundamental task within the field of computer vision, necessitating that models accurately identify and locate objects within images or videos in accordance with specified instructions. These perception tasks can be categorized along two dimensions: prediction type (box vs. mask) and instruction type (word-based vs. sentence-based). This classification yields four different groups." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 277, + 552, + 399 + ], + "blocks": [ + { + "bbox": [ + 313, + 239, + 554, + 272 + ], + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 272 + ], + "type": "text", + "content": "Table 1. The comparison of capabilities of different methods. Current works can only address a subset of the combinations, while our MVP-LM can cover all tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 277, + 552, + 399 + ], + "lines": [ + { + "bbox": [ + 316, + 277, + 552, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 552, + 399 + ], + "type": "table", + "html": "
MethodOutput TypeInstruction Type
BoxMaskWordSentence
Specialists
OWL[52], PromptDet[14], OV-DINO[66]✓✓
X-Decoder[90], OpenSeeD[85], Mask DINO[36]✓✓✓
MDETR[29], GLIP[38], Grounding DINO[46]✓✓✓
VLT[10], LAVT[79], PolyFormer[45]✓✓
VLLM-based Generalists
QwenVL[1], InternVL[6], DeepseekVL[51]✓✓
LISA[32], PixelLM[58], Glamm[57]✓✓
OMG-LLaVA[87], PSALM[88], HyperSeg[71]✓✓✓
MVP-LM (Ours)✓✓✓✓
", + "image_path": "69b304799a32a07b9bc17a257e4adaed2a29d581c2eb0cb30a379bf4e6960d70.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 411, + 555, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 555, + 626 + ], + "type": "text", + "content": "The distinction between coordinate-level prediction and pixel-level prediction has been extensively explored in the literature. Given that mask annotations can be converted to box annotations without loss of information, the concurrent optimization of box predictions and mask predictions within the same dataset has emerged as a prominent training paradigm. This approach has demonstrated mutual benefits for both prediction types [23, 36]. In contrast to the costly annotations required for segmentation tasks, box annotations are significantly easier to obtain, resulting in a considerably larger dataset. Consequently, another line of research focuses on enhancing segmentation performance by leveraging additional box-annotated data. Numerous methods [4, 23, 34, 35, 39, 64, 85] have been developed to achieve this, ranging from the utilization of pre-trained models to natively annotate box data, to the design of specialized segmentation loss functions tailored for box annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 556, + 715 + ], + "type": "text", + "content": "Word-based perception employs individual words to denote the categories of targets, which often results in a degree of vagueness and ambiguity. A category name can correspond to none or multiple objects within an image. In contrast, sentence-based perception utilizes complete sentences to describe the objects of interest, resulting in greater precision. In most cases, only one object within the im" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 228, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 228, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 228, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22153" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "age satisfies the specified conditions. A common approach in the literature [29, 38, 46] involves the deployment of a two-stream framework, wherein visual queries extracted by any vision backbone are utilized to decode object locations and compute similarity with the text features generated by a language encoder to identify the objects. However, many of these methods overlook the underlying semantic concepts inherent in the descriptions, resulting in suboptimal outcomes, particularly for sentence-based perception." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 181, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 294, + 277 + ], + "type": "text", + "content": "Recently, Large Language Models (LLMs) have demonstrated a remarkable ability to comprehend user queries and generate appropriate responses, positioning them as a viable option for sentence-based perception [1, 5, 6, 33, 51, 57, 59]. Additionally, LLMs possess the capability to extract language features through forwarding, which aligns well with word-based perception. Therefore, LLMs reveal substantial potential for unifying both perceptions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 421 + ], + "type": "text", + "content": "The representative of sentence-based location tasks, Referring Expression Comprehension (REC), has emerged as the standard benchmark for Visual Large Language Models (VLLMs), as box coordinates can be expressed as a sequence of numbers in text form, rendering them compatible with traditional VLLMs. However, many existing models [1, 6, 51] encounter difficulties in achieving pixel-level predictions. Recent advances [5, 33, 57, 59] have successfully output masks by incorporating an additional mask decoder. Nevertheless, these approaches have not adequately addressed the challenges associated with word-based perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 483 + ], + "type": "text", + "content": "Alternatively, other studies [87, 88] have demonstrated comparable performance on standard word-based perception benchmarks by employing the VLLM as a versatile decoder. Nonetheless, these models focus solely on pixel-level prediction and overlook box prediction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 532 + ], + "type": "text", + "content": "As illustrated in Tab. 1, although joint training has been examined for certain combinations of tasks, the collaborative effects of joint training across all four groups have not been thoroughly explored in the existing literature." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 700 + ], + "type": "text", + "content": "To this end, we introduce a Multi-granular and Versatile Perception framework incorporating VLLM, termed MVPLM, which is capable of detecting and segmenting targeted objects in images according to various types of user instructions within a single model. By employing a Chain-of-Thought (CoT)-inspired data curation [72], we transform multiple box- and mask-annotated datasets into a unified supervised fine-tuning (SFT) dataset format, thereby accommodating a diverse set of tasks, including panoptic segmentation, object detection, visual grounding, and referring expression segmentation. Furthermore, to leverage the decoding and generative capabilities of LLMs, we enhance the queries for object identification and localization by incorporating features derived from LLM-generated sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 286, + 713 + ], + "type": "text", + "content": "In summary, our contributions are outlined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 120 + ], + "type": "text", + "content": "- We introduce MVP-LM, a framework that harnesses both the decoding and generative capabilities of VLLM to perform both word-based and sentence-based perception tasks across varying granularities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 168 + ], + "type": "text", + "content": "- We have developed a CoT-inspired data curation that unifies datasets from diverse tasks into a single SFT dataset, thereby encouraging the model to adopt a \"thinking-then-perceiving\" paradigm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 168, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 553, + 239 + ], + "type": "text", + "content": "- MVP-LM demonstrates competitive performance across various benchmarks in both word-based and sentence-based perception, thereby validating the potential of our framework. Notably, our MVP-LM achieves remarkable results on both closed-set and open-set tasks compared with other specialists and VLLMs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 400, + 264 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "type": "text", + "content": "Perception in computer vision can be categorized into four groups along two dimensions: prediction type and instruction type. Below, we review both the related work within each of these four groups and CoT, and highlight the gaps MVP-LM aims to fill." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 337, + 447, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 337, + 447, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 337, + 447, + 350 + ], + "type": "text", + "content": "2.1. Word-based Preception" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 354, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 486 + ], + "type": "text", + "content": "Detection and Grounding. Word-based location tasks encompass object detection, grounding, and instance detection, focusing on identifying and localizing objects via bounding boxes with categorical labels. Conventional frameworks [13-15, 17, 18, 31, 37, 73, 83] utilize CLIP-like encoders to compute visual-linguistic similarities, enabling zero-shot capabilities. Significant advances include GLIP [38], which unifies phrase grounding with object detection, and OV-DINO [66], which implements an end-to-end training pipeline for enhanced detection performance and generalization capability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 553, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 630 + ], + "type": "text", + "content": "Segmentation. Word-based segmentation tasks, encompassing panoptic, semantic, and instance segmentation, demand precise pixel-level predictions. However, generating such masks from textual inputs remains underexplored, largely due to the scarcity of fine-grained annotated data. Existing methods [2, 22, 40, 65, 75, 82, 85, 90] primarily align word embeddings with visual features, resulting in coarse segmentation masks. Nevertheless, these approaches [85, 90] often demonstrate limited textual understanding and lack the requisite fine-grained control for high-quality mask generation, which ultimately restricts their performance in both closed-set and open-set tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 636, + 462, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 636, + 462, + 649 + ], + "spans": [ + { + "bbox": [ + 314, + 636, + 462, + 649 + ], + "type": "text", + "content": "2.2. Sentence-based Preception" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Referring Expression Comprehension. Sentence-based location tasks, which require a deeper multi-modality understanding, have become a standard benchmark for VLLMs. Specifically, Referring Expression Comprehension (REC) is often the preferred task, as box coordinates" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22154" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 527, + 222 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 527, + 222 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 527, + 222 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 527, + 222 + ], + "type": "image", + "image_path": "b4a92503a5159d1dde21a5c44042b32c28964e78512e74880dc0bcbf71808ee7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 231, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 231, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 555, + 299 + ], + "type": "text", + "content": "Figure 1. Overview of MVP-LM. MVP-LM implements perception by integrating a multi-granularity decoder into the existing VLLM framework. We utilize a unified prompt template to construct the input sequence for the LLM across different tasks. The base query is derived from the summary token of the generated response. Concurrently, we extract the instruction embeddings from the input sequence (denoted by the same color) and select the corresponding residuals from the multi-scale visual features based on these embeddings. After aggregating the base query vector with the residuals, we decode the bounding box and segmentation mask by inputting the resulting queries into the multi-granularity decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "type": "text", + "content": "can be represented as a text sequence of numbers, making them compatible with conventional VLLMs [1, 6, 41, 47, 48, 51, 60, 67, 77, 84]. However, many existing models struggle to achieve pixel-level predictions, and even more so, to facilitate joint training with sentence-based mask perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 296, + 538 + ], + "type": "text", + "content": "Referring Expression Segmentation. Sentence-based segmentation tasks demand a fine-grained understanding of descriptive sentences to accurately generate segmentation masks for target objects. Traditional approaches [3, 9, 24-27, 56] have made strides in fusing visual and textual features to improve segmentation performance, but often fall short in fully leveraging the richness of sentence descriptions, resulting in suboptimal performance in both closed-set and open-set scenarios. Recent efforts [5, 33, 43, 57, 59] have successfully implemented Referring Expression Segmentation (RES) by incorporating an additional mask decoder. However, these approaches fall short in addressing word-based perception." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 163, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 163, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 163, + 572 + ], + "type": "text", + "content": "2.3. Chain-of-Thought" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Chain-of-Thought prompting [72] was initially introduced during inference to enhance the reasoning capabilities of LLMs. Subsequently, CoT has evolved into a prevalent paradigm in the post-training phase, wherein models first undergo supervised fine-tuning on CoT datasets, followed by reinforcement learning [62]. A similar trend has also been observed in VLLMs, which have demonstrated promising performance in multi-modal understanding and reasoning tasks. Various meticulously designed rewards [16, 21, 50, 68, 78, 86], ranging from format to geometry, have been proposed to handle diverse scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 304, + 515, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 515, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 515, + 317 + ], + "type": "text", + "content": "2.4. Identified Gaps and Our Contribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 323, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 419 + ], + "type": "text", + "content": "While significant progress has been made in addressing some of the combinations [6, 33, 51, 59, 88], a comprehensive solution effectively handling all four groups remains elusive. MVP-LM bridges this gap by offering a solution capable of simultaneous perception across all four combinations, marking a significant advancement in the field, addressing the current limitations, and setting a new benchmark for multi-dimensional visual perception." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 434, + 397, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 397, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 397, + 448 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 503 + ], + "type": "text", + "content": "This section provides an in-depth description of our proposed methodology. We begin by detailing the architecture of our MVP-LM in Sec. 3.1, followed by an exposition of our strategies for dataset processing in Sec. 3.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 514, + 479, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 479, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 479, + 527 + ], + "type": "text", + "content": "3.1. MVP-LM Architecture Design" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": "System Overview. As illustrated in Fig. 1, our architecture comprises four essential components: an image encoder (Swin-Base Transformer [49]), a connector module, an LLM (Phi-1.5 [20]), and a multi-granularity decoder. The image encoder extracts rich visual features from input images, serving as the foundation for subsequent processing. The connector module facilitates seamless integration between visual and textual modalities by aligning features from the image encoder with the token embeddings for the LLM. The multi-granularity decoder is responsible for producing both bounding boxes and segmentation masks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Dynamic Query Generation. The dynamic generation of visual queries is pivotal for the model's ability to handle diverse perception tasks effectively. As depicted in Fig. 2, each visual query is composed of two components: a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22155" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 281, + 155 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 281, + 155 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 281, + 155 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 281, + 155 + ], + "type": "image", + "image_path": "dc119ff177d4b859279c9495fc6e2fc84ff92a1ea65c270c53de66f99ae09716.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "text", + "content": "Figure 2. Dynamic query selection. MVP-LM expands the embedding (hidden states) of the summary token into " + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "text", + "content": " base query vectors. Concurrently, multi-scale visual features are flattened, and their similarity with the instruction embedding from the input sequence is computed. The top " + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "text", + "content": " similar features are then selected and integrated with the corresponding " + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "text", + "content": " base query vectors, resulting in " + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 164, + 296, + 275 + ], + "type": "text", + "content": " final queries tailored for multi-granularity perception tasks. This dynamic selection mechanism ensures that the most relevant visual features are incorporated into each query, enhancing the precision and adaptability of the decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 295, + 306 + ], + "type": "text", + "content": "context-aware base query vector and language-guided residuals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 378 + ], + "type": "text", + "content": "Initially, MVP-LM leverages the VLLM to generate a specialized summary token that encapsulates the essential information from the input sequences. The hidden states corresponding to this summary token are projected into the base query vectors using an MLP, providing a robust foundation for query formulation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "text", + "content": "Simultaneously, we compute the similarity between multi-scale visual features and the text embedding derived from the input instructions (i.e., category names for word-based perception and referring expressions for sentence-based perception). To facilitate this comparison, an MLP aligns the dimensionality of the text embeddings with that of the multi-scale visual features. The top " + }, + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "text", + "content": " most similar visual features are then identified, ensuring that the queries are informed by the most relevant visual information. These selected features are integrated with the corresponding base query vectors, resulting in " + }, + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 379, + 295, + 569 + ], + "type": "text", + "content": " refined final queries that encapsulate both contextual and visual clues. This dynamic query generation process enables MVP-LM to adaptively focus on relevant regions of the image, thereby enhancing the model's ability to generate accurate bounding boxes and segmentation masks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Multi-granularity Decoder. The multi-granularity decoder is the cornerstone of our architecture, facilitating the simultaneous output of bounding boxes and segmentation masks. Our decoder is inspired by the design principles of OpenSeeD [85], which emphasizes the profound integration of multi-scale features and comprehensive query processing. Initially, a set of content queries and their corresponding reference points are generated through a query selection mechanism as seen in Fig. 2. These queries are then iteratively cross-attended with multi-scale visual features via deformable attention layers, enabling the decoder to capture both global contextual information and detailed" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "local features. The outputs from each MSDeform layer are subsequently processed by three shared heads—a cross-modal similarity computation head, a box head, and a mask head—to produce predictions. The unified output streamlines the training process and allows the model to learn shared representations that benefit both detection and segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 167, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 228 + ], + "type": "text", + "content": "Training Objectives. To fully exploit the capabilities of MVP-LM, we employ a joint training approach that simultaneously optimizes multiple tasks across diverse datasets. The training objective is formulated as a composite loss function comprising several components:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 353, + 236, + 515, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 236, + 515, + 264 + ], + "spans": [ + { + "bbox": [ + 353, + 236, + 515, + 264 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {l l m} + \\lambda_ {w o r d} \\mathcal {L} _ {w o r d} + \\lambda_ {s e n t} \\mathcal {L} _ {s e n t} \\\\ + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {b o x} \\\\ \\end{array}", + "image_path": "44a3d0449d0c3f906ddd77a1cef5cbef6991fdebeb4ef18298d5e8cd0e7fa543.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{llm}" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": " represents the standard loss for next-token prediction, ensuring that the LLM accurately generates coherent and contextually appropriate summaries. The term " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{word}" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": " corresponds to the Cross-Entropy (CE) loss for word-based perception conditions, facilitating precise classification based on single-word instructions. Conversely, " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sent}" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": " denotes the Binary Cross Entropy (BCE) loss for sentence-based perception conditions, enabling the model to handle more complex, descriptive instructions effectively. The mask loss, " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": ", is composed of a pixel-level BCE loss and a Dice loss, which together promote accurate and smooth segmentation masks by penalizing discrepancies between predicted and ground-truth masks at both the pixel and region levels. The box loss, " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": ", includes the " + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "inline_equation", + "content": "l1" + }, + { + "bbox": [ + 313, + 273, + 555, + 511 + ], + "type": "text", + "content": " loss to minimize the absolute differences between predicted and ground-truth bounding box coordinates, as well as the Generalized Intersection over Union (GIoU) loss to enhance the alignment and overlap between predicted and actual boxes. Formally, these loss components are defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 521, + 502, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 521, + 502, + 550 + ], + "spans": [ + { + "bbox": [ + 365, + 521, + 502, + 550 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {m a s k}} = \\lambda_ {b c e} \\mathcal {L} _ {b c e} + \\lambda_ {d i c e} \\mathcal {L} _ {d i c e}, \\\\ \\mathcal {L} _ {b o x} = \\lambda_ {l 1} \\mathcal {L} _ {l 1} + \\lambda_ {g i o u} \\mathcal {L} _ {g i o u}, \\\\ \\end{array}", + "image_path": "dcb6f65e4f6906dda4265f0d5d2ca0dc2e6d5fdfd2f89c2734af7cd2f054480c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 665 + ], + "type": "text", + "content": "Following the methodology of OpenSeeD, we employ bipartite matching to match the model's predictions with their corresponding ground truth annotations during training. Predictions from each layer are assigned to their corresponding annotations to calculate losses through Hungarian matching. Moreover, a denoising strategy is incorporated, similar to that used in MaskDINO, to stabilize optimization and accelerate convergence. However, during inference, only the final layer's output is utilized." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 405, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 405, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 405, + 684 + ], + "type": "text", + "content": "3.2. Data Curation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "CoT-inspired Dataset Unification. We unified heterogeneous datasets into an SFT dataset specially tailored for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22156" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "the training of VLLM. For each sample in the perception datasets, we construct a question-answer pair based on the original annotations. Each training pair consists of three key components: a task description, an instruction, and a response that embodies a \"thinking-then-perceiving\" process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 294, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 294, + 229 + ], + "type": "text", + "content": "The task description specifies the particular task associated with the data. In the case of word-based perception, the description may be articulated as follows: \"Please identify all objects according to the given phrase list. This is all the candidate phrases.\" Conversely, for sentence-based perception, the description is framed as: \"Please identify the target according to the following instruction.\"" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 231, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 294, + 374 + ], + "type": "text", + "content": "The instruction varies according to the specific tasks involved. For word-based perception, we concatenate the provided word lists into a single sentence, using a comma as a separator. For instance, given three categories: apple, banana, and orange, the resultant sentence would be \"apple, banana, orange.\" After passing this sentence through VLLM, we select the corresponding output embeddings and apply average pooling to obtain the text embedding for each word, while for sentence-based perception, the referring expression directly serves as the final instruction, requiring no additional processing. The text embedding is derived simply by applying pooling over the entire expression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 294, + 483 + ], + "type": "text", + "content": "The response format is standardized across various tasks. We adopt a \"thinking-then-perceiving\" strategy by preponding the input image caption to the summary, thereby encouraging the VLLM to effectively capture image details. Specifically, the response is structured as follows: \"[image caption]. The perception result is " + }, + { + "bbox": [ + 55, + 376, + 294, + 483 + ], + "type": "inline_equation", + "content": "<\\text{PER}>" + }, + { + "bbox": [ + 55, + 376, + 294, + 483 + ], + "type": "text", + "content": "\", where \"<\\text{PER}>\" denotes a special token representing the summary. The hidden states associated with this summary token are subsequently utilized to derive the visual queries." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 294, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 294, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 294, + 605 + ], + "type": "text", + "content": "Multi-caption Auto-labeling. To fully leverage the generative prowess of VLLMs, we design our system to generate comprehensive image captions in conjunction with summary tokens. This generation approach enriches the contextual information available to the model, facilitating a deeper understanding of the visual content. However, a significant challenge arises as the majority of our training datasets lack image captions. To address this, we integrate open-source VLLMs to autonomously generate descriptive captions for each image prior to the training phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 714 + ], + "type": "text", + "content": "Specifically, we employ a diverse set of VLLMs, including VILA-3B, VILA-13B, InternVL2-8B, and InternVL2-26B, to produce multiple unique captions for the same image. The utilization of distinct VLLMs not only enhances the accuracy of the generated captions but also introduces substantial diversity, mitigating the risk of model overfitting to a single captioning style or vocabulary. By incorporating multi-caption auto-labeling, our approach ensures a more comprehensive and resilient training paradigm, effec" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 102, + 553, + 231 + ], + "blocks": [ + { + "bbox": [ + 313, + 65, + 553, + 99 + ], + "lines": [ + { + "bbox": [ + 313, + 65, + 553, + 99 + ], + "spans": [ + { + "bbox": [ + 313, + 65, + 553, + 99 + ], + "type": "text", + "content": "Table 2. Hyperparameters for both training stages. \"CP\", \"RC\", \"O\", and \"GG\" denote COCO-Panoptic, RefCOCO, Objects365, and GoldG, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 102, + 553, + 231 + ], + "lines": [ + { + "bbox": [ + 316, + 102, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 102, + 553, + 231 + ], + "type": "table", + "html": "
ParametersStage1Stage2
Training ComponentsConnectorConnector + LLM + Multi-granularity Decoder
OptimizerAdamW
Training Rate2 × 10-34 × 10-5
Batch Size12864
Number of Steps465080000
Learning Rate ScheduleCosine Decay
Weight Decay0.00.05
Warmup Ratio0.03
Training DataCC3MCP(33.3%) RC(33.3%) O(16.7%) GG(16.7%)
LossLLLMLLLM + 2 · Lword/sent + 5 · LL1 + 2 · LGIoU + 5 · LBCE + 5 · LDICE
Image Size1024 × 1024
Image ProcessingResize longer to 1024 and pad shorter to 1024
", + "image_path": "e8256d8e90baff093c2564635e1f696d2d470b827208a99bfbbe44a3ca38dcb0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 244, + 553, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 268 + ], + "type": "text", + "content": "tively bridging the gap between visual inputs and linguistic descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 273, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 554, + 405 + ], + "type": "text", + "content": "Data Refinement. During the data generation process, the occurrence of irregular or erroneous data is inevitable. Consequently, we implement a straightforward data-cleaning procedure for the dataset. First, we utilize the length and information entropy of the generated captions to filter out meaningless results. Second, we remove sentences containing suggestive keywords such as \"indicate\", \"might\", \"may\", and \"imply\", as well as other similar terms, to mitigate the hallucinations that are inherent in large language models. These two mechanisms ensure the quality of the generated captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 517 + ], + "type": "text", + "content": "Randomness in Training Pairs. The variability of training pairs is essential for preventing the model from learning spurious correlations and for enhancing its generalization capabilities. In our framework, such diversity is achieved by randomly selecting one of four generated captions for each training response. However, fixed category name ordering in word-based datasets can still introduce spurious causal relationships, leading to overfitting and impaired performance on unseen data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 713 + ], + "type": "text", + "content": "To counteract these issues, our mitigation strategy is twofold. Firstly, we introduce randomness by shuffling the order of the word list in each iteration. This disrupts fixed positional biases, preventing the model from relying on word order for predictions. Secondly, word lists are dynamically constructed by randomly incorporating a subset of negative category names (categories absent from the image). This selective inclusion diversifies training prompts, proving particularly beneficial for datasets with extensive vocabularies. By carefully balancing positive and negative categories within VLLM input sequence length constraints, the model's ability to discern relevant from irrelevant objects is enhanced. This approach fosters a more adaptive learning process, significantly reducing overfitting and ensuring robust performance across both closed-set and open-set detection scenarios." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22157" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 92, + 518, + 233 + ], + "blocks": [ + { + "bbox": [ + 55, + 65, + 555, + 87 + ], + "lines": [ + { + "bbox": [ + 55, + 65, + 555, + 87 + ], + "spans": [ + { + "bbox": [ + 55, + 65, + 555, + 87 + ], + "type": "text", + "content": "Table 3. Quantitative results on the closed-set COCO-Panoptic segmentation and open-set segmentation tasks. Our model MVP-LM achieves remarkable performance among previous VLLM-based methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 92, + 518, + 233 + ], + "lines": [ + { + "bbox": [ + 90, + 92, + 518, + 233 + ], + "spans": [ + { + "bbox": [ + 90, + 92, + 518, + 233 + ], + "type": "table", + "html": "
TypeMethodBackboneCOCO-PanopticADE-OVCitys-OVPC59-OVPAS20-OV
PQmIoUPQmIoUPQmIoUmIoU
SpecialistsMask2former [7]Swin-B55.165.1-----
OneFormer [28]Swin-L57.967.4-----
OpenSeeD [85]Swin-L59.568.619.723.441.4--
SEEM [91]DaViT-B56.166.3-----
MaskCLIP [11]ViT-L30.947.615.123.7-45.9-
SimBaseline [76]ViT-B---20.5-47.788.4
DeOP [22]ResNet-101c---22.9-48.891.7
DaTaSeg [19]ViTDet-B52.862.712.318.328.051.1-
VLLM-based GeneralistsOMG-LLaVA [87]ConvNeXt-L53.8------
PSALM [32]Swin-B55.966.613.718.228.848.581.3
MVP-LMSwin-B56.166.819.420.535.344.185.7
", + "image_path": "d49f4138d048c396083526bdde60e8869859d5ad4bcd964fdcf080907a5c0d8c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 85, + 277, + 525, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 273 + ], + "type": "text", + "content": "Table 4. Comparison with the state-of-the-art models on the closed-set Referring Segmentation benchmarks (RefCOCO series) and more challenging Generalized Referring Expression Segmentation benchmark gRefCOCO[43]. Gray numbers denote the method using gRefCOCO for training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 277, + 525, + 423 + ], + "lines": [ + { + "bbox": [ + 85, + 277, + 525, + 423 + ], + "spans": [ + { + "bbox": [ + 85, + 277, + 525, + 423 + ], + "type": "table", + "html": "
TypeMethodRefCOCORefCOCO+RefCOCOggRefCOCO
valtestAtestBvaltestAtestBval(U)test(U)valtestAtestB
SpecialistsVLT [10]67.570.565.256.361.050.155.057.752.562.250.5
CRIS [69]70.573.266.162.368.153.759.960.455.363.851.0
LAVT [79]72.775.868.862.168.455.161.262.157.665.355.0
PolyFormer-B [45]74.876.671.167.672.959.367.869.1---
VLLM-based GeneralistsLISA-7B [32]74.176.571.162.467.456.566.468.538.752.644.8
PixelLM-7B [58]73.076.568.266.371.758.369.370.5---
GSVA-7B [74]76.477.472.864.567.758.671.172.061.769.260.3
LaSagnA-7B [70]76.878.773.866.470.660.170.671.938.150.442.1
OMG-LLaVA [87]78.080.374.169.173.163.072.972.9---
PSALM [88]83.684.781.672.975.570.173.874.442.052.450.6
MVP-LM83.685.182.573.976.865.475.175.637.650.449.9
", + "image_path": "5b78b6cf1cc6c9010bc3b6f28b56a7a77c67f380cd7303426ae5ae4c6abfac1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 137, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 137, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 137, + 450 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 458, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 602 + ], + "type": "text", + "content": "Datasets. We train MVP-LM in a multi-dataset and multitask manner like [19, 88]. For generic perception, we use COCO Panoptic Segmentation [42](with 118K samples in total) and O365 [61](with 1.7M samples in total). We use RefCOCO series [54, 80](with 126K samples in total) for referring segmentation tasks while utilizing the GoldG [30](with 770K samples in total) dataset for grounding. For evaluation, we report the main results on conventional COCO image segmentation tasks (panoptic/semantic) and RefCOCO series. Besides, we conduct evaluation on more challenging perception tasks like open-set segmentation and detection [8, 12, 53, 89]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 676 + ], + "type": "text", + "content": "Evaluation Metrics. We report results in widely used evaluation metrics: panoptic quality (PQ) and mean intersection-over-union (mIoU) for panoptic and semantic segmentation, respectively. Additionally, we use cumulative Intersection-over-Union (cIoU) for referring segmentation evaluation consistent with previous studies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. The model architecture integrates Phi1.5 [20] as the language model and Swin-B [49] as visual encoder, with our multi-granularity decoder building" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": "on OpenSeeD [85]. For most tasks, we initialize 100 queries for subsequent mask decoding, and for panoptic segmentation, we introduce an additional 100 learnable queries specifically for stuff categories. Our training proceeds in two phases following LLaVA [44]: 1) visual-language alignment trains exclusively the connector on CC3M [63] while freezes other components; 2) subsequent fine-tuning stage optimizes the whole model except the vision encoder. Detailed hyperparameters and training configurations are listed in Tab. 2. Training spans 80k iterations on the aggregated datasets and 9k for ablation studies, which utilize only the COCO Panoptic and RefCOCO datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 397, + 600 + ], + "type": "text", + "content": "4.1. Main results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Word-based Perception Results. We present the performance of MVP-LM on the COCO-Panoptic [42] and open-set segmentation [8, 12, 53, 89] tasks as detailed in Tab. 3. MVP-LM demonstrates competitive performance when compared to both specialized models and VLLM-based methods in closed-set tasks. Notably, it exhibits exceptional performance in the realm of open-set segmentation within the context of VLLM. Specifically, when evaluated on the open-set ADE20K dataset [89], our" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22158" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 92, + 294, + 160 + ], + "blocks": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "lines": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "spans": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "type": "text", + "content": "Table 5. Results on referring expression comprehension tasks. DS-VL2 is the abbreviation for DeepSeek-VL2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 92, + 294, + 160 + ], + "lines": [ + { + "bbox": [ + 57, + 92, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 57, + 92, + 294, + 160 + ], + "type": "table", + "html": "
MethodParamsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra13B87.891.181.882.987.874.482.683.1
MiniGPTv27B88.791.785.380.085.174.584.484.7
Octopus7B89.092.683.483.589.476.084.486.2
MVP-LM1.3B93.594.591.684.988.879.386.787.4
DS-VL2200B+95.196.795.191.294.987.492.892.9
", + "image_path": "af10024bab862a3a9c57f07ac4ffc081cb4860a2e438146429d38c6b519b8b84.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 295, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 246 + ], + "type": "text", + "content": "MVP-LM demonstrates substantial improvements over the VLLM-based PSALM [88] and achieves gains of 5.7 and 2.3 in PQ and mIoU metrics, respectively. Additionally, it surpasses PSALM by 6.5 and 4.4 on the mIoU metric for the open-set datasets Cityscapes [8] and PascalVOC-20 [12], respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 438 + ], + "type": "text", + "content": "Sentence-based Perception Results. We evaluate MVP-LM against state-of-the-art methods on the referring expression segmentation (RES) benchmarks, RefCOCO+/g[55, 81] as well as the more challenging generalized referring expression segmentation benchmark, gRefCOCO[43] in Tab. 4. We also include RefCOCO series for referring expression comprehension (REC) in Tab. 5. In the domain of RES, MVP-LM achieves state-of-the-art performance on the RefCOCO and RefCOCOg datasets across all subsets. In the RefCOCO+ dataset, MVP-LM attains scores of 73.9 and 76.8 on val and testA, respectively. Moreover, our MVP-LM, only 1.3B parameters, outperforms listed 7B and 13B models, and achieves metrics that are second only to DeepSeek-VL2 with over 200B parameters in REC. These results highlight the robustness and generalization capabilities of MVP-LM when compared to previous methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 148, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 148, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 148, + 458 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 521 + ], + "type": "text", + "content": "This section presents an ablation study investigating critical architectural decisions, specifically: multi-dataset joint learning, performance across diverse response configurations, the impact of query quantity, and the effect of query selection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 594 + ], + "type": "text", + "content": "Joint Training. The progressive incorporation of training data sources is analyzed in Tab. 6. Initial training on COCO and RefCOCO datasets yields a MVP-LM mIoU of 77.6 on the RefCOCO validation set. Subsequent inclusion of O365 significantly increases cIoU to 81.8, with further enhancement to 83.6 cIoU upon integrating the Grounding dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "For COCO evaluation, sequential integration of O365 yields a PQ of 55.8 and mIoU of 65.9. Subsequent integration of the Grounding dataset marginally improves PQ to 56.1 and mIoU to 66.8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Crucially, sentence-based perception is enhanced by this multi-dataset approach by 6.0 points while word-based perception remains consistent. These findings empirically validate the effectiveness of our unified multi-task training strategy in coordinating multi-modal representations and significantly enhancing cross-task generalization." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 363, + 114, + 504, + 168 + ], + "blocks": [ + { + "bbox": [ + 313, + 65, + 553, + 109 + ], + "lines": [ + { + "bbox": [ + 313, + 65, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 65, + 553, + 109 + ], + "type": "text", + "content": "Table 6. Ablation of the datasets involved in the joint training. We use the abbreviations C, R, O, and G to represent COCO, Ref-COCO, O365, and Grounding, respectively. Both experiments are trained for 56k iterations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 363, + 114, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 363, + 114, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 363, + 114, + 504, + 168 + ], + "type": "table", + "html": "
DatasetRefCOCO val cIoUCOCO
PQmIoU
C, R77.656.466.3
C, R, O81.855.865.9
C, R, O, G83.656.166.8
", + "image_path": "63a29edc1634511024a4bf67593c7bb9cc08f220a65f29286174723766eca910.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 353, + 260, + 515, + 326 + ], + "blocks": [ + { + "bbox": [ + 313, + 175, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 253 + ], + "type": "text", + "content": "Table 7. Ablation of different response settings. 'None' denotes the basic response setting containing the summary token only. 'Existing Obj' means generating existing objects' names or descriptions in the image before summary tokens. 'Caption' indicates the regression of the image caption prior to summary tokens. " + }, + { + "bbox": [ + 313, + 175, + 555, + 253 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 313, + 175, + 555, + 253 + ], + "type": "text", + "content": " denotes the direct use of similar visual features as the final queries without LLM-generated base query vectors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 260, + 515, + 326 + ], + "lines": [ + { + "bbox": [ + 353, + 260, + 515, + 326 + ], + "spans": [ + { + "bbox": [ + 353, + 260, + 515, + 326 + ], + "type": "table", + "html": "
Response settingRefCOCO val cIoUCOCO
PQmIoU
None75.655.365.7
Existing Obj.75.654.965.6
Caption75.755.666.2
Caption†75.955.066.1
", + "image_path": "05bc8ed1c55b0ae871b3c2169719067fa1b5ada91523c0b2133179eb0c210af9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 508 + ], + "type": "text", + "content": "Response Settings. We provide a comprehensive ablation study on various response settings in Tab. 7. The baseline response, \"The perception result is " + }, + { + "bbox": [ + 313, + 340, + 555, + 508 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 340, + 555, + 508 + ], + "type": "text", + "content": "\", remains consistent across all inputs, achieving 75.6 cIoU on the RefCOCO validation set and 55.3 PQ/65.7 mIoU on COCO. When the LLM is instructed to output existing category names or object descriptions from the input image prior to the baseline response, the model maintains its performance on RefCOCO (75.6 cIoU) but exhibits a decline in COCO metrics, resulting in 54.9 PQ and 65.6 mIoU. Conversely, our final configuration, which directs the LLM to generate the image caption before delivering the baseline response, results in simultaneous improvements across all metrics (75.7 cIoU, 55.6 PQ, and 66.2 mIoU)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 592 + ], + "type": "text", + "content": "Additionally, we evaluate the performance of utilizing the most similar visual features as the final queries. Compared to the approach of combining these features with the base query vectors generated by the LLM, this configuration yields a modest improvement in the RefCOCO validation set (+0.2). However, it is accompanied by a noticeable decline in COCO PQ, resulting in a decrease of 0.6." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 665 + ], + "type": "text", + "content": "The incorporation of descriptive captions in the responses facilitates the mutual enhancement of both sentence-level (RefCOCO) and word-level (COCO) perception tasks. This observation verifies our hypothesis that structured semantic prompting significantly improves visual-language understanding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Number of Queries. As presented in Tab. 8, the response to variations in the number of queries differs between the RefCOCO and COCO datasets. Specifically, performance on the RefCOCO validation set deteriorates with an increasing" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22159" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 72, + 519, + 305 + ], + "blocks": [ + { + "bbox": [ + 290, + 61, + 314, + 70 + ], + "lines": [ + { + "bbox": [ + 290, + 61, + 314, + 70 + ], + "spans": [ + { + "bbox": [ + 290, + 61, + 314, + 70 + ], + "type": "text", + "content": "COCO" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 89, + 72, + 519, + 305 + ], + "lines": [ + { + "bbox": [ + 89, + 72, + 519, + 305 + ], + "spans": [ + { + "bbox": [ + 89, + 72, + 519, + 305 + ], + "type": "image", + "image_path": "3e9b7b06f213aabed6e6096e71a8b5fec4acfa25638b4039e03241eddd5bb914.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 312, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 555, + 336 + ], + "type": "text", + "content": "Figure 3. Visualization of the segmentation results on COCO, ADE20K and gRefCOCO datasets. We plot the predicted mask with a confidence larger than 0.5 for a clear visualization. The text above each image in the gRefCOCO denotes the corresponding reference text." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 106, + 388, + 244, + 443 + ], + "blocks": [ + { + "bbox": [ + 55, + 346, + 296, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 379 + ], + "type": "text", + "content": "Table 8. Ablation of the number of queries. 100 queries is the default setting, considering the balance between performance and computational costs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 388, + 244, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 388, + 244, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 244, + 443 + ], + "type": "table", + "html": "
# queriesRefCOCO val cIoUCOCO
PQmIoU
3076.353.165.2
10075.755.666.2
30072.454.965.5
", + "image_path": "7cf3d7fc19d49319883a821070af5fea32ae12cff39c09df76b7689598cf5594.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 85, + 470, + 264, + 516 + ], + "blocks": [ + { + "bbox": [ + 75, + 450, + 274, + 462 + ], + "lines": [ + { + "bbox": [ + 75, + 450, + 274, + 462 + ], + "spans": [ + { + "bbox": [ + 75, + 450, + 274, + 462 + ], + "type": "text", + "content": "Table 9. Ablation of whether applying query selection." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 470, + 264, + 516 + ], + "lines": [ + { + "bbox": [ + 85, + 470, + 264, + 516 + ], + "spans": [ + { + "bbox": [ + 85, + 470, + 264, + 516 + ], + "type": "table", + "html": "
Apply Query SelectionRefCOCO val cIoUCOCO
PQmIoU
✓75.755.666.2
78.354.465.1
", + "image_path": "600b64eef6d7ee204b3a8394fab4d6dc2cc4f6614d0545b08ce8219ac8831e02.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "number of queries, decreasing from 76.3 cIoU to 72.4 cIoU. We hypothesize that, due to the presence of only one target in the RefCOCO images, an increase in the number of candidates may complicate the model's task. In contrast, for the COCO dataset, the use of 100 queries yields optimal performance among the three settings evaluated. Additionally, it is important to note that an increased number of queries incurs additional computational costs. Consequently, we adopt 100 queries as our default setting to achieve an optimal balance between performance and efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Query Selection. As conveyed in Tab. 9, applying separate learnable query embeddings instead of query selection improves RefCOCO val results (from 75.7 to 78.3) but lowers COCO-Panoptic results (from 55.6 to 54.4). It reflects that RefCOCO's simpler targets suit fixed queries, and that COCO's complexity needs more flexible query selection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 425, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 425, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 425, + 360 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 413 + ], + "type": "text", + "content": "We present visualizations for the COCO, ADE20K, and gRefCOCO datasets in Fig. 3. The results demonstrate that MVP-LM effectively generates precise masks and is proficient in identifying a comprehensive range of objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 423, + 388, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 388, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 388, + 435 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 555, + 635 + ], + "type": "text", + "content": "In this study, we categorize perception tasks into four groups based on two dimensions: prediction type and instruction type. We observe that current works typically address only a subset of the possible combinations. To address this limitation, we propose a unified Visual Large Language Model framework, MVP-LM, encompassing all these tasks. By incorporating a multi-granularity decoder and a CoT-inspired dataset unification strategy, our model enables joint training across multiple datasets and adopts a \"thinking-then-perceiving\" paradigm. Furthermore, we design a query enhancement strategy to exploit the decoding and generative capabilities of VLLM. Extensive experiments demonstrate the potential across various benchmarks. Considering the improvement brought by our CoT-inspired data curation, we would investigate the application of R1-like training in the realm of perception in our future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 424, + 659 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "This work was partly supported by the National Key Research and Development Program of China (No. 2024YFB2808903), the Shenzhen Science and Technology Program (JSGG20220831093004008)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22160" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 192 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "text", + "content": "[3] Ding-Jie Chen, Songhao Jia, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. See-through-text grouping for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7454-7463, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 315 + ], + "type": "text", + "content": "[4] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4974-4983, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 361 + ], + "type": "text", + "content": "[5] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 295, + 430 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185-24198, 2024. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "type": "text", + "content": "[8] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 295, + 621 + ], + "type": "text", + "content": "[9] Henghui Ding, Scott Cohen, Brian Price, and Xudong Jiang. Phraseclick: toward achieving flexible interactive segmentation by phrase and click. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 417-435. Springer, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 624, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 295, + 677 + ], + "type": "text", + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16321-16330, 2021. 1, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 712 + ], + "type": "text", + "content": "[11] Zheng Ding, Jieke Wang, and Zhuowen Tu. Open-vocabulary universal image segmentation with maskclip. 2023. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 119, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 173 + ], + "type": "text", + "content": "[13] Chengjian Feng, Yujie Zhong, Yu Gao, Matthew R. Scott, and Weilin Huang. TOOD: Task-aligned One-stage Object Detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3490-3499, Los Alamitos, CA, USA, 2021. IEEE Computer Society. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 175, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 229 + ], + "type": "text", + "content": "[14] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In Computer Vision - ECCV 2022, pages 701-717, Cham, 2022. Springer Nature Switzerland. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 231, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 284 + ], + "type": "text", + "content": "[15] Chengjian Feng, Yujie Zhong, Zequn Jie, Weidi Xie, and Lin Ma. Instagen: Enhancing object detection by training on synthetic dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14121-14130, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[16] Kaituo Feng, Kaixiong Gong, Bohao Li, Zonghao Guo, Yibing Wang, Tianshuo Peng, Junfei Wu, Xiaoying Zhang, Benyou Wang, and Xiangyu Yue. Video-r1: Reinforcing video reasoning in mllms, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[17] Shenghao Fu, Junkai Yan, Qize Yang, Xihan Wei, Xiaohua Xie, and Wei-Shi Zheng. Frozen-detr: Enhancing detr with image understanding from frozen foundation models. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 409 + ], + "type": "text", + "content": "[18] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In ICLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "type": "text", + "content": "[19] Xiuye Gu, Yin Cui, Jonathan Huang, Abdullah Rashwan, Xuan Yang, Xingyi Zhou, Golnaz Ghiasi, Weicheng Kuo, Huizhong Chen, Liang-Chieh Chen, et al. Dataseg: Taming a universal multi-dataset multi-task segmentation model. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "text", + "content": "[20] Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Cao César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need. arXiv preprint arXiv:2306.11644, 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[21] Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: An Imm perceiving any aspect ratio and high-resolution images. In Computer Vision - ECCV 2024, pages 390-406, Cham, 2025. Springer Nature Switzerland. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 643 + ], + "type": "text", + "content": "[22] Cong Han, Yujie Zhong, Dengjie Li, Kai Han, and Lin Ma. Open-vocabulary semantic segmentation with decoupled one-pass network. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1086-1096, 2023. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[24] Zhiwei Hu, Guang Feng, Jiayu Sun, Lihe Zhang, and Huchuan Lu. Bi-directional relationship inferring network for referring image segmentation. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22161" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 4424-4433, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[25] Jiaqi Huang, Zunnan Xu, Ting Liu, Yong Liu, Haonan Han, Kehong Yuan, and Xiu Li. Densely connected parameter-efficient tuning for referring image segmentation, 2025." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 193 + ], + "type": "text", + "content": "[26] Shaofei Huang, Tianrui Hui, Si Liu, Guanbin Li, Yunchao Wei, Jizhong Han, Luoji Liu, and Bo Li. Referring image segmentation via cross-modal progressive comprehension. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10488-10497, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 195, + 295, + 249 + ], + "type": "text", + "content": "[27] Tianrui Hui, Si Liu, Shaofei Huang, Guanbin Li, Sansi Yu, Faxi Zhang, and Jizhong Han. Linguistic structure guided context modeling for referring image segmentation. In European Conference on Computer Vision, pages 59-75. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "type": "text", + "content": "[28] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 306, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 306, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 306, + 295, + 360 + ], + "type": "text", + "content": "[29] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 446 + ], + "type": "text", + "content": "[30] Gaoussou Youssouf Kebe, Padraig Higgins, Patrick Jenkins, Kasra Darvish, Rishabh Sachdeva, Ryan Barron, John Winder, Donald Engel, Edward Raff, Francis Ferraro, and Cynthia Matuszek. A spoken language dataset of descriptions for speech-based grounded language learning. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1), 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 294, + 481 + ], + "type": "text", + "content": "[31] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 483, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 483, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 483, + 294, + 524 + ], + "type": "text", + "content": "[32] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. ArXiv, abs/2308.00692, 2023. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 294, + 581 + ], + "type": "text", + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "type": "text", + "content": "[34] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. Discobox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3406-3416, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[35] Shiyi Lan, Xitong Yang, Zhiding Yu, Zuxuan Wu, Jose M Alvarez, and Anima Anandkumar. Vision transformers are good mask auto-labelers. arXiv preprint arXiv:2301.03992, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[36] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3041-3050, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[37] Liangqi Li, Jiaxu Miao, Dahu Shi, Wenming Tan, Ye Ren, Yi Yang, and Shiliang Pu. Distilling detr with visual-linguistic knowledge for open-vocabulary object detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 229 + ], + "type": "text", + "content": "[38] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 231, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 297 + ], + "type": "text", + "content": "[39] Wentong Li, Wenyu Liu, Jianke Zhu, Miaomiao Cui, XianSheng Hua, and Lei Zhang. Box-supervised instance segmentation with level set evolution. In Computer VisionECCV 2022:17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXIX, pages 1-18. Springer, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 353 + ], + "type": "text", + "content": "[40] Yanwei Li, Hengshuang Zhao, Xiaojuan Qi, Yukang Chen, Lu Qi, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Fully convolutional networks for panoptic segmentation with point-based supervision. IEEE transactions on pattern analysis and machine intelligence, 45(4):4552-4568, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 409 + ], + "type": "text", + "content": "[41] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 463 + ], + "type": "text", + "content": "[42] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, 2014. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 555, + 510 + ], + "type": "text", + "content": "[43] Chang Liu, Henghui Ding, and Xudong Jiang. Gres: Generalized referring expression segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23592-23601, 2023. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 610 + ], + "type": "text", + "content": "[45] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "text", + "content": "[46] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[47] Xuejing Liu, Liang Li, Shuhui Wang, Zheng-Jun Zha, Li Su, and Qingming Huang. Knowledge-guided pairwise reconstruction network for weakly supervised referring expression" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22162" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 95 + ], + "type": "text", + "content": "grounding. In Proceedings of the 27th ACM International Conference on Multimedia, pages 539-547, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[48] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[49] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 231 + ], + "type": "text", + "content": "[50] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 285 + ], + "type": "text", + "content": "[51] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "text", + "content": "[52] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983-73007, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 398 + ], + "type": "text", + "content": "[53] Roozbeh Mottaghi, Xianjie Chen, Xiaobai Liu, Nam-Gyu Cho, Seong-Whan Lee, Sanja Fidler, Raquel Urtasun, and Alan Yuille. The role of context for object detection and semantic segmentation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 891–898, 2014. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "text", + "content": "[54] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 532 + ], + "type": "text", + "content": "[55] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[56] Hai Nguyen-Truong, E-Ro Nguyen, Tuan-Anh Vu, Minh-Triet Tran, Binh-Son Hua, and Sai-Kit Yeung. Vision-aware text features in referring image segmentation: From object understanding to context understanding, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[57] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Erix Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. arXiv preprint arXiv:2311.03356, 2023. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[58] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. ArXiv, abs/2312.02228, 2023. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[59] Zhongwei Ren, Zhicheng Huang, Yunchao Wei, Yao Zhao, Dongmei Fu, Jiashi Feng, and Xiaojie Jin. Pixel reasoning with large multimodal model. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26374-26383, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 161 + ], + "type": "text", + "content": "[60] Anna Rohrbach, Marcus Rohrbach, Ronghang Hu, Trevor Darrell, and Bernt Schiele. Grounding of textual phrases in images by reconstruction. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 817-834. Springer, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 217 + ], + "type": "text", + "content": "[61] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "text", + "content": "[62] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 326 + ], + "type": "text", + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[64] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[65] Huiyu Wang, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Max-deeplab: End-to-end panoptic segmentation with mask transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5463-5474, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[66] Hao Wang, Pengzhen Ren, Zequn Jie, Xiao Dong, Chengjian Feng, Yinlong Qian, Lin Ma, Dongmei Jiang, Yaowei Wang, Xiangyuan Lan, et al. Ov-dino: Unified open-vocabulary detection with language-aware selective fusion. arXiv preprint arXiv:2407.07844, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "type": "text", + "content": "[67] Peng Wang, Qi Wu, Jiewei Cao, Chunhua Shen, Lianli Gao, and Anton van den Hengel. Neighbourhood watch: Referring expression comprehension via language-guided graph attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1960-1968, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[68] Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, Xiangnan Fang, Zewen He, Zhenbo Luo, Wenxuan Wang, Junqi Lin, Jian Luan, and Qin Jin. Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding. arXiv e-prints, art. arXiv:2503.13377, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "text", + "content": "[69] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 714 + ], + "type": "text", + "content": "[70] Cong Wei, Haoxian Tan, Yujie Zhong, Yujiu Yang, and Lin Ma. Lasagna: Language-based segmentation assistant for complex queries. arXiv preprint arXiv:2404.08506, 2024. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22163" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[71] Cong Wei, Yujie Zhong, Haoxian Tan, Yong Liu, Jie Hu, Dengjie Li, Zheng Zhao, and Yujiu Yang. Hyperseg: Hybrid segmentation assistant with fine-grained visual perceiver. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8931-8941, 2025. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[72] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[73] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[74] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. arXiv preprint arXiv:2312.10103, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[75] Yuwen Xiong, Renjie Liao, Hengshuang Zhao, Rui Hu, Min Bai, Ersin Yumer, and Raquel Urtasun. Upsnet: A unified panoptic segmentation network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8818-8826, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 371 + ], + "type": "text", + "content": "[76] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A simple baseline for open-vocabulary semantic segmentation with pre-trained vision-language model. In European Conference on Computer Vision, pages 736-753. Springer, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[77] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[78] Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, Bo Zhang, and Wei Chen. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[79] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. Levt: Language-aware vision transformer for referring image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18155-18165, 2022. 1, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "text", + "content": "[80] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "text", + "content": "[81] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[82] Chang-Bin Zhang, Jinhong Ni, Yujie Zhong, and Kai Han. v-clr: View-consistent learning for open-world instance segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 20307-20317, 2025. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 520 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[83] Chang-Bin Zhang, Yujie Zhong, and Kai Han. Mr. detr: Instructive multi-route training for detection transformers. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 9933-9943, 2025. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[84] Hanwang Zhang, Yulei Niu, and Shih-Fu Chang. Grounding referring expressions in images by variational context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4158-4166, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[85] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 1, 2, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[86] Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via stepwise group relative policy optimization, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[87] Tao Zhang, Xiangtai Li, Hao Fei, Haobo Yuan, Shengqiong Wu, Shunping Ji, Chen Change Loy, and Shuicheng Yan. Omg-llava: Bridging image-level, object-level, pixel-level reasoning and understanding. arXiv preprint arXiv:2406.19389, 2024. 1, 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[88] Zheng Zhang, Yeyao Ma, Enming Zhang, and Xiang Bai. Psalm: Pixelwise segmentation with large multi-modal model. In European Conference on Computer Vision, pages 74-91. Springer, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "text", + "content": "[89] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127(3):302-321, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 473 + ], + "type": "text", + "content": "[90] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15116-15127, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "type": "text", + "content": "[91] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22164" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_content_list.json b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aac99bbd3b5aa69aa922fff80059f7b64e6c810f --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_content_list.json @@ -0,0 +1,1361 @@ +[ + { + "type": "text", + "text": "Adversarial Attention Perturbations for Large Object Detection Transformers", + "text_level": 1, + "bbox": [ + 91, + 108, + 879, + 131 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zachary Yahn1, Selim Furkan Tekin1, Fatih Ilhan1, Sihao Hu1, Tiansheng Huang1, Yichang Xu1, Margaret Loper2, Ling Liu1 \n1Georgia Institute of Technology, Atlanta, GA \n2Georgia Tech Research Institute, Atlanta, USA", + "bbox": [ + 238, + 157, + 730, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zachary.yahn, stekin6, filhan, sihaohu, thuang, xuyichang}@gatech.edu, margaret.loper@gtri.gatech.edu, ling.liu@cc.gatech.edu", + "bbox": [ + 169, + 232, + 795, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 299, + 313, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial perturbations are useful tools for exposing vulnerabilities in neural networks. Existing adversarial perturbation methods for object detection are either limited to attacking CNN-based detectors or weak against transformer-based detectors. This paper presents an Attention-Focused Offensive Gradient (AFOG) attack against object detection transformers. By design, AFOG is neural-architecture agnostic and effective for attacking both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. This paper makes three original contributions. First, AFOG utilizes a learnable attention mechanism that focuses perturbations on vulnerable image regions in multi-box detection tasks, increasing performance over non-attention baselines by up to $30.6\\%$ . Second, AFOG's attack loss is formulated by integrating two types of feature loss through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is an efficient and stealthy adversarial perturbation method. It probes the weak spots of detection transformers by adding strategically generated and visually imperceptible perturbations which can cause well-trained object detection models to fail. Extensive experiments conducted with twelve large detection transformers on COCO demonstrate the efficacy of AFOG. Our empirical results also show that AFOG outperforms existing attacks on transformer-based and CNN-based object detectors by up to $83\\%$ with superior speed and imperceptibility. Code is available at: Link.", + "bbox": [ + 76, + 330, + 473, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 784, + 209, + 800 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based neural architectures and algorithms have blossomed in recent years, enhancing computer vision tasks including object detection. Attention is the core of the transformer architecture [30]. Attention allows a detector to focus on specific regions of images based on objects of interest, effectively predicting the presence and lo", + "bbox": [ + 73, + 809, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9be411cf3f959f4f01c2ca65169ed11fa1012491174efb1ba935f79ac70433be.jpg", + "image_caption": [ + "Figure 1. Successful AFOG attacks against three object detection transformers. White pixels in the attention map indicate learned areas of high importance, often corresponding to objects in the input image. This example illustrates that AFOG's adversarial attention learns which regions of an image are most susceptible to adversarial perturbations. Even for the same input, learnable adversarial attention maps may differ under different models." + ], + "image_footnote": [], + "bbox": [ + 501, + 297, + 890, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cation of each potential object. Powered by various attention mechanisms, detection transformers can handle overlapping objects more effectively than some older methods because they can capture long-range dependencies in an image, allowing models to reason about the relationships between objects. Since modern object detection transformers such as Swin [20] and DETR [4] significantly outperform traditional convolutional neural network (CNN)-based models [8], represented by Faster R-CNN [28], SSD [19] and YOLO-v3 [27], there is a pressing need to investigate and understand the vulnerabilities of large detection transformers in the presence of adversarial perturbations. Adversarial perturbations are helpful instrumentation for exposing vulnerabilities of large detection transformers, making them useful mechanisms for motivating more robust transformer models for object detection.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3184", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing attack methods struggle to disrupt transformer-based object detectors. Surrogate-based attacks (also known as black-box attacks), such as UEA [33] and RAD [5] generate perturbations on surrogate models and then test the adversarial effect against victim detectors that have similar detection architecture at inference. However, surrogate-based attacks suffer poor attack performance when the victim architectures are not similar to the surrogate models used to generate adversarial perturbations. Victim-based (also known as white-box) attacks, such as EBAD [3] and OATB [12], generate adversarial attacks by performing direct inference against victim models. Recent studies, such as AttentionFool [22], specialize in transformer-based victim models, but these transformer-only attacks are not applicable to convolutional detectors such as YOLO [27].", + "bbox": [ + 75, + 90, + 472, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present an Attention-Focused Offensive Gradient (AFOG) attack, which is a victim architecture agnostic and effective in attacking both advanced object detection transformers and traditional CNN-based detectors. AFOG by design offers three novel characteristics. First, inspired by a transformer's self-attention, we utilize a learnable attention mechanism to enable AFOG to adaptively focus its adversarial perturbations on vulnerable image areas (those areas where perturbations have the greatest effect on the output) in multi-box detection tasks. Second, we formulate the attack loss function of AFOG by integrating learnable feature-map-based attention updates with iterative injection of adversarial perturbations. Finally, we design AFOG to be an efficient and yet stealthy adversarial perturbation method. By efficient, we want AFOG to rapidly generate adversarial perturbations in minimal iterations. By stealthy, we want AFOG to generate small amounts of perturbations that are visually imperceptible and yet can cause well-trained object detectors to produce erroneous detection results. Figure 1 shows an example of successful AFOG attacks against three detection transformers. Row 1 shows benign detection from three state-of-the-art detectors, Row 2 and Row 3 show AFOG adversarial perturbations on the three detectors and their respective AFOG adversarial attention maps, and Row 4 shows the detection results under our AFOG attack. We validate AFOG with extensive experiments on twelve state-of-the-art object detection transformers with the COCO benchmark [18] and three popular families of traditional CNN-based object detectors. The results show that AFOG is consistently effective across all twelve detection transformers compared to existing methods, achieving high attack success rate and reducing benign mAP (mean Average Precision) by up to $37.79 \\times$", + "bbox": [ + 75, + 333, + 472, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 845, + 218, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing adversarial perturbations against object detectors can be broadly categorized into surrogate-model based", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(also known as black-box) and victim-model based (white-box) methods. Surrogate-based approaches in literature tend to rely on the assumption that the surrogate model is similar to the victim model. RAD [6], GHFD [32], and UEA [33] generate adversarial examples on a surrogate FR-CNN model, then attack other CNN-based detectors such as YOLO and SSD. However, these attacks show poor performance against object detection transformers [32]. For victim-model based methods, GARSDC [17] is more effective but can require more than 3000 iterations to converge. GALD [13] first attacks vision transformer classifiers and then transfers the classification-based adversarial perturbations to object detectors of similar transformer architecture, and hence is less effective by comparison. We classify EBAD [3] as a victim-based attack because it uses an ensemble of surrogate models to attack a victim, but requires access to the victim's loss function for its nested ensemble optimization. EBAD shows poor performance when transferring perturbations from a DETR surrogate to a DETR victim [24]. AttentionFool [22] is a recent victim-based attack targeting dot-product self-attention against DETR [14] with a ResNet-101 backbone. However, AttentionFool shows inconsistent performance against DETR with a ResNet-50 backbone [22]. AttentionFool is also not applicable to convolutional models such as YOLO [27] because it exclusively targets self-attention mechanisms. TOG [7] targets a prediction's objectness score and introduces vanishing and fabrication attack modes. OATB [12] uses a \"division map\" that statically emphasizes image regions during perturbation based on object location priors. DBA [16] prioritizes perturbations against image backgrounds to enhance imperceptibility, though it shows almost no effect against a Swin transformer for object detection [20]. In comparison, AFOG adversarial perturbations show strong performance against a wide variety of transformer- and CNN-based detector victims while maintaining superior imperceptibility and efficiency.", + "bbox": [ + 496, + 90, + 893, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 667, + 635, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 500, + 694, + 683, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a victim detector $f_{D}(\\vartheta, x)$ , where $x \\in \\mathcal{D}$ is a victim image $x$ and $\\mathcal{D}$ is the test set, let $x$ contain $N_{x}$ objects to be detected, denoted by $\\mathcal{O}_x = \\{O_1, O_2, \\ldots, O_{N_x}\\}$ . Each object $O_i$ ( $1 \\leq i \\leq N_x$ ) is a recognition target for the detector $f_{D}(\\vartheta, x)$ . Under benign scenarios, let $(B_i, C_i)$ denote a ground truth object $O_i$ with bounding box $B_i$ and class label $C_i$ . Let $K$ denote the total number of ground truth classes, and $C_i \\in \\{1, 2, \\ldots, K\\}$ , e.g., $K = 21$ for the VOC dataset [9], including the background class. Given input image $x$ , $f_{D}(\\vartheta, x)$ outputs the benign prediction of $N$ detected objects, denoted by $\\mathcal{R}[f_{D}(\\vartheta, x)] = \\{(b_i, c_i) | i = 1, \\ldots, N_x\\}$ . Each detected object $o_i$ is associated with", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3185", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "its predicted bounding box $b_{i}$ and predicted class label $c_{i}$ . $(b_{i},c_{i})$ is evaluated as a correct prediction if the intersection over union (IoU) of $B_{i}$ and $b_{i}$ is greater than the detection threshold $\\gamma$ (usually set to 0.5), and $C_i = c_i$ $(1\\leq i\\leq N_x,$ $x\\in \\mathcal{D})$ . The overall detection accuracy is measured using mAP (mean Average Precision) on the entire test dataset $\\mathcal{D}$", + "bbox": [ + 75, + 90, + 467, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $x_{adv}$ denote the adversarial example generated by injecting a sequence of adversarial perturbations to $x$ through an iterative attention-based learning mechanism. The goal of our AFOG attack on detector $f_{D}(\\vartheta, x)$ is to find $x_{adv}$ that maximizes the success rate of falsifying the predictions of all object recognitions for all images in $\\mathcal{D}$ , i.e.", + "bbox": [ + 73, + 181, + 468, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\arg \\max_{x\\in \\mathcal{D},i\\in N_x},\\{(b_i,c_i)\\in \\mathcal{R}[f_D(\\vartheta ,x_{adv})]$", + "bbox": [ + 99, + 273, + 408, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$(IOU(B_i, b_i) < \\gamma \\lor C_i \\neq c_i), \\min ||x - x_{adv}||_p\\}$ .", + "bbox": [ + 117, + 289, + 452, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The formula indicates that for each detected object $o_i$ , denoted by $(b_i, c_i)$ in victim image $x$ , the AFOG attack is successful on $o_i$ if the IoU of $B_i$ and $b_i$ is less than the detection threshold $\\gamma$ (usually set to 0.5) or $c_i \\in \\{1, 2, \\dots, K\\} (C_i \\neq c_i)$ , and $x_{adv}$ also satisfies the distortion constraint of $min||x - x_{adv}||_p$ , where $p$ is typically defined by $L_2$ norm, $L_0$ norm, or $L_\\infty$ norm.", + "bbox": [ + 75, + 306, + 468, + 411 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95b784c0a6270e675ace62a5dfe730cb9f1043280133c4d34f074479516dd625.jpg", + "image_caption": [ + "Benign Model Training:", + "Freeze (※) Input, Update (♀) Parameters", + "$\\partial \\mathcal{L}(\\pmb {x},\\mathcal{O};\\pmb {\\theta})$", + "Pk+1←Pk+αΓ[∂LAFOG(xad,OE;θ)]", + "$x_{adv_k}\\gets \\prod_{x\\in}\\left(x + A_k\\odot P_k\\right)$", + "AFOG Attack:", + "Freeze (k) Parameters, Update (C) Attention and Perturbation,", + "Compute (O) Adversarial Image", + "Figure 2. AFOG attack framework. During attack iteration the adversarial perturbation is updated via bounding-box and class loss gradient propagation. An adversarial attention map concurrently learns to apply per-pixel amplification or dampening to the perturbation, focusing its effects on vulnerable image regions." + ], + "image_footnote": [], + "bbox": [ + 254, + 444, + 431, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Adversarial Loss Optimization", + "text_level": 1, + "bbox": [ + 76, + 709, + 349, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AFOG attacks the victim model via an iterative projected gradient descent method [23]. Figure 2 provides an illustration of the AFOG attack framework. At the initialization step, the attack first propagates the unaltered victim image $x$ through the feed-forward network (FFN) of the victim model $f_{D}(\\vartheta ,x)$ to obtain its benign prediction $\\mathcal{O}_x = \\{(b_i,c_i)|i = 1,\\dots ,N_x\\}$ .", + "bbox": [ + 75, + 733, + 468, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a scenario where no ground truth is available, the attack assumes these benign predictions to be correct labels for the input image $x$ . For the remaining section, we will use $(B_i, C_i)$ in the context of performing an attack on $f_D(\\vartheta, x)$", + "bbox": [ + 76, + 839, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to obtain $x_{adv}$ . Upon initialization for each input image $x \\in \\mathcal{D}$ , the AFOG attack corrupts the victim image $x$ by adding perturbations generated via an element-wise multiplication of two components using Equation 1:", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {a d v _ {k}} \\leftarrow \\prod_ {x, \\epsilon} \\left(x + A _ {k} \\odot P _ {k}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 155, + 890, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$A$ is the attention map, and $P$ is the perturbation map. Here $\\odot$ represents the Hadamard matrix product and $\\Pi$ is a projection onto a hypersphere centered on unaltered victim image $x$ with the radius $\\epsilon$ as the maximum perturbation budget. $A$ and $P$ are initialized according to Equation 2.", + "bbox": [ + 496, + 190, + 890, + 266 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {0} \\sim 1, P _ {0} \\sim \\text {R a n d o m} (- \\epsilon , \\epsilon) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 270, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Random $(-\\epsilon, \\epsilon)$ is a uniform random distribution. The perturbed image $x_{adv}$ propagates through the FFN of the victim model. The attack loss function $\\mathcal{L}_{AFOG}$ assesses the adversarial output $\\mathcal{O}_{x_{adv}}$ and benign output $\\mathcal{O}_x$ , computing a loss that reflects the attack's progress in corrupting image $x$ . This attack loss is formulated in Equations 3, 4, and 5.", + "bbox": [ + 496, + 290, + 890, + 381 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {A F O G} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\tag {3} \\\\ + \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 392, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, o _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, o _ {a d v _ {i}}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 455, + 890, + 510 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, c _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, c _ {a d v _ {i}}\\right) \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 532, + 890, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a victim image $x$ , the optimized attack loss $\\mathcal{L}_{AFOG}$ is attained by making the model incorrectly predict every target object. We achieve this loss optimization by both falsifying every target object's bounding box prediction and its class label prediction. This can be represented by optimizing an adversarial bounding box loss $\\mathcal{L}_{bbox}$ and an adversarial class label prediction loss $\\mathcal{L}_{cls}$ for all $N_x$ target objects. Conceptually, this suppresses the confidence of the original correct bounding box and class label prediction, while increasing the confidence of the adversarial prediction of an incorrect bounding box or class label.", + "bbox": [ + 496, + 585, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recall Figure 2: the AFOG attack freezes the model parameters $\\vartheta$ and uses the gradient of the AFOG attack loss to update the attention map $A$ and the perturbation $P$ by backpropagation, generating the next iteration of the adversarial perturbation. A newly corrupted image $x_{adv}$ is created upon injecting the updated adversarial perturbation by following Equation 1, where A and P are updated via Equations 6 and 7:", + "bbox": [ + 496, + 751, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {k + 1} \\leftarrow A _ {k} + \\alpha_ {A} \\sigma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial A _ {k}} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 872, + 890, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3186", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3096f28c10076b522da0a6ecc93a28d43b2fa7770c0fba59996f64f83ac0110f.jpg", + "image_caption": [ + "Figure 3. Example attention map generation on DETR-R101 for $k = 10$ iterations. AFOG attention initially learns important regions of the image, then iteratively updates to achieve an effective attack. It finds vulnerable pixels in foreground objects as well as background regions, including in unexpected or unintuitive areas such as the sky above the boat." + ], + "image_footnote": [], + "bbox": [ + 222, + 95, + 750, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {k + 1} \\leftarrow P _ {k} + \\alpha_ {P} \\Gamma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial P _ {k}} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 353, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\alpha_{A}$ and $\\alpha_{P}$ denote the attention map learning rate and the perturbation learning rate respectively, $\\sigma$ is a normalization function, $\\Gamma$ is a sign function, $\\mathcal{L}_{AFOG}$ is the attack loss, $\\vartheta$ is the model parameters, and $\\mathcal{O}_x$ is the model's benign predictions from input $x$ . This attack process repeats until the number of attack iterations is reached. Algorithm 1 provides the pseudocode.", + "bbox": [ + 75, + 402, + 472, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adversarial Attention Mechanism", + "text_level": 1, + "bbox": [ + 76, + 530, + 375, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key innovation of our AFOG method is empowering perturbation generation with a learnable attention mechanism. Motivated by the intuition that certain parts of a victim image are more susceptible to adversarial perturbation than others, we add an attention map to focus the perturbation on vulnerable pixels. The AFOG attack concurrently learns the adversarial attention map alongside the perturbation to iteratively maximize the attack loss $\\mathcal{L}_{\\text{AFOG}}$ (recall Equation 3). Figure 3 gives an illustrative example of the iterative learning of AFOG's adversarial attention map and the corresponding perturbation. Unlike focusing mechanisms in other methods such as [33], AFOG's attention map dynamically updates during attack iteration, leaving it unconstrained by assumptions about foreground [12], background [16], or region proposal [33] importance from static methods. We posit that pixel importance may not correspond to human intuition, and design AFOG to iteratively learn pixel importance instead. We observe that in early iterations the attention mechanism tends to focus on the primary objects in an image, and then branches out to affect surrounding areas as the attack progresses. This adaptability is a key feature of learnable attention, showcasing an advantage over static attention maps.", + "bbox": [ + 75, + 554, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 AFOG attack on an input image.", + "text_level": 1, + "bbox": [ + 500, + 340, + 807, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: Victim image $x \\in \\mathcal{D}$ , test-set $\\mathcal{D}$ , Victim pretrained model $f_{D}(\\vartheta)$ , Perturbation step size $\\alpha_{P}$ , Attention step size $\\alpha_{A}$ , Number of iterations $T$ , Maximum perturbation $\\epsilon$ .", + "bbox": [ + 500, + 357, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize $\\mathcal{O}_x\\gets f_D(x;\\vartheta)$", + "2: Initialize attention map $A_0 \\gets 1$", + "3: Initialize perturbation $P_0 \\gets \\text{Random}(-\\epsilon, \\epsilon)$ ;", + "4: Initialize step variable $k \\gets 1$" + ], + "bbox": [ + 509, + 417, + 833, + 477 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5: while $k < T$ do", + "6: Attack image $x_{advk} \\gets \\prod_{x,\\epsilon}(x + A_k \\odot P_k)$ ;", + "7: Forward propagate $x_{adv}$ through $f_{D}(\\vartheta, x_{adv})$ ;", + "8: Compute bbox-loss $\\mathcal{L}_{bbox}(x_{adv},\\mathcal{O}_x;\\vartheta)$", + "9: Compute $\\operatorname{cls}$ -loss $\\mathcal{L}_{cls}(x_{adv},\\mathcal{O}_x;\\vartheta)$", + "10: $\\mathcal{L}_{AFOG}(x_{adv},\\mathcal{O}_x;\\vartheta) = \\mathtt{bbox - loss} + \\mathtt{cls - loss};$", + "11: Calculate losses with respect to $A_{k}$ and $P_{k}$ : $\\mathcal{L}_A(x_{adv},\\mathcal{O}_x;\\vartheta),\\mathcal{L}_P(x_{adv},\\mathcal{O}_x;\\vartheta)$ ;", + "12: Normalize attention loss $\\mathcal{L}_A\\gets \\mathrm{Norm}(\\mathcal{L}_A)$", + "13: Take sign of perturbation loss $\\mathcal{L}_P\\gets \\mathrm{Sign}(\\mathcal{L}_P)$", + "14: $A_{k + 1}\\gets A_k - \\alpha_A\\mathcal{L}_A;$", + "15: $P_{k + 1}\\gets P_k - \\alpha_P\\mathcal{L}_P;$", + "16: $k\\gets k + 1;$", + "17: end while", + "18: $x_{advk + 1}\\gets \\prod_{x,\\epsilon}(x + A_{k + 1}\\odot P_{k + 1})$", + "19: return $x_{adv}$" + ], + "bbox": [ + 504, + 417, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Special Cases of the AFOG Attack", + "text_level": 1, + "bbox": [ + 500, + 771, + 800, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We explore two special cases of our AFOG attack, each of which targets a specific malicious detection behavior in the victim detector. The first special case is an object vanishing attack, coined AFOG-V, which attempts to attack the objectness detection task of multi-box object detection. The goal of this special case AFOG attack is to make the victim model unable to detect any object, making all object", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detections vanish for victim image $x$ . We implement the AFOG-V attack by changing the initialization: we use an empty set instead of forward propagation of $x$ to obtain the benign detection results for the $N_{x}$ objects as the assumed ground-truth in AFOG-V attack. Let $\\varnothing$ denote an altered version of $\\mathcal{O}_x$ that contains no predictions. The formulation of the AFOG-V attack loss function is given by Equation 8:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {A F O G _ {V}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\varnothing ; \\vartheta\\right) \\tag {8} \\\\ - \\mathcal {L} _ {c l s} (x _ {a d v}, \\varnothing ; \\vartheta) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 208, + 468, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second special case is object fabrication, coined as AFOG-F, which attempts to attack the bounding-box detection task by generating perturbations that lead to spurious detections (i.e., false positives). Similarly, we revise the AFOG attack process: instead of keeping the benign predictions with the confidence score above a certain tunable threshold (0.5 by default), for AFOG-F, we remove this threshold and allow the benign detection set $\\mathcal{O}_x$ to include a much larger set of \"ground truth\" targets, which inevitably include erroneous detections. Accordingly the AFOG-F loss function is given in Equation 9:", + "bbox": [ + 75, + 252, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {A F O G _ {F}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\tag {9} \\\\ - \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 417, + 468, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{O}_F$ is a modified version such that the likelihood score of each prediction is set to 1.0. We explore these two specialized cases of our AFOG attack to gain a deeper understanding of the adverse effects of adversarial perturbations on different detection transformers.", + "bbox": [ + 75, + 460, + 468, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 550, + 209, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 575, + 266, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the Common Objects in Context (COCO) 2017 [18] test-dev split to evaluate the performance of our attack on contemporary object detectors. COCO 2017 is a standard benchmark for evaluating object detector performance, and its test-dev set has 80 object categories across 5,000 images. We use the PASCAL Visual Objects Challenge (VOC) 2007 [9] dataset to compare our attack with the state-of-the-art attacks on CNN-based object detectors, because existing attacks designed only for CNN-based detectors all report their evaluation with VOC. The VOC devkit split contains 4,952 images with 20 object classes. All experiments were performed on an NVIDIA A100. We list our models and their benign performance in Table 1. We provide model details in Section ?? of the Supplementary Material.", + "bbox": [ + 75, + 598, + 468, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We select a variety of model sizes, ranging from Detection Transformer (DETR) [4], a lightweight 40 million parameter model, to EVA [10], a vision-focused foundation model with over one billion parameters. We use the Detrex framework [29], built on top of Detectron2 [34], to standardize our model implementations via DINO [37]. DINO", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and Detrex adapt many general purpose vision models to object detection. The Detrex framework standardizes all images in [0, 1]. We also attack the original versions of several transformer models to demonstrate that the AFOG attack also works beyond the Detrex framework. All 12 transformer models and FRCNN use PyTorch [26] implementations, whereas SSD300 [19] and YOLOv3 [27] use Tensorflow [1]. We use mean average precision (mAP) to evaluate both benign and attacked performance. The Average Precision (AP) is given by interpolating the product of precision and recall at several decision thresholds. The mean AP (mAP) is the average of APs over the number of object classes. A lower mAP indicates that the detector is less effective for the object detection task. The victim detector's mAP under adversarial perturbations reflects the degree of degradation in detection performance for the victim detector when compared with its benign mAP. We measure imperceptibility by reporting the average distortion of the attack via four metrics: $L_{2}$ norm, $L_{0}$ norm, semantic structural similarity index (SSIM), and mean distortion $\\mu_{\\Delta}$ . We compute each metric per image and report the average over all images in COCO 2017 test-dev. Our experiments in the next section show that setting the total number of attack iterations to 10 is effective for all 12 state-of-the-art detection models on the COCO benchmark.", + "bbox": [ + 496, + 90, + 893, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Effectiveness Comparison of AFOG Attack", + "text_level": 1, + "bbox": [ + 498, + 481, + 867, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows the effectiveness of the AFOG attack and its two special extensions AFOG-V and AFOG-F on each of the twelve detection transformers with varying model sizes (Col. 2). We use their benign mAP scores (Col. 3) as the reference for the evaluation. We make three observations: (i) under AFOG attacks all 12 transformers suffer a drastic performance drop in mAP scores, ranging from $2.5 \\times$ reduction to $37.8 \\times$ reduction. (ii) Compared to the AFOG generic attack, the special extension AFOG-V offers stronger attack success rates in terms of mAP reduction for 11 out of 12 transformers, except DETR-R50 [4]; and the special extension AFOG-F offers stronger attack success rates for 7 out of 12 transformers, except DETR-R50 [4], DETR-R101 [4], Deformable-DETR [38], R50 (DINO) [11], Swin-L (DINO) [20] (iii) For AlignDETR [2], AFOG-F achieves a mAP of 1.36 and outperforms both AFOG (mAP of 18.13) and AFOG-V (mAP of 1.64), and all three attacks significantly reduce the benign mAP by a factor ranging from $2.8 \\times$ to $37.8 \\times$ .", + "bbox": [ + 496, + 506, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare AFOG against eleven benchmark attacks on DETR and Swin in Table 2. We observe that AFOG achieves superior performance on both DETR-R50 and Swin-L, notably recording an $82.7\\%$ improvement over the next strongest attack on Swin. AFOG also uses the smallest perturbation budget and fewest iterations, indicating its superior imperceptibility and speed.", + "bbox": [ + 496, + 795, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3188", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63f6277155868b5513c5aa1e60099b6be22f87958231a8b1b63bd896f9d0f265.jpg", + "image_caption": [ + "Figure 4. Visualization with six examples, comparing benign scenario and AFOG attack results on Detection Transformer (DETR) with ResNet-50 backbone. Six diverse test images contain objects with varying classes, scales, perspectives, lighting, motion, and textures." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 222, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ff461ff15b503589fc4d7798b6cc9e4a3c591555f0baf2ea28c39a2f4a57990.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 88, + 367, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff1d53d2bcd8423a5e4684fa4ac683dacd3655a1c5ca86ed5d4dfb921f49182e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 89, + 493, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d00a828367bd278d88d795845f4bb35b53efdbfb3956521ac6690da4c0515b66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 625, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80d72ae204bda3c8d7e809d55187021db01e411f7e1a650ca4d1442bf0d7313d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 90, + 743, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb2722a37a08ac01f2d52e29a8da55f72bde869fd5fb48b908266fb7a65e92a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 89, + 893, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/087d4cbc7490429984f55c533ed5fc29c4e0f1b5202bd94cd9435c31daf23213.jpg", + "table_caption": [ + "Table 1. AFOG effectiveness over 12 detection transformers, measured by mAP on perturbed images. (*) indicates DINO framework used with the corresponding backbone for object detection." + ], + "table_footnote": [], + "table_body": "
ModelParams (M)BenignAFOGAFOG-VAFOG-F
DETR-R50 [4]39.842.14.14.59.8
DETR-R101 [4]76.043.55.25.111.3
Deform.-DETR [38]40.044.54.81.57.1
R50*[11]47.649.25.31.56.3
AlignDETR [2]47.651.418.11.61.4
ViTDet* [14]108.154.93.80.92.8
ConvNeXt* [21]219.055.43.91.93.1
Swin-L* [20]217.256.87.32.48.6
InternImage* [31]241.056.97.32.85.1
FocalNet* [36]228.958.57.32.55.1
EVA* [10]1037.262.112.24.18.7
DETA [25]218.862.925.63.74.3
", + "bbox": [ + 78, + 319, + 475, + 450 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e68bb84be2fc700d77a4e3fe424253063524ab4b074bb2edcbdaa246c0e9609.jpg", + "table_caption": [ + "Table 2. Comparison benchmark of AFOG against other state-of-the-art object detection attacks on DETR and Swin. Results are theirs. (*) indicates results from [32]. (†) indicates results from [24]. (-) indicates no result." + ], + "table_footnote": [], + "table_body": "
AttackTypePert. BudgetIters.Adversarial mAP
DETR-R50Swin
GARSDC [17]Surrogate0.053000+6.0-
GALD [13]Surrogate0.0631020.6-
RAD* [6]Surrogate0.0631027.247.2
GHFD* [32]Surrogate0.0635012.742.3
UEA* [33]Surrogate0.0635028.550.7
DAG* [35]Surrogate0.0635028.650.7
RAP* [15]Surrogate0.0635024.749.5
EBAD†[3]Victim0.0391034.9-
AttentionFool [22]Victim-10-15021.0-
OATB [12]Victim0.0782026.6-
DBA [16]Victim-50-56.7
AFOGVictim0.031104.17.3
", + "bbox": [ + 78, + 521, + 478, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 shows six visual examples and the comparison of DETR results for benign detection (Row 1) and adversarial detection under the AFOG attack (Row 2) scenarios. The first example shows that DETR detects the person and surfboard correctly under no attack (Row 1) but outputs erroneous detections under the AFOG attack (Row 2). Additional visual examples for all twelve transformers are available in the Supplementary Material.", + "bbox": [ + 75, + 717, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the distortion effect and timing cost for each of the 12 transformer models in Table 3. We note several interesting observations: (i) Considering the $L_{2}$ and SSIM metrics, AFOG attacks induce very similar levels of", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distortion for 10 out of 12 transformer models. The two largest detection transformers, DETA [25] and EVA [10], show worse $L_{2}$ and SSIM scores for AFOG, AFOG-V and AFOG-F. (ii) Considering the $L_{0}$ metric, only DETA shows the lowest $L_{0}$ value consistently for AFOG, AFOG-V and AFOG-F. (iii) As expected, the time required to attack a detection transformer model is strongly correlated to the number of model parameters. For instance, EVA [10] has over 1 billion parameters, and took much longer on average to attack an input image over 10 iterations compared to the other 11 models. (iii) DETA [25] has 218.8 million parameters and is similar to other models within the Detrex framework: FocalNet [36], InternImage [31], Swin-L [20] and ConvNext [21]. However, we observe that it consistently takes about $1.5 - 2 \\times$ longer time to attack one input image on average with our AFOG attack or its special case extensions AFOG-V or AFOG-F. (iv) AFOG, AFOG-V and AFOG-F have similar average perturbation magnitude.", + "bbox": [ + 496, + 282, + 893, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Effectiveness on CNN-Based Detectors", + "text_level": 1, + "bbox": [ + 498, + 571, + 831, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section compares AFOG with four attacks designed for CNN-based models: TOG [7], UEA [33], RAP [15], and DAG [35]. They are surrogate-based attacks, and a majority of them (e.g., DAG, RAP, UEA) can only directly attack two-stage detectors (exemplified by Faster-R-CNN [28]) and rely on adversarial transferability to attack single-stage CNN-based detectors like YOLOv3 [27] and SSD [19]. Table 4 reports the comparison results. We make two observations: (i) AFOG and TOG are the only two victim-based attacks against single-stage detectors, represented by YOLOv3 and SSD. Both AFOG and TOG can drastically reduce the benign mAP of YOLOv3 $(83.43\\%)$ and SSD $(76.11\\%)$ . For YOLOv3, TOG is a stronger attack, achieving mAP of 0.56 compared to mAP of AFOG being 2.62. However, for SSD-300, AFOG is a stronger attack, achieving mAP of 0.50, compared to mAP of TOG (0.86). (ii) For two-stage CNN-based object detectors, exemplified by Faster-R-CNN, AFOG outperforms all four other victim-based attacks, achieving the lowest adversarial mAP score of $2.07\\%$ . (iii) AFOG also uses the same distortion mag", + "bbox": [ + 496, + 598, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3189", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a5e49bc014f5ec7b7f977976f041af48c1a612e388d53b5fb5e7cf54d8ceb511.jpg", + "table_caption": [ + "Table 3. Timing and Imperceptibility Results. $L_{2}$ represents the average $L_{2}$ norm difference between perturbed and clean images, $L_{0}$ is the average proportion of perturbed pixels, SSIM is the structural similarity index measure, $\\mu_{\\Delta}$ is the average perturbation magnitude, and t is average total attack time for all ten iterations in seconds." + ], + "table_footnote": [], + "table_body": "
ModelL2AFOGtimeL2AFOG-VtimeL2AFOG-Ftime
L0SSIMμΔL0SSIMμΔL0SSIMμΔ
DETR-R50 [4]0.03220.97070.87150.01731.450.03230.97070.87160.01720.990.03230.97100.87170.01721.21
DETR-R101 [4]0.03230.97070.87210.01731.470.03230.97060.87240.0131.160.03230.97080.87240.01721.70
Deform.-DETR [38]0.03230.97190.87110.01741.630.03230.97130.87160.01731.630.00120.97140.87170.01731.87
R50 [11]0.03170.96580.83430.01712.700.03170.96500.83480.01702.340.03170.96540.83460.01703.16
AlignDETR [2]0.03190.96570.83470.01702.410.03190.96460.83490.01702.330.03190.96470.83490.01703.29
ViTDet [14]0.03180.96710.83530.01716.880.03180.96570.83610.01706.670.03180.96640.83550.01717.33
ConvNext [21]0.03180.96660.83420.01715.380.03180.96540.83490.01705.260.03180.96630.83470.01705.86
Swin-L [20]0.03270.97240.86730.01757.130.03270.97160.86800.01737.280.03270.97220.86780.01748.99
InternImage [31]0.03180.96650.83600.01706.350.03180.96530.83670.01706.230.03180.96600.83640.01716.70
FocalNet [36]0.03200.96650.83650.01728.960.03200.96570.83780.01718.840.03200.96590.83740.01719.74
EVA [10]0.03700.96660.82400.017254.340.03700.96650.82460.017154.530.03700.96650.82420.017151.18
DETA [25]0.04810.96620.80960.017013.200.04810.96540.80990.017013.100.04810.96540.80990.017013.13
", + "bbox": [ + 76, + 128, + 916, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nitude budget $L_{\\infty}$ as TOG, while achieving higher attack success rate (lower mAP) on SSD-300 and FRCNN with decreased $L_{2}$ distortion cost. Hence, AFOG excels at attacking both transformer-based detectors and CNN-based detectors.", + "bbox": [ + 75, + 297, + 468, + 375 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a251e0f43cdc67ee5f4dd8d9d9225472c2684158ca8b1e44d58e013259221350.jpg", + "table_caption": [ + "Table 4. Comparing AFOG with four state-of-the-art attacks on representative CNN-based object detectors. mAPs of existing attacks were taken from respective papers [7]. (-) indicates N/A." + ], + "table_footnote": [], + "table_body": "
ModelAttackmAPtDistortion Cost
\\(L_{\\infty}\\)\\(L_{2}\\)\\(L_{0}\\)SSIM
YOLOv3Benign83.430.00.00.00.01.0
TOG [7]0.560.980.0310.0830.9840.875
AFOG2.281.310.0310.0130.8550.801
SSD-300Benign76.110.00.00.00.01.0
UEA [33]20.0-----
DAG [35]64.0-----
TOG [7]0.860.390.0310.1200.9750.879
AFOG0.500.490.0310.0220.8580.793
FRCNNBenign67.370.00.00.00.01.0
UEA [33]5.00.170.3430.1910.9590.652
RAP [15]4.784.040.0820.0100.5310.994
DAG [35]3.567.990.0240.0020.4930.999
TOG [7]2.641.680.0310.0580.9760.862
AFOG2.382.110.0310.0190.8540.788
", + "bbox": [ + 84, + 422, + 459, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Effect of Learnable Self-Attention", + "text_level": 1, + "bbox": [ + 76, + 619, + 375, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We isolate and analyze the role of the learnable self-attention mechanism in empowering AFOG to successfully corrupt transformer-based models by probing for weak spots and performing fast and human-imperceptible perturbations. Figure 5 shows a comparison between AFOG with and without its learnable attention mechanism, measured by percent difference in adversarial mAP scores. Ablation results show that our learnable attention mechanism improves performance by up to $30.6\\%$ on InternImage, with an average of $15.1\\%$ improvement across all models.", + "bbox": [ + 75, + 643, + 468, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 shows the self-attention map for the last encoder layer of DETR [4] at initialization and three progressive attack iterations $(3^{rd}, 4^{th}$ , and $8^{th}$ ). We observe that, as the attack progresses in iterations (indicated by grey arrows), the self-attention map becomes increasingly less associated with the benign detected objects. By the $8^{th}$ iteration of the AFOG attack, the self-attention maps have become almost", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "entirely disjoint from the benign detections of the victim model (DETR) prior to the attack. We note that, when the model's self-attention map that reflects learned detection knowledge is changed, it is unlikely for prior knowledge to be kept intact. During the AFOG attack, the perturbed inputs focused by AFOG's own adversarial attention corrupt the pixel-level relationships learned by a well-trained victim model prior to adversarial injection (e.g., $k = 1$ ), causing catastrophic forgetting to occur.", + "bbox": [ + 496, + 297, + 893, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2dd6e1409fa5c9c44f15609da7c034edba9500a9ca8b2c390a724a9e7a2be09c.jpg", + "image_caption": [ + "Figure 5. Improvement (\\% difference) of AFOG with attention over AFOG without attention. The learnable attention mechanism improves performance by as much as $30.6\\%$ on InternImage, with an average improvement of $15.1\\%$ across all models." + ], + "image_footnote": [], + "bbox": [ + 509, + 449, + 883, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Worst Case Analysis", + "text_level": 1, + "bbox": [ + 498, + 698, + 694, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experimental results on 12 transformer-based detectors and 3 representative CNN-based detectors have shown that AFOG is a fast and effective attack. However, there are still cases where AFOG fails in its adversarial perturbations within a total of 10 iterations, one setting of our attack termination hyper-parameter. By investigating those cases where AFOG did not succeed, we found that the most common situation is where the attack fails to corrupt the central subject of the victim input image, e.g., a person. Figure 7 gives three visual examples, with the top one (Rows 1-2) being a successful attack followed by two failed cases (Rows 3-6) when attacking DETR-R50. Columns 2-3 in all six", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3190", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cb4ab033a6219e383c0db53918950e3a520a5676b12164acadb535148d5965a6.jpg", + "image_caption": [ + "Figure 6. DETR self-attention weights analysis. Each blue frame shows the self-attention weights of DETR-R50's last encoder layer at the red indicated point in the image at a given attack iteration $k$ . In early iterations, the model understands a strong association between the indicated red point on the elephant and the rest of the elephant shape. As the attack iterates (grey arrows), this association becomes corrupted, evidenced by the lack of encoder self-attention structure in later stages. By the $8^{th}$ iteration, the selected point has lost association with the rest of the elephant object, demonstrating that AFOG has disrupted learned knowledge." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 468, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rows show DETR's encoder's last layer's self-attention map for the first and second locations indicated by red points in the first column. Rows 1, 3, and 5 show benign situations and their corresponding self-attention maps, and Rows 2, 4, 6 show the self-attention maps after the final attack iteration. The fourth column shows AFOG's adversarial attention map after the final attack iteration.", + "bbox": [ + 75, + 393, + 468, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We make three observations: (i) For the successful attack case (Rows 1-2), the self-attention maps of victim model DETR-R50 lose association between the indicated red point and the rest of the foreground object it comprises. (ii) In both unsuccessful cases (Rows 3-6), AFOG fails to disrupt this association. (iii) Consider AFOG's adversarial attention maps from the successful case in Figure 7 (Row 2, Col. 4), we observe a clear focus on the key objects in the victim input image. In comparison, AFOG's adversarial attention maps for the two failure cases (Row 4, Col. 4 and Row 6, Col. 4) both fail to focus on foreground objects. The first failure case (Rows 3 and 4) shows that the AFOG attention map attributes significant weight to the distant person in the left of the victim image, resulting in a large number of fabricated predictions in that region. Similarly, the attention map for the second victim image (Rows 5 and 6) appears to entirely miss the fire hydrant, resulting in a failure to disrupt this object. In these cases, AFOG's adversarial attention has learned to focus on the wrong pixels.", + "bbox": [ + 75, + 500, + 470, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 800, + 194, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented AFOG, an Attention-Focused Offensive Gradient attack. AFOG is effective for attacking both advanced object detection transformers and traditional CNN-based detectors with a unified, architecture-agnostic framework. AFOG utilizes a learnable attention mecha", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf13afda4d23d1131fe77bb746aa8a22bf3a95849100c969f46f57648fbc60e4.jpg", + "image_caption": [ + "Figure 7. Worst case analysis with DETR-R50 as victim model. Rows 1-2 show a benign image and the successful AFOG perturbed image. Rows 3-6 show two cases where AFOG attack did not succeed. Columns 2-3 show DETR's encoder's self-attention map for the two locations marked in red in column one. Column 4 shows AFOG's internal attention map, where white pixels indicate high attention and blue indicates low. In the successful case (Rows 1-2), encoder self-attention maps (Cols. 2-3) lose associative structure. In the unsuccessful cases (Rows 4 and 6), AFOG fails to disrupt this structure and predictions are preserved. AFOG adversarial attention maps (Col. 4) show focus on important object structure in the successful case (Row 2) and lack of focus in the unsuccessful cases (Rows. 4, 6)." + ], + "image_footnote": [], + "bbox": [ + 503, + 85, + 890, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nism to enable its adversarial perturbations to focus on vulnerable areas of images in multi-box detection tasks. AFOG's attack loss function integrates multiple feature losses (e.g., bounding-box loss, class loss) through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is efficient and stealthy. Adversarial perturbations generated by AFOG are visually imperceptible and yet can cause well-trained detectors to fail miserably. Extensive experiments on state-of-the-art object detectors show that AFOG is consistently effective across twelve object detection transformers. Comparative evaluation with nearly a dozen SOTA methods shows that AFOG significantly outperforms surrogate-based and victim-based attacks on both object detection transformers and CNN-based object detectors.", + "bbox": [ + 496, + 613, + 890, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is partially sponsored by the NSF CISE grants 2302720 and 2312758, CISCO Edge AI program, and GTRI PhD Fellowship.", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3191", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 5", + "[2] Zhi Cai, Songtao Liu, Guodong Wang, Zheng Ge, Xiangyu Zhang, and Di Huang. Align-detr: Improving detr with simple iou-aware bce loss, 2023. 5, 6, 7", + "[3]Zikui Cai, Yaoteng Tan, and M. Salman Asif. Ensemble-based blackbox attacks on dense prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4045-4055, 2023. 2, 6", + "[4] Nicolas Carrion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-223, 2020. 1, 5, 6, 7", + "[5] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2", + "[6] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2, 6", + "[7] Ka-Ho Chow, Ling Liu, Margaret Loper, Juhyun Bae, Mehmet Emre Gursoy, Stacey Truex, Wenqi Wei, and Yanzhao Wu. Adversarial objectness gradient attacks in real-time object detection systems. In IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications, pages 263-272. IEEE, 2020. 2, 6, 7", + "[8] Papers With Code. Object detection on coco test-dev. https://paperswithcode.com/sota/object-detection-on-coco.1", + "[9] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 2, 5", + "[10] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022.5,6,7", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Shu. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 06 2016. 5, 6, 7", + "[12] Zhuo Leng, Zesen Cheng, Pengxu Wei, and Jie Chen. Object-aware transfer-based black-box adversarial attack on object detector. In Qingshan Liu, Hanzi Wang, Zhanyu" + ], + "bbox": [ + 78, + 114, + 468, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ma, Weishi Zheng, Hongbin Zha, Xilin Chen, Liang Wang, and Rongrong Ji, editors, Pattern Recognition and Computer Vision, pages 278-289, Singapore, 2024. Springer Nature Singapore. 2, 4, 6", + "[13] Tuo Li and Yahong Han. Improving transferable adversarial attack for vision transformers via global attention and local drop. Multimedia Systems, 29:3467 - 3480, 2023. 2, 6", + "[14] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 2, 6, 7", + "[15] Yuezun Li, Daniel Tian, Mingching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. In BMVC, 2018. 6, 7", + "[16] Jiawei Lian, Shaohui Mei, Xiaofei Wang, Yi Wang, Lefan Wang, Yingjie Lu, Mingyang Ma, and Lap-Pui Chau. Attack anything: Blind dnns via universal background adversarial attack, 2024. 2, 4, 6", + "[17] Siyuan Liang, Longkang Li, Yanbo Fan, and Xiaojun Jia. A large-scale multiple-objective method for black-box attack against object detection. European Conference on Computer Vision (ECCV) 2022, pages 619-636, 2022. 2, 6", + "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, ECCV 2014, pages 740-755. Springer International Publishing, 2014. 2, 5", + "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing. 1, 5, 6", + "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6, 7", + "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6, 7", + "[22] Giulio Lovisotto, Nicole Finnie, Maurizio Munoz, Chaithanya Kumar Murnmadi, and Jan Hendrik Metzen. Give me your attention: Dot-product attention considered harmful for adversarial patch robustness. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15213-15222, 2022. 2, 6", + "[23] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 3", + "[24] Khoi Nguyen Tiet Nguyen, Wenyu Zhang, Kangkang Lu, Yuhuan Wu, Xingjian Zheng, Hui Li Tan, and Liangli Zhen. A Survey and Evaluation of Adversarial Attacks for Object Detection, Aug. 2024. arXiv:2408.01934 [cs]. 2, 6", + "[25] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbihl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 6, 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3192", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[27] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 1, 2, 5, 6", + "[28] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 1, 6", + "[29] Tianhe Ren, Shilong Liu, Feng Li, Hao Zhang, Ailing Zeng, Jie Yang, Xingyu Liao, Ding Jia, Hongyang Li, He Cao, Jianan Wang, Zhaoyang Zeng, Xianbiao Qi, Yuhui Yuan, Jianwei Yang, and Lei Zhang. detrex: Benchmarking detection transformers, 2023. 5", + "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1", + "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 6, 7", + "[32] Yang Wang, Yunfei Zheng, Lei Chen, Zhen Yang, Jingwei Wu, and Tieyong Cao. Gradient-guided hierarchical feature attack for object detector. Journal of King Saud University - Computer and Information Sciences, 36(1):101901, 2024. 2, 6", + "[33] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 954–960. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 4, 6, 7", + "[34] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5", + "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie1, and Alan Yuill. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision, pages 1378-1387, 2017. 6, 7", + "[36] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 6, 7", + "[37] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection, 2022. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5, 6, 7", + "bbox": [ + 501, + 90, + 890, + 146 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3193", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_model.json b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d6eb3f2a9ee14ebd51c51eceee802c8083ec2e22 --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.109, + 0.88, + 0.132 + ], + "angle": 0, + "content": "Adversarial Attention Perturbations for Large Object Detection Transformers" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.159, + 0.731, + 0.23 + ], + "angle": 0, + "content": "Zachary Yahn1, Selim Furkan Tekin1, Fatih Ilhan1, Sihao Hu1, Tiansheng Huang1, Yichang Xu1, Margaret Loper2, Ling Liu1 \n1Georgia Institute of Technology, Atlanta, GA \n2Georgia Tech Research Institute, Atlanta, USA" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.233, + 0.796, + 0.266 + ], + "angle": 0, + "content": "{zachary.yahn, stekin6, filhan, sihaohu, thuang, xuyichang}@gatech.edu, margaret.loper@gtri.gatech.edu, ling.liu@cc.gatech.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.3, + 0.314, + 0.316 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.474, + 0.758 + ], + "angle": 0, + "content": "Adversarial perturbations are useful tools for exposing vulnerabilities in neural networks. Existing adversarial perturbation methods for object detection are either limited to attacking CNN-based detectors or weak against transformer-based detectors. This paper presents an Attention-Focused Offensive Gradient (AFOG) attack against object detection transformers. By design, AFOG is neural-architecture agnostic and effective for attacking both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. This paper makes three original contributions. First, AFOG utilizes a learnable attention mechanism that focuses perturbations on vulnerable image regions in multi-box detection tasks, increasing performance over non-attention baselines by up to \\(30.6\\%\\). Second, AFOG's attack loss is formulated by integrating two types of feature loss through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is an efficient and stealthy adversarial perturbation method. It probes the weak spots of detection transformers by adding strategically generated and visually imperceptible perturbations which can cause well-trained object detection models to fail. Extensive experiments conducted with twelve large detection transformers on COCO demonstrate the efficacy of AFOG. Our empirical results also show that AFOG outperforms existing attacks on transformer-based and CNN-based object detectors by up to \\(83\\%\\) with superior speed and imperceptibility. Code is available at: Link." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.785, + 0.21, + 0.801 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Transformer-based neural architectures and algorithms have blossomed in recent years, enhancing computer vision tasks including object detection. Attention is the core of the transformer architecture [30]. Attention allows a detector to focus on specific regions of images based on objects of interest, effectively predicting the presence and lo" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.555, + 0.895, + 0.652 + ], + "angle": 0, + "content": "Figure 1. Successful AFOG attacks against three object detection transformers. White pixels in the attention map indicate learned areas of high importance, often corresponding to objects in the input image. This example illustrates that AFOG's adversarial attention learns which regions of an image are most susceptible to adversarial perturbations. Even for the same input, learnable adversarial attention maps may differ under different models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "cation of each potential object. Powered by various attention mechanisms, detection transformers can handle overlapping objects more effectively than some older methods because they can capture long-range dependencies in an image, allowing models to reason about the relationships between objects. Since modern object detection transformers such as Swin [20] and DETR [4] significantly outperform traditional convolutional neural network (CNN)-based models [8], represented by Faster R-CNN [28], SSD [19] and YOLO-v3 [27], there is a pressing need to investigate and understand the vulnerabilities of large detection transformers in the presence of adversarial perturbations. Adversarial perturbations are helpful instrumentation for exposing vulnerabilities of large detection transformers, making them useful mechanisms for motivating more robust transformer models for object detection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3184" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.333 + ], + "angle": 0, + "content": "Existing attack methods struggle to disrupt transformer-based object detectors. Surrogate-based attacks (also known as black-box attacks), such as UEA [33] and RAD [5] generate perturbations on surrogate models and then test the adversarial effect against victim detectors that have similar detection architecture at inference. However, surrogate-based attacks suffer poor attack performance when the victim architectures are not similar to the surrogate models used to generate adversarial perturbations. Victim-based (also known as white-box) attacks, such as EBAD [3] and OATB [12], generate adversarial attacks by performing direct inference against victim models. Recent studies, such as AttentionFool [22], specialize in transformer-based victim models, but these transformer-only attacks are not applicable to convolutional detectors such as YOLO [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.834 + ], + "angle": 0, + "content": "In this paper, we present an Attention-Focused Offensive Gradient (AFOG) attack, which is a victim architecture agnostic and effective in attacking both advanced object detection transformers and traditional CNN-based detectors. AFOG by design offers three novel characteristics. First, inspired by a transformer's self-attention, we utilize a learnable attention mechanism to enable AFOG to adaptively focus its adversarial perturbations on vulnerable image areas (those areas where perturbations have the greatest effect on the output) in multi-box detection tasks. Second, we formulate the attack loss function of AFOG by integrating learnable feature-map-based attention updates with iterative injection of adversarial perturbations. Finally, we design AFOG to be an efficient and yet stealthy adversarial perturbation method. By efficient, we want AFOG to rapidly generate adversarial perturbations in minimal iterations. By stealthy, we want AFOG to generate small amounts of perturbations that are visually imperceptible and yet can cause well-trained object detectors to produce erroneous detection results. Figure 1 shows an example of successful AFOG attacks against three detection transformers. Row 1 shows benign detection from three state-of-the-art detectors, Row 2 and Row 3 show AFOG adversarial perturbations on the three detectors and their respective AFOG adversarial attention maps, and Row 4 shows the detection results under our AFOG attack. We validate AFOG with extensive experiments on twelve state-of-the-art object detection transformers with the COCO benchmark [18] and three popular families of traditional CNN-based object detectors. The results show that AFOG is consistently effective across all twelve detection transformers compared to existing methods, achieving high attack success rate and reducing benign mAP (mean Average Precision) by up to \\(37.79 \\times\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.22, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Existing adversarial perturbations against object detectors can be broadly categorized into surrogate-model based" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.651 + ], + "angle": 0, + "content": "(also known as black-box) and victim-model based (white-box) methods. Surrogate-based approaches in literature tend to rely on the assumption that the surrogate model is similar to the victim model. RAD [6], GHFD [32], and UEA [33] generate adversarial examples on a surrogate FR-CNN model, then attack other CNN-based detectors such as YOLO and SSD. However, these attacks show poor performance against object detection transformers [32]. For victim-model based methods, GARSDC [17] is more effective but can require more than 3000 iterations to converge. GALD [13] first attacks vision transformer classifiers and then transfers the classification-based adversarial perturbations to object detectors of similar transformer architecture, and hence is less effective by comparison. We classify EBAD [3] as a victim-based attack because it uses an ensemble of surrogate models to attack a victim, but requires access to the victim's loss function for its nested ensemble optimization. EBAD shows poor performance when transferring perturbations from a DETR surrogate to a DETR victim [24]. AttentionFool [22] is a recent victim-based attack targeting dot-product self-attention against DETR [14] with a ResNet-101 backbone. However, AttentionFool shows inconsistent performance against DETR with a ResNet-50 backbone [22]. AttentionFool is also not applicable to convolutional models such as YOLO [27] because it exclusively targets self-attention mechanisms. TOG [7] targets a prediction's objectness score and introduces vanishing and fabrication attack modes. OATB [12] uses a \"division map\" that statically emphasizes image regions during perturbation based on object location priors. DBA [16] prioritizes perturbations against image backgrounds to enhance imperceptibility, though it shows almost no effect against a Swin transformer for object detection [20]. In comparison, AFOG adversarial perturbations show strong performance against a wide variety of transformer- and CNN-based detector victims while maintaining superior imperceptibility and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.669, + 0.637, + 0.687 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.684, + 0.71 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a victim detector \\( f_{D}(\\vartheta, x) \\), where \\( x \\in \\mathcal{D} \\) is a victim image \\( x \\) and \\( \\mathcal{D} \\) is the test set, let \\( x \\) contain \\( N_{x} \\) objects to be detected, denoted by \\( \\mathcal{O}_x = \\{O_1, O_2, \\ldots, O_{N_x}\\} \\). Each object \\( O_i \\) (\\( 1 \\leq i \\leq N_x \\)) is a recognition target for the detector \\( f_{D}(\\vartheta, x) \\). Under benign scenarios, let \\( (B_i, C_i) \\) denote a ground truth object \\( O_i \\) with bounding box \\( B_i \\) and class label \\( C_i \\). Let \\( K \\) denote the total number of ground truth classes, and \\( C_i \\in \\{1, 2, \\ldots, K\\} \\), e.g., \\( K = 21 \\) for the VOC dataset [9], including the background class. Given input image \\( x \\), \\( f_{D}(\\vartheta, x) \\) outputs the benign prediction of \\( N \\) detected objects, denoted by \\( \\mathcal{R}[f_{D}(\\vartheta, x)] = \\{(b_i, c_i) | i = 1, \\ldots, N_x\\} \\). Each detected object \\( o_i \\) is associated with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3185" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.182 + ], + "angle": 0, + "content": "its predicted bounding box \\( b_{i} \\) and predicted class label \\( c_{i} \\). \\( (b_{i},c_{i}) \\) is evaluated as a correct prediction if the intersection over union (IoU) of \\( B_{i} \\) and \\( b_{i} \\) is greater than the detection threshold \\( \\gamma \\) (usually set to 0.5), and \\( C_i = c_i \\) \\( (1\\leq i\\leq N_x, \\) \\( x\\in \\mathcal{D}) \\). The overall detection accuracy is measured using mAP (mean Average Precision) on the entire test dataset \\( \\mathcal{D} \\)" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.183, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Let \\( x_{adv} \\) denote the adversarial example generated by injecting a sequence of adversarial perturbations to \\( x \\) through an iterative attention-based learning mechanism. The goal of our AFOG attack on detector \\( f_{D}(\\vartheta, x) \\) is to find \\( x_{adv} \\) that maximizes the success rate of falsifying the predictions of all object recognitions for all images in \\( \\mathcal{D} \\), i.e." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.274, + 0.41, + 0.29 + ], + "angle": 0, + "content": "\\(\\arg \\max_{x\\in \\mathcal{D},i\\in N_x},\\{(b_i,c_i)\\in \\mathcal{R}[f_D(\\vartheta ,x_{adv})]\\)" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.29, + 0.453, + 0.307 + ], + "angle": 0, + "content": "\\((IOU(B_i, b_i) < \\gamma \\lor C_i \\neq c_i), \\min ||x - x_{adv}||_p\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.469, + 0.412 + ], + "angle": 0, + "content": "The formula indicates that for each detected object \\(o_i\\), denoted by \\((b_i, c_i)\\) in victim image \\(x\\), the AFOG attack is successful on \\(o_i\\) if the IoU of \\(B_i\\) and \\(b_i\\) is less than the detection threshold \\(\\gamma\\) (usually set to 0.5) or \\(c_i \\in \\{1, 2, \\dots, K\\} (C_i \\neq c_i)\\), and \\(x_{adv}\\) also satisfies the distortion constraint of \\(min||x - x_{adv}||_p\\), where \\(p\\) is typically defined by \\(L_2\\) norm, \\(L_0\\) norm, or \\(L_\\infty\\) norm." + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.432, + 0.203, + 0.441 + ], + "angle": 0, + "content": "Benign Model Training:" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.441, + 0.269, + 0.451 + ], + "angle": 0, + "content": "Freeze (※) Input, Update (♀) Parameters" + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.436, + 0.389, + 0.445 + ], + "angle": 0, + "content": "\\(\\partial \\mathcal{L}(\\pmb {x},\\mathcal{O};\\pmb {\\theta})\\)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.445, + 0.432, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.256, + 0.549, + 0.432, + 0.566 + ], + "angle": 0, + "content": "Pk+1←Pk+αΓ[∂LAFOG(xad,OE;θ)]" + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.566, + 0.235, + 0.586 + ], + "angle": 0, + "content": "\\(x_{adv_k}\\gets \\prod_{x\\in}\\left(x + A_k\\odot P_k\\right)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.588, + 0.16, + 0.596 + ], + "angle": 0, + "content": "AFOG Attack:" + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.597, + 0.357, + 0.606 + ], + "angle": 0, + "content": "Freeze (k) Parameters, Update (C) Attention and Perturbation," + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.607, + 0.23, + 0.616 + ], + "angle": 0, + "content": "Compute (O) Adversarial Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.625, + 0.469, + 0.696 + ], + "angle": 0, + "content": "Figure 2. AFOG attack framework. During attack iteration the adversarial perturbation is updated via bounding-box and class loss gradient propagation. An adversarial attention map concurrently learns to apply per-pixel amplification or dampening to the perturbation, focusing its effects on vulnerable image regions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.71, + 0.35, + 0.726 + ], + "angle": 0, + "content": "3.2. Adversarial Loss Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.84 + ], + "angle": 0, + "content": "AFOG attacks the victim model via an iterative projected gradient descent method [23]. Figure 2 provides an illustration of the AFOG attack framework. At the initialization step, the attack first propagates the unaltered victim image \\( x \\) through the feed-forward network (FFN) of the victim model \\( f_{D}(\\vartheta ,x) \\) to obtain its benign prediction \\( \\mathcal{O}_x = \\{(b_i,c_i)|i = 1,\\dots ,N_x\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.469, + 0.902 + ], + "angle": 0, + "content": "In a scenario where no ground truth is available, the attack assumes these benign predictions to be correct labels for the input image \\( x \\). For the remaining section, we will use \\( (B_i, C_i) \\) in the context of performing an attack on \\( f_D(\\vartheta, x) \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "to obtain \\( x_{adv} \\). Upon initialization for each input image \\( x \\in \\mathcal{D} \\), the AFOG attack corrupts the victim image \\( x \\) by adding perturbations generated via an element-wise multiplication of two components using Equation 1:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.156, + 0.892, + 0.188 + ], + "angle": 0, + "content": "\\[\nx _ {a d v _ {k}} \\leftarrow \\prod_ {x, \\epsilon} \\left(x + A _ {k} \\odot P _ {k}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.191, + 0.892, + 0.267 + ], + "angle": 0, + "content": "\\(A\\) is the attention map, and \\(P\\) is the perturbation map. Here \\(\\odot\\) represents the Hadamard matrix product and \\(\\Pi\\) is a projection onto a hypersphere centered on unaltered victim image \\(x\\) with the radius \\(\\epsilon\\) as the maximum perturbation budget. \\(A\\) and \\(P\\) are initialized according to Equation 2." + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.271, + 0.892, + 0.288 + ], + "angle": 0, + "content": "\\[\nA _ {0} \\sim 1, P _ {0} \\sim \\text {R a n d o m} (- \\epsilon , \\epsilon) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Random \\((-\\epsilon, \\epsilon)\\) is a uniform random distribution. The perturbed image \\(x_{adv}\\) propagates through the FFN of the victim model. The attack loss function \\(\\mathcal{L}_{AFOG}\\) assesses the adversarial output \\(\\mathcal{O}_{x_{adv}}\\) and benign output \\(\\mathcal{O}_x\\), computing a loss that reflects the attack's progress in corrupting image \\(x\\). This attack loss is formulated in Equations 3, 4, and 5." + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.393, + 0.892, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {A F O G} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\tag {3} \\\\ + \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, o _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, o _ {a d v _ {i}}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.534, + 0.892, + 0.575 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, c _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, c _ {a d v _ {i}}\\right) \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.751 + ], + "angle": 0, + "content": "Given a victim image \\( x \\), the optimized attack loss \\( \\mathcal{L}_{AFOG} \\) is attained by making the model incorrectly predict every target object. We achieve this loss optimization by both falsifying every target object's bounding box prediction and its class label prediction. This can be represented by optimizing an adversarial bounding box loss \\( \\mathcal{L}_{bbox} \\) and an adversarial class label prediction loss \\( \\mathcal{L}_{cls} \\) for all \\( N_x \\) target objects. Conceptually, this suppresses the confidence of the original correct bounding box and class label prediction, while increasing the confidence of the adversarial prediction of an incorrect bounding box or class label." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Recall Figure 2: the AFOG attack freezes the model parameters \\(\\vartheta\\) and uses the gradient of the AFOG attack loss to update the attention map \\(A\\) and the perturbation \\(P\\) by backpropagation, generating the next iteration of the adversarial perturbation. A newly corrupted image \\(x_{adv}\\) is created upon injecting the updated adversarial perturbation by following Equation 1, where A and P are updated via Equations 6 and 7:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.873, + 0.892, + 0.906 + ], + "angle": 0, + "content": "\\[\nA _ {k + 1} \\leftarrow A _ {k} + \\alpha_ {A} \\sigma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial A _ {k}} \\right] \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.223, + 0.096, + 0.75, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.274, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Example attention map generation on DETR-R101 for \\( k = 10 \\) iterations. AFOG attention initially learns important regions of the image, then iteratively updates to achieve an effective attack. It finds vulnerable pixels in foreground objects as well as background regions, including in unexpected or unintuitive areas such as the sky above the boat." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.354, + 0.47, + 0.387 + ], + "angle": 0, + "content": "\\[\nP _ {k + 1} \\leftarrow P _ {k} + \\alpha_ {P} \\Gamma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial P _ {k}} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.404, + 0.473, + 0.51 + ], + "angle": 0, + "content": "\\(\\alpha_{A}\\) and \\(\\alpha_{P}\\) denote the attention map learning rate and the perturbation learning rate respectively, \\(\\sigma\\) is a normalization function, \\(\\Gamma\\) is a sign function, \\(\\mathcal{L}_{AFOG}\\) is the attack loss, \\(\\vartheta\\) is the model parameters, and \\(\\mathcal{O}_x\\) is the model's benign predictions from input \\(x\\). This attack process repeats until the number of attack iterations is reached. Algorithm 1 provides the pseudocode." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.531, + 0.376, + 0.546 + ], + "angle": 0, + "content": "3.3. Adversarial Attention Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.473, + 0.903 + ], + "angle": 0, + "content": "A key innovation of our AFOG method is empowering perturbation generation with a learnable attention mechanism. Motivated by the intuition that certain parts of a victim image are more susceptible to adversarial perturbation than others, we add an attention map to focus the perturbation on vulnerable pixels. The AFOG attack concurrently learns the adversarial attention map alongside the perturbation to iteratively maximize the attack loss \\(\\mathcal{L}_{\\text{AFOG}}\\) (recall Equation 3). Figure 3 gives an illustrative example of the iterative learning of AFOG's adversarial attention map and the corresponding perturbation. Unlike focusing mechanisms in other methods such as [33], AFOG's attention map dynamically updates during attack iteration, leaving it unconstrained by assumptions about foreground [12], background [16], or region proposal [33] importance from static methods. We posit that pixel importance may not correspond to human intuition, and design AFOG to iteratively learn pixel importance instead. We observe that in early iterations the attention mechanism tends to focus on the primary objects in an image, and then branches out to affect surrounding areas as the attack progresses. This adaptability is a key feature of learnable attention, showcasing an advantage over static attention maps." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.341, + 0.808, + 0.357 + ], + "angle": 0, + "content": "Algorithm 1 AFOG attack on an input image." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.358, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Require: Victim image \\( x \\in \\mathcal{D} \\), test-set \\( \\mathcal{D} \\), Victim pretrained model \\( f_{D}(\\vartheta) \\), Perturbation step size \\( \\alpha_{P} \\), Attention step size \\( \\alpha_{A} \\), Number of iterations \\( T \\), Maximum perturbation \\( \\epsilon \\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.419, + 0.7, + 0.433 + ], + "angle": 0, + "content": "1: Initialize \\(\\mathcal{O}_x\\gets f_D(x;\\vartheta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.434, + 0.746, + 0.448 + ], + "angle": 0, + "content": "2: Initialize attention map \\( A_0 \\gets 1 \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.449, + 0.834, + 0.463 + ], + "angle": 0, + "content": "3: Initialize perturbation \\(P_0 \\gets \\text{Random}(-\\epsilon, \\epsilon)\\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.464, + 0.729, + 0.478 + ], + "angle": 0, + "content": "4: Initialize step variable \\( k \\gets 1 \\)" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.419, + 0.834, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.479, + 0.639, + 0.492 + ], + "angle": 0, + "content": "5: while \\( k < T \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.493, + 0.834, + 0.517 + ], + "angle": 0, + "content": "6: Attack image \\(x_{advk} \\gets \\prod_{x,\\epsilon}(x + A_k \\odot P_k)\\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.519, + 0.858, + 0.533 + ], + "angle": 0, + "content": "7: Forward propagate \\( x_{adv} \\) through \\( f_{D}(\\vartheta, x_{adv}) \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.534, + 0.821, + 0.548 + ], + "angle": 0, + "content": "8: Compute bbox-loss \\(\\mathcal{L}_{bbox}(x_{adv},\\mathcal{O}_x;\\vartheta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.549, + 0.797, + 0.563 + ], + "angle": 0, + "content": "9: Compute \\(\\operatorname{cls}\\)-loss \\(\\mathcal{L}_{cls}(x_{adv},\\mathcal{O}_x;\\vartheta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.564, + 0.854, + 0.578 + ], + "angle": 0, + "content": "10: \\(\\mathcal{L}_{AFOG}(x_{adv},\\mathcal{O}_x;\\vartheta) = \\mathtt{bbox - loss} + \\mathtt{cls - loss};\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.579, + 0.891, + 0.609 + ], + "angle": 0, + "content": "11: Calculate losses with respect to \\(A_{k}\\) and \\(P_{k}\\): \\(\\mathcal{L}_A(x_{adv},\\mathcal{O}_x;\\vartheta),\\mathcal{L}_P(x_{adv},\\mathcal{O}_x;\\vartheta)\\);" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.61, + 0.844, + 0.624 + ], + "angle": 0, + "content": "12: Normalize attention loss \\(\\mathcal{L}_A\\gets \\mathrm{Norm}(\\mathcal{L}_A)\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.625, + 0.871, + 0.639 + ], + "angle": 0, + "content": "13: Take sign of perturbation loss \\(\\mathcal{L}_P\\gets \\mathrm{Sign}(\\mathcal{L}_P)\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.64, + 0.706, + 0.654 + ], + "angle": 0, + "content": "14: \\(A_{k + 1}\\gets A_k - \\alpha_A\\mathcal{L}_A;\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.655, + 0.704, + 0.669 + ], + "angle": 0, + "content": "15: \\(P_{k + 1}\\gets P_k - \\alpha_P\\mathcal{L}_P;\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.67, + 0.634, + 0.683 + ], + "angle": 0, + "content": "16: \\(k\\gets k + 1;\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.685, + 0.6, + 0.697 + ], + "angle": 0, + "content": "17: end while" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.699, + 0.768, + 0.723 + ], + "angle": 0, + "content": "18: \\(x_{advk + 1}\\gets \\prod_{x,\\epsilon}(x + A_{k + 1}\\odot P_{k + 1})\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.724, + 0.612, + 0.737 + ], + "angle": 0, + "content": "19: return \\(x_{adv}\\)" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.419, + 0.891, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.801, + 0.789 + ], + "angle": 0, + "content": "3.4. Special Cases of the AFOG Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We explore two special cases of our AFOG attack, each of which targets a specific malicious detection behavior in the victim detector. The first special case is an object vanishing attack, coined AFOG-V, which attempts to attack the objectness detection task of multi-box object detection. The goal of this special case AFOG attack is to make the victim model unable to detect any object, making all object" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3187" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "detections vanish for victim image \\( x \\). We implement the AFOG-V attack by changing the initialization: we use an empty set instead of forward propagation of \\( x \\) to obtain the benign detection results for the \\( N_{x} \\) objects as the assumed ground-truth in AFOG-V attack. Let \\( \\varnothing \\) denote an altered version of \\( \\mathcal{O}_x \\) that contains no predictions. The formulation of the AFOG-V attack loss function is given by Equation 8:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.209, + 0.47, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {A F O G _ {V}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\varnothing ; \\vartheta\\right) \\tag {8} \\\\ - \\mathcal {L} _ {c l s} (x _ {a d v}, \\varnothing ; \\vartheta) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.47, + 0.418 + ], + "angle": 0, + "content": "The second special case is object fabrication, coined as AFOG-F, which attempts to attack the bounding-box detection task by generating perturbations that lead to spurious detections (i.e., false positives). Similarly, we revise the AFOG attack process: instead of keeping the benign predictions with the confidence score above a certain tunable threshold (0.5 by default), for AFOG-F, we remove this threshold and allow the benign detection set \\(\\mathcal{O}_x\\) to include a much larger set of \"ground truth\" targets, which inevitably include erroneous detections. Accordingly the AFOG-F loss function is given in Equation 9:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.419, + 0.47, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {A F O G _ {F}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\tag {9} \\\\ - \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.539 + ], + "angle": 0, + "content": "where \\(\\mathcal{O}_F\\) is a modified version such that the likelihood score of each prediction is set to 1.0. We explore these two specialized cases of our AFOG attack to gain a deeper understanding of the adverse effects of adversarial perturbations on different detection transformers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.551, + 0.21, + 0.568 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.576, + 0.267, + 0.592 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.81 + ], + "angle": 0, + "content": "We use the Common Objects in Context (COCO) 2017 [18] test-dev split to evaluate the performance of our attack on contemporary object detectors. COCO 2017 is a standard benchmark for evaluating object detector performance, and its test-dev set has 80 object categories across 5,000 images. We use the PASCAL Visual Objects Challenge (VOC) 2007 [9] dataset to compare our attack with the state-of-the-art attacks on CNN-based object detectors, because existing attacks designed only for CNN-based detectors all report their evaluation with VOC. The VOC devkit split contains 4,952 images with 20 object classes. All experiments were performed on an NVIDIA A100. We list our models and their benign performance in Table 1. We provide model details in Section ?? of the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We select a variety of model sizes, ranging from Detection Transformer (DETR) [4], a lightweight 40 million parameter model, to EVA [10], a vision-focused foundation model with over one billion parameters. We use the Detrex framework [29], built on top of Detectron2 [34], to standardize our model implementations via DINO [37]. DINO" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.469 + ], + "angle": 0, + "content": "and Detrex adapt many general purpose vision models to object detection. The Detrex framework standardizes all images in [0, 1]. We also attack the original versions of several transformer models to demonstrate that the AFOG attack also works beyond the Detrex framework. All 12 transformer models and FRCNN use PyTorch [26] implementations, whereas SSD300 [19] and YOLOv3 [27] use Tensorflow [1]. We use mean average precision (mAP) to evaluate both benign and attacked performance. The Average Precision (AP) is given by interpolating the product of precision and recall at several decision thresholds. The mean AP (mAP) is the average of APs over the number of object classes. A lower mAP indicates that the detector is less effective for the object detection task. The victim detector's mAP under adversarial perturbations reflects the degree of degradation in detection performance for the victim detector when compared with its benign mAP. We measure imperceptibility by reporting the average distortion of the attack via four metrics: \\( L_{2} \\) norm, \\( L_{0} \\) norm, semantic structural similarity index (SSIM), and mean distortion \\( \\mu_{\\Delta} \\). We compute each metric per image and report the average over all images in COCO 2017 test-dev. Our experiments in the next section show that setting the total number of attack iterations to 10 is effective for all 12 state-of-the-art detection models on the COCO benchmark." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.482, + 0.868, + 0.499 + ], + "angle": 0, + "content": "4.2. Effectiveness Comparison of AFOG Attack" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Table 1 shows the effectiveness of the AFOG attack and its two special extensions AFOG-V and AFOG-F on each of the twelve detection transformers with varying model sizes (Col. 2). We use their benign mAP scores (Col. 3) as the reference for the evaluation. We make three observations: (i) under AFOG attacks all 12 transformers suffer a drastic performance drop in mAP scores, ranging from \\(2.5 \\times\\) reduction to \\(37.8 \\times\\) reduction. (ii) Compared to the AFOG generic attack, the special extension AFOG-V offers stronger attack success rates in terms of mAP reduction for 11 out of 12 transformers, except DETR-R50 [4]; and the special extension AFOG-F offers stronger attack success rates for 7 out of 12 transformers, except DETR-R50 [4], DETR-R101 [4], Deformable-DETR [38], R50 (DINO) [11], Swin-L (DINO) [20] (iii) For AlignDETR [2], AFOG-F achieves a mAP of 1.36 and outperforms both AFOG (mAP of 18.13) and AFOG-V (mAP of 1.64), and all three attacks significantly reduce the benign mAP by a factor ranging from \\(2.8 \\times\\) to \\(37.8 \\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.902 + ], + "angle": 0, + "content": "We compare AFOG against eleven benchmark attacks on DETR and Swin in Table 2. We observe that AFOG achieves superior performance on both DETR-R50 and Swin-L, notably recording an \\(82.7\\%\\) improvement over the next strongest attack on Swin. AFOG also uses the smallest perturbation budget and fewest iterations, indicating its superior imperceptibility and speed." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.223, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.089, + 0.368, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.09, + 0.495, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.627, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.091, + 0.745, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.09, + 0.894, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.24, + 0.893, + 0.269 + ], + "angle": 0, + "content": "Figure 4. Visualization with six examples, comparing benign scenario and AFOG attack results on Detection Transformer (DETR) with ResNet-50 backbone. Six diverse test images contain objects with varying classes, scales, perspectives, lighting, motion, and textures." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.279, + 0.471, + 0.32 + ], + "angle": 0, + "content": "Table 1. AFOG effectiveness over 12 detection transformers, measured by mAP on perturbed images. (*) indicates DINO framework used with the corresponding backbone for object detection." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.32, + 0.477, + 0.451 + ], + "angle": 0, + "content": "
ModelParams (M)BenignAFOGAFOG-VAFOG-F
DETR-R50 [4]39.842.14.14.59.8
DETR-R101 [4]76.043.55.25.111.3
Deform.-DETR [38]40.044.54.81.57.1
R50*[11]47.649.25.31.56.3
AlignDETR [2]47.651.418.11.61.4
ViTDet* [14]108.154.93.80.92.8
ConvNeXt* [21]219.055.43.91.93.1
Swin-L* [20]217.256.87.32.48.6
InternImage* [31]241.056.97.32.85.1
FocalNet* [36]228.958.57.32.55.1
EVA* [10]1037.262.112.24.18.7
DETA [25]218.862.925.63.74.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.466, + 0.471, + 0.522 + ], + "angle": 0, + "content": "Table 2. Comparison benchmark of AFOG against other state-of-the-art object detection attacks on DETR and Swin. Results are theirs. (*) indicates results from [32]. (†) indicates results from [24]. (-) indicates no result." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.522, + 0.48, + 0.687 + ], + "angle": 0, + "content": "
AttackTypePert. BudgetIters.Adversarial mAP
DETR-R50Swin
GARSDC [17]Surrogate0.053000+6.0-
GALD [13]Surrogate0.0631020.6-
RAD* [6]Surrogate0.0631027.247.2
GHFD* [32]Surrogate0.0635012.742.3
UEA* [33]Surrogate0.0635028.550.7
DAG* [35]Surrogate0.0635028.650.7
RAP* [15]Surrogate0.0635024.749.5
EBAD†[3]Victim0.0391034.9-
AttentionFool [22]Victim-10-15021.0-
OATB [12]Victim0.0782026.6-
DBA [16]Victim-50-56.7
AFOGVictim0.031104.17.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Figure 4 shows six visual examples and the comparison of DETR results for benign detection (Row 1) and adversarial detection under the AFOG attack (Row 2) scenarios. The first example shows that DETR detects the person and surfboard correctly under no attack (Row 1) but outputs erroneous detections under the AFOG attack (Row 2). Additional visual examples for all twelve transformers are available in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We report the distortion effect and timing cost for each of the 12 transformer models in Table 3. We note several interesting observations: (i) Considering the \\( L_{2} \\) and SSIM metrics, AFOG attacks induce very similar levels of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.283, + 0.895, + 0.556 + ], + "angle": 0, + "content": "distortion for 10 out of 12 transformer models. The two largest detection transformers, DETA [25] and EVA [10], show worse \\( L_{2} \\) and SSIM scores for AFOG, AFOG-V and AFOG-F. (ii) Considering the \\( L_{0} \\) metric, only DETA shows the lowest \\( L_{0} \\) value consistently for AFOG, AFOG-V and AFOG-F. (iii) As expected, the time required to attack a detection transformer model is strongly correlated to the number of model parameters. For instance, EVA [10] has over 1 billion parameters, and took much longer on average to attack an input image over 10 iterations compared to the other 11 models. (iii) DETA [25] has 218.8 million parameters and is similar to other models within the Detrex framework: FocalNet [36], InternImage [31], Swin-L [20] and ConvNext [21]. However, we observe that it consistently takes about \\( 1.5 - 2 \\times \\) longer time to attack one input image on average with our AFOG attack or its special case extensions AFOG-V or AFOG-F. (iv) AFOG, AFOG-V and AFOG-F have similar average perturbation magnitude." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.573, + 0.833, + 0.589 + ], + "angle": 0, + "content": "4.3. Effectiveness on CNN-Based Detectors" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.893, + 0.901 + ], + "angle": 0, + "content": "This section compares AFOG with four attacks designed for CNN-based models: TOG [7], UEA [33], RAP [15], and DAG [35]. They are surrogate-based attacks, and a majority of them (e.g., DAG, RAP, UEA) can only directly attack two-stage detectors (exemplified by Faster-R-CNN [28]) and rely on adversarial transferability to attack single-stage CNN-based detectors like YOLOv3 [27] and SSD [19]. Table 4 reports the comparison results. We make two observations: (i) AFOG and TOG are the only two victim-based attacks against single-stage detectors, represented by YOLOv3 and SSD. Both AFOG and TOG can drastically reduce the benign mAP of YOLOv3 \\((83.43\\%)\\) and SSD \\((76.11\\%)\\). For YOLOv3, TOG is a stronger attack, achieving mAP of 0.56 compared to mAP of AFOG being 2.62. However, for SSD-300, AFOG is a stronger attack, achieving mAP of 0.50, compared to mAP of TOG (0.86). (ii) For two-stage CNN-based object detectors, exemplified by Faster-R-CNN, AFOG outperforms all four other victim-based attacks, achieving the lowest adversarial mAP score of \\(2.07\\%\\). (iii) AFOG also uses the same distortion mag" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3189" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.088, + 0.892, + 0.129 + ], + "angle": 0, + "content": "Table 3. Timing and Imperceptibility Results. \\( L_{2} \\) represents the average \\( L_{2} \\) norm difference between perturbed and clean images, \\( L_{0} \\) is the average proportion of perturbed pixels, SSIM is the structural similarity index measure, \\( \\mu_{\\Delta} \\) is the average perturbation magnitude, and t is average total attack time for all ten iterations in seconds." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.13, + 0.917, + 0.273 + ], + "angle": 0, + "content": "
ModelL2AFOGtimeL2AFOG-VtimeL2AFOG-Ftime
L0SSIMμΔL0SSIMμΔL0SSIMμΔ
DETR-R50 [4]0.03220.97070.87150.01731.450.03230.97070.87160.01720.990.03230.97100.87170.01721.21
DETR-R101 [4]0.03230.97070.87210.01731.470.03230.97060.87240.0131.160.03230.97080.87240.01721.70
Deform.-DETR [38]0.03230.97190.87110.01741.630.03230.97130.87160.01731.630.00120.97140.87170.01731.87
R50 [11]0.03170.96580.83430.01712.700.03170.96500.83480.01702.340.03170.96540.83460.01703.16
AlignDETR [2]0.03190.96570.83470.01702.410.03190.96460.83490.01702.330.03190.96470.83490.01703.29
ViTDet [14]0.03180.96710.83530.01716.880.03180.96570.83610.01706.670.03180.96640.83550.01717.33
ConvNext [21]0.03180.96660.83420.01715.380.03180.96540.83490.01705.260.03180.96630.83470.01705.86
Swin-L [20]0.03270.97240.86730.01757.130.03270.97160.86800.01737.280.03270.97220.86780.01748.99
InternImage [31]0.03180.96650.83600.01706.350.03180.96530.83670.01706.230.03180.96600.83640.01716.70
FocalNet [36]0.03200.96650.83650.01728.960.03200.96570.83780.01718.840.03200.96590.83740.01719.74
EVA [10]0.03700.96660.82400.017254.340.03700.96650.82460.017154.530.03700.96650.82420.017151.18
DETA [25]0.04810.96620.80960.017013.200.04810.96540.80990.017013.100.04810.96540.80990.017013.13
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.375 + ], + "angle": 0, + "content": "nitude budget \\( L_{\\infty} \\) as TOG, while achieving higher attack success rate (lower mAP) on SSD-300 and FRCNN with decreased \\( L_{2} \\) distortion cost. Hence, AFOG excels at attacking both transformer-based detectors and CNN-based detectors." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Table 4. Comparing AFOG with four state-of-the-art attacks on representative CNN-based object detectors. mAPs of existing attacks were taken from respective papers [7]. (-) indicates N/A." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.424, + 0.46, + 0.596 + ], + "angle": 0, + "content": "
ModelAttackmAPtDistortion Cost
\\(L_{\\infty}\\)\\(L_{2}\\)\\(L_{0}\\)SSIM
YOLOv3Benign83.430.00.00.00.01.0
TOG [7]0.560.980.0310.0830.9840.875
AFOG2.281.310.0310.0130.8550.801
SSD-300Benign76.110.00.00.00.01.0
UEA [33]20.0-----
DAG [35]64.0-----
TOG [7]0.860.390.0310.1200.9750.879
AFOG0.500.490.0310.0220.8580.793
FRCNNBenign67.370.00.00.00.01.0
UEA [33]5.00.170.3430.1910.9590.652
RAP [15]4.784.040.0820.0100.5310.994
DAG [35]3.567.990.0240.0020.4930.999
TOG [7]2.641.680.0310.0580.9760.862
AFOG2.382.110.0310.0190.8540.788
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.377, + 0.637 + ], + "angle": 0, + "content": "4.4. Effect of Learnable Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.795 + ], + "angle": 0, + "content": "We isolate and analyze the role of the learnable self-attention mechanism in empowering AFOG to successfully corrupt transformer-based models by probing for weak spots and performing fast and human-imperceptible perturbations. Figure 5 shows a comparison between AFOG with and without its learnable attention mechanism, measured by percent difference in adversarial mAP scores. Ablation results show that our learnable attention mechanism improves performance by up to \\(30.6\\%\\) on InternImage, with an average of \\(15.1\\%\\) improvement across all models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Figure 6 shows the self-attention map for the last encoder layer of DETR [4] at initialization and three progressive attack iterations \\((3^{rd}, 4^{th}\\), and \\(8^{th}\\)). We observe that, as the attack progresses in iterations (indicated by grey arrows), the self-attention map becomes increasingly less associated with the benign detected objects. By the \\(8^{th}\\) iteration of the AFOG attack, the self-attention maps have become almost" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.435 + ], + "angle": 0, + "content": "entirely disjoint from the benign detections of the victim model (DETR) prior to the attack. We note that, when the model's self-attention map that reflects learned detection knowledge is changed, it is unlikely for prior knowledge to be kept intact. During the AFOG attack, the perturbed inputs focused by AFOG's own adversarial attention corrupt the pixel-level relationships learned by a well-trained victim model prior to adversarial injection (e.g., \\( k = 1 \\)), causing catastrophic forgetting to occur." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.45, + 0.885, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.634, + 0.895, + 0.691 + ], + "angle": 0, + "content": "Figure 5. Improvement (\\% difference) of AFOG with attention over AFOG without attention. The learnable attention mechanism improves performance by as much as \\(30.6\\%\\) on InternImage, with an average improvement of \\(15.1\\%\\) across all models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.699, + 0.695, + 0.716 + ], + "angle": 0, + "content": "4.5. Worst Case Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our experimental results on 12 transformer-based detectors and 3 representative CNN-based detectors have shown that AFOG is a fast and effective attack. However, there are still cases where AFOG fails in its adversarial perturbations within a total of 10 iterations, one setting of our attack termination hyper-parameter. By investigating those cases where AFOG did not succeed, we found that the most common situation is where the attack fails to corrupt the central subject of the victim input image, e.g., a person. Figure 7 gives three visual examples, with the top one (Rows 1-2) being a successful attack followed by two failed cases (Rows 3-6) when attacking DETR-R50. Columns 2-3 in all six" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3190" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.47, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.47, + 0.389 + ], + "angle": 0, + "content": "Figure 6. DETR self-attention weights analysis. Each blue frame shows the self-attention weights of DETR-R50's last encoder layer at the red indicated point in the image at a given attack iteration \\( k \\). In early iterations, the model understands a strong association between the indicated red point on the elephant and the rest of the elephant shape. As the attack iterates (grey arrows), this association becomes corrupted, evidenced by the lack of encoder self-attention structure in later stages. By the \\( 8^{th} \\) iteration, the selected point has lost association with the rest of the elephant object, demonstrating that AFOG has disrupted learned knowledge." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.47, + 0.5 + ], + "angle": 0, + "content": "rows show DETR's encoder's last layer's self-attention map for the first and second locations indicated by red points in the first column. Rows 1, 3, and 5 show benign situations and their corresponding self-attention maps, and Rows 2, 4, 6 show the self-attention maps after the final attack iteration. The fourth column shows AFOG's adversarial attention map after the final attack iteration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.471, + 0.789 + ], + "angle": 0, + "content": "We make three observations: (i) For the successful attack case (Rows 1-2), the self-attention maps of victim model DETR-R50 lose association between the indicated red point and the rest of the foreground object it comprises. (ii) In both unsuccessful cases (Rows 3-6), AFOG fails to disrupt this association. (iii) Consider AFOG's adversarial attention maps from the successful case in Figure 7 (Row 2, Col. 4), we observe a clear focus on the key objects in the victim input image. In comparison, AFOG's adversarial attention maps for the two failure cases (Row 4, Col. 4 and Row 6, Col. 4) both fail to focus on foreground objects. The first failure case (Rows 3 and 4) shows that the AFOG attention map attributes significant weight to the distant person in the left of the victim image, resulting in a large number of fabricated predictions in that region. Similarly, the attention map for the second victim image (Rows 5 and 6) appears to entirely miss the fire hydrant, resulting in a failure to disrupt this object. In these cases, AFOG's adversarial attention has learned to focus on the wrong pixels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.196, + 0.816 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.9 + ], + "angle": 0, + "content": "We have presented AFOG, an Attention-Focused Offensive Gradient attack. AFOG is effective for attacking both advanced object detection transformers and traditional CNN-based detectors with a unified, architecture-agnostic framework. AFOG utilizes a learnable attention mecha" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.086, + 0.892, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.596 + ], + "angle": 0, + "content": "Figure 7. Worst case analysis with DETR-R50 as victim model. Rows 1-2 show a benign image and the successful AFOG perturbed image. Rows 3-6 show two cases where AFOG attack did not succeed. Columns 2-3 show DETR's encoder's self-attention map for the two locations marked in red in column one. Column 4 shows AFOG's internal attention map, where white pixels indicate high attention and blue indicates low. In the successful case (Rows 1-2), encoder self-attention maps (Cols. 2-3) lose associative structure. In the unsuccessful cases (Rows 4 and 6), AFOG fails to disrupt this structure and predictions are preserved. AFOG adversarial attention maps (Col. 4) show focus on important object structure in the successful case (Row 2) and lack of focus in the unsuccessful cases (Rows. 4, 6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.841 + ], + "angle": 0, + "content": "nism to enable its adversarial perturbations to focus on vulnerable areas of images in multi-box detection tasks. AFOG's attack loss function integrates multiple feature losses (e.g., bounding-box loss, class loss) through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is efficient and stealthy. Adversarial perturbations generated by AFOG are visually imperceptible and yet can cause well-trained detectors to fail miserably. Extensive experiments on state-of-the-art object detectors show that AFOG is consistently effective across twelve object detection transformers. Comparative evaluation with nearly a dozen SOTA methods shows that AFOG significantly outperforms surrogate-based and victim-based attacks on both object detection transformers and CNN-based object detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This research is partially sponsored by the NSF CISE grants 2302720 and 2312758, CISCO Edge AI program, and GTRI PhD Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3191" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.295 + ], + "angle": 0, + "content": "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.297, + 0.47, + 0.338 + ], + "angle": 0, + "content": "[2] Zhi Cai, Songtao Liu, Guodong Wang, Zheng Ge, Xiangyu Zhang, and Di Huang. Align-detr: Improving detr with simple iou-aware bce loss, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.339, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[3]Zikui Cai, Yaoteng Tan, and M. Salman Asif. Ensemble-based blackbox attacks on dense prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4045-4055, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.47, + 0.463 + ], + "angle": 0, + "content": "[4] Nicolas Carrion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-223, 2020. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.465, + 0.469, + 0.506 + ], + "angle": 0, + "content": "[5] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.469, + 0.548 + ], + "angle": 0, + "content": "[6] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[7] Ka-Ho Chow, Ling Liu, Margaret Loper, Juhyun Bae, Mehmet Emre Gursoy, Stacey Truex, Wenqi Wei, and Yanzhao Wu. Adversarial objectness gradient attacks in real-time object detection systems. In IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications, pages 263-272. IEEE, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.635, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[8] Papers With Code. Object detection on coco test-dev. https://paperswithcode.com/sota/object-detection-on-coco.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[9] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[10] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022.5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Shu. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 06 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[12] Zhuo Leng, Zesen Cheng, Pengxu Wei, and Jie Chen. Object-aware transfer-based black-box adversarial attack on object detector. In Qingshan Liu, Hanzi Wang, Zhanyu" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Ma, Weishi Zheng, Hongbin Zha, Xilin Chen, Liang Wang, and Rongrong Ji, editors, Pattern Recognition and Computer Vision, pages 278-289, Singapore, 2024. Springer Nature Singapore. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[13] Tuo Li and Yahong Han. Improving transferable adversarial attack for vision transformers via global attention and local drop. Multimedia Systems, 29:3467 - 3480, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[14] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[15] Yuezun Li, Daniel Tian, Mingching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. In BMVC, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[16] Jiawei Lian, Shaohui Mei, Xiaofei Wang, Yi Wang, Lefan Wang, Yingjie Lu, Mingyang Ma, and Lap-Pui Chau. Attack anything: Blind dnns via universal background adversarial attack, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[17] Siyuan Liang, Longkang Li, Yanbo Fan, and Xiaojun Jia. A large-scale multiple-objective method for black-box attack against object detection. European Conference on Computer Vision (ECCV) 2022, pages 619-636, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, ECCV 2014, pages 740-755. Springer International Publishing, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[22] Giulio Lovisotto, Nicole Finnie, Maurizio Munoz, Chaithanya Kumar Murnmadi, and Jan Hendrik Metzen. Give me your attention: Dot-product attention considered harmful for adversarial patch robustness. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15213-15222, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[23] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[24] Khoi Nguyen Tiet Nguyen, Wenyu Zhang, Kangkang Lu, Yuhuan Wu, Xingjian Zheng, Hui Li Tan, and Liangli Zhen. A Survey and Evaluation of Adversarial Attacks for Object Detection, Aug. 2024. arXiv:2408.01934 [cs]. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbihl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3192" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.244 + ], + "angle": 0, + "content": "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.246, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[27] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[28] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.316, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[29] Tianhe Ren, Shilong Liu, Feng Li, Hao Zhang, Ailing Zeng, Jie Yang, Xingyu Liao, Ding Jia, Hongyang Li, He Cao, Jianan Wang, Zhaoyang Zeng, Xianbiao Qi, Yuhui Yuan, Jianwei Yang, and Lei Zhang. detrex: Benchmarking detection transformers, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.385, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.483, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.553, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[32] Yang Wang, Yunfei Zheng, Lei Chen, Zhen Yang, Jingwei Wu, and Tieyong Cao. Gradient-guided hierarchical feature attack for object detector. Journal of King Saud University - Computer and Information Sciences, 36(1):101901, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.622, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[33] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 954–960. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.706, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[34] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.748, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie1, and Alan Yuill. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision, pages 1378-1387, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.818, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[36] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[37] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5, 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3193" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_origin.pdf b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4ff1d75b66e2d25c53b73fd880b415f74ccd453f --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/88192fda-87cb-4027-ba07-95fbfb69b360_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dedaf8b6adf6a4c0383b987b0ae2e0fda130a0b6eda84a6893266873af3effa +size 4627642 diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/full.md b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93176ec0ad109df2bed780e760176b625e5e302f --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/full.md @@ -0,0 +1,288 @@ +# Adversarial Attention Perturbations for Large Object Detection Transformers + +Zachary Yahn1, Selim Furkan Tekin1, Fatih Ilhan1, Sihao Hu1, Tiansheng Huang1, Yichang Xu1, Margaret Loper2, Ling Liu1 +1Georgia Institute of Technology, Atlanta, GA +2Georgia Tech Research Institute, Atlanta, USA + +{zachary.yahn, stekin6, filhan, sihaohu, thuang, xuyichang}@gatech.edu, margaret.loper@gtri.gatech.edu, ling.liu@cc.gatech.edu + +# Abstract + +Adversarial perturbations are useful tools for exposing vulnerabilities in neural networks. Existing adversarial perturbation methods for object detection are either limited to attacking CNN-based detectors or weak against transformer-based detectors. This paper presents an Attention-Focused Offensive Gradient (AFOG) attack against object detection transformers. By design, AFOG is neural-architecture agnostic and effective for attacking both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. This paper makes three original contributions. First, AFOG utilizes a learnable attention mechanism that focuses perturbations on vulnerable image regions in multi-box detection tasks, increasing performance over non-attention baselines by up to $30.6\%$ . Second, AFOG's attack loss is formulated by integrating two types of feature loss through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is an efficient and stealthy adversarial perturbation method. It probes the weak spots of detection transformers by adding strategically generated and visually imperceptible perturbations which can cause well-trained object detection models to fail. Extensive experiments conducted with twelve large detection transformers on COCO demonstrate the efficacy of AFOG. Our empirical results also show that AFOG outperforms existing attacks on transformer-based and CNN-based object detectors by up to $83\%$ with superior speed and imperceptibility. Code is available at: Link. + +# 1. Introduction + +Transformer-based neural architectures and algorithms have blossomed in recent years, enhancing computer vision tasks including object detection. Attention is the core of the transformer architecture [30]. Attention allows a detector to focus on specific regions of images based on objects of interest, effectively predicting the presence and lo + +![](images/9be411cf3f959f4f01c2ca65169ed11fa1012491174efb1ba935f79ac70433be.jpg) +Figure 1. Successful AFOG attacks against three object detection transformers. White pixels in the attention map indicate learned areas of high importance, often corresponding to objects in the input image. This example illustrates that AFOG's adversarial attention learns which regions of an image are most susceptible to adversarial perturbations. Even for the same input, learnable adversarial attention maps may differ under different models. + +cation of each potential object. Powered by various attention mechanisms, detection transformers can handle overlapping objects more effectively than some older methods because they can capture long-range dependencies in an image, allowing models to reason about the relationships between objects. Since modern object detection transformers such as Swin [20] and DETR [4] significantly outperform traditional convolutional neural network (CNN)-based models [8], represented by Faster R-CNN [28], SSD [19] and YOLO-v3 [27], there is a pressing need to investigate and understand the vulnerabilities of large detection transformers in the presence of adversarial perturbations. Adversarial perturbations are helpful instrumentation for exposing vulnerabilities of large detection transformers, making them useful mechanisms for motivating more robust transformer models for object detection. + +Existing attack methods struggle to disrupt transformer-based object detectors. Surrogate-based attacks (also known as black-box attacks), such as UEA [33] and RAD [5] generate perturbations on surrogate models and then test the adversarial effect against victim detectors that have similar detection architecture at inference. However, surrogate-based attacks suffer poor attack performance when the victim architectures are not similar to the surrogate models used to generate adversarial perturbations. Victim-based (also known as white-box) attacks, such as EBAD [3] and OATB [12], generate adversarial attacks by performing direct inference against victim models. Recent studies, such as AttentionFool [22], specialize in transformer-based victim models, but these transformer-only attacks are not applicable to convolutional detectors such as YOLO [27]. + +In this paper, we present an Attention-Focused Offensive Gradient (AFOG) attack, which is a victim architecture agnostic and effective in attacking both advanced object detection transformers and traditional CNN-based detectors. AFOG by design offers three novel characteristics. First, inspired by a transformer's self-attention, we utilize a learnable attention mechanism to enable AFOG to adaptively focus its adversarial perturbations on vulnerable image areas (those areas where perturbations have the greatest effect on the output) in multi-box detection tasks. Second, we formulate the attack loss function of AFOG by integrating learnable feature-map-based attention updates with iterative injection of adversarial perturbations. Finally, we design AFOG to be an efficient and yet stealthy adversarial perturbation method. By efficient, we want AFOG to rapidly generate adversarial perturbations in minimal iterations. By stealthy, we want AFOG to generate small amounts of perturbations that are visually imperceptible and yet can cause well-trained object detectors to produce erroneous detection results. Figure 1 shows an example of successful AFOG attacks against three detection transformers. Row 1 shows benign detection from three state-of-the-art detectors, Row 2 and Row 3 show AFOG adversarial perturbations on the three detectors and their respective AFOG adversarial attention maps, and Row 4 shows the detection results under our AFOG attack. We validate AFOG with extensive experiments on twelve state-of-the-art object detection transformers with the COCO benchmark [18] and three popular families of traditional CNN-based object detectors. The results show that AFOG is consistently effective across all twelve detection transformers compared to existing methods, achieving high attack success rate and reducing benign mAP (mean Average Precision) by up to $37.79 \times$ + +# 2. Related Work + +Existing adversarial perturbations against object detectors can be broadly categorized into surrogate-model based + +(also known as black-box) and victim-model based (white-box) methods. Surrogate-based approaches in literature tend to rely on the assumption that the surrogate model is similar to the victim model. RAD [6], GHFD [32], and UEA [33] generate adversarial examples on a surrogate FR-CNN model, then attack other CNN-based detectors such as YOLO and SSD. However, these attacks show poor performance against object detection transformers [32]. For victim-model based methods, GARSDC [17] is more effective but can require more than 3000 iterations to converge. GALD [13] first attacks vision transformer classifiers and then transfers the classification-based adversarial perturbations to object detectors of similar transformer architecture, and hence is less effective by comparison. We classify EBAD [3] as a victim-based attack because it uses an ensemble of surrogate models to attack a victim, but requires access to the victim's loss function for its nested ensemble optimization. EBAD shows poor performance when transferring perturbations from a DETR surrogate to a DETR victim [24]. AttentionFool [22] is a recent victim-based attack targeting dot-product self-attention against DETR [14] with a ResNet-101 backbone. However, AttentionFool shows inconsistent performance against DETR with a ResNet-50 backbone [22]. AttentionFool is also not applicable to convolutional models such as YOLO [27] because it exclusively targets self-attention mechanisms. TOG [7] targets a prediction's objectness score and introduces vanishing and fabrication attack modes. OATB [12] uses a "division map" that statically emphasizes image regions during perturbation based on object location priors. DBA [16] prioritizes perturbations against image backgrounds to enhance imperceptibility, though it shows almost no effect against a Swin transformer for object detection [20]. In comparison, AFOG adversarial perturbations show strong performance against a wide variety of transformer- and CNN-based detector victims while maintaining superior imperceptibility and efficiency. + +# 3. Methodology + +# 3.1. Problem Definition + +Given a victim detector $f_{D}(\vartheta, x)$ , where $x \in \mathcal{D}$ is a victim image $x$ and $\mathcal{D}$ is the test set, let $x$ contain $N_{x}$ objects to be detected, denoted by $\mathcal{O}_x = \{O_1, O_2, \ldots, O_{N_x}\}$ . Each object $O_i$ ( $1 \leq i \leq N_x$ ) is a recognition target for the detector $f_{D}(\vartheta, x)$ . Under benign scenarios, let $(B_i, C_i)$ denote a ground truth object $O_i$ with bounding box $B_i$ and class label $C_i$ . Let $K$ denote the total number of ground truth classes, and $C_i \in \{1, 2, \ldots, K\}$ , e.g., $K = 21$ for the VOC dataset [9], including the background class. Given input image $x$ , $f_{D}(\vartheta, x)$ outputs the benign prediction of $N$ detected objects, denoted by $\mathcal{R}[f_{D}(\vartheta, x)] = \{(b_i, c_i) | i = 1, \ldots, N_x\}$ . Each detected object $o_i$ is associated with + +its predicted bounding box $b_{i}$ and predicted class label $c_{i}$ . $(b_{i},c_{i})$ is evaluated as a correct prediction if the intersection over union (IoU) of $B_{i}$ and $b_{i}$ is greater than the detection threshold $\gamma$ (usually set to 0.5), and $C_i = c_i$ $(1\leq i\leq N_x,$ $x\in \mathcal{D})$ . The overall detection accuracy is measured using mAP (mean Average Precision) on the entire test dataset $\mathcal{D}$ + +Let $x_{adv}$ denote the adversarial example generated by injecting a sequence of adversarial perturbations to $x$ through an iterative attention-based learning mechanism. The goal of our AFOG attack on detector $f_{D}(\vartheta, x)$ is to find $x_{adv}$ that maximizes the success rate of falsifying the predictions of all object recognitions for all images in $\mathcal{D}$ , i.e. + +$\arg \max_{x\in \mathcal{D},i\in N_x},\{(b_i,c_i)\in \mathcal{R}[f_D(\vartheta ,x_{adv})]$ + +$(IOU(B_i, b_i) < \gamma \lor C_i \neq c_i), \min ||x - x_{adv}||_p\}$ . + +The formula indicates that for each detected object $o_i$ , denoted by $(b_i, c_i)$ in victim image $x$ , the AFOG attack is successful on $o_i$ if the IoU of $B_i$ and $b_i$ is less than the detection threshold $\gamma$ (usually set to 0.5) or $c_i \in \{1, 2, \dots, K\} (C_i \neq c_i)$ , and $x_{adv}$ also satisfies the distortion constraint of $min||x - x_{adv}||_p$ , where $p$ is typically defined by $L_2$ norm, $L_0$ norm, or $L_\infty$ norm. + +![](images/95b784c0a6270e675ace62a5dfe730cb9f1043280133c4d34f074479516dd625.jpg) +Benign Model Training: +Freeze (※) Input, Update (♀) Parameters +$\partial \mathcal{L}(\pmb {x},\mathcal{O};\pmb {\theta})$ +Pk+1←Pk+αΓ[∂LAFOG(xad,OE;θ)] +$x_{adv_k}\gets \prod_{x\in}\left(x + A_k\odot P_k\right)$ +AFOG Attack: +Freeze (k) Parameters, Update (C) Attention and Perturbation, +Compute (O) Adversarial Image +Figure 2. AFOG attack framework. During attack iteration the adversarial perturbation is updated via bounding-box and class loss gradient propagation. An adversarial attention map concurrently learns to apply per-pixel amplification or dampening to the perturbation, focusing its effects on vulnerable image regions. + +# 3.2. Adversarial Loss Optimization + +AFOG attacks the victim model via an iterative projected gradient descent method [23]. Figure 2 provides an illustration of the AFOG attack framework. At the initialization step, the attack first propagates the unaltered victim image $x$ through the feed-forward network (FFN) of the victim model $f_{D}(\vartheta ,x)$ to obtain its benign prediction $\mathcal{O}_x = \{(b_i,c_i)|i = 1,\dots ,N_x\}$ . + +In a scenario where no ground truth is available, the attack assumes these benign predictions to be correct labels for the input image $x$ . For the remaining section, we will use $(B_i, C_i)$ in the context of performing an attack on $f_D(\vartheta, x)$ + +to obtain $x_{adv}$ . Upon initialization for each input image $x \in \mathcal{D}$ , the AFOG attack corrupts the victim image $x$ by adding perturbations generated via an element-wise multiplication of two components using Equation 1: + +$$ +x _ {a d v _ {k}} \leftarrow \prod_ {x, \epsilon} \left(x + A _ {k} \odot P _ {k}\right) \tag {1} +$$ + +$A$ is the attention map, and $P$ is the perturbation map. Here $\odot$ represents the Hadamard matrix product and $\Pi$ is a projection onto a hypersphere centered on unaltered victim image $x$ with the radius $\epsilon$ as the maximum perturbation budget. $A$ and $P$ are initialized according to Equation 2. + +$$ +A _ {0} \sim 1, P _ {0} \sim \text {R a n d o m} (- \epsilon , \epsilon) \tag {2} +$$ + +Random $(-\epsilon, \epsilon)$ is a uniform random distribution. The perturbed image $x_{adv}$ propagates through the FFN of the victim model. The attack loss function $\mathcal{L}_{AFOG}$ assesses the adversarial output $\mathcal{O}_{x_{adv}}$ and benign output $\mathcal{O}_x$ , computing a loss that reflects the attack's progress in corrupting image $x$ . This attack loss is formulated in Equations 3, 4, and 5. + +$$ +\begin{array}{l} \mathcal {L} _ {A F O G} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) = \mathcal {L} _ {b b o x} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) \tag {3} \\ + \mathcal {L} _ {c l s} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) \\ \end{array} +$$ + +$$ +\mathcal {L} _ {b b o x} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) = \sum_ {i = 1} ^ {N _ {x}} \left[ f _ {\vartheta} \left(x, o _ {i}\right) - f _ {\vartheta} \left(x _ {a d v}, o _ {a d v _ {i}}\right) \right] \tag {4} +$$ + +$$ +\mathcal {L} _ {c l s} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) = \sum_ {i = 1} ^ {N _ {x}} \left[ f _ {\vartheta} \left(x, c _ {i}\right) - f _ {\vartheta} \left(x _ {a d v}, c _ {a d v _ {i}}\right) \right] \tag {5} +$$ + +Given a victim image $x$ , the optimized attack loss $\mathcal{L}_{AFOG}$ is attained by making the model incorrectly predict every target object. We achieve this loss optimization by both falsifying every target object's bounding box prediction and its class label prediction. This can be represented by optimizing an adversarial bounding box loss $\mathcal{L}_{bbox}$ and an adversarial class label prediction loss $\mathcal{L}_{cls}$ for all $N_x$ target objects. Conceptually, this suppresses the confidence of the original correct bounding box and class label prediction, while increasing the confidence of the adversarial prediction of an incorrect bounding box or class label. + +Recall Figure 2: the AFOG attack freezes the model parameters $\vartheta$ and uses the gradient of the AFOG attack loss to update the attention map $A$ and the perturbation $P$ by backpropagation, generating the next iteration of the adversarial perturbation. A newly corrupted image $x_{adv}$ is created upon injecting the updated adversarial perturbation by following Equation 1, where A and P are updated via Equations 6 and 7: + +$$ +A _ {k + 1} \leftarrow A _ {k} + \alpha_ {A} \sigma \left[ \frac {\partial \mathcal {L} _ {A F O G} \left(x _ {a d v} , \mathcal {O} _ {x} ; \vartheta\right)}{\partial A _ {k}} \right] \tag {6} +$$ + +![](images/3096f28c10076b522da0a6ecc93a28d43b2fa7770c0fba59996f64f83ac0110f.jpg) +Figure 3. Example attention map generation on DETR-R101 for $k = 10$ iterations. AFOG attention initially learns important regions of the image, then iteratively updates to achieve an effective attack. It finds vulnerable pixels in foreground objects as well as background regions, including in unexpected or unintuitive areas such as the sky above the boat. + +$$ +P _ {k + 1} \leftarrow P _ {k} + \alpha_ {P} \Gamma \left[ \frac {\partial \mathcal {L} _ {A F O G} \left(x _ {a d v} , \mathcal {O} _ {x} ; \vartheta\right)}{\partial P _ {k}} \right] \tag {7} +$$ + +$\alpha_{A}$ and $\alpha_{P}$ denote the attention map learning rate and the perturbation learning rate respectively, $\sigma$ is a normalization function, $\Gamma$ is a sign function, $\mathcal{L}_{AFOG}$ is the attack loss, $\vartheta$ is the model parameters, and $\mathcal{O}_x$ is the model's benign predictions from input $x$ . This attack process repeats until the number of attack iterations is reached. Algorithm 1 provides the pseudocode. + +# 3.3. Adversarial Attention Mechanism + +A key innovation of our AFOG method is empowering perturbation generation with a learnable attention mechanism. Motivated by the intuition that certain parts of a victim image are more susceptible to adversarial perturbation than others, we add an attention map to focus the perturbation on vulnerable pixels. The AFOG attack concurrently learns the adversarial attention map alongside the perturbation to iteratively maximize the attack loss $\mathcal{L}_{\text{AFOG}}$ (recall Equation 3). Figure 3 gives an illustrative example of the iterative learning of AFOG's adversarial attention map and the corresponding perturbation. Unlike focusing mechanisms in other methods such as [33], AFOG's attention map dynamically updates during attack iteration, leaving it unconstrained by assumptions about foreground [12], background [16], or region proposal [33] importance from static methods. We posit that pixel importance may not correspond to human intuition, and design AFOG to iteratively learn pixel importance instead. We observe that in early iterations the attention mechanism tends to focus on the primary objects in an image, and then branches out to affect surrounding areas as the attack progresses. This adaptability is a key feature of learnable attention, showcasing an advantage over static attention maps. + +# Algorithm 1 AFOG attack on an input image. + +Require: Victim image $x \in \mathcal{D}$ , test-set $\mathcal{D}$ , Victim pretrained model $f_{D}(\vartheta)$ , Perturbation step size $\alpha_{P}$ , Attention step size $\alpha_{A}$ , Number of iterations $T$ , Maximum perturbation $\epsilon$ . + +1: Initialize $\mathcal{O}_x\gets f_D(x;\vartheta)$ +2: Initialize attention map $A_0 \gets 1$ +3: Initialize perturbation $P_0 \gets \text{Random}(-\epsilon, \epsilon)$ ; +4: Initialize step variable $k \gets 1$ + +5: while $k < T$ do +6: Attack image $x_{advk} \gets \prod_{x,\epsilon}(x + A_k \odot P_k)$ ; +7: Forward propagate $x_{adv}$ through $f_{D}(\vartheta, x_{adv})$ ; +8: Compute bbox-loss $\mathcal{L}_{bbox}(x_{adv},\mathcal{O}_x;\vartheta)$ +9: Compute $\operatorname{cls}$ -loss $\mathcal{L}_{cls}(x_{adv},\mathcal{O}_x;\vartheta)$ +10: $\mathcal{L}_{AFOG}(x_{adv},\mathcal{O}_x;\vartheta) = \mathtt{bbox - loss} + \mathtt{cls - loss};$ +11: Calculate losses with respect to $A_{k}$ and $P_{k}$ : $\mathcal{L}_A(x_{adv},\mathcal{O}_x;\vartheta),\mathcal{L}_P(x_{adv},\mathcal{O}_x;\vartheta)$ ; +12: Normalize attention loss $\mathcal{L}_A\gets \mathrm{Norm}(\mathcal{L}_A)$ +13: Take sign of perturbation loss $\mathcal{L}_P\gets \mathrm{Sign}(\mathcal{L}_P)$ +14: $A_{k + 1}\gets A_k - \alpha_A\mathcal{L}_A;$ +15: $P_{k + 1}\gets P_k - \alpha_P\mathcal{L}_P;$ +16: $k\gets k + 1;$ +17: end while +18: $x_{advk + 1}\gets \prod_{x,\epsilon}(x + A_{k + 1}\odot P_{k + 1})$ +19: return $x_{adv}$ + +# 3.4. Special Cases of the AFOG Attack + +We explore two special cases of our AFOG attack, each of which targets a specific malicious detection behavior in the victim detector. The first special case is an object vanishing attack, coined AFOG-V, which attempts to attack the objectness detection task of multi-box object detection. The goal of this special case AFOG attack is to make the victim model unable to detect any object, making all object + +detections vanish for victim image $x$ . We implement the AFOG-V attack by changing the initialization: we use an empty set instead of forward propagation of $x$ to obtain the benign detection results for the $N_{x}$ objects as the assumed ground-truth in AFOG-V attack. Let $\varnothing$ denote an altered version of $\mathcal{O}_x$ that contains no predictions. The formulation of the AFOG-V attack loss function is given by Equation 8: + +$$ +\begin{array}{l} \mathcal {L} _ {A F O G _ {V}} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) = - \mathcal {L} _ {b b o x} \left(x _ {a d v}, \varnothing ; \vartheta\right) \tag {8} \\ - \mathcal {L} _ {c l s} (x _ {a d v}, \varnothing ; \vartheta) \\ \end{array} +$$ + +The second special case is object fabrication, coined as AFOG-F, which attempts to attack the bounding-box detection task by generating perturbations that lead to spurious detections (i.e., false positives). Similarly, we revise the AFOG attack process: instead of keeping the benign predictions with the confidence score above a certain tunable threshold (0.5 by default), for AFOG-F, we remove this threshold and allow the benign detection set $\mathcal{O}_x$ to include a much larger set of "ground truth" targets, which inevitably include erroneous detections. Accordingly the AFOG-F loss function is given in Equation 9: + +$$ +\begin{array}{l} \mathcal {L} _ {A F O G _ {F}} \left(x _ {a d v}, \mathcal {O} _ {x}; \vartheta\right) = - \mathcal {L} _ {b b o x} \left(x _ {a d v}, \mathcal {O} _ {F}; \vartheta\right) \tag {9} \\ - \mathcal {L} _ {c l s} \left(x _ {a d v}, \mathcal {O} _ {F}; \vartheta\right) \\ \end{array} +$$ + +where $\mathcal{O}_F$ is a modified version such that the likelihood score of each prediction is set to 1.0. We explore these two specialized cases of our AFOG attack to gain a deeper understanding of the adverse effects of adversarial perturbations on different detection transformers. + +# 4. Experiments + +# 4.1. Experimental Setup + +We use the Common Objects in Context (COCO) 2017 [18] test-dev split to evaluate the performance of our attack on contemporary object detectors. COCO 2017 is a standard benchmark for evaluating object detector performance, and its test-dev set has 80 object categories across 5,000 images. We use the PASCAL Visual Objects Challenge (VOC) 2007 [9] dataset to compare our attack with the state-of-the-art attacks on CNN-based object detectors, because existing attacks designed only for CNN-based detectors all report their evaluation with VOC. The VOC devkit split contains 4,952 images with 20 object classes. All experiments were performed on an NVIDIA A100. We list our models and their benign performance in Table 1. We provide model details in Section ?? of the Supplementary Material. + +We select a variety of model sizes, ranging from Detection Transformer (DETR) [4], a lightweight 40 million parameter model, to EVA [10], a vision-focused foundation model with over one billion parameters. We use the Detrex framework [29], built on top of Detectron2 [34], to standardize our model implementations via DINO [37]. DINO + +and Detrex adapt many general purpose vision models to object detection. The Detrex framework standardizes all images in [0, 1]. We also attack the original versions of several transformer models to demonstrate that the AFOG attack also works beyond the Detrex framework. All 12 transformer models and FRCNN use PyTorch [26] implementations, whereas SSD300 [19] and YOLOv3 [27] use Tensorflow [1]. We use mean average precision (mAP) to evaluate both benign and attacked performance. The Average Precision (AP) is given by interpolating the product of precision and recall at several decision thresholds. The mean AP (mAP) is the average of APs over the number of object classes. A lower mAP indicates that the detector is less effective for the object detection task. The victim detector's mAP under adversarial perturbations reflects the degree of degradation in detection performance for the victim detector when compared with its benign mAP. We measure imperceptibility by reporting the average distortion of the attack via four metrics: $L_{2}$ norm, $L_{0}$ norm, semantic structural similarity index (SSIM), and mean distortion $\mu_{\Delta}$ . We compute each metric per image and report the average over all images in COCO 2017 test-dev. Our experiments in the next section show that setting the total number of attack iterations to 10 is effective for all 12 state-of-the-art detection models on the COCO benchmark. + +# 4.2. Effectiveness Comparison of AFOG Attack + +Table 1 shows the effectiveness of the AFOG attack and its two special extensions AFOG-V and AFOG-F on each of the twelve detection transformers with varying model sizes (Col. 2). We use their benign mAP scores (Col. 3) as the reference for the evaluation. We make three observations: (i) under AFOG attacks all 12 transformers suffer a drastic performance drop in mAP scores, ranging from $2.5 \times$ reduction to $37.8 \times$ reduction. (ii) Compared to the AFOG generic attack, the special extension AFOG-V offers stronger attack success rates in terms of mAP reduction for 11 out of 12 transformers, except DETR-R50 [4]; and the special extension AFOG-F offers stronger attack success rates for 7 out of 12 transformers, except DETR-R50 [4], DETR-R101 [4], Deformable-DETR [38], R50 (DINO) [11], Swin-L (DINO) [20] (iii) For AlignDETR [2], AFOG-F achieves a mAP of 1.36 and outperforms both AFOG (mAP of 18.13) and AFOG-V (mAP of 1.64), and all three attacks significantly reduce the benign mAP by a factor ranging from $2.8 \times$ to $37.8 \times$ . + +We compare AFOG against eleven benchmark attacks on DETR and Swin in Table 2. We observe that AFOG achieves superior performance on both DETR-R50 and Swin-L, notably recording an $82.7\%$ improvement over the next strongest attack on Swin. AFOG also uses the smallest perturbation budget and fewest iterations, indicating its superior imperceptibility and speed. + +![](images/63f6277155868b5513c5aa1e60099b6be22f87958231a8b1b63bd896f9d0f265.jpg) +Figure 4. Visualization with six examples, comparing benign scenario and AFOG attack results on Detection Transformer (DETR) with ResNet-50 backbone. Six diverse test images contain objects with varying classes, scales, perspectives, lighting, motion, and textures. + +![](images/9ff461ff15b503589fc4d7798b6cc9e4a3c591555f0baf2ea28c39a2f4a57990.jpg) + +![](images/ff1d53d2bcd8423a5e4684fa4ac683dacd3655a1c5ca86ed5d4dfb921f49182e.jpg) + +![](images/d00a828367bd278d88d795845f4bb35b53efdbfb3956521ac6690da4c0515b66.jpg) + +![](images/80d72ae204bda3c8d7e809d55187021db01e411f7e1a650ca4d1442bf0d7313d.jpg) + +![](images/cb2722a37a08ac01f2d52e29a8da55f72bde869fd5fb48b908266fb7a65e92a1.jpg) + +Table 1. AFOG effectiveness over 12 detection transformers, measured by mAP on perturbed images. (*) indicates DINO framework used with the corresponding backbone for object detection. + +
ModelParams (M)BenignAFOGAFOG-VAFOG-F
DETR-R50 [4]39.842.14.14.59.8
DETR-R101 [4]76.043.55.25.111.3
Deform.-DETR [38]40.044.54.81.57.1
R50*[11]47.649.25.31.56.3
AlignDETR [2]47.651.418.11.61.4
ViTDet* [14]108.154.93.80.92.8
ConvNeXt* [21]219.055.43.91.93.1
Swin-L* [20]217.256.87.32.48.6
InternImage* [31]241.056.97.32.85.1
FocalNet* [36]228.958.57.32.55.1
EVA* [10]1037.262.112.24.18.7
DETA [25]218.862.925.63.74.3
+ +Table 2. Comparison benchmark of AFOG against other state-of-the-art object detection attacks on DETR and Swin. Results are theirs. (*) indicates results from [32]. (†) indicates results from [24]. (-) indicates no result. + +
AttackTypePert. BudgetIters.Adversarial mAP
DETR-R50Swin
GARSDC [17]Surrogate0.053000+6.0-
GALD [13]Surrogate0.0631020.6-
RAD* [6]Surrogate0.0631027.247.2
GHFD* [32]Surrogate0.0635012.742.3
UEA* [33]Surrogate0.0635028.550.7
DAG* [35]Surrogate0.0635028.650.7
RAP* [15]Surrogate0.0635024.749.5
EBAD†[3]Victim0.0391034.9-
AttentionFool [22]Victim-10-15021.0-
OATB [12]Victim0.0782026.6-
DBA [16]Victim-50-56.7
AFOGVictim0.031104.17.3
+ +Figure 4 shows six visual examples and the comparison of DETR results for benign detection (Row 1) and adversarial detection under the AFOG attack (Row 2) scenarios. The first example shows that DETR detects the person and surfboard correctly under no attack (Row 1) but outputs erroneous detections under the AFOG attack (Row 2). Additional visual examples for all twelve transformers are available in the Supplementary Material. + +We report the distortion effect and timing cost for each of the 12 transformer models in Table 3. We note several interesting observations: (i) Considering the $L_{2}$ and SSIM metrics, AFOG attacks induce very similar levels of + +distortion for 10 out of 12 transformer models. The two largest detection transformers, DETA [25] and EVA [10], show worse $L_{2}$ and SSIM scores for AFOG, AFOG-V and AFOG-F. (ii) Considering the $L_{0}$ metric, only DETA shows the lowest $L_{0}$ value consistently for AFOG, AFOG-V and AFOG-F. (iii) As expected, the time required to attack a detection transformer model is strongly correlated to the number of model parameters. For instance, EVA [10] has over 1 billion parameters, and took much longer on average to attack an input image over 10 iterations compared to the other 11 models. (iii) DETA [25] has 218.8 million parameters and is similar to other models within the Detrex framework: FocalNet [36], InternImage [31], Swin-L [20] and ConvNext [21]. However, we observe that it consistently takes about $1.5 - 2 \times$ longer time to attack one input image on average with our AFOG attack or its special case extensions AFOG-V or AFOG-F. (iv) AFOG, AFOG-V and AFOG-F have similar average perturbation magnitude. + +# 4.3. Effectiveness on CNN-Based Detectors + +This section compares AFOG with four attacks designed for CNN-based models: TOG [7], UEA [33], RAP [15], and DAG [35]. They are surrogate-based attacks, and a majority of them (e.g., DAG, RAP, UEA) can only directly attack two-stage detectors (exemplified by Faster-R-CNN [28]) and rely on adversarial transferability to attack single-stage CNN-based detectors like YOLOv3 [27] and SSD [19]. Table 4 reports the comparison results. We make two observations: (i) AFOG and TOG are the only two victim-based attacks against single-stage detectors, represented by YOLOv3 and SSD. Both AFOG and TOG can drastically reduce the benign mAP of YOLOv3 $(83.43\%)$ and SSD $(76.11\%)$ . For YOLOv3, TOG is a stronger attack, achieving mAP of 0.56 compared to mAP of AFOG being 2.62. However, for SSD-300, AFOG is a stronger attack, achieving mAP of 0.50, compared to mAP of TOG (0.86). (ii) For two-stage CNN-based object detectors, exemplified by Faster-R-CNN, AFOG outperforms all four other victim-based attacks, achieving the lowest adversarial mAP score of $2.07\%$ . (iii) AFOG also uses the same distortion mag + +Table 3. Timing and Imperceptibility Results. $L_{2}$ represents the average $L_{2}$ norm difference between perturbed and clean images, $L_{0}$ is the average proportion of perturbed pixels, SSIM is the structural similarity index measure, $\mu_{\Delta}$ is the average perturbation magnitude, and t is average total attack time for all ten iterations in seconds. + +
ModelL2AFOGtimeL2AFOG-VtimeL2AFOG-Ftime
L0SSIMμΔL0SSIMμΔL0SSIMμΔ
DETR-R50 [4]0.03220.97070.87150.01731.450.03230.97070.87160.01720.990.03230.97100.87170.01721.21
DETR-R101 [4]0.03230.97070.87210.01731.470.03230.97060.87240.0131.160.03230.97080.87240.01721.70
Deform.-DETR [38]0.03230.97190.87110.01741.630.03230.97130.87160.01731.630.00120.97140.87170.01731.87
R50 [11]0.03170.96580.83430.01712.700.03170.96500.83480.01702.340.03170.96540.83460.01703.16
AlignDETR [2]0.03190.96570.83470.01702.410.03190.96460.83490.01702.330.03190.96470.83490.01703.29
ViTDet [14]0.03180.96710.83530.01716.880.03180.96570.83610.01706.670.03180.96640.83550.01717.33
ConvNext [21]0.03180.96660.83420.01715.380.03180.96540.83490.01705.260.03180.96630.83470.01705.86
Swin-L [20]0.03270.97240.86730.01757.130.03270.97160.86800.01737.280.03270.97220.86780.01748.99
InternImage [31]0.03180.96650.83600.01706.350.03180.96530.83670.01706.230.03180.96600.83640.01716.70
FocalNet [36]0.03200.96650.83650.01728.960.03200.96570.83780.01718.840.03200.96590.83740.01719.74
EVA [10]0.03700.96660.82400.017254.340.03700.96650.82460.017154.530.03700.96650.82420.017151.18
DETA [25]0.04810.96620.80960.017013.200.04810.96540.80990.017013.100.04810.96540.80990.017013.13
+ +nitude budget $L_{\infty}$ as TOG, while achieving higher attack success rate (lower mAP) on SSD-300 and FRCNN with decreased $L_{2}$ distortion cost. Hence, AFOG excels at attacking both transformer-based detectors and CNN-based detectors. + +Table 4. Comparing AFOG with four state-of-the-art attacks on representative CNN-based object detectors. mAPs of existing attacks were taken from respective papers [7]. (-) indicates N/A. + +
ModelAttackmAPtDistortion Cost
\(L_{\infty}\)\(L_{2}\)\(L_{0}\)SSIM
YOLOv3Benign83.430.00.00.00.01.0
TOG [7]0.560.980.0310.0830.9840.875
AFOG2.281.310.0310.0130.8550.801
SSD-300Benign76.110.00.00.00.01.0
UEA [33]20.0-----
DAG [35]64.0-----
TOG [7]0.860.390.0310.1200.9750.879
AFOG0.500.490.0310.0220.8580.793
FRCNNBenign67.370.00.00.00.01.0
UEA [33]5.00.170.3430.1910.9590.652
RAP [15]4.784.040.0820.0100.5310.994
DAG [35]3.567.990.0240.0020.4930.999
TOG [7]2.641.680.0310.0580.9760.862
AFOG2.382.110.0310.0190.8540.788
+ +# 4.4. Effect of Learnable Self-Attention + +We isolate and analyze the role of the learnable self-attention mechanism in empowering AFOG to successfully corrupt transformer-based models by probing for weak spots and performing fast and human-imperceptible perturbations. Figure 5 shows a comparison between AFOG with and without its learnable attention mechanism, measured by percent difference in adversarial mAP scores. Ablation results show that our learnable attention mechanism improves performance by up to $30.6\%$ on InternImage, with an average of $15.1\%$ improvement across all models. + +Figure 6 shows the self-attention map for the last encoder layer of DETR [4] at initialization and three progressive attack iterations $(3^{rd}, 4^{th}$ , and $8^{th}$ ). We observe that, as the attack progresses in iterations (indicated by grey arrows), the self-attention map becomes increasingly less associated with the benign detected objects. By the $8^{th}$ iteration of the AFOG attack, the self-attention maps have become almost + +entirely disjoint from the benign detections of the victim model (DETR) prior to the attack. We note that, when the model's self-attention map that reflects learned detection knowledge is changed, it is unlikely for prior knowledge to be kept intact. During the AFOG attack, the perturbed inputs focused by AFOG's own adversarial attention corrupt the pixel-level relationships learned by a well-trained victim model prior to adversarial injection (e.g., $k = 1$ ), causing catastrophic forgetting to occur. + +![](images/2dd6e1409fa5c9c44f15609da7c034edba9500a9ca8b2c390a724a9e7a2be09c.jpg) +Figure 5. Improvement (\% difference) of AFOG with attention over AFOG without attention. The learnable attention mechanism improves performance by as much as $30.6\%$ on InternImage, with an average improvement of $15.1\%$ across all models. + +# 4.5. Worst Case Analysis + +Our experimental results on 12 transformer-based detectors and 3 representative CNN-based detectors have shown that AFOG is a fast and effective attack. However, there are still cases where AFOG fails in its adversarial perturbations within a total of 10 iterations, one setting of our attack termination hyper-parameter. By investigating those cases where AFOG did not succeed, we found that the most common situation is where the attack fails to corrupt the central subject of the victim input image, e.g., a person. Figure 7 gives three visual examples, with the top one (Rows 1-2) being a successful attack followed by two failed cases (Rows 3-6) when attacking DETR-R50. Columns 2-3 in all six + +![](images/cb4ab033a6219e383c0db53918950e3a520a5676b12164acadb535148d5965a6.jpg) +Figure 6. DETR self-attention weights analysis. Each blue frame shows the self-attention weights of DETR-R50's last encoder layer at the red indicated point in the image at a given attack iteration $k$ . In early iterations, the model understands a strong association between the indicated red point on the elephant and the rest of the elephant shape. As the attack iterates (grey arrows), this association becomes corrupted, evidenced by the lack of encoder self-attention structure in later stages. By the $8^{th}$ iteration, the selected point has lost association with the rest of the elephant object, demonstrating that AFOG has disrupted learned knowledge. + +rows show DETR's encoder's last layer's self-attention map for the first and second locations indicated by red points in the first column. Rows 1, 3, and 5 show benign situations and their corresponding self-attention maps, and Rows 2, 4, 6 show the self-attention maps after the final attack iteration. The fourth column shows AFOG's adversarial attention map after the final attack iteration. + +We make three observations: (i) For the successful attack case (Rows 1-2), the self-attention maps of victim model DETR-R50 lose association between the indicated red point and the rest of the foreground object it comprises. (ii) In both unsuccessful cases (Rows 3-6), AFOG fails to disrupt this association. (iii) Consider AFOG's adversarial attention maps from the successful case in Figure 7 (Row 2, Col. 4), we observe a clear focus on the key objects in the victim input image. In comparison, AFOG's adversarial attention maps for the two failure cases (Row 4, Col. 4 and Row 6, Col. 4) both fail to focus on foreground objects. The first failure case (Rows 3 and 4) shows that the AFOG attention map attributes significant weight to the distant person in the left of the victim image, resulting in a large number of fabricated predictions in that region. Similarly, the attention map for the second victim image (Rows 5 and 6) appears to entirely miss the fire hydrant, resulting in a failure to disrupt this object. In these cases, AFOG's adversarial attention has learned to focus on the wrong pixels. + +# 5. Conclusion + +We have presented AFOG, an Attention-Focused Offensive Gradient attack. AFOG is effective for attacking both advanced object detection transformers and traditional CNN-based detectors with a unified, architecture-agnostic framework. AFOG utilizes a learnable attention mecha + +![](images/bf13afda4d23d1131fe77bb746aa8a22bf3a95849100c969f46f57648fbc60e4.jpg) +Figure 7. Worst case analysis with DETR-R50 as victim model. Rows 1-2 show a benign image and the successful AFOG perturbed image. Rows 3-6 show two cases where AFOG attack did not succeed. Columns 2-3 show DETR's encoder's self-attention map for the two locations marked in red in column one. Column 4 shows AFOG's internal attention map, where white pixels indicate high attention and blue indicates low. In the successful case (Rows 1-2), encoder self-attention maps (Cols. 2-3) lose associative structure. In the unsuccessful cases (Rows 4 and 6), AFOG fails to disrupt this structure and predictions are preserved. AFOG adversarial attention maps (Col. 4) show focus on important object structure in the successful case (Row 2) and lack of focus in the unsuccessful cases (Rows. 4, 6). + +nism to enable its adversarial perturbations to focus on vulnerable areas of images in multi-box detection tasks. AFOG's attack loss function integrates multiple feature losses (e.g., bounding-box loss, class loss) through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is efficient and stealthy. Adversarial perturbations generated by AFOG are visually imperceptible and yet can cause well-trained detectors to fail miserably. Extensive experiments on state-of-the-art object detectors show that AFOG is consistently effective across twelve object detection transformers. Comparative evaluation with nearly a dozen SOTA methods shows that AFOG significantly outperforms surrogate-based and victim-based attacks on both object detection transformers and CNN-based object detectors. + +Acknowledgement. This research is partially sponsored by the NSF CISE grants 2302720 and 2312758, CISCO Edge AI program, and GTRI PhD Fellowship. + +# References + +[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 5 +[2] Zhi Cai, Songtao Liu, Guodong Wang, Zheng Ge, Xiangyu Zhang, and Di Huang. Align-detr: Improving detr with simple iou-aware bce loss, 2023. 5, 6, 7 +[3]Zikui Cai, Yaoteng Tan, and M. Salman Asif. Ensemble-based blackbox attacks on dense prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4045-4055, 2023. 2, 6 +[4] Nicolas Carrion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-223, 2020. 1, 5, 6, 7 +[5] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2 +[6] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2, 6 +[7] Ka-Ho Chow, Ling Liu, Margaret Loper, Juhyun Bae, Mehmet Emre Gursoy, Stacey Truex, Wenqi Wei, and Yanzhao Wu. Adversarial objectness gradient attacks in real-time object detection systems. In IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications, pages 263-272. IEEE, 2020. 2, 6, 7 +[8] Papers With Code. Object detection on coco test-dev. https://paperswithcode.com/sota/object-detection-on-coco.1 +[9] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 2, 5 +[10] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022.5,6,7 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Shu. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 06 2016. 5, 6, 7 +[12] Zhuo Leng, Zesen Cheng, Pengxu Wei, and Jie Chen. Object-aware transfer-based black-box adversarial attack on object detector. In Qingshan Liu, Hanzi Wang, Zhanyu + +Ma, Weishi Zheng, Hongbin Zha, Xilin Chen, Liang Wang, and Rongrong Ji, editors, Pattern Recognition and Computer Vision, pages 278-289, Singapore, 2024. Springer Nature Singapore. 2, 4, 6 +[13] Tuo Li and Yahong Han. Improving transferable adversarial attack for vision transformers via global attention and local drop. Multimedia Systems, 29:3467 - 3480, 2023. 2, 6 +[14] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 2, 6, 7 +[15] Yuezun Li, Daniel Tian, Mingching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. In BMVC, 2018. 6, 7 +[16] Jiawei Lian, Shaohui Mei, Xiaofei Wang, Yi Wang, Lefan Wang, Yingjie Lu, Mingyang Ma, and Lap-Pui Chau. Attack anything: Blind dnns via universal background adversarial attack, 2024. 2, 4, 6 +[17] Siyuan Liang, Longkang Li, Yanbo Fan, and Xiaojun Jia. A large-scale multiple-objective method for black-box attack against object detection. European Conference on Computer Vision (ECCV) 2022, pages 619-636, 2022. 2, 6 +[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, ECCV 2014, pages 740-755. Springer International Publishing, 2014. 2, 5 +[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing. 1, 5, 6 +[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6, 7 +[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6, 7 +[22] Giulio Lovisotto, Nicole Finnie, Maurizio Munoz, Chaithanya Kumar Murnmadi, and Jan Hendrik Metzen. Give me your attention: Dot-product attention considered harmful for adversarial patch robustness. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15213-15222, 2022. 2, 6 +[23] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 3 +[24] Khoi Nguyen Tiet Nguyen, Wenyu Zhang, Kangkang Lu, Yuhuan Wu, Xingjian Zheng, Hui Li Tan, and Liangli Zhen. A Survey and Evaluation of Adversarial Attacks for Object Detection, Aug. 2024. arXiv:2408.01934 [cs]. 2, 6 +[25] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbihl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 6, 7 + +[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[27] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 1, 2, 5, 6 +[28] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 1, 6 +[29] Tianhe Ren, Shilong Liu, Feng Li, Hao Zhang, Ailing Zeng, Jie Yang, Xingyu Liao, Ding Jia, Hongyang Li, He Cao, Jianan Wang, Zhaoyang Zeng, Xianbiao Qi, Yuhui Yuan, Jianwei Yang, and Lei Zhang. detrex: Benchmarking detection transformers, 2023. 5 +[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1 +[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 6, 7 +[32] Yang Wang, Yunfei Zheng, Lei Chen, Zhen Yang, Jingwei Wu, and Tieyong Cao. Gradient-guided hierarchical feature attack for object detector. Journal of King Saud University - Computer and Information Sciences, 36(1):101901, 2024. 2, 6 +[33] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 954–960. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 4, 6, 7 +[34] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5 +[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie1, and Alan Yuill. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision, pages 1378-1387, 2017. 6, 7 +[36] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 6, 7 +[37] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection, 2022. 5 + +[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5, 6, 7 \ No newline at end of file diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/images.zip b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..20f09dd08839558c4aba8ee1b376d6d01cfc6cf8 --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f622ecdaba730cadf296f0d3b8fccdb88539929aa7fbf29162d6ebab1ac58648 +size 673512 diff --git a/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/layout.json b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..325c91859946c9fa94299d00fe9afcb961151288 --- /dev/null +++ b/2025/Adversarial Attention Perturbations for Large Object Detection Transformers/layout.json @@ -0,0 +1,9421 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 86, + 538, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 86, + 538, + 104 + ], + "spans": [ + { + "bbox": [ + 56, + 86, + 538, + 104 + ], + "type": "text", + "content": "Adversarial Attention Perturbations for Large Object Detection Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 125, + 447, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 125, + 447, + 182 + ], + "spans": [ + { + "bbox": [ + 146, + 125, + 447, + 182 + ], + "type": "text", + "content": "Zachary Yahn1, Selim Furkan Tekin1, Fatih Ilhan1, Sihao Hu1, Tiansheng Huang1, Yichang Xu1, Margaret Loper2, Ling Liu1 \n1Georgia Institute of Technology, Atlanta, GA \n2Georgia Tech Research Institute, Atlanta, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 184, + 487, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 184, + 487, + 210 + ], + "spans": [ + { + "bbox": [ + 104, + 184, + 487, + 210 + ], + "type": "text", + "content": "{zachary.yahn, stekin6, filhan, sihaohu, thuang, xuyichang}@gatech.edu, margaret.loper@gtri.gatech.edu, ling.liu@cc.gatech.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "spans": [ + { + "bbox": [ + 143, + 237, + 192, + 250 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "text", + "content": "Adversarial perturbations are useful tools for exposing vulnerabilities in neural networks. Existing adversarial perturbation methods for object detection are either limited to attacking CNN-based detectors or weak against transformer-based detectors. This paper presents an Attention-Focused Offensive Gradient (AFOG) attack against object detection transformers. By design, AFOG is neural-architecture agnostic and effective for attacking both large transformer-based object detectors and conventional CNN-based detectors with a unified adversarial attention framework. This paper makes three original contributions. First, AFOG utilizes a learnable attention mechanism that focuses perturbations on vulnerable image regions in multi-box detection tasks, increasing performance over non-attention baselines by up to " + }, + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "inline_equation", + "content": "30.6\\%" + }, + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "text", + "content": ". Second, AFOG's attack loss is formulated by integrating two types of feature loss through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is an efficient and stealthy adversarial perturbation method. It probes the weak spots of detection transformers by adding strategically generated and visually imperceptible perturbations which can cause well-trained object detection models to fail. Extensive experiments conducted with twelve large detection transformers on COCO demonstrate the efficacy of AFOG. Our empirical results also show that AFOG outperforms existing attacks on transformer-based and CNN-based object detectors by up to " + }, + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 47, + 262, + 290, + 600 + ], + "type": "text", + "content": " with superior speed and imperceptibility. Code is available at: Link." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 128, + 634 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 641, + 288, + 715 + ], + "type": "text", + "content": "Transformer-based neural architectures and algorithms have blossomed in recent years, enhancing computer vision tasks including object detection. Attention is the core of the transformer architecture [30]. Attention allows a detector to focus on specific regions of images based on objects of interest, effectively predicting the presence and lo" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 236, + 545, + 437 + ], + "blocks": [ + { + "bbox": [ + 307, + 236, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 437 + ], + "type": "image", + "image_path": "9be411cf3f959f4f01c2ca65169ed11fa1012491174efb1ba935f79ac70433be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 439, + 547, + 516 + ], + "lines": [ + { + "bbox": [ + 304, + 439, + 547, + 516 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 547, + 516 + ], + "type": "text", + "content": "Figure 1. Successful AFOG attacks against three object detection transformers. White pixels in the attention map indicate learned areas of high importance, often corresponding to objects in the input image. This example illustrates that AFOG's adversarial attention learns which regions of an image are most susceptible to adversarial perturbations. Even for the same input, learnable adversarial attention maps may differ under different models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "cation of each potential object. Powered by various attention mechanisms, detection transformers can handle overlapping objects more effectively than some older methods because they can capture long-range dependencies in an image, allowing models to reason about the relationships between objects. Since modern object detection transformers such as Swin [20] and DETR [4] significantly outperform traditional convolutional neural network (CNN)-based models [8], represented by Faster R-CNN [28], SSD [19] and YOLO-v3 [27], there is a pressing need to investigate and understand the vulnerabilities of large detection transformers in the presence of adversarial perturbations. Adversarial perturbations are helpful instrumentation for exposing vulnerabilities of large detection transformers, making them useful mechanisms for motivating more robust transformer models for object detection." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3184" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 263 + ], + "type": "text", + "content": "Existing attack methods struggle to disrupt transformer-based object detectors. Surrogate-based attacks (also known as black-box attacks), such as UEA [33] and RAD [5] generate perturbations on surrogate models and then test the adversarial effect against victim detectors that have similar detection architecture at inference. However, surrogate-based attacks suffer poor attack performance when the victim architectures are not similar to the surrogate models used to generate adversarial perturbations. Victim-based (also known as white-box) attacks, such as EBAD [3] and OATB [12], generate adversarial attacks by performing direct inference against victim models. Recent studies, such as AttentionFool [22], specialize in transformer-based victim models, but these transformer-only attacks are not applicable to convolutional detectors such as YOLO [27]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 264, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 660 + ], + "type": "text", + "content": "In this paper, we present an Attention-Focused Offensive Gradient (AFOG) attack, which is a victim architecture agnostic and effective in attacking both advanced object detection transformers and traditional CNN-based detectors. AFOG by design offers three novel characteristics. First, inspired by a transformer's self-attention, we utilize a learnable attention mechanism to enable AFOG to adaptively focus its adversarial perturbations on vulnerable image areas (those areas where perturbations have the greatest effect on the output) in multi-box detection tasks. Second, we formulate the attack loss function of AFOG by integrating learnable feature-map-based attention updates with iterative injection of adversarial perturbations. Finally, we design AFOG to be an efficient and yet stealthy adversarial perturbation method. By efficient, we want AFOG to rapidly generate adversarial perturbations in minimal iterations. By stealthy, we want AFOG to generate small amounts of perturbations that are visually imperceptible and yet can cause well-trained object detectors to produce erroneous detection results. Figure 1 shows an example of successful AFOG attacks against three detection transformers. Row 1 shows benign detection from three state-of-the-art detectors, Row 2 and Row 3 show AFOG adversarial perturbations on the three detectors and their respective AFOG adversarial attention maps, and Row 4 shows the detection results under our AFOG attack. We validate AFOG with extensive experiments on twelve state-of-the-art object detection transformers with the COCO benchmark [18] and three popular families of traditional CNN-based object detectors. The results show that AFOG is consistently effective across all twelve detection transformers compared to existing methods, achieving high attack success rate and reducing benign mAP (mean Average Precision) by up to " + }, + { + "bbox": [ + 46, + 264, + 289, + 660 + ], + "type": "inline_equation", + "content": "37.79 \\times" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 134, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "Existing adversarial perturbations against object detectors can be broadly categorized into surrogate-model based" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 515 + ], + "type": "text", + "content": "(also known as black-box) and victim-model based (white-box) methods. Surrogate-based approaches in literature tend to rely on the assumption that the surrogate model is similar to the victim model. RAD [6], GHFD [32], and UEA [33] generate adversarial examples on a surrogate FR-CNN model, then attack other CNN-based detectors such as YOLO and SSD. However, these attacks show poor performance against object detection transformers [32]. For victim-model based methods, GARSDC [17] is more effective but can require more than 3000 iterations to converge. GALD [13] first attacks vision transformer classifiers and then transfers the classification-based adversarial perturbations to object detectors of similar transformer architecture, and hence is less effective by comparison. We classify EBAD [3] as a victim-based attack because it uses an ensemble of surrogate models to attack a victim, but requires access to the victim's loss function for its nested ensemble optimization. EBAD shows poor performance when transferring perturbations from a DETR surrogate to a DETR victim [24]. AttentionFool [22] is a recent victim-based attack targeting dot-product self-attention against DETR [14] with a ResNet-101 backbone. However, AttentionFool shows inconsistent performance against DETR with a ResNet-50 backbone [22]. AttentionFool is also not applicable to convolutional models such as YOLO [27] because it exclusively targets self-attention mechanisms. TOG [7] targets a prediction's objectness score and introduces vanishing and fabrication attack modes. OATB [12] uses a \"division map\" that statically emphasizes image regions during perturbation based on object location priors. DBA [16] prioritizes perturbations against image backgrounds to enhance imperceptibility, though it shows almost no effect against a Swin transformer for object detection [20]. In comparison, AFOG adversarial perturbations show strong performance against a wide variety of transformer- and CNN-based detector victims while maintaining superior imperceptibility and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 529, + 389, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 389, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 389, + 544 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 550, + 418, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 418, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 418, + 562 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Given a victim detector " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta, x)" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is a victim image " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is the test set, let " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " contain " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "N_{x}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " objects to be detected, denoted by " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x = \\{O_1, O_2, \\ldots, O_{N_x}\\}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ". Each object " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq N_x" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ") is a recognition target for the detector " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta, x)" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ". Under benign scenarios, let " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "(B_i, C_i)" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " denote a ground truth object " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " with bounding box " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " and class label " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " denote the total number of ground truth classes, and " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_i \\in \\{1, 2, \\ldots, K\\}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", e.g., " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "K = 21" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " for the VOC dataset [9], including the background class. Given input image " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta, x)" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " outputs the benign prediction of " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " detected objects, denoted by " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}[f_{D}(\\vartheta, x)] = \\{(b_i, c_i) | i = 1, \\ldots, N_x\\}" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": ". Each detected object " + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": " is associated with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3185" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": "its predicted bounding box " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " and predicted class label " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "(b_{i},c_{i})" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " is evaluated as a correct prediction if the intersection over union (IoU) of " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " is greater than the detection threshold " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": " (usually set to 0.5), and " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "C_i = c_i" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "(1\\leq i\\leq N_x," + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{D})" + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "text", + "content": ". The overall detection accuracy is measured using mAP (mean Average Precision) on the entire test dataset " + }, + { + "bbox": [ + 46, + 72, + 286, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": " denote the adversarial example generated by injecting a sequence of adversarial perturbations to " + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": " through an iterative attention-based learning mechanism. The goal of our AFOG attack on detector " + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta, x)" + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": " is to find " + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": " that maximizes the success rate of falsifying the predictions of all object recognitions for all images in " + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 45, + 144, + 287, + 216 + ], + "type": "text", + "content": ", i.e." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 217, + 250, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 250, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 250, + 229 + ], + "type": "inline_equation", + "content": "\\arg \\max_{x\\in \\mathcal{D},i\\in N_x},\\{(b_i,c_i)\\in \\mathcal{R}[f_D(\\vartheta ,x_{adv})]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 229, + 277, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 229, + 277, + 243 + ], + "spans": [ + { + "bbox": [ + 72, + 229, + 277, + 243 + ], + "type": "inline_equation", + "content": "(IOU(B_i, b_i) < \\gamma \\lor C_i \\neq c_i), \\min ||x - x_{adv}||_p\\}" + }, + { + "bbox": [ + 72, + 229, + 277, + 243 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": "The formula indicates that for each detected object " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "(b_i, c_i)" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " in victim image " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": ", the AFOG attack is successful on " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " if the IoU of " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "b_i" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " is less than the detection threshold " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " (usually set to 0.5) or " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "c_i \\in \\{1, 2, \\dots, K\\} (C_i \\neq c_i)" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " also satisfies the distortion constraint of " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "min||x - x_{adv}||_p" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " is typically defined by " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " norm, " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "L_0" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " norm, or " + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "inline_equation", + "content": "L_\\infty" + }, + { + "bbox": [ + 46, + 243, + 287, + 326 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 352, + 264, + 434 + ], + "blocks": [ + { + "bbox": [ + 58, + 342, + 124, + 349 + ], + "lines": [ + { + "bbox": [ + 58, + 342, + 124, + 349 + ], + "spans": [ + { + "bbox": [ + 58, + 342, + 124, + 349 + ], + "type": "text", + "content": "Benign Model Training:" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 349, + 164, + 357 + ], + "lines": [ + { + "bbox": [ + 59, + 349, + 164, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 349, + 164, + 357 + ], + "type": "text", + "content": "Freeze (※) Input, Update (♀) Parameters" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 180, + 345, + 238, + 352 + ], + "lines": [ + { + "bbox": [ + 180, + 345, + 238, + 352 + ], + "spans": [ + { + "bbox": [ + 180, + 345, + 238, + 352 + ], + "type": "inline_equation", + "content": "\\partial \\mathcal{L}(\\pmb {x},\\mathcal{O};\\pmb {\\theta})" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 156, + 352, + 264, + 434 + ], + "lines": [ + { + "bbox": [ + 156, + 352, + 264, + 434 + ], + "spans": [ + { + "bbox": [ + 156, + 352, + 264, + 434 + ], + "type": "image", + "image_path": "95b784c0a6270e675ace62a5dfe730cb9f1043280133c4d34f074479516dd625.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 434, + 264, + 448 + ], + "lines": [ + { + "bbox": [ + 156, + 434, + 264, + 448 + ], + "spans": [ + { + "bbox": [ + 156, + 434, + 264, + 448 + ], + "type": "text", + "content": "Pk+1←Pk+αΓ[∂LAFOG(xad,OE;θ)]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 448, + 143, + 464 + ], + "lines": [ + { + "bbox": [ + 70, + 448, + 143, + 464 + ], + "spans": [ + { + "bbox": [ + 70, + 448, + 143, + 464 + ], + "type": "inline_equation", + "content": "x_{adv_k}\\gets \\prod_{x\\in}\\left(x + A_k\\odot P_k\\right)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 465, + 97, + 472 + ], + "lines": [ + { + "bbox": [ + 59, + 465, + 97, + 472 + ], + "spans": [ + { + "bbox": [ + 59, + 465, + 97, + 472 + ], + "type": "text", + "content": "AFOG Attack:" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 472, + 218, + 479 + ], + "lines": [ + { + "bbox": [ + 59, + 472, + 218, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 472, + 218, + 479 + ], + "type": "text", + "content": "Freeze (k) Parameters, Update (C) Attention and Perturbation," + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 480, + 140, + 487 + ], + "lines": [ + { + "bbox": [ + 59, + 480, + 140, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 480, + 140, + 487 + ], + "type": "text", + "content": "Compute (O) Adversarial Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 551 + ], + "type": "text", + "content": "Figure 2. AFOG attack framework. During attack iteration the adversarial perturbation is updated via bounding-box and class loss gradient propagation. An adversarial attention map concurrently learns to apply per-pixel amplification or dampening to the perturbation, focusing its effects on vulnerable image regions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 562, + 214, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 214, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 214, + 574 + ], + "type": "text", + "content": "3.2. Adversarial Loss Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": "AFOG attacks the victim model via an iterative projected gradient descent method [23]. Figure 2 provides an illustration of the AFOG attack framework. At the initialization step, the attack first propagates the unaltered victim image " + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": " through the feed-forward network (FFN) of the victim model " + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta ,x)" + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": " to obtain its benign prediction " + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x = \\{(b_i,c_i)|i = 1,\\dots ,N_x\\}" + }, + { + "bbox": [ + 46, + 581, + 287, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": "In a scenario where no ground truth is available, the attack assumes these benign predictions to be correct labels for the input image " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": ". For the remaining section, we will use " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "(B_i, C_i)" + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "text", + "content": " in the context of performing an attack on " + }, + { + "bbox": [ + 47, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_D(\\vartheta, x)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "to obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". Upon initialization for each input image " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", the AFOG attack corrupts the victim image " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " by adding perturbations generated via an element-wise multiplication of two components using Equation 1:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 123, + 545, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 123, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 367, + 123, + 545, + 148 + ], + "type": "interline_equation", + "content": "x _ {a d v _ {k}} \\leftarrow \\prod_ {x, \\epsilon} \\left(x + A _ {k} \\odot P _ {k}\\right) \\tag {1}", + "image_path": "4a6cdbcb512a63a6dc5593b88133ce337d65a4fd5c52f7b5aac534ec4b79bad5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " is the attention map, and " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " is the perturbation map. Here " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " represents the Hadamard matrix product and " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " is a projection onto a hypersphere centered on unaltered victim image " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " with the radius " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " as the maximum perturbation budget. " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 151, + 545, + 211 + ], + "type": "text", + "content": " are initialized according to Equation 2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 361, + 214, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 214, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 361, + 214, + 545, + 228 + ], + "type": "interline_equation", + "content": "A _ {0} \\sim 1, P _ {0} \\sim \\text {R a n d o m} (- \\epsilon , \\epsilon) \\tag {2}", + "image_path": "9050e9abb47f10a3540f154a4cc0ac4fa76e8878840858a9da1e50972ebfff0f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": "Random " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "(-\\epsilon, \\epsilon)" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": " is a uniform random distribution. The perturbed image " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": " propagates through the FFN of the victim model. The attack loss function " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AFOG}" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": " assesses the adversarial output " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_{x_{adv}}" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": " and benign output " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": ", computing a loss that reflects the attack's progress in corrupting image " + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": ". This attack loss is formulated in Equations 3, 4, and 5." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 335, + 311, + 545, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 311, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 335, + 311, + 545, + 338 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {A F O G} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\tag {3} \\\\ + \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) \\\\ \\end{array}", + "image_path": "86dd9458283ec08ed410ff2166ccb8de353a815bb98e2dfe0bf1fe9d5b05651f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 361, + 545, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 545, + 404 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, o _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, o _ {a d v _ {i}}\\right) \\right] \\tag {4}", + "image_path": "6dc0052caf8fe2475cb2056d0c769fad43076a4688de4bdbfc8de53079753761.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 311, + 422, + 545, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 311, + 422, + 545, + 455 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = \\sum_ {i = 1} ^ {N _ {x}} \\left[ f _ {\\vartheta} \\left(x, c _ {i}\\right) - f _ {\\vartheta} \\left(x _ {a d v}, c _ {a d v _ {i}}\\right) \\right] \\tag {5}", + "image_path": "f7182cd295ac1b7dcaf73285a04c8150f3190d19f959996feca310d7df934337.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": "Given a victim image " + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": ", the optimized attack loss " + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AFOG}" + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": " is attained by making the model incorrectly predict every target object. We achieve this loss optimization by both falsifying every target object's bounding box prediction and its class label prediction. This can be represented by optimizing an adversarial bounding box loss " + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbox}" + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": " and an adversarial class label prediction loss " + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "inline_equation", + "content": "N_x" + }, + { + "bbox": [ + 304, + 464, + 545, + 594 + ], + "type": "text", + "content": " target objects. Conceptually, this suppresses the confidence of the original correct bounding box and class label prediction, while increasing the confidence of the adversarial prediction of an incorrect bounding box or class label." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "content": "Recall Figure 2: the AFOG attack freezes the model parameters " + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "content": " and uses the gradient of the AFOG attack loss to update the attention map " + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "content": " and the perturbation " + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "content": " by backpropagation, generating the next iteration of the adversarial perturbation. A newly corrupted image " + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 304, + 595, + 545, + 689 + ], + "type": "text", + "content": " is created upon injecting the updated adversarial perturbation by following Equation 1, where A and P are updated via Equations 6 and 7:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 333, + 691, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 691, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 333, + 691, + 545, + 717 + ], + "type": "interline_equation", + "content": "A _ {k + 1} \\leftarrow A _ {k} + \\alpha_ {A} \\sigma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial A _ {k}} \\right] \\tag {6}", + "image_path": "5caebb515a5b819b9edcefe278a48197fe7c11a97160b6337bc842056cf33058.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3186" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 136, + 76, + 459, + 210 + ], + "blocks": [ + { + "bbox": [ + 136, + 76, + 459, + 210 + ], + "lines": [ + { + "bbox": [ + 136, + 76, + 459, + 210 + ], + "spans": [ + { + "bbox": [ + 136, + 76, + 459, + 210 + ], + "type": "image", + "image_path": "3096f28c10076b522da0a6ecc93a28d43b2fa7770c0fba59996f64f83ac0110f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": "Figure 3. Example attention map generation on DETR-R101 for " + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 46, + 217, + 547, + 251 + ], + "type": "text", + "content": " iterations. AFOG attention initially learns important regions of the image, then iteratively updates to achieve an effective attack. It finds vulnerable pixels in foreground objects as well as background regions, including in unexpected or unintuitive areas such as the sky above the boat." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 280, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 280, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 75, + 280, + 287, + 306 + ], + "type": "interline_equation", + "content": "P _ {k + 1} \\leftarrow P _ {k} + \\alpha_ {P} \\Gamma \\left[ \\frac {\\partial \\mathcal {L} _ {A F O G} \\left(x _ {a d v} , \\mathcal {O} _ {x} ; \\vartheta\\right)}{\\partial P _ {k}} \\right] \\tag {7}", + "image_path": "415983b529479d7c61c016e6d7364fdab5b08d536662918efe6a3a393debdce4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\alpha_{A}" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\alpha_{P}" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " denote the attention map learning rate and the perturbation learning rate respectively, " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " is a normalization function, " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " is a sign function, " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AFOG}" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " is the attack loss, " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " is the model parameters, and " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": " is the model's benign predictions from input " + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 319, + 289, + 403 + ], + "type": "text", + "content": ". This attack process repeats until the number of attack iterations is reached. Algorithm 1 provides the pseudocode." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 420, + 230, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 420, + 230, + 432 + ], + "spans": [ + { + "bbox": [ + 47, + 420, + 230, + 432 + ], + "type": "text", + "content": "3.3. Adversarial Attention Mechanism" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 439, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 289, + 715 + ], + "type": "text", + "content": "A key innovation of our AFOG method is empowering perturbation generation with a learnable attention mechanism. Motivated by the intuition that certain parts of a victim image are more susceptible to adversarial perturbation than others, we add an attention map to focus the perturbation on vulnerable pixels. The AFOG attack concurrently learns the adversarial attention map alongside the perturbation to iteratively maximize the attack loss " + }, + { + "bbox": [ + 46, + 439, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{AFOG}}" + }, + { + "bbox": [ + 46, + 439, + 289, + 715 + ], + "type": "text", + "content": " (recall Equation 3). Figure 3 gives an illustrative example of the iterative learning of AFOG's adversarial attention map and the corresponding perturbation. Unlike focusing mechanisms in other methods such as [33], AFOG's attention map dynamically updates during attack iteration, leaving it unconstrained by assumptions about foreground [12], background [16], or region proposal [33] importance from static methods. We posit that pixel importance may not correspond to human intuition, and design AFOG to iteratively learn pixel importance instead. We observe that in early iterations the attention mechanism tends to focus on the primary objects in an image, and then branches out to affect surrounding areas as the attack progresses. This adaptability is a key feature of learnable attention, showcasing an advantage over static attention maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 270, + 494, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 494, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 494, + 282 + ], + "type": "text", + "content": "Algorithm 1 AFOG attack on an input image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": "Require: Victim image " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", test-set " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", Victim pretrained model " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta)" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", Perturbation step size " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\alpha_{P}" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", Attention step size " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\alpha_{A}" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", Number of iterations " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": ", Maximum perturbation " + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 283, + 545, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 331, + 510, + 378 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 312, + 331, + 428, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 331, + 428, + 342 + ], + "spans": [ + { + "bbox": [ + 312, + 331, + 428, + 342 + ], + "type": "text", + "content": "1: Initialize " + }, + { + "bbox": [ + 312, + 331, + 428, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x\\gets f_D(x;\\vartheta)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 343, + 456, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 343, + 456, + 354 + ], + "spans": [ + { + "bbox": [ + 312, + 343, + 456, + 354 + ], + "type": "text", + "content": "2: Initialize attention map " + }, + { + "bbox": [ + 312, + 343, + 456, + 354 + ], + "type": "inline_equation", + "content": "A_0 \\gets 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 355, + 510, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 355, + 510, + 366 + ], + "spans": [ + { + "bbox": [ + 312, + 355, + 510, + 366 + ], + "type": "text", + "content": "3: Initialize perturbation " + }, + { + "bbox": [ + 312, + 355, + 510, + 366 + ], + "type": "inline_equation", + "content": "P_0 \\gets \\text{Random}(-\\epsilon, \\epsilon)" + }, + { + "bbox": [ + 312, + 355, + 510, + 366 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 367, + 446, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 367, + 446, + 378 + ], + "spans": [ + { + "bbox": [ + 312, + 367, + 446, + 378 + ], + "type": "text", + "content": "4: Initialize step variable " + }, + { + "bbox": [ + 312, + 367, + 446, + 378 + ], + "type": "inline_equation", + "content": "k \\gets 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 309, + 331, + 545, + 583 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 379, + 391, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 379, + 391, + 389 + ], + "spans": [ + { + "bbox": [ + 312, + 379, + 391, + 389 + ], + "type": "text", + "content": "5: while " + }, + { + "bbox": [ + 312, + 379, + 391, + 389 + ], + "type": "inline_equation", + "content": "k < T" + }, + { + "bbox": [ + 312, + 379, + 391, + 389 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 390, + 510, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 390, + 510, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 390, + 510, + 409 + ], + "type": "text", + "content": "6: Attack image " + }, + { + "bbox": [ + 312, + 390, + 510, + 409 + ], + "type": "inline_equation", + "content": "x_{advk} \\gets \\prod_{x,\\epsilon}(x + A_k \\odot P_k)" + }, + { + "bbox": [ + 312, + 390, + 510, + 409 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "text", + "content": "7: Forward propagate " + }, + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "inline_equation", + "content": "x_{adv}" + }, + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "inline_equation", + "content": "f_{D}(\\vartheta, x_{adv})" + }, + { + "bbox": [ + 312, + 411, + 525, + 422 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 422, + 502, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 422, + 502, + 434 + ], + "spans": [ + { + "bbox": [ + 312, + 422, + 502, + 434 + ], + "type": "text", + "content": "8: Compute bbox-loss " + }, + { + "bbox": [ + 312, + 422, + 502, + 434 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bbox}(x_{adv},\\mathcal{O}_x;\\vartheta)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "type": "text", + "content": "9: Compute " + }, + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "type": "inline_equation", + "content": "\\operatorname{cls}" + }, + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "type": "text", + "content": "-loss " + }, + { + "bbox": [ + 312, + 434, + 487, + 445 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}(x_{adv},\\mathcal{O}_x;\\vartheta)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 309, + 446, + 522, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 446, + 522, + 457 + ], + "spans": [ + { + "bbox": [ + 309, + 446, + 522, + 457 + ], + "type": "text", + "content": "10: " + }, + { + "bbox": [ + 309, + 446, + 522, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AFOG}(x_{adv},\\mathcal{O}_x;\\vartheta) = \\mathtt{bbox - loss} + \\mathtt{cls - loss};" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "text", + "content": "11: Calculate losses with respect to " + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "inline_equation", + "content": "A_{k}" + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "inline_equation", + "content": "P_{k}" + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A(x_{adv},\\mathcal{O}_x;\\vartheta),\\mathcal{L}_P(x_{adv},\\mathcal{O}_x;\\vartheta)" + }, + { + "bbox": [ + 309, + 458, + 545, + 482 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 309, + 483, + 516, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 483, + 516, + 494 + ], + "spans": [ + { + "bbox": [ + 309, + 483, + 516, + 494 + ], + "type": "text", + "content": "12: Normalize attention loss " + }, + { + "bbox": [ + 309, + 483, + 516, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_A\\gets \\mathrm{Norm}(\\mathcal{L}_A)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 309, + 495, + 533, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 495, + 533, + 506 + ], + "spans": [ + { + "bbox": [ + 309, + 495, + 533, + 506 + ], + "type": "text", + "content": "13: Take sign of perturbation loss " + }, + { + "bbox": [ + 309, + 495, + 533, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_P\\gets \\mathrm{Sign}(\\mathcal{L}_P)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 309, + 506, + 432, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 506, + 432, + 517 + ], + "spans": [ + { + "bbox": [ + 309, + 506, + 432, + 517 + ], + "type": "text", + "content": "14: " + }, + { + "bbox": [ + 309, + 506, + 432, + 517 + ], + "type": "inline_equation", + "content": "A_{k + 1}\\gets A_k - \\alpha_A\\mathcal{L}_A;" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 309, + 518, + 430, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 518, + 430, + 529 + ], + "spans": [ + { + "bbox": [ + 309, + 518, + 430, + 529 + ], + "type": "text", + "content": "15: " + }, + { + "bbox": [ + 309, + 518, + 430, + 529 + ], + "type": "inline_equation", + "content": "P_{k + 1}\\gets P_k - \\alpha_P\\mathcal{L}_P;" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 530, + 388, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 530, + 388, + 540 + ], + "spans": [ + { + "bbox": [ + 309, + 530, + 388, + 540 + ], + "type": "text", + "content": "16: " + }, + { + "bbox": [ + 309, + 530, + 388, + 540 + ], + "type": "inline_equation", + "content": "k\\gets k + 1;" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 309, + 542, + 367, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 542, + 367, + 552 + ], + "spans": [ + { + "bbox": [ + 309, + 542, + 367, + 552 + ], + "type": "text", + "content": "17: end while" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 309, + 553, + 470, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 553, + 470, + 572 + ], + "spans": [ + { + "bbox": [ + 309, + 553, + 470, + 572 + ], + "type": "text", + "content": "18: " + }, + { + "bbox": [ + 309, + 553, + 470, + 572 + ], + "type": "inline_equation", + "content": "x_{advk + 1}\\gets \\prod_{x,\\epsilon}(x + A_{k + 1}\\odot P_{k + 1})" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 309, + 573, + 374, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 573, + 374, + 583 + ], + "spans": [ + { + "bbox": [ + 309, + 573, + 374, + 583 + ], + "type": "text", + "content": "19: return " + }, + { + "bbox": [ + 309, + 573, + 374, + 583 + ], + "type": "inline_equation", + "content": "x_{adv}" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 611, + 490, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 490, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 490, + 624 + ], + "type": "text", + "content": "3.4. Special Cases of the AFOG Attack" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "We explore two special cases of our AFOG attack, each of which targets a specific malicious detection behavior in the victim detector. The first special case is an object vanishing attack, coined AFOG-V, which attempts to attack the objectness detection task of multi-box object detection. The goal of this special case AFOG attack is to make the victim model unable to detect any object, making all object" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3187" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "detections vanish for victim image " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". We implement the AFOG-V attack by changing the initialization: we use an empty set instead of forward propagation of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to obtain the benign detection results for the " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "N_{x}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " objects as the assumed ground-truth in AFOG-V attack. Let " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\varnothing" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " denote an altered version of " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " that contains no predictions. The formulation of the AFOG-V attack loss function is given by Equation 8:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 165, + 287, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 165, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 73, + 165, + 287, + 193 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {A F O G _ {V}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\varnothing ; \\vartheta\\right) \\tag {8} \\\\ - \\mathcal {L} _ {c l s} (x _ {a d v}, \\varnothing ; \\vartheta) \\\\ \\end{array}", + "image_path": "95cb75d9c796ea8882c623933dc429b9e96c3b2a196cd17e8064f37d74a764b5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 200, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 287, + 331 + ], + "type": "text", + "content": "The second special case is object fabrication, coined as AFOG-F, which attempts to attack the bounding-box detection task by generating perturbations that lead to spurious detections (i.e., false positives). Similarly, we revise the AFOG attack process: instead of keeping the benign predictions with the confidence score above a certain tunable threshold (0.5 by default), for AFOG-F, we remove this threshold and allow the benign detection set " + }, + { + "bbox": [ + 46, + 200, + 287, + 331 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_x" + }, + { + "bbox": [ + 46, + 200, + 287, + 331 + ], + "type": "text", + "content": " to include a much larger set of \"ground truth\" targets, which inevitably include erroneous detections. Accordingly the AFOG-F loss function is given in Equation 9:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 331, + 287, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 331, + 287, + 359 + ], + "spans": [ + { + "bbox": [ + 64, + 331, + 287, + 359 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {A F O G _ {F}} \\left(x _ {a d v}, \\mathcal {O} _ {x}; \\vartheta\\right) = - \\mathcal {L} _ {b b o x} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\tag {9} \\\\ - \\mathcal {L} _ {c l s} \\left(x _ {a d v}, \\mathcal {O} _ {F}; \\vartheta\\right) \\\\ \\end{array}", + "image_path": "b4ac2f5d8e29f0bb7864ab4ebd2de2c2a1384b9ea2dad4eea6d5d918a447e400.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_F" + }, + { + "bbox": [ + 46, + 365, + 287, + 426 + ], + "type": "text", + "content": " is a modified version such that the likelihood score of each prediction is set to 1.0. We explore these two specialized cases of our AFOG attack to gain a deeper understanding of the adverse effects of adversarial perturbations on different detection transformers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 128, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 456, + 163, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 163, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 163, + 468 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 641 + ], + "type": "text", + "content": "We use the Common Objects in Context (COCO) 2017 [18] test-dev split to evaluate the performance of our attack on contemporary object detectors. COCO 2017 is a standard benchmark for evaluating object detector performance, and its test-dev set has 80 object categories across 5,000 images. We use the PASCAL Visual Objects Challenge (VOC) 2007 [9] dataset to compare our attack with the state-of-the-art attacks on CNN-based object detectors, because existing attacks designed only for CNN-based detectors all report their evaluation with VOC. The VOC devkit split contains 4,952 images with 20 object classes. All experiments were performed on an NVIDIA A100. We list our models and their benign performance in Table 1. We provide model details in Section ?? of the Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "We select a variety of model sizes, ranging from Detection Transformer (DETR) [4], a lightweight 40 million parameter model, to EVA [10], a vision-focused foundation model with over one billion parameters. We use the Detrex framework [29], built on top of Detectron2 [34], to standardize our model implementations via DINO [37]. DINO" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "content": "and Detrex adapt many general purpose vision models to object detection. The Detrex framework standardizes all images in [0, 1]. We also attack the original versions of several transformer models to demonstrate that the AFOG attack also works beyond the Detrex framework. All 12 transformer models and FRCNN use PyTorch [26] implementations, whereas SSD300 [19] and YOLOv3 [27] use Tensorflow [1]. We use mean average precision (mAP) to evaluate both benign and attacked performance. The Average Precision (AP) is given by interpolating the product of precision and recall at several decision thresholds. The mean AP (mAP) is the average of APs over the number of object classes. A lower mAP indicates that the detector is less effective for the object detection task. The victim detector's mAP under adversarial perturbations reflects the degree of degradation in detection performance for the victim detector when compared with its benign mAP. We measure imperceptibility by reporting the average distortion of the attack via four metrics: " + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "content": " norm, " + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "content": " norm, semantic structural similarity index (SSIM), and mean distortion " + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "inline_equation", + "content": "\\mu_{\\Delta}" + }, + { + "bbox": [ + 304, + 72, + 547, + 371 + ], + "type": "text", + "content": ". We compute each metric per image and report the average over all images in COCO 2017 test-dev. Our experiments in the next section show that setting the total number of attack iterations to 10 is effective for all 12 state-of-the-art detection models on the COCO benchmark." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 381, + 531, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 381, + 531, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 381, + 531, + 395 + ], + "type": "text", + "content": "4.2. Effectiveness Comparison of AFOG Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": "Table 1 shows the effectiveness of the AFOG attack and its two special extensions AFOG-V and AFOG-F on each of the twelve detection transformers with varying model sizes (Col. 2). We use their benign mAP scores (Col. 3) as the reference for the evaluation. We make three observations: (i) under AFOG attacks all 12 transformers suffer a drastic performance drop in mAP scores, ranging from " + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "inline_equation", + "content": "2.5 \\times" + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": " reduction to " + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "inline_equation", + "content": "37.8 \\times" + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": " reduction. (ii) Compared to the AFOG generic attack, the special extension AFOG-V offers stronger attack success rates in terms of mAP reduction for 11 out of 12 transformers, except DETR-R50 [4]; and the special extension AFOG-F offers stronger attack success rates for 7 out of 12 transformers, except DETR-R50 [4], DETR-R101 [4], Deformable-DETR [38], R50 (DINO) [11], Swin-L (DINO) [20] (iii) For AlignDETR [2], AFOG-F achieves a mAP of 1.36 and outperforms both AFOG (mAP of 18.13) and AFOG-V (mAP of 1.64), and all three attacks significantly reduce the benign mAP by a factor ranging from " + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "inline_equation", + "content": "2.8 \\times" + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "inline_equation", + "content": "37.8 \\times" + }, + { + "bbox": [ + 304, + 401, + 545, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": "We compare AFOG against eleven benchmark attacks on DETR and Swin in Table 2. We observe that AFOG achieves superior performance on both DETR-R50 and Swin-L, notably recording an " + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "inline_equation", + "content": "82.7\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 714 + ], + "type": "text", + "content": " improvement over the next strongest attack on Swin. AFOG also uses the smallest perturbation budget and fewest iterations, indicating its superior imperceptibility and speed." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3188" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 136, + 178 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 136, + 178 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 136, + 178 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 136, + 178 + ], + "type": "image", + "image_path": "63f6277155868b5513c5aa1e60099b6be22f87958231a8b1b63bd896f9d0f265.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 546, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 546, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 546, + 213 + ], + "type": "text", + "content": "Figure 4. Visualization with six examples, comparing benign scenario and AFOG attack results on Detection Transformer (DETR) with ResNet-50 backbone. Six diverse test images contain objects with varying classes, scales, perspectives, lighting, motion, and textures." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 225, + 178 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 225, + 178 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 225, + 178 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 225, + 178 + ], + "type": "image", + "image_path": "9ff461ff15b503589fc4d7798b6cc9e4a3c591555f0baf2ea28c39a2f4a57990.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 302, + 178 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 302, + 178 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 302, + 178 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 302, + 178 + ], + "type": "image", + "image_path": "ff1d53d2bcd8423a5e4684fa4ac683dacd3655a1c5ca86ed5d4dfb921f49182e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 383, + 179 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 383, + 179 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 383, + 179 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 383, + 179 + ], + "type": "image", + "image_path": "d00a828367bd278d88d795845f4bb35b53efdbfb3956521ac6690da4c0515b66.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 385, + 72, + 455, + 179 + ], + "blocks": [ + { + "bbox": [ + 385, + 72, + 455, + 179 + ], + "lines": [ + { + "bbox": [ + 385, + 72, + 455, + 179 + ], + "spans": [ + { + "bbox": [ + 385, + 72, + 455, + 179 + ], + "type": "image", + "image_path": "80d72ae204bda3c8d7e809d55187021db01e411f7e1a650ca4d1442bf0d7313d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 458, + 71, + 547, + 178 + ], + "blocks": [ + { + "bbox": [ + 458, + 71, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 458, + 71, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 458, + 71, + 547, + 178 + ], + "type": "image", + "image_path": "cb2722a37a08ac01f2d52e29a8da55f72bde869fd5fb48b908266fb7a65e92a1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 48, + 253, + 291, + 357 + ], + "blocks": [ + { + "bbox": [ + 47, + 220, + 288, + 253 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 288, + 253 + ], + "type": "text", + "content": "Table 1. AFOG effectiveness over 12 detection transformers, measured by mAP on perturbed images. (*) indicates DINO framework used with the corresponding backbone for object detection." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 253, + 291, + 357 + ], + "lines": [ + { + "bbox": [ + 48, + 253, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 291, + 357 + ], + "type": "table", + "html": "
ModelParams (M)BenignAFOGAFOG-VAFOG-F
DETR-R50 [4]39.842.14.14.59.8
DETR-R101 [4]76.043.55.25.111.3
Deform.-DETR [38]40.044.54.81.57.1
R50*[11]47.649.25.31.56.3
AlignDETR [2]47.651.418.11.61.4
ViTDet* [14]108.154.93.80.92.8
ConvNeXt* [21]219.055.43.91.93.1
Swin-L* [20]217.256.87.32.48.6
InternImage* [31]241.056.97.32.85.1
FocalNet* [36]228.958.57.32.55.1
EVA* [10]1037.262.112.24.18.7
DETA [25]218.862.925.63.74.3
", + "image_path": "087d4cbc7490429984f55c533ed5fc29c4e0f1b5202bd94cd9435c31daf23213.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 48, + 413, + 293, + 544 + ], + "blocks": [ + { + "bbox": [ + 46, + 369, + 288, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 288, + 413 + ], + "type": "text", + "content": "Table 2. Comparison benchmark of AFOG against other state-of-the-art object detection attacks on DETR and Swin. Results are theirs. (*) indicates results from [32]. (†) indicates results from [24]. (-) indicates no result." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 413, + 293, + 544 + ], + "lines": [ + { + "bbox": [ + 48, + 413, + 293, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 413, + 293, + 544 + ], + "type": "table", + "html": "
AttackTypePert. BudgetIters.Adversarial mAP
DETR-R50Swin
GARSDC [17]Surrogate0.053000+6.0-
GALD [13]Surrogate0.0631020.6-
RAD* [6]Surrogate0.0631027.247.2
GHFD* [32]Surrogate0.0635012.742.3
UEA* [33]Surrogate0.0635028.550.7
DAG* [35]Surrogate0.0635028.650.7
RAP* [15]Surrogate0.0635024.749.5
EBAD†[3]Victim0.0391034.9-
AttentionFool [22]Victim-10-15021.0-
OATB [12]Victim0.0782026.6-
DBA [16]Victim-50-56.7
AFOGVictim0.031104.17.3
", + "image_path": "8e68bb84be2fc700d77a4e3fe424253063524ab4b074bb2edcbdaa246c0e9609.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 664 + ], + "type": "text", + "content": "Figure 4 shows six visual examples and the comparison of DETR results for benign detection (Row 1) and adversarial detection under the AFOG attack (Row 2) scenarios. The first example shows that DETR detects the person and surfboard correctly under no attack (Row 1) but outputs erroneous detections under the AFOG attack (Row 2). Additional visual examples for all twelve transformers are available in the Supplementary Material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "We report the distortion effect and timing cost for each of the 12 transformer models in Table 3. We note several interesting observations: (i) Considering the " + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": " and SSIM metrics, AFOG attacks induce very similar levels of" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "content": "distortion for 10 out of 12 transformer models. The two largest detection transformers, DETA [25] and EVA [10], show worse " + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "content": " and SSIM scores for AFOG, AFOG-V and AFOG-F. (ii) Considering the " + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "content": " metric, only DETA shows the lowest " + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "content": " value consistently for AFOG, AFOG-V and AFOG-F. (iii) As expected, the time required to attack a detection transformer model is strongly correlated to the number of model parameters. For instance, EVA [10] has over 1 billion parameters, and took much longer on average to attack an input image over 10 iterations compared to the other 11 models. (iii) DETA [25] has 218.8 million parameters and is similar to other models within the Detrex framework: FocalNet [36], InternImage [31], Swin-L [20] and ConvNext [21]. However, we observe that it consistently takes about " + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "inline_equation", + "content": "1.5 - 2 \\times" + }, + { + "bbox": [ + 304, + 224, + 547, + 440 + ], + "type": "text", + "content": " longer time to attack one input image on average with our AFOG attack or its special case extensions AFOG-V or AFOG-F. (iv) AFOG, AFOG-V and AFOG-F have similar average perturbation magnitude." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 453, + 509, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 453, + 509, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 453, + 509, + 466 + ], + "type": "text", + "content": "4.3. Effectiveness on CNN-Based Detectors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": "This section compares AFOG with four attacks designed for CNN-based models: TOG [7], UEA [33], RAP [15], and DAG [35]. They are surrogate-based attacks, and a majority of them (e.g., DAG, RAP, UEA) can only directly attack two-stage detectors (exemplified by Faster-R-CNN [28]) and rely on adversarial transferability to attack single-stage CNN-based detectors like YOLOv3 [27] and SSD [19]. Table 4 reports the comparison results. We make two observations: (i) AFOG and TOG are the only two victim-based attacks against single-stage detectors, represented by YOLOv3 and SSD. Both AFOG and TOG can drastically reduce the benign mAP of YOLOv3 " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "(83.43\\%)" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": " and SSD " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "(76.11\\%)" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": ". For YOLOv3, TOG is a stronger attack, achieving mAP of 0.56 compared to mAP of AFOG being 2.62. However, for SSD-300, AFOG is a stronger attack, achieving mAP of 0.50, compared to mAP of TOG (0.86). (ii) For two-stage CNN-based object detectors, exemplified by Faster-R-CNN, AFOG outperforms all four other victim-based attacks, achieving the lowest adversarial mAP score of " + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 304, + 474, + 546, + 713 + ], + "type": "text", + "content": ". (iii) AFOG also uses the same distortion mag" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3189" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 102, + 561, + 216 + ], + "blocks": [ + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "lines": [ + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "spans": [ + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "text", + "content": "Table 3. Timing and Imperceptibility Results. " + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "text", + "content": " represents the average " + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "text", + "content": " norm difference between perturbed and clean images, " + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "text", + "content": " is the average proportion of perturbed pixels, SSIM is the structural similarity index measure, " + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "inline_equation", + "content": "\\mu_{\\Delta}" + }, + { + "bbox": [ + 46, + 69, + 545, + 102 + ], + "type": "text", + "content": " is the average perturbation magnitude, and t is average total attack time for all ten iterations in seconds." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 102, + 561, + 216 + ], + "lines": [ + { + "bbox": [ + 47, + 102, + 561, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 102, + 561, + 216 + ], + "type": "table", + "html": "
ModelL2AFOGtimeL2AFOG-VtimeL2AFOG-Ftime
L0SSIMμΔL0SSIMμΔL0SSIMμΔ
DETR-R50 [4]0.03220.97070.87150.01731.450.03230.97070.87160.01720.990.03230.97100.87170.01721.21
DETR-R101 [4]0.03230.97070.87210.01731.470.03230.97060.87240.0131.160.03230.97080.87240.01721.70
Deform.-DETR [38]0.03230.97190.87110.01741.630.03230.97130.87160.01731.630.00120.97140.87170.01731.87
R50 [11]0.03170.96580.83430.01712.700.03170.96500.83480.01702.340.03170.96540.83460.01703.16
AlignDETR [2]0.03190.96570.83470.01702.410.03190.96460.83490.01702.330.03190.96470.83490.01703.29
ViTDet [14]0.03180.96710.83530.01716.880.03180.96570.83610.01706.670.03180.96640.83550.01717.33
ConvNext [21]0.03180.96660.83420.01715.380.03180.96540.83490.01705.260.03180.96630.83470.01705.86
Swin-L [20]0.03270.97240.86730.01757.130.03270.97160.86800.01737.280.03270.97220.86780.01748.99
InternImage [31]0.03180.96650.83600.01706.350.03180.96530.83670.01706.230.03180.96600.83640.01716.70
FocalNet [36]0.03200.96650.83650.01728.960.03200.96570.83780.01718.840.03200.96590.83740.01719.74
EVA [10]0.03700.96660.82400.017254.340.03700.96650.82460.017154.530.03700.96650.82420.017151.18
DETA [25]0.04810.96620.80960.017013.200.04810.96540.80990.017013.100.04810.96540.80990.017013.13
", + "image_path": "a5e49bc014f5ec7b7f977976f041af48c1a612e388d53b5fb5e7cf54d8ceb511.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "text", + "content": "nitude budget " + }, + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "text", + "content": " as TOG, while achieving higher attack success rate (lower mAP) on SSD-300 and FRCNN with decreased " + }, + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 46, + 236, + 287, + 297 + ], + "type": "text", + "content": " distortion cost. Hence, AFOG excels at attacking both transformer-based detectors and CNN-based detectors." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 335, + 281, + 472 + ], + "blocks": [ + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 335 + ], + "type": "text", + "content": "Table 4. Comparing AFOG with four state-of-the-art attacks on representative CNN-based object detectors. mAPs of existing attacks were taken from respective papers [7]. (-) indicates N/A." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 335, + 281, + 472 + ], + "lines": [ + { + "bbox": [ + 52, + 335, + 281, + 472 + ], + "spans": [ + { + "bbox": [ + 52, + 335, + 281, + 472 + ], + "type": "table", + "html": "
ModelAttackmAPtDistortion Cost
\\(L_{\\infty}\\)\\(L_{2}\\)\\(L_{0}\\)SSIM
YOLOv3Benign83.430.00.00.00.01.0
TOG [7]0.560.980.0310.0830.9840.875
AFOG2.281.310.0310.0130.8550.801
SSD-300Benign76.110.00.00.00.01.0
UEA [33]20.0-----
DAG [35]64.0-----
TOG [7]0.860.390.0310.1200.9750.879
AFOG0.500.490.0310.0220.8580.793
FRCNNBenign67.370.00.00.00.01.0
UEA [33]5.00.170.3430.1910.9590.652
RAP [15]4.784.040.0820.0100.5310.994
DAG [35]3.567.990.0240.0020.4930.999
TOG [7]2.641.680.0310.0580.9760.862
AFOG2.382.110.0310.0190.8540.788
", + "image_path": "a251e0f43cdc67ee5f4dd8d9d9225472c2684158ca8b1e44d58e013259221350.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 491, + 230, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 230, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 230, + 504 + ], + "type": "text", + "content": "4.4. Effect of Learnable Self-Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "content": "We isolate and analyze the role of the learnable self-attention mechanism in empowering AFOG to successfully corrupt transformer-based models by probing for weak spots and performing fast and human-imperceptible perturbations. Figure 5 shows a comparison between AFOG with and without its learnable attention mechanism, measured by percent difference in adversarial mAP scores. Ablation results show that our learnable attention mechanism improves performance by up to " + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "inline_equation", + "content": "30.6\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "content": " on InternImage, with an average of " + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 46, + 510, + 287, + 629 + ], + "type": "text", + "content": " improvement across all models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Figure 6 shows the self-attention map for the last encoder layer of DETR [4] at initialization and three progressive attack iterations " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "(3^{rd}, 4^{th}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "8^{th}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "). We observe that, as the attack progresses in iterations (indicated by grey arrows), the self-attention map becomes increasingly less associated with the benign detected objects. By the " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "8^{th}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " iteration of the AFOG attack, the self-attention maps have become almost" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "content": "entirely disjoint from the benign detections of the victim model (DETR) prior to the attack. We note that, when the model's self-attention map that reflects learned detection knowledge is changed, it is unlikely for prior knowledge to be kept intact. During the AFOG attack, the perturbed inputs focused by AFOG's own adversarial attention corrupt the pixel-level relationships learned by a well-trained victim model prior to adversarial injection (e.g., " + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 304, + 236, + 547, + 344 + ], + "type": "text", + "content": "), causing catastrophic forgetting to occur." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 312, + 356, + 541, + 495 + ], + "blocks": [ + { + "bbox": [ + 312, + 356, + 541, + 495 + ], + "lines": [ + { + "bbox": [ + 312, + 356, + 541, + 495 + ], + "spans": [ + { + "bbox": [ + 312, + 356, + 541, + 495 + ], + "type": "image", + "image_path": "2dd6e1409fa5c9c44f15609da7c034edba9500a9ca8b2c390a724a9e7a2be09c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "type": "text", + "content": "Figure 5. Improvement (\\% difference) of AFOG with attention over AFOG without attention. The learnable attention mechanism improves performance by as much as " + }, + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "type": "inline_equation", + "content": "30.6\\%" + }, + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "type": "text", + "content": " on InternImage, with an average improvement of " + }, + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 304, + 502, + 547, + 547 + ], + "type": "text", + "content": " across all models." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 553, + 425, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 553, + 425, + 567 + ], + "spans": [ + { + "bbox": [ + 305, + 553, + 425, + 567 + ], + "type": "text", + "content": "4.5. Worst Case Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "Our experimental results on 12 transformer-based detectors and 3 representative CNN-based detectors have shown that AFOG is a fast and effective attack. However, there are still cases where AFOG fails in its adversarial perturbations within a total of 10 iterations, one setting of our attack termination hyper-parameter. By investigating those cases where AFOG did not succeed, we found that the most common situation is where the attack fails to corrupt the central subject of the victim input image, e.g., a person. Figure 7 gives three visual examples, with the top one (Rows 1-2) being a successful attack followed by two failed cases (Rows 3-6) when attacking DETR-R50. Columns 2-3 in all six" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3190" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 287, + 195 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 287, + 195 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 287, + 195 + ], + "type": "image", + "image_path": "cb4ab033a6219e383c0db53918950e3a520a5676b12164acadb535148d5965a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "type": "text", + "content": "Figure 6. DETR self-attention weights analysis. Each blue frame shows the self-attention weights of DETR-R50's last encoder layer at the red indicated point in the image at a given attack iteration " + }, + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "type": "text", + "content": ". In early iterations, the model understands a strong association between the indicated red point on the elephant and the rest of the elephant shape. As the attack iterates (grey arrows), this association becomes corrupted, evidenced by the lack of encoder self-attention structure in later stages. By the " + }, + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "type": "inline_equation", + "content": "8^{th}" + }, + { + "bbox": [ + 46, + 198, + 287, + 308 + ], + "type": "text", + "content": " iteration, the selected point has lost association with the rest of the elephant object, demonstrating that AFOG has disrupted learned knowledge." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 287, + 396 + ], + "type": "text", + "content": "rows show DETR's encoder's last layer's self-attention map for the first and second locations indicated by red points in the first column. Rows 1, 3, and 5 show benign situations and their corresponding self-attention maps, and Rows 2, 4, 6 show the self-attention maps after the final attack iteration. The fourth column shows AFOG's adversarial attention map after the final attack iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 624 + ], + "type": "text", + "content": "We make three observations: (i) For the successful attack case (Rows 1-2), the self-attention maps of victim model DETR-R50 lose association between the indicated red point and the rest of the foreground object it comprises. (ii) In both unsuccessful cases (Rows 3-6), AFOG fails to disrupt this association. (iii) Consider AFOG's adversarial attention maps from the successful case in Figure 7 (Row 2, Col. 4), we observe a clear focus on the key objects in the victim input image. In comparison, AFOG's adversarial attention maps for the two failure cases (Row 4, Col. 4 and Row 6, Col. 4) both fail to focus on foreground objects. The first failure case (Rows 3 and 4) shows that the AFOG attention map attributes significant weight to the distant person in the left of the victim image, resulting in a large number of fabricated predictions in that region. Similarly, the attention map for the second victim image (Rows 5 and 6) appears to entirely miss the fire hydrant, resulting in a failure to disrupt this object. In these cases, AFOG's adversarial attention has learned to focus on the wrong pixels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 119, + 646 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "We have presented AFOG, an Attention-Focused Offensive Gradient attack. AFOG is effective for attacking both advanced object detection transformers and traditional CNN-based detectors with a unified, architecture-agnostic framework. AFOG utilizes a learnable attention mecha" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 68, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 308, + 68, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 308, + 68, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 68, + 545, + 327 + ], + "type": "image", + "image_path": "bf13afda4d23d1131fe77bb746aa8a22bf3a95849100c969f46f57648fbc60e4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 329, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 472 + ], + "type": "text", + "content": "Figure 7. Worst case analysis with DETR-R50 as victim model. Rows 1-2 show a benign image and the successful AFOG perturbed image. Rows 3-6 show two cases where AFOG attack did not succeed. Columns 2-3 show DETR's encoder's self-attention map for the two locations marked in red in column one. Column 4 shows AFOG's internal attention map, where white pixels indicate high attention and blue indicates low. In the successful case (Rows 1-2), encoder self-attention maps (Cols. 2-3) lose associative structure. In the unsuccessful cases (Rows 4 and 6), AFOG fails to disrupt this structure and predictions are preserved. AFOG adversarial attention maps (Col. 4) show focus on important object structure in the successful case (Row 2) and lack of focus in the unsuccessful cases (Rows. 4, 6)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 486, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 666 + ], + "type": "text", + "content": "nism to enable its adversarial perturbations to focus on vulnerable areas of images in multi-box detection tasks. AFOG's attack loss function integrates multiple feature losses (e.g., bounding-box loss, class loss) through learnable attention updates with iterative injection of adversarial perturbations. Finally, AFOG is efficient and stealthy. Adversarial perturbations generated by AFOG are visually imperceptible and yet can cause well-trained detectors to fail miserably. Extensive experiments on state-of-the-art object detectors show that AFOG is consistently effective across twelve object detection transformers. Comparative evaluation with nearly a dozen SOTA methods shows that AFOG significantly outperforms surrogate-based and victim-based attacks on both object detection transformers and CNN-based object detectors." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement. This research is partially sponsored by the NSF CISE grants 2302720 and 2312758, CISCO Edge AI program, and GTRI PhD Fellowship." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3191" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 233 + ], + "type": "text", + "content": "[1] Martin Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dandelion Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 235, + 287, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 235, + 287, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 235, + 287, + 267 + ], + "type": "text", + "content": "[2] Zhi Cai, Songtao Liu, Guodong Wang, Zheng Ge, Xiangyu Zhang, and Di Huang. Align-detr: Improving detr with simple iou-aware bce loss, 2023. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 268, + 287, + 312 + ], + "type": "text", + "content": "[3]Zikui Cai, Yaoteng Tan, and M. Salman Asif. Ensemble-based blackbox attacks on dense prediction. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4045-4055, 2023. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 366 + ], + "type": "text", + "content": "[4] Nicolas Carrion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-223, 2020. 1, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 368, + 287, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 368, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 53, + 368, + 287, + 400 + ], + "type": "text", + "content": "[5] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 434 + ], + "type": "text", + "content": "[6] Sizhe Chen, Fan He, Xiaolin Huang, and Kun Zhang. Relevance attack on detectors. Pattern Recognition, 124:108491, 2022. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 435, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 501 + ], + "type": "text", + "content": "[7] Ka-Ho Chow, Ling Liu, Margaret Loper, Juhyun Bae, Mehmet Emre Gursoy, Stacey Truex, Wenqi Wei, and Yanzhao Wu. Adversarial objectness gradient attacks in real-time object detection systems. In IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications, pages 263-272. IEEE, 2020. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 502, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 502, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 502, + 287, + 533 + ], + "type": "text", + "content": "[8] Papers With Code. Object detection on coco test-dev. https://paperswithcode.com/sota/object-detection-on-coco.1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "type": "text", + "content": "[9] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John Winn, and Andrew Zisserman. The Pascal Visual Object Classes (VOC) Challenge. International Journal of Computer Vision, 88(2):303-338, June 2010. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "text", + "content": "[10] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. arXiv preprint arXiv:2211.07636, 2022.5,6,7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Shu. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition, pages 770-778, 06 2016. 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 714 + ], + "type": "text", + "content": "[12] Zhuo Leng, Zesen Cheng, Pengxu Wei, and Jie Chen. Object-aware transfer-based black-box adversarial attack on object detector. In Qingshan Liu, Hanzi Wang, Zhanyu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "Ma, Weishi Zheng, Hongbin Zha, Xilin Chen, Liang Wang, and Rongrong Ji, editors, Pattern Recognition and Computer Vision, pages 278-289, Singapore, 2024. Springer Nature Singapore. 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 150 + ], + "type": "text", + "content": "[13] Tuo Li and Yahong Han. Improving transferable adversarial attack for vision transformers via global attention and local drop. Multimedia Systems, 29:3467 - 3480, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 545, + 183 + ], + "type": "text", + "content": "[14] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection, 2022. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "text", + "content": "[15] Yuezun Li, Daniel Tian, Mingching Chang, Xiao Bian, and Siwei Lyu. Robust adversarial perturbation on deep proposal-based models. In BMVC, 2018. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 259 + ], + "type": "text", + "content": "[16] Jiawei Lian, Shaohui Mei, Xiaofei Wang, Yi Wang, Lefan Wang, Yingjie Lu, Mingyang Ma, and Lap-Pui Chau. Attack anything: Blind dnns via universal background adversarial attack, 2024. 2, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[17] Siyuan Liang, Longkang Li, Yanbo Fan, and Xiaojun Jia. A large-scale multiple-objective method for black-box attack against object detection. European Conference on Computer Vision (ECCV) 2022, pages 619-636, 2022. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 305, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 371 + ], + "type": "text", + "content": "[18] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, ECCV 2014, pages 740-755. Springer International Publishing, 2014. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 437 + ], + "type": "text", + "content": "[19] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. Ssd: Single shot multibox detector. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors, Computer Vision - ECCV 2016, pages 21-37, Cham, 2016. Springer International Publishing. 1, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 491 + ], + "type": "text", + "content": "[20] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 525 + ], + "type": "text", + "content": "[21] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 525, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 592 + ], + "type": "text", + "content": "[22] Giulio Lovisotto, Nicole Finnie, Maurizio Munoz, Chaithanya Kumar Murnmadi, and Jan Hendrik Metzen. Give me your attention: Dot-product attention considered harmful for adversarial patch robustness. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15213-15222, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[23] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[24] Khoi Nguyen Tiet Nguyen, Wenyu Zhang, Kangkang Lu, Yuhuan Wu, Xingjian Zheng, Hui Li Tan, and Liangli Zhen. A Survey and Evaluation of Adversarial Attacks for Object Detection, Aug. 2024. arXiv:2408.01934 [cs]. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[25] Jeffrey Ouyang-Zhang, Jang Hyun Cho, Xingyi Zhou, and Philipp Krahenbihl. Nms strikes back. arXiv preprint arXiv:2212.06137, 2022. 6, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3192" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 193 + ], + "type": "text", + "content": "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 194, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 194, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 194, + 287, + 216 + ], + "type": "text", + "content": "[27] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 250 + ], + "type": "text", + "content": "[28] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 250, + 287, + 304 + ], + "type": "text", + "content": "[29] Tianhe Ren, Shilong Liu, Feng Li, Hao Zhang, Ailing Zeng, Jie Yang, Xingyu Liao, Ding Jia, Hongyang Li, He Cao, Jianan Wang, Zhaoyang Zeng, Xianbiao Qi, Yuhui Yuan, Jianwei Yang, and Lei Zhang. detrex: Benchmarking detection transformers, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 304, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 304, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 49, + 304, + 287, + 380 + ], + "type": "text", + "content": "[30] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 382, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 382, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 49, + 382, + 287, + 437 + ], + "type": "text", + "content": "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. arXiv preprint arXiv:2211.05778, 2022. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 437, + 287, + 491 + ], + "type": "text", + "content": "[32] Yang Wang, Yunfei Zheng, Lei Chen, Zhen Yang, Jingwei Wu, and Tieyong Cao. Gradient-guided hierarchical feature attack for object detector. Journal of King Saud University - Computer and Information Sciences, 36(1):101901, 2024. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 492, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 492, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 49, + 492, + 287, + 557 + ], + "type": "text", + "content": "[33] Xingxing Wei, Siyuan Liang, Ning Chen, and Xiaochun Cao. Transferable adversarial attacks for image and video object detection. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19, pages 954–960. International Joint Conferences on Artificial Intelligence Organization, 7 2019. 2, 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 559, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 590 + ], + "type": "text", + "content": "[34] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 49, + 592, + 287, + 645 + ], + "type": "text", + "content": "[35] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie1, and Alan Yuill. Adversarial examples for semantic segmentation and object detection. In International Conference on Computer Vision, pages 1378-1387, 2017. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 647, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 647, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 647, + 287, + 668 + ], + "type": "text", + "content": "[36] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 712 + ], + "type": "text", + "content": "[37] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection, 2022. 5" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[38] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 5, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3193" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_content_list.json b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..700bf9fdf4f1c07c0c2449ac79f1d06b68676238 --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_content_list.json @@ -0,0 +1,1713 @@ +[ + { + "type": "text", + "text": "Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization", + "text_level": 1, + "bbox": [ + 94, + 128, + 901, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuyu Zhang $^{1,2,3}$ Ning Chen $^{1,3}$ Yongshan Liu $^{4}$ Qinghua Zhang $^{2,3}$ Xu Zhang $^{1,2,3,*}$", + "bbox": [ + 143, + 202, + 857, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Artificial Intelligence, Chongqing University of Posts and Telecommunications", + "bbox": [ + 228, + 223, + 767, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications", + "bbox": [ + 171, + 239, + 823, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism, China", + "bbox": [ + 163, + 258, + 831, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Mathematics and Statistics, Chongqing University of Posts and Telecommunications", + "bbox": [ + 210, + 276, + 781, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to $9.47\\%$ improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks.", + "bbox": [ + 89, + 358, + 485, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 719, + 220, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks have shown remarkable progress across various tasks in standard supervised learning settings where the training and test data come from the same distribution [11, 18]. However, their performance often suffers significantly when encountering domain shift, a prevalent issue in real-world applications [23, 27, 50]. For instance, a model trained on DSLR images may struggle when tested on smartphone images due to differences in reso", + "bbox": [ + 89, + 744, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lution, noise, and other factors. These domain shifts are pervasive in real-world scenarios, emphasizing the need for models that can generalize well to unseen distributions.", + "bbox": [ + 511, + 328, + 906, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single Domain Generalization (SDG) methods aim to address these domain shifts when only a single source domain is available for training. Current SDG approaches primarily employ data augmentation techniques to generate diverse training samples [16, 25, 33, 37, 40, 47]. Methods such as adversarial data augmentation [25, 33, 47] generate perturbed samples that simulate potential domain shifts. Despite their contributions, as shown in Figure 1, these adversarial data augmentation approaches often fail to adequately capture the global structure of domain shifts in the parameter space due to their emphasis on localized perturbations. As a result, they may underexplore the parameter space, limiting generalization across diverse distributions.", + "bbox": [ + 511, + 377, + 908, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we draw inspiration from dynamical systems theory [1, 6] and propose a novel perspective of model perturbations based on the edge of chaos. This critical transition point maintains a balance between order and chaos in dynamical systems. In SDG, training models near the edge of chaos can help capture features that better generalize across domains by maintaining both stability and adaptability [44]. To quantitatively measure a model's degree of chaos, we employ the Lyapunov Exponent (LE) [10] as a metric to estimate the model's proximity to this edge of chaos. A high LE indicates chaotic behavior and sensitivity to perturbations, whereas a low LE suggests stable dynamics that may restrict the model's generalization capability. We frame the training of neural networks as a discrete-time dynamical system in the parameter space, where each parameter update represents a state transition in this system. By dynamically adjusting the LE, we aim to guide adversarial data augmentation in a way that encourages broader exploration of the parameter space, balancing stability and adaptability for better domain generalization.", + "bbox": [ + 511, + 579, + 908, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building on this insight, we introduce LE-Aware", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*The corresponding author, email: zhangx@cqupt.edu.cn", + "bbox": [ + 106, + 875, + 413, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The code is available at: https://github.com/ChannySwift/LEAwareSGD.", + "bbox": [ + 106, + 887, + 485, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "552", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12c7c405a8ef6d5dbd4c0792897b579e0750d9ebfe0f6525b14c5521c5dea037.jpg", + "image_caption": [ + "(a) ADA" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 281, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d9c07935d20914b168c0256d8aadc6d5a960b59ee988473d7ff6bba8a4a3aede.jpg", + "image_caption": [ + "(b) ME-ADA" + ], + "image_footnote": [], + "bbox": [ + 292, + 88, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8184d81b360efdb6cd29cd76aa754439114d749be2add7363d09ed663edc8c55.jpg", + "image_caption": [ + "(c) AdvST", + "Figure 1. t-SNE visualization of parameter trajectories in initial epochs for a model trained on the PACS dataset [15]. Each point represents the model's parameters after an iteration. Comparison between (a) ADA [33], (b) ME-ADA [47], (c) AdvST [48], and (d) our method demonstrates that our approach can explore a broader parameter space, facilitating the learning of more generalizable features across unseen domains." + ], + "image_footnote": [], + "bbox": [ + 93, + 260, + 279, + 393 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a690f1defb557c6b73500c326de9c3f45cb8f21c95fafe864469a82448cd7f1.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 292, + 261, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stochastic Gradient Descent (LEAwareSGD), a novel optimizer that dynamically adjusts the learning rate based on LE computations. By integrating adversarial data augmentation with LE-aware optimization, LEAwareSGD generates diverse perturbations that enhance the model's ability to generalize to unseen domains while guiding it to operate near the edge of chaos. Unlike conventional optimizers [14, 20, 26, 30] without considering the system's state, LEAwareSGD leverages LE as a feedback mechanism, adjusting the learning rate in response to the system's state and enabling the model to explore the parameter space effectively for capturing more generalizable features.", + "bbox": [ + 89, + 550, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 733, + 413, + 747 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conceptualize the training of neural networks as a discrete-time dynamical system and leverage the LE to assess the system's proximity to the edge of chaos.", + "- We propose LEAwareSGD, an optimizer for adversarial data augmentation that dynamically adjusts the learning rate based on LE, guiding the training process near the edge of chaos to enhance generalization.", + "- Extensive experiments on three standard SDG benchmarks, including PACS [15], OfficeHome [31], and DomainNet [24], validate the effectiveness of the proposed" + ], + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LEAwareSGD to improve cross-domain generalization.", + "bbox": [ + 527, + 90, + 890, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 122, + 653, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 148, + 779, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single domain generalization seeks to improve model robustness by enabling generalization from a single source domain to unseen target domains, eliminating the need for target data during training. Most SDG methods rely on data augmentation [25, 43] to increase diversity within the source domain, while generative model-based approaches [16, 37, 40] synthesize samples to extend the source distribution. Although these generative techniques introduce sample variety, they are often computationally demanding and complex to train, frequently producing samples that lack sufficient quality and diversity. This limitation can cause models to overfit to synthetic data rather than enhancing generalization to real-world domains. Adversarial data augmentation [2, 47, 48] offers an alternative by generating adversarial examples that mimic domain shifts to improve model robustness. However, these examples often introduce only localized perturbations, failing to explore the parameter space extensively. In contrast, our proposed LE-guided optimization introduces a dynamic SDG framework that modulates model sensitivity to perturbations. This LE-aware optimization promotes broader exploration of the parameter space, effectively capturing complex domain shifts and improving cross-domain generalization capability.", + "bbox": [ + 511, + 171, + 906, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Optimizers in Domain Generalization", + "text_level": 1, + "bbox": [ + 511, + 530, + 836, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimizers play a critical role in model generalization, with different strategies yielding varied outcomes depending on the loss landscape's properties. Widely-used optimizers like Stochastic Gradient Descent (SGD) [22] and Adam [14] are commonly used in deep learning tasks, yet they exhibit different behaviors in terms of generalization. Some prior works [9, 46] show that Adam often results in worse generalization than SGD due to its susceptibility to sharp minima. In contrast, SGD tends to find flatter minima more conducive to robust generalization. This distinction highlights the importance of the loss surface's geometry, as flatter minima are generally associated with improved generalization performance, especially in out-of-distribution settings. In light of this, recent optimization strategies have been explicitly designed to address the sharpness of minima. Sharpness-Aware Minimization (SAM) [7] represents a key development in this direction, where the objective is to minimize sharpness and seek flatter regions of the loss landscape. SAM and its variants [46, 51] have demonstrated notable improvements in generalization, particularly in distribution. Recent improvements such as GSAM [51], which introduces a surrogate gap to better distinguish between sharp and flat minima, further enhance generalization", + "bbox": [ + 511, + 553, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "553", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "by guiding the model toward regions with both low sharpness and low loss. Our proposed LEAwareSGD differs from these optimizers by dynamically adjusting the learning rate based on the LE, allowing the model toward the edge of chaos and exploring a broader parameter space for improving generalization capability.", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Dynamical Systems in Deep Learning", + "text_level": 1, + "bbox": [ + 89, + 190, + 415, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The application of dynamical systems theory to deep learning has introduced new perspectives on neural network training, particularly in understanding stability, and generalization [19, 29]. By modeling neural networks as discrete-time systems, researchers have used concepts such as chaos theory to enhance generalization, especially within residual networks via flow maps [17]. This perspective underscores the significance of training near the edge of chaos, a state that balances stability and adaptability [44, 45]. Lyapunov exponent (LE) is a key tool to quantitatively measure the model's training stability, particularly in recurrent architectures [8, 32]. However, existing methods have not fully leveraged LE for direct control of training under domain shifts, particularly in the context of adversarial data augmentation for domain generalization tasks. Our proposed LEAwareSGD addresses this limitation by using LE to dynamically guide the optimization process, effectively enhancing model generalization capability.", + "bbox": [ + 89, + 212, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 496, + 225, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview of LEAwareSGD", + "text_level": 1, + "bbox": [ + 89, + 521, + 334, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose LEAwareSGD, a novel optimization method for adversarial data augmentation to enhance generalization by exploiting the edge of chaos. In this approach, the LE serves as a metric to identify and explore regions in the parameter space where generalizable features are more likely to emerge.", + "bbox": [ + 89, + 542, + 483, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. LE-Based Model Perturbation", + "text_level": 1, + "bbox": [ + 89, + 642, + 361, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The LE quantifies the rate at which small perturbations in model parameters evolve during training, offering insights into the stability of the optimization process. Here, we denote the model parameters $\\theta_{t}$ at iteration $t$ and the standard gradient descent update can be formulated as follows:", + "bbox": [ + 89, + 664, + 483, + 739 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {t + 1} = \\theta_ {t} - \\eta_ {t} \\nabla L (\\theta_ {t}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 748, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta_t$ denotes the learning rate and $L(\\theta_t)$ represents the loss function.", + "bbox": [ + 89, + 768, + 482, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To quantify how perturbations evolve during training, we introduce an initial perturbation $\\delta \\theta_0$ to the model parameters. Parameters after perturbation is $\\widetilde{\\theta}_t\\coloneqq \\theta_t + \\delta \\theta_t$ , where $\\delta$ is a tiny real-valued scalar hyperparameter. The perturbation at step $t + 1$ propagates according to the following formula:", + "bbox": [ + 89, + 800, + 483, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\theta} _ {t + 1} = \\widetilde {\\theta} _ {t} - \\eta_ {t} \\nabla L (\\widetilde {\\theta} _ {t}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 882, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Combining Equation 1 and Equation 2, we get the following update for the perturbation:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\theta_ {t + 1} = \\delta \\theta_ {t} - \\eta_ {t} (\\nabla L (\\widetilde {\\theta_ {t}}) - \\nabla L (\\theta_ {t})). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 126, + 903, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further analyze the propagation of this perturbation, we perform a first-order Taylor expansion of the loss function $L(\\widetilde{\\theta_t})$ at $\\theta_{t}$ and calculate the gradient:", + "bbox": [ + 511, + 157, + 906, + 204 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla L \\left(\\widetilde {\\theta} _ {t}\\right) = \\nabla L \\left(\\theta_ {t}\\right) + H \\left[ L \\left(\\theta_ {t}\\right) \\right] \\delta \\theta_ {t} + o \\left(\\| \\delta \\theta_ {t} \\| ^ {2}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 223, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $H[L(\\theta_t)]$ is the Hessian matrix, and $o(\\| \\delta \\theta_t\\|^2)$ represents higher-order terms. Ignoring these higher-order terms, and substituting Equation 4 into Equation 3, we approximate the evolution of the perturbation as:", + "bbox": [ + 511, + 253, + 903, + 315 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\delta \\theta_ {t + 1} = \\left(I - \\eta_ {t} H \\left[ L \\left(\\theta_ {t}\\right) \\right]\\right) \\delta \\theta_ {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 324, + 903, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By recursively applying this relationship, we obtain the propagation of the perturbation over time:", + "bbox": [ + 511, + 348, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\delta \\theta_ {t} = \\left(I - \\eta_ {t - 1} H \\left[ L \\left(\\theta_ {t - 1}\\right) \\right]\\right) \\left(I - \\eta_ {t - 2} H \\left[ L \\left(\\theta_ {t - 2}\\right) \\right]\\right) \\tag {6} \\\\ \\cdot \\cdot \\cdot (I - \\eta_ {0} H [ L (\\theta_ {0}) ]) \\delta \\theta_ {0}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 396, + 903, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The definition of the Lyapunov exponent is:", + "bbox": [ + 532, + 444, + 823, + 459 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L E} = \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\ln \\left(\\frac {\\| \\delta \\theta_ {t} \\|}{\\| \\delta \\theta_ {0} \\|}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 470, + 903, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By substituting Equation 6 into the definition of LE, we can derive the connection between the model's learning rate $\\eta$ and LE:", + "bbox": [ + 511, + 513, + 903, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L E} \\geq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {\\substack {i = 0\\\\t - 1}} ^ {t - 1} \\ln (1 - \\eta_ {i} \\| H [ L (\\theta_ {i}) ] \\|) \\tag{8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 566, + 903, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L E} \\leq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {i = 0} ^ {t - 1} \\ln \\left(\\| I - \\eta_ {i} H [ L (\\theta_ {i}) ] \\|\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 611, + 846, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the LE is determined by the learning rate $\\eta_{i}$ and the Hessian matrix $H[L(\\theta_i)]$", + "bbox": [ + 511, + 656, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. LE-Guided Optimization", + "text_level": 1, + "bbox": [ + 513, + 696, + 743, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A positive LE indicates the exponential growth of perturbations, signifying instability, while a negative LE suggests convergence and stability. However, extreme stability or instability can lead to either overfitting or convergence difficulties. Our approach aims to guide model parameters toward an LE value close to zero but slightly negative, positioning the system near the edge of chaos. This critical state balances stability with adaptability, enhancing the model's generalization capacity. To achieve this, we dynamically adjust the learning rate to amplify perturbations, thereby steering the LE toward closer to the edge of chaos. Specifically, the innovation of the proposed LEAwareSGD lies in", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "554", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "its learning rate adjustment based on LE dynamics, allowing the model to adapt effectively across training phases. We modulate the learning rate according to changes in the LE, $\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}$ .", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\eta_ {t + 1} = \\eta_ {t} \\cdot \\exp (- \\beta \\cdot \\Delta \\mathrm {L E} _ {t}) \\quad \\text {i f} \\quad \\Delta \\mathrm {L E} _ {t} > 0 \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 167, + 482, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is a weighting parameter that controls the sensitivity of the learning rate adjustment. When $\\Delta \\mathrm{LE}_t > 0$ , indicating that the model is approaching regions near the edge of chaos where generalizable features are more likely to emerge, we reduce the learning rate to allow for deeper exploration of these regions. This approach ensures that the model explores parameter regions with strong generalization potential while avoiding over-stabilization that could hinder the learning process. When $\\Delta \\mathrm{LE}_t \\leq 0$ , we leave the learning rate unchanged, maintaining the current exploration rate.", + "bbox": [ + 89, + 198, + 483, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. LE-Based Adversarial Data Augmentation", + "text_level": 1, + "bbox": [ + 89, + 375, + 452, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further enhance the model's robustness to domain shifts, we integrate LEAwareSGD with adversarial data augmentation. Adversarial data augmentation generates samples that simulate domain shifts, exposing the model to a broader range of data variations. Given a model $f_{\\theta}$ with parameters $\\theta$ , we define adversarial samples as transformations of the original data. Let $\\tau(x; \\omega)$ denote a set of semantic transformations parameterized by $\\omega$ . The joint optimization objective is:", + "bbox": [ + 89, + 397, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\theta} \\max _ {\\omega} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {S}} \\left[ \\ell (\\theta ; \\tau (x; \\omega), y) - \\lambda d _ {\\theta} (\\tau (x; \\omega), x) \\right] \\\\ + \\frac {\\gamma}{2} \\cdot \\| \\theta \\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 546, + 480, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ell (\\theta ;\\cdot)$ is the prediction loss, $\\lambda$ balances the adversarial loss and feature consistency, and $d_{\\theta}(\\cdot)$ represents the feature distance between the original and transformed samples. Here, the regularization term $\\frac{\\gamma}{2}\\cdot \\| \\theta \\| _2^2$ denotes a weight decay, ensuring that the Hessian matrix remains approximately positive definite. This regularization encourages LE to be negative, facilitating the stability of model training. Note that this adversarial training procedure is guided by LE-based optimization, which encourages the model to generalize to unseen domains by learning robust features invariant to domain-specific perturbations.", + "bbox": [ + 89, + 609, + 483, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Training Procedure", + "text_level": 1, + "bbox": [ + 89, + 787, + 277, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training procedure involves iterative steps alternating between adversarial example generation and parameter updates using LEAwareSGD. In each epoch, adversarial examples are first generated to simulate unseen domain shifts, fostering the model to learn invariant features. Subsequently, LEAwareSGD adjusts the model's parameters with", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a LE-guided learning rate to encourage the parameters' updating towards the edge of chaos. This LE-guided adaptation encourages the exploration of parameter regions conducive to generalization while controlling sensitivity to perturbations. The complete algorithm is shown in Algorithm 1. We also provide a detailed theoretical analysis of the proposed LEAwareSGD in Suppl. Sec-A.1.", + "bbox": [ + 511, + 90, + 906, + 198 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 LEAwareSGD" + ], + "code_body": "1: Input: Initial parameters $\\theta_0$ , learning rate $\\eta_0$ , weighting parameter $\\beta$ , number of iterations $N$ \n2: Output: Optimized parameters $\\theta^*$ \n3: for $t = 0, 1, 2, \\dots, N - 1$ do \n4: Compute gradient $\\nabla_{\\theta} L(\\theta_t)$ \n5: Update parameters: $\\theta_{t+1} = \\theta_t - \\eta_t \\nabla_{\\theta} L(\\theta_t)$ \n6: Compute updated model perturbation using Eq. 6 \n7: Calculate LE using Eq. 7 \n8: Update $\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}$ \n9: if $\\Delta \\mathrm{LE}_t > 0$ then \n10: $\\eta_{t+1} = \\eta_t \\cdot \\exp(-\\beta \\cdot \\Delta \\mathrm{LE}_t)$ \n11: else \n12: $\\eta_{t+1} = \\eta_t$ \n13: end if \n14: end for \n15: Return: $\\theta^*$", + "bbox": [ + 521, + 229, + 903, + 473 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 491, + 637, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Datasets and Settings", + "text_level": 1, + "bbox": [ + 511, + 516, + 714, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three widely-used domain generalization benchmarks: PACS [15], OfficeHome [31], and DomainNet [24], arranging them in increasing levels of difficulty to comprehensively test the model's robustness across varying domain shifts. PACS consists of four domains, namely, Photo (P), Art Painting (A), Cartoon (C), and Sketch (S), comprising 9,991 images from 7 categories. OfficeHome includes four domain, namely, Art (A), Clipart (C), Product (P), and Real-World (R), with 15,500 images across 65 classes. Compared to PACS, OfficeHome introduces greater diversity in both object categories and domain styles. DomainNet is the most challenging, containing six domains, namely, Real (R), Infograph (I), Clipart (C), Painting (P), Quickdraw (Q), and Sketch (S), with 596,010 images from 345 classes. The substantial domain shifts and the high number of classes in this dataset present considerable challenges for domain generalization.", + "bbox": [ + 511, + 537, + 906, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. In all experiments, we employed ResNet-18 [11] as the backbone model and resized all input images to $224 \\times 224$ . The network was trained with LEAwareSGD, setting the momentum to 0.9 and $\\beta$ to 1e-1. To ensure a fair comparison, we excluded the test-time learning stage of PSDG[41] in all of our experiments. For PACS and OfficeHome datasets, we follow the standard", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "555", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "protocol of using one domain as the source domain and the remaining three domains as the target domains [47, 48]. For the PACS dataset, we adopted a batch size of 16, $\\gamma$ of 5e-4, and an initial learning rate of 5e-4, training for 50 epochs. For the OfficeHome dataset, we used a batch size of 32, set $\\gamma$ to 1e-4, and initialized the learning rate at 1e-4, training for 50 epochs. For experiments conducted on the DomainNet dataset, we follow [37, 48] and use Real as the source domain and the remaining five domains as the target domains. The batch size was increased to 128, $\\gamma$ was set to 1e-5, and the initial learning rate was set to 1e-3, with a training duration of 200 epochs. For hyperparameter tuning, we fix all training settings (e.g., random seed, learning rate) and tune only $\\beta$ and $\\gamma$ , with the best setting shown in Fig.2. All experiments were performed on an NVIDIA A100 GPU, and each experiment was repeated 10 times with different random seeds. The results are reported as the average accuracy alongside the standard deviation.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with SDG Approaches", + "text_level": 1, + "bbox": [ + 89, + 378, + 397, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on PACS. This adjustment aligns the evaluation setup with our method, which operates solely during the training phase. The evaluation results on the PACS dataset, presented in Table 1, indicate that our method achieves the highest average accuracy of $69.46\\%$ , outperforming all compared methods. Specifically, our model achieves the highest performance gains in both the A and P domains, with classification accuracies of $79.17\\%$ and $65.05\\%$ , respectively. Our approach outperforms the second-best method, PSDG $(67.14\\%)$ , with an improvement of $2.32\\%$ in average accuracy. This result underscores the effectiveness of our LE-guided adversarial data augmentation strategy, which enhances model generalization by facilitating the exploration of a broader parameter space and reducing overfitting to domain-specific features.", + "bbox": [ + 89, + 402, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/71bf92bc005611fb165169dd8eb563c3a6b6287bf1edcd06a12bfe6475c7ef77.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsACPSAvg.
ERM70.4973.5645.9241.2157.80
MixUp [39]67.6067.1351.5947.0458.34
RSC [13]73.4075.9056.2041.6061.78
ADA [33]71.5676.8452.3643.6661.11
ME-ADA [47]71.5276.8346.2246.3260.22
SAGM [35]65.9470.1042.1453.6557.96
RandConv [38]73.6874.8855.4246.7762.69
MetaCNN [34]76.5074.5056.7055.7065.80
L2D [36]77.0076.5553.2254.4465.30
PDEN [16]76.4373.8758.5253.9265.69
PSDG [41]77.8875.2455.4959.9667.14
AdvST [48]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
", + "bbox": [ + 93, + 646, + 480, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on OfficeHome. Table 2 shows the classification accuracy comparison of our method with existing baselines on the OfficeHome dataset. Our proposed method consistently outperforms other approaches, achieving the highest accuracy across all domains (A, C, P, and R) with an average accuracy of $54.38\\%$ . Notably, our method surpasses the previous best method, AdvST, which obtained an average accuracy of $52.60\\%$ , with an improvement of $1.78\\%$ . This gain highlights the effectiveness of our LE-guided optimization, which facilitates the exploration of a broader parameter space and enhances the model's ability to generalize to unseen domains.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/df133d6bd18f9e46f7bd94f7babe5c7363581a3a165c26b5baa29d2b690d0f0a.jpg", + "table_caption": [ + "Table 1. Classification accuracy (%) comparison on the PACS dataset. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsACPRAvg.
ERM42.7439.3139.8652.5043.60
MixUp [39]43.6940.6838.5652.1243.76
SagNet [21]49.8443.0141.4255.6147.47
RandConv [38]45.7940.2037.4552.8744.08
Fish [28]45.1039.7436.6752.0943.40
ADA [33]45.0736.0744.4353.4444.75
ME-ADA [47]45.2745.7938.6551.6945.35
GSAM [51]28.9741.4041.7251.5040.90
SAGM [35]47.4243.2741.3654.9746.76
Mixstyle [49]51.1948.7346.8555.8850.66
PCGrad [42]49.9146.6146.1156.3149.74
PSDG [41]42.8046.4044.3154.6847.05
AdvST [48]51.3252.0848.8958.1152.60
Ours52.8955.1851.6958.9554.38±0.30
", + "bbox": [ + 516, + 282, + 903, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Classification accuracy (%) comparison on the Office-Home dataset. The best results are in bold.", + "bbox": [ + 511, + 517, + 903, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on DomainNet. Table 3 shows the classification accuracy comparison on the DomainNet dataset. Our proposed method achieves the highest average accuracy of $28.15\\%$ , surpassing all other methods, including the second-best method, AdvST, which reaches an average accuracy of $27.22\\%$ . Notably, our method performs best in the C and S domains, with accuracies of $44.67\\%$ and $33.74\\%$ , respectively. In the Q domain, our method achieves an accuracy of $6.70\\%$ , slightly lower than SimDE's $6.85\\%$ . This may be due to the highly abstract nature of the Quickdraw domain, which can benefit from specialized augmentation techniques to better capture its distinct features. Despite this, our method's overall performance advantage demonstrates the effectiveness of the LE-guided adversarial data augmentation approach in handling the diversity and complexity of DomainNet.", + "bbox": [ + 511, + 551, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Comparison with Optimization Approaches", + "text_level": 1, + "bbox": [ + 511, + 801, + 883, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further validate the effectiveness of our proposed LEAwareSGD optimizer, we conducted a comparison with four widely used optimization approaches, including Adam [14], AdamW [20], RMSprop [30], and SGD [26], on the PACS and OfficeHome datasets. As shown in Table 4,", + "bbox": [ + 511, + 824, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "556", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f1aca5b1e029d4df1c443b5561535bc3a7b89e6a18c96785870b2aeae8d30a4b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPICSQAvg.
ERM38.0513.3137.8926.263.3623.77
MixUp [39]38.6013.9238.0126.013.7124.05
CutOut [5]38.3013.7038.4026.203.7024.06
AugMix [12]40.8013.9041.7029.806.3026.50
RandAug [3]41.3013.5741.1130.405.3126.34
ADA [33]38.2013.8040.2024.804.3024.26
ME-ADA [47]38.4813.5640.6626.054.4224.63
ACVC [4]41.3012.9042.8030.906.6026.90
PDEN [16]38.4511.2538.9931.715.5825.20
SimDE [37]39.9612.9141.7333.466.8526.98
PSDG [41]41.1013.0041.2931.104.8926.28
AdvST [48]42.4014.9041.7031.006.1027.22
Ours42.7715.9144.6733.746.7028.15 ±0.61
", + "bbox": [ + 91, + 88, + 496, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our proposed optimizer achieves the highest average accuracy of $69.46\\%$ , surpassing all other optimizers on the PACS dataset. In contrast, widely used optimizers such as Adam, AdamW, and RMSprop perform significantly worse, with average accuracies of $66.43\\%$ , $66.83\\%$ , and $62.34\\%$ , respectively. This disparity likely stems from the reliance of these adaptive optimizers on variable learning rates, which may cause overfitting to domain-specific features and reduce their generalization ability in domain generalization tasks. On the other hand, both our LEAwareSGD and standard SGD achieve substantially higher accuracies. While SGD achieves an average accuracy of $67.06\\%$ , our LEAwareSGD optimizer further improves this to $69.46\\%$ . This enhancement underscores the advantage of integrating LE guidance into the optimization process, which enables the model to better capture generalizable features and avoid overfitting to domain-specific characteristics. We present the experimental results on the OfficeHome dataset in Suppl. Sec-C.2, while analyses of optimizers' performance under different learning rates are provided in Suppl. Sec-C.5.", + "bbox": [ + 88, + 345, + 486, + 662 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c38222a6bfaaa5cdb2d0f5efece7db33d1f0b3e8a6826fd211c6dc602ac1e6a1.jpg", + "table_caption": [ + "Table 3. Classification accuracy (%) comparison on the Domain-Net dataset. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsACPSAvg.
Adam [14]76.5271.1564.0653.9866.43
AdamW [20]76.6871.9362.0356.6866.83
RMSprop [30]71.6271.0859.0947.5562.34
SGD [26]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
", + "bbox": [ + 91, + 675, + 495, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Extension to Multi-Source DG Approaches", + "text_level": 1, + "bbox": [ + 89, + 832, + 457, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further validate the scalability of our solution, we perform experiments under the leave-one-out setting of general domain generalization (DG) with ResNet-18 as the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "backbone, where three domains are combined for training and the remaining domain is reserved for evaluation. Table 5 presents the results on the PACS dataset, showing that our method delivers at least comparable performance, highlighting its versatility. This improvement underscores the effectiveness of our LE-guided optimization, which enables a broader exploration of the parameter space and enhances the model's generalization to unseen domains.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6369fa0bc0b33d0152e80b668a2422ce479479b8099194f773abeb32ee92fbd6.jpg", + "table_caption": [ + "Table 4. Comparison results $(\\%)$ using different optimizers on the PACS dataset. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsACPSAvg.
RSC [13]82.8179.7493.9580.8584.34
ADA [33]82.8178.3395.6375.2983.02
ME-ADA [47]77.8878.5895.3378.0782.47
Mixstyle [49]83.1179.4396.3172.9582.95
L2D [36]81.4479.5695.5180.5884.27
AdvST [48]78.6972.9791.6280.9181.05
Ours81.3578.8895.4582.3984.52
", + "bbox": [ + 532, + 222, + 887, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Classification accuracy (\\%) comparison for the leave-one-domain-out using the ResNet-18 backbone on the PACS dataset. The column name denotes the test domain, with the other three used for training. The best results are in bold.", + "bbox": [ + 511, + 359, + 906, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Additional Analysis", + "text_level": 1, + "bbox": [ + 511, + 425, + 702, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance comparison using different backbone networks. To further validate the universality of our solution, we conduct experiments with different ResNet backbones, including ResNet-34, ResNet-50, ResNet-101, and ResNet-152, on the PACS and OfficeHome datasets. As shown in Table 6 on the PACS dataset, our approach reaches a performance of $73.68\\%$ on ResNet-34, $71.37\\%$ on ResNet-50, $74.32\\%$ on ResNet-101, and $75.34\\%$ on ResNet-152. These results demonstrate the consistent effectiveness of our method across various backbone architectures, highlighting its superior accuracy, versatility, and robustness. The experimental results for the OfficeHome dataset are presented in Suppl. Sec-C.6.", + "bbox": [ + 511, + 446, + 908, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance comparison with limited training data. We compare the performance of our proposed method with AdvST on the PACS and OfficeHome datasets under varying source domain data ratios (10%, 20%, and 50%). As shown in Tables 7 on the PACS dataset, our method consistently outperforms AdvST across all data ratios, demonstrating robust generalization even with limited source domain data for training. At the 10% data ratio, our method achieves an average accuracy of 58.73%, showing a notable improvement of 9.47% over AdvST's 49.26%. With the 20% data ratio, our approach reaches an average accuracy of 61.78%, surpassing AdvST by 8.23%. At the 50% data ratio, the performance gap narrows slightly, with our method achieving an average accuracy of 66.44%, which is 6.26% higher than the 60.18% achieved by AdvST. The experimental results on the OfficeHome dataset are provided in Suppl. Sec-C.7. These findings suggest that our", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "557", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3b548e14611decb8ebdd47cbb7aefe0c9cf5c653a08496a3c525c2228aa6c54f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BackboneMethodsACPSAvg.
ResNet34PSDG79.4680.7957.4358.2268.98
AdvST79.1676.6765.2860.5570.42
Ours81.8981.8565.6565.3173.68
ResNet50PSDG79.8579.7258.0865.6870.83
AdvST79.6575.0562.6255.7568.27
Ours81.3278.7065.0860.3971.37
ResNet101PSDG79.8479.9759.1770.7972.44
AdvST80.9375.3566.2958.7070.32
Ours82.9580.7968.5764.9574.32
ResNet152PSDG82.9281.5462.3871.2574.52
AdvST80.7679.0265.9061.9971.92
Ours83.4482.1967.6368.1175.34
", + "bbox": [ + 91, + 88, + 486, + 285 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method effectively reduces overfitting to specific domains with limited training data, particularly in low-data settings, by encouraging broader exploration of the parameter space to capture generalizable features.", + "bbox": [ + 89, + 351, + 483, + 412 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c6235090ebf25f7bc5dc346c50ff5929408d27f5d47dd926171477e7899bf41.jpg", + "table_caption": [ + "Table 6. Comparison results (\\%) using different ResNet series backbones on the PACS dataset. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
RatioMethodsACPSAvg.ΔAvg.
10%AdvST58.7456.9440.6240.7349.26
Ours65.4364.4349.4055.6658.73+9.47
20%AdvST61.4864.3143.4244.9853.55
Ours68.8168.7853.2856.2761.78+8.23
50%AdvST70.2772.7449.5248.2060.18
Ours75.8873.5561.6054.7266.44+6.26
", + "bbox": [ + 101, + 425, + 473, + 540 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computational complexity analysis. To evaluate the computational efficiency of our method, we compared its training time cost with three adversarial data augmentation methods on the PACS and OfficeHome datasets. As shown in Table 8, our approach requires an average training time of 1.99 hours on PACS, slightly higher than AdvST (1.90 hours) but comparable to ADA (2.13 hours) and ME-ADA (2.10 hours). This marginal increase is due to the Lyapunov-guided optimization, which enhances model robustness and generalization. On OfficeHome, our method completes training in 1.19 hours, surpassing AdvST (0.75 hours) but remaining faster than ADA (1.95 hours) and ME-ADA (1.90 hours). While our method introduces a slight time increase compared to AdvST, it achieves significant performance improvements, demonstrating a favorable balance between training time and model performance.", + "bbox": [ + 88, + 583, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizability analysis with LEAwareSGD using different adversarial data augmentation approaches. To assess the impact of LEAwareSGD on the generalization performance of adversarial data augmentation methods, we integrated it with ADA, ME-ADA, and AdvST", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a71e1631f6c48141eac37ac0c2eea8dc9eef8ae32c84c570046592de8a66b65.jpg", + "table_caption": [ + "Table 7. Performance comparison (%) using different data ratios on PACS dataset." + ], + "table_footnote": [], + "table_body": "
DatasetMethodsACPSAvg.
PACSADA1.802.101.403.202.13
ME-ADA1.802.101.303.202.10
AdvST1.581.871.402.761.90
Ours1.622.021.352.981.99
OfficeHomeMethodsACPRAvg.
ADA1.202.102.202.301.95
ME-ADA1.201.902.202.301.90
AdvST0.750.740.750.760.75
Ours1.101.171.231.271.19
", + "bbox": [ + 516, + 88, + 916, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Training time (hours) on PACS and OfficeHome datasets.", + "bbox": [ + 513, + 266, + 903, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8d74aed33b4c6a386436373fb02e3ab27dcfd35852fddcb1aaad349992cf67f.jpg", + "image_caption": [ + "(a) Parameter $\\beta$ on the PACS dataset" + ], + "image_footnote": [], + "bbox": [ + 516, + 292, + 689, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3303dbc38c79211304f06470f21e1fea538b26a250935f19a24aa0e3984c1814.jpg", + "image_caption": [ + "(b) Parameter $\\beta$ on the OfficeHome dataset" + ], + "image_footnote": [], + "bbox": [ + 718, + 292, + 888, + 388 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41b79aa285f528be448f92d0f9763bdbfd64e02a211d52b42704b7ef383a66f8.jpg", + "image_caption": [ + "(c) Parameter $\\gamma$ on the PACS dataset", + "Figure 2. Parameter sensitivity study." + ], + "image_footnote": [], + "bbox": [ + 517, + 429, + 689, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/660838a09ed295292ab4cb24909166b78664bc36db40e39913bdad9aee2cdc2c.jpg", + "image_caption": [ + "(d) Parameter $\\gamma$ on the OfficeHome dataset" + ], + "image_footnote": [], + "bbox": [ + 718, + 429, + 888, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on the PACS and OfficeHome datasets. Table 9 shows consistent improvements across all methods, underscoring LEAwareSGD's effectiveness in enhancing model generalization. On the PACS dataset, integrating LEAwareSGD with AdvST led to a $2.40\\%$ increase in average accuracy, reaching $69.46\\%$ . This improvement highlights LEAwareSGD's ability to facilitate more effective parameter space exploration, enhancing AdvST's generalization capacity. For ADA and ME-ADA, LEAwareSGD achieved gains of $0.52\\%$ and $2.30\\%$ , respectively, with ME-ADA reaching an average accuracy of $62.52\\%$ . On the OfficeHome dataset, LEAwareSGD also improved the performance of all methods. ME-ADA saw the highest gain, with a $7.00\\%$ increase over ME-ADA alone, achieving an average accuracy of $52.35\\%$ . ADA with LEAwareSGD increased by $5.15\\%$ , reaching an average accuracy of $49.90\\%$ . For AdvST, which already demonstrated strong performance, LEAwareSGD provided an additional $1.78\\%$ gain, bringing the average accuracy to $54.38\\%$ . These findings demonstrate that LEAwareSGD enhances the generalization", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "558", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "capacity of adversarial data augmentation techniques by enabling broader exploration of the parameter space. This exploration can facilitate learning more generalizable features across unseen domains, underscoring the effectiveness of LEAwareSGD as a robust optimization framework for SDG.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd9cee1f44e11bf1315147bd782d7ca5fd22a1d89de89cc519c31d4a2ccdbd4d.jpg", + "table_caption": [], + "table_footnote": [ + "Note: $\\dagger$ indicates methods integrated with LEAwareSGD." + ], + "table_body": "
DatasetMethodsACPSAvg.ΔAvg.
PACSADA73.7471.6540.7160.3961.62
\\( \\mathrm{ADA}^{\\dagger} \\)71.8269.1551.9455.6662.14+0.52
ME-ADA71.5276.8346.2246.3260.22
\\( \\mathrm{ME-ADA}^{\\dagger} \\)72.8869.1647.9860.0562.52+2.30
AdvST76.6574.9262.4754.1867.06
\\( \\mathrm{AdvST}^{\\dagger} \\)79.1777.1665.0557.7869.46+2.40
Office HomeMethodsACPRAvg.ΔAvg.
ADA45.0736.0744.4353.4444.75
\\( \\mathrm{ADA}^{\\dagger} \\)50.6948.5246.1154.2849.90+5.15
ME-ADA45.2745.7938.6551.6945.35
\\( \\mathrm{ME-ADA}^{\\dagger} \\)53.6150.0949.1756.5152.35+7.00
AdvST51.3252.0848.8958.1152.60
\\( \\mathrm{AdvST}^{\\dagger} \\)52.8955.1851.6958.9554.38+1.78
", + "bbox": [ + 91, + 181, + 488, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Performance improvement (%) of adversarial data augmentation approaches with LEAwareSGD on the PACS and OfficeHome datasets.", + "bbox": [ + 89, + 446, + 483, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter sensitivity analysis. To assess the impact of hyperparameters on the performance of LEAwareSGD, we conducted a sensitivity analysis on the parameters beta $\\beta$ and $\\gamma$ over the PACS and OfficeHome datasets, as shown in Figure 2. Figures 2a and 2b show that increasing $\\beta$ from 1e-4 to 1.0 resulted in fluctuating accuracy, with peak performance observed at $\\beta = 1\\mathrm{e} - 3$ on PACS and $\\beta = 1\\mathrm{e} - 2$ on OfficeHome. This result suggests an optimal range between 1e-3 and 1e-2.", + "bbox": [ + 89, + 506, + 483, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Figure 2c and 2d, for the parameter $\\gamma$ in Equation 10, which controls the strength of regularization, the best accuracy on the PACS dataset was achieved at $\\gamma = 5\\mathrm{e - }4$ . Increasing $\\gamma$ beyond 5e-4 caused slight declines in accuracy, indicating that higher regularization may overly restrict the model. On the OfficeHome dataset, $\\gamma = 1\\mathrm{e - }5$ yielded the best accuracy across most domains, while higher values generally led to reduced performance.", + "bbox": [ + 89, + 643, + 483, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of LE results of different methods on the PACS dataset. To evaluate our proposed method's stability and generalization performance, we compare the LE values obtained from different adversarial data augmentation methods on the PACS dataset. Figure 3 shows that we track LE dynamics across four domains: Art Painting, Cartoon, Photo, and Sketch. Our method consistently yields LE values closer to zero, indicating that the model operates near the edge of chaos, where an optimal balance between sta", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb46c638218fd55a7d3c6cc88aa0e4f9bb91b369ea61eebcfd392bbe7c00a66c.jpg", + "image_caption": [ + "(a) Trained on Art Painting domain" + ], + "image_footnote": [], + "bbox": [ + 517, + 85, + 689, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/621e693643a1dc9051e61af3c73388c3f51970f17193486b94c78cb3d4f3c14a.jpg", + "image_caption": [ + "(b) Trained on Cartoon domain" + ], + "image_footnote": [], + "bbox": [ + 718, + 85, + 888, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/301cb89cdfdd85dbe72653782f3193ddc47584d291cfae76a287af06fc982da0.jpg", + "image_caption": [ + "(c) Trained on Photo domain" + ], + "image_footnote": [], + "bbox": [ + 517, + 210, + 687, + 310 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed2c0b4c9754e345ea61f72cffd82292bd53c39af3e5cfd723cd8d4cab684d11.jpg", + "image_caption": [ + "(d) Trained on Sketch domain", + "Figure 3. LE dynamics over training epochs across four domains (Art Painting, Cartoon, Photo, Sketch) on the PACS dataset." + ], + "image_footnote": [], + "bbox": [ + 714, + 210, + 885, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "bility and adaptability is achieved.", + "bbox": [ + 511, + 378, + 741, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the Art Painting domain (Figure 3a), our method maintains LE values closer to zero throughout training compared to ADA, ME-ADA, and AdvST, suggesting a better balance between stability and adaptability, leading to improved generalization. In contrast, ADA exhibits the lowest LE values and large fluctuations, indicating instability and convergence to local optima. Similar results can be derived from the three remaining domains (Figure 3b-3d), where our method consistently stabilizes near the edge of chaos, supporting better generalization. Comparison methods, by contrast, show greater fluctuations or lower LE values, indicating tendencies towards overfitting and limited adaptability. These results reveal that our method consistently achieves LE values near zero across all domains and operates effectively near the edge of chaos. Additional detailed experimental analyses are provided in Suppl. Sec-A.2.", + "bbox": [ + 511, + 393, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce LEAwareSGD, a novel LE-guided optimization integrated with an adversarial data augmentation method to address the challenge of unseen domain shifts in SDG. By leveraging principles from dynamical systems theory, our approach encourages model training near the edge of chaos, enabling the exploration of a broader parameter space and substantially enhancing generalization to new domains. LEAwareSGD dynamically adjusts the learning rate based on LE values, encouraging the model to capture more generalizable features. Experiments on three standard SDG benchmarks demonstrate that LEAwareSGD outperforms state-of-the-art SDG methods, underscoring its effectiveness. A promising future work is to investigate its scalability to larger datasets and more complex domain generalization tasks.", + "bbox": [ + 511, + 674, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "559", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported in part by the National Natural Science Foundation of China (No. 62276038), and in part by the Chongqing Municipal Education Commission (No. KJZD-M202400603, No. KJQN202500612), and in part by the Project of Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism (No. H2023009).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kathleen T Alligood, Tim D Sauer, James A Yorke, and David Chillingworth. Chaos: an introduction to dynamical systems. SIAM Review, 40(3):732-732, 1998. 1", + "[2] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11400-11410, 2023. 2", + "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6", + "[4] Iike Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4174, 2022. 6", + "[5] Terrance DeVries. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6", + "[6] Ling Feng, Lin Zhang, and Choy Heng Lai. Optimal machine intelligence at the edge of chaos. arXiv preprint arXiv:1909.05176, 2019. 1", + "[7] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2", + "[8] Alejandro Guarneros-Sandoval, Mariana Ballesteros, Ivan Salgado, Julia Rodríguez-Santillán, and Isaac Chairez. Lyapunov stable learning laws for multilayer recurrent neural networks. Neurocomputing, 491:644-657, 2022. 3", + "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International Conference on Machine Learning, pages 1225-1234. PMLR, 2016. 2", + "[10] Kazuko Hayashi. Chaotic nature of the electroencephalogram during shallow and deep anesthesia: From analysis of the lyapunov exponent. *Neuroscience*, 557:116-123, 2024. 1", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 4", + "[12] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A" + ], + "bbox": [ + 93, + 260, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019.6", + "[13] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124-140, 2020. 5, 6", + "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 2, 5, 6", + "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 2, 4", + "[16] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 1, 2, 5, 6", + "[17] Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dynamical systems: An approximation perspective. Journal of the European Mathematical Society, 25(5):1671-1709, 2022. 3", + "[18] Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hyperspectral image classification: An overview. IEEE Transactions on Geoscience and Remote Sensing, 57(9):6690-6709, 2019. 1", + "[19] Guan-Horng Liu and Evangelos A Theodorou. Deep learning theory review: An optimal control and dynamical systems perspective. arXiv preprint arXiv:1908.10920, 2019. 3", + "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2, 5, 6", + "[21] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 5", + "[22] Yurii Nesterov. A method for solving the convex programming problem with convergence rate o (1/k2). In Proceedings of the USSR Academy of Sciences, pages 543-547, 1983. 2", + "[23] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2009. 1", + "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 2, 4", + "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 2", + "[26] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2, 5, 6" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "560", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ling Shao, Fan Zhu, and Xuelong Li. Transfer learning for visual categorization: A survey. IEEE Transactions on Neural Networks and Learning Systems, 26(5):1019-1034, 2014. 1", + "[28] Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937, 2021.5", + "[29] Yang Tang, Jürgen Kurths, Wei Lin, Edward Ott, and Ljupco Kocarev. Introduction to focus issue: When machine learning meets complex systems: Networks, chaos, and nonlinear dynamics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(6), 2020. 3", + "[30] Tijmen Tieleman and Geoffrey Hinton. Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn, 17, 2012. 2, 5, 6", + "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 2, 4", + "[32] Ryan Vogt, Maximilian Puelma Touzel, Eli Shlizerman, and Guillaume Lajoie. On lyapunov exponents for rnns: Understanding information propagation using dynamical systems tools. Frontiers in Applied Mathematics and Statistics, 8: 818799, 2022. 3", + "[33] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 5339-5349, 2018. 1, 2, 5, 6", + "[34] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4682-4691, 2022. 5", + "[35] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769-3778, 2023. 5", + "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 5, 6", + "[37] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4798-4808, 2023. 1, 2, 5, 6", + "[38] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual represent" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tation learning via random convolutions. arXiv preprint arXiv:2007.13003, 2020.5", + "[39] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 5, 6", + "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Causality-inspired domain expansion network for single domain generalization. Knowledge-Based Systems, 301:112269, 2024. 1, 2", + "[41] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3794-3805, 2024. 4, 5, 6", + "[42] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 5", + "[43] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE Transactions on Medical Imaging, 39(7):2531-2540, 2020. 2", + "[44] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Edge of chaos as a guiding principle for modern neural network training. arXiv preprint arXiv:2107.09437, 2021. 1, 3", + "[45] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Asymptotic edge of chaos as guiding principle for neural network training. International Journal of Artificial Intelligence and Robotics Research, 1(01):2350001, 2024. 3", + "[46] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5189-5202, 2023. 2", + "[47] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. Advances in Neural Information Processing Systems, 33:14435–14447, 2020. 1, 2, 5, 6", + "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21832-21840, 2024. 2, 5, 6", + "[49] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision, 132 (3):822-836, 2024. 5, 6", + "[50] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 1", + "[51] Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training. arXiv preprint arXiv:2203.08065, 2022. 2, 5" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "561", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_model.json b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8cbe724a0988fcc8c9d92c438cbd24e9bee990e3 --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.903, + 0.177 + ], + "angle": 0, + "content": "Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.203, + 0.858, + 0.223 + ], + "angle": 0, + "content": "Zuyu Zhang\\(^{1,2,3}\\) Ning Chen\\(^{1,3}\\) Yongshan Liu\\(^{4}\\) Qinghua Zhang\\(^{2,3}\\) Xu Zhang\\(^{1,2,3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.224, + 0.769, + 0.239 + ], + "angle": 0, + "content": "1 School of Artificial Intelligence, Chongqing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.241, + 0.825, + 0.256 + ], + "angle": 0, + "content": "\\(^{2}\\) Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.259, + 0.833, + 0.274 + ], + "angle": 0, + "content": "3 Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism, China" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.277, + 0.782, + 0.292 + ], + "angle": 0, + "content": "\\(^{4}\\) School of Mathematics and Statistics, Chongqing University of Posts and Telecommunications" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.486, + 0.691 + ], + "angle": 0, + "content": "Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to \\(9.47\\%\\) improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.222, + 0.736 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Deep neural networks have shown remarkable progress across various tasks in standard supervised learning settings where the training and test data come from the same distribution [11, 18]. However, their performance often suffers significantly when encountering domain shift, a prevalent issue in real-world applications [23, 27, 50]. For instance, a model trained on DSLR images may struggle when tested on smartphone images due to differences in reso" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.907, + 0.373 + ], + "angle": 0, + "content": "lution, noise, and other factors. These domain shifts are pervasive in real-world scenarios, emphasizing the need for models that can generalize well to unseen distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.909, + 0.576 + ], + "angle": 0, + "content": "Single Domain Generalization (SDG) methods aim to address these domain shifts when only a single source domain is available for training. Current SDG approaches primarily employ data augmentation techniques to generate diverse training samples [16, 25, 33, 37, 40, 47]. Methods such as adversarial data augmentation [25, 33, 47] generate perturbed samples that simulate potential domain shifts. Despite their contributions, as shown in Figure 1, these adversarial data augmentation approaches often fail to adequately capture the global structure of domain shifts in the parameter space due to their emphasis on localized perturbations. As a result, they may underexplore the parameter space, limiting generalization across diverse distributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.882 + ], + "angle": 0, + "content": "In this work, we draw inspiration from dynamical systems theory [1, 6] and propose a novel perspective of model perturbations based on the edge of chaos. This critical transition point maintains a balance between order and chaos in dynamical systems. In SDG, training models near the edge of chaos can help capture features that better generalize across domains by maintaining both stability and adaptability [44]. To quantitatively measure a model's degree of chaos, we employ the Lyapunov Exponent (LE) [10] as a metric to estimate the model's proximity to this edge of chaos. A high LE indicates chaotic behavior and sensitivity to perturbations, whereas a low LE suggests stable dynamics that may restrict the model's generalization capability. We frame the training of neural networks as a discrete-time dynamical system in the parameter space, where each parameter update represents a state transition in this system. By dynamically adjusting the LE, we aim to guide adversarial data augmentation in a way that encourages broader exploration of the parameter space, balancing stability and adaptability for better domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Building on this insight, we introduce LE-Aware" + }, + { + "type": "page_footnote", + "bbox": [ + 0.107, + 0.876, + 0.414, + 0.888 + ], + "angle": 0, + "content": "*The corresponding author, email: zhangx@cqupt.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.107, + 0.888, + 0.486, + 0.9 + ], + "angle": 0, + "content": "The code is available at: https://github.com/ChannySwift/LEAwareSGD." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.282, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.229, + 0.212, + 0.241 + ], + "angle": 0, + "content": "(a) ADA" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.089, + 0.481, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.229, + 0.424, + 0.241 + ], + "angle": 0, + "content": "(b) ME-ADA" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.261, + 0.28, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.399, + 0.217, + 0.411 + ], + "angle": 0, + "content": "(c) AdvST" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.262, + 0.481, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.4, + 0.411, + 0.411 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Figure 1. t-SNE visualization of parameter trajectories in initial epochs for a model trained on the PACS dataset [15]. Each point represents the model's parameters after an iteration. Comparison between (a) ADA [33], (b) ME-ADA [47], (c) AdvST [48], and (d) our method demonstrates that our approach can explore a broader parameter space, facilitating the learning of more generalizable features across unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Stochastic Gradient Descent (LEAwareSGD), a novel optimizer that dynamically adjusts the learning rate based on LE computations. By integrating adversarial data augmentation with LE-aware optimization, LEAwareSGD generates diverse perturbations that enhance the model's ability to generalize to unseen domains while guiding it to operate near the edge of chaos. Unlike conventional optimizers [14, 20, 26, 30] without considering the system's state, LEAwareSGD leverages LE as a feedback mechanism, adjusting the learning rate in response to the system's state and enabling the model to explore the parameter space effectively for capturing more generalizable features." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.734, + 0.414, + 0.748 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- We conceptualize the training of neural networks as a discrete-time dynamical system and leverage the LE to assess the system's proximity to the edge of chaos." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- We propose LEAwareSGD, an optimizer for adversarial data augmentation that dynamically adjusts the learning rate based on LE, guiding the training process near the edge of chaos to enhance generalization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Extensive experiments on three standard SDG benchmarks, including PACS [15], OfficeHome [31], and DomainNet [24], validate the effectiveness of the proposed" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.892, + 0.107 + ], + "angle": 0, + "content": "LEAwareSGD to improve cross-domain generalization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.123, + 0.655, + 0.138 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.149, + 0.78, + 0.164 + ], + "angle": 0, + "content": "2.1. Single Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.172, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Single domain generalization seeks to improve model robustness by enabling generalization from a single source domain to unseen target domains, eliminating the need for target data during training. Most SDG methods rely on data augmentation [25, 43] to increase diversity within the source domain, while generative model-based approaches [16, 37, 40] synthesize samples to extend the source distribution. Although these generative techniques introduce sample variety, they are often computationally demanding and complex to train, frequently producing samples that lack sufficient quality and diversity. This limitation can cause models to overfit to synthetic data rather than enhancing generalization to real-world domains. Adversarial data augmentation [2, 47, 48] offers an alternative by generating adversarial examples that mimic domain shifts to improve model robustness. However, these examples often introduce only localized perturbations, failing to explore the parameter space extensively. In contrast, our proposed LE-guided optimization introduces a dynamic SDG framework that modulates model sensitivity to perturbations. This LE-aware optimization promotes broader exploration of the parameter space, effectively capturing complex domain shifts and improving cross-domain generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.838, + 0.547 + ], + "angle": 0, + "content": "2.2. Optimizers in Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Optimizers play a critical role in model generalization, with different strategies yielding varied outcomes depending on the loss landscape's properties. Widely-used optimizers like Stochastic Gradient Descent (SGD) [22] and Adam [14] are commonly used in deep learning tasks, yet they exhibit different behaviors in terms of generalization. Some prior works [9, 46] show that Adam often results in worse generalization than SGD due to its susceptibility to sharp minima. In contrast, SGD tends to find flatter minima more conducive to robust generalization. This distinction highlights the importance of the loss surface's geometry, as flatter minima are generally associated with improved generalization performance, especially in out-of-distribution settings. In light of this, recent optimization strategies have been explicitly designed to address the sharpness of minima. Sharpness-Aware Minimization (SAM) [7] represents a key development in this direction, where the objective is to minimize sharpness and seek flatter regions of the loss landscape. SAM and its variants [46, 51] have demonstrated notable improvements in generalization, particularly in distribution. Recent improvements such as GSAM [51], which introduces a surrogate gap to better distinguish between sharp and flat minima, further enhance generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "553" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "by guiding the model toward regions with both low sharpness and low loss. Our proposed LEAwareSGD differs from these optimizers by dynamically adjusting the learning rate based on the LE, allowing the model toward the edge of chaos and exploring a broader parameter space for improving generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.416, + 0.208 + ], + "angle": 0, + "content": "2.3. Dynamical Systems in Deep Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.485, + 0.485 + ], + "angle": 0, + "content": "The application of dynamical systems theory to deep learning has introduced new perspectives on neural network training, particularly in understanding stability, and generalization [19, 29]. By modeling neural networks as discrete-time systems, researchers have used concepts such as chaos theory to enhance generalization, especially within residual networks via flow maps [17]. This perspective underscores the significance of training near the edge of chaos, a state that balances stability and adaptability [44, 45]. Lyapunov exponent (LE) is a key tool to quantitatively measure the model's training stability, particularly in recurrent architectures [8, 32]. However, existing methods have not fully leveraged LE for direct control of training under domain shifts, particularly in the context of adversarial data augmentation for domain generalization tasks. Our proposed LEAwareSGD addresses this limitation by using LE to dynamically guide the optimization process, effectively enhancing model generalization capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.497, + 0.226, + 0.515 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.336, + 0.537 + ], + "angle": 0, + "content": "3.1. Overview of LEAwareSGD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.635 + ], + "angle": 0, + "content": "We propose LEAwareSGD, a novel optimization method for adversarial data augmentation to enhance generalization by exploiting the edge of chaos. In this approach, the LE serves as a metric to identify and explore regions in the parameter space where generalizable features are more likely to emerge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.643, + 0.362, + 0.658 + ], + "angle": 0, + "content": "3.2. LE-Based Model Perturbation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.741 + ], + "angle": 0, + "content": "The LE quantifies the rate at which small perturbations in model parameters evolve during training, offering insights into the stability of the optimization process. Here, we denote the model parameters \\(\\theta_{t}\\) at iteration \\(t\\) and the standard gradient descent update can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.749, + 0.483, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\theta_ {t + 1} = \\theta_ {t} - \\eta_ {t} \\nabla L (\\theta_ {t}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.483, + 0.8 + ], + "angle": 0, + "content": "where \\(\\eta_t\\) denotes the learning rate and \\(L(\\theta_t)\\) represents the loss function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.484, + 0.878 + ], + "angle": 0, + "content": "To quantify how perturbations evolve during training, we introduce an initial perturbation \\(\\delta \\theta_0\\) to the model parameters. Parameters after perturbation is \\(\\widetilde{\\theta}_t\\coloneqq \\theta_t + \\delta \\theta_t\\), where \\(\\delta\\) is a tiny real-valued scalar hyperparameter. The perturbation at step \\(t + 1\\) propagates according to the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.883, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\theta} _ {t + 1} = \\widetilde {\\theta} _ {t} - \\eta_ {t} \\nabla L (\\widetilde {\\theta} _ {t}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Combining Equation 1 and Equation 2, we get the following update for the perturbation:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.127, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\delta \\theta_ {t + 1} = \\delta \\theta_ {t} - \\eta_ {t} (\\nabla L (\\widetilde {\\theta_ {t}}) - \\nabla L (\\theta_ {t})). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.158, + 0.907, + 0.205 + ], + "angle": 0, + "content": "To further analyze the propagation of this perturbation, we perform a first-order Taylor expansion of the loss function \\( L(\\widetilde{\\theta_t}) \\) at \\( \\theta_{t} \\) and calculate the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.224, + 0.905, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\nabla L \\left(\\widetilde {\\theta} _ {t}\\right) = \\nabla L \\left(\\theta_ {t}\\right) + H \\left[ L \\left(\\theta_ {t}\\right) \\right] \\delta \\theta_ {t} + o \\left(\\| \\delta \\theta_ {t} \\| ^ {2}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.254, + 0.905, + 0.316 + ], + "angle": 0, + "content": "where \\(H[L(\\theta_t)]\\) is the Hessian matrix, and \\(o(\\| \\delta \\theta_t\\|^2)\\) represents higher-order terms. Ignoring these higher-order terms, and substituting Equation 4 into Equation 3, we approximate the evolution of the perturbation as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.325, + 0.905, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\delta \\theta_ {t + 1} = \\left(I - \\eta_ {t} H \\left[ L \\left(\\theta_ {t}\\right) \\right]\\right) \\delta \\theta_ {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.38 + ], + "angle": 0, + "content": "By recursively applying this relationship, we obtain the propagation of the perturbation over time:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.397, + 0.904, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\delta \\theta_ {t} = \\left(I - \\eta_ {t - 1} H \\left[ L \\left(\\theta_ {t - 1}\\right) \\right]\\right) \\left(I - \\eta_ {t - 2} H \\left[ L \\left(\\theta_ {t - 2}\\right) \\right]\\right) \\tag {6} \\\\ \\cdot \\cdot \\cdot (I - \\eta_ {0} H [ L (\\theta_ {0}) ]) \\delta \\theta_ {0}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.445, + 0.824, + 0.46 + ], + "angle": 0, + "content": "The definition of the Lyapunov exponent is:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.471, + 0.905, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L E} = \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\ln \\left(\\frac {\\| \\delta \\theta_ {t} \\|}{\\| \\delta \\theta_ {0} \\|}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.905, + 0.56 + ], + "angle": 0, + "content": "By substituting Equation 6 into the definition of LE, we can derive the connection between the model's learning rate \\(\\eta\\) and LE:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.567, + 0.905, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L E} \\geq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {\\substack {i = 0\\\\t - 1}} ^ {t - 1} \\ln (1 - \\eta_ {i} \\| H [ L (\\theta_ {i}) ] \\|) \\tag{8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.612, + 0.848, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L E} \\leq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {i = 0} ^ {t - 1} \\ln \\left(\\| I - \\eta_ {i} H [ L (\\theta_ {i}) ] \\|\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.657, + 0.905, + 0.688 + ], + "angle": 0, + "content": "where the LE is determined by the learning rate \\(\\eta_{i}\\) and the Hessian matrix \\(H[L(\\theta_i)]\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.698, + 0.744, + 0.714 + ], + "angle": 0, + "content": "3.3. LE-Guided Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "A positive LE indicates the exponential growth of perturbations, signifying instability, while a negative LE suggests convergence and stability. However, extreme stability or instability can lead to either overfitting or convergence difficulties. Our approach aims to guide model parameters toward an LE value close to zero but slightly negative, positioning the system near the edge of chaos. This critical state balances stability with adaptability, enhancing the model's generalization capacity. To achieve this, we dynamically adjust the learning rate to amplify perturbations, thereby steering the LE toward closer to the edge of chaos. Specifically, the innovation of the proposed LEAwareSGD lies in" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "554" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "its learning rate adjustment based on LE dynamics, allowing the model to adapt effectively across training phases. We modulate the learning rate according to changes in the LE, \\(\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.169, + 0.483, + 0.187 + ], + "angle": 0, + "content": "\\[\n\\eta_ {t + 1} = \\eta_ {t} \\cdot \\exp (- \\beta \\cdot \\Delta \\mathrm {L E} _ {t}) \\quad \\text {i f} \\quad \\Delta \\mathrm {L E} _ {t} > 0 \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.484, + 0.364 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a weighting parameter that controls the sensitivity of the learning rate adjustment. When \\(\\Delta \\mathrm{LE}_t > 0\\), indicating that the model is approaching regions near the edge of chaos where generalizable features are more likely to emerge, we reduce the learning rate to allow for deeper exploration of these regions. This approach ensures that the model explores parameter regions with strong generalization potential while avoiding over-stabilization that could hinder the learning process. When \\(\\Delta \\mathrm{LE}_t \\leq 0\\), we leave the learning rate unchanged, maintaining the current exploration rate." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.453, + 0.392 + ], + "angle": 0, + "content": "3.4. LE-Based Adversarial Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.534 + ], + "angle": 0, + "content": "To further enhance the model's robustness to domain shifts, we integrate LEAwareSGD with adversarial data augmentation. Adversarial data augmentation generates samples that simulate domain shifts, exposing the model to a broader range of data variations. Given a model \\( f_{\\theta} \\) with parameters \\( \\theta \\), we define adversarial samples as transformations of the original data. Let \\( \\tau(x; \\omega) \\) denote a set of semantic transformations parameterized by \\( \\omega \\). The joint optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.547, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\theta} \\max _ {\\omega} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {S}} \\left[ \\ell (\\theta ; \\tau (x; \\omega), y) - \\lambda d _ {\\theta} (\\tau (x; \\omega), x) \\right] \\\\ + \\frac {\\gamma}{2} \\cdot \\| \\theta \\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.777 + ], + "angle": 0, + "content": "where \\(\\ell (\\theta ;\\cdot)\\) is the prediction loss, \\(\\lambda\\) balances the adversarial loss and feature consistency, and \\(d_{\\theta}(\\cdot)\\) represents the feature distance between the original and transformed samples. Here, the regularization term \\(\\frac{\\gamma}{2}\\cdot \\| \\theta \\| _2^2\\) denotes a weight decay, ensuring that the Hessian matrix remains approximately positive definite. This regularization encourages LE to be negative, facilitating the stability of model training. Note that this adversarial training procedure is guided by LE-based optimization, which encourages the model to generalize to unseen domains by learning robust features invariant to domain-specific perturbations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.279, + 0.804 + ], + "angle": 0, + "content": "3.5. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The training procedure involves iterative steps alternating between adversarial example generation and parameter updates using LEAwareSGD. In each epoch, adversarial examples are first generated to simulate unseen domain shifts, fostering the model to learn invariant features. Subsequently, LEAwareSGD adjusts the model's parameters with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.199 + ], + "angle": 0, + "content": "a LE-guided learning rate to encourage the parameters' updating towards the edge of chaos. This LE-guided adaptation encourages the exploration of parameter regions conducive to generalization while controlling sensitivity to perturbations. The complete algorithm is shown in Algorithm 1. We also provide a detailed theoretical analysis of the proposed LEAwareSGD in Suppl. Sec-A.1." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.211, + 0.704, + 0.226 + ], + "angle": 0, + "content": "Algorithm 1 LEAwareSGD" + }, + { + "type": "algorithm", + "bbox": [ + 0.522, + 0.23, + 0.905, + 0.474 + ], + "angle": 0, + "content": "1: Input: Initial parameters \\(\\theta_0\\), learning rate \\(\\eta_0\\), weighting parameter \\(\\beta\\), number of iterations \\(N\\) \n2: Output: Optimized parameters \\(\\theta^*\\) \n3: for \\(t = 0, 1, 2, \\dots, N - 1\\) do \n4: Compute gradient \\(\\nabla_{\\theta} L(\\theta_t)\\) \n5: Update parameters: \\(\\theta_{t+1} = \\theta_t - \\eta_t \\nabla_{\\theta} L(\\theta_t)\\) \n6: Compute updated model perturbation using Eq. 6 \n7: Calculate LE using Eq. 7 \n8: Update \\(\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}\\) \n9: if \\(\\Delta \\mathrm{LE}_t > 0\\) then \n10: \\(\\eta_{t+1} = \\eta_t \\cdot \\exp(-\\beta \\cdot \\Delta \\mathrm{LE}_t)\\) \n11: else \n12: \\(\\eta_{t+1} = \\eta_t\\) \n13: end if \n14: end for \n15: Return: \\(\\theta^*\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.492, + 0.638, + 0.509 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.715, + 0.533 + ], + "angle": 0, + "content": "4.1. Datasets and Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three widely-used domain generalization benchmarks: PACS [15], OfficeHome [31], and DomainNet [24], arranging them in increasing levels of difficulty to comprehensively test the model's robustness across varying domain shifts. PACS consists of four domains, namely, Photo (P), Art Painting (A), Cartoon (C), and Sketch (S), comprising 9,991 images from 7 categories. OfficeHome includes four domain, namely, Art (A), Clipart (C), Product (P), and Real-World (R), with 15,500 images across 65 classes. Compared to PACS, OfficeHome introduces greater diversity in both object categories and domain styles. DomainNet is the most challenging, containing six domains, namely, Real (R), Infograph (I), Clipart (C), Painting (P), Quickdraw (Q), and Sketch (S), with 596,010 images from 345 classes. The substantial domain shifts and the high number of classes in this dataset present considerable challenges for domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation details. In all experiments, we employed ResNet-18 [11] as the backbone model and resized all input images to \\(224 \\times 224\\). The network was trained with LEAwareSGD, setting the momentum to 0.9 and \\(\\beta\\) to 1e-1. To ensure a fair comparison, we excluded the test-time learning stage of PSDG[41] in all of our experiments. For PACS and OfficeHome datasets, we follow the standard" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "555" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.364 + ], + "angle": 0, + "content": "protocol of using one domain as the source domain and the remaining three domains as the target domains [47, 48]. For the PACS dataset, we adopted a batch size of 16, \\(\\gamma\\) of 5e-4, and an initial learning rate of 5e-4, training for 50 epochs. For the OfficeHome dataset, we used a batch size of 32, set \\(\\gamma\\) to 1e-4, and initialized the learning rate at 1e-4, training for 50 epochs. For experiments conducted on the DomainNet dataset, we follow [37, 48] and use Real as the source domain and the remaining five domains as the target domains. The batch size was increased to 128, \\(\\gamma\\) was set to 1e-5, and the initial learning rate was set to 1e-3, with a training duration of 200 epochs. For hyperparameter tuning, we fix all training settings (e.g., random seed, learning rate) and tune only \\(\\beta\\) and \\(\\gamma\\), with the best setting shown in Fig.2. All experiments were performed on an NVIDIA A100 GPU, and each experiment was repeated 10 times with different random seeds. The results are reported as the average accuracy alongside the standard deviation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.38, + 0.398, + 0.396 + ], + "angle": 0, + "content": "4.2. Comparison with SDG Approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.631 + ], + "angle": 0, + "content": "Results on PACS. This adjustment aligns the evaluation setup with our method, which operates solely during the training phase. The evaluation results on the PACS dataset, presented in Table 1, indicate that our method achieves the highest average accuracy of \\(69.46\\%\\), outperforming all compared methods. Specifically, our model achieves the highest performance gains in both the A and P domains, with classification accuracies of \\(79.17\\%\\) and \\(65.05\\%\\), respectively. Our approach outperforms the second-best method, PSDG \\((67.14\\%)\\), with an improvement of \\(2.32\\%\\) in average accuracy. This result underscores the effectiveness of our LE-guided adversarial data augmentation strategy, which enhances model generalization by facilitating the exploration of a broader parameter space and reducing overfitting to domain-specific features." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.647, + 0.481, + 0.858 + ], + "angle": 0, + "content": "
MethodsACPSAvg.
ERM70.4973.5645.9241.2157.80
MixUp [39]67.6067.1351.5947.0458.34
RSC [13]73.4075.9056.2041.6061.78
ADA [33]71.5676.8452.3643.6661.11
ME-ADA [47]71.5276.8346.2246.3260.22
SAGM [35]65.9470.1042.1453.6557.96
RandConv [38]73.6874.8855.4246.7762.69
MetaCNN [34]76.5074.5056.7055.7065.80
L2D [36]77.0076.5553.2254.4465.30
PDEN [16]76.4373.8758.5253.9265.69
PSDG [41]77.8875.2455.4959.9667.14
AdvST [48]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.868, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Table 1. Classification accuracy (%) comparison on the PACS dataset. The best results are in bold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Results on OfficeHome. Table 2 shows the classification accuracy comparison of our method with existing baselines on the OfficeHome dataset. Our proposed method consistently outperforms other approaches, achieving the highest accuracy across all domains (A, C, P, and R) with an average accuracy of \\(54.38\\%\\). Notably, our method surpasses the previous best method, AdvST, which obtained an average accuracy of \\(52.60\\%\\), with an improvement of \\(1.78\\%\\). This gain highlights the effectiveness of our LE-guided optimization, which facilitates the exploration of a broader parameter space and enhances the model's ability to generalize to unseen domains." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.283, + 0.905, + 0.508 + ], + "angle": 0, + "content": "
MethodsACPRAvg.
ERM42.7439.3139.8652.5043.60
MixUp [39]43.6940.6838.5652.1243.76
SagNet [21]49.8443.0141.4255.6147.47
RandConv [38]45.7940.2037.4552.8744.08
Fish [28]45.1039.7436.6752.0943.40
ADA [33]45.0736.0744.4353.4444.75
ME-ADA [47]45.2745.7938.6551.6945.35
GSAM [51]28.9741.4041.7251.5040.90
SAGM [35]47.4243.2741.3654.9746.76
Mixstyle [49]51.1948.7346.8555.8850.66
PCGrad [42]49.9146.6146.1156.3149.74
PSDG [41]42.8046.4044.3154.6847.05
AdvST [48]51.3252.0848.8958.1152.60
Ours52.8955.1851.6958.9554.38±0.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.518, + 0.905, + 0.545 + ], + "angle": 0, + "content": "Table 2. Classification accuracy (%) comparison on the Office-Home dataset. The best results are in bold." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Results on DomainNet. Table 3 shows the classification accuracy comparison on the DomainNet dataset. Our proposed method achieves the highest average accuracy of \\(28.15\\%\\), surpassing all other methods, including the second-best method, AdvST, which reaches an average accuracy of \\(27.22\\%\\). Notably, our method performs best in the C and S domains, with accuracies of \\(44.67\\%\\) and \\(33.74\\%\\), respectively. In the Q domain, our method achieves an accuracy of \\(6.70\\%\\), slightly lower than SimDE's \\(6.85\\%\\). This may be due to the highly abstract nature of the Quickdraw domain, which can benefit from specialized augmentation techniques to better capture its distinct features. Despite this, our method's overall performance advantage demonstrates the effectiveness of the LE-guided adversarial data augmentation approach in handling the diversity and complexity of DomainNet." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.885, + 0.82 + ], + "angle": 0, + "content": "4.3. Comparison with Optimization Approaches" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.9 + ], + "angle": 0, + "content": "To further validate the effectiveness of our proposed LEAwareSGD optimizer, we conducted a comparison with four widely used optimization approaches, including Adam [14], AdamW [20], RMSprop [30], and SGD [26], on the PACS and OfficeHome datasets. As shown in Table 4," + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "556" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.498, + 0.3 + ], + "angle": 0, + "content": "
MethodsPICSQAvg.
ERM38.0513.3137.8926.263.3623.77
MixUp [39]38.6013.9238.0126.013.7124.05
CutOut [5]38.3013.7038.4026.203.7024.06
AugMix [12]40.8013.9041.7029.806.3026.50
RandAug [3]41.3013.5741.1130.405.3126.34
ADA [33]38.2013.8040.2024.804.3024.26
ME-ADA [47]38.4813.5640.6626.054.4224.63
ACVC [4]41.3012.9042.8030.906.6026.90
PDEN [16]38.4511.2538.9931.715.5825.20
SimDE [37]39.9612.9141.7333.466.8526.98
PSDG [41]41.1013.0041.2931.104.8926.28
AdvST [48]42.4014.9041.7031.006.1027.22
Ours42.7715.9144.6733.746.7028.15 ±0.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.337 + ], + "angle": 0, + "content": "Table 3. Classification accuracy (%) comparison on the Domain-Net dataset. The best results are in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.347, + 0.487, + 0.664 + ], + "angle": 0, + "content": "our proposed optimizer achieves the highest average accuracy of \\(69.46\\%\\), surpassing all other optimizers on the PACS dataset. In contrast, widely used optimizers such as Adam, AdamW, and RMSprop perform significantly worse, with average accuracies of \\(66.43\\%\\), \\(66.83\\%\\), and \\(62.34\\%\\), respectively. This disparity likely stems from the reliance of these adaptive optimizers on variable learning rates, which may cause overfitting to domain-specific features and reduce their generalization ability in domain generalization tasks. On the other hand, both our LEAwareSGD and standard SGD achieve substantially higher accuracies. While SGD achieves an average accuracy of \\(67.06\\%\\), our LEAwareSGD optimizer further improves this to \\(69.46\\%\\). This enhancement underscores the advantage of integrating LE guidance into the optimization process, which enables the model to better capture generalizable features and avoid overfitting to domain-specific characteristics. We present the experimental results on the OfficeHome dataset in Suppl. Sec-C.2, while analyses of optimizers' performance under different learning rates are provided in Suppl. Sec-C.5." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.676, + 0.496, + 0.776 + ], + "angle": 0, + "content": "
MethodsACPSAvg.
Adam [14]76.5271.1564.0653.9866.43
AdamW [20]76.6871.9362.0356.6866.83
RMSprop [30]71.6271.0859.0947.5562.34
SGD [26]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Table 4. Comparison results \\((\\%)\\) using different optimizers on the PACS dataset. The best results are in bold." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.458, + 0.851 + ], + "angle": 0, + "content": "4.4. Extension to Multi-Source DG Approaches" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To further validate the scalability of our solution, we perform experiments under the leave-one-out setting of general domain generalization (DG) with ResNet-18 as the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "backbone, where three domains are combined for training and the remaining domain is reserved for evaluation. Table 5 presents the results on the PACS dataset, showing that our method delivers at least comparable performance, highlighting its versatility. This improvement underscores the effectiveness of our LE-guided optimization, which enables a broader exploration of the parameter space and enhances the model's generalization to unseen domains." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.223, + 0.888, + 0.351 + ], + "angle": 0, + "content": "
MethodsACPSAvg.
RSC [13]82.8179.7493.9580.8584.34
ADA [33]82.8178.3395.6375.2983.02
ME-ADA [47]77.8878.5895.3378.0782.47
Mixstyle [49]83.1179.4396.3172.9582.95
L2D [36]81.4479.5695.5180.5884.27
AdvST [48]78.6972.9791.6280.9181.05
Ours81.3578.8895.4582.3984.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.361, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Table 5. Classification accuracy (\\%) comparison for the leave-one-domain-out using the ResNet-18 backbone on the PACS dataset. The column name denotes the test domain, with the other three used for training. The best results are in bold." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.704, + 0.442 + ], + "angle": 0, + "content": "4.5. Additional Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.644 + ], + "angle": 0, + "content": "Performance comparison using different backbone networks. To further validate the universality of our solution, we conduct experiments with different ResNet backbones, including ResNet-34, ResNet-50, ResNet-101, and ResNet-152, on the PACS and OfficeHome datasets. As shown in Table 6 on the PACS dataset, our approach reaches a performance of \\(73.68\\%\\) on ResNet-34, \\(71.37\\%\\) on ResNet-50, \\(74.32\\%\\) on ResNet-101, and \\(75.34\\%\\) on ResNet-152. These results demonstrate the consistent effectiveness of our method across various backbone architectures, highlighting its superior accuracy, versatility, and robustness. The experimental results for the OfficeHome dataset are presented in Suppl. Sec-C.6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Performance comparison with limited training data. We compare the performance of our proposed method with AdvST on the PACS and OfficeHome datasets under varying source domain data ratios (10%, 20%, and 50%). As shown in Tables 7 on the PACS dataset, our method consistently outperforms AdvST across all data ratios, demonstrating robust generalization even with limited source domain data for training. At the 10% data ratio, our method achieves an average accuracy of 58.73%, showing a notable improvement of 9.47% over AdvST's 49.26%. With the 20% data ratio, our approach reaches an average accuracy of 61.78%, surpassing AdvST by 8.23%. At the 50% data ratio, the performance gap narrows slightly, with our method achieving an average accuracy of 66.44%, which is 6.26% higher than the 60.18% achieved by AdvST. The experimental results on the OfficeHome dataset are provided in Suppl. Sec-C.7. These findings suggest that our" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "557" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.487, + 0.286 + ], + "angle": 0, + "content": "
BackboneMethodsACPSAvg.
ResNet34PSDG79.4680.7957.4358.2268.98
AdvST79.1676.6765.2860.5570.42
Ours81.8981.8565.6565.3173.68
ResNet50PSDG79.8579.7258.0865.6870.83
AdvST79.6575.0562.6255.7568.27
Ours81.3278.7065.0860.3971.37
ResNet101PSDG79.8479.9759.1770.7972.44
AdvST80.9375.3566.2958.7070.32
Ours82.9580.7968.5764.9574.32
ResNet152PSDG82.9281.5462.3871.2574.52
AdvST80.7679.0265.9061.9971.92
Ours83.4482.1967.6368.1175.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.325 + ], + "angle": 0, + "content": "Table 6. Comparison results (\\%) using different ResNet series backbones on the PACS dataset. The best results are in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.414 + ], + "angle": 0, + "content": "method effectively reduces overfitting to specific domains with limited training data, particularly in low-data settings, by encouraging broader exploration of the parameter space to capture generalizable features." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.426, + 0.475, + 0.541 + ], + "angle": 0, + "content": "
RatioMethodsACPSAvg.ΔAvg.
10%AdvST58.7456.9440.6240.7349.26
Ours65.4364.4349.4055.6658.73+9.47
20%AdvST61.4864.3143.4244.9853.55
Ours68.8168.7853.2856.2761.78+8.23
50%AdvST70.2772.7449.5248.2060.18
Ours75.8873.5561.6054.7266.44+6.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Table 7. Performance comparison (%) using different data ratios on PACS dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Computational complexity analysis. To evaluate the computational efficiency of our method, we compared its training time cost with three adversarial data augmentation methods on the PACS and OfficeHome datasets. As shown in Table 8, our approach requires an average training time of 1.99 hours on PACS, slightly higher than AdvST (1.90 hours) but comparable to ADA (2.13 hours) and ME-ADA (2.10 hours). This marginal increase is due to the Lyapunov-guided optimization, which enhances model robustness and generalization. On OfficeHome, our method completes training in 1.19 hours, surpassing AdvST (0.75 hours) but remaining faster than ADA (1.95 hours) and ME-ADA (1.90 hours). While our method introduces a slight time increase compared to AdvST, it achieves significant performance improvements, demonstrating a favorable balance between training time and model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Generalizability analysis with LEAwareSGD using different adversarial data augmentation approaches. To assess the impact of LEAwareSGD on the generalization performance of adversarial data augmentation methods, we integrated it with ADA, ME-ADA, and AdvST" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.917, + 0.256 + ], + "angle": 0, + "content": "
DatasetMethodsACPSAvg.
PACSADA1.802.101.403.202.13
ME-ADA1.802.101.303.202.10
AdvST1.581.871.402.761.90
Ours1.622.021.352.981.99
OfficeHomeMethodsACPRAvg.
ADA1.202.102.202.301.95
ME-ADA1.201.902.202.301.90
AdvST0.750.740.750.760.75
Ours1.101.171.231.271.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.267, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Table 8. Training time (hours) on PACS and OfficeHome datasets." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.293, + 0.691, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.396, + 0.706, + 0.42 + ], + "angle": 0, + "content": "(a) Parameter \\(\\beta\\) on the PACS dataset" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.293, + 0.89, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.396, + 0.906, + 0.42 + ], + "angle": 0, + "content": "(b) Parameter \\(\\beta\\) on the OfficeHome dataset" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.43, + 0.691, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.533, + 0.705, + 0.545 + ], + "angle": 0, + "content": "(c) Parameter \\(\\gamma\\) on the PACS dataset" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.43, + 0.89, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.533, + 0.905, + 0.556 + ], + "angle": 0, + "content": "(d) Parameter \\(\\gamma\\) on the OfficeHome dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.57, + 0.822, + 0.584 + ], + "angle": 0, + "content": "Figure 2. Parameter sensitivity study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "on the PACS and OfficeHome datasets. Table 9 shows consistent improvements across all methods, underscoring LEAwareSGD's effectiveness in enhancing model generalization. On the PACS dataset, integrating LEAwareSGD with AdvST led to a \\(2.40\\%\\) increase in average accuracy, reaching \\(69.46\\%\\). This improvement highlights LEAwareSGD's ability to facilitate more effective parameter space exploration, enhancing AdvST's generalization capacity. For ADA and ME-ADA, LEAwareSGD achieved gains of \\(0.52\\%\\) and \\(2.30\\%\\), respectively, with ME-ADA reaching an average accuracy of \\(62.52\\%\\). On the OfficeHome dataset, LEAwareSGD also improved the performance of all methods. ME-ADA saw the highest gain, with a \\(7.00\\%\\) increase over ME-ADA alone, achieving an average accuracy of \\(52.35\\%\\). ADA with LEAwareSGD increased by \\(5.15\\%\\), reaching an average accuracy of \\(49.90\\%\\). For AdvST, which already demonstrated strong performance, LEAwareSGD provided an additional \\(1.78\\%\\) gain, bringing the average accuracy to \\(54.38\\%\\). These findings demonstrate that LEAwareSGD enhances the generalization" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "capacity of adversarial data augmentation techniques by enabling broader exploration of the parameter space. This exploration can facilitate learning more generalizable features across unseen domains, underscoring the effectiveness of LEAwareSGD as a robust optimization framework for SDG." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.182, + 0.489, + 0.42 + ], + "angle": 0, + "content": "
DatasetMethodsACPSAvg.ΔAvg.
PACSADA73.7471.6540.7160.3961.62
\\( \\mathrm{ADA}^{\\dagger} \\)71.8269.1551.9455.6662.14+0.52
ME-ADA71.5276.8346.2246.3260.22
\\( \\mathrm{ME-ADA}^{\\dagger} \\)72.8869.1647.9860.0562.52+2.30
AdvST76.6574.9262.4754.1867.06
\\( \\mathrm{AdvST}^{\\dagger} \\)79.1777.1665.0557.7869.46+2.40
Office HomeMethodsACPRAvg.ΔAvg.
ADA45.0736.0744.4353.4444.75
\\( \\mathrm{ADA}^{\\dagger} \\)50.6948.5246.1154.2849.90+5.15
ME-ADA45.2745.7938.6551.6945.35
\\( \\mathrm{ME-ADA}^{\\dagger} \\)53.6150.0949.1756.5152.35+7.00
AdvST51.3252.0848.8958.1152.60
\\( \\mathrm{AdvST}^{\\dagger} \\)52.8955.1851.6958.9554.38+1.78
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.113, + 0.422, + 0.455, + 0.436 + ], + "angle": 0, + "content": "Note: \\(\\dagger\\) indicates methods integrated with LEAwareSGD." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Table 9. Performance improvement (%) of adversarial data augmentation approaches with LEAwareSGD on the PACS and OfficeHome datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.642 + ], + "angle": 0, + "content": "Parameter sensitivity analysis. To assess the impact of hyperparameters on the performance of LEAwareSGD, we conducted a sensitivity analysis on the parameters beta \\(\\beta\\) and \\(\\gamma\\) over the PACS and OfficeHome datasets, as shown in Figure 2. Figures 2a and 2b show that increasing \\(\\beta\\) from 1e-4 to 1.0 resulted in fluctuating accuracy, with peak performance observed at \\(\\beta = 1\\mathrm{e} - 3\\) on PACS and \\(\\beta = 1\\mathrm{e} - 2\\) on OfficeHome. This result suggests an optimal range between 1e-3 and 1e-2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.764 + ], + "angle": 0, + "content": "As illustrated in Figure 2c and 2d, for the parameter \\(\\gamma\\) in Equation 10, which controls the strength of regularization, the best accuracy on the PACS dataset was achieved at \\(\\gamma = 5\\mathrm{e - }4\\). Increasing \\(\\gamma\\) beyond 5e-4 caused slight declines in accuracy, indicating that higher regularization may overly restrict the model. On the OfficeHome dataset, \\(\\gamma = 1\\mathrm{e - }5\\) yielded the best accuracy across most domains, while higher values generally led to reduced performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparison of LE results of different methods on the PACS dataset. To evaluate our proposed method's stability and generalization performance, we compare the LE values obtained from different adversarial data augmentation methods on the PACS dataset. Figure 3 shows that we track LE dynamics across four domains: Art Painting, Cartoon, Photo, and Sketch. Our method consistently yields LE values closer to zero, indicating that the model operates near the edge of chaos, where an optimal balance between sta" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.086, + 0.691, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.188, + 0.704, + 0.201 + ], + "angle": 0, + "content": "(a) Trained on Art Painting domain" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.086, + 0.89, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.188, + 0.893, + 0.2 + ], + "angle": 0, + "content": "(b) Trained on Cartoon domain" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.212, + 0.688, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.314, + 0.687, + 0.325 + ], + "angle": 0, + "content": "(c) Trained on Photo domain" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.212, + 0.886, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.314, + 0.886, + 0.325 + ], + "angle": 0, + "content": "(d) Trained on Sketch domain" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.337, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Figure 3. LE dynamics over training epochs across four domains (Art Painting, Cartoon, Photo, Sketch) on the PACS dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.742, + 0.394 + ], + "angle": 0, + "content": "bility and adaptability is achieved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.635 + ], + "angle": 0, + "content": "In the Art Painting domain (Figure 3a), our method maintains LE values closer to zero throughout training compared to ADA, ME-ADA, and AdvST, suggesting a better balance between stability and adaptability, leading to improved generalization. In contrast, ADA exhibits the lowest LE values and large fluctuations, indicating instability and convergence to local optima. Similar results can be derived from the three remaining domains (Figure 3b-3d), where our method consistently stabilizes near the edge of chaos, supporting better generalization. Comparison methods, by contrast, show greater fluctuations or lower LE values, indicating tendencies towards overfitting and limited adaptability. These results reveal that our method consistently achieves LE values near zero across all domains and operates effectively near the edge of chaos. Additional detailed experimental analyses are provided in Suppl. Sec-A.2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we introduce LEAwareSGD, a novel LE-guided optimization integrated with an adversarial data augmentation method to address the challenge of unseen domain shifts in SDG. By leveraging principles from dynamical systems theory, our approach encourages model training near the edge of chaos, enabling the exploration of a broader parameter space and substantially enhancing generalization to new domains. LEAwareSGD dynamically adjusts the learning rate based on LE values, encouraging the model to capture more generalizable features. Experiments on three standard SDG benchmarks demonstrate that LEAwareSGD outperforms state-of-the-art SDG methods, underscoring its effectiveness. A promising future work is to investigate its scalability to larger datasets and more complex domain generalization tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "559" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work is supported in part by the National Natural Science Foundation of China (No. 62276038), and in part by the Chongqing Municipal Education Commission (No. KJZD-M202400603, No. KJQN202500612), and in part by the Project of Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism (No. H2023009)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Kathleen T Alligood, Tim D Sauer, James A Yorke, and David Chillingworth. Chaos: an introduction to dynamical systems. SIAM Review, 40(3):732-732, 1998. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11400-11410, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[4] Iike Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4174, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Terrance DeVries. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[6] Ling Feng, Lin Zhang, and Choy Heng Lai. Optimal machine intelligence at the edge of chaos. arXiv preprint arXiv:1909.05176, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[7] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[8] Alejandro Guarneros-Sandoval, Mariana Ballesteros, Ivan Salgado, Julia Rodríguez-Santillán, and Isaac Chairez. Lyapunov stable learning laws for multilayer recurrent neural networks. Neurocomputing, 491:644-657, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International Conference on Machine Learning, pages 1225-1234. PMLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Kazuko Hayashi. Chaotic nature of the electroencephalogram during shallow and deep anesthesia: From analysis of the lyapunov exponent. *Neuroscience*, 557:116-123, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.261, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124-140, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.194 + ], + "angle": 0, + "content": "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.197, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.255, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[16] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[17] Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dynamical systems: An approximation perspective. Journal of the European Mathematical Society, 25(5):1671-1709, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[18] Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hyperspectral image classification: An overview. IEEE Transactions on Geoscience and Remote Sensing, 57(9):6690-6709, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[19] Guan-Horng Liu and Evangelos A Theodorou. Deep learning theory review: An optimal control and dynamical systems perspective. arXiv preprint arXiv:1908.10920, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.557, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[21] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[22] Yurii Nesterov. A method for solving the convex programming problem with convergence rate o (1/k2). In Proceedings of the USSR Academy of Sciences, pages 543-547, 1983. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[23] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "560" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Ling Shao, Fan Zhu, and Xuelong Li. Transfer learning for visual categorization: A survey. IEEE Transactions on Neural Networks and Learning Systems, 26(5):1019-1034, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[28] Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937, 2021.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[29] Yang Tang, Jürgen Kurths, Wei Lin, Edward Ott, and Ljupco Kocarev. Introduction to focus issue: When machine learning meets complex systems: Networks, chaos, and nonlinear dynamics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(6), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[30] Tijmen Tieleman and Geoffrey Hinton. Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn, 17, 2012. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[32] Ryan Vogt, Maximilian Puelma Touzel, Eli Shlizerman, and Guillaume Lajoie. On lyapunov exponents for rnns: Understanding information propagation using dynamical systems tools. Frontiers in Applied Mathematics and Statistics, 8: 818799, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[33] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 5339-5349, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[34] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4682-4691, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[35] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769-3778, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[37] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4798-4808, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[38] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual represent" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "tation learning via random convolutions. arXiv preprint arXiv:2007.13003, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[39] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Causality-inspired domain expansion network for single domain generalization. Knowledge-Based Systems, 301:112269, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[41] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3794-3805, 2024. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[42] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.427 + ], + "angle": 0, + "content": "[43] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE Transactions on Medical Imaging, 39(7):2531-2540, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.469 + ], + "angle": 0, + "content": "[44] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Edge of chaos as a guiding principle for modern neural network training. arXiv preprint arXiv:2107.09437, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[45] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Asymptotic edge of chaos as guiding principle for neural network training. International Journal of Artificial Intelligence and Robotics Research, 1(01):2350001, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[46] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5189-5202, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.594, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[47] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. Advances in Neural Information Processing Systems, 33:14435–14447, 2020. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.907, + 0.719 + ], + "angle": 0, + "content": "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21832-21840, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[49] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision, 132 (3):822-836, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[50] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[51] Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training. arXiv preprint arXiv:2203.08065, 2022. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "561" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_origin.pdf b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..76eec31d6e83bd98584fd3f5d2b1a404ad363c44 --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/f0ac30f5-bb31-44e2-94e3-b0f8627b8df9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fe33094c8a26b99272c8d45dd527fc786393d9ac6c38a85895fcc6ef2dee332 +size 648843 diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/full.md b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9f0ff368fb321945ff83d0cb301ed6175677ec56 --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/full.md @@ -0,0 +1,362 @@ +# Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization + +Zuyu Zhang $^{1,2,3}$ Ning Chen $^{1,3}$ Yongshan Liu $^{4}$ Qinghua Zhang $^{2,3}$ Xu Zhang $^{1,2,3,*}$ + +1 School of Artificial Intelligence, Chongqing University of Posts and Telecommunications + +$^{2}$ Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications + +3 Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism, China + +$^{4}$ School of Mathematics and Statistics, Chongqing University of Posts and Telecommunications + +# Abstract + +Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to $9.47\%$ improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks. + +# 1. Introduction + +Deep neural networks have shown remarkable progress across various tasks in standard supervised learning settings where the training and test data come from the same distribution [11, 18]. However, their performance often suffers significantly when encountering domain shift, a prevalent issue in real-world applications [23, 27, 50]. For instance, a model trained on DSLR images may struggle when tested on smartphone images due to differences in reso + +lution, noise, and other factors. These domain shifts are pervasive in real-world scenarios, emphasizing the need for models that can generalize well to unseen distributions. + +Single Domain Generalization (SDG) methods aim to address these domain shifts when only a single source domain is available for training. Current SDG approaches primarily employ data augmentation techniques to generate diverse training samples [16, 25, 33, 37, 40, 47]. Methods such as adversarial data augmentation [25, 33, 47] generate perturbed samples that simulate potential domain shifts. Despite their contributions, as shown in Figure 1, these adversarial data augmentation approaches often fail to adequately capture the global structure of domain shifts in the parameter space due to their emphasis on localized perturbations. As a result, they may underexplore the parameter space, limiting generalization across diverse distributions. + +In this work, we draw inspiration from dynamical systems theory [1, 6] and propose a novel perspective of model perturbations based on the edge of chaos. This critical transition point maintains a balance between order and chaos in dynamical systems. In SDG, training models near the edge of chaos can help capture features that better generalize across domains by maintaining both stability and adaptability [44]. To quantitatively measure a model's degree of chaos, we employ the Lyapunov Exponent (LE) [10] as a metric to estimate the model's proximity to this edge of chaos. A high LE indicates chaotic behavior and sensitivity to perturbations, whereas a low LE suggests stable dynamics that may restrict the model's generalization capability. We frame the training of neural networks as a discrete-time dynamical system in the parameter space, where each parameter update represents a state transition in this system. By dynamically adjusting the LE, we aim to guide adversarial data augmentation in a way that encourages broader exploration of the parameter space, balancing stability and adaptability for better domain generalization. + +Building on this insight, we introduce LE-Aware + +![](images/12c7c405a8ef6d5dbd4c0792897b579e0750d9ebfe0f6525b14c5521c5dea037.jpg) +(a) ADA + +![](images/d9c07935d20914b168c0256d8aadc6d5a960b59ee988473d7ff6bba8a4a3aede.jpg) +(b) ME-ADA + +![](images/8184d81b360efdb6cd29cd76aa754439114d749be2add7363d09ed663edc8c55.jpg) +(c) AdvST +Figure 1. t-SNE visualization of parameter trajectories in initial epochs for a model trained on the PACS dataset [15]. Each point represents the model's parameters after an iteration. Comparison between (a) ADA [33], (b) ME-ADA [47], (c) AdvST [48], and (d) our method demonstrates that our approach can explore a broader parameter space, facilitating the learning of more generalizable features across unseen domains. + +![](images/0a690f1defb557c6b73500c326de9c3f45cb8f21c95fafe864469a82448cd7f1.jpg) +(d) Ours + +Stochastic Gradient Descent (LEAwareSGD), a novel optimizer that dynamically adjusts the learning rate based on LE computations. By integrating adversarial data augmentation with LE-aware optimization, LEAwareSGD generates diverse perturbations that enhance the model's ability to generalize to unseen domains while guiding it to operate near the edge of chaos. Unlike conventional optimizers [14, 20, 26, 30] without considering the system's state, LEAwareSGD leverages LE as a feedback mechanism, adjusting the learning rate in response to the system's state and enabling the model to explore the parameter space effectively for capturing more generalizable features. + +Our contributions are summarized as follows: + +- We conceptualize the training of neural networks as a discrete-time dynamical system and leverage the LE to assess the system's proximity to the edge of chaos. +- We propose LEAwareSGD, an optimizer for adversarial data augmentation that dynamically adjusts the learning rate based on LE, guiding the training process near the edge of chaos to enhance generalization. +- Extensive experiments on three standard SDG benchmarks, including PACS [15], OfficeHome [31], and DomainNet [24], validate the effectiveness of the proposed + +LEAwareSGD to improve cross-domain generalization. + +# 2. Related Work + +# 2.1. Single Domain Generalization + +Single domain generalization seeks to improve model robustness by enabling generalization from a single source domain to unseen target domains, eliminating the need for target data during training. Most SDG methods rely on data augmentation [25, 43] to increase diversity within the source domain, while generative model-based approaches [16, 37, 40] synthesize samples to extend the source distribution. Although these generative techniques introduce sample variety, they are often computationally demanding and complex to train, frequently producing samples that lack sufficient quality and diversity. This limitation can cause models to overfit to synthetic data rather than enhancing generalization to real-world domains. Adversarial data augmentation [2, 47, 48] offers an alternative by generating adversarial examples that mimic domain shifts to improve model robustness. However, these examples often introduce only localized perturbations, failing to explore the parameter space extensively. In contrast, our proposed LE-guided optimization introduces a dynamic SDG framework that modulates model sensitivity to perturbations. This LE-aware optimization promotes broader exploration of the parameter space, effectively capturing complex domain shifts and improving cross-domain generalization capability. + +# 2.2. Optimizers in Domain Generalization + +Optimizers play a critical role in model generalization, with different strategies yielding varied outcomes depending on the loss landscape's properties. Widely-used optimizers like Stochastic Gradient Descent (SGD) [22] and Adam [14] are commonly used in deep learning tasks, yet they exhibit different behaviors in terms of generalization. Some prior works [9, 46] show that Adam often results in worse generalization than SGD due to its susceptibility to sharp minima. In contrast, SGD tends to find flatter minima more conducive to robust generalization. This distinction highlights the importance of the loss surface's geometry, as flatter minima are generally associated with improved generalization performance, especially in out-of-distribution settings. In light of this, recent optimization strategies have been explicitly designed to address the sharpness of minima. Sharpness-Aware Minimization (SAM) [7] represents a key development in this direction, where the objective is to minimize sharpness and seek flatter regions of the loss landscape. SAM and its variants [46, 51] have demonstrated notable improvements in generalization, particularly in distribution. Recent improvements such as GSAM [51], which introduces a surrogate gap to better distinguish between sharp and flat minima, further enhance generalization + +by guiding the model toward regions with both low sharpness and low loss. Our proposed LEAwareSGD differs from these optimizers by dynamically adjusting the learning rate based on the LE, allowing the model toward the edge of chaos and exploring a broader parameter space for improving generalization capability. + +# 2.3. Dynamical Systems in Deep Learning + +The application of dynamical systems theory to deep learning has introduced new perspectives on neural network training, particularly in understanding stability, and generalization [19, 29]. By modeling neural networks as discrete-time systems, researchers have used concepts such as chaos theory to enhance generalization, especially within residual networks via flow maps [17]. This perspective underscores the significance of training near the edge of chaos, a state that balances stability and adaptability [44, 45]. Lyapunov exponent (LE) is a key tool to quantitatively measure the model's training stability, particularly in recurrent architectures [8, 32]. However, existing methods have not fully leveraged LE for direct control of training under domain shifts, particularly in the context of adversarial data augmentation for domain generalization tasks. Our proposed LEAwareSGD addresses this limitation by using LE to dynamically guide the optimization process, effectively enhancing model generalization capability. + +# 3. Methodology + +# 3.1. Overview of LEAwareSGD + +We propose LEAwareSGD, a novel optimization method for adversarial data augmentation to enhance generalization by exploiting the edge of chaos. In this approach, the LE serves as a metric to identify and explore regions in the parameter space where generalizable features are more likely to emerge. + +# 3.2. LE-Based Model Perturbation + +The LE quantifies the rate at which small perturbations in model parameters evolve during training, offering insights into the stability of the optimization process. Here, we denote the model parameters $\theta_{t}$ at iteration $t$ and the standard gradient descent update can be formulated as follows: + +$$ +\theta_ {t + 1} = \theta_ {t} - \eta_ {t} \nabla L (\theta_ {t}), \tag {1} +$$ + +where $\eta_t$ denotes the learning rate and $L(\theta_t)$ represents the loss function. + +To quantify how perturbations evolve during training, we introduce an initial perturbation $\delta \theta_0$ to the model parameters. Parameters after perturbation is $\widetilde{\theta}_t\coloneqq \theta_t + \delta \theta_t$ , where $\delta$ is a tiny real-valued scalar hyperparameter. The perturbation at step $t + 1$ propagates according to the following formula: + +$$ +\widetilde {\theta} _ {t + 1} = \widetilde {\theta} _ {t} - \eta_ {t} \nabla L (\widetilde {\theta} _ {t}). \tag {2} +$$ + +Combining Equation 1 and Equation 2, we get the following update for the perturbation: + +$$ +\delta \theta_ {t + 1} = \delta \theta_ {t} - \eta_ {t} (\nabla L (\widetilde {\theta_ {t}}) - \nabla L (\theta_ {t})). \tag {3} +$$ + +To further analyze the propagation of this perturbation, we perform a first-order Taylor expansion of the loss function $L(\widetilde{\theta_t})$ at $\theta_{t}$ and calculate the gradient: + +$$ +\nabla L \left(\widetilde {\theta} _ {t}\right) = \nabla L \left(\theta_ {t}\right) + H \left[ L \left(\theta_ {t}\right) \right] \delta \theta_ {t} + o \left(\| \delta \theta_ {t} \| ^ {2}\right), \tag {4} +$$ + +where $H[L(\theta_t)]$ is the Hessian matrix, and $o(\| \delta \theta_t\|^2)$ represents higher-order terms. Ignoring these higher-order terms, and substituting Equation 4 into Equation 3, we approximate the evolution of the perturbation as: + +$$ +\delta \theta_ {t + 1} = \left(I - \eta_ {t} H \left[ L \left(\theta_ {t}\right) \right]\right) \delta \theta_ {t}. \tag {5} +$$ + +By recursively applying this relationship, we obtain the propagation of the perturbation over time: + +$$ +\begin{array}{l} \delta \theta_ {t} = \left(I - \eta_ {t - 1} H \left[ L \left(\theta_ {t - 1}\right) \right]\right) \left(I - \eta_ {t - 2} H \left[ L \left(\theta_ {t - 2}\right) \right]\right) \tag {6} \\ \cdot \cdot \cdot (I - \eta_ {0} H [ L (\theta_ {0}) ]) \delta \theta_ {0}. \\ \end{array} +$$ + +The definition of the Lyapunov exponent is: + +$$ +\mathrm {L E} = \lim _ {t \rightarrow \infty} \frac {1}{t} \ln \left(\frac {\| \delta \theta_ {t} \|}{\| \delta \theta_ {0} \|}\right) \tag {7} +$$ + +By substituting Equation 6 into the definition of LE, we can derive the connection between the model's learning rate $\eta$ and LE: + +$$ +\mathrm {L E} \geq \lim _ {t \rightarrow \infty} \frac {1}{t} \sum_ {\substack {i = 0\\t - 1}} ^ {t - 1} \ln (1 - \eta_ {i} \| H [ L (\theta_ {i}) ] \|) \tag{8} +$$ + +$$ +\mathrm {L E} \leq \lim _ {t \rightarrow \infty} \frac {1}{t} \sum_ {i = 0} ^ {t - 1} \ln \left(\| I - \eta_ {i} H [ L (\theta_ {i}) ] \|\right), +$$ + +where the LE is determined by the learning rate $\eta_{i}$ and the Hessian matrix $H[L(\theta_i)]$ + +# 3.3. LE-Guided Optimization + +A positive LE indicates the exponential growth of perturbations, signifying instability, while a negative LE suggests convergence and stability. However, extreme stability or instability can lead to either overfitting or convergence difficulties. Our approach aims to guide model parameters toward an LE value close to zero but slightly negative, positioning the system near the edge of chaos. This critical state balances stability with adaptability, enhancing the model's generalization capacity. To achieve this, we dynamically adjust the learning rate to amplify perturbations, thereby steering the LE toward closer to the edge of chaos. Specifically, the innovation of the proposed LEAwareSGD lies in + +its learning rate adjustment based on LE dynamics, allowing the model to adapt effectively across training phases. We modulate the learning rate according to changes in the LE, $\Delta \mathrm{LE}_t = \mathrm{LE}_t - \mathrm{LE}_{t-1}$ . + +$$ +\eta_ {t + 1} = \eta_ {t} \cdot \exp (- \beta \cdot \Delta \mathrm {L E} _ {t}) \quad \text {i f} \quad \Delta \mathrm {L E} _ {t} > 0 \tag {9} +$$ + +where $\beta$ is a weighting parameter that controls the sensitivity of the learning rate adjustment. When $\Delta \mathrm{LE}_t > 0$ , indicating that the model is approaching regions near the edge of chaos where generalizable features are more likely to emerge, we reduce the learning rate to allow for deeper exploration of these regions. This approach ensures that the model explores parameter regions with strong generalization potential while avoiding over-stabilization that could hinder the learning process. When $\Delta \mathrm{LE}_t \leq 0$ , we leave the learning rate unchanged, maintaining the current exploration rate. + +# 3.4. LE-Based Adversarial Data Augmentation + +To further enhance the model's robustness to domain shifts, we integrate LEAwareSGD with adversarial data augmentation. Adversarial data augmentation generates samples that simulate domain shifts, exposing the model to a broader range of data variations. Given a model $f_{\theta}$ with parameters $\theta$ , we define adversarial samples as transformations of the original data. Let $\tau(x; \omega)$ denote a set of semantic transformations parameterized by $\omega$ . The joint optimization objective is: + +$$ +\begin{array}{l} \min _ {\theta} \max _ {\omega} \mathbb {E} _ {(x, y) \sim \mathcal {D} _ {S}} \left[ \ell (\theta ; \tau (x; \omega), y) - \lambda d _ {\theta} (\tau (x; \omega), x) \right] \\ + \frac {\gamma}{2} \cdot \| \theta \| _ {2} ^ {2}, \tag {10} \\ \end{array} +$$ + +where $\ell (\theta ;\cdot)$ is the prediction loss, $\lambda$ balances the adversarial loss and feature consistency, and $d_{\theta}(\cdot)$ represents the feature distance between the original and transformed samples. Here, the regularization term $\frac{\gamma}{2}\cdot \| \theta \| _2^2$ denotes a weight decay, ensuring that the Hessian matrix remains approximately positive definite. This regularization encourages LE to be negative, facilitating the stability of model training. Note that this adversarial training procedure is guided by LE-based optimization, which encourages the model to generalize to unseen domains by learning robust features invariant to domain-specific perturbations. + +# 3.5. Training Procedure + +The training procedure involves iterative steps alternating between adversarial example generation and parameter updates using LEAwareSGD. In each epoch, adversarial examples are first generated to simulate unseen domain shifts, fostering the model to learn invariant features. Subsequently, LEAwareSGD adjusts the model's parameters with + +a LE-guided learning rate to encourage the parameters' updating towards the edge of chaos. This LE-guided adaptation encourages the exploration of parameter regions conducive to generalization while controlling sensitivity to perturbations. The complete algorithm is shown in Algorithm 1. We also provide a detailed theoretical analysis of the proposed LEAwareSGD in Suppl. Sec-A.1. + +Algorithm 1 LEAwareSGD +1: Input: Initial parameters $\theta_0$ , learning rate $\eta_0$ , weighting parameter $\beta$ , number of iterations $N$ +2: Output: Optimized parameters $\theta^*$ +3: for $t = 0, 1, 2, \dots, N - 1$ do +4: Compute gradient $\nabla_{\theta} L(\theta_t)$ +5: Update parameters: $\theta_{t+1} = \theta_t - \eta_t \nabla_{\theta} L(\theta_t)$ +6: Compute updated model perturbation using Eq. 6 +7: Calculate LE using Eq. 7 +8: Update $\Delta \mathrm{LE}_t = \mathrm{LE}_t - \mathrm{LE}_{t-1}$ +9: if $\Delta \mathrm{LE}_t > 0$ then +10: $\eta_{t+1} = \eta_t \cdot \exp(-\beta \cdot \Delta \mathrm{LE}_t)$ +11: else +12: $\eta_{t+1} = \eta_t$ +13: end if +14: end for +15: Return: $\theta^*$ + +# 4. Experiment + +# 4.1. Datasets and Settings + +Datasets. We evaluate our method on three widely-used domain generalization benchmarks: PACS [15], OfficeHome [31], and DomainNet [24], arranging them in increasing levels of difficulty to comprehensively test the model's robustness across varying domain shifts. PACS consists of four domains, namely, Photo (P), Art Painting (A), Cartoon (C), and Sketch (S), comprising 9,991 images from 7 categories. OfficeHome includes four domain, namely, Art (A), Clipart (C), Product (P), and Real-World (R), with 15,500 images across 65 classes. Compared to PACS, OfficeHome introduces greater diversity in both object categories and domain styles. DomainNet is the most challenging, containing six domains, namely, Real (R), Infograph (I), Clipart (C), Painting (P), Quickdraw (Q), and Sketch (S), with 596,010 images from 345 classes. The substantial domain shifts and the high number of classes in this dataset present considerable challenges for domain generalization. + +Implementation details. In all experiments, we employed ResNet-18 [11] as the backbone model and resized all input images to $224 \times 224$ . The network was trained with LEAwareSGD, setting the momentum to 0.9 and $\beta$ to 1e-1. To ensure a fair comparison, we excluded the test-time learning stage of PSDG[41] in all of our experiments. For PACS and OfficeHome datasets, we follow the standard + +protocol of using one domain as the source domain and the remaining three domains as the target domains [47, 48]. For the PACS dataset, we adopted a batch size of 16, $\gamma$ of 5e-4, and an initial learning rate of 5e-4, training for 50 epochs. For the OfficeHome dataset, we used a batch size of 32, set $\gamma$ to 1e-4, and initialized the learning rate at 1e-4, training for 50 epochs. For experiments conducted on the DomainNet dataset, we follow [37, 48] and use Real as the source domain and the remaining five domains as the target domains. The batch size was increased to 128, $\gamma$ was set to 1e-5, and the initial learning rate was set to 1e-3, with a training duration of 200 epochs. For hyperparameter tuning, we fix all training settings (e.g., random seed, learning rate) and tune only $\beta$ and $\gamma$ , with the best setting shown in Fig.2. All experiments were performed on an NVIDIA A100 GPU, and each experiment was repeated 10 times with different random seeds. The results are reported as the average accuracy alongside the standard deviation. + +# 4.2. Comparison with SDG Approaches + +Results on PACS. This adjustment aligns the evaluation setup with our method, which operates solely during the training phase. The evaluation results on the PACS dataset, presented in Table 1, indicate that our method achieves the highest average accuracy of $69.46\%$ , outperforming all compared methods. Specifically, our model achieves the highest performance gains in both the A and P domains, with classification accuracies of $79.17\%$ and $65.05\%$ , respectively. Our approach outperforms the second-best method, PSDG $(67.14\%)$ , with an improvement of $2.32\%$ in average accuracy. This result underscores the effectiveness of our LE-guided adversarial data augmentation strategy, which enhances model generalization by facilitating the exploration of a broader parameter space and reducing overfitting to domain-specific features. + +
MethodsACPSAvg.
ERM70.4973.5645.9241.2157.80
MixUp [39]67.6067.1351.5947.0458.34
RSC [13]73.4075.9056.2041.6061.78
ADA [33]71.5676.8452.3643.6661.11
ME-ADA [47]71.5276.8346.2246.3260.22
SAGM [35]65.9470.1042.1453.6557.96
RandConv [38]73.6874.8855.4246.7762.69
MetaCNN [34]76.5074.5056.7055.7065.80
L2D [36]77.0076.5553.2254.4465.30
PDEN [16]76.4373.8758.5253.9265.69
PSDG [41]77.8875.2455.4959.9667.14
AdvST [48]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
+ +Results on OfficeHome. Table 2 shows the classification accuracy comparison of our method with existing baselines on the OfficeHome dataset. Our proposed method consistently outperforms other approaches, achieving the highest accuracy across all domains (A, C, P, and R) with an average accuracy of $54.38\%$ . Notably, our method surpasses the previous best method, AdvST, which obtained an average accuracy of $52.60\%$ , with an improvement of $1.78\%$ . This gain highlights the effectiveness of our LE-guided optimization, which facilitates the exploration of a broader parameter space and enhances the model's ability to generalize to unseen domains. + +Table 1. Classification accuracy (%) comparison on the PACS dataset. The best results are in bold. + +
MethodsACPRAvg.
ERM42.7439.3139.8652.5043.60
MixUp [39]43.6940.6838.5652.1243.76
SagNet [21]49.8443.0141.4255.6147.47
RandConv [38]45.7940.2037.4552.8744.08
Fish [28]45.1039.7436.6752.0943.40
ADA [33]45.0736.0744.4353.4444.75
ME-ADA [47]45.2745.7938.6551.6945.35
GSAM [51]28.9741.4041.7251.5040.90
SAGM [35]47.4243.2741.3654.9746.76
Mixstyle [49]51.1948.7346.8555.8850.66
PCGrad [42]49.9146.6146.1156.3149.74
PSDG [41]42.8046.4044.3154.6847.05
AdvST [48]51.3252.0848.8958.1152.60
Ours52.8955.1851.6958.9554.38±0.30
+ +Table 2. Classification accuracy (%) comparison on the Office-Home dataset. The best results are in bold. + +Results on DomainNet. Table 3 shows the classification accuracy comparison on the DomainNet dataset. Our proposed method achieves the highest average accuracy of $28.15\%$ , surpassing all other methods, including the second-best method, AdvST, which reaches an average accuracy of $27.22\%$ . Notably, our method performs best in the C and S domains, with accuracies of $44.67\%$ and $33.74\%$ , respectively. In the Q domain, our method achieves an accuracy of $6.70\%$ , slightly lower than SimDE's $6.85\%$ . This may be due to the highly abstract nature of the Quickdraw domain, which can benefit from specialized augmentation techniques to better capture its distinct features. Despite this, our method's overall performance advantage demonstrates the effectiveness of the LE-guided adversarial data augmentation approach in handling the diversity and complexity of DomainNet. + +# 4.3. Comparison with Optimization Approaches + +To further validate the effectiveness of our proposed LEAwareSGD optimizer, we conducted a comparison with four widely used optimization approaches, including Adam [14], AdamW [20], RMSprop [30], and SGD [26], on the PACS and OfficeHome datasets. As shown in Table 4, + +
MethodsPICSQAvg.
ERM38.0513.3137.8926.263.3623.77
MixUp [39]38.6013.9238.0126.013.7124.05
CutOut [5]38.3013.7038.4026.203.7024.06
AugMix [12]40.8013.9041.7029.806.3026.50
RandAug [3]41.3013.5741.1130.405.3126.34
ADA [33]38.2013.8040.2024.804.3024.26
ME-ADA [47]38.4813.5640.6626.054.4224.63
ACVC [4]41.3012.9042.8030.906.6026.90
PDEN [16]38.4511.2538.9931.715.5825.20
SimDE [37]39.9612.9141.7333.466.8526.98
PSDG [41]41.1013.0041.2931.104.8926.28
AdvST [48]42.4014.9041.7031.006.1027.22
Ours42.7715.9144.6733.746.7028.15 ±0.61
+ +our proposed optimizer achieves the highest average accuracy of $69.46\%$ , surpassing all other optimizers on the PACS dataset. In contrast, widely used optimizers such as Adam, AdamW, and RMSprop perform significantly worse, with average accuracies of $66.43\%$ , $66.83\%$ , and $62.34\%$ , respectively. This disparity likely stems from the reliance of these adaptive optimizers on variable learning rates, which may cause overfitting to domain-specific features and reduce their generalization ability in domain generalization tasks. On the other hand, both our LEAwareSGD and standard SGD achieve substantially higher accuracies. While SGD achieves an average accuracy of $67.06\%$ , our LEAwareSGD optimizer further improves this to $69.46\%$ . This enhancement underscores the advantage of integrating LE guidance into the optimization process, which enables the model to better capture generalizable features and avoid overfitting to domain-specific characteristics. We present the experimental results on the OfficeHome dataset in Suppl. Sec-C.2, while analyses of optimizers' performance under different learning rates are provided in Suppl. Sec-C.5. + +Table 3. Classification accuracy (%) comparison on the Domain-Net dataset. The best results are in bold. + +
MethodsACPSAvg.
Adam [14]76.5271.1564.0653.9866.43
AdamW [20]76.6871.9362.0356.6866.83
RMSprop [30]71.6271.0859.0947.5562.34
SGD [26]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
+ +# 4.4. Extension to Multi-Source DG Approaches + +To further validate the scalability of our solution, we perform experiments under the leave-one-out setting of general domain generalization (DG) with ResNet-18 as the + +backbone, where three domains are combined for training and the remaining domain is reserved for evaluation. Table 5 presents the results on the PACS dataset, showing that our method delivers at least comparable performance, highlighting its versatility. This improvement underscores the effectiveness of our LE-guided optimization, which enables a broader exploration of the parameter space and enhances the model's generalization to unseen domains. + +Table 4. Comparison results $(\%)$ using different optimizers on the PACS dataset. The best results are in bold. + +
MethodsACPSAvg.
RSC [13]82.8179.7493.9580.8584.34
ADA [33]82.8178.3395.6375.2983.02
ME-ADA [47]77.8878.5895.3378.0782.47
Mixstyle [49]83.1179.4396.3172.9582.95
L2D [36]81.4479.5695.5180.5884.27
AdvST [48]78.6972.9791.6280.9181.05
Ours81.3578.8895.4582.3984.52
+ +Table 5. Classification accuracy (\%) comparison for the leave-one-domain-out using the ResNet-18 backbone on the PACS dataset. The column name denotes the test domain, with the other three used for training. The best results are in bold. + +# 4.5. Additional Analysis + +Performance comparison using different backbone networks. To further validate the universality of our solution, we conduct experiments with different ResNet backbones, including ResNet-34, ResNet-50, ResNet-101, and ResNet-152, on the PACS and OfficeHome datasets. As shown in Table 6 on the PACS dataset, our approach reaches a performance of $73.68\%$ on ResNet-34, $71.37\%$ on ResNet-50, $74.32\%$ on ResNet-101, and $75.34\%$ on ResNet-152. These results demonstrate the consistent effectiveness of our method across various backbone architectures, highlighting its superior accuracy, versatility, and robustness. The experimental results for the OfficeHome dataset are presented in Suppl. Sec-C.6. + +Performance comparison with limited training data. We compare the performance of our proposed method with AdvST on the PACS and OfficeHome datasets under varying source domain data ratios (10%, 20%, and 50%). As shown in Tables 7 on the PACS dataset, our method consistently outperforms AdvST across all data ratios, demonstrating robust generalization even with limited source domain data for training. At the 10% data ratio, our method achieves an average accuracy of 58.73%, showing a notable improvement of 9.47% over AdvST's 49.26%. With the 20% data ratio, our approach reaches an average accuracy of 61.78%, surpassing AdvST by 8.23%. At the 50% data ratio, the performance gap narrows slightly, with our method achieving an average accuracy of 66.44%, which is 6.26% higher than the 60.18% achieved by AdvST. The experimental results on the OfficeHome dataset are provided in Suppl. Sec-C.7. These findings suggest that our + +
BackboneMethodsACPSAvg.
ResNet34PSDG79.4680.7957.4358.2268.98
AdvST79.1676.6765.2860.5570.42
Ours81.8981.8565.6565.3173.68
ResNet50PSDG79.8579.7258.0865.6870.83
AdvST79.6575.0562.6255.7568.27
Ours81.3278.7065.0860.3971.37
ResNet101PSDG79.8479.9759.1770.7972.44
AdvST80.9375.3566.2958.7070.32
Ours82.9580.7968.5764.9574.32
ResNet152PSDG82.9281.5462.3871.2574.52
AdvST80.7679.0265.9061.9971.92
Ours83.4482.1967.6368.1175.34
+ +method effectively reduces overfitting to specific domains with limited training data, particularly in low-data settings, by encouraging broader exploration of the parameter space to capture generalizable features. + +Table 6. Comparison results (\%) using different ResNet series backbones on the PACS dataset. The best results are in bold. + +
RatioMethodsACPSAvg.ΔAvg.
10%AdvST58.7456.9440.6240.7349.26
Ours65.4364.4349.4055.6658.73+9.47
20%AdvST61.4864.3143.4244.9853.55
Ours68.8168.7853.2856.2761.78+8.23
50%AdvST70.2772.7449.5248.2060.18
Ours75.8873.5561.6054.7266.44+6.26
+ +Computational complexity analysis. To evaluate the computational efficiency of our method, we compared its training time cost with three adversarial data augmentation methods on the PACS and OfficeHome datasets. As shown in Table 8, our approach requires an average training time of 1.99 hours on PACS, slightly higher than AdvST (1.90 hours) but comparable to ADA (2.13 hours) and ME-ADA (2.10 hours). This marginal increase is due to the Lyapunov-guided optimization, which enhances model robustness and generalization. On OfficeHome, our method completes training in 1.19 hours, surpassing AdvST (0.75 hours) but remaining faster than ADA (1.95 hours) and ME-ADA (1.90 hours). While our method introduces a slight time increase compared to AdvST, it achieves significant performance improvements, demonstrating a favorable balance between training time and model performance. + +Generalizability analysis with LEAwareSGD using different adversarial data augmentation approaches. To assess the impact of LEAwareSGD on the generalization performance of adversarial data augmentation methods, we integrated it with ADA, ME-ADA, and AdvST + +Table 7. Performance comparison (%) using different data ratios on PACS dataset. + +
DatasetMethodsACPSAvg.
PACSADA1.802.101.403.202.13
ME-ADA1.802.101.303.202.10
AdvST1.581.871.402.761.90
Ours1.622.021.352.981.99
OfficeHomeMethodsACPRAvg.
ADA1.202.102.202.301.95
ME-ADA1.201.902.202.301.90
AdvST0.750.740.750.760.75
Ours1.101.171.231.271.19
+ +Table 8. Training time (hours) on PACS and OfficeHome datasets. + +![](images/b8d74aed33b4c6a386436373fb02e3ab27dcfd35852fddcb1aaad349992cf67f.jpg) +(a) Parameter $\beta$ on the PACS dataset + +![](images/3303dbc38c79211304f06470f21e1fea538b26a250935f19a24aa0e3984c1814.jpg) +(b) Parameter $\beta$ on the OfficeHome dataset + +![](images/41b79aa285f528be448f92d0f9763bdbfd64e02a211d52b42704b7ef383a66f8.jpg) +(c) Parameter $\gamma$ on the PACS dataset +Figure 2. Parameter sensitivity study. + +![](images/660838a09ed295292ab4cb24909166b78664bc36db40e39913bdad9aee2cdc2c.jpg) +(d) Parameter $\gamma$ on the OfficeHome dataset + +on the PACS and OfficeHome datasets. Table 9 shows consistent improvements across all methods, underscoring LEAwareSGD's effectiveness in enhancing model generalization. On the PACS dataset, integrating LEAwareSGD with AdvST led to a $2.40\%$ increase in average accuracy, reaching $69.46\%$ . This improvement highlights LEAwareSGD's ability to facilitate more effective parameter space exploration, enhancing AdvST's generalization capacity. For ADA and ME-ADA, LEAwareSGD achieved gains of $0.52\%$ and $2.30\%$ , respectively, with ME-ADA reaching an average accuracy of $62.52\%$ . On the OfficeHome dataset, LEAwareSGD also improved the performance of all methods. ME-ADA saw the highest gain, with a $7.00\%$ increase over ME-ADA alone, achieving an average accuracy of $52.35\%$ . ADA with LEAwareSGD increased by $5.15\%$ , reaching an average accuracy of $49.90\%$ . For AdvST, which already demonstrated strong performance, LEAwareSGD provided an additional $1.78\%$ gain, bringing the average accuracy to $54.38\%$ . These findings demonstrate that LEAwareSGD enhances the generalization + +capacity of adversarial data augmentation techniques by enabling broader exploration of the parameter space. This exploration can facilitate learning more generalizable features across unseen domains, underscoring the effectiveness of LEAwareSGD as a robust optimization framework for SDG. + +
DatasetMethodsACPSAvg.ΔAvg.
PACSADA73.7471.6540.7160.3961.62
\( \mathrm{ADA}^{\dagger} \)71.8269.1551.9455.6662.14+0.52
ME-ADA71.5276.8346.2246.3260.22
\( \mathrm{ME-ADA}^{\dagger} \)72.8869.1647.9860.0562.52+2.30
AdvST76.6574.9262.4754.1867.06
\( \mathrm{AdvST}^{\dagger} \)79.1777.1665.0557.7869.46+2.40
Office HomeMethodsACPRAvg.ΔAvg.
ADA45.0736.0744.4353.4444.75
\( \mathrm{ADA}^{\dagger} \)50.6948.5246.1154.2849.90+5.15
ME-ADA45.2745.7938.6551.6945.35
\( \mathrm{ME-ADA}^{\dagger} \)53.6150.0949.1756.5152.35+7.00
AdvST51.3252.0848.8958.1152.60
\( \mathrm{AdvST}^{\dagger} \)52.8955.1851.6958.9554.38+1.78
+ +Note: $\dagger$ indicates methods integrated with LEAwareSGD. + +Table 9. Performance improvement (%) of adversarial data augmentation approaches with LEAwareSGD on the PACS and OfficeHome datasets. + +Parameter sensitivity analysis. To assess the impact of hyperparameters on the performance of LEAwareSGD, we conducted a sensitivity analysis on the parameters beta $\beta$ and $\gamma$ over the PACS and OfficeHome datasets, as shown in Figure 2. Figures 2a and 2b show that increasing $\beta$ from 1e-4 to 1.0 resulted in fluctuating accuracy, with peak performance observed at $\beta = 1\mathrm{e} - 3$ on PACS and $\beta = 1\mathrm{e} - 2$ on OfficeHome. This result suggests an optimal range between 1e-3 and 1e-2. + +As illustrated in Figure 2c and 2d, for the parameter $\gamma$ in Equation 10, which controls the strength of regularization, the best accuracy on the PACS dataset was achieved at $\gamma = 5\mathrm{e - }4$ . Increasing $\gamma$ beyond 5e-4 caused slight declines in accuracy, indicating that higher regularization may overly restrict the model. On the OfficeHome dataset, $\gamma = 1\mathrm{e - }5$ yielded the best accuracy across most domains, while higher values generally led to reduced performance. + +Comparison of LE results of different methods on the PACS dataset. To evaluate our proposed method's stability and generalization performance, we compare the LE values obtained from different adversarial data augmentation methods on the PACS dataset. Figure 3 shows that we track LE dynamics across four domains: Art Painting, Cartoon, Photo, and Sketch. Our method consistently yields LE values closer to zero, indicating that the model operates near the edge of chaos, where an optimal balance between sta + +![](images/eb46c638218fd55a7d3c6cc88aa0e4f9bb91b369ea61eebcfd392bbe7c00a66c.jpg) +(a) Trained on Art Painting domain + +![](images/621e693643a1dc9051e61af3c73388c3f51970f17193486b94c78cb3d4f3c14a.jpg) +(b) Trained on Cartoon domain + +![](images/301cb89cdfdd85dbe72653782f3193ddc47584d291cfae76a287af06fc982da0.jpg) +(c) Trained on Photo domain + +![](images/ed2c0b4c9754e345ea61f72cffd82292bd53c39af3e5cfd723cd8d4cab684d11.jpg) +(d) Trained on Sketch domain +Figure 3. LE dynamics over training epochs across four domains (Art Painting, Cartoon, Photo, Sketch) on the PACS dataset. + +bility and adaptability is achieved. + +In the Art Painting domain (Figure 3a), our method maintains LE values closer to zero throughout training compared to ADA, ME-ADA, and AdvST, suggesting a better balance between stability and adaptability, leading to improved generalization. In contrast, ADA exhibits the lowest LE values and large fluctuations, indicating instability and convergence to local optima. Similar results can be derived from the three remaining domains (Figure 3b-3d), where our method consistently stabilizes near the edge of chaos, supporting better generalization. Comparison methods, by contrast, show greater fluctuations or lower LE values, indicating tendencies towards overfitting and limited adaptability. These results reveal that our method consistently achieves LE values near zero across all domains and operates effectively near the edge of chaos. Additional detailed experimental analyses are provided in Suppl. Sec-A.2. + +# 5. Conclusion + +In this work, we introduce LEAwareSGD, a novel LE-guided optimization integrated with an adversarial data augmentation method to address the challenge of unseen domain shifts in SDG. By leveraging principles from dynamical systems theory, our approach encourages model training near the edge of chaos, enabling the exploration of a broader parameter space and substantially enhancing generalization to new domains. LEAwareSGD dynamically adjusts the learning rate based on LE values, encouraging the model to capture more generalizable features. Experiments on three standard SDG benchmarks demonstrate that LEAwareSGD outperforms state-of-the-art SDG methods, underscoring its effectiveness. A promising future work is to investigate its scalability to larger datasets and more complex domain generalization tasks. + +# Acknowledgments + +This work is supported in part by the National Natural Science Foundation of China (No. 62276038), and in part by the Chongqing Municipal Education Commission (No. KJZD-M202400603, No. KJQN202500612), and in part by the Project of Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism (No. H2023009). + +# References + +[1] Kathleen T Alligood, Tim D Sauer, James A Yorke, and David Chillingworth. Chaos: an introduction to dynamical systems. SIAM Review, 40(3):732-732, 1998. 1 +[2] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11400-11410, 2023. 2 +[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6 +[4] Iike Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4174, 2022. 6 +[5] Terrance DeVries. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6 +[6] Ling Feng, Lin Zhang, and Choy Heng Lai. Optimal machine intelligence at the edge of chaos. arXiv preprint arXiv:1909.05176, 2019. 1 +[7] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2 +[8] Alejandro Guarneros-Sandoval, Mariana Ballesteros, Ivan Salgado, Julia Rodríguez-Santillán, and Isaac Chairez. Lyapunov stable learning laws for multilayer recurrent neural networks. Neurocomputing, 491:644-657, 2022. 3 +[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International Conference on Machine Learning, pages 1225-1234. PMLR, 2016. 2 +[10] Kazuko Hayashi. Chaotic nature of the electroencephalogram during shallow and deep anesthesia: From analysis of the lyapunov exponent. *Neuroscience*, 557:116-123, 2024. 1 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 4 +[12] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A + +simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019.6 +[13] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124-140, 2020. 5, 6 +[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 2, 5, 6 +[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 2, 4 +[16] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 1, 2, 5, 6 +[17] Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dynamical systems: An approximation perspective. Journal of the European Mathematical Society, 25(5):1671-1709, 2022. 3 +[18] Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hyperspectral image classification: An overview. IEEE Transactions on Geoscience and Remote Sensing, 57(9):6690-6709, 2019. 1 +[19] Guan-Horng Liu and Evangelos A Theodorou. Deep learning theory review: An optimal control and dynamical systems perspective. arXiv preprint arXiv:1908.10920, 2019. 3 +[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2, 5, 6 +[21] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 5 +[22] Yurii Nesterov. A method for solving the convex programming problem with convergence rate o (1/k2). In Proceedings of the USSR Academy of Sciences, pages 543-547, 1983. 2 +[23] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2009. 1 +[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 2, 4 +[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 2 +[26] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2, 5, 6 + +[27] Ling Shao, Fan Zhu, and Xuelong Li. Transfer learning for visual categorization: A survey. IEEE Transactions on Neural Networks and Learning Systems, 26(5):1019-1034, 2014. 1 +[28] Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937, 2021.5 +[29] Yang Tang, Jürgen Kurths, Wei Lin, Edward Ott, and Ljupco Kocarev. Introduction to focus issue: When machine learning meets complex systems: Networks, chaos, and nonlinear dynamics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(6), 2020. 3 +[30] Tijmen Tieleman and Geoffrey Hinton. Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn, 17, 2012. 2, 5, 6 +[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 2, 4 +[32] Ryan Vogt, Maximilian Puelma Touzel, Eli Shlizerman, and Guillaume Lajoie. On lyapunov exponents for rnns: Understanding information propagation using dynamical systems tools. Frontiers in Applied Mathematics and Statistics, 8: 818799, 2022. 3 +[33] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 5339-5349, 2018. 1, 2, 5, 6 +[34] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4682-4691, 2022. 5 +[35] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769-3778, 2023. 5 +[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 5, 6 +[37] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4798-4808, 2023. 1, 2, 5, 6 +[38] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual represent + +tation learning via random convolutions. arXiv preprint arXiv:2007.13003, 2020.5 +[39] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 5, 6 +[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Causality-inspired domain expansion network for single domain generalization. Knowledge-Based Systems, 301:112269, 2024. 1, 2 +[41] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3794-3805, 2024. 4, 5, 6 +[42] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 5 +[43] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE Transactions on Medical Imaging, 39(7):2531-2540, 2020. 2 +[44] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Edge of chaos as a guiding principle for modern neural network training. arXiv preprint arXiv:2107.09437, 2021. 1, 3 +[45] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Asymptotic edge of chaos as guiding principle for neural network training. International Journal of Artificial Intelligence and Robotics Research, 1(01):2350001, 2024. 3 +[46] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5189-5202, 2023. 2 +[47] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. Advances in Neural Information Processing Systems, 33:14435–14447, 2020. 1, 2, 5, 6 +[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21832-21840, 2024. 2, 5, 6 +[49] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision, 132 (3):822-836, 2024. 5, 6 +[50] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 1 +[51] Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training. arXiv preprint arXiv:2203.08065, 2022. 2, 5 \ No newline at end of file diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/images.zip b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f7fe198cd1d2345df6d1524c5ca5e5f4d08e96c --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a21cde961efddd02eceb802f374a3cdad30874bd90d3b949ae40b1e4e77eb2df +size 587945 diff --git a/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/layout.json b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4d0a68634aa3c4ff64fb923c8961a0652df9e96c --- /dev/null +++ b/2025/Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization/layout.json @@ -0,0 +1,9484 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "text", + "content": "Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "content": "Zuyu Zhang" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "content": " Ning Chen" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "content": " Yongshan Liu" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "content": " Qinghua Zhang" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "text", + "content": " Xu Zhang" + }, + { + "bbox": [ + 88, + 160, + 525, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 177, + 470, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 177, + 470, + 189 + ], + "spans": [ + { + "bbox": [ + 140, + 177, + 470, + 189 + ], + "type": "text", + "content": "1 School of Artificial Intelligence, Chongqing University of Posts and Telecommunications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 190, + 504, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 190, + 504, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 504, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 105, + 190, + 504, + 202 + ], + "type": "text", + "content": " Key Laboratory of Big Data Intelligent Computing, Chongqing University of Posts and Telecommunications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 205, + 509, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 205, + 509, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 205, + 509, + 217 + ], + "type": "text", + "content": "3 Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 219, + 478, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 219, + 478, + 231 + ], + "spans": [ + { + "bbox": [ + 129, + 219, + 478, + 231 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 129, + 219, + 478, + 231 + ], + "type": "text", + "content": " School of Mathematics and Statistics, Chongqing University of Posts and Telecommunications" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 284, + 297, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 297, + 547 + ], + "type": "text", + "content": "Single Domain Generalization (SDG) aims to develop models capable of generalizing to unseen target domains using only one source domain, a task complicated by substantial domain shifts and limited data diversity. Existing SDG approaches primarily rely on data augmentation techniques, which struggle to effectively adapt training dynamics to accommodate large domain shifts. To address this, we propose LEAwareSGD, a novel Lyapunov Exponent (LE)-guided optimization approach inspired by dynamical systems theory. By leveraging LE measurements to modulate the learning rate, LEAwareSGD encourages model training near the edge of chaos, a critical state that optimally balances stability and adaptability. This dynamic adjustment allows the model to explore a wider parameter space and capture more generalizable features, ultimately enhancing the model's generalization capability. Extensive experiments on PACS, OfficeHome, and DomainNet demonstrate that LEAwareSGD yields substantial generalization gains, achieving up to " + }, + { + "bbox": [ + 55, + 284, + 297, + 547 + ], + "type": "inline_equation", + "content": "9.47\\%" + }, + { + "bbox": [ + 55, + 284, + 297, + 547 + ], + "type": "text", + "content": " improvement on PACS in low-data regimes. These results underscore the effectiveness of training near the edge of chaos for enhancing model generalization capability in SDG tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 135, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 135, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 135, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 590, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 685 + ], + "type": "text", + "content": "Deep neural networks have shown remarkable progress across various tasks in standard supervised learning settings where the training and test data come from the same distribution [11, 18]. However, their performance often suffers significantly when encountering domain shift, a prevalent issue in real-world applications [23, 27, 50]. For instance, a model trained on DSLR images may struggle when tested on smartphone images due to differences in reso" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 260, + 555, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 295 + ], + "type": "text", + "content": "lution, noise, and other factors. These domain shifts are pervasive in real-world scenarios, emphasizing the need for models that can generalize well to unseen distributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 299, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 556, + 456 + ], + "type": "text", + "content": "Single Domain Generalization (SDG) methods aim to address these domain shifts when only a single source domain is available for training. Current SDG approaches primarily employ data augmentation techniques to generate diverse training samples [16, 25, 33, 37, 40, 47]. Methods such as adversarial data augmentation [25, 33, 47] generate perturbed samples that simulate potential domain shifts. Despite their contributions, as shown in Figure 1, these adversarial data augmentation approaches often fail to adequately capture the global structure of domain shifts in the parameter space due to their emphasis on localized perturbations. As a result, they may underexplore the parameter space, limiting generalization across diverse distributions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 459, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 698 + ], + "type": "text", + "content": "In this work, we draw inspiration from dynamical systems theory [1, 6] and propose a novel perspective of model perturbations based on the edge of chaos. This critical transition point maintains a balance between order and chaos in dynamical systems. In SDG, training models near the edge of chaos can help capture features that better generalize across domains by maintaining both stability and adaptability [44]. To quantitatively measure a model's degree of chaos, we employ the Lyapunov Exponent (LE) [10] as a metric to estimate the model's proximity to this edge of chaos. A high LE indicates chaotic behavior and sensitivity to perturbations, whereas a low LE suggests stable dynamics that may restrict the model's generalization capability. We frame the training of neural networks as a discrete-time dynamical system in the parameter space, where each parameter update represents a state transition in this system. By dynamically adjusting the LE, we aim to guide adversarial data augmentation in a way that encourages broader exploration of the parameter space, balancing stability and adaptability for better domain generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Building on this insight, we introduce LE-Aware" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 693, + 253, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 693, + 253, + 703 + ], + "spans": [ + { + "bbox": [ + 65, + 693, + 253, + 703 + ], + "type": "text", + "content": "*The corresponding author, email: zhangx@cqupt.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 65, + 703, + 297, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 703, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 65, + 703, + 297, + 712 + ], + "type": "text", + "content": "The code is available at: https://github.com/ChannySwift/LEAwareSGD." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "552" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 172, + 177 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 172, + 177 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 172, + 177 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 172, + 177 + ], + "type": "image", + "image_path": "12c7c405a8ef6d5dbd4c0792897b579e0750d9ebfe0f6525b14c5521c5dea037.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 181, + 129, + 190 + ], + "lines": [ + { + "bbox": [ + 99, + 181, + 129, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 181, + 129, + 190 + ], + "type": "text", + "content": "(a) ADA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 70, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 179, + 70, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 179, + 70, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 179, + 70, + 294, + 180 + ], + "type": "image", + "image_path": "d9c07935d20914b168c0256d8aadc6d5a960b59ee988473d7ff6bba8a4a3aede.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 181, + 259, + 190 + ], + "lines": [ + { + "bbox": [ + 215, + 181, + 259, + 190 + ], + "spans": [ + { + "bbox": [ + 215, + 181, + 259, + 190 + ], + "type": "text", + "content": "(b) ME-ADA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 206, + 171, + 312 + ], + "blocks": [ + { + "bbox": [ + 57, + 206, + 171, + 312 + ], + "lines": [ + { + "bbox": [ + 57, + 206, + 171, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 171, + 312 + ], + "type": "image", + "image_path": "8184d81b360efdb6cd29cd76aa754439114d749be2add7363d09ed663edc8c55.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 316, + 132, + 325 + ], + "lines": [ + { + "bbox": [ + 97, + 316, + 132, + 325 + ], + "spans": [ + { + "bbox": [ + 97, + 316, + 132, + 325 + ], + "type": "text", + "content": "(c) AdvST" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 335, + 295, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 411 + ], + "type": "text", + "content": "Figure 1. t-SNE visualization of parameter trajectories in initial epochs for a model trained on the PACS dataset [15]. Each point represents the model's parameters after an iteration. Comparison between (a) ADA [33], (b) ME-ADA [47], (c) AdvST [48], and (d) our method demonstrates that our approach can explore a broader parameter space, facilitating the learning of more generalizable features across unseen domains." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 179, + 207, + 294, + 312 + ], + "blocks": [ + { + "bbox": [ + 179, + 207, + 294, + 312 + ], + "lines": [ + { + "bbox": [ + 179, + 207, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 179, + 207, + 294, + 312 + ], + "type": "image", + "image_path": "0a690f1defb557c6b73500c326de9c3f45cb8f21c95fafe864469a82448cd7f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 316, + 251, + 325 + ], + "lines": [ + { + "bbox": [ + 223, + 316, + 251, + 325 + ], + "spans": [ + { + "bbox": [ + 223, + 316, + 251, + 325 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": "Stochastic Gradient Descent (LEAwareSGD), a novel optimizer that dynamically adjusts the learning rate based on LE computations. By integrating adversarial data augmentation with LE-aware optimization, LEAwareSGD generates diverse perturbations that enhance the model's ability to generalize to unseen domains while guiding it to operate near the edge of chaos. Unlike conventional optimizers [14, 20, 26, 30] without considering the system's state, LEAwareSGD leverages LE as a feedback mechanism, adjusting the learning rate in response to the system's state and enabling the model to explore the parameter space effectively for capturing more generalizable features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 581, + 253, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 581, + 253, + 592 + ], + "spans": [ + { + "bbox": [ + 67, + 581, + 253, + 592 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": "- We conceptualize the training of neural networks as a discrete-time dynamical system and leverage the LE to assess the system's proximity to the edge of chaos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We propose LEAwareSGD, an optimizer for adversarial data augmentation that dynamically adjusts the learning rate based on LE, guiding the training process near the edge of chaos to enhance generalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "- Extensive experiments on three standard SDG benchmarks, including PACS [15], OfficeHome [31], and DomainNet [24], validate the effectiveness of the proposed" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 545, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 545, + 84 + ], + "type": "text", + "content": "LEAwareSGD to improve cross-domain generalization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 97, + 400, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 97, + 400, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 97, + 400, + 109 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 118, + 477, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 477, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 477, + 129 + ], + "type": "text", + "content": "2.1. Single Domain Generalization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 136, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 411 + ], + "type": "text", + "content": "Single domain generalization seeks to improve model robustness by enabling generalization from a single source domain to unseen target domains, eliminating the need for target data during training. Most SDG methods rely on data augmentation [25, 43] to increase diversity within the source domain, while generative model-based approaches [16, 37, 40] synthesize samples to extend the source distribution. Although these generative techniques introduce sample variety, they are often computationally demanding and complex to train, frequently producing samples that lack sufficient quality and diversity. This limitation can cause models to overfit to synthetic data rather than enhancing generalization to real-world domains. Adversarial data augmentation [2, 47, 48] offers an alternative by generating adversarial examples that mimic domain shifts to improve model robustness. However, these examples often introduce only localized perturbations, failing to explore the parameter space extensively. In contrast, our proposed LE-guided optimization introduces a dynamic SDG framework that modulates model sensitivity to perturbations. This LE-aware optimization promotes broader exploration of the parameter space, effectively capturing complex domain shifts and improving cross-domain generalization capability." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 420, + 512, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 512, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 512, + 433 + ], + "type": "text", + "content": "2.2. Optimizers in Domain Generalization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 715 + ], + "type": "text", + "content": "Optimizers play a critical role in model generalization, with different strategies yielding varied outcomes depending on the loss landscape's properties. Widely-used optimizers like Stochastic Gradient Descent (SGD) [22] and Adam [14] are commonly used in deep learning tasks, yet they exhibit different behaviors in terms of generalization. Some prior works [9, 46] show that Adam often results in worse generalization than SGD due to its susceptibility to sharp minima. In contrast, SGD tends to find flatter minima more conducive to robust generalization. This distinction highlights the importance of the loss surface's geometry, as flatter minima are generally associated with improved generalization performance, especially in out-of-distribution settings. In light of this, recent optimization strategies have been explicitly designed to address the sharpness of minima. Sharpness-Aware Minimization (SAM) [7] represents a key development in this direction, where the objective is to minimize sharpness and seek flatter regions of the loss landscape. SAM and its variants [46, 51] have demonstrated notable improvements in generalization, particularly in distribution. Recent improvements such as GSAM [51], which introduces a surrogate gap to better distinguish between sharp and flat minima, further enhance generalization" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "553" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "by guiding the model toward regions with both low sharpness and low loss. Our proposed LEAwareSGD differs from these optimizers by dynamically adjusting the learning rate based on the LE, allowing the model toward the edge of chaos and exploring a broader parameter space for improving generalization capability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 254, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 254, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 254, + 164 + ], + "type": "text", + "content": "2.3. Dynamical Systems in Deep Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 384 + ], + "type": "text", + "content": "The application of dynamical systems theory to deep learning has introduced new perspectives on neural network training, particularly in understanding stability, and generalization [19, 29]. By modeling neural networks as discrete-time systems, researchers have used concepts such as chaos theory to enhance generalization, especially within residual networks via flow maps [17]. This perspective underscores the significance of training near the edge of chaos, a state that balances stability and adaptability [44, 45]. Lyapunov exponent (LE) is a key tool to quantitatively measure the model's training stability, particularly in recurrent architectures [8, 32]. However, existing methods have not fully leveraged LE for direct control of training under domain shifts, particularly in the context of adversarial data augmentation for domain generalization tasks. Our proposed LEAwareSGD addresses this limitation by using LE to dynamically guide the optimization process, effectively enhancing model generalization capability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 393, + 138, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 138, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 138, + 407 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 205, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 205, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 205, + 425 + ], + "type": "text", + "content": "3.1. Overview of LEAwareSGD" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "content": "We propose LEAwareSGD, a novel optimization method for adversarial data augmentation to enhance generalization by exploiting the edge of chaos. In this approach, the LE serves as a metric to identify and explore regions in the parameter space where generalizable features are more likely to emerge." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 221, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 221, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 221, + 521 + ], + "type": "text", + "content": "3.2. LE-Based Model Perturbation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "text", + "content": "The LE quantifies the rate at which small perturbations in model parameters evolve during training, offering insights into the stability of the optimization process. Here, we denote the model parameters " + }, + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "text", + "content": " at iteration " + }, + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 526, + 296, + 586 + ], + "type": "text", + "content": " and the standard gradient descent update can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 593, + 295, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 593, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 126, + 593, + 295, + 605 + ], + "type": "interline_equation", + "content": "\\theta_ {t + 1} = \\theta_ {t} - \\eta_ {t} \\nabla L (\\theta_ {t}), \\tag {1}", + "image_path": "58687385fda03cec0893a3e05e1484bf078c32a55058f9e95e347940390248ba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "inline_equation", + "content": "\\eta_t" + }, + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "text", + "content": " denotes the learning rate and " + }, + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "inline_equation", + "content": "L(\\theta_t)" + }, + { + "bbox": [ + 55, + 609, + 295, + 633 + ], + "type": "text", + "content": " represents the loss function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "content": "To quantify how perturbations evolve during training, we introduce an initial perturbation " + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "inline_equation", + "content": "\\delta \\theta_0" + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "content": " to the model parameters. Parameters after perturbation is " + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "inline_equation", + "content": "\\widetilde{\\theta}_t\\coloneqq \\theta_t + \\delta \\theta_t" + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "content": " is a tiny real-valued scalar hyperparameter. The perturbation at step " + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 634, + 296, + 695 + ], + "type": "text", + "content": " propagates according to the following formula:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 699, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 699, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\widetilde {\\theta} _ {t + 1} = \\widetilde {\\theta} _ {t} - \\eta_ {t} \\nabla L (\\widetilde {\\theta} _ {t}). \\tag {2}", + "image_path": "261f0db9c29e3dbe92406998bd74042ae77f0607d11ccc474ad00cce369e5694.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "Combining Equation 1 and Equation 2, we get the following update for the perturbation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 100, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 100, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 351, + 100, + 553, + 121 + ], + "type": "interline_equation", + "content": "\\delta \\theta_ {t + 1} = \\delta \\theta_ {t} - \\eta_ {t} (\\nabla L (\\widetilde {\\theta_ {t}}) - \\nabla L (\\theta_ {t})). \\tag {3}", + "image_path": "220a047548d6d6aba6ca92b30bd54b2f83264170cc42d25499d470619164927a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "text", + "content": "To further analyze the propagation of this perturbation, we perform a first-order Taylor expansion of the loss function " + }, + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "inline_equation", + "content": "L(\\widetilde{\\theta_t})" + }, + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 313, + 125, + 555, + 162 + ], + "type": "text", + "content": " and calculate the gradient:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 177, + 553, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 177, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 330, + 177, + 553, + 192 + ], + "type": "interline_equation", + "content": "\\nabla L \\left(\\widetilde {\\theta} _ {t}\\right) = \\nabla L \\left(\\theta_ {t}\\right) + H \\left[ L \\left(\\theta_ {t}\\right) \\right] \\delta \\theta_ {t} + o \\left(\\| \\delta \\theta_ {t} \\| ^ {2}\\right), \\tag {4}", + "image_path": "3bd320759cede4f3cc5c4cb1e5d9cfd29b95f82379707cb08b9739a7c8e65cf3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "inline_equation", + "content": "H[L(\\theta_t)]" + }, + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "text", + "content": " is the Hessian matrix, and " + }, + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "inline_equation", + "content": "o(\\| \\delta \\theta_t\\|^2)" + }, + { + "bbox": [ + 313, + 201, + 553, + 250 + ], + "type": "text", + "content": " represents higher-order terms. Ignoring these higher-order terms, and substituting Equation 4 into Equation 3, we approximate the evolution of the perturbation as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 370, + 257, + 553, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 257, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 370, + 257, + 553, + 271 + ], + "type": "interline_equation", + "content": "\\delta \\theta_ {t + 1} = \\left(I - \\eta_ {t} H \\left[ L \\left(\\theta_ {t}\\right) \\right]\\right) \\delta \\theta_ {t}. \\tag {5}", + "image_path": "133f9f079470493aba72633f7c93cc0936b3a0ec1d40ed5fd7aa2171551b34ac.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "content": "By recursively applying this relationship, we obtain the propagation of the perturbation over time:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 314, + 553, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 314, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 314, + 553, + 342 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\delta \\theta_ {t} = \\left(I - \\eta_ {t - 1} H \\left[ L \\left(\\theta_ {t - 1}\\right) \\right]\\right) \\left(I - \\eta_ {t - 2} H \\left[ L \\left(\\theta_ {t - 2}\\right) \\right]\\right) \\tag {6} \\\\ \\cdot \\cdot \\cdot (I - \\eta_ {0} H [ L (\\theta_ {0}) ]) \\delta \\theta_ {0}. \\\\ \\end{array}", + "image_path": "053d2251d676282d55109b5115d47aca5a3f4c93bf33fb9cdc70343c86f8034c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 352, + 504, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 352, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 326, + 352, + 504, + 364 + ], + "type": "text", + "content": "The definition of the Lyapunov exponent is:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 378, + 373, + 553, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 373, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 378, + 373, + 553, + 399 + ], + "type": "interline_equation", + "content": "\\mathrm {L E} = \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\ln \\left(\\frac {\\| \\delta \\theta_ {t} \\|}{\\| \\delta \\theta_ {0} \\|}\\right) \\tag {7}", + "image_path": "66566a802b3b227d79ee907ed86ebb4566b615bd748914601c2f7f93dc07d330.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 407, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 443 + ], + "type": "text", + "content": "By substituting Equation 6 into the definition of LE, we can derive the connection between the model's learning rate " + }, + { + "bbox": [ + 313, + 407, + 553, + 443 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 407, + 553, + 443 + ], + "type": "text", + "content": " and LE:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 449, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 449, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 348, + 449, + 553, + 487 + ], + "type": "interline_equation", + "content": "\\mathrm {L E} \\geq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {\\substack {i = 0\\\\t - 1}} ^ {t - 1} \\ln (1 - \\eta_ {i} \\| H [ L (\\theta_ {i}) ] \\|) \\tag{8}", + "image_path": "7bb7a1ec3f04cb2d927a958a7a67e13321f78cd8ab10dbd5bb15aeaee2538ba8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 350, + 484, + 518, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 484, + 518, + 515 + ], + "spans": [ + { + "bbox": [ + 350, + 484, + 518, + 515 + ], + "type": "interline_equation", + "content": "\\mathrm {L E} \\leq \\lim _ {t \\rightarrow \\infty} \\frac {1}{t} \\sum_ {i = 0} ^ {t - 1} \\ln \\left(\\| I - \\eta_ {i} H [ L (\\theta_ {i}) ] \\|\\right),", + "image_path": "895ee6dab89a5575b955457498a633a6af41f4ce3b3ad8e845da1de51e738846.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "content": "where the LE is determined by the learning rate " + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "inline_equation", + "content": "\\eta_{i}" + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "text", + "content": " and the Hessian matrix " + }, + { + "bbox": [ + 313, + 520, + 553, + 544 + ], + "type": "inline_equation", + "content": "H[L(\\theta_i)]" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 552, + 455, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 455, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 455, + 565 + ], + "type": "text", + "content": "3.3. LE-Guided Optimization" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "A positive LE indicates the exponential growth of perturbations, signifying instability, while a negative LE suggests convergence and stability. However, extreme stability or instability can lead to either overfitting or convergence difficulties. Our approach aims to guide model parameters toward an LE value close to zero but slightly negative, positioning the system near the edge of chaos. This critical state balances stability with adaptability, enhancing the model's generalization capacity. To achieve this, we dynamically adjust the learning rate to amplify perturbations, thereby steering the LE toward closer to the edge of chaos. Specifically, the innovation of the proposed LEAwareSGD lies in" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "554" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "its learning rate adjustment based on LE dynamics, allowing the model to adapt effectively across training phases. We modulate the learning rate according to changes in the LE, " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 133, + 295, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 133, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 80, + 133, + 295, + 148 + ], + "type": "interline_equation", + "content": "\\eta_ {t + 1} = \\eta_ {t} \\cdot \\exp (- \\beta \\cdot \\Delta \\mathrm {L E} _ {t}) \\quad \\text {i f} \\quad \\Delta \\mathrm {L E} _ {t} > 0 \\tag {9}", + "image_path": "7815a0e9572a2006bad6fb6117a8d8bce0d25b220c58ea56ba477d9a9eb85cf2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "text", + "content": " is a weighting parameter that controls the sensitivity of the learning rate adjustment. When " + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{LE}_t > 0" + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "text", + "content": ", indicating that the model is approaching regions near the edge of chaos where generalizable features are more likely to emerge, we reduce the learning rate to allow for deeper exploration of these regions. This approach ensures that the model explores parameter regions with strong generalization potential while avoiding over-stabilization that could hinder the learning process. When " + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{LE}_t \\leq 0" + }, + { + "bbox": [ + 55, + 157, + 296, + 288 + ], + "type": "text", + "content": ", we leave the learning rate unchanged, maintaining the current exploration rate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 297, + 277, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 277, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 277, + 310 + ], + "type": "text", + "content": "3.4. LE-Based Adversarial Data Augmentation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "content": "To further enhance the model's robustness to domain shifts, we integrate LEAwareSGD with adversarial data augmentation. Adversarial data augmentation generates samples that simulate domain shifts, exposing the model to a broader range of data variations. Given a model " + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "content": ", we define adversarial samples as transformations of the original data. Let " + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\tau(x; \\omega)" + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "content": " denote a set of semantic transformations parameterized by " + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 315, + 296, + 422 + ], + "type": "text", + "content": ". The joint optimization objective is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 433, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 433, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 59, + 433, + 294, + 482 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\theta} \\max _ {\\omega} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D} _ {S}} \\left[ \\ell (\\theta ; \\tau (x; \\omega), y) - \\lambda d _ {\\theta} (\\tau (x; \\omega), x) \\right] \\\\ + \\frac {\\gamma}{2} \\cdot \\| \\theta \\| _ {2} ^ {2}, \\tag {10} \\\\ \\end{array}", + "image_path": "46dc6d4c93efc70d16b86cd0bc1aefb1510a95414a578a88923c58bb434ae912.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\ell (\\theta ;\\cdot)" + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": " is the prediction loss, " + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": " balances the adversarial loss and feature consistency, and " + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "inline_equation", + "content": "d_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": " represents the feature distance between the original and transformed samples. Here, the regularization term " + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\frac{\\gamma}{2}\\cdot \\| \\theta \\| _2^2" + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": " denotes a weight decay, ensuring that the Hessian matrix remains approximately positive definite. This regularization encourages LE to be negative, facilitating the stability of model training. Note that this adversarial training procedure is guided by LE-based optimization, which encourages the model to generalize to unseen domains by learning robust features invariant to domain-specific perturbations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 170, + 636 + ], + "type": "text", + "content": "3.5. Training Procedure" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "The training procedure involves iterative steps alternating between adversarial example generation and parameter updates using LEAwareSGD. In each epoch, adversarial examples are first generated to simulate unseen domain shifts, fostering the model to learn invariant features. Subsequently, LEAwareSGD adjusts the model's parameters with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 157 + ], + "type": "text", + "content": "a LE-guided learning rate to encourage the parameters' updating towards the edge of chaos. This LE-guided adaptation encourages the exploration of parameter regions conducive to generalization while controlling sensitivity to perturbations. The complete algorithm is shown in Algorithm 1. We also provide a detailed theoretical analysis of the proposed LEAwareSGD in Suppl. Sec-A.1." + } + ] + } + ], + "index": 9 + }, + { + "type": "code", + "bbox": [ + 319, + 182, + 553, + 375 + ], + "blocks": [ + { + "bbox": [ + 315, + 167, + 430, + 178 + ], + "lines": [ + { + "bbox": [ + 315, + 167, + 430, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 430, + 178 + ], + "type": "text", + "content": "Algorithm 1 LEAwareSGD" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "lines": [ + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": "1: Input: Initial parameters " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\eta_0" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": ", weighting parameter " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": ", number of iterations " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n2: Output: Optimized parameters " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\theta^*" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "t = 0, 1, 2, \\dots, N - 1" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " do \n4: Compute gradient " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta} L(\\theta_t)" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n5: Update parameters: " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\theta_{t+1} = \\theta_t - \\eta_t \\nabla_{\\theta} L(\\theta_t)" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n6: Compute updated model perturbation using Eq. 6 \n7: Calculate LE using Eq. 7 \n8: Update " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{LE}_t = \\mathrm{LE}_t - \\mathrm{LE}_{t-1}" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n9: if " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\Delta \\mathrm{LE}_t > 0" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " then \n10: " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\eta_{t+1} = \\eta_t \\cdot \\exp(-\\beta \\cdot \\Delta \\mathrm{LE}_t)" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n11: else \n12: " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\eta_{t+1} = \\eta_t" + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "text", + "content": " \n13: end if \n14: end for \n15: Return: " + }, + { + "bbox": [ + 319, + 182, + 553, + 375 + ], + "type": "inline_equation", + "content": "\\theta^*" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 389, + 390, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 389, + 390, + 403 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 390, + 403 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 409, + 437, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 437, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 437, + 422 + ], + "type": "text", + "content": "4.1. Datasets and Settings" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 629 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three widely-used domain generalization benchmarks: PACS [15], OfficeHome [31], and DomainNet [24], arranging them in increasing levels of difficulty to comprehensively test the model's robustness across varying domain shifts. PACS consists of four domains, namely, Photo (P), Art Painting (A), Cartoon (C), and Sketch (S), comprising 9,991 images from 7 categories. OfficeHome includes four domain, namely, Art (A), Clipart (C), Product (P), and Real-World (R), with 15,500 images across 65 classes. Compared to PACS, OfficeHome introduces greater diversity in both object categories and domain styles. DomainNet is the most challenging, containing six domains, namely, Real (R), Infograph (I), Clipart (C), Painting (P), Quickdraw (Q), and Sketch (S), with 596,010 images from 345 classes. The substantial domain shifts and the high number of classes in this dataset present considerable challenges for domain generalization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Implementation details. In all experiments, we employed ResNet-18 [11] as the backbone model and resized all input images to " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": ". The network was trained with LEAwareSGD, setting the momentum to 0.9 and " + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": " to 1e-1. To ensure a fair comparison, we excluded the test-time learning stage of PSDG[41] in all of our experiments. For PACS and OfficeHome datasets, we follow the standard" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "555" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "protocol of using one domain as the source domain and the remaining three domains as the target domains [47, 48]. For the PACS dataset, we adopted a batch size of 16, " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " of 5e-4, and an initial learning rate of 5e-4, training for 50 epochs. For the OfficeHome dataset, we used a batch size of 32, set " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " to 1e-4, and initialized the learning rate at 1e-4, training for 50 epochs. For experiments conducted on the DomainNet dataset, we follow [37, 48] and use Real as the source domain and the remaining five domains as the target domains. The batch size was increased to 128, " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " was set to 1e-5, and the initial learning rate was set to 1e-3, with a training duration of 200 epochs. For hyperparameter tuning, we fix all training settings (e.g., random seed, learning rate) and tune only " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": ", with the best setting shown in Fig.2. All experiments were performed on an NVIDIA A100 GPU, and each experiment was repeated 10 times with different random seeds. The results are reported as the average accuracy alongside the standard deviation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 243, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 243, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 243, + 313 + ], + "type": "text", + "content": "4.2. Comparison with SDG Approaches" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": "Results on PACS. This adjustment aligns the evaluation setup with our method, which operates solely during the training phase. The evaluation results on the PACS dataset, presented in Table 1, indicate that our method achieves the highest average accuracy of " + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "inline_equation", + "content": "69.46\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": ", outperforming all compared methods. Specifically, our model achieves the highest performance gains in both the A and P domains, with classification accuracies of " + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "inline_equation", + "content": "79.17\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "inline_equation", + "content": "65.05\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": ", respectively. Our approach outperforms the second-best method, PSDG " + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "inline_equation", + "content": "(67.14\\%)" + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": ", with an improvement of " + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "inline_equation", + "content": "2.32\\%" + }, + { + "bbox": [ + 55, + 319, + 295, + 499 + ], + "type": "text", + "content": " in average accuracy. This result underscores the effectiveness of our LE-guided adversarial data augmentation strategy, which enhances model generalization by facilitating the exploration of a broader parameter space and reducing overfitting to domain-specific features." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 512, + 294, + 679 + ], + "blocks": [ + { + "bbox": [ + 57, + 512, + 294, + 679 + ], + "lines": [ + { + "bbox": [ + 57, + 512, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 294, + 679 + ], + "type": "table", + "html": "
MethodsACPSAvg.
ERM70.4973.5645.9241.2157.80
MixUp [39]67.6067.1351.5947.0458.34
RSC [13]73.4075.9056.2041.6061.78
ADA [33]71.5676.8452.3643.6661.11
ME-ADA [47]71.5276.8346.2246.3260.22
SAGM [35]65.9470.1042.1453.6557.96
RandConv [38]73.6874.8855.4246.7762.69
MetaCNN [34]76.5074.5056.7055.7065.80
L2D [36]77.0076.5553.2254.4465.30
PDEN [16]76.4373.8758.5253.9265.69
PSDG [41]77.8875.2455.4959.9667.14
AdvST [48]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
", + "image_path": "71bf92bc005611fb165169dd8eb563c3a6b6287bf1edcd06a12bfe6475c7ef77.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "Results on OfficeHome. Table 2 shows the classification accuracy comparison of our method with existing baselines on the OfficeHome dataset. Our proposed method consistently outperforms other approaches, achieving the highest accuracy across all domains (A, C, P, and R) with an average accuracy of " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "54.38\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ". Notably, our method surpasses the previous best method, AdvST, which obtained an average accuracy of " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "52.60\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ", with an improvement of " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "1.78\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ". This gain highlights the effectiveness of our LE-guided optimization, which facilitates the exploration of a broader parameter space and enhances the model's ability to generalize to unseen domains." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 224, + 553, + 402 + ], + "blocks": [ + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 687, + 295, + 710 + ], + "type": "text", + "content": "Table 1. Classification accuracy (%) comparison on the PACS dataset. The best results are in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 224, + 553, + 402 + ], + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 402 + ], + "type": "table", + "html": "
MethodsACPRAvg.
ERM42.7439.3139.8652.5043.60
MixUp [39]43.6940.6838.5652.1243.76
SagNet [21]49.8443.0141.4255.6147.47
RandConv [38]45.7940.2037.4552.8744.08
Fish [28]45.1039.7436.6752.0943.40
ADA [33]45.0736.0744.4353.4444.75
ME-ADA [47]45.2745.7938.6551.6945.35
GSAM [51]28.9741.4041.7251.5040.90
SAGM [35]47.4243.2741.3654.9746.76
Mixstyle [49]51.1948.7346.8555.8850.66
PCGrad [42]49.9146.6146.1156.3149.74
PSDG [41]42.8046.4044.3154.6847.05
AdvST [48]51.3252.0848.8958.1152.60
Ours52.8955.1851.6958.9554.38±0.30
", + "image_path": "df133d6bd18f9e46f7bd94f7babe5c7363581a3a165c26b5baa29d2b690d0f0a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 410, + 553, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 553, + 431 + ], + "type": "text", + "content": "Table 2. Classification accuracy (%) comparison on the Office-Home dataset. The best results are in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": "Results on DomainNet. Table 3 shows the classification accuracy comparison on the DomainNet dataset. Our proposed method achieves the highest average accuracy of " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "28.15\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ", surpassing all other methods, including the second-best method, AdvST, which reaches an average accuracy of " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "27.22\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ". Notably, our method performs best in the C and S domains, with accuracies of " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "44.67\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "33.74\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ", respectively. In the Q domain, our method achieves an accuracy of " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "6.70\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ", slightly lower than SimDE's " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "6.85\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ". This may be due to the highly abstract nature of the Quickdraw domain, which can benefit from specialized augmentation techniques to better capture its distinct features. Despite this, our method's overall performance advantage demonstrates the effectiveness of the LE-guided adversarial data augmentation approach in handling the diversity and complexity of DomainNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 635, + 541, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 541, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 541, + 649 + ], + "type": "text", + "content": "4.3. Comparison with Optimization Approaches" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 712 + ], + "type": "text", + "content": "To further validate the effectiveness of our proposed LEAwareSGD optimizer, we conducted a comparison with four widely used optimization approaches, including Adam [14], AdamW [20], RMSprop [30], and SGD [26], on the PACS and OfficeHome datasets. As shown in Table 4," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "556" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 304, + 237 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 304, + 237 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 304, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 304, + 237 + ], + "type": "table", + "html": "
MethodsPICSQAvg.
ERM38.0513.3137.8926.263.3623.77
MixUp [39]38.6013.9238.0126.013.7124.05
CutOut [5]38.3013.7038.4026.203.7024.06
AugMix [12]40.8013.9041.7029.806.3026.50
RandAug [3]41.3013.5741.1130.405.3126.34
ADA [33]38.2013.8040.2024.804.3024.26
ME-ADA [47]38.4813.5640.6626.054.4224.63
ACVC [4]41.3012.9042.8030.906.6026.90
PDEN [16]38.4511.2538.9931.715.5825.20
SimDE [37]39.9612.9141.7333.466.8526.98
PSDG [41]41.1013.0041.2931.104.8926.28
AdvST [48]42.4014.9041.7031.006.1027.22
Ours42.7715.9144.6733.746.7028.15 ±0.61
", + "image_path": "f1aca5b1e029d4df1c443b5561535bc3a7b89e6a18c96785870b2aeae8d30a4b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": "our proposed optimizer achieves the highest average accuracy of " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "69.46\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ", surpassing all other optimizers on the PACS dataset. In contrast, widely used optimizers such as Adam, AdamW, and RMSprop perform significantly worse, with average accuracies of " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "66.43\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "66.83\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "62.34\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ", respectively. This disparity likely stems from the reliance of these adaptive optimizers on variable learning rates, which may cause overfitting to domain-specific features and reduce their generalization ability in domain generalization tasks. On the other hand, both our LEAwareSGD and standard SGD achieve substantially higher accuracies. While SGD achieves an average accuracy of " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "67.06\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ", our LEAwareSGD optimizer further improves this to " + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "inline_equation", + "content": "69.46\\%" + }, + { + "bbox": [ + 54, + 274, + 298, + 525 + ], + "type": "text", + "content": ". This enhancement underscores the advantage of integrating LE guidance into the optimization process, which enables the model to better capture generalizable features and avoid overfitting to domain-specific characteristics. We present the experimental results on the OfficeHome dataset in Suppl. Sec-C.2, while analyses of optimizers' performance under different learning rates are provided in Suppl. Sec-C.5." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 535, + 303, + 614 + ], + "blocks": [ + { + "bbox": [ + 55, + 245, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 266 + ], + "type": "text", + "content": "Table 3. Classification accuracy (%) comparison on the Domain-Net dataset. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 535, + 303, + 614 + ], + "lines": [ + { + "bbox": [ + 56, + 535, + 303, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 303, + 614 + ], + "type": "table", + "html": "
MethodsACPSAvg.
Adam [14]76.5271.1564.0653.9866.43
AdamW [20]76.6871.9362.0356.6866.83
RMSprop [30]71.6271.0859.0947.5562.34
SGD [26]76.6574.9262.4754.1867.06
Ours79.1777.1665.0557.7869.46±0.33
", + "image_path": "c38222a6bfaaa5cdb2d0f5efece7db33d1f0b3e8a6826fd211c6dc602ac1e6a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 280, + 673 + ], + "type": "text", + "content": "4.4. Extension to Multi-Source DG Approaches" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "To further validate the scalability of our solution, we perform experiments under the leave-one-out setting of general domain generalization (DG) with ResNet-18 as the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "backbone, where three domains are combined for training and the remaining domain is reserved for evaluation. Table 5 presents the results on the PACS dataset, showing that our method delivers at least comparable performance, highlighting its versatility. This improvement underscores the effectiveness of our LE-guided optimization, which enables a broader exploration of the parameter space and enhances the model's generalization to unseen domains." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 176, + 543, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 622, + 295, + 644 + ], + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 644 + ], + "type": "text", + "content": "Table 4. Comparison results " + }, + { + "bbox": [ + 55, + 622, + 295, + 644 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 622, + 295, + 644 + ], + "type": "text", + "content": " using different optimizers on the PACS dataset. The best results are in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 176, + 543, + 277 + ], + "lines": [ + { + "bbox": [ + 326, + 176, + 543, + 277 + ], + "spans": [ + { + "bbox": [ + 326, + 176, + 543, + 277 + ], + "type": "table", + "html": "
MethodsACPSAvg.
RSC [13]82.8179.7493.9580.8584.34
ADA [33]82.8178.3395.6375.2983.02
ME-ADA [47]77.8878.5895.3378.0782.47
Mixstyle [49]83.1179.4396.3172.9582.95
L2D [36]81.4479.5695.5180.5884.27
AdvST [48]78.6972.9791.6280.9181.05
Ours81.3578.8895.4582.3984.52
", + "image_path": "6369fa0bc0b33d0152e80b668a2422ce479479b8099194f773abeb32ee92fbd6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 330 + ], + "type": "text", + "content": "Table 5. Classification accuracy (\\%) comparison for the leave-one-domain-out using the ResNet-18 backbone on the PACS dataset. The column name denotes the test domain, with the other three used for training. The best results are in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 337, + 430, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 430, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 430, + 350 + ], + "type": "text", + "content": "4.5. Additional Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": "Performance comparison using different backbone networks. To further validate the universality of our solution, we conduct experiments with different ResNet backbones, including ResNet-34, ResNet-50, ResNet-101, and ResNet-152, on the PACS and OfficeHome datasets. As shown in Table 6 on the PACS dataset, our approach reaches a performance of " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "73.68\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " on ResNet-34, " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "71.37\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " on ResNet-50, " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "74.32\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " on ResNet-101, and " + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "inline_equation", + "content": "75.34\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 510 + ], + "type": "text", + "content": " on ResNet-152. These results demonstrate the consistent effectiveness of our method across various backbone architectures, highlighting its superior accuracy, versatility, and robustness. The experimental results for the OfficeHome dataset are presented in Suppl. Sec-C.6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Performance comparison with limited training data. We compare the performance of our proposed method with AdvST on the PACS and OfficeHome datasets under varying source domain data ratios (10%, 20%, and 50%). As shown in Tables 7 on the PACS dataset, our method consistently outperforms AdvST across all data ratios, demonstrating robust generalization even with limited source domain data for training. At the 10% data ratio, our method achieves an average accuracy of 58.73%, showing a notable improvement of 9.47% over AdvST's 49.26%. With the 20% data ratio, our approach reaches an average accuracy of 61.78%, surpassing AdvST by 8.23%. At the 50% data ratio, the performance gap narrows slightly, with our method achieving an average accuracy of 66.44%, which is 6.26% higher than the 60.18% achieved by AdvST. The experimental results on the OfficeHome dataset are provided in Suppl. Sec-C.7. These findings suggest that our" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "557" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 226 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 226 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 226 + ], + "type": "table", + "html": "
BackboneMethodsACPSAvg.
ResNet34PSDG79.4680.7957.4358.2268.98
AdvST79.1676.6765.2860.5570.42
Ours81.8981.8565.6565.3173.68
ResNet50PSDG79.8579.7258.0865.6870.83
AdvST79.6575.0562.6255.7568.27
Ours81.3278.7065.0860.3971.37
ResNet101PSDG79.8479.9759.1770.7972.44
AdvST80.9375.3566.2958.7070.32
Ours82.9580.7968.5764.9574.32
ResNet152PSDG82.9281.5462.3871.2574.52
AdvST80.7679.0265.9061.9971.92
Ours83.4482.1967.6368.1175.34
", + "image_path": "3b548e14611decb8ebdd47cbb7aefe0c9cf5c653a08496a3c525c2228aa6c54f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 327 + ], + "type": "text", + "content": "method effectively reduces overfitting to specific domains with limited training data, particularly in low-data settings, by encouraging broader exploration of the parameter space to capture generalizable features." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 62, + 337, + 290, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 234, + 296, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 257 + ], + "type": "text", + "content": "Table 6. Comparison results (\\%) using different ResNet series backbones on the PACS dataset. The best results are in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 337, + 290, + 428 + ], + "lines": [ + { + "bbox": [ + 62, + 337, + 290, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 290, + 428 + ], + "type": "table", + "html": "
RatioMethodsACPSAvg.ΔAvg.
10%AdvST58.7456.9440.6240.7349.26
Ours65.4364.4349.4055.6658.73+9.47
20%AdvST61.4864.3143.4244.9853.55
Ours68.8168.7853.2856.2761.78+8.23
50%AdvST70.2772.7449.5248.2060.18
Ours75.8873.5561.6054.7266.44+6.26
", + "image_path": "3c6235090ebf25f7bc5dc346c50ff5929408d27f5d47dd926171477e7899bf41.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 462, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 653 + ], + "type": "text", + "content": "Computational complexity analysis. To evaluate the computational efficiency of our method, we compared its training time cost with three adversarial data augmentation methods on the PACS and OfficeHome datasets. As shown in Table 8, our approach requires an average training time of 1.99 hours on PACS, slightly higher than AdvST (1.90 hours) but comparable to ADA (2.13 hours) and ME-ADA (2.10 hours). This marginal increase is due to the Lyapunov-guided optimization, which enhances model robustness and generalization. On OfficeHome, our method completes training in 1.19 hours, surpassing AdvST (0.75 hours) but remaining faster than ADA (1.95 hours) and ME-ADA (1.90 hours). While our method introduces a slight time increase compared to AdvST, it achieves significant performance improvements, demonstrating a favorable balance between training time and model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Generalizability analysis with LEAwareSGD using different adversarial data augmentation approaches. To assess the impact of LEAwareSGD on the generalization performance of adversarial data augmentation methods, we integrated it with ADA, ME-ADA, and AdvST" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 561, + 202 + ], + "blocks": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 458 + ], + "type": "text", + "content": "Table 7. Performance comparison (%) using different data ratios on PACS dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 561, + 202 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 561, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 561, + 202 + ], + "type": "table", + "html": "
DatasetMethodsACPSAvg.
PACSADA1.802.101.403.202.13
ME-ADA1.802.101.303.202.10
AdvST1.581.871.402.761.90
Ours1.622.021.352.981.99
OfficeHomeMethodsACPRAvg.
ADA1.202.102.202.301.95
ME-ADA1.201.902.202.301.90
AdvST0.750.740.750.760.75
Ours1.101.171.231.271.19
", + "image_path": "4a71e1631f6c48141eac37ac0c2eea8dc9eef8ae32c84c570046592de8a66b65.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 211, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 314, + 211, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 314, + 211, + 553, + 222 + ], + "type": "text", + "content": "Table 8. Training time (hours) on PACS and OfficeHome datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 232, + 422, + 308 + ], + "blocks": [ + { + "bbox": [ + 316, + 232, + 422, + 308 + ], + "lines": [ + { + "bbox": [ + 316, + 232, + 422, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 422, + 308 + ], + "type": "image", + "image_path": "b8d74aed33b4c6a386436373fb02e3ab27dcfd35852fddcb1aaad349992cf67f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 313, + 432, + 332 + ], + "lines": [ + { + "bbox": [ + 315, + 313, + 432, + 332 + ], + "spans": [ + { + "bbox": [ + 315, + 313, + 432, + 332 + ], + "type": "text", + "content": "(a) Parameter " + }, + { + "bbox": [ + 315, + 313, + 432, + 332 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 315, + 313, + 432, + 332 + ], + "type": "text", + "content": " on the PACS dataset" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 440, + 232, + 544, + 308 + ], + "blocks": [ + { + "bbox": [ + 440, + 232, + 544, + 308 + ], + "lines": [ + { + "bbox": [ + 440, + 232, + 544, + 308 + ], + "spans": [ + { + "bbox": [ + 440, + 232, + 544, + 308 + ], + "type": "image", + "image_path": "3303dbc38c79211304f06470f21e1fea538b26a250935f19a24aa0e3984c1814.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 313, + 554, + 332 + ], + "lines": [ + { + "bbox": [ + 437, + 313, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 437, + 313, + 554, + 332 + ], + "type": "text", + "content": "(b) Parameter " + }, + { + "bbox": [ + 437, + 313, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 437, + 313, + 554, + 332 + ], + "type": "text", + "content": " on the OfficeHome dataset" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 340, + 422, + 416 + ], + "blocks": [ + { + "bbox": [ + 317, + 340, + 422, + 416 + ], + "lines": [ + { + "bbox": [ + 317, + 340, + 422, + 416 + ], + "spans": [ + { + "bbox": [ + 317, + 340, + 422, + 416 + ], + "type": "image", + "image_path": "41b79aa285f528be448f92d0f9763bdbfd64e02a211d52b42704b7ef383a66f8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 422, + 431, + 431 + ], + "lines": [ + { + "bbox": [ + 315, + 422, + 431, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 431, + 431 + ], + "type": "text", + "content": "(c) Parameter " + }, + { + "bbox": [ + 315, + 422, + 431, + 431 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 315, + 422, + 431, + 431 + ], + "type": "text", + "content": " on the PACS dataset" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 451, + 503, + 462 + ], + "lines": [ + { + "bbox": [ + 365, + 451, + 503, + 462 + ], + "spans": [ + { + "bbox": [ + 365, + 451, + 503, + 462 + ], + "type": "text", + "content": "Figure 2. Parameter sensitivity study." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 440, + 340, + 544, + 416 + ], + "blocks": [ + { + "bbox": [ + 440, + 340, + 544, + 416 + ], + "lines": [ + { + "bbox": [ + 440, + 340, + 544, + 416 + ], + "spans": [ + { + "bbox": [ + 440, + 340, + 544, + 416 + ], + "type": "image", + "image_path": "660838a09ed295292ab4cb24909166b78664bc36db40e39913bdad9aee2cdc2c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 422, + 553, + 440 + ], + "lines": [ + { + "bbox": [ + 437, + 422, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 437, + 422, + 553, + 440 + ], + "type": "text", + "content": "(d) Parameter " + }, + { + "bbox": [ + 437, + 422, + 553, + 440 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 437, + 422, + 553, + 440 + ], + "type": "text", + "content": " on the OfficeHome dataset" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "on the PACS and OfficeHome datasets. Table 9 shows consistent improvements across all methods, underscoring LEAwareSGD's effectiveness in enhancing model generalization. On the PACS dataset, integrating LEAwareSGD with AdvST led to a " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.40\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " increase in average accuracy, reaching " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "69.46\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". This improvement highlights LEAwareSGD's ability to facilitate more effective parameter space exploration, enhancing AdvST's generalization capacity. For ADA and ME-ADA, LEAwareSGD achieved gains of " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "0.52\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.30\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ", respectively, with ME-ADA reaching an average accuracy of " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "62.52\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". On the OfficeHome dataset, LEAwareSGD also improved the performance of all methods. ME-ADA saw the highest gain, with a " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "7.00\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " increase over ME-ADA alone, achieving an average accuracy of " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "52.35\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". ADA with LEAwareSGD increased by " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "5.15\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ", reaching an average accuracy of " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "49.90\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". For AdvST, which already demonstrated strong performance, LEAwareSGD provided an additional " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.78\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " gain, bringing the average accuracy to " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "54.38\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". These findings demonstrate that LEAwareSGD enhances the generalization" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "558" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "capacity of adversarial data augmentation techniques by enabling broader exploration of the parameter space. This exploration can facilitate learning more generalizable features across unseen domains, underscoring the effectiveness of LEAwareSGD as a robust optimization framework for SDG." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 144, + 299, + 332 + ], + "blocks": [ + { + "bbox": [ + 56, + 144, + 299, + 332 + ], + "lines": [ + { + "bbox": [ + 56, + 144, + 299, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 299, + 332 + ], + "type": "table", + "html": "
DatasetMethodsACPSAvg.ΔAvg.
PACSADA73.7471.6540.7160.3961.62
\\( \\mathrm{ADA}^{\\dagger} \\)71.8269.1551.9455.6662.14+0.52
ME-ADA71.5276.8346.2246.3260.22
\\( \\mathrm{ME-ADA}^{\\dagger} \\)72.8869.1647.9860.0562.52+2.30
AdvST76.6574.9262.4754.1867.06
\\( \\mathrm{AdvST}^{\\dagger} \\)79.1777.1665.0557.7869.46+2.40
Office HomeMethodsACPRAvg.ΔAvg.
ADA45.0736.0744.4353.4444.75
\\( \\mathrm{ADA}^{\\dagger} \\)50.6948.5246.1154.2849.90+5.15
ME-ADA45.2745.7938.6551.6945.35
\\( \\mathrm{ME-ADA}^{\\dagger} \\)53.6150.0949.1756.5152.35+7.00
AdvST51.3252.0848.8958.1152.60
\\( \\mathrm{AdvST}^{\\dagger} \\)52.8955.1851.6958.9554.38+1.78
", + "image_path": "fd9cee1f44e11bf1315147bd782d7ca5fd22a1d89de89cc519c31d4a2ccdbd4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 69, + 334, + 278, + 345 + ], + "lines": [ + { + "bbox": [ + 69, + 334, + 278, + 345 + ], + "spans": [ + { + "bbox": [ + 69, + 334, + 278, + 345 + ], + "type": "text", + "content": "Note: " + }, + { + "bbox": [ + 69, + 334, + 278, + 345 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 334, + 278, + 345 + ], + "type": "text", + "content": " indicates methods integrated with LEAwareSGD." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 354, + 296, + 386 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 386 + ], + "type": "text", + "content": "Table 9. Performance improvement (%) of adversarial data augmentation approaches with LEAwareSGD on the PACS and OfficeHome datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": "Parameter sensitivity analysis. To assess the impact of hyperparameters on the performance of LEAwareSGD, we conducted a sensitivity analysis on the parameters beta " + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": " over the PACS and OfficeHome datasets, as shown in Figure 2. Figures 2a and 2b show that increasing " + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": " from 1e-4 to 1.0 resulted in fluctuating accuracy, with peak performance observed at " + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\beta = 1\\mathrm{e} - 3" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": " on PACS and " + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "inline_equation", + "content": "\\beta = 1\\mathrm{e} - 2" + }, + { + "bbox": [ + 55, + 401, + 296, + 508 + ], + "type": "text", + "content": " on OfficeHome. This result suggests an optimal range between 1e-3 and 1e-2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": "As illustrated in Figure 2c and 2d, for the parameter " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " in Equation 10, which controls the strength of regularization, the best accuracy on the PACS dataset was achieved at " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\gamma = 5\\mathrm{e - }4" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": ". Increasing " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " beyond 5e-4 caused slight declines in accuracy, indicating that higher regularization may overly restrict the model. On the OfficeHome dataset, " + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\gamma = 1\\mathrm{e - }5" + }, + { + "bbox": [ + 55, + 510, + 296, + 605 + ], + "type": "text", + "content": " yielded the best accuracy across most domains, while higher values generally led to reduced performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Comparison of LE results of different methods on the PACS dataset. To evaluate our proposed method's stability and generalization performance, we compare the LE values obtained from different adversarial data augmentation methods on the PACS dataset. Figure 3 shows that we track LE dynamics across four domains: Art Painting, Cartoon, Photo, and Sketch. Our method consistently yields LE values closer to zero, indicating that the model operates near the edge of chaos, where an optimal balance between sta" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 68, + 422, + 146 + ], + "blocks": [ + { + "bbox": [ + 317, + 68, + 422, + 146 + ], + "lines": [ + { + "bbox": [ + 317, + 68, + 422, + 146 + ], + "spans": [ + { + "bbox": [ + 317, + 68, + 422, + 146 + ], + "type": "image", + "image_path": "eb46c638218fd55a7d3c6cc88aa0e4f9bb91b369ea61eebcfd392bbe7c00a66c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 148, + 430, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 148, + 430, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 148, + 430, + 159 + ], + "type": "text", + "content": "(a) Trained on Art Painting domain" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 68, + 544, + 146 + ], + "blocks": [ + { + "bbox": [ + 440, + 68, + 544, + 146 + ], + "lines": [ + { + "bbox": [ + 440, + 68, + 544, + 146 + ], + "spans": [ + { + "bbox": [ + 440, + 68, + 544, + 146 + ], + "type": "image", + "image_path": "621e693643a1dc9051e61af3c73388c3f51970f17193486b94c78cb3d4f3c14a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 148, + 546, + 158 + ], + "lines": [ + { + "bbox": [ + 445, + 148, + 546, + 158 + ], + "spans": [ + { + "bbox": [ + 445, + 148, + 546, + 158 + ], + "type": "text", + "content": "(b) Trained on Cartoon domain" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 167, + 421, + 246 + ], + "blocks": [ + { + "bbox": [ + 317, + 167, + 421, + 246 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 421, + 246 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 421, + 246 + ], + "type": "image", + "image_path": "301cb89cdfdd85dbe72653782f3193ddc47584d291cfae76a287af06fc982da0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 248, + 420, + 257 + ], + "lines": [ + { + "bbox": [ + 327, + 248, + 420, + 257 + ], + "spans": [ + { + "bbox": [ + 327, + 248, + 420, + 257 + ], + "type": "text", + "content": "(c) Trained on Photo domain" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 437, + 167, + 542, + 246 + ], + "blocks": [ + { + "bbox": [ + 437, + 167, + 542, + 246 + ], + "lines": [ + { + "bbox": [ + 437, + 167, + 542, + 246 + ], + "spans": [ + { + "bbox": [ + 437, + 167, + 542, + 246 + ], + "type": "image", + "image_path": "ed2c0b4c9754e345ea61f72cffd82292bd53c39af3e5cfd723cd8d4cab684d11.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 248, + 542, + 257 + ], + "lines": [ + { + "bbox": [ + 445, + 248, + 542, + 257 + ], + "spans": [ + { + "bbox": [ + 445, + 248, + 542, + 257 + ], + "type": "text", + "content": "(d) Trained on Sketch domain" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 266, + 553, + 288 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 288 + ], + "type": "text", + "content": "Figure 3. LE dynamics over training epochs across four domains (Art Painting, Cartoon, Photo, Sketch) on the PACS dataset." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 454, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 454, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 454, + 312 + ], + "type": "text", + "content": "bility and adaptability is achieved." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "content": "In the Art Painting domain (Figure 3a), our method maintains LE values closer to zero throughout training compared to ADA, ME-ADA, and AdvST, suggesting a better balance between stability and adaptability, leading to improved generalization. In contrast, ADA exhibits the lowest LE values and large fluctuations, indicating instability and convergence to local optima. Similar results can be derived from the three remaining domains (Figure 3b-3d), where our method consistently stabilizes near the edge of chaos, supporting better generalization. Comparison methods, by contrast, show greater fluctuations or lower LE values, indicating tendencies towards overfitting and limited adaptability. These results reveal that our method consistently achieves LE values near zero across all domains and operates effectively near the edge of chaos. Additional detailed experimental analyses are provided in Suppl. Sec-A.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 534, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 712 + ], + "type": "text", + "content": "In this work, we introduce LEAwareSGD, a novel LE-guided optimization integrated with an adversarial data augmentation method to address the challenge of unseen domain shifts in SDG. By leveraging principles from dynamical systems theory, our approach encourages model training near the edge of chaos, enabling the exploration of a broader parameter space and substantially enhancing generalization to new domains. LEAwareSGD dynamically adjusts the learning rate based on LE values, encouraging the model to capture more generalizable features. Experiments on three standard SDG benchmarks demonstrate that LEAwareSGD outperforms state-of-the-art SDG methods, underscoring its effectiveness. A promising future work is to investigate its scalability to larger datasets and more complex domain generalization tasks." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "559" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work is supported in part by the National Natural Science Foundation of China (No. 62276038), and in part by the Chongqing Municipal Education Commission (No. KJZD-M202400603, No. KJQN202500612), and in part by the Project of Key Laboratory of Tourism Multisource Data Perception and Decision, Ministry of Culture and Tourism (No. H2023009)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 239 + ], + "type": "text", + "content": "[1] Kathleen T Alligood, Tim D Sauer, James A Yorke, and David Chillingworth. Chaos: an introduction to dynamical systems. SIAM Review, 40(3):732-732, 1998. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[2] Sheng Cheng, Tejas Gokhale, and Yezhou Yang. Adversarial bayesian augmentation for single-source domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11400-11410, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 341 + ], + "type": "text", + "content": "[3] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 396 + ], + "type": "text", + "content": "[4] Iike Cugu, Massimiliano Mancini, Yanbei Chen, and Zeynep Akata. Attention consistency on visual corruptions for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4165-4174, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 430 + ], + "type": "text", + "content": "[5] Terrance DeVries. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 464 + ], + "type": "text", + "content": "[6] Ling Feng, Lin Zhang, and Choy Heng Lai. Optimal machine intelligence at the edge of chaos. arXiv preprint arXiv:1909.05176, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 508 + ], + "type": "text", + "content": "[7] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. arXiv preprint arXiv:2010.01412, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 555 + ], + "type": "text", + "content": "[8] Alejandro Guarneros-Sandoval, Mariana Ballesteros, Ivan Salgado, Julia Rodríguez-Santillán, and Isaac Chairez. Lyapunov stable learning laws for multilayer recurrent neural networks. Neurocomputing, 491:644-657, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 599 + ], + "type": "text", + "content": "[9] Moritz Hardt, Ben Recht, and Yoram Singer. Train faster, generalize better: Stability of stochastic gradient descent. In International Conference on Machine Learning, pages 1225-1234. PMLR, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[10] Kazuko Hayashi. Chaotic nature of the electroencephalogram during shallow and deep anesthesia: From analysis of the lyapunov exponent. *Neuroscience*, 557:116-123, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 129 + ], + "type": "text", + "content": "[13] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In ECCV, pages 124-140, 2020. 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 132, + 553, + 153 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 153 + ], + "type": "text", + "content": "[14] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 156, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 156, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 553, + 199 + ], + "type": "text", + "content": "[15] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 5542-5550, 2017. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 201, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 266 + ], + "type": "text", + "content": "[16] Lei Li, Ke Gao, Juan Cao, Ziyao Huang, Yepeng Weng, Xiaoyue Mi, Zhengze Yu, Xiaoya Li, and Boyang Xia. Progressive domain expansion network for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 312 + ], + "type": "text", + "content": "[17] Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dynamical systems: An approximation perspective. Journal of the European Mathematical Society, 25(5):1671-1709, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 369 + ], + "type": "text", + "content": "[18] Shutao Li, Weiwei Song, Leyuan Fang, Yushi Chen, Pedram Ghamisi, and Jon Atli Benediktsson. Deep learning for hyperspectral image classification: An overview. IEEE Transactions on Geoscience and Remote Sensing, 57(9):6690-6709, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 413 + ], + "type": "text", + "content": "[19] Guan-Horng Liu and Evangelos A Theodorou. Deep learning theory review: An optimal control and dynamical systems perspective. arXiv preprint arXiv:1908.10920, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 438 + ], + "type": "text", + "content": "[20] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 441, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 495 + ], + "type": "text", + "content": "[21] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8690-8699, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[22] Yurii Nesterov. A method for solving the convex programming problem with convergence rate o (1/k2). In Proceedings of the USSR Academy of Sciences, pages 543-547, 1983. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 544, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 575 + ], + "type": "text", + "content": "[23] Sinno Jialin Pan and Qiang Yang. A survey on transfer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345-1359, 2009. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "text", + "content": "[24] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1406-1415, 2019. 2, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[25] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12556-12565, 2020. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[26] Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400-407, 1951. 2, 5, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "560" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[27] Ling Shao, Fan Zhu, and Xuelong Li. Transfer learning for visual categorization: A survey. IEEE Transactions on Neural Networks and Learning Systems, 26(5):1019-1034, 2014. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[28] Yuge Shi, Jeffrey Seely, Philip HS Torr, N Siddharth, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. arXiv preprint arXiv:2104.09937, 2021.5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "text", + "content": "[29] Yang Tang, Jürgen Kurths, Wei Lin, Edward Ott, and Ljupco Kocarev. Introduction to focus issue: When machine learning meets complex systems: Networks, chaos, and nonlinear dynamics. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30(6), 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[30] Tijmen Tieleman and Geoffrey Hinton. Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn, 17, 2012. 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[31] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "text", + "content": "[32] Ryan Vogt, Maximilian Puelma Touzel, Eli Shlizerman, and Guillaume Lajoie. On lyapunov exponents for rnns: Understanding information propagation using dynamical systems tools. Frontiers in Applied Mathematics and Statistics, 8: 818799, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 442 + ], + "type": "text", + "content": "[33] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, pages 5339-5349, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 510 + ], + "type": "text", + "content": "[34] Chaoqun Wan, Xu Shen, Yonggang Zhang, Zhiheng Yin, Xinmei Tian, Feng Gao, Jianqiang Huang, and Xian-Sheng Hua. Meta convolutional neural networks for single domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4682-4691, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "type": "text", + "content": "[35] Pengfei Wang, Zhaoxiang Zhang, Zhen Lei, and Lei Zhang. Sharpness-aware gradient matching for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3769-3778, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "type": "text", + "content": "[36] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 834-843, 2021. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "text", + "content": "[37] Qinwei Xu, Ruipeng Zhang, Yi-Yan Wu, Ya Zhang, Ning Liu, and Yanfeng Wang. Simde: A simple domain expansion approach for single-source domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4798-4808, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "[38] Zhenlin Xu, Deyi Liu, Junlin Yang, Colin Raffel, and Marc Niethammer. Robust and generalizable visual represent" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "tation learning via random convolutions. arXiv preprint arXiv:2007.13003, 2020.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 95, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 128 + ], + "type": "text", + "content": "[39] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 171 + ], + "type": "text", + "content": "[40] Shuai Yang, Zhen Zhang, and Lichuan Gu. Causality-inspired domain expansion network for single domain generalization. Knowledge-Based Systems, 301:112269, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "text", + "content": "[41] Shuai Yang, Zhen Zhang, and Lichuan Gu. Practical single domain generalization via training-time and test-time learning. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3794-3805, 2024. 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[42] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 338 + ], + "type": "text", + "content": "[43] Ling Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Holger Roth, Andriy Myronenko, Daguang Xu, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation. IEEE Transactions on Medical Imaging, 39(7):2531-2540, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 371 + ], + "type": "text", + "content": "[44] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Edge of chaos as a guiding principle for modern neural network training. arXiv preprint arXiv:2107.09437, 2021. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 415 + ], + "type": "text", + "content": "[45] Lin Zhang, Ling Feng, Kan Chen, and Choy Heng Lai. Asymptotic edge of chaos as guiding principle for neural network training. International Journal of Artificial Intelligence and Robotics Research, 1(01):2350001, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 469 + ], + "type": "text", + "content": "[46] Xingxuan Zhang, Renzhe Xu, Han Yu, Yancheng Dong, Pengfei Tian, and Peng Cui. Flatness-aware minimization for domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5189-5202, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 470, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 470, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 470, + 555, + 525 + ], + "type": "text", + "content": "[47] Long Zhao, Ting Liu, Xi Peng, and Dimitris Metaxas. Maximum-entropy adversarial data augmentation for improved generalization and robustness. Advances in Neural Information Processing Systems, 33:14435–14447, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 525, + 555, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 555, + 569 + ], + "type": "text", + "content": "[48] Guangtao Zheng, Mengdi Huai, and Aidong Zhang. Advst: Revisiting data augmentations for single domain generalization. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 21832-21840, 2024. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 613 + ], + "type": "text", + "content": "[49] Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation. International Journal of Computer Vision, 132 (3):822-836, 2024. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 555, + 658 + ], + "type": "text", + "content": "[50] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[51] Juntang Zhuang, Boqing Gong, Liangzhe Yuan, Yin Cui, Hartwig Adam, Nicha Dvornek, Sekhar Tatikonda, James Duncan, and Ting Liu. Surrogate gap minimization improves sharpness-aware training. arXiv preprint arXiv:2203.08065, 2022. 2, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "561" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_content_list.json b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..779eb7025a9a575e7649ab83de34386e57e60c10 --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_content_list.json @@ -0,0 +1,1899 @@ +[ + { + "type": "text", + "text": "Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis", + "text_level": 1, + "bbox": [ + 197, + 130, + 802, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yanzuo Lu $^{1,2}$ , Yuxi Ren $^{2}$ , Xin Xia $^{2}$ , Shanchuan Lin $^{2}$ , Xing Wang $^{2}$ , Xuefeng Xiao $^{2*}$ , Andy J. Ma $^{1,3,4\\dagger}$ , Xiaohua Xie $^{1,3,4,5}$ , Jian-Huang Lai $^{1,3,4,5}$", + "bbox": [ + 199, + 202, + 800, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Sun Yat-Sen University 2ByteDance Seed Vision \n3Guangdong Provincial Key Laboratory of Information Security Technology, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Pazhou Lab (HuangPu), Guangzhou, China", + "bbox": [ + 109, + 255, + 887, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oliveryanzulu@gmail.com, xiaoxuefeng.ailab@bytedance.com, majh8@mail.sysu.edu.cn", + "bbox": [ + 143, + 330, + 844, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distribution Matching Distillation (DMD) is a promising score distillation technique that compresses pre-trained teacher diffusion models into efficient one-step or multi-step student generators. Nevertheless, its reliance on the reverse Kullback-Leibler (KL) divergence minimization potentially induces mode collapse (or mode-seeking) in certain applications. To circumvent this inherent drawback, we propose Adversarial Distribution Matching (ADM), a novel framework that leverages diffusion-based discriminators to align the latent predictions between real and fake score estimators for score distillation in an adversarial manner. In the context of extremely challenging one-step distillation, we further improve the pre-trained generator by adversarial distillation with hybrid discriminators in both latent and pixel spaces. Different from the mean squared error used in DMD2 pre-training, our method incorporates the distributional loss on ODE pairs collected from the teacher model, and thus providing a better initialization for score distillation fine-tuning in the next stage. By combining the adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our proposed method achieves superior one-step performance on SDXL compared to DMD2 while consuming less GPU time. Additional experiments that apply multi-step ADM distillation on SD3-Medium, SD3.5-Large, and CogVideoX set a new benchmark towards efficient image and video synthesis.", + "bbox": [ + 88, + 393, + 485, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent methods for accelerating diffusion models [12, 17, 63, 65] have concentrated on reducing the sampling steps through distillation. The distillation process generally trains", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee0a03c5ae49927ae0ea66fe1ac89fedf63a579444d366e3fc72ec2823526b01.jpg", + "image_caption": [ + "(a) A white polar bear cub wearing a red sitta in a meadow with flowers." + ], + "image_footnote": [], + "bbox": [ + 516, + 361, + 614, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a86eb6eb51f6704986d4549274c849c33d0622bfea3bd6e8472a9badd9d82f97.jpg", + "image_caption": [ + "(b) mage, wizard, female, demoness, forest, D&D, fantasy, incricate, elegant, highly detailed, digital painting, artstation, [...]" + ], + "image_footnote": [], + "bbox": [ + 614, + 361, + 705, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/516db7b264c7fbb3e87edbf3fd8f759cb41461506d5f4f3c353dbfcb0480f04d.jpg", + "image_caption": [ + "(c) beautiful nighttime landscape photography of the Rocky Mountains with a crystal blue lake, milky way galaxy sky, serene, dramatic lightning." + ], + "image_footnote": [], + "bbox": [ + 707, + 361, + 805, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36b49336c818ac960121d7ee4716ae2cefa5de355cffd08e566333b26f667651.jpg", + "image_caption": [ + "(d) A planeplane lacavillo glows over the surface of a sand with swirling mist around it.", + "Figure 1. In these images, some are generated by the baseline SDXL via 50NFE, the others with our DMDX in 1NFE. Can you tell which is the accelerated one? Answers in the footnote $\\ddagger$ ." + ], + "image_footnote": [], + "bbox": [ + 807, + 361, + 903, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a more efficient generator (a.k.a student model) that approximates the output distribution of the pretrained teacher. This field has observed multiple pathways developing in parallel with progressive distillation [23, 58], consistency distillation [18, 34, 35, 43, 55, 66, 69, 70, 92], score distillation [3, 16, 21, 38, 55, 59, 61, 84-87, 93], rectified flow [27, 28, 30, 71, 81] and adversarial distillation [23, 24, 43, 55, 60, 61, 85, 91]. They are currently appearing as distinct yet non-exclusive research directions.", + "bbox": [ + 509, + 587, + 906, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Distribution Matching Distillation (DMD) [85, 86] is a promising score distillation [51] approach, which distills the powerful text-to-image diffusion model SDXL-Base [56] into a one-step generator with great fidelity. While DMD's primary contribution lies in introducing additional regularizers to constrain the distribution matching loss, its predecessor Diff-Instruct [37] pioneered the use of a fake score estimator to approximate the student model's output distribution. This contrasts with the distillation loss in Adversarial Diffusion Distillation (ADD) [61], which directly em", + "bbox": [ + 509, + 723, + 908, + 875 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Project Lead. † Corresponding Author.", + "bbox": [ + 114, + 886, + 336, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{\\dagger}$ Our DMDX (left to right): bottom, top, bottom, top.", + "bbox": [ + 537, + 886, + 826, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ploys the student model itself as score estimator. Theoretically, an intrinsic correspondence for DMD and the distillation loss in ADD can be equivalently found between Variational Score Distillation (VSD) [74] and Score Distillation Sampling (SDS) [51] in text-to-3D generation, where SDS represents a specialized instance of VSD through the use of a single-point Dirac distribution as the variational distribution [74]. And thus the VSD loss in DMD performs better than the SDS loss in ADD. Intuitively, since the capacity of a diffusion model decreases significantly when distilled for few-step generations [23], the student model no longer serves as a decent score estimator as the teacher. Therefore, the score-based distillation loss in ADD hardly contributes to its final performance.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the optimization of DMD loss relies on the reverse Kullback-Leibler (KL) divergence minimization which is zero-forcing that drives low-probability regions to zero, causing the model to focus on only a few dominant modes and potentially leading to mode collapse [45]. To enhance sample diversity, DMD [86] employs an ODE-based (Ordinary Differential Equation) regularizer with synthetic data, while DMD2 [85] introduces a GAN-based (Generative Adversarial Network [8]) regularizer with real data to counterbalance this side effect. Subsequent efforts such as Moment Matching Distillation (MMD) [59], Score identity Distillation (SiD) [93] and Score Implicit Matching (SIM) [38] all used a variant of Fisher divergence to align the fake score estimator with pre-trained real score estimator [16]. Despite great success, their ability for distribution matching is limited by their dependence on a predefined form of explicit divergence metric. In this work, we would like to raise and investigate a question: Can we bypass the limitations of a predefined divergence by developing a framework that learns an implicit, data-driven discrepancy measure, thereby enabling more flexible and fine-grained matching of complex, high-dimensional distributions?", + "bbox": [ + 91, + 305, + 482, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the multifaceted alignment requirements in complex multimodal text-conditioned image or even video generation may not be fully captured, this motivates our exploration of more adaptive discrepancy learning paradigms. As our first contribution, we show how to align the latent predictions between real and fake score estimators to realize score distillation in an adversarial manner with prior knowledge of teacher model and dynamically learnable parameters, termed Adversarial Distribution Matching (ADM). This is different from the ODE-based or GAN-based regularizer in DMD and DMD2 that is used to counterbalance the mode collapse effect of reverse KL divergence. In contrast, we are performing distribution matching by means of GAN training to replace and circumvent the use of DMD loss, which we will provide more discussion in Sec. 4.1.3.", + "bbox": [ + 89, + 641, + 482, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our second contribution concerns about the extremely challenging one-step distillation, where we observe that the", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "score distillation has a higher risk of gradient exploding and vanishing. We attribute the issue more to the lower overlap of support sets between the student and teacher distributions, not just the approximation errors in the fake score estimator as attributed in DMD2 [85]. In other words, while score distillation yields superior generation quality, it imposes higher requirements on initialization especially when distilled in extremely few steps, which we will provide more in-depth analysis in Sec. 4.3.2. Although we notice that an ODE-based pre-training on synthetic data is used for SDXL one-step distillation in DMD2 implementation, the mean squared error loss was probably still not enough to provide more overlapping regions of support sets between the student and teacher distributions. Our experiments demonstrate that when providing a better initialization for distribution matching, the effect of Two Time-scale Update Rule (TTUR) [2] to the final performance is very limited.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Naturally, our third contribution focuses on providing a better initialization for further score-based fine-tuning. We employ adversarial distillation to trade off sample quality and mode converge inspired by SDXL-Lightning [23] and LADD [60]. With this distribution-level loss optimization, we can pre-train the student model to capture more potential modes of the teacher model distribution, especially through a hybrid discriminator in both latent and pixel spaces which we will introduce later in Sec. 4.2. To facilitate the diversity, we also propose employing a cubic timestep schedule for the generator to bias towards higher noise levels.", + "bbox": [ + 511, + 347, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By combining adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our one-step SDXL provides competitive fidelity compared to the baseline with $50 \\times$ acceleration in the Number of Function Evaluation (NFE) as shown in Fig. 1. More experiments on multi-step ADM distillation across the best existing diffusion models including SD3-Medium, SD3.5-Large [6], and CogVideoX [83] consistently achieve new benchmarks for efficient image and video synthesis.", + "bbox": [ + 511, + 513, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 662, + 653, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Progressive Distillation was proposed in [58] to distill multi-step prediction into one-step prediction of the same distance along the trajectory. SDXL-Lightning [23] extended this idea by utilizing GAN training and achieved one-step high-resolution (1024px) generation for the first time. This approach involving multi-stage process can be quite cumbersome, as it requires iteratively distilling from its predecessor, halving the sampling steps each time.", + "bbox": [ + 511, + 688, + 903, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consistency Distillation proposed [7, 34, 35, 64, 66] to enforce the consistency property into diffusion model, i.e. predictions towards original sample are consistent for arbitrary pairs of noisy timesteps belong to the same trajectory. Subsequent efforts extended it into the trajectory consistency to relax the training objectives, i.e. predictions towards noisy", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16819", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "samples of arbitrary subsequent timesteps are consistent, including CTM [18], TCD [92], TSCD [55] and PCM [69].", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rectified Flow [27, 28] aims to obtain faster straight trajectories through multiple reflow processes that iteratively learn the velocity of many ODE pairs from its predecessor. PeRFlow [81] tried splitting the trajectory into pieces and applying piece-wise rectification with real data samples. In this paper, we empirically found the straightness can also be satisfied via an adversarial distillation paradigm without the need for repeatedly collecting tremendous synthetic data.", + "bbox": [ + 88, + 121, + 482, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Score Distillation intuitively tries to keep that a sample appearing in the student model distribution at a specific noise level is with the same probability as it does in the teacher model distribution. In light of the motivation that relying on a single divergence might be problematic, a concurrent work Score-of-Mixture Distillation (SMD) [16] shares our perspective and explicitly designed a class of $\\alpha$ -skew Jensen-Shannon (JS) divergences for optimization. In contrast, we implicitly measure the discrepancy between the fake and real score estimators in an adversarial manner, facilitating a more capable and adaptive discrepancy learning paradigm along the distillation process.", + "bbox": [ + 89, + 242, + 483, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Distillation employs a discriminator to align the student model distribution with a specific target distribution [40, 94]. While methods UFOGen [79], DMD2 [85], and APT [24] directly align with the real data distribution, SDXL-Lightning [23] and Hyper-SD [55] utilize intermediate timestep predictions from the teacher model as approximation objectives. Inspired by LADD [60], our adversarial distillation pre-training aligns with ODE-based synthetic data generated from the teacher model.", + "bbox": [ + 89, + 422, + 483, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human Feedback Learning was firstly considered by ReFL [78] in diffusion models that includes two stages of reward model training and preference fine-tuning. HyperSD [55] treated this as a standalone technique, ultimately applying LoRA insertion to shift the output distribution of the low-step generator toward human preferences. Subsequent studies, notably [36, 39], further attempted to integrate CFG and Score-based divergence.", + "bbox": [ + 89, + 559, + 483, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 691, + 228, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 717, + 251, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a training data distribution $p_{\\mathrm{data}}$ with standard deviation $\\sigma_{\\mathrm{data}}$ , the diffusion model [12] generates samples by reversing the forward diffusion process that progressively adds noise to a data sample $x_0 \\sim p_{\\mathrm{data}}$ as,", + "bbox": [ + 89, + 739, + 482, + 800 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x}\\right) \\sim \\mathcal {N} \\left(\\alpha_ {t} \\boldsymbol {x}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 804, + 480, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\geq 0, \\sigma_{t} > 0$ are specified noise schedules such that $\\alpha_{t} / \\sigma_{t}$ satisfies monotonically decreasing w.r.t. $t$ and larger $t$ indicates greater noise. We consider two different formulations for denoising models.", + "bbox": [ + 89, + 825, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DDPM and DDIM [12, 63] assume discrete-time schedules", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $t \\in [1, T]$ (typically $T = 1000$ ) and noise prediction parameterization* [58]. The training objective is given by,", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 125, + 903, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function and $\\epsilon_{\\theta}$ is a neural network with parameters $\\theta$ . The noise schedule is defined as $\\alpha_{t} = \\sqrt{\\bar{\\alpha}_{t}}, \\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}}$ such that $\\boldsymbol{x}_t = \\sqrt{\\bar{\\alpha}_t} \\boldsymbol{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon$ . For DDIM sampling, it solves the Probability-Flow Ordinary Differential Equations (PF-ODE) [65] by $d\\bar{\\boldsymbol{x}}_t = \\epsilon_\\theta (\\frac{\\bar{\\boldsymbol{x}}_t}{\\sqrt{\\bar{\\sigma}_t^2 + 1}}) d\\bar{\\sigma}_t$ , where $\\bar{\\boldsymbol{x}}_t = \\frac{\\boldsymbol{x}_t}{\\sqrt{\\bar{\\alpha}_t}}$ and $\\bar{\\sigma}_t = \\sqrt{\\frac{1 - \\bar{\\alpha}_t}{\\bar{\\alpha}_t}}$ , starting from $\\boldsymbol{x}_T \\sim \\mathcal{N}(\\boldsymbol{0}, \\boldsymbol{I})$ and stopping at $\\boldsymbol{x}_0$ .", + "bbox": [ + 511, + 146, + 906, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flow Matching [26, 27, 30] uses velocity prediction parameterization [58] and continuous-time coefficients (typically $\\alpha_{t} = 1 - t$ , $\\sigma_{t} = t$ with $t \\in [0, T = 1]$ ). The conditional probability path or the velocity we say can be given as $\\pmb{v}_{t} = \\frac{d\\alpha_{t}}{dt}\\pmb{x}_{0} + \\frac{d\\sigma_{t}}{dt}\\epsilon$ , such that the training objective is,", + "bbox": [ + 511, + 263, + 905, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\boldsymbol {v} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {v} _ {t} \\| _ {2} ^ {2} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 343, + 903, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w(t)$ is a weighting function and $\\pmb{v}_{\\theta}$ is a neural network parameterized by $\\theta$ . The sampling procedure starts from $t = T$ with $\\pmb{x}_T \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I})$ and stops at $t = 0$ , solving the PF-ODE by $d\\pmb{x}_t = \\pmb{v}_{\\theta}(\\pmb{x}_t, t)dt$ .", + "bbox": [ + 511, + 364, + 905, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Distribution Matching Distillation", + "text_level": 1, + "bbox": [ + 511, + 436, + 810, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DMD [85, 86] distills pretrained diffusion models $\\pmb{F}_{\\phi}(\\pmb{x}_t,t)$ into one-step or multi-step efficient generators $G_{\\theta}(\\pmb{x}_t,t)$ by minimizing the reverse KL divergence between the target distribution $p_{\\mathrm{real}}$ and the efficient generator output distribution $p_{\\mathrm{fake}}$ . The gradient of DMD objective w.r.t. $\\theta$ is,", + "bbox": [ + 511, + 458, + 905, + 534 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D M D}} = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} - \\left[ \\left(s _ {\\text {r e a l}} \\left(\\boldsymbol {x} _ {t}\\right) - s _ {\\text {f a k e}} \\left(\\boldsymbol {x} _ {t}\\right)\\right) \\frac {d \\boldsymbol {G} _ {\\theta} \\left(\\boldsymbol {z} , t ^ {\\prime}\\right)}{d \\theta} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 537, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $t'$ is randomly selected from predefined generator schedule, $t \\sim \\mathcal{U}(0,T)$ , and noisy samples $\\boldsymbol{x}_t = \\boldsymbol{q}(\\boldsymbol{x}_t|\\hat{\\boldsymbol{x}}_0)$ are obtained by randomly diffusing the generator output $\\hat{\\boldsymbol{x}}_0 = \\boldsymbol{G}_{\\theta}(\\boldsymbol{z},t')$ . Score functions $s_{\\mathrm{real}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{real}}(\\boldsymbol{x}_t)$ , $s_{\\mathrm{fake}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{fake}}(\\boldsymbol{x}_t)$ are vector fields that point towards higher density of data at a given noise level [17, 65] for $p_{\\mathrm{real}}$ and $p_{\\mathrm{fake}}$ , respectively.", + "bbox": [ + 511, + 568, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the real score estimator is the teacher model $\\mathbf{F}_{\\phi}(\\mathbf{x}_t,t)$ itself, the fake score estimator $\\pmb{f}_{\\psi}(\\pmb{x}_t,t)$ is initialized the same as $\\pmb{F}_{\\phi}(\\pmb{x}_t,t)$ and dynamically learned to describe $p_{\\mathrm{fake}}$ with pretrain loss as Eqs. (2) and (3). In practice, the gradient in Eq. (4) is computed as,", + "bbox": [ + 511, + 675, + 905, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {g r a d} \\left(\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t\\right) = \\frac {\\boldsymbol {f} _ {\\psi} \\left(\\boldsymbol {x} _ {t} , t\\right) - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\left\\| \\hat {\\boldsymbol {x}} _ {0} - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right) \\right\\| _ {1}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 753, + 903, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such that the training loss is implemented like,", + "bbox": [ + 511, + 789, + 821, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D M D}} (\\theta) = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} [ \\| \\hat {\\boldsymbol {x}} _ {0} - s g (\\hat {\\boldsymbol {x}} _ {0} - \\operatorname {g r a d} (\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t)) \\| _ {2} ^ {2} ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 808, + 903, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $sg(\\cdot)$ denotes the stop gradient operation.", + "bbox": [ + 511, + 837, + 831, + 852 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*This is not mandatory for works like CogVideoX [83] which also used DDIM sampling but parameterized in velocity. The different parameterizations can be equivalently transformed into each other [58].", + "bbox": [ + 511, + 862, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16820", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e44a13f3cc8690a0fafb9e59c1881a073abc346d726524c940fcb433605a9033.jpg", + "image_caption": [ + "ADM" + ], + "image_footnote": [], + "bbox": [ + 143, + 107, + 844, + 228 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51df300708192b09c7df88f42838958f4f8e97d8e6a1c9cc8c73c573f7edf2be.jpg", + "image_caption": [ + "ADP", + "Figure 2. Overall pipeline of our proposed Adversarial Distribution Matching (ADM) and Adversarial Distillation Pre-training (ADP)." + ], + "image_footnote": [], + "bbox": [ + 143, + 244, + 844, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 470, + 225, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Adversarial Distribution Matching", + "text_level": 1, + "bbox": [ + 89, + 498, + 393, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of using a predefined divergence between the fake and real distributions, we use an implicit, data-driven discrepancy measure through an adversarial discriminator. Specifically, our discriminator $D_{\\tau}(\\pmb{x}_t,t)$ consists of a frozen latent diffusion model initialized the same as teacher model $F_{\\phi}(\\pmb{x}_t,t)$ and multiple trainable heads added upon different UNet [57] or DiT [49] blocks. Given the noisy sample $\\pmb{x}_t = \\pmb{q}(\\pmb{x}_t|\\hat{\\pmb{x}}_0)$ diffused from the output of few-step generator $\\hat{\\pmb{x}}_0 = \\pmb{G}_{\\theta}(\\pmb{z},t')$ , the score estimators no longer solve the PF-ODE w.r.t. the end point $\\pmb{x}_0^{\\mathrm{fake}} = \\pmb{f}_{\\psi}(\\pmb{x}_t,t)$ and $\\pmb{x}_0^{\\mathrm{real}} = \\pmb{F}_{\\phi}(\\pmb{x}_t,t)$ as used in Eq. (5).", + "bbox": [ + 89, + 523, + 483, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead, we set a fixed timestep interval $\\Delta t$ (defaults to $T / 64$ ) and solve the PF-ODE w.r.t. $(t - \\Delta t)$ , such that the fake sample $\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}$ and real sample $\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}$ can be obtained to serve as score predictions and sent into the discriminator. The discriminator hierarchically aggregates features from frozen backbone layers and dynamically weights them through multiple learnable heads, establishing an adaptive discrepancy metric that leverages both diffusion priors and data-driven trainable dynamics. We use Hinge loss [22] to train the generator $\\pmb{G}_{\\theta}(\\pmb{x}_t,t)$ and discriminator $D_{\\tau}(\\pmb{x}_t,t)$ alternatively. This encourages the fake score prediction $\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}$ to be closer to the real score prediction $\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}$ :", + "bbox": [ + 89, + 691, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}} {\\mathbb {E}} \\left[ - \\boldsymbol {D} _ {\\tau} \\left(\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t\\right) \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 877, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} (\\tau) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, \\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {r e a l}}} {\\mathbb {E}} [ \\max (0, 1 + D _ {\\tau} (\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t)) \\tag {8} \\\\ \\left. + \\max \\left(0, 1 - D _ {\\tau} \\left(x _ {t - \\Delta t} ^ {\\text {r e a l}}, t - \\Delta t\\right)\\right) \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 468, + 905, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In conjunction with the dynamically learned fake model, we clarify our training procedure in Appendix A. The overall pipeline is demonstrated in Fig. 2.", + "bbox": [ + 511, + 521, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Motivation of Discriminator Timestep $(t - \\Delta t)$", + "text_level": 1, + "bbox": [ + 511, + 577, + 883, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the final objective of score distillation is to match the probability flows varying with noise levels of student and teacher models to exactly the same, timestep information must be considered when measuring the discrepancy between distributions. This coincides with our discriminator design that uses a pre-trained diffusion model, and we take a small step alongside the PF-ODE, succeeding in preserving the input timestep information of score estimators.", + "bbox": [ + 511, + 597, + 906, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Data-driven Effect", + "text_level": 1, + "bbox": [ + 511, + 729, + 689, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The flexibility of using discriminators for distributional discrepancy measure is not only in the noise level of score function, but also within the distillation process. As distillation iterates, the model is exposed to increasingly diverse data, causing the discrepancies in modes between the two distributions to change. In the early phase of training where the discrepancy is significant, a more global evaluation is necessary, whereas in the later phases, when the discrepancy becomes minor, a more localized, fine-grained optimization might become essential. In other words, driven", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/45d0869ca4553ec88ba79e6c155f4aa6a56f3aa919ecde8d7203fb67071d8a89.jpg", + "image_caption": [ + "Figure 3. Changes of DMD loss over multi-step ADM distillation for CogVideoX. Note that we did not optimize this objective directly during ADM distillation but recorded it over iterations." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 478, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by the volume of data, the divergence measure employed across different training phases can vary.", + "bbox": [ + 89, + 239, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Relationship with DMD and DMD2", + "text_level": 1, + "bbox": [ + 89, + 277, + 385, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the mode collapse issue in DMD loss, an ODE-based regularizer and a GAN-based regularizer are additionally used for distillation in DMD [85] and DMD2 [85], respectively. However, these two regularizers do not fundamentally address the mode-seeking behavior introduced by reverse KL divergence as shown in Fig. 4(a), but rather counterbalance it by trade-offs between losses. In ADM, our adversarial loss is actually playing the role of DMD loss to realize score distillation with an implicit, data-driven discrepancy measure instead of a predefined divergence. Therefore, our motivation for using GAN training in ADM is different from that of DMD2 [85] and we don't require additional regularizers.", + "bbox": [ + 89, + 296, + 483, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, the learnable discriminator can approximate any nonlinear function to implicitly measure distribution divergence, which probably inherently encompasses the reverse KL divergence in DMD loss. As shown in Fig. 3, we visualize the changes of DMD loss in Eq. (6) during the multi-step ADM distillation on CogVideoX [83]. Though not directly optimize on Eq. (6), the results indicate a very steady downward trend that supports our assumption. More theoretical discussion is provided in Sec. 4.3.3.", + "bbox": [ + 89, + 492, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Adversarial Distillation Pre-training", + "text_level": 1, + "bbox": [ + 89, + 637, + 406, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To stabilize the extremely difficult one-step distillation, we opt to provide a better initialization for ADM fine-tuning with adversarial distillation pre-training on synthetic data. Our pre-training configuration refers to Rectified Flow [27] in several aspects, where we 1) collect the ODE pairs from teacher model in an offline manner, 2) construct noisy samples by linearly interpolation between pure noise and clean data sample of the ODE pair, 3) alter the prediction target of the generator to the velocity of ODE pair.", + "bbox": [ + 89, + 659, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for adversarial training, we formulate a latent-space discriminator $D_{\\tau_1}(\\boldsymbol{x}_t,t)$ initialized from teacher model and a pixel-space discriminator $D_{\\tau_2}(\\boldsymbol{x})$ initialized from the vision encoder of SAM [19] model, respectively, as shown in Fig. 2. We also append multiple trainable heads to both the backbone networks similar to the practice in ADM. All these contribute to increasing the discriminative capability,", + "bbox": [ + 89, + 796, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "facilitating the student model in discovering more potential modes in the teacher model distribution. Specifically, let $\\tilde{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)$ denote the predicted PF-ODE endpoint of generator, where $\\pmb{x}_{t}$ represents the noisy sample interpolated between an ODE pair $(\\pmb {x}_T,\\pmb {x}_0)$ with random timestep $t\\in [0,T]$ . For latent-space discriminator, we diffuse the generator output with another random noise and timestep $t^{\\prime}\\in (0,T]$ , getting $\\tilde{\\pmb{x}}_{t^{\\prime}} = \\pmb {q}(\\tilde{\\pmb{x}}_{t^{\\prime}}|\\tilde{\\pmb{x}}_0)$ as its input. For pixel-space discriminator, the generator output will be first decoded via VAE decoder and then sent into the vision encoder. The training objective, also based on Hinge loss [22], encourages the generator output $\\tilde{\\pmb{x}}_0$ to be closer to the synthetic data sample $\\pmb{x}_0$ :", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} - \\left[ \\lambda_ {1} D _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) + \\lambda_ {2} D _ {\\tau_ {2}} \\left(\\tilde {\\boldsymbol {x}} _ {0}\\right) \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 290, + 906, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} \\left(\\tau_ {1}, \\tau_ {2}\\right) = \\underset {\\boldsymbol {x} _ {0}, \\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} \\left[ \\lambda_ {1} \\cdot \\max \\left(0, 1 + \\boldsymbol {D} _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right. \\\\ + \\lambda_ {2} \\cdot \\max \\left(0, 1 + D _ {\\tau_ {2}} \\left(\\tilde {x} _ {0}\\right)\\right) \\tag {10} \\\\ + \\lambda_ {1} \\cdot \\max (0, 1 - D _ {\\tau_ {1}} \\left(\\boldsymbol {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)) \\\\ \\left. \\left. + \\lambda_ {2} \\cdot \\max \\left(0, 1 - D _ {\\tau_ {2}} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right] \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 323, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We empirically found that setting balancing coefficients $\\lambda_{1} = 0.85, \\lambda_{2} = 0.15$ produces visually coherent results.", + "bbox": [ + 511, + 407, + 906, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Cubic Generator Timestep Schedule", + "text_level": 1, + "bbox": [ + 511, + 454, + 812, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitively, higher noise levels encourage exploration of new modes by weakening the restrictive information encoded in latent representations. Therefore, we propose employing a cubic timestep schedule for the generator. The schedule maps uniform $[0,T)$ samples through $[1 - (t / T)^3 ]*T$ , non-linearly concentrating values near $T$ with heavy noise similar to LADD [60].", + "bbox": [ + 511, + 474, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Uniform Discriminator Timestep Schedule", + "text_level": 1, + "bbox": [ + 511, + 598, + 854, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike in ADM where the discriminator is firstly used for score distillation, the utilization of latent-space discriminator is common for adversarial distillation works. Inspired by SDXL-Lightning [23], where they found that the diffusion encoder is trained to focus on high-frequency details at lower timesteps and low-frequency structures at higher timesteps, we set a uniform $(0,T]$ for discriminator timestep $t'$ to capture both advantages during pre-training.", + "bbox": [ + 511, + 619, + 906, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3. Relationship with LADD", + "text_level": 1, + "bbox": [ + 511, + 757, + 733, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our motivation to perform adversarial distillation on synthetic data is inspired by LADD [60], but differs in a lot that we 1) construct noisy samples via ODE pairs in the style of Rectified Flow [27] rather than random noise, 2) develop a cubic generator timestep schedule that facilitate deterministic Euler sampling rather than consistency ones, 3) introduce an additional pixel-space encoder to increase the capability of discriminator and find more modes.", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16822", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Discussion", + "text_level": 1, + "bbox": [ + 89, + 90, + 207, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Difference between ADM and ADP", + "text_level": 1, + "bbox": [ + 89, + 113, + 382, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "One question may appear like what is the difference between these two adversarial approaches with latent-space discriminators? The effectiveness in score distillation correlates with the fact that score function $\\nabla_{\\pmb{x}}\\log p(\\pmb{x};\\sigma (t))$ is defined at different noise levels $\\sigma (t)$ . In contrast, adversarial distillation only aligns the distribution of clean data samples when $t = 0$ . While the latent-space discriminator in pre-training is capturing information at different scales and details by randomly diffusing the generator output, the crux for ADM is more than that by solving the PF-ODE of both score estimators. In other words, ADM additionally supervises the complete denoising process through noisy samples that are with higher density at different noise levels of respective distribution.", + "bbox": [ + 88, + 132, + 482, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This leads to a situation in ADM when two distributions initialized with less overlap in support sets, the noisy samples remain in regions unfamiliar to each other, and the discriminator can easily distinguish between them thus leading to extreme gradient signals. However, since Gaussian noise is isotropic, we artificially create overlapping regions for the randomly diffused samples in ADP to make the discrimination more difficult, resulting in relatively smooth gradients. Therefore, our ADM still falls under score distillation given that it encourages the entire probability flows to be closer, while the pre-training belongs to adversarial distillation because it only cares about the clean data distribution at $t = 0$ .", + "bbox": [ + 89, + 344, + 483, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2. Importance of Pre-training", + "text_level": 1, + "bbox": [ + 89, + 534, + 326, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another question we haven't discussed so far is why do we need pre-training for one-step score distillation? Taking the reverse KL divergence used by DMD loss as an example:", + "bbox": [ + 89, + 553, + 483, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {D} _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\int p _ {\\text {f a k e}} (\\boldsymbol {x}) \\log \\frac {p _ {\\text {f a k e}} (\\boldsymbol {x})}{p _ {\\text {r e a l}} (\\boldsymbol {x})} d \\boldsymbol {x}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 602, + 482, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When employing one-step distillation, the generator output is worse in visual fidelity and structural integrity compared to multi-step sampling, leading $p_{\\mathrm{fake}}(\\pmb{x}) \\to 0$ where $p_{\\mathrm{real}}(\\pmb{x}) > 0$ in regions. The integrand $p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{p_{\\mathrm{real}}(\\pmb{x})}$ approaching zero $0 \\cdot (-\\infty)$ causes the optimization to avoid regions where $p_{\\mathrm{real}}(\\pmb{x}) > 0$ but $p_{\\mathrm{fake}}(\\pmb{x})$ has negligible density, a phenomenon called zero-forcing. Instead of fully covering $p_{\\mathrm{real}}$ 's support, $p_{\\mathrm{fake}}$ collapses to a subset of modes of $p_{\\mathrm{real}}$ , inducing mode-seeking behavior as illustrated in Fig. 4(a). During training this manifests itself as gradient vanishing sometimes. And conversely, the fuzzy samples that the diffusion model typically produces through one step are also not within the teacher model distribution, yielding $p_{\\mathrm{real}}(\\pmb{x}) \\to 0$ where $p_{\\mathrm{fake}}(\\pmb{x}) > 0$ in regions and the integrand $p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{0}$ diverging to $+\\infty$ , and thus resulting in numerical instability and gradient exploding.", + "bbox": [ + 89, + 637, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similarly, when the support sets of the student and", + "bbox": [ + 109, + 886, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/562e6a4f4d7a2a7a42b3338e938721c3c930037b14d245b68b42d0041bb47780.jpg", + "image_caption": [ + "Figure 4. Illustration for theoretical discussion." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 906, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "teacher distributions overlap hardly at all, forward KL divergence approaches $+\\infty$ where $p_{\\mathrm{fake}}(\\boldsymbol{x}) > 0$ , while JS divergence saturates to a constant log 2 and Fisher divergence may degenerate without definition. Therefore, when this assumption is undermined, many of the single divergences no longer apply and a better initialization with more overlapping regions becomes essential as exemplified in Fig. 4(b).", + "bbox": [ + 511, + 392, + 906, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.3. Theoretical Objective", + "text_level": 1, + "bbox": [ + 511, + 510, + 710, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The last question is why ADM is better than DMD loss theoretically? In fact, the Hinge GAN [22] we use has been proven to minimize Total-Variation Distance (TVD) [68]:", + "bbox": [ + 511, + 531, + 906, + 577 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT V \\left(p _ {\\text {f a k e}}, p _ {\\text {r e a l}}\\right) = \\int \\left| p _ {\\text {f a k e}} (\\boldsymbol {x}) - p _ {\\text {r e a l}} (\\boldsymbol {x}) \\right| d \\boldsymbol {x} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 580, + 905, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "That is to say, when the discriminator is sufficiently rich and well trained, the theoretical optimum of Hinge loss upon convergence minimizes TVD. When the support sets of fake and real distributions have minimal overlap, TVD provides two key advantages over reverse KL divergence: 1) Symmetry: TVD yields the same discrepancy measure regardless of the initial distribution, whereas the asymmetric reverse KL may exhibit mode-seeking behavior and neglect other portions of the overall distribution. For example, as illustrated in Fig. 4(c), TVD maintains substantial loss values and provides optimization directions covering the mode when $p_{\\mathrm{fake}}(\\pmb{x}) \\to 0$ while $p_{\\mathrm{real}}(\\pmb{x}) > 0$ , whereas the reverse KL divergence suffers from gradient vanishing issues in such scenarios, as discussed in Sec. 4.3.2. 2) Boundedness: TVD is bounded within [0,1], and thus mitigating disruptions from outliers during training, especially in our high-dimensional multimodal text-conditioned image and video distributions, avoiding the numerical instability inherent to reverse KL divergence caused by gradient exploding.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/891ff9a75ab171bc5ebd6a3e8d67cf8292b85ab3f52c8faba05c8a0054b19910.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodStep NFECLIP ScorePick ScoreHPSv2MPS
ADD [61] (512px)1135.008822.152427.097110.4340
LCM [34]1228.466920.126723.82464.8134
Lightning [23]1133.498521.919427.155710.2285
DMD2 [85]1135.215322.097827.452310.6947
DMDX (Ours)1135.255722.273627.704611.1978
SDXL-Base [56]255035.030922.249427.374310.7042
", + "bbox": [ + 93, + 89, + 478, + 203 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f1d568af66c55eb0b1b112c6111a81be75b732fc9af8c84c6b4ae156801688c7.jpg", + "table_caption": [ + "Table 1. Quantitative results on fully fine-tuning SDXL-Base." + ], + "table_footnote": [], + "table_body": "
MethodStepNFECLIP ScorePick ScoreHPSv2MPS
TSCD [61]4834.018521.966527.272810.8600
PCM [69] (Shift=1)4433.504221.970327.368010.5707
PCM [69] (Shift=3)4433.381821.939627.114610.5635
PCM [69] (Stoch.)4433.418521.882227.317710.5200
Flash [3]4434.397822.090427.258610.6634
ADM (Ours)4434.907622.547128.449211.9543
SD3-Medium [6]255034.763322.296127.973311.3652
LADD [60]4434.739522.395827.492311.4372
ADM (Ours)4434.973022.884227.733112.2350
SD3.5-Large [6]255034.966822.508727.968811.5826
", + "bbox": [ + 93, + 229, + 478, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/699efdce4f2c37a22ff69b58fc9ec44afa7f59dfc48a523cb40139d8b4c5e9e3.jpg", + "table_caption": [ + "Table 2. Quantitative results on LoRA fine-tuning SD3-Medium and fully fine-tuning SD3.5-Large." + ], + "table_footnote": [], + "table_body": "
MethodStepNFEFinal ScoreQuality ScoreSemantic Score
ADM8878.58480.82569.621
+ Longer Training ×28880.76483.03171.693
ADM w/ CFG81679.86580.93875.569
+ Longer Training ×281681.79683.00876.947
CogVideoX-2b [83]10020080.03680.80176.974
ADM8882.06783.22777.423
ADM w/ CFG81680.98282.16576.251
CogVideoX-5b [83]10020081.22681.78578.987
", + "bbox": [ + 93, + 438, + 482, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative results on fully fine-tuning CogVideoX.", + "bbox": [ + 102, + 588, + 468, + 602 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 614, + 223, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models. For one-step distillation, we employ Adversarial Distillation Pre-training (ADP) and ADM fine-tuning on SDXL-Base [56], termed DMDX. For multi-step distillation, we only employ ADM training on both text-to-image models SD3-Medium, SD3.5-Large [6], and text-to-video models CogVideoX-2b, CogVideoX-5b [83]. Following most concurrent works, we didn't use Classifier-Free Guidance (CFG) [11] in our text-to-image model. We tried conducting CFG-integrated experiments on the text-to-video model, with the approach detailed in Sec. 5.2.", + "bbox": [ + 89, + 641, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. No visual data is required for both ADP and ADM proposed in this work. For image generators, we utilize the text prompts from JourneyDB [67] that exhibits high-level of detail and specificity for training. For video generators, we collect training prompts from OpenVid-1M [47], Vript [82] and Open-Sora-Plan-v1.1.0 [50].", + "bbox": [ + 89, + 792, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. The image generators are evaluated on 10K", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/242459914feae3653c71bb6d02d3a2a9f2dd3825aa50466cd4eb0e0f04af7879.jpg", + "image_caption": [ + "Figure 5. Qualitative results on fully fine-tuning SDXL-Base." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prompts from COCO 2014 [25] following DMD2 [85]. We report the CLIP score [52] and human preference benchmarks PickScore [20], HPSv2 [75] and MPS [90] as many concurrent works including Hyper-SD [55] and Emu3 [73]. But we don't include Hyper-SD in the one-step quantitative comparison because one-step Hyper-SDXL has been optimized directly on human feedback with ReFL [78]. Instead, we compare with the TSCD algorithm proposed therein on SD3-Medium [13], since 4-step Hyper-SD3 LoRA is without ReFL optimization. The video generators are evaluated on VBenchmark [14], which consists of multiple quality and semantic dimensions comprehensively.", + "bbox": [ + 511, + 484, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters. Despite having multiple models to train in our proposed ADP and ADM, we achieve satisfactory visual fidelity and structural integrity without extensive hyperparameter tuning. For the remainder of this work, we only adjust the learning rate of generator for different experiments. The optimizer settings for discriminators and fake models are uniform across all the experiments. Unless longer training is specifically noted, we train the same for only 8K iterations with a batch size of 128 and 8 for text-to-image and text-to-video models, respectively. More specific implementation details are provided in Appendix B.", + "bbox": [ + 511, + 667, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Efficient Image Synthesis", + "text_level": 1, + "bbox": [ + 511, + 847, + 743, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1 quantitatively compares our two-stage approach combing ADP and single-step ADM distillation with ex", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16824", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/14577d22c29c8d9c944921ca95a76d884d47fbce21d3a140faf305c7fe9a7b6e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AblationCLIP ScorePick ScoreHPSv2MPS
Ablation on adversarial distillation.
A1: Rectified Flow [27]27.437620.021123.60934.4518
A2: DINOv2 as pixel-space34.183621.875027.103910.2407
A3: λ1 = 0.7, λ2 = 0.333.694321.634426.89029.9633
A4: λ1 = 1.0, λ2 = 0.033.892921.739526.786910.0757
A5: w/o ADM (ADP only)35.772322.009527.349910.6646
Ablation on score distillation.
B1: ADM w/o ADP32.502021.763126.873210.8986
B2: DMD Loss w/o ADP32.748221.034125.96808.8977
B3: DMD Loss w/ ADP34.511921.936627.398510.6046
B4: DMDX (Ours)35.255722.273627.704611.1978
", + "bbox": [ + 99, + 88, + 470, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "isting one-step distillation methods on fully fine-tuning SDXL-Base. The results show that our method achieves excellent performance on both image-text alignment and human preference, which is consistent with the qualitative comparisons in Fig. 5, including better portrait aesthetic, animal hair details, subject-background separation and physical structure.", + "bbox": [ + 89, + 296, + 482, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For multi-step ADM distillation, it can serve as a standalone score distillation method. We tried both fully finetuning and LoRA fine-tuning [13] configurations and the quantitative results in Tab. 2 demonstrate our superior performance. Qualitative results are provided in Appendix C.", + "bbox": [ + 89, + 402, + 483, + 478 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Efficient Video Synthesis", + "text_level": 1, + "bbox": [ + 89, + 484, + 316, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As quantitatively shown in Tab. 3, except for the regular 8-step ADM distillation for both sizes of CogVideoX, we also try integrating Classifier-Free Guidance (CFG) [11] for text-to-video task. Specifically, we assign the CFG scale for real model by randomly sampling within the range [5.0, 7.0], while assigning the few-step generator's scale through explicit subtraction of 2.0 from the real model's value. Empirically, we found that this subtractive offset requires progressive enlargement as the targeted sampling steps decreases. There is no CFG for the fake model. The VBench [14] results demonstrate that our few-step generators achieve comparable performance to the base model with $92 - 96\\%$ acceleration. We conduct additional evaluations on the longer training of 2B model, motivated by the observation that DMD loss did't converge well at 8K iterations as Fig. 3 indicates. Experimental results demonstrate that the DMD loss is also approximately optimized by the learnable discriminator during ADM distillation. More quantitative and qualitative results refer to Appendix C.", + "bbox": [ + 89, + 507, + 483, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 803, + 254, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 4, we conduct extensive ablation studies on fully fine-tuning SDXL-Base to validate our effectiveness. Qualitative comparisons are provided in Appendix C.", + "bbox": [ + 89, + 825, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of ADP. The following conclusions can be drawn: 1) Single reflow process provides very limited effectiveness", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e76aec7f5110c166a038ace359e3b50a1002c6aba0b5feb8f0620e9d93c553e.jpg", + "table_caption": [ + "Table 4. Quantitative results on ablation studies." + ], + "table_footnote": [], + "table_body": "
TTURTraining TimeCLIP ScorePick ScoreHPSv2MPS
1×1.0035.255722.273627.704611.1978
4×1.8535.258322.277327.725511.2720
8×2.5335.329922.288327.758611.2838
", + "bbox": [ + 535, + 89, + 880, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7146d726e25b72b9b2990c4b037a4d5a1545eb07134d77a74b7e80866957399b.jpg", + "table_caption": [ + "Table 5. Ablation study on two time-scale update rule." + ], + "table_footnote": [], + "table_body": "
ADDLCMLightningDMD2OursTeacher
LPIPS↑0.60710.62570.67070.67150.71560.6936
", + "bbox": [ + 517, + 194, + 898, + 238 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Quantitative diversity evaluation on PartiPrompts [88].", + "bbox": [ + 519, + 239, + 897, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(A1), while multiple processes achieving effective rectification requires considerable computational expense. 2) Using SAM [19] offers more imaging fidelity compared to the widely adopted DINOv2 [48] (A2/A5), which is likely due to SAM's higher resolution of 1024px (compared to DINOv2's 518px). 3) The weighting $\\lambda_{2}$ for pixel space should neither be too large nor too small. Excessive weighting leads to degraded structural integrity (A3/A5), while insufficient weighting results in blurriness (A4/A5). Qualitative results in Appendix C provide more visible distinction.", + "bbox": [ + 511, + 267, + 903, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of ADM. We summarize the key findings: 1) The absence of ADP exerts substantial performance degradation (B1/B4), which is aligned with our analysis in Sec. 4.3.2. 2) Without a regularizer, the DMD loss underperforms compared to standalone ADM (B1/B2), indicating its poor robustness. 3) Although DMD loss optimization also benefits from ADP (B2/B3), its distribution matching capability remains inferior to ADM (B3/B4).", + "bbox": [ + 511, + 419, + 903, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of TTUR. Tab. 5 presents the impact of different TTUR settings on final performance and training duration. The results demonstrate that increasing TTUR yields only marginal performance gains while nearly doubling training time, rendering this trade-off clearly unwarranted. This highlights the critical role of our proposed ADP in one-step distillation and suggests that the training instability in DMD2 likely stems from insufficient support set overlap.", + "bbox": [ + 511, + 540, + 905, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diversity evaluation. Following DMD2 [85], we generate 4 samples per prompt on Partiprompts [88] with different seeds, and report the averaged pairwise LPIPS similarities [89] in Tab. 6. The results indicate that our method significantly outperforms others in diversity. More randomly curated multi-seed samples can be found in Appendix C.", + "bbox": [ + 511, + 662, + 903, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 768, + 635, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We realize that one weakness is that teacher model might require CFG to produce accurate score prediction. Our experiments suggest this is a common characteristic of score distillation methods generally, not a limitation unique to our approach. This limits the application to guidance-distilled models such as FLUX.1-dev [1], which could be a promising research topic for future work.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (U22A2095, 12326618, 62276281), Guangdong Basic and Applied Basic Research Foundation, China (2024A1515011882) and the Project of Guangdong Provincial Key Laboratory of Information Security Technology (2023B1212060026).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024.8", + "[2] Naresh Babu Bynagari. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2", + "[3] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 1, 7, 2", + "[4] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 7, 1", + "[7] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. In ICLR, 2025. 2", + "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1", + "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 1", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 7, 8", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3", + "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 7, 8" + ], + "bbox": [ + 93, + 244, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 7, 8, 2, 3", + "[15] Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models. arXiv preprint arXiv:2309.14509, 2023. 2", + "[16] Tejas Jayashankar, J. Jon Ryu, and Gregory Wornell. Score-of-mixture training: Training one-step generative models made simple via score estimation of mixture distributions. arXiv preprint arXiv:2502.09609, 2025. 1, 2, 3", + "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 1, 3", + "[18] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 1, 3", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In ICCV, pages 3992-4003, 2023. 5, 8, 1", + "[20] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeuriPS, 2023. 7", + "[21] Jonas Kohler, Albert Pumarola, Edgar Schonfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Thabet. Imagine flash: Accelerating emu diffusion models with backward distillation. arXiv preprint arXiv:2405.05224, 2024. 1", + "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 4, 5, 6", + "[23] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2, 3, 5, 7", + "[24] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, Lu Jiang, and ByteDance Seed. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 1, 3", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 7", + "[26] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2023. 3", + "[27] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2022. 1, 3, 5, 8", + "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion-based text-to-image generation. In ICLR, 2024. 1, 3", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 2", + "[30] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 3", + "[31] Yanzuo Lu, Manlin Zhang, Yiqi Lin, Andy J. Ma, Xiaohua Xie, and Jianhuang Lai. Improving pre-trained masked autoencoder via locality enhancement for person re-identification. In PRCV, pages 509–521, 2022. 5", + "[32] Yanzuo Lu, Meng Shen, Andy J Ma, Xiaohua Xie, and JianHuang Lai. Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation. In AAAI, pages 3900-3908, 2024. 5", + "[33] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In CVPR, pages 6420-6429, 2024. 5", + "[34] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2, 7", + "[35] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 1, 2", + "[36] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. In TMLR, 2024. 3", + "[37] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, pages 76525-76546, 2023. 1", + "[38] Weijian Luo, Zemin Huang, Zhengyang Geng, J. Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching. In NeurIPS, 2024. 1, 2", + "[39] Weijian Luo, Colin Zhang, Debing Zhang, and Zhengyang Geng. David and goliath: Small one-step model beats large diffusion with score post-training. In ICML, 2025. 3", + "[40] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. In ICLR, 2025. 3", + "[41] Hongxu Ma, Guanshuo Wang, Fufu Yu, Qiong Jia, and Shouhong Ding. Ms-detr: Towards effective video moment retrieval and highlight detection by joint motion-semantic learning. In ACMMM, 2025. 5", + "[42] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Ji-hong Guan, and Shuiqeng Zhou. Fine-grained zero-shot object detection. In ACMMM, 2025. 5", + "[43] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. arXiv preprint arXiv:2409.11367, 2024. 1", + "[44] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In CVPR, pages 21203-21214, 2025. 5", + "[45] Thomas Minka. Divergence measures and message passing. Microsoft Research, Technical Report, 2005. 2", + "[46] Movie Gen Team. Movie gen: A cast of media foundation models. https://ai.meta.com/static-resource/movie-gen-research-paper, 2024.2", + "[47] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 7", + "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 8, 1, 5", + "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 4, 1", + "[50] PKU-Yuan Lab and Tuzhan AI. Open-sora-plan-v1.1.0. https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0, 2024.7", + "[51] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 7, 6", + "[53] Samyam Rajbhandari, Jeff Rasley, Olatunj Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC, 2020. 2", + "[54] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 1", + "[55] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. In NeurIPS, 2024. 1, 3, 7", + "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 7, 5", + "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 1", + "[58] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 1, 2, 3", + "[59] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. In NeurIPS, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 1, 2, 3, 5, 7", + "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 1, 7", + "[62] Meng Shen, Yanzuo Lu, Yanxu Hu, and Andy J. Ma. Collaborative learning of diverse experts for source-free universal domain adaptation. In ACM MM, pages 2054–2065, 2023. 5", + "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3", + "[64] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In ICLR, 2024. 2", + "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3", + "[66] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 1, 2", + "[67] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 7", + "[68] Zhiqiang Tan, Yunfu Song, and Zhijian Ou. Calibrated adversarial algorithms for generative modelling. Stat, 8:e224, 2019. 6", + "[69] Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Hongsheng Li, and Xiaogang Wang. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024. 1, 3, 7, 2", + "[70] Fu-Yun Wang, Zhaoyang Huang, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Computation-efficient personalized style video generation without personalized video data*. In SIGGRAPH ASIA Technical Communications, 2024. 1", + "[71] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. arXiv preprint arXiv:2410.07303, 2024. 1", + "[72] Shibo Wang and Pankaj Kanwar. Bfloat16: The secret to high performance on cloud tpus. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus, 2019.2", + "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 7", + "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2", + "[75] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7", + "[76] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 1", + "[77] Xuefeng Xiao, Lianwen Jin, Yafeng Yang, Weixin Yang, Jun Sun, and Tianhai Chang. Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition. Pattern Recognition, 72:72-81, 2017. 5", + "[78] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurlPS, 2023. 3, 7", + "[79] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3", + "[80] Zunnan Xu, Zhentao Yu, Zixiang Zhou, Jun Zhou, Xiaoyu Jin, Fa-Ting Hong, Xiaozhong Ji, Junwei Zhu, Chengfei Cai, Shiyu Tang, et al. Hunyuanportrait: Implicit condition control for enhanced portrait animation. In CVPR, pages 15909-15919, 2025. 5", + "[81] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. In NeurIPS, 2024. 1, 3", + "[82] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 7", + "[83] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5, 7", + "[84] Hongwei Yi, Shitong Shao, Tian Ye, Jiantong Zhao, Qingyu Yin, Michael Lingelbach, Li Yuan, Yonghong Tian, Enze Xie, and Daquan Zhou. Magic 1-for-1: Generating one minute video clips within one minute. arXiv preprint arXiv:2502.07701, 2025. 1", + "[85] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 1, 2, 3, 5, 7, 8", + "[86] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, pages 6613-6623, 2024. 1, 2, 3", + "[87] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16828", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022. 8", + "[89] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8", + "[90] Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multidimensional human preference for text-to-image generation. In CVPR, pages 8018-8027, 2024. 7", + "[91] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, Sergey Tulyakov, and Jian Ren. Sf-v: Single forward video generation model. In NeurIPS, 2024. 1", + "[92] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation: Improved latent consistency distillation by semi-linear consistency function with trajectory mapping. arXiv preprint arXiv:2402.19159, 2024. 1, 3", + "[93] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In ICML, 2024. 1, 2", + "[94] Mingyuan Zhou, Huangjie Zheng, Yi Gu, Zhendong Wang, and Hai Huang. Adversarial score identity distillation: Rapidly surpassing the teacher in one step. In ICLR, 2025. 3" + ], + "bbox": [ + 91, + 92, + 482, + 542 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "16829", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_model.json b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfb494afaafc0b17da0113a082d6517ce77df343 --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_model.json @@ -0,0 +1,2941 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.131, + 0.803, + 0.177 + ], + "angle": 0, + "content": "Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.203, + 0.802, + 0.241 + ], + "angle": 0, + "content": "Yanzuo Lu\\(^{1,2}\\), Yuxi Ren\\(^{2}\\), Xin Xia\\(^{2}\\), Shanchuan Lin\\(^{2}\\), Xing Wang\\(^{2}\\), Xuefeng Xiao\\(^{2*}\\), Andy J. Ma\\(^{1,3,4\\dagger}\\), Xiaohua Xie\\(^{1,3,4,5}\\), Jian-Huang Lai\\(^{1,3,4,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.256, + 0.888, + 0.328 + ], + "angle": 0, + "content": "1Sun Yat-Sen University 2ByteDance Seed Vision \n3Guangdong Provincial Key Laboratory of Information Security Technology, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Pazhou Lab (HuangPu), Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.331, + 0.846, + 0.345 + ], + "angle": 0, + "content": "oliveryanzulu@gmail.com, xiaoxuefeng.ailab@bytedance.com, majh8@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.486, + 0.789 + ], + "angle": 0, + "content": "Distribution Matching Distillation (DMD) is a promising score distillation technique that compresses pre-trained teacher diffusion models into efficient one-step or multi-step student generators. Nevertheless, its reliance on the reverse Kullback-Leibler (KL) divergence minimization potentially induces mode collapse (or mode-seeking) in certain applications. To circumvent this inherent drawback, we propose Adversarial Distribution Matching (ADM), a novel framework that leverages diffusion-based discriminators to align the latent predictions between real and fake score estimators for score distillation in an adversarial manner. In the context of extremely challenging one-step distillation, we further improve the pre-trained generator by adversarial distillation with hybrid discriminators in both latent and pixel spaces. Different from the mean squared error used in DMD2 pre-training, our method incorporates the distributional loss on ODE pairs collected from the teacher model, and thus providing a better initialization for score distillation fine-tuning in the next stage. By combining the adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our proposed method achieves superior one-step performance on SDXL compared to DMD2 while consuming less GPU time. Additional experiments that apply multi-step ADM distillation on SD3-Medium, SD3.5-Large, and CogVideoX set a new benchmark towards efficient image and video synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Recent methods for accelerating diffusion models [12, 17, 63, 65] have concentrated on reducing the sampling steps through distillation. The distillation process generally trains" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.362, + 0.616, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.513, + 0.602, + 0.529 + ], + "angle": 0, + "content": "(a) A white polar bear cub wearing a red sitta in a meadow with flowers." + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.362, + 0.707, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.513, + 0.704, + 0.535 + ], + "angle": 0, + "content": "(b) mage, wizard, female, demoness, forest, D&D, fantasy, incricate, elegant, highly detailed, digital painting, artstation, [...]" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.362, + 0.807, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.513, + 0.803, + 0.534 + ], + "angle": 0, + "content": "(c) beautiful nighttime landscape photography of the Rocky Mountains with a crystal blue lake, milky way galaxy sky, serene, dramatic lightning." + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.362, + 0.905, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.513, + 0.896, + 0.53 + ], + "angle": 0, + "content": "(d) A planeplane lacavillo glows over the surface of a sand with swirling mist around it." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.536, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 1. In these images, some are generated by the baseline SDXL via 50NFE, the others with our DMDX in 1NFE. Can you tell which is the accelerated one? Answers in the footnote\\(\\ddagger\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.588, + 0.907, + 0.723 + ], + "angle": 0, + "content": "a more efficient generator (a.k.a student model) that approximates the output distribution of the pretrained teacher. This field has observed multiple pathways developing in parallel with progressive distillation [23, 58], consistency distillation [18, 34, 35, 43, 55, 66, 69, 70, 92], score distillation [3, 16, 21, 38, 55, 59, 61, 84-87, 93], rectified flow [27, 28, 30, 71, 81] and adversarial distillation [23, 24, 43, 55, 60, 61, 85, 91]. They are currently appearing as distinct yet non-exclusive research directions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.724, + 0.909, + 0.876 + ], + "angle": 0, + "content": "Distribution Matching Distillation (DMD) [85, 86] is a promising score distillation [51] approach, which distills the powerful text-to-image diffusion model SDXL-Base [56] into a one-step generator with great fidelity. While DMD's primary contribution lies in introducing additional regularizers to constrain the distribution matching loss, its predecessor Diff-Instruct [37] pioneered the use of a fake score estimator to approximate the student model's output distribution. This contrasts with the distillation loss in Adversarial Diffusion Distillation (ADD) [61], which directly em" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.338, + 0.901 + ], + "angle": 0, + "content": "*Project Lead. † Corresponding Author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.539, + 0.887, + 0.828, + 0.901 + ], + "angle": 0, + "content": "\\(^{\\dagger}\\)Our DMDX (left to right): bottom, top, bottom, top." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "ploys the student model itself as score estimator. Theoretically, an intrinsic correspondence for DMD and the distillation loss in ADD can be equivalently found between Variational Score Distillation (VSD) [74] and Score Distillation Sampling (SDS) [51] in text-to-3D generation, where SDS represents a specialized instance of VSD through the use of a single-point Dirac distribution as the variational distribution [74]. And thus the VSD loss in DMD performs better than the SDS loss in ADD. Intuitively, since the capacity of a diffusion model decreases significantly when distilled for few-step generations [23], the student model no longer serves as a decent score estimator as the teacher. Therefore, the score-based distillation loss in ADD hardly contributes to its final performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.306, + 0.483, + 0.638 + ], + "angle": 0, + "content": "However, the optimization of DMD loss relies on the reverse Kullback-Leibler (KL) divergence minimization which is zero-forcing that drives low-probability regions to zero, causing the model to focus on only a few dominant modes and potentially leading to mode collapse [45]. To enhance sample diversity, DMD [86] employs an ODE-based (Ordinary Differential Equation) regularizer with synthetic data, while DMD2 [85] introduces a GAN-based (Generative Adversarial Network [8]) regularizer with real data to counterbalance this side effect. Subsequent efforts such as Moment Matching Distillation (MMD) [59], Score identity Distillation (SiD) [93] and Score Implicit Matching (SIM) [38] all used a variant of Fisher divergence to align the fake score estimator with pre-trained real score estimator [16]. Despite great success, their ability for distribution matching is limited by their dependence on a predefined form of explicit divergence metric. In this work, we would like to raise and investigate a question: Can we bypass the limitations of a predefined divergence by developing a framework that learns an implicit, data-driven discrepancy measure, thereby enabling more flexible and fine-grained matching of complex, high-dimensional distributions?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.867 + ], + "angle": 0, + "content": "Since the multifaceted alignment requirements in complex multimodal text-conditioned image or even video generation may not be fully captured, this motivates our exploration of more adaptive discrepancy learning paradigms. As our first contribution, we show how to align the latent predictions between real and fake score estimators to realize score distillation in an adversarial manner with prior knowledge of teacher model and dynamically learnable parameters, termed Adversarial Distribution Matching (ADM). This is different from the ODE-based or GAN-based regularizer in DMD and DMD2 that is used to counterbalance the mode collapse effect of reverse KL divergence. In contrast, we are performing distribution matching by means of GAN training to replace and circumvent the use of DMD loss, which we will provide more discussion in Sec. 4.1.3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our second contribution concerns about the extremely challenging one-step distillation, where we observe that the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "score distillation has a higher risk of gradient exploding and vanishing. We attribute the issue more to the lower overlap of support sets between the student and teacher distributions, not just the approximation errors in the fake score estimator as attributed in DMD2 [85]. In other words, while score distillation yields superior generation quality, it imposes higher requirements on initialization especially when distilled in extremely few steps, which we will provide more in-depth analysis in Sec. 4.3.2. Although we notice that an ODE-based pre-training on synthetic data is used for SDXL one-step distillation in DMD2 implementation, the mean squared error loss was probably still not enough to provide more overlapping regions of support sets between the student and teacher distributions. Our experiments demonstrate that when providing a better initialization for distribution matching, the effect of Two Time-scale Update Rule (TTUR) [2] to the final performance is very limited." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.905, + 0.514 + ], + "angle": 0, + "content": "Naturally, our third contribution focuses on providing a better initialization for further score-based fine-tuning. We employ adversarial distillation to trade off sample quality and mode converge inspired by SDXL-Lightning [23] and LADD [60]. With this distribution-level loss optimization, we can pre-train the student model to capture more potential modes of the teacher model distribution, especially through a hybrid discriminator in both latent and pixel spaces which we will introduce later in Sec. 4.2. To facilitate the diversity, we also propose employing a cubic timestep schedule for the generator to bias towards higher noise levels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.905, + 0.651 + ], + "angle": 0, + "content": "By combining adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our one-step SDXL provides competitive fidelity compared to the baseline with \\(50 \\times\\) acceleration in the Number of Function Evaluation (NFE) as shown in Fig. 1. More experiments on multi-step ADM distillation across the best existing diffusion models including SD3-Medium, SD3.5-Large [6], and CogVideoX [83] consistently achieve new benchmarks for efficient image and video synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.655, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.81 + ], + "angle": 0, + "content": "Progressive Distillation was proposed in [58] to distill multi-step prediction into one-step prediction of the same distance along the trajectory. SDXL-Lightning [23] extended this idea by utilizing GAN training and achieved one-step high-resolution (1024px) generation for the first time. This approach involving multi-stage process can be quite cumbersome, as it requires iteratively distilling from its predecessor, halving the sampling steps each time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Consistency Distillation proposed [7, 34, 35, 64, 66] to enforce the consistency property into diffusion model, i.e. predictions towards original sample are consistent for arbitrary pairs of noisy timesteps belong to the same trajectory. Subsequent efforts extended it into the trajectory consistency to relax the training objectives, i.e. predictions towards noisy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16819" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "samples of arbitrary subsequent timesteps are consistent, including CTM [18], TCD [92], TSCD [55] and PCM [69]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.242 + ], + "angle": 0, + "content": "Rectified Flow [27, 28] aims to obtain faster straight trajectories through multiple reflow processes that iteratively learn the velocity of many ODE pairs from its predecessor. PeRFlow [81] tried splitting the trajectory into pieces and applying piece-wise rectification with real data samples. In this paper, we empirically found the straightness can also be satisfied via an adversarial distillation paradigm without the need for repeatedly collecting tremendous synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Score Distillation intuitively tries to keep that a sample appearing in the student model distribution at a specific noise level is with the same probability as it does in the teacher model distribution. In light of the motivation that relying on a single divergence might be problematic, a concurrent work Score-of-Mixture Distillation (SMD) [16] shares our perspective and explicitly designed a class of \\(\\alpha\\)-skew Jensen-Shannon (JS) divergences for optimization. In contrast, we implicitly measure the discrepancy between the fake and real score estimators in an adversarial manner, facilitating a more capable and adaptive discrepancy learning paradigm along the distillation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.559 + ], + "angle": 0, + "content": "Adversarial Distillation employs a discriminator to align the student model distribution with a specific target distribution [40, 94]. While methods UFOGen [79], DMD2 [85], and APT [24] directly align with the real data distribution, SDXL-Lightning [23] and Hyper-SD [55] utilize intermediate timestep predictions from the teacher model as approximation objectives. Inspired by LADD [60], our adversarial distillation pre-training aligns with ODE-based synthetic data generated from the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Human Feedback Learning was firstly considered by ReFL [78] in diffusion models that includes two stages of reward model training and preference fine-tuning. HyperSD [55] treated this as a standalone technique, ultimately applying LoRA insertion to shift the output distribution of the low-step generator toward human preferences. Subsequent studies, notably [36, 39], further attempted to integrate CFG and Score-based divergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.229, + 0.708 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.253, + 0.733 + ], + "angle": 0, + "content": "3.1. Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.801 + ], + "angle": 0, + "content": "Given a training data distribution \\( p_{\\mathrm{data}} \\) with standard deviation \\( \\sigma_{\\mathrm{data}} \\), the diffusion model [12] generates samples by reversing the forward diffusion process that progressively adds noise to a data sample \\( x_0 \\sim p_{\\mathrm{data}} \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.805, + 0.482, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x}\\right) \\sim \\mathcal {N} \\left(\\alpha_ {t} \\boldsymbol {x}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.885 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\geq 0, \\sigma_{t} > 0\\) are specified noise schedules such that \\(\\alpha_{t} / \\sigma_{t}\\) satisfies monotonically decreasing w.r.t. \\(t\\) and larger \\(t\\) indicates greater noise. We consider two different formulations for denoising models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "DDPM and DDIM [12, 63] assume discrete-time schedules" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "with \\( t \\in [1, T] \\) (typically \\( T = 1000 \\)) and noise prediction parameterization* [58]. The training objective is given by," + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.126, + 0.905, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.147, + 0.907, + 0.263 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function and \\( \\epsilon_{\\theta} \\) is a neural network with parameters \\( \\theta \\). The noise schedule is defined as \\( \\alpha_{t} = \\sqrt{\\bar{\\alpha}_{t}}, \\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}} \\) such that \\( \\boldsymbol{x}_t = \\sqrt{\\bar{\\alpha}_t} \\boldsymbol{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon \\). For DDIM sampling, it solves the Probability-Flow Ordinary Differential Equations (PF-ODE) [65] by \\( d\\bar{\\boldsymbol{x}}_t = \\epsilon_\\theta (\\frac{\\bar{\\boldsymbol{x}}_t}{\\sqrt{\\bar{\\sigma}_t^2 + 1}}) d\\bar{\\sigma}_t \\), where \\( \\bar{\\boldsymbol{x}}_t = \\frac{\\boldsymbol{x}_t}{\\sqrt{\\bar{\\alpha}_t}} \\) and \\( \\bar{\\sigma}_t = \\sqrt{\\frac{1 - \\bar{\\alpha}_t}{\\bar{\\alpha}_t}} \\), starting from \\( \\boldsymbol{x}_T \\sim \\mathcal{N}(\\boldsymbol{0}, \\boldsymbol{I}) \\) and stopping at \\( \\boldsymbol{x}_0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.264, + 0.906, + 0.341 + ], + "angle": 0, + "content": "Flow Matching [26, 27, 30] uses velocity prediction parameterization [58] and continuous-time coefficients (typically \\(\\alpha_{t} = 1 - t\\), \\(\\sigma_{t} = t\\) with \\(t \\in [0, T = 1]\\)). The conditional probability path or the velocity we say can be given as \\(\\pmb{v}_{t} = \\frac{d\\alpha_{t}}{dt}\\pmb{x}_{0} + \\frac{d\\sigma_{t}}{dt}\\epsilon\\), such that the training objective is," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.344, + 0.905, + 0.362 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\boldsymbol {v} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {v} _ {t} \\| _ {2} ^ {2} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.906, + 0.426 + ], + "angle": 0, + "content": "where \\( w(t) \\) is a weighting function and \\( \\pmb{v}_{\\theta} \\) is a neural network parameterized by \\( \\theta \\). The sampling procedure starts from \\( t = T \\) with \\( \\pmb{x}_T \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I}) \\) and stops at \\( t = 0 \\), solving the PF-ODE by \\( d\\pmb{x}_t = \\pmb{v}_{\\theta}(\\pmb{x}_t, t)dt \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.812, + 0.453 + ], + "angle": 0, + "content": "3.2. Distribution Matching Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.535 + ], + "angle": 0, + "content": "DMD [85, 86] distills pretrained diffusion models \\( \\pmb{F}_{\\phi}(\\pmb{x}_t,t) \\) into one-step or multi-step efficient generators \\( G_{\\theta}(\\pmb{x}_t,t) \\) by minimizing the reverse KL divergence between the target distribution \\( p_{\\mathrm{real}} \\) and the efficient generator output distribution \\( p_{\\mathrm{fake}} \\). The gradient of DMD objective w.r.t. \\( \\theta \\) is," + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.539, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D M D}} = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} - \\left[ \\left(s _ {\\text {r e a l}} \\left(\\boldsymbol {x} _ {t}\\right) - s _ {\\text {f a k e}} \\left(\\boldsymbol {x} _ {t}\\right)\\right) \\frac {d \\boldsymbol {G} _ {\\theta} \\left(\\boldsymbol {z} , t ^ {\\prime}\\right)}{d \\theta} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.675 + ], + "angle": 0, + "content": "where \\(z \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(t'\\) is randomly selected from predefined generator schedule, \\(t \\sim \\mathcal{U}(0,T)\\), and noisy samples \\(\\boldsymbol{x}_t = \\boldsymbol{q}(\\boldsymbol{x}_t|\\hat{\\boldsymbol{x}}_0)\\) are obtained by randomly diffusing the generator output \\(\\hat{\\boldsymbol{x}}_0 = \\boldsymbol{G}_{\\theta}(\\boldsymbol{z},t')\\). Score functions \\(s_{\\mathrm{real}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{real}}(\\boldsymbol{x}_t)\\), \\(s_{\\mathrm{fake}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{fake}}(\\boldsymbol{x}_t)\\) are vector fields that point towards higher density of data at a given noise level [17, 65] for \\(p_{\\mathrm{real}}\\) and \\(p_{\\mathrm{fake}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.676, + 0.906, + 0.751 + ], + "angle": 0, + "content": "While the real score estimator is the teacher model \\( \\mathbf{F}_{\\phi}(\\mathbf{x}_t,t) \\) itself, the fake score estimator \\( \\pmb{f}_{\\psi}(\\pmb{x}_t,t) \\) is initialized the same as \\( \\pmb{F}_{\\phi}(\\pmb{x}_t,t) \\) and dynamically learned to describe \\( p_{\\mathrm{fake}} \\) with pretrain loss as Eqs. (2) and (3). In practice, the gradient in Eq. (4) is computed as," + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.754, + 0.905, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\operatorname {g r a d} \\left(\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t\\right) = \\frac {\\boldsymbol {f} _ {\\psi} \\left(\\boldsymbol {x} _ {t} , t\\right) - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\left\\| \\hat {\\boldsymbol {x}} _ {0} - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right) \\right\\| _ {1}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.79, + 0.822, + 0.805 + ], + "angle": 0, + "content": "such that the training loss is implemented like," + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.809, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D M D}} (\\theta) = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} [ \\| \\hat {\\boldsymbol {x}} _ {0} - s g (\\hat {\\boldsymbol {x}} _ {0} - \\operatorname {g r a d} (\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t)) \\| _ {2} ^ {2} ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.838, + 0.832, + 0.853 + ], + "angle": 0, + "content": "where \\( sg(\\cdot) \\) denotes the stop gradient operation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.906, + 0.901 + ], + "angle": 0, + "content": "*This is not mandatory for works like CogVideoX [83] which also used DDIM sampling but parameterized in velocity. The different parameterizations can be equivalently transformed into each other [58]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16820" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.095, + 0.514, + 0.106 + ], + "angle": 0, + "content": "ADM" + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.108, + 0.845, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.233, + 0.512, + 0.244 + ], + "angle": 0, + "content": "ADP" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.245, + 0.846, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.446, + 0.896, + 0.461 + ], + "angle": 0, + "content": "Figure 2. Overall pipeline of our proposed Adversarial Distribution Matching (ADM) and Adversarial Distillation Pre-training (ADP)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.471, + 0.226, + 0.489 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.499, + 0.395, + 0.515 + ], + "angle": 0, + "content": "4.1. Adversarial Distribution Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.69 + ], + "angle": 0, + "content": "Instead of using a predefined divergence between the fake and real distributions, we use an implicit, data-driven discrepancy measure through an adversarial discriminator. Specifically, our discriminator \\( D_{\\tau}(\\pmb{x}_t,t) \\) consists of a frozen latent diffusion model initialized the same as teacher model \\( F_{\\phi}(\\pmb{x}_t,t) \\) and multiple trainable heads added upon different UNet [57] or DiT [49] blocks. Given the noisy sample \\( \\pmb{x}_t = \\pmb{q}(\\pmb{x}_t|\\hat{\\pmb{x}}_0) \\) diffused from the output of few-step generator \\( \\hat{\\pmb{x}}_0 = \\pmb{G}_{\\theta}(\\pmb{z},t') \\), the score estimators no longer solve the PF-ODE w.r.t. the end point \\( \\pmb{x}_0^{\\mathrm{fake}} = \\pmb{f}_{\\psi}(\\pmb{x}_t,t) \\) and \\( \\pmb{x}_0^{\\mathrm{real}} = \\pmb{F}_{\\phi}(\\pmb{x}_t,t) \\) as used in Eq. (5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Instead, we set a fixed timestep interval \\(\\Delta t\\) (defaults to \\(T / 64\\)) and solve the PF-ODE w.r.t. \\((t - \\Delta t)\\), such that the fake sample \\(\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}\\) and real sample \\(\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}\\) can be obtained to serve as score predictions and sent into the discriminator. The discriminator hierarchically aggregates features from frozen backbone layers and dynamically weights them through multiple learnable heads, establishing an adaptive discrepancy metric that leverages both diffusion priors and data-driven trainable dynamics. We use Hinge loss [22] to train the generator \\(\\pmb{G}_{\\theta}(\\pmb{x}_t,t)\\) and discriminator \\(D_{\\tau}(\\pmb{x}_t,t)\\) alternatively. This encourages the fake score prediction \\(\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}\\) to be closer to the real score prediction \\(\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.878, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}} {\\mathbb {E}} \\left[ - \\boldsymbol {D} _ {\\tau} \\left(\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t\\right) \\right] \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.469, + 0.906, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} (\\tau) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, \\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {r e a l}}} {\\mathbb {E}} [ \\max (0, 1 + D _ {\\tau} (\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t)) \\tag {8} \\\\ \\left. + \\max \\left(0, 1 - D _ {\\tau} \\left(x _ {t - \\Delta t} ^ {\\text {r e a l}}, t - \\Delta t\\right)\\right) \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.568 + ], + "angle": 0, + "content": "In conjunction with the dynamically learned fake model, we clarify our training procedure in Appendix A. The overall pipeline is demonstrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.885, + 0.594 + ], + "angle": 0, + "content": "4.1.1. Motivation of Discriminator Timestep \\((t - \\Delta t)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Given that the final objective of score distillation is to match the probability flows varying with noise levels of student and teacher models to exactly the same, timestep information must be considered when measuring the discrepancy between distributions. This coincides with our discriminator design that uses a pre-trained diffusion model, and we take a small step alongside the PF-ODE, succeeding in preserving the input timestep information of score estimators." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.691, + 0.744 + ], + "angle": 0, + "content": "4.1.2. Data-driven Effect" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The flexibility of using discriminators for distributional discrepancy measure is not only in the noise level of score function, but also within the distillation process. As distillation iterates, the model is exposed to increasingly diverse data, causing the discrepancies in modes between the two distributions to change. In the early phase of training where the discrepancy is significant, a more global evaluation is necessary, whereas in the later phases, when the discrepancy becomes minor, a more localized, fine-grained optimization might become essential. In other words, driven" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16821" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.48, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.192, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Figure 3. Changes of DMD loss over multi-step ADM distillation for CogVideoX. Note that we did not optimize this objective directly during ADM distillation but recorded it over iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.483, + 0.272 + ], + "angle": 0, + "content": "by the volume of data, the divergence measure employed across different training phases can vary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.386, + 0.293 + ], + "angle": 0, + "content": "4.1.3. Relationship with DMD and DMD2" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.492 + ], + "angle": 0, + "content": "To mitigate the mode collapse issue in DMD loss, an ODE-based regularizer and a GAN-based regularizer are additionally used for distillation in DMD [85] and DMD2 [85], respectively. However, these two regularizers do not fundamentally address the mode-seeking behavior introduced by reverse KL divergence as shown in Fig. 4(a), but rather counterbalance it by trade-offs between losses. In ADM, our adversarial loss is actually playing the role of DMD loss to realize score distillation with an implicit, data-driven discrepancy measure instead of a predefined divergence. Therefore, our motivation for using GAN training in ADM is different from that of DMD2 [85] and we don't require additional regularizers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Intuitively, the learnable discriminator can approximate any nonlinear function to implicitly measure distribution divergence, which probably inherently encompasses the reverse KL divergence in DMD loss. As shown in Fig. 3, we visualize the changes of DMD loss in Eq. (6) during the multi-step ADM distillation on CogVideoX [83]. Though not directly optimize on Eq. (6), the results indicate a very steady downward trend that supports our assumption. More theoretical discussion is provided in Sec. 4.3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.407, + 0.654 + ], + "angle": 0, + "content": "4.2. Adversarial Distillation Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.796 + ], + "angle": 0, + "content": "To stabilize the extremely difficult one-step distillation, we opt to provide a better initialization for ADM fine-tuning with adversarial distillation pre-training on synthetic data. Our pre-training configuration refers to Rectified Flow [27] in several aspects, where we 1) collect the ODE pairs from teacher model in an offline manner, 2) construct noisy samples by linearly interpolation between pure noise and clean data sample of the ODE pair, 3) alter the prediction target of the generator to the velocity of ODE pair." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As for adversarial training, we formulate a latent-space discriminator \\(D_{\\tau_1}(\\boldsymbol{x}_t,t)\\) initialized from teacher model and a pixel-space discriminator \\(D_{\\tau_2}(\\boldsymbol{x})\\) initialized from the vision encoder of SAM [19] model, respectively, as shown in Fig. 2. We also append multiple trainable heads to both the backbone networks similar to the practice in ADM. All these contribute to increasing the discriminative capability," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.288 + ], + "angle": 0, + "content": "facilitating the student model in discovering more potential modes in the teacher model distribution. Specifically, let \\(\\tilde{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)\\) denote the predicted PF-ODE endpoint of generator, where \\(\\pmb{x}_{t}\\) represents the noisy sample interpolated between an ODE pair \\((\\pmb {x}_T,\\pmb {x}_0)\\) with random timestep \\(t\\in [0,T]\\). For latent-space discriminator, we diffuse the generator output with another random noise and timestep \\(t^{\\prime}\\in (0,T]\\), getting \\(\\tilde{\\pmb{x}}_{t^{\\prime}} = \\pmb {q}(\\tilde{\\pmb{x}}_{t^{\\prime}}|\\tilde{\\pmb{x}}_0)\\) as its input. For pixel-space discriminator, the generator output will be first decoded via VAE decoder and then sent into the vision encoder. The training objective, also based on Hinge loss [22], encourages the generator output \\(\\tilde{\\pmb{x}}_0\\) to be closer to the synthetic data sample \\(\\pmb{x}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.291, + 0.907, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} - \\left[ \\lambda_ {1} D _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) + \\lambda_ {2} D _ {\\tau_ {2}} \\left(\\tilde {\\boldsymbol {x}} _ {0}\\right) \\right] \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.324, + 0.905, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} \\left(\\tau_ {1}, \\tau_ {2}\\right) = \\underset {\\boldsymbol {x} _ {0}, \\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} \\left[ \\lambda_ {1} \\cdot \\max \\left(0, 1 + \\boldsymbol {D} _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right. \\\\ + \\lambda_ {2} \\cdot \\max \\left(0, 1 + D _ {\\tau_ {2}} \\left(\\tilde {x} _ {0}\\right)\\right) \\tag {10} \\\\ + \\lambda_ {1} \\cdot \\max (0, 1 - D _ {\\tau_ {1}} \\left(\\boldsymbol {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)) \\\\ \\left. \\left. + \\lambda_ {2} \\cdot \\max \\left(0, 1 - D _ {\\tau_ {2}} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right] \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.408, + 0.907, + 0.438 + ], + "angle": 0, + "content": "We empirically found that setting balancing coefficients \\(\\lambda_{1} = 0.85, \\lambda_{2} = 0.15\\) produces visually coherent results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.813, + 0.47 + ], + "angle": 0, + "content": "4.2.1. Cubic Generator Timestep Schedule" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Intuitively, higher noise levels encourage exploration of new modes by weakening the restrictive information encoded in latent representations. Therefore, we propose employing a cubic timestep schedule for the generator. The schedule maps uniform \\([0,T)\\) samples through \\([1 - (t / T)^3 ]*T\\), non-linearly concentrating values near \\(T\\) with heavy noise similar to LADD [60]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.856, + 0.615 + ], + "angle": 0, + "content": "4.2.2. Uniform Discriminator Timestep Schedule" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.743 + ], + "angle": 0, + "content": "Unlike in ADM where the discriminator is firstly used for score distillation, the utilization of latent-space discriminator is common for adversarial distillation works. Inspired by SDXL-Lightning [23], where they found that the diffusion encoder is trained to focus on high-frequency details at lower timesteps and low-frequency structures at higher timesteps, we set a uniform \\((0,T]\\) for discriminator timestep \\(t'\\) to capture both advantages during pre-training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.734, + 0.773 + ], + "angle": 0, + "content": "4.2.3. Relationship with LADD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our motivation to perform adversarial distillation on synthetic data is inspired by LADD [60], but differs in a lot that we 1) construct noisy samples via ODE pairs in the style of Rectified Flow [27] rather than random noise, 2) develop a cubic generator timestep schedule that facilitate deterministic Euler sampling rather than consistency ones, 3) introduce an additional pixel-space encoder to increase the capability of discriminator and find more modes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.209, + 0.105 + ], + "angle": 0, + "content": "4.3. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.114, + 0.383, + 0.127 + ], + "angle": 0, + "content": "4.3.1. Difference between ADM and ADP" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.133, + 0.483, + 0.344 + ], + "angle": 0, + "content": "One question may appear like what is the difference between these two adversarial approaches with latent-space discriminators? The effectiveness in score distillation correlates with the fact that score function \\(\\nabla_{\\pmb{x}}\\log p(\\pmb{x};\\sigma (t))\\) is defined at different noise levels \\(\\sigma (t)\\). In contrast, adversarial distillation only aligns the distribution of clean data samples when \\(t = 0\\). While the latent-space discriminator in pre-training is capturing information at different scales and details by randomly diffusing the generator output, the crux for ADM is more than that by solving the PF-ODE of both score estimators. In other words, ADM additionally supervises the complete denoising process through noisy samples that are with higher density at different noise levels of respective distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.526 + ], + "angle": 0, + "content": "This leads to a situation in ADM when two distributions initialized with less overlap in support sets, the noisy samples remain in regions unfamiliar to each other, and the discriminator can easily distinguish between them thus leading to extreme gradient signals. However, since Gaussian noise is isotropic, we artificially create overlapping regions for the randomly diffused samples in ADP to make the discrimination more difficult, resulting in relatively smooth gradients. Therefore, our ADM still falls under score distillation given that it encourages the entire probability flows to be closer, while the pre-training belongs to adversarial distillation because it only cares about the clean data distribution at \\( t = 0 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.327, + 0.55 + ], + "angle": 0, + "content": "4.3.2. Importance of Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.6 + ], + "angle": 0, + "content": "Another question we haven't discussed so far is why do we need pre-training for one-step score distillation? Taking the reverse KL divergence used by DMD loss as an example:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.603, + 0.483, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathbb {D} _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\int p _ {\\text {f a k e}} (\\boldsymbol {x}) \\log \\frac {p _ {\\text {f a k e}} (\\boldsymbol {x})}{p _ {\\text {r e a l}} (\\boldsymbol {x})} d \\boldsymbol {x}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.886 + ], + "angle": 0, + "content": "When employing one-step distillation, the generator output is worse in visual fidelity and structural integrity compared to multi-step sampling, leading \\( p_{\\mathrm{fake}}(\\pmb{x}) \\to 0 \\) where \\( p_{\\mathrm{real}}(\\pmb{x}) > 0 \\) in regions. The integrand \\( p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{p_{\\mathrm{real}}(\\pmb{x})} \\) approaching zero \\( 0 \\cdot (-\\infty) \\) causes the optimization to avoid regions where \\( p_{\\mathrm{real}}(\\pmb{x}) > 0 \\) but \\( p_{\\mathrm{fake}}(\\pmb{x}) \\) has negligible density, a phenomenon called zero-forcing. Instead of fully covering \\( p_{\\mathrm{real}} \\)'s support, \\( p_{\\mathrm{fake}} \\) collapses to a subset of modes of \\( p_{\\mathrm{real}} \\), inducing mode-seeking behavior as illustrated in Fig. 4(a). During training this manifests itself as gradient vanishing sometimes. And conversely, the fuzzy samples that the diffusion model typically produces through one step are also not within the teacher model distribution, yielding \\( p_{\\mathrm{real}}(\\pmb{x}) \\to 0 \\) where \\( p_{\\mathrm{fake}}(\\pmb{x}) > 0 \\) in regions and the integrand \\( p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{0} \\) diverging to \\( +\\infty \\), and thus resulting in numerical instability and gradient exploding." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Similarly, when the support sets of the student and" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.091, + 0.907, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.362, + 0.851, + 0.375 + ], + "angle": 0, + "content": "Figure 4. Illustration for theoretical discussion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.5 + ], + "angle": 0, + "content": "teacher distributions overlap hardly at all, forward KL divergence approaches \\(+\\infty\\) where \\(p_{\\mathrm{fake}}(\\boldsymbol{x}) > 0\\), while JS divergence saturates to a constant log 2 and Fisher divergence may degenerate without definition. Therefore, when this assumption is undermined, many of the single divergences no longer apply and a better initialization with more overlapping regions becomes essential as exemplified in Fig. 4(b)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.712, + 0.527 + ], + "angle": 0, + "content": "4.3.3. Theoretical Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.578 + ], + "angle": 0, + "content": "The last question is why ADM is better than DMD loss theoretically? In fact, the Hinge GAN [22] we use has been proven to minimize Total-Variation Distance (TVD) [68]:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.581, + 0.906, + 0.613 + ], + "angle": 0, + "content": "\\[\nT V \\left(p _ {\\text {f a k e}}, p _ {\\text {r e a l}}\\right) = \\int \\left| p _ {\\text {f a k e}} (\\boldsymbol {x}) - p _ {\\text {r e a l}} (\\boldsymbol {x}) \\right| d \\boldsymbol {x} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "That is to say, when the discriminator is sufficiently rich and well trained, the theoretical optimum of Hinge loss upon convergence minimizes TVD. When the support sets of fake and real distributions have minimal overlap, TVD provides two key advantages over reverse KL divergence: 1) Symmetry: TVD yields the same discrepancy measure regardless of the initial distribution, whereas the asymmetric reverse KL may exhibit mode-seeking behavior and neglect other portions of the overall distribution. For example, as illustrated in Fig. 4(c), TVD maintains substantial loss values and provides optimization directions covering the mode when \\( p_{\\mathrm{fake}}(\\pmb{x}) \\to 0 \\) while \\( p_{\\mathrm{real}}(\\pmb{x}) > 0 \\), whereas the reverse KL divergence suffers from gradient vanishing issues in such scenarios, as discussed in Sec. 4.3.2. 2) Boundedness: TVD is bounded within [0,1], and thus mitigating disruptions from outliers during training, especially in our high-dimensional multimodal text-conditioned image and video distributions, avoiding the numerical instability inherent to reverse KL divergence caused by gradient exploding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16823" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.48, + 0.204 + ], + "angle": 0, + "content": "
MethodStep NFECLIP ScorePick ScoreHPSv2MPS
ADD [61] (512px)1135.008822.152427.097110.4340
LCM [34]1228.466920.126723.82464.8134
Lightning [23]1133.498521.919427.155710.2285
DMD2 [85]1135.215322.097827.452310.6947
DMDX (Ours)1135.255722.273627.704611.1978
SDXL-Base [56]255035.030922.249427.374310.7042
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.208, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Table 1. Quantitative results on fully fine-tuning SDXL-Base." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.23, + 0.48, + 0.399 + ], + "angle": 0, + "content": "
MethodStepNFECLIP ScorePick ScoreHPSv2MPS
TSCD [61]4834.018521.966527.272810.8600
PCM [69] (Shift=1)4433.504221.970327.368010.5707
PCM [69] (Shift=3)4433.381821.939627.114610.5635
PCM [69] (Stoch.)4433.418521.882227.317710.5200
Flash [3]4434.397822.090427.258610.6634
ADM (Ours)4434.907622.547128.449211.9543
SD3-Medium [6]255034.763322.296127.973311.3652
LADD [60]4434.739522.395827.492311.4372
ADM (Ours)4434.973022.884227.733112.2350
SD3.5-Large [6]255034.966822.508727.968811.5826
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.403, + 0.483, + 0.432 + ], + "angle": 0, + "content": "Table 2. Quantitative results on LoRA fine-tuning SD3-Medium and fully fine-tuning SD3.5-Large." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.439, + 0.483, + 0.584 + ], + "angle": 0, + "content": "
MethodStepNFEFinal ScoreQuality ScoreSemantic Score
ADM8878.58480.82569.621
+ Longer Training ×28880.76483.03171.693
ADM w/ CFG81679.86580.93875.569
+ Longer Training ×281681.79683.00876.947
CogVideoX-2b [83]10020080.03680.80176.974
ADM8882.06783.22777.423
ADM w/ CFG81680.98282.16576.251
CogVideoX-5b [83]10020081.22681.78578.987
" + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.589, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Table 3. Quantitative results on fully fine-tuning CogVideoX." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.615, + 0.224, + 0.632 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Models. For one-step distillation, we employ Adversarial Distillation Pre-training (ADP) and ADM fine-tuning on SDXL-Base [56], termed DMDX. For multi-step distillation, we only employ ADM training on both text-to-image models SD3-Medium, SD3.5-Large [6], and text-to-video models CogVideoX-2b, CogVideoX-5b [83]. Following most concurrent works, we didn't use Classifier-Free Guidance (CFG) [11] in our text-to-image model. We tried conducting CFG-integrated experiments on the text-to-video model, with the approach detailed in Sec. 5.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Datasets. No visual data is required for both ADP and ADM proposed in this work. For image generators, we utilize the text prompts from JourneyDB [67] that exhibits high-level of detail and specificity for training. For video generators, we collect training prompts from OpenVid-1M [47], Vript [82] and Open-Sora-Plan-v1.1.0 [50]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation. The image generators are evaluated on 10K" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.453, + 0.892, + 0.468 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on fully fine-tuning SDXL-Base." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.668 + ], + "angle": 0, + "content": "prompts from COCO 2014 [25] following DMD2 [85]. We report the CLIP score [52] and human preference benchmarks PickScore [20], HPSv2 [75] and MPS [90] as many concurrent works including Hyper-SD [55] and Emu3 [73]. But we don't include Hyper-SD in the one-step quantitative comparison because one-step Hyper-SDXL has been optimized directly on human feedback with ReFL [78]. Instead, we compare with the TSCD algorithm proposed therein on SD3-Medium [13], since 4-step Hyper-SD3 LoRA is without ReFL optimization. The video generators are evaluated on VBenchmark [14], which consists of multiple quality and semantic dimensions comprehensively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Hyperparameters. Despite having multiple models to train in our proposed ADP and ADM, we achieve satisfactory visual fidelity and structural integrity without extensive hyperparameter tuning. For the remainder of this work, we only adjust the learning rate of generator for different experiments. The optimizer settings for discriminators and fake models are uniform across all the experiments. Unless longer training is specifically noted, we train the same for only 8K iterations with a batch size of 128 and 8 for text-to-image and text-to-video models, respectively. More specific implementation details are provided in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.744, + 0.864 + ], + "angle": 0, + "content": "5.1. Efficient Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Tab. 1 quantitatively compares our two-stage approach combing ADP and single-step ADM distillation with ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16824" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.089, + 0.472, + 0.27 + ], + "angle": 0, + "content": "
AblationCLIP ScorePick ScoreHPSv2MPS
Ablation on adversarial distillation.
A1: Rectified Flow [27]27.437620.021123.60934.4518
A2: DINOv2 as pixel-space34.183621.875027.103910.2407
A3: λ1 = 0.7, λ2 = 0.333.694321.634426.89029.9633
A4: λ1 = 1.0, λ2 = 0.033.892921.739526.786910.0757
A5: w/o ADM (ADP only)35.772322.009527.349910.6646
Ablation on score distillation.
B1: ADM w/o ADP32.502021.763126.873210.8986
B2: DMD Loss w/o ADP32.748221.034125.96808.8977
B3: DMD Loss w/ ADP34.511921.936627.398510.6046
B4: DMDX (Ours)35.255722.273627.704611.1978
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.274, + 0.43, + 0.287 + ], + "angle": 0, + "content": "Table 4. Quantitative results on ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.401 + ], + "angle": 0, + "content": "isting one-step distillation methods on fully fine-tuning SDXL-Base. The results show that our method achieves excellent performance on both image-text alignment and human preference, which is consistent with the qualitative comparisons in Fig. 5, including better portrait aesthetic, animal hair details, subject-background separation and physical structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.479 + ], + "angle": 0, + "content": "For multi-step ADM distillation, it can serve as a standalone score distillation method. We tried both fully finetuning and LoRA fine-tuning [13] configurations and the quantitative results in Tab. 2 demonstrate our superior performance. Qualitative results are provided in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.486, + 0.318, + 0.501 + ], + "angle": 0, + "content": "5.2. Efficient Video Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.795 + ], + "angle": 0, + "content": "As quantitatively shown in Tab. 3, except for the regular 8-step ADM distillation for both sizes of CogVideoX, we also try integrating Classifier-Free Guidance (CFG) [11] for text-to-video task. Specifically, we assign the CFG scale for real model by randomly sampling within the range [5.0, 7.0], while assigning the few-step generator's scale through explicit subtraction of 2.0 from the real model's value. Empirically, we found that this subtractive offset requires progressive enlargement as the targeted sampling steps decreases. There is no CFG for the fake model. The VBench [14] results demonstrate that our few-step generators achieve comparable performance to the base model with \\(92 - 96\\%\\) acceleration. We conduct additional evaluations on the longer training of 2B model, motivated by the observation that DMD loss did't converge well at 8K iterations as Fig. 3 indicates. Experimental results demonstrate that the DMD loss is also approximately optimized by the learnable discriminator during ADM distillation. More quantitative and qualitative results refer to Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.255, + 0.818 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "In Tab. 4, we conduct extensive ablation studies on fully fine-tuning SDXL-Base to validate our effectiveness. Qualitative comparisons are provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of ADP. The following conclusions can be drawn: 1) Single reflow process provides very limited effectiveness" + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.09, + 0.882, + 0.168 + ], + "angle": 0, + "content": "
TTURTraining TimeCLIP ScorePick ScoreHPSv2MPS
1×1.0035.255722.273627.704611.1978
4×1.8535.258322.277327.725511.2720
8×2.5335.329922.288327.758611.2838
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.172, + 0.871, + 0.186 + ], + "angle": 0, + "content": "Table 5. Ablation study on two time-scale update rule." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.195, + 0.899, + 0.239 + ], + "angle": 0, + "content": "
ADDLCMLightningDMD2OursTeacher
LPIPS↑0.60710.62570.67070.67150.71560.6936
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.241, + 0.898, + 0.256 + ], + "angle": 0, + "content": "Table 6. Quantitative diversity evaluation on PartiPrompts [88]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.418 + ], + "angle": 0, + "content": "(A1), while multiple processes achieving effective rectification requires considerable computational expense. 2) Using SAM [19] offers more imaging fidelity compared to the widely adopted DINOv2 [48] (A2/A5), which is likely due to SAM's higher resolution of 1024px (compared to DINOv2's 518px). 3) The weighting \\(\\lambda_{2}\\) for pixel space should neither be too large nor too small. Excessive weighting leads to degraded structural integrity (A3/A5), while insufficient weighting results in blurriness (A4/A5). Qualitative results in Appendix C provide more visible distinction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.905, + 0.54 + ], + "angle": 0, + "content": "Effect of ADM. We summarize the key findings: 1) The absence of ADP exerts substantial performance degradation (B1/B4), which is aligned with our analysis in Sec. 4.3.2. 2) Without a regularizer, the DMD loss underperforms compared to standalone ADM (B1/B2), indicating its poor robustness. 3) Although DMD loss optimization also benefits from ADP (B2/B3), its distribution matching capability remains inferior to ADM (B3/B4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Effect of TTUR. Tab. 5 presents the impact of different TTUR settings on final performance and training duration. The results demonstrate that increasing TTUR yields only marginal performance gains while nearly doubling training time, rendering this trade-off clearly unwarranted. This highlights the critical role of our proposed ADP in one-step distillation and suggests that the training instability in DMD2 likely stems from insufficient support set overlap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.905, + 0.754 + ], + "angle": 0, + "content": "Diversity evaluation. Following DMD2 [85], we generate 4 samples per prompt on Partiprompts [88] with different seeds, and report the averaged pairwise LPIPS similarities [89] in Tab. 6. The results indicate that our method significantly outperforms others in diversity. More randomly curated multi-seed samples can be found in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.636, + 0.785 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We realize that one weakness is that teacher model might require CFG to produce accurate score prediction. Our experiments suggest this is a common characteristic of score distillation methods generally, not a limitation unique to our approach. This limits the application to guidance-distilled models such as FLUX.1-dev [1], which could be a promising research topic for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16825" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (U22A2095, 12326618, 62276281), Guangdong Basic and Applied Basic Research Foundation, China (2024A1515011882) and the Project of Guangdong Provincial Key Laboratory of Information Security Technology (2023B1212060026)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[1] Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[2] Naresh Babu Bynagari. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.331, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[3] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 1, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[4] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.433, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.517, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.615, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[7] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.484, + 0.712 + ], + "angle": 0, + "content": "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.714, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[14] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 7, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[15] Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models. arXiv preprint arXiv:2309.14509, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Tejas Jayashankar, J. Jon Ryu, and Gregory Wornell. Score-of-mixture training: Training one-step generative models made simple via score estimation of mixture distributions. arXiv preprint arXiv:2502.09609, 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[18] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.481 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In ICCV, pages 3992-4003, 2023. 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[20] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeuriPS, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.606 + ], + "angle": 0, + "content": "[21] Jonas Kohler, Albert Pumarola, Edgar Schonfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Thabet. Imagine flash: Accelerating emu diffusion models with backward distillation. arXiv preprint arXiv:2405.05224, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[23] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[24] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, Lu Jiang, and ByteDance Seed. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[26] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[27] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2022. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16826" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "diffusion-based text-to-image generation. In ICLR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.482, + 0.149 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[30] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[31] Yanzuo Lu, Manlin Zhang, Yiqi Lin, Andy J. Ma, Xiaohua Xie, and Jianhuang Lai. Improving pre-trained masked autoencoder via locality enhancement for person re-identification. In PRCV, pages 509–521, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[32] Yanzuo Lu, Meng Shen, Andy J Ma, Xiaohua Xie, and JianHuang Lai. Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation. In AAAI, pages 3900-3908, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[33] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In CVPR, pages 6420-6429, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[34] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[35] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[36] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. In TMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[37] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, pages 76525-76546, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[38] Weijian Luo, Zemin Huang, Zhengyang Geng, J. Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching. In NeurIPS, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[39] Weijian Luo, Colin Zhang, Debing Zhang, and Zhengyang Geng. David and goliath: Small one-step model beats large diffusion with score post-training. In ICML, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[40] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[41] Hongxu Ma, Guanshuo Wang, Fufu Yu, Qiong Jia, and Shouhong Ding. Ms-detr: Towards effective video moment retrieval and highlight detection by joint motion-semantic learning. In ACMMM, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[42] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Ji-hong Guan, and Shuiqeng Zhou. Fine-grained zero-shot object detection. In ACMMM, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[43] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. arXiv preprint arXiv:2409.11367, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In CVPR, pages 21203-21214, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.904, + 0.163 + ], + "angle": 0, + "content": "[45] Thomas Minka. Divergence measures and message passing. Microsoft Research, Technical Report, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[46] Movie Gen Team. Movie gen: A cast of media foundation models. https://ai.meta.com/static-resource/movie-gen-research-paper, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[47] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 8, 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[50] PKU-Yuan Lab and Tuzhan AI. Open-sora-plan-v1.1.0. https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[51] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[53] Samyam Rajbhandari, Jeff Rasley, Olatunj Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[54] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[55] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. In NeurIPS, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[58] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[59] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. In NeurIPS, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[62] Meng Shen, Yanzuo Lu, Yanxu Hu, and Andy J. Ma. Collaborative learning of diverse experts for source-free universal domain adaptation. In ACM MM, pages 2054–2065, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[64] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.386 + ], + "angle": 0, + "content": "[66] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[67] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[68] Zhiqiang Tan, Yunfu Song, and Zhijian Ou. Calibrated adversarial algorithms for generative modelling. Stat, 8:e224, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[69] Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Hongsheng Li, and Xiaogang Wang. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024. 1, 3, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[70] Fu-Yun Wang, Zhaoyang Huang, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Computation-efficient personalized style video generation without personalized video data*. In SIGGRAPH ASIA Technical Communications, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[71] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. arXiv preprint arXiv:2410.07303, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[72] Shibo Wang and Pankaj Kanwar. Bfloat16: The secret to high performance on cloud tpus. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[75] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[76] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[77] Xuefeng Xiao, Lianwen Jin, Yafeng Yang, Weixin Yang, Jun Sun, and Tianhai Chang. Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition. Pattern Recognition, 72:72-81, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[78] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurlPS, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[79] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[80] Zunnan Xu, Zhentao Yu, Zixiang Zhou, Jun Zhou, Xiaoyu Jin, Fa-Ting Hong, Xiaozhong Ji, Junwei Zhu, Chengfei Cai, Shiyu Tang, et al. Hunyuanportrait: Implicit condition control for enhanced portrait animation. In CVPR, pages 15909-15919, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[81] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. In NeurIPS, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[82] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[83] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[84] Hongwei Yi, Shitong Shao, Tian Ye, Jiantong Zhao, Qingyu Yin, Michael Lingelbach, Li Yuan, Yonghong Tian, Enze Xie, and Daquan Zhou. Magic 1-for-1: Generating one minute video clips within one minute. arXiv preprint arXiv:2502.07701, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[85] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[86] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, pages 6613-6623, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[87] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16828" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[88] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[89] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[90] Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multidimensional human preference for text-to-image generation. In CVPR, pages 8018-8027, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[91] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, Sergey Tulyakov, and Jian Ren. Sf-v: Single forward video generation model. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[92] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation: Improved latent consistency distillation by semi-linear consistency function with trajectory mapping. arXiv preprint arXiv:2402.19159, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[93] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In ICML, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[94] Mingyuan Zhou, Huangjie Zheng, Yi Gu, Zhendong Wang, and Hai Huang. Adversarial score identity distillation: Rapidly surpassing the teacher in one step. In ICLR, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16829" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_origin.pdf b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c231d571c763125681ea37efc71e9bf36e6685ca --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/b23b5268-40ae-4571-a9ee-075cda737453_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:853409ac2a854cbeaf755cafc4023598d65e5eafd6896716b6f05f6ef3d9ee40 +size 6829761 diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/full.md b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f0e5a4be41e9bb25342761d6f879597c18ca82db --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/full.md @@ -0,0 +1,407 @@ +# Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis + +Yanzuo Lu $^{1,2}$ , Yuxi Ren $^{2}$ , Xin Xia $^{2}$ , Shanchuan Lin $^{2}$ , Xing Wang $^{2}$ , Xuefeng Xiao $^{2*}$ , Andy J. Ma $^{1,3,4\dagger}$ , Xiaohua Xie $^{1,3,4,5}$ , Jian-Huang Lai $^{1,3,4,5}$ + +1Sun Yat-Sen University 2ByteDance Seed Vision +3Guangdong Provincial Key Laboratory of Information Security Technology, China +4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China +5Pazhou Lab (HuangPu), Guangzhou, China + +oliveryanzulu@gmail.com, xiaoxuefeng.ailab@bytedance.com, majh8@mail.sysu.edu.cn + +# Abstract + +Distribution Matching Distillation (DMD) is a promising score distillation technique that compresses pre-trained teacher diffusion models into efficient one-step or multi-step student generators. Nevertheless, its reliance on the reverse Kullback-Leibler (KL) divergence minimization potentially induces mode collapse (or mode-seeking) in certain applications. To circumvent this inherent drawback, we propose Adversarial Distribution Matching (ADM), a novel framework that leverages diffusion-based discriminators to align the latent predictions between real and fake score estimators for score distillation in an adversarial manner. In the context of extremely challenging one-step distillation, we further improve the pre-trained generator by adversarial distillation with hybrid discriminators in both latent and pixel spaces. Different from the mean squared error used in DMD2 pre-training, our method incorporates the distributional loss on ODE pairs collected from the teacher model, and thus providing a better initialization for score distillation fine-tuning in the next stage. By combining the adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our proposed method achieves superior one-step performance on SDXL compared to DMD2 while consuming less GPU time. Additional experiments that apply multi-step ADM distillation on SD3-Medium, SD3.5-Large, and CogVideoX set a new benchmark towards efficient image and video synthesis. + +# 1. Introduction + +Recent methods for accelerating diffusion models [12, 17, 63, 65] have concentrated on reducing the sampling steps through distillation. The distillation process generally trains + +![](images/ee0a03c5ae49927ae0ea66fe1ac89fedf63a579444d366e3fc72ec2823526b01.jpg) +(a) A white polar bear cub wearing a red sitta in a meadow with flowers. + +![](images/a86eb6eb51f6704986d4549274c849c33d0622bfea3bd6e8472a9badd9d82f97.jpg) +(b) mage, wizard, female, demoness, forest, D&D, fantasy, incricate, elegant, highly detailed, digital painting, artstation, [...] + +![](images/516db7b264c7fbb3e87edbf3fd8f759cb41461506d5f4f3c353dbfcb0480f04d.jpg) +(c) beautiful nighttime landscape photography of the Rocky Mountains with a crystal blue lake, milky way galaxy sky, serene, dramatic lightning. + +![](images/36b49336c818ac960121d7ee4716ae2cefa5de355cffd08e566333b26f667651.jpg) +(d) A planeplane lacavillo glows over the surface of a sand with swirling mist around it. +Figure 1. In these images, some are generated by the baseline SDXL via 50NFE, the others with our DMDX in 1NFE. Can you tell which is the accelerated one? Answers in the footnote $\ddagger$ . + +a more efficient generator (a.k.a student model) that approximates the output distribution of the pretrained teacher. This field has observed multiple pathways developing in parallel with progressive distillation [23, 58], consistency distillation [18, 34, 35, 43, 55, 66, 69, 70, 92], score distillation [3, 16, 21, 38, 55, 59, 61, 84-87, 93], rectified flow [27, 28, 30, 71, 81] and adversarial distillation [23, 24, 43, 55, 60, 61, 85, 91]. They are currently appearing as distinct yet non-exclusive research directions. + +Distribution Matching Distillation (DMD) [85, 86] is a promising score distillation [51] approach, which distills the powerful text-to-image diffusion model SDXL-Base [56] into a one-step generator with great fidelity. While DMD's primary contribution lies in introducing additional regularizers to constrain the distribution matching loss, its predecessor Diff-Instruct [37] pioneered the use of a fake score estimator to approximate the student model's output distribution. This contrasts with the distillation loss in Adversarial Diffusion Distillation (ADD) [61], which directly em + +ploys the student model itself as score estimator. Theoretically, an intrinsic correspondence for DMD and the distillation loss in ADD can be equivalently found between Variational Score Distillation (VSD) [74] and Score Distillation Sampling (SDS) [51] in text-to-3D generation, where SDS represents a specialized instance of VSD through the use of a single-point Dirac distribution as the variational distribution [74]. And thus the VSD loss in DMD performs better than the SDS loss in ADD. Intuitively, since the capacity of a diffusion model decreases significantly when distilled for few-step generations [23], the student model no longer serves as a decent score estimator as the teacher. Therefore, the score-based distillation loss in ADD hardly contributes to its final performance. + +However, the optimization of DMD loss relies on the reverse Kullback-Leibler (KL) divergence minimization which is zero-forcing that drives low-probability regions to zero, causing the model to focus on only a few dominant modes and potentially leading to mode collapse [45]. To enhance sample diversity, DMD [86] employs an ODE-based (Ordinary Differential Equation) regularizer with synthetic data, while DMD2 [85] introduces a GAN-based (Generative Adversarial Network [8]) regularizer with real data to counterbalance this side effect. Subsequent efforts such as Moment Matching Distillation (MMD) [59], Score identity Distillation (SiD) [93] and Score Implicit Matching (SIM) [38] all used a variant of Fisher divergence to align the fake score estimator with pre-trained real score estimator [16]. Despite great success, their ability for distribution matching is limited by their dependence on a predefined form of explicit divergence metric. In this work, we would like to raise and investigate a question: Can we bypass the limitations of a predefined divergence by developing a framework that learns an implicit, data-driven discrepancy measure, thereby enabling more flexible and fine-grained matching of complex, high-dimensional distributions? + +Since the multifaceted alignment requirements in complex multimodal text-conditioned image or even video generation may not be fully captured, this motivates our exploration of more adaptive discrepancy learning paradigms. As our first contribution, we show how to align the latent predictions between real and fake score estimators to realize score distillation in an adversarial manner with prior knowledge of teacher model and dynamically learnable parameters, termed Adversarial Distribution Matching (ADM). This is different from the ODE-based or GAN-based regularizer in DMD and DMD2 that is used to counterbalance the mode collapse effect of reverse KL divergence. In contrast, we are performing distribution matching by means of GAN training to replace and circumvent the use of DMD loss, which we will provide more discussion in Sec. 4.1.3. + +Our second contribution concerns about the extremely challenging one-step distillation, where we observe that the + +score distillation has a higher risk of gradient exploding and vanishing. We attribute the issue more to the lower overlap of support sets between the student and teacher distributions, not just the approximation errors in the fake score estimator as attributed in DMD2 [85]. In other words, while score distillation yields superior generation quality, it imposes higher requirements on initialization especially when distilled in extremely few steps, which we will provide more in-depth analysis in Sec. 4.3.2. Although we notice that an ODE-based pre-training on synthetic data is used for SDXL one-step distillation in DMD2 implementation, the mean squared error loss was probably still not enough to provide more overlapping regions of support sets between the student and teacher distributions. Our experiments demonstrate that when providing a better initialization for distribution matching, the effect of Two Time-scale Update Rule (TTUR) [2] to the final performance is very limited. + +Naturally, our third contribution focuses on providing a better initialization for further score-based fine-tuning. We employ adversarial distillation to trade off sample quality and mode converge inspired by SDXL-Lightning [23] and LADD [60]. With this distribution-level loss optimization, we can pre-train the student model to capture more potential modes of the teacher model distribution, especially through a hybrid discriminator in both latent and pixel spaces which we will introduce later in Sec. 4.2. To facilitate the diversity, we also propose employing a cubic timestep schedule for the generator to bias towards higher noise levels. + +By combining adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our one-step SDXL provides competitive fidelity compared to the baseline with $50 \times$ acceleration in the Number of Function Evaluation (NFE) as shown in Fig. 1. More experiments on multi-step ADM distillation across the best existing diffusion models including SD3-Medium, SD3.5-Large [6], and CogVideoX [83] consistently achieve new benchmarks for efficient image and video synthesis. + +# 2. Related Work + +Progressive Distillation was proposed in [58] to distill multi-step prediction into one-step prediction of the same distance along the trajectory. SDXL-Lightning [23] extended this idea by utilizing GAN training and achieved one-step high-resolution (1024px) generation for the first time. This approach involving multi-stage process can be quite cumbersome, as it requires iteratively distilling from its predecessor, halving the sampling steps each time. + +Consistency Distillation proposed [7, 34, 35, 64, 66] to enforce the consistency property into diffusion model, i.e. predictions towards original sample are consistent for arbitrary pairs of noisy timesteps belong to the same trajectory. Subsequent efforts extended it into the trajectory consistency to relax the training objectives, i.e. predictions towards noisy + +samples of arbitrary subsequent timesteps are consistent, including CTM [18], TCD [92], TSCD [55] and PCM [69]. + +Rectified Flow [27, 28] aims to obtain faster straight trajectories through multiple reflow processes that iteratively learn the velocity of many ODE pairs from its predecessor. PeRFlow [81] tried splitting the trajectory into pieces and applying piece-wise rectification with real data samples. In this paper, we empirically found the straightness can also be satisfied via an adversarial distillation paradigm without the need for repeatedly collecting tremendous synthetic data. + +Score Distillation intuitively tries to keep that a sample appearing in the student model distribution at a specific noise level is with the same probability as it does in the teacher model distribution. In light of the motivation that relying on a single divergence might be problematic, a concurrent work Score-of-Mixture Distillation (SMD) [16] shares our perspective and explicitly designed a class of $\alpha$ -skew Jensen-Shannon (JS) divergences for optimization. In contrast, we implicitly measure the discrepancy between the fake and real score estimators in an adversarial manner, facilitating a more capable and adaptive discrepancy learning paradigm along the distillation process. + +Adversarial Distillation employs a discriminator to align the student model distribution with a specific target distribution [40, 94]. While methods UFOGen [79], DMD2 [85], and APT [24] directly align with the real data distribution, SDXL-Lightning [23] and Hyper-SD [55] utilize intermediate timestep predictions from the teacher model as approximation objectives. Inspired by LADD [60], our adversarial distillation pre-training aligns with ODE-based synthetic data generated from the teacher model. + +Human Feedback Learning was firstly considered by ReFL [78] in diffusion models that includes two stages of reward model training and preference fine-tuning. HyperSD [55] treated this as a standalone technique, ultimately applying LoRA insertion to shift the output distribution of the low-step generator toward human preferences. Subsequent studies, notably [36, 39], further attempted to integrate CFG and Score-based divergence. + +# 3. Preliminaries + +# 3.1. Diffusion Model + +Given a training data distribution $p_{\mathrm{data}}$ with standard deviation $\sigma_{\mathrm{data}}$ , the diffusion model [12] generates samples by reversing the forward diffusion process that progressively adds noise to a data sample $x_0 \sim p_{\mathrm{data}}$ as, + +$$ +\boldsymbol {q} \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x}\right) \sim \mathcal {N} \left(\alpha_ {t} \boldsymbol {x}, \sigma_ {t} ^ {2} \boldsymbol {I}\right), \tag {1} +$$ + +where $\alpha_{t} \geq 0, \sigma_{t} > 0$ are specified noise schedules such that $\alpha_{t} / \sigma_{t}$ satisfies monotonically decreasing w.r.t. $t$ and larger $t$ indicates greater noise. We consider two different formulations for denoising models. + +DDPM and DDIM [12, 63] assume discrete-time schedules + +with $t \in [1, T]$ (typically $T = 1000$ ) and noise prediction parameterization* [58]. The training objective is given by, + +$$ +\mathbb {E} _ {\boldsymbol {x} _ {0}, t, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I})} [ w (t) \| \epsilon_ {\theta} (\boldsymbol {x} _ {t}, t) - \epsilon \| _ {2} ^ {2} ], \tag {2} +$$ + +where $w(t)$ is a weighting function and $\epsilon_{\theta}$ is a neural network with parameters $\theta$ . The noise schedule is defined as $\alpha_{t} = \sqrt{\bar{\alpha}_{t}}, \sigma_{t} = \sqrt{1 - \bar{\alpha}_{t}}$ such that $\boldsymbol{x}_t = \sqrt{\bar{\alpha}_t} \boldsymbol{x}_0 + \sqrt{1 - \bar{\alpha}_t} \epsilon$ . For DDIM sampling, it solves the Probability-Flow Ordinary Differential Equations (PF-ODE) [65] by $d\bar{\boldsymbol{x}}_t = \epsilon_\theta (\frac{\bar{\boldsymbol{x}}_t}{\sqrt{\bar{\sigma}_t^2 + 1}}) d\bar{\sigma}_t$ , where $\bar{\boldsymbol{x}}_t = \frac{\boldsymbol{x}_t}{\sqrt{\bar{\alpha}_t}}$ and $\bar{\sigma}_t = \sqrt{\frac{1 - \bar{\alpha}_t}{\bar{\alpha}_t}}$ , starting from $\boldsymbol{x}_T \sim \mathcal{N}(\boldsymbol{0}, \boldsymbol{I})$ and stopping at $\boldsymbol{x}_0$ . + +Flow Matching [26, 27, 30] uses velocity prediction parameterization [58] and continuous-time coefficients (typically $\alpha_{t} = 1 - t$ , $\sigma_{t} = t$ with $t \in [0, T = 1]$ ). The conditional probability path or the velocity we say can be given as $\pmb{v}_{t} = \frac{d\alpha_{t}}{dt}\pmb{x}_{0} + \frac{d\sigma_{t}}{dt}\epsilon$ , such that the training objective is, + +$$ +\mathbb {E} _ {\boldsymbol {x} _ {0}, t, \epsilon \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I})} [ w (t) \| \boldsymbol {v} _ {\theta} (\boldsymbol {x} _ {t}, t) - \boldsymbol {v} _ {t} \| _ {2} ^ {2} ], \tag {3} +$$ + +where $w(t)$ is a weighting function and $\pmb{v}_{\theta}$ is a neural network parameterized by $\theta$ . The sampling procedure starts from $t = T$ with $\pmb{x}_T \sim \mathcal{N}(\pmb{0}, \pmb{I})$ and stops at $t = 0$ , solving the PF-ODE by $d\pmb{x}_t = \pmb{v}_{\theta}(\pmb{x}_t, t)dt$ . + +# 3.2. Distribution Matching Distillation + +DMD [85, 86] distills pretrained diffusion models $\pmb{F}_{\phi}(\pmb{x}_t,t)$ into one-step or multi-step efficient generators $G_{\theta}(\pmb{x}_t,t)$ by minimizing the reverse KL divergence between the target distribution $p_{\mathrm{real}}$ and the efficient generator output distribution $p_{\mathrm{fake}}$ . The gradient of DMD objective w.r.t. $\theta$ is, + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {D M D}} = \underset {\boldsymbol {z}, t ^ {\prime}, t, \boldsymbol {x} _ {t}} {\mathbb {E}} - \left[ \left(s _ {\text {r e a l}} \left(\boldsymbol {x} _ {t}\right) - s _ {\text {f a k e}} \left(\boldsymbol {x} _ {t}\right)\right) \frac {d \boldsymbol {G} _ {\theta} \left(\boldsymbol {z} , t ^ {\prime}\right)}{d \theta} \right], \tag {4} +$$ + +where $z \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $t'$ is randomly selected from predefined generator schedule, $t \sim \mathcal{U}(0,T)$ , and noisy samples $\boldsymbol{x}_t = \boldsymbol{q}(\boldsymbol{x}_t|\hat{\boldsymbol{x}}_0)$ are obtained by randomly diffusing the generator output $\hat{\boldsymbol{x}}_0 = \boldsymbol{G}_{\theta}(\boldsymbol{z},t')$ . Score functions $s_{\mathrm{real}}(\boldsymbol{x}_t) = \nabla_{\boldsymbol{x}_t}\log p_{\mathrm{real}}(\boldsymbol{x}_t)$ , $s_{\mathrm{fake}}(\boldsymbol{x}_t) = \nabla_{\boldsymbol{x}_t}\log p_{\mathrm{fake}}(\boldsymbol{x}_t)$ are vector fields that point towards higher density of data at a given noise level [17, 65] for $p_{\mathrm{real}}$ and $p_{\mathrm{fake}}$ , respectively. + +While the real score estimator is the teacher model $\mathbf{F}_{\phi}(\mathbf{x}_t,t)$ itself, the fake score estimator $\pmb{f}_{\psi}(\pmb{x}_t,t)$ is initialized the same as $\pmb{F}_{\phi}(\pmb{x}_t,t)$ and dynamically learned to describe $p_{\mathrm{fake}}$ with pretrain loss as Eqs. (2) and (3). In practice, the gradient in Eq. (4) is computed as, + +$$ +\operatorname {g r a d} \left(\hat {\boldsymbol {x}} _ {0}, \boldsymbol {x} _ {t}, t\right) = \frac {\boldsymbol {f} _ {\psi} \left(\boldsymbol {x} _ {t} , t\right) - \boldsymbol {F} _ {\phi} \left(\boldsymbol {x} _ {t} , t\right)}{\left\| \hat {\boldsymbol {x}} _ {0} - \boldsymbol {F} _ {\phi} \left(\boldsymbol {x} _ {t} , t\right) \right\| _ {1}} \tag {5} +$$ + +such that the training loss is implemented like, + +$$ +\mathcal {L} _ {\mathrm {D M D}} (\theta) = \underset {\boldsymbol {z}, t ^ {\prime}, t, \boldsymbol {x} _ {t}} {\mathbb {E}} [ \| \hat {\boldsymbol {x}} _ {0} - s g (\hat {\boldsymbol {x}} _ {0} - \operatorname {g r a d} (\hat {\boldsymbol {x}} _ {0}, \boldsymbol {x} _ {t}, t)) \| _ {2} ^ {2} ], \tag {6} +$$ + +where $sg(\cdot)$ denotes the stop gradient operation. + +![](images/e44a13f3cc8690a0fafb9e59c1881a073abc346d726524c940fcb433605a9033.jpg) +ADM + +![](images/51df300708192b09c7df88f42838958f4f8e97d8e6a1c9cc8c73c573f7edf2be.jpg) +ADP +Figure 2. Overall pipeline of our proposed Adversarial Distribution Matching (ADM) and Adversarial Distillation Pre-training (ADP). + +# 4. Methodology + +# 4.1. Adversarial Distribution Matching + +Instead of using a predefined divergence between the fake and real distributions, we use an implicit, data-driven discrepancy measure through an adversarial discriminator. Specifically, our discriminator $D_{\tau}(\pmb{x}_t,t)$ consists of a frozen latent diffusion model initialized the same as teacher model $F_{\phi}(\pmb{x}_t,t)$ and multiple trainable heads added upon different UNet [57] or DiT [49] blocks. Given the noisy sample $\pmb{x}_t = \pmb{q}(\pmb{x}_t|\hat{\pmb{x}}_0)$ diffused from the output of few-step generator $\hat{\pmb{x}}_0 = \pmb{G}_{\theta}(\pmb{z},t')$ , the score estimators no longer solve the PF-ODE w.r.t. the end point $\pmb{x}_0^{\mathrm{fake}} = \pmb{f}_{\psi}(\pmb{x}_t,t)$ and $\pmb{x}_0^{\mathrm{real}} = \pmb{F}_{\phi}(\pmb{x}_t,t)$ as used in Eq. (5). + +Instead, we set a fixed timestep interval $\Delta t$ (defaults to $T / 64$ ) and solve the PF-ODE w.r.t. $(t - \Delta t)$ , such that the fake sample $\pmb{x}_{t - \Delta t}^{\mathrm{fake}}$ and real sample $\pmb{x}_{t - \Delta t}^{\mathrm{real}}$ can be obtained to serve as score predictions and sent into the discriminator. The discriminator hierarchically aggregates features from frozen backbone layers and dynamically weights them through multiple learnable heads, establishing an adaptive discrepancy metric that leverages both diffusion priors and data-driven trainable dynamics. We use Hinge loss [22] to train the generator $\pmb{G}_{\theta}(\pmb{x}_t,t)$ and discriminator $D_{\tau}(\pmb{x}_t,t)$ alternatively. This encourages the fake score prediction $\pmb{x}_{t - \Delta t}^{\mathrm{fake}}$ to be closer to the real score prediction $\pmb{x}_{t - \Delta t}^{\mathrm{real}}$ : + +$$ +\mathcal {L} _ {\mathrm {G A N}} (\theta) = \underset {\boldsymbol {x} _ {t - \Delta t} ^ {\text {f a k e}}} {\mathbb {E}} \left[ - \boldsymbol {D} _ {\tau} \left(\boldsymbol {x} _ {t - \Delta t} ^ {\text {f a k e}}, t - \Delta t\right) \right] \tag {7} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {G A N}} (\tau) = \underset {\boldsymbol {x} _ {t - \Delta t} ^ {\text {f a k e}}, \boldsymbol {x} _ {t - \Delta t} ^ {\text {r e a l}}} {\mathbb {E}} [ \max (0, 1 + D _ {\tau} (\boldsymbol {x} _ {t - \Delta t} ^ {\text {f a k e}}, t - \Delta t)) \tag {8} \\ \left. + \max \left(0, 1 - D _ {\tau} \left(x _ {t - \Delta t} ^ {\text {r e a l}}, t - \Delta t\right)\right) \right] \\ \end{array} +$$ + +In conjunction with the dynamically learned fake model, we clarify our training procedure in Appendix A. The overall pipeline is demonstrated in Fig. 2. + +# 4.1.1. Motivation of Discriminator Timestep $(t - \Delta t)$ + +Given that the final objective of score distillation is to match the probability flows varying with noise levels of student and teacher models to exactly the same, timestep information must be considered when measuring the discrepancy between distributions. This coincides with our discriminator design that uses a pre-trained diffusion model, and we take a small step alongside the PF-ODE, succeeding in preserving the input timestep information of score estimators. + +# 4.1.2. Data-driven Effect + +The flexibility of using discriminators for distributional discrepancy measure is not only in the noise level of score function, but also within the distillation process. As distillation iterates, the model is exposed to increasingly diverse data, causing the discrepancies in modes between the two distributions to change. In the early phase of training where the discrepancy is significant, a more global evaluation is necessary, whereas in the later phases, when the discrepancy becomes minor, a more localized, fine-grained optimization might become essential. In other words, driven + +![](images/45d0869ca4553ec88ba79e6c155f4aa6a56f3aa919ecde8d7203fb67071d8a89.jpg) +Figure 3. Changes of DMD loss over multi-step ADM distillation for CogVideoX. Note that we did not optimize this objective directly during ADM distillation but recorded it over iterations. + +by the volume of data, the divergence measure employed across different training phases can vary. + +# 4.1.3. Relationship with DMD and DMD2 + +To mitigate the mode collapse issue in DMD loss, an ODE-based regularizer and a GAN-based regularizer are additionally used for distillation in DMD [85] and DMD2 [85], respectively. However, these two regularizers do not fundamentally address the mode-seeking behavior introduced by reverse KL divergence as shown in Fig. 4(a), but rather counterbalance it by trade-offs between losses. In ADM, our adversarial loss is actually playing the role of DMD loss to realize score distillation with an implicit, data-driven discrepancy measure instead of a predefined divergence. Therefore, our motivation for using GAN training in ADM is different from that of DMD2 [85] and we don't require additional regularizers. + +Intuitively, the learnable discriminator can approximate any nonlinear function to implicitly measure distribution divergence, which probably inherently encompasses the reverse KL divergence in DMD loss. As shown in Fig. 3, we visualize the changes of DMD loss in Eq. (6) during the multi-step ADM distillation on CogVideoX [83]. Though not directly optimize on Eq. (6), the results indicate a very steady downward trend that supports our assumption. More theoretical discussion is provided in Sec. 4.3.3. + +# 4.2. Adversarial Distillation Pre-training + +To stabilize the extremely difficult one-step distillation, we opt to provide a better initialization for ADM fine-tuning with adversarial distillation pre-training on synthetic data. Our pre-training configuration refers to Rectified Flow [27] in several aspects, where we 1) collect the ODE pairs from teacher model in an offline manner, 2) construct noisy samples by linearly interpolation between pure noise and clean data sample of the ODE pair, 3) alter the prediction target of the generator to the velocity of ODE pair. + +As for adversarial training, we formulate a latent-space discriminator $D_{\tau_1}(\boldsymbol{x}_t,t)$ initialized from teacher model and a pixel-space discriminator $D_{\tau_2}(\boldsymbol{x})$ initialized from the vision encoder of SAM [19] model, respectively, as shown in Fig. 2. We also append multiple trainable heads to both the backbone networks similar to the practice in ADM. All these contribute to increasing the discriminative capability, + +facilitating the student model in discovering more potential modes in the teacher model distribution. Specifically, let $\tilde{\pmb{x}}_0 = G_\theta (\pmb {x}_t,t)$ denote the predicted PF-ODE endpoint of generator, where $\pmb{x}_{t}$ represents the noisy sample interpolated between an ODE pair $(\pmb {x}_T,\pmb {x}_0)$ with random timestep $t\in [0,T]$ . For latent-space discriminator, we diffuse the generator output with another random noise and timestep $t^{\prime}\in (0,T]$ , getting $\tilde{\pmb{x}}_{t^{\prime}} = \pmb {q}(\tilde{\pmb{x}}_{t^{\prime}}|\tilde{\pmb{x}}_0)$ as its input. For pixel-space discriminator, the generator output will be first decoded via VAE decoder and then sent into the vision encoder. The training objective, also based on Hinge loss [22], encourages the generator output $\tilde{\pmb{x}}_0$ to be closer to the synthetic data sample $\pmb{x}_0$ : + +$$ +\mathcal {L} _ {\mathrm {G A N}} (\theta) = \underset {\tilde {\boldsymbol {x}} _ {0}, t ^ {\prime}} {\mathbb {E}} - \left[ \lambda_ {1} D _ {\tau_ {1}} \left(\tilde {\boldsymbol {x}} _ {t ^ {\prime}}, t ^ {\prime}\right) + \lambda_ {2} D _ {\tau_ {2}} \left(\tilde {\boldsymbol {x}} _ {0}\right) \right] \tag {9} +$$ + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {G A N}} \left(\tau_ {1}, \tau_ {2}\right) = \underset {\boldsymbol {x} _ {0}, \tilde {\boldsymbol {x}} _ {0}, t ^ {\prime}} {\mathbb {E}} \left[ \lambda_ {1} \cdot \max \left(0, 1 + \boldsymbol {D} _ {\tau_ {1}} \left(\tilde {\boldsymbol {x}} _ {t ^ {\prime}}, t ^ {\prime}\right)\right) \right. \\ + \lambda_ {2} \cdot \max \left(0, 1 + D _ {\tau_ {2}} \left(\tilde {x} _ {0}\right)\right) \tag {10} \\ + \lambda_ {1} \cdot \max (0, 1 - D _ {\tau_ {1}} \left(\boldsymbol {x} _ {t ^ {\prime}}, t ^ {\prime}\right)) \\ \left. \left. + \lambda_ {2} \cdot \max \left(0, 1 - D _ {\tau_ {2}} \left(\boldsymbol {x} _ {0}\right)\right) \right] \right. \\ \end{array} +$$ + +We empirically found that setting balancing coefficients $\lambda_{1} = 0.85, \lambda_{2} = 0.15$ produces visually coherent results. + +# 4.2.1. Cubic Generator Timestep Schedule + +Intuitively, higher noise levels encourage exploration of new modes by weakening the restrictive information encoded in latent representations. Therefore, we propose employing a cubic timestep schedule for the generator. The schedule maps uniform $[0,T)$ samples through $[1 - (t / T)^3 ]*T$ , non-linearly concentrating values near $T$ with heavy noise similar to LADD [60]. + +# 4.2.2. Uniform Discriminator Timestep Schedule + +Unlike in ADM where the discriminator is firstly used for score distillation, the utilization of latent-space discriminator is common for adversarial distillation works. Inspired by SDXL-Lightning [23], where they found that the diffusion encoder is trained to focus on high-frequency details at lower timesteps and low-frequency structures at higher timesteps, we set a uniform $(0,T]$ for discriminator timestep $t'$ to capture both advantages during pre-training. + +# 4.2.3. Relationship with LADD + +Our motivation to perform adversarial distillation on synthetic data is inspired by LADD [60], but differs in a lot that we 1) construct noisy samples via ODE pairs in the style of Rectified Flow [27] rather than random noise, 2) develop a cubic generator timestep schedule that facilitate deterministic Euler sampling rather than consistency ones, 3) introduce an additional pixel-space encoder to increase the capability of discriminator and find more modes. + +# 4.3. Discussion + +# 4.3.1. Difference between ADM and ADP + +One question may appear like what is the difference between these two adversarial approaches with latent-space discriminators? The effectiveness in score distillation correlates with the fact that score function $\nabla_{\pmb{x}}\log p(\pmb{x};\sigma (t))$ is defined at different noise levels $\sigma (t)$ . In contrast, adversarial distillation only aligns the distribution of clean data samples when $t = 0$ . While the latent-space discriminator in pre-training is capturing information at different scales and details by randomly diffusing the generator output, the crux for ADM is more than that by solving the PF-ODE of both score estimators. In other words, ADM additionally supervises the complete denoising process through noisy samples that are with higher density at different noise levels of respective distribution. + +This leads to a situation in ADM when two distributions initialized with less overlap in support sets, the noisy samples remain in regions unfamiliar to each other, and the discriminator can easily distinguish between them thus leading to extreme gradient signals. However, since Gaussian noise is isotropic, we artificially create overlapping regions for the randomly diffused samples in ADP to make the discrimination more difficult, resulting in relatively smooth gradients. Therefore, our ADM still falls under score distillation given that it encourages the entire probability flows to be closer, while the pre-training belongs to adversarial distillation because it only cares about the clean data distribution at $t = 0$ . + +# 4.3.2. Importance of Pre-training + +Another question we haven't discussed so far is why do we need pre-training for one-step score distillation? Taking the reverse KL divergence used by DMD loss as an example: + +$$ +\mathbb {D} _ {\mathrm {K L}} \left(p _ {\text {f a k e}} \| p _ {\text {r e a l}}\right) = \int p _ {\text {f a k e}} (\boldsymbol {x}) \log \frac {p _ {\text {f a k e}} (\boldsymbol {x})}{p _ {\text {r e a l}} (\boldsymbol {x})} d \boldsymbol {x}. \tag {11} +$$ + +When employing one-step distillation, the generator output is worse in visual fidelity and structural integrity compared to multi-step sampling, leading $p_{\mathrm{fake}}(\pmb{x}) \to 0$ where $p_{\mathrm{real}}(\pmb{x}) > 0$ in regions. The integrand $p_{\mathrm{fake}}(\pmb{x}) \log \frac{p_{\mathrm{fake}}(\pmb{x})}{p_{\mathrm{real}}(\pmb{x})}$ approaching zero $0 \cdot (-\infty)$ causes the optimization to avoid regions where $p_{\mathrm{real}}(\pmb{x}) > 0$ but $p_{\mathrm{fake}}(\pmb{x})$ has negligible density, a phenomenon called zero-forcing. Instead of fully covering $p_{\mathrm{real}}$ 's support, $p_{\mathrm{fake}}$ collapses to a subset of modes of $p_{\mathrm{real}}$ , inducing mode-seeking behavior as illustrated in Fig. 4(a). During training this manifests itself as gradient vanishing sometimes. And conversely, the fuzzy samples that the diffusion model typically produces through one step are also not within the teacher model distribution, yielding $p_{\mathrm{real}}(\pmb{x}) \to 0$ where $p_{\mathrm{fake}}(\pmb{x}) > 0$ in regions and the integrand $p_{\mathrm{fake}}(\pmb{x}) \log \frac{p_{\mathrm{fake}}(\pmb{x})}{0}$ diverging to $+\infty$ , and thus resulting in numerical instability and gradient exploding. + +Similarly, when the support sets of the student and + +![](images/562e6a4f4d7a2a7a42b3338e938721c3c930037b14d245b68b42d0041bb47780.jpg) +Figure 4. Illustration for theoretical discussion. + +teacher distributions overlap hardly at all, forward KL divergence approaches $+\infty$ where $p_{\mathrm{fake}}(\boldsymbol{x}) > 0$ , while JS divergence saturates to a constant log 2 and Fisher divergence may degenerate without definition. Therefore, when this assumption is undermined, many of the single divergences no longer apply and a better initialization with more overlapping regions becomes essential as exemplified in Fig. 4(b). + +# 4.3.3. Theoretical Objective + +The last question is why ADM is better than DMD loss theoretically? In fact, the Hinge GAN [22] we use has been proven to minimize Total-Variation Distance (TVD) [68]: + +$$ +T V \left(p _ {\text {f a k e}}, p _ {\text {r e a l}}\right) = \int \left| p _ {\text {f a k e}} (\boldsymbol {x}) - p _ {\text {r e a l}} (\boldsymbol {x}) \right| d \boldsymbol {x} \tag {12} +$$ + +That is to say, when the discriminator is sufficiently rich and well trained, the theoretical optimum of Hinge loss upon convergence minimizes TVD. When the support sets of fake and real distributions have minimal overlap, TVD provides two key advantages over reverse KL divergence: 1) Symmetry: TVD yields the same discrepancy measure regardless of the initial distribution, whereas the asymmetric reverse KL may exhibit mode-seeking behavior and neglect other portions of the overall distribution. For example, as illustrated in Fig. 4(c), TVD maintains substantial loss values and provides optimization directions covering the mode when $p_{\mathrm{fake}}(\pmb{x}) \to 0$ while $p_{\mathrm{real}}(\pmb{x}) > 0$ , whereas the reverse KL divergence suffers from gradient vanishing issues in such scenarios, as discussed in Sec. 4.3.2. 2) Boundedness: TVD is bounded within [0,1], and thus mitigating disruptions from outliers during training, especially in our high-dimensional multimodal text-conditioned image and video distributions, avoiding the numerical instability inherent to reverse KL divergence caused by gradient exploding. + +
MethodStep NFECLIP ScorePick ScoreHPSv2MPS
ADD [61] (512px)1135.008822.152427.097110.4340
LCM [34]1228.466920.126723.82464.8134
Lightning [23]1133.498521.919427.155710.2285
DMD2 [85]1135.215322.097827.452310.6947
DMDX (Ours)1135.255722.273627.704611.1978
SDXL-Base [56]255035.030922.249427.374310.7042
+ +Table 1. Quantitative results on fully fine-tuning SDXL-Base. + +
MethodStepNFECLIP ScorePick ScoreHPSv2MPS
TSCD [61]4834.018521.966527.272810.8600
PCM [69] (Shift=1)4433.504221.970327.368010.5707
PCM [69] (Shift=3)4433.381821.939627.114610.5635
PCM [69] (Stoch.)4433.418521.882227.317710.5200
Flash [3]4434.397822.090427.258610.6634
ADM (Ours)4434.907622.547128.449211.9543
SD3-Medium [6]255034.763322.296127.973311.3652
LADD [60]4434.739522.395827.492311.4372
ADM (Ours)4434.973022.884227.733112.2350
SD3.5-Large [6]255034.966822.508727.968811.5826
+ +Table 2. Quantitative results on LoRA fine-tuning SD3-Medium and fully fine-tuning SD3.5-Large. + +
MethodStepNFEFinal ScoreQuality ScoreSemantic Score
ADM8878.58480.82569.621
+ Longer Training ×28880.76483.03171.693
ADM w/ CFG81679.86580.93875.569
+ Longer Training ×281681.79683.00876.947
CogVideoX-2b [83]10020080.03680.80176.974
ADM8882.06783.22777.423
ADM w/ CFG81680.98282.16576.251
CogVideoX-5b [83]10020081.22681.78578.987
+ +Table 3. Quantitative results on fully fine-tuning CogVideoX. + +# 5. Experiments + +Models. For one-step distillation, we employ Adversarial Distillation Pre-training (ADP) and ADM fine-tuning on SDXL-Base [56], termed DMDX. For multi-step distillation, we only employ ADM training on both text-to-image models SD3-Medium, SD3.5-Large [6], and text-to-video models CogVideoX-2b, CogVideoX-5b [83]. Following most concurrent works, we didn't use Classifier-Free Guidance (CFG) [11] in our text-to-image model. We tried conducting CFG-integrated experiments on the text-to-video model, with the approach detailed in Sec. 5.2. + +Datasets. No visual data is required for both ADP and ADM proposed in this work. For image generators, we utilize the text prompts from JourneyDB [67] that exhibits high-level of detail and specificity for training. For video generators, we collect training prompts from OpenVid-1M [47], Vript [82] and Open-Sora-Plan-v1.1.0 [50]. + +Evaluation. The image generators are evaluated on 10K + +![](images/242459914feae3653c71bb6d02d3a2a9f2dd3825aa50466cd4eb0e0f04af7879.jpg) +Figure 5. Qualitative results on fully fine-tuning SDXL-Base. + +prompts from COCO 2014 [25] following DMD2 [85]. We report the CLIP score [52] and human preference benchmarks PickScore [20], HPSv2 [75] and MPS [90] as many concurrent works including Hyper-SD [55] and Emu3 [73]. But we don't include Hyper-SD in the one-step quantitative comparison because one-step Hyper-SDXL has been optimized directly on human feedback with ReFL [78]. Instead, we compare with the TSCD algorithm proposed therein on SD3-Medium [13], since 4-step Hyper-SD3 LoRA is without ReFL optimization. The video generators are evaluated on VBenchmark [14], which consists of multiple quality and semantic dimensions comprehensively. + +Hyperparameters. Despite having multiple models to train in our proposed ADP and ADM, we achieve satisfactory visual fidelity and structural integrity without extensive hyperparameter tuning. For the remainder of this work, we only adjust the learning rate of generator for different experiments. The optimizer settings for discriminators and fake models are uniform across all the experiments. Unless longer training is specifically noted, we train the same for only 8K iterations with a batch size of 128 and 8 for text-to-image and text-to-video models, respectively. More specific implementation details are provided in Appendix B. + +# 5.1. Efficient Image Synthesis + +Tab. 1 quantitatively compares our two-stage approach combing ADP and single-step ADM distillation with ex + +
AblationCLIP ScorePick ScoreHPSv2MPS
Ablation on adversarial distillation.
A1: Rectified Flow [27]27.437620.021123.60934.4518
A2: DINOv2 as pixel-space34.183621.875027.103910.2407
A3: λ1 = 0.7, λ2 = 0.333.694321.634426.89029.9633
A4: λ1 = 1.0, λ2 = 0.033.892921.739526.786910.0757
A5: w/o ADM (ADP only)35.772322.009527.349910.6646
Ablation on score distillation.
B1: ADM w/o ADP32.502021.763126.873210.8986
B2: DMD Loss w/o ADP32.748221.034125.96808.8977
B3: DMD Loss w/ ADP34.511921.936627.398510.6046
B4: DMDX (Ours)35.255722.273627.704611.1978
+ +isting one-step distillation methods on fully fine-tuning SDXL-Base. The results show that our method achieves excellent performance on both image-text alignment and human preference, which is consistent with the qualitative comparisons in Fig. 5, including better portrait aesthetic, animal hair details, subject-background separation and physical structure. + +For multi-step ADM distillation, it can serve as a standalone score distillation method. We tried both fully finetuning and LoRA fine-tuning [13] configurations and the quantitative results in Tab. 2 demonstrate our superior performance. Qualitative results are provided in Appendix C. + +# 5.2. Efficient Video Synthesis + +As quantitatively shown in Tab. 3, except for the regular 8-step ADM distillation for both sizes of CogVideoX, we also try integrating Classifier-Free Guidance (CFG) [11] for text-to-video task. Specifically, we assign the CFG scale for real model by randomly sampling within the range [5.0, 7.0], while assigning the few-step generator's scale through explicit subtraction of 2.0 from the real model's value. Empirically, we found that this subtractive offset requires progressive enlargement as the targeted sampling steps decreases. There is no CFG for the fake model. The VBench [14] results demonstrate that our few-step generators achieve comparable performance to the base model with $92 - 96\%$ acceleration. We conduct additional evaluations on the longer training of 2B model, motivated by the observation that DMD loss did't converge well at 8K iterations as Fig. 3 indicates. Experimental results demonstrate that the DMD loss is also approximately optimized by the learnable discriminator during ADM distillation. More quantitative and qualitative results refer to Appendix C. + +# 5.3. Ablation Studies + +In Tab. 4, we conduct extensive ablation studies on fully fine-tuning SDXL-Base to validate our effectiveness. Qualitative comparisons are provided in Appendix C. + +Effect of ADP. The following conclusions can be drawn: 1) Single reflow process provides very limited effectiveness + +Table 4. Quantitative results on ablation studies. + +
TTURTraining TimeCLIP ScorePick ScoreHPSv2MPS
1×1.0035.255722.273627.704611.1978
4×1.8535.258322.277327.725511.2720
8×2.5335.329922.288327.758611.2838
+ +Table 5. Ablation study on two time-scale update rule. + +
ADDLCMLightningDMD2OursTeacher
LPIPS↑0.60710.62570.67070.67150.71560.6936
+ +Table 6. Quantitative diversity evaluation on PartiPrompts [88]. + +(A1), while multiple processes achieving effective rectification requires considerable computational expense. 2) Using SAM [19] offers more imaging fidelity compared to the widely adopted DINOv2 [48] (A2/A5), which is likely due to SAM's higher resolution of 1024px (compared to DINOv2's 518px). 3) The weighting $\lambda_{2}$ for pixel space should neither be too large nor too small. Excessive weighting leads to degraded structural integrity (A3/A5), while insufficient weighting results in blurriness (A4/A5). Qualitative results in Appendix C provide more visible distinction. + +Effect of ADM. We summarize the key findings: 1) The absence of ADP exerts substantial performance degradation (B1/B4), which is aligned with our analysis in Sec. 4.3.2. 2) Without a regularizer, the DMD loss underperforms compared to standalone ADM (B1/B2), indicating its poor robustness. 3) Although DMD loss optimization also benefits from ADP (B2/B3), its distribution matching capability remains inferior to ADM (B3/B4). + +Effect of TTUR. Tab. 5 presents the impact of different TTUR settings on final performance and training duration. The results demonstrate that increasing TTUR yields only marginal performance gains while nearly doubling training time, rendering this trade-off clearly unwarranted. This highlights the critical role of our proposed ADP in one-step distillation and suggests that the training instability in DMD2 likely stems from insufficient support set overlap. + +Diversity evaluation. Following DMD2 [85], we generate 4 samples per prompt on Partiprompts [88] with different seeds, and report the averaged pairwise LPIPS similarities [89] in Tab. 6. The results indicate that our method significantly outperforms others in diversity. More randomly curated multi-seed samples can be found in Appendix C. + +# 6. Limitations + +We realize that one weakness is that teacher model might require CFG to produce accurate score prediction. Our experiments suggest this is a common characteristic of score distillation methods generally, not a limitation unique to our approach. This limits the application to guidance-distilled models such as FLUX.1-dev [1], which could be a promising research topic for future work. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China (U22A2095, 12326618, 62276281), Guangdong Basic and Applied Basic Research Foundation, China (2024A1515011882) and the Project of Guangdong Provincial Key Laboratory of Information Security Technology (2023B1212060026). + +# References + +[1] Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024.8 +[2] Naresh Babu Bynagari. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2 +[3] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 1, 7, 2 +[4] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 7, 1 +[7] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. In ICLR, 2025. 2 +[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1 +[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 1 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 7, 8 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3 +[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 7, 8 + +[14] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 7, 8, 2, 3 +[15] Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models. arXiv preprint arXiv:2309.14509, 2023. 2 +[16] Tejas Jayashankar, J. Jon Ryu, and Gregory Wornell. Score-of-mixture training: Training one-step generative models made simple via score estimation of mixture distributions. arXiv preprint arXiv:2502.09609, 2025. 1, 2, 3 +[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 1, 3 +[18] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 1, 3 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In ICCV, pages 3992-4003, 2023. 5, 8, 1 +[20] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeuriPS, 2023. 7 +[21] Jonas Kohler, Albert Pumarola, Edgar Schonfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Thabet. Imagine flash: Accelerating emu diffusion models with backward distillation. arXiv preprint arXiv:2405.05224, 2024. 1 +[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 4, 5, 6 +[23] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2, 3, 5, 7 +[24] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, Lu Jiang, and ByteDance Seed. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 1, 3 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 7 +[26] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2023. 3 +[27] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2022. 1, 3, 5, 8 +[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality + +diffusion-based text-to-image generation. In ICLR, 2024. 1, 3 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 2 +[30] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 3 +[31] Yanzuo Lu, Manlin Zhang, Yiqi Lin, Andy J. Ma, Xiaohua Xie, and Jianhuang Lai. Improving pre-trained masked autoencoder via locality enhancement for person re-identification. In PRCV, pages 509–521, 2022. 5 +[32] Yanzuo Lu, Meng Shen, Andy J Ma, Xiaohua Xie, and JianHuang Lai. Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation. In AAAI, pages 3900-3908, 2024. 5 +[33] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In CVPR, pages 6420-6429, 2024. 5 +[34] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2, 7 +[35] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 1, 2 +[36] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. In TMLR, 2024. 3 +[37] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, pages 76525-76546, 2023. 1 +[38] Weijian Luo, Zemin Huang, Zhengyang Geng, J. Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching. In NeurIPS, 2024. 1, 2 +[39] Weijian Luo, Colin Zhang, Debing Zhang, and Zhengyang Geng. David and goliath: Small one-step model beats large diffusion with score post-training. In ICML, 2025. 3 +[40] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. In ICLR, 2025. 3 +[41] Hongxu Ma, Guanshuo Wang, Fufu Yu, Qiong Jia, and Shouhong Ding. Ms-detr: Towards effective video moment retrieval and highlight detection by joint motion-semantic learning. In ACMMM, 2025. 5 +[42] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Ji-hong Guan, and Shuiqeng Zhou. Fine-grained zero-shot object detection. In ACMMM, 2025. 5 +[43] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. arXiv preprint arXiv:2409.11367, 2024. 1 +[44] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen + +Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In CVPR, pages 21203-21214, 2025. 5 +[45] Thomas Minka. Divergence measures and message passing. Microsoft Research, Technical Report, 2005. 2 +[46] Movie Gen Team. Movie gen: A cast of media foundation models. https://ai.meta.com/static-resource/movie-gen-research-paper, 2024.2 +[47] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 7 +[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 8, 1, 5 +[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 4, 1 +[50] PKU-Yuan Lab and Tuzhan AI. Open-sora-plan-v1.1.0. https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0, 2024.7 +[51] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 7, 6 +[53] Samyam Rajbhandari, Jeff Rasley, Olatunj Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC, 2020. 2 +[54] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 1 +[55] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. In NeurIPS, 2024. 1, 3, 7 +[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 7, 5 +[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 1 +[58] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 1, 2, 3 +[59] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. In NeurIPS, 2024. 1, 2 + +[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 1, 2, 3, 5, 7 +[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 1, 7 +[62] Meng Shen, Yanzuo Lu, Yanxu Hu, and Andy J. Ma. Collaborative learning of diverse experts for source-free universal domain adaptation. In ACM MM, pages 2054–2065, 2023. 5 +[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3 +[64] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In ICLR, 2024. 2 +[65] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3 +[66] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 1, 2 +[67] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 7 +[68] Zhiqiang Tan, Yunfu Song, and Zhijian Ou. Calibrated adversarial algorithms for generative modelling. Stat, 8:e224, 2019. 6 +[69] Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Hongsheng Li, and Xiaogang Wang. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024. 1, 3, 7, 2 +[70] Fu-Yun Wang, Zhaoyang Huang, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Computation-efficient personalized style video generation without personalized video data*. In SIGGRAPH ASIA Technical Communications, 2024. 1 +[71] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. arXiv preprint arXiv:2410.07303, 2024. 1 +[72] Shibo Wang and Pankaj Kanwar. Bfloat16: The secret to high performance on cloud tpus. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus, 2019.2 +[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 7 +[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and + +diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2 +[75] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7 +[76] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 1 +[77] Xuefeng Xiao, Lianwen Jin, Yafeng Yang, Weixin Yang, Jun Sun, and Tianhai Chang. Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition. Pattern Recognition, 72:72-81, 2017. 5 +[78] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurlPS, 2023. 3, 7 +[79] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3 +[80] Zunnan Xu, Zhentao Yu, Zixiang Zhou, Jun Zhou, Xiaoyu Jin, Fa-Ting Hong, Xiaozhong Ji, Junwei Zhu, Chengfei Cai, Shiyu Tang, et al. Hunyuanportrait: Implicit condition control for enhanced portrait animation. In CVPR, pages 15909-15919, 2025. 5 +[81] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. In NeurIPS, 2024. 1, 3 +[82] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 7 +[83] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5, 7 +[84] Hongwei Yi, Shitong Shao, Tian Ye, Jiantong Zhao, Qingyu Yin, Michael Lingelbach, Li Yuan, Yonghong Tian, Enze Xie, and Daquan Zhou. Magic 1-for-1: Generating one minute video clips within one minute. arXiv preprint arXiv:2502.07701, 2025. 1 +[85] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 1, 2, 3, 5, 7, 8 +[86] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, pages 6613-6623, 2024. 1, 2, 3 +[87] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 1 + +[88] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022. 8 +[89] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8 +[90] Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multidimensional human preference for text-to-image generation. In CVPR, pages 8018-8027, 2024. 7 +[91] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, Sergey Tulyakov, and Jian Ren. Sf-v: Single forward video generation model. In NeurIPS, 2024. 1 +[92] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation: Improved latent consistency distillation by semi-linear consistency function with trajectory mapping. arXiv preprint arXiv:2402.19159, 2024. 1, 3 +[93] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In ICML, 2024. 1, 2 +[94] Mingyuan Zhou, Huangjie Zheng, Yi Gu, Zhendong Wang, and Hai Huang. Adversarial score identity distillation: Rapidly surpassing the teacher in one step. In ICLR, 2025. 3 \ No newline at end of file diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/images.zip b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2d0964a489c494d9c5c7cf53794e199473d58002 --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b80dc2443a30219a427aa55a75a43775a3ab7923f1dcf8111beac9452035738 +size 607941 diff --git a/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/layout.json b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3ee77a2fcefcbdcfb4c6fdfc32a0c1f00be83f7c --- /dev/null +++ b/2025/Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis/layout.json @@ -0,0 +1,11258 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 491, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 491, + 140 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 491, + 140 + ], + "type": "text", + "content": "Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "spans": [ + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": "Yanzuo Lu" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Yuxi Ren" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xin Xia" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Shanchuan Lin" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xing Wang" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xuefeng Xiao" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Andy J. Ma" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1,3,4\\dagger}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Xiaohua Xie" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1,3,4,5}" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "text", + "content": ", Jian-Huang Lai" + }, + { + "bbox": [ + 122, + 160, + 490, + 190 + ], + "type": "inline_equation", + "content": "^{1,3,4,5}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 202, + 543, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 202, + 543, + 259 + ], + "spans": [ + { + "bbox": [ + 67, + 202, + 543, + 259 + ], + "type": "text", + "content": "1Sun Yat-Sen University 2ByteDance Seed Vision \n3Guangdong Provincial Key Laboratory of Information Security Technology, China \n4Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China \n5Pazhou Lab (HuangPu), Guangzhou, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 262, + 517, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 262, + 517, + 273 + ], + "spans": [ + { + "bbox": [ + 88, + 262, + 517, + 273 + ], + "type": "text", + "content": "oliveryanzulu@gmail.com, xiaoxuefeng.ailab@bytedance.com, majh8@mail.sysu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 297, + 624 + ], + "type": "text", + "content": "Distribution Matching Distillation (DMD) is a promising score distillation technique that compresses pre-trained teacher diffusion models into efficient one-step or multi-step student generators. Nevertheless, its reliance on the reverse Kullback-Leibler (KL) divergence minimization potentially induces mode collapse (or mode-seeking) in certain applications. To circumvent this inherent drawback, we propose Adversarial Distribution Matching (ADM), a novel framework that leverages diffusion-based discriminators to align the latent predictions between real and fake score estimators for score distillation in an adversarial manner. In the context of extremely challenging one-step distillation, we further improve the pre-trained generator by adversarial distillation with hybrid discriminators in both latent and pixel spaces. Different from the mean squared error used in DMD2 pre-training, our method incorporates the distributional loss on ODE pairs collected from the teacher model, and thus providing a better initialization for score distillation fine-tuning in the next stage. By combining the adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our proposed method achieves superior one-step performance on SDXL compared to DMD2 while consuming less GPU time. Additional experiments that apply multi-step ADM distillation on SD3-Medium, SD3.5-Large, and CogVideoX set a new benchmark towards efficient image and video synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "Recent methods for accelerating diffusion models [12, 17, 63, 65] have concentrated on reducing the sampling steps through distillation. The distillation process generally trains" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 286, + 376, + 405 + ], + "blocks": [ + { + "bbox": [ + 316, + 286, + 376, + 405 + ], + "lines": [ + { + "bbox": [ + 316, + 286, + 376, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 376, + 405 + ], + "type": "image", + "image_path": "ee0a03c5ae49927ae0ea66fe1ac89fedf63a579444d366e3fc72ec2823526b01.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 406, + 368, + 418 + ], + "lines": [ + { + "bbox": [ + 317, + 406, + 368, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 406, + 368, + 418 + ], + "type": "text", + "content": "(a) A white polar bear cub wearing a red sitta in a meadow with flowers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 376, + 286, + 432, + 405 + ], + "blocks": [ + { + "bbox": [ + 376, + 286, + 432, + 405 + ], + "lines": [ + { + "bbox": [ + 376, + 286, + 432, + 405 + ], + "spans": [ + { + "bbox": [ + 376, + 286, + 432, + 405 + ], + "type": "image", + "image_path": "a86eb6eb51f6704986d4549274c849c33d0622bfea3bd6e8472a9badd9d82f97.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 406, + 430, + 423 + ], + "lines": [ + { + "bbox": [ + 377, + 406, + 430, + 423 + ], + "spans": [ + { + "bbox": [ + 377, + 406, + 430, + 423 + ], + "type": "text", + "content": "(b) mage, wizard, female, demoness, forest, D&D, fantasy, incricate, elegant, highly detailed, digital painting, artstation, [...]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 433, + 286, + 493, + 404 + ], + "blocks": [ + { + "bbox": [ + 433, + 286, + 493, + 404 + ], + "lines": [ + { + "bbox": [ + 433, + 286, + 493, + 404 + ], + "spans": [ + { + "bbox": [ + 433, + 286, + 493, + 404 + ], + "type": "image", + "image_path": "516db7b264c7fbb3e87edbf3fd8f759cb41461506d5f4f3c353dbfcb0480f04d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 406, + 491, + 422 + ], + "lines": [ + { + "bbox": [ + 436, + 406, + 491, + 422 + ], + "spans": [ + { + "bbox": [ + 436, + 406, + 491, + 422 + ], + "type": "text", + "content": "(c) beautiful nighttime landscape photography of the Rocky Mountains with a crystal blue lake, milky way galaxy sky, serene, dramatic lightning." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 494, + 286, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 494, + 286, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 494, + 286, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 494, + 286, + 553, + 405 + ], + "type": "image", + "image_path": "36b49336c818ac960121d7ee4716ae2cefa5de355cffd08e566333b26f667651.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 406, + 548, + 419 + ], + "lines": [ + { + "bbox": [ + 495, + 406, + 548, + 419 + ], + "spans": [ + { + "bbox": [ + 495, + 406, + 548, + 419 + ], + "type": "text", + "content": "(d) A planeplane lacavillo glows over the surface of a sand with swirling mist around it." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 424, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 424, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. In these images, some are generated by the baseline SDXL via 50NFE, the others with our DMDX in 1NFE. Can you tell which is the accelerated one? Answers in the footnote" + }, + { + "bbox": [ + 313, + 424, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 313, + 424, + 555, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 465, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 465, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 312, + 465, + 555, + 572 + ], + "type": "text", + "content": "a more efficient generator (a.k.a student model) that approximates the output distribution of the pretrained teacher. This field has observed multiple pathways developing in parallel with progressive distillation [23, 58], consistency distillation [18, 34, 35, 43, 55, 66, 69, 70, 92], score distillation [3, 16, 21, 38, 55, 59, 61, 84-87, 93], rectified flow [27, 28, 30, 71, 81] and adversarial distillation [23, 24, 43, 55, 60, 61, 85, 91]. They are currently appearing as distinct yet non-exclusive research directions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 573, + 556, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 573, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 312, + 573, + 556, + 693 + ], + "type": "text", + "content": "Distribution Matching Distillation (DMD) [85, 86] is a promising score distillation [51] approach, which distills the powerful text-to-image diffusion model SDXL-Base [56] into a one-step generator with great fidelity. While DMD's primary contribution lies in introducing additional regularizers to constrain the distribution matching loss, its predecessor Diff-Instruct [37] pioneered the use of a fake score estimator to approximate the student model's output distribution. This contrasts with the distillation loss in Adversarial Diffusion Distillation (ADD) [61], which directly em" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 206, + 713 + ], + "type": "text", + "content": "*Project Lead. † Corresponding Author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 329, + 702, + 506, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 702, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 329, + 702, + 506, + 713 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 329, + 702, + 506, + 713 + ], + "type": "text", + "content": "Our DMDX (left to right): bottom, top, bottom, top." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16818" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "ploys the student model itself as score estimator. Theoretically, an intrinsic correspondence for DMD and the distillation loss in ADD can be equivalently found between Variational Score Distillation (VSD) [74] and Score Distillation Sampling (SDS) [51] in text-to-3D generation, where SDS represents a specialized instance of VSD through the use of a single-point Dirac distribution as the variational distribution [74]. And thus the VSD loss in DMD performs better than the SDS loss in ADD. Intuitively, since the capacity of a diffusion model decreases significantly when distilled for few-step generations [23], the student model no longer serves as a decent score estimator as the teacher. Therefore, the score-based distillation loss in ADD hardly contributes to its final performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 242, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 295, + 505 + ], + "type": "text", + "content": "However, the optimization of DMD loss relies on the reverse Kullback-Leibler (KL) divergence minimization which is zero-forcing that drives low-probability regions to zero, causing the model to focus on only a few dominant modes and potentially leading to mode collapse [45]. To enhance sample diversity, DMD [86] employs an ODE-based (Ordinary Differential Equation) regularizer with synthetic data, while DMD2 [85] introduces a GAN-based (Generative Adversarial Network [8]) regularizer with real data to counterbalance this side effect. Subsequent efforts such as Moment Matching Distillation (MMD) [59], Score identity Distillation (SiD) [93] and Score Implicit Matching (SIM) [38] all used a variant of Fisher divergence to align the fake score estimator with pre-trained real score estimator [16]. Despite great success, their ability for distribution matching is limited by their dependence on a predefined form of explicit divergence metric. In this work, we would like to raise and investigate a question: Can we bypass the limitations of a predefined divergence by developing a framework that learns an implicit, data-driven discrepancy measure, thereby enabling more flexible and fine-grained matching of complex, high-dimensional distributions?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 508, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 686 + ], + "type": "text", + "content": "Since the multifaceted alignment requirements in complex multimodal text-conditioned image or even video generation may not be fully captured, this motivates our exploration of more adaptive discrepancy learning paradigms. As our first contribution, we show how to align the latent predictions between real and fake score estimators to realize score distillation in an adversarial manner with prior knowledge of teacher model and dynamically learnable parameters, termed Adversarial Distribution Matching (ADM). This is different from the ODE-based or GAN-based regularizer in DMD and DMD2 that is used to counterbalance the mode collapse effect of reverse KL divergence. In contrast, we are performing distribution matching by means of GAN training to replace and circumvent the use of DMD loss, which we will provide more discussion in Sec. 4.1.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our second contribution concerns about the extremely challenging one-step distillation, where we observe that the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "score distillation has a higher risk of gradient exploding and vanishing. We attribute the issue more to the lower overlap of support sets between the student and teacher distributions, not just the approximation errors in the fake score estimator as attributed in DMD2 [85]. In other words, while score distillation yields superior generation quality, it imposes higher requirements on initialization especially when distilled in extremely few steps, which we will provide more in-depth analysis in Sec. 4.3.2. Although we notice that an ODE-based pre-training on synthetic data is used for SDXL one-step distillation in DMD2 implementation, the mean squared error loss was probably still not enough to provide more overlapping regions of support sets between the student and teacher distributions. Our experiments demonstrate that when providing a better initialization for distribution matching, the effect of Two Time-scale Update Rule (TTUR) [2] to the final performance is very limited." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 275, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 407 + ], + "type": "text", + "content": "Naturally, our third contribution focuses on providing a better initialization for further score-based fine-tuning. We employ adversarial distillation to trade off sample quality and mode converge inspired by SDXL-Lightning [23] and LADD [60]. With this distribution-level loss optimization, we can pre-train the student model to capture more potential modes of the teacher model distribution, especially through a hybrid discriminator in both latent and pixel spaces which we will introduce later in Sec. 4.2. To facilitate the diversity, we also propose employing a cubic timestep schedule for the generator to bias towards higher noise levels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 407, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 515 + ], + "type": "text", + "content": "By combining adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our one-step SDXL provides competitive fidelity compared to the baseline with " + }, + { + "bbox": [ + 313, + 407, + 553, + 515 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 313, + 407, + 553, + 515 + ], + "type": "text", + "content": " acceleration in the Number of Function Evaluation (NFE) as shown in Fig. 1. More experiments on multi-step ADM distillation across the best existing diffusion models including SD3-Medium, SD3.5-Large [6], and CogVideoX [83] consistently achieve new benchmarks for efficient image and video synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 525, + 400, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 400, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 400, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 641 + ], + "type": "text", + "content": "Progressive Distillation was proposed in [58] to distill multi-step prediction into one-step prediction of the same distance along the trajectory. SDXL-Lightning [23] extended this idea by utilizing GAN training and achieved one-step high-resolution (1024px) generation for the first time. This approach involving multi-stage process can be quite cumbersome, as it requires iteratively distilling from its predecessor, halving the sampling steps each time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Consistency Distillation proposed [7, 34, 35, 64, 66] to enforce the consistency property into diffusion model, i.e. predictions towards original sample are consistent for arbitrary pairs of noisy timesteps belong to the same trajectory. Subsequent efforts extended it into the trajectory consistency to relax the training objectives, i.e. predictions towards noisy" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16819" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "samples of arbitrary subsequent timesteps are consistent, including CTM [18], TCD [92], TSCD [55] and PCM [69]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 191 + ], + "type": "text", + "content": "Rectified Flow [27, 28] aims to obtain faster straight trajectories through multiple reflow processes that iteratively learn the velocity of many ODE pairs from its predecessor. PeRFlow [81] tried splitting the trajectory into pieces and applying piece-wise rectification with real data samples. In this paper, we empirically found the straightness can also be satisfied via an adversarial distillation paradigm without the need for repeatedly collecting tremendous synthetic data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 335 + ], + "type": "text", + "content": "Score Distillation intuitively tries to keep that a sample appearing in the student model distribution at a specific noise level is with the same probability as it does in the teacher model distribution. In light of the motivation that relying on a single divergence might be problematic, a concurrent work Score-of-Mixture Distillation (SMD) [16] shares our perspective and explicitly designed a class of " + }, + { + "bbox": [ + 55, + 192, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 192, + 296, + 335 + ], + "type": "text", + "content": "-skew Jensen-Shannon (JS) divergences for optimization. In contrast, we implicitly measure the discrepancy between the fake and real score estimators in an adversarial manner, facilitating a more capable and adaptive discrepancy learning paradigm along the distillation process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 442 + ], + "type": "text", + "content": "Adversarial Distillation employs a discriminator to align the student model distribution with a specific target distribution [40, 94]. While methods UFOGen [79], DMD2 [85], and APT [24] directly align with the real data distribution, SDXL-Lightning [23] and Hyper-SD [55] utilize intermediate timestep predictions from the teacher model as approximation objectives. Inspired by LADD [60], our adversarial distillation pre-training aligns with ODE-based synthetic data generated from the teacher model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 539 + ], + "type": "text", + "content": "Human Feedback Learning was firstly considered by ReFL [78] in diffusion models that includes two stages of reward model training and preference fine-tuning. HyperSD [55] treated this as a standalone technique, ultimately applying LoRA insertion to shift the output distribution of the low-step generator toward human preferences. Subsequent studies, notably [36, 39], further attempted to integrate CFG and Score-based divergence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 548, + 140, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 140, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 140, + 560 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 154, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 154, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 154, + 580 + ], + "type": "text", + "content": "3.1. Diffusion Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "content": "Given a training data distribution " + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}" + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "content": " with standard deviation " + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{data}}" + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "content": ", the diffusion model [12] generates samples by reversing the forward diffusion process that progressively adds noise to a data sample " + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "inline_equation", + "content": "x_0 \\sim p_{\\mathrm{data}}" + }, + { + "bbox": [ + 55, + 586, + 295, + 634 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 637, + 294, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 637, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 121, + 637, + 294, + 651 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x}\\right) \\sim \\mathcal {N} \\left(\\alpha_ {t} \\boldsymbol {x}, \\sigma_ {t} ^ {2} \\boldsymbol {I}\\right), \\tag {1}", + "image_path": "62c0a24dbf4689878740a0dd630a3c2ce377cda09bb2bce687676d5440964541.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\geq 0, \\sigma_{t} > 0" + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": " are specified noise schedules such that " + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\alpha_{t} / \\sigma_{t}" + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": " satisfies monotonically decreasing w.r.t. " + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": " and larger " + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": " indicates greater noise. We consider two different formulations for denoising models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "DDPM and DDIM [12, 63] assume discrete-time schedules" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "t \\in [1, T]" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " (typically " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ") and noise prediction parameterization* [58]. The training objective is given by," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 99, + 553, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 99, + 553, + 113 + ], + "spans": [ + { + "bbox": [ + 356, + 99, + 553, + 113 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} ], \\tag {2}", + "image_path": "182a3f7f4bf308b191089587c5f79761de92ee425ae0cbe51ff2ae208f866048.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": " is a weighting function and " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": " is a neural network with parameters " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": ". The noise schedule is defined as " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\sqrt{\\bar{\\alpha}_{t}}, \\sigma_{t} = \\sqrt{1 - \\bar{\\alpha}_{t}}" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t = \\sqrt{\\bar{\\alpha}_t} \\boldsymbol{x}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": ". For DDIM sampling, it solves the Probability-Flow Ordinary Differential Equations (PF-ODE) [65] by " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "d\\bar{\\boldsymbol{x}}_t = \\epsilon_\\theta (\\frac{\\bar{\\boldsymbol{x}}_t}{\\sqrt{\\bar{\\sigma}_t^2 + 1}}) d\\bar{\\sigma}_t" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\bar{\\boldsymbol{x}}_t = \\frac{\\boldsymbol{x}_t}{\\sqrt{\\bar{\\alpha}_t}}" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\bar{\\sigma}_t = \\sqrt{\\frac{1 - \\bar{\\alpha}_t}{\\bar{\\alpha}_t}}" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": ", starting from " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T \\sim \\mathcal{N}(\\boldsymbol{0}, \\boldsymbol{I})" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": " and stopping at " + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 313, + 116, + 555, + 208 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": "Flow Matching [26, 27, 30] uses velocity prediction parameterization [58] and continuous-time coefficients (typically " + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - t" + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\sigma_{t} = t" + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "inline_equation", + "content": "t \\in [0, T = 1]" + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": "). The conditional probability path or the velocity we say can be given as " + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{t} = \\frac{d\\alpha_{t}}{dt}\\pmb{x}_{0} + \\frac{d\\sigma_{t}}{dt}\\epsilon" + }, + { + "bbox": [ + 313, + 209, + 554, + 270 + ], + "type": "text", + "content": ", such that the training objective is," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 272, + 553, + 286 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 272, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 353, + 272, + 553, + 286 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\boldsymbol {x} _ {0}, t, \\epsilon \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I})} [ w (t) \\| \\boldsymbol {v} _ {\\theta} (\\boldsymbol {x} _ {t}, t) - \\boldsymbol {v} _ {t} \\| _ {2} ^ {2} ], \\tag {3}", + "image_path": "bd67511571fa5a7326c0000c74571fe1561bf03aea05185bd904457c56509bbf.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": " is a weighting function and " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{\\theta}" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": " is a neural network parameterized by " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": ". The sampling procedure starts from " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "\\pmb{x}_T \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I})" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": " and stops at " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": ", solving the PF-ODE by " + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "inline_equation", + "content": "d\\pmb{x}_t = \\pmb{v}_{\\theta}(\\pmb{x}_t, t)dt" + }, + { + "bbox": [ + 313, + 289, + 554, + 337 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 346, + 496, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 496, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 496, + 358 + ], + "type": "text", + "content": "3.2. Distribution Matching Distillation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": "DMD [85, 86] distills pretrained diffusion models " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\phi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " into one-step or multi-step efficient generators " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "G_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " by minimizing the reverse KL divergence between the target distribution " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " and the efficient generator output distribution " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": ". The gradient of DMD objective w.r.t. " + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 363, + 554, + 423 + ], + "type": "text", + "content": " is," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 426, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {D M D}} = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} - \\left[ \\left(s _ {\\text {r e a l}} \\left(\\boldsymbol {x} _ {t}\\right) - s _ {\\text {f a k e}} \\left(\\boldsymbol {x} _ {t}\\right)\\right) \\frac {d \\boldsymbol {G} _ {\\theta} \\left(\\boldsymbol {z} , t ^ {\\prime}\\right)}{d \\theta} \\right], \\tag {4}", + "image_path": "819201087a17facf9114e25d188a94086afa5d9d28c2699d4c6f8c8da5ce79ac.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "z \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": " is randomly selected from predefined generator schedule, " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "t \\sim \\mathcal{U}(0,T)" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": ", and noisy samples " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_t = \\boldsymbol{q}(\\boldsymbol{x}_t|\\hat{\\boldsymbol{x}}_0)" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": " are obtained by randomly diffusing the generator output " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}}_0 = \\boldsymbol{G}_{\\theta}(\\boldsymbol{z},t')" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": ". Score functions " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{real}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{real}}(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{fake}}(\\boldsymbol{x}_t) = \\nabla_{\\boldsymbol{x}_t}\\log p_{\\mathrm{fake}}(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": " are vector fields that point towards higher density of data at a given noise level [17, 65] for " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 450, + 554, + 534 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "content": "While the real score estimator is the teacher model " + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\phi}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "content": " itself, the fake score estimator " + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\psi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "content": " is initialized the same as " + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{F}_{\\phi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "content": " and dynamically learned to describe " + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 313, + 535, + 554, + 594 + ], + "type": "text", + "content": " with pretrain loss as Eqs. (2) and (3). In practice, the gradient in Eq. (4) is computed as," + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 349, + 597, + 553, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 597, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 349, + 597, + 553, + 623 + ], + "type": "interline_equation", + "content": "\\operatorname {g r a d} \\left(\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t\\right) = \\frac {\\boldsymbol {f} _ {\\psi} \\left(\\boldsymbol {x} _ {t} , t\\right) - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right)}{\\left\\| \\hat {\\boldsymbol {x}} _ {0} - \\boldsymbol {F} _ {\\phi} \\left(\\boldsymbol {x} _ {t} , t\\right) \\right\\| _ {1}} \\tag {5}", + "image_path": "f0ea6349aec80f1c6c3a258033573cd770317a0d44c874bbc6d8dbbdfd07df58.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 625, + 503, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 503, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 503, + 637 + ], + "type": "text", + "content": "such that the training loss is implemented like," + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 640, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 640, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 320, + 640, + 553, + 659 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D M D}} (\\theta) = \\underset {\\boldsymbol {z}, t ^ {\\prime}, t, \\boldsymbol {x} _ {t}} {\\mathbb {E}} [ \\| \\hat {\\boldsymbol {x}} _ {0} - s g (\\hat {\\boldsymbol {x}} _ {0} - \\operatorname {g r a d} (\\hat {\\boldsymbol {x}} _ {0}, \\boldsymbol {x} _ {t}, t)) \\| _ {2} ^ {2} ], \\tag {6}", + "image_path": "6fd62ee256bf117da980b3842bf80ea051c8e6ec723a22ce4e911f95a5445647.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 663, + 509, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 509, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 509, + 675 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 663, + 509, + 675 + ], + "type": "inline_equation", + "content": "sg(\\cdot)" + }, + { + "bbox": [ + 313, + 663, + 509, + 675 + ], + "type": "text", + "content": " denotes the stop gradient operation." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": "*This is not mandatory for works like CogVideoX [83] which also used DDIM sampling but parameterized in velocity. The different parameterizations can be equivalently transformed into each other [58]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16820" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 85, + 517, + 181 + ], + "blocks": [ + { + "bbox": [ + 290, + 75, + 314, + 83 + ], + "lines": [ + { + "bbox": [ + 290, + 75, + 314, + 83 + ], + "spans": [ + { + "bbox": [ + 290, + 75, + 314, + 83 + ], + "type": "text", + "content": "ADM" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 85, + 517, + 181 + ], + "lines": [ + { + "bbox": [ + 88, + 85, + 517, + 181 + ], + "spans": [ + { + "bbox": [ + 88, + 85, + 517, + 181 + ], + "type": "image", + "image_path": "e44a13f3cc8690a0fafb9e59c1881a073abc346d726524c940fcb433605a9033.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 88, + 194, + 517, + 352 + ], + "blocks": [ + { + "bbox": [ + 290, + 184, + 313, + 193 + ], + "lines": [ + { + "bbox": [ + 290, + 184, + 313, + 193 + ], + "spans": [ + { + "bbox": [ + 290, + 184, + 313, + 193 + ], + "type": "text", + "content": "ADP" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 194, + 517, + 352 + ], + "lines": [ + { + "bbox": [ + 88, + 194, + 517, + 352 + ], + "spans": [ + { + "bbox": [ + 88, + 194, + 517, + 352 + ], + "type": "image", + "image_path": "51df300708192b09c7df88f42838958f4f8e97d8e6a1c9cc8c73c573f7edf2be.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 353, + 548, + 365 + ], + "lines": [ + { + "bbox": [ + 60, + 353, + 548, + 365 + ], + "spans": [ + { + "bbox": [ + 60, + 353, + 548, + 365 + ], + "type": "text", + "content": "Figure 2. Overall pipeline of our proposed Adversarial Distribution Matching (ADM) and Adversarial Distillation Pre-training (ADP)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 138, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 138, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 138, + 387 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 395, + 241, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 241, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 241, + 407 + ], + "type": "text", + "content": "4.1. Adversarial Distribution Matching" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": "Instead of using a predefined divergence between the fake and real distributions, we use an implicit, data-driven discrepancy measure through an adversarial discriminator. Specifically, our discriminator " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "D_{\\tau}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": " consists of a frozen latent diffusion model initialized the same as teacher model " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "F_{\\phi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": " and multiple trainable heads added upon different UNet [57] or DiT [49] blocks. Given the noisy sample " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\pmb{q}(\\pmb{x}_t|\\hat{\\pmb{x}}_0)" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": " diffused from the output of few-step generator " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0 = \\pmb{G}_{\\theta}(\\pmb{z},t')" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": ", the score estimators no longer solve the PF-ODE w.r.t. the end point " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{\\mathrm{fake}} = \\pmb{f}_{\\psi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0^{\\mathrm{real}} = \\pmb{F}_{\\phi}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 415, + 296, + 546 + ], + "type": "text", + "content": " as used in Eq. (5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": "Instead, we set a fixed timestep interval " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " (defaults to " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "T / 64" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": ") and solve the PF-ODE w.r.t. " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "(t - \\Delta t)" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": ", such that the fake sample " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " and real sample " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " can be obtained to serve as score predictions and sent into the discriminator. The discriminator hierarchically aggregates features from frozen backbone layers and dynamically weights them through multiple learnable heads, establishing an adaptive discrepancy metric that leverages both diffusion priors and data-driven trainable dynamics. We use Hinge loss [22] to train the generator " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{G}_{\\theta}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " and discriminator " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "D_{\\tau}(\\pmb{x}_t,t)" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " alternatively. This encourages the fake score prediction " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\Delta t}^{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": " to be closer to the real score prediction " + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t - \\Delta t}^{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 548, + 296, + 693 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 102, + 695, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 695, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 102, + 695, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}} {\\mathbb {E}} \\left[ - \\boldsymbol {D} _ {\\tau} \\left(\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t\\right) \\right] \\tag {7}", + "image_path": "7a0fb0ff41772c3600f60656a6a7eb9678695064034c49e8c5750053afb827dc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 371, + 554, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 371, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 326, + 371, + 554, + 406 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} (\\tau) = \\underset {\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, \\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {r e a l}}} {\\mathbb {E}} [ \\max (0, 1 + D _ {\\tau} (\\boldsymbol {x} _ {t - \\Delta t} ^ {\\text {f a k e}}, t - \\Delta t)) \\tag {8} \\\\ \\left. + \\max \\left(0, 1 - D _ {\\tau} \\left(x _ {t - \\Delta t} ^ {\\text {r e a l}}, t - \\Delta t\\right)\\right) \\right] \\\\ \\end{array}", + "image_path": "f46fd21a9034ad77b17f214bf9181a30e3b083b5090df9a98d65b65e42d2ae6e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 413, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 449 + ], + "type": "text", + "content": "In conjunction with the dynamically learned fake model, we clarify our training procedure in Appendix A. The overall pipeline is demonstrated in Fig. 2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 457, + 541, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 541, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 541, + 470 + ], + "type": "text", + "content": "4.1.1. Motivation of Discriminator Timestep " + }, + { + "bbox": [ + 313, + 457, + 541, + 470 + ], + "type": "inline_equation", + "content": "(t - \\Delta t)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 570 + ], + "type": "text", + "content": "Given that the final objective of score distillation is to match the probability flows varying with noise levels of student and teacher models to exactly the same, timestep information must be considered when measuring the discrepancy between distributions. This coincides with our discriminator design that uses a pre-trained diffusion model, and we take a small step alongside the PF-ODE, succeeding in preserving the input timestep information of score estimators." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 422, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 422, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 422, + 589 + ], + "type": "text", + "content": "4.1.2. Data-driven Effect" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "The flexibility of using discriminators for distributional discrepancy measure is not only in the noise level of score function, but also within the distillation process. As distillation iterates, the model is exposed to increasingly diverse data, causing the discrepancies in modes between the two distributions to change. In the early phase of training where the discrepancy is significant, a more global evaluation is necessary, whereas in the later phases, when the discrepancy becomes minor, a more localized, fine-grained optimization might become essential. In other words, driven" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16821" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 293, + 151 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 293, + 151 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 293, + 151 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 293, + 151 + ], + "type": "image", + "image_path": "45d0869ca4553ec88ba79e6c155f4aa6a56f3aa919ecde8d7203fb67071d8a89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 185 + ], + "type": "text", + "content": "Figure 3. Changes of DMD loss over multi-step ADM distillation for CogVideoX. Note that we did not optimize this objective directly during ADM distillation but recorded it over iterations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 215 + ], + "type": "text", + "content": "by the volume of data, the divergence measure employed across different training phases can vary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 220, + 236, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 236, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 236, + 232 + ], + "type": "text", + "content": "4.1.3. Relationship with DMD and DMD2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 235, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 389 + ], + "type": "text", + "content": "To mitigate the mode collapse issue in DMD loss, an ODE-based regularizer and a GAN-based regularizer are additionally used for distillation in DMD [85] and DMD2 [85], respectively. However, these two regularizers do not fundamentally address the mode-seeking behavior introduced by reverse KL divergence as shown in Fig. 4(a), but rather counterbalance it by trade-offs between losses. In ADM, our adversarial loss is actually playing the role of DMD loss to realize score distillation with an implicit, data-driven discrepancy measure instead of a predefined divergence. Therefore, our motivation for using GAN training in ADM is different from that of DMD2 [85] and we don't require additional regularizers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 498 + ], + "type": "text", + "content": "Intuitively, the learnable discriminator can approximate any nonlinear function to implicitly measure distribution divergence, which probably inherently encompasses the reverse KL divergence in DMD loss. As shown in Fig. 3, we visualize the changes of DMD loss in Eq. (6) during the multi-step ADM distillation on CogVideoX [83]. Though not directly optimize on Eq. (6), the results indicate a very steady downward trend that supports our assumption. More theoretical discussion is provided in Sec. 4.3.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 249, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 249, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 249, + 517 + ], + "type": "text", + "content": "4.2. Adversarial Distillation Pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 630 + ], + "type": "text", + "content": "To stabilize the extremely difficult one-step distillation, we opt to provide a better initialization for ADM fine-tuning with adversarial distillation pre-training on synthetic data. Our pre-training configuration refers to Rectified Flow [27] in several aspects, where we 1) collect the ODE pairs from teacher model in an offline manner, 2) construct noisy samples by linearly interpolation between pure noise and clean data sample of the ODE pair, 3) alter the prediction target of the generator to the velocity of ODE pair." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "content": "As for adversarial training, we formulate a latent-space discriminator " + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\tau_1}(\\boldsymbol{x}_t,t)" + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "content": " initialized from teacher model and a pixel-space discriminator " + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{\\tau_2}(\\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 631, + 296, + 714 + ], + "type": "text", + "content": " initialized from the vision encoder of SAM [19] model, respectively, as shown in Fig. 2. We also append multiple trainable heads to both the backbone networks similar to the practice in ADM. All these contribute to increasing the discriminative capability," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "facilitating the student model in discovering more potential modes in the teacher model distribution. Specifically, let " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_0 = G_\\theta (\\pmb {x}_t,t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " denote the predicted PF-ODE endpoint of generator, where " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " represents the noisy sample interpolated between an ODE pair " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "(\\pmb {x}_T,\\pmb {x}_0)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " with random timestep " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ". For latent-space discriminator, we diffuse the generator output with another random noise and timestep " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "t^{\\prime}\\in (0,T]" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ", getting " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_{t^{\\prime}} = \\pmb {q}(\\tilde{\\pmb{x}}_{t^{\\prime}}|\\tilde{\\pmb{x}}_0)" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " as its input. For pixel-space discriminator, the generator output will be first decoded via VAE decoder and then sent into the vision encoder. The training objective, also based on Hinge loss [22], encourages the generator output " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " to be closer to the synthetic data sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 230, + 555, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 230, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 349, + 230, + 555, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G A N}} (\\theta) = \\underset {\\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} - \\left[ \\lambda_ {1} D _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) + \\lambda_ {2} D _ {\\tau_ {2}} \\left(\\tilde {\\boldsymbol {x}} _ {0}\\right) \\right] \\tag {9}", + "image_path": "cadf611e1267420ac5d4f22bf7d90a4d53386e45e5ac8c74b50a334c3ef90122.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 256, + 553, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 256, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 335, + 256, + 553, + 311 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {G A N}} \\left(\\tau_ {1}, \\tau_ {2}\\right) = \\underset {\\boldsymbol {x} _ {0}, \\tilde {\\boldsymbol {x}} _ {0}, t ^ {\\prime}} {\\mathbb {E}} \\left[ \\lambda_ {1} \\cdot \\max \\left(0, 1 + \\boldsymbol {D} _ {\\tau_ {1}} \\left(\\tilde {\\boldsymbol {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right. \\\\ + \\lambda_ {2} \\cdot \\max \\left(0, 1 + D _ {\\tau_ {2}} \\left(\\tilde {x} _ {0}\\right)\\right) \\tag {10} \\\\ + \\lambda_ {1} \\cdot \\max (0, 1 - D _ {\\tau_ {1}} \\left(\\boldsymbol {x} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)) \\\\ \\left. \\left. + \\lambda_ {2} \\cdot \\max \\left(0, 1 - D _ {\\tau_ {2}} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right] \\right. \\\\ \\end{array}", + "image_path": "6db873953e389d7094f86e2fe97cf8097f2e28b841d9bdc2c7dec02d011b3cba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 323, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 346 + ], + "type": "text", + "content": "We empirically found that setting balancing coefficients " + }, + { + "bbox": [ + 313, + 323, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.85, \\lambda_{2} = 0.15" + }, + { + "bbox": [ + 313, + 323, + 555, + 346 + ], + "type": "text", + "content": " produces visually coherent results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 360, + 497, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 497, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 497, + 372 + ], + "type": "text", + "content": "4.2.1. Cubic Generator Timestep Schedule" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": "Intuitively, higher noise levels encourage exploration of new modes by weakening the restrictive information encoded in latent representations. Therefore, we propose employing a cubic timestep schedule for the generator. The schedule maps uniform " + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "inline_equation", + "content": "[0,T)" + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": " samples through " + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "inline_equation", + "content": "[1 - (t / T)^3 ]*T" + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": ", non-linearly concentrating values near " + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": " with heavy noise similar to LADD [60]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 523, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 523, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 523, + 487 + ], + "type": "text", + "content": "4.2.2. Uniform Discriminator Timestep Schedule" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "text", + "content": "Unlike in ADM where the discriminator is firstly used for score distillation, the utilization of latent-space discriminator is common for adversarial distillation works. Inspired by SDXL-Lightning [23], where they found that the diffusion encoder is trained to focus on high-frequency details at lower timesteps and low-frequency structures at higher timesteps, we set a uniform " + }, + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "inline_equation", + "content": "(0,T]" + }, + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "text", + "content": " for discriminator timestep " + }, + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 491, + 555, + 588 + ], + "type": "text", + "content": " to capture both advantages during pre-training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 449, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 449, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 449, + 612 + ], + "type": "text", + "content": "4.2.3. Relationship with LADD" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Our motivation to perform adversarial distillation on synthetic data is inspired by LADD [60], but differs in a lot that we 1) construct noisy samples via ODE pairs in the style of Rectified Flow [27] rather than random noise, 2) develop a cubic generator timestep schedule that facilitate deterministic Euler sampling rather than consistency ones, 3) introduce an additional pixel-space encoder to increase the capability of discriminator and find more modes." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16822" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 127, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 127, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 127, + 83 + ], + "type": "text", + "content": "4.3. Discussion" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 234, + 100 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 234, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 234, + 100 + ], + "type": "text", + "content": "4.3.1. Difference between ADM and ADP" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "text", + "content": "One question may appear like what is the difference between these two adversarial approaches with latent-space discriminators? The effectiveness in score distillation correlates with the fact that score function " + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "inline_equation", + "content": "\\nabla_{\\pmb{x}}\\log p(\\pmb{x};\\sigma (t))" + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "text", + "content": " is defined at different noise levels " + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "inline_equation", + "content": "\\sigma (t)" + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "text", + "content": ". In contrast, adversarial distillation only aligns the distribution of clean data samples when " + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 54, + 105, + 295, + 272 + ], + "type": "text", + "content": ". While the latent-space discriminator in pre-training is capturing information at different scales and details by randomly diffusing the generator output, the crux for ADM is more than that by solving the PF-ODE of both score estimators. In other words, ADM additionally supervises the complete denoising process through noisy samples that are with higher density at different noise levels of respective distribution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 273, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 416 + ], + "type": "text", + "content": "This leads to a situation in ADM when two distributions initialized with less overlap in support sets, the noisy samples remain in regions unfamiliar to each other, and the discriminator can easily distinguish between them thus leading to extreme gradient signals. However, since Gaussian noise is isotropic, we artificially create overlapping regions for the randomly diffused samples in ADP to make the discrimination more difficult, resulting in relatively smooth gradients. Therefore, our ADM still falls under score distillation given that it encourages the entire probability flows to be closer, while the pre-training belongs to adversarial distillation because it only cares about the clean data distribution at " + }, + { + "bbox": [ + 55, + 273, + 296, + 416 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 273, + 296, + 416 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 200, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 200, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 200, + 435 + ], + "type": "text", + "content": "4.3.2. Importance of Pre-training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 475 + ], + "type": "text", + "content": "Another question we haven't discussed so far is why do we need pre-training for one-step score distillation? Taking the reverse KL divergence used by DMD loss as an example:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 477, + 295, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 477, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 73, + 477, + 295, + 503 + ], + "type": "interline_equation", + "content": "\\mathbb {D} _ {\\mathrm {K L}} \\left(p _ {\\text {f a k e}} \\| p _ {\\text {r e a l}}\\right) = \\int p _ {\\text {f a k e}} (\\boldsymbol {x}) \\log \\frac {p _ {\\text {f a k e}} (\\boldsymbol {x})}{p _ {\\text {r e a l}} (\\boldsymbol {x})} d \\boldsymbol {x}. \\tag {11}", + "image_path": "01f6a5399b8dee98457cb2be2869332c5df3f294f4912433f5affee38b3c61e8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": "When employing one-step distillation, the generator output is worse in visual fidelity and structural integrity compared to multi-step sampling, leading " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x}) \\to 0" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}(\\pmb{x}) > 0" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " in regions. The integrand " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{p_{\\mathrm{real}}(\\pmb{x})}" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " approaching zero " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "0 \\cdot (-\\infty)" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " causes the optimization to avoid regions where " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}(\\pmb{x}) > 0" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " has negligible density, a phenomenon called zero-forcing. Instead of fully covering " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": "'s support, " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " collapses to a subset of modes of " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": ", inducing mode-seeking behavior as illustrated in Fig. 4(a). During training this manifests itself as gradient vanishing sometimes. And conversely, the fuzzy samples that the diffusion model typically produces through one step are also not within the teacher model distribution, yielding " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}(\\pmb{x}) \\to 0" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x}) > 0" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " in regions and the integrand " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x}) \\log \\frac{p_{\\mathrm{fake}}(\\pmb{x})}{0}" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": " diverging to " + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 55, + 505, + 295, + 701 + ], + "type": "text", + "content": ", and thus resulting in numerical instability and gradient exploding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "content": "Similarly, when the support sets of the student and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 555, + 284 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 284 + ], + "type": "image", + "image_path": "562e6a4f4d7a2a7a42b3338e938721c3c930037b14d245b68b42d0041bb47780.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 286, + 520, + 297 + ], + "lines": [ + { + "bbox": [ + 348, + 286, + 520, + 297 + ], + "spans": [ + { + "bbox": [ + 348, + 286, + 520, + 297 + ], + "type": "text", + "content": "Figure 4. Illustration for theoretical discussion." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "text", + "content": "teacher distributions overlap hardly at all, forward KL divergence approaches " + }, + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "inline_equation", + "content": "+\\infty" + }, + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\boldsymbol{x}) > 0" + }, + { + "bbox": [ + 313, + 311, + 555, + 396 + ], + "type": "text", + "content": ", while JS divergence saturates to a constant log 2 and Fisher divergence may degenerate without definition. Therefore, when this assumption is undermined, many of the single divergences no longer apply and a better initialization with more overlapping regions becomes essential as exemplified in Fig. 4(b)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 404, + 435, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 435, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 435, + 417 + ], + "type": "text", + "content": "4.3.3. Theoretical Objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 421, + 555, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 457 + ], + "type": "text", + "content": "The last question is why ADM is better than DMD loss theoretically? In fact, the Hinge GAN [22] we use has been proven to minimize Total-Variation Distance (TVD) [68]:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 460, + 554, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 460, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 335, + 460, + 554, + 485 + ], + "type": "interline_equation", + "content": "T V \\left(p _ {\\text {f a k e}}, p _ {\\text {r e a l}}\\right) = \\int \\left| p _ {\\text {f a k e}} (\\boldsymbol {x}) - p _ {\\text {r e a l}} (\\boldsymbol {x}) \\right| d \\boldsymbol {x} \\tag {12}", + "image_path": "bca5c2c9840e21e5d99a8d788192af05b7b6b7e8afb209ecbd01957c62197808.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "That is to say, when the discriminator is sufficiently rich and well trained, the theoretical optimum of Hinge loss upon convergence minimizes TVD. When the support sets of fake and real distributions have minimal overlap, TVD provides two key advantages over reverse KL divergence: 1) Symmetry: TVD yields the same discrepancy measure regardless of the initial distribution, whereas the asymmetric reverse KL may exhibit mode-seeking behavior and neglect other portions of the overall distribution. For example, as illustrated in Fig. 4(c), TVD maintains substantial loss values and provides optimization directions covering the mode when " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{fake}}(\\pmb{x}) \\to 0" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " while " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{real}}(\\pmb{x}) > 0" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ", whereas the reverse KL divergence suffers from gradient vanishing issues in such scenarios, as discussed in Sec. 4.3.2. 2) Boundedness: TVD is bounded within [0,1], and thus mitigating disruptions from outliers during training, especially in our high-dimensional multimodal text-conditioned image and video distributions, avoiding the numerical instability inherent to reverse KL divergence caused by gradient exploding." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16823" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 293, + 161 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 293, + 161 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 293, + 161 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 293, + 161 + ], + "type": "table", + "html": "
MethodStep NFECLIP ScorePick ScoreHPSv2MPS
ADD [61] (512px)1135.008822.152427.097110.4340
LCM [34]1228.466920.126723.82464.8134
Lightning [23]1133.498521.919427.155710.2285
DMD2 [85]1135.215322.097827.452310.6947
DMDX (Ours)1135.255722.273627.704611.1978
SDXL-Base [56]255035.030922.249427.374310.7042
", + "image_path": "891ff9a75ab171bc5ebd6a3e8d67cf8292b85ab3f52c8faba05c8a0054b19910.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 182, + 293, + 316 + ], + "blocks": [ + { + "bbox": [ + 63, + 164, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 63, + 164, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 63, + 164, + 287, + 175 + ], + "type": "text", + "content": "Table 1. Quantitative results on fully fine-tuning SDXL-Base." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 182, + 293, + 316 + ], + "lines": [ + { + "bbox": [ + 57, + 182, + 293, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 182, + 293, + 316 + ], + "type": "table", + "html": "
MethodStepNFECLIP ScorePick ScoreHPSv2MPS
TSCD [61]4834.018521.966527.272810.8600
PCM [69] (Shift=1)4433.504221.970327.368010.5707
PCM [69] (Shift=3)4433.381821.939627.114610.5635
PCM [69] (Stoch.)4433.418521.882227.317710.5200
Flash [3]4434.397822.090427.258610.6634
ADM (Ours)4434.907622.547128.449211.9543
SD3-Medium [6]255034.763322.296127.973311.3652
LADD [60]4434.739522.395827.492311.4372
ADM (Ours)4434.973022.884227.733112.2350
SD3.5-Large [6]255034.966822.508727.968811.5826
", + "image_path": "f1d568af66c55eb0b1b112c6111a81be75b732fc9af8c84c6b4ae156801688c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 347, + 295, + 462 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 342 + ], + "type": "text", + "content": "Table 2. Quantitative results on LoRA fine-tuning SD3-Medium and fully fine-tuning SD3.5-Large." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 347, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 57, + 347, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 57, + 347, + 295, + 462 + ], + "type": "table", + "html": "
MethodStepNFEFinal ScoreQuality ScoreSemantic Score
ADM8878.58480.82569.621
+ Longer Training ×28880.76483.03171.693
ADM w/ CFG81679.86580.93875.569
+ Longer Training ×281681.79683.00876.947
CogVideoX-2b [83]10020080.03680.80176.974
ADM8882.06783.22777.423
ADM w/ CFG81680.98282.16576.251
CogVideoX-5b [83]10020081.22681.78578.987
", + "image_path": "699efdce4f2c37a22ff69b58fc9ec44afa7f59dfc48a523cb40139d8b4c5e9e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 466, + 287, + 477 + ], + "lines": [ + { + "bbox": [ + 63, + 466, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 63, + 466, + 287, + 477 + ], + "type": "text", + "content": "Table 3. Quantitative results on fully fine-tuning CogVideoX." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 487, + 137, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 137, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 137, + 500 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 628 + ], + "type": "text", + "content": "Models. For one-step distillation, we employ Adversarial Distillation Pre-training (ADP) and ADM fine-tuning on SDXL-Base [56], termed DMDX. For multi-step distillation, we only employ ADM training on both text-to-image models SD3-Medium, SD3.5-Large [6], and text-to-video models CogVideoX-2b, CogVideoX-5b [83]. Following most concurrent works, we didn't use Classifier-Free Guidance (CFG) [11] in our text-to-image model. We tried conducting CFG-integrated experiments on the text-to-video model, with the approach detailed in Sec. 5.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 700 + ], + "type": "text", + "content": "Datasets. No visual data is required for both ADP and ADM proposed in this work. For image generators, we utilize the text prompts from JourneyDB [67] that exhibits high-level of detail and specificity for training. For video generators, we collect training prompts from OpenVid-1M [47], Vript [82] and Open-Sora-Plan-v1.1.0 [50]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Evaluation. The image generators are evaluated on 10K" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 356 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 356 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 356 + ], + "type": "image", + "image_path": "242459914feae3653c71bb6d02d3a2a9f2dd3825aa50466cd4eb0e0f04af7879.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 358, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 321, + 358, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 321, + 358, + 545, + 370 + ], + "type": "text", + "content": "Figure 5. Qualitative results on fully fine-tuning SDXL-Base." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 529 + ], + "type": "text", + "content": "prompts from COCO 2014 [25] following DMD2 [85]. We report the CLIP score [52] and human preference benchmarks PickScore [20], HPSv2 [75] and MPS [90] as many concurrent works including Hyper-SD [55] and Emu3 [73]. But we don't include Hyper-SD in the one-step quantitative comparison because one-step Hyper-SDXL has been optimized directly on human feedback with ReFL [78]. Instead, we compare with the TSCD algorithm proposed therein on SD3-Medium [13], since 4-step Hyper-SD3 LoRA is without ReFL optimization. The video generators are evaluated on VBenchmark [14], which consists of multiple quality and semantic dimensions comprehensively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 556, + 662 + ], + "type": "text", + "content": "Hyperparameters. Despite having multiple models to train in our proposed ADP and ADM, we achieve satisfactory visual fidelity and structural integrity without extensive hyperparameter tuning. For the remainder of this work, we only adjust the learning rate of generator for different experiments. The optimizer settings for discriminators and fake models are uniform across all the experiments. Unless longer training is specifically noted, we train the same for only 8K iterations with a batch size of 128 and 8 for text-to-image and text-to-video models, respectively. More specific implementation details are provided in Appendix B." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 671, + 455, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 455, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 455, + 684 + ], + "type": "text", + "content": "5.1. Efficient Image Synthesis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Tab. 1 quantitatively compares our two-stage approach combing ADP and single-step ADM distillation with ex" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16824" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 288, + 213 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 288, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 288, + 213 + ], + "type": "table", + "html": "
AblationCLIP ScorePick ScoreHPSv2MPS
Ablation on adversarial distillation.
A1: Rectified Flow [27]27.437620.021123.60934.4518
A2: DINOv2 as pixel-space34.183621.875027.103910.2407
A3: λ1 = 0.7, λ2 = 0.333.694321.634426.89029.9633
A4: λ1 = 1.0, λ2 = 0.033.892921.739526.786910.0757
A5: w/o ADM (ADP only)35.772322.009527.349910.6646
Ablation on score distillation.
B1: ADM w/o ADP32.502021.763126.873210.8986
B2: DMD Loss w/o ADP32.748221.034125.96808.8977
B3: DMD Loss w/ ADP34.511921.936627.398510.6046
B4: DMDX (Ours)35.255722.273627.704611.1978
", + "image_path": "14577d22c29c8d9c944921ca95a76d884d47fbce21d3a140faf305c7fe9a7b6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 235, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 317 + ], + "type": "text", + "content": "isting one-step distillation methods on fully fine-tuning SDXL-Base. The results show that our method achieves excellent performance on both image-text alignment and human preference, which is consistent with the qualitative comparisons in Fig. 5, including better portrait aesthetic, animal hair details, subject-background separation and physical structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 379 + ], + "type": "text", + "content": "For multi-step ADM distillation, it can serve as a standalone score distillation method. We tried both fully finetuning and LoRA fine-tuning [13] configurations and the quantitative results in Tab. 2 demonstrate our superior performance. Qualitative results are provided in Appendix C." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 194, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 194, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 194, + 396 + ], + "type": "text", + "content": "5.2. Efficient Video Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "text", + "content": "As quantitatively shown in Tab. 3, except for the regular 8-step ADM distillation for both sizes of CogVideoX, we also try integrating Classifier-Free Guidance (CFG) [11] for text-to-video task. Specifically, we assign the CFG scale for real model by randomly sampling within the range [5.0, 7.0], while assigning the few-step generator's scale through explicit subtraction of 2.0 from the real model's value. Empirically, we found that this subtractive offset requires progressive enlargement as the targeted sampling steps decreases. There is no CFG for the fake model. The VBench [14] results demonstrate that our few-step generators achieve comparable performance to the base model with " + }, + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "inline_equation", + "content": "92 - 96\\%" + }, + { + "bbox": [ + 55, + 402, + 296, + 629 + ], + "type": "text", + "content": " acceleration. We conduct additional evaluations on the longer training of 2B model, motivated by the observation that DMD loss did't converge well at 8K iterations as Fig. 3 indicates. Experimental results demonstrate that the DMD loss is also approximately optimized by the learnable discriminator during ADM distillation. More quantitative and qualitative results refer to Appendix C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 636, + 156, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 156, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 156, + 647 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "In Tab. 4, we conduct extensive ablation studies on fully fine-tuning SDXL-Base to validate our effectiveness. Qualitative comparisons are provided in Appendix C." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Effect of ADP. The following conclusions can be drawn: 1) Single reflow process provides very limited effectiveness" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 328, + 71, + 539, + 133 + ], + "blocks": [ + { + "bbox": [ + 87, + 217, + 263, + 227 + ], + "lines": [ + { + "bbox": [ + 87, + 217, + 263, + 227 + ], + "spans": [ + { + "bbox": [ + 87, + 217, + 263, + 227 + ], + "type": "text", + "content": "Table 4. Quantitative results on ablation studies." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 71, + 539, + 133 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 539, + 133 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 539, + 133 + ], + "type": "table", + "html": "
TTURTraining TimeCLIP ScorePick ScoreHPSv2MPS
1×1.0035.255722.273627.704611.1978
4×1.8535.258322.277327.725511.2720
8×2.5335.329922.288327.758611.2838
", + "image_path": "0e76aec7f5110c166a038ace359e3b50a1002c6aba0b5feb8f0620e9d93c553e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 154, + 550, + 189 + ], + "blocks": [ + { + "bbox": [ + 334, + 136, + 533, + 147 + ], + "lines": [ + { + "bbox": [ + 334, + 136, + 533, + 147 + ], + "spans": [ + { + "bbox": [ + 334, + 136, + 533, + 147 + ], + "type": "text", + "content": "Table 5. Ablation study on two time-scale update rule." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 154, + 550, + 189 + ], + "lines": [ + { + "bbox": [ + 317, + 154, + 550, + 189 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 550, + 189 + ], + "type": "table", + "html": "
ADDLCMLightningDMD2OursTeacher
LPIPS↑0.60710.62570.67070.67150.71560.6936
", + "image_path": "7146d726e25b72b9b2990c4b037a4d5a1545eb07134d77a74b7e80866957399b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 190, + 549, + 202 + ], + "lines": [ + { + "bbox": [ + 318, + 190, + 549, + 202 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 549, + 202 + ], + "type": "text", + "content": "Table 6. Quantitative diversity evaluation on PartiPrompts [88]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 212, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 331 + ], + "type": "text", + "content": "(A1), while multiple processes achieving effective rectification requires considerable computational expense. 2) Using SAM [19] offers more imaging fidelity compared to the widely adopted DINOv2 [48] (A2/A5), which is likely due to SAM's higher resolution of 1024px (compared to DINOv2's 518px). 3) The weighting " + }, + { + "bbox": [ + 313, + 212, + 553, + 331 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 212, + 553, + 331 + ], + "type": "text", + "content": " for pixel space should neither be too large nor too small. Excessive weighting leads to degraded structural integrity (A3/A5), while insufficient weighting results in blurriness (A4/A5). Qualitative results in Appendix C provide more visible distinction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 332, + 553, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 553, + 427 + ], + "type": "text", + "content": "Effect of ADM. We summarize the key findings: 1) The absence of ADP exerts substantial performance degradation (B1/B4), which is aligned with our analysis in Sec. 4.3.2. 2) Without a regularizer, the DMD loss underperforms compared to standalone ADM (B1/B2), indicating its poor robustness. 3) Although DMD loss optimization also benefits from ADP (B2/B3), its distribution matching capability remains inferior to ADM (B3/B4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 428, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 554, + 524 + ], + "type": "text", + "content": "Effect of TTUR. Tab. 5 presents the impact of different TTUR settings on final performance and training duration. The results demonstrate that increasing TTUR yields only marginal performance gains while nearly doubling training time, rendering this trade-off clearly unwarranted. This highlights the critical role of our proposed ADP in one-step distillation and suggests that the training instability in DMD2 likely stems from insufficient support set overlap." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 525, + 553, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 553, + 597 + ], + "type": "text", + "content": "Diversity evaluation. Following DMD2 [85], we generate 4 samples per prompt on Partiprompts [88] with different seeds, and report the averaged pairwise LPIPS similarities [89] in Tab. 6. The results indicate that our method significantly outperforms others in diversity. More randomly curated multi-seed samples can be found in Appendix C." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 609, + 389, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 389, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 389, + 621 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "We realize that one weakness is that teacher model might require CFG to produce accurate score prediction. Our experiments suggest this is a common characteristic of score distillation methods generally, not a limitation unique to our approach. This limits the application to guidance-distilled models such as FLUX.1-dev [1], which could be a promising research topic for future work." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16825" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (U22A2095, 12326618, 62276281), Guangdong Basic and Applied Basic Research Foundation, China (2024A1515011882) and the Project of Guangdong Provincial Key Laboratory of Information Security Technology (2023B1212060026)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 228 + ], + "type": "text", + "content": "[1] Black Forest Labs. Flux.1-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024.8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 260 + ], + "type": "text", + "content": "[2] Naresh Babu Bynagari. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 262, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 295, + 306 + ], + "type": "text", + "content": "[3] Clement Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin. Flash diffusion: Accelerating any conditional diffusion model for few steps image generation. arXiv preprint arXiv:2406.02347, 2024. 1, 7, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 341 + ], + "type": "text", + "content": "[4] Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost. arXiv preprint arXiv:1604.06174, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 342, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 296, + 407 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 409, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 409, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 61, + 409, + 296, + 485 + ], + "type": "text", + "content": "[6] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 7, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 518 + ], + "type": "text", + "content": "[7] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. In ICLR, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 520, + 296, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 296, + 563 + ], + "type": "text", + "content": "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 565, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 565, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 565, + 296, + 599 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 622 + ], + "type": "text", + "content": "[10] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 645 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS Workshops, 2021. 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 668 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[13] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 138 + ], + "type": "text", + "content": "[14] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models. In CVPR, 2024. 7, 8, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 194 + ], + "type": "text", + "content": "[15] Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models. arXiv preprint arXiv:2309.14509, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[16] Tejas Jayashankar, J. Jon Ryu, and Gregory Wornell. Score-of-mixture training: Training one-step generative models made simple via score estimation of mixture distributions. arXiv preprint arXiv:2502.09609, 2025. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[17] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 326 + ], + "type": "text", + "content": "[18] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 380 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In ICCV, pages 3992-4003, 2023. 5, 8, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 382, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 554, + 425 + ], + "type": "text", + "content": "[20] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Martiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. In NeuriPS, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 554, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 479 + ], + "type": "text", + "content": "[21] Jonas Kohler, Albert Pumarola, Edgar Schonfeld, Artsiom Sanakoyeu, Roshan Sumbaly, Peter Vajda, and Ali Thabet. Imagine flash: Accelerating emu diffusion models with backward distillation. arXiv preprint arXiv:2405.05224, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 481, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 503 + ], + "type": "text", + "content": "[22] Jae Hyun Lim and Jong Chul Ye. Geometric gan. arXiv preprint arXiv:1705.02894, 2017. 4, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[23] Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxlightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 537, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 580 + ], + "type": "text", + "content": "[24] Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, Lu Jiang, and ByteDance Seed. Diffusion adversarial post-training for one-step video generation. arXiv preprint arXiv:2501.08316, 2025. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ECCV, 2014. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[26] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[27] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2022. 1, 3, 5, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[28] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16826" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "diffusion-based text-to-image generation. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "text", + "content": "[30] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 196 + ], + "type": "text", + "content": "[31] Yanzuo Lu, Manlin Zhang, Yiqi Lin, Andy J. Ma, Xiaohua Xie, and Jianhuang Lai. Improving pre-trained masked autoencoder via locality enhancement for person re-identification. In PRCV, pages 509–521, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[32] Yanzuo Lu, Meng Shen, Andy J Ma, Xiaohua Xie, and JianHuang Lai. Mlnet: Mutual learning network with neighborhood invariance for universal domain adaptation. In AAAI, pages 3900-3908, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[33] Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. Coarse-to-fine latent diffusion for pose-guided person image synthesis. In CVPR, pages 6420-6429, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[34] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 1, 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[35] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 409 + ], + "type": "text", + "content": "[36] Weijian Luo. Diff-instruct++: Training one-step text-to-image generator model to align with human preferences. In TMLR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 454 + ], + "type": "text", + "content": "[37] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. In NeurIPS, pages 76525-76546, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[38] Weijian Luo, Zemin Huang, Zhengyang Geng, J. Zico Kolter, and Guo-jun Qi. One-step diffusion distillation through score implicit matching. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "text", + "content": "[39] Weijian Luo, Colin Zhang, Debing Zhang, and Zhengyang Geng. David and goliath: Small one-step model beats large diffusion with score post-training. In ICML, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[40] Yihong Luo, Xiaolong Chen, Xinghua Qu, Tianyang Hu, and Jing Tang. You only sample once: Taming one-step text-to-image synthesis by self-cooperative diffusion gans. In ICLR, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 611 + ], + "type": "text", + "content": "[41] Hongxu Ma, Guanshuo Wang, Fufu Yu, Qiong Jia, and Shouhong Ding. Ms-detr: Towards effective video moment retrieval and highlight detection by joint motion-semantic learning. In ACMMM, 2025. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 645 + ], + "type": "text", + "content": "[42] Hongxu Ma, Chenbo Zhang, Lu Zhang, Jiaogen Zhou, Ji-hong Guan, and Shuiqeng Zhou. Fine-grained zero-shot object detection. In ACMMM, 2025. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[43] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. arXiv preprint arXiv:2409.11367, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[44] Yuxi Mi, Zhizhou Zhong, Yuge Huang, Qiuyang Yuan, Xuan Zhao, Jianqing Xu, Shouhong Ding, Shaoming Wang, Rizen" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Guo, and Shuigeng Zhou. Data synthesis with diverse styles for face recognition via 3dmm-guided diffusion. In CVPR, pages 21203-21214, 2025. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[45] Thomas Minka. Divergence measures and message passing. Microsoft Research, Technical Report, 2005. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[46] Movie Gen Team. Movie gen: A cast of media foundation models. https://ai.meta.com/static-resource/movie-gen-research-paper, 2024.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "text", + "content": "[47] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 219, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 317 + ], + "type": "text", + "content": "[48] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 8, 1, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 319, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 341 + ], + "type": "text", + "content": "[49] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 4, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[50] PKU-Yuan Lab and Tuzhan AI. Open-sora-plan-v1.1.0. https://huggingface.co/datasets/LanguageBind/Open-Sora-Plan-v1.1.0, 2024.7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "text", + "content": "[51] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 474 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 7, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 477, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 509 + ], + "type": "text", + "content": "[53] Samyam Rajbhandari, Jeff Rasley, Olatunj Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. In SC, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 543 + ], + "type": "text", + "content": "[54] Prajit Ramachandran, Barret Zoph, and Quoc V. Le. Searching for activation functions. arXiv preprint arXiv:1710.05941, 2017. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "text", + "content": "[55] Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-sd: Trajectory segmented consistency model for efficient image synthesis. In NeurIPS, 2024. 1, 3, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "text", + "content": "[56] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 7, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[57] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 4, 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[58] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *ICLR*, 2022. 1, 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[59] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. In NeurIPS, 2024. 1, 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16827" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 160 + ], + "type": "text", + "content": "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "text", + "content": "[62] Meng Shen, Yanzuo Lu, Yanxu Hu, and Andy J. Ma. Collaborative learning of diverse experts for source-free universal domain adaptation. In ACM MM, pages 2054–2065, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 217 + ], + "type": "text", + "content": "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "type": "text", + "content": "[64] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 283, + 294, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 305 + ], + "type": "text", + "content": "[66] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 306, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 294, + 360 + ], + "type": "text", + "content": "[67] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, Jifeng Dai, Yu Qiao, Limin Wang, and Hongsheng Li. Journeydb: A benchmark for generative image understanding. arXiv preprint arXiv:2307.00716, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 392 + ], + "type": "text", + "content": "[68] Zhiqiang Tan, Yunfu Song, and Zhijian Ou. Calibrated adversarial algorithms for generative modelling. Stat, 8:e224, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "text", + "content": "[69] Fu-Yun Wang, Zhaoyang Huang, Alexander William Bergman, Dazhong Shen, Peng Gao, Michael Lingelbach, Keqiang Sun, Weikang Bian, Guanglu Song, Yu Liu, Hongsheng Li, and Xiaogang Wang. Phased consistency model. arXiv preprint arXiv:2405.18407, 2024. 1, 3, 7, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 503 + ], + "type": "text", + "content": "[70] Fu-Yun Wang, Zhaoyang Huang, Weikang Bian, Xiaoyu Shi, Keqiang Sun, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Computation-efficient personalized style video generation without personalized video data*. In SIGGRAPH ASIA Technical Communications, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "text", + "content": "[71] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. arXiv preprint arXiv:2410.07303, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[72] Shibo Wang and Pankaj Kanwar. Bfloat16: The secret to high performance on cloud tpus. https://cloud.google.com/blog/products/ai-machine-learning/bfloat16-the-secret-to-high-performance-on-cloud-tpus, 2019.2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 604, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 689 + ], + "type": "text", + "content": "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[74] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "diverse text-to-3d generation with variational score distillation. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[75] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 173 + ], + "type": "text", + "content": "[76] Yuxin Wu and Kaiming He. Group normalization. In ECCV, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[77] Xuefeng Xiao, Lianwen Jin, Yafeng Yang, Weixin Yang, Jun Sun, and Tianhai Chang. Building fast and compact convolutional neural networks for offline handwritten Chinese character recognition. Pattern Recognition, 72:72-81, 2017. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[78] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurlPS, 2023. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "type": "text", + "content": "[79] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In CVPR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 300, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 553, + 354 + ], + "type": "text", + "content": "[80] Zunnan Xu, Zhentao Yu, Zixiang Zhou, Jun Zhou, Xiaoyu Jin, Fa-Ting Hong, Xiaozhong Ji, Junwei Zhu, Chengfei Cai, Shiyu Tang, et al. Hunyuanportrait: Implicit condition control for enhanced portrait animation. In CVPR, pages 15909-15919, 2025. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "text", + "content": "[81] Hanshu Yan, Xingchao Liu, Jiachun Pan, Jun Hao Liew, Qiang Liu, and Jiashi Feng. Perflow: Piecewise rectified flow as universal plug-and-play accelerator. In NeurIPS, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 443 + ], + "type": "text", + "content": "[82] Dongjie Yang, Suyuan Huang, Chengqiang Lu, Xiaodong Han, Haoxin Zhang, Yan Gao, Yao Hu, and Hai Zhao. Vript: A video is worth thousands of words. arXiv preprint arXiv:2406.06040, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 521 + ], + "type": "text", + "content": "[83] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[84] Hongwei Yi, Shitong Shao, Tian Ye, Jiantong Zhao, Qingyu Yin, Michael Lingelbach, Li Yuan, Yonghong Tian, Enze Xie, and Daquan Zhou. Magic 1-for-1: Generating one minute video clips within one minute. arXiv preprint arXiv:2502.07701, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[85] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[86] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, pages 6613-6623, 2024. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[87] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16828" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 430 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[88] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. TMLR, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "text", + "content": "[89] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[90] Sixian Zhang, Bohan Wang, Junqiang Wu, Yan Li, Tingting Gao, Di Zhang, and Zhongyuan Wang. Learning multidimensional human preference for text-to-image generation. In CVPR, pages 8018-8027, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 283 + ], + "type": "text", + "content": "[91] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, Sergey Tulyakov, and Jian Ren. Sf-v: Single forward video generation model. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 350 + ], + "type": "text", + "content": "[92] Jianbin Zheng, Minghui Hu, Zhongyi Fan, Chaoyue Wang, Changxing Ding, Dacheng Tao, and Tat-Jen Cham. Trajectory consistency distillation: Improved latent consistency distillation by semi-linear consistency function with trajectory mapping. arXiv preprint arXiv:2402.19159, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[93] Mingyuan Zhou, Huangjie Zheng, Zhendong Wang, Mingzhang Yin, and Hai Huang. Score identity distillation: Exponentially fast distillation of pretrained diffusion models for one-step generation. In ICML, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 430 + ], + "type": "text", + "content": "[94] Mingyuan Zhou, Huangjie Zheng, Yi Gu, Zhendong Wang, and Hai Huang. Adversarial score identity distillation: Rapidly surpassing the teacher in one step. In ICLR, 2025. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16829" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_content_list.json b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..95eb95547c6199f36720b3d0c7e0c81bdd7fd8d2 --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_content_list.json @@ -0,0 +1,1626 @@ +[ + { + "type": "text", + "text": "Adversarial Exploitation of Data Diversity Improves Visual Localization", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sihang Li* Siqi Tan* Bowen Chang Jing Zhang Chen Feng Yiming Li", + "bbox": [ + 161, + 179, + 830, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "New York University", + "bbox": [ + 411, + 205, + 583, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sihangli, tan.k, cfeng, yimingli}@nyu.edu", + "bbox": [ + 313, + 224, + 683, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ai4ce.github.io/RAP", + "bbox": [ + 375, + 243, + 614, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual localization, which estimates a camera's pose within a known scene, is a fundamental capability for autonomous systems. While absolute pose regression (APR) methods have shown promise for efficient inference, they often struggle with generalization. Recent approaches attempt to address this through data augmentation with varied viewpoints, yet they overlook a critical factor: appearance diversity. In this work, we identify appearance variation as the key to robust localization. Specifically, we first lift real 2D images into 3D Gaussian Splats with varying appearance and deblurring ability, enabling the synthesis of diverse training data that varies not just in poses but also in environmental conditions such as lighting and weather. To fully unleash the potential of the appearance-diverse data, we build a two-branch joint training pipeline with an adversarial discriminator to bridge the syn-to-real gap. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, reducing translation and rotation errors by $50\\%$ and $33\\%$ on indoor datasets, and $38\\%$ and $44\\%$ on outdoor datasets. Most notably, our method shows remarkable robustness in dynamic driving scenarios under varying weather conditions and in day-to-night scenarios, where previous APR methods fail.", + "bbox": [ + 89, + 323, + 483, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 686, + 220, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual localization, the task of calculating a 6-DoF camera pose—its translation and rotation—based on a query image within a given environment, is essential for various applications, including robotics [1], autonomous vehicles [18], and virtual reality [11]. Besides traditional geometry-based approaches, recent learning-based visual localization methods adopt absolute pose regression (APR) [5, 8, 24, 53], scene coordinate regression (SCR) [3, 4, 43, 61], or post pose refinement (PPR) [9, 19, 33, 43, 59, 64, 66]. SCR methods focus on learning-based 2D-3D correspondences followed by subsequent Perspective-n-Point (PnP) for pose estimation.", + "bbox": [ + 89, + 713, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/61b4c710face4e57011a77d867b8ed225c07bac95a93e4145e0c2c5241857bc4.jpg", + "image_caption": [ + "Figure 1. We propose RAP, a novel pipeline to train robust APR models. We lift real-world 2D images into 3D Gaussian Splats [25] to synthesize images with diverse appearances and poses, improving model generalizability. We also introduce an adversarial discriminator, mitigating the syn-to-real gap to learn robust features. Together, we achieve state-of-the-art performance." + ], + "image_footnote": [], + "bbox": [ + 514, + 295, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "PPR methods heavily rely on a pose prior, usually obtained from image retrieval, followed by iterative refinement. In contrast, APR methods employ a supervised framework to train a regression neural network on image-pose pairs, enabling direct pose estimation during inference. APR offers faster runtime and lower error in challenging scenes with sparse views, significant lighting variations, or numerous dynamic objects, making it a promising method for ensuring robustness in real-world applications.", + "bbox": [ + 511, + 609, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite promises, there is a performance gap in localization accuracy between APR and other methods. A well-known pivotal work [49] attributes this to APR performing image-based memorization, i.e., retrieving poses seen during training. Driven by this crucial finding, to improve such memorization while avoiding the need for denser real-world training samples, recent methods leverage Neural Radiance Fields (NeRF) [40] to synthesize additional posed images for APR training [8, 30, 42]. LENS [42] tried to employ appearance perturbation using NeRF-W [39], but found the", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 114, + 886, + 367, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26848", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "improvements minimal [42]. As appearance augmentation is common in other learning tasks, why does it fail in APR?", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hypothesize a learning gap: Limitations in previous training pipelines prevented the effective use of diverse data to boost performance. Artifacts always exist in images rendered by common novel view synthesis (NVS) methods, which might disturb the model feature space. Inspired by generative adversarial networks (GAN) [20], where a discriminator is trained to distinguish between real and generated samples, we propose adversarial training for APR, designing a discriminator to align the features of synthetic and real images, thereby reducing the syn-to-real domain gap and mitigating the impact of rendering artifacts. Augmentation quality also matters. To efficiently synthesize diverse high-quality images with controllable varying appearance, we extend the vanilla 3D Gaussian Splitting (3DGS) [25] to appearance-varying 3DGS with deblurring ability.", + "bbox": [ + 88, + 121, + 482, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These form our two-branch joint training framework for robust absolute pose regression (RAP). The first branch coarsely trains our Transformer-based pose regressor with both real data and data synthesized at the original real pose, together with an adversarial discriminator to reduce the synto-real domain gap. The second branch progressively generates randomly perturbed poses and appearances, providing additional supervision to the same APR Transformer. Through extensive experiments, we demonstrate that exploiting data diversity using adversarial training significantly increases localization accuracy in APR. Meanwhile, our results indicate that APR consistently benefits from more diverse visual data, and we observe clear signs of a more generalizable APR emerging with its localization performance cannot be explained merely by memorization.", + "bbox": [ + 91, + 348, + 483, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 109, + 574, + 413, + 587 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify the crucial role of appearance diversity for APR, and develop a 3DGS-based appearance-varying data augmentation framework to efficiently generate diverse synthetic data with controllable lighting conditions.", + "- We propose an adversarial discriminator to reduce the syn-to-real gap. Together with progressive data synthesis, we form a robust two-branch joint training pipeline that fully unleashes the power of data diversity.", + "- We conduct extensive experiments showing our method outperforms state-of-the-art approaches on challenging datasets with significant appearance change. Ablation studies further analyze key factors affecting performance." + ], + "bbox": [ + 89, + 589, + 482, + 770 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 776, + 240, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Localization. Visual localization aims to estimate a camera's translation and rotation within a 3D scene. Traditional geometry-based methods [6, 15, 32, 35, 44-47, 57] accomplish this by using point clouds and a reference image database, relying on stored descriptors and image retrieval to establish 2D-3D correspondences. In contrast, scene coordinate regression (SCR) methods [2-4, 61] em", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bed map information within neural networks to directly predict 2D-3D correspondences. Both approaches generally require PnP [17] and RANSAC [16] to output camera poses at test time, which adds additional computation cost. Alternatively, absolute pose regression (APR) [5, 7, 22, 24, 41, 52] aims to directly regress the camera pose from a query image using neural networks. Although the performance is suboptimal compared with geometry-based methods, APR remains a promising approach due to its fast inference.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data Augmentation for Pose Regression. End-to-end pose regression methods rely heavily on the amount and diversity of training data. Previous work [49] shows that APR implicitly learns image retrieval in the given environment. Therefore, the following works LENS [42], DFNet [8] and PMNet [30] enhance APR performance by spatially enriching training views with NeRF. However, these approaches fail to address the generalizability of APR models and exhibit several limitations: (1) The efficiency of training and novel view synthesis (NVS) in NeRF is severely restricted, hindering scalability. (2) They limit NVS to geometric (pose) transformations while neglecting photometric (appearance) variations, thereby decreasing APR robustness to changes in visual appearance. (3) The augmented data is underutilized in their learning frameworks, leaving its potential for improving APR largely untapped. Differently, our framework switches to 3DGS [25] as the scene representation to efficiently generate novel posed images with controllable appearances and introduce adversarial training to unleash the power of such diverse data.", + "bbox": [ + 511, + 227, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Handling and Synthesizing Challenging Scenarios. Visual localization often encounters unstructured photo collections [56], where visual appearance varies due to moving objects, lighting changes, and inconsistent camera exposure settings. To tackle these in-the-wild challenges, NeRF-W [38] uses per-image transient and appearance embeddings. In 3DGS [25], VastGaussian [31] applies a CNN to 3DGS outputs but still struggles with significant appearance variations. SWAG [12] mitigates this issue by storing appearance information in an external hash-grid-based implicit field, while GS-W [63] enhances flexibility by separating intrinsic and dynamic appearance features for each Gaussian point. 3DGM [29] leverages consensus across multiple sequences as the self-supervision signal to remove transient and moving objects without human annotations. Deblur-GS [62] addresses motion blur—another challenge in localization datasets—by modeling camera motion to yield sharper edges in rendered scenes. Our method incorporates appearance modeling and edge refinement to handle and synthesize diverse indoor, outdoor, and driving scenes.", + "bbox": [ + 511, + 529, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 835, + 604, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Pre-Processing with 3DGS", + "text_level": 1, + "bbox": [ + 511, + 854, + 754, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A robust pose regressor should focus on intrinsic scene attributes, not appearance variations. Therefore, we first", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26849", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ded7d99ee2f9af0dfe52d63070baaaed376e73807554e871bd73f06b20609cb5.jpg", + "image_caption": [ + "Figure 2. Pipeline of RAP. We lift multiple RGB video sequences into 3D Gaussian Splats, which serve as our data engine. The branch-1 (see Sec. 3.3.1) inputs paired real and synthetic images to regress poses, with a discriminator to bridge the syn-to-real gap. The branch-2 (see Sec. 3.3.2) generates views with novel poses and appearances, which are fed into the same pose regressor as additional supervision." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further analyze the behavior of the $K$ in the $K \\times 16 \\times 3$ system, we assume that the camera is located at the center of the image. The camera is positioned at the center of the image by a camera position vector $\\mathbf{C} = [0, 1]$ . To achieve this, we use the following procedure to generate the camera positions:", + "bbox": [ + 91, + 338, + 483, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {C}} = \\operatorname {M L P} (\\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\omega \\boldsymbol {\\mathcal {E}}, \\theta), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 563, + 482, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega$ is the blending weight that controls the dynamic appearance of the rendered image.", + "bbox": [ + 89, + 593, + 482, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Another significant challenge in visual localization is motion blur, often caused by slow shutter speeds during video capture, leading to pose ambiguity and degraded rendering quality, further decreasing localization accuracy. Inspired by Deblur-GS [62], we model camera motion blur as the inverse of scene motion, i.e., the transformation in Gaussian position denoted by $\\mathcal{T} \\in \\mathrm{SE}(3)$ . For each training image, we sample a certain time step along a linear trajectory with a sampling weight $\\phi \\in \\mathbb{R}^n$ and blend them to compute loss $\\mathcal{L}$ with the original blurry image $\\mathcal{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}$ , optimizing $\\mathcal{T}$ , $\\phi$ , $\\mathcal{C}$ and other 3DGS parameters $\\Theta$ :", + "bbox": [ + 89, + 625, + 482, + 790 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\phi , \\boldsymbol {\\mathcal {T}}, \\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\boldsymbol {\\mathcal {E}}, \\boldsymbol {\\Theta}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\mathcal {I}} _ {b}, \\sum_ {i = 1} ^ {n} \\phi_ {i} \\operatorname {R e n d e r} \\left(\\boldsymbol {\\mathcal {T}} _ {i} (\\boldsymbol {\\mu}), \\boldsymbol {\\mathcal {C}}, \\boldsymbol {\\Theta}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 801, + 483, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the details of $\\mathcal{L}$ are in supplemental materials. After training, our 3DGS can efficiently render posed images given $\\theta$ and $\\omega$ .", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture of Pose Regressor", + "text_level": 1, + "bbox": [ + 511, + 337, + 785, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of images and their associated camera poses $\\{(I_i, P_i)\\}_{i=1}^n$ , our goal is to train a neural network to directly output a homogeneous camera pose $P \\in \\mathbb{R}^{3 \\times 4}$ for a query image $I \\in \\mathbb{R}^{H \\times W \\times C}$ . Our network architecture is shown as the pose regressor in Fig. 2.", + "bbox": [ + 511, + 359, + 906, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Extractor. Pose regression networks typically extract features using a common backbone $\\varphi$ , such as VGG [55] or EfficientNet [58], leveraging multiple deeper layers for translation and rotation regression:", + "bbox": [ + 511, + 435, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi (\\boldsymbol {I}) = \\left\\{\\mathcal {F} _ {0} (\\boldsymbol {I}), \\dots , \\mathcal {F} _ {N - 1} (\\boldsymbol {I}), \\mathcal {F} _ {N} (\\boldsymbol {I}) \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 503, + 903, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{F}_{*}(\\cdot)$ denotes features extracted from the $*$ -th layer of a backbone with $N$ layers. $\\mathcal{F}_t(\\pmb {I})$ and $\\mathcal{F}_r(\\pmb {I})$ denote features for translation and rotation regression, respectively.", + "bbox": [ + 511, + 529, + 905, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose Transformer. Unlike CNN-based regression models [8, 30], where fine-grained local features can introduce noise and harm performance, we propose Pose Transformer to leverage the strong ability of Vision Transformer (ViT) [14] for modeling long-range dependencies. Each Transformer generates a global token (Trans for translation and Rot for rotation) to provide a comprehensive context for pose regression, inspired by the CLS token in ViT. Given $\\mathcal{F}_r(I)$ and $\\mathcal{F}_t(I)$ , the translation token is then concatenated with the flattened input features*:", + "bbox": [ + 511, + 574, + 905, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\operatorname {C a t} (\\text {F l a t t e n} (\\mathcal {F} _ {t} (\\boldsymbol {I})), \\text {T r a n s}) \\in \\mathbb {R} ^ {\\left(H _ {t} W _ {t} + 1\\right) \\times C _ {t}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 732, + 903, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The positional encodings are then added to the flattened feature $(\\mathsf{PE} + \\widetilde{\\mathcal{F}}_t(\\pmb {I}))\\in \\mathbb{R}^{(H_tW_t + 1)\\times C_t}$ . Multi-head Self-Attention (MSA) is then conducted through a stack of multiple layers with the post-processing as follows:", + "bbox": [ + 511, + 766, + 906, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {F} _ {t} ^ {\\prime}} (\\boldsymbol {I}) = \\operatorname {M S A} \\left(\\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I})\\right) + \\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 835, + 903, + 863 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\mathrm {L N} (\\mathrm {F F N} (\\mathrm {L N} (\\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I}))) + \\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I})),\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 859, + 846, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*We only present the translation regression for simplicity.", + "bbox": [ + 531, + 887, + 836, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26850", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where LN indicates layer normalization and FFN denotes the fully connected feed-forward network, consisting of two linear layers with a ReLU. The final output is flattened back to $(H_{t}W_{t} + 1)\\times c_{t}$ . See supplementary for more details.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regression Head. Only the processed translation token, Trans, capturing global features for regression, is fed into the regression head. This regression head consists of two MLPs, each with a hidden layer and GeLU activation:", + "bbox": [ + 89, + 152, + 482, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {t}} = \\operatorname {L i n e a r} (\\operatorname {G e L U} (\\operatorname {L i n e a r} (\\widehat {\\text {T r a n s}}))). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 215, + 480, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\hat{t}$ represents the final prediction for translation. Similarly, we obtain the rotation prediction denoted by $\\hat{r}$ .", + "bbox": [ + 89, + 239, + 482, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Two-Branch Joint Training Paradigm", + "text_level": 1, + "bbox": [ + 89, + 277, + 416, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Branch-1: Aligning Features via Discriminator", + "text_level": 1, + "bbox": [ + 89, + 299, + 464, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthetic images from 3DGS provide novel viewpoints and appearances but often contain artifacts, leading to a syn-to-real domain gap. To align features from rendered and real images of the same pose, we introduce an adversarial training mechanism besides the basic pose regression training.", + "bbox": [ + 89, + 316, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Regression Loss. For basic training, we render the synthetic image $I^{\\prime}$ with the same pose label $\\pmb{P}$ as the real image $\\pmb{I}$ , both used as supervision for the pose regressor. The training objective consists of translation loss $\\mathcal{L}_t$ and the rotation loss $\\mathcal{L}_r$ , which are measured by the Euclidean distance between the ground truth pose $\\pmb{P} = \\{\\hat{t},\\hat{r}\\}$ and the estimated pose $\\widehat{P} = \\{\\hat{t},\\hat{r}\\}$ :", + "bbox": [ + 89, + 393, + 483, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\| \\boldsymbol {t} - \\hat {\\boldsymbol {t}} \\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 520, + 480, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r} = \\left\\| \\boldsymbol {r} - \\frac {\\hat {\\boldsymbol {r}}}{\\| \\hat {\\boldsymbol {r}} \\|} \\right\\| _ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 541, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s e}} ^ {1} = \\mathcal {L} _ {t} \\exp (- s _ {t}) + s _ {t} + \\mathcal {L} _ {r} \\exp (- s _ {r}) + s _ {r}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 577, + 480, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_t$ and $s_r$ are learned parameters for balancing the optimization between rotation and translation [23].", + "bbox": [ + 89, + 599, + 482, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial Loss. The adversarial training mechanism optimizes the discriminator to distinguish real from rendered image features, while training the feature extractor to fool the discriminator, effectively bridging the domain gaps. To prevent vanishing gradients, we propose a novel adversarial objective for pose regression, inspired by LSGAN [37]:", + "bbox": [ + 89, + 630, + 483, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {D} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {D i s}} (D) = \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} \\sim p _ {\\mathrm {d a t a}} (\\boldsymbol {I})} \\left[ \\left(D (\\operatorname {A d j} (\\mathcal {F} _ {t} (\\boldsymbol {I}))) - 1\\right) ^ {2} \\right] \\\\ + \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} ^ {\\prime}} \\left[ D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {I} ^ {\\prime}\\right)\\right) ^ {2} \\right], \\right. \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 726, + 480, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {G} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {G e n}} (G) = \\frac {1}{2} \\mathbb {E} _ {I ^ {\\prime}} \\left[ \\left(D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(I ^ {\\prime}\\right)\\right) - 1\\right) ^ {2} \\right]. \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 789, + 480, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, Adj and $\\mathrm{Adj}'$ are the adjustment layers, consisting of Conv-ReLU-BN layers. The feature extractor $\\varphi$ acts as the generator $G$ , while $D$ is the discriminator, composed of several convolutional layers with ReLU activations. More details are in supplemental materials.", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Branch-2: Training while Synthesizing Data", + "text_level": 1, + "bbox": [ + 511, + 90, + 867, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the proposed appearance-varying 3DGS, more posed images are generated to enrich the training data for better generalizability. Specifically, our data synthesis is categorized into two dimensions: pose augmentation and appearance augmentation, as illustrated in Fig. 1. For pose augmentation, given a training pose $P$ , a perturbed pose $P_{\\mathrm{syn}}$ is generated around $P$ by the translation noise of $\\delta t$ and rotation noise of $\\delta r$ . For appearance augmentation, we randomly adjust the appearance of rendered images using random blending weights $\\omega$ , and then render the synthetic image $I_{\\mathrm{syn}}$ using the Gaussian Splats trained in Sec. 3.1. The novel image-pose pair $(I_{\\mathrm{syn}}, P_{\\mathrm{syn}})$ , online generated every 20 epochs during training until the validation MSE loss and median errors cease to decrease, serves as additional supervision for the training. Given the estimated pose of the synthesized image denoted by $\\widehat{P_{\\mathrm{syn}}}$ , the loss function $\\mathcal{L}_{\\mathrm{pose}}^2(\\widehat{P_{\\mathrm{syn}}}, P_{\\mathrm{syn}})$ is same as $\\mathcal{L}_{\\mathrm{pose}}^1$ .", + "bbox": [ + 511, + 109, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Overall Objective", + "text_level": 1, + "bbox": [ + 513, + 381, + 683, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss for the pose regressor is:", + "bbox": [ + 513, + 401, + 769, + 416 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {1} \\mathcal {L} _ {\\text {p o s e}} ^ {1} + \\beta_ {2} \\mathcal {L} _ {\\text {p o s e}} ^ {2} + \\beta_ {3} \\left(\\mathcal {L} _ {\\text {G e n}} + \\mathcal {L} _ {\\text {D i s}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 428, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{1},\\beta_{2},\\beta_{3}$ are loss weights. The total loss will optimize the pose regressor, adjustment layers, and discriminator. Only the pose regressor will be deployed in the inference phase, while the other two components are discarded.", + "bbox": [ + 511, + 458, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 535, + 645, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Evaluation Setup", + "text_level": 1, + "bbox": [ + 513, + 560, + 681, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets. We follow previous works [8, 30] to mainly use four scenes in the Cambridge Landmarks dataset [24] with spatial extents around $875\\mathrm{m}^2$ . Moreover, we evaluate our method on MARS [28], a self-driving dataset featuring challenges like moving objects, lighting changes, and motion blur. To investigate the robustness of our model under extreme lighting changes, such as the transition from day to night, we also prepared a subset of the Aachen Day-Night dataset [48]. The training data includes images captured using various camera models with differing resolutions, which renders direct evaluation with APR methods infeasible. Thus, we standardized the camera models through center cropping and built a COLMAP [50] model as pose annotations, including 13 nighttime images for evaluation and 246 daytime images for training 3DGS and RAP. We also employ the 7-Scenes dataset [54], which provides seven indoor scenes with volumes spanning $1\\mathrm{m}^3 - 18\\mathrm{m}^3$ and follow the original training and testing splits with more accurate SfM pose annotations [4, 9]. Although it is an indoor dataset, it is still non-trivial as it includes texture-less surfaces, object occlusions, and motion blur.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26851", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15ebf38683dbd6908706749e21ed47014c0be1f8e9f74ee3345b54c5b2a5dc5a.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison of camera pose estimation errors between a) DFNet [8] and b) our RAP framework across five scenes on the Cambridge Landmarks dataset [24]. Our RAP framework estimates trajectories that more closely follow the ground truth, with significantly reduced rotation and position errors compared to DFNet [8]." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 911, + 294 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/707b802e0a70c1f7f8c9395601504ea31d5aaf6a79139cf3f0eaf523765aab02.jpg", + "table_caption": [ + "Table 1. Median translation (cm) and rotation $(^{\\circ})$ errors on the Cambridge Landmarks dataset [24]. The best results† in pure APR and PPR are highlighted in bold." + ], + "table_footnote": [ + "Since most methods did not report results on Court, it is excluded from the average error calculation. As Coordinet + LENS [42] does not provide open-source code, it is unclear whether any post-processing is used." + ], + "table_body": "
MethodsCollegeHospitalShopChurchAverage‡Court
Pure APRPN [24]166/4.86262/4.90141/7.18245/7.95204/6.23683/3.50
MapNet [5]107/1.89194/3.91149/4.22200/4.53163/3.64N/A
MS-Trans. [53]83/1.47181/2.3986/3.07162/3.99128/2.73N/A
PAE [51]90/1.49207/2.5899/3.88164/4.16140/3.03N/A
LENS† [42]33/0.5044/0.9027/1.6053/1.6039/1.20N/A
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90217/4.11
PMNet [30]68/1.97103/1.3158/2.10133/3.7390/2.27N/A
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28115/1.68
SCRDSAC* [2]18/0.321/0.45/0.315/0.615/0.434/0.2
ACE [4]28/0.431/0.65/0.318/0.621/0.543/0.2
GLACE [61]19/0.317/0.44/0.29/0.312/0.319/0.1
PPRFQN-MN [19]28/0.3854/0.8213/0.6358/2.0038/0.964253/39.16
LENS [42]34/0.5445/0.9628/1.6654/1.6640/1.21N/A
CrossFire [43]47/0.743/0.720/1.239/1.437/1.0N/A
NeFeS50 [9]37/0.5452/0.8815/0.5337/1.1435/0.77N/A
HR-APR [33]36/0.5853/0.8913/0.5138/1.1635/0.78N/A
MCLoc [59]31/0.4239/0.7312/0.4526/0.8827/0.62N/A
DFNetGS-CPR [34]23/0.3242/0.7410/0.3627/0.6226/0.51N/A
ACEGS-CPR [34]20/0.2921/0.405/0.2413/0.4015/0.33N/A
DFNetref(Ours)16/0.2421/0.418/0.4210/0.2614/0.3325/0.13
RAPref (Ours)15/0.2518/0.385/0.239/0.2312/0.2723/0.15
", + "bbox": [ + 94, + 398, + 480, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We first compare our proposed RAP against common APR-only approaches on the four datasets, where PMNet [30] and DFNet [8] are the most related and advanced methods based on data augmentation. We split the remaining methods into two categories based on whether they rely on extra novel view synthesis in test time, including SCR [3, 4, 61] and PPR (Post Pose Refinement) [9, 19, 33, 43, 59, 64, 66], which involves rendering images, querying features in novel views by the initial pose, iterative refinement or sequential refinement [41].", + "bbox": [ + 89, + 718, + 482, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. First, we optimize our 3DGS for each scene without masking moving objects. We", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c57aa03fd01ce4f456179d649cd6a2a01a90affa509ee3d9b08a653d61298417.jpg", + "image_caption": [ + "Figure 4. Visualization of $\\mathrm{RAP}_{\\mathrm{ref}}$ on MARS [28]. In each subfigure, a diagonal line separates the \"Predicted\" (rendered from the refined pose) and \"GT\" (ground truth) sections. Smooth alignment along this boundary shows $\\mathrm{RAP}_{\\mathrm{ref}}$ 's improved pose accuracy." + ], + "image_footnote": [], + "bbox": [ + 514, + 352, + 906, + 535 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/95cc2a1ca9e7dfb03a4d4835f9ff48014b46a20798abfdae83d9b40e6cdaeacf.jpg", + "table_caption": [ + "Table 2. Median translation (cm) and rotation $(^{\\circ})$ errors on the MARS dataset [28]." + ], + "table_footnote": [], + "table_body": "
Methods“0011”“0015”“0037”“0041”Average
PoseNet [24]149/1.80136/2.34123/1.6075/0.92121/1.67
RAP (Ours)32/0.6137/1.0815/0.3528/0.3528/0.60
RAPref (Ours)8.5/0.138.2/0.208.7/0.097.6/0.118.3/0.13
", + "bbox": [ + 545, + 638, + 877, + 696 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3398e30e20ff627d4ecf30f724f6d702750b29ef74da0c7109874b46e2e30b81.jpg", + "table_caption": [ + "Table 3. Median translation (COLMAP [65] unit) and rotation $(^{\\circ})$ errors on the Aachen Day-Night Dataset [48]." + ], + "table_footnote": [], + "table_body": "
APR-BasedSCR-Based
PoseNet [24]DFNet [8]RAP w/o App.RAPRAPrefACE [4]GLACE [61]
217/74.30174/85.80134/75.99130/13.7067/4.64914/90.50482/36.4
", + "bbox": [ + 516, + 746, + 911, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then train our RAP network, which uses an Efficient-B0 backbone [36] pre-trained on ImageNet [13], optimized with Adam [26] at a learning rate of $10^{-4}$ . Only the features from the third (reduction_3) and fourth (reduction_4) layers are used respectively for translation and rotation regression, and both layers are utilized for", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26852", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c2e0d3a703c676af4979a21260e04cec28f2293f6fbad517b34edcf025eaabbd.jpg", + "image_caption": [ + "Figure 5. Visualization of the localization errors of $\\mathrm{RAP}_{\\mathrm{ref}}$ on the 7-Scenes dataset [54]." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/edf6bccc4647ad8798d7869d96b6fac0619d09812bd21c01bfe949b3644b9077.jpg", + "table_caption": [ + "Table 4. Quantitative results on the 7-Scenes dataset [54]. The best results in pure APR and PPR are highlighted in bold. DSLAM GT and SfM GT refer to different sets of ground truth. More visualizations and details are in supplemental materials." + ], + "table_footnote": [ + "As marepo [10] combines SCR and APR, we classify it as SCR." + ], + "table_body": "
CategoryMethodsChessFireHeadsOfficePumpkinKitchenStairsAverage
Pure APRPoseNet (PN) [24]32/8.1247/14.429/1.2048/7.6847/8.4259/8.6447/13.8044/10.4
MapNet [5]8/3.2527/11.718/13.317/51.522/4.0223/4.9330/12.121/7.77
AtLoc+ [60]10/3.1826/10.814/11.417/5.1620/3.9416/4.9029/10.219/7.08
MS-Transformer [53]11/4.6624/9.6014/12.217/5.6618/4.4417/5.9417/5.9418/7.28
PAE [51]12/4.9524/9.3114/12.519/5.7918/4.8918/6.1925/8.7419/7.48
CoordiNet + LENS [42]4/1.3811/3.778/5.868/1.989/2.2710/2.2715/3.679/3.07
DFNet [8]5/1.8817/6.456/3.638/2.4810/2.7822/5.4516/3.2912/3.71
PMNet [30]4/1.7010/4.517/4.237/1.9614/3.3314/3.3616/3.6210/3.24
RAP (Ours, DSLAM GT)3/1.417/3.466/6.025/1.976/1.967/2.1810/2.146/2.73
RAP (Ours, SfM GT)2/0.856/2.844/4.525/1.944/1.707/2.1210/1.305/2.18
SCRDSAC [2]0.5/0.170.8/0.280.5/0.341.2/0.341.2/0.280.7/0.212.7/0.781.1/0.34
ACE [4]0.5/0.180.8/0.330.5/0.331.0/0.291.0/0.220.8/0.202.9/0.811.1/0.34
GLACE [61]0.6/0.180.9/0.340.6/0.341.1/0.290.9/0.230.8/0.203.2/0.931.2/0.36
marepo† [10]2.6/1.352.5/1.422.3/2.213.6/1.444.2/1.555.1/1.996.7/1.833.9/1.68
PPRFQN-MN [19]4.1/1.3110.5/2.979.2/2.453.6/2.364.6/1.7616.1/4.42139.5/34.6728/7.3
CrossFire [43]1/0.45/1.93/2.35/1.63/0.82/0.812/1.94.4/1.38
DFNet + NeFeS50 [9]2/0.572/0.742/1.282/0.562/0.552/0.575/1.282.4/0.79
HR-APR [33]2/0.552/0.752/1.452/0.642/0.622/0.675/1.302.4/0.85
MCLoc [59]2/0.83/1.43/1.34/1.35/1.66/1.66/2.04.1/1.43
DFNet + GS-CPR (SfM GT) [34]0.7/0.200.9/0.320.6/0.361.2/0.321.3/0.310.9/0.252.2/0.611.1/0.34
ACE + GS-CPR (SfM GT) [34]0.5/0.150.6/0.250.4/0.280.9/0.261.0/0.230.7/0.171.4/0.420.8/0.25
RAPref (Ours, DSLAM GT)2.83/1.432.07/1.231.54/1.872.55/1.204.59/1.544.23/1.833.48/1.233.04/1.48
RAPref (Ours, SfM GT)0.33/0.110.51/0.210.40/0.290.59/0.170.83/0.200.50/0.121.11/0.320.61/0.20
", + "bbox": [ + 174, + 309, + 823, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "narrowing the domain gap via the discriminator, which is also optimized with Adam [26], using a learning rate of $10^{-4}$ and betas set to (0.5, 0.999). More details about training are in supplemental materials. For generating random views, we apply random normalized perturbations to each training pose: $\\delta t = 20\\mathrm{cm}$ and $\\delta r = 10^{\\circ}$ for indoor scenes, and $\\delta t = 150\\mathrm{cm}$ and $\\delta r = 4^{\\circ}$ for outdoor scenes.", + "bbox": [ + 88, + 592, + 482, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To allow for comparison with SCR methods and leverage 3DGS's efficient rendering for PPR, we extend the APR pipeline with match-based refinement similar to GS-CPR [34], denoted as $\\mathrm{RAP}_{\\mathrm{ref}}$ . At test time, RAP's initial pose is used to render an RGB-D image via 3DGS. Together with MASt3R [27], we can obtain 2D-3D correspondences to perform RANSAC-PnP [16, 17], resulting in a refined pose. More details are in supplementary materials.", + "bbox": [ + 88, + 699, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Results", + "text_level": 1, + "bbox": [ + 89, + 832, + 277, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cambridge Landmarks [24]. In the challenging outdoor Cambridge Landmarks dataset (Table 1), our RAP reduces both translation and rotation errors across all scenes by over", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$30\\%$ compared to other APR-only methods. The visualization in Fig. 3 shows that our method produces fewer outliers than DFNet [8]. In the three larger-scale scenes with significant appearance diversity (College, Church, and Court), rotation error is even halved compared to DFNet. Table 1 also shows the effectiveness of our $\\mathrm{RAP}_{\\mathrm{ref}}$ in further reducing pose errors through refinement. $\\mathrm{RAP}_{\\mathrm{ref}}$ outperforms $\\mathrm{CoordiNet} + \\mathrm{LENS}$ [42], which assumes a continuous trajectory when an Extended Kalman Filter [21] is required for refinement [41]. $\\mathrm{RAP}_{\\mathrm{ref}}$ even surpasses ACE [4] and its post-refinement variant, $\\mathrm{ACE} + \\mathrm{GS}$ -CPR [34], despite GS-CPR manually masking dynamic objects when building 3DGS. This demonstrates the strong representation capability of our appearance-varying 3DGS with deblurring.", + "bbox": [ + 511, + 592, + 906, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MARS [28]. Autonomous driving scenarios present unique challenges, including moving objects and frequent changes in lighting conditions, as illustrated in Fig. 4. Our RAP demonstrates effective and robust performance across four challenging scenes, as shown in Table 2, achieving an average of $28~\\mathrm{cm} / 0.60^{\\circ}$ localization error. This significantly", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26853", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ca43c06e3f28ee1d55e70a74d6fa13ba8fa2d9ae2b53259b4ab5bff36ff623ac.jpg", + "table_caption": [ + "Table 5. Ablation study." + ], + "table_footnote": [], + "table_body": "
Setupon ShopTrans. (cm) ↓Rot. (°) ↓
I (Baseline): φ = VGG161745.45
II: φ = Efficient-B01034.64
III: II + Pose Aug.753.52
IV: III + Appearance Aug.603.14
V: IV + Decoder (ConvNet)522.51
VI: V + Decoder (Transformer)401.98
VII (Ours): VI + Discriminator331.48
", + "bbox": [ + 145, + 107, + 426, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "outperforms the baseline† PoseNet [24]. With one-shot refinement, our $\\mathrm{RAP}_{\\mathrm{ref}}$ further reduces outdoor localization errors to below $10~\\mathrm{cm}$ .", + "bbox": [ + 89, + 229, + 482, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aachen Day-Night [48]. Benefiting from appearance augmentation, our RAP significantly reduces the localization error from 134 unit $/75.99^{\\circ}$ to 130 unit $/13.70^{\\circ}$ , outperforming other APR [8, 24] and SCR [4, 61] baselines, as shown in Table 3. This demonstrates the effectiveness of appearance diversity in handling extreme lighting changes.", + "bbox": [ + 89, + 275, + 482, + 366 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "7-Scenes [54]. As shown in Table 4, our RAP reduces translation error by $50\\%$ ( $10\\mathrm{cm} \\rightarrow 5\\mathrm{cm}$ ) and rotation error by $32.72\\%$ ( $3.24^{\\circ} \\rightarrow 2.18^{\\circ}$ ) on average compared to previous state-of-the-art single-frame APR methods. The only exception is Heads, where the rotation error is suboptimal. This scene consists of just two sequences—one for training and one for testing—potentially limiting the effectiveness of our augmentation in capturing scene variability. Meanwhile, $\\mathrm{RAP}_{\\mathrm{ref}}$ further reduces localization error below 1 cm with one-shot refinement using our 3DGS. It also surpasses ACE [4] and its post-refinement variant, $\\mathrm{ACE} + \\mathrm{GSCPR}$ [34]. Qualitative examples are shown in Fig. 5.", + "bbox": [ + 89, + 366, + 482, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 559, + 243, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on the validation set of Shop in the Cambridge Landmarks dataset to investigate the impact of all the components in our RAP. Setup I, our baseline, consists of the same components as in PoseNet [24] and has been retrained for our experiments. In Setup II, we replace the feature extraction from VGG16 [55] to Efficient-B0 [36], which enhances performance due to its superior feature representation, while they both exhibit poor performance due to the lack of data synthesis. In Setup III and IV, we explore the effectiveness of the designed pose augmentation and appearance augmentation, which bring notable improvements: translation error reduces from $103~\\mathrm{cm}$ to $75~\\mathrm{cm}$ , and rotation error from $3.52^{\\circ}$ to $3.14^{\\circ}$ . In Setup V and VI, we add regular convolutional layers and Pose Transformer between feature extraction and pose regression. Both improve performance due to the increasing parameters, but the Transformer achieves superior results by effectively handling long-term dependencies through attenuating the effect of noise.", + "bbox": [ + 89, + 580, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5eff30fa95553b18efeb01e66b48e3fb59a73df1e300fb37fcd0488624d81f92.jpg", + "image_caption": [ + "Figure 6. Visualization of the training set distribution and results on St. George's Basilica Building [49]. The red hollow spheres, centered on the real images in the training set, indicate the potential locations of all synthetic images during training." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 906, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion mechanisms. Finally, in Setup VII, our adversarial discriminator effectively reduces the syn-to-real domain gap, allowing the model to learn better pose regression features from synthetic data and further reduce localization error.", + "bbox": [ + 511, + 335, + 905, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussion on Data Synthesis", + "text_level": 1, + "bbox": [ + 511, + 409, + 772, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Emerging Generalizability. Previously, APR has been understood to implicitly learn image retrieval [49], lacking the ability to successfully interpolate between training samples and generalize beyond them. To investigate how APR training is affected by increasing synthetic data, we trained RAP on the St. George's Basilica Building [49] and visualized the results in Fig. 6. Here, the translation perturbation was set to $\\delta t = 350$ cm and the rotation perturbation to $\\delta r = 60^{\\circ}$ . Notably, the test set contained two regions entirely uncovered by the training set. Despite this, the model still closely predicts the test camera poses, demonstrating generalization ability beyond the original training positions.", + "bbox": [ + 511, + 431, + 905, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also learn from our experiments that reducing the rotation perturbation, such that the overlap between test and training views remains minimal, leads to high localization error. This is because the translation and rotation parameter space is inherently a SE(3) manifold. Even if the translation remains fixed, significant rotation changes result in entirely different visual content in the images, naturally preventing the model from estimating poses of such unseen views, which correspond to a large distance on the SE(3) manifold. Therefore, enabling generalization across a broader range of space is an important direction for future work.", + "bbox": [ + 511, + 613, + 905, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analyzing Generalization Boundaries. To evaluate the model's generalizability, we designed an experiment introducing a \"void zone\" centered on the test camera, where all real and synthetic data within this zone were excluded. The void zone was progressively expanded to determine the critical threshold at which the localization performance declines most significantly. Specifically, for Shop, we used $100\\%$ of the training set to ensure complete scene cov", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ DFNet [8] results are omitted as we were unable to successfully train its NeRF component, likely due to the need for manual scene scaling within $[- \\pi, \\pi]$ , which is tedious for diverse outdoor scenes.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26854", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9dcbd3d6adcc93c60cb9bcfd1089087c573a055c079cb1b0b346da363b0fb5dc.jpg", + "table_caption": [ + "Table 6. Exploring the generalization boundaries of the model with synthetic data. Green, blue, and red percentages indicate the relative change in localization error (Med Err) compared to the scenario without a void zone." + ], + "table_footnote": [], + "table_body": "
w/o Void Zonew/ Void Zone (cm/°)
10/0.520/130/1.550/280/2.5
Med Err ↓ (rel. change)33/1.2630/1.3432/1.3240/1.8439/2.0749/2.20
0%/0%-9%/6.3%-3%/4.7%21%/46.0%18%/64.3%48%/74.6%
Avg Err ↓41/1.5138/1.7541/1.6351/2.4048/2.4061/2.84
Max Err ↓155/4.52147/6.56219/6.01192/8.38246/9.80242/13.12
Min Err ↓3/0.178/0.204/0.227/0.304/0.168/0.60
", + "bbox": [ + 94, + 148, + 486, + 234 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/54cb0580225220446d27947504acf5844afdcd38e62421dc34ba3b14832f4747.jpg", + "table_caption": [ + "Table 7. Ablation on different pose augmentation policies." + ], + "table_footnote": [], + "table_body": "
MethodsCollegeHospitalShopChurchAverage
RAP (LENS [42])73/1.15126/1.8771/3.37128/3.50100/22.47
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
", + "bbox": [ + 114, + 267, + 459, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erage, with void zone ranges set as [10/0.5, 20/1, 30/1.5, 50/2, 80/2.5, 100/3] $(\\mathrm{cm} / {}^{\\circ})$ . The results in Table 6 demonstrate a stepwise decline in performance. Initially, expanding the void zone has minimal impact on localization accuracy. However, at $30~\\mathrm{cm} / 1.5^{\\circ}$ , a sharp decrease in performance marks the model's generalization boundary.", + "bbox": [ + 89, + 325, + 482, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pose Augmentation Policy. We conduct experiments using a modified version of RAP, with the pose augmentation approach identical to that of LENS [42], as shown in the Table 7, where our method obtains superior performance. This may be because, although LENS's pose augmentation policy covers a broader spatial area than the training set, its synthetic data may have lower NVS image quality in many unseen regions, which could negatively affect APR training. Density of Training Data. As shown in Table 8, our method with the proposed augmentation significantly reduces errors as the density of real training data increases from $20\\%$ to $80\\%$ . However, the localization accuracy remains almost unchanged from $80\\%$ to $100\\%$ , as the scene is already sufficiently covered. Notably, using $100\\%$ of the training data without augmentation can result in a significantly higher maximum error in translation, nearly double that with only $20\\%$ of the training data with augmentation, despite its limited spatial coverage. This suggests that our augmentation method successfully prevents overfitting to the training data, improving generalization to the test set.", + "bbox": [ + 89, + 416, + 482, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quality of Training Data. We evaluate the impact of synthetic image quality on model performance in Table 9, using $20\\%$ and $50\\%$ of the real data for pose regression. For Shop, it is evident that fewer training samples in 3DGS result in lower-quality rendered views (as indicated by lower PSNR), leading to suboptimal localization performance, particularly for rotation. Surprisingly, localization performance using only $20\\%$ of the data for training suboptimal 3DGS and pose regression surpasses the results obtained with $100\\%$ of the data without augmentation, as shown in Table 8. This experiment confirms the need for a robust NVS model and the proposed augmentation method in APR training.", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eafafd4cd513f8ac6d8bc20943b9247540589b4778c4f8023e3d6934dea90b05.jpg", + "table_caption": [ + "Table 8. Impact of the density of real training data. Our augmentation improves the model's ability to generalize across the entire scene, although this effect has an upper limit." + ], + "table_footnote": [], + "table_body": "
Training Pose %w/ Appearance & Pose Aug. (cm/°)w/o Aug. (cm/°)
100%80%60%40%20%100%50%
Med Err ↓33/1.2632/1.2737/1.9057/2.2387/3.6598/3.75104/4.17
Avg Err ↓41/1.5140/1.5048/2.1762/2.8191/4.65128/4.49139/5.33
Max Err ↓155/4.52158/4.09193/9.39230/11.06231/15.45490/20.73500/21.02
Min Err ↓3/0.177/0.207/0.186/0.3812/0.4613/0.639/0.48
", + "bbox": [ + 517, + 136, + 903, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/45e0359316464b9bb281168b90438162a72ea8ccffffd17c0e0c36956d5d68cb.jpg", + "table_caption": [ + "Table 9. Impact of synthetic image quality. Training with higher-quality synthetic images from advanced NVS models enhances localization performance." + ], + "table_footnote": [], + "table_body": "
3DGS PerformanceLocalization Performance (cm/°)
% Images (Train)PSNR ↑ (Train)PSNR ↑ (Test)% Images (Train)Med Err ↓Avg Err ↓Max Err ↓Min Err ↓
20%29.0815.9820%58/3.5968/4.31211/12.1914/0.51
20%29.0815.9850%43/2.4755/3.26196/21.117/0.40
50%26.8817.5550%37/1.8848/2.37184/10.179/0.52
100%24.6018.3050%35/1.6441/2.12130/11.074/0.38
", + "bbox": [ + 517, + 266, + 903, + 348 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8ce8de5eb20aa37b479c0eadb2d4cbb095246ba575b0ae668953b8ae2c8f858f.jpg", + "table_caption": [ + "Table 10. Ablation on different 3D representations." + ], + "table_footnote": [], + "table_body": "
MethodsCollegeHospitalShopChurchAverage
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90
DFNetGS102/2.31137/8.0877/3.92123/4.68110/4.75
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
", + "bbox": [ + 550, + 380, + 867, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different 3D Representations. We evaluate the impact of different 3D representations on the performance in Table 10. Trivially replacing NeRF with 3DGS in existing frameworks degrades performance due to 3DGS's inferior 3D consistency. This shows the effectiveness of our proposed joint training paradigm in RAP, which better utilizes diverse synthetic data to learn appearance-invariant features, rather than naively transferring from NeRF to 3DGS.", + "bbox": [ + 511, + 450, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 587, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Summary. We address absolute pose regression with a robust two-branch joint training framework based on Transformer, coupled with an efficient data synthesis pipeline leveraging 3D Gaussian Splats (3DGS) to synthesize numerous posed images with diverse appearances as additional supervision. Our RAP achieves state-of-the-art localization performance, even under challenging appearance variations. Moreover, we thoroughly investigate the impact of synthesizing diverse data and present a novel perspective on APR: generalizability can emerge if the learning gap in APR is effectively addressed together with diverse data. We believe our RAP could be a promising starting point, and the experiments presented in the paper can provide useful insights for future research in this field.", + "bbox": [ + 511, + 612, + 906, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part through NSF grants 2238968, 2121391, and 2024882, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025).", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26855", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Joydeep Biswas and Manuela Veloso. Depth camera based indoor mobile robot localization and navigation. In 2012 IEEE International Conference on Robotics and Automation, pages 1697-1702, 2012. 1", + "[2] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5847-5865, 2021. 2, 5, 6", + "[3] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6684-6692, 2017. 1, 5", + "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044-5053, 2023. 1, 2, 4, 5, 6, 7", + "[5] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2018. 1, 2, 5, 6", + "[6] David M. Chen, Georges Baatz, Kevin Koser, Sam S. Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, Bernd Girod, and Radek Grzesczuk. City-scale landmark identification on mobile devices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2011. 2", + "[7] Shuai Chen, Zirui Wang, and Victor Prisacariu. Directposenet: Absolute pose regression with photometric consistency. In 2021 International Conference on 3D Vision (3DV), 2021. 2", + "[8] Shuai Chen, Xinghui Li, Zirui Wang, and Victor A Prisacariu. Dfnet: Enhance absolute pose regression with direct feature matching. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[9] Shuai Chen, Yash Bhalgat, Xinghui Li, Jia-Wang Bian, Kejie Li, Zirui Wang, and Victor Adrian Prisacariu. Neural refinement for absolute pose regression with feature synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20987-20996, 2024. 1, 4, 5, 6", + "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 6", + "[11] Manuela Chessa, Chiara Bassano, and Fabio Solari. Detection and localization of changes in immersive virtual reality. In International Conference on Image Analysis and Processing, pages 121-132, 2023. 1", + "[12] Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. Swag: Splitting in the wild images with appearance-conditioned gausians. In European Conference on Computer Vision, pages 325-340, 2025. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3", + "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 2", + "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 6", + "[17] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8):930–943, 2003. 2, 6", + "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1", + "[19] Hugo Germain, Daniel DeTone, Geoffrey Pascoe, Tanner Schmidt, David Novotny, Richard Newcombe, Chris Sweeney, Richard Szeliski, and Vasileios Balntas. Feature query networks: Neural surface description for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5071-5081, 2022. 1, 5, 6", + "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[21] Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. In Signal Processing, Sensor Fusion, and Target Recognition VI, pages 182 - 193, 1997. 6", + "[22] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[23] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5974-5983, 2017. 4", + "[24] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In IEEE international conference on computer vision, pages 2938-2946, 2015. 1, 2, 4, 5, 6, 7", + "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26856", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "radiance field rendering. ACM Transaction on Graphics, 42 (4), 2023. 1, 2, 3", + "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 6", + "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 6", + "[28] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 4, 5, 6", + "[29] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems, 2024. 2", + "[30] Jingyu Lin, Jiaqi Gu, Bojian Wu, Lubin Fan, Renjie Chen, Ligang Liu, and Jieping Ye. Learning neural volumetric pose features for camera localization. In European Conference on Computer Vision, 2024. 1, 2, 3, 4, 5, 6", + "[31] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5166-5175, 2024. 2", + "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 2", + "[33] Changkun Liu, Shuai Chen, Yukun Zhao, Huajian Huang, Victor Prisacariu, and Tristan Braud. Hr-apr: Apr-agnostic framework with uncertainty estimation and hierarchical refinement for camera relocalisation. In 2024 IEEE International Conference on Robotics and Automation, pages 8544–8550, 2024. 1, 5, 6", + "[34] Changkun Liu, Shuai Chen, Yash Sanjay Bhalgat, Siyan HU, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient camera pose refinement via 3d gaussian splatting. In The Thirteenth International Conference on Learning Representations, 2025. 5, 6, 7", + "[35] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[36] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE International Conference on Computer Vision, pages 2372-2381, 2017. 5, 7", + "[37] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 2794-2802, 2017. 4", + "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1", + "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1", + "[41] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Coordinate: uncertainty-aware pose regressor for reliable vehicle localization. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2229-2238, 2022. 2, 5, 6", + "[42] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In Conference on Robot Learning, pages 1347-1356, 2022. 1, 2, 5, 6, 8", + "[43] Arthur Moreau, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Crossfire: Camera relocalization on self-supervised features from an implicit representation. In IEEE/CVF International Conference on Computer Vision, pages 252-262, 2023. 1, 5, 6", + "[44] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In IEEE international conference on computer vision, pages 3456-3465, 2017. 2", + "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019.", + "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020.", + "[47] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2016. 2", + "[48] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6 dof outdoor visual localization in changing conditions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 4, 5, 7", + "[49] Torsten Sattler, Qunjie Zhou, Marc Pollefeys, and Laura Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3302-3312, 2019. 1, 2, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26857", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 4", + "[51] Yoli Shavit and Yosi Keller. Camera pose auto-encoders for improving pose regression. In European Conference on Computer Vision, pages 140–157, 2022. 5, 6", + "[52] Yoli Shavit, Ron Ferens, and Yoshi Keller. Paying attention to activation maps in camera pose regression. arXiv preprint arXiv:2103.11477, 2021. 2", + "[53] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 1, 5, 6", + "[54] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 4, 6, 7", + "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 7", + "[56] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM Transactions on Graphics, pages 835-846, 2006. 2", + "[57] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7199-7209, 2018. 2", + "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114, 2019. 3", + "[59] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The unreasonable effectiveness of pretrained features for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12798, 2024. 1, 5, 6", + "[60] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 6", + "[61] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 5, 6, 7", + "[62] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. In ACM in Computer Graphics and Interactive Techniques, 2024. 2, 3", + "[63] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, and Zhaopeng Cui. Pnerfloc: Visual localization with point-based neural radiance fields. In AAAI Conference on Artificial Intelligence, pages 7450-7459, 2024. 1, 5", + "[65] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In IEEE International Conference on Computer Vision, pages 2075-2083, 2015. 5", + "[66] Qunjie Zhou, Maxim Maximov, Or Litany, and Laura Leal-Taixe. The nefrect match: Exploring nef features for visual localization. arXiv preprint arXiv:2403.09577, 2024. 1, 5" + ], + "bbox": [ + 516, + 90, + 903, + 233 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26858", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_model.json b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_model.json new file mode 100644 index 0000000000000000000000000000000000000000..42dfe55a28d0160e0a91a3e831bc635117f242d4 --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Adversarial Exploitation of Data Diversity Improves Visual Localization" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.18, + 0.831, + 0.201 + ], + "angle": 0, + "content": "Sihang Li* Siqi Tan* Bowen Chang Jing Zhang Chen Feng Yiming Li" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.206, + 0.584, + 0.224 + ], + "angle": 0, + "content": "New York University" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.226, + 0.684, + 0.241 + ], + "angle": 0, + "content": "{sihangli, tan.k, cfeng, yimingli}@nyu.edu" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.244, + 0.616, + 0.258 + ], + "angle": 0, + "content": "https://ai4ce.github.io/RAP" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.293, + 0.327, + 0.309 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Visual localization, which estimates a camera's pose within a known scene, is a fundamental capability for autonomous systems. While absolute pose regression (APR) methods have shown promise for efficient inference, they often struggle with generalization. Recent approaches attempt to address this through data augmentation with varied viewpoints, yet they overlook a critical factor: appearance diversity. In this work, we identify appearance variation as the key to robust localization. Specifically, we first lift real 2D images into 3D Gaussian Splats with varying appearance and deblurring ability, enabling the synthesis of diverse training data that varies not just in poses but also in environmental conditions such as lighting and weather. To fully unleash the potential of the appearance-diverse data, we build a two-branch joint training pipeline with an adversarial discriminator to bridge the syn-to-real gap. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, reducing translation and rotation errors by \\(50\\%\\) and \\(33\\%\\) on indoor datasets, and \\(38\\%\\) and \\(44\\%\\) on outdoor datasets. Most notably, our method shows remarkable robustness in dynamic driving scenarios under varying weather conditions and in day-to-night scenarios, where previous APR methods fail." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.688, + 0.222, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Visual localization, the task of calculating a 6-DoF camera pose—its translation and rotation—based on a query image within a given environment, is essential for various applications, including robotics [1], autonomous vehicles [18], and virtual reality [11]. Besides traditional geometry-based approaches, recent learning-based visual localization methods adopt absolute pose regression (APR) [5, 8, 24, 53], scene coordinate regression (SCR) [3, 4, 43, 61], or post pose refinement (PPR) [9, 19, 33, 43, 59, 64, 66]. SCR methods focus on learning-based 2D-3D correspondences followed by subsequent Perspective-n-Point (PnP) for pose estimation." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.296, + 0.907, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.593 + ], + "angle": 0, + "content": "Figure 1. We propose RAP, a novel pipeline to train robust APR models. We lift real-world 2D images into 3D Gaussian Splats [25] to synthesize images with diverse appearances and poses, improving model generalizability. We also introduce an adversarial discriminator, mitigating the syn-to-real gap to learn robust features. Together, we achieve state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.748 + ], + "angle": 0, + "content": "PPR methods heavily rely on a pose prior, usually obtained from image retrieval, followed by iterative refinement. In contrast, APR methods employ a supervised framework to train a regression neural network on image-pose pairs, enabling direct pose estimation during inference. APR offers faster runtime and lower error in challenging scenes with sparse views, significant lighting variations, or numerous dynamic objects, making it a promising method for ensuring robustness in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite promises, there is a performance gap in localization accuracy between APR and other methods. A well-known pivotal work [49] attributes this to APR performing image-based memorization, i.e., retrieving poses seen during training. Driven by this crucial finding, to improve such memorization while avoiding the need for denser real-world training samples, recent methods leverage Neural Radiance Fields (NeRF) [40] to synthesize additional posed images for APR training [8, 30, 42]. LENS [42] tried to employ appearance perturbation using NeRF-W [39], but found the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.368, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26848" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "improvements minimal [42]. As appearance augmentation is common in other learning tasks, why does it fail in APR?" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.483, + 0.348 + ], + "angle": 0, + "content": "We hypothesize a learning gap: Limitations in previous training pipelines prevented the effective use of diverse data to boost performance. Artifacts always exist in images rendered by common novel view synthesis (NVS) methods, which might disturb the model feature space. Inspired by generative adversarial networks (GAN) [20], where a discriminator is trained to distinguish between real and generated samples, we propose adversarial training for APR, designing a discriminator to align the features of synthetic and real images, thereby reducing the syn-to-real domain gap and mitigating the impact of rendering artifacts. Augmentation quality also matters. To efficiently synthesize diverse high-quality images with controllable varying appearance, we extend the vanilla 3D Gaussian Splitting (3DGS) [25] to appearance-varying 3DGS with deblurring ability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.349, + 0.485, + 0.575 + ], + "angle": 0, + "content": "These form our two-branch joint training framework for robust absolute pose regression (RAP). The first branch coarsely trains our Transformer-based pose regressor with both real data and data synthesized at the original real pose, together with an adversarial discriminator to reduce the synto-real domain gap. The second branch progressively generates randomly perturbed poses and appearances, providing additional supervision to the same APR Transformer. Through extensive experiments, we demonstrate that exploiting data diversity using adversarial training significantly increases localization accuracy in APR. Meanwhile, our results indicate that APR consistently benefits from more diverse visual data, and we observe clear signs of a more generalizable APR emerging with its localization performance cannot be explained merely by memorization." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.575, + 0.414, + 0.588 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.59, + 0.483, + 0.65 + ], + "angle": 0, + "content": "- We identify the crucial role of appearance diversity for APR, and develop a 3DGS-based appearance-varying data augmentation framework to efficiently generate diverse synthetic data with controllable lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.651, + 0.483, + 0.71 + ], + "angle": 0, + "content": "- We propose an adversarial discriminator to reduce the syn-to-real gap. Together with progressive data synthesis, we form a robust two-branch joint training pipeline that fully unleashes the power of data diversity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.711, + 0.483, + 0.771 + ], + "angle": 0, + "content": "- We conduct extensive experiments showing our method outperforms state-of-the-art approaches on challenging datasets with significant appearance change. Ablation studies further analyze key factors affecting performance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.59, + 0.483, + 0.771 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.241, + 0.792 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Visual Localization. Visual localization aims to estimate a camera's translation and rotation within a 3D scene. Traditional geometry-based methods [6, 15, 32, 35, 44-47, 57] accomplish this by using point clouds and a reference image database, relying on stored descriptors and image retrieval to establish 2D-3D correspondences. In contrast, scene coordinate regression (SCR) methods [2-4, 61] em" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "bed map information within neural networks to directly predict 2D-3D correspondences. Both approaches generally require PnP [17] and RANSAC [16] to output camera poses at test time, which adds additional computation cost. Alternatively, absolute pose regression (APR) [5, 7, 22, 24, 41, 52] aims to directly regress the camera pose from a query image using neural networks. Although the performance is suboptimal compared with geometry-based methods, APR remains a promising approach due to its fast inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.529 + ], + "angle": 0, + "content": "Data Augmentation for Pose Regression. End-to-end pose regression methods rely heavily on the amount and diversity of training data. Previous work [49] shows that APR implicitly learns image retrieval in the given environment. Therefore, the following works LENS [42], DFNet [8] and PMNet [30] enhance APR performance by spatially enriching training views with NeRF. However, these approaches fail to address the generalizability of APR models and exhibit several limitations: (1) The efficiency of training and novel view synthesis (NVS) in NeRF is severely restricted, hindering scalability. (2) They limit NVS to geometric (pose) transformations while neglecting photometric (appearance) variations, thereby decreasing APR robustness to changes in visual appearance. (3) The augmented data is underutilized in their learning frameworks, leaving its potential for improving APR largely untapped. Differently, our framework switches to 3DGS [25] as the scene representation to efficiently generate novel posed images with controllable appearances and introduce adversarial training to unleash the power of such diverse data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.905, + 0.832 + ], + "angle": 0, + "content": "Handling and Synthesizing Challenging Scenarios. Visual localization often encounters unstructured photo collections [56], where visual appearance varies due to moving objects, lighting changes, and inconsistent camera exposure settings. To tackle these in-the-wild challenges, NeRF-W [38] uses per-image transient and appearance embeddings. In 3DGS [25], VastGaussian [31] applies a CNN to 3DGS outputs but still struggles with significant appearance variations. SWAG [12] mitigates this issue by storing appearance information in an external hash-grid-based implicit field, while GS-W [63] enhances flexibility by separating intrinsic and dynamic appearance features for each Gaussian point. 3DGM [29] leverages consensus across multiple sequences as the self-supervision signal to remove transient and moving objects without human annotations. Deblur-GS [62] addresses motion blur—another challenge in localization datasets—by modeling camera motion to yield sharper edges in rendered scenes. Our method incorporates appearance modeling and edge refinement to handle and synthesize diverse indoor, outdoor, and driving scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.605, + 0.851 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.855, + 0.755, + 0.87 + ], + "angle": 0, + "content": "3.1. Pre-Processing with 3DGS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "A robust pose regressor should focus on intrinsic scene attributes, not appearance variations. Therefore, we first" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26849" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.905, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Pipeline of RAP. We lift multiple RGB video sequences into 3D Gaussian Splats, which serve as our data engine. The branch-1 (see Sec. 3.3.1) inputs paired real and synthetic images to regress poses, with a discriminator to bridge the syn-to-real gap. The branch-2 (see Sec. 3.3.2) generates views with novel poses and appearances, which are fed into the same pose regressor as additional supervision." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.339, + 0.485, + 0.55 + ], + "angle": 0, + "content": "To further analyze the behavior of the \\(K\\) in the \\(K \\times 16 \\times 3\\) system, we assume that the camera is located at the center of the image. The camera is positioned at the center of the image by a camera position vector \\(\\mathbf{C} = [0, 1]\\). To achieve this, we use the following procedure to generate the camera positions:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.564, + 0.483, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {C}} = \\operatorname {M L P} (\\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\omega \\boldsymbol {\\mathcal {E}}, \\theta), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.623 + ], + "angle": 0, + "content": "where \\(\\omega\\) is the blending weight that controls the dynamic appearance of the rendered image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Another significant challenge in visual localization is motion blur, often caused by slow shutter speeds during video capture, leading to pose ambiguity and degraded rendering quality, further decreasing localization accuracy. Inspired by Deblur-GS [62], we model camera motion blur as the inverse of scene motion, i.e., the transformation in Gaussian position denoted by \\(\\mathcal{T} \\in \\mathrm{SE}(3)\\). For each training image, we sample a certain time step along a linear trajectory with a sampling weight \\(\\phi \\in \\mathbb{R}^n\\) and blend them to compute loss \\(\\mathcal{L}\\) with the original blurry image \\(\\mathcal{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}\\), optimizing \\(\\mathcal{T}\\), \\(\\phi\\), \\(\\mathcal{C}\\) and other 3DGS parameters \\(\\Theta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\underset {\\phi , \\boldsymbol {\\mathcal {T}}, \\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\boldsymbol {\\mathcal {E}}, \\boldsymbol {\\Theta}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\mathcal {I}} _ {b}, \\sum_ {i = 1} ^ {n} \\phi_ {i} \\operatorname {R e n d e r} \\left(\\boldsymbol {\\mathcal {T}} _ {i} (\\boldsymbol {\\mu}), \\boldsymbol {\\mathcal {C}}, \\boldsymbol {\\Theta}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where the details of \\(\\mathcal{L}\\) are in supplemental materials. After training, our 3DGS can efficiently render posed images given \\(\\theta\\) and \\(\\omega\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.338, + 0.787, + 0.354 + ], + "angle": 0, + "content": "3.2. Architecture of Pose Regressor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Given a set of images and their associated camera poses \\(\\{(I_i, P_i)\\}_{i=1}^n\\), our goal is to train a neural network to directly output a homogeneous camera pose \\(P \\in \\mathbb{R}^{3 \\times 4}\\) for a query image \\(I \\in \\mathbb{R}^{H \\times W \\times C}\\). Our network architecture is shown as the pose regressor in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Feature Extractor. Pose regression networks typically extract features using a common backbone \\(\\varphi\\), such as VGG [55] or EfficientNet [58], leveraging multiple deeper layers for translation and rotation regression:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.505, + 0.905, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\varphi (\\boldsymbol {I}) = \\left\\{\\mathcal {F} _ {0} (\\boldsymbol {I}), \\dots , \\mathcal {F} _ {N - 1} (\\boldsymbol {I}), \\mathcal {F} _ {N} (\\boldsymbol {I}) \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.906, + 0.575 + ], + "angle": 0, + "content": "\\(\\mathcal{F}_{*}(\\cdot)\\) denotes features extracted from the \\(*\\)-th layer of a backbone with \\(N\\) layers. \\(\\mathcal{F}_t(\\pmb {I})\\) and \\(\\mathcal{F}_r(\\pmb {I})\\) denote features for translation and rotation regression, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.727 + ], + "angle": 0, + "content": "Pose Transformer. Unlike CNN-based regression models [8, 30], where fine-grained local features can introduce noise and harm performance, we propose Pose Transformer to leverage the strong ability of Vision Transformer (ViT) [14] for modeling long-range dependencies. Each Transformer generates a global token (Trans for translation and Rot for rotation) to provide a comprehensive context for pose regression, inspired by the CLS token in ViT. Given \\(\\mathcal{F}_r(I)\\) and \\(\\mathcal{F}_t(I)\\), the translation token is then concatenated with the flattened input features*:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.733, + 0.905, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\operatorname {C a t} (\\text {F l a t t e n} (\\mathcal {F} _ {t} (\\boldsymbol {I})), \\text {T r a n s}) \\in \\mathbb {R} ^ {\\left(H _ {t} W _ {t} + 1\\right) \\times C _ {t}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.829 + ], + "angle": 0, + "content": "The positional encodings are then added to the flattened feature \\((\\mathsf{PE} + \\widetilde{\\mathcal{F}}_t(\\pmb {I}))\\in \\mathbb{R}^{(H_tW_t + 1)\\times C_t}\\). Multi-head Self-Attention (MSA) is then conducted through a stack of multiple layers with the post-processing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.837, + 0.905, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {F} _ {t} ^ {\\prime}} (\\boldsymbol {I}) = \\operatorname {M S A} \\left(\\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I})\\right) + \\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.86, + 0.848, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\mathrm {L N} (\\mathrm {F F N} (\\mathrm {L N} (\\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I}))) + \\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I})),\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.838, + 0.901 + ], + "angle": 0, + "content": "*We only present the translation regression for simplicity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "where LN indicates layer normalization and FFN denotes the fully connected feed-forward network, consisting of two linear layers with a ReLU. The final output is flattened back to \\((H_{t}W_{t} + 1)\\times c_{t}\\). See supplementary for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.212 + ], + "angle": 0, + "content": "Regression Head. Only the processed translation token, Trans, capturing global features for regression, is fed into the regression head. This regression head consists of two MLPs, each with a hidden layer and GeLU activation:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.217, + 0.482, + 0.235 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {t}} = \\operatorname {L i n e a r} (\\operatorname {G e L U} (\\operatorname {L i n e a r} (\\widehat {\\text {T r a n s}}))). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.483, + 0.271 + ], + "angle": 0, + "content": "The \\(\\hat{t}\\) represents the final prediction for translation. Similarly, we obtain the rotation prediction denoted by \\(\\hat{r}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.278, + 0.417, + 0.294 + ], + "angle": 0, + "content": "3.3. Two-Branch Joint Training Paradigm" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.465, + 0.314 + ], + "angle": 0, + "content": "3.3.1. Branch-1: Aligning Features via Discriminator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.394 + ], + "angle": 0, + "content": "Synthetic images from 3DGS provide novel viewpoints and appearances but often contain artifacts, leading to a syn-to-real domain gap. To align features from rendered and real images of the same pose, we introduce an adversarial training mechanism besides the basic pose regression training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Pose Regression Loss. For basic training, we render the synthetic image \\(I^{\\prime}\\) with the same pose label \\(\\pmb{P}\\) as the real image \\(\\pmb{I}\\), both used as supervision for the pose regressor. The training objective consists of translation loss \\(\\mathcal{L}_t\\) and the rotation loss \\(\\mathcal{L}_r\\), which are measured by the Euclidean distance between the ground truth pose \\(\\pmb{P} = \\{\\hat{t},\\hat{r}\\}\\) and the estimated pose \\(\\widehat{P} = \\{\\hat{t},\\hat{r}\\}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.521, + 0.482, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\| \\boldsymbol {t} - \\hat {\\boldsymbol {t}} \\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.542, + 0.483, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r} = \\left\\| \\boldsymbol {r} - \\frac {\\hat {\\boldsymbol {r}}}{\\| \\hat {\\boldsymbol {r}} \\|} \\right\\| _ {2}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.578, + 0.482, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s e}} ^ {1} = \\mathcal {L} _ {t} \\exp (- s _ {t}) + s _ {t} + \\mathcal {L} _ {r} \\exp (- s _ {r}) + s _ {r}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.631 + ], + "angle": 0, + "content": "where \\( s_t \\) and \\( s_r \\) are learned parameters for balancing the optimization between rotation and translation [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Adversarial Loss. The adversarial training mechanism optimizes the discriminator to distinguish real from rendered image features, while training the feature extractor to fool the discriminator, effectively bridging the domain gaps. To prevent vanishing gradients, we propose a novel adversarial objective for pose regression, inspired by LSGAN [37]:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.727, + 0.482, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {D} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {D i s}} (D) = \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} \\sim p _ {\\mathrm {d a t a}} (\\boldsymbol {I})} \\left[ \\left(D (\\operatorname {A d j} (\\mathcal {F} _ {t} (\\boldsymbol {I}))) - 1\\right) ^ {2} \\right] \\\\ + \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} ^ {\\prime}} \\left[ D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {I} ^ {\\prime}\\right)\\right) ^ {2} \\right], \\right. \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.79, + 0.482, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\underset {G} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {G e n}} (G) = \\frac {1}{2} \\mathbb {E} _ {I ^ {\\prime}} \\left[ \\left(D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(I ^ {\\prime}\\right)\\right) - 1\\right) ^ {2} \\right]. \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here, Adj and \\(\\mathrm{Adj}'\\) are the adjustment layers, consisting of Conv-ReLU-BN layers. The feature extractor \\(\\varphi\\) acts as the generator \\(G\\), while \\(D\\) is the discriminator, composed of several convolutional layers with ReLU activations. More details are in supplemental materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.868, + 0.107 + ], + "angle": 0, + "content": "3.3.2. Branch-2: Training while Synthesizing Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.111, + 0.907, + 0.376 + ], + "angle": 0, + "content": "With the proposed appearance-varying 3DGS, more posed images are generated to enrich the training data for better generalizability. Specifically, our data synthesis is categorized into two dimensions: pose augmentation and appearance augmentation, as illustrated in Fig. 1. For pose augmentation, given a training pose \\( P \\), a perturbed pose \\( P_{\\mathrm{syn}} \\) is generated around \\( P \\) by the translation noise of \\( \\delta t \\) and rotation noise of \\( \\delta r \\). For appearance augmentation, we randomly adjust the appearance of rendered images using random blending weights \\( \\omega \\), and then render the synthetic image \\( I_{\\mathrm{syn}} \\) using the Gaussian Splats trained in Sec. 3.1. The novel image-pose pair \\( (I_{\\mathrm{syn}}, P_{\\mathrm{syn}}) \\), online generated every 20 epochs during training until the validation MSE loss and median errors cease to decrease, serves as additional supervision for the training. Given the estimated pose of the synthesized image denoted by \\( \\widehat{P_{\\mathrm{syn}}} \\), the loss function \\( \\mathcal{L}_{\\mathrm{pose}}^2(\\widehat{P_{\\mathrm{syn}}}, P_{\\mathrm{syn}}) \\) is same as \\( \\mathcal{L}_{\\mathrm{pose}}^1 \\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.382, + 0.684, + 0.397 + ], + "angle": 0, + "content": "3.3.3. Overall Objective" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.402, + 0.771, + 0.417 + ], + "angle": 0, + "content": "The total loss for the pose regressor is:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.429, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {1} \\mathcal {L} _ {\\text {p o s e}} ^ {1} + \\beta_ {2} \\mathcal {L} _ {\\text {p o s e}} ^ {2} + \\beta_ {3} \\left(\\mathcal {L} _ {\\text {G e n}} + \\mathcal {L} _ {\\text {D i s}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.52 + ], + "angle": 0, + "content": "where \\(\\beta_{1},\\beta_{2},\\beta_{3}\\) are loss weights. The total loss will optimize the pose regressor, adjustment layers, and discriminator. Only the pose regressor will be deployed in the inference phase, while the other two components are discarded." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.536, + 0.646, + 0.552 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.561, + 0.682, + 0.577 + ], + "angle": 0, + "content": "4.1. Evaluation Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. We follow previous works [8, 30] to mainly use four scenes in the Cambridge Landmarks dataset [24] with spatial extents around \\(875\\mathrm{m}^2\\). Moreover, we evaluate our method on MARS [28], a self-driving dataset featuring challenges like moving objects, lighting changes, and motion blur. To investigate the robustness of our model under extreme lighting changes, such as the transition from day to night, we also prepared a subset of the Aachen Day-Night dataset [48]. The training data includes images captured using various camera models with differing resolutions, which renders direct evaluation with APR methods infeasible. Thus, we standardized the camera models through center cropping and built a COLMAP [50] model as pose annotations, including 13 nighttime images for evaluation and 246 daytime images for training 3DGS and RAP. We also employ the 7-Scenes dataset [54], which provides seven indoor scenes with volumes spanning \\(1\\mathrm{m}^3 - 18\\mathrm{m}^3\\) and follow the original training and testing splits with more accurate SfM pose annotations [4, 9]. Although it is an indoor dataset, it is still non-trivial as it includes texture-less surfaces, object occlusions, and motion blur." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.912, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.343 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of camera pose estimation errors between a) DFNet [8] and b) our RAP framework across five scenes on the Cambridge Landmarks dataset [24]. Our RAP framework estimates trajectories that more closely follow the ground truth, with significantly reduced rotation and position errors compared to DFNet [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Table 1. Median translation (cm) and rotation \\((^{\\circ})\\) errors on the Cambridge Landmarks dataset [24]. The best results† in pure APR and PPR are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.4, + 0.481, + 0.668 + ], + "angle": 0, + "content": "
MethodsCollegeHospitalShopChurchAverage‡Court
Pure APRPN [24]166/4.86262/4.90141/7.18245/7.95204/6.23683/3.50
MapNet [5]107/1.89194/3.91149/4.22200/4.53163/3.64N/A
MS-Trans. [53]83/1.47181/2.3986/3.07162/3.99128/2.73N/A
PAE [51]90/1.49207/2.5899/3.88164/4.16140/3.03N/A
LENS† [42]33/0.5044/0.9027/1.6053/1.6039/1.20N/A
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90217/4.11
PMNet [30]68/1.97103/1.3158/2.10133/3.7390/2.27N/A
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28115/1.68
SCRDSAC* [2]18/0.321/0.45/0.315/0.615/0.434/0.2
ACE [4]28/0.431/0.65/0.318/0.621/0.543/0.2
GLACE [61]19/0.317/0.44/0.29/0.312/0.319/0.1
PPRFQN-MN [19]28/0.3854/0.8213/0.6358/2.0038/0.964253/39.16
LENS [42]34/0.5445/0.9628/1.6654/1.6640/1.21N/A
CrossFire [43]47/0.743/0.720/1.239/1.437/1.0N/A
NeFeS50 [9]37/0.5452/0.8815/0.5337/1.1435/0.77N/A
HR-APR [33]36/0.5853/0.8913/0.5138/1.1635/0.78N/A
MCLoc [59]31/0.4239/0.7312/0.4526/0.8827/0.62N/A
DFNetGS-CPR [34]23/0.3242/0.7410/0.3627/0.6226/0.51N/A
ACEGS-CPR [34]20/0.2921/0.405/0.2413/0.4015/0.33N/A
DFNetref(Ours)16/0.2421/0.418/0.4210/0.2614/0.3325/0.13
RAPref (Ours)15/0.2518/0.385/0.239/0.2312/0.2723/0.15
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.67, + 0.483, + 0.708 + ], + "angle": 0, + "content": "Since most methods did not report results on Court, it is excluded from the average error calculation. As Coordinet + LENS [42] does not provide open-source code, it is unclear whether any post-processing is used." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Baselines. We first compare our proposed RAP against common APR-only approaches on the four datasets, where PMNet [30] and DFNet [8] are the most related and advanced methods based on data augmentation. We split the remaining methods into two categories based on whether they rely on extra novel view synthesis in test time, including SCR [3, 4, 61] and PPR (Post Pose Refinement) [9, 19, 33, 43, 59, 64, 66], which involves rendering images, querying features in novel views by the initial pose, iterative refinement or sequential refinement [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. First, we optimize our 3DGS for each scene without masking moving objects. We" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.353, + 0.907, + 0.536 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Figure 4. Visualization of \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) on MARS [28]. In each subfigure, a diagonal line separates the \"Predicted\" (rendered from the refined pose) and \"GT\" (ground truth) sections. Smooth alignment along this boundary shows \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) 's improved pose accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.607, + 0.906, + 0.634 + ], + "angle": 0, + "content": "Table 2. Median translation (cm) and rotation \\((^{\\circ})\\) errors on the MARS dataset [28]." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.64, + 0.878, + 0.698 + ], + "angle": 0, + "content": "
Methods“0011”“0015”“0037”“0041”Average
PoseNet [24]149/1.80136/2.34123/1.6075/0.92121/1.67
RAP (Ours)32/0.6137/1.0815/0.3528/0.3528/0.60
RAPref (Ours)8.5/0.138.2/0.208.7/0.097.6/0.118.3/0.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.714, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Table 3. Median translation (COLMAP [65] unit) and rotation \\((^{\\circ})\\) errors on the Aachen Day-Night Dataset [48]." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.747, + 0.912, + 0.793 + ], + "angle": 0, + "content": "
APR-BasedSCR-Based
PoseNet [24]DFNet [8]RAP w/o App.RAPRAPrefACE [4]GLACE [61]
217/74.30174/85.80134/75.99130/13.7067/4.64914/90.50482/36.4
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "then train our RAP network, which uses an Efficient-B0 backbone [36] pre-trained on ImageNet [13], optimized with Adam [26] at a learning rate of \\( 10^{-4} \\). Only the features from the third (reduction_3) and fourth (reduction_4) layers are used respectively for translation and rotation regression, and both layers are utilized for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26852" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.251, + 0.772, + 0.265 + ], + "angle": 0, + "content": "Figure 5. Visualization of the localization errors of \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) on the 7-Scenes dataset [54]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.277, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 4. Quantitative results on the 7-Scenes dataset [54]. The best results in pure APR and PPR are highlighted in bold. DSLAM GT and SfM GT refer to different sets of ground truth. More visualizations and details are in supplemental materials." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.31, + 0.825, + 0.566 + ], + "angle": 0, + "content": "
CategoryMethodsChessFireHeadsOfficePumpkinKitchenStairsAverage
Pure APRPoseNet (PN) [24]32/8.1247/14.429/1.2048/7.6847/8.4259/8.6447/13.8044/10.4
MapNet [5]8/3.2527/11.718/13.317/51.522/4.0223/4.9330/12.121/7.77
AtLoc+ [60]10/3.1826/10.814/11.417/5.1620/3.9416/4.9029/10.219/7.08
MS-Transformer [53]11/4.6624/9.6014/12.217/5.6618/4.4417/5.9417/5.9418/7.28
PAE [51]12/4.9524/9.3114/12.519/5.7918/4.8918/6.1925/8.7419/7.48
CoordiNet + LENS [42]4/1.3811/3.778/5.868/1.989/2.2710/2.2715/3.679/3.07
DFNet [8]5/1.8817/6.456/3.638/2.4810/2.7822/5.4516/3.2912/3.71
PMNet [30]4/1.7010/4.517/4.237/1.9614/3.3314/3.3616/3.6210/3.24
RAP (Ours, DSLAM GT)3/1.417/3.466/6.025/1.976/1.967/2.1810/2.146/2.73
RAP (Ours, SfM GT)2/0.856/2.844/4.525/1.944/1.707/2.1210/1.305/2.18
SCRDSAC [2]0.5/0.170.8/0.280.5/0.341.2/0.341.2/0.280.7/0.212.7/0.781.1/0.34
ACE [4]0.5/0.180.8/0.330.5/0.331.0/0.291.0/0.220.8/0.202.9/0.811.1/0.34
GLACE [61]0.6/0.180.9/0.340.6/0.341.1/0.290.9/0.230.8/0.203.2/0.931.2/0.36
marepo† [10]2.6/1.352.5/1.422.3/2.213.6/1.444.2/1.555.1/1.996.7/1.833.9/1.68
PPRFQN-MN [19]4.1/1.3110.5/2.979.2/2.453.6/2.364.6/1.7616.1/4.42139.5/34.6728/7.3
CrossFire [43]1/0.45/1.93/2.35/1.63/0.82/0.812/1.94.4/1.38
DFNet + NeFeS50 [9]2/0.572/0.742/1.282/0.562/0.552/0.575/1.282.4/0.79
HR-APR [33]2/0.552/0.752/1.452/0.642/0.622/0.675/1.302.4/0.85
MCLoc [59]2/0.83/1.43/1.34/1.35/1.66/1.66/2.04.1/1.43
DFNet + GS-CPR (SfM GT) [34]0.7/0.200.9/0.320.6/0.361.2/0.321.3/0.310.9/0.252.2/0.611.1/0.34
ACE + GS-CPR (SfM GT) [34]0.5/0.150.6/0.250.4/0.280.9/0.261.0/0.230.7/0.171.4/0.420.8/0.25
RAPref (Ours, DSLAM GT)2.83/1.432.07/1.231.54/1.872.55/1.204.59/1.544.23/1.833.48/1.233.04/1.48
RAPref (Ours, SfM GT)0.33/0.110.51/0.210.40/0.290.59/0.170.83/0.200.50/0.121.11/0.320.61/0.20
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.324, + 0.567, + 0.67, + 0.58 + ], + "angle": 0, + "content": "As marepo [10] combines SCR and APR, we classify it as SCR." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.593, + 0.483, + 0.698 + ], + "angle": 0, + "content": "narrowing the domain gap via the discriminator, which is also optimized with Adam [26], using a learning rate of \\(10^{-4}\\) and betas set to (0.5, 0.999). More details about training are in supplemental materials. For generating random views, we apply random normalized perturbations to each training pose: \\(\\delta t = 20\\mathrm{cm}\\) and \\(\\delta r = 10^{\\circ}\\) for indoor scenes, and \\(\\delta t = 150\\mathrm{cm}\\) and \\(\\delta r = 4^{\\circ}\\) for outdoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.7, + 0.483, + 0.822 + ], + "angle": 0, + "content": "To allow for comparison with SCR methods and leverage 3DGS's efficient rendering for PPR, we extend the APR pipeline with match-based refinement similar to GS-CPR [34], denoted as \\(\\mathrm{RAP}_{\\mathrm{ref}}\\). At test time, RAP's initial pose is used to render an RGB-D image via 3DGS. Together with MASt3R [27], we can obtain 2D-3D correspondences to perform RANSAC-PnP [16, 17], resulting in a refined pose. More details are in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.279, + 0.847 + ], + "angle": 0, + "content": "4.2. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Cambridge Landmarks [24]. In the challenging outdoor Cambridge Landmarks dataset (Table 1), our RAP reduces both translation and rotation errors across all scenes by over" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.804 + ], + "angle": 0, + "content": "\\(30\\%\\) compared to other APR-only methods. The visualization in Fig. 3 shows that our method produces fewer outliers than DFNet [8]. In the three larger-scale scenes with significant appearance diversity (College, Church, and Court), rotation error is even halved compared to DFNet. Table 1 also shows the effectiveness of our \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) in further reducing pose errors through refinement. \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) outperforms \\(\\mathrm{CoordiNet} + \\mathrm{LENS}\\) [42], which assumes a continuous trajectory when an Extended Kalman Filter [21] is required for refinement [41]. \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) even surpasses ACE [4] and its post-refinement variant, \\(\\mathrm{ACE} + \\mathrm{GS}\\)-CPR [34], despite GS-CPR manually masking dynamic objects when building 3DGS. This demonstrates the strong representation capability of our appearance-varying 3DGS with deblurring." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "MARS [28]. Autonomous driving scenarios present unique challenges, including moving objects and frequent changes in lighting conditions, as illustrated in Fig. 4. Our RAP demonstrates effective and robust performance across four challenging scenes, as shown in Table 2, achieving an average of \\(28~\\mathrm{cm} / 0.60^{\\circ}\\) localization error. This significantly" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26853" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.212, + 0.09, + 0.362, + 0.103 + ], + "angle": 0, + "content": "Table 5. Ablation study." + }, + { + "type": "table", + "bbox": [ + 0.147, + 0.108, + 0.428, + 0.212 + ], + "angle": 0, + "content": "
Setupon ShopTrans. (cm) ↓Rot. (°) ↓
I (Baseline): φ = VGG161745.45
II: φ = Efficient-B01034.64
III: II + Pose Aug.753.52
IV: III + Appearance Aug.603.14
V: IV + Decoder (ConvNet)522.51
VI: V + Decoder (Transformer)401.98
VII (Ours): VI + Discriminator331.48
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.275 + ], + "angle": 0, + "content": "outperforms the baseline† PoseNet [24]. With one-shot refinement, our \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) further reduces outdoor localization errors to below \\(10~\\mathrm{cm}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.367 + ], + "angle": 0, + "content": "Aachen Day-Night [48]. Benefiting from appearance augmentation, our RAP significantly reduces the localization error from 134 unit \\( /75.99^{\\circ} \\) to 130 unit \\( /13.70^{\\circ} \\), outperforming other APR [8, 24] and SCR [4, 61] baselines, as shown in Table 3. This demonstrates the effectiveness of appearance diversity in handling extreme lighting changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.549 + ], + "angle": 0, + "content": "7-Scenes [54]. As shown in Table 4, our RAP reduces translation error by \\(50\\%\\) (\\(10\\mathrm{cm} \\rightarrow 5\\mathrm{cm}\\)) and rotation error by \\(32.72\\%\\) (\\(3.24^{\\circ} \\rightarrow 2.18^{\\circ}\\)) on average compared to previous state-of-the-art single-frame APR methods. The only exception is Heads, where the rotation error is suboptimal. This scene consists of just two sequences—one for training and one for testing—potentially limiting the effectiveness of our augmentation in capturing scene variability. Meanwhile, \\(\\mathrm{RAP}_{\\mathrm{ref}}\\) further reduces localization error below 1 cm with one-shot refinement using our 3DGS. It also surpasses ACE [4] and its post-refinement variant, \\(\\mathrm{ACE} + \\mathrm{GSCPR}\\) [34]. Qualitative examples are shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.245, + 0.575 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We conduct ablation studies on the validation set of Shop in the Cambridge Landmarks dataset to investigate the impact of all the components in our RAP. Setup I, our baseline, consists of the same components as in PoseNet [24] and has been retrained for our experiments. In Setup II, we replace the feature extraction from VGG16 [55] to Efficient-B0 [36], which enhances performance due to its superior feature representation, while they both exhibit poor performance due to the lack of data synthesis. In Setup III and IV, we explore the effectiveness of the designed pose augmentation and appearance augmentation, which bring notable improvements: translation error reduces from \\(103~\\mathrm{cm}\\) to \\(75~\\mathrm{cm}\\), and rotation error from \\(3.52^{\\circ}\\) to \\(3.14^{\\circ}\\). In Setup V and VI, we add regular convolutional layers and Pose Transformer between feature extraction and pose regression. Both improve performance due to the increasing parameters, but the Transformer achieves superior results by effectively handling long-term dependencies through attenuating the effect of noise." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.907, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Figure 6. Visualization of the training set distribution and results on St. George's Basilica Building [49]. The red hollow spheres, centered on the real images in the training set, indicate the potential locations of all synthetic images during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.398 + ], + "angle": 0, + "content": "tion mechanisms. Finally, in Setup VII, our adversarial discriminator effectively reduces the syn-to-real domain gap, allowing the model to learn better pose regression features from synthetic data and further reduce localization error." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.774, + 0.424 + ], + "angle": 0, + "content": "4.4. Discussion on Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.613 + ], + "angle": 0, + "content": "Emerging Generalizability. Previously, APR has been understood to implicitly learn image retrieval [49], lacking the ability to successfully interpolate between training samples and generalize beyond them. To investigate how APR training is affected by increasing synthetic data, we trained RAP on the St. George's Basilica Building [49] and visualized the results in Fig. 6. Here, the translation perturbation was set to \\(\\delta t = 350\\) cm and the rotation perturbation to \\(\\delta r = 60^{\\circ}\\). Notably, the test set contained two regions entirely uncovered by the training set. Despite this, the model still closely predicts the test camera poses, demonstrating generalization ability beyond the original training positions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.78 + ], + "angle": 0, + "content": "We also learn from our experiments that reducing the rotation perturbation, such that the overlap between test and training views remains minimal, leads to high localization error. This is because the translation and rotation parameter space is inherently a SE(3) manifold. Even if the translation remains fixed, significant rotation changes result in entirely different visual content in the images, naturally preventing the model from estimating poses of such unseen views, which correspond to a large distance on the SE(3) manifold. Therefore, enabling generalization across a broader range of space is an important direction for future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Analyzing Generalization Boundaries. To evaluate the model's generalizability, we designed an experiment introducing a \"void zone\" centered on the test camera, where all real and synthetic data within this zone were excluded. The void zone was progressively expanded to determine the critical threshold at which the localization performance declines most significantly. Specifically, for Shop, we used \\(100\\%\\) of the training set to ensure complete scene cov" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\)DFNet [8] results are omitted as we were unable to successfully train its NeRF component, likely due to the need for manual scene scaling within \\([- \\pi, \\pi]\\), which is tedious for diverse outdoor scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26854" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.145 + ], + "angle": 0, + "content": "Table 6. Exploring the generalization boundaries of the model with synthetic data. Green, blue, and red percentages indicate the relative change in localization error (Med Err) compared to the scenario without a void zone." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.15, + 0.488, + 0.236 + ], + "angle": 0, + "content": "
w/o Void Zonew/ Void Zone (cm/°)
10/0.520/130/1.550/280/2.5
Med Err ↓ (rel. change)33/1.2630/1.3432/1.3240/1.8439/2.0749/2.20
0%/0%-9%/6.3%-3%/4.7%21%/46.0%18%/64.3%48%/74.6%
Avg Err ↓41/1.5138/1.7541/1.6351/2.4048/2.4061/2.84
Max Err ↓155/4.52147/6.56219/6.01192/8.38246/9.80242/13.12
Min Err ↓3/0.178/0.204/0.227/0.304/0.168/0.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.249, + 0.468, + 0.263 + ], + "angle": 0, + "content": "Table 7. Ablation on different pose augmentation policies." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.268, + 0.46, + 0.307 + ], + "angle": 0, + "content": "
MethodsCollegeHospitalShopChurchAverage
RAP (LENS [42])73/1.15126/1.8771/3.37128/3.50100/22.47
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.483, + 0.417 + ], + "angle": 0, + "content": "erage, with void zone ranges set as [10/0.5, 20/1, 30/1.5, 50/2, 80/2.5, 100/3] \\((\\mathrm{cm} / {}^{\\circ})\\). The results in Table 6 demonstrate a stepwise decline in performance. Initially, expanding the void zone has minimal impact on localization accuracy. However, at \\(30~\\mathrm{cm} / 1.5^{\\circ}\\), a sharp decrease in performance marks the model's generalization boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Pose Augmentation Policy. We conduct experiments using a modified version of RAP, with the pose augmentation approach identical to that of LENS [42], as shown in the Table 7, where our method obtains superior performance. This may be because, although LENS's pose augmentation policy covers a broader spatial area than the training set, its synthetic data may have lower NVS image quality in many unseen regions, which could negatively affect APR training. Density of Training Data. As shown in Table 8, our method with the proposed augmentation significantly reduces errors as the density of real training data increases from \\(20\\%\\) to \\(80\\%\\). However, the localization accuracy remains almost unchanged from \\(80\\%\\) to \\(100\\%\\), as the scene is already sufficiently covered. Notably, using \\(100\\%\\) of the training data without augmentation can result in a significantly higher maximum error in translation, nearly double that with only \\(20\\%\\) of the training data with augmentation, despite its limited spatial coverage. This suggests that our augmentation method successfully prevents overfitting to the training data, improving generalization to the test set." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quality of Training Data. We evaluate the impact of synthetic image quality on model performance in Table 9, using \\(20\\%\\) and \\(50\\%\\) of the real data for pose regression. For Shop, it is evident that fewer training samples in 3DGS result in lower-quality rendered views (as indicated by lower PSNR), leading to suboptimal localization performance, particularly for rotation. Surprisingly, localization performance using only \\(20\\%\\) of the data for training suboptimal 3DGS and pose regression surpasses the results obtained with \\(100\\%\\) of the data without augmentation, as shown in Table 8. This experiment confirms the need for a robust NVS model and the proposed augmentation method in APR training." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 8. Impact of the density of real training data. Our augmentation improves the model's ability to generalize across the entire scene, although this effect has an upper limit." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.207 + ], + "angle": 0, + "content": "
Training Pose %w/ Appearance & Pose Aug. (cm/°)w/o Aug. (cm/°)
100%80%60%40%20%100%50%
Med Err ↓33/1.2632/1.2737/1.9057/2.2387/3.6598/3.75104/4.17
Avg Err ↓41/1.5140/1.5048/2.1762/2.8191/4.65128/4.49139/5.33
Max Err ↓155/4.52158/4.09193/9.39230/11.06231/15.45490/20.73500/21.02
Min Err ↓3/0.177/0.207/0.186/0.3812/0.4613/0.639/0.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 9. Impact of synthetic image quality. Training with higher-quality synthetic images from advanced NVS models enhances localization performance." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.267, + 0.905, + 0.349 + ], + "angle": 0, + "content": "
3DGS PerformanceLocalization Performance (cm/°)
% Images (Train)PSNR ↑ (Train)PSNR ↑ (Test)% Images (Train)Med Err ↓Avg Err ↓Max Err ↓Min Err ↓
20%29.0815.9820%58/3.5968/4.31211/12.1914/0.51
20%29.0815.9850%43/2.4755/3.26196/21.117/0.40
50%26.8817.5550%37/1.8848/2.37184/10.179/0.52
100%24.6018.3050%35/1.6441/2.12130/11.074/0.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.362, + 0.871, + 0.376 + ], + "angle": 0, + "content": "Table 10. Ablation on different 3D representations." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.381, + 0.869, + 0.432 + ], + "angle": 0, + "content": "
MethodsCollegeHospitalShopChurchAverage
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90
DFNetGS102/2.31137/8.0877/3.92123/4.68110/4.75
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Different 3D Representations. We evaluate the impact of different 3D representations on the performance in Table 10. Trivially replacing NeRF with 3DGS in existing frameworks degrades performance due to 3DGS's inferior 3D consistency. This shows the effectiveness of our proposed joint training paradigm in RAP, which better utilizes diverse synthetic data to learn appearance-invariant features, rather than naively transferring from NeRF to 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.634, + 0.604 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Summary. We address absolute pose regression with a robust two-branch joint training framework based on Transformer, coupled with an efficient data synthesis pipeline leveraging 3D Gaussian Splats (3DGS) to synthesize numerous posed images with diverse appearances as additional supervision. Our RAP achieves state-of-the-art localization performance, even under challenging appearance variations. Moreover, we thoroughly investigate the impact of synthesizing diverse data and present a novel perspective on APR: generalizability can emerge if the learning gap in APR is effectively addressed together with diverse data. We believe our RAP could be a promising starting point, and the experiments presented in the paper can provide useful insights for future research in this field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part through NSF grants 2238968, 2121391, and 2024882, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26855" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Joydeep Biswas and Manuela Veloso. Depth camera based indoor mobile robot localization and navigation. In 2012 IEEE International Conference on Robotics and Automation, pages 1697-1702, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5847-5865, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6684-6692, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044-5053, 2023. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[5] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[6] David M. Chen, Georges Baatz, Kevin Koser, Sam S. Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, Bernd Girod, and Radek Grzesczuk. City-scale landmark identification on mobile devices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[7] Shuai Chen, Zirui Wang, and Victor Prisacariu. Directposenet: Absolute pose regression with photometric consistency. In 2021 International Conference on 3D Vision (3DV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[8] Shuai Chen, Xinghui Li, Zirui Wang, and Victor A Prisacariu. Dfnet: Enhance absolute pose regression with direct feature matching. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[9] Shuai Chen, Yash Bhalgat, Xinghui Li, Jia-Wang Bian, Kejie Li, Zirui Wang, and Victor Adrian Prisacariu. Neural refinement for absolute pose regression with feature synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20987-20996, 2024. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Manuela Chessa, Chiara Bassano, and Fabio Solari. Detection and localization of changes in immersive virtual reality. In International Conference on Image Analysis and Processing, pages 121-132, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. Swag: Splitting in the wild images with appearance-conditioned gausians. In European Conference on Computer Vision, pages 325-340, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[17] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8):930–943, 2003. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[19] Hugo Germain, Daniel DeTone, Geoffrey Pascoe, Tanner Schmidt, David Novotny, Richard Newcombe, Chris Sweeney, Richard Szeliski, and Vasileios Balntas. Feature query networks: Neural surface description for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5071-5081, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[21] Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. In Signal Processing, Sensor Fusion, and Target Recognition VI, pages 182 - 193, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[22] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[23] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5974-5983, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In IEEE international conference on computer vision, pages 2938-2946, 2015. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26856" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "radiance field rendering. ACM Transaction on Graphics, 42 (4), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[28] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[29] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[30] Jingyu Lin, Jiaqi Gu, Bojian Wu, Lubin Fan, Renjie Chen, Ligang Liu, and Jieping Ye. Learning neural volumetric pose features for camera localization. In European Conference on Computer Vision, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[31] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5166-5175, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[33] Changkun Liu, Shuai Chen, Yukun Zhao, Huajian Huang, Victor Prisacariu, and Tristan Braud. Hr-apr: Apr-agnostic framework with uncertainty estimation and hierarchical refinement for camera relocalisation. In 2024 IEEE International Conference on Robotics and Automation, pages 8544–8550, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[34] Changkun Liu, Shuai Chen, Yash Sanjay Bhalgat, Siyan HU, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient camera pose refinement via 3d gaussian splatting. In The Thirteenth International Conference on Learning Representations, 2025. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[36] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE International Conference on Computer Vision, pages 2372-2381, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 2794-2802, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[41] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Coordinate: uncertainty-aware pose regressor for reliable vehicle localization. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2229-2238, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[42] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In Conference on Robot Learning, pages 1347-1356, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[43] Arthur Moreau, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Crossfire: Camera relocalization on self-supervised features from an implicit representation. In IEEE/CVF International Conference on Computer Vision, pages 252-262, 2023. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.904, + 0.546 + ], + "angle": 0, + "content": "[44] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In IEEE international conference on computer vision, pages 3456-3465, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[47] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[48] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6 dof outdoor visual localization in changing conditions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Torsten Sattler, Qunjie Zhou, Marc Pollefeys, and Laura Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3302-3312, 2019. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26857" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[51] Yoli Shavit and Yosi Keller. Camera pose auto-encoders for improving pose regression. In European Conference on Computer Vision, pages 140–157, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[52] Yoli Shavit, Ron Ferens, and Yoshi Keller. Paying attention to activation maps in camera pose regression. arXiv preprint arXiv:2103.11477, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[53] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.297, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[54] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[56] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM Transactions on Graphics, pages 835-846, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[57] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7199-7209, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[59] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The unreasonable effectiveness of pretrained features for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12798, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[60] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[61] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[62] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. In ACM in Computer Graphics and Interactive Techniques, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[64] Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, and Zhaopeng Cui. Pnerfloc: Visual localization with point-based neural radiance fields. In AAAI Conference on Artificial Intelligence, pages 7450-7459, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[65] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In IEEE International Conference on Computer Vision, pages 2075-2083, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[66] Qunjie Zhou, Maxim Maximov, Or Litany, and Laura Leal-Taixe. The nefrect match: Exploring nef features for visual localization. arXiv preprint arXiv:2403.09577, 2024. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26858" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_origin.pdf b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..328a406c44f10327cabef3f5dc873d1f198e4f4a --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/b8dc72b7-af77-4d02-8bf0-d21936019f63_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04271a36a38f5157618c3d9803b5331d75544021bb549ef91d1169018b7bd4ba +size 2470249 diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/full.md b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..46bf898aadf44b3f8588692bbc609ff22b1dfbf9 --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/full.md @@ -0,0 +1,347 @@ +# Adversarial Exploitation of Data Diversity Improves Visual Localization + +Sihang Li* Siqi Tan* Bowen Chang Jing Zhang Chen Feng Yiming Li + +New York University + +{sihangli, tan.k, cfeng, yimingli}@nyu.edu + +https://ai4ce.github.io/RAP + +# Abstract + +Visual localization, which estimates a camera's pose within a known scene, is a fundamental capability for autonomous systems. While absolute pose regression (APR) methods have shown promise for efficient inference, they often struggle with generalization. Recent approaches attempt to address this through data augmentation with varied viewpoints, yet they overlook a critical factor: appearance diversity. In this work, we identify appearance variation as the key to robust localization. Specifically, we first lift real 2D images into 3D Gaussian Splats with varying appearance and deblurring ability, enabling the synthesis of diverse training data that varies not just in poses but also in environmental conditions such as lighting and weather. To fully unleash the potential of the appearance-diverse data, we build a two-branch joint training pipeline with an adversarial discriminator to bridge the syn-to-real gap. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, reducing translation and rotation errors by $50\%$ and $33\%$ on indoor datasets, and $38\%$ and $44\%$ on outdoor datasets. Most notably, our method shows remarkable robustness in dynamic driving scenarios under varying weather conditions and in day-to-night scenarios, where previous APR methods fail. + +# 1. Introduction + +Visual localization, the task of calculating a 6-DoF camera pose—its translation and rotation—based on a query image within a given environment, is essential for various applications, including robotics [1], autonomous vehicles [18], and virtual reality [11]. Besides traditional geometry-based approaches, recent learning-based visual localization methods adopt absolute pose regression (APR) [5, 8, 24, 53], scene coordinate regression (SCR) [3, 4, 43, 61], or post pose refinement (PPR) [9, 19, 33, 43, 59, 64, 66]. SCR methods focus on learning-based 2D-3D correspondences followed by subsequent Perspective-n-Point (PnP) for pose estimation. + +![](images/61b4c710face4e57011a77d867b8ed225c07bac95a93e4145e0c2c5241857bc4.jpg) +Figure 1. We propose RAP, a novel pipeline to train robust APR models. We lift real-world 2D images into 3D Gaussian Splats [25] to synthesize images with diverse appearances and poses, improving model generalizability. We also introduce an adversarial discriminator, mitigating the syn-to-real gap to learn robust features. Together, we achieve state-of-the-art performance. + +PPR methods heavily rely on a pose prior, usually obtained from image retrieval, followed by iterative refinement. In contrast, APR methods employ a supervised framework to train a regression neural network on image-pose pairs, enabling direct pose estimation during inference. APR offers faster runtime and lower error in challenging scenes with sparse views, significant lighting variations, or numerous dynamic objects, making it a promising method for ensuring robustness in real-world applications. + +Despite promises, there is a performance gap in localization accuracy between APR and other methods. A well-known pivotal work [49] attributes this to APR performing image-based memorization, i.e., retrieving poses seen during training. Driven by this crucial finding, to improve such memorization while avoiding the need for denser real-world training samples, recent methods leverage Neural Radiance Fields (NeRF) [40] to synthesize additional posed images for APR training [8, 30, 42]. LENS [42] tried to employ appearance perturbation using NeRF-W [39], but found the + +improvements minimal [42]. As appearance augmentation is common in other learning tasks, why does it fail in APR? + +We hypothesize a learning gap: Limitations in previous training pipelines prevented the effective use of diverse data to boost performance. Artifacts always exist in images rendered by common novel view synthesis (NVS) methods, which might disturb the model feature space. Inspired by generative adversarial networks (GAN) [20], where a discriminator is trained to distinguish between real and generated samples, we propose adversarial training for APR, designing a discriminator to align the features of synthetic and real images, thereby reducing the syn-to-real domain gap and mitigating the impact of rendering artifacts. Augmentation quality also matters. To efficiently synthesize diverse high-quality images with controllable varying appearance, we extend the vanilla 3D Gaussian Splitting (3DGS) [25] to appearance-varying 3DGS with deblurring ability. + +These form our two-branch joint training framework for robust absolute pose regression (RAP). The first branch coarsely trains our Transformer-based pose regressor with both real data and data synthesized at the original real pose, together with an adversarial discriminator to reduce the synto-real domain gap. The second branch progressively generates randomly perturbed poses and appearances, providing additional supervision to the same APR Transformer. Through extensive experiments, we demonstrate that exploiting data diversity using adversarial training significantly increases localization accuracy in APR. Meanwhile, our results indicate that APR consistently benefits from more diverse visual data, and we observe clear signs of a more generalizable APR emerging with its localization performance cannot be explained merely by memorization. + +Our contributions are summarized as follows: + +- We identify the crucial role of appearance diversity for APR, and develop a 3DGS-based appearance-varying data augmentation framework to efficiently generate diverse synthetic data with controllable lighting conditions. +- We propose an adversarial discriminator to reduce the syn-to-real gap. Together with progressive data synthesis, we form a robust two-branch joint training pipeline that fully unleashes the power of data diversity. +- We conduct extensive experiments showing our method outperforms state-of-the-art approaches on challenging datasets with significant appearance change. Ablation studies further analyze key factors affecting performance. + +# 2. Related Works + +Visual Localization. Visual localization aims to estimate a camera's translation and rotation within a 3D scene. Traditional geometry-based methods [6, 15, 32, 35, 44-47, 57] accomplish this by using point clouds and a reference image database, relying on stored descriptors and image retrieval to establish 2D-3D correspondences. In contrast, scene coordinate regression (SCR) methods [2-4, 61] em + +bed map information within neural networks to directly predict 2D-3D correspondences. Both approaches generally require PnP [17] and RANSAC [16] to output camera poses at test time, which adds additional computation cost. Alternatively, absolute pose regression (APR) [5, 7, 22, 24, 41, 52] aims to directly regress the camera pose from a query image using neural networks. Although the performance is suboptimal compared with geometry-based methods, APR remains a promising approach due to its fast inference. + +Data Augmentation for Pose Regression. End-to-end pose regression methods rely heavily on the amount and diversity of training data. Previous work [49] shows that APR implicitly learns image retrieval in the given environment. Therefore, the following works LENS [42], DFNet [8] and PMNet [30] enhance APR performance by spatially enriching training views with NeRF. However, these approaches fail to address the generalizability of APR models and exhibit several limitations: (1) The efficiency of training and novel view synthesis (NVS) in NeRF is severely restricted, hindering scalability. (2) They limit NVS to geometric (pose) transformations while neglecting photometric (appearance) variations, thereby decreasing APR robustness to changes in visual appearance. (3) The augmented data is underutilized in their learning frameworks, leaving its potential for improving APR largely untapped. Differently, our framework switches to 3DGS [25] as the scene representation to efficiently generate novel posed images with controllable appearances and introduce adversarial training to unleash the power of such diverse data. + +Handling and Synthesizing Challenging Scenarios. Visual localization often encounters unstructured photo collections [56], where visual appearance varies due to moving objects, lighting changes, and inconsistent camera exposure settings. To tackle these in-the-wild challenges, NeRF-W [38] uses per-image transient and appearance embeddings. In 3DGS [25], VastGaussian [31] applies a CNN to 3DGS outputs but still struggles with significant appearance variations. SWAG [12] mitigates this issue by storing appearance information in an external hash-grid-based implicit field, while GS-W [63] enhances flexibility by separating intrinsic and dynamic appearance features for each Gaussian point. 3DGM [29] leverages consensus across multiple sequences as the self-supervision signal to remove transient and moving objects without human annotations. Deblur-GS [62] addresses motion blur—another challenge in localization datasets—by modeling camera motion to yield sharper edges in rendered scenes. Our method incorporates appearance modeling and edge refinement to handle and synthesize diverse indoor, outdoor, and driving scenes. + +# 3. Method + +# 3.1. Pre-Processing with 3DGS + +A robust pose regressor should focus on intrinsic scene attributes, not appearance variations. Therefore, we first + +![](images/ded7d99ee2f9af0dfe52d63070baaaed376e73807554e871bd73f06b20609cb5.jpg) +Figure 2. Pipeline of RAP. We lift multiple RGB video sequences into 3D Gaussian Splats, which serve as our data engine. The branch-1 (see Sec. 3.3.1) inputs paired real and synthetic images to regress poses, with a discriminator to bridge the syn-to-real gap. The branch-2 (see Sec. 3.3.2) generates views with novel poses and appearances, which are fed into the same pose regressor as additional supervision. + +To further analyze the behavior of the $K$ in the $K \times 16 \times 3$ system, we assume that the camera is located at the center of the image. The camera is positioned at the center of the image by a camera position vector $\mathbf{C} = [0, 1]$ . To achieve this, we use the following procedure to generate the camera positions: + +$$ +\boldsymbol {\mathcal {C}} = \operatorname {M L P} (\boldsymbol {\mu}, \boldsymbol {\mathcal {Y}}, \omega \boldsymbol {\mathcal {E}}, \theta), \tag {1} +$$ + +where $\omega$ is the blending weight that controls the dynamic appearance of the rendered image. + +Another significant challenge in visual localization is motion blur, often caused by slow shutter speeds during video capture, leading to pose ambiguity and degraded rendering quality, further decreasing localization accuracy. Inspired by Deblur-GS [62], we model camera motion blur as the inverse of scene motion, i.e., the transformation in Gaussian position denoted by $\mathcal{T} \in \mathrm{SE}(3)$ . For each training image, we sample a certain time step along a linear trajectory with a sampling weight $\phi \in \mathbb{R}^n$ and blend them to compute loss $\mathcal{L}$ with the original blurry image $\mathcal{I}_b \in \mathbb{R}^{H \times W \times 3}$ , optimizing $\mathcal{T}$ , $\phi$ , $\mathcal{C}$ and other 3DGS parameters $\Theta$ : + +$$ +\underset {\phi , \boldsymbol {\mathcal {T}}, \boldsymbol {\mu}, \boldsymbol {\mathcal {Y}}, \boldsymbol {\mathcal {E}}, \boldsymbol {\Theta}} {\operatorname {a r g m i n}} \mathcal {L} \left(\boldsymbol {\mathcal {I}} _ {b}, \sum_ {i = 1} ^ {n} \phi_ {i} \operatorname {R e n d e r} \left(\boldsymbol {\mathcal {T}} _ {i} (\boldsymbol {\mu}), \boldsymbol {\mathcal {C}}, \boldsymbol {\Theta}\right)\right), \tag {2} +$$ + +where the details of $\mathcal{L}$ are in supplemental materials. After training, our 3DGS can efficiently render posed images given $\theta$ and $\omega$ . + +# 3.2. Architecture of Pose Regressor + +Given a set of images and their associated camera poses $\{(I_i, P_i)\}_{i=1}^n$ , our goal is to train a neural network to directly output a homogeneous camera pose $P \in \mathbb{R}^{3 \times 4}$ for a query image $I \in \mathbb{R}^{H \times W \times C}$ . Our network architecture is shown as the pose regressor in Fig. 2. + +Feature Extractor. Pose regression networks typically extract features using a common backbone $\varphi$ , such as VGG [55] or EfficientNet [58], leveraging multiple deeper layers for translation and rotation regression: + +$$ +\varphi (\boldsymbol {I}) = \left\{\mathcal {F} _ {0} (\boldsymbol {I}), \dots , \mathcal {F} _ {N - 1} (\boldsymbol {I}), \mathcal {F} _ {N} (\boldsymbol {I}) \right\}, \tag {3} +$$ + +$\mathcal{F}_{*}(\cdot)$ denotes features extracted from the $*$ -th layer of a backbone with $N$ layers. $\mathcal{F}_t(\pmb {I})$ and $\mathcal{F}_r(\pmb {I})$ denote features for translation and rotation regression, respectively. + +Pose Transformer. Unlike CNN-based regression models [8, 30], where fine-grained local features can introduce noise and harm performance, we propose Pose Transformer to leverage the strong ability of Vision Transformer (ViT) [14] for modeling long-range dependencies. Each Transformer generates a global token (Trans for translation and Rot for rotation) to provide a comprehensive context for pose regression, inspired by the CLS token in ViT. Given $\mathcal{F}_r(I)$ and $\mathcal{F}_t(I)$ , the translation token is then concatenated with the flattened input features*: + +$$ +\widetilde {\mathcal {F}} _ {t} (\boldsymbol {I}) = \operatorname {C a t} (\text {F l a t t e n} (\mathcal {F} _ {t} (\boldsymbol {I})), \text {T r a n s}) \in \mathbb {R} ^ {\left(H _ {t} W _ {t} + 1\right) \times C _ {t}}. \tag {4} +$$ + +The positional encodings are then added to the flattened feature $(\mathsf{PE} + \widetilde{\mathcal{F}}_t(\pmb {I}))\in \mathbb{R}^{(H_tW_t + 1)\times C_t}$ . Multi-head Self-Attention (MSA) is then conducted through a stack of multiple layers with the post-processing as follows: + +$$ +\widehat {\mathcal {F} _ {t} ^ {\prime}} (\boldsymbol {I}) = \operatorname {M S A} \left(\mathrm {P E} + \widetilde {\mathcal {F} _ {t}} (\boldsymbol {I})\right) + \mathrm {P E} + \widetilde {\mathcal {F} _ {t}} (\boldsymbol {I}), \tag {5} +$$ + +$$ +\widehat {\mathcal {F}} _ {t} (\boldsymbol {I}) = \mathrm {L N} (\mathrm {F F N} (\mathrm {L N} (\widehat {\mathcal {F}} _ {t} ^ {\prime} (\boldsymbol {I}))) + \widehat {\mathcal {F}} _ {t} ^ {\prime} (\boldsymbol {I})), +$$ + +where LN indicates layer normalization and FFN denotes the fully connected feed-forward network, consisting of two linear layers with a ReLU. The final output is flattened back to $(H_{t}W_{t} + 1)\times c_{t}$ . See supplementary for more details. + +Regression Head. Only the processed translation token, Trans, capturing global features for regression, is fed into the regression head. This regression head consists of two MLPs, each with a hidden layer and GeLU activation: + +$$ +\hat {\boldsymbol {t}} = \operatorname {L i n e a r} (\operatorname {G e L U} (\operatorname {L i n e a r} (\widehat {\text {T r a n s}}))). \tag {6} +$$ + +The $\hat{t}$ represents the final prediction for translation. Similarly, we obtain the rotation prediction denoted by $\hat{r}$ . + +# 3.3. Two-Branch Joint Training Paradigm + +# 3.3.1. Branch-1: Aligning Features via Discriminator + +Synthetic images from 3DGS provide novel viewpoints and appearances but often contain artifacts, leading to a syn-to-real domain gap. To align features from rendered and real images of the same pose, we introduce an adversarial training mechanism besides the basic pose regression training. + +Pose Regression Loss. For basic training, we render the synthetic image $I^{\prime}$ with the same pose label $\pmb{P}$ as the real image $\pmb{I}$ , both used as supervision for the pose regressor. The training objective consists of translation loss $\mathcal{L}_t$ and the rotation loss $\mathcal{L}_r$ , which are measured by the Euclidean distance between the ground truth pose $\pmb{P} = \{\hat{t},\hat{r}\}$ and the estimated pose $\widehat{P} = \{\hat{t},\hat{r}\}$ : + +$$ +\mathcal {L} _ {t} = \| \boldsymbol {t} - \hat {\boldsymbol {t}} \| _ {2}, \tag {7} +$$ + +$$ +\mathcal {L} _ {r} = \left\| \boldsymbol {r} - \frac {\hat {\boldsymbol {r}}}{\| \hat {\boldsymbol {r}} \|} \right\| _ {2}, \tag {8} +$$ + +$$ +\mathcal {L} _ {\text {p o s e}} ^ {1} = \mathcal {L} _ {t} \exp (- s _ {t}) + s _ {t} + \mathcal {L} _ {r} \exp (- s _ {r}) + s _ {r}, \tag {9} +$$ + +where $s_t$ and $s_r$ are learned parameters for balancing the optimization between rotation and translation [23]. + +Adversarial Loss. The adversarial training mechanism optimizes the discriminator to distinguish real from rendered image features, while training the feature extractor to fool the discriminator, effectively bridging the domain gaps. To prevent vanishing gradients, we propose a novel adversarial objective for pose regression, inspired by LSGAN [37]: + +$$ +\begin{array}{l} \underset {D} {\operatorname {a r g m i n}} \mathcal {L} _ {\mathrm {D i s}} (D) = \frac {1}{2} \mathbb {E} _ {\boldsymbol {I} \sim p _ {\mathrm {d a t a}} (\boldsymbol {I})} \left[ \left(D (\operatorname {A d j} (\mathcal {F} _ {t} (\boldsymbol {I}))) - 1\right) ^ {2} \right] \\ + \frac {1}{2} \mathbb {E} _ {\boldsymbol {I} ^ {\prime}} \left[ D \left(\operatorname {A d j} ^ {\prime} \left(\mathcal {F} _ {t} \left(\boldsymbol {I} ^ {\prime}\right)\right) ^ {2} \right], \right. \tag {10} \\ \end{array} +$$ + +$$ +\underset {G} {\operatorname {a r g m i n}} \mathcal {L} _ {\text {G e n}} (G) = \frac {1}{2} \mathbb {E} _ {I ^ {\prime}} \left[ \left(D \left(\operatorname {A d j} ^ {\prime} \left(\mathcal {F} _ {t} \left(I ^ {\prime}\right)\right) - 1\right) ^ {2} \right]. \right. \tag {11} +$$ + +Here, Adj and $\mathrm{Adj}'$ are the adjustment layers, consisting of Conv-ReLU-BN layers. The feature extractor $\varphi$ acts as the generator $G$ , while $D$ is the discriminator, composed of several convolutional layers with ReLU activations. More details are in supplemental materials. + +# 3.3.2. Branch-2: Training while Synthesizing Data + +With the proposed appearance-varying 3DGS, more posed images are generated to enrich the training data for better generalizability. Specifically, our data synthesis is categorized into two dimensions: pose augmentation and appearance augmentation, as illustrated in Fig. 1. For pose augmentation, given a training pose $P$ , a perturbed pose $P_{\mathrm{syn}}$ is generated around $P$ by the translation noise of $\delta t$ and rotation noise of $\delta r$ . For appearance augmentation, we randomly adjust the appearance of rendered images using random blending weights $\omega$ , and then render the synthetic image $I_{\mathrm{syn}}$ using the Gaussian Splats trained in Sec. 3.1. The novel image-pose pair $(I_{\mathrm{syn}}, P_{\mathrm{syn}})$ , online generated every 20 epochs during training until the validation MSE loss and median errors cease to decrease, serves as additional supervision for the training. Given the estimated pose of the synthesized image denoted by $\widehat{P_{\mathrm{syn}}}$ , the loss function $\mathcal{L}_{\mathrm{pose}}^2(\widehat{P_{\mathrm{syn}}}, P_{\mathrm{syn}})$ is same as $\mathcal{L}_{\mathrm{pose}}^1$ . + +# 3.3.3. Overall Objective + +The total loss for the pose regressor is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \beta_ {1} \mathcal {L} _ {\text {p o s e}} ^ {1} + \beta_ {2} \mathcal {L} _ {\text {p o s e}} ^ {2} + \beta_ {3} \left(\mathcal {L} _ {\text {G e n}} + \mathcal {L} _ {\text {D i s}}\right), \tag {12} +$$ + +where $\beta_{1},\beta_{2},\beta_{3}$ are loss weights. The total loss will optimize the pose regressor, adjustment layers, and discriminator. Only the pose regressor will be deployed in the inference phase, while the other two components are discarded. + +# 4. Experiments + +# 4.1. Evaluation Setup + +Datasets. We follow previous works [8, 30] to mainly use four scenes in the Cambridge Landmarks dataset [24] with spatial extents around $875\mathrm{m}^2$ . Moreover, we evaluate our method on MARS [28], a self-driving dataset featuring challenges like moving objects, lighting changes, and motion blur. To investigate the robustness of our model under extreme lighting changes, such as the transition from day to night, we also prepared a subset of the Aachen Day-Night dataset [48]. The training data includes images captured using various camera models with differing resolutions, which renders direct evaluation with APR methods infeasible. Thus, we standardized the camera models through center cropping and built a COLMAP [50] model as pose annotations, including 13 nighttime images for evaluation and 246 daytime images for training 3DGS and RAP. We also employ the 7-Scenes dataset [54], which provides seven indoor scenes with volumes spanning $1\mathrm{m}^3 - 18\mathrm{m}^3$ and follow the original training and testing splits with more accurate SfM pose annotations [4, 9]. Although it is an indoor dataset, it is still non-trivial as it includes texture-less surfaces, object occlusions, and motion blur. + +![](images/15ebf38683dbd6908706749e21ed47014c0be1f8e9f74ee3345b54c5b2a5dc5a.jpg) +Figure 3. Qualitative comparison of camera pose estimation errors between a) DFNet [8] and b) our RAP framework across five scenes on the Cambridge Landmarks dataset [24]. Our RAP framework estimates trajectories that more closely follow the ground truth, with significantly reduced rotation and position errors compared to DFNet [8]. + +Table 1. Median translation (cm) and rotation $(^{\circ})$ errors on the Cambridge Landmarks dataset [24]. The best results† in pure APR and PPR are highlighted in bold. + +
MethodsCollegeHospitalShopChurchAverage‡Court
Pure APRPN [24]166/4.86262/4.90141/7.18245/7.95204/6.23683/3.50
MapNet [5]107/1.89194/3.91149/4.22200/4.53163/3.64N/A
MS-Trans. [53]83/1.47181/2.3986/3.07162/3.99128/2.73N/A
PAE [51]90/1.49207/2.5899/3.88164/4.16140/3.03N/A
LENS† [42]33/0.5044/0.9027/1.6053/1.6039/1.20N/A
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90217/4.11
PMNet [30]68/1.97103/1.3158/2.10133/3.7390/2.27N/A
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28115/1.68
SCRDSAC* [2]18/0.321/0.45/0.315/0.615/0.434/0.2
ACE [4]28/0.431/0.65/0.318/0.621/0.543/0.2
GLACE [61]19/0.317/0.44/0.29/0.312/0.319/0.1
PPRFQN-MN [19]28/0.3854/0.8213/0.6358/2.0038/0.964253/39.16
LENS [42]34/0.5445/0.9628/1.6654/1.6640/1.21N/A
CrossFire [43]47/0.743/0.720/1.239/1.437/1.0N/A
NeFeS50 [9]37/0.5452/0.8815/0.5337/1.1435/0.77N/A
HR-APR [33]36/0.5853/0.8913/0.5138/1.1635/0.78N/A
MCLoc [59]31/0.4239/0.7312/0.4526/0.8827/0.62N/A
DFNetGS-CPR [34]23/0.3242/0.7410/0.3627/0.6226/0.51N/A
ACEGS-CPR [34]20/0.2921/0.405/0.2413/0.4015/0.33N/A
DFNetref(Ours)16/0.2421/0.418/0.4210/0.2614/0.3325/0.13
RAPref (Ours)15/0.2518/0.385/0.239/0.2312/0.2723/0.15
+ +Since most methods did not report results on Court, it is excluded from the average error calculation. As Coordinet + LENS [42] does not provide open-source code, it is unclear whether any post-processing is used. + +Baselines. We first compare our proposed RAP against common APR-only approaches on the four datasets, where PMNet [30] and DFNet [8] are the most related and advanced methods based on data augmentation. We split the remaining methods into two categories based on whether they rely on extra novel view synthesis in test time, including SCR [3, 4, 61] and PPR (Post Pose Refinement) [9, 19, 33, 43, 59, 64, 66], which involves rendering images, querying features in novel views by the initial pose, iterative refinement or sequential refinement [41]. + +Implementation Details. First, we optimize our 3DGS for each scene without masking moving objects. We + +![](images/c57aa03fd01ce4f456179d649cd6a2a01a90affa509ee3d9b08a653d61298417.jpg) +Figure 4. Visualization of $\mathrm{RAP}_{\mathrm{ref}}$ on MARS [28]. In each subfigure, a diagonal line separates the "Predicted" (rendered from the refined pose) and "GT" (ground truth) sections. Smooth alignment along this boundary shows $\mathrm{RAP}_{\mathrm{ref}}$ 's improved pose accuracy. + +Table 2. Median translation (cm) and rotation $(^{\circ})$ errors on the MARS dataset [28]. + +
Methods“0011”“0015”“0037”“0041”Average
PoseNet [24]149/1.80136/2.34123/1.6075/0.92121/1.67
RAP (Ours)32/0.6137/1.0815/0.3528/0.3528/0.60
RAPref (Ours)8.5/0.138.2/0.208.7/0.097.6/0.118.3/0.13
+ +Table 3. Median translation (COLMAP [65] unit) and rotation $(^{\circ})$ errors on the Aachen Day-Night Dataset [48]. + +
APR-BasedSCR-Based
PoseNet [24]DFNet [8]RAP w/o App.RAPRAPrefACE [4]GLACE [61]
217/74.30174/85.80134/75.99130/13.7067/4.64914/90.50482/36.4
+ +then train our RAP network, which uses an Efficient-B0 backbone [36] pre-trained on ImageNet [13], optimized with Adam [26] at a learning rate of $10^{-4}$ . Only the features from the third (reduction_3) and fourth (reduction_4) layers are used respectively for translation and rotation regression, and both layers are utilized for + +![](images/c2e0d3a703c676af4979a21260e04cec28f2293f6fbad517b34edcf025eaabbd.jpg) +Figure 5. Visualization of the localization errors of $\mathrm{RAP}_{\mathrm{ref}}$ on the 7-Scenes dataset [54]. + +Table 4. Quantitative results on the 7-Scenes dataset [54]. The best results in pure APR and PPR are highlighted in bold. DSLAM GT and SfM GT refer to different sets of ground truth. More visualizations and details are in supplemental materials. + +
CategoryMethodsChessFireHeadsOfficePumpkinKitchenStairsAverage
Pure APRPoseNet (PN) [24]32/8.1247/14.429/1.2048/7.6847/8.4259/8.6447/13.8044/10.4
MapNet [5]8/3.2527/11.718/13.317/51.522/4.0223/4.9330/12.121/7.77
AtLoc+ [60]10/3.1826/10.814/11.417/5.1620/3.9416/4.9029/10.219/7.08
MS-Transformer [53]11/4.6624/9.6014/12.217/5.6618/4.4417/5.9417/5.9418/7.28
PAE [51]12/4.9524/9.3114/12.519/5.7918/4.8918/6.1925/8.7419/7.48
CoordiNet + LENS [42]4/1.3811/3.778/5.868/1.989/2.2710/2.2715/3.679/3.07
DFNet [8]5/1.8817/6.456/3.638/2.4810/2.7822/5.4516/3.2912/3.71
PMNet [30]4/1.7010/4.517/4.237/1.9614/3.3314/3.3616/3.6210/3.24
RAP (Ours, DSLAM GT)3/1.417/3.466/6.025/1.976/1.967/2.1810/2.146/2.73
RAP (Ours, SfM GT)2/0.856/2.844/4.525/1.944/1.707/2.1210/1.305/2.18
SCRDSAC [2]0.5/0.170.8/0.280.5/0.341.2/0.341.2/0.280.7/0.212.7/0.781.1/0.34
ACE [4]0.5/0.180.8/0.330.5/0.331.0/0.291.0/0.220.8/0.202.9/0.811.1/0.34
GLACE [61]0.6/0.180.9/0.340.6/0.341.1/0.290.9/0.230.8/0.203.2/0.931.2/0.36
marepo† [10]2.6/1.352.5/1.422.3/2.213.6/1.444.2/1.555.1/1.996.7/1.833.9/1.68
PPRFQN-MN [19]4.1/1.3110.5/2.979.2/2.453.6/2.364.6/1.7616.1/4.42139.5/34.6728/7.3
CrossFire [43]1/0.45/1.93/2.35/1.63/0.82/0.812/1.94.4/1.38
DFNet + NeFeS50 [9]2/0.572/0.742/1.282/0.562/0.552/0.575/1.282.4/0.79
HR-APR [33]2/0.552/0.752/1.452/0.642/0.622/0.675/1.302.4/0.85
MCLoc [59]2/0.83/1.43/1.34/1.35/1.66/1.66/2.04.1/1.43
DFNet + GS-CPR (SfM GT) [34]0.7/0.200.9/0.320.6/0.361.2/0.321.3/0.310.9/0.252.2/0.611.1/0.34
ACE + GS-CPR (SfM GT) [34]0.5/0.150.6/0.250.4/0.280.9/0.261.0/0.230.7/0.171.4/0.420.8/0.25
RAPref (Ours, DSLAM GT)2.83/1.432.07/1.231.54/1.872.55/1.204.59/1.544.23/1.833.48/1.233.04/1.48
RAPref (Ours, SfM GT)0.33/0.110.51/0.210.40/0.290.59/0.170.83/0.200.50/0.121.11/0.320.61/0.20
+ +As marepo [10] combines SCR and APR, we classify it as SCR. + +narrowing the domain gap via the discriminator, which is also optimized with Adam [26], using a learning rate of $10^{-4}$ and betas set to (0.5, 0.999). More details about training are in supplemental materials. For generating random views, we apply random normalized perturbations to each training pose: $\delta t = 20\mathrm{cm}$ and $\delta r = 10^{\circ}$ for indoor scenes, and $\delta t = 150\mathrm{cm}$ and $\delta r = 4^{\circ}$ for outdoor scenes. + +To allow for comparison with SCR methods and leverage 3DGS's efficient rendering for PPR, we extend the APR pipeline with match-based refinement similar to GS-CPR [34], denoted as $\mathrm{RAP}_{\mathrm{ref}}$ . At test time, RAP's initial pose is used to render an RGB-D image via 3DGS. Together with MASt3R [27], we can obtain 2D-3D correspondences to perform RANSAC-PnP [16, 17], resulting in a refined pose. More details are in supplementary materials. + +# 4.2. Benchmark Results + +Cambridge Landmarks [24]. In the challenging outdoor Cambridge Landmarks dataset (Table 1), our RAP reduces both translation and rotation errors across all scenes by over + +$30\%$ compared to other APR-only methods. The visualization in Fig. 3 shows that our method produces fewer outliers than DFNet [8]. In the three larger-scale scenes with significant appearance diversity (College, Church, and Court), rotation error is even halved compared to DFNet. Table 1 also shows the effectiveness of our $\mathrm{RAP}_{\mathrm{ref}}$ in further reducing pose errors through refinement. $\mathrm{RAP}_{\mathrm{ref}}$ outperforms $\mathrm{CoordiNet} + \mathrm{LENS}$ [42], which assumes a continuous trajectory when an Extended Kalman Filter [21] is required for refinement [41]. $\mathrm{RAP}_{\mathrm{ref}}$ even surpasses ACE [4] and its post-refinement variant, $\mathrm{ACE} + \mathrm{GS}$ -CPR [34], despite GS-CPR manually masking dynamic objects when building 3DGS. This demonstrates the strong representation capability of our appearance-varying 3DGS with deblurring. + +MARS [28]. Autonomous driving scenarios present unique challenges, including moving objects and frequent changes in lighting conditions, as illustrated in Fig. 4. Our RAP demonstrates effective and robust performance across four challenging scenes, as shown in Table 2, achieving an average of $28~\mathrm{cm} / 0.60^{\circ}$ localization error. This significantly + +Table 5. Ablation study. + +
Setupon ShopTrans. (cm) ↓Rot. (°) ↓
I (Baseline): φ = VGG161745.45
II: φ = Efficient-B01034.64
III: II + Pose Aug.753.52
IV: III + Appearance Aug.603.14
V: IV + Decoder (ConvNet)522.51
VI: V + Decoder (Transformer)401.98
VII (Ours): VI + Discriminator331.48
+ +outperforms the baseline† PoseNet [24]. With one-shot refinement, our $\mathrm{RAP}_{\mathrm{ref}}$ further reduces outdoor localization errors to below $10~\mathrm{cm}$ . + +Aachen Day-Night [48]. Benefiting from appearance augmentation, our RAP significantly reduces the localization error from 134 unit $/75.99^{\circ}$ to 130 unit $/13.70^{\circ}$ , outperforming other APR [8, 24] and SCR [4, 61] baselines, as shown in Table 3. This demonstrates the effectiveness of appearance diversity in handling extreme lighting changes. + +7-Scenes [54]. As shown in Table 4, our RAP reduces translation error by $50\%$ ( $10\mathrm{cm} \rightarrow 5\mathrm{cm}$ ) and rotation error by $32.72\%$ ( $3.24^{\circ} \rightarrow 2.18^{\circ}$ ) on average compared to previous state-of-the-art single-frame APR methods. The only exception is Heads, where the rotation error is suboptimal. This scene consists of just two sequences—one for training and one for testing—potentially limiting the effectiveness of our augmentation in capturing scene variability. Meanwhile, $\mathrm{RAP}_{\mathrm{ref}}$ further reduces localization error below 1 cm with one-shot refinement using our 3DGS. It also surpasses ACE [4] and its post-refinement variant, $\mathrm{ACE} + \mathrm{GSCPR}$ [34]. Qualitative examples are shown in Fig. 5. + +# 4.3. Ablation Study + +We conduct ablation studies on the validation set of Shop in the Cambridge Landmarks dataset to investigate the impact of all the components in our RAP. Setup I, our baseline, consists of the same components as in PoseNet [24] and has been retrained for our experiments. In Setup II, we replace the feature extraction from VGG16 [55] to Efficient-B0 [36], which enhances performance due to its superior feature representation, while they both exhibit poor performance due to the lack of data synthesis. In Setup III and IV, we explore the effectiveness of the designed pose augmentation and appearance augmentation, which bring notable improvements: translation error reduces from $103~\mathrm{cm}$ to $75~\mathrm{cm}$ , and rotation error from $3.52^{\circ}$ to $3.14^{\circ}$ . In Setup V and VI, we add regular convolutional layers and Pose Transformer between feature extraction and pose regression. Both improve performance due to the increasing parameters, but the Transformer achieves superior results by effectively handling long-term dependencies through attenuating the effect of noise. + +![](images/5eff30fa95553b18efeb01e66b48e3fb59a73df1e300fb37fcd0488624d81f92.jpg) +Figure 6. Visualization of the training set distribution and results on St. George's Basilica Building [49]. The red hollow spheres, centered on the real images in the training set, indicate the potential locations of all synthetic images during training. + +tion mechanisms. Finally, in Setup VII, our adversarial discriminator effectively reduces the syn-to-real domain gap, allowing the model to learn better pose regression features from synthetic data and further reduce localization error. + +# 4.4. Discussion on Data Synthesis + +Emerging Generalizability. Previously, APR has been understood to implicitly learn image retrieval [49], lacking the ability to successfully interpolate between training samples and generalize beyond them. To investigate how APR training is affected by increasing synthetic data, we trained RAP on the St. George's Basilica Building [49] and visualized the results in Fig. 6. Here, the translation perturbation was set to $\delta t = 350$ cm and the rotation perturbation to $\delta r = 60^{\circ}$ . Notably, the test set contained two regions entirely uncovered by the training set. Despite this, the model still closely predicts the test camera poses, demonstrating generalization ability beyond the original training positions. + +We also learn from our experiments that reducing the rotation perturbation, such that the overlap between test and training views remains minimal, leads to high localization error. This is because the translation and rotation parameter space is inherently a SE(3) manifold. Even if the translation remains fixed, significant rotation changes result in entirely different visual content in the images, naturally preventing the model from estimating poses of such unseen views, which correspond to a large distance on the SE(3) manifold. Therefore, enabling generalization across a broader range of space is an important direction for future work. + +Analyzing Generalization Boundaries. To evaluate the model's generalizability, we designed an experiment introducing a "void zone" centered on the test camera, where all real and synthetic data within this zone were excluded. The void zone was progressively expanded to determine the critical threshold at which the localization performance declines most significantly. Specifically, for Shop, we used $100\%$ of the training set to ensure complete scene cov + +Table 6. Exploring the generalization boundaries of the model with synthetic data. Green, blue, and red percentages indicate the relative change in localization error (Med Err) compared to the scenario without a void zone. + +
w/o Void Zonew/ Void Zone (cm/°)
10/0.520/130/1.550/280/2.5
Med Err ↓ (rel. change)33/1.2630/1.3432/1.3240/1.8439/2.0749/2.20
0%/0%-9%/6.3%-3%/4.7%21%/46.0%18%/64.3%48%/74.6%
Avg Err ↓41/1.5138/1.7541/1.6351/2.4048/2.4061/2.84
Max Err ↓155/4.52147/6.56219/6.01192/8.38246/9.80242/13.12
Min Err ↓3/0.178/0.204/0.227/0.304/0.168/0.60
+ +Table 7. Ablation on different pose augmentation policies. + +
MethodsCollegeHospitalShopChurchAverage
RAP (LENS [42])73/1.15126/1.8771/3.37128/3.50100/22.47
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
+ +erage, with void zone ranges set as [10/0.5, 20/1, 30/1.5, 50/2, 80/2.5, 100/3] $(\mathrm{cm} / {}^{\circ})$ . The results in Table 6 demonstrate a stepwise decline in performance. Initially, expanding the void zone has minimal impact on localization accuracy. However, at $30~\mathrm{cm} / 1.5^{\circ}$ , a sharp decrease in performance marks the model's generalization boundary. + +Pose Augmentation Policy. We conduct experiments using a modified version of RAP, with the pose augmentation approach identical to that of LENS [42], as shown in the Table 7, where our method obtains superior performance. This may be because, although LENS's pose augmentation policy covers a broader spatial area than the training set, its synthetic data may have lower NVS image quality in many unseen regions, which could negatively affect APR training. Density of Training Data. As shown in Table 8, our method with the proposed augmentation significantly reduces errors as the density of real training data increases from $20\%$ to $80\%$ . However, the localization accuracy remains almost unchanged from $80\%$ to $100\%$ , as the scene is already sufficiently covered. Notably, using $100\%$ of the training data without augmentation can result in a significantly higher maximum error in translation, nearly double that with only $20\%$ of the training data with augmentation, despite its limited spatial coverage. This suggests that our augmentation method successfully prevents overfitting to the training data, improving generalization to the test set. + +Quality of Training Data. We evaluate the impact of synthetic image quality on model performance in Table 9, using $20\%$ and $50\%$ of the real data for pose regression. For Shop, it is evident that fewer training samples in 3DGS result in lower-quality rendered views (as indicated by lower PSNR), leading to suboptimal localization performance, particularly for rotation. Surprisingly, localization performance using only $20\%$ of the data for training suboptimal 3DGS and pose regression surpasses the results obtained with $100\%$ of the data without augmentation, as shown in Table 8. This experiment confirms the need for a robust NVS model and the proposed augmentation method in APR training. + +Table 8. Impact of the density of real training data. Our augmentation improves the model's ability to generalize across the entire scene, although this effect has an upper limit. + +
Training Pose %w/ Appearance & Pose Aug. (cm/°)w/o Aug. (cm/°)
100%80%60%40%20%100%50%
Med Err ↓33/1.2632/1.2737/1.9057/2.2387/3.6598/3.75104/4.17
Avg Err ↓41/1.5140/1.5048/2.1762/2.8191/4.65128/4.49139/5.33
Max Err ↓155/4.52158/4.09193/9.39230/11.06231/15.45490/20.73500/21.02
Min Err ↓3/0.177/0.207/0.186/0.3812/0.4613/0.639/0.48
+ +Table 9. Impact of synthetic image quality. Training with higher-quality synthetic images from advanced NVS models enhances localization performance. + +
3DGS PerformanceLocalization Performance (cm/°)
% Images (Train)PSNR ↑ (Train)PSNR ↑ (Test)% Images (Train)Med Err ↓Avg Err ↓Max Err ↓Min Err ↓
20%29.0815.9820%58/3.5968/4.31211/12.1914/0.51
20%29.0815.9850%43/2.4755/3.26196/21.117/0.40
50%26.8817.5550%37/1.8848/2.37184/10.179/0.52
100%24.6018.3050%35/1.6441/2.12130/11.074/0.38
+ +Table 10. Ablation on different 3D representations. + +
MethodsCollegeHospitalShopChurchAverage
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90
DFNetGS102/2.31137/8.0877/3.92123/4.68110/4.75
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
+ +Different 3D Representations. We evaluate the impact of different 3D representations on the performance in Table 10. Trivially replacing NeRF with 3DGS in existing frameworks degrades performance due to 3DGS's inferior 3D consistency. This shows the effectiveness of our proposed joint training paradigm in RAP, which better utilizes diverse synthetic data to learn appearance-invariant features, rather than naively transferring from NeRF to 3DGS. + +# 5. Conclusion + +Summary. We address absolute pose regression with a robust two-branch joint training framework based on Transformer, coupled with an efficient data synthesis pipeline leveraging 3D Gaussian Splats (3DGS) to synthesize numerous posed images with diverse appearances as additional supervision. Our RAP achieves state-of-the-art localization performance, even under challenging appearance variations. Moreover, we thoroughly investigate the impact of synthesizing diverse data and present a novel perspective on APR: generalizability can emerge if the learning gap in APR is effectively addressed together with diverse data. We believe our RAP could be a promising starting point, and the experiments presented in the paper can provide useful insights for future research in this field. + +Acknowledgment. This work was supported in part through NSF grants 2238968, 2121391, and 2024882, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025). + +# References + +[1] Joydeep Biswas and Manuela Veloso. Depth camera based indoor mobile robot localization and navigation. In 2012 IEEE International Conference on Robotics and Automation, pages 1697-1702, 2012. 1 +[2] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5847-5865, 2021. 2, 5, 6 +[3] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6684-6692, 2017. 1, 5 +[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044-5053, 2023. 1, 2, 4, 5, 6, 7 +[5] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2018. 1, 2, 5, 6 +[6] David M. Chen, Georges Baatz, Kevin Koser, Sam S. Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, Bernd Girod, and Radek Grzesczuk. City-scale landmark identification on mobile devices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2011. 2 +[7] Shuai Chen, Zirui Wang, and Victor Prisacariu. Directposenet: Absolute pose regression with photometric consistency. In 2021 International Conference on 3D Vision (3DV), 2021. 2 +[8] Shuai Chen, Xinghui Li, Zirui Wang, and Victor A Prisacariu. Dfnet: Enhance absolute pose regression with direct feature matching. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[9] Shuai Chen, Yash Bhalgat, Xinghui Li, Jia-Wang Bian, Kejie Li, Zirui Wang, and Victor Adrian Prisacariu. Neural refinement for absolute pose regression with feature synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20987-20996, 2024. 1, 4, 5, 6 +[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 6 +[11] Manuela Chessa, Chiara Bassano, and Fabio Solari. Detection and localization of changes in immersive virtual reality. In International Conference on Image Analysis and Processing, pages 121-132, 2023. 1 +[12] Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. Swag: Splitting in the wild images with appearance-conditioned gausians. In European Conference on Computer Vision, pages 325-340, 2025. 2 + +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3 +[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 2 +[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 6 +[17] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8):930–943, 2003. 2, 6 +[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1 +[19] Hugo Germain, Daniel DeTone, Geoffrey Pascoe, Tanner Schmidt, David Novotny, Richard Newcombe, Chris Sweeney, Richard Szeliski, and Vasileios Balntas. Feature query networks: Neural surface description for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5071-5081, 2022. 1, 5, 6 +[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[21] Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. In Signal Processing, Sensor Fusion, and Target Recognition VI, pages 182 - 193, 1997. 6 +[22] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[23] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5974-5983, 2017. 4 +[24] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In IEEE international conference on computer vision, pages 2938-2946, 2015. 1, 2, 4, 5, 6, 7 +[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time + +radiance field rendering. ACM Transaction on Graphics, 42 (4), 2023. 1, 2, 3 +[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 6 +[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 6 +[28] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 4, 5, 6 +[29] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems, 2024. 2 +[30] Jingyu Lin, Jiaqi Gu, Bojian Wu, Lubin Fan, Renjie Chen, Ligang Liu, and Jieping Ye. Learning neural volumetric pose features for camera localization. In European Conference on Computer Vision, 2024. 1, 2, 3, 4, 5, 6 +[31] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5166-5175, 2024. 2 +[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 2 +[33] Changkun Liu, Shuai Chen, Yukun Zhao, Huajian Huang, Victor Prisacariu, and Tristan Braud. Hr-apr: Apr-agnostic framework with uncertainty estimation and hierarchical refinement for camera relocalisation. In 2024 IEEE International Conference on Robotics and Automation, pages 8544–8550, 2024. 1, 5, 6 +[34] Changkun Liu, Shuai Chen, Yash Sanjay Bhalgat, Siyan HU, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient camera pose refinement via 3d gaussian splatting. In The Thirteenth International Conference on Learning Representations, 2025. 5, 6, 7 +[35] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2 +[36] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE International Conference on Computer Vision, pages 2372-2381, 2017. 5, 7 +[37] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 2794-2802, 2017. 4 +[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel + +Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1 +[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1 +[41] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Coordinate: uncertainty-aware pose regressor for reliable vehicle localization. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2229-2238, 2022. 2, 5, 6 +[42] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In Conference on Robot Learning, pages 1347-1356, 2022. 1, 2, 5, 6, 8 +[43] Arthur Moreau, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Crossfire: Camera relocalization on self-supervised features from an implicit representation. In IEEE/CVF International Conference on Computer Vision, pages 252-262, 2023. 1, 5, 6 +[44] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In IEEE international conference on computer vision, pages 3456-3465, 2017. 2 +[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019. +[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020. +[47] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2016. 2 +[48] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6 dof outdoor visual localization in changing conditions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 4, 5, 7 +[49] Torsten Sattler, Qunjie Zhou, Marc Pollefeys, and Laura Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3302-3312, 2019. 1, 2, 7 + +[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 4 +[51] Yoli Shavit and Yosi Keller. Camera pose auto-encoders for improving pose regression. In European Conference on Computer Vision, pages 140–157, 2022. 5, 6 +[52] Yoli Shavit, Ron Ferens, and Yoshi Keller. Paying attention to activation maps in camera pose regression. arXiv preprint arXiv:2103.11477, 2021. 2 +[53] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 1, 5, 6 +[54] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 4, 6, 7 +[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 7 +[56] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM Transactions on Graphics, pages 835-846, 2006. 2 +[57] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7199-7209, 2018. 2 +[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114, 2019. 3 +[59] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The unreasonable effectiveness of pretrained features for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12798, 2024. 1, 5, 6 +[60] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 6 +[61] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 5, 6, 7 +[62] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. In ACM in Computer Graphics and Interactive Techniques, 2024. 2, 3 +[63] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, 2024. 2, 3 + +[64] Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, and Zhaopeng Cui. Pnerfloc: Visual localization with point-based neural radiance fields. In AAAI Conference on Artificial Intelligence, pages 7450-7459, 2024. 1, 5 +[65] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In IEEE International Conference on Computer Vision, pages 2075-2083, 2015. 5 +[66] Qunjie Zhou, Maxim Maximov, Or Litany, and Laura Leal-Taixe. The nefrect match: Exploring nef features for visual localization. arXiv preprint arXiv:2403.09577, 2024. 1, 5 \ No newline at end of file diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/images.zip b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9c2beb245f65f7fd0e73b6d1fbeb28cd377808d8 --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0728effc5d388b5039d7a90ef124608159e13f5288f155576cf2003ba8c8857 +size 855744 diff --git a/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/layout.json b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7d0586999f3148115fa427601b5db3bffc383817 --- /dev/null +++ b/2025/Adversarial Exploitation of Data Diversity Improves Visual Localization/layout.json @@ -0,0 +1,9646 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "text", + "content": "Adversarial Exploitation of Data Diversity Improves Visual Localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 508, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 508, + 159 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 508, + 159 + ], + "type": "text", + "content": "Sihang Li* Siqi Tan* Bowen Chang Jing Zhang Chen Feng Yiming Li" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 163, + 357, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 163, + 357, + 177 + ], + "spans": [ + { + "bbox": [ + 252, + 163, + 357, + 177 + ], + "type": "text", + "content": "New York University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 178, + 418, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 178, + 418, + 190 + ], + "spans": [ + { + "bbox": [ + 192, + 178, + 418, + 190 + ], + "type": "text", + "content": "{sihangli, tan.k, cfeng, yimingli}@nyu.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 230, + 193, + 376, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 193, + 376, + 204 + ], + "spans": [ + { + "bbox": [ + 230, + 193, + 376, + 204 + ], + "type": "text", + "content": "https://ai4ce.github.io/RAP" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 232, + 200, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 244 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 244 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": "Visual localization, which estimates a camera's pose within a known scene, is a fundamental capability for autonomous systems. While absolute pose regression (APR) methods have shown promise for efficient inference, they often struggle with generalization. Recent approaches attempt to address this through data augmentation with varied viewpoints, yet they overlook a critical factor: appearance diversity. In this work, we identify appearance variation as the key to robust localization. Specifically, we first lift real 2D images into 3D Gaussian Splats with varying appearance and deblurring ability, enabling the synthesis of diverse training data that varies not just in poses but also in environmental conditions such as lighting and weather. To fully unleash the potential of the appearance-diverse data, we build a two-branch joint training pipeline with an adversarial discriminator to bridge the syn-to-real gap. Extensive experiments demonstrate that our approach significantly outperforms state-of-the-art methods, reducing translation and rotation errors by " + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": " on indoor datasets, and " + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 55, + 256, + 296, + 533 + ], + "type": "text", + "content": " on outdoor datasets. Most notably, our method shows remarkable robustness in dynamic driving scenarios under varying weather conditions and in day-to-night scenarios, where previous APR methods fail." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 135, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 565, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 696 + ], + "type": "text", + "content": "Visual localization, the task of calculating a 6-DoF camera pose—its translation and rotation—based on a query image within a given environment, is essential for various applications, including robotics [1], autonomous vehicles [18], and virtual reality [11]. Besides traditional geometry-based approaches, recent learning-based visual localization methods adopt absolute pose regression (APR) [5, 8, 24, 53], scene coordinate regression (SCR) [3, 4, 43, 61], or post pose refinement (PPR) [9, 19, 33, 43, 59, 64, 66]. SCR methods focus on learning-based 2D-3D correspondences followed by subsequent Perspective-n-Point (PnP) for pose estimation." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 234, + 555, + 396 + ], + "blocks": [ + { + "bbox": [ + 315, + 234, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 315, + 234, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 234, + 555, + 396 + ], + "type": "image", + "image_path": "61b4c710face4e57011a77d867b8ed225c07bac95a93e4145e0c2c5241857bc4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 469 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 469 + ], + "type": "text", + "content": "Figure 1. We propose RAP, a novel pipeline to train robust APR models. We lift real-world 2D images into 3D Gaussian Splats [25] to synthesize images with diverse appearances and poses, improving model generalizability. We also introduce an adversarial discriminator, mitigating the syn-to-real gap to learn robust features. Together, we achieve state-of-the-art performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 592 + ], + "type": "text", + "content": "PPR methods heavily rely on a pose prior, usually obtained from image retrieval, followed by iterative refinement. In contrast, APR methods employ a supervised framework to train a regression neural network on image-pose pairs, enabling direct pose estimation during inference. APR offers faster runtime and lower error in challenging scenes with sparse views, significant lighting variations, or numerous dynamic objects, making it a promising method for ensuring robustness in real-world applications." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Despite promises, there is a performance gap in localization accuracy between APR and other methods. A well-known pivotal work [49] attributes this to APR performing image-based memorization, i.e., retrieving poses seen during training. Driven by this crucial finding, to improve such memorization while avoiding the need for denser real-world training samples, recent methods leverage Neural Radiance Fields (NeRF) [40] to synthesize additional posed images for APR training [8, 30, 42]. LENS [42] tried to employ appearance perturbation using NeRF-W [39], but found the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 225, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 225, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 225, + 714 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26848" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "improvements minimal [42]. As appearance augmentation is common in other learning tasks, why does it fail in APR?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 295, + 275 + ], + "type": "text", + "content": "We hypothesize a learning gap: Limitations in previous training pipelines prevented the effective use of diverse data to boost performance. Artifacts always exist in images rendered by common novel view synthesis (NVS) methods, which might disturb the model feature space. Inspired by generative adversarial networks (GAN) [20], where a discriminator is trained to distinguish between real and generated samples, we propose adversarial training for APR, designing a discriminator to align the features of synthetic and real images, thereby reducing the syn-to-real domain gap and mitigating the impact of rendering artifacts. Augmentation quality also matters. To efficiently synthesize diverse high-quality images with controllable varying appearance, we extend the vanilla 3D Gaussian Splitting (3DGS) [25] to appearance-varying 3DGS with deblurring ability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 276, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 455 + ], + "type": "text", + "content": "These form our two-branch joint training framework for robust absolute pose regression (RAP). The first branch coarsely trains our Transformer-based pose regressor with both real data and data synthesized at the original real pose, together with an adversarial discriminator to reduce the synto-real domain gap. The second branch progressively generates randomly perturbed poses and appearances, providing additional supervision to the same APR Transformer. Through extensive experiments, we demonstrate that exploiting data diversity using adversarial training significantly increases localization accuracy in APR. Meanwhile, our results indicate that APR consistently benefits from more diverse visual data, and we observe clear signs of a more generalizable APR emerging with its localization performance cannot be explained merely by memorization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 455, + 253, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 455, + 253, + 465 + ], + "spans": [ + { + "bbox": [ + 67, + 455, + 253, + 465 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 295, + 610 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 467, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 514 + ], + "type": "text", + "content": "- We identify the crucial role of appearance diversity for APR, and develop a 3DGS-based appearance-varying data augmentation framework to efficiently generate diverse synthetic data with controllable lighting conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 295, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 295, + 562 + ], + "type": "text", + "content": "- We propose an adversarial discriminator to reduce the syn-to-real gap. Together with progressive data synthesis, we form a robust two-branch joint training pipeline that fully unleashes the power of data diversity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 610 + ], + "type": "text", + "content": "- We conduct extensive experiments showing our method outperforms state-of-the-art approaches on challenging datasets with significant appearance change. Ablation studies further analyze key factors affecting performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 615, + 147, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 147, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 147, + 627 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Visual Localization. Visual localization aims to estimate a camera's translation and rotation within a 3D scene. Traditional geometry-based methods [6, 15, 32, 35, 44-47, 57] accomplish this by using point clouds and a reference image database, relying on stored descriptors and image retrieval to establish 2D-3D correspondences. In contrast, scene coordinate regression (SCR) methods [2-4, 61] em" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "bed map information within neural networks to directly predict 2D-3D correspondences. Both approaches generally require PnP [17] and RANSAC [16] to output camera poses at test time, which adds additional computation cost. Alternatively, absolute pose regression (APR) [5, 7, 22, 24, 41, 52] aims to directly regress the camera pose from a query image using neural networks. Although the performance is suboptimal compared with geometry-based methods, APR remains a promising approach due to its fast inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 180, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 418 + ], + "type": "text", + "content": "Data Augmentation for Pose Regression. End-to-end pose regression methods rely heavily on the amount and diversity of training data. Previous work [49] shows that APR implicitly learns image retrieval in the given environment. Therefore, the following works LENS [42], DFNet [8] and PMNet [30] enhance APR performance by spatially enriching training views with NeRF. However, these approaches fail to address the generalizability of APR models and exhibit several limitations: (1) The efficiency of training and novel view synthesis (NVS) in NeRF is severely restricted, hindering scalability. (2) They limit NVS to geometric (pose) transformations while neglecting photometric (appearance) variations, thereby decreasing APR robustness to changes in visual appearance. (3) The augmented data is underutilized in their learning frameworks, leaving its potential for improving APR largely untapped. Differently, our framework switches to 3DGS [25] as the scene representation to efficiently generate novel posed images with controllable appearances and introduce adversarial training to unleash the power of such diverse data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 658 + ], + "type": "text", + "content": "Handling and Synthesizing Challenging Scenarios. Visual localization often encounters unstructured photo collections [56], where visual appearance varies due to moving objects, lighting changes, and inconsistent camera exposure settings. To tackle these in-the-wild challenges, NeRF-W [38] uses per-image transient and appearance embeddings. In 3DGS [25], VastGaussian [31] applies a CNN to 3DGS outputs but still struggles with significant appearance variations. SWAG [12] mitigates this issue by storing appearance information in an external hash-grid-based implicit field, while GS-W [63] enhances flexibility by separating intrinsic and dynamic appearance features for each Gaussian point. 3DGM [29] leverages consensus across multiple sequences as the self-supervision signal to remove transient and moving objects without human annotations. Deblur-GS [62] addresses motion blur—another challenge in localization datasets—by modeling camera motion to yield sharper edges in rendered scenes. Our method incorporates appearance modeling and edge refinement to handle and synthesize diverse indoor, outdoor, and driving scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 662, + 370, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 370, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 370, + 673 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 462, + 689 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 462, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 462, + 689 + ], + "type": "text", + "content": "3.1. Pre-Processing with 3DGS" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "A robust pose regressor should focus on intrinsic scene attributes, not appearance variations. Therefore, we first" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26849" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 210 + ], + "type": "image", + "image_path": "ded7d99ee2f9af0dfe52d63070baaaed376e73807554e871bd73f06b20609cb5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 253 + ], + "type": "text", + "content": "Figure 2. Pipeline of RAP. We lift multiple RGB video sequences into 3D Gaussian Splats, which serve as our data engine. The branch-1 (see Sec. 3.3.1) inputs paired real and synthetic images to regress poses, with a discriminator to bridge the syn-to-real gap. The branch-2 (see Sec. 3.3.2) generates views with novel poses and appearances, which are fed into the same pose regressor as additional supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "text", + "content": "To further analyze the behavior of the " + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "inline_equation", + "content": "K \\times 16 \\times 3" + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "text", + "content": " system, we assume that the camera is located at the center of the image. The camera is positioned at the center of the image by a camera position vector " + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{C} = [0, 1]" + }, + { + "bbox": [ + 56, + 268, + 296, + 435 + ], + "type": "text", + "content": ". To achieve this, we use the following procedure to generate the camera positions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 123, + 446, + 295, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 446, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 123, + 446, + 295, + 459 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {C}} = \\operatorname {M L P} (\\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\omega \\boldsymbol {\\mathcal {E}}, \\theta), \\tag {1}", + "image_path": "4b38aab494884b8b6cc3177ca04dad8644fd386ec796f5dea8ddd07cfd8d0620.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 470, + 295, + 493 + ], + "type": "text", + "content": " is the blending weight that controls the dynamic appearance of the rendered image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": "Another significant challenge in visual localization is motion blur, often caused by slow shutter speeds during video capture, leading to pose ambiguity and degraded rendering quality, further decreasing localization accuracy. Inspired by Deblur-GS [62], we model camera motion blur as the inverse of scene motion, i.e., the transformation in Gaussian position denoted by " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": ". For each training image, we sample a certain time step along a linear trajectory with a sampling weight " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": " and blend them to compute loss " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": " with the original blurry image " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_b \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": ", optimizing " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": " and other 3DGS parameters " + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 495, + 295, + 626 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 635, + 296, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 296, + 669 + ], + "type": "interline_equation", + "content": "\\underset {\\phi , \\boldsymbol {\\mathcal {T}}, \\boldsymbol {\\mu}, \\boldsymbol {\\mathcal {Y}}, \\boldsymbol {\\mathcal {E}}, \\boldsymbol {\\Theta}} {\\operatorname {a r g m i n}} \\mathcal {L} \\left(\\boldsymbol {\\mathcal {I}} _ {b}, \\sum_ {i = 1} ^ {n} \\phi_ {i} \\operatorname {R e n d e r} \\left(\\boldsymbol {\\mathcal {T}} _ {i} (\\boldsymbol {\\mu}), \\boldsymbol {\\mathcal {C}}, \\boldsymbol {\\Theta}\\right)\\right), \\tag {2}", + "image_path": "714321e56f4ea8e17d886f6d5f70c473cd31ff6a984299f07f258279d8c7cada.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where the details of " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " are in supplemental materials. After training, our 3DGS can efficiently render posed images given " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 267, + 481, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 481, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 481, + 280 + ], + "type": "text", + "content": "3.2. Architecture of Pose Regressor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "content": "Given a set of images and their associated camera poses " + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "inline_equation", + "content": "\\{(I_i, P_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "content": ", our goal is to train a neural network to directly output a homogeneous camera pose " + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "content": " for a query image " + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 285, + 555, + 344 + ], + "type": "text", + "content": ". Our network architecture is shown as the pose regressor in Fig. 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 345, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 393 + ], + "type": "text", + "content": "Feature Extractor. Pose regression networks typically extract features using a common backbone " + }, + { + "bbox": [ + 313, + 345, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 345, + 555, + 393 + ], + "type": "text", + "content": ", such as VGG [55] or EfficientNet [58], leveraging multiple deeper layers for translation and rotation regression:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 352, + 399, + 553, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 399, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 352, + 399, + 553, + 414 + ], + "type": "interline_equation", + "content": "\\varphi (\\boldsymbol {I}) = \\left\\{\\mathcal {F} _ {0} (\\boldsymbol {I}), \\dots , \\mathcal {F} _ {N - 1} (\\boldsymbol {I}), \\mathcal {F} _ {N} (\\boldsymbol {I}) \\right\\}, \\tag {3}", + "image_path": "d1e3891e28ced6820685570118a3137cbbf614f4f01b36d79110caa320906beb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{*}(\\cdot)" + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "content": " denotes features extracted from the " + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "content": "-th layer of a backbone with " + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "content": " layers. " + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(\\pmb {I})" + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r(\\pmb {I})" + }, + { + "bbox": [ + 313, + 419, + 554, + 455 + ], + "type": "text", + "content": " denote features for translation and rotation regression, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "text", + "content": "Pose Transformer. Unlike CNN-based regression models [8, 30], where fine-grained local features can introduce noise and harm performance, we propose Pose Transformer to leverage the strong ability of Vision Transformer (ViT) [14] for modeling long-range dependencies. Each Transformer generates a global token (Trans for translation and Rot for rotation) to provide a comprehensive context for pose regression, inspired by the CLS token in ViT. Given " + }, + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_r(I)" + }, + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t(I)" + }, + { + "bbox": [ + 313, + 455, + 554, + 575 + ], + "type": "text", + "content": ", the translation token is then concatenated with the flattened input features*:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 580, + 553, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 580, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 315, + 580, + 553, + 607 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\operatorname {C a t} (\\text {F l a t t e n} (\\mathcal {F} _ {t} (\\boldsymbol {I})), \\text {T r a n s}) \\in \\mathbb {R} ^ {\\left(H _ {t} W _ {t} + 1\\right) \\times C _ {t}}. \\tag {4}", + "image_path": "0bc59864e0373e827fa4d934d7a41fd904c83dfef317e05e43947ae80f1df0c5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 607, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 656 + ], + "type": "text", + "content": "The positional encodings are then added to the flattened feature " + }, + { + "bbox": [ + 313, + 607, + 555, + 656 + ], + "type": "inline_equation", + "content": "(\\mathsf{PE} + \\widetilde{\\mathcal{F}}_t(\\pmb {I}))\\in \\mathbb{R}^{(H_tW_t + 1)\\times C_t}" + }, + { + "bbox": [ + 313, + 607, + 555, + 656 + ], + "type": "text", + "content": ". Multi-head Self-Attention (MSA) is then conducted through a stack of multiple layers with the post-processing as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 662, + 553, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 662, + 553, + 684 + ], + "spans": [ + { + "bbox": [ + 342, + 662, + 553, + 684 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {F} _ {t} ^ {\\prime}} (\\boldsymbol {I}) = \\operatorname {M S A} \\left(\\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I})\\right) + \\mathrm {P E} + \\widetilde {\\mathcal {F} _ {t}} (\\boldsymbol {I}), \\tag {5}", + "image_path": "4c78eb15e8a47c0f82bce3db2d08ed760e8e52b0a328e349903407046ce45ed9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 681, + 518, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 681, + 518, + 696 + ], + "spans": [ + { + "bbox": [ + 345, + 681, + 518, + 696 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {F}} _ {t} (\\boldsymbol {I}) = \\mathrm {L N} (\\mathrm {F F N} (\\mathrm {L N} (\\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I}))) + \\widehat {\\mathcal {F}} _ {t} ^ {\\prime} (\\boldsymbol {I})),", + "image_path": "24da5fe6e97a5247c6fd0466abdf0872c98feb02c3f6adee08413dd60d31f973.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 512, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 512, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 512, + 713 + ], + "type": "text", + "content": "*We only present the translation regression for simplicity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26850" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "where LN indicates layer normalization and FFN denotes the fully connected feed-forward network, consisting of two linear layers with a ReLU. The final output is flattened back to " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "(H_{t}W_{t} + 1)\\times c_{t}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ". See supplementary for more details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 167 + ], + "type": "text", + "content": "Regression Head. Only the processed translation token, Trans, capturing global features for regression, is fed into the regression head. This regression head consists of two MLPs, each with a hidden layer and GeLU activation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 96, + 171, + 294, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 171, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 96, + 171, + 294, + 186 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {t}} = \\operatorname {L i n e a r} (\\operatorname {G e L U} (\\operatorname {L i n e a r} (\\widehat {\\text {T r a n s}}))). \\tag {6}", + "image_path": "2fcaf2dd6da8d8ca015dd88f9f30c1e0f830ab36b38272ba90cb3c6dec72227b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "text", + "content": " represents the final prediction for translation. Similarly, we obtain the rotation prediction denoted by " + }, + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "inline_equation", + "content": "\\hat{r}" + }, + { + "bbox": [ + 55, + 190, + 295, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 220, + 255, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 255, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 255, + 232 + ], + "type": "text", + "content": "3.3. Two-Branch Joint Training Paradigm" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 237, + 284, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 284, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 284, + 248 + ], + "type": "text", + "content": "3.3.1. Branch-1: Aligning Features via Discriminator" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 312 + ], + "type": "text", + "content": "Synthetic images from 3DGS provide novel viewpoints and appearances but often contain artifacts, leading to a syn-to-real domain gap. To align features from rendered and real images of the same pose, we introduce an adversarial training mechanism besides the basic pose regression training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": "Pose Regression Loss. For basic training, we render the synthetic image " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": " with the same pose label " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": " as the real image " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": ", both used as supervision for the pose regressor. The training objective consists of translation loss " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": " and the rotation loss " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_r" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": ", which are measured by the Euclidean distance between the ground truth pose " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\pmb{P} = \\{\\hat{t},\\hat{r}\\}" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": " and the estimated pose " + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "inline_equation", + "content": "\\widehat{P} = \\{\\hat{t},\\hat{r}\\}" + }, + { + "bbox": [ + 55, + 312, + 296, + 397 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 144, + 412, + 294, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 412, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 144, + 412, + 294, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\| \\boldsymbol {t} - \\hat {\\boldsymbol {t}} \\| _ {2}, \\tag {7}", + "image_path": "b8e7f861907cca4e68727709df0ce0e0d4702de89624be6c6cf0b70ac24de1f3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 134, + 429, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 429, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 134, + 429, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r} = \\left\\| \\boldsymbol {r} - \\frac {\\hat {\\boldsymbol {r}}}{\\| \\hat {\\boldsymbol {r}} \\|} \\right\\| _ {2}, \\tag {8}", + "image_path": "7e31e49f936bd7b9ebcfe54b77e76a69e3ca9645db077e45d4b0563be424536d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 457, + 294, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 457, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 72, + 457, + 294, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s e}} ^ {1} = \\mathcal {L} _ {t} \\exp (- s _ {t}) + s _ {t} + \\mathcal {L} _ {r} \\exp (- s _ {r}) + s _ {r}, \\tag {9}", + "image_path": "1b909f51e2d4cff5d779d77121b4d8291d346bda3859a5554aba46753311e8bf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " are learned parameters for balancing the optimization between rotation and translation [23]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": "Adversarial Loss. The adversarial training mechanism optimizes the discriminator to distinguish real from rendered image features, while training the feature extractor to fool the discriminator, effectively bridging the domain gaps. To prevent vanishing gradients, we propose a novel adversarial objective for pose regression, inspired by LSGAN [37]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 64, + 575, + 294, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 575, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 64, + 575, + 294, + 622 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {D} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\mathrm {D i s}} (D) = \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} \\sim p _ {\\mathrm {d a t a}} (\\boldsymbol {I})} \\left[ \\left(D (\\operatorname {A d j} (\\mathcal {F} _ {t} (\\boldsymbol {I}))) - 1\\right) ^ {2} \\right] \\\\ + \\frac {1}{2} \\mathbb {E} _ {\\boldsymbol {I} ^ {\\prime}} \\left[ D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(\\boldsymbol {I} ^ {\\prime}\\right)\\right) ^ {2} \\right], \\right. \\tag {10} \\\\ \\end{array}", + "image_path": "3edf91002820f43b20fe5441486e6e6032d70e0e104942e6af96bf063ca5f50e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 625, + 294, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 294, + 649 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 294, + 649 + ], + "type": "interline_equation", + "content": "\\underset {G} {\\operatorname {a r g m i n}} \\mathcal {L} _ {\\text {G e n}} (G) = \\frac {1}{2} \\mathbb {E} _ {I ^ {\\prime}} \\left[ \\left(D \\left(\\operatorname {A d j} ^ {\\prime} \\left(\\mathcal {F} _ {t} \\left(I ^ {\\prime}\\right)\\right) - 1\\right) ^ {2} \\right]. \\right. \\tag {11}", + "image_path": "2d1b4623b7dfe4e2e9b016288007aac3a3db89445543ecbd5397d58995892ac7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Here, Adj and " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Adj}'" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " are the adjustment layers, consisting of Conv-ReLU-BN layers. The feature extractor " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " acts as the generator " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " is the discriminator, composed of several convolutional layers with ReLU activations. More details are in supplemental materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "text", + "content": "3.3.2. Branch-2: Training while Synthesizing Data" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": "With the proposed appearance-varying 3DGS, more posed images are generated to enrich the training data for better generalizability. Specifically, our data synthesis is categorized into two dimensions: pose augmentation and appearance augmentation, as illustrated in Fig. 1. For pose augmentation, given a training pose " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": ", a perturbed pose " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": " is generated around " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": " by the translation noise of " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\delta t" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": " and rotation noise of " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\delta r" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": ". For appearance augmentation, we randomly adjust the appearance of rendered images using random blending weights " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": ", and then render the synthetic image " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{syn}}" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": " using the Gaussian Splats trained in Sec. 3.1. The novel image-pose pair " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{syn}}, P_{\\mathrm{syn}})" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": ", online generated every 20 epochs during training until the validation MSE loss and median errors cease to decrease, serves as additional supervision for the training. Given the estimated pose of the synthesized image denoted by " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\widehat{P_{\\mathrm{syn}}}" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": ", the loss function " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pose}}^2(\\widehat{P_{\\mathrm{syn}}}, P_{\\mathrm{syn}})" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": " is same as " + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pose}}^1" + }, + { + "bbox": [ + 313, + 87, + 555, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 302, + 418, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 418, + 314 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 418, + 314 + ], + "type": "text", + "content": "3.3.3. Overall Objective" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 318, + 471, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 318, + 471, + 330 + ], + "spans": [ + { + "bbox": [ + 314, + 318, + 471, + 330 + ], + "type": "text", + "content": "The total loss for the pose regressor is:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 331, + 339, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 339, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 331, + 339, + 553, + 354 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\beta_ {1} \\mathcal {L} _ {\\text {p o s e}} ^ {1} + \\beta_ {2} \\mathcal {L} _ {\\text {p o s e}} ^ {2} + \\beta_ {3} \\left(\\mathcal {L} _ {\\text {G e n}} + \\mathcal {L} _ {\\text {D i s}}\\right), \\tag {12}", + "image_path": "6666c10c98a0ff9f22efd23f7dba98854e73556e8f9d31ef9feb607c2741a57a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 363, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 363, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2},\\beta_{3}" + }, + { + "bbox": [ + 313, + 363, + 554, + 411 + ], + "type": "text", + "content": " are loss weights. The total loss will optimize the pose regressor, adjustment layers, and discriminator. Only the pose regressor will be deployed in the inference phase, while the other two components are discarded." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 424, + 395, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 424, + 395, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 424, + 395, + 437 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 444, + 417, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 444, + 417, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 444, + 417, + 456 + ], + "type": "text", + "content": "4.1. Evaluation Setup" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "Datasets. We follow previous works [8, 30] to mainly use four scenes in the Cambridge Landmarks dataset [24] with spatial extents around " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "875\\mathrm{m}^2" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": ". Moreover, we evaluate our method on MARS [28], a self-driving dataset featuring challenges like moving objects, lighting changes, and motion blur. To investigate the robustness of our model under extreme lighting changes, such as the transition from day to night, we also prepared a subset of the Aachen Day-Night dataset [48]. The training data includes images captured using various camera models with differing resolutions, which renders direct evaluation with APR methods infeasible. Thus, we standardized the camera models through center cropping and built a COLMAP [50] model as pose annotations, including 13 nighttime images for evaluation and 246 daytime images for training 3DGS and RAP. We also employ the 7-Scenes dataset [54], which provides seven indoor scenes with volumes spanning " + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}^3 - 18\\mathrm{m}^3" + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": " and follow the original training and testing splits with more accurate SfM pose annotations [4, 9]. Although it is an indoor dataset, it is still non-trivial as it includes texture-less surfaces, object occlusions, and motion blur." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26851" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 558, + 233 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 558, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 558, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 558, + 233 + ], + "type": "image", + "image_path": "15ebf38683dbd6908706749e21ed47014c0be1f8e9f74ee3345b54c5b2a5dc5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of camera pose estimation errors between a) DFNet [8] and b) our RAP framework across five scenes on the Cambridge Landmarks dataset [24]. Our RAP framework estimates trajectories that more closely follow the ground truth, with significantly reduced rotation and position errors compared to DFNet [8]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 316, + 294, + 529 + ], + "blocks": [ + { + "bbox": [ + 55, + 281, + 295, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 314 + ], + "type": "text", + "content": "Table 1. Median translation (cm) and rotation " + }, + { + "bbox": [ + 55, + 281, + 295, + 314 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 55, + 281, + 295, + 314 + ], + "type": "text", + "content": " errors on the Cambridge Landmarks dataset [24]. The best results† in pure APR and PPR are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 316, + 294, + 529 + ], + "lines": [ + { + "bbox": [ + 58, + 316, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 58, + 316, + 294, + 529 + ], + "type": "table", + "html": "
MethodsCollegeHospitalShopChurchAverage‡Court
Pure APRPN [24]166/4.86262/4.90141/7.18245/7.95204/6.23683/3.50
MapNet [5]107/1.89194/3.91149/4.22200/4.53163/3.64N/A
MS-Trans. [53]83/1.47181/2.3986/3.07162/3.99128/2.73N/A
PAE [51]90/1.49207/2.5899/3.88164/4.16140/3.03N/A
LENS† [42]33/0.5044/0.9027/1.6053/1.6039/1.20N/A
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90217/4.11
PMNet [30]68/1.97103/1.3158/2.10133/3.7390/2.27N/A
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28115/1.68
SCRDSAC* [2]18/0.321/0.45/0.315/0.615/0.434/0.2
ACE [4]28/0.431/0.65/0.318/0.621/0.543/0.2
GLACE [61]19/0.317/0.44/0.29/0.312/0.319/0.1
PPRFQN-MN [19]28/0.3854/0.8213/0.6358/2.0038/0.964253/39.16
LENS [42]34/0.5445/0.9628/1.6654/1.6640/1.21N/A
CrossFire [43]47/0.743/0.720/1.239/1.437/1.0N/A
NeFeS50 [9]37/0.5452/0.8815/0.5337/1.1435/0.77N/A
HR-APR [33]36/0.5853/0.8913/0.5138/1.1635/0.78N/A
MCLoc [59]31/0.4239/0.7312/0.4526/0.8827/0.62N/A
DFNetGS-CPR [34]23/0.3242/0.7410/0.3627/0.6226/0.51N/A
ACEGS-CPR [34]20/0.2921/0.405/0.2413/0.4015/0.33N/A
DFNetref(Ours)16/0.2421/0.418/0.4210/0.2614/0.3325/0.13
RAPref (Ours)15/0.2518/0.385/0.239/0.2312/0.2723/0.15
", + "image_path": "707b802e0a70c1f7f8c9395601504ea31d5aaf6a79139cf3f0eaf523765aab02.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 530, + 295, + 560 + ], + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 560 + ], + "type": "text", + "content": "Since most methods did not report results on Court, it is excluded from the average error calculation. As Coordinet + LENS [42] does not provide open-source code, it is unclear whether any post-processing is used." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "Baselines. We first compare our proposed RAP against common APR-only approaches on the four datasets, where PMNet [30] and DFNet [8] are the most related and advanced methods based on data augmentation. We split the remaining methods into two categories based on whether they rely on extra novel view synthesis in test time, including SCR [3, 4, 61] and PPR (Post Pose Refinement) [9, 19, 33, 43, 59, 64, 66], which involves rendering images, querying features in novel views by the initial pose, iterative refinement or sequential refinement [41]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. First, we optimize our 3DGS for each scene without masking moving objects. We" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 279, + 555, + 424 + ], + "blocks": [ + { + "bbox": [ + 315, + 279, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 315, + 279, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 315, + 279, + 555, + 424 + ], + "type": "image", + "image_path": "c57aa03fd01ce4f456179d649cd6a2a01a90affa509ee3d9b08a653d61298417.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "text", + "content": "Figure 4. Visualization of " + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "text", + "content": " on MARS [28]. In each subfigure, a diagonal line separates the \"Predicted\" (rendered from the refined pose) and \"GT\" (ground truth) sections. Smooth alignment along this boundary shows " + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "text", + "content": " 's improved pose accuracy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 334, + 506, + 537, + 552 + ], + "blocks": [ + { + "bbox": [ + 313, + 480, + 554, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 502 + ], + "type": "text", + "content": "Table 2. Median translation (cm) and rotation " + }, + { + "bbox": [ + 313, + 480, + 554, + 502 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 313, + 480, + 554, + 502 + ], + "type": "text", + "content": " errors on the MARS dataset [28]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 506, + 537, + 552 + ], + "lines": [ + { + "bbox": [ + 334, + 506, + 537, + 552 + ], + "spans": [ + { + "bbox": [ + 334, + 506, + 537, + 552 + ], + "type": "table", + "html": "
Methods“0011”“0015”“0037”“0041”Average
PoseNet [24]149/1.80136/2.34123/1.6075/0.92121/1.67
RAP (Ours)32/0.6137/1.0815/0.3528/0.3528/0.60
RAPref (Ours)8.5/0.138.2/0.208.7/0.097.6/0.118.3/0.13
", + "image_path": "95cc2a1ca9e7dfb03a4d4835f9ff48014b46a20798abfdae83d9b40e6cdaeacf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 316, + 591, + 558, + 628 + ], + "blocks": [ + { + "bbox": [ + 313, + 565, + 554, + 588 + ], + "lines": [ + { + "bbox": [ + 313, + 565, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 554, + 588 + ], + "type": "text", + "content": "Table 3. Median translation (COLMAP [65] unit) and rotation " + }, + { + "bbox": [ + 313, + 565, + 554, + 588 + ], + "type": "inline_equation", + "content": "(^{\\circ})" + }, + { + "bbox": [ + 313, + 565, + 554, + 588 + ], + "type": "text", + "content": " errors on the Aachen Day-Night Dataset [48]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 591, + 558, + 628 + ], + "lines": [ + { + "bbox": [ + 316, + 591, + 558, + 628 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 558, + 628 + ], + "type": "table", + "html": "
APR-BasedSCR-Based
PoseNet [24]DFNet [8]RAP w/o App.RAPRAPrefACE [4]GLACE [61]
217/74.30174/85.80134/75.99130/13.7067/4.64914/90.50482/36.4
", + "image_path": "3398e30e20ff627d4ecf30f724f6d702750b29ef74da0c7109874b46e2e30b81.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "then train our RAP network, which uses an Efficient-B0 backbone [36] pre-trained on ImageNet [13], optimized with Adam [26] at a learning rate of " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ". Only the features from the third (reduction_3) and fourth (reduction_4) layers are used respectively for translation and rotation regression, and both layers are utilized for" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26852" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 194 + ], + "type": "image", + "image_path": "c2e0d3a703c676af4979a21260e04cec28f2293f6fbad517b34edcf025eaabbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 198, + 472, + 209 + ], + "lines": [ + { + "bbox": [ + 138, + 198, + 472, + 209 + ], + "spans": [ + { + "bbox": [ + 138, + 198, + 472, + 209 + ], + "type": "text", + "content": "Figure 5. Visualization of the localization errors of " + }, + { + "bbox": [ + 138, + 198, + 472, + 209 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 138, + 198, + 472, + 209 + ], + "type": "text", + "content": " on the 7-Scenes dataset [54]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 107, + 245, + 504, + 448 + ], + "blocks": [ + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 555, + 243 + ], + "type": "text", + "content": "Table 4. Quantitative results on the 7-Scenes dataset [54]. The best results in pure APR and PPR are highlighted in bold. DSLAM GT and SfM GT refer to different sets of ground truth. More visualizations and details are in supplemental materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 245, + 504, + 448 + ], + "lines": [ + { + "bbox": [ + 107, + 245, + 504, + 448 + ], + "spans": [ + { + "bbox": [ + 107, + 245, + 504, + 448 + ], + "type": "table", + "html": "
CategoryMethodsChessFireHeadsOfficePumpkinKitchenStairsAverage
Pure APRPoseNet (PN) [24]32/8.1247/14.429/1.2048/7.6847/8.4259/8.6447/13.8044/10.4
MapNet [5]8/3.2527/11.718/13.317/51.522/4.0223/4.9330/12.121/7.77
AtLoc+ [60]10/3.1826/10.814/11.417/5.1620/3.9416/4.9029/10.219/7.08
MS-Transformer [53]11/4.6624/9.6014/12.217/5.6618/4.4417/5.9417/5.9418/7.28
PAE [51]12/4.9524/9.3114/12.519/5.7918/4.8918/6.1925/8.7419/7.48
CoordiNet + LENS [42]4/1.3811/3.778/5.868/1.989/2.2710/2.2715/3.679/3.07
DFNet [8]5/1.8817/6.456/3.638/2.4810/2.7822/5.4516/3.2912/3.71
PMNet [30]4/1.7010/4.517/4.237/1.9614/3.3314/3.3616/3.6210/3.24
RAP (Ours, DSLAM GT)3/1.417/3.466/6.025/1.976/1.967/2.1810/2.146/2.73
RAP (Ours, SfM GT)2/0.856/2.844/4.525/1.944/1.707/2.1210/1.305/2.18
SCRDSAC [2]0.5/0.170.8/0.280.5/0.341.2/0.341.2/0.280.7/0.212.7/0.781.1/0.34
ACE [4]0.5/0.180.8/0.330.5/0.331.0/0.291.0/0.220.8/0.202.9/0.811.1/0.34
GLACE [61]0.6/0.180.9/0.340.6/0.341.1/0.290.9/0.230.8/0.203.2/0.931.2/0.36
marepo† [10]2.6/1.352.5/1.422.3/2.213.6/1.444.2/1.555.1/1.996.7/1.833.9/1.68
PPRFQN-MN [19]4.1/1.3110.5/2.979.2/2.453.6/2.364.6/1.7616.1/4.42139.5/34.6728/7.3
CrossFire [43]1/0.45/1.93/2.35/1.63/0.82/0.812/1.94.4/1.38
DFNet + NeFeS50 [9]2/0.572/0.742/1.282/0.562/0.552/0.575/1.282.4/0.79
HR-APR [33]2/0.552/0.752/1.452/0.642/0.622/0.675/1.302.4/0.85
MCLoc [59]2/0.83/1.43/1.34/1.35/1.66/1.66/2.04.1/1.43
DFNet + GS-CPR (SfM GT) [34]0.7/0.200.9/0.320.6/0.361.2/0.321.3/0.310.9/0.252.2/0.611.1/0.34
ACE + GS-CPR (SfM GT) [34]0.5/0.150.6/0.250.4/0.280.9/0.261.0/0.230.7/0.171.4/0.420.8/0.25
RAPref (Ours, DSLAM GT)2.83/1.432.07/1.231.54/1.872.55/1.204.59/1.544.23/1.833.48/1.233.04/1.48
RAPref (Ours, SfM GT)0.33/0.110.51/0.210.40/0.290.59/0.170.83/0.200.50/0.121.11/0.320.61/0.20
", + "image_path": "edf6bccc4647ad8798d7869d96b6fac0619d09812bd21c01bfe949b3644b9077.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 198, + 449, + 410, + 459 + ], + "lines": [ + { + "bbox": [ + 198, + 449, + 410, + 459 + ], + "spans": [ + { + "bbox": [ + 198, + 449, + 410, + 459 + ], + "type": "text", + "content": "As marepo [10] combines SCR and APR, we classify it as SCR." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": "narrowing the domain gap via the discriminator, which is also optimized with Adam [26], using a learning rate of " + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": " and betas set to (0.5, 0.999). More details about training are in supplemental materials. For generating random views, we apply random normalized perturbations to each training pose: " + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\delta t = 20\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\delta r = 10^{\\circ}" + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": " for indoor scenes, and " + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\delta t = 150\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\delta r = 4^{\\circ}" + }, + { + "bbox": [ + 54, + 469, + 295, + 552 + ], + "type": "text", + "content": " for outdoor scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 554, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 554, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 554, + 295, + 651 + ], + "type": "text", + "content": "To allow for comparison with SCR methods and leverage 3DGS's efficient rendering for PPR, we extend the APR pipeline with match-based refinement similar to GS-CPR [34], denoted as " + }, + { + "bbox": [ + 54, + 554, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 54, + 554, + 295, + 651 + ], + "type": "text", + "content": ". At test time, RAP's initial pose is used to render an RGB-D image via 3DGS. Together with MASt3R [27], we can obtain 2D-3D correspondences to perform RANSAC-PnP [16, 17], resulting in a refined pose. More details are in supplementary materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 170, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 170, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 170, + 670 + ], + "type": "text", + "content": "4.2. Benchmark Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Cambridge Landmarks [24]. In the challenging outdoor Cambridge Landmarks dataset (Table 1), our RAP reduces both translation and rotation errors across all scenes by over" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": " compared to other APR-only methods. The visualization in Fig. 3 shows that our method produces fewer outliers than DFNet [8]. In the three larger-scale scenes with significant appearance diversity (College, Church, and Court), rotation error is even halved compared to DFNet. Table 1 also shows the effectiveness of our " + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": " in further reducing pose errors through refinement. " + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{CoordiNet} + \\mathrm{LENS}" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": " [42], which assumes a continuous trajectory when an Extended Kalman Filter [21] is required for refinement [41]. " + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": " even surpasses ACE [4] and its post-refinement variant, " + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathrm{ACE} + \\mathrm{GS}" + }, + { + "bbox": [ + 313, + 469, + 555, + 636 + ], + "type": "text", + "content": "-CPR [34], despite GS-CPR manually masking dynamic objects when building 3DGS. This demonstrates the strong representation capability of our appearance-varying 3DGS with deblurring." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "MARS [28]. Autonomous driving scenarios present unique challenges, including moving objects and frequent changes in lighting conditions, as illustrated in Fig. 4. Our RAP demonstrates effective and robust performance across four challenging scenes, as shown in Table 2, achieving an average of " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "28~\\mathrm{cm} / 0.60^{\\circ}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " localization error. This significantly" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26853" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 85, + 261, + 167 + ], + "blocks": [ + { + "bbox": [ + 129, + 71, + 221, + 81 + ], + "lines": [ + { + "bbox": [ + 129, + 71, + 221, + 81 + ], + "spans": [ + { + "bbox": [ + 129, + 71, + 221, + 81 + ], + "type": "text", + "content": "Table 5. Ablation study." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 89, + 85, + 261, + 167 + ], + "lines": [ + { + "bbox": [ + 89, + 85, + 261, + 167 + ], + "spans": [ + { + "bbox": [ + 89, + 85, + 261, + 167 + ], + "type": "table", + "html": "
Setupon ShopTrans. (cm) ↓Rot. (°) ↓
I (Baseline): φ = VGG161745.45
II: φ = Efficient-B01034.64
III: II + Pose Aug.753.52
IV: III + Appearance Aug.603.14
V: IV + Decoder (ConvNet)522.51
VI: V + Decoder (Transformer)401.98
VII (Ours): VI + Discriminator331.48
", + "image_path": "ca43c06e3f28ee1d55e70a74d6fa13ba8fa2d9ae2b53259b4ab5bff36ff623ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": "outperforms the baseline† PoseNet [24]. With one-shot refinement, our " + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": " further reduces outdoor localization errors to below " + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "inline_equation", + "content": "10~\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 182, + 295, + 217 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "text", + "content": "Aachen Day-Night [48]. Benefiting from appearance augmentation, our RAP significantly reduces the localization error from 134 unit " + }, + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "inline_equation", + "content": "/75.99^{\\circ}" + }, + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "text", + "content": " to 130 unit " + }, + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "inline_equation", + "content": "/13.70^{\\circ}" + }, + { + "bbox": [ + 55, + 218, + 295, + 290 + ], + "type": "text", + "content": ", outperforming other APR [8, 24] and SCR [4, 61] baselines, as shown in Table 3. This demonstrates the effectiveness of appearance diversity in handling extreme lighting changes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": "7-Scenes [54]. As shown in Table 4, our RAP reduces translation error by " + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm} \\rightarrow 5\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": ") and rotation error by " + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "32.72\\%" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "3.24^{\\circ} \\rightarrow 2.18^{\\circ}" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": ") on average compared to previous state-of-the-art single-frame APR methods. The only exception is Heads, where the rotation error is suboptimal. This scene consists of just two sequences—one for training and one for testing—potentially limiting the effectiveness of our augmentation in capturing scene variability. Meanwhile, " + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{RAP}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": " further reduces localization error below 1 cm with one-shot refinement using our 3DGS. It also surpasses ACE [4] and its post-refinement variant, " + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "inline_equation", + "content": "\\mathrm{ACE} + \\mathrm{GSCPR}" + }, + { + "bbox": [ + 55, + 290, + 295, + 434 + ], + "type": "text", + "content": " [34]. Qualitative examples are shown in Fig. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 443, + 149, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 149, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 149, + 455 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": "We conduct ablation studies on the validation set of Shop in the Cambridge Landmarks dataset to investigate the impact of all the components in our RAP. Setup I, our baseline, consists of the same components as in PoseNet [24] and has been retrained for our experiments. In Setup II, we replace the feature extraction from VGG16 [55] to Efficient-B0 [36], which enhances performance due to its superior feature representation, while they both exhibit poor performance due to the lack of data synthesis. In Setup III and IV, we explore the effectiveness of the designed pose augmentation and appearance augmentation, which bring notable improvements: translation error reduces from " + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "inline_equation", + "content": "103~\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "inline_equation", + "content": "75~\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": ", and rotation error from " + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "inline_equation", + "content": "3.52^{\\circ}" + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "inline_equation", + "content": "3.14^{\\circ}" + }, + { + "bbox": [ + 55, + 460, + 295, + 676 + ], + "type": "text", + "content": ". In Setup V and VI, we add regular convolutional layers and Pose Transformer between feature extraction and pose regression. Both improve performance due to the increasing parameters, but the Transformer achieves superior results by effectively handling long-term dependencies through attenuating the effect of noise." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 555, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 555, + 200 + ], + "type": "image", + "image_path": "5eff30fa95553b18efeb01e66b48e3fb59a73df1e300fb37fcd0488624d81f92.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "type": "text", + "content": "Figure 6. Visualization of the training set distribution and results on St. George's Basilica Building [49]. The red hollow spheres, centered on the real images in the training set, indicate the potential locations of all synthetic images during training." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 315 + ], + "type": "text", + "content": "tion mechanisms. Finally, in Setup VII, our adversarial discriminator effectively reduces the syn-to-real domain gap, allowing the model to learn better pose regression features from synthetic data and further reduce localization error." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 473, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 473, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 473, + 335 + ], + "type": "text", + "content": "4.4. Discussion on Data Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "content": "Emerging Generalizability. Previously, APR has been understood to implicitly learn image retrieval [49], lacking the ability to successfully interpolate between training samples and generalize beyond them. To investigate how APR training is affected by increasing synthetic data, we trained RAP on the St. George's Basilica Building [49] and visualized the results in Fig. 6. Here, the translation perturbation was set to " + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\delta t = 350" + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "content": " cm and the rotation perturbation to " + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\delta r = 60^{\\circ}" + }, + { + "bbox": [ + 313, + 342, + 554, + 485 + ], + "type": "text", + "content": ". Notably, the test set contained two regions entirely uncovered by the training set. Despite this, the model still closely predicts the test camera poses, demonstrating generalization ability beyond the original training positions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 617 + ], + "type": "text", + "content": "We also learn from our experiments that reducing the rotation perturbation, such that the overlap between test and training views remains minimal, leads to high localization error. This is because the translation and rotation parameter space is inherently a SE(3) manifold. Even if the translation remains fixed, significant rotation changes result in entirely different visual content in the images, naturally preventing the model from estimating poses of such unseen views, which correspond to a large distance on the SE(3) manifold. Therefore, enabling generalization across a broader range of space is an important direction for future work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Analyzing Generalization Boundaries. To evaluate the model's generalizability, we designed an experiment introducing a \"void zone\" centered on the test camera, where all real and synthetic data within this zone were excluded. The void zone was progressively expanded to determine the critical threshold at which the localization performance declines most significantly. Specifically, for Shop, we used " + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": " of the training set to ensure complete scene cov" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "DFNet [8] results are omitted as we were unable to successfully train its NeRF component, likely due to the need for manual scene scaling within " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "[- \\pi, \\pi]" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": ", which is tedious for diverse outdoor scenes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26854" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 298, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": "Table 6. Exploring the generalization boundaries of the model with synthetic data. Green, blue, and red percentages indicate the relative change in localization error (Med Err) compared to the scenario without a void zone." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 298, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 298, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 298, + 186 + ], + "type": "table", + "html": "
w/o Void Zonew/ Void Zone (cm/°)
10/0.520/130/1.550/280/2.5
Med Err ↓ (rel. change)33/1.2630/1.3432/1.3240/1.8439/2.0749/2.20
0%/0%-9%/6.3%-3%/4.7%21%/46.0%18%/64.3%48%/74.6%
Avg Err ↓41/1.5138/1.7541/1.6351/2.4048/2.4061/2.84
Max Err ↓155/4.52147/6.56219/6.01192/8.38246/9.80242/13.12
Min Err ↓3/0.178/0.204/0.227/0.304/0.168/0.60
", + "image_path": "9dcbd3d6adcc93c60cb9bcfd1089087c573a055c079cb1b0b346da363b0fb5dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 212, + 281, + 243 + ], + "blocks": [ + { + "bbox": [ + 64, + 197, + 286, + 208 + ], + "lines": [ + { + "bbox": [ + 64, + 197, + 286, + 208 + ], + "spans": [ + { + "bbox": [ + 64, + 197, + 286, + 208 + ], + "type": "text", + "content": "Table 7. Ablation on different pose augmentation policies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 212, + 281, + 243 + ], + "lines": [ + { + "bbox": [ + 70, + 212, + 281, + 243 + ], + "spans": [ + { + "bbox": [ + 70, + 212, + 281, + 243 + ], + "type": "table", + "html": "
MethodsCollegeHospitalShopChurchAverage
RAP (LENS [42])73/1.15126/1.8771/3.37128/3.50100/22.47
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
", + "image_path": "54cb0580225220446d27947504acf5844afdcd38e62421dc34ba3b14832f4747.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "text", + "content": "erage, with void zone ranges set as [10/0.5, 20/1, 30/1.5, 50/2, 80/2.5, 100/3] " + }, + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "inline_equation", + "content": "(\\mathrm{cm} / {}^{\\circ})" + }, + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "text", + "content": ". The results in Table 6 demonstrate a stepwise decline in performance. Initially, expanding the void zone has minimal impact on localization accuracy. However, at " + }, + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "inline_equation", + "content": "30~\\mathrm{cm} / 1.5^{\\circ}" + }, + { + "bbox": [ + 55, + 258, + 295, + 330 + ], + "type": "text", + "content": ", a sharp decrease in performance marks the model's generalization boundary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": "Pose Augmentation Policy. We conduct experiments using a modified version of RAP, with the pose augmentation approach identical to that of LENS [42], as shown in the Table 7, where our method obtains superior performance. This may be because, although LENS's pose augmentation policy covers a broader spatial area than the training set, its synthetic data may have lower NVS image quality in many unseen regions, which could negatively affect APR training. Density of Training Data. As shown in Table 8, our method with the proposed augmentation significantly reduces errors as the density of real training data increases from " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": ". However, the localization accuracy remains almost unchanged from " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": ", as the scene is already sufficiently covered. Notably, using " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": " of the training data without augmentation can result in a significantly higher maximum error in translation, nearly double that with only " + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 330, + 295, + 570 + ], + "type": "text", + "content": " of the training data with augmentation, despite its limited spatial coverage. This suggests that our augmentation method successfully prevents overfitting to the training data, improving generalization to the test set." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "Quality of Training Data. We evaluate the impact of synthetic image quality on model performance in Table 9, using " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " of the real data for pose regression. For Shop, it is evident that fewer training samples in 3DGS result in lower-quality rendered views (as indicated by lower PSNR), leading to suboptimal localization performance, particularly for rotation. Surprisingly, localization performance using only " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " of the data for training suboptimal 3DGS and pose regression surpasses the results obtained with " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " of the data without augmentation, as shown in Table 8. This experiment confirms the need for a robust NVS model and the proposed augmentation method in APR training." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 108, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 8. Impact of the density of real training data. Our augmentation improves the model's ability to generalize across the entire scene, although this effect has an upper limit." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 108, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 163 + ], + "type": "table", + "html": "
Training Pose %w/ Appearance & Pose Aug. (cm/°)w/o Aug. (cm/°)
100%80%60%40%20%100%50%
Med Err ↓33/1.2632/1.2737/1.9057/2.2387/3.6598/3.75104/4.17
Avg Err ↓41/1.5140/1.5048/2.1762/2.8191/4.65128/4.49139/5.33
Max Err ↓155/4.52158/4.09193/9.39230/11.06231/15.45490/20.73500/21.02
Min Err ↓3/0.177/0.207/0.186/0.3812/0.4613/0.639/0.48
", + "image_path": "eafafd4cd513f8ac6d8bc20943b9247540589b4778c4f8023e3d6934dea90b05.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 211, + 553, + 276 + ], + "blocks": [ + { + "bbox": [ + 313, + 174, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 208 + ], + "type": "text", + "content": "Table 9. Impact of synthetic image quality. Training with higher-quality synthetic images from advanced NVS models enhances localization performance." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 211, + 553, + 276 + ], + "lines": [ + { + "bbox": [ + 317, + 211, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 553, + 276 + ], + "type": "table", + "html": "
3DGS PerformanceLocalization Performance (cm/°)
% Images (Train)PSNR ↑ (Train)PSNR ↑ (Test)% Images (Train)Med Err ↓Avg Err ↓Max Err ↓Min Err ↓
20%29.0815.9820%58/3.5968/4.31211/12.1914/0.51
20%29.0815.9850%43/2.4755/3.26196/21.117/0.40
50%26.8817.5550%37/1.8848/2.37184/10.179/0.52
100%24.6018.3050%35/1.6441/2.12130/11.074/0.38
", + "image_path": "45e0359316464b9bb281168b90438162a72ea8ccffffd17c0e0c36956d5d68cb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 337, + 301, + 531, + 342 + ], + "blocks": [ + { + "bbox": [ + 335, + 286, + 533, + 297 + ], + "lines": [ + { + "bbox": [ + 335, + 286, + 533, + 297 + ], + "spans": [ + { + "bbox": [ + 335, + 286, + 533, + 297 + ], + "type": "text", + "content": "Table 10. Ablation on different 3D representations." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 301, + 531, + 342 + ], + "lines": [ + { + "bbox": [ + 337, + 301, + 531, + 342 + ], + "spans": [ + { + "bbox": [ + 337, + 301, + 531, + 342 + ], + "type": "table", + "html": "
MethodsCollegeHospitalShopChurchAverage
DFNet [8]73/2.37200/2.9867/2.21137/4.03119/2.90
DFNetGS102/2.31137/8.0877/3.92123/4.68110/4.75
RAP (Ours)52/0.9087/1.2133/1.4853/1.5256/1.28
", + "image_path": "8ce8de5eb20aa37b479c0eadb2d4cbb095246ba575b0ae668953b8ae2c8f858f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": "Different 3D Representations. We evaluate the impact of different 3D representations on the performance in Table 10. Trivially replacing NeRF with 3DGS in existing frameworks degrades performance due to 3DGS's inferior 3D consistency. This shows the effectiveness of our proposed joint training paradigm in RAP, which better utilizes diverse synthetic data to learn appearance-invariant features, rather than naively transferring from NeRF to 3DGS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 653 + ], + "type": "text", + "content": "Summary. We address absolute pose regression with a robust two-branch joint training framework based on Transformer, coupled with an efficient data synthesis pipeline leveraging 3D Gaussian Splats (3DGS) to synthesize numerous posed images with diverse appearances as additional supervision. Our RAP achieves state-of-the-art localization performance, even under challenging appearance variations. Moreover, we thoroughly investigate the impact of synthesizing diverse data and present a novel perspective on APR: generalizability can emerge if the learning gap in APR is effectively addressed together with diverse data. We believe our RAP could be a promising starting point, and the experiments presented in the paper can provide useful insights for future research in this field." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part through NSF grants 2238968, 2121391, and 2024882, and the NYU IT High Performance Computing resources, services, and staff expertise. Yiming Li is supported by NVIDIA Graduate Fellowship (2024-2025)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26855" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Joydeep Biswas and Manuela Veloso. Depth camera based indoor mobile robot localization and navigation. In 2012 IEEE International Conference on Robotics and Automation, pages 1697-1702, 2012. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Eric Brachmann and Carsten Rother. Visual camera relocalization from rgb and rgb-d images using dsac. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):5847-5865, 2021. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "text", + "content": "[3] Eric Brachmann, Alexander Krull, Sebastian Nowozin, Jamie Shotton, Frank Michel, Stefan Gumhold, and Carsten Rother. Dsac-differentiable ransac for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6684-6692, 2017. 1, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] Eric Brachmann, Tommaso Cavallari, and Victor Adrian Prisacariu. Accelerated coordinate encoding: Learning to relocalize in minutes using rgb and poses. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5044-5053, 2023. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 339 + ], + "type": "text", + "content": "[5] Samarth Brahmbhatt, Jinwei Gu, Kihwan Kim, James Hays, and Jan Kautz. Geometry-aware learning of maps for camera localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2616-2625, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 407 + ], + "type": "text", + "content": "[6] David M. Chen, Georges Baatz, Kevin Koser, Sam S. Tsai, Ramakrishna Vedantham, Timo Pylvanainen, Kimmo Roimela, Xin Chen, Jeff Bach, Marc Pollefeys, Bernd Girod, and Radek Grzesczuk. City-scale landmark identification on mobile devices. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 452 + ], + "type": "text", + "content": "[7] Shuai Chen, Zirui Wang, and Victor Prisacariu. Directposenet: Absolute pose regression with photometric consistency. In 2021 International Conference on 3D Vision (3DV), 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "text", + "content": "[8] Shuai Chen, Xinghui Li, Zirui Wang, and Victor A Prisacariu. Dfnet: Enhance absolute pose regression with direct feature matching. In European Conference on Computer Vision, pages 1-17. Springer, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 566 + ], + "type": "text", + "content": "[9] Shuai Chen, Yash Bhalgat, Xinghui Li, Jia-Wang Bian, Kejie Li, Zirui Wang, and Victor Adrian Prisacariu. Neural refinement for absolute pose regression with feature synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20987-20996, 2024. 1, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 621 + ], + "type": "text", + "content": "[10] Shuai Chen, Tommaso Cavallari, Victor Adrian Prisacariu, and Eric Brachmann. Map-relative pose regression for visual re-localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20665-20674, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[11] Manuela Chessa, Chiara Bassano, and Fabio Solari. Detection and localization of changes in immersive virtual reality. In International Conference on Image Analysis and Processing, pages 121-132, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Hiba Dahmani, Moussab Bennehar, Nathan Piasco, Luis Roldao, and Dzmitry Tsishkou. Swag: Splitting in the wild images with appearance-conditioned gausians. In European Conference on Computer Vision, pages 325-340, 2025. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 194 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint description and detection of local features. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8092-8101, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "text", + "content": "[16] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 298, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 351 + ], + "type": "text", + "content": "[17] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8):930–943, 2003. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "text", + "content": "[18] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 464 + ], + "type": "text", + "content": "[19] Hugo Germain, Daniel DeTone, Geoffrey Pascoe, Tanner Schmidt, David Novotny, Richard Newcombe, Chris Sweeney, Richard Szeliski, and Vasileios Balntas. Feature query networks: Neural surface description for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5071-5081, 2022. 1, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 466, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 509 + ], + "type": "text", + "content": "[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 553 + ], + "type": "text", + "content": "[21] Simon J. Julier and Jeffrey K. Uhlmann. New extension of the Kalman filter to nonlinear systems. In Signal Processing, Sensor Fusion, and Target Recognition VI, pages 182 - 193, 1997. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 599 + ], + "type": "text", + "content": "[22] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 644 + ], + "type": "text", + "content": "[23] Alex Kendall and Roberto Cipolla. Geometric loss functions for camera pose regression with deep learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5974-5983, 2017. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In IEEE international conference on computer vision, pages 2938-2946, 2015. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26856" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "radiance field rendering. ACM Transaction on Graphics, 42 (4), 2023. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 160 + ], + "type": "text", + "content": "[27] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. arXiv preprint arXiv:2406.09756, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[28] Yiming Li, Zhiheng Li, Nuo Chen, Moonjun Gong, Zonglin Lyu, Zehong Wang, Peili Jiang, and Chen Feng. Multiagent multitraversal multimodal self-driving: Open mars dataset. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22041-22051, 2024. 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[29] Yiming Li, Zehong Wang, Yue Wang, Zhiding Yu, Zan Gojcic, Marco Pavone, Chen Feng, and Jose M. Alvarez. Memorize what matters: Emergent scene decomposition from multitraverse. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[30] Jingyu Lin, Jiaqi Gu, Bojian Wu, Lubin Fan, Renjie Chen, Ligang Liu, and Jieping Ye. Learning neural volumetric pose features for camera localization. In European Conference on Computer Vision, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 380 + ], + "type": "text", + "content": "[31] Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, and Wenming Yang. Vastgaussian: Vast 3d gaussians for large scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5166-5175, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[32] Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. Lightglue: Local feature matching at light speed. In IEEE/CVF International Conference on Computer Vision, pages 17627-17638, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 491 + ], + "type": "text", + "content": "[33] Changkun Liu, Shuai Chen, Yukun Zhao, Huajian Huang, Victor Prisacariu, and Tristan Braud. Hr-apr: Apr-agnostic framework with uncertainty estimation and hierarchical refinement for camera relocalisation. In 2024 IEEE International Conference on Robotics and Automation, pages 8544–8550, 2024. 1, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "text", + "content": "[34] Changkun Liu, Shuai Chen, Yash Sanjay Bhalgat, Siyan HU, Ming Cheng, Zirui Wang, Victor Adrian Prisacariu, and Tristan Braud. GS-CPR: Efficient camera pose refinement via 3d gaussian splatting. In The Thirteenth International Conference on Learning Representations, 2025. 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[35] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[36] Liu Liu, Hongdong Li, and Yuchao Dai. Efficient global 2d-3d matching for camera localization in a large-scale 3d map. In IEEE International Conference on Computer Vision, pages 2372-2381, 2017. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[37] Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition, pages 2794-2802, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[39] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[41] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Coordinate: uncertainty-aware pose regressor for reliable vehicle localization. In IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2229-2238, 2022. 2, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[42] Arthur Moreau, Nathan Piasco, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Lens: Localization enhanced by nerf synthesis. In Conference on Robot Learning, pages 1347-1356, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 386 + ], + "type": "text", + "content": "[43] Arthur Moreau, Nathan Piasco, Moussab Bennehar, Dzmitry Tsishkou, Bogdan Stanciulescu, and Arnaud de La Fortelle. Crossfire: Camera relocalization on self-supervised features from an implicit representation. In IEEE/CVF International Conference on Computer Vision, pages 252-262, 2023. 1, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[44] Hyeonwoo Noh, Andre Araujo, Jack Sim, Tobias Weyand, and Bohyung Han. Large-scale image retrieval with attentive deep local features. In IEEE international conference on computer vision, pages 3456-3465, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12716–12725, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4938–4947, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 590 + ], + "type": "text", + "content": "[47] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient & effective prioritized matching for large-scale image-based localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(9):1744-1756, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 591, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 657 + ], + "type": "text", + "content": "[48] Torsten Sattler, Will Maddern, Carl Toft, Akihiko Torii, Lars Hammarstrand, Erik Stenberg, Daniel Safari, Masatoshi Okutomi, Marc Pollefeys, Josef Sivic, et al. Benchmarking 6 dof outdoor visual localization in changing conditions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8601-8610, 2018. 4, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[49] Torsten Sattler, Qunjie Zhou, Marc Pollefeys, and Laura Leal-Taixe. Understanding the limitations of cnn-based absolute camera pose regression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3302-3312, 2019. 1, 2, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26857" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4104-4113, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 152 + ], + "type": "text", + "content": "[51] Yoli Shavit and Yosi Keller. Camera pose auto-encoders for improving pose regression. In European Conference on Computer Vision, pages 140–157, 2022. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "text", + "content": "[52] Yoli Shavit, Ron Ferens, and Yoshi Keller. Paying attention to activation maps in camera pose regression. arXiv preprint arXiv:2103.11477, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 232 + ], + "type": "text", + "content": "[53] Yoli Shavit, Ron Ferens, and Yoshi Keller. Learning multiscene absolute pose regression with transformers. In IEEE/CVF International Conference on Computer Vision, pages 2733-2742, 2021. 1, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 235, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 235, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 294, + 289 + ], + "type": "text", + "content": "[54] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2930-2937, 2013. 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 324 + ], + "type": "text", + "content": "[55] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[56] Noah Snavely, Steven M Seitz, and Richard Szeliski. Photo tourism: exploring photo collections in 3d. In ACM Transactions on Graphics, pages 835-846, 2006. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 426 + ], + "type": "text", + "content": "[57] Hajime Taira, Masatoshi Okutomi, Torsten Sattler, Mircea Cimpoi, Marc Pollefeys, Josef Sivic, Tomas Pajdla, and Akihiko Torii. Inloc: Indoor visual localization with dense matching and view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7199-7209, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "text", + "content": "[58] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, pages 6105-6114, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 529 + ], + "type": "text", + "content": "[59] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The unreasonable effectiveness of pretrained features for camera pose refinement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12798, 2024. 1, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 575 + ], + "type": "text", + "content": "[60] Bing Wang, Changhao Chen, Chris Xiaoxuan Lu, Peijun Zhao, Niki Trigoni, and Andrew Markham. Atloc: Attention guided camera localization. In AAAI Conference on Artificial Intelligence, pages 10393-10401, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 632 + ], + "type": "text", + "content": "[61] Fangjinhua Wang, Xudong Jiang, Silvano Galliani, Christoph Vogel, and Marc Pollefeys. Glace: Global local accelerated coordinate encoding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21562-21571, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[62] Chen Wenbo and Liu Ligang. Deblur-gs: 3d gaussian splattering from camera motion blurred images. In ACM in Computer Graphics and Interactive Techniques, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[63] Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. Gaussian in the wild: 3d gaussian splatting for unconstrained image collections. In European Conference on Computer Vision, 2024. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 185 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[64] Boming Zhao, Luwei Yang, Mao Mao, Hujun Bao, and Zhaopeng Cui. Pnerfloc: Visual localization with point-based neural radiance fields. In AAAI Conference on Artificial Intelligence, pages 7450-7459, 2024. 1, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[65] Enliang Zheng and Changchang Wu. Structure from motion using structure-less resection. In IEEE International Conference on Computer Vision, pages 2075-2083, 2015. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[66] Qunjie Zhou, Maxim Maximov, Or Litany, and Laura Leal-Taixe. The nefrect match: Exploring nef features for visual localization. arXiv preprint arXiv:2403.09577, 2024. 1, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26858" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_content_list.json b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a114b9318e5b8e765dc254a17dc0fd036fbc814f --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "Adversarial Purification via Super-Resolution and Diffusion", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mincheol Park $^{1,2}$ , Cheonjun Park $^{5}$ , Seungseop Lim $^{4}$ , Mijin Koo $^{6}$ , Hyunwuk Lee $^{2}$ , Won Woo Ro $^{3}$ , Suhyun Kim $^{7*}$", + "bbox": [ + 225, + 179, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Samsung Advanced Institute of Technology, $^{2}$ Samsung Electronics, $^{3}$ Yonsei University, $^{4}$ AITRICS, $^{5}$ Hankuk University of Foreign Studies, $^{6}$ Seoul National University, $^{7}$ Kyung Hee University", + "bbox": [ + 104, + 215, + 893, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mincheol.park, cheonjun.park, hyunwuk.lee, wro}@yonsei.ac.kr, ss.lim@aitrics.com, starmj09@snu.ac.kr, dr.suhyun.kim@gmail.com", + "bbox": [ + 215, + 253, + 776, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks are widely used in various computer vision tasks, but their vulnerability to adversarial perturbations remains a significant challenge for reliable decision-making. Adversarial purification, a test-time defense strategy, has shown potential in countering these threats by removing noise through diffusion models. This plug-and-play method, using off-the-shelf models, appears highly effective. However, the purified data from diffusion often deviates more from the original data than the adversarial examples, leading to missing critical information and causing misclassification. In this study, we propose that upsampling with Super-Resolution (SR), followed by downsampling, can also aid in eliminating adversarial noise, similar to the noise addition and removal process in diffusion models. While SR alone is not as effective as the diffusion process, it better restores the original features typically associated with the early layers of networks. By combining SR, which initially mitigates damage to early-layer information from adversarial attacks, with diffusion, we observe a synergistic effect, leading to enhanced performance over diffusion models alone. Our comprehensive evaluations demonstrate that this combined approach, PuriFlow, significantly improves accuracy and robustness, working synergistically with state-of-the-art methods.", + "bbox": [ + 88, + 353, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 223, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Neural Networks (DNNs) have been widely accepted in various computer vision tasks. However, their susceptibility to small, human-imperceptible noises can lead to untrustworthy decision-making. In particular, the subtle, maliciously manipulated noises, known as adversarial perturbations, produce fatal images called adversarial examples [10]. This intentional threat appears across numerous", + "bbox": [ + 89, + 770, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdeb382978ed846143fa244c75867cea893c2e77b8031bfb21122b14ab67877d.jpg", + "image_caption": [ + "Figure 1. The overview of PuriFlow, which begins with restoring features of adversarial examples using the SR and drifting downsampled SR images within the forward diffusion. The perturbed images are then denoised in the reverse diffusion for classification." + ], + "image_footnote": [], + "bbox": [ + 516, + 323, + 903, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "types of attacks [3, 9, 13, 34], leaving many DNNs vulnerable and posing substantial challenges to their deployment in AI-powered systems.", + "bbox": [ + 511, + 672, + 906, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To counter the adversarial threats, adversarial purification [38, 58] has emerged as a test-time defense strategy. This approach is powered by leveraging off-the-shelf diffusion models. At its core, noise removal through stochastic denoising effectively eliminates artifacts introduced by various attacks. Unlike adversarial training and certified methods [27, 30, 34], which require prior knowledge of target attacks, this simple application of generative power excels in protecting classifier without the need for training, pointing to a new direction.", + "bbox": [ + 511, + 718, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While diffusion models appear effective in purification, a key problem remains widely discussed in studies [5, 11, 32,", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4605", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1be3d78e3114267a8b8b170e1df928f79d951efe6e2d9274d47c3b657811f079.jpg", + "image_caption": [ + "Figure 2. The example of purification on the clean images. The top row exhibits the ground truth images. The second and third rows depict purified images obtained using diffusion and our proposed PuriFlow. The fourth row visualizes Grad-CAM [44] as represented by ResNet-50. Compared to Diffpure, PuriFlow enhances visual fidelity and accuracy by preserving DNN's activation." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 486, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "38, 60]: how to lower the upper bound of differences between purified and original data, which can exceed those caused by adversarial attacks. This concern also arises when clean data is input, as shown in Figure 2. However, resolving it solely within the diffusion model is not straightforward. The noise predictor in off-the-shelf diffusion models is optimized to predict only the mean under isotropic covariance [26], limiting its ability to fully restore subtle features tied to true data covariates. To improve the restoration of details, diffusion models may require retraining with a proper objective [37], which may not align well with test-time defenses. Recently, guidance during noise removal has been proposed for test-time defenses [5]; however, this approach, despite allowing extensive processing time, is limited to large-scale datasets, impacting its generality.", + "bbox": [ + 89, + 625, + 483, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve purification without retraining, we note that adversarial attacks are designed to maximize changes in model output with minimal alterations to the input. This", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "results in alterations in the feature map that can be substantial, while changes in the image might be minor. Diffusion-based purification processes, which lack prior knowledge of the original image, risk deviating in the wrong direction during the noise addition and recovery phases, especially if the starting image contains feature information significantly distorted by adversarial noises. Therefore, it is crucial to eliminate as much distorted information as possible before applying the diffusion process.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We hypothesize that using Super-Resolution (SR) to upsample and then downsample an image could help reduce noise. Our tests confirm that while SR does reduce adversarial noise, it is not as effective as diffusion models. However, an in-depth experiment reveals that SR performs better than diffusion models in restoring features to the early layers of a network. This insight suggests that initially, using SR to reduce adversarial noise before applying diffusion could enhance overall performance by restoring the features the diffusion process alone cannot recover.", + "bbox": [ + 511, + 227, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper proposes PuriFlow, an enhanced purification flow that leverages the synergy of SR and the diffusion process. As shown in Figure 1, PuriFlow integrates SR before the forward diffusion process to restore distorted features of adversarial images early on. By boosting content similarity [28], SR generates enhanced samples with features closer to the originals, even when mapped into a high-dimensional space. Our findings show that content proximity increases when SR images are resized back to their original dimensions, positioning resized SR images in regions with a higher probability for correct labels. In other words, resizing SR effectively reduces the initial cross-entropy on adversarial examples, allowing better diffusion. Our theoretical analysis and comprehensive evaluations demonstrate that this initialization aids in restoring features by diffusion, further reduces cross-entropy, and improves accuracy.", + "bbox": [ + 511, + 378, + 906, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as three-fold:", + "bbox": [ + 532, + 619, + 872, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We find the purification effectiveness of the SR process. While SR alone is less effective in fully eliminating adversarial noises, it excels at preserving original features in the early layers of the network.", + "- Our approach, PuriFlow, finds a synergistic effect by integrating SR with diffusion. Our insight into geometrical analysis and empirical study finds that this integration can reduce cross-entropy in adversarial examples.", + "- Our pipeline utilizes only off-the-shelf SR and diffusion models. With negligible execution time required for SR, this plug-and-play method demonstrates improved accuracy and robustness in comprehensive evaluations." + ], + "bbox": [ + 513, + 635, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 829, + 653, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising defenses. Traditional test-time adversarial defenses often apply denoising techniques to input images. JPEG compression has been employed to remove high", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4606", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frequency noise [14, 16], but Guo et al. [22] found it insufficient, proposing total variance minimization with image quilting as an alternative. Likewise, Prakash et al. [40] use wavelet denoising combined with BayesShrink to mitigate artifacts leading to misclassification, which surpasses total variance minimization and Wiener filtering in adversarial defenses [54]. Xie et al. [55] denoise feature maps through non-local means, supporting adversarial training. While practical, these methods often distort image details and face a core challenge: the inverse problem, which results in non-unique solutions.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SR defenses. Modern SR models have gained attention to enhance perceptual quality in adversarial examples [36]. Recently, Bhardwaj et al. [7] raised concerns about the computational demands of the SR models, which can limit real-time applicability. They demonstrate the effectiveness of a highly efficient SR model [8], showing robust performance even on compact neural processing units. PuriFlow shares a similar purpose on rapid SR applications but adds weight to its synergistic use for restoring refined features with minimal overhead, assisting diffusion to enhance defense.", + "bbox": [ + 89, + 268, + 483, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial purification. Purifying adversarial noise using generative models has shown promise in strengthening classifier robustness as a test-time defense. PixelDefend [47] utilizes an autoregressive generative model, while Defense-GAN [43] relies on GANs. Langevin Dynamics (LD) sampling has also been effective for defenses through energy-based models [15, 21, 25]. Yoon et al. [58] introduce a denoising score-based model with an LD variant for denoising processes. DiffPure [38] and Lee and Kim [32] employ forward and reverse-time Stochastic Differential Equations (SDEs) [26, 46] for purification, but the loss of class information remains a risk due to the dependence on diffusion time steps. Bai et al. [5] mitigate this problem with contrastive guidance, while ScoreOpt [60] uses a one-shot denoiser. This work attempts to tackle losing crucial information in purified images by combining SR with diffusion.", + "bbox": [ + 89, + 430, + 483, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background for Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 684, + 397, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The score-based diffusion model is widely accepted in adversarial purification. It performs with bidirectional processes: forward and reverse time transition of $\\mathbf{x}_t\\in \\mathbb{R}^d$ over the time interval $t\\in [0,T]$ . Starting from an original sample $\\mathbf{x}_0\\sim p_0(\\mathbf{x})$ where $p_0(\\mathbf{x})$ is the unknown, true data distribution, the diffusion model progressively transforms $p_0(\\mathbf{x})$ into a nearly spherical Gaussian distribution $p_T(\\mathbf{x})\\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ . This converted distribution $p_T(\\mathbf{x})$ retains no information about $p_0(\\mathbf{x})$ .", + "bbox": [ + 89, + 709, + 482, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This forward process of $\\{\\mathbf{x}_t\\}_{t = 0}^T$ , as defined by Ito SDE, is given with the positive increment of $t$ ..", + "bbox": [ + 89, + 844, + 482, + 875 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} = f (\\mathbf {x}, t) d t + g (t) d \\mathbf {w}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 885, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the initial data $\\mathbf{x}_0\\sim p_0(\\mathbf{x}),f(\\cdot ,t):\\mathbb{R}^d\\mapsto \\mathbb{R}^d$ is the drift coefficient of $\\mathbf{x}_t$ and $g(\\cdot):\\mathbb{R}\\mapsto \\mathbb{R}$ is the diffusion coefficient tied with Brownian motion $\\mathbf{w}\\in \\mathbb{R}^d$ . Here, when $f(\\mathbf{x},t)$ is affine, the transition kernel always becomes a Gaussian distribution $p_{t^{\\prime}t}(\\mathbf{x}_t|\\mathbf{x}_{t^{\\prime}})$ where $0\\leq t^{\\prime} < t\\leq T$ [49]. Thus, the forward diffusion process allows for a direct transition with a closed form using the Gaussian kernel $p_{0t}(\\mathbf{x}_t|\\mathbf{x}_0)$ from $t = 0$ to a certain $t$ , which prevents the need for neural network estimations.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a perturbed sample $\\mathbf{x}_T\\sim p_T(\\mathbf{x})$ , the reverse-time SDE defines that it drifts backward in the time steps:", + "bbox": [ + 511, + 227, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] d t + g (t) d \\bar {\\mathbf {w}}, \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 268, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\mathbf{w}}$ represents a standard Wiener process and the infinitesimal time step $dt$ is negative. A key component in Eq. 2 is the score function $\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})$ , which offers how the density $p_t(\\mathbf{x})$ changes with respect to $\\mathbf{x}$ . Instead of directly knowing this function, one approach is to train a neural network, $\\mathbf{s}_{\\theta}(\\mathbf{x},t)$ , to estimate it. This training is achieved by solving a denoising score-matching [51] problem:", + "bbox": [ + 511, + 297, + 906, + 404 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\ell (\\theta) = \\sum_ {t = 1} ^ {T} \\lambda (t) \\mathbb {E} _ {\\mathbf {x}, \\tilde {\\mathbf {x}} | \\mathbf {x}} [ \\mathbf {s} _ {\\theta} (\\tilde {\\mathbf {x}}, t) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {0 t} (\\tilde {\\mathbf {x}} | \\mathbf {x}) | | _ {2} ^ {2} ]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 412, + 903, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, Gaussian kernel $p_{0t}(\\tilde{\\mathbf{x}}|\\mathbf{x})$ depicts the transition from $\\tilde{\\mathbf{x}} = \\mathbf{x}_t$ to $\\mathbf{x} = \\mathbf{x}_0$ and $\\lambda(t)$ is weighting function ensuring that the contributions at different times are appropriately balanced. This optimized model serves as $\\mathbf{s}_{\\theta^*}(\\mathbf{x}, t)$ where $\\theta^* = \\mathrm{argmin}_{\\theta} \\ell(\\theta)$ to replace the elusive score function, enabling an effective reverse diffusion process.", + "bbox": [ + 511, + 453, + 905, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 556, + 648, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our motivation, design methodology, and geometrical analysis of the proposed approach, PuriFlow. To begin, we briefly review adversarial attacks.", + "bbox": [ + 511, + 582, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial attack. Given a clean image $\\mathbf{x}_0$ with label $y$ , adversarial attacks generally seek perturbations to mislead a classifier $\\mathcal{F}_{\\phi}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{K}$ into incorrectly classifying $\\mathbf{x}_0$ . This manipulated image $\\mathbf{x}_0^a$ is generated by maximizing cross-entropy loss within a tiny region around $\\mathbf{x}_0$ :", + "bbox": [ + 511, + 638, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {0} ^ {a} = \\lim _ {i \\rightarrow N} \\prod_ {\\mathcal {B} \\left(\\mathbf {x} _ {0} ^ {i}, \\epsilon\\right)} \\left(\\mathbf {x} _ {0} ^ {i} + \\mu \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {0} ^ {i}} \\ell_ {c e} (\\phi ; \\mathbf {x} _ {0} ^ {i}, y)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 726, + 906, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_0^0$ is the initial clean image $\\mathbf{x}_0$ and $N$ is the number of attack iterations. $\\mathcal{B}(\\mathbf{x}_0^i,\\epsilon)$ is the $\\ell_p$ -norm ball around $\\mathbf{x}_0^i$ within a radius of $\\epsilon$ , $\\Pi (\\cdot)$ is the projection to the norm ball, and $\\mu$ denotes the step size. The gradient can be obtained from surrogate networks when $\\mathcal{F}_{\\phi}(\\cdot)$ is unknown.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. PuriFlow: Enhanced Purification Flow", + "text_level": 1, + "bbox": [ + 511, + 848, + 848, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an adversarial example, $\\mathbf{x}_0^a$ , VP-SDE [48] is widely accepted approach for purification. VP-SDE controls noise", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4607", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/064cdc7a169202867996906cef60368462d4c9ad82f0b919aae52962c1658932.jpg", + "image_caption": [ + "Figure 3. Top: visualization of different image types. AE denotes the adversarial example generated by a PGD attack on ResNet-50 using an ImageNet-1k image. SR indicates the downsampled high-resolution image. Diffusion and PuriFlow use the same diffusion method, VP-SDE [38]. Bottom: corresponding feature maps from the Conv2_2 layer of VGG-19, as described in [18]." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 228 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "addition through drift coefficients $f(\\mathbf{x}, t) = -\\frac{1}{2} \\beta_t \\mathbf{x}$ and $g(t) = \\sqrt{\\beta_t}$ , where $\\beta_t = \\frac{1}{t} \\beta_{\\text{max}} + (1 - \\frac{1}{t}) \\beta_{\\text{min}}$ [38]. This model then diffuses the data by solving the forward-time SDE in Eq. 1 with a closed form:", + "bbox": [ + 89, + 349, + 483, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} ^ {a} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} ^ {a} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 422, + 482, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)$ and $\\bar{\\alpha}_{t} = e^{-\\int_{0}^{t}\\beta_{t}(s)ds}$ .", + "bbox": [ + 89, + 450, + 387, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The corresponding diffusion process is then performed by solving the reverse-time SDE in Eq. 2:", + "bbox": [ + 89, + 467, + 482, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd \\mathbf {x} = \\left[ f \\left(\\mathbf {x} _ {t} ^ {a}, t\\right) - g ^ {2} (t) \\nabla_ {\\mathbf {x} _ {t} ^ {a}} \\log p _ {t} \\left(\\mathbf {x} _ {t} ^ {a}\\right) \\right] d t + g (t) d \\bar {\\mathbf {w}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 508, + 480, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the drift coefficients, $f(\\cdot ,t)$ and $g^{2}(t)$ , follows [5, 48].", + "bbox": [ + 89, + 537, + 482, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During this reverse process, the crucial step is computing the score, defined as $\\nabla_{\\mathbf{x}_t^a}\\log p_t(\\mathbf{x}_t^a)$ . Here, we raise a key question: Can the neural network $s_{\\theta^*}(\\cdot ,t)$ , trained for score estimation, precisely predict $\\mathbf{x}_0$ from $\\mathbf{x}_t^a$ ? Even if $s_{\\theta^*}(\\cdot ,t)$ is well-trained, accurately estimating the direction toward $\\mathbf{x}_0$ is not straightforward when dealing with the noised data $\\mathbf{x}_t^a$ that has been perturbed from $\\mathbf{x}_0^a$ , since $\\mathbf{s}_{\\theta^*}(\\cdot ,t)$ has not been exposed to such perturbed data during training with Eq. 3. In other words, the trained neural network lacks prior knowledge of guiding $\\mathbf{x}_t^a$ , risking deviations in the wrong direction during the noise addition and recovery phases.", + "bbox": [ + 88, + 553, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This concern is seemingly explicit. Figure 3 shows that the perturbations introduced by Eq. 4 appear subtle in pixel values but cause significant damage to the underlying features of $\\mathbf{x}_0$ . These observations indicate that diffusion alone struggles to fully reconstruct the original features. To address this, we seek an existing method that positions the perturbed data closer to its ground-truth image, aiming to reduce feature discrepancies early on and thus ensure a more precise projection of the score function.", + "bbox": [ + 88, + 719, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we focus on a model that can effectively reduce content distance [18, 28] to capture feature proximity. Reducing feature-level shifts that lead to increased cross", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/08191acf1a96c1871cb1a916d5fd07bd4be85225b0cfd30317837edd1c0d88e4.jpg", + "image_caption": [ + "(a) Impact of SR compared to diffusion with increments of $t^\\prime$" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 193 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98fac25418ba935d5814b383779760961859ec93d8a6b76311131600a9ae5c5a.jpg", + "image_caption": [ + "(b) Synergistic effect in feature restoration of SR combined with diffusion. Diffusion and SR+Diffusion use the same diffusion time $t^{\\prime} = 90$", + "Figure 4. Normalized content distance from the ground truth for each image type, measured across specific CONV layers of VGG-19 as shown in Figure 5. Real values are provided in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 517, + 214, + 903, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entropy aligns with the goal of purification, ultimately improving classification accuracy. This approach also benefits clean data in preserving features. In this context, the SR model suits this purpose. The SR model, $\\mathcal{H}_{\\omega^{*}}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{D}$ such that $D > d$ , is designed to upsample and interpolate low-resolution images into high-resolution counterparts. To optimize its parameters $\\omega$ , various SR methods [1, 28, 31, 45, 52] minimize both pixel distance and VGG loss [31], which captures mutual similarity in feature maps across different layers of the VGG model:", + "bbox": [ + 511, + 426, + 906, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega^ {*} = \\min _ {\\omega} \\frac {1}{W ^ {l} H ^ {l}} \\left(\\sum_ {u = 1} ^ {W ^ {l}} \\sum_ {v = 1} ^ {H ^ {l}} \\left[ \\Phi^ {l} (\\mathbf {X}) _ {u, v} - \\Phi^ {l} \\left(\\mathcal {H} _ {\\omega} (\\mathbf {x})\\right) _ {u, v} \\right] ^ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 588, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X} \\in \\mathbb{R}^{D}$ and $\\mathbf{x} \\in \\mathbb{R}^{d}$ , with $W^{l}$ and $H^{l}$ denoting the dimensions of extracted features $\\Phi^l (\\cdot)$ in the $l$ -th convolution layer. Thus, SR models can naturally reduce the content distance by enhancing feature similarity while estimating unknown pixels and their covariates in $\\mathbb{R}^{D}$ .", + "bbox": [ + 511, + 643, + 905, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our findings, as illustrated in Figure 4, show that SR images downsampled from $\\mathbb{R}^D$ to $\\mathbb{R}^d$ can also reduce the content distance. This reduction covers all layers of the VGG network, demonstrating that even a one-shot application can effectively mitigate feature discrepancies present in adversarial examples. Notably, as observed in Figure 4a, increasing $t'$ when applying diffusion alone cannot sufficiently aid in restoring features tied to the early layers of the CNN. This observation suggests that a single application of SR can be more effective than diffusion in recovering detailed features [59] captured by the early layers.", + "bbox": [ + 511, + 719, + 906, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we propose an enhanced purification flow,", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4608", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65d3259a8c6b94a5d010d531599c38fa0d74012a26969bd13341c5140fbfe2cb.jpg", + "image_caption": [ + "Figure 5. Transition of diffusion only and SR-integrated diffusion with the same diffusion time of $t' = 90$ under a PGD attack ( $\\epsilon = 8/255$ , $N = 100$ ) on ResNet-50 using ImageNet-1k. \"SR\" represents downsampled high-dimensional images, while \"AE\" denotes adversarial examples. Content distance [18] from the ground truth is measured in two convolutional layers of VGG-19 and plotted with cross-entropy." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 295, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5ed4042115827e922264d25cff9871e1ccf8d7e09d9ab966ba79a7deccae5439.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 88, + 496, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/05214416f75d2afea1bf3104c9aa99c2fc564a11fca8b0576facfb1b8709b773.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 700, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc21e9c162230ce689e6692bc018e2ffd9a7ce85c7a50a0f8bd5a2022731c7cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 88, + 901, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PuriFlow, where SR works before the start of the forward-time SDE. As shown in Figure 4b, this approach demonstrates the effectiveness of using SR early before solving both SDEs. By correcting feature distortions that diffusion alone struggles to recover, SR allows for a synergistic effect, enhancing feature restoration across all layers through diffusion. As observed in Figure 5, the early reduction in content distance leads to a notable decrease in cross-entropy. Thus, the subsequent diffusion aligns the data more closely with the ground truth. To delve into this synergistic effect, we provide a theoretical study from a manifold perspective. We examine how SR influences data positioning and allows for more precise use of the score function within diffusion.", + "bbox": [ + 88, + 303, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Geometrical Analysis on PuriFlow", + "text_level": 1, + "bbox": [ + 89, + 508, + 392, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose $\\mathcal{M} \\subset \\mathbb{R}^d$ is the set of all data points, i.e., $\\forall \\mathbf{x} \\in \\mathcal{M}$ , defining $\\mathcal{M}$ as the data manifold. That is, the data distribution $p_0$ is uniform on the data manifold $\\mathcal{M}$ . Then, within a local neighborhood, the manifold coincides with its tangent space of dimension $k \\ll d$ , expressed as follows:", + "bbox": [ + 89, + 530, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right) = T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right), \\quad T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\simeq \\mathbb {R} ^ {k}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 617, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{\\mathbf{x}_0}\\mathcal{M}$ denotes the tangent space to the manifold $\\mathcal{M}$ at $\\mathbf{x}_0$ and $dr$ is an infinitesimal radius within which the manifold can be locally approximated by its tangent space. Here, we take note of a property of the trained score function in diffusion models [12].", + "bbox": [ + 89, + 643, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1. The noisy data become increasingly concentrated on spherical manifolds as diffusion time $t$ increases. The score function is trained via denoising score matching (Eq. 3), using data on $\\mathcal{M}$ . Let $\\mathcal{P}_t(\\cdot)$ denote a mapping from noisy data $\\pmb{x}_t$ to its estimated clean counterpart $\\hat{\\pmb{x}}_0$ :", + "bbox": [ + 89, + 729, + 483, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {t} (\\cdot): \\boldsymbol {x} _ {t} \\mapsto \\hat {\\boldsymbol {x}} _ {0} = - \\frac {1}{2} (1 - \\alpha_ {t}) \\left(\\boldsymbol {x} _ {t} + 2 \\boldsymbol {s} _ {\\theta^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 815, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{P}_t(\\pmb{x}_t) \\in \\mathcal{M}$ and it satisfies $\\pmb{J}_{\\mathcal{P}_t}^2 = \\pmb{J}_{\\mathcal{P}_t} = \\pmb{J}_{\\mathcal{P}_t}^T : \\mathbb{R}^d \\to T_{\\mathcal{P}_t(\\pmb{x}_t)}\\mathcal{M}$ . That is, $\\mathcal{P}_t(\\cdot)$ acts as an orthonormal projection onto $\\mathcal{M}$ .", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2. The score function is trained solely on the data concentrated on both the noisy and data manifolds. Therefore, applying it to data outside these manifolds may lead to inaccurate inference.", + "bbox": [ + 511, + 303, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both premises suggest that when the score function is applied to off-manifold data, such as adversarial examples [7], its inaccurate inference may lead to misaligned orthonormal projections onto an unintended tangent space. Thus, the off-manifold data should be corrected to better align with its underlying manifold. Now, we show that the applying SR followed by downsampling can typically position off-manifold data, degraded from $\\mathbf{x}_0$ , closer to the local manifold, which is the tangent space of $\\mathbf{x}_0$ .", + "bbox": [ + 511, + 369, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 1. Suppose $\\mathcal{H}_{\\omega^*}$ is an ideal SR and smooth continuity, and let $\\mathcal{R}$ be a linear downsampler. Given an offmanifold data $\\mathfrak{D}(\\pmb{x}_0)$ degraded from $\\pmb{x}_0$ , the processed data $\\mathcal{R}\\mathcal{H}_{\\omega^*}(\\mathfrak{D}(\\pmb{x}_0))$ is projected near the tangent space $T_{\\pmb{x}_0}\\mathcal{M}$ of the data manifold $\\mathcal{M}$ at $\\pmb{x}_0$ with a tiny $\\delta > 0$ :", + "bbox": [ + 511, + 510, + 905, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\mathcal {R H} _ {\\omega^ {*}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) - \\Pi_ {T _ {x _ {0}} \\mathcal {M}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right\\| _ {F} < \\delta . \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 592, + 903, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Section 5, we evaluate the effectiveness of using the score function more accurately, supported by Proposition 1, against various adversarial attacks.", + "bbox": [ + 511, + 614, + 905, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 672, + 638, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of PuriFlow in two configurations: integrating SR with solving SDEs [38] and combining SR with a recent alternative, iterative One-Shot Denoising (OSD) [60]. OSD is defined by Tweedie's formula [50], expressed as $\\mathbb{E}[\\mathbf{x}|\\tilde{\\mathbf{x}}] = \\tilde{\\mathbf{x}} +\\sigma^2\\nabla_{\\tilde{\\mathbf{x}}}\\log p_\\sigma (\\tilde{\\mathbf{x}})$ , where $\\tilde{\\mathbf{x}}$ denotes perturbed data and $p_{\\sigma}(\\tilde{\\mathbf{x}})$ represents marginal Gaussian perturbation of $p(\\mathbf{x})$ . This formulation avoids repeated noise removal steps for varying time $t$ .", + "bbox": [ + 511, + 696, + 905, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this context, the denoising score-matching problem defined by Eq. 3 optimizes the neural network $\\mathbf{s}_{\\theta^{*}}(\\cdot ,t)$ given $\\mathbf{x}_t$ to predict a fully denoised $\\mathbf{x}_0$ per a specific time $t$ . Thus, Tweedie's formula can be represented as:", + "bbox": [ + 511, + 818, + 905, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} \\left[ \\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t ^ {\\prime}} \\right] = \\mathbf {x} _ {t ^ {\\prime}} + \\sigma^ {2} \\nabla_ {\\mathbf {x} _ {t ^ {\\prime}}} \\log p _ {\\sigma} \\left(\\mathbf {x} _ {t ^ {\\prime}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 883, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4609", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57072df7ee4890fea5cf1d1aa6595111b85fa4515f82fdbc6a98bc6a7556b4ee.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparison of different adversarial purification methods against BPDA+EOT attacks with $\\ell_{\\infty}$ perturbations. Our evaluations use WRN-28-10 on CIFAR-10, with consistent experimental settings as outlined in [38, 60] for $\\epsilon = 8 / 255$ . (* This purification utilizes a variant of the LD sampling.)" + ], + "table_body": "
MethodCriterionStandard (%)Robust (%)
Yang et al. [56]Gibbs Update94.8040.80
Song et al. [47]Mask+Recon.95.009.00
Hill et al. [24]EBM+LD84.1254.90
Yoon et al. [58]DSM+LD*86.1470.01
Nie et al. [38]SDEs89.0281.40
Bai et al. [5]Guide+SDEs92.6181.94
PuriFlow([38])SR+SDEs90.0683.00
Zhang et al. [60]iOSD93.7588.08
PuriFlow([60])SR+iOSD94.1491.01
", + "bbox": [ + 91, + 88, + 483, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The alternative proposed by [60] involves iteratively applying OSD using Eq. 11 during noise removal. Therefore, it can serve as a suitable baseline to confirm synergies with SR. We assess these configurations, detailed as algorithms in the Appendix, across key test-time defense scenarios, including adaptive white-box, preprocessor-blind [58], and black-box attacks. This section also includes an ablation study to demonstrate the impact of incorporating SR. Detailed experimental settings and additional evaluations for certified robustness are provided in the Appendix. For adaptive attacks, we adopt MDSR [33] for efficiency and ESRGAN [52] otherwise. We denote $t'$ on a discrete scale [26] from 1 to 1000 to represent the continuous interval [0, 1].", + "bbox": [ + 88, + 338, + 485, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Defense on Adaptive Attacks", + "text_level": 1, + "bbox": [ + 89, + 545, + 349, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of PuriFlow against four strong adaptive attacks: BPDA+EOT [3, 4], an adaptive ensemble white-box attack; AutoAttack [13], an adaptive white-box attack; PGD+EOT [32], and DiffAttack [29], a diffusion-based purification targeted attack. Here, the evaluations focus on projections within an $\\ell_{\\infty}$ -norm ball, as the projection in $\\ell_{2}$ -norm ball is less effective.", + "bbox": [ + 89, + 566, + 483, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BPDA+EOT. Backward Pass Differentiable Approximation (BPDA) combined with Expectation over Transformation (EOT) is used to evaluate randomized adversarial purification methods. Table 1 demonstrates that PuriFlow outperforms standalone diffusion on CIFAR-10 with WRN-28-10, achieving a more balanced standard and robust accuracy than the guiding method. In addition, integrating SR with iterative OSD further enhances performance, emphasizing its critical role in diffusion-based purification.", + "bbox": [ + 89, + 681, + 483, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "AutoAttack. Tables 2 demonstrates the effectiveness of PuriFlow against AutoAttack, on ImageNet-1k. PuriFlow outperforms adversarial training methods and a diffusion-only approach across ResNet-50 and DeiT-S. Compared to diffusion only, it achieves notable robust accuracy improve", + "bbox": [ + 89, + 824, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4c45244a0181a31891b6bc2abdba9e560cb046d0f2677f40bb2153bc9c8606ba.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Evaluation comparing adversarial training (AT) and purification (AP) methods against AutoAttack with $\\ell_{\\infty}$ perturbations at $\\epsilon = 4 / 255$ on ImageNet-1k. Diffusion time is set to $t' = 150$ for [38] as specified its study and $t' = 130$ for PuriFlow." + ], + "table_body": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5631.06
Wong et al. [53]55.6226.95
Salman et al. [42]64.0237.89
Bai et al. [6]67.3835.51
APNie et al. [38]67.7940.93
PuriFlow([38])71.6852.92
DeiT-S79.900.00
ATBai et al. [6]66.5035.50
APNie et al. [38]73.6343.18
PuriFlow([38])75.6953.12
", + "bbox": [ + 526, + 89, + 895, + 276 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02ff93cb6170e8349a78626c9814099957d5d03a64fd657431574092922d2f4e.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Evaluation comparing adversarial training (AT) and purification (AP) methods against PGD+EOT within an $\\ell_{\\infty}$ -norm ball of radius $\\epsilon = 8/255$ on CIFAR-10. Results are sourced from [32]. (* For a fair comparison, PuriFlow and [60] are directly evaluated on the same settings with $t' = 250$ , as stated in [60].)" + ], + "table_body": "
TypeMethodStandard (%)Robust (%)
WRN-28-1095.630.00
ATPang et al. [39]88.6264.95
Gowal et al. [19]88.5465.93
Gowal et al. [20]87.5166.01
APYoon et al. [58]85.6633.48
Zhang et al. [60]*93.7559.37
PuriFlow([60])*94.1460.76
WRN-70-1695.790.00
ATGowal et al. [19]91.1068.66
Gowal et al. [20]88.7469.03
Rebuffi et al. [41]92.2269.97
APYoon et al. [58]86.7637.11
Zhang et al. [60]*95.1163.28
PuriFlow([60])*95.3164.86
", + "bbox": [ + 526, + 357, + 895, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ments, including $11.99\\%$ (ResNet-50) and $9.94\\%$ (DeiT-S), and standard accuracy. PuriFlow also shows efficiency with reduced diffusion time $t'$ on this large-scale dataset.", + "bbox": [ + 511, + 700, + 906, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{PGD} + \\mathbf{EOT}$ . Motivated by [32], we assess PuriFlow's robustness against $\\mathrm{PGD} + \\mathrm{EOT}$ that utilizes gradient signs within the entire randomized defense framework. Table 3 shows PuriFlow's strength on CIFAR-10, achieving the highest standard accuracy for WRN-28-10 and WRN-70-16. Despite the substantial impact of $\\mathrm{PGD} + \\mathrm{EOT}$ on randomized purification, PuriFlow achieves superior robust accuracy among purification methods, demonstrating a synergistic effect in which SR effectively acts as the initialization before starting the iterative one-shot denoising process.", + "bbox": [ + 511, + 757, + 908, + 910 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4610", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f5288071aec4dec5b79d89a0053f33a337958abe2f2b834ad26e0ee3578c1b85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodStandard (%)Robust (%)
WRN-70-1695.790.00
APNie et al. [38]90.0745.31
PuriFlow([38])90.2159.47
ResNet-5076.130.00
APNie et al. [38]67.7928.13
PuriFlow([38])71.6843.75
", + "bbox": [ + 106, + 89, + 470, + 210 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4416ba052962728e8edc7ebf679c701be74d53fbb4b8e21ad33fc613051fa9ad.jpg", + "table_caption": [ + "Table 4. Evaluation against DiffAttack, targeting WRN-70-16 on CIFAR-10 and ResNet-50 on ImageNet-1k. Each diffusion time $t'$ is the same as those used for other attacks on PuriFlow." + ], + "table_footnote": [], + "table_body": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5638.97
Wong et al. [53]55.6229.15
Salman et al. [42]64.0238.01
Bai et al. [6]67.3840.27
APNie et al. [38]74.5238.87
PuriFlow([38])74.0549.92
DeiT-S79.900.00
ATBai et al. [6]66.5040.32
APNie et al. [38]78.9434.59
PuriFlow([38])78.4449.85
", + "bbox": [ + 102, + 273, + 470, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DiffAttack. DiffAttack neutralizes vanishing and exploding gradients, high memory costs, and increased randomness—natural defensive effects from lengthy $t'$ of diffusion-based purification. Table 4 demonstrates that while purification methods using $t' = 100$ for CIFAR-10 and $t' = 150$ for ImageNet-1k exhibit limitations in robustness by this intentional attack, integrating SR demonstrates a notable improvement by $14.16\\%$ and $15.62\\%$ . We hypothesize that SR's early enhancement of feature proximity helps mitigate diverging feature maps from the ground truth, even under the reconstruction-deviating impacts of DiffAttack.", + "bbox": [ + 88, + 536, + 482, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Defense on Preprocessor-Blind Attack", + "text_level": 1, + "bbox": [ + 89, + 712, + 419, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate PuriFlow against the PGD ( $N = 100$ ) attack, targeting only the classifier on ImageNet-1k. Table 5 shows that PuriFlow, leveraging only off-the-shelf models, surpasses adversarial training methods in both standard and robust accuracy for ResNet-50 and DeiT-S. Notably, PuriFlow significantly outperforms diffusion only [38] at the same practical diffusion time $t' = 10$ , achieving comparable standard accuracy while enhancing robust accuracy by $11.05\\%$ for ResNet-50 and $15.26\\%$ for DeiT-S. This improvement signifies the synergistic importance of SR in adversarial purification utilizing diffusion models.", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bf280e4b7f6d1224f20ce1732dd352df65c54020dcbb828a49f61080312d896.jpg", + "table_caption": [ + "Table 5. Comparison under a preprocessor-blind PGD attack targeting only the classifier, within an $\\ell_{\\infty}$ perturbations at $\\epsilon = 4/255$ on ImageNet-1k. The time $t'$ for both AP methods is set as 10." + ], + "table_footnote": [], + "table_body": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.139.25
APNie et al. [38]67.7962.88
PuriFlow [38]74.0068.36
", + "bbox": [ + 527, + 89, + 893, + 161 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f72e8892f71090e296c21b49e47526444b950798826a917ffa454312b9ee7316.jpg", + "table_caption": [ + "Table 6. Evaluation of ResNet-50 under a black-box attack using Square Attack ( $N = 5000$ ) within an $\\ell_{\\infty}$ -norm ball of radius $\\epsilon = 4/255$ on ImageNet-1k. Diffusion time is set to $t' = 150$ for [38] as specified in their study and $t' = 10$ for PuriFlow. Evaluation settings follow those in [38]." + ], + "table_footnote": [], + "table_body": "
MethodTypeStandard (%)Robust (%)
ResNet-5076.130.00
WaveletDenoising71.1568.63
TVMDenoising57.8056.96
NL-meansDenoising64.3262.67
DRLN×2 [2]SR71.6668.72
EDSR×2 [33]SR71.7768.73
ESRGAN×2 [52]SR72.9768.75
MDSR×2 [33]SR73.3269.47
", + "bbox": [ + 514, + 260, + 906, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Comparison of SR with other denoising methods as an initialization approach for diffusion [38] ( $t' = 90$ ) under a PGD attack on ImageNet-1k. \"TVM\" refers to total variance minimization, while \"NL-means\" denotes the non-local means algorithm.", + "bbox": [ + 511, + 414, + 906, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Defense on Black-Box Attack", + "text_level": 1, + "bbox": [ + 511, + 502, + 774, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 shows that PuriFlow, using $t' = 10$ approximately $6.67\\%$ of the diffusion time required by diffusion-only method [38]—achieves notable improvements in both standard and robust accuracy for ResNet-50. Specifically, it shows enhancements of $6.21\\%$ and $5.48\\%$ in each accuracy on ImageNet-1k. These results highlight the synergy of initialization for diffusion, indicating better performance and greater efficiency in this black-box attack.", + "bbox": [ + 511, + 526, + 906, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 664, + 666, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of SR compared to various candidates. Puri-Flow can be configured by combining various initialization methods. We test several SR models and traditional denoising techniques by selectively replacing each method used in diffusion-based purification. This evaluation, conducted with PGD targeting the classifier on ImageNet-1k, is shown in Table 7. While Wavelet denoising is practical for PGD, it stands less effective than SR techniques. Among SR models tested, MDSR $\\times 2$ outperforms DRLN $\\times 2$ , EDSR $\\times 2$ , and ESRGAN $\\times 2$ , achieving the highest standard and robust accuracies. Furthermore, we explore different upsample ratios $(\\times 2, \\times 4, \\times 8)$ using ESRGAN. Table 10 indicates that $\\times 2$ represents the best balance between efficiency and performance, making it the optimal choice for PuriFlow.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4611", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a863487964602bc927b57e31af7aab08f3acaa111cdcb910bf8ec5561b30071f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SRDiffusionClassifierStandard (%)Robust (%)Wall time (s /img)
✓76.1300.00≈0.00
✓✓73.6830.700.01
✓✓74.5238.371.08
✓✓✓74.0049.921.09
", + "bbox": [ + 99, + 89, + 575, + 166 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02dc552cb589addb9523ada30f6322d94fbac07830d37582c60bcae80cb33059.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTypeStandard (%)Robust (%)
ResNet-50VP-SDE72.9768.75
VE-SDE73.3966.51
DeiT-SVP-SDE77.2972.37
VE-SDE77.7269.33
", + "bbox": [ + 576, + 89, + 895, + 166 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7d629769932d338344cadaa13017dc42fb8261dc498e9fc7b7442e17843f5f5d.jpg", + "table_caption": [ + "Table 8. Evaluation of PuriFlow under a preprocessor-blind PGD attack on ImageNet-1k selectively employing off-the-shelf models. Left: ResNet-50, ESRGAN×2, and VP-SDE are the classifier, SR model, and diffusion [38] model, respectively, with diffusion time $t' = 10$ . Right: Impact of diffusion models within PuriFlow, evaluated at diffusion time $t' = 90$ using ResNet-50 and DeiT-S on ImageNet-1k." + ], + "table_footnote": [], + "table_body": "
Off-the-shelf modelsESRGAN×2EDSR×2MDSR×2
SR (each column)0.010.01≈0.00
Diffusion [38]1.080.810.73
ResNet-50≈0.00≈0.00≈0.00
Total wall time (s / img)1.090.820.73
ConfigurationCriterionIterationsPurif. time
Diffusion [38]SDEst' = 1005.27
SR+DiffusionSR+SDEst' = 1000.005+5.27
ScoreOpt-O [60]iOSDM = 50.51
SR+ScoreOpt-OSR+iOSDM = 50.005+0.51
", + "bbox": [ + 91, + 236, + 483, + 383 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d680720cc2e68b0f3be98d294e8b28448c9acb70be9ac50ec3184e9841267a9f.jpg", + "table_caption": [ + "Table 9. Top: Wall time for PuriFlow with different SR models. Bottom: Examples of purification times (s/img) for configurations using WRN-70-16 [38] and WRN-28-10 [60] on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
RatioStandard (%)Robust (%)FLOPsParams
×871.26%66.80%1110G16.7M
×472.41%67.32%899G16.7M
×272.97%68.75%224G16.7M
", + "bbox": [ + 102, + 446, + 470, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Impact of SR upsampling ratios from ESRGAN $\\times 2$ combined with diffusion [38] ( $t' = 90$ ) using ResNet-50 on ImageNet-1k. FLOPs represent computational complexity [23], and Params refer to the number of trainable parameters.", + "bbox": [ + 89, + 521, + 483, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Off-the-shelf model adoption and latency analysis. Table 8 shows the enhanced defensive capabilities achieved by integrating off-the-shelf models into PuriFlow's framework. The table highlights the individual and combined effects of SR and diffusion. When used together, SR and diffusion improve robust accuracy by $19.22\\%$ and $11.55\\%$ , respectively, illustrating their strong synergistic effect on adversarial examples. Although combining SR with diffusion slightly decreases standard accuracy compared to diffusion alone, the robustness gains make this trade-off useful, with a tolerable latency of $1.09~\\mathrm{s / img}$ when executed on a single NVIDIA RTX 4090 GPU. Table 9 further confirms that integrating SR models adds minimal latency to the overall purification process. Specifically, SR's execution time is $1,054\\times$ faster than the diffusion's, and its overhead is only $1/102$ compared to one-shot denoising methods. This result emphasizes SR's role as an effective and efficient initialization.", + "bbox": [ + 89, + 604, + 482, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of diffusion types. We explore VE-SDE [46], which can act as an alternative to diffusion. Table 8 demon", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59dbb534d522936d3ff20b6b46c71b63ae0c4133cad62412958869e6f977e395.jpg", + "image_caption": [ + "Figure 6. Impact of iterative SR on cross-entropy and MSEs under a PGD $(N = 100)$ attack targeting ResNet-50 within an $\\ell_{\\infty}$ -norm ball $(\\epsilon = 4 / 255)$ on ImageNet-1k. SR, SR', and SR\" indicate one, two, and three rounds of SR using ESRGAN $\\times 2$ , respectively." + ], + "image_footnote": [], + "bbox": [ + 519, + 237, + 728, + 373 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85d77e4784240ce9cfa162e1157267cbf7396a147b154f717fc79d492661890b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 238, + 900, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "strates that VP-SDE slightly reduces standard accuracy (approximately $0.42\\%$ for ResNet-50 and $0.43\\%$ for DeiT-S) but enhances robust accuracy by $2.24\\%$ and $3.04\\%$ , respectively. Despite the minor drop in standard accuracy, integrating VP-SDE shows overall robustness, synergizing a balanced performance with SR.", + "bbox": [ + 511, + 462, + 906, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Iterative applications of SR without diffusion. Figure 6 shows the impact of repeated SR applications for adversarial examples relative to their originals, with changes in MSE and cross-entropy and their impact on standard and robust accuracy. Up to three rounds of SR demonstrate a trade-off: MSE increases while cross-entropy decreases and eventually saturates. Here, robust accuracy rises but converges while standard accuracy decreases. We conjecture that repeated SR may behave as seen in model inversion [35, 57]: Images are generated to minimize cross-entropy loss, but this steady direction moves them farther from the originals.", + "bbox": [ + 511, + 560, + 908, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 739, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce PuriFlow, a novel purification flow integrating Super-Resolution (SR) with diffusion. While SR alone is less effective at mitigating adversarial noise, its image upsampling and downsampling excel at restoring subtle features captured by early layers. Diffusion alone lacks this ability. However, their mixture synergistically reduces overall feature proximity and cross-entropy. PuriFlow outperforms leading methods in extensive evaluations, demonstrating efficiency for minimal SR overhead.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4612", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was partly supported by the IITP(Institute of Information & Communications Technology Planning & Evaluation)-ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT)(IITP-2025-RS-2023-00258649, $50\\%$ ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-00562437, $40\\%$ ), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), $10\\%$ ).", + "bbox": [ + 89, + 116, + 483, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 314, + 187, + 330 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yasin Almalioglu, Kutsev Bengisu Ozyoruk, Abdulkadir Gokce, Kagan Incetan, Guliz Irem Gokceler, Muhammed Ali Simsek, Kivanc Ararat, Richard J Chen, Nicholas J Durr, Faisal Mahmood, et al. Endol2h: deep super-resolution for capsule endoscopy. IEEE Transactions on Medical Imaging, 2020. 4", + "[2] Saeed Anwar and Nick Barnes. Densely residual laplacian super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7", + "[3] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018. 1, 6", + "[4] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International Conference on Machine Learning, 2018. 6", + "[5] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In International Conference on Machine Learning, 2024. 1, 2, 3, 4, 6", + "[6] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 2021. 6, 7", + "[7] Kartikeya Bhardwaj, Dibakar Gope, James Ward, Paul Whatmough, and Danny Loh. Super-efficient super resolution for fast adversarial defense at the edge. In 2022 Design, Automation & Test in Europe Conference & Exhibition, 2022. 3, 5", + "[8] Kartikeya Bhardwaj, Milos Milosavljevic, Liam O'Neil, Dibakar Gope, Ramon Matas, Alex Chalfin, Naveen Suda, Lingchuan Meng, and Danny Loh. Collapsible linear blocks for super-efficient super resolution. Proceedings of Machine Learning and Systems, 2022. 3", + "[9] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + ], + "bbox": [ + 99, + 339, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1", + "[11] Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, and J Zico Kolter. (certified!!) adversarial robustness for free! In International Conference on Learning Representations, 2023. 1", + "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 2022. 5", + "[13] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, 2020. 1, 6", + "[14] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Li Chen, Michael E Kounavis, and Duen Horng Chau. Keeping the bad guys out: Protecting and vaccinating deep learning withJPEG compression. arXiv preprint arXiv:1705.02900, 2017.3", + "[15] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems, 2019. 3", + "[16] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 3", + "[17] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. 6, 7", + "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 4, 5", + "[19] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6", + "[20] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 2021. 6", + "[21] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3", + "[22] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 3", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Mitch Hill, Jonathan Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. International Conference on Learning Representations, 2020. 6", + "[25] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 3", + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3, 6", + "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 2020. 1", + "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, 2016. 2, 4", + "[29] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. Advances in Neural Information Processing Systems, 2024. 6", + "[30] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 1", + "[31] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 4", + "[32] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6", + "[33] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6, 7", + "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1", + "[35] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. Google Research Blog, 2015. 8", + "[36] Aamir Mustafa, Salman H. Khan, Munawar Hayat, Jianbing Shen, and Ling Shao. Image super-resolution as a defense against adversarial attacks. IEEE Transactions on Image Processing, 2020. 3", + "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2", + "[38] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, 2022. 6", + "[40] Aaditya Prakash, Nick Moran, Solomon Garber, Antonella DiLillo, and James Storer. Deflecting adversarial attacks with pixel deflection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 3", + "[41] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6", + "[42] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 2020. 6, 7", + "[43] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 3", + "[44] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2", + "[45] Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo. Perceptual extreme super-resolution network with receptive field block. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 4", + "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 8", + "[47] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 3, 6", + "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 3, 4", + "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3", + "[50] Soobin Um, Suhyeon Lee, and Jong Chul Ye. Don't play favorites: Minority guidance for diffusion models. In International Conference on Learning Representations, 2024. 5", + "[51] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 2011. 3", + "[52] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision Workshops, 2018. 4, 6, 7", + "[53] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 6, 7", + "[54] Fei Wu, Wenxue Yang, Limin Xiao, and Jinbin Zhu. Adaptive wiener filter and natural noise to eliminate adversarial perturbation. *Electronics*, 2020. 3", + "[55] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 3", + "[56] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, 2019. 6", + "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 8", + "[58] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 1, 3, 6", + "[59] MD Zeiler. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, 2014. 4", + "[60] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 2023. 2, 3, 5, 6, 8" + ], + "bbox": [ + 91, + 90, + 482, + 542 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4615", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_model.json b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_model.json new file mode 100644 index 0000000000000000000000000000000000000000..798da754c0b625587a4fbf425b240dd51cd673e0 --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_model.json @@ -0,0 +1,2411 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Adversarial Purification via Super-Resolution and Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.18, + 0.771, + 0.217 + ], + "angle": 0, + "content": "Mincheol Park\\(^{1,2}\\), Cheonjun Park\\(^{5}\\), Seungseop Lim\\(^{4}\\), Mijin Koo\\(^{6}\\), Hyunwuk Lee\\(^{2}\\), Won Woo Ro\\(^{3}\\), Suhyun Kim\\(^{7*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.217, + 0.895, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Samsung Advanced Institute of Technology, \\(^{2}\\)Samsung Electronics, \\(^{3}\\)Yonsei University, \\(^{4}\\)AITRICS, \\(^{5}\\)Hankuk University of Foreign Studies, \\(^{6}\\)Seoul National University, \\(^{7}\\)Kyung Hee University" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.255, + 0.777, + 0.287 + ], + "angle": 0, + "content": "{mincheol.park, cheonjun.park, hyunwuk.lee, wro}@yonsei.ac.kr, ss.lim@aitrics.com, starmj09@snu.ac.kr, dr.suhyun.kim@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.328, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Deep neural networks are widely used in various computer vision tasks, but their vulnerability to adversarial perturbations remains a significant challenge for reliable decision-making. Adversarial purification, a test-time defense strategy, has shown potential in countering these threats by removing noise through diffusion models. This plug-and-play method, using off-the-shelf models, appears highly effective. However, the purified data from diffusion often deviates more from the original data than the adversarial examples, leading to missing critical information and causing misclassification. In this study, we propose that upsampling with Super-Resolution (SR), followed by downsampling, can also aid in eliminating adversarial noise, similar to the noise addition and removal process in diffusion models. While SR alone is not as effective as the diffusion process, it better restores the original features typically associated with the early layers of networks. By combining SR, which initially mitigates damage to early-layer information from adversarial attacks, with diffusion, we observe a synergistic effect, leading to enhanced performance over diffusion models alone. Our comprehensive evaluations demonstrate that this combined approach, PuriFlow, significantly improves accuracy and robustness, working synergistically with state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.224, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Deep Neural Networks (DNNs) have been widely accepted in various computer vision tasks. However, their susceptibility to small, human-imperceptible noises can lead to untrustworthy decision-making. In particular, the subtle, maliciously manipulated noises, known as adversarial perturbations, produce fatal images called adversarial examples [10]. This intentional threat appears across numerous" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.589, + 0.908, + 0.645 + ], + "angle": 0, + "content": "Figure 1. The overview of PuriFlow, which begins with restoring features of adversarial examples using the SR and drifting downsampled SR images within the forward diffusion. The perturbed images are then denoised in the reverse diffusion for classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.718 + ], + "angle": 0, + "content": "types of attacks [3, 9, 13, 34], leaving many DNNs vulnerable and posing substantial challenges to their deployment in AI-powered systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.87 + ], + "angle": 0, + "content": "To counter the adversarial threats, adversarial purification [38, 58] has emerged as a test-time defense strategy. This approach is powered by leveraging off-the-shelf diffusion models. At its core, noise removal through stochastic denoising effectively eliminates artifacts introduced by various attacks. Unlike adversarial training and certified methods [27, 30, 34], which require prior knowledge of target attacks, this simple application of generative power excels in protecting classifier without the need for training, pointing to a new direction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "While diffusion models appear effective in purification, a key problem remains widely discussed in studies [5, 11, 32," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.488, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.486, + 0.592 + ], + "angle": 0, + "content": "Figure 2. The example of purification on the clean images. The top row exhibits the ground truth images. The second and third rows depict purified images obtained using diffusion and our proposed PuriFlow. The fourth row visualizes Grad-CAM [44] as represented by ResNet-50. Compared to Diffpure, PuriFlow enhances visual fidelity and accuracy by preserving DNN's activation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.484, + 0.852 + ], + "angle": 0, + "content": "38, 60]: how to lower the upper bound of differences between purified and original data, which can exceed those caused by adversarial attacks. This concern also arises when clean data is input, as shown in Figure 2. However, resolving it solely within the diffusion model is not straightforward. The noise predictor in off-the-shelf diffusion models is optimized to predict only the mean under isotropic covariance [26], limiting its ability to fully restore subtle features tied to true data covariates. To improve the restoration of details, diffusion models may require retraining with a proper objective [37], which may not align well with test-time defenses. Recently, guidance during noise removal has been proposed for test-time defenses [5]; however, this approach, despite allowing extensive processing time, is limited to large-scale datasets, impacting its generality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To improve purification without retraining, we note that adversarial attacks are designed to maximize changes in model output with minimal alterations to the input. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "results in alterations in the feature map that can be substantial, while changes in the image might be minor. Diffusion-based purification processes, which lack prior knowledge of the original image, risk deviating in the wrong direction during the noise addition and recovery phases, especially if the starting image contains feature information significantly distorted by adversarial noises. Therefore, it is crucial to eliminate as much distorted information as possible before applying the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.378 + ], + "angle": 0, + "content": "We hypothesize that using Super-Resolution (SR) to upsample and then downsample an image could help reduce noise. Our tests confirm that while SR does reduce adversarial noise, it is not as effective as diffusion models. However, an in-depth experiment reveals that SR performs better than diffusion models in restoring features to the early layers of a network. This insight suggests that initially, using SR to reduce adversarial noise before applying diffusion could enhance overall performance by restoring the features the diffusion process alone cannot recover." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.62 + ], + "angle": 0, + "content": "This paper proposes PuriFlow, an enhanced purification flow that leverages the synergy of SR and the diffusion process. As shown in Figure 1, PuriFlow integrates SR before the forward diffusion process to restore distorted features of adversarial images early on. By boosting content similarity [28], SR generates enhanced samples with features closer to the originals, even when mapped into a high-dimensional space. Our findings show that content proximity increases when SR images are resized back to their original dimensions, positioning resized SR images in regions with a higher probability for correct labels. In other words, resizing SR effectively reduces the initial cross-entropy on adversarial examples, allowing better diffusion. Our theoretical analysis and comprehensive evaluations demonstrate that this initialization aids in restoring features by diffusion, further reduces cross-entropy, and improves accuracy." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.62, + 0.874, + 0.634 + ], + "angle": 0, + "content": "Our contributions can be summarized as three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.636, + 0.905, + 0.695 + ], + "angle": 0, + "content": "- We find the purification effectiveness of the SR process. While SR alone is less effective in fully eliminating adversarial noises, it excels at preserving original features in the early layers of the network." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.697, + 0.905, + 0.756 + ], + "angle": 0, + "content": "- Our approach, PuriFlow, finds a synergistic effect by integrating SR with diffusion. Our insight into geometrical analysis and empirical study finds that this integration can reduce cross-entropy in adversarial examples." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.757, + 0.905, + 0.817 + ], + "angle": 0, + "content": "- Our pipeline utilizes only off-the-shelf SR and diffusion models. With negligible execution time required for SR, this plug-and-play method demonstrates improved accuracy and robustness in comprehensive evaluations." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.636, + 0.905, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.655, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Denoising defenses. Traditional test-time adversarial defenses often apply denoising techniques to input images. JPEG compression has been employed to remove high" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "frequency noise [14, 16], but Guo et al. [22] found it insufficient, proposing total variance minimization with image quilting as an alternative. Likewise, Prakash et al. [40] use wavelet denoising combined with BayesShrink to mitigate artifacts leading to misclassification, which surpasses total variance minimization and Wiener filtering in adversarial defenses [54]. Xie et al. [55] denoise feature maps through non-local means, supporting adversarial training. While practical, these methods often distort image details and face a core challenge: the inverse problem, which results in non-unique solutions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.484, + 0.421 + ], + "angle": 0, + "content": "SR defenses. Modern SR models have gained attention to enhance perceptual quality in adversarial examples [36]. Recently, Bhardwaj et al. [7] raised concerns about the computational demands of the SR models, which can limit real-time applicability. They demonstrate the effectiveness of a highly efficient SR model [8], showing robust performance even on compact neural processing units. PuriFlow shares a similar purpose on rapid SR applications but adds weight to its synergistic use for restoring refined features with minimal overhead, assisting diffusion to enhance defense." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.673 + ], + "angle": 0, + "content": "Adversarial purification. Purifying adversarial noise using generative models has shown promise in strengthening classifier robustness as a test-time defense. PixelDefend [47] utilizes an autoregressive generative model, while Defense-GAN [43] relies on GANs. Langevin Dynamics (LD) sampling has also been effective for defenses through energy-based models [15, 21, 25]. Yoon et al. [58] introduce a denoising score-based model with an LD variant for denoising processes. DiffPure [38] and Lee and Kim [32] employ forward and reverse-time Stochastic Differential Equations (SDEs) [26, 46] for purification, but the loss of class information remains a risk due to the dependence on diffusion time steps. Bai et al. [5] mitigate this problem with contrastive guidance, while ScoreOpt [60] uses a one-shot denoiser. This work attempts to tackle losing crucial information in purified images by combining SR with diffusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.398, + 0.701 + ], + "angle": 0, + "content": "3. Background for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.483, + 0.845 + ], + "angle": 0, + "content": "The score-based diffusion model is widely accepted in adversarial purification. It performs with bidirectional processes: forward and reverse time transition of \\(\\mathbf{x}_t\\in \\mathbb{R}^d\\) over the time interval \\(t\\in [0,T]\\). Starting from an original sample \\(\\mathbf{x}_0\\sim p_0(\\mathbf{x})\\) where \\(p_0(\\mathbf{x})\\) is the unknown, true data distribution, the diffusion model progressively transforms \\(p_0(\\mathbf{x})\\) into a nearly spherical Gaussian distribution \\(p_T(\\mathbf{x})\\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\). This converted distribution \\(p_T(\\mathbf{x})\\) retains no information about \\(p_0(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.845, + 0.483, + 0.876 + ], + "angle": 0, + "content": "This forward process of \\(\\{\\mathbf{x}_t\\}_{t = 0}^T\\) , as defined by Ito SDE, is given with the positive increment of \\(t\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} = f (\\mathbf {x}, t) d t + g (t) d \\mathbf {w}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.227 + ], + "angle": 0, + "content": "where the initial data \\(\\mathbf{x}_0\\sim p_0(\\mathbf{x}),f(\\cdot ,t):\\mathbb{R}^d\\mapsto \\mathbb{R}^d\\) is the drift coefficient of \\(\\mathbf{x}_t\\) and \\(g(\\cdot):\\mathbb{R}\\mapsto \\mathbb{R}\\) is the diffusion coefficient tied with Brownian motion \\(\\mathbf{w}\\in \\mathbb{R}^d\\). Here, when \\(f(\\mathbf{x},t)\\) is affine, the transition kernel always becomes a Gaussian distribution \\(p_{t^{\\prime}t}(\\mathbf{x}_t|\\mathbf{x}_{t^{\\prime}})\\) where \\(0\\leq t^{\\prime} < t\\leq T\\) [49]. Thus, the forward diffusion process allows for a direct transition with a closed form using the Gaussian kernel \\(p_{0t}(\\mathbf{x}_t|\\mathbf{x}_0)\\) from \\(t = 0\\) to a certain \\(t\\), which prevents the need for neural network estimations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Given a perturbed sample \\(\\mathbf{x}_T\\sim p_T(\\mathbf{x})\\) , the reverse-time SDE defines that it drifts backward in the time steps:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.269, + 0.907, + 0.288 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] d t + g (t) d \\bar {\\mathbf {w}}, \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.298, + 0.907, + 0.405 + ], + "angle": 0, + "content": "where \\(\\bar{\\mathbf{w}}\\) represents a standard Wiener process and the infinitesimal time step \\(dt\\) is negative. A key component in Eq. 2 is the score function \\(\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})\\), which offers how the density \\(p_t(\\mathbf{x})\\) changes with respect to \\(\\mathbf{x}\\). Instead of directly knowing this function, one approach is to train a neural network, \\(\\mathbf{s}_{\\theta}(\\mathbf{x},t)\\), to estimate it. This training is achieved by solving a denoising score-matching [51] problem:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.414, + 0.905, + 0.454 + ], + "angle": 0, + "content": "\\[\n\\ell (\\theta) = \\sum_ {t = 1} ^ {T} \\lambda (t) \\mathbb {E} _ {\\mathbf {x}, \\tilde {\\mathbf {x}} | \\mathbf {x}} [ \\mathbf {s} _ {\\theta} (\\tilde {\\mathbf {x}}, t) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {0 t} (\\tilde {\\mathbf {x}} | \\mathbf {x}) | | _ {2} ^ {2} ]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.906, + 0.545 + ], + "angle": 0, + "content": "Here, Gaussian kernel \\(p_{0t}(\\tilde{\\mathbf{x}}|\\mathbf{x})\\) depicts the transition from \\(\\tilde{\\mathbf{x}} = \\mathbf{x}_t\\) to \\(\\mathbf{x} = \\mathbf{x}_0\\) and \\(\\lambda(t)\\) is weighting function ensuring that the contributions at different times are appropriately balanced. This optimized model serves as \\(\\mathbf{s}_{\\theta^*}(\\mathbf{x}, t)\\) where \\(\\theta^* = \\mathrm{argmin}_{\\theta} \\ell(\\theta)\\) to replace the elusive score function, enabling an effective reverse diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.649, + 0.575 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.63 + ], + "angle": 0, + "content": "In this section, we introduce our motivation, design methodology, and geometrical analysis of the proposed approach, PuriFlow. To begin, we briefly review adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.64, + 0.906, + 0.716 + ], + "angle": 0, + "content": "Adversarial attack. Given a clean image \\(\\mathbf{x}_0\\) with label \\(y\\), adversarial attacks generally seek perturbations to mislead a classifier \\(\\mathcal{F}_{\\phi}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{K}\\) into incorrectly classifying \\(\\mathbf{x}_0\\). This manipulated image \\(\\mathbf{x}_0^a\\) is generated by maximizing cross-entropy loss within a tiny region around \\(\\mathbf{x}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.727, + 0.907, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {0} ^ {a} = \\lim _ {i \\rightarrow N} \\prod_ {\\mathcal {B} \\left(\\mathbf {x} _ {0} ^ {i}, \\epsilon\\right)} \\left(\\mathbf {x} _ {0} ^ {i} + \\mu \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {0} ^ {i}} \\ell_ {c e} (\\phi ; \\mathbf {x} _ {0} ^ {i}, y)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_0^0\\) is the initial clean image \\(\\mathbf{x}_0\\) and \\(N\\) is the number of attack iterations. \\(\\mathcal{B}(\\mathbf{x}_0^i,\\epsilon)\\) is the \\(\\ell_p\\)-norm ball around \\(\\mathbf{x}_0^i\\) within a radius of \\(\\epsilon\\), \\(\\Pi (\\cdot)\\) is the projection to the norm ball, and \\(\\mu\\) denotes the step size. The gradient can be obtained from surrogate networks when \\(\\mathcal{F}_{\\phi}(\\cdot)\\) is unknown." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.849, + 0.864 + ], + "angle": 0, + "content": "4.1. PuriFlow: Enhanced Purification Flow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Given an adversarial example, \\(\\mathbf{x}_0^a\\), VP-SDE [48] is widely accepted approach for purification. VP-SDE controls noise" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Top: visualization of different image types. AE denotes the adversarial example generated by a PGD attack on ResNet-50 using an ImageNet-1k image. SR indicates the downsampled high-resolution image. Diffusion and PuriFlow use the same diffusion method, VP-SDE [38]. Bottom: corresponding feature maps from the Conv2_2 layer of VGG-19, as described in [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.412 + ], + "angle": 0, + "content": "addition through drift coefficients \\( f(\\mathbf{x}, t) = -\\frac{1}{2} \\beta_t \\mathbf{x} \\) and \\( g(t) = \\sqrt{\\beta_t} \\), where \\( \\beta_t = \\frac{1}{t} \\beta_{\\text{max}} + (1 - \\frac{1}{t}) \\beta_{\\text{min}} \\) [38]. This model then diffuses the data by solving the forward-time SDE in Eq. 1 with a closed form:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.423, + 0.483, + 0.44 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} ^ {a} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} ^ {a} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.388, + 0.468 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)\\) and \\(\\bar{\\alpha}_{t} = e^{-\\int_{0}^{t}\\beta_{t}(s)ds}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.499 + ], + "angle": 0, + "content": "The corresponding diffusion process is then performed by solving the reverse-time SDE in Eq. 2:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.51, + 0.482, + 0.528 + ], + "angle": 0, + "content": "\\[\nd \\mathbf {x} = \\left[ f \\left(\\mathbf {x} _ {t} ^ {a}, t\\right) - g ^ {2} (t) \\nabla_ {\\mathbf {x} _ {t} ^ {a}} \\log p _ {t} \\left(\\mathbf {x} _ {t} ^ {a}\\right) \\right] d t + g (t) d \\bar {\\mathbf {w}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.554 + ], + "angle": 0, + "content": "where the drift coefficients, \\( f(\\cdot ,t) \\) and \\( g^{2}(t) \\), follows [5, 48]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.719 + ], + "angle": 0, + "content": "During this reverse process, the crucial step is computing the score, defined as \\(\\nabla_{\\mathbf{x}_t^a}\\log p_t(\\mathbf{x}_t^a)\\). Here, we raise a key question: Can the neural network \\(s_{\\theta^*}(\\cdot ,t)\\), trained for score estimation, precisely predict \\(\\mathbf{x}_0\\) from \\(\\mathbf{x}_t^a\\)? Even if \\(s_{\\theta^*}(\\cdot ,t)\\) is well-trained, accurately estimating the direction toward \\(\\mathbf{x}_0\\) is not straightforward when dealing with the noised data \\(\\mathbf{x}_t^a\\) that has been perturbed from \\(\\mathbf{x}_0^a\\), since \\(\\mathbf{s}_{\\theta^*}(\\cdot ,t)\\) has not been exposed to such perturbed data during training with Eq. 3. In other words, the trained neural network lacks prior knowledge of guiding \\(\\mathbf{x}_t^a\\), risking deviations in the wrong direction during the noise addition and recovery phases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.856 + ], + "angle": 0, + "content": "This concern is seemingly explicit. Figure 3 shows that the perturbations introduced by Eq. 4 appear subtle in pixel values but cause significant damage to the underlying features of \\(\\mathbf{x}_0\\). These observations indicate that diffusion alone struggles to fully reconstruct the original features. To address this, we seek an existing method that positions the perturbed data closer to its ground-truth image, aiming to reduce feature discrepancies early on and thus ensure a more precise projection of the score function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To this end, we focus on a model that can effectively reduce content distance [18, 28] to capture feature proximity. Reducing feature-level shifts that lead to increased cross" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.198, + 0.874, + 0.211 + ], + "angle": 0, + "content": "(a) Impact of SR compared to diffusion with increments of \\(t^\\prime\\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.215, + 0.905, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.323, + 0.905, + 0.347 + ], + "angle": 0, + "content": "(b) Synergistic effect in feature restoration of SR combined with diffusion. Diffusion and SR+Diffusion use the same diffusion time \\(t^{\\prime} = 90\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.36, + 0.906, + 0.403 + ], + "angle": 0, + "content": "Figure 4. Normalized content distance from the ground truth for each image type, measured across specific CONV layers of VGG-19 as shown in Figure 5. Real values are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.579 + ], + "angle": 0, + "content": "entropy aligns with the goal of purification, ultimately improving classification accuracy. This approach also benefits clean data in preserving features. In this context, the SR model suits this purpose. The SR model, \\(\\mathcal{H}_{\\omega^{*}}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{D}\\) such that \\(D > d\\) , is designed to upsample and interpolate low-resolution images into high-resolution counterparts. To optimize its parameters \\(\\omega\\) , various SR methods [1, 28, 31, 45, 52] minimize both pixel distance and VGG loss [31], which captures mutual similarity in feature maps across different layers of the VGG model:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.589, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\omega^ {*} = \\min _ {\\omega} \\frac {1}{W ^ {l} H ^ {l}} \\left(\\sum_ {u = 1} ^ {W ^ {l}} \\sum_ {v = 1} ^ {H ^ {l}} \\left[ \\Phi^ {l} (\\mathbf {X}) _ {u, v} - \\Phi^ {l} \\left(\\mathcal {H} _ {\\omega} (\\mathbf {x})\\right) _ {u, v} \\right] ^ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.719 + ], + "angle": 0, + "content": "where \\(\\mathbf{X} \\in \\mathbb{R}^{D}\\) and \\(\\mathbf{x} \\in \\mathbb{R}^{d}\\), with \\(W^{l}\\) and \\(H^{l}\\) denoting the dimensions of extracted features \\(\\Phi^l (\\cdot)\\) in the \\(l\\)-th convolution layer. Thus, SR models can naturally reduce the content distance by enhancing feature similarity while estimating unknown pixels and their covariates in \\(\\mathbb{R}^{D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.886 + ], + "angle": 0, + "content": "Our findings, as illustrated in Figure 4, show that SR images downsampled from \\(\\mathbb{R}^D\\) to \\(\\mathbb{R}^d\\) can also reduce the content distance. This reduction covers all layers of the VGG network, demonstrating that even a one-shot application can effectively mitigate feature discrepancies present in adversarial examples. Notably, as observed in Figure 4a, increasing \\(t'\\) when applying diffusion alone cannot sufficiently aid in restoring features tied to the early layers of the CNN. This observation suggests that a single application of SR can be more effective than diffusion in recovering detailed features [59] captured by the early layers." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose an enhanced purification flow," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.297, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.089, + 0.498, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.701, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.089, + 0.903, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Figure 5. Transition of diffusion only and SR-integrated diffusion with the same diffusion time of \\( t' = 90 \\) under a PGD attack (\\( \\epsilon = 8/255 \\), \\( N = 100 \\)) on ResNet-50 using ImageNet-1k. \"SR\" represents downsampled high-dimensional images, while \"AE\" denotes adversarial examples. Content distance [18] from the ground truth is measured in two convolutional layers of VGG-19 and plotted with cross-entropy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.304, + 0.485, + 0.501 + ], + "angle": 0, + "content": "PuriFlow, where SR works before the start of the forward-time SDE. As shown in Figure 4b, this approach demonstrates the effectiveness of using SR early before solving both SDEs. By correcting feature distortions that diffusion alone struggles to recover, SR allows for a synergistic effect, enhancing feature restoration across all layers through diffusion. As observed in Figure 5, the early reduction in content distance leads to a notable decrease in cross-entropy. Thus, the subsequent diffusion aligns the data more closely with the ground truth. To delve into this synergistic effect, we provide a theoretical study from a manifold perspective. We examine how SR influences data positioning and allows for more precise use of the score function within diffusion." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.51, + 0.393, + 0.525 + ], + "angle": 0, + "content": "4.2. Geometrical Analysis on PuriFlow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Suppose \\(\\mathcal{M} \\subset \\mathbb{R}^d\\) is the set of all data points, i.e., \\(\\forall \\mathbf{x} \\in \\mathcal{M}\\), defining \\(\\mathcal{M}\\) as the data manifold. That is, the data distribution \\(p_0\\) is uniform on the data manifold \\(\\mathcal{M}\\). Then, within a local neighborhood, the manifold coincides with its tangent space of dimension \\(k \\ll d\\), expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.618, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right) = T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right), \\quad T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\simeq \\mathbb {R} ^ {k}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.721 + ], + "angle": 0, + "content": "where \\( T_{\\mathbf{x}_0}\\mathcal{M} \\) denotes the tangent space to the manifold \\( \\mathcal{M} \\) at \\( \\mathbf{x}_0 \\) and \\( dr \\) is an infinitesimal radius within which the manifold can be locally approximated by its tangent space. Here, we take note of a property of the trained score function in diffusion models [12]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.807 + ], + "angle": 0, + "content": "Remark 1. The noisy data become increasingly concentrated on spherical manifolds as diffusion time \\(t\\) increases. The score function is trained via denoising score matching (Eq. 3), using data on \\(\\mathcal{M}\\). Let \\(\\mathcal{P}_t(\\cdot)\\) denote a mapping from noisy data \\(\\pmb{x}_t\\) to its estimated clean counterpart \\(\\hat{\\pmb{x}}_0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.816, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {t} (\\cdot): \\boldsymbol {x} _ {t} \\mapsto \\hat {\\boldsymbol {x}} _ {0} = - \\frac {1}{2} (1 - \\alpha_ {t}) \\left(\\boldsymbol {x} _ {t} + 2 \\boldsymbol {s} _ {\\theta^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{P}_t(\\pmb{x}_t) \\in \\mathcal{M}\\) and it satisfies \\(\\pmb{J}_{\\mathcal{P}_t}^2 = \\pmb{J}_{\\mathcal{P}_t} = \\pmb{J}_{\\mathcal{P}_t}^T : \\mathbb{R}^d \\to T_{\\mathcal{P}_t(\\pmb{x}_t)}\\mathcal{M}\\). That is, \\(\\mathcal{P}_t(\\cdot)\\) acts as an orthonormal projection onto \\(\\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Remark 2. The score function is trained solely on the data concentrated on both the noisy and data manifolds. Therefore, applying it to data outside these manifolds may lead to inaccurate inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Both premises suggest that when the score function is applied to off-manifold data, such as adversarial examples [7], its inaccurate inference may lead to misaligned orthonormal projections onto an unintended tangent space. Thus, the off-manifold data should be corrected to better align with its underlying manifold. Now, we show that the applying SR followed by downsampling can typically position off-manifold data, degraded from \\(\\mathbf{x}_0\\), closer to the local manifold, which is the tangent space of \\(\\mathbf{x}_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.511, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Proposition 1. Suppose \\(\\mathcal{H}_{\\omega^*}\\) is an ideal SR and smooth continuity, and let \\(\\mathcal{R}\\) be a linear downsampler. Given an offmanifold data \\(\\mathfrak{D}(\\pmb{x}_0)\\) degraded from \\(\\pmb{x}_0\\), the processed data \\(\\mathcal{R}\\mathcal{H}_{\\omega^*}(\\mathfrak{D}(\\pmb{x}_0))\\) is projected near the tangent space \\(T_{\\pmb{x}_0}\\mathcal{M}\\) of the data manifold \\(\\mathcal{M}\\) at \\(\\pmb{x}_0\\) with a tiny \\(\\delta > 0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.593, + 0.905, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\mathcal {R H} _ {\\omega^ {*}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) - \\Pi_ {T _ {x _ {0}} \\mathcal {M}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right\\| _ {F} < \\delta . \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.661 + ], + "angle": 0, + "content": "In Section 5, we evaluate the effectiveness of using the score function more accurately, supported by Proposition 1, against various adversarial attacks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.673, + 0.64, + 0.69 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.819 + ], + "angle": 0, + "content": "We evaluate the performance of PuriFlow in two configurations: integrating SR with solving SDEs [38] and combining SR with a recent alternative, iterative One-Shot Denoising (OSD) [60]. OSD is defined by Tweedie's formula [50], expressed as \\(\\mathbb{E}[\\mathbf{x}|\\tilde{\\mathbf{x}}] = \\tilde{\\mathbf{x}} +\\sigma^2\\nabla_{\\tilde{\\mathbf{x}}}\\log p_\\sigma (\\tilde{\\mathbf{x}})\\), where \\(\\tilde{\\mathbf{x}}\\) denotes perturbed data and \\(p_{\\sigma}(\\tilde{\\mathbf{x}})\\) represents marginal Gaussian perturbation of \\(p(\\mathbf{x})\\). This formulation avoids repeated noise removal steps for varying time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.819, + 0.906, + 0.88 + ], + "angle": 0, + "content": "In this context, the denoising score-matching problem defined by Eq. 3 optimizes the neural network \\(\\mathbf{s}_{\\theta^{*}}(\\cdot ,t)\\) given \\(\\mathbf{x}_t\\) to predict a fully denoised \\(\\mathbf{x}_0\\) per a specific time \\(t\\). Thus, Tweedie's formula can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} \\left[ \\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t ^ {\\prime}} \\right] = \\mathbf {x} _ {t ^ {\\prime}} + \\sigma^ {2} \\nabla_ {\\mathbf {x} _ {t ^ {\\prime}}} \\log p _ {\\sigma} \\left(\\mathbf {x} _ {t ^ {\\prime}}\\right). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4609" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.236 + ], + "angle": 0, + "content": "
MethodCriterionStandard (%)Robust (%)
Yang et al. [56]Gibbs Update94.8040.80
Song et al. [47]Mask+Recon.95.009.00
Hill et al. [24]EBM+LD84.1254.90
Yoon et al. [58]DSM+LD*86.1470.01
Nie et al. [38]SDEs89.0281.40
Bai et al. [5]Guide+SDEs92.6181.94
PuriFlow([38])SR+SDEs90.0683.00
Zhang et al. [60]iOSD93.7588.08
PuriFlow([60])SR+iOSD94.1491.01
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.312 + ], + "angle": 0, + "content": "Table 1. Comparison of different adversarial purification methods against BPDA+EOT attacks with \\(\\ell_{\\infty}\\) perturbations. Our evaluations use WRN-28-10 on CIFAR-10, with consistent experimental settings as outlined in [38, 60] for \\(\\epsilon = 8 / 255\\). (* This purification utilizes a variant of the LD sampling.)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.486, + 0.538 + ], + "angle": 0, + "content": "The alternative proposed by [60] involves iteratively applying OSD using Eq. 11 during noise removal. Therefore, it can serve as a suitable baseline to confirm synergies with SR. We assess these configurations, detailed as algorithms in the Appendix, across key test-time defense scenarios, including adaptive white-box, preprocessor-blind [58], and black-box attacks. This section also includes an ablation study to demonstrate the impact of incorporating SR. Detailed experimental settings and additional evaluations for certified robustness are provided in the Appendix. For adaptive attacks, we adopt MDSR [33] for efficiency and ESRGAN [52] otherwise. We denote \\( t' \\) on a discrete scale [26] from 1 to 1000 to represent the continuous interval [0, 1]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.35, + 0.563 + ], + "angle": 0, + "content": "5.1. Defense on Adaptive Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.485, + 0.675 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of PuriFlow against four strong adaptive attacks: BPDA+EOT [3, 4], an adaptive ensemble white-box attack; AutoAttack [13], an adaptive white-box attack; PGD+EOT [32], and DiffAttack [29], a diffusion-based purification targeted attack. Here, the evaluations focus on projections within an \\(\\ell_{\\infty}\\)-norm ball, as the projection in \\(\\ell_{2}\\)-norm ball is less effective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.82 + ], + "angle": 0, + "content": "BPDA+EOT. Backward Pass Differentiable Approximation (BPDA) combined with Expectation over Transformation (EOT) is used to evaluate randomized adversarial purification methods. Table 1 demonstrates that PuriFlow outperforms standalone diffusion on CIFAR-10 with WRN-28-10, achieving a more balanced standard and robust accuracy than the guiding method. In addition, integrating SR with iterative OSD further enhances performance, emphasizing its critical role in diffusion-based purification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "AutoAttack. Tables 2 demonstrates the effectiveness of PuriFlow against AutoAttack, on ImageNet-1k. PuriFlow outperforms adversarial training methods and a diffusion-only approach across ResNet-50 and DeiT-S. Compared to diffusion only, it achieves notable robust accuracy improve" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.09, + 0.897, + 0.277 + ], + "angle": 0, + "content": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5631.06
Wong et al. [53]55.6226.95
Salman et al. [42]64.0237.89
Bai et al. [6]67.3835.51
APNie et al. [38]67.7940.93
PuriFlow([38])71.6852.92
DeiT-S79.900.00
ATBai et al. [6]66.5035.50
APNie et al. [38]73.6343.18
PuriFlow([38])75.6953.12
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.286, + 0.908, + 0.343 + ], + "angle": 0, + "content": "Table 2. Evaluation comparing adversarial training (AT) and purification (AP) methods against AutoAttack with \\(\\ell_{\\infty}\\) perturbations at \\(\\epsilon = 4 / 255\\) on ImageNet-1k. Diffusion time is set to \\(t' = 150\\) for [38] as specified its study and \\(t' = 130\\) for PuriFlow." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.358, + 0.897, + 0.595 + ], + "angle": 0, + "content": "
TypeMethodStandard (%)Robust (%)
WRN-28-1095.630.00
ATPang et al. [39]88.6264.95
Gowal et al. [19]88.5465.93
Gowal et al. [20]87.5166.01
APYoon et al. [58]85.6633.48
Zhang et al. [60]*93.7559.37
PuriFlow([60])*94.1460.76
WRN-70-1695.790.00
ATGowal et al. [19]91.1068.66
Gowal et al. [20]88.7469.03
Rebuffi et al. [41]92.2269.97
APYoon et al. [58]86.7637.11
Zhang et al. [60]*95.1163.28
PuriFlow([60])*95.3164.86
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.604, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Table 3. Evaluation comparing adversarial training (AT) and purification (AP) methods against PGD+EOT within an \\(\\ell_{\\infty}\\)-norm ball of radius \\(\\epsilon = 8/255\\) on CIFAR-10. Results are sourced from [32]. (* For a fair comparison, PuriFlow and [60] are directly evaluated on the same settings with \\(t' = 250\\), as stated in [60].)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.748 + ], + "angle": 0, + "content": "ments, including \\(11.99\\%\\) (ResNet-50) and \\(9.94\\%\\) (DeiT-S), and standard accuracy. PuriFlow also shows efficiency with reduced diffusion time \\(t'\\) on this large-scale dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.909, + 0.911 + ], + "angle": 0, + "content": "\\(\\mathbf{PGD} + \\mathbf{EOT}\\). Motivated by [32], we assess PuriFlow's robustness against \\(\\mathrm{PGD} + \\mathrm{EOT}\\) that utilizes gradient signs within the entire randomized defense framework. Table 3 shows PuriFlow's strength on CIFAR-10, achieving the highest standard accuracy for WRN-28-10 and WRN-70-16. Despite the substantial impact of \\(\\mathrm{PGD} + \\mathrm{EOT}\\) on randomized purification, PuriFlow achieves superior robust accuracy among purification methods, demonstrating a synergistic effect in which SR effectively acts as the initialization before starting the iterative one-shot denoising process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4610" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.09, + 0.472, + 0.212 + ], + "angle": 0, + "content": "
TypeMethodStandard (%)Robust (%)
WRN-70-1695.790.00
APNie et al. [38]90.0745.31
PuriFlow([38])90.2159.47
ResNet-5076.130.00
APNie et al. [38]67.7928.13
PuriFlow([38])71.6843.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.219, + 0.483, + 0.261 + ], + "angle": 0, + "content": "Table 4. Evaluation against DiffAttack, targeting WRN-70-16 on CIFAR-10 and ResNet-50 on ImageNet-1k. Each diffusion time \\( t' \\) is the same as those used for other attacks on PuriFlow." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.275, + 0.472, + 0.461 + ], + "angle": 0, + "content": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5638.97
Wong et al. [53]55.6229.15
Salman et al. [42]64.0238.01
Bai et al. [6]67.3840.27
APNie et al. [38]74.5238.87
PuriFlow([38])74.0549.92
DeiT-S79.900.00
ATBai et al. [6]66.5040.32
APNie et al. [38]78.9434.59
PuriFlow([38])78.4449.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.512 + ], + "angle": 0, + "content": "Table 5. Comparison under a preprocessor-blind PGD attack targeting only the classifier, within an \\( \\ell_{\\infty} \\) perturbations at \\( \\epsilon = 4/255 \\) on ImageNet-1k. The time \\( t' \\) for both AP methods is set as 10." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.705 + ], + "angle": 0, + "content": "DiffAttack. DiffAttack neutralizes vanishing and exploding gradients, high memory costs, and increased randomness—natural defensive effects from lengthy \\( t' \\) of diffusion-based purification. Table 4 demonstrates that while purification methods using \\( t' = 100 \\) for CIFAR-10 and \\( t' = 150 \\) for ImageNet-1k exhibit limitations in robustness by this intentional attack, integrating SR demonstrates a notable improvement by \\( 14.16\\% \\) and \\( 15.62\\% \\). We hypothesize that SR's early enhancement of feature proximity helps mitigate diverging feature maps from the ground truth, even under the reconstruction-deviating impacts of DiffAttack." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.421, + 0.729 + ], + "angle": 0, + "content": "5.2. Defense on Preprocessor-Blind Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate PuriFlow against the PGD (\\(N = 100\\)) attack, targeting only the classifier on ImageNet-1k. Table 5 shows that PuriFlow, leveraging only off-the-shelf models, surpasses adversarial training methods in both standard and robust accuracy for ResNet-50 and DeiT-S. Notably, PuriFlow significantly outperforms diffusion only [38] at the same practical diffusion time \\(t' = 10\\), achieving comparable standard accuracy while enhancing robust accuracy by \\(11.05\\%\\) for ResNet-50 and \\(15.26\\%\\) for DeiT-S. This improvement signifies the synergistic importance of SR in adversarial purification utilizing diffusion models." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.09, + 0.895, + 0.162 + ], + "angle": 0, + "content": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.139.25
APNie et al. [38]67.7962.88
PuriFlow [38]74.0068.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.172, + 0.907, + 0.242 + ], + "angle": 0, + "content": "Table 6. Evaluation of ResNet-50 under a black-box attack using Square Attack (\\(N = 5000\\)) within an \\(\\ell_{\\infty}\\)-norm ball of radius \\(\\epsilon = 4/255\\) on ImageNet-1k. Diffusion time is set to \\(t' = 150\\) for [38] as specified in their study and \\(t' = 10\\) for PuriFlow. Evaluation settings follow those in [38]." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.261, + 0.907, + 0.405 + ], + "angle": 0, + "content": "
MethodTypeStandard (%)Robust (%)
ResNet-5076.130.00
WaveletDenoising71.1568.63
TVMDenoising57.8056.96
NL-meansDenoising64.3262.67
DRLN×2 [2]SR71.6668.72
EDSR×2 [33]SR71.7768.73
ESRGAN×2 [52]SR72.9768.75
MDSR×2 [33]SR73.3269.47
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.415, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Table 7. Comparison of SR with other denoising methods as an initialization approach for diffusion [38] (\\( t' = 90 \\)) under a PGD attack on ImageNet-1k. \"TVM\" refers to total variance minimization, while \"NL-means\" denotes the non-local means algorithm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.776, + 0.518 + ], + "angle": 0, + "content": "5.3. Defense on Black-Box Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Table 6 shows that PuriFlow, using \\( t' = 10 \\) approximately \\( 6.67\\% \\) of the diffusion time required by diffusion-only method [38]—achieves notable improvements in both standard and robust accuracy for ResNet-50. Specifically, it shows enhancements of \\( 6.21\\% \\) and \\( 5.48\\% \\) in each accuracy on ImageNet-1k. These results highlight the synergy of initialization for diffusion, indicating better performance and greater efficiency in this black-box attack." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.667, + 0.681 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Impact of SR compared to various candidates. Puri-Flow can be configured by combining various initialization methods. We test several SR models and traditional denoising techniques by selectively replacing each method used in diffusion-based purification. This evaluation, conducted with PGD targeting the classifier on ImageNet-1k, is shown in Table 7. While Wavelet denoising is practical for PGD, it stands less effective than SR techniques. Among SR models tested, MDSR \\(\\times 2\\) outperforms DRLN \\(\\times 2\\), EDSR \\(\\times 2\\), and ESRGAN \\(\\times 2\\), achieving the highest standard and robust accuracies. Furthermore, we explore different upsample ratios \\((\\times 2, \\times 4, \\times 8)\\) using ESRGAN. Table 10 indicates that \\(\\times 2\\) represents the best balance between efficiency and performance, making it the optimal choice for PuriFlow." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4611" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.09, + 0.576, + 0.167 + ], + "angle": 0, + "content": "
SRDiffusionClassifierStandard (%)Robust (%)Wall time (s /img)
✓76.1300.00≈0.00
✓✓73.6830.700.01
✓✓74.5238.371.08
✓✓✓74.0049.921.09
" + }, + { + "type": "table", + "bbox": [ + 0.577, + 0.09, + 0.897, + 0.167 + ], + "angle": 0, + "content": "
ModelTypeStandard (%)Robust (%)
ResNet-50VP-SDE72.9768.75
VE-SDE73.3966.51
DeiT-SVP-SDE77.2972.37
VE-SDE77.7269.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.171, + 0.905, + 0.214 + ], + "angle": 0, + "content": "Table 8. Evaluation of PuriFlow under a preprocessor-blind PGD attack on ImageNet-1k selectively employing off-the-shelf models. Left: ResNet-50, ESRGAN×2, and VP-SDE are the classifier, SR model, and diffusion [38] model, respectively, with diffusion time \\( t' = 10 \\). Right: Impact of diffusion models within PuriFlow, evaluated at diffusion time \\( t' = 90 \\) using ResNet-50 and DeiT-S on ImageNet-1k." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.237, + 0.485, + 0.384 + ], + "angle": 0, + "content": "
Off-the-shelf modelsESRGAN×2EDSR×2MDSR×2
SR (each column)0.010.01≈0.00
Diffusion [38]1.080.810.73
ResNet-50≈0.00≈0.00≈0.00
Total wall time (s / img)1.090.820.73
ConfigurationCriterionIterationsPurif. time
Diffusion [38]SDEst' = 1005.27
SR+DiffusionSR+SDEst' = 1000.005+5.27
ScoreOpt-O [60]iOSDM = 50.51
SR+ScoreOpt-OSR+iOSDM = 50.005+0.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.39, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Table 9. Top: Wall time for PuriFlow with different SR models. Bottom: Examples of purification times (s/img) for configurations using WRN-70-16 [38] and WRN-28-10 [60] on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.447, + 0.472, + 0.516 + ], + "angle": 0, + "content": "
RatioStandard (%)Robust (%)FLOPsParams
×871.26%66.80%1110G16.7M
×472.41%67.32%899G16.7M
×272.97%68.75%224G16.7M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Table 10. Impact of SR upsampling ratios from ESRGAN \\(\\times 2\\) combined with diffusion [38] (\\(t' = 90\\)) using ResNet-50 on ImageNet-1k. FLOPs represent computational complexity [23], and Params refer to the number of trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.862 + ], + "angle": 0, + "content": "Off-the-shelf model adoption and latency analysis. Table 8 shows the enhanced defensive capabilities achieved by integrating off-the-shelf models into PuriFlow's framework. The table highlights the individual and combined effects of SR and diffusion. When used together, SR and diffusion improve robust accuracy by \\(19.22\\%\\) and \\(11.55\\%\\), respectively, illustrating their strong synergistic effect on adversarial examples. Although combining SR with diffusion slightly decreases standard accuracy compared to diffusion alone, the robustness gains make this trade-off useful, with a tolerable latency of \\(1.09~\\mathrm{s / img}\\) when executed on a single NVIDIA RTX 4090 GPU. Table 9 further confirms that integrating SR models adds minimal latency to the overall purification process. Specifically, SR's execution time is \\(1,054\\times\\) faster than the diffusion's, and its overhead is only \\(1/102\\) compared to one-shot denoising methods. This result emphasizes SR's role as an effective and efficient initialization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Impact of diffusion types. We explore VE-SDE [46], which can act as an alternative to diffusion. Table 8 demon" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.238, + 0.73, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.239, + 0.901, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.382, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Figure 6. Impact of iterative SR on cross-entropy and MSEs under a PGD \\((N = 100)\\) attack targeting ResNet-50 within an \\(\\ell_{\\infty}\\)-norm ball \\((\\epsilon = 4 / 255)\\) on ImageNet-1k. SR, SR', and SR\" indicate one, two, and three rounds of SR using ESRGAN \\(\\times 2\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.554 + ], + "angle": 0, + "content": "strates that VP-SDE slightly reduces standard accuracy (approximately \\(0.42\\%\\) for ResNet-50 and \\(0.43\\%\\) for DeiT-S) but enhances robust accuracy by \\(2.24\\%\\) and \\(3.04\\%\\), respectively. Despite the minor drop in standard accuracy, integrating VP-SDE shows overall robustness, synergizing a balanced performance with SR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.909, + 0.728 + ], + "angle": 0, + "content": "Iterative applications of SR without diffusion. Figure 6 shows the impact of repeated SR applications for adversarial examples relative to their originals, with changes in MSE and cross-entropy and their impact on standard and robust accuracy. Up to three rounds of SR demonstrate a trade-off: MSE increases while cross-entropy decreases and eventually saturates. Here, robust accuracy rises but converges while standard accuracy decreases. We conjecture that repeated SR may behave as seen in model inversion [35, 57]: Images are generated to minimize cross-entropy loss, but this steady direction moves them farther from the originals." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.634, + 0.755 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce PuriFlow, a novel purification flow integrating Super-Resolution (SR) with diffusion. While SR alone is less effective at mitigating adversarial noise, its image upsampling and downsampling excel at restoring subtle features captured by early layers. Diffusion alone lacks this ability. However, their mixture synergistically reduces overall feature proximity and cross-entropy. PuriFlow outperforms leading methods in extensive evaluations, demonstrating efficiency for minimal SR overhead." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4612" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.299 + ], + "angle": 0, + "content": "This research was partly supported by the IITP(Institute of Information & Communications Technology Planning & Evaluation)-ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT)(IITP-2025-RS-2023-00258649, \\(50\\%\\) ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-00562437, \\(40\\%\\) ), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), \\(10\\%\\))." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.315, + 0.188, + 0.332 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.34, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[1] Yasin Almalioglu, Kutsev Bengisu Ozyoruk, Abdulkadir Gokce, Kagan Incetan, Guliz Irem Gokceler, Muhammed Ali Simsek, Kivanc Ararat, Richard J Chen, Nicholas J Durr, Faisal Mahmood, et al. Endol2h: deep super-resolution for capsule endoscopy. IEEE Transactions on Medical Imaging, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.426, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[2] Saeed Anwar and Nick Barnes. Densely residual laplacian super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.483, + 0.526 + ], + "angle": 0, + "content": "[3] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.529, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[4] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International Conference on Machine Learning, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[5] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In International Conference on Machine Learning, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[6] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[7] Kartikeya Bhardwaj, Dibakar Gope, James Ward, Paul Whatmough, and Danny Loh. Super-efficient super resolution for fast adversarial defense at the edge. In 2022 Design, Automation & Test in Europe Conference & Exhibition, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[8] Kartikeya Bhardwaj, Milos Milosavljevic, Liam O'Neil, Dibakar Gope, Ramon Matas, Alex Chalfin, Naveen Suda, Lingchuan Meng, and Danny Loh. Collapsible linear blocks for super-efficient super resolution. Proceedings of Machine Learning and Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.34, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[10] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[11] Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, and J Zico Kolter. (certified!!) adversarial robustness for free! In International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[13] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[14] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Li Chen, Michael E Kounavis, and Duen Horng Chau. Keeping the bad guys out: Protecting and vaccinating deep learning withJPEG compression. arXiv preprint arXiv:1705.02900, 2017.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.904, + 0.455 + ], + "angle": 0, + "content": "[15] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[16] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[17] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[19] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[20] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[21] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[22] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[24] Mitch Hill, Jonathan Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. International Conference on Learning Representations, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[29] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. Advances in Neural Information Processing Systems, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[30] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[31] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[32] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[33] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[35] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. Google Research Blog, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[36] Aamir Mustafa, Salman H. Khan, Munawar Hayat, Jianbing Shen, and Ling Shao. Image super-resolution as a defense against adversarial attacks. IEEE Transactions on Image Processing, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[38] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[39] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[40] Aaditya Prakash, Nick Moran, Solomon Garber, Antonella DiLillo, and James Storer. Deflecting adversarial attacks with pixel deflection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[41] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[42] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[43] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[44] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[45] Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo. Perceptual extreme super-resolution network with receptive field block. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[47] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.906, + 0.714 + ], + "angle": 0, + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[50] Soobin Um, Suhyeon Lee, and Jong Chul Ye. Don't play favorites: Minority guidance for diffusion models. In International Conference on Learning Representations, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[51] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[52] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision Workshops, 2018. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[53] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[54] Fei Wu, Wenxue Yang, Limin Xiao, and Jinbin Zhu. Adaptive wiener filter and natural noise to eliminate adversarial perturbation. *Electronics*, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[55] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[56] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[58] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[59] MD Zeiler. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[60] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4615" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_origin.pdf b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88e220c705ac69563aa0d8d3ed3d004d9466c5d1 --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/b9eb49c2-a69c-48c0-bba6-c208edb9f012_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7f274c7c1e95aed1b28b285be412ed18f108e060373e1406c38f76a1e3f0c1cb +size 1427727 diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/full.md b/2025/Adversarial Purification via Super-Resolution and Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..438be70c097e4bfc2a2fdb9cf748b790883571a3 --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/full.md @@ -0,0 +1,349 @@ +# Adversarial Purification via Super-Resolution and Diffusion + +Mincheol Park $^{1,2}$ , Cheonjun Park $^{5}$ , Seungseop Lim $^{4}$ , Mijin Koo $^{6}$ , Hyunwuk Lee $^{2}$ , Won Woo Ro $^{3}$ , Suhyun Kim $^{7*}$ + +$^{1}$ Samsung Advanced Institute of Technology, $^{2}$ Samsung Electronics, $^{3}$ Yonsei University, $^{4}$ AITRICS, $^{5}$ Hankuk University of Foreign Studies, $^{6}$ Seoul National University, $^{7}$ Kyung Hee University + +{mincheol.park, cheonjun.park, hyunwuk.lee, wro}@yonsei.ac.kr, ss.lim@aitrics.com, starmj09@snu.ac.kr, dr.suhyun.kim@gmail.com + +# Abstract + +Deep neural networks are widely used in various computer vision tasks, but their vulnerability to adversarial perturbations remains a significant challenge for reliable decision-making. Adversarial purification, a test-time defense strategy, has shown potential in countering these threats by removing noise through diffusion models. This plug-and-play method, using off-the-shelf models, appears highly effective. However, the purified data from diffusion often deviates more from the original data than the adversarial examples, leading to missing critical information and causing misclassification. In this study, we propose that upsampling with Super-Resolution (SR), followed by downsampling, can also aid in eliminating adversarial noise, similar to the noise addition and removal process in diffusion models. While SR alone is not as effective as the diffusion process, it better restores the original features typically associated with the early layers of networks. By combining SR, which initially mitigates damage to early-layer information from adversarial attacks, with diffusion, we observe a synergistic effect, leading to enhanced performance over diffusion models alone. Our comprehensive evaluations demonstrate that this combined approach, PuriFlow, significantly improves accuracy and robustness, working synergistically with state-of-the-art methods. + +# 1. Introduction + +Deep Neural Networks (DNNs) have been widely accepted in various computer vision tasks. However, their susceptibility to small, human-imperceptible noises can lead to untrustworthy decision-making. In particular, the subtle, maliciously manipulated noises, known as adversarial perturbations, produce fatal images called adversarial examples [10]. This intentional threat appears across numerous + +![](images/fdeb382978ed846143fa244c75867cea893c2e77b8031bfb21122b14ab67877d.jpg) +Figure 1. The overview of PuriFlow, which begins with restoring features of adversarial examples using the SR and drifting downsampled SR images within the forward diffusion. The perturbed images are then denoised in the reverse diffusion for classification. + +types of attacks [3, 9, 13, 34], leaving many DNNs vulnerable and posing substantial challenges to their deployment in AI-powered systems. + +To counter the adversarial threats, adversarial purification [38, 58] has emerged as a test-time defense strategy. This approach is powered by leveraging off-the-shelf diffusion models. At its core, noise removal through stochastic denoising effectively eliminates artifacts introduced by various attacks. Unlike adversarial training and certified methods [27, 30, 34], which require prior knowledge of target attacks, this simple application of generative power excels in protecting classifier without the need for training, pointing to a new direction. + +While diffusion models appear effective in purification, a key problem remains widely discussed in studies [5, 11, 32, + +![](images/1be3d78e3114267a8b8b170e1df928f79d951efe6e2d9274d47c3b657811f079.jpg) +Figure 2. The example of purification on the clean images. The top row exhibits the ground truth images. The second and third rows depict purified images obtained using diffusion and our proposed PuriFlow. The fourth row visualizes Grad-CAM [44] as represented by ResNet-50. Compared to Diffpure, PuriFlow enhances visual fidelity and accuracy by preserving DNN's activation. + +38, 60]: how to lower the upper bound of differences between purified and original data, which can exceed those caused by adversarial attacks. This concern also arises when clean data is input, as shown in Figure 2. However, resolving it solely within the diffusion model is not straightforward. The noise predictor in off-the-shelf diffusion models is optimized to predict only the mean under isotropic covariance [26], limiting its ability to fully restore subtle features tied to true data covariates. To improve the restoration of details, diffusion models may require retraining with a proper objective [37], which may not align well with test-time defenses. Recently, guidance during noise removal has been proposed for test-time defenses [5]; however, this approach, despite allowing extensive processing time, is limited to large-scale datasets, impacting its generality. + +To improve purification without retraining, we note that adversarial attacks are designed to maximize changes in model output with minimal alterations to the input. This + +results in alterations in the feature map that can be substantial, while changes in the image might be minor. Diffusion-based purification processes, which lack prior knowledge of the original image, risk deviating in the wrong direction during the noise addition and recovery phases, especially if the starting image contains feature information significantly distorted by adversarial noises. Therefore, it is crucial to eliminate as much distorted information as possible before applying the diffusion process. + +We hypothesize that using Super-Resolution (SR) to upsample and then downsample an image could help reduce noise. Our tests confirm that while SR does reduce adversarial noise, it is not as effective as diffusion models. However, an in-depth experiment reveals that SR performs better than diffusion models in restoring features to the early layers of a network. This insight suggests that initially, using SR to reduce adversarial noise before applying diffusion could enhance overall performance by restoring the features the diffusion process alone cannot recover. + +This paper proposes PuriFlow, an enhanced purification flow that leverages the synergy of SR and the diffusion process. As shown in Figure 1, PuriFlow integrates SR before the forward diffusion process to restore distorted features of adversarial images early on. By boosting content similarity [28], SR generates enhanced samples with features closer to the originals, even when mapped into a high-dimensional space. Our findings show that content proximity increases when SR images are resized back to their original dimensions, positioning resized SR images in regions with a higher probability for correct labels. In other words, resizing SR effectively reduces the initial cross-entropy on adversarial examples, allowing better diffusion. Our theoretical analysis and comprehensive evaluations demonstrate that this initialization aids in restoring features by diffusion, further reduces cross-entropy, and improves accuracy. + +Our contributions can be summarized as three-fold: + +- We find the purification effectiveness of the SR process. While SR alone is less effective in fully eliminating adversarial noises, it excels at preserving original features in the early layers of the network. +- Our approach, PuriFlow, finds a synergistic effect by integrating SR with diffusion. Our insight into geometrical analysis and empirical study finds that this integration can reduce cross-entropy in adversarial examples. +- Our pipeline utilizes only off-the-shelf SR and diffusion models. With negligible execution time required for SR, this plug-and-play method demonstrates improved accuracy and robustness in comprehensive evaluations. + +# 2. Related Work + +Denoising defenses. Traditional test-time adversarial defenses often apply denoising techniques to input images. JPEG compression has been employed to remove high + +frequency noise [14, 16], but Guo et al. [22] found it insufficient, proposing total variance minimization with image quilting as an alternative. Likewise, Prakash et al. [40] use wavelet denoising combined with BayesShrink to mitigate artifacts leading to misclassification, which surpasses total variance minimization and Wiener filtering in adversarial defenses [54]. Xie et al. [55] denoise feature maps through non-local means, supporting adversarial training. While practical, these methods often distort image details and face a core challenge: the inverse problem, which results in non-unique solutions. + +SR defenses. Modern SR models have gained attention to enhance perceptual quality in adversarial examples [36]. Recently, Bhardwaj et al. [7] raised concerns about the computational demands of the SR models, which can limit real-time applicability. They demonstrate the effectiveness of a highly efficient SR model [8], showing robust performance even on compact neural processing units. PuriFlow shares a similar purpose on rapid SR applications but adds weight to its synergistic use for restoring refined features with minimal overhead, assisting diffusion to enhance defense. + +Adversarial purification. Purifying adversarial noise using generative models has shown promise in strengthening classifier robustness as a test-time defense. PixelDefend [47] utilizes an autoregressive generative model, while Defense-GAN [43] relies on GANs. Langevin Dynamics (LD) sampling has also been effective for defenses through energy-based models [15, 21, 25]. Yoon et al. [58] introduce a denoising score-based model with an LD variant for denoising processes. DiffPure [38] and Lee and Kim [32] employ forward and reverse-time Stochastic Differential Equations (SDEs) [26, 46] for purification, but the loss of class information remains a risk due to the dependence on diffusion time steps. Bai et al. [5] mitigate this problem with contrastive guidance, while ScoreOpt [60] uses a one-shot denoiser. This work attempts to tackle losing crucial information in purified images by combining SR with diffusion. + +# 3. Background for Diffusion Models + +The score-based diffusion model is widely accepted in adversarial purification. It performs with bidirectional processes: forward and reverse time transition of $\mathbf{x}_t\in \mathbb{R}^d$ over the time interval $t\in [0,T]$ . Starting from an original sample $\mathbf{x}_0\sim p_0(\mathbf{x})$ where $p_0(\mathbf{x})$ is the unknown, true data distribution, the diffusion model progressively transforms $p_0(\mathbf{x})$ into a nearly spherical Gaussian distribution $p_T(\mathbf{x})\approx \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ . This converted distribution $p_T(\mathbf{x})$ retains no information about $p_0(\mathbf{x})$ . + +This forward process of $\{\mathbf{x}_t\}_{t = 0}^T$ , as defined by Ito SDE, is given with the positive increment of $t$ .. + +$$ +d \mathbf {x} = f (\mathbf {x}, t) d t + g (t) d \mathbf {w}, \tag {1} +$$ + +where the initial data $\mathbf{x}_0\sim p_0(\mathbf{x}),f(\cdot ,t):\mathbb{R}^d\mapsto \mathbb{R}^d$ is the drift coefficient of $\mathbf{x}_t$ and $g(\cdot):\mathbb{R}\mapsto \mathbb{R}$ is the diffusion coefficient tied with Brownian motion $\mathbf{w}\in \mathbb{R}^d$ . Here, when $f(\mathbf{x},t)$ is affine, the transition kernel always becomes a Gaussian distribution $p_{t^{\prime}t}(\mathbf{x}_t|\mathbf{x}_{t^{\prime}})$ where $0\leq t^{\prime} < t\leq T$ [49]. Thus, the forward diffusion process allows for a direct transition with a closed form using the Gaussian kernel $p_{0t}(\mathbf{x}_t|\mathbf{x}_0)$ from $t = 0$ to a certain $t$ , which prevents the need for neural network estimations. + +Given a perturbed sample $\mathbf{x}_T\sim p_T(\mathbf{x})$ , the reverse-time SDE defines that it drifts backward in the time steps: + +$$ +d \mathbf {x} = [ f (\mathbf {x}, t) - g ^ {2} (t) \nabla_ {\mathbf {x}} \log p _ {t} (\mathbf {x}) ] d t + g (t) d \bar {\mathbf {w}}, \quad (2) +$$ + +where $\bar{\mathbf{w}}$ represents a standard Wiener process and the infinitesimal time step $dt$ is negative. A key component in Eq. 2 is the score function $\nabla_{\mathbf{x}}\log p_t(\mathbf{x})$ , which offers how the density $p_t(\mathbf{x})$ changes with respect to $\mathbf{x}$ . Instead of directly knowing this function, one approach is to train a neural network, $\mathbf{s}_{\theta}(\mathbf{x},t)$ , to estimate it. This training is achieved by solving a denoising score-matching [51] problem: + +$$ +\ell (\theta) = \sum_ {t = 1} ^ {T} \lambda (t) \mathbb {E} _ {\mathbf {x}, \tilde {\mathbf {x}} | \mathbf {x}} [ \mathbf {s} _ {\theta} (\tilde {\mathbf {x}}, t) - \nabla_ {\tilde {\mathbf {x}}} \log p _ {0 t} (\tilde {\mathbf {x}} | \mathbf {x}) | | _ {2} ^ {2} ]. \tag {3} +$$ + +Here, Gaussian kernel $p_{0t}(\tilde{\mathbf{x}}|\mathbf{x})$ depicts the transition from $\tilde{\mathbf{x}} = \mathbf{x}_t$ to $\mathbf{x} = \mathbf{x}_0$ and $\lambda(t)$ is weighting function ensuring that the contributions at different times are appropriately balanced. This optimized model serves as $\mathbf{s}_{\theta^*}(\mathbf{x}, t)$ where $\theta^* = \mathrm{argmin}_{\theta} \ell(\theta)$ to replace the elusive score function, enabling an effective reverse diffusion process. + +# 4. Methodology + +In this section, we introduce our motivation, design methodology, and geometrical analysis of the proposed approach, PuriFlow. To begin, we briefly review adversarial attacks. + +Adversarial attack. Given a clean image $\mathbf{x}_0$ with label $y$ , adversarial attacks generally seek perturbations to mislead a classifier $\mathcal{F}_{\phi}(\cdot):\mathbb{R}^{d}\mapsto \mathbb{R}^{K}$ into incorrectly classifying $\mathbf{x}_0$ . This manipulated image $\mathbf{x}_0^a$ is generated by maximizing cross-entropy loss within a tiny region around $\mathbf{x}_0$ : + +$$ +\mathbf {x} _ {0} ^ {a} = \lim _ {i \rightarrow N} \prod_ {\mathcal {B} \left(\mathbf {x} _ {0} ^ {i}, \epsilon\right)} \left(\mathbf {x} _ {0} ^ {i} + \mu \operatorname {s i g n} \left(\nabla_ {\mathbf {x} _ {0} ^ {i}} \ell_ {c e} (\phi ; \mathbf {x} _ {0} ^ {i}, y)\right)\right), \tag {4} +$$ + +where $\mathbf{x}_0^0$ is the initial clean image $\mathbf{x}_0$ and $N$ is the number of attack iterations. $\mathcal{B}(\mathbf{x}_0^i,\epsilon)$ is the $\ell_p$ -norm ball around $\mathbf{x}_0^i$ within a radius of $\epsilon$ , $\Pi (\cdot)$ is the projection to the norm ball, and $\mu$ denotes the step size. The gradient can be obtained from surrogate networks when $\mathcal{F}_{\phi}(\cdot)$ is unknown. + +# 4.1. PuriFlow: Enhanced Purification Flow + +Given an adversarial example, $\mathbf{x}_0^a$ , VP-SDE [48] is widely accepted approach for purification. VP-SDE controls noise + +![](images/064cdc7a169202867996906cef60368462d4c9ad82f0b919aae52962c1658932.jpg) +Figure 3. Top: visualization of different image types. AE denotes the adversarial example generated by a PGD attack on ResNet-50 using an ImageNet-1k image. SR indicates the downsampled high-resolution image. Diffusion and PuriFlow use the same diffusion method, VP-SDE [38]. Bottom: corresponding feature maps from the Conv2_2 layer of VGG-19, as described in [18]. + +addition through drift coefficients $f(\mathbf{x}, t) = -\frac{1}{2} \beta_t \mathbf{x}$ and $g(t) = \sqrt{\beta_t}$ , where $\beta_t = \frac{1}{t} \beta_{\text{max}} + (1 - \frac{1}{t}) \beta_{\text{min}}$ [38]. This model then diffuses the data by solving the forward-time SDE in Eq. 1 with a closed form: + +$$ +\mathbf {x} _ {t} ^ {a} = \sqrt {\bar {\alpha} _ {t}} \mathbf {x} _ {0} ^ {a} + \sqrt {1 - \bar {\alpha} _ {t}} \mathbf {z} _ {t}, \tag {5} +$$ + +where $\mathbf{z}_t\sim \mathcal{N}(\mathbf{0},\mathbf{I}_d)$ and $\bar{\alpha}_{t} = e^{-\int_{0}^{t}\beta_{t}(s)ds}$ . + +The corresponding diffusion process is then performed by solving the reverse-time SDE in Eq. 2: + +$$ +d \mathbf {x} = \left[ f \left(\mathbf {x} _ {t} ^ {a}, t\right) - g ^ {2} (t) \nabla_ {\mathbf {x} _ {t} ^ {a}} \log p _ {t} \left(\mathbf {x} _ {t} ^ {a}\right) \right] d t + g (t) d \bar {\mathbf {w}}, \tag {6} +$$ + +where the drift coefficients, $f(\cdot ,t)$ and $g^{2}(t)$ , follows [5, 48]. + +During this reverse process, the crucial step is computing the score, defined as $\nabla_{\mathbf{x}_t^a}\log p_t(\mathbf{x}_t^a)$ . Here, we raise a key question: Can the neural network $s_{\theta^*}(\cdot ,t)$ , trained for score estimation, precisely predict $\mathbf{x}_0$ from $\mathbf{x}_t^a$ ? Even if $s_{\theta^*}(\cdot ,t)$ is well-trained, accurately estimating the direction toward $\mathbf{x}_0$ is not straightforward when dealing with the noised data $\mathbf{x}_t^a$ that has been perturbed from $\mathbf{x}_0^a$ , since $\mathbf{s}_{\theta^*}(\cdot ,t)$ has not been exposed to such perturbed data during training with Eq. 3. In other words, the trained neural network lacks prior knowledge of guiding $\mathbf{x}_t^a$ , risking deviations in the wrong direction during the noise addition and recovery phases. + +This concern is seemingly explicit. Figure 3 shows that the perturbations introduced by Eq. 4 appear subtle in pixel values but cause significant damage to the underlying features of $\mathbf{x}_0$ . These observations indicate that diffusion alone struggles to fully reconstruct the original features. To address this, we seek an existing method that positions the perturbed data closer to its ground-truth image, aiming to reduce feature discrepancies early on and thus ensure a more precise projection of the score function. + +To this end, we focus on a model that can effectively reduce content distance [18, 28] to capture feature proximity. Reducing feature-level shifts that lead to increased cross + +![](images/08191acf1a96c1871cb1a916d5fd07bd4be85225b0cfd30317837edd1c0d88e4.jpg) +(a) Impact of SR compared to diffusion with increments of $t^\prime$ + +![](images/98fac25418ba935d5814b383779760961859ec93d8a6b76311131600a9ae5c5a.jpg) +(b) Synergistic effect in feature restoration of SR combined with diffusion. Diffusion and SR+Diffusion use the same diffusion time $t^{\prime} = 90$ +Figure 4. Normalized content distance from the ground truth for each image type, measured across specific CONV layers of VGG-19 as shown in Figure 5. Real values are provided in the Appendix. + +entropy aligns with the goal of purification, ultimately improving classification accuracy. This approach also benefits clean data in preserving features. In this context, the SR model suits this purpose. The SR model, $\mathcal{H}_{\omega^{*}}(\cdot):\mathbb{R}^{d}\mapsto \mathbb{R}^{D}$ such that $D > d$ , is designed to upsample and interpolate low-resolution images into high-resolution counterparts. To optimize its parameters $\omega$ , various SR methods [1, 28, 31, 45, 52] minimize both pixel distance and VGG loss [31], which captures mutual similarity in feature maps across different layers of the VGG model: + +$$ +\omega^ {*} = \min _ {\omega} \frac {1}{W ^ {l} H ^ {l}} \left(\sum_ {u = 1} ^ {W ^ {l}} \sum_ {v = 1} ^ {H ^ {l}} \left[ \Phi^ {l} (\mathbf {X}) _ {u, v} - \Phi^ {l} \left(\mathcal {H} _ {\omega} (\mathbf {x})\right) _ {u, v} \right] ^ {2}\right), \tag {7} +$$ + +where $\mathbf{X} \in \mathbb{R}^{D}$ and $\mathbf{x} \in \mathbb{R}^{d}$ , with $W^{l}$ and $H^{l}$ denoting the dimensions of extracted features $\Phi^l (\cdot)$ in the $l$ -th convolution layer. Thus, SR models can naturally reduce the content distance by enhancing feature similarity while estimating unknown pixels and their covariates in $\mathbb{R}^{D}$ . + +Our findings, as illustrated in Figure 4, show that SR images downsampled from $\mathbb{R}^D$ to $\mathbb{R}^d$ can also reduce the content distance. This reduction covers all layers of the VGG network, demonstrating that even a one-shot application can effectively mitigate feature discrepancies present in adversarial examples. Notably, as observed in Figure 4a, increasing $t'$ when applying diffusion alone cannot sufficiently aid in restoring features tied to the early layers of the CNN. This observation suggests that a single application of SR can be more effective than diffusion in recovering detailed features [59] captured by the early layers. + +In this paper, we propose an enhanced purification flow, + +![](images/65d3259a8c6b94a5d010d531599c38fa0d74012a26969bd13341c5140fbfe2cb.jpg) +Figure 5. Transition of diffusion only and SR-integrated diffusion with the same diffusion time of $t' = 90$ under a PGD attack ( $\epsilon = 8/255$ , $N = 100$ ) on ResNet-50 using ImageNet-1k. "SR" represents downsampled high-dimensional images, while "AE" denotes adversarial examples. Content distance [18] from the ground truth is measured in two convolutional layers of VGG-19 and plotted with cross-entropy. + +![](images/5ed4042115827e922264d25cff9871e1ccf8d7e09d9ab966ba79a7deccae5439.jpg) + +![](images/05214416f75d2afea1bf3104c9aa99c2fc564a11fca8b0576facfb1b8709b773.jpg) + +![](images/fc21e9c162230ce689e6692bc018e2ffd9a7ce85c7a50a0f8bd5a2022731c7cb.jpg) + +PuriFlow, where SR works before the start of the forward-time SDE. As shown in Figure 4b, this approach demonstrates the effectiveness of using SR early before solving both SDEs. By correcting feature distortions that diffusion alone struggles to recover, SR allows for a synergistic effect, enhancing feature restoration across all layers through diffusion. As observed in Figure 5, the early reduction in content distance leads to a notable decrease in cross-entropy. Thus, the subsequent diffusion aligns the data more closely with the ground truth. To delve into this synergistic effect, we provide a theoretical study from a manifold perspective. We examine how SR influences data positioning and allows for more precise use of the score function within diffusion. + +# 4.2. Geometrical Analysis on PuriFlow + +Suppose $\mathcal{M} \subset \mathbb{R}^d$ is the set of all data points, i.e., $\forall \mathbf{x} \in \mathcal{M}$ , defining $\mathcal{M}$ as the data manifold. That is, the data distribution $p_0$ is uniform on the data manifold $\mathcal{M}$ . Then, within a local neighborhood, the manifold coincides with its tangent space of dimension $k \ll d$ , expressed as follows: + +$$ +\mathcal {M} \cap \mathcal {B} \left(\mathbf {x} _ {0}, d r\right) = T _ {\mathbf {x} _ {0}} \mathcal {M} \cap \mathcal {B} \left(\mathbf {x} _ {0}, d r\right), \quad T _ {\mathbf {x} _ {0}} \mathcal {M} \simeq \mathbb {R} ^ {k}, \tag {8} +$$ + +where $T_{\mathbf{x}_0}\mathcal{M}$ denotes the tangent space to the manifold $\mathcal{M}$ at $\mathbf{x}_0$ and $dr$ is an infinitesimal radius within which the manifold can be locally approximated by its tangent space. Here, we take note of a property of the trained score function in diffusion models [12]. + +Remark 1. The noisy data become increasingly concentrated on spherical manifolds as diffusion time $t$ increases. The score function is trained via denoising score matching (Eq. 3), using data on $\mathcal{M}$ . Let $\mathcal{P}_t(\cdot)$ denote a mapping from noisy data $\pmb{x}_t$ to its estimated clean counterpart $\hat{\pmb{x}}_0$ : + +$$ +\mathcal {P} _ {t} (\cdot): \boldsymbol {x} _ {t} \mapsto \hat {\boldsymbol {x}} _ {0} = - \frac {1}{2} (1 - \alpha_ {t}) \left(\boldsymbol {x} _ {t} + 2 \boldsymbol {s} _ {\theta^ {*}} \left(\boldsymbol {x} _ {t}, t\right)\right), \tag {9} +$$ + +where $\mathcal{P}_t(\pmb{x}_t) \in \mathcal{M}$ and it satisfies $\pmb{J}_{\mathcal{P}_t}^2 = \pmb{J}_{\mathcal{P}_t} = \pmb{J}_{\mathcal{P}_t}^T : \mathbb{R}^d \to T_{\mathcal{P}_t(\pmb{x}_t)}\mathcal{M}$ . That is, $\mathcal{P}_t(\cdot)$ acts as an orthonormal projection onto $\mathcal{M}$ . + +Remark 2. The score function is trained solely on the data concentrated on both the noisy and data manifolds. Therefore, applying it to data outside these manifolds may lead to inaccurate inference. + +Both premises suggest that when the score function is applied to off-manifold data, such as adversarial examples [7], its inaccurate inference may lead to misaligned orthonormal projections onto an unintended tangent space. Thus, the off-manifold data should be corrected to better align with its underlying manifold. Now, we show that the applying SR followed by downsampling can typically position off-manifold data, degraded from $\mathbf{x}_0$ , closer to the local manifold, which is the tangent space of $\mathbf{x}_0$ . + +Proposition 1. Suppose $\mathcal{H}_{\omega^*}$ is an ideal SR and smooth continuity, and let $\mathcal{R}$ be a linear downsampler. Given an offmanifold data $\mathfrak{D}(\pmb{x}_0)$ degraded from $\pmb{x}_0$ , the processed data $\mathcal{R}\mathcal{H}_{\omega^*}(\mathfrak{D}(\pmb{x}_0))$ is projected near the tangent space $T_{\pmb{x}_0}\mathcal{M}$ of the data manifold $\mathcal{M}$ at $\pmb{x}_0$ with a tiny $\delta > 0$ : + +$$ +\left\| \mathcal {R H} _ {\omega^ {*}} \left(\mathfrak {D} \left(\boldsymbol {x} _ {0}\right)\right) - \Pi_ {T _ {x _ {0}} \mathcal {M}} \left(\mathfrak {D} \left(\boldsymbol {x} _ {0}\right)\right) \right\| _ {F} < \delta . \tag {10} +$$ + +In Section 5, we evaluate the effectiveness of using the score function more accurately, supported by Proposition 1, against various adversarial attacks. + +# 5. Experiment + +We evaluate the performance of PuriFlow in two configurations: integrating SR with solving SDEs [38] and combining SR with a recent alternative, iterative One-Shot Denoising (OSD) [60]. OSD is defined by Tweedie's formula [50], expressed as $\mathbb{E}[\mathbf{x}|\tilde{\mathbf{x}}] = \tilde{\mathbf{x}} +\sigma^2\nabla_{\tilde{\mathbf{x}}}\log p_\sigma (\tilde{\mathbf{x}})$ , where $\tilde{\mathbf{x}}$ denotes perturbed data and $p_{\sigma}(\tilde{\mathbf{x}})$ represents marginal Gaussian perturbation of $p(\mathbf{x})$ . This formulation avoids repeated noise removal steps for varying time $t$ . + +In this context, the denoising score-matching problem defined by Eq. 3 optimizes the neural network $\mathbf{s}_{\theta^{*}}(\cdot ,t)$ given $\mathbf{x}_t$ to predict a fully denoised $\mathbf{x}_0$ per a specific time $t$ . Thus, Tweedie's formula can be represented as: + +$$ +\mathbb {E} \left[ \mathbf {x} _ {0} \mid \mathbf {x} _ {t ^ {\prime}} \right] = \mathbf {x} _ {t ^ {\prime}} + \sigma^ {2} \nabla_ {\mathbf {x} _ {t ^ {\prime}}} \log p _ {\sigma} \left(\mathbf {x} _ {t ^ {\prime}}\right). \tag {11} +$$ + +
MethodCriterionStandard (%)Robust (%)
Yang et al. [56]Gibbs Update94.8040.80
Song et al. [47]Mask+Recon.95.009.00
Hill et al. [24]EBM+LD84.1254.90
Yoon et al. [58]DSM+LD*86.1470.01
Nie et al. [38]SDEs89.0281.40
Bai et al. [5]Guide+SDEs92.6181.94
PuriFlow([38])SR+SDEs90.0683.00
Zhang et al. [60]iOSD93.7588.08
PuriFlow([60])SR+iOSD94.1491.01
+ +Table 1. Comparison of different adversarial purification methods against BPDA+EOT attacks with $\ell_{\infty}$ perturbations. Our evaluations use WRN-28-10 on CIFAR-10, with consistent experimental settings as outlined in [38, 60] for $\epsilon = 8 / 255$ . (* This purification utilizes a variant of the LD sampling.) + +The alternative proposed by [60] involves iteratively applying OSD using Eq. 11 during noise removal. Therefore, it can serve as a suitable baseline to confirm synergies with SR. We assess these configurations, detailed as algorithms in the Appendix, across key test-time defense scenarios, including adaptive white-box, preprocessor-blind [58], and black-box attacks. This section also includes an ablation study to demonstrate the impact of incorporating SR. Detailed experimental settings and additional evaluations for certified robustness are provided in the Appendix. For adaptive attacks, we adopt MDSR [33] for efficiency and ESRGAN [52] otherwise. We denote $t'$ on a discrete scale [26] from 1 to 1000 to represent the continuous interval [0, 1]. + +# 5.1. Defense on Adaptive Attacks + +In this section, we evaluate the performance of PuriFlow against four strong adaptive attacks: BPDA+EOT [3, 4], an adaptive ensemble white-box attack; AutoAttack [13], an adaptive white-box attack; PGD+EOT [32], and DiffAttack [29], a diffusion-based purification targeted attack. Here, the evaluations focus on projections within an $\ell_{\infty}$ -norm ball, as the projection in $\ell_{2}$ -norm ball is less effective. + +BPDA+EOT. Backward Pass Differentiable Approximation (BPDA) combined with Expectation over Transformation (EOT) is used to evaluate randomized adversarial purification methods. Table 1 demonstrates that PuriFlow outperforms standalone diffusion on CIFAR-10 with WRN-28-10, achieving a more balanced standard and robust accuracy than the guiding method. In addition, integrating SR with iterative OSD further enhances performance, emphasizing its critical role in diffusion-based purification. + +AutoAttack. Tables 2 demonstrates the effectiveness of PuriFlow against AutoAttack, on ImageNet-1k. PuriFlow outperforms adversarial training methods and a diffusion-only approach across ResNet-50 and DeiT-S. Compared to diffusion only, it achieves notable robust accuracy improve + +
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5631.06
Wong et al. [53]55.6226.95
Salman et al. [42]64.0237.89
Bai et al. [6]67.3835.51
APNie et al. [38]67.7940.93
PuriFlow([38])71.6852.92
DeiT-S79.900.00
ATBai et al. [6]66.5035.50
APNie et al. [38]73.6343.18
PuriFlow([38])75.6953.12
+ +Table 2. Evaluation comparing adversarial training (AT) and purification (AP) methods against AutoAttack with $\ell_{\infty}$ perturbations at $\epsilon = 4 / 255$ on ImageNet-1k. Diffusion time is set to $t' = 150$ for [38] as specified its study and $t' = 130$ for PuriFlow. + +
TypeMethodStandard (%)Robust (%)
WRN-28-1095.630.00
ATPang et al. [39]88.6264.95
Gowal et al. [19]88.5465.93
Gowal et al. [20]87.5166.01
APYoon et al. [58]85.6633.48
Zhang et al. [60]*93.7559.37
PuriFlow([60])*94.1460.76
WRN-70-1695.790.00
ATGowal et al. [19]91.1068.66
Gowal et al. [20]88.7469.03
Rebuffi et al. [41]92.2269.97
APYoon et al. [58]86.7637.11
Zhang et al. [60]*95.1163.28
PuriFlow([60])*95.3164.86
+ +Table 3. Evaluation comparing adversarial training (AT) and purification (AP) methods against PGD+EOT within an $\ell_{\infty}$ -norm ball of radius $\epsilon = 8/255$ on CIFAR-10. Results are sourced from [32]. (* For a fair comparison, PuriFlow and [60] are directly evaluated on the same settings with $t' = 250$ , as stated in [60].) + +ments, including $11.99\%$ (ResNet-50) and $9.94\%$ (DeiT-S), and standard accuracy. PuriFlow also shows efficiency with reduced diffusion time $t'$ on this large-scale dataset. + +$\mathbf{PGD} + \mathbf{EOT}$ . Motivated by [32], we assess PuriFlow's robustness against $\mathrm{PGD} + \mathrm{EOT}$ that utilizes gradient signs within the entire randomized defense framework. Table 3 shows PuriFlow's strength on CIFAR-10, achieving the highest standard accuracy for WRN-28-10 and WRN-70-16. Despite the substantial impact of $\mathrm{PGD} + \mathrm{EOT}$ on randomized purification, PuriFlow achieves superior robust accuracy among purification methods, demonstrating a synergistic effect in which SR effectively acts as the initialization before starting the iterative one-shot denoising process. + +
TypeMethodStandard (%)Robust (%)
WRN-70-1695.790.00
APNie et al. [38]90.0745.31
PuriFlow([38])90.2159.47
ResNet-5076.130.00
APNie et al. [38]67.7928.13
PuriFlow([38])71.6843.75
+ +Table 4. Evaluation against DiffAttack, targeting WRN-70-16 on CIFAR-10 and ResNet-50 on ImageNet-1k. Each diffusion time $t'$ is the same as those used for other attacks on PuriFlow. + +
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5638.97
Wong et al. [53]55.6229.15
Salman et al. [42]64.0238.01
Bai et al. [6]67.3840.27
APNie et al. [38]74.5238.87
PuriFlow([38])74.0549.92
DeiT-S79.900.00
ATBai et al. [6]66.5040.32
APNie et al. [38]78.9434.59
PuriFlow([38])78.4449.85
+ +DiffAttack. DiffAttack neutralizes vanishing and exploding gradients, high memory costs, and increased randomness—natural defensive effects from lengthy $t'$ of diffusion-based purification. Table 4 demonstrates that while purification methods using $t' = 100$ for CIFAR-10 and $t' = 150$ for ImageNet-1k exhibit limitations in robustness by this intentional attack, integrating SR demonstrates a notable improvement by $14.16\%$ and $15.62\%$ . We hypothesize that SR's early enhancement of feature proximity helps mitigate diverging feature maps from the ground truth, even under the reconstruction-deviating impacts of DiffAttack. + +# 5.2. Defense on Preprocessor-Blind Attack + +We evaluate PuriFlow against the PGD ( $N = 100$ ) attack, targeting only the classifier on ImageNet-1k. Table 5 shows that PuriFlow, leveraging only off-the-shelf models, surpasses adversarial training methods in both standard and robust accuracy for ResNet-50 and DeiT-S. Notably, PuriFlow significantly outperforms diffusion only [38] at the same practical diffusion time $t' = 10$ , achieving comparable standard accuracy while enhancing robust accuracy by $11.05\%$ for ResNet-50 and $15.26\%$ for DeiT-S. This improvement signifies the synergistic importance of SR in adversarial purification utilizing diffusion models. + +Table 5. Comparison under a preprocessor-blind PGD attack targeting only the classifier, within an $\ell_{\infty}$ perturbations at $\epsilon = 4/255$ on ImageNet-1k. The time $t'$ for both AP methods is set as 10. + +
TypeMethodStandard (%)Robust (%)
ResNet-5076.139.25
APNie et al. [38]67.7962.88
PuriFlow [38]74.0068.36
+ +Table 6. Evaluation of ResNet-50 under a black-box attack using Square Attack ( $N = 5000$ ) within an $\ell_{\infty}$ -norm ball of radius $\epsilon = 4/255$ on ImageNet-1k. Diffusion time is set to $t' = 150$ for [38] as specified in their study and $t' = 10$ for PuriFlow. Evaluation settings follow those in [38]. + +
MethodTypeStandard (%)Robust (%)
ResNet-5076.130.00
WaveletDenoising71.1568.63
TVMDenoising57.8056.96
NL-meansDenoising64.3262.67
DRLN×2 [2]SR71.6668.72
EDSR×2 [33]SR71.7768.73
ESRGAN×2 [52]SR72.9768.75
MDSR×2 [33]SR73.3269.47
+ +Table 7. Comparison of SR with other denoising methods as an initialization approach for diffusion [38] ( $t' = 90$ ) under a PGD attack on ImageNet-1k. "TVM" refers to total variance minimization, while "NL-means" denotes the non-local means algorithm. + +# 5.3. Defense on Black-Box Attack + +Table 6 shows that PuriFlow, using $t' = 10$ approximately $6.67\%$ of the diffusion time required by diffusion-only method [38]—achieves notable improvements in both standard and robust accuracy for ResNet-50. Specifically, it shows enhancements of $6.21\%$ and $5.48\%$ in each accuracy on ImageNet-1k. These results highlight the synergy of initialization for diffusion, indicating better performance and greater efficiency in this black-box attack. + +# 5.4. Ablation Study + +Impact of SR compared to various candidates. Puri-Flow can be configured by combining various initialization methods. We test several SR models and traditional denoising techniques by selectively replacing each method used in diffusion-based purification. This evaluation, conducted with PGD targeting the classifier on ImageNet-1k, is shown in Table 7. While Wavelet denoising is practical for PGD, it stands less effective than SR techniques. Among SR models tested, MDSR $\times 2$ outperforms DRLN $\times 2$ , EDSR $\times 2$ , and ESRGAN $\times 2$ , achieving the highest standard and robust accuracies. Furthermore, we explore different upsample ratios $(\times 2, \times 4, \times 8)$ using ESRGAN. Table 10 indicates that $\times 2$ represents the best balance between efficiency and performance, making it the optimal choice for PuriFlow. + +
SRDiffusionClassifierStandard (%)Robust (%)Wall time (s /img)
✓76.1300.00≈0.00
✓✓73.6830.700.01
✓✓74.5238.371.08
✓✓✓74.0049.921.09
+ +
ModelTypeStandard (%)Robust (%)
ResNet-50VP-SDE72.9768.75
VE-SDE73.3966.51
DeiT-SVP-SDE77.2972.37
VE-SDE77.7269.33
+ +Table 8. Evaluation of PuriFlow under a preprocessor-blind PGD attack on ImageNet-1k selectively employing off-the-shelf models. Left: ResNet-50, ESRGAN×2, and VP-SDE are the classifier, SR model, and diffusion [38] model, respectively, with diffusion time $t' = 10$ . Right: Impact of diffusion models within PuriFlow, evaluated at diffusion time $t' = 90$ using ResNet-50 and DeiT-S on ImageNet-1k. + +
Off-the-shelf modelsESRGAN×2EDSR×2MDSR×2
SR (each column)0.010.01≈0.00
Diffusion [38]1.080.810.73
ResNet-50≈0.00≈0.00≈0.00
Total wall time (s / img)1.090.820.73
ConfigurationCriterionIterationsPurif. time
Diffusion [38]SDEst' = 1005.27
SR+DiffusionSR+SDEst' = 1000.005+5.27
ScoreOpt-O [60]iOSDM = 50.51
SR+ScoreOpt-OSR+iOSDM = 50.005+0.51
+ +Table 9. Top: Wall time for PuriFlow with different SR models. Bottom: Examples of purification times (s/img) for configurations using WRN-70-16 [38] and WRN-28-10 [60] on CIFAR-10. + +
RatioStandard (%)Robust (%)FLOPsParams
×871.26%66.80%1110G16.7M
×472.41%67.32%899G16.7M
×272.97%68.75%224G16.7M
+ +Table 10. Impact of SR upsampling ratios from ESRGAN $\times 2$ combined with diffusion [38] ( $t' = 90$ ) using ResNet-50 on ImageNet-1k. FLOPs represent computational complexity [23], and Params refer to the number of trainable parameters. + +Off-the-shelf model adoption and latency analysis. Table 8 shows the enhanced defensive capabilities achieved by integrating off-the-shelf models into PuriFlow's framework. The table highlights the individual and combined effects of SR and diffusion. When used together, SR and diffusion improve robust accuracy by $19.22\%$ and $11.55\%$ , respectively, illustrating their strong synergistic effect on adversarial examples. Although combining SR with diffusion slightly decreases standard accuracy compared to diffusion alone, the robustness gains make this trade-off useful, with a tolerable latency of $1.09~\mathrm{s / img}$ when executed on a single NVIDIA RTX 4090 GPU. Table 9 further confirms that integrating SR models adds minimal latency to the overall purification process. Specifically, SR's execution time is $1,054\times$ faster than the diffusion's, and its overhead is only $1/102$ compared to one-shot denoising methods. This result emphasizes SR's role as an effective and efficient initialization. + +Impact of diffusion types. We explore VE-SDE [46], which can act as an alternative to diffusion. Table 8 demon + +![](images/59dbb534d522936d3ff20b6b46c71b63ae0c4133cad62412958869e6f977e395.jpg) +Figure 6. Impact of iterative SR on cross-entropy and MSEs under a PGD $(N = 100)$ attack targeting ResNet-50 within an $\ell_{\infty}$ -norm ball $(\epsilon = 4 / 255)$ on ImageNet-1k. SR, SR', and SR" indicate one, two, and three rounds of SR using ESRGAN $\times 2$ , respectively. + +![](images/85d77e4784240ce9cfa162e1157267cbf7396a147b154f717fc79d492661890b.jpg) + +strates that VP-SDE slightly reduces standard accuracy (approximately $0.42\%$ for ResNet-50 and $0.43\%$ for DeiT-S) but enhances robust accuracy by $2.24\%$ and $3.04\%$ , respectively. Despite the minor drop in standard accuracy, integrating VP-SDE shows overall robustness, synergizing a balanced performance with SR. + +Iterative applications of SR without diffusion. Figure 6 shows the impact of repeated SR applications for adversarial examples relative to their originals, with changes in MSE and cross-entropy and their impact on standard and robust accuracy. Up to three rounds of SR demonstrate a trade-off: MSE increases while cross-entropy decreases and eventually saturates. Here, robust accuracy rises but converges while standard accuracy decreases. We conjecture that repeated SR may behave as seen in model inversion [35, 57]: Images are generated to minimize cross-entropy loss, but this steady direction moves them farther from the originals. + +# 6. Conclusion + +We introduce PuriFlow, a novel purification flow integrating Super-Resolution (SR) with diffusion. While SR alone is less effective at mitigating adversarial noise, its image upsampling and downsampling excel at restoring subtle features captured by early layers. Diffusion alone lacks this ability. However, their mixture synergistically reduces overall feature proximity and cross-entropy. PuriFlow outperforms leading methods in extensive evaluations, demonstrating efficiency for minimal SR overhead. + +# Acknowledgment + +This research was partly supported by the IITP(Institute of Information & Communications Technology Planning & Evaluation)-ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT)(IITP-2025-RS-2023-00258649, $50\%$ ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-00562437, $40\%$ ), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), $10\%$ ). + +# References + +[1] Yasin Almalioglu, Kutsev Bengisu Ozyoruk, Abdulkadir Gokce, Kagan Incetan, Guliz Irem Gokceler, Muhammed Ali Simsek, Kivanc Ararat, Richard J Chen, Nicholas J Durr, Faisal Mahmood, et al. Endol2h: deep super-resolution for capsule endoscopy. IEEE Transactions on Medical Imaging, 2020. 4 +[2] Saeed Anwar and Nick Barnes. Densely residual laplacian super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7 +[3] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018. 1, 6 +[4] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International Conference on Machine Learning, 2018. 6 +[5] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In International Conference on Machine Learning, 2024. 1, 2, 3, 4, 6 +[6] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 2021. 6, 7 +[7] Kartikeya Bhardwaj, Dibakar Gope, James Ward, Paul Whatmough, and Danny Loh. Super-efficient super resolution for fast adversarial defense at the edge. In 2022 Design, Automation & Test in Europe Conference & Exhibition, 2022. 3, 5 +[8] Kartikeya Bhardwaj, Milos Milosavljevic, Liam O'Neil, Dibakar Gope, Ramon Matas, Alex Chalfin, Naveen Suda, Lingchuan Meng, and Danny Loh. Collapsible linear blocks for super-efficient super resolution. Proceedings of Machine Learning and Systems, 2022. 3 +[9] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1 + +[10] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1 +[11] Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, and J Zico Kolter. (certified!!) adversarial robustness for free! In International Conference on Learning Representations, 2023. 1 +[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 2022. 5 +[13] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, 2020. 1, 6 +[14] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Li Chen, Michael E Kounavis, and Duen Horng Chau. Keeping the bad guys out: Protecting and vaccinating deep learning withJPEG compression. arXiv preprint arXiv:1705.02900, 2017.3 +[15] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems, 2019. 3 +[16] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 3 +[17] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. 6, 7 +[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 4, 5 +[19] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6 +[20] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 2021. 6 +[21] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3 +[22] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 3 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 8 + +[24] Mitch Hill, Jonathan Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. International Conference on Learning Representations, 2020. 6 +[25] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 3 +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3, 6 +[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 2020. 1 +[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, 2016. 2, 4 +[29] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. Advances in Neural Information Processing Systems, 2024. 6 +[30] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 1 +[31] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 4 +[32] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6 +[33] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6, 7 +[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1 +[35] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. Google Research Blog, 2015. 8 +[36] Aamir Mustafa, Salman H. Khan, Munawar Hayat, Jianbing Shen, and Ling Shao. Image super-resolution as a defense against adversarial attacks. IEEE Transactions on Image Processing, 2020. 3 +[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2 +[38] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, 2022. 1, 2, 3, 4, 5, 6, 7, 8 + +[39] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, 2022. 6 +[40] Aaditya Prakash, Nick Moran, Solomon Garber, Antonella DiLillo, and James Storer. Deflecting adversarial attacks with pixel deflection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 3 +[41] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6 +[42] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 2020. 6, 7 +[43] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 3 +[44] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2 +[45] Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo. Perceptual extreme super-resolution network with receptive field block. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 4 +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 8 +[47] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 3, 6 +[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 3, 4 +[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[50] Soobin Um, Suhyeon Lee, and Jong Chul Ye. Don't play favorites: Minority guidance for diffusion models. In International Conference on Learning Representations, 2024. 5 +[51] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 2011. 3 +[52] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In + +European Conference on Computer Vision Workshops, 2018. 4, 6, 7 +[53] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 6, 7 +[54] Fei Wu, Wenxue Yang, Limin Xiao, and Jinbin Zhu. Adaptive wiener filter and natural noise to eliminate adversarial perturbation. *Electronics*, 2020. 3 +[55] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 3 +[56] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, 2019. 6 +[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 8 +[58] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 1, 3, 6 +[59] MD Zeiler. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, 2014. 4 +[60] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 2023. 2, 3, 5, 6, 8 \ No newline at end of file diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/images.zip b/2025/Adversarial Purification via Super-Resolution and Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..113917b5e089c39fd531ae18cdd1dc74ef22e649 --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa1afa39b5bfc41199b563c3fcb3892cf85147cfdf1b909eea74ca382e1d2158 +size 805346 diff --git a/2025/Adversarial Purification via Super-Resolution and Diffusion/layout.json b/2025/Adversarial Purification via Super-Resolution and Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfbf4b8e9b49d2f63e8a402f80aa58355f4756ff --- /dev/null +++ b/2025/Adversarial Purification via Super-Resolution and Diffusion/layout.json @@ -0,0 +1,11530 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "Adversarial Purification via Super-Resolution and Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": "Mincheol Park" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Cheonjun Park" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Seungseop Lim" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Mijin Koo" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Hyunwuk Lee" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Won Woo Ro" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": ", Suhyun Kim" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{7*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Samsung Advanced Institute of Technology, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Samsung Electronics, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Yonsei University, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "AITRICS, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Hankuk University of Foreign Studies, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Seoul National University, " + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 64, + 171, + 547, + 200 + ], + "type": "text", + "content": "Kyung Hee University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 201, + 475, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 475, + 227 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 475, + 227 + ], + "type": "text", + "content": "{mincheol.park, cheonjun.park, hyunwuk.lee, wro}@yonsei.ac.kr, ss.lim@aitrics.com, starmj09@snu.ac.kr, dr.suhyun.kim@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "type": "text", + "content": "Deep neural networks are widely used in various computer vision tasks, but their vulnerability to adversarial perturbations remains a significant challenge for reliable decision-making. Adversarial purification, a test-time defense strategy, has shown potential in countering these threats by removing noise through diffusion models. This plug-and-play method, using off-the-shelf models, appears highly effective. However, the purified data from diffusion often deviates more from the original data than the adversarial examples, leading to missing critical information and causing misclassification. In this study, we propose that upsampling with Super-Resolution (SR), followed by downsampling, can also aid in eliminating adversarial noise, similar to the noise addition and removal process in diffusion models. While SR alone is not as effective as the diffusion process, it better restores the original features typically associated with the early layers of networks. By combining SR, which initially mitigates damage to early-layer information from adversarial attacks, with diffusion, we observe a synergistic effect, leading to enhanced performance over diffusion models alone. Our comprehensive evaluations demonstrate that this combined approach, PuriFlow, significantly improves accuracy and robustness, working synergistically with state-of-the-art methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 137, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "content": "Deep Neural Networks (DNNs) have been widely accepted in various computer vision tasks. However, their susceptibility to small, human-imperceptible noises can lead to untrustworthy decision-making. In particular, the subtle, maliciously manipulated noises, known as adversarial perturbations, produce fatal images called adversarial examples [10]. This intentional threat appears across numerous" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 256, + 553, + 462 + ], + "blocks": [ + { + "bbox": [ + 316, + 256, + 553, + 462 + ], + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 462 + ], + "type": "image", + "image_path": "fdeb382978ed846143fa244c75867cea893c2e77b8031bfb21122b14ab67877d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 466, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 510 + ], + "type": "text", + "content": "Figure 1. The overview of PuriFlow, which begins with restoring features of adversarial examples using the SR and drifting downsampled SR images within the forward diffusion. The perturbed images are then denoised in the reverse diffusion for classification." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 568 + ], + "type": "text", + "content": "types of attacks [3, 9, 13, 34], leaving many DNNs vulnerable and posing substantial challenges to their deployment in AI-powered systems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": "To counter the adversarial threats, adversarial purification [38, 58] has emerged as a test-time defense strategy. This approach is powered by leveraging off-the-shelf diffusion models. At its core, noise removal through stochastic denoising effectively eliminates artifacts introduced by various attacks. Unlike adversarial training and certified methods [27, 30, 34], which require prior knowledge of target attacks, this simple application of generative power excels in protecting classifier without the need for training, pointing to a new direction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "While diffusion models appear effective in purification, a key problem remains widely discussed in studies [5, 11, 32," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4605" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 298, + 393 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 298, + 393 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 298, + 393 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 298, + 393 + ], + "type": "image", + "image_path": "1be3d78e3114267a8b8b170e1df928f79d951efe6e2d9274d47c3b657811f079.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 401, + 297, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 297, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 297, + 468 + ], + "type": "text", + "content": "Figure 2. The example of purification on the clean images. The top row exhibits the ground truth images. The second and third rows depict purified images obtained using diffusion and our proposed PuriFlow. The fourth row visualizes Grad-CAM [44] as represented by ResNet-50. Compared to Diffpure, PuriFlow enhances visual fidelity and accuracy by preserving DNN's activation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 495, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 296, + 674 + ], + "type": "text", + "content": "38, 60]: how to lower the upper bound of differences between purified and original data, which can exceed those caused by adversarial attacks. This concern also arises when clean data is input, as shown in Figure 2. However, resolving it solely within the diffusion model is not straightforward. The noise predictor in off-the-shelf diffusion models is optimized to predict only the mean under isotropic covariance [26], limiting its ability to fully restore subtle features tied to true data covariates. To improve the restoration of details, diffusion models may require retraining with a proper objective [37], which may not align well with test-time defenses. Recently, guidance during noise removal has been proposed for test-time defenses [5]; however, this approach, despite allowing extensive processing time, is limited to large-scale datasets, impacting its generality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "To improve purification without retraining, we note that adversarial attacks are designed to maximize changes in model output with minimal alterations to the input. This" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "results in alterations in the feature map that can be substantial, while changes in the image might be minor. Diffusion-based purification processes, which lack prior knowledge of the original image, risk deviating in the wrong direction during the noise addition and recovery phases, especially if the starting image contains feature information significantly distorted by adversarial noises. Therefore, it is crucial to eliminate as much distorted information as possible before applying the diffusion process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 299 + ], + "type": "text", + "content": "We hypothesize that using Super-Resolution (SR) to upsample and then downsample an image could help reduce noise. Our tests confirm that while SR does reduce adversarial noise, it is not as effective as diffusion models. However, an in-depth experiment reveals that SR performs better than diffusion models in restoring features to the early layers of a network. This insight suggests that initially, using SR to reduce adversarial noise before applying diffusion could enhance overall performance by restoring the features the diffusion process alone cannot recover." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 491 + ], + "type": "text", + "content": "This paper proposes PuriFlow, an enhanced purification flow that leverages the synergy of SR and the diffusion process. As shown in Figure 1, PuriFlow integrates SR before the forward diffusion process to restore distorted features of adversarial images early on. By boosting content similarity [28], SR generates enhanced samples with features closer to the originals, even when mapped into a high-dimensional space. Our findings show that content proximity increases when SR images are resized back to their original dimensions, positioning resized SR images in regions with a higher probability for correct labels. In other words, resizing SR effectively reduces the initial cross-entropy on adversarial examples, allowing better diffusion. Our theoretical analysis and comprehensive evaluations demonstrate that this initialization aids in restoring features by diffusion, further reduces cross-entropy, and improves accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 491, + 534, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 491, + 534, + 502 + ], + "spans": [ + { + "bbox": [ + 326, + 491, + 534, + 502 + ], + "type": "text", + "content": "Our contributions can be summarized as three-fold:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 503, + 553, + 647 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 503, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 553, + 550 + ], + "type": "text", + "content": "- We find the purification effectiveness of the SR process. While SR alone is less effective in fully eliminating adversarial noises, it excels at preserving original features in the early layers of the network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 552, + 553, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 553, + 598 + ], + "type": "text", + "content": "- Our approach, PuriFlow, finds a synergistic effect by integrating SR with diffusion. Our insight into geometrical analysis and empirical study finds that this integration can reduce cross-entropy in adversarial examples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 599, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 599, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 599, + 553, + 647 + ], + "type": "text", + "content": "- Our pipeline utilizes only off-the-shelf SR and diffusion models. With negligible execution time required for SR, this plug-and-play method demonstrates improved accuracy and robustness in comprehensive evaluations." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 657, + 400, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 400, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 400, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Denoising defenses. Traditional test-time adversarial defenses often apply denoising techniques to input images. JPEG compression has been employed to remove high" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4606" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "frequency noise [14, 16], but Guo et al. [22] found it insufficient, proposing total variance minimization with image quilting as an alternative. Likewise, Prakash et al. [40] use wavelet denoising combined with BayesShrink to mitigate artifacts leading to misclassification, which surpasses total variance minimization and Wiener filtering in adversarial defenses [54]. Xie et al. [55] denoise feature maps through non-local means, supporting adversarial training. While practical, these methods often distort image details and face a core challenge: the inverse problem, which results in non-unique solutions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 213, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 333 + ], + "type": "text", + "content": "SR defenses. Modern SR models have gained attention to enhance perceptual quality in adversarial examples [36]. Recently, Bhardwaj et al. [7] raised concerns about the computational demands of the SR models, which can limit real-time applicability. They demonstrate the effectiveness of a highly efficient SR model [8], showing robust performance even on compact neural processing units. PuriFlow shares a similar purpose on rapid SR applications but adds weight to its synergistic use for restoring refined features with minimal overhead, assisting diffusion to enhance defense." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 341, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 533 + ], + "type": "text", + "content": "Adversarial purification. Purifying adversarial noise using generative models has shown promise in strengthening classifier robustness as a test-time defense. PixelDefend [47] utilizes an autoregressive generative model, while Defense-GAN [43] relies on GANs. Langevin Dynamics (LD) sampling has also been effective for defenses through energy-based models [15, 21, 25]. Yoon et al. [58] introduce a denoising score-based model with an LD variant for denoising processes. DiffPure [38] and Lee and Kim [32] employ forward and reverse-time Stochastic Differential Equations (SDEs) [26, 46] for purification, but the loss of class information remains a risk due to the dependence on diffusion time steps. Bai et al. [5] mitigate this problem with contrastive guidance, while ScoreOpt [60] uses a one-shot denoiser. This work attempts to tackle losing crucial information in purified images by combining SR with diffusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 542, + 243, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 243, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 243, + 555 + ], + "type": "text", + "content": "3. Background for Diffusion Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": "The score-based diffusion model is widely accepted in adversarial purification. It performs with bidirectional processes: forward and reverse time transition of " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": " over the time interval " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": ". Starting from an original sample " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim p_0(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "p_0(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": " is the unknown, true data distribution, the diffusion model progressively transforms " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "p_0(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": " into a nearly spherical Gaussian distribution " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "p_T(\\mathbf{x})\\approx \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": ". This converted distribution " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "p_T(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": " retains no information about " + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "inline_equation", + "content": "p_0(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 562, + 295, + 669 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": "This forward process of " + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t = 0}^T" + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": " , as defined by Ito SDE, is given with the positive increment of " + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 669, + 295, + 693 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 122, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} = f (\\mathbf {x}, t) d t + g (t) d \\mathbf {w}, \\tag {1}", + "image_path": "3a38ecea33d685b04d457ef5afab7ba6094ea87d6d4dd6b346cd36b16b3ab5c2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "where the initial data " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0\\sim p_0(\\mathbf{x}),f(\\cdot ,t):\\mathbb{R}^d\\mapsto \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is the drift coefficient of " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "g(\\cdot):\\mathbb{R}\\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is the diffusion coefficient tied with Brownian motion " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{w}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ". Here, when " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x},t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " is affine, the transition kernel always becomes a Gaussian distribution " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "p_{t^{\\prime}t}(\\mathbf{x}_t|\\mathbf{x}_{t^{\\prime}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "0\\leq t^{\\prime} < t\\leq T" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " [49]. Thus, the forward diffusion process allows for a direct transition with a closed form using the Gaussian kernel " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "p_{0t}(\\mathbf{x}_t|\\mathbf{x}_0)" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " to a certain " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", which prevents the need for neural network estimations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "content": "Given a perturbed sample " + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T\\sim p_T(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "content": " , the reverse-time SDE defines that it drifts backward in the time steps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 213, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 213, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 337, + 213, + 555, + 228 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} = [ f (\\mathbf {x}, t) - g ^ {2} (t) \\nabla_ {\\mathbf {x}} \\log p _ {t} (\\mathbf {x}) ] d t + g (t) d \\bar {\\mathbf {w}}, \\quad (2)", + "image_path": "e0233320f81f9e9a32a5a5c60ce251a8554ada9628b4678bfe824a52d405502f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{w}}" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": " represents a standard Wiener process and the infinitesimal time step " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "dt" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": " is negative. A key component in Eq. 2 is the score function " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}}\\log p_t(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": ", which offers how the density " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "p_t(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": " changes with respect to " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": ". Instead of directly knowing this function, one approach is to train a neural network, " + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta}(\\mathbf{x},t)" + }, + { + "bbox": [ + 313, + 236, + 555, + 320 + ], + "type": "text", + "content": ", to estimate it. This training is achieved by solving a denoising score-matching [51] problem:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 327, + 553, + 359 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 327, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 319, + 327, + 553, + 359 + ], + "type": "interline_equation", + "content": "\\ell (\\theta) = \\sum_ {t = 1} ^ {T} \\lambda (t) \\mathbb {E} _ {\\mathbf {x}, \\tilde {\\mathbf {x}} | \\mathbf {x}} [ \\mathbf {s} _ {\\theta} (\\tilde {\\mathbf {x}}, t) - \\nabla_ {\\tilde {\\mathbf {x}}} \\log p _ {0 t} (\\tilde {\\mathbf {x}} | \\mathbf {x}) | | _ {2} ^ {2} ]. \\tag {3}", + "image_path": "6f5aa37914034c159ef6735d3addc78b3ef7978b678003c398f29bb06652f345.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": "Here, Gaussian kernel " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "p_{0t}(\\tilde{\\mathbf{x}}|\\mathbf{x})" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " depicts the transition from " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}} = \\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\lambda(t)" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " is weighting function ensuring that the contributions at different times are appropriately balanced. This optimized model serves as " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta^*}(\\mathbf{x}, t)" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "inline_equation", + "content": "\\theta^* = \\mathrm{argmin}_{\\theta} \\ell(\\theta)" + }, + { + "bbox": [ + 313, + 359, + 554, + 431 + ], + "type": "text", + "content": " to replace the elusive score function, enabling an effective reverse diffusion process." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 397, + 455 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 498 + ], + "type": "text", + "content": "In this section, we introduce our motivation, design methodology, and geometrical analysis of the proposed approach, PuriFlow. To begin, we briefly review adversarial attacks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": "Adversarial attack. Given a clean image " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": " with label " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": ", adversarial attacks generally seek perturbations to mislead a classifier " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\phi}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{K}" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": " into incorrectly classifying " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": ". This manipulated image " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^a" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": " is generated by maximizing cross-entropy loss within a tiny region around " + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 506, + 554, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 575, + 555, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 575, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 321, + 575, + 555, + 596 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {0} ^ {a} = \\lim _ {i \\rightarrow N} \\prod_ {\\mathcal {B} \\left(\\mathbf {x} _ {0} ^ {i}, \\epsilon\\right)} \\left(\\mathbf {x} _ {0} ^ {i} + \\mu \\operatorname {s i g n} \\left(\\nabla_ {\\mathbf {x} _ {0} ^ {i}} \\ell_ {c e} (\\phi ; \\mathbf {x} _ {0} ^ {i}, y)\\right)\\right), \\tag {4}", + "image_path": "50b6d55ad67d80356197df9d3fbfdf3a57bb90e7a14791bdb8a6541b19850208.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^0" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is the initial clean image " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is the number of attack iterations. " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathbf{x}_0^i,\\epsilon)" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "-norm ball around " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " within a radius of " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\Pi (\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is the projection to the norm ball, and " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " denotes the step size. The gradient can be obtained from surrogate networks when " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\phi}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " is unknown." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 672, + 519, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 519, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 519, + 684 + ], + "type": "text", + "content": "4.1. PuriFlow: Enhanced Purification Flow" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Given an adversarial example, " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^a" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", VP-SDE [48] is widely accepted approach for purification. VP-SDE controls noise" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4607" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 181 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 181 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 181 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 181 + ], + "type": "image", + "image_path": "064cdc7a169202867996906cef60368462d4c9ad82f0b919aae52962c1658932.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 296, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 257 + ], + "type": "text", + "content": "Figure 3. Top: visualization of different image types. AE denotes the adversarial example generated by a PGD attack on ResNet-50 using an ImageNet-1k image. SR indicates the downsampled high-resolution image. Diffusion and PuriFlow use the same diffusion method, VP-SDE [38]. Bottom: corresponding feature maps from the Conv2_2 layer of VGG-19, as described in [18]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": "addition through drift coefficients " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}, t) = -\\frac{1}{2} \\beta_t \\mathbf{x}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "g(t) = \\sqrt{\\beta_t}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\beta_t = \\frac{1}{t} \\beta_{\\text{max}} + (1 - \\frac{1}{t}) \\beta_{\\text{min}}" + }, + { + "bbox": [ + 55, + 277, + 296, + 326 + ], + "type": "text", + "content": " [38]. This model then diffuses the data by solving the forward-time SDE in Eq. 1 with a closed form:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 335, + 295, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 335, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 118, + 335, + 295, + 348 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} ^ {a} = \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {x} _ {0} ^ {a} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\mathbf {z} _ {t}, \\tag {5}", + "image_path": "e9b0ae36a7b828dd933da4a48cb2e16c09b807553525ed43132e6ee4f5c9c8f9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I}_d)" + }, + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = e^{-\\int_{0}^{t}\\beta_{t}(s)ds}" + }, + { + "bbox": [ + 55, + 357, + 237, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 370, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 395 + ], + "type": "text", + "content": "The corresponding diffusion process is then performed by solving the reverse-time SDE in Eq. 2:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 403, + 294, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 403, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 294, + 418 + ], + "type": "interline_equation", + "content": "d \\mathbf {x} = \\left[ f \\left(\\mathbf {x} _ {t} ^ {a}, t\\right) - g ^ {2} (t) \\nabla_ {\\mathbf {x} _ {t} ^ {a}} \\log p _ {t} \\left(\\mathbf {x} _ {t} ^ {a}\\right) \\right] d t + g (t) d \\bar {\\mathbf {w}}, \\tag {6}", + "image_path": "c45fce56665935bd9fa0e463dd6bc091bd9ca26b21116eea1957d83721b12435.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "text", + "content": "where the drift coefficients, " + }, + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "inline_equation", + "content": "f(\\cdot ,t)" + }, + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "inline_equation", + "content": "g^{2}(t)" + }, + { + "bbox": [ + 55, + 426, + 295, + 438 + ], + "type": "text", + "content": ", follows [5, 48]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": "During this reverse process, the crucial step is computing the score, defined as " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathbf{x}_t^a}\\log p_t(\\mathbf{x}_t^a)" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": ". Here, we raise a key question: Can the neural network " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "s_{\\theta^*}(\\cdot ,t)" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": ", trained for score estimation, precisely predict " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^a" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": "? Even if " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "s_{\\theta^*}(\\cdot ,t)" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": " is well-trained, accurately estimating the direction toward " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": " is not straightforward when dealing with the noised data " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^a" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": " that has been perturbed from " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^a" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta^*}(\\cdot ,t)" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": " has not been exposed to such perturbed data during training with Eq. 3. In other words, the trained neural network lacks prior knowledge of guiding " + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t^a" + }, + { + "bbox": [ + 54, + 438, + 295, + 569 + ], + "type": "text", + "content": ", risking deviations in the wrong direction during the noise addition and recovery phases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": "This concern is seemingly explicit. Figure 3 shows that the perturbations introduced by Eq. 4 appear subtle in pixel values but cause significant damage to the underlying features of " + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 54, + 570, + 295, + 677 + ], + "type": "text", + "content": ". These observations indicate that diffusion alone struggles to fully reconstruct the original features. To address this, we seek an existing method that positions the perturbed data closer to its ground-truth image, aiming to reduce feature discrepancies early on and thus ensure a more precise projection of the score function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "To this end, we focus on a model that can effectively reduce content distance [18, 28] to capture feature proximity. Reducing feature-level shifts that lead to increased cross" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 153 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 153 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 153 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 153 + ], + "type": "image", + "image_path": "08191acf1a96c1871cb1a916d5fd07bd4be85225b0cfd30317837edd1c0d88e4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 156, + 534, + 167 + ], + "lines": [ + { + "bbox": [ + 334, + 156, + 534, + 167 + ], + "spans": [ + { + "bbox": [ + 334, + 156, + 534, + 167 + ], + "type": "text", + "content": "(a) Impact of SR compared to diffusion with increments of " + }, + { + "bbox": [ + 334, + 156, + 534, + 167 + ], + "type": "inline_equation", + "content": "t^\\prime" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 170, + 553, + 253 + ], + "blocks": [ + { + "bbox": [ + 317, + 170, + 553, + 253 + ], + "lines": [ + { + "bbox": [ + 317, + 170, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 170, + 553, + 253 + ], + "type": "image", + "image_path": "98fac25418ba935d5814b383779760961859ec93d8a6b76311131600a9ae5c5a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 255, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 274 + ], + "type": "text", + "content": "(b) Synergistic effect in feature restoration of SR combined with diffusion. Diffusion and SR+Diffusion use the same diffusion time " + }, + { + "bbox": [ + 314, + 255, + 553, + 274 + ], + "type": "inline_equation", + "content": "t^{\\prime} = 90" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 285, + 554, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 319 + ], + "type": "text", + "content": "Figure 4. Normalized content distance from the ground truth for each image type, measured across specific CONV layers of VGG-19 as shown in Figure 5. Real values are provided in the Appendix." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "text", + "content": "entropy aligns with the goal of purification, ultimately improving classification accuracy. This approach also benefits clean data in preserving features. In this context, the SR model suits this purpose. The SR model, " + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\omega^{*}}(\\cdot):\\mathbb{R}^{d}\\mapsto \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "inline_equation", + "content": "D > d" + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "text", + "content": " , is designed to upsample and interpolate low-resolution images into high-resolution counterparts. To optimize its parameters " + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 338, + 555, + 458 + ], + "type": "text", + "content": " , various SR methods [1, 28, 31, 45, 52] minimize both pixel distance and VGG loss [31], which captures mutual similarity in feature maps across different layers of the VGG model:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 466, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\omega^ {*} = \\min _ {\\omega} \\frac {1}{W ^ {l} H ^ {l}} \\left(\\sum_ {u = 1} ^ {W ^ {l}} \\sum_ {v = 1} ^ {H ^ {l}} \\left[ \\Phi^ {l} (\\mathbf {X}) _ {u, v} - \\Phi^ {l} \\left(\\mathcal {H} _ {\\omega} (\\mathbf {x})\\right) _ {u, v} \\right] ^ {2}\\right), \\tag {7}", + "image_path": "ac48d15d8b408d42deb6ae459c7a008e1e199c3713e330a18171e3a8d423abca.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "W^{l}" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "H^{l}" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": " denoting the dimensions of extracted features " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\Phi^l (\\cdot)" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": "-th convolution layer. Thus, SR models can naturally reduce the content distance by enhancing feature similarity while estimating unknown pixels and their covariates in " + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 510, + 554, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "text", + "content": "Our findings, as illustrated in Figure 4, show that SR images downsampled from " + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "text", + "content": " can also reduce the content distance. This reduction covers all layers of the VGG network, demonstrating that even a one-shot application can effectively mitigate feature discrepancies present in adversarial examples. Notably, as observed in Figure 4a, increasing " + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 570, + 555, + 701 + ], + "type": "text", + "content": " when applying diffusion alone cannot sufficiently aid in restoring features tied to the early layers of the CNN. This observation suggests that a single application of SR can be more effective than diffusion in recovering detailed features [59] captured by the early layers." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "In this paper, we propose an enhanced purification flow," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4608" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 181, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 181, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 181, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 181, + 175 + ], + "type": "image", + "image_path": "65d3259a8c6b94a5d010d531599c38fa0d74012a26969bd13341c5140fbfe2cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "text", + "content": "Figure 5. Transition of diffusion only and SR-integrated diffusion with the same diffusion time of " + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "inline_equation", + "content": "t' = 90" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "text", + "content": " under a PGD attack (" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 55, + 186, + 555, + 220 + ], + "type": "text", + "content": ") on ResNet-50 using ImageNet-1k. \"SR\" represents downsampled high-dimensional images, while \"AE\" denotes adversarial examples. Content distance [18] from the ground truth is measured in two convolutional layers of VGG-19 and plotted with cross-entropy." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 70, + 304, + 175 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 304, + 175 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 304, + 175 + ], + "type": "image", + "image_path": "5ed4042115827e922264d25cff9871e1ccf8d7e09d9ab966ba79a7deccae5439.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 429, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 429, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 429, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 429, + 175 + ], + "type": "image", + "image_path": "05214416f75d2afea1bf3104c9aa99c2fc564a11fca8b0576facfb1b8709b773.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 431, + 70, + 552, + 175 + ], + "blocks": [ + { + "bbox": [ + 431, + 70, + 552, + 175 + ], + "lines": [ + { + "bbox": [ + 431, + 70, + 552, + 175 + ], + "spans": [ + { + "bbox": [ + 431, + 70, + 552, + 175 + ], + "type": "image", + "image_path": "fc21e9c162230ce689e6692bc018e2ffd9a7ce85c7a50a0f8bd5a2022731c7cb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 240, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 240, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 296, + 396 + ], + "type": "text", + "content": "PuriFlow, where SR works before the start of the forward-time SDE. As shown in Figure 4b, this approach demonstrates the effectiveness of using SR early before solving both SDEs. By correcting feature distortions that diffusion alone struggles to recover, SR allows for a synergistic effect, enhancing feature restoration across all layers through diffusion. As observed in Figure 5, the early reduction in content distance leads to a notable decrease in cross-entropy. Thus, the subsequent diffusion aligns the data more closely with the ground truth. To delve into this synergistic effect, we provide a theoretical study from a manifold perspective. We examine how SR influences data positioning and allows for more precise use of the score function within diffusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 240, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 240, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 240, + 415 + ], + "type": "text", + "content": "4.2. Geometrical Analysis on PuriFlow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": " is the set of all data points, i.e., " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\forall \\mathbf{x} \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": ", defining " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": " as the data manifold. That is, the data distribution " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": " is uniform on the data manifold " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": ". Then, within a local neighborhood, the manifold coincides with its tangent space of dimension " + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "inline_equation", + "content": "k \\ll d" + }, + { + "bbox": [ + 55, + 420, + 296, + 480 + ], + "type": "text", + "content": ", expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 489, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 489, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 61, + 489, + 295, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right) = T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\cap \\mathcal {B} \\left(\\mathbf {x} _ {0}, d r\\right), \\quad T _ {\\mathbf {x} _ {0}} \\mathcal {M} \\simeq \\mathbb {R} ^ {k}, \\tag {8}", + "image_path": "7a7b04c48851fdabfcdc9b8dfcb3e63106dfa2b97883a0b6235938f03c616803.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{x}_0}\\mathcal{M}" + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "content": " denotes the tangent space to the manifold " + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "inline_equation", + "content": "dr" + }, + { + "bbox": [ + 55, + 510, + 296, + 571 + ], + "type": "text", + "content": " is an infinitesimal radius within which the manifold can be locally approximated by its tangent space. Here, we take note of a property of the trained score function in diffusion models [12]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": "Remark 1. The noisy data become increasingly concentrated on spherical manifolds as diffusion time " + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": " increases. The score function is trained via denoising score matching (Eq. 3), using data on " + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t(\\cdot)" + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": " denote a mapping from noisy data " + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": " to its estimated clean counterpart " + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_0" + }, + { + "bbox": [ + 55, + 578, + 296, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 646, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 646, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 67, + 646, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {t} (\\cdot): \\boldsymbol {x} _ {t} \\mapsto \\hat {\\boldsymbol {x}} _ {0} = - \\frac {1}{2} (1 - \\alpha_ {t}) \\left(\\boldsymbol {x} _ {t} + 2 \\boldsymbol {s} _ {\\theta^ {*}} \\left(\\boldsymbol {x} _ {t}, t\\right)\\right), \\tag {9}", + "image_path": "9dab2d41fff965a4c55b6801d311fa9dce05ee666f3f91b6a56fe03afbb6e38e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t(\\pmb{x}_t) \\in \\mathcal{M}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and it satisfies " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{J}_{\\mathcal{P}_t}^2 = \\pmb{J}_{\\mathcal{P}_t} = \\pmb{J}_{\\mathcal{P}_t}^T : \\mathbb{R}^d \\to T_{\\mathcal{P}_t(\\pmb{x}_t)}\\mathcal{M}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ". That is, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " acts as an orthonormal projection onto " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "content": "Remark 2. The score function is trained solely on the data concentrated on both the noisy and data manifolds. Therefore, applying it to data outside these manifolds may lead to inaccurate inference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "content": "Both premises suggest that when the score function is applied to off-manifold data, such as adversarial examples [7], its inaccurate inference may lead to misaligned orthonormal projections onto an unintended tangent space. Thus, the off-manifold data should be corrected to better align with its underlying manifold. Now, we show that the applying SR followed by downsampling can typically position off-manifold data, degraded from " + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "content": ", closer to the local manifold, which is the tangent space of " + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 293, + 555, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": "Proposition 1. Suppose " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\omega^*}" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " is an ideal SR and smooth continuity, and let " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " be a linear downsampler. Given an offmanifold data " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathfrak{D}(\\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " degraded from " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": ", the processed data " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{R}\\mathcal{H}_{\\omega^*}(\\mathfrak{D}(\\pmb{x}_0))" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " is projected near the tangent space " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "T_{\\pmb{x}_0}\\mathcal{M}" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " of the data manifold " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": " with a tiny " + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 313, + 404, + 554, + 464 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 469, + 553, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 469, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 339, + 469, + 553, + 484 + ], + "type": "interline_equation", + "content": "\\left\\| \\mathcal {R H} _ {\\omega^ {*}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) - \\Pi_ {T _ {x _ {0}} \\mathcal {M}} \\left(\\mathfrak {D} \\left(\\boldsymbol {x} _ {0}\\right)\\right) \\right\\| _ {F} < \\delta . \\tag {10}", + "image_path": "6d0125454a236d15cf15ac1570cda24060be1f6fdc847ff7f0feff2c1055297d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 523 + ], + "type": "text", + "content": "In Section 5, we evaluate the effectiveness of using the score function more accurately, supported by Proposition 1, against various adversarial attacks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 533, + 391, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 391, + 546 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 391, + 546 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": "We evaluate the performance of PuriFlow in two configurations: integrating SR with solving SDEs [38] and combining SR with a recent alternative, iterative One-Shot Denoising (OSD) [60]. OSD is defined by Tweedie's formula [50], expressed as " + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\mathbf{x}|\\tilde{\\mathbf{x}}] = \\tilde{\\mathbf{x}} +\\sigma^2\\nabla_{\\tilde{\\mathbf{x}}}\\log p_\\sigma (\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": " denotes perturbed data and " + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "inline_equation", + "content": "p_{\\sigma}(\\tilde{\\mathbf{x}})" + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": " represents marginal Gaussian perturbation of " + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": ". This formulation avoids repeated noise removal steps for varying time " + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 552, + 554, + 648 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "content": "In this context, the denoising score-matching problem defined by Eq. 3 optimizes the neural network " + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\theta^{*}}(\\cdot ,t)" + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "content": " to predict a fully denoised " + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "content": " per a specific time " + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 648, + 554, + 696 + ], + "type": "text", + "content": ". Thus, Tweedie's formula can be represented as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 359, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathbb {E} \\left[ \\mathbf {x} _ {0} \\mid \\mathbf {x} _ {t ^ {\\prime}} \\right] = \\mathbf {x} _ {t ^ {\\prime}} + \\sigma^ {2} \\nabla_ {\\mathbf {x} _ {t ^ {\\prime}}} \\log p _ {\\sigma} \\left(\\mathbf {x} _ {t ^ {\\prime}}\\right). \\tag {11}", + "image_path": "dbfb91a0c8efcc8190f396272b9cadc96756b98975f8e8169ed0215d57bdbd43.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4609" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 186 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 186 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 186 + ], + "type": "table", + "html": "
MethodCriterionStandard (%)Robust (%)
Yang et al. [56]Gibbs Update94.8040.80
Song et al. [47]Mask+Recon.95.009.00
Hill et al. [24]EBM+LD84.1254.90
Yoon et al. [58]DSM+LD*86.1470.01
Nie et al. [38]SDEs89.0281.40
Bai et al. [5]Guide+SDEs92.6181.94
PuriFlow([38])SR+SDEs90.0683.00
Zhang et al. [60]iOSD93.7588.08
PuriFlow([60])SR+iOSD94.1491.01
", + "image_path": "57072df7ee4890fea5cf1d1aa6595111b85fa4515f82fdbc6a98bc6a7556b4ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "type": "text", + "content": "Table 1. Comparison of different adversarial purification methods against BPDA+EOT attacks with " + }, + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "type": "text", + "content": " perturbations. Our evaluations use WRN-28-10 on CIFAR-10, with consistent experimental settings as outlined in [38, 60] for " + }, + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 55, + 191, + 296, + 247 + ], + "type": "text", + "content": ". (* This purification utilizes a variant of the LD sampling.)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 268, + 297, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 297, + 426 + ], + "type": "text", + "content": "The alternative proposed by [60] involves iteratively applying OSD using Eq. 11 during noise removal. Therefore, it can serve as a suitable baseline to confirm synergies with SR. We assess these configurations, detailed as algorithms in the Appendix, across key test-time defense scenarios, including adaptive white-box, preprocessor-blind [58], and black-box attacks. This section also includes an ablation study to demonstrate the impact of incorporating SR. Detailed experimental settings and additional evaluations for certified robustness are provided in the Appendix. For adaptive attacks, we adopt MDSR [33] for efficiency and ESRGAN [52] otherwise. We denote " + }, + { + "bbox": [ + 54, + 268, + 297, + 426 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 54, + 268, + 297, + 426 + ], + "type": "text", + "content": " on a discrete scale [26] from 1 to 1000 to represent the continuous interval [0, 1]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 432, + 214, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 214, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 214, + 445 + ], + "type": "text", + "content": "5.1. Defense on Adaptive Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "text", + "content": "In this section, we evaluate the performance of PuriFlow against four strong adaptive attacks: BPDA+EOT [3, 4], an adaptive ensemble white-box attack; AutoAttack [13], an adaptive white-box attack; PGD+EOT [32], and DiffAttack [29], a diffusion-based purification targeted attack. Here, the evaluations focus on projections within an " + }, + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "text", + "content": "-norm ball, as the projection in " + }, + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\ell_{2}" + }, + { + "bbox": [ + 55, + 449, + 296, + 534 + ], + "type": "text", + "content": "-norm ball is less effective." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 649 + ], + "type": "text", + "content": "BPDA+EOT. Backward Pass Differentiable Approximation (BPDA) combined with Expectation over Transformation (EOT) is used to evaluate randomized adversarial purification methods. Table 1 demonstrates that PuriFlow outperforms standalone diffusion on CIFAR-10 with WRN-28-10, achieving a more balanced standard and robust accuracy than the guiding method. In addition, integrating SR with iterative OSD further enhances performance, emphasizing its critical role in diffusion-based purification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 715 + ], + "type": "text", + "content": "AutoAttack. Tables 2 demonstrates the effectiveness of PuriFlow against AutoAttack, on ImageNet-1k. PuriFlow outperforms adversarial training methods and a diffusion-only approach across ResNet-50 and DeiT-S. Compared to diffusion only, it achieves notable robust accuracy improve" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 71, + 548, + 219 + ], + "blocks": [ + { + "bbox": [ + 322, + 71, + 548, + 219 + ], + "lines": [ + { + "bbox": [ + 322, + 71, + 548, + 219 + ], + "spans": [ + { + "bbox": [ + 322, + 71, + 548, + 219 + ], + "type": "table", + "html": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5631.06
Wong et al. [53]55.6226.95
Salman et al. [42]64.0237.89
Bai et al. [6]67.3835.51
APNie et al. [38]67.7940.93
PuriFlow([38])71.6852.92
DeiT-S79.900.00
ATBai et al. [6]66.5035.50
APNie et al. [38]73.6343.18
PuriFlow([38])75.6953.12
", + "image_path": "4c45244a0181a31891b6bc2abdba9e560cb046d0f2677f40bb2153bc9c8606ba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": "Table 2. Evaluation comparing adversarial training (AT) and purification (AP) methods against AutoAttack with " + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": " perturbations at " + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": " on ImageNet-1k. Diffusion time is set to " + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "inline_equation", + "content": "t' = 150" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": " for [38] as specified its study and " + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "inline_equation", + "content": "t' = 130" + }, + { + "bbox": [ + 313, + 226, + 555, + 271 + ], + "type": "text", + "content": " for PuriFlow." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 283, + 548, + 471 + ], + "blocks": [ + { + "bbox": [ + 322, + 283, + 548, + 471 + ], + "lines": [ + { + "bbox": [ + 322, + 283, + 548, + 471 + ], + "spans": [ + { + "bbox": [ + 322, + 283, + 548, + 471 + ], + "type": "table", + "html": "
TypeMethodStandard (%)Robust (%)
WRN-28-1095.630.00
ATPang et al. [39]88.6264.95
Gowal et al. [19]88.5465.93
Gowal et al. [20]87.5166.01
APYoon et al. [58]85.6633.48
Zhang et al. [60]*93.7559.37
PuriFlow([60])*94.1460.76
WRN-70-1695.790.00
ATGowal et al. [19]91.1068.66
Gowal et al. [20]88.7469.03
Rebuffi et al. [41]92.2269.97
APYoon et al. [58]86.7637.11
Zhang et al. [60]*95.1163.28
PuriFlow([60])*95.3164.86
", + "image_path": "02ff93cb6170e8349a78626c9814099957d5d03a64fd657431574092922d2f4e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "text", + "content": "Table 3. Evaluation comparing adversarial training (AT) and purification (AP) methods against PGD+EOT within an " + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "text", + "content": "-norm ball of radius " + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\epsilon = 8/255" + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "text", + "content": " on CIFAR-10. Results are sourced from [32]. (* For a fair comparison, PuriFlow and [60] are directly evaluated on the same settings with " + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "inline_equation", + "content": "t' = 250" + }, + { + "bbox": [ + 313, + 478, + 555, + 536 + ], + "type": "text", + "content": ", as stated in [60].)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "text", + "content": "ments, including " + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "inline_equation", + "content": "11.99\\%" + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "text", + "content": " (ResNet-50) and " + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "inline_equation", + "content": "9.94\\%" + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "text", + "content": " (DeiT-S), and standard accuracy. PuriFlow also shows efficiency with reduced diffusion time " + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 313, + 555, + 555, + 592 + ], + "type": "text", + "content": " on this large-scale dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "inline_equation", + "content": "\\mathbf{PGD} + \\mathbf{EOT}" + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "text", + "content": ". Motivated by [32], we assess PuriFlow's robustness against " + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{PGD} + \\mathrm{EOT}" + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "text", + "content": " that utilizes gradient signs within the entire randomized defense framework. Table 3 shows PuriFlow's strength on CIFAR-10, achieving the highest standard accuracy for WRN-28-10 and WRN-70-16. Despite the substantial impact of " + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "inline_equation", + "content": "\\mathrm{PGD} + \\mathrm{EOT}" + }, + { + "bbox": [ + 313, + 600, + 556, + 721 + ], + "type": "text", + "content": " on randomized purification, PuriFlow achieves superior robust accuracy among purification methods, demonstrating a synergistic effect in which SR effectively acts as the initialization before starting the iterative one-shot denoising process." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4610" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 71, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 288, + 167 + ], + "type": "table", + "html": "
TypeMethodStandard (%)Robust (%)
WRN-70-1695.790.00
APNie et al. [38]90.0745.31
PuriFlow([38])90.2159.47
ResNet-5076.130.00
APNie et al. [38]67.7928.13
PuriFlow([38])71.6843.75
", + "image_path": "f5288071aec4dec5b79d89a0053f33a337958abe2f2b834ad26e0ee3578c1b85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 217, + 288, + 365 + ], + "blocks": [ + { + "bbox": [ + 55, + 173, + 295, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 295, + 206 + ], + "type": "text", + "content": "Table 4. Evaluation against DiffAttack, targeting WRN-70-16 on CIFAR-10 and ResNet-50 on ImageNet-1k. Each diffusion time " + }, + { + "bbox": [ + 55, + 173, + 295, + 206 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 173, + 295, + 206 + ], + "type": "text", + "content": " is the same as those used for other attacks on PuriFlow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 217, + 288, + 365 + ], + "lines": [ + { + "bbox": [ + 63, + 217, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 63, + 217, + 288, + 365 + ], + "type": "table", + "html": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.130.00
ATEngstrom et al. [17]62.5638.97
Wong et al. [53]55.6229.15
Salman et al. [42]64.0238.01
Bai et al. [6]67.3840.27
APNie et al. [38]74.5238.87
PuriFlow([38])74.0549.92
DeiT-S79.900.00
ATBai et al. [6]66.5040.32
APNie et al. [38]78.9434.59
PuriFlow([38])78.4449.85
", + "image_path": "4416ba052962728e8edc7ebf679c701be74d53fbb4b8e21ad33fc613051fa9ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": "DiffAttack. DiffAttack neutralizes vanishing and exploding gradients, high memory costs, and increased randomness—natural defensive effects from lengthy " + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": " of diffusion-based purification. Table 4 demonstrates that while purification methods using " + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "inline_equation", + "content": "t' = 100" + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": " for CIFAR-10 and " + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "inline_equation", + "content": "t' = 150" + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": " for ImageNet-1k exhibit limitations in robustness by this intentional attack, integrating SR demonstrates a notable improvement by " + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "inline_equation", + "content": "14.16\\%" + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "inline_equation", + "content": "15.62\\%" + }, + { + "bbox": [ + 54, + 425, + 295, + 558 + ], + "type": "text", + "content": ". We hypothesize that SR's early enhancement of feature proximity helps mitigate diverging feature maps from the ground truth, even under the reconstruction-deviating impacts of DiffAttack." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 257, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 257, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 257, + 577 + ], + "type": "text", + "content": "5.2. Defense on Preprocessor-Blind Attack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "We evaluate PuriFlow against the PGD (" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ") attack, targeting only the classifier on ImageNet-1k. Table 5 shows that PuriFlow, leveraging only off-the-shelf models, surpasses adversarial training methods in both standard and robust accuracy for ResNet-50 and DeiT-S. Notably, PuriFlow significantly outperforms diffusion only [38] at the same practical diffusion time " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "t' = 10" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": ", achieving comparable standard accuracy while enhancing robust accuracy by " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "11.05\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for ResNet-50 and " + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "inline_equation", + "content": "15.26\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": " for DeiT-S. This improvement signifies the synergistic importance of SR in adversarial purification utilizing diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 71, + 547, + 128 + ], + "blocks": [ + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "text", + "content": "Table 5. Comparison under a preprocessor-blind PGD attack targeting only the classifier, within an " + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "text", + "content": " perturbations at " + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "text", + "content": " on ImageNet-1k. The time " + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 371, + 295, + 405 + ], + "type": "text", + "content": " for both AP methods is set as 10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 71, + 547, + 128 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 547, + 128 + ], + "type": "table", + "html": "
TypeMethodStandard (%)Robust (%)
ResNet-5076.139.25
APNie et al. [38]67.7962.88
PuriFlow [38]74.0068.36
", + "image_path": "3bf280e4b7f6d1224f20ce1732dd352df65c54020dcbb828a49f61080312d896.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 206, + 555, + 320 + ], + "blocks": [ + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": "Table 6. Evaluation of ResNet-50 under a black-box attack using Square Attack (" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "inline_equation", + "content": "N = 5000" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": ") within an " + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": "-norm ball of radius " + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\epsilon = 4/255" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": " on ImageNet-1k. Diffusion time is set to " + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "inline_equation", + "content": "t' = 150" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": " for [38] as specified in their study and " + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "inline_equation", + "content": "t' = 10" + }, + { + "bbox": [ + 313, + 136, + 555, + 191 + ], + "type": "text", + "content": " for PuriFlow. Evaluation settings follow those in [38]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 206, + 555, + 320 + ], + "lines": [ + { + "bbox": [ + 315, + 206, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 206, + 555, + 320 + ], + "type": "table", + "html": "
MethodTypeStandard (%)Robust (%)
ResNet-5076.130.00
WaveletDenoising71.1568.63
TVMDenoising57.8056.96
NL-meansDenoising64.3262.67
DRLN×2 [2]SR71.6668.72
EDSR×2 [33]SR71.7768.73
ESRGAN×2 [52]SR72.9768.75
MDSR×2 [33]SR73.3269.47
", + "image_path": "f72e8892f71090e296c21b49e47526444b950798826a917ffa454312b9ee7316.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 328, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 373 + ], + "type": "text", + "content": "Table 7. Comparison of SR with other denoising methods as an initialization approach for diffusion [38] (" + }, + { + "bbox": [ + 313, + 328, + 555, + 373 + ], + "type": "inline_equation", + "content": "t' = 90" + }, + { + "bbox": [ + 313, + 328, + 555, + 373 + ], + "type": "text", + "content": ") under a PGD attack on ImageNet-1k. \"TVM\" refers to total variance minimization, while \"NL-means\" denotes the non-local means algorithm." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 398, + 474, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 474, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 474, + 410 + ], + "type": "text", + "content": "5.3. Defense on Black-Box Attack" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": "Table 6 shows that PuriFlow, using " + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "inline_equation", + "content": "t' = 10" + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": " approximately " + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "inline_equation", + "content": "6.67\\%" + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": " of the diffusion time required by diffusion-only method [38]—achieves notable improvements in both standard and robust accuracy for ResNet-50. Specifically, it shows enhancements of " + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "inline_equation", + "content": "6.21\\%" + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "inline_equation", + "content": "5.48\\%" + }, + { + "bbox": [ + 313, + 417, + 555, + 514 + ], + "type": "text", + "content": " in each accuracy on ImageNet-1k. These results highlight the synergy of initialization for diffusion, indicating better performance and greater efficiency in this black-box attack." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 526, + 408, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 408, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 408, + 539 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Impact of SR compared to various candidates. Puri-Flow can be configured by combining various initialization methods. We test several SR models and traditional denoising techniques by selectively replacing each method used in diffusion-based purification. This evaluation, conducted with PGD targeting the classifier on ImageNet-1k, is shown in Table 7. While Wavelet denoising is practical for PGD, it stands less effective than SR techniques. Among SR models tested, MDSR " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " outperforms DRLN " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ", EDSR " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ", and ESRGAN " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": ", achieving the highest standard and robust accuracies. Furthermore, we explore different upsample ratios " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "(\\times 2, \\times 4, \\times 8)" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " using ESRGAN. Table 10 indicates that " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " represents the best balance between efficiency and performance, making it the optimal choice for PuriFlow." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4611" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 352, + 132 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 352, + 132 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 352, + 132 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 352, + 132 + ], + "type": "table", + "html": "
SRDiffusionClassifierStandard (%)Robust (%)Wall time (s /img)
✓76.1300.00≈0.00
✓✓73.6830.700.01
✓✓74.5238.371.08
✓✓✓74.0049.921.09
", + "image_path": "a863487964602bc927b57e31af7aab08f3acaa111cdcb910bf8ec5561b30071f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 353, + 71, + 548, + 132 + ], + "blocks": [ + { + "bbox": [ + 353, + 71, + 548, + 132 + ], + "lines": [ + { + "bbox": [ + 353, + 71, + 548, + 132 + ], + "spans": [ + { + "bbox": [ + 353, + 71, + 548, + 132 + ], + "type": "table", + "html": "
ModelTypeStandard (%)Robust (%)
ResNet-50VP-SDE72.9768.75
VE-SDE73.3966.51
DeiT-SVP-SDE77.2972.37
VE-SDE77.7269.33
", + "image_path": "02dc552cb589addb9523ada30f6322d94fbac07830d37582c60bcae80cb33059.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 187, + 296, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "type": "text", + "content": "Table 8. Evaluation of PuriFlow under a preprocessor-blind PGD attack on ImageNet-1k selectively employing off-the-shelf models. Left: ResNet-50, ESRGAN×2, and VP-SDE are the classifier, SR model, and diffusion [38] model, respectively, with diffusion time " + }, + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "type": "inline_equation", + "content": "t' = 10" + }, + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "type": "text", + "content": ". Right: Impact of diffusion models within PuriFlow, evaluated at diffusion time " + }, + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "type": "inline_equation", + "content": "t' = 90" + }, + { + "bbox": [ + 55, + 135, + 553, + 169 + ], + "type": "text", + "content": " using ResNet-50 and DeiT-S on ImageNet-1k." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 187, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 187, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 296, + 304 + ], + "type": "table", + "html": "
Off-the-shelf modelsESRGAN×2EDSR×2MDSR×2
SR (each column)0.010.01≈0.00
Diffusion [38]1.080.810.73
ResNet-50≈0.00≈0.00≈0.00
Total wall time (s / img)1.090.820.73
ConfigurationCriterionIterationsPurif. time
Diffusion [38]SDEst' = 1005.27
SR+DiffusionSR+SDEst' = 1000.005+5.27
ScoreOpt-O [60]iOSDM = 50.51
SR+ScoreOpt-OSR+iOSDM = 50.005+0.51
", + "image_path": "7d629769932d338344cadaa13017dc42fb8261dc498e9fc7b7442e17843f5f5d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 63, + 354, + 288, + 408 + ], + "blocks": [ + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 295, + 343 + ], + "type": "text", + "content": "Table 9. Top: Wall time for PuriFlow with different SR models. Bottom: Examples of purification times (s/img) for configurations using WRN-70-16 [38] and WRN-28-10 [60] on CIFAR-10." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 354, + 288, + 408 + ], + "lines": [ + { + "bbox": [ + 63, + 354, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 63, + 354, + 288, + 408 + ], + "type": "table", + "html": "
RatioStandard (%)Robust (%)FLOPsParams
×871.26%66.80%1110G16.7M
×472.41%67.32%899G16.7M
×272.97%68.75%224G16.7M
", + "image_path": "d680720cc2e68b0f3be98d294e8b28448c9acb70be9ac50ec3184e9841267a9f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "type": "text", + "content": "Table 10. Impact of SR upsampling ratios from ESRGAN " + }, + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "type": "text", + "content": " combined with diffusion [38] (" + }, + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "type": "inline_equation", + "content": "t' = 90" + }, + { + "bbox": [ + 55, + 413, + 296, + 460 + ], + "type": "text", + "content": ") using ResNet-50 on ImageNet-1k. FLOPs represent computational complexity [23], and Params refer to the number of trainable parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": "Off-the-shelf model adoption and latency analysis. Table 8 shows the enhanced defensive capabilities achieved by integrating off-the-shelf models into PuriFlow's framework. The table highlights the individual and combined effects of SR and diffusion. When used together, SR and diffusion improve robust accuracy by " + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "inline_equation", + "content": "19.22\\%" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "inline_equation", + "content": "11.55\\%" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": ", respectively, illustrating their strong synergistic effect on adversarial examples. Although combining SR with diffusion slightly decreases standard accuracy compared to diffusion alone, the robustness gains make this trade-off useful, with a tolerable latency of " + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "inline_equation", + "content": "1.09~\\mathrm{s / img}" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": " when executed on a single NVIDIA RTX 4090 GPU. Table 9 further confirms that integrating SR models adds minimal latency to the overall purification process. Specifically, SR's execution time is " + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "inline_equation", + "content": "1,054\\times" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": " faster than the diffusion's, and its overhead is only " + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "inline_equation", + "content": "1/102" + }, + { + "bbox": [ + 55, + 479, + 295, + 682 + ], + "type": "text", + "content": " compared to one-shot denoising methods. This result emphasizes SR's role as an effective and efficient initialization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Impact of diffusion types. We explore VE-SDE [46], which can act as an alternative to diffusion. Table 8 demon" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 188, + 446, + 296 + ], + "blocks": [ + { + "bbox": [ + 318, + 188, + 446, + 296 + ], + "lines": [ + { + "bbox": [ + 318, + 188, + 446, + 296 + ], + "spans": [ + { + "bbox": [ + 318, + 188, + 446, + 296 + ], + "type": "image", + "image_path": "59dbb534d522936d3ff20b6b46c71b63ae0c4133cad62412958869e6f977e395.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": "Figure 6. Impact of iterative SR on cross-entropy and MSEs under a PGD " + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "inline_equation", + "content": "(N = 100)" + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": " attack targeting ResNet-50 within an " + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": "-norm ball " + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "inline_equation", + "content": "(\\epsilon = 4 / 255)" + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": " on ImageNet-1k. SR, SR', and SR\" indicate one, two, and three rounds of SR using ESRGAN " + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 313, + 302, + 555, + 346 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 450, + 189, + 551, + 297 + ], + "blocks": [ + { + "bbox": [ + 450, + 189, + 551, + 297 + ], + "lines": [ + { + "bbox": [ + 450, + 189, + 551, + 297 + ], + "spans": [ + { + "bbox": [ + 450, + 189, + 551, + 297 + ], + "type": "image", + "image_path": "85d77e4784240ce9cfa162e1157267cbf7396a147b154f717fc79d492661890b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "content": "strates that VP-SDE slightly reduces standard accuracy (approximately " + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "inline_equation", + "content": "0.42\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "content": " for ResNet-50 and " + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "inline_equation", + "content": "0.43\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "content": " for DeiT-S) but enhances robust accuracy by " + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "inline_equation", + "content": "2.24\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "inline_equation", + "content": "3.04\\%" + }, + { + "bbox": [ + 313, + 366, + 555, + 438 + ], + "type": "text", + "content": ", respectively. Despite the minor drop in standard accuracy, integrating VP-SDE shows overall robustness, synergizing a balanced performance with SR." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 556, + 576 + ], + "type": "text", + "content": "Iterative applications of SR without diffusion. Figure 6 shows the impact of repeated SR applications for adversarial examples relative to their originals, with changes in MSE and cross-entropy and their impact on standard and robust accuracy. Up to three rounds of SR demonstrate a trade-off: MSE increases while cross-entropy decreases and eventually saturates. Here, robust accuracy rises but converges while standard accuracy decreases. We conjecture that repeated SR may behave as seen in model inversion [35, 57]: Images are generated to minimize cross-entropy loss, but this steady direction moves them farther from the originals." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 388, + 597 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We introduce PuriFlow, a novel purification flow integrating Super-Resolution (SR) with diffusion. While SR alone is less effective at mitigating adversarial noise, its image upsampling and downsampling excel at restoring subtle features captured by early layers. Diffusion alone lacks this ability. However, their mixture synergistically reduces overall feature proximity and cross-entropy. PuriFlow outperforms leading methods in extensive evaluations, demonstrating efficiency for minimal SR overhead." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4612" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "text", + "content": "This research was partly supported by the IITP(Institute of Information & Communications Technology Planning & Evaluation)-ITRC(Information Technology Research Center) grant funded by the Korea government(MSIT)(IITP-2025-RS-2023-00258649, " + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "text", + "content": " ), the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2025-00562437, " + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "text", + "content": " ), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (No.RS-2022-00155911, Artificial Intelligence Convergence Innovation Human Resources Development (Kyung Hee University), " + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 92, + 296, + 236 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 249, + 115, + 262 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 269, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 269, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 295, + 334 + ], + "type": "text", + "content": "[1] Yasin Almalioglu, Kutsev Bengisu Ozyoruk, Abdulkadir Gokce, Kagan Incetan, Guliz Irem Gokceler, Muhammed Ali Simsek, Kivanc Ararat, Richard J Chen, Nicholas J Durr, Faisal Mahmood, et al. Endol2h: deep super-resolution for capsule endoscopy. IEEE Transactions on Medical Imaging, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 337, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 337, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 296, + 370 + ], + "type": "text", + "content": "[2] Saeed Anwar and Nick Barnes. Densely residual laplacian super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 372, + 295, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 295, + 416 + ], + "type": "text", + "content": "[3] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, 2018. 1, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 418, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 418, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 294, + 452 + ], + "type": "text", + "content": "[4] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In International Conference on Machine Learning, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 294, + 507 + ], + "type": "text", + "content": "[5] Mingyuan Bai, Wei Huang, Tenghui Li, Andong Wang, Junbin Gao, Cesar F Caiafa, and Qibin Zhao. Diffusion models demand contrastive guidance for adversarial purification to advance. In International Conference on Machine Learning, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 542 + ], + "type": "text", + "content": "[6] Yutong Bai, Jieru Mei, Alan L Yuille, and Cihang Xie. Are transformers more robust than cnns? Advances in Neural Information Processing Systems, 2021. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 599 + ], + "type": "text", + "content": "[7] Kartikeya Bhardwaj, Dibakar Gope, James Ward, Paul Whatmough, and Danny Loh. Super-efficient super resolution for fast adversarial defense at the edge. In 2022 Design, Automation & Test in Europe Conference & Exhibition, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 294, + 656 + ], + "type": "text", + "content": "[8] Kartikeya Bhardwaj, Milos Milosavljevic, Liam O'Neil, Dibakar Gope, Ramon Matas, Alex Chalfin, Naveen Suda, Lingchuan Meng, and Danny Loh. Collapsible linear blocks for super-efficient super resolution. Proceedings of Machine Learning and Systems, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "text", + "content": "[9] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[10] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 183 + ], + "type": "text", + "content": "[11] Nicholas Carlini, Florian Tramer, Krishnamurthy Dj Dvijotham, Leslie Rice, Mingjie Sun, and J Zico Kolter. (certified!!) adversarial robustness for free! In International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "text", + "content": "[12] Hyungjin Chung, Byeongsu Sim, Dohoon Ryu, and Jong Chul Ye. Improving diffusion models for inverse problems using manifold constraints. Advances in Neural Information Processing Systems, 2022. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "text", + "content": "[13] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International Conference on Machine Learning, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "text", + "content": "[14] Nilaksh Das, Madhuri Shanbhogue, Shang-Tse Chen, Fred Hohman, Li Chen, Michael E Kounavis, and Duen Horng Chau. Keeping the bad guys out: Protecting and vaccinating deep learning withJPEG compression. arXiv preprint arXiv:1705.02900, 2017.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 360 + ], + "type": "text", + "content": "[15] Yilun Du and Igor Mordatch. Implicit generation and modeling with energy based models. Advances in Neural Information Processing Systems, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 361, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 402 + ], + "type": "text", + "content": "[16] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M Roy. A study of the effect of jpg compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 436 + ], + "type": "text", + "content": "[17] Logan Engstrom, Andrew Ilyas, Hadi Salman, Shibani Santurkar, and Dimitris Tsipras. Robustness (python library), 2019. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 480 + ], + "type": "text", + "content": "[18] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[19] Sven Gowal, Chongli Qin, Jonathan Uesato, Timothy Mann, and Pushmeet Kohli. Uncovering the limits of adversarial training against norm-bounded adversarial examples. arXiv preprint arXiv:2010.03593, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "text", + "content": "[20] Sven Gowal, Sylvestre-Alvise Rebuffi, Olivia Wiles, Florian Stimberg, Dan Andrei Calian, and Timothy A Mann. Improving robustness using generated data. Advances in Neural Information Processing Systems, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[21] Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 668 + ], + "type": "text", + "content": "[22] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens van der Maaten. Countering adversarial images using input transformations. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4613" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[24] Mitch Hill, Jonathan Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. International Conference on Learning Representations, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[25] Mitch Hill, Jonathan Craig Mitchell, and Song-Chun Zhu. Stochastic security: Adversarial defense using long-run dynamics of energy-based models. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 239 + ], + "type": "text", + "content": "[27] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning. Advances in Neural Information Processing Systems, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 274 + ], + "type": "text", + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, 2016. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "text", + "content": "[29] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. Advances in Neural Information Processing Systems, 2024. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "type": "text", + "content": "[30] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 410 + ], + "type": "text", + "content": "[31] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 453 + ], + "type": "text", + "content": "[32] Minjong Lee and Dongwoo Kim. Robust evaluation of diffusion-based adversarial purification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[33] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2017. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 555 + ], + "type": "text", + "content": "[34] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 589 + ], + "type": "text", + "content": "[35] Alexander Mordvintsev, Christopher Olah, and Mike Tyka. Inceptionism: Going deeper into neural networks. Google Research Blog, 2015. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 634 + ], + "type": "text", + "content": "[36] Aamir Mustafa, Salman H. Khan, Munawar Hayat, Jianbing Shen, and Ling Shao. Image super-resolution as a defense against adversarial attacks. IEEE Transactions on Image Processing, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[38] Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In International Conference on Machine Learning, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[39] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. In International Conference on Machine Learning, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 171 + ], + "type": "text", + "content": "[40] Aaditya Prakash, Nick Moran, Solomon Garber, Antonella DiLillo, and James Storer. Deflecting adversarial attacks with pixel deflection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 217 + ], + "type": "text", + "content": "[41] Sylvestre-Alvise Rebuffi, Sven Gowal, Dan A Calian, Florian Stimberg, Olivia Wiles, and Timothy Mann. Fixing data augmentation to improve adversarial robustness. arXiv preprint arXiv:2103.01946, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[42] Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do adversarially robust imagenet models transfer better? Advances in Neural Information Processing Systems, 2020. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[43] Pouya Samangouei, Maya Kabbab, and Rama Chellappa. Defense-GAN: Protecting classifiers against adversarial attacks using generative models. In International Conference on Learning Representations, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "text", + "content": "[44] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "text", + "content": "[45] Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo. Perceptual extreme super-resolution network with receptive field block. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 421, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 554, + 453 + ], + "type": "text", + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 509 + ], + "type": "text", + "content": "[47] Yang Song, Taesup Kim, Sebastian Nowozin, Stefano Ermon, and Nate Kushman. Pixeldefend: Leveraging generative models to understand and defend against adversarial examples. In International Conference on Learning Representations, 2018. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 511, + 554, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 554, + 565 + ], + "type": "text", + "content": "[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 621 + ], + "type": "text", + "content": "[49] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 554, + 655 + ], + "type": "text", + "content": "[50] Soobin Um, Suhyeon Lee, and Jong Chul Ye. Don't play favorites: Minority guidance for diffusion models. In International Conference on Learning Representations, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 657, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 554, + 678 + ], + "type": "text", + "content": "[51] Pascal Vincent. A connection between score matching and denoising autoencoders. *Neural Computation*, 2011. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[52] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4614" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 430 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision Workshops, 2018. 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[53] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[54] Fei Wu, Wenxue Yang, Limin Xiao, and Jinbin Zhu. Adaptive wiener filter and natural noise to eliminate adversarial perturbation. *Electronics*, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "text", + "content": "[55] Cihang Xie, Yuxin Wu, Laurens van der Maaten, Alan L Yuille, and Kaiming He. Feature denoising for improving adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[56] Yuzhe Yang, Guo Zhang, Dina Katabi, and Zhi Xu. Me-net: Towards effective adversarial robustness with matrix estimation. In International Conference on Machine Learning, 2019. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[57] Hongxu Yin, Pavlo Molchanov, Jose M Alvarez, Zhizhong Li, Arun Mallya, Derek Hoiem, Niraj K Jha, and Jan Kautz. Dreaming to distill: Data-free knowledge transfer via deep-inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 354 + ], + "type": "text", + "content": "[58] Jongmin Yoon, Sung Ju Hwang, and Juho Lee. Adversarial purification with score-based generative models. In International Conference on Machine Learning, 2021. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 386 + ], + "type": "text", + "content": "[59] MD Zeiler. Visualizing and understanding convolutional networks. In European Conference on Computer Vision, 2014. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[60] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4615" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_content_list.json b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..275e37af6cd510dbb9f4ff1cd2406559c0bb9039 --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "Adversarial Reconstruction Feedback for Robust Fine-grained Generalization", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shijie Wang $^{1}$ , Jian Shi $^{2}$ , Haojie Li $^{1*}$ $^{1}$ College of Computer and Engineering, Shandong University of Science and Technology, China \n $^{2}$ School of Software, Dalian University of Technology, China", + "bbox": [ + 119, + 179, + 877, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing fine-grained image retrieval (FGIR) methods predominantly rely on supervision from predefined categories to learn discriminative representations for retrieving fine-grained objects. However, they inadvertently introduce category-specific semantics into the retrieval representation, creating semantic dependencies on predefined classes that critically hinder generalization to unseen categories. To tackle this, we propose AdvRF, a novel adversarial reconstruction feedback framework aimed at learning category-agnostic discrepancy representations. Specifically, AdvRF reformulates FGIR as a visual discrepancy reconstruction task via synergizing category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models. The reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. Consequently, the retrieval model localizes visual differences, while the reconstruction model encodes these differences into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Quantitative and qualitative evaluations demonstrate that our AdvRF achieves impressive performance on both widely-used fine-grained and coarse-grained datasets.", + "bbox": [ + 89, + 301, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fine-grained image retrieval (FGIR) aims to retrieve visually similar subcategories, even those that were unseen during the training phase. It plays a vital role in numerous vision applications from fashion industry, e.g., retrieval of diverse types of clothes [1, 18, 21, 41], to environmental conservation, e.g., retrieving endangered species [5, 37, 39, 42]. Given its significance, a substantial body of research has focused on learning discriminative and generalizable embed", + "bbox": [ + 89, + 752, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5238bdc197d484a7c1e124e264865e9493dcfea68333971c446d71926c0d534.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 268, + 903, + 333 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e817585f07671b66f24a9eca6de95e7e92c007ceca278153112850bdedc77360.jpg", + "image_caption": [ + "(a) Classification-based Representation Learning", + "(b) Reconstruction-based Representation Learning", + "Figure 1. Motivation of the proposed AdvRF. (a) Classification-based representation learning easily embeds predefined category semantics into representations, causing visually similar cues across subcategories to appear dissimilar in feature space while visually dissimilar cues within the same subcategory appear similar. (b) Reconstruction-based representation learning allows for a focus on object details and the contextual semantics of both the object and its parts, based on their appearance, thereby capturing category-agnostic representations. This facilitates a deeper understanding of unseen categories by enabling precise appearance modeling using category-agnostic visual descriptions." + ], + "image_footnote": [], + "bbox": [ + 516, + 349, + 901, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dings to enhance the performance of FGIR.", + "bbox": [ + 511, + 625, + 802, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current work on FGIR tasks [19, 20, 23, 26, 35, 45] has achieved promising results by introducing metric constraints or designing localization schemes to capture diverse visual discrepancies from visually similar objects. However, these approaches fundamentally couple discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations. As illustrated in Fig. 1a, this coupling leads to two paradoxical phenomena: (1) visually similar bird heads from different subcategories are represented differently due to category divergence, whereas (2) dissimilar bird heads and wings within the same subcategory are clustered due to sharing the same category. Consequently, the model's representation power relies heavily on predefined category semantics, struggling to interpret unseen subcategories based on actual visual appearances.", + "bbox": [ + 511, + 642, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fortunately, object reconstruction tasks inherently focus", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: hjli@sdust.edu.cn", + "bbox": [ + 107, + 886, + 334, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3080", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on pixel-level fidelity and contextual coherence, requiring the model to preserve appearance details and their contextual semantics without embedding category-specific semantics [9, 12]. As demonstrated in Fig. 1b, this property enables the reconstruction model's encoder to generate similar feature representations for visually similar parts across different subcategories. Building on this insight, we consider to leverage the inherent category-agnostic representation capability of object reconstruction models to eliminate dependencies on predefined category semantics in visual discrepancy modeling. However, naively training a fine-grained reconstruction model proves inadequate for FGIR, as it focuses on modeling the entire image appearance, including irrelevant background information, instead of emphasizing key visual discrepancies. Therefore, one natural question arises: is it possible to synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations?", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer this, we propose AdvRF, a novel Adversarial Reconstruction Feedback framework that reformulates FGIR as a visual discrepancy reconstruction task. AdvRF synergizes a retrieval model and a reconstruction model within an adversarial learning paradigm inspired by Generative Adversarial Networks (GANs) [8]. The framework operates through alternating optimization: while the retrieval model pinpoints subtle discrepancies within objects, the reconstruction model represents these discrepancies with category-agnostic representations, creating a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Specifically, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. This adversarial interplay progressively achieves precise localization of visual discrepancies, while eliminating predefined category semantics in discrepancy modeling. For efficient deployment, AdvRF distills category-agnostic discrepancy representations purified by the reconstruction model into the retrieval model through lightweight knowledge distillation.", + "bbox": [ + 91, + 395, + 482, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contribution are summarized below:", + "bbox": [ + 109, + 744, + 415, + 757 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to reformulate FGIR tasks as a visual discrepancy reconstruction process, improving its generalization.", + "- An adversarial reconstruction feedback framework, i.e., AdvRF, is proposed. AdvRF establishes an adversarial pipeline to alternately train the retrieval and reconstruction models, effectively capturing category-agnostic discrepancies essential for representing unseen categories.", + "- Extensive experiments show that our AdvRF achieves" + ], + "bbox": [ + 91, + 763, + 478, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "state-of-the-art performance on widely-used fine-grained and coarse-grained retrieval benchmarks.", + "bbox": [ + 527, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 138, + 653, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-grained image retrieval can be broadly categorized into two groups. The first group, localization-based scheme, focuses on localizing object or its details to facilitate the retrieval of visually similar objects [23, 33, 44, 51]. CaRA [38] implements a rectified activation strategy to enhance the localization of object details. The second group, metric-based schemes, seeks to learn an embedding space where similar examples are closely aligned, while dissimilar ones are pushed apart [14, 15, 25, 49, 50]. NIA [26] enforces unique translatability of samples from their respective class proxies to bring the distance of samples with the same subcategory closer. Unlike these approaches, PLEor [34] incorporates category-specific language descriptions based on the CLIP model to guide the model in representing visual discrepancies. However, such methods, which rely on representing similar objects through predefined categories, unintentionally embed category-specific semantic information into the retrieval representations, thereby limiting their ability to generalize to unseen categories. To address this issue, we propose AdvRF, which incorporates an adversarial pipeline to acquire category-agnostic discrepancies derived from the reconstructed features of objects.", + "bbox": [ + 511, + 164, + 903, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial learning has gained extensive application across multiple domains, including generative adversarial networks (GANs) [4, 6, 43], person re-identification [53], and domain adaptation [7, 11]. The essence of adversarial learning lies in minimizing distributional discrepancies between target and source domains by counteracting adversarial attacks. Furthermore, existing work such as OpenGAN [16] seeks to enhance the model's generalization to unseen categories by training a robust open-vs-closed discriminator that distinguishes between synthesized fake data and real data. In contrast to OpenGAN, which improves generalization through the generation of fake samples, AdvRF enhances generalization by capturing class-agnostic discrepancies via adversarial learning between retrieval and reconstruction models from a feature perspective.", + "bbox": [ + 511, + 498, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Adversarial Reconstruction Feedback", + "text_level": 1, + "bbox": [ + 511, + 739, + 852, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The core of AdvRF, as depicted in Fig. 2, lies in its innovative collaboration between a retrieval model and a reconstruction model. The retrieval model pinpoints discrepancy locations within objects, while the reconstruction model, leveraging these locations, generates category-agnostic representations of these discrepancies. This is achieved through adversarial feedback learning, where the two models create a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Fur", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3081", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6afc28cc1f19ff5b6790c98a698953be76ff960f1e35f579b373225dc55899fd.jpg", + "image_caption": [ + "Figure 2. Detailed illustration of adversarial reconstruction feedback. See §3 for more details." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 903, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thermore, these discrepancies are back into the retrieval model via knowledge distillation, significantly boosting computational efficiency during test.", + "bbox": [ + 89, + 373, + 483, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 434, + 294, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-grained Retrieval Model $\\mathcal{R}$ . It is designed to extract robust object representations and generate the final retrieval embeddings. Given an input image $\\mathbf{X}$ , we denote $\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}$ as the $C$ -dimensional feature tensor with $H \\times W$ spatial dimensions, encoded by a backbone network $\\mathbf{F} = \\mathcal{R}(\\mathbf{X})$ . Traditionally, the most prevalent approach for fine-grained retrieval tasks embeds the full feature tensor $\\mathbf{F}$ through global average pooling (GAP, $g(\\cdot)$ ), which computes the mean across the $H \\times W$ spatial planes. This process yields the final retrieval embeddings $\\mathbf{E}_{\\mathbf{R}} \\in \\mathbb{R}^{C} = g(\\mathbf{F})$ . Importantly, our proposed AdvRF framework maintains computational efficiency, as only the retrieval model is needed at inference time.", + "bbox": [ + 88, + 458, + 482, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-grained Reconstruction Model $\\mathcal{G}$ . It consists of an encoder $\\mathcal{G}_E$ and a decoder $\\mathcal{G}_D$ , with the detailed architecture outlined below. For the encoder $\\mathcal{G}_E$ , we use a lightweight network, specifically ResNet34, with pooling layers omitted from the last two blocks in our experimental setup. To emphasize subtle discrepancies within fine-grained objects, the encoder aggregates feature maps from the last three blocks into a final representation, enabling efficient encoding of input signals. This aggregation process applies $1 \\times 1$ convolutional layers, which distills subtle details from low-level features while capturing semantic information from high-level representations.", + "bbox": [ + 88, + 656, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The decoder, $\\mathcal{G}_D$ , adopts a U-Net architecture [24] with eight downsampling blocks, seven upsampling blocks, and a final colorization block to ensure high-fidelity reconstruction of the inputs. Each downsampling block consists of", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a $4 \\times 4$ convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation. Similarly, each upsampling block contains a transposed convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation.", + "bbox": [ + 511, + 373, + 906, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Category-agnostic Discrepancy Acquisition", + "text_level": 1, + "bbox": [ + 511, + 459, + 880, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that the reconstruction model is primarily designed to recover pixel-level details from inputs, it faces difficulties in identifying which visual cues represent meaningful discrepancies for identifying visually similar objects. To address this, we design a category-agnostic discrepancy acquisition module that captures visual discrepancies through category-agnostic representations produced by the reconstruction model.", + "bbox": [ + 511, + 481, + 906, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discrepancy Decoupling. Since the retrieval model is designed to capture visual discrepancies within images, we leverage its representation as a foundation to identify and localize both visual discrepancy and non-discrepancy regions. Therefore, we map the retrieval representation $\\mathbf{F}$ into a pattern map $\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}$ , which serves to indicate the locations of discrepancies. $\\hat{\\mathbf{M}}$ can be generated by a lightweight generator $\\mathcal{T}(\\cdot)$ as below:", + "bbox": [ + 511, + 602, + 905, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} = \\sigma (\\mathcal {T} (\\mathbf {F})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 733, + 903, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the sigmoid activation function, $\\mathcal{T}(\\cdot)$ is a convolutional layer with kernel size 1.", + "bbox": [ + 511, + 763, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, optimizing the pattern map via backpropagation of the loss function primarily tunes the parameters of the lightweight generator, with minimal impact on the retrieval model's parameters. Consequently, the output of the retrieval model may still include non-discriminative information, potentially impairing retrieval performance. To ensure the retrieval model is optimized to focus exclu", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3082", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sively on visual discrepancies, we introduce an mean generator with the same architecture to produce a refined pattern map. In this way, Eq. 1 can be rewritten as:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {M}} = \\sigma (\\operatorname {E} (\\mathcal {T}) (\\mathbf {F})), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 143, + 482, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{E}(\\mathcal{T})$ denotes the mean generator without learnable parameters. Its parameters can be updated in a temporal average manner. Concretely, at the $t$ -th iteration, parameters $\\operatorname{E}(\\mathcal{T})$ are accumulated by:", + "bbox": [ + 89, + 167, + 483, + 229 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {E} ^ {(t)} (\\mathcal {T}) [ \\theta ] = (1 - \\delta) \\cdot \\operatorname {E} ^ {(t - 1)} (\\mathcal {T}) [ \\theta ] + \\delta \\cdot \\theta , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 234, + 482, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{E}^{(t)}(\\mathcal{T})[\\theta]$ and $\\operatorname{E}^{(t-1)}(\\mathcal{T})[\\theta]$ denote the parameters of the mean generator in current iteration and last iteration, respectively. The mean generator is initialized as $\\operatorname{E}^{(0)}(\\mathcal{T})[\\theta] = \\theta$ . The hyper-parameter $\\delta$ is the updating ratio within the range of $(0, 1]$ .", + "bbox": [ + 89, + 258, + 483, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Category-agnostic Discrepancy Transfer. We feed the input image $\\mathbf{X}$ into the encoder $\\mathcal{G}_E$ of the reconstruction model to obtain the category-agnostic representation $\\hat{\\mathbf{F}}_{\\mathbf{I}} = \\mathcal{G}_{E}(\\mathbf{X})\\in \\mathbb{R}^{C\\times H\\times W}$ , and then resize both the pattern map $\\hat{\\mathbf{M}}$ and $\\hat{\\mathbf{F}}_{\\mathbf{I}}$ to match the size of the original input images, ensuring high fidelity during the subsequent reconstruction process. Using the amplified pattern map $\\mathbf{M}$ , we decompose the amplified image representation $\\mathbf{F}_{\\mathbf{I}}$ into the category-agnostic visual discrepancies $\\mathbf{C}_{\\mathbf{A}}$ and the non-discrepancy representation $\\mathbf{C}_{\\mathbf{R}}$ as follows:", + "bbox": [ + 89, + 335, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {\\mathbf {A}} = \\mathbf {F} _ {\\mathbf {I}} \\odot \\mathbf {M}, \\quad \\mathbf {C} _ {\\mathbf {R}} = \\mathbf {F} _ {\\mathbf {I}} \\odot (1 - \\mathbf {M}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 494, + 482, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\odot$ denotes element-wise multiplication.", + "bbox": [ + 89, + 518, + 393, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering that employing both models concurrently is time-consuming and memory-intensive for retrieval evaluation, we design a category-agnostic discrepancy parameterization constraint. Formally, the category-agnostic discrepancy serves as the supervisory signal and eliminates the need for gradient updates to adjust retrieval representations:", + "bbox": [ + 89, + 534, + 483, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {P}} = \\left| \\left| \\mathbf {E} _ {\\mathbf {R}} - g \\left(\\mathbf {C} _ {\\mathbf {A}}\\right) \\right| \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 638, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $||\\cdot||$ refers to the Frobenius norm. This constraint directly optimizes the parameters within the retrieval model, ensuring that its output representations exclusively capture category-agnostic visual discrepancies while eliminating category-specific semantics. As a result, the retrieval model gains the ability to pinpoint object discrepancies and characterize them based solely on their visual appearance, even when encountering unseen categories. Importantly, the contextual semantics of the discrepancies are still preserved, as the reconstruction process inherently considers the contextual semantics of both the object and its parts.", + "bbox": [ + 89, + 659, + 483, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adversarial Feedback Learning", + "text_level": 1, + "bbox": [ + 89, + 833, + 372, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To acquire category-agnostic discrepancy representations, we introduce an adversarial feedback learning strategy inspired by GANs, which synergizes a retrieval model with", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a reconstruction model. Through adversarial feedback interplay, the reconstruction model exploits residual discrepancies overlooked by the retrieval model for reconstruction, thereby challenging the retrieval model and enhancing object recovery. Concurrently, the retrieval model dynamically refines its discrepancy localization based on feedback from the reconstruction model, further challenging the reconstruction model and improving discrepancy localization.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction Feedback. Given the category-agnostic discrepancies $\\mathbf{C}_{\\mathbf{A}}$ and the non-discrepancy representation $\\mathbf{C}_{\\mathbf{R}}$ , we input them separately into the decoder $\\mathcal{G}_D(\\cdot)$ of the reconstruction model to obtain the reconstructed regions:", + "bbox": [ + 511, + 212, + 903, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 281, + 903, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}$ and $\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}$ represent the reconstructed images, respectively. Considering that the reconstruction model utilizes the residual discrepancies overlooked by the retrieval model to reconstruct discrepancies, we should evaluate the quality of discrepancy reconstruction using the image generated by non-discrepancy representation $\\mathbf{C}_{\\mathbf{R}}$ . Therefore, we minimize the difference between the reconstructed image $\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}$ and the input image $\\mathbf{X}$ , within the discrepancy region:", + "bbox": [ + 511, + 306, + 903, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} ^ {R} = \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}}\\right) \\right\\|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 441, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathbf{M}$ acts as a spatial attention map to localize discrepancies, maintaining intrinsic invariance across feature/pixel spaces. It can be regarded as a soft validation mask to direct the reconstruction model's focus to critical discrepancy regions during regeneration.", + "bbox": [ + 511, + 465, + 903, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, the reconstruction model also leverages the residual non-discrepancy information from the category-agnostic representation $\\mathbf{C}_{\\mathbf{A}}$ to reconstruct the non-discrepancy regions. Hence, we also impose another reconstruction constraint for minimizing their differences between the reconstructed image $\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}$ and the original images $\\mathbf{X}$ :", + "bbox": [ + 511, + 541, + 905, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} ^ {A} = \\left| \\left| (1 - \\mathbf {M}) \\odot (\\mathbf {X} - \\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}}) \\right| \\right|, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 645, + 903, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(1 - \\mathbf{M})$ indicates the non-discrepancy region.", + "bbox": [ + 511, + 667, + 867, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the total loss for training the reconstruction model could be integrated as:", + "bbox": [ + 511, + 683, + 903, + 713 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} ^ {R e c F} = \\mathcal {L} _ {\\mathcal {G}} ^ {A} + \\mathcal {L} _ {\\mathcal {G}} ^ {R}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 720, + 903, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent the retrieval model from optimizing towards less accurate discrepancy localization, we freeze its parameters and exclusively back-propagate the loss gradients to the reconstruction model, including both its encoder and decoder.", + "bbox": [ + 511, + 750, + 903, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Retrieval Feedback. Unlike the reconstruction feedback, which aims to improve the reconstruction ability of the reconstruction model, the retrieval feedback focuses on precisely locating discrepancies within objects. It leverages the reconstruction model as an evaluator to assess the accuracy of discrepancy localization. The above process is much", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3083", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "like how the generator receives feedback from the discriminator in GANs. Formally, we also need to feed $\\mathbf{C}_{\\mathbf{A}}$ and $\\mathbf{C}_{\\mathbf{R}}$ into $\\mathcal{G}_D$ to obtain the reconstructed images as below:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 146, + 482, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}_{\\mathbf{A}}^{\\mathcal{R}}$ and $\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}$ represent the reconstructed images, respectively.", + "bbox": [ + 89, + 174, + 482, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This optimization strategy is similar to optimizing the generator based on feedback from the discriminator. In other words, the discrepancy localization provided by the retrieval model makes it challenging for the reconstruction model to accurately reconstruct the object. Therefore, given the reconstructed images $\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}$ and the original image $\\mathbf{X}$ on the discrepancy regions, we impose a reconstruction constraint to maximize their differences:", + "bbox": [ + 89, + 205, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R}} ^ {R} = - \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 335, + 482, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, the minus sign means that minimizing the loss leads the framework to maximize the difference between the $\\mathbf{X}$ and $\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}$ . Similarly, we impose an additional reconstruction constraint to maximize the differences between the reconstructed and original non-discrepancy regions:", + "bbox": [ + 89, + 366, + 483, + 443 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R}} ^ {A} = - \\left\\| \\left(1 - \\mathbf {M}\\right) \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 450, + 482, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the total loss for training the retrieval model to produce the accurate pattern map is:", + "bbox": [ + 89, + 481, + 482, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R}} ^ {R e t F} = \\mathcal {L} _ {\\mathcal {R}} ^ {A} + \\mathcal {L} _ {\\mathcal {R}} ^ {R}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 520, + 482, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, to prevent compromising the reconstruction capabilities of the reconstruction model, we freeze its parameters and exclusively backpropagate the loss gradients to the retrieval model.", + "bbox": [ + 89, + 551, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Alternating Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 619, + 356, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AdvRF implements an iterative alternating protocol where the reconstruction and retrieval models cyclically enhance each other's improvement. In each training epoch, the following steps are performed:", + "bbox": [ + 89, + 642, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Reconstruction Updating Phase (RecU): The parameters of the retrieval model $\\Theta_{\\mathcal{R}}$ are frozen, and the reconstruction model $\\Theta_{\\mathcal{G}}$ is updated using the loss function:", + "bbox": [ + 91, + 702, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {G}} ^ {R e c U} [ \\Theta_ {\\mathcal {G}} ] = \\alpha \\cdot \\mathcal {L} _ {\\mathcal {G}} ^ {R e c F}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 761, + 482, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2. Retrieval Updating Phase (RetU): Freezing $\\Theta_{\\mathcal{G}}$ , we then refine $\\Theta_{\\mathcal{R}}$ using multi-task learning with:", + "bbox": [ + 89, + 785, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {R}} ^ {R e t U} \\left[ \\Theta_ {\\mathcal {R}} \\right] = \\mathcal {L} _ {\\mathcal {C}} + \\beta \\cdot \\mathcal {L} _ {\\mathcal {P}} + \\gamma \\cdot \\mathcal {L} _ {\\mathcal {R}} ^ {R e t F}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 825, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This alternating training strategy repeats until joint convergence, with $\\alpha$ , $\\beta$ and $\\gamma$ dynamically balancing task-specific gradients.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of performance and efficiency on CUB-200-2011 using different combinations of constraints. The first row indicates that we use classification-based feedback as supervision, to replace the proposed AdvRF for comparison. \"T\" is the time of extracted retrieval embeddings.", + "bbox": [ + 511, + 89, + 906, + 159 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c6cffc2cb8e45abaeaaac7be544088372a8a9f5c32a616c69ff7b396058b0c8d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\mathcal{G}}^{RecF} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{RetF} \\)\\( \\mathcal{L}_{\\mathcal{P}} \\)Performance
\\( \\mathcal{L}_{\\mathcal{G}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{G}}^{R} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{R} \\)R@1T
66.3%21.1ms
✓✓73.7%36.7ms
✓✓73.4%36.7ms
✓✓✓✓76.8%36.7ms
✓✓✓✓✓76.6%21.1ms
", + "bbox": [ + 526, + 169, + 893, + 290 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e425d2db622c116d3266e3779a489ee714712f2a4fa2267c1f99d72dc7ab2f08.jpg", + "table_caption": [ + "Table 2. Evaluation results of retrieval performance on CUB-200-2011 dataset with diverse feedback." + ], + "table_footnote": [], + "table_body": "
Feedback TypeRecall@1
Classification-based Feedback66.3%
Reconstruction-based Feedback62.2%
Our AdvRF76.6%
", + "bbox": [ + 526, + 342, + 893, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 444, + 648, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 470, + 702, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. CUB-200-2011 [3] consists of 200 bird species. We use the first 100 subcategories (5,864 images) for training and the consists of (5,924 images) for testing. The Stanford Cars [17] includes 196 car models. Similarly, we use the first 98 classes, which contain 8,054 images, for training and the remaining classes, which contain 8,131 images, for testing. Finally, FGVC Aircraft [22] is split into first 50 classes, containing 5,000 images, for training and the remaining 50 classes with 5,000 images, for testing. Stanford Online Products (SOP) [27] is divided into the 11, 318 subcategories (59, 551 images) in training, and the rest 11, 316 classes (60, 502 images) in testing. This split ensures no category overlap between training and testing sets, where all testing categories are strictly unseen during training to evaluate cross-category generalization.", + "bbox": [ + 511, + 492, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our retrieval model is built upon a ResNet-50 backbone [10] initialized with ImageNet pretrained weights. Input images are resized to $256 \\times 256$ pixels and randomly cropped to $224 \\times 224$ during training. We employ Stochastic Gradient Descent with an initial learning rate of $10^{-5}$ , weight decay of 0.0001, and momentum of 0.9, using a batch size of 32 distributed across four NVIDIA A100 GPUs. To enhance robustness, standard data augmentations including random cropping, horizontal flipping, and color jittering are applied. The learning rate follows an exponential decay schedule (factor=0.9 every 5 epochs) over 200 training epochs, ensuring stable convergence while mit", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3084", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b1e2fce6673487c84940063c14b3f74dfa0a3e160f04e8728b36abbee8a93687.jpg", + "table_caption": [ + "Table 3. Compared with competitive methods on CUB-200-2011, Stanford Cars 196 and FGVC Aircraft datasets. \"Arch\" represents the architecture of utilizing backbone network. \"R50\" denotes Resnet50 [10] backbone network." + ], + "table_footnote": [], + "table_body": "
MethodArchCUB-200-2011Stanford Cars 196FGVC Aircraft
124812481248
SCDA TIP17 [44]R5057.370.281.088.448.360.271.881.856.567.777.685.7
CRL IJCAI18 [51]R5062.574.282.989.757.869.178.686.661.171.680.988.2
HDCL IJON21 [47]R5069.579.686.892.484.490.194.196.571.181.088.393.3
CEP ECCV20 [2]R5069.279.286.991.689.393.996.698.1----
CaRA TPAMI24 [38]R5073.982.289.493.694.196.998.298.984.390.494.296.3
FRPT AAAI23 [36]R5074.383.789.894.391.195.197.398.677.685.791.495.6
DGCRL AAAI19 [52]R5067.979.186.291.875.983.989.794.070.179.688.093.0
DAS ECCV22 [20]R5069.279.387.192.687.893.296.097.9----
CBML TPAMI23 [13]R5069.980.487.292.588.192.695.497.4----
NIR CVPR22 [26]R5070.580.6--89.193.4------
HIST CVPR22 [19]R5071.481.188.1-89.693.996.4-----
IDML TPAMI24 [32]R5070.780.2--90.694.5------
HSE ICCV23 [46]R5070.680.187.1-89.693.896.0-----
PNCA++ ECCV20 [31]R5072.282.089.293.590.194.597.098.4----
PLEor CVPR23 [34]R5074.884.591.394.994.496.998.398.986.391.795.196.7
Our AdvRFR5076.685.391.795.094.997.298.698.988.092.595.596.9
", + "bbox": [ + 91, + 128, + 903, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "igating overfitting to category-specific patterns.", + "bbox": [ + 89, + 453, + 403, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocols. We evaluate the retrieval performance by $Recall@K$ with cosine distance, which is average recall scores over all query images in the test set and strictly follows the setting in previous work [30]. Specifically, for each query, our model returns the top $K$ similar images. In the top $K$ returning images, the score will be 1 if there exists at least one positive image, and 0 otherwise.", + "bbox": [ + 88, + 469, + 482, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Experiments", + "text_level": 1, + "bbox": [ + 89, + 589, + 295, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficacy of various constraints. The proposed AdvRF, as described in Sec. 3, is optimized through a combination of four loss functions, each playing a distinct role in guiding AdvRF to capture category-agnostic discrepancies. Tab. 1 presents quantitative comparisons across various constraint combinations. Initially, we use ResNet-50 [10] with only the classification loss $\\mathcal{L}_{\\mathcal{C}}$ , achieving $66.3\\%$ Recall@1 accuracy on the CUB-200-2011 dataset. By introducing the reconstruction feedback loss $\\mathcal{L}_{\\mathcal{G}}^{RecF}$ and the retrieval feedback loss $\\mathcal{L}_{\\mathcal{R}}^{RetF}$ , we synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models, obtaining a performance of $76.8\\%$ . When removing $\\mathcal{L}_{\\mathcal{G}}^{A}$ or $\\mathcal{L}_{\\mathcal{G}}^{R}$ , the reconstruction model's sensitivity to residual discrepancies from the retrieval model decreases, leading to reduced performance. Additionally, when removing $\\mathcal{L}_{\\mathcal{R}}^{A}$ or $\\mathcal{L}_{\\mathcal{R}}^{R}$ , the retrieval model struggles to accurately evaluate the discrepancy localization, which also results in decreased performance. Finally, for efficient deployment, we introduce", + "bbox": [ + 91, + 613, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a category-agnostic discrepancy parameterization loss $\\mathcal{L}_{\\mathcal{P}}$ to distill category-agnostic discrepancy representations into the retrieval model, enabling real-time retrieval without acceptable accuracy loss.", + "bbox": [ + 511, + 452, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different types of feedback. Tab. 2 presents a comparison of the retrieval performance for retrieving visually similar objects using different types of feedback. Directly using classification-based feedback fundamentally couples discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations, resulting in a performance of $66.3\\%$ . When directly training a reconstruction model and using its encoder outputs for retrieval, the model focuses on modeling the entire image appearance, including irrelevant background information, rather than emphasizing key visual discrepancies, which leads to a lower performance. In contrast, our AdvRF effectively combines the advantages of category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations, thus achieving a performance of $76.6\\%$ .", + "bbox": [ + 511, + 515, + 908, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with the State-of-the-Arts", + "text_level": 1, + "bbox": [ + 511, + 801, + 852, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fine-grained image retrieval. Our AdvRF demonstrates superior performance across all three FGIR benchmarks (CUB-200-2011, Stanford Cars-196, FGVC Aircraft), significantly outperforming existing state-of-the-art methods (Tab. 3). Localization-based approaches (e.g., CaRA [38],", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3085", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/df6b0c3e420ca27a8600116ac444198b19eafdba7ee860bd61612fefacce342f.jpg", + "table_caption": [ + "Table 4. Recall@k for $\\mathrm{k} = 1,{10},{100},{1000}$ on Stanford Online Products (SOP)." + ], + "table_footnote": [], + "table_body": "
Method1101001000
MS [40] CVPR1978.290.596.098.7
NSM [48] BMVC2179.591.596.7-
DCML [49] CVPR2179.890.895.895.8
ETLR [14] CVPR2179.891.196.3-
MRML-PA [50] ICCV2179.990.796.1-
HSE [46] ICCV2380.091.496.3-
DAS [20] ECCV2280.691.896.799.0
CEP [2] ECCV2081.191.796.398.8
PNCA++ [31] ECCV2081.492.496.999.0
IBC [28] ICML2181.491.395.9-
HIST [19] CVPR2281.492.096.7-
IDML [32] TPAMI2481.592.354.851.3
CaRA [38] TPAMI2482.492.697.099.0
Our AdvRF84.293.797.699.1
", + "bbox": [ + 94, + 128, + 478, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FRPT [36]) and metric-learning frameworks (e.g., IDML [32], HIST [19]) demonstrate effectiveness in capturing fine-grained visual discrepancies. However, their inherent coupling of discrepancy modeling with predefined category supervision embeds category-specific semantics into retrieval representations, fundamentally limiting performance breakthroughs in unseen category generalization. Therefore, AdvRF introduces an adversarial reconstruction mechanism that decouples discrepancy modeling from categorical supervision through iterative training between the retrieval and reconstruction models. This mechanism explicitly grounds visual discrepancies in appearance cues rather than seen category semantics, thereby achieving significant performance gains in generalization to unseen categories.", + "bbox": [ + 88, + 400, + 482, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse-grained image retrieval. To further validate AdvRF's generalization capability, we evaluate it on a largescale coarse-grained benchmark, i.e., Stanford Online Products, in Tab. 4. The framework's synergy between the precise discrepancy localization of the retrieval model and the category-agnostic representation learning capability of the reconstruction model can represent objects using category-agnostic description. Hence, AdvRF not only captures subtle inter-class differences in fine-grained settings but also maintains robustness to coarse-grained semantic gaps, thus obtaining a better performance on SOP.", + "bbox": [ + 88, + 612, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 787, + 259, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Investigation on the updating ratio $\\delta$ . Tab. 5 showcases the accuracy of various updating ratios in Eqn. 3. Notably, as the ratio increases, retrieval performance declines, indicating that excessive updates to the generator cause it to rely too heavily on the current learning parameters, making it harder to fine-tune the parameters within the retrieval", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/07cb9830013daaf9746a3ef20d9244e033f0068428f41c7442381d8d28d63960.jpg", + "table_caption": [ + "Table 5. Evaluation results on CUB-200-2011 of light-weight generator trained with different updating ratio $\\delta$ in Eqn. 3." + ], + "table_footnote": [], + "table_body": "
Ratio δ0.10.20.40.60.8
R@175.4%76.6%74.8%74.4%73.9%
", + "bbox": [ + 517, + 128, + 901, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1927f720658cd549a4cb0fc9cb9deb5c298555be3b3c6f3286e2fe4a8b82e37.jpg", + "table_caption": [ + "Table 6. Results comparing to various pattern maps based on Recall@K on CUB-200-2011." + ], + "table_footnote": [], + "table_body": "
MethodR@1R@2R@4R@8
CAM [29]69.8%79.7%84.2%91.6%
Bounding box73.9%82.6%90.5%94.2%
Our AdvRF76.6%85.3%91.7%95.0%
", + "bbox": [ + 514, + 229, + 903, + 309 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a8d73637416e8e6c6439e5410e2b92ad277fa2b866f1943828d6de973ea5a78.jpg", + "table_caption": [ + "Table 7. Effect on the reconstruction ability with different reconstructed manners on CUB-200-2011." + ], + "table_footnote": [], + "table_body": "
Reconstruction MannerRecall@1Recall@2
Non-adversarial Recon.72.6%82.4%
Adversarial Recon.76.6%+4.085.3%+2.9
", + "bbox": [ + 514, + 369, + 906, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model. Conversely, a lower ratio preserves sensitivity to prior knowledge, forcing the generator to rely more on the retrieval model's features, thereby providing more accurate features by modifying the retrieval model's parameters.", + "bbox": [ + 511, + 460, + 906, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual discrepancy localization with different manners. Switching the visual discrepancy localization method provides insights into acquiring category-agnostic discrepancies. As Tab. 6 indicates, shifting from our discrepancy decoupling strategy to a fixed localization method causes a significant performance drop, nearing the accuracy of fine-tuning a pre-trained model. Specifically, using class activation maps or dataset-provided bounding boxes for localization often leads to imprecise results, including background and missing critical discrepancies. In contrast, our decoupling strategy allows AdvRF to accurately locate visual discrepancies and generate precise category-agnostic representations, consistently improving performance.", + "bbox": [ + 511, + 521, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial learning between the retrieval and reconstruction models. In AdvRF, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to enhance its reconstruction capability. When the reconstruction model only uses the discrepancies localized by the retrieval model, rather than its residual discrepancies, it struggles to provide more comprehensive feedback, resulting in a performance of $72.6\\%$ as shown in the first row of Tab. 7. This implicitly indicates that adversarial learning creates a self-reinforcing cycle, where both", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3086", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bbc69070db3eff6bbf95fabe805e8f409b74d656e184074e54a20544e96c7be.jpg", + "image_caption": [ + "Figure 3. Analyses of hyper-parameters $\\alpha$ , $\\beta$ and $\\gamma$ in Eq. 14 and 15. Results denote Recall@1 accuracy on CUB-200-2011." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8bc19777535b8954896cc7c7421fcf894950d8fc63c78ba94880735a029b1e87.jpg", + "image_caption": [ + "(a)", + "Figure 4. Visualizations of pattern maps using different feedback: (a) inputs, (b) classification-based feedback, (c) non-discrepancy reconstruction-based feedback, (d) discrepancy reconstruction-based feedback, and (e) our AdvRF." + ], + "image_footnote": [], + "bbox": [ + 94, + 300, + 173, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19ee7885f4c1bb97e1b05b945cdcf269bbba059539d4b85fe0098c3e09a201f5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 173, + 300, + 248, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0014a467debb4bc07c8aa489a90e66c93bb3a743d161366c8a6bf0ebee4ad766.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 250, + 300, + 325, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3fa0953cf128b86d4ca2e1408bbc0fa17073c2b91070b321a418a9ffa085577.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 326, + 300, + 401, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6664c4eea81105b15aeb67f5393612b9abb38de1587e1ce9d781874d58926261.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 401, + 300, + 478, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the retrieval and reconstruction models iteratively challenge and reinforce each other.", + "bbox": [ + 89, + 580, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyper-parameter analyses. We conduct sensitivity analyses of the hyperparameters in Eq. 14 and 15, with evaluation results presented in Fig. 3. The performance of our AdvRF shows slight sensitivity to variations in $\\alpha$ , $\\beta$ , $\\gamma$ , and $\\delta$ . In our experiments, the default values are set to $\\alpha = 0.7$ , $\\beta = 0.5$ , $\\gamma = 0.6$ , respectively.", + "bbox": [ + 89, + 612, + 482, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of pattern maps with various feedback. Fig. 4 illustrates the impact of various feedback signals on discrepancy localization. When using classification-based feedback, the model struggles to accurately localize discrepancies. However, formulating FGIR as a visual discrepancy reconstruction task enhances discrepancy localization through reconstruction-based feedback. Notably, limited reconstruction signals, such as using only discrepancy-based feedback ( $\\mathbf{C}_{\\mathbf{A}}$ in Eq. 4) or non-discrepancy-based feedback ( $\\mathbf{C}_{\\mathbf{R}}$ in Eq. 4), may cause pattern maps to overlook certain discrepancies. Our results suggest that combining comprehensive reconstruction feedback creates a self-reinforcing cycle, where both the retrieval and reconstruc", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f55175f45a83489e6e34a7c370f3bc6ad3b59090ac3607bf8d1b2aceecbd651.jpg", + "image_caption": [ + "(a) Category-related Embeddings" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 676, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c57401f9b835fed62a18d0654ee57290683d405b28db86189c1465fcd03fc9bc.jpg", + "image_caption": [ + "Figure 5. Evaluating the efficacy of category-related embeddings from baseline [10] versus category-agnostic embeddings from our AdvRF, with their similarity in grid formats." + ], + "image_footnote": [], + "bbox": [ + 679, + 90, + 705, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/df6e6371e14f4ac7bc37d1e9769bb961f37c16cf38498095f529b0cd175b9f39.jpg", + "image_caption": [ + "(b) Category-agnostic Embeddings" + ], + "image_footnote": [], + "bbox": [ + 707, + 90, + 893, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion models iteratively challenge and strengthen each other. Analysis of category-agnostic discrepancies. We employ an indirect method to interpret category-agnostic discrepancies by comparing the similarities between category-agnostic retrieval embeddings produced by our AdvRF and category-related retrieval embeddings generated by the classification-based feedback illustrated in Fig. 1a. As illustrated in Fig. 5, we compute these similarities for the ten high-similarity images across five novel subcategories. Our analysis reveals that category-agnostic embeddings effectively highlight nearest sample pairs that belong to the same subcategory. In contrast, category-related retrieval embeddings struggle to identify these high-similarity images, as they tend to capture visual discrepancies associated with the semantics of base categories. Overall, our AdvRF successfully learns category-agnostic descriptions for unseen subcategories by formulating FGIR tasks as a visual discrepancy reconstruction process.", + "bbox": [ + 511, + 303, + 906, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce AdvRF, which acquires category-agnostic visual discrepancies by formulating FGIR as a visual discrepancy reconstruction task. AdvRF designs an adversarial pipeline: the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. As a result, AdvRF precisely localizes visual differences and encodes them into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Importantly, our algorithm is end-to-end trainable and achieves state-of-the-art performance on the widely-used fine-grained and coarse-grained retrieval datasets.", + "bbox": [ + 511, + 613, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (NSFC) under Grant (No.61932020), and the Taishan Scholar Program of Shandong Province (tstp20221128).", + "bbox": [ + 511, + 840, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3087", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kenan E. Ak, Ashraf A. Kassim, Joo-Hwee Lim, and Jo Yew Tham. Learning attribute representations with localization for flexible fashion search. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7708-7717. Computer Vision Foundation / IEEE Computer Society, 2018. 1", + "[2] Malik Boudiaf, Jérôme Rony, Imtiaz Masud Ziko, Eric Granger, Marco Pedersoli, Pablo Piantanida, and Ismail Ben Ayed. A unifying mutual information view of metric learning: Cross-entropy vs. pairwise losses. In ECCV, pages 548–564. Springer, 2020. 6, 7", + "[3] Steve Branson, Grant Van Horn, Serge J. Belongie, and Pietro Perona. Bird species categorization using pose normalized deep convolutional nets. CoRR, abs/1406.2952, 2014. 5", + "[4] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Chojui Hsieh, and Boqing Gong. Robust and accurate object detection via adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16622-16631. Computer Vision Foundation / IEEE, 2021. 2", + "[5] Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed M. Elgammal. Link the head to the \"beak\": Zero shot learning from noisy text description at part precision. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6288-6297. IEEE Computer Society, 2017. 1", + "[6] Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xiaoming Wei, and Xiaolin Wei. Masked auto-encoders meet generative adversarial networks and beyond. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24449-24459. IEEE, 2023. 2", + "[7] Yaroslav Ganin and Victor S. Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1180-1189. JMLR.org, 2015. 2", + "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2", + "[9] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2021. 2", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778, 2016. 5, 6, 8", + "[11] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised domain adaptation. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "UT, USA, June 18-22, 2018, pages 1498-1507. Computer Vision Foundation / IEEE Computer Society, 2018. 2", + "[12] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In International Conference on 3D Vision, 3DV 2024, Davos, Switzerland, March 18-21, 2024, pages 31-41. IEEE, 2024. 2", + "[13] Shichao Kan, Zhiquan He, Yigang Cen, Yang Li, Vladimir Mladenovic, and Zhihai He. Contrastive bayesian analysis for deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7220-7238, 2023. 6", + "[14] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Embedding transfer with label relaxation for improved metric learning. In CVPR, pages 3967-3976. Computer Vision Foundation / IEEE, 2021. 2, 7", + "[15] ByungSoo Ko, Geonmo Gu, Han-Gyu Kim, and ByungSoo Ko. Learning with memory-based virtual classes for deep metric learning. In ICCV, pages 11772-11781. IEEE, 2021. 2", + "[16] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 793-802. IEEE, 2021. 2", + "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561, 2013. 5", + "[18] Haojie Li, Mingxuan Li, Qijie Peng, Shijie Wang, Hong Yu, and Zhihui Wang. Correlation-guided semantic consistency network for visible-infrared person re-identification. IEEE Trans. Circuits Syst. Video Technol., 34(6):4503-4515, 2024. 1", + "[19] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 212-222. IEEE, 2022. 1, 6, 7", + "[20] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. DAS: densely-anchored sampling for deep metric learning. In ECCV, pages 399–417. Springer, 2022. 1, 6, 7", + "[21] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1096-1104. IEEE Computer Society, 2016. 1", + "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5", + "[23] Olga Moskvyak, Frédéric Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Keypoint-aligned embeddings for image retrieval and re-identification. In WACV, pages 676–685. IEEE, 2021. 1, 2", + "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + ], + "bbox": [ + 93, + 92, + 903, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3088", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, pages 234-241. Springer, 2015. 3", + "[25] Karsten Roth, Timo Milbich, Björn Ommer, Joseph Paul Cohen, and Marzyeh Ghassemi. Simultaneous similarity-based self-distillation for deep metric learning. In ICML, pages 9095–9106. PMLR, 2021. 2", + "[26] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7410-7420. IEEE, 2022. 1, 2, 6", + "[27] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823. IEEE Computer Society, 2015. 5", + "[28] Jenny Seidenschwarz. Learning intra-batch connections for deep metric learning. In ICML, pages 9410-9421. PMLR, 2021. 7", + "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 7", + "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, pages 4004-4012. IEEE Computer Society, 2016. 6", + "[31] Eu Wern Teh, Terrance DeVries, Graham W. Taylor, and Graham. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, pages 448-464. Springer, 2020. 6, 7", + "[32] Chengkun Wang, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Introspective deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 46(4):1964-1980, 2024. 6, 7", + "[33] Shijie Wang, Zhihui Wang, Haojie Li, and Wanli Ouyang. Category-specific nuance exploration network for fine-grained object retrieval. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2513-2521. AAAI Press, 2022. 2", + "[34] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19381-19391. IEEE, 2023. 2, 6", + "[35] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Learning to parameterize visual attributes for open-set fine-grained retrieval. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Fine-grained retrieval prompt tuning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 2644-2652. AAAI Press, 2023. 6, 7", + "[37] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Semantic-guided information alignment network for fine-grained image recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1", + "[38] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Content-aware rectified activation for zero-shot fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 46(6):4366-4380, 2024. 2, 6, 7", + "[39] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Accurate fine-grained object recognition with structure-driven relation graph networks. Int. J. Comput. Vis., 132(1):137-160, 2024. 1", + "[40] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030. Computer Vision Foundation / IEEE, 2019. 7", + "[41] Zhuhui Wang, Shijie Wang, Haojie Li, Zhi Dou, and Jianjun Li. Graph-propagation based correlation learning for weakly supervised fine-grained image classification. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12289-12296. AAAI Press, 2020. 1", + "[42] Zhihui Wang, Shijie Wang, Shuhui Yang, Haojie Li, Jianjun Li, and Zezhou Li. Weakly supervised fine-grained image classification via gaussian mixture model oriented discriminative learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9746-9755. IEEE, 2020.", + "[43] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15054-15063. IEEE, 2022. 2", + "[44] Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, and Zhi-Hua Zhou. Selective convolutional descriptor aggregation for fine-grained image retrieval. IEEE Trans. Image Process., 26(6):2868-2881, 2017. 2, 6", + "[45] Xiu-Shen Wei, Yang Shen, Xuhao Sun, Peng Wang, and Yuxin Peng. Attribute-aware deep hashing with self-consistency for large-scale fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):13904-13920, 2023. 1", + "[46] Bailin Yang, Haoqiang Sun, Frederick W. B. Li, Zheng Chen, Jianlu Cai, and Chao Song. HSE: hybrid species embedding" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3089", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for deep metric learning. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11013-11023. IEEE, 2023. 6, 7", + "[47] Xianxian Zeng, Shun Liu, Xiaodong Wang, Yun Zhang, Kairui Chen, and Dong Li. Hard decorrelated centralized loss for fine-grained image retrieval. Neurocomputing, 453: 26-37, 2021. 6", + "[48] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In BMVC, page 91. BMVA Press, 2019. 7", + "[49] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In CVPR, pages 9320-9329. Computer Vision Foundation / IEEE, 2021. 2, 7", + "[50] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In ICCV, pages 12045-12054. IEEE, 2021. 2, 7", + "[51] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, Feiyue Huang, and Yanhua Yang. Centralized ranking loss with weakly supervised localization for fine-grained object retrieval. In IJCAI, pages 1226-1233. ijcai.org, 2018. 2, 6", + "[52] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Baochang Zhang, Yongjian Wu, and Feiyue Huang. Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer. In AAAI, pages 9291-9298. AAAI Press, 2019. 6", + "[53] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2138-2147. Computer Vision Foundation / IEEE, 2019. 2" + ], + "bbox": [ + 91, + 90, + 480, + 529 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3090", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_model.json b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7eaa5fafbdd75539c36e868795e08d1e4fdf869 --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "Adversarial Reconstruction Feedback for Robust Fine-grained Generalization" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.18, + 0.878, + 0.237 + ], + "angle": 0, + "content": "Shijie Wang\\(^{1}\\), Jian Shi\\(^{2}\\), Haojie Li\\(^{1*}\\) \n\\(^{1}\\)College of Computer and Engineering, Shandong University of Science and Technology, China \n\\(^{2}\\)School of Software, Dalian University of Technology, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Existing fine-grained image retrieval (FGIR) methods predominantly rely on supervision from predefined categories to learn discriminative representations for retrieving fine-grained objects. However, they inadvertently introduce category-specific semantics into the retrieval representation, creating semantic dependencies on predefined classes that critically hinder generalization to unseen categories. To tackle this, we propose AdvRF, a novel adversarial reconstruction feedback framework aimed at learning category-agnostic discrepancy representations. Specifically, AdvRF reformulates FGIR as a visual discrepancy reconstruction task via synergizing category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models. The reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. Consequently, the retrieval model localizes visual differences, while the reconstruction model encodes these differences into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Quantitative and qualitative evaluations demonstrate that our AdvRF achieves impressive performance on both widely-used fine-grained and coarse-grained datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Fine-grained image retrieval (FGIR) aims to retrieve visually similar subcategories, even those that were unseen during the training phase. It plays a vital role in numerous vision applications from fashion industry, e.g., retrieval of diverse types of clothes [1, 18, 21, 41], to environmental conservation, e.g., retrieving endangered species [5, 37, 39, 42]. Given its significance, a substantial body of research has focused on learning discriminative and generalizable embed" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.27, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.337, + 0.839, + 0.35 + ], + "angle": 0, + "content": "(a) Classification-based Representation Learning" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.35, + 0.903, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.418, + 0.844, + 0.431 + ], + "angle": 0, + "content": "(b) Reconstruction-based Representation Learning" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Motivation of the proposed AdvRF. (a) Classification-based representation learning easily embeds predefined category semantics into representations, causing visually similar cues across subcategories to appear dissimilar in feature space while visually dissimilar cues within the same subcategory appear similar. (b) Reconstruction-based representation learning allows for a focus on object details and the contextual semantics of both the object and its parts, based on their appearance, thereby capturing category-agnostic representations. This facilitates a deeper understanding of unseen categories by enabling precise appearance modeling using category-agnostic visual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.626, + 0.803, + 0.641 + ], + "angle": 0, + "content": "dings to enhance the performance of FGIR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Current work on FGIR tasks [19, 20, 23, 26, 35, 45] has achieved promising results by introducing metric constraints or designing localization schemes to capture diverse visual discrepancies from visually similar objects. However, these approaches fundamentally couple discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations. As illustrated in Fig. 1a, this coupling leads to two paradoxical phenomena: (1) visually similar bird heads from different subcategories are represented differently due to category divergence, whereas (2) dissimilar bird heads and wings within the same subcategory are clustered due to sharing the same category. Consequently, the model's representation power relies heavily on predefined category semantics, struggling to interpret unseen subcategories based on actual visual appearances." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Fortunately, object reconstruction tasks inherently focus" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.336, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: hjli@sdust.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3080" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "on pixel-level fidelity and contextual coherence, requiring the model to preserve appearance details and their contextual semantics without embedding category-specific semantics [9, 12]. As demonstrated in Fig. 1b, this property enables the reconstruction model's encoder to generate similar feature representations for visually similar parts across different subcategories. Building on this insight, we consider to leverage the inherent category-agnostic representation capability of object reconstruction models to eliminate dependencies on predefined category semantics in visual discrepancy modeling. However, naively training a fine-grained reconstruction model proves inadequate for FGIR, as it focuses on modeling the entire image appearance, including irrelevant background information, instead of emphasizing key visual discrepancies. Therefore, one natural question arises: is it possible to synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.396, + 0.483, + 0.742 + ], + "angle": 0, + "content": "To answer this, we propose AdvRF, a novel Adversarial Reconstruction Feedback framework that reformulates FGIR as a visual discrepancy reconstruction task. AdvRF synergizes a retrieval model and a reconstruction model within an adversarial learning paradigm inspired by Generative Adversarial Networks (GANs) [8]. The framework operates through alternating optimization: while the retrieval model pinpoints subtle discrepancies within objects, the reconstruction model represents these discrepancies with category-agnostic representations, creating a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Specifically, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. This adversarial interplay progressively achieves precise localization of visual discrepancies, while eliminating predefined category semantics in discrepancy modeling. For efficient deployment, AdvRF distills category-agnostic discrepancy representations purified by the reconstruction model into the retrieval model through lightweight knowledge distillation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.745, + 0.416, + 0.758 + ], + "angle": 0, + "content": "Our main contribution are summarized below:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.765, + 0.48, + 0.808 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to reformulate FGIR tasks as a visual discrepancy reconstruction process, improving its generalization." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.48, + 0.882 + ], + "angle": 0, + "content": "- An adversarial reconstruction feedback framework, i.e., AdvRF, is proposed. AdvRF establishes an adversarial pipeline to alternately train the retrieval and reconstruction models, effectively capturing category-agnostic discrepancies essential for representing unseen categories." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.887, + 0.48, + 0.899 + ], + "angle": 0, + "content": "- Extensive experiments show that our AdvRF achieves" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.765, + 0.48, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "state-of-the-art performance on widely-used fine-grained and coarse-grained retrieval benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.14, + 0.655, + 0.155 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.165, + 0.905, + 0.497 + ], + "angle": 0, + "content": "Fine-grained image retrieval can be broadly categorized into two groups. The first group, localization-based scheme, focuses on localizing object or its details to facilitate the retrieval of visually similar objects [23, 33, 44, 51]. CaRA [38] implements a rectified activation strategy to enhance the localization of object details. The second group, metric-based schemes, seeks to learn an embedding space where similar examples are closely aligned, while dissimilar ones are pushed apart [14, 15, 25, 49, 50]. NIA [26] enforces unique translatability of samples from their respective class proxies to bring the distance of samples with the same subcategory closer. Unlike these approaches, PLEor [34] incorporates category-specific language descriptions based on the CLIP model to guide the model in representing visual discrepancies. However, such methods, which rely on representing similar objects through predefined categories, unintentionally embed category-specific semantic information into the retrieval representations, thereby limiting their ability to generalize to unseen categories. To address this issue, we propose AdvRF, which incorporates an adversarial pipeline to acquire category-agnostic discrepancies derived from the reconstructed features of objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.499, + 0.905, + 0.725 + ], + "angle": 0, + "content": "Adversarial learning has gained extensive application across multiple domains, including generative adversarial networks (GANs) [4, 6, 43], person re-identification [53], and domain adaptation [7, 11]. The essence of adversarial learning lies in minimizing distributional discrepancies between target and source domains by counteracting adversarial attacks. Furthermore, existing work such as OpenGAN [16] seeks to enhance the model's generalization to unseen categories by training a robust open-vs-closed discriminator that distinguishes between synthesized fake data and real data. In contrast to OpenGAN, which improves generalization through the generation of fake samples, AdvRF enhances generalization by capturing class-agnostic discrepancies via adversarial learning between retrieval and reconstruction models from a feature perspective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.854, + 0.755 + ], + "angle": 0, + "content": "3. Adversarial Reconstruction Feedback" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The core of AdvRF, as depicted in Fig. 2, lies in its innovative collaboration between a retrieval model and a reconstruction model. The retrieval model pinpoints discrepancy locations within objects, while the reconstruction model, leveraging these locations, generates category-agnostic representations of these discrepancies. This is achieved through adversarial feedback learning, where the two models create a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Fur" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "3081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.905, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.333, + 0.788, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Detailed illustration of adversarial reconstruction feedback. See §3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.421 + ], + "angle": 0, + "content": "thermore, these discrepancies are back into the retrieval model via knowledge distillation, significantly boosting computational efficiency during test." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.295, + 0.451 + ], + "angle": 0, + "content": "3.1. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.655 + ], + "angle": 0, + "content": "Fine-grained Retrieval Model \\(\\mathcal{R}\\). It is designed to extract robust object representations and generate the final retrieval embeddings. Given an input image \\(\\mathbf{X}\\), we denote \\(\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}\\) as the \\(C\\)-dimensional feature tensor with \\(H \\times W\\) spatial dimensions, encoded by a backbone network \\(\\mathbf{F} = \\mathcal{R}(\\mathbf{X})\\). Traditionally, the most prevalent approach for fine-grained retrieval tasks embeds the full feature tensor \\(\\mathbf{F}\\) through global average pooling (GAP, \\(g(\\cdot)\\)), which computes the mean across the \\(H \\times W\\) spatial planes. This process yields the final retrieval embeddings \\(\\mathbf{E}_{\\mathbf{R}} \\in \\mathbb{R}^{C} = g(\\mathbf{F})\\). Importantly, our proposed AdvRF framework maintains computational efficiency, as only the retrieval model is needed at inference time." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.657, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Fine-grained Reconstruction Model \\(\\mathcal{G}\\). It consists of an encoder \\(\\mathcal{G}_E\\) and a decoder \\(\\mathcal{G}_D\\), with the detailed architecture outlined below. For the encoder \\(\\mathcal{G}_E\\), we use a lightweight network, specifically ResNet34, with pooling layers omitted from the last two blocks in our experimental setup. To emphasize subtle discrepancies within fine-grained objects, the encoder aggregates feature maps from the last three blocks into a final representation, enabling efficient encoding of input signals. This aggregation process applies \\(1 \\times 1\\) convolutional layers, which distills subtle details from low-level features while capturing semantic information from high-level representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The decoder, \\(\\mathcal{G}_D\\), adopts a U-Net architecture [24] with eight downsampling blocks, seven upsampling blocks, and a final colorization block to ensure high-fidelity reconstruction of the inputs. Each downsampling block consists of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.374, + 0.907, + 0.45 + ], + "angle": 0, + "content": "a \\(4 \\times 4\\) convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation. Similarly, each upsampling block contains a transposed convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.46, + 0.882, + 0.477 + ], + "angle": 0, + "content": "3.2. Category-agnostic Discrepancy Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Considering that the reconstruction model is primarily designed to recover pixel-level details from inputs, it faces difficulties in identifying which visual cues represent meaningful discrepancies for identifying visually similar objects. To address this, we design a category-agnostic discrepancy acquisition module that captures visual discrepancies through category-agnostic representations produced by the reconstruction model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.725 + ], + "angle": 0, + "content": "Discrepancy Decoupling. Since the retrieval model is designed to capture visual discrepancies within images, we leverage its representation as a foundation to identify and localize both visual discrepancy and non-discrepancy regions. Therefore, we map the retrieval representation \\(\\mathbf{F}\\) into a pattern map \\(\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}\\), which serves to indicate the locations of discrepancies. \\(\\hat{\\mathbf{M}}\\) can be generated by a lightweight generator \\(\\mathcal{T}(\\cdot)\\) as below:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.734, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} = \\sigma (\\mathcal {T} (\\mathbf {F})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.794 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the sigmoid activation function, \\(\\mathcal{T}(\\cdot)\\) is a convolutional layer with kernel size 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "However, optimizing the pattern map via backpropagation of the loss function primarily tunes the parameters of the lightweight generator, with minimal impact on the retrieval model's parameters. Consequently, the output of the retrieval model may still include non-discriminative information, potentially impairing retrieval performance. To ensure the retrieval model is optimized to focus exclu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3082" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "sively on visual discrepancies, we introduce an mean generator with the same architecture to produce a refined pattern map. In this way, Eq. 1 can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.144, + 0.483, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {M}} = \\sigma (\\operatorname {E} (\\mathcal {T}) (\\mathbf {F})), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.169, + 0.484, + 0.23 + ], + "angle": 0, + "content": "where \\(\\operatorname{E}(\\mathcal{T})\\) denotes the mean generator without learnable parameters. Its parameters can be updated in a temporal average manner. Concretely, at the \\(t\\)-th iteration, parameters \\(\\operatorname{E}(\\mathcal{T})\\) are accumulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.235, + 0.483, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\operatorname {E} ^ {(t)} (\\mathcal {T}) [ \\theta ] = (1 - \\delta) \\cdot \\operatorname {E} ^ {(t - 1)} (\\mathcal {T}) [ \\theta ] + \\delta \\cdot \\theta , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.336 + ], + "angle": 0, + "content": "where \\(\\operatorname{E}^{(t)}(\\mathcal{T})[\\theta]\\) and \\(\\operatorname{E}^{(t-1)}(\\mathcal{T})[\\theta]\\) denote the parameters of the mean generator in current iteration and last iteration, respectively. The mean generator is initialized as \\(\\operatorname{E}^{(0)}(\\mathcal{T})[\\theta] = \\theta\\). The hyper-parameter \\(\\delta\\) is the updating ratio within the range of \\((0, 1]\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Category-agnostic Discrepancy Transfer. We feed the input image \\(\\mathbf{X}\\) into the encoder \\(\\mathcal{G}_E\\) of the reconstruction model to obtain the category-agnostic representation \\(\\hat{\\mathbf{F}}_{\\mathbf{I}} = \\mathcal{G}_{E}(\\mathbf{X})\\in \\mathbb{R}^{C\\times H\\times W}\\), and then resize both the pattern map \\(\\hat{\\mathbf{M}}\\) and \\(\\hat{\\mathbf{F}}_{\\mathbf{I}}\\) to match the size of the original input images, ensuring high fidelity during the subsequent reconstruction process. Using the amplified pattern map \\(\\mathbf{M}\\), we decompose the amplified image representation \\(\\mathbf{F}_{\\mathbf{I}}\\) into the category-agnostic visual discrepancies \\(\\mathbf{C}_{\\mathbf{A}}\\) and the non-discrepancy representation \\(\\mathbf{C}_{\\mathbf{R}}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.495, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {\\mathbf {A}} = \\mathbf {F} _ {\\mathbf {I}} \\odot \\mathbf {M}, \\quad \\mathbf {C} _ {\\mathbf {R}} = \\mathbf {F} _ {\\mathbf {I}} \\odot (1 - \\mathbf {M}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.395, + 0.534 + ], + "angle": 0, + "content": "Here, \\(\\odot\\) denotes element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.626 + ], + "angle": 0, + "content": "Considering that employing both models concurrently is time-consuming and memory-intensive for retrieval evaluation, we design a category-agnostic discrepancy parameterization constraint. Formally, the category-agnostic discrepancy serves as the supervisory signal and eliminates the need for gradient updates to adjust retrieval representations:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.64, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {P}} = \\left| \\left| \\mathbf {E} _ {\\mathbf {R}} - g \\left(\\mathbf {C} _ {\\mathbf {A}}\\right) \\right| \\right|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.827 + ], + "angle": 0, + "content": "where \\( ||\\cdot|| \\) refers to the Frobenius norm. This constraint directly optimizes the parameters within the retrieval model, ensuring that its output representations exclusively capture category-agnostic visual discrepancies while eliminating category-specific semantics. As a result, the retrieval model gains the ability to pinpoint object discrepancies and characterize them based solely on their visual appearance, even when encountering unseen categories. Importantly, the contextual semantics of the discrepancies are still preserved, as the reconstruction process inherently considers the contextual semantics of both the object and its parts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.373, + 0.851 + ], + "angle": 0, + "content": "3.3. Adversarial Feedback Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To acquire category-agnostic discrepancy representations, we introduce an adversarial feedback learning strategy inspired by GANs, which synergizes a retrieval model with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "a reconstruction model. Through adversarial feedback interplay, the reconstruction model exploits residual discrepancies overlooked by the retrieval model for reconstruction, thereby challenging the retrieval model and enhancing object recovery. Concurrently, the retrieval model dynamically refines its discrepancy localization based on feedback from the reconstruction model, further challenging the reconstruction model and improving discrepancy localization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Reconstruction Feedback. Given the category-agnostic discrepancies \\(\\mathbf{C}_{\\mathbf{A}}\\) and the non-discrepancy representation \\(\\mathbf{C}_{\\mathbf{R}}\\), we input them separately into the decoder \\(\\mathcal{G}_D(\\cdot)\\) of the reconstruction model to obtain the reconstructed regions:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.282, + 0.905, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}\\) and \\(\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}\\) represent the reconstructed images, respectively. Considering that the reconstruction model utilizes the residual discrepancies overlooked by the retrieval model to reconstruct discrepancies, we should evaluate the quality of discrepancy reconstruction using the image generated by non-discrepancy representation \\(\\mathbf{C}_{\\mathbf{R}}\\). Therefore, we minimize the difference between the reconstructed image \\(\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}\\) and the input image \\(\\mathbf{X}\\), within the discrepancy region:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.443, + 0.905, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} ^ {R} = \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}}\\right) \\right\\|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.905, + 0.542 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{M}\\) acts as a spatial attention map to localize discrepancies, maintaining intrinsic invariance across feature/pixel spaces. It can be regarded as a soft validation mask to direct the reconstruction model's focus to critical discrepancy regions during regeneration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Similarly, the reconstruction model also leverages the residual non-discrepancy information from the category-agnostic representation \\(\\mathbf{C}_{\\mathbf{A}}\\) to reconstruct the non-discrepancy regions. Hence, we also impose another reconstruction constraint for minimizing their differences between the reconstructed image \\(\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}\\) and the original images \\(\\mathbf{X}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.646, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} ^ {A} = \\left| \\left| (1 - \\mathbf {M}) \\odot (\\mathbf {X} - \\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}}) \\right| \\right|, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.669, + 0.868, + 0.684 + ], + "angle": 0, + "content": "where \\((1 - \\mathbf{M})\\) indicates the non-discrepancy region." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.684, + 0.905, + 0.714 + ], + "angle": 0, + "content": "Therefore, the total loss for training the reconstruction model could be integrated as:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.722, + 0.905, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} ^ {R e c F} = \\mathcal {L} _ {\\mathcal {G}} ^ {A} + \\mathcal {L} _ {\\mathcal {G}} ^ {R}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.81 + ], + "angle": 0, + "content": "To prevent the retrieval model from optimizing towards less accurate discrepancy localization, we freeze its parameters and exclusively back-propagate the loss gradients to the reconstruction model, including both its encoder and decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Retrieval Feedback. Unlike the reconstruction feedback, which aims to improve the reconstruction ability of the reconstruction model, the retrieval feedback focuses on precisely locating discrepancies within objects. It leverages the reconstruction model as an evaluator to assess the accuracy of discrepancy localization. The above process is much" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3083" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "like how the generator receives feedback from the discriminator in GANs. Formally, we also need to feed \\(\\mathbf{C}_{\\mathbf{A}}\\) and \\(\\mathbf{C}_{\\mathbf{R}}\\) into \\(\\mathcal{G}_D\\) to obtain the reconstructed images as below:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.147, + 0.483, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.207 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}_{\\mathbf{A}}^{\\mathcal{R}}\\) and \\(\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}\\) represent the reconstructed images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.327 + ], + "angle": 0, + "content": "This optimization strategy is similar to optimizing the generator based on feedback from the discriminator. In other words, the discrepancy localization provided by the retrieval model makes it challenging for the reconstruction model to accurately reconstruct the object. Therefore, given the reconstructed images \\(\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}\\) and the original image \\(\\mathbf{X}\\) on the discrepancy regions, we impose a reconstruction constraint to maximize their differences:" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.337, + 0.483, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R}} ^ {R} = - \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.444 + ], + "angle": 0, + "content": "Importantly, the minus sign means that minimizing the loss leads the framework to maximize the difference between the \\(\\mathbf{X}\\) and \\(\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}\\). Similarly, we impose an additional reconstruction constraint to maximize the differences between the reconstructed and original non-discrepancy regions:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.452, + 0.483, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R}} ^ {A} = - \\left\\| \\left(1 - \\mathbf {M}\\right) \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.483, + 0.513 + ], + "angle": 0, + "content": "Finally, the total loss for training the retrieval model to produce the accurate pattern map is:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.521, + 0.483, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R}} ^ {R e t F} = \\mathcal {L} _ {\\mathcal {R}} ^ {A} + \\mathcal {L} _ {\\mathcal {R}} ^ {R}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Similarly, to prevent compromising the reconstruction capabilities of the reconstruction model, we freeze its parameters and exclusively backpropagate the loss gradients to the retrieval model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.357, + 0.638 + ], + "angle": 0, + "content": "3.4. Alternating Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.703 + ], + "angle": 0, + "content": "AdvRF implements an iterative alternating protocol where the reconstruction and retrieval models cyclically enhance each other's improvement. In each training epoch, the following steps are performed:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.703, + 0.484, + 0.75 + ], + "angle": 0, + "content": "1. Reconstruction Updating Phase (RecU): The parameters of the retrieval model \\(\\Theta_{\\mathcal{R}}\\) are frozen, and the reconstruction model \\(\\Theta_{\\mathcal{G}}\\) is updated using the loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.762, + 0.483, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {G}} ^ {R e c U} [ \\Theta_ {\\mathcal {G}} ] = \\alpha \\cdot \\mathcal {L} _ {\\mathcal {G}} ^ {R e c F}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.483, + 0.817 + ], + "angle": 0, + "content": "2. Retrieval Updating Phase (RetU): Freezing \\(\\Theta_{\\mathcal{G}}\\), we then refine \\(\\Theta_{\\mathcal{R}}\\) using multi-task learning with:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.826, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {R}} ^ {R e t U} \\left[ \\Theta_ {\\mathcal {R}} \\right] = \\mathcal {L} _ {\\mathcal {C}} + \\beta \\cdot \\mathcal {L} _ {\\mathcal {P}} + \\gamma \\cdot \\mathcal {L} _ {\\mathcal {R}} ^ {R e t F}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This alternating training strategy repeats until joint convergence, with \\(\\alpha\\), \\(\\beta\\) and \\(\\gamma\\) dynamically balancing task-specific gradients." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison of performance and efficiency on CUB-200-2011 using different combinations of constraints. The first row indicates that we use classification-based feedback as supervision, to replace the proposed AdvRF for comparison. \"T\" is the time of extracted retrieval embeddings." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.17, + 0.895, + 0.291 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\mathcal{G}}^{RecF} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{RetF} \\)\\( \\mathcal{L}_{\\mathcal{P}} \\)Performance
\\( \\mathcal{L}_{\\mathcal{G}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{G}}^{R} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{R} \\)R@1T
66.3%21.1ms
✓✓73.7%36.7ms
✓✓73.4%36.7ms
✓✓✓✓76.8%36.7ms
✓✓✓✓✓76.6%21.1ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.304, + 0.907, + 0.332 + ], + "angle": 0, + "content": "Table 2. Evaluation results of retrieval performance on CUB-200-2011 dataset with diverse feedback." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.343, + 0.895, + 0.418 + ], + "angle": 0, + "content": "
Feedback TypeRecall@1
Classification-based Feedback66.3%
Reconstruction-based Feedback62.2%
Our AdvRF76.6%
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.445, + 0.649, + 0.463 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.704, + 0.487 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Datasets. CUB-200-2011 [3] consists of 200 bird species. We use the first 100 subcategories (5,864 images) for training and the consists of (5,924 images) for testing. The Stanford Cars [17] includes 196 car models. Similarly, we use the first 98 classes, which contain 8,054 images, for training and the remaining classes, which contain 8,131 images, for testing. Finally, FGVC Aircraft [22] is split into first 50 classes, containing 5,000 images, for training and the remaining 50 classes with 5,000 images, for testing. Stanford Online Products (SOP) [27] is divided into the 11, 318 subcategories (59, 551 images) in training, and the rest 11, 316 classes (60, 502 images) in testing. This split ensures no category overlap between training and testing sets, where all testing categories are strictly unseen during training to evaluate cross-category generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our retrieval model is built upon a ResNet-50 backbone [10] initialized with ImageNet pretrained weights. Input images are resized to \\(256 \\times 256\\) pixels and randomly cropped to \\(224 \\times 224\\) during training. We employ Stochastic Gradient Descent with an initial learning rate of \\(10^{-5}\\), weight decay of 0.0001, and momentum of 0.9, using a batch size of 32 distributed across four NVIDIA A100 GPUs. To enhance robustness, standard data augmentations including random cropping, horizontal flipping, and color jittering are applied. The learning rate follows an exponential decay schedule (factor=0.9 every 5 epochs) over 200 training epochs, ensuring stable convergence while mit" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3084" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 3. Compared with competitive methods on CUB-200-2011, Stanford Cars 196 and FGVC Aircraft datasets. \"Arch\" represents the architecture of utilizing backbone network. \"R50\" denotes Resnet50 [10] backbone network." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.905, + 0.427 + ], + "angle": 0, + "content": "
MethodArchCUB-200-2011Stanford Cars 196FGVC Aircraft
124812481248
SCDA TIP17 [44]R5057.370.281.088.448.360.271.881.856.567.777.685.7
CRL IJCAI18 [51]R5062.574.282.989.757.869.178.686.661.171.680.988.2
HDCL IJON21 [47]R5069.579.686.892.484.490.194.196.571.181.088.393.3
CEP ECCV20 [2]R5069.279.286.991.689.393.996.698.1----
CaRA TPAMI24 [38]R5073.982.289.493.694.196.998.298.984.390.494.296.3
FRPT AAAI23 [36]R5074.383.789.894.391.195.197.398.677.685.791.495.6
DGCRL AAAI19 [52]R5067.979.186.291.875.983.989.794.070.179.688.093.0
DAS ECCV22 [20]R5069.279.387.192.687.893.296.097.9----
CBML TPAMI23 [13]R5069.980.487.292.588.192.695.497.4----
NIR CVPR22 [26]R5070.580.6--89.193.4------
HIST CVPR22 [19]R5071.481.188.1-89.693.996.4-----
IDML TPAMI24 [32]R5070.780.2--90.694.5------
HSE ICCV23 [46]R5070.680.187.1-89.693.896.0-----
PNCA++ ECCV20 [31]R5072.282.089.293.590.194.597.098.4----
PLEor CVPR23 [34]R5074.884.591.394.994.496.998.398.986.391.795.196.7
Our AdvRFR5076.685.391.795.094.997.298.698.988.092.595.596.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.404, + 0.469 + ], + "angle": 0, + "content": "igating overfitting to category-specific patterns." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Evaluation protocols. We evaluate the retrieval performance by \\( Recall@K \\) with cosine distance, which is average recall scores over all query images in the test set and strictly follows the setting in previous work [30]. Specifically, for each query, our model returns the top \\( K \\) similar images. In the top \\( K \\) returning images, the score will be 1 if there exists at least one positive image, and 0 otherwise." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.59, + 0.296, + 0.607 + ], + "angle": 0, + "content": "4.2. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Efficacy of various constraints. The proposed AdvRF, as described in Sec. 3, is optimized through a combination of four loss functions, each playing a distinct role in guiding AdvRF to capture category-agnostic discrepancies. Tab. 1 presents quantitative comparisons across various constraint combinations. Initially, we use ResNet-50 [10] with only the classification loss \\(\\mathcal{L}_{\\mathcal{C}}\\), achieving \\(66.3\\%\\) Recall@1 accuracy on the CUB-200-2011 dataset. By introducing the reconstruction feedback loss \\(\\mathcal{L}_{\\mathcal{G}}^{RecF}\\) and the retrieval feedback loss \\(\\mathcal{L}_{\\mathcal{R}}^{RetF}\\), we synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models, obtaining a performance of \\(76.8\\%\\). When removing \\(\\mathcal{L}_{\\mathcal{G}}^{A}\\) or \\(\\mathcal{L}_{\\mathcal{G}}^{R}\\), the reconstruction model's sensitivity to residual discrepancies from the retrieval model decreases, leading to reduced performance. Additionally, when removing \\(\\mathcal{L}_{\\mathcal{R}}^{A}\\) or \\(\\mathcal{L}_{\\mathcal{R}}^{R}\\), the retrieval model struggles to accurately evaluate the discrepancy localization, which also results in decreased performance. Finally, for efficient deployment, we introduce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.514 + ], + "angle": 0, + "content": "a category-agnostic discrepancy parameterization loss \\(\\mathcal{L}_{\\mathcal{P}}\\) to distill category-agnostic discrepancy representations into the retrieval model, enabling real-time retrieval without acceptable accuracy loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.786 + ], + "angle": 0, + "content": "Different types of feedback. Tab. 2 presents a comparison of the retrieval performance for retrieving visually similar objects using different types of feedback. Directly using classification-based feedback fundamentally couples discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations, resulting in a performance of \\(66.3\\%\\). When directly training a reconstruction model and using its encoder outputs for retrieval, the model focuses on modeling the entire image appearance, including irrelevant background information, rather than emphasizing key visual discrepancies, which leads to a lower performance. In contrast, our AdvRF effectively combines the advantages of category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations, thus achieving a performance of \\(76.6\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.854, + 0.818 + ], + "angle": 0, + "content": "4.3. Comparisons with the State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Fine-grained image retrieval. Our AdvRF demonstrates superior performance across all three FGIR benchmarks (CUB-200-2011, Stanford Cars-196, FGVC Aircraft), significantly outperforming existing state-of-the-art methods (Tab. 3). Localization-based approaches (e.g., CaRA [38]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3085" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 4. Recall@k for \\( \\mathrm{k} = 1,{10},{100},{1000} \\) on Stanford Online Products (SOP)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.48, + 0.374 + ], + "angle": 0, + "content": "
Method1101001000
MS [40] CVPR1978.290.596.098.7
NSM [48] BMVC2179.591.596.7-
DCML [49] CVPR2179.890.895.895.8
ETLR [14] CVPR2179.891.196.3-
MRML-PA [50] ICCV2179.990.796.1-
HSE [46] ICCV2380.091.496.3-
DAS [20] ECCV2280.691.896.799.0
CEP [2] ECCV2081.191.796.398.8
PNCA++ [31] ECCV2081.492.496.999.0
IBC [28] ICML2181.491.395.9-
HIST [19] CVPR2281.492.096.7-
IDML [32] TPAMI2481.592.354.851.3
CaRA [38] TPAMI2482.492.697.099.0
Our AdvRF84.293.797.699.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.401, + 0.483, + 0.612 + ], + "angle": 0, + "content": "FRPT [36]) and metric-learning frameworks (e.g., IDML [32], HIST [19]) demonstrate effectiveness in capturing fine-grained visual discrepancies. However, their inherent coupling of discrepancy modeling with predefined category supervision embeds category-specific semantics into retrieval representations, fundamentally limiting performance breakthroughs in unseen category generalization. Therefore, AdvRF introduces an adversarial reconstruction mechanism that decouples discrepancy modeling from categorical supervision through iterative training between the retrieval and reconstruction models. This mechanism explicitly grounds visual discrepancies in appearance cues rather than seen category semantics, thereby achieving significant performance gains in generalization to unseen categories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Coarse-grained image retrieval. To further validate AdvRF's generalization capability, we evaluate it on a largescale coarse-grained benchmark, i.e., Stanford Online Products, in Tab. 4. The framework's synergy between the precise discrepancy localization of the retrieval model and the category-agnostic representation learning capability of the reconstruction model can represent objects using category-agnostic description. Hence, AdvRF not only captures subtle inter-class differences in fine-grained settings but also maintains robustness to coarse-grained semantic gaps, thus obtaining a better performance on SOP." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.789, + 0.26, + 0.805 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Investigation on the updating ratio \\(\\delta\\). Tab. 5 showcases the accuracy of various updating ratios in Eqn. 3. Notably, as the ratio increases, retrieval performance declines, indicating that excessive updates to the generator cause it to rely too heavily on the current learning parameters, making it harder to fine-tune the parameters within the retrieval" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 5. Evaluation results on CUB-200-2011 of light-weight generator trained with different updating ratio \\( \\delta \\) in Eqn. 3." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.129, + 0.902, + 0.173 + ], + "angle": 0, + "content": "
Ratio δ0.10.20.40.60.8
R@175.4%76.6%74.8%74.4%73.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.192, + 0.907, + 0.22 + ], + "angle": 0, + "content": "Table 6. Results comparing to various pattern maps based on Recall@K on CUB-200-2011." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.231, + 0.905, + 0.31 + ], + "angle": 0, + "content": "
MethodR@1R@2R@4R@8
CAM [29]69.8%79.7%84.2%91.6%
Bounding box73.9%82.6%90.5%94.2%
Our AdvRF76.6%85.3%91.7%95.0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.33, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Table 7. Effect on the reconstruction ability with different reconstructed manners on CUB-200-2011." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.37, + 0.907, + 0.429 + ], + "angle": 0, + "content": "
Reconstruction MannerRecall@1Recall@2
Non-adversarial Recon.72.6%82.4%
Adversarial Recon.76.6%+4.085.3%+2.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.521 + ], + "angle": 0, + "content": "model. Conversely, a lower ratio preserves sensitivity to prior knowledge, forcing the generator to rely more on the retrieval model's features, thereby providing more accurate features by modifying the retrieval model's parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Visual discrepancy localization with different manners. Switching the visual discrepancy localization method provides insights into acquiring category-agnostic discrepancies. As Tab. 6 indicates, shifting from our discrepancy decoupling strategy to a fixed localization method causes a significant performance drop, nearing the accuracy of fine-tuning a pre-trained model. Specifically, using class activation maps or dataset-provided bounding boxes for localization often leads to imprecise results, including background and missing critical discrepancies. In contrast, our decoupling strategy allows AdvRF to accurately locate visual discrepancies and generate precise category-agnostic representations, consistently improving performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Adversarial learning between the retrieval and reconstruction models. In AdvRF, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to enhance its reconstruction capability. When the reconstruction model only uses the discrepancies localized by the retrieval model, rather than its residual discrepancies, it struggles to provide more comprehensive feedback, resulting in a performance of \\(72.6\\%\\) as shown in the first row of Tab. 7. This implicitly indicates that adversarial learning creates a self-reinforcing cycle, where both" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "3086" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.255, + 0.483, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Analyses of hyper-parameters \\(\\alpha\\), \\(\\beta\\) and \\(\\gamma\\) in Eq. 14 and 15. Results denote Recall@1 accuracy on CUB-200-2011." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.301, + 0.174, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.48, + 0.147, + 0.489 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.301, + 0.249, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.481, + 0.218, + 0.489 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.301, + 0.326, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.481, + 0.296, + 0.489 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.301, + 0.402, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.481, + 0.372, + 0.489 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.301, + 0.48, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.481, + 0.448, + 0.489 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.552 + ], + "angle": 0, + "content": "Figure 4. Visualizations of pattern maps using different feedback: (a) inputs, (b) classification-based feedback, (c) non-discrepancy reconstruction-based feedback, (d) discrepancy reconstruction-based feedback, and (e) our AdvRF." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.611 + ], + "angle": 0, + "content": "the retrieval and reconstruction models iteratively challenge and reinforce each other." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Hyper-parameter analyses. We conduct sensitivity analyses of the hyperparameters in Eq. 14 and 15, with evaluation results presented in Fig. 3. The performance of our AdvRF shows slight sensitivity to variations in \\(\\alpha\\), \\(\\beta\\), \\(\\gamma\\), and \\(\\delta\\). In our experiments, the default values are set to \\(\\alpha = 0.7\\), \\(\\beta = 0.5\\), \\(\\gamma = 0.6\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of pattern maps with various feedback. Fig. 4 illustrates the impact of various feedback signals on discrepancy localization. When using classification-based feedback, the model struggles to accurately localize discrepancies. However, formulating FGIR as a visual discrepancy reconstruction task enhances discrepancy localization through reconstruction-based feedback. Notably, limited reconstruction signals, such as using only discrepancy-based feedback (\\(\\mathbf{C}_{\\mathbf{A}}\\) in Eq. 4) or non-discrepancy-based feedback (\\(\\mathbf{C}_{\\mathbf{R}}\\) in Eq. 4), may cause pattern maps to overlook certain discrepancies. Our results suggest that combining comprehensive reconstruction feedback creates a self-reinforcing cycle, where both the retrieval and reconstruc" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.677, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.212, + 0.698, + 0.224 + ], + "angle": 0, + "content": "(a) Category-related Embeddings" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.091, + 0.707, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.091, + 0.894, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.212, + 0.892, + 0.224 + ], + "angle": 0, + "content": "(b) Category-agnostic Embeddings" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.236, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Figure 5. Evaluating the efficacy of category-related embeddings from baseline [10] versus category-agnostic embeddings from our AdvRF, with their similarity in grid formats." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.576 + ], + "angle": 0, + "content": "tion models iteratively challenge and strengthen each other. Analysis of category-agnostic discrepancies. We employ an indirect method to interpret category-agnostic discrepancies by comparing the similarities between category-agnostic retrieval embeddings produced by our AdvRF and category-related retrieval embeddings generated by the classification-based feedback illustrated in Fig. 1a. As illustrated in Fig. 5, we compute these similarities for the ten high-similarity images across five novel subcategories. Our analysis reveals that category-agnostic embeddings effectively highlight nearest sample pairs that belong to the same subcategory. In contrast, category-related retrieval embeddings struggle to identify these high-similarity images, as they tend to capture visual discrepancies associated with the semantics of base categories. Overall, our AdvRF successfully learns category-agnostic descriptions for unseen subcategories by formulating FGIR tasks as a visual discrepancy reconstruction process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this paper, we introduce AdvRF, which acquires category-agnostic visual discrepancies by formulating FGIR as a visual discrepancy reconstruction task. AdvRF designs an adversarial pipeline: the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. As a result, AdvRF precisely localizes visual differences and encodes them into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Importantly, our algorithm is end-to-end trainable and achieves state-of-the-art performance on the widely-used fine-grained and coarse-grained retrieval datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (NSFC) under Grant (No.61932020), and the Taishan Scholar Program of Shandong Province (tstp20221128)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3087" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[1] Kenan E. Ak, Ashraf A. Kassim, Joo-Hwee Lim, and Jo Yew Tham. Learning attribute representations with localization for flexible fashion search. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7708-7717. Computer Vision Foundation / IEEE Computer Society, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.215, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[2] Malik Boudiaf, Jérôme Rony, Imtiaz Masud Ziko, Eric Granger, Marco Pedersoli, Pablo Piantanida, and Ismail Ben Ayed. A unifying mutual information view of metric learning: Cross-entropy vs. pairwise losses. In ECCV, pages 548–564. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.287, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Steve Branson, Grant Van Horn, Serge J. Belongie, and Pietro Perona. Bird species categorization using pose normalized deep convolutional nets. CoRR, abs/1406.2952, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[4] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Chojui Hsieh, and Boqing Gong. Robust and accurate object detection via adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16622-16631. Computer Vision Foundation / IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.429, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[5] Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed M. Elgammal. Link the head to the \"beak\": Zero shot learning from noisy text description at part precision. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6288-6297. IEEE Computer Society, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.515, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[6] Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xiaoming Wei, and Xiaolin Wei. Masked auto-encoders meet generative adversarial networks and beyond. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24449-24459. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.485, + 0.669 + ], + "angle": 0, + "content": "[7] Yaroslav Ganin and Victor S. Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1180-1189. JMLR.org, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[9] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778, 2016. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised domain adaptation. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "UT, USA, June 18-22, 2018, pages 1498-1507. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[12] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In International Conference on 3D Vision, 3DV 2024, Davos, Switzerland, March 18-21, 2024, pages 31-41. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Shichao Kan, Zhiquan He, Yigang Cen, Yang Li, Vladimir Mladenovic, and Zhihai He. Contrastive bayesian analysis for deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7220-7238, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[14] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Embedding transfer with label relaxation for improved metric learning. In CVPR, pages 3967-3976. Computer Vision Foundation / IEEE, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[15] ByungSoo Ko, Geonmo Gu, Han-Gyu Kim, and ByungSoo Ko. Learning with memory-based virtual classes for deep metric learning. In ICCV, pages 11772-11781. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[16] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 793-802. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[18] Haojie Li, Mingxuan Li, Qijie Peng, Shijie Wang, Hong Yu, and Zhihui Wang. Correlation-guided semantic consistency network for visible-infrared person re-identification. IEEE Trans. Circuits Syst. Video Technol., 34(6):4503-4515, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[19] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 212-222. IEEE, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[20] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. DAS: densely-anchored sampling for deep metric learning. In ECCV, pages 399–417. Springer, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[21] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1096-1104. IEEE Computer Society, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Olga Moskvyak, Frédéric Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Keypoint-aligned embeddings for image retrieval and re-identification. In WACV, pages 676–685. IEEE, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3088" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Karsten Roth, Timo Milbich, Björn Ommer, Joseph Paul Cohen, and Marzyeh Ghassemi. Simultaneous similarity-based self-distillation for deep metric learning. In ICML, pages 9095–9106. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[26] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7410-7420. IEEE, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[27] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823. IEEE Computer Society, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[28] Jenny Seidenschwarz. Learning intra-batch connections for deep metric learning. In ICML, pages 9410-9421. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.505 + ], + "angle": 0, + "content": "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, pages 4004-4012. IEEE Computer Society, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.483, + 0.562 + ], + "angle": 0, + "content": "[31] Eu Wern Teh, Terrance DeVries, Graham W. Taylor, and Graham. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, pages 448-464. Springer, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.564, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[32] Chengkun Wang, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Introspective deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 46(4):1964-1980, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.607, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[33] Shijie Wang, Zhihui Wang, Haojie Li, and Wanli Ouyang. Category-specific nuance exploration network for fine-grained object retrieval. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2513-2521. AAAI Press, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[34] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19381-19391. IEEE, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[35] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Learning to parameterize visual attributes for open-set fine-grained retrieval. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Fine-grained retrieval prompt tuning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 2644-2652. AAAI Press, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[37] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Semantic-guided information alignment network for fine-grained image recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Content-aware rectified activation for zero-shot fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 46(6):4366-4380, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[39] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Accurate fine-grained object recognition with structure-driven relation graph networks. Int. J. Comput. Vis., 132(1):137-160, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[40] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030. Computer Vision Foundation / IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[41] Zhuhui Wang, Shijie Wang, Haojie Li, Zhi Dou, and Jianjun Li. Graph-propagation based correlation learning for weakly supervised fine-grained image classification. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12289-12296. AAAI Press, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[42] Zhihui Wang, Shijie Wang, Shuhui Yang, Haojie Li, Jianjun Li, and Zezhou Li. Weakly supervised fine-grained image classification via gaussian mixture model oriented discriminative learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9746-9755. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[43] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15054-15063. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[44] Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, and Zhi-Hua Zhou. Selective convolutional descriptor aggregation for fine-grained image retrieval. IEEE Trans. Image Process., 26(6):2868-2881, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Xiu-Shen Wei, Yang Shen, Xuhao Sun, Peng Wang, and Yuxin Peng. Attribute-aware deep hashing with self-consistency for large-scale fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):13904-13920, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[46] Bailin Yang, Haoqiang Sun, Frederick W. B. Li, Zheng Chen, Jianlu Cai, and Chao Song. HSE: hybrid species embedding" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3089" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "for deep metric learning. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11013-11023. IEEE, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[47] Xianxian Zeng, Shun Liu, Xiaodong Wang, Yun Zhang, Kairui Chen, and Dong Li. Hard decorrelated centralized loss for fine-grained image retrieval. Neurocomputing, 453: 26-37, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[48] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In BMVC, page 91. BMVA Press, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[49] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In CVPR, pages 9320-9329. Computer Vision Foundation / IEEE, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[50] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In ICCV, pages 12045-12054. IEEE, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[51] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, Feiyue Huang, and Yanhua Yang. Centralized ranking loss with weakly supervised localization for fine-grained object retrieval. In IJCAI, pages 1226-1233. ijcai.org, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[52] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Baochang Zhang, Yongjian Wu, and Feiyue Huang. Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer. In AAAI, pages 9291-9298. AAAI Press, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[53] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2138-2147. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3090" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_origin.pdf b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84bcc08768817b1fa22637df44149fb2a782a8cd --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/2762e731-fb17-490b-a90d-7e63ff84cb06_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1dc2319cc927844caa8d38904ad06013126fcc3307921026312f8b392d023b0 +size 1396289 diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/full.md b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..17790ec89eece6ea3fd4b5715b02342775d8da7b --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/full.md @@ -0,0 +1,361 @@ +# Adversarial Reconstruction Feedback for Robust Fine-grained Generalization + +Shijie Wang $^{1}$ , Jian Shi $^{2}$ , Haojie Li $^{1*}$ $^{1}$ College of Computer and Engineering, Shandong University of Science and Technology, China + $^{2}$ School of Software, Dalian University of Technology, China + +# Abstract + +Existing fine-grained image retrieval (FGIR) methods predominantly rely on supervision from predefined categories to learn discriminative representations for retrieving fine-grained objects. However, they inadvertently introduce category-specific semantics into the retrieval representation, creating semantic dependencies on predefined classes that critically hinder generalization to unseen categories. To tackle this, we propose AdvRF, a novel adversarial reconstruction feedback framework aimed at learning category-agnostic discrepancy representations. Specifically, AdvRF reformulates FGIR as a visual discrepancy reconstruction task via synergizing category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models. The reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. Consequently, the retrieval model localizes visual differences, while the reconstruction model encodes these differences into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Quantitative and qualitative evaluations demonstrate that our AdvRF achieves impressive performance on both widely-used fine-grained and coarse-grained datasets. + +# 1. Introduction + +Fine-grained image retrieval (FGIR) aims to retrieve visually similar subcategories, even those that were unseen during the training phase. It plays a vital role in numerous vision applications from fashion industry, e.g., retrieval of diverse types of clothes [1, 18, 21, 41], to environmental conservation, e.g., retrieving endangered species [5, 37, 39, 42]. Given its significance, a substantial body of research has focused on learning discriminative and generalizable embed + +![](images/d5238bdc197d484a7c1e124e264865e9493dcfea68333971c446d71926c0d534.jpg) + +![](images/e817585f07671b66f24a9eca6de95e7e92c007ceca278153112850bdedc77360.jpg) +(a) Classification-based Representation Learning +(b) Reconstruction-based Representation Learning +Figure 1. Motivation of the proposed AdvRF. (a) Classification-based representation learning easily embeds predefined category semantics into representations, causing visually similar cues across subcategories to appear dissimilar in feature space while visually dissimilar cues within the same subcategory appear similar. (b) Reconstruction-based representation learning allows for a focus on object details and the contextual semantics of both the object and its parts, based on their appearance, thereby capturing category-agnostic representations. This facilitates a deeper understanding of unseen categories by enabling precise appearance modeling using category-agnostic visual descriptions. + +dings to enhance the performance of FGIR. + +Current work on FGIR tasks [19, 20, 23, 26, 35, 45] has achieved promising results by introducing metric constraints or designing localization schemes to capture diverse visual discrepancies from visually similar objects. However, these approaches fundamentally couple discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations. As illustrated in Fig. 1a, this coupling leads to two paradoxical phenomena: (1) visually similar bird heads from different subcategories are represented differently due to category divergence, whereas (2) dissimilar bird heads and wings within the same subcategory are clustered due to sharing the same category. Consequently, the model's representation power relies heavily on predefined category semantics, struggling to interpret unseen subcategories based on actual visual appearances. + +Fortunately, object reconstruction tasks inherently focus + +on pixel-level fidelity and contextual coherence, requiring the model to preserve appearance details and their contextual semantics without embedding category-specific semantics [9, 12]. As demonstrated in Fig. 1b, this property enables the reconstruction model's encoder to generate similar feature representations for visually similar parts across different subcategories. Building on this insight, we consider to leverage the inherent category-agnostic representation capability of object reconstruction models to eliminate dependencies on predefined category semantics in visual discrepancy modeling. However, naively training a fine-grained reconstruction model proves inadequate for FGIR, as it focuses on modeling the entire image appearance, including irrelevant background information, instead of emphasizing key visual discrepancies. Therefore, one natural question arises: is it possible to synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations? + +To answer this, we propose AdvRF, a novel Adversarial Reconstruction Feedback framework that reformulates FGIR as a visual discrepancy reconstruction task. AdvRF synergizes a retrieval model and a reconstruction model within an adversarial learning paradigm inspired by Generative Adversarial Networks (GANs) [8]. The framework operates through alternating optimization: while the retrieval model pinpoints subtle discrepancies within objects, the reconstruction model represents these discrepancies with category-agnostic representations, creating a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Specifically, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. This adversarial interplay progressively achieves precise localization of visual discrepancies, while eliminating predefined category semantics in discrepancy modeling. For efficient deployment, AdvRF distills category-agnostic discrepancy representations purified by the reconstruction model into the retrieval model through lightweight knowledge distillation. + +Our main contribution are summarized below: + +- To the best of our knowledge, we are the first to reformulate FGIR tasks as a visual discrepancy reconstruction process, improving its generalization. +- An adversarial reconstruction feedback framework, i.e., AdvRF, is proposed. AdvRF establishes an adversarial pipeline to alternately train the retrieval and reconstruction models, effectively capturing category-agnostic discrepancies essential for representing unseen categories. +- Extensive experiments show that our AdvRF achieves + +state-of-the-art performance on widely-used fine-grained and coarse-grained retrieval benchmarks. + +# 2. Related Work + +Fine-grained image retrieval can be broadly categorized into two groups. The first group, localization-based scheme, focuses on localizing object or its details to facilitate the retrieval of visually similar objects [23, 33, 44, 51]. CaRA [38] implements a rectified activation strategy to enhance the localization of object details. The second group, metric-based schemes, seeks to learn an embedding space where similar examples are closely aligned, while dissimilar ones are pushed apart [14, 15, 25, 49, 50]. NIA [26] enforces unique translatability of samples from their respective class proxies to bring the distance of samples with the same subcategory closer. Unlike these approaches, PLEor [34] incorporates category-specific language descriptions based on the CLIP model to guide the model in representing visual discrepancies. However, such methods, which rely on representing similar objects through predefined categories, unintentionally embed category-specific semantic information into the retrieval representations, thereby limiting their ability to generalize to unseen categories. To address this issue, we propose AdvRF, which incorporates an adversarial pipeline to acquire category-agnostic discrepancies derived from the reconstructed features of objects. + +Adversarial learning has gained extensive application across multiple domains, including generative adversarial networks (GANs) [4, 6, 43], person re-identification [53], and domain adaptation [7, 11]. The essence of adversarial learning lies in minimizing distributional discrepancies between target and source domains by counteracting adversarial attacks. Furthermore, existing work such as OpenGAN [16] seeks to enhance the model's generalization to unseen categories by training a robust open-vs-closed discriminator that distinguishes between synthesized fake data and real data. In contrast to OpenGAN, which improves generalization through the generation of fake samples, AdvRF enhances generalization by capturing class-agnostic discrepancies via adversarial learning between retrieval and reconstruction models from a feature perspective. + +# 3. Adversarial Reconstruction Feedback + +The core of AdvRF, as depicted in Fig. 2, lies in its innovative collaboration between a retrieval model and a reconstruction model. The retrieval model pinpoints discrepancy locations within objects, while the reconstruction model, leveraging these locations, generates category-agnostic representations of these discrepancies. This is achieved through adversarial feedback learning, where the two models create a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Fur + +![](images/6afc28cc1f19ff5b6790c98a698953be76ff960f1e35f579b373225dc55899fd.jpg) +Figure 2. Detailed illustration of adversarial reconstruction feedback. See §3 for more details. + +thermore, these discrepancies are back into the retrieval model via knowledge distillation, significantly boosting computational efficiency during test. + +# 3.1. Network Architecture + +Fine-grained Retrieval Model $\mathcal{R}$ . It is designed to extract robust object representations and generate the final retrieval embeddings. Given an input image $\mathbf{X}$ , we denote $\mathbf{F} \in \mathbb{R}^{C \times H \times W}$ as the $C$ -dimensional feature tensor with $H \times W$ spatial dimensions, encoded by a backbone network $\mathbf{F} = \mathcal{R}(\mathbf{X})$ . Traditionally, the most prevalent approach for fine-grained retrieval tasks embeds the full feature tensor $\mathbf{F}$ through global average pooling (GAP, $g(\cdot)$ ), which computes the mean across the $H \times W$ spatial planes. This process yields the final retrieval embeddings $\mathbf{E}_{\mathbf{R}} \in \mathbb{R}^{C} = g(\mathbf{F})$ . Importantly, our proposed AdvRF framework maintains computational efficiency, as only the retrieval model is needed at inference time. + +Fine-grained Reconstruction Model $\mathcal{G}$ . It consists of an encoder $\mathcal{G}_E$ and a decoder $\mathcal{G}_D$ , with the detailed architecture outlined below. For the encoder $\mathcal{G}_E$ , we use a lightweight network, specifically ResNet34, with pooling layers omitted from the last two blocks in our experimental setup. To emphasize subtle discrepancies within fine-grained objects, the encoder aggregates feature maps from the last three blocks into a final representation, enabling efficient encoding of input signals. This aggregation process applies $1 \times 1$ convolutional layers, which distills subtle details from low-level features while capturing semantic information from high-level representations. + +The decoder, $\mathcal{G}_D$ , adopts a U-Net architecture [24] with eight downsampling blocks, seven upsampling blocks, and a final colorization block to ensure high-fidelity reconstruction of the inputs. Each downsampling block consists of + +a $4 \times 4$ convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation. Similarly, each upsampling block contains a transposed convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation. + +# 3.2. Category-agnostic Discrepancy Acquisition + +Considering that the reconstruction model is primarily designed to recover pixel-level details from inputs, it faces difficulties in identifying which visual cues represent meaningful discrepancies for identifying visually similar objects. To address this, we design a category-agnostic discrepancy acquisition module that captures visual discrepancies through category-agnostic representations produced by the reconstruction model. + +Discrepancy Decoupling. Since the retrieval model is designed to capture visual discrepancies within images, we leverage its representation as a foundation to identify and localize both visual discrepancy and non-discrepancy regions. Therefore, we map the retrieval representation $\mathbf{F}$ into a pattern map $\hat{\mathbf{M}}\in \mathbb{R}^{H\times W}$ , which serves to indicate the locations of discrepancies. $\hat{\mathbf{M}}$ can be generated by a lightweight generator $\mathcal{T}(\cdot)$ as below: + +$$ +\hat {\mathbf {M}} = \sigma (\mathcal {T} (\mathbf {F})), \tag {1} +$$ + +where $\sigma (\cdot)$ is the sigmoid activation function, $\mathcal{T}(\cdot)$ is a convolutional layer with kernel size 1. + +However, optimizing the pattern map via backpropagation of the loss function primarily tunes the parameters of the lightweight generator, with minimal impact on the retrieval model's parameters. Consequently, the output of the retrieval model may still include non-discriminative information, potentially impairing retrieval performance. To ensure the retrieval model is optimized to focus exclu + +sively on visual discrepancies, we introduce an mean generator with the same architecture to produce a refined pattern map. In this way, Eq. 1 can be rewritten as: + +$$ +\hat {\mathbf {M}} = \sigma (\operatorname {E} (\mathcal {T}) (\mathbf {F})), \tag {2} +$$ + +where $\operatorname{E}(\mathcal{T})$ denotes the mean generator without learnable parameters. Its parameters can be updated in a temporal average manner. Concretely, at the $t$ -th iteration, parameters $\operatorname{E}(\mathcal{T})$ are accumulated by: + +$$ +\operatorname {E} ^ {(t)} (\mathcal {T}) [ \theta ] = (1 - \delta) \cdot \operatorname {E} ^ {(t - 1)} (\mathcal {T}) [ \theta ] + \delta \cdot \theta , \tag {3} +$$ + +where $\operatorname{E}^{(t)}(\mathcal{T})[\theta]$ and $\operatorname{E}^{(t-1)}(\mathcal{T})[\theta]$ denote the parameters of the mean generator in current iteration and last iteration, respectively. The mean generator is initialized as $\operatorname{E}^{(0)}(\mathcal{T})[\theta] = \theta$ . The hyper-parameter $\delta$ is the updating ratio within the range of $(0, 1]$ . + +Category-agnostic Discrepancy Transfer. We feed the input image $\mathbf{X}$ into the encoder $\mathcal{G}_E$ of the reconstruction model to obtain the category-agnostic representation $\hat{\mathbf{F}}_{\mathbf{I}} = \mathcal{G}_{E}(\mathbf{X})\in \mathbb{R}^{C\times H\times W}$ , and then resize both the pattern map $\hat{\mathbf{M}}$ and $\hat{\mathbf{F}}_{\mathbf{I}}$ to match the size of the original input images, ensuring high fidelity during the subsequent reconstruction process. Using the amplified pattern map $\mathbf{M}$ , we decompose the amplified image representation $\mathbf{F}_{\mathbf{I}}$ into the category-agnostic visual discrepancies $\mathbf{C}_{\mathbf{A}}$ and the non-discrepancy representation $\mathbf{C}_{\mathbf{R}}$ as follows: + +$$ +\mathbf {C} _ {\mathbf {A}} = \mathbf {F} _ {\mathbf {I}} \odot \mathbf {M}, \quad \mathbf {C} _ {\mathbf {R}} = \mathbf {F} _ {\mathbf {I}} \odot (1 - \mathbf {M}). \tag {4} +$$ + +Here, $\odot$ denotes element-wise multiplication. + +Considering that employing both models concurrently is time-consuming and memory-intensive for retrieval evaluation, we design a category-agnostic discrepancy parameterization constraint. Formally, the category-agnostic discrepancy serves as the supervisory signal and eliminates the need for gradient updates to adjust retrieval representations: + +$$ +\mathcal {L} _ {\mathcal {P}} = \left| \left| \mathbf {E} _ {\mathbf {R}} - g \left(\mathbf {C} _ {\mathbf {A}}\right) \right| \right|, \tag {5} +$$ + +where $||\cdot||$ refers to the Frobenius norm. This constraint directly optimizes the parameters within the retrieval model, ensuring that its output representations exclusively capture category-agnostic visual discrepancies while eliminating category-specific semantics. As a result, the retrieval model gains the ability to pinpoint object discrepancies and characterize them based solely on their visual appearance, even when encountering unseen categories. Importantly, the contextual semantics of the discrepancies are still preserved, as the reconstruction process inherently considers the contextual semantics of both the object and its parts. + +# 3.3. Adversarial Feedback Learning + +To acquire category-agnostic discrepancy representations, we introduce an adversarial feedback learning strategy inspired by GANs, which synergizes a retrieval model with + +a reconstruction model. Through adversarial feedback interplay, the reconstruction model exploits residual discrepancies overlooked by the retrieval model for reconstruction, thereby challenging the retrieval model and enhancing object recovery. Concurrently, the retrieval model dynamically refines its discrepancy localization based on feedback from the reconstruction model, further challenging the reconstruction model and improving discrepancy localization. + +Reconstruction Feedback. Given the category-agnostic discrepancies $\mathbf{C}_{\mathbf{A}}$ and the non-discrepancy representation $\mathbf{C}_{\mathbf{R}}$ , we input them separately into the decoder $\mathcal{G}_D(\cdot)$ of the reconstruction model to obtain the reconstructed regions: + +$$ +\mathbf {X} _ {\mathbf {A}} ^ {\mathcal {G}} = \mathcal {G} _ {D} (\mathbf {C} _ {\mathbf {A}}), \quad \mathbf {X} _ {\mathbf {R}} ^ {\mathcal {G}} = \mathcal {G} _ {D} (\mathbf {C} _ {\mathbf {R}}). \tag {6} +$$ + +Here, $\mathbf{X}_{\mathbf{A}}^{\mathcal{G}}$ and $\mathbf{X}_{\mathbf{R}}^{\mathcal{G}}$ represent the reconstructed images, respectively. Considering that the reconstruction model utilizes the residual discrepancies overlooked by the retrieval model to reconstruct discrepancies, we should evaluate the quality of discrepancy reconstruction using the image generated by non-discrepancy representation $\mathbf{C}_{\mathbf{R}}$ . Therefore, we minimize the difference between the reconstructed image $\mathbf{X}_{\mathbf{R}}^{\mathcal{G}}$ and the input image $\mathbf{X}$ , within the discrepancy region: + +$$ +\mathcal {L} _ {\mathcal {G}} ^ {R} = \left\| \mathbf {M} \odot \left(\mathbf {X} - \mathbf {X} _ {\mathbf {R}} ^ {\mathcal {G}}\right) \right\|. \tag {7} +$$ + +Here, $\mathbf{M}$ acts as a spatial attention map to localize discrepancies, maintaining intrinsic invariance across feature/pixel spaces. It can be regarded as a soft validation mask to direct the reconstruction model's focus to critical discrepancy regions during regeneration. + +Similarly, the reconstruction model also leverages the residual non-discrepancy information from the category-agnostic representation $\mathbf{C}_{\mathbf{A}}$ to reconstruct the non-discrepancy regions. Hence, we also impose another reconstruction constraint for minimizing their differences between the reconstructed image $\mathbf{X}_{\mathbf{A}}^{\mathcal{G}}$ and the original images $\mathbf{X}$ : + +$$ +\mathcal {L} _ {\mathcal {G}} ^ {A} = \left| \left| (1 - \mathbf {M}) \odot (\mathbf {X} - \mathbf {X} _ {\mathbf {A}} ^ {\mathcal {G}}) \right| \right|, \tag {8} +$$ + +where $(1 - \mathbf{M})$ indicates the non-discrepancy region. + +Therefore, the total loss for training the reconstruction model could be integrated as: + +$$ +\mathcal {L} _ {\mathcal {G}} ^ {R e c F} = \mathcal {L} _ {\mathcal {G}} ^ {A} + \mathcal {L} _ {\mathcal {G}} ^ {R}. \tag {9} +$$ + +To prevent the retrieval model from optimizing towards less accurate discrepancy localization, we freeze its parameters and exclusively back-propagate the loss gradients to the reconstruction model, including both its encoder and decoder. + +Retrieval Feedback. Unlike the reconstruction feedback, which aims to improve the reconstruction ability of the reconstruction model, the retrieval feedback focuses on precisely locating discrepancies within objects. It leverages the reconstruction model as an evaluator to assess the accuracy of discrepancy localization. The above process is much + +like how the generator receives feedback from the discriminator in GANs. Formally, we also need to feed $\mathbf{C}_{\mathbf{A}}$ and $\mathbf{C}_{\mathbf{R}}$ into $\mathcal{G}_D$ to obtain the reconstructed images as below: + +$$ +\hat {\mathbf {X}} _ {\mathbf {A}} ^ {\mathcal {R}} = \mathcal {G} _ {D} (\mathbf {C} _ {\mathbf {A}}), \quad \hat {\mathbf {X}} _ {\mathbf {R}} ^ {\mathcal {R}} = \mathcal {G} _ {D} (\mathbf {C} _ {\mathbf {R}}), \tag {10} +$$ + +where $\hat{\mathbf{X}}_{\mathbf{A}}^{\mathcal{R}}$ and $\hat{\mathbf{X}}_{\mathbf{R}}^{\mathcal{R}}$ represent the reconstructed images, respectively. + +This optimization strategy is similar to optimizing the generator based on feedback from the discriminator. In other words, the discrepancy localization provided by the retrieval model makes it challenging for the reconstruction model to accurately reconstruct the object. Therefore, given the reconstructed images $\hat{\mathbf{X}}_{\mathbf{R}}^{\mathcal{R}}$ and the original image $\mathbf{X}$ on the discrepancy regions, we impose a reconstruction constraint to maximize their differences: + +$$ +\mathcal {L} _ {\mathcal {R}} ^ {R} = - \left\| \mathbf {M} \odot \left(\mathbf {X} - \hat {\mathbf {X}} _ {\mathbf {R}} ^ {\mathcal {R}}\right) \right\|. \tag {11} +$$ + +Importantly, the minus sign means that minimizing the loss leads the framework to maximize the difference between the $\mathbf{X}$ and $\hat{\mathbf{X}}_{\mathbf{R}}^{\mathcal{R}}$ . Similarly, we impose an additional reconstruction constraint to maximize the differences between the reconstructed and original non-discrepancy regions: + +$$ +\mathcal {L} _ {\mathcal {R}} ^ {A} = - \left\| \left(1 - \mathbf {M}\right) \odot \left(\mathbf {X} - \hat {\mathbf {X}} _ {\mathbf {A}} ^ {\mathcal {R}}\right) \right\|. \tag {12} +$$ + +Finally, the total loss for training the retrieval model to produce the accurate pattern map is: + +$$ +\mathcal {L} _ {\mathcal {R}} ^ {R e t F} = \mathcal {L} _ {\mathcal {R}} ^ {A} + \mathcal {L} _ {\mathcal {R}} ^ {R}. \tag {13} +$$ + +Similarly, to prevent compromising the reconstruction capabilities of the reconstruction model, we freeze its parameters and exclusively backpropagate the loss gradients to the retrieval model. + +# 3.4. Alternating Training Strategy + +AdvRF implements an iterative alternating protocol where the reconstruction and retrieval models cyclically enhance each other's improvement. In each training epoch, the following steps are performed: + +1. Reconstruction Updating Phase (RecU): The parameters of the retrieval model $\Theta_{\mathcal{R}}$ are frozen, and the reconstruction model $\Theta_{\mathcal{G}}$ is updated using the loss function: + +$$ +\mathcal {L} _ {\mathcal {G}} ^ {R e c U} [ \Theta_ {\mathcal {G}} ] = \alpha \cdot \mathcal {L} _ {\mathcal {G}} ^ {R e c F}. \tag {14} +$$ + +2. Retrieval Updating Phase (RetU): Freezing $\Theta_{\mathcal{G}}$ , we then refine $\Theta_{\mathcal{R}}$ using multi-task learning with: + +$$ +\mathcal {L} _ {\mathcal {R}} ^ {R e t U} \left[ \Theta_ {\mathcal {R}} \right] = \mathcal {L} _ {\mathcal {C}} + \beta \cdot \mathcal {L} _ {\mathcal {P}} + \gamma \cdot \mathcal {L} _ {\mathcal {R}} ^ {R e t F}. \tag {15} +$$ + +This alternating training strategy repeats until joint convergence, with $\alpha$ , $\beta$ and $\gamma$ dynamically balancing task-specific gradients. + +Table 1. Comparison of performance and efficiency on CUB-200-2011 using different combinations of constraints. The first row indicates that we use classification-based feedback as supervision, to replace the proposed AdvRF for comparison. "T" is the time of extracted retrieval embeddings. + +
\( \mathcal{L}_{\mathcal{G}}^{RecF} \)\( \mathcal{L}_{\mathcal{R}}^{RetF} \)\( \mathcal{L}_{\mathcal{P}} \)Performance
\( \mathcal{L}_{\mathcal{G}}^{A} \)\( \mathcal{L}_{\mathcal{G}}^{R} \)\( \mathcal{L}_{\mathcal{R}}^{A} \)\( \mathcal{L}_{\mathcal{R}}^{R} \)R@1T
66.3%21.1ms
✓✓73.7%36.7ms
✓✓73.4%36.7ms
✓✓✓✓76.8%36.7ms
✓✓✓✓✓76.6%21.1ms
+ +Table 2. Evaluation results of retrieval performance on CUB-200-2011 dataset with diverse feedback. + +
Feedback TypeRecall@1
Classification-based Feedback66.3%
Reconstruction-based Feedback62.2%
Our AdvRF76.6%
+ +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. CUB-200-2011 [3] consists of 200 bird species. We use the first 100 subcategories (5,864 images) for training and the consists of (5,924 images) for testing. The Stanford Cars [17] includes 196 car models. Similarly, we use the first 98 classes, which contain 8,054 images, for training and the remaining classes, which contain 8,131 images, for testing. Finally, FGVC Aircraft [22] is split into first 50 classes, containing 5,000 images, for training and the remaining 50 classes with 5,000 images, for testing. Stanford Online Products (SOP) [27] is divided into the 11, 318 subcategories (59, 551 images) in training, and the rest 11, 316 classes (60, 502 images) in testing. This split ensures no category overlap between training and testing sets, where all testing categories are strictly unseen during training to evaluate cross-category generalization. + +Implementation Details. Our retrieval model is built upon a ResNet-50 backbone [10] initialized with ImageNet pretrained weights. Input images are resized to $256 \times 256$ pixels and randomly cropped to $224 \times 224$ during training. We employ Stochastic Gradient Descent with an initial learning rate of $10^{-5}$ , weight decay of 0.0001, and momentum of 0.9, using a batch size of 32 distributed across four NVIDIA A100 GPUs. To enhance robustness, standard data augmentations including random cropping, horizontal flipping, and color jittering are applied. The learning rate follows an exponential decay schedule (factor=0.9 every 5 epochs) over 200 training epochs, ensuring stable convergence while mit + +Table 3. Compared with competitive methods on CUB-200-2011, Stanford Cars 196 and FGVC Aircraft datasets. "Arch" represents the architecture of utilizing backbone network. "R50" denotes Resnet50 [10] backbone network. + +
MethodArchCUB-200-2011Stanford Cars 196FGVC Aircraft
124812481248
SCDA TIP17 [44]R5057.370.281.088.448.360.271.881.856.567.777.685.7
CRL IJCAI18 [51]R5062.574.282.989.757.869.178.686.661.171.680.988.2
HDCL IJON21 [47]R5069.579.686.892.484.490.194.196.571.181.088.393.3
CEP ECCV20 [2]R5069.279.286.991.689.393.996.698.1----
CaRA TPAMI24 [38]R5073.982.289.493.694.196.998.298.984.390.494.296.3
FRPT AAAI23 [36]R5074.383.789.894.391.195.197.398.677.685.791.495.6
DGCRL AAAI19 [52]R5067.979.186.291.875.983.989.794.070.179.688.093.0
DAS ECCV22 [20]R5069.279.387.192.687.893.296.097.9----
CBML TPAMI23 [13]R5069.980.487.292.588.192.695.497.4----
NIR CVPR22 [26]R5070.580.6--89.193.4------
HIST CVPR22 [19]R5071.481.188.1-89.693.996.4-----
IDML TPAMI24 [32]R5070.780.2--90.694.5------
HSE ICCV23 [46]R5070.680.187.1-89.693.896.0-----
PNCA++ ECCV20 [31]R5072.282.089.293.590.194.597.098.4----
PLEor CVPR23 [34]R5074.884.591.394.994.496.998.398.986.391.795.196.7
Our AdvRFR5076.685.391.795.094.997.298.698.988.092.595.596.9
+ +igating overfitting to category-specific patterns. + +Evaluation protocols. We evaluate the retrieval performance by $Recall@K$ with cosine distance, which is average recall scores over all query images in the test set and strictly follows the setting in previous work [30]. Specifically, for each query, our model returns the top $K$ similar images. In the top $K$ returning images, the score will be 1 if there exists at least one positive image, and 0 otherwise. + +# 4.2. Ablation Experiments + +Efficacy of various constraints. The proposed AdvRF, as described in Sec. 3, is optimized through a combination of four loss functions, each playing a distinct role in guiding AdvRF to capture category-agnostic discrepancies. Tab. 1 presents quantitative comparisons across various constraint combinations. Initially, we use ResNet-50 [10] with only the classification loss $\mathcal{L}_{\mathcal{C}}$ , achieving $66.3\%$ Recall@1 accuracy on the CUB-200-2011 dataset. By introducing the reconstruction feedback loss $\mathcal{L}_{\mathcal{G}}^{RecF}$ and the retrieval feedback loss $\mathcal{L}_{\mathcal{R}}^{RetF}$ , we synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models, obtaining a performance of $76.8\%$ . When removing $\mathcal{L}_{\mathcal{G}}^{A}$ or $\mathcal{L}_{\mathcal{G}}^{R}$ , the reconstruction model's sensitivity to residual discrepancies from the retrieval model decreases, leading to reduced performance. Additionally, when removing $\mathcal{L}_{\mathcal{R}}^{A}$ or $\mathcal{L}_{\mathcal{R}}^{R}$ , the retrieval model struggles to accurately evaluate the discrepancy localization, which also results in decreased performance. Finally, for efficient deployment, we introduce + +a category-agnostic discrepancy parameterization loss $\mathcal{L}_{\mathcal{P}}$ to distill category-agnostic discrepancy representations into the retrieval model, enabling real-time retrieval without acceptable accuracy loss. + +Different types of feedback. Tab. 2 presents a comparison of the retrieval performance for retrieving visually similar objects using different types of feedback. Directly using classification-based feedback fundamentally couples discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations, resulting in a performance of $66.3\%$ . When directly training a reconstruction model and using its encoder outputs for retrieval, the model focuses on modeling the entire image appearance, including irrelevant background information, rather than emphasizing key visual discrepancies, which leads to a lower performance. In contrast, our AdvRF effectively combines the advantages of category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations, thus achieving a performance of $76.6\%$ . + +# 4.3. Comparisons with the State-of-the-Arts + +Fine-grained image retrieval. Our AdvRF demonstrates superior performance across all three FGIR benchmarks (CUB-200-2011, Stanford Cars-196, FGVC Aircraft), significantly outperforming existing state-of-the-art methods (Tab. 3). Localization-based approaches (e.g., CaRA [38], + +Table 4. Recall@k for $\mathrm{k} = 1,{10},{100},{1000}$ on Stanford Online Products (SOP). + +
Method1101001000
MS [40] CVPR1978.290.596.098.7
NSM [48] BMVC2179.591.596.7-
DCML [49] CVPR2179.890.895.895.8
ETLR [14] CVPR2179.891.196.3-
MRML-PA [50] ICCV2179.990.796.1-
HSE [46] ICCV2380.091.496.3-
DAS [20] ECCV2280.691.896.799.0
CEP [2] ECCV2081.191.796.398.8
PNCA++ [31] ECCV2081.492.496.999.0
IBC [28] ICML2181.491.395.9-
HIST [19] CVPR2281.492.096.7-
IDML [32] TPAMI2481.592.354.851.3
CaRA [38] TPAMI2482.492.697.099.0
Our AdvRF84.293.797.699.1
+ +FRPT [36]) and metric-learning frameworks (e.g., IDML [32], HIST [19]) demonstrate effectiveness in capturing fine-grained visual discrepancies. However, their inherent coupling of discrepancy modeling with predefined category supervision embeds category-specific semantics into retrieval representations, fundamentally limiting performance breakthroughs in unseen category generalization. Therefore, AdvRF introduces an adversarial reconstruction mechanism that decouples discrepancy modeling from categorical supervision through iterative training between the retrieval and reconstruction models. This mechanism explicitly grounds visual discrepancies in appearance cues rather than seen category semantics, thereby achieving significant performance gains in generalization to unseen categories. + +Coarse-grained image retrieval. To further validate AdvRF's generalization capability, we evaluate it on a largescale coarse-grained benchmark, i.e., Stanford Online Products, in Tab. 4. The framework's synergy between the precise discrepancy localization of the retrieval model and the category-agnostic representation learning capability of the reconstruction model can represent objects using category-agnostic description. Hence, AdvRF not only captures subtle inter-class differences in fine-grained settings but also maintains robustness to coarse-grained semantic gaps, thus obtaining a better performance on SOP. + +# 4.4. Further Analysis + +Investigation on the updating ratio $\delta$ . Tab. 5 showcases the accuracy of various updating ratios in Eqn. 3. Notably, as the ratio increases, retrieval performance declines, indicating that excessive updates to the generator cause it to rely too heavily on the current learning parameters, making it harder to fine-tune the parameters within the retrieval + +Table 5. Evaluation results on CUB-200-2011 of light-weight generator trained with different updating ratio $\delta$ in Eqn. 3. + +
Ratio δ0.10.20.40.60.8
R@175.4%76.6%74.8%74.4%73.9%
+ +Table 6. Results comparing to various pattern maps based on Recall@K on CUB-200-2011. + +
MethodR@1R@2R@4R@8
CAM [29]69.8%79.7%84.2%91.6%
Bounding box73.9%82.6%90.5%94.2%
Our AdvRF76.6%85.3%91.7%95.0%
+ +Table 7. Effect on the reconstruction ability with different reconstructed manners on CUB-200-2011. + +
Reconstruction MannerRecall@1Recall@2
Non-adversarial Recon.72.6%82.4%
Adversarial Recon.76.6%+4.085.3%+2.9
+ +model. Conversely, a lower ratio preserves sensitivity to prior knowledge, forcing the generator to rely more on the retrieval model's features, thereby providing more accurate features by modifying the retrieval model's parameters. + +Visual discrepancy localization with different manners. Switching the visual discrepancy localization method provides insights into acquiring category-agnostic discrepancies. As Tab. 6 indicates, shifting from our discrepancy decoupling strategy to a fixed localization method causes a significant performance drop, nearing the accuracy of fine-tuning a pre-trained model. Specifically, using class activation maps or dataset-provided bounding boxes for localization often leads to imprecise results, including background and missing critical discrepancies. In contrast, our decoupling strategy allows AdvRF to accurately locate visual discrepancies and generate precise category-agnostic representations, consistently improving performance. + +Adversarial learning between the retrieval and reconstruction models. In AdvRF, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to enhance its reconstruction capability. When the reconstruction model only uses the discrepancies localized by the retrieval model, rather than its residual discrepancies, it struggles to provide more comprehensive feedback, resulting in a performance of $72.6\%$ as shown in the first row of Tab. 7. This implicitly indicates that adversarial learning creates a self-reinforcing cycle, where both + +![](images/8bbc69070db3eff6bbf95fabe805e8f409b74d656e184074e54a20544e96c7be.jpg) +Figure 3. Analyses of hyper-parameters $\alpha$ , $\beta$ and $\gamma$ in Eq. 14 and 15. Results denote Recall@1 accuracy on CUB-200-2011. + +![](images/8bc19777535b8954896cc7c7421fcf894950d8fc63c78ba94880735a029b1e87.jpg) +(a) +Figure 4. Visualizations of pattern maps using different feedback: (a) inputs, (b) classification-based feedback, (c) non-discrepancy reconstruction-based feedback, (d) discrepancy reconstruction-based feedback, and (e) our AdvRF. + +![](images/19ee7885f4c1bb97e1b05b945cdcf269bbba059539d4b85fe0098c3e09a201f5.jpg) +(b) + +![](images/0014a467debb4bc07c8aa489a90e66c93bb3a743d161366c8a6bf0ebee4ad766.jpg) +(c) + +![](images/a3fa0953cf128b86d4ca2e1408bbc0fa17073c2b91070b321a418a9ffa085577.jpg) +(d) + +![](images/6664c4eea81105b15aeb67f5393612b9abb38de1587e1ce9d781874d58926261.jpg) +(e) + +the retrieval and reconstruction models iteratively challenge and reinforce each other. + +Hyper-parameter analyses. We conduct sensitivity analyses of the hyperparameters in Eq. 14 and 15, with evaluation results presented in Fig. 3. The performance of our AdvRF shows slight sensitivity to variations in $\alpha$ , $\beta$ , $\gamma$ , and $\delta$ . In our experiments, the default values are set to $\alpha = 0.7$ , $\beta = 0.5$ , $\gamma = 0.6$ , respectively. + +Effect of pattern maps with various feedback. Fig. 4 illustrates the impact of various feedback signals on discrepancy localization. When using classification-based feedback, the model struggles to accurately localize discrepancies. However, formulating FGIR as a visual discrepancy reconstruction task enhances discrepancy localization through reconstruction-based feedback. Notably, limited reconstruction signals, such as using only discrepancy-based feedback ( $\mathbf{C}_{\mathbf{A}}$ in Eq. 4) or non-discrepancy-based feedback ( $\mathbf{C}_{\mathbf{R}}$ in Eq. 4), may cause pattern maps to overlook certain discrepancies. Our results suggest that combining comprehensive reconstruction feedback creates a self-reinforcing cycle, where both the retrieval and reconstruc + +![](images/2f55175f45a83489e6e34a7c370f3bc6ad3b59090ac3607bf8d1b2aceecbd651.jpg) +(a) Category-related Embeddings + +![](images/c57401f9b835fed62a18d0654ee57290683d405b28db86189c1465fcd03fc9bc.jpg) +Figure 5. Evaluating the efficacy of category-related embeddings from baseline [10] versus category-agnostic embeddings from our AdvRF, with their similarity in grid formats. + +![](images/df6e6371e14f4ac7bc37d1e9769bb961f37c16cf38498095f529b0cd175b9f39.jpg) +(b) Category-agnostic Embeddings + +tion models iteratively challenge and strengthen each other. Analysis of category-agnostic discrepancies. We employ an indirect method to interpret category-agnostic discrepancies by comparing the similarities between category-agnostic retrieval embeddings produced by our AdvRF and category-related retrieval embeddings generated by the classification-based feedback illustrated in Fig. 1a. As illustrated in Fig. 5, we compute these similarities for the ten high-similarity images across five novel subcategories. Our analysis reveals that category-agnostic embeddings effectively highlight nearest sample pairs that belong to the same subcategory. In contrast, category-related retrieval embeddings struggle to identify these high-similarity images, as they tend to capture visual discrepancies associated with the semantics of base categories. Overall, our AdvRF successfully learns category-agnostic descriptions for unseen subcategories by formulating FGIR tasks as a visual discrepancy reconstruction process. + +# 5. Conclusion + +In this paper, we introduce AdvRF, which acquires category-agnostic visual discrepancies by formulating FGIR as a visual discrepancy reconstruction task. AdvRF designs an adversarial pipeline: the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. As a result, AdvRF precisely localizes visual differences and encodes them into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Importantly, our algorithm is end-to-end trainable and achieves state-of-the-art performance on the widely-used fine-grained and coarse-grained retrieval datasets. + +Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (NSFC) under Grant (No.61932020), and the Taishan Scholar Program of Shandong Province (tstp20221128). + +# References + +[1] Kenan E. Ak, Ashraf A. Kassim, Joo-Hwee Lim, and Jo Yew Tham. Learning attribute representations with localization for flexible fashion search. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7708-7717. Computer Vision Foundation / IEEE Computer Society, 2018. 1 +[2] Malik Boudiaf, Jérôme Rony, Imtiaz Masud Ziko, Eric Granger, Marco Pedersoli, Pablo Piantanida, and Ismail Ben Ayed. A unifying mutual information view of metric learning: Cross-entropy vs. pairwise losses. In ECCV, pages 548–564. Springer, 2020. 6, 7 +[3] Steve Branson, Grant Van Horn, Serge J. Belongie, and Pietro Perona. Bird species categorization using pose normalized deep convolutional nets. CoRR, abs/1406.2952, 2014. 5 +[4] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Chojui Hsieh, and Boqing Gong. Robust and accurate object detection via adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16622-16631. Computer Vision Foundation / IEEE, 2021. 2 +[5] Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed M. Elgammal. Link the head to the "beak": Zero shot learning from noisy text description at part precision. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6288-6297. IEEE Computer Society, 2017. 1 +[6] Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xiaoming Wei, and Xiaolin Wei. Masked auto-encoders meet generative adversarial networks and beyond. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24449-24459. IEEE, 2023. 2 +[7] Yaroslav Ganin and Victor S. Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1180-1189. JMLR.org, 2015. 2 +[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2 +[9] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2021. 2 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778, 2016. 5, 6, 8 +[11] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised domain adaptation. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, + +UT, USA, June 18-22, 2018, pages 1498-1507. Computer Vision Foundation / IEEE Computer Society, 2018. 2 +[12] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In International Conference on 3D Vision, 3DV 2024, Davos, Switzerland, March 18-21, 2024, pages 31-41. IEEE, 2024. 2 +[13] Shichao Kan, Zhiquan He, Yigang Cen, Yang Li, Vladimir Mladenovic, and Zhihai He. Contrastive bayesian analysis for deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7220-7238, 2023. 6 +[14] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Embedding transfer with label relaxation for improved metric learning. In CVPR, pages 3967-3976. Computer Vision Foundation / IEEE, 2021. 2, 7 +[15] ByungSoo Ko, Geonmo Gu, Han-Gyu Kim, and ByungSoo Ko. Learning with memory-based virtual classes for deep metric learning. In ICCV, pages 11772-11781. IEEE, 2021. 2 +[16] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 793-802. IEEE, 2021. 2 +[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561, 2013. 5 +[18] Haojie Li, Mingxuan Li, Qijie Peng, Shijie Wang, Hong Yu, and Zhihui Wang. Correlation-guided semantic consistency network for visible-infrared person re-identification. IEEE Trans. Circuits Syst. Video Technol., 34(6):4503-4515, 2024. 1 +[19] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 212-222. IEEE, 2022. 1, 6, 7 +[20] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. DAS: densely-anchored sampling for deep metric learning. In ECCV, pages 399–417. Springer, 2022. 1, 6, 7 +[21] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1096-1104. IEEE Computer Society, 2016. 1 +[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5 +[23] Olga Moskvyak, Frédéric Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Keypoint-aligned embeddings for image retrieval and re-identification. In WACV, pages 676–685. IEEE, 2021. 1, 2 +[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. + +In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, pages 234-241. Springer, 2015. 3 +[25] Karsten Roth, Timo Milbich, Björn Ommer, Joseph Paul Cohen, and Marzyeh Ghassemi. Simultaneous similarity-based self-distillation for deep metric learning. In ICML, pages 9095–9106. PMLR, 2021. 2 +[26] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7410-7420. IEEE, 2022. 1, 2, 6 +[27] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823. IEEE Computer Society, 2015. 5 +[28] Jenny Seidenschwarz. Learning intra-batch connections for deep metric learning. In ICML, pages 9410-9421. PMLR, 2021. 7 +[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 7 +[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, pages 4004-4012. IEEE Computer Society, 2016. 6 +[31] Eu Wern Teh, Terrance DeVries, Graham W. Taylor, and Graham. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, pages 448-464. Springer, 2020. 6, 7 +[32] Chengkun Wang, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Introspective deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 46(4):1964-1980, 2024. 6, 7 +[33] Shijie Wang, Zhihui Wang, Haojie Li, and Wanli Ouyang. Category-specific nuance exploration network for fine-grained object retrieval. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2513-2521. AAAI Press, 2022. 2 +[34] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19381-19391. IEEE, 2023. 2, 6 +[35] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Learning to parameterize visual attributes for open-set fine-grained retrieval. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1 + +[36] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Fine-grained retrieval prompt tuning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 2644-2652. AAAI Press, 2023. 6, 7 +[37] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Semantic-guided information alignment network for fine-grained image recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1 +[38] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Content-aware rectified activation for zero-shot fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 46(6):4366-4380, 2024. 2, 6, 7 +[39] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Accurate fine-grained object recognition with structure-driven relation graph networks. Int. J. Comput. Vis., 132(1):137-160, 2024. 1 +[40] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030. Computer Vision Foundation / IEEE, 2019. 7 +[41] Zhuhui Wang, Shijie Wang, Haojie Li, Zhi Dou, and Jianjun Li. Graph-propagation based correlation learning for weakly supervised fine-grained image classification. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12289-12296. AAAI Press, 2020. 1 +[42] Zhihui Wang, Shijie Wang, Shuhui Yang, Haojie Li, Jianjun Li, and Zezhou Li. Weakly supervised fine-grained image classification via gaussian mixture model oriented discriminative learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9746-9755. IEEE, 2020. +[43] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15054-15063. IEEE, 2022. 2 +[44] Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, and Zhi-Hua Zhou. Selective convolutional descriptor aggregation for fine-grained image retrieval. IEEE Trans. Image Process., 26(6):2868-2881, 2017. 2, 6 +[45] Xiu-Shen Wei, Yang Shen, Xuhao Sun, Peng Wang, and Yuxin Peng. Attribute-aware deep hashing with self-consistency for large-scale fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):13904-13920, 2023. 1 +[46] Bailin Yang, Haoqiang Sun, Frederick W. B. Li, Zheng Chen, Jianlu Cai, and Chao Song. HSE: hybrid species embedding + +for deep metric learning. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11013-11023. IEEE, 2023. 6, 7 +[47] Xianxian Zeng, Shun Liu, Xiaodong Wang, Yun Zhang, Kairui Chen, and Dong Li. Hard decorrelated centralized loss for fine-grained image retrieval. Neurocomputing, 453: 26-37, 2021. 6 +[48] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In BMVC, page 91. BMVA Press, 2019. 7 +[49] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In CVPR, pages 9320-9329. Computer Vision Foundation / IEEE, 2021. 2, 7 +[50] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In ICCV, pages 12045-12054. IEEE, 2021. 2, 7 +[51] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, Feiyue Huang, and Yanhua Yang. Centralized ranking loss with weakly supervised localization for fine-grained object retrieval. In IJCAI, pages 1226-1233. ijcai.org, 2018. 2, 6 +[52] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Baochang Zhang, Yongjian Wu, and Feiyue Huang. Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer. In AAAI, pages 9291-9298. AAAI Press, 2019. 6 +[53] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2138-2147. Computer Vision Foundation / IEEE, 2019. 2 \ No newline at end of file diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/images.zip b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7469017029ded8ede732fe5624959382c9a33806 --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:157479e8b4f3f81537d6b557fd401eb1d5164ceb64bee86ba2665ade34ca14ce +size 606819 diff --git a/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/layout.json b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b834b67e5812c9abf944d0d6e0abe1adb0c4471 --- /dev/null +++ b/2025/Adversarial Reconstruction Feedback for Robust Fine-grained Generalization/layout.json @@ -0,0 +1,9680 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "type": "text", + "content": "Adversarial Reconstruction Feedback for Robust Fine-grained Generalization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "content": "Shijie Wang" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "content": ", Jian Shi" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "content": ", Haojie Li" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "content": "College of Computer and Engineering, Shandong University of Science and Technology, China \n" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 142, + 537, + 187 + ], + "type": "text", + "content": "School of Software, Dalian University of Technology, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 550 + ], + "type": "text", + "content": "Existing fine-grained image retrieval (FGIR) methods predominantly rely on supervision from predefined categories to learn discriminative representations for retrieving fine-grained objects. However, they inadvertently introduce category-specific semantics into the retrieval representation, creating semantic dependencies on predefined classes that critically hinder generalization to unseen categories. To tackle this, we propose AdvRF, a novel adversarial reconstruction feedback framework aimed at learning category-agnostic discrepancy representations. Specifically, AdvRF reformulates FGIR as a visual discrepancy reconstruction task via synergizing category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models. The reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. Consequently, the retrieval model localizes visual differences, while the reconstruction model encodes these differences into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Quantitative and qualitative evaluations demonstrate that our AdvRF achieves impressive performance on both widely-used fine-grained and coarse-grained datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 596, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 295, + 693 + ], + "type": "text", + "content": "Fine-grained image retrieval (FGIR) aims to retrieve visually similar subcategories, even those that were unseen during the training phase. It plays a vital role in numerous vision applications from fashion industry, e.g., retrieval of diverse types of clothes [1, 18, 21, 41], to environmental conservation, e.g., retrieving endangered species [5, 37, 39, 42]. Given its significance, a substantial body of research has focused on learning discriminative and generalizable embed" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 213, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 316, + 213, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 264 + ], + "type": "image", + "image_path": "d5238bdc197d484a7c1e124e264865e9493dcfea68333971c446d71926c0d534.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 277, + 552, + 329 + ], + "blocks": [ + { + "bbox": [ + 356, + 266, + 513, + 277 + ], + "lines": [ + { + "bbox": [ + 356, + 266, + 513, + 277 + ], + "spans": [ + { + "bbox": [ + 356, + 266, + 513, + 277 + ], + "type": "text", + "content": "(a) Classification-based Representation Learning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 277, + 552, + 329 + ], + "lines": [ + { + "bbox": [ + 316, + 277, + 552, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 552, + 329 + ], + "type": "image", + "image_path": "e817585f07671b66f24a9eca6de95e7e92c007ceca278153112850bdedc77360.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 331, + 516, + 341 + ], + "lines": [ + { + "bbox": [ + 353, + 331, + 516, + 341 + ], + "spans": [ + { + "bbox": [ + 353, + 331, + 516, + 341 + ], + "type": "text", + "content": "(b) Reconstruction-based Representation Learning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 350, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Motivation of the proposed AdvRF. (a) Classification-based representation learning easily embeds predefined category semantics into representations, causing visually similar cues across subcategories to appear dissimilar in feature space while visually dissimilar cues within the same subcategory appear similar. (b) Reconstruction-based representation learning allows for a focus on object details and the contextual semantics of both the object and its parts, based on their appearance, thereby capturing category-agnostic representations. This facilitates a deeper understanding of unseen categories by enabling precise appearance modeling using category-agnostic visual descriptions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 495, + 491, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 491, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 491, + 507 + ], + "type": "text", + "content": "dings to enhance the performance of FGIR." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 700 + ], + "type": "text", + "content": "Current work on FGIR tasks [19, 20, 23, 26, 35, 45] has achieved promising results by introducing metric constraints or designing localization schemes to capture diverse visual discrepancies from visually similar objects. However, these approaches fundamentally couple discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations. As illustrated in Fig. 1a, this coupling leads to two paradoxical phenomena: (1) visually similar bird heads from different subcategories are represented differently due to category divergence, whereas (2) dissimilar bird heads and wings within the same subcategory are clustered due to sharing the same category. Consequently, the model's representation power relies heavily on predefined category semantics, struggling to interpret unseen subcategories based on actual visual appearances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "Fortunately, object reconstruction tasks inherently focus" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 205, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 205, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 205, + 713 + ], + "type": "text", + "content": "*Corresponding author: hjli@sdust.edu.cn" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3080" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "on pixel-level fidelity and contextual coherence, requiring the model to preserve appearance details and their contextual semantics without embedding category-specific semantics [9, 12]. As demonstrated in Fig. 1b, this property enables the reconstruction model's encoder to generate similar feature representations for visually similar parts across different subcategories. Building on this insight, we consider to leverage the inherent category-agnostic representation capability of object reconstruction models to eliminate dependencies on predefined category semantics in visual discrepancy modeling. However, naively training a fine-grained reconstruction model proves inadequate for FGIR, as it focuses on modeling the entire image appearance, including irrelevant background information, instead of emphasizing key visual discrepancies. Therefore, one natural question arises: is it possible to synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 313, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 295, + 587 + ], + "type": "text", + "content": "To answer this, we propose AdvRF, a novel Adversarial Reconstruction Feedback framework that reformulates FGIR as a visual discrepancy reconstruction task. AdvRF synergizes a retrieval model and a reconstruction model within an adversarial learning paradigm inspired by Generative Adversarial Networks (GANs) [8]. The framework operates through alternating optimization: while the retrieval model pinpoints subtle discrepancies within objects, the reconstruction model represents these discrepancies with category-agnostic representations, creating a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Specifically, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. This adversarial interplay progressively achieves precise localization of visual discrepancies, while eliminating predefined category semantics in discrepancy modeling. For efficient deployment, AdvRF distills category-agnostic discrepancy representations purified by the reconstruction model into the retrieval model through lightweight knowledge distillation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 590, + 254, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 590, + 254, + 600 + ], + "spans": [ + { + "bbox": [ + 67, + 590, + 254, + 600 + ], + "type": "text", + "content": "Our main contribution are summarized below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 605, + 293, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 605, + 293, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 293, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 293, + 639 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to reformulate FGIR tasks as a visual discrepancy reconstruction process, improving its generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 642, + 293, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 293, + 698 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 293, + 698 + ], + "type": "text", + "content": "- An adversarial reconstruction feedback framework, i.e., AdvRF, is proposed. AdvRF establishes an adversarial pipeline to alternately train the retrieval and reconstruction models, effectively capturing category-agnostic discrepancies essential for representing unseen categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 702, + 293, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 293, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 293, + 712 + ], + "type": "text", + "content": "- Extensive experiments show that our AdvRF achieves" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "content": "state-of-the-art performance on widely-used fine-grained and coarse-grained retrieval benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 130, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 553, + 393 + ], + "type": "text", + "content": "Fine-grained image retrieval can be broadly categorized into two groups. The first group, localization-based scheme, focuses on localizing object or its details to facilitate the retrieval of visually similar objects [23, 33, 44, 51]. CaRA [38] implements a rectified activation strategy to enhance the localization of object details. The second group, metric-based schemes, seeks to learn an embedding space where similar examples are closely aligned, while dissimilar ones are pushed apart [14, 15, 25, 49, 50]. NIA [26] enforces unique translatability of samples from their respective class proxies to bring the distance of samples with the same subcategory closer. Unlike these approaches, PLEor [34] incorporates category-specific language descriptions based on the CLIP model to guide the model in representing visual discrepancies. However, such methods, which rely on representing similar objects through predefined categories, unintentionally embed category-specific semantic information into the retrieval representations, thereby limiting their ability to generalize to unseen categories. To address this issue, we propose AdvRF, which incorporates an adversarial pipeline to acquire category-agnostic discrepancies derived from the reconstructed features of objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 395, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 553, + 574 + ], + "type": "text", + "content": "Adversarial learning has gained extensive application across multiple domains, including generative adversarial networks (GANs) [4, 6, 43], person re-identification [53], and domain adaptation [7, 11]. The essence of adversarial learning lies in minimizing distributional discrepancies between target and source domains by counteracting adversarial attacks. Furthermore, existing work such as OpenGAN [16] seeks to enhance the model's generalization to unseen categories by training a robust open-vs-closed discriminator that distinguishes between synthesized fake data and real data. In contrast to OpenGAN, which improves generalization through the generation of fake samples, AdvRF enhances generalization by capturing class-agnostic discrepancies via adversarial learning between retrieval and reconstruction models from a feature perspective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 586, + 522, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 522, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 522, + 597 + ], + "type": "text", + "content": "3. Adversarial Reconstruction Feedback" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "The core of AdvRF, as depicted in Fig. 2, lies in its innovative collaboration between a retrieval model and a reconstruction model. The retrieval model pinpoints discrepancy locations within objects, while the reconstruction model, leveraging these locations, generates category-agnostic representations of these discrepancies. This is achieved through adversarial feedback learning, where the two models create a self-reinforcing cycle where each component iteratively challenges and reinforces each other. Fur" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "3081" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 553, + 258 + ], + "type": "image", + "image_path": "6afc28cc1f19ff5b6790c98a698953be76ff960f1e35f579b373225dc55899fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 263, + 482, + 275 + ], + "lines": [ + { + "bbox": [ + 127, + 263, + 482, + 275 + ], + "spans": [ + { + "bbox": [ + 127, + 263, + 482, + 275 + ], + "type": "text", + "content": "Figure 2. Detailed illustration of adversarial reconstruction feedback. See §3 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "type": "text", + "content": "thermore, these discrepancies are back into the retrieval model via knowledge distillation, significantly boosting computational efficiency during test." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 180, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 180, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 180, + 357 + ], + "type": "text", + "content": "3.1. Network Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": "Fine-grained Retrieval Model " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": ". It is designed to extract robust object representations and generate the final retrieval embeddings. Given an input image " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": "-dimensional feature tensor with " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": " spatial dimensions, encoded by a backbone network " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{F} = \\mathcal{R}(\\mathbf{X})" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": ". Traditionally, the most prevalent approach for fine-grained retrieval tasks embeds the full feature tensor " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": " through global average pooling (GAP, " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": "), which computes the mean across the " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": " spatial planes. This process yields the final retrieval embeddings " + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{R}} \\in \\mathbb{R}^{C} = g(\\mathbf{F})" + }, + { + "bbox": [ + 54, + 363, + 295, + 518 + ], + "type": "text", + "content": ". Importantly, our proposed AdvRF framework maintains computational efficiency, as only the retrieval model is needed at inference time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": "Fine-grained Reconstruction Model " + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": ". It consists of an encoder " + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_E" + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_D" + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": ", with the detailed architecture outlined below. For the encoder " + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_E" + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": ", we use a lightweight network, specifically ResNet34, with pooling layers omitted from the last two blocks in our experimental setup. To emphasize subtle discrepancies within fine-grained objects, the encoder aggregates feature maps from the last three blocks into a final representation, enabling efficient encoding of input signals. This aggregation process applies " + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 54, + 520, + 295, + 664 + ], + "type": "text", + "content": " convolutional layers, which distills subtle details from low-level features while capturing semantic information from high-level representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "The decoder, " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_D" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", adopts a U-Net architecture [24] with eight downsampling blocks, seven upsampling blocks, and a final colorization block to ensure high-fidelity reconstruction of the inputs. Each downsampling block consists of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 296, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 555, + 356 + ], + "type": "text", + "content": "a " + }, + { + "bbox": [ + 313, + 296, + 555, + 356 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 296, + 555, + 356 + ], + "type": "text", + "content": " convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation. Similarly, each upsampling block contains a transposed convolutional layer with stride 2, followed by a normalization layer and LeakyReLU activation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 364, + 539, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 539, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 539, + 377 + ], + "type": "text", + "content": "3.2. Category-agnostic Discrepancy Acquisition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 476 + ], + "type": "text", + "content": "Considering that the reconstruction model is primarily designed to recover pixel-level details from inputs, it faces difficulties in identifying which visual cues represent meaningful discrepancies for identifying visually similar objects. To address this, we design a category-agnostic discrepancy acquisition module that captures visual discrepancies through category-agnostic representations produced by the reconstruction model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "content": "Discrepancy Decoupling. Since the retrieval model is designed to capture visual discrepancies within images, we leverage its representation as a foundation to identify and localize both visual discrepancy and non-discrepancy regions. Therefore, we map the retrieval representation " + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "content": " into a pattern map " + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "content": ", which serves to indicate the locations of discrepancies. " + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "content": " can be generated by a lightweight generator " + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 477, + 554, + 574 + ], + "type": "text", + "content": " as below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 400, + 581, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 581, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 400, + 581, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} = \\sigma (\\mathcal {T} (\\mathbf {F})), \\tag {1}", + "image_path": "0b1a7a2499c0974c794f5fa5d19af4c31083d305cff15e31e628558e84f0552c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": " is the sigmoid activation function, " + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": " is a convolutional layer with kernel size 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "However, optimizing the pattern map via backpropagation of the loss function primarily tunes the parameters of the lightweight generator, with minimal impact on the retrieval model's parameters. Consequently, the output of the retrieval model may still include non-discriminative information, potentially impairing retrieval performance. To ensure the retrieval model is optimized to focus exclu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3082" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "sively on visual discrepancies, we introduce an mean generator with the same architecture to produce a refined pattern map. In this way, Eq. 1 can be rewritten as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 114, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 114, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 134, + 114, + 295, + 129 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {M}} = \\sigma (\\operatorname {E} (\\mathcal {T}) (\\mathbf {F})), \\tag {2}", + "image_path": "265a1274f6a4f861ce21878694c62e2414f98b743fea941973ed92f7c7112003.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\operatorname{E}(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "text", + "content": " denotes the mean generator without learnable parameters. Its parameters can be updated in a temporal average manner. Concretely, at the " + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "text", + "content": "-th iteration, parameters " + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\operatorname{E}(\\mathcal{T})" + }, + { + "bbox": [ + 55, + 133, + 296, + 182 + ], + "type": "text", + "content": " are accumulated by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 186, + 295, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 186, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 85, + 186, + 295, + 201 + ], + "type": "interline_equation", + "content": "\\operatorname {E} ^ {(t)} (\\mathcal {T}) [ \\theta ] = (1 - \\delta) \\cdot \\operatorname {E} ^ {(t - 1)} (\\mathcal {T}) [ \\theta ] + \\delta \\cdot \\theta , \\tag {3}", + "image_path": "a4561f34e90cc08e43a5901e0db0e894b690c4b42f81c8de183b42bb21f63327.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\operatorname{E}^{(t)}(\\mathcal{T})[\\theta]" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\operatorname{E}^{(t-1)}(\\mathcal{T})[\\theta]" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " denote the parameters of the mean generator in current iteration and last iteration, respectively. The mean generator is initialized as " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\operatorname{E}^{(0)}(\\mathcal{T})[\\theta] = \\theta" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": ". The hyper-parameter " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " is the updating ratio within the range of " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "(0, 1]" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": "Category-agnostic Discrepancy Transfer. We feed the input image " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " into the encoder " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_E" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " of the reconstruction model to obtain the category-agnostic representation " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathbf{I}} = \\mathcal{G}_{E}(\\mathbf{X})\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": ", and then resize both the pattern map " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{F}}_{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " to match the size of the original input images, ensuring high fidelity during the subsequent reconstruction process. Using the amplified pattern map " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": ", we decompose the amplified image representation " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " into the category-agnostic visual discrepancies " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " and the non-discrepancy representation " + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 266, + 296, + 386 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 392, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 392, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 89, + 392, + 295, + 405 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {\\mathbf {A}} = \\mathbf {F} _ {\\mathbf {I}} \\odot \\mathbf {M}, \\quad \\mathbf {C} _ {\\mathbf {R}} = \\mathbf {F} _ {\\mathbf {I}} \\odot (1 - \\mathbf {M}). \\tag {4}", + "image_path": "0d6e060603110dcafccec3299e8ae0edc9633aa686d3c2454eace5fed0dff812.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 241, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 241, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 241, + 422 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 411, + 241, + 422 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 411, + 241, + 422 + ], + "type": "text", + "content": " denotes element-wise multiplication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 495 + ], + "type": "text", + "content": "Considering that employing both models concurrently is time-consuming and memory-intensive for retrieval evaluation, we design a category-agnostic discrepancy parameterization constraint. Formally, the category-agnostic discrepancy serves as the supervisory signal and eliminates the need for gradient updates to adjust retrieval representations:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 506, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 506, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 125, + 506, + 295, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {P}} = \\left| \\left| \\mathbf {E} _ {\\mathbf {R}} - g \\left(\\mathbf {C} _ {\\mathbf {A}}\\right) \\right| \\right|, \\tag {5}", + "image_path": "0f37274d8b8e6e14564fd2650e922575a39c657c6bd8a29a62a1944a89ca8b71.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 522, + 296, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 522, + 296, + 654 + ], + "type": "inline_equation", + "content": "||\\cdot||" + }, + { + "bbox": [ + 55, + 522, + 296, + 654 + ], + "type": "text", + "content": " refers to the Frobenius norm. This constraint directly optimizes the parameters within the retrieval model, ensuring that its output representations exclusively capture category-agnostic visual discrepancies while eliminating category-specific semantics. As a result, the retrieval model gains the ability to pinpoint object discrepancies and characterize them based solely on their visual appearance, even when encountering unseen categories. Importantly, the contextual semantics of the discrepancies are still preserved, as the reconstruction process inherently considers the contextual semantics of both the object and its parts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 660, + 228, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 228, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 228, + 673 + ], + "type": "text", + "content": "3.3. Adversarial Feedback Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "To acquire category-agnostic discrepancy representations, we introduce an adversarial feedback learning strategy inspired by GANs, which synergizes a retrieval model with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "a reconstruction model. Through adversarial feedback interplay, the reconstruction model exploits residual discrepancies overlooked by the retrieval model for reconstruction, thereby challenging the retrieval model and enhancing object recovery. Concurrently, the retrieval model dynamically refines its discrepancy localization based on feedback from the reconstruction model, further challenging the reconstruction model and improving discrepancy localization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": "Reconstruction Feedback. Given the category-agnostic discrepancies " + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": " and the non-discrepancy representation " + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": ", we input them separately into the decoder " + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_D(\\cdot)" + }, + { + "bbox": [ + 313, + 168, + 553, + 216 + ], + "type": "text", + "content": " of the reconstruction model to obtain the reconstructed regions:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 223, + 553, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 223, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 358, + 223, + 553, + 237 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}). \\tag {6}", + "image_path": "1fde3d7aed9fbea2fca78e436a7ee4c25673bc7c0af2c71011217ddb7769fab8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": " represent the reconstructed images, respectively. Considering that the reconstruction model utilizes the residual discrepancies overlooked by the retrieval model to reconstruct discrepancies, we should evaluate the quality of discrepancy reconstruction using the image generated by non-discrepancy representation " + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": ". Therefore, we minimize the difference between the reconstructed image " + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{R}}^{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": " and the input image " + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 553, + 350 + ], + "type": "text", + "content": ", within the discrepancy region:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 378, + 350, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 350, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 378, + 350, + 553, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} ^ {R} = \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\mathbf {X} _ {\\mathbf {R}} ^ {\\mathcal {G}}\\right) \\right\\|. \\tag {7}", + "image_path": "f57751005c3ad70afe61fb37a5f24121ecc322d83f764ab6f7899a070bd0e2c1.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 429 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 369, + 553, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 369, + 553, + 429 + ], + "type": "text", + "content": " acts as a spatial attention map to localize discrepancies, maintaining intrinsic invariance across feature/pixel spaces. It can be regarded as a soft validation mask to direct the reconstruction model's focus to critical discrepancy regions during regeneration." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "text", + "content": "Similarly, the reconstruction model also leverages the residual non-discrepancy information from the category-agnostic representation " + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{A}}" + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "text", + "content": " to reconstruct the non-discrepancy regions. Hence, we also impose another reconstruction constraint for minimizing their differences between the reconstructed image " + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathbf{A}}^{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "text", + "content": " and the original images " + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 429, + 554, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 366, + 511, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 511, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 366, + 511, + 553, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} ^ {A} = \\left| \\left| (1 - \\mathbf {M}) \\odot (\\mathbf {X} - \\mathbf {X} _ {\\mathbf {A}} ^ {\\mathcal {G}}) \\right| \\right|, \\tag {8}", + "image_path": "1c37c972969a2386e52acdee002cf51e72c18530a53a048b6ddc4c1542ba9195.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 529, + 531, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 531, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 531, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 529, + 531, + 541 + ], + "type": "inline_equation", + "content": "(1 - \\mathbf{M})" + }, + { + "bbox": [ + 313, + 529, + 531, + 541 + ], + "type": "text", + "content": " indicates the non-discrepancy region." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 553, + 565 + ], + "type": "text", + "content": "Therefore, the total loss for training the reconstruction model could be integrated as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 392, + 571, + 553, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 571, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 392, + 571, + 553, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} ^ {R e c F} = \\mathcal {L} _ {\\mathcal {G}} ^ {A} + \\mathcal {L} _ {\\mathcal {G}} ^ {R}. \\tag {9}", + "image_path": "abc1ab72e65fc96a088afc94b12d39d79317051d89a7a0c2792e2f6057e99270.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 641 + ], + "type": "text", + "content": "To prevent the retrieval model from optimizing towards less accurate discrepancy localization, we freeze its parameters and exclusively back-propagate the loss gradients to the reconstruction model, including both its encoder and decoder." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Retrieval Feedback. Unlike the reconstruction feedback, which aims to improve the reconstruction ability of the reconstruction model, the retrieval feedback focuses on precisely locating discrepancies within objects. It leverages the reconstruction model as an evaluator to assess the accuracy of discrepancy localization. The above process is much" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3083" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "like how the generator receives feedback from the discriminator in GANs. Formally, we also need to feed " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_D" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " to obtain the reconstructed images as below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 116, + 295, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 116, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 99, + 116, + 295, + 131 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {A}}), \\quad \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}} = \\mathcal {G} _ {D} (\\mathbf {C} _ {\\mathbf {R}}), \\tag {10}", + "image_path": "a239245e84b1ab4c39e1b4d7503cf3ffad2e90f3833bba08d8a5665a91843200.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathbf{A}}^{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 138, + 295, + 163 + ], + "type": "text", + "content": " represent the reconstructed images, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "text", + "content": "This optimization strategy is similar to optimizing the generator based on feedback from the discriminator. In other words, the discrepancy localization provided by the retrieval model makes it challenging for the reconstruction model to accurately reconstruct the object. Therefore, given the reconstructed images " + }, + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "text", + "content": " and the original image " + }, + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 163, + 296, + 258 + ], + "type": "text", + "content": " on the discrepancy regions, we impose a reconstruction constraint to maximize their differences:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 266, + 295, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 266, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 295, + 282 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R}} ^ {R} = - \\left\\| \\mathbf {M} \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {R}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {11}", + "image_path": "521a108efde4a66fe1df26841e3ebb387617a33ce9b0aa3e3c019d31f8e7cbda.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "text", + "content": "Importantly, the minus sign means that minimizing the loss leads the framework to maximize the difference between the " + }, + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathbf{R}}^{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 290, + 296, + 351 + ], + "type": "text", + "content": ". Similarly, we impose an additional reconstruction constraint to maximize the differences between the reconstructed and original non-discrepancy regions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 357, + 295, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 357, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 102, + 357, + 295, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R}} ^ {A} = - \\left\\| \\left(1 - \\mathbf {M}\\right) \\odot \\left(\\mathbf {X} - \\hat {\\mathbf {X}} _ {\\mathbf {A}} ^ {\\mathcal {R}}\\right) \\right\\|. \\tag {12}", + "image_path": "2649f99e52c4659c5e10c93f0ac03ddae7deb73b5a43a30fad3b6dc05f8f96fa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 381, + 295, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 406 + ], + "type": "text", + "content": "Finally, the total loss for training the retrieval model to produce the accurate pattern map is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 132, + 412, + 295, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 412, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 295, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R}} ^ {R e t F} = \\mathcal {L} _ {\\mathcal {R}} ^ {A} + \\mathcal {L} _ {\\mathcal {R}} ^ {R}. \\tag {13}", + "image_path": "c217d04b33d98ca81c6c3aaba9e9a8012ccd51d4c462c700d8f49599952f37cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 483 + ], + "type": "text", + "content": "Similarly, to prevent compromising the reconstruction capabilities of the reconstruction model, we freeze its parameters and exclusively backpropagate the loss gradients to the retrieval model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 491, + 218, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 218, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 218, + 505 + ], + "type": "text", + "content": "3.4. Alternating Training Strategy" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 556 + ], + "type": "text", + "content": "AdvRF implements an iterative alternating protocol where the reconstruction and retrieval models cyclically enhance each other's improvement. In each training epoch, the following steps are performed:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "text", + "content": "1. Reconstruction Updating Phase (RecU): The parameters of the retrieval model " + }, + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{R}}" + }, + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "text", + "content": " are frozen, and the reconstruction model " + }, + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 56, + 556, + 296, + 594 + ], + "type": "text", + "content": " is updated using the loss function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 603, + 295, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 603, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 129, + 603, + 295, + 618 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {G}} ^ {R e c U} [ \\Theta_ {\\mathcal {G}} ] = \\alpha \\cdot \\mathcal {L} _ {\\mathcal {G}} ^ {R e c F}. \\tag {14}", + "image_path": "07a333a8ba80a05fe2e9b0c24eba4a80314f2533a0b39c4f27cc1e8619c4b0a2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "text", + "content": "2. Retrieval Updating Phase (RetU): Freezing " + }, + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{G}}" + }, + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "text", + "content": ", we then refine " + }, + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "inline_equation", + "content": "\\Theta_{\\mathcal{R}}" + }, + { + "bbox": [ + 55, + 622, + 295, + 647 + ], + "type": "text", + "content": " using multi-task learning with:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 88, + 654, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 654, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 88, + 654, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {R}} ^ {R e t U} \\left[ \\Theta_ {\\mathcal {R}} \\right] = \\mathcal {L} _ {\\mathcal {C}} + \\beta \\cdot \\mathcal {L} _ {\\mathcal {P}} + \\gamma \\cdot \\mathcal {L} _ {\\mathcal {R}} ^ {R e t F}. \\tag {15}", + "image_path": "182bc0f0e4ba6da2aaf08876ffa39832bebb9cde09bb1a76e68df42232c0380e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "This alternating training strategy repeats until joint convergence, with " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " dynamically balancing task-specific gradients." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 71, + 555, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Comparison of performance and efficiency on CUB-200-2011 using different combinations of constraints. The first row indicates that we use classification-based feedback as supervision, to replace the proposed AdvRF for comparison. \"T\" is the time of extracted retrieval embeddings." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 322, + 134, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 322, + 134, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 322, + 134, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 322, + 134, + 547, + 230 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\mathcal{G}}^{RecF} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{RetF} \\)\\( \\mathcal{L}_{\\mathcal{P}} \\)Performance
\\( \\mathcal{L}_{\\mathcal{G}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{G}}^{R} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{A} \\)\\( \\mathcal{L}_{\\mathcal{R}}^{R} \\)R@1T
66.3%21.1ms
✓✓73.7%36.7ms
✓✓73.4%36.7ms
✓✓✓✓76.8%36.7ms
✓✓✓✓✓76.6%21.1ms
", + "image_path": "c6cffc2cb8e45abaeaaac7be544088372a8a9f5c32a616c69ff7b396058b0c8d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 322, + 271, + 547, + 331 + ], + "blocks": [ + { + "bbox": [ + 313, + 240, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 262 + ], + "type": "text", + "content": "Table 2. Evaluation results of retrieval performance on CUB-200-2011 dataset with diverse feedback." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 271, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 322, + 271, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 322, + 271, + 547, + 331 + ], + "type": "table", + "html": "
Feedback TypeRecall@1
Classification-based Feedback66.3%
Reconstruction-based Feedback62.2%
Our AdvRF76.6%
", + "image_path": "e425d2db622c116d3266e3779a489ee714712f2a4fa2267c1f99d72dc7ab2f08.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 352, + 397, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 397, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 397, + 366 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 373, + 430, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 430, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 430, + 385 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 390, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 570 + ], + "type": "text", + "content": "Datasets. CUB-200-2011 [3] consists of 200 bird species. We use the first 100 subcategories (5,864 images) for training and the consists of (5,924 images) for testing. The Stanford Cars [17] includes 196 car models. Similarly, we use the first 98 classes, which contain 8,054 images, for training and the remaining classes, which contain 8,131 images, for testing. Finally, FGVC Aircraft [22] is split into first 50 classes, containing 5,000 images, for training and the remaining 50 classes with 5,000 images, for testing. Stanford Online Products (SOP) [27] is divided into the 11, 318 subcategories (59, 551 images) in training, and the rest 11, 316 classes (60, 502 images) in testing. This split ensures no category overlap between training and testing sets, where all testing categories are strictly unseen during training to evaluate cross-category generalization." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Implementation Details. Our retrieval model is built upon a ResNet-50 backbone [10] initialized with ImageNet pretrained weights. Input images are resized to " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " pixels and randomly cropped to " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " during training. We employ Stochastic Gradient Descent with an initial learning rate of " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": ", weight decay of 0.0001, and momentum of 0.9, using a batch size of 32 distributed across four NVIDIA A100 GPUs. To enhance robustness, standard data augmentations including random cropping, horizontal flipping, and color jittering are applied. The learning rate follows an exponential decay schedule (factor=0.9 every 5 epochs) over 200 training epochs, ensuring stable convergence while mit" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3084" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 553, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Compared with competitive methods on CUB-200-2011, Stanford Cars 196 and FGVC Aircraft datasets. \"Arch\" represents the architecture of utilizing backbone network. \"R50\" denotes Resnet50 [10] backbone network." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 553, + 338 + ], + "type": "table", + "html": "
MethodArchCUB-200-2011Stanford Cars 196FGVC Aircraft
124812481248
SCDA TIP17 [44]R5057.370.281.088.448.360.271.881.856.567.777.685.7
CRL IJCAI18 [51]R5062.574.282.989.757.869.178.686.661.171.680.988.2
HDCL IJON21 [47]R5069.579.686.892.484.490.194.196.571.181.088.393.3
CEP ECCV20 [2]R5069.279.286.991.689.393.996.698.1----
CaRA TPAMI24 [38]R5073.982.289.493.694.196.998.298.984.390.494.296.3
FRPT AAAI23 [36]R5074.383.789.894.391.195.197.398.677.685.791.495.6
DGCRL AAAI19 [52]R5067.979.186.291.875.983.989.794.070.179.688.093.0
DAS ECCV22 [20]R5069.279.387.192.687.893.296.097.9----
CBML TPAMI23 [13]R5069.980.487.292.588.192.695.497.4----
NIR CVPR22 [26]R5070.580.6--89.193.4------
HIST CVPR22 [19]R5071.481.188.1-89.693.996.4-----
IDML TPAMI24 [32]R5070.780.2--90.694.5------
HSE ICCV23 [46]R5070.680.187.1-89.693.896.0-----
PNCA++ ECCV20 [31]R5072.282.089.293.590.194.597.098.4----
PLEor CVPR23 [34]R5074.884.591.394.994.496.998.398.986.391.795.196.7
Our AdvRFR5076.685.391.795.094.997.298.698.988.092.595.596.9
", + "image_path": "b1e2fce6673487c84940063c14b3f74dfa0a3e160f04e8728b36abbee8a93687.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 359, + 247, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 247, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 247, + 371 + ], + "type": "text", + "content": "igating overfitting to category-specific patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "text", + "content": "Evaluation protocols. We evaluate the retrieval performance by " + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "inline_equation", + "content": "Recall@K" + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "text", + "content": " with cosine distance, which is average recall scores over all query images in the test set and strictly follows the setting in previous work [30]. Specifically, for each query, our model returns the top " + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "text", + "content": " similar images. In the top " + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 372, + 295, + 456 + ], + "type": "text", + "content": " returning images, the score will be 1 if there exists at least one positive image, and 0 otherwise." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 181, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 181, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 181, + 480 + ], + "type": "text", + "content": "4.2. Ablation Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": "Efficacy of various constraints. The proposed AdvRF, as described in Sec. 3, is optimized through a combination of four loss functions, each playing a distinct role in guiding AdvRF to capture category-agnostic discrepancies. Tab. 1 presents quantitative comparisons across various constraint combinations. Initially, we use ResNet-50 [10] with only the classification loss " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{C}}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ", achieving " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "66.3\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": " Recall@1 accuracy on the CUB-200-2011 dataset. By introducing the reconstruction feedback loss " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}^{RecF}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": " and the retrieval feedback loss " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{R}}^{RetF}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ", we synergize category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models, obtaining a performance of " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "76.8\\%" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ". When removing " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}^{A}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{G}}^{R}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ", the reconstruction model's sensitivity to residual discrepancies from the retrieval model decreases, leading to reduced performance. Additionally, when removing " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{R}}^{A}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{R}}^{R}" + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": ", the retrieval model struggles to accurately evaluate the discrepancy localization, which also results in decreased performance. Finally, for efficient deployment, we introduce" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 358, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 407 + ], + "type": "text", + "content": "a category-agnostic discrepancy parameterization loss " + }, + { + "bbox": [ + 313, + 358, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{P}}" + }, + { + "bbox": [ + 313, + 358, + 555, + 407 + ], + "type": "text", + "content": " to distill category-agnostic discrepancy representations into the retrieval model, enabling real-time retrieval without acceptable accuracy loss." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "text", + "content": "Different types of feedback. Tab. 2 presents a comparison of the retrieval performance for retrieving visually similar objects using different types of feedback. Directly using classification-based feedback fundamentally couples discrepancy modeling with predefined category supervision, inadvertently embedding category-specific semantics into the retrieval representations, resulting in a performance of " + }, + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "inline_equation", + "content": "66.3\\%" + }, + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "text", + "content": ". When directly training a reconstruction model and using its encoder outputs for retrieval, the model focuses on modeling the entire image appearance, including irrelevant background information, rather than emphasizing key visual discrepancies, which leads to a lower performance. In contrast, our AdvRF effectively combines the advantages of category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models to model visual discrepancies using category-agnostic representations, thus achieving a performance of " + }, + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "inline_equation", + "content": "76.6\\%" + }, + { + "bbox": [ + 313, + 408, + 556, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 635, + 522, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 522, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 522, + 647 + ], + "type": "text", + "content": "4.3. Comparisons with the State-of-the-Arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Fine-grained image retrieval. Our AdvRF demonstrates superior performance across all three FGIR benchmarks (CUB-200-2011, Stanford Cars-196, FGVC Aircraft), significantly outperforming existing state-of-the-art methods (Tab. 3). Localization-based approaches (e.g., CaRA [38]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3085" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 293, + 296 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 4. Recall@k for " + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "inline_equation", + "content": "\\mathrm{k} = 1,{10},{100},{1000}" + }, + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": " on Stanford Online Products (SOP)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 293, + 296 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 293, + 296 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 293, + 296 + ], + "type": "table", + "html": "
Method1101001000
MS [40] CVPR1978.290.596.098.7
NSM [48] BMVC2179.591.596.7-
DCML [49] CVPR2179.890.895.895.8
ETLR [14] CVPR2179.891.196.3-
MRML-PA [50] ICCV2179.990.796.1-
HSE [46] ICCV2380.091.496.3-
DAS [20] ECCV2280.691.896.799.0
CEP [2] ECCV2081.191.796.398.8
PNCA++ [31] ECCV2081.492.496.999.0
IBC [28] ICML2181.491.395.9-
HIST [19] CVPR2281.492.096.7-
IDML [32] TPAMI2481.592.354.851.3
CaRA [38] TPAMI2482.492.697.099.0
Our AdvRF84.293.797.699.1
", + "image_path": "df6b0c3e420ca27a8600116ac444198b19eafdba7ee860bd61612fefacce342f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 317, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 317, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 317, + 295, + 484 + ], + "type": "text", + "content": "FRPT [36]) and metric-learning frameworks (e.g., IDML [32], HIST [19]) demonstrate effectiveness in capturing fine-grained visual discrepancies. However, their inherent coupling of discrepancy modeling with predefined category supervision embeds category-specific semantics into retrieval representations, fundamentally limiting performance breakthroughs in unseen category generalization. Therefore, AdvRF introduces an adversarial reconstruction mechanism that decouples discrepancy modeling from categorical supervision through iterative training between the retrieval and reconstruction models. This mechanism explicitly grounds visual discrepancies in appearance cues rather than seen category semantics, thereby achieving significant performance gains in generalization to unseen categories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 485, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 617 + ], + "type": "text", + "content": "Coarse-grained image retrieval. To further validate AdvRF's generalization capability, we evaluate it on a largescale coarse-grained benchmark, i.e., Stanford Online Products, in Tab. 4. The framework's synergy between the precise discrepancy localization of the retrieval model and the category-agnostic representation learning capability of the reconstruction model can represent objects using category-agnostic description. Hence, AdvRF not only captures subtle inter-class differences in fine-grained settings but also maintains robustness to coarse-grained semantic gaps, thus obtaining a better performance on SOP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 159, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 159, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 159, + 637 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Investigation on the updating ratio " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": ". Tab. 5 showcases the accuracy of various updating ratios in Eqn. 3. Notably, as the ratio increases, retrieval performance declines, indicating that excessive updates to the generator cause it to rely too heavily on the current learning parameters, making it harder to fine-tune the parameters within the retrieval" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 102, + 552, + 137 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 5. Evaluation results on CUB-200-2011 of light-weight generator trained with different updating ratio " + }, + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "type": "text", + "content": " in Eqn. 3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 102, + 552, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 102, + 552, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 102, + 552, + 137 + ], + "type": "table", + "html": "
Ratio δ0.10.20.40.60.8
R@175.4%76.6%74.8%74.4%73.9%
", + "image_path": "07cb9830013daaf9746a3ef20d9244e033f0068428f41c7442381d8d28d63960.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 182, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 313, + 152, + 555, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 174 + ], + "type": "text", + "content": "Table 6. Results comparing to various pattern maps based on Recall@K on CUB-200-2011." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 182, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 315, + 182, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 315, + 182, + 553, + 245 + ], + "type": "table", + "html": "
MethodR@1R@2R@4R@8
CAM [29]69.8%79.7%84.2%91.6%
Bounding box73.9%82.6%90.5%94.2%
Our AdvRF76.6%85.3%91.7%95.0%
", + "image_path": "b1927f720658cd549a4cb0fc9cb9deb5c298555be3b3c6f3286e2fe4a8b82e37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 293, + 555, + 339 + ], + "blocks": [ + { + "bbox": [ + 313, + 261, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 283 + ], + "type": "text", + "content": "Table 7. Effect on the reconstruction ability with different reconstructed manners on CUB-200-2011." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 293, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 315, + 293, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 555, + 339 + ], + "type": "table", + "html": "
Reconstruction MannerRecall@1Recall@2
Non-adversarial Recon.72.6%82.4%
Adversarial Recon.76.6%+4.085.3%+2.9
", + "image_path": "8a8d73637416e8e6c6439e5410e2b92ad277fa2b866f1943828d6de973ea5a78.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 555, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 412 + ], + "type": "text", + "content": "model. Conversely, a lower ratio preserves sensitivity to prior knowledge, forcing the generator to rely more on the retrieval model's features, thereby providing more accurate features by modifying the retrieval model's parameters." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 413, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 569 + ], + "type": "text", + "content": "Visual discrepancy localization with different manners. Switching the visual discrepancy localization method provides insights into acquiring category-agnostic discrepancies. As Tab. 6 indicates, shifting from our discrepancy decoupling strategy to a fixed localization method causes a significant performance drop, nearing the accuracy of fine-tuning a pre-trained model. Specifically, using class activation maps or dataset-provided bounding boxes for localization often leads to imprecise results, including background and missing critical discrepancies. In contrast, our decoupling strategy allows AdvRF to accurately locate visual discrepancies and generate precise category-agnostic representations, consistently improving performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Adversarial learning between the retrieval and reconstruction models. In AdvRF, the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to enhance its reconstruction capability. When the reconstruction model only uses the discrepancies localized by the retrieval model, rather than its residual discrepancies, it struggles to provide more comprehensive feedback, resulting in a performance of " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "72.6\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " as shown in the first row of Tab. 7. This implicitly indicates that adversarial learning creates a self-reinforcing cycle, where both" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "3086" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 194 + ], + "type": "image", + "image_path": "8bbc69070db3eff6bbf95fabe805e8f409b74d656e184074e54a20544e96c7be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "text", + "content": "Figure 3. Analyses of hyper-parameters " + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 56, + 201, + 295, + 224 + ], + "type": "text", + "content": " in Eq. 14 and 15. Results denote Recall@1 accuracy on CUB-200-2011." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 238, + 106, + 380 + ], + "blocks": [ + { + "bbox": [ + 58, + 238, + 106, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 238, + 106, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 238, + 106, + 380 + ], + "type": "image", + "image_path": "8bc19777535b8954896cc7c7421fcf894950d8fc63c78ba94880735a029b1e87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 380, + 89, + 387 + ], + "lines": [ + { + "bbox": [ + 80, + 380, + 89, + 387 + ], + "spans": [ + { + "bbox": [ + 80, + 380, + 89, + 387 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 392, + 295, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 437 + ], + "type": "text", + "content": "Figure 4. Visualizations of pattern maps using different feedback: (a) inputs, (b) classification-based feedback, (c) non-discrepancy reconstruction-based feedback, (d) discrepancy reconstruction-based feedback, and (e) our AdvRF." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 106, + 238, + 152, + 380 + ], + "blocks": [ + { + "bbox": [ + 106, + 238, + 152, + 380 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 152, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 152, + 380 + ], + "type": "image", + "image_path": "19ee7885f4c1bb97e1b05b945cdcf269bbba059539d4b85fe0098c3e09a201f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 380, + 133, + 387 + ], + "lines": [ + { + "bbox": [ + 125, + 380, + 133, + 387 + ], + "spans": [ + { + "bbox": [ + 125, + 380, + 133, + 387 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 153, + 238, + 199, + 380 + ], + "blocks": [ + { + "bbox": [ + 153, + 238, + 199, + 380 + ], + "lines": [ + { + "bbox": [ + 153, + 238, + 199, + 380 + ], + "spans": [ + { + "bbox": [ + 153, + 238, + 199, + 380 + ], + "type": "image", + "image_path": "0014a467debb4bc07c8aa489a90e66c93bb3a743d161366c8a6bf0ebee4ad766.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 380, + 181, + 387 + ], + "lines": [ + { + "bbox": [ + 173, + 380, + 181, + 387 + ], + "spans": [ + { + "bbox": [ + 173, + 380, + 181, + 387 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 200, + 238, + 246, + 380 + ], + "blocks": [ + { + "bbox": [ + 200, + 238, + 246, + 380 + ], + "lines": [ + { + "bbox": [ + 200, + 238, + 246, + 380 + ], + "spans": [ + { + "bbox": [ + 200, + 238, + 246, + 380 + ], + "type": "image", + "image_path": "a3fa0953cf128b86d4ca2e1408bbc0fa17073c2b91070b321a418a9ffa085577.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 380, + 227, + 387 + ], + "lines": [ + { + "bbox": [ + 219, + 380, + 227, + 387 + ], + "spans": [ + { + "bbox": [ + 219, + 380, + 227, + 387 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 246, + 238, + 293, + 380 + ], + "blocks": [ + { + "bbox": [ + 246, + 238, + 293, + 380 + ], + "lines": [ + { + "bbox": [ + 246, + 238, + 293, + 380 + ], + "spans": [ + { + "bbox": [ + 246, + 238, + 293, + 380 + ], + "type": "image", + "image_path": "6664c4eea81105b15aeb67f5393612b9abb38de1587e1ce9d781874d58926261.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 380, + 274, + 387 + ], + "lines": [ + { + "bbox": [ + 266, + 380, + 274, + 387 + ], + "spans": [ + { + "bbox": [ + 266, + 380, + 274, + 387 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 483 + ], + "type": "text", + "content": "the retrieval and reconstruction models iteratively challenge and reinforce each other." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": "Hyper-parameter analyses. We conduct sensitivity analyses of the hyperparameters in Eq. 14 and 15, with evaluation results presented in Fig. 3. The performance of our AdvRF shows slight sensitivity to variations in " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ". In our experiments, the default values are set to " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\alpha = 0.7" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\gamma = 0.6" + }, + { + "bbox": [ + 55, + 485, + 295, + 556 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Effect of pattern maps with various feedback. Fig. 4 illustrates the impact of various feedback signals on discrepancy localization. When using classification-based feedback, the model struggles to accurately localize discrepancies. However, formulating FGIR as a visual discrepancy reconstruction task enhances discrepancy localization through reconstruction-based feedback. Notably, limited reconstruction signals, such as using only discrepancy-based feedback (" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{A}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " in Eq. 4) or non-discrepancy-based feedback (" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{\\mathbf{R}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": " in Eq. 4), may cause pattern maps to overlook certain discrepancies. Our results suggest that combining comprehensive reconstruction feedback creates a self-reinforcing cycle, where both the retrieval and reconstruc" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 414, + 163 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 414, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 414, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 414, + 163 + ], + "type": "image", + "image_path": "2f55175f45a83489e6e34a7c370f3bc6ad3b59090ac3607bf8d1b2aceecbd651.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 167, + 427, + 177 + ], + "lines": [ + { + "bbox": [ + 320, + 167, + 427, + 177 + ], + "spans": [ + { + "bbox": [ + 320, + 167, + 427, + 177 + ], + "type": "text", + "content": "(a) Category-related Embeddings" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 416, + 72, + 432, + 163 + ], + "blocks": [ + { + "bbox": [ + 416, + 72, + 432, + 163 + ], + "lines": [ + { + "bbox": [ + 416, + 72, + 432, + 163 + ], + "spans": [ + { + "bbox": [ + 416, + 72, + 432, + 163 + ], + "type": "image", + "image_path": "c57401f9b835fed62a18d0654ee57290683d405b28db86189c1465fcd03fc9bc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "type": "text", + "content": "Figure 5. Evaluating the efficacy of category-related embeddings from baseline [10] versus category-agnostic embeddings from our AdvRF, with their similarity in grid formats." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 547, + 163 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 547, + 163 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 547, + 163 + ], + "type": "image", + "image_path": "df6e6371e14f4ac7bc37d1e9769bb961f37c16cf38498095f529b0cd175b9f39.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 167, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 433, + 167, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 433, + 167, + 545, + 177 + ], + "type": "text", + "content": "(b) Category-agnostic Embeddings" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 240, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 456 + ], + "type": "text", + "content": "tion models iteratively challenge and strengthen each other. Analysis of category-agnostic discrepancies. We employ an indirect method to interpret category-agnostic discrepancies by comparing the similarities between category-agnostic retrieval embeddings produced by our AdvRF and category-related retrieval embeddings generated by the classification-based feedback illustrated in Fig. 1a. As illustrated in Fig. 5, we compute these similarities for the ten high-similarity images across five novel subcategories. Our analysis reveals that category-agnostic embeddings effectively highlight nearest sample pairs that belong to the same subcategory. In contrast, category-related retrieval embeddings struggle to identify these high-similarity images, as they tend to capture visual discrepancies associated with the semantics of base categories. Overall, our AdvRF successfully learns category-agnostic descriptions for unseen subcategories by formulating FGIR tasks as a visual discrepancy reconstruction process." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "content": "In this paper, we introduce AdvRF, which acquires category-agnostic visual discrepancies by formulating FGIR as a visual discrepancy reconstruction task. AdvRF designs an adversarial pipeline: the reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. As a result, AdvRF precisely localizes visual differences and encodes them into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Importantly, our algorithm is end-to-end trainable and achieves state-of-the-art performance on the widely-used fine-grained and coarse-grained retrieval datasets." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (NSFC) under Grant (No.61932020), and the Taishan Scholar Program of Shandong Province (tstp20221128)." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3087" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "text", + "content": "[1] Kenan E. Ak, Ashraf A. Kassim, Joo-Hwee Lim, and Jo Yew Tham. Learning attribute representations with localization for flexible fashion search. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7708-7717. Computer Vision Foundation / IEEE Computer Society, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 225 + ], + "type": "text", + "content": "[2] Malik Boudiaf, Jérôme Rony, Imtiaz Masud Ziko, Eric Granger, Marco Pedersoli, Pablo Piantanida, and Ismail Ben Ayed. A unifying mutual information view of metric learning: Cross-entropy vs. pairwise losses. In ECCV, pages 548–564. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 227, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 227, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 227, + 295, + 270 + ], + "type": "text", + "content": "[3] Steve Branson, Grant Van Horn, Serge J. Belongie, and Pietro Perona. Bird species categorization using pose normalized deep convolutional nets. CoRR, abs/1406.2952, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 338 + ], + "type": "text", + "content": "[4] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Chojui Hsieh, and Boqing Gong. Robust and accurate object detection via adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 16622-16631. Computer Vision Foundation / IEEE, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 339, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 339, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 61, + 339, + 296, + 407 + ], + "type": "text", + "content": "[5] Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed M. Elgammal. Link the head to the \"beak\": Zero shot learning from noisy text description at part precision. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 6288-6297. IEEE Computer Society, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 407, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 296, + 474 + ], + "type": "text", + "content": "[6] Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xiaoming Wei, and Xiaolin Wei. Masked auto-encoders meet generative adversarial networks and beyond. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 24449-24459. IEEE, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 529 + ], + "type": "text", + "content": "[7] Yaroslav Ganin and Victor S. Lempitsky. Unsupervised domain adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1180-1189. JMLR.org, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 532, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 296, + 576 + ], + "type": "text", + "content": "[8] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "text", + "content": "[9] Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578-1604, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 58, + 624, + 296, + 667 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778, 2016. 5, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised domain adaptation. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "UT, USA, June 18-22, 2018, pages 1498-1507. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 151 + ], + "type": "text", + "content": "[12] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown categories and camera poses. In International Conference on 3D Vision, 3DV 2024, Davos, Switzerland, March 18-21, 2024, pages 31-41. IEEE, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[13] Shichao Kan, Zhiquan He, Yigang Cen, Yang Li, Vladimir Mladenovic, and Zhihai He. Contrastive bayesian analysis for deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7220-7238, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[14] Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Embedding transfer with label relaxation for improved metric learning. In CVPR, pages 3967-3976. Computer Vision Foundation / IEEE, 2021. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[15] ByungSoo Ko, Geonmo Gu, Han-Gyu Kim, and ByungSoo Ko. Learning with memory-based virtual classes for deep metric learning. In ICCV, pages 11772-11781. IEEE, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "text", + "content": "[16] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 793-802. IEEE, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 388 + ], + "type": "text", + "content": "[17] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554-561, 2013. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 441 + ], + "type": "text", + "content": "[18] Haojie Li, Mingxuan Li, Qijie Peng, Shijie Wang, Hong Yu, and Zhihui Wang. Correlation-guided semantic consistency network for visible-infrared person re-identification. IEEE Trans. Circuits Syst. Video Technol., 34(6):4503-4515, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 499 + ], + "type": "text", + "content": "[19] Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 212-222. IEEE, 2022. 1, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[20] Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. DAS: densely-anchored sampling for deep metric learning. In ECCV, pages 399–417. Springer, 2022. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "text", + "content": "[21] Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1096-1104. IEEE Computer Society, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[22] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. CoRR, abs/1306.5151, 2013. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[23] Olga Moskvyak, Frédéric Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Keypoint-aligned embeddings for image retrieval and re-identification. In WACV, pages 676–685. IEEE, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[24] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3088" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 117 + ], + "type": "text", + "content": "In Medical Image Computing and Computer-Assisted Intervention - MICCAI 2015 - 18th International Conference Munich, Germany, October 5 - 9, 2015, Proceedings, Part III, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[25] Karsten Roth, Timo Milbich, Björn Ommer, Joseph Paul Cohen, and Marzyeh Ghassemi. Simultaneous similarity-based self-distillation for deep metric learning. In ICML, pages 9095–9106. PMLR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 218 + ], + "type": "text", + "content": "[26] Karsten Roth, Oriol Vinyals, and Zeynep Akata. Nonisotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7410-7420. IEEE, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "text", + "content": "[27] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In CVPR, pages 815-823. IEEE Computer Society, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 297 + ], + "type": "text", + "content": "[28] Jenny Seidenschwarz. Learning intra-batch connections for deep metric learning. In ICML, pages 9410-9421. PMLR, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 354 + ], + "type": "text", + "content": "[29] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 399 + ], + "type": "text", + "content": "[30] Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, pages 4004-4012. IEEE Computer Society, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 401, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 295, + 445 + ], + "type": "text", + "content": "[31] Eu Wern Teh, Terrance DeVries, Graham W. Taylor, and Graham. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, pages 448-464. Springer, 2020. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 446, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 446, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 446, + 295, + 479 + ], + "type": "text", + "content": "[32] Chengkun Wang, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Introspective deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 46(4):1964-1980, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 480, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 480, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 480, + 295, + 568 + ], + "type": "text", + "content": "[33] Shijie Wang, Zhihui Wang, Haojie Li, and Wanli Ouyang. Category-specific nuance exploration network for fine-grained object retrieval. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Intelligence, IAAI 2022, The Twelveth Symposium on Educational Advances in Artificial Intelligence, EAAI 2022 Virtual Event, February 22 - March 1, 2022, pages 2513-2521. AAAI Press, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "text", + "content": "[34] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 19381-19391. IEEE, 2023. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 712 + ], + "type": "text", + "content": "[35] Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Learning to parameterize visual attributes for open-set fine-grained retrieval. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "text", + "content": "[36] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Fine-grained retrieval prompt tuning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 2644-2652. AAAI Press, 2023. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[37] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Semantic-guided information alignment network for fine-grained image recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 1-1, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[38] Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Content-aware rectified activation for zero-shot fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 46(6):4366-4380, 2024. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[39] Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Accurate fine-grained object recognition with structure-driven relation graph networks. Int. J. Comput. Vis., 132(1):137-160, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[40] Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022-5030. Computer Vision Foundation / IEEE, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 350, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 448 + ], + "type": "text", + "content": "[41] Zhuhui Wang, Shijie Wang, Haojie Li, Zhi Dou, and Jianjun Li. Graph-propagation based correlation learning for weakly supervised fine-grained image classification. In The Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence Conference, IAAI 2020, The Tenth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2020, New York, NY, USA, February 7-12, 2020, pages 12289-12296. AAAI Press, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 449, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 524 + ], + "type": "text", + "content": "[42] Zhihui Wang, Shijie Wang, Shuhui Yang, Haojie Li, Jianjun Li, and Zezhou Li. Weakly supervised fine-grained image classification via gaussian mixture model oriented discriminative learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 13-19, 2020, pages 9746-9755. IEEE, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 525, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 591 + ], + "type": "text", + "content": "[43] Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural networks via image quantization and contrastive adversarial learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 15054-15063. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[44] Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, and Zhi-Hua Zhou. Selective convolutional descriptor aggregation for fine-grained image retrieval. IEEE Trans. Image Process., 26(6):2868-2881, 2017. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[45] Xiu-Shen Wei, Yang Shen, Xuhao Sun, Peng Wang, and Yuxin Peng. Attribute-aware deep hashing with self-consistency for large-scale fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 45(11):13904-13920, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[46] Bailin Yang, Haoqiang Sun, Frederick W. B. Li, Zheng Chen, Jianlu Cai, and Chao Song. HSE: hybrid species embedding" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3089" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 419 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "for deep metric learning. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 11013-11023. IEEE, 2023. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[47] Xianxian Zeng, Shun Liu, Xiaodong Wang, Yun Zhang, Kairui Chen, and Dong Li. Hard decorrelated centralized loss for fine-grained image retrieval. Neurocomputing, 453: 26-37, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[48] Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. In BMVC, page 91. BMVA Press, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[49] Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In CVPR, pages 9320-9329. Computer Vision Foundation / IEEE, 2021. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[50] Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In ICCV, pages 12045-12054. IEEE, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[51] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, Feiyue Huang, and Yanhua Yang. Centralized ranking loss with weakly supervised localization for fine-grained object retrieval. In IJCAI, pages 1226-1233. ijcai.org, 2018. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "text", + "content": "[52] Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Baochang Zhang, Yongjian Wu, and Feiyue Huang. Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer. In AAAI, pages 9291-9298. AAAI Press, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 419 + ], + "type": "text", + "content": "[53] Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2138-2147. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3090" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_content_list.json b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..32439dd2e076b0b50ac0df1138398bc6daf67992 --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "Adversarial Robust Memory-Based Continual Learner", + "text_level": 1, + "bbox": [ + 220, + 130, + 777, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyue $\\mathrm{Mi}^{1,2}$ , Fan Tang $^{1,2*}$ , Zonghan Yang $^{3,4}$ , Danding Wang $^{1,2}$ , Juan Cao $^{1,2}$ , Peng Li $^{3}$ , Yang Liu $^{3,4}$", + "bbox": [ + 102, + 178, + 895, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences", + "bbox": [ + 228, + 198, + 761, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Institute for AI Industry Research (AIR), Tsinghua University", + "bbox": [ + 220, + 233, + 748, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Department of Computer Science & Technology, Tsinghua University", + "bbox": [ + 217, + 252, + 774, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 314, + 326, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the remarkable advances that have been made in continual learning, the adversarial vulnerability of such methods has not been fully discussed. We delve into the adversarial robustness of memory-based continual learning algorithms and observe limited improvement in robustness by directly applying adversarial training techniques. Our preliminary studies reveal the twin challenges for building adversarial robust continual learners: accelerated forgetting in continual learning and gradient obfuscation in adversarial robustness. In this study, we put forward a novel adversarial robust memory-based continual learner that adjusts data logits to mitigate the forgetting of pasts caused by adversarial samples. Furthermore, we devise a gradient-based data selection mechanism to alleviate the gradient obfuscation caused by limited stored data. The proposed approach can widely integrate with existing memory-based continual learning and adversarial training algorithms. Extensive experiments on Split-CIFAR10/100 and Split-TinyImageNet demonstrate the effectiveness of our approach, achieving a maximum forgetting reduction of $34.17\\%$ in adversarial data for ResNet, and $20.10\\%$ for ViT.", + "bbox": [ + 89, + 347, + 483, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 695, + 220, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Continual learning [48, 62] enables intelligent models to continually acquire new knowledge and adapt to changing environments while maintaining previous capabilities. It has recently emerged as a promising direction to achieve ideal intelligent models, and some pioneer works have investigated the robustness of continual learning in various aspects [42, 53, 56]. Achieving robust continual learners represents a crucial advancement in enabling intelligent models to adapt to dynamic scenarios effectively.", + "bbox": [ + 89, + 720, + 482, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, when dealing with streaming non-i.i.d. data,", + "bbox": [ + 109, + 859, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1cc626a1108478708580c34b8405ebfe88e9f524f95937419d79f94d89cd1df6.jpg", + "image_caption": [ + "Figure 1. A preliminary analysis of the combination of continual learning and adversarial training on Split-CIFAR10 with buffer size 200. We consider the final average accuracy (FAA) and forgetting of adversarial and clean samples in both class incremental and task incremental settings. The red numbers indicate the performance after adding our approach." + ], + "image_footnote": [], + "bbox": [ + 532, + 311, + 888, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "continual learners are usually vulnerable to adversarial examples [23]. Chen et al. [12] first attempt at preserving adversarial robustness in continual learning [29] by leveraging adversarial training [52, 57] and a large amount of unlabeled data. Despite the impressive results of Chen et al. [12], the inherent challenges for conducting adversarial robust continual learners are still under disclosure.", + "bbox": [ + 511, + 579, + 906, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this study, we investigate building adversarial robust continual learners without abundant unlabeled data. We select two representative approaches for continual learning and adversarial defenses: memory-based continual learning and adversarial training. Figure 1 shows the direct combination of memory-based continual learning (ER [10]) with an adversarial training approach (TRADES [57]). The results for adversarial data reconfirm that continual learners risk adversarial attacks [12, 20, 23, 24], and demonstrate a critical trade-off: while adversarial training can strengthen continual learners against attacks [12, 20, 23, 24], it simultaneously increases forgetting on clean data, thus compromising the core objective of continual learning.", + "bbox": [ + 511, + 686, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Further experimental analysis in Sec. 3 reveals two key", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "562", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenges when combining adversarial training with continual learning: (1) Adversarial data increase negative gradients toward past classes compared with clean data, accelerating catastrophic forgetting in continual learning; (2) Insufficient historical data in continual learning leads to gradient obfuscation [4], compromising the effectiveness of adversarial training compared to joint training. These challenges emerge naturally from the direct integration of the two approaches.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the analysis, we propose an adversarial robust memory-based continual learner without a large amount of additional data, which consists of an anti-forgettable logit calibration (AFLC) module and a robustness-aware experience replay (RAER) strategy. The AFLC module designs different logit calibration strategies for past, current, and future classes, thus mitigating the negative gradients of adversarial samples for continual learning. By leveraging attacking difficulty, RAER selects data that are adversarial safety and maintains the diversity of data distribution, rather than storing data overfitting to the decision boundary [9]. Our approach can be combined with most memory-based continual learning as well as adversarial training algorithms, and we have demonstrated the effectiveness of our approach on the Split-CIFAR10/100 and more challenging dataset Split-Tiny-ImageNet across ResNet and ViT architectures.", + "bbox": [ + 91, + 227, + 483, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 469, + 434, + 482 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the twin critical challenges to carrying out adversarial robust memory-based continual learning algorithms: accelerated forgetting for continual learning and gradient obfuscation for adversarial robustness.", + "- We propose antiforgettable logit calibration to reduce negative influences from adversarial training through task-adaptive logit calibration, and robustness-aware experience replay by storing robust and distribution-representative past data to alleviate gradient obfuscation.", + "- Experiments on Split-CIFAR10/100 and Split-TinyImageNet across ResNet and ViT demonstrate our effectiveness, in improving both clean accuracy and adversarial robustness of continual learning and mitigating forgetting in both class-incremental and task-incremental settings without additional data." + ], + "bbox": [ + 91, + 484, + 482, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 724, + 240, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual learning. Continual learning [38, 48, 62] enables models to adapt to evolving data distributions while preserving past knowledge. To alleviate catastrophic forgetting, existing approaches [34, 39] can be categorized into three types: memory-based [7, 8, 10], regularization-based [25, 29, 30], and dynamic architecture [17, 41, 50, 55, 61]. Continual learning can be categorized into class incremental (class-IL) and task-incremental (task-IL) settings [33] based on the availability of task information during inference. In both settings, memory-based meth", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods have demonstrated superior performance without requiring model expansion [7, 60]. Researchers have explored continual learning robustness under various constraints [19, 53, 56], and saft risks about backdoor attacks [1, 22, 47] and privacy concerns [21]. Recent studies have revealed both the vulnerability of continual models to adversarial attacks and the potential of adversarial samples for mitigating forgetting [20, 23, 24, 27, 49]. Chen et al. [12] combined LwF [29] with adversarial training using additional data, while our work uniquely analyzes the core challenges in achieving adversarial robustness for continual learning without such requirements.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial defense. The vulnerability of deep neural networks to adversarial examples [18, 46, 52, 59] has led to numerous defense methods [2, 5], with adversarial training emerging as the most effective approach [14, 15, 28, 31, 32, 36, 37, 57]. Recent research has extended to real-world scenarios, such as long-tailed distributions [54] and open-world settings [43]. Meanwhile some studies show the potential of continual algorithms in adapting to new attacks [13, 40]. Different from them, we specifically address adversarial robustness in class-incremental and task-incremental settings.", + "bbox": [ + 511, + 275, + 906, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Baselines and Their Inherent Problems", + "text_level": 1, + "bbox": [ + 511, + 465, + 864, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 494, + 653, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memory-based continual learning. Following [6, 7], we consider continual learning across $T$ distinct classification tasks. We pre-allocate classification heads for all tasks at initialization rather than expanding them incrementally. For the $t$ -th task ( $1 \\leq t \\leq T$ ), we denote its data distribution as $D_{t}$ , with input samples $x_{t}$ and corresponding labels $y_{t}$ . A model $\\mathrm{M}_{\\theta}(\\cdot)$ with parameters $\\theta$ is optimized sequentially in the given task order. Let $h_{\\theta}(x_{t}) \\in \\mathbb{R}^{C}$ denote the output logits where $h_{\\theta}(x_{t}) = \\mathrm{M}_{\\theta}(x_{t})$ , and $C$ is the total number of classes. The softmax output is defined as $f_{\\theta}(x_{t}) \\triangleq \\operatorname{softmax}(h_{\\theta}(x_{t}))$ . Given a logit vector $h_{\\theta}(x_{t}) = [h_{\\theta}(x_{t})_{1}, \\dots, h_{\\theta}(x_{t})_{C}]$ where $h_{\\theta}(x_{t})_{i}$ represents the logit for the $i$ -th class, we partition the classes into three sets based on the current task $t$ :", + "bbox": [ + 511, + 520, + 906, + 731 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np = \\{1, 2, \\dots , (t - 1) b \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 748, + 733, + 765 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nc = \\{(t - 1) b + 1, (t - 1) b + 2, \\dots , t b \\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 768, + 903, + 784 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nu = \\{t b + 1, t b + 2, \\dots , C \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 787, + 759, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $b$ is the number of classes per task. Accordingly, we divide $h_{\\theta}(x_t)$ into three parts: $h_{\\theta}(x_t)_p$ , $h_{\\theta}(x_t)_c$ , and $h_{\\theta}(x_t)_u$ , representing logits for past, current, and future classes respectively.", + "bbox": [ + 511, + 821, + 903, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Continual learner aims to correctly classify data of all", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "563", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/45dc768fbc4142940637d0cadb7dda8f3c88608138c9030dd14446c1122f5d5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Buffer SizeMethodW/ ATClass Incremental SettingTask Incremental SettingCRD↓FRI↓RRD↓
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
-Joint (upper bound)X91.81±0.07-0.00±0.00-98.17±0.12-0.00±0.00-7.92--
✓79.33±0.36-50.93±1.28-94.80±0.30-74.63±1.13-
-SGD (lower bound)X19.66±0.0796.39±0.675.17±2.5833.31±5.1966.84±6.5037.42±7.394.44±3.7833.34±3.490.91-3.7341.10
✓19.34±0.0591.93±0.2015.65±0.1869.71±1.7765.33±5.9134.43±7.1627.71±1.8854.56±3.47
200ERX48.80±1.8460.00±2.250.27±0.1812.60±2.1392.89±0.265.00±0.260.01±0.0113.78±0.4114.5112.9131.70
✓28.18±0.6980.58±1.0517.86±0.2969.58±1.1584.49±0.6110.23±0.9844.30±1.0536.89±1.11
5120X83.34±1.3815.35±1.710.08±0.1113.65±2.3496.84±0.170.71±0.190.06±0.0421.83±8.6913.5312.1118.80
✓61.88±0.7437.72±0.7327.28±0.5641.66±1.2291.24±0.192.56±0.1256.59±0.8819.34±1.11
200DERX62.71±2.7638.75±4.110.01±0.007.42±2.1391.32±0.836.95±1.440.09±0.057.39±2.1817.7215.7334.80
✓35.79±2.8565.67±4.1616.44±0.2152.84±3.7182.79±0.2111.48±0.6639.52±0.3740.64±1.38
5120X83.99±0.9910.57±1.510.07±0.077.05±2.0395.67±0.131.63±0.130.05±0.057.39±0.4816.7415.6624.32
✓56.58±3.8740.40±4.0723.26±0.4445.06±1.4889.62±0.673.12±0.2353.66±2.0122.95±3.68
200DER++X65.55±2.8132.17±4.260.24±0.327.38±3.3792.11±0.558.85±2.980.22±0.3310.45±0.9414.478.2437.14
✓45.43±3.2846.67±5.0914.46±1.3535.09±2.7083.29±1.0710.83±1.7736.82±0.5934.09±2.23
5120X85.74±0.657.32±0.881.51±1.276.10±0.5396.13±0.531.14±0.350.82±0.526.06±0.4910.638.2122.18
✓68.67±1.1123.68±2.6626.95±0.2927.68±2.4891.95±0.331.20±0.9054.26±0.4116.87±1.72
200X-DERX60.96±1.0912.29±0.090.00±0.003.84±1.1894.33±0.632.49±0.170.13±0.032.59±0.6420.237.6524.02
✓34.04±0.9025.13±4.0316.82±0.9827.84±4.1680.8±0.664.96±1.6660.83±0.5610.99±0.81
5120X69.93±0.634.33±0.260.21±0.124.52±1.3996.94±0.020.29±0.040.13±0.022.04±0.4324.9618.5221.27
✓34.29±0.6538.92±0.0319.24±0.1038.79±0.4182.67±1.682.74±0.8364.11±0.128.14±0.98
", + "bbox": [ + 94, + 88, + 906, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. We conduct a systematic experimental analysis combining adversarial training (specifically Vanilla AT [31]) with memory-based continual learning algorithms. See Sec. 3.2 for a detailed summary of observations. Performance is evaluated using several metrics (detailed in Sec. 5.1): Final Average Accuracy (FAA), Forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase on clean data (FRI), and Robust Relative Decrease (RRD) compared to joint training. All adversarial samples are generated using PGD-20. Note that the reported CRD, FRI, and RRD values are averaged across both task-incremental and class-incremental settings.", + "bbox": [ + 89, + 409, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "observed tasks, and its objective can be formalized as:", + "bbox": [ + 89, + 484, + 447, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {t}, \\text {w h e r e} \\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {\\left(x _ {t}, y _ {t}\\right) \\sim D _ {t}} \\left[ \\ell \\left(f _ {\\theta} \\left(x _ {t}\\right), y _ {t}\\right) \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 506, + 480, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell$ is the task-specific loss. The unavailability of past task data poses great challenges for continual learning. However, direct sequential optimization can lead to severe catastrophic forgetting. To address the above challenge, memory-based continual learners usually store a few past task data and replay them during the current task training phase [7], and become leading-edge approaches. When training the $t$ -th task, its loss function can be formalized as the following:", + "bbox": [ + 89, + 561, + 482, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} [ \\ell (f _ {\\theta} (x), y) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 704, + 482, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is the episodic memory of the current task, which preserves a small subset of observed past data $\\langle x_{\\mathcal{M}},y_{\\mathcal{M}}\\rangle$", + "bbox": [ + 89, + 729, + 482, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial training. Adversarial training is currently the most reliable way [5] to improve the adversarial robustness of a model. Following [31], it can be formally defined as a min-max optimization problem:", + "bbox": [ + 89, + 758, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 827, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta$ is an adversarial perturbation, $D$ is data distribution, and $S$ is a perturbation set. $S = \\left\\{\\delta \\mid \\| \\delta \\|_p \\leq \\epsilon \\right\\}$ , $\\epsilon$ is", + "bbox": [ + 89, + 864, + 483, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the perturbation size. The goal of adversarial training is to correctly classify all adversarial examples $(\\widetilde{x} = x + \\delta)$ .", + "bbox": [ + 511, + 484, + 903, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we integrate memory-based continual learner and adversarial training as adversarial robust memory-based continual learner baselines, and the new optimization objective during the training stage of $t$ -th task is:", + "bbox": [ + 511, + 515, + 905, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 582, + 903, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Vanilla Solutions", + "text_level": 1, + "bbox": [ + 511, + 621, + 679, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate the challenges in adversarial robust continual learning, we analyze several popular memory-based continual learning methods (ER [10], DER [7], DER++ [7], and X-DER [6]) on Split-CIFAR10 [26] under both class and task incremental settings. We employ Vanilla AT [31] for adversarial training due to its classic and strong performance [36]. Performance is evaluated against an upper bound (\"Joint\") and a lower bound (\"SGD\") using multiple metrics: Final Average Accuracy (FAA), forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase in clean data (FRI), and Robust Relative Decrease (RRD). Detailed metric definitions are provided in Sec. 5.1, and additional implementation details and hyper-parameters are provided in the supplementary material.", + "bbox": [ + 509, + 643, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Results in Table 1 reveal that direct integration of adversarial training with continual learning is suboptimal. First, the results validate our observations from Sec. 1: while", + "bbox": [ + 511, + 854, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "564", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adversarial training enhances robustness, it compromises clean performance (positive CRDs) and accelerates forgetting (positive FRIs). Larger memory sizes consistently improve performance across all metrics. Moreover, continual models show inferior robustness improvement compared to joint models (positive RRDs).", + "bbox": [ + 88, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To investigate the underlying mechanisms, we select ER as our base model for further analysis. While DER/DER++ and X-DER demonstrate better stability with lower forgetting, they struggle with new task learning, resulting in poor robust FAA. In contrast, ER combined with adversarial training exhibits strong robustness.", + "bbox": [ + 89, + 181, + 482, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. The Inherent Problems", + "text_level": 1, + "bbox": [ + 89, + 279, + 305, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By analyzing gradients during training, we identify two critical challenges in robust continual learning: accelerated forgetting in continual learning and gradient obfuscation in adversarial training.", + "bbox": [ + 89, + 301, + 482, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accelerated forgetting in continual learning. Adversarial training hastens the forgetting of clean data in continual learners. According to [31], adversarial examples are designed to mislead model predictions by increasing probabilities for incorrect classes while suppressing those of the true class. In the continual learning setting, adversarial examples from the current task tend to increase probabilities for past classes while decreasing probabilities for their true class. Specifically, for an adversarial example $\\widetilde{x}_t$ generated from a clean sample $x_t$ of current task $t$ :", + "bbox": [ + 89, + 362, + 482, + 512 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {p} > f _ {\\theta} \\left(x _ {t}\\right) _ {p}, f _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {\\text {t r u e c l a s s}} < f _ {\\theta} \\left(x _ {t}\\right) _ {\\text {t r u e c l a s s}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 518, + 480, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When using cross-entropy as the loss function, adversarial data amplify negative gradients for current data $x_{t}$ for past tasks:", + "bbox": [ + 89, + 541, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (\\widetilde {x} _ {t})} \\right] _ {p} = f _ {\\theta} (\\widetilde {x} _ {t}) _ {p} > f _ {\\theta} (x _ {t}) _ {p} = \\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (x _ {t})} \\right] _ {p}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 588, + 480, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our empirical analysis reveals that these adversarial examples generate stronger negative gradients toward past tasks, especially during the initial training phases. In Figure 2, experiments with $\\mathrm{ER + AT}$ on Split-CIFAR10 using a buffer size of 200 show adversarial data consistently produces larger gradient norms compared to clean data. The larger gradient norms indicate stronger positive gradients for the true class and stronger negative gradients for other classes, including past classes, thus accelerating forgetting of previously learned tasks. While memory-based methods using stored adversarial samples can help mitigate catastrophic forgetting, the limited buffer size (typically small in practical applications) means that adversarial training on the entire dataset still accelerates forgetting.", + "bbox": [ + 89, + 628, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gradient obfuscation in adversarial training. Gradient obfuscation [4] occurs when gradient information is obscured by non-differentiable operators or randomness, leading to inferior robustness. We observe a specific form of", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/388c31bd51e2ed97f3e1bfe631bb4f02b505d540bcaaca5f035f2dccd035589d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 87, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bac76602735a5d52be05b919cc8d47fc54f62544c4858256ea4b9df9e3b2ea2c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 220, + 903, + 344 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b44940afc77cdc7474f845a958bf459d85850c453c74c6be9c157ded6c23237b.jpg", + "image_caption": [ + "Figure 2. Gradient norms of adversarial data and clean data during the training stage. We experiment ER+AT with a 200 buffer size on the Split-CIFAR10 dataset.", + "Figure 3. The gradient cosine similarities between robust joint model and robust continual model, which indicates a special gradient obfuscation only in continual learning. We experiment with buffer size=200/5,120 on the Split-CIFAR10 dataset. Gradient obfuscation weakens as the buffer size gets larger, and our method also mitigates gradient obfuscation." + ], + "image_footnote": [], + "bbox": [ + 550, + 397, + 857, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gradient obfuscation, termed \"shattered gradients\", in adversarial training under continual learning, resulting in less effective robustness compared to joint models.", + "bbox": [ + 511, + 598, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In continual learning, gradient information is biased due to limited replay data. With few old task samples, the model tends to overfit, causing direction-biased gradients when generating adversarial examples. Our analysis of gradient directions between adversarial robust joint and continual models reveals significant differences, contributing to lower adversarial robustness in continual learners (see Figure 3).", + "bbox": [ + 511, + 643, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 762, + 604, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 787, + 624, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the identified challenges, we propose a novel framework consisting of two key components: antiforgettable logit calibration (AFLC) to mitigate forgetting acceleration, and robustness-aware experience replay (RAER) to reduce gradient obfuscation. The full pipeline is illustrated in Figure 4: during training each task, the", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "565", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a493aef0c46f4adcd6adc6b6bd32607dc81091936c0bf940dab7ef0a1615dbbe.jpg", + "image_caption": [ + "Figure 4. The pipeline of our adversarial robust memory-based continual learner. First, generate adversarial samples $(\\widetilde{X}_t,\\widetilde{X}_{\\mathcal{M}})$ and robust difficulty factors $(K_{t})$ for both current task data $(X_{t},Y_{t})$ and memory data $(X_{\\mathcal{M}},Y_{\\mathcal{M}})$ . The model then processes these samples through AFLC (Sec. 4.2), which calibrates logits based on task order to reduce negative gradients toward past tasks. Meanwhile, RAER (Sec. 4.3) leverages the robust difficulty factors $K_{t}$ to select representative samples for memory storage, mitigating gradient obfuscation." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 671, + 306 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0b7d85f31fde0b9ee1c57faeae9cef9a37694110bbc4e526a4700c20d5926981.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 89, + 897, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model processes both current task data and memory samples through the model. AFLC then calibrates logits based on task order (past/current/future) before loss computation. Simultaneously, RAER evaluates current task samples using a robustness difficulty factor $k$ to select representative data for memory storage.", + "bbox": [ + 89, + 381, + 483, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Anti-Forgettable Logit Calibration", + "text_level": 1, + "bbox": [ + 89, + 479, + 393, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce AFLC to regulate the influence of adversarial examples on past task knowledge. The core idea is to calibrate logits differently for past and current tasks to control gradient flow.", + "bbox": [ + 89, + 502, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an input $x$ (either from current task $x_{t}$ or memory buffer $x_{\\mathcal{M}}$ ), AFLC applies a task-dependent calibration:", + "bbox": [ + 89, + 563, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nh _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = h _ {\\theta} (x) _ {i} - v _ {i}, v _ {p} > v _ {c}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 601, + 480, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $v_{i}$ represents the calibration factor for class i:", + "bbox": [ + 89, + 627, + 434, + 642 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = - \\log \\left(\\frac {n _ {i}}{\\sum_ {j = 1} ^ {C} n _ {j}}\\right) - \\alpha_ {i}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 648, + 482, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_i$ denotes sample count for class $i$ , and $\\alpha_i$ is a hyperparameter. When all tasks share the classification head, we realize that making logit changes to past and current classes only is equivalent to negative gradient increases. So we make a Further Prior (FP) adjustment that essentially averages the adjustments applied to current and past classes and applies this value to future classes, helping to balance the gradients across all periods.", + "bbox": [ + 89, + 690, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = \\frac {1}{t \\cdot b} \\sum_ {j = 1} ^ {t \\cdot b} \\left(v _ {j}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 820, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t$ is the current task order and $b$ is the number of classes per task.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The calibrated probability distribution is computed as:", + "bbox": [ + 532, + 381, + 890, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = \\frac {\\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i}\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {j}\\right)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 407, + 903, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and leads to two key properties: For current task samples,", + "bbox": [ + 511, + 455, + 893, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x})} \\right] _ {p} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {t}) _ {p} < f _ {\\theta} (\\widetilde {x} _ {t}) _ {p}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 479, + 903, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and for memory buffer samples,", + "bbox": [ + 513, + 527, + 725, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x _ {\\mathcal {M}}})} \\right] _ {c} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {\\mathcal {M}}) _ {c} > f _ {\\theta} (\\widetilde {x} _ {\\mathcal {M}}) _ {c}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 551, + 903, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In that way, AFLC can mitigate forgetting acceleration by reducing the negative gradient to past classes from current adversarial data and improving the negative gradient to current classes from memory adversarial data.", + "bbox": [ + 511, + 597, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparisons with other logit calibration CL. X-DER [6] also adjusts the negative gradients of the new task data affecting the past tasks by updating the new task with the logit masking. Logit masking can be seen as a special case of our method at the softmax layer, which can be viewed as $\\mathrm{v}_p = \\mathrm{v}_c = 0$ for past data, and $\\mathrm{v}_p = +\\infty$ , $\\mathrm{v}_c = 0$ for current data. However, their extreme margin $\\mathbf{V}$ inhibits the model's ability to learn new knowledge. X-DER further adds a safeguard threshold to limit the activations of past and future heads, which also can be combined with us.", + "bbox": [ + 511, + 657, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Robustness-Aware Experience Replay", + "text_level": 1, + "bbox": [ + 511, + 818, + 839, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address gradient obfuscation, we propose RAER, which selects samples based on their robustness characteristics. For each sample $x$ , we measure its robustness difficulty factor $k$ during PGD-10 attack, without additional forward or", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "566", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "backward calculation, where $k \\in [0,10] \\cap \\mathbb{N}$ indicates the number of successful attacks. Samples meeting the threshold condition $k < \\rho$ are candidates for memory storage. The memory update process can be formalized as:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} \\leftarrow \\text {R a n d o m S a m p l e} \\left(\\{x \\mid k _ {x} < \\rho \\}, m\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 165, + 483, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m$ is the memory size and $\\rho$ is set to 5 in practice.", + "bbox": [ + 89, + 195, + 464, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Why RAER works. The robustness difficulty factor $k$ shows how easily a sample can be attacked: a high $k$ suggests vulnerability and proximity to decision boundaries, while a low $k$ implies robustness and distance from boundaries. We selectively store samples with $k < \\rho$ in memory, excluding boundary-overfitting outliers. Unlike prior continual learning [9] and adversarial training [58] methods, RAER emphasizes distribution-representative and robust learnable samples while maintaining model reliability.", + "bbox": [ + 89, + 210, + 483, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The computational overhead of RAER is minimal, as $k$ is calculated during the existing adversarial example generation process, without additional forward or backward calculation.", + "bbox": [ + 89, + 348, + 483, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with other data-selection CL. Existing works [3, 44] alleviate catastrophic forgetting using data selection and focus solely on clean samples. RAER uniquely targets gradient obfuscation in continual robust settings by: Selecting adversarially safe samples for better retention; excluding high- $k$ boundary samples; and maintaining representative data distribution. As demonstrated in Sec. 5, RAER outperforms traditional data selection strategies combined with adversarial training, particularly in adversarial robustness.", + "bbox": [ + 89, + 409, + 483, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Overall Objective", + "text_level": 1, + "bbox": [ + 89, + 571, + 264, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To sum up, we design AFLC to reduce negative gradients of the current adversarial sample to past tasks, and RAER to select data that alleviate gradient obfuscation to store. We set $\\alpha$ of AFLC is 3.5 in practice and $x_{\\mathcal{M}}$ and $y_{\\mathcal{M}}$ are selected by RAER with $\\rho = 5$ . When $\\mathrm{ER + AT}$ combines with ours, the loss of task t is:", + "bbox": [ + 89, + 595, + 483, + 685 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} \\triangleq \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {t}\\right), y _ {t}\\right) + \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {\\mathcal {M}}\\right), y _ {\\mathcal {M}}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 698, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 731, + 223, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 756, + 297, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and training details. Following standard practices in adversarial training and continual learning [6, 37, 63], we evaluate our approach on Split-CIFAR10 and Split-CIFAR100 [26] and the model architecture is ResNet18, trained using SGD optimizer with learning rate 0.1. Split-CIFAR10 comprises ten classes (5,000 training, 1,000 test samples per class) divided into five binary classification tasks. Split-CIFAR100 contains 100 classes (500 training,", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "100 test samples per class) split into ten-way classification tasks. Evaluations on Split-Tiny-ImageNet [45] and ViT [16] in supplementary materials.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For adversarial training, we use perturbations of 8/255 and step size of 2/255. Data augmentation includes random cropping (4-pixel padding) and horizontal flipping for all data. Unlike Chen et al. [12], we achieve robustness without additional unlabeled data1. All experiments are conducted with three different random seeds.", + "bbox": [ + 511, + 136, + 905, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate both adversarial robustness and continual learning performance using Final Average Accuracy (FAA) and forgetting metrics on clean and adversarial samples. For adversarial evaluation, we employ PGD-20 (white-box attack) and Auto Attack (AA) [14], which combines black-box and white-box attacks for reliable robustness assessment [36, 37, 51]. Extended robustness verification using additional black-box (RayS [11]) and adaptive attacks in the supplementary material. Let $a_{i}^{t}$ denote the accuracy for task $i$ after training on task $t$ . The metrics are defined as:", + "bbox": [ + 511, + 227, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F A A} \\triangleq \\frac {1}{T} \\sum_ {i = 1} ^ {T} a _ {i} ^ {T}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 402, + 906, + 445 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {f o r g e t t i n g} \\triangleq \\frac {1}{T - 1} \\sum_ {j = 1} ^ {T - 1} f _ {j}, \\text {s . t .} f _ {j} = \\max _ {l \\in \\{1, \\dots , T - 1 \\}} a _ {i} ^ {l} - a _ {j} ^ {T}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 468, + 911, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Forgetting, ranging from $[-100, 100]$ , measures the average decrease in accuracy across tasks.", + "bbox": [ + 511, + 523, + 905, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For analysis, we further define:", + "bbox": [ + 532, + 555, + 741, + 569 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C R D} \\triangleq \\mathrm {F A A} _ {\\text {c l e a n}} - \\widetilde {\\mathrm {F A A}} _ {\\text {c l e a n}}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 579, + 906, + 598 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F R I} \\triangleq \\widetilde {\\mathrm {F o r g e t t i n g}} _ {\\mathrm {c l e a n}} - \\mathrm {F o r g e t t i n g} _ {\\mathrm {c l e a n}}, \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 621, + 906, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R R D} \\triangleq \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}} - \\mathrm {F A A} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}}\\right) - \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} - \\mathrm {F A A} _ {\\mathrm {a d v}}\\right), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 667, + 903, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where \"clean\" and \"adv\" denote metrics on clean and adversarial data, respectively, tilde $(\\sim)$ indicates robust models, and \"Joint\" refers to a joint training model.", + "bbox": [ + 511, + 703, + 903, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 757, + 653, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method against various adversarial training and memory-based continual learning combinations on Split-CIFAR10/100. Evaluations on Split-TinyImageNet [45] and ViT [16] in supplementary materials. Due to computational constraints, we focus on integrating our approach with ER+AT and ER+TRADES ( $\\beta = 6.0$ ), which showed superior adversarial FAA performance among baselines.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "567", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8f85153e81efdb0f268e1ad4f2e3abf85e34197e53ebf346dfec2050a61ba2e9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
200ER+AT28.1880.5817.8616.9469.5884.4910.2344.3044.6936.89
ER+TRADES22.4277.2515.7215.5364.9578.798.1051.3351.5021.97
ER+FAT33.6169.2115.1414.8149.0483.4010.3543.6943.9628.56
ER+LBGAT25.6884.4716.6516.5670.5078.1918.8540.6940.7340.83
ER+SCORE48.6556.792.400.9318.9688.909.827.256.728.70
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)18.16(↑ 1.22)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)47.61(↑ 2.92)34.04(↓ 2.85)
ER+TRADES+Ours43.34(↑ 20.92)33.40(↓ 43.85)19.85(↑ 4.13)18.35(↑ 2.82)30.78(↓ 34.17)82.59(↑ 3.80)7.53(↓ 0.57)59.41(↑ 8.08)59.59(↑ 8.09)14.16(↓ 7.81)
5,120ER+AT61.8837.7227.2826.6941.6691.242.5656.5956.9019.34
ER+TRADES20.3685.1416.3016.1872.8588.481.5964.3664.5212.47
ER+FAT54.5543.1819.6818.9142.1591.502.1256.7256.8714.79
ER+LBGAT62.4537.7327.4226.6647.8391.103.2556.5756.3119.38
ER+SCORE51.3752.843.141.1020.2995.631.5814.6613.922.53
ER+AT+Ours64.34(↑ 2.46)23.64(↓ 14.08)31.31(↑ 4.03)30.49(↑ 3.80)20.46(↓ 21.20)91.00(↓ 0.24)3.51(↑ 0.95)60.49(↑ 3.90)60.61(↑ 3.71)13.27(↓ 6.07)
ER+TRADES+Ours39.80(↑ 19.44)44.08(↓ 41.06)23.07(↑ 6.77)21.98(↑ 5.80)41.37(↓ 31.48)86.48(↓ 2.00)1.71(↑ 0.12)69.25(↑ 4.89)69.38(↑ 4.86)5.33(↓ 7.14)
", + "bbox": [ + 96, + 88, + 903, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2945f76b4b5c521621958229774591b5a00e3a0f0bfdfbd3bcc5113b501e9a3e.jpg", + "table_caption": [ + "(a) Results on Split-CIFAR10. We chose two buffer sizes of 200 and 5, 120." + ], + "table_footnote": [], + "table_body": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
500ER+AT11.9473.545.665.5438.0352.7128.4417.3519.5625.67
ER+TRADES7.5968.135.435.1144.1350.7520.2226.8927.6920.34
ER+FAT11.4873.995.355.2337.2656.1024.7120.0122.9922.31
ER+LBGAT11.7775.106.165.8239.7742.0226.4513.9914.4323.14
ER+SCORE16.2277.710.910.626.1368.8711.213.237.954.94
ER+AT+Ours24.14(↑ 12.20)53.03(↓ 20.51)7.13(↑ 1.47)6.68(↑ 1.14)25.81(↓ 12.22)56.00(↑ 3.29)26.19(↓ 2.25)17.95(↑ 0.60)20.26(↑ 0.70)24.87(↓ 0.80)
ER+TRADES+Ours23.24(↑ 15.65)24.29(↓ 43.84)9.93(↑ 4.50)7.50(↑ 2.39)11.70(↓ 32.43)56.68(↑ 5.93)21.39(↑ 1.17)27.39(↑ 0.50)28.50(↑ 0.81)16.76(↓ 3.58)
2,000ER+AT18.7765.067.207.0133.5662.0117.9721.1624.0420.47
ER+TRADES9.5070.495.355.0142.0860.6313.7826.6829.1918.60
ER+FAT17.0966.916.125.9333.6263.8815.9923.4826.7517.12
ER+LBGAT20.5863.317.937.0530.0955.7725.1117.9518.8819.18
ER+SCORE30.1061.510.750.514.1076.9019.604.8811.975.25
ER+AT+Ours31.93(↑ 13.16)40.50(↓ 24.56)9.61(↑ 2.41)9.16(↑ 2.15)17.78(↓ 15.78)63.77(↑ 1.76)17.16(↓ 0.81)23.11(↑ 1.95)25.59(↑ 1.55)17.57(↓ 2.90)
ER+TRADES+Ours28.73(↑ 19.23)24.16(↓ 46.33)12.75(↑ 7.40)11.02(↑ 6.01)14.56(↓ 27.52)62.01(↑ 1.38)13.16(↓ 0.62)34.81(↑ 8.13)35.36(↑ 6.17)11.60(↓ 7.00)
", + "bbox": [ + 96, + 297, + 906, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e4d3881bf6f5a1964230b2ca06aa53536eba6cca815a0e733889244dfed6f3f7.jpg", + "table_caption": [ + "(b) Results on Split-CIFAR100. We choose two buffer sizes of 500 and 2,000.", + "Table 2. Experiment results on Split-CIFAR10/100 datasets and model architecture is ResNet18. Here, PGD-20 and AA are adversarial data Final Average Accuracy (FAA) generated by PGD-20 and Auto Attack (AA), respectively. Forgetting of adversarial data is computed based on PGD-20. Bold represents the best experimental results for the same settings, underline indicates the second-best result, green signifies relative improvement and red indicates relative degradation. By applying our methods, various model performances can be improved across the board." + ], + "table_footnote": [], + "table_body": "
MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
ER+AT28.1880.5817.8669.5884.4910.2344.3036.89
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)34.04(↓ 2.85)
GSS+AT27.5980.7816.6768.5384.419.8344.2534.79
GSS+AT+Ours36.93(↑ 9.34)67.72(↓ 13.06)16.84(↑ 0.17)60.04(↓ 8.49)85.57(↑ 1.16)8.86(↓ 0.97)47.11(↑ 2.86)31.83(↓ 2.97)
ASER+AT18.8587.7814.0665.5773.8719.0130.6544.85
ASER+AT+Ours24.45(↑ 5.61)81.73(↓ 6.05)14.91(↑ 0.85)62.74(↓ 2.83)77.70(↑ 3.83)15.21(↓ 3.80)34.50(↑ 3.85)38.55(↓ 6.30)
X-DER+AT34.0425.1316.8227.8480.804.9660.8310.99
X-DER+AT+Ours43.25(↑ 9.21)20.77(↓ 4.36)17.22(↑ 0.41)18.56(↓ 9.28)84.87(↑ 4.07)1.74(↓ 3.21)61.68(↑ 0.85)7.03(↓ 3.97)
", + "bbox": [ + 96, + 577, + 903, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Experiments with other data selection-based and logit masking-based continual learning methods.", + "bbox": [ + 181, + 756, + 812, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on different adversarial training methods. Table 2 reveals several key findings. Buffer size significantly impacts performance, with larger buffers generally improving results, except for $\\mathrm{ER + TRADES}$ on Split-CIFAR10 at buffer size 5, 120. Regarding robustness, $\\mathrm{ER + AT}$ demon", + "bbox": [ + 89, + 786, + 480, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strates superior performance in class incremental settings, while ER+TRADES achieves optimal results under task incremental learning, albeit with reduced clean sample accuracy. ER+LBGAT exhibits balanced performance in class incremental settings, while ER+FAT and ER+SCORE excel on clean samples but underperform on adversarial ones.", + "bbox": [ + 511, + 786, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method significantly enhances both ER+TRADES", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "1The 80M-TinyImage dataset used in [12] has been withdrawn due to privacy concerns", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "568", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f235e515984687e31c55a1108d89b2b142c94142c235cd2b83d4d6084e779850.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AFLCRAERFPClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
22.4277.2515.7264.9578.798.1051.3321.97
✓37.8821.7618.5821.7781.6610.8653.6417.60
✓✓37.4510.2919.819.6178.1315.3955.7114.13
✓✓✓43.3433.4019.8530.7882.597.5359.4114.16
", + "bbox": [ + 133, + 88, + 862, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation experiments on the Split-CIFAR10 dataset with the buffer size of 200, using ER+TRADES as the baseline. Bold represents the best experimental results for the same settings. Experiments have demonstrated that AFLC mitigates clean-sample accelerated forgetting from adversarial samples, RAER mitigates gradient obfuscation (Adversarial FAA has a boost) and robust forgetting; adding FP improves the model's ability to learn new tasks (FAA has an overall increase).", + "bbox": [ + 89, + 198, + 906, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and ER+AT across Split-CIFAR10/100 datasets, particularly in class incremental settings. For Split-CIFAR10, we achieve FAA improvements of $\\Delta_{clean} = 20.92\\%$ and $\\Delta_{adv} = 8.09\\%$ , with forgetting reductions of $\\Delta_{f_{clean}} = 43.85\\%$ and $\\Delta_{f_{adv}} = 34.17\\%$ . On Split-CIFAR100, improvements reach $\\Delta_{clean} = 19.23\\%$ and $\\Delta_{adv} = 8.13\\%$ for FAA, with forgetting reductions of $\\Delta_{f_{clean}} = 46.33\\%$ and $\\Delta_{f_{adv}} = 32.43\\%$ .", + "bbox": [ + 89, + 262, + 480, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach demonstrates positive improvements in $93\\%$ of evaluation metrics (Table 2). Minor performance decreases in task incremental settings with buffer size 5, 120 reflect AFLC's diminishing returns and the known clean-robust accuracy trade-off [31, 35, 37, 57].", + "bbox": [ + 89, + 390, + 482, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on different CL methods. We evaluate our method against state-of-the-art continual learning approaches on Split-CIFAR10 with buffer size 200, including gradient-based GSS, Shapley value-based ASER, and logit masking-based X-DER. These methods primarily address clean-data continual learning, whereas our approach specifically targets adversarial robustness.", + "bbox": [ + 89, + 472, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 demonstrates that existing data selection methods (GSS, ASER) show limited effectiveness in adversarial settings, with robustness performance not exceeding $\\mathrm{ER + AT}$ . Our method, when integrated with these approaches, enhances both robustness and clean accuracy across class and task incremental settings.", + "bbox": [ + 89, + 585, + 482, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "X-DER employs logit masking, a specialized case of our AFLC, for memory updates and future preparation. While effective at forgetting mitigation, its aggressive gradient suppression limits model learning capacity, resulting in lower robust FAA than $\\mathrm{ER + AT}$ in class incremental settings. However, integration with our method yields significant improvements in both clean and adversarial metrics.", + "bbox": [ + 89, + 681, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The consistent performance gains across all evaluation metrics validate our method's effectiveness in simultaneously addressing adversarial robustness and catastrophic forgetting. These empirical results demonstrate the broad applicability of our approach across different continual learning frameworks. Accuracy curves during continual training can be found in the supplementary material.", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Experiments", + "text_level": 1, + "bbox": [ + 511, + 262, + 720, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 4, we have performed ablation experiments based on ER+TRADES under the Split-CIFAR10 dataset. The results demonstrate that AFLC (Sec. 4.2) can effectively mitigate the increased forgetting caused by adversarial training under class incremental setting (55.49% for clean samples and 43.18% for adversarial samples forgetting, with corresponding FAA improvements of 15.46% and 2.86%, respectively). AFLC does not show significant improvement in the task incremental setting due to excessive suppression of future task classification heads and the use of the same calibration value for classes within the same task.", + "bbox": [ + 511, + 284, + 903, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "RAER (Sec. 4.3) can further improve the robust accuracy of AFLC by $1.23\\%$ for class incremental setting and $2.07\\%$ for task incremental setting and reduce the robust forgetting by $12.16\\%$ and $3.47\\%$ respectively. That proves the data selected by RAER describe the overall data distribution more accurately and effectively mitigate the gradient obfuscation.", + "bbox": [ + 511, + 450, + 903, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When considering the future prior adjustment (FP in Table 4), we find that although the forgetting of the class incremental setting is higher, the FAA of both clean samples and adversarial samples has been significantly improved, and the forgetting of the task incremental setting has been further reduced, which proves that FP can reduce negative gradients to future classes and help learn new tasks. We also provide sensitivity analysis of hyperparameters ( $\\alpha$ in AFLC and $\\rho$ in RAER) in supplementary materials.", + "bbox": [ + 511, + 542, + 905, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 693, + 782, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper explores achieving adversarial robust continual learning without additional data. We analyze the challenges of directly combining continual learning and adversarial training - accelerated forgetting and gradient obfuscation. To address these, we propose: anti-forgettable logit calibration to reduce negative effects of past knowledge, and a robust difficulty-based data selection mechanism to mitigate gradient obfuscation. Experiments demonstrate the effectiveness of our approach. As AI systems increasingly need to learn continuously in dynamic environments, our work represents a further step toward more resilient and adaptable machine learning systems.", + "bbox": [ + 511, + 719, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "569", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Beijing Natural Science Foundation under No. L221013, Strategic Priority Research Program of the Chinese Academy of Sciences (No. XDB0680202), and the Innovation Funding of ICT, CAS under Grant No. E561160, E561090.", + "bbox": [ + 89, + 112, + 483, + 188 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 200, + 187, + 215 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ali Abbasi, Parsa Nooralinejad, Hamed Pirsiavash, and Soheil Kolouri. Brainwash: A poisoning attack to forget in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24057-24067, 2024. 2", + "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. 2", + "[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 6", + "[4] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283. PMLR, 2018. 2, 4", + "[5] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356, 2021. 2, 3", + "[6] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3, 5, 6", + "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3", + "[8] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. In International Conference on Learning Representations, 2022. 2", + "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 532-547. Springer, 2018. 2, 6", + "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019, 2019. 1, 2, 3", + "[11] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In Proceedings of the 26rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020. 6" + ], + "bbox": [ + 93, + 224, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Tianlong Chen, Sijia Liu, Shiyu Chang, Lisa Amini, and Zhangyang Wang. Queried unlabeled data improves and robustifies class-incremental learning. Transactions on Machine Learning and Data Mining, 2022. 1, 2, 6, 7", + "[13] Ting-Chun Chou, Jhih-Yuan Huang, and Wei-Po Lee. Continual learning with adversarial training to enhance robustness of image recognition models. In 2022 International Conference on Cyberworlds, pages 236-242. IEEE, 2022. 2", + "[14] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2, 6", + "[15] Jiequan Cui, Shu Liu, Liwei Wang, and Jiaya Jia. Learnable boundary guided adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15721-15730, 2021. 2", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 6", + "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2", + "[18] Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A Kai Qin, and Yuan He. Advdrop: Adversarial attack to dnns by dropping information. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7506-7515, 2021. 2", + "[19] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2", + "[20] Yunhui Guo, Mingrui Liu, Yandong Li, Liqiang Wang, Tianbao Yang, and Tajana Rosing. Attacking lifelong learning models with gradient reversion. 2020. 1, 2", + "[21] Ahmad Hassanpour, Majid Moradikia, Bian Yang, Ahmed Abdelhadi, Christoph Busch, and Julian Fierrez. Differential privacy preservation in robust continual learning. IEEE Access, 10:24273-24287, 2022. 2", + "[22] Siteng Kang, Zhan Shi, and Xinhua Zhang. Poisoning generative replay in continual learning to promote forgetting. In International Conference on Machine Learning, pages 15769-15785. PMLR, 2023. 2", + "[23] Hikmat Khan, Nidhal Carla Bouaynaya, and Ghulam Rasool. Adversarily robust continual learning. In 2022 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2022. 1, 2", + "[24] Hikmat Khan, Pir Masoom Shah, Syed Farhan Alam Zaidi, et al. Susceptibility of continual learning against adversarial attacks. arXiv preprint arXiv:2207.05225, 2022. 1, 2", + "[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "570", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2", + "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 3, 6", + "[27] Lilly Kumari, Shengjie Wang, Tianyi Zhou, and Jeff Bilmes. Retrospective adversarial replay for continual learning. In Advances in Neural Information Processing Systems, 2022. 2", + "[28] Sungyoon Lee, Hoki Kim, and Jaewook Lee. Graddiv: Adversarial robustness of randomized neural networks via gradient diversity regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2645-2651, 2023. 2", + "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2", + "[30] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 2", + "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 2, 3, 4, 8", + "[32] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 2", + "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, and Joost van de Weijer. Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2023. 2", + "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 2", + "[35] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In International Conference on Learning Representations, 2020. 8", + "[36] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 2, 3, 6", + "[37] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. International Conference on Machine Learning, 2022. 2, 6, 8", + "[38] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural Networks*, 113:54–71, 2019. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97(2):285, 1990. 2", + "[40] Mohammad Rostami, Leonidas Spinoulas, Mohamed Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14851-14860, 2021. 2", + "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2", + "[42] Fahad Sarfraz, Elahe Arani, and Bahram Zonooz. Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning. In The Eleventh International Conference on Learning Representations, 2023, 1", + "[43] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In European Conference on Computer Vision, pages 682-698. Springer, 2020. 2", + "[44] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9630-9638, 2021. 6", + "[45] Stanford. Tiny imagenet challenge (cs231n). http://tiny-imagenet.herokuapp.com/, 2015. Accessed: Feb 21, 2023. 6", + "[46] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2", + "[47] Muhammad Umer, Glenn Dawson, and Robi Polikar. Targeted forgetting and false memory formation in continual learners through adversarial backdoor attacks. In 2020 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2020. 2", + "[48] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2", + "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998. PMLR, 2022. 2", + "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2", + "[51] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, 2023. 6", + "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "571", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 2", + "[53] Tongtong Wu, Massimo Caccia, Zhuang Li, Yuan-Fang Li, Guilin Qi, and Gholamreza Haffari. Pretrained language model in continual learning: A comparative study. In International Conference on Learning Representations, 2021. 1, 2", + "[54] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8659-8668, 2021. 2", + "[55] Ju Xu, Jin Ma, Xuesong Gao, and Zhanxing Zhu. Adaptive progressive continual learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6715-6728, 2022. 2", + "[56] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 1, 2", + "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1, 2, 8", + "[58] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In International Conference on Learning Representations, 2021. 6", + "[59] Jie Zhang, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, Lei Zhang, and Chao Wu. Towards efficient data free black-box adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15115-15125, 2022. 2", + "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 2", + "[61] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2", + "[62] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2", + "[63] Kaijie Zhu, Jindong Wang, Xixu Hu, Xing Xie, and Ge Yang. Improving generalization of adversarial training via robust critical fine-tuning. arXiv preprint arXiv:2308.02533, 2023. 6" + ], + "bbox": [ + 91, + 92, + 482, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "572", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_model.json b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0faa062fbca767c00e958d3a35edf34e23bcff80 --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.221, + 0.131, + 0.778, + 0.154 + ], + "angle": 0, + "content": "Adversarial Robust Memory-Based Continual Learner" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.179, + 0.896, + 0.199 + ], + "angle": 0, + "content": "Xiaoyue \\(\\mathrm{Mi}^{1,2}\\), Fan Tang\\(^{1,2*}\\), Zonghan Yang\\(^{3,4}\\), Danding Wang\\(^{1,2}\\), Juan Cao\\(^{1,2}\\), Peng Li\\(^{3}\\), Yang Liu\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.199, + 0.763, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Computing Technology, Chinese Academy of Sciences \\(^{2}\\)University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.234, + 0.749, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Institute for AI Industry Research (AIR), Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.253, + 0.776, + 0.271 + ], + "angle": 0, + "content": "\\(^{4}\\)Department of Computer Science & Technology, Tsinghua University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.315, + 0.327, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.665 + ], + "angle": 0, + "content": "Despite the remarkable advances that have been made in continual learning, the adversarial vulnerability of such methods has not been fully discussed. We delve into the adversarial robustness of memory-based continual learning algorithms and observe limited improvement in robustness by directly applying adversarial training techniques. Our preliminary studies reveal the twin challenges for building adversarial robust continual learners: accelerated forgetting in continual learning and gradient obfuscation in adversarial robustness. In this study, we put forward a novel adversarial robust memory-based continual learner that adjusts data logits to mitigate the forgetting of pasts caused by adversarial samples. Furthermore, we devise a gradient-based data selection mechanism to alleviate the gradient obfuscation caused by limited stored data. The proposed approach can widely integrate with existing memory-based continual learning and adversarial training algorithms. Extensive experiments on Split-CIFAR10/100 and Split-TinyImageNet demonstrate the effectiveness of our approach, achieving a maximum forgetting reduction of \\(34.17\\%\\) in adversarial data for ResNet, and \\(20.10\\%\\) for ViT." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.222, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.859 + ], + "angle": 0, + "content": "Continual learning [48, 62] enables intelligent models to continually acquire new knowledge and adapt to changing environments while maintaining previous capabilities. It has recently emerged as a promising direction to achieve ideal intelligent models, and some pioneer works have investigated the robustness of continual learning in various aspects [42, 53, 56]. Achieving robust continual learners represents a crucial advancement in enabling intelligent models to adapt to dynamic scenarios effectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.86, + 0.483, + 0.875 + ], + "angle": 0, + "content": "However, when dealing with streaming non-i.i.d. data," + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.312, + 0.889, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.477, + 0.907, + 0.562 + ], + "angle": 0, + "content": "Figure 1. A preliminary analysis of the combination of continual learning and adversarial training on Split-CIFAR10 with buffer size 200. We consider the final average accuracy (FAA) and forgetting of adversarial and clean samples in both class incremental and task incremental settings. The red numbers indicate the performance after adding our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.685 + ], + "angle": 0, + "content": "continual learners are usually vulnerable to adversarial examples [23]. Chen et al. [12] first attempt at preserving adversarial robustness in continual learning [29] by leveraging adversarial training [52, 57] and a large amount of unlabeled data. Despite the impressive results of Chen et al. [12], the inherent challenges for conducting adversarial robust continual learners are still under disclosure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.909, + 0.884 + ], + "angle": 0, + "content": "In this study, we investigate building adversarial robust continual learners without abundant unlabeled data. We select two representative approaches for continual learning and adversarial defenses: memory-based continual learning and adversarial training. Figure 1 shows the direct combination of memory-based continual learning (ER [10]) with an adversarial training approach (TRADES [57]). The results for adversarial data reconfirm that continual learners risk adversarial attacks [12, 20, 23, 24], and demonstrate a critical trade-off: while adversarial training can strengthen continual learners against attacks [12, 20, 23, 24], it simultaneously increases forgetting on clean data, thus compromising the core objective of continual learning." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Further experimental analysis in Sec. 3 reveals two key" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "562" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "challenges when combining adversarial training with continual learning: (1) Adversarial data increase negative gradients toward past classes compared with clean data, accelerating catastrophic forgetting in continual learning; (2) Insufficient historical data in continual learning leads to gradient obfuscation [4], compromising the effectiveness of adversarial training compared to joint training. These challenges emerge naturally from the direct integration of the two approaches." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.484, + 0.469 + ], + "angle": 0, + "content": "Based on the analysis, we propose an adversarial robust memory-based continual learner without a large amount of additional data, which consists of an anti-forgettable logit calibration (AFLC) module and a robustness-aware experience replay (RAER) strategy. The AFLC module designs different logit calibration strategies for past, current, and future classes, thus mitigating the negative gradients of adversarial samples for continual learning. By leveraging attacking difficulty, RAER selects data that are adversarial safety and maintains the diversity of data distribution, rather than storing data overfitting to the decision boundary [9]. Our approach can be combined with most memory-based continual learning as well as adversarial training algorithms, and we have demonstrated the effectiveness of our approach on the Split-CIFAR10/100 and more challenging dataset Split-Tiny-ImageNet across ResNet and ViT architectures." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.47, + 0.435, + 0.483 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.485, + 0.483, + 0.545 + ], + "angle": 0, + "content": "- We reveal the twin critical challenges to carrying out adversarial robust memory-based continual learning algorithms: accelerated forgetting for continual learning and gradient obfuscation for adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.546, + 0.483, + 0.62 + ], + "angle": 0, + "content": "- We propose antiforgettable logit calibration to reduce negative influences from adversarial training through task-adaptive logit calibration, and robustness-aware experience replay by storing robust and distribution-representative past data to alleviate gradient obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.62, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- Experiments on Split-CIFAR10/100 and Split-TinyImageNet across ResNet and ViT demonstrate our effectiveness, in improving both clean accuracy and adversarial robustness of continual learning and mitigating forgetting in both class-incremental and task-incremental settings without additional data." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.485, + 0.483, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.241, + 0.74 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Continual learning. Continual learning [38, 48, 62] enables models to adapt to evolving data distributions while preserving past knowledge. To alleviate catastrophic forgetting, existing approaches [34, 39] can be categorized into three types: memory-based [7, 8, 10], regularization-based [25, 29, 30], and dynamic architecture [17, 41, 50, 55, 61]. Continual learning can be categorized into class incremental (class-IL) and task-incremental (task-IL) settings [33] based on the availability of task information during inference. In both settings, memory-based meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "ods have demonstrated superior performance without requiring model expansion [7, 60]. Researchers have explored continual learning robustness under various constraints [19, 53, 56], and saft risks about backdoor attacks [1, 22, 47] and privacy concerns [21]. Recent studies have revealed both the vulnerability of continual models to adversarial attacks and the potential of adversarial samples for mitigating forgetting [20, 23, 24, 27, 49]. Chen et al. [12] combined LwF [29] with adversarial training using additional data, while our work uniquely analyzes the core challenges in achieving adversarial robustness for continual learning without such requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Adversarial defense. The vulnerability of deep neural networks to adversarial examples [18, 46, 52, 59] has led to numerous defense methods [2, 5], with adversarial training emerging as the most effective approach [14, 15, 28, 31, 32, 36, 37, 57]. Recent research has extended to real-world scenarios, such as long-tailed distributions [54] and open-world settings [43]. Meanwhile some studies show the potential of continual algorithms in adapting to new attacks [13, 40]. Different from them, we specifically address adversarial robustness in class-incremental and task-incremental settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.866, + 0.483 + ], + "angle": 0, + "content": "3. Baselines and Their Inherent Problems" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.655, + 0.51 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Memory-based continual learning. Following [6, 7], we consider continual learning across \\(T\\) distinct classification tasks. We pre-allocate classification heads for all tasks at initialization rather than expanding them incrementally. For the \\(t\\)-th task (\\(1 \\leq t \\leq T\\)), we denote its data distribution as \\(D_{t}\\), with input samples \\(x_{t}\\) and corresponding labels \\(y_{t}\\). A model \\(\\mathrm{M}_{\\theta}(\\cdot)\\) with parameters \\(\\theta\\) is optimized sequentially in the given task order. Let \\(h_{\\theta}(x_{t}) \\in \\mathbb{R}^{C}\\) denote the output logits where \\(h_{\\theta}(x_{t}) = \\mathrm{M}_{\\theta}(x_{t})\\), and \\(C\\) is the total number of classes. The softmax output is defined as \\(f_{\\theta}(x_{t}) \\triangleq \\operatorname{softmax}(h_{\\theta}(x_{t}))\\). Given a logit vector \\(h_{\\theta}(x_{t}) = [h_{\\theta}(x_{t})_{1}, \\dots, h_{\\theta}(x_{t})_{C}]\\) where \\(h_{\\theta}(x_{t})_{i}\\) represents the logit for the \\(i\\)-th class, we partition the classes into three sets based on the current task \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.749, + 0.735, + 0.766 + ], + "angle": 0, + "content": "\\[\np = \\{1, 2, \\dots , (t - 1) b \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.769, + 0.905, + 0.785 + ], + "angle": 0, + "content": "\\[\nc = \\{(t - 1) b + 1, (t - 1) b + 2, \\dots , t b \\}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.788, + 0.76, + 0.803 + ], + "angle": 0, + "content": "\\[\nu = \\{t b + 1, t b + 2, \\dots , C \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.822, + 0.905, + 0.882 + ], + "angle": 0, + "content": "where \\( b \\) is the number of classes per task. Accordingly, we divide \\( h_{\\theta}(x_t) \\) into three parts: \\( h_{\\theta}(x_t)_p \\), \\( h_{\\theta}(x_t)_c \\), and \\( h_{\\theta}(x_t)_u \\), representing logits for past, current, and future classes respectively." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Continual learner aims to correctly classify data of all" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "563" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.406 + ], + "angle": 0, + "content": "
Buffer SizeMethodW/ ATClass Incremental SettingTask Incremental SettingCRD↓FRI↓RRD↓
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
-Joint (upper bound)X91.81±0.07-0.00±0.00-98.17±0.12-0.00±0.00-7.92--
✓79.33±0.36-50.93±1.28-94.80±0.30-74.63±1.13-
-SGD (lower bound)X19.66±0.0796.39±0.675.17±2.5833.31±5.1966.84±6.5037.42±7.394.44±3.7833.34±3.490.91-3.7341.10
✓19.34±0.0591.93±0.2015.65±0.1869.71±1.7765.33±5.9134.43±7.1627.71±1.8854.56±3.47
200ERX48.80±1.8460.00±2.250.27±0.1812.60±2.1392.89±0.265.00±0.260.01±0.0113.78±0.4114.5112.9131.70
✓28.18±0.6980.58±1.0517.86±0.2969.58±1.1584.49±0.6110.23±0.9844.30±1.0536.89±1.11
5120X83.34±1.3815.35±1.710.08±0.1113.65±2.3496.84±0.170.71±0.190.06±0.0421.83±8.6913.5312.1118.80
✓61.88±0.7437.72±0.7327.28±0.5641.66±1.2291.24±0.192.56±0.1256.59±0.8819.34±1.11
200DERX62.71±2.7638.75±4.110.01±0.007.42±2.1391.32±0.836.95±1.440.09±0.057.39±2.1817.7215.7334.80
✓35.79±2.8565.67±4.1616.44±0.2152.84±3.7182.79±0.2111.48±0.6639.52±0.3740.64±1.38
5120X83.99±0.9910.57±1.510.07±0.077.05±2.0395.67±0.131.63±0.130.05±0.057.39±0.4816.7415.6624.32
✓56.58±3.8740.40±4.0723.26±0.4445.06±1.4889.62±0.673.12±0.2353.66±2.0122.95±3.68
200DER++X65.55±2.8132.17±4.260.24±0.327.38±3.3792.11±0.558.85±2.980.22±0.3310.45±0.9414.478.2437.14
✓45.43±3.2846.67±5.0914.46±1.3535.09±2.7083.29±1.0710.83±1.7736.82±0.5934.09±2.23
5120X85.74±0.657.32±0.881.51±1.276.10±0.5396.13±0.531.14±0.350.82±0.526.06±0.4910.638.2122.18
✓68.67±1.1123.68±2.6626.95±0.2927.68±2.4891.95±0.331.20±0.9054.26±0.4116.87±1.72
200X-DERX60.96±1.0912.29±0.090.00±0.003.84±1.1894.33±0.632.49±0.170.13±0.032.59±0.6420.237.6524.02
✓34.04±0.9025.13±4.0316.82±0.9827.84±4.1680.8±0.664.96±1.6660.83±0.5610.99±0.81
5120X69.93±0.634.33±0.260.21±0.124.52±1.3996.94±0.020.29±0.040.13±0.022.04±0.4324.9618.5221.27
✓34.29±0.6538.92±0.0319.24±0.1038.79±0.4182.67±1.682.74±0.8364.11±0.128.14±0.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.41, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Table 1. We conduct a systematic experimental analysis combining adversarial training (specifically Vanilla AT [31]) with memory-based continual learning algorithms. See Sec. 3.2 for a detailed summary of observations. Performance is evaluated using several metrics (detailed in Sec. 5.1): Final Average Accuracy (FAA), Forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase on clean data (FRI), and Robust Relative Decrease (RRD) compared to joint training. All adversarial samples are generated using PGD-20. Note that the reported CRD, FRI, and RRD values are averaged across both task-incremental and class-incremental settings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.485, + 0.449, + 0.5 + ], + "angle": 0, + "content": "observed tasks, and its objective can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.507, + 0.482, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {t}, \\text {w h e r e} \\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {\\left(x _ {t}, y _ {t}\\right) \\sim D _ {t}} \\left[ \\ell \\left(f _ {\\theta} \\left(x _ {t}\\right), y _ {t}\\right) \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.697 + ], + "angle": 0, + "content": "where \\(\\ell\\) is the task-specific loss. The unavailability of past task data poses great challenges for continual learning. However, direct sequential optimization can lead to severe catastrophic forgetting. To address the above challenge, memory-based continual learners usually store a few past task data and replay them during the current task training phase [7], and become leading-edge approaches. When training the \\(t\\)-th task, its loss function can be formalized as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.705, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} [ \\ell (f _ {\\theta} (x), y) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.759 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is the episodic memory of the current task, which preserves a small subset of observed past data \\(\\langle x_{\\mathcal{M}},y_{\\mathcal{M}}\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.82 + ], + "angle": 0, + "content": "Adversarial training. Adversarial training is currently the most reliable way [5] to improve the adversarial robustness of a model. Following [31], it can be formally defined as a min-max optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.828, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.866, + 0.484, + 0.905 + ], + "angle": 0, + "content": "where \\(\\delta\\) is an adversarial perturbation, \\(D\\) is data distribution, and \\(S\\) is a perturbation set. \\(S = \\left\\{\\delta \\mid \\| \\delta \\|_p \\leq \\epsilon \\right\\}\\), \\(\\epsilon\\) is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.905, + 0.515 + ], + "angle": 0, + "content": "the perturbation size. The goal of adversarial training is to correctly classify all adversarial examples \\((\\widetilde{x} = x + \\delta)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Here we integrate memory-based continual learner and adversarial training as adversarial robust memory-based continual learner baselines, and the new optimization objective during the training stage of \\( t \\)-th task is:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.583, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right). \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.681, + 0.637 + ], + "angle": 0, + "content": "3.2. Vanilla Solutions" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.854 + ], + "angle": 0, + "content": "To investigate the challenges in adversarial robust continual learning, we analyze several popular memory-based continual learning methods (ER [10], DER [7], DER++ [7], and X-DER [6]) on Split-CIFAR10 [26] under both class and task incremental settings. We employ Vanilla AT [31] for adversarial training due to its classic and strong performance [36]. Performance is evaluated against an upper bound (\"Joint\") and a lower bound (\"SGD\") using multiple metrics: Final Average Accuracy (FAA), forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase in clean data (FRI), and Robust Relative Decrease (RRD). Detailed metric definitions are provided in Sec. 5.1, and additional implementation details and hyper-parameters are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Results in Table 1 reveal that direct integration of adversarial training with continual learning is suboptimal. First, the results validate our observations from Sec. 1: while" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "564" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "adversarial training enhances robustness, it compromises clean performance (positive CRDs) and accelerates forgetting (positive FRIs). Larger memory sizes consistently improve performance across all metrics. Moreover, continual models show inferior robustness improvement compared to joint models (positive RRDs)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.273 + ], + "angle": 0, + "content": "To investigate the underlying mechanisms, we select ER as our base model for further analysis. While DER/DER++ and X-DER demonstrate better stability with lower forgetting, they struggle with new task learning, resulting in poor robust FAA. In contrast, ER combined with adversarial training exhibits strong robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.28, + 0.306, + 0.294 + ], + "angle": 0, + "content": "3.3. The Inherent Problems" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.483, + 0.361 + ], + "angle": 0, + "content": "By analyzing gradients during training, we identify two critical challenges in robust continual learning: accelerated forgetting in continual learning and gradient obfuscation in adversarial training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.483, + 0.513 + ], + "angle": 0, + "content": "Accelerated forgetting in continual learning. Adversarial training hastens the forgetting of clean data in continual learners. According to [31], adversarial examples are designed to mislead model predictions by increasing probabilities for incorrect classes while suppressing those of the true class. In the continual learning setting, adversarial examples from the current task tend to increase probabilities for past classes while decreasing probabilities for their true class. Specifically, for an adversarial example \\(\\widetilde{x}_t\\) generated from a clean sample \\(x_t\\) of current task \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.52, + 0.482, + 0.537 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {p} > f _ {\\theta} \\left(x _ {t}\\right) _ {p}, f _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {\\text {t r u e c l a s s}} < f _ {\\theta} \\left(x _ {t}\\right) _ {\\text {t r u e c l a s s}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.586 + ], + "angle": 0, + "content": "When using cross-entropy as the loss function, adversarial data amplify negative gradients for current data \\( x_{t} \\) for past tasks:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.589, + 0.482, + 0.626 + ], + "angle": 0, + "content": "\\[\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (\\widetilde {x} _ {t})} \\right] _ {p} = f _ {\\theta} (\\widetilde {x} _ {t}) _ {p} > f _ {\\theta} (x _ {t}) _ {p} = \\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (x _ {t})} \\right] _ {p}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Our empirical analysis reveals that these adversarial examples generate stronger negative gradients toward past tasks, especially during the initial training phases. In Figure 2, experiments with \\(\\mathrm{ER + AT}\\) on Split-CIFAR10 using a buffer size of 200 show adversarial data consistently produces larger gradient norms compared to clean data. The larger gradient norms indicate stronger positive gradients for the true class and stronger negative gradients for other classes, including past classes, thus accelerating forgetting of previously learned tasks. While memory-based methods using stored adversarial samples can help mitigate catastrophic forgetting, the limited buffer size (typically small in practical applications) means that adversarial training on the entire dataset still accelerates forgetting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Gradient obfuscation in adversarial training. Gradient obfuscation [4] occurs when gradient information is obscured by non-differentiable operators or randomness, leading to inferior robustness. We observe a specific form of" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.221, + 0.905, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.354, + 0.905, + 0.396 + ], + "angle": 0, + "content": "Figure 2. Gradient norms of adversarial data and clean data during the training stage. We experiment ER+AT with a 200 buffer size on the Split-CIFAR10 dataset." + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.398, + 0.858, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.593 + ], + "angle": 0, + "content": "Figure 3. The gradient cosine similarities between robust joint model and robust continual model, which indicates a special gradient obfuscation only in continual learning. We experiment with buffer size=200/5,120 on the Split-CIFAR10 dataset. Gradient obfuscation weakens as the buffer size gets larger, and our method also mitigates gradient obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.644 + ], + "angle": 0, + "content": "gradient obfuscation, termed \"shattered gradients\", in adversarial training under continual learning, resulting in less effective robustness compared to joint models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.75 + ], + "angle": 0, + "content": "In continual learning, gradient information is biased due to limited replay data. With few old task samples, the model tends to overfit, causing direction-biased gradients when generating adversarial examples. Our analysis of gradient directions between adversarial robust joint and continual models reveals significant differences, contributing to lower adversarial robustness in continual learners (see Figure 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.605, + 0.779 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.625, + 0.803 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address the identified challenges, we propose a novel framework consisting of two key components: antiforgettable logit calibration (AFLC) to mitigate forgetting acceleration, and robustness-aware experience replay (RAER) to reduce gradient obfuscation. The full pipeline is illustrated in Figure 4: during training each task, the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.672, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.09, + 0.898, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 4. The pipeline of our adversarial robust memory-based continual learner. First, generate adversarial samples \\((\\widetilde{X}_t,\\widetilde{X}_{\\mathcal{M}})\\) and robust difficulty factors \\((K_{t})\\) for both current task data \\((X_{t},Y_{t})\\) and memory data \\((X_{\\mathcal{M}},Y_{\\mathcal{M}})\\). The model then processes these samples through AFLC (Sec. 4.2), which calibrates logits based on task order to reduce negative gradients toward past tasks. Meanwhile, RAER (Sec. 4.3) leverages the robust difficulty factors \\(K_{t}\\) to select representative samples for memory storage, mitigating gradient obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.473 + ], + "angle": 0, + "content": "model processes both current task data and memory samples through the model. AFLC then calibrates logits based on task order (past/current/future) before loss computation. Simultaneously, RAER evaluates current task samples using a robustness difficulty factor \\( k \\) to select representative data for memory storage." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.481, + 0.395, + 0.497 + ], + "angle": 0, + "content": "4.2. Anti-Forgettable Logit Calibration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.562 + ], + "angle": 0, + "content": "We introduce AFLC to regulate the influence of adversarial examples on past task knowledge. The core idea is to calibrate logits differently for past and current tasks to control gradient flow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Given an input \\( x \\) (either from current task \\( x_{t} \\) or memory buffer \\( x_{\\mathcal{M}} \\)), AFLC applies a task-dependent calibration:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.602, + 0.482, + 0.62 + ], + "angle": 0, + "content": "\\[\nh _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = h _ {\\theta} (x) _ {i} - v _ {i}, v _ {p} > v _ {c}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.628, + 0.436, + 0.643 + ], + "angle": 0, + "content": "where \\( v_{i} \\) represents the calibration factor for class i:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.65, + 0.483, + 0.684 + ], + "angle": 0, + "content": "\\[\nv _ {i} = - \\log \\left(\\frac {n _ {i}}{\\sum_ {j = 1} ^ {C} n _ {j}}\\right) - \\alpha_ {i}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.484, + 0.812 + ], + "angle": 0, + "content": "where \\( n_i \\) denotes sample count for class \\( i \\), and \\( \\alpha_i \\) is a hyperparameter. When all tasks share the classification head, we realize that making logit changes to past and current classes only is equivalent to negative gradient increases. So we make a Further Prior (FP) adjustment that essentially averages the adjustments applied to current and past classes and applies this value to future classes, helping to balance the gradients across all periods." + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.821, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nv _ {i} = \\frac {1}{t \\cdot b} \\sum_ {j = 1} ^ {t \\cdot b} \\left(v _ {j}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( t \\) is the current task order and \\( b \\) is the number of classes per task." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.382, + 0.892, + 0.398 + ], + "angle": 0, + "content": "The calibrated probability distribution is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.408, + 0.905, + 0.446 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = \\frac {\\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i}\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {j}\\right)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.895, + 0.472 + ], + "angle": 0, + "content": "and leads to two key properties: For current task samples," + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.481, + 0.905, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x})} \\right] _ {p} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {t}) _ {p} < f _ {\\theta} (\\widetilde {x} _ {t}) _ {p}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.528, + 0.727, + 0.543 + ], + "angle": 0, + "content": "and for memory buffer samples," + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.553, + 0.905, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x _ {\\mathcal {M}}})} \\right] _ {c} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {\\mathcal {M}}) _ {c} > f _ {\\theta} (\\widetilde {x} _ {\\mathcal {M}}) _ {c}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.658 + ], + "angle": 0, + "content": "In that way, AFLC can mitigate forgetting acceleration by reducing the negative gradient to past classes from current adversarial data and improving the negative gradient to current classes from memory adversarial data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Comparisons with other logit calibration CL. X-DER [6] also adjusts the negative gradients of the new task data affecting the past tasks by updating the new task with the logit masking. Logit masking can be seen as a special case of our method at the softmax layer, which can be viewed as \\(\\mathrm{v}_p = \\mathrm{v}_c = 0\\) for past data, and \\(\\mathrm{v}_p = +\\infty\\), \\(\\mathrm{v}_c = 0\\) for current data. However, their extreme margin \\(\\mathbf{V}\\) inhibits the model's ability to learn new knowledge. X-DER further adds a safeguard threshold to limit the activations of past and future heads, which also can be combined with us." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.84, + 0.835 + ], + "angle": 0, + "content": "4.3. Robustness-Aware Experience Replay" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To address gradient obfuscation, we propose RAER, which selects samples based on their robustness characteristics. For each sample \\( x \\), we measure its robustness difficulty factor \\( k \\) during PGD-10 attack, without additional forward or" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "566" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.152 + ], + "angle": 0, + "content": "backward calculation, where \\( k \\in [0,10] \\cap \\mathbb{N} \\) indicates the number of successful attacks. Samples meeting the threshold condition \\( k < \\rho \\) are candidates for memory storage. The memory update process can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.166, + 0.484, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} \\leftarrow \\text {R a n d o m S a m p l e} \\left(\\{x \\mid k _ {x} < \\rho \\}, m\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.465, + 0.211 + ], + "angle": 0, + "content": "where \\(m\\) is the memory size and \\(\\rho\\) is set to 5 in practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.348 + ], + "angle": 0, + "content": "Why RAER works. The robustness difficulty factor \\( k \\) shows how easily a sample can be attacked: a high \\( k \\) suggests vulnerability and proximity to decision boundaries, while a low \\( k \\) implies robustness and distance from boundaries. We selectively store samples with \\( k < \\rho \\) in memory, excluding boundary-overfitting outliers. Unlike prior continual learning [9] and adversarial training [58] methods, RAER emphasizes distribution-representative and robust learnable samples while maintaining model reliability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.407 + ], + "angle": 0, + "content": "The computational overhead of RAER is minimal, as \\( k \\) is calculated during the existing adversarial example generation process, without additional forward or backward calculation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.56 + ], + "angle": 0, + "content": "Comparisons with other data-selection CL. Existing works [3, 44] alleviate catastrophic forgetting using data selection and focus solely on clean samples. RAER uniquely targets gradient obfuscation in continual robust settings by: Selecting adversarially safe samples for better retention; excluding high-\\(k\\) boundary samples; and maintaining representative data distribution. As demonstrated in Sec. 5, RAER outperforms traditional data selection strategies combined with adversarial training, particularly in adversarial robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.572, + 0.265, + 0.589 + ], + "angle": 0, + "content": "4.4. Overall Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.686 + ], + "angle": 0, + "content": "To sum up, we design AFLC to reduce negative gradients of the current adversarial sample to past tasks, and RAER to select data that alleviate gradient obfuscation to store. We set \\(\\alpha\\) of AFLC is 3.5 in practice and \\(x_{\\mathcal{M}}\\) and \\(y_{\\mathcal{M}}\\) are selected by RAER with \\(\\rho = 5\\). When \\(\\mathrm{ER + AT}\\) combines with ours, the loss of task t is:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.699, + 0.483, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} \\triangleq \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {t}\\right), y _ {t}\\right) + \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {\\mathcal {M}}\\right), y _ {\\mathcal {M}}\\right). \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.224, + 0.749 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.299, + 0.774 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets and training details. Following standard practices in adversarial training and continual learning [6, 37, 63], we evaluate our approach on Split-CIFAR10 and Split-CIFAR100 [26] and the model architecture is ResNet18, trained using SGD optimizer with learning rate 0.1. Split-CIFAR10 comprises ten classes (5,000 training, 1,000 test samples per class) divided into five binary classification tasks. Split-CIFAR100 contains 100 classes (500 training," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "100 test samples per class) split into ten-way classification tasks. Evaluations on Split-Tiny-ImageNet [45] and ViT [16] in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.227 + ], + "angle": 0, + "content": "For adversarial training, we use perturbations of 8/255 and step size of 2/255. Data augmentation includes random cropping (4-pixel padding) and horizontal flipping for all data. Unlike Chen et al. [12], we achieve robustness without additional unlabeled data1. All experiments are conducted with three different random seeds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate both adversarial robustness and continual learning performance using Final Average Accuracy (FAA) and forgetting metrics on clean and adversarial samples. For adversarial evaluation, we employ PGD-20 (white-box attack) and Auto Attack (AA) [14], which combines black-box and white-box attacks for reliable robustness assessment [36, 37, 51]. Extended robustness verification using additional black-box (RayS [11]) and adaptive attacks in the supplementary material. Let \\( a_{i}^{t} \\) denote the accuracy for task \\( i \\) after training on task \\( t \\). The metrics are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.404, + 0.907, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F A A} \\triangleq \\frac {1}{T} \\sum_ {i = 1} ^ {T} a _ {i} ^ {T}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.469, + 0.912, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\text {f o r g e t t i n g} \\triangleq \\frac {1}{T - 1} \\sum_ {j = 1} ^ {T - 1} f _ {j}, \\text {s . t .} f _ {j} = \\max _ {l \\in \\{1, \\dots , T - 1 \\}} a _ {i} ^ {l} - a _ {j} ^ {T}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.525, + 0.906, + 0.555 + ], + "angle": 0, + "content": "Forgetting, ranging from \\([-100, 100]\\), measures the average decrease in accuracy across tasks." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.556, + 0.742, + 0.57 + ], + "angle": 0, + "content": "For analysis, we further define:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.58, + 0.907, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C R D} \\triangleq \\mathrm {F A A} _ {\\text {c l e a n}} - \\widetilde {\\mathrm {F A A}} _ {\\text {c l e a n}}, \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.622, + 0.907, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F R I} \\triangleq \\widetilde {\\mathrm {F o r g e t t i n g}} _ {\\mathrm {c l e a n}} - \\mathrm {F o r g e t t i n g} _ {\\mathrm {c l e a n}}, \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.669, + 0.905, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R R D} \\triangleq \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}} - \\mathrm {F A A} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}}\\right) - \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} - \\mathrm {F A A} _ {\\mathrm {a d v}}\\right), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.75 + ], + "angle": 0, + "content": "where \"clean\" and \"adv\" denote metrics on clean and adversarial data, respectively, tilde \\((\\sim)\\) indicates robust models, and \"Joint\" refers to a joint training model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.654, + 0.773 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method against various adversarial training and memory-based continual learning combinations on Split-CIFAR10/100. Evaluations on Split-TinyImageNet [45] and ViT [16] in supplementary materials. Due to computational constraints, we focus on integrating our approach with ER+AT and ER+TRADES (\\(\\beta = 6.0\\)), which showed superior adversarial FAA performance among baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "567" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.282 + ], + "angle": 0, + "content": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
200ER+AT28.1880.5817.8616.9469.5884.4910.2344.3044.6936.89
ER+TRADES22.4277.2515.7215.5364.9578.798.1051.3351.5021.97
ER+FAT33.6169.2115.1414.8149.0483.4010.3543.6943.9628.56
ER+LBGAT25.6884.4716.6516.5670.5078.1918.8540.6940.7340.83
ER+SCORE48.6556.792.400.9318.9688.909.827.256.728.70
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)18.16(↑ 1.22)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)47.61(↑ 2.92)34.04(↓ 2.85)
ER+TRADES+Ours43.34(↑ 20.92)33.40(↓ 43.85)19.85(↑ 4.13)18.35(↑ 2.82)30.78(↓ 34.17)82.59(↑ 3.80)7.53(↓ 0.57)59.41(↑ 8.08)59.59(↑ 8.09)14.16(↓ 7.81)
5,120ER+AT61.8837.7227.2826.6941.6691.242.5656.5956.9019.34
ER+TRADES20.3685.1416.3016.1872.8588.481.5964.3664.5212.47
ER+FAT54.5543.1819.6818.9142.1591.502.1256.7256.8714.79
ER+LBGAT62.4537.7327.4226.6647.8391.103.2556.5756.3119.38
ER+SCORE51.3752.843.141.1020.2995.631.5814.6613.922.53
ER+AT+Ours64.34(↑ 2.46)23.64(↓ 14.08)31.31(↑ 4.03)30.49(↑ 3.80)20.46(↓ 21.20)91.00(↓ 0.24)3.51(↑ 0.95)60.49(↑ 3.90)60.61(↑ 3.71)13.27(↓ 6.07)
ER+TRADES+Ours39.80(↑ 19.44)44.08(↓ 41.06)23.07(↑ 6.77)21.98(↑ 5.80)41.37(↓ 31.48)86.48(↓ 2.00)1.71(↑ 0.12)69.25(↑ 4.89)69.38(↑ 4.86)5.33(↓ 7.14)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.298, + 0.284, + 0.7, + 0.296 + ], + "angle": 0, + "content": "(a) Results on Split-CIFAR10. We chose two buffer sizes of 200 and 5, 120." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.298, + 0.907, + 0.495 + ], + "angle": 0, + "content": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
500ER+AT11.9473.545.665.5438.0352.7128.4417.3519.5625.67
ER+TRADES7.5968.135.435.1144.1350.7520.2226.8927.6920.34
ER+FAT11.4873.995.355.2337.2656.1024.7120.0122.9922.31
ER+LBGAT11.7775.106.165.8239.7742.0226.4513.9914.4323.14
ER+SCORE16.2277.710.910.626.1368.8711.213.237.954.94
ER+AT+Ours24.14(↑ 12.20)53.03(↓ 20.51)7.13(↑ 1.47)6.68(↑ 1.14)25.81(↓ 12.22)56.00(↑ 3.29)26.19(↓ 2.25)17.95(↑ 0.60)20.26(↑ 0.70)24.87(↓ 0.80)
ER+TRADES+Ours23.24(↑ 15.65)24.29(↓ 43.84)9.93(↑ 4.50)7.50(↑ 2.39)11.70(↓ 32.43)56.68(↑ 5.93)21.39(↑ 1.17)27.39(↑ 0.50)28.50(↑ 0.81)16.76(↓ 3.58)
2,000ER+AT18.7765.067.207.0133.5662.0117.9721.1624.0420.47
ER+TRADES9.5070.495.355.0142.0860.6313.7826.6829.1918.60
ER+FAT17.0966.916.125.9333.6263.8815.9923.4826.7517.12
ER+LBGAT20.5863.317.937.0530.0955.7725.1117.9518.8819.18
ER+SCORE30.1061.510.750.514.1076.9019.604.8811.975.25
ER+AT+Ours31.93(↑ 13.16)40.50(↓ 24.56)9.61(↑ 2.41)9.16(↑ 2.15)17.78(↓ 15.78)63.77(↑ 1.76)17.16(↓ 0.81)23.11(↑ 1.95)25.59(↑ 1.55)17.57(↓ 2.90)
ER+TRADES+Ours28.73(↑ 19.23)24.16(↓ 46.33)12.75(↑ 7.40)11.02(↑ 6.01)14.56(↓ 27.52)62.01(↑ 1.38)13.16(↓ 0.62)34.81(↑ 8.13)35.36(↑ 6.17)11.60(↓ 7.00)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.289, + 0.496, + 0.707, + 0.508 + ], + "angle": 0, + "content": "(b) Results on Split-CIFAR100. We choose two buffer sizes of 500 and 2,000." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.511, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Table 2. Experiment results on Split-CIFAR10/100 datasets and model architecture is ResNet18. Here, PGD-20 and AA are adversarial data Final Average Accuracy (FAA) generated by PGD-20 and Auto Attack (AA), respectively. Forgetting of adversarial data is computed based on PGD-20. Bold represents the best experimental results for the same settings, underline indicates the second-best result, green signifies relative improvement and red indicates relative degradation. By applying our methods, various model performances can be improved across the board." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.578, + 0.905, + 0.757 + ], + "angle": 0, + "content": "
MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
ER+AT28.1880.5817.8669.5884.4910.2344.3036.89
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)34.04(↓ 2.85)
GSS+AT27.5980.7816.6768.5384.419.8344.2534.79
GSS+AT+Ours36.93(↑ 9.34)67.72(↓ 13.06)16.84(↑ 0.17)60.04(↓ 8.49)85.57(↑ 1.16)8.86(↓ 0.97)47.11(↑ 2.86)31.83(↓ 2.97)
ASER+AT18.8587.7814.0665.5773.8719.0130.6544.85
ASER+AT+Ours24.45(↑ 5.61)81.73(↓ 6.05)14.91(↑ 0.85)62.74(↓ 2.83)77.70(↑ 3.83)15.21(↓ 3.80)34.50(↑ 3.85)38.55(↓ 6.30)
X-DER+AT34.0425.1316.8227.8480.804.9660.8310.99
X-DER+AT+Ours43.25(↑ 9.21)20.77(↓ 4.36)17.22(↑ 0.41)18.56(↓ 9.28)84.87(↑ 4.07)1.74(↓ 3.21)61.68(↑ 0.85)7.03(↓ 3.97)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.182, + 0.757, + 0.813, + 0.771 + ], + "angle": 0, + "content": "Table 3. Experiments with other data selection-based and logit masking-based continual learning methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.482, + 0.863 + ], + "angle": 0, + "content": "Results on different adversarial training methods. Table 2 reveals several key findings. Buffer size significantly impacts performance, with larger buffers generally improving results, except for \\(\\mathrm{ER + TRADES}\\) on Split-CIFAR10 at buffer size 5, 120. Regarding robustness, \\(\\mathrm{ER + AT}\\) demon" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.907, + 0.879 + ], + "angle": 0, + "content": "strates superior performance in class incremental settings, while ER+TRADES achieves optimal results under task incremental learning, albeit with reduced clean sample accuracy. ER+LBGAT exhibits balanced performance in class incremental settings, while ER+FAT and ER+SCORE excel on clean samples but underperform on adversarial ones." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our method significantly enhances both ER+TRADES" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1The 80M-TinyImage dataset used in [12] has been withdrawn due to privacy concerns" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "568" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.134, + 0.089, + 0.863, + 0.198 + ], + "angle": 0, + "content": "
AFLCRAERFPClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
22.4277.2515.7264.9578.798.1051.3321.97
✓37.8821.7618.5821.7781.6610.8653.6417.60
✓✓37.4510.2919.819.6178.1315.3955.7114.13
✓✓✓43.3433.4019.8530.7882.597.5359.4114.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.199, + 0.907, + 0.255 + ], + "angle": 0, + "content": "Table 4. Ablation experiments on the Split-CIFAR10 dataset with the buffer size of 200, using ER+TRADES as the baseline. Bold represents the best experimental results for the same settings. Experiments have demonstrated that AFLC mitigates clean-sample accelerated forgetting from adversarial samples, RAER mitigates gradient obfuscation (Adversarial FAA has a boost) and robust forgetting; adding FP improves the model's ability to learn new tasks (FAA has an overall increase)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.482, + 0.386 + ], + "angle": 0, + "content": "and ER+AT across Split-CIFAR10/100 datasets, particularly in class incremental settings. For Split-CIFAR10, we achieve FAA improvements of \\(\\Delta_{clean} = 20.92\\%\\) and \\(\\Delta_{adv} = 8.09\\%\\), with forgetting reductions of \\(\\Delta_{f_{clean}} = 43.85\\%\\) and \\(\\Delta_{f_{adv}} = 34.17\\%\\). On Split-CIFAR100, improvements reach \\(\\Delta_{clean} = 19.23\\%\\) and \\(\\Delta_{adv} = 8.13\\%\\) for FAA, with forgetting reductions of \\(\\Delta_{f_{clean}} = 46.33\\%\\) and \\(\\Delta_{f_{adv}} = 32.43\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.468 + ], + "angle": 0, + "content": "Our approach demonstrates positive improvements in \\(93\\%\\) of evaluation metrics (Table 2). Minor performance decreases in task incremental settings with buffer size 5, 120 reflect AFLC's diminishing returns and the known clean-robust accuracy trade-off [31, 35, 37, 57]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Results on different CL methods. We evaluate our method against state-of-the-art continual learning approaches on Split-CIFAR10 with buffer size 200, including gradient-based GSS, Shapley value-based ASER, and logit masking-based X-DER. These methods primarily address clean-data continual learning, whereas our approach specifically targets adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 3 demonstrates that existing data selection methods (GSS, ASER) show limited effectiveness in adversarial settings, with robustness performance not exceeding \\(\\mathrm{ER + AT}\\). Our method, when integrated with these approaches, enhances both robustness and clean accuracy across class and task incremental settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.79 + ], + "angle": 0, + "content": "X-DER employs logit masking, a specialized case of our AFLC, for memory updates and future preparation. While effective at forgetting mitigation, its aggressive gradient suppression limits model learning capacity, resulting in lower robust FAA than \\(\\mathrm{ER + AT}\\) in class incremental settings. However, integration with our method yields significant improvements in both clean and adversarial metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The consistent performance gains across all evaluation metrics validate our method's effectiveness in simultaneously addressing adversarial robustness and catastrophic forgetting. These empirical results demonstrate the broad applicability of our approach across different continual learning frameworks. Accuracy curves during continual training can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.263, + 0.722, + 0.279 + ], + "angle": 0, + "content": "5.3. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.905, + 0.451 + ], + "angle": 0, + "content": "In Table 4, we have performed ablation experiments based on ER+TRADES under the Split-CIFAR10 dataset. The results demonstrate that AFLC (Sec. 4.2) can effectively mitigate the increased forgetting caused by adversarial training under class incremental setting (55.49% for clean samples and 43.18% for adversarial samples forgetting, with corresponding FAA improvements of 15.46% and 2.86%, respectively). AFLC does not show significant improvement in the task incremental setting due to excessive suppression of future task classification heads and the use of the same calibration value for classes within the same task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.905, + 0.543 + ], + "angle": 0, + "content": "RAER (Sec. 4.3) can further improve the robust accuracy of AFLC by \\(1.23\\%\\) for class incremental setting and \\(2.07\\%\\) for task incremental setting and reduce the robust forgetting by \\(12.16\\%\\) and \\(3.47\\%\\) respectively. That proves the data selected by RAER describe the overall data distribution more accurately and effectively mitigate the gradient obfuscation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.68 + ], + "angle": 0, + "content": "When considering the future prior adjustment (FP in Table 4), we find that although the forgetting of the class incremental setting is higher, the FAA of both clean samples and adversarial samples has been significantly improved, and the forgetting of the task incremental setting has been further reduced, which proves that FP can reduce negative gradients to future classes and help learn new tasks. We also provide sensitivity analysis of hyperparameters (\\(\\alpha\\) in AFLC and \\(\\rho\\) in RAER) in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.783, + 0.71 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This paper explores achieving adversarial robust continual learning without additional data. We analyze the challenges of directly combining continual learning and adversarial training - accelerated forgetting and gradient obfuscation. To address these, we propose: anti-forgettable logit calibration to reduce negative effects of past knowledge, and a robust difficulty-based data selection mechanism to mitigate gradient obfuscation. Experiments demonstrate the effectiveness of our approach. As AI systems increasingly need to learn continuously in dynamic environments, our work represents a further step toward more resilient and adaptable machine learning systems." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "569" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.113, + 0.484, + 0.189 + ], + "angle": 0, + "content": "This work was supported in part by the Beijing Natural Science Foundation under No. L221013, Strategic Priority Research Program of the Chinese Academy of Sciences (No. XDB0680202), and the Innovation Funding of ICT, CAS under Grant No. E561160, E561090." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.202, + 0.188, + 0.217 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.226, + 0.484, + 0.296 + ], + "angle": 0, + "content": "[1] Ali Abbasi, Parsa Nooralinejad, Hamed Pirsiavash, and Soheil Kolouri. Brainwash: A poisoning attack to forget in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24057-24067, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.297, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.34, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.396, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[4] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283. PMLR, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[5] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[6] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[8] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 532-547. Springer, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In Proceedings of the 26rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.226, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[12] Tianlong Chen, Sijia Liu, Shiyu Chang, Lisa Amini, and Zhangyang Wang. Queried unlabeled data improves and robustifies class-incremental learning. Transactions on Machine Learning and Data Mining, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[13] Ting-Chun Chou, Jhih-Yuan Huang, and Wei-Po Lee. Continual learning with adversarial training to enhance robustness of image recognition models. In 2022 International Conference on Cyberworlds, pages 236-242. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[14] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[15] Jiequan Cui, Shu Liu, Liwei Wang, and Jiaya Jia. Learnable boundary guided adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15721-15730, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.908, + 0.501 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.906, + 0.571 + ], + "angle": 0, + "content": "[18] Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A Kai Qin, and Yuan He. Advdrop: Adversarial attack to dnns by dropping information. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7506-7515, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[19] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[20] Yunhui Guo, Mingrui Liu, Yandong Li, Liqiang Wang, Tianbao Yang, and Tajana Rosing. Attacking lifelong learning models with gradient reversion. 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[21] Ahmad Hassanpour, Majid Moradikia, Bian Yang, Ahmed Abdelhadi, Christoph Busch, and Julian Fierrez. Differential privacy preservation in robust continual learning. IEEE Access, 10:24273-24287, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[22] Siteng Kang, Zhan Shi, and Xinhua Zhang. Poisoning generative replay in continual learning to promote forgetting. In International Conference on Machine Learning, pages 15769-15785. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[23] Hikmat Khan, Nidhal Carla Bouaynaya, and Ghulam Rasool. Adversarily robust continual learning. In 2022 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Hikmat Khan, Pir Masoom Shah, Syed Farhan Alam Zaidi, et al. Susceptibility of continual learning against adversarial attacks. arXiv preprint arXiv:2207.05225, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "570" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[27] Lilly Kumari, Shengjie Wang, Tianyi Zhou, and Jeff Bilmes. Retrospective adversarial replay for continual learning. In Advances in Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[28] Sungyoon Lee, Hoki Kim, and Jaewook Lee. Graddiv: Adversarial robustness of randomized neural networks via gradient diversity regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2645-2651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[30] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[32] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, and Joost van de Weijer. Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[35] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In International Conference on Learning Representations, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[36] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. International Conference on Machine Learning, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[38] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural Networks*, 113:54–71, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97(2):285, 1990. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[40] Mohammad Rostami, Leonidas Spinoulas, Mohamed Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14851-14860, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[42] Fahad Sarfraz, Elahe Arani, and Bahram Zonooz. Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning. In The Eleventh International Conference on Learning Representations, 2023, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[43] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In European Conference on Computer Vision, pages 682-698. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[44] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9630-9638, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[45] Stanford. Tiny imagenet challenge (cs231n). http://tiny-imagenet.herokuapp.com/, 2015. Accessed: Feb 21, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[46] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[47] Muhammad Umer, Glenn Dawson, and Robi Polikar. Targeted forgetting and false memory formation in continual learners through adversarial backdoor attacks. In 2020 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[48] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[51] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "571" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[53] Tongtong Wu, Massimo Caccia, Zhuang Li, Yuan-Fang Li, Guilin Qi, and Gholamreza Haffari. Pretrained language model in continual learning: A comparative study. In International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[54] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8659-8668, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[55] Ju Xu, Jin Ma, Xuesong Gao, and Zhanxing Zhu. Adaptive progressive continual learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6715-6728, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[56] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[58] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[59] Jie Zhang, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, Lei Zhang, and Chao Wu. Towards efficient data free black-box adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15115-15125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[61] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[62] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[63] Kaijie Zhu, Jindong Wang, Xixu Hu, Xing Xie, and Ge Yang. Improving generalization of adversarial training via robust critical fine-tuning. arXiv preprint arXiv:2308.02533, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "572" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_origin.pdf b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3353805b5c1b918333e102dcce730a22c1436829 --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/8c1dfab1-b341-43f5-bde7-c2f8e9dc2261_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad37e134c79afde6121590a16e3b81213ca8fa72a40b4f7258f1e914c7a733d +size 3903989 diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/full.md b/2025/Adversarial Robust Memory-Based Continual Learner/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1430a1ba1045e0e8bf8627402ade869b9549e221 --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/full.md @@ -0,0 +1,394 @@ +# Adversarial Robust Memory-Based Continual Learner + +Xiaoyue $\mathrm{Mi}^{1,2}$ , Fan Tang $^{1,2*}$ , Zonghan Yang $^{3,4}$ , Danding Wang $^{1,2}$ , Juan Cao $^{1,2}$ , Peng Li $^{3}$ , Yang Liu $^{3,4}$ + +$^{1}$ Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences + +$^{3}$ Institute for AI Industry Research (AIR), Tsinghua University + +$^{4}$ Department of Computer Science & Technology, Tsinghua University + +# Abstract + +Despite the remarkable advances that have been made in continual learning, the adversarial vulnerability of such methods has not been fully discussed. We delve into the adversarial robustness of memory-based continual learning algorithms and observe limited improvement in robustness by directly applying adversarial training techniques. Our preliminary studies reveal the twin challenges for building adversarial robust continual learners: accelerated forgetting in continual learning and gradient obfuscation in adversarial robustness. In this study, we put forward a novel adversarial robust memory-based continual learner that adjusts data logits to mitigate the forgetting of pasts caused by adversarial samples. Furthermore, we devise a gradient-based data selection mechanism to alleviate the gradient obfuscation caused by limited stored data. The proposed approach can widely integrate with existing memory-based continual learning and adversarial training algorithms. Extensive experiments on Split-CIFAR10/100 and Split-TinyImageNet demonstrate the effectiveness of our approach, achieving a maximum forgetting reduction of $34.17\%$ in adversarial data for ResNet, and $20.10\%$ for ViT. + +# 1. Introduction + +Continual learning [48, 62] enables intelligent models to continually acquire new knowledge and adapt to changing environments while maintaining previous capabilities. It has recently emerged as a promising direction to achieve ideal intelligent models, and some pioneer works have investigated the robustness of continual learning in various aspects [42, 53, 56]. Achieving robust continual learners represents a crucial advancement in enabling intelligent models to adapt to dynamic scenarios effectively. + +However, when dealing with streaming non-i.i.d. data, + +![](images/1cc626a1108478708580c34b8405ebfe88e9f524f95937419d79f94d89cd1df6.jpg) +Figure 1. A preliminary analysis of the combination of continual learning and adversarial training on Split-CIFAR10 with buffer size 200. We consider the final average accuracy (FAA) and forgetting of adversarial and clean samples in both class incremental and task incremental settings. The red numbers indicate the performance after adding our approach. + +continual learners are usually vulnerable to adversarial examples [23]. Chen et al. [12] first attempt at preserving adversarial robustness in continual learning [29] by leveraging adversarial training [52, 57] and a large amount of unlabeled data. Despite the impressive results of Chen et al. [12], the inherent challenges for conducting adversarial robust continual learners are still under disclosure. + +In this study, we investigate building adversarial robust continual learners without abundant unlabeled data. We select two representative approaches for continual learning and adversarial defenses: memory-based continual learning and adversarial training. Figure 1 shows the direct combination of memory-based continual learning (ER [10]) with an adversarial training approach (TRADES [57]). The results for adversarial data reconfirm that continual learners risk adversarial attacks [12, 20, 23, 24], and demonstrate a critical trade-off: while adversarial training can strengthen continual learners against attacks [12, 20, 23, 24], it simultaneously increases forgetting on clean data, thus compromising the core objective of continual learning. + +Further experimental analysis in Sec. 3 reveals two key + +challenges when combining adversarial training with continual learning: (1) Adversarial data increase negative gradients toward past classes compared with clean data, accelerating catastrophic forgetting in continual learning; (2) Insufficient historical data in continual learning leads to gradient obfuscation [4], compromising the effectiveness of adversarial training compared to joint training. These challenges emerge naturally from the direct integration of the two approaches. + +Based on the analysis, we propose an adversarial robust memory-based continual learner without a large amount of additional data, which consists of an anti-forgettable logit calibration (AFLC) module and a robustness-aware experience replay (RAER) strategy. The AFLC module designs different logit calibration strategies for past, current, and future classes, thus mitigating the negative gradients of adversarial samples for continual learning. By leveraging attacking difficulty, RAER selects data that are adversarial safety and maintains the diversity of data distribution, rather than storing data overfitting to the decision boundary [9]. Our approach can be combined with most memory-based continual learning as well as adversarial training algorithms, and we have demonstrated the effectiveness of our approach on the Split-CIFAR10/100 and more challenging dataset Split-Tiny-ImageNet across ResNet and ViT architectures. + +Our contributions can be summarized as follows: + +- We reveal the twin critical challenges to carrying out adversarial robust memory-based continual learning algorithms: accelerated forgetting for continual learning and gradient obfuscation for adversarial robustness. +- We propose antiforgettable logit calibration to reduce negative influences from adversarial training through task-adaptive logit calibration, and robustness-aware experience replay by storing robust and distribution-representative past data to alleviate gradient obfuscation. +- Experiments on Split-CIFAR10/100 and Split-TinyImageNet across ResNet and ViT demonstrate our effectiveness, in improving both clean accuracy and adversarial robustness of continual learning and mitigating forgetting in both class-incremental and task-incremental settings without additional data. + +# 2. Related Works + +Continual learning. Continual learning [38, 48, 62] enables models to adapt to evolving data distributions while preserving past knowledge. To alleviate catastrophic forgetting, existing approaches [34, 39] can be categorized into three types: memory-based [7, 8, 10], regularization-based [25, 29, 30], and dynamic architecture [17, 41, 50, 55, 61]. Continual learning can be categorized into class incremental (class-IL) and task-incremental (task-IL) settings [33] based on the availability of task information during inference. In both settings, memory-based meth + +ods have demonstrated superior performance without requiring model expansion [7, 60]. Researchers have explored continual learning robustness under various constraints [19, 53, 56], and saft risks about backdoor attacks [1, 22, 47] and privacy concerns [21]. Recent studies have revealed both the vulnerability of continual models to adversarial attacks and the potential of adversarial samples for mitigating forgetting [20, 23, 24, 27, 49]. Chen et al. [12] combined LwF [29] with adversarial training using additional data, while our work uniquely analyzes the core challenges in achieving adversarial robustness for continual learning without such requirements. + +Adversarial defense. The vulnerability of deep neural networks to adversarial examples [18, 46, 52, 59] has led to numerous defense methods [2, 5], with adversarial training emerging as the most effective approach [14, 15, 28, 31, 32, 36, 37, 57]. Recent research has extended to real-world scenarios, such as long-tailed distributions [54] and open-world settings [43]. Meanwhile some studies show the potential of continual algorithms in adapting to new attacks [13, 40]. Different from them, we specifically address adversarial robustness in class-incremental and task-incremental settings. + +# 3. Baselines and Their Inherent Problems + +# 3.1. Preliminaries + +Memory-based continual learning. Following [6, 7], we consider continual learning across $T$ distinct classification tasks. We pre-allocate classification heads for all tasks at initialization rather than expanding them incrementally. For the $t$ -th task ( $1 \leq t \leq T$ ), we denote its data distribution as $D_{t}$ , with input samples $x_{t}$ and corresponding labels $y_{t}$ . A model $\mathrm{M}_{\theta}(\cdot)$ with parameters $\theta$ is optimized sequentially in the given task order. Let $h_{\theta}(x_{t}) \in \mathbb{R}^{C}$ denote the output logits where $h_{\theta}(x_{t}) = \mathrm{M}_{\theta}(x_{t})$ , and $C$ is the total number of classes. The softmax output is defined as $f_{\theta}(x_{t}) \triangleq \operatorname{softmax}(h_{\theta}(x_{t}))$ . Given a logit vector $h_{\theta}(x_{t}) = [h_{\theta}(x_{t})_{1}, \dots, h_{\theta}(x_{t})_{C}]$ where $h_{\theta}(x_{t})_{i}$ represents the logit for the $i$ -th class, we partition the classes into three sets based on the current task $t$ : + +$$ +p = \{1, 2, \dots , (t - 1) b \}, +$$ + +$$ +c = \{(t - 1) b + 1, (t - 1) b + 2, \dots , t b \}, \tag {1} +$$ + +$$ +u = \{t b + 1, t b + 2, \dots , C \}, +$$ + +where $b$ is the number of classes per task. Accordingly, we divide $h_{\theta}(x_t)$ into three parts: $h_{\theta}(x_t)_p$ , $h_{\theta}(x_t)_c$ , and $h_{\theta}(x_t)_u$ , representing logits for past, current, and future classes respectively. + +Continual learner aims to correctly classify data of all + +
Buffer SizeMethodW/ ATClass Incremental SettingTask Incremental SettingCRD↓FRI↓RRD↓
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
-Joint (upper bound)X91.81±0.07-0.00±0.00-98.17±0.12-0.00±0.00-7.92--
✓79.33±0.36-50.93±1.28-94.80±0.30-74.63±1.13-
-SGD (lower bound)X19.66±0.0796.39±0.675.17±2.5833.31±5.1966.84±6.5037.42±7.394.44±3.7833.34±3.490.91-3.7341.10
✓19.34±0.0591.93±0.2015.65±0.1869.71±1.7765.33±5.9134.43±7.1627.71±1.8854.56±3.47
200ERX48.80±1.8460.00±2.250.27±0.1812.60±2.1392.89±0.265.00±0.260.01±0.0113.78±0.4114.5112.9131.70
✓28.18±0.6980.58±1.0517.86±0.2969.58±1.1584.49±0.6110.23±0.9844.30±1.0536.89±1.11
5120X83.34±1.3815.35±1.710.08±0.1113.65±2.3496.84±0.170.71±0.190.06±0.0421.83±8.6913.5312.1118.80
✓61.88±0.7437.72±0.7327.28±0.5641.66±1.2291.24±0.192.56±0.1256.59±0.8819.34±1.11
200DERX62.71±2.7638.75±4.110.01±0.007.42±2.1391.32±0.836.95±1.440.09±0.057.39±2.1817.7215.7334.80
✓35.79±2.8565.67±4.1616.44±0.2152.84±3.7182.79±0.2111.48±0.6639.52±0.3740.64±1.38
5120X83.99±0.9910.57±1.510.07±0.077.05±2.0395.67±0.131.63±0.130.05±0.057.39±0.4816.7415.6624.32
✓56.58±3.8740.40±4.0723.26±0.4445.06±1.4889.62±0.673.12±0.2353.66±2.0122.95±3.68
200DER++X65.55±2.8132.17±4.260.24±0.327.38±3.3792.11±0.558.85±2.980.22±0.3310.45±0.9414.478.2437.14
✓45.43±3.2846.67±5.0914.46±1.3535.09±2.7083.29±1.0710.83±1.7736.82±0.5934.09±2.23
5120X85.74±0.657.32±0.881.51±1.276.10±0.5396.13±0.531.14±0.350.82±0.526.06±0.4910.638.2122.18
✓68.67±1.1123.68±2.6626.95±0.2927.68±2.4891.95±0.331.20±0.9054.26±0.4116.87±1.72
200X-DERX60.96±1.0912.29±0.090.00±0.003.84±1.1894.33±0.632.49±0.170.13±0.032.59±0.6420.237.6524.02
✓34.04±0.9025.13±4.0316.82±0.9827.84±4.1680.8±0.664.96±1.6660.83±0.5610.99±0.81
5120X69.93±0.634.33±0.260.21±0.124.52±1.3996.94±0.020.29±0.040.13±0.022.04±0.4324.9618.5221.27
✓34.29±0.6538.92±0.0319.24±0.1038.79±0.4182.67±1.682.74±0.8364.11±0.128.14±0.98
+ +Table 1. We conduct a systematic experimental analysis combining adversarial training (specifically Vanilla AT [31]) with memory-based continual learning algorithms. See Sec. 3.2 for a detailed summary of observations. Performance is evaluated using several metrics (detailed in Sec. 5.1): Final Average Accuracy (FAA), Forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase on clean data (FRI), and Robust Relative Decrease (RRD) compared to joint training. All adversarial samples are generated using PGD-20. Note that the reported CRD, FRI, and RRD values are averaged across both task-incremental and class-incremental settings. + +observed tasks, and its objective can be formalized as: + +$$ +\underset {\theta} {\arg \min } \sum_ {t = 1} ^ {T} \mathcal {L} _ {t}, \text {w h e r e} \mathcal {L} _ {t} \triangleq \mathbb {E} _ {\left(x _ {t}, y _ {t}\right) \sim D _ {t}} \left[ \ell \left(f _ {\theta} \left(x _ {t}\right), y _ {t}\right) \right], \tag {2} +$$ + +where $\ell$ is the task-specific loss. The unavailability of past task data poses great challenges for continual learning. However, direct sequential optimization can lead to severe catastrophic forgetting. To address the above challenge, memory-based continual learners usually store a few past task data and replay them during the current task training phase [7], and become leading-edge approaches. When training the $t$ -th task, its loss function can be formalized as the following: + +$$ +\mathcal {L} _ {t} \triangleq \mathbb {E} _ {(x, y) \sim D _ {t} \cup \mathcal {M}} [ \ell (f _ {\theta} (x), y) ], \tag {3} +$$ + +where $\mathcal{M}$ is the episodic memory of the current task, which preserves a small subset of observed past data $\langle x_{\mathcal{M}},y_{\mathcal{M}}\rangle$ + +Adversarial training. Adversarial training is currently the most reliable way [5] to improve the adversarial robustness of a model. Following [31], it can be formally defined as a min-max optimization problem: + +$$ +\underset {\theta} {\arg \min } \mathbb {E} _ {(x, y) \sim D} \left(\underset {\delta \in S} {\max } \ell \left(f _ {\theta} (x + \delta), y\right)\right), \tag {4} +$$ + +where $\delta$ is an adversarial perturbation, $D$ is data distribution, and $S$ is a perturbation set. $S = \left\{\delta \mid \| \delta \|_p \leq \epsilon \right\}$ , $\epsilon$ is + +the perturbation size. The goal of adversarial training is to correctly classify all adversarial examples $(\widetilde{x} = x + \delta)$ . + +Here we integrate memory-based continual learner and adversarial training as adversarial robust memory-based continual learner baselines, and the new optimization objective during the training stage of $t$ -th task is: + +$$ +\underset {\theta} {\arg \min } \mathbb {E} _ {(x, y) \sim D _ {t} \cup \mathcal {M}} \left(\underset {\delta \in S} {\max } \ell \left(f _ {\theta} (x + \delta), y\right)\right). \tag {5} +$$ + +# 3.2. Vanilla Solutions + +To investigate the challenges in adversarial robust continual learning, we analyze several popular memory-based continual learning methods (ER [10], DER [7], DER++ [7], and X-DER [6]) on Split-CIFAR10 [26] under both class and task incremental settings. We employ Vanilla AT [31] for adversarial training due to its classic and strong performance [36]. Performance is evaluated against an upper bound ("Joint") and a lower bound ("SGD") using multiple metrics: Final Average Accuracy (FAA), forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase in clean data (FRI), and Robust Relative Decrease (RRD). Detailed metric definitions are provided in Sec. 5.1, and additional implementation details and hyper-parameters are provided in the supplementary material. + +Results in Table 1 reveal that direct integration of adversarial training with continual learning is suboptimal. First, the results validate our observations from Sec. 1: while + +adversarial training enhances robustness, it compromises clean performance (positive CRDs) and accelerates forgetting (positive FRIs). Larger memory sizes consistently improve performance across all metrics. Moreover, continual models show inferior robustness improvement compared to joint models (positive RRDs). + +To investigate the underlying mechanisms, we select ER as our base model for further analysis. While DER/DER++ and X-DER demonstrate better stability with lower forgetting, they struggle with new task learning, resulting in poor robust FAA. In contrast, ER combined with adversarial training exhibits strong robustness. + +# 3.3. The Inherent Problems + +By analyzing gradients during training, we identify two critical challenges in robust continual learning: accelerated forgetting in continual learning and gradient obfuscation in adversarial training. + +Accelerated forgetting in continual learning. Adversarial training hastens the forgetting of clean data in continual learners. According to [31], adversarial examples are designed to mislead model predictions by increasing probabilities for incorrect classes while suppressing those of the true class. In the continual learning setting, adversarial examples from the current task tend to increase probabilities for past classes while decreasing probabilities for their true class. Specifically, for an adversarial example $\widetilde{x}_t$ generated from a clean sample $x_t$ of current task $t$ : + +$$ +f _ {\theta} \left(\widetilde {x} _ {t}\right) _ {p} > f _ {\theta} \left(x _ {t}\right) _ {p}, f _ {\theta} \left(\widetilde {x} _ {t}\right) _ {\text {t r u e c l a s s}} < f _ {\theta} \left(x _ {t}\right) _ {\text {t r u e c l a s s}}. \tag {6} +$$ + +When using cross-entropy as the loss function, adversarial data amplify negative gradients for current data $x_{t}$ for past tasks: + +$$ +\left[ \frac {\partial \ell}{\partial h _ {\theta} (\widetilde {x} _ {t})} \right] _ {p} = f _ {\theta} (\widetilde {x} _ {t}) _ {p} > f _ {\theta} (x _ {t}) _ {p} = \left[ \frac {\partial \ell}{\partial h _ {\theta} (x _ {t})} \right] _ {p}. \tag {7} +$$ + +Our empirical analysis reveals that these adversarial examples generate stronger negative gradients toward past tasks, especially during the initial training phases. In Figure 2, experiments with $\mathrm{ER + AT}$ on Split-CIFAR10 using a buffer size of 200 show adversarial data consistently produces larger gradient norms compared to clean data. The larger gradient norms indicate stronger positive gradients for the true class and stronger negative gradients for other classes, including past classes, thus accelerating forgetting of previously learned tasks. While memory-based methods using stored adversarial samples can help mitigate catastrophic forgetting, the limited buffer size (typically small in practical applications) means that adversarial training on the entire dataset still accelerates forgetting. + +Gradient obfuscation in adversarial training. Gradient obfuscation [4] occurs when gradient information is obscured by non-differentiable operators or randomness, leading to inferior robustness. We observe a specific form of + +![](images/388c31bd51e2ed97f3e1bfe631bb4f02b505d540bcaaca5f035f2dccd035589d.jpg) + +![](images/bac76602735a5d52be05b919cc8d47fc54f62544c4858256ea4b9df9e3b2ea2c.jpg) + +![](images/b44940afc77cdc7474f845a958bf459d85850c453c74c6be9c157ded6c23237b.jpg) +Figure 2. Gradient norms of adversarial data and clean data during the training stage. We experiment ER+AT with a 200 buffer size on the Split-CIFAR10 dataset. +Figure 3. The gradient cosine similarities between robust joint model and robust continual model, which indicates a special gradient obfuscation only in continual learning. We experiment with buffer size=200/5,120 on the Split-CIFAR10 dataset. Gradient obfuscation weakens as the buffer size gets larger, and our method also mitigates gradient obfuscation. + +gradient obfuscation, termed "shattered gradients", in adversarial training under continual learning, resulting in less effective robustness compared to joint models. + +In continual learning, gradient information is biased due to limited replay data. With few old task samples, the model tends to overfit, causing direction-biased gradients when generating adversarial examples. Our analysis of gradient directions between adversarial robust joint and continual models reveals significant differences, contributing to lower adversarial robustness in continual learners (see Figure 3). + +# 4. Method + +# 4.1. Overview + +To address the identified challenges, we propose a novel framework consisting of two key components: antiforgettable logit calibration (AFLC) to mitigate forgetting acceleration, and robustness-aware experience replay (RAER) to reduce gradient obfuscation. The full pipeline is illustrated in Figure 4: during training each task, the + +![](images/a493aef0c46f4adcd6adc6b6bd32607dc81091936c0bf940dab7ef0a1615dbbe.jpg) +Figure 4. The pipeline of our adversarial robust memory-based continual learner. First, generate adversarial samples $(\widetilde{X}_t,\widetilde{X}_{\mathcal{M}})$ and robust difficulty factors $(K_{t})$ for both current task data $(X_{t},Y_{t})$ and memory data $(X_{\mathcal{M}},Y_{\mathcal{M}})$ . The model then processes these samples through AFLC (Sec. 4.2), which calibrates logits based on task order to reduce negative gradients toward past tasks. Meanwhile, RAER (Sec. 4.3) leverages the robust difficulty factors $K_{t}$ to select representative samples for memory storage, mitigating gradient obfuscation. + +![](images/0b7d85f31fde0b9ee1c57faeae9cef9a37694110bbc4e526a4700c20d5926981.jpg) + +model processes both current task data and memory samples through the model. AFLC then calibrates logits based on task order (past/current/future) before loss computation. Simultaneously, RAER evaluates current task samples using a robustness difficulty factor $k$ to select representative data for memory storage. + +# 4.2. Anti-Forgettable Logit Calibration + +We introduce AFLC to regulate the influence of adversarial examples on past task knowledge. The core idea is to calibrate logits differently for past and current tasks to control gradient flow. + +Given an input $x$ (either from current task $x_{t}$ or memory buffer $x_{\mathcal{M}}$ ), AFLC applies a task-dependent calibration: + +$$ +h _ {\theta} ^ {\mathrm {l c}} (x) _ {i} = h _ {\theta} (x) _ {i} - v _ {i}, v _ {p} > v _ {c}, \tag {8} +$$ + +where $v_{i}$ represents the calibration factor for class i: + +$$ +v _ {i} = - \log \left(\frac {n _ {i}}{\sum_ {j = 1} ^ {C} n _ {j}}\right) - \alpha_ {i}, \tag {9} +$$ + +where $n_i$ denotes sample count for class $i$ , and $\alpha_i$ is a hyperparameter. When all tasks share the classification head, we realize that making logit changes to past and current classes only is equivalent to negative gradient increases. So we make a Further Prior (FP) adjustment that essentially averages the adjustments applied to current and past classes and applies this value to future classes, helping to balance the gradients across all periods. + +$$ +v _ {i} = \frac {1}{t \cdot b} \sum_ {j = 1} ^ {t \cdot b} \left(v _ {j}\right), \tag {10} +$$ + +where $t$ is the current task order and $b$ is the number of classes per task. + +The calibrated probability distribution is computed as: + +$$ +f _ {\theta} ^ {\mathrm {l c}} (x) _ {i} = \frac {\exp \left(h _ {\theta} ^ {\mathrm {l c}} (x) _ {i}\right)}{\sum_ {j = 1} ^ {C} \exp \left(h _ {\theta} ^ {\mathrm {l c}} (x) _ {j}\right)}, \tag {11} +$$ + +and leads to two key properties: For current task samples, + +$$ +\left[ \frac {\partial \ell}{\partial h _ {\theta} ^ {\mathrm {l c}} (\widetilde {x})} \right] _ {p} = f _ {\theta} ^ {\mathrm {l c}} (\widetilde {x} _ {t}) _ {p} < f _ {\theta} (\widetilde {x} _ {t}) _ {p}, \tag {12} +$$ + +and for memory buffer samples, + +$$ +\left[ \frac {\partial \ell}{\partial h _ {\theta} ^ {\mathrm {l c}} (\widetilde {x _ {\mathcal {M}}})} \right] _ {c} = f _ {\theta} ^ {\mathrm {l c}} (\widetilde {x} _ {\mathcal {M}}) _ {c} > f _ {\theta} (\widetilde {x} _ {\mathcal {M}}) _ {c}. \tag {13} +$$ + +In that way, AFLC can mitigate forgetting acceleration by reducing the negative gradient to past classes from current adversarial data and improving the negative gradient to current classes from memory adversarial data. + +Comparisons with other logit calibration CL. X-DER [6] also adjusts the negative gradients of the new task data affecting the past tasks by updating the new task with the logit masking. Logit masking can be seen as a special case of our method at the softmax layer, which can be viewed as $\mathrm{v}_p = \mathrm{v}_c = 0$ for past data, and $\mathrm{v}_p = +\infty$ , $\mathrm{v}_c = 0$ for current data. However, their extreme margin $\mathbf{V}$ inhibits the model's ability to learn new knowledge. X-DER further adds a safeguard threshold to limit the activations of past and future heads, which also can be combined with us. + +# 4.3. Robustness-Aware Experience Replay + +To address gradient obfuscation, we propose RAER, which selects samples based on their robustness characteristics. For each sample $x$ , we measure its robustness difficulty factor $k$ during PGD-10 attack, without additional forward or + +backward calculation, where $k \in [0,10] \cap \mathbb{N}$ indicates the number of successful attacks. Samples meeting the threshold condition $k < \rho$ are candidates for memory storage. The memory update process can be formalized as: + +$$ +\mathcal {M} \leftarrow \text {R a n d o m S a m p l e} \left(\{x \mid k _ {x} < \rho \}, m\right), \tag {14} +$$ + +where $m$ is the memory size and $\rho$ is set to 5 in practice. + +Why RAER works. The robustness difficulty factor $k$ shows how easily a sample can be attacked: a high $k$ suggests vulnerability and proximity to decision boundaries, while a low $k$ implies robustness and distance from boundaries. We selectively store samples with $k < \rho$ in memory, excluding boundary-overfitting outliers. Unlike prior continual learning [9] and adversarial training [58] methods, RAER emphasizes distribution-representative and robust learnable samples while maintaining model reliability. + +The computational overhead of RAER is minimal, as $k$ is calculated during the existing adversarial example generation process, without additional forward or backward calculation. + +Comparisons with other data-selection CL. Existing works [3, 44] alleviate catastrophic forgetting using data selection and focus solely on clean samples. RAER uniquely targets gradient obfuscation in continual robust settings by: Selecting adversarially safe samples for better retention; excluding high- $k$ boundary samples; and maintaining representative data distribution. As demonstrated in Sec. 5, RAER outperforms traditional data selection strategies combined with adversarial training, particularly in adversarial robustness. + +# 4.4. Overall Objective + +To sum up, we design AFLC to reduce negative gradients of the current adversarial sample to past tasks, and RAER to select data that alleviate gradient obfuscation to store. We set $\alpha$ of AFLC is 3.5 in practice and $x_{\mathcal{M}}$ and $y_{\mathcal{M}}$ are selected by RAER with $\rho = 5$ . When $\mathrm{ER + AT}$ combines with ours, the loss of task t is: + +$$ +\mathcal {L} _ {t} \triangleq \mathrm {C E} \left(f _ {\theta} ^ {l c} \left(\tilde {x} _ {t}\right), y _ {t}\right) + \mathrm {C E} \left(f _ {\theta} ^ {l c} \left(\tilde {x} _ {\mathcal {M}}\right), y _ {\mathcal {M}}\right). \tag {15} +$$ + +# 5. Experiments + +# 5.1. Experimental Settings + +Datasets and training details. Following standard practices in adversarial training and continual learning [6, 37, 63], we evaluate our approach on Split-CIFAR10 and Split-CIFAR100 [26] and the model architecture is ResNet18, trained using SGD optimizer with learning rate 0.1. Split-CIFAR10 comprises ten classes (5,000 training, 1,000 test samples per class) divided into five binary classification tasks. Split-CIFAR100 contains 100 classes (500 training, + +100 test samples per class) split into ten-way classification tasks. Evaluations on Split-Tiny-ImageNet [45] and ViT [16] in supplementary materials. + +For adversarial training, we use perturbations of 8/255 and step size of 2/255. Data augmentation includes random cropping (4-pixel padding) and horizontal flipping for all data. Unlike Chen et al. [12], we achieve robustness without additional unlabeled data1. All experiments are conducted with three different random seeds. + +Evaluation metrics. We evaluate both adversarial robustness and continual learning performance using Final Average Accuracy (FAA) and forgetting metrics on clean and adversarial samples. For adversarial evaluation, we employ PGD-20 (white-box attack) and Auto Attack (AA) [14], which combines black-box and white-box attacks for reliable robustness assessment [36, 37, 51]. Extended robustness verification using additional black-box (RayS [11]) and adaptive attacks in the supplementary material. Let $a_{i}^{t}$ denote the accuracy for task $i$ after training on task $t$ . The metrics are defined as: + +$$ +\mathrm {F A A} \triangleq \frac {1}{T} \sum_ {i = 1} ^ {T} a _ {i} ^ {T}, \tag {16} +$$ + +$$ +\text {f o r g e t t i n g} \triangleq \frac {1}{T - 1} \sum_ {j = 1} ^ {T - 1} f _ {j}, \text {s . t .} f _ {j} = \max _ {l \in \{1, \dots , T - 1 \}} a _ {i} ^ {l} - a _ {j} ^ {T}. \tag {17} +$$ + +Forgetting, ranging from $[-100, 100]$ , measures the average decrease in accuracy across tasks. + +For analysis, we further define: + +$$ +\mathrm {C R D} \triangleq \mathrm {F A A} _ {\text {c l e a n}} - \widetilde {\mathrm {F A A}} _ {\text {c l e a n}}, \tag {18} +$$ + +$$ +\mathrm {F R I} \triangleq \widetilde {\mathrm {F o r g e t t i n g}} _ {\mathrm {c l e a n}} - \mathrm {F o r g e t t i n g} _ {\mathrm {c l e a n}}, \tag {19} +$$ + +$$ +\mathrm {R R D} \triangleq \left(\widetilde {\mathrm {F A A}} _ {\mathrm {a d v}} ^ {\text {J o i n t}} - \mathrm {F A A} _ {\mathrm {a d v}} ^ {\text {J o i n t}}\right) - \left(\widetilde {\mathrm {F A A}} _ {\mathrm {a d v}} - \mathrm {F A A} _ {\mathrm {a d v}}\right), \tag {20} +$$ + +where "clean" and "adv" denote metrics on clean and adversarial data, respectively, tilde $(\sim)$ indicates robust models, and "Joint" refers to a joint training model. + +# 5.2. Main Results + +We evaluate our method against various adversarial training and memory-based continual learning combinations on Split-CIFAR10/100. Evaluations on Split-TinyImageNet [45] and ViT [16] in supplementary materials. Due to computational constraints, we focus on integrating our approach with ER+AT and ER+TRADES ( $\beta = 6.0$ ), which showed superior adversarial FAA performance among baselines. + +
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
200ER+AT28.1880.5817.8616.9469.5884.4910.2344.3044.6936.89
ER+TRADES22.4277.2515.7215.5364.9578.798.1051.3351.5021.97
ER+FAT33.6169.2115.1414.8149.0483.4010.3543.6943.9628.56
ER+LBGAT25.6884.4716.6516.5670.5078.1918.8540.6940.7340.83
ER+SCORE48.6556.792.400.9318.9688.909.827.256.728.70
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)18.16(↑ 1.22)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)47.61(↑ 2.92)34.04(↓ 2.85)
ER+TRADES+Ours43.34(↑ 20.92)33.40(↓ 43.85)19.85(↑ 4.13)18.35(↑ 2.82)30.78(↓ 34.17)82.59(↑ 3.80)7.53(↓ 0.57)59.41(↑ 8.08)59.59(↑ 8.09)14.16(↓ 7.81)
5,120ER+AT61.8837.7227.2826.6941.6691.242.5656.5956.9019.34
ER+TRADES20.3685.1416.3016.1872.8588.481.5964.3664.5212.47
ER+FAT54.5543.1819.6818.9142.1591.502.1256.7256.8714.79
ER+LBGAT62.4537.7327.4226.6647.8391.103.2556.5756.3119.38
ER+SCORE51.3752.843.141.1020.2995.631.5814.6613.922.53
ER+AT+Ours64.34(↑ 2.46)23.64(↓ 14.08)31.31(↑ 4.03)30.49(↑ 3.80)20.46(↓ 21.20)91.00(↓ 0.24)3.51(↑ 0.95)60.49(↑ 3.90)60.61(↑ 3.71)13.27(↓ 6.07)
ER+TRADES+Ours39.80(↑ 19.44)44.08(↓ 41.06)23.07(↑ 6.77)21.98(↑ 5.80)41.37(↓ 31.48)86.48(↓ 2.00)1.71(↑ 0.12)69.25(↑ 4.89)69.38(↑ 4.86)5.33(↓ 7.14)
+ +(a) Results on Split-CIFAR10. We chose two buffer sizes of 200 and 5, 120. + +
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
500ER+AT11.9473.545.665.5438.0352.7128.4417.3519.5625.67
ER+TRADES7.5968.135.435.1144.1350.7520.2226.8927.6920.34
ER+FAT11.4873.995.355.2337.2656.1024.7120.0122.9922.31
ER+LBGAT11.7775.106.165.8239.7742.0226.4513.9914.4323.14
ER+SCORE16.2277.710.910.626.1368.8711.213.237.954.94
ER+AT+Ours24.14(↑ 12.20)53.03(↓ 20.51)7.13(↑ 1.47)6.68(↑ 1.14)25.81(↓ 12.22)56.00(↑ 3.29)26.19(↓ 2.25)17.95(↑ 0.60)20.26(↑ 0.70)24.87(↓ 0.80)
ER+TRADES+Ours23.24(↑ 15.65)24.29(↓ 43.84)9.93(↑ 4.50)7.50(↑ 2.39)11.70(↓ 32.43)56.68(↑ 5.93)21.39(↑ 1.17)27.39(↑ 0.50)28.50(↑ 0.81)16.76(↓ 3.58)
2,000ER+AT18.7765.067.207.0133.5662.0117.9721.1624.0420.47
ER+TRADES9.5070.495.355.0142.0860.6313.7826.6829.1918.60
ER+FAT17.0966.916.125.9333.6263.8815.9923.4826.7517.12
ER+LBGAT20.5863.317.937.0530.0955.7725.1117.9518.8819.18
ER+SCORE30.1061.510.750.514.1076.9019.604.8811.975.25
ER+AT+Ours31.93(↑ 13.16)40.50(↓ 24.56)9.61(↑ 2.41)9.16(↑ 2.15)17.78(↓ 15.78)63.77(↑ 1.76)17.16(↓ 0.81)23.11(↑ 1.95)25.59(↑ 1.55)17.57(↓ 2.90)
ER+TRADES+Ours28.73(↑ 19.23)24.16(↓ 46.33)12.75(↑ 7.40)11.02(↑ 6.01)14.56(↓ 27.52)62.01(↑ 1.38)13.16(↓ 0.62)34.81(↑ 8.13)35.36(↑ 6.17)11.60(↓ 7.00)
+ +(b) Results on Split-CIFAR100. We choose two buffer sizes of 500 and 2,000. +Table 2. Experiment results on Split-CIFAR10/100 datasets and model architecture is ResNet18. Here, PGD-20 and AA are adversarial data Final Average Accuracy (FAA) generated by PGD-20 and Auto Attack (AA), respectively. Forgetting of adversarial data is computed based on PGD-20. Bold represents the best experimental results for the same settings, underline indicates the second-best result, green signifies relative improvement and red indicates relative degradation. By applying our methods, various model performances can be improved across the board. + +
MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
ER+AT28.1880.5817.8669.5884.4910.2344.3036.89
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)34.04(↓ 2.85)
GSS+AT27.5980.7816.6768.5384.419.8344.2534.79
GSS+AT+Ours36.93(↑ 9.34)67.72(↓ 13.06)16.84(↑ 0.17)60.04(↓ 8.49)85.57(↑ 1.16)8.86(↓ 0.97)47.11(↑ 2.86)31.83(↓ 2.97)
ASER+AT18.8587.7814.0665.5773.8719.0130.6544.85
ASER+AT+Ours24.45(↑ 5.61)81.73(↓ 6.05)14.91(↑ 0.85)62.74(↓ 2.83)77.70(↑ 3.83)15.21(↓ 3.80)34.50(↑ 3.85)38.55(↓ 6.30)
X-DER+AT34.0425.1316.8227.8480.804.9660.8310.99
X-DER+AT+Ours43.25(↑ 9.21)20.77(↓ 4.36)17.22(↑ 0.41)18.56(↓ 9.28)84.87(↑ 4.07)1.74(↓ 3.21)61.68(↑ 0.85)7.03(↓ 3.97)
+ +Table 3. Experiments with other data selection-based and logit masking-based continual learning methods. + +Results on different adversarial training methods. Table 2 reveals several key findings. Buffer size significantly impacts performance, with larger buffers generally improving results, except for $\mathrm{ER + TRADES}$ on Split-CIFAR10 at buffer size 5, 120. Regarding robustness, $\mathrm{ER + AT}$ demon + +strates superior performance in class incremental settings, while ER+TRADES achieves optimal results under task incremental learning, albeit with reduced clean sample accuracy. ER+LBGAT exhibits balanced performance in class incremental settings, while ER+FAT and ER+SCORE excel on clean samples but underperform on adversarial ones. + +Our method significantly enhances both ER+TRADES + +
AFLCRAERFPClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
22.4277.2515.7264.9578.798.1051.3321.97
✓37.8821.7618.5821.7781.6610.8653.6417.60
✓✓37.4510.2919.819.6178.1315.3955.7114.13
✓✓✓43.3433.4019.8530.7882.597.5359.4114.16
+ +Table 4. Ablation experiments on the Split-CIFAR10 dataset with the buffer size of 200, using ER+TRADES as the baseline. Bold represents the best experimental results for the same settings. Experiments have demonstrated that AFLC mitigates clean-sample accelerated forgetting from adversarial samples, RAER mitigates gradient obfuscation (Adversarial FAA has a boost) and robust forgetting; adding FP improves the model's ability to learn new tasks (FAA has an overall increase). + +and ER+AT across Split-CIFAR10/100 datasets, particularly in class incremental settings. For Split-CIFAR10, we achieve FAA improvements of $\Delta_{clean} = 20.92\%$ and $\Delta_{adv} = 8.09\%$ , with forgetting reductions of $\Delta_{f_{clean}} = 43.85\%$ and $\Delta_{f_{adv}} = 34.17\%$ . On Split-CIFAR100, improvements reach $\Delta_{clean} = 19.23\%$ and $\Delta_{adv} = 8.13\%$ for FAA, with forgetting reductions of $\Delta_{f_{clean}} = 46.33\%$ and $\Delta_{f_{adv}} = 32.43\%$ . + +Our approach demonstrates positive improvements in $93\%$ of evaluation metrics (Table 2). Minor performance decreases in task incremental settings with buffer size 5, 120 reflect AFLC's diminishing returns and the known clean-robust accuracy trade-off [31, 35, 37, 57]. + +Results on different CL methods. We evaluate our method against state-of-the-art continual learning approaches on Split-CIFAR10 with buffer size 200, including gradient-based GSS, Shapley value-based ASER, and logit masking-based X-DER. These methods primarily address clean-data continual learning, whereas our approach specifically targets adversarial robustness. + +Table 3 demonstrates that existing data selection methods (GSS, ASER) show limited effectiveness in adversarial settings, with robustness performance not exceeding $\mathrm{ER + AT}$ . Our method, when integrated with these approaches, enhances both robustness and clean accuracy across class and task incremental settings. + +X-DER employs logit masking, a specialized case of our AFLC, for memory updates and future preparation. While effective at forgetting mitigation, its aggressive gradient suppression limits model learning capacity, resulting in lower robust FAA than $\mathrm{ER + AT}$ in class incremental settings. However, integration with our method yields significant improvements in both clean and adversarial metrics. + +The consistent performance gains across all evaluation metrics validate our method's effectiveness in simultaneously addressing adversarial robustness and catastrophic forgetting. These empirical results demonstrate the broad applicability of our approach across different continual learning frameworks. Accuracy curves during continual training can be found in the supplementary material. + +# 5.3. Ablation Experiments + +In Table 4, we have performed ablation experiments based on ER+TRADES under the Split-CIFAR10 dataset. The results demonstrate that AFLC (Sec. 4.2) can effectively mitigate the increased forgetting caused by adversarial training under class incremental setting (55.49% for clean samples and 43.18% for adversarial samples forgetting, with corresponding FAA improvements of 15.46% and 2.86%, respectively). AFLC does not show significant improvement in the task incremental setting due to excessive suppression of future task classification heads and the use of the same calibration value for classes within the same task. + +RAER (Sec. 4.3) can further improve the robust accuracy of AFLC by $1.23\%$ for class incremental setting and $2.07\%$ for task incremental setting and reduce the robust forgetting by $12.16\%$ and $3.47\%$ respectively. That proves the data selected by RAER describe the overall data distribution more accurately and effectively mitigate the gradient obfuscation. + +When considering the future prior adjustment (FP in Table 4), we find that although the forgetting of the class incremental setting is higher, the FAA of both clean samples and adversarial samples has been significantly improved, and the forgetting of the task incremental setting has been further reduced, which proves that FP can reduce negative gradients to future classes and help learn new tasks. We also provide sensitivity analysis of hyperparameters ( $\alpha$ in AFLC and $\rho$ in RAER) in supplementary materials. + +# 6. Conclusion and Future Work + +This paper explores achieving adversarial robust continual learning without additional data. We analyze the challenges of directly combining continual learning and adversarial training - accelerated forgetting and gradient obfuscation. To address these, we propose: anti-forgettable logit calibration to reduce negative effects of past knowledge, and a robust difficulty-based data selection mechanism to mitigate gradient obfuscation. Experiments demonstrate the effectiveness of our approach. As AI systems increasingly need to learn continuously in dynamic environments, our work represents a further step toward more resilient and adaptable machine learning systems. + +# Acknowledgements + +This work was supported in part by the Beijing Natural Science Foundation under No. L221013, Strategic Priority Research Program of the Chinese Academy of Sciences (No. XDB0680202), and the Innovation Funding of ICT, CAS under Grant No. E561160, E561090. + +# References + +[1] Ali Abbasi, Parsa Nooralinejad, Hamed Pirsiavash, and Soheil Kolouri. Brainwash: A poisoning attack to forget in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24057-24067, 2024. 2 +[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. 2 +[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 6 +[4] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283. PMLR, 2018. 2, 4 +[5] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356, 2021. 2, 3 +[6] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3, 5, 6 +[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3 +[8] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. In International Conference on Learning Representations, 2022. 2 +[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 532-547. Springer, 2018. 2, 6 +[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019, 2019. 1, 2, 3 +[11] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In Proceedings of the 26rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020. 6 + +[12] Tianlong Chen, Sijia Liu, Shiyu Chang, Lisa Amini, and Zhangyang Wang. Queried unlabeled data improves and robustifies class-incremental learning. Transactions on Machine Learning and Data Mining, 2022. 1, 2, 6, 7 +[13] Ting-Chun Chou, Jhih-Yuan Huang, and Wei-Po Lee. Continual learning with adversarial training to enhance robustness of image recognition models. In 2022 International Conference on Cyberworlds, pages 236-242. IEEE, 2022. 2 +[14] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2, 6 +[15] Jiequan Cui, Shu Liu, Liwei Wang, and Jiaya Jia. Learnable boundary guided adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15721-15730, 2021. 2 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 6 +[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2 +[18] Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A Kai Qin, and Yuan He. Advdrop: Adversarial attack to dnns by dropping information. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7506-7515, 2021. 2 +[19] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2 +[20] Yunhui Guo, Mingrui Liu, Yandong Li, Liqiang Wang, Tianbao Yang, and Tajana Rosing. Attacking lifelong learning models with gradient reversion. 2020. 1, 2 +[21] Ahmad Hassanpour, Majid Moradikia, Bian Yang, Ahmed Abdelhadi, Christoph Busch, and Julian Fierrez. Differential privacy preservation in robust continual learning. IEEE Access, 10:24273-24287, 2022. 2 +[22] Siteng Kang, Zhan Shi, and Xinhua Zhang. Poisoning generative replay in continual learning to promote forgetting. In International Conference on Machine Learning, pages 15769-15785. PMLR, 2023. 2 +[23] Hikmat Khan, Nidhal Carla Bouaynaya, and Ghulam Rasool. Adversarily robust continual learning. In 2022 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2022. 1, 2 +[24] Hikmat Khan, Pir Masoom Shah, Syed Farhan Alam Zaidi, et al. Susceptibility of continual learning against adversarial attacks. arXiv preprint arXiv:2207.05225, 2022. 1, 2 +[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran + +Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2 +[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 3, 6 +[27] Lilly Kumari, Shengjie Wang, Tianyi Zhou, and Jeff Bilmes. Retrospective adversarial replay for continual learning. In Advances in Neural Information Processing Systems, 2022. 2 +[28] Sungyoon Lee, Hoki Kim, and Jaewook Lee. Graddiv: Adversarial robustness of randomized neural networks via gradient diversity regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2645-2651, 2023. 2 +[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2 +[30] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 2 +[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 2, 3, 4, 8 +[32] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 2 +[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, and Joost van de Weijer. Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2023. 2 +[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 2 +[35] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In International Conference on Learning Representations, 2020. 8 +[36] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 2, 3, 6 +[37] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. International Conference on Machine Learning, 2022. 2, 6, 8 +[38] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural Networks*, 113:54–71, 2019. 2 + +[39] Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97(2):285, 1990. 2 +[40] Mohammad Rostami, Leonidas Spinoulas, Mohamed Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14851-14860, 2021. 2 +[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2 +[42] Fahad Sarfraz, Elahe Arani, and Bahram Zonooz. Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning. In The Eleventh International Conference on Learning Representations, 2023, 1 +[43] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In European Conference on Computer Vision, pages 682-698. Springer, 2020. 2 +[44] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9630-9638, 2021. 6 +[45] Stanford. Tiny imagenet challenge (cs231n). http://tiny-imagenet.herokuapp.com/, 2015. Accessed: Feb 21, 2023. 6 +[46] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2 +[47] Muhammad Umer, Glenn Dawson, and Robi Polikar. Targeted forgetting and false memory formation in continual learners through adversarial backdoor attacks. In 2020 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2020. 2 +[48] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998. PMLR, 2022. 2 +[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2 +[51] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, 2023. 6 +[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 2 +[53] Tongtong Wu, Massimo Caccia, Zhuang Li, Yuan-Fang Li, Guilin Qi, and Gholamreza Haffari. Pretrained language model in continual learning: A comparative study. In International Conference on Learning Representations, 2021. 1, 2 +[54] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8659-8668, 2021. 2 +[55] Ju Xu, Jin Ma, Xuesong Gao, and Zhanxing Zhu. Adaptive progressive continual learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6715-6728, 2022. 2 +[56] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 1, 2 +[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1, 2, 8 +[58] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In International Conference on Learning Representations, 2021. 6 +[59] Jie Zhang, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, Lei Zhang, and Chao Wu. Towards efficient data free black-box adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15115-15125, 2022. 2 +[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 2 +[61] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2 +[62] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[63] Kaijie Zhu, Jindong Wang, Xixu Hu, Xing Xie, and Ge Yang. Improving generalization of adversarial training via robust critical fine-tuning. arXiv preprint arXiv:2308.02533, 2023. 6 \ No newline at end of file diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/images.zip b/2025/Adversarial Robust Memory-Based Continual Learner/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..811033a839ea7bf13a5dcc4d74fab1d709e84947 --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:549798671e56a34b5a376715066d14d01e75ae1687bbd1a357def702f3682e69 +size 753086 diff --git a/2025/Adversarial Robust Memory-Based Continual Learner/layout.json b/2025/Adversarial Robust Memory-Based Continual Learner/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c12153816592581e26a7f029414b219bf516080a --- /dev/null +++ b/2025/Adversarial Robust Memory-Based Continual Learner/layout.json @@ -0,0 +1,9932 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 135, + 103, + 476, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 103, + 476, + 121 + ], + "spans": [ + { + "bbox": [ + 135, + 103, + 476, + 121 + ], + "type": "text", + "content": "Adversarial Robust Memory-Based Continual Learner" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "spans": [ + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": "Xiaoyue " + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{Mi}^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Fan Tang" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Zonghan Yang" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Danding Wang" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Juan Cao" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Peng Li" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "text", + "content": ", Yang Liu" + }, + { + "bbox": [ + 63, + 141, + 548, + 157 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "type": "text", + "content": "Institute of Computing Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 157, + 466, + 185 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 185, + 458, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 185, + 458, + 200 + ], + "spans": [ + { + "bbox": [ + 135, + 185, + 458, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 185, + 458, + 200 + ], + "type": "text", + "content": "Institute for AI Industry Research (AIR), Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 200, + 474, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 200, + 474, + 214 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 474, + 214 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 200, + 474, + 214 + ], + "type": "text", + "content": "Department of Computer Science & Technology, Tsinghua University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 249, + 200, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 249, + 200, + 261 + ], + "spans": [ + { + "bbox": [ + 151, + 249, + 200, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "text", + "content": "Despite the remarkable advances that have been made in continual learning, the adversarial vulnerability of such methods has not been fully discussed. We delve into the adversarial robustness of memory-based continual learning algorithms and observe limited improvement in robustness by directly applying adversarial training techniques. Our preliminary studies reveal the twin challenges for building adversarial robust continual learners: accelerated forgetting in continual learning and gradient obfuscation in adversarial robustness. In this study, we put forward a novel adversarial robust memory-based continual learner that adjusts data logits to mitigate the forgetting of pasts caused by adversarial samples. Furthermore, we devise a gradient-based data selection mechanism to alleviate the gradient obfuscation caused by limited stored data. The proposed approach can widely integrate with existing memory-based continual learning and adversarial training algorithms. Extensive experiments on Split-CIFAR10/100 and Split-TinyImageNet demonstrate the effectiveness of our approach, achieving a maximum forgetting reduction of " + }, + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "inline_equation", + "content": "34.17\\%" + }, + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "text", + "content": " in adversarial data for ResNet, and " + }, + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "inline_equation", + "content": "20.10\\%" + }, + { + "bbox": [ + 55, + 275, + 296, + 526 + ], + "type": "text", + "content": " for ViT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 551, + 135, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 135, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 135, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 571, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 680 + ], + "type": "text", + "content": "Continual learning [48, 62] enables intelligent models to continually acquire new knowledge and adapt to changing environments while maintaining previous capabilities. It has recently emerged as a promising direction to achieve ideal intelligent models, and some pioneer works have investigated the robustness of continual learning in various aspects [42, 53, 56]. Achieving robust continual learners represents a crucial advancement in enabling intelligent models to adapt to dynamic scenarios effectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 681, + 295, + 693 + ], + "type": "text", + "content": "However, when dealing with streaming non-i.i.d. data," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 326, + 247, + 544, + 369 + ], + "blocks": [ + { + "bbox": [ + 326, + 247, + 544, + 369 + ], + "lines": [ + { + "bbox": [ + 326, + 247, + 544, + 369 + ], + "spans": [ + { + "bbox": [ + 326, + 247, + 544, + 369 + ], + "type": "image", + "image_path": "1cc626a1108478708580c34b8405ebfe88e9f524f95937419d79f94d89cd1df6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 445 + ], + "type": "text", + "content": "Figure 1. A preliminary analysis of the combination of continual learning and adversarial training on Split-CIFAR10 with buffer size 200. We consider the final average accuracy (FAA) and forgetting of adversarial and clean samples in both class incremental and task incremental settings. The red numbers indicate the performance after adding our approach." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 459, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 542 + ], + "type": "text", + "content": "continual learners are usually vulnerable to adversarial examples [23]. Chen et al. [12] first attempt at preserving adversarial robustness in continual learning [29] by leveraging adversarial training [52, 57] and a large amount of unlabeled data. Despite the impressive results of Chen et al. [12], the inherent challenges for conducting adversarial robust continual learners are still under disclosure." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 700 + ], + "type": "text", + "content": "In this study, we investigate building adversarial robust continual learners without abundant unlabeled data. We select two representative approaches for continual learning and adversarial defenses: memory-based continual learning and adversarial training. Figure 1 shows the direct combination of memory-based continual learning (ER [10]) with an adversarial training approach (TRADES [57]). The results for adversarial data reconfirm that continual learners risk adversarial attacks [12, 20, 23, 24], and demonstrate a critical trade-off: while adversarial training can strengthen continual learners against attacks [12, 20, 23, 24], it simultaneously increases forgetting on clean data, thus compromising the core objective of continual learning." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Further experimental analysis in Sec. 3 reveals two key" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "562" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "challenges when combining adversarial training with continual learning: (1) Adversarial data increase negative gradients toward past classes compared with clean data, accelerating catastrophic forgetting in continual learning; (2) Insufficient historical data in continual learning leads to gradient obfuscation [4], compromising the effectiveness of adversarial training compared to joint training. These challenges emerge naturally from the direct integration of the two approaches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 296, + 371 + ], + "type": "text", + "content": "Based on the analysis, we propose an adversarial robust memory-based continual learner without a large amount of additional data, which consists of an anti-forgettable logit calibration (AFLC) module and a robustness-aware experience replay (RAER) strategy. The AFLC module designs different logit calibration strategies for past, current, and future classes, thus mitigating the negative gradients of adversarial samples for continual learning. By leveraging attacking difficulty, RAER selects data that are adversarial safety and maintains the diversity of data distribution, rather than storing data overfitting to the decision boundary [9]. Our approach can be combined with most memory-based continual learning as well as adversarial training algorithms, and we have demonstrated the effectiveness of our approach on the Split-CIFAR10/100 and more challenging dataset Split-Tiny-ImageNet across ResNet and ViT architectures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 372, + 266, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 372, + 266, + 382 + ], + "spans": [ + { + "bbox": [ + 67, + 372, + 266, + 382 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 384, + 295, + 563 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 384, + 295, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 295, + 431 + ], + "type": "text", + "content": "- We reveal the twin critical challenges to carrying out adversarial robust memory-based continual learning algorithms: accelerated forgetting for continual learning and gradient obfuscation for adversarial robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 432, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 295, + 491 + ], + "type": "text", + "content": "- We propose antiforgettable logit calibration to reduce negative influences from adversarial training through task-adaptive logit calibration, and robustness-aware experience replay by storing robust and distribution-representative past data to alleviate gradient obfuscation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 491, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 295, + 563 + ], + "type": "text", + "content": "- Experiments on Split-CIFAR10/100 and Split-TinyImageNet across ResNet and ViT demonstrate our effectiveness, in improving both clean accuracy and adversarial robustness of continual learning and mitigating forgetting in both class-incremental and task-incremental settings without additional data." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 574, + 147, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 147, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 147, + 586 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Continual learning. Continual learning [38, 48, 62] enables models to adapt to evolving data distributions while preserving past knowledge. To alleviate catastrophic forgetting, existing approaches [34, 39] can be categorized into three types: memory-based [7, 8, 10], regularization-based [25, 29, 30], and dynamic architecture [17, 41, 50, 55, 61]. Continual learning can be categorized into class incremental (class-IL) and task-incremental (task-IL) settings [33] based on the availability of task information during inference. In both settings, memory-based meth" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "ods have demonstrated superior performance without requiring model expansion [7, 60]. Researchers have explored continual learning robustness under various constraints [19, 53, 56], and saft risks about backdoor attacks [1, 22, 47] and privacy concerns [21]. Recent studies have revealed both the vulnerability of continual models to adversarial attacks and the potential of adversarial samples for mitigating forgetting [20, 23, 24, 27, 49]. Chen et al. [12] combined LwF [29] with adversarial training using additional data, while our work uniquely analyzes the core challenges in achieving adversarial robustness for continual learning without such requirements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 218, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 350 + ], + "type": "text", + "content": "Adversarial defense. The vulnerability of deep neural networks to adversarial examples [18, 46, 52, 59] has led to numerous defense methods [2, 5], with adversarial training emerging as the most effective approach [14, 15, 28, 31, 32, 36, 37, 57]. Recent research has extended to real-world scenarios, such as long-tailed distributions [54] and open-world settings [43]. Meanwhile some studies show the potential of continual algorithms in adapting to new attacks [13, 40]. Different from them, we specifically address adversarial robustness in class-incremental and task-incremental settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 529, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 529, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 529, + 382 + ], + "type": "text", + "content": "3. Baselines and Their Inherent Problems" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 400, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 400, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 400, + 403 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": "Memory-based continual learning. Following [6, 7], we consider continual learning across " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " distinct classification tasks. We pre-allocate classification heads for all tasks at initialization rather than expanding them incrementally. For the " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": "-th task (" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "1 \\leq t \\leq T" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": "), we denote its data distribution as " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "D_{t}" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": ", with input samples " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " and corresponding labels " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": ". A model " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\mathrm{M}_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " is optimized sequentially in the given task order. Let " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_{t}) \\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " denote the output logits where " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_{t}) = \\mathrm{M}_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " is the total number of classes. The softmax output is defined as " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_{t}) \\triangleq \\operatorname{softmax}(h_{\\theta}(x_{t}))" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": ". Given a logit vector " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_{t}) = [h_{\\theta}(x_{t})_{1}, \\dots, h_{\\theta}(x_{t})_{C}]" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_{t})_{i}" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": " represents the logit for the " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": "-th class, we partition the classes into three sets based on the current task " + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 412, + 555, + 579 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 593, + 449, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 593, + 449, + 606 + ], + "spans": [ + { + "bbox": [ + 353, + 593, + 449, + 606 + ], + "type": "interline_equation", + "content": "p = \\{1, 2, \\dots , (t - 1) b \\},", + "image_path": "89a0594882c4170349f4572eb61908eea143fa26147f253c56a0a0a6eef2170a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 609, + 553, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 609, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 356, + 609, + 553, + 621 + ], + "type": "interline_equation", + "content": "c = \\{(t - 1) b + 1, (t - 1) b + 2, \\dots , t b \\}, \\tag {1}", + "image_path": "a64b80116e2d6b20519b6da51177bc5c267af8a1c3b924c57927b660622c44ce.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 624, + 465, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 624, + 465, + 635 + ], + "spans": [ + { + "bbox": [ + 355, + 624, + 465, + 635 + ], + "type": "interline_equation", + "content": "u = \\{t b + 1, t b + 2, \\dots , C \\},", + "image_path": "54ff104d8879b1b32a2d5089456e6e33211f20f9bb920f76a457c2f0cdc6bc0e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": " is the number of classes per task. Accordingly, we divide " + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_t)" + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": " into three parts: " + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_t)_p" + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_t)_c" + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "inline_equation", + "content": "h_{\\theta}(x_t)_u" + }, + { + "bbox": [ + 313, + 651, + 553, + 698 + ], + "type": "text", + "content": ", representing logits for past, current, and future classes respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "Continual learner aims to correctly classify data of all" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "563" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 321 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 321 + ], + "type": "table", + "html": "
Buffer SizeMethodW/ ATClass Incremental SettingTask Incremental SettingCRD↓FRI↓RRD↓
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
-Joint (upper bound)X91.81±0.07-0.00±0.00-98.17±0.12-0.00±0.00-7.92--
✓79.33±0.36-50.93±1.28-94.80±0.30-74.63±1.13-
-SGD (lower bound)X19.66±0.0796.39±0.675.17±2.5833.31±5.1966.84±6.5037.42±7.394.44±3.7833.34±3.490.91-3.7341.10
✓19.34±0.0591.93±0.2015.65±0.1869.71±1.7765.33±5.9134.43±7.1627.71±1.8854.56±3.47
200ERX48.80±1.8460.00±2.250.27±0.1812.60±2.1392.89±0.265.00±0.260.01±0.0113.78±0.4114.5112.9131.70
✓28.18±0.6980.58±1.0517.86±0.2969.58±1.1584.49±0.6110.23±0.9844.30±1.0536.89±1.11
5120X83.34±1.3815.35±1.710.08±0.1113.65±2.3496.84±0.170.71±0.190.06±0.0421.83±8.6913.5312.1118.80
✓61.88±0.7437.72±0.7327.28±0.5641.66±1.2291.24±0.192.56±0.1256.59±0.8819.34±1.11
200DERX62.71±2.7638.75±4.110.01±0.007.42±2.1391.32±0.836.95±1.440.09±0.057.39±2.1817.7215.7334.80
✓35.79±2.8565.67±4.1616.44±0.2152.84±3.7182.79±0.2111.48±0.6639.52±0.3740.64±1.38
5120X83.99±0.9910.57±1.510.07±0.077.05±2.0395.67±0.131.63±0.130.05±0.057.39±0.4816.7415.6624.32
✓56.58±3.8740.40±4.0723.26±0.4445.06±1.4889.62±0.673.12±0.2353.66±2.0122.95±3.68
200DER++X65.55±2.8132.17±4.260.24±0.327.38±3.3792.11±0.558.85±2.980.22±0.3310.45±0.9414.478.2437.14
✓45.43±3.2846.67±5.0914.46±1.3535.09±2.7083.29±1.0710.83±1.7736.82±0.5934.09±2.23
5120X85.74±0.657.32±0.881.51±1.276.10±0.5396.13±0.531.14±0.350.82±0.526.06±0.4910.638.2122.18
✓68.67±1.1123.68±2.6626.95±0.2927.68±2.4891.95±0.331.20±0.9054.26±0.4116.87±1.72
200X-DERX60.96±1.0912.29±0.090.00±0.003.84±1.1894.33±0.632.49±0.170.13±0.032.59±0.6420.237.6524.02
✓34.04±0.9025.13±4.0316.82±0.9827.84±4.1680.8±0.664.96±1.6660.83±0.5610.99±0.81
5120X69.93±0.634.33±0.260.21±0.124.52±1.3996.94±0.020.29±0.040.13±0.022.04±0.4324.9618.5221.27
✓34.29±0.6538.92±0.0319.24±0.1038.79±0.4182.67±1.682.74±0.8364.11±0.128.14±0.98
", + "image_path": "45dc768fbc4142940637d0cadb7dda8f3c88608138c9030dd14446c1122f5d5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 324, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 555, + 380 + ], + "type": "text", + "content": "Table 1. We conduct a systematic experimental analysis combining adversarial training (specifically Vanilla AT [31]) with memory-based continual learning algorithms. See Sec. 3.2 for a detailed summary of observations. Performance is evaluated using several metrics (detailed in Sec. 5.1): Final Average Accuracy (FAA), Forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase on clean data (FRI), and Robust Relative Decrease (RRD) compared to joint training. All adversarial samples are generated using PGD-20. Note that the reported CRD, FRI, and RRD values are averaged across both task-incremental and class-incremental settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 384, + 274, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 274, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 274, + 396 + ], + "type": "text", + "content": "observed tasks, and its objective can be formalized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 401, + 294, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 401, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 58, + 401, + 294, + 443 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\sum_ {t = 1} ^ {T} \\mathcal {L} _ {t}, \\text {w h e r e} \\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {\\left(x _ {t}, y _ {t}\\right) \\sim D _ {t}} \\left[ \\ell \\left(f _ {\\theta} \\left(x _ {t}\\right), y _ {t}\\right) \\right], \\tag {2}", + "image_path": "b3c9a28791b4bba73e03a010f0c9d3624d414e0410a206140ff906e27a181e36.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": " is the task-specific loss. The unavailability of past task data poses great challenges for continual learning. However, direct sequential optimization can lead to severe catastrophic forgetting. To address the above challenge, memory-based continual learners usually store a few past task data and replay them during the current task training phase [7], and become leading-edge approaches. When training the " + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 445, + 295, + 552 + ], + "type": "text", + "content": "-th task, its loss function can be formalized as the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 108, + 558, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 558, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 108, + 558, + 295, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} \\triangleq \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} [ \\ell (f _ {\\theta} (x), y) ], \\tag {3}", + "image_path": "f41f4a19a3ddc8297eb180d80913f84de879bd5fad72d518cfc6d66ccbfaeefb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "text", + "content": " is the episodic memory of the current task, which preserves a small subset of observed past data " + }, + { + "bbox": [ + 55, + 578, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\langle x_{\\mathcal{M}},y_{\\mathcal{M}}\\rangle" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 649 + ], + "type": "text", + "content": "Adversarial training. Adversarial training is currently the most reliable way [5] to improve the adversarial robustness of a model. Following [31], it can be formally defined as a min-max optimization problem:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 655, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 655, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 86, + 655, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right), \\tag {4}", + "image_path": "fb4b794491b94e11fc5a17d1a40b3296eeeeefba78c9481b2f512de3d360ba3a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": " is an adversarial perturbation, " + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": " is data distribution, and " + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": " is a perturbation set. " + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "inline_equation", + "content": "S = \\left\\{\\delta \\mid \\| \\delta \\|_p \\leq \\epsilon \\right\\}" + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 685, + 296, + 716 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 407 + ], + "type": "text", + "content": "the perturbation size. The goal of adversarial training is to correctly classify all adversarial examples " + }, + { + "bbox": [ + 313, + 384, + 553, + 407 + ], + "type": "inline_equation", + "content": "(\\widetilde{x} = x + \\delta)" + }, + { + "bbox": [ + 313, + 384, + 553, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "text", + "content": "Here we integrate memory-based continual learner and adversarial training as adversarial robust memory-based continual learner baselines, and the new optimization objective during the training stage of " + }, + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 408, + 554, + 456 + ], + "type": "text", + "content": "-th task is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 330, + 461, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 461, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 330, + 461, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, y) \\sim D _ {t} \\cup \\mathcal {M}} \\left(\\underset {\\delta \\in S} {\\max } \\ell \\left(f _ {\\theta} (x + \\delta), y\\right)\\right). \\tag {5}", + "image_path": "0b24425d1bd73c7c5051344ec13bfcc5afe90692a7d60f4329724bd8f752c5f5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 492, + 416, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 416, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 416, + 504 + ], + "type": "text", + "content": "3.2. Vanilla Solutions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 510, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 676 + ], + "type": "text", + "content": "To investigate the challenges in adversarial robust continual learning, we analyze several popular memory-based continual learning methods (ER [10], DER [7], DER++ [7], and X-DER [6]) on Split-CIFAR10 [26] under both class and task incremental settings. We employ Vanilla AT [31] for adversarial training due to its classic and strong performance [36]. Performance is evaluated against an upper bound (\"Joint\") and a lower bound (\"SGD\") using multiple metrics: Final Average Accuracy (FAA), forgetting, Clean Relative Decrease (CRD), Forgetting Relative Increase in clean data (FRI), and Robust Relative Decrease (RRD). Detailed metric definitions are provided in Sec. 5.1, and additional implementation details and hyper-parameters are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 712 + ], + "type": "text", + "content": "Results in Table 1 reveal that direct integration of adversarial training with continual learning is suboptimal. First, the results validate our observations from Sec. 1: while" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "564" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 144 + ], + "type": "text", + "content": "adversarial training enhances robustness, it compromises clean performance (positive CRDs) and accelerates forgetting (positive FRIs). Larger memory sizes consistently improve performance across all metrics. Moreover, continual models show inferior robustness improvement compared to joint models (positive RRDs)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 216 + ], + "type": "text", + "content": "To investigate the underlying mechanisms, we select ER as our base model for further analysis. While DER/DER++ and X-DER demonstrate better stability with lower forgetting, they struggle with new task learning, resulting in poor robust FAA. In contrast, ER combined with adversarial training exhibits strong robustness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 187, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 187, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 187, + 232 + ], + "type": "text", + "content": "3.3. The Inherent Problems" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 285 + ], + "type": "text", + "content": "By analyzing gradients during training, we identify two critical challenges in robust continual learning: accelerated forgetting in continual learning and gradient obfuscation in adversarial training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "text", + "content": "Accelerated forgetting in continual learning. Adversarial training hastens the forgetting of clean data in continual learners. According to [31], adversarial examples are designed to mislead model predictions by increasing probabilities for incorrect classes while suppressing those of the true class. In the continual learning setting, adversarial examples from the current task tend to increase probabilities for past classes while decreasing probabilities for their true class. Specifically, for an adversarial example " + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_t" + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "text", + "content": " generated from a clean sample " + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "text", + "content": " of current task " + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 287, + 295, + 406 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 411, + 294, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 411, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 67, + 411, + 294, + 425 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {p} > f _ {\\theta} \\left(x _ {t}\\right) _ {p}, f _ {\\theta} \\left(\\widetilde {x} _ {t}\\right) _ {\\text {t r u e c l a s s}} < f _ {\\theta} \\left(x _ {t}\\right) _ {\\text {t r u e c l a s s}}. \\tag {6}", + "image_path": "16f63b6f1e9c3e00af98ecf5cc30c006e2de40817cd75ee9fd4d6e7ee6179aad.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 429, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 464 + ], + "type": "text", + "content": "When using cross-entropy as the loss function, adversarial data amplify negative gradients for current data " + }, + { + "bbox": [ + 55, + 429, + 295, + 464 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 429, + 295, + 464 + ], + "type": "text", + "content": " for past tasks:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 466, + 294, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 466, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 66, + 466, + 294, + 495 + ], + "type": "interline_equation", + "content": "\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (\\widetilde {x} _ {t})} \\right] _ {p} = f _ {\\theta} (\\widetilde {x} _ {t}) _ {p} > f _ {\\theta} (x _ {t}) _ {p} = \\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} (x _ {t})} \\right] _ {p}. \\tag {7}", + "image_path": "f7beb243acd96c8bc48c9cec92c67190b87a4c6dcc769b8971859e906b547ac4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "Our empirical analysis reveals that these adversarial examples generate stronger negative gradients toward past tasks, especially during the initial training phases. In Figure 2, experiments with " + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + AT}" + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": " on Split-CIFAR10 using a buffer size of 200 show adversarial data consistently produces larger gradient norms compared to clean data. The larger gradient norms indicate stronger positive gradients for the true class and stronger negative gradients for other classes, including past classes, thus accelerating forgetting of previously learned tasks. While memory-based methods using stored adversarial samples can help mitigate catastrophic forgetting, the limited buffer size (typically small in practical applications) means that adversarial training on the entire dataset still accelerates forgetting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Gradient obfuscation in adversarial training. Gradient obfuscation [4] occurs when gradient information is obscured by non-differentiable operators or randomness, leading to inferior robustness. We observe a specific form of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 553, + 156 + ], + "type": "image", + "image_path": "388c31bd51e2ed97f3e1bfe631bb4f02b505d540bcaaca5f035f2dccd035589d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 175, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 313, + 175, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 273 + ], + "type": "image", + "image_path": "bac76602735a5d52be05b919cc8d47fc54f62544c4858256ea4b9df9e3b2ea2c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 337, + 315, + 525, + 397 + ], + "blocks": [ + { + "bbox": [ + 313, + 280, + 553, + 313 + ], + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 313 + ], + "type": "text", + "content": "Figure 2. Gradient norms of adversarial data and clean data during the training stage. We experiment ER+AT with a 200 buffer size on the Split-CIFAR10 dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 315, + 525, + 397 + ], + "lines": [ + { + "bbox": [ + 337, + 315, + 525, + 397 + ], + "spans": [ + { + "bbox": [ + 337, + 315, + 525, + 397 + ], + "type": "image", + "image_path": "b44940afc77cdc7474f845a958bf459d85850c453c74c6be9c157ded6c23237b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 554, + 469 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 469 + ], + "type": "text", + "content": "Figure 3. The gradient cosine similarities between robust joint model and robust continual model, which indicates a special gradient obfuscation only in continual learning. We experiment with buffer size=200/5,120 on the Split-CIFAR10 dataset. Gradient obfuscation weakens as the buffer size gets larger, and our method also mitigates gradient obfuscation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 510 + ], + "type": "text", + "content": "gradient obfuscation, termed \"shattered gradients\", in adversarial training under continual learning, resulting in less effective robustness compared to joint models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "type": "text", + "content": "In continual learning, gradient information is biased due to limited replay data. With few old task samples, the model tends to overfit, causing direction-biased gradients when generating adversarial examples. Our analysis of gradient directions between adversarial robust joint and continual models reveals significant differences, contributing to lower adversarial robustness in continual learners (see Figure 3)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 604, + 370, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 370, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 370, + 616 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 382, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 382, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 382, + 635 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "To address the identified challenges, we propose a novel framework consisting of two key components: antiforgettable logit calibration (AFLC) to mitigate forgetting acceleration, and robustness-aware experience replay (RAER) to reduce gradient obfuscation. The full pipeline is illustrated in Figure 4: during training each task, the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "565" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 411, + 243 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 411, + 243 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 411, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 411, + 243 + ], + "type": "image", + "image_path": "a493aef0c46f4adcd6adc6b6bd32607dc81091936c0bf940dab7ef0a1615dbbe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": "Figure 4. The pipeline of our adversarial robust memory-based continual learner. First, generate adversarial samples " + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "(\\widetilde{X}_t,\\widetilde{X}_{\\mathcal{M}})" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": " and robust difficulty factors " + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "(K_{t})" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": " for both current task data " + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "(X_{t},Y_{t})" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": " and memory data " + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "(X_{\\mathcal{M}},Y_{\\mathcal{M}})" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": ". The model then processes these samples through AFLC (Sec. 4.2), which calibrates logits based on task order to reduce negative gradients toward past tasks. Meanwhile, RAER (Sec. 4.3) leverages the robust difficulty factors " + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 55, + 247, + 555, + 293 + ], + "type": "text", + "content": " to select representative samples for memory storage, mitigating gradient obfuscation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 414, + 71, + 549, + 242 + ], + "blocks": [ + { + "bbox": [ + 414, + 71, + 549, + 242 + ], + "lines": [ + { + "bbox": [ + 414, + 71, + 549, + 242 + ], + "spans": [ + { + "bbox": [ + 414, + 71, + 549, + 242 + ], + "type": "image", + "image_path": "0b7d85f31fde0b9ee1c57faeae9cef9a37694110bbc4e526a4700c20d5926981.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "text", + "content": "model processes both current task data and memory samples through the model. AFLC then calibrates logits based on task order (past/current/future) before loss computation. Simultaneously, RAER evaluates current task samples using a robustness difficulty factor " + }, + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 302, + 296, + 374 + ], + "type": "text", + "content": " to select representative data for memory storage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 241, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 241, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 241, + 393 + ], + "type": "text", + "content": "4.2. Anti-Forgettable Logit Calibration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 398, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 445 + ], + "type": "text", + "content": "We introduce AFLC to regulate the influence of adversarial examples on past task knowledge. The core idea is to calibrate logits differently for past and current tasks to control gradient flow." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "text", + "content": "Given an input " + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "text", + "content": " (either from current task " + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "text", + "content": " or memory buffer " + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "inline_equation", + "content": "x_{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 446, + 296, + 471 + ], + "type": "text", + "content": "), AFLC applies a task-dependent calibration:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 476, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 476, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 110, + 476, + 294, + 491 + ], + "type": "interline_equation", + "content": "h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = h _ {\\theta} (x) _ {i} - v _ {i}, v _ {p} > v _ {c}, \\tag {8}", + "image_path": "e5b67c02f1ab2c9f3689ab5cfba7ffe6857f86bcfceae5d1e72ed9b4acbd5671.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 266, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 266, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 266, + 509 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 266, + 509 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 497, + 266, + 509 + ], + "type": "text", + "content": " represents the calibration factor for class i:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 514, + 295, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 514, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 116, + 514, + 295, + 541 + ], + "type": "interline_equation", + "content": "v _ {i} = - \\log \\left(\\frac {n _ {i}}{\\sum_ {j = 1} ^ {C} n _ {j}}\\right) - \\alpha_ {i}, \\tag {9}", + "image_path": "ac45af9d80bb5924f9d6cbbef7d44eb33ed11f750caa64a5d9c188660624170b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "text", + "content": " denotes sample count for class " + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 55, + 547, + 296, + 643 + ], + "type": "text", + "content": " is a hyperparameter. When all tasks share the classification head, we realize that making logit changes to past and current classes only is equivalent to negative gradient increases. So we make a Further Prior (FP) adjustment that essentially averages the adjustments applied to current and past classes and applies this value to future classes, helping to balance the gradients across all periods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 136, + 650, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 650, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 136, + 650, + 295, + 682 + ], + "type": "interline_equation", + "content": "v _ {i} = \\frac {1}{t \\cdot b} \\sum_ {j = 1} ^ {t \\cdot b} \\left(v _ {j}\\right), \\tag {10}", + "image_path": "eced215913496dd4a3cf14790ce33e9e3983424907c5aa74de6b3e914fcb5f27.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the current task order and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the number of classes per task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 302, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 302, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 326, + 302, + 545, + 315 + ], + "type": "text", + "content": "The calibrated probability distribution is computed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 323, + 553, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 323, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 369, + 323, + 553, + 353 + ], + "type": "interline_equation", + "content": "f _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i} = \\frac {\\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {i}\\right)}{\\sum_ {j = 1} ^ {C} \\exp \\left(h _ {\\theta} ^ {\\mathrm {l c}} (x) _ {j}\\right)}, \\tag {11}", + "image_path": "59259db6203fc17b0b332322f6a967e92a62d21f0740a01ba9980fe760aeb0f1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 547, + 373 + ], + "type": "text", + "content": "and leads to two key properties: For current task samples," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 380, + 553, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 380, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 362, + 380, + 553, + 410 + ], + "type": "interline_equation", + "content": "\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x})} \\right] _ {p} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {t}) _ {p} < f _ {\\theta} (\\widetilde {x} _ {t}) _ {p}, \\tag {12}", + "image_path": "ac3d558013abacfc1f3fe3454e75ea41124076b310da63c301cd9dc74999c128.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 418, + 444, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 444, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 444, + 430 + ], + "type": "text", + "content": "and for memory buffer samples," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 352, + 437, + 553, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 437, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 352, + 437, + 553, + 466 + ], + "type": "interline_equation", + "content": "\\left[ \\frac {\\partial \\ell}{\\partial h _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x _ {\\mathcal {M}}})} \\right] _ {c} = f _ {\\theta} ^ {\\mathrm {l c}} (\\widetilde {x} _ {\\mathcal {M}}) _ {c} > f _ {\\theta} (\\widetilde {x} _ {\\mathcal {M}}) _ {c}. \\tag {13}", + "image_path": "8b2b992589d017038241658d10a667f08728bf856a3164780852080f51832572.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 521 + ], + "type": "text", + "content": "In that way, AFLC can mitigate forgetting acceleration by reducing the negative gradient to past classes from current adversarial data and improving the negative gradient to current classes from memory adversarial data." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": "Comparisons with other logit calibration CL. X-DER [6] also adjusts the negative gradients of the new task data affecting the past tasks by updating the new task with the logit masking. Logit masking can be seen as a special case of our method at the softmax layer, which can be viewed as " + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_p = \\mathrm{v}_c = 0" + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": " for past data, and " + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_p = +\\infty" + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{v}_c = 0" + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": " for current data. However, their extreme margin " + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 521, + 555, + 640 + ], + "type": "text", + "content": " inhibits the model's ability to learn new knowledge. X-DER further adds a safeguard threshold to limit the activations of past and future heads, which also can be combined with us." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 648, + 514, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 514, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 514, + 661 + ], + "type": "text", + "content": "4.3. Robustness-Aware Experience Replay" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "To address gradient obfuscation, we propose RAER, which selects samples based on their robustness characteristics. For each sample " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": ", we measure its robustness difficulty factor " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " during PGD-10 attack, without additional forward or" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "566" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "backward calculation, where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "k \\in [0,10] \\cap \\mathbb{N}" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " indicates the number of successful attacks. Samples meeting the threshold condition " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "k < \\rho" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " are candidates for memory storage. The memory update process can be formalized as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 131, + 296, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 131, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 91, + 131, + 296, + 145 + ], + "type": "interline_equation", + "content": "\\mathcal {M} \\leftarrow \\text {R a n d o m S a m p l e} \\left(\\{x \\mid k _ {x} < \\rho \\}, m\\right), \\tag {14}", + "image_path": "4c510ca5791fe7eb1dddef0b08bd03359735a841ea5de4feb5a67774b3ab8714.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "text", + "content": " is the memory size and " + }, + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 155, + 284, + 167 + ], + "type": "text", + "content": " is set to 5 in practice." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "content": "Why RAER works. The robustness difficulty factor " + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "content": " shows how easily a sample can be attacked: a high " + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "content": " suggests vulnerability and proximity to decision boundaries, while a low " + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "content": " implies robustness and distance from boundaries. We selectively store samples with " + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "inline_equation", + "content": "k < \\rho" + }, + { + "bbox": [ + 55, + 167, + 296, + 275 + ], + "type": "text", + "content": " in memory, excluding boundary-overfitting outliers. Unlike prior continual learning [9] and adversarial training [58] methods, RAER emphasizes distribution-representative and robust learnable samples while maintaining model reliability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 276, + 296, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 322 + ], + "type": "text", + "content": "The computational overhead of RAER is minimal, as " + }, + { + "bbox": [ + 55, + 276, + 296, + 322 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 276, + 296, + 322 + ], + "type": "text", + "content": " is calculated during the existing adversarial example generation process, without additional forward or backward calculation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 324, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 443 + ], + "type": "text", + "content": "Comparisons with other data-selection CL. Existing works [3, 44] alleviate catastrophic forgetting using data selection and focus solely on clean samples. RAER uniquely targets gradient obfuscation in continual robust settings by: Selecting adversarially safe samples for better retention; excluding high-" + }, + { + "bbox": [ + 55, + 324, + 296, + 443 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 324, + 296, + 443 + ], + "type": "text", + "content": " boundary samples; and maintaining representative data distribution. As demonstrated in Sec. 5, RAER outperforms traditional data selection strategies combined with adversarial training, particularly in adversarial robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 162, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 162, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 162, + 466 + ], + "type": "text", + "content": "4.4. Overall Objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": "To sum up, we design AFLC to reduce negative gradients of the current adversarial sample to past tasks, and RAER to select data that alleviate gradient obfuscation to store. We set " + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": " of AFLC is 3.5 in practice and " + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "inline_equation", + "content": "x_{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "inline_equation", + "content": "y_{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": " are selected by RAER with " + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\rho = 5" + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + AT}" + }, + { + "bbox": [ + 55, + 472, + 296, + 543 + ], + "type": "text", + "content": " combines with ours, the loss of task t is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 553, + 295, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 553, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 69, + 553, + 295, + 569 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} \\triangleq \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {t}\\right), y _ {t}\\right) + \\mathrm {C E} \\left(f _ {\\theta} ^ {l c} \\left(\\tilde {x} _ {\\mathcal {M}}\\right), y _ {\\mathcal {M}}\\right). \\tag {15}", + "image_path": "68dd53c3c8971166c1788e9040e7b4a951c3cf4117b888c881cfa21771319bb3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 137, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 137, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 137, + 593 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 599, + 182, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 182, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 182, + 613 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Datasets and training details. Following standard practices in adversarial training and continual learning [6, 37, 63], we evaluate our approach on Split-CIFAR10 and Split-CIFAR100 [26] and the model architecture is ResNet18, trained using SGD optimizer with learning rate 0.1. Split-CIFAR10 comprises ten classes (5,000 training, 1,000 test samples per class) divided into five binary classification tasks. Split-CIFAR100 contains 100 classes (500 training," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "100 test samples per class) split into ten-way classification tasks. Evaluations on Split-Tiny-ImageNet [45] and ViT [16] in supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 108, + 554, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 179 + ], + "type": "text", + "content": "For adversarial training, we use perturbations of 8/255 and step size of 2/255. Data augmentation includes random cropping (4-pixel padding) and horizontal flipping for all data. Unlike Chen et al. [12], we achieve robustness without additional unlabeled data1. All experiments are conducted with three different random seeds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate both adversarial robustness and continual learning performance using Final Average Accuracy (FAA) and forgetting metrics on clean and adversarial samples. For adversarial evaluation, we employ PGD-20 (white-box attack) and Auto Attack (AA) [14], which combines black-box and white-box attacks for reliable robustness assessment [36, 37, 51]. Extended robustness verification using additional black-box (RayS [11]) and adaptive attacks in the supplementary material. Let " + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "inline_equation", + "content": "a_{i}^{t}" + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": " denote the accuracy for task " + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": " after training on task " + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": ". The metrics are defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 394, + 319, + 555, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 319, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 394, + 319, + 555, + 353 + ], + "type": "interline_equation", + "content": "\\mathrm {F A A} \\triangleq \\frac {1}{T} \\sum_ {i = 1} ^ {T} a _ {i} ^ {T}, \\tag {16}", + "image_path": "519c7e9191014fccac83e460c7523182135762ab33436fd34fd1c77a32f7e2e6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 371, + 558, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 558, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 558, + 415 + ], + "type": "interline_equation", + "content": "\\text {f o r g e t t i n g} \\triangleq \\frac {1}{T - 1} \\sum_ {j = 1} ^ {T - 1} f _ {j}, \\text {s . t .} f _ {j} = \\max _ {l \\in \\{1, \\dots , T - 1 \\}} a _ {i} ^ {l} - a _ {j} ^ {T}. \\tag {17}", + "image_path": "550a02fb7a4e18d270674208cb6e238dd6c9f410fe6b93f8e4b43a89f89dd0f9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 554, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 439 + ], + "type": "text", + "content": "Forgetting, ranging from " + }, + { + "bbox": [ + 313, + 415, + 554, + 439 + ], + "type": "inline_equation", + "content": "[-100, 100]" + }, + { + "bbox": [ + 313, + 415, + 554, + 439 + ], + "type": "text", + "content": ", measures the average decrease in accuracy across tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 440, + 454, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 440, + 454, + 451 + ], + "spans": [ + { + "bbox": [ + 326, + 440, + 454, + 451 + ], + "type": "text", + "content": "For analysis, we further define:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 459, + 555, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 459, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 367, + 459, + 555, + 474 + ], + "type": "interline_equation", + "content": "\\mathrm {C R D} \\triangleq \\mathrm {F A A} _ {\\text {c l e a n}} - \\widetilde {\\mathrm {F A A}} _ {\\text {c l e a n}}, \\tag {18}", + "image_path": "183785fc03e078573207014520289208ae6e92edeff0b72195b679a3cb59513b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 337, + 492, + 555, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 492, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 337, + 492, + 555, + 509 + ], + "type": "interline_equation", + "content": "\\mathrm {F R I} \\triangleq \\widetilde {\\mathrm {F o r g e t t i n g}} _ {\\mathrm {c l e a n}} - \\mathrm {F o r g e t t i n g} _ {\\mathrm {c l e a n}}, \\tag {19}", + "image_path": "4a5632d27f25571e98aa4d38fb6f171df6ecf030ced0302e917cab4e1d0c42be.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 529, + 553, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 553, + 557 + ], + "type": "interline_equation", + "content": "\\mathrm {R R D} \\triangleq \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}} - \\mathrm {F A A} _ {\\mathrm {a d v}} ^ {\\text {J o i n t}}\\right) - \\left(\\widetilde {\\mathrm {F A A}} _ {\\mathrm {a d v}} - \\mathrm {F A A} _ {\\mathrm {a d v}}\\right), \\tag {20}", + "image_path": "6df2cee42cfec4940708bd7612bcdba92f041e5b387b0f8753d36c13df5f2b48.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 557, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 594 + ], + "type": "text", + "content": "where \"clean\" and \"adv\" denote metrics on clean and adversarial data, respectively, tilde " + }, + { + "bbox": [ + 313, + 557, + 553, + 594 + ], + "type": "inline_equation", + "content": "(\\sim)" + }, + { + "bbox": [ + 313, + 557, + 553, + 594 + ], + "type": "text", + "content": " indicates robust models, and \"Joint\" refers to a joint training model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 600, + 400, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 400, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 400, + 612 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "We evaluate our method against various adversarial training and memory-based continual learning combinations on Split-CIFAR10/100. Evaluations on Split-TinyImageNet [45] and ViT [16] in supplementary materials. Due to computational constraints, we focus on integrating our approach with ER+AT and ER+TRADES (" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 6.0" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "), which showed superior adversarial FAA performance among baselines." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "567" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 223 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 223 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 223 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 223 + ], + "type": "table", + "html": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
200ER+AT28.1880.5817.8616.9469.5884.4910.2344.3044.6936.89
ER+TRADES22.4277.2515.7215.5364.9578.798.1051.3351.5021.97
ER+FAT33.6169.2115.1414.8149.0483.4010.3543.6943.9628.56
ER+LBGAT25.6884.4716.6516.5670.5078.1918.8540.6940.7340.83
ER+SCORE48.6556.792.400.9318.9688.909.827.256.728.70
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)18.16(↑ 1.22)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)47.61(↑ 2.92)34.04(↓ 2.85)
ER+TRADES+Ours43.34(↑ 20.92)33.40(↓ 43.85)19.85(↑ 4.13)18.35(↑ 2.82)30.78(↓ 34.17)82.59(↑ 3.80)7.53(↓ 0.57)59.41(↑ 8.08)59.59(↑ 8.09)14.16(↓ 7.81)
5,120ER+AT61.8837.7227.2826.6941.6691.242.5656.5956.9019.34
ER+TRADES20.3685.1416.3016.1872.8588.481.5964.3664.5212.47
ER+FAT54.5543.1819.6818.9142.1591.502.1256.7256.8714.79
ER+LBGAT62.4537.7327.4226.6647.8391.103.2556.5756.3119.38
ER+SCORE51.3752.843.141.1020.2995.631.5814.6613.922.53
ER+AT+Ours64.34(↑ 2.46)23.64(↓ 14.08)31.31(↑ 4.03)30.49(↑ 3.80)20.46(↓ 21.20)91.00(↓ 0.24)3.51(↑ 0.95)60.49(↑ 3.90)60.61(↑ 3.71)13.27(↓ 6.07)
ER+TRADES+Ours39.80(↑ 19.44)44.08(↓ 41.06)23.07(↑ 6.77)21.98(↑ 5.80)41.37(↓ 31.48)86.48(↓ 2.00)1.71(↑ 0.12)69.25(↑ 4.89)69.38(↑ 4.86)5.33(↓ 7.14)
", + "image_path": "8f85153e81efdb0f268e1ad4f2e3abf85e34197e53ebf346dfec2050a61ba2e9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 236, + 555, + 392 + ], + "blocks": [ + { + "bbox": [ + 182, + 224, + 428, + 234 + ], + "lines": [ + { + "bbox": [ + 182, + 224, + 428, + 234 + ], + "spans": [ + { + "bbox": [ + 182, + 224, + 428, + 234 + ], + "type": "text", + "content": "(a) Results on Split-CIFAR10. We chose two buffer sizes of 200 and 5, 120." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 236, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 59, + 236, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 59, + 236, + 555, + 392 + ], + "type": "table", + "html": "
Buffer Size MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓FAA↑Forgetting↓PGD-20↑AA↑Forgetting↓
500ER+AT11.9473.545.665.5438.0352.7128.4417.3519.5625.67
ER+TRADES7.5968.135.435.1144.1350.7520.2226.8927.6920.34
ER+FAT11.4873.995.355.2337.2656.1024.7120.0122.9922.31
ER+LBGAT11.7775.106.165.8239.7742.0226.4513.9914.4323.14
ER+SCORE16.2277.710.910.626.1368.8711.213.237.954.94
ER+AT+Ours24.14(↑ 12.20)53.03(↓ 20.51)7.13(↑ 1.47)6.68(↑ 1.14)25.81(↓ 12.22)56.00(↑ 3.29)26.19(↓ 2.25)17.95(↑ 0.60)20.26(↑ 0.70)24.87(↓ 0.80)
ER+TRADES+Ours23.24(↑ 15.65)24.29(↓ 43.84)9.93(↑ 4.50)7.50(↑ 2.39)11.70(↓ 32.43)56.68(↑ 5.93)21.39(↑ 1.17)27.39(↑ 0.50)28.50(↑ 0.81)16.76(↓ 3.58)
2,000ER+AT18.7765.067.207.0133.5662.0117.9721.1624.0420.47
ER+TRADES9.5070.495.355.0142.0860.6313.7826.6829.1918.60
ER+FAT17.0966.916.125.9333.6263.8815.9923.4826.7517.12
ER+LBGAT20.5863.317.937.0530.0955.7725.1117.9518.8819.18
ER+SCORE30.1061.510.750.514.1076.9019.604.8811.975.25
ER+AT+Ours31.93(↑ 13.16)40.50(↓ 24.56)9.61(↑ 2.41)9.16(↑ 2.15)17.78(↓ 15.78)63.77(↑ 1.76)17.16(↓ 0.81)23.11(↑ 1.95)25.59(↑ 1.55)17.57(↓ 2.90)
ER+TRADES+Ours28.73(↑ 19.23)24.16(↓ 46.33)12.75(↑ 7.40)11.02(↑ 6.01)14.56(↓ 27.52)62.01(↑ 1.38)13.16(↓ 0.62)34.81(↑ 8.13)35.36(↑ 6.17)11.60(↓ 7.00)
", + "image_path": "2945f76b4b5c521621958229774591b5a00e3a0f0bfdfbd3bcc5113b501e9a3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 457, + 553, + 599 + ], + "blocks": [ + { + "bbox": [ + 176, + 392, + 432, + 402 + ], + "lines": [ + { + "bbox": [ + 176, + 392, + 432, + 402 + ], + "spans": [ + { + "bbox": [ + 176, + 392, + 432, + 402 + ], + "type": "text", + "content": "(b) Results on Split-CIFAR100. We choose two buffer sizes of 500 and 2,000." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 404, + 555, + 459 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 459 + ], + "type": "text", + "content": "Table 2. Experiment results on Split-CIFAR10/100 datasets and model architecture is ResNet18. Here, PGD-20 and AA are adversarial data Final Average Accuracy (FAA) generated by PGD-20 and Auto Attack (AA), respectively. Forgetting of adversarial data is computed based on PGD-20. Bold represents the best experimental results for the same settings, underline indicates the second-best result, green signifies relative improvement and red indicates relative degradation. By applying our methods, various model performances can be improved across the board." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 457, + 553, + 599 + ], + "lines": [ + { + "bbox": [ + 59, + 457, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 59, + 457, + 553, + 599 + ], + "type": "table", + "html": "
MethodClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
ER+AT28.1880.5817.8669.5884.4910.2344.3036.89
ER+AT+Ours35.68(↑ 7.50)71.18(↓ 9.40)18.40(↑ 0.54)67.85(↓ 1.73)84.87(↑ 0.38)9.93(↓ 0.30)47.30(↑ 3.00)34.04(↓ 2.85)
GSS+AT27.5980.7816.6768.5384.419.8344.2534.79
GSS+AT+Ours36.93(↑ 9.34)67.72(↓ 13.06)16.84(↑ 0.17)60.04(↓ 8.49)85.57(↑ 1.16)8.86(↓ 0.97)47.11(↑ 2.86)31.83(↓ 2.97)
ASER+AT18.8587.7814.0665.5773.8719.0130.6544.85
ASER+AT+Ours24.45(↑ 5.61)81.73(↓ 6.05)14.91(↑ 0.85)62.74(↓ 2.83)77.70(↑ 3.83)15.21(↓ 3.80)34.50(↑ 3.85)38.55(↓ 6.30)
X-DER+AT34.0425.1316.8227.8480.804.9660.8310.99
X-DER+AT+Ours43.25(↑ 9.21)20.77(↓ 4.36)17.22(↑ 0.41)18.56(↓ 9.28)84.87(↑ 4.07)1.74(↓ 3.21)61.68(↑ 0.85)7.03(↓ 3.97)
", + "image_path": "e4d3881bf6f5a1964230b2ca06aa53536eba6cca815a0e733889244dfed6f3f7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 599, + 497, + 610 + ], + "lines": [ + { + "bbox": [ + 111, + 599, + 497, + 610 + ], + "spans": [ + { + "bbox": [ + 111, + 599, + 497, + 610 + ], + "type": "text", + "content": "Table 3. Experiments with other data selection-based and logit masking-based continual learning methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "text", + "content": "Results on different adversarial training methods. Table 2 reveals several key findings. Buffer size significantly impacts performance, with larger buffers generally improving results, except for " + }, + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + TRADES}" + }, + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "text", + "content": " on Split-CIFAR10 at buffer size 5, 120. Regarding robustness, " + }, + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + AT}" + }, + { + "bbox": [ + 55, + 623, + 294, + 683 + ], + "type": "text", + "content": " demon" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 623, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 696 + ], + "type": "text", + "content": "strates superior performance in class incremental settings, while ER+TRADES achieves optimal results under task incremental learning, albeit with reduced clean sample accuracy. ER+LBGAT exhibits balanced performance in class incremental settings, while ER+FAT and ER+SCORE excel on clean samples but underperform on adversarial ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Our method significantly enhances both ER+TRADES" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1The 80M-TinyImage dataset used in [12] has been withdrawn due to privacy concerns" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "568" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 70, + 528, + 156 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 528, + 156 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 528, + 156 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 528, + 156 + ], + "type": "table", + "html": "
AFLCRAERFPClass Incremental SettingTask Incremental Setting
Clean DataAdversarial DataClean DataAdversarial Data
FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓FAA↑Forgetting↓
22.4277.2515.7264.9578.798.1051.3321.97
✓37.8821.7618.5821.7781.6610.8653.6417.60
✓✓37.4510.2919.819.6178.1315.3955.7114.13
✓✓✓43.3433.4019.8530.7882.597.5359.4114.16
", + "image_path": "f235e515984687e31c55a1108d89b2b142c94142c235cd2b83d4d6084e779850.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 555, + 201 + ], + "type": "text", + "content": "Table 4. Ablation experiments on the Split-CIFAR10 dataset with the buffer size of 200, using ER+TRADES as the baseline. Bold represents the best experimental results for the same settings. Experiments have demonstrated that AFLC mitigates clean-sample accelerated forgetting from adversarial samples, RAER mitigates gradient obfuscation (Adversarial FAA has a boost) and robust forgetting; adding FP improves the model's ability to learn new tasks (FAA has an overall increase)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": "and ER+AT across Split-CIFAR10/100 datasets, particularly in class incremental settings. For Split-CIFAR10, we achieve FAA improvements of " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{clean} = 20.92\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{adv} = 8.09\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": ", with forgetting reductions of " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{f_{clean}} = 43.85\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{f_{adv}} = 34.17\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": ". On Split-CIFAR100, improvements reach " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{clean} = 19.23\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{adv} = 8.13\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": " for FAA, with forgetting reductions of " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{f_{clean}} = 46.33\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "inline_equation", + "content": "\\Delta_{f_{adv}} = 32.43\\%" + }, + { + "bbox": [ + 55, + 208, + 294, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 295, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 370 + ], + "type": "text", + "content": "Our approach demonstrates positive improvements in " + }, + { + "bbox": [ + 55, + 309, + 295, + 370 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 55, + 309, + 295, + 370 + ], + "type": "text", + "content": " of evaluation metrics (Table 2). Minor performance decreases in task incremental settings with buffer size 5, 120 reflect AFLC's diminishing returns and the known clean-robust accuracy trade-off [31, 35, 37, 57]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 460 + ], + "type": "text", + "content": "Results on different CL methods. We evaluate our method against state-of-the-art continual learning approaches on Split-CIFAR10 with buffer size 200, including gradient-based GSS, Shapley value-based ASER, and logit masking-based X-DER. These methods primarily address clean-data continual learning, whereas our approach specifically targets adversarial robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 536 + ], + "type": "text", + "content": "Table 3 demonstrates that existing data selection methods (GSS, ASER) show limited effectiveness in adversarial settings, with robustness performance not exceeding " + }, + { + "bbox": [ + 55, + 464, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + AT}" + }, + { + "bbox": [ + 55, + 464, + 295, + 536 + ], + "type": "text", + "content": ". Our method, when integrated with these approaches, enhances both robustness and clean accuracy across class and task incremental settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 625 + ], + "type": "text", + "content": "X-DER employs logit masking, a specialized case of our AFLC, for memory updates and future preparation. While effective at forgetting mitigation, its aggressive gradient suppression limits model learning capacity, resulting in lower robust FAA than " + }, + { + "bbox": [ + 55, + 540, + 295, + 625 + ], + "type": "inline_equation", + "content": "\\mathrm{ER + AT}" + }, + { + "bbox": [ + 55, + 540, + 295, + 625 + ], + "type": "text", + "content": " in class incremental settings. However, integration with our method yields significant improvements in both clean and adversarial metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "The consistent performance gains across all evaluation metrics validate our method's effectiveness in simultaneously addressing adversarial robustness and catastrophic forgetting. These empirical results demonstrate the broad applicability of our approach across different continual learning frameworks. Accuracy curves during continual training can be found in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 208, + 441, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 441, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 441, + 220 + ], + "type": "text", + "content": "5.3. Ablation Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 225, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 553, + 357 + ], + "type": "text", + "content": "In Table 4, we have performed ablation experiments based on ER+TRADES under the Split-CIFAR10 dataset. The results demonstrate that AFLC (Sec. 4.2) can effectively mitigate the increased forgetting caused by adversarial training under class incremental setting (55.49% for clean samples and 43.18% for adversarial samples forgetting, with corresponding FAA improvements of 15.46% and 2.86%, respectively). AFLC does not show significant improvement in the task incremental setting due to excessive suppression of future task classification heads and the use of the same calibration value for classes within the same task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "content": "RAER (Sec. 4.3) can further improve the robust accuracy of AFLC by " + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "inline_equation", + "content": "1.23\\%" + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "content": " for class incremental setting and " + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "content": " for task incremental setting and reduce the robust forgetting by " + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "inline_equation", + "content": "12.16\\%" + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "inline_equation", + "content": "3.47\\%" + }, + { + "bbox": [ + 313, + 357, + 553, + 430 + ], + "type": "text", + "content": " respectively. That proves the data selected by RAER describe the overall data distribution more accurately and effectively mitigate the gradient obfuscation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "text", + "content": "When considering the future prior adjustment (FP in Table 4), we find that although the forgetting of the class incremental setting is higher, the FAA of both clean samples and adversarial samples has been significantly improved, and the forgetting of the task incremental setting has been further reduced, which proves that FP can reduce negative gradients to future classes and help learn new tasks. We also provide sensitivity analysis of hyperparameters (" + }, + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "text", + "content": " in AFLC and " + }, + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 430, + 554, + 538 + ], + "type": "text", + "content": " in RAER) in supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 549, + 479, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 479, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 479, + 562 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 714 + ], + "type": "text", + "content": "This paper explores achieving adversarial robust continual learning without additional data. We analyze the challenges of directly combining continual learning and adversarial training - accelerated forgetting and gradient obfuscation. To address these, we propose: anti-forgettable logit calibration to reduce negative effects of past knowledge, and a robust difficulty-based data selection mechanism to mitigate gradient obfuscation. Experiments demonstrate the effectiveness of our approach. As AI systems increasingly need to learn continuously in dynamic environments, our work represents a further step toward more resilient and adaptable machine learning systems." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "569" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 149 + ], + "type": "text", + "content": "This work was supported in part by the Beijing Natural Science Foundation under No. L221013, Strategic Priority Research Program of the Chinese Academy of Sciences (No. XDB0680202), and the Innovation Funding of ICT, CAS under Grant No. E561160, E561090." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 115, + 171 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 178, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 178, + 296, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 178, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 296, + 234 + ], + "type": "text", + "content": "[1] Ali Abbasi, Parsa Nooralinejad, Hamed Pirsiavash, and Soheil Kolouri. Brainwash: A poisoning attack to forget in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24057-24067, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 235, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 296, + 267 + ], + "type": "text", + "content": "[2] Naveed Akhtar and Ajmal Mian. Threat of adversarial attacks on deep learning in computer vision: A survey. IEEE Access, 6:14410-14430, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 313, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 313, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 313, + 295, + 367 + ], + "type": "text", + "content": "[4] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In International Conference on Machine Learning, pages 274-283. PMLR, 2018. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 402 + ], + "type": "text", + "content": "[5] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. arXiv preprint arXiv:2102.01356, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "text", + "content": "[6] Matteo Boschini, Lorenzo Bonicelli, Pietro Buzzega, Angelo Porrello, and Simone Calderara. Class-incremental continual learning into the extended der-verse. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "text", + "content": "[7] Pietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati, and Simone Calderara. Dark experience for general continual learning: a strong, simple baseline. Advances in neural information processing systems, 33:15920-15930, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "text", + "content": "[8] Lucas Caccia, Rahaf Aljundi, Nader Asadi, Tinne Tuytelaars, Joelle Pineau, and Eugene Belilovsky. New insights on reducing abrupt representation change in online continual learning. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 613 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajanthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In European Conference on Computer Vision, pages 532-547. Springer, 2018. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[10] Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. Continual learning with tiny episodic memories. arXiv preprint arXiv:1902.10486, 2019, 2019. 1, 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[11] Jinghui Chen and Quanquan Gu. Rays: A ray searching method for hard-label adversarial attack. In Proceedings of the 26rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2020. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[12] Tianlong Chen, Sijia Liu, Shiyu Chang, Lisa Amini, and Zhangyang Wang. Queried unlabeled data improves and robustifies class-incremental learning. Transactions on Machine Learning and Data Mining, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 162 + ], + "type": "text", + "content": "[13] Ting-Chun Chou, Jhih-Yuan Huang, and Wei-Po Lee. Continual learning with adversarial training to enhance robustness of image recognition models. In 2022 International Conference on Cyberworlds, pages 236-242. IEEE, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 228 + ], + "type": "text", + "content": "[14] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 274 + ], + "type": "text", + "content": "[15] Jiequan Cui, Shu Liu, Liwei Wang, and Jiaya Jia. Learnable boundary guided adversarial training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15721-15730, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 554, + 341 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 342, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 555, + 396 + ], + "type": "text", + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9285-9295, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 554, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 554, + 452 + ], + "type": "text", + "content": "[18] Ranjie Duan, Yuefeng Chen, Dantong Niu, Yun Yang, A Kai Qin, and Yuan He. Advdrop: Adversarial attack to dnns by dropping information. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7506-7515, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 486 + ], + "type": "text", + "content": "[19] Sebastian Farquhar and Yarin Gal. Towards robust evaluations of continual learning. arXiv preprint arXiv:1805.09733, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 520 + ], + "type": "text", + "content": "[20] Yunhui Guo, Mingrui Liu, Yandong Li, Liqiang Wang, Tianbao Yang, and Tajana Rosing. Attacking lifelong learning models with gradient reversion. 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[21] Ahmad Hassanpour, Majid Moradikia, Bian Yang, Ahmed Abdelhadi, Christoph Busch, and Julian Fierrez. Differential privacy preservation in robust continual learning. IEEE Access, 10:24273-24287, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 610 + ], + "type": "text", + "content": "[22] Siteng Kang, Zhan Shi, and Xinhua Zhang. Poisoning generative replay in continual learning to promote forgetting. In International Conference on Machine Learning, pages 15769-15785. PMLR, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 655 + ], + "type": "text", + "content": "[23] Hikmat Khan, Nidhal Carla Bouaynaya, and Ghulam Rasool. Adversarily robust continual learning. In 2022 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 690 + ], + "type": "text", + "content": "[24] Hikmat Khan, Pir Masoom Shah, Syed Farhan Alam Zaidi, et al. Susceptibility of continual learning against adversarial attacks. arXiv preprint arXiv:2207.05225, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "570" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 116 + ], + "type": "text", + "content": "Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 151 + ], + "type": "text", + "content": "[26] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical Report 0, University of Toronto, Toronto, Ontario, 2009. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 194 + ], + "type": "text", + "content": "[27] Lilly Kumari, Shengjie Wang, Tianyi Zhou, and Jeff Bilmes. Retrospective adversarial replay for continual learning. In Advances in Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[28] Sungyoon Lee, Hoki Kim, and Jaewook Lee. Graddiv: Adversarial robustness of randomized neural networks via gradient diversity regularization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):2645-2651, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[29] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40(12):2935-2947, 2017. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 341 + ], + "type": "text", + "content": "[30] Guoliang Lin, Hanlu Chu, and Hanjiang Lai. Towards better plasticity-stability trade-off in incremental learning: A simple linear connector. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 89-98, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[31] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 2, 3, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[32] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 497 + ], + "type": "text", + "content": "[33] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, and Joost van de Weijer. Class-incremental learning: Survey and performance evaluation on image classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5):5513-5533, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[34] Michael McCloskey and Neal J Cohen. Catastrophic interference in connectionist networks: The sequential learning problem. In Psychology of learning and motivation, pages 109-165. Elsevier, 1989. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 589 + ], + "type": "text", + "content": "[35] Tianyu Pang, Kun Xu, Yinpeng Dong, Chao Du, Ning Chen, and Jun Zhu. Rethinking softmax cross-entropy loss for adversarial robustness. In International Conference on Learning Representations, 2020. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "text", + "content": "[36] Tianyu Pang, Xiao Yang, Yinpeng Dong, Hang Su, and Jun Zhu. Bag of tricks for adversarial training. In International Conference on Learning Representations, 2021. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[37] Tianyu Pang, Min Lin, Xiao Yang, Jun Zhu, and Shuicheng Yan. Robustness and accuracy could be reconcilable by (proper) definition. International Conference on Machine Learning, 2022. 2, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[38] German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. *Neural Networks*, 113:54–71, 2019. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[39] Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions. Psychological review, 97(2):285, 1990. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[40] Mohammad Rostami, Leonidas Spinoulas, Mohamed Hussein, Joe Mathai, and Wael Abd-Almageed. Detection and continual learning of novel face presentation attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14851-14860, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[41] Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "text", + "content": "[42] Fahad Sarfraz, Elahe Arani, and Bahram Zonooz. Error sensitivity modulation based experience replay: Mitigating abrupt representation drift in continual learning. In The Eleventh International Conference on Learning Representations, 2023, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[43] Rui Shao, Pramuditha Perera, Pong C Yuen, and Vishal M Patel. Open-set adversarial defense. In European Conference on Computer Vision, pages 682-698. Springer, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[44] Dongsub Shim, Zheda Mai, Jihwan Jeong, Scott Sanner, Hyunwoo Kim, and Jongseong Jang. Online class incremental continual learning with adversarial shapley value. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9630-9638, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[45] Stanford. Tiny imagenet challenge (cs231n). http://tiny-imagenet.herokuapp.com/, 2015. Accessed: Feb 21, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[46] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[47] Muhammad Umer, Glenn Dawson, and Robi Polikar. Targeted forgetting and false memory formation in continual learners through adversarial backdoor attacks. In 2020 International Joint Conference on Neural Networks, pages 1-8. IEEE, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[48] Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 579 + ], + "type": "text", + "content": "[49] Zhenyi Wang, Li Shen, Le Fang, Qiuling Suo, Tiehang Duan, and Mingchen Gao. Improving task-free continual learning by distributionally robust memory evolution. In International Conference on Machine Learning, pages 22985-22998. PMLR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 581, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 644 + ], + "type": "text", + "content": "[50] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 139-149, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[51] Zekai Wang, Tianyu Pang, Chao Du, Min Lin, Weiwei Liu, and Shuicheng Yan. Better diffusion models further improve adversarial training. In International Conference on Machine Learning, 2023. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[52] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "571" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 620 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[53] Tongtong Wu, Massimo Caccia, Zhuang Li, Yuan-Fang Li, Guilin Qi, and Gholamreza Haffari. Pretrained language model in continual learning: A comparative study. In International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[54] Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robustness under long-tailed distribution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8659-8668, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 239 + ], + "type": "text", + "content": "[55] Ju Xu, Jin Ma, Xuesong Gao, and Zhanxing Zhu. Adaptive progressive continual learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6715-6728, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[56] Jaehong Yoon, Saehoon Kim, Eunho Yang, and Sung Ju Hwang. Scalable and order-robust continual learning with additive parameter decomposition. arXiv preprint arXiv:1902.09432, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[57] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International Conference on Machine Learning, pages 7472-7482. PMLR, 2019. 1, 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "text", + "content": "[58] Jingfeng Zhang, Jianing Zhu, Gang Niu, Bo Han, Masashi Sugiyama, and Mohan Kankanhalli. Geometry-aware instance-reweighted adversarial training. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[59] Jie Zhang, Bo Li, Jianghe Xu, Shuang Wu, Shouhong Ding, Lei Zhang, and Chao Wu. Towards efficient data free black-box adversarial attack. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15115-15125, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 476 + ], + "type": "text", + "content": "[60] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Deep class-incremental learning: A survey. arXiv preprint arXiv:2302.03648, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 477, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 531 + ], + "type": "text", + "content": "[61] Da-Wei Zhou, Hai-Long Sun, Han-Jia Ye, and De-Chuan Zhan. Expandable subspace ensemble for pre-trained model-based class-incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23554–23564, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "text", + "content": "[62] Da-Wei Zhou, Qi-Wei Wang, Zhi-Hong Qi, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Class-incremental learning: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 620 + ], + "type": "text", + "content": "[63] Kaijie Zhu, Jindong Wang, Xixu Hu, Xing Xie, and Ge Yang. Improving generalization of adversarial training via robust critical fine-tuning. arXiv preprint arXiv:2308.02533, 2023. 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "572" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_content_list.json b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5ec881a27de41db2ffaaf78293c3bf97dbf265eb --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_content_list.json @@ -0,0 +1,1444 @@ +[ + { + "type": "text", + "text": "Adversarial Robustness of Discriminative Self-Supervised Learning in Vision", + "text_level": 1, + "bbox": [ + 111, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ömer Veysel Çağatan $^{1,2}$ 1Ömer Faruk Tal ${}^{1}\\mathrm{M}$ . Emre Gürsoy \n $^{1}$ Department of Computer Engineering, Koç University \n $^{2}$ KUIS AI Center, Koç University", + "bbox": [ + 240, + 179, + 754, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ocagatan19@ku.edu.tr, otal19@ku.edu.tr, emregursoy@ku.edu.tr", + "bbox": [ + 228, + 236, + 761, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems.", + "bbox": [ + 89, + 316, + 485, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 648, + 220, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) [4], particularly discriminative approaches, has emerged as a foundational method for training models with remarkable capabilities in areas such as language [80], vision [65], and decision-making [50]. As these models become increasingly widespread and integrated into various applications, ensuring their reliability and safety has become a critical concern [6, 7].", + "bbox": [ + 89, + 672, + 482, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One particular challenge is the surprising vulnerability of deep learning models to adversarial examples, where slight input alterations can significantly impact model performance [35, 78]. This phenomenon has sparked significant debate, seeking to understand and mitigate these vulnerabilities [3, 32, 75, 76, 79, 83-85]. One prominent theory [47] suggests that adversarial examples arise from the model's sensitivity to non-robust features in the input data.", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "According to this view, both robust (stable) and non-robust (vulnerable) features contribute to classification, with adversarial attacks manipulating the latter to cause misclassification. However, this theory, developed primarily in the context of supervised learning, faces challenges when extended to other self-supervised paradigms. [56] indicates that non-robust features are less effective in SSL methods such as contrastive learning [16], masked image modeling [40], or diffusion models [43]. This discrepancy suggests that non-robust features may lack the transferability across learning paradigms that robust or natural features possess. Thus, it becomes essential to systematically evaluate and compare how different SSL approaches respond to adversarial attacks, particularly given the theoretical evidence suggesting their feature representations may differ fundamentally from supervised models.", + "bbox": [ + 511, + 287, + 906, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These theoretical insights into how adversarial examples affect different learning paradigms highlight several critical gaps in our understanding of SSL's adversarial robustness. Notwithstanding the progress made in understanding the adversarial robustness of SSL, particularly contrastive learning, which we extensively discuss in Section 2, several key questions remain unresolved. First, with the wide variety of self-supervised representations available, employing different pretext tasks and data augmentations, which approaches demonstrate the greatest adversarial robustness? This remains unclear since most methods don't provide any results on adversarial robustness unless it is a specific focus of the proposed approach. Secondly, robustness is typically assessed by the model's accuracy on the pretraining dataset. Still, its adversarial impact on transfer learning or downstream tasks like detection and segmentation has not been thoroughly investigated [52].", + "bbox": [ + 511, + 532, + 908, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The choice of model architecture also raises questions about robustness. Standard vision SSL pretraining typically utilizes a ResNet [38] as the backbone, but more recently, larger and more powerful models [12, 20, 65] have been developed using vision transformers [28]. This leads to the question: Which architecture demonstrates greater robustness under the same SSL objective and with comparable", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2313", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "parameter sizes?", + "bbox": [ + 89, + 90, + 205, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another factor to consider is the training duration. State-of-the-art SSL models are trained for longer durations compared to their supervised counterparts. Several studies indicate that this extended training consistently enhances performance, raising the question of whether this might compromise the models' adversarial robustness.", + "bbox": [ + 89, + 111, + 480, + 202 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While previous work has examined aspects of adversarial robustness in SSL, our study provides the first comprehensive cross-model comparison across multiple tasks, architectures, and training regimes. We assess seven different SSL models (Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]) alongside a supervised model against various adversarial attacks on ImageNet [74] and nine other image-recognition datasets. We also evaluate their robustness in segmentation and detection tasks. Our investigation addresses the following key questions:", + "bbox": [ + 89, + 207, + 482, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. How does the adversarial robustness of various SSL models compare to that of supervised models on ImageNet?", + "bbox": [ + 91, + 387, + 482, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSL models consistently demonstrate greater adversarial robustness than supervised models on ImageNet. Noncontrastive methods show particular resilience against IAA [13] attacks, while all SSL approaches exhibit strong resistance to UAP [14], with MoCoV3 demonstrating the strongest overall performance.", + "bbox": [ + 109, + 433, + 482, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Does SSL robustness transfer to downstream tasks like transfer learning, segmentation, and detection?", + "bbox": [ + 89, + 523, + 482, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The robustness advantages transfer effectively to classification tasks via linear probing and fine-tuning. However, in segmentation and detection, all models exhibit similar vulnerability regardless of pretraining methodology, suggesting task-specific architectural components may override backbone robustness properties.", + "bbox": [ + 109, + 553, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. How does model architecture influence adversarial robustness under the same SSL objective?", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Architecture impact is highly objective dependent. MoCoV3 shows reduced robustness with Vision Transformers, whereas DINO demonstrates improved performance with ViT compared to ResNet, challenging the notion that architectural effects are uniform across SSL paradigms. Additionally, we have evaluated DINOv2 [65] and MAE [40] that do not demonstrate a consistently high or low level of robustness.", + "bbox": [ + 109, + 674, + 482, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. Does extending training duration compromise adversarial robustness in SSL models?", + "bbox": [ + 89, + 795, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extended training either maintains or slightly enhances adversarial performance. For UAP attacks, performance improves meaningfully after more epochs in both SwAV and MoCoV3, indicating that longer training periods do not create a performance-robustness tradeoff.", + "bbox": [ + 109, + 825, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self Supervised Learning Self-supervised learning(SSL) seeks to extract meaningful and general representations from unlabeled data by leveraging pretext tasks. These tasks can vary, such as predicting the next word [71] or neighboring words [26] in a text, reconstructing masked sections of an image [40], or ensuring that two different perspectives of the same image result in similar visual representations [16].", + "bbox": [ + 511, + 114, + 903, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Avoiding collapse is a key challenge in SSL for computer vision, and various methods can be classified based on how they address this issue. Contrastive approaches like SimCLR [16] and MoCo [19, 20, 39] use an objective that pushes apart representations of different inputs (negative samples) while bringing together those of the same input (positive samples). The performance and scalability of these methods heavily depend on the number and selection of negative samples. In another category, distillation methods such as BYOL [36], SimSiam [18], and DINO [12], prevent collapse by introducing asymmetry between different encoder branches and employing algorithmic adjustments. Additional SSL techniques, including DeepCluster [10], SeLa [1], and SwAV [11], enforce a clustering structure in the feature space to avoid constant representations. Meanwhile, methods like Barlow Twins [89], Whitening MSE (W-MSE) [30], VICReg [5], CorInfoMax [67] prevent collapse by using feature decorrelation.", + "bbox": [ + 511, + 220, + 903, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial Self-Supervised Learning While self-supervised learning (SSL) has outperformed supervised training [16], numerous studies highlight that contrastive learning remains susceptible to adversarial attacks when transferring the learned features to downstream classification tasks [42, 49]. To improve the robustness of contrastive learning, adversarial training has been adapted to self-supervised settings. In the absence of labels, adversarial examples are generated by maximizing the contrastive loss with respect to all input samples. Several prior works, such as ACL [48], RoCL [49], and CLAE [42], adopt this approach. Additionally, ACL incorporates the dual-BN technique [87] to further enhance performance. DeACL [90] introduces a two-stage approach, distilling a standard pretrained encoder through adversarial training. Nguyen et al. [64] establishes an upper bound on the adversarial loss of a prediction model, which is based on the learned representations, for any downstream task. This upper bound is determined using the model's loss on clean data and a robustness regularization term, which helps make the prediction model more resistant to adversarial attacks. [37] demonstrates that adversarial sensitivity stems from the uniform distribution of data representations on a unit hypersphere in the representation space. The presence of false negative pairs during training contributes to this effect, increasing the model's vulnerability to input perturbations.", + "bbox": [ + 511, + 493, + 903, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although self-supervised adversarial training has made", + "bbox": [ + 532, + 886, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2314", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/686b75e84ddcb7e1a609561bbf037aa119ae94a72acec203dc338d2f666d0923.jpg", + "image_caption": [ + "Figure 1. Performance scores for tasks such as ImageNet classification, transfer learning (with linear probing), segmentation, and detection (both with frozen backbones) are shown with the percentage drop in adversarial robustness. The shaded regions indicate the $95\\%$ confidence interval around the regression line. Note the consistent pattern of higher robustness (lower percentage drop) among SSL models compared to supervised approaches in classification tasks." + ], + "image_footnote": [], + "bbox": [ + 176, + 90, + 821, + 234 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a0d1b6b3b33dddc23a25c85fce4056a4fc5765fd7ed6c945de509913f41e7d67.jpg", + "image_caption": [ + "Figure 2. Averaged scores of SSL models on ImageNet across various attack types, including Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). Adv Avg refers to the average score across all attacks combined. The shaded regions indicate the $95\\%$ confidence interval around the regression line." + ], + "image_footnote": [], + "bbox": [ + 215, + 318, + 784, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "progress, it still does not match the performance of supervised methods. Luo et al. [58] suggests that this shortfall is due to data augmentation and proposes a dynamic data augmentation scheduler to achieve comparable results to supervised training. Xu et al. [88] efficiently applies ACL on the ImageNet [74] to obtain a robust representation using robustness-aware core set selection.", + "bbox": [ + 88, + 544, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robustness of Self-Supervised Learning", + "text_level": 1, + "bbox": [ + 109, + 654, + 393, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[41] found that incorporating an extra self-supervised task in a multi-task framework can enhance the adversarial robustness of supervised models. In a similar vein, Carmon et al. [9] discovered that using additional unlabeled data also strengthens the model's adversarial resilience. Furthermore, Chen et al. [17] created robust variants of pretext-based SSL tasks, showing that their integration with robust fine-tuning leads to a notable increase in robustness compared to standard adversarial training. Chhipa et al. [21] demonstrates a clear relationship between the performance of learned representations within SSL paradigms and the severity of distribution shifts and corruptions, and highlights the critical impact of distribution shifts and image corruptions on the performance and resilience of SSL methods. Similarly, Zhong et al. [91] conduct robustness tests to assess", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the behavioral differences between contrastive and supervised learning under changes in downstream or pre-training data distributions, while also exploring the effects of data augmentation and feature space characteristics. Kowalczuk et al. [52] conducts a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks, revealing the need for broader enhancements in encoder robustness. Goldblum et al. [34] benchmarks diverse pretrained models across multiple computer vision tasks, finding that supervised convolutional neural networks still outperform newer architectures on most metrics, while revealing self-supervised learning backbones show competitive potential when compared under equivalent conditions.", + "bbox": [ + 511, + 544, + 906, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike prior studies that primarily focus on individual SSL methods, specific tasks, or limited adversarial scenarios, our work provides a comprehensive, unified benchmark across multiple SSL paradigms, architectures, and tasks—including classification, transfer learning, segmentation, and detection—under a diverse set of adversarial attacks, offering a broader and deeper understanding of adversarial robustness in SSL.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 89, + 282, + 108 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. SSL Models", + "text_level": 1, + "bbox": [ + 89, + 117, + 220, + 132 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While numerous SSL approaches have been proposed [66], we focus on a subset of widely used models due to computational constraints: Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]. We primarily use ResNet50 [38] backbones, as most SSL checkpoints are released in this format, with DINO and MoCoV3 also offering ViT [28] variants. For BYOL, DINO, MoCoV3, SimCLR, and SwAV, complete model checkpoints were available. In contrast, Barlow Twins and VICReg only provided backbone weights, requiring linear evaluation via official code, which led to a slight drop in performance. For comparison, we also include a supervised ResNet50 baseline from PyTorch [68].", + "bbox": [ + 89, + 143, + 482, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. ImageNet and Transfer Learning", + "text_level": 1, + "bbox": [ + 89, + 359, + 382, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the benchmark suite introduced in the transfer learning study [45], which encompasses the target datasets like FGVC Aircraft [60], Caltech-101 [33], Stanford Cars [53], CIFAR 10 [54], CIFAR 100 [54], DTD [22], Oxford 102 Flowers [22], and Food-101 [8]. We follow Ericsson et al. [29] for both linear evaluation and fine-tuning of these datasets. We prioritized linear evaluation in our analysis as the backbone remains frozen during this process, allowing for a more equitable comparison of objectives within this setup. We apply the same adversarial techniques to ImageNet and transfer learning: Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). In brief, instance-based methods generate unique perturbations for each image, while UAP creates a single perturbation that applies across the entire dataset. Comprehensive details and categorizations of all attack methods—such as white-box, black-box, gradient-free, among others—are provided in Section 1 of the Supp Mat..", + "bbox": [ + 89, + 385, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Segmentation", + "text_level": 1, + "bbox": [ + 89, + 678, + 232, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For segmentation, we use the Pascal VOC 2012 dataset [31] and CityScapes [23] dataset, training a DeepLabV3+ model [15]. To conduct the attacks, we follow the setup from Rony et al. [73], utilizing Alma [73], Asma [73], DAG [86], DDN [73], FGSM [35], FMN [70], and PGD [59]. While our primary metric is the mean Intersection Over Union (mIOU), we also report the Attack Pixel Success Rate (APSR) introduced by [73]. Although our main focus is on using a frozen backbone, we also perform training following the standard procedure. Our CityScapes results include only the frozen backbone approach, while our Pascal VOC results include both frozen and unfrozen backbone configurations.", + "bbox": [ + 89, + 703, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/125d7db239fc4165ac1b0f9ac32e020375ad52b182f4791c22edda82b79db2c7.jpg", + "table_caption": [ + "Table 1. Performance of various models on ImageNet, Transfer Learning, Segmentation (Pascal VOC), and Detection (INRIA Person) tasks with frozen backbones, showing original (Orig.) and adversarial (Adv.) scores with performance drops in red." + ], + "table_footnote": [], + "table_body": "
ModelImageNetTransfer LearningSegmentationDetection
Orig.Adv.Orig.Adv.Orig.Adv.Orig.Adv.
Barlow Twins71.236.5 ↓49%80.337.8 ↓54%76.920.5 ↓73%88.421.9 ↓75%
BYOL74.638.9 ↓48%78.736.6 ↓55%76.719.0 ↓75%87.417.3 ↓80%
DINO75.334.2 ↓55%80.735.6 ↓57%77.018.9 ↓76%87.622.0 ↓75%
MoCoV374.640.5 ↓46%80.540.3 ↓51%76.219.9 ↓74%87.318.5 ↓79%
SimCLR68.932.1 ↓53%73.133.0 ↓57%75.619.3 ↓74%87.713.1 ↓85%
Supervised76.134.5 ↓55%74.631.2 ↓60%74.216.5 ↓78%86.118.0 ↓79%
SwAV75.332.6 ↓57%79.232.4 ↓60%76.519.2 ↓75%86.620.5 ↓76%
VICReg71.337.0 ↓48%79.937.7 ↓54%77.920.5 ↓74%88.414.0 ↓84%
", + "bbox": [ + 514, + 155, + 903, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Detection", + "text_level": 1, + "bbox": [ + 511, + 287, + 624, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For object detection, we utilized the INRIA Person [24] and CoCo [57] datasets, and trained a Faster R-CNN [72]. To perform adversarial attacks, we followed the setup described by [44], employing the Transfer-based Self-Ensemble Attack (T-SEA). The T-SEA attack can be deployed using various methods and optimizers. In our experiments, we employed BIM [44], MIM [27], PGD [59], and Optim [44] methods. Additionally, we explored simpler methods that rely on common optimizers, such as Adam [51], SGD, and Nesterov [63]. Throughout our evaluation, we report the mean average precision (mAP) scores as the primary performance metric. While our primary focus was on employing a frozen backbone, we also conducted training experiments following the standard training procedures for comparative analysis. Our COCO results include only the frozen backbone approach.", + "bbox": [ + 511, + 309, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Results and Discussion", + "text_level": 1, + "bbox": [ + 511, + 568, + 728, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our experimental findings on ImageNet, transfer learning, and detection, and discuss each in turn. Figure 1 and table 1 summarize the performance of various SSL models compared to supervised learning across our main evaluation tasks in the frozen backbone setup. While we address the results individually, the full detailed results are provided in Section 4 of the Supp. Mat..", + "bbox": [ + 511, + 594, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. ImageNet", + "text_level": 1, + "bbox": [ + 511, + 710, + 625, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SSL vs Supervised. Most robustness studies on contrastive learning [42, 48, 49, 64, 87, 90] focus on small datasets like CIFAR10 [54] and primarily evaluate robustness using adversarial attacks such as FGSM [35] and PGD [59]. While computational constraints explain the reluctance to scale to larger datasets like ImageNet [74], many evaluations inadequately incorporate Universal Adversarial Perturbations (UAP). Our findings, as shown in Figure 2 contradict previous research by Gupta et al. [37]. Under IAA, MoCoV3 demonstrates the strongest robustness (54% drop), while SimCLR shows the weakest performance (61% drop). For", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2316", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "UAP attacks, MoCoV3 again leads (38% drop), while the supervised model and SwAV both show 50% drops. Notably, our results challenge the conclusion about contrastive vs. non-contrastive methods. MoCoV3, a contrastive model, consistently demonstrates the highest adversarial robustness with ResNet architecture, while DINO, which was claimed to perform better due to its non-contrastive nature, shows the weakest IAA robustness in our ResNet evaluation (64% drop). Our data reveals that Barlow, BYOL, MoCoV3, and VICReg all demonstrate comparable resilience against IAA (around 54% drop), contradicting the simple categorization of robustness based on contrastive versus non-contrastive approaches. Our comprehensive evaluation using diverse attack methods demonstrates that the relationship between self-supervised learning approaches and adversarial robustness is more complex than previously suggested. The full results are in Section 4.1 of the Supp. Mat..", + "bbox": [ + 89, + 90, + 480, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "What makes MoCoV3 robust? Although MoCoV3 and SimCLR both utilize the InfoNCE [77, 81] objective, there is a stark contrast in their adversarial robustness. To understand this disparity, we evaluate MoCoV1 [39], MoCoV2 [19], and MoCoV3.", + "bbox": [ + 89, + 349, + 480, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A brief MoCo History. MoCoV1 introduced a dynamic dictionary with a queue and momentum-updated encoder to maintain consistent negative samples. MoCoV2 enhanced this with a multi-layer projection head and stronger data augmentation. MoCoV3 further evolved by eliminating the memory bank and incorporating a prediction head similar to BYOL and SimSiam [18].", + "bbox": [ + 89, + 429, + 480, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our analysis reveals a clear progression in robustness across MoCo versions. While MoCoV1 demonstrates limited resilience with a $71\\%$ drop in overall adversarial accuracy, MoCoV2 shows significant improvement (52% drop), and MoCoV3 achieves the strongest performance (46% drop). The most dramatic enhancement occurs in UAP resistance, where MoCoV1's performance drops by $77\\%$ , compared to MoCoV2's $39\\%$ and MoCoV3's $36\\%$ . The substantial improvement from MoCoV1 to MoCoV2 (7% in clean accuracy, 27 percentage points in UAP robustness) primarily stems from the non-linear projector, with data augmentation providing marginal benefits. While [46] suggests non-linear projectors aren't always essential, our results indicate they significantly enhance both performance and adversarial robustness. MoCoV3's superior performance over MoCoV2 (an additional $7\\%$ in clean accuracy and 7 percentage points in UAP robustness) can be attributed to its prediction head and larger batch size. Unlike the transition from V1 to V2, MoCoV3 shows substantial improvements in both IAA (34% vs 25%) and UAP (48% vs 41%), highlighting the prediction head's critical role in enhancing overall robustness. Our findings suggest that momentum, a common feature in robust models like MoCoV3 and BYOL, significantly contributes to adversar", + "bbox": [ + 89, + 537, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ial resilience, while MoCoV2's performance more closely resembles that of SimCLR, which lacks this feature. The full results are in Section 4.6 of the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Augmentations vs Algorithms.", + "text_level": 1, + "bbox": [ + 532, + 138, + 746, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Morningstar et al. [61] challenges the notion that SSL progress is primarily driven by algorithmic advancements, suggesting instead that augmentation diversity, along with data and model scale, play more critical roles. Their analysis argues that many algorithmic improvements, such as prediction networks or new loss functions, had minimal impact on downstream task performance compared to stronger augmentation techniques.", + "bbox": [ + 511, + 154, + 903, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our comprehensive analysis of the MoCo family evolution provides a more nuanced perspective on this debate. The substantial progression in adversarial robustness from MoCoV1 to MoCoV2 and further to MoCoV3 suggests that architectural innovations like non-linear projectors and prediction heads significantly impact robustness. MoCoV2's dramatic improvement over MoCoV1, particularly in UAP performance, indicates that the multi-layer projection head provides substantial benefits beyond mere augmentation changes. Similarly, MoCoV3's further enhancements in both IAA and UAP performance relative to MoCoV2 highlight the crucial role of the prediction head in overall robustness.", + "bbox": [ + 511, + 276, + 903, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our comparison lacks perfectly controlled baselines for augmentations across different objectives (with slight variations in augmentation between MoCoV2 and V3), this limitation stems from the unavailability of public checkpoints rather than our experimental design. Despite this constraint, the consistent improvements in adversarial performance strongly suggest that algorithmic innovations significantly contribute to adversarial robustness. Importantly, our findings demonstrate that higher clean accuracy doesn't automatically translate to improved robustness on ImageNet.", + "bbox": [ + 511, + 474, + 903, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although the checkpoints from [61] are not publicly available, we conducted an ablation study on augmentation types and batch sizes using BYOL which is the only model in our SSL pool that includes configurations with varied augmentations and batch sizes. For this analysis, we consider four distinct models: BYOL-NC (without color distortions), BYOL-CC (with only color and crop augmentations), and the standard BYOL models with batch sizes of 128 (BYOL-128) and 512 (BYOL-512).", + "bbox": [ + 511, + 642, + 903, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that batch size and augmentation choices have varying effects on the adversarial robustness of BYOL variants. For IAA robustness, both BYOL-NC and BYOLCC show similar performance (60% drop), while BYOL128, BYOL-512, and standard BYOL demonstrate slightly better resilience (57%, 56%, and 56% drops, respectively). These modest differences suggest that batch size and augmentation have a limited impact on instance-level attack ro", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bustness.", + "bbox": [ + 89, + 90, + 156, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For UAP attacks, we see more substantial variations: BYOL-128 (41% drop) and BYOL-512 (43% drop) perform comparably, while BYOL-NC and BYOL-CC show notably weaker performance (46% and 48% drops). Standard BYOL demonstrates a 39% drop, which is marginally better than the batch size variants but substantially better than the limited-augmentation models. The 7-9 percentage point difference between standard BYOL and the limited-augmentation variants suggests that comprehensive augmentation strategies may contribute to improved UAP robustness.", + "bbox": [ + 88, + 108, + 480, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interestingly, we note that BYOL-CC performs slightly worse than BYOL-NC despite having more augmentations, though this difference is too small to draw meaningful conclusions. Similarly, the differences between BYOL-128, BYOL-512, and standard BYOL in UAP robustness are relatively minor and should be interpreted cautiously. Our results indicate that the relationship between augmentation and adversarial robustness is complex and that meaningful improvements likely require more than just incremental changes to batch size or augmentation strategies. While our data hints at potential benefits from comprehensive augmentation for UAP robustness, the effects are modest and require further investigation with more controlled experiments. The full results are in Section 4.7 of the supplementary material.", + "bbox": [ + 91, + 276, + 483, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ResNet vs ViT in Adversarial Robustness. While ViTs are generally seen as more robust than CNNs [62], Bai et al. [2], Pinto et al. [69] demonstrate that with the right training methods, CNNs [55] can achieve comparable robustness. Despite ViT's success [12, 20, 25, 28, 65], most SSL methods still use ResNet for validation. We examine MoCoV3 and DINO, as they are the only models that include ViT training, with our analysis covering both standard ViTs comparable to ResNet50 and larger ViT-B variants with approximately 4x more parameters.", + "bbox": [ + 89, + 503, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our results reveal notable architectural differences. DINO performs significantly better with ViT architectures than ResNet, with DINO-ViT-B achieving $45.19\\%$ adversarial average (42% drop from clean accuracy) compared to DINO-ResNet's $55\\%$ drop. In contrast, MoCo struggles with transformer architectures, with MoCo-ViT showing poor performance at a $61\\%$ drop, significantly worse than MoCoV3-ResNet at a $46\\%$ drop. Even MoCoV3-ViT-B $51\\%$ drop underperforms its ResNet counterpart despite having 4x more parameters.", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These differences are especially evident in UAP results, where DINO-ViT-B shows greater resilience (38% drop) than DINO-ResNet (44% drop), while MoCo-ViT performs poorly (26.31%, 64% drop) compared to MoCoV3-ResNet (36% drop). These findings suggest that the interaction between self-supervised approaches and model architectures", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "significantly impacts adversarial robustness, with DINO benefiting from the ViT architecture while MoCo struggles with transformer models. The full results are in Section 4.6 of the Supp. Mat..", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, we have evaluated DINOv2's [65] small and base and MAE's [40] base versions. We discuss them in the section 2 in Supp. Mat..", + "bbox": [ + 511, + 152, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Training Duration. SSL models tend to demonstrate better performance as training epochs increase [11, 16, 20]. However, due to computational constraints, many models are reported with different numbers of epochs. This prompts the question of whether longer training durations enhance or reduce adversarial robustness. As ViT models do not have checkpoints at various epochs, we focus on ResNet-based SSL models, specifically SwAV and MoCoV3, which offer multiple checkpoints throughout the training process. We find that both SwAV and MoCo show very marginal improvement of about $1\\%$ on IAA across various epochs, which is minimal compared to the rise in original accuracy. In contrast, both methods exhibit a modest increase in UAP performance after surpassing 100 epochs, with SwAV improving from $33\\%$ to $40\\%$ and MoCoV3 from $41\\%$ to $48\\%$ . Overall, our results suggest that despite differences in reported checkpoints, robustness generally remains largely stable during training for IAA, with slightly more noticeable but still modest gains for UAP. This indicates that training duration has a limited impact on adversarial robustness, even as clean accuracy continues to improve with extended training. The full results are in Section 4.5 of the Supp. Mat.", + "bbox": [ + 511, + 199, + 903, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Transfer Learning", + "text_level": 1, + "bbox": [ + 511, + 559, + 692, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we analyze how adversarial robustness transfers from ImageNet pre-training to downstream tasks. We examine whether vulnerability patterns established during pre-training persist when models are evaluated through linear probing or fine-tuning across various ResNet-based self-supervised and supervised learning approaches. This analysis quantifies robustness transfer relationships through both Spearman rank and Pearson correlations.", + "bbox": [ + 511, + 582, + 903, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LINEAR. Our correlation analysis shows strong relationships between ImageNet and linear evaluation vulnerability. For performance drops, both Spearman and Pearson correlations are high across all attack types (Spearman—Adv: 0.93, UAP: 0.83, IAA: 0.79; Pearson—Adv: 0.94, UAP: 0.94, IAA: 0.88). These consistently high correlations suggest that robustness characteristics established during pre-training largely persist through linear probing. The performance drop analysis reveals method-specific robustness characteristics. MoCoV3 consistently exhibits smaller performance drops across both ImageNet (IAA: $54\\%$ , UAP: $36\\%$ ) and linear evaluation (IAA: $56\\%$ , UAP: $43\\%$ ), indicating its contrastive learning approach with mo", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2318", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fcca2a0eb816a57576b4afec2d274588315351dd955c17f0f405aa8f83b892d6.jpg", + "table_caption": [ + "Table 2. Component Contribution Analysis Across Datasets. Metrics include: Original Accuracy (Orig), Adversarial Accuracy (Adv) with relative performance drop percentage, Direction Ratio (D) and Magnitude Ratio (M). Ratio interpretation: Head dominant (<0.67), Balanced (0.67-1.5), Backbone dominant (>1.5)." + ], + "table_footnote": [], + "table_body": "
ModelCIFAR-10CIFAR-100ImageNetCIFAR-10 (FT)CIFAR-100 (FT)
OrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DM
Barlow92.333.0 ↓64%2.73.177.920.5 ↓74%1.45.071.242.4 ↓40%0.99.497.166.9 ↓31%1.52.484.645.0 ↓57%0.82.4
BYOL93.031.0 ↓67%1.64.278.219.0 ↓76%1.076.074.639.4 ↓47%0.912.596.967.0 ↓31%1.21.483.961.2 ↓27%0.72.7
DINO93.927.6 ↓71%3.23.379.716.0 ↓80%1.66.275.324.7 ↓67%1.214.196.977.8 ↓20%1.92.984.752.9 ↓38%1.794.6
MoCoV394.733.0 ↓65%2.51.780.219.3 ↓76%1.14.574.642.7 ↓43%0.710.996.972.0 ↓26%1.11.384.562.6 ↓26%0.72.6
SimCLR91.037.9 ↓58%1.51.973.019.5 ↓73%1.33.568.924.3 ↓65%1.07.797.267.3 ↓31%1.01.984.443.0 ↓50%0.62.0
Supervised91.442.8 ↓53%1.80.773.924.5 ↓67%1.31.576.138.8 ↓49%1.43.096.262.3 ↓35%2.00.582.659.3 ↓28%1.31.1
SwAV93.919.4 ↓79%3.13.079.411.1 ↓86%1.86.475.324.7 ↓67%1.217.496.879.9 ↓17%1.72.784.454.6 ↓35%1.03.0
VICReg92.833.0 ↓64%2.73.177.822.3 ↓71%1.34.371.342.4 ↓40%0.99.397.168.6 ↓29%1.42.484.343.4 ↓48%0.82.4
", + "bbox": [ + 119, + 142, + 879, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mentum encoders develops features with inherently better transferable robustness properties. In contrast, supervised pre-training and clustering-based approaches like SwAV experience more severe drops in both settings, suggesting these methods may create more brittle representations.", + "bbox": [ + 88, + 303, + 480, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The high correlation in adversarial vulnerability between ImageNet and linear evaluation indicates that linear probing provides a reliable assessment of a pre-trained model's downstream robustness characteristics without requiring extensive adaptation. This finding has practical implications for model selection, suggesting that robustness evaluations on ImageNet can effectively predict linear transfer performance. The full results are in Section 4.8 of the Supp. Mat..", + "bbox": [ + 88, + 382, + 482, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FINETUNE. Fine-tuning reveals distinct transfer patterns by attack type. For IAA, the correlation between ImageNet and fine-tuned vulnerability nearly vanishes for performance drops (Spearman: 0.12, Pearson: 0.36), indicating that fine-tuning substantially reshapes defense against instance-specific attacks. Conversely, UAP vulnerability correlation remains high (Spearman: 0.88, Pearson: 0.93), suggesting that susceptibility to universal perturbations is more persistently encoded in the network regardless of parameter adaptation. This attack-dependent correlation disparity reveals that universal perturbation vulnerabilities, which exploit systematic weaknesses across the feature space, are deeply encoded in the network's architecture and learning approach, while instance-specific vulnerabilities are more malleable through fine-tuning. The absolute magnitude of performance drops diminishes considerably after fine-tuning across all methods (IAA: $49\\%$ vs. $60\\%$ in linear, UAP: $43\\%$ vs. $49\\%$ in linear), highlighting fine-tuning's effectiveness in mitigating vulnerability. MoCoV3 continues to demonstrate superior robustness with the smallest drops (IAA: $45\\%$ , UAP: $38\\%$ ), while BYOL shows dramatic improvement from ImageNet to fine-tuning for IAA attacks (from $56\\%$ to $47\\%$ ). These findings suggest different mechanisms for robustness transfer: UAP vulnerability appears tied to fundamental architectural and al", + "bbox": [ + 91, + 523, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gorithmic properties that persist across transfer paradigms, while instance-specific attack vulnerability depends more on the fine-tuning process than initial representation properties. This distinction may be valuable for practical applications, suggesting that pre-training method selection strongly impacts universal attack robustness, while defense against instance-specific attacks can be substantially improved through appropriate fine-tuning strategies. The full results are in Section 4.9 of the Supp. Mat..", + "bbox": [ + 511, + 303, + 903, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attributing Adversarial Vulnerability. Having established the correlations between ImageNet and transfer learning robustness patterns, we now seek to understand the underlying mechanisms causing these disparities across models and training paradigms. Specifically, we investigate whether adversarial vulnerability stems primarily from the backbone feature extractor or the classification head and how this attribution changes between linear probing and fine-tuning.", + "bbox": [ + 511, + 439, + 903, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our analysis investigates the relationship between adversarial robustness and component-specific vulnerability under $\\mathrm{FGSM}_1$ attack (detail in Supp. Mat.), focusing on two metrics: the Direction Ratio $D = \\frac{\\mathrm{mean}(1 - \\cos(\\mathbf{l}_{adv}))}{\\mathrm{mean}(1 - \\cos(\\mathbf{f},\\mathbf{f}_{adv}))}$ , which compares directional shifts in logits (l) versus backbone features (f), and the Magnitude Ratio $M = \\frac{\\mathrm{mean}(\\|\\mathbf{l} - \\mathbf{l}_{adv}\\|_2)}{\\mathrm{mean}(\\|\\mathbf{f} - \\mathbf{f}_{adv}\\|_2)}$ , quantifying relative sensitivity to perturbation magnitudes. We fit linear regression models using these ratios to predict adversarial accuracy drop across models.", + "bbox": [ + 511, + 577, + 903, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For probed models, Direction Ratio strongly correlates with vulnerability $(R^2 = 0.56 - 0.85)$ , suggesting that directional instability in the head dominates FGSM robustness. SwAV, with $D = 3.1$ on CIFAR-10, suffers a $79\\%$ accuracy drop, while SimCLR $(D = 1.5)$ drops $58\\%$ . This may arise from the linear head's limited capacity to compensate for adversarial noise, amplifying directional shifts in logit space.", + "bbox": [ + 511, + 718, + 903, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After fine-tuning, this correlation weakens significantly $(R^2 = 0.05 - 0.52)$ : SwAV achieves only a $17\\%$ drop despite retaining $D = 1.7$ , implying that fine-tuning enables the head to stabilize predictions even under directional feature", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2319", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "shifts. This pattern aligns with our transfer learning analysis, where fine-tuning disrupted the correlation between ImageNet and downstream instance-specific attack robustness. Regression figures are in Section 2 of the Supp. Mat.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As in Table 2, self-supervised models exhibit higher pre-fine-tuning Magnitude Ratios (e.g., SwAV $M = 17.4$ on ImageNet vs. supervised $M = 3.0$ ), potentially due to their reliance on globally normalized feature spaces, where FGSM perturbations propagate more aggressively to logits. Fine-tuning reduces $M$ across models (SwAV $M = 2.7$ post-FT), aligning with improved robustness, possibly by suppressing logit magnitude distortions.", + "bbox": [ + 89, + 152, + 480, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These patterns are specific to FGSM, as its single-step gradient reliance may emphasize head-layer instability, whereas iterative attacks could exploit backbone vulnerabilities differently. While our analysis tentatively links directional logit shifts $(D)$ and magnitude sensitivity $(M)$ to robustness, dataset-dependent variations—such as weaker correlations on CIFAR-100 FT—highlight the need for broader evaluation across threat models to generalize these insights.", + "bbox": [ + 89, + 273, + 480, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also measured inter-class and intra-class distances for both probed and fine-tuned CIFAR-10 representations; however, these metrics did not yield meaningful correlations that explain the observed differences in adversarial robustness. The detailed regression figures, inter/intra-class distance measurements, and t-SNE [82] visualizations of feature spaces are provided in Section 3 of the Supp. Mat.", + "bbox": [ + 89, + 411, + 480, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Segmentation", + "text_level": 1, + "bbox": [ + 89, + 529, + 232, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unlike in classification, we observe no strong correlation between ImageNet robustness and segmentation performance across both PASCAL VOC and CityScapes datasets. Self-supervised approaches demonstrate competitive performance in clean conditions, with VICReg leading on PASCAL VOC (77.89 mIOU with frozen backbone) and SwAV on CityScapes (66.48 mIOU). All models suffer catastrophic performance degradation (74-84%) under adversarial attacks regardless of training methodology. This uniform degradation pattern suggests attacks primarily target segmentation modules rather than backbones.", + "bbox": [ + 89, + 551, + 480, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments with frozen versus unfrozen backbones reveal that frozen backbones generally achieve both higher clean performance and slightly better adversarial robustness on PASCAL VOC. The supervised model's exception to this pattern stems from using the standard MMSegmentation model for the unfrozen case due to reproduction challenges. These findings indicate that while SSL models produce competitive segmentation performance in clean conditions, they offer minimal advantage in terms of adversarial robustness for segmentation tasks, unlike their significant impact on object recognition. This highlights the need for robustness techniques specifically designed for segmenta", + "bbox": [ + 89, + 719, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion architectures rather than focusing solely on backbone improvements. The full results are in Section 4.2 of the Supp. Mat.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Detection", + "text_level": 1, + "bbox": [ + 513, + 150, + 624, + 164 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Detection results show distinct patterns from recognition and segmentation tasks. On INRIA Person with unfrozen backbones, SwAV demonstrates the highest robustness (72% decrease), while the Supervised model shows the poorest performance (89% decrease). With frozen backbones, Barlow Twins and DINO lead in robustness (both 75% decrease), while VICReg becomes unexpectedly vulnerable (84% decrease), contradicting its recognition performance.", + "bbox": [ + 511, + 172, + 903, + 308 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "COCO dataset results show improved robustness overall, with Barlow Twins maintaining the strongest performance (50% decrease) and the Supervised model remaining the least robust (67% decrease). This consistent weakness in the Supervised model suggests inherent robustness benefits from self-supervised pretraining.", + "bbox": [ + 511, + 311, + 903, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings indicate that task-specific architectures significantly influence adversarial robustness, and robustness in ImageNet classification doesn't necessarily transfer to detection tasks. This highlights the importance of task-specific evaluations and suggests that backbone architecture becomes less critical than overall model design for downstream applications. The full results are in Section 4.3 of the Supp. Mat.", + "bbox": [ + 511, + 402, + 903, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 542, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our evaluation shows that SSL models generally offer greater adversarial robustness than supervised counterparts in image classification, with MoCoV3 performing particularly well—likely due to its momentum encoder and prediction head. This robustness advantage is less evident in segmentation and detection, where task-specific architectures play a larger role. Architectural effects also vary by objective: DINO benefits from ViTs, while MoCoV3 performs better with CNNs. Extended training does not compromise robustness and may slightly improve it. We find that vulnerability to UAPs transfers more consistently across learning paradigms, while susceptibility to IAAs can be mitigated through fine-tuning. These results highlight the complex interplay between SSL objectives, architectures, and downstream tasks, underscoring the need for further study. We hope our findings support continued progress toward more robust visual representation learning.", + "bbox": [ + 511, + 568, + 903, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 844, + 681, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was funded by the KUIS AI Center at Koç University, Turkey.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2320", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning, 2020. 2", + "[2] Yutong Bai, Jieru Mei, Alan Yuille, and Cihang Xie. Are transformers more robust than cnns?, 2021. 6", + "[3] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing, 2022. 1", + "[4] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari S. Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A cookbook of self-supervised learning. ArXiv, abs/2304.12210, 2023. 1", + "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning, 2022. 2, 4", + "[6] Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian Hadfield, Jeff Clune, Tegan Maharaj, Frank Hutter, Atilim Gunes Baydin, Sheila McIlraith, Qiqi Gao, Ashwin Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Brauner, and Soren Mindermann. Managing extreme ai risks amid rapid progress. Science, 384(6698): 842-845, 2024. 1", + "[7] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga" + ], + "bbox": [ + 99, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models, 2022. 1", + "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In Computer Vision – ECCV 2014, pages 446–461, Cham, 2014. Springer International Publishing. 4", + "[9] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, Percy Liang, and John C. Duchi. Unlabeled data improves adversarial robustness, 2022. 3", + "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features, 2019. 2", + "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 4, 6", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 4, 6", + "[13] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey, 2018. 2", + "[14] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020. 2", + "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation, 2018. 4", + "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. *ArXiv*, abs/2002.05709, 2020. 1, 2, 4, 6", + "[17] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning, 2020. 3", + "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2020. 2, 5", + "[19] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5", + "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 1, 2, 4, 6", + "[21] Prakash Chandra Chhipa, Johan Rodahl Holmgren, Kanjar De, Rajkumar Saini, and Marcus Liwicki. Can self-supervised representation learning methods withstand distribution shifts and corruptions?, 2023. 3", + "[22] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild, 2013. 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2321", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. 4", + "[24] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 886-893 vol. 1, 2005. 4", + "[25] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 6", + "[26] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2", + "[27] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum, 2018. 4", + "[28] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 1, 4, 6", + "[29] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. How well do self-supervised models transfer?, 2021. 4", + "[30] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning, 2021. 2", + "[31] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html", + "[32] Alhussein Fawzi, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Robustness of classifiers: from adversarial to random noise, 2016. 1", + "[33] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 4", + "[34] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, Rama Chellappa, Andrew Gordon Wilson, and Tom Goldstein. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks, 2023. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 4", + "[36] Jean-Bastien Grill, Florian Strub, Florent Altch'e, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. ArXiv, abs/2006.07733, 2020. 2, 4", + "[37] Rohit Gupta, Naveed Akhtar, Ajmal Mian, and Mubarak Shah. Contrastive self-supervised learning leads to higher adversarial susceptibility, 2022. 2, 4", + "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 1, 4", + "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 5", + "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2021. 1, 2, 6", + "[41] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty, 2019. 3", + "[42] Chih-Hui Ho and Nuno Vasconcelos. Contrastive learning with adversarial examples, 2020. 2, 4", + "[43] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1", + "[44] Hao Huang, Ziyan Chen, Huanran Chen, Yongtao Wang, and Kevin Zhang. T-sea: Transfer-based self-ensemble attack on object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20514–20523, 2023. 4", + "[45] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes imagenet good for transfer learning?, 2016. 4", + "[46] Mark Ibrahim, David Klindt, and Randall Balestriero. Occam's razor for self supervised learning: What is sufficient to learn good representations?, 2024. 5", + "[47] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features, 2019. 1", + "[48] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning, 2020. 2, 4", + "[49] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning, 2020. 2, 4", + "[50] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. 1", + "[51] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2322", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Antoni Kowalczuk, Jan Dubinski, Atiyeh Ashari Ghomi, Yi Sui, George Stein, Jiapeng Wu, Jesse C. Cresswell, Franziska Boenisch, and Adam Dziedzic. Benchmarking robust self-supervised learning across diverse downstream tasks, 2024. 1, 3", + "[53] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 4", + "[54] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 4", + "[55] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[56] Ang Li, Yifei Wang, Yiwen Guo, and Yisen Wang. Adversarial examples are not real features, 2024. 1", + "[57] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 4", + "[58] Rundong Luo, Yifei Wang, and Yisen Wang. Rethinking the effect of data augmentation in adversarial contrastive learning, 2023. 3", + "[59] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 4", + "[60] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft, 2013. 4", + "[61] Warren Morningstar, Alex Bijamov, Chris Duvarney, Luke Friedman, Neha Kalibhat, Luyang Liu, Philip Mansfield, Renan Rojas-Gomez, Karan Singhal, Bradley Green, and Sushant Prakash. Augmentations vs algorithms: What works in self-supervised learning, 2024. 5", + "[62] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 262-271, 2020. 6", + "[63] Yurii Nesterov. A method for solving the convex programming problem with convergence rate $o(1 / k^2)$ . Proceedings of the USSR Academy of Sciences, 269:543-547, 1983. 4", + "[64] A. Tuan Nguyen, Ser Nam Lim, and Philip Torr. Task-agnostic robust representation learning, 2022. 2, 4", + "[65] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 1, 2, 6", + "[66] Utku Ozbulak, Hyun Jung Lee, Beril Boga, Esla Timothy Anzaku, Homin Park, Arnout Van Messem, Wesley De Neve," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Joris Vankerschaver. Know your self-supervised learning: A survey on image-based generative and discriminative training, 2023. 4", + "[67] Serdar Ozsoy, Shadi Hamdan, Sercan Ö. Arik, Deniz Yuret, and Alper T. Erdogan. Self-supervised learning with an information maximization criterion, 2022. 2", + "[68] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 4", + "[69] Francesco Pinto, Philip H. S. Torr, and Puneet K. Dokania. An impartial take to the cnn vs transformer robustness contest, 2022. 6", + "[70] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints, 2021. 4", + "[71] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2", + "[72] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 4", + "[73] Jérôme Rony, Jean-Christophe Pesquet, and Ismail Ben Ayed. Proximal splitting adversarial attacks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4", + "[74] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 2, 3, 4", + "[75] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Mady. Adversarily robust generalization requires more data, 2018. 1", + "[76] Ali Shafahi, W. Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable?, 2020. 1", + "[77] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Neural Information Processing Systems, 2016. 5", + "[78] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[79] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples, 2016. 1", + "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2323", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5", + "[82] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8", + "[83] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1", + "[84] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training, 2022.", + "[85] Dongxian Wu, Shu tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization, 2020. 1", + "[86] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection, 2017. 4", + "[87] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan Yuille, and Quoc V. Le. Adversarial examples improve image recognition, 2020. 2, 4", + "[88] Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, and Mohan Kankanhalli. Efficient adversarial contrastive learning via robustness-aware coreset selection, 2023. 3", + "[89] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction, 2021. 2, 4", + "[90] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D. Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness, 2022. 2, 4", + "[91] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning?, 2022. 3" + ], + "bbox": [ + 91, + 92, + 480, + 561 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2324", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_model.json b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f884903eed6f058e4710f0169e53841edfce2d78 --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_model.json @@ -0,0 +1,2391 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.131, + 0.885, + 0.153 + ], + "angle": 0, + "content": "Adversarial Robustness of Discriminative Self-Supervised Learning in Vision" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.18, + 0.756, + 0.235 + ], + "angle": 0, + "content": "Ömer Veysel Çağatan\\(^{1,2}\\) 1Ömer Faruk Tal \\({}^{1}\\mathrm{M}\\). Emre Gürsoy \n\\(^{1}\\)Department of Computer Engineering, Koç University \n\\(^{2}\\)KUIS AI Center, Koç University" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.237, + 0.762, + 0.251 + ], + "angle": 0, + "content": "ocagatan19@ku.edu.tr, otal19@ku.edu.tr, emregursoy@ku.edu.tr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.621 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.649, + 0.222, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) [4], particularly discriminative approaches, has emerged as a foundational method for training models with remarkable capabilities in areas such as language [80], vision [65], and decision-making [50]. As these models become increasingly widespread and integrated into various applications, ensuring their reliability and safety has become a critical concern [6, 7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "One particular challenge is the surprising vulnerability of deep learning models to adversarial examples, where slight input alterations can significantly impact model performance [35, 78]. This phenomenon has sparked significant debate, seeking to understand and mitigate these vulnerabilities [3, 32, 75, 76, 79, 83-85]. One prominent theory [47] suggests that adversarial examples arise from the model's sensitivity to non-robust features in the input data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.53 + ], + "angle": 0, + "content": "According to this view, both robust (stable) and non-robust (vulnerable) features contribute to classification, with adversarial attacks manipulating the latter to cause misclassification. However, this theory, developed primarily in the context of supervised learning, faces challenges when extended to other self-supervised paradigms. [56] indicates that non-robust features are less effective in SSL methods such as contrastive learning [16], masked image modeling [40], or diffusion models [43]. This discrepancy suggests that non-robust features may lack the transferability across learning paradigms that robust or natural features possess. Thus, it becomes essential to systematically evaluate and compare how different SSL approaches respond to adversarial attacks, particularly given the theoretical evidence suggesting their feature representations may differ fundamentally from supervised models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.909, + 0.791 + ], + "angle": 0, + "content": "These theoretical insights into how adversarial examples affect different learning paradigms highlight several critical gaps in our understanding of SSL's adversarial robustness. Notwithstanding the progress made in understanding the adversarial robustness of SSL, particularly contrastive learning, which we extensively discuss in Section 2, several key questions remain unresolved. First, with the wide variety of self-supervised representations available, employing different pretext tasks and data augmentations, which approaches demonstrate the greatest adversarial robustness? This remains unclear since most methods don't provide any results on adversarial robustness unless it is a specific focus of the proposed approach. Secondly, robustness is typically assessed by the model's accuracy on the pretraining dataset. Still, its adversarial impact on transfer learning or downstream tasks like detection and segmentation has not been thoroughly investigated [52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The choice of model architecture also raises questions about robustness. Standard vision SSL pretraining typically utilizes a ResNet [38] as the backbone, but more recently, larger and more powerful models [12, 20, 65] have been developed using vision transformers [28]. This leads to the question: Which architecture demonstrates greater robustness under the same SSL objective and with comparable" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2313" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.206, + 0.106 + ], + "angle": 0, + "content": "parameter sizes?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.112, + 0.482, + 0.203 + ], + "angle": 0, + "content": "Another factor to consider is the training duration. State-of-the-art SSL models are trained for longer durations compared to their supervised counterparts. Several studies indicate that this extended training consistently enhances performance, raising the question of whether this might compromise the models' adversarial robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.483, + 0.375 + ], + "angle": 0, + "content": "While previous work has examined aspects of adversarial robustness in SSL, our study provides the first comprehensive cross-model comparison across multiple tasks, architectures, and training regimes. We assess seven different SSL models (Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]) alongside a supervised model against various adversarial attacks on ImageNet [74] and nine other image-recognition datasets. We also evaluate their robustness in segmentation and detection tasks. Our investigation addresses the following key questions:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.388, + 0.483, + 0.432 + ], + "angle": 0, + "content": "1. How does the adversarial robustness of various SSL models compare to that of supervised models on ImageNet?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.434, + 0.483, + 0.523 + ], + "angle": 0, + "content": "SSL models consistently demonstrate greater adversarial robustness than supervised models on ImageNet. Noncontrastive methods show particular resilience against IAA [13] attacks, while all SSL approaches exhibit strong resistance to UAP [14], with MoCoV3 demonstrating the strongest overall performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.524, + 0.483, + 0.553 + ], + "angle": 0, + "content": "2. Does SSL robustness transfer to downstream tasks like transfer learning, segmentation, and detection?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.554, + 0.483, + 0.644 + ], + "angle": 0, + "content": "The robustness advantages transfer effectively to classification tasks via linear probing and fine-tuning. However, in segmentation and detection, all models exhibit similar vulnerability regardless of pretraining methodology, suggesting task-specific architectural components may override backbone robustness properties." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.645, + 0.483, + 0.674 + ], + "angle": 0, + "content": "3. How does model architecture influence adversarial robustness under the same SSL objective?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.675, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Architecture impact is highly objective dependent. MoCoV3 shows reduced robustness with Vision Transformers, whereas DINO demonstrates improved performance with ViT compared to ResNet, challenging the notion that architectural effects are uniform across SSL paradigms. Additionally, we have evaluated DINOv2 [65] and MAE [40] that do not demonstrate a consistently high or low level of robustness." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.824 + ], + "angle": 0, + "content": "4. Does extending training duration compromise adversarial robustness in SSL models?" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Extended training either maintains or slightly enhances adversarial performance. For UAP attacks, performance improves meaningfully after more epochs in both SwAV and MoCoV3, indicating that longer training periods do not create a performance-robustness tradeoff." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.221 + ], + "angle": 0, + "content": "Self Supervised Learning Self-supervised learning(SSL) seeks to extract meaningful and general representations from unlabeled data by leveraging pretext tasks. These tasks can vary, such as predicting the next word [71] or neighboring words [26] in a text, reconstructing masked sections of an image [40], or ensuring that two different perspectives of the same image result in similar visual representations [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.905, + 0.493 + ], + "angle": 0, + "content": "Avoiding collapse is a key challenge in SSL for computer vision, and various methods can be classified based on how they address this issue. Contrastive approaches like SimCLR [16] and MoCo [19, 20, 39] use an objective that pushes apart representations of different inputs (negative samples) while bringing together those of the same input (positive samples). The performance and scalability of these methods heavily depend on the number and selection of negative samples. In another category, distillation methods such as BYOL [36], SimSiam [18], and DINO [12], prevent collapse by introducing asymmetry between different encoder branches and employing algorithmic adjustments. Additional SSL techniques, including DeepCluster [10], SeLa [1], and SwAV [11], enforce a clustering structure in the feature space to avoid constant representations. Meanwhile, methods like Barlow Twins [89], Whitening MSE (W-MSE) [30], VICReg [5], CorInfoMax [67] prevent collapse by using feature decorrelation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Adversarial Self-Supervised Learning While self-supervised learning (SSL) has outperformed supervised training [16], numerous studies highlight that contrastive learning remains susceptible to adversarial attacks when transferring the learned features to downstream classification tasks [42, 49]. To improve the robustness of contrastive learning, adversarial training has been adapted to self-supervised settings. In the absence of labels, adversarial examples are generated by maximizing the contrastive loss with respect to all input samples. Several prior works, such as ACL [48], RoCL [49], and CLAE [42], adopt this approach. Additionally, ACL incorporates the dual-BN technique [87] to further enhance performance. DeACL [90] introduces a two-stage approach, distilling a standard pretrained encoder through adversarial training. Nguyen et al. [64] establishes an upper bound on the adversarial loss of a prediction model, which is based on the learned representations, for any downstream task. This upper bound is determined using the model's loss on clean data and a robustness regularization term, which helps make the prediction model more resistant to adversarial attacks. [37] demonstrates that adversarial sensitivity stems from the uniform distribution of data representations on a unit hypersphere in the representation space. The presence of false negative pairs during training contributes to this effect, increasing the model's vulnerability to input perturbations." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Although self-supervised adversarial training has made" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2314" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.092, + 0.823, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.247, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 1. Performance scores for tasks such as ImageNet classification, transfer learning (with linear probing), segmentation, and detection (both with frozen backbones) are shown with the percentage drop in adversarial robustness. The shaded regions indicate the \\(95\\%\\) confidence interval around the regression line. Note the consistent pattern of higher robustness (lower percentage drop) among SSL models compared to supervised approaches in classification tasks." + }, + { + "type": "image", + "bbox": [ + 0.216, + 0.319, + 0.785, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.476, + 0.908, + 0.517 + ], + "angle": 0, + "content": "Figure 2. Averaged scores of SSL models on ImageNet across various attack types, including Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). Adv Avg refers to the average score across all attacks combined. The shaded regions indicate the \\(95\\%\\) confidence interval around the regression line." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.483, + 0.65 + ], + "angle": 0, + "content": "progress, it still does not match the performance of supervised methods. Luo et al. [58] suggests that this shortfall is due to data augmentation and proposes a dynamic data augmentation scheduler to achieve comparable results to supervised training. Xu et al. [88] efficiently applies ACL on the ImageNet [74] to obtain a robust representation using robustness-aware core set selection." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.655, + 0.395, + 0.67 + ], + "angle": 0, + "content": "Robustness of Self-Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[41] found that incorporating an extra self-supervised task in a multi-task framework can enhance the adversarial robustness of supervised models. In a similar vein, Carmon et al. [9] discovered that using additional unlabeled data also strengthens the model's adversarial resilience. Furthermore, Chen et al. [17] created robust variants of pretext-based SSL tasks, showing that their integration with robust fine-tuning leads to a notable increase in robustness compared to standard adversarial training. Chhipa et al. [21] demonstrates a clear relationship between the performance of learned representations within SSL paradigms and the severity of distribution shifts and corruptions, and highlights the critical impact of distribution shifts and image corruptions on the performance and resilience of SSL methods. Similarly, Zhong et al. [91] conduct robustness tests to assess" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.756 + ], + "angle": 0, + "content": "the behavioral differences between contrastive and supervised learning under changes in downstream or pre-training data distributions, while also exploring the effects of data augmentation and feature space characteristics. Kowalczuk et al. [52] conducts a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks, revealing the need for broader enhancements in encoder robustness. Goldblum et al. [34] benchmarks diverse pretrained models across multiple computer vision tasks, finding that supervised convolutional neural networks still outperform newer architectures on most metrics, while revealing self-supervised learning backbones show competitive potential when compared under equivalent conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Unlike prior studies that primarily focus on individual SSL methods, specific tasks, or limited adversarial scenarios, our work provides a comprehensive, unified benchmark across multiple SSL paradigms, architectures, and tasks—including classification, transfer learning, segmentation, and detection—under a diverse set of adversarial attacks, offering a broader and deeper understanding of adversarial robustness in SSL." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2315" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.284, + 0.109 + ], + "angle": 0, + "content": "3. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.118, + 0.221, + 0.133 + ], + "angle": 0, + "content": "3.1. SSL Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.483, + 0.342 + ], + "angle": 0, + "content": "While numerous SSL approaches have been proposed [66], we focus on a subset of widely used models due to computational constraints: Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]. We primarily use ResNet50 [38] backbones, as most SSL checkpoints are released in this format, with DINO and MoCoV3 also offering ViT [28] variants. For BYOL, DINO, MoCoV3, SimCLR, and SwAV, complete model checkpoints were available. In contrast, Barlow Twins and VICReg only provided backbone weights, requiring linear evaluation via official code, which led to a slight drop in performance. For comparison, we also include a supervised ResNet50 baseline from PyTorch [68]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.383, + 0.379 + ], + "angle": 0, + "content": "3.2. ImageNet and Transfer Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.659 + ], + "angle": 0, + "content": "We use the benchmark suite introduced in the transfer learning study [45], which encompasses the target datasets like FGVC Aircraft [60], Caltech-101 [33], Stanford Cars [53], CIFAR 10 [54], CIFAR 100 [54], DTD [22], Oxford 102 Flowers [22], and Food-101 [8]. We follow Ericsson et al. [29] for both linear evaluation and fine-tuning of these datasets. We prioritized linear evaluation in our analysis as the backbone remains frozen during this process, allowing for a more equitable comparison of objectives within this setup. We apply the same adversarial techniques to ImageNet and transfer learning: Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). In brief, instance-based methods generate unique perturbations for each image, while UAP creates a single perturbation that applies across the entire dataset. Comprehensive details and categorizations of all attack methods—such as white-box, black-box, gradient-free, among others—are provided in Section 1 of the Supp Mat.." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.233, + 0.695 + ], + "angle": 0, + "content": "3.3. Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For segmentation, we use the Pascal VOC 2012 dataset [31] and CityScapes [23] dataset, training a DeepLabV3+ model [15]. To conduct the attacks, we follow the setup from Rony et al. [73], utilizing Alma [73], Asma [73], DAG [86], DDN [73], FGSM [35], FMN [70], and PGD [59]. While our primary metric is the mean Intersection Over Union (mIOU), we also report the Attack Pixel Success Rate (APSR) introduced by [73]. Although our main focus is on using a frozen backbone, we also perform training following the standard procedure. Our CityScapes results include only the frozen backbone approach, while our Pascal VOC results include both frozen and unfrozen backbone configurations." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Table 1. Performance of various models on ImageNet, Transfer Learning, Segmentation (Pascal VOC), and Detection (INRIA Person) tasks with frozen backbones, showing original (Orig.) and adversarial (Adv.) scores with performance drops in red." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.156, + 0.905, + 0.263 + ], + "angle": 0, + "content": "
ModelImageNetTransfer LearningSegmentationDetection
Orig.Adv.Orig.Adv.Orig.Adv.Orig.Adv.
Barlow Twins71.236.5 ↓49%80.337.8 ↓54%76.920.5 ↓73%88.421.9 ↓75%
BYOL74.638.9 ↓48%78.736.6 ↓55%76.719.0 ↓75%87.417.3 ↓80%
DINO75.334.2 ↓55%80.735.6 ↓57%77.018.9 ↓76%87.622.0 ↓75%
MoCoV374.640.5 ↓46%80.540.3 ↓51%76.219.9 ↓74%87.318.5 ↓79%
SimCLR68.932.1 ↓53%73.133.0 ↓57%75.619.3 ↓74%87.713.1 ↓85%
Supervised76.134.5 ↓55%74.631.2 ↓60%74.216.5 ↓78%86.118.0 ↓79%
SwAV75.332.6 ↓57%79.232.4 ↓60%76.519.2 ↓75%86.620.5 ↓76%
VICReg71.337.0 ↓48%79.937.7 ↓54%77.920.5 ↓74%88.414.0 ↓84%
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.288, + 0.625, + 0.303 + ], + "angle": 0, + "content": "3.4. Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.553 + ], + "angle": 0, + "content": "For object detection, we utilized the INRIA Person [24] and CoCo [57] datasets, and trained a Faster R-CNN [72]. To perform adversarial attacks, we followed the setup described by [44], employing the Transfer-based Self-Ensemble Attack (T-SEA). The T-SEA attack can be deployed using various methods and optimizers. In our experiments, we employed BIM [44], MIM [27], PGD [59], and Optim [44] methods. Additionally, we explored simpler methods that rely on common optimizers, such as Adam [51], SGD, and Nesterov [63]. Throughout our evaluation, we report the mean average precision (mAP) scores as the primary performance metric. While our primary focus was on employing a frozen backbone, we also conducted training experiments following the standard training procedures for comparative analysis. Our COCO results include only the frozen backbone approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.729, + 0.585 + ], + "angle": 0, + "content": "4. Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.701 + ], + "angle": 0, + "content": "In this section, we present our experimental findings on ImageNet, transfer learning, and detection, and discuss each in turn. Figure 1 and table 1 summarize the performance of various SSL models compared to supervised learning across our main evaluation tasks in the frozen backbone setup. While we address the results individually, the full detailed results are provided in Section 4 of the Supp. Mat.." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.627, + 0.729 + ], + "angle": 0, + "content": "4.1. ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "SSL vs Supervised. Most robustness studies on contrastive learning [42, 48, 49, 64, 87, 90] focus on small datasets like CIFAR10 [54] and primarily evaluate robustness using adversarial attacks such as FGSM [35] and PGD [59]. While computational constraints explain the reluctance to scale to larger datasets like ImageNet [74], many evaluations inadequately incorporate Universal Adversarial Perturbations (UAP). Our findings, as shown in Figure 2 contradict previous research by Gupta et al. [37]. Under IAA, MoCoV3 demonstrates the strongest robustness (54% drop), while SimCLR shows the weakest performance (61% drop). For" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2316" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.349 + ], + "angle": 0, + "content": "UAP attacks, MoCoV3 again leads (38% drop), while the supervised model and SwAV both show 50% drops. Notably, our results challenge the conclusion about contrastive vs. non-contrastive methods. MoCoV3, a contrastive model, consistently demonstrates the highest adversarial robustness with ResNet architecture, while DINO, which was claimed to perform better due to its non-contrastive nature, shows the weakest IAA robustness in our ResNet evaluation (64% drop). Our data reveals that Barlow, BYOL, MoCoV3, and VICReg all demonstrate comparable resilience against IAA (around 54% drop), contradicting the simple categorization of robustness based on contrastive versus non-contrastive approaches. Our comprehensive evaluation using diverse attack methods demonstrates that the relationship between self-supervised learning approaches and adversarial robustness is more complex than previously suggested. The full results are in Section 4.1 of the Supp. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.482, + 0.426 + ], + "angle": 0, + "content": "What makes MoCoV3 robust? Although MoCoV3 and SimCLR both utilize the InfoNCE [77, 81] objective, there is a stark contrast in their adversarial robustness. To understand this disparity, we evaluate MoCoV1 [39], MoCoV2 [19], and MoCoV3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.482, + 0.535 + ], + "angle": 0, + "content": "A brief MoCo History. MoCoV1 introduced a dynamic dictionary with a queue and momentum-updated encoder to maintain consistent negative samples. MoCoV2 enhanced this with a multi-layer projection head and stronger data augmentation. MoCoV3 further evolved by eliminating the memory bank and incorporating a prediction head similar to BYOL and SimSiam [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our analysis reveals a clear progression in robustness across MoCo versions. While MoCoV1 demonstrates limited resilience with a \\(71\\%\\) drop in overall adversarial accuracy, MoCoV2 shows significant improvement (52% drop), and MoCoV3 achieves the strongest performance (46% drop). The most dramatic enhancement occurs in UAP resistance, where MoCoV1's performance drops by \\(77\\%\\), compared to MoCoV2's \\(39\\%\\) and MoCoV3's \\(36\\%\\). The substantial improvement from MoCoV1 to MoCoV2 (7% in clean accuracy, 27 percentage points in UAP robustness) primarily stems from the non-linear projector, with data augmentation providing marginal benefits. While [46] suggests non-linear projectors aren't always essential, our results indicate they significantly enhance both performance and adversarial robustness. MoCoV3's superior performance over MoCoV2 (an additional \\(7\\%\\) in clean accuracy and 7 percentage points in UAP robustness) can be attributed to its prediction head and larger batch size. Unlike the transition from V1 to V2, MoCoV3 shows substantial improvements in both IAA (34% vs 25%) and UAP (48% vs 41%), highlighting the prediction head's critical role in enhancing overall robustness. Our findings suggest that momentum, a common feature in robust models like MoCoV3 and BYOL, significantly contributes to adversar" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ial resilience, while MoCoV2's performance more closely resembles that of SimCLR, which lacks this feature. The full results are in Section 4.6 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.139, + 0.748, + 0.153 + ], + "angle": 0, + "content": "Augmentations vs Algorithms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.905, + 0.276 + ], + "angle": 0, + "content": "Morningstar et al. [61] challenges the notion that SSL progress is primarily driven by algorithmic advancements, suggesting instead that augmentation diversity, along with data and model scale, play more critical roles. Their analysis argues that many algorithmic improvements, such as prediction networks or new loss functions, had minimal impact on downstream task performance compared to stronger augmentation techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.905, + 0.473 + ], + "angle": 0, + "content": "Our comprehensive analysis of the MoCo family evolution provides a more nuanced perspective on this debate. The substantial progression in adversarial robustness from MoCoV1 to MoCoV2 and further to MoCoV3 suggests that architectural innovations like non-linear projectors and prediction heads significantly impact robustness. MoCoV2's dramatic improvement over MoCoV1, particularly in UAP performance, indicates that the multi-layer projection head provides substantial benefits beyond mere augmentation changes. Similarly, MoCoV3's further enhancements in both IAA and UAP performance relative to MoCoV2 highlight the crucial role of the prediction head in overall robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.642 + ], + "angle": 0, + "content": "While our comparison lacks perfectly controlled baselines for augmentations across different objectives (with slight variations in augmentation between MoCoV2 and V3), this limitation stems from the unavailability of public checkpoints rather than our experimental design. Despite this constraint, the consistent improvements in adversarial performance strongly suggest that algorithmic innovations significantly contribute to adversarial robustness. Importantly, our findings demonstrate that higher clean accuracy doesn't automatically translate to improved robustness on ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.905, + 0.779 + ], + "angle": 0, + "content": "Although the checkpoints from [61] are not publicly available, we conducted an ablation study on augmentation types and batch sizes using BYOL which is the only model in our SSL pool that includes configurations with varied augmentations and batch sizes. For this analysis, we consider four distinct models: BYOL-NC (without color distortions), BYOL-CC (with only color and crop augmentations), and the standard BYOL models with batch sizes of 128 (BYOL-128) and 512 (BYOL-512)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We observe that batch size and augmentation choices have varying effects on the adversarial robustness of BYOL variants. For IAA robustness, both BYOL-NC and BYOLCC show similar performance (60% drop), while BYOL128, BYOL-512, and standard BYOL demonstrate slightly better resilience (57%, 56%, and 56% drops, respectively). These modest differences suggest that batch size and augmentation have a limited impact on instance-level attack ro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.158, + 0.105 + ], + "angle": 0, + "content": "bustness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.109, + 0.482, + 0.273 + ], + "angle": 0, + "content": "For UAP attacks, we see more substantial variations: BYOL-128 (41% drop) and BYOL-512 (43% drop) perform comparably, while BYOL-NC and BYOL-CC show notably weaker performance (46% and 48% drops). Standard BYOL demonstrates a 39% drop, which is marginally better than the batch size variants but substantially better than the limited-augmentation models. The 7-9 percentage point difference between standard BYOL and the limited-augmentation variants suggests that comprehensive augmentation strategies may contribute to improved UAP robustness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.502 + ], + "angle": 0, + "content": "Interestingly, we note that BYOL-CC performs slightly worse than BYOL-NC despite having more augmentations, though this difference is too small to draw meaningful conclusions. Similarly, the differences between BYOL-128, BYOL-512, and standard BYOL in UAP robustness are relatively minor and should be interpreted cautiously. Our results indicate that the relationship between augmentation and adversarial robustness is complex and that meaningful improvements likely require more than just incremental changes to batch size or augmentation strategies. While our data hints at potential benefits from comprehensive augmentation for UAP robustness, the effects are modest and require further investigation with more controlled experiments. The full results are in Section 4.7 of the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.656 + ], + "angle": 0, + "content": "ResNet vs ViT in Adversarial Robustness. While ViTs are generally seen as more robust than CNNs [62], Bai et al. [2], Pinto et al. [69] demonstrate that with the right training methods, CNNs [55] can achieve comparable robustness. Despite ViT's success [12, 20, 25, 28, 65], most SSL methods still use ResNet for validation. We examine MoCoV3 and DINO, as they are the only models that include ViT training, with our analysis covering both standard ViTs comparable to ResNet50 and larger ViT-B variants with approximately 4x more parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Our results reveal notable architectural differences. DINO performs significantly better with ViT architectures than ResNet, with DINO-ViT-B achieving \\(45.19\\%\\) adversarial average (42% drop from clean accuracy) compared to DINO-ResNet's \\(55\\%\\) drop. In contrast, MoCo struggles with transformer architectures, with MoCo-ViT showing poor performance at a \\(61\\%\\) drop, significantly worse than MoCoV3-ResNet at a \\(46\\%\\) drop. Even MoCoV3-ViT-B \\(51\\%\\) drop underperforms its ResNet counterpart despite having 4x more parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These differences are especially evident in UAP results, where DINO-ViT-B shows greater resilience (38% drop) than DINO-ResNet (44% drop), while MoCo-ViT performs poorly (26.31%, 64% drop) compared to MoCoV3-ResNet (36% drop). These findings suggest that the interaction between self-supervised approaches and model architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "significantly impacts adversarial robustness, with DINO benefiting from the ViT architecture while MoCo struggles with transformer models. The full results are in Section 4.6 of the Supp. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Lastly, we have evaluated DINOv2's [65] small and base and MAE's [40] base versions. We discuss them in the section 2 in Supp. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.905, + 0.547 + ], + "angle": 0, + "content": "Impact of Training Duration. SSL models tend to demonstrate better performance as training epochs increase [11, 16, 20]. However, due to computational constraints, many models are reported with different numbers of epochs. This prompts the question of whether longer training durations enhance or reduce adversarial robustness. As ViT models do not have checkpoints at various epochs, we focus on ResNet-based SSL models, specifically SwAV and MoCoV3, which offer multiple checkpoints throughout the training process. We find that both SwAV and MoCo show very marginal improvement of about \\(1\\%\\) on IAA across various epochs, which is minimal compared to the rise in original accuracy. In contrast, both methods exhibit a modest increase in UAP performance after surpassing 100 epochs, with SwAV improving from \\(33\\%\\) to \\(40\\%\\) and MoCoV3 from \\(41\\%\\) to \\(48\\%\\). Overall, our results suggest that despite differences in reported checkpoints, robustness generally remains largely stable during training for IAA, with slightly more noticeable but still modest gains for UAP. This indicates that training duration has a limited impact on adversarial robustness, even as clean accuracy continues to improve with extended training. The full results are in Section 4.5 of the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.694, + 0.575 + ], + "angle": 0, + "content": "4.2. Transfer Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.703 + ], + "angle": 0, + "content": "In this section, we analyze how adversarial robustness transfers from ImageNet pre-training to downstream tasks. We examine whether vulnerability patterns established during pre-training persist when models are evaluated through linear probing or fine-tuning across various ResNet-based self-supervised and supervised learning approaches. This analysis quantifies robustness transfer relationships through both Spearman rank and Pearson correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "LINEAR. Our correlation analysis shows strong relationships between ImageNet and linear evaluation vulnerability. For performance drops, both Spearman and Pearson correlations are high across all attack types (Spearman—Adv: 0.93, UAP: 0.83, IAA: 0.79; Pearson—Adv: 0.94, UAP: 0.94, IAA: 0.88). These consistently high correlations suggest that robustness characteristics established during pre-training largely persist through linear probing. The performance drop analysis reveals method-specific robustness characteristics. MoCoV3 consistently exhibits smaller performance drops across both ImageNet (IAA: \\(54\\%\\), UAP: \\(36\\%\\)) and linear evaluation (IAA: \\(56\\%\\), UAP: \\(43\\%\\)), indicating its contrastive learning approach with mo" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "2318" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Component Contribution Analysis Across Datasets. Metrics include: Original Accuracy (Orig), Adversarial Accuracy (Adv) with relative performance drop percentage, Direction Ratio (D) and Magnitude Ratio (M). Ratio interpretation: Head dominant (<0.67), Balanced (0.67-1.5), Backbone dominant (>1.5)." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.143, + 0.88, + 0.281 + ], + "angle": 0, + "content": "
ModelCIFAR-10CIFAR-100ImageNetCIFAR-10 (FT)CIFAR-100 (FT)
OrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DM
Barlow92.333.0 ↓64%2.73.177.920.5 ↓74%1.45.071.242.4 ↓40%0.99.497.166.9 ↓31%1.52.484.645.0 ↓57%0.82.4
BYOL93.031.0 ↓67%1.64.278.219.0 ↓76%1.076.074.639.4 ↓47%0.912.596.967.0 ↓31%1.21.483.961.2 ↓27%0.72.7
DINO93.927.6 ↓71%3.23.379.716.0 ↓80%1.66.275.324.7 ↓67%1.214.196.977.8 ↓20%1.92.984.752.9 ↓38%1.794.6
MoCoV394.733.0 ↓65%2.51.780.219.3 ↓76%1.14.574.642.7 ↓43%0.710.996.972.0 ↓26%1.11.384.562.6 ↓26%0.72.6
SimCLR91.037.9 ↓58%1.51.973.019.5 ↓73%1.33.568.924.3 ↓65%1.07.797.267.3 ↓31%1.01.984.443.0 ↓50%0.62.0
Supervised91.442.8 ↓53%1.80.773.924.5 ↓67%1.31.576.138.8 ↓49%1.43.096.262.3 ↓35%2.00.582.659.3 ↓28%1.31.1
SwAV93.919.4 ↓79%3.13.079.411.1 ↓86%1.86.475.324.7 ↓67%1.217.496.879.9 ↓17%1.72.784.454.6 ↓35%1.03.0
VICReg92.833.0 ↓64%2.73.177.822.3 ↓71%1.34.371.342.4 ↓40%0.99.397.168.6 ↓29%1.42.484.343.4 ↓48%0.82.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.304, + 0.482, + 0.382 + ], + "angle": 0, + "content": "mentum encoders develops features with inherently better transferable robustness properties. In contrast, supervised pre-training and clustering-based approaches like SwAV experience more severe drops in both settings, suggesting these methods may create more brittle representations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.383, + 0.483, + 0.519 + ], + "angle": 0, + "content": "The high correlation in adversarial vulnerability between ImageNet and linear evaluation indicates that linear probing provides a reliable assessment of a pre-trained model's downstream robustness characteristics without requiring extensive adaptation. This finding has practical implications for model selection, suggesting that robustness evaluations on ImageNet can effectively predict linear transfer performance. The full results are in Section 4.8 of the Supp. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.524, + 0.483, + 0.901 + ], + "angle": 0, + "content": "FINETUNE. Fine-tuning reveals distinct transfer patterns by attack type. For IAA, the correlation between ImageNet and fine-tuned vulnerability nearly vanishes for performance drops (Spearman: 0.12, Pearson: 0.36), indicating that fine-tuning substantially reshapes defense against instance-specific attacks. Conversely, UAP vulnerability correlation remains high (Spearman: 0.88, Pearson: 0.93), suggesting that susceptibility to universal perturbations is more persistently encoded in the network regardless of parameter adaptation. This attack-dependent correlation disparity reveals that universal perturbation vulnerabilities, which exploit systematic weaknesses across the feature space, are deeply encoded in the network's architecture and learning approach, while instance-specific vulnerabilities are more malleable through fine-tuning. The absolute magnitude of performance drops diminishes considerably after fine-tuning across all methods (IAA: \\(49\\%\\) vs. \\(60\\%\\) in linear, UAP: \\(43\\%\\) vs. \\(49\\%\\) in linear), highlighting fine-tuning's effectiveness in mitigating vulnerability. MoCoV3 continues to demonstrate superior robustness with the smallest drops (IAA: \\(45\\%\\), UAP: \\(38\\%\\)), while BYOL shows dramatic improvement from ImageNet to fine-tuning for IAA attacks (from \\(56\\%\\) to \\(47\\%\\)). These findings suggest different mechanisms for robustness transfer: UAP vulnerability appears tied to fundamental architectural and al" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.44 + ], + "angle": 0, + "content": "gorithmic properties that persist across transfer paradigms, while instance-specific attack vulnerability depends more on the fine-tuning process than initial representation properties. This distinction may be valuable for practical applications, suggesting that pre-training method selection strongly impacts universal attack robustness, while defense against instance-specific attacks can be substantially improved through appropriate fine-tuning strategies. The full results are in Section 4.9 of the Supp. Mat.." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.905, + 0.577 + ], + "angle": 0, + "content": "Attributing Adversarial Vulnerability. Having established the correlations between ImageNet and transfer learning robustness patterns, we now seek to understand the underlying mechanisms causing these disparities across models and training paradigms. Specifically, we investigate whether adversarial vulnerability stems primarily from the backbone feature extractor or the classification head and how this attribution changes between linear probing and fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.905, + 0.718 + ], + "angle": 0, + "content": "Our analysis investigates the relationship between adversarial robustness and component-specific vulnerability under \\(\\mathrm{FGSM}_1\\) attack (detail in Supp. Mat.), focusing on two metrics: the Direction Ratio \\(D = \\frac{\\mathrm{mean}(1 - \\cos(\\mathbf{l}_{adv}))}{\\mathrm{mean}(1 - \\cos(\\mathbf{f},\\mathbf{f}_{adv}))}\\), which compares directional shifts in logits (l) versus backbone features (f), and the Magnitude Ratio \\(M = \\frac{\\mathrm{mean}(\\|\\mathbf{l} - \\mathbf{l}_{adv}\\|_2)}{\\mathrm{mean}(\\|\\mathbf{f} - \\mathbf{f}_{adv}\\|_2)}\\), quantifying relative sensitivity to perturbation magnitudes. We fit linear regression models using these ratios to predict adversarial accuracy drop across models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.84 + ], + "angle": 0, + "content": "For probed models, Direction Ratio strongly correlates with vulnerability \\((R^2 = 0.56 - 0.85)\\), suggesting that directional instability in the head dominates FGSM robustness. SwAV, with \\(D = 3.1\\) on CIFAR-10, suffers a \\(79\\%\\) accuracy drop, while SimCLR \\((D = 1.5)\\) drops \\(58\\%\\). This may arise from the linear head's limited capacity to compensate for adversarial noise, amplifying directional shifts in logit space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "After fine-tuning, this correlation weakens significantly \\((R^2 = 0.05 - 0.52)\\): SwAV achieves only a \\(17\\%\\) drop despite retaining \\(D = 1.7\\), implying that fine-tuning enables the head to stabilize predictions even under directional feature" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2319" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "shifts. This pattern aligns with our transfer learning analysis, where fine-tuning disrupted the correlation between ImageNet and downstream instance-specific attack robustness. Regression figures are in Section 2 of the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.273 + ], + "angle": 0, + "content": "As in Table 2, self-supervised models exhibit higher pre-fine-tuning Magnitude Ratios (e.g., SwAV \\( M = 17.4 \\) on ImageNet vs. supervised \\( M = 3.0 \\)), potentially due to their reliance on globally normalized feature spaces, where FGSM perturbations propagate more aggressively to logits. Fine-tuning reduces \\( M \\) across models (SwAV \\( M = 2.7 \\) post-FT), aligning with improved robustness, possibly by suppressing logit magnitude distortions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.482, + 0.41 + ], + "angle": 0, + "content": "These patterns are specific to FGSM, as its single-step gradient reliance may emphasize head-layer instability, whereas iterative attacks could exploit backbone vulnerabilities differently. While our analysis tentatively links directional logit shifts \\((D)\\) and magnitude sensitivity \\((M)\\) to robustness, dataset-dependent variations—such as weaker correlations on CIFAR-100 FT—highlight the need for broader evaluation across threat models to generalize these insights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.482, + 0.518 + ], + "angle": 0, + "content": "We also measured inter-class and intra-class distances for both probed and fine-tuned CIFAR-10 representations; however, these metrics did not yield meaningful correlations that explain the observed differences in adversarial robustness. The detailed regression figures, inter/intra-class distance measurements, and t-SNE [82] visualizations of feature spaces are provided in Section 3 of the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.53, + 0.233, + 0.545 + ], + "angle": 0, + "content": "4.3. Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.482, + 0.718 + ], + "angle": 0, + "content": "Unlike in classification, we observe no strong correlation between ImageNet robustness and segmentation performance across both PASCAL VOC and CityScapes datasets. Self-supervised approaches demonstrate competitive performance in clean conditions, with VICReg leading on PASCAL VOC (77.89 mIOU with frozen backbone) and SwAV on CityScapes (66.48 mIOU). All models suffer catastrophic performance degradation (74-84%) under adversarial attacks regardless of training methodology. This uniform degradation pattern suggests attacks primarily target segmentation modules rather than backbones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Our experiments with frozen versus unfrozen backbones reveal that frozen backbones generally achieve both higher clean performance and slightly better adversarial robustness on PASCAL VOC. The supervised model's exception to this pattern stems from using the standard MMSegmentation model for the unfrozen case due to reproduction challenges. These findings indicate that while SSL models produce competitive segmentation performance in clean conditions, they offer minimal advantage in terms of adversarial robustness for segmentation tasks, unlike their significant impact on object recognition. This highlights the need for robustness techniques specifically designed for segmenta" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "tion architectures rather than focusing solely on backbone improvements. The full results are in Section 4.2 of the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.151, + 0.625, + 0.165 + ], + "angle": 0, + "content": "4.4. Detection" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.174, + 0.905, + 0.309 + ], + "angle": 0, + "content": "Detection results show distinct patterns from recognition and segmentation tasks. On INRIA Person with unfrozen backbones, SwAV demonstrates the highest robustness (72% decrease), while the Supervised model shows the poorest performance (89% decrease). With frozen backbones, Barlow Twins and DINO lead in robustness (both 75% decrease), while VICReg becomes unexpectedly vulnerable (84% decrease), contradicting its recognition performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.905, + 0.403 + ], + "angle": 0, + "content": "COCO dataset results show improved robustness overall, with Barlow Twins maintaining the strongest performance (50% decrease) and the Supervised model remaining the least robust (67% decrease). This consistent weakness in the Supervised model suggests inherent robustness benefits from self-supervised pretraining." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.404, + 0.905, + 0.525 + ], + "angle": 0, + "content": "These findings indicate that task-specific architectures significantly influence adversarial robustness, and robustness in ImageNet classification doesn't necessarily transfer to detection tasks. This highlights the importance of task-specific evaluations and suggests that backbone architecture becomes less critical than overall model design for downstream applications. The full results are in Section 4.3 of the Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.543, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.905, + 0.827 + ], + "angle": 0, + "content": "Our evaluation shows that SSL models generally offer greater adversarial robustness than supervised counterparts in image classification, with MoCoV3 performing particularly well—likely due to its momentum encoder and prediction head. This robustness advantage is less evident in segmentation and detection, where task-specific architectures play a larger role. Architectural effects also vary by objective: DINO benefits from ViTs, while MoCoV3 performs better with CNNs. Extended training does not compromise robustness and may slightly improve it. We find that vulnerability to UAPs transfers more consistently across learning paradigms, while susceptibility to IAAs can be mitigated through fine-tuning. These results highlight the complex interplay between SSL objectives, architectures, and downstream tasks, underscoring the need for further study. We hope our findings support continued progress toward more robust visual representation learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.682, + 0.861 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "This work was funded by the KUIS AI Center at Koç University, Turkey." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2320" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[2] Yutong Bai, Jieru Mei, Alan Yuille, and Cihang Xie. Are transformers more robust than cnns?, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.482, + 0.229 + ], + "angle": 0, + "content": "[3] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.232, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[4] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari S. Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A cookbook of self-supervised learning. ArXiv, abs/2304.12210, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[6] Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian Hadfield, Jeff Clune, Tegan Maharaj, Frank Hutter, Atilim Gunes Baydin, Sheila McIlraith, Qiqi Gao, Ashwin Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Brauner, and Soren Mindermann. Managing extreme ai risks amid rapid progress. Science, 384(6698): 842-845, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[7] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In Computer Vision – ECCV 2014, pages 446–461, Cham, 2014. Springer International Publishing. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[9] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, Percy Liang, and John C. Duchi. Unlabeled data improves adversarial robustness, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[13] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[14] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. *ArXiv*, abs/2002.05709, 2020. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[17] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[19] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[21] Prakash Chandra Chhipa, Johan Rodahl Holmgren, Kanjar De, Rajkumar Saini, and Marcus Liwicki. Can self-supervised representation learning methods withstand distribution shifts and corruptions?, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[22] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild, 2013. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2321" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[24] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 886-893 vol. 1, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[25] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[26] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[27] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[28] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[29] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. How well do self-supervised models transfer?, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[30] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[31] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[32] Alhussein Fawzi, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Robustness of classifiers: from adversarial to random noise, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[33] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[34] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, Rama Chellappa, Andrew Gordon Wilson, and Tom Goldstein. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[35] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[36] Jean-Bastien Grill, Florian Strub, Florent Altch'e, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. ArXiv, abs/2006.07733, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[37] Rohit Gupta, Naveed Akhtar, Ajmal Mian, and Mubarak Shah. Contrastive self-supervised learning leads to higher adversarial susceptibility, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[41] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[42] Chih-Hui Ho and Nuno Vasconcelos. Contrastive learning with adversarial examples, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[43] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[44] Hao Huang, Ziyan Chen, Huanran Chen, Yongtao Wang, and Kevin Zhang. T-sea: Transfer-based self-ensemble attack on object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20514–20523, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[45] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes imagenet good for transfer learning?, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[46] Mark Ibrahim, David Klindt, and Randall Balestriero. Occam's razor for self supervised learning: What is sufficient to learn good representations?, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[47] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[48] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[49] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2322" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[52] Antoni Kowalczuk, Jan Dubinski, Atiyeh Ashari Ghomi, Yi Sui, George Stein, Jiapeng Wu, Jesse C. Cresswell, Franziska Boenisch, and Adam Dziedzic. Benchmarking robust self-supervised learning across diverse downstream tasks, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[53] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[54] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[55] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[56] Ang Li, Yifei Wang, Yiwen Guo, and Yisen Wang. Adversarial examples are not real features, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[57] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.384, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[58] Rundong Luo, Yifei Wang, and Yisen Wang. Rethinking the effect of data augmentation in adversarial contrastive learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[59] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[60] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.529, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[61] Warren Morningstar, Alex Bijamov, Chris Duvarney, Luke Friedman, Neha Kalibhat, Luyang Liu, Philip Mansfield, Renan Rojas-Gomez, Karan Singhal, Bradley Green, and Sushant Prakash. Augmentations vs algorithms: What works in self-supervised learning, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[62] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 262-271, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[63] Yurii Nesterov. A method for solving the convex programming problem with convergence rate \\( o(1 / k^2) \\). Proceedings of the USSR Academy of Sciences, 269:543-547, 1983. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[64] A. Tuan Nguyen, Ser Nam Lim, and Philip Torr. Task-agnostic robust representation learning, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[65] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[66] Utku Ozbulak, Hyun Jung Lee, Beril Boga, Esla Timothy Anzaku, Homin Park, Arnout Van Messem, Wesley De Neve," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "and Joris Vankerschaver. Know your self-supervised learning: A survey on image-based generative and discriminative training, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[67] Serdar Ozsoy, Shadi Hamdan, Sercan Ö. Arik, Deniz Yuret, and Alper T. Erdogan. Self-supervised learning with an information maximization criterion, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[68] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[69] Francesco Pinto, Philip H. S. Torr, and Puneet K. Dokania. An impartial take to the cnn vs transformer robustness contest, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[70] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.905, + 0.409 + ], + "angle": 0, + "content": "[71] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[72] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[73] Jérôme Rony, Jean-Christophe Pesquet, and Ismail Ben Ayed. Proximal splitting adversarial attacks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[74] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[75] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Mady. Adversarily robust generalization requires more data, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[76] Ali Shafahi, W. Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable?, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[77] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Neural Information Processing Systems, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[78] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[79] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2323" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[81] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[82] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[83] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[84] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[85] Dongxian Wu, Shu tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.294, + 0.482, + 0.334 + ], + "angle": 0, + "content": "[86] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[87] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan Yuille, and Quoc V. Le. Adversarial examples improve image recognition, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[88] Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, and Mohan Kankanhalli. Efficient adversarial contrastive learning via robustness-aware coreset selection, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[89] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[90] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D. Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.482, + 0.562 + ], + "angle": 0, + "content": "[91] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning?, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2324" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_origin.pdf b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1c82cc6a95bece1ce1887be7bbd01b48e140467 --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/793fb0cc-d29e-4453-8e4e-2aca5e095268_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd331b0c181ea2c115691af7170cb8dc66ee35fa5301d56f96b5c4b39b33e852 +size 315911 diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/full.md b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e31860a61090dd27e0651defc800480a88aab729 --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/full.md @@ -0,0 +1,301 @@ +# Adversarial Robustness of Discriminative Self-Supervised Learning in Vision + +Ömer Veysel Çağatan $^{1,2}$ 1Ömer Faruk Tal ${}^{1}\mathrm{M}$ . Emre Gürsoy + $^{1}$ Department of Computer Engineering, Koç University + $^{2}$ KUIS AI Center, Koç University + +ocagatan19@ku.edu.tr, otal19@ku.edu.tr, emregursoy@ku.edu.tr + +# Abstract + +Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems. + +# 1. Introduction + +Self-supervised learning (SSL) [4], particularly discriminative approaches, has emerged as a foundational method for training models with remarkable capabilities in areas such as language [80], vision [65], and decision-making [50]. As these models become increasingly widespread and integrated into various applications, ensuring their reliability and safety has become a critical concern [6, 7]. + +One particular challenge is the surprising vulnerability of deep learning models to adversarial examples, where slight input alterations can significantly impact model performance [35, 78]. This phenomenon has sparked significant debate, seeking to understand and mitigate these vulnerabilities [3, 32, 75, 76, 79, 83-85]. One prominent theory [47] suggests that adversarial examples arise from the model's sensitivity to non-robust features in the input data. + +According to this view, both robust (stable) and non-robust (vulnerable) features contribute to classification, with adversarial attacks manipulating the latter to cause misclassification. However, this theory, developed primarily in the context of supervised learning, faces challenges when extended to other self-supervised paradigms. [56] indicates that non-robust features are less effective in SSL methods such as contrastive learning [16], masked image modeling [40], or diffusion models [43]. This discrepancy suggests that non-robust features may lack the transferability across learning paradigms that robust or natural features possess. Thus, it becomes essential to systematically evaluate and compare how different SSL approaches respond to adversarial attacks, particularly given the theoretical evidence suggesting their feature representations may differ fundamentally from supervised models. + +These theoretical insights into how adversarial examples affect different learning paradigms highlight several critical gaps in our understanding of SSL's adversarial robustness. Notwithstanding the progress made in understanding the adversarial robustness of SSL, particularly contrastive learning, which we extensively discuss in Section 2, several key questions remain unresolved. First, with the wide variety of self-supervised representations available, employing different pretext tasks and data augmentations, which approaches demonstrate the greatest adversarial robustness? This remains unclear since most methods don't provide any results on adversarial robustness unless it is a specific focus of the proposed approach. Secondly, robustness is typically assessed by the model's accuracy on the pretraining dataset. Still, its adversarial impact on transfer learning or downstream tasks like detection and segmentation has not been thoroughly investigated [52]. + +The choice of model architecture also raises questions about robustness. Standard vision SSL pretraining typically utilizes a ResNet [38] as the backbone, but more recently, larger and more powerful models [12, 20, 65] have been developed using vision transformers [28]. This leads to the question: Which architecture demonstrates greater robustness under the same SSL objective and with comparable + +parameter sizes? + +Another factor to consider is the training duration. State-of-the-art SSL models are trained for longer durations compared to their supervised counterparts. Several studies indicate that this extended training consistently enhances performance, raising the question of whether this might compromise the models' adversarial robustness. + +While previous work has examined aspects of adversarial robustness in SSL, our study provides the first comprehensive cross-model comparison across multiple tasks, architectures, and training regimes. We assess seven different SSL models (Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]) alongside a supervised model against various adversarial attacks on ImageNet [74] and nine other image-recognition datasets. We also evaluate their robustness in segmentation and detection tasks. Our investigation addresses the following key questions: + +1. How does the adversarial robustness of various SSL models compare to that of supervised models on ImageNet? + +SSL models consistently demonstrate greater adversarial robustness than supervised models on ImageNet. Noncontrastive methods show particular resilience against IAA [13] attacks, while all SSL approaches exhibit strong resistance to UAP [14], with MoCoV3 demonstrating the strongest overall performance. + +2. Does SSL robustness transfer to downstream tasks like transfer learning, segmentation, and detection? + +The robustness advantages transfer effectively to classification tasks via linear probing and fine-tuning. However, in segmentation and detection, all models exhibit similar vulnerability regardless of pretraining methodology, suggesting task-specific architectural components may override backbone robustness properties. + +3. How does model architecture influence adversarial robustness under the same SSL objective? + +Architecture impact is highly objective dependent. MoCoV3 shows reduced robustness with Vision Transformers, whereas DINO demonstrates improved performance with ViT compared to ResNet, challenging the notion that architectural effects are uniform across SSL paradigms. Additionally, we have evaluated DINOv2 [65] and MAE [40] that do not demonstrate a consistently high or low level of robustness. + +4. Does extending training duration compromise adversarial robustness in SSL models? + +Extended training either maintains or slightly enhances adversarial performance. For UAP attacks, performance improves meaningfully after more epochs in both SwAV and MoCoV3, indicating that longer training periods do not create a performance-robustness tradeoff. + +# 2. Related Work + +Self Supervised Learning Self-supervised learning(SSL) seeks to extract meaningful and general representations from unlabeled data by leveraging pretext tasks. These tasks can vary, such as predicting the next word [71] or neighboring words [26] in a text, reconstructing masked sections of an image [40], or ensuring that two different perspectives of the same image result in similar visual representations [16]. + +Avoiding collapse is a key challenge in SSL for computer vision, and various methods can be classified based on how they address this issue. Contrastive approaches like SimCLR [16] and MoCo [19, 20, 39] use an objective that pushes apart representations of different inputs (negative samples) while bringing together those of the same input (positive samples). The performance and scalability of these methods heavily depend on the number and selection of negative samples. In another category, distillation methods such as BYOL [36], SimSiam [18], and DINO [12], prevent collapse by introducing asymmetry between different encoder branches and employing algorithmic adjustments. Additional SSL techniques, including DeepCluster [10], SeLa [1], and SwAV [11], enforce a clustering structure in the feature space to avoid constant representations. Meanwhile, methods like Barlow Twins [89], Whitening MSE (W-MSE) [30], VICReg [5], CorInfoMax [67] prevent collapse by using feature decorrelation. + +Adversarial Self-Supervised Learning While self-supervised learning (SSL) has outperformed supervised training [16], numerous studies highlight that contrastive learning remains susceptible to adversarial attacks when transferring the learned features to downstream classification tasks [42, 49]. To improve the robustness of contrastive learning, adversarial training has been adapted to self-supervised settings. In the absence of labels, adversarial examples are generated by maximizing the contrastive loss with respect to all input samples. Several prior works, such as ACL [48], RoCL [49], and CLAE [42], adopt this approach. Additionally, ACL incorporates the dual-BN technique [87] to further enhance performance. DeACL [90] introduces a two-stage approach, distilling a standard pretrained encoder through adversarial training. Nguyen et al. [64] establishes an upper bound on the adversarial loss of a prediction model, which is based on the learned representations, for any downstream task. This upper bound is determined using the model's loss on clean data and a robustness regularization term, which helps make the prediction model more resistant to adversarial attacks. [37] demonstrates that adversarial sensitivity stems from the uniform distribution of data representations on a unit hypersphere in the representation space. The presence of false negative pairs during training contributes to this effect, increasing the model's vulnerability to input perturbations. + +Although self-supervised adversarial training has made + +![](images/686b75e84ddcb7e1a609561bbf037aa119ae94a72acec203dc338d2f666d0923.jpg) +Figure 1. Performance scores for tasks such as ImageNet classification, transfer learning (with linear probing), segmentation, and detection (both with frozen backbones) are shown with the percentage drop in adversarial robustness. The shaded regions indicate the $95\%$ confidence interval around the regression line. Note the consistent pattern of higher robustness (lower percentage drop) among SSL models compared to supervised approaches in classification tasks. + +![](images/a0d1b6b3b33dddc23a25c85fce4056a4fc5765fd7ed6c945de509913f41e7d67.jpg) +Figure 2. Averaged scores of SSL models on ImageNet across various attack types, including Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). Adv Avg refers to the average score across all attacks combined. The shaded regions indicate the $95\%$ confidence interval around the regression line. + +progress, it still does not match the performance of supervised methods. Luo et al. [58] suggests that this shortfall is due to data augmentation and proposes a dynamic data augmentation scheduler to achieve comparable results to supervised training. Xu et al. [88] efficiently applies ACL on the ImageNet [74] to obtain a robust representation using robustness-aware core set selection. + +# Robustness of Self-Supervised Learning + +[41] found that incorporating an extra self-supervised task in a multi-task framework can enhance the adversarial robustness of supervised models. In a similar vein, Carmon et al. [9] discovered that using additional unlabeled data also strengthens the model's adversarial resilience. Furthermore, Chen et al. [17] created robust variants of pretext-based SSL tasks, showing that their integration with robust fine-tuning leads to a notable increase in robustness compared to standard adversarial training. Chhipa et al. [21] demonstrates a clear relationship between the performance of learned representations within SSL paradigms and the severity of distribution shifts and corruptions, and highlights the critical impact of distribution shifts and image corruptions on the performance and resilience of SSL methods. Similarly, Zhong et al. [91] conduct robustness tests to assess + +the behavioral differences between contrastive and supervised learning under changes in downstream or pre-training data distributions, while also exploring the effects of data augmentation and feature space characteristics. Kowalczuk et al. [52] conducts a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks, revealing the need for broader enhancements in encoder robustness. Goldblum et al. [34] benchmarks diverse pretrained models across multiple computer vision tasks, finding that supervised convolutional neural networks still outperform newer architectures on most metrics, while revealing self-supervised learning backbones show competitive potential when compared under equivalent conditions. + +Unlike prior studies that primarily focus on individual SSL methods, specific tasks, or limited adversarial scenarios, our work provides a comprehensive, unified benchmark across multiple SSL paradigms, architectures, and tasks—including classification, transfer learning, segmentation, and detection—under a diverse set of adversarial attacks, offering a broader and deeper understanding of adversarial robustness in SSL. + +# 3. Experimental Setup + +# 3.1. SSL Models + +While numerous SSL approaches have been proposed [66], we focus on a subset of widely used models due to computational constraints: Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]. We primarily use ResNet50 [38] backbones, as most SSL checkpoints are released in this format, with DINO and MoCoV3 also offering ViT [28] variants. For BYOL, DINO, MoCoV3, SimCLR, and SwAV, complete model checkpoints were available. In contrast, Barlow Twins and VICReg only provided backbone weights, requiring linear evaluation via official code, which led to a slight drop in performance. For comparison, we also include a supervised ResNet50 baseline from PyTorch [68]. + +# 3.2. ImageNet and Transfer Learning + +We use the benchmark suite introduced in the transfer learning study [45], which encompasses the target datasets like FGVC Aircraft [60], Caltech-101 [33], Stanford Cars [53], CIFAR 10 [54], CIFAR 100 [54], DTD [22], Oxford 102 Flowers [22], and Food-101 [8]. We follow Ericsson et al. [29] for both linear evaluation and fine-tuning of these datasets. We prioritized linear evaluation in our analysis as the backbone remains frozen during this process, allowing for a more equitable comparison of objectives within this setup. We apply the same adversarial techniques to ImageNet and transfer learning: Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). In brief, instance-based methods generate unique perturbations for each image, while UAP creates a single perturbation that applies across the entire dataset. Comprehensive details and categorizations of all attack methods—such as white-box, black-box, gradient-free, among others—are provided in Section 1 of the Supp Mat.. + +# 3.3. Segmentation + +For segmentation, we use the Pascal VOC 2012 dataset [31] and CityScapes [23] dataset, training a DeepLabV3+ model [15]. To conduct the attacks, we follow the setup from Rony et al. [73], utilizing Alma [73], Asma [73], DAG [86], DDN [73], FGSM [35], FMN [70], and PGD [59]. While our primary metric is the mean Intersection Over Union (mIOU), we also report the Attack Pixel Success Rate (APSR) introduced by [73]. Although our main focus is on using a frozen backbone, we also perform training following the standard procedure. Our CityScapes results include only the frozen backbone approach, while our Pascal VOC results include both frozen and unfrozen backbone configurations. + +Table 1. Performance of various models on ImageNet, Transfer Learning, Segmentation (Pascal VOC), and Detection (INRIA Person) tasks with frozen backbones, showing original (Orig.) and adversarial (Adv.) scores with performance drops in red. + +
ModelImageNetTransfer LearningSegmentationDetection
Orig.Adv.Orig.Adv.Orig.Adv.Orig.Adv.
Barlow Twins71.236.5 ↓49%80.337.8 ↓54%76.920.5 ↓73%88.421.9 ↓75%
BYOL74.638.9 ↓48%78.736.6 ↓55%76.719.0 ↓75%87.417.3 ↓80%
DINO75.334.2 ↓55%80.735.6 ↓57%77.018.9 ↓76%87.622.0 ↓75%
MoCoV374.640.5 ↓46%80.540.3 ↓51%76.219.9 ↓74%87.318.5 ↓79%
SimCLR68.932.1 ↓53%73.133.0 ↓57%75.619.3 ↓74%87.713.1 ↓85%
Supervised76.134.5 ↓55%74.631.2 ↓60%74.216.5 ↓78%86.118.0 ↓79%
SwAV75.332.6 ↓57%79.232.4 ↓60%76.519.2 ↓75%86.620.5 ↓76%
VICReg71.337.0 ↓48%79.937.7 ↓54%77.920.5 ↓74%88.414.0 ↓84%
+ +# 3.4. Detection + +For object detection, we utilized the INRIA Person [24] and CoCo [57] datasets, and trained a Faster R-CNN [72]. To perform adversarial attacks, we followed the setup described by [44], employing the Transfer-based Self-Ensemble Attack (T-SEA). The T-SEA attack can be deployed using various methods and optimizers. In our experiments, we employed BIM [44], MIM [27], PGD [59], and Optim [44] methods. Additionally, we explored simpler methods that rely on common optimizers, such as Adam [51], SGD, and Nesterov [63]. Throughout our evaluation, we report the mean average precision (mAP) scores as the primary performance metric. While our primary focus was on employing a frozen backbone, we also conducted training experiments following the standard training procedures for comparative analysis. Our COCO results include only the frozen backbone approach. + +# 4. Results and Discussion + +In this section, we present our experimental findings on ImageNet, transfer learning, and detection, and discuss each in turn. Figure 1 and table 1 summarize the performance of various SSL models compared to supervised learning across our main evaluation tasks in the frozen backbone setup. While we address the results individually, the full detailed results are provided in Section 4 of the Supp. Mat.. + +# 4.1. ImageNet + +SSL vs Supervised. Most robustness studies on contrastive learning [42, 48, 49, 64, 87, 90] focus on small datasets like CIFAR10 [54] and primarily evaluate robustness using adversarial attacks such as FGSM [35] and PGD [59]. While computational constraints explain the reluctance to scale to larger datasets like ImageNet [74], many evaluations inadequately incorporate Universal Adversarial Perturbations (UAP). Our findings, as shown in Figure 2 contradict previous research by Gupta et al. [37]. Under IAA, MoCoV3 demonstrates the strongest robustness (54% drop), while SimCLR shows the weakest performance (61% drop). For + +UAP attacks, MoCoV3 again leads (38% drop), while the supervised model and SwAV both show 50% drops. Notably, our results challenge the conclusion about contrastive vs. non-contrastive methods. MoCoV3, a contrastive model, consistently demonstrates the highest adversarial robustness with ResNet architecture, while DINO, which was claimed to perform better due to its non-contrastive nature, shows the weakest IAA robustness in our ResNet evaluation (64% drop). Our data reveals that Barlow, BYOL, MoCoV3, and VICReg all demonstrate comparable resilience against IAA (around 54% drop), contradicting the simple categorization of robustness based on contrastive versus non-contrastive approaches. Our comprehensive evaluation using diverse attack methods demonstrates that the relationship between self-supervised learning approaches and adversarial robustness is more complex than previously suggested. The full results are in Section 4.1 of the Supp. Mat.. + +What makes MoCoV3 robust? Although MoCoV3 and SimCLR both utilize the InfoNCE [77, 81] objective, there is a stark contrast in their adversarial robustness. To understand this disparity, we evaluate MoCoV1 [39], MoCoV2 [19], and MoCoV3. + +A brief MoCo History. MoCoV1 introduced a dynamic dictionary with a queue and momentum-updated encoder to maintain consistent negative samples. MoCoV2 enhanced this with a multi-layer projection head and stronger data augmentation. MoCoV3 further evolved by eliminating the memory bank and incorporating a prediction head similar to BYOL and SimSiam [18]. + +Our analysis reveals a clear progression in robustness across MoCo versions. While MoCoV1 demonstrates limited resilience with a $71\%$ drop in overall adversarial accuracy, MoCoV2 shows significant improvement (52% drop), and MoCoV3 achieves the strongest performance (46% drop). The most dramatic enhancement occurs in UAP resistance, where MoCoV1's performance drops by $77\%$ , compared to MoCoV2's $39\%$ and MoCoV3's $36\%$ . The substantial improvement from MoCoV1 to MoCoV2 (7% in clean accuracy, 27 percentage points in UAP robustness) primarily stems from the non-linear projector, with data augmentation providing marginal benefits. While [46] suggests non-linear projectors aren't always essential, our results indicate they significantly enhance both performance and adversarial robustness. MoCoV3's superior performance over MoCoV2 (an additional $7\%$ in clean accuracy and 7 percentage points in UAP robustness) can be attributed to its prediction head and larger batch size. Unlike the transition from V1 to V2, MoCoV3 shows substantial improvements in both IAA (34% vs 25%) and UAP (48% vs 41%), highlighting the prediction head's critical role in enhancing overall robustness. Our findings suggest that momentum, a common feature in robust models like MoCoV3 and BYOL, significantly contributes to adversar + +ial resilience, while MoCoV2's performance more closely resembles that of SimCLR, which lacks this feature. The full results are in Section 4.6 of the supplementary material. + +# Augmentations vs Algorithms. + +Morningstar et al. [61] challenges the notion that SSL progress is primarily driven by algorithmic advancements, suggesting instead that augmentation diversity, along with data and model scale, play more critical roles. Their analysis argues that many algorithmic improvements, such as prediction networks or new loss functions, had minimal impact on downstream task performance compared to stronger augmentation techniques. + +Our comprehensive analysis of the MoCo family evolution provides a more nuanced perspective on this debate. The substantial progression in adversarial robustness from MoCoV1 to MoCoV2 and further to MoCoV3 suggests that architectural innovations like non-linear projectors and prediction heads significantly impact robustness. MoCoV2's dramatic improvement over MoCoV1, particularly in UAP performance, indicates that the multi-layer projection head provides substantial benefits beyond mere augmentation changes. Similarly, MoCoV3's further enhancements in both IAA and UAP performance relative to MoCoV2 highlight the crucial role of the prediction head in overall robustness. + +While our comparison lacks perfectly controlled baselines for augmentations across different objectives (with slight variations in augmentation between MoCoV2 and V3), this limitation stems from the unavailability of public checkpoints rather than our experimental design. Despite this constraint, the consistent improvements in adversarial performance strongly suggest that algorithmic innovations significantly contribute to adversarial robustness. Importantly, our findings demonstrate that higher clean accuracy doesn't automatically translate to improved robustness on ImageNet. + +Although the checkpoints from [61] are not publicly available, we conducted an ablation study on augmentation types and batch sizes using BYOL which is the only model in our SSL pool that includes configurations with varied augmentations and batch sizes. For this analysis, we consider four distinct models: BYOL-NC (without color distortions), BYOL-CC (with only color and crop augmentations), and the standard BYOL models with batch sizes of 128 (BYOL-128) and 512 (BYOL-512). + +We observe that batch size and augmentation choices have varying effects on the adversarial robustness of BYOL variants. For IAA robustness, both BYOL-NC and BYOLCC show similar performance (60% drop), while BYOL128, BYOL-512, and standard BYOL demonstrate slightly better resilience (57%, 56%, and 56% drops, respectively). These modest differences suggest that batch size and augmentation have a limited impact on instance-level attack ro + +bustness. + +For UAP attacks, we see more substantial variations: BYOL-128 (41% drop) and BYOL-512 (43% drop) perform comparably, while BYOL-NC and BYOL-CC show notably weaker performance (46% and 48% drops). Standard BYOL demonstrates a 39% drop, which is marginally better than the batch size variants but substantially better than the limited-augmentation models. The 7-9 percentage point difference between standard BYOL and the limited-augmentation variants suggests that comprehensive augmentation strategies may contribute to improved UAP robustness. + +Interestingly, we note that BYOL-CC performs slightly worse than BYOL-NC despite having more augmentations, though this difference is too small to draw meaningful conclusions. Similarly, the differences between BYOL-128, BYOL-512, and standard BYOL in UAP robustness are relatively minor and should be interpreted cautiously. Our results indicate that the relationship between augmentation and adversarial robustness is complex and that meaningful improvements likely require more than just incremental changes to batch size or augmentation strategies. While our data hints at potential benefits from comprehensive augmentation for UAP robustness, the effects are modest and require further investigation with more controlled experiments. The full results are in Section 4.7 of the supplementary material. + +ResNet vs ViT in Adversarial Robustness. While ViTs are generally seen as more robust than CNNs [62], Bai et al. [2], Pinto et al. [69] demonstrate that with the right training methods, CNNs [55] can achieve comparable robustness. Despite ViT's success [12, 20, 25, 28, 65], most SSL methods still use ResNet for validation. We examine MoCoV3 and DINO, as they are the only models that include ViT training, with our analysis covering both standard ViTs comparable to ResNet50 and larger ViT-B variants with approximately 4x more parameters. + +Our results reveal notable architectural differences. DINO performs significantly better with ViT architectures than ResNet, with DINO-ViT-B achieving $45.19\%$ adversarial average (42% drop from clean accuracy) compared to DINO-ResNet's $55\%$ drop. In contrast, MoCo struggles with transformer architectures, with MoCo-ViT showing poor performance at a $61\%$ drop, significantly worse than MoCoV3-ResNet at a $46\%$ drop. Even MoCoV3-ViT-B $51\%$ drop underperforms its ResNet counterpart despite having 4x more parameters. + +These differences are especially evident in UAP results, where DINO-ViT-B shows greater resilience (38% drop) than DINO-ResNet (44% drop), while MoCo-ViT performs poorly (26.31%, 64% drop) compared to MoCoV3-ResNet (36% drop). These findings suggest that the interaction between self-supervised approaches and model architectures + +significantly impacts adversarial robustness, with DINO benefiting from the ViT architecture while MoCo struggles with transformer models. The full results are in Section 4.6 of the Supp. Mat.. + +Lastly, we have evaluated DINOv2's [65] small and base and MAE's [40] base versions. We discuss them in the section 2 in Supp. Mat.. + +Impact of Training Duration. SSL models tend to demonstrate better performance as training epochs increase [11, 16, 20]. However, due to computational constraints, many models are reported with different numbers of epochs. This prompts the question of whether longer training durations enhance or reduce adversarial robustness. As ViT models do not have checkpoints at various epochs, we focus on ResNet-based SSL models, specifically SwAV and MoCoV3, which offer multiple checkpoints throughout the training process. We find that both SwAV and MoCo show very marginal improvement of about $1\%$ on IAA across various epochs, which is minimal compared to the rise in original accuracy. In contrast, both methods exhibit a modest increase in UAP performance after surpassing 100 epochs, with SwAV improving from $33\%$ to $40\%$ and MoCoV3 from $41\%$ to $48\%$ . Overall, our results suggest that despite differences in reported checkpoints, robustness generally remains largely stable during training for IAA, with slightly more noticeable but still modest gains for UAP. This indicates that training duration has a limited impact on adversarial robustness, even as clean accuracy continues to improve with extended training. The full results are in Section 4.5 of the Supp. Mat. + +# 4.2. Transfer Learning + +In this section, we analyze how adversarial robustness transfers from ImageNet pre-training to downstream tasks. We examine whether vulnerability patterns established during pre-training persist when models are evaluated through linear probing or fine-tuning across various ResNet-based self-supervised and supervised learning approaches. This analysis quantifies robustness transfer relationships through both Spearman rank and Pearson correlations. + +LINEAR. Our correlation analysis shows strong relationships between ImageNet and linear evaluation vulnerability. For performance drops, both Spearman and Pearson correlations are high across all attack types (Spearman—Adv: 0.93, UAP: 0.83, IAA: 0.79; Pearson—Adv: 0.94, UAP: 0.94, IAA: 0.88). These consistently high correlations suggest that robustness characteristics established during pre-training largely persist through linear probing. The performance drop analysis reveals method-specific robustness characteristics. MoCoV3 consistently exhibits smaller performance drops across both ImageNet (IAA: $54\%$ , UAP: $36\%$ ) and linear evaluation (IAA: $56\%$ , UAP: $43\%$ ), indicating its contrastive learning approach with mo + +Table 2. Component Contribution Analysis Across Datasets. Metrics include: Original Accuracy (Orig), Adversarial Accuracy (Adv) with relative performance drop percentage, Direction Ratio (D) and Magnitude Ratio (M). Ratio interpretation: Head dominant (<0.67), Balanced (0.67-1.5), Backbone dominant (>1.5). + +
ModelCIFAR-10CIFAR-100ImageNetCIFAR-10 (FT)CIFAR-100 (FT)
OrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DM
Barlow92.333.0 ↓64%2.73.177.920.5 ↓74%1.45.071.242.4 ↓40%0.99.497.166.9 ↓31%1.52.484.645.0 ↓57%0.82.4
BYOL93.031.0 ↓67%1.64.278.219.0 ↓76%1.076.074.639.4 ↓47%0.912.596.967.0 ↓31%1.21.483.961.2 ↓27%0.72.7
DINO93.927.6 ↓71%3.23.379.716.0 ↓80%1.66.275.324.7 ↓67%1.214.196.977.8 ↓20%1.92.984.752.9 ↓38%1.794.6
MoCoV394.733.0 ↓65%2.51.780.219.3 ↓76%1.14.574.642.7 ↓43%0.710.996.972.0 ↓26%1.11.384.562.6 ↓26%0.72.6
SimCLR91.037.9 ↓58%1.51.973.019.5 ↓73%1.33.568.924.3 ↓65%1.07.797.267.3 ↓31%1.01.984.443.0 ↓50%0.62.0
Supervised91.442.8 ↓53%1.80.773.924.5 ↓67%1.31.576.138.8 ↓49%1.43.096.262.3 ↓35%2.00.582.659.3 ↓28%1.31.1
SwAV93.919.4 ↓79%3.13.079.411.1 ↓86%1.86.475.324.7 ↓67%1.217.496.879.9 ↓17%1.72.784.454.6 ↓35%1.03.0
VICReg92.833.0 ↓64%2.73.177.822.3 ↓71%1.34.371.342.4 ↓40%0.99.397.168.6 ↓29%1.42.484.343.4 ↓48%0.82.4
+ +mentum encoders develops features with inherently better transferable robustness properties. In contrast, supervised pre-training and clustering-based approaches like SwAV experience more severe drops in both settings, suggesting these methods may create more brittle representations. + +The high correlation in adversarial vulnerability between ImageNet and linear evaluation indicates that linear probing provides a reliable assessment of a pre-trained model's downstream robustness characteristics without requiring extensive adaptation. This finding has practical implications for model selection, suggesting that robustness evaluations on ImageNet can effectively predict linear transfer performance. The full results are in Section 4.8 of the Supp. Mat.. + +FINETUNE. Fine-tuning reveals distinct transfer patterns by attack type. For IAA, the correlation between ImageNet and fine-tuned vulnerability nearly vanishes for performance drops (Spearman: 0.12, Pearson: 0.36), indicating that fine-tuning substantially reshapes defense against instance-specific attacks. Conversely, UAP vulnerability correlation remains high (Spearman: 0.88, Pearson: 0.93), suggesting that susceptibility to universal perturbations is more persistently encoded in the network regardless of parameter adaptation. This attack-dependent correlation disparity reveals that universal perturbation vulnerabilities, which exploit systematic weaknesses across the feature space, are deeply encoded in the network's architecture and learning approach, while instance-specific vulnerabilities are more malleable through fine-tuning. The absolute magnitude of performance drops diminishes considerably after fine-tuning across all methods (IAA: $49\%$ vs. $60\%$ in linear, UAP: $43\%$ vs. $49\%$ in linear), highlighting fine-tuning's effectiveness in mitigating vulnerability. MoCoV3 continues to demonstrate superior robustness with the smallest drops (IAA: $45\%$ , UAP: $38\%$ ), while BYOL shows dramatic improvement from ImageNet to fine-tuning for IAA attacks (from $56\%$ to $47\%$ ). These findings suggest different mechanisms for robustness transfer: UAP vulnerability appears tied to fundamental architectural and al + +gorithmic properties that persist across transfer paradigms, while instance-specific attack vulnerability depends more on the fine-tuning process than initial representation properties. This distinction may be valuable for practical applications, suggesting that pre-training method selection strongly impacts universal attack robustness, while defense against instance-specific attacks can be substantially improved through appropriate fine-tuning strategies. The full results are in Section 4.9 of the Supp. Mat.. + +Attributing Adversarial Vulnerability. Having established the correlations between ImageNet and transfer learning robustness patterns, we now seek to understand the underlying mechanisms causing these disparities across models and training paradigms. Specifically, we investigate whether adversarial vulnerability stems primarily from the backbone feature extractor or the classification head and how this attribution changes between linear probing and fine-tuning. + +Our analysis investigates the relationship between adversarial robustness and component-specific vulnerability under $\mathrm{FGSM}_1$ attack (detail in Supp. Mat.), focusing on two metrics: the Direction Ratio $D = \frac{\mathrm{mean}(1 - \cos(\mathbf{l}_{adv}))}{\mathrm{mean}(1 - \cos(\mathbf{f},\mathbf{f}_{adv}))}$ , which compares directional shifts in logits (l) versus backbone features (f), and the Magnitude Ratio $M = \frac{\mathrm{mean}(\|\mathbf{l} - \mathbf{l}_{adv}\|_2)}{\mathrm{mean}(\|\mathbf{f} - \mathbf{f}_{adv}\|_2)}$ , quantifying relative sensitivity to perturbation magnitudes. We fit linear regression models using these ratios to predict adversarial accuracy drop across models. + +For probed models, Direction Ratio strongly correlates with vulnerability $(R^2 = 0.56 - 0.85)$ , suggesting that directional instability in the head dominates FGSM robustness. SwAV, with $D = 3.1$ on CIFAR-10, suffers a $79\%$ accuracy drop, while SimCLR $(D = 1.5)$ drops $58\%$ . This may arise from the linear head's limited capacity to compensate for adversarial noise, amplifying directional shifts in logit space. + +After fine-tuning, this correlation weakens significantly $(R^2 = 0.05 - 0.52)$ : SwAV achieves only a $17\%$ drop despite retaining $D = 1.7$ , implying that fine-tuning enables the head to stabilize predictions even under directional feature + +shifts. This pattern aligns with our transfer learning analysis, where fine-tuning disrupted the correlation between ImageNet and downstream instance-specific attack robustness. Regression figures are in Section 2 of the Supp. Mat. + +As in Table 2, self-supervised models exhibit higher pre-fine-tuning Magnitude Ratios (e.g., SwAV $M = 17.4$ on ImageNet vs. supervised $M = 3.0$ ), potentially due to their reliance on globally normalized feature spaces, where FGSM perturbations propagate more aggressively to logits. Fine-tuning reduces $M$ across models (SwAV $M = 2.7$ post-FT), aligning with improved robustness, possibly by suppressing logit magnitude distortions. + +These patterns are specific to FGSM, as its single-step gradient reliance may emphasize head-layer instability, whereas iterative attacks could exploit backbone vulnerabilities differently. While our analysis tentatively links directional logit shifts $(D)$ and magnitude sensitivity $(M)$ to robustness, dataset-dependent variations—such as weaker correlations on CIFAR-100 FT—highlight the need for broader evaluation across threat models to generalize these insights. + +We also measured inter-class and intra-class distances for both probed and fine-tuned CIFAR-10 representations; however, these metrics did not yield meaningful correlations that explain the observed differences in adversarial robustness. The detailed regression figures, inter/intra-class distance measurements, and t-SNE [82] visualizations of feature spaces are provided in Section 3 of the Supp. Mat. + +# 4.3. Segmentation + +Unlike in classification, we observe no strong correlation between ImageNet robustness and segmentation performance across both PASCAL VOC and CityScapes datasets. Self-supervised approaches demonstrate competitive performance in clean conditions, with VICReg leading on PASCAL VOC (77.89 mIOU with frozen backbone) and SwAV on CityScapes (66.48 mIOU). All models suffer catastrophic performance degradation (74-84%) under adversarial attacks regardless of training methodology. This uniform degradation pattern suggests attacks primarily target segmentation modules rather than backbones. + +Our experiments with frozen versus unfrozen backbones reveal that frozen backbones generally achieve both higher clean performance and slightly better adversarial robustness on PASCAL VOC. The supervised model's exception to this pattern stems from using the standard MMSegmentation model for the unfrozen case due to reproduction challenges. These findings indicate that while SSL models produce competitive segmentation performance in clean conditions, they offer minimal advantage in terms of adversarial robustness for segmentation tasks, unlike their significant impact on object recognition. This highlights the need for robustness techniques specifically designed for segmenta + +tion architectures rather than focusing solely on backbone improvements. The full results are in Section 4.2 of the Supp. Mat. + +# 4.4. Detection + +Detection results show distinct patterns from recognition and segmentation tasks. On INRIA Person with unfrozen backbones, SwAV demonstrates the highest robustness (72% decrease), while the Supervised model shows the poorest performance (89% decrease). With frozen backbones, Barlow Twins and DINO lead in robustness (both 75% decrease), while VICReg becomes unexpectedly vulnerable (84% decrease), contradicting its recognition performance. + +COCO dataset results show improved robustness overall, with Barlow Twins maintaining the strongest performance (50% decrease) and the Supervised model remaining the least robust (67% decrease). This consistent weakness in the Supervised model suggests inherent robustness benefits from self-supervised pretraining. + +These findings indicate that task-specific architectures significantly influence adversarial robustness, and robustness in ImageNet classification doesn't necessarily transfer to detection tasks. This highlights the importance of task-specific evaluations and suggests that backbone architecture becomes less critical than overall model design for downstream applications. The full results are in Section 4.3 of the Supp. Mat. + +# 5. Conclusion + +Our evaluation shows that SSL models generally offer greater adversarial robustness than supervised counterparts in image classification, with MoCoV3 performing particularly well—likely due to its momentum encoder and prediction head. This robustness advantage is less evident in segmentation and detection, where task-specific architectures play a larger role. Architectural effects also vary by objective: DINO benefits from ViTs, while MoCoV3 performs better with CNNs. Extended training does not compromise robustness and may slightly improve it. We find that vulnerability to UAPs transfers more consistently across learning paradigms, while susceptibility to IAAs can be mitigated through fine-tuning. These results highlight the complex interplay between SSL objectives, architectures, and downstream tasks, underscoring the need for further study. We hope our findings support continued progress toward more robust visual representation learning. + +# Acknowledgements + +This work was funded by the KUIS AI Center at Koç University, Turkey. + +# References + +[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning, 2020. 2 +[2] Yutong Bai, Jieru Mei, Alan Yuille, and Cihang Xie. Are transformers more robust than cnns?, 2021. 6 +[3] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing, 2022. 1 +[4] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari S. Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A cookbook of self-supervised learning. ArXiv, abs/2304.12210, 2023. 1 +[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning, 2022. 2, 4 +[6] Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian Hadfield, Jeff Clune, Tegan Maharaj, Frank Hutter, Atilim Gunes Baydin, Sheila McIlraith, Qiqi Gao, Ashwin Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Brauner, and Soren Mindermann. Managing extreme ai risks amid rapid progress. Science, 384(6698): 842-845, 2024. 1 +[7] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga + +Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models, 2022. 1 +[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In Computer Vision – ECCV 2014, pages 446–461, Cham, 2014. Springer International Publishing. 4 +[9] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, Percy Liang, and John C. Duchi. Unlabeled data improves adversarial robustness, 2022. 3 +[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features, 2019. 2 +[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 4, 6 +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 4, 6 +[13] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey, 2018. 2 +[14] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020. 2 +[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation, 2018. 4 +[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. *ArXiv*, abs/2002.05709, 2020. 1, 2, 4, 6 +[17] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning, 2020. 3 +[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2020. 2, 5 +[19] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5 +[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 1, 2, 4, 6 +[21] Prakash Chandra Chhipa, Johan Rodahl Holmgren, Kanjar De, Rajkumar Saini, and Marcus Liwicki. Can self-supervised representation learning methods withstand distribution shifts and corruptions?, 2023. 3 +[22] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild, 2013. 4 + +[23] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. 4 +[24] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 886-893 vol. 1, 2005. 4 +[25] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 6 +[26] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2 +[27] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum, 2018. 4 +[28] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 1, 4, 6 +[29] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. How well do self-supervised models transfer?, 2021. 4 +[30] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning, 2021. 2 +[31] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html +[32] Alhussein Fawzi, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Robustness of classifiers: from adversarial to random noise, 2016. 1 +[33] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 4 +[34] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, Rama Chellappa, Andrew Gordon Wilson, and Tom Goldstein. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks, 2023. 3 + +[35] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 4 +[36] Jean-Bastien Grill, Florian Strub, Florent Altch'e, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. ArXiv, abs/2006.07733, 2020. 2, 4 +[37] Rohit Gupta, Naveed Akhtar, Ajmal Mian, and Mubarak Shah. Contrastive self-supervised learning leads to higher adversarial susceptibility, 2022. 2, 4 +[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 1, 4 +[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 5 +[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2021. 1, 2, 6 +[41] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty, 2019. 3 +[42] Chih-Hui Ho and Nuno Vasconcelos. Contrastive learning with adversarial examples, 2020. 2, 4 +[43] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1 +[44] Hao Huang, Ziyan Chen, Huanran Chen, Yongtao Wang, and Kevin Zhang. T-sea: Transfer-based self-ensemble attack on object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20514–20523, 2023. 4 +[45] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes imagenet good for transfer learning?, 2016. 4 +[46] Mark Ibrahim, David Klindt, and Randall Balestriero. Occam's razor for self supervised learning: What is sufficient to learn good representations?, 2024. 5 +[47] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features, 2019. 1 +[48] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning, 2020. 2, 4 +[49] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning, 2020. 2, 4 +[50] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. 1 +[51] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 4 + +[52] Antoni Kowalczuk, Jan Dubinski, Atiyeh Ashari Ghomi, Yi Sui, George Stein, Jiapeng Wu, Jesse C. Cresswell, Franziska Boenisch, and Adam Dziedzic. Benchmarking robust self-supervised learning across diverse downstream tasks, 2024. 1, 3 +[53] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 4 +[54] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 4 +[55] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[56] Ang Li, Yifei Wang, Yiwen Guo, and Yisen Wang. Adversarial examples are not real features, 2024. 1 +[57] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 4 +[58] Rundong Luo, Yifei Wang, and Yisen Wang. Rethinking the effect of data augmentation in adversarial contrastive learning, 2023. 3 +[59] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 4 +[60] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft, 2013. 4 +[61] Warren Morningstar, Alex Bijamov, Chris Duvarney, Luke Friedman, Neha Kalibhat, Luyang Liu, Philip Mansfield, Renan Rojas-Gomez, Karan Singhal, Bradley Green, and Sushant Prakash. Augmentations vs algorithms: What works in self-supervised learning, 2024. 5 +[62] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 262-271, 2020. 6 +[63] Yurii Nesterov. A method for solving the convex programming problem with convergence rate $o(1 / k^2)$ . Proceedings of the USSR Academy of Sciences, 269:543-547, 1983. 4 +[64] A. Tuan Nguyen, Ser Nam Lim, and Philip Torr. Task-agnostic robust representation learning, 2022. 2, 4 +[65] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 1, 2, 6 +[66] Utku Ozbulak, Hyun Jung Lee, Beril Boga, Esla Timothy Anzaku, Homin Park, Arnout Van Messem, Wesley De Neve, + +and Joris Vankerschaver. Know your self-supervised learning: A survey on image-based generative and discriminative training, 2023. 4 +[67] Serdar Ozsoy, Shadi Hamdan, Sercan Ö. Arik, Deniz Yuret, and Alper T. Erdogan. Self-supervised learning with an information maximization criterion, 2022. 2 +[68] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 4 +[69] Francesco Pinto, Philip H. S. Torr, and Puneet K. Dokania. An impartial take to the cnn vs transformer robustness contest, 2022. 6 +[70] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints, 2021. 4 +[71] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2 +[72] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 4 +[73] Jérôme Rony, Jean-Christophe Pesquet, and Ismail Ben Ayed. Proximal splitting adversarial attacks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4 +[74] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 2, 3, 4 +[75] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Mady. Adversarily robust generalization requires more data, 2018. 1 +[76] Ali Shafahi, W. Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable?, 2020. 1 +[77] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Neural Information Processing Systems, 2016. 5 +[78] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[79] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples, 2016. 1 +[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1 + +[81] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5 +[82] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8 +[83] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1 +[84] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training, 2022. +[85] Dongxian Wu, Shu tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization, 2020. 1 +[86] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection, 2017. 4 +[87] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan Yuille, and Quoc V. Le. Adversarial examples improve image recognition, 2020. 2, 4 +[88] Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, and Mohan Kankanhalli. Efficient adversarial contrastive learning via robustness-aware coreset selection, 2023. 3 +[89] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction, 2021. 2, 4 +[90] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D. Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness, 2022. 2, 4 +[91] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning?, 2022. 3 \ No newline at end of file diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/images.zip b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..58fbb6ee50476de8ce605bf6dc331d1b02e4b68c --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:520a2601dddf66a31da84a9c56e5b19b602ac723d30d4142594cbeca0825fa95 +size 232985 diff --git a/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/layout.json b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..041fe1db1f1278295a7a5004afee5a7f58d166c1 --- /dev/null +++ b/2025/Adversarial Robustness of Discriminative Self-Supervised Learning in Vision/layout.json @@ -0,0 +1,8170 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 68, + 103, + 541, + 121 + ], + "type": "text", + "content": "Adversarial Robustness of Discriminative Self-Supervised Learning in Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "spans": [ + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "content": "Ömer Veysel Çağatan" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "content": " 1Ömer Faruk Tal " + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "inline_equation", + "content": "{}^{1}\\mathrm{M}" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "content": ". Emre Gürsoy \n" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "content": "Department of Computer Engineering, Koç University \n" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 147, + 142, + 462, + 186 + ], + "type": "text", + "content": "KUIS AI Center, Koç University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 187, + 466, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 187, + 466, + 198 + ], + "spans": [ + { + "bbox": [ + 140, + 187, + 466, + 198 + ], + "type": "text", + "content": "ocagatan19@ku.edu.tr, otal19@ku.edu.tr, emregursoy@ku.edu.tr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 297, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 491 + ], + "type": "text", + "content": "Self-supervised learning (SSL) has advanced significantly in visual representation learning, yet comprehensive evaluations of its adversarial robustness remain limited. In this study, we evaluate the adversarial robustness of seven discriminative self-supervised models and one supervised model across diverse tasks, including ImageNet classification, transfer learning, segmentation, and detection. Our findings suggest that discriminative SSL models generally exhibit better robustness to adversarial attacks compared to their supervised counterpart on ImageNet, with this advantage extending to transfer learning when using linear evaluation. However, when fine-tuning is applied, the robustness gap between SSL and supervised models narrows considerably. Similarly, this robustness advantage diminishes in segmentation and detection tasks. We also investigate how various factors might influence adversarial robustness, including architectural choices, training duration, data augmentations, and batch sizes. Our analysis contributes to the ongoing exploration of adversarial robustness in visual self-supervised representation systems." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 135, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "content": "Self-supervised learning (SSL) [4], particularly discriminative approaches, has emerged as a foundational method for training models with remarkable capabilities in areas such as language [80], vision [65], and decision-making [50]. As these models become increasingly widespread and integrated into various applications, ensuring their reliability and safety has become a critical concern [6, 7]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "One particular challenge is the surprising vulnerability of deep learning models to adversarial examples, where slight input alterations can significantly impact model performance [35, 78]. This phenomenon has sparked significant debate, seeking to understand and mitigate these vulnerabilities [3, 32, 75, 76, 79, 83-85]. One prominent theory [47] suggests that adversarial examples arise from the model's sensitivity to non-robust features in the input data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 419 + ], + "type": "text", + "content": "According to this view, both robust (stable) and non-robust (vulnerable) features contribute to classification, with adversarial attacks manipulating the latter to cause misclassification. However, this theory, developed primarily in the context of supervised learning, faces challenges when extended to other self-supervised paradigms. [56] indicates that non-robust features are less effective in SSL methods such as contrastive learning [16], masked image modeling [40], or diffusion models [43]. This discrepancy suggests that non-robust features may lack the transferability across learning paradigms that robust or natural features possess. Thus, it becomes essential to systematically evaluate and compare how different SSL approaches respond to adversarial attacks, particularly given the theoretical evidence suggesting their feature representations may differ fundamentally from supervised models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 626 + ], + "type": "text", + "content": "These theoretical insights into how adversarial examples affect different learning paradigms highlight several critical gaps in our understanding of SSL's adversarial robustness. Notwithstanding the progress made in understanding the adversarial robustness of SSL, particularly contrastive learning, which we extensively discuss in Section 2, several key questions remain unresolved. First, with the wide variety of self-supervised representations available, employing different pretext tasks and data augmentations, which approaches demonstrate the greatest adversarial robustness? This remains unclear since most methods don't provide any results on adversarial robustness unless it is a specific focus of the proposed approach. Secondly, robustness is typically assessed by the model's accuracy on the pretraining dataset. Still, its adversarial impact on transfer learning or downstream tasks like detection and segmentation has not been thoroughly investigated [52]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "The choice of model architecture also raises questions about robustness. Standard vision SSL pretraining typically utilizes a ResNet [38] as the backbone, but more recently, larger and more powerful models [12, 20, 65] have been developed using vision transformers [28]. This leads to the question: Which architecture demonstrates greater robustness under the same SSL objective and with comparable" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2313" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 126, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 126, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 126, + 83 + ], + "type": "text", + "content": "parameter sizes?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 88, + 294, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 294, + 160 + ], + "type": "text", + "content": "Another factor to consider is the training duration. State-of-the-art SSL models are trained for longer durations compared to their supervised counterparts. Several studies indicate that this extended training consistently enhances performance, raising the question of whether this might compromise the models' adversarial robustness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 164, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 295, + 297 + ], + "type": "text", + "content": "While previous work has examined aspects of adversarial robustness in SSL, our study provides the first comprehensive cross-model comparison across multiple tasks, architectures, and training regimes. We assess seven different SSL models (Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]) alongside a supervised model against various adversarial attacks on ImageNet [74] and nine other image-recognition datasets. We also evaluate their robustness in segmentation and detection tasks. Our investigation addresses the following key questions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 307, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 295, + 342 + ], + "type": "text", + "content": "1. How does the adversarial robustness of various SSL models compare to that of supervised models on ImageNet?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 343, + 295, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 343, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 343, + 295, + 414 + ], + "type": "text", + "content": "SSL models consistently demonstrate greater adversarial robustness than supervised models on ImageNet. Noncontrastive methods show particular resilience against IAA [13] attacks, while all SSL approaches exhibit strong resistance to UAP [14], with MoCoV3 demonstrating the strongest overall performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 437 + ], + "type": "text", + "content": "2. Does SSL robustness transfer to downstream tasks like transfer learning, segmentation, and detection?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 438, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 438, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 67, + 438, + 295, + 510 + ], + "type": "text", + "content": "The robustness advantages transfer effectively to classification tasks via linear probing and fine-tuning. However, in segmentation and detection, all models exhibit similar vulnerability regardless of pretraining methodology, suggesting task-specific architectural components may override backbone robustness properties." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "3. How does model architecture influence adversarial robustness under the same SSL objective?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 534, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 534, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 295, + 629 + ], + "type": "text", + "content": "Architecture impact is highly objective dependent. MoCoV3 shows reduced robustness with Vision Transformers, whereas DINO demonstrates improved performance with ViT compared to ResNet, challenging the notion that architectural effects are uniform across SSL paradigms. Additionally, we have evaluated DINOv2 [65] and MAE [40] that do not demonstrate a consistently high or low level of robustness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 652 + ], + "type": "text", + "content": "4. Does extending training duration compromise adversarial robustness in SSL models?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 654, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 654, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 654, + 295, + 712 + ], + "type": "text", + "content": "Extended training either maintains or slightly enhances adversarial performance. For UAP attacks, performance improves meaningfully after more epochs in both SwAV and MoCoV3, indicating that longer training periods do not create a performance-robustness tradeoff." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 175 + ], + "type": "text", + "content": "Self Supervised Learning Self-supervised learning(SSL) seeks to extract meaningful and general representations from unlabeled data by leveraging pretext tasks. These tasks can vary, such as predicting the next word [71] or neighboring words [26] in a text, reconstructing masked sections of an image [40], or ensuring that two different perspectives of the same image result in similar visual representations [16]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 390 + ], + "type": "text", + "content": "Avoiding collapse is a key challenge in SSL for computer vision, and various methods can be classified based on how they address this issue. Contrastive approaches like SimCLR [16] and MoCo [19, 20, 39] use an objective that pushes apart representations of different inputs (negative samples) while bringing together those of the same input (positive samples). The performance and scalability of these methods heavily depend on the number and selection of negative samples. In another category, distillation methods such as BYOL [36], SimSiam [18], and DINO [12], prevent collapse by introducing asymmetry between different encoder branches and employing algorithmic adjustments. Additional SSL techniques, including DeepCluster [10], SeLa [1], and SwAV [11], enforce a clustering structure in the feature space to avoid constant representations. Meanwhile, methods like Barlow Twins [89], Whitening MSE (W-MSE) [30], VICReg [5], CorInfoMax [67] prevent collapse by using feature decorrelation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 391, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 701 + ], + "type": "text", + "content": "Adversarial Self-Supervised Learning While self-supervised learning (SSL) has outperformed supervised training [16], numerous studies highlight that contrastive learning remains susceptible to adversarial attacks when transferring the learned features to downstream classification tasks [42, 49]. To improve the robustness of contrastive learning, adversarial training has been adapted to self-supervised settings. In the absence of labels, adversarial examples are generated by maximizing the contrastive loss with respect to all input samples. Several prior works, such as ACL [48], RoCL [49], and CLAE [42], adopt this approach. Additionally, ACL incorporates the dual-BN technique [87] to further enhance performance. DeACL [90] introduces a two-stage approach, distilling a standard pretrained encoder through adversarial training. Nguyen et al. [64] establishes an upper bound on the adversarial loss of a prediction model, which is based on the learned representations, for any downstream task. This upper bound is determined using the model's loss on clean data and a robustness regularization term, which helps make the prediction model more resistant to adversarial attacks. [37] demonstrates that adversarial sensitivity stems from the uniform distribution of data representations on a unit hypersphere in the representation space. The presence of false negative pairs during training contributes to this effect, increasing the model's vulnerability to input perturbations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 702, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 702, + 553, + 713 + ], + "type": "text", + "content": "Although self-supervised adversarial training has made" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2314" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 72, + 503, + 186 + ], + "blocks": [ + { + "bbox": [ + 108, + 72, + 503, + 186 + ], + "lines": [ + { + "bbox": [ + 108, + 72, + 503, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 72, + 503, + 186 + ], + "type": "image", + "image_path": "686b75e84ddcb7e1a609561bbf037aa119ae94a72acec203dc338d2f666d0923.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "type": "text", + "content": "Figure 1. Performance scores for tasks such as ImageNet classification, transfer learning (with linear probing), segmentation, and detection (both with frozen backbones) are shown with the percentage drop in adversarial robustness. The shaded regions indicate the " + }, + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 195, + 555, + 240 + ], + "type": "text", + "content": " confidence interval around the regression line. Note the consistent pattern of higher robustness (lower percentage drop) among SSL models compared to supervised approaches in classification tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 252, + 480, + 366 + ], + "blocks": [ + { + "bbox": [ + 132, + 252, + 480, + 366 + ], + "lines": [ + { + "bbox": [ + 132, + 252, + 480, + 366 + ], + "spans": [ + { + "bbox": [ + 132, + 252, + 480, + 366 + ], + "type": "image", + "image_path": "a0d1b6b3b33dddc23a25c85fce4056a4fc5765fd7ed6c945de509913f41e7d67.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 376, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 555, + 409 + ], + "type": "text", + "content": "Figure 2. Averaged scores of SSL models on ImageNet across various attack types, including Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). Adv Avg refers to the average score across all attacks combined. The shaded regions indicate the " + }, + { + "bbox": [ + 54, + 376, + 555, + 409 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 376, + 555, + 409 + ], + "type": "text", + "content": " confidence interval around the regression line." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 431, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 295, + 514 + ], + "type": "text", + "content": "progress, it still does not match the performance of supervised methods. Luo et al. [58] suggests that this shortfall is due to data augmentation and proposes a dynamic data augmentation scheduler to achieve comparable results to supervised training. Xu et al. [88] efficiently applies ACL on the ImageNet [74] to obtain a robust representation using robustness-aware core set selection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 518, + 241, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 518, + 241, + 530 + ], + "spans": [ + { + "bbox": [ + 67, + 518, + 241, + 530 + ], + "type": "text", + "content": "Robustness of Self-Supervised Learning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "[41] found that incorporating an extra self-supervised task in a multi-task framework can enhance the adversarial robustness of supervised models. In a similar vein, Carmon et al. [9] discovered that using additional unlabeled data also strengthens the model's adversarial resilience. Furthermore, Chen et al. [17] created robust variants of pretext-based SSL tasks, showing that their integration with robust fine-tuning leads to a notable increase in robustness compared to standard adversarial training. Chhipa et al. [21] demonstrates a clear relationship between the performance of learned representations within SSL paradigms and the severity of distribution shifts and corruptions, and highlights the critical impact of distribution shifts and image corruptions on the performance and resilience of SSL methods. Similarly, Zhong et al. [91] conduct robustness tests to assess" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 598 + ], + "type": "text", + "content": "the behavioral differences between contrastive and supervised learning under changes in downstream or pre-training data distributions, while also exploring the effects of data augmentation and feature space characteristics. Kowalczuk et al. [52] conducts a comprehensive empirical evaluation of the adversarial robustness of self-supervised vision encoders across multiple downstream tasks, revealing the need for broader enhancements in encoder robustness. Goldblum et al. [34] benchmarks diverse pretrained models across multiple computer vision tasks, finding that supervised convolutional neural networks still outperform newer architectures on most metrics, while revealing self-supervised learning backbones show competitive potential when compared under equivalent conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Unlike prior studies that primarily focus on individual SSL methods, specific tasks, or limited adversarial scenarios, our work provides a comprehensive, unified benchmark across multiple SSL paradigms, architectures, and tasks—including classification, transfer learning, segmentation, and detection—under a diverse set of adversarial attacks, offering a broader and deeper understanding of adversarial robustness in SSL." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2315" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 173, + 86 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 173, + 86 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 173, + 86 + ], + "type": "text", + "content": "3. Experimental Setup" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 93, + 135, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 93, + 135, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 93, + 135, + 105 + ], + "type": "text", + "content": "3.1. SSL Models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 270 + ], + "type": "text", + "content": "While numerous SSL approaches have been proposed [66], we focus on a subset of widely used models due to computational constraints: Barlow Twins [89], BYOL [36], DINO [12], MoCoV3 [20], SimCLR [16], SwAV [11], and VICReg [5]. We primarily use ResNet50 [38] backbones, as most SSL checkpoints are released in this format, with DINO and MoCoV3 also offering ViT [28] variants. For BYOL, DINO, MoCoV3, SimCLR, and SwAV, complete model checkpoints were available. In contrast, Barlow Twins and VICReg only provided backbone weights, requiring linear evaluation via official code, which led to a slight drop in performance. For comparison, we also include a supervised ResNet50 baseline from PyTorch [68]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 234, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 234, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 234, + 300 + ], + "type": "text", + "content": "3.2. ImageNet and Transfer Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 305, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 521 + ], + "type": "text", + "content": "We use the benchmark suite introduced in the transfer learning study [45], which encompasses the target datasets like FGVC Aircraft [60], Caltech-101 [33], Stanford Cars [53], CIFAR 10 [54], CIFAR 100 [54], DTD [22], Oxford 102 Flowers [22], and Food-101 [8]. We follow Ericsson et al. [29] for both linear evaluation and fine-tuning of these datasets. We prioritized linear evaluation in our analysis as the backbone remains frozen during this process, allowing for a more equitable comparison of objectives within this setup. We apply the same adversarial techniques to ImageNet and transfer learning: Instance Adversarial Attacks (IAA) and Universal Adversarial Perturbations (UAP). In brief, instance-based methods generate unique perturbations for each image, while UAP creates a single perturbation that applies across the entire dataset. Comprehensive details and categorizations of all attack methods—such as white-box, black-box, gradient-free, among others—are provided in Section 1 of the Supp Mat.." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 537, + 142, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 142, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 142, + 550 + ], + "type": "text", + "content": "3.3. Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 713 + ], + "type": "text", + "content": "For segmentation, we use the Pascal VOC 2012 dataset [31] and CityScapes [23] dataset, training a DeepLabV3+ model [15]. To conduct the attacks, we follow the setup from Rony et al. [73], utilizing Alma [73], Asma [73], DAG [86], DDN [73], FGSM [35], FMN [70], and PGD [59]. While our primary metric is the mean Intersection Over Union (mIOU), we also report the Attack Pixel Success Rate (APSR) introduced by [73]. Although our main focus is on using a frozen backbone, we also perform training following the standard procedure. Our CityScapes results include only the frozen backbone approach, while our Pascal VOC results include both frozen and unfrozen backbone configurations." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 123, + 553, + 208 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 116 + ], + "type": "text", + "content": "Table 1. Performance of various models on ImageNet, Transfer Learning, Segmentation (Pascal VOC), and Detection (INRIA Person) tasks with frozen backbones, showing original (Orig.) and adversarial (Adv.) scores with performance drops in red." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 123, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 315, + 123, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 123, + 553, + 208 + ], + "type": "table", + "html": "
ModelImageNetTransfer LearningSegmentationDetection
Orig.Adv.Orig.Adv.Orig.Adv.Orig.Adv.
Barlow Twins71.236.5 ↓49%80.337.8 ↓54%76.920.5 ↓73%88.421.9 ↓75%
BYOL74.638.9 ↓48%78.736.6 ↓55%76.719.0 ↓75%87.417.3 ↓80%
DINO75.334.2 ↓55%80.735.6 ↓57%77.018.9 ↓76%87.622.0 ↓75%
MoCoV374.640.5 ↓46%80.540.3 ↓51%76.219.9 ↓74%87.318.5 ↓79%
SimCLR68.932.1 ↓53%73.133.0 ↓57%75.619.3 ↓74%87.713.1 ↓85%
Supervised76.134.5 ↓55%74.631.2 ↓60%74.216.5 ↓78%86.118.0 ↓79%
SwAV75.332.6 ↓57%79.232.4 ↓60%76.519.2 ↓75%86.620.5 ↓76%
VICReg71.337.0 ↓48%79.937.7 ↓54%77.920.5 ↓74%88.414.0 ↓84%
", + "image_path": "125d7db239fc4165ac1b0f9ac32e020375ad52b182f4791c22edda82b79db2c7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 382, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 382, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 382, + 239 + ], + "type": "text", + "content": "3.4. Detection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": "For object detection, we utilized the INRIA Person [24] and CoCo [57] datasets, and trained a Faster R-CNN [72]. To perform adversarial attacks, we followed the setup described by [44], employing the Transfer-based Self-Ensemble Attack (T-SEA). The T-SEA attack can be deployed using various methods and optimizers. In our experiments, we employed BIM [44], MIM [27], PGD [59], and Optim [44] methods. Additionally, we explored simpler methods that rely on common optimizers, such as Adam [51], SGD, and Nesterov [63]. Throughout our evaluation, we report the mean average precision (mAP) scores as the primary performance metric. While our primary focus was on employing a frozen backbone, we also conducted training experiments following the standard training procedures for comparative analysis. Our COCO results include only the frozen backbone approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 446, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 446, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 446, + 463 + ], + "type": "text", + "content": "4. Results and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 555 + ], + "type": "text", + "content": "In this section, we present our experimental findings on ImageNet, transfer learning, and detection, and discuss each in turn. Figure 1 and table 1 summarize the performance of various SSL models compared to supervised learning across our main evaluation tasks in the frozen backbone setup. While we address the results individually, the full detailed results are provided in Section 4 of the Supp. Mat.." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 383, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 383, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 383, + 577 + ], + "type": "text", + "content": "4.1. ImageNet" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "SSL vs Supervised. Most robustness studies on contrastive learning [42, 48, 49, 64, 87, 90] focus on small datasets like CIFAR10 [54] and primarily evaluate robustness using adversarial attacks such as FGSM [35] and PGD [59]. While computational constraints explain the reluctance to scale to larger datasets like ImageNet [74], many evaluations inadequately incorporate Universal Adversarial Perturbations (UAP). Our findings, as shown in Figure 2 contradict previous research by Gupta et al. [37]. Under IAA, MoCoV3 demonstrates the strongest robustness (54% drop), while SimCLR shows the weakest performance (61% drop). For" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2316" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "content": "UAP attacks, MoCoV3 again leads (38% drop), while the supervised model and SwAV both show 50% drops. Notably, our results challenge the conclusion about contrastive vs. non-contrastive methods. MoCoV3, a contrastive model, consistently demonstrates the highest adversarial robustness with ResNet architecture, while DINO, which was claimed to perform better due to its non-contrastive nature, shows the weakest IAA robustness in our ResNet evaluation (64% drop). Our data reveals that Barlow, BYOL, MoCoV3, and VICReg all demonstrate comparable resilience against IAA (around 54% drop), contradicting the simple categorization of robustness based on contrastive versus non-contrastive approaches. Our comprehensive evaluation using diverse attack methods demonstrates that the relationship between self-supervised learning approaches and adversarial robustness is more complex than previously suggested. The full results are in Section 4.1 of the Supp. Mat.." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 294, + 337 + ], + "type": "text", + "content": "What makes MoCoV3 robust? Although MoCoV3 and SimCLR both utilize the InfoNCE [77, 81] objective, there is a stark contrast in their adversarial robustness. To understand this disparity, we evaluate MoCoV1 [39], MoCoV2 [19], and MoCoV3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 294, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 294, + 423 + ], + "type": "text", + "content": "A brief MoCo History. MoCoV1 introduced a dynamic dictionary with a queue and momentum-updated encoder to maintain consistent negative samples. MoCoV2 enhanced this with a multi-layer projection head and stronger data augmentation. MoCoV3 further evolved by eliminating the memory bank and incorporating a prediction head similar to BYOL and SimSiam [18]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": "Our analysis reveals a clear progression in robustness across MoCo versions. While MoCoV1 demonstrates limited resilience with a " + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "inline_equation", + "content": "71\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": " drop in overall adversarial accuracy, MoCoV2 shows significant improvement (52% drop), and MoCoV3 achieves the strongest performance (46% drop). The most dramatic enhancement occurs in UAP resistance, where MoCoV1's performance drops by " + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "inline_equation", + "content": "77\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": ", compared to MoCoV2's " + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": " and MoCoV3's " + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": ". The substantial improvement from MoCoV1 to MoCoV2 (7% in clean accuracy, 27 percentage points in UAP robustness) primarily stems from the non-linear projector, with data augmentation providing marginal benefits. While [46] suggests non-linear projectors aren't always essential, our results indicate they significantly enhance both performance and adversarial robustness. MoCoV3's superior performance over MoCoV2 (an additional " + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 713 + ], + "type": "text", + "content": " in clean accuracy and 7 percentage points in UAP robustness) can be attributed to its prediction head and larger batch size. Unlike the transition from V1 to V2, MoCoV3 shows substantial improvements in both IAA (34% vs 25%) and UAP (48% vs 41%), highlighting the prediction head's critical role in enhancing overall robustness. Our findings suggest that momentum, a common feature in robust models like MoCoV3 and BYOL, significantly contributes to adversar" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ial resilience, while MoCoV2's performance more closely resembles that of SimCLR, which lacks this feature. The full results are in Section 4.6 of the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 110, + 457, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 110, + 457, + 121 + ], + "spans": [ + { + "bbox": [ + 326, + 110, + 457, + 121 + ], + "type": "text", + "content": "Augmentations vs Algorithms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 122, + 553, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 553, + 218 + ], + "type": "text", + "content": "Morningstar et al. [61] challenges the notion that SSL progress is primarily driven by algorithmic advancements, suggesting instead that augmentation diversity, along with data and model scale, play more critical roles. Their analysis argues that many algorithmic improvements, such as prediction networks or new loss functions, had minimal impact on downstream task performance compared to stronger augmentation techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 219, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 553, + 374 + ], + "type": "text", + "content": "Our comprehensive analysis of the MoCo family evolution provides a more nuanced perspective on this debate. The substantial progression in adversarial robustness from MoCoV1 to MoCoV2 and further to MoCoV3 suggests that architectural innovations like non-linear projectors and prediction heads significantly impact robustness. MoCoV2's dramatic improvement over MoCoV1, particularly in UAP performance, indicates that the multi-layer projection head provides substantial benefits beyond mere augmentation changes. Similarly, MoCoV3's further enhancements in both IAA and UAP performance relative to MoCoV2 highlight the crucial role of the prediction head in overall robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 508 + ], + "type": "text", + "content": "While our comparison lacks perfectly controlled baselines for augmentations across different objectives (with slight variations in augmentation between MoCoV2 and V3), this limitation stems from the unavailability of public checkpoints rather than our experimental design. Despite this constraint, the consistent improvements in adversarial performance strongly suggest that algorithmic innovations significantly contribute to adversarial robustness. Importantly, our findings demonstrate that higher clean accuracy doesn't automatically translate to improved robustness on ImageNet." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 509, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 616 + ], + "type": "text", + "content": "Although the checkpoints from [61] are not publicly available, we conducted an ablation study on augmentation types and batch sizes using BYOL which is the only model in our SSL pool that includes configurations with varied augmentations and batch sizes. For this analysis, we consider four distinct models: BYOL-NC (without color distortions), BYOL-CC (with only color and crop augmentations), and the standard BYOL models with batch sizes of 128 (BYOL-128) and 512 (BYOL-512)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "We observe that batch size and augmentation choices have varying effects on the adversarial robustness of BYOL variants. For IAA robustness, both BYOL-NC and BYOLCC show similar performance (60% drop), while BYOL128, BYOL-512, and standard BYOL demonstrate slightly better resilience (57%, 56%, and 56% drops, respectively). These modest differences suggest that batch size and augmentation have a limited impact on instance-level attack ro" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2317" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 96, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 96, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 96, + 83 + ], + "type": "text", + "content": "bustness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 86, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 86, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 86, + 294, + 216 + ], + "type": "text", + "content": "For UAP attacks, we see more substantial variations: BYOL-128 (41% drop) and BYOL-512 (43% drop) perform comparably, while BYOL-NC and BYOL-CC show notably weaker performance (46% and 48% drops). Standard BYOL demonstrates a 39% drop, which is marginally better than the batch size variants but substantially better than the limited-augmentation models. The 7-9 percentage point difference between standard BYOL and the limited-augmentation variants suggests that comprehensive augmentation strategies may contribute to improved UAP robustness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 219, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 397 + ], + "type": "text", + "content": "Interestingly, we note that BYOL-CC performs slightly worse than BYOL-NC despite having more augmentations, though this difference is too small to draw meaningful conclusions. Similarly, the differences between BYOL-128, BYOL-512, and standard BYOL in UAP robustness are relatively minor and should be interpreted cautiously. Our results indicate that the relationship between augmentation and adversarial robustness is complex and that meaningful improvements likely require more than just incremental changes to batch size or augmentation strategies. While our data hints at potential benefits from comprehensive augmentation for UAP robustness, the effects are modest and require further investigation with more controlled experiments. The full results are in Section 4.7 of the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 519 + ], + "type": "text", + "content": "ResNet vs ViT in Adversarial Robustness. While ViTs are generally seen as more robust than CNNs [62], Bai et al. [2], Pinto et al. [69] demonstrate that with the right training methods, CNNs [55] can achieve comparable robustness. Despite ViT's success [12, 20, 25, 28, 65], most SSL methods still use ResNet for validation. We examine MoCoV3 and DINO, as they are the only models that include ViT training, with our analysis covering both standard ViTs comparable to ResNet50 and larger ViT-B variants with approximately 4x more parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "Our results reveal notable architectural differences. DINO performs significantly better with ViT architectures than ResNet, with DINO-ViT-B achieving " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "45.19\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " adversarial average (42% drop from clean accuracy) compared to DINO-ResNet's " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " drop. In contrast, MoCo struggles with transformer architectures, with MoCo-ViT showing poor performance at a " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "61\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " drop, significantly worse than MoCoV3-ResNet at a " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " drop. Even MoCoV3-ViT-B " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " drop underperforms its ResNet counterpart despite having 4x more parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "These differences are especially evident in UAP results, where DINO-ViT-B shows greater resilience (38% drop) than DINO-ResNet (44% drop), while MoCo-ViT performs poorly (26.31%, 64% drop) compared to MoCoV3-ResNet (36% drop). These findings suggest that the interaction between self-supervised approaches and model architectures" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "significantly impacts adversarial robustness, with DINO benefiting from the ViT architecture while MoCo struggles with transformer models. The full results are in Section 4.6 of the Supp. Mat.." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "type": "text", + "content": "Lastly, we have evaluated DINOv2's [65] small and base and MAE's [40] base versions. We discuss them in the section 2 in Supp. Mat.." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": "Impact of Training Duration. SSL models tend to demonstrate better performance as training epochs increase [11, 16, 20]. However, due to computational constraints, many models are reported with different numbers of epochs. This prompts the question of whether longer training durations enhance or reduce adversarial robustness. As ViT models do not have checkpoints at various epochs, we focus on ResNet-based SSL models, specifically SwAV and MoCoV3, which offer multiple checkpoints throughout the training process. We find that both SwAV and MoCo show very marginal improvement of about " + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": " on IAA across various epochs, which is minimal compared to the rise in original accuracy. In contrast, both methods exhibit a modest increase in UAP performance after surpassing 100 epochs, with SwAV improving from " + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "inline_equation", + "content": "33\\%" + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": " and MoCoV3 from " + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "inline_equation", + "content": "41\\%" + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "inline_equation", + "content": "48\\%" + }, + { + "bbox": [ + 313, + 158, + 553, + 433 + ], + "type": "text", + "content": ". Overall, our results suggest that despite differences in reported checkpoints, robustness generally remains largely stable during training for IAA, with slightly more noticeable but still modest gains for UAP. This indicates that training duration has a limited impact on adversarial robustness, even as clean accuracy continues to improve with extended training. The full results are in Section 4.5 of the Supp. Mat." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 443, + 424, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 424, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 424, + 455 + ], + "type": "text", + "content": "4.2. Transfer Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 553, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 556 + ], + "type": "text", + "content": "In this section, we analyze how adversarial robustness transfers from ImageNet pre-training to downstream tasks. We examine whether vulnerability patterns established during pre-training persist when models are evaluated through linear probing or fine-tuning across various ResNet-based self-supervised and supervised learning approaches. This analysis quantifies robustness transfer relationships through both Spearman rank and Pearson correlations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "LINEAR. Our correlation analysis shows strong relationships between ImageNet and linear evaluation vulnerability. For performance drops, both Spearman and Pearson correlations are high across all attack types (Spearman—Adv: 0.93, UAP: 0.83, IAA: 0.79; Pearson—Adv: 0.94, UAP: 0.94, IAA: 0.88). These consistently high correlations suggest that robustness characteristics established during pre-training largely persist through linear probing. The performance drop analysis reveals method-specific robustness characteristics. MoCoV3 consistently exhibits smaller performance drops across both ImageNet (IAA: " + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "inline_equation", + "content": "54\\%" + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": ", UAP: " + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": ") and linear evaluation (IAA: " + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": ", UAP: " + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "), indicating its contrastive learning approach with mo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "2318" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 113, + 538, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Component Contribution Analysis Across Datasets. Metrics include: Original Accuracy (Orig), Adversarial Accuracy (Adv) with relative performance drop percentage, Direction Ratio (D) and Magnitude Ratio (M). Ratio interpretation: Head dominant (<0.67), Balanced (0.67-1.5), Backbone dominant (>1.5)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 113, + 538, + 222 + ], + "lines": [ + { + "bbox": [ + 73, + 113, + 538, + 222 + ], + "spans": [ + { + "bbox": [ + 73, + 113, + 538, + 222 + ], + "type": "table", + "html": "
ModelCIFAR-10CIFAR-100ImageNetCIFAR-10 (FT)CIFAR-100 (FT)
OrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DMOrigAdv ↓DM
Barlow92.333.0 ↓64%2.73.177.920.5 ↓74%1.45.071.242.4 ↓40%0.99.497.166.9 ↓31%1.52.484.645.0 ↓57%0.82.4
BYOL93.031.0 ↓67%1.64.278.219.0 ↓76%1.076.074.639.4 ↓47%0.912.596.967.0 ↓31%1.21.483.961.2 ↓27%0.72.7
DINO93.927.6 ↓71%3.23.379.716.0 ↓80%1.66.275.324.7 ↓67%1.214.196.977.8 ↓20%1.92.984.752.9 ↓38%1.794.6
MoCoV394.733.0 ↓65%2.51.780.219.3 ↓76%1.14.574.642.7 ↓43%0.710.996.972.0 ↓26%1.11.384.562.6 ↓26%0.72.6
SimCLR91.037.9 ↓58%1.51.973.019.5 ↓73%1.33.568.924.3 ↓65%1.07.797.267.3 ↓31%1.01.984.443.0 ↓50%0.62.0
Supervised91.442.8 ↓53%1.80.773.924.5 ↓67%1.31.576.138.8 ↓49%1.43.096.262.3 ↓35%2.00.582.659.3 ↓28%1.31.1
SwAV93.919.4 ↓79%3.13.079.411.1 ↓86%1.86.475.324.7 ↓67%1.217.496.879.9 ↓17%1.72.784.454.6 ↓35%1.03.0
VICReg92.833.0 ↓64%2.73.177.822.3 ↓71%1.34.371.342.4 ↓40%0.99.397.168.6 ↓29%1.42.484.343.4 ↓48%0.82.4
", + "image_path": "fcca2a0eb816a57576b4afec2d274588315351dd955c17f0f405aa8f83b892d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 240, + 294, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 240, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 294, + 302 + ], + "type": "text", + "content": "mentum encoders develops features with inherently better transferable robustness properties. In contrast, supervised pre-training and clustering-based approaches like SwAV experience more severe drops in both settings, suggesting these methods may create more brittle representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 303, + 295, + 411 + ], + "type": "text", + "content": "The high correlation in adversarial vulnerability between ImageNet and linear evaluation indicates that linear probing provides a reliable assessment of a pre-trained model's downstream robustness characteristics without requiring extensive adaptation. This finding has practical implications for model selection, suggesting that robustness evaluations on ImageNet can effectively predict linear transfer performance. The full results are in Section 4.8 of the Supp. Mat.." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": "FINETUNE. Fine-tuning reveals distinct transfer patterns by attack type. For IAA, the correlation between ImageNet and fine-tuned vulnerability nearly vanishes for performance drops (Spearman: 0.12, Pearson: 0.36), indicating that fine-tuning substantially reshapes defense against instance-specific attacks. Conversely, UAP vulnerability correlation remains high (Spearman: 0.88, Pearson: 0.93), suggesting that susceptibility to universal perturbations is more persistently encoded in the network regardless of parameter adaptation. This attack-dependent correlation disparity reveals that universal perturbation vulnerabilities, which exploit systematic weaknesses across the feature space, are deeply encoded in the network's architecture and learning approach, while instance-specific vulnerabilities are more malleable through fine-tuning. The absolute magnitude of performance drops diminishes considerably after fine-tuning across all methods (IAA: " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": " in linear, UAP: " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": " in linear), highlighting fine-tuning's effectiveness in mitigating vulnerability. MoCoV3 continues to demonstrate superior robustness with the smallest drops (IAA: " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": ", UAP: " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": "), while BYOL shows dramatic improvement from ImageNet to fine-tuning for IAA attacks (from " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "56\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "47\\%" + }, + { + "bbox": [ + 56, + 415, + 295, + 713 + ], + "type": "text", + "content": "). These findings suggest different mechanisms for robustness transfer: UAP vulnerability appears tied to fundamental architectural and al" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 240, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 348 + ], + "type": "text", + "content": "gorithmic properties that persist across transfer paradigms, while instance-specific attack vulnerability depends more on the fine-tuning process than initial representation properties. This distinction may be valuable for practical applications, suggesting that pre-training method selection strongly impacts universal attack robustness, while defense against instance-specific attacks can be substantially improved through appropriate fine-tuning strategies. The full results are in Section 4.9 of the Supp. Mat.." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 348, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 456 + ], + "type": "text", + "content": "Attributing Adversarial Vulnerability. Having established the correlations between ImageNet and transfer learning robustness patterns, we now seek to understand the underlying mechanisms causing these disparities across models and training paradigms. Specifically, we investigate whether adversarial vulnerability stems primarily from the backbone feature extractor or the classification head and how this attribution changes between linear probing and fine-tuning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "text", + "content": "Our analysis investigates the relationship between adversarial robustness and component-specific vulnerability under " + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{FGSM}_1" + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "text", + "content": " attack (detail in Supp. Mat.), focusing on two metrics: the Direction Ratio " + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "inline_equation", + "content": "D = \\frac{\\mathrm{mean}(1 - \\cos(\\mathbf{l}_{adv}))}{\\mathrm{mean}(1 - \\cos(\\mathbf{f},\\mathbf{f}_{adv}))}" + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "text", + "content": ", which compares directional shifts in logits (l) versus backbone features (f), and the Magnitude Ratio " + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "inline_equation", + "content": "M = \\frac{\\mathrm{mean}(\\|\\mathbf{l} - \\mathbf{l}_{adv}\\|_2)}{\\mathrm{mean}(\\|\\mathbf{f} - \\mathbf{f}_{adv}\\|_2)}" + }, + { + "bbox": [ + 313, + 457, + 553, + 568 + ], + "type": "text", + "content": ", quantifying relative sensitivity to perturbation magnitudes. We fit linear regression models using these ratios to predict adversarial accuracy drop across models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": "For probed models, Direction Ratio strongly correlates with vulnerability " + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "inline_equation", + "content": "(R^2 = 0.56 - 0.85)" + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": ", suggesting that directional instability in the head dominates FGSM robustness. SwAV, with " + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "inline_equation", + "content": "D = 3.1" + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": " on CIFAR-10, suffers a " + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "inline_equation", + "content": "79\\%" + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": " accuracy drop, while SimCLR " + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "inline_equation", + "content": "(D = 1.5)" + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": " drops " + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "inline_equation", + "content": "58\\%" + }, + { + "bbox": [ + 313, + 569, + 553, + 665 + ], + "type": "text", + "content": ". This may arise from the linear head's limited capacity to compensate for adversarial noise, amplifying directional shifts in logit space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "After fine-tuning, this correlation weakens significantly " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "(R^2 = 0.05 - 0.52)" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ": SwAV achieves only a " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " drop despite retaining " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "D = 1.7" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", implying that fine-tuning enables the head to stabilize predictions even under directional feature" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2319" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "shifts. This pattern aligns with our transfer learning analysis, where fine-tuning disrupted the correlation between ImageNet and downstream instance-specific attack robustness. Regression figures are in Section 2 of the Supp. Mat." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": "As in Table 2, self-supervised models exhibit higher pre-fine-tuning Magnitude Ratios (e.g., SwAV " + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "inline_equation", + "content": "M = 17.4" + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": " on ImageNet vs. supervised " + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "inline_equation", + "content": "M = 3.0" + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": "), potentially due to their reliance on globally normalized feature spaces, where FGSM perturbations propagate more aggressively to logits. Fine-tuning reduces " + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": " across models (SwAV " + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "inline_equation", + "content": "M = 2.7" + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": " post-FT), aligning with improved robustness, possibly by suppressing logit magnitude distortions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "text", + "content": "These patterns are specific to FGSM, as its single-step gradient reliance may emphasize head-layer instability, whereas iterative attacks could exploit backbone vulnerabilities differently. While our analysis tentatively links directional logit shifts " + }, + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "inline_equation", + "content": "(D)" + }, + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "text", + "content": " and magnitude sensitivity " + }, + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 55, + 217, + 294, + 324 + ], + "type": "text", + "content": " to robustness, dataset-dependent variations—such as weaker correlations on CIFAR-100 FT—highlight the need for broader evaluation across threat models to generalize these insights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 326, + 294, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 294, + 410 + ], + "type": "text", + "content": "We also measured inter-class and intra-class distances for both probed and fine-tuned CIFAR-10 representations; however, these metrics did not yield meaningful correlations that explain the observed differences in adversarial robustness. The detailed regression figures, inter/intra-class distance measurements, and t-SNE [82] visualizations of feature spaces are provided in Section 3 of the Supp. Mat." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 142, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 142, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 142, + 431 + ], + "type": "text", + "content": "4.3. Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 294, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 294, + 568 + ], + "type": "text", + "content": "Unlike in classification, we observe no strong correlation between ImageNet robustness and segmentation performance across both PASCAL VOC and CityScapes datasets. Self-supervised approaches demonstrate competitive performance in clean conditions, with VICReg leading on PASCAL VOC (77.89 mIOU with frozen backbone) and SwAV on CityScapes (66.48 mIOU). All models suffer catastrophic performance degradation (74-84%) under adversarial attacks regardless of training methodology. This uniform degradation pattern suggests attacks primarily target segmentation modules rather than backbones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "content": "Our experiments with frozen versus unfrozen backbones reveal that frozen backbones generally achieve both higher clean performance and slightly better adversarial robustness on PASCAL VOC. The supervised model's exception to this pattern stems from using the standard MMSegmentation model for the unfrozen case due to reproduction challenges. These findings indicate that while SSL models produce competitive segmentation performance in clean conditions, they offer minimal advantage in terms of adversarial robustness for segmentation tasks, unlike their significant impact on object recognition. This highlights the need for robustness techniques specifically designed for segmenta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "tion architectures rather than focusing solely on backbone improvements. The full results are in Section 4.2 of the Supp. Mat." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 119, + 382, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 119, + 382, + 130 + ], + "spans": [ + { + "bbox": [ + 314, + 119, + 382, + 130 + ], + "type": "text", + "content": "4.4. Detection" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 137, + 553, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 553, + 244 + ], + "type": "text", + "content": "Detection results show distinct patterns from recognition and segmentation tasks. On INRIA Person with unfrozen backbones, SwAV demonstrates the highest robustness (72% decrease), while the Supervised model shows the poorest performance (89% decrease). With frozen backbones, Barlow Twins and DINO lead in robustness (both 75% decrease), while VICReg becomes unexpectedly vulnerable (84% decrease), contradicting its recognition performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 319 + ], + "type": "text", + "content": "COCO dataset results show improved robustness overall, with Barlow Twins maintaining the strongest performance (50% decrease) and the Supervised model remaining the least robust (67% decrease). This consistent weakness in the Supervised model suggests inherent robustness benefits from self-supervised pretraining." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 415 + ], + "type": "text", + "content": "These findings indicate that task-specific architectures significantly influence adversarial robustness, and robustness in ImageNet classification doesn't necessarily transfer to detection tasks. This highlights the importance of task-specific evaluations and suggests that backbone architecture becomes less critical than overall model design for downstream applications. The full results are in Section 4.3 of the Supp. Mat." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 430, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 654 + ], + "type": "text", + "content": "Our evaluation shows that SSL models generally offer greater adversarial robustness than supervised counterparts in image classification, with MoCoV3 performing particularly well—likely due to its momentum encoder and prediction head. This robustness advantage is less evident in segmentation and detection, where task-specific architectures play a larger role. Architectural effects also vary by objective: DINO benefits from ViTs, while MoCoV3 performs better with CNNs. Extended training does not compromise robustness and may slightly improve it. We find that vulnerability to UAPs transfers more consistently across learning paradigms, while susceptibility to IAAs can be mitigated through fine-tuning. These results highlight the complex interplay between SSL objectives, architectures, and downstream tasks, underscoring the need for further study. We hope our findings support continued progress toward more robust visual representation learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 669, + 417, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 417, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 417, + 681 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "This work was funded by the KUIS AI Center at Koç University, Turkey." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2320" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 146 + ], + "type": "text", + "content": "[2] Yutong Bai, Jieru Mei, Alan Yuille, and Cihang Xie. Are transformers more robust than cnns?, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 294, + 181 + ], + "type": "text", + "content": "[3] Yang Bai, Yuyuan Zeng, Yong Jiang, Shu-Tao Xia, Xingjun Ma, and Yisen Wang. Improving adversarial robustness via channel-wise activation suppressing, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 183, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 183, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 183, + 296, + 258 + ], + "type": "text", + "content": "[4] Randall Balestriero, Mark Ibrahim, Vlad Sobal, Ari S. Morcos, Shashank Shekhar, Tom Goldstein, Florian Bordes, Adrien Bardes, Gregoire Mialon, Yuandong Tian, Avi Schwarzschild, Andrew Gordon Wilson, Jonas Geiping, Quentin Garrido, Pierre Fernandez, Amir Bar, Hamed Pirsiavash, Yann LeCun, and Micah Goldblum. A cookbook of self-supervised learning. ArXiv, abs/2304.12210, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "type": "text", + "content": "[5] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 392 + ], + "type": "text", + "content": "[6] Yoshua Bengio, Geoffrey Hinton, Andrew Yao, Dawn Song, Pieter Abbeel, Trevor Darrell, Yuval Noah Harari, Ya-Qin Zhang, Lan Xue, Shai Shalev-Shwartz, Gillian Hadfield, Jeff Clune, Tegan Maharaj, Frank Hutter, Atilim Gunes Baydin, Sheila McIlraith, Qiqi Gao, Ashwin Acharya, David Krueger, Anca Dragan, Philip Torr, Stuart Russell, Daniel Kahneman, Jan Brauner, and Soren Mindermann. Managing extreme ai risks amid rapid progress. Science, 384(6698): 842-845, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 713 + ], + "type": "text", + "content": "[7] Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson, Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel, Jared Quincy Davis, Dora Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Stefano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani, Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar, Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Carlos Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou, Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich, Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Re, Dorsa Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan Taori, Armin W. Thomas, Florian Tramér, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu, Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Jiaxuan You, Matei Zaharia, Michael Zhang, Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the opportunities and risks of foundation models, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "text", + "content": "[8] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101 – mining discriminative components with random forests. In Computer Vision – ECCV 2014, pages 446–461, Cham, 2014. Springer International Publishing. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 553, + 194 + ], + "type": "text", + "content": "[9] Yair Carmon, Aditi Raghunathan, Ludwig Schmidt, Percy Liang, and John C. Duchi. Unlabeled data improves adversarial robustness, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 226 + ], + "type": "text", + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 2, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herv'e J'egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9630-9640, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[13] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. Adversarial attacks and defences: A survey, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 361, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 402 + ], + "type": "text", + "content": "[14] Ashutosh Chaubey, Nikhil Agrawal, Kavya Barnwal, Keerat K Guliani, and Pramod Mehta. Universal adversarial perturbations: A survey. arXiv preprint arXiv:2005.08087, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[15] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[16] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. *ArXiv*, abs/2002.05709, 2020. 1, 2, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[17] Tianlong Chen, Sijia Liu, Shiyu Chang, Yu Cheng, Lisa Amini, and Zhangyang Wang. Adversarial robustness: From self-supervised pre-training to fine-tuning, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "text", + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15745-15753, 2020. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "text", + "content": "[19] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 635 + ], + "type": "text", + "content": "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. arXiv preprint arXiv:2104.02057, 2021. 1, 2, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[21] Prakash Chandra Chhipa, Johan Rodahl Holmgren, Kanjar De, Rajkumar Saini, and Marcus Liwicki. Can self-supervised representation learning methods withstand distribution shifts and corruptions?, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[22] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild, 2013. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "2321" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[23] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding, 2016. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[24] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), pages 886-893 vol. 1, 2005. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 314 + ], + "type": "text", + "content": "[25] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey Gritsenko, Vighnesh Birodkar, Cristina Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lučić, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, and Neil Houlsby. Scaling vision transformers to 22 billion parameters, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 348 + ], + "type": "text", + "content": "[26] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "text", + "content": "[27] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum, 2018. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "type": "text", + "content": "[28] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 1, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 449, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 469 + ], + "type": "text", + "content": "[29] Linus Ericsson, Henry Gouk, and Timothy M. Hospedales. How well do self-supervised models transfer?, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 502 + ], + "type": "text", + "content": "[30] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[31] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "text", + "content": "[32] Alhussein Fawzi, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Robustness of classifiers: from adversarial to random noise, 2016. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 646 + ], + "type": "text", + "content": "[33] Li Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In 2004 Conference on Computer Vision and Pattern Recognition Workshop, pages 178–178, 2004. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[34] Micah Goldblum, Hossein Souri, Renkun Ni, Manli Shu, Viraj Prabhu, Gowthami Somepalli, Prithvijit Chattopadhyay, Mark Ibrahim, Adrien Bardes, Judy Hoffman, Rama Chellappa, Andrew Gordon Wilson, and Tom Goldstein. Battle of the backbones: A large-scale comparison of pretrained models across computer vision tasks, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[35] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 182 + ], + "type": "text", + "content": "[36] Jean-Bastien Grill, Florian Strub, Florent Altch'e, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. ArXiv, abs/2006.07733, 2020. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "text", + "content": "[37] Rohit Gupta, Naveed Akhtar, Ajmal Mian, and Mubarak Shah. Contrastive self-supervised learning leads to higher adversarial susceptibility, 2022. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 240 + ], + "type": "text", + "content": "[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015. 1, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 283 + ], + "type": "text", + "content": "[39] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. arXiv preprint arXiv:1911.05722, 2019. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 340 + ], + "type": "text", + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll'ar, and Ross B. Girshick. Masked autoencoders are scalable vision learners. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979-15988, 2021. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[41] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 397 + ], + "type": "text", + "content": "[42] Chih-Hui Ho and Nuno Vasconcelos. Contrastive learning with adversarial examples, 2020. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "text", + "content": "[43] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[44] Hao Huang, Ziyan Chen, Huanran Chen, Yongtao Wang, and Kevin Zhang. T-sea: Transfer-based self-ensemble attack on object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20514–20523, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 477, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 499 + ], + "type": "text", + "content": "[45] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes imagenet good for transfer learning?, 2016. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 532 + ], + "type": "text", + "content": "[46] Mark Ibrahim, David Klindt, and Randall Balestriero. Occam's razor for self supervised learning: What is sufficient to learn good representations?, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "text", + "content": "[47] Andrew Ilyas, Shibani Santurkar, Dimitris Tsipras, Logan Engstrom, Brandon Tran, and Aleksander Madry. Adversarial examples are not bugs, they are features, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 601 + ], + "type": "text", + "content": "[48] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning, 2020. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "text", + "content": "[49] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning, 2020. 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[50] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[51] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2322" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[52] Antoni Kowalczuk, Jan Dubinski, Atiyeh Ashari Ghomi, Yi Sui, George Stein, Jiapeng Wu, Jesse C. Cresswell, Franziska Boenisch, and Adam Dziedzic. Benchmarking robust self-supervised learning across diverse downstream tasks, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[53] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, 2013. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 198 + ], + "type": "text", + "content": "[54] Alex Krizhevsky. Learning multiple layers of features from tiny images. 2009. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "text", + "content": "[55] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 256 + ], + "type": "text", + "content": "[56] Ang Li, Yifei Wang, Yiwen Guo, and Yisen Wang. Adversarial examples are not real features, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 301 + ], + "type": "text", + "content": "[57] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 335 + ], + "type": "text", + "content": "[58] Rundong Luo, Yifei Wang, and Yisen Wang. Rethinking the effect of data augmentation in adversarial contrastive learning, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[59] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 416 + ], + "type": "text", + "content": "[60] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft, 2013. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 474 + ], + "type": "text", + "content": "[61] Warren Morningstar, Alex Bijamov, Chris Duvarney, Luke Friedman, Neha Kalibhat, Luyang Liu, Philip Mansfield, Renan Rojas-Gomez, Karan Singhal, Bradley Green, and Sushant Prakash. Augmentations vs algorithms: What works in self-supervised learning, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 530 + ], + "type": "text", + "content": "[62] Muzammal Naseer, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Fatih Porikli. A self-supervised approach for adversarial robustness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 262-271, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 532, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 564 + ], + "type": "text", + "content": "[63] Yurii Nesterov. A method for solving the convex programming problem with convergence rate " + }, + { + "bbox": [ + 56, + 532, + 294, + 564 + ], + "type": "inline_equation", + "content": "o(1 / k^2)" + }, + { + "bbox": [ + 56, + 532, + 294, + 564 + ], + "type": "text", + "content": ". Proceedings of the USSR Academy of Sciences, 269:543-547, 1983. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "type": "text", + "content": "[64] A. Tuan Nguyen, Ser Nam Lim, and Philip Torr. Task-agnostic robust representation learning, 2022. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 689 + ], + "type": "text", + "content": "[65] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 1, 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[66] Utku Ozbulak, Hyun Jung Lee, Beril Boga, Esla Timothy Anzaku, Homin Park, Arnout Van Messem, Wesley De Neve," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "and Joris Vankerschaver. Know your self-supervised learning: A survey on image-based generative and discriminative training, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "text", + "content": "[67] Serdar Ozsoy, Shadi Hamdan, Sercan Ö. Arik, Deniz Yuret, and Alper T. Erdogan. Self-supervised learning with an information maximization criterion, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 143, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 228 + ], + "type": "text", + "content": "[68] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "text", + "content": "[69] Francesco Pinto, Philip H. S. Torr, and Puneet K. Dokania. An impartial take to the cnn vs transformer robustness contest, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "type": "text", + "content": "[70] Maura Pintor, Fabio Roli, Wieland Brendel, and Battista Biggio. Fast minimum-norm adversarial attacks through adaptive norm constraints, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 301, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 553, + 323 + ], + "type": "text", + "content": "[71] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "text", + "content": "[72] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks, 2016. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[73] Jérôme Rony, Jean-Christophe Pesquet, and Ismail Ben Ayed. Proximal splitting adversarial attacks for semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "type": "text", + "content": "[74] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge, 2015. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "text", + "content": "[75] Ludwig Schmidt, Shibani Santurkar, Dimitris Tsipras, Kunal Talwar, and Aleksander Mady. Adversarily robust generalization requires more data, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "text", + "content": "[76] Ali Shafahi, W. Ronny Huang, Christoph Studer, Soheil Feizi, and Tom Goldstein. Are adversarial examples inevitable?, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 564 + ], + "type": "text", + "content": "[77] Kihyuk Sohn. Improved deep metric learning with multiclass n-pair loss objective. In Neural Information Processing Systems, 2016. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[78] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 644 + ], + "type": "text", + "content": "[79] Thomas Tanay and Lewis Griffin. A boundary tilting perspective on the phenomenon of adversarial examples, 2016. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[80] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2323" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 445 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[81] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[82] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9 (86):2579-2605, 2008. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[83] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[84] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 231 + ], + "type": "text", + "content": "[85] Dongxian Wu, Shu tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 264 + ], + "type": "text", + "content": "[86] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 298 + ], + "type": "text", + "content": "[87] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan Yuille, and Quoc V. Le. Adversarial examples improve image recognition, 2020. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "type": "text", + "content": "[88] Xilie Xu, Jingfeng Zhang, Feng Liu, Masashi Sugiyama, and Mohan Kankanhalli. Efficient adversarial contrastive learning via robustness-aware coreset selection, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "text", + "content": "[89] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction, 2021. 2, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 409 + ], + "type": "text", + "content": "[90] Chaoning Zhang, Kang Zhang, Chenshuang Zhang, Xi Niu, Jiu Feng, Chang D. Yoo, and In So Kweon. Decoupled adversarial contrastive learning for self-supervised adversarial robustness, 2022. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 412, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 294, + 445 + ], + "type": "text", + "content": "[91] Yuanyi Zhong, Haoran Tang, Junkun Chen, Jian Peng, and Yu-Xiong Wang. Is self-supervised learning more robust than supervised learning?, 2022. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2324" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_content_list.json b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4a6a64e44847ba8c1c7b3c32d0c4447da8e2c2c3 --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "Adversarial Training for Probabilistic Robustness", + "text_level": 1, + "bbox": [ + 246, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Zhang $^{1}$ , Yuhang Chen $^{1}$ , Zhen Chen $^{2}$ , Wenjie Ruan $^{2}$ , Xiaowei Huang $^{2}$ , Siddartha Khastgir $^{1}$ , Xingyu Zhao $^{1*}$ $^{1}$ WMG, University of Warwick $^{2}$ University of Liverpool {yi.zhang.16,yuhang.chen.1,s.khastgir.1,xingyu.zhao}@warwick.ac.uk; {zhen.chen2,xiaowei.huang}@liverpool.ac.uk; w.ruan@trustai.uk", + "bbox": [ + 181, + 179, + 815, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning (DL) has shown transformative potential across industries, yet its sensitivity to adversarial examples (AEs) limits its reliability and broader deployment. Research on DL robustness has developed various techniques, with adversarial training (AT) established as a leading approach to counter AEs. Traditional AT focuses on worst-case robustness (WCR), but recent work has introduced probabilistic robustness (PR), which evaluates the likelihood of AEs within a local perturbation range, providing an overall assessment of the model's robustness and acknowledging residual risks that are more practical to manage. However, existing AT methods are fundamentally designed to improve WCR, and no dedicated methods currently target PR. To bridge this gap, we formulate a new min-max optimization as the theoretical foundation for PR-focused AT, and introduce an AT-PR training scheme with numerical algorithms to solve the new optimization problem. Our experiments, based on 70 DL models trained on common datasets and diverse architectures, demonstrate that: i) AT-PR achieves higher improvements in PR than AT-WCR methods; ii) it shows more consistent effectiveness across varying local inputs; iii) it exhibits a reduced trade-off in model's generalization.", + "bbox": [ + 89, + 335, + 485, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 714, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep Learning (DL) has shown potential for transformative impacts across many industries including transportation, healthcare and energy. The reality that DL is not as dependable as required now becomes a major impediment to their broader deployment. Since the first discovery of the phenomenon of Adversarial Examples (AEs) for DL [21, 39, 48], robustness is arguably the property in the limelight, encompassing research topics like adversarial attacks in malicious contexts [7, 8, 14, 21, 36, 64, 65, 75],", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AE detection during testing [19, 23, 24, 35, 42], verification with formal guarantees [20, 26, 30, 37, 44], and various robustness improvement techniques, including distillation [40], randomized smoothing [12], gradient masking [2], and the more widely used adversarial training (AT) [4, 21, 22, 36, 52]. As large language models and multimodal foundation models gain popularity, more recent research has increasingly focused on understanding their robustness as well [11, 18, 28, 32, 33, 58, 71, 72, 74].", + "bbox": [ + 511, + 305, + 905, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Regardless of specific DL tasks or model architectures, robustness broadly refers to a model's ability to maintain consistent decisions under small input perturbations, with local robustness focusing on stability around individual inputs, as opposed to global robustness across the entire input distribution. More specifically, when evaluating local robustness, common problem formulations include [16, 72]: i) a binary question, \"does an AE exist within a specified perturbation norm-ball?\" [20, 30, 44, 63]; ii) a maximum safe distance question, \"what is the largest perturbation radius within which no AE exists?\" [1, 38, 60, 61]; and iii) a maximum prediction loss question, \"which AE within the norm-ball yields the highest prediction loss?\", which is often applied in AT to enhance robustness [36, 56]. It is evident that the three formulations focus on extreme scenarios, where local robustness is evaluated based on the existence of a deterministic AE, reflecting a worst-case perspective.", + "bbox": [ + 511, + 443, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A more recent and distinct view, however, considers probabilistic robustness (PR), employing statistical approaches to its evaluation [5, 13, 41, 49, 50, 55, 59, 60, 69, 72], it answers the question of \"what is the likelihood of AEs in the given perturbation norm-ball?\" [73]. However, existing AT methods are primarily designed for WCR via the well-established \"min-max\" optimization [3, 21, 36] (cf. Sec. 2.2), and are thus referred to as \"AT-WCR\" in this paper. No dedicated AT methods exist for enhancing PR, which motivates our work.", + "bbox": [ + 511, + 702, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we first formulate AT for PR as a new min-max optimization problem, and then devise AT-PR training scheme based on numerical algorithms to effectively", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding to X. Zhao, xingyu.zhao@warwick.ac.uk.", + "bbox": [ + 107, + 887, + 462, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1675", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "solve this training optimization. Our experiments on 70 DL models trained on common datasets and models demonstrate key strengths of AT-PR over AT-WCR: $i$ while traditional AT-WCR can also improve PR, our new AT-PR training scheme is more effective, especially when enhancing PR becomes challenging for traditional AT-WCR; $ii$ AT-PR not only outperforms AT-WCR in improving PR on average, but also offers more stable and consistent performance across different local inputs; $iii$ AT-PR has less negative impact on model's generalization compared to AT-WCR.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, key contributions of this paper include:", + "bbox": [ + 109, + 243, + 455, + 258 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Problem formulation: We lay the foundation of AT for PR by formulating a new min-max optimization problem.", + "- Numerical algorithms: We develop effective numerical algorithms to solve the AT-PR min-max optimization.", + "- Empirical evaluation: We validate the effectiveness of AT-PR through extensive experiments, showing its superior improvement in PR, more consistent effectiveness across different local inputs, and reduced impact on generalization compared to AT-WCR.", + "- Open-source repository: A repository is provided at https://github.com/wellzline/AT-PR with code, datasets, models, and all experiment details." + ], + "bbox": [ + 89, + 258, + 483, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries and Related Works", + "text_level": 1, + "bbox": [ + 89, + 455, + 393, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Worst-Case vs Probabilistic Robustness", + "text_level": 1, + "bbox": [ + 89, + 482, + 429, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although definitions of robustness vary across different DL tasks and model types, it generally refers to a DL model's ability to maintain consistent predictions despite small input perturbations. Typically it is defined as all inputs in a region $\\eta$ have the same prediction, where $\\eta$ is a small norm ball (in a $L_{p}$ -norm distance) of radius $\\gamma$ around an input $\\mathbf{x}$ . A perturbed input (e.g., by adding noise on $\\mathbf{x}$ ) $\\mathbf{x}'$ within $\\eta$ is an AE if its prediction label differs from $\\mathbf{x}$ .", + "bbox": [ + 89, + 505, + 483, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation methods for robustness are typically based on metrics defined around AEs [27, 75]. As discussed earlier in Sec. 1, four types of robustness metrics, as summarized in [16, 72], address: (1) a binary question, (2) a maximum safe distance question, (3) a maximum prediction loss question (Fig. 1 (a)), and (4) a statistical inference question estimating the likelihood of AEs within a given perturbation norm-ball (Fig. 1 (b)). The first three metrics focus on extreme scenarios, where local robustness is assessed based on the existence of a deterministic AE, reflecting a worst-case perspective referred to as WCR. In contrast, the fourth approach adopts a probabilistic view, evaluating overall local robustness in the presence of AEs.", + "bbox": [ + 89, + 628, + 483, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PR is arguably more practical than WCR, as extreme-case focus is often unnecessary [73]. Instead, knowing the likelihood of AEs is more relevant for applications, where it is sufficient to show that the risk of AE remains below an acceptable threshold rather than exactly zero [59, 69, 72].", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by [72], we illustrate the difference between the maximum loss WCR and PR in Fig. 1, omitting the binary and maximum safe distance cases as they are less relevant to the AT focus of this paper.", + "bbox": [ + 511, + 90, + 906, + 152 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/697ef97724b8801d9b0c18d3955ce50176773663612f33571c447d95f526fad7.jpg", + "image_caption": [ + "Class A" + ], + "image_footnote": [], + "bbox": [ + 516, + 162, + 709, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7ea8f056e35a19db73b4dea3e51e56a5cf261303c81d35f1aacd72659f419106.jpg", + "image_caption": [ + "Figure 1. Worst-case (a) vs. Probabilistic (b) Robustness." + ], + "image_footnote": [], + "bbox": [ + 540, + 290, + 568, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/705aca3d16f6fe087b1984d786ea29dbb9960ec3ffa2043cd949721350f58b5c.jpg", + "image_caption": [ + "#" + ], + "image_footnote": [], + "bbox": [ + 622, + 290, + 661, + 299 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c10ef37e22f6aba6edf80d440dd76afbc145614a6e3a7a51de6935727acb24c8.jpg", + "image_caption": [ + "Non-AE" + ], + "image_footnote": [], + "bbox": [ + 720, + 162, + 901, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6d162242fbfe60e28f0b56daa25a7761ca8c42b90a4b5728025c344954358ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 290, + 785, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To evaluate the WCR, we normally formulate it as a question of maximizing the prediction loss, cf. Fig. 1 (a), where adversarial attacks are typically introduced to find such a worst-case AE with perturbation:", + "bbox": [ + 511, + 345, + 906, + 406 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} = \\arg \\max _ {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 417, + 906, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\delta$ is a small perturbation constrained within an $L_{p}$ -norm ball of radius $\\gamma$ . This optimization causes the model to misclassify $\\boldsymbol{x}^{\\prime} = \\boldsymbol{x} + \\delta$ as a different label $y^\\prime \\neq y$ . There are some popular attacks following this formulation, including FGSM [21], PGD [36], and ZOO [10], among others. Following common practice, we measure WCR by the success rate of some selected attacks on a test dataset. Estimation of PR is normally formalized as the statistical inference problem on the following metric:", + "bbox": [ + 511, + 452, + 906, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Probabilistic Robustness) For a DL classifier $f_{\\theta}$ that takes input $x$ and returns a prediction label, the PR of an input $x$ in a norm ball of radius $\\gamma$ is:", + "bbox": [ + 511, + 598, + 905, + 645 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nP R (\\boldsymbol {x}, \\gamma) = \\underset {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} {\\mathbb {E}} _ {\\delta \\sim P r (\\cdot | \\boldsymbol {x})} [ I _ {\\{f _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) = y \\}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 655, + 906, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $I_{S}(\\pmb{x})$ is an indicator function—it is equal to 1 when $S$ is true and 0 otherwise; $Pr(\\cdot)$ is the local distribution of inputs representing how perturbations $\\delta$ are generated, which is precisely the \"input model\" used by [59, 72].", + "bbox": [ + 511, + 694, + 905, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Def. 1 suggests PR is the probability that the model prediction remains unchanged from a random perturbation $x'$ . A \"frequentist\" interpretation of this probability is: it is the limiting relative frequency of perturbations where the output label is preserved, in an infinite sequence of independently generated perturbations [72]. That is, the \"proportion\" of non-AEs in such infinite set of perturbations.", + "bbox": [ + 511, + 763, + 905, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The work [59] is the first to define PR and introduces a black-box statistical estimator, particularly useful when", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1676", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PR is extremely high (i.e., detecting AEs is a rare event). A more efficient white-box statistical estimator for PR was later proposed in [50]. Huang et al. [25] first extend PR to explainable AI (XAI). Zhang et al. [69] explore PR under functional perturbations, while the recent study [72] is the first to estimate PR in the text-to-image DL task. The recent work [73] provides an overview of PR, including its definition, estimation, enhancement and system-level integration. Without loss of generality, we adopt the estimator in [59] to estimate PR in later experiments.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Training", + "text_level": 1, + "bbox": [ + 89, + 250, + 290, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "AT [4, 21, 22, 36, 52] is a widely used empirical approach for improving robustness, typically formulated as a \"min-max\" optimization problem. It introduces a set of input perturbations $\\delta$ within a norm-ball $\\eta$ of radius $\\gamma$ , designed to maximize the training loss $\\mathcal{L}$ . The model is then trained with these perturbations by minimizing the overall loss of the training dataset. These perturbations are often approximated using empirical adversarial attacks, such as FGSM [21] and PGD [36] due to their high efficiency. Taking a classification task as an example, the min-max optimization target of AT can be formalised as:", + "bbox": [ + 89, + 272, + 483, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (AT for WCR) For a DL model $f_{\\theta}$ , trained on a dataset $\\mathcal{D}$ , consisting of pairs $(\\pmb{x}, y)$ where $\\pmb{x} \\in \\mathcal{X}$ is the input and $y \\in \\mathcal{Y}$ denotes the label, with loss function $\\mathcal{L}(\\pmb{x}, y; \\theta)$ , the objective of AT-WCR can be formulated as:", + "bbox": [ + 89, + 446, + 483, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 517, + 482, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta$ represents the perturbation that maximizes classification loss $\\mathcal{L}$ , while the outer minimization aims to train $f_{\\theta}$ to be robust against such adversarial perturbations.", + "bbox": [ + 88, + 559, + 482, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The inner maximization is precisely the adversarial attack formulation outlined in Eq. (1). As highlighted in [3], the effectiveness of AT largely depends on identifying strong AEs within this inner maximization. Notably, the PGD attack [36] has become a critical benchmark and is widely regarded as the standard method for AT in practice [3]. Accordingly, we also use PGD in later experiments. Based on the standard min-max optimization, various AT variants have been proposed to achieve more stable and efficient training, such as [6, 15, 29, 47, 52, 57, 58, 66, 67]. Comprehensive overviews of recent AT variants can be found in surveys [3, 27], all of which fundamentally rely on the min-max problem for WCR.", + "bbox": [ + 89, + 613, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, to the best of our knowledge, only three works [43, 55, 70] have explicitly aimed at improving PR [73]. Although Wang et al. [55] demonstrate that simply adding randomly perturbed inputs significantly enhances PR, the resulting AR is nearly zero. The methods of [43] and [70] propose loss functions based on common risk", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "measures (i.e., $\\rho$ -essential supremum, Conditional Value-at-Risk, and Entropic Value-at-Risk) to train models to minimize these risks. However, again, these approaches provide little to no improvement in AR. We argue that, akin to the well-established AT framework for AR, a dedicated AT framework for PR is necessary—one that does not compromise AR performance. Accordingly, our proposed AT-PR is the first to explicitly reformulate the min-max optimization principle for PR, while also providing comparable AR performance through practical training algorithms.", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Approach: AT-PR", + "text_level": 1, + "bbox": [ + 511, + 253, + 728, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 277, + 718, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, the goal of AT-WCR in Def. 2 is to identify the AE that maximizes model loss and incorporate it into AT to lower the maximum loss, cf. the top-middle illustration in Fig. 2. However, to effectively improve local PR, we aim to locate the \"widest peak\"—the region with the highest proportion of AEs in the loss landscape (cf. the top-right illustration in Fig. 2), rather than the tallest peak. Adding the AE from this peak during AT may ideally reduce the loss for AEs around that area, thereby lowering the overall AE proportion (cf. the intuitive interpretation of PR in Def. 1). In some cases, AT-WCR and AT-PR may identify the same AE for training when the tallest peak also happens to be the widest, as shown in the second row of Fig. 2. Notably, the two cases illustrated in Fig. 2 representatively cover all possible scenarios, though additional peaks may exist in reality. Given the difference between the goals of improving WCR and PR, we formulate a new min-max problem for AT-PR:", + "bbox": [ + 511, + 299, + 906, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (AT for PR) For a DL model $f_{\\theta}$ , trained on a dataset $\\mathcal{D}$ consisting of pairs $(x, y)$ , the objective of AT-PR can be formulated as:", + "bbox": [ + 511, + 563, + 906, + 609 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma , P R (\\boldsymbol {x} + \\delta , k) = 0} k \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 614, + 905, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $k$ and $\\delta$ are the variables in the new inner max optimization, $PR(\\cdot, \\cdot)$ is the metric defined in Eq. (2). As shown in Fig. 3, the intuition behind the maximization is to find an optimal $\\delta^*$ that produces an $AE$ $x' = x + \\delta^*$ and to maximize the radius $k$ of a smaller norm-ball centered on $x'$ such that the $PR$ within this $k$ -sized norm-ball is 0 (i.e., all inputs within this smaller norm-ball are AEs).", + "bbox": [ + 511, + 652, + 905, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The high non-convexity of DL models makes it intractable to analytically solve both the established min-max optimization in Eq. (3) and our new formulation in Eq. (4). Thus, similar to existing AT-WCR, we propose algorithms to numerically solve Eq. (4) in AT-PR in the next section.", + "bbox": [ + 511, + 765, + 905, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Optimization Algorithm", + "text_level": 1, + "bbox": [ + 511, + 848, + 736, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the outer minimization in Eq. (4) can be effectively handled with standard DL training methods as in AT-WCR,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1677", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/45e4646f4632831e825b287d49e36f3ff221b061eaaeee41001d36066dfec770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 356, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0bc69b4220e5a18c1848fbc1a91af8d0edd253f3ae476d3f8ff863131525636.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 89, + 630, + 214 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f2638fb90f667e4c111681f00ca826b384bdb6362f1cdc3095aa6e0bb95d0d13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 88, + 903, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e9125c9065c840c5551d6c0e4c3ad46f6fa230015a81219c64d1b96cbb5bff4d.jpg", + "image_caption": [ + "Figure 2. Two cases of the local loss landscape and their AEs identified by AT-WCR and AT-PR for AT." + ], + "image_footnote": [], + "bbox": [ + 93, + 215, + 357, + 339 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7f4eac3de2261e7295b0c44f242116eed48d38466123235dd0f9994e2e5a2d98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 215, + 653, + 339 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8239a33d05b6f04482cc45ce6cf6271c4340114e5088b746541ecf33e6e4a05c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 229, + 898, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b1e3431703387c08f43c69b982e145e3aa4ad1c3362acf2805616a8f009fa88.jpg", + "image_caption": [ + "Figure 3. Illustration of the new inner-max optimization in Def. 3." + ], + "image_footnote": [], + "bbox": [ + 91, + 393, + 475, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we devise the Algo. 1 to solve the inner optimization in our AT-PR. The gist of Algo. 1 is shown in Fig. 4, which essentially consists of 3 main steps:", + "bbox": [ + 89, + 609, + 482, + 654 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Step 1 (lines 1-9): Apply PGD attacks in different regions of the local norm-ball under study to find a diverse set of AEs at various local optima $x_{i}^{\\prime}$ .", + "- Step 2 (lines 10–20): For each locally optimal AE found in Step 1, perform a gradient descent search to locate the nearest decision boundary.", + "- Step 3 (lines 21-25): Calculate the distance $d_{i}$ from each $x_{i}^{\\prime}$ to its respective decision boundary and return the AE with the largest $d_{i}$ value." + ], + "bbox": [ + 89, + 656, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to AT-WCR, the returned AE will be added into the training dataset in AT-PR. We note, as an numerical algorithm to the optimization, its effectiveness may depends on certain assumptions include: $i$ ) the PGD attack in Step 1 may effectively find local optima by adjusting hyperparameters such as attack steps and learning rate, which aligns with the insight of \"PGD will converge to local maxima\"", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[36]; ii) The gradient descent based boundary search with adaptive hyperparameters may efficiently find the decision boundary as used in some boundary attacks [9, 31, 68]. We also provide an initial theoretical analysis of the effectiveness and convergence properties of the proposed optimization algorithm in Appendix A.", + "bbox": [ + 511, + 392, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 1: Apply PGD in different regions of the norm-ball to get a set of diverse AEs at different local optimal points $x_{i}^{\\prime}$ .", + "bbox": [ + 517, + 496, + 759, + 541 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86c16682dcb38119d2433013afa9f2335d278cd739a57b3ec64c40fc4d9d3671.jpg", + "image_caption": [ + "Figure 4. Illustration of the 3 main steps of Algo. 1." + ], + "image_footnote": [], + "bbox": [ + 516, + 545, + 903, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We analyze the time complexity of the proposed algorithm. There are several computational steps to be considered with the assumption that the dimension of the input $\\pmb{x}$ is $n$ . First, sampling $\\alpha$ and steps from uniform distributions (lines 5-6) is independent of the input and thus the time complexity is $O(1)$ . Second, the PGD algorithm (line 7) consists of steps gradient calculations with complexity $O(\\text{grad})^1$ and one projection calculation with complexity", + "bbox": [ + 511, + 757, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The time complexity of gradient calculation depends on the model", + "bbox": [ + 531, + 886, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1678", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 PGD and gradient-based search for $x_{pr}^{\\prime}$ \nRequire: Inputs $[X,Y],N,\\alpha_{min},\\alpha_{max},step_{min},step_{max}$ 1: Initialize $AE_{s}\\gets []$ Initialize AE candidate set 2: Apply PGD to get different AE candidates 3: for idx = 0 to N do 4: $x_{init}\\gets x + \\mathrm{Uniform}(-\\gamma ,\\gamma)$ 5: $\\alpha \\leftarrow$ Uniform $(\\alpha_{min},\\alpha_{max})$ 6: steps $\\leftarrow$ Uniform $(step_{min},step_{max})$ 7: $x_{idx}^{\\prime}\\gets pgd(x_{init},y,\\gamma ,\\alpha ,steps)$ 8: Append $x_{idx}^{\\prime}$ to $AE_{s}$ 9: end for \nRequire: $AE_s,x,y,C$ \n10: Initialize Max_d $\\leftarrow 0$ . $x_{pr}^{\\prime}\\gets$ None 11: for each $x^{\\prime}\\in AE_{s}$ do 12: $\\tilde{x} = x^{\\prime}$ 13: while iter $< C$ do 14: $y^\\prime = f(\\tilde{x})$ 15: if $y^\\prime = y$ then 16: break $\\triangleright$ Exit if $\\tilde{x}$ is classified correctly 17: end if 18: $g\\gets \\nabla_{\\tilde{x}}L(\\tilde{x},y)$ $\\triangleright$ Compute gradient 19: $\\tilde{x} = \\tilde{x} -\\alpha \\cdot g$ $\\triangleright$ Update $\\tilde{x}$ \n20: end while \n21: $d\\gets D(\\tilde{x},x^{\\prime})$ $\\triangleright$ Distance between $\\tilde{x}\\& x^{\\prime}$ \n22: if $d > Max\\_ d$ then \n23: $x_{pr}^{\\prime} = x^{\\prime}$ \n24: Max_d=d \n25: end if \n26: end for \n27: Output $x_{pr}^{\\prime}$ Farthest AE from decision boundary.", + "bbox": [ + 91, + 89, + 485, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$O(n)$ . Third, in the effort to locate the decision boundary (lines 18-19), one gradient calculation step is performed. Last, a $L_{\\infty}$ distance between $\\tilde{\\pmb{x}}$ and $\\pmb{x}^{\\prime}$ is estimated (line 21), with complexity $O(n)$ . In Step 1, the algorithm uses different random starts via random initialization [46] to ensure it finds different local optima within norm-ball $\\eta$ (line 4), and generates one AE (lines 5-8) for each combination of input $(x,y)$ , radius $\\gamma$ , randomly sampled $\\alpha$ and steps. A total of $N$ AEs are generated. Thus the complexity upper bound of Step 1 is then $O(N*(steps*O(grad)+O(n)))$ . In Step 2, with the maximum iteration number $C$ for the nearest decision boundary locating process of $N$ AEs, the complexity upper bound of Step 2 is then $O(C*N*O(grad))$ . In Step 3, the complexity of distance estimation is $O(n)$ as discussed above. Assuming $O(grad) \\gg O(n)$ , which", + "bbox": [ + 89, + 627, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "architecture and the specific loss function employed during training. E.g., DL models with more layers or complex loss functions may require more computational resources for gradient computation due to backpropagation.", + "bbox": [ + 89, + 862, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "typically holds in DL scenarios, the complexity simplifies to $O(N * (steps + C) \\cdot O(grad))$ . As analyzed, the computational cost of AT-PR primarily comes from locating the optimal AE. However, it remains compatible with other techniques designed to enhance AT efficiency, such as [46, 62]. We provide comparative empirical results by replacing PGD with Fast-AT [62] to demonstrate such compatibility in Table 3, refer to Sec. 4 for details.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 224, + 645, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train a total of 70 DL models on four datasets: CIFAR-10, CIFAR-100, SVHN, and Tiny/ImageNet, using models such as ResNet, Wide-ResNet (WRN), Vision Transformer (ViT) [17, 34], and DeiT [51]. To evaluate AT-PR, we compare it with two standard AT-WCR (PGD and FGSM) and four advanced methods: TRADES [66], MART [57], ALP, and CLP [29]. Implementation details are in Appendix B.", + "bbox": [ + 511, + 250, + 906, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Effectiveness and Efficiency of AT-PR", + "text_level": 1, + "bbox": [ + 511, + 364, + 839, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assess AT-PR's effectiveness against AT-WCR from two perspectives: 1) improvement in PR and 2) impact on generalization. We compare models trained with AT-PR, AT-WCR (FGSM and PGD) and a baseline without AT, using three metrics: (i) Accuracy on clean test data; (ii) Adversarial Accuracy under PGD, C&W, and AA attacks to reflect WCR; and (iii) PR, computed as $PR(x, \\gamma)$ for various $\\gamma$ using the estimator from [59]. We report the mean and standard deviation of PR on the test set, along with the generalization error (GE) to measure train-test robustness gaps.", + "bbox": [ + 511, + 387, + 906, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Key findings summarized from Table 1:", + "bbox": [ + 532, + 539, + 795, + 551 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i) All original models without AT show low WCR. As expected, PGD is the most effective in improving WCR, though it compromises generalization the most. AT-PR also enhances WCR, outperforming FGSM and, in some cases, even surpassing PGD, as seen with SVHN/ResNet-18, Tiny-ImageNet/ResNet-34 for PGD.Ac., and SVHN/WRN50-2, Tiny-ImageNet/ResNet-18, ResNet-34 for CW.Ac.", + "ii) Both AT-WCR with FGSM and PGD can improve PR compared to the original models without AT, but they are less effective when compared to the performance of AT-PR. Specifically, the mean PR evaluations of AT-PR-trained models are higher than those trained with AT-WCR across most datasets, model architectures and norm-ball sizes. For instance, on CIFAR-100 with ResNet-18, AT-PR improves $PR(\\cdot, 0.15)$ from 44.71 (FGSM) and 80.47 (PGD) to 87.38.", + "iii) Compared to AT-WCR, the improvement in PR becomes more pronounced as the norm-ball size $\\gamma$ increases, which is anticipated since larger norm-balls generally lead to less robust models. This demonstrates that our AT-PR approach is particularly useful when enhancing PR becomes more challenging for traditional AT-WCR.", + "iv) Each local input is associated with an estimated PR, and we report the mean and standard deviation over the test" + ], + "bbox": [ + 511, + 553, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1679", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20f26f73625add2a7c250335597fcb2aadb9edb9a7eb8847306d164fb4617848.jpg", + "table_caption": [ + "Table 1. Evaluation results of DL models trained with different AT schemes, showing standard accuracy, adversarial accuracy (indicating WCR: PGD-10, C&W-10), and PR estimated (mean ± std) for four norm-ball sizes $\\gamma$ . All results are calculated over the test datasets." + ], + "table_footnote": [], + "table_body": "
DatasetModelAT SchemeAccu. %PGD Ac. %CW Ac. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18W/O AT94.960.020.5178.38 ± 33.1466.3 ± 38.0653.41 ± 40.7637.01 ± 39.77
AT-WCR(FGSM)88.60.090.0993.99 ± 17.588.02 ± 25.0380.42 ± 31.8167.77 ± 38.6
AT-WCR(PGD)8145.0144.5697.28 ± 13.2995.77 ± 17.0794.1 ± 20.289.78 ± 26.53
AT-PR80.1439.7739.0398.22±10.3897.28±13.1696.05±16.1893.21±21.47
WRN-50-2W/O AT94.140.230.1476.96±32.7665.65±37.4154.43±39.8840.19±39.49
AT-WCR(FGSM)92.040.080.0495.68±14.5391.9±20.2586.67±25.9876.79±33.1
AT-WCR(PGD)82.6249.5148.1397.86±12.6197.08±14.5495.84±17.1492.62±22.87
AT-PR82.7143.2743.0798.09±10.8197.33±13.0396.22±15.5893.56±20.89
CIFAR-100ResNet-18W/O AT75.50.020.044.58 ± 42.6432.1 ± 40.123.32 ± 36.4115.62 ± 31.57
AT-WCR(FGSM)67.801.410.6683.90 ± 31.1172.89 ± 37.9461.06 ± 42.2344.71 ± 43.04
AT-WCR(PGD)52.8519.9319.5195.4 ± 18.1592.7 ± 22.7288.98 ± 28.1980.47 ± 36.4
AT-PR53.2119.3818.3997.59±12.9695.96±16.6293.15±21.7787.38±29.63
WRN-50-2W/O AT75.460.490.4551.11±40.5437.96±39.2328.33±36.7219.11±31.97
AT-WCR(FGSM)73.170.290.185.68±27.4576.62±33.8966.37/38.6652.04±41.42
AT-WCR(PGD)57.4623.1722.595.31±17.3892.0±23.7988.14±29.1580.67±36.21
AT-PR59.4622.0621.5996.05±16.1793.72±20.4789.48±26.8781.6±34.86
SVHNResNet-18W/O AT94.960.080.1297.11 ± 9.6695.63 ± 11.9893.74 ± 14.4890.12 ± 18.40
AT-WCR(FGSM)93.494.685.3298.79 ± 7.0698.42 ± 7.7097.86 ± 8.6796.57 ±10.78
AT-WCR(PGD)89.2144.5743.7098.96 ± 5.8598.50 ± 7.4097.72 ± 9.6195.41 ± 14.41
AT-PR90.2945.8142.2199.12 ± 5.1598.87 ± 5.9398.54 ± 6.8497.84 ± 8.60
WRN-50-2W/O AT95.911.741.2996.60 ± 10.3694.77 ± 13.0192.40 ± 15.8787.97 ± 20.04
AT-WCR(FGSM)94.318.138.5498.34 ± 7.4997.71 ± 8.3996.80 ± 9.6995.12 ± 12.06
AT-WCR(PGD)88.9154.9950.5898.72 ± 6.4498.28 ± 7.6697.76 ± 8.9596.77 ± 11.18
AT-PR91.9752.2251.5898.88 ± 6.1298.38 ± 7.5397.72 ± 9.1996.26 ± 12.35
Tiny-ImageNetResNet-18W/O AT59.260.00.080.78 ± 33.172.95 ± 37.6964.17 ± 40.9151.14 ± 42.91
AT-WCR(FGSM)54.810.050.0189.9 ± 24.9785.33 ± 29.9580.18 ± 33.7469.03 ± 39.39
AT-WCR(PGD)47.4321.319.3196.32 ± 17.0193.79 ± 22.190.16 ± 27.7682.46 ± 35.97
AT-PR51.6121.119.8998.23 ± 11.0696.1 ± 17.8493.12 ± 23.3286.65 ± 32.26
ResNet-34W/O AT61.660.00.082.39 ± 31.375.2 ± 36.066.95 ± 39.4253.7 ± 42.17
AT-WCR(FGSM)570.130.0290.67 ± 23.9285.81 ± 29.580.57 ± 34.1370.22 ± 39.36
AT-WCR(PGD)49.3320.9719.5895.45 ± 18.5593.08 ± 23.4889.85 ± 28.383.53 ± 34.99
AT-PR53.9821.7220.9596.71 ± 15.7495.02 ± 19.8992.5 ± 23.8786.23 ± 32.29
", + "bbox": [ + 94, + 127, + 901, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set in Table 1. Notably, AT-PR consistently achieves higher mean PR compared to models trained with AT-WCR, while also exhibiting lower standard deviations. This indicates that AT-PR not only improves PR but also offers more stable and consistent performance across different local inputs.", + "bbox": [ + 89, + 527, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\nu$ ) While PGD significantly improves robustness, it often reduces accuracy [66]. In contrast, AT-PR has a smaller negative impact on accuracy compared to PGD, with almost all models achieving higher accuracy than those trained using PGD while also enhancing both WCR and PR. This effect is more pronounced on larger datasets—for example, on Tiny-ImageNet with ResNet-18, AT-PR achieves an accuracy of $53.98\\%$ compared to $49.33\\%$ for PGD.", + "bbox": [ + 89, + 606, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows that combining AT-PR with TRADES improves clean accuracy and PR on more complex attention-based models ViT (Vanilla ViT and DeiT) across different scales (DeiT-Tiny, ViT-Small, ViT-Base) for CIFAR10. With just five epochs of AT-PR fine-tuning, all three ViT models achieve higher clean accuracy and PR while maintaining comparable WCR, showing a slight decrease on PGD-20 and AA but improved performance on CW-20.", + "bbox": [ + 89, + 729, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents a comparison of AT-PR with four advanced AT methods (TRADES, MART, ALP and CLP) across three datasets using ResNet-18 and WRN-50-2. Re", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sults show that AT-PR consistently achieves the best PR performance across four different norm-ball sizes $\\gamma$ . Additionally, to verify that AT-PR is compatible with efficient AT methods [46, 62], we replace PGD with Fast-AT [62] in our framework (step 1 in Algo. 1). The result shows that Fast-AT-PR achieves $\\sim 40\\%$ speedup over the original AT-PR, with only $\\sim 50\\%$ extra cost compared to other AT baselines. Notably, Fast-AT-PR retains comparable PR performance to AT-PR and continues to outperform other AT baselines.", + "bbox": [ + 511, + 527, + 906, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Generalization Error of AT-PR", + "text_level": 1, + "bbox": [ + 511, + 676, + 790, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate the generalization error (GE) (train-test gap) of AT-PR to gain insights into its effectiveness. Table 4 compares the PR GE between AT-PR and PGD. Models trained with AT-PR consistently yields lower GE of up $2.2\\%$ on CIFAR-10, $2.87\\%$ on CIFAR-100, and $9.66\\%$ on TinyImageNet, significantly lower than PGD's $4.67\\%$ , $8.54\\%$ , and $11.94\\%$ , respectively. We further analyze the underlying reasons based on the PAC-Bayes analysis for Adversarial Robustness [54], see Appendix C for details.", + "bbox": [ + 511, + 699, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. AEs Used in AT-WCR vs. AT-PR", + "text_level": 1, + "bbox": [ + 511, + 847, + 803, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the AEs used during training to understand how AT-PR differs from AT-WCR. By comparing the AEs gen", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1680", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0132cb98599face1bc51aad2917e9b2cebc0ba4ad1a77bfb7457a78d0cd5265c.jpg", + "table_caption": [ + "Table 2. ViT models trained on CIFAR-10 with TRADES and AT-PR, showing standard accuracy, adversarial accuracy (indicating WCR: PGD, C&W, AA), and PR estimated (mean ± std) for four norm-ball sizes $\\gamma$ ." + ], + "table_footnote": [], + "table_body": "
DatasetModelAT SchemeAccu. %PGD-20. %CW-20. %AA. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10DeiT-TiTRADES79.4449.7846.4545.9696.61 ± 15.2394.73 ± 19.3692.69 ± 23.0489.18 ± 27.9
+AT-PR81.0749.0846.9245.9597.26 ± 14.095.94 ± 17.1294.16 ± 20.590.65 ± 25.88
Vit-STRADES81.0350.9648.1447.4096.93 ± 14.9995.48 ± 18.493.58 ± 21.6290.06 ± 26.81
+AT-PR83.2950.4448.2047.3497.76 ± 11.8996.03 ± 16.4693.98 ± 20.6990.48 ± 26.19
Vit-BTRADES83.5553.1450.9449.9697.34 ± 13.5795.87 ± 16.9993.78 ± 21.1790.77 ± 25.88
+AT-PR86.0352.9451.2449.3497.9 ± 11.896.49 ± 15.294.44 ± 19.6291.02 ± 25.27
", + "bbox": [ + 94, + 127, + 903, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec51dd576a30f3c52df87d336afe93a11e2262fb7f2167027f67ae36b276ef74.jpg", + "table_caption": [ + "Table 3. Comparison of PR(, γ) and training time (sec/epoch) for more recent AT methods (TRADES, MART, ALP, CLP) and Fast-AT-PR." + ], + "table_footnote": [], + "table_body": "
DataModelMetricAT Scheme
PGDTRADESMARTALPCLPFast-AT-PRAT-PR
CIFAR-10ResNet-18PR(,0.08)%97.28 ± 13.2997.74 ± 12.4297.4 ± 13.796.73 ± 15.997.61 ± 13.0698.12 ± 11.5298.22 ± 10.38
PR(,0.10)%95.77 ± 17.0796.33 ± 16.3995.92 ± 17.4395.16 ± 19.5396.29 ± 16.6697.1 ± 14.3497.28 ± 13.16
PR(,0.12)%94.10 ± 20.2094.55 ± 20.3193.93 ± 21.3893.24 ± 23.0894.54 ± 20.2895.4 ± 17.9096.05 ± 16.18
PR(,0.15)%89.78 ± 26.5391.22 ± 25.7090.35 ± 26.9189.76 ± 28.1490.86 ± 26.2392.71 ± 23.5493.21 ± 21.47
Time (s/ep.)546144646497184
WRN-50-2PR(,0.08)%97.86 ± 12.6197.55 ± 13.1497.40 ± 13.6197.75 ± 12.4397.40 ± 13.9197.96 ± 10.8298.09 ± 10.81
PR(,0.10)%97.08 ± 14.5496.14 ± 16.7096.15 ± 16.7896.42 ± 16.1196.20 ± 17.5897.22 ± 12.9697.33 ± 13.03
PR(,0.12)%95.84 ± 17.1494.33 ± 20.4194.41 ± 20.3794.64 ± 20.0194.95 ± 19.8896.22 ± 15.6296.22 ± 15.58
PR(,0.15)%92.62 ± 22.8791.01 ± 25.7991.11 ± 25.6591.22 ± 25.7091.48 ± 25.6994.74 ± 19.8893.56 ± 20.89
Time (s/ep.)13110571155157229305
CIFAR-100ResNet-18PR(,0.08)%95.40 ± 18.1595.02 ± 19.2795.45 ± 18.1196.20 ± 16.0796.16 ± 16.3697.09 ± 13.4897.59 ± 12.96
PR(,0.10)%92.70 ± 22.7292.09 ± 24.2292.73 ± 23.3994.40 ± 20.0893.81 ± 21.4294.67 ± 18.9195.96 ± 16.62
PR(,0.12)%88.98 ± 28.1987.80 ± 29.8089.35 ± 28.3292.05 ± 24.1291.05 ± 25.8692.63 ± 23.9193.15 ± 21.77
PR(,0.15)%80.47 ± 36.4079.20 ± 37.7381.93 ± 35.7285.89 ± 31.6584.97 ± 33.0085.87 ± 31.9987.38 ± 29.63
Time (s/ep.)556143646597180
WRN-50-2PR(,0.08)%95.31 ± 17.3895.24 ± 18.6195.30 ± 16.8795.34 ± 17.7995.29 ± 17.7096.52 ± 15.1196.05 ± 16.17
PR(,0.10)%92.00 ± 23.7992.30 ± 23.6192.59 ± 22.8192.36 ± 23.1991.79 ± 23.8793.49 ± 20.9393.72 ± 20.47
PR(,0.12)%88.14 ± 29.1588.65 ± 28.4688.38 ± 28.6588.51 ± 28.2287.54 ± 29.9288.82 ± 28.3889.48 ± 26.87
PR(,0.15)%80.67 ± 36.2180.35 ± 36.5480.79 ± 36.0180.37 ± 36.8679.71 ± 36.7681.02 ± 37.0981.60 ± 34.86
Time (s/ep.)13110270155156225312
SVHNResNet-18PR(,0.08)%98.96 ± 5.8599.03 ± 5.7698.99 ± 6.1599.05 ± 5.7399.12 ± 5.3398.90 ± 5.9799.12 ± 5.15
PR(,0.10)%98.50 ± 7.4098.46 ± 8.2798.31 ± 8.2098.63 ± 7.1598.60 ± 7.2198.53 ± 7.2498.87 ± 5.93
PR(,0.12)%97.72 ± 9.6197.30 ± 11.0897.10 ± 11.2697.85 ± 9.1397.66 ± 9.7898.11 ± 8.1998.54 ± 6.84
PR(,0.15)%95.41 ± 14.4194.15 ± 16.5794.12 ± 16.6595.54 ± 14.1094.93 ± 15.0097.10 ± 10.0597.84 ± 8.60
Time (s/ep.)7886619193141202
WRN-50-2PR(,0.08)%98.72 ± 6.4498.69 ± 6.9098.61 ± 7.6698.54 ± 7.5298.59 ± 7.4698.43 ± 6.6998.88 ± 6.12
PR(,0.10)%98.28 ± 7.6698.11 ± 8.6698.07 ± 9.2798.12 ± 8.7498.00 ± 8.8998.15 ± 7.2998.38 ± 7.53
PR(,0.12)%97.76 ± 8.9597.26 ± 10.8997.34 ± 10.9097.65 ± 8.8596.98 ± 11.1397.79 ± 8.1297.72 ± 9.19
PR(,0.15)%96.77 ± 11.1895.20 ± 14.8495.45 ± 14.7295.67 ± 14.3294.74 ± 15.3796.94 ± 9.8096.26 ± 12.35
Time (s/ep.)190146102223225321420
", + "bbox": [ + 112, + 253, + 883, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "erated by PGD and AT-PR, we find that about $10\\%$ of training inputs produce different AEs (e.g., $10.05\\% / 8.39\\%$ for ResNet-18 and $8.83\\% / 10.42\\%$ for WRN50-2 on CIFAR-10/CIFAR-100). This means that for roughly $90\\%$ of inputs, the loss landscapes align with Case 2 (\"tall & wide vs. short & tight\") in Fig. 2, where both methods select the same AEs. However, for the remaining $10\\%$ -ish, where the landscape fits Case 1 (\"tall & tight vs. short & wide\"), AT-PR identifies different AEs that better optimize PR.", + "bbox": [ + 88, + 637, + 482, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 (top row) compares the loss of AEs generated by PGD and AT-PR. For all models, PGD AEs consistently have higher losses, reflecting its objective to find AE with the maximum loss for AT. This aligns with the goal of AT-WCR, which seeks to minimize the loss on such worst-case AEs (see Fig.2, top-middle). In contrast, AT-PR generates AEs with lower loss. Combined with the fact that our Algorithm 1 searches for examples farthest from the deci", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sion boundary, suggests that AT-PR effectively locates AEs within the widest peak of the loss landscape. Training with these AEs helps reduce the \"short & wide\" peak (see Fig. 2, top-right).", + "bbox": [ + 511, + 637, + 906, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 (b) visualizes the distributions of all generated AEs $\\{x_i^{\\prime}\\}_{i = 1}^{n}$ by analyzing the distance $(L_{2}$ norm) between each $x_{i}^{\\prime}$ and the original input $x$ (e.g., $d_0,d_1$ in Fig. 5 (a)), along with their angular relationships. The 2D coordinate in Fig. 5 (a) introduces a fixed anchor $x_{a}$ near $x$ , for an AE $x_0^\\prime$ , the angle $\\theta_0$ is computed between $\\vec{v}_{x\\rightarrow x_0'}$ and $\\vec{v}_{x\\rightarrow x_a}$ , representing $x_0'$ 's relative position, given by: $\\theta_0 = \\arccos \\left(\\frac{\\vec{v}_{x\\rightarrow x_0'}\\cdot\\vec{v}_{x\\rightarrow x_a}}{\\|\\vec{v}_{x\\rightarrow x_0'}\\||\\vec{v}_{x\\rightarrow x_a}\\|}\\right)$ . The resulting angles range from approximately $88^{\\circ}$ to $92^{\\circ}$ Fig. 5 (b), reflecting the unique geometric properties of high-dimensional spaces, where random vectors highly likely to be orthogonal [53]. To further capture the variability in relative positions,", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1681", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad1331b8c313fa988b50992fd8fbb578b49e8a9d1dfd0d443453730bd8e99e5a.jpg", + "table_caption": [ + "Table 4. Comparative GE experiment on DL models across different datasets. Train/test set evaluations of PR estimates (mean) for four norm-ball sizes $(\\gamma)$ across models trained with PGD and AT-PR, presented as $PR_{train} / PR_{test} / (PR_{train} - PR_{test})$" + ], + "table_footnote": [], + "table_body": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
WRN-50-2PGD98.88/97.86/1.0198.19/97.08/1.197.49/95.84/1.6495.3/92.62/2.67
AT-PR98.63/98.09/0.5497.8/97.33/0.4796.71/96.22/0.4694.9/93.56/1.34
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
WRN-50-2PGD98.35/95.31/3.0496.71/92.0/4.7193.67/88.14/5.5386.69/80.67/6.02
AT-PR97.3/96.05/1.2595.41/93.72/1.6991.93/89.48/2.4583.25/81.6/1.65
Tiny-ImageNetResNet-18PGD98.72/96.32/2.497.86/93.79/4.0796.94/90.16/6.7894.4/82.46/11.94
AT-PR99.01/98.23/0.7898.02/96.1/1.9296.68/93.12/3.5694.31/86.65/7.66
ResNet-34PGD98.67/95.45/3.2298.02/93.08/6.9496.9/89.85/7.0593.36/83.53/9.83
AT-PR99.28/96.71/2.5798.82/95.02/3.898.11/92.5/5.6195.89/86.23/9.66
", + "bbox": [ + 96, + 127, + 903, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/778121e37fcf5c7b726e33e62aceab1291063f57c4cb38f830a2bc45b263caf2.jpg", + "table_caption": [ + "Table 5. PR (mean) on train/test sets, reported as: ${PR}_{\\text{train }}/{PR}_{\\text{test }}/\\left( {{PR}_{\\text{train }} - P{R}_{\\text{test }}}\\right)$ ,for model trained with PGD, Random-AE, AT-PR." + ], + "table_footnote": [], + "table_body": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
Random98.62/97.71/0.9197.73/96.78/0.9596.58/95.48/1.194.19/92.64/1.54
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
Random96.6/95.87/0.7394.85/93.77/1.0892.18/90.92/1.2685.39/85.21/0.18
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
", + "bbox": [ + 94, + 353, + 903, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce2f8a4c449f5f3d9bf5a24c5f87d0ddfe062327b258dc554d9233bde8b8847a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 483, + 215, + 563 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c53a0c6169b3a8975f751a0993ad809813e9ee9ca9d59962975d0b870946566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 484, + 344, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d38565850766f48f6710a0e15fb24e758a3b910ecaf73ef27e54942e04a01449.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 483, + 478, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6944ecc0522a75de4e80259b8222c967df4093025b1850393d951237bc4db04d.jpg", + "image_caption": [ + "Figure 5. (Top row) Averaged loss of AEs used during AT for AT-PR and PGD. (Bottom row) (a) Illustration of angles and distances of AEs relative to $x$ . (b) Distribution of generated AEs. (c) Pairwise angular distribution among AEs." + ], + "image_footnote": [], + "bbox": [ + 94, + 568, + 207, + 659 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c11df29aff62af65b87111c9d103e94deb370b57456449e113dfa9066ddab0e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 563, + 349, + 660 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ee19c9c6a7cd7cf84f3e63d92f457a3e2f9a3a8b57392338500c5b1526c2ab2.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 352, + 563, + 480, + 646 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we computed the Pairwise Angular Distribution among all vectors $\\{\\vec{v}_{x\\rightarrow x_i'}\\}_{i = 1}^n$ . The angle $\\beta$ for each pair of vectors is calculated as illustrated in Fig. 5 (a), with the resulting distribution shown in Fig. 5 (c). This analysis shows that the angles between all pairs of vectors range from $60^{\\circ}$ to $85^{\\circ}$ , indicating that the generated AEs are widely spread within the norm-ball. The final AE selected by AT-PR, highlighted in orange in Fig. 5 (b), has the largest distance to nearest decision boundary (1.0 for the L2 norm), compared to other", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AEs, which range from 0.76 to 0.92.", + "bbox": [ + 513, + 486, + 756, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To empirically validate that AEs generated by AT-PR are most effective for improving PR, we conduct an ablation study by randomly sampling AEs from those generated in Step 1 of Algo. 1 (also illustrated as random blue spots in Fig.5 (b)). Table 5 shows that while randomly selected AEs improve PR and GE compared to PGD, AT-PR consistently performs better. As Fig. 2 illustrates, although random AEs may fall within a short and wide peak (explaining their improved PR over PGD), they do not necessarily lie within the \"widest peak\" that AT-PR specifically targets. This distinction explains AT-PR's superior PR performance.", + "bbox": [ + 511, + 501, + 906, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, for the first time, we introduce a novel AT method targeting PR, named as AT-PR. By reformulating a new min-max optimization, we lay the theoretical foundation for AT-PR and develop specialized numerical algorithms to solve this optimization. Extensive experimental evaluation demonstrates that AT-PR consistently outperforms traditional AT methods designed for WCR in improving PR. Additionally, we show that AT-PR incurs a smaller trade-off in model generalization compared to common AT-WCR methods.", + "bbox": [ + 511, + 703, + 906, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the future, we plan to address the challenges identified in [73], such as benchmarking PR and extending it to diverse generative AI tasks.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1682", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements SK and XZ have received funding from the European Union's EU Framework Program for Research and Innovation Europe Horizon (grant agreement No 101202457). XZ's contribution is also supported by the UK EPSRC New Investigator Award [EP/Z536568/1] and NVIDIA Academic Grant Program. SK's contribution is supported by the UKRI Future Leaders Fellowship Grant [MR/S035176/1], and Innovate UK through the DriveSafeAI Project (10062594). XH has received funding from the European Union's Horizon Europe Research and Innovation Action under grant agreement No 101212818. YZ's contribution is supported by China Scholarship Council. We thank Zheng Wang for his valuable comments.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them.", + "bbox": [ + 89, + 289, + 480, + 364 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 382, + 187, + 398 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amir Aminifar. Universal adversarial perturbations in epileptic seizure detection. In 2020 Int. Joint Conference on Neural Networks (IJCNN), pages 1-6. IEEE, 2020. 1", + "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Proc. of the 35th International Conference on Machine Learning, pages 274-283. PMLR, 2018. 1", + "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Proc. of the 30th Int. Joint Conference on Artificial Intelligence, pages 4312-4321, 2021. 1, 3", + "[4] Mislav Balunović and Martin Vechev. Adversarial training and provable defenses: Bridging the gap. In 8th International Conference on Learning Representations (ICLR'20'20), 2020. 1, 3", + "[5] Teodora Baluta, Zheng Leong Chua, Kuldeep S Meel, and Prateek Saxena. Scalable quantitative verification for deep neural networks. In ICSE'21, pages 312-323, 2021. 1", + "[6] Qi-Zhi Cai, Chang Liu, and Dawn Song. Curriculum adversarial training. In Proc. of the 27th International Joint Conference on Artificial Intelligence, page 3740-3747. AAAI Press, 2018. 3", + "[7] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE symposium on security and privacy (sp), pages 39-57, 2017. 1", + "[8] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. A survey on adversarial attacks and defences. CAAI Transactions on Intelligence Technology, 6(1):25-45, 2021. 1", + "[9] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In IEEE symposium on security and privacy (SP), pages 1277-1294. IEEE, 2020. 4, 12" + ], + "bbox": [ + 99, + 407, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. ZOO: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proc. of the 10th ACM Workshop on Artificial Intelligence and Security, page 15-26. ACM, 2017. 2", + "[11] Zhen Chen, Yi Zhang, Fu Wang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. TARP-VP: Towards evaluation of transferred adversarial robustness and privacy on label mapping visual prompting models. Advances in Neural Information Processing Systems, 37:6776-6796, 2024. 1", + "[12] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In International Conference on Machine Learning, pages 1310-1320. PMLR, 2019. 1", + "[13] Nicolas Couellan. Probabilistic robustness estimates for feed-forward neural networks. Neural networks, 142:138-147, 2021. 1", + "[14] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1", + "[15] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2020. 3", + "[16] Yi Dong, Wei Huang, Vibhav Bharti, Victoria Cox, Alec Banks, Sen Wang, Xingyu Zhao, Sven Schewe, and Xiaowei Huang. Reliability assessment and safety arguments for machine learning components in system assurance. ACM Transactions on Embedded Computing Systems, 22(3):1-48, 2023. 1, 2", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5", + "[18] Chengbin Du, Yanxi Li, Zhongwei Qiu, and Chang Xu. Stable diffusion is unstable. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[19] Xiaoning Du, Xiaofei Xie, Yi Li, Lei Ma, Yang Liu, and Jianjun Zhao. Deepstellar: Model-based quantitative analysis of stateful deep learning systems. In Proc. of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 477–487, 2019. 1", + "[20] Timon Gehr, Matthew Mirman, Dana Drachsler-Cohen, Petar Tsankov, Swarat Chaudhuri, and Martin Vechev. Ai2: Safety and robustness certification of neural networks with abstract interpretation. In IEEE symposium on security and privacy (SP), pages 3-18, 2018. 1", + "[21] Ian J Goodfellow. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR'15), 2015. 1, 2, 3", + "[22] Sven Gowal, Krishnamurthy Dj Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1683", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Arandjelovic, Timothy Mann, and Pushmeet Kohli. Scalable verified training for provably robust image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4842-4851, 2019. 1, 3", + "[23] Wei Huang, Youcheng Sun, Xingyu Zhao, James Sharp, Wenjie Ruan, Jie Meng, and Xiaowei Huang. Coverage-Guided Testing for Recurrent Neural Networks. IEEE Transactions on Reliability, 71(3):1191-1206, 2022. 1", + "[24] Wei Huang, Xingyu Zhao, Alec Banks, Victoria Cox, and Xiaowei Huang. Hierarchical distribution-aware testing of deep learning. ACM Transactions on Software Engineering and Methodology, 33(2):1-35, 2023. 1", + "[25] Wei Huang, Xingyu Zhao, Gaojie Jin, and Xiaowei Huang. Safari: Versatile and efficient evaluations for robustness of interpretability. In IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 1988-1998, 2023. 3", + "[26] Xiaowei Huang, Marta Kwiatkowska, Sen Wang, and Min Wu. Safety verification of deep neural networks. In Computer Aided Verification: 29th International Conference, CAV 2017, Heidelberg, Germany, July 24-28, 2017, Proceedings, Part I 30, pages 3-29. Springer, 2017. 1", + "[27] Xiaowei Huang, Daniel Kroening, et al. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 37:100270, 2020. 2, 3", + "[28] Xiaowei Huang, Wenjie Ruan, et al. A survey of safety and trustworthiness of large language models through the lens of verification and validation. Artificial Intelligence Review, 57 (7):175, 2024. 1", + "[29] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. Adversarial logit pairing. arXiv preprint arXiv:1803.06373, 2018. 3, 5, 12", + "[30] Guy Katz, Clark Barrett, David L Dill, Kyle Julian, and Mykel J Kochenderfer. Reluplex: An efficient smt solver for verifying deep neural networks. In 29th Int. Conf. Computer Aided Verification, pages 97-117. Springer, 2017. 1", + "[31] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In Proc. of the IEEE/CVF conference on computer vision and pattern recognition, pages 1221-1230, 2020. 4", + "[32] Han Liu, Yuhao Wu, Shixuan Zhai, Bo Yuan, and Ning Zhang. Riatig: Reliable and imperceptible adversarial text-to-image generation with natural prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20585-20594, 2023. 1", + "[33] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Discovering failure modes of text-guided diffusion models via adversarial search. In The 12th International Conference on Learning Representations, 2024. 1", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5", + "[35] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deepauge: Multi-granularity testing criteria for deep" + ], + "bbox": [ + 91, + 90, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning systems. In Proce. of the 33rd ACM/IEEE Int. Conference on Automated Software Engineering (ASE'18), pages 120-131, 2018. 1", + "[36] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 4, 12", + "[37] Mark Huasong Meng, Guangdong Bai, Sin Gee Teo, Zhe Hou, Yan Xiao, Yun Lin, and Jin Song Dong. Adversarial robustness of deep neural networks: A survey from a formal verification perspective. IEEE Transactions on Dependable and Secure Computing, 2022. 1", + "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 1", + "[39] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 1", + "[40] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1", + "[41] Mikhail Pautov, Nurislam Tursynbek, Marina Munkhoeva, Nikita Muravev, Aleksandr Petiushko, and Ivan Oseledets. Cc-cert: A probabilistic approach to certify general robustness of neural networks. In AAAI'22, pages 7975-7983, 2022. 1", + "[42] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepexplore: Automated whitebox testing of deep learning systems. In Proceedings of the 26th Symposium on Operating Systems Principles, pages 1-18. ACM, 2017. 1", + "[43] Alexander Robey, Luiz Chamon, George J Pappas, and Hamed Hassani. Probabilistically robust learning: Balancing average and worst-case performance. In Int. Conference on Machine Learning, pages 18667-18686. PMLR, 2022. 3", + "[44] Wenjie Ruan, Xiaowei Huang, and Marta Kwiatkowska. Reachability analysis of deep neural networks with provable guarantees. In Proc. of the 27th Int. Joint Conference on Artificial Intelligence (IJCAI'18), pages 2651-2659, 2018. 1", + "[45] Sebastian Ruder. An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747, 2016. 12", + "[46] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 5, 6", + "[47] Uri Shaham, Yutaro Yamada, and Sahand Negahban. Understanding adversarial training: Increasing local stability of supervised models through robust optimization. Neurocomputing, 307:195-204, 2018. 3", + "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In In Proc. of 2nd Int. Conf. on Learning Representations, 2014. 1" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1684", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Karim Tit, Teddy Furon, and Mathias Rousset. Efficient statistical assessment of neural network corruption robustness. NeurIPS'21, 34:9253-9263, 2021. 1", + "[50] Karim TIT, Teddy Furon, and Mathias Rousset. Gradient-informed neural network statistical robustness estimation. In Proc. of The 26th Int. Conf. on Artificial Intelligence and Statistics, pages 323-334. PMLR, 2023. 1, 3", + "[51] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5", + "[52] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 1, 3", + "[53] Princeton University. Lecture 11: Rademacher complexity, 2014. Accessed: 2025-01-05. 7", + "[54] Paul Viallard, Eric Guillaume VIDOT, Amaury Habrard, and Emilie Morvant. A pac-bayes analysis of adversarial robustness. Advances in Neural Information Processing Systems, 34:14421-14433, 2021. 6, 12", + "[55] Benjie Wang, Stefan Webb, and Tom Rainforth. Statistically robust neural network classification. In Uncertainty in Artificial Intelligence, pages 1735-1745. PMLR, 2021. 1, 3", + "[56] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. In ICML'19, pages 6586-6595. PMLR, 2019. 1", + "[57] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 3, 5, 12", + "[58] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 3", + "[59] Stefan Webb, Tom Rainforth, Yee Whye Teh, and M Pawan Kumar. A statistical approach to assessing neural network robustness. In Int. Conf. on Learning Representations, 2019, 1, 2, 3, 5", + "[60] Lily Weng, Pin-Yu Chen, Lam Nguyen, Mark Squillante, Akhilan Boopathy, Ivan Oseledets, and Luca Daniel. Proven: Verifying robustness of neural networks with a probabilistic approach. In Int. Conf. on Machine Learning, pages 6727-6736. PMLR, 2019. 1", + "[61] Tsui Wei Weng, Huan Zhang, Pin-Yu Chen, Jinfeng Yi, Dong Su, Yupeng Gao, Cho-Jui Hsieh, and Luca Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. In Int. Conf. on Learning Representations, 2018. 1", + "[62] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 5, 6", + "[63] Weiming Xiang, Hoang-Dung Tran, and Taylor T. Johnson. Output reachable set estimation and verification for multilayer neural networks. IEEE Tran. on Neural Networks and Learning Systems, 29(11):5777-5783, 2018. 1" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Xiangyu Yin, Sihao Wu, Jiaxu Liu, Meng Fang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. Representation-based robustness in goal-conditioned reinforcement learning. In Proc. of the AAAI Conference on Artificial Intelligence, pages 21761-21769, 2024. 1", + "[65] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE Tran. on neural networks and learning systems, 30(9):2805-2824, 2019. 1", + "[66] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3, 5, 6, 12", + "[67] Jingfeng Zhang, Xilie Xu, Bo Han, Gang Niu, Lizhen Cui, Masashi Sugiyama, and Mohan Kankanhalli. Attacks which do not kill training make adversarial learning stronger. In International conference on machine learning, pages 11278-11287. PMLR, 2020. 3", + "[68] Jiawei Zhang, Linyi Li, Huichen Li, Xiaolu Zhang, Shuang Yang, and Bo Li. Progressive-scale boundary blackbox attack via projective gradient estimation. In Int. Conference on Machine Learning, pages 12479-12490. PMLR, 2021. 4", + "[69] Tianle Zhang, Wenjie Ruan, and Jonathan E. Fieldsend. Proa: A probabilistic robustness assessment against functional perturbations. In Machine Learning and Knowledge Discovery in Databases (ECML/PKDD'22), page 154-170. Springer-Verlag, 2023. 1, 2, 3", + "[70] Tianle Zhang, Yanghao Zhang, Ronghui Mu, Jiaxu Liu, Jonathan Fieldsend, and Wenjie Ruan. Prass: probabilistic risk-averse robust learning with stochastic search. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence, pages 559–567, 2024. 3", + "[71] Yi Zhang, Zhen Chen, Chih-Hong Cheng, Wenjie Ruan, Xiaowei Huang, Dezong Zhao, David Flynn, Siddartha Khastgir, and Xingyu Zhao. Trustworthy text-to-image diffusion models: A timely and focused survey. arXiv preprint arXiv:2409.18214, 2024. 1", + "[72] Yi Zhang, Yun Tang, Wenjie Ruan, Xiaowei Huang, Siddartha Khastgir, Paul Jennings, and Xingyu Zhao. Protip: Probabilistic robustness verification on text-to-image diffusion models against stochastic perturbation. In European Conference on Computer Vision, 2024. 1, 2, 3", + "[73] Xingyu Zhao. Probabilistic robustness in deep learning: A concise yet comprehensive guide. Adversarial Example Detection and Mitigation Using Machine Learning, pages 1-13, 2025. 1, 2, 3, 8", + "[74] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2385-2392, 2023. 1", + "[75] Monty-Maximilian Zuhlke and Daniel Kudenko. Adversarial robustness of neural networks from the perspective of Lipschitz calculus: A survey. ACM Computing Surveys, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 866 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1685", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_model.json b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ad03be59979a44f44aa4b493890aa740217b327 --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "Adversarial Training for Probabilistic Robustness" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.18, + 0.816, + 0.27 + ], + "angle": 0, + "content": "Yi Zhang\\(^{1}\\), Yuhang Chen\\(^{1}\\), Zhen Chen\\(^{2}\\), Wenjie Ruan\\(^{2}\\), Xiaowei Huang\\(^{2}\\), Siddartha Khastgir\\(^{1}\\), Xingyu Zhao\\(^{1*}\\) \n\\(^{1}\\)WMG, University of Warwick \\(^{2}\\)University of Liverpool {yi.zhang.16,yuhang.chen.1,s.khastgir.1,xingyu.zhao}@warwick.ac.uk; {zhen.chen2,xiaowei.huang}@liverpool.ac.uk; w.ruan@trustai.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.684 + ], + "angle": 0, + "content": "Deep learning (DL) has shown transformative potential across industries, yet its sensitivity to adversarial examples (AEs) limits its reliability and broader deployment. Research on DL robustness has developed various techniques, with adversarial training (AT) established as a leading approach to counter AEs. Traditional AT focuses on worst-case robustness (WCR), but recent work has introduced probabilistic robustness (PR), which evaluates the likelihood of AEs within a local perturbation range, providing an overall assessment of the model's robustness and acknowledging residual risks that are more practical to manage. However, existing AT methods are fundamentally designed to improve WCR, and no dedicated methods currently target PR. To bridge this gap, we formulate a new min-max optimization as the theoretical foundation for PR-focused AT, and introduce an AT-PR training scheme with numerical algorithms to solve the new optimization problem. Our experiments, based on 70 DL models trained on common datasets and diverse architectures, demonstrate that: i) AT-PR achieves higher improvements in PR than AT-WCR methods; ii) it shows more consistent effectiveness across varying local inputs; iii) it exhibits a reduced trade-off in model's generalization." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.715, + 0.222, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Deep Learning (DL) has shown potential for transformative impacts across many industries including transportation, healthcare and energy. The reality that DL is not as dependable as required now becomes a major impediment to their broader deployment. Since the first discovery of the phenomenon of Adversarial Examples (AEs) for DL [21, 39, 48], robustness is arguably the property in the limelight, encompassing research topics like adversarial attacks in malicious contexts [7, 8, 14, 21, 36, 64, 65, 75]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.906, + 0.442 + ], + "angle": 0, + "content": "AE detection during testing [19, 23, 24, 35, 42], verification with formal guarantees [20, 26, 30, 37, 44], and various robustness improvement techniques, including distillation [40], randomized smoothing [12], gradient masking [2], and the more widely used adversarial training (AT) [4, 21, 22, 36, 52]. As large language models and multimodal foundation models gain popularity, more recent research has increasingly focused on understanding their robustness as well [11, 18, 28, 32, 33, 58, 71, 72, 74]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Regardless of specific DL tasks or model architectures, robustness broadly refers to a model's ability to maintain consistent decisions under small input perturbations, with local robustness focusing on stability around individual inputs, as opposed to global robustness across the entire input distribution. More specifically, when evaluating local robustness, common problem formulations include [16, 72]: i) a binary question, \"does an AE exist within a specified perturbation norm-ball?\" [20, 30, 44, 63]; ii) a maximum safe distance question, \"what is the largest perturbation radius within which no AE exists?\" [1, 38, 60, 61]; and iii) a maximum prediction loss question, \"which AE within the norm-ball yields the highest prediction loss?\", which is often applied in AT to enhance robustness [36, 56]. It is evident that the three formulations focus on extreme scenarios, where local robustness is evaluated based on the existence of a deterministic AE, reflecting a worst-case perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.853 + ], + "angle": 0, + "content": "A more recent and distinct view, however, considers probabilistic robustness (PR), employing statistical approaches to its evaluation [5, 13, 41, 49, 50, 55, 59, 60, 69, 72], it answers the question of \"what is the likelihood of AEs in the given perturbation norm-ball?\" [73]. However, existing AT methods are primarily designed for WCR via the well-established \"min-max\" optimization [3, 21, 36] (cf. Sec. 2.2), and are thus referred to as \"AT-WCR\" in this paper. No dedicated AT methods exist for enhancing PR, which motivates our work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we first formulate AT for PR as a new min-max optimization problem, and then devise AT-PR training scheme based on numerical algorithms to effectively" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.463, + 0.901 + ], + "angle": 0, + "content": "*Corresponding to X. Zhao, xingyu.zhao@warwick.ac.uk." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "solve this training optimization. Our experiments on 70 DL models trained on common datasets and models demonstrate key strengths of AT-PR over AT-WCR: \\(i\\) while traditional AT-WCR can also improve PR, our new AT-PR training scheme is more effective, especially when enhancing PR becomes challenging for traditional AT-WCR; \\(ii\\) AT-PR not only outperforms AT-WCR in improving PR on average, but also offers more stable and consistent performance across different local inputs; \\(iii\\) AT-PR has less negative impact on model's generalization compared to AT-WCR." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.244, + 0.457, + 0.259 + ], + "angle": 0, + "content": "In summary, key contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.26, + 0.483, + 0.29 + ], + "angle": 0, + "content": "- Problem formulation: We lay the foundation of AT for PR by formulating a new min-max optimization problem." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.29, + 0.484, + 0.32 + ], + "angle": 0, + "content": "- Numerical algorithms: We develop effective numerical algorithms to solve the AT-PR min-max optimization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.32, + 0.484, + 0.395 + ], + "angle": 0, + "content": "- Empirical evaluation: We validate the effectiveness of AT-PR through extensive experiments, showing its superior improvement in PR, more consistent effectiveness across different local inputs, and reduced impact on generalization compared to AT-WCR." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.396, + 0.484, + 0.442 + ], + "angle": 0, + "content": "- Open-source repository: A repository is provided at https://github.com/wellzline/AT-PR with code, datasets, models, and all experiment details." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.26, + 0.484, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.395, + 0.473 + ], + "angle": 0, + "content": "2. Preliminaries and Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.431, + 0.499 + ], + "angle": 0, + "content": "2.1. Worst-Case vs Probabilistic Robustness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Although definitions of robustness vary across different DL tasks and model types, it generally refers to a DL model's ability to maintain consistent predictions despite small input perturbations. Typically it is defined as all inputs in a region \\(\\eta\\) have the same prediction, where \\(\\eta\\) is a small norm ball (in a \\(L_{p}\\)-norm distance) of radius \\(\\gamma\\) around an input \\(\\mathbf{x}\\). A perturbed input (e.g., by adding noise on \\(\\mathbf{x}\\)) \\(\\mathbf{x}'\\) within \\(\\eta\\) is an AE if its prediction label differs from \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Evaluation methods for robustness are typically based on metrics defined around AEs [27, 75]. As discussed earlier in Sec. 1, four types of robustness metrics, as summarized in [16, 72], address: (1) a binary question, (2) a maximum safe distance question, (3) a maximum prediction loss question (Fig. 1 (a)), and (4) a statistical inference question estimating the likelihood of AEs within a given perturbation norm-ball (Fig. 1 (b)). The first three metrics focus on extreme scenarios, where local robustness is assessed based on the existence of a deterministic AE, reflecting a worst-case perspective referred to as WCR. In contrast, the fourth approach adopts a probabilistic view, evaluating overall local robustness in the presence of AEs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "PR is arguably more practical than WCR, as extreme-case focus is often unnecessary [73]. Instead, knowing the likelihood of AEs is more relevant for applications, where it is sufficient to show that the risk of AE remains below an acceptable threshold rather than exactly zero [59, 69, 72]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.153 + ], + "angle": 0, + "content": "Inspired by [72], we illustrate the difference between the maximum loss WCR and PR in Fig. 1, omitting the binary and maximum safe distance cases as they are less relevant to the AT focus of this paper." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.164, + 0.71, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.291, + 0.569, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.291, + 0.598, + 0.3 + ], + "angle": 0, + "content": "Class A" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.291, + 0.662, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.67, + 0.291, + 0.688, + 0.3 + ], + "angle": 0, + "content": "#" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.164, + 0.902, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.291, + 0.787, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.291, + 0.856, + 0.299 + ], + "angle": 0, + "content": "Non-AE" + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.315, + 0.88, + 0.33 + ], + "angle": 0, + "content": "Figure 1. Worst-case (a) vs. Probabilistic (b) Robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.407 + ], + "angle": 0, + "content": "To evaluate the WCR, we normally formulate it as a question of maximizing the prediction loss, cf. Fig. 1 (a), where adversarial attacks are typically introduced to find such a worst-case AE with perturbation:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.418, + 0.907, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} = \\arg \\max _ {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.59 + ], + "angle": 0, + "content": "where \\(\\delta\\) is a small perturbation constrained within an \\(L_{p}\\)-norm ball of radius \\(\\gamma\\). This optimization causes the model to misclassify \\(\\boldsymbol{x}^{\\prime} = \\boldsymbol{x} + \\delta\\) as a different label \\(y^\\prime \\neq y\\). There are some popular attacks following this formulation, including FGSM [21], PGD [36], and ZOO [10], among others. Following common practice, we measure WCR by the success rate of some selected attacks on a test dataset. Estimation of PR is normally formalized as the statistical inference problem on the following metric:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.646 + ], + "angle": 0, + "content": "Definition 1 (Probabilistic Robustness) For a DL classifier \\( f_{\\theta} \\) that takes input \\( x \\) and returns a prediction label, the PR of an input \\( x \\) in a norm ball of radius \\( \\gamma \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.656, + 0.907, + 0.685 + ], + "angle": 0, + "content": "\\[\nP R (\\boldsymbol {x}, \\gamma) = \\underset {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} {\\mathbb {E}} _ {\\delta \\sim P r (\\cdot | \\boldsymbol {x})} [ I _ {\\{f _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) = y \\}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.756 + ], + "angle": 0, + "content": "where \\( I_{S}(\\pmb{x}) \\) is an indicator function—it is equal to 1 when \\( S \\) is true and 0 otherwise; \\( Pr(\\cdot) \\) is the local distribution of inputs representing how perturbations \\( \\delta \\) are generated, which is precisely the \"input model\" used by [59, 72]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Def. 1 suggests PR is the probability that the model prediction remains unchanged from a random perturbation \\( x' \\). A \"frequentist\" interpretation of this probability is: it is the limiting relative frequency of perturbations where the output label is preserved, in an infinite sequence of independently generated perturbations [72]. That is, the \"proportion\" of non-AEs in such infinite set of perturbations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The work [59] is the first to define PR and introduces a black-box statistical estimator, particularly useful when" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1676" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.243 + ], + "angle": 0, + "content": "PR is extremely high (i.e., detecting AEs is a rare event). A more efficient white-box statistical estimator for PR was later proposed in [50]. Huang et al. [25] first extend PR to explainable AI (XAI). Zhang et al. [69] explore PR under functional perturbations, while the recent study [72] is the first to estimate PR in the text-to-image DL task. The recent work [73] provides an overview of PR, including its definition, estimation, enhancement and system-level integration. Without loss of generality, we adopt the estimator in [59] to estimate PR in later experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.292, + 0.267 + ], + "angle": 0, + "content": "2.2. Adversarial Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.44 + ], + "angle": 0, + "content": "AT [4, 21, 22, 36, 52] is a widely used empirical approach for improving robustness, typically formulated as a \"min-max\" optimization problem. It introduces a set of input perturbations \\(\\delta\\) within a norm-ball \\(\\eta\\) of radius \\(\\gamma\\), designed to maximize the training loss \\(\\mathcal{L}\\). The model is then trained with these perturbations by minimizing the overall loss of the training dataset. These perturbations are often approximated using empirical adversarial attacks, such as FGSM [21] and PGD [36] due to their high efficiency. Taking a classification task as an example, the min-max optimization target of AT can be formalised as:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Definition 2 (AT for WCR) For a DL model \\( f_{\\theta} \\), trained on a dataset \\( \\mathcal{D} \\), consisting of pairs \\( (\\pmb{x}, y) \\) where \\( \\pmb{x} \\in \\mathcal{X} \\) is the input and \\( y \\in \\mathcal{Y} \\) denotes the label, with loss function \\( \\mathcal{L}(\\pmb{x}, y; \\theta) \\), the objective of AT-WCR can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.518, + 0.483, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.56, + 0.483, + 0.607 + ], + "angle": 0, + "content": "where \\(\\delta\\) represents the perturbation that maximizes classification loss \\(\\mathcal{L}\\), while the outer minimization aims to train \\(f_{\\theta}\\) to be robust against such adversarial perturbations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The inner maximization is precisely the adversarial attack formulation outlined in Eq. (1). As highlighted in [3], the effectiveness of AT largely depends on identifying strong AEs within this inner maximization. Notably, the PGD attack [36] has become a critical benchmark and is widely regarded as the standard method for AT in practice [3]. Accordingly, we also use PGD in later experiments. Based on the standard min-max optimization, various AT variants have been proposed to achieve more stable and efficient training, such as [6, 15, 29, 47, 52, 57, 58, 66, 67]. Comprehensive overviews of recent AT variants can be found in surveys [3, 27], all of which fundamentally rely on the min-max problem for WCR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In contrast, to the best of our knowledge, only three works [43, 55, 70] have explicitly aimed at improving PR [73]. Although Wang et al. [55] demonstrate that simply adding randomly perturbed inputs significantly enhances PR, the resulting AR is nearly zero. The methods of [43] and [70] propose loss functions based on common risk" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "measures (i.e., \\(\\rho\\)-essential supremum, Conditional Value-at-Risk, and Entropic Value-at-Risk) to train models to minimize these risks. However, again, these approaches provide little to no improvement in AR. We argue that, akin to the well-established AT framework for AR, a dedicated AT framework for PR is necessary—one that does not compromise AR performance. Accordingly, our proposed AT-PR is the first to explicitly reformulate the min-max optimization principle for PR, while also providing comparable AR performance through practical training algorithms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.255, + 0.729, + 0.272 + ], + "angle": 0, + "content": "3. The Approach: AT-PR" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.279, + 0.719, + 0.294 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Intuitively, the goal of AT-WCR in Def. 2 is to identify the AE that maximizes model loss and incorporate it into AT to lower the maximum loss, cf. the top-middle illustration in Fig. 2. However, to effectively improve local PR, we aim to locate the \"widest peak\"—the region with the highest proportion of AEs in the loss landscape (cf. the top-right illustration in Fig. 2), rather than the tallest peak. Adding the AE from this peak during AT may ideally reduce the loss for AEs around that area, thereby lowering the overall AE proportion (cf. the intuitive interpretation of PR in Def. 1). In some cases, AT-WCR and AT-PR may identify the same AE for training when the tallest peak also happens to be the widest, as shown in the second row of Fig. 2. Notably, the two cases illustrated in Fig. 2 representatively cover all possible scenarios, though additional peaks may exist in reality. Given the difference between the goals of improving WCR and PR, we formulate a new min-max problem for AT-PR:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Definition 3 (AT for PR) For a DL model \\( f_{\\theta} \\), trained on a dataset \\( \\mathcal{D} \\) consisting of pairs \\( (x, y) \\), the objective of AT-PR can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.615, + 0.906, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma , P R (\\boldsymbol {x} + \\delta , k) = 0} k \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.76 + ], + "angle": 0, + "content": "where \\( k \\) and \\( \\delta \\) are the variables in the new inner max optimization, \\( PR(\\cdot, \\cdot) \\) is the metric defined in Eq. (2). As shown in Fig. 3, the intuition behind the maximization is to find an optimal \\( \\delta^* \\) that produces an \\( AE \\) \\( x' = x + \\delta^* \\) and to maximize the radius \\( k \\) of a smaller norm-ball centered on \\( x' \\) such that the \\( PR \\) within this \\( k \\)-sized norm-ball is 0 (i.e., all inputs within this smaller norm-ball are AEs)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.843 + ], + "angle": 0, + "content": "The high non-convexity of DL models makes it intractable to analytically solve both the established min-max optimization in Eq. (3) and our new formulation in Eq. (4). Thus, similar to existing AT-WCR, we propose algorithms to numerically solve Eq. (4) in AT-PR in the next section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.738, + 0.866 + ], + "angle": 0, + "content": "3.2. Optimization Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "While the outer minimization in Eq. (4) can be effectively handled with standard DL training methods as in AT-WCR," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.357, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.09, + 0.631, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.089, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.217, + 0.359, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.216, + 0.655, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.23, + 0.9, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.352, + 0.807, + 0.368 + ], + "angle": 0, + "content": "Figure 2. Two cases of the local loss landscape and their AEs identified by AT-WCR and AT-PR for AT." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.395, + 0.476, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.579 + ], + "angle": 0, + "content": "Figure 3. Illustration of the new inner-max optimization in Def. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.655 + ], + "angle": 0, + "content": "we devise the Algo. 1 to solve the inner optimization in our AT-PR. The gist of Algo. 1 is shown in Fig. 4, which essentially consists of 3 main steps:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.702 + ], + "angle": 0, + "content": "- Step 1 (lines 1-9): Apply PGD attacks in different regions of the local norm-ball under study to find a diverse set of AEs at various local optima \\( x_{i}^{\\prime} \\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.702, + 0.483, + 0.747 + ], + "angle": 0, + "content": "- Step 2 (lines 10–20): For each locally optimal AE found in Step 1, perform a gradient descent search to locate the nearest decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.748, + 0.483, + 0.793 + ], + "angle": 0, + "content": "- Step 3 (lines 21-25): Calculate the distance \\( d_{i} \\) from each \\( x_{i}^{\\prime} \\) to its respective decision boundary and return the AE with the largest \\( d_{i} \\) value." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Similar to AT-WCR, the returned AE will be added into the training dataset in AT-PR. We note, as an numerical algorithm to the optimization, its effectiveness may depends on certain assumptions include: \\(i\\)) the PGD attack in Step 1 may effectively find local optima by adjusting hyperparameters such as attack steps and learning rate, which aligns with the insight of \"PGD will converge to local maxima\"" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.907, + 0.485 + ], + "angle": 0, + "content": "[36]; ii) The gradient descent based boundary search with adaptive hyperparameters may efficiently find the decision boundary as used in some boundary attacks [9, 31, 68]. We also provide an initial theoretical analysis of the effectiveness and convergence properties of the proposed optimization algorithm in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.497, + 0.761, + 0.542 + ], + "angle": 0, + "content": "Step 1: Apply PGD in different regions of the norm-ball to get a set of diverse AEs at different local optimal points \\( x_{i}^{\\prime} \\)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.729, + 0.864, + 0.744 + ], + "angle": 0, + "content": "Figure 4. Illustration of the 3 main steps of Algo. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.907, + 0.88 + ], + "angle": 0, + "content": "We analyze the time complexity of the proposed algorithm. There are several computational steps to be considered with the assumption that the dimension of the input \\( \\pmb{x} \\) is \\( n \\). First, sampling \\( \\alpha \\) and steps from uniform distributions (lines 5-6) is independent of the input and thus the time complexity is \\( O(1) \\). Second, the PGD algorithm (line 7) consists of steps gradient calculations with complexity \\( O(\\text{grad})^1 \\) and one projection calculation with complexity" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.907, + 0.902 + ], + "angle": 0, + "content": "1The time complexity of gradient calculation depends on the model" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1678" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.09, + 0.486, + 0.601 + ], + "angle": 0, + "content": "Algorithm 1 PGD and gradient-based search for \\(x_{pr}^{\\prime}\\) \nRequire: Inputs \\([X,Y],N,\\alpha_{min},\\alpha_{max},step_{min},step_{max}\\) 1: Initialize \\(AE_{s}\\gets []\\) Initialize AE candidate set 2: Apply PGD to get different AE candidates 3: for idx = 0 to N do 4: \\(x_{init}\\gets x + \\mathrm{Uniform}(-\\gamma ,\\gamma)\\) 5: \\(\\alpha \\leftarrow\\) Uniform \\((\\alpha_{min},\\alpha_{max})\\) 6: steps \\(\\leftarrow\\) Uniform \\((step_{min},step_{max})\\) 7: \\(x_{idx}^{\\prime}\\gets pgd(x_{init},y,\\gamma ,\\alpha ,steps)\\) 8: Append \\(x_{idx}^{\\prime}\\) to \\(AE_{s}\\) 9: end for \nRequire: \\(AE_s,x,y,C\\) \n10: Initialize Max_d \\(\\leftarrow 0\\) . \\(x_{pr}^{\\prime}\\gets\\) None 11: for each \\(x^{\\prime}\\in AE_{s}\\) do 12: \\(\\tilde{x} = x^{\\prime}\\) 13: while iter \\(< C\\) do 14: \\(y^\\prime = f(\\tilde{x})\\) 15: if \\(y^\\prime = y\\) then 16: break \\(\\triangleright\\) Exit if \\(\\tilde{x}\\) is classified correctly 17: end if 18: \\(g\\gets \\nabla_{\\tilde{x}}L(\\tilde{x},y)\\) \\(\\triangleright\\) Compute gradient 19: \\(\\tilde{x} = \\tilde{x} -\\alpha \\cdot g\\) \\(\\triangleright\\) Update \\(\\tilde{x}\\) \n20: end while \n21: \\(d\\gets D(\\tilde{x},x^{\\prime})\\) \\(\\triangleright\\) Distance between \\(\\tilde{x}\\& x^{\\prime}\\) \n22: if \\(d > Max\\_ d\\) then \n23: \\(x_{pr}^{\\prime} = x^{\\prime}\\) \n24: Max_d=d \n25: end if \n26: end for \n27: Output \\(x_{pr}^{\\prime}\\) Farthest AE from decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.856 + ], + "angle": 0, + "content": "\\(O(n)\\). Third, in the effort to locate the decision boundary (lines 18-19), one gradient calculation step is performed. Last, a \\(L_{\\infty}\\) distance between \\(\\tilde{\\pmb{x}}\\) and \\(\\pmb{x}^{\\prime}\\) is estimated (line 21), with complexity \\(O(n)\\). In Step 1, the algorithm uses different random starts via random initialization [46] to ensure it finds different local optima within norm-ball \\(\\eta\\) (line 4), and generates one AE (lines 5-8) for each combination of input \\((x,y)\\), radius \\(\\gamma\\), randomly sampled \\(\\alpha\\) and steps. A total of \\(N\\) AEs are generated. Thus the complexity upper bound of Step 1 is then \\(O(N*(steps*O(grad)+O(n)))\\). In Step 2, with the maximum iteration number \\(C\\) for the nearest decision boundary locating process of \\(N\\) AEs, the complexity upper bound of Step 2 is then \\(O(C*N*O(grad))\\). In Step 3, the complexity of distance estimation is \\(O(n)\\) as discussed above. Assuming \\(O(grad) \\gg O(n)\\), which" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.903 + ], + "angle": 0, + "content": "architecture and the specific loss function employed during training. E.g., DL models with more layers or complex loss functions may require more computational resources for gradient computation due to backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "typically holds in DL scenarios, the complexity simplifies to \\( O(N * (steps + C) \\cdot O(grad)) \\). As analyzed, the computational cost of AT-PR primarily comes from locating the optimal AE. However, it remains compatible with other techniques designed to enhance AT efficiency, such as [46, 62]. We provide comparative empirical results by replacing PGD with Fast-AT [62] to demonstrate such compatibility in Table 3, refer to Sec. 4 for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.646, + 0.243 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.358 + ], + "angle": 0, + "content": "We train a total of 70 DL models on four datasets: CIFAR-10, CIFAR-100, SVHN, and Tiny/ImageNet, using models such as ResNet, Wide-ResNet (WRN), Vision Transformer (ViT) [17, 34], and DeiT [51]. To evaluate AT-PR, we compare it with two standard AT-WCR (PGD and FGSM) and four advanced methods: TRADES [66], MART [57], ALP, and CLP [29]. Implementation details are in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.84, + 0.382 + ], + "angle": 0, + "content": "4.1. Effectiveness and Efficiency of AT-PR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.54 + ], + "angle": 0, + "content": "We assess AT-PR's effectiveness against AT-WCR from two perspectives: 1) improvement in PR and 2) impact on generalization. We compare models trained with AT-PR, AT-WCR (FGSM and PGD) and a baseline without AT, using three metrics: (i) Accuracy on clean test data; (ii) Adversarial Accuracy under PGD, C&W, and AA attacks to reflect WCR; and (iii) PR, computed as \\( PR(x, \\gamma) \\) for various \\( \\gamma \\) using the estimator from [59]. We report the mean and standard deviation of PR on the test set, along with the generalization error (GE) to measure train-test robustness gaps." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.54, + 0.797, + 0.553 + ], + "angle": 0, + "content": "Key findings summarized from Table 1:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.659 + ], + "angle": 0, + "content": "i) All original models without AT show low WCR. As expected, PGD is the most effective in improving WCR, though it compromises generalization the most. AT-PR also enhances WCR, outperforming FGSM and, in some cases, even surpassing PGD, as seen with SVHN/ResNet-18, Tiny-ImageNet/ResNet-34 for PGD.Ac., and SVHN/WRN50-2, Tiny-ImageNet/ResNet-18, ResNet-34 for CW.Ac." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.78 + ], + "angle": 0, + "content": "ii) Both AT-WCR with FGSM and PGD can improve PR compared to the original models without AT, but they are less effective when compared to the performance of AT-PR. Specifically, the mean PR evaluations of AT-PR-trained models are higher than those trained with AT-WCR across most datasets, model architectures and norm-ball sizes. For instance, on CIFAR-100 with ResNet-18, AT-PR improves \\( PR(\\cdot, 0.15) \\) from 44.71 (FGSM) and 80.47 (PGD) to 87.38." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.87 + ], + "angle": 0, + "content": "iii) Compared to AT-WCR, the improvement in PR becomes more pronounced as the norm-ball size \\(\\gamma\\) increases, which is anticipated since larger norm-balls generally lead to less robust models. This demonstrates that our AT-PR approach is particularly useful when enhancing PR becomes more challenging for traditional AT-WCR." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "iv) Each local input is associated with an estimated PR, and we report the mean and standard deviation over the test" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1679" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Evaluation results of DL models trained with different AT schemes, showing standard accuracy, adversarial accuracy (indicating WCR: PGD-10, C&W-10), and PR estimated (mean ± std) for four norm-ball sizes \\(\\gamma\\). All results are calculated over the test datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.903, + 0.502 + ], + "angle": 0, + "content": "
DatasetModelAT SchemeAccu. %PGD Ac. %CW Ac. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18W/O AT94.960.020.5178.38 ± 33.1466.3 ± 38.0653.41 ± 40.7637.01 ± 39.77
AT-WCR(FGSM)88.60.090.0993.99 ± 17.588.02 ± 25.0380.42 ± 31.8167.77 ± 38.6
AT-WCR(PGD)8145.0144.5697.28 ± 13.2995.77 ± 17.0794.1 ± 20.289.78 ± 26.53
AT-PR80.1439.7739.0398.22±10.3897.28±13.1696.05±16.1893.21±21.47
WRN-50-2W/O AT94.140.230.1476.96±32.7665.65±37.4154.43±39.8840.19±39.49
AT-WCR(FGSM)92.040.080.0495.68±14.5391.9±20.2586.67±25.9876.79±33.1
AT-WCR(PGD)82.6249.5148.1397.86±12.6197.08±14.5495.84±17.1492.62±22.87
AT-PR82.7143.2743.0798.09±10.8197.33±13.0396.22±15.5893.56±20.89
CIFAR-100ResNet-18W/O AT75.50.020.044.58 ± 42.6432.1 ± 40.123.32 ± 36.4115.62 ± 31.57
AT-WCR(FGSM)67.801.410.6683.90 ± 31.1172.89 ± 37.9461.06 ± 42.2344.71 ± 43.04
AT-WCR(PGD)52.8519.9319.5195.4 ± 18.1592.7 ± 22.7288.98 ± 28.1980.47 ± 36.4
AT-PR53.2119.3818.3997.59±12.9695.96±16.6293.15±21.7787.38±29.63
WRN-50-2W/O AT75.460.490.4551.11±40.5437.96±39.2328.33±36.7219.11±31.97
AT-WCR(FGSM)73.170.290.185.68±27.4576.62±33.8966.37/38.6652.04±41.42
AT-WCR(PGD)57.4623.1722.595.31±17.3892.0±23.7988.14±29.1580.67±36.21
AT-PR59.4622.0621.5996.05±16.1793.72±20.4789.48±26.8781.6±34.86
SVHNResNet-18W/O AT94.960.080.1297.11 ± 9.6695.63 ± 11.9893.74 ± 14.4890.12 ± 18.40
AT-WCR(FGSM)93.494.685.3298.79 ± 7.0698.42 ± 7.7097.86 ± 8.6796.57 ±10.78
AT-WCR(PGD)89.2144.5743.7098.96 ± 5.8598.50 ± 7.4097.72 ± 9.6195.41 ± 14.41
AT-PR90.2945.8142.2199.12 ± 5.1598.87 ± 5.9398.54 ± 6.8497.84 ± 8.60
WRN-50-2W/O AT95.911.741.2996.60 ± 10.3694.77 ± 13.0192.40 ± 15.8787.97 ± 20.04
AT-WCR(FGSM)94.318.138.5498.34 ± 7.4997.71 ± 8.3996.80 ± 9.6995.12 ± 12.06
AT-WCR(PGD)88.9154.9950.5898.72 ± 6.4498.28 ± 7.6697.76 ± 8.9596.77 ± 11.18
AT-PR91.9752.2251.5898.88 ± 6.1298.38 ± 7.5397.72 ± 9.1996.26 ± 12.35
Tiny-ImageNetResNet-18W/O AT59.260.00.080.78 ± 33.172.95 ± 37.6964.17 ± 40.9151.14 ± 42.91
AT-WCR(FGSM)54.810.050.0189.9 ± 24.9785.33 ± 29.9580.18 ± 33.7469.03 ± 39.39
AT-WCR(PGD)47.4321.319.3196.32 ± 17.0193.79 ± 22.190.16 ± 27.7682.46 ± 35.97
AT-PR51.6121.119.8998.23 ± 11.0696.1 ± 17.8493.12 ± 23.3286.65 ± 32.26
ResNet-34W/O AT61.660.00.082.39 ± 31.375.2 ± 36.066.95 ± 39.4253.7 ± 42.17
AT-WCR(FGSM)570.130.0290.67 ± 23.9285.81 ± 29.580.57 ± 34.1370.22 ± 39.36
AT-WCR(PGD)49.3320.9719.5895.45 ± 18.5593.08 ± 23.4889.85 ± 28.383.53 ± 34.99
AT-PR53.9821.7220.9596.71 ± 15.7495.02 ± 19.8992.5 ± 23.8786.23 ± 32.29
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.604 + ], + "angle": 0, + "content": "set in Table 1. Notably, AT-PR consistently achieves higher mean PR compared to models trained with AT-WCR, while also exhibiting lower standard deviations. This indicates that AT-PR not only improves PR but also offers more stable and consistent performance across different local inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.729 + ], + "angle": 0, + "content": "\\(\\nu\\) ) While PGD significantly improves robustness, it often reduces accuracy [66]. In contrast, AT-PR has a smaller negative impact on accuracy compared to PGD, with almost all models achieving higher accuracy than those trained using PGD while also enhancing both WCR and PR. This effect is more pronounced on larger datasets—for example, on Tiny-ImageNet with ResNet-18, AT-PR achieves an accuracy of \\(53.98\\%\\) compared to \\(49.33\\%\\) for PGD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Table 2 shows that combining AT-PR with TRADES improves clean accuracy and PR on more complex attention-based models ViT (Vanilla ViT and DeiT) across different scales (DeiT-Tiny, ViT-Small, ViT-Base) for CIFAR10. With just five epochs of AT-PR fine-tuning, all three ViT models achieve higher clean accuracy and PR while maintaining comparable WCR, showing a slight decrease on PGD-20 and AA but improved performance on CW-20." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Table 3 presents a comparison of AT-PR with four advanced AT methods (TRADES, MART, ALP and CLP) across three datasets using ResNet-18 and WRN-50-2. Re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.907, + 0.665 + ], + "angle": 0, + "content": "sults show that AT-PR consistently achieves the best PR performance across four different norm-ball sizes \\(\\gamma\\). Additionally, to verify that AT-PR is compatible with efficient AT methods [46, 62], we replace PGD with Fast-AT [62] in our framework (step 1 in Algo. 1). The result shows that Fast-AT-PR achieves \\(\\sim 40\\%\\) speedup over the original AT-PR, with only \\(\\sim 50\\%\\) extra cost compared to other AT baselines. Notably, Fast-AT-PR retains comparable PR performance to AT-PR and continues to outperform other AT baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.791, + 0.692 + ], + "angle": 0, + "content": "4.2. Generalization Error of AT-PR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.837 + ], + "angle": 0, + "content": "We investigate the generalization error (GE) (train-test gap) of AT-PR to gain insights into its effectiveness. Table 4 compares the PR GE between AT-PR and PGD. Models trained with AT-PR consistently yields lower GE of up \\(2.2\\%\\) on CIFAR-10, \\(2.87\\%\\) on CIFAR-100, and \\(9.66\\%\\) on TinyImageNet, significantly lower than PGD's \\(4.67\\%\\), \\(8.54\\%\\), and \\(11.94\\%\\), respectively. We further analyze the underlying reasons based on the PAC-Bayes analysis for Adversarial Robustness [54], see Appendix C for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.805, + 0.864 + ], + "angle": 0, + "content": "4.3. AEs Used in AT-WCR vs. AT-PR" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We analyze the AEs used during training to understand how AT-PR differs from AT-WCR. By comparing the AEs gen" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1680" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. ViT models trained on CIFAR-10 with TRADES and AT-PR, showing standard accuracy, adversarial accuracy (indicating WCR: PGD, C&W, AA), and PR estimated (mean ± std) for four norm-ball sizes \\(\\gamma\\)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.905, + 0.215 + ], + "angle": 0, + "content": "
DatasetModelAT SchemeAccu. %PGD-20. %CW-20. %AA. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10DeiT-TiTRADES79.4449.7846.4545.9696.61 ± 15.2394.73 ± 19.3692.69 ± 23.0489.18 ± 27.9
+AT-PR81.0749.0846.9245.9597.26 ± 14.095.94 ± 17.1294.16 ± 20.590.65 ± 25.88
Vit-STRADES81.0350.9648.1447.4096.93 ± 14.9995.48 ± 18.493.58 ± 21.6290.06 ± 26.81
+AT-PR83.2950.4448.2047.3497.76 ± 11.8996.03 ± 16.4693.98 ± 20.6990.48 ± 26.19
Vit-BTRADES83.5553.1450.9449.9697.34 ± 13.5795.87 ± 16.9993.78 ± 21.1790.77 ± 25.88
+AT-PR86.0352.9451.2449.3497.9 ± 11.896.49 ± 15.294.44 ± 19.6291.02 ± 25.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.228, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Table 3. Comparison of PR(, γ) and training time (sec/epoch) for more recent AT methods (TRADES, MART, ALP, CLP) and Fast-AT-PR." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.254, + 0.885, + 0.613 + ], + "angle": 0, + "content": "
DataModelMetricAT Scheme
PGDTRADESMARTALPCLPFast-AT-PRAT-PR
CIFAR-10ResNet-18PR(,0.08)%97.28 ± 13.2997.74 ± 12.4297.4 ± 13.796.73 ± 15.997.61 ± 13.0698.12 ± 11.5298.22 ± 10.38
PR(,0.10)%95.77 ± 17.0796.33 ± 16.3995.92 ± 17.4395.16 ± 19.5396.29 ± 16.6697.1 ± 14.3497.28 ± 13.16
PR(,0.12)%94.10 ± 20.2094.55 ± 20.3193.93 ± 21.3893.24 ± 23.0894.54 ± 20.2895.4 ± 17.9096.05 ± 16.18
PR(,0.15)%89.78 ± 26.5391.22 ± 25.7090.35 ± 26.9189.76 ± 28.1490.86 ± 26.2392.71 ± 23.5493.21 ± 21.47
Time (s/ep.)546144646497184
WRN-50-2PR(,0.08)%97.86 ± 12.6197.55 ± 13.1497.40 ± 13.6197.75 ± 12.4397.40 ± 13.9197.96 ± 10.8298.09 ± 10.81
PR(,0.10)%97.08 ± 14.5496.14 ± 16.7096.15 ± 16.7896.42 ± 16.1196.20 ± 17.5897.22 ± 12.9697.33 ± 13.03
PR(,0.12)%95.84 ± 17.1494.33 ± 20.4194.41 ± 20.3794.64 ± 20.0194.95 ± 19.8896.22 ± 15.6296.22 ± 15.58
PR(,0.15)%92.62 ± 22.8791.01 ± 25.7991.11 ± 25.6591.22 ± 25.7091.48 ± 25.6994.74 ± 19.8893.56 ± 20.89
Time (s/ep.)13110571155157229305
CIFAR-100ResNet-18PR(,0.08)%95.40 ± 18.1595.02 ± 19.2795.45 ± 18.1196.20 ± 16.0796.16 ± 16.3697.09 ± 13.4897.59 ± 12.96
PR(,0.10)%92.70 ± 22.7292.09 ± 24.2292.73 ± 23.3994.40 ± 20.0893.81 ± 21.4294.67 ± 18.9195.96 ± 16.62
PR(,0.12)%88.98 ± 28.1987.80 ± 29.8089.35 ± 28.3292.05 ± 24.1291.05 ± 25.8692.63 ± 23.9193.15 ± 21.77
PR(,0.15)%80.47 ± 36.4079.20 ± 37.7381.93 ± 35.7285.89 ± 31.6584.97 ± 33.0085.87 ± 31.9987.38 ± 29.63
Time (s/ep.)556143646597180
WRN-50-2PR(,0.08)%95.31 ± 17.3895.24 ± 18.6195.30 ± 16.8795.34 ± 17.7995.29 ± 17.7096.52 ± 15.1196.05 ± 16.17
PR(,0.10)%92.00 ± 23.7992.30 ± 23.6192.59 ± 22.8192.36 ± 23.1991.79 ± 23.8793.49 ± 20.9393.72 ± 20.47
PR(,0.12)%88.14 ± 29.1588.65 ± 28.4688.38 ± 28.6588.51 ± 28.2287.54 ± 29.9288.82 ± 28.3889.48 ± 26.87
PR(,0.15)%80.67 ± 36.2180.35 ± 36.5480.79 ± 36.0180.37 ± 36.8679.71 ± 36.7681.02 ± 37.0981.60 ± 34.86
Time (s/ep.)13110270155156225312
SVHNResNet-18PR(,0.08)%98.96 ± 5.8599.03 ± 5.7698.99 ± 6.1599.05 ± 5.7399.12 ± 5.3398.90 ± 5.9799.12 ± 5.15
PR(,0.10)%98.50 ± 7.4098.46 ± 8.2798.31 ± 8.2098.63 ± 7.1598.60 ± 7.2198.53 ± 7.2498.87 ± 5.93
PR(,0.12)%97.72 ± 9.6197.30 ± 11.0897.10 ± 11.2697.85 ± 9.1397.66 ± 9.7898.11 ± 8.1998.54 ± 6.84
PR(,0.15)%95.41 ± 14.4194.15 ± 16.5794.12 ± 16.6595.54 ± 14.1094.93 ± 15.0097.10 ± 10.0597.84 ± 8.60
Time (s/ep.)7886619193141202
WRN-50-2PR(,0.08)%98.72 ± 6.4498.69 ± 6.9098.61 ± 7.6698.54 ± 7.5298.59 ± 7.4698.43 ± 6.6998.88 ± 6.12
PR(,0.10)%98.28 ± 7.6698.11 ± 8.6698.07 ± 9.2798.12 ± 8.7498.00 ± 8.8998.15 ± 7.2998.38 ± 7.53
PR(,0.12)%97.76 ± 8.9597.26 ± 10.8997.34 ± 10.9097.65 ± 8.8596.98 ± 11.1397.79 ± 8.1297.72 ± 9.19
PR(,0.15)%96.77 ± 11.1895.20 ± 14.8495.45 ± 14.7295.67 ± 14.3294.74 ± 15.3796.94 ± 9.8096.26 ± 12.35
Time (s/ep.)190146102223225321420
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.483, + 0.775 + ], + "angle": 0, + "content": "erated by PGD and AT-PR, we find that about \\(10\\%\\) of training inputs produce different AEs (e.g., \\(10.05\\% / 8.39\\%\\) for ResNet-18 and \\(8.83\\% / 10.42\\%\\) for WRN50-2 on CIFAR-10/CIFAR-100). This means that for roughly \\(90\\%\\) of inputs, the loss landscapes align with Case 2 (\"tall & wide vs. short & tight\") in Fig. 2, where both methods select the same AEs. However, for the remaining \\(10\\%\\) -ish, where the landscape fits Case 1 (\"tall & tight vs. short & wide\"), AT-PR identifies different AEs that better optimize PR." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fig. 5 (top row) compares the loss of AEs generated by PGD and AT-PR. For all models, PGD AEs consistently have higher losses, reflecting its objective to find AE with the maximum loss for AT. This aligns with the goal of AT-WCR, which seeks to minimize the loss on such worst-case AEs (see Fig.2, top-middle). In contrast, AT-PR generates AEs with lower loss. Combined with the fact that our Algorithm 1 searches for examples farthest from the deci" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.699 + ], + "angle": 0, + "content": "sion boundary, suggests that AT-PR effectively locates AEs within the widest peak of the loss landscape. Training with these AEs helps reduce the \"short & wide\" peak (see Fig. 2, top-right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 (b) visualizes the distributions of all generated AEs \\(\\{x_i^{\\prime}\\}_{i = 1}^{n}\\) by analyzing the distance \\((L_{2}\\) norm) between each \\(x_{i}^{\\prime}\\) and the original input \\(x\\) (e.g., \\(d_0,d_1\\) in Fig. 5 (a)), along with their angular relationships. The 2D coordinate in Fig. 5 (a) introduces a fixed anchor \\(x_{a}\\) near \\(x\\), for an AE \\(x_0^\\prime\\), the angle \\(\\theta_0\\) is computed between \\(\\vec{v}_{x\\rightarrow x_0'}\\) and \\(\\vec{v}_{x\\rightarrow x_a}\\), representing \\(x_0'\\)'s relative position, given by: \\(\\theta_0 = \\arccos \\left(\\frac{\\vec{v}_{x\\rightarrow x_0'}\\cdot\\vec{v}_{x\\rightarrow x_a}}{\\|\\vec{v}_{x\\rightarrow x_0'}\\||\\vec{v}_{x\\rightarrow x_a}\\|}\\right)\\). The resulting angles range from approximately \\(88^{\\circ}\\) to \\(92^{\\circ}\\) Fig. 5 (b), reflecting the unique geometric properties of high-dimensional spaces, where random vectors highly likely to be orthogonal [53]. To further capture the variability in relative positions," + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "1681" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 4. Comparative GE experiment on DL models across different datasets. Train/test set evaluations of PR estimates (mean) for four norm-ball sizes \\((\\gamma)\\) across models trained with PGD and AT-PR, presented as \\(PR_{train} / PR_{test} / (PR_{train} - PR_{test})\\)" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.128, + 0.905, + 0.317 + ], + "angle": 0, + "content": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
WRN-50-2PGD98.88/97.86/1.0198.19/97.08/1.197.49/95.84/1.6495.3/92.62/2.67
AT-PR98.63/98.09/0.5497.8/97.33/0.4796.71/96.22/0.4694.9/93.56/1.34
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
WRN-50-2PGD98.35/95.31/3.0496.71/92.0/4.7193.67/88.14/5.5386.69/80.67/6.02
AT-PR97.3/96.05/1.2595.41/93.72/1.6991.93/89.48/2.4583.25/81.6/1.65
Tiny-ImageNetResNet-18PGD98.72/96.32/2.497.86/93.79/4.0796.94/90.16/6.7894.4/82.46/11.94
AT-PR99.01/98.23/0.7898.02/96.1/1.9296.68/93.12/3.5694.31/86.65/7.66
ResNet-34PGD98.67/95.45/3.2298.02/93.08/6.9496.9/89.85/7.0593.36/83.53/9.83
AT-PR99.28/96.71/2.5798.82/95.02/3.898.11/92.5/5.6195.89/86.23/9.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.094, + 0.329, + 0.902, + 0.345 + ], + "angle": 0, + "content": "Table 5. PR (mean) on train/test sets, reported as: \\( {PR}_{\\text{train }}/{PR}_{\\text{test }}/\\left( {{PR}_{\\text{train }} - P{R}_{\\text{test }}}\\right) \\) ,for model trained with PGD, Random-AE, AT-PR." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.354, + 0.905, + 0.464 + ], + "angle": 0, + "content": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
Random98.62/97.71/0.9197.73/96.78/0.9596.58/95.48/1.194.19/92.64/1.54
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
Random96.6/95.87/0.7394.85/93.77/1.0892.18/90.92/1.2685.39/85.21/0.18
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.484, + 0.216, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.485, + 0.345, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.484, + 0.48, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.569, + 0.209, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.564, + 0.351, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.564, + 0.482, + 0.647 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.422, + 0.651, + 0.435, + 0.66 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Figure 5. (Top row) Averaged loss of AEs used during AT for AT-PR and PGD. (Bottom row) (a) Illustration of angles and distances of AEs relative to \\( x \\). (b) Distribution of generated AEs. (c) Pairwise angular distribution among AEs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "we computed the Pairwise Angular Distribution among all vectors \\(\\{\\vec{v}_{x\\rightarrow x_i'}\\}_{i = 1}^n\\). The angle \\(\\beta\\) for each pair of vectors is calculated as illustrated in Fig. 5 (a), with the resulting distribution shown in Fig. 5 (c). This analysis shows that the angles between all pairs of vectors range from \\(60^{\\circ}\\) to \\(85^{\\circ}\\), indicating that the generated AEs are widely spread within the norm-ball. The final AE selected by AT-PR, highlighted in orange in Fig. 5 (b), has the largest distance to nearest decision boundary (1.0 for the L2 norm), compared to other" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.487, + 0.757, + 0.502 + ], + "angle": 0, + "content": "AEs, which range from 0.76 to 0.92." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.669 + ], + "angle": 0, + "content": "To empirically validate that AEs generated by AT-PR are most effective for improving PR, we conduct an ablation study by randomly sampling AEs from those generated in Step 1 of Algo. 1 (also illustrated as random blue spots in Fig.5 (b)). Table 5 shows that while randomly selected AEs improve PR and GE compared to PGD, AT-PR consistently performs better. As Fig. 2 illustrates, although random AEs may fall within a short and wide peak (explaining their improved PR over PGD), they do not necessarily lie within the \"widest peak\" that AT-PR specifically targets. This distinction explains AT-PR's superior PR performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.854 + ], + "angle": 0, + "content": "In this work, for the first time, we introduce a novel AT method targeting PR, named as AT-PR. By reformulating a new min-max optimization, we lay the theoretical foundation for AT-PR and develop specialized numerical algorithms to solve this optimization. Extensive experimental evaluation demonstrates that AT-PR consistently outperforms traditional AT methods designed for WCR in improving PR. Additionally, we show that AT-PR incurs a smaller trade-off in model generalization compared to common AT-WCR methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In the future, we plan to address the challenges identified in [73], such as benchmarking PR and extending it to diverse generative AI tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Acknowledgements SK and XZ have received funding from the European Union's EU Framework Program for Research and Innovation Europe Horizon (grant agreement No 101202457). XZ's contribution is also supported by the UK EPSRC New Investigator Award [EP/Z536568/1] and NVIDIA Academic Grant Program. SK's contribution is supported by the UKRI Future Leaders Fellowship Grant [MR/S035176/1], and Innovate UK through the DriveSafeAI Project (10062594). XH has received funding from the European Union's Horizon Europe Research and Innovation Action under grant agreement No 101212818. YZ's contribution is supported by China Scholarship Council. We thank Zheng Wang for his valuable comments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.482, + 0.366 + ], + "angle": 0, + "content": "Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.383, + 0.188, + 0.399 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.408, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[1] Amir Aminifar. Universal adversarial perturbations in epileptic seizure detection. In 2020 Int. Joint Conference on Neural Networks (IJCNN), pages 1-6. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.452, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Proc. of the 35th International Conference on Machine Learning, pages 274-283. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Proc. of the 30th Int. Joint Conference on Artificial Intelligence, pages 4312-4321, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.482, + 0.638 + ], + "angle": 0, + "content": "[4] Mislav Balunović and Martin Vechev. Adversarial training and provable defenses: Bridging the gap. In 8th International Conference on Learning Representations (ICLR'20'20), 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.641, + 0.482, + 0.682 + ], + "angle": 0, + "content": "[5] Teodora Baluta, Zheng Leong Chua, Kuldeep S Meel, and Prateek Saxena. Scalable quantitative verification for deep neural networks. In ICSE'21, pages 312-323, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.685, + 0.482, + 0.739 + ], + "angle": 0, + "content": "[6] Qi-Zhi Cai, Chang Liu, and Dawn Song. Curriculum adversarial training. In Proc. of the 27th International Joint Conference on Artificial Intelligence, page 3740-3747. AAAI Press, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.743, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[7] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE symposium on security and privacy (sp), pages 39-57, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[8] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. A survey on adversarial attacks and defences. CAAI Transactions on Intelligence Technology, 6(1):25-45, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[9] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In IEEE symposium on security and privacy (SP), pages 1277-1294. IEEE, 2020. 4, 12" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.408, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[10] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. ZOO: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proc. of the 10th ACM Workshop on Artificial Intelligence and Security, page 15-26. ACM, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] Zhen Chen, Yi Zhang, Fu Wang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. TARP-VP: Towards evaluation of transferred adversarial robustness and privacy on label mapping visual prompting models. Advances in Neural Information Processing Systems, 37:6776-6796, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[12] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In International Conference on Machine Learning, pages 1310-1320. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[13] Nicolas Couellan. Probabilistic robustness estimates for feed-forward neural networks. Neural networks, 142:138-147, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[14] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[15] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[16] Yi Dong, Wei Huang, Vibhav Bharti, Victoria Cox, Alec Banks, Sen Wang, Xingyu Zhao, Sven Schewe, and Xiaowei Huang. Reliability assessment and safety arguments for machine learning components in system assurance. ACM Transactions on Embedded Computing Systems, 22(3):1-48, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[18] Chengbin Du, Yanxi Li, Zhongwei Qiu, and Chang Xu. Stable diffusion is unstable. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[19] Xiaoning Du, Xiaofei Xie, Yi Li, Lei Ma, Yang Liu, and Jianjun Zhao. Deepstellar: Model-based quantitative analysis of stateful deep learning systems. In Proc. of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 477–487, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[20] Timon Gehr, Matthew Mirman, Dana Drachsler-Cohen, Petar Tsankov, Swarat Chaudhuri, and Martin Vechev. Ai2: Safety and robustness certification of neural networks with abstract interpretation. In IEEE symposium on security and privacy (SP), pages 3-18, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[21] Ian J Goodfellow. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR'15), 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Sven Gowal, Krishnamurthy Dj Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1683" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "Arandjelovic, Timothy Mann, and Pushmeet Kohli. Scalable verified training for provably robust image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4842-4851, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[23] Wei Huang, Youcheng Sun, Xingyu Zhao, James Sharp, Wenjie Ruan, Jie Meng, and Xiaowei Huang. Coverage-Guided Testing for Recurrent Neural Networks. IEEE Transactions on Reliability, 71(3):1191-1206, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[24] Wei Huang, Xingyu Zhao, Alec Banks, Victoria Cox, and Xiaowei Huang. Hierarchical distribution-aware testing of deep learning. ACM Transactions on Software Engineering and Methodology, 33(2):1-35, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[25] Wei Huang, Xingyu Zhao, Gaojie Jin, and Xiaowei Huang. Safari: Versatile and efficient evaluations for robustness of interpretability. In IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 1988-1998, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[26] Xiaowei Huang, Marta Kwiatkowska, Sen Wang, and Min Wu. Safety verification of deep neural networks. In Computer Aided Verification: 29th International Conference, CAV 2017, Heidelberg, Germany, July 24-28, 2017, Proceedings, Part I 30, pages 3-29. Springer, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.446 + ], + "angle": 0, + "content": "[27] Xiaowei Huang, Daniel Kroening, et al. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 37:100270, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[28] Xiaowei Huang, Wenjie Ruan, et al. A survey of safety and trustworthiness of large language models through the lens of verification and validation. Artificial Intelligence Review, 57 (7):175, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.545 + ], + "angle": 0, + "content": "[29] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. Adversarial logit pairing. arXiv preprint arXiv:1803.06373, 2018. 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[30] Guy Katz, Clark Barrett, David L Dill, Kyle Julian, and Mykel J Kochenderfer. Reluplex: An efficient smt solver for verifying deep neural networks. In 29th Int. Conf. Computer Aided Verification, pages 97-117. Springer, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[31] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In Proc. of the IEEE/CVF conference on computer vision and pattern recognition, pages 1221-1230, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[32] Han Liu, Yuhao Wu, Shixuan Zhai, Bo Yuan, and Ning Zhang. Riatig: Reliable and imperceptible adversarial text-to-image generation with natural prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20585-20594, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[33] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Discovering failure modes of text-guided diffusion models via adversarial search. In The 12th International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[35] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deepauge: Multi-granularity testing criteria for deep" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "learning systems. In Proce. of the 33rd ACM/IEEE Int. Conference on Automated Software Engineering (ASE'18), pages 120-131, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 4, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[37] Mark Huasong Meng, Guangdong Bai, Sin Gee Teo, Zhe Hou, Yan Xiao, Yun Lin, and Jin Song Dong. Adversarial robustness of deep neural networks: A survey from a formal verification perspective. IEEE Transactions on Dependable and Secure Computing, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[39] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[40] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[41] Mikhail Pautov, Nurislam Tursynbek, Marina Munkhoeva, Nikita Muravev, Aleksandr Petiushko, and Ivan Oseledets. Cc-cert: A probabilistic approach to certify general robustness of neural networks. In AAAI'22, pages 7975-7983, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[42] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepexplore: Automated whitebox testing of deep learning systems. In Proceedings of the 26th Symposium on Operating Systems Principles, pages 1-18. ACM, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[43] Alexander Robey, Luiz Chamon, George J Pappas, and Hamed Hassani. Probabilistically robust learning: Balancing average and worst-case performance. In Int. Conference on Machine Learning, pages 18667-18686. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[44] Wenjie Ruan, Xiaowei Huang, and Marta Kwiatkowska. Reachability analysis of deep neural networks with provable guarantees. In Proc. of the 27th Int. Joint Conference on Artificial Intelligence (IJCAI'18), pages 2651-2659, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[45] Sebastian Ruder. An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747, 2016. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[46] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[47] Uri Shaham, Yutaro Yamada, and Sahand Negahban. Understanding adversarial training: Increasing local stability of supervised models through robust optimization. Neurocomputing, 307:195-204, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In In Proc. of 2nd Int. Conf. on Learning Representations, 2014. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1684" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[49] Karim Tit, Teddy Furon, and Mathias Rousset. Efficient statistical assessment of neural network corruption robustness. NeurIPS'21, 34:9253-9263, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[50] Karim TIT, Teddy Furon, and Mathias Rousset. Gradient-informed neural network statistical robustness estimation. In Proc. of The 26th Int. Conf. on Artificial Intelligence and Statistics, pages 323-334. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[51] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[52] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[53] Princeton University. Lecture 11: Rademacher complexity, 2014. Accessed: 2025-01-05. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[54] Paul Viallard, Eric Guillaume VIDOT, Amaury Habrard, and Emilie Morvant. A pac-bayes analysis of adversarial robustness. Advances in Neural Information Processing Systems, 34:14421-14433, 2021. 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[55] Benjie Wang, Stefan Webb, and Tom Rainforth. Statistically robust neural network classification. In Uncertainty in Artificial Intelligence, pages 1735-1745. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[56] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. In ICML'19, pages 6586-6595. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[57] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 3, 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[58] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[59] Stefan Webb, Tom Rainforth, Yee Whye Teh, and M Pawan Kumar. A statistical approach to assessing neural network robustness. In Int. Conf. on Learning Representations, 2019, 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[60] Lily Weng, Pin-Yu Chen, Lam Nguyen, Mark Squillante, Akhilan Boopathy, Ivan Oseledets, and Luca Daniel. Proven: Verifying robustness of neural networks with a probabilistic approach. In Int. Conf. on Machine Learning, pages 6727-6736. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[61] Tsui Wei Weng, Huan Zhang, Pin-Yu Chen, Jinfeng Yi, Dong Su, Yupeng Gao, Cho-Jui Hsieh, and Luca Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. In Int. Conf. on Learning Representations, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[62] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[63] Weiming Xiang, Hoang-Dung Tran, and Taylor T. Johnson. Output reachable set estimation and verification for multilayer neural networks. IEEE Tran. on Neural Networks and Learning Systems, 29(11):5777-5783, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[64] Xiangyu Yin, Sihao Wu, Jiaxu Liu, Meng Fang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. Representation-based robustness in goal-conditioned reinforcement learning. In Proc. of the AAAI Conference on Artificial Intelligence, pages 21761-21769, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[65] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE Tran. on neural networks and learning systems, 30(9):2805-2824, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[66] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[67] Jingfeng Zhang, Xilie Xu, Bo Han, Gang Niu, Lizhen Cui, Masashi Sugiyama, and Mohan Kankanhalli. Attacks which do not kill training make adversarial learning stronger. In International conference on machine learning, pages 11278-11287. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[68] Jiawei Zhang, Linyi Li, Huichen Li, Xiaolu Zhang, Shuang Yang, and Bo Li. Progressive-scale boundary blackbox attack via projective gradient estimation. In Int. Conference on Machine Learning, pages 12479-12490. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.486 + ], + "angle": 0, + "content": "[69] Tianle Zhang, Wenjie Ruan, and Jonathan E. Fieldsend. Proa: A probabilistic robustness assessment against functional perturbations. In Machine Learning and Knowledge Discovery in Databases (ECML/PKDD'22), page 154-170. Springer-Verlag, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[70] Tianle Zhang, Yanghao Zhang, Ronghui Mu, Jiaxu Liu, Jonathan Fieldsend, and Wenjie Ruan. Prass: probabilistic risk-averse robust learning with stochastic search. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence, pages 559–567, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[71] Yi Zhang, Zhen Chen, Chih-Hong Cheng, Wenjie Ruan, Xiaowei Huang, Dezong Zhao, David Flynn, Siddartha Khastgir, and Xingyu Zhao. Trustworthy text-to-image diffusion models: A timely and focused survey. arXiv preprint arXiv:2409.18214, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[72] Yi Zhang, Yun Tang, Wenjie Ruan, Xiaowei Huang, Siddartha Khastgir, Paul Jennings, and Xingyu Zhao. Protip: Probabilistic robustness verification on text-to-image diffusion models against stochastic perturbation. In European Conference on Computer Vision, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.754 + ], + "angle": 0, + "content": "[73] Xingyu Zhao. Probabilistic robustness in deep learning: A concise yet comprehensive guide. Adversarial Example Detection and Mitigation Using Machine Learning, pages 1-13, 2025. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.755, + 0.905, + 0.811 + ], + "angle": 0, + "content": "[74] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2385-2392, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.812, + 0.905, + 0.867 + ], + "angle": 0, + "content": "[75] Monty-Maximilian Zuhlke and Daniel Kudenko. Adversarial robustness of neural networks from the perspective of Lipschitz calculus: A survey. ACM Computing Surveys, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.867 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1685" + } + ] +] \ No newline at end of file diff --git a/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_origin.pdf b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..005c0b839e34c34c576b5aa61ff2a8efe14cd508 --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/47fc789f-39f4-483e-ba24-6d818ddae7f8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:acb6ac3c8bbc984b73033989d73c9345418219216a5b7e43d3dafb8d3d667d95 +size 1582915 diff --git a/2025/Adversarial Training for Probabilistic Robustness/full.md b/2025/Adversarial Training for Probabilistic Robustness/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c006521a8c46a99a1e123614edb06dc55d3b209a --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/full.md @@ -0,0 +1,342 @@ +# Adversarial Training for Probabilistic Robustness + +Yi Zhang $^{1}$ , Yuhang Chen $^{1}$ , Zhen Chen $^{2}$ , Wenjie Ruan $^{2}$ , Xiaowei Huang $^{2}$ , Siddartha Khastgir $^{1}$ , Xingyu Zhao $^{1*}$ $^{1}$ WMG, University of Warwick $^{2}$ University of Liverpool {yi.zhang.16,yuhang.chen.1,s.khastgir.1,xingyu.zhao}@warwick.ac.uk; {zhen.chen2,xiaowei.huang}@liverpool.ac.uk; w.ruan@trustai.uk + +# Abstract + +Deep learning (DL) has shown transformative potential across industries, yet its sensitivity to adversarial examples (AEs) limits its reliability and broader deployment. Research on DL robustness has developed various techniques, with adversarial training (AT) established as a leading approach to counter AEs. Traditional AT focuses on worst-case robustness (WCR), but recent work has introduced probabilistic robustness (PR), which evaluates the likelihood of AEs within a local perturbation range, providing an overall assessment of the model's robustness and acknowledging residual risks that are more practical to manage. However, existing AT methods are fundamentally designed to improve WCR, and no dedicated methods currently target PR. To bridge this gap, we formulate a new min-max optimization as the theoretical foundation for PR-focused AT, and introduce an AT-PR training scheme with numerical algorithms to solve the new optimization problem. Our experiments, based on 70 DL models trained on common datasets and diverse architectures, demonstrate that: i) AT-PR achieves higher improvements in PR than AT-WCR methods; ii) it shows more consistent effectiveness across varying local inputs; iii) it exhibits a reduced trade-off in model's generalization. + +# 1. Introduction + +Deep Learning (DL) has shown potential for transformative impacts across many industries including transportation, healthcare and energy. The reality that DL is not as dependable as required now becomes a major impediment to their broader deployment. Since the first discovery of the phenomenon of Adversarial Examples (AEs) for DL [21, 39, 48], robustness is arguably the property in the limelight, encompassing research topics like adversarial attacks in malicious contexts [7, 8, 14, 21, 36, 64, 65, 75], + +AE detection during testing [19, 23, 24, 35, 42], verification with formal guarantees [20, 26, 30, 37, 44], and various robustness improvement techniques, including distillation [40], randomized smoothing [12], gradient masking [2], and the more widely used adversarial training (AT) [4, 21, 22, 36, 52]. As large language models and multimodal foundation models gain popularity, more recent research has increasingly focused on understanding their robustness as well [11, 18, 28, 32, 33, 58, 71, 72, 74]. + +Regardless of specific DL tasks or model architectures, robustness broadly refers to a model's ability to maintain consistent decisions under small input perturbations, with local robustness focusing on stability around individual inputs, as opposed to global robustness across the entire input distribution. More specifically, when evaluating local robustness, common problem formulations include [16, 72]: i) a binary question, "does an AE exist within a specified perturbation norm-ball?" [20, 30, 44, 63]; ii) a maximum safe distance question, "what is the largest perturbation radius within which no AE exists?" [1, 38, 60, 61]; and iii) a maximum prediction loss question, "which AE within the norm-ball yields the highest prediction loss?", which is often applied in AT to enhance robustness [36, 56]. It is evident that the three formulations focus on extreme scenarios, where local robustness is evaluated based on the existence of a deterministic AE, reflecting a worst-case perspective. + +A more recent and distinct view, however, considers probabilistic robustness (PR), employing statistical approaches to its evaluation [5, 13, 41, 49, 50, 55, 59, 60, 69, 72], it answers the question of "what is the likelihood of AEs in the given perturbation norm-ball?" [73]. However, existing AT methods are primarily designed for WCR via the well-established "min-max" optimization [3, 21, 36] (cf. Sec. 2.2), and are thus referred to as "AT-WCR" in this paper. No dedicated AT methods exist for enhancing PR, which motivates our work. + +In this paper, we first formulate AT for PR as a new min-max optimization problem, and then devise AT-PR training scheme based on numerical algorithms to effectively + +solve this training optimization. Our experiments on 70 DL models trained on common datasets and models demonstrate key strengths of AT-PR over AT-WCR: $i$ while traditional AT-WCR can also improve PR, our new AT-PR training scheme is more effective, especially when enhancing PR becomes challenging for traditional AT-WCR; $ii$ AT-PR not only outperforms AT-WCR in improving PR on average, but also offers more stable and consistent performance across different local inputs; $iii$ AT-PR has less negative impact on model's generalization compared to AT-WCR. + +In summary, key contributions of this paper include: + +- Problem formulation: We lay the foundation of AT for PR by formulating a new min-max optimization problem. +- Numerical algorithms: We develop effective numerical algorithms to solve the AT-PR min-max optimization. +- Empirical evaluation: We validate the effectiveness of AT-PR through extensive experiments, showing its superior improvement in PR, more consistent effectiveness across different local inputs, and reduced impact on generalization compared to AT-WCR. +- Open-source repository: A repository is provided at https://github.com/wellzline/AT-PR with code, datasets, models, and all experiment details. + +# 2. Preliminaries and Related Works + +# 2.1. Worst-Case vs Probabilistic Robustness + +Although definitions of robustness vary across different DL tasks and model types, it generally refers to a DL model's ability to maintain consistent predictions despite small input perturbations. Typically it is defined as all inputs in a region $\eta$ have the same prediction, where $\eta$ is a small norm ball (in a $L_{p}$ -norm distance) of radius $\gamma$ around an input $\mathbf{x}$ . A perturbed input (e.g., by adding noise on $\mathbf{x}$ ) $\mathbf{x}'$ within $\eta$ is an AE if its prediction label differs from $\mathbf{x}$ . + +Evaluation methods for robustness are typically based on metrics defined around AEs [27, 75]. As discussed earlier in Sec. 1, four types of robustness metrics, as summarized in [16, 72], address: (1) a binary question, (2) a maximum safe distance question, (3) a maximum prediction loss question (Fig. 1 (a)), and (4) a statistical inference question estimating the likelihood of AEs within a given perturbation norm-ball (Fig. 1 (b)). The first three metrics focus on extreme scenarios, where local robustness is assessed based on the existence of a deterministic AE, reflecting a worst-case perspective referred to as WCR. In contrast, the fourth approach adopts a probabilistic view, evaluating overall local robustness in the presence of AEs. + +PR is arguably more practical than WCR, as extreme-case focus is often unnecessary [73]. Instead, knowing the likelihood of AEs is more relevant for applications, where it is sufficient to show that the risk of AE remains below an acceptable threshold rather than exactly zero [59, 69, 72]. + +Inspired by [72], we illustrate the difference between the maximum loss WCR and PR in Fig. 1, omitting the binary and maximum safe distance cases as they are less relevant to the AT focus of this paper. + +![](images/697ef97724b8801d9b0c18d3955ce50176773663612f33571c447d95f526fad7.jpg) +Class A + +![](images/7ea8f056e35a19db73b4dea3e51e56a5cf261303c81d35f1aacd72659f419106.jpg) +Figure 1. Worst-case (a) vs. Probabilistic (b) Robustness. + +![](images/705aca3d16f6fe087b1984d786ea29dbb9960ec3ffa2043cd949721350f58b5c.jpg) +# + +![](images/c10ef37e22f6aba6edf80d440dd76afbc145614a6e3a7a51de6935727acb24c8.jpg) +Non-AE + +![](images/6d162242fbfe60e28f0b56daa25a7761ca8c42b90a4b5728025c344954358ec6.jpg) + +To evaluate the WCR, we normally formulate it as a question of maximizing the prediction loss, cf. Fig. 1 (a), where adversarial attacks are typically introduced to find such a worst-case AE with perturbation: + +$$ +\boldsymbol {\delta} = \arg \max _ {\| \boldsymbol {\delta} \| \leq \gamma} \mathcal {L} (\boldsymbol {x} + \boldsymbol {\delta}, y; \boldsymbol {\theta}), \tag {1} +$$ + +where $\delta$ is a small perturbation constrained within an $L_{p}$ -norm ball of radius $\gamma$ . This optimization causes the model to misclassify $\boldsymbol{x}^{\prime} = \boldsymbol{x} + \delta$ as a different label $y^\prime \neq y$ . There are some popular attacks following this formulation, including FGSM [21], PGD [36], and ZOO [10], among others. Following common practice, we measure WCR by the success rate of some selected attacks on a test dataset. Estimation of PR is normally formalized as the statistical inference problem on the following metric: + +Definition 1 (Probabilistic Robustness) For a DL classifier $f_{\theta}$ that takes input $x$ and returns a prediction label, the PR of an input $x$ in a norm ball of radius $\gamma$ is: + +$$ +P R (\boldsymbol {x}, \gamma) = \underset {\| \boldsymbol {\delta} \| \leq \gamma} {\mathbb {E}} _ {\delta \sim P r (\cdot | \boldsymbol {x})} [ I _ {\{f _ {\boldsymbol {\theta}} (\boldsymbol {x} + \boldsymbol {\delta}) = y \}} (\boldsymbol {x} + \boldsymbol {\delta}) ], \tag {2} +$$ + +where $I_{S}(\pmb{x})$ is an indicator function—it is equal to 1 when $S$ is true and 0 otherwise; $Pr(\cdot)$ is the local distribution of inputs representing how perturbations $\delta$ are generated, which is precisely the "input model" used by [59, 72]. + +Def. 1 suggests PR is the probability that the model prediction remains unchanged from a random perturbation $x'$ . A "frequentist" interpretation of this probability is: it is the limiting relative frequency of perturbations where the output label is preserved, in an infinite sequence of independently generated perturbations [72]. That is, the "proportion" of non-AEs in such infinite set of perturbations. + +The work [59] is the first to define PR and introduces a black-box statistical estimator, particularly useful when + +PR is extremely high (i.e., detecting AEs is a rare event). A more efficient white-box statistical estimator for PR was later proposed in [50]. Huang et al. [25] first extend PR to explainable AI (XAI). Zhang et al. [69] explore PR under functional perturbations, while the recent study [72] is the first to estimate PR in the text-to-image DL task. The recent work [73] provides an overview of PR, including its definition, estimation, enhancement and system-level integration. Without loss of generality, we adopt the estimator in [59] to estimate PR in later experiments. + +# 2.2. Adversarial Training + +AT [4, 21, 22, 36, 52] is a widely used empirical approach for improving robustness, typically formulated as a "min-max" optimization problem. It introduces a set of input perturbations $\delta$ within a norm-ball $\eta$ of radius $\gamma$ , designed to maximize the training loss $\mathcal{L}$ . The model is then trained with these perturbations by minimizing the overall loss of the training dataset. These perturbations are often approximated using empirical adversarial attacks, such as FGSM [21] and PGD [36] due to their high efficiency. Taking a classification task as an example, the min-max optimization target of AT can be formalised as: + +Definition 2 (AT for WCR) For a DL model $f_{\theta}$ , trained on a dataset $\mathcal{D}$ , consisting of pairs $(\pmb{x}, y)$ where $\pmb{x} \in \mathcal{X}$ is the input and $y \in \mathcal{Y}$ denotes the label, with loss function $\mathcal{L}(\pmb{x}, y; \theta)$ , the objective of AT-WCR can be formulated as: + +$$ +\min _ {\boldsymbol {\theta}} \mathbb {E} _ {(\boldsymbol {x}, y) \sim \mathcal {D}} \left[ \max _ {\| \delta \| \leq \gamma} \mathcal {L} (\boldsymbol {x} + \boldsymbol {\delta}, y; \boldsymbol {\theta}) \right], \tag {3} +$$ + +where $\delta$ represents the perturbation that maximizes classification loss $\mathcal{L}$ , while the outer minimization aims to train $f_{\theta}$ to be robust against such adversarial perturbations. + +The inner maximization is precisely the adversarial attack formulation outlined in Eq. (1). As highlighted in [3], the effectiveness of AT largely depends on identifying strong AEs within this inner maximization. Notably, the PGD attack [36] has become a critical benchmark and is widely regarded as the standard method for AT in practice [3]. Accordingly, we also use PGD in later experiments. Based on the standard min-max optimization, various AT variants have been proposed to achieve more stable and efficient training, such as [6, 15, 29, 47, 52, 57, 58, 66, 67]. Comprehensive overviews of recent AT variants can be found in surveys [3, 27], all of which fundamentally rely on the min-max problem for WCR. + +In contrast, to the best of our knowledge, only three works [43, 55, 70] have explicitly aimed at improving PR [73]. Although Wang et al. [55] demonstrate that simply adding randomly perturbed inputs significantly enhances PR, the resulting AR is nearly zero. The methods of [43] and [70] propose loss functions based on common risk + +measures (i.e., $\rho$ -essential supremum, Conditional Value-at-Risk, and Entropic Value-at-Risk) to train models to minimize these risks. However, again, these approaches provide little to no improvement in AR. We argue that, akin to the well-established AT framework for AR, a dedicated AT framework for PR is necessary—one that does not compromise AR performance. Accordingly, our proposed AT-PR is the first to explicitly reformulate the min-max optimization principle for PR, while also providing comparable AR performance through practical training algorithms. + +# 3. The Approach: AT-PR + +# 3.1. Problem Formulation + +Intuitively, the goal of AT-WCR in Def. 2 is to identify the AE that maximizes model loss and incorporate it into AT to lower the maximum loss, cf. the top-middle illustration in Fig. 2. However, to effectively improve local PR, we aim to locate the "widest peak"—the region with the highest proportion of AEs in the loss landscape (cf. the top-right illustration in Fig. 2), rather than the tallest peak. Adding the AE from this peak during AT may ideally reduce the loss for AEs around that area, thereby lowering the overall AE proportion (cf. the intuitive interpretation of PR in Def. 1). In some cases, AT-WCR and AT-PR may identify the same AE for training when the tallest peak also happens to be the widest, as shown in the second row of Fig. 2. Notably, the two cases illustrated in Fig. 2 representatively cover all possible scenarios, though additional peaks may exist in reality. Given the difference between the goals of improving WCR and PR, we formulate a new min-max problem for AT-PR: + +Definition 3 (AT for PR) For a DL model $f_{\theta}$ , trained on a dataset $\mathcal{D}$ consisting of pairs $(x, y)$ , the objective of AT-PR can be formulated as: + +$$ +\min _ {\boldsymbol {\theta}} \mathbb {E} _ {(\boldsymbol {x}, y) \sim \mathcal {D}} \left[ \max _ {\| \delta \| \leq \gamma , P R (\boldsymbol {x} + \delta , k) = 0} k \right], \tag {4} +$$ + +where $k$ and $\delta$ are the variables in the new inner max optimization, $PR(\cdot, \cdot)$ is the metric defined in Eq. (2). As shown in Fig. 3, the intuition behind the maximization is to find an optimal $\delta^*$ that produces an $AE$ $x' = x + \delta^*$ and to maximize the radius $k$ of a smaller norm-ball centered on $x'$ such that the $PR$ within this $k$ -sized norm-ball is 0 (i.e., all inputs within this smaller norm-ball are AEs). + +The high non-convexity of DL models makes it intractable to analytically solve both the established min-max optimization in Eq. (3) and our new formulation in Eq. (4). Thus, similar to existing AT-WCR, we propose algorithms to numerically solve Eq. (4) in AT-PR in the next section. + +# 3.2. Optimization Algorithm + +While the outer minimization in Eq. (4) can be effectively handled with standard DL training methods as in AT-WCR, + +![](images/45e4646f4632831e825b287d49e36f3ff221b061eaaeee41001d36066dfec770.jpg) + +![](images/d0bc69b4220e5a18c1848fbc1a91af8d0edd253f3ae476d3f8ff863131525636.jpg) + +![](images/f2638fb90f667e4c111681f00ca826b384bdb6362f1cdc3095aa6e0bb95d0d13.jpg) + +![](images/e9125c9065c840c5551d6c0e4c3ad46f6fa230015a81219c64d1b96cbb5bff4d.jpg) +Figure 2. Two cases of the local loss landscape and their AEs identified by AT-WCR and AT-PR for AT. + +![](images/7f4eac3de2261e7295b0c44f242116eed48d38466123235dd0f9994e2e5a2d98.jpg) + +![](images/8239a33d05b6f04482cc45ce6cf6271c4340114e5088b746541ecf33e6e4a05c.jpg) + +![](images/0b1e3431703387c08f43c69b982e145e3aa4ad1c3362acf2805616a8f009fa88.jpg) +Figure 3. Illustration of the new inner-max optimization in Def. 3. + +we devise the Algo. 1 to solve the inner optimization in our AT-PR. The gist of Algo. 1 is shown in Fig. 4, which essentially consists of 3 main steps: + +- Step 1 (lines 1-9): Apply PGD attacks in different regions of the local norm-ball under study to find a diverse set of AEs at various local optima $x_{i}^{\prime}$ . +- Step 2 (lines 10–20): For each locally optimal AE found in Step 1, perform a gradient descent search to locate the nearest decision boundary. +- Step 3 (lines 21-25): Calculate the distance $d_{i}$ from each $x_{i}^{\prime}$ to its respective decision boundary and return the AE with the largest $d_{i}$ value. + +Similar to AT-WCR, the returned AE will be added into the training dataset in AT-PR. We note, as an numerical algorithm to the optimization, its effectiveness may depends on certain assumptions include: $i$ ) the PGD attack in Step 1 may effectively find local optima by adjusting hyperparameters such as attack steps and learning rate, which aligns with the insight of "PGD will converge to local maxima" + +[36]; ii) The gradient descent based boundary search with adaptive hyperparameters may efficiently find the decision boundary as used in some boundary attacks [9, 31, 68]. We also provide an initial theoretical analysis of the effectiveness and convergence properties of the proposed optimization algorithm in Appendix A. + +Step 1: Apply PGD in different regions of the norm-ball to get a set of diverse AEs at different local optimal points $x_{i}^{\prime}$ . + +![](images/86c16682dcb38119d2433013afa9f2335d278cd739a57b3ec64c40fc4d9d3671.jpg) +Figure 4. Illustration of the 3 main steps of Algo. 1. + +We analyze the time complexity of the proposed algorithm. There are several computational steps to be considered with the assumption that the dimension of the input $\pmb{x}$ is $n$ . First, sampling $\alpha$ and steps from uniform distributions (lines 5-6) is independent of the input and thus the time complexity is $O(1)$ . Second, the PGD algorithm (line 7) consists of steps gradient calculations with complexity $O(\text{grad})^1$ and one projection calculation with complexity + +Algorithm 1 PGD and gradient-based search for $x_{pr}^{\prime}$ +Require: Inputs $[X,Y],N,\alpha_{min},\alpha_{max},step_{min},step_{max}$ 1: Initialize $AE_{s}\gets []$ Initialize AE candidate set 2: Apply PGD to get different AE candidates 3: for idx = 0 to N do 4: $x_{init}\gets x + \mathrm{Uniform}(-\gamma ,\gamma)$ 5: $\alpha \leftarrow$ Uniform $(\alpha_{min},\alpha_{max})$ 6: steps $\leftarrow$ Uniform $(step_{min},step_{max})$ 7: $x_{idx}^{\prime}\gets pgd(x_{init},y,\gamma ,\alpha ,steps)$ 8: Append $x_{idx}^{\prime}$ to $AE_{s}$ 9: end for +Require: $AE_s,x,y,C$ +10: Initialize Max_d $\leftarrow 0$ . $x_{pr}^{\prime}\gets$ None 11: for each $x^{\prime}\in AE_{s}$ do 12: $\tilde{x} = x^{\prime}$ 13: while iter $< C$ do 14: $y^\prime = f(\tilde{x})$ 15: if $y^\prime = y$ then 16: break $\triangleright$ Exit if $\tilde{x}$ is classified correctly 17: end if 18: $g\gets \nabla_{\tilde{x}}L(\tilde{x},y)$ $\triangleright$ Compute gradient 19: $\tilde{x} = \tilde{x} -\alpha \cdot g$ $\triangleright$ Update $\tilde{x}$ +20: end while +21: $d\gets D(\tilde{x},x^{\prime})$ $\triangleright$ Distance between $\tilde{x}\& x^{\prime}$ +22: if $d > Max\_ d$ then +23: $x_{pr}^{\prime} = x^{\prime}$ +24: Max_d=d +25: end if +26: end for +27: Output $x_{pr}^{\prime}$ Farthest AE from decision boundary. + +$O(n)$ . Third, in the effort to locate the decision boundary (lines 18-19), one gradient calculation step is performed. Last, a $L_{\infty}$ distance between $\tilde{\pmb{x}}$ and $\pmb{x}^{\prime}$ is estimated (line 21), with complexity $O(n)$ . In Step 1, the algorithm uses different random starts via random initialization [46] to ensure it finds different local optima within norm-ball $\eta$ (line 4), and generates one AE (lines 5-8) for each combination of input $(x,y)$ , radius $\gamma$ , randomly sampled $\alpha$ and steps. A total of $N$ AEs are generated. Thus the complexity upper bound of Step 1 is then $O(N*(steps*O(grad)+O(n)))$ . In Step 2, with the maximum iteration number $C$ for the nearest decision boundary locating process of $N$ AEs, the complexity upper bound of Step 2 is then $O(C*N*O(grad))$ . In Step 3, the complexity of distance estimation is $O(n)$ as discussed above. Assuming $O(grad) \gg O(n)$ , which + +architecture and the specific loss function employed during training. E.g., DL models with more layers or complex loss functions may require more computational resources for gradient computation due to backpropagation. + +typically holds in DL scenarios, the complexity simplifies to $O(N * (steps + C) \cdot O(grad))$ . As analyzed, the computational cost of AT-PR primarily comes from locating the optimal AE. However, it remains compatible with other techniques designed to enhance AT efficiency, such as [46, 62]. We provide comparative empirical results by replacing PGD with Fast-AT [62] to demonstrate such compatibility in Table 3, refer to Sec. 4 for details. + +# 4. Experiments + +We train a total of 70 DL models on four datasets: CIFAR-10, CIFAR-100, SVHN, and Tiny/ImageNet, using models such as ResNet, Wide-ResNet (WRN), Vision Transformer (ViT) [17, 34], and DeiT [51]. To evaluate AT-PR, we compare it with two standard AT-WCR (PGD and FGSM) and four advanced methods: TRADES [66], MART [57], ALP, and CLP [29]. Implementation details are in Appendix B. + +# 4.1. Effectiveness and Efficiency of AT-PR + +We assess AT-PR's effectiveness against AT-WCR from two perspectives: 1) improvement in PR and 2) impact on generalization. We compare models trained with AT-PR, AT-WCR (FGSM and PGD) and a baseline without AT, using three metrics: (i) Accuracy on clean test data; (ii) Adversarial Accuracy under PGD, C&W, and AA attacks to reflect WCR; and (iii) PR, computed as $PR(x, \gamma)$ for various $\gamma$ using the estimator from [59]. We report the mean and standard deviation of PR on the test set, along with the generalization error (GE) to measure train-test robustness gaps. + +Key findings summarized from Table 1: + +i) All original models without AT show low WCR. As expected, PGD is the most effective in improving WCR, though it compromises generalization the most. AT-PR also enhances WCR, outperforming FGSM and, in some cases, even surpassing PGD, as seen with SVHN/ResNet-18, Tiny-ImageNet/ResNet-34 for PGD.Ac., and SVHN/WRN50-2, Tiny-ImageNet/ResNet-18, ResNet-34 for CW.Ac. +ii) Both AT-WCR with FGSM and PGD can improve PR compared to the original models without AT, but they are less effective when compared to the performance of AT-PR. Specifically, the mean PR evaluations of AT-PR-trained models are higher than those trained with AT-WCR across most datasets, model architectures and norm-ball sizes. For instance, on CIFAR-100 with ResNet-18, AT-PR improves $PR(\cdot, 0.15)$ from 44.71 (FGSM) and 80.47 (PGD) to 87.38. +iii) Compared to AT-WCR, the improvement in PR becomes more pronounced as the norm-ball size $\gamma$ increases, which is anticipated since larger norm-balls generally lead to less robust models. This demonstrates that our AT-PR approach is particularly useful when enhancing PR becomes more challenging for traditional AT-WCR. +iv) Each local input is associated with an estimated PR, and we report the mean and standard deviation over the test + +Table 1. Evaluation results of DL models trained with different AT schemes, showing standard accuracy, adversarial accuracy (indicating WCR: PGD-10, C&W-10), and PR estimated (mean ± std) for four norm-ball sizes $\gamma$ . All results are calculated over the test datasets. + +
DatasetModelAT SchemeAccu. %PGD Ac. %CW Ac. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18W/O AT94.960.020.5178.38 ± 33.1466.3 ± 38.0653.41 ± 40.7637.01 ± 39.77
AT-WCR(FGSM)88.60.090.0993.99 ± 17.588.02 ± 25.0380.42 ± 31.8167.77 ± 38.6
AT-WCR(PGD)8145.0144.5697.28 ± 13.2995.77 ± 17.0794.1 ± 20.289.78 ± 26.53
AT-PR80.1439.7739.0398.22±10.3897.28±13.1696.05±16.1893.21±21.47
WRN-50-2W/O AT94.140.230.1476.96±32.7665.65±37.4154.43±39.8840.19±39.49
AT-WCR(FGSM)92.040.080.0495.68±14.5391.9±20.2586.67±25.9876.79±33.1
AT-WCR(PGD)82.6249.5148.1397.86±12.6197.08±14.5495.84±17.1492.62±22.87
AT-PR82.7143.2743.0798.09±10.8197.33±13.0396.22±15.5893.56±20.89
CIFAR-100ResNet-18W/O AT75.50.020.044.58 ± 42.6432.1 ± 40.123.32 ± 36.4115.62 ± 31.57
AT-WCR(FGSM)67.801.410.6683.90 ± 31.1172.89 ± 37.9461.06 ± 42.2344.71 ± 43.04
AT-WCR(PGD)52.8519.9319.5195.4 ± 18.1592.7 ± 22.7288.98 ± 28.1980.47 ± 36.4
AT-PR53.2119.3818.3997.59±12.9695.96±16.6293.15±21.7787.38±29.63
WRN-50-2W/O AT75.460.490.4551.11±40.5437.96±39.2328.33±36.7219.11±31.97
AT-WCR(FGSM)73.170.290.185.68±27.4576.62±33.8966.37/38.6652.04±41.42
AT-WCR(PGD)57.4623.1722.595.31±17.3892.0±23.7988.14±29.1580.67±36.21
AT-PR59.4622.0621.5996.05±16.1793.72±20.4789.48±26.8781.6±34.86
SVHNResNet-18W/O AT94.960.080.1297.11 ± 9.6695.63 ± 11.9893.74 ± 14.4890.12 ± 18.40
AT-WCR(FGSM)93.494.685.3298.79 ± 7.0698.42 ± 7.7097.86 ± 8.6796.57 ±10.78
AT-WCR(PGD)89.2144.5743.7098.96 ± 5.8598.50 ± 7.4097.72 ± 9.6195.41 ± 14.41
AT-PR90.2945.8142.2199.12 ± 5.1598.87 ± 5.9398.54 ± 6.8497.84 ± 8.60
WRN-50-2W/O AT95.911.741.2996.60 ± 10.3694.77 ± 13.0192.40 ± 15.8787.97 ± 20.04
AT-WCR(FGSM)94.318.138.5498.34 ± 7.4997.71 ± 8.3996.80 ± 9.6995.12 ± 12.06
AT-WCR(PGD)88.9154.9950.5898.72 ± 6.4498.28 ± 7.6697.76 ± 8.9596.77 ± 11.18
AT-PR91.9752.2251.5898.88 ± 6.1298.38 ± 7.5397.72 ± 9.1996.26 ± 12.35
Tiny-ImageNetResNet-18W/O AT59.260.00.080.78 ± 33.172.95 ± 37.6964.17 ± 40.9151.14 ± 42.91
AT-WCR(FGSM)54.810.050.0189.9 ± 24.9785.33 ± 29.9580.18 ± 33.7469.03 ± 39.39
AT-WCR(PGD)47.4321.319.3196.32 ± 17.0193.79 ± 22.190.16 ± 27.7682.46 ± 35.97
AT-PR51.6121.119.8998.23 ± 11.0696.1 ± 17.8493.12 ± 23.3286.65 ± 32.26
ResNet-34W/O AT61.660.00.082.39 ± 31.375.2 ± 36.066.95 ± 39.4253.7 ± 42.17
AT-WCR(FGSM)570.130.0290.67 ± 23.9285.81 ± 29.580.57 ± 34.1370.22 ± 39.36
AT-WCR(PGD)49.3320.9719.5895.45 ± 18.5593.08 ± 23.4889.85 ± 28.383.53 ± 34.99
AT-PR53.9821.7220.9596.71 ± 15.7495.02 ± 19.8992.5 ± 23.8786.23 ± 32.29
+ +set in Table 1. Notably, AT-PR consistently achieves higher mean PR compared to models trained with AT-WCR, while also exhibiting lower standard deviations. This indicates that AT-PR not only improves PR but also offers more stable and consistent performance across different local inputs. + +$\nu$ ) While PGD significantly improves robustness, it often reduces accuracy [66]. In contrast, AT-PR has a smaller negative impact on accuracy compared to PGD, with almost all models achieving higher accuracy than those trained using PGD while also enhancing both WCR and PR. This effect is more pronounced on larger datasets—for example, on Tiny-ImageNet with ResNet-18, AT-PR achieves an accuracy of $53.98\%$ compared to $49.33\%$ for PGD. + +Table 2 shows that combining AT-PR with TRADES improves clean accuracy and PR on more complex attention-based models ViT (Vanilla ViT and DeiT) across different scales (DeiT-Tiny, ViT-Small, ViT-Base) for CIFAR10. With just five epochs of AT-PR fine-tuning, all three ViT models achieve higher clean accuracy and PR while maintaining comparable WCR, showing a slight decrease on PGD-20 and AA but improved performance on CW-20. + +Table 3 presents a comparison of AT-PR with four advanced AT methods (TRADES, MART, ALP and CLP) across three datasets using ResNet-18 and WRN-50-2. Re + +sults show that AT-PR consistently achieves the best PR performance across four different norm-ball sizes $\gamma$ . Additionally, to verify that AT-PR is compatible with efficient AT methods [46, 62], we replace PGD with Fast-AT [62] in our framework (step 1 in Algo. 1). The result shows that Fast-AT-PR achieves $\sim 40\%$ speedup over the original AT-PR, with only $\sim 50\%$ extra cost compared to other AT baselines. Notably, Fast-AT-PR retains comparable PR performance to AT-PR and continues to outperform other AT baselines. + +# 4.2. Generalization Error of AT-PR + +We investigate the generalization error (GE) (train-test gap) of AT-PR to gain insights into its effectiveness. Table 4 compares the PR GE between AT-PR and PGD. Models trained with AT-PR consistently yields lower GE of up $2.2\%$ on CIFAR-10, $2.87\%$ on CIFAR-100, and $9.66\%$ on TinyImageNet, significantly lower than PGD's $4.67\%$ , $8.54\%$ , and $11.94\%$ , respectively. We further analyze the underlying reasons based on the PAC-Bayes analysis for Adversarial Robustness [54], see Appendix C for details. + +# 4.3. AEs Used in AT-WCR vs. AT-PR + +We analyze the AEs used during training to understand how AT-PR differs from AT-WCR. By comparing the AEs gen + +Table 2. ViT models trained on CIFAR-10 with TRADES and AT-PR, showing standard accuracy, adversarial accuracy (indicating WCR: PGD, C&W, AA), and PR estimated (mean ± std) for four norm-ball sizes $\gamma$ . + +
DatasetModelAT SchemeAccu. %PGD-20. %CW-20. %AA. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10DeiT-TiTRADES79.4449.7846.4545.9696.61 ± 15.2394.73 ± 19.3692.69 ± 23.0489.18 ± 27.9
+AT-PR81.0749.0846.9245.9597.26 ± 14.095.94 ± 17.1294.16 ± 20.590.65 ± 25.88
Vit-STRADES81.0350.9648.1447.4096.93 ± 14.9995.48 ± 18.493.58 ± 21.6290.06 ± 26.81
+AT-PR83.2950.4448.2047.3497.76 ± 11.8996.03 ± 16.4693.98 ± 20.6990.48 ± 26.19
Vit-BTRADES83.5553.1450.9449.9697.34 ± 13.5795.87 ± 16.9993.78 ± 21.1790.77 ± 25.88
+AT-PR86.0352.9451.2449.3497.9 ± 11.896.49 ± 15.294.44 ± 19.6291.02 ± 25.27
+ +Table 3. Comparison of PR(, γ) and training time (sec/epoch) for more recent AT methods (TRADES, MART, ALP, CLP) and Fast-AT-PR. + +
DataModelMetricAT Scheme
PGDTRADESMARTALPCLPFast-AT-PRAT-PR
CIFAR-10ResNet-18PR(,0.08)%97.28 ± 13.2997.74 ± 12.4297.4 ± 13.796.73 ± 15.997.61 ± 13.0698.12 ± 11.5298.22 ± 10.38
PR(,0.10)%95.77 ± 17.0796.33 ± 16.3995.92 ± 17.4395.16 ± 19.5396.29 ± 16.6697.1 ± 14.3497.28 ± 13.16
PR(,0.12)%94.10 ± 20.2094.55 ± 20.3193.93 ± 21.3893.24 ± 23.0894.54 ± 20.2895.4 ± 17.9096.05 ± 16.18
PR(,0.15)%89.78 ± 26.5391.22 ± 25.7090.35 ± 26.9189.76 ± 28.1490.86 ± 26.2392.71 ± 23.5493.21 ± 21.47
Time (s/ep.)546144646497184
WRN-50-2PR(,0.08)%97.86 ± 12.6197.55 ± 13.1497.40 ± 13.6197.75 ± 12.4397.40 ± 13.9197.96 ± 10.8298.09 ± 10.81
PR(,0.10)%97.08 ± 14.5496.14 ± 16.7096.15 ± 16.7896.42 ± 16.1196.20 ± 17.5897.22 ± 12.9697.33 ± 13.03
PR(,0.12)%95.84 ± 17.1494.33 ± 20.4194.41 ± 20.3794.64 ± 20.0194.95 ± 19.8896.22 ± 15.6296.22 ± 15.58
PR(,0.15)%92.62 ± 22.8791.01 ± 25.7991.11 ± 25.6591.22 ± 25.7091.48 ± 25.6994.74 ± 19.8893.56 ± 20.89
Time (s/ep.)13110571155157229305
CIFAR-100ResNet-18PR(,0.08)%95.40 ± 18.1595.02 ± 19.2795.45 ± 18.1196.20 ± 16.0796.16 ± 16.3697.09 ± 13.4897.59 ± 12.96
PR(,0.10)%92.70 ± 22.7292.09 ± 24.2292.73 ± 23.3994.40 ± 20.0893.81 ± 21.4294.67 ± 18.9195.96 ± 16.62
PR(,0.12)%88.98 ± 28.1987.80 ± 29.8089.35 ± 28.3292.05 ± 24.1291.05 ± 25.8692.63 ± 23.9193.15 ± 21.77
PR(,0.15)%80.47 ± 36.4079.20 ± 37.7381.93 ± 35.7285.89 ± 31.6584.97 ± 33.0085.87 ± 31.9987.38 ± 29.63
Time (s/ep.)556143646597180
WRN-50-2PR(,0.08)%95.31 ± 17.3895.24 ± 18.6195.30 ± 16.8795.34 ± 17.7995.29 ± 17.7096.52 ± 15.1196.05 ± 16.17
PR(,0.10)%92.00 ± 23.7992.30 ± 23.6192.59 ± 22.8192.36 ± 23.1991.79 ± 23.8793.49 ± 20.9393.72 ± 20.47
PR(,0.12)%88.14 ± 29.1588.65 ± 28.4688.38 ± 28.6588.51 ± 28.2287.54 ± 29.9288.82 ± 28.3889.48 ± 26.87
PR(,0.15)%80.67 ± 36.2180.35 ± 36.5480.79 ± 36.0180.37 ± 36.8679.71 ± 36.7681.02 ± 37.0981.60 ± 34.86
Time (s/ep.)13110270155156225312
SVHNResNet-18PR(,0.08)%98.96 ± 5.8599.03 ± 5.7698.99 ± 6.1599.05 ± 5.7399.12 ± 5.3398.90 ± 5.9799.12 ± 5.15
PR(,0.10)%98.50 ± 7.4098.46 ± 8.2798.31 ± 8.2098.63 ± 7.1598.60 ± 7.2198.53 ± 7.2498.87 ± 5.93
PR(,0.12)%97.72 ± 9.6197.30 ± 11.0897.10 ± 11.2697.85 ± 9.1397.66 ± 9.7898.11 ± 8.1998.54 ± 6.84
PR(,0.15)%95.41 ± 14.4194.15 ± 16.5794.12 ± 16.6595.54 ± 14.1094.93 ± 15.0097.10 ± 10.0597.84 ± 8.60
Time (s/ep.)7886619193141202
WRN-50-2PR(,0.08)%98.72 ± 6.4498.69 ± 6.9098.61 ± 7.6698.54 ± 7.5298.59 ± 7.4698.43 ± 6.6998.88 ± 6.12
PR(,0.10)%98.28 ± 7.6698.11 ± 8.6698.07 ± 9.2798.12 ± 8.7498.00 ± 8.8998.15 ± 7.2998.38 ± 7.53
PR(,0.12)%97.76 ± 8.9597.26 ± 10.8997.34 ± 10.9097.65 ± 8.8596.98 ± 11.1397.79 ± 8.1297.72 ± 9.19
PR(,0.15)%96.77 ± 11.1895.20 ± 14.8495.45 ± 14.7295.67 ± 14.3294.74 ± 15.3796.94 ± 9.8096.26 ± 12.35
Time (s/ep.)190146102223225321420
+ +erated by PGD and AT-PR, we find that about $10\%$ of training inputs produce different AEs (e.g., $10.05\% / 8.39\%$ for ResNet-18 and $8.83\% / 10.42\%$ for WRN50-2 on CIFAR-10/CIFAR-100). This means that for roughly $90\%$ of inputs, the loss landscapes align with Case 2 ("tall & wide vs. short & tight") in Fig. 2, where both methods select the same AEs. However, for the remaining $10\%$ -ish, where the landscape fits Case 1 ("tall & tight vs. short & wide"), AT-PR identifies different AEs that better optimize PR. + +Fig. 5 (top row) compares the loss of AEs generated by PGD and AT-PR. For all models, PGD AEs consistently have higher losses, reflecting its objective to find AE with the maximum loss for AT. This aligns with the goal of AT-WCR, which seeks to minimize the loss on such worst-case AEs (see Fig.2, top-middle). In contrast, AT-PR generates AEs with lower loss. Combined with the fact that our Algorithm 1 searches for examples farthest from the deci + +sion boundary, suggests that AT-PR effectively locates AEs within the widest peak of the loss landscape. Training with these AEs helps reduce the "short & wide" peak (see Fig. 2, top-right). + +Fig. 5 (b) visualizes the distributions of all generated AEs $\{x_i^{\prime}\}_{i = 1}^{n}$ by analyzing the distance $(L_{2}$ norm) between each $x_{i}^{\prime}$ and the original input $x$ (e.g., $d_0,d_1$ in Fig. 5 (a)), along with their angular relationships. The 2D coordinate in Fig. 5 (a) introduces a fixed anchor $x_{a}$ near $x$ , for an AE $x_0^\prime$ , the angle $\theta_0$ is computed between $\vec{v}_{x\rightarrow x_0'}$ and $\vec{v}_{x\rightarrow x_a}$ , representing $x_0'$ 's relative position, given by: $\theta_0 = \arccos \left(\frac{\vec{v}_{x\rightarrow x_0'}\cdot\vec{v}_{x\rightarrow x_a}}{\|\vec{v}_{x\rightarrow x_0'}\||\vec{v}_{x\rightarrow x_a}\|}\right)$ . The resulting angles range from approximately $88^{\circ}$ to $92^{\circ}$ Fig. 5 (b), reflecting the unique geometric properties of high-dimensional spaces, where random vectors highly likely to be orthogonal [53]. To further capture the variability in relative positions, + +Table 4. Comparative GE experiment on DL models across different datasets. Train/test set evaluations of PR estimates (mean) for four norm-ball sizes $(\gamma)$ across models trained with PGD and AT-PR, presented as $PR_{train} / PR_{test} / (PR_{train} - PR_{test})$ + +
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
WRN-50-2PGD98.88/97.86/1.0198.19/97.08/1.197.49/95.84/1.6495.3/92.62/2.67
AT-PR98.63/98.09/0.5497.8/97.33/0.4796.71/96.22/0.4694.9/93.56/1.34
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
WRN-50-2PGD98.35/95.31/3.0496.71/92.0/4.7193.67/88.14/5.5386.69/80.67/6.02
AT-PR97.3/96.05/1.2595.41/93.72/1.6991.93/89.48/2.4583.25/81.6/1.65
Tiny-ImageNetResNet-18PGD98.72/96.32/2.497.86/93.79/4.0796.94/90.16/6.7894.4/82.46/11.94
AT-PR99.01/98.23/0.7898.02/96.1/1.9296.68/93.12/3.5694.31/86.65/7.66
ResNet-34PGD98.67/95.45/3.2298.02/93.08/6.9496.9/89.85/7.0593.36/83.53/9.83
AT-PR99.28/96.71/2.5798.82/95.02/3.898.11/92.5/5.6195.89/86.23/9.66
+ +Table 5. PR (mean) on train/test sets, reported as: ${PR}_{\text{train }}/{PR}_{\text{test }}/\left( {{PR}_{\text{train }} - P{R}_{\text{test }}}\right)$ ,for model trained with PGD, Random-AE, AT-PR. + +
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
Random98.62/97.71/0.9197.73/96.78/0.9596.58/95.48/1.194.19/92.64/1.54
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
Random96.6/95.87/0.7394.85/93.77/1.0892.18/90.92/1.2685.39/85.21/0.18
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
+ +![](images/ce2f8a4c449f5f3d9bf5a24c5f87d0ddfe062327b258dc554d9233bde8b8847a.jpg) + +![](images/9c53a0c6169b3a8975f751a0993ad809813e9ee9ca9d59962975d0b870946566.jpg) + +![](images/d38565850766f48f6710a0e15fb24e758a3b910ecaf73ef27e54942e04a01449.jpg) + +![](images/6944ecc0522a75de4e80259b8222c967df4093025b1850393d951237bc4db04d.jpg) +Figure 5. (Top row) Averaged loss of AEs used during AT for AT-PR and PGD. (Bottom row) (a) Illustration of angles and distances of AEs relative to $x$ . (b) Distribution of generated AEs. (c) Pairwise angular distribution among AEs. + +![](images/c11df29aff62af65b87111c9d103e94deb370b57456449e113dfa9066ddab0e7.jpg) + +![](images/0ee19c9c6a7cd7cf84f3e63d92f457a3e2f9a3a8b57392338500c5b1526c2ab2.jpg) +(c) + +we computed the Pairwise Angular Distribution among all vectors $\{\vec{v}_{x\rightarrow x_i'}\}_{i = 1}^n$ . The angle $\beta$ for each pair of vectors is calculated as illustrated in Fig. 5 (a), with the resulting distribution shown in Fig. 5 (c). This analysis shows that the angles between all pairs of vectors range from $60^{\circ}$ to $85^{\circ}$ , indicating that the generated AEs are widely spread within the norm-ball. The final AE selected by AT-PR, highlighted in orange in Fig. 5 (b), has the largest distance to nearest decision boundary (1.0 for the L2 norm), compared to other + +AEs, which range from 0.76 to 0.92. + +To empirically validate that AEs generated by AT-PR are most effective for improving PR, we conduct an ablation study by randomly sampling AEs from those generated in Step 1 of Algo. 1 (also illustrated as random blue spots in Fig.5 (b)). Table 5 shows that while randomly selected AEs improve PR and GE compared to PGD, AT-PR consistently performs better. As Fig. 2 illustrates, although random AEs may fall within a short and wide peak (explaining their improved PR over PGD), they do not necessarily lie within the "widest peak" that AT-PR specifically targets. This distinction explains AT-PR's superior PR performance. + +# 5. Conclusion + +In this work, for the first time, we introduce a novel AT method targeting PR, named as AT-PR. By reformulating a new min-max optimization, we lay the theoretical foundation for AT-PR and develop specialized numerical algorithms to solve this optimization. Extensive experimental evaluation demonstrates that AT-PR consistently outperforms traditional AT methods designed for WCR in improving PR. Additionally, we show that AT-PR incurs a smaller trade-off in model generalization compared to common AT-WCR methods. + +In the future, we plan to address the challenges identified in [73], such as benchmarking PR and extending it to diverse generative AI tasks. + +Acknowledgements SK and XZ have received funding from the European Union's EU Framework Program for Research and Innovation Europe Horizon (grant agreement No 101202457). XZ's contribution is also supported by the UK EPSRC New Investigator Award [EP/Z536568/1] and NVIDIA Academic Grant Program. SK's contribution is supported by the UKRI Future Leaders Fellowship Grant [MR/S035176/1], and Innovate UK through the DriveSafeAI Project (10062594). XH has received funding from the European Union's Horizon Europe Research and Innovation Action under grant agreement No 101212818. YZ's contribution is supported by China Scholarship Council. We thank Zheng Wang for his valuable comments. + +Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them. + +# References + +[1] Amir Aminifar. Universal adversarial perturbations in epileptic seizure detection. In 2020 Int. Joint Conference on Neural Networks (IJCNN), pages 1-6. IEEE, 2020. 1 +[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Proc. of the 35th International Conference on Machine Learning, pages 274-283. PMLR, 2018. 1 +[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Proc. of the 30th Int. Joint Conference on Artificial Intelligence, pages 4312-4321, 2021. 1, 3 +[4] Mislav Balunović and Martin Vechev. Adversarial training and provable defenses: Bridging the gap. In 8th International Conference on Learning Representations (ICLR'20'20), 2020. 1, 3 +[5] Teodora Baluta, Zheng Leong Chua, Kuldeep S Meel, and Prateek Saxena. Scalable quantitative verification for deep neural networks. In ICSE'21, pages 312-323, 2021. 1 +[6] Qi-Zhi Cai, Chang Liu, and Dawn Song. Curriculum adversarial training. In Proc. of the 27th International Joint Conference on Artificial Intelligence, page 3740-3747. AAAI Press, 2018. 3 +[7] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE symposium on security and privacy (sp), pages 39-57, 2017. 1 +[8] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. A survey on adversarial attacks and defences. CAAI Transactions on Intelligence Technology, 6(1):25-45, 2021. 1 +[9] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In IEEE symposium on security and privacy (SP), pages 1277-1294. IEEE, 2020. 4, 12 + +[10] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. ZOO: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proc. of the 10th ACM Workshop on Artificial Intelligence and Security, page 15-26. ACM, 2017. 2 +[11] Zhen Chen, Yi Zhang, Fu Wang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. TARP-VP: Towards evaluation of transferred adversarial robustness and privacy on label mapping visual prompting models. Advances in Neural Information Processing Systems, 37:6776-6796, 2024. 1 +[12] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In International Conference on Machine Learning, pages 1310-1320. PMLR, 2019. 1 +[13] Nicolas Couellan. Probabilistic robustness estimates for feed-forward neural networks. Neural networks, 142:138-147, 2021. 1 +[14] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1 +[15] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2020. 3 +[16] Yi Dong, Wei Huang, Vibhav Bharti, Victoria Cox, Alec Banks, Sen Wang, Xingyu Zhao, Sven Schewe, and Xiaowei Huang. Reliability assessment and safety arguments for machine learning components in system assurance. ACM Transactions on Embedded Computing Systems, 22(3):1-48, 2023. 1, 2 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5 +[18] Chengbin Du, Yanxi Li, Zhongwei Qiu, and Chang Xu. Stable diffusion is unstable. Advances in Neural Information Processing Systems, 36, 2024. 1 +[19] Xiaoning Du, Xiaofei Xie, Yi Li, Lei Ma, Yang Liu, and Jianjun Zhao. Deepstellar: Model-based quantitative analysis of stateful deep learning systems. In Proc. of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 477–487, 2019. 1 +[20] Timon Gehr, Matthew Mirman, Dana Drachsler-Cohen, Petar Tsankov, Swarat Chaudhuri, and Martin Vechev. Ai2: Safety and robustness certification of neural networks with abstract interpretation. In IEEE symposium on security and privacy (SP), pages 3-18, 2018. 1 +[21] Ian J Goodfellow. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR'15), 2015. 1, 2, 3 +[22] Sven Gowal, Krishnamurthy Dj Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja + +Arandjelovic, Timothy Mann, and Pushmeet Kohli. Scalable verified training for provably robust image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4842-4851, 2019. 1, 3 +[23] Wei Huang, Youcheng Sun, Xingyu Zhao, James Sharp, Wenjie Ruan, Jie Meng, and Xiaowei Huang. Coverage-Guided Testing for Recurrent Neural Networks. IEEE Transactions on Reliability, 71(3):1191-1206, 2022. 1 +[24] Wei Huang, Xingyu Zhao, Alec Banks, Victoria Cox, and Xiaowei Huang. Hierarchical distribution-aware testing of deep learning. ACM Transactions on Software Engineering and Methodology, 33(2):1-35, 2023. 1 +[25] Wei Huang, Xingyu Zhao, Gaojie Jin, and Xiaowei Huang. Safari: Versatile and efficient evaluations for robustness of interpretability. In IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 1988-1998, 2023. 3 +[26] Xiaowei Huang, Marta Kwiatkowska, Sen Wang, and Min Wu. Safety verification of deep neural networks. In Computer Aided Verification: 29th International Conference, CAV 2017, Heidelberg, Germany, July 24-28, 2017, Proceedings, Part I 30, pages 3-29. Springer, 2017. 1 +[27] Xiaowei Huang, Daniel Kroening, et al. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 37:100270, 2020. 2, 3 +[28] Xiaowei Huang, Wenjie Ruan, et al. A survey of safety and trustworthiness of large language models through the lens of verification and validation. Artificial Intelligence Review, 57 (7):175, 2024. 1 +[29] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. Adversarial logit pairing. arXiv preprint arXiv:1803.06373, 2018. 3, 5, 12 +[30] Guy Katz, Clark Barrett, David L Dill, Kyle Julian, and Mykel J Kochenderfer. Reluplex: An efficient smt solver for verifying deep neural networks. In 29th Int. Conf. Computer Aided Verification, pages 97-117. Springer, 2017. 1 +[31] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In Proc. of the IEEE/CVF conference on computer vision and pattern recognition, pages 1221-1230, 2020. 4 +[32] Han Liu, Yuhao Wu, Shixuan Zhai, Bo Yuan, and Ning Zhang. Riatig: Reliable and imperceptible adversarial text-to-image generation with natural prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20585-20594, 2023. 1 +[33] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Discovering failure modes of text-guided diffusion models via adversarial search. In The 12th International Conference on Learning Representations, 2024. 1 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5 +[35] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deepauge: Multi-granularity testing criteria for deep + +learning systems. In Proce. of the 33rd ACM/IEEE Int. Conference on Automated Software Engineering (ASE'18), pages 120-131, 2018. 1 +[36] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 4, 12 +[37] Mark Huasong Meng, Guangdong Bai, Sin Gee Teo, Zhe Hou, Yan Xiao, Yun Lin, and Jin Song Dong. Adversarial robustness of deep neural networks: A survey from a formal verification perspective. IEEE Transactions on Dependable and Secure Computing, 2022. 1 +[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 1 +[39] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 1 +[40] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1 +[41] Mikhail Pautov, Nurislam Tursynbek, Marina Munkhoeva, Nikita Muravev, Aleksandr Petiushko, and Ivan Oseledets. Cc-cert: A probabilistic approach to certify general robustness of neural networks. In AAAI'22, pages 7975-7983, 2022. 1 +[42] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepexplore: Automated whitebox testing of deep learning systems. In Proceedings of the 26th Symposium on Operating Systems Principles, pages 1-18. ACM, 2017. 1 +[43] Alexander Robey, Luiz Chamon, George J Pappas, and Hamed Hassani. Probabilistically robust learning: Balancing average and worst-case performance. In Int. Conference on Machine Learning, pages 18667-18686. PMLR, 2022. 3 +[44] Wenjie Ruan, Xiaowei Huang, and Marta Kwiatkowska. Reachability analysis of deep neural networks with provable guarantees. In Proc. of the 27th Int. Joint Conference on Artificial Intelligence (IJCAI'18), pages 2651-2659, 2018. 1 +[45] Sebastian Ruder. An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747, 2016. 12 +[46] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 5, 6 +[47] Uri Shaham, Yutaro Yamada, and Sahand Negahban. Understanding adversarial training: Increasing local stability of supervised models through robust optimization. Neurocomputing, 307:195-204, 2018. 3 +[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In In Proc. of 2nd Int. Conf. on Learning Representations, 2014. 1 + +[49] Karim Tit, Teddy Furon, and Mathias Rousset. Efficient statistical assessment of neural network corruption robustness. NeurIPS'21, 34:9253-9263, 2021. 1 +[50] Karim TIT, Teddy Furon, and Mathias Rousset. Gradient-informed neural network statistical robustness estimation. In Proc. of The 26th Int. Conf. on Artificial Intelligence and Statistics, pages 323-334. PMLR, 2023. 1, 3 +[51] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5 +[52] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 1, 3 +[53] Princeton University. Lecture 11: Rademacher complexity, 2014. Accessed: 2025-01-05. 7 +[54] Paul Viallard, Eric Guillaume VIDOT, Amaury Habrard, and Emilie Morvant. A pac-bayes analysis of adversarial robustness. Advances in Neural Information Processing Systems, 34:14421-14433, 2021. 6, 12 +[55] Benjie Wang, Stefan Webb, and Tom Rainforth. Statistically robust neural network classification. In Uncertainty in Artificial Intelligence, pages 1735-1745. PMLR, 2021. 1, 3 +[56] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. In ICML'19, pages 6586-6595. PMLR, 2019. 1 +[57] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 3, 5, 12 +[58] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 3 +[59] Stefan Webb, Tom Rainforth, Yee Whye Teh, and M Pawan Kumar. A statistical approach to assessing neural network robustness. In Int. Conf. on Learning Representations, 2019, 1, 2, 3, 5 +[60] Lily Weng, Pin-Yu Chen, Lam Nguyen, Mark Squillante, Akhilan Boopathy, Ivan Oseledets, and Luca Daniel. Proven: Verifying robustness of neural networks with a probabilistic approach. In Int. Conf. on Machine Learning, pages 6727-6736. PMLR, 2019. 1 +[61] Tsui Wei Weng, Huan Zhang, Pin-Yu Chen, Jinfeng Yi, Dong Su, Yupeng Gao, Cho-Jui Hsieh, and Luca Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. In Int. Conf. on Learning Representations, 2018. 1 +[62] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 5, 6 +[63] Weiming Xiang, Hoang-Dung Tran, and Taylor T. Johnson. Output reachable set estimation and verification for multilayer neural networks. IEEE Tran. on Neural Networks and Learning Systems, 29(11):5777-5783, 2018. 1 + +[64] Xiangyu Yin, Sihao Wu, Jiaxu Liu, Meng Fang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. Representation-based robustness in goal-conditioned reinforcement learning. In Proc. of the AAAI Conference on Artificial Intelligence, pages 21761-21769, 2024. 1 +[65] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE Tran. on neural networks and learning systems, 30(9):2805-2824, 2019. 1 +[66] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3, 5, 6, 12 +[67] Jingfeng Zhang, Xilie Xu, Bo Han, Gang Niu, Lizhen Cui, Masashi Sugiyama, and Mohan Kankanhalli. Attacks which do not kill training make adversarial learning stronger. In International conference on machine learning, pages 11278-11287. PMLR, 2020. 3 +[68] Jiawei Zhang, Linyi Li, Huichen Li, Xiaolu Zhang, Shuang Yang, and Bo Li. Progressive-scale boundary blackbox attack via projective gradient estimation. In Int. Conference on Machine Learning, pages 12479-12490. PMLR, 2021. 4 +[69] Tianle Zhang, Wenjie Ruan, and Jonathan E. Fieldsend. Proa: A probabilistic robustness assessment against functional perturbations. In Machine Learning and Knowledge Discovery in Databases (ECML/PKDD'22), page 154-170. Springer-Verlag, 2023. 1, 2, 3 +[70] Tianle Zhang, Yanghao Zhang, Ronghui Mu, Jiaxu Liu, Jonathan Fieldsend, and Wenjie Ruan. Prass: probabilistic risk-averse robust learning with stochastic search. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence, pages 559–567, 2024. 3 +[71] Yi Zhang, Zhen Chen, Chih-Hong Cheng, Wenjie Ruan, Xiaowei Huang, Dezong Zhao, David Flynn, Siddartha Khastgir, and Xingyu Zhao. Trustworthy text-to-image diffusion models: A timely and focused survey. arXiv preprint arXiv:2409.18214, 2024. 1 +[72] Yi Zhang, Yun Tang, Wenjie Ruan, Xiaowei Huang, Siddartha Khastgir, Paul Jennings, and Xingyu Zhao. Protip: Probabilistic robustness verification on text-to-image diffusion models against stochastic perturbation. In European Conference on Computer Vision, 2024. 1, 2, 3 +[73] Xingyu Zhao. Probabilistic robustness in deep learning: A concise yet comprehensive guide. Adversarial Example Detection and Mitigation Using Machine Learning, pages 1-13, 2025. 1, 2, 3, 8 +[74] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2385-2392, 2023. 1 +[75] Monty-Maximilian Zuhlke and Daniel Kudenko. Adversarial robustness of neural networks from the perspective of Lipschitz calculus: A survey. ACM Computing Surveys, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Adversarial Training for Probabilistic Robustness/images.zip b/2025/Adversarial Training for Probabilistic Robustness/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b1c5ce951374956406db35282b009ac9d8b6c395 --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:649f688969fc9596965777f2f539c3311b503ec461a78f1681f113826fe0b576 +size 1015385 diff --git a/2025/Adversarial Training for Probabilistic Robustness/layout.json b/2025/Adversarial Training for Probabilistic Robustness/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2a09db7ab02e9bc5556d341faf6586ee7d02b9d --- /dev/null +++ b/2025/Adversarial Training for Probabilistic Robustness/layout.json @@ -0,0 +1,11136 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 460, + 121 + ], + "type": "text", + "content": "Adversarial Training for Probabilistic Robustness" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "spans": [ + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": "Yi Zhang" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Yuhang Chen" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Zhen Chen" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Wenjie Ruan" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Xiaowei Huang" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Siddartha Khastgir" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": ", Xingyu Zhao" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": "WMG, University of Warwick " + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 142, + 499, + 213 + ], + "type": "text", + "content": "University of Liverpool {yi.zhang.16,yuhang.chen.1,s.khastgir.1,xingyu.zhao}@warwick.ac.uk; {zhen.chen2,xiaowei.huang}@liverpool.ac.uk; w.ruan@trustai.uk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 541 + ], + "type": "text", + "content": "Deep learning (DL) has shown transformative potential across industries, yet its sensitivity to adversarial examples (AEs) limits its reliability and broader deployment. Research on DL robustness has developed various techniques, with adversarial training (AT) established as a leading approach to counter AEs. Traditional AT focuses on worst-case robustness (WCR), but recent work has introduced probabilistic robustness (PR), which evaluates the likelihood of AEs within a local perturbation range, providing an overall assessment of the model's robustness and acknowledging residual risks that are more practical to manage. However, existing AT methods are fundamentally designed to improve WCR, and no dedicated methods currently target PR. To bridge this gap, we formulate a new min-max optimization as the theoretical foundation for PR-focused AT, and introduce an AT-PR training scheme with numerical algorithms to solve the new optimization problem. Our experiments, based on 70 DL models trained on common datasets and diverse architectures, demonstrate that: i) AT-PR achieves higher improvements in PR than AT-WCR methods; ii) it shows more consistent effectiveness across varying local inputs; iii) it exhibits a reduced trade-off in model's generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Deep Learning (DL) has shown potential for transformative impacts across many industries including transportation, healthcare and energy. The reality that DL is not as dependable as required now becomes a major impediment to their broader deployment. Since the first discovery of the phenomenon of Adversarial Examples (AEs) for DL [21, 39, 48], robustness is arguably the property in the limelight, encompassing research topics like adversarial attacks in malicious contexts [7, 8, 14, 21, 36, 64, 65, 75]," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 554, + 350 + ], + "type": "text", + "content": "AE detection during testing [19, 23, 24, 35, 42], verification with formal guarantees [20, 26, 30, 37, 44], and various robustness improvement techniques, including distillation [40], randomized smoothing [12], gradient masking [2], and the more widely used adversarial training (AT) [4, 21, 22, 36, 52]. As large language models and multimodal foundation models gain popularity, more recent research has increasingly focused on understanding their robustness as well [11, 18, 28, 32, 33, 58, 71, 72, 74]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "type": "text", + "content": "Regardless of specific DL tasks or model architectures, robustness broadly refers to a model's ability to maintain consistent decisions under small input perturbations, with local robustness focusing on stability around individual inputs, as opposed to global robustness across the entire input distribution. More specifically, when evaluating local robustness, common problem formulations include [16, 72]: i) a binary question, \"does an AE exist within a specified perturbation norm-ball?\" [20, 30, 44, 63]; ii) a maximum safe distance question, \"what is the largest perturbation radius within which no AE exists?\" [1, 38, 60, 61]; and iii) a maximum prediction loss question, \"which AE within the norm-ball yields the highest prediction loss?\", which is often applied in AT to enhance robustness [36, 56]. It is evident that the three formulations focus on extreme scenarios, where local robustness is evaluated based on the existence of a deterministic AE, reflecting a worst-case perspective." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 556, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 675 + ], + "type": "text", + "content": "A more recent and distinct view, however, considers probabilistic robustness (PR), employing statistical approaches to its evaluation [5, 13, 41, 49, 50, 55, 59, 60, 69, 72], it answers the question of \"what is the likelihood of AEs in the given perturbation norm-ball?\" [73]. However, existing AT methods are primarily designed for WCR via the well-established \"min-max\" optimization [3, 21, 36] (cf. Sec. 2.2), and are thus referred to as \"AT-WCR\" in this paper. No dedicated AT methods exist for enhancing PR, which motivates our work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we first formulate AT for PR as a new min-max optimization problem, and then devise AT-PR training scheme based on numerical algorithms to effectively" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 283, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 283, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 283, + 713 + ], + "type": "text", + "content": "*Corresponding to X. Zhao, xingyu.zhao@warwick.ac.uk." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1675" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "solve this training optimization. Our experiments on 70 DL models trained on common datasets and models demonstrate key strengths of AT-PR over AT-WCR: " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " while traditional AT-WCR can also improve PR, our new AT-PR training scheme is more effective, especially when enhancing PR becomes challenging for traditional AT-WCR; " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " AT-PR not only outperforms AT-WCR in improving PR on average, but also offers more stable and consistent performance across different local inputs; " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "iii" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " AT-PR has less negative impact on model's generalization compared to AT-WCR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 193, + 279, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 193, + 279, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 193, + 279, + 205 + ], + "type": "text", + "content": "In summary, key contributions of this paper include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 350 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 229 + ], + "type": "text", + "content": "- Problem formulation: We lay the foundation of AT for PR by formulating a new min-max optimization problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "type": "text", + "content": "- Numerical algorithms: We develop effective numerical algorithms to solve the AT-PR min-max optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 253, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 312 + ], + "type": "text", + "content": "- Empirical evaluation: We validate the effectiveness of AT-PR through extensive experiments, showing its superior improvement in PR, more consistent effectiveness across different local inputs, and reduced impact on generalization compared to AT-WCR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 313, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 296, + 350 + ], + "type": "text", + "content": "- Open-source repository: A repository is provided at https://github.com/wellzline/AT-PR with code, datasets, models, and all experiment details." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 361, + 241, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 241, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 241, + 374 + ], + "type": "text", + "content": "2. Preliminaries and Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 382, + 263, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 263, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 263, + 395 + ], + "type": "text", + "content": "2.1. Worst-Case vs Probabilistic Robustness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "Although definitions of robustness vary across different DL tasks and model types, it generally refers to a DL model's ability to maintain consistent predictions despite small input perturbations. Typically it is defined as all inputs in a region " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " have the same prediction, where " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " is a small norm ball (in a " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "-norm distance) of radius " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " around an input " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": ". A perturbed input (e.g., by adding noise on " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": ") " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}'" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " is an AE if its prediction label differs from " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 498, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 652 + ], + "type": "text", + "content": "Evaluation methods for robustness are typically based on metrics defined around AEs [27, 75]. As discussed earlier in Sec. 1, four types of robustness metrics, as summarized in [16, 72], address: (1) a binary question, (2) a maximum safe distance question, (3) a maximum prediction loss question (Fig. 1 (a)), and (4) a statistical inference question estimating the likelihood of AEs within a given perturbation norm-ball (Fig. 1 (b)). The first three metrics focus on extreme scenarios, where local robustness is assessed based on the existence of a deterministic AE, reflecting a worst-case perspective referred to as WCR. In contrast, the fourth approach adopts a probabilistic view, evaluating overall local robustness in the presence of AEs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "PR is arguably more practical than WCR, as extreme-case focus is often unnecessary [73]. Instead, knowing the likelihood of AEs is more relevant for applications, where it is sufficient to show that the risk of AE remains below an acceptable threshold rather than exactly zero [59, 69, 72]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 121 + ], + "type": "text", + "content": "Inspired by [72], we illustrate the difference between the maximum loss WCR and PR in Fig. 1, omitting the binary and maximum safe distance cases as they are less relevant to the AT focus of this paper." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 129, + 434, + 228 + ], + "blocks": [ + { + "bbox": [ + 316, + 129, + 434, + 228 + ], + "lines": [ + { + "bbox": [ + 316, + 129, + 434, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 434, + 228 + ], + "type": "image", + "image_path": "697ef97724b8801d9b0c18d3955ce50176773663612f33571c447d95f526fad7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 350, + 230, + 365, + 237 + ], + "lines": [ + { + "bbox": [ + 350, + 230, + 365, + 237 + ], + "spans": [ + { + "bbox": [ + 350, + 230, + 365, + 237 + ], + "type": "text", + "content": "Class A" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 331, + 230, + 348, + 237 + ], + "blocks": [ + { + "bbox": [ + 331, + 230, + 348, + 237 + ], + "lines": [ + { + "bbox": [ + 331, + 230, + 348, + 237 + ], + "spans": [ + { + "bbox": [ + 331, + 230, + 348, + 237 + ], + "type": "image", + "image_path": "7ea8f056e35a19db73b4dea3e51e56a5cf261303c81d35f1aacd72659f419106.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 249, + 538, + 261 + ], + "lines": [ + { + "bbox": [ + 329, + 249, + 538, + 261 + ], + "spans": [ + { + "bbox": [ + 329, + 249, + 538, + 261 + ], + "type": "text", + "content": "Figure 1. Worst-case (a) vs. Probabilistic (b) Robustness." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 381, + 230, + 405, + 237 + ], + "blocks": [ + { + "bbox": [ + 381, + 230, + 405, + 237 + ], + "lines": [ + { + "bbox": [ + 381, + 230, + 405, + 237 + ], + "spans": [ + { + "bbox": [ + 381, + 230, + 405, + 237 + ], + "type": "image", + "image_path": "705aca3d16f6fe087b1984d786ea29dbb9960ec3ffa2043cd949721350f58b5c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 230, + 421, + 237 + ], + "lines": [ + { + "bbox": [ + 410, + 230, + 421, + 237 + ], + "spans": [ + { + "bbox": [ + 410, + 230, + 421, + 237 + ], + "type": "text", + "content": "#" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 441, + 129, + 552, + 228 + ], + "blocks": [ + { + "bbox": [ + 441, + 129, + 552, + 228 + ], + "lines": [ + { + "bbox": [ + 441, + 129, + 552, + 228 + ], + "spans": [ + { + "bbox": [ + 441, + 129, + 552, + 228 + ], + "type": "image", + "image_path": "c10ef37e22f6aba6edf80d440dd76afbc145614a6e3a7a51de6935727acb24c8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 230, + 523, + 236 + ], + "lines": [ + { + "bbox": [ + 505, + 230, + 523, + 236 + ], + "spans": [ + { + "bbox": [ + 505, + 230, + 523, + 236 + ], + "type": "text", + "content": "Non-AE" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 475, + 230, + 481, + 236 + ], + "blocks": [ + { + "bbox": [ + 475, + 230, + 481, + 236 + ], + "lines": [ + { + "bbox": [ + 475, + 230, + 481, + 236 + ], + "spans": [ + { + "bbox": [ + 475, + 230, + 481, + 236 + ], + "type": "image", + "image_path": "6d162242fbfe60e28f0b56daa25a7761ca8c42b90a4b5728025c344954358ec6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 274, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 322 + ], + "type": "text", + "content": "To evaluate the WCR, we normally formulate it as a question of maximizing the prediction loss, cf. Fig. 1 (a), where adversarial attacks are typically introduced to find such a worst-case AE with perturbation:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 371, + 331, + 555, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 331, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 371, + 331, + 555, + 350 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} = \\arg \\max _ {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}), \\tag {1}", + "image_path": "da9f3f58d4221d079262f0edb5dc2f83d9cab089c28ba0308f53f5ad81ce6082.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": " is a small perturbation constrained within an " + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": "-norm ball of radius " + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": ". This optimization causes the model to misclassify " + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}^{\\prime} = \\boldsymbol{x} + \\delta" + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": " as a different label " + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "inline_equation", + "content": "y^\\prime \\neq y" + }, + { + "bbox": [ + 313, + 358, + 555, + 467 + ], + "type": "text", + "content": ". There are some popular attacks following this formulation, including FGSM [21], PGD [36], and ZOO [10], among others. Following common practice, we measure WCR by the success rate of some selected attacks on a test dataset. Estimation of PR is normally formalized as the statistical inference problem on the following metric:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "content": "Definition 1 (Probabilistic Robustness) For a DL classifier " + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "content": " that takes input " + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "content": " and returns a prediction label, the PR of an input " + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "content": " in a norm ball of radius " + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 474, + 554, + 511 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 338, + 519, + 555, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 519, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 338, + 519, + 555, + 542 + ], + "type": "interline_equation", + "content": "P R (\\boldsymbol {x}, \\gamma) = \\underset {\\| \\boldsymbol {\\delta} \\| \\leq \\gamma} {\\mathbb {E}} _ {\\delta \\sim P r (\\cdot | \\boldsymbol {x})} [ I _ {\\{f _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) = y \\}} (\\boldsymbol {x} + \\boldsymbol {\\delta}) ], \\tag {2}", + "image_path": "6bd808ace8d0ffd323c97ba21bf2e9bee75bf285798887e23068542b48700768.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "inline_equation", + "content": "I_{S}(\\pmb{x})" + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "content": " is an indicator function—it is equal to 1 when " + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "content": " is true and 0 otherwise; " + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "inline_equation", + "content": "Pr(\\cdot)" + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "content": " is the local distribution of inputs representing how perturbations " + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 550, + 554, + 598 + ], + "type": "text", + "content": " are generated, which is precisely the \"input model\" used by [59, 72]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": "Def. 1 suggests PR is the probability that the model prediction remains unchanged from a random perturbation " + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 605, + 554, + 689 + ], + "type": "text", + "content": ". A \"frequentist\" interpretation of this probability is: it is the limiting relative frequency of perturbations where the output label is preserved, in an infinite sequence of independently generated perturbations [72]. That is, the \"proportion\" of non-AEs in such infinite set of perturbations." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "The work [59] is the first to define PR and introduces a black-box statistical estimator, particularly useful when" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1676" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "PR is extremely high (i.e., detecting AEs is a rare event). A more efficient white-box statistical estimator for PR was later proposed in [50]. Huang et al. [25] first extend PR to explainable AI (XAI). Zhang et al. [69] explore PR under functional perturbations, while the recent study [72] is the first to estimate PR in the text-to-image DL task. The recent work [73] provides an overview of PR, including its definition, estimation, enhancement and system-level integration. Without loss of generality, we adopt the estimator in [59] to estimate PR in later experiments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 178, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 178, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 178, + 211 + ], + "type": "text", + "content": "2.2. Adversarial Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "content": "AT [4, 21, 22, 36, 52] is a widely used empirical approach for improving robustness, typically formulated as a \"min-max\" optimization problem. It introduces a set of input perturbations " + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "content": " within a norm-ball " + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "content": " of radius " + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "content": ", designed to maximize the training loss " + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 216, + 296, + 348 + ], + "type": "text", + "content": ". The model is then trained with these perturbations by minimizing the overall loss of the training dataset. These perturbations are often approximated using empirical adversarial attacks, such as FGSM [21] and PGD [36] due to their high efficiency. Taking a classification task as an example, the min-max optimization target of AT can be formalised as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": "Definition 2 (AT for WCR) For a DL model " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": ", trained on a dataset " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": ", consisting of pairs " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "(\\pmb{x}, y)" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " is the input and " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " denotes the label, with loss function " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\pmb{x}, y; \\theta)" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": ", the objective of AT-WCR can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 410, + 295, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 410, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 94, + 410, + 295, + 439 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma} \\mathcal {L} (\\boldsymbol {x} + \\boldsymbol {\\delta}, y; \\boldsymbol {\\theta}) \\right], \\tag {3}", + "image_path": "92b2b3dc12d98f2e2c32172ee1daeb41f236f7ae8f655cb43af2d4ad6e17a774.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "text", + "content": " represents the perturbation that maximizes classification loss " + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "text", + "content": ", while the outer minimization aims to train " + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 54, + 443, + 295, + 480 + ], + "type": "text", + "content": " to be robust against such adversarial perturbations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 641 + ], + "type": "text", + "content": "The inner maximization is precisely the adversarial attack formulation outlined in Eq. (1). As highlighted in [3], the effectiveness of AT largely depends on identifying strong AEs within this inner maximization. Notably, the PGD attack [36] has become a critical benchmark and is widely regarded as the standard method for AT in practice [3]. Accordingly, we also use PGD in later experiments. Based on the standard min-max optimization, various AT variants have been proposed to achieve more stable and efficient training, such as [6, 15, 29, 47, 52, 57, 58, 66, 67]. Comprehensive overviews of recent AT variants can be found in surveys [3, 27], all of which fundamentally rely on the min-max problem for WCR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "In contrast, to the best of our knowledge, only three works [43, 55, 70] have explicitly aimed at improving PR [73]. Although Wang et al. [55] demonstrate that simply adding randomly perturbed inputs significantly enhances PR, the resulting AR is nearly zero. The methods of [43] and [70] propose loss functions based on common risk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "measures (i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "-essential supremum, Conditional Value-at-Risk, and Entropic Value-at-Risk) to train models to minimize these risks. However, again, these approaches provide little to no improvement in AR. We argue that, akin to the well-established AT framework for AR, a dedicated AT framework for PR is necessary—one that does not compromise AR performance. Accordingly, our proposed AT-PR is the first to explicitly reformulate the min-max optimization principle for PR, while also providing comparable AR performance through practical training algorithms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 201, + 446, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 446, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 446, + 215 + ], + "type": "text", + "content": "3. The Approach: AT-PR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 440, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 440, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 440, + 232 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 237, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 441 + ], + "type": "text", + "content": "Intuitively, the goal of AT-WCR in Def. 2 is to identify the AE that maximizes model loss and incorporate it into AT to lower the maximum loss, cf. the top-middle illustration in Fig. 2. However, to effectively improve local PR, we aim to locate the \"widest peak\"—the region with the highest proportion of AEs in the loss landscape (cf. the top-right illustration in Fig. 2), rather than the tallest peak. Adding the AE from this peak during AT may ideally reduce the loss for AEs around that area, thereby lowering the overall AE proportion (cf. the intuitive interpretation of PR in Def. 1). In some cases, AT-WCR and AT-PR may identify the same AE for training when the tallest peak also happens to be the widest, as shown in the second row of Fig. 2. Notably, the two cases illustrated in Fig. 2 representatively cover all possible scenarios, though additional peaks may exist in reality. Given the difference between the goals of improving WCR and PR, we formulate a new min-max problem for AT-PR:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "text", + "content": "Definition 3 (AT for PR) For a DL model " + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "text", + "content": ", trained on a dataset " + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "text", + "content": " consisting of pairs " + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 313, + 446, + 555, + 483 + ], + "type": "text", + "content": ", the objective of AT-PR can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 355, + 487, + 554, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 487, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 355, + 487, + 554, + 514 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta}} \\mathbb {E} _ {(\\boldsymbol {x}, y) \\sim \\mathcal {D}} \\left[ \\max _ {\\| \\delta \\| \\leq \\gamma , P R (\\boldsymbol {x} + \\delta , k) = 0} k \\right], \\tag {4}", + "image_path": "a99e940bf5b1a0ba7c772abfcb6692f94fd928e74dcb6c0bd6bd652427aa0486.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " are the variables in the new inner max optimization, " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "PR(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " is the metric defined in Eq. (2). As shown in Fig. 3, the intuition behind the maximization is to find an optimal " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "\\delta^*" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " that produces an " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "AE" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "x' = x + \\delta^*" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " and to maximize the radius " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " of a smaller norm-ball centered on " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " such that the " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "PR" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": " within this " + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 517, + 554, + 601 + ], + "type": "text", + "content": "-sized norm-ball is 0 (i.e., all inputs within this smaller norm-ball are AEs)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": "The high non-convexity of DL models makes it intractable to analytically solve both the established min-max optimization in Eq. (3) and our new formulation in Eq. (4). Thus, similar to existing AT-WCR, we propose algorithms to numerically solve Eq. (4) in AT-PR in the next section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 451, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 451, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 451, + 685 + ], + "type": "text", + "content": "3.2. Optimization Algorithm" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "While the outer minimization in Eq. (4) can be effectively handled with standard DL training methods as in AT-WCR," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1677" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 218, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 218, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 218, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 218, + 170 + ], + "type": "image", + "image_path": "45e4646f4632831e825b287d49e36f3ff221b061eaaeee41001d36066dfec770.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 386, + 170 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 386, + 170 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 386, + 170 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 386, + 170 + ], + "type": "image", + "image_path": "d0bc69b4220e5a18c1848fbc1a91af8d0edd253f3ae476d3f8ff863131525636.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 553, + 171 + ], + "type": "image", + "image_path": "f2638fb90f667e4c111681f00ca826b384bdb6362f1cdc3095aa6e0bb95d0d13.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 171, + 219, + 269 + ], + "blocks": [ + { + "bbox": [ + 57, + 171, + 219, + 269 + ], + "lines": [ + { + "bbox": [ + 57, + 171, + 219, + 269 + ], + "spans": [ + { + "bbox": [ + 57, + 171, + 219, + 269 + ], + "type": "image", + "image_path": "e9125c9065c840c5551d6c0e4c3ad46f6fa230015a81219c64d1b96cbb5bff4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 278, + 493, + 291 + ], + "lines": [ + { + "bbox": [ + 116, + 278, + 493, + 291 + ], + "spans": [ + { + "bbox": [ + 116, + 278, + 493, + 291 + ], + "type": "text", + "content": "Figure 2. Two cases of the local loss landscape and their AEs identified by AT-WCR and AT-PR for AT." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 171, + 400, + 269 + ], + "blocks": [ + { + "bbox": [ + 225, + 171, + 400, + 269 + ], + "lines": [ + { + "bbox": [ + 225, + 171, + 400, + 269 + ], + "spans": [ + { + "bbox": [ + 225, + 171, + 400, + 269 + ], + "type": "image", + "image_path": "7f4eac3de2261e7295b0c44f242116eed48d38466123235dd0f9994e2e5a2d98.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 434, + 182, + 550, + 260 + ], + "blocks": [ + { + "bbox": [ + 434, + 182, + 550, + 260 + ], + "lines": [ + { + "bbox": [ + 434, + 182, + 550, + 260 + ], + "spans": [ + { + "bbox": [ + 434, + 182, + 550, + 260 + ], + "type": "image", + "image_path": "8239a33d05b6f04482cc45ce6cf6271c4340114e5088b746541ecf33e6e4a05c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 312, + 291, + 438 + ], + "blocks": [ + { + "bbox": [ + 56, + 312, + 291, + 438 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 291, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 291, + 438 + ], + "type": "image", + "image_path": "0b1e3431703387c08f43c69b982e145e3aa4ad1c3362acf2805616a8f009fa88.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 445, + 295, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 458 + ], + "type": "text", + "content": "Figure 3. Illustration of the new inner-max optimization in Def. 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 518 + ], + "type": "text", + "content": "we devise the Algo. 1 to solve the inner optimization in our AT-PR. The gist of Algo. 1 is shown in Fig. 4, which essentially consists of 3 main steps:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 520, + 295, + 628 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": "- Step 1 (lines 1-9): Apply PGD attacks in different regions of the local norm-ball under study to find a diverse set of AEs at various local optima " + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 520, + 295, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 555, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 591 + ], + "type": "text", + "content": "- Step 2 (lines 10–20): For each locally optimal AE found in Step 1, perform a gradient descent search to locate the nearest decision boundary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": "- Step 3 (lines 21-25): Calculate the distance " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " from each " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " to its respective decision boundary and return the AE with the largest " + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 592, + 295, + 628 + ], + "type": "text", + "content": " value." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Similar to AT-WCR, the returned AE will be added into the training dataset in AT-PR. We note, as an numerical algorithm to the optimization, its effectiveness may depends on certain assumptions include: " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": ") the PGD attack in Step 1 may effectively find local optima by adjusting hyperparameters such as attack steps and learning rate, which aligns with the insight of \"PGD will converge to local maxima\"" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "type": "text", + "content": "[36]; ii) The gradient descent based boundary search with adaptive hyperparameters may efficiently find the decision boundary as used in some boundary attacks [9, 31, 68]. We also provide an initial theoretical analysis of the effectiveness and convergence properties of the proposed optimization algorithm in Appendix A." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 393, + 465, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 465, + 429 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 465, + 429 + ], + "type": "text", + "content": "Step 1: Apply PGD in different regions of the norm-ball to get a set of diverse AEs at different local optimal points " + }, + { + "bbox": [ + 317, + 393, + 465, + 429 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 317, + 393, + 465, + 429 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 432, + 553, + 570 + ], + "blocks": [ + { + "bbox": [ + 316, + 432, + 553, + 570 + ], + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 570 + ], + "type": "image", + "image_path": "86c16682dcb38119d2433013afa9f2335d278cd739a57b3ec64c40fc4d9d3671.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 577, + 528, + 589 + ], + "lines": [ + { + "bbox": [ + 339, + 577, + 528, + 589 + ], + "spans": [ + { + "bbox": [ + 339, + 577, + 528, + 589 + ], + "type": "text", + "content": "Figure 4. Illustration of the 3 main steps of Algo. 1." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": "We analyze the time complexity of the proposed algorithm. There are several computational steps to be considered with the assumption that the dimension of the input " + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": ". First, sampling " + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": " and steps from uniform distributions (lines 5-6) is independent of the input and thus the time complexity is " + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "inline_equation", + "content": "O(1)" + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": ". Second, the PGD algorithm (line 7) consists of steps gradient calculations with complexity " + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "inline_equation", + "content": "O(\\text{grad})^1" + }, + { + "bbox": [ + 313, + 600, + 555, + 696 + ], + "type": "text", + "content": " and one projection calculation with complexity" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 555, + 714 + ], + "type": "text", + "content": "1The time complexity of gradient calculation depends on the model" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1678" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 71, + 297, + 475 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": "Algorithm 1 PGD and gradient-based search for " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{pr}^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " \nRequire: Inputs " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "[X,Y],N,\\alpha_{min},\\alpha_{max},step_{min},step_{max}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 1: Initialize " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "AE_{s}\\gets []" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Initialize AE candidate set 2: Apply PGD to get different AE candidates 3: for idx = 0 to N do 4: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{init}\\gets x + \\mathrm{Uniform}(-\\gamma ,\\gamma)" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 5: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\alpha \\leftarrow" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Uniform " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "(\\alpha_{min},\\alpha_{max})" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 6: steps " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Uniform " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "(step_{min},step_{max})" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 7: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{idx}^{\\prime}\\gets pgd(x_{init},y,\\gamma ,\\alpha ,steps)" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 8: Append " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{idx}^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "AE_{s}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 9: end for \nRequire: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "AE_s,x,y,C" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " \n10: Initialize Max_d " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\leftarrow 0" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{pr}^{\\prime}\\gets" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " None 11: for each " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x^{\\prime}\\in AE_{s}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " do 12: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x} = x^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 13: while iter " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "< C" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " do 14: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "y^\\prime = f(\\tilde{x})" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " 15: if " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "y^\\prime = y" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " then 16: break " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Exit if " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " is classified correctly 17: end if 18: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "g\\gets \\nabla_{\\tilde{x}}L(\\tilde{x},y)" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Compute gradient 19: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x} = \\tilde{x} -\\alpha \\cdot g" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Update " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " \n20: end while \n21: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "d\\gets D(\\tilde{x},x^{\\prime})" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Distance between " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{x}\\& x^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " \n22: if " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "d > Max\\_ d" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " then \n23: " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{pr}^{\\prime} = x^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " \n24: Max_d=d \n25: end if \n26: end for \n27: Output " + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "inline_equation", + "content": "x_{pr}^{\\prime}" + }, + { + "bbox": [ + 56, + 71, + 297, + 475 + ], + "type": "text", + "content": " Farthest AE from decision boundary." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ". Third, in the effort to locate the decision boundary (lines 18-19), one gradient calculation step is performed. Last, a " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " distance between " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " is estimated (line 21), with complexity " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ". In Step 1, the algorithm uses different random starts via random initialization [46] to ensure it finds different local optima within norm-ball " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " (line 4), and generates one AE (lines 5-8) for each combination of input " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ", radius " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ", randomly sampled " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " and steps. A total of " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " AEs are generated. Thus the complexity upper bound of Step 1 is then " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(N*(steps*O(grad)+O(n)))" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ". In Step 2, with the maximum iteration number " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " for the nearest decision boundary locating process of " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " AEs, the complexity upper bound of Step 2 is then " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(C*N*O(grad))" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ". In Step 3, the complexity of distance estimation is " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": " as discussed above. Assuming " + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "inline_equation", + "content": "O(grad) \\gg O(n)" + }, + { + "bbox": [ + 55, + 497, + 296, + 677 + ], + "type": "text", + "content": ", which" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 715 + ], + "type": "text", + "content": "architecture and the specific loss function employed during training. E.g., DL models with more layers or complex loss functions may require more computational resources for gradient computation due to backpropagation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "typically holds in DL scenarios, the complexity simplifies to " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "O(N * (steps + C) \\cdot O(grad))" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ". As analyzed, the computational cost of AT-PR primarily comes from locating the optimal AE. However, it remains compatible with other techniques designed to enhance AT efficiency, such as [46, 62]. We provide comparative empirical results by replacing PGD with Fast-AT [62] to demonstrate such compatibility in Table 3, refer to Sec. 4 for details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 198, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 283 + ], + "type": "text", + "content": "We train a total of 70 DL models on four datasets: CIFAR-10, CIFAR-100, SVHN, and Tiny/ImageNet, using models such as ResNet, Wide-ResNet (WRN), Vision Transformer (ViT) [17, 34], and DeiT [51]. To evaluate AT-PR, we compare it with two standard AT-WCR (PGD and FGSM) and four advanced methods: TRADES [66], MART [57], ALP, and CLP [29]. Implementation details are in Appendix B." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 514, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 514, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 514, + 302 + ], + "type": "text", + "content": "4.1. Effectiveness and Efficiency of AT-PR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "text", + "content": "We assess AT-PR's effectiveness against AT-WCR from two perspectives: 1) improvement in PR and 2) impact on generalization. We compare models trained with AT-PR, AT-WCR (FGSM and PGD) and a baseline without AT, using three metrics: (i) Accuracy on clean test data; (ii) Adversarial Accuracy under PGD, C&W, and AA attacks to reflect WCR; and (iii) PR, computed as " + }, + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "inline_equation", + "content": "PR(x, \\gamma)" + }, + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "text", + "content": " for various " + }, + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 307, + 555, + 427 + ], + "type": "text", + "content": " using the estimator from [59]. We report the mean and standard deviation of PR on the test set, along with the generalization error (GE) to measure train-test robustness gaps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 427, + 487, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 427, + 487, + 437 + ], + "spans": [ + { + "bbox": [ + 326, + 427, + 487, + 437 + ], + "type": "text", + "content": "Key findings summarized from Table 1:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 438, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 521 + ], + "type": "text", + "content": "i) All original models without AT show low WCR. As expected, PGD is the most effective in improving WCR, though it compromises generalization the most. AT-PR also enhances WCR, outperforming FGSM and, in some cases, even surpassing PGD, as seen with SVHN/ResNet-18, Tiny-ImageNet/ResNet-34 for PGD.Ac., and SVHN/WRN50-2, Tiny-ImageNet/ResNet-18, ResNet-34 for CW.Ac." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "text", + "content": "ii) Both AT-WCR with FGSM and PGD can improve PR compared to the original models without AT, but they are less effective when compared to the performance of AT-PR. Specifically, the mean PR evaluations of AT-PR-trained models are higher than those trained with AT-WCR across most datasets, model architectures and norm-ball sizes. For instance, on CIFAR-100 with ResNet-18, AT-PR improves " + }, + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "inline_equation", + "content": "PR(\\cdot, 0.15)" + }, + { + "bbox": [ + 313, + 522, + 554, + 617 + ], + "type": "text", + "content": " from 44.71 (FGSM) and 80.47 (PGD) to 87.38." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "content": "iii) Compared to AT-WCR, the improvement in PR becomes more pronounced as the norm-ball size " + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "content": " increases, which is anticipated since larger norm-balls generally lead to less robust models. This demonstrates that our AT-PR approach is particularly useful when enhancing PR becomes more challenging for traditional AT-WCR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "iv) Each local input is associated with an estimated PR, and we report the mean and standard deviation over the test" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1679" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 552, + 397 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Evaluation results of DL models trained with different AT schemes, showing standard accuracy, adversarial accuracy (indicating WCR: PGD-10, C&W-10), and PR estimated (mean ± std) for four norm-ball sizes " + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": ". All results are calculated over the test datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 552, + 397 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 552, + 397 + ], + "type": "table", + "html": "
DatasetModelAT SchemeAccu. %PGD Ac. %CW Ac. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18W/O AT94.960.020.5178.38 ± 33.1466.3 ± 38.0653.41 ± 40.7637.01 ± 39.77
AT-WCR(FGSM)88.60.090.0993.99 ± 17.588.02 ± 25.0380.42 ± 31.8167.77 ± 38.6
AT-WCR(PGD)8145.0144.5697.28 ± 13.2995.77 ± 17.0794.1 ± 20.289.78 ± 26.53
AT-PR80.1439.7739.0398.22±10.3897.28±13.1696.05±16.1893.21±21.47
WRN-50-2W/O AT94.140.230.1476.96±32.7665.65±37.4154.43±39.8840.19±39.49
AT-WCR(FGSM)92.040.080.0495.68±14.5391.9±20.2586.67±25.9876.79±33.1
AT-WCR(PGD)82.6249.5148.1397.86±12.6197.08±14.5495.84±17.1492.62±22.87
AT-PR82.7143.2743.0798.09±10.8197.33±13.0396.22±15.5893.56±20.89
CIFAR-100ResNet-18W/O AT75.50.020.044.58 ± 42.6432.1 ± 40.123.32 ± 36.4115.62 ± 31.57
AT-WCR(FGSM)67.801.410.6683.90 ± 31.1172.89 ± 37.9461.06 ± 42.2344.71 ± 43.04
AT-WCR(PGD)52.8519.9319.5195.4 ± 18.1592.7 ± 22.7288.98 ± 28.1980.47 ± 36.4
AT-PR53.2119.3818.3997.59±12.9695.96±16.6293.15±21.7787.38±29.63
WRN-50-2W/O AT75.460.490.4551.11±40.5437.96±39.2328.33±36.7219.11±31.97
AT-WCR(FGSM)73.170.290.185.68±27.4576.62±33.8966.37/38.6652.04±41.42
AT-WCR(PGD)57.4623.1722.595.31±17.3892.0±23.7988.14±29.1580.67±36.21
AT-PR59.4622.0621.5996.05±16.1793.72±20.4789.48±26.8781.6±34.86
SVHNResNet-18W/O AT94.960.080.1297.11 ± 9.6695.63 ± 11.9893.74 ± 14.4890.12 ± 18.40
AT-WCR(FGSM)93.494.685.3298.79 ± 7.0698.42 ± 7.7097.86 ± 8.6796.57 ±10.78
AT-WCR(PGD)89.2144.5743.7098.96 ± 5.8598.50 ± 7.4097.72 ± 9.6195.41 ± 14.41
AT-PR90.2945.8142.2199.12 ± 5.1598.87 ± 5.9398.54 ± 6.8497.84 ± 8.60
WRN-50-2W/O AT95.911.741.2996.60 ± 10.3694.77 ± 13.0192.40 ± 15.8787.97 ± 20.04
AT-WCR(FGSM)94.318.138.5498.34 ± 7.4997.71 ± 8.3996.80 ± 9.6995.12 ± 12.06
AT-WCR(PGD)88.9154.9950.5898.72 ± 6.4498.28 ± 7.6697.76 ± 8.9596.77 ± 11.18
AT-PR91.9752.2251.5898.88 ± 6.1298.38 ± 7.5397.72 ± 9.1996.26 ± 12.35
Tiny-ImageNetResNet-18W/O AT59.260.00.080.78 ± 33.172.95 ± 37.6964.17 ± 40.9151.14 ± 42.91
AT-WCR(FGSM)54.810.050.0189.9 ± 24.9785.33 ± 29.9580.18 ± 33.7469.03 ± 39.39
AT-WCR(PGD)47.4321.319.3196.32 ± 17.0193.79 ± 22.190.16 ± 27.7682.46 ± 35.97
AT-PR51.6121.119.8998.23 ± 11.0696.1 ± 17.8493.12 ± 23.3286.65 ± 32.26
ResNet-34W/O AT61.660.00.082.39 ± 31.375.2 ± 36.066.95 ± 39.4253.7 ± 42.17
AT-WCR(FGSM)570.130.0290.67 ± 23.9285.81 ± 29.580.57 ± 34.1370.22 ± 39.36
AT-WCR(PGD)49.3320.9719.5895.45 ± 18.5593.08 ± 23.4889.85 ± 28.383.53 ± 34.99
AT-PR53.9821.7220.9596.71 ± 15.7495.02 ± 19.8992.5 ± 23.8786.23 ± 32.29
", + "image_path": "20f26f73625add2a7c250335597fcb2aadb9edb9a7eb8847306d164fb4617848.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 478 + ], + "type": "text", + "content": "set in Table 1. Notably, AT-PR consistently achieves higher mean PR compared to models trained with AT-WCR, while also exhibiting lower standard deviations. This indicates that AT-PR not only improves PR but also offers more stable and consistent performance across different local inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "text", + "content": " ) While PGD significantly improves robustness, it often reduces accuracy [66]. In contrast, AT-PR has a smaller negative impact on accuracy compared to PGD, with almost all models achieving higher accuracy than those trained using PGD while also enhancing both WCR and PR. This effect is more pronounced on larger datasets—for example, on Tiny-ImageNet with ResNet-18, AT-PR achieves an accuracy of " + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "inline_equation", + "content": "53.98\\%" + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "inline_equation", + "content": "49.33\\%" + }, + { + "bbox": [ + 55, + 480, + 295, + 577 + ], + "type": "text", + "content": " for PGD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 675 + ], + "type": "text", + "content": "Table 2 shows that combining AT-PR with TRADES improves clean accuracy and PR on more complex attention-based models ViT (Vanilla ViT and DeiT) across different scales (DeiT-Tiny, ViT-Small, ViT-Base) for CIFAR10. With just five epochs of AT-PR fine-tuning, all three ViT models achieve higher clean accuracy and PR while maintaining comparable WCR, showing a slight decrease on PGD-20 and AA but improved performance on CW-20." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Table 3 presents a comparison of AT-PR with four advanced AT methods (TRADES, MART, ALP and CLP) across three datasets using ResNet-18 and WRN-50-2. Re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "text", + "content": "sults show that AT-PR consistently achieves the best PR performance across four different norm-ball sizes " + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "text", + "content": ". Additionally, to verify that AT-PR is compatible with efficient AT methods [46, 62], we replace PGD with Fast-AT [62] in our framework (step 1 in Algo. 1). The result shows that Fast-AT-PR achieves " + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\sim 40\\%" + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "text", + "content": " speedup over the original AT-PR, with only " + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\sim 50\\%" + }, + { + "bbox": [ + 313, + 418, + 555, + 526 + ], + "type": "text", + "content": " extra cost compared to other AT baselines. Notably, Fast-AT-PR retains comparable PR performance to AT-PR and continues to outperform other AT baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 536, + 484, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 484, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 484, + 548 + ], + "type": "text", + "content": "4.2. Generalization Error of AT-PR" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": "We investigate the generalization error (GE) (train-test gap) of AT-PR to gain insights into its effectiveness. Table 4 compares the PR GE between AT-PR and PGD. Models trained with AT-PR consistently yields lower GE of up " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": " on CIFAR-10, " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "2.87\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": " on CIFAR-100, and " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "9.66\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": " on TinyImageNet, significantly lower than PGD's " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "4.67\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "8.54\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "inline_equation", + "content": "11.94\\%" + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": ", respectively. We further analyze the underlying reasons based on the PAC-Bayes analysis for Adversarial Robustness [54], see Appendix C for details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 671, + 492, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 492, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 492, + 684 + ], + "type": "text", + "content": "4.3. AEs Used in AT-WCR vs. AT-PR" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "We analyze the AEs used during training to understand how AT-PR differs from AT-WCR. By comparing the AEs gen" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1680" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 553, + 170 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. ViT models trained on CIFAR-10 with TRADES and AT-PR, showing standard accuracy, adversarial accuracy (indicating WCR: PGD, C&W, AA), and PR estimated (mean ± std) for four norm-ball sizes " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 553, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 553, + 170 + ], + "type": "table", + "html": "
DatasetModelAT SchemeAccu. %PGD-20. %CW-20. %AA. %PR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10DeiT-TiTRADES79.4449.7846.4545.9696.61 ± 15.2394.73 ± 19.3692.69 ± 23.0489.18 ± 27.9
+AT-PR81.0749.0846.9245.9597.26 ± 14.095.94 ± 17.1294.16 ± 20.590.65 ± 25.88
Vit-STRADES81.0350.9648.1447.4096.93 ± 14.9995.48 ± 18.493.58 ± 21.6290.06 ± 26.81
+AT-PR83.2950.4448.2047.3497.76 ± 11.8996.03 ± 16.4693.98 ± 20.6990.48 ± 26.19
Vit-BTRADES83.5553.1450.9449.9697.34 ± 13.5795.87 ± 16.9993.78 ± 21.1790.77 ± 25.88
+AT-PR86.0352.9451.2449.3497.9 ± 11.896.49 ± 15.294.44 ± 19.6291.02 ± 25.27
", + "image_path": "0132cb98599face1bc51aad2917e9b2cebc0ba4ad1a77bfb7457a78d0cd5265c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 201, + 541, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 180, + 555, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 555, + 192 + ], + "type": "text", + "content": "Table 3. Comparison of PR(, γ) and training time (sec/epoch) for more recent AT methods (TRADES, MART, ALP, CLP) and Fast-AT-PR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 201, + 541, + 485 + ], + "lines": [ + { + "bbox": [ + 69, + 201, + 541, + 485 + ], + "spans": [ + { + "bbox": [ + 69, + 201, + 541, + 485 + ], + "type": "table", + "html": "
DataModelMetricAT Scheme
PGDTRADESMARTALPCLPFast-AT-PRAT-PR
CIFAR-10ResNet-18PR(,0.08)%97.28 ± 13.2997.74 ± 12.4297.4 ± 13.796.73 ± 15.997.61 ± 13.0698.12 ± 11.5298.22 ± 10.38
PR(,0.10)%95.77 ± 17.0796.33 ± 16.3995.92 ± 17.4395.16 ± 19.5396.29 ± 16.6697.1 ± 14.3497.28 ± 13.16
PR(,0.12)%94.10 ± 20.2094.55 ± 20.3193.93 ± 21.3893.24 ± 23.0894.54 ± 20.2895.4 ± 17.9096.05 ± 16.18
PR(,0.15)%89.78 ± 26.5391.22 ± 25.7090.35 ± 26.9189.76 ± 28.1490.86 ± 26.2392.71 ± 23.5493.21 ± 21.47
Time (s/ep.)546144646497184
WRN-50-2PR(,0.08)%97.86 ± 12.6197.55 ± 13.1497.40 ± 13.6197.75 ± 12.4397.40 ± 13.9197.96 ± 10.8298.09 ± 10.81
PR(,0.10)%97.08 ± 14.5496.14 ± 16.7096.15 ± 16.7896.42 ± 16.1196.20 ± 17.5897.22 ± 12.9697.33 ± 13.03
PR(,0.12)%95.84 ± 17.1494.33 ± 20.4194.41 ± 20.3794.64 ± 20.0194.95 ± 19.8896.22 ± 15.6296.22 ± 15.58
PR(,0.15)%92.62 ± 22.8791.01 ± 25.7991.11 ± 25.6591.22 ± 25.7091.48 ± 25.6994.74 ± 19.8893.56 ± 20.89
Time (s/ep.)13110571155157229305
CIFAR-100ResNet-18PR(,0.08)%95.40 ± 18.1595.02 ± 19.2795.45 ± 18.1196.20 ± 16.0796.16 ± 16.3697.09 ± 13.4897.59 ± 12.96
PR(,0.10)%92.70 ± 22.7292.09 ± 24.2292.73 ± 23.3994.40 ± 20.0893.81 ± 21.4294.67 ± 18.9195.96 ± 16.62
PR(,0.12)%88.98 ± 28.1987.80 ± 29.8089.35 ± 28.3292.05 ± 24.1291.05 ± 25.8692.63 ± 23.9193.15 ± 21.77
PR(,0.15)%80.47 ± 36.4079.20 ± 37.7381.93 ± 35.7285.89 ± 31.6584.97 ± 33.0085.87 ± 31.9987.38 ± 29.63
Time (s/ep.)556143646597180
WRN-50-2PR(,0.08)%95.31 ± 17.3895.24 ± 18.6195.30 ± 16.8795.34 ± 17.7995.29 ± 17.7096.52 ± 15.1196.05 ± 16.17
PR(,0.10)%92.00 ± 23.7992.30 ± 23.6192.59 ± 22.8192.36 ± 23.1991.79 ± 23.8793.49 ± 20.9393.72 ± 20.47
PR(,0.12)%88.14 ± 29.1588.65 ± 28.4688.38 ± 28.6588.51 ± 28.2287.54 ± 29.9288.82 ± 28.3889.48 ± 26.87
PR(,0.15)%80.67 ± 36.2180.35 ± 36.5480.79 ± 36.0180.37 ± 36.8679.71 ± 36.7681.02 ± 37.0981.60 ± 34.86
Time (s/ep.)13110270155156225312
SVHNResNet-18PR(,0.08)%98.96 ± 5.8599.03 ± 5.7698.99 ± 6.1599.05 ± 5.7399.12 ± 5.3398.90 ± 5.9799.12 ± 5.15
PR(,0.10)%98.50 ± 7.4098.46 ± 8.2798.31 ± 8.2098.63 ± 7.1598.60 ± 7.2198.53 ± 7.2498.87 ± 5.93
PR(,0.12)%97.72 ± 9.6197.30 ± 11.0897.10 ± 11.2697.85 ± 9.1397.66 ± 9.7898.11 ± 8.1998.54 ± 6.84
PR(,0.15)%95.41 ± 14.4194.15 ± 16.5794.12 ± 16.6595.54 ± 14.1094.93 ± 15.0097.10 ± 10.0597.84 ± 8.60
Time (s/ep.)7886619193141202
WRN-50-2PR(,0.08)%98.72 ± 6.4498.69 ± 6.9098.61 ± 7.6698.54 ± 7.5298.59 ± 7.4698.43 ± 6.6998.88 ± 6.12
PR(,0.10)%98.28 ± 7.6698.11 ± 8.6698.07 ± 9.2798.12 ± 8.7498.00 ± 8.8998.15 ± 7.2998.38 ± 7.53
PR(,0.12)%97.76 ± 8.9597.26 ± 10.8997.34 ± 10.9097.65 ± 8.8596.98 ± 11.1397.79 ± 8.1297.72 ± 9.19
PR(,0.15)%96.77 ± 11.1895.20 ± 14.8495.45 ± 14.7295.67 ± 14.3294.74 ± 15.3796.94 ± 9.8096.26 ± 12.35
Time (s/ep.)190146102223225321420
", + "image_path": "ec51dd576a30f3c52df87d336afe93a11e2262fb7f2167027f67ae36b276ef74.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": "erated by PGD and AT-PR, we find that about " + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": " of training inputs produce different AEs (e.g., " + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "inline_equation", + "content": "10.05\\% / 8.39\\%" + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": " for ResNet-18 and " + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "inline_equation", + "content": "8.83\\% / 10.42\\%" + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": " for WRN50-2 on CIFAR-10/CIFAR-100). This means that for roughly " + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": " of inputs, the loss landscapes align with Case 2 (\"tall & wide vs. short & tight\") in Fig. 2, where both methods select the same AEs. However, for the remaining " + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 505, + 295, + 613 + ], + "type": "text", + "content": " -ish, where the landscape fits Case 1 (\"tall & tight vs. short & wide\"), AT-PR identifies different AEs that better optimize PR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Fig. 5 (top row) compares the loss of AEs generated by PGD and AT-PR. For all models, PGD AEs consistently have higher losses, reflecting its objective to find AE with the maximum loss for AT. This aligns with the goal of AT-WCR, which seeks to minimize the loss on such worst-case AEs (see Fig.2, top-middle). In contrast, AT-PR generates AEs with lower loss. Combined with the fact that our Algorithm 1 searches for examples farthest from the deci" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 553 + ], + "type": "text", + "content": "sion boundary, suggests that AT-PR effectively locates AEs within the widest peak of the loss landscape. Training with these AEs helps reduce the \"short & wide\" peak (see Fig. 2, top-right)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Fig. 5 (b) visualizes the distributions of all generated AEs " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{x_i^{\\prime}\\}_{i = 1}^{n}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " by analyzing the distance " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "(L_{2}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " norm) between each " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " and the original input " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_0,d_1" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " in Fig. 5 (a)), along with their angular relationships. The 2D coordinate in Fig. 5 (a) introduces a fixed anchor " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{a}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " near " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", for an AE " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_0^\\prime" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", the angle " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_0" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " is computed between " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\vec{v}_{x\\rightarrow x_0'}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\vec{v}_{x\\rightarrow x_a}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", representing " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "'s relative position, given by: " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta_0 = \\arccos \\left(\\frac{\\vec{v}_{x\\rightarrow x_0'}\\cdot\\vec{v}_{x\\rightarrow x_a}}{\\|\\vec{v}_{x\\rightarrow x_0'}\\||\\vec{v}_{x\\rightarrow x_a}\\|}\\right)" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ". The resulting angles range from approximately " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "88^{\\circ}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "92^{\\circ}" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " Fig. 5 (b), reflecting the unique geometric properties of high-dimensional spaces, where random vectors highly likely to be orthogonal [53]. To further capture the variability in relative positions," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1681" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 101, + 553, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 4. Comparative GE experiment on DL models across different datasets. Train/test set evaluations of PR estimates (mean) for four norm-ball sizes " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " across models trained with PGD and AT-PR, presented as " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "PR_{train} / PR_{test} / (PR_{train} - PR_{test})" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 101, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 59, + 101, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 101, + 553, + 251 + ], + "type": "table", + "html": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
WRN-50-2PGD98.88/97.86/1.0198.19/97.08/1.197.49/95.84/1.6495.3/92.62/2.67
AT-PR98.63/98.09/0.5497.8/97.33/0.4796.71/96.22/0.4694.9/93.56/1.34
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
WRN-50-2PGD98.35/95.31/3.0496.71/92.0/4.7193.67/88.14/5.5386.69/80.67/6.02
AT-PR97.3/96.05/1.2595.41/93.72/1.6991.93/89.48/2.4583.25/81.6/1.65
Tiny-ImageNetResNet-18PGD98.72/96.32/2.497.86/93.79/4.0796.94/90.16/6.7894.4/82.46/11.94
AT-PR99.01/98.23/0.7898.02/96.1/1.9296.68/93.12/3.5694.31/86.65/7.66
ResNet-34PGD98.67/95.45/3.2298.02/93.08/6.9496.9/89.85/7.0593.36/83.53/9.83
AT-PR99.28/96.71/2.5798.82/95.02/3.898.11/92.5/5.6195.89/86.23/9.66
", + "image_path": "ad1331b8c313fa988b50992fd8fbb578b49e8a9d1dfd0d443453730bd8e99e5a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 280, + 553, + 367 + ], + "blocks": [ + { + "bbox": [ + 57, + 260, + 552, + 273 + ], + "lines": [ + { + "bbox": [ + 57, + 260, + 552, + 273 + ], + "spans": [ + { + "bbox": [ + 57, + 260, + 552, + 273 + ], + "type": "text", + "content": "Table 5. PR (mean) on train/test sets, reported as: " + }, + { + "bbox": [ + 57, + 260, + 552, + 273 + ], + "type": "inline_equation", + "content": "{PR}_{\\text{train }}/{PR}_{\\text{test }}/\\left( {{PR}_{\\text{train }} - P{R}_{\\text{test }}}\\right)" + }, + { + "bbox": [ + 57, + 260, + 552, + 273 + ], + "type": "text", + "content": " ,for model trained with PGD, Random-AE, AT-PR." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 280, + 553, + 367 + ], + "lines": [ + { + "bbox": [ + 58, + 280, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 58, + 280, + 553, + 367 + ], + "type": "table", + "html": "
DatasetModelAT SchemePR(·,0.08) %PR(·,0.1) %PR(·,0.12) %PR(·,0.15) %
CIFAR-10ResNet-18PGD99.02/97.28/1.7498.36/95.77/2.5997.33/94.1/3.2394.45/89.78/4.67
Random98.62/97.71/0.9197.73/96.78/0.9596.58/95.48/1.194.19/92.64/1.54
AT-PR98.58/98.22/0.3697.92/97.28/0.6496.99/96.05/0.9495.41/93.21/2.2
CIFAR-100ResNet-18PGD99.02/95.4/3.6197.98/92.7/5.2895.49/88.98/6.589.01/80.47/8.54
Random96.6/95.87/0.7394.85/93.77/1.0892.18/90.92/1.2685.39/85.21/0.18
AT-PR97.64/97.59/0.0596.59/95.96/0.6394.78/93.15/1.6290.25/87.38/2.87
", + "image_path": "778121e37fcf5c7b726e33e62aceab1291063f57c4cb38f830a2bc45b263caf2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 383, + 132, + 446 + ], + "blocks": [ + { + "bbox": [ + 56, + 383, + 132, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 383, + 132, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 132, + 446 + ], + "type": "image", + "image_path": "ce2f8a4c449f5f3d9bf5a24c5f87d0ddfe062327b258dc554d9233bde8b8847a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 384, + 211, + 445 + ], + "blocks": [ + { + "bbox": [ + 134, + 384, + 211, + 445 + ], + "lines": [ + { + "bbox": [ + 134, + 384, + 211, + 445 + ], + "spans": [ + { + "bbox": [ + 134, + 384, + 211, + 445 + ], + "type": "image", + "image_path": "9c53a0c6169b3a8975f751a0993ad809813e9ee9ca9d59962975d0b870946566.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 217, + 383, + 293, + 445 + ], + "blocks": [ + { + "bbox": [ + 217, + 383, + 293, + 445 + ], + "lines": [ + { + "bbox": [ + 217, + 383, + 293, + 445 + ], + "spans": [ + { + "bbox": [ + 217, + 383, + 293, + 445 + ], + "type": "image", + "image_path": "d38565850766f48f6710a0e15fb24e758a3b910ecaf73ef27e54942e04a01449.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 450, + 127, + 522 + ], + "blocks": [ + { + "bbox": [ + 58, + 450, + 127, + 522 + ], + "lines": [ + { + "bbox": [ + 58, + 450, + 127, + 522 + ], + "spans": [ + { + "bbox": [ + 58, + 450, + 127, + 522 + ], + "type": "image", + "image_path": "6944ecc0522a75de4e80259b8222c967df4093025b1850393d951237bc4db04d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 532, + 295, + 577 + ], + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 577 + ], + "type": "text", + "content": "Figure 5. (Top row) Averaged loss of AEs used during AT for AT-PR and PGD. (Bottom row) (a) Illustration of angles and distances of AEs relative to " + }, + { + "bbox": [ + 55, + 532, + 295, + 577 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 532, + 295, + 577 + ], + "type": "text", + "content": ". (b) Distribution of generated AEs. (c) Pairwise angular distribution among AEs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 132, + 446, + 214, + 523 + ], + "blocks": [ + { + "bbox": [ + 132, + 446, + 214, + 523 + ], + "lines": [ + { + "bbox": [ + 132, + 446, + 214, + 523 + ], + "spans": [ + { + "bbox": [ + 132, + 446, + 214, + 523 + ], + "type": "image", + "image_path": "c11df29aff62af65b87111c9d103e94deb370b57456449e113dfa9066ddab0e7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 216, + 446, + 294, + 512 + ], + "blocks": [ + { + "bbox": [ + 216, + 446, + 294, + 512 + ], + "lines": [ + { + "bbox": [ + 216, + 446, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 216, + 446, + 294, + 512 + ], + "type": "image", + "image_path": "0ee19c9c6a7cd7cf84f3e63d92f457a3e2f9a3a8b57392338500c5b1526c2ab2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 258, + 515, + 266, + 522 + ], + "lines": [ + { + "bbox": [ + 258, + 515, + 266, + 522 + ], + "spans": [ + { + "bbox": [ + 258, + 515, + 266, + 522 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "we computed the Pairwise Angular Distribution among all vectors " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\vec{v}_{x\\rightarrow x_i'}\\}_{i = 1}^n" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". The angle " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " for each pair of vectors is calculated as illustrated in Fig. 5 (a), with the resulting distribution shown in Fig. 5 (c). This analysis shows that the angles between all pairs of vectors range from " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "85^{\\circ}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", indicating that the generated AEs are widely spread within the norm-ball. The final AE selected by AT-PR, highlighted in orange in Fig. 5 (b), has the largest distance to nearest decision boundary (1.0 for the L2 norm), compared to other" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 385, + 463, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 463, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 463, + 397 + ], + "type": "text", + "content": "AEs, which range from 0.76 to 0.92." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 529 + ], + "type": "text", + "content": "To empirically validate that AEs generated by AT-PR are most effective for improving PR, we conduct an ablation study by randomly sampling AEs from those generated in Step 1 of Algo. 1 (also illustrated as random blue spots in Fig.5 (b)). Table 5 shows that while randomly selected AEs improve PR and GE compared to PGD, AT-PR consistently performs better. As Fig. 2 illustrates, although random AEs may fall within a short and wide peak (explaining their improved PR over PGD), they do not necessarily lie within the \"widest peak\" that AT-PR specifically targets. This distinction explains AT-PR's superior PR performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 557, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 676 + ], + "type": "text", + "content": "In this work, for the first time, we introduce a novel AT method targeting PR, named as AT-PR. By reformulating a new min-max optimization, we lay the theoretical foundation for AT-PR and develop specialized numerical algorithms to solve this optimization. Extensive experimental evaluation demonstrates that AT-PR consistently outperforms traditional AT methods designed for WCR in improving PR. Additionally, we show that AT-PR incurs a smaller trade-off in model generalization compared to common AT-WCR methods." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In the future, we plan to address the challenges identified in [73], such as benchmarking PR and extending it to diverse generative AI tasks." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1682" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "Acknowledgements SK and XZ have received funding from the European Union's EU Framework Program for Research and Innovation Europe Horizon (grant agreement No 101202457). XZ's contribution is also supported by the UK EPSRC New Investigator Award [EP/Z536568/1] and NVIDIA Academic Grant Program. SK's contribution is supported by the UKRI Future Leaders Fellowship Grant [MR/S035176/1], and Innovate UK through the DriveSafeAI Project (10062594). XH has received funding from the European Union's Horizon Europe Research and Innovation Action under grant agreement No 101212818. YZ's contribution is supported by China Scholarship Council. We thank Zheng Wang for his valuable comments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 229, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 294, + 289 + ], + "type": "text", + "content": "Views and opinions expressed are those of the authors only and do not necessarily reflect those of the European Union or European Research Executive Agency (REA). Neither the European Union nor the granting authority can be held responsible for them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 303, + 115, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 115, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 115, + 316 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 323, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 323, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 295, + 355 + ], + "type": "text", + "content": "[1] Amir Aminifar. Universal adversarial perturbations in epileptic seizure detection. In 2020 Int. Joint Conference on Neural Networks (IJCNN), pages 1-6. IEEE, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 357, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 357, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 357, + 296, + 411 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In Proc. of the 35th International Conference on Machine Learning, pages 274-283. PMLR, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 415, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 459 + ], + "type": "text", + "content": "[3] Tao Bai, Jinqi Luo, Jun Zhao, Bihan Wen, and Qian Wang. Recent advances in adversarial training for adversarial robustness. In Proc. of the 30th Int. Joint Conference on Artificial Intelligence, pages 4312-4321, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 460, + 294, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 505 + ], + "type": "text", + "content": "[4] Mislav Balunović and Martin Vechev. Adversarial training and provable defenses: Bridging the gap. In 8th International Conference on Learning Representations (ICLR'20'20), 2020. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 507, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 507, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 507, + 294, + 540 + ], + "type": "text", + "content": "[5] Teodora Baluta, Zheng Leong Chua, Kuldeep S Meel, and Prateek Saxena. Scalable quantitative verification for deep neural networks. In ICSE'21, pages 312-323, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 542, + 294, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 542, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 62, + 542, + 294, + 585 + ], + "type": "text", + "content": "[6] Qi-Zhi Cai, Chang Liu, and Dawn Song. Curriculum adversarial training. In Proc. of the 27th International Joint Conference on Artificial Intelligence, page 3740-3747. AAAI Press, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 588, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 588, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 588, + 294, + 621 + ], + "type": "text", + "content": "[7] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In IEEE symposium on security and privacy (sp), pages 39-57, 2017. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 294, + 666 + ], + "type": "text", + "content": "[8] Anirban Chakraborty, Manaar Alam, Vishal Dey, Anupam Chattopadhyay, and Debdeep Mukhopadhyay. A survey on adversarial attacks and defences. CAAI Transactions on Intelligence Technology, 6(1):25-45, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[9] Jianbo Chen, Michael I Jordan, and Martin J Wainwright. Hopskipjumpattack: A query-efficient decision-based attack. In IEEE symposium on security and privacy (SP), pages 1277-1294. IEEE, 2020. 4, 12" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[10] Pin-Yu Chen, Huan Zhang, Yash Sharma, Jinfeng Yi, and Cho-Jui Hsieh. ZOO: Zeroth order optimization based blackbox attacks to deep neural networks without training substitute models. In Proc. of the 10th ACM Workshop on Artificial Intelligence and Security, page 15-26. ACM, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[11] Zhen Chen, Yi Zhang, Fu Wang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. TARP-VP: Towards evaluation of transferred adversarial robustness and privacy on label mapping visual prompting models. Advances in Neural Information Processing Systems, 37:6776-6796, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 554, + 228 + ], + "type": "text", + "content": "[12] Jeremy Cohen, Elan Rosenfeld, and Zico Kolter. Certified adversarial robustness via randomized smoothing. In International Conference on Machine Learning, pages 1310-1320. PMLR, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 262 + ], + "type": "text", + "content": "[13] Nicolas Couellan. Probabilistic robustness estimates for feed-forward neural networks. Neural networks, 142:138-147, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 308 + ], + "type": "text", + "content": "[14] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In International conference on machine learning, pages 2206-2216. PMLR, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "text", + "content": "[15] Gavin Weiguang Ding, Yash Sharma, Kry Yik Chau Lui, and Ruitong Huang. Mma training: Direct input space margin maximization through adversarial training. In International Conference on Learning Representations, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 419 + ], + "type": "text", + "content": "[16] Yi Dong, Wei Huang, Vibhav Bharti, Victoria Cox, Alec Banks, Sen Wang, Xingyu Zhao, Sven Schewe, and Xiaowei Huang. Reliability assessment and safety arguments for machine learning components in system assurance. ACM Transactions on Embedded Computing Systems, 22(3):1-48, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 422, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 498 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 553, + 533 + ], + "type": "text", + "content": "[18] Chengbin Du, Yanxi Li, Zhongwei Qiu, and Chang Xu. Stable diffusion is unstable. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 534, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 554, + 599 + ], + "type": "text", + "content": "[19] Xiaoning Du, Xiaofei Xie, Yi Li, Lei Ma, Yang Liu, and Jianjun Zhao. Deepstellar: Model-based quantitative analysis of stateful deep learning systems. In Proc. of the 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 477–487, 2019. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 601, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 656 + ], + "type": "text", + "content": "[20] Timon Gehr, Matthew Mirman, Dana Drachsler-Cohen, Petar Tsankov, Swarat Chaudhuri, and Martin Vechev. Ai2: Safety and robustness certification of neural networks with abstract interpretation. In IEEE symposium on security and privacy (SP), pages 3-18, 2018. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 689 + ], + "type": "text", + "content": "[21] Ian J Goodfellow. Explaining and harnessing adversarial examples. In International Conference on Learning Representations (ICLR'15), 2015. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[22] Sven Gowal, Krishnamurthy Dj Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1683" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "Arandjelovic, Timothy Mann, and Pushmeet Kohli. Scalable verified training for provably robust image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4842-4851, 2019. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[23] Wei Huang, Youcheng Sun, Xingyu Zhao, James Sharp, Wenjie Ruan, Jie Meng, and Xiaowei Huang. Coverage-Guided Testing for Recurrent Neural Networks. IEEE Transactions on Reliability, 71(3):1191-1206, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[24] Wei Huang, Xingyu Zhao, Alec Banks, Victoria Cox, and Xiaowei Huang. Hierarchical distribution-aware testing of deep learning. ACM Transactions on Software Engineering and Methodology, 33(2):1-35, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "text", + "content": "[25] Wei Huang, Xingyu Zhao, Gaojie Jin, and Xiaowei Huang. Safari: Versatile and efficient evaluations for robustness of interpretability. In IEEE/CVF International Conference on Computer Vision (ICCV'23), pages 1988-1998, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 308 + ], + "type": "text", + "content": "[26] Xiaowei Huang, Marta Kwiatkowska, Sen Wang, and Min Wu. Safety verification of deep neural networks. In Computer Aided Verification: 29th International Conference, CAV 2017, Heidelberg, Germany, July 24-28, 2017, Proceedings, Part I 30, pages 3-29. Springer, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 353 + ], + "type": "text", + "content": "[27] Xiaowei Huang, Daniel Kroening, et al. A survey of safety and trustworthiness of deep neural networks: Verification, testing, adversarial attack and defence, and interpretability. Computer Science Review, 37:100270, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 398 + ], + "type": "text", + "content": "[28] Xiaowei Huang, Wenjie Ruan, et al. A survey of safety and trustworthiness of large language models through the lens of verification and validation. Artificial Intelligence Review, 57 (7):175, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 431 + ], + "type": "text", + "content": "[29] Harini Kannan, Alexey Kurakin, and Ian Goodfellow. Adversarial logit pairing. arXiv preprint arXiv:1803.06373, 2018. 3, 5, 12" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "text", + "content": "[30] Guy Katz, Clark Barrett, David L Dill, Kyle Julian, and Mykel J Kochenderfer. Reluplex: An efficient smt solver for verifying deep neural networks. In 29th Int. Conf. Computer Aided Verification, pages 97-117. Springer, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 522 + ], + "type": "text", + "content": "[31] Huichen Li, Xiaojun Xu, Xiaolu Zhang, Shuang Yang, and Bo Li. Qeba: Query-efficient boundary-based blackbox attack. In Proc. of the IEEE/CVF conference on computer vision and pattern recognition, pages 1221-1230, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 578 + ], + "type": "text", + "content": "[32] Han Liu, Yuhao Wu, Shixuan Zhai, Bo Yuan, and Ning Zhang. Riatig: Reliable and imperceptible adversarial text-to-image generation with natural prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20585-20594, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 624 + ], + "type": "text", + "content": "[33] Qihao Liu, Adam Kortylewski, Yutong Bai, Song Bai, and Alan Yuille. Discovering failure modes of text-guided diffusion models via adversarial search. In The 12th International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "text", + "content": "[35] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deepauge: Multi-granularity testing criteria for deep" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "learning systems. In Proce. of the 33rd ACM/IEEE Int. Conference on Automated Software Engineering (ASE'18), pages 120-131, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[36] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, 2018. 1, 2, 3, 4, 12" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 205 + ], + "type": "text", + "content": "[37] Mark Huasong Meng, Guangdong Bai, Sin Gee Teo, Zhe Hou, Yan Xiao, Yun Lin, and Jin Song Dong. Adversarial robustness of deep neural networks: A survey from a formal verification perspective. IEEE Transactions on Dependable and Secure Computing, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "text", + "content": "[38] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard. Universal adversarial perturbations. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1765-1773, 2017. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 304 + ], + "type": "text", + "content": "[39] Nicolas Papernot, Patrick McDaniel, Somesh Jha, Matt Fredrikson, Z Berkay Celik, and Ananthram Swami. The limitations of deep learning in adversarial settings. In IEEE European symposium on security and privacy (EuroS&P), pages 372-387. IEEE, 2016. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "text", + "content": "[40] Nicolas Papernot, Patrick McDaniel, Xi Wu, Somesh Jha, and Ananthram Swami. Distillation as a defense to adversarial perturbations against deep neural networks. In 2016 IEEE symposium on security and privacy (SP), pages 582-597. IEEE, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 414 + ], + "type": "text", + "content": "[41] Mikhail Pautov, Nurislam Tursynbek, Marina Munkhoeva, Nikita Muravev, Aleksandr Petiushko, and Ivan Oseledets. Cc-cert: A probabilistic approach to certify general robustness of neural networks. In AAAI'22, pages 7975-7983, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 459 + ], + "type": "text", + "content": "[42] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepexplore: Automated whitebox testing of deep learning systems. In Proceedings of the 26th Symposium on Operating Systems Principles, pages 1-18. ACM, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[43] Alexander Robey, Luiz Chamon, George J Pappas, and Hamed Hassani. Probabilistically robust learning: Balancing average and worst-case performance. In Int. Conference on Machine Learning, pages 18667-18686. PMLR, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[44] Wenjie Ruan, Xiaowei Huang, and Marta Kwiatkowska. Reachability analysis of deep neural networks with provable guarantees. In Proc. of the 27th Int. Joint Conference on Artificial Intelligence (IJCAI'18), pages 2651-2659, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "type": "text", + "content": "[45] Sebastian Ruder. An overview of gradient descent optimization algorithms. ArXiv, abs/1609.04747, 2016. 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[46] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! Advances in neural information processing systems, 32, 2019. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[47] Uri Shaham, Yutaro Yamada, and Sahand Negahban. Understanding adversarial training: Increasing local stability of supervised models through robust optimization. Neurocomputing, 307:195-204, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In In Proc. of 2nd Int. Conf. on Learning Representations, 2014. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1684" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[49] Karim Tit, Teddy Furon, and Mathias Rousset. Efficient statistical assessment of neural network corruption robustness. NeurIPS'21, 34:9253-9263, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[50] Karim TIT, Teddy Furon, and Mathias Rousset. Gradient-informed neural network statistical robustness estimation. In Proc. of The 26th Int. Conf. on Artificial Intelligence and Statistics, pages 323-334. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[51] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[52] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. In International Conference on Learning Representations, 2018. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 271 + ], + "type": "text", + "content": "[53] Princeton University. Lecture 11: Rademacher complexity, 2014. Accessed: 2025-01-05. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[54] Paul Viallard, Eric Guillaume VIDOT, Amaury Habrard, and Emilie Morvant. A pac-bayes analysis of adversarial robustness. Advances in Neural Information Processing Systems, 34:14421-14433, 2021. 6, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 349 + ], + "type": "text", + "content": "[55] Benjie Wang, Stefan Webb, and Tom Rainforth. Statistically robust neural network classification. In Uncertainty in Artificial Intelligence, pages 1735-1745. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[56] Yisen Wang, Xingjun Ma, James Bailey, Jinfeng Yi, Bowen Zhou, and Quanquan Gu. On the convergence and robustness of adversarial training. In ICML'19, pages 6586-6595. PMLR, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[57] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In International conference on learning representations, 2019. 3, 5, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "text", + "content": "[58] Zeyu Wang, Xianhang Li, Hongru Zhu, and Cihang Xie. Revisiting adversarial training at scale. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24675-24685, 2024. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "text", + "content": "[59] Stefan Webb, Tom Rainforth, Yee Whye Teh, and M Pawan Kumar. A statistical approach to assessing neural network robustness. In Int. Conf. on Learning Representations, 2019, 1, 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 581 + ], + "type": "text", + "content": "[60] Lily Weng, Pin-Yu Chen, Lam Nguyen, Mark Squillante, Akhilan Boopathy, Ivan Oseledets, and Luca Daniel. Proven: Verifying robustness of neural networks with a probabilistic approach. In Int. Conf. on Machine Learning, pages 6727-6736. PMLR, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[61] Tsui Wei Weng, Huan Zhang, Pin-Yu Chen, Jinfeng Yi, Dong Su, Yupeng Gao, Cho-Jui Hsieh, and Luca Daniel. Evaluating the robustness of neural networks: An extreme value theory approach. In Int. Conf. on Learning Representations, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 669 + ], + "type": "text", + "content": "[62] Eric Wong, Leslie Rice, and J. Zico Kolter. Fast is better than free: Revisiting adversarial training. In International Conference on Learning Representations, 2020. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[63] Weiming Xiang, Hoang-Dung Tran, and Taylor T. Johnson. Output reachable set estimation and verification for multilayer neural networks. IEEE Tran. on Neural Networks and Learning Systems, 29(11):5777-5783, 2018. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 686 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[64] Xiangyu Yin, Sihao Wu, Jiaxu Liu, Meng Fang, Xingyu Zhao, Xiaowei Huang, and Wenjie Ruan. Representation-based robustness in goal-conditioned reinforcement learning. In Proc. of the AAAI Conference on Artificial Intelligence, pages 21761-21769, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[65] Xiaoyong Yuan, Pan He, Qile Zhu, and Xiaolin Li. Adversarial examples: Attacks and defenses for deep learning. IEEE Tran. on neural networks and learning systems, 30(9):2805-2824, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "text", + "content": "[66] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In International conference on machine learning, pages 7472-7482. PMLR, 2019. 3, 5, 6, 12" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 283 + ], + "type": "text", + "content": "[67] Jingfeng Zhang, Xilie Xu, Bo Han, Gang Niu, Lizhen Cui, Masashi Sugiyama, and Mohan Kankanhalli. Attacks which do not kill training make adversarial learning stronger. In International conference on machine learning, pages 11278-11287. PMLR, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[68] Jiawei Zhang, Linyi Li, Huichen Li, Xiaolu Zhang, Shuang Yang, and Bo Li. Progressive-scale boundary blackbox attack via projective gradient estimation. In Int. Conference on Machine Learning, pages 12479-12490. PMLR, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[69] Tianle Zhang, Wenjie Ruan, and Jonathan E. Fieldsend. Proa: A probabilistic robustness assessment against functional perturbations. In Machine Learning and Knowledge Discovery in Databases (ECML/PKDD'22), page 154-170. Springer-Verlag, 2023. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 441 + ], + "type": "text", + "content": "[70] Tianle Zhang, Yanghao Zhang, Ronghui Mu, Jiaxu Liu, Jonathan Fieldsend, and Wenjie Ruan. Prass: probabilistic risk-averse robust learning with stochastic search. In Proceedings of the 33rd International Joint Conference on Artificial Intelligence, pages 559–567, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 496 + ], + "type": "text", + "content": "[71] Yi Zhang, Zhen Chen, Chih-Hong Cheng, Wenjie Ruan, Xiaowei Huang, Dezong Zhao, David Flynn, Siddartha Khastgir, and Xingyu Zhao. Trustworthy text-to-image diffusion models: A timely and focused survey. arXiv preprint arXiv:2409.18214, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[72] Yi Zhang, Yun Tang, Wenjie Ruan, Xiaowei Huang, Siddartha Khastgir, Paul Jennings, and Xingyu Zhao. Protip: Probabilistic robustness verification on text-to-image diffusion models against stochastic perturbation. In European Conference on Computer Vision, 2024. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[73] Xingyu Zhao. Probabilistic robustness in deep learning: A concise yet comprehensive guide. Adversarial Example Detection and Mitigation Using Machine Learning, pages 1-13, 2025. 1, 2, 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 597, + 553, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 597, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 316, + 597, + 553, + 642 + ], + "type": "text", + "content": "[74] Haomin Zhuang, Yihua Zhang, and Sijia Liu. A pilot study of query-free adversarial attack against stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2385-2392, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 316, + 643, + 553, + 686 + ], + "type": "text", + "content": "[75] Monty-Maximilian Zuhlke and Daniel Kudenko. Adversarial robustness of neural networks from the perspective of Lipschitz calculus: A survey. ACM Computing Surveys, 2024. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1685" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_content_list.json b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5183789594978e37774cae6ad6b43b5b99a577b6 --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_content_list.json @@ -0,0 +1,1701 @@ +[ + { + "type": "text", + "text": "AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations", + "text_level": 1, + "bbox": [ + 151, + 130, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junli Liu\\*1,2 Qizhi Chen\\*2,3 Zhigang Wang\\*2 Yiwen Tang1,2 Yiting Zhang2 Chi Yan2 Dong Wang2 Xuelong Li4 Bin Zhao+1,2", + "bbox": [ + 163, + 200, + 831, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwestern Polytechnical University $^{2}$ Shanghai AI Laboratory $^{3}$ Zhejiang University $^{4}$ TeleAI", + "bbox": [ + 156, + 239, + 836, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b218763d3a467aadcefb54e6c41196e0f01b577e07cd170e988f9fc842c2b40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 270, + 307, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19fa1d712a5c0b08add20267d0bf40dce4bd2acfe3150becd488a5a460a16298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 353, + 307, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fa850aa2a7778806fa37f912e486242f488df5b1a2ad71a10bdba912a6b6db8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 435, + 307, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7e0337ba00e5ab7e7281f64f48ceed83ca2c6540def82e2e7a0efecb460c7e0.jpg", + "image_caption": [ + "Figure 1. Overview of AerialVG. This work introduces aerial visual grounding (AerialVG), a new task for intelligent drones. We provide the first dataset for this task, consisting of 50K annotations and 103K objects. Additionally, an AerialVG model is proposed to explore the positional relations among objects and achieve superior context-aware visual grounding in drone applications." + ], + "image_footnote": [], + "bbox": [ + 120, + 517, + 307, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7680ff1f87d71a5285eabc78ef24107466370d0d343d3b1c4032d067cdee4d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 270, + 496, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59b4fd44115f67d09ded023a4b05562714d5608f50be9d00d34321876a0e64ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 353, + 496, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b62f54ef27fa623c2d0a341ed924370a5d14bafb9594d12678d3b8db5eee2872.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 435, + 496, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/900f50284ace838e4e62e5190a18e215628cd6a66c380fa7618a8d63a84d6d2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 517, + 496, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3752a4fd729a136778f2e39b6fad323d57250297c15796ce12ba87543956bbf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 270, + 686, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd5a8d2f549929ce89505607dd7663828e78051e3313b0ac6e3943d4a8a8e06a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 353, + 686, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6735695b6dd05b4bd390a0337d6e13fadea7aa79b40a2dadbbc6dad5f43dee62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 435, + 686, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b584175ecc471a7f16d405ad283d7d853698d011e334cfe708cdb1ea29de179.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 517, + 686, + 598 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2aa53f424bfb86a72aa6e0a94b12aca6611836d3153683569d18366c12620dfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 270, + 875, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e325b8e2d8e505b5a9d5fec2f9ec37e38747b7bcde1807d03e5d636931f8633.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 353, + 875, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6123157bda65f05d2efdc5467638e1ac8e96efcabc132ba767c0fb6777d6df36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 435, + 875, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c095c75e6db2bb190435e03e3a49842b84f3f37065862afa09c0d0e3706de897.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 517, + 875, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 665, + 326, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual grounding (VG) aims to localize target objects in an image based on natural language descriptions. In this paper, we propose AerialVG, a new task focusing on visual grounding from aerial views. Compared to traditional VG, AerialVG poses new challenges, e.g., appearance-based grounding is insufficient to distinguish among multiple visually similar objects, and positional relations should be emphasized. Besides, existing VG models struggle when applied to aerial imagery, where high-resolution images cause significant difficulties. To address these challenges, we introduce the first AerialVG dataset, consisting of 5K", + "bbox": [ + 88, + 696, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "real-world aerial images, 50K manually annotated descriptions, and 103K objects. Particularly, each annotation in AerialVG dataset contains multiple target objects annotated with relative spatial relations, requiring models to perform comprehensive spatial reasoning. Furthermore, we propose an innovative model especially for the AerialVG task, where a Hierarchical Cross-Attention is devised to focus on target regions, and a Relation-Aware Grounding module is designed to infer positional relations. Experimental results validate the effectiveness of our dataset and method, highlighting the importance of spatial reasoning in aerial visual grounding. The code will be released at https://github.com/Ideal-ljl/AerialVG.", + "bbox": [ + 511, + 666, + 908, + 862 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally", + "bbox": [ + 107, + 875, + 264, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5177", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Grounding (VG) is a challenging task that involves identifying specific regions or objects within an image following natural language instructions. This task requires models to deeply understand both visual and textual modalities, effectively aligning image context with corresponding linguistic expressions. VG is critical for a wide range of applications, including visual question answering, vision-and-language navigation, and human-robot interaction, etc [22, 33, 35, 36, 49, 55].", + "bbox": [ + 89, + 114, + 480, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the continuous progress of multimodal models, the performance of VG tasks has been significantly improved. However, existing models and datasets still have obvious limitations. Current research focuses primarily on ground objects, and the exploration of aerial visual grounding remains underdeveloped. Aerial visual grounding has significant potential applications in areas such as emergency rescue, logistics, and ecological monitoring, which calls for further research and development. Specifically, the ability to manage multiple objects of the same type and their spatial relationships in complex environments remains underdeveloped, leading to limited generalization capabilities in real-world scenarios.", + "bbox": [ + 89, + 251, + 482, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a more challenging task, termed Aerial Visual Grounding (AerialVG). AerialVG aims to apply the visual grounding task to UAV platforms, and promote the development of UAV intelligent perception and autonomous navigation technologies. Compared with traditional VG tasks, AerialVG faces more challenges, mainly in the following two aspects:", + "bbox": [ + 89, + 448, + 482, + 553 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Broad field of view and small objects. Unlike images from the ground perspective, when UAVs capture images from the air, the field of view is usually much broader, and the size of objects in the image tends to be relatively small. While a larger field of view provides more context and coverage, it also imposes challenges to locate and identify objects at a finer scale.", + "- Complex object distribution and spatial relationship." + ], + "bbox": [ + 89, + 553, + 482, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From the perspective of UAVs, object types are mainly pedestrians and vehicles, which are usually distributed in crowded urban environments. Especially in busy urban streets or traffic-intensive areas, there may be multiple similar or identical objects in the field of view at the same time. For example, multiple similar vehicles are often simultaneously presented in one street. In this case, it is difficult to accurately distinguish the target object based on appearance features alone, and it is necessary to rely on the spatial relationship between objects, such as relative position, distance, and direction, to achieve precise positioning and recognition.", + "bbox": [ + 102, + 674, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address above issues, we propose the first dataset designed specifically for the AerialVG task. The dataset consists of 5,000 high-resolution images covering a variety of", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "complex urban environments and contains a rich variety of object types, such as pedestrians, vehicles, and other common urban objects. In addition, the dataset includes nearly 50k manual annotations. These annotations not only provide the bounding box of each object but also include natural language descriptions related to the object, which can help the model understand the positional relationship between objects. With these high-quality annotations, a comprehensive and efficient test platform is provided for the AerialVG task.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aiming at the core challenges in the AerialVG task, we also design an innovative model that can accurately locate objects based on the positional relation between them. Specifically, a Hierarchical Cross-Attention is first proposed to help the model focus more effectively on regions where objects are most likely to appear, thus addressing the issue of vast visual areas. Furthermore, we devise a Relation-Aware Grounding module to understand spatial information such as the relative position, direction, and distance between objects, effectively distinguishing multiple similar objects.", + "bbox": [ + 511, + 242, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, the main contributions of this work are summarized as follows:", + "bbox": [ + 511, + 407, + 903, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the AerialVG task and introduce the first dataset, providing a valuable resource for advancing research in aerial visual grounding.", + "- We present a novel AerialVG model that effectively addresses two key challenges, i.e., the large field of view with small-sized objects and the complex distribution and spatial relationships between objects.", + "- We conduct extensive experiments to thoroughly validate the effectiveness of our proposed model, demonstrating its superior performance." + ], + "bbox": [ + 513, + 439, + 903, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 603, + 653, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Grounding. Visual Grounding (VG) is a fundamental task that requires identifying specific objects in an image based on natural language descriptions. This task has made some progress in both 2D [5, 25] and 3D [17, 18, 40-42, 50, 53] fields. Existing visual grounding methods can be broadly divided into two categories, i.e., (1) two-stage methods[13, 28, 45] and (2) one-stage methods[27, 38, 39]. The former first generates region-level candidate boxes and then uses language to select the best matching region. The latter directly interacts with visual and language cues through bounding boxes, avoiding the computationally intensive proposal box generation and region feature extraction steps present in two-stage methods. Several recent methods[8, 10, 23, 25] have adopted the transformer structure for multimodal fusion, thus achieving better performance. In addition, with the booming development of vision language models (VLMs), some studies [46, 47, 51] have improved the performance of vision grounding by fine", + "bbox": [ + 511, + 628, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5178", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/774d18a672ad3737cfeea57f1d4c1320139d43c62141fa616016c4c99c3b0e65.jpg", + "image_caption": [ + "Figure 2. Example of AerialVG Dataset. There are many red cars in the picture. In order to achieve accurate object positioning, we need to use the positional relationship of surrounding auxiliary objects to assist reasoning. To this end, we accurately annotated the data to ensure that the spatial relationship between objects can be fully captured, thereby improving the accuracy of AerialVG." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tuning pre-trained VLMs, making full use of the knowledge and advantages of pre-trained models. These advances have improved both the accuracy and efficiency of VG systems, enabling more complex applications. However, they still cannot solve the spatial relationship problem in AerialVG.", + "bbox": [ + 88, + 470, + 480, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language Models. Vision-Language Models (VLMs) integrate visual and linguistic information to perform tasks that require both image understanding and natural language processing. Typically, they use pretrained visual backbones to encode images, and large-scale language models to interpret textual inputs. The visual and textual information is then fused through mechanisms like cross-attention layers. Recent advancements, such as GPT4o [32], LLaVA [24], and InstructBLIP [6], have achieved significant progress in language-guided navigation and visual reasoning. VLMs have also been adapted to a variety of domains. Specifically, robot action generation produces robot actions based on textual instructions [37, 54], vision language navigation guides UAVs in aerial environments [14, 26], 3D generation creates and understands 3D data based on language inputs [43, 48]. Moreover, VLMs are increasingly used in video understanding, biomedical imaging, and remote sensing, demonstrating their versatility across different tasks. These developments highlight the growing potential of VLMs in various real-world applications.", + "bbox": [ + 89, + 549, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Reasoning Representation. Spatial reasoning, traditionally part of broader tasks like SLAM [1, 11] or", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "depth estimation [12], has seen increasing interest in the context of VLMs. Earlier research focused on explicit spatial scene representations, such as spatial memory [15, 16] or spatial scene graphs [19, 44], to help machines understand and reason about spatial relations. Recent work has further advanced the spatial reasoning capabilities of VLMs. Some approaches leverage 3D information to enhance VLM performance. [3, 20, 21, 34] focus on reconstructing scenes from multi-view images and integrating dense semantic features to improve these representations. These 3D representations are then fused into LLMs for better spatial understanding. However, multi-view images are not always available, and the explicit construction of scenes using dense semantic features can be computationally intensive. Furthermore, the gap between 3D representations and language can often degrade performance. In contrast, other approaches utilize 2D VLMs to understand spatial relationships and measure distances. Unlike the 3D methods, spatial understanding in these models is implicitly encoded, allowing the VLM to directly address spatial reasoning tasks without requiring explicit 3D representations or scene graphs, as seen in works like [2, 4].", + "bbox": [ + 511, + 470, + 906, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 511, + 815, + 602, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we develop the AerialVG dataset as shown in Figure 2. This dataset contains 5,000 high-definition images taken from UAVs. We annotated them to ensure that they are suitable for aerial visual grounding task. In addition, we", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5179", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db38ad66f5a9790cf22b7864ff1ae6d839bd8fdf5212aaad48be5d69ca7eb536.jpg", + "image_caption": [ + "Figure 3. Resolution distribution and word cloud of AerialVG dataset. Most of the images in AerialVG dataset are high-resolution, and the most frequently appearing words are vehicle type, color, and location." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 480, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "provide a specially designed natural language description for each object, which details the objects in the image and their relative positions. The AerialVG dataset provides a solid foundation for fine-grained object grounding in complex scenes from a high-altitude perspective, and especially lays a new research direction for the efficient grounding and understanding tasks of intelligent UAVs.", + "bbox": [ + 89, + 382, + 483, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Dataset generation", + "text_level": 1, + "bbox": [ + 89, + 500, + 272, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The images in our AerialVG dataset are sourced from the VisDrone2019 dataset [56]. VisDrone consists of 288 video clips (261,908 frames) and 10,209 static images captured by various UAV-mounted cameras across 14 cities, covering a range of environments such as urban and rural areas. This diversity in scene type, location, and conditions (including different weather and lighting) offers a robust foundation for visual grounding tasks. The dataset includes more than 2.6 million manually annotated bounding boxes, highlighting common objects like pedestrians, vehicles, and bicycles, with attributes such as visibility, object class, and occlusion.", + "bbox": [ + 89, + 521, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In UAV-view images, buildings and roads occupy most of the visual space, while the primary objects of interest are pedestrians and vehicles. They often exhibit similar appearances. Relying solely on coarse-grained categories is insufficient for accurate object localization. Therefore, the annotations must capture fine-grained object features, e.g., color, vehicle type, clothing, and consider the spatial relationships between target objects and prominent surrounding elements. Due to the complexity of visual annotation, we opted for manual annotation to ensure precision.", + "bbox": [ + 89, + 688, + 482, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our manual annotation process follows these principles:", + "bbox": [ + 109, + 839, + 480, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Observer-perspective spatial relationships. For the convenience of annotation and unification, our position relation is described from the first-person perspective and", + "bbox": [ + 91, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/57d33ed39dda29c4b25cb4d7f8897c8b5ed8ecab13eb9a78c64c83f07bb4ad04.jpg", + "image_caption": [ + "Figure 4. Comparison between different datasets. Words refers to the average number of words in the text; Objects refer to the average number of objects in each image; Area refers to the proportion of irrelevant areas; Resolution refers to the maximum resolution of the dataset images and Captions refers to the average number of annotations contained in each image." + ], + "image_footnote": [], + "bbox": [ + 555, + 90, + 864, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is divided into eight categories, i.e., above, below, left, right, top left, top right, bottom left, and bottom right. As shown in Figure 2 A, the white bus is above the red car with a white roof, not on the right.", + "bbox": [ + 524, + 446, + 906, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Selection of prominent auxiliary objects. When locating the target object, we give priority to selecting the closest object as the auxiliary object to improve positioning accuracy. To ensure simplicity, if one auxiliary object is sufficient to locate the target, no redundant description will be provided. For instance, as shown in Figure 2 D, the target object is the white sedan, and we select the blue SUV as the auxiliary object, while disregarding the red car at the top left.", + "bbox": [ + 513, + 507, + 908, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset statistics", + "text_level": 1, + "bbox": [ + 513, + 652, + 679, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This subsection presents key statistics of the AerialVG dataset and offers a comparative comparison with existing visual grounding datasets.", + "bbox": [ + 511, + 674, + 906, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows the resolution distribution and word cloud for the AerialVG dataset. Unlike the RefCOCO series, where over $70\\%$ of images have a resolution around $600 \\times 650$ , the AerialVG dataset consists mostly of images with resolutions greater than $1000 \\times 500$ . This higher resolution and wider field of view make the Aerial VG task more challenging.", + "bbox": [ + 511, + 719, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The word cloud analysis highlights frequent mentions of vehicle types, colors, and spatial positions. These annotations emphasize the importance of understanding not only object identification but also their relative locations in the environment. This rich annotation style is critical for tasks", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5180", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "like UAV navigation, emergency response, and large-scale monitoring, where spatial relationships are essential for accurate decision-making.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 compares the AerialVG dataset with existing visual grounding datasets across three key aspects, i.e., the average number of words in the text, the average number of objects located in the image, and the proportion of irrelevant regions. In all three aspects, the AerialVG dataset demonstrates a higher level of challenge, highlighting its complexity and suitability for more advanced visual grounding tasks.", + "bbox": [ + 89, + 136, + 480, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 270, + 181, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 5, we take GroundingDINO [25] as the baseline and design an end-to-end model for AerialVG. GroundingDINO [25] is built on a transformer architecture, where the encoder stacks self-attention and cross-attention layers to fuse image and text features. After fusing these multimodal features, it selects the most relevant features as queries, which are highly similar to the text. These queries are then decoded in the decoder, and the final bounding boxes and logits are generated by the bbox head and class head, respectively.", + "bbox": [ + 89, + 295, + 482, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike GroundingDINO [25], after embedding the image into hierarchical features using Swin-Transformer [29], we fuse these features with text features encoded by BERT [9] via Hierarchical Cross Attention, enabling the model to focus more on potential regions where objects are likely to appear. After the Decoder, we introduce the Relation-Aware Grounding module to capture the positional relationships between different objects.", + "bbox": [ + 89, + 446, + 482, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Hierarchical Cross-Attention", + "text_level": 1, + "bbox": [ + 89, + 575, + 349, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In aerial view images, high resolution often comes with large portions of the scene occupied by buildings, roads, and other background elements, reducing the focus on key objects. To address this, we propose Hierarchical Cross-Attention, a mechanism designed to help the model more efficiently attend to relevant areas within the scene.", + "bbox": [ + 89, + 598, + 480, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the hierarchical feature representation enabled by Swin-Transformer[29], image features are organized in a layered manner. However, traditional cross-attention methods typically compute a complete attention map by flattening the features for processing, which can lead to information loss during the subsequent fusion process. In hierarchical structures, features at the same spatial location across different layers still represent the same object. Therefore, if a high-attention region is identified in one layer based on textual cues, this region should receive increased attention in subsequent layers. Conversely, if a region shows low attention, it can be partially disregarded in the next layer, enabling the model to prioritize more relevant areas effectively.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 5, after calculating the attention map between textual and image features, we retain their hierarchical structure. As the layer number increases, the attention map will also become smaller, making small objects receive insufficient attention in the higher layers. To alleviate this problem, we first convolve the attention map of the first layer to make it consistent with the shape of the last layer map, and fuse the result with it.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {n} ^ {*} = (1 - \\beta) \\times A _ {n} + \\beta \\times C \\left(A _ {1}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 227, + 903, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C(\\cdot)$ denotes the convolution operation, $\\beta$ is a hyperparameter, $A_{j}$ represents the attention map of the j-th layer. The processed attention map is then combined with the highest layer's attention map $A_{n}$ , ensuring that even small objects are given adequate focus. After that, we use Equation 2 to enhance the attention map in lower layers based on the higher layers.", + "bbox": [ + 511, + 251, + 903, + 357 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} ^ {*} = (1 - \\alpha) \\times A _ {i} + \\alpha \\times F \\left(A _ {i + 1}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 369, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we use bilinear interpolation $F(\\cdot)$ , to expand the attention map $A_{i+1}$ of the $i+1$ layer to the shape of the $i$ layer $A_i$ . The expanded attention map is then weighted by the hyper-parameter $\\alpha$ and combined with $A_i$ to form the updated $A_i^*$ . This approach effectively achieves our goal without introducing additional parameters, yielding improvements in performance.", + "bbox": [ + 511, + 398, + 905, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Relation-Aware Grounding", + "text_level": 1, + "bbox": [ + 511, + 515, + 759, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following GroundingDINO [25], we first perform language-guided query selection after the encoder, and then pass the selected queries into the decoder. In the decoder, we perform cross-attention on the query separately with the image features and text features to enhance the model's ability to localize objects. However, experiments showed that these features did not adequately capture the positional relations described in the text.", + "bbox": [ + 511, + 537, + 905, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We draw inspiration from certain 3D visual grounding methods for determining spatial relationships. However, those methods generally require explicit bounding boxes and clearly defined spatial relationships, rather than leveraging full textual descriptions. Our approach circumvents these requirements while still achieving competitive results.", + "bbox": [ + 511, + 657, + 905, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we design a Relation-Aware Grounding module to capture positional relation information. As shown in figure 5, after passing through the decoder, the fused features already contain the positional information of each object. Assuming we have $m$ features before the Relation-Aware Grounding module, we then concatenate every two of the $m$ features in pairs to obtain $m^2$ features, which effectively captures the pairwise relationships between the objects and enhances the model's ability to learn these interactions. The self-attention operation is performed", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/65af3e60eaadcf55f40f3575371d2e01d190cbbb934a1cda77f8e88a1fd73487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 88, + 880, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97c68f1df2aab60de409f007425f4051c968bc3c853cfbd096b0523681380ba4.jpg", + "image_caption": [ + "Figure 5. The architecture of AerialVG model. Hierarchical Cross Attention directs the model's focus to potential object locations, while the Relation-Aware Grounding module enables the model to perceive spatial relationships between objects." + ], + "image_footnote": [], + "bbox": [ + 112, + 266, + 377, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03908c1c6a45192f2301b9632b527344740af3ad632cf4e5d209ad969d305d17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 266, + 867, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "on $m$ concatenated features in the same column to perceive the positional relation between objects. Subsequently, we conduct cross-attention with $m^2$ queries and the text features. This step enables the model to better align and integrate both visual and textual information, enhancing its understanding of the contextual relationships between objects.", + "bbox": [ + 88, + 450, + 480, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we first map the last state of the hidden layer to the specified feature dimension through the Relation Head. On this basis, we calculate the similarity between these features and the text features. The calculated similarity scores are then subjected to a softmax operation to generate a set of normalized similarity values. These normalized values represent the relationship score between each pair of objects, that is, how well they match the relationship described in the text. In this way, we can get an $m \\times m$ relationship matrix, which represents the relationship score between each pair of objects. We take the maximum value from each row of this relationship matrix as the final relation score for each object. We combine the relation scores with the scores obtained from the query through the class head to form the final logits.", + "bbox": [ + 89, + 556, + 482, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Loss Function", + "text_level": 1, + "bbox": [ + 89, + 790, + 235, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our loss function consists of three parts, i.e., $\\mathcal{L}_{\\mathrm{reg}}$ , $\\mathcal{L}_{\\mathrm{cls}}$ and $\\mathcal{L}_{\\mathrm{rel}}$ as shown in Equation 3.", + "bbox": [ + 89, + 811, + 482, + 843 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e l}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 849, + 480, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous work on GroundingDINO, we employ L1 loss and GIOU loss for bounding box regression in $\\mathcal{L}_{\\mathrm{reg}}$ .", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For classification, we adopt a contrastive loss between predicted objects and language tokens as $\\mathcal{L}_{\\mathrm{cls}}$ , in line with the GLIP [52] approach.", + "bbox": [ + 511, + 450, + 905, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we compute a Relation Loss based on the predicted relationships between object pairs.", + "bbox": [ + 511, + 497, + 905, + 529 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e l}} = - \\log \\left(\\frac {e ^ {m _ {i , j}}}{\\sum_ {i , j} e ^ {m _ {i , j}}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 532, + 903, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where Hungarian matching is performed to identify the corresponding target object $m_{i}$ and the auxiliary object $m_{j}$ . In this way, the element $m_{i,j}$ of the $m \\times m$ relationship matrix represents the ground truth relation between the objects. We then use Equation 4 to maximize the score at the corresponding position $m_{i,j}$ in the relation matrix.", + "bbox": [ + 511, + 577, + 905, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 685, + 637, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 710, + 730, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed AerialVG model, utilizes Swin-L [29] as the image backbone. For text representation, we employ the BERT-base [9] from Hugging Face as the text backbone. Following Grounding DINO, we extract four levels of image feature scales from the image backbone, ranging from $4 \\times$ to $32 \\times$ . We set the maximum number of text tokens to 256. We set $N = 6$ for Encoder and Decoder layers, while the Relation-Aware Grounding module consists of three layers. The final loss function consists of multiple components, i.e., classification loss (or contrastive loss), L1 loss, GIoU loss, and relation loss, with corresponding weights set", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5182", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to 1.0, 5.0, 2.0, and 1.0, respectively. The hyper-parameters $\\alpha$ and $\\beta$ are set to 0.2 and 0.3. We conduct extensive experiments on the AerialVG dataset and the RefCOCO, RefCOCO+, and RefCOCOg datasets [30]. Additionally, we perform ablation studies to demonstrate the effectiveness of our model design.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model training consists of two stages. Initially, we train the model without the Relation-aware Grounding module, using existing VG datasets following Grounding DINO [25] to achieve convergence. Once the model has converged, we freeze all other parameters and train the Relation-Aware Grounding module using our AerialVG dataset only, allowing it to specifically capture inter-object relationships. More details can be found in our supplementary material.", + "bbox": [ + 89, + 181, + 480, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluation metric", + "text_level": 1, + "bbox": [ + 89, + 328, + 264, + 343 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In AerialVG, object homogeneity is a significant challenge, as many objects share nearly identical visual characteristics. A model that performs well on Top-5 Accuracy but poorly on Top-1 Accuracy likely relies primarily on visual similarity rather than understanding the spatial relationships or contextual cues that distinguish between objects. By analyzing both Top-1 and Top-5 Accuracy, we can diagnose specific failure modes of the model. Model with both high Top-1 and Top-5 Accuracy demonstrates strong object discrimination and contextual understanding ability, making it more reliable for AerialVG-related applications.", + "bbox": [ + 89, + 349, + 482, + 517 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/56ccb445e8aee6d3876f24bc79d42dd9e961264fd391066e605ae0fbaecbca85.jpg", + "table_caption": [ + "Table 1. Performance comparison on the AerialVG dataset." + ], + "table_footnote": [], + "table_body": "
MethodSetting
Zero-ShotFine-Tuning
Top 1Top 5Top 1Top 5
TranVG[8]2.573.6511.5313.68
Dynamic-MDETR[7]2.614.6319.8729.87
SimVG[5]3.614.6820.9431.32
Grounding DINO[25]11.1013.7929.3678.87
AerialVG(Ours)--50.0187.00
", + "bbox": [ + 91, + 556, + 491, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 89, + 726, + 184, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents the evaluation results of different models on the AerialVG dataset. In the Zero-Shot setting, existing VG models exhibit limited performance, with both Top-1 Accuracy and Top-5 Accuracy hovering around $10\\%$ . This result highlights the challenges posed by high-resolution aerial imagery and the unique characteristics of the aerial viewpoint, which are not well-represented in existing datasets. The poor performance of these models further underscores the necessity of AerialVG as a benchmark for evaluating visual grounding in aerial scenes.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After fine-tuning, Top-5 Accuracy shows a significant improvement, indicating that models can better recognize and retrieve relevant object candidates when adapted to the AerialVG. However, Top-1 Accuracy remains relatively low, suggesting that while models are able to identify visually similar objects within the scene, they struggle to distinguish the correct target based on spatial relationships and contextual cues.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison, the proposed AerialVG model demonstrates superior performance over previous approaches. It demonstrates a better understanding of object relationships in aerial images, resulting in higher Top-1 and Top-5 accuracy. This suggests that our model effectively integrates the Hierarchical Cross-Attention mechanisms and the Relation-Aware Grounding module, which are crucial for distinguishing visually similar objects based on spatial dependencies. These results validate our architectural choices and demonstrate the effectiveness of our model in addressing the unique challenges of aerial visual grounding.", + "bbox": [ + 511, + 212, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the AerialVG dataset, we also evaluate our AerialVG model on the RefCOCO[30], RefCOCO+[30], and RefCOCOg[31] datasets to assess its performance on standard visual grounding benchmarks. As shown in Table 2, our AerialVG model also presents good performance across multiple metrics on the RefCOCO series datasets, achieving comparable results with previous state-of-the-art approaches.", + "bbox": [ + 511, + 380, + 903, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings confirm that the integration of the Hierarchical Cross-Attention and Relation-Aware Grounding module does not compromise the model's ability to accurately localize individual objects. Despite being designed to enhance spatial reasoning and relational understanding, these components maintain strong object-level grounding capabilities. This highlights the versatility of our model, making it effective in both aerial imagery and conventional ground-based VG tasks.", + "bbox": [ + 511, + 501, + 903, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 650, + 663, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further validate the effectiveness of the Hierarchical Cross-Attention and the Relation-Aware Grounding module, we conduct ablation experiments, as shown in Table 3. The results demonstrate that each component plays a crucial role in improving different aspects of the model's performance.", + "bbox": [ + 511, + 672, + 903, + 762 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Hierarchical Cross-Attention significantly enhances Top-5 Accuracy, indicating that this module helps the model focus on fine-grained details within the image. By attending to multi-level image features, the model can better capture subtle differences between similar objects, leading to improved candidate retrieval.", + "- Relation-Aware Grounding module notably improves Top-1 Accuracy, demonstrating its ability to refine object selection by incorporating relational reasoning. This con" + ], + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5183", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/84938423097d2d2f7c955a68f38ef4acd8d19876ded2f0856fd7cfc45ee77cdf.jpg", + "table_caption": [ + "Table 2. Performance comparison on RefCOCO, RefCOCO+, and RefCOCOg datasets." + ], + "table_footnote": [], + "table_body": "
MethodBackboneRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
TranVG [8]ResNet-10180.3282.6778.1263.568.1855.6367.1167.44
Dynamic-MDETR [7]ResNet-5085.9788.8280.1274.8381.763.4473.1874.49
SimVG [5]ViT-L90.6192.5387.6885.3689.6179.7482.6786.8
Grounding DINO[25]Swin-L90.5693.1988.2482.7588.9575.9286.1387.02
AerialVG (Ours)Swin-L90.5893.2388.5682.8389.0276.0386.5288.04
", + "bbox": [ + 181, + 114, + 823, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cbfa7208381afa830c03e7fadee2ac001e22228898238e5660bf34de0fe983a8.jpg", + "table_caption": [ + "Table 3. Ablation study. Attn refers to hierarchical cross attention, Relation refers to relation aware grounding." + ], + "table_footnote": [], + "table_body": "
MethodAerialVG
Fine-Tuning
Top 1Top 5
Grounding DINO (Baseline)29.3678.87
Grounding DINO + Attn29.7882.32
Grounding DINO + Relation45.6379.63
Grounding DINO + Attn +Relation50.0187.00
", + "bbox": [ + 102, + 320, + 475, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "firms that the module helps disambiguate objects based on spatial and contextual relationships, rather than relying solely on visual appearance.", + "bbox": [ + 102, + 464, + 482, + 510 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4390bf43736934e2b9cbc143c7e71e175f0ae036e5469d5a0b848fd19e29ba33.jpg", + "image_caption": [ + "Figure 6. Example of Qualitative Test. Annotation: A blue car is parking on the right road with a white sedan above it. The model output score for this result is 0.70." + ], + "image_footnote": [], + "bbox": [ + 93, + 523, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 274, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better validate the generalization ability of our model, we conducted real-world experiments by capturing a set of aerial images using the DJI M30T drone. The images were taken at an altitude of 50 meters with a resolution of", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$4000 \\times 3000$ . The images are filled with various types of vehicles, posing a significant challenge to object localization. However, as shown in Figure 6, our model still successfully locates the objects accurately based on spatial relationship information, demonstrating its strong generalization capability. We will show more visualization results in the supplementary material.", + "bbox": [ + 511, + 282, + 906, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 511, + 412, + 785, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary limitation of current visual grounding models, including our AerialVG, is their inability to effectively recognize unstructured background information, such as roads and other environmental elements. While these models excel at detecting foreground objects, they often struggle to incorporate contextual information from the background, which is crucial for tasks requiring a more holistic understanding of the scene. Promoting the spatial intelligence of perception models will be a key focus of our future work.", + "bbox": [ + 511, + 440, + 908, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce AerialVG, a visual grounding task designed specifically for aerial perspectives. We construct the first real-world dataset tailored for this task, featuring high-resolution aerial imagery with densely populated scenes and complex spatial relationships. To address the challenges of fine-grained object differentiation and spatial reasoning, we propose an effective model, integrating Hierarchical Cross-Attention and a Relation-Aware Grounding module to enhance object localization accuracy. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art performance on the AerialVG dataset and competitive results on standard Ref-COCO benchmarks. Our findings highlight the importance of spatial reasoning in aerial visual grounding and establish AerialVG as a challenging benchmark for future research. We hope that our dataset and the proposed method will serve as a solid foundation for advancing aerial view perception and autonomous vision-and-language applications.", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5184", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Shanghai AI Laboratory, the National Natural Science Foundation of China (62376222), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 3", + "[2] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024. 3", + "[3] Qizhi Chen, Delin Qu, Yiwen Tang, Haoming Song, Yiting Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Freegaussian: Guidance-free controllable 3d gaussian splats with flow derivatives. arXiv preprint arXiv:2410.22070, 2024. 3", + "[4] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 3", + "[5] Ming Dai, Lingfeng Yang, Yihao Xu, Zhenhua Feng, and Wankou Yang. Simvg: A simple framework for visual grounding with decoupled multi-modal fusion. arXiv preprint arXiv:2409.17531, 2024. 2, 7, 8", + "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3", + "[7] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 7, 8", + "[8] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers, 2022. 2, 7, 8", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 5, 6", + "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation, 2021. 2", + "[11] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 3", + "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3" + ], + "bbox": [ + 91, + 212, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2", + "[14] Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. Openfly: A versatile toolchain and large-scale benchmark for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025.3", + "[15] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 3", + "[16] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4089-4098, 2018. 3", + "[17] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance, 2023. 2", + "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2", + "[19] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering, 2020. 3", + "[20] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images, 2023. 3", + "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3", + "[22] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2", + "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2", + "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3", + "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 5, 7, 8", + "[26] Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yan-ing Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs, 2023. 3", + "[27] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5185", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 2", + "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 2", + "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 5, 6", + "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 7", + "[31] Varun K. Nagaraja, Vlad I. Morariu, and Larry S. Davis. Modeling context between objects for referring expression understanding, 2016. 7", + "[32] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar Tabarak Khan, Logan Kilpatrick, Jong Wook Kim Christina Kim Yongjik Kim Jan Hendrik Kirchner Jamie Kiros Matt Knight Daniel Kokotajlo Lukasz Kondraciuk Andrew Kondrich Aris Konstantinidis Kyle Kosic Gretchen Krueger,Vishal Kuo Michael Lampe,Ikai LanTeddy Lee Jan Leike,Jade Leung,Daniel LevyChak Ming Li,Rachel LimMolly Lin Stephanie Lin Mateusz Litwin Theresa Lopez,Ryan Lowe Patricia LueAnna Makanju,Kim Mal" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "facini, Sam Manning, Todor Markov, Yaniv Markovsky, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024.3", + "[33] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11778-11795, 2023. 2", + "[34] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene rendering and control. arXiv preprint arXiv:2406.16038, 2024. 3", + "[35] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene control and rendering. In Advances in Neural Information Processing Systems, pages 12271-12292. Curran Associates, Inc., 2024. 2", + "[36] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Qizhi Chen, Dan Xu, Yiting Zhang, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19584-19594, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5186", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, and Xuelong Li. Spatialvia: Exploring spatial representations for visual-language-action model, 2025. 3", + "[38] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 2", + "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2", + "[40] Yiwen Tang, Hongbo Chen, Liyue Qian, Songhan Ge, Mingbo Zhang, and Rui Zheng. Detection of spine curve and vertebral level on ultrasound images using detr. In 2022 IEEE international ultrasonics symposium (IUS), pages 1-4. IEEE, 2022. 2", + "[41] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024.", + "[42] Yiwen Tang, Ray Zhang, Jiaming Liu, Zoey Guo, Bin Zhao, Zhigang Wang, Peng Gao, Hongsheng Li, Dong Wang, and Xuelong Li. Any2point: Empowering any-modality large models for efficient 3d understanding. In European Conference on Computer Vision, pages 456-473. Springer, 2024. 2", + "[43] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Exploring the potential of encoder-free architectures in 3d lmm's, 2025. 3", + "[44] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[45] Liwei Wang, Yin Li, Jing Huang, and Svetlana Lazebnik. Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):394–407, 2018. 2", + "[46] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 2", + "[47] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video, 2023. 2", + "[48] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. 3", + "[49] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding, 2024. 2", + "[51] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. $\\mathbf{X}^2$ -vlm: All-in-one pretrained model for vision-language tasks, 2023. 2", + "[52] Haotian* Zhang, Pengchuan* Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6", + "[53] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2", + "[54] Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé III, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2024. 3", + "[55] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10012-10022, 2020. 2", + "[56] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 4" + ], + "bbox": [ + 516, + 90, + 903, + 515 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_model.json b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6d897016aee2b9bb705af0c5c5b8b25424619975 --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.131, + 0.853, + 0.175 + ], + "angle": 0, + "content": "AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.202, + 0.832, + 0.24 + ], + "angle": 0, + "content": "Junli Liu\\*1,2 Qizhi Chen\\*2,3 Zhigang Wang\\*2 Yiwen Tang1,2 Yiting Zhang2 Chi Yan2 Dong Wang2 Xuelong Li4 Bin Zhao+1,2" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.24, + 0.838, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Northwestern Polytechnical University \\(^{2}\\)Shanghai AI Laboratory \\(^{3}\\)Zhejiang University \\(^{4}\\)TeleAI" + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.271, + 0.308, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.354, + 0.308, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.436, + 0.308, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.518, + 0.308, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.271, + 0.497, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.354, + 0.498, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.436, + 0.498, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.518, + 0.498, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.271, + 0.687, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.354, + 0.687, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.436, + 0.687, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.518, + 0.687, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.271, + 0.876, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.354, + 0.876, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.436, + 0.876, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.518, + 0.876, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.611, + 0.908, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Overview of AerialVG. This work introduces aerial visual grounding (AerialVG), a new task for intelligent drones. We provide the first dataset for this task, consisting of 50K annotations and 103K objects. Additionally, an AerialVG model is proposed to explore the positional relations among objects and achieve superior context-aware visual grounding in drone applications." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.666, + 0.327, + 0.681 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.698, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Visual grounding (VG) aims to localize target objects in an image based on natural language descriptions. In this paper, we propose AerialVG, a new task focusing on visual grounding from aerial views. Compared to traditional VG, AerialVG poses new challenges, e.g., appearance-based grounding is insufficient to distinguish among multiple visually similar objects, and positional relations should be emphasized. Besides, existing VG models struggle when applied to aerial imagery, where high-resolution images cause significant difficulties. To address these challenges, we introduce the first AerialVG dataset, consisting of 5K" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.909, + 0.863 + ], + "angle": 0, + "content": "real-world aerial images, 50K manually annotated descriptions, and 103K objects. Particularly, each annotation in AerialVG dataset contains multiple target objects annotated with relative spatial relations, requiring models to perform comprehensive spatial reasoning. Furthermore, we propose an innovative model especially for the AerialVG task, where a Hierarchical Cross-Attention is devised to focus on target regions, and a Relation-Aware Grounding module is designed to infer positional relations. Experimental results validate the effectiveness of our dataset and method, highlighting the importance of spatial reasoning in aerial visual grounding. The code will be released at https://github.com/Ideal-ljl/AerialVG." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.266, + 0.888 + ], + "angle": 0, + "content": "*Authors contributed equally" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.266, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5177" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.252 + ], + "angle": 0, + "content": "Visual Grounding (VG) is a challenging task that involves identifying specific regions or objects within an image following natural language instructions. This task requires models to deeply understand both visual and textual modalities, effectively aligning image context with corresponding linguistic expressions. VG is critical for a wide range of applications, including visual question answering, vision-and-language navigation, and human-robot interaction, etc [22, 33, 35, 36, 49, 55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.448 + ], + "angle": 0, + "content": "With the continuous progress of multimodal models, the performance of VG tasks has been significantly improved. However, existing models and datasets still have obvious limitations. Current research focuses primarily on ground objects, and the exploration of aerial visual grounding remains underdeveloped. Aerial visual grounding has significant potential applications in areas such as emergency rescue, logistics, and ecological monitoring, which calls for further research and development. Specifically, the ability to manage multiple objects of the same type and their spatial relationships in complex environments remains underdeveloped, leading to limited generalization capabilities in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.554 + ], + "angle": 0, + "content": "To this end, we propose a more challenging task, termed Aerial Visual Grounding (AerialVG). AerialVG aims to apply the visual grounding task to UAV platforms, and promote the development of UAV intelligent perception and autonomous navigation technologies. Compared with traditional VG tasks, AerialVG faces more challenges, mainly in the following two aspects:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.659 + ], + "angle": 0, + "content": "- Broad field of view and small objects. Unlike images from the ground perspective, when UAVs capture images from the air, the field of view is usually much broader, and the size of objects in the image tends to be relatively small. While a larger field of view provides more context and coverage, it also imposes challenges to locate and identify objects at a finer scale." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.66, + 0.483, + 0.675 + ], + "angle": 0, + "content": "- Complex object distribution and spatial relationship." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.554, + 0.483, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.675, + 0.483, + 0.856 + ], + "angle": 0, + "content": "From the perspective of UAVs, object types are mainly pedestrians and vehicles, which are usually distributed in crowded urban environments. Especially in busy urban streets or traffic-intensive areas, there may be multiple similar or identical objects in the field of view at the same time. For example, multiple similar vehicles are often simultaneously presented in one street. In this case, it is difficult to accurately distinguish the target object based on appearance features alone, and it is necessary to rely on the spatial relationship between objects, such as relative position, distance, and direction, to achieve precise positioning and recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address above issues, we propose the first dataset designed specifically for the AerialVG task. The dataset consists of 5,000 high-resolution images covering a variety of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "complex urban environments and contains a rich variety of object types, such as pedestrians, vehicles, and other common urban objects. In addition, the dataset includes nearly 50k manual annotations. These annotations not only provide the bounding box of each object but also include natural language descriptions related to the object, which can help the model understand the positional relationship between objects. With these high-quality annotations, a comprehensive and efficient test platform is provided for the AerialVG task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.409 + ], + "angle": 0, + "content": "Aiming at the core challenges in the AerialVG task, we also design an innovative model that can accurately locate objects based on the positional relation between them. Specifically, a Hierarchical Cross-Attention is first proposed to help the model focus more effectively on regions where objects are most likely to appear, thus addressing the issue of vast visual areas. Furthermore, we devise a Relation-Aware Grounding module to understand spatial information such as the relative position, direction, and distance between objects, effectively distinguishing multiple similar objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Overall, the main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.484 + ], + "angle": 0, + "content": "- We propose the AerialVG task and introduce the first dataset, providing a valuable resource for advancing research in aerial visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.485, + 0.905, + 0.544 + ], + "angle": 0, + "content": "- We present a novel AerialVG model that effectively addresses two key challenges, i.e., the large field of view with small-sized objects and the complex distribution and spatial relationships between objects." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.59 + ], + "angle": 0, + "content": "- We conduct extensive experiments to thoroughly validate the effectiveness of our proposed model, demonstrating its superior performance." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.44, + 0.905, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.604, + 0.655, + 0.62 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Visual Grounding. Visual Grounding (VG) is a fundamental task that requires identifying specific objects in an image based on natural language descriptions. This task has made some progress in both 2D [5, 25] and 3D [17, 18, 40-42, 50, 53] fields. Existing visual grounding methods can be broadly divided into two categories, i.e., (1) two-stage methods[13, 28, 45] and (2) one-stage methods[27, 38, 39]. The former first generates region-level candidate boxes and then uses language to select the best matching region. The latter directly interacts with visual and language cues through bounding boxes, avoiding the computationally intensive proposal box generation and region feature extraction steps present in two-stage methods. Several recent methods[8, 10, 23, 25] have adopted the transformer structure for multimodal fusion, thus achieving better performance. In addition, with the booming development of vision language models (VLMs), some studies [46, 47, 51] have improved the performance of vision grounding by fine" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.402, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Figure 2. Example of AerialVG Dataset. There are many red cars in the picture. In order to achieve accurate object positioning, we need to use the positional relationship of surrounding auxiliary objects to assist reasoning. To this end, we accurately annotated the data to ensure that the spatial relationship between objects can be fully captured, thereby improving the accuracy of AerialVG." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.482, + 0.547 + ], + "angle": 0, + "content": "tuning pre-trained VLMs, making full use of the knowledge and advantages of pre-trained models. These advances have improved both the accuracy and efficiency of VG systems, enabling more complex applications. However, they still cannot solve the spatial relationship problem in AerialVG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Vision-Language Models. Vision-Language Models (VLMs) integrate visual and linguistic information to perform tasks that require both image understanding and natural language processing. Typically, they use pretrained visual backbones to encode images, and large-scale language models to interpret textual inputs. The visual and textual information is then fused through mechanisms like cross-attention layers. Recent advancements, such as GPT4o [32], LLaVA [24], and InstructBLIP [6], have achieved significant progress in language-guided navigation and visual reasoning. VLMs have also been adapted to a variety of domains. Specifically, robot action generation produces robot actions based on textual instructions [37, 54], vision language navigation guides UAVs in aerial environments [14, 26], 3D generation creates and understands 3D data based on language inputs [43, 48]. Moreover, VLMs are increasingly used in video understanding, biomedical imaging, and remote sensing, demonstrating their versatility across different tasks. These developments highlight the growing potential of VLMs in various real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Spatial Reasoning Representation. Spatial reasoning, traditionally part of broader tasks like SLAM [1, 11] or" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.804 + ], + "angle": 0, + "content": "depth estimation [12], has seen increasing interest in the context of VLMs. Earlier research focused on explicit spatial scene representations, such as spatial memory [15, 16] or spatial scene graphs [19, 44], to help machines understand and reason about spatial relations. Recent work has further advanced the spatial reasoning capabilities of VLMs. Some approaches leverage 3D information to enhance VLM performance. [3, 20, 21, 34] focus on reconstructing scenes from multi-view images and integrating dense semantic features to improve these representations. These 3D representations are then fused into LLMs for better spatial understanding. However, multi-view images are not always available, and the explicit construction of scenes using dense semantic features can be computationally intensive. Furthermore, the gap between 3D representations and language can often degrade performance. In contrast, other approaches utilize 2D VLMs to understand spatial relationships and measure distances. Unlike the 3D methods, spatial understanding in these models is implicitly encoded, allowing the VLM to directly address spatial reasoning tasks without requiring explicit 3D representations or scene graphs, as seen in works like [2, 4]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.603, + 0.831 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In this paper, we develop the AerialVG dataset as shown in Figure 2. This dataset contains 5,000 high-definition images taken from UAVs. We annotated them to ensure that they are suitable for aerial visual grounding task. In addition, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.482, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.31, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Resolution distribution and word cloud of AerialVG dataset. Most of the images in AerialVG dataset are high-resolution, and the most frequently appearing words are vehicle type, color, and location." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.489 + ], + "angle": 0, + "content": "provide a specially designed natural language description for each object, which details the objects in the image and their relative positions. The AerialVG dataset provides a solid foundation for fine-grained object grounding in complex scenes from a high-altitude perspective, and especially lays a new research direction for the efficient grounding and understanding tasks of intelligent UAVs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.273, + 0.515 + ], + "angle": 0, + "content": "3.1. Dataset generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.688 + ], + "angle": 0, + "content": "The images in our AerialVG dataset are sourced from the VisDrone2019 dataset [56]. VisDrone consists of 288 video clips (261,908 frames) and 10,209 static images captured by various UAV-mounted cameras across 14 cities, covering a range of environments such as urban and rural areas. This diversity in scene type, location, and conditions (including different weather and lighting) offers a robust foundation for visual grounding tasks. The dataset includes more than 2.6 million manually annotated bounding boxes, highlighting common objects like pedestrians, vehicles, and bicycles, with attributes such as visibility, object class, and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.839 + ], + "angle": 0, + "content": "In UAV-view images, buildings and roads occupy most of the visual space, while the primary objects of interest are pedestrians and vehicles. They often exhibit similar appearances. Relying solely on coarse-grained categories is insufficient for accurate object localization. Therefore, the annotations must capture fine-grained object features, e.g., color, vehicle type, clothing, and consider the spatial relationships between target objects and prominent surrounding elements. Due to the complexity of visual annotation, we opted for manual annotation to ensure precision." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.84, + 0.482, + 0.855 + ], + "angle": 0, + "content": "Our manual annotation process follows these principles:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Observer-perspective spatial relationships. For the convenience of annotation and unification, our position relation is described from the first-person perspective and" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.091, + 0.865, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 4. Comparison between different datasets. Words refers to the average number of words in the text; Objects refer to the average number of objects in each image; Area refers to the proportion of irrelevant areas; Resolution refers to the maximum resolution of the dataset images and Captions refers to the average number of annotations contained in each image." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.447, + 0.907, + 0.508 + ], + "angle": 0, + "content": "is divided into eight categories, i.e., above, below, left, right, top left, top right, bottom left, and bottom right. As shown in Figure 2 A, the white bus is above the red car with a white roof, not on the right." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.508, + 0.909, + 0.645 + ], + "angle": 0, + "content": "- Selection of prominent auxiliary objects. When locating the target object, we give priority to selecting the closest object as the auxiliary object to improve positioning accuracy. To ensure simplicity, if one auxiliary object is sufficient to locate the target, no redundant description will be provided. For instance, as shown in Figure 2 D, the target object is the white sedan, and we select the blue SUV as the auxiliary object, while disregarding the red car at the top left." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.653, + 0.681, + 0.667 + ], + "angle": 0, + "content": "3.2. Dataset statistics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.719 + ], + "angle": 0, + "content": "This subsection presents key statistics of the AerialVG dataset and offers a comparative comparison with existing visual grounding datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Figure 3 shows the resolution distribution and word cloud for the AerialVG dataset. Unlike the RefCOCO series, where over \\(70\\%\\) of images have a resolution around \\(600 \\times 650\\), the AerialVG dataset consists mostly of images with resolutions greater than \\(1000 \\times 500\\). This higher resolution and wider field of view make the Aerial VG task more challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The word cloud analysis highlights frequent mentions of vehicle types, colors, and spatial positions. These annotations emphasize the importance of understanding not only object identification but also their relative locations in the environment. This rich annotation style is critical for tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5180" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "like UAV navigation, emergency response, and large-scale monitoring, where spatial relationships are essential for accurate decision-making." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Figure 4 compares the AerialVG dataset with existing visual grounding datasets across three key aspects, i.e., the average number of words in the text, the average number of objects located in the image, and the proportion of irrelevant regions. In all three aspects, the AerialVG dataset demonstrates a higher level of challenge, highlighting its complexity and suitability for more advanced visual grounding tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.271, + 0.182, + 0.287 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.446 + ], + "angle": 0, + "content": "As shown in Figure 5, we take GroundingDINO [25] as the baseline and design an end-to-end model for AerialVG. GroundingDINO [25] is built on a transformer architecture, where the encoder stacks self-attention and cross-attention layers to fuse image and text features. After fusing these multimodal features, it selects the most relevant features as queries, which are highly similar to the text. These queries are then decoded in the decoder, and the final bounding boxes and logits are generated by the bbox head and class head, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Unlike GroundingDINO [25], after embedding the image into hierarchical features using Swin-Transformer [29], we fuse these features with text features encoded by BERT [9] via Hierarchical Cross Attention, enabling the model to focus more on potential regions where objects are likely to appear. After the Decoder, we introduce the Relation-Aware Grounding module to capture the positional relationships between different objects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.351, + 0.592 + ], + "angle": 0, + "content": "4.1. Hierarchical Cross-Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.482, + 0.689 + ], + "angle": 0, + "content": "In aerial view images, high resolution often comes with large portions of the scene occupied by buildings, roads, and other background elements, reducing the focus on key objects. To address this, we propose Hierarchical Cross-Attention, a mechanism designed to help the model more efficiently attend to relevant areas within the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With the hierarchical feature representation enabled by Swin-Transformer[29], image features are organized in a layered manner. However, traditional cross-attention methods typically compute a complete attention map by flattening the features for processing, which can lead to information loss during the subsequent fusion process. In hierarchical structures, features at the same spatial location across different layers still represent the same object. Therefore, if a high-attention region is identified in one layer based on textual cues, this region should receive increased attention in subsequent layers. Conversely, if a region shows low attention, it can be partially disregarded in the next layer, enabling the model to prioritize more relevant areas effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "As shown in Figure 5, after calculating the attention map between textual and image features, we retain their hierarchical structure. As the layer number increases, the attention map will also become smaller, making small objects receive insufficient attention in the higher layers. To alleviate this problem, we first convolve the attention map of the first layer to make it consistent with the shape of the last layer map, and fuse the result with it." + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.228, + 0.905, + 0.246 + ], + "angle": 0, + "content": "\\[\nA _ {n} ^ {*} = (1 - \\beta) \\times A _ {n} + \\beta \\times C \\left(A _ {1}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.358 + ], + "angle": 0, + "content": "where \\( C(\\cdot) \\) denotes the convolution operation, \\( \\beta \\) is a hyperparameter, \\( A_{j} \\) represents the attention map of the j-th layer. The processed attention map is then combined with the highest layer's attention map \\( A_{n} \\), ensuring that even small objects are given adequate focus. After that, we use Equation 2 to enhance the attention map in lower layers based on the higher layers." + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.371, + 0.905, + 0.389 + ], + "angle": 0, + "content": "\\[\nA _ {i} ^ {*} = (1 - \\alpha) \\times A _ {i} + \\alpha \\times F \\left(A _ {i + 1}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.906, + 0.506 + ], + "angle": 0, + "content": "Specifically, we use bilinear interpolation \\( F(\\cdot) \\), to expand the attention map \\( A_{i+1} \\) of the \\( i+1 \\) layer to the shape of the \\( i \\) layer \\( A_i \\). The expanded attention map is then weighted by the hyper-parameter \\( \\alpha \\) and combined with \\( A_i \\) to form the updated \\( A_i^* \\). This approach effectively achieves our goal without introducing additional parameters, yielding improvements in performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.761, + 0.532 + ], + "angle": 0, + "content": "4.2. Relation-Aware Grounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Following GroundingDINO [25], we first perform language-guided query selection after the encoder, and then pass the selected queries into the decoder. In the decoder, we perform cross-attention on the query separately with the image features and text features to enhance the model's ability to localize objects. However, experiments showed that these features did not adequately capture the positional relations described in the text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.75 + ], + "angle": 0, + "content": "We draw inspiration from certain 3D visual grounding methods for determining spatial relationships. However, those methods generally require explicit bounding boxes and clearly defined spatial relationships, rather than leveraging full textual descriptions. Our approach circumvents these requirements while still achieving competitive results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Specifically, we design a Relation-Aware Grounding module to capture positional relation information. As shown in figure 5, after passing through the decoder, the fused features already contain the positional information of each object. Assuming we have \\( m \\) features before the Relation-Aware Grounding module, we then concatenate every two of the \\( m \\) features in pairs to obtain \\( m^2 \\) features, which effectively captures the pairwise relationships between the objects and enhances the model's ability to learn these interactions. The self-attention operation is performed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5181" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.881, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.267, + 0.378, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.267, + 0.868, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.408, + 0.908, + 0.438 + ], + "angle": 0, + "content": "Figure 5. The architecture of AerialVG model. Hierarchical Cross Attention directs the model's focus to potential object locations, while the Relation-Aware Grounding module enables the model to perceive spatial relationships between objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.482, + 0.557 + ], + "angle": 0, + "content": "on \\( m \\) concatenated features in the same column to perceive the positional relation between objects. Subsequently, we conduct cross-attention with \\( m^2 \\) queries and the text features. This step enables the model to better align and integrate both visual and textual information, enhancing its understanding of the contextual relationships between objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Finally, we first map the last state of the hidden layer to the specified feature dimension through the Relation Head. On this basis, we calculate the similarity between these features and the text features. The calculated similarity scores are then subjected to a softmax operation to generate a set of normalized similarity values. These normalized values represent the relationship score between each pair of objects, that is, how well they match the relationship described in the text. In this way, we can get an \\( m \\times m \\) relationship matrix, which represents the relationship score between each pair of objects. We take the maximum value from each row of this relationship matrix as the final relation score for each object. We combine the relation scores with the scores obtained from the query through the class head to form the final logits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.236, + 0.805 + ], + "angle": 0, + "content": "4.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.483, + 0.844 + ], + "angle": 0, + "content": "Our loss function consists of three parts, i.e., \\(\\mathcal{L}_{\\mathrm{reg}}\\), \\(\\mathcal{L}_{\\mathrm{cls}}\\) and \\(\\mathcal{L}_{\\mathrm{rel}}\\) as shown in Equation 3." + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.85, + 0.482, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e l}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Following previous work on GroundingDINO, we employ L1 loss and GIOU loss for bounding box regression in \\(\\mathcal{L}_{\\mathrm{reg}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.497 + ], + "angle": 0, + "content": "For classification, we adopt a contrastive loss between predicted objects and language tokens as \\(\\mathcal{L}_{\\mathrm{cls}}\\), in line with the GLIP [52] approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.53 + ], + "angle": 0, + "content": "Additionally, we compute a Relation Loss based on the predicted relationships between object pairs." + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e l}} = - \\log \\left(\\frac {e ^ {m _ {i , j}}}{\\sum_ {i , j} e ^ {m _ {i , j}}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.67 + ], + "angle": 0, + "content": "where Hungarian matching is performed to identify the corresponding target object \\( m_{i} \\) and the auxiliary object \\( m_{j} \\). In this way, the element \\( m_{i,j} \\) of the \\( m \\times m \\) relationship matrix represents the ground truth relation between the objects. We then use Equation 4 to maximize the score at the corresponding position \\( m_{i,j} \\) in the relation matrix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.686, + 0.638, + 0.703 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.731, + 0.727 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our proposed AerialVG model, utilizes Swin-L [29] as the image backbone. For text representation, we employ the BERT-base [9] from Hugging Face as the text backbone. Following Grounding DINO, we extract four levels of image feature scales from the image backbone, ranging from \\(4 \\times\\) to \\(32 \\times\\). We set the maximum number of text tokens to 256. We set \\(N = 6\\) for Encoder and Decoder layers, while the Relation-Aware Grounding module consists of three layers. The final loss function consists of multiple components, i.e., classification loss (or contrastive loss), L1 loss, GIoU loss, and relation loss, with corresponding weights set" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "to 1.0, 5.0, 2.0, and 1.0, respectively. The hyper-parameters \\(\\alpha\\) and \\(\\beta\\) are set to 0.2 and 0.3. We conduct extensive experiments on the AerialVG dataset and the RefCOCO, RefCOCO+, and RefCOCOg datasets [30]. Additionally, we perform ablation studies to demonstrate the effectiveness of our model design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Our model training consists of two stages. Initially, we train the model without the Relation-aware Grounding module, using existing VG datasets following Grounding DINO [25] to achieve convergence. Once the model has converged, we freeze all other parameters and train the Relation-Aware Grounding module using our AerialVG dataset only, allowing it to specifically capture inter-object relationships. More details can be found in our supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.266, + 0.344 + ], + "angle": 0, + "content": "5.2. Evaluation metric" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.518 + ], + "angle": 0, + "content": "In AerialVG, object homogeneity is a significant challenge, as many objects share nearly identical visual characteristics. A model that performs well on Top-5 Accuracy but poorly on Top-1 Accuracy likely relies primarily on visual similarity rather than understanding the spatial relationships or contextual cues that distinguish between objects. By analyzing both Top-1 and Top-5 Accuracy, we can diagnose specific failure modes of the model. Model with both high Top-1 and Top-5 Accuracy demonstrates strong object discrimination and contextual understanding ability, making it more reliable for AerialVG-related applications." + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.532, + 0.462, + 0.545 + ], + "angle": 0, + "content": "Table 1. Performance comparison on the AerialVG dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.557, + 0.493, + 0.701 + ], + "angle": 0, + "content": "
MethodSetting
Zero-ShotFine-Tuning
Top 1Top 5Top 1Top 5
TranVG[8]2.573.6511.5313.68
Dynamic-MDETR[7]2.614.6319.8729.87
SimVG[5]3.614.6820.9431.32
Grounding DINO[25]11.1013.7929.3678.87
AerialVG(Ours)--50.0187.00
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.185, + 0.743 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the evaluation results of different models on the AerialVG dataset. In the Zero-Shot setting, existing VG models exhibit limited performance, with both Top-1 Accuracy and Top-5 Accuracy hovering around \\(10\\%\\). This result highlights the challenges posed by high-resolution aerial imagery and the unique characteristics of the aerial viewpoint, which are not well-represented in existing datasets. The poor performance of these models further underscores the necessity of AerialVG as a benchmark for evaluating visual grounding in aerial scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "After fine-tuning, Top-5 Accuracy shows a significant improvement, indicating that models can better recognize and retrieve relevant object candidates when adapted to the AerialVG. However, Top-1 Accuracy remains relatively low, suggesting that while models are able to identify visually similar objects within the scene, they struggle to distinguish the correct target based on spatial relationships and contextual cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.379 + ], + "angle": 0, + "content": "In comparison, the proposed AerialVG model demonstrates superior performance over previous approaches. It demonstrates a better understanding of object relationships in aerial images, resulting in higher Top-1 and Top-5 accuracy. This suggests that our model effectively integrates the Hierarchical Cross-Attention mechanisms and the Relation-Aware Grounding module, which are crucial for distinguishing visually similar objects based on spatial dependencies. These results validate our architectural choices and demonstrate the effectiveness of our model in addressing the unique challenges of aerial visual grounding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.905, + 0.501 + ], + "angle": 0, + "content": "In addition to the AerialVG dataset, we also evaluate our AerialVG model on the RefCOCO[30], RefCOCO+[30], and RefCOCOg[31] datasets to assess its performance on standard visual grounding benchmarks. As shown in Table 2, our AerialVG model also presents good performance across multiple metrics on the RefCOCO series datasets, achieving comparable results with previous state-of-the-art approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.905, + 0.638 + ], + "angle": 0, + "content": "These findings confirm that the integration of the Hierarchical Cross-Attention and Relation-Aware Grounding module does not compromise the model's ability to accurately localize individual objects. Despite being designed to enhance spatial reasoning and relational understanding, these components maintain strong object-level grounding capabilities. This highlights the versatility of our model, making it effective in both aerial imagery and conventional ground-based VG tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.665, + 0.667 + ], + "angle": 0, + "content": "5.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.905, + 0.763 + ], + "angle": 0, + "content": "To further validate the effectiveness of the Hierarchical Cross-Attention and the Relation-Aware Grounding module, we conduct ablation experiments, as shown in Table 3. The results demonstrate that each component plays a crucial role in improving different aspects of the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.854 + ], + "angle": 0, + "content": "- Hierarchical Cross-Attention significantly enhances Top-5 Accuracy, indicating that this module helps the model focus on fine-grained details within the image. By attending to multi-level image features, the model can better capture subtle differences between similar objects, leading to improved candidate retrieval." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Relation-Aware Grounding module notably improves Top-1 Accuracy, demonstrating its ability to refine object selection by incorporating relational reasoning. This con" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5183" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.238, + 0.089, + 0.759, + 0.104 + ], + "angle": 0, + "content": "Table 2. Performance comparison on RefCOCO, RefCOCO+, and RefCOCOg datasets." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.115, + 0.825, + 0.26 + ], + "angle": 0, + "content": "
MethodBackboneRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
TranVG [8]ResNet-10180.3282.6778.1263.568.1855.6367.1167.44
Dynamic-MDETR [7]ResNet-5085.9788.8280.1274.8381.763.4473.1874.49
SimVG [5]ViT-L90.6192.5387.6885.3689.6179.7482.6786.8
Grounding DINO[25]Swin-L90.5693.1988.2482.7588.9575.9286.1387.02
AerialVG (Ours)Swin-L90.5893.2388.5682.8389.0276.0386.5288.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Table 3. Ablation study. Attn refers to hierarchical cross attention, Relation refers to relation aware grounding." + }, + { + "type": "table", + "bbox": [ + 0.103, + 0.321, + 0.476, + 0.451 + ], + "angle": 0, + "content": "
MethodAerialVG
Fine-Tuning
Top 1Top 5
Grounding DINO (Baseline)29.3678.87
Grounding DINO + Attn29.7882.32
Grounding DINO + Relation45.6379.63
Grounding DINO + Attn +Relation50.0187.00
" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.465, + 0.483, + 0.511 + ], + "angle": 0, + "content": "firms that the module helps disambiguate objects based on spatial and contextual relationships, rather than relying solely on visual appearance." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.524, + 0.483, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Figure 6. Example of Qualitative Test. Annotation: A blue car is parking on the right road with a white sedan above it. The model output score for this result is 0.70." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.275, + 0.833 + ], + "angle": 0, + "content": "5.5. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To better validate the generalization ability of our model, we conducted real-world experiments by capturing a set of aerial images using the DJI M30T drone. The images were taken at an altitude of 50 meters with a resolution of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.39 + ], + "angle": 0, + "content": "\\(4000 \\times 3000\\). The images are filled with various types of vehicles, posing a significant challenge to object localization. However, as shown in Figure 6, our model still successfully locates the objects accurately based on spatial relationship information, demonstrating its strong generalization capability. We will show more visualization results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.787, + 0.429 + ], + "angle": 0, + "content": "6. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.909, + 0.579 + ], + "angle": 0, + "content": "The primary limitation of current visual grounding models, including our AerialVG, is their inability to effectively recognize unstructured background information, such as roads and other environmental elements. While these models excel at detecting foreground objects, they often struggle to incorporate contextual information from the background, which is crucial for tasks requiring a more holistic understanding of the scene. Promoting the spatial intelligence of perception models will be a key focus of our future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.601, + 0.634, + 0.617 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce AerialVG, a visual grounding task designed specifically for aerial perspectives. We construct the first real-world dataset tailored for this task, featuring high-resolution aerial imagery with densely populated scenes and complex spatial relationships. To address the challenges of fine-grained object differentiation and spatial reasoning, we propose an effective model, integrating Hierarchical Cross-Attention and a Relation-Aware Grounding module to enhance object localization accuracy. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art performance on the AerialVG dataset and competitive results on standard Ref-COCO benchmarks. Our findings highlight the importance of spatial reasoning in aerial visual grounding and establish AerialVG as a challenging benchmark for future research. We hope that our dataset and the proposed method will serve as a solid foundation for advancing aerial view perception and autonomous vision-and-language applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5184" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work is supported by the Shanghai AI Laboratory, the National Natural Science Foundation of China (62376222), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.284, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.34, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[3] Qizhi Chen, Delin Qu, Yiwen Tang, Haoming Song, Yiting Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Freegaussian: Guidance-free controllable 3d gaussian splats with flow derivatives. arXiv preprint arXiv:2410.22070, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.396, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[4] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.466, + 0.485, + 0.522 + ], + "angle": 0, + "content": "[5] Ming Dai, Lingfeng Yang, Yihao Xu, Zhenhua Feng, and Wankou Yang. Simvg: A simple framework for visual grounding with decoupled multi-modal fusion. arXiv preprint arXiv:2409.17531, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.522, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.592, + 0.485, + 0.662 + ], + "angle": 0, + "content": "[7] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.663, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.705, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[11] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.213, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[14] Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. Openfly: A versatile toolchain and large-scale benchmark for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[15] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[16] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4089-4098, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[17] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.906, + 0.472 + ], + "angle": 0, + "content": "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[19] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[20] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.628 + ], + "angle": 0, + "content": "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[22] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[26] Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yan-ing Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[27] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.483, + 0.331 + ], + "angle": 0, + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[31] Varun K. Nagaraja, Vlad I. Morariu, and Larry S. Davis. Modeling context between objects for referring expression understanding, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[32] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar Tabarak Khan, Logan Kilpatrick, Jong Wook Kim Christina Kim Yongjik Kim Jan Hendrik Kirchner Jamie Kiros Matt Knight Daniel Kokotajlo Lukasz Kondraciuk Andrew Kondrich Aris Konstantinidis Kyle Kosic Gretchen Krueger,Vishal Kuo Michael Lampe,Ikai LanTeddy Lee Jan Leike,Jade Leung,Daniel LevyChak Ming Li,Rachel LimMolly Lin Stephanie Lin Mateusz Litwin Theresa Lopez,Ryan Lowe Patricia LueAnna Makanju,Kim Mal" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.546, + 0.092, + 0.905, + 0.619 + ], + "angle": 0, + "content": "facini, Sam Manning, Todor Markov, Yaniv Markovsky, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[33] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11778-11795, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[34] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene rendering and control. arXiv preprint arXiv:2406.16038, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[35] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene control and rendering. In Advances in Neural Information Processing Systems, pages 12271-12292. Curran Associates, Inc., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[36] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Qizhi Chen, Dan Xu, Yiting Zhang, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19584-19594, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5186" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[37] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, and Xuelong Li. Spatialvia: Exploring spatial representations for visual-language-action model, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[38] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.253, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[40] Yiwen Tang, Hongbo Chen, Liyue Qian, Songhan Ge, Mingbo Zhang, and Rui Zheng. Detection of spine curve and vertebral level on ultrasound images using detr. In 2022 IEEE international ultrasonics symposium (IUS), pages 1-4. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[41] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.479 + ], + "angle": 0, + "content": "[42] Yiwen Tang, Ray Zhang, Jiaming Liu, Zoey Guo, Bin Zhao, Zhigang Wang, Peng Gao, Hongsheng Li, Dong Wang, and Xuelong Li. Any2point: Empowering any-modality large models for efficient 3d understanding. In European Conference on Computer Vision, pages 456-473. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[43] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Exploring the potential of encoder-free architectures in 3d lmm's, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[44] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[45] Liwei Wang, Yin Li, Jing Huang, and Svetlana Lazebnik. Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):394–407, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[46] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[47] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[48] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[49] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[50] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[51] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. \\(\\mathbf{X}^2\\)-vlm: All-in-one pretrained model for vision-language tasks, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[52] Haotian* Zhang, Pengchuan* Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[53] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[54] Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé III, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.458 + ], + "angle": 0, + "content": "[55] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10012-10022, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[56] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5187" + } + ] +] \ No newline at end of file diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_origin.pdf b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..de4bf6f1052bd4ad67debb8ef1acee13392eb58c --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/b41713a2-2be5-479f-9826-9a9a4762ace5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c6299560517d105ee3927f02cc274420cfe125e1be32b3441794afbb5ed706f +size 35019208 diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/full.md b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f55dfd670b1e58e03cac03280e81e2c29947771b --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/full.md @@ -0,0 +1,321 @@ +# AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations + +Junli Liu\*1,2 Qizhi Chen\*2,3 Zhigang Wang\*2 Yiwen Tang1,2 Yiting Zhang2 Chi Yan2 Dong Wang2 Xuelong Li4 Bin Zhao+1,2 + +$^{1}$ Northwestern Polytechnical University $^{2}$ Shanghai AI Laboratory $^{3}$ Zhejiang University $^{4}$ TeleAI + +![](images/0b218763d3a467aadcefb54e6c41196e0f01b577e07cd170e988f9fc842c2b40.jpg) + +![](images/19fa1d712a5c0b08add20267d0bf40dce4bd2acfe3150becd488a5a460a16298.jpg) + +![](images/4fa850aa2a7778806fa37f912e486242f488df5b1a2ad71a10bdba912a6b6db8.jpg) + +![](images/b7e0337ba00e5ab7e7281f64f48ceed83ca2c6540def82e2e7a0efecb460c7e0.jpg) +Figure 1. Overview of AerialVG. This work introduces aerial visual grounding (AerialVG), a new task for intelligent drones. We provide the first dataset for this task, consisting of 50K annotations and 103K objects. Additionally, an AerialVG model is proposed to explore the positional relations among objects and achieve superior context-aware visual grounding in drone applications. + +![](images/f7680ff1f87d71a5285eabc78ef24107466370d0d343d3b1c4032d067cdee4d6.jpg) + +![](images/59b4fd44115f67d09ded023a4b05562714d5608f50be9d00d34321876a0e64ba.jpg) + +![](images/b62f54ef27fa623c2d0a341ed924370a5d14bafb9594d12678d3b8db5eee2872.jpg) + +![](images/900f50284ace838e4e62e5190a18e215628cd6a66c380fa7618a8d63a84d6d2f.jpg) + +![](images/3752a4fd729a136778f2e39b6fad323d57250297c15796ce12ba87543956bbf9.jpg) + +![](images/dd5a8d2f549929ce89505607dd7663828e78051e3313b0ac6e3943d4a8a8e06a.jpg) + +![](images/6735695b6dd05b4bd390a0337d6e13fadea7aa79b40a2dadbbc6dad5f43dee62.jpg) + +![](images/9b584175ecc471a7f16d405ad283d7d853698d011e334cfe708cdb1ea29de179.jpg) + +![](images/2aa53f424bfb86a72aa6e0a94b12aca6611836d3153683569d18366c12620dfe.jpg) + +![](images/3e325b8e2d8e505b5a9d5fec2f9ec37e38747b7bcde1807d03e5d636931f8633.jpg) + +![](images/6123157bda65f05d2efdc5467638e1ac8e96efcabc132ba767c0fb6777d6df36.jpg) + +![](images/c095c75e6db2bb190435e03e3a49842b84f3f37065862afa09c0d0e3706de897.jpg) + +# Abstract + +Visual grounding (VG) aims to localize target objects in an image based on natural language descriptions. In this paper, we propose AerialVG, a new task focusing on visual grounding from aerial views. Compared to traditional VG, AerialVG poses new challenges, e.g., appearance-based grounding is insufficient to distinguish among multiple visually similar objects, and positional relations should be emphasized. Besides, existing VG models struggle when applied to aerial imagery, where high-resolution images cause significant difficulties. To address these challenges, we introduce the first AerialVG dataset, consisting of 5K + +real-world aerial images, 50K manually annotated descriptions, and 103K objects. Particularly, each annotation in AerialVG dataset contains multiple target objects annotated with relative spatial relations, requiring models to perform comprehensive spatial reasoning. Furthermore, we propose an innovative model especially for the AerialVG task, where a Hierarchical Cross-Attention is devised to focus on target regions, and a Relation-Aware Grounding module is designed to infer positional relations. Experimental results validate the effectiveness of our dataset and method, highlighting the importance of spatial reasoning in aerial visual grounding. The code will be released at https://github.com/Ideal-ljl/AerialVG. + +# 1. Introduction + +Visual Grounding (VG) is a challenging task that involves identifying specific regions or objects within an image following natural language instructions. This task requires models to deeply understand both visual and textual modalities, effectively aligning image context with corresponding linguistic expressions. VG is critical for a wide range of applications, including visual question answering, vision-and-language navigation, and human-robot interaction, etc [22, 33, 35, 36, 49, 55]. + +With the continuous progress of multimodal models, the performance of VG tasks has been significantly improved. However, existing models and datasets still have obvious limitations. Current research focuses primarily on ground objects, and the exploration of aerial visual grounding remains underdeveloped. Aerial visual grounding has significant potential applications in areas such as emergency rescue, logistics, and ecological monitoring, which calls for further research and development. Specifically, the ability to manage multiple objects of the same type and their spatial relationships in complex environments remains underdeveloped, leading to limited generalization capabilities in real-world scenarios. + +To this end, we propose a more challenging task, termed Aerial Visual Grounding (AerialVG). AerialVG aims to apply the visual grounding task to UAV platforms, and promote the development of UAV intelligent perception and autonomous navigation technologies. Compared with traditional VG tasks, AerialVG faces more challenges, mainly in the following two aspects: + +- Broad field of view and small objects. Unlike images from the ground perspective, when UAVs capture images from the air, the field of view is usually much broader, and the size of objects in the image tends to be relatively small. While a larger field of view provides more context and coverage, it also imposes challenges to locate and identify objects at a finer scale. +- Complex object distribution and spatial relationship. + +From the perspective of UAVs, object types are mainly pedestrians and vehicles, which are usually distributed in crowded urban environments. Especially in busy urban streets or traffic-intensive areas, there may be multiple similar or identical objects in the field of view at the same time. For example, multiple similar vehicles are often simultaneously presented in one street. In this case, it is difficult to accurately distinguish the target object based on appearance features alone, and it is necessary to rely on the spatial relationship between objects, such as relative position, distance, and direction, to achieve precise positioning and recognition. + +To address above issues, we propose the first dataset designed specifically for the AerialVG task. The dataset consists of 5,000 high-resolution images covering a variety of + +complex urban environments and contains a rich variety of object types, such as pedestrians, vehicles, and other common urban objects. In addition, the dataset includes nearly 50k manual annotations. These annotations not only provide the bounding box of each object but also include natural language descriptions related to the object, which can help the model understand the positional relationship between objects. With these high-quality annotations, a comprehensive and efficient test platform is provided for the AerialVG task. + +Aiming at the core challenges in the AerialVG task, we also design an innovative model that can accurately locate objects based on the positional relation between them. Specifically, a Hierarchical Cross-Attention is first proposed to help the model focus more effectively on regions where objects are most likely to appear, thus addressing the issue of vast visual areas. Furthermore, we devise a Relation-Aware Grounding module to understand spatial information such as the relative position, direction, and distance between objects, effectively distinguishing multiple similar objects. + +Overall, the main contributions of this work are summarized as follows: + +- We propose the AerialVG task and introduce the first dataset, providing a valuable resource for advancing research in aerial visual grounding. +- We present a novel AerialVG model that effectively addresses two key challenges, i.e., the large field of view with small-sized objects and the complex distribution and spatial relationships between objects. +- We conduct extensive experiments to thoroughly validate the effectiveness of our proposed model, demonstrating its superior performance. + +# 2. Related Work + +Visual Grounding. Visual Grounding (VG) is a fundamental task that requires identifying specific objects in an image based on natural language descriptions. This task has made some progress in both 2D [5, 25] and 3D [17, 18, 40-42, 50, 53] fields. Existing visual grounding methods can be broadly divided into two categories, i.e., (1) two-stage methods[13, 28, 45] and (2) one-stage methods[27, 38, 39]. The former first generates region-level candidate boxes and then uses language to select the best matching region. The latter directly interacts with visual and language cues through bounding boxes, avoiding the computationally intensive proposal box generation and region feature extraction steps present in two-stage methods. Several recent methods[8, 10, 23, 25] have adopted the transformer structure for multimodal fusion, thus achieving better performance. In addition, with the booming development of vision language models (VLMs), some studies [46, 47, 51] have improved the performance of vision grounding by fine + +![](images/774d18a672ad3737cfeea57f1d4c1320139d43c62141fa616016c4c99c3b0e65.jpg) +Figure 2. Example of AerialVG Dataset. There are many red cars in the picture. In order to achieve accurate object positioning, we need to use the positional relationship of surrounding auxiliary objects to assist reasoning. To this end, we accurately annotated the data to ensure that the spatial relationship between objects can be fully captured, thereby improving the accuracy of AerialVG. + +tuning pre-trained VLMs, making full use of the knowledge and advantages of pre-trained models. These advances have improved both the accuracy and efficiency of VG systems, enabling more complex applications. However, they still cannot solve the spatial relationship problem in AerialVG. + +Vision-Language Models. Vision-Language Models (VLMs) integrate visual and linguistic information to perform tasks that require both image understanding and natural language processing. Typically, they use pretrained visual backbones to encode images, and large-scale language models to interpret textual inputs. The visual and textual information is then fused through mechanisms like cross-attention layers. Recent advancements, such as GPT4o [32], LLaVA [24], and InstructBLIP [6], have achieved significant progress in language-guided navigation and visual reasoning. VLMs have also been adapted to a variety of domains. Specifically, robot action generation produces robot actions based on textual instructions [37, 54], vision language navigation guides UAVs in aerial environments [14, 26], 3D generation creates and understands 3D data based on language inputs [43, 48]. Moreover, VLMs are increasingly used in video understanding, biomedical imaging, and remote sensing, demonstrating their versatility across different tasks. These developments highlight the growing potential of VLMs in various real-world applications. + +Spatial Reasoning Representation. Spatial reasoning, traditionally part of broader tasks like SLAM [1, 11] or + +depth estimation [12], has seen increasing interest in the context of VLMs. Earlier research focused on explicit spatial scene representations, such as spatial memory [15, 16] or spatial scene graphs [19, 44], to help machines understand and reason about spatial relations. Recent work has further advanced the spatial reasoning capabilities of VLMs. Some approaches leverage 3D information to enhance VLM performance. [3, 20, 21, 34] focus on reconstructing scenes from multi-view images and integrating dense semantic features to improve these representations. These 3D representations are then fused into LLMs for better spatial understanding. However, multi-view images are not always available, and the explicit construction of scenes using dense semantic features can be computationally intensive. Furthermore, the gap between 3D representations and language can often degrade performance. In contrast, other approaches utilize 2D VLMs to understand spatial relationships and measure distances. Unlike the 3D methods, spatial understanding in these models is implicitly encoded, allowing the VLM to directly address spatial reasoning tasks without requiring explicit 3D representations or scene graphs, as seen in works like [2, 4]. + +# 3. Dataset + +In this paper, we develop the AerialVG dataset as shown in Figure 2. This dataset contains 5,000 high-definition images taken from UAVs. We annotated them to ensure that they are suitable for aerial visual grounding task. In addition, we + +![](images/db38ad66f5a9790cf22b7864ff1ae6d839bd8fdf5212aaad48be5d69ca7eb536.jpg) +Figure 3. Resolution distribution and word cloud of AerialVG dataset. Most of the images in AerialVG dataset are high-resolution, and the most frequently appearing words are vehicle type, color, and location. + +provide a specially designed natural language description for each object, which details the objects in the image and their relative positions. The AerialVG dataset provides a solid foundation for fine-grained object grounding in complex scenes from a high-altitude perspective, and especially lays a new research direction for the efficient grounding and understanding tasks of intelligent UAVs. + +# 3.1. Dataset generation + +The images in our AerialVG dataset are sourced from the VisDrone2019 dataset [56]. VisDrone consists of 288 video clips (261,908 frames) and 10,209 static images captured by various UAV-mounted cameras across 14 cities, covering a range of environments such as urban and rural areas. This diversity in scene type, location, and conditions (including different weather and lighting) offers a robust foundation for visual grounding tasks. The dataset includes more than 2.6 million manually annotated bounding boxes, highlighting common objects like pedestrians, vehicles, and bicycles, with attributes such as visibility, object class, and occlusion. + +In UAV-view images, buildings and roads occupy most of the visual space, while the primary objects of interest are pedestrians and vehicles. They often exhibit similar appearances. Relying solely on coarse-grained categories is insufficient for accurate object localization. Therefore, the annotations must capture fine-grained object features, e.g., color, vehicle type, clothing, and consider the spatial relationships between target objects and prominent surrounding elements. Due to the complexity of visual annotation, we opted for manual annotation to ensure precision. + +Our manual annotation process follows these principles: + +- Observer-perspective spatial relationships. For the convenience of annotation and unification, our position relation is described from the first-person perspective and + +![](images/57d33ed39dda29c4b25cb4d7f8897c8b5ed8ecab13eb9a78c64c83f07bb4ad04.jpg) +Figure 4. Comparison between different datasets. Words refers to the average number of words in the text; Objects refer to the average number of objects in each image; Area refers to the proportion of irrelevant areas; Resolution refers to the maximum resolution of the dataset images and Captions refers to the average number of annotations contained in each image. + +is divided into eight categories, i.e., above, below, left, right, top left, top right, bottom left, and bottom right. As shown in Figure 2 A, the white bus is above the red car with a white roof, not on the right. + +- Selection of prominent auxiliary objects. When locating the target object, we give priority to selecting the closest object as the auxiliary object to improve positioning accuracy. To ensure simplicity, if one auxiliary object is sufficient to locate the target, no redundant description will be provided. For instance, as shown in Figure 2 D, the target object is the white sedan, and we select the blue SUV as the auxiliary object, while disregarding the red car at the top left. + +# 3.2. Dataset statistics + +This subsection presents key statistics of the AerialVG dataset and offers a comparative comparison with existing visual grounding datasets. + +Figure 3 shows the resolution distribution and word cloud for the AerialVG dataset. Unlike the RefCOCO series, where over $70\%$ of images have a resolution around $600 \times 650$ , the AerialVG dataset consists mostly of images with resolutions greater than $1000 \times 500$ . This higher resolution and wider field of view make the Aerial VG task more challenging. + +The word cloud analysis highlights frequent mentions of vehicle types, colors, and spatial positions. These annotations emphasize the importance of understanding not only object identification but also their relative locations in the environment. This rich annotation style is critical for tasks + +like UAV navigation, emergency response, and large-scale monitoring, where spatial relationships are essential for accurate decision-making. + +Figure 4 compares the AerialVG dataset with existing visual grounding datasets across three key aspects, i.e., the average number of words in the text, the average number of objects located in the image, and the proportion of irrelevant regions. In all three aspects, the AerialVG dataset demonstrates a higher level of challenge, highlighting its complexity and suitability for more advanced visual grounding tasks. + +# 4. Method + +As shown in Figure 5, we take GroundingDINO [25] as the baseline and design an end-to-end model for AerialVG. GroundingDINO [25] is built on a transformer architecture, where the encoder stacks self-attention and cross-attention layers to fuse image and text features. After fusing these multimodal features, it selects the most relevant features as queries, which are highly similar to the text. These queries are then decoded in the decoder, and the final bounding boxes and logits are generated by the bbox head and class head, respectively. + +Unlike GroundingDINO [25], after embedding the image into hierarchical features using Swin-Transformer [29], we fuse these features with text features encoded by BERT [9] via Hierarchical Cross Attention, enabling the model to focus more on potential regions where objects are likely to appear. After the Decoder, we introduce the Relation-Aware Grounding module to capture the positional relationships between different objects. + +# 4.1. Hierarchical Cross-Attention + +In aerial view images, high resolution often comes with large portions of the scene occupied by buildings, roads, and other background elements, reducing the focus on key objects. To address this, we propose Hierarchical Cross-Attention, a mechanism designed to help the model more efficiently attend to relevant areas within the scene. + +With the hierarchical feature representation enabled by Swin-Transformer[29], image features are organized in a layered manner. However, traditional cross-attention methods typically compute a complete attention map by flattening the features for processing, which can lead to information loss during the subsequent fusion process. In hierarchical structures, features at the same spatial location across different layers still represent the same object. Therefore, if a high-attention region is identified in one layer based on textual cues, this region should receive increased attention in subsequent layers. Conversely, if a region shows low attention, it can be partially disregarded in the next layer, enabling the model to prioritize more relevant areas effectively. + +As shown in Figure 5, after calculating the attention map between textual and image features, we retain their hierarchical structure. As the layer number increases, the attention map will also become smaller, making small objects receive insufficient attention in the higher layers. To alleviate this problem, we first convolve the attention map of the first layer to make it consistent with the shape of the last layer map, and fuse the result with it. + +$$ +A _ {n} ^ {*} = (1 - \beta) \times A _ {n} + \beta \times C \left(A _ {1}\right), \tag {1} +$$ + +where $C(\cdot)$ denotes the convolution operation, $\beta$ is a hyperparameter, $A_{j}$ represents the attention map of the j-th layer. The processed attention map is then combined with the highest layer's attention map $A_{n}$ , ensuring that even small objects are given adequate focus. After that, we use Equation 2 to enhance the attention map in lower layers based on the higher layers. + +$$ +A _ {i} ^ {*} = (1 - \alpha) \times A _ {i} + \alpha \times F \left(A _ {i + 1}\right). \tag {2} +$$ + +Specifically, we use bilinear interpolation $F(\cdot)$ , to expand the attention map $A_{i+1}$ of the $i+1$ layer to the shape of the $i$ layer $A_i$ . The expanded attention map is then weighted by the hyper-parameter $\alpha$ and combined with $A_i$ to form the updated $A_i^*$ . This approach effectively achieves our goal without introducing additional parameters, yielding improvements in performance. + +# 4.2. Relation-Aware Grounding + +Following GroundingDINO [25], we first perform language-guided query selection after the encoder, and then pass the selected queries into the decoder. In the decoder, we perform cross-attention on the query separately with the image features and text features to enhance the model's ability to localize objects. However, experiments showed that these features did not adequately capture the positional relations described in the text. + +We draw inspiration from certain 3D visual grounding methods for determining spatial relationships. However, those methods generally require explicit bounding boxes and clearly defined spatial relationships, rather than leveraging full textual descriptions. Our approach circumvents these requirements while still achieving competitive results. + +Specifically, we design a Relation-Aware Grounding module to capture positional relation information. As shown in figure 5, after passing through the decoder, the fused features already contain the positional information of each object. Assuming we have $m$ features before the Relation-Aware Grounding module, we then concatenate every two of the $m$ features in pairs to obtain $m^2$ features, which effectively captures the pairwise relationships between the objects and enhances the model's ability to learn these interactions. The self-attention operation is performed + +![](images/65af3e60eaadcf55f40f3575371d2e01d190cbbb934a1cda77f8e88a1fd73487.jpg) + +![](images/97c68f1df2aab60de409f007425f4051c968bc3c853cfbd096b0523681380ba4.jpg) +Figure 5. The architecture of AerialVG model. Hierarchical Cross Attention directs the model's focus to potential object locations, while the Relation-Aware Grounding module enables the model to perceive spatial relationships between objects. + +![](images/03908c1c6a45192f2301b9632b527344740af3ad632cf4e5d209ad969d305d17.jpg) + +on $m$ concatenated features in the same column to perceive the positional relation between objects. Subsequently, we conduct cross-attention with $m^2$ queries and the text features. This step enables the model to better align and integrate both visual and textual information, enhancing its understanding of the contextual relationships between objects. + +Finally, we first map the last state of the hidden layer to the specified feature dimension through the Relation Head. On this basis, we calculate the similarity between these features and the text features. The calculated similarity scores are then subjected to a softmax operation to generate a set of normalized similarity values. These normalized values represent the relationship score between each pair of objects, that is, how well they match the relationship described in the text. In this way, we can get an $m \times m$ relationship matrix, which represents the relationship score between each pair of objects. We take the maximum value from each row of this relationship matrix as the final relation score for each object. We combine the relation scores with the scores obtained from the query through the class head to form the final logits. + +# 4.3. Loss Function + +Our loss function consists of three parts, i.e., $\mathcal{L}_{\mathrm{reg}}$ , $\mathcal{L}_{\mathrm{cls}}$ and $\mathcal{L}_{\mathrm{rel}}$ as shown in Equation 3. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e g}} + \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {r e l}}. \tag {3} +$$ + +Following previous work on GroundingDINO, we employ L1 loss and GIOU loss for bounding box regression in $\mathcal{L}_{\mathrm{reg}}$ . + +For classification, we adopt a contrastive loss between predicted objects and language tokens as $\mathcal{L}_{\mathrm{cls}}$ , in line with the GLIP [52] approach. + +Additionally, we compute a Relation Loss based on the predicted relationships between object pairs. + +$$ +\mathcal {L} _ {\text {r e l}} = - \log \left(\frac {e ^ {m _ {i , j}}}{\sum_ {i , j} e ^ {m _ {i , j}}}\right), \tag {4} +$$ + +where Hungarian matching is performed to identify the corresponding target object $m_{i}$ and the auxiliary object $m_{j}$ . In this way, the element $m_{i,j}$ of the $m \times m$ relationship matrix represents the ground truth relation between the objects. We then use Equation 4 to maximize the score at the corresponding position $m_{i,j}$ in the relation matrix. + +# 5. Experiment + +# 5.1. Implementation Details + +Our proposed AerialVG model, utilizes Swin-L [29] as the image backbone. For text representation, we employ the BERT-base [9] from Hugging Face as the text backbone. Following Grounding DINO, we extract four levels of image feature scales from the image backbone, ranging from $4 \times$ to $32 \times$ . We set the maximum number of text tokens to 256. We set $N = 6$ for Encoder and Decoder layers, while the Relation-Aware Grounding module consists of three layers. The final loss function consists of multiple components, i.e., classification loss (or contrastive loss), L1 loss, GIoU loss, and relation loss, with corresponding weights set + +to 1.0, 5.0, 2.0, and 1.0, respectively. The hyper-parameters $\alpha$ and $\beta$ are set to 0.2 and 0.3. We conduct extensive experiments on the AerialVG dataset and the RefCOCO, RefCOCO+, and RefCOCOg datasets [30]. Additionally, we perform ablation studies to demonstrate the effectiveness of our model design. + +Our model training consists of two stages. Initially, we train the model without the Relation-aware Grounding module, using existing VG datasets following Grounding DINO [25] to achieve convergence. Once the model has converged, we freeze all other parameters and train the Relation-Aware Grounding module using our AerialVG dataset only, allowing it to specifically capture inter-object relationships. More details can be found in our supplementary material. + +# 5.2. Evaluation metric + +In AerialVG, object homogeneity is a significant challenge, as many objects share nearly identical visual characteristics. A model that performs well on Top-5 Accuracy but poorly on Top-1 Accuracy likely relies primarily on visual similarity rather than understanding the spatial relationships or contextual cues that distinguish between objects. By analyzing both Top-1 and Top-5 Accuracy, we can diagnose specific failure modes of the model. Model with both high Top-1 and Top-5 Accuracy demonstrates strong object discrimination and contextual understanding ability, making it more reliable for AerialVG-related applications. + +Table 1. Performance comparison on the AerialVG dataset. + +
MethodSetting
Zero-ShotFine-Tuning
Top 1Top 5Top 1Top 5
TranVG[8]2.573.6511.5313.68
Dynamic-MDETR[7]2.614.6319.8729.87
SimVG[5]3.614.6820.9431.32
Grounding DINO[25]11.1013.7929.3678.87
AerialVG(Ours)--50.0187.00
+ +# 5.3. Results + +Table 1 presents the evaluation results of different models on the AerialVG dataset. In the Zero-Shot setting, existing VG models exhibit limited performance, with both Top-1 Accuracy and Top-5 Accuracy hovering around $10\%$ . This result highlights the challenges posed by high-resolution aerial imagery and the unique characteristics of the aerial viewpoint, which are not well-represented in existing datasets. The poor performance of these models further underscores the necessity of AerialVG as a benchmark for evaluating visual grounding in aerial scenes. + +After fine-tuning, Top-5 Accuracy shows a significant improvement, indicating that models can better recognize and retrieve relevant object candidates when adapted to the AerialVG. However, Top-1 Accuracy remains relatively low, suggesting that while models are able to identify visually similar objects within the scene, they struggle to distinguish the correct target based on spatial relationships and contextual cues. + +In comparison, the proposed AerialVG model demonstrates superior performance over previous approaches. It demonstrates a better understanding of object relationships in aerial images, resulting in higher Top-1 and Top-5 accuracy. This suggests that our model effectively integrates the Hierarchical Cross-Attention mechanisms and the Relation-Aware Grounding module, which are crucial for distinguishing visually similar objects based on spatial dependencies. These results validate our architectural choices and demonstrate the effectiveness of our model in addressing the unique challenges of aerial visual grounding. + +In addition to the AerialVG dataset, we also evaluate our AerialVG model on the RefCOCO[30], RefCOCO+[30], and RefCOCOg[31] datasets to assess its performance on standard visual grounding benchmarks. As shown in Table 2, our AerialVG model also presents good performance across multiple metrics on the RefCOCO series datasets, achieving comparable results with previous state-of-the-art approaches. + +These findings confirm that the integration of the Hierarchical Cross-Attention and Relation-Aware Grounding module does not compromise the model's ability to accurately localize individual objects. Despite being designed to enhance spatial reasoning and relational understanding, these components maintain strong object-level grounding capabilities. This highlights the versatility of our model, making it effective in both aerial imagery and conventional ground-based VG tasks. + +# 5.4. Ablation study + +To further validate the effectiveness of the Hierarchical Cross-Attention and the Relation-Aware Grounding module, we conduct ablation experiments, as shown in Table 3. The results demonstrate that each component plays a crucial role in improving different aspects of the model's performance. + +- Hierarchical Cross-Attention significantly enhances Top-5 Accuracy, indicating that this module helps the model focus on fine-grained details within the image. By attending to multi-level image features, the model can better capture subtle differences between similar objects, leading to improved candidate retrieval. +- Relation-Aware Grounding module notably improves Top-1 Accuracy, demonstrating its ability to refine object selection by incorporating relational reasoning. This con + +Table 2. Performance comparison on RefCOCO, RefCOCO+, and RefCOCOg datasets. + +
MethodBackboneRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
TranVG [8]ResNet-10180.3282.6778.1263.568.1855.6367.1167.44
Dynamic-MDETR [7]ResNet-5085.9788.8280.1274.8381.763.4473.1874.49
SimVG [5]ViT-L90.6192.5387.6885.3689.6179.7482.6786.8
Grounding DINO[25]Swin-L90.5693.1988.2482.7588.9575.9286.1387.02
AerialVG (Ours)Swin-L90.5893.2388.5682.8389.0276.0386.5288.04
+ +Table 3. Ablation study. Attn refers to hierarchical cross attention, Relation refers to relation aware grounding. + +
MethodAerialVG
Fine-Tuning
Top 1Top 5
Grounding DINO (Baseline)29.3678.87
Grounding DINO + Attn29.7882.32
Grounding DINO + Relation45.6379.63
Grounding DINO + Attn +Relation50.0187.00
+ +firms that the module helps disambiguate objects based on spatial and contextual relationships, rather than relying solely on visual appearance. + +![](images/4390bf43736934e2b9cbc143c7e71e175f0ae036e5469d5a0b848fd19e29ba33.jpg) +Figure 6. Example of Qualitative Test. Annotation: A blue car is parking on the right road with a white sedan above it. The model output score for this result is 0.70. + +# 5.5. Qualitative Results + +To better validate the generalization ability of our model, we conducted real-world experiments by capturing a set of aerial images using the DJI M30T drone. The images were taken at an altitude of 50 meters with a resolution of + +$4000 \times 3000$ . The images are filled with various types of vehicles, posing a significant challenge to object localization. However, as shown in Figure 6, our model still successfully locates the objects accurately based on spatial relationship information, demonstrating its strong generalization capability. We will show more visualization results in the supplementary material. + +# 6. Limitations and Future Work + +The primary limitation of current visual grounding models, including our AerialVG, is their inability to effectively recognize unstructured background information, such as roads and other environmental elements. While these models excel at detecting foreground objects, they often struggle to incorporate contextual information from the background, which is crucial for tasks requiring a more holistic understanding of the scene. Promoting the spatial intelligence of perception models will be a key focus of our future work. + +# 7. Conclusion + +In this paper, we introduce AerialVG, a visual grounding task designed specifically for aerial perspectives. We construct the first real-world dataset tailored for this task, featuring high-resolution aerial imagery with densely populated scenes and complex spatial relationships. To address the challenges of fine-grained object differentiation and spatial reasoning, we propose an effective model, integrating Hierarchical Cross-Attention and a Relation-Aware Grounding module to enhance object localization accuracy. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art performance on the AerialVG dataset and competitive results on standard Ref-COCO benchmarks. Our findings highlight the importance of spatial reasoning in aerial visual grounding and establish AerialVG as a challenging benchmark for future research. We hope that our dataset and the proposed method will serve as a solid foundation for advancing aerial view perception and autonomous vision-and-language applications. + +# Acknowledgments + +This work is supported by the Shanghai AI Laboratory, the National Natural Science Foundation of China (62376222), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001). + +# References + +[1] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 3 +[2] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024. 3 +[3] Qizhi Chen, Delin Qu, Yiwen Tang, Haoming Song, Yiting Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Freegaussian: Guidance-free controllable 3d gaussian splats with flow derivatives. arXiv preprint arXiv:2410.22070, 2024. 3 +[4] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 3 +[5] Ming Dai, Lingfeng Yang, Yihao Xu, Zhenhua Feng, and Wankou Yang. Simvg: A simple framework for visual grounding with decoupled multi-modal fusion. arXiv preprint arXiv:2409.17531, 2024. 2, 7, 8 +[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3 +[7] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 7, 8 +[8] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers, 2022. 2, 7, 8 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 5, 6 +[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation, 2021. 2 +[11] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 3 +[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3 + +[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2 +[14] Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. Openfly: A versatile toolchain and large-scale benchmark for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025.3 +[15] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 3 +[16] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4089-4098, 2018. 3 +[17] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance, 2023. 2 +[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2 +[19] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering, 2020. 3 +[20] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images, 2023. 3 +[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3 +[22] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2 +[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3 +[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 5, 7, 8 +[26] Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yan-ing Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs, 2023. 3 +[27] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C + +Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 2 +[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 2 +[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 5, 6 +[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 7 +[31] Varun K. Nagaraja, Vlad I. Morariu, and Larry S. Davis. Modeling context between objects for referring expression understanding, 2016. 7 +[32] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar Tabarak Khan, Logan Kilpatrick, Jong Wook Kim Christina Kim Yongjik Kim Jan Hendrik Kirchner Jamie Kiros Matt Knight Daniel Kokotajlo Lukasz Kondraciuk Andrew Kondrich Aris Konstantinidis Kyle Kosic Gretchen Krueger,Vishal Kuo Michael Lampe,Ikai LanTeddy Lee Jan Leike,Jade Leung,Daniel LevyChak Ming Li,Rachel LimMolly Lin Stephanie Lin Mateusz Litwin Theresa Lopez,Ryan Lowe Patricia LueAnna Makanju,Kim Mal + +facini, Sam Manning, Todor Markov, Yaniv Markovsky, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024.3 +[33] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11778-11795, 2023. 2 +[34] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene rendering and control. arXiv preprint arXiv:2406.16038, 2024. 3 +[35] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene control and rendering. In Advances in Neural Information Processing Systems, pages 12271-12292. Curran Associates, Inc., 2024. 2 +[36] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Qizhi Chen, Dan Xu, Yiting Zhang, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19584-19594, 2024. 2 + +[37] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, and Xuelong Li. Spatialvia: Exploring spatial representations for visual-language-action model, 2025. 3 +[38] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 2 +[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2 +[40] Yiwen Tang, Hongbo Chen, Liyue Qian, Songhan Ge, Mingbo Zhang, and Rui Zheng. Detection of spine curve and vertebral level on ultrasound images using detr. In 2022 IEEE international ultrasonics symposium (IUS), pages 1-4. IEEE, 2022. 2 +[41] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024. +[42] Yiwen Tang, Ray Zhang, Jiaming Liu, Zoey Guo, Bin Zhao, Zhigang Wang, Peng Gao, Hongsheng Li, Dong Wang, and Xuelong Li. Any2point: Empowering any-modality large models for efficient 3d understanding. In European Conference on Computer Vision, pages 456-473. Springer, 2024. 2 +[43] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Exploring the potential of encoder-free architectures in 3d lmm's, 2025. 3 +[44] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[45] Liwei Wang, Yin Li, Jing Huang, and Svetlana Lazebnik. Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):394–407, 2018. 2 +[46] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 2 +[47] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video, 2023. 2 +[48] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. 3 +[49] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 2 + +[50] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding, 2024. 2 +[51] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. $\mathbf{X}^2$ -vlm: All-in-one pretrained model for vision-language tasks, 2023. 2 +[52] Haotian* Zhang, Pengchuan* Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6 +[53] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2 +[54] Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé III, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2024. 3 +[55] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10012-10022, 2020. 2 +[56] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 4 \ No newline at end of file diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/images.zip b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0e34ac6bd0892e7936309d8c2182e7f614ba3165 --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dee7f51c236069922c7cfebfb1346125d1d86e5990eae751022812d54484ea0 +size 726395 diff --git a/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/layout.json b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1591e3fd4fc4bb62b7a287964b4af90a5e03a638 --- /dev/null +++ b/2025/AerialVG_ A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations/layout.json @@ -0,0 +1,8111 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 103, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 103, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 93, + 103, + 522, + 138 + ], + "type": "text", + "content": "AerialVG: A Challenging Benchmark for Aerial Visual Grounding by Exploring Positional Relations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 159, + 509, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 159, + 509, + 190 + ], + "spans": [ + { + "bbox": [ + 100, + 159, + 509, + 190 + ], + "type": "text", + "content": "Junli Liu\\*1,2 Qizhi Chen\\*2,3 Zhigang Wang\\*2 Yiwen Tang1,2 Yiting Zhang2 Chi Yan2 Dong Wang2 Xuelong Li4 Bin Zhao+1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "spans": [ + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "text", + "content": "Northwestern Polytechnical University " + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 190, + 512, + 203 + ], + "type": "text", + "content": "TeleAI" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 74, + 214, + 188, + 279 + ], + "blocks": [ + { + "bbox": [ + 74, + 214, + 188, + 279 + ], + "lines": [ + { + "bbox": [ + 74, + 214, + 188, + 279 + ], + "spans": [ + { + "bbox": [ + 74, + 214, + 188, + 279 + ], + "type": "image", + "image_path": "0b218763d3a467aadcefb54e6c41196e0f01b577e07cd170e988f9fc842c2b40.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 74, + 280, + 188, + 344 + ], + "blocks": [ + { + "bbox": [ + 74, + 280, + 188, + 344 + ], + "lines": [ + { + "bbox": [ + 74, + 280, + 188, + 344 + ], + "spans": [ + { + "bbox": [ + 74, + 280, + 188, + 344 + ], + "type": "image", + "image_path": "19fa1d712a5c0b08add20267d0bf40dce4bd2acfe3150becd488a5a460a16298.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 74, + 345, + 188, + 409 + ], + "blocks": [ + { + "bbox": [ + 74, + 345, + 188, + 409 + ], + "lines": [ + { + "bbox": [ + 74, + 345, + 188, + 409 + ], + "spans": [ + { + "bbox": [ + 74, + 345, + 188, + 409 + ], + "type": "image", + "image_path": "4fa850aa2a7778806fa37f912e486242f488df5b1a2ad71a10bdba912a6b6db8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 74, + 410, + 188, + 474 + ], + "blocks": [ + { + "bbox": [ + 74, + 410, + 188, + 474 + ], + "lines": [ + { + "bbox": [ + 74, + 410, + 188, + 474 + ], + "spans": [ + { + "bbox": [ + 74, + 410, + 188, + 474 + ], + "type": "image", + "image_path": "b7e0337ba00e5ab7e7281f64f48ceed83ca2c6540def82e2e7a0efecb460c7e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 54, + 483, + 555, + 518 + ], + "type": "text", + "content": "Figure 1. Overview of AerialVG. This work introduces aerial visual grounding (AerialVG), a new task for intelligent drones. We provide the first dataset for this task, consisting of 50K annotations and 103K objects. Additionally, an AerialVG model is proposed to explore the positional relations among objects and achieve superior context-aware visual grounding in drone applications." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 214, + 304, + 279 + ], + "blocks": [ + { + "bbox": [ + 190, + 214, + 304, + 279 + ], + "lines": [ + { + "bbox": [ + 190, + 214, + 304, + 279 + ], + "spans": [ + { + "bbox": [ + 190, + 214, + 304, + 279 + ], + "type": "image", + "image_path": "f7680ff1f87d71a5285eabc78ef24107466370d0d343d3b1c4032d067cdee4d6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 190, + 280, + 304, + 344 + ], + "blocks": [ + { + "bbox": [ + 190, + 280, + 304, + 344 + ], + "lines": [ + { + "bbox": [ + 190, + 280, + 304, + 344 + ], + "spans": [ + { + "bbox": [ + 190, + 280, + 304, + 344 + ], + "type": "image", + "image_path": "59b4fd44115f67d09ded023a4b05562714d5608f50be9d00d34321876a0e64ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 190, + 345, + 304, + 408 + ], + "blocks": [ + { + "bbox": [ + 190, + 345, + 304, + 408 + ], + "lines": [ + { + "bbox": [ + 190, + 345, + 304, + 408 + ], + "spans": [ + { + "bbox": [ + 190, + 345, + 304, + 408 + ], + "type": "image", + "image_path": "b62f54ef27fa623c2d0a341ed924370a5d14bafb9594d12678d3b8db5eee2872.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 190, + 410, + 304, + 474 + ], + "blocks": [ + { + "bbox": [ + 190, + 410, + 304, + 474 + ], + "lines": [ + { + "bbox": [ + 190, + 410, + 304, + 474 + ], + "spans": [ + { + "bbox": [ + 190, + 410, + 304, + 474 + ], + "type": "image", + "image_path": "900f50284ace838e4e62e5190a18e215628cd6a66c380fa7618a8d63a84d6d2f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 214, + 420, + 279 + ], + "blocks": [ + { + "bbox": [ + 306, + 214, + 420, + 279 + ], + "lines": [ + { + "bbox": [ + 306, + 214, + 420, + 279 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 420, + 279 + ], + "type": "image", + "image_path": "3752a4fd729a136778f2e39b6fad323d57250297c15796ce12ba87543956bbf9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 280, + 420, + 344 + ], + "blocks": [ + { + "bbox": [ + 306, + 280, + 420, + 344 + ], + "lines": [ + { + "bbox": [ + 306, + 280, + 420, + 344 + ], + "spans": [ + { + "bbox": [ + 306, + 280, + 420, + 344 + ], + "type": "image", + "image_path": "dd5a8d2f549929ce89505607dd7663828e78051e3313b0ac6e3943d4a8a8e06a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 306, + 345, + 420, + 408 + ], + "blocks": [ + { + "bbox": [ + 306, + 345, + 420, + 408 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 420, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 420, + 408 + ], + "type": "image", + "image_path": "6735695b6dd05b4bd390a0337d6e13fadea7aa79b40a2dadbbc6dad5f43dee62.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 306, + 410, + 420, + 474 + ], + "blocks": [ + { + "bbox": [ + 306, + 410, + 420, + 474 + ], + "lines": [ + { + "bbox": [ + 306, + 410, + 420, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 410, + 420, + 474 + ], + "type": "image", + "image_path": "9b584175ecc471a7f16d405ad283d7d853698d011e334cfe708cdb1ea29de179.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 422, + 214, + 536, + 279 + ], + "blocks": [ + { + "bbox": [ + 422, + 214, + 536, + 279 + ], + "lines": [ + { + "bbox": [ + 422, + 214, + 536, + 279 + ], + "spans": [ + { + "bbox": [ + 422, + 214, + 536, + 279 + ], + "type": "image", + "image_path": "2aa53f424bfb86a72aa6e0a94b12aca6611836d3153683569d18366c12620dfe.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 422, + 280, + 536, + 344 + ], + "blocks": [ + { + "bbox": [ + 422, + 280, + 536, + 344 + ], + "lines": [ + { + "bbox": [ + 422, + 280, + 536, + 344 + ], + "spans": [ + { + "bbox": [ + 422, + 280, + 536, + 344 + ], + "type": "image", + "image_path": "3e325b8e2d8e505b5a9d5fec2f9ec37e38747b7bcde1807d03e5d636931f8633.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 422, + 345, + 536, + 409 + ], + "blocks": [ + { + "bbox": [ + 422, + 345, + 536, + 409 + ], + "lines": [ + { + "bbox": [ + 422, + 345, + 536, + 409 + ], + "spans": [ + { + "bbox": [ + 422, + 345, + 536, + 409 + ], + "type": "image", + "image_path": "6123157bda65f05d2efdc5467638e1ac8e96efcabc132ba767c0fb6777d6df36.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 422, + 410, + 536, + 474 + ], + "blocks": [ + { + "bbox": [ + 422, + 410, + 536, + 474 + ], + "lines": [ + { + "bbox": [ + 422, + 410, + 536, + 474 + ], + "spans": [ + { + "bbox": [ + 422, + 410, + 536, + 474 + ], + "type": "image", + "image_path": "c095c75e6db2bb190435e03e3a49842b84f3f37065862afa09c0d0e3706de897.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 151, + 527, + 200, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 527, + 200, + 539 + ], + "spans": [ + { + "bbox": [ + 151, + 527, + 200, + 539 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 552, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 552, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 552, + 297, + 684 + ], + "type": "text", + "content": "Visual grounding (VG) aims to localize target objects in an image based on natural language descriptions. In this paper, we propose AerialVG, a new task focusing on visual grounding from aerial views. Compared to traditional VG, AerialVG poses new challenges, e.g., appearance-based grounding is insufficient to distinguish among multiple visually similar objects, and positional relations should be emphasized. Besides, existing VG models struggle when applied to aerial imagery, where high-resolution images cause significant difficulties. To address these challenges, we introduce the first AerialVG dataset, consisting of 5K" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 528, + 556, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 556, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 556, + 683 + ], + "type": "text", + "content": "real-world aerial images, 50K manually annotated descriptions, and 103K objects. Particularly, each annotation in AerialVG dataset contains multiple target objects annotated with relative spatial relations, requiring models to perform comprehensive spatial reasoning. Furthermore, we propose an innovative model especially for the AerialVG task, where a Hierarchical Cross-Attention is devised to focus on target regions, and a Relation-Aware Grounding module is designed to infer positional relations. Experimental results validate the effectiveness of our dataset and method, highlighting the importance of spatial reasoning in aerial visual grounding. The code will be released at https://github.com/Ideal-ljl/AerialVG." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 162, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 162, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 162, + 703 + ], + "type": "text", + "content": "*Authors contributed equally" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5177" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 199 + ], + "type": "text", + "content": "Visual Grounding (VG) is a challenging task that involves identifying specific regions or objects within an image following natural language instructions. This task requires models to deeply understand both visual and textual modalities, effectively aligning image context with corresponding linguistic expressions. VG is critical for a wide range of applications, including visual question answering, vision-and-language navigation, and human-robot interaction, etc [22, 33, 35, 36, 49, 55]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 199, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 354 + ], + "type": "text", + "content": "With the continuous progress of multimodal models, the performance of VG tasks has been significantly improved. However, existing models and datasets still have obvious limitations. Current research focuses primarily on ground objects, and the exploration of aerial visual grounding remains underdeveloped. Aerial visual grounding has significant potential applications in areas such as emergency rescue, logistics, and ecological monitoring, which calls for further research and development. Specifically, the ability to manage multiple objects of the same type and their spatial relationships in complex environments remains underdeveloped, leading to limited generalization capabilities in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 438 + ], + "type": "text", + "content": "To this end, we propose a more challenging task, termed Aerial Visual Grounding (AerialVG). AerialVG aims to apply the visual grounding task to UAV platforms, and promote the development of UAV intelligent perception and autonomous navigation technologies. Compared with traditional VG tasks, AerialVG faces more challenges, mainly in the following two aspects:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 295, + 534 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 438, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 521 + ], + "type": "text", + "content": "- Broad field of view and small objects. Unlike images from the ground perspective, when UAVs capture images from the air, the field of view is usually much broader, and the size of objects in the image tends to be relatively small. While a larger field of view provides more context and coverage, it also imposes challenges to locate and identify objects at a finer scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 534 + ], + "type": "text", + "content": "- Complex object distribution and spatial relationship." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 63, + 534, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 534, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 63, + 534, + 295, + 677 + ], + "type": "text", + "content": "From the perspective of UAVs, object types are mainly pedestrians and vehicles, which are usually distributed in crowded urban environments. Especially in busy urban streets or traffic-intensive areas, there may be multiple similar or identical objects in the field of view at the same time. For example, multiple similar vehicles are often simultaneously presented in one street. In this case, it is difficult to accurately distinguish the target object based on appearance features alone, and it is necessary to rely on the spatial relationship between objects, such as relative position, distance, and direction, to achieve precise positioning and recognition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To address above issues, we propose the first dataset designed specifically for the AerialVG task. The dataset consists of 5,000 high-resolution images covering a variety of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "complex urban environments and contains a rich variety of object types, such as pedestrians, vehicles, and other common urban objects. In addition, the dataset includes nearly 50k manual annotations. These annotations not only provide the bounding box of each object but also include natural language descriptions related to the object, which can help the model understand the positional relationship between objects. With these high-quality annotations, a comprehensive and efficient test platform is provided for the AerialVG task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 192, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 323 + ], + "type": "text", + "content": "Aiming at the core challenges in the AerialVG task, we also design an innovative model that can accurately locate objects based on the positional relation between them. Specifically, a Hierarchical Cross-Attention is first proposed to help the model focus more effectively on regions where objects are most likely to appear, thus addressing the issue of vast visual areas. Furthermore, we devise a Relation-Aware Grounding module to understand spatial information such as the relative position, direction, and distance between objects, effectively distinguishing multiple similar objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 346 + ], + "type": "text", + "content": "Overall, the main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 348, + 553, + 467 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 348, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 383 + ], + "type": "text", + "content": "- We propose the AerialVG task and introduce the first dataset, providing a valuable resource for advancing research in aerial visual grounding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 384, + 553, + 430 + ], + "type": "text", + "content": "- We present a novel AerialVG model that effectively addresses two key challenges, i.e., the large field of view with small-sized objects and the complex distribution and spatial relationships between objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 431, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 467 + ], + "type": "text", + "content": "- We conduct extensive experiments to thoroughly validate the effectiveness of our proposed model, demonstrating its superior performance." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 478, + 400, + 491 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 715 + ], + "type": "text", + "content": "Visual Grounding. Visual Grounding (VG) is a fundamental task that requires identifying specific objects in an image based on natural language descriptions. This task has made some progress in both 2D [5, 25] and 3D [17, 18, 40-42, 50, 53] fields. Existing visual grounding methods can be broadly divided into two categories, i.e., (1) two-stage methods[13, 28, 45] and (2) one-stage methods[27, 38, 39]. The former first generates region-level candidate boxes and then uses language to select the best matching region. The latter directly interacts with visual and language cues through bounding boxes, avoiding the computationally intensive proposal box generation and region feature extraction steps present in two-stage methods. Several recent methods[8, 10, 23, 25] have adopted the transformer structure for multimodal fusion, thus achieving better performance. In addition, with the booming development of vision language models (VLMs), some studies [46, 47, 51] have improved the performance of vision grounding by fine" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5178" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 311 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 311 + ], + "type": "image", + "image_path": "774d18a672ad3737cfeea57f1d4c1320139d43c62141fa616016c4c99c3b0e65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 318, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 54, + 318, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 555, + 353 + ], + "type": "text", + "content": "Figure 2. Example of AerialVG Dataset. There are many red cars in the picture. In order to achieve accurate object positioning, we need to use the positional relationship of surrounding auxiliary objects to assist reasoning. To this end, we accurately annotated the data to ensure that the spatial relationship between objects can be fully captured, thereby improving the accuracy of AerialVG." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 373, + 294, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 294, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 294, + 433 + ], + "type": "text", + "content": "tuning pre-trained VLMs, making full use of the knowledge and advantages of pre-trained models. These advances have improved both the accuracy and efficiency of VG systems, enabling more complex applications. However, they still cannot solve the spatial relationship problem in AerialVG." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 685 + ], + "type": "text", + "content": "Vision-Language Models. Vision-Language Models (VLMs) integrate visual and linguistic information to perform tasks that require both image understanding and natural language processing. Typically, they use pretrained visual backbones to encode images, and large-scale language models to interpret textual inputs. The visual and textual information is then fused through mechanisms like cross-attention layers. Recent advancements, such as GPT4o [32], LLaVA [24], and InstructBLIP [6], have achieved significant progress in language-guided navigation and visual reasoning. VLMs have also been adapted to a variety of domains. Specifically, robot action generation produces robot actions based on textual instructions [37, 54], vision language navigation guides UAVs in aerial environments [14, 26], 3D generation creates and understands 3D data based on language inputs [43, 48]. Moreover, VLMs are increasingly used in video understanding, biomedical imaging, and remote sensing, demonstrating their versatility across different tasks. These developments highlight the growing potential of VLMs in various real-world applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Spatial Reasoning Representation. Spatial reasoning, traditionally part of broader tasks like SLAM [1, 11] or" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 373, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 636 + ], + "type": "text", + "content": "depth estimation [12], has seen increasing interest in the context of VLMs. Earlier research focused on explicit spatial scene representations, such as spatial memory [15, 16] or spatial scene graphs [19, 44], to help machines understand and reason about spatial relations. Recent work has further advanced the spatial reasoning capabilities of VLMs. Some approaches leverage 3D information to enhance VLM performance. [3, 20, 21, 34] focus on reconstructing scenes from multi-view images and integrating dense semantic features to improve these representations. These 3D representations are then fused into LLMs for better spatial understanding. However, multi-view images are not always available, and the explicit construction of scenes using dense semantic features can be computationally intensive. Furthermore, the gap between 3D representations and language can often degrade performance. In contrast, other approaches utilize 2D VLMs to understand spatial relationships and measure distances. Unlike the 3D methods, spatial understanding in these models is implicitly encoded, allowing the VLM to directly address spatial reasoning tasks without requiring explicit 3D representations or scene graphs, as seen in works like [2, 4]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 646, + 369, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 369, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 369, + 658 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we develop the AerialVG dataset as shown in Figure 2. This dataset contains 5,000 high-definition images taken from UAVs. We annotated them to ensure that they are suitable for aerial visual grounding task. In addition, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5179" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 294, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 294, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 294, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 294, + 234 + ], + "type": "image", + "image_path": "db38ad66f5a9790cf22b7864ff1ae6d839bd8fdf5212aaad48be5d69ca7eb536.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 245, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 295, + 291 + ], + "type": "text", + "content": "Figure 3. Resolution distribution and word cloud of AerialVG dataset. Most of the images in AerialVG dataset are high-resolution, and the most frequently appearing words are vehicle type, color, and location." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 387 + ], + "type": "text", + "content": "provide a specially designed natural language description for each object, which details the objects in the image and their relative positions. The AerialVG dataset provides a solid foundation for fine-grained object grounding in complex scenes from a high-altitude perspective, and especially lays a new research direction for the efficient grounding and understanding tasks of intelligent UAVs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 167, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 167, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 167, + 407 + ], + "type": "text", + "content": "3.1. Dataset generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 544 + ], + "type": "text", + "content": "The images in our AerialVG dataset are sourced from the VisDrone2019 dataset [56]. VisDrone consists of 288 video clips (261,908 frames) and 10,209 static images captured by various UAV-mounted cameras across 14 cities, covering a range of environments such as urban and rural areas. This diversity in scene type, location, and conditions (including different weather and lighting) offers a robust foundation for visual grounding tasks. The dataset includes more than 2.6 million manually annotated bounding boxes, highlighting common objects like pedestrians, vehicles, and bicycles, with attributes such as visibility, object class, and occlusion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 664 + ], + "type": "text", + "content": "In UAV-view images, buildings and roads occupy most of the visual space, while the primary objects of interest are pedestrians and vehicles. They often exhibit similar appearances. Relying solely on coarse-grained categories is insufficient for accurate object localization. Therefore, the annotations must capture fine-grained object features, e.g., color, vehicle type, clothing, and consider the spatial relationships between target objects and prominent surrounding elements. Due to the complexity of visual annotation, we opted for manual annotation to ensure precision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 665, + 294, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 294, + 677 + ], + "type": "text", + "content": "Our manual annotation process follows these principles:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "content": "- Observer-perspective spatial relationships. For the convenience of annotation and unification, our position relation is described from the first-person perspective and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 72, + 529, + 266 + ], + "blocks": [ + { + "bbox": [ + 340, + 72, + 529, + 266 + ], + "lines": [ + { + "bbox": [ + 340, + 72, + 529, + 266 + ], + "spans": [ + { + "bbox": [ + 340, + 72, + 529, + 266 + ], + "type": "image", + "image_path": "57d33ed39dda29c4b25cb4d7f8897c8b5ed8ecab13eb9a78c64c83f07bb4ad04.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 344 + ], + "type": "text", + "content": "Figure 4. Comparison between different datasets. Words refers to the average number of words in the text; Objects refer to the average number of objects in each image; Area refers to the proportion of irrelevant areas; Resolution refers to the maximum resolution of the dataset images and Captions refers to the average number of annotations contained in each image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 354, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 354, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 321, + 354, + 555, + 402 + ], + "type": "text", + "content": "is divided into eight categories, i.e., above, below, left, right, top left, top right, bottom left, and bottom right. As shown in Figure 2 A, the white bus is above the red car with a white roof, not on the right." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 402, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 402, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 402, + 556, + 510 + ], + "type": "text", + "content": "- Selection of prominent auxiliary objects. When locating the target object, we give priority to selecting the closest object as the auxiliary object to improve positioning accuracy. To ensure simplicity, if one auxiliary object is sufficient to locate the target, no redundant description will be provided. For instance, as shown in Figure 2 D, the target object is the white sedan, and we select the blue SUV as the auxiliary object, while disregarding the red car at the top left." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 517, + 416, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 416, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 416, + 528 + ], + "type": "text", + "content": "3.2. Dataset statistics" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 569 + ], + "type": "text", + "content": "This subsection presents key statistics of the AerialVG dataset and offers a comparative comparison with existing visual grounding datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "text", + "content": "Figure 3 shows the resolution distribution and word cloud for the AerialVG dataset. Unlike the RefCOCO series, where over " + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "text", + "content": " of images have a resolution around " + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "inline_equation", + "content": "600 \\times 650" + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "text", + "content": ", the AerialVG dataset consists mostly of images with resolutions greater than " + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "inline_equation", + "content": "1000 \\times 500" + }, + { + "bbox": [ + 313, + 570, + 555, + 653 + ], + "type": "text", + "content": ". This higher resolution and wider field of view make the Aerial VG task more challenging." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "The word cloud analysis highlights frequent mentions of vehicle types, colors, and spatial positions. These annotations emphasize the importance of understanding not only object identification but also their relative locations in the environment. This rich annotation style is critical for tasks" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5180" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "like UAV navigation, emergency response, and large-scale monitoring, where spatial relationships are essential for accurate decision-making." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 204 + ], + "type": "text", + "content": "Figure 4 compares the AerialVG dataset with existing visual grounding datasets across three key aspects, i.e., the average number of words in the text, the average number of objects located in the image, and the proportion of irrelevant regions. In all three aspects, the AerialVG dataset demonstrates a higher level of challenge, highlighting its complexity and suitability for more advanced visual grounding tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 214, + 111, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 111, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 111, + 227 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 234, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 353 + ], + "type": "text", + "content": "As shown in Figure 5, we take GroundingDINO [25] as the baseline and design an end-to-end model for AerialVG. GroundingDINO [25] is built on a transformer architecture, where the encoder stacks self-attention and cross-attention layers to fuse image and text features. After fusing these multimodal features, it selects the most relevant features as queries, which are highly similar to the text. These queries are then decoded in the decoder, and the final bounding boxes and logits are generated by the bbox head and class head, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 450 + ], + "type": "text", + "content": "Unlike GroundingDINO [25], after embedding the image into hierarchical features using Swin-Transformer [29], we fuse these features with text features encoded by BERT [9] via Hierarchical Cross Attention, enabling the model to focus more on potential regions where objects are likely to appear. After the Decoder, we introduce the Relation-Aware Grounding module to capture the positional relationships between different objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 214, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 214, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 214, + 468 + ], + "type": "text", + "content": "4.1. Hierarchical Cross-Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 474, + 294, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 294, + 545 + ], + "type": "text", + "content": "In aerial view images, high resolution often comes with large portions of the scene occupied by buildings, roads, and other background elements, reducing the focus on key objects. To address this, we propose Hierarchical Cross-Attention, a mechanism designed to help the model more efficiently attend to relevant areas within the scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "With the hierarchical feature representation enabled by Swin-Transformer[29], image features are organized in a layered manner. However, traditional cross-attention methods typically compute a complete attention map by flattening the features for processing, which can lead to information loss during the subsequent fusion process. In hierarchical structures, features at the same spatial location across different layers still represent the same object. Therefore, if a high-attention region is identified in one layer based on textual cues, this region should receive increased attention in subsequent layers. Conversely, if a region shows low attention, it can be partially disregarded in the next layer, enabling the model to prioritize more relevant areas effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "As shown in Figure 5, after calculating the attention map between textual and image features, we retain their hierarchical structure. As the layer number increases, the attention map will also become smaller, making small objects receive insufficient attention in the higher layers. To alleviate this problem, we first convolve the attention map of the first layer to make it consistent with the shape of the last layer map, and fuse the result with it." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 180, + 553, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 180, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 361, + 180, + 553, + 194 + ], + "type": "interline_equation", + "content": "A _ {n} ^ {*} = (1 - \\beta) \\times A _ {n} + \\beta \\times C \\left(A _ {1}\\right), \\tag {1}", + "image_path": "3180ff7232d6f836c67da771833200e079a58af25d79e9d5d962fe9104b468f8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "inline_equation", + "content": "C(\\cdot)" + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "content": " denotes the convolution operation, " + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "content": " is a hyperparameter, " + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "inline_equation", + "content": "A_{j}" + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "content": " represents the attention map of the j-th layer. The processed attention map is then combined with the highest layer's attention map " + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "inline_equation", + "content": "A_{n}" + }, + { + "bbox": [ + 313, + 199, + 553, + 283 + ], + "type": "text", + "content": ", ensuring that even small objects are given adequate focus. After that, we use Equation 2 to enhance the attention map in lower layers based on the higher layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 358, + 293, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 293, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 358, + 293, + 553, + 308 + ], + "type": "interline_equation", + "content": "A _ {i} ^ {*} = (1 - \\alpha) \\times A _ {i} + \\alpha \\times F \\left(A _ {i + 1}\\right). \\tag {2}", + "image_path": "bf90f7b7d48d9fa4ebc77bd8585f3c0598f239c3e7001bb3e0bbb68bb695e76e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": "Specifically, we use bilinear interpolation " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": ", to expand the attention map " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "A_{i+1}" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "i+1" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": " layer to the shape of the " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": " layer " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": ". The expanded attention map is then weighted by the hyper-parameter " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": " and combined with " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": " to form the updated " + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "inline_equation", + "content": "A_i^*" + }, + { + "bbox": [ + 313, + 316, + 554, + 400 + ], + "type": "text", + "content": ". This approach effectively achieves our goal without introducing additional parameters, yielding improvements in performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 408, + 465, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 465, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 465, + 421 + ], + "type": "text", + "content": "4.2. Relation-Aware Grounding" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 520 + ], + "type": "text", + "content": "Following GroundingDINO [25], we first perform language-guided query selection after the encoder, and then pass the selected queries into the decoder. In the decoder, we perform cross-attention on the query separately with the image features and text features to enhance the model's ability to localize objects. However, experiments showed that these features did not adequately capture the positional relations described in the text." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 594 + ], + "type": "text", + "content": "We draw inspiration from certain 3D visual grounding methods for determining spatial relationships. However, those methods generally require explicit bounding boxes and clearly defined spatial relationships, rather than leveraging full textual descriptions. Our approach circumvents these requirements while still achieving competitive results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Specifically, we design a Relation-Aware Grounding module to capture positional relation information. As shown in figure 5, after passing through the decoder, the fused features already contain the positional information of each object. Assuming we have " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " features before the Relation-Aware Grounding module, we then concatenate every two of the " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " features in pairs to obtain " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " features, which effectively captures the pairwise relationships between the objects and enhances the model's ability to learn these interactions. The self-attention operation is performed" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5181" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 539, + 209 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 209 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 209 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 209 + ], + "type": "image", + "image_path": "65af3e60eaadcf55f40f3575371d2e01d190cbbb934a1cda77f8e88a1fd73487.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 211, + 231, + 319 + ], + "blocks": [ + { + "bbox": [ + 69, + 211, + 231, + 319 + ], + "lines": [ + { + "bbox": [ + 69, + 211, + 231, + 319 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 231, + 319 + ], + "type": "image", + "image_path": "97c68f1df2aab60de409f007425f4051c968bc3c853cfbd096b0523681380ba4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 346 + ], + "type": "text", + "content": "Figure 5. The architecture of AerialVG model. Hierarchical Cross Attention directs the model's focus to potential object locations, while the Relation-Aware Grounding module enables the model to perceive spatial relationships between objects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 248, + 211, + 531, + 319 + ], + "blocks": [ + { + "bbox": [ + 248, + 211, + 531, + 319 + ], + "lines": [ + { + "bbox": [ + 248, + 211, + 531, + 319 + ], + "spans": [ + { + "bbox": [ + 248, + 211, + 531, + 319 + ], + "type": "image", + "image_path": "03908c1c6a45192f2301b9632b527344740af3ad632cf4e5d209ad969d305d17.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "text", + "content": "on " + }, + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "text", + "content": " concatenated features in the same column to perceive the positional relation between objects. Subsequently, we conduct cross-attention with " + }, + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "inline_equation", + "content": "m^2" + }, + { + "bbox": [ + 54, + 357, + 294, + 441 + ], + "type": "text", + "content": " queries and the text features. This step enables the model to better align and integrate both visual and textual information, enhancing its understanding of the contextual relationships between objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 620 + ], + "type": "text", + "content": "Finally, we first map the last state of the hidden layer to the specified feature dimension through the Relation Head. On this basis, we calculate the similarity between these features and the text features. The calculated similarity scores are then subjected to a softmax operation to generate a set of normalized similarity values. These normalized values represent the relationship score between each pair of objects, that is, how well they match the relationship described in the text. In this way, we can get an " + }, + { + "bbox": [ + 55, + 441, + 295, + 620 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 55, + 441, + 295, + 620 + ], + "type": "text", + "content": " relationship matrix, which represents the relationship score between each pair of objects. We take the maximum value from each row of this relationship matrix as the final relation score for each object. We combine the relation scores with the scores obtained from the query through the class head to form the final logits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 626, + 144, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 144, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 144, + 637 + ], + "type": "text", + "content": "4.3. Loss Function" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "text", + "content": "Our loss function consists of three parts, i.e., " + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rel}}" + }, + { + "bbox": [ + 55, + 643, + 295, + 668 + ], + "type": "text", + "content": " as shown in Equation 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 673, + 294, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 673, + 294, + 685 + ], + "spans": [ + { + "bbox": [ + 127, + 673, + 294, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e g}} + \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e l}}. \\tag {3}", + "image_path": "dbecf2a25b57b9d6fe4ec1213dd220198d5eaf2d54baf7555bf1d4e964d90d1e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Following previous work on GroundingDINO, we employ L1 loss and GIOU loss for bounding box regression in " + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 357, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 393 + ], + "type": "text", + "content": "For classification, we adopt a contrastive loss between predicted objects and language tokens as " + }, + { + "bbox": [ + 313, + 357, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 313, + 357, + 554, + 393 + ], + "type": "text", + "content": ", in line with the GLIP [52] approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 394, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 419 + ], + "type": "text", + "content": "Additionally, we compute a Relation Loss based on the predicted relationships between object pairs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 422, + 553, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 373, + 422, + 553, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e l}} = - \\log \\left(\\frac {e ^ {m _ {i , j}}}{\\sum_ {i , j} e ^ {m _ {i , j}}}\\right), \\tag {4}", + "image_path": "7396edea3b8710ddbbd2a2a7d9abdd9510c47adc0c6d0f9d2efb1d98f5db2a93.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": "where Hungarian matching is performed to identify the corresponding target object " + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "inline_equation", + "content": "m_{i}" + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": " and the auxiliary object " + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "inline_equation", + "content": "m_{j}" + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": ". In this way, the element " + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "inline_equation", + "content": "m_{i,j}" + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "inline_equation", + "content": "m \\times m" + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": " relationship matrix represents the ground truth relation between the objects. We then use Equation 4 to maximize the score at the corresponding position " + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "inline_equation", + "content": "m_{i,j}" + }, + { + "bbox": [ + 313, + 457, + 554, + 530 + ], + "type": "text", + "content": " in the relation matrix." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 543, + 390, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 390, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 390, + 556 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 563, + 447, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 447, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 447, + 575 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Our proposed AerialVG model, utilizes Swin-L [29] as the image backbone. For text representation, we employ the BERT-base [9] from Hugging Face as the text backbone. Following Grounding DINO, we extract four levels of image feature scales from the image backbone, ranging from " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". We set the maximum number of text tokens to 256. We set " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "N = 6" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " for Encoder and Decoder layers, while the Relation-Aware Grounding module consists of three layers. The final loss function consists of multiple components, i.e., classification loss (or contrastive loss), L1 loss, GIoU loss, and relation loss, with corresponding weights set" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5182" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "to 1.0, 5.0, 2.0, and 1.0, respectively. The hyper-parameters " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " are set to 0.2 and 0.3. We conduct extensive experiments on the AerialVG dataset and the RefCOCO, RefCOCO+, and RefCOCOg datasets [30]. Additionally, we perform ablation studies to demonstrate the effectiveness of our model design." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 294, + 251 + ], + "type": "text", + "content": "Our model training consists of two stages. Initially, we train the model without the Relation-aware Grounding module, using existing VG datasets following Grounding DINO [25] to achieve convergence. Once the model has converged, we freeze all other parameters and train the Relation-Aware Grounding module using our AerialVG dataset only, allowing it to specifically capture inter-object relationships. More details can be found in our supplementary material." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 260, + 162, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 162, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 162, + 272 + ], + "type": "text", + "content": "5.2. Evaluation metric" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 410 + ], + "type": "text", + "content": "In AerialVG, object homogeneity is a significant challenge, as many objects share nearly identical visual characteristics. A model that performs well on Top-5 Accuracy but poorly on Top-1 Accuracy likely relies primarily on visual similarity rather than understanding the spatial relationships or contextual cues that distinguish between objects. By analyzing both Top-1 and Top-5 Accuracy, we can diagnose specific failure modes of the model. Model with both high Top-1 and Top-5 Accuracy demonstrates strong object discrimination and contextual understanding ability, making it more reliable for AerialVG-related applications." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 441, + 301, + 555 + ], + "blocks": [ + { + "bbox": [ + 67, + 421, + 282, + 431 + ], + "lines": [ + { + "bbox": [ + 67, + 421, + 282, + 431 + ], + "spans": [ + { + "bbox": [ + 67, + 421, + 282, + 431 + ], + "type": "text", + "content": "Table 1. Performance comparison on the AerialVG dataset." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 441, + 301, + 555 + ], + "lines": [ + { + "bbox": [ + 56, + 441, + 301, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 301, + 555 + ], + "type": "table", + "html": "
MethodSetting
Zero-ShotFine-Tuning
Top 1Top 5Top 1Top 5
TranVG[8]2.573.6511.5313.68
Dynamic-MDETR[7]2.614.6319.8729.87
SimVG[5]3.614.6820.9431.32
Grounding DINO[25]11.1013.7929.3678.87
AerialVG(Ours)--50.0187.00
", + "image_path": "56ccb445e8aee6d3876f24bc79d42dd9e961264fd391066e605ae0fbaecbca85.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 113, + 588 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Table 1 presents the evaluation results of different models on the AerialVG dataset. In the Zero-Shot setting, existing VG models exhibit limited performance, with both Top-1 Accuracy and Top-5 Accuracy hovering around " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ". This result highlights the challenges posed by high-resolution aerial imagery and the unique characteristics of the aerial viewpoint, which are not well-represented in existing datasets. The poor performance of these models further underscores the necessity of AerialVG as a benchmark for evaluating visual grounding in aerial scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "After fine-tuning, Top-5 Accuracy shows a significant improvement, indicating that models can better recognize and retrieve relevant object candidates when adapted to the AerialVG. However, Top-1 Accuracy remains relatively low, suggesting that while models are able to identify visually similar objects within the scene, they struggle to distinguish the correct target based on spatial relationships and contextual cues." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 168, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 300 + ], + "type": "text", + "content": "In comparison, the proposed AerialVG model demonstrates superior performance over previous approaches. It demonstrates a better understanding of object relationships in aerial images, resulting in higher Top-1 and Top-5 accuracy. This suggests that our model effectively integrates the Hierarchical Cross-Attention mechanisms and the Relation-Aware Grounding module, which are crucial for distinguishing visually similar objects based on spatial dependencies. These results validate our architectural choices and demonstrate the effectiveness of our model in addressing the unique challenges of aerial visual grounding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 301, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 396 + ], + "type": "text", + "content": "In addition to the AerialVG dataset, we also evaluate our AerialVG model on the RefCOCO[30], RefCOCO+[30], and RefCOCOg[31] datasets to assess its performance on standard visual grounding benchmarks. As shown in Table 2, our AerialVG model also presents good performance across multiple metrics on the RefCOCO series datasets, achieving comparable results with previous state-of-the-art approaches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 397, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 553, + 505 + ], + "type": "text", + "content": "These findings confirm that the integration of the Hierarchical Cross-Attention and Relation-Aware Grounding module does not compromise the model's ability to accurately localize individual objects. Despite being designed to enhance spatial reasoning and relational understanding, these components maintain strong object-level grounding capabilities. This highlights the versatility of our model, making it effective in both aerial imagery and conventional ground-based VG tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 515, + 406, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 406, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 406, + 528 + ], + "type": "text", + "content": "5.4. Ablation study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 604 + ], + "type": "text", + "content": "To further validate the effectiveness of the Hierarchical Cross-Attention and the Relation-Aware Grounding module, we conduct ablation experiments, as shown in Table 3. The results demonstrate that each component plays a crucial role in improving different aspects of the model's performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 676 + ], + "type": "text", + "content": "- Hierarchical Cross-Attention significantly enhances Top-5 Accuracy, indicating that this module helps the model focus on fine-grained details within the image. By attending to multi-level image features, the model can better capture subtle differences between similar objects, leading to improved candidate retrieval." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "- Relation-Aware Grounding module notably improves Top-1 Accuracy, demonstrating its ability to refine object selection by incorporating relational reasoning. This con" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5183" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 91, + 504, + 205 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 464, + 82 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 464, + 82 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 464, + 82 + ], + "type": "text", + "content": "Table 2. Performance comparison on RefCOCO, RefCOCO+, and RefCOCOg datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 91, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 111, + 91, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 111, + 91, + 504, + 205 + ], + "type": "table", + "html": "
MethodBackboneRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
TranVG [8]ResNet-10180.3282.6778.1263.568.1855.6367.1167.44
Dynamic-MDETR [7]ResNet-5085.9788.8280.1274.8381.763.4473.1874.49
SimVG [5]ViT-L90.6192.5387.6885.3689.6179.7482.6786.8
Grounding DINO[25]Swin-L90.5693.1988.2482.7588.9575.9286.1387.02
AerialVG (Ours)Swin-L90.5893.2388.5682.8389.0276.0386.5288.04
", + "image_path": "84938423097d2d2f7c955a68f38ef4acd8d19876ded2f0856fd7cfc45ee77cdf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 63, + 254, + 291, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 246 + ], + "type": "text", + "content": "Table 3. Ablation study. Attn refers to hierarchical cross attention, Relation refers to relation aware grounding." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 254, + 291, + 357 + ], + "lines": [ + { + "bbox": [ + 63, + 254, + 291, + 357 + ], + "spans": [ + { + "bbox": [ + 63, + 254, + 291, + 357 + ], + "type": "table", + "html": "
MethodAerialVG
Fine-Tuning
Top 1Top 5
Grounding DINO (Baseline)29.3678.87
Grounding DINO + Attn29.7882.32
Grounding DINO + Relation45.6379.63
Grounding DINO + Attn +Relation50.0187.00
", + "image_path": "cbfa7208381afa830c03e7fadee2ac001e22228898238e5660bf34de0fe983a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 368, + 295, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 368, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 63, + 368, + 295, + 404 + ], + "type": "text", + "content": "firms that the module helps disambiguate objects based on spatial and contextual relationships, rather than relying solely on visual appearance." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 415, + 295, + 594 + ], + "blocks": [ + { + "bbox": [ + 57, + 415, + 295, + 594 + ], + "lines": [ + { + "bbox": [ + 57, + 415, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 295, + 594 + ], + "type": "image", + "image_path": "4390bf43736934e2b9cbc143c7e71e175f0ae036e5469d5a0b848fd19e29ba33.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 601, + 295, + 635 + ], + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 635 + ], + "type": "text", + "content": "Figure 6. Example of Qualitative Test. Annotation: A blue car is parking on the right road with a white sedan above it. The model output score for this result is 0.70." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "text", + "content": "5.5. Qualitative Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "To better validate the generalization ability of our model, we conducted real-world experiments by capturing a set of aerial images using the DJI M30T drone. The images were taken at an altitude of 50 meters with a resolution of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "type": "inline_equation", + "content": "4000 \\times 3000" + }, + { + "bbox": [ + 313, + 224, + 555, + 308 + ], + "type": "text", + "content": ". The images are filled with various types of vehicles, posing a significant challenge to object localization. However, as shown in Figure 6, our model still successfully locates the objects accurately based on spatial relationship information, demonstrating its strong generalization capability. We will show more visualization results in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 481, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 481, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 481, + 339 + ], + "type": "text", + "content": "6. Limitations and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 349, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 556, + 458 + ], + "type": "text", + "content": "The primary limitation of current visual grounding models, including our AerialVG, is their inability to effectively recognize unstructured background information, such as roads and other environmental elements. While these models excel at detecting foreground objects, they often struggle to incorporate contextual information from the background, which is crucial for tasks requiring a more holistic understanding of the scene. Promoting the spatial intelligence of perception models will be a key focus of our future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 488 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we introduce AerialVG, a visual grounding task designed specifically for aerial perspectives. We construct the first real-world dataset tailored for this task, featuring high-resolution aerial imagery with densely populated scenes and complex spatial relationships. To address the challenges of fine-grained object differentiation and spatial reasoning, we propose an effective model, integrating Hierarchical Cross-Attention and a Relation-Aware Grounding module to enhance object localization accuracy. Extensive experiments validate the effectiveness of our approach, demonstrating state-of-the-art performance on the AerialVG dataset and competitive results on standard Ref-COCO benchmarks. Our findings highlight the importance of spatial reasoning in aerial visual grounding and establish AerialVG as a challenging benchmark for future research. We hope that our dataset and the proposed method will serve as a solid foundation for advancing aerial view perception and autonomous vision-and-language applications." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5184" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work is supported by the Shanghai AI Laboratory, the National Natural Science Foundation of China (62376222), and Young Elite Scientists Sponsorship Program by CAST (2023QNRC001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 168, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 296, + 224 + ], + "type": "text", + "content": "[1] Cesar Cadena, Luca Carlone, Henry Carrillo, Yasir Latif, Davide Scaramuzza, José Neira, Ian Reid, and John J Leonard. Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age. IEEE Transactions on robotics, 32(6):1309-1332, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 224, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 296, + 269 + ], + "type": "text", + "content": "[2] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brian Ichter, Danny Driess, Pete Florence, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 269, + 295, + 312 + ], + "type": "text", + "content": "[3] Qizhi Chen, Delin Qu, Yiwen Tang, Haoming Song, Yiting Zhang, Dong Wang, Bin Zhao, and Xuelong Li. Freegaussian: Guidance-free controllable 3d gaussian splats with flow derivatives. arXiv preprint arXiv:2410.22070, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 313, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 313, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 61, + 313, + 296, + 368 + ], + "type": "text", + "content": "[4] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 369, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 369, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 369, + 296, + 413 + ], + "type": "text", + "content": "[5] Ming Dai, Lingfeng Yang, Yihao Xu, Zhenhua Feng, and Wankou Yang. Simvg: A simple framework for visual grounding with decoupled multi-modal fusion. arXiv preprint arXiv:2409.17531, 2024. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 413, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 413, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 61, + 413, + 296, + 467 + ], + "type": "text", + "content": "[6] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 468, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 468, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 61, + 468, + 296, + 524 + ], + "type": "text", + "content": "[7] Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2988-2997, 2021. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 525, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 525, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 61, + 525, + 296, + 557 + ], + "type": "text", + "content": "[8] Jiajun Deng, Zhengyuan Yang, Tianlang Chen, Wengang Zhou, and Houqiang Li. Transvg: End-to-end visual grounding with transformers, 2022. 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 558, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 558, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 61, + 558, + 296, + 591 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 624 + ], + "type": "text", + "content": "[10] Henghui Ding, Chang Liu, Suchen Wang, and Xudong Jiang. Vision-language transformer and query generation for referring segmentation, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 658 + ], + "type": "text", + "content": "[11] Hugh Durrant-Whyte and Tim Bailey. Simultaneous localization and mapping: part i. IEEE robotics & automation magazine, 13(2):99–110, 2006. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 714 + ], + "type": "text", + "content": "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[13] Akira Fukui, Dong Huk Park, Daylen Yang, Anna Rohrbach, Trevor Darrell, and Marcus Rohrbach. Multimodal compact bilinear pooling for visual question answering and visual grounding. arXiv preprint arXiv:1606.01847, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[14] Yunpeng Gao, Chenhui Li, Zhongrui You, Junli Liu, Zhen Li, Pengan Chen, Qizhi Chen, Zhonghan Tang, Liansheng Wang, Penghui Yang, et al. Openfly: A versatile toolchain and large-scale benchmark for aerial vision-language navigation. arXiv preprint arXiv:2502.18041, 2025.3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 207 + ], + "type": "text", + "content": "[15] Theophile Gervet, Soumith Chintala, Dhruv Batra, Jitendra Malik, and Devendra Singh Chaplot. Navigating to objects in the real world. Science Robotics, 8(79):eadf6991, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 262 + ], + "type": "text", + "content": "[16] Daniel Gordon, Aniruddha Kembhavi, Mohammad Rastegari, Joseph Redmon, Dieter Fox, and Ali Farhadi. Iqa: Visual question answering in interactive environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4089-4098, 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 308 + ], + "type": "text", + "content": "[17] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding with gpt and prototype guidance, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 554, + 373 + ], + "type": "text", + "content": "[18] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xi-anzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xi-anzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 375, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 554, + 409 + ], + "type": "text", + "content": "[19] Marcel Hildebrandt, Hang Li, Rajat Koner, Volker Tresp, and Stephan Gunnemann. Scene graph reasoning for visual question answering, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 443 + ], + "type": "text", + "content": "[20] Yining Hong, Chunru Lin, Yilun Du, Zhenfang Chen, Joshua B. Tenenbaum, and Chuang Gan. 3d concept learning and reasoning from multi-view images, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 554, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 497 + ], + "type": "text", + "content": "[21] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 498, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 544 + ], + "type": "text", + "content": "[22] Ke Li, Di Wang, Haojie Xu, Haodi Zhong, and Cong Wang. Language-guided progressive attention for visual grounding in remote sensing images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-13, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 554, + 578 + ], + "type": "text", + "content": "[23] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 554, + 600 + ], + "type": "text", + "content": "[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 656 + ], + "type": "text", + "content": "[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 5, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 554, + 690 + ], + "type": "text", + "content": "[26] Shubo Liu, Hongsheng Zhang, Yuankai Qi, Peng Wang, Yan-ing Zhang, and Qi Wu. Aerialvln: Vision-and-language navigation for uavs, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 714 + ], + "type": "text", + "content": "[27] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5185" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "Berg. Ssd: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[28] Yongfei Liu, Bo Wan, Xiaodan Zhu, and Xuming He. Learning cross-modal context graph for visual grounding. In Proceedings of the AAAI conference on artificial intelligence, pages 11645-11652, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[29] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 295, + 262 + ], + "type": "text", + "content": "[30] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 262, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 295, + 296 + ], + "type": "text", + "content": "[31] Varun K. Nagaraja, Vlad I. Morariu, and Larry S. Davis. Modeling context between objects for referring expression understanding, 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 295, + 713 + ], + "type": "text", + "content": "[32] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heating, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar Tabarak Khan, Logan Kilpatrick, Jong Wook Kim Christina Kim Yongjik Kim Jan Hendrik Kirchner Jamie Kiros Matt Knight Daniel Kokotajlo Lukasz Kondraciuk Andrew Kondrich Aris Konstantinidis Kyle Kosic Gretchen Krueger,Vishal Kuo Michael Lampe,Ikai LanTeddy Lee Jan Leike,Jade Leung,Daniel LevyChak Ming Li,Rachel LimMolly Lin Stephanie Lin Mateusz Litwin Theresa Lopez,Ryan Lowe Patricia LueAnna Makanju,Kim Mal" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 334, + 72, + 553, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 553, + 490 + ], + "type": "text", + "content": "facini, Sam Manning, Todor Markov, Yaniv Markovsky, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mely, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Filipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 534 + ], + "type": "text", + "content": "[33] Delin Qu, Bangyan Liao, Huiqing Zhang, Omar Ait-Aider, and Yizhen Lao. Fast rolling shutter correction in the wild. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11778-11795, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "text", + "content": "[34] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene rendering and control. arXiv preprint arXiv:2406.16038, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 657 + ], + "type": "text", + "content": "[35] Delin Qu, Qizhi Chen, Pingrui Zhang, Xianqiang Gao, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Livescene: Language embedding interactive radiance fields for physical scene control and rendering. In Advances in Neural Information Processing Systems, pages 12271-12292. Curran Associates, Inc., 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[36] Delin Qu, Chi Yan, Dong Wang, Jie Yin, Qizhi Chen, Dan Xu, Yiting Zhang, Bin Zhao, and Xuelong Li. Implicit event-rgbd neural slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19584-19594, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5186" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[37] Delin Qu, Haoming Song, Qizhi Chen, Yuanqi Yao, Xinyi Ye, Yan Ding, Zhigang Wang, JiaYuan Gu, Bin Zhao, Dong Wang, and Xuelong Li. Spatialvia: Exploring spatial representations for visual-language-action model, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "text", + "content": "[38] J Redmon. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 198 + ], + "type": "text", + "content": "[39] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 294, + 255 + ], + "type": "text", + "content": "[40] Yiwen Tang, Hongbo Chen, Liyue Qian, Songhan Ge, Mingbo Zhang, and Rui Zheng. Detection of spine curve and vertebral level on ultrasound images using detr. In 2022 IEEE international ultrasonics symposium (IUS), pages 1-4. IEEE, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "type": "text", + "content": "[41] Yiwen Tang, Ray Zhang, Zoey Guo, Xianzheng Ma, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Pointpeft: Parameter-efficient fine-tuning for 3d pre-trained models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5171-5179, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 294, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 379 + ], + "type": "text", + "content": "[42] Yiwen Tang, Ray Zhang, Jiaming Liu, Zoey Guo, Bin Zhao, Zhigang Wang, Peng Gao, Hongsheng Li, Dong Wang, and Xuelong Li. Any2point: Empowering any-modality large models for efficient 3d understanding. In European Conference on Computer Vision, pages 456-473. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "text", + "content": "[43] Yiwen Tang, Zoey Guo, Zhuhao Wang, Ray Zhang, Qizhi Chen, Junli Liu, Delin Qu, Zhigang Wang, Dong Wang, Xuelong Li, and Bin Zhao. Exploring the potential of encoder-free architectures in 3d lmm's, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 472 + ], + "type": "text", + "content": "[44] Johanna Wald, Helisa Dhamo, Nassir Navab, and Federico Tombari. Learning 3d semantic scene graphs from 3d indoor reconstructions. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "text", + "content": "[45] Liwei Wang, Yin Li, Jing Huang, and Svetlana Lazebnik. Learning two-branch neural networks for image-text matching tasks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):394–407, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[46] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "text", + "content": "[47] Haiyang Xu, Qinghao Ye, Ming Yan, Yaya Shi, Jiabo Ye, Yuanhong Xu, Chenliang Li, Bin Bi, Qi Qian, Wei Wang, Guohai Xu, Ji Zhang, Songfang Huang, Fei Huang, and Jingren Zhou. mplug-2: A modularized multi-modal foundation model across text, image and video, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 657 + ], + "type": "text", + "content": "[48] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[49] Bin Yan, Yi Jiang, Jiannan Wu, Dong Wang, Ping Luo, Zehuan Yuan, and Huchuan Lu. Universal instance perception as object discovery and retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15325-15336, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[50] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[51] Yan Zeng, Xinsong Zhang, Hang Li, Jiawei Wang, Jipeng Zhang, and Wangchunshu Zhou. " + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^2" + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "-vlm: All-in-one pretrained model for vision-language tasks, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[52] Haotian* Zhang, Pengchuan* Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[53] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[54] Ruijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao, Hal Daumé III, Andrey Kolobov, Furong Huang, and Jianwei Yang. Tracevla: Visual trace prompting enhances spatial-temporal awareness for generalist robotic policies, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 362 + ], + "type": "text", + "content": "[55] Fengda Zhu, Yi Zhu, Xiaojun Chang, and Xiaodan Liang. Vision-language navigation with self-supervised auxiliary reasoning tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10012-10022, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "text", + "content": "[56] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 4" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5187" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_content_list.json b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c647bb8260b6e1d8b76db5212706b1718914a0e0 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_content_list.json @@ -0,0 +1,1490 @@ +[ + { + "type": "text", + "text": "AETHER: Geometric-Aware Unified World Modeling", + "text_level": 1, + "bbox": [ + 230, + 130, + 767, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoyi Zhu* Yifan Wang* Jianjun Zhou* Wenzheng Chang* Yang Zhou* Zizun Li* Junyi Chen* Chunhua Shen Jiangmiao Pang Tong He USTC Shanghai AI Lab SII SJTU ZJU FDU *Equal Contribution † Corresponding Author https://aether-world.github.io/", + "bbox": [ + 184, + 180, + 810, + 277 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94cd943bbc5d4b733ea437470020d9295c18436b009cb913494dce09c0b22851.jpg", + "image_caption": [ + "Figure 1. An overview of AETHER, trained entirely on synthetic data. The figure highlights its three key capabilities: 4D reconstruction, action-conditioned 4D prediction, and visual planning, all demonstrated on unseen real-world data. The 4D reconstruction examples are derived from MovieGen [48] and Veo 2 [62] generated videos, while the action-conditioned prediction uses an observation image from a university classroom. The visual planning example utilizes observation and goal images from an office building. Better viewed when zoomed in. Additional visualizations can be found on our website." + ], + "image_footnote": [], + "bbox": [ + 91, + 310, + 906, + 843 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 179, + 40 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8535", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 90, + 326, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The integration of geometric reconstruction and generative modeling remains a critical challenge in developing AI systems capable of human-like spatial reasoning. This paper proposes AETHER, a unified framework that enables geometry-aware reasoning in world models by jointly optimizing three core capabilities: (1) 4D dynamic reconstruction, (2) action-conditioned video prediction, and (3) goal-conditioned visual planning. Through task-interleaved feature learning, AETHER achieves synergistic knowledge sharing across reconstruction, prediction, and planning objectives. Building upon video generation models, our framework demonstrates zero-shot synthetic-to-real generalization despite never observing real-world data during training. Furthermore, our approach achieves zero-shot generalization in both action following and reconstruction tasks, thanks to its intrinsic geometric modeling. Notably, even without real-world data, its reconstruction performance is comparable with or even better than that of domain-specific models. Additionally, AETHER employs camera trajectories as geometry-informed action spaces, enabling effective action-conditioned prediction and visual planning. We hope our work inspires the community to explore new frontiers in physically-reasonable world modeling and its applications.", + "bbox": [ + 91, + 123, + 483, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 503, + 220, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "\"Prediction is not just one of the things your brain does. It is the primary function of the neocortex.\"", + "bbox": [ + 86, + 530, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Jeff Hawkins, On Intelligence (2004)", + "bbox": [ + 210, + 561, + 477, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of visual intelligence systems capable of comprehending and forecasting the physical world remains a cornerstone of AI research. World models have emerged as a foundational paradigm for building autonomous systems that not only perceive but also anticipate environmental dynamics to make reasonable actions. At their core, three capabilities stand out: First, perception equips the system with the ability to capture the intricate four-dimensional (4D) changes—integrating spatial and temporal information—that are essential for understanding the physical world [37, 63, 65, 66, 82, 86]. This continuous sensing of dynamic cues enables a geometric representation of the environment. Second, prediction leverages this perceptual information to forecast how the environment will evolve under specific actions, thereby providing a foresight into future states [3, 24, 28, 32, 35, 60, 77]. Finally, planning uses these predictive insights to determine the optimal sequence of actions required to achieve a given goal. Together, these three aspects empower world models to not only represent the current state of the environment but also to anticipate and navigate its future dynamics effectively.", + "bbox": [ + 91, + 583, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these principles, we introduce AETHER, a unified framework that, for the first time, bridges reconstruction, prediction, and planning, as shown in Fig. 1. AETHER leverages pre-trained video generation models [28, 77] and is further refined via post-training with synthetic 4D data. Although multiple action modalities exist, ranging from keyboard inputs [2, 11, 15, 46, 79] to human or robotic motions [16, 84, 89, 90] and point flows [22, 69], we choose camera pose trajectories as our global action representation. This choice is particularly effective for ego-view tasks: in navigation, camera trajectories directly correspond to the navigation paths, while in robotic manipulation, the movement of an in-hand camera captures the 6D motion of the end effector. To address the scarcity of 4D data, we utilize RGB-D synthetic video data and propose a robust camera pose annotation pipeline to reconstruct full 4D dynamics.", + "bbox": [ + 511, + 90, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through a simple training strategy that randomly combines input and output modalities, our method transforms the base video generation model into a unified, multi-task world model with three key capabilities: (1) Depth and camera pose estimation from full video sequences; (2) Video prediction conditioned on an initial observation—with the option to incorporate a camera trajectory action; and (3) Goal-conditioned visual planning based on observation-goal image pairs. We transform depth videos into scale-invariant normalized disparity representations to meet the tokenization requirements of video VAEs. Simultaneously, we encode camera trajectories as scale-invariant raymap sequence representations, structured to align with the spatiotemporal framework of diffusion transformers (DiTs). By dynamically integrating cross-task and cross-modal conditioning signals during training, our framework enables synergistic knowledge transfer across heterogeneous inputs, facilitating joint optimization for multi-task generative modeling.", + "bbox": [ + 511, + 333, + 908, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work introduces AETHER, a unified world model that integrates reconstruction, prediction, and planning through multi-task learning on synthetic 4D data. We propose a robust automatic data annotation pipeline to obtain accurate 4D geometry knowledge. By combining geometric reasoning with generative priors, our framework achieves robust zero-shot transfer to real-world tasks, demonstrating accuracy comparable to SOTA reconstruction models while enabling actionable planning capabilities. The results underscore the value of synergistic 4D modeling for advancing spatial intelligence in AI systems. We hope that AETHER will serve as an effective starter framework for the community to explore post-training world models with scalable synthetic data.", + "bbox": [ + 511, + 604, + 908, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. 4D Synthetic Data Annotation Pipeline", + "text_level": 1, + "bbox": [ + 511, + 829, + 862, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For the synthetic data source, we follow DA-V [74] and TheMatrix [17] to collect large-scale synthetic data with high-quality video depth data. With high-resolution RGB videos", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8536", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7efdfaea582c4f64a08371bfce5c7e861f975166dbefa05849b7a7aae90a8ae8.jpg", + "image_caption": [ + "Figure 2. Some visualization results of data annotated through our pipeline. Better viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 864, + 287 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1df129fba7f8a3074d614967e4545622814889cf489599b48c3a8f3cecacc4f8.jpg", + "image_caption": [ + "Figure 3. Our robust automatic camera annotation pipeline." + ], + "image_footnote": [], + "bbox": [ + 99, + 306, + 470, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and corresponding per-frame depth maps collected, we built a robust and fully automatic camera annotation pipeline for both camera extrinsics and intrinsics. As illustrated in Fig. 3, the pipeline has four stages: (1) object-level dynamic masking, (2) reconstruction-friendly video slicing, (3) coarse camera localization and calibration, and (4) tracking-based camera refinement with bundle adjustment. We present several visualizations of our annotated data in Fig. 2, ranging from indoor to outdoor scenes, and from static to dynamic scenarios, demonstrating the robustness and accuracy of our annotation method.", + "bbox": [ + 89, + 460, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Masking. Distinguishing between dynamic and static regions is crucial for accurate camera parameters estimation. Here, we utilize semantic categories that are potentially dynamic (e.g., cars, people) to segment dynamic objects. Although this may occasionally misclassify static objects, such as stationary parked cars, as dynamic, we find it more robust than flow-based segmentation methods. Specifically, we use Grounded SAM 2 [50] to ensure the temporal consistency of dynamic masks over long sequences.", + "bbox": [ + 89, + 627, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Slicing. Video slicing plays a critical role in 3D reconstruction by serving two key purposes: First, it eliminates unsuitable video segments (such as scene cuts or motion-blurred frames) that could compromise reconstruction quality. Second, it segments long videos into shorter, temporally coherent clips to enhance processing efficiency. The specific criteria for frame removal are as follows: (1) Insufficient Feature Points: We employ the SIFT [39] feature descriptor to extract keypoints from each frame. Frames exhibiting", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "insufficient SIFT keypoints are discarded to ensure robust correspondence estimation. Additionally, frames containing regions with insufficient texture due to low illumination are excluded, as such areas typically exhibit poor feature discriminability and pose challenges for reliable matching. (2) Large Areas of Dynamic Regions: Frames where dynamic regions (obtained from dynamic annotation) dominate over static regions can introduce ambiguity in camera pose estimation. Such frames are filtered out to ensure robust results. (3) Large Motion or Inaccurate Correspondence: Using an off-the-shelf optical flow estimator, RAFT [61], we estimate the magnitude of motion. If these magnitudes exceed a predefined threshold, we truncate the sequence at the current frame, retaining all preceding frames as a valid segment. Similarly, if the ratio of forward-to-backward optical flow errors surpasses a threshold value, we truncate the current frames to ensure temporal coherence.", + "bbox": [ + 511, + 309, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Coarse Camera Estimation. For each video slice, we first use DroidCalib [25] to perform a coarse estimation of the camera parameters, leveraging the depth information from static regions. However, due to the lower input resolution of the DroidCalib model and the limited accuracy of its correspondence estimation, a refinement process is necessary to obtain precise camera parameters.", + "bbox": [ + 511, + 566, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Camera Refinement. We begin camera refinement by employing the state-of-the-art tracker, CoTracker3 [33], to capture accurate long-term correspondences across the entire slice. SIFT [39] and SuperPoint [12] feature points are extracted from static regions, and then tracked to form correspondences. Subsequently, bundle adjustment is performed on all frames to minimize the accumulated reprojection error of all correspondences. With access to high-quality dense depth, we apply forward-backward reprojection to estimate and minimize errors in 3D space [8], which improves per-frame camera accuracy while preserving inter-frame geometric consistency. Specifically, we solve the nonlinear optimization problem by Ceres Solver [1], and the Cauchy loss function is applied to measure correspondence residuals, which accounts for the problem's sparsity.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8537", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b1d7d09061d558099383590e9091c6342f1cecfbd3eea132a6239d3aef80a6a5.jpg", + "image_caption": [ + "Figure 4. The overall pipeline of AETHER. With different condition combinations, AETHER can serve different tasks." + ], + "image_footnote": [ + "Image latent", + "Action latent", + "Image prediction", + "Action prediction", + "Depth prediction", + "Zero-padding", + "Noise" + ], + "bbox": [ + 130, + 85, + 867, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. AETHER Multi-Task World Model", + "text_level": 1, + "bbox": [ + 89, + 328, + 405, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce how we post-train a base video diffusion model into a unified multi-task world model AETHER. We use CogVideoX-5b-I2V [77] as our base model. We first give an overview of our framework in Sec. 3.1, then we detail on the input process of depth videos and camera pose trajectories in Sec. 3.2 and Sec. 3.3. Finally, we show how we do model training in Sec. 3.4.", + "bbox": [ + 89, + 354, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Method Overview", + "text_level": 1, + "bbox": [ + 89, + 468, + 264, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mainstream video diffusion models [27, 40] typically involve two processes: a forward (noising) process and a reverse (denoising) process. The forward process incrementally adds Gaussian noise, denoted as $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ , to a clean latent sample $\\mathbf{z}_0 \\in \\mathbb{R}^{k \\times c \\times h \\times w}$ , where $k, c, h, w$ represent the dimensions of the video latents. Through this process, the clean $\\mathbf{z}_0$ is gradually transformed into a noisy latent $\\mathbf{z}_t$ . In the reverse process, a learned denoising model $\\epsilon_{\\theta}$ progressively removes the noise from $\\mathbf{z}_t$ to reconstruct the original latent representation. The denoising model $\\epsilon_{\\theta}$ is conditioned on auxiliary inputs $\\mathbf{c}$ and the diffusion timestep $t$ .", + "bbox": [ + 89, + 489, + 483, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our method, the target latent $\\mathbf{z}_0$ comprises three modalities: color video latents $\\mathbf{z}_{\\mathrm{c0}}$ , depth video latents $\\mathbf{z}_{\\mathrm{d0}}$ , and action latents $\\mathbf{z}_{\\mathrm{a0}}$ . The model additionally takes two types of conditions as input: color video conditions $\\mathbf{c}_{\\mathrm{c}}$ and action conditions $\\mathbf{c}_{\\mathrm{a}}$ . For the action modality, we choose camera pose trajectory as a global action, facilitated by our automated camera pose annotation pipeline described earlier. All latents and conditions are channel-wise concatenated. The training objective of AETHER can be expressed as:", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {\\substack {\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\\\ t \\sim \\mathcal {U} (1, \\mathcal {T}) \\\\ \\mathbf {z} _ {0} = \\mathbf {z} _ {\\mathrm {c} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {d} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {a} _ {0}} \\\\ \\mathbf {c} = \\mathbf {c} _ {\\mathrm {c}} \\otimes \\mathbf {c} _ {\\mathrm {a}}}} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| ^ {2}, \\tag{1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 805, + 483, + 852 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes the channel-wise concatenation operation, $\\mathcal{U}(\\cdot)$ represents a uniform distribution, and $\\mathcal{T}$ denotes the denoising steps.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multi-task objective of AETHER is determined by the specific conditions $\\mathbf{c}$ for different tasks. (1) Reconstruction: $\\mathbf{c}_{\\mathrm{c}}$ represents the input video latents. (2) Video prediction: $\\mathbf{c}_{\\mathrm{c}}$ takes the latent of observation image as the first frame, while other latents are zero-masked. (3) Goal-conditioned visual planning: The first and last latents of $\\mathbf{c}_{\\mathrm{c}}$ correspond to the observation and goal images, respectively, with all intermediate latents zero-padded. For the action condition $\\mathbf{c}_{\\mathrm{a}}$ , it is either entirely zero-masked or contains the full target camera pose trajectory in action-free or action-conditioned control cases. Illustrations are show in Fig. 4.", + "bbox": [ + 511, + 329, + 906, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Depth Videos Process", + "text_level": 1, + "bbox": [ + 511, + 505, + 714, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a depth video $\\mathbf{x}_d$ , we first clip the depth values to a predefined range $[d_{\\min}, d_{\\max}]$ . Next, we apply a square root transformation and subsequently compute the reciprocal to convert the depth values into disparity, as described in [57]. Each disparity video clip is then normalized in a scale-invariant manner. Subsequently, the normalized disparity values are linearly mapped from $[0,1]$ to $[-1,1]$ . To meet the input requirements of the VAE, the single-channel disparity map is replicated across three channels, as done in prior works [34, 74]. The final depth latent is computed as:", + "bbox": [ + 511, + 526, + 906, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {\\text {d i s p}} = \\frac {1}{\\sqrt {\\operatorname {c l i p} \\left(\\mathbf {x} _ {d} , d _ {\\min } , d _ {\\max }\\right)}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 685, + 906, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {\\text {d i s p}} = \\frac {\\mathbf {x} _ {\\text {d i s p}}}{\\operatorname* {m a x} \\left(\\mathbf {x} _ {\\text {d i s p}}\\right)} \\times 2 - 1, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 720, + 906, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {d} = \\mathcal {E} \\left(\\hat {\\mathbf {x}} _ {\\text {d i s p}} \\otimes \\mathbf {1} _ {3}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 753, + 906, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ denotes the 3D VAE, and $\\otimes \\mathbf{1}_3$ represents the channel-wise replication of 3 times. The above operations are designed to be compatible with the pretrained 3D VAE model, ensuring minimal reconstruction error.", + "bbox": [ + 511, + 779, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Camera Trajectories Process", + "text_level": 1, + "bbox": [ + 511, + 848, + 769, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We transform camera parameters into raymap videos [7] so that video diffusion can process them compatibly. Specifi", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8538", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cally, given the intrinsic matrix $\\mathbf{K} \\in \\mathbb{R}^{T \\times 3 \\times 3}$ and the extrinsic matrix $\\mathbf{E} \\in \\mathbb{R}^{T \\times 4 \\times 4}$ , the transformation process can be described as follows.", + "bbox": [ + 89, + 90, + 485, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Translation Scaling and Normalization. The translation component of the camera pose (inverse of extrinsic matrix), $\\mathbf{t} \\in \\mathbb{R}^3$ , is first scaled by a constant factor $s_{\\mathrm{ray}}$ and normalized using the maximum disparity value $d_{\\mathrm{max}}$ . To suppress large values, we then pass it through a signed $\\log(1 + \\cdot)$ transformation:", + "bbox": [ + 89, + 136, + 485, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} ^ {\\prime} = \\frac {\\mathbf {t}}{\\operatorname* {m a x} \\left(x _ {\\text {d i s p}}\\right)} \\cdot s _ {\\text {r a y}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 233, + 482, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} _ {\\log} = \\operatorname {s i g n} \\left(\\mathbf {t} ^ {\\prime}\\right) \\cdot \\log \\left(1 + \\left| \\mathbf {t} ^ {\\prime} \\right|\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 267, + 483, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_{\\mathrm{ray}}$ is a predefined scaling factor.", + "bbox": [ + 89, + 292, + 359, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Raymap Construction. Using the intrinsic matrix $\\mathbf{K}$ , we compute the camera ray directions $\\mathbf{r}_{\\mathrm{d}}$ in homogeneous coordinates for each pixel. Note that we do not unit normalize it but let it have a unit value along $z$ axis. The ray origins $\\mathbf{r}_{\\mathrm{o}}$ are set to the translation $\\mathbf{t}_{\\log}$ , replicated across the spatial dimensions. The raymap in the world coordinate system is obtained by transforming the ray directions $\\mathbf{r}_{\\mathrm{d}}$ using the extrinsic matrix $\\mathbf{E}$ . The final raymap $\\mathbf{r}$ consists of 6 channels: 3 for the ray directions $\\mathbf{r}_{\\mathrm{d}}$ and 3 for the ray origins $\\mathbf{r}_{\\mathrm{o}}$ .", + "bbox": [ + 89, + 308, + 485, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Resolution Downsampling. To align the raymap with the latent feature dimensions from the VAE, we perform adjustments both spatially and temporally. Spatially, the raymap is downsampled by a factor of 8 using bilinear interpolation. Temporally, every consecutive group of 4 frames is concatenated along the channel dimension. The resulting rearranged tensor is denoted as $\\mathbf{z}_a$", + "bbox": [ + 89, + 444, + 483, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Converting raymap back to camera matrix. Given generated raymap sequences rearranged by the time axis $\\hat{\\mathbf{r}}\\in$ $\\mathbb{R}^{T\\times 6\\times h\\times w} = [\\hat{\\mathbf{r}}_{\\mathrm{d}},\\hat{\\mathbf{r}}_{\\mathrm{o}}]$ , we first recover the ray origins by:", + "bbox": [ + 89, + 550, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {r} _ {\\mathrm {o}}} ^ {\\prime} = \\frac {1}{s _ {\\text {r a y}}} \\cdot \\operatorname {s i g n} \\left(\\hat {\\mathbf {r} _ {\\mathrm {o}}}\\right) \\cdot \\left(\\exp \\left(\\left| \\hat {\\mathbf {r} _ {\\mathrm {o}}} \\right|\\right) - 1\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 603, + 483, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we can recover both the intrinsic and extrinsic through Alg. 1 in the supplementary material.", + "bbox": [ + 89, + 643, + 482, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 89, + 681, + 246, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize AETHER with pre-trained CogVideo-5b-I2V [77] weights, excluding the additional input and output projection layer channels for depth and raymap action trajectories, which are initialized to zero. Since text prompt conditions are not used, an empty text embedding is provided during both training and inference.", + "bbox": [ + 89, + 703, + 483, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the dataset we use contains video clips with variable lengths and frames per second (FPS), we randomly select $T \\in \\{17, 25, 33, 41\\}$ frames, and the FPS is randomly sampled from $\\{8, 10, 12, 15, 24\\}$ . The RoPE [59] coefficients are linearly interpolated to align with them.", + "bbox": [ + 89, + 795, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, conditional inputs are randomly masked to generalize across tasks. For $\\mathbf{c}_c$ , masking probabilities are:", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$30\\%$ for both observation and goal images (visual planning tasks), $40\\%$ for observation images only (video prediction), $28\\%$ for full-color video latents (4D reconstruction), and $2\\%$ for masking all of $\\mathbf{c}_c$ . For $\\mathbf{c}_a$ , trajectory latents are either kept or fully masked with equal probability (supporting action-free or action-conditioned tasks with raymap conditions). This strategy enables the model to adapt to diverse tasks and input condition settings.", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training process consists of two stages. In the first stage, we adopt the loss function of a standard latent diffusion model, which minimizes the mean squared error (MSE) in the latent space. In the second stage, we refine the generated outputs by decoding them into the image space. Specifically, we introduce three additional loss terms: a Multi-Scale Structure Similarity (MS-SSIM) loss [67] for color video, a scale- and shift-invariant loss [49] for depth videos, and a scale- and shift-invariant pointmap loss [66] for pointmaps projected from the generated depths and raymaps. Further details on the stage 2 loss functions are provided in the supplementary material. Notably, the second stage takes about $\\frac{1}{4}$ of the training steps used in the first stage.", + "bbox": [ + 511, + 212, + 908, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ a hybrid training strategy combining Fully Sharded Data Parallel (FSDP) [87] with Zero-2 optimization within compute nodes and Distributed Data Parallel (DDP) across nodes. Since depth videos require online normalization, the VAE encoder is also run online during training and operates under DDP. Our implementation processes a local batch size of 4 per GPU, resulting in an effective batch size of 320 samples across 80 A100-80GB GPUs. Training is conducted over two weeks using the AdamW [38] optimizer with a OneCycle [56] learning rate scheduler.", + "bbox": [ + 511, + 407, + 908, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Reconstruction Experiments", + "text_level": 1, + "bbox": [ + 511, + 571, + 777, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate that AETHER can achieve zero-shot reconstruction metrics comparable to or even better than SOTA reconstruction methods. We mainly consider two zero-shot reconstruction tasks: video depth estimation and camera pose estimation. Note that we only denoise for 4 steps for reconstruction tasks.", + "bbox": [ + 511, + 598, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Zero-Shot Video Depth Estimation", + "text_level": 1, + "bbox": [ + 511, + 696, + 815, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Video depth estimation is evaluated based on two key aspects: per-frame depth quality and inter-frame depth consistency. These evaluations are performed by aligning the predicted depth maps with the ground truth using a per-sequence scale. We use absolute relative error (Abs Rel) and $\\delta < 1.25$ (percentage of predicted depths within a 1.25-factor of true depth) as metrics. For implementation, we adopt the settings outlined in CUT3R [65]. Our baselines include both reconstruction-based methods—such as DUSt3R [66], MASt3R [37], MonST3R [82], Spann3R [63], and CUT3R [65]—and diffusion-based depth estimators, including ChronoDepth [55], DepthCrafter [29],", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8539", + "bbox": [ + 482, + 954, + 514, + 965 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26844e565e3dd0d2b445af6cdca1e0bac3f64aaf1e3bfa700d646580efb69b42.jpg", + "table_caption": [ + "Table 1. Video depth Evaluation. Methods requiring global alignment are marked \"GA\"." + ], + "table_footnote": [], + "table_body": "
MethodSintel [6]BONN [44]KITTI [21]
Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑
Reconstruction Methods. Alignment: per-sequence scale
DUSt3R-GA [66]0.65645.20.15583.30.14481.3
MASt3R-GA [37]0.64143.90.25270.10.18374.5
MonST3R-GA [82]0.37855.80.06796.30.16874.4
Spann3R [63]0.62242.60.14481.30.19873.7
CUT3R [65]0.42147.90.07893.70.11888.1
AETHER (Ours)0.32450.20.27359.40.05697.8
Diffusion-Based Methods. Alignment: per-sequence scale&shift
ChronoDepth [55]0.42938.30.31851.80.25254.3
DepthCrafter [29]0.59055.50.25356.30.12486.5
DA-V [74]1.25243.70.45731.10.09493.0
AETHER (Ours)0.31460.40.30860.20.05497.7
", + "bbox": [ + 93, + 103, + 906, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2162ce6860718b2cad0a7b157b188b622065f02bfa6f350b35cf46a3e632a1b4.jpg", + "table_caption": [ + "Table 2. Evaluation on Camera Pose Estimation." + ], + "table_footnote": [], + "table_body": "
MethodSintel [6]TUM-dynamics [58]ScanNet [10]
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Optimization-based Methods
Particle-SfM [86]0.1290.0310.535---0.1360.0230.836
Robust-CVD [36]0.3600.1543.4430.1530.0263.5280.2270.0647.374
CasualSAM [85]0.1410.0350.6150.0710.0101.7120.1580.0341.618
DUSt3R-GA [66]0.4170.2505.7960.0830.0173.5670.0810.0280.784
MASt3R-GA [37]0.1850.0601.4960.0380.0120.4480.0780.0200.475
MonST3R-GA [82]0.1110.0440.8960.0980.0190.9350.0770.0180.529
Feed-forward Methods
DUSt3R [66]0.2900.1327.8690.1400.1063.2860.2460.1088.210
Spann3R [63]0.3290.1104.4710.0560.0210.5910.0960.0230.661
CUT3R [65]0.2130.0660.6210.0460.0150.4730.0990.0220.600
AETHER (Ours)0.1890.0540.6940.0920.0121.1060.1760.0281.204
", + "bbox": [ + 91, + 313, + 911, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and DepthAnyVideo (DA-V) [74]. It is important to note that when comparing with diffusion-based depth estimators, we apply scale and shift alignment to the ground truth, as most of these methods are not inherently scale-invariant. All videos are resized with original aspect ratios kept to make the short side align with our model's input size. For videos that exceed the maximum forward processing spatial or temporal size of our model, we employ a sliding window strategy with a stride size of 8. In regions of overlap between windows, we first estimate a relative scale by calculating the average of element-wise division. This relative scale is then used to adjust the latter window's depth predictions. Finally, a linspace-weighted average is applied to the overlapping areas, following approaches similar to prior methods [29, 80].", + "bbox": [ + 88, + 518, + 485, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and Analysis. Table 1 summarizes the video depth estimation results across Sintel [6], BONN [44], and KITTI [21] datasets. For reconstruction-based methods, AETHER outperforms or is comparable with prior approaches. On Sintel, AETHER achieves the lowest Abs Rel (0.324), surpassing MonST3R-GA (0.378), and competitive $\\delta < 1.25$ (50.2). On KITTI, AETHER sets a new benchmark with Abs Rel of 0.056 and $\\delta < 1.25$ of 97.8, outperforming the previous SOTA CUT3R (Abs Rel: 0.118, $\\delta < 1.25$ : 88.1). Among diffusion-based methods, AETHER shows consistent superiority. It achieves the best performance on", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sintel (Abs Rel: 0.314, $\\delta < 1.25$ : 60.4) and KITTI (Abs Rel: 0.054, $\\delta < 1.25$ : 97.7), significantly outperforming ChronoDepth [55], DepthCrafter [29], and DA-V [74]. On BONN, AETHER achieves the highest $\\delta < 1.25$ (60.2) with competitive Abs Rel (0.308).", + "bbox": [ + 511, + 517, + 906, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Zero-Shot Camera Pose Estimation", + "text_level": 1, + "bbox": [ + 511, + 604, + 821, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Following MonST3R [82] and CUT3R [65], we evaluate camera pose estimation accuracy on the Sintel [6], TUM Dynamics [58], and ScanNet [10] datasets. Notably, both Sintel and TUM Dynamics contain highly dynamic objects, presenting significant challenges for traditional Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM) systems. We report Absolute Translation Error (ATE), Relative Translation Error (RPE Trans), and Relative Rotation Error (RPE Rot) after Sim(3) alignment with the ground truth, following the methodology in [65]. The implementation settings are consistent with those used in CUT3R [65]. All videos are resized with original aspect ratios kept and then center cropped to align with our model's input size. For long videos exceeding our model's maximum temporal forward processing length, a sliding window strategy with a stride size of 32 is employed. In overlapping regions between windows, camera poses are aligned following prior methods [64]. Transla", + "bbox": [ + 509, + 628, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8540", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion alignment is performed using linear interpolation, while quaternion rotations are interpolated with spherical linear interpolation. Additionally, we observed that the generated camera trajectories exhibit noise, likely due to the limited number of denoising steps. To mitigate this, we apply a simple Kalman filter [71] to smooth the trajectories.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Analysis. Table 2 shows the evaluation results. Among feed-forward methods, AETHER achieves the best ATE (0.189) and RPE Trans (0.054) on Sintel [6], while remaining competitive in RPE Rot (0.694) compared to CUT3R (0.621). On TUM Dynamics [58], AETHER achieves the best RPE Trans (0.012). For other metrics, AETHER is also comparable with other specialist models.", + "bbox": [ + 89, + 181, + 485, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Generation and Planning Experiments", + "text_level": 1, + "bbox": [ + 89, + 301, + 437, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we first show video prediction, with or without action conditioning, quantitatively or qualitatively, in Sec. 5.1. We then show visual planning abilities in Sec. 5.2. More visualizations are in the supplementary material.", + "bbox": [ + 89, + 325, + 485, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Video Prediction", + "text_level": 1, + "bbox": [ + 89, + 395, + 256, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We use CogVideoX-5b-I2V [77] as our baseline. To ensure a fair comparison, we construct a validation dataset comprising two subsets: in-domain and out-domain data. The in-domain subset includes novel, unseen scenes from the same synthetic environments as the training dataset, while the out-domain subset consists of data from entirely new synthetic environments. Both models are provided with the first frame as the observation image. For action-free prediction, since CogVideoX depends heavily on text prompts, we use GPT-4o [31] to generate image descriptions and predictions of future scenes as prompts for CogVideoX. In contrast, AETHER is evaluated using empty text prompts. For action-conditioned prediction, we also labeled camera trajectories in the validation dataset and generated corresponding raymap sequences as action conditions for AETHER. For the baseline, in addition to the prompts used for action-free prediction, we use GPT-4o [31] to generate detailed descriptions of object and camera movements, enabling the baseline to use language as action conditions. We use the default classifier-free guidance value of 6 on text prompts for CogVideoX and a value of 3 on the observation image for AETHER. No classifier-free guidance is applied to action conditions to ensure fairness. Evaluation metrics follow VBenchmark [30], a standard benchmark for video generation, with additional details on prompts and evaluation metrics provided in the supplementary material.", + "bbox": [ + 89, + 417, + 485, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image-to-Video Prediction. We first evaluate image-to-video prediction without action conditions. The results, presented in Tab. 3, show that AETHER consistently outperforms the baseline on both in-domain and out-domain validation sets. Notably, AETHER demonstrates a larger performance improvement on out-domain data, which can likely be at", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tributed to the baseline model's pre-training data containing domains similar to the in-domain dataset.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Action-Conditioned Video Prediction. To assess the effectiveness of our post-training in improving action control and action-following capabilities, we conduct action-conditioned video prediction experiments. The results, shown in Tab. 4, indicate that AETHER consistently outperforms the baseline in both in-domain and out-domain settings. Notably, CogVideoX tends to generate static scenes with high visual and aesthetic quality, while AETHER accurately follows the action conditions, producing highly dynamic scenes. These results validate the effectiveness of our framework and the advantages of using camera trajectories as action conditions.", + "bbox": [ + 511, + 121, + 908, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Visual Planning", + "text_level": 1, + "bbox": [ + 511, + 296, + 671, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. We evaluate the action-conditioned navigation capability of AETHER on our validation set. To demonstrate the effectiveness of our multi-task objective, particularly the incorporation of the reconstruction objective, we also post-train an ablation model without the video depth objective, denoted as AETHER-no-depth. Given the observation image, goal image, and camera trajectory, the resulting video should be highly determined. Thus, we report pixel-wise reconstruction metrics, including PSNR, SSIM [68], MS-SSIM [67], and LPIPS [83], for action-conditioned navigation. For the action-free case, which represents a visual path navigation task, we also report the VBench metrics. We do not use any classifier-free guidance on both tasks.", + "bbox": [ + 511, + 318, + 908, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Action-Conditioned Navigation. The quantitative results for action-conditioned navigation are presented in Tab. 5. Aether consistently outperforms the ablation model, demonstrating the significant benefits of incorporating the reconstruction objective into generative models.", + "bbox": [ + 511, + 530, + 908, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual Path Planning. In the absence of action conditions, this task evaluates the model's ability to function as a \"world model as an agent,\" requiring it to plan a path from the observation image to the goal image. The results, shown in Tab. 6, indicate that the reconstruction objective substantially improves the model's visual path planning capability. Additionally, qualitative visualizations on completely in-the-wild data are provided in supplementary material.", + "bbox": [ + 511, + 604, + 908, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Related Work", + "text_level": 1, + "bbox": [ + 511, + 739, + 656, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "World Models. World models have emerged as a critical framework in artificial intelligence, enabling agents to simulate, understand, and predict environmental dynamics. Early work [23] introduced latent representations and recurrent neural networks for decision-making. Recent advancements include Cat3D [20] for 3D scene generation, Cat4D [72] for dynamic 4D environments, and Genie 2 [46], a large-scale model for interactive 3D worlds. Motion Prompting [22] further enables precise video generation control. These ad", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8541", + "bbox": [ + 482, + 955, + 513, + 967 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/baacf544f1d57e77dde07620482f80e8fe54b2b5d9aef374fd2dfe13fd6842b5.jpg", + "table_caption": [ + "Table 3. VBenchmark [30] Metrics of Video Prediction without Action Conditions. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each group, the better performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX89.36/84.61/87.7792.72/91.43/92.2998.24/96.93/97.8188.75/95.00/90.8354.49/53.58/54.1855.38/52.29/54.3579.01/77.52/78.51
AETHER91.50/87.55/90.1894.29/93.62/94.0798.54/98.19/98.4296.25/100.00/97.5054.36/52.58/53.7755.08/54.88/55.0180.34/79.42/80.04
", + "bbox": [ + 91, + 116, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5449ad1792616d2b5a2228f7f3d1e6bc095ca98d86c54af90ba2a8a77b80bae0.jpg", + "table_caption": [ + "Table 4. VBenchmark [30] Metrics of Action-Conditioned Video Prediction. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each metric group, the better performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX91.56/88.23/90.5192.98/92.29/92.7798.44/97.81/98.2483.87/93.02/86.7656.19/57.43/56.5856.48/61.60/58.1079.56/80.70/79.92
AETHER90.73/93.27/91.5493.61/95.03/94.0698.53/98.62/98.56100.00/83.72/94.8555.04/56.50/55.5053.89/63.23/56.8480.33/81.55/80.71
", + "bbox": [ + 91, + 199, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa2aa8f6e5fdabf4ba898fb89555e7bb42fc23071b304c87153913193a808015.jpg", + "table_caption": [ + "Table 5. Pixel-wise Metrics of Action-Conditioned Navigation. Comparison of performance between AETHER-no-depth and AETHER on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓
AETHER-no-depth19.13/18.67/18.970.5630/0.4830/0.53530.5467/0.5204/0.53760.3116/0.2995/0.3074
AETHER19.87/19.37/19.700.5803/0.5058/0.55450.5830/0.5627/0.57600.2691/0.2599/0.2659
", + "bbox": [ + 91, + 282, + 903, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2c9bcbfc0c51b0409a14764144dfe6e57906bd8e0c62629001b97846d855b8f5.jpg", + "table_caption": [ + "Table 6. Quantitative Results of Action-Free Visual Path Planning. Comparison of performance between Aether and Aether-no-depth on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
Aether-no-depth88.68/89.61/88.6193.62/93.92/93.6698.37/98.31/98.3297.06/91.67/96.1554.12/56.26/54.7851.77/58.46/54.2979.11/80.43/79.59
Aether (Ours)89.69/91.61/90.3693.88/94.58/94.1398.50/98.40/98.4697.06/91.67/95.1955.83/56.87/56.1954.71/61.13/56.9380.21/81.53/80.67
", + "bbox": [ + 91, + 364, + 916, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vancements demonstrate the evolution of world models toward dynamic, interactive, and controllable applications in robotics, gaming, and simulation.", + "bbox": [ + 88, + 431, + 482, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reconstruction. Reconstruction has been a long-standing topic in computer vision, with notable progress in both traditional and learning-based methods. Classical approaches, such as Structure-from-Motion (SfM) [9, 26, 45, 53] and Multi-View Stereo (MVS) [19, 54], rely on multi-view geometry for feature matching, pose estimation, and dense point cloud generation, demonstrating robust performance in controlled settings. Deep learning has introduced powerful alternatives, tackling sub-tasks like feature matching [14, 52], point tracking [13, 64], triangulation [42], and MVS [78, 81]. End-to-end methods now directly predict point maps [37, 66] or depth maps from images [4, 76], often incorporating camera parameters [70]. Recently, diffusion models have achieved breakthroughs in image and video generation [27, 35, 43, 73, 77], inspiring novel 3D reconstruction approaches that leverage rich 2D priors [18, 29, 34, 41, 74, 75, 88, 90]. These methods demonstrate the potential of integrating diffusion-based 2D knowledge into 3D modeling.", + "bbox": [ + 88, + 477, + 482, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Generation. Video generation has evolved from foundational techniques like DDPM [27, 43] to modern frameworks leveraging diffusion-based techniques. Advances such as latent diffusion [51] and diffusion transformers [47] have improved generation quality, while models like Sora [5] and Stable Video Diffusion (SVD) [3] emphasize temporal consistency. Open-source models, including LTX Video [24], CogVideoX [77], and Hunyuan Video [35], offer increased flexibility, and techniques like multi-scale architectures (e.g.,", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pyramid Flow [32]) enhance motion dynamics. These advancements highlight rapid progress, with ongoing efforts to improve scalability and temporal stability.", + "bbox": [ + 511, + 431, + 906, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 489, + 769, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce AETHER, a geometry-aware multitask world model that reconstructs 4D dynamic videos, predicts future frames conditioned on observation images and actions, and performs visual planning based on observation and goal images. We propose an automatic 4D synthetic data labeling pipeline, enabling AETHER to train on synthetic data and generalize to unseen real-world data in a zero-shot manner. Post-trained on the CogVideoX base model, AETHER achieves state-of-the-art or competitive reconstruction performance and outperforms baselines in generation and planning tasks, demonstrating the value of incorporating reconstruction objectives into world modeling.", + "bbox": [ + 509, + 515, + 908, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, limitations remain. Camera pose estimation is less accurate, likely due to incompatibilities between raymap representation and prior video diffusion models. Indoor scene reconstruction also lags behind outdoor performance, likely due to the predominance of outdoor training data. Additionally, predictions without language prompts often fail in highly dynamic scenes. Future work can address these by exploring novel action representations, co-training with real-world data, and retaining the language prompting capabilities of the base model.", + "bbox": [ + 509, + 696, + 908, + 848 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8542", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Keir Mierle, et al. Ceres solver: Tutorial & reference. Google Inc, 2(72):8, 2012. 3", + "[2] Eloi Alonso, Adam Jolley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37:58757-58791, 2025. 2", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 8", + "[4] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8", + "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8", + "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 6, 7", + "[7] Junyi Chen, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Where am i and what will i see: An auto-regressive model for spatial localization and view prediction, 2024. 4", + "[8] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 3", + "[9] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 8", + "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6", + "[11] Decart, Julian Quevedo, Quinn McIntyre, Spruce Campbell, Xinlei Chen, and Robert Wachen. Oasis: A universe in a Transformer. 2024. 2", + "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 93, + 200, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition workshops, pages 224-236, 2018. 3", + "[13] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 8", + "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 8", + "[15] Tom Erez, Yuval Tassa, and Emanuel Todorov. Infinite-horizon model predictive control for periodic tasks with contacts. Robotics: Science and systems VII, 73, 2012. 2", + "[16] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2", + "[17] Ruili Feng, Han Zhang, Zhantao Yang, Jie Xiao, Zhilei Shu, Zhiheng Liu, Andy Zheng, Yukun Huang, Yu Liu, and Hongyang Zhang. The matrix: Infinite-horizon world generation with real-time moving control. arXiv preprint arXiv:2412.03568, 2024. 2", + "[18] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2024. 8", + "[19] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 8", + "[20] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 7", + "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6", + "[22] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, Chen Sun, Oliver Wang, Andrew Owens, and Deqing Sun. Motion prompting: Controlling video generation with motion trajectories, 2024. 2, 7", + "[23] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 7", + "[24] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 2, 8", + "[25] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8543", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 8", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4, 8", + "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[29] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 8", + "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7, 8", + "[31] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7", + "[32] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2, 8", + "[33] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3", + "[34] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4, 8", + "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Junkun Yuan, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yanxin Long, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. Hunyuanvideo: A systematic framework for large video generative models, 2024. 2, 8", + "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 6" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 5, 6, 8", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3", + "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4", + "[41] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 8", + "[42] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 8", + "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 8", + "[44] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 6", + "[45] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, pages 58-77. Springer, 2024. 8", + "[46] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Roktaschel. Genie 2: A large-scale foundation world model. 2024. 2, 7", + "[47] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 8", + "[48] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8544", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Eliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2025. 1", + "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 5", + "[50] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 3", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8", + "[52] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 8", + "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 8", + "[54] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 8", + "[55] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Vitor Guizilini, Yue Wang, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv preprint arXiv:2406.01493, 2024. 5, 6", + "[56] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5", + "[57] Ziyang Song, Zerong Wang, Bo Li, Hao Zhang, Ruijie Zhu, Li Liu, Peng-Tao Jiang, and Tianzhu Zhang. Depthmaster:" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Taming diffusion models for monocular depth estimation. arXiv preprint arXiv:2501.02576, 2025.4", + "[58] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 6, 7", + "[59] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[60] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2", + "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3", + "[62] Veo-Team,.; Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathewson, Kurtis David, Matthieu Kim Lorraine, Marc van Zee, Medhini Narasimhan, Miaosen Wang, Mohammad Babaeizadeh, Nelly Papalampidi, Nick Pezzotti, Nilpa Jha, Parker Barnes, Pieter-Jan Kindermans, Rachel Hornung, Ruben Villegas, Ryan Poplin, Salah Zaiem, Sander Dieleman, Sayna Ebrahimi, Scott Wisdom, Serena Zhang, Shlomi Fruchter, Signe Nørly, Weizhe Hua, Xinchen Yan, Yuqing Du, and Yutian Chen. Veo 2. 2024. 1", + "[63] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 5, 6", + "[64] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 6, 8", + "[65] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025. 2, 5, 6", + "[66] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 5, 6, 8", + "[67] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 5, 7", + "[68] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Mo" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8545", + "bbox": [ + 482, + 955, + 514, + 965 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tionctrl: A unified and flexible motion controller for video generation. 2023. 2", + "[70] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xiangyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 230-247. Springer, 2020. 8", + "[71] Greg Welch, Gary Bishop, et al. An introduction to the kalman filter. 1995. 7", + "[72] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 7", + "[73] Wanghan Xu, Xiaoyu Yue, Zidong Wang, Yao Teng, Wenlong Zhang, Xihui Liu, Luping Zhou, Wanli Ouyang, and Lei Bai. Exploring representation-aligned latent space for better generation. arXiv preprint arXiv:2502.00359, 2025. 8", + "[74] Honghui Yang, Di Huang, Wei Yin, Chunhua Shen, Haifeng Liu, Xiaofei He, Binbin Lin, Wanli Ouyang, and Tong He. Depth any video with scalable synthetic data. arXiv preprint arXiv:2410.10815, 2024. 2, 4, 6, 8", + "[75] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, et al. Unipad: A universal pre-training paradigm for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15238-15250, 2024. 8", + "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 8", + "[77] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5, 7, 8", + "[78] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 8", + "[79] Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Creating new games with generative interactive videos, 2025. 2", + "[80] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 6", + "[81] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1):199-214, 2023. 8", + "[82] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[84] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance, 2024. 2", + "[85] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 6", + "[86] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2, 6", + "[87] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023.5", + "[88] Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Hengshuang Zhao, Chunhua Shen, Yu Qiao, Tong He, et al. Ponderv2: Pave the way for 3d foundation model with a universal pre-training paradigm. arXiv preprint arXiv:2310.08586, 2023. 8", + "[89] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2", + "[90] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2, 8" + ], + "bbox": [ + 516, + 90, + 906, + 612 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8546", + "bbox": [ + 482, + 955, + 516, + 965 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_model.json b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7b8df4231a645687634a5c485fa4b84dda126019 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_model.json @@ -0,0 +1,2545 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.18, + 0.041 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.24, + 0.001, + 0.807, + 0.045 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.768, + 0.154 + ], + "angle": 0, + "content": "AETHER: Geometric-Aware Unified World Modeling" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.181, + 0.811, + 0.278 + ], + "angle": 0, + "content": "Haoyi Zhu* Yifan Wang* Jianjun Zhou* Wenzheng Chang* Yang Zhou* Zizun Li* Junyi Chen* Chunhua Shen Jiangmiao Pang Tong He USTC Shanghai AI Lab SII SJTU ZJU FDU *Equal Contribution † Corresponding Author https://aether-world.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.311, + 0.907, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.853, + 0.907, + 0.922 + ], + "angle": 0, + "content": "Figure 1. An overview of AETHER, trained entirely on synthetic data. The figure highlights its three key capabilities: 4D reconstruction, action-conditioned 4D prediction, and visual planning, all demonstrated on unseen real-world data. The 4D reconstruction examples are derived from MovieGen [48] and Veo 2 [62] generated videos, while the action-conditioned prediction uses an observation image from a university classroom. The visual planning example utilizes observation and goal images from an office building. Better viewed when zoomed in. Additional visualizations can be found on our website." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.515, + 0.967 + ], + "angle": 0, + "content": "8535" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.091, + 0.327, + 0.108 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.124, + 0.485, + 0.471 + ], + "angle": 0, + "content": "The integration of geometric reconstruction and generative modeling remains a critical challenge in developing AI systems capable of human-like spatial reasoning. This paper proposes AETHER, a unified framework that enables geometry-aware reasoning in world models by jointly optimizing three core capabilities: (1) 4D dynamic reconstruction, (2) action-conditioned video prediction, and (3) goal-conditioned visual planning. Through task-interleaved feature learning, AETHER achieves synergistic knowledge sharing across reconstruction, prediction, and planning objectives. Building upon video generation models, our framework demonstrates zero-shot synthetic-to-real generalization despite never observing real-world data during training. Furthermore, our approach achieves zero-shot generalization in both action following and reconstruction tasks, thanks to its intrinsic geometric modeling. Notably, even without real-world data, its reconstruction performance is comparable with or even better than that of domain-specific models. Additionally, AETHER employs camera trajectories as geometry-informed action spaces, enabling effective action-conditioned prediction and visual planning. We hope our work inspires the community to explore new frontiers in physically-reasonable world modeling and its applications." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.504, + 0.222, + 0.52 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.531, + 0.483, + 0.56 + ], + "angle": 0, + "content": "\"Prediction is not just one of the things your brain does. It is the primary function of the neocortex.\"" + }, + { + "type": "text", + "bbox": [ + 0.211, + 0.562, + 0.478, + 0.575 + ], + "angle": 0, + "content": "Jeff Hawkins, On Intelligence (2004)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The development of visual intelligence systems capable of comprehending and forecasting the physical world remains a cornerstone of AI research. World models have emerged as a foundational paradigm for building autonomous systems that not only perceive but also anticipate environmental dynamics to make reasonable actions. At their core, three capabilities stand out: First, perception equips the system with the ability to capture the intricate four-dimensional (4D) changes—integrating spatial and temporal information—that are essential for understanding the physical world [37, 63, 65, 66, 82, 86]. This continuous sensing of dynamic cues enables a geometric representation of the environment. Second, prediction leverages this perceptual information to forecast how the environment will evolve under specific actions, thereby providing a foresight into future states [3, 24, 28, 32, 35, 60, 77]. Finally, planning uses these predictive insights to determine the optimal sequence of actions required to achieve a given goal. Together, these three aspects empower world models to not only represent the current state of the environment but also to anticipate and navigate its future dynamics effectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.333 + ], + "angle": 0, + "content": "Motivated by these principles, we introduce AETHER, a unified framework that, for the first time, bridges reconstruction, prediction, and planning, as shown in Fig. 1. AETHER leverages pre-trained video generation models [28, 77] and is further refined via post-training with synthetic 4D data. Although multiple action modalities exist, ranging from keyboard inputs [2, 11, 15, 46, 79] to human or robotic motions [16, 84, 89, 90] and point flows [22, 69], we choose camera pose trajectories as our global action representation. This choice is particularly effective for ego-view tasks: in navigation, camera trajectories directly correspond to the navigation paths, while in robotic manipulation, the movement of an in-hand camera captures the 6D motion of the end effector. To address the scarcity of 4D data, we utilize RGB-D synthetic video data and propose a robust camera pose annotation pipeline to reconstruct full 4D dynamics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.909, + 0.605 + ], + "angle": 0, + "content": "Through a simple training strategy that randomly combines input and output modalities, our method transforms the base video generation model into a unified, multi-task world model with three key capabilities: (1) Depth and camera pose estimation from full video sequences; (2) Video prediction conditioned on an initial observation—with the option to incorporate a camera trajectory action; and (3) Goal-conditioned visual planning based on observation-goal image pairs. We transform depth videos into scale-invariant normalized disparity representations to meet the tokenization requirements of video VAEs. Simultaneously, we encode camera trajectories as scale-invariant raymap sequence representations, structured to align with the spatiotemporal framework of diffusion transformers (DiTs). By dynamically integrating cross-task and cross-modal conditioning signals during training, our framework enables synergistic knowledge transfer across heterogeneous inputs, facilitating joint optimization for multi-task generative modeling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.606, + 0.909, + 0.817 + ], + "angle": 0, + "content": "In summary, this work introduces AETHER, a unified world model that integrates reconstruction, prediction, and planning through multi-task learning on synthetic 4D data. We propose a robust automatic data annotation pipeline to obtain accurate 4D geometry knowledge. By combining geometric reasoning with generative priors, our framework achieves robust zero-shot transfer to real-world tasks, demonstrating accuracy comparable to SOTA reconstruction models while enabling actionable planning capabilities. The results underscore the value of synergistic 4D modeling for advancing spatial intelligence in AI systems. We hope that AETHER will serve as an effective starter framework for the community to explore post-training world models with scalable synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.864, + 0.848 + ], + "angle": 0, + "content": "2. 4D Synthetic Data Annotation Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "For the synthetic data source, we follow DA-V [74] and TheMatrix [17] to collect large-scale synthetic data with high-quality video depth data. With high-resolution RGB videos" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8536" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.865, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.29, + 0.816, + 0.303 + ], + "angle": 0, + "content": "Figure 2. Some visualization results of data annotated through our pipeline. Better viewed when zoomed in." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.308, + 0.471, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.447, + 0.462, + 0.46 + ], + "angle": 0, + "content": "Figure 3. Our robust automatic camera annotation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.626 + ], + "angle": 0, + "content": "and corresponding per-frame depth maps collected, we built a robust and fully automatic camera annotation pipeline for both camera extrinsics and intrinsics. As illustrated in Fig. 3, the pipeline has four stages: (1) object-level dynamic masking, (2) reconstruction-friendly video slicing, (3) coarse camera localization and calibration, and (4) tracking-based camera refinement with bundle adjustment. We present several visualizations of our annotated data in Fig. 2, ranging from indoor to outdoor scenes, and from static to dynamic scenarios, demonstrating the robustness and accuracy of our annotation method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Dynamic Masking. Distinguishing between dynamic and static regions is crucial for accurate camera parameters estimation. Here, we utilize semantic categories that are potentially dynamic (e.g., cars, people) to segment dynamic objects. Although this may occasionally misclassify static objects, such as stationary parked cars, as dynamic, we find it more robust than flow-based segmentation methods. Specifically, we use Grounded SAM 2 [50] to ensure the temporal consistency of dynamic masks over long sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Video Slicing. Video slicing plays a critical role in 3D reconstruction by serving two key purposes: First, it eliminates unsuitable video segments (such as scene cuts or motion-blurred frames) that could compromise reconstruction quality. Second, it segments long videos into shorter, temporally coherent clips to enhance processing efficiency. The specific criteria for frame removal are as follows: (1) Insufficient Feature Points: We employ the SIFT [39] feature descriptor to extract keypoints from each frame. Frames exhibiting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.908, + 0.567 + ], + "angle": 0, + "content": "insufficient SIFT keypoints are discarded to ensure robust correspondence estimation. Additionally, frames containing regions with insufficient texture due to low illumination are excluded, as such areas typically exhibit poor feature discriminability and pose challenges for reliable matching. (2) Large Areas of Dynamic Regions: Frames where dynamic regions (obtained from dynamic annotation) dominate over static regions can introduce ambiguity in camera pose estimation. Such frames are filtered out to ensure robust results. (3) Large Motion or Inaccurate Correspondence: Using an off-the-shelf optical flow estimator, RAFT [61], we estimate the magnitude of motion. If these magnitudes exceed a predefined threshold, we truncate the sequence at the current frame, retaining all preceding frames as a valid segment. Similarly, if the ratio of forward-to-backward optical flow errors surpasses a threshold value, we truncate the current frames to ensure temporal coherence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Coarse Camera Estimation. For each video slice, we first use DroidCalib [25] to perform a coarse estimation of the camera parameters, leveraging the depth information from static regions. However, due to the lower input resolution of the DroidCalib model and the limited accuracy of its correspondence estimation, a refinement process is necessary to obtain precise camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Camera Refinement. We begin camera refinement by employing the state-of-the-art tracker, CoTracker3 [33], to capture accurate long-term correspondences across the entire slice. SIFT [39] and SuperPoint [12] feature points are extracted from static regions, and then tracked to form correspondences. Subsequently, bundle adjustment is performed on all frames to minimize the accumulated reprojection error of all correspondences. With access to high-quality dense depth, we apply forward-backward reprojection to estimate and minimize errors in 3D space [8], which improves per-frame camera accuracy while preserving inter-frame geometric consistency. Specifically, we solve the nonlinear optimization problem by Ceres Solver [1], and the Cauchy loss function is applied to measure correspondence residuals, which accounts for the problem's sparsity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.515, + 0.967 + ], + "angle": 0, + "content": "8537" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.087, + 0.868, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.173, + 0.293, + 0.264, + 0.308 + ], + "angle": 0, + "content": "Image latent" + }, + { + "type": "image_footnote", + "bbox": [ + 0.264, + 0.293, + 0.344, + 0.307 + ], + "angle": 0, + "content": "Action latent" + }, + { + "type": "image_footnote", + "bbox": [ + 0.345, + 0.293, + 0.45, + 0.308 + ], + "angle": 0, + "content": "Image prediction" + }, + { + "type": "image_footnote", + "bbox": [ + 0.45, + 0.293, + 0.554, + 0.308 + ], + "angle": 0, + "content": "Action prediction" + }, + { + "type": "image_footnote", + "bbox": [ + 0.554, + 0.293, + 0.657, + 0.308 + ], + "angle": 0, + "content": "Depth prediction" + }, + { + "type": "image_footnote", + "bbox": [ + 0.657, + 0.293, + 0.758, + 0.308 + ], + "angle": 0, + "content": "Zero-padding" + }, + { + "type": "image_footnote", + "bbox": [ + 0.759, + 0.293, + 0.816, + 0.308 + ], + "angle": 0, + "content": "Noise" + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.309, + 0.845, + 0.323 + ], + "angle": 0, + "content": "Figure 4. The overall pipeline of AETHER. With different condition combinations, AETHER can serve different tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.329, + 0.406, + 0.345 + ], + "angle": 0, + "content": "3. AETHER Multi-Task World Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.461 + ], + "angle": 0, + "content": "In this section, we introduce how we post-train a base video diffusion model into a unified multi-task world model AETHER. We use CogVideoX-5b-I2V [77] as our base model. We first give an overview of our framework in Sec. 3.1, then we detail on the input process of depth videos and camera pose trajectories in Sec. 3.2 and Sec. 3.3. Finally, we show how we do model training in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.266, + 0.484 + ], + "angle": 0, + "content": "3.1. Method Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Mainstream video diffusion models [27, 40] typically involve two processes: a forward (noising) process and a reverse (denoising) process. The forward process incrementally adds Gaussian noise, denoted as \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\), to a clean latent sample \\(\\mathbf{z}_0 \\in \\mathbb{R}^{k \\times c \\times h \\times w}\\), where \\(k, c, h, w\\) represent the dimensions of the video latents. Through this process, the clean \\(\\mathbf{z}_0\\) is gradually transformed into a noisy latent \\(\\mathbf{z}_t\\). In the reverse process, a learned denoising model \\(\\epsilon_{\\theta}\\) progressively removes the noise from \\(\\mathbf{z}_t\\) to reconstruct the original latent representation. The denoising model \\(\\epsilon_{\\theta}\\) is conditioned on auxiliary inputs \\(\\mathbf{c}\\) and the diffusion timestep \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.793 + ], + "angle": 0, + "content": "In our method, the target latent \\(\\mathbf{z}_0\\) comprises three modalities: color video latents \\(\\mathbf{z}_{\\mathrm{c0}}\\), depth video latents \\(\\mathbf{z}_{\\mathrm{d0}}\\), and action latents \\(\\mathbf{z}_{\\mathrm{a0}}\\). The model additionally takes two types of conditions as input: color video conditions \\(\\mathbf{c}_{\\mathrm{c}}\\) and action conditions \\(\\mathbf{c}_{\\mathrm{a}}\\). For the action modality, we choose camera pose trajectory as a global action, facilitated by our automated camera pose annotation pipeline described earlier. All latents and conditions are channel-wise concatenated. The training objective of AETHER can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.806, + 0.484, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {\\substack {\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\\\ t \\sim \\mathcal {U} (1, \\mathcal {T}) \\\\ \\mathbf {z} _ {0} = \\mathbf {z} _ {\\mathrm {c} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {d} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {a} _ {0}} \\\\ \\mathbf {c} = \\mathbf {c} _ {\\mathrm {c}} \\otimes \\mathbf {c} _ {\\mathrm {a}}}} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| ^ {2}, \\tag{1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes the channel-wise concatenation operation, \\(\\mathcal{U}(\\cdot)\\) represents a uniform distribution, and \\(\\mathcal{T}\\) denotes the denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.497 + ], + "angle": 0, + "content": "The multi-task objective of AETHER is determined by the specific conditions \\(\\mathbf{c}\\) for different tasks. (1) Reconstruction: \\(\\mathbf{c}_{\\mathrm{c}}\\) represents the input video latents. (2) Video prediction: \\(\\mathbf{c}_{\\mathrm{c}}\\) takes the latent of observation image as the first frame, while other latents are zero-masked. (3) Goal-conditioned visual planning: The first and last latents of \\(\\mathbf{c}_{\\mathrm{c}}\\) correspond to the observation and goal images, respectively, with all intermediate latents zero-padded. For the action condition \\(\\mathbf{c}_{\\mathrm{a}}\\), it is either entirely zero-masked or contains the full target camera pose trajectory in action-free or action-conditioned control cases. Illustrations are show in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.715, + 0.521 + ], + "angle": 0, + "content": "3.2. Depth Videos Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Given a depth video \\(\\mathbf{x}_d\\), we first clip the depth values to a predefined range \\([d_{\\min}, d_{\\max}]\\). Next, we apply a square root transformation and subsequently compute the reciprocal to convert the depth values into disparity, as described in [57]. Each disparity video clip is then normalized in a scale-invariant manner. Subsequently, the normalized disparity values are linearly mapped from \\([0,1]\\) to \\([-1,1]\\). To meet the input requirements of the VAE, the single-channel disparity map is replicated across three channels, as done in prior works [34, 74]. The final depth latent is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.686, + 0.907, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {\\text {d i s p}} = \\frac {1}{\\sqrt {\\operatorname {c l i p} \\left(\\mathbf {x} _ {d} , d _ {\\min } , d _ {\\max }\\right)}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.722, + 0.907, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {\\text {d i s p}} = \\frac {\\mathbf {x} _ {\\text {d i s p}}}{\\operatorname* {m a x} \\left(\\mathbf {x} _ {\\text {d i s p}}\\right)} \\times 2 - 1, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.755, + 0.907, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {d} = \\mathcal {E} \\left(\\hat {\\mathbf {x}} _ {\\text {d i s p}} \\otimes \\mathbf {1} _ {3}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) denotes the 3D VAE, and \\(\\otimes \\mathbf{1}_3\\) represents the channel-wise replication of 3 times. The above operations are designed to be compatible with the pretrained 3D VAE model, ensuring minimal reconstruction error." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.771, + 0.865 + ], + "angle": 0, + "content": "3.3. Camera Trajectories Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We transform camera parameters into raymap videos [7] so that video diffusion can process them compatibly. Specifi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8538" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.136 + ], + "angle": 0, + "content": "cally, given the intrinsic matrix \\(\\mathbf{K} \\in \\mathbb{R}^{T \\times 3 \\times 3}\\) and the extrinsic matrix \\(\\mathbf{E} \\in \\mathbb{R}^{T \\times 4 \\times 4}\\), the transformation process can be described as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Translation Scaling and Normalization. The translation component of the camera pose (inverse of extrinsic matrix), \\(\\mathbf{t} \\in \\mathbb{R}^3\\), is first scaled by a constant factor \\(s_{\\mathrm{ray}}\\) and normalized using the maximum disparity value \\(d_{\\mathrm{max}}\\). To suppress large values, we then pass it through a signed \\(\\log(1 + \\cdot)\\) transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.234, + 0.483, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} ^ {\\prime} = \\frac {\\mathbf {t}}{\\operatorname* {m a x} \\left(x _ {\\text {d i s p}}\\right)} \\cdot s _ {\\text {r a y}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.268, + 0.484, + 0.285 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} _ {\\log} = \\operatorname {s i g n} \\left(\\mathbf {t} ^ {\\prime}\\right) \\cdot \\log \\left(1 + \\left| \\mathbf {t} ^ {\\prime} \\right|\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.361, + 0.308 + ], + "angle": 0, + "content": "where \\( s_{\\mathrm{ray}} \\) is a predefined scaling factor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.486, + 0.445 + ], + "angle": 0, + "content": "Raymap Construction. Using the intrinsic matrix \\(\\mathbf{K}\\), we compute the camera ray directions \\(\\mathbf{r}_{\\mathrm{d}}\\) in homogeneous coordinates for each pixel. Note that we do not unit normalize it but let it have a unit value along \\(z\\) axis. The ray origins \\(\\mathbf{r}_{\\mathrm{o}}\\) are set to the translation \\(\\mathbf{t}_{\\log}\\), replicated across the spatial dimensions. The raymap in the world coordinate system is obtained by transforming the ray directions \\(\\mathbf{r}_{\\mathrm{d}}\\) using the extrinsic matrix \\(\\mathbf{E}\\). The final raymap \\(\\mathbf{r}\\) consists of 6 channels: 3 for the ray directions \\(\\mathbf{r}_{\\mathrm{d}}\\) and 3 for the ray origins \\(\\mathbf{r}_{\\mathrm{o}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.485, + 0.55 + ], + "angle": 0, + "content": "Resolution Downsampling. To align the raymap with the latent feature dimensions from the VAE, we perform adjustments both spatially and temporally. Spatially, the raymap is downsampled by a factor of 8 using bilinear interpolation. Temporally, every consecutive group of 4 frames is concatenated along the channel dimension. The resulting rearranged tensor is denoted as \\(\\mathbf{z}_a\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.485, + 0.597 + ], + "angle": 0, + "content": "Converting raymap back to camera matrix. Given generated raymap sequences rearranged by the time axis \\(\\hat{\\mathbf{r}}\\in\\) \\(\\mathbb{R}^{T\\times 6\\times h\\times w} = [\\hat{\\mathbf{r}}_{\\mathrm{d}},\\hat{\\mathbf{r}}_{\\mathrm{o}}]\\) , we first recover the ray origins by:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.604, + 0.484, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {r} _ {\\mathrm {o}}} ^ {\\prime} = \\frac {1}{s _ {\\text {r a y}}} \\cdot \\operatorname {s i g n} \\left(\\hat {\\mathbf {r} _ {\\mathrm {o}}}\\right) \\cdot \\left(\\exp \\left(\\left| \\hat {\\mathbf {r} _ {\\mathrm {o}}} \\right|\\right) - 1\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Then, we can recover both the intrinsic and extrinsic through Alg. 1 in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.248, + 0.699 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.485, + 0.796 + ], + "angle": 0, + "content": "We initialize AETHER with pre-trained CogVideo-5b-I2V [77] weights, excluding the additional input and output projection layer channels for depth and raymap action trajectories, which are initialized to zero. Since text prompt conditions are not used, an empty text embedding is provided during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.485, + 0.871 + ], + "angle": 0, + "content": "As the dataset we use contains video clips with variable lengths and frames per second (FPS), we randomly select \\( T \\in \\{17, 25, 33, 41\\} \\) frames, and the FPS is randomly sampled from \\( \\{8, 10, 12, 15, 24\\} \\). The RoPE [59] coefficients are linearly interpolated to align with them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "During training, conditional inputs are randomly masked to generalize across tasks. For \\(\\mathbf{c}_c\\), masking probabilities are:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.212 + ], + "angle": 0, + "content": "\\(30\\%\\) for both observation and goal images (visual planning tasks), \\(40\\%\\) for observation images only (video prediction), \\(28\\%\\) for full-color video latents (4D reconstruction), and \\(2\\%\\) for masking all of \\(\\mathbf{c}_c\\). For \\(\\mathbf{c}_a\\), trajectory latents are either kept or fully masked with equal probability (supporting action-free or action-conditioned tasks with raymap conditions). This strategy enables the model to adapt to diverse tasks and input condition settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.909, + 0.409 + ], + "angle": 0, + "content": "Our training process consists of two stages. In the first stage, we adopt the loss function of a standard latent diffusion model, which minimizes the mean squared error (MSE) in the latent space. In the second stage, we refine the generated outputs by decoding them into the image space. Specifically, we introduce three additional loss terms: a Multi-Scale Structure Similarity (MS-SSIM) loss [67] for color video, a scale- and shift-invariant loss [49] for depth videos, and a scale- and shift-invariant pointmap loss [66] for pointmaps projected from the generated depths and raymaps. Further details on the stage 2 loss functions are provided in the supplementary material. Notably, the second stage takes about \\(\\frac{1}{4}\\) of the training steps used in the first stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.909, + 0.56 + ], + "angle": 0, + "content": "We employ a hybrid training strategy combining Fully Sharded Data Parallel (FSDP) [87] with Zero-2 optimization within compute nodes and Distributed Data Parallel (DDP) across nodes. Since depth videos require online normalization, the VAE encoder is also run online during training and operates under DDP. Our implementation processes a local batch size of 4 per GPU, resulting in an effective batch size of 320 samples across 80 A100-80GB GPUs. Training is conducted over two weeks using the AdamW [38] optimizer with a OneCycle [56] learning rate scheduler." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.779, + 0.591 + ], + "angle": 0, + "content": "4. Reconstruction Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.908, + 0.689 + ], + "angle": 0, + "content": "In this section, we demonstrate that AETHER can achieve zero-shot reconstruction metrics comparable to or even better than SOTA reconstruction methods. We mainly consider two zero-shot reconstruction tasks: video depth estimation and camera pose estimation. Note that we only denoise for 4 steps for reconstruction tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.816, + 0.714 + ], + "angle": 0, + "content": "4.1. Zero-Shot Video Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Video depth estimation is evaluated based on two key aspects: per-frame depth quality and inter-frame depth consistency. These evaluations are performed by aligning the predicted depth maps with the ground truth using a per-sequence scale. We use absolute relative error (Abs Rel) and \\(\\delta < 1.25\\) (percentage of predicted depths within a 1.25-factor of true depth) as metrics. For implementation, we adopt the settings outlined in CUT3R [65]. Our baselines include both reconstruction-based methods—such as DUSt3R [66], MASt3R [37], MonST3R [82], Spann3R [63], and CUT3R [65]—and diffusion-based depth estimators, including ChronoDepth [55], DepthCrafter [29]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.955, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8539" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.089, + 0.764, + 0.103 + ], + "angle": 0, + "content": "Table 1. Video depth Evaluation. Methods requiring global alignment are marked \"GA\"." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.104, + 0.907, + 0.296 + ], + "angle": 0, + "content": "
MethodSintel [6]BONN [44]KITTI [21]
Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑
Reconstruction Methods. Alignment: per-sequence scale
DUSt3R-GA [66]0.65645.20.15583.30.14481.3
MASt3R-GA [37]0.64143.90.25270.10.18374.5
MonST3R-GA [82]0.37855.80.06796.30.16874.4
Spann3R [63]0.62242.60.14481.30.19873.7
CUT3R [65]0.42147.90.07893.70.11888.1
AETHER (Ours)0.32450.20.27359.40.05697.8
Diffusion-Based Methods. Alignment: per-sequence scale&shift
ChronoDepth [55]0.42938.30.31851.80.25254.3
DepthCrafter [29]0.59055.50.25356.30.12486.5
DA-V [74]1.25243.70.45731.10.09493.0
AETHER (Ours)0.31460.40.30860.20.05497.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.346, + 0.301, + 0.651, + 0.313 + ], + "angle": 0, + "content": "Table 2. Evaluation on Camera Pose Estimation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.314, + 0.912, + 0.508 + ], + "angle": 0, + "content": "
MethodSintel [6]TUM-dynamics [58]ScanNet [10]
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Optimization-based Methods
Particle-SfM [86]0.1290.0310.535---0.1360.0230.836
Robust-CVD [36]0.3600.1543.4430.1530.0263.5280.2270.0647.374
CasualSAM [85]0.1410.0350.6150.0710.0101.7120.1580.0341.618
DUSt3R-GA [66]0.4170.2505.7960.0830.0173.5670.0810.0280.784
MASt3R-GA [37]0.1850.0601.4960.0380.0120.4480.0780.0200.475
MonST3R-GA [82]0.1110.0440.8960.0980.0190.9350.0770.0180.529
Feed-forward Methods
DUSt3R [66]0.2900.1327.8690.1400.1063.2860.2460.1088.210
Spann3R [63]0.3290.1104.4710.0560.0210.5910.0960.0230.661
CUT3R [65]0.2130.0660.6210.0460.0150.4730.0990.0220.600
AETHER (Ours)0.1890.0540.6940.0920.0121.1060.1760.0281.204
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.519, + 0.486, + 0.732 + ], + "angle": 0, + "content": "and DepthAnyVideo (DA-V) [74]. It is important to note that when comparing with diffusion-based depth estimators, we apply scale and shift alignment to the ground truth, as most of these methods are not inherently scale-invariant. All videos are resized with original aspect ratios kept to make the short side align with our model's input size. For videos that exceed the maximum forward processing spatial or temporal size of our model, we employ a sliding window strategy with a stride size of 8. In regions of overlap between windows, we first estimate a relative scale by calculating the average of element-wise division. This relative scale is then used to adjust the latter window's depth predictions. Finally, a linspace-weighted average is applied to the overlapping areas, following approaches similar to prior methods [29, 80]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Results and Analysis. Table 1 summarizes the video depth estimation results across Sintel [6], BONN [44], and KITTI [21] datasets. For reconstruction-based methods, AETHER outperforms or is comparable with prior approaches. On Sintel, AETHER achieves the lowest Abs Rel (0.324), surpassing MonST3R-GA (0.378), and competitive \\(\\delta < 1.25\\) (50.2). On KITTI, AETHER sets a new benchmark with Abs Rel of 0.056 and \\(\\delta < 1.25\\) of 97.8, outperforming the previous SOTA CUT3R (Abs Rel: 0.118, \\(\\delta < 1.25\\): 88.1). Among diffusion-based methods, AETHER shows consistent superiority. It achieves the best performance on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Sintel (Abs Rel: 0.314, \\(\\delta < 1.25\\): 60.4) and KITTI (Abs Rel: 0.054, \\(\\delta < 1.25\\): 97.7), significantly outperforming ChronoDepth [55], DepthCrafter [29], and DA-V [74]. On BONN, AETHER achieves the highest \\(\\delta < 1.25\\) (60.2) with competitive Abs Rel (0.308)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.822, + 0.621 + ], + "angle": 0, + "content": "4.2. Zero-Shot Camera Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. Following MonST3R [82] and CUT3R [65], we evaluate camera pose estimation accuracy on the Sintel [6], TUM Dynamics [58], and ScanNet [10] datasets. Notably, both Sintel and TUM Dynamics contain highly dynamic objects, presenting significant challenges for traditional Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM) systems. We report Absolute Translation Error (ATE), Relative Translation Error (RPE Trans), and Relative Rotation Error (RPE Rot) after Sim(3) alignment with the ground truth, following the methodology in [65]. The implementation settings are consistent with those used in CUT3R [65]. All videos are resized with original aspect ratios kept and then center cropped to align with our model's input size. For long videos exceeding our model's maximum temporal forward processing length, a sliding window strategy with a stride size of 32 is employed. In overlapping regions between windows, camera poses are aligned following prior methods [64]. Transla" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "tion alignment is performed using linear interpolation, while quaternion rotations are interpolated with spherical linear interpolation. Additionally, we observed that the generated camera trajectories exhibit noise, likely due to the limited number of denoising steps. To mitigate this, we apply a simple Kalman filter [71] to smooth the trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.486, + 0.288 + ], + "angle": 0, + "content": "Results and Analysis. Table 2 shows the evaluation results. Among feed-forward methods, AETHER achieves the best ATE (0.189) and RPE Trans (0.054) on Sintel [6], while remaining competitive in RPE Rot (0.694) compared to CUT3R (0.621). On TUM Dynamics [58], AETHER achieves the best RPE Trans (0.012). For other metrics, AETHER is also comparable with other specialist models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.302, + 0.438, + 0.318 + ], + "angle": 0, + "content": "5. Generation and Planning Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.388 + ], + "angle": 0, + "content": "In this section, we first show video prediction, with or without action conditioning, quantitatively or qualitatively, in Sec. 5.1. We then show visual planning abilities in Sec. 5.2. More visualizations are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.396, + 0.258, + 0.411 + ], + "angle": 0, + "content": "5.1. Video Prediction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.486, + 0.81 + ], + "angle": 0, + "content": "Implementation Details. We use CogVideoX-5b-I2V [77] as our baseline. To ensure a fair comparison, we construct a validation dataset comprising two subsets: in-domain and out-domain data. The in-domain subset includes novel, unseen scenes from the same synthetic environments as the training dataset, while the out-domain subset consists of data from entirely new synthetic environments. Both models are provided with the first frame as the observation image. For action-free prediction, since CogVideoX depends heavily on text prompts, we use GPT-4o [31] to generate image descriptions and predictions of future scenes as prompts for CogVideoX. In contrast, AETHER is evaluated using empty text prompts. For action-conditioned prediction, we also labeled camera trajectories in the validation dataset and generated corresponding raymap sequences as action conditions for AETHER. For the baseline, in addition to the prompts used for action-free prediction, we use GPT-4o [31] to generate detailed descriptions of object and camera movements, enabling the baseline to use language as action conditions. We use the default classifier-free guidance value of 6 on text prompts for CogVideoX and a value of 3 on the observation image for AETHER. No classifier-free guidance is applied to action conditions to ensure fairness. Evaluation metrics follow VBenchmark [30], a standard benchmark for video generation, with additional details on prompts and evaluation metrics provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Image-to-Video Prediction. We first evaluate image-to-video prediction without action conditions. The results, presented in Tab. 3, show that AETHER consistently outperforms the baseline on both in-domain and out-domain validation sets. Notably, AETHER demonstrates a larger performance improvement on out-domain data, which can likely be at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tributed to the baseline model's pre-training data containing domains similar to the in-domain dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.909, + 0.289 + ], + "angle": 0, + "content": "Action-Conditioned Video Prediction. To assess the effectiveness of our post-training in improving action control and action-following capabilities, we conduct action-conditioned video prediction experiments. The results, shown in Tab. 4, indicate that AETHER consistently outperforms the baseline in both in-domain and out-domain settings. Notably, CogVideoX tends to generate static scenes with high visual and aesthetic quality, while AETHER accurately follows the action conditions, producing highly dynamic scenes. These results validate the effectiveness of our framework and the advantages of using camera trajectories as action conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.673, + 0.314 + ], + "angle": 0, + "content": "5.2. Visual Planning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.909, + 0.529 + ], + "angle": 0, + "content": "Implementation Details. We evaluate the action-conditioned navigation capability of AETHER on our validation set. To demonstrate the effectiveness of our multi-task objective, particularly the incorporation of the reconstruction objective, we also post-train an ablation model without the video depth objective, denoted as AETHER-no-depth. Given the observation image, goal image, and camera trajectory, the resulting video should be highly determined. Thus, we report pixel-wise reconstruction metrics, including PSNR, SSIM [68], MS-SSIM [67], and LPIPS [83], for action-conditioned navigation. For the action-free case, which represents a visual path navigation task, we also report the VBench metrics. We do not use any classifier-free guidance on both tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.909, + 0.606 + ], + "angle": 0, + "content": "Action-Conditioned Navigation. The quantitative results for action-conditioned navigation are presented in Tab. 5. Aether consistently outperforms the ablation model, demonstrating the significant benefits of incorporating the reconstruction objective into generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.909, + 0.728 + ], + "angle": 0, + "content": "Visual Path Planning. In the absence of action conditions, this task evaluates the model's ability to function as a \"world model as an agent,\" requiring it to plan a path from the observation image to the goal image. The results, shown in Tab. 6, indicate that the reconstruction objective substantially improves the model's visual path planning capability. Additionally, qualitative visualizations on completely in-the-wild data are provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.657, + 0.755 + ], + "angle": 0, + "content": "6. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "World Models. World models have emerged as a critical framework in artificial intelligence, enabling agents to simulate, understand, and predict environmental dynamics. Early work [23] introduced latent representations and recurrent neural networks for decision-making. Recent advancements include Cat3D [20] for 3D scene generation, Cat4D [72] for dynamic 4D environments, and Genie 2 [46], a large-scale model for interactive 3D worlds. Motion Prompting [22] further enables precise video generation control. These ad" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.514, + 0.968 + ], + "angle": 0, + "content": "8541" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 3. VBenchmark [30] Metrics of Video Prediction without Action Conditions. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each group, the better performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.117, + 0.905, + 0.168 + ], + "angle": 0, + "content": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX89.36/84.61/87.7792.72/91.43/92.2998.24/96.93/97.8188.75/95.00/90.8354.49/53.58/54.1855.38/52.29/54.3579.01/77.52/78.51
AETHER91.50/87.55/90.1894.29/93.62/94.0798.54/98.19/98.4296.25/100.00/97.5054.36/52.58/53.7755.08/54.88/55.0180.34/79.42/80.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.172, + 0.907, + 0.2 + ], + "angle": 0, + "content": "Table 4. VBenchmark [30] Metrics of Action-Conditioned Video Prediction. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each metric group, the better performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.2, + 0.905, + 0.25 + ], + "angle": 0, + "content": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX91.56/88.23/90.5192.98/92.29/92.7798.44/97.81/98.2483.87/93.02/86.7656.19/57.43/56.5856.48/61.60/58.1079.56/80.70/79.92
AETHER90.73/93.27/91.5493.61/95.03/94.0698.53/98.62/98.56100.00/83.72/94.8555.04/56.50/55.5053.89/63.23/56.8480.33/81.55/80.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.254, + 0.907, + 0.282 + ], + "angle": 0, + "content": "Table 5. Pixel-wise Metrics of Action-Conditioned Navigation. Comparison of performance between AETHER-no-depth and AETHER on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.283, + 0.905, + 0.334 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓
AETHER-no-depth19.13/18.67/18.970.5630/0.4830/0.53530.5467/0.5204/0.53760.3116/0.2995/0.3074
AETHER19.87/19.37/19.700.5803/0.5058/0.55450.5830/0.5627/0.57600.2691/0.2599/0.2659
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.337, + 0.907, + 0.365 + ], + "angle": 0, + "content": "Table 6. Quantitative Results of Action-Free Visual Path Planning. Comparison of performance between Aether and Aether-no-depth on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.366, + 0.918, + 0.414 + ], + "angle": 0, + "content": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
Aether-no-depth88.68/89.61/88.6193.62/93.92/93.6698.37/98.31/98.3297.06/91.67/96.1554.12/56.26/54.7851.77/58.46/54.2979.11/80.43/79.59
Aether (Ours)89.69/91.61/90.3693.88/94.58/94.1398.50/98.40/98.4697.06/91.67/95.1955.83/56.87/56.1954.71/61.13/56.9380.21/81.53/80.67
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.483, + 0.476 + ], + "angle": 0, + "content": "vancements demonstrate the evolution of world models toward dynamic, interactive, and controllable applications in robotics, gaming, and simulation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Reconstruction. Reconstruction has been a long-standing topic in computer vision, with notable progress in both traditional and learning-based methods. Classical approaches, such as Structure-from-Motion (SfM) [9, 26, 45, 53] and Multi-View Stereo (MVS) [19, 54], rely on multi-view geometry for feature matching, pose estimation, and dense point cloud generation, demonstrating robust performance in controlled settings. Deep learning has introduced powerful alternatives, tackling sub-tasks like feature matching [14, 52], point tracking [13, 64], triangulation [42], and MVS [78, 81]. End-to-end methods now directly predict point maps [37, 66] or depth maps from images [4, 76], often incorporating camera parameters [70]. Recently, diffusion models have achieved breakthroughs in image and video generation [27, 35, 43, 73, 77], inspiring novel 3D reconstruction approaches that leverage rich 2D priors [18, 29, 34, 41, 74, 75, 88, 90]. These methods demonstrate the potential of integrating diffusion-based 2D knowledge into 3D modeling." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Video Generation. Video generation has evolved from foundational techniques like DDPM [27, 43] to modern frameworks leveraging diffusion-based techniques. Advances such as latent diffusion [51] and diffusion transformers [47] have improved generation quality, while models like Sora [5] and Stable Video Diffusion (SVD) [3] emphasize temporal consistency. Open-source models, including LTX Video [24], CogVideoX [77], and Hunyuan Video [35], offer increased flexibility, and techniques like multi-scale architectures (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Pyramid Flow [32]) enhance motion dynamics. These advancements highlight rapid progress, with ongoing efforts to improve scalability and temporal stability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.491, + 0.771, + 0.507 + ], + "angle": 0, + "content": "7. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.516, + 0.909, + 0.697 + ], + "angle": 0, + "content": "In this work, we introduce AETHER, a geometry-aware multitask world model that reconstructs 4D dynamic videos, predicts future frames conditioned on observation images and actions, and performs visual planning based on observation and goal images. We propose an automatic 4D synthetic data labeling pipeline, enabling AETHER to train on synthetic data and generalize to unseen real-world data in a zero-shot manner. Post-trained on the CogVideoX base model, AETHER achieves state-of-the-art or competitive reconstruction performance and outperforms baselines in generation and planning tasks, demonstrating the value of incorporating reconstruction objectives into world modeling." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.697, + 0.909, + 0.849 + ], + "angle": 0, + "content": "However, limitations remain. Camera pose estimation is less accurate, likely due to incompatibilities between raymap representation and prior video diffusion models. Indoor scene reconstruction also lags behind outdoor performance, likely due to the predominance of outdoor training data. Additionally, predictions without language prompts often fail in highly dynamic scenes. Future work can address these by exploring novel action representations, co-training with real-world data, and retaining the language prompting capabilities of the base model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8542" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.193 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.202, + 0.484, + 0.229 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Keir Mierle, et al. Ceres solver: Tutorial & reference. Google Inc, 2(72):8, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.231, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[2] Eloi Alonso, Adam Jolley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37:58757-58791, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.303, + 0.484, + 0.372 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[4] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.486 + ], + "angle": 0, + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.57 + ], + "angle": 0, + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[7] Junyi Chen, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Where am i and what will i see: An auto-regressive model for spatial localization and view prediction, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[8] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.742 + ], + "angle": 0, + "content": "[9] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Decart, Julian Quevedo, Quinn McIntyre, Spruce Campbell, Xinlei Chen, and Robert Wachen. Oasis: A universe in a Transformer. 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.202, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition workshops, pages 224-236, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[13] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[15] Tom Erez, Yuval Tassa, and Emanuel Todorov. Infinite-horizon model predictive control for periodic tasks with contacts. Robotics: Science and systems VII, 73, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[16] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[17] Ruili Feng, Han Zhang, Zhantao Yang, Jie Xiao, Zhilei Shu, Zhiheng Liu, Andy Zheng, Yukun Huang, Yu Liu, and Hongyang Zhang. The matrix: Infinite-horizon world generation with real-time moving control. arXiv preprint arXiv:2412.03568, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[18] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[19] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[20] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.907, + 0.649 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[22] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, Chen Sun, Oliver Wang, Andrew Owens, and Deqing Sun. Motion prompting: Controlling video generation with motion trajectories, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[23] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[24] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.516, + 0.967 + ], + "angle": 0, + "content": "8543" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[26] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.178 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.237, + 0.484, + 0.292 + ], + "angle": 0, + "content": "[29] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.38, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[31] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.437, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[32] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.508, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[33] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.579, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[34] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Junkun Yuan, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yanxin Long, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. Hunyuanvideo: A systematic framework for large video generative models, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[37] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.908, + 0.22 + ], + "angle": 0, + "content": "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.908, + 0.335 + ], + "angle": 0, + "content": "[41] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.908, + 0.405 + ], + "angle": 0, + "content": "[42] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.408, + 0.908, + 0.462 + ], + "angle": 0, + "content": "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.465, + 0.908, + 0.534 + ], + "angle": 0, + "content": "[44] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.536, + 0.908, + 0.589 + ], + "angle": 0, + "content": "[45] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, pages 58-77. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.593, + 0.908, + 0.731 + ], + "angle": 0, + "content": "[46] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Roktaschel. Genie 2: A large-scale foundation world model. 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.908, + 0.787 + ], + "angle": 0, + "content": "[47] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[48] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.515, + 0.967 + ], + "angle": 0, + "content": "8544" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.486, + 0.313 + ], + "angle": 0, + "content": "Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Eliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[50] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.459, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.53, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[52] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.484, + 0.742 + ], + "angle": 0, + "content": "[54] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.484, + 0.813 + ], + "angle": 0, + "content": "[55] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Vitor Guizilini, Yue Wang, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv preprint arXiv:2406.01493, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[56] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[57] Ziyang Song, Zerong Wang, Bo Li, Hao Zhang, Ruijie Zhu, Li Liu, Peng-Tao Jiang, and Tianzhu Zhang. Depthmaster:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Taming diffusion models for monocular depth estimation. arXiv preprint arXiv:2501.02576, 2025.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.188 + ], + "angle": 0, + "content": "[58] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[59] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[60] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.907, + 0.342 + ], + "angle": 0, + "content": "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.908, + 0.523 + ], + "angle": 0, + "content": "[62] Veo-Team,.; Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathewson, Kurtis David, Matthieu Kim Lorraine, Marc van Zee, Medhini Narasimhan, Miaosen Wang, Mohammad Babaeizadeh, Nelly Papalampidi, Nick Pezzotti, Nilpa Jha, Parker Barnes, Pieter-Jan Kindermans, Rachel Hornung, Ruben Villegas, Ryan Poplin, Salah Zaiem, Sander Dieleman, Sayna Ebrahimi, Scott Wisdom, Serena Zhang, Shlomi Fruchter, Signe Nørly, Weizhe Hua, Xinchen Yan, Yuqing Du, and Yutian Chen. Veo 2. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.564 + ], + "angle": 0, + "content": "[63] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[64] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.691 + ], + "angle": 0, + "content": "[65] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[66] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[67] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[68] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Mo" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.515, + 0.967 + ], + "angle": 0, + "content": "8545" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tionctrl: A unified and flexible motion controller for video generation. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[70] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xiangyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 230-247. Springer, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[71] Greg Welch, Gary Bishop, et al. An introduction to the kalman filter. 1995. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[72] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[73] Wanghan Xu, Xiaoyu Yue, Zidong Wang, Yao Teng, Wenlong Zhang, Xihui Liu, Luping Zhou, Wanli Ouyang, and Lei Bai. Exploring representation-aligned latent space for better generation. arXiv preprint arXiv:2502.00359, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[74] Honghui Yang, Di Huang, Wei Yin, Chunhua Shen, Haifeng Liu, Xiaofei He, Binbin Lin, Wanli Ouyang, and Tong He. Depth any video with scalable synthetic data. arXiv preprint arXiv:2410.10815, 2024. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[75] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, et al. Unipad: A universal pre-training paradigm for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15238-15250, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[77] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[78] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.705 + ], + "angle": 0, + "content": "[79] Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Creating new games with generative interactive videos, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[80] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[81] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1):199-214, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[82] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[84] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[85] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.907, + 0.344 + ], + "angle": 0, + "content": "[86] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.907, + 0.415 + ], + "angle": 0, + "content": "[87] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.907, + 0.487 + ], + "angle": 0, + "content": "[88] Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Hengshuang Zhao, Chunhua Shen, Yu Qiao, Tong He, et al. Ponderv2: Pave the way for 3d foundation model with a universal pre-training paradigm. arXiv preprint arXiv:2310.08586, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.487, + 0.907, + 0.556 + ], + "angle": 0, + "content": "[89] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.907, + 0.613 + ], + "angle": 0, + "content": "[90] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.956, + 0.517, + 0.967 + ], + "angle": 0, + "content": "8546" + } + ] +] \ No newline at end of file diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_origin.pdf b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b636031031e9e77e2658761e63f148eb884c6d2 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/d408fc86-f2f5-4340-bdc7-9915c8ed0e52_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f3242ed32db492a80d8eb980bfbee2cd1ea530187f41f56d150605110c59fe8 +size 7220326 diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/full.md b/2025/Aether_ Geometric-Aware Unified World Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..589a7ec7d11bf39b88f9274792cc14b6d5c564d2 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/full.md @@ -0,0 +1,330 @@ +# AETHER: Geometric-Aware Unified World Modeling + +Haoyi Zhu* Yifan Wang* Jianjun Zhou* Wenzheng Chang* Yang Zhou* Zizun Li* Junyi Chen* Chunhua Shen Jiangmiao Pang Tong He USTC Shanghai AI Lab SII SJTU ZJU FDU *Equal Contribution † Corresponding Author https://aether-world.github.io/ + +![](images/94cd943bbc5d4b733ea437470020d9295c18436b009cb913494dce09c0b22851.jpg) +Figure 1. An overview of AETHER, trained entirely on synthetic data. The figure highlights its three key capabilities: 4D reconstruction, action-conditioned 4D prediction, and visual planning, all demonstrated on unseen real-world data. The 4D reconstruction examples are derived from MovieGen [48] and Veo 2 [62] generated videos, while the action-conditioned prediction uses an observation image from a university classroom. The visual planning example utilizes observation and goal images from an office building. Better viewed when zoomed in. Additional visualizations can be found on our website. + +# Abstract + +The integration of geometric reconstruction and generative modeling remains a critical challenge in developing AI systems capable of human-like spatial reasoning. This paper proposes AETHER, a unified framework that enables geometry-aware reasoning in world models by jointly optimizing three core capabilities: (1) 4D dynamic reconstruction, (2) action-conditioned video prediction, and (3) goal-conditioned visual planning. Through task-interleaved feature learning, AETHER achieves synergistic knowledge sharing across reconstruction, prediction, and planning objectives. Building upon video generation models, our framework demonstrates zero-shot synthetic-to-real generalization despite never observing real-world data during training. Furthermore, our approach achieves zero-shot generalization in both action following and reconstruction tasks, thanks to its intrinsic geometric modeling. Notably, even without real-world data, its reconstruction performance is comparable with or even better than that of domain-specific models. Additionally, AETHER employs camera trajectories as geometry-informed action spaces, enabling effective action-conditioned prediction and visual planning. We hope our work inspires the community to explore new frontiers in physically-reasonable world modeling and its applications. + +# 1. Introduction + +"Prediction is not just one of the things your brain does. It is the primary function of the neocortex." + +Jeff Hawkins, On Intelligence (2004) + +The development of visual intelligence systems capable of comprehending and forecasting the physical world remains a cornerstone of AI research. World models have emerged as a foundational paradigm for building autonomous systems that not only perceive but also anticipate environmental dynamics to make reasonable actions. At their core, three capabilities stand out: First, perception equips the system with the ability to capture the intricate four-dimensional (4D) changes—integrating spatial and temporal information—that are essential for understanding the physical world [37, 63, 65, 66, 82, 86]. This continuous sensing of dynamic cues enables a geometric representation of the environment. Second, prediction leverages this perceptual information to forecast how the environment will evolve under specific actions, thereby providing a foresight into future states [3, 24, 28, 32, 35, 60, 77]. Finally, planning uses these predictive insights to determine the optimal sequence of actions required to achieve a given goal. Together, these three aspects empower world models to not only represent the current state of the environment but also to anticipate and navigate its future dynamics effectively. + +Motivated by these principles, we introduce AETHER, a unified framework that, for the first time, bridges reconstruction, prediction, and planning, as shown in Fig. 1. AETHER leverages pre-trained video generation models [28, 77] and is further refined via post-training with synthetic 4D data. Although multiple action modalities exist, ranging from keyboard inputs [2, 11, 15, 46, 79] to human or robotic motions [16, 84, 89, 90] and point flows [22, 69], we choose camera pose trajectories as our global action representation. This choice is particularly effective for ego-view tasks: in navigation, camera trajectories directly correspond to the navigation paths, while in robotic manipulation, the movement of an in-hand camera captures the 6D motion of the end effector. To address the scarcity of 4D data, we utilize RGB-D synthetic video data and propose a robust camera pose annotation pipeline to reconstruct full 4D dynamics. + +Through a simple training strategy that randomly combines input and output modalities, our method transforms the base video generation model into a unified, multi-task world model with three key capabilities: (1) Depth and camera pose estimation from full video sequences; (2) Video prediction conditioned on an initial observation—with the option to incorporate a camera trajectory action; and (3) Goal-conditioned visual planning based on observation-goal image pairs. We transform depth videos into scale-invariant normalized disparity representations to meet the tokenization requirements of video VAEs. Simultaneously, we encode camera trajectories as scale-invariant raymap sequence representations, structured to align with the spatiotemporal framework of diffusion transformers (DiTs). By dynamically integrating cross-task and cross-modal conditioning signals during training, our framework enables synergistic knowledge transfer across heterogeneous inputs, facilitating joint optimization for multi-task generative modeling. + +In summary, this work introduces AETHER, a unified world model that integrates reconstruction, prediction, and planning through multi-task learning on synthetic 4D data. We propose a robust automatic data annotation pipeline to obtain accurate 4D geometry knowledge. By combining geometric reasoning with generative priors, our framework achieves robust zero-shot transfer to real-world tasks, demonstrating accuracy comparable to SOTA reconstruction models while enabling actionable planning capabilities. The results underscore the value of synergistic 4D modeling for advancing spatial intelligence in AI systems. We hope that AETHER will serve as an effective starter framework for the community to explore post-training world models with scalable synthetic data. + +# 2. 4D Synthetic Data Annotation Pipeline + +For the synthetic data source, we follow DA-V [74] and TheMatrix [17] to collect large-scale synthetic data with high-quality video depth data. With high-resolution RGB videos + +![](images/7efdfaea582c4f64a08371bfce5c7e861f975166dbefa05849b7a7aae90a8ae8.jpg) +Figure 2. Some visualization results of data annotated through our pipeline. Better viewed when zoomed in. + +![](images/1df129fba7f8a3074d614967e4545622814889cf489599b48c3a8f3cecacc4f8.jpg) +Figure 3. Our robust automatic camera annotation pipeline. + +and corresponding per-frame depth maps collected, we built a robust and fully automatic camera annotation pipeline for both camera extrinsics and intrinsics. As illustrated in Fig. 3, the pipeline has four stages: (1) object-level dynamic masking, (2) reconstruction-friendly video slicing, (3) coarse camera localization and calibration, and (4) tracking-based camera refinement with bundle adjustment. We present several visualizations of our annotated data in Fig. 2, ranging from indoor to outdoor scenes, and from static to dynamic scenarios, demonstrating the robustness and accuracy of our annotation method. + +Dynamic Masking. Distinguishing between dynamic and static regions is crucial for accurate camera parameters estimation. Here, we utilize semantic categories that are potentially dynamic (e.g., cars, people) to segment dynamic objects. Although this may occasionally misclassify static objects, such as stationary parked cars, as dynamic, we find it more robust than flow-based segmentation methods. Specifically, we use Grounded SAM 2 [50] to ensure the temporal consistency of dynamic masks over long sequences. + +Video Slicing. Video slicing plays a critical role in 3D reconstruction by serving two key purposes: First, it eliminates unsuitable video segments (such as scene cuts or motion-blurred frames) that could compromise reconstruction quality. Second, it segments long videos into shorter, temporally coherent clips to enhance processing efficiency. The specific criteria for frame removal are as follows: (1) Insufficient Feature Points: We employ the SIFT [39] feature descriptor to extract keypoints from each frame. Frames exhibiting + +insufficient SIFT keypoints are discarded to ensure robust correspondence estimation. Additionally, frames containing regions with insufficient texture due to low illumination are excluded, as such areas typically exhibit poor feature discriminability and pose challenges for reliable matching. (2) Large Areas of Dynamic Regions: Frames where dynamic regions (obtained from dynamic annotation) dominate over static regions can introduce ambiguity in camera pose estimation. Such frames are filtered out to ensure robust results. (3) Large Motion or Inaccurate Correspondence: Using an off-the-shelf optical flow estimator, RAFT [61], we estimate the magnitude of motion. If these magnitudes exceed a predefined threshold, we truncate the sequence at the current frame, retaining all preceding frames as a valid segment. Similarly, if the ratio of forward-to-backward optical flow errors surpasses a threshold value, we truncate the current frames to ensure temporal coherence. + +Coarse Camera Estimation. For each video slice, we first use DroidCalib [25] to perform a coarse estimation of the camera parameters, leveraging the depth information from static regions. However, due to the lower input resolution of the DroidCalib model and the limited accuracy of its correspondence estimation, a refinement process is necessary to obtain precise camera parameters. + +Camera Refinement. We begin camera refinement by employing the state-of-the-art tracker, CoTracker3 [33], to capture accurate long-term correspondences across the entire slice. SIFT [39] and SuperPoint [12] feature points are extracted from static regions, and then tracked to form correspondences. Subsequently, bundle adjustment is performed on all frames to minimize the accumulated reprojection error of all correspondences. With access to high-quality dense depth, we apply forward-backward reprojection to estimate and minimize errors in 3D space [8], which improves per-frame camera accuracy while preserving inter-frame geometric consistency. Specifically, we solve the nonlinear optimization problem by Ceres Solver [1], and the Cauchy loss function is applied to measure correspondence residuals, which accounts for the problem's sparsity. + +Figure 4. The overall pipeline of AETHER. With different condition combinations, AETHER can serve different tasks. +![](images/b1d7d09061d558099383590e9091c6342f1cecfbd3eea132a6239d3aef80a6a5.jpg) +Image latent +Action latent +Image prediction +Action prediction +Depth prediction +Zero-padding +Noise + +# 3. AETHER Multi-Task World Model + +In this section, we introduce how we post-train a base video diffusion model into a unified multi-task world model AETHER. We use CogVideoX-5b-I2V [77] as our base model. We first give an overview of our framework in Sec. 3.1, then we detail on the input process of depth videos and camera pose trajectories in Sec. 3.2 and Sec. 3.3. Finally, we show how we do model training in Sec. 3.4. + +# 3.1. Method Overview + +Mainstream video diffusion models [27, 40] typically involve two processes: a forward (noising) process and a reverse (denoising) process. The forward process incrementally adds Gaussian noise, denoted as $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ , to a clean latent sample $\mathbf{z}_0 \in \mathbb{R}^{k \times c \times h \times w}$ , where $k, c, h, w$ represent the dimensions of the video latents. Through this process, the clean $\mathbf{z}_0$ is gradually transformed into a noisy latent $\mathbf{z}_t$ . In the reverse process, a learned denoising model $\epsilon_{\theta}$ progressively removes the noise from $\mathbf{z}_t$ to reconstruct the original latent representation. The denoising model $\epsilon_{\theta}$ is conditioned on auxiliary inputs $\mathbf{c}$ and the diffusion timestep $t$ . + +In our method, the target latent $\mathbf{z}_0$ comprises three modalities: color video latents $\mathbf{z}_{\mathrm{c0}}$ , depth video latents $\mathbf{z}_{\mathrm{d0}}$ , and action latents $\mathbf{z}_{\mathrm{a0}}$ . The model additionally takes two types of conditions as input: color video conditions $\mathbf{c}_{\mathrm{c}}$ and action conditions $\mathbf{c}_{\mathrm{a}}$ . For the action modality, we choose camera pose trajectory as a global action, facilitated by our automated camera pose annotation pipeline described earlier. All latents and conditions are channel-wise concatenated. The training objective of AETHER can be expressed as: + +$$ +\mathcal {L} _ {\theta} = \mathbb {E} _ {\substack {\epsilon \sim \mathcal {N} (0, \mathbf {I}) \\ t \sim \mathcal {U} (1, \mathcal {T}) \\ \mathbf {z} _ {0} = \mathbf {z} _ {\mathrm {c} _ {0}} \otimes \mathbf {z} _ {\mathrm {d} _ {0}} \otimes \mathbf {z} _ {\mathrm {a} _ {0}} \\ \mathbf {c} = \mathbf {c} _ {\mathrm {c}} \otimes \mathbf {c} _ {\mathrm {a}}}} \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t, \mathbf {c}) \| ^ {2}, \tag{1} +$$ + +where $\otimes$ denotes the channel-wise concatenation operation, $\mathcal{U}(\cdot)$ represents a uniform distribution, and $\mathcal{T}$ denotes the denoising steps. + +The multi-task objective of AETHER is determined by the specific conditions $\mathbf{c}$ for different tasks. (1) Reconstruction: $\mathbf{c}_{\mathrm{c}}$ represents the input video latents. (2) Video prediction: $\mathbf{c}_{\mathrm{c}}$ takes the latent of observation image as the first frame, while other latents are zero-masked. (3) Goal-conditioned visual planning: The first and last latents of $\mathbf{c}_{\mathrm{c}}$ correspond to the observation and goal images, respectively, with all intermediate latents zero-padded. For the action condition $\mathbf{c}_{\mathrm{a}}$ , it is either entirely zero-masked or contains the full target camera pose trajectory in action-free or action-conditioned control cases. Illustrations are show in Fig. 4. + +# 3.2. Depth Videos Process + +Given a depth video $\mathbf{x}_d$ , we first clip the depth values to a predefined range $[d_{\min}, d_{\max}]$ . Next, we apply a square root transformation and subsequently compute the reciprocal to convert the depth values into disparity, as described in [57]. Each disparity video clip is then normalized in a scale-invariant manner. Subsequently, the normalized disparity values are linearly mapped from $[0,1]$ to $[-1,1]$ . To meet the input requirements of the VAE, the single-channel disparity map is replicated across three channels, as done in prior works [34, 74]. The final depth latent is computed as: + +$$ +\mathbf {x} _ {\text {d i s p}} = \frac {1}{\sqrt {\operatorname {c l i p} \left(\mathbf {x} _ {d} , d _ {\min } , d _ {\max }\right)}}, \tag {2} +$$ + +$$ +\hat {\mathbf {x}} _ {\text {d i s p}} = \frac {\mathbf {x} _ {\text {d i s p}}}{\operatorname* {m a x} \left(\mathbf {x} _ {\text {d i s p}}\right)} \times 2 - 1, \tag {3} +$$ + +$$ +\mathbf {z} _ {d} = \mathcal {E} \left(\hat {\mathbf {x}} _ {\text {d i s p}} \otimes \mathbf {1} _ {3}\right), \tag {4} +$$ + +where $\mathcal{E}$ denotes the 3D VAE, and $\otimes \mathbf{1}_3$ represents the channel-wise replication of 3 times. The above operations are designed to be compatible with the pretrained 3D VAE model, ensuring minimal reconstruction error. + +# 3.3. Camera Trajectories Process + +We transform camera parameters into raymap videos [7] so that video diffusion can process them compatibly. Specifi + +cally, given the intrinsic matrix $\mathbf{K} \in \mathbb{R}^{T \times 3 \times 3}$ and the extrinsic matrix $\mathbf{E} \in \mathbb{R}^{T \times 4 \times 4}$ , the transformation process can be described as follows. + +Translation Scaling and Normalization. The translation component of the camera pose (inverse of extrinsic matrix), $\mathbf{t} \in \mathbb{R}^3$ , is first scaled by a constant factor $s_{\mathrm{ray}}$ and normalized using the maximum disparity value $d_{\mathrm{max}}$ . To suppress large values, we then pass it through a signed $\log(1 + \cdot)$ transformation: + +$$ +\mathbf {t} ^ {\prime} = \frac {\mathbf {t}}{\operatorname* {m a x} \left(x _ {\text {d i s p}}\right)} \cdot s _ {\text {r a y}}, \tag {5} +$$ + +$$ +\mathbf {t} _ {\log} = \operatorname {s i g n} \left(\mathbf {t} ^ {\prime}\right) \cdot \log \left(1 + \left| \mathbf {t} ^ {\prime} \right|\right), \tag {6} +$$ + +where $s_{\mathrm{ray}}$ is a predefined scaling factor. + +Raymap Construction. Using the intrinsic matrix $\mathbf{K}$ , we compute the camera ray directions $\mathbf{r}_{\mathrm{d}}$ in homogeneous coordinates for each pixel. Note that we do not unit normalize it but let it have a unit value along $z$ axis. The ray origins $\mathbf{r}_{\mathrm{o}}$ are set to the translation $\mathbf{t}_{\log}$ , replicated across the spatial dimensions. The raymap in the world coordinate system is obtained by transforming the ray directions $\mathbf{r}_{\mathrm{d}}$ using the extrinsic matrix $\mathbf{E}$ . The final raymap $\mathbf{r}$ consists of 6 channels: 3 for the ray directions $\mathbf{r}_{\mathrm{d}}$ and 3 for the ray origins $\mathbf{r}_{\mathrm{o}}$ . + +Resolution Downsampling. To align the raymap with the latent feature dimensions from the VAE, we perform adjustments both spatially and temporally. Spatially, the raymap is downsampled by a factor of 8 using bilinear interpolation. Temporally, every consecutive group of 4 frames is concatenated along the channel dimension. The resulting rearranged tensor is denoted as $\mathbf{z}_a$ + +Converting raymap back to camera matrix. Given generated raymap sequences rearranged by the time axis $\hat{\mathbf{r}}\in$ $\mathbb{R}^{T\times 6\times h\times w} = [\hat{\mathbf{r}}_{\mathrm{d}},\hat{\mathbf{r}}_{\mathrm{o}}]$ , we first recover the ray origins by: + +$$ +\hat {\mathbf {r} _ {\mathrm {o}}} ^ {\prime} = \frac {1}{s _ {\text {r a y}}} \cdot \operatorname {s i g n} \left(\hat {\mathbf {r} _ {\mathrm {o}}}\right) \cdot \left(\exp \left(\left| \hat {\mathbf {r} _ {\mathrm {o}}} \right|\right) - 1\right), \tag {7} +$$ + +Then, we can recover both the intrinsic and extrinsic through Alg. 1 in the supplementary material. + +# 3.4. Model Training + +We initialize AETHER with pre-trained CogVideo-5b-I2V [77] weights, excluding the additional input and output projection layer channels for depth and raymap action trajectories, which are initialized to zero. Since text prompt conditions are not used, an empty text embedding is provided during both training and inference. + +As the dataset we use contains video clips with variable lengths and frames per second (FPS), we randomly select $T \in \{17, 25, 33, 41\}$ frames, and the FPS is randomly sampled from $\{8, 10, 12, 15, 24\}$ . The RoPE [59] coefficients are linearly interpolated to align with them. + +During training, conditional inputs are randomly masked to generalize across tasks. For $\mathbf{c}_c$ , masking probabilities are: + +$30\%$ for both observation and goal images (visual planning tasks), $40\%$ for observation images only (video prediction), $28\%$ for full-color video latents (4D reconstruction), and $2\%$ for masking all of $\mathbf{c}_c$ . For $\mathbf{c}_a$ , trajectory latents are either kept or fully masked with equal probability (supporting action-free or action-conditioned tasks with raymap conditions). This strategy enables the model to adapt to diverse tasks and input condition settings. + +Our training process consists of two stages. In the first stage, we adopt the loss function of a standard latent diffusion model, which minimizes the mean squared error (MSE) in the latent space. In the second stage, we refine the generated outputs by decoding them into the image space. Specifically, we introduce three additional loss terms: a Multi-Scale Structure Similarity (MS-SSIM) loss [67] for color video, a scale- and shift-invariant loss [49] for depth videos, and a scale- and shift-invariant pointmap loss [66] for pointmaps projected from the generated depths and raymaps. Further details on the stage 2 loss functions are provided in the supplementary material. Notably, the second stage takes about $\frac{1}{4}$ of the training steps used in the first stage. + +We employ a hybrid training strategy combining Fully Sharded Data Parallel (FSDP) [87] with Zero-2 optimization within compute nodes and Distributed Data Parallel (DDP) across nodes. Since depth videos require online normalization, the VAE encoder is also run online during training and operates under DDP. Our implementation processes a local batch size of 4 per GPU, resulting in an effective batch size of 320 samples across 80 A100-80GB GPUs. Training is conducted over two weeks using the AdamW [38] optimizer with a OneCycle [56] learning rate scheduler. + +# 4. Reconstruction Experiments + +In this section, we demonstrate that AETHER can achieve zero-shot reconstruction metrics comparable to or even better than SOTA reconstruction methods. We mainly consider two zero-shot reconstruction tasks: video depth estimation and camera pose estimation. Note that we only denoise for 4 steps for reconstruction tasks. + +# 4.1. Zero-Shot Video Depth Estimation + +Implementation Details. Video depth estimation is evaluated based on two key aspects: per-frame depth quality and inter-frame depth consistency. These evaluations are performed by aligning the predicted depth maps with the ground truth using a per-sequence scale. We use absolute relative error (Abs Rel) and $\delta < 1.25$ (percentage of predicted depths within a 1.25-factor of true depth) as metrics. For implementation, we adopt the settings outlined in CUT3R [65]. Our baselines include both reconstruction-based methods—such as DUSt3R [66], MASt3R [37], MonST3R [82], Spann3R [63], and CUT3R [65]—and diffusion-based depth estimators, including ChronoDepth [55], DepthCrafter [29], + +Table 1. Video depth Evaluation. Methods requiring global alignment are marked "GA". + +
MethodSintel [6]BONN [44]KITTI [21]
Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑
Reconstruction Methods. Alignment: per-sequence scale
DUSt3R-GA [66]0.65645.20.15583.30.14481.3
MASt3R-GA [37]0.64143.90.25270.10.18374.5
MonST3R-GA [82]0.37855.80.06796.30.16874.4
Spann3R [63]0.62242.60.14481.30.19873.7
CUT3R [65]0.42147.90.07893.70.11888.1
AETHER (Ours)0.32450.20.27359.40.05697.8
Diffusion-Based Methods. Alignment: per-sequence scale&shift
ChronoDepth [55]0.42938.30.31851.80.25254.3
DepthCrafter [29]0.59055.50.25356.30.12486.5
DA-V [74]1.25243.70.45731.10.09493.0
AETHER (Ours)0.31460.40.30860.20.05497.7
+ +Table 2. Evaluation on Camera Pose Estimation. + +
MethodSintel [6]TUM-dynamics [58]ScanNet [10]
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Optimization-based Methods
Particle-SfM [86]0.1290.0310.535---0.1360.0230.836
Robust-CVD [36]0.3600.1543.4430.1530.0263.5280.2270.0647.374
CasualSAM [85]0.1410.0350.6150.0710.0101.7120.1580.0341.618
DUSt3R-GA [66]0.4170.2505.7960.0830.0173.5670.0810.0280.784
MASt3R-GA [37]0.1850.0601.4960.0380.0120.4480.0780.0200.475
MonST3R-GA [82]0.1110.0440.8960.0980.0190.9350.0770.0180.529
Feed-forward Methods
DUSt3R [66]0.2900.1327.8690.1400.1063.2860.2460.1088.210
Spann3R [63]0.3290.1104.4710.0560.0210.5910.0960.0230.661
CUT3R [65]0.2130.0660.6210.0460.0150.4730.0990.0220.600
AETHER (Ours)0.1890.0540.6940.0920.0121.1060.1760.0281.204
+ +and DepthAnyVideo (DA-V) [74]. It is important to note that when comparing with diffusion-based depth estimators, we apply scale and shift alignment to the ground truth, as most of these methods are not inherently scale-invariant. All videos are resized with original aspect ratios kept to make the short side align with our model's input size. For videos that exceed the maximum forward processing spatial or temporal size of our model, we employ a sliding window strategy with a stride size of 8. In regions of overlap between windows, we first estimate a relative scale by calculating the average of element-wise division. This relative scale is then used to adjust the latter window's depth predictions. Finally, a linspace-weighted average is applied to the overlapping areas, following approaches similar to prior methods [29, 80]. + +Results and Analysis. Table 1 summarizes the video depth estimation results across Sintel [6], BONN [44], and KITTI [21] datasets. For reconstruction-based methods, AETHER outperforms or is comparable with prior approaches. On Sintel, AETHER achieves the lowest Abs Rel (0.324), surpassing MonST3R-GA (0.378), and competitive $\delta < 1.25$ (50.2). On KITTI, AETHER sets a new benchmark with Abs Rel of 0.056 and $\delta < 1.25$ of 97.8, outperforming the previous SOTA CUT3R (Abs Rel: 0.118, $\delta < 1.25$ : 88.1). Among diffusion-based methods, AETHER shows consistent superiority. It achieves the best performance on + +Sintel (Abs Rel: 0.314, $\delta < 1.25$ : 60.4) and KITTI (Abs Rel: 0.054, $\delta < 1.25$ : 97.7), significantly outperforming ChronoDepth [55], DepthCrafter [29], and DA-V [74]. On BONN, AETHER achieves the highest $\delta < 1.25$ (60.2) with competitive Abs Rel (0.308). + +# 4.2. Zero-Shot Camera Pose Estimation + +Implementation Details. Following MonST3R [82] and CUT3R [65], we evaluate camera pose estimation accuracy on the Sintel [6], TUM Dynamics [58], and ScanNet [10] datasets. Notably, both Sintel and TUM Dynamics contain highly dynamic objects, presenting significant challenges for traditional Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM) systems. We report Absolute Translation Error (ATE), Relative Translation Error (RPE Trans), and Relative Rotation Error (RPE Rot) after Sim(3) alignment with the ground truth, following the methodology in [65]. The implementation settings are consistent with those used in CUT3R [65]. All videos are resized with original aspect ratios kept and then center cropped to align with our model's input size. For long videos exceeding our model's maximum temporal forward processing length, a sliding window strategy with a stride size of 32 is employed. In overlapping regions between windows, camera poses are aligned following prior methods [64]. Transla + +tion alignment is performed using linear interpolation, while quaternion rotations are interpolated with spherical linear interpolation. Additionally, we observed that the generated camera trajectories exhibit noise, likely due to the limited number of denoising steps. To mitigate this, we apply a simple Kalman filter [71] to smooth the trajectories. + +Results and Analysis. Table 2 shows the evaluation results. Among feed-forward methods, AETHER achieves the best ATE (0.189) and RPE Trans (0.054) on Sintel [6], while remaining competitive in RPE Rot (0.694) compared to CUT3R (0.621). On TUM Dynamics [58], AETHER achieves the best RPE Trans (0.012). For other metrics, AETHER is also comparable with other specialist models. + +# 5. Generation and Planning Experiments + +In this section, we first show video prediction, with or without action conditioning, quantitatively or qualitatively, in Sec. 5.1. We then show visual planning abilities in Sec. 5.2. More visualizations are in the supplementary material. + +# 5.1. Video Prediction + +Implementation Details. We use CogVideoX-5b-I2V [77] as our baseline. To ensure a fair comparison, we construct a validation dataset comprising two subsets: in-domain and out-domain data. The in-domain subset includes novel, unseen scenes from the same synthetic environments as the training dataset, while the out-domain subset consists of data from entirely new synthetic environments. Both models are provided with the first frame as the observation image. For action-free prediction, since CogVideoX depends heavily on text prompts, we use GPT-4o [31] to generate image descriptions and predictions of future scenes as prompts for CogVideoX. In contrast, AETHER is evaluated using empty text prompts. For action-conditioned prediction, we also labeled camera trajectories in the validation dataset and generated corresponding raymap sequences as action conditions for AETHER. For the baseline, in addition to the prompts used for action-free prediction, we use GPT-4o [31] to generate detailed descriptions of object and camera movements, enabling the baseline to use language as action conditions. We use the default classifier-free guidance value of 6 on text prompts for CogVideoX and a value of 3 on the observation image for AETHER. No classifier-free guidance is applied to action conditions to ensure fairness. Evaluation metrics follow VBenchmark [30], a standard benchmark for video generation, with additional details on prompts and evaluation metrics provided in the supplementary material. + +Image-to-Video Prediction. We first evaluate image-to-video prediction without action conditions. The results, presented in Tab. 3, show that AETHER consistently outperforms the baseline on both in-domain and out-domain validation sets. Notably, AETHER demonstrates a larger performance improvement on out-domain data, which can likely be at + +tributed to the baseline model's pre-training data containing domains similar to the in-domain dataset. + +Action-Conditioned Video Prediction. To assess the effectiveness of our post-training in improving action control and action-following capabilities, we conduct action-conditioned video prediction experiments. The results, shown in Tab. 4, indicate that AETHER consistently outperforms the baseline in both in-domain and out-domain settings. Notably, CogVideoX tends to generate static scenes with high visual and aesthetic quality, while AETHER accurately follows the action conditions, producing highly dynamic scenes. These results validate the effectiveness of our framework and the advantages of using camera trajectories as action conditions. + +# 5.2. Visual Planning + +Implementation Details. We evaluate the action-conditioned navigation capability of AETHER on our validation set. To demonstrate the effectiveness of our multi-task objective, particularly the incorporation of the reconstruction objective, we also post-train an ablation model without the video depth objective, denoted as AETHER-no-depth. Given the observation image, goal image, and camera trajectory, the resulting video should be highly determined. Thus, we report pixel-wise reconstruction metrics, including PSNR, SSIM [68], MS-SSIM [67], and LPIPS [83], for action-conditioned navigation. For the action-free case, which represents a visual path navigation task, we also report the VBench metrics. We do not use any classifier-free guidance on both tasks. + +Action-Conditioned Navigation. The quantitative results for action-conditioned navigation are presented in Tab. 5. Aether consistently outperforms the ablation model, demonstrating the significant benefits of incorporating the reconstruction objective into generative models. + +Visual Path Planning. In the absence of action conditions, this task evaluates the model's ability to function as a "world model as an agent," requiring it to plan a path from the observation image to the goal image. The results, shown in Tab. 6, indicate that the reconstruction objective substantially improves the model's visual path planning capability. Additionally, qualitative visualizations on completely in-the-wild data are provided in supplementary material. + +# 6. Related Work + +World Models. World models have emerged as a critical framework in artificial intelligence, enabling agents to simulate, understand, and predict environmental dynamics. Early work [23] introduced latent representations and recurrent neural networks for decision-making. Recent advancements include Cat3D [20] for 3D scene generation, Cat4D [72] for dynamic 4D environments, and Genie 2 [46], a large-scale model for interactive 3D worlds. Motion Prompting [22] further enables precise video generation control. These ad + +Table 3. VBenchmark [30] Metrics of Video Prediction without Action Conditions. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each group, the better performance is highlighted in bold. + +
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX89.36/84.61/87.7792.72/91.43/92.2998.24/96.93/97.8188.75/95.00/90.8354.49/53.58/54.1855.38/52.29/54.3579.01/77.52/78.51
AETHER91.50/87.55/90.1894.29/93.62/94.0798.54/98.19/98.4296.25/100.00/97.5054.36/52.58/53.7755.08/54.88/55.0180.34/79.42/80.04
+ +Table 4. VBenchmark [30] Metrics of Action-Conditioned Video Prediction. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each metric group, the better performance is highlighted in bold. + +
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX91.56/88.23/90.5192.98/92.29/92.7798.44/97.81/98.2483.87/93.02/86.7656.19/57.43/56.5856.48/61.60/58.1079.56/80.70/79.92
AETHER90.73/93.27/91.5493.61/95.03/94.0698.53/98.62/98.56100.00/83.72/94.8555.04/56.50/55.5053.89/63.23/56.8480.33/81.55/80.71
+ +Table 5. Pixel-wise Metrics of Action-Conditioned Navigation. Comparison of performance between AETHER-no-depth and AETHER on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold. + +
PSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓
AETHER-no-depth19.13/18.67/18.970.5630/0.4830/0.53530.5467/0.5204/0.53760.3116/0.2995/0.3074
AETHER19.87/19.37/19.700.5803/0.5058/0.55450.5830/0.5627/0.57600.2691/0.2599/0.2659
+ +Table 6. Quantitative Results of Action-Free Visual Path Planning. Comparison of performance between Aether and Aether-no-depth on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold. + +
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
Aether-no-depth88.68/89.61/88.6193.62/93.92/93.6698.37/98.31/98.3297.06/91.67/96.1554.12/56.26/54.7851.77/58.46/54.2979.11/80.43/79.59
Aether (Ours)89.69/91.61/90.3693.88/94.58/94.1398.50/98.40/98.4697.06/91.67/95.1955.83/56.87/56.1954.71/61.13/56.9380.21/81.53/80.67
+ +vancements demonstrate the evolution of world models toward dynamic, interactive, and controllable applications in robotics, gaming, and simulation. + +Reconstruction. Reconstruction has been a long-standing topic in computer vision, with notable progress in both traditional and learning-based methods. Classical approaches, such as Structure-from-Motion (SfM) [9, 26, 45, 53] and Multi-View Stereo (MVS) [19, 54], rely on multi-view geometry for feature matching, pose estimation, and dense point cloud generation, demonstrating robust performance in controlled settings. Deep learning has introduced powerful alternatives, tackling sub-tasks like feature matching [14, 52], point tracking [13, 64], triangulation [42], and MVS [78, 81]. End-to-end methods now directly predict point maps [37, 66] or depth maps from images [4, 76], often incorporating camera parameters [70]. Recently, diffusion models have achieved breakthroughs in image and video generation [27, 35, 43, 73, 77], inspiring novel 3D reconstruction approaches that leverage rich 2D priors [18, 29, 34, 41, 74, 75, 88, 90]. These methods demonstrate the potential of integrating diffusion-based 2D knowledge into 3D modeling. + +Video Generation. Video generation has evolved from foundational techniques like DDPM [27, 43] to modern frameworks leveraging diffusion-based techniques. Advances such as latent diffusion [51] and diffusion transformers [47] have improved generation quality, while models like Sora [5] and Stable Video Diffusion (SVD) [3] emphasize temporal consistency. Open-source models, including LTX Video [24], CogVideoX [77], and Hunyuan Video [35], offer increased flexibility, and techniques like multi-scale architectures (e.g., + +Pyramid Flow [32]) enhance motion dynamics. These advancements highlight rapid progress, with ongoing efforts to improve scalability and temporal stability. + +# 7. Conclusion and Limitations + +In this work, we introduce AETHER, a geometry-aware multitask world model that reconstructs 4D dynamic videos, predicts future frames conditioned on observation images and actions, and performs visual planning based on observation and goal images. We propose an automatic 4D synthetic data labeling pipeline, enabling AETHER to train on synthetic data and generalize to unseen real-world data in a zero-shot manner. Post-trained on the CogVideoX base model, AETHER achieves state-of-the-art or competitive reconstruction performance and outperforms baselines in generation and planning tasks, demonstrating the value of incorporating reconstruction objectives into world modeling. + +However, limitations remain. Camera pose estimation is less accurate, likely due to incompatibilities between raymap representation and prior video diffusion models. Indoor scene reconstruction also lags behind outdoor performance, likely due to the predominance of outdoor training data. Additionally, predictions without language prompts often fail in highly dynamic scenes. Future work can address these by exploring novel action representations, co-training with real-world data, and retaining the language prompting capabilities of the base model. + +# Acknowledgments + +This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Sameer Agarwal, Keir Mierle, et al. Ceres solver: Tutorial & reference. Google Inc, 2(72):8, 2012. 3 +[2] Eloi Alonso, Adam Jolley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37:58757-58791, 2025. 2 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 8 +[4] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8 +[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8 +[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 6, 7 +[7] Junyi Chen, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Where am i and what will i see: An auto-regressive model for spatial localization and view prediction, 2024. 4 +[8] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 3 +[9] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 8 +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6 +[11] Decart, Julian Quevedo, Quinn McIntyre, Spruce Campbell, Xinlei Chen, and Robert Wachen. Oasis: A universe in a Transformer. 2024. 2 +[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on + +computer vision and pattern recognition workshops, pages 224-236, 2018. 3 +[13] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 8 +[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 8 +[15] Tom Erez, Yuval Tassa, and Emanuel Todorov. Infinite-horizon model predictive control for periodic tasks with contacts. Robotics: Science and systems VII, 73, 2012. 2 +[16] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2 +[17] Ruili Feng, Han Zhang, Zhantao Yang, Jie Xiao, Zhilei Shu, Zhiheng Liu, Andy Zheng, Yukun Huang, Yu Liu, and Hongyang Zhang. The matrix: Infinite-horizon world generation with real-time moving control. arXiv preprint arXiv:2412.03568, 2024. 2 +[18] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2024. 8 +[19] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 8 +[20] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 7 +[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6 +[22] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, Chen Sun, Oliver Wang, Andrew Owens, and Deqing Sun. Motion prompting: Controlling video generation with motion trajectories, 2024. 2, 7 +[23] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 7 +[24] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 2, 8 +[25] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3 + +[26] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 8 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4, 8 +[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[29] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 8 +[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7, 8 +[31] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7 +[32] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2, 8 +[33] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3 +[34] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4, 8 +[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Junkun Yuan, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yanxin Long, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. Hunyuanvideo: A systematic framework for large video generative models, 2024. 2, 8 +[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 6 + +[37] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 5, 6, 8 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[39] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3 +[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4 +[41] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 8 +[42] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 8 +[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 8 +[44] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 6 +[45] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, pages 58-77. Springer, 2024. 8 +[46] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Roktaschel. Genie 2: A large-scale foundation world model. 2024. 2, 7 +[47] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 8 +[48] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, + +Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Eliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2025. 1 +[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 5 +[50] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 3 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8 +[52] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 8 +[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 8 +[54] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 8 +[55] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Vitor Guizilini, Yue Wang, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv preprint arXiv:2406.01493, 2024. 5, 6 +[56] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5 +[57] Ziyang Song, Zerong Wang, Bo Li, Hao Zhang, Ruijie Zhu, Li Liu, Peng-Tao Jiang, and Tianzhu Zhang. Depthmaster: + +Taming diffusion models for monocular depth estimation. arXiv preprint arXiv:2501.02576, 2025.4 +[58] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 6, 7 +[59] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[60] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2 +[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3 +[62] Veo-Team,.; Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathewson, Kurtis David, Matthieu Kim Lorraine, Marc van Zee, Medhini Narasimhan, Miaosen Wang, Mohammad Babaeizadeh, Nelly Papalampidi, Nick Pezzotti, Nilpa Jha, Parker Barnes, Pieter-Jan Kindermans, Rachel Hornung, Ruben Villegas, Ryan Poplin, Salah Zaiem, Sander Dieleman, Sayna Ebrahimi, Scott Wisdom, Serena Zhang, Shlomi Fruchter, Signe Nørly, Weizhe Hua, Xinchen Yan, Yuqing Du, and Yutian Chen. Veo 2. 2024. 1 +[63] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 5, 6 +[64] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 6, 8 +[65] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025. 2, 5, 6 +[66] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 5, 6, 8 +[67] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 5, 7 +[68] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Mo + +tionctrl: A unified and flexible motion controller for video generation. 2023. 2 +[70] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xiangyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 230-247. Springer, 2020. 8 +[71] Greg Welch, Gary Bishop, et al. An introduction to the kalman filter. 1995. 7 +[72] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 7 +[73] Wanghan Xu, Xiaoyu Yue, Zidong Wang, Yao Teng, Wenlong Zhang, Xihui Liu, Luping Zhou, Wanli Ouyang, and Lei Bai. Exploring representation-aligned latent space for better generation. arXiv preprint arXiv:2502.00359, 2025. 8 +[74] Honghui Yang, Di Huang, Wei Yin, Chunhua Shen, Haifeng Liu, Xiaofei He, Binbin Lin, Wanli Ouyang, and Tong He. Depth any video with scalable synthetic data. arXiv preprint arXiv:2410.10815, 2024. 2, 4, 6, 8 +[75] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, et al. Unipad: A universal pre-training paradigm for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15238-15250, 2024. 8 +[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 8 +[77] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5, 7, 8 +[78] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 8 +[79] Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Creating new games with generative interactive videos, 2025. 2 +[80] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 6 +[81] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1):199-214, 2023. 8 +[82] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 5, 6 + +[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[84] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance, 2024. 2 +[85] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 6 +[86] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2, 6 +[87] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023.5 +[88] Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Hengshuang Zhao, Chunhua Shen, Yu Qiao, Tong He, et al. Ponderv2: Pave the way for 3d foundation model with a universal pre-training paradigm. arXiv preprint arXiv:2310.08586, 2023. 8 +[89] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2 +[90] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2, 8 \ No newline at end of file diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/images.zip b/2025/Aether_ Geometric-Aware Unified World Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..afbd60760078e449ca73062ce411683719bdce23 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3b7c0ffc55c52e0be0b3e4e7a16689d88d1f3666ab44a90b305ca11b8aa6d24 +size 805815 diff --git a/2025/Aether_ Geometric-Aware Unified World Modeling/layout.json b/2025/Aether_ Geometric-Aware Unified World Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0949acc610a7a661578af8a2066abc599eb68f10 --- /dev/null +++ b/2025/Aether_ Geometric-Aware Unified World Modeling/layout.json @@ -0,0 +1,8986 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 470, + 121 + ], + "type": "text", + "content": "AETHER: Geometric-Aware Unified World Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 143, + 496, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 143, + 496, + 220 + ], + "spans": [ + { + "bbox": [ + 113, + 143, + 496, + 220 + ], + "type": "text", + "content": "Haoyi Zhu* Yifan Wang* Jianjun Zhou* Wenzheng Chang* Yang Zhou* Zizun Li* Junyi Chen* Chunhua Shen Jiangmiao Pang Tong He USTC Shanghai AI Lab SII SJTU ZJU FDU *Equal Contribution † Corresponding Author https://aether-world.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 246, + 555, + 668 + ], + "blocks": [ + { + "bbox": [ + 56, + 246, + 555, + 668 + ], + "lines": [ + { + "bbox": [ + 56, + 246, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 555, + 668 + ], + "type": "image", + "image_path": "94cd943bbc5d4b733ea437470020d9295c18436b009cb913494dce09c0b22851.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 675, + 555, + 730 + ], + "lines": [ + { + "bbox": [ + 55, + 675, + 555, + 730 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 555, + 730 + ], + "type": "text", + "content": "Figure 1. An overview of AETHER, trained entirely on synthetic data. The figure highlights its three key capabilities: 4D reconstruction, action-conditioned 4D prediction, and visual planning, all demonstrated on unseen real-world data. The 4D reconstruction examples are derived from MovieGen [48] and Veo 2 [62] generated videos, while the action-conditioned prediction uses an observation image from a university classroom. The visual planning example utilizes observation and goal images from an office building. Better viewed when zoomed in. Additional visualizations can be found on our website." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 110, + 32 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 110, + 32 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 110, + 32 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 35 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 35 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8535" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 200, + 85 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 373 + ], + "type": "text", + "content": "The integration of geometric reconstruction and generative modeling remains a critical challenge in developing AI systems capable of human-like spatial reasoning. This paper proposes AETHER, a unified framework that enables geometry-aware reasoning in world models by jointly optimizing three core capabilities: (1) 4D dynamic reconstruction, (2) action-conditioned video prediction, and (3) goal-conditioned visual planning. Through task-interleaved feature learning, AETHER achieves synergistic knowledge sharing across reconstruction, prediction, and planning objectives. Building upon video generation models, our framework demonstrates zero-shot synthetic-to-real generalization despite never observing real-world data during training. Furthermore, our approach achieves zero-shot generalization in both action following and reconstruction tasks, thanks to its intrinsic geometric modeling. Notably, even without real-world data, its reconstruction performance is comparable with or even better than that of domain-specific models. Additionally, AETHER employs camera trajectories as geometry-informed action spaces, enabling effective action-conditioned prediction and visual planning. We hope our work inspires the community to explore new frontiers in physically-reasonable world modeling and its applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 399, + 135, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 135, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 135, + 411 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 420, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 295, + 443 + ], + "type": "text", + "content": "\"Prediction is not just one of the things your brain does. It is the primary function of the neocortex.\"" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 445, + 292, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 445, + 292, + 455 + ], + "spans": [ + { + "bbox": [ + 129, + 445, + 292, + 455 + ], + "type": "text", + "content": "Jeff Hawkins, On Intelligence (2004)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 295, + 713 + ], + "type": "text", + "content": "The development of visual intelligence systems capable of comprehending and forecasting the physical world remains a cornerstone of AI research. World models have emerged as a foundational paradigm for building autonomous systems that not only perceive but also anticipate environmental dynamics to make reasonable actions. At their core, three capabilities stand out: First, perception equips the system with the ability to capture the intricate four-dimensional (4D) changes—integrating spatial and temporal information—that are essential for understanding the physical world [37, 63, 65, 66, 82, 86]. This continuous sensing of dynamic cues enables a geometric representation of the environment. Second, prediction leverages this perceptual information to forecast how the environment will evolve under specific actions, thereby providing a foresight into future states [3, 24, 28, 32, 35, 60, 77]. Finally, planning uses these predictive insights to determine the optimal sequence of actions required to achieve a given goal. Together, these three aspects empower world models to not only represent the current state of the environment but also to anticipate and navigate its future dynamics effectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 263 + ], + "type": "text", + "content": "Motivated by these principles, we introduce AETHER, a unified framework that, for the first time, bridges reconstruction, prediction, and planning, as shown in Fig. 1. AETHER leverages pre-trained video generation models [28, 77] and is further refined via post-training with synthetic 4D data. Although multiple action modalities exist, ranging from keyboard inputs [2, 11, 15, 46, 79] to human or robotic motions [16, 84, 89, 90] and point flows [22, 69], we choose camera pose trajectories as our global action representation. This choice is particularly effective for ego-view tasks: in navigation, camera trajectories directly correspond to the navigation paths, while in robotic manipulation, the movement of an in-hand camera captures the 6D motion of the end effector. To address the scarcity of 4D data, we utilize RGB-D synthetic video data and propose a robust camera pose annotation pipeline to reconstruct full 4D dynamics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 264, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 556, + 479 + ], + "type": "text", + "content": "Through a simple training strategy that randomly combines input and output modalities, our method transforms the base video generation model into a unified, multi-task world model with three key capabilities: (1) Depth and camera pose estimation from full video sequences; (2) Video prediction conditioned on an initial observation—with the option to incorporate a camera trajectory action; and (3) Goal-conditioned visual planning based on observation-goal image pairs. We transform depth videos into scale-invariant normalized disparity representations to meet the tokenization requirements of video VAEs. Simultaneously, we encode camera trajectories as scale-invariant raymap sequence representations, structured to align with the spatiotemporal framework of diffusion transformers (DiTs). By dynamically integrating cross-task and cross-modal conditioning signals during training, our framework enables synergistic knowledge transfer across heterogeneous inputs, facilitating joint optimization for multi-task generative modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 647 + ], + "type": "text", + "content": "In summary, this work introduces AETHER, a unified world model that integrates reconstruction, prediction, and planning through multi-task learning on synthetic 4D data. We propose a robust automatic data annotation pipeline to obtain accurate 4D geometry knowledge. By combining geometric reasoning with generative priors, our framework achieves robust zero-shot transfer to real-world tasks, demonstrating accuracy comparable to SOTA reconstruction models while enabling actionable planning capabilities. The results underscore the value of synergistic 4D modeling for advancing spatial intelligence in AI systems. We hope that AETHER will serve as an effective starter framework for the community to explore post-training world models with scalable synthetic data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 657, + 528, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 528, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 528, + 671 + ], + "type": "text", + "content": "2. 4D Synthetic Data Annotation Pipeline" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "For the synthetic data source, we follow DA-V [74] and TheMatrix [17] to collect large-scale synthetic data with high-quality video depth data. With high-resolution RGB videos" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8536" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 529, + 228 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 529, + 228 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 529, + 228 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 529, + 228 + ], + "type": "image", + "image_path": "7efdfaea582c4f64a08371bfce5c7e861f975166dbefa05849b7a7aae90a8ae8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 229, + 499, + 239 + ], + "lines": [ + { + "bbox": [ + 109, + 229, + 499, + 239 + ], + "spans": [ + { + "bbox": [ + 109, + 229, + 499, + 239 + ], + "type": "text", + "content": "Figure 2. Some visualization results of data annotated through our pipeline. Better viewed when zoomed in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 243, + 288, + 353 + ], + "blocks": [ + { + "bbox": [ + 61, + 243, + 288, + 353 + ], + "lines": [ + { + "bbox": [ + 61, + 243, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 288, + 353 + ], + "type": "image", + "image_path": "1df129fba7f8a3074d614967e4545622814889cf489599b48c3a8f3cecacc4f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 354, + 282, + 364 + ], + "lines": [ + { + "bbox": [ + 67, + 354, + 282, + 364 + ], + "spans": [ + { + "bbox": [ + 67, + 354, + 282, + 364 + ], + "type": "text", + "content": "Figure 3. Our robust automatic camera annotation pipeline." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 495 + ], + "type": "text", + "content": "and corresponding per-frame depth maps collected, we built a robust and fully automatic camera annotation pipeline for both camera extrinsics and intrinsics. As illustrated in Fig. 3, the pipeline has four stages: (1) object-level dynamic masking, (2) reconstruction-friendly video slicing, (3) coarse camera localization and calibration, and (4) tracking-based camera refinement with bundle adjustment. We present several visualizations of our annotated data in Fig. 2, ranging from indoor to outdoor scenes, and from static to dynamic scenarios, demonstrating the robustness and accuracy of our annotation method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 605 + ], + "type": "text", + "content": "Dynamic Masking. Distinguishing between dynamic and static regions is crucial for accurate camera parameters estimation. Here, we utilize semantic categories that are potentially dynamic (e.g., cars, people) to segment dynamic objects. Although this may occasionally misclassify static objects, such as stationary parked cars, as dynamic, we find it more robust than flow-based segmentation methods. Specifically, we use Grounded SAM 2 [50] to ensure the temporal consistency of dynamic masks over long sequences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Video Slicing. Video slicing plays a critical role in 3D reconstruction by serving two key purposes: First, it eliminates unsuitable video segments (such as scene cuts or motion-blurred frames) that could compromise reconstruction quality. Second, it segments long videos into shorter, temporally coherent clips to enhance processing efficiency. The specific criteria for frame removal are as follows: (1) Insufficient Feature Points: We employ the SIFT [39] feature descriptor to extract keypoints from each frame. Frames exhibiting" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 449 + ], + "type": "text", + "content": "insufficient SIFT keypoints are discarded to ensure robust correspondence estimation. Additionally, frames containing regions with insufficient texture due to low illumination are excluded, as such areas typically exhibit poor feature discriminability and pose challenges for reliable matching. (2) Large Areas of Dynamic Regions: Frames where dynamic regions (obtained from dynamic annotation) dominate over static regions can introduce ambiguity in camera pose estimation. Such frames are filtered out to ensure robust results. (3) Large Motion or Inaccurate Correspondence: Using an off-the-shelf optical flow estimator, RAFT [61], we estimate the magnitude of motion. If these magnitudes exceed a predefined threshold, we truncate the sequence at the current frame, retaining all preceding frames as a valid segment. Similarly, if the ratio of forward-to-backward optical flow errors surpasses a threshold value, we truncate the current frames to ensure temporal coherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 533 + ], + "type": "text", + "content": "Coarse Camera Estimation. For each video slice, we first use DroidCalib [25] to perform a coarse estimation of the camera parameters, leveraging the depth information from static regions. However, due to the lower input resolution of the DroidCalib model and the limited accuracy of its correspondence estimation, a refinement process is necessary to obtain precise camera parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Camera Refinement. We begin camera refinement by employing the state-of-the-art tracker, CoTracker3 [33], to capture accurate long-term correspondences across the entire slice. SIFT [39] and SuperPoint [12] feature points are extracted from static regions, and then tracked to form correspondences. Subsequently, bundle adjustment is performed on all frames to minimize the accumulated reprojection error of all correspondences. With access to high-quality dense depth, we apply forward-backward reprojection to estimate and minimize errors in 3D space [8], which improves per-frame camera accuracy while preserving inter-frame geometric consistency. Specifically, we solve the nonlinear optimization problem by Ceres Solver [1], and the Cauchy loss function is applied to measure correspondence residuals, which accounts for the problem's sparsity." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8537" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 531, + 229 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 531, + 229 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 531, + 229 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 531, + 229 + ], + "type": "image", + "image_path": "b1d7d09061d558099383590e9091c6342f1cecfbd3eea132a6239d3aef80a6a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 232, + 161, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 232, + 161, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 161, + 243 + ], + "type": "text", + "content": "Image latent" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 161, + 232, + 210, + 243 + ], + "lines": [ + { + "bbox": [ + 161, + 232, + 210, + 243 + ], + "spans": [ + { + "bbox": [ + 161, + 232, + 210, + 243 + ], + "type": "text", + "content": "Action latent" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 211, + 232, + 275, + 243 + ], + "lines": [ + { + "bbox": [ + 211, + 232, + 275, + 243 + ], + "spans": [ + { + "bbox": [ + 211, + 232, + 275, + 243 + ], + "type": "text", + "content": "Image prediction" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 275, + 232, + 339, + 243 + ], + "lines": [ + { + "bbox": [ + 275, + 232, + 339, + 243 + ], + "spans": [ + { + "bbox": [ + 275, + 232, + 339, + 243 + ], + "type": "text", + "content": "Action prediction" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 339, + 232, + 402, + 243 + ], + "lines": [ + { + "bbox": [ + 339, + 232, + 402, + 243 + ], + "spans": [ + { + "bbox": [ + 339, + 232, + 402, + 243 + ], + "type": "text", + "content": "Depth prediction" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 402, + 232, + 463, + 243 + ], + "lines": [ + { + "bbox": [ + 402, + 232, + 463, + 243 + ], + "spans": [ + { + "bbox": [ + 402, + 232, + 463, + 243 + ], + "type": "text", + "content": "Zero-padding" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 464, + 232, + 499, + 243 + ], + "lines": [ + { + "bbox": [ + 464, + 232, + 499, + 243 + ], + "spans": [ + { + "bbox": [ + 464, + 232, + 499, + 243 + ], + "type": "text", + "content": "Noise" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 92, + 244, + 517, + 255 + ], + "lines": [ + { + "bbox": [ + 92, + 244, + 517, + 255 + ], + "spans": [ + { + "bbox": [ + 92, + 244, + 517, + 255 + ], + "type": "text", + "content": "Figure 4. The overall pipeline of AETHER. With different condition combinations, AETHER can serve different tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 260, + 248, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 248, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 248, + 273 + ], + "type": "text", + "content": "3. AETHER Multi-Task World Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 365 + ], + "type": "text", + "content": "In this section, we introduce how we post-train a base video diffusion model into a unified multi-task world model AETHER. We use CogVideoX-5b-I2V [77] as our base model. We first give an overview of our framework in Sec. 3.1, then we detail on the input process of depth videos and camera pose trajectories in Sec. 3.2 and Sec. 3.3. Finally, we show how we do model training in Sec. 3.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 371, + 162, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 162, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 162, + 383 + ], + "type": "text", + "content": "3.1. Method Overview" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": "Mainstream video diffusion models [27, 40] typically involve two processes: a forward (noising) process and a reverse (denoising) process. The forward process incrementally adds Gaussian noise, denoted as " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": ", to a clean latent sample " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 \\in \\mathbb{R}^{k \\times c \\times h \\times w}" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "k, c, h, w" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " represent the dimensions of the video latents. Through this process, the clean " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " is gradually transformed into a noisy latent " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": ". In the reverse process, a learned denoising model " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " progressively removes the noise from " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " to reconstruct the original latent representation. The denoising model " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " is conditioned on auxiliary inputs " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": " and the diffusion timestep " + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 388, + 296, + 519 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "In our method, the target latent " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " comprises three modalities: color video latents " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{c0}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", depth video latents " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{d0}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", and action latents " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{a0}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ". The model additionally takes two types of conditions as input: color video conditions " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " and action conditions " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ". For the action modality, we choose camera pose trajectory as a global action, facilitated by our automated camera pose annotation pipeline described earlier. All latents and conditions are channel-wise concatenated. The training objective of AETHER can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 638, + 296, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 638, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 88, + 638, + 296, + 675 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\theta} = \\mathbb {E} _ {\\substack {\\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\\\ t \\sim \\mathcal {U} (1, \\mathcal {T}) \\\\ \\mathbf {z} _ {0} = \\mathbf {z} _ {\\mathrm {c} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {d} _ {0}} \\otimes \\mathbf {z} _ {\\mathrm {a} _ {0}} \\\\ \\mathbf {c} = \\mathbf {c} _ {\\mathrm {c}} \\otimes \\mathbf {c} _ {\\mathrm {a}}}} \\| \\epsilon - \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t, \\mathbf {c}) \\| ^ {2}, \\tag{1}", + "image_path": "735d832a6d5c672879a2fd660d523766d78e77d9ba8a24722c7d04461d340559.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denotes the channel-wise concatenation operation, " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(\\cdot)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represents a uniform distribution, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denotes the denoising steps." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": "The multi-task objective of AETHER is determined by the specific conditions " + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": " for different tasks. (1) Reconstruction: " + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": " represents the input video latents. (2) Video prediction: " + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": " takes the latent of observation image as the first frame, while other latents are zero-masked. (3) Goal-conditioned visual planning: The first and last latents of " + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": " correspond to the observation and goal images, respectively, with all intermediate latents zero-padded. For the action condition " + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 261, + 555, + 393 + ], + "type": "text", + "content": ", it is either entirely zero-masked or contains the full target camera pose trajectory in action-free or action-conditioned control cases. Illustrations are show in Fig. 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 400, + 437, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 437, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 437, + 412 + ], + "type": "text", + "content": "3.2. Depth Videos Process" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": "Given a depth video " + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_d" + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": ", we first clip the depth values to a predefined range " + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "inline_equation", + "content": "[d_{\\min}, d_{\\max}]" + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": ". Next, we apply a square root transformation and subsequently compute the reciprocal to convert the depth values into disparity, as described in [57]. Each disparity video clip is then normalized in a scale-invariant manner. Subsequently, the normalized disparity values are linearly mapped from " + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 417, + 555, + 537 + ], + "type": "text", + "content": ". To meet the input requirements of the VAE, the single-channel disparity map is replicated across three channels, as done in prior works [34, 74]. The final depth latent is computed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 543, + 555, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 543, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 365, + 543, + 555, + 570 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {\\text {d i s p}} = \\frac {1}{\\sqrt {\\operatorname {c l i p} \\left(\\mathbf {x} _ {d} , d _ {\\min } , d _ {\\max }\\right)}}, \\tag {2}", + "image_path": "981b08a904065232d784300a599c1d972fb5f389be2f9f7e318c0c5b45996a31.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 366, + 571, + 555, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 571, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 366, + 571, + 555, + 595 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {\\text {d i s p}} = \\frac {\\mathbf {x} _ {\\text {d i s p}}}{\\operatorname* {m a x} \\left(\\mathbf {x} _ {\\text {d i s p}}\\right)} \\times 2 - 1, \\tag {3}", + "image_path": "09e27c3ea3ea90605857feb401a94511d0727046644aa94190e28a0c202a57ef.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 378, + 597, + 555, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 597, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 378, + 597, + 555, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {d} = \\mathcal {E} \\left(\\hat {\\mathbf {x}} _ {\\text {d i s p}} \\otimes \\mathbf {1} _ {3}\\right), \\tag {4}", + "image_path": "0cb82e72afcab9419f5b3a9aeb344486c8af33a55ecc01888a2047894ea540b1.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "text", + "content": " denotes the 3D VAE, and " + }, + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\otimes \\mathbf{1}_3" + }, + { + "bbox": [ + 313, + 617, + 554, + 665 + ], + "type": "text", + "content": " represents the channel-wise replication of 3 times. The above operations are designed to be compatible with the pretrained 3D VAE model, ensuring minimal reconstruction error." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "text", + "content": "3.3. Camera Trajectories Process" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We transform camera parameters into raymap videos [7] so that video diffusion can process them compatibly. Specifi" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8538" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": "cally, given the intrinsic matrix " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}^{T \\times 3 \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": " and the extrinsic matrix " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{T \\times 4 \\times 4}" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": ", the transformation process can be described as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "content": "Translation Scaling and Normalization. The translation component of the camera pose (inverse of extrinsic matrix), " + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "content": ", is first scaled by a constant factor " + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{ray}}" + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "content": " and normalized using the maximum disparity value " + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "content": ". To suppress large values, we then pass it through a signed " + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "inline_equation", + "content": "\\log(1 + \\cdot)" + }, + { + "bbox": [ + 55, + 108, + 297, + 180 + ], + "type": "text", + "content": " transformation:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 121, + 185, + 295, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 185, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 121, + 185, + 295, + 209 + ], + "type": "interline_equation", + "content": "\\mathbf {t} ^ {\\prime} = \\frac {\\mathbf {t}}{\\operatorname* {m a x} \\left(x _ {\\text {d i s p}}\\right)} \\cdot s _ {\\text {r a y}}, \\tag {5}", + "image_path": "3b09fd1d8fe0ff765a8d751bc44d03f0019c3b4ac9b18ace073976dbb0240ea1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 212, + 296, + 225 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 212, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 113, + 212, + 296, + 225 + ], + "type": "interline_equation", + "content": "\\mathbf {t} _ {\\log} = \\operatorname {s i g n} \\left(\\mathbf {t} ^ {\\prime}\\right) \\cdot \\log \\left(1 + \\left| \\mathbf {t} ^ {\\prime} \\right|\\right), \\tag {6}", + "image_path": "8f2a8cacd85b420f6dda89cc6154316f2d79ead0bb22ede1041b86bc7f668243.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 220, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 220, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 220, + 243 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 232, + 220, + 243 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{ray}}" + }, + { + "bbox": [ + 55, + 232, + 220, + 243 + ], + "type": "text", + "content": " is a predefined scaling factor." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": "Raymap Construction. Using the intrinsic matrix " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": ", we compute the camera ray directions " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " in homogeneous coordinates for each pixel. Note that we do not unit normalize it but let it have a unit value along " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " axis. The ray origins " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " are set to the translation " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{\\log}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": ", replicated across the spatial dimensions. The raymap in the world coordinate system is obtained by transforming the ray directions " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " using the extrinsic matrix " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": ". The final raymap " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " consists of 6 channels: 3 for the ray directions " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": " and 3 for the ray origins " + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{o}}" + }, + { + "bbox": [ + 55, + 244, + 297, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "type": "text", + "content": "Resolution Downsampling. To align the raymap with the latent feature dimensions from the VAE, we perform adjustments both spatially and temporally. Spatially, the raymap is downsampled by a factor of 8 using bilinear interpolation. Temporally, every consecutive group of 4 frames is concatenated along the channel dimension. The resulting rearranged tensor is denoted as " + }, + { + "bbox": [ + 55, + 352, + 296, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "content": "Converting raymap back to camera matrix. Given generated raymap sequences rearranged by the time axis " + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{r}}\\in" + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times 6\\times h\\times w} = [\\hat{\\mathbf{r}}_{\\mathrm{d}},\\hat{\\mathbf{r}}_{\\mathrm{o}}]" + }, + { + "bbox": [ + 55, + 436, + 296, + 472 + ], + "type": "text", + "content": " , we first recover the ray origins by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 478, + 296, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 478, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 94, + 478, + 296, + 504 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {r} _ {\\mathrm {o}}} ^ {\\prime} = \\frac {1}{s _ {\\text {r a y}}} \\cdot \\operatorname {s i g n} \\left(\\hat {\\mathbf {r} _ {\\mathrm {o}}}\\right) \\cdot \\left(\\exp \\left(\\left| \\hat {\\mathbf {r} _ {\\mathrm {o}}} \\right|\\right) - 1\\right), \\tag {7}", + "image_path": "5e7f019bc02d681487a53e95aab901307ac538234ed6dee3395f7c686eca38c1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 534 + ], + "type": "text", + "content": "Then, we can recover both the intrinsic and extrinsic through Alg. 1 in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 540, + 151, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 151, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 151, + 553 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 557, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 630 + ], + "type": "text", + "content": "We initialize AETHER with pre-trained CogVideo-5b-I2V [77] weights, excluding the additional input and output projection layer channels for depth and raymap action trajectories, which are initialized to zero. Since text prompt conditions are not used, an empty text embedding is provided during both training and inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": "As the dataset we use contains video clips with variable lengths and frames per second (FPS), we randomly select " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "T \\in \\{17, 25, 33, 41\\}" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": " frames, and the FPS is randomly sampled from " + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\{8, 10, 12, 15, 24\\}" + }, + { + "bbox": [ + 55, + 630, + 296, + 689 + ], + "type": "text", + "content": ". The RoPE [59] coefficients are linearly interpolated to align with them." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "During training, conditional inputs are randomly masked to generalize across tasks. For " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_c" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", masking probabilities are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for both observation and goal images (visual planning tasks), " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for observation images only (video prediction), " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for full-color video latents (4D reconstruction), and " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " for masking all of " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_c" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_a" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", trajectory latents are either kept or fully masked with equal probability (supporting action-free or action-conditioned tasks with raymap conditions). This strategy enables the model to adapt to diverse tasks and input condition settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 168, + 556, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 556, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 556, + 323 + ], + "type": "text", + "content": "Our training process consists of two stages. In the first stage, we adopt the loss function of a standard latent diffusion model, which minimizes the mean squared error (MSE) in the latent space. In the second stage, we refine the generated outputs by decoding them into the image space. Specifically, we introduce three additional loss terms: a Multi-Scale Structure Similarity (MS-SSIM) loss [67] for color video, a scale- and shift-invariant loss [49] for depth videos, and a scale- and shift-invariant pointmap loss [66] for pointmaps projected from the generated depths and raymaps. Further details on the stage 2 loss functions are provided in the supplementary material. Notably, the second stage takes about " + }, + { + "bbox": [ + 313, + 168, + 556, + 323 + ], + "type": "inline_equation", + "content": "\\frac{1}{4}" + }, + { + "bbox": [ + 313, + 168, + 556, + 323 + ], + "type": "text", + "content": " of the training steps used in the first stage." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 323, + 556, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 556, + 443 + ], + "type": "text", + "content": "We employ a hybrid training strategy combining Fully Sharded Data Parallel (FSDP) [87] with Zero-2 optimization within compute nodes and Distributed Data Parallel (DDP) across nodes. Since depth videos require online normalization, the VAE encoder is also run online during training and operates under DDP. Our implementation processes a local batch size of 4 per GPU, resulting in an effective batch size of 320 samples across 80 A100-80GB GPUs. Training is conducted over two weeks using the AdamW [38] optimizer with a OneCycle [56] learning rate scheduler." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 453, + 476, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 476, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 476, + 468 + ], + "type": "text", + "content": "4. Reconstruction Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 545 + ], + "type": "text", + "content": "In this section, we demonstrate that AETHER can achieve zero-shot reconstruction metrics comparable to or even better than SOTA reconstruction methods. We mainly consider two zero-shot reconstruction tasks: video depth estimation and camera pose estimation. Note that we only denoise for 4 steps for reconstruction tasks." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 552, + 499, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 499, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 499, + 565 + ], + "type": "text", + "content": "4.1. Zero-Shot Video Depth Estimation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Implementation Details. Video depth estimation is evaluated based on two key aspects: per-frame depth quality and inter-frame depth consistency. These evaluations are performed by aligning the predicted depth maps with the ground truth using a per-sequence scale. We use absolute relative error (Abs Rel) and " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " (percentage of predicted depths within a 1.25-factor of true depth) as metrics. For implementation, we adopt the settings outlined in CUT3R [65]. Our baselines include both reconstruction-based methods—such as DUSt3R [66], MASt3R [37], MonST3R [82], Spann3R [63], and CUT3R [65]—and diffusion-based depth estimators, including ChronoDepth [55], DepthCrafter [29]," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 756, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 756, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 756, + 315, + 765 + ], + "type": "text", + "content": "8539" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 82, + 555, + 234 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 467, + 81 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 467, + 81 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 467, + 81 + ], + "type": "text", + "content": "Table 1. Video depth Evaluation. Methods requiring global alignment are marked \"GA\"." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 82, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 57, + 82, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 57, + 82, + 555, + 234 + ], + "type": "table", + "html": "
MethodSintel [6]BONN [44]KITTI [21]
Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑Abs Rel ↓δ < 1.25 ↑
Reconstruction Methods. Alignment: per-sequence scale
DUSt3R-GA [66]0.65645.20.15583.30.14481.3
MASt3R-GA [37]0.64143.90.25270.10.18374.5
MonST3R-GA [82]0.37855.80.06796.30.16874.4
Spann3R [63]0.62242.60.14481.30.19873.7
CUT3R [65]0.42147.90.07893.70.11888.1
AETHER (Ours)0.32450.20.27359.40.05697.8
Diffusion-Based Methods. Alignment: per-sequence scale&shift
ChronoDepth [55]0.42938.30.31851.80.25254.3
DepthCrafter [29]0.59055.50.25356.30.12486.5
DA-V [74]1.25243.70.45731.10.09493.0
AETHER (Ours)0.31460.40.30860.20.05497.7
", + "image_path": "26844e565e3dd0d2b445af6cdca1e0bac3f64aaf1e3bfa700d646580efb69b42.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 248, + 558, + 402 + ], + "blocks": [ + { + "bbox": [ + 211, + 238, + 398, + 247 + ], + "lines": [ + { + "bbox": [ + 211, + 238, + 398, + 247 + ], + "spans": [ + { + "bbox": [ + 211, + 238, + 398, + 247 + ], + "type": "text", + "content": "Table 2. Evaluation on Camera Pose Estimation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 248, + 558, + 402 + ], + "lines": [ + { + "bbox": [ + 56, + 248, + 558, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 558, + 402 + ], + "type": "table", + "html": "
MethodSintel [6]TUM-dynamics [58]ScanNet [10]
ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓ATE ↓RPE trans ↓RPE rot ↓
Optimization-based Methods
Particle-SfM [86]0.1290.0310.535---0.1360.0230.836
Robust-CVD [36]0.3600.1543.4430.1530.0263.5280.2270.0647.374
CasualSAM [85]0.1410.0350.6150.0710.0101.7120.1580.0341.618
DUSt3R-GA [66]0.4170.2505.7960.0830.0173.5670.0810.0280.784
MASt3R-GA [37]0.1850.0601.4960.0380.0120.4480.0780.0200.475
MonST3R-GA [82]0.1110.0440.8960.0980.0190.9350.0770.0180.529
Feed-forward Methods
DUSt3R [66]0.2900.1327.8690.1400.1063.2860.2460.1088.210
Spann3R [63]0.3290.1104.4710.0560.0210.5910.0960.0230.661
CUT3R [65]0.2130.0660.6210.0460.0150.4730.0990.0220.600
AETHER (Ours)0.1890.0540.6940.0920.0121.1060.1760.0281.204
", + "image_path": "2162ce6860718b2cad0a7b157b188b622065f02bfa6f350b35cf46a3e632a1b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 411, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 579 + ], + "type": "text", + "content": "and DepthAnyVideo (DA-V) [74]. It is important to note that when comparing with diffusion-based depth estimators, we apply scale and shift alignment to the ground truth, as most of these methods are not inherently scale-invariant. All videos are resized with original aspect ratios kept to make the short side align with our model's input size. For videos that exceed the maximum forward processing spatial or temporal size of our model, we employ a sliding window strategy with a stride size of 8. In regions of overlap between windows, we first estimate a relative scale by calculating the average of element-wise division. This relative scale is then used to adjust the latter window's depth predictions. Finally, a linspace-weighted average is applied to the overlapping areas, following approaches similar to prior methods [29, 80]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Results and Analysis. Table 1 summarizes the video depth estimation results across Sintel [6], BONN [44], and KITTI [21] datasets. For reconstruction-based methods, AETHER outperforms or is comparable with prior approaches. On Sintel, AETHER achieves the lowest Abs Rel (0.324), surpassing MonST3R-GA (0.378), and competitive " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " (50.2). On KITTI, AETHER sets a new benchmark with Abs Rel of 0.056 and " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " of 97.8, outperforming the previous SOTA CUT3R (Abs Rel: 0.118, " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": ": 88.1). Among diffusion-based methods, AETHER shows consistent superiority. It achieves the best performance on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "text", + "content": "Sintel (Abs Rel: 0.314, " + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "text", + "content": ": 60.4) and KITTI (Abs Rel: 0.054, " + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "text", + "content": ": 97.7), significantly outperforming ChronoDepth [55], DepthCrafter [29], and DA-V [74]. On BONN, AETHER achieves the highest " + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\delta < 1.25" + }, + { + "bbox": [ + 313, + 410, + 555, + 471 + ], + "type": "text", + "content": " (60.2) with competitive Abs Rel (0.308)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 479, + 503, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 503, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 503, + 491 + ], + "type": "text", + "content": "4.2. Zero-Shot Camera Pose Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. Following MonST3R [82] and CUT3R [65], we evaluate camera pose estimation accuracy on the Sintel [6], TUM Dynamics [58], and ScanNet [10] datasets. Notably, both Sintel and TUM Dynamics contain highly dynamic objects, presenting significant challenges for traditional Structure-from-Motion (SfM) and Simultaneous Localization and Mapping (SLAM) systems. We report Absolute Translation Error (ATE), Relative Translation Error (RPE Trans), and Relative Rotation Error (RPE Rot) after Sim(3) alignment with the ground truth, following the methodology in [65]. The implementation settings are consistent with those used in CUT3R [65]. All videos are resized with original aspect ratios kept and then center cropped to align with our model's input size. For long videos exceeding our model's maximum temporal forward processing length, a sliding window strategy with a stride size of 32 is employed. In overlapping regions between windows, camera poses are aligned following prior methods [64]. Transla" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8540" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "tion alignment is performed using linear interpolation, while quaternion rotations are interpolated with spherical linear interpolation. Additionally, we observed that the generated camera trajectories exhibit noise, likely due to the limited number of denoising steps. To mitigate this, we apply a simple Kalman filter [71] to smooth the trajectories." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 297, + 228 + ], + "type": "text", + "content": "Results and Analysis. Table 2 shows the evaluation results. Among feed-forward methods, AETHER achieves the best ATE (0.189) and RPE Trans (0.054) on Sintel [6], while remaining competitive in RPE Rot (0.694) compared to CUT3R (0.621). On TUM Dynamics [58], AETHER achieves the best RPE Trans (0.012). For other metrics, AETHER is also comparable with other specialist models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 268, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 268, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 268, + 251 + ], + "type": "text", + "content": "5. Generation and Planning Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 297, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 307 + ], + "type": "text", + "content": "In this section, we first show video prediction, with or without action conditioning, quantitatively or qualitatively, in Sec. 5.1. We then show visual planning abilities in Sec. 5.2. More visualizations are in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 313, + 157, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 157, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 157, + 325 + ], + "type": "text", + "content": "5.1. Video Prediction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 297, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 297, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 297, + 641 + ], + "type": "text", + "content": "Implementation Details. We use CogVideoX-5b-I2V [77] as our baseline. To ensure a fair comparison, we construct a validation dataset comprising two subsets: in-domain and out-domain data. The in-domain subset includes novel, unseen scenes from the same synthetic environments as the training dataset, while the out-domain subset consists of data from entirely new synthetic environments. Both models are provided with the first frame as the observation image. For action-free prediction, since CogVideoX depends heavily on text prompts, we use GPT-4o [31] to generate image descriptions and predictions of future scenes as prompts for CogVideoX. In contrast, AETHER is evaluated using empty text prompts. For action-conditioned prediction, we also labeled camera trajectories in the validation dataset and generated corresponding raymap sequences as action conditions for AETHER. For the baseline, in addition to the prompts used for action-free prediction, we use GPT-4o [31] to generate detailed descriptions of object and camera movements, enabling the baseline to use language as action conditions. We use the default classifier-free guidance value of 6 on text prompts for CogVideoX and a value of 3 on the observation image for AETHER. No classifier-free guidance is applied to action conditions to ensure fairness. Evaluation metrics follow VBenchmark [30], a standard benchmark for video generation, with additional details on prompts and evaluation metrics provided in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "Image-to-Video Prediction. We first evaluate image-to-video prediction without action conditions. The results, presented in Tab. 3, show that AETHER consistently outperforms the baseline on both in-domain and out-domain validation sets. Notably, AETHER demonstrates a larger performance improvement on out-domain data, which can likely be at" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "tributed to the baseline model's pre-training data containing domains similar to the in-domain dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 96, + 556, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 556, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 556, + 228 + ], + "type": "text", + "content": "Action-Conditioned Video Prediction. To assess the effectiveness of our post-training in improving action control and action-following capabilities, we conduct action-conditioned video prediction experiments. The results, shown in Tab. 4, indicate that AETHER consistently outperforms the baseline in both in-domain and out-domain settings. Notably, CogVideoX tends to generate static scenes with high visual and aesthetic quality, while AETHER accurately follows the action conditions, producing highly dynamic scenes. These results validate the effectiveness of our framework and the advantages of using camera trajectories as action conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 235, + 411, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 411, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 411, + 248 + ], + "type": "text", + "content": "5.2. Visual Planning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 252, + 556, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 556, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 556, + 418 + ], + "type": "text", + "content": "Implementation Details. We evaluate the action-conditioned navigation capability of AETHER on our validation set. To demonstrate the effectiveness of our multi-task objective, particularly the incorporation of the reconstruction objective, we also post-train an ablation model without the video depth objective, denoted as AETHER-no-depth. Given the observation image, goal image, and camera trajectory, the resulting video should be highly determined. Thus, we report pixel-wise reconstruction metrics, including PSNR, SSIM [68], MS-SSIM [67], and LPIPS [83], for action-conditioned navigation. For the action-free case, which represents a visual path navigation task, we also report the VBench metrics. We do not use any classifier-free guidance on both tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 420, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 556, + 479 + ], + "type": "text", + "content": "Action-Conditioned Navigation. The quantitative results for action-conditioned navigation are presented in Tab. 5. Aether consistently outperforms the ablation model, demonstrating the significant benefits of incorporating the reconstruction objective into generative models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 576 + ], + "type": "text", + "content": "Visual Path Planning. In the absence of action conditions, this task evaluates the model's ability to function as a \"world model as an agent,\" requiring it to plan a path from the observation image to the goal image. The results, shown in Tab. 6, indicate that the reconstruction objective substantially improves the model's visual path planning capability. Additionally, qualitative visualizations on completely in-the-wild data are provided in supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 402, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 402, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 402, + 597 + ], + "type": "text", + "content": "6. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "World Models. World models have emerged as a critical framework in artificial intelligence, enabling agents to simulate, understand, and predict environmental dynamics. Early work [23] introduced latent representations and recurrent neural networks for decision-making. Recent advancements include Cat3D [20] for 3D scene generation, Cat4D [72] for dynamic 4D environments, and Genie 2 [46], a large-scale model for interactive 3D worlds. Motion Prompting [22] further enables precise video generation control. These ad" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 314, + 766 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 314, + 766 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 314, + 766 + ], + "type": "text", + "content": "8541" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 3. VBenchmark [30] Metrics of Video Prediction without Action Conditions. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each group, the better performance is highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 553, + 133 + ], + "type": "table", + "html": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX89.36/84.61/87.7792.72/91.43/92.2998.24/96.93/97.8188.75/95.00/90.8354.49/53.58/54.1855.38/52.29/54.3579.01/77.52/78.51
AETHER91.50/87.55/90.1894.29/93.62/94.0798.54/98.19/98.4296.25/100.00/97.5054.36/52.58/53.7755.08/54.88/55.0180.34/79.42/80.04
", + "image_path": "baacf544f1d57e77dde07620482f80e8fe54b2b5d9aef374fd2dfe13fd6842b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 158, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 136, + 555, + 158 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 555, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 555, + 158 + ], + "type": "text", + "content": "Table 4. VBenchmark [30] Metrics of Action-Conditioned Video Prediction. Comparison between CogVideoX and AETHER (Ours) on in-domain/out-domain/overall performance on the validation set. For each metric group, the better performance is highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 158, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 158, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 553, + 198 + ], + "type": "table", + "html": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
CogVideoX91.56/88.23/90.5192.98/92.29/92.7798.44/97.81/98.2483.87/93.02/86.7656.19/57.43/56.5856.48/61.60/58.1079.56/80.70/79.92
AETHER90.73/93.27/91.5493.61/95.03/94.0698.53/98.62/98.56100.00/83.72/94.8555.04/56.50/55.5053.89/63.23/56.8480.33/81.55/80.71
", + "image_path": "5449ad1792616d2b5a2228f7f3d1e6bc095ca98d86c54af90ba2a8a77b80bae0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 224, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 555, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 223 + ], + "type": "text", + "content": "Table 5. Pixel-wise Metrics of Action-Conditioned Navigation. Comparison of performance between AETHER-no-depth and AETHER on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 224, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 224, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 553, + 264 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑MS-SSIM ↑LPIPS ↓
AETHER-no-depth19.13/18.67/18.970.5630/0.4830/0.53530.5467/0.5204/0.53760.3116/0.2995/0.3074
AETHER19.87/19.37/19.700.5803/0.5058/0.55450.5830/0.5627/0.57600.2691/0.2599/0.2659
", + "image_path": "fa2aa8f6e5fdabf4ba898fb89555e7bb42fc23071b304c87153913193a808015.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 289, + 561, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 289 + ], + "type": "text", + "content": "Table 6. Quantitative Results of Action-Free Visual Path Planning. Comparison of performance between Aether and Aether-no-depth on in-domain/out-domain/overall performance. For each metric group, the better performance is highlighted in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 289, + 561, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 289, + 561, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 561, + 327 + ], + "type": "table", + "html": "
subject consistencyb.g. consistencymotion smoothnessdynamic degreeaesthetic qualityimaging qualityweighted average
Aether-no-depth88.68/89.61/88.6193.62/93.92/93.6698.37/98.31/98.3297.06/91.67/96.1554.12/56.26/54.7851.77/58.46/54.2979.11/80.43/79.59
Aether (Ours)89.69/91.61/90.3693.88/94.58/94.1398.50/98.40/98.4697.06/91.67/95.1955.83/56.87/56.1954.71/61.13/56.9380.21/81.53/80.67
", + "image_path": "2c9bcbfc0c51b0409a14764144dfe6e57906bd8e0c62629001b97846d855b8f5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 342, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 295, + 376 + ], + "type": "text", + "content": "vancements demonstrate the evolution of world models toward dynamic, interactive, and controllable applications in robotics, gaming, and simulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 378, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 604 + ], + "type": "text", + "content": "Reconstruction. Reconstruction has been a long-standing topic in computer vision, with notable progress in both traditional and learning-based methods. Classical approaches, such as Structure-from-Motion (SfM) [9, 26, 45, 53] and Multi-View Stereo (MVS) [19, 54], rely on multi-view geometry for feature matching, pose estimation, and dense point cloud generation, demonstrating robust performance in controlled settings. Deep learning has introduced powerful alternatives, tackling sub-tasks like feature matching [14, 52], point tracking [13, 64], triangulation [42], and MVS [78, 81]. End-to-end methods now directly predict point maps [37, 66] or depth maps from images [4, 76], often incorporating camera parameters [70]. Recently, diffusion models have achieved breakthroughs in image and video generation [27, 35, 43, 73, 77], inspiring novel 3D reconstruction approaches that leverage rich 2D priors [18, 29, 34, 41, 74, 75, 88, 90]. These methods demonstrate the potential of integrating diffusion-based 2D knowledge into 3D modeling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Video Generation. Video generation has evolved from foundational techniques like DDPM [27, 43] to modern frameworks leveraging diffusion-based techniques. Advances such as latent diffusion [51] and diffusion transformers [47] have improved generation quality, while models like Sora [5] and Stable Video Diffusion (SVD) [3] emphasize temporal consistency. Open-source models, including LTX Video [24], CogVideoX [77], and Hunyuan Video [35], offer increased flexibility, and techniques like multi-scale architectures (e.g.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 378 + ], + "type": "text", + "content": "Pyramid Flow [32]) enhance motion dynamics. These advancements highlight rapid progress, with ongoing efforts to improve scalability and temporal stability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 471, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 471, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 471, + 401 + ], + "type": "text", + "content": "7. Conclusion and Limitations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 408, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 408, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 408, + 556, + 552 + ], + "type": "text", + "content": "In this work, we introduce AETHER, a geometry-aware multitask world model that reconstructs 4D dynamic videos, predicts future frames conditioned on observation images and actions, and performs visual planning based on observation and goal images. We propose an automatic 4D synthetic data labeling pipeline, enabling AETHER to train on synthetic data and generalize to unseen real-world data in a zero-shot manner. Post-trained on the CogVideoX base model, AETHER achieves state-of-the-art or competitive reconstruction performance and outperforms baselines in generation and planning tasks, demonstrating the value of incorporating reconstruction objectives into world modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 552, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 552, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 312, + 552, + 556, + 672 + ], + "type": "text", + "content": "However, limitations remain. Camera pose estimation is less accurate, likely due to incompatibilities between raymap representation and prior video diffusion models. Indoor scene reconstruction also lags behind outdoor performance, likely due to the predominance of outdoor training data. Additionally, predictions without language prompts often fail in highly dynamic scenes. Future work can address these by exploring novel action representations, co-training with real-world data, and retaining the language prompting capabilities of the base model." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8542" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work is supported by the National Key R&D Program of China (NO.2022ZD0160102) and Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 181 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 181 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Keir Mierle, et al. Ceres solver: Tutorial & reference. Google Inc, 2(72):8, 2012. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 182, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 296, + 237 + ], + "type": "text", + "content": "[2] Eloi Alonso, Adam Jolley, Vincent Micheli, Anssi Kanervisto, Amos J Storkey, Tim Pearce, and François Fleuret. Diffusion for world modeling: Visual details matter in atari. Advances in Neural Information Processing Systems, 37:58757-58791, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 239, + 296, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 294 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 339 + ], + "type": "text", + "content": "[4] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "text", + "content": "[5] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 296, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 451 + ], + "type": "text", + "content": "[6] Daniel J Butler, Jonas Wulff, Garrett B Stanley, and Michael J Black. A naturalistic open source movie for optical flow evaluation. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 611-625. Springer, 2012. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 486 + ], + "type": "text", + "content": "[7] Junyi Chen, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Where am i and what will i see: An auto-regressive model for spatial localization and view prediction, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "text", + "content": "[8] Yuhua Chen, Cordelia Schmid, and Cristian Sminchisescu. Self-supervised learning with geometric constraints in monocular video: Connecting flow, depth, and camera. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7063-7072, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 296, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 587 + ], + "type": "text", + "content": "[9] Hainan Cui, Xiang Gao, Shuhan Shen, and Zhanyi Hu. Hsfm: Hybrid structure-from-motion. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1212-1221, 2017. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 678 + ], + "type": "text", + "content": "[11] Decart, Julian Quevedo, Quinn McIntyre, Spruce Campbell, Xinlei Chen, and Robert Wachen. Oasis: A universe in a Transformer. 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition workshops, pages 224-236, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[13] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Recasens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. Advances in Neural Information Processing Systems, 35:13610-13626, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "type": "text", + "content": "[14] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 227 + ], + "type": "text", + "content": "[15] Tom Erez, Yuval Tassa, and Emanuel Todorov. Infinite-horizon model predictive control for periodic tasks with contacts. Robotics: Science and systems VII, 73, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 554, + 271 + ], + "type": "text", + "content": "[16] Hao-Shu Fang, Hongjie Fang, Zhenyu Tang, Jirong Liu, Chenxi Wang, Junbo Wang, Haoyi Zhu, and Cewu Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot. arXiv preprint arXiv:2307.00595, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 326 + ], + "type": "text", + "content": "[17] Ruili Feng, Han Zhang, Zhantao Yang, Jie Xiao, Zhilei Shu, Zhiheng Liu, Andy Zheng, Yukun Huang, Yu Liu, and Hongyang Zhang. The matrix: Infinite-horizon world generation with real-time moving control. arXiv preprint arXiv:2412.03568, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 554, + 382 + ], + "type": "text", + "content": "[18] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. In European Conference on Computer Vision, pages 241-258. Springer, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 415 + ], + "type": "text", + "content": "[19] Yasutaka Furukawa, Carlos Hernández, et al. Multi-view stereo: A tutorial. Foundations and trends® in Computer Graphics and Vision, 9(1-2):1-148, 2015. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 469 + ], + "type": "text", + "content": "[20] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 514 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The international journal of robotics research, 32(11):1231-1237, 2013. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 515, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 555, + 579 + ], + "type": "text", + "content": "[22] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, Chen Sun, Oliver Wang, Andrew Owens, and Deqing Sun. Motion prompting: Controlling video generation with motion trajectories, 2024. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 581, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 554, + 602 + ], + "type": "text", + "content": "[23] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 604, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 669 + ], + "type": "text", + "content": "[24] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. Ltx-video: Realtime video latent diffusion. arXiv preprint arXiv:2501.00103, 2024. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[25] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8543" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "text", + "content": "[26] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 4, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 296, + 184 + ], + "type": "text", + "content": "[28] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 187, + 296, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 296, + 231 + ], + "type": "text", + "content": "[29] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 298 + ], + "type": "text", + "content": "[30] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 296, + 343 + ], + "type": "text", + "content": "[31] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 346, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 346, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 346, + 296, + 399 + ], + "type": "text", + "content": "[32] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 402, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 402, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 402, + 296, + 455 + ], + "type": "text", + "content": "[33] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv preprint arXiv:2410.11831, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 458, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 458, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 458, + 296, + 523 + ], + "type": "text", + "content": "[34] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 667 + ], + "type": "text", + "content": "[35] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Junkun Yuan, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yanxin Long, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Dax Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. Hunyuanvideo: A systematic framework for large video generative models, 2024. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[36] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1611-1621, 2021. 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[37] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 118, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 140 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 142, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 555, + 174 + ], + "type": "text", + "content": "[39] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 554, + 219 + ], + "type": "text", + "content": "[40] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 555, + 265 + ], + "type": "text", + "content": "[41] Yuanxun Lu, Jingyang Zhang, Tian Fang, Jean-Daniel Nahmias, Yanghai Tsin, Long Quan, Xun Cao, Yao Yao, and Shiwei Li. Matrix3d: Large photogrammetry model all-in-one. arXiv preprint arXiv:2502.07685, 2025. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 266, + 555, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 555, + 320 + ], + "type": "text", + "content": "[42] Dror Moran, Hodaya Koslowsky, Yoni Kasten, Haggai Maron, Meirav Galun, and Ronen Basri. Deep permutation equivariant structure from motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5976-5986, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 323, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 323, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 555, + 365 + ], + "type": "text", + "content": "[43] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 368, + 555, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 368, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 368, + 555, + 422 + ], + "type": "text", + "content": "[44] Emanuele Palazzolo, Jens Behley, Philipp Lottes, Philippe Giguere, and Cyril Stachniss. Refusion: 3d reconstruction in dynamic environments for rgb-d cameras exploiting residuals. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7855-7862. IEEE, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 424, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 424, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 424, + 555, + 466 + ], + "type": "text", + "content": "[45] Linfei Pan, Daniel Barath, Marc Pollefeys, and Johannes L Schonberger. Global structure-from-motion revisited. In European Conference on Computer Vision, pages 58-77. Springer, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 469, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 469, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 469, + 555, + 578 + ], + "type": "text", + "content": "[46] Jack Parker-Holder, Philip Ball, Jake Bruce, Vibhavari Dasagi, Kristian Holsheimer, Christos Kaplanis, Alexandre Moufarek, Guy Scully, Jeremy Shar, Jimmy Shi, Stephen Spencer, Jessica Yung, Michael Dennis, Sultan Kenjeyev, Shangbang Long, Vlad Mnih, Harris Chan, Maxime Gazeau, Bonnie Li, Fabio Pardo, Luyu Wang, Lei Zhang, Frederic Besse, Tim Harley, Anna Mitenkova, Jane Wang, Jeff Clune, Demis Hassabis, Raia Hadsell, Adrian Bolton, Satinder Singh, and Tim Roktaschel. Genie 2: A large-scale foundation world model. 2024. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "text", + "content": "[47] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 713 + ], + "type": "text", + "content": "[48] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly," + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8544" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 297, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 297, + 247 + ], + "type": "text", + "content": "Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Eliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models, 2025. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 296, + 304 + ], + "type": "text", + "content": "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 306, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 296, + 361 + ], + "type": "text", + "content": "[50] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. Grounded sam: Assembling open-world models for diverse visual tasks, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 363, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 363, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 296, + 418 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 419, + 296, + 474 + ], + "type": "text", + "content": "[52] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 475, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 520 + ], + "type": "text", + "content": "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 521, + 296, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 296, + 587 + ], + "type": "text", + "content": "[54] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part III 14, pages 501-518. Springer, 2016. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 589, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 296, + 643 + ], + "type": "text", + "content": "[55] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Vitor Guizilini, Yue Wang, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors. arXiv preprint arXiv:2406.01493, 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 646, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 296, + 689 + ], + "type": "text", + "content": "[56] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[57] Ziyang Song, Zerong Wang, Bo Li, Hao Zhang, Ruijie Zhu, Li Liu, Peng-Tao Jiang, and Tianzhu Zhang. Depthmaster:" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Taming diffusion models for monocular depth estimation. arXiv preprint arXiv:2501.02576, 2025.4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 148 + ], + "type": "text", + "content": "[58] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In 2012 IEEE/RSJ international conference on intelligent robots and systems, pages 573-580. IEEE, 2012. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "text", + "content": "[59] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 195, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 216 + ], + "type": "text", + "content": "[60] Wan Team. Wan: Open and advanced large-scale video generative models. 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 270 + ], + "type": "text", + "content": "[61] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 272, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 414 + ], + "type": "text", + "content": "[62] Veo-Team,.; Agrim Gupta, Ali Razavi, Andeep Toor, Ankush Gupta, Dumitru Erhan, Eleni Shaw, Eric Lau, Frank Belletti, Gabe Barth-Maron, Gregory Shaw, Hakan Erdogan, Hakim Sidahmed, Henna Nandwani, Hernan Moraldo, Hyunjik Kim, Irina Blok, Jeff Donahue, José Lezama, Kory Mathewson, Kurtis David, Matthieu Kim Lorraine, Marc van Zee, Medhini Narasimhan, Miaosen Wang, Mohammad Babaeizadeh, Nelly Papalampidi, Nick Pezzotti, Nilpa Jha, Parker Barnes, Pieter-Jan Kindermans, Rachel Hornung, Ruben Villegas, Ryan Poplin, Salah Zaiem, Sander Dieleman, Sayna Ebrahimi, Scott Wisdom, Serena Zhang, Shlomi Fruchter, Signe Nørly, Weizhe Hua, Xinchen Yan, Yuqing Du, and Yutian Chen. Veo 2. 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 446 + ], + "type": "text", + "content": "[63] Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory. arXiv preprint arXiv:2408.16061, 2024. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 502 + ], + "type": "text", + "content": "[64] Jianyuan Wang, Nikita Karaev, Christian Rupprecht, and David Novotny. Vggsfm: Visual geometry grounded deep structure from motion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21686-21697, 2024. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 547 + ], + "type": "text", + "content": "[65] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. arXiv preprint arXiv:2501.12387, 2025. 2, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "text", + "content": "[66] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2, 5, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "type": "text", + "content": "[67] Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In The Thirty-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398-1402. IEEE, 2003. 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[68] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[69] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Mo" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 315, + 765 + ], + "type": "text", + "content": "8545" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "tionctrl: A unified and flexible motion controller for video generation. 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 150 + ], + "type": "text", + "content": "[70] Xingkui Wei, Yinda Zhang, Zhuwen Li, Yanwei Fu, and Xiangyang Xue. Deepsfm: Structure from motion via deep bundle adjustment. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I 16, pages 230-247. Springer, 2020. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 171 + ], + "type": "text", + "content": "[71] Greg Welch, Gary Bishop, et al. An introduction to the kalman filter. 1995. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 295, + 216 + ], + "type": "text", + "content": "[72] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 261 + ], + "type": "text", + "content": "[73] Wanghan Xu, Xiaoyu Yue, Zidong Wang, Yao Teng, Wenlong Zhang, Xihui Liu, Luping Zhou, Wanli Ouyang, and Lei Bai. Exploring representation-aligned latent space for better generation. arXiv preprint arXiv:2502.00359, 2025. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[74] Honghui Yang, Di Huang, Wei Yin, Chunhua Shen, Haifeng Liu, Xiaofei He, Binbin Lin, Wanli Ouyang, and Tong He. Depth any video with scalable synthetic data. arXiv preprint arXiv:2410.10815, 2024. 2, 4, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 371 + ], + "type": "text", + "content": "[75] Honghui Yang, Sha Zhang, Di Huang, Xiaoyang Wu, Haoyi Zhu, Tong He, Shixiang Tang, Hengshuang Zhao, Qibo Qiu, Binbin Lin, et al. Unipad: A universal pre-training paradigm for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15238-15250, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 426 + ], + "type": "text", + "content": "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[77] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "text", + "content": "[78] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 558 + ], + "type": "text", + "content": "[79] Jiwen Yu, Yiran Qin, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Gamefactory: Creating new games with generative interactive videos, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 612 + ], + "type": "text", + "content": "[80] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 657 + ], + "type": "text", + "content": "[81] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1):199-214, 2023. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[82] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 2, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 485 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 127 + ], + "type": "text", + "content": "[83] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 128, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 555, + 173 + ], + "type": "text", + "content": "[84] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimic motion: High-quality human motion video generation with confidence-aware pose guidance, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 217 + ], + "type": "text", + "content": "[85] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 218, + 555, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 555, + 272 + ], + "type": "text", + "content": "[86] Wang Zhao, Shaohui Liu, Hengkai Guo, Wenping Wang, and Yong-Jin Liu. *Particlesfm: Exploiting dense point trajectories for localizing moving cameras in the wild*. In European Conference on Computer Vision, pages 523-542. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 328 + ], + "type": "text", + "content": "[87] Yanli Zhao, Andrew Gu, Rohan Varma, Liang Luo, Chien-Chin Huang, Min Xu, Less Wright, Hamid Shojanazeri, Myle Ott, Sam Shleifer, et al. Pytorch fsdp: experiences on scaling fully sharded data parallel. arXiv preprint arXiv:2304.11277, 2023.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 555, + 385 + ], + "type": "text", + "content": "[88] Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Hengshuang Zhao, Chunhua Shen, Yu Qiao, Tong He, et al. Ponderv2: Pave the way for 3d foundation model with a universal pre-training paradigm. arXiv preprint arXiv:2310.08586, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 385, + 555, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 385, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 385, + 555, + 440 + ], + "type": "text", + "content": "[89] Haoyi Zhu, Yating Wang, Di Huang, Weicai Ye, Wanli Ouyang, and Tong He. Point cloud matters: Rethinking the impact of different observation spaces on robot learning. Advances in Neural Information Processing Systems, 37: 77799-77830, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 555, + 485 + ], + "type": "text", + "content": "[90] Haoyi Zhu, Honghui Yang, Yating Wang, Jiange Yang, Limin Wang, and Tong He. Spa: 3d spatial-awareness enables effective embodied representation. arXiv preprint arXiv:2410.08208, 2024. 2, 8" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 757, + 316, + 765 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 757, + 316, + 765 + ], + "spans": [ + { + "bbox": [ + 295, + 757, + 316, + 765 + ], + "type": "text", + "content": "8546" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_content_list.json b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f4a5db15ae9b7590582dd142150d8d2946633c35 --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_content_list.json @@ -0,0 +1,2078 @@ +[ + { + "type": "text", + "text": "AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi-Lin Wei\\*, Mu Lin\\*, Yuhao Lin\\*, Jian-Jian Jiang\\*, Xiao-Ming Wu\\*, Ling-An Zeng\\*, Wei-Shi Zheng†,2", + "bbox": [ + 238, + 202, + 756, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 School of Computer Science and Engineering, Sun Yat-sen University, China", + "bbox": [ + 186, + 238, + 808, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China", + "bbox": [ + 106, + 256, + 888, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{weiylin5, linm67, linyh96, jiangjj35, wuxm65, zenglan3}@mail2.sysu.edu.cn wszheng@ieee.org", + "bbox": [ + 133, + 277, + 861, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://isee-laboratory.github.io/AffordDexGrasp/", + "bbox": [ + 299, + 292, + 691, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/851ee5ad1020d907eb07085e67cd7c7004423b3e13caab0a47d222611228b900.jpg", + "image_caption": [ + "Training Data", + "Use the teapot with grasping the handle from the right." + ], + "image_footnote": [], + "bbox": [ + 96, + 340, + 191, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0b67ec285c8d3e430274e6f693d2a855cc9ed20761f1e39c26ff3f918963574.jpg", + "image_caption": [ + "Generalizable-Instructive Affordance", + "Use the mug by contacting the handle from the right." + ], + "image_footnote": [], + "bbox": [ + 197, + 340, + 393, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dadea84cdf4e1e584861d2ed6c439c2ffb3f7b23822cb36b8cfebabbca47d01f.jpg", + "image_caption": [ + "Unseen Category" + ], + "image_footnote": [], + "bbox": [ + 398, + 340, + 493, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the handle from the right.", + "bbox": [ + 398, + 396, + 491, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d1284c5a41b999c1fa0d9a876aa7b67a5155114d740146f66d70f82efe6f90f.jpg", + "image_caption": [ + "Use the mug by contacting the handle from the right." + ], + "image_footnote": [], + "bbox": [ + 503, + 340, + 596, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62f0115f32b1dadb8a267bd82ddab83bc8d8d235dbc614cccd7d20f96c3af95e.jpg", + "image_caption": [ + "Real World Open set" + ], + "image_footnote": [], + "bbox": [ + 604, + 340, + 697, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the handle from the right.", + "bbox": [ + 604, + 396, + 694, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36f051009c4c808f8bee163ea1020aa4c64b2b3944f830f5be3cd900860fb987.jpg", + "image_caption": [ + "Use the kettle by contacting the handle from the left." + ], + "image_footnote": [], + "bbox": [ + 705, + 340, + 800, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62c1e933b3455da5c5fb91bdeb551d79a96a0cd29356b6750e0e51a6177a730a.jpg", + "image_caption": [ + "the handle from the left." + ], + "image_footnote": [], + "bbox": [ + 805, + 340, + 900, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/016528dfbbad3517223b8ddcfa626f092ec44f3cdece38a5012b582eaafdab33.jpg", + "image_caption": [ + "From the right, grasp the handle to use the knife." + ], + "image_footnote": [], + "bbox": [ + 96, + 412, + 191, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f96d2bf8632d5929131b44065824020067f63d86cd6fd2d35d5e756d28759f1b.jpg", + "image_caption": [ + "Use the frying pan and grasp handle from the right." + ], + "image_footnote": [], + "bbox": [ + 196, + 412, + 393, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/abeb627b88a11ef8decd7dcea31fc4220c6d6cb0e5a618971168f67889b8d04f.jpg", + "image_caption": [ + "Use the hammer and grasp handle from the back." + ], + "image_footnote": [], + "bbox": [ + 503, + 412, + 596, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36b2889101237f181baaf44b132872e00bfe07dd6ab466ba51793e24018835bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 412, + 697, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$p$ handle from the back.", + "bbox": [ + 604, + 470, + 689, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0c565a626041c2e0fc3dfe39d567661895d7c1f3da795b65a45c924ea50d405.jpg", + "image_caption": [ + "Use the frying pan and grasp handle from the back." + ], + "image_footnote": [], + "bbox": [ + 705, + 412, + 800, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a498eae91c628b0323958faa85ef6be48b285cbbb1dbacc41c260b5c08b20889.jpg", + "image_caption": [ + "Hold the bottle body from the left.", + "Figure 1. Open-set Language-guided Dexterous Grasp. Our framework bridges the gap between language and grasp actions through Generalizable-Instructive Affordance, which enables cross-category generalization via category-agnostic cues and graspable local structure. Remarkably, our framework demonstrates strong generalization without requiring extra real training data in real-world experiments." + ], + "image_footnote": [], + "bbox": [ + 96, + 488, + 191, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0dde91e9c69ce33f9ced60433b0c46507f840a2e36a30cf285f1bc9486980b59.jpg", + "image_caption": [ + "Hold the body of trigger sprayer from the left." + ], + "image_footnote": [], + "bbox": [ + 197, + 489, + 393, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b6e6acbc46cc46a2d511f8e9035eb56c0b873c5d846abdef490571e9b457a28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 488, + 493, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af28b5de35957a31e3e694d3944b8b899dd3b8ff815f56d30bc6614c0b3a1141.jpg", + "image_caption": [ + "Hold the body of trigger sprayer from the left." + ], + "image_footnote": [], + "bbox": [ + 503, + 488, + 596, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ea3f3f8709b245ac99e621e6ccc04b7b29b8a7afa3aa85dd76d55a88114bcec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 488, + 697, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sprayer from the left.", + "bbox": [ + 604, + 544, + 681, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b31d3b61d8aadbe3b900b8483b8f9a14ce5a244ad45f3adc4de9e2c7b882131a.jpg", + "image_caption": [ + "Hold the body of the lotion pump from the left." + ], + "image_footnote": [], + "bbox": [ + 705, + 488, + 800, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 619, + 326, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language-guided robot dexterous generation enables robots to grasp and manipulate objects based on human commands. However, previous data-driven methods are hard to understand intention and execute grasping with unseen categories in the open set. In this work, we explore a new task, Open-set Language-guided Dexterous Grasp, and find that the main challenge is the huge gap between high-level human language semantics and low-level robot actions. To solve this problem, we propose an Affordance Dexterous Grasp (AffordDexGrasp) framework, with the insight of bridging the gap with a new generalizable-instructive affordance representation. This affordance can generalize to unseen categories by leveraging the object's local structure and category-agnostic semantic attributes, thereby ef", + "bbox": [ + 88, + 652, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fectively guiding dexterous grasp generation. Built upon the affordance, our framework introduces Affordance Flow Matching (AFM) for affordance generation with language as input, and Grasp Flow Matching (GFM) for generating dexterous grasp with affordance as input. To evaluate our framework, we build an open-set table-top language-guided dexterous grasp dataset. Extensive experiments in the simulation and real worlds show that our framework surpasses all previous methods in open-set generalization.", + "bbox": [ + 511, + 622, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 785, + 643, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Achieving generalizable robot dexterous grasping is an important goal in the fields of robotics and computer vision, with exciting potential applications in human-robot interaction and robot manipulation.", + "bbox": [ + 511, + 811, + 906, + 871 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent works explore the task of language-guided dexterous grasp generation[20, 21, 51, 59], aiming to enable", + "bbox": [ + 511, + 872, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 876, + 220, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 888, + 235, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dexterous hands to perform actions based on language instructions, going beyond previous works that focus on stable grasping [26, 38, 46, 54]. The typical language-guided approaches employ language as the condition of generative models to predict hand parameters, achieving impressive performance on objects within known categories [4, 51]. However, in the open real world, there are many categories that may not appear during training, and the cost of collecting data for dexterous hands is quite expensive. Therefore, the open-set generalization on unseen category samples is crucial for robot grasp. While previous works on parallel grippers explore open-set tasks [23, 36], there is limited research on open-set generalization for dexterous hands, as their significantly higher degrees of freedom present complex challenges.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore a novel and challenging task: open-set language-guided dexterous grasping, as shown in Figure 1, where models are evaluated on objects and language instructions from both seen and unseen categories. This task poses a great challenge in ensuring that grasps are intention consistent with corresponding language instructions for unseen categories. We find that the main challenge lies in the huge gap between high-level natural language and low-level robot action spaces, which makes it difficult to generalize the ability of understanding intentions and grasping from the training domain to unseen categories.", + "bbox": [ + 89, + 319, + 480, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the above challenge, we propose the Affordance Dexterous Grasp (AffordDexGrasp) framework, with the key insight of employing a new affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. Our affordance representation combines two characteristics: generalizable (generalizing to unseen categories based on language) and instructive (guiding dexterous grasp generation effectively). However, achieving these two characteristics is not trivial. For example, as shown in Figure 2, fine-grained contact information can effectively guide grasp generation or optimization [2, 9, 17, 55], but it is difficult to generalize to unseen categories. In contrast, coarse information, such as object parts, can be obtained from a pre-trained computer vision model [27, 41, 43], but it is too coarse to guide dexterous hand actions with higher degrees of freedom.", + "bbox": [ + 89, + 488, + 482, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve these characteristics, we propose the Generalizable-Instructive Affordance by defining a general dexterous affordance that aligns with category-independent information, such as intention, object parts, and direction. As shown in Figure 2, the affordance represents the potential graspable regions of all grasps with the same semantics. In this way, the models do not need to learn complex dexterous contact patterns but instead focus on a general graspable area, which can be well aligned with category-agnostic semantic attributes and guide grasp generation effectively.", + "bbox": [ + 89, + 732, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate affordance and employ it to guide grasp gen", + "bbox": [ + 109, + 885, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eration, our framework consists of two cascaded generative models. The Affordance Flow Matching generates affordance maps in a generalizable manner based on language. And the Grasp Flow Matching generates dexterous grasp poses under the effective guidance of affordance. Moreover, we introduce a pre-understanding stage and a post-optimization stage to further boost generalization. Specifically, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intention to enhance generalization across diverse user commands. And we introduce an affordance-guided optimization to improve grasp quality while preserving consistency with user intentions.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For evaluating our framework, we build a open-set tabletop language-guided dexterous grasp dataset, based on the language guided dexterous grasp dataset [51, 56]. We exclude specific categories from the training set to test the model's open-set generalization. Moreover, we provide high-quality rendered images to facilitate the usage by MLLM, and we extend the dataset to scene-level data to better simulate real-world environments. The comprehensive experiments are conducted on both the simulation open-set dataset and real-world environments. The results show that our framework can generate dexterous grasp with consistent intention and high quality in open set.", + "bbox": [ + 511, + 273, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 470, + 651, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dexterous Grasp", + "text_level": 1, + "bbox": [ + 511, + 494, + 679, + 511 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dexterous grasping is critical in robotics, which equips robots with human-like grasping capabilities. Some methods [29, 44, 48, 58] focus on grasp stability and quality, while recent studies explore task-oriented [50, 61] or language-guided grasping [21, 51] with specific semantic intention. For language-guided task, the data-driven methods achieve impressive performance in unseen object within seen categories by leveraging language conditioned generative model [4, 51]. However, we find that these models are hard to generalize for samples from unseen categories. And we find that the main challenge is the huge gap between language and grasp action. To solve this problems, we propose AffordDexGrasp framework with the insight that bridge the gap by generalizable-instructive affordance.", + "bbox": [ + 511, + 518, + 903, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Open-set Robot Grasp", + "text_level": 1, + "bbox": [ + 511, + 742, + 723, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Exploring the performance of robotic models in open-set scenarios is crucial due to the diverse object categories in the real world and the high cost of data collection. For tasks that only focus on stable grasping, both parallel grippers [6, 52] and dexterous hands [47, 58] achieve impressive performance on open set. However, when it comes to considering task-oriented [45] or language-guided grasping [33], it becomes much more difficult. Some works in parallel grippers achieve this by using pre-trained visual ground", + "bbox": [ + 511, + 763, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11819", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9193775e0a6e1d7738d4c13e90f3d49834d3a841ac65775a8e753b31ef9a2707.jpg", + "image_caption": [ + "(a) Visualization of different affordance representations." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 566, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/011b755e22e552f5c4dc0863e4f757f92db30d9628035918f7dc6f7ea0fa733d.jpg", + "image_caption": [ + "(b) Quantitative Results Across Affordances.", + "Figure 2. Different affordance representations. (a) While contact map are too elaborate to generalize and object part are too coarse to guide grasping, our affrodacne achieve a balance. (b) While object part has a lower upper bound and contact shows significant degradation in generalization, only our affordance effectively achieve the balance (Top-1 indicates grasp intention consistency)." + ], + "image_footnote": [], + "bbox": [ + 571, + 98, + 895, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing models [22, 27, 41], implicit feature semantic fields [18, 32, 36], or multimodal large models [13, 14]. Compared to parallel grippers, dexterous hands have a significantly higher number of degrees of freedom, making the methods of parallel grippers difficult to transfer to dexterous hand [51]. In this paper, we explore the open-set language-guided dexterous grasping task and propose a novel framework to address it.", + "bbox": [ + 88, + 369, + 483, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Grasp Affordance", + "text_level": 1, + "bbox": [ + 89, + 503, + 266, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Affordance is first introduced in [8], referring to environmental action possibilities. There are currently two types of grasp affordance: fine-grained contact [3, 42] and coarse-grained object segmentation [27]. The fine-grained information, such as contact areas, normals, or key points, is commonly used in hand-object interaction[2, 9, 59, 62]. However, this fine-grained information is difficult to generalize well to unseen categories. As a result, inaccurate contact information would lead to unreasonable grasping. On the other hand, We find through experiments that the coarse object segmentation [27] is too coarse to guide dexterous grasp generation with relatively high degrees of freedom. To solve this problem, we propose Generalizable-Instructive Affordance, which can generalize through the local structure of objects using category-independent clues and effectively guide grasp generation.", + "bbox": [ + 91, + 525, + 483, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Affordance Dexterous Grasp Framework", + "text_level": 1, + "bbox": [ + 89, + 782, + 460, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the scene point cloud $\\mathcal{O}$ , RGB images $\\mathcal{I}$ and language command $\\mathcal{C}$ as input, our goal is to generate intention aligned and high-quality dexterous grasps poses $\\mathcal{G}^{dex} = (r,t,q)$ , where $r$ denotes the rotation, $t$ denotes the translation, and $q$ represents the joint angles of the dexterous hand.", + "bbox": [ + 89, + 809, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce a novel generalizable-", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "instructive affordance representation (Section 3.1). Built on this affordance representation, we then propose the Affordance Dexterous Grasp framework (Section 3.2) including intention pre-understanding stage, Affordance and Grasp Flow Matching, and grasp post-optimization stage.", + "bbox": [ + 511, + 369, + 906, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generalizable-Instructive Affordance", + "text_level": 1, + "bbox": [ + 511, + 455, + 836, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the generalizable-instructive affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. The affordances serve two primary objectives: 1) enable generalization by aligning object local structures with category-agnostic language semantics; and 2) provide instructive guidance for grasp generation through object affordance cues. However, there exists a trade-off in the generalization and instruction, and it is not trivial to achieve both objectives.", + "bbox": [ + 511, + 477, + 906, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generalization vs. Instruction. While more fine-grained contact information can better guide grasp generation, it becomes harder to generalize to unseen categories. Fine-grained affordances typically compute the distance between object points and predefined elements, such as hand surface points or hand key points [19, 28]. However, we observe that these fine-gained representations are difficult to generalize to unseen category samples as shown in Figure 2. As the dexterous hand structure is complex and contact modes are varied, the contact maps are also elaborate and varies in space, which make it difficult to generalize.", + "bbox": [ + 511, + 613, + 908, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the other hand, the information agnostic to dexterous hands, like object part segmentation, can be obtained in a generalizable manner by vision models [27, 60] or semantic feature fields [36]. However, the object part is too coarse for dexterous hand, as it may struggle to define the clear grasping area. For example, grasping the body of a mug from above and from the side involves the same part, but with different intentions. Moreover, this coarse-grained", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11820", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ce3bdadeea1d70e692f9a6fba428df19dc30ba03ff06e911153ef684b23341f4.jpg", + "image_caption": [ + "Figure 3. The pipeline of Affordance Dexterous Grasp framework. The inference pipeline includes three stages: 1) intention pre- understanding assisted by MLLM; 2) affordance flow matching for generating affordance base on MLLM output; 3) Grasp Flow Matching and Optimization for outputing grasp poses based on the affordance and MLLM outputs. In the training time, AFM and GFM are independently trained one after another. Transformer and Perceiver are attention-based interaction module for velocity vector field prediction." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learning may hinder the model from learning an accurate grasping method during training.", + "bbox": [ + 88, + 419, + 480, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generalizable-Instructive Affordance. To address this challenge, we propose a novel affordance representation that simultaneously enables both generalization and instruction. The key insight is to establish a correspondence between the general graspable affordance of objects and category-agnostic semantic attributes. Unlike contact maps that describe a particular hand contact area, our affordance represents all potential graspable regions that share the same semantic intention. This approach allows the affordance to generalize more effectively with category-agnostic semantic attributes from language, as it doesn't have to learn the complex distribution of hand contact. Meanwhile, our affordance serves as priors to enhance grasp generation by providing valuable graspable area cues.", + "bbox": [ + 88, + 450, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the ground truth of generalizable-instructive affordance, we first organize those grasps of one object with the same intention, contact parts and grasp direction into a semantic group. Then we calculate the distances between each scene points $\\{o_i\\}^M$ and hand surface points $\\{h_i\\}^N$ , to obtain the contact map $\\Omega_{contact} = \\{d_i\\}_1^M$ for each grasp in the group, where $d_i = \\min_{j=1\\dots N}\\| o_i - h_j\\|_2$ . Then we unite contact areas of $k$ grasps in each group by calculating the minimum value of contact areas: $d_i^{min} = \\{\\min_k(d_i^k)\\}^M$ . In order to make the affordance map more smooth, we apply Gaussian filtering to the union contact map: $a_i = \\sum_{j\\in \\mathcal{N}(i)}\\hat{w}_{ij}d_j^{min}$ and the weights are obtained by:", + "bbox": [ + 89, + 661, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} _ {i j} = \\frac {\\exp \\left(- \\| o _ {i} - o _ {j} \\| ^ {2} / (2 \\sigma^ {2})\\right)}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\left(- \\| o _ {i} - o _ {k} \\| ^ {2} / (2 \\sigma^ {2})\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 866, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(i)$ is the neighborhood of $o_i$ , $\\sigma$ is set as the average nearest neighbor distance. Finally, we obtain the ground truth of generalizable-instructive affordance $\\Omega = \\{a_i\\}^M$ .", + "bbox": [ + 511, + 419, + 906, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. AffordDexGrasp Framework", + "text_level": 1, + "bbox": [ + 511, + 478, + 774, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Intention Pre-Understanding", + "text_level": 1, + "bbox": [ + 511, + 502, + 761, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The variability of user expression in the open world presents a challenge for models in generalizing and understanding user intentions. To address this, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intentions, by imputing the user language and the rendered RGB image into GPT-4o [15] for intention understanding and key cues extraction. We empirically find that the following cues are both helpful for grasping and easily reasoned by the MLLM: the object category, user intention (e.g., use it), the contacting part of the object (e.g., the handle of a mug), and the grasp direction. The key cues can be extract from language when the commands are clear, otherwise can be reasoned from the image.", + "bbox": [ + 511, + 521, + 906, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To improve the understanding ability of MLLM, we employ chain of thought (CoT) and visual prompt, and the prompts can be found in supplementary material A.4. Additionally, the grasp direction is simplified as a discretized direction in six predefined coordinates (front, back, left, right, up, down). We first ask MLLM to obtain the direction in image coordinates, and transform it to world coordinates by the camera pose. Then the direction vector is derived from the index of the nearest coordinate axis to form discretized direction. Finally, MLLM organizes this information into a concise sentence (e.g., \"use the mug from the left by contacting the handle\"). This compact and information-dense", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentence structure allows subsequent models to better understand users' intentions.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Affordance Flow Matching", + "text_level": 1, + "bbox": [ + 89, + 128, + 326, + 145 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Affordance Flow Matching (AFM) to learn the affordance distribution efficiently, achieving intention aligned and generalizable affordance generation. AFM is built on the conditional flow matching model [1, 25, 30] aiming to learn the velocity vector field from the noised affordance map $\\Omega_0$ to the target affordance map $\\Omega_1$ , and the objective loss of AFM could be:", + "bbox": [ + 89, + 148, + 483, + 255 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {A F M} = \\mathbb {E} _ {t, \\Omega_ {0} \\sim p _ {0}, \\Omega_ {1} \\sim p _ {1}} \\| v _ {\\eta} (\\Omega_ {t}, t | c) - (\\Omega_ {1} - \\Omega_ {0}) \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 265, + 482, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_0$ is Gaussian distribution with zero mean and unit variance at time $t = 0$ , and $p_1$ is the target distribution as time $t = 1$ . $v_{\\eta}(\\Omega_t, t|c)$ is parameterized as a neural network with weights $\\eta$ and $c$ is the input condition features.", + "bbox": [ + 89, + 297, + 483, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the input scene object point cloud $\\mathcal{O}$ is encoded by Pointnet++ [34], language is encoded by a pre-trained CLIP model [35], and the direction vector and time embedding are encoded by the Multilayer Perceptrons (MLP). Due to the permutation invariance of point level affordance, the noisy affordance are concatenated with object features to aware the position and semantic information. Then all features are fed into a Perceiver IO Transformer, following [16, 49]. The noisy affordance features are served as input and output array, and the features of language and direction are served as latent array. The outputs of the decoder go through an MLP to obtain the prediction of the flow $v_{\\eta}(\\Omega_t,t|c)\\in R^M$ . Finally, in the inference time, we adopt multiple steps $1 / \\Delta t$ following the equation:", + "bbox": [ + 89, + 358, + 483, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Omega} _ {t + \\Delta t} = \\Omega_ {t} + \\Delta t v _ {\\eta} \\left(\\Omega_ {t}, t \\mid c _ {1}\\right), \\quad t \\in [ 0, 1). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 580, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Grasp Flow Matching", + "text_level": 1, + "bbox": [ + 89, + 611, + 290, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce Grasp Flow Matching (GFM) to learn the grasp distribution, conditioned on the output of MLLM and AFM. GFM is also built upon the flow matching model, and the objective is to generate a parameterized dexterous hand pose $\\mathcal{G}^{dex} = (r,t,q)$ , which represents the global rotation, translation, and joint angles. In the training time, grasp generation requires calculating explicit losses in 3D space, which involves obtaining the target dexterous poses, which go through forward kinematics to obtain hand configuration [53]. To obtain the target pose, we estimate it in one step during the training stage:", + "bbox": [ + 89, + 630, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {1} ^ {\\hat {d e x}} = \\mathcal {G} _ {t} ^ {d e x} + (1 - t) * v _ {\\theta} (\\mathcal {G} _ {t} ^ {d e x}, t | c _ {2}), \\quad t \\in [ 0, 1), (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 808, + 482, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the condition features $c_{2}$ are the concatenation of object affordance, language and direction features. And the object affordance features are obtained by feeding affordance with point clouds into PointNet++ [34].", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The losses of grasp generation include 1) the grasp pose regression L2 loss; 2) the hand chamfer loss to minimize the discrepancies between the actual hand shapes; 3) the hand fingertip key point loss to minimize the distance between fingertip position. The loss function can be formulated as:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {G F M} = \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {c h a m f e r}} \\mathcal {L} _ {\\text {c h a m f e r}} + \\lambda_ {\\text {t i p}} \\mathcal {L} _ {\\text {t i p}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 174, + 905, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\text{pose}}, \\lambda_{\\text{chamfer}}$ and $\\lambda_{\\text{tip}}$ are loss weights.", + "bbox": [ + 511, + 205, + 839, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The object penetration loss, which is used to penalize the penetration of hands into objects, is excluded from grasp generation training, as it leads to significant challenges in the training process [51]. Employing the penetration loss can reduce object penetration, but negatively impacts both intention alignment and generation diversity. Therefore, to achieve semantic-aligned grasp generation, we exclude object penetration loss and introduce affordance-guided grasp optimization to reduce object penetration.", + "bbox": [ + 511, + 220, + 906, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.4. Affordance-guided Grasp Optimization", + "text_level": 1, + "bbox": [ + 511, + 362, + 834, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a non-parametric test-time adaptation (TTA), affordance-guided grasp optimization, to improve the grasp quality while maintaining intention consistency. Compared to learning-based TTA methods [17, 51], our nonparametric optimization does not suffer from performance degradation caused by out-of-domain open-set samples.", + "bbox": [ + 511, + 380, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design optimization objectives based on that high-quality grasps should exhibit high grasp stability and intention alignment in grasping posture and contact areas. Specifically, our optimization objectives include: 1) affordance contact loss for constraining hand contact candidate points in contact with the affordance area:", + "bbox": [ + 511, + 472, + 905, + 561 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a f f - d i s t}} = \\sum_ {i} \\mathbb {I} \\left(\\left(d \\left(p _ {i} ^ {c}\\right) < \\tau_ {1}\\right) \\vee \\left(d \\left(p _ {i} ^ {r}\\right) < \\tau_ {1}\\right)\\right) \\cdot d \\left(p _ {i} ^ {r}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 569, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d(p) = \\min_{o \\in \\{o_j \\in \\mathcal{O} | a_j > \\tau_2\\}} \\| p - o_j \\|_2$ , and $o_j$ and $a_j$ represent the position and affordance values of object points, and $p_i^c$ and $p_i^r$ represent the hand contact candidates of the initial coarse hand and current refined hand. 2) affordance fingertip loss to keep the fingertip positions that are in contact with the affordance area unchanged.", + "bbox": [ + 511, + 607, + 905, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3) object penetration loss for punish hand-object penetration, 4) self-penetration loss to punish the penetration of hand it self. 5) joint limitation loss to punish out-of-limit joint angles. The optimization objectives can be formulated as:", + "bbox": [ + 511, + 699, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} ^ {d e x}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {a f f - d i s t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a f f - f i n g e r}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s p e n}} + \\lambda_ {5} \\mathcal {L} _ {\\text {j o i n t}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 781, + 910, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Open-set Dexterous Grasp Dataset", + "text_level": 1, + "bbox": [ + 511, + 830, + 828, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To support the evaluation of our framework, we build an open-set dexterous grasp dataset in table-top environment based on the language-guided dexterous grasp dataset", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11822", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d1f866de7273b6da08c4652c32ff9ebbf481d5d1c9a72bba8e858761b4f42002.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IntentionQualityDiversity
FID ↓CD ↓R - Precision ↑Suc. ↑Q1 ↑Pen. ↓δt ↑δr ↑δq ↑
Top1Top2Top3
Open Set A
ContactGen[28]0.4289.380.1640.2320.28911.6%1.71e-41.092.8618.439.2
Contact2Grasp[19]0.42610.20.2570.3370.39816.5%0.01720.6685.734.8448.3
GraspCVAE[40]0.3785.550.3090.4020.46921.9%0.01500.7094.364.6021.2
SceneDiffuser[12]0.3035.020.3970.4730.52329.1%0.01510.4877.379.4468.9
DexGYS[51]0.2974.630.3170.4010.46344.2%0.05120.3626.306.7954.7
Ours0.2313.810.4800.5880.66645.1%0.05310.2937.486.9861.5
Open Set B
ContactGen[28]0.4928.950.1960.2570.2996.04%1.30e-41.113.2618.745.5
Contact2Grasp[19]0.36911.90.2480.3180.36716.2%0.008010.7985.224.6340.7
GraspCVAE[40]0.3655.350.2970.3570.40324.2%0.004210.7383.965.4725.6
SceneDiffuser[12]0.2716.500.3020.3550.39330.7%0.008510.6835.416.1372.9
DexGYS[51]0.3583.400.2940.3580.40335.2%0.02200.6915.406.1159.0
Ours0.1622.950.5320.6090.65538.9%0.03520.3616.866.8456.3
", + "bbox": [ + 125, + 88, + 870, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results on open-set datasets compared with the SOTA methods.", + "bbox": [ + 281, + 376, + 712, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[51, 56]. Our open-set dataset consists of 33 categories, 1536 objects, 1909 scenes, and 43,504 dexterous grasps for Shadow Hand [37] and Leap Hand [39].", + "bbox": [ + 89, + 404, + 480, + 450 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Open Set Data Split. To enable comprehensive evaluation, we construct Open sets A and B using two independent dataset splits, each with its own set of unseen categories. For each split, all categories are first labeled as either seen or unseen. Then, $80\\%$ of the objects from the seen categories are used for training, while the remaining $20\\%$ and all objects from the unseen categories are used for testing.", + "- Scene Construction. To make our dataset more practical, we generate scene data by placing objects in a tabletop environment using Blenderproc [5]. To prevent collisions between the hand and the table, objects are elevated using a shelf, and collision detection is performed to filter out invalid grasps. The physics engine is activated to ensure physically plausible object placements and grasping poses, enhancing dataset quality.", + "- Scene Point-Cloud and Image Rendering. The scene was captured using five RGB-D cameras: four positioned at elevated lateral angles and one directly above the object. The partial point clouds are merged into a global point cloud. To render realistic RGB images for MLLM processing, we use the texture generation model Paint3D [57] to apply realistic textures to all objects in our dataset." + ], + "bbox": [ + 89, + 450, + 482, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 792, + 215, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 816, + 274, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employ three types of metric to evaluate the ability of intention consistency, grasp quality and grasp diversity.", + "bbox": [ + 89, + 839, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Evaluating Intention Consistency. We employ several metrics, including FID, R-Precision and Chamfer Distance.", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1) FID (Frechet Inception Distance) [11] measures the distribution similarity between generated grasps and ground truth. To extract grasp and instruction features, we train an object-grasp point cloud encoder and a language encoder by contrastive learning [10]. 2) R-Precision evaluates the semantic alignment between language instructions and generated grasps [10]. Specifically, for each generated grasp, we construct a pool of 32 samples, including its ground-truth instruction and randomly selected samples. We then compute and rank the cosine distances between the point cloud features and the language features of all samples in the pool. The average accuracy is computed at the top-1, top-2, and top-3 positions, that a retrieval is considered successful if the ground truth entry appears among the top-k candidates. 3) Chamfer Distance (CD) quantifies the discrepancy between the predicted hand point cloud and the closest target with same intention [51].", + "bbox": [ + 511, + 404, + 906, + 661 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Evaluating Grasp Quality. Following [46], we use success rate, denoted as $Suc$ , in Issac Gym [24] and Q1 [7] to assess grasp stability. Maximal penetration depth (cm), denoted as $Pen$ , is used to calculate the maximal penetration depth from the object point cloud to hand meshes.", + "bbox": [ + 511, + 662, + 905, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Evaluating Grasp Diversity. We compute the Standard deviation of translation $\\delta_t$ , rotation $\\delta_r$ and joint angle $\\delta_q$ of samples within the same scene observation.", + "bbox": [ + 511, + 739, + 905, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 800, + 730, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For training our framework, the number of epochs is set to 50 for AFM and 200 for GFM. For AFM, the loss weight of L2 loss is set to 1. For GFM, we set $\\lambda_{\\mathrm{pose}} = 10$ , $\\lambda_{\\mathrm{chamfer}} = 1$ , and $\\lambda_{\\mathrm{tip}} = 2$ . The batch size is set to 16 for AFM and 64 for GFM. The initial learning rate is $2.0 \\times 10^{-4}$ , decaying", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea58a6d902f45c0aacd628d48fbd9d8e39a3cdf612a69f5294b499e2a99a2631.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Open Set AOpen Set B
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
Necessity of Generalizable-Instructive Affordance
w/o affordance0.3385.090.3690.01800.5690.3224.440.3080.01940.452
w object part (pred)0.3135.060.3200.01380.5940.3163.790.3448.12e-30.721
w contact map (pred)0.3204.890.3800.01310.6050.2863.250.4089.25e-30.636
w our affordance (pred)0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w object part (GT)0.3065.050.3480.01980.5230.2713.790.3536.68e-30.526
w contact map (GT)0.1132.010.5980.02850.4940.1382.680.5940.01430.643
w our affordance (GT)0.1693.520.5090.01810.5450.1642.610.5500.01200.584
Effectiveness of Pre-Understanding Stage
w/o key cues extraction0.2583.990.4630.02220.5100.1863.420.5158.33e-30.702
w/o direction0.2544.070.4320.01620.5310.1852.850.5298.80e-30.679
w MLLM0.2423.790.4800.02400.5010.1762.760.5380.01500.612
Effectiveness of Affordance-based Optimization
w/o our optimization0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w penetration loss0.3367.070.3180.05050.2990.2415.760.4450.01470.719
w ContactNet [17]0.34912.30.2790.04340.1240.35511.10.2400.02060.129
w RefineNet [51]0.2493.960.4360.04870.3990.2032.920.4930.01490.657
w our optimization0.2313.810.4770.05310.2930.1622.950.5320.03500.361
", + "bbox": [ + 122, + 88, + 874, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to $2.0 \\times 10^{-5}$ using a cosine learning rate scheduler [31]. The Adam optimizer is used with a weight decay rate of $5.0 \\times 10^{-6}$ . In the inference, the time step is set to 10 for AFM and 20 for GFM. For the optimization, $\\lambda_{1} = 100$ , $\\lambda_{2} = 10$ , $\\lambda_{3} = 100$ , $\\lambda_{4} = 10$ , $\\lambda_{5} = 100$ . The number of optimization iterations is set to 200. All experiments are implemented using PyTorch on a single RTX 4090 GPU.", + "bbox": [ + 89, + 462, + 485, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 587, + 382, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our methods with the SOTA methods, as shown in Table 1. The generic grasp methods are reproduced by concatenating the point cloud features and features of same language guidance with ours. The same encoders are employed for fair comparison, and the penetration loss is excluded to avoid learning difficulties according to [51]. The results show that our framework significantly outperforms all previous methods in terms of intention consistency and grasp quality. Our framework also achieves a reasonably high level of diversity, as excessive diversity may lead to unnatural postures. Previous language-guided methods fail to generalize well to unseen categories due to the huge gap between language and grasping actions. Similarly, contact-based methods don't perform well, as the contact maps are difficult to generalize. Additionally, our method outperforms other methods in the close set, as shown in Table 3. Overall, the results indicate that our framework achieves strong performance in generating intention-aligned, high-quality, and diverse grasps.", + "bbox": [ + 88, + 613, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/76f77b5afa0e6db87d89fe0954969da1b8a9e9c55a1dbef672adca535dacf55c.jpg", + "table_caption": [ + "Table 2. Ablation study for our framework. The results of first two experiment groups are obtained from model outputs without optimization for an intuitive and reasonable comparison. The results in each group should be compared with light yellow line (our default setting)." + ], + "table_footnote": [], + "table_body": "
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
GraspCVAE0.2082.400.3950.01000.771
DexGYS0.08041.740.5900.05510.397
Ours0.02861.130.7790.05620.193
", + "bbox": [ + 535, + 460, + 880, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Experiment of close set with SOTA methods.", + "bbox": [ + 547, + 549, + 870, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Necessity of Our Affordance", + "text_level": 1, + "bbox": [ + 511, + 573, + 769, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results shown in Table 2 validate the key insight of our framework: using the generalizable-instructive affordance to bridge the gap between high-level language and low-level grasp actions. The first row presents the results without affordance, which fails to generalize to unseen categories. The subsequent two rows show the results under the condition of object parts and contact maps. Prediction refers to maps generated by models, while GT refers to using the corresponding ground truth. As discussed in Section 3.1, the results demonstrate that object parts are too coarse to guide the grasp effectively, even when using the ground truth. On the other hand, the contact map is too fine-grained to generalize to unseen categories. Only our affordance achieves a balance between generalization and instruction.", + "bbox": [ + 511, + 597, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, our affordance representation and framework demonstrate good generalization in the one-shot setting, as shown in Table 4. For the one-shot experiments, we introduce several novel categories to the test set that differ significantly from the training set and add one object from each unseen category to the training set. The results show", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11824", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff9a2f66754552d456ac72bbbff60e6905a37318b256e6f3dd9423ebe98e928d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 90, + 187, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f519ccdeb3f8cce1538d3a8d844d1749f8b2eca6b89b3581c5217958209bdd1.jpg", + "image_caption": [ + "Figure 4. The visualization of generated affordance and dexterous grasp. The left top shows the zero-shot samples and the left bottom shows the one-shot samples in real world. The right top and right bottom show the zero-shot samples in simulation open set A and B." + ], + "image_footnote": [], + "bbox": [ + 93, + 167, + 186, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/20e2a9ba5a3591a20b0edc4145b86ee5a4ee497d4aab6d4442b326c9b0445e98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 93, + 287, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28b195ad32641f84d0f9bde82224acb8275311f2a8d1cb4e448e26120b7e6e81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 167, + 287, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac1873f58b02f2801fe3a262136a431df91ebba8cec69c0176c533ab0d678b72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 92, + 390, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dd62dd124ddf7e64f1955ff8173d8767a04ed439a1245326a9536fc3af15034.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 167, + 390, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a00adb8a0018a38db997a5970933dc866aff9dbab9265be93f27287912aef22e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 92, + 491, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/84b7fd3620518bd178c053dcaf6f15a212d29d68958b0693bf3fd9a637a4ddf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 167, + 491, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96c2c28e389b2de8bbf57d1efa143c6c32bf22e01ce16a2b11334d35f57a15b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 92, + 593, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b300cb4db437cc4cc8e5ecb879262e4df3f28b483a1e046e5f7212bfa2eb08c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 167, + 593, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/60abc94610e4bbe048c189954d6febeffd4a21c5837fe88ee3fe5e1ec42db156.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 599, + 92, + 694, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ca27846da0469bc14a458e7ca54a29fbea4d7bd49fcdadf343f702007d02c9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 167, + 694, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c6e5b63ce2eb82118e3d34a3431b21ac4d86652c837b1f2fdc673a8abdc2c2f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 92, + 797, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4c442c48f4bd7c7fa5e61c17c4aea7f8763c9515e6d4b7d05ba6354a2813538.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 167, + 797, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4730945ebc7f3f20ec9572a97eff42a85aa112af89e346b347b33c939f6482c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 92, + 900, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edff7035c278721fa4962391ed5e5982dc8427c3bfdab4715e651eb231a5d801.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 167, + 900, + 239 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/10bc1fbcc9f91dcd0050c4d6c34fab515ccdc4513f634f64cf06ad30ecd587c4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Seen CateSimilar CateNovel Cate
Top1 ↑ CD ↓Top1 ↑ CD ↓Top1 ↑ CD ↓
Zero-Shot
w/o afford0.8272.610.2955.710.2197.82
Ours0.8801.060.4323.650.2466.42
One-Shot
w/o afford0.7673.170.4293.670.3426.64
Ours0.8700.990.6561.730.5863.03
", + "bbox": [ + 94, + 297, + 478, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that our framework achieves a significant performance improvement, not only on samples from different categories with similar structures but also for novel categories.", + "bbox": [ + 89, + 486, + 483, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Effectiveness of Each Component", + "text_level": 1, + "bbox": [ + 89, + 542, + 383, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further ablation studies are conducted to evaluate each component of our framework. For the pre-understanding stage assisted by MLLM, w/o key cues extraction refers to directly inputting the user's commands into the model, while w/o direction means that the direction information is not utilized. Both designs enhance performance, and we believe that MLLM would be more beneficial in real-world applications due to its powerful generalization. For affordance-based optimization, our optimization improves grasp quality and maintains intention consistency, outperforming the learning-based methods ContactNet [17] and RefineNet [51], as well as the model which is trained with penetration loss.", + "bbox": [ + 89, + 565, + 483, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Real World Experiments", + "text_level": 1, + "bbox": [ + 89, + 772, + 316, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The real-word experiments are conducted to verify the simulation-to-reality ability of our framework. We employ a Leap Hand, a Kinova Gen3 6DOF arms and an original wrist RGB-D camera of Kinova arm. And we collect several common objects in daily life, including unseen objects and unseen categories. In experiment, we synthesize the scene point cloud by taking several partial depth maps", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a89dce9a47945e39a3997f05be7b926cfa7b92cdcb44b6a253afe9622277348d.jpg", + "table_caption": [ + "Table 4. Zero-shot and one-shot generalization of our framework." + ], + "table_footnote": [], + "table_body": "
DexGYS273647010
Ours105910610610
", + "bbox": [ + 514, + 297, + 910, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Real world experiments: Successes in 10 attempts.", + "bbox": [ + 531, + 371, + 885, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "around the object. Then the scene point cloud, a RGB image and the user language command are fed into our framework to obtain the dexterous grasp pose. During execution, we first move the arm to a pre-grasp position, then synchronously move the joints of the robotic arm and the dexterous hand to reach the target pose. In evaluation, a grasp is considered as success if the hand can lift it up and the action is consistent with the given command. The results shown in Table 5 shows that our framework achieve higher success rate than previous method.", + "bbox": [ + 511, + 402, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 571, + 638, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We believe that achieving generalizable dexterous grasps in open set is crucial within the deep learning and robotics communities. In this paper, we explore a novel task of open-set language-guided dexterous grasp. This task is challenging due to the huge gap between language and grasping actions, which hinders the model's generalization ability. We propose the AffordDexGrasp framework with the insight that using a new affordance representation, generalizable-instructive affordance, to solve this challenge. This affordance enables generalization to unseen categories by utilizing the object's local structure and category-agnostic semantic attributes, thus facilitating effective dexterous grasp generation. Based on this affordance, our framework introduces two flow matching based models for affordance generation and affordance-guided grasp generation. Moreover, we introduce a pre-understand stage and a pose-optimization stage to further boost generalization. We conduct extensive open-set experiments in both simulation and the real world, and the results show that our framework outperforms all SOTA methods.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11825", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported partially by NSFC (92470202, U21A20471), Guangdong NSF Project (No. 2023B151504 0025), Guangdong Key Research and Development Program (No. 2024B0101040004).", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 5", + "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 2, 3", + "[3] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 3", + "[4] Xiaoyun Chang and Yi Sun. Text2grasp: Grasp synthesis by text prompts of object grasping parts. arXiv preprint arXiv:2404.15189, 2024. 2", + "[5] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Wendelin Knauer, Klaus H Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6", + "[6] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 39(5):3929-3945, 2023. 2", + "[7] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In Proceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., pages 2290-2295. IEEE, 1992. 6", + "[8] James J Gibson. The theory of affordances:(1979). In *The people, place, and space reader*, pages 56-60. Routledge, 2014. 3", + "[9] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1471-1481, 2021. 2, 3", + "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 6", + "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + ], + "bbox": [ + 93, + 215, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rium. Advances in neural information processing systems, 30, 2017. 6", + "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 6", + "[13] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 3", + "[14] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3", + "[15] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4", + "[16] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 5", + "[17] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2, 5, 7, 8", + "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 3", + "[19] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3, 6", + "[20] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 1", + "[21] Kailin Li, Jingbo Wang, Lixin Yang, Cewu Lu, and Bo Dai. Semgrasp: Semantic grasp generation via language aligned discretization. In European Conference on Computer Vision, pages 109-127. Springer, 2024. 1, 2", + "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3", + "[23] Samuel Li, Sarthak Bhagat, Joseph Campbell, Yaqi Xie, Woojun Kim, Katia Sycara, and Simon Stepputtis. Shape-grasp: Zero-shot task-oriented grasping with large language models through geometric decomposition. In 2024" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11826", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10527-10534. IEEE, 2024. 2", + "[24] Jacky Liang, Viktor Makoviychuk, Ankur Handa, Nuttapong Chentanez, Miles Macklin, and Dieter Fox. Gpu-accelerated robotic simulation for distributed reinforcement learning. In Conference on Robot Learning, pages 270-282. PMLR, 2018. 6", + "[25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 5", + "[26] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2", + "[27] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2, 3", + "[28] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 6", + "[29] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2", + "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5", + "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7", + "[32] Teli Ma, Zifan Wang, Jiaming Zhou, Mengmeng Wang, and Junwei Liang. Glover: Generalizable open-vocabulary affordance reasoning for task-oriented grasping. arXiv preprint arXiv:2411.12286, 2024. 3", + "[33] Teli Ma, Jiaming Zhou, Zifan Wang, Ronghe Qiu, and Junwei Liang. Contrastive imitation learning for language-guided multi-task robotic manipulation. arXiv preprint arXiv:2406.09738, 2024. 2", + "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5", + "[36] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2, 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] ShadowHand. Shadowrobot. https://www.).shadowrobot.com/dexterous-hand-series/, 2005.6", + "[38] Lin Shao, Fabio Ferreira, Mikael Jorda, Varun Nambiar, Jianlan Luo, Eugen Solowjow, Juan Aparicio Ojea, Oussama Khatib, and Jeannette Bohg. Unigrasp: Learning a unified model to grasp with multifingered robotic hands. IEEE Robotics and Automation Letters, 5(2):2286-2293, 2020. 2", + "[39] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv preprint arXiv:2309.06440, 2023. 6", + "[40] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 6", + "[41] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 2, 3", + "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3", + "[43] Ardian Umam, Cheng-Kun Yang, Min-Hung Chen, Jen-Hui Chuang, and Yen-Yu Lin. Partdistill: 3d shape part segmentation by vision-language model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2024. 2", + "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3891-3902, 2023. 2", + "[45] An-Lan Wang, Nuo Chen, Kun-Yu Lin, Li Yuan-Ming, and Wei-Shi Zheng. Task-oriented 6-dof grasp pose detection in clutters. arXiv preprint arXiv:2502.16976, 2025. 2", + "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2, 6", + "[47] Wenbo Wang, Fangyun Wei, Lei Zhou, Xi Chen, Lin Luo, Xiaohan Yi, Yizhong Zhang, Yaobo Liang, Chang Xu, Yan Lu, et al. Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping. arXiv preprint arXiv:2412.02699, 2024. 2", + "[48] Yan-Kang Wang, Chengyi Xing, Yi-Lin Wei, Xiao-Ming Wu, and Wei-Shi Zheng. Single-view scene point cloud human grasp generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-841, 2024. 2", + "[49] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11827", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 5", + "[50] Wei Wei, Peng Wang, Sizhe Wang, Yongkang Luo, Wanyi Li, Daheng Li, Yayu Huang, and Haonan Duan. Learning human-like functional grasping for multi-finger hands from few demonstrations. IEEE Transactions on Robotics, 2024. 2", + "[51] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. Advances in Neural Information Processing Systems, 37:46881-46907, 2025. 1, 2, 3, 5, 6, 7, 8", + "[52] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-dof grasp detection. In European Conference on Computer Vision, pages 357-375. Springer, 2024. 2", + "[53] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5", + "[54] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4737-4746, 2023. 2", + "[55] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 2", + "[56] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20953-20962, 2022. 2, 6", + "[57] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4252-4262, 2024. 6", + "[58] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. In 8th Annual Conference on Robot Learning, 2024. 2", + "[59] Zhongqun Zhang, Hengfei Wang, Ziwei Yu, Yihua Cheng, Angela Yao, and Hyung Jin Chang. Nl2contact: Natural language guided 3d hand-object contact modeling with diffusion model. In European Conference on Computer Vision, pages 284-300. Springer, 2024. 1, 3", + "[60] Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yun-hao Fang, and Hao Su. Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmenta" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion and maximum likelihood estimation. arXiv preprint arXiv:2312.03015, 2023. 3", + "[61] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15741-15751, 2021. 2", + "[62] Binghui Zuo, Zimeng Zhao, Wenqian Sun, Xiaohan Yuan, Zhipeng Yu, and Yangang Wang. Graspondiff: Grasping generation for hand-object interaction with multimodal guided diffusion. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11828", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_model.json b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4f613ef825f3f4eb6acf74d2e15f97a3eb39a28 --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_model.json @@ -0,0 +1,2895 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.819, + 0.175 + ], + "angle": 0, + "content": "AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.203, + 0.758, + 0.239 + ], + "angle": 0, + "content": "Yi-Lin Wei\\*, Mu Lin\\*, Yuhao Lin\\*, Jian-Jian Jiang\\*, Xiao-Ming Wu\\*, Ling-An Zeng\\*, Wei-Shi Zheng†,2" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.239, + 0.809, + 0.257 + ], + "angle": 0, + "content": "1 School of Computer Science and Engineering, Sun Yat-sen University, China" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.257, + 0.89, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.278, + 0.862, + 0.292 + ], + "angle": 0, + "content": "{weiylin5, linm67, linyh96, jiangjj35, wuxm65, zenglan3}@mail2.sysu.edu.cn wszheng@ieee.org" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.294, + 0.692, + 0.31 + ], + "angle": 0, + "content": "https://isee-laboratory.github.io/AffordDexGrasp/" + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.329, + 0.179, + 0.341 + ], + "angle": 0, + "content": "Training Data" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.342, + 0.192, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.397, + 0.293, + 0.408 + ], + "angle": 0, + "content": "Use the teapot with grasping the handle from the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.329, + 0.391, + 0.34 + ], + "angle": 0, + "content": "Generalizable-Instructive Affordance" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.342, + 0.395, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.397, + 0.493, + 0.408 + ], + "angle": 0, + "content": "Use the mug by contacting the handle from the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.329, + 0.49, + 0.341 + ], + "angle": 0, + "content": "Unseen Category" + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.342, + 0.495, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.397, + 0.493, + 0.408 + ], + "angle": 0, + "content": "the handle from the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.329, + 0.753, + 0.34 + ], + "angle": 0, + "content": "Real World Open set" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.341, + 0.597, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.397, + 0.695, + 0.407 + ], + "angle": 0, + "content": "Use the mug by contacting the handle from the right." + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.342, + 0.698, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.397, + 0.695, + 0.407 + ], + "angle": 0, + "content": "the handle from the right." + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.341, + 0.801, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.397, + 0.897, + 0.407 + ], + "angle": 0, + "content": "Use the kettle by contacting the handle from the left." + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.342, + 0.901, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.808, + 0.397, + 0.898, + 0.407 + ], + "angle": 0, + "content": "the handle from the left." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.413, + 0.192, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.472, + 0.283, + 0.482 + ], + "angle": 0, + "content": "From the right, grasp the handle to use the knife." + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.413, + 0.395, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.472, + 0.49, + 0.482 + ], + "angle": 0, + "content": "Use the frying pan and grasp handle from the right." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.413, + 0.597, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.471, + 0.691, + 0.482 + ], + "angle": 0, + "content": "Use the hammer and grasp handle from the back." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.413, + 0.698, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.471, + 0.69, + 0.481 + ], + "angle": 0, + "content": "\\(p\\) handle from the back." + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.413, + 0.801, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.471, + 0.897, + 0.482 + ], + "angle": 0, + "content": "Use the frying pan and grasp handle from the back." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.489, + 0.192, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.545, + 0.256, + 0.555 + ], + "angle": 0, + "content": "Hold the bottle body from the left." + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.491, + 0.395, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.545, + 0.479, + 0.555 + ], + "angle": 0, + "content": "Hold the body of trigger sprayer from the left." + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.489, + 0.495, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.489, + 0.597, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.545, + 0.683, + 0.555 + ], + "angle": 0, + "content": "Hold the body of trigger sprayer from the left." + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.489, + 0.698, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.545, + 0.683, + 0.555 + ], + "angle": 0, + "content": "sprayer from the left." + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.489, + 0.801, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.545, + 0.887, + 0.555 + ], + "angle": 0, + "content": "Hold the body of the lotion pump from the left." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.907, + 0.608 + ], + "angle": 0, + "content": "Figure 1. Open-set Language-guided Dexterous Grasp. Our framework bridges the gap between language and grasp actions through Generalizable-Instructive Affordance, which enables cross-category generalization via category-agnostic cues and graspable local structure. Remarkably, our framework demonstrates strong generalization without requiring extra real training data in real-world experiments." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.621, + 0.327, + 0.637 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.654, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Language-guided robot dexterous generation enables robots to grasp and manipulate objects based on human commands. However, previous data-driven methods are hard to understand intention and execute grasping with unseen categories in the open set. In this work, we explore a new task, Open-set Language-guided Dexterous Grasp, and find that the main challenge is the huge gap between high-level human language semantics and low-level robot actions. To solve this problem, we propose an Affordance Dexterous Grasp (AffordDexGrasp) framework, with the insight of bridging the gap with a new generalizable-instructive affordance representation. This affordance can generalize to unseen categories by leveraging the object's local structure and category-agnostic semantic attributes, thereby ef" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.76 + ], + "angle": 0, + "content": "fectively guiding dexterous grasp generation. Built upon the affordance, our framework introduces Affordance Flow Matching (AFM) for affordance generation with language as input, and Grasp Flow Matching (GFM) for generating dexterous grasp with affordance as input. To evaluate our framework, we build an open-set table-top language-guided dexterous grasp dataset. Extensive experiments in the simulation and real worlds show that our framework surpasses all previous methods in open-set generalization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.786, + 0.645, + 0.802 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.907, + 0.872 + ], + "angle": 0, + "content": "Achieving generalizable robot dexterous grasping is an important goal in the fields of robotics and computer vision, with exciting potential applications in human-robot interaction and robot manipulation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.873, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recent works explore the task of language-guided dexterous grasp generation[20, 21, 51, 59], aiming to enable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.877, + 0.221, + 0.889 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.236, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.877, + 0.236, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11818" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "dexterous hands to perform actions based on language instructions, going beyond previous works that focus on stable grasping [26, 38, 46, 54]. The typical language-guided approaches employ language as the condition of generative models to predict hand parameters, achieving impressive performance on objects within known categories [4, 51]. However, in the open real world, there are many categories that may not appear during training, and the cost of collecting data for dexterous hands is quite expensive. Therefore, the open-set generalization on unseen category samples is crucial for robot grasp. While previous works on parallel grippers explore open-set tasks [23, 36], there is limited research on open-set generalization for dexterous hands, as their significantly higher degrees of freedom present complex challenges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.482, + 0.487 + ], + "angle": 0, + "content": "In this work, we explore a novel and challenging task: open-set language-guided dexterous grasping, as shown in Figure 1, where models are evaluated on objects and language instructions from both seen and unseen categories. This task poses a great challenge in ensuring that grasps are intention consistent with corresponding language instructions for unseen categories. We find that the main challenge lies in the huge gap between high-level natural language and low-level robot action spaces, which makes it difficult to generalize the ability of understanding intentions and grasping from the training domain to unseen categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.731 + ], + "angle": 0, + "content": "To solve the above challenge, we propose the Affordance Dexterous Grasp (AffordDexGrasp) framework, with the key insight of employing a new affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. Our affordance representation combines two characteristics: generalizable (generalizing to unseen categories based on language) and instructive (guiding dexterous grasp generation effectively). However, achieving these two characteristics is not trivial. For example, as shown in Figure 2, fine-grained contact information can effectively guide grasp generation or optimization [2, 9, 17, 55], but it is difficult to generalize to unseen categories. In contrast, coarse information, such as object parts, can be obtained from a pre-trained computer vision model [27, 41, 43], but it is too coarse to guide dexterous hand actions with higher degrees of freedom." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.884 + ], + "angle": 0, + "content": "To achieve these characteristics, we propose the Generalizable-Instructive Affordance by defining a general dexterous affordance that aligns with category-independent information, such as intention, object parts, and direction. As shown in Figure 2, the affordance represents the potential graspable regions of all grasps with the same semantics. In this way, the models do not need to learn complex dexterous contact patterns but instead focus on a general graspable area, which can be well aligned with category-agnostic semantic attributes and guide grasp generation effectively." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.886, + 0.482, + 0.902 + ], + "angle": 0, + "content": "To generate affordance and employ it to guide grasp gen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "eration, our framework consists of two cascaded generative models. The Affordance Flow Matching generates affordance maps in a generalizable manner based on language. And the Grasp Flow Matching generates dexterous grasp poses under the effective guidance of affordance. Moreover, we introduce a pre-understanding stage and a post-optimization stage to further boost generalization. Specifically, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intention to enhance generalization across diverse user commands. And we introduce an affordance-guided optimization to improve grasp quality while preserving consistency with user intentions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.456 + ], + "angle": 0, + "content": "For evaluating our framework, we build a open-set tabletop language-guided dexterous grasp dataset, based on the language guided dexterous grasp dataset [51, 56]. We exclude specific categories from the training set to test the model's open-set generalization. Moreover, we provide high-quality rendered images to facilitate the usage by MLLM, and we extend the dataset to scene-level data to better simulate real-world environments. The comprehensive experiments are conducted on both the simulation open-set dataset and real-world environments. The results show that our framework can generate dexterous grasp with consistent intention and high quality in open set." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.471, + 0.652, + 0.487 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.496, + 0.681, + 0.512 + ], + "angle": 0, + "content": "2.1. Dexterous Grasp" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Dexterous grasping is critical in robotics, which equips robots with human-like grasping capabilities. Some methods [29, 44, 48, 58] focus on grasp stability and quality, while recent studies explore task-oriented [50, 61] or language-guided grasping [21, 51] with specific semantic intention. For language-guided task, the data-driven methods achieve impressive performance in unseen object within seen categories by leveraging language conditioned generative model [4, 51]. However, we find that these models are hard to generalize for samples from unseen categories. And we find that the main challenge is the huge gap between language and grasp action. To solve this problems, we propose AffordDexGrasp framework with the insight that bridge the gap by generalizable-instructive affordance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.724, + 0.759 + ], + "angle": 0, + "content": "2.2. Open-set Robot Grasp" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Exploring the performance of robotic models in open-set scenarios is crucial due to the diverse object categories in the real world and the high cost of data collection. For tasks that only focus on stable grasping, both parallel grippers [6, 52] and dexterous hands [47, 58] achieve impressive performance on open set. However, when it comes to considering task-oriented [45] or language-guided grasping [33], it becomes much more difficult. Some works in parallel grippers achieve this by using pre-trained visual ground" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11819" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.568, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.185, + 0.278, + 0.483, + 0.29 + ], + "angle": 0, + "content": "(a) Visualization of different affordance representations." + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.099, + 0.897, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.277, + 0.866, + 0.29 + ], + "angle": 0, + "content": "(b) Quantitative Results Across Affordances." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.302, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Different affordance representations. (a) While contact map are too elaborate to generalize and object part are too coarse to guide grasping, our affrodacne achieve a balance. (b) While object part has a lower upper bound and contact shows significant degradation in generalization, only our affordance effectively achieve the balance (Top-1 indicates grasp intention consistency)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.484, + 0.491 + ], + "angle": 0, + "content": "ing models [22, 27, 41], implicit feature semantic fields [18, 32, 36], or multimodal large models [13, 14]. Compared to parallel grippers, dexterous hands have a significantly higher number of degrees of freedom, making the methods of parallel grippers difficult to transfer to dexterous hand [51]. In this paper, we explore the open-set language-guided dexterous grasping task and propose a novel framework to address it." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.504, + 0.267, + 0.519 + ], + "angle": 0, + "content": "2.3. Grasp Affordance" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.526, + 0.484, + 0.768 + ], + "angle": 0, + "content": "Affordance is first introduced in [8], referring to environmental action possibilities. There are currently two types of grasp affordance: fine-grained contact [3, 42] and coarse-grained object segmentation [27]. The fine-grained information, such as contact areas, normals, or key points, is commonly used in hand-object interaction[2, 9, 59, 62]. However, this fine-grained information is difficult to generalize well to unseen categories. As a result, inaccurate contact information would lead to unreasonable grasping. On the other hand, We find through experiments that the coarse object segmentation [27] is too coarse to guide dexterous grasp generation with relatively high degrees of freedom. To solve this problem, we propose Generalizable-Instructive Affordance, which can generalize through the local structure of objects using category-independent clues and effectively guide grasp generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.784, + 0.461, + 0.801 + ], + "angle": 0, + "content": "3. Affordance Dexterous Grasp Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Given the scene point cloud \\(\\mathcal{O}\\), RGB images \\(\\mathcal{I}\\) and language command \\(\\mathcal{C}\\) as input, our goal is to generate intention aligned and high-quality dexterous grasps poses \\(\\mathcal{G}^{dex} = (r,t,q)\\), where \\(r\\) denotes the rotation, \\(t\\) denotes the translation, and \\(q\\) represents the joint angles of the dexterous hand." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we first introduce a novel generalizable-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.448 + ], + "angle": 0, + "content": "instructive affordance representation (Section 3.1). Built on this affordance representation, we then propose the Affordance Dexterous Grasp framework (Section 3.2) including intention pre-understanding stage, Affordance and Grasp Flow Matching, and grasp post-optimization stage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.456, + 0.838, + 0.47 + ], + "angle": 0, + "content": "3.1. Generalizable-Instructive Affordance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.613 + ], + "angle": 0, + "content": "We propose the generalizable-instructive affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. The affordances serve two primary objectives: 1) enable generalization by aligning object local structures with category-agnostic language semantics; and 2) provide instructive guidance for grasp generation through object affordance cues. However, there exists a trade-off in the generalization and instruction, and it is not trivial to achieve both objectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Generalization vs. Instruction. While more fine-grained contact information can better guide grasp generation, it becomes harder to generalize to unseen categories. Fine-grained affordances typically compute the distance between object points and predefined elements, such as hand surface points or hand key points [19, 28]. However, we observe that these fine-gained representations are difficult to generalize to unseen category samples as shown in Figure 2. As the dexterous hand structure is complex and contact modes are varied, the contact maps are also elaborate and varies in space, which make it difficult to generalize." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "On the other hand, the information agnostic to dexterous hands, like object part segmentation, can be obtained in a generalizable manner by vision models [27, 60] or semantic feature fields [36]. However, the object part is too coarse for dexterous hand, as it may struggle to define the clear grasping area. For example, grasping the body of a mug from above and from the side involves the same part, but with different intentions. Moreover, this coarse-grained" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 3. The pipeline of Affordance Dexterous Grasp framework. The inference pipeline includes three stages: 1) intention pre- understanding assisted by MLLM; 2) affordance flow matching for generating affordance base on MLLM output; 3) Grasp Flow Matching and Optimization for outputing grasp poses based on the affordance and MLLM outputs. In the training time, AFM and GFM are independently trained one after another. Transformer and Perceiver are attention-based interaction module for velocity vector field prediction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.42, + 0.482, + 0.451 + ], + "angle": 0, + "content": "learning may hinder the model from learning an accurate grasping method during training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Generalizable-Instructive Affordance. To address this challenge, we propose a novel affordance representation that simultaneously enables both generalization and instruction. The key insight is to establish a correspondence between the general graspable affordance of objects and category-agnostic semantic attributes. Unlike contact maps that describe a particular hand contact area, our affordance represents all potential graspable regions that share the same semantic intention. This approach allows the affordance to generalize more effectively with category-agnostic semantic attributes from language, as it doesn't have to learn the complex distribution of hand contact. Meanwhile, our affordance serves as priors to enhance grasp generation by providing valuable graspable area cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.483, + 0.856 + ], + "angle": 0, + "content": "To obtain the ground truth of generalizable-instructive affordance, we first organize those grasps of one object with the same intention, contact parts and grasp direction into a semantic group. Then we calculate the distances between each scene points \\(\\{o_i\\}^M\\) and hand surface points \\(\\{h_i\\}^N\\), to obtain the contact map \\(\\Omega_{contact} = \\{d_i\\}_1^M\\) for each grasp in the group, where \\(d_i = \\min_{j=1\\dots N}\\| o_i - h_j\\|_2\\). Then we unite contact areas of \\(k\\) grasps in each group by calculating the minimum value of contact areas: \\(d_i^{min} = \\{\\min_k(d_i^k)\\}^M\\). In order to make the affordance map more smooth, we apply Gaussian filtering to the union contact map: \\(a_i = \\sum_{j\\in \\mathcal{N}(i)}\\hat{w}_{ij}d_j^{min}\\) and the weights are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.867, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\hat {w} _ {i j} = \\frac {\\exp \\left(- \\| o _ {i} - o _ {j} \\| ^ {2} / (2 \\sigma^ {2})\\right)}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\left(- \\| o _ {i} - o _ {k} \\| ^ {2} / (2 \\sigma^ {2})\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.467 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(i)\\) is the neighborhood of \\(o_i\\), \\(\\sigma\\) is set as the average nearest neighbor distance. Finally, we obtain the ground truth of generalizable-instructive affordance \\(\\Omega = \\{a_i\\}^M\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.776, + 0.495 + ], + "angle": 0, + "content": "3.2. AffordDexGrasp Framework" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.763, + 0.518 + ], + "angle": 0, + "content": "3.2.1. Intention Pre-Understanding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.719 + ], + "angle": 0, + "content": "The variability of user expression in the open world presents a challenge for models in generalizing and understanding user intentions. To address this, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intentions, by imputing the user language and the rendered RGB image into GPT-4o [15] for intention understanding and key cues extraction. We empirically find that the following cues are both helpful for grasping and easily reasoned by the MLLM: the object category, user intention (e.g., use it), the contacting part of the object (e.g., the handle of a mug), and the grasp direction. The key cues can be extract from language when the commands are clear, otherwise can be reasoned from the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To improve the understanding ability of MLLM, we employ chain of thought (CoT) and visual prompt, and the prompts can be found in supplementary material A.4. Additionally, the grasp direction is simplified as a discretized direction in six predefined coordinates (front, back, left, right, up, down). We first ask MLLM to obtain the direction in image coordinates, and transform it to world coordinates by the camera pose. Then the direction vector is derived from the index of the nearest coordinate axis to form discretized direction. Finally, MLLM organizes this information into a concise sentence (e.g., \"use the mug from the left by contacting the handle\"). This compact and information-dense" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11821" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "sentence structure allows subsequent models to better understand users' intentions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.327, + 0.146 + ], + "angle": 0, + "content": "3.2.2. Affordance Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.256 + ], + "angle": 0, + "content": "We introduce Affordance Flow Matching (AFM) to learn the affordance distribution efficiently, achieving intention aligned and generalizable affordance generation. AFM is built on the conditional flow matching model [1, 25, 30] aiming to learn the velocity vector field from the noised affordance map \\(\\Omega_0\\) to the target affordance map \\(\\Omega_1\\), and the objective loss of AFM could be:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.266, + 0.483, + 0.298 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {A F M} = \\mathbb {E} _ {t, \\Omega_ {0} \\sim p _ {0}, \\Omega_ {1} \\sim p _ {1}} \\| v _ {\\eta} (\\Omega_ {t}, t | c) - (\\Omega_ {1} - \\Omega_ {0}) \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.359 + ], + "angle": 0, + "content": "where \\( p_0 \\) is Gaussian distribution with zero mean and unit variance at time \\( t = 0 \\), and \\( p_1 \\) is the target distribution as time \\( t = 1 \\). \\( v_{\\eta}(\\Omega_t, t|c) \\) is parameterized as a neural network with weights \\( \\eta \\) and \\( c \\) is the input condition features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.571 + ], + "angle": 0, + "content": "Specifically, the input scene object point cloud \\(\\mathcal{O}\\) is encoded by Pointnet++ [34], language is encoded by a pre-trained CLIP model [35], and the direction vector and time embedding are encoded by the Multilayer Perceptrons (MLP). Due to the permutation invariance of point level affordance, the noisy affordance are concatenated with object features to aware the position and semantic information. Then all features are fed into a Perceiver IO Transformer, following [16, 49]. The noisy affordance features are served as input and output array, and the features of language and direction are served as latent array. The outputs of the decoder go through an MLP to obtain the prediction of the flow \\(v_{\\eta}(\\Omega_t,t|c)\\in R^M\\). Finally, in the inference time, we adopt multiple steps \\(1 / \\Delta t\\) following the equation:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.582, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Omega} _ {t + \\Delta t} = \\Omega_ {t} + \\Delta t v _ {\\eta} \\left(\\Omega_ {t}, t \\mid c _ {1}\\right), \\quad t \\in [ 0, 1). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.612, + 0.292, + 0.627 + ], + "angle": 0, + "content": "3.2.3. Grasp Flow Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.798 + ], + "angle": 0, + "content": "We introduce Grasp Flow Matching (GFM) to learn the grasp distribution, conditioned on the output of MLLM and AFM. GFM is also built upon the flow matching model, and the objective is to generate a parameterized dexterous hand pose \\(\\mathcal{G}^{dex} = (r,t,q)\\), which represents the global rotation, translation, and joint angles. In the training time, grasp generation requires calculating explicit losses in 3D space, which involves obtaining the target dexterous poses, which go through forward kinematics to obtain hand configuration [53]. To obtain the target pose, we estimate it in one step during the training stage:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.809, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {1} ^ {\\hat {d e x}} = \\mathcal {G} _ {t} ^ {d e x} + (1 - t) * v _ {\\theta} (\\mathcal {G} _ {t} ^ {d e x}, t | c _ {2}), \\quad t \\in [ 0, 1), (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where the condition features \\( c_{2} \\) are the concatenation of object affordance, language and direction features. And the object affordance features are obtained by feeding affordance with point clouds into PointNet++ [34]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "The losses of grasp generation include 1) the grasp pose regression L2 loss; 2) the hand chamfer loss to minimize the discrepancies between the actual hand shapes; 3) the hand fingertip key point loss to minimize the distance between fingertip position. The loss function can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.175, + 0.906, + 0.206 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {G F M} = \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {c h a m f e r}} \\mathcal {L} _ {\\text {c h a m f e r}} + \\lambda_ {\\text {t i p}} \\mathcal {L} _ {\\text {t i p}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.84, + 0.221 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\text{pose}}, \\lambda_{\\text{chamfer}}\\) and \\(\\lambda_{\\text{tip}}\\) are loss weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.907, + 0.356 + ], + "angle": 0, + "content": "The object penetration loss, which is used to penalize the penetration of hands into objects, is excluded from grasp generation training, as it leads to significant challenges in the training process [51]. Employing the penetration loss can reduce object penetration, but negatively impacts both intention alignment and generation diversity. Therefore, to achieve semantic-aligned grasp generation, we exclude object penetration loss and introduce affordance-guided grasp optimization to reduce object penetration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.836, + 0.378 + ], + "angle": 0, + "content": "3.2.4. Affordance-guided Grasp Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.472 + ], + "angle": 0, + "content": "We introduce a non-parametric test-time adaptation (TTA), affordance-guided grasp optimization, to improve the grasp quality while maintaining intention consistency. Compared to learning-based TTA methods [17, 51], our nonparametric optimization does not suffer from performance degradation caused by out-of-domain open-set samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.563 + ], + "angle": 0, + "content": "We design optimization objectives based on that high-quality grasps should exhibit high grasp stability and intention alignment in grasping posture and contact areas. Specifically, our optimization objectives include: 1) affordance contact loss for constraining hand contact candidate points in contact with the affordance area:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.57, + 0.906, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a f f - d i s t}} = \\sum_ {i} \\mathbb {I} \\left(\\left(d \\left(p _ {i} ^ {c}\\right) < \\tau_ {1}\\right) \\vee \\left(d \\left(p _ {i} ^ {r}\\right) < \\tau_ {1}\\right)\\right) \\cdot d \\left(p _ {i} ^ {r}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.699 + ], + "angle": 0, + "content": "where \\( d(p) = \\min_{o \\in \\{o_j \\in \\mathcal{O} | a_j > \\tau_2\\}} \\| p - o_j \\|_2 \\), and \\( o_j \\) and \\( a_j \\) represent the position and affordance values of object points, and \\( p_i^c \\) and \\( p_i^r \\) represent the hand contact candidates of the initial coarse hand and current refined hand. 2) affordance fingertip loss to keep the fingertip positions that are in contact with the affordance area unchanged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.773 + ], + "angle": 0, + "content": "3) object penetration loss for punish hand-object penetration, 4) self-penetration loss to punish the penetration of hand it self. 5) joint limitation loss to punish out-of-limit joint angles. The optimization objectives can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.782, + 0.911, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} ^ {d e x}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {a f f - d i s t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a f f - f i n g e r}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s p e n}} + \\lambda_ {5} \\mathcal {L} _ {\\text {j o i n t}}\\right). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.829, + 0.848 + ], + "angle": 0, + "content": "4. Open-set Dexterous Grasp Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To support the evaluation of our framework, we build an open-set dexterous grasp dataset in table-top environment based on the language-guided dexterous grasp dataset" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11822" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.872, + 0.367 + ], + "angle": 0, + "content": "
IntentionQualityDiversity
FID ↓CD ↓R - Precision ↑Suc. ↑Q1 ↑Pen. ↓δt ↑δr ↑δq ↑
Top1Top2Top3
Open Set A
ContactGen[28]0.4289.380.1640.2320.28911.6%1.71e-41.092.8618.439.2
Contact2Grasp[19]0.42610.20.2570.3370.39816.5%0.01720.6685.734.8448.3
GraspCVAE[40]0.3785.550.3090.4020.46921.9%0.01500.7094.364.6021.2
SceneDiffuser[12]0.3035.020.3970.4730.52329.1%0.01510.4877.379.4468.9
DexGYS[51]0.2974.630.3170.4010.46344.2%0.05120.3626.306.7954.7
Ours0.2313.810.4800.5880.66645.1%0.05310.2937.486.9861.5
Open Set B
ContactGen[28]0.4928.950.1960.2570.2996.04%1.30e-41.113.2618.745.5
Contact2Grasp[19]0.36911.90.2480.3180.36716.2%0.008010.7985.224.6340.7
GraspCVAE[40]0.3655.350.2970.3570.40324.2%0.004210.7383.965.4725.6
SceneDiffuser[12]0.2716.500.3020.3550.39330.7%0.008510.6835.416.1372.9
DexGYS[51]0.3583.400.2940.3580.40335.2%0.02200.6915.406.1159.0
Ours0.1622.950.5320.6090.65538.9%0.03520.3616.866.8456.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.282, + 0.377, + 0.713, + 0.391 + ], + "angle": 0, + "content": "Table 1. Results on open-set datasets compared with the SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[51, 56]. Our open-set dataset consists of 33 categories, 1536 objects, 1909 scenes, and 43,504 dexterous grasps for Shadow Hand [37] and Leap Hand [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.556 + ], + "angle": 0, + "content": "- Open Set Data Split. To enable comprehensive evaluation, we construct Open sets A and B using two independent dataset splits, each with its own set of unseen categories. For each split, all categories are first labeled as either seen or unseen. Then, \\(80\\%\\) of the objects from the seen categories are used for training, while the remaining \\(20\\%\\) and all objects from the unseen categories are used for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.677 + ], + "angle": 0, + "content": "- Scene Construction. To make our dataset more practical, we generate scene data by placing objects in a tabletop environment using Blenderproc [5]. To prevent collisions between the hand and the table, objects are elevated using a shelf, and collision detection is performed to filter out invalid grasps. The physics engine is activated to ensure physically plausible object placements and grasping poses, enhancing dataset quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.783 + ], + "angle": 0, + "content": "- Scene Point-Cloud and Image Rendering. The scene was captured using five RGB-D cameras: four positioned at elevated lateral angles and one directly above the object. The partial point clouds are merged into a global point cloud. To render realistic RGB images for MLLM processing, we use the texture generation model Paint3D [57] to apply realistic textures to all objects in our dataset." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.216, + 0.81 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.275, + 0.833 + ], + "angle": 0, + "content": "5.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "We employ three types of metric to evaluate the ability of intention consistency, grasp quality and grasp diversity." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Evaluating Intention Consistency. We employ several metrics, including FID, R-Precision and Chamfer Distance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.907, + 0.662 + ], + "angle": 0, + "content": "1) FID (Frechet Inception Distance) [11] measures the distribution similarity between generated grasps and ground truth. To extract grasp and instruction features, we train an object-grasp point cloud encoder and a language encoder by contrastive learning [10]. 2) R-Precision evaluates the semantic alignment between language instructions and generated grasps [10]. Specifically, for each generated grasp, we construct a pool of 32 samples, including its ground-truth instruction and randomly selected samples. We then compute and rank the cosine distances between the point cloud features and the language features of all samples in the pool. The average accuracy is computed at the top-1, top-2, and top-3 positions, that a retrieval is considered successful if the ground truth entry appears among the top-k candidates. 3) Chamfer Distance (CD) quantifies the discrepancy between the predicted hand point cloud and the closest target with same intention [51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.906, + 0.739 + ], + "angle": 0, + "content": "- Evaluating Grasp Quality. Following [46], we use success rate, denoted as \\( Suc \\), in Issac Gym [24] and Q1 [7] to assess grasp stability. Maximal penetration depth (cm), denoted as \\( Pen \\), is used to calculate the maximal penetration depth from the object point cloud to hand meshes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.906, + 0.786 + ], + "angle": 0, + "content": "- Evaluating Grasp Diversity. We compute the Standard deviation of translation \\(\\delta_t\\), rotation \\(\\delta_r\\) and joint angle \\(\\delta_q\\) of samples within the same scene observation." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.801, + 0.731, + 0.817 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For training our framework, the number of epochs is set to 50 for AFM and 200 for GFM. For AFM, the loss weight of L2 loss is set to 1. For GFM, we set \\(\\lambda_{\\mathrm{pose}} = 10\\), \\(\\lambda_{\\mathrm{chamfer}} = 1\\), and \\(\\lambda_{\\mathrm{tip}} = 2\\). The batch size is set to 16 for AFM and 64 for GFM. The initial learning rate is \\(2.0 \\times 10^{-4}\\), decaying" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11823" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.123, + 0.089, + 0.875, + 0.403 + ], + "angle": 0, + "content": "
Open Set AOpen Set B
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
Necessity of Generalizable-Instructive Affordance
w/o affordance0.3385.090.3690.01800.5690.3224.440.3080.01940.452
w object part (pred)0.3135.060.3200.01380.5940.3163.790.3448.12e-30.721
w contact map (pred)0.3204.890.3800.01310.6050.2863.250.4089.25e-30.636
w our affordance (pred)0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w object part (GT)0.3065.050.3480.01980.5230.2713.790.3536.68e-30.526
w contact map (GT)0.1132.010.5980.02850.4940.1382.680.5940.01430.643
w our affordance (GT)0.1693.520.5090.01810.5450.1642.610.5500.01200.584
Effectiveness of Pre-Understanding Stage
w/o key cues extraction0.2583.990.4630.02220.5100.1863.420.5158.33e-30.702
w/o direction0.2544.070.4320.01620.5310.1852.850.5298.80e-30.679
w MLLM0.2423.790.4800.02400.5010.1762.760.5380.01500.612
Effectiveness of Affordance-based Optimization
w/o our optimization0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w penetration loss0.3367.070.3180.05050.2990.2415.760.4450.01470.719
w ContactNet [17]0.34912.30.2790.04340.1240.35511.10.2400.02060.129
w RefineNet [51]0.2493.960.4360.04870.3990.2032.920.4930.01490.657
w our optimization0.2313.810.4770.05310.2930.1622.950.5320.03500.361
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.412, + 0.907, + 0.444 + ], + "angle": 0, + "content": "Table 2. Ablation study for our framework. The results of first two experiment groups are obtained from model outputs without optimization for an intuitive and reasonable comparison. The results in each group should be compared with light yellow line (our default setting)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.486, + 0.57 + ], + "angle": 0, + "content": "to \\(2.0 \\times 10^{-5}\\) using a cosine learning rate scheduler [31]. The Adam optimizer is used with a weight decay rate of \\(5.0 \\times 10^{-6}\\). In the inference, the time step is set to 10 for AFM and 20 for GFM. For the optimization, \\(\\lambda_{1} = 100\\), \\(\\lambda_{2} = 10\\), \\(\\lambda_{3} = 100\\), \\(\\lambda_{4} = 10\\), \\(\\lambda_{5} = 100\\). The number of optimization iterations is set to 200. All experiments are implemented using PyTorch on a single RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.383, + 0.604 + ], + "angle": 0, + "content": "5.3. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We compare our methods with the SOTA methods, as shown in Table 1. The generic grasp methods are reproduced by concatenating the point cloud features and features of same language guidance with ours. The same encoders are employed for fair comparison, and the penetration loss is excluded to avoid learning difficulties according to [51]. The results show that our framework significantly outperforms all previous methods in terms of intention consistency and grasp quality. Our framework also achieves a reasonably high level of diversity, as excessive diversity may lead to unnatural postures. Previous language-guided methods fail to generalize well to unseen categories due to the huge gap between language and grasping actions. Similarly, contact-based methods don't perform well, as the contact maps are difficult to generalize. Additionally, our method outperforms other methods in the close set, as shown in Table 3. Overall, the results indicate that our framework achieves strong performance in generating intention-aligned, high-quality, and diverse grasps." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.461, + 0.882, + 0.54 + ], + "angle": 0, + "content": "
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
GraspCVAE0.2082.400.3950.01000.771
DexGYS0.08041.740.5900.05510.397
Ours0.02861.130.7790.05620.193
" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.55, + 0.871, + 0.564 + ], + "angle": 0, + "content": "Table 3. Experiment of close set with SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.771, + 0.59 + ], + "angle": 0, + "content": "5.4. Necessity of Our Affordance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.809 + ], + "angle": 0, + "content": "The results shown in Table 2 validate the key insight of our framework: using the generalizable-instructive affordance to bridge the gap between high-level language and low-level grasp actions. The first row presents the results without affordance, which fails to generalize to unseen categories. The subsequent two rows show the results under the condition of object parts and contact maps. Prediction refers to maps generated by models, while GT refers to using the corresponding ground truth. As discussed in Section 3.1, the results demonstrate that object parts are too coarse to guide the grasp effectively, even when using the ground truth. On the other hand, the contact map is too fine-grained to generalize to unseen categories. Only our affordance achieves a balance between generalization and instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Furthermore, our affordance representation and framework demonstrate good generalization in the one-shot setting, as shown in Table 4. For the one-shot experiments, we introduce several novel categories to the test set that differ significantly from the training set and add one object from each unseen category to the training set. The results show" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11824" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.188, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.169, + 0.187, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.094, + 0.288, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.169, + 0.288, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.093, + 0.391, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.169, + 0.391, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.093, + 0.492, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.169, + 0.492, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.093, + 0.594, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.169, + 0.594, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.093, + 0.695, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.169, + 0.695, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.093, + 0.799, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.169, + 0.798, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.093, + 0.901, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.169, + 0.901, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 4. The visualization of generated affordance and dexterous grasp. The left top shows the zero-shot samples and the left bottom shows the one-shot samples in real world. The right top and right bottom show the zero-shot samples in simulation open set A and B." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.298, + 0.48, + 0.442 + ], + "angle": 0, + "content": "
Seen CateSimilar CateNovel Cate
Top1 ↑ CD ↓Top1 ↑ CD ↓Top1 ↑ CD ↓
Zero-Shot
w/o afford0.8272.610.2955.710.2197.82
Ours0.8801.060.4323.650.2466.42
One-Shot
w/o afford0.7673.170.4293.670.3426.64
Ours0.8700.990.6561.730.5863.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.453, + 0.482, + 0.467 + ], + "angle": 0, + "content": "Table 4. Zero-shot and one-shot generalization of our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.534 + ], + "angle": 0, + "content": "that our framework achieves a significant performance improvement, not only on samples from different categories with similar structures but also for novel categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.385, + 0.56 + ], + "angle": 0, + "content": "5.5. Effectiveness of Each Component" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Further ablation studies are conducted to evaluate each component of our framework. For the pre-understanding stage assisted by MLLM, w/o key cues extraction refers to directly inputting the user's commands into the model, while w/o direction means that the direction information is not utilized. Both designs enhance performance, and we believe that MLLM would be more beneficial in real-world applications due to its powerful generalization. For affordance-based optimization, our optimization improves grasp quality and maintains intention consistency, outperforming the learning-based methods ContactNet [17] and RefineNet [51], as well as the model which is trained with penetration loss." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.318, + 0.79 + ], + "angle": 0, + "content": "5.6. Real World Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The real-word experiments are conducted to verify the simulation-to-reality ability of our framework. We employ a Leap Hand, a Kinova Gen3 6DOF arms and an original wrist RGB-D camera of Kinova arm. And we collect several common objects in daily life, including unseen objects and unseen categories. In experiment, we synthesize the scene point cloud by taking several partial depth maps" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.298, + 0.911, + 0.361 + ], + "angle": 0, + "content": "
DexGYS273647010
Ours105910610610
" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.372, + 0.886, + 0.387 + ], + "angle": 0, + "content": "Table 5. Real world experiments: Successes in 10 attempts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.553 + ], + "angle": 0, + "content": "around the object. Then the scene point cloud, a RGB image and the user language command are fed into our framework to obtain the dexterous grasp pose. During execution, we first move the arm to a pre-grasp position, then synchronously move the joints of the robotic arm and the dexterous hand to reach the target pose. In evaluation, a grasp is considered as success if the hand can lift it up and the action is consistent with the given command. The results shown in Table 5 shows that our framework achieve higher success rate than previous method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.64, + 0.588 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We believe that achieving generalizable dexterous grasps in open set is crucial within the deep learning and robotics communities. In this paper, we explore a novel task of open-set language-guided dexterous grasp. This task is challenging due to the huge gap between language and grasping actions, which hinders the model's generalization ability. We propose the AffordDexGrasp framework with the insight that using a new affordance representation, generalizable-instructive affordance, to solve this challenge. This affordance enables generalization to unseen categories by utilizing the object's local structure and category-agnostic semantic attributes, thus facilitating effective dexterous grasp generation. Based on this affordance, our framework introduces two flow matching based models for affordance generation and affordance-guided grasp generation. Moreover, we introduce a pre-understand stage and a pose-optimization stage to further boost generalization. We conduct extensive open-set experiments in both simulation and the real world, and the results show that our framework outperforms all SOTA methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11825" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was supported partially by NSFC (92470202, U21A20471), Guangdong NSF Project (No. 2023B151504 0025), Guangdong Key Research and Development Program (No. 2024B0101040004)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[3] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.458 + ], + "angle": 0, + "content": "[4] Xiaoyun Chang and Yi Sun. Text2grasp: Grasp synthesis by text prompts of object grasping parts. arXiv preprint arXiv:2404.15189, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[5] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Wendelin Knauer, Klaus H Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[6] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 39(5):3929-3945, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In Proceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., pages 2290-2295. IEEE, 1992. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[8] James J Gibson. The theory of affordances:(1979). In *The people, place, and space reader*, pages 56-60. Routledge, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1471-1481, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "rium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[13] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[14] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[15] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[16] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[17] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[19] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[20] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[21] Kailin Li, Jingbo Wang, Lixin Yang, Cewu Lu, and Bo Dai. Semgrasp: Semantic grasp generation via language aligned discretization. In European Conference on Computer Vision, pages 109-127. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[23] Samuel Li, Sarthak Bhagat, Joseph Campbell, Yaqi Xie, Woojun Kim, Katia Sycara, and Simon Stepputtis. Shape-grasp: Zero-shot task-oriented grasping with large language models through geometric decomposition. In 2024" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11826" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10527-10534. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[24] Jacky Liang, Viktor Makoviychuk, Ankur Handa, Nuttapong Chentanez, Miles Macklin, and Dieter Fox. Gpu-accelerated robotic simulation for distributed reinforcement learning. In Conference on Robot Learning, pages 270-282. PMLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[26] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[27] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[28] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[29] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.635 + ], + "angle": 0, + "content": "[32] Teli Ma, Zifan Wang, Jiaming Zhou, Mengmeng Wang, and Junwei Liang. Glover: Generalizable open-vocabulary affordance reasoning for task-oriented grasping. arXiv preprint arXiv:2411.12286, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[33] Teli Ma, Jiaming Zhou, Zifan Wang, Ronghe Qiu, and Junwei Liang. Contrastive imitation learning for language-guided multi-task robotic manipulation. arXiv preprint arXiv:2406.09738, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[36] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[37] ShadowHand. Shadowrobot. https://www.).shadowrobot.com/dexterous-hand-series/, 2005.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Lin Shao, Fabio Ferreira, Mikael Jorda, Varun Nambiar, Jianlan Luo, Eugen Solowjow, Juan Aparicio Ojea, Oussama Khatib, and Jeannette Bohg. Unigrasp: Learning a unified model to grasp with multifingered robotic hands. IEEE Robotics and Automation Letters, 5(2):2286-2293, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[39] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv preprint arXiv:2309.06440, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[40] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[41] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[43] Ardian Umam, Cheng-Kun Yang, Min-Hung Chen, Jen-Hui Chuang, and Yen-Yu Lin. Partdistill: 3d shape part segmentation by vision-language model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3891-3902, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[45] An-Lan Wang, Nuo Chen, Kun-Yu Lin, Li Yuan-Ming, and Wei-Shi Zheng. Task-oriented 6-dof grasp pose detection in clutters. arXiv preprint arXiv:2502.16976, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[47] Wenbo Wang, Fangyun Wei, Lei Zhou, Xi Chen, Lin Luo, Xiaohan Yi, Yizhong Zhang, Yaobo Liang, Chang Xu, Yan Lu, et al. Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping. arXiv preprint arXiv:2412.02699, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[48] Yan-Kang Wang, Chengyi Xing, Yi-Lin Wei, Xiao-Ming Wu, and Wei-Shi Zheng. Single-view scene point cloud human grasp generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-841, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11827" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[50] Wei Wei, Peng Wang, Sizhe Wang, Yongkang Luo, Wanyi Li, Daheng Li, Yayu Huang, and Haonan Duan. Learning human-like functional grasping for multi-finger hands from few demonstrations. IEEE Transactions on Robotics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[51] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. Advances in Neural Information Processing Systems, 37:46881-46907, 2025. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[52] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-dof grasp detection. In European Conference on Computer Vision, pages 357-375. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[53] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[54] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4737-4746, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[55] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[56] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20953-20962, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[57] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4252-4262, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[58] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. In 8th Annual Conference on Robot Learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[59] Zhongqun Zhang, Hengfei Wang, Ziwei Yu, Yihua Cheng, Angela Yao, and Hyung Jin Chang. Nl2contact: Natural language guided 3d hand-object contact modeling with diffusion model. In European Conference on Computer Vision, pages 284-300. Springer, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[60] Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yun-hao Fang, and Hao Su. Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmenta" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tion and maximum likelihood estimation. arXiv preprint arXiv:2312.03015, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[61] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15741-15751, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[62] Binghui Zuo, Zimeng Zhao, Wenqian Sun, Xiaohan Yuan, Zhipeng Yu, and Yangang Wang. Graspondiff: Grasping generation for hand-object interaction with multimodal guided diffusion. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11828" + } + ] +] \ No newline at end of file diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_origin.pdf b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..485f06895d3d43bc34021f0f898443735338df05 --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/428be4db-41fb-452f-ae6c-be43892c88ef_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ccbfadeb68d12f8f8919fe343c6d32258054bdde1b473fcc100e03e03f22cd1 +size 2958965 diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/full.md b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1d490c8a773f80b9e6b6b5395b95cb6b91c21ef1 --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/full.md @@ -0,0 +1,404 @@ +# AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance + +Yi-Lin Wei\*, Mu Lin\*, Yuhao Lin\*, Jian-Jian Jiang\*, Xiao-Ming Wu\*, Ling-An Zeng\*, Wei-Shi Zheng†,2 + +1 School of Computer Science and Engineering, Sun Yat-sen University, China + +$^{2}$ Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China + +{weiylin5, linm67, linyh96, jiangjj35, wuxm65, zenglan3}@mail2.sysu.edu.cn wszheng@ieee.org + +https://isee-laboratory.github.io/AffordDexGrasp/ + +![](images/851ee5ad1020d907eb07085e67cd7c7004423b3e13caab0a47d222611228b900.jpg) +Training Data +Use the teapot with grasping the handle from the right. + +![](images/b0b67ec285c8d3e430274e6f693d2a855cc9ed20761f1e39c26ff3f918963574.jpg) +Generalizable-Instructive Affordance +Use the mug by contacting the handle from the right. + +![](images/dadea84cdf4e1e584861d2ed6c439c2ffb3f7b23822cb36b8cfebabbca47d01f.jpg) +Unseen Category + +the handle from the right. + +![](images/9d1284c5a41b999c1fa0d9a876aa7b67a5155114d740146f66d70f82efe6f90f.jpg) +Use the mug by contacting the handle from the right. + +![](images/62f0115f32b1dadb8a267bd82ddab83bc8d8d235dbc614cccd7d20f96c3af95e.jpg) +Real World Open set + +the handle from the right. + +![](images/36f051009c4c808f8bee163ea1020aa4c64b2b3944f830f5be3cd900860fb987.jpg) +Use the kettle by contacting the handle from the left. + +![](images/62c1e933b3455da5c5fb91bdeb551d79a96a0cd29356b6750e0e51a6177a730a.jpg) +the handle from the left. + +![](images/016528dfbbad3517223b8ddcfa626f092ec44f3cdece38a5012b582eaafdab33.jpg) +From the right, grasp the handle to use the knife. + +![](images/f96d2bf8632d5929131b44065824020067f63d86cd6fd2d35d5e756d28759f1b.jpg) +Use the frying pan and grasp handle from the right. + +![](images/abeb627b88a11ef8decd7dcea31fc4220c6d6cb0e5a618971168f67889b8d04f.jpg) +Use the hammer and grasp handle from the back. + +![](images/36b2889101237f181baaf44b132872e00bfe07dd6ab466ba51793e24018835bb.jpg) + +$p$ handle from the back. + +![](images/c0c565a626041c2e0fc3dfe39d567661895d7c1f3da795b65a45c924ea50d405.jpg) +Use the frying pan and grasp handle from the back. + +![](images/a498eae91c628b0323958faa85ef6be48b285cbbb1dbacc41c260b5c08b20889.jpg) +Hold the bottle body from the left. +Figure 1. Open-set Language-guided Dexterous Grasp. Our framework bridges the gap between language and grasp actions through Generalizable-Instructive Affordance, which enables cross-category generalization via category-agnostic cues and graspable local structure. Remarkably, our framework demonstrates strong generalization without requiring extra real training data in real-world experiments. + +![](images/0dde91e9c69ce33f9ced60433b0c46507f840a2e36a30cf285f1bc9486980b59.jpg) +Hold the body of trigger sprayer from the left. + +![](images/1b6e6acbc46cc46a2d511f8e9035eb56c0b873c5d846abdef490571e9b457a28.jpg) + +![](images/af28b5de35957a31e3e694d3944b8b899dd3b8ff815f56d30bc6614c0b3a1141.jpg) +Hold the body of trigger sprayer from the left. + +![](images/3ea3f3f8709b245ac99e621e6ccc04b7b29b8a7afa3aa85dd76d55a88114bcec.jpg) + +sprayer from the left. + +![](images/b31d3b61d8aadbe3b900b8483b8f9a14ce5a244ad45f3adc4de9e2c7b882131a.jpg) +Hold the body of the lotion pump from the left. + +# Abstract + +Language-guided robot dexterous generation enables robots to grasp and manipulate objects based on human commands. However, previous data-driven methods are hard to understand intention and execute grasping with unseen categories in the open set. In this work, we explore a new task, Open-set Language-guided Dexterous Grasp, and find that the main challenge is the huge gap between high-level human language semantics and low-level robot actions. To solve this problem, we propose an Affordance Dexterous Grasp (AffordDexGrasp) framework, with the insight of bridging the gap with a new generalizable-instructive affordance representation. This affordance can generalize to unseen categories by leveraging the object's local structure and category-agnostic semantic attributes, thereby ef + +fectively guiding dexterous grasp generation. Built upon the affordance, our framework introduces Affordance Flow Matching (AFM) for affordance generation with language as input, and Grasp Flow Matching (GFM) for generating dexterous grasp with affordance as input. To evaluate our framework, we build an open-set table-top language-guided dexterous grasp dataset. Extensive experiments in the simulation and real worlds show that our framework surpasses all previous methods in open-set generalization. + +# 1. Introduction + +Achieving generalizable robot dexterous grasping is an important goal in the fields of robotics and computer vision, with exciting potential applications in human-robot interaction and robot manipulation. + +Recent works explore the task of language-guided dexterous grasp generation[20, 21, 51, 59], aiming to enable + +dexterous hands to perform actions based on language instructions, going beyond previous works that focus on stable grasping [26, 38, 46, 54]. The typical language-guided approaches employ language as the condition of generative models to predict hand parameters, achieving impressive performance on objects within known categories [4, 51]. However, in the open real world, there are many categories that may not appear during training, and the cost of collecting data for dexterous hands is quite expensive. Therefore, the open-set generalization on unseen category samples is crucial for robot grasp. While previous works on parallel grippers explore open-set tasks [23, 36], there is limited research on open-set generalization for dexterous hands, as their significantly higher degrees of freedom present complex challenges. + +In this work, we explore a novel and challenging task: open-set language-guided dexterous grasping, as shown in Figure 1, where models are evaluated on objects and language instructions from both seen and unseen categories. This task poses a great challenge in ensuring that grasps are intention consistent with corresponding language instructions for unseen categories. We find that the main challenge lies in the huge gap between high-level natural language and low-level robot action spaces, which makes it difficult to generalize the ability of understanding intentions and grasping from the training domain to unseen categories. + +To solve the above challenge, we propose the Affordance Dexterous Grasp (AffordDexGrasp) framework, with the key insight of employing a new affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. Our affordance representation combines two characteristics: generalizable (generalizing to unseen categories based on language) and instructive (guiding dexterous grasp generation effectively). However, achieving these two characteristics is not trivial. For example, as shown in Figure 2, fine-grained contact information can effectively guide grasp generation or optimization [2, 9, 17, 55], but it is difficult to generalize to unseen categories. In contrast, coarse information, such as object parts, can be obtained from a pre-trained computer vision model [27, 41, 43], but it is too coarse to guide dexterous hand actions with higher degrees of freedom. + +To achieve these characteristics, we propose the Generalizable-Instructive Affordance by defining a general dexterous affordance that aligns with category-independent information, such as intention, object parts, and direction. As shown in Figure 2, the affordance represents the potential graspable regions of all grasps with the same semantics. In this way, the models do not need to learn complex dexterous contact patterns but instead focus on a general graspable area, which can be well aligned with category-agnostic semantic attributes and guide grasp generation effectively. + +To generate affordance and employ it to guide grasp gen + +eration, our framework consists of two cascaded generative models. The Affordance Flow Matching generates affordance maps in a generalizable manner based on language. And the Grasp Flow Matching generates dexterous grasp poses under the effective guidance of affordance. Moreover, we introduce a pre-understanding stage and a post-optimization stage to further boost generalization. Specifically, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intention to enhance generalization across diverse user commands. And we introduce an affordance-guided optimization to improve grasp quality while preserving consistency with user intentions. + +For evaluating our framework, we build a open-set tabletop language-guided dexterous grasp dataset, based on the language guided dexterous grasp dataset [51, 56]. We exclude specific categories from the training set to test the model's open-set generalization. Moreover, we provide high-quality rendered images to facilitate the usage by MLLM, and we extend the dataset to scene-level data to better simulate real-world environments. The comprehensive experiments are conducted on both the simulation open-set dataset and real-world environments. The results show that our framework can generate dexterous grasp with consistent intention and high quality in open set. + +# 2. Related work + +# 2.1. Dexterous Grasp + +Dexterous grasping is critical in robotics, which equips robots with human-like grasping capabilities. Some methods [29, 44, 48, 58] focus on grasp stability and quality, while recent studies explore task-oriented [50, 61] or language-guided grasping [21, 51] with specific semantic intention. For language-guided task, the data-driven methods achieve impressive performance in unseen object within seen categories by leveraging language conditioned generative model [4, 51]. However, we find that these models are hard to generalize for samples from unseen categories. And we find that the main challenge is the huge gap between language and grasp action. To solve this problems, we propose AffordDexGrasp framework with the insight that bridge the gap by generalizable-instructive affordance. + +# 2.2. Open-set Robot Grasp + +Exploring the performance of robotic models in open-set scenarios is crucial due to the diverse object categories in the real world and the high cost of data collection. For tasks that only focus on stable grasping, both parallel grippers [6, 52] and dexterous hands [47, 58] achieve impressive performance on open set. However, when it comes to considering task-oriented [45] or language-guided grasping [33], it becomes much more difficult. Some works in parallel grippers achieve this by using pre-trained visual ground + +![](images/9193775e0a6e1d7738d4c13e90f3d49834d3a841ac65775a8e753b31ef9a2707.jpg) +(a) Visualization of different affordance representations. + +![](images/011b755e22e552f5c4dc0863e4f757f92db30d9628035918f7dc6f7ea0fa733d.jpg) +(b) Quantitative Results Across Affordances. +Figure 2. Different affordance representations. (a) While contact map are too elaborate to generalize and object part are too coarse to guide grasping, our affrodacne achieve a balance. (b) While object part has a lower upper bound and contact shows significant degradation in generalization, only our affordance effectively achieve the balance (Top-1 indicates grasp intention consistency). + +ing models [22, 27, 41], implicit feature semantic fields [18, 32, 36], or multimodal large models [13, 14]. Compared to parallel grippers, dexterous hands have a significantly higher number of degrees of freedom, making the methods of parallel grippers difficult to transfer to dexterous hand [51]. In this paper, we explore the open-set language-guided dexterous grasping task and propose a novel framework to address it. + +# 2.3. Grasp Affordance + +Affordance is first introduced in [8], referring to environmental action possibilities. There are currently two types of grasp affordance: fine-grained contact [3, 42] and coarse-grained object segmentation [27]. The fine-grained information, such as contact areas, normals, or key points, is commonly used in hand-object interaction[2, 9, 59, 62]. However, this fine-grained information is difficult to generalize well to unseen categories. As a result, inaccurate contact information would lead to unreasonable grasping. On the other hand, We find through experiments that the coarse object segmentation [27] is too coarse to guide dexterous grasp generation with relatively high degrees of freedom. To solve this problem, we propose Generalizable-Instructive Affordance, which can generalize through the local structure of objects using category-independent clues and effectively guide grasp generation. + +# 3. Affordance Dexterous Grasp Framework + +Given the scene point cloud $\mathcal{O}$ , RGB images $\mathcal{I}$ and language command $\mathcal{C}$ as input, our goal is to generate intention aligned and high-quality dexterous grasps poses $\mathcal{G}^{dex} = (r,t,q)$ , where $r$ denotes the rotation, $t$ denotes the translation, and $q$ represents the joint angles of the dexterous hand. + +In this section, we first introduce a novel generalizable- + +instructive affordance representation (Section 3.1). Built on this affordance representation, we then propose the Affordance Dexterous Grasp framework (Section 3.2) including intention pre-understanding stage, Affordance and Grasp Flow Matching, and grasp post-optimization stage. + +# 3.1. Generalizable-Instructive Affordance + +We propose the generalizable-instructive affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. The affordances serve two primary objectives: 1) enable generalization by aligning object local structures with category-agnostic language semantics; and 2) provide instructive guidance for grasp generation through object affordance cues. However, there exists a trade-off in the generalization and instruction, and it is not trivial to achieve both objectives. + +Generalization vs. Instruction. While more fine-grained contact information can better guide grasp generation, it becomes harder to generalize to unseen categories. Fine-grained affordances typically compute the distance between object points and predefined elements, such as hand surface points or hand key points [19, 28]. However, we observe that these fine-gained representations are difficult to generalize to unseen category samples as shown in Figure 2. As the dexterous hand structure is complex and contact modes are varied, the contact maps are also elaborate and varies in space, which make it difficult to generalize. + +On the other hand, the information agnostic to dexterous hands, like object part segmentation, can be obtained in a generalizable manner by vision models [27, 60] or semantic feature fields [36]. However, the object part is too coarse for dexterous hand, as it may struggle to define the clear grasping area. For example, grasping the body of a mug from above and from the side involves the same part, but with different intentions. Moreover, this coarse-grained + +![](images/ce3bdadeea1d70e692f9a6fba428df19dc30ba03ff06e911153ef684b23341f4.jpg) +Figure 3. The pipeline of Affordance Dexterous Grasp framework. The inference pipeline includes three stages: 1) intention pre- understanding assisted by MLLM; 2) affordance flow matching for generating affordance base on MLLM output; 3) Grasp Flow Matching and Optimization for outputing grasp poses based on the affordance and MLLM outputs. In the training time, AFM and GFM are independently trained one after another. Transformer and Perceiver are attention-based interaction module for velocity vector field prediction. + +learning may hinder the model from learning an accurate grasping method during training. + +Generalizable-Instructive Affordance. To address this challenge, we propose a novel affordance representation that simultaneously enables both generalization and instruction. The key insight is to establish a correspondence between the general graspable affordance of objects and category-agnostic semantic attributes. Unlike contact maps that describe a particular hand contact area, our affordance represents all potential graspable regions that share the same semantic intention. This approach allows the affordance to generalize more effectively with category-agnostic semantic attributes from language, as it doesn't have to learn the complex distribution of hand contact. Meanwhile, our affordance serves as priors to enhance grasp generation by providing valuable graspable area cues. + +To obtain the ground truth of generalizable-instructive affordance, we first organize those grasps of one object with the same intention, contact parts and grasp direction into a semantic group. Then we calculate the distances between each scene points $\{o_i\}^M$ and hand surface points $\{h_i\}^N$ , to obtain the contact map $\Omega_{contact} = \{d_i\}_1^M$ for each grasp in the group, where $d_i = \min_{j=1\dots N}\| o_i - h_j\|_2$ . Then we unite contact areas of $k$ grasps in each group by calculating the minimum value of contact areas: $d_i^{min} = \{\min_k(d_i^k)\}^M$ . In order to make the affordance map more smooth, we apply Gaussian filtering to the union contact map: $a_i = \sum_{j\in \mathcal{N}(i)}\hat{w}_{ij}d_j^{min}$ and the weights are obtained by: + +$$ +\hat {w} _ {i j} = \frac {\exp \left(- \| o _ {i} - o _ {j} \| ^ {2} / (2 \sigma^ {2})\right)}{\sum_ {k \in \mathcal {N} (i)} \exp \left(- \| o _ {i} - o _ {k} \| ^ {2} / (2 \sigma^ {2})\right)}, \tag {1} +$$ + +where $\mathcal{N}(i)$ is the neighborhood of $o_i$ , $\sigma$ is set as the average nearest neighbor distance. Finally, we obtain the ground truth of generalizable-instructive affordance $\Omega = \{a_i\}^M$ . + +# 3.2. AffordDexGrasp Framework + +# 3.2.1. Intention Pre-Understanding + +The variability of user expression in the open world presents a challenge for models in generalizing and understanding user intentions. To address this, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intentions, by imputing the user language and the rendered RGB image into GPT-4o [15] for intention understanding and key cues extraction. We empirically find that the following cues are both helpful for grasping and easily reasoned by the MLLM: the object category, user intention (e.g., use it), the contacting part of the object (e.g., the handle of a mug), and the grasp direction. The key cues can be extract from language when the commands are clear, otherwise can be reasoned from the image. + +To improve the understanding ability of MLLM, we employ chain of thought (CoT) and visual prompt, and the prompts can be found in supplementary material A.4. Additionally, the grasp direction is simplified as a discretized direction in six predefined coordinates (front, back, left, right, up, down). We first ask MLLM to obtain the direction in image coordinates, and transform it to world coordinates by the camera pose. Then the direction vector is derived from the index of the nearest coordinate axis to form discretized direction. Finally, MLLM organizes this information into a concise sentence (e.g., "use the mug from the left by contacting the handle"). This compact and information-dense + +sentence structure allows subsequent models to better understand users' intentions. + +# 3.2.2. Affordance Flow Matching + +We introduce Affordance Flow Matching (AFM) to learn the affordance distribution efficiently, achieving intention aligned and generalizable affordance generation. AFM is built on the conditional flow matching model [1, 25, 30] aiming to learn the velocity vector field from the noised affordance map $\Omega_0$ to the target affordance map $\Omega_1$ , and the objective loss of AFM could be: + +$$ +\mathcal {L} _ {A F M} = \mathbb {E} _ {t, \Omega_ {0} \sim p _ {0}, \Omega_ {1} \sim p _ {1}} \| v _ {\eta} (\Omega_ {t}, t | c) - (\Omega_ {1} - \Omega_ {0}) \| ^ {2}, \tag {2} +$$ + +where $p_0$ is Gaussian distribution with zero mean and unit variance at time $t = 0$ , and $p_1$ is the target distribution as time $t = 1$ . $v_{\eta}(\Omega_t, t|c)$ is parameterized as a neural network with weights $\eta$ and $c$ is the input condition features. + +Specifically, the input scene object point cloud $\mathcal{O}$ is encoded by Pointnet++ [34], language is encoded by a pre-trained CLIP model [35], and the direction vector and time embedding are encoded by the Multilayer Perceptrons (MLP). Due to the permutation invariance of point level affordance, the noisy affordance are concatenated with object features to aware the position and semantic information. Then all features are fed into a Perceiver IO Transformer, following [16, 49]. The noisy affordance features are served as input and output array, and the features of language and direction are served as latent array. The outputs of the decoder go through an MLP to obtain the prediction of the flow $v_{\eta}(\Omega_t,t|c)\in R^M$ . Finally, in the inference time, we adopt multiple steps $1 / \Delta t$ following the equation: + +$$ +\hat {\Omega} _ {t + \Delta t} = \Omega_ {t} + \Delta t v _ {\eta} \left(\Omega_ {t}, t \mid c _ {1}\right), \quad t \in [ 0, 1). \tag {3} +$$ + +# 3.2.3. Grasp Flow Matching + +We introduce Grasp Flow Matching (GFM) to learn the grasp distribution, conditioned on the output of MLLM and AFM. GFM is also built upon the flow matching model, and the objective is to generate a parameterized dexterous hand pose $\mathcal{G}^{dex} = (r,t,q)$ , which represents the global rotation, translation, and joint angles. In the training time, grasp generation requires calculating explicit losses in 3D space, which involves obtaining the target dexterous poses, which go through forward kinematics to obtain hand configuration [53]. To obtain the target pose, we estimate it in one step during the training stage: + +$$ +\mathcal {G} _ {1} ^ {\hat {d e x}} = \mathcal {G} _ {t} ^ {d e x} + (1 - t) * v _ {\theta} (\mathcal {G} _ {t} ^ {d e x}, t | c _ {2}), \quad t \in [ 0, 1), (4) +$$ + +where the condition features $c_{2}$ are the concatenation of object affordance, language and direction features. And the object affordance features are obtained by feeding affordance with point clouds into PointNet++ [34]. + +The losses of grasp generation include 1) the grasp pose regression L2 loss; 2) the hand chamfer loss to minimize the discrepancies between the actual hand shapes; 3) the hand fingertip key point loss to minimize the distance between fingertip position. The loss function can be formulated as: + +$$ +\mathcal {L} _ {G F M} = \lambda_ {\text {p o s e}} \mathcal {L} _ {\text {p o s e}} + \lambda_ {\text {c h a m f e r}} \mathcal {L} _ {\text {c h a m f e r}} + \lambda_ {\text {t i p}} \mathcal {L} _ {\text {t i p}}, \tag {5} +$$ + +where $\lambda_{\text{pose}}, \lambda_{\text{chamfer}}$ and $\lambda_{\text{tip}}$ are loss weights. + +The object penetration loss, which is used to penalize the penetration of hands into objects, is excluded from grasp generation training, as it leads to significant challenges in the training process [51]. Employing the penetration loss can reduce object penetration, but negatively impacts both intention alignment and generation diversity. Therefore, to achieve semantic-aligned grasp generation, we exclude object penetration loss and introduce affordance-guided grasp optimization to reduce object penetration. + +# 3.2.4. Affordance-guided Grasp Optimization + +We introduce a non-parametric test-time adaptation (TTA), affordance-guided grasp optimization, to improve the grasp quality while maintaining intention consistency. Compared to learning-based TTA methods [17, 51], our nonparametric optimization does not suffer from performance degradation caused by out-of-domain open-set samples. + +We design optimization objectives based on that high-quality grasps should exhibit high grasp stability and intention alignment in grasping posture and contact areas. Specifically, our optimization objectives include: 1) affordance contact loss for constraining hand contact candidate points in contact with the affordance area: + +$$ +\mathcal {L} _ {\text {a f f - d i s t}} = \sum_ {i} \mathbb {I} \left(\left(d \left(p _ {i} ^ {c}\right) < \tau_ {1}\right) \vee \left(d \left(p _ {i} ^ {r}\right) < \tau_ {1}\right)\right) \cdot d \left(p _ {i} ^ {r}\right), \tag {6} +$$ + +where $d(p) = \min_{o \in \{o_j \in \mathcal{O} | a_j > \tau_2\}} \| p - o_j \|_2$ , and $o_j$ and $a_j$ represent the position and affordance values of object points, and $p_i^c$ and $p_i^r$ represent the hand contact candidates of the initial coarse hand and current refined hand. 2) affordance fingertip loss to keep the fingertip positions that are in contact with the affordance area unchanged. + +3) object penetration loss for punish hand-object penetration, 4) self-penetration loss to punish the penetration of hand it self. 5) joint limitation loss to punish out-of-limit joint angles. The optimization objectives can be formulated as: + +$$ +\min _ {\mathcal {G} ^ {d e x}} \left(\lambda_ {1} \mathcal {L} _ {\text {a f f - d i s t}} + \lambda_ {2} \mathcal {L} _ {\text {a f f - f i n g e r}} + \lambda_ {3} \mathcal {L} _ {\text {p e n}} + \lambda_ {4} \mathcal {L} _ {\text {s p e n}} + \lambda_ {5} \mathcal {L} _ {\text {j o i n t}}\right). \tag {7} +$$ + +# 4. Open-set Dexterous Grasp Dataset + +To support the evaluation of our framework, we build an open-set dexterous grasp dataset in table-top environment based on the language-guided dexterous grasp dataset + +
IntentionQualityDiversity
FID ↓CD ↓R - Precision ↑Suc. ↑Q1 ↑Pen. ↓δt ↑δr ↑δq ↑
Top1Top2Top3
Open Set A
ContactGen[28]0.4289.380.1640.2320.28911.6%1.71e-41.092.8618.439.2
Contact2Grasp[19]0.42610.20.2570.3370.39816.5%0.01720.6685.734.8448.3
GraspCVAE[40]0.3785.550.3090.4020.46921.9%0.01500.7094.364.6021.2
SceneDiffuser[12]0.3035.020.3970.4730.52329.1%0.01510.4877.379.4468.9
DexGYS[51]0.2974.630.3170.4010.46344.2%0.05120.3626.306.7954.7
Ours0.2313.810.4800.5880.66645.1%0.05310.2937.486.9861.5
Open Set B
ContactGen[28]0.4928.950.1960.2570.2996.04%1.30e-41.113.2618.745.5
Contact2Grasp[19]0.36911.90.2480.3180.36716.2%0.008010.7985.224.6340.7
GraspCVAE[40]0.3655.350.2970.3570.40324.2%0.004210.7383.965.4725.6
SceneDiffuser[12]0.2716.500.3020.3550.39330.7%0.008510.6835.416.1372.9
DexGYS[51]0.3583.400.2940.3580.40335.2%0.02200.6915.406.1159.0
Ours0.1622.950.5320.6090.65538.9%0.03520.3616.866.8456.3
+ +Table 1. Results on open-set datasets compared with the SOTA methods. + +[51, 56]. Our open-set dataset consists of 33 categories, 1536 objects, 1909 scenes, and 43,504 dexterous grasps for Shadow Hand [37] and Leap Hand [39]. + +- Open Set Data Split. To enable comprehensive evaluation, we construct Open sets A and B using two independent dataset splits, each with its own set of unseen categories. For each split, all categories are first labeled as either seen or unseen. Then, $80\%$ of the objects from the seen categories are used for training, while the remaining $20\%$ and all objects from the unseen categories are used for testing. +- Scene Construction. To make our dataset more practical, we generate scene data by placing objects in a tabletop environment using Blenderproc [5]. To prevent collisions between the hand and the table, objects are elevated using a shelf, and collision detection is performed to filter out invalid grasps. The physics engine is activated to ensure physically plausible object placements and grasping poses, enhancing dataset quality. +- Scene Point-Cloud and Image Rendering. The scene was captured using five RGB-D cameras: four positioned at elevated lateral angles and one directly above the object. The partial point clouds are merged into a global point cloud. To render realistic RGB images for MLLM processing, we use the texture generation model Paint3D [57] to apply realistic textures to all objects in our dataset. + +# 5. Experiment + +# 5.1. Evaluation Metrics + +We employ three types of metric to evaluate the ability of intention consistency, grasp quality and grasp diversity. + +- Evaluating Intention Consistency. We employ several metrics, including FID, R-Precision and Chamfer Distance. + +1) FID (Frechet Inception Distance) [11] measures the distribution similarity between generated grasps and ground truth. To extract grasp and instruction features, we train an object-grasp point cloud encoder and a language encoder by contrastive learning [10]. 2) R-Precision evaluates the semantic alignment between language instructions and generated grasps [10]. Specifically, for each generated grasp, we construct a pool of 32 samples, including its ground-truth instruction and randomly selected samples. We then compute and rank the cosine distances between the point cloud features and the language features of all samples in the pool. The average accuracy is computed at the top-1, top-2, and top-3 positions, that a retrieval is considered successful if the ground truth entry appears among the top-k candidates. 3) Chamfer Distance (CD) quantifies the discrepancy between the predicted hand point cloud and the closest target with same intention [51]. + +- Evaluating Grasp Quality. Following [46], we use success rate, denoted as $Suc$ , in Issac Gym [24] and Q1 [7] to assess grasp stability. Maximal penetration depth (cm), denoted as $Pen$ , is used to calculate the maximal penetration depth from the object point cloud to hand meshes. + +- Evaluating Grasp Diversity. We compute the Standard deviation of translation $\delta_t$ , rotation $\delta_r$ and joint angle $\delta_q$ of samples within the same scene observation. + +# 5.2. Implementation Details + +For training our framework, the number of epochs is set to 50 for AFM and 200 for GFM. For AFM, the loss weight of L2 loss is set to 1. For GFM, we set $\lambda_{\mathrm{pose}} = 10$ , $\lambda_{\mathrm{chamfer}} = 1$ , and $\lambda_{\mathrm{tip}} = 2$ . The batch size is set to 16 for AFM and 64 for GFM. The initial learning rate is $2.0 \times 10^{-4}$ , decaying + +
Open Set AOpen Set B
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
Necessity of Generalizable-Instructive Affordance
w/o affordance0.3385.090.3690.01800.5690.3224.440.3080.01940.452
w object part (pred)0.3135.060.3200.01380.5940.3163.790.3448.12e-30.721
w contact map (pred)0.3204.890.3800.01310.6050.2863.250.4089.25e-30.636
w our affordance (pred)0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w object part (GT)0.3065.050.3480.01980.5230.2713.790.3536.68e-30.526
w contact map (GT)0.1132.010.5980.02850.4940.1382.680.5940.01430.643
w our affordance (GT)0.1693.520.5090.01810.5450.1642.610.5500.01200.584
Effectiveness of Pre-Understanding Stage
w/o key cues extraction0.2583.990.4630.02220.5100.1863.420.5158.33e-30.702
w/o direction0.2544.070.4320.01620.5310.1852.850.5298.80e-30.679
w MLLM0.2423.790.4800.02400.5010.1762.760.5380.01500.612
Effectiveness of Affordance-based Optimization
w/o our optimization0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w penetration loss0.3367.070.3180.05050.2990.2415.760.4450.01470.719
w ContactNet [17]0.34912.30.2790.04340.1240.35511.10.2400.02060.129
w RefineNet [51]0.2493.960.4360.04870.3990.2032.920.4930.01490.657
w our optimization0.2313.810.4770.05310.2930.1622.950.5320.03500.361
+ +to $2.0 \times 10^{-5}$ using a cosine learning rate scheduler [31]. The Adam optimizer is used with a weight decay rate of $5.0 \times 10^{-6}$ . In the inference, the time step is set to 10 for AFM and 20 for GFM. For the optimization, $\lambda_{1} = 100$ , $\lambda_{2} = 10$ , $\lambda_{3} = 100$ , $\lambda_{4} = 10$ , $\lambda_{5} = 100$ . The number of optimization iterations is set to 200. All experiments are implemented using PyTorch on a single RTX 4090 GPU. + +# 5.3. Comparison with SOTA Methods + +We compare our methods with the SOTA methods, as shown in Table 1. The generic grasp methods are reproduced by concatenating the point cloud features and features of same language guidance with ours. The same encoders are employed for fair comparison, and the penetration loss is excluded to avoid learning difficulties according to [51]. The results show that our framework significantly outperforms all previous methods in terms of intention consistency and grasp quality. Our framework also achieves a reasonably high level of diversity, as excessive diversity may lead to unnatural postures. Previous language-guided methods fail to generalize well to unseen categories due to the huge gap between language and grasping actions. Similarly, contact-based methods don't perform well, as the contact maps are difficult to generalize. Additionally, our method outperforms other methods in the close set, as shown in Table 3. Overall, the results indicate that our framework achieves strong performance in generating intention-aligned, high-quality, and diverse grasps. + +Table 2. Ablation study for our framework. The results of first two experiment groups are obtained from model outputs without optimization for an intuitive and reasonable comparison. The results in each group should be compared with light yellow line (our default setting). + +
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
GraspCVAE0.2082.400.3950.01000.771
DexGYS0.08041.740.5900.05510.397
Ours0.02861.130.7790.05620.193
+ +Table 3. Experiment of close set with SOTA methods. + +# 5.4. Necessity of Our Affordance + +The results shown in Table 2 validate the key insight of our framework: using the generalizable-instructive affordance to bridge the gap between high-level language and low-level grasp actions. The first row presents the results without affordance, which fails to generalize to unseen categories. The subsequent two rows show the results under the condition of object parts and contact maps. Prediction refers to maps generated by models, while GT refers to using the corresponding ground truth. As discussed in Section 3.1, the results demonstrate that object parts are too coarse to guide the grasp effectively, even when using the ground truth. On the other hand, the contact map is too fine-grained to generalize to unseen categories. Only our affordance achieves a balance between generalization and instruction. + +Furthermore, our affordance representation and framework demonstrate good generalization in the one-shot setting, as shown in Table 4. For the one-shot experiments, we introduce several novel categories to the test set that differ significantly from the training set and add one object from each unseen category to the training set. The results show + +![](images/ff9a2f66754552d456ac72bbbff60e6905a37318b256e6f3dd9423ebe98e928d.jpg) + +![](images/3f519ccdeb3f8cce1538d3a8d844d1749f8b2eca6b89b3581c5217958209bdd1.jpg) +Figure 4. The visualization of generated affordance and dexterous grasp. The left top shows the zero-shot samples and the left bottom shows the one-shot samples in real world. The right top and right bottom show the zero-shot samples in simulation open set A and B. + +![](images/20e2a9ba5a3591a20b0edc4145b86ee5a4ee497d4aab6d4442b326c9b0445e98.jpg) + +![](images/28b195ad32641f84d0f9bde82224acb8275311f2a8d1cb4e448e26120b7e6e81.jpg) + +![](images/ac1873f58b02f2801fe3a262136a431df91ebba8cec69c0176c533ab0d678b72.jpg) + +![](images/1dd62dd124ddf7e64f1955ff8173d8767a04ed439a1245326a9536fc3af15034.jpg) + +![](images/a00adb8a0018a38db997a5970933dc866aff9dbab9265be93f27287912aef22e.jpg) + +![](images/84b7fd3620518bd178c053dcaf6f15a212d29d68958b0693bf3fd9a637a4ddf2.jpg) + +![](images/96c2c28e389b2de8bbf57d1efa143c6c32bf22e01ce16a2b11334d35f57a15b6.jpg) + +![](images/b300cb4db437cc4cc8e5ecb879262e4df3f28b483a1e046e5f7212bfa2eb08c4.jpg) + +![](images/60abc94610e4bbe048c189954d6febeffd4a21c5837fe88ee3fe5e1ec42db156.jpg) + +![](images/4ca27846da0469bc14a458e7ca54a29fbea4d7bd49fcdadf343f702007d02c9e.jpg) + +![](images/c6e5b63ce2eb82118e3d34a3431b21ac4d86652c837b1f2fdc673a8abdc2c2f2.jpg) + +![](images/a4c442c48f4bd7c7fa5e61c17c4aea7f8763c9515e6d4b7d05ba6354a2813538.jpg) + +![](images/4730945ebc7f3f20ec9572a97eff42a85aa112af89e346b347b33c939f6482c3.jpg) + +![](images/edff7035c278721fa4962391ed5e5982dc8427c3bfdab4715e651eb231a5d801.jpg) + +
Seen CateSimilar CateNovel Cate
Top1 ↑ CD ↓Top1 ↑ CD ↓Top1 ↑ CD ↓
Zero-Shot
w/o afford0.8272.610.2955.710.2197.82
Ours0.8801.060.4323.650.2466.42
One-Shot
w/o afford0.7673.170.4293.670.3426.64
Ours0.8700.990.6561.730.5863.03
+ +that our framework achieves a significant performance improvement, not only on samples from different categories with similar structures but also for novel categories. + +# 5.5. Effectiveness of Each Component + +Further ablation studies are conducted to evaluate each component of our framework. For the pre-understanding stage assisted by MLLM, w/o key cues extraction refers to directly inputting the user's commands into the model, while w/o direction means that the direction information is not utilized. Both designs enhance performance, and we believe that MLLM would be more beneficial in real-world applications due to its powerful generalization. For affordance-based optimization, our optimization improves grasp quality and maintains intention consistency, outperforming the learning-based methods ContactNet [17] and RefineNet [51], as well as the model which is trained with penetration loss. + +# 5.6. Real World Experiments + +The real-word experiments are conducted to verify the simulation-to-reality ability of our framework. We employ a Leap Hand, a Kinova Gen3 6DOF arms and an original wrist RGB-D camera of Kinova arm. And we collect several common objects in daily life, including unseen objects and unseen categories. In experiment, we synthesize the scene point cloud by taking several partial depth maps + +Table 4. Zero-shot and one-shot generalization of our framework. + +
DexGYS273647010
Ours105910610610
+ +Table 5. Real world experiments: Successes in 10 attempts. + +around the object. Then the scene point cloud, a RGB image and the user language command are fed into our framework to obtain the dexterous grasp pose. During execution, we first move the arm to a pre-grasp position, then synchronously move the joints of the robotic arm and the dexterous hand to reach the target pose. In evaluation, a grasp is considered as success if the hand can lift it up and the action is consistent with the given command. The results shown in Table 5 shows that our framework achieve higher success rate than previous method. + +# 6. Conclusions + +We believe that achieving generalizable dexterous grasps in open set is crucial within the deep learning and robotics communities. In this paper, we explore a novel task of open-set language-guided dexterous grasp. This task is challenging due to the huge gap between language and grasping actions, which hinders the model's generalization ability. We propose the AffordDexGrasp framework with the insight that using a new affordance representation, generalizable-instructive affordance, to solve this challenge. This affordance enables generalization to unseen categories by utilizing the object's local structure and category-agnostic semantic attributes, thus facilitating effective dexterous grasp generation. Based on this affordance, our framework introduces two flow matching based models for affordance generation and affordance-guided grasp generation. Moreover, we introduce a pre-understand stage and a pose-optimization stage to further boost generalization. We conduct extensive open-set experiments in both simulation and the real world, and the results show that our framework outperforms all SOTA methods. + +# 7. Acknowledgements + +This work was supported partially by NSFC (92470202, U21A20471), Guangdong NSF Project (No. 2023B151504 0025), Guangdong Key Research and Development Program (No. 2024B0101040004). + +# References + +[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 5 +[2] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 2, 3 +[3] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 3 +[4] Xiaoyun Chang and Yi Sun. Text2grasp: Grasp synthesis by text prompts of object grasping parts. arXiv preprint arXiv:2404.15189, 2024. 2 +[5] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Wendelin Knauer, Klaus H Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6 +[6] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 39(5):3929-3945, 2023. 2 +[7] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In Proceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., pages 2290-2295. IEEE, 1992. 6 +[8] James J Gibson. The theory of affordances:(1979). In *The people, place, and space reader*, pages 56-60. Routledge, 2014. 3 +[9] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1471-1481, 2021. 2, 3 +[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 6 +[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- + +rium. Advances in neural information processing systems, 30, 2017. 6 +[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 6 +[13] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 3 +[14] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3 +[15] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4 +[16] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 5 +[17] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2, 5, 7, 8 +[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 3 +[19] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3, 6 +[20] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 1 +[21] Kailin Li, Jingbo Wang, Lixin Yang, Cewu Lu, and Bo Dai. Semgrasp: Semantic grasp generation via language aligned discretization. In European Conference on Computer Vision, pages 109-127. Springer, 2024. 1, 2 +[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3 +[23] Samuel Li, Sarthak Bhagat, Joseph Campbell, Yaqi Xie, Woojun Kim, Katia Sycara, and Simon Stepputtis. Shape-grasp: Zero-shot task-oriented grasping with large language models through geometric decomposition. In 2024 + +IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10527-10534. IEEE, 2024. 2 +[24] Jacky Liang, Viktor Makoviychuk, Ankur Handa, Nuttapong Chentanez, Miles Macklin, and Dieter Fox. Gpu-accelerated robotic simulation for distributed reinforcement learning. In Conference on Robot Learning, pages 270-282. PMLR, 2018. 6 +[25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 5 +[26] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2 +[27] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2, 3 +[28] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 6 +[29] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2 +[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5 +[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7 +[32] Teli Ma, Zifan Wang, Jiaming Zhou, Mengmeng Wang, and Junwei Liang. Glover: Generalizable open-vocabulary affordance reasoning for task-oriented grasping. arXiv preprint arXiv:2411.12286, 2024. 3 +[33] Teli Ma, Jiaming Zhou, Zifan Wang, Ronghe Qiu, and Junwei Liang. Contrastive imitation learning for language-guided multi-task robotic manipulation. arXiv preprint arXiv:2406.09738, 2024. 2 +[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5 +[36] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2, 3 + +[37] ShadowHand. Shadowrobot. https://www.).shadowrobot.com/dexterous-hand-series/, 2005.6 +[38] Lin Shao, Fabio Ferreira, Mikael Jorda, Varun Nambiar, Jianlan Luo, Eugen Solowjow, Juan Aparicio Ojea, Oussama Khatib, and Jeannette Bohg. Unigrasp: Learning a unified model to grasp with multifingered robotic hands. IEEE Robotics and Automation Letters, 5(2):2286-2293, 2020. 2 +[39] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv preprint arXiv:2309.06440, 2023. 6 +[40] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 6 +[41] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 2, 3 +[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3 +[43] Ardian Umam, Cheng-Kun Yang, Min-Hung Chen, Jen-Hui Chuang, and Yen-Yu Lin. Partdistill: 3d shape part segmentation by vision-language model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2024. 2 +[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3891-3902, 2023. 2 +[45] An-Lan Wang, Nuo Chen, Kun-Yu Lin, Li Yuan-Ming, and Wei-Shi Zheng. Task-oriented 6-dof grasp pose detection in clutters. arXiv preprint arXiv:2502.16976, 2025. 2 +[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2, 6 +[47] Wenbo Wang, Fangyun Wei, Lei Zhou, Xi Chen, Lin Luo, Xiaohan Yi, Yizhong Zhang, Yaobo Liang, Chang Xu, Yan Lu, et al. Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping. arXiv preprint arXiv:2412.02699, 2024. 2 +[48] Yan-Kang Wang, Chengyi Xing, Yi-Lin Wei, Xiao-Ming Wu, and Wei-Shi Zheng. Single-view scene point cloud human grasp generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-841, 2024. 2 +[49] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: + +Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 5 +[50] Wei Wei, Peng Wang, Sizhe Wang, Yongkang Luo, Wanyi Li, Daheng Li, Yayu Huang, and Haonan Duan. Learning human-like functional grasping for multi-finger hands from few demonstrations. IEEE Transactions on Robotics, 2024. 2 +[51] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. Advances in Neural Information Processing Systems, 37:46881-46907, 2025. 1, 2, 3, 5, 6, 7, 8 +[52] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-dof grasp detection. In European Conference on Computer Vision, pages 357-375. Springer, 2024. 2 +[53] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5 +[54] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4737-4746, 2023. 2 +[55] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 2 +[56] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20953-20962, 2022. 2, 6 +[57] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4252-4262, 2024. 6 +[58] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. In 8th Annual Conference on Robot Learning, 2024. 2 +[59] Zhongqun Zhang, Hengfei Wang, Ziwei Yu, Yihua Cheng, Angela Yao, and Hyung Jin Chang. Nl2contact: Natural language guided 3d hand-object contact modeling with diffusion model. In European Conference on Computer Vision, pages 284-300. Springer, 2024. 1, 3 +[60] Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yun-hao Fang, and Hao Su. Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmenta + +tion and maximum likelihood estimation. arXiv preprint arXiv:2312.03015, 2023. 3 +[61] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15741-15751, 2021. 2 +[62] Binghui Zuo, Zimeng Zhao, Wenqian Sun, Xiaohan Yuan, Zhipeng Yu, and Yangang Wang. Graspondiff: Grasping generation for hand-object interaction with multimodal guided diffusion. IEEE Transactions on Visualization and Computer Graphics, 2024. 3 \ No newline at end of file diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/images.zip b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c18442d8cad1b4901f16e8823c2a5adfd26b668f --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d752548c5aa4ef1fe4aeae94e25c3c09d4f6bbdded633b8443d1423ae81e18b +size 834839 diff --git a/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/layout.json b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f37ee96195cf1b212cf67a23faa9c566e1bc07f8 --- /dev/null +++ b/2025/AffordDexGrasp_ Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance/layout.json @@ -0,0 +1,10250 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 138 + ], + "type": "text", + "content": "AffordDexGrasp: Open-set Language-guided Dexterous Grasp with Generalizable-Instructive Affordance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 160, + 463, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 160, + 463, + 189 + ], + "spans": [ + { + "bbox": [ + 146, + 160, + 463, + 189 + ], + "type": "text", + "content": "Yi-Lin Wei\\*, Mu Lin\\*, Yuhao Lin\\*, Jian-Jian Jiang\\*, Xiao-Ming Wu\\*, Ling-An Zeng\\*, Wei-Shi Zheng†,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 114, + 189, + 495, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 189, + 495, + 203 + ], + "spans": [ + { + "bbox": [ + 114, + 189, + 495, + 203 + ], + "type": "text", + "content": "1 School of Computer Science and Engineering, Sun Yat-sen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 203, + 544, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 203, + 544, + 217 + ], + "spans": [ + { + "bbox": [ + 65, + 203, + 544, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 203, + 544, + 217 + ], + "type": "text", + "content": " Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 220, + 527, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 220, + 527, + 231 + ], + "spans": [ + { + "bbox": [ + 82, + 220, + 527, + 231 + ], + "type": "text", + "content": "{weiylin5, linm67, linyh96, jiangjj35, wuxm65, zenglan3}@mail2.sysu.edu.cn wszheng@ieee.org" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 183, + 232, + 423, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 232, + 423, + 245 + ], + "spans": [ + { + "bbox": [ + 183, + 232, + 423, + 245 + ], + "type": "text", + "content": "https://isee-laboratory.github.io/AffordDexGrasp/" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 270, + 117, + 312 + ], + "blocks": [ + { + "bbox": [ + 66, + 260, + 109, + 270 + ], + "lines": [ + { + "bbox": [ + 66, + 260, + 109, + 270 + ], + "spans": [ + { + "bbox": [ + 66, + 260, + 109, + 270 + ], + "type": "text", + "content": "Training Data" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 270, + 117, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 270, + 117, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 270, + 117, + 312 + ], + "type": "image", + "image_path": "851ee5ad1020d907eb07085e67cd7c7004423b3e13caab0a47d222611228b900.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 314, + 179, + 323 + ], + "lines": [ + { + "bbox": [ + 59, + 314, + 179, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 314, + 179, + 323 + ], + "type": "text", + "content": "Use the teapot with grasping the handle from the right." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 121, + 270, + 241, + 312 + ], + "blocks": [ + { + "bbox": [ + 121, + 260, + 239, + 269 + ], + "lines": [ + { + "bbox": [ + 121, + 260, + 239, + 269 + ], + "spans": [ + { + "bbox": [ + 121, + 260, + 239, + 269 + ], + "type": "text", + "content": "Generalizable-Instructive Affordance" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 270, + 241, + 312 + ], + "lines": [ + { + "bbox": [ + 121, + 270, + 241, + 312 + ], + "spans": [ + { + "bbox": [ + 121, + 270, + 241, + 312 + ], + "type": "image", + "image_path": "b0b67ec285c8d3e430274e6f693d2a855cc9ed20761f1e39c26ff3f918963574.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 314, + 301, + 323 + ], + "lines": [ + { + "bbox": [ + 183, + 314, + 301, + 323 + ], + "spans": [ + { + "bbox": [ + 183, + 314, + 301, + 323 + ], + "type": "text", + "content": "Use the mug by contacting the handle from the right." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 244, + 270, + 302, + 312 + ], + "blocks": [ + { + "bbox": [ + 247, + 260, + 299, + 270 + ], + "lines": [ + { + "bbox": [ + 247, + 260, + 299, + 270 + ], + "spans": [ + { + "bbox": [ + 247, + 260, + 299, + 270 + ], + "type": "text", + "content": "Unseen Category" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 270, + 302, + 312 + ], + "lines": [ + { + "bbox": [ + 244, + 270, + 302, + 312 + ], + "spans": [ + { + "bbox": [ + 244, + 270, + 302, + 312 + ], + "type": "image", + "image_path": "dadea84cdf4e1e584861d2ed6c439c2ffb3f7b23822cb36b8cfebabbca47d01f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 244, + 314, + 301, + 323 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 314, + 301, + 323 + ], + "spans": [ + { + "bbox": [ + 244, + 314, + 301, + 323 + ], + "type": "text", + "content": "the handle from the right." + } + ] + } + ], + "index": 16, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 270, + 365, + 312 + ], + "blocks": [ + { + "bbox": [ + 308, + 270, + 365, + 312 + ], + "lines": [ + { + "bbox": [ + 308, + 270, + 365, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 270, + 365, + 312 + ], + "type": "image", + "image_path": "9d1284c5a41b999c1fa0d9a876aa7b67a5155114d740146f66d70f82efe6f90f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 314, + 425, + 322 + ], + "lines": [ + { + "bbox": [ + 309, + 314, + 425, + 322 + ], + "spans": [ + { + "bbox": [ + 309, + 314, + 425, + 322 + ], + "type": "text", + "content": "Use the mug by contacting the handle from the right." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 370, + 270, + 427, + 312 + ], + "blocks": [ + { + "bbox": [ + 398, + 260, + 460, + 269 + ], + "lines": [ + { + "bbox": [ + 398, + 260, + 460, + 269 + ], + "spans": [ + { + "bbox": [ + 398, + 260, + 460, + 269 + ], + "type": "text", + "content": "Real World Open set" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 270, + 427, + 312 + ], + "lines": [ + { + "bbox": [ + 370, + 270, + 427, + 312 + ], + "spans": [ + { + "bbox": [ + 370, + 270, + 427, + 312 + ], + "type": "image", + "image_path": "62f0115f32b1dadb8a267bd82ddab83bc8d8d235dbc614cccd7d20f96c3af95e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 314, + 425, + 322 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 314, + 425, + 322 + ], + "spans": [ + { + "bbox": [ + 370, + 314, + 425, + 322 + ], + "type": "text", + "content": "the handle from the right." + } + ] + } + ], + "index": 21, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 432, + 270, + 490, + 312 + ], + "blocks": [ + { + "bbox": [ + 432, + 270, + 490, + 312 + ], + "lines": [ + { + "bbox": [ + 432, + 270, + 490, + 312 + ], + "spans": [ + { + "bbox": [ + 432, + 270, + 490, + 312 + ], + "type": "image", + "image_path": "36f051009c4c808f8bee163ea1020aa4c64b2b3944f830f5be3cd900860fb987.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 314, + 548, + 322 + ], + "lines": [ + { + "bbox": [ + 433, + 314, + 548, + 322 + ], + "spans": [ + { + "bbox": [ + 433, + 314, + 548, + 322 + ], + "type": "text", + "content": "Use the kettle by contacting the handle from the left." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 493, + 270, + 551, + 312 + ], + "blocks": [ + { + "bbox": [ + 493, + 270, + 551, + 312 + ], + "lines": [ + { + "bbox": [ + 493, + 270, + 551, + 312 + ], + "spans": [ + { + "bbox": [ + 493, + 270, + 551, + 312 + ], + "type": "image", + "image_path": "62c1e933b3455da5c5fb91bdeb551d79a96a0cd29356b6750e0e51a6177a730a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 494, + 314, + 549, + 322 + ], + "lines": [ + { + "bbox": [ + 494, + 314, + 549, + 322 + ], + "spans": [ + { + "bbox": [ + 494, + 314, + 549, + 322 + ], + "type": "text", + "content": "the handle from the left." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 59, + 327, + 117, + 370 + ], + "blocks": [ + { + "bbox": [ + 59, + 327, + 117, + 370 + ], + "lines": [ + { + "bbox": [ + 59, + 327, + 117, + 370 + ], + "spans": [ + { + "bbox": [ + 59, + 327, + 117, + 370 + ], + "type": "image", + "image_path": "016528dfbbad3517223b8ddcfa626f092ec44f3cdece38a5012b582eaafdab33.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 373, + 173, + 381 + ], + "lines": [ + { + "bbox": [ + 64, + 373, + 173, + 381 + ], + "spans": [ + { + "bbox": [ + 64, + 373, + 173, + 381 + ], + "type": "text", + "content": "From the right, grasp the handle to use the knife." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 120, + 327, + 241, + 370 + ], + "blocks": [ + { + "bbox": [ + 120, + 327, + 241, + 370 + ], + "lines": [ + { + "bbox": [ + 120, + 327, + 241, + 370 + ], + "spans": [ + { + "bbox": [ + 120, + 327, + 241, + 370 + ], + "type": "image", + "image_path": "f96d2bf8632d5929131b44065824020067f63d86cd6fd2d35d5e756d28759f1b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 373, + 299, + 381 + ], + "lines": [ + { + "bbox": [ + 184, + 373, + 299, + 381 + ], + "spans": [ + { + "bbox": [ + 184, + 373, + 299, + 381 + ], + "type": "text", + "content": "Use the frying pan and grasp handle from the right." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 308, + 327, + 365, + 371 + ], + "blocks": [ + { + "bbox": [ + 308, + 327, + 365, + 371 + ], + "lines": [ + { + "bbox": [ + 308, + 327, + 365, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 365, + 371 + ], + "type": "image", + "image_path": "abeb627b88a11ef8decd7dcea31fc4220c6d6cb0e5a618971168f67889b8d04f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 373, + 422, + 381 + ], + "lines": [ + { + "bbox": [ + 311, + 373, + 422, + 381 + ], + "spans": [ + { + "bbox": [ + 311, + 373, + 422, + 381 + ], + "type": "text", + "content": "Use the hammer and grasp handle from the back." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 369, + 327, + 427, + 371 + ], + "blocks": [ + { + "bbox": [ + 369, + 327, + 427, + 371 + ], + "lines": [ + { + "bbox": [ + 369, + 327, + 427, + 371 + ], + "spans": [ + { + "bbox": [ + 369, + 327, + 427, + 371 + ], + "type": "image", + "image_path": "36b2889101237f181baaf44b132872e00bfe07dd6ab466ba51793e24018835bb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 370, + 373, + 422, + 380 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 373, + 422, + 380 + ], + "spans": [ + { + "bbox": [ + 370, + 373, + 422, + 380 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 370, + 373, + 422, + 380 + ], + "type": "text", + "content": " handle from the back." + } + ] + } + ], + "index": 33, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 432, + 327, + 490, + 371 + ], + "blocks": [ + { + "bbox": [ + 432, + 327, + 490, + 371 + ], + "lines": [ + { + "bbox": [ + 432, + 327, + 490, + 371 + ], + "spans": [ + { + "bbox": [ + 432, + 327, + 490, + 371 + ], + "type": "image", + "image_path": "c0c565a626041c2e0fc3dfe39d567661895d7c1f3da795b65a45c924ea50d405.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 373, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 433, + 373, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 433, + 373, + 548, + 381 + ], + "type": "text", + "content": "Use the frying pan and grasp handle from the back." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 59, + 387, + 117, + 430 + ], + "blocks": [ + { + "bbox": [ + 59, + 387, + 117, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 387, + 117, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 387, + 117, + 430 + ], + "type": "image", + "image_path": "a498eae91c628b0323958faa85ef6be48b285cbbb1dbacc41c260b5c08b20889.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 431, + 156, + 439 + ], + "lines": [ + { + "bbox": [ + 81, + 431, + 156, + 439 + ], + "spans": [ + { + "bbox": [ + 81, + 431, + 156, + 439 + ], + "type": "text", + "content": "Hold the bottle body from the left." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "type": "text", + "content": "Figure 1. Open-set Language-guided Dexterous Grasp. Our framework bridges the gap between language and grasp actions through Generalizable-Instructive Affordance, which enables cross-category generalization via category-agnostic cues and graspable local structure. Remarkably, our framework demonstrates strong generalization without requiring extra real training data in real-world experiments." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 121, + 388, + 241, + 430 + ], + "blocks": [ + { + "bbox": [ + 121, + 388, + 241, + 430 + ], + "lines": [ + { + "bbox": [ + 121, + 388, + 241, + 430 + ], + "spans": [ + { + "bbox": [ + 121, + 388, + 241, + 430 + ], + "type": "image", + "image_path": "0dde91e9c69ce33f9ced60433b0c46507f840a2e36a30cf285f1bc9486980b59.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 431, + 293, + 439 + ], + "lines": [ + { + "bbox": [ + 192, + 431, + 293, + 439 + ], + "spans": [ + { + "bbox": [ + 192, + 431, + 293, + 439 + ], + "type": "text", + "content": "Hold the body of trigger sprayer from the left." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 244, + 387, + 302, + 430 + ], + "blocks": [ + { + "bbox": [ + 244, + 387, + 302, + 430 + ], + "lines": [ + { + "bbox": [ + 244, + 387, + 302, + 430 + ], + "spans": [ + { + "bbox": [ + 244, + 387, + 302, + 430 + ], + "type": "image", + "image_path": "1b6e6acbc46cc46a2d511f8e9035eb56c0b873c5d846abdef490571e9b457a28.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 308, + 387, + 365, + 430 + ], + "blocks": [ + { + "bbox": [ + 308, + 387, + 365, + 430 + ], + "lines": [ + { + "bbox": [ + 308, + 387, + 365, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 365, + 430 + ], + "type": "image", + "image_path": "af28b5de35957a31e3e694d3944b8b899dd3b8ff815f56d30bc6614c0b3a1141.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 431, + 417, + 439 + ], + "lines": [ + { + "bbox": [ + 316, + 431, + 417, + 439 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 417, + 439 + ], + "type": "text", + "content": "Hold the body of trigger sprayer from the left." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 369, + 387, + 427, + 430 + ], + "blocks": [ + { + "bbox": [ + 369, + 387, + 427, + 430 + ], + "lines": [ + { + "bbox": [ + 369, + 387, + 427, + 430 + ], + "spans": [ + { + "bbox": [ + 369, + 387, + 427, + 430 + ], + "type": "image", + "image_path": "3ea3f3f8709b245ac99e621e6ccc04b7b29b8a7afa3aa85dd76d55a88114bcec.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 370, + 431, + 417, + 439 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 431, + 417, + 439 + ], + "spans": [ + { + "bbox": [ + 370, + 431, + 417, + 439 + ], + "type": "text", + "content": "sprayer from the left." + } + ] + } + ], + "index": 44, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 432, + 387, + 490, + 430 + ], + "blocks": [ + { + "bbox": [ + 432, + 387, + 490, + 430 + ], + "lines": [ + { + "bbox": [ + 432, + 387, + 490, + 430 + ], + "spans": [ + { + "bbox": [ + 432, + 387, + 490, + 430 + ], + "type": "image", + "image_path": "b31d3b61d8aadbe3b900b8483b8f9a14ce5a244ad45f3adc4de9e2c7b882131a.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 431, + 542, + 439 + ], + "lines": [ + { + "bbox": [ + 440, + 431, + 542, + 439 + ], + "spans": [ + { + "bbox": [ + 440, + 431, + 542, + 439 + ], + "type": "text", + "content": "Hold the body of the lotion pump from the left." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "bbox": [ + 152, + 491, + 200, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 491, + 200, + 504 + ], + "spans": [ + { + "bbox": [ + 152, + 491, + 200, + 504 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 54, + 517, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 517, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 297, + 685 + ], + "type": "text", + "content": "Language-guided robot dexterous generation enables robots to grasp and manipulate objects based on human commands. However, previous data-driven methods are hard to understand intention and execute grasping with unseen categories in the open set. In this work, we explore a new task, Open-set Language-guided Dexterous Grasp, and find that the main challenge is the huge gap between high-level human language semantics and low-level robot actions. To solve this problem, we propose an Affordance Dexterous Grasp (AffordDexGrasp) framework, with the insight of bridging the gap with a new generalizable-instructive affordance representation. This affordance can generalize to unseen categories by leveraging the object's local structure and category-agnostic semantic attributes, thereby ef" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 493, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 601 + ], + "type": "text", + "content": "fectively guiding dexterous grasp generation. Built upon the affordance, our framework introduces Affordance Flow Matching (AFM) for affordance generation with language as input, and Grasp Flow Matching (GFM) for generating dexterous grasp with affordance as input. To evaluate our framework, we build an open-set table-top language-guided dexterous grasp dataset. Extensive experiments in the simulation and real worlds show that our framework surpasses all previous methods in open-set generalization." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 314, + 622, + 394, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 622, + 394, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 622, + 394, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "text", + "content": "Achieving generalizable robot dexterous grasping is an important goal in the fields of robotics and computer vision, with exciting potential applications in human-robot interaction and robot manipulation." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 691, + 555, + 715 + ], + "type": "text", + "content": "Recent works explore the task of language-guided dexterous grasp generation[20, 21, 51, 59], aiming to enable" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 694, + 135, + 704 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 144, + 714 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11818" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "dexterous hands to perform actions based on language instructions, going beyond previous works that focus on stable grasping [26, 38, 46, 54]. The typical language-guided approaches employ language as the condition of generative models to predict hand parameters, achieving impressive performance on objects within known categories [4, 51]. However, in the open real world, there are many categories that may not appear during training, and the cost of collecting data for dexterous hands is quite expensive. Therefore, the open-set generalization on unseen category samples is crucial for robot grasp. While previous works on parallel grippers explore open-set tasks [23, 36], there is limited research on open-set generalization for dexterous hands, as their significantly higher degrees of freedom present complex challenges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 294, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 294, + 385 + ], + "type": "text", + "content": "In this work, we explore a novel and challenging task: open-set language-guided dexterous grasping, as shown in Figure 1, where models are evaluated on objects and language instructions from both seen and unseen categories. This task poses a great challenge in ensuring that grasps are intention consistent with corresponding language instructions for unseen categories. We find that the main challenge lies in the huge gap between high-level natural language and low-level robot action spaces, which makes it difficult to generalize the ability of understanding intentions and grasping from the training domain to unseen categories." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 387, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 578 + ], + "type": "text", + "content": "To solve the above challenge, we propose the Affordance Dexterous Grasp (AffordDexGrasp) framework, with the key insight of employing a new affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. Our affordance representation combines two characteristics: generalizable (generalizing to unseen categories based on language) and instructive (guiding dexterous grasp generation effectively). However, achieving these two characteristics is not trivial. For example, as shown in Figure 2, fine-grained contact information can effectively guide grasp generation or optimization [2, 9, 17, 55], but it is difficult to generalize to unseen categories. In contrast, coarse information, such as object parts, can be obtained from a pre-trained computer vision model [27, 41, 43], but it is too coarse to guide dexterous hand actions with higher degrees of freedom." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "content": "To achieve these characteristics, we propose the Generalizable-Instructive Affordance by defining a general dexterous affordance that aligns with category-independent information, such as intention, object parts, and direction. As shown in Figure 2, the affordance represents the potential graspable regions of all grasps with the same semantics. In this way, the models do not need to learn complex dexterous contact patterns but instead focus on a general graspable area, which can be well aligned with category-agnostic semantic attributes and guide grasp generation effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 714 + ], + "type": "text", + "content": "To generate affordance and employ it to guide grasp gen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "eration, our framework consists of two cascaded generative models. The Affordance Flow Matching generates affordance maps in a generalizable manner based on language. And the Grasp Flow Matching generates dexterous grasp poses under the effective guidance of affordance. Moreover, we introduce a pre-understanding stage and a post-optimization stage to further boost generalization. Specifically, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intention to enhance generalization across diverse user commands. And we introduce an affordance-guided optimization to improve grasp quality while preserving consistency with user intentions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 217, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 361 + ], + "type": "text", + "content": "For evaluating our framework, we build a open-set tabletop language-guided dexterous grasp dataset, based on the language guided dexterous grasp dataset [51, 56]. We exclude specific categories from the training set to test the model's open-set generalization. Moreover, we provide high-quality rendered images to facilitate the usage by MLLM, and we extend the dataset to scene-level data to better simulate real-world environments. The comprehensive experiments are conducted on both the simulation open-set dataset and real-world environments. The results show that our framework can generate dexterous grasp with consistent intention and high quality in open set." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 373, + 399, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 399, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 399, + 385 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 392, + 416, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 416, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 416, + 405 + ], + "type": "text", + "content": "2.1. Dexterous Grasp" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 411, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 578 + ], + "type": "text", + "content": "Dexterous grasping is critical in robotics, which equips robots with human-like grasping capabilities. Some methods [29, 44, 48, 58] focus on grasp stability and quality, while recent studies explore task-oriented [50, 61] or language-guided grasping [21, 51] with specific semantic intention. For language-guided task, the data-driven methods achieve impressive performance in unseen object within seen categories by leveraging language conditioned generative model [4, 51]. However, we find that these models are hard to generalize for samples from unseen categories. And we find that the main challenge is the huge gap between language and grasp action. To solve this problems, we propose AffordDexGrasp framework with the insight that bridge the gap by generalizable-instructive affordance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 443, + 601 + ], + "type": "text", + "content": "2.2. Open-set Robot Grasp" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 715 + ], + "type": "text", + "content": "Exploring the performance of robotic models in open-set scenarios is crucial due to the diverse object categories in the real world and the high cost of data collection. For tasks that only focus on stable grasping, both parallel grippers [6, 52] and dexterous hands [47, 58] achieve impressive performance on open set. However, when it comes to considering task-oriented [45] or language-guided grasping [33], it becomes much more difficult. Some works in parallel grippers achieve this by using pre-trained visual ground" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11819" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 347, + 217 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 347, + 217 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 347, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 347, + 217 + ], + "type": "image", + "image_path": "9193775e0a6e1d7738d4c13e90f3d49834d3a841ac65775a8e753b31ef9a2707.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 220, + 295, + 229 + ], + "lines": [ + { + "bbox": [ + 113, + 220, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 113, + 220, + 295, + 229 + ], + "type": "text", + "content": "(a) Visualization of different affordance representations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 350, + 78, + 548, + 216 + ], + "blocks": [ + { + "bbox": [ + 350, + 78, + 548, + 216 + ], + "lines": [ + { + "bbox": [ + 350, + 78, + 548, + 216 + ], + "spans": [ + { + "bbox": [ + 350, + 78, + 548, + 216 + ], + "type": "image", + "image_path": "011b755e22e552f5c4dc0863e4f757f92db30d9628035918f7dc6f7ea0fa733d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 219, + 529, + 229 + ], + "lines": [ + { + "bbox": [ + 384, + 219, + 529, + 229 + ], + "spans": [ + { + "bbox": [ + 384, + 219, + 529, + 229 + ], + "type": "text", + "content": "(b) Quantitative Results Across Affordances." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 239, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 273 + ], + "type": "text", + "content": "Figure 2. Different affordance representations. (a) While contact map are too elaborate to generalize and object part are too coarse to guide grasping, our affrodacne achieve a balance. (b) While object part has a lower upper bound and contact shows significant degradation in generalization, only our affordance effectively achieve the balance (Top-1 indicates grasp intention consistency)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 293, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 296, + 388 + ], + "type": "text", + "content": "ing models [22, 27, 41], implicit feature semantic fields [18, 32, 36], or multimodal large models [13, 14]. Compared to parallel grippers, dexterous hands have a significantly higher number of degrees of freedom, making the methods of parallel grippers difficult to transfer to dexterous hand [51]. In this paper, we explore the open-set language-guided dexterous grasping task and propose a novel framework to address it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 163, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 163, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 163, + 411 + ], + "type": "text", + "content": "2.3. Grasp Affordance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 296, + 608 + ], + "type": "text", + "content": "Affordance is first introduced in [8], referring to environmental action possibilities. There are currently two types of grasp affordance: fine-grained contact [3, 42] and coarse-grained object segmentation [27]. The fine-grained information, such as contact areas, normals, or key points, is commonly used in hand-object interaction[2, 9, 59, 62]. However, this fine-grained information is difficult to generalize well to unseen categories. As a result, inaccurate contact information would lead to unreasonable grasping. On the other hand, We find through experiments that the coarse object segmentation [27] is too coarse to guide dexterous grasp generation with relatively high degrees of freedom. To solve this problem, we propose Generalizable-Instructive Affordance, which can generalize through the local structure of objects using category-independent clues and effectively guide grasp generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 282, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 282, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 282, + 634 + ], + "type": "text", + "content": "3. Affordance Dexterous Grasp Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": "Given the scene point cloud " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": ", RGB images " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " and language command " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " as input, our goal is to generate intention aligned and high-quality dexterous grasps poses " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{dex} = (r,t,q)" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " denotes the rotation, " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " denotes the translation, and " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " represents the joint angles of the dexterous hand." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "In this section, we first introduce a novel generalizable-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 354 + ], + "type": "text", + "content": "instructive affordance representation (Section 3.1). Built on this affordance representation, we then propose the Affordance Dexterous Grasp framework (Section 3.2) including intention pre-understanding stage, Affordance and Grasp Flow Matching, and grasp post-optimization stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 512, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 512, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 512, + 372 + ], + "type": "text", + "content": "3.1. Generalizable-Instructive Affordance" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 485 + ], + "type": "text", + "content": "We propose the generalizable-instructive affordance as an intermediate representation to bridge the gap between high-level language and low-level grasp actions. The affordances serve two primary objectives: 1) enable generalization by aligning object local structures with category-agnostic language semantics; and 2) provide instructive guidance for grasp generation through object affordance cues. However, there exists a trade-off in the generalization and instruction, and it is not trivial to achieve both objectives." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "content": "Generalization vs. Instruction. While more fine-grained contact information can better guide grasp generation, it becomes harder to generalize to unseen categories. Fine-grained affordances typically compute the distance between object points and predefined elements, such as hand surface points or hand key points [19, 28]. However, we observe that these fine-gained representations are difficult to generalize to unseen category samples as shown in Figure 2. As the dexterous hand structure is complex and contact modes are varied, the contact maps are also elaborate and varies in space, which make it difficult to generalize." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "On the other hand, the information agnostic to dexterous hands, like object part segmentation, can be obtained in a generalizable manner by vision models [27, 60] or semantic feature fields [36]. However, the object part is too coarse for dexterous hand, as it may struggle to define the clear grasping area. For example, grasping the body of a mug from above and from the side involves the same part, but with different intentions. Moreover, this coarse-grained" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11820" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 259 + ], + "type": "image", + "image_path": "ce3bdadeea1d70e692f9a6fba428df19dc30ba03ff06e911153ef684b23341f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 313 + ], + "type": "text", + "content": "Figure 3. The pipeline of Affordance Dexterous Grasp framework. The inference pipeline includes three stages: 1) intention pre- understanding assisted by MLLM; 2) affordance flow matching for generating affordance base on MLLM output; 3) Grasp Flow Matching and Optimization for outputing grasp poses based on the affordance and MLLM outputs. In the training time, AFM and GFM are independently trained one after another. Transformer and Perceiver are attention-based interaction module for velocity vector field prediction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 294, + 357 + ], + "type": "text", + "content": "learning may hinder the model from learning an accurate grasping method during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 357, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 523 + ], + "type": "text", + "content": "Generalizable-Instructive Affordance. To address this challenge, we propose a novel affordance representation that simultaneously enables both generalization and instruction. The key insight is to establish a correspondence between the general graspable affordance of objects and category-agnostic semantic attributes. Unlike contact maps that describe a particular hand contact area, our affordance represents all potential graspable regions that share the same semantic intention. This approach allows the affordance to generalize more effectively with category-agnostic semantic attributes from language, as it doesn't have to learn the complex distribution of hand contact. Meanwhile, our affordance serves as priors to enhance grasp generation by providing valuable graspable area cues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": "To obtain the ground truth of generalizable-instructive affordance, we first organize those grasps of one object with the same intention, contact parts and grasp direction into a semantic group. Then we calculate the distances between each scene points " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\{o_i\\}^M" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": " and hand surface points " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\{h_i\\}^N" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": ", to obtain the contact map " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Omega_{contact} = \\{d_i\\}_1^M" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": " for each grasp in the group, where " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "d_i = \\min_{j=1\\dots N}\\| o_i - h_j\\|_2" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": ". Then we unite contact areas of " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": " grasps in each group by calculating the minimum value of contact areas: " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "d_i^{min} = \\{\\min_k(d_i^k)\\}^M" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": ". In order to make the affordance map more smooth, we apply Gaussian filtering to the union contact map: " + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "inline_equation", + "content": "a_i = \\sum_{j\\in \\mathcal{N}(i)}\\hat{w}_{ij}d_j^{min}" + }, + { + "bbox": [ + 55, + 524, + 295, + 677 + ], + "type": "text", + "content": " and the weights are obtained by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 686, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 686, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 85, + 686, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\hat {w} _ {i j} = \\frac {\\exp \\left(- \\| o _ {i} - o _ {j} \\| ^ {2} / (2 \\sigma^ {2})\\right)}{\\sum_ {k \\in \\mathcal {N} (i)} \\exp \\left(- \\| o _ {i} - o _ {k} \\| ^ {2} / (2 \\sigma^ {2})\\right)}, \\tag {1}", + "image_path": "dd174de9392e91f0e1999d329cc3052634829af3d22b7c967b00c9b2d610bfcd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(i)" + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "content": " is the neighborhood of " + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "content": " is set as the average nearest neighbor distance. Finally, we obtain the ground truth of generalizable-instructive affordance " + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\Omega = \\{a_i\\}^M" + }, + { + "bbox": [ + 313, + 332, + 555, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 379, + 474, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 474, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 474, + 392 + ], + "type": "text", + "content": "3.2. AffordDexGrasp Framework" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 398, + 466, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 466, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 466, + 410 + ], + "type": "text", + "content": "3.2.1. Intention Pre-Understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 569 + ], + "type": "text", + "content": "The variability of user expression in the open world presents a challenge for models in generalizing and understanding user intentions. To address this, we employ the Multimodal Large Language Model (MLLM) to pre-understand user intentions, by imputing the user language and the rendered RGB image into GPT-4o [15] for intention understanding and key cues extraction. We empirically find that the following cues are both helpful for grasping and easily reasoned by the MLLM: the object category, user intention (e.g., use it), the contacting part of the object (e.g., the handle of a mug), and the grasp direction. The key cues can be extract from language when the commands are clear, otherwise can be reasoned from the image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "To improve the understanding ability of MLLM, we employ chain of thought (CoT) and visual prompt, and the prompts can be found in supplementary material A.4. Additionally, the grasp direction is simplified as a discretized direction in six predefined coordinates (front, back, left, right, up, down). We first ask MLLM to obtain the direction in image coordinates, and transform it to world coordinates by the camera pose. Then the direction vector is derived from the index of the nearest coordinate axis to form discretized direction. Finally, MLLM organizes this information into a concise sentence (e.g., \"use the mug from the left by contacting the handle\"). This compact and information-dense" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11821" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "sentence structure allows subsequent models to better understand users' intentions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 200, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 200, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 200, + 115 + ], + "type": "text", + "content": "3.2.2. Affordance Flow Matching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "text", + "content": "We introduce Affordance Flow Matching (AFM) to learn the affordance distribution efficiently, achieving intention aligned and generalizable affordance generation. AFM is built on the conditional flow matching model [1, 25, 30] aiming to learn the velocity vector field from the noised affordance map " + }, + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\Omega_0" + }, + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "text", + "content": " to the target affordance map " + }, + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\Omega_1" + }, + { + "bbox": [ + 55, + 118, + 296, + 202 + ], + "type": "text", + "content": ", and the objective loss of AFM could be:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 210, + 295, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 210, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 210, + 295, + 236 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {A F M} = \\mathbb {E} _ {t, \\Omega_ {0} \\sim p _ {0}, \\Omega_ {1} \\sim p _ {1}} \\| v _ {\\eta} (\\Omega_ {t}, t | c) - (\\Omega_ {1} - \\Omega_ {0}) \\| ^ {2}, \\tag {2}", + "image_path": "064f7af7348538ff1f2bfa397901304ddc60ce169e38a447feb05b308b37f111.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": " is Gaussian distribution with zero mean and unit variance at time " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": " is the target distribution as time " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "v_{\\eta}(\\Omega_t, t|c)" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": " is parameterized as a neural network with weights " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 236, + 296, + 284 + ], + "type": "text", + "content": " is the input condition features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "text", + "content": "Specifically, the input scene object point cloud " + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "text", + "content": " is encoded by Pointnet++ [34], language is encoded by a pre-trained CLIP model [35], and the direction vector and time embedding are encoded by the Multilayer Perceptrons (MLP). Due to the permutation invariance of point level affordance, the noisy affordance are concatenated with object features to aware the position and semantic information. Then all features are fed into a Perceiver IO Transformer, following [16, 49]. The noisy affordance features are served as input and output array, and the features of language and direction are served as latent array. The outputs of the decoder go through an MLP to obtain the prediction of the flow " + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "inline_equation", + "content": "v_{\\eta}(\\Omega_t,t|c)\\in R^M" + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "text", + "content": ". Finally, in the inference time, we adopt multiple steps " + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "inline_equation", + "content": "1 / \\Delta t" + }, + { + "bbox": [ + 55, + 284, + 296, + 452 + ], + "type": "text", + "content": " following the equation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 460, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 460, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 83, + 460, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\hat {\\Omega} _ {t + \\Delta t} = \\Omega_ {t} + \\Delta t v _ {\\eta} \\left(\\Omega_ {t}, t \\mid c _ {1}\\right), \\quad t \\in [ 0, 1). \\tag {3}", + "image_path": "672aaf9ed4def56cd1edb397578bfe6e6e3bd14b1b49e92f3afbfa731bdca75c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 484, + 178, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 178, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 178, + 496 + ], + "type": "text", + "content": "3.2.3. Grasp Flow Matching" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 499, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 632 + ], + "type": "text", + "content": "We introduce Grasp Flow Matching (GFM) to learn the grasp distribution, conditioned on the output of MLLM and AFM. GFM is also built upon the flow matching model, and the objective is to generate a parameterized dexterous hand pose " + }, + { + "bbox": [ + 55, + 499, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^{dex} = (r,t,q)" + }, + { + "bbox": [ + 55, + 499, + 296, + 632 + ], + "type": "text", + "content": ", which represents the global rotation, translation, and joint angles. In the training time, grasp generation requires calculating explicit losses in 3D space, which involves obtaining the target dexterous poses, which go through forward kinematics to obtain hand configuration [53]. To obtain the target pose, we estimate it in one step during the training stage:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 640, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 640, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 640, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {1} ^ {\\hat {d e x}} = \\mathcal {G} _ {t} ^ {d e x} + (1 - t) * v _ {\\theta} (\\mathcal {G} _ {t} ^ {d e x}, t | c _ {2}), \\quad t \\in [ 0, 1), (4)", + "image_path": "66f006c99532fa49b355040f90d253afe3e994e0707065f23bce6270dd72601b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "where the condition features " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{2}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " are the concatenation of object affordance, language and direction features. And the object affordance features are obtained by feeding affordance with point clouds into PointNet++ [34]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "The losses of grasp generation include 1) the grasp pose regression L2 loss; 2) the hand chamfer loss to minimize the discrepancies between the actual hand shapes; 3) the hand fingertip key point loss to minimize the distance between fingertip position. The loss function can be formulated as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 319, + 138, + 554, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 138, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 319, + 138, + 554, + 163 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {G F M} = \\lambda_ {\\text {p o s e}} \\mathcal {L} _ {\\text {p o s e}} + \\lambda_ {\\text {c h a m f e r}} \\mathcal {L} _ {\\text {c h a m f e r}} + \\lambda_ {\\text {t i p}} \\mathcal {L} _ {\\text {t i p}}, \\tag {5}", + "image_path": "a9050388c0b186bf7f0c7522bbd3d24bf1e58f8cbe1b602a7663c953a469d80a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{pose}}, \\lambda_{\\text{chamfer}}" + }, + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{tip}}" + }, + { + "bbox": [ + 313, + 163, + 514, + 175 + ], + "type": "text", + "content": " are loss weights." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 175, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 281 + ], + "type": "text", + "content": "The object penetration loss, which is used to penalize the penetration of hands into objects, is excluded from grasp generation training, as it leads to significant challenges in the training process [51]. Employing the penetration loss can reduce object penetration, but negatively impacts both intention alignment and generation diversity. Therefore, to achieve semantic-aligned grasp generation, we exclude object penetration loss and introduce affordance-guided grasp optimization to reduce object penetration." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 287, + 511, + 299 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 511, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 511, + 299 + ], + "type": "text", + "content": "3.2.4. Affordance-guided Grasp Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "content": "We introduce a non-parametric test-time adaptation (TTA), affordance-guided grasp optimization, to improve the grasp quality while maintaining intention consistency. Compared to learning-based TTA methods [17, 51], our nonparametric optimization does not suffer from performance degradation caused by out-of-domain open-set samples." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 374, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 445 + ], + "type": "text", + "content": "We design optimization objectives based on that high-quality grasps should exhibit high grasp stability and intention alignment in grasping posture and contact areas. Specifically, our optimization objectives include: 1) affordance contact loss for constraining hand contact candidate points in contact with the affordance area:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 451, + 554, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 451, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 319, + 451, + 554, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a f f - d i s t}} = \\sum_ {i} \\mathbb {I} \\left(\\left(d \\left(p _ {i} ^ {c}\\right) < \\tau_ {1}\\right) \\vee \\left(d \\left(p _ {i} ^ {r}\\right) < \\tau_ {1}\\right)\\right) \\cdot d \\left(p _ {i} ^ {r}\\right), \\tag {6}", + "image_path": "286a31f5ab92eb49501e00fac27cfbffb641d1d72b764fb9d5d4df05397f5bb5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "inline_equation", + "content": "d(p) = \\min_{o \\in \\{o_j \\in \\mathcal{O} | a_j > \\tau_2\\}} \\| p - o_j \\|_2" + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "inline_equation", + "content": "o_j" + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "inline_equation", + "content": "a_j" + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": " represent the position and affordance values of object points, and " + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "inline_equation", + "content": "p_i^c" + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "inline_equation", + "content": "p_i^r" + }, + { + "bbox": [ + 313, + 481, + 554, + 553 + ], + "type": "text", + "content": " represent the hand contact candidates of the initial coarse hand and current refined hand. 2) affordance fingertip loss to keep the fingertip positions that are in contact with the affordance area unchanged." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 554, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 612 + ], + "type": "text", + "content": "3) object penetration loss for punish hand-object penetration, 4) self-penetration loss to punish the penetration of hand it self. 5) joint limitation loss to punish out-of-limit joint angles. The optimization objectives can be formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 619, + 557, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 557, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 557, + 647 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} ^ {d e x}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {a f f - d i s t}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a f f - f i n g e r}} + \\lambda_ {3} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {4} \\mathcal {L} _ {\\text {s p e n}} + \\lambda_ {5} \\mathcal {L} _ {\\text {j o i n t}}\\right). \\tag {7}", + "image_path": "c9708c31dcb7823a30b0fd393a6419eb1c80dd5ed9fcd2641cf99530242963e4.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "text", + "content": "4. Open-set Dexterous Grasp Dataset" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "To support the evaluation of our framework, we build an open-set dexterous grasp dataset in table-top environment based on the language-guided dexterous grasp dataset" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11822" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 533, + 290 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 533, + 290 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 533, + 290 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 533, + 290 + ], + "type": "table", + "html": "
IntentionQualityDiversity
FID ↓CD ↓R - Precision ↑Suc. ↑Q1 ↑Pen. ↓δt ↑δr ↑δq ↑
Top1Top2Top3
Open Set A
ContactGen[28]0.4289.380.1640.2320.28911.6%1.71e-41.092.8618.439.2
Contact2Grasp[19]0.42610.20.2570.3370.39816.5%0.01720.6685.734.8448.3
GraspCVAE[40]0.3785.550.3090.4020.46921.9%0.01500.7094.364.6021.2
SceneDiffuser[12]0.3035.020.3970.4730.52329.1%0.01510.4877.379.4468.9
DexGYS[51]0.2974.630.3170.4010.46344.2%0.05120.3626.306.7954.7
Ours0.2313.810.4800.5880.66645.1%0.05310.2937.486.9861.5
Open Set B
ContactGen[28]0.4928.950.1960.2570.2996.04%1.30e-41.113.2618.745.5
Contact2Grasp[19]0.36911.90.2480.3180.36716.2%0.008010.7985.224.6340.7
GraspCVAE[40]0.3655.350.2970.3570.40324.2%0.004210.7383.965.4725.6
SceneDiffuser[12]0.2716.500.3020.3550.39330.7%0.008510.6835.416.1372.9
DexGYS[51]0.3583.400.2940.3580.40335.2%0.02200.6915.406.1159.0
Ours0.1622.950.5320.6090.65538.9%0.03520.3616.866.8456.3
", + "image_path": "d1f866de7273b6da08c4652c32ff9ebbf481d5d1c9a72bba8e858761b4f42002.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 172, + 298, + 436, + 309 + ], + "lines": [ + { + "bbox": [ + 172, + 298, + 436, + 309 + ], + "spans": [ + { + "bbox": [ + 172, + 298, + 436, + 309 + ], + "type": "text", + "content": "Table 1. Results on open-set datasets compared with the SOTA methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 320, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 294, + 357 + ], + "type": "text", + "content": "[51, 56]. Our open-set dataset consists of 33 categories, 1536 objects, 1909 scenes, and 43,504 dexterous grasps for Shadow Hand [37] and Leap Hand [39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 295, + 620 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "text", + "content": "- Open Set Data Split. To enable comprehensive evaluation, we construct Open sets A and B using two independent dataset splits, each with its own set of unseen categories. For each split, all categories are first labeled as either seen or unseen. Then, " + }, + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "text", + "content": " of the objects from the seen categories are used for training, while the remaining " + }, + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 357, + 295, + 440 + ], + "type": "text", + "content": " and all objects from the unseen categories are used for testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 441, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 536 + ], + "type": "text", + "content": "- Scene Construction. To make our dataset more practical, we generate scene data by placing objects in a tabletop environment using Blenderproc [5]. To prevent collisions between the hand and the table, objects are elevated using a shelf, and collision detection is performed to filter out invalid grasps. The physics engine is activated to ensure physically plausible object placements and grasping poses, enhancing dataset quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 536, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 620 + ], + "type": "text", + "content": "- Scene Point-Cloud and Image Rendering. The scene was captured using five RGB-D cameras: four positioned at elevated lateral angles and one directly above the object. The partial point clouds are merged into a global point cloud. To render realistic RGB images for MLLM processing, we use the texture generation model Paint3D [57] to apply realistic textures to all objects in our dataset." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 628, + 132, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 132, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 132, + 641 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 168, + 659 + ], + "type": "text", + "content": "5.1. Evaluation Metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 689 + ], + "type": "text", + "content": "We employ three types of metric to evaluate the ability of intention consistency, grasp quality and grasp diversity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "- Evaluating Intention Consistency. We employ several metrics, including FID, R-Precision and Chamfer Distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 320, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 524 + ], + "type": "text", + "content": "1) FID (Frechet Inception Distance) [11] measures the distribution similarity between generated grasps and ground truth. To extract grasp and instruction features, we train an object-grasp point cloud encoder and a language encoder by contrastive learning [10]. 2) R-Precision evaluates the semantic alignment between language instructions and generated grasps [10]. Specifically, for each generated grasp, we construct a pool of 32 samples, including its ground-truth instruction and randomly selected samples. We then compute and rank the cosine distances between the point cloud features and the language features of all samples in the pool. The average accuracy is computed at the top-1, top-2, and top-3 positions, that a retrieval is considered successful if the ground truth entry appears among the top-k candidates. 3) Chamfer Distance (CD) quantifies the discrepancy between the predicted hand point cloud and the closest target with same intention [51]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "text", + "content": "- Evaluating Grasp Quality. Following [46], we use success rate, denoted as " + }, + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "inline_equation", + "content": "Suc" + }, + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "text", + "content": ", in Issac Gym [24] and Q1 [7] to assess grasp stability. Maximal penetration depth (cm), denoted as " + }, + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "inline_equation", + "content": "Pen" + }, + { + "bbox": [ + 313, + 525, + 554, + 585 + ], + "type": "text", + "content": ", is used to calculate the maximal penetration depth from the object point cloud to hand meshes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "text", + "content": "- Evaluating Grasp Diversity. We compute the Standard deviation of translation " + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "inline_equation", + "content": "\\delta_t" + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "inline_equation", + "content": "\\delta_r" + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "text", + "content": " and joint angle " + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "inline_equation", + "content": "\\delta_q" + }, + { + "bbox": [ + 313, + 586, + 554, + 622 + ], + "type": "text", + "content": " of samples within the same scene observation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 634, + 447, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 447, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 447, + 647 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "For training our framework, the number of epochs is set to 50 for AFM and 200 for GFM. For AFM, the loss weight of L2 loss is set to 1. For GFM, we set " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{pose}} = 10" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{chamfer}} = 1" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tip}} = 2" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". The batch size is set to 16 for AFM and 64 for GFM. The initial learning rate is " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "2.0 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", decaying" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11823" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 535, + 319 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 535, + 319 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 535, + 319 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 535, + 319 + ], + "type": "table", + "html": "
Open Set AOpen Set B
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
Necessity of Generalizable-Instructive Affordance
w/o affordance0.3385.090.3690.01800.5690.3224.440.3080.01940.452
w object part (pred)0.3135.060.3200.01380.5940.3163.790.3448.12e-30.721
w contact map (pred)0.3204.890.3800.01310.6050.2863.250.4089.25e-30.636
w our affordance (pred)0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w object part (GT)0.3065.050.3480.01980.5230.2713.790.3536.68e-30.526
w contact map (GT)0.1132.010.5980.02850.4940.1382.680.5940.01430.643
w our affordance (GT)0.1693.520.5090.01810.5450.1642.610.5500.01200.584
Effectiveness of Pre-Understanding Stage
w/o key cues extraction0.2583.990.4630.02220.5100.1863.420.5158.33e-30.702
w/o direction0.2544.070.4320.01620.5310.1852.850.5298.80e-30.679
w MLLM0.2423.790.4800.02400.5010.1762.760.5380.01500.612
Effectiveness of Affordance-based Optimization
w/o our optimization0.2423.790.4800.02400.5010.1762.760.5380.01500.612
w penetration loss0.3367.070.3180.05050.2990.2415.760.4450.01470.719
w ContactNet [17]0.34912.30.2790.04340.1240.35511.10.2400.02060.129
w RefineNet [51]0.2493.960.4360.04870.3990.2032.920.4930.01490.657
w our optimization0.2313.810.4770.05310.2930.1622.950.5320.03500.361
", + "image_path": "ea58a6d902f45c0aacd628d48fbd9d8e39a3cdf612a69f5294b499e2a99a2631.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "2.0 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": " using a cosine learning rate scheduler [31]. The Adam optimizer is used with a weight decay rate of " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "5.0 \\times 10^{-6}" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ". In the inference, the time step is set to 10 for AFM and 20 for GFM. For the optimization, " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 100" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 10" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{3} = 100" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{4} = 10" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "inline_equation", + "content": "\\lambda_{5} = 100" + }, + { + "bbox": [ + 55, + 366, + 297, + 451 + ], + "type": "text", + "content": ". The number of optimization iterations is set to 200. All experiments are implemented using PyTorch on a single RTX 4090 GPU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 465, + 234, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 234, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 234, + 478 + ], + "type": "text", + "content": "5.3. Comparison with SOTA Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 486, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 715 + ], + "type": "text", + "content": "We compare our methods with the SOTA methods, as shown in Table 1. The generic grasp methods are reproduced by concatenating the point cloud features and features of same language guidance with ours. The same encoders are employed for fair comparison, and the penetration loss is excluded to avoid learning difficulties according to [51]. The results show that our framework significantly outperforms all previous methods in terms of intention consistency and grasp quality. Our framework also achieves a reasonably high level of diversity, as excessive diversity may lead to unnatural postures. Previous language-guided methods fail to generalize well to unseen categories due to the huge gap between language and grasping actions. Similarly, contact-based methods don't perform well, as the contact maps are difficult to generalize. Additionally, our method outperforms other methods in the close set, as shown in Table 3. Overall, the results indicate that our framework achieves strong performance in generating intention-aligned, high-quality, and diverse grasps." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 328, + 365, + 539, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 326, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 555, + 351 + ], + "type": "text", + "content": "Table 2. Ablation study for our framework. The results of first two experiment groups are obtained from model outputs without optimization for an intuitive and reasonable comparison. The results in each group should be compared with light yellow line (our default setting)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 365, + 539, + 427 + ], + "lines": [ + { + "bbox": [ + 328, + 365, + 539, + 427 + ], + "spans": [ + { + "bbox": [ + 328, + 365, + 539, + 427 + ], + "type": "table", + "html": "
FID ↓CD ↓Top1 ↑Q1 ↑Pen. ↓
GraspCVAE0.2082.400.3950.01000.771
DexGYS0.08041.740.5900.05510.397
Ours0.02861.130.7790.05620.193
", + "image_path": "76f77b5afa0e6db87d89fe0954969da1b8a9e9c55a1dbef672adca535dacf55c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 335, + 435, + 533, + 446 + ], + "lines": [ + { + "bbox": [ + 335, + 435, + 533, + 446 + ], + "spans": [ + { + "bbox": [ + 335, + 435, + 533, + 446 + ], + "type": "text", + "content": "Table 3. Experiment of close set with SOTA methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 471, + 467 + ], + "type": "text", + "content": "5.4. Necessity of Our Affordance" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 640 + ], + "type": "text", + "content": "The results shown in Table 2 validate the key insight of our framework: using the generalizable-instructive affordance to bridge the gap between high-level language and low-level grasp actions. The first row presents the results without affordance, which fails to generalize to unseen categories. The subsequent two rows show the results under the condition of object parts and contact maps. Prediction refers to maps generated by models, while GT refers to using the corresponding ground truth. As discussed in Section 3.1, the results demonstrate that object parts are too coarse to guide the grasp effectively, even when using the ground truth. On the other hand, the contact map is too fine-grained to generalize to unseen categories. Only our affordance achieves a balance between generalization and instruction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Furthermore, our affordance representation and framework demonstrate good generalization in the one-shot setting, as shown in Table 4. For the one-shot experiments, we introduce several novel categories to the test set that differ significantly from the training set and add one object from each unseen category to the training set. The results show" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11824" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 115, + 132 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 115, + 132 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 115, + 132 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 115, + 132 + ], + "type": "image", + "image_path": "ff9a2f66754552d456ac72bbbff60e6905a37318b256e6f3dd9423ebe98e928d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 133, + 114, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 133, + 114, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 133, + 114, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 133, + 114, + 190 + ], + "type": "image", + "image_path": "3f519ccdeb3f8cce1538d3a8d844d1749f8b2eca6b89b3581c5217958209bdd1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 555, + 217 + ], + "type": "text", + "content": "Figure 4. The visualization of generated affordance and dexterous grasp. The left top shows the zero-shot samples and the left bottom shows the one-shot samples in real world. The right top and right bottom show the zero-shot samples in simulation open set A and B." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 118, + 74, + 176, + 132 + ], + "blocks": [ + { + "bbox": [ + 118, + 74, + 176, + 132 + ], + "lines": [ + { + "bbox": [ + 118, + 74, + 176, + 132 + ], + "spans": [ + { + "bbox": [ + 118, + 74, + 176, + 132 + ], + "type": "image", + "image_path": "20e2a9ba5a3591a20b0edc4145b86ee5a4ee497d4aab6d4442b326c9b0445e98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 118, + 133, + 176, + 190 + ], + "blocks": [ + { + "bbox": [ + 118, + 133, + 176, + 190 + ], + "lines": [ + { + "bbox": [ + 118, + 133, + 176, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 133, + 176, + 190 + ], + "type": "image", + "image_path": "28b195ad32641f84d0f9bde82224acb8275311f2a8d1cb4e448e26120b7e6e81.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 73, + 239, + 132 + ], + "blocks": [ + { + "bbox": [ + 181, + 73, + 239, + 132 + ], + "lines": [ + { + "bbox": [ + 181, + 73, + 239, + 132 + ], + "spans": [ + { + "bbox": [ + 181, + 73, + 239, + 132 + ], + "type": "image", + "image_path": "ac1873f58b02f2801fe3a262136a431df91ebba8cec69c0176c533ab0d678b72.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 133, + 239, + 190 + ], + "blocks": [ + { + "bbox": [ + 181, + 133, + 239, + 190 + ], + "lines": [ + { + "bbox": [ + 181, + 133, + 239, + 190 + ], + "spans": [ + { + "bbox": [ + 181, + 133, + 239, + 190 + ], + "type": "image", + "image_path": "1dd62dd124ddf7e64f1955ff8173d8767a04ed439a1245326a9536fc3af15034.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 244, + 73, + 301, + 132 + ], + "blocks": [ + { + "bbox": [ + 244, + 73, + 301, + 132 + ], + "lines": [ + { + "bbox": [ + 244, + 73, + 301, + 132 + ], + "spans": [ + { + "bbox": [ + 244, + 73, + 301, + 132 + ], + "type": "image", + "image_path": "a00adb8a0018a38db997a5970933dc866aff9dbab9265be93f27287912aef22e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 244, + 133, + 301, + 190 + ], + "blocks": [ + { + "bbox": [ + 244, + 133, + 301, + 190 + ], + "lines": [ + { + "bbox": [ + 244, + 133, + 301, + 190 + ], + "spans": [ + { + "bbox": [ + 244, + 133, + 301, + 190 + ], + "type": "image", + "image_path": "84b7fd3620518bd178c053dcaf6f15a212d29d68958b0693bf3fd9a637a4ddf2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 73, + 363, + 132 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 363, + 132 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 363, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 363, + 132 + ], + "type": "image", + "image_path": "96c2c28e389b2de8bbf57d1efa143c6c32bf22e01ce16a2b11334d35f57a15b6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 133, + 363, + 190 + ], + "blocks": [ + { + "bbox": [ + 306, + 133, + 363, + 190 + ], + "lines": [ + { + "bbox": [ + 306, + 133, + 363, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 133, + 363, + 190 + ], + "type": "image", + "image_path": "b300cb4db437cc4cc8e5ecb879262e4df3f28b483a1e046e5f7212bfa2eb08c4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 367, + 73, + 425, + 132 + ], + "blocks": [ + { + "bbox": [ + 367, + 73, + 425, + 132 + ], + "lines": [ + { + "bbox": [ + 367, + 73, + 425, + 132 + ], + "spans": [ + { + "bbox": [ + 367, + 73, + 425, + 132 + ], + "type": "image", + "image_path": "60abc94610e4bbe048c189954d6febeffd4a21c5837fe88ee3fe5e1ec42db156.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 368, + 133, + 425, + 190 + ], + "blocks": [ + { + "bbox": [ + 368, + 133, + 425, + 190 + ], + "lines": [ + { + "bbox": [ + 368, + 133, + 425, + 190 + ], + "spans": [ + { + "bbox": [ + 368, + 133, + 425, + 190 + ], + "type": "image", + "image_path": "4ca27846da0469bc14a458e7ca54a29fbea4d7bd49fcdadf343f702007d02c9e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 431, + 73, + 488, + 132 + ], + "blocks": [ + { + "bbox": [ + 431, + 73, + 488, + 132 + ], + "lines": [ + { + "bbox": [ + 431, + 73, + 488, + 132 + ], + "spans": [ + { + "bbox": [ + 431, + 73, + 488, + 132 + ], + "type": "image", + "image_path": "c6e5b63ce2eb82118e3d34a3431b21ac4d86652c837b1f2fdc673a8abdc2c2f2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 431, + 133, + 488, + 190 + ], + "blocks": [ + { + "bbox": [ + 431, + 133, + 488, + 190 + ], + "lines": [ + { + "bbox": [ + 431, + 133, + 488, + 190 + ], + "spans": [ + { + "bbox": [ + 431, + 133, + 488, + 190 + ], + "type": "image", + "image_path": "a4c442c48f4bd7c7fa5e61c17c4aea7f8763c9515e6d4b7d05ba6354a2813538.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 495, + 73, + 551, + 132 + ], + "blocks": [ + { + "bbox": [ + 495, + 73, + 551, + 132 + ], + "lines": [ + { + "bbox": [ + 495, + 73, + 551, + 132 + ], + "spans": [ + { + "bbox": [ + 495, + 73, + 551, + 132 + ], + "type": "image", + "image_path": "4730945ebc7f3f20ec9572a97eff42a85aa112af89e346b347b33c939f6482c3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 495, + 133, + 551, + 190 + ], + "blocks": [ + { + "bbox": [ + 495, + 133, + 551, + 190 + ], + "lines": [ + { + "bbox": [ + 495, + 133, + 551, + 190 + ], + "spans": [ + { + "bbox": [ + 495, + 133, + 551, + 190 + ], + "type": "image", + "image_path": "edff7035c278721fa4962391ed5e5982dc8427c3bfdab4715e651eb231a5d801.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 58, + 236, + 293, + 350 + ], + "blocks": [ + { + "bbox": [ + 58, + 236, + 293, + 350 + ], + "lines": [ + { + "bbox": [ + 58, + 236, + 293, + 350 + ], + "spans": [ + { + "bbox": [ + 58, + 236, + 293, + 350 + ], + "type": "table", + "html": "
Seen CateSimilar CateNovel Cate
Top1 ↑ CD ↓Top1 ↑ CD ↓Top1 ↑ CD ↓
Zero-Shot
w/o afford0.8272.610.2955.710.2197.82
Ours0.8801.060.4323.650.2466.42
One-Shot
w/o afford0.7673.170.4293.670.3426.64
Ours0.8700.990.6561.730.5863.03
", + "image_path": "10bc1fbcc9f91dcd0050c4d6c34fab515ccdc4513f634f64cf06ad30ecd587c4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "type": "text", + "content": "that our framework achieves a significant performance improvement, not only on samples from different categories with similar structures but also for novel categories." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "type": "text", + "content": "5.5. Effectiveness of Each Component" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 448, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 604 + ], + "type": "text", + "content": "Further ablation studies are conducted to evaluate each component of our framework. For the pre-understanding stage assisted by MLLM, w/o key cues extraction refers to directly inputting the user's commands into the model, while w/o direction means that the direction information is not utilized. Both designs enhance performance, and we believe that MLLM would be more beneficial in real-world applications due to its powerful generalization. For affordance-based optimization, our optimization improves grasp quality and maintains intention consistency, outperforming the learning-based methods ContactNet [17] and RefineNet [51], as well as the model which is trained with penetration loss." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 612, + 194, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 194, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 194, + 625 + ], + "type": "text", + "content": "5.6. Real World Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "The real-word experiments are conducted to verify the simulation-to-reality ability of our framework. We employ a Leap Hand, a Kinova Gen3 6DOF arms and an original wrist RGB-D camera of Kinova arm. And we collect several common objects in daily life, including unseen objects and unseen categories. In experiment, we synthesize the scene point cloud by taking several partial depth maps" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 315, + 236, + 557, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 358, + 294, + 369 + ], + "lines": [ + { + "bbox": [ + 56, + 358, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 294, + 369 + ], + "type": "text", + "content": "Table 4. Zero-shot and one-shot generalization of our framework." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 236, + 557, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 236, + 557, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 236, + 557, + 285 + ], + "type": "table", + "html": "
DexGYS273647010
Ours105910610610
", + "image_path": "a89dce9a47945e39a3997f05be7b926cfa7b92cdcb44b6a253afe9622277348d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 294, + 542, + 306 + ], + "lines": [ + { + "bbox": [ + 325, + 294, + 542, + 306 + ], + "spans": [ + { + "bbox": [ + 325, + 294, + 542, + 306 + ], + "type": "text", + "content": "Table 5. Real world experiments: Successes in 10 attempts." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "type": "text", + "content": "around the object. Then the scene point cloud, a RGB image and the user language command are fed into our framework to obtain the dexterous grasp pose. During execution, we first move the arm to a pre-grasp position, then synchronously move the joints of the robotic arm and the dexterous hand to reach the target pose. In evaluation, a grasp is considered as success if the hand can lift it up and the action is consistent with the given command. The results shown in Table 5 shows that our framework achieve higher success rate than previous method." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 453, + 391, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 391, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 391, + 465 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "We believe that achieving generalizable dexterous grasps in open set is crucial within the deep learning and robotics communities. In this paper, we explore a novel task of open-set language-guided dexterous grasp. This task is challenging due to the huge gap between language and grasping actions, which hinders the model's generalization ability. We propose the AffordDexGrasp framework with the insight that using a new affordance representation, generalizable-instructive affordance, to solve this challenge. This affordance enables generalization to unseen categories by utilizing the object's local structure and category-agnostic semantic attributes, thus facilitating effective dexterous grasp generation. Based on this affordance, our framework introduces two flow matching based models for affordance generation and affordance-guided grasp generation. Moreover, we introduce a pre-understand stage and a pose-optimization stage to further boost generalization. We conduct extensive open-set experiments in both simulation and the real world, and the results show that our framework outperforms all SOTA methods." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11825" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "7. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported partially by NSFC (92470202, U21A20471), Guangdong NSF Project (No. 2023B151504 0025), Guangdong Key Research and Development Program (No. 2024B0101040004)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 204 + ], + "type": "text", + "content": "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 261 + ], + "type": "text", + "content": "[2] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 296, + 328 + ], + "type": "text", + "content": "[3] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 361-378. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 362 + ], + "type": "text", + "content": "[4] Xiaoyun Chang and Yi Sun. Text2grasp: Grasp synthesis by text prompts of object grasping parts. arXiv preprint arXiv:2404.15189, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 430 + ], + "type": "text", + "content": "[5] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Wendelin Knauer, Klaus H Strobl, Matthias Hunt, and Rudolph Triebel. Blenderproc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "text", + "content": "[6] Hao-Shu Fang, Chenxi Wang, Hongjie Fang, Minghao Gou, Jirong Liu, Hengxu Yan, Wenhai Liu, Yichen Xie, and Cewu Lu. Anygrasp: Robust and efficient grasp perception in spatial and temporal domains. IEEE Transactions on Robotics, 39(5):3929-3945, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 531 + ], + "type": "text", + "content": "[7] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In Proceedings., 1992 IEEE International Conference on Robotics and Automation, 1992., pages 2290-2295. IEEE, 1992. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 565 + ], + "type": "text", + "content": "[8] James J Gibson. The theory of affordances:(1979). In *The people, place, and space reader*, pages 56-60. Routledge, 2014. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 622 + ], + "type": "text", + "content": "[9] Patrick Grady, Chengcheng Tang, Christopher D Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C Kemp. Contactopt: Optimizing contact to improve grasps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1471-1481, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 678 + ], + "type": "text", + "content": "[10] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[11] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "rium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16750-16761, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[13] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. arXiv preprint arXiv:2307.05973, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "text", + "content": "[14] Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, and Li Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation. arXiv preprint arXiv:2409.01652, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 294 + ], + "type": "text", + "content": "[15] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[16] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[17] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11107-11116, 2021. 2, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 458 + ], + "type": "text", + "content": "[18] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[19] Haoming Li, Xinzhuo Lin, Yang Zhou, Xiang Li, Yuchi Huo, Jiming Chen, and Qi Ye. Contact2grasp: 3d grasp synthesis via hand-object contact constraint. arXiv preprint arXiv:2210.09245, 2022. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 557 + ], + "type": "text", + "content": "[20] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 603 + ], + "type": "text", + "content": "[21] Kailin Li, Jingbo Wang, Lixin Yang, Cewu Lu, and Bo Dai. Semgrasp: Semantic grasp generation via language aligned discretization. In European Conference on Computer Vision, pages 109-127. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 604, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 669 + ], + "type": "text", + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[23] Samuel Li, Sarthak Bhagat, Joseph Campbell, Yaqi Xie, Woojun Kim, Katia Sycara, and Simon Stepputtis. Shape-grasp: Zero-shot task-oriented grasping with large language models through geometric decomposition. In 2024" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11826" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10527-10534. IEEE, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 148 + ], + "type": "text", + "content": "[24] Jacky Liang, Viktor Makoviychuk, Ankur Handa, Nuttapong Chentanez, Miles Macklin, and Dieter Fox. Gpu-accelerated robotic simulation for distributed reinforcement learning. In Conference on Robot Learning, pages 270-282. PMLR, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 150, + 294, + 183 + ], + "type": "text", + "content": "[25] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 216 + ], + "type": "text", + "content": "[26] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 281 + ], + "type": "text", + "content": "[27] Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part segmentation for 3d point clouds via pretrained image-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21736-21746, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 337 + ], + "type": "text", + "content": "[28] Shaowei Liu, Yang Zhou, Jimei Yang, Saurabh Gupta, and Shenlong Wang. Contactgen: Generative contact modeling for grasp generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20609-20620, 2023. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[29] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 426 + ], + "type": "text", + "content": "[30] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 502 + ], + "type": "text", + "content": "[32] Teli Ma, Zifan Wang, Jiaming Zhou, Mengmeng Wang, and Junwei Liang. Glover: Generalizable open-vocabulary affordance reasoning for task-oriented grasping. arXiv preprint arXiv:2411.12286, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 547 + ], + "type": "text", + "content": "[33] Teli Ma, Jiaming Zhou, Zifan Wang, Ronghe Qiu, and Junwei Liang. Contrastive imitation learning for language-guided multi-task robotic manipulation. arXiv preprint arXiv:2406.09738, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "text", + "content": "[34] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[36] Adam Rashid, Satvik Sharma, Chung Min Kim, Justin Kerr, Lawrence Yunliang Chen, Angjoo Kanazawa, and Ken Goldberg. Language embedded radiance fields for zero-shot task-oriented grasping. In 7th Annual Conference on Robot Learning, 2023. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[37] ShadowHand. Shadowrobot. https://www.).shadowrobot.com/dexterous-hand-series/, 2005.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[38] Lin Shao, Fabio Ferreira, Mikael Jorda, Varun Nambiar, Jianlan Luo, Eugen Solowjow, Juan Aparicio Ojea, Oussama Khatib, and Jeannette Bohg. Unigrasp: Learning a unified model to grasp with multifingered robotic hands. IEEE Robotics and Automation Letters, 5(2):2286-2293, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[39] Kenneth Shaw, Ananye Agarwal, and Deepak Pathak. Leap hand: Low-cost, efficient, and anthropomorphic hand for robot learning. arXiv preprint arXiv:2309.06440, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[40] Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. Advances in neural information processing systems, 28, 2015. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[41] Peize Sun, Shoufa Chen, Chenchen Zhu, Fanyi Xiao, Ping Luo, Saining Xie, and Zhicheng Yan. Going denser with open-vocabulary part segmentation. arXiv preprint arXiv:2305.11173, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "text", + "content": "[42] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 398 + ], + "type": "text", + "content": "[43] Ardian Umam, Cheng-Kun Yang, Min-Hung Chen, Jen-Hui Chuang, and Yen-Yu Lin. Partdistill: 3d shape part segmentation by vision-language model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 465 + ], + "type": "text", + "content": "[44] Weikang Wan, Haoran Geng, Yun Liu, Zikang Shan, Yaodong Yang, Li Yi, and He Wang. Unidxgrasp++: Improving dexterous grasping policy learning via geometry-aware curriculum and iterative generalist-specialist learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3891-3902, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[45] An-Lan Wang, Nuo Chen, Kun-Yu Lin, Li Yuan-Ming, and Wei-Shi Zheng. Task-oriented 6-dof grasp pose detection in clutters. arXiv preprint arXiv:2502.16976, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 566 + ], + "type": "text", + "content": "[46] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11359-11366. IEEE, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 621 + ], + "type": "text", + "content": "[47] Wenbo Wang, Fangyun Wei, Lei Zhou, Xi Chen, Lin Luo, Xiaohan Yi, Yizhong Zhang, Yaobo Liang, Chang Xu, Yan Lu, et al. Unigrasptransformer: Simplified policy distillation for scalable dexterous robotic grasping. arXiv preprint arXiv:2412.02699, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[48] Yan-Kang Wang, Chengyi Xing, Yi-Lin Wei, Xiao-Ming Wu, and Wei-Shi Zheng. Single-view scene point cloud human grasp generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 831-841, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11827" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[50] Wei Wei, Peng Wang, Sizhe Wang, Yongkang Luo, Wanyi Li, Daheng Li, Yayu Huang, and Haonan Duan. Learning human-like functional grasping for multi-finger hands from few demonstrations. IEEE Transactions on Robotics, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 229 + ], + "type": "text", + "content": "[51] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. Advances in Neural Information Processing Systems, 37:46881-46907, 2025. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 275 + ], + "type": "text", + "content": "[52] Xiao-Ming Wu, Jia-Feng Cai, Jian-Jian Jiang, Dian Zheng, Yi-Lin Wei, and Wei-Shi Zheng. An economic framework for 6-dof grasp detection. In European Conference on Computer Vision, pages 357-375. Springer, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 319 + ], + "type": "text", + "content": "[53] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 398 + ], + "type": "text", + "content": "[54] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4737-4746, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 454 + ], + "type": "text", + "content": "[55] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 511 + ], + "type": "text", + "content": "[56] Lixin Yang, Kailin Li, Xinyu Zhan, Fei Wu, Anran Xu, Liu Liu, and Cewu Lu. Oakink: A large-scale knowledge repository for understanding hand-object interaction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20953-20962, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[57] Xianfang Zeng, Xin Chen, Zhongqi Qi, Wen Liu, Zibo Zhao, Zhibin Wang, Bin Fu, Yong Liu, and Gang Yu. Paint3d: Paint anything 3d with lighting-less texture diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4252-4262, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 623 + ], + "type": "text", + "content": "[58] Jialiang Zhang, Haoran Liu, Danshi Li, XinQiang Yu, Hao ran Geng, Yufei Ding, Jiayi Chen, and He Wang. Dexgrasp-net 2.0: Learning generative dexterous grasping in largescale synthetic cluttered scenes. In 8th Annual Conference on Robot Learning, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[59] Zhongqun Zhang, Hengfei Wang, Ziwei Yu, Yihua Cheng, Angela Yao, and Hyung Jin Chang. Nl2contact: Natural language guided 3d hand-object contact modeling with diffusion model. In European Conference on Computer Vision, pages 284-300. Springer, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[60] Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yun-hao Fang, and Hao Su. Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmenta" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 206 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "tion and maximum likelihood estimation. arXiv preprint arXiv:2312.03015, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[61] Tianqiang Zhu, Rina Wu, Xiangbo Lin, and Yi Sun. Toward human-like grasp: Dexterous grasping via semantic representation of object-hand. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15741-15751, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[62] Binghui Zuo, Zimeng Zhao, Wenqian Sun, Xiaohan Yuan, Zhipeng Yu, and Yangang Wang. Graspondiff: Grasping generation for hand-object interaction with multimodal guided diffusion. IEEE Transactions on Visualization and Computer Graphics, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11828" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_content_list.json b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d9a19f46bf723b4ddf8dbab8ab145c935ec15390 --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_content_list.json @@ -0,0 +1,1821 @@ +[ + { + "type": "text", + "text": "After the Party: Navigating the Mapping From Color to Ambient Lighting", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Florin-Alexandru Vasluianu Tim Seizinger Zongwei Wu* Radu Timofte Computer Vision Lab, CAIDAS & IFI, University of Wurzburg", + "bbox": [ + 189, + 180, + 807, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{firstname.lastname}@uni-wuerzburg.de", + "bbox": [ + 331, + 219, + 660, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Illumination in practical scenarios is inherently complex, involving colored light sources, occlusions, and diverse material interactions that produce intricate reflectance and shading effects. However, existing methods often oversimplify this challenge by assuming a single light source or uniform, white-balanced lighting, leaving many of these complexities unaddressed. In this paper, we introduce CL3AN, the first large-scale, high-resolution dataset of its kind designed to facilitate the restoration of images captured under multiple Colored Light sources to their Ambient-Normalized counterparts. Through benchmarking, we find that leading approaches often produce artifacts—such as illumination inconsistencies, texture leakage, and color distortion—primarily due to their limited ability to precisely disentangle illumination from reflectance. Motivated by this insight, we achieve such a desired decomposition through a novel learning framework that leverages explicit chromaticity-luminance components guidance, drawing inspiration from the principles of the Retinex model. Extensive evaluations on existing benchmarks and our dataset demonstrate the effectiveness of our approach, showcasing enhanced robustness under non-homogeneous color lighting and material-specific reflectance variations, all while maintaining a highly competitive computational cost. The benchmark, codes, and models are available at www.github.com/fvasluianu97/RLN2.", + "bbox": [ + 89, + 300, + 485, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 720, + 220, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The goal of Ambient Lighting Normalization is recovering an ambient-lit image, under uniform lighting distribution, regardless of the lighting scenario characteristic to the depicted contents [65]. The ability to recover ambient-normalized images—free from spatially varying illumination artifacts—is critical for robust visual understanding [33, 51, 53]. In both practical imaging scenarios [27, 32] and the rapidly evolving domain of image generation and", + "bbox": [ + 89, + 746, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/75f88847bb339ab0df3037fb024699d37acc278be393a5245022b8346c1a6fe7.jpg", + "image_caption": [ + "Figure 1. Effective deployment of Ambient Lighting Normalization (ALN) as a pre-processing step in Neural Image Editing [19]. We compare the effectiveness of our proposed RLN² against state-of-the-art models in fields like Shadow Removal (DES3 [36]), or ALN (IFBlend [65]), trained on settings including single (1) [68] or multiple (3) white-domain-tuned [65] or color light sources." + ], + "image_footnote": [], + "bbox": [ + 516, + 271, + 903, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "synthesis [57, 81], illumination always involves complex interactions of multiple colored light sources, occlusions, and material properties, resulting in intricate shadows, glare, color shifts, and texture distortions. Lighting and material characteristics are the driving factors for explaining the manifestation of such artifacts, thus explaining the need for data-level representativity prior to proposing image restoration solutions aiming for improved degradation robustness or restoration performance.", + "bbox": [ + 511, + 561, + 906, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While a wide range of applications require illumination-invariant representations [7, 31, 59, 67], existing normalization methods struggle to disentangle these intertwined factors [19, 35, 62, 81], as illustrated in Fig. 1. Even if data contributions further away from the natural light in $\\mathrm{STD + }$ [44, 68], or multiple white-light-tuned light sources from AMBIENT6K [65] can improve general effectiveness, model design can further drive robustness.", + "bbox": [ + 511, + 700, + 908, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This limitation stems primarily from oversimplified assumptions, such as single white light sources [64, 68] or homogeneous color fields [56, 68, 87], which fail to capture the richness of complex illumination. A key technical challenge lies in the lack of comprehensive datasets that", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding", + "bbox": [ + 109, + 886, + 197, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9218", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adequately represent the complexity of real-world lighting, especially outside the realm of natural white-color-tuned lighting. Prior efforts in shadow removal [64, 68], color constancy [6, 15, 16], or color correction [1, 22, 34, 52, 73] address isolated aspects of illumination but fall short of modeling multi-light interactions. Recent attempts [38, 65] introduce variable-intensity lighting datasets but lack chromatic diversity and material-specific reflectance variations. Furthermore, wide access to ambient-lit images outside the scope of scene-rendering software remains a challenge. As a result, learning-based methods often produce artifacts.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address these limitations through two core contributions. First, we introduce CL3AN, the first large-scale dataset capturing high-resolution scenes under multiple colored directional lights paired with ambient-lit references. Each scene triplet includes variations beyond white-balanced lighting, colored lighting, and a diffuse-lit ground truth, enabling systematic study of illumination effects like self-shadows, highlights, glare areas, color spill or saturation, and other material-dependent artifacts. By benchmarking leading image restoration methods [9, 23, 65, 80], we observe that these models frequently suffer from critical shortcomings, such as the conflation of chromatic variations caused by illumination, texture leakage, inconsistent partial normalization, etc.", + "bbox": [ + 89, + 260, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify the fundamental limitation of existing methods as their inability to systematically disentangle reflectance and illumination—a critical flaw exacerbated by the lack of explicit guidance for separating chromaticity and luminance under multi-colored lighting. Most current approaches [23, 80, 84] rely on single-stream, end-to-end feature modeling, which implicitly maps input images to ambient-normalized counterparts without incorporating physical awareness of light-material interactions. While some recent works attempt to address this by introducing pixel- and frequency-domain supervision [12, 65], more flexible towards reflectance- and chromaticity-information extraction, they remain limited in learning appropriate frequency filters — a challenging task without explicit guidance — resulting in suboptimal performance under the challenging multi-colored lighting conditions.", + "bbox": [ + 89, + 474, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, from a novel perspective, we propose $\\mathbf{RLN}^2$ , a Retinex-inspired framework that introduces parallel, physics-grounded streams to independently model reflectance (material properties) and illumination (lighting geometry/color). We observe that these intrinsic specificities are naturally encoded in the cylindrical representation of the input image, where the Hue (H) and Saturation (S) maps align with the observed chromaticity under the effect of the incoming light, while the Value (V) map provides strong priors for materials-specific reflectance. By embedding these domain-specific clues — such as reflectance information entropy (V), and smoothness constraints (H, S)", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into feature refinement, our framework achieves precise decomposition, effectively eliminating undesirable artifacts while preserving fine-grained details.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 513, + 136, + 836, + 148 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The introduction of the CL3AN, a first-of-its-kind dataset enabling the study of ambient lighting normalization under multiple RGB lights;", + "- The extensive study of a wide range of models, seminal works in various image restoration tasks, when trained and tested on the CL3AN dataset;", + "- We propose $\\mathbf{RLN}^2$ , a robust algorithm built on the principle of simultaneous image reflection and lightness components compensation to recover an ambient lit image, characterized by uniform lighting distribution. $\\mathbf{RLN}^2$ is proposed as a strong baseline for the ambient lighting normalization task, achieving state-of-the-art results on both the public AMBIENT6K dataset and the proposed CL3AN benchmark." + ], + "bbox": [ + 511, + 151, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 376, + 653, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work expands Ambient Lighting Normalization to multicolor direct lighting, addressing previously unexplored effects appearing in cluttered scenes, and analyzing shadow interactions and highlights appearing under multicolor direct lighting.", + "bbox": [ + 511, + 402, + 903, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lighting and Color: Color analysis is central to image processing and understanding, with various downstream tasks such as facial recognition [61, 76, 77], or tracking and matching applications [5, 14, 41, 59, 63] relying on the color similarity principle, thus benefiting from increased appearance robustness, especially under challenging lighting.", + "bbox": [ + 511, + 478, + 903, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Quality Assessment (IQA) methods [3, 79] show a strong correlation between details-abundance and image quality, motivating the focus on professional scene lighting rigs in all expert image photography applications. Given the scarcity of such equipment, alternative strategies [1, 2, 6, 45] are developed around image sensor sensor properties and cross-device constancy [8, 20].", + "bbox": [ + 511, + 568, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Naturally, registered image stacks either for tasks such as multiple view geometry estimation [17, 90] or different image stitching applications require consistent color and content appearance [4, 11, 42, 60]. In such applications, the effort is focused on consistent lighting in the image acquisition, rather than post-processing correction. In-the-wild applications focusing on widely available amateur images [17, 24, 60] document extensive data curation strategies prior to model estimations.", + "bbox": [ + 511, + 674, + 905, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Later applications, given the newly-found popularity of neural 3D modeling, approach scene relighting as part of their rendering pipeline [58, 62]. Alternatively, approaches based on complex generative image models [35, 55] successfully tackle lighting manipulation, but at a high computational cost. Unfortunately, the limited availability of", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9219", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c5d6867ece602091ba1d438df7b6ea0a5a3ef712b75b698daa42d55febdafec.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 272, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3febe8bc9579f52c5f669791b0ef5d09df51674914218a7419c43e262dee3310.jpg", + "image_caption": [ + "B", + "ISTD/WSRD", + "AMBIENT6K", + "CL3AN (ours)" + ], + "image_footnote": [], + "bbox": [ + 279, + 92, + 506, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1b8ff6d2a50f0fc4d47c05960bfefb803e8858eb4278b104df9aa3795c8f9d2c.jpg", + "image_caption": [ + "C", + "Figure 2. A comparison between the lighting scenarios represented in datasets targeting different applications of Ambient Lighting Normalization (ISTD [68], WSRD[64], or AMBIENT6K [65]). In (A), ambient lighting is used to acquire a uniformly-lit reference image. In (B), direct lighting is used for shadow casting, while constraining consistency to the reference image in shadow-free segments. In (C), more directional lights are added to the direct lighting system. In our proposed setup (D), the direct lighting setup is based on RGB lights, dropping the color consistency constraint, and showcasing complex material-light interactions under various conditions." + ], + "image_footnote": [], + "bbox": [ + 513, + 93, + 699, + 215 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7af12f05f4f136016eb35f410ddb40aa2bcb29667905c2e376d5b04ee671f0f1.jpg", + "image_caption": [ + "D" + ], + "image_footnote": [], + "bbox": [ + 702, + 93, + 903, + 215 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "real-world ambient lit images, primarily due to the complexity of an ambient lighting setup, explains the difficulty in addressing real-world lighting normalization, even with the obvious availability of large-complexity models.", + "bbox": [ + 88, + 371, + 480, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reliance on natural light for ambient-lit images proved a viable option, but hindered the variety of possible observed degradations, especially in terms of chromaticity shifts. Unfortunately, the large number of factors involved in light occlusion interaction greatly constrains a representative simulation in general settings. While ISTD/ISTD+ [44, 68], SRD [56] or USR [26] show exclusively flat surfaces under the influence of natural light, the influence of object self-shadows was just recently represented in WSRD [64], only possible under a laboratory setup. Recently, multiple source lighting gained in data-level representation. For example, in LSMI [38], the authors target White-Balance correction robustness under natural and artificial lighting. Alternatively, effects characteristic to multiple illuminants were captured in works such as [30, 54], or in the context of flash photography [29]. However, a uniformly-lit ambient normalized image can not be acquired in these settings.", + "bbox": [ + 89, + 431, + 482, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Alternatively, the introduction of AMBIENT6K [65] coupled multiple light sources of varying intensity and orientation for complex interactions, while also providing an ambient-lit image based on an alternative lighting setup. Unfortunately, the strong connection between the operating parameters set for the direct and ambient lighting systems limits the study to exposure correction for white-aligned scene lighting (see Fig. 2 (A), (B), and (C)).", + "bbox": [ + 89, + 688, + 482, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Restoration (IR): Joint color and exposure correction, while restoring details lost to occlusion or saturation, defines Ambient Lighting Normalization as an Image Restoration task. Various Convolution Neural Networks (CNNs) [10, 37, 83, 85] became strong baselines for the family of IR tasks. The later introduction of the attention", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mechanism [66] in image-based applications [13] enabled consistent performance advantages in the wide IR field. Focusing on different feature fusion techniques through specifically tailored attention mechanisms, based on the IR task particularities, defines the success of multiple models [10, 47, 48, 71, 80, 84] easily becoming benchmark solutions for various IR problems. Separate refinement for image reflectance and lightness under the Retinex model [43] proved particularly effective in low-light image enhancement (LLIE) [9, 72, 78, 88], or image shadow removal [28].", + "bbox": [ + 511, + 371, + 903, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global range dependencies benefits drive the development of State Space Models (SSM) [21], with promising results in general IR [23]. Alternatively, global knowledge can be leveraged through the integration of alternative frequency image representations in IR solutions. Solutions such as FFTFormer [39], DCFormer [46], SFNet [12], DW-GAN [18], and IFBlend [65] successfully extract priors based on Fourier Transform, Discrete Cosine Transform, or other Wavelet Transform variants.", + "bbox": [ + 511, + 523, + 905, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, diffusion-based iterative recurrent solutions [25] started to mature as benchmark solutions in Image Restoration [50, 69, 75, 89], especially in terms of perceived image quality. However, their characteristic increased computational effort hinders their deployment.", + "bbox": [ + 511, + 660, + 905, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 752, + 604, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core of our work is extending the study of Ambient Lighting Normalization to direct color lighting.", + "bbox": [ + 511, + 777, + 905, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 compares the properties of a scene depicted under multi-source color direct lighting, and an equivalent image under white-aligned direct lighting. The goal to recover the ambient-lit reference image. Provided statistics, such as the RGB image histogram, and the Hue/Saturation footprint showcase the complexity of the ALN transformation, espe", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/524aa8bdc7ef33427e115e065022f7192616a5b5f75a203606c166af218e5053.jpg", + "image_caption": [ + "Figure 3. A graphical representation of the Ambient Lighting Normalization transformation in the context of the white-aligned lighting, compared to the challenges posed by the introduction of a multicolor direct lighting system. The RGB image histogram, coupled with the Hue-Saturation rough projection is provided as a reference. Note the complexity of the change in image statistics." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 480, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cially increasing with the introduction of variable intensity multicolor direct light sources. Naturally, the most notable difference comes from the color hue deviation. Materials adopt the color of incoming light, with the reflected light wavelength depending on the reflectance defining the represented materials, and the incoming light properties.", + "bbox": [ + 89, + 364, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, the complex geometry of the scene drives light occlusions, with numerous self-shadows at the image level. In this case, given the non-uniform color distribution, the color of the observed shadows depends on the local lighting availability and their light source-defining properties. A successful restoration of the ambient-lit image can not be performed without a full understanding of both scene geometry (for local exposure compensation) and lighting system configuration (for chromaticity uniformization, given the arithmetic of the locally available light radiation given the material properties).", + "bbox": [ + 89, + 455, + 482, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. CL3AN", + "text_level": 1, + "bbox": [ + 89, + 633, + 187, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CL3AN represents an image collection representing Triplets consisting of images acquired under both Color and White Lighting, and Ambient Normalized images, under uniformly distributed diffuse light.", + "bbox": [ + 89, + 657, + 482, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As introduced in Fig. 2, CL3AN is the only dataset providing equivalent representations of the same scene under white-aligned direct lighting (setups (B), (C)), varying intensity multicolor lighting (setup (D)), while also providing an ambient-lit reference image (setup (A)). This makes CL3AN a first-of-its-kind dataset.", + "bbox": [ + 89, + 719, + 482, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The capturing system relies on a professional photography studio setup, in which a dark room is used to ensure that all image-capturing parameters become trackable and controllable. Two separate scene lighting systems are concurrently developed, optimizing for uniform diffuse light distribution, respectively direct lighting, being triggered only", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for the acquisition of the corresponding samples. Note that directional lights are fully controllable, so they can be tuned to the white natural light domain, producing images similar to those available in datasets such as AMBIENT6K [65].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The images are captured with a tripod-mounted and remotely operated Canon R6 MKII camera, at constant focal range. This leads to a perfect pixel alignment between the images characteristic to each triplet. For consistency and traceability, the camera-related parameters are kept fixed during the image-capturing process. All hardware parameters relevant to data acquisition are documented in Tab. 1.", + "bbox": [ + 511, + 152, + 903, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the direct lighting system, three programmable mobile RGB lights are used to capture scene variants under variable light color, intensity, position, and orientation. The direct consequence of this set of parameters is a nonhomogeneous distribution of the incoming light color across the scene, implying that global color correction can not be used for restoration in this scenario (see Fig. 2 and Fig. 3).", + "bbox": [ + 511, + 258, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diffuse ambient lighting system is composed of five white softbox lights, covering all sides of the cubic acquisition environment, except the base, which is the surface supporting the captured contents. Thus, the possibility for self-shadows in the ambient-lit images is almost fully mitigated, being furthermore enforced through geometry-based object filtering. Calibration over a 24-sample color checker under full white light intensity enables the alignment of both lighting setups, by determining the set of operating parameters producing a similar colors under optimal exposure.", + "bbox": [ + 511, + 366, + 905, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CL3AN is a scene-based dataset, covering 105 cluttered scenes representing an extensive set of materials, both conductors and dielectrics, with varying transparency, color, and surface roughness. The data is available as both RAW and RGB images, at $24\\mathrm{MP}$ , the camera sensor resolution. Data splitting is performed based on the represented contents, with 10 scenes used exclusively for testing, and 10 scenes for validation. Both testing and validation splits represent objects unseen during training, but under lighting conditions consistent with the training split. Additional metadata listing the objects within a scene is provided. The training split consists of a number of 3667 samples, with 437 images used for validation and 431 for testing.", + "bbox": [ + 511, + 517, + 905, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. RLN", + "text_level": 1, + "bbox": [ + 513, + 726, + 594, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 and Fig. 3 show the challenging environment posed by the introduction of multicolor direct lighting in the ALN field. Dramatic changes can be observed when mapping from the input image to the ambient-lit reference, especially in terms of observed chromaticity. Moreover, light occlusion by the many rough surfaces corresponding to the represented objects drives a wide variety of intricate effects, with a strong information loss. Learning a robust ALN mapping requires a robust design, able to disentangle significant information compensation in confined image segments from", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9221", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d5c4bb563cf260bf05c38f5d3b0cb788cbcce04a4ce5ff303094cb3fbfc36399.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Camera parametersDiffuse LightingDirect Lighting
Exposure time (s)ApertureISOWhite balanceFocal dist. (mm)# Active lightsWhite light temp. (K)White light temp. rangeIntensity# Active lightsWhite light temp. (K)Intensity rangeColor Hue range
1/60f111006400K5056400±5%100%≥ 1640030-100%0°-360°
", + "bbox": [ + 96, + 78, + 903, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Dataset capturing parameters regarding the setup of the camera, the diffuse lighting system, and the direct lighting hardware.", + "bbox": [ + 96, + 154, + 898, + 169 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cfaea4fd6f24a885337942fc952b347d0581a2612a8f2b74d9887db7c0d1f8cf.jpg", + "image_caption": [ + "Figure 4. A graphical representation of the $\\mathrm{RLN}^2$ architecture, including detailed representations for an Encoder (E) - Decoder (D) blocks pair on the Luminance L-branch. The Reflectance R-branch design is similar, being omitted for clarity." + ], + "image_footnote": [], + "bbox": [ + 109, + 178, + 460, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the large magnitude chromaticity shifts.", + "bbox": [ + 89, + 481, + 352, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given these observations, problem sub-division is the core of our proposed design $\\mathrm{RLN}^2$ , aiming for joint independent compensation in both Image Reflectance and Luminance components. Under the Retinex theory [43], a given image $I$ can be decomposed as a product of two multiplicative terms, representing image illumination or lightness $L$ , and the material specific image reflectance $R$ .", + "bbox": [ + 89, + 496, + 482, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, an image pair $(I,I^{\\prime})$ in which $I^{\\prime}$ is the degraded version of the reference image $I$ , can be represented by the component set $\\{(L,R),(L',R')\\}$ , in line with Eq. (1).", + "bbox": [ + 89, + 601, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} I (x, y) = L (x, y) \\cdot R (x, y) \\\\ I ^ {\\prime} (x) = I ^ {\\prime} (x) - R ^ {\\prime} (x) \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 654, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI ^ {\\prime} (x, y) = L ^ {\\prime} (x, y) \\cdot R ^ {\\prime} (x, y)\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 672, + 385, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, our goal is to determine the pair $(\\overline{L},\\overline{R})$ in which $\\overline{L}$ restores the details lost due to inadequate lighting, while $\\overline{R}$ aligns the input chromaticity with the white-aligned ambient lit setup.", + "bbox": [ + 89, + 698, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL (x, y) = L ^ {\\prime} (x, y) + \\bar {L} (x, y) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 765, + 482, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR (x, y) = R ^ {\\prime} (x, y) + \\bar {R} (x, y)\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 787, + 385, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 provides a graphical representation of the proposed $\\mathrm{RLN}^2$ , in which the $(R', L')$ components of the input $I'$ are subjected to joint independent compensation, while the alternative HSV image representation information is used for feature refinement. Aligned with the properties of the HSV image representation, which codes color", + "bbox": [ + 89, + 810, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in a cylindrical coordinate system, in which the Value (V) channel is similar to explicit illumination intensity, and the Hue-Saturation (HS) represents color tones, the V channel is used as guidance on the L-branch of our model, while the HS pair drives feature refinement on the R-branch of RLN $^2$ .", + "bbox": [ + 511, + 181, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the encoder level, the information goes through successive decomposition, in which Haar-DWT is used for spectral dissemination, alongside RGB feature downsampling. The dual-domain information will be altered in the feature refinement stage, at the $R$ block level, through the Cross-Domain Feature Fusion Attention module (CDFFA).", + "bbox": [ + 511, + 257, + 903, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the low-frequency information $X_{lf}$ , and the RGB extracted features $X_{L'}$ , projected to an n-dimensional space, are then subjected to relevance enhancement, through a set of parameters based on dot product vector similarity (see Fig. 5) to features extracted from the guidance component. The V channel controls the estimation of the $\\overline{L}$ residual, while H and S drive the reflectance estimation $\\overline{R}$ .", + "bbox": [ + 511, + 347, + 906, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {L ^ {\\prime}} = p r o j _ {n} \\left(X _ {L ^ {\\prime}}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 479, + 756, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {l f} = \\operatorname {p r o j} _ {n} \\left(X _ {l f}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 498, + 903, + 523 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = s i m (X _ {L ^ {\\prime}, i}, X _ {V, i})\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 517, + 792, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {i} = s i m (X _ {l f, i}, X _ {V, i})\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 536, + 790, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The set of similarity coefficients is used to define a relevance probability distribution, then used to enhance the relevant information by filtering outlier features. The enhanced information is further refined before output.", + "bbox": [ + 511, + 564, + 905, + 625 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\operatorname {s o f t m a x} (\\alpha)\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 637, + 714, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = \\operatorname {s o f t m a x} (\\beta) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 656, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {o u t} = \\left\\{\\alpha_ {i} \\cdot X _ {L ^ {\\prime}, i} \\right\\} _ {i = 0} ^ {n} + \\left\\{\\beta_ {i} \\cdot X _ {l f, i} \\right\\} _ {i = 0} ^ {n}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 674, + 849, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the decoder stage, the refined information is concatenated to the high-frequency spectral information from the corresponding encoder block. Moreover, an ImageNet [40] pretrained ConvNeXt-based feature extractor module [49], provides wide contextual information in feature fusion. A cross-attention module handles information exchange between the wide context information and the dual-domain features refined in the previous stage.", + "bbox": [ + 511, + 703, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a fair comparison, at this level, $\\mathrm{RLN}^2$ benefits from priors identical to [65], even if the implemented strategies are different. After linear projections, irrelevant features are suppressed through Channel Attention [74], being further refined for Reflectance/Luminance compensation [82].", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9222", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5911921078cd2997832e0e46fc94616aae90386352e67c0c7b58314d0470d5c6.jpg", + "image_caption": [ + "Figure 5. A detailed Reflectance/Luminance Refinement Block (R.Block), including the CDFFA attention module." + ], + "image_footnote": [], + "bbox": [ + 205, + 85, + 366, + 258 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9afcf01e9b8e49a782ae4b0b8d6045adbeecf0d4a17cc68a30e8e2baf3eb4df9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFusionGuidanceMACs (G.)PSNR↑SSIM↑LPIPS↓
\\( RLN^2 \\)NoneNone20.9519.8920.6890.271
\\( RLN^2 \\)concat.RGB23.6720.0180.7030.233
\\( RLN^2 \\)concat.L*a*b*23.6820.1490.7240.226
\\( RLN^2 \\)concat.HSV23.6820.2160.7310.220
\\( RLN^2 \\)(CDFFA)HSV22.7220.5230.7460.208
", + "bbox": [ + 91, + 306, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Inner-stage guidance feature fusion ablation study conducted on the testing split of the CL3AN dataset.", + "bbox": [ + 89, + 397, + 482, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablative study: Tab. 2 compares the standard RLN $^2$ configuration, as represented in Fig. 4, to three similar variants in which the control mechanism implemented in the middle stage is modified. Reducing the $R.$ block to simple feature refinement by removing any control mechanism shows an important performance deficit, consistent in image restoration fidelity and perceptual quality. Alternatively, using the RGB domain images, passed through a dynamical filter and then split based on the filtered output and its alternative residual shows a slight improvement, at a relatively low induced computational cost.", + "bbox": [ + 88, + 438, + 482, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Task-aligned image representations, with explicit coding of image luminance and chromaticity, prove rich in relevant information, adding only the image conversion cost to the general computation. Focusing on relevant feature enhancement through the CDFFA introduction further increases model robustness and restoration performance.", + "bbox": [ + 89, + 609, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation: $\\mathrm{RLN}^2$ was developed in PyTorch, and the documented experiments were performed on up to three NVIDIA L40s 48 GB VRAM GPUs running a CUDA 12.6 environment. The optimization is split into multiple stages, with the resolution of the training patches gradually increasing as the training procedure progresses. The optimizer used was Adam, with the learning rate $lr = 0.0002$ , and the loss optimized was the L1 distance. A cosine learning rate scheduler was used, with two oscillation periods. Optimization deploys gradient clipping, with a standard value of 0.01 under the L1 norm. Further details regarding the training procedure description can be found in the attached supplementary material.", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 513, + 89, + 599, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Setup: Additionally to the introduced CL3AN dataset, we perform experiments on the AMBI-ENT6K image database [65], which consists of 6000 images representing scenes under the effect of up to three white-aligned LED directional lights (see Fig. 3).", + "bbox": [ + 511, + 114, + 903, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As compared metrics, we report restoration fidelity through Peak Signal to Noise Ratio (PSNR), and Structure Similarity Index Measure (SSIM) [70]. To quantify the perceptual quality of the restored images, we deploy the AlexNet-based Learned Perceptual Image Patch Similarity score (LPIPS) [86].", + "bbox": [ + 511, + 191, + 903, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of compared methods, we chose various types of solutions proposed in the image restoration field, which became seminal work after their introduction. For an extensive benchmark, we combine refined CNN solutions such as MPRNet [83] and HINet [37], with transformer-based architectures like SwinIR [47], NAFNet [10], Uformer [71], Restormer [84], and GRL [80]. Additionally, we consider solutions integrating state-space-models (e.g. MAMBAIR [21]), or solutions optimizing joint Spatial-Frequency entropy such as SFNet [12] and IFBlend [65]. Ultimately, we include Retinexformer [9], representing the class of solutions based on Retinex image decomposition. We train all the aforementioned methods using their official public implementations and settings.", + "bbox": [ + 511, + 281, + 903, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results: In Tab. 3, we compare the aforementioned solutions with different variants of the proposed $\\mathrm{RLN}^2$ . The different variants represent our proposed models at different complexity levels, such that when comparing to similar existing solutions, advantages become easily identifiable. The computational complexity, as the Multiplications-Additions count (MAC), is reported for a patch of $128\\times 128$ pixels, using ptflops for consistency.", + "bbox": [ + 511, + 492, + 903, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The naming convention can be easily followed. Firstly, the difference between the RLN $^2$ -S and the RLN $^2$ -L is given by the complexity of the Wide Context Feature Extractor found in the feature fusion module, at the decoder stage, with RLN $^2$ -S using a ConvNeXt-T configuration, while RLN $^2$ -L benefits from a large ConvNeXt-XL model [49]. Secondly, we distinguish between RLN $^2$ variants that have/have not access to the dual image frequency domain representation. Thus RLN $^2$ -S and RLN $^2$ -L drop the frequency domain feature at the encoder, refinement stage, and decoder stage, using cross-attention for fusion, while RLN $^2$ -Sf and RLN $^2$ -Lf use all the available information.", + "bbox": [ + 511, + 613, + 903, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All the compared RLN $^2$ variants are trained under the same conditions, including the optimizer, batch size, data augmentation strategies, optimized objective, and the number of samples seen during training. For AMBIENT6K [65], we adopt their publicly available benchmark.", + "bbox": [ + 511, + 794, + 903, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reading Tab. 3, it becomes obvious that $\\mathrm{RLN}^2$ is a high-performance solution, improving over similar state-of-the", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9223", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/84370e84a46d04b026a9061eed8b7db0076a71406aee594e75b204c73bb9ae3b.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on scene #1 from test split of the CL3AN dataset. Best viewed in the electronic version." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 908, + 340 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1cab81132df09f5dc84aec075ea4802c72de1e6f6a3ad73373abad1acde2c1f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method nameTypeGeneral informationCL3AN (ours)AMBIENT6K [65]
Restoration TaskPriorMACs (G.)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
unprocessed----10.8370.4470.51813.4030.6520.250
NAFNet[10]Conv.Multi-task IRRGB15.9219.4760.7090.24920.5800.8080.142
MPRNet [83]Conv.Multi-task IRRGB37.2118.4530.6880.29120.9470.8200.129
SFNet [12]Transf.Multi-task IRRGB + Freq.31.2718.3820.6860.29120.5190.8120.141
SwinIR [47]Transf.Multi-task IRRGB37.8116.3860.6430.37220.5280.8170.131
Uformer [71]Transf.Multi-task IRRGB19.3317.5080.6550.31320.7760.8180.131
Restormer [84]Transf.Multi-task IRRGB35.3118.5600.6910.27821.1410.8170.132
HINet [37]Conv.Multi-task IRRGB42.6819.3880.7070.24820.8560.8210.129
IFBlend [65]ConvALNRGB + Freq.26.0120.3700.7200.22821.4430.8190.128
MAMBAIR [23]Transf. + SSMMulti-task IRRGB34.3218.9700.7040.254---
GRL [80]Transf.Multi-task IRRGB2.1618.0890.6720.308---
Retinexformer [9]Transf.LLIERGB4.8618.6490.6830.281---
\\( RLN^2-S \\) (ours)ConvALNRGB3.9519.9920.7180.23621.1810.8150.131
\\( RLN^2-Sf \\) (ours)ConvALNRGB + Freq.4.2420.1280.7300.22321.3330.8190.128
\\( RLN^2-L \\) (ours)ConvALNRGB22.4520.3830.7230.22221.5530.8210.123
\\( RLN^2-Lf \\) (ours)ConvALNRGB + Freq.22.7220.5230.7460.20821.7120.8250.120
", + "bbox": [ + 94, + 359, + 906, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Various RLN $^2$ variants versus current SOTA image restoration models, on the proposed CL3AN dataset test split (1920 × 1440 px.), and AMBIENT6K benchmark [65] (1280 × 960 px. default size). The best and the second best performing models are emphasized.", + "bbox": [ + 89, + 646, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "art solutions at all complexity levels, on both CL3AN and AMBIENT6K [65]. While $\\mathrm{RLN}^2$ -S largely outperforms RetinexFormer [9], at a lower computational cost, $\\mathrm{RLN}^2$ -Sf improves over the perceptual performance of IFBlend [65], with a fraction of its required floating point operations.", + "bbox": [ + 89, + 683, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Against IFBlend [65], a strong recently proposed solution in the ALN field, $\\mathrm{RLN}^2$ -L shows an advantage, especially in joint RGB-Frequency domains processing, with an $\\approx 15\\%$ complexity reduction, while outperforming it.", + "bbox": [ + 89, + 761, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Another important observation is regarding the architectural core of the compared methods. Under both white-aligned and RGB direct lighting, ALN tends to favor solutions showing an extensive receptive field [37, 83], rather than transformers optimized for local dependencies [47, 71,", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "84], even when balancing to wider context blocks [80]. However, MPRNet [83] and HINet [37] are larger models, with an additional cost of approx. $63\\% - 87\\%$ , compared to the highest complexity RLN²-Lf variant. While state-state models prove computationally ineffective for the level of performance achieved, frequency-domain information access shows a consistent advantage, especially in terms of observed perceptual quality.", + "bbox": [ + 511, + 683, + 906, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results: In Fig. 6, we provide for qualitative evaluation a set of samples representing a cluttered scene under the effect of non-homogeneous RGB direct lighting. Note the complex geometry and the wide RGB spectrum represented at the ambient lit image level. $\\mathrm{RLN}^2$ -Lf shows improved behavior in terms of color normalization (row", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9224", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d78f011811514e37254f59a4686e5871d6252b64c8b02dde83267fae95fcdb3.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 102, + 259, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc79a030fe1552ecff7329773b69dfe881f9d0139781ef004b117c0d5d8ac069.jpg", + "image_caption": [ + "Restormer [84]" + ], + "image_footnote": [], + "bbox": [ + 259, + 102, + 423, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/911dcd32925decfa6917cb4c8e710558af1a89e1dc0f849723c425ed092c052d.jpg", + "image_caption": [ + "IFBlend [65]" + ], + "image_footnote": [], + "bbox": [ + 424, + 102, + 586, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b25e940749d3e4d41461e8685f3bf17247e00e4a76f22f180241f0c21170d135.jpg", + "image_caption": [ + "$\\mathrm{RLN}^2$ -Lf (Ours)" + ], + "image_footnote": [], + "bbox": [ + 588, + 102, + 753, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac627ee46216cadc8170a9b557820aa4bdfc9fd2e9efe3d48bbe85fad075985d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 102, + 919, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f7cf4fe9daeea95e6354c8bdd4de511c799801bb7210fa5eb2d38d87e59d36d.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison challenging samples from the AMBIENT6K testing set [65]. Our $\\mathrm{RLN}^2$ produces improved quality ambient normalized images, with sharper images, and a lower amount of visible artifacts." + ], + "image_footnote": [], + "bbox": [ + 94, + 200, + 258, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c564554c3feb8b4576da27f42da1e4ebae64a8ca4bb33a1a0f2b5e4480fa72d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 200, + 423, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e8b0b6ae3af887a83a546cf328935ef4a65e5088436275c2cb7bb02037e5eb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 200, + 588, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b530f3ba765428d3f9f6504f6c6cdb178e7796815103c86f340c8a72afdb05f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 200, + 753, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a2a4086cf0e714c57920886416500b4edf81d4e0bfc135cb8965d4cc65b3d0f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 200, + 919, + 294 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/23f5cb000b6f1a367762fdb30f508ecffd5b966eac8bbe21db3bb7ec48320ff5.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 93, + 330, + 287, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8deaa16b60f6fb1936ce7c745081118a29eb0e617197d9366dafae8870ab643f.jpg", + "image_caption": [ + "Input", + "RLN²-Lf", + "Figure 8. IFBlend[65] and $\\mathrm{RLN}^2$ -Lf on real-world images from LSMI [38] (left), or real product photography samples (right)." + ], + "image_footnote": [], + "bbox": [ + 287, + 330, + 480, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1), superior restorations in highlights-rich areas, and better high-frequency shadow areas compensation (row 2).", + "bbox": [ + 89, + 491, + 482, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 provides for visual comparison equivalent renderings against the top performing state-of-the-art methods the AMBIENT6K [65] benchmark. Consistently, the restored ambient normalized images are characterized by more natural colors, an increased amount of restored details, and geometrically correct renderings even in areas severely affected by underexposure (second row). Overcoming the challenges specific to non-homogeneous direct color lighting, RLN $^2$ shows consistent performance, confirming the advantages quantified through evaluations provided in Tab. 3.", + "bbox": [ + 88, + 521, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To test $\\mathrm{RLN}^2$ on out-of-domain data, in Fig. 8 we compare IFBlend [65] and our proposed $\\mathrm{RLN}^2$ -Lf on inputs from LSMI [38], or complex lighting scenarios from product photography. Our $\\mathrm{RLN}^2$ -Lf shows increased robustness, with plausible rendered colors under uniform white ambient lighting conditions. Moreover, in Fig. 9 we deploy it on AI-generated images representing various contents in challenging lighting environments. Applying our ALN solution as a pre-processing step in Neural Image Editing proves beneficial for IC-Light [19], with improved layer separation leading to less semantic discontinuities (rows 1, 2, and 4), and improves the rendered results quality in term of shadow casting (row 3) and foreground clarity (row 2).", + "bbox": [ + 88, + 672, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5195b0f5809b0f174ee7d2a3c710d956905c7ce436a213f2eb939c6958f370b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 330, + 709, + 486 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0dc31a86c19e9d7c73ea3a682caddcaf6768d2b8b0a650e7a28ef00c6203f908.jpg", + "image_caption": [ + "Input", + "$\\mathbf{RLN}^2$ -lfout" + ], + "image_footnote": [], + "bbox": [ + 517, + 486, + 707, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76bb35738fc727ad942ba23e1aeceabda5f446a32b30bd3bff75e2e6176527f8.jpg", + "image_caption": [ + "w.o. RLN²-Lf", + "W. RLN²-Lf", + "Figure 9. ALN as a preprocessing step before image relighting with foreground condition. Results rendered with ICLight [19] for the input image (col. 3) vs. the $\\mathrm{RLN}^2$ output (col. 4)." + ], + "image_footnote": [], + "bbox": [ + 710, + 330, + 903, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 707, + 633, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose extending the current ALN study to the challenges added by multicolor and multi-intensity RGB direct lighting. We propose CL3AN, a high-resolution image database showcasing a multitude of materials under complex RGB lighting scenarios while providing white-light ambient-lit reference images using a professional photography setup. We introduce RLN², a high-performance ALN algorithm setting a new standard for the ALN task, showing how alternative image representations can provide valuable control signals in feature refinement, or offer natural bypassing options for detail preservation.", + "bbox": [ + 509, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "deepai.org/machine-learning-model/text2img", + "bbox": [ + 114, + 886, + 455, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9225", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We thank the reviewers and ACs for their sustained effort and helpful feedback. This research was partially supported by the Alexander von Humboldt Foundation.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1397-1406, 2020. 2", + "[2] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1981-1990, 2021. 2", + "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 2", + "[4] Alexandre Araujo, Jean Ponce, and Julien Mairal. Towards real-world focus stacking with deep learning. arXiv preprint arXiv:2311.17846, 2023. 2", + "[5] Monika Bansal, Munish Kumar, and Manish Kumar. 2d object recognition: a comparative analysis of sift, surf and orb feature descriptors. Multimedia Tools and Applications, 80 (12):18839-18857, 2021. 2", + "[6] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2", + "[7] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 1", + "[8] Ivaylo Boyadzhiev, Kavita Bala, Sylvain Paris, and Frédo Durand. User-guided white balance for mixed lighting conditions. ACM Trans. Graph., 31(6):200-1, 2012. 2", + "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 2, 3, 6, 7", + "[10] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 3, 6, 7", + "[11] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 2", + "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 6, 7" + ], + "bbox": [ + 93, + 194, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[14] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 2", + "[15] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2", + "[16] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 2", + "[17] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, and Marc Pollefeys. Building rome on a cloudless day. In Computer Vision - ECCV 2010, pages 368-381, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg. 2", + "[18] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 203-212, 2021. 3", + "[19] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 1, 8", + "[20] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2", + "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 6", + "[22] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5917-5926, 2023. 2", + "[23] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2, 3, 7", + "[24] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 2", + "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3", + "[26] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF interna" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9226", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional conference on computer vision, pages 2472-2481, 2019.3", + "[27] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0-0, 2018. 1", + "[28] Yan Huang, Xinchang Lu, Yuhui Quan, Yong Xu, and Hui Ji. Image shadow removal via multi-scale deep retina decomposition. Pattern Recognition, page 111126, 2024. 3", + "[29] Zhuo Hui, Kalyan Sunkavalli, Sunil Hadap, and Aswin C. Sankaranarayanan. Illuminant spectra-based source separation using flash photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6209-6218, 2018. 3", + "[30] Zhuo Hui, Ayan Chakrabarti, Kalyan Sunkavalli, and Aswin C Sankaranarayanan. Learning to separate multiple illuminants in a single image. In Computer Vision and Pattern Recognition (CVPR 2019), 2019. 3", + "[31] Anne Humeau-Heurtier. Texture feature extraction methods: A survey. IEEE access, 7:8975–9000, 2019. 1", + "[32] Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 536-537, 2020. 1", + "[33] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 1", + "[34] Jong Ju Jeon, Jun Young Park, and Il Kyu Eom. Low-light image enhancement using gamma correction prior in mixed color spaces. Pattern Recognition, 146:110001, 2024. 2", + "[35] Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion. Advances in Neural Information Processing Systems, 37:141129-141152, 2025. 1, 2", + "[36] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. arXiv preprint arXiv:2211.08089, 2022. 1", + "[37] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2021. 3, 6, 7", + "[38] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 2, 3, 8", + "[39] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3", + "[40] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 5", + "[41] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 2", + "[42] Wei-Sheng Lai, Orazio Gallo, Jinwei Gu, Deqing Sun, Ming-Hsuan Yang, and Jan Kautz. Video stitching for linear camera arrays. arXiv preprint arXiv:1907.13622, 2019. 2", + "[43] EDWIN H. LAND. The retina. American Scientist, 52(2): 247-264, 1964. 3, 5", + "[44] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 1, 3", + "[45] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Mimt: Multi-illuminant color constancy via multitask local surface and light color learning. arXiv preprint arXiv:2211.08772, 2022. 2", + "[46] Xinyu Li, Yanyi Zhang, Jianbo Yuan, Hanlin Lu, and Yibo Zhu. Discrete cosin transformer: Image modeling from frequency domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5468-5478, 2023. 3", + "[47] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 3, 6, 7", + "[48] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 3", + "[49] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 6", + "[50] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1680-1691, 2023. 3", + "[51] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 1", + "[52] Radoslaw Mantiuk, Rafal Mantiuk, Anna Tomaszewska, and Wolfgang Heidrich. Color correction for tone mapping. In Computer graphics forum, pages 193-202. Wiley Online Library, 2009. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9227", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523-3542, 2021. 1", + "[54] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 3", + "[55]Punatawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22646-22657, 2023. 2", + "[56] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 1, 3", + "[57] Pramod Rao, Gereon Fox, Abhinitra Meka, Mallikarjun BR, Fangneng Zhan, Tim Weyrich, Bernd Bickel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, et al. Lite2relight: 3d-aware single image portrait relighting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1", + "[58] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 2", + "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2", + "[60] Richard Szeliski and Richard Szeliski. Image alignment and stitching. Computer Vision: Algorithms and Applications, pages 401-441, 2022. 2", + "[61] Hironobu Takano, Hiroki Kobayashi, and Kiyomi Nakamura. Rotation invariant iris recognition method adaptive to ambient lighting variation. IEICE transactions on information and systems, 90(6):955-962, 2007. 2", + "[62] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20762-20772, 2023. 1, 2", + "[63] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2", + "[64] Florin-Alexandru Vasluianu, Tim Seizinger, and Radu Timofte. Wsrd: A novel benchmark for high resolution image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1826-1835, 2023. 1, 2, 3", + "[65] Florin-Alexandru Vasluianu, Tim Seizinger, Zongwei Wu, Rakesh Ranjan, and Radu Timofte. Towards image ambient" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "lighting normalization. arXiv preprint arXiv:2403.18730, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[66] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[67] Qiyang Wan, Ruiping Wang, and Xilin Chen. Interpretable object recognition by semantic prototype analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 800-809, 2024. 1", + "[68] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 3", + "[69] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 3", + "[70] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[71] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 3, 6, 7", + "[72] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 3", + "[73] Ching-Chih Weng, Homer Chen, and Chiou-Shann Fuh. A novel automatic white balance method for digital still cameras. In 2005 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3801-3804. IEEE, 2005. 2", + "[74] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5", + "[75] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 3", + "[76] Xudong Xie and Kin-Man Lam. An efficient illumination normalization method for face recognition. Pattern Recognition Letters, 27(6):609-617, 2006. 2", + "[77] Xiaohua Xie, Wei-Shi Zheng, Jianhuang Lai, Pong C. Yuen, and Ching Y. Suen. Normalization of face illumination based on large-and small-scale features. IEEE Transactions on Image Processing, 20(7):1807-1821, 2011. 2", + "[78] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. International Journal of Computer Vision, 132(9):3610-3632, 2024. 3", + "[79] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9228", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[80] Xiaoyu Xiang Denis Demandolx Rakesh Ranjan Radu Timofte Yawei Li, Yuchen Fan and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 6, 7", + "[81] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 1", + "[82] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 5", + "[83] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3, 6, 7", + "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3, 6, 7, 8", + "[85] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 3", + "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[87] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 1", + "[88] Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed images via robust retina decomposition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3", + "[89] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1219-1229, 2023. 3", + "[90] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12786-12796, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 795 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9229", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_model.json b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94f16329ee55e55e34069afc062cfba13f315eff --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_model.json @@ -0,0 +1,2963 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "After the Party: Navigating the Mapping From Color to Ambient Lighting" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.181, + 0.808, + 0.218 + ], + "angle": 0, + "content": "Florin-Alexandru Vasluianu Tim Seizinger Zongwei Wu* Radu Timofte Computer Vision Lab, CAIDAS & IFI, University of Wurzburg" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.22, + 0.661, + 0.234 + ], + "angle": 0, + "content": "{firstname.lastname}@uni-wuerzburg.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.692 + ], + "angle": 0, + "content": "Illumination in practical scenarios is inherently complex, involving colored light sources, occlusions, and diverse material interactions that produce intricate reflectance and shading effects. However, existing methods often oversimplify this challenge by assuming a single light source or uniform, white-balanced lighting, leaving many of these complexities unaddressed. In this paper, we introduce CL3AN, the first large-scale, high-resolution dataset of its kind designed to facilitate the restoration of images captured under multiple Colored Light sources to their Ambient-Normalized counterparts. Through benchmarking, we find that leading approaches often produce artifacts—such as illumination inconsistencies, texture leakage, and color distortion—primarily due to their limited ability to precisely disentangle illumination from reflectance. Motivated by this insight, we achieve such a desired decomposition through a novel learning framework that leverages explicit chromaticity-luminance components guidance, drawing inspiration from the principles of the Retinex model. Extensive evaluations on existing benchmarks and our dataset demonstrate the effectiveness of our approach, showcasing enhanced robustness under non-homogeneous color lighting and material-specific reflectance variations, all while maintaining a highly competitive computational cost. The benchmark, codes, and models are available at www.github.com/fvasluianu97/RLN2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.722, + 0.222, + 0.737 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.484, + 0.869 + ], + "angle": 0, + "content": "The goal of Ambient Lighting Normalization is recovering an ambient-lit image, under uniform lighting distribution, regardless of the lighting scenario characteristic to the depicted contents [65]. The ability to recover ambient-normalized images—free from spatially varying illumination artifacts—is critical for robust visual understanding [33, 51, 53]. In both practical imaging scenarios [27, 32] and the rapidly evolving domain of image generation and" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.272, + 0.905, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.466, + 0.908, + 0.551 + ], + "angle": 0, + "content": "Figure 1. Effective deployment of Ambient Lighting Normalization (ALN) as a pre-processing step in Neural Image Editing [19]. We compare the effectiveness of our proposed RLN² against state-of-the-art models in fields like Shadow Removal (DES3 [36]), or ALN (IFBlend [65]), trained on settings including single (1) [68] or multiple (3) white-domain-tuned [65] or color light sources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.699 + ], + "angle": 0, + "content": "synthesis [57, 81], illumination always involves complex interactions of multiple colored light sources, occlusions, and material properties, resulting in intricate shadows, glare, color shifts, and texture distortions. Lighting and material characteristics are the driving factors for explaining the manifestation of such artifacts, thus explaining the need for data-level representativity prior to proposing image restoration solutions aiming for improved degradation robustness or restoration performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.822 + ], + "angle": 0, + "content": "While a wide range of applications require illumination-invariant representations [7, 31, 59, 67], existing normalization methods struggle to disentangle these intertwined factors [19, 35, 62, 81], as illustrated in Fig. 1. Even if data contributions further away from the natural light in \\(\\mathrm{STD + }\\) [44, 68], or multiple white-light-tuned light sources from AMBIENT6K [65] can improve general effectiveness, model design can further drive robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This limitation stems primarily from oversimplified assumptions, such as single white light sources [64, 68] or homogeneous color fields [56, 68, 87], which fail to capture the richness of complex illumination. A key technical challenge lies in the lack of comprehensive datasets that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.199, + 0.901 + ], + "angle": 0, + "content": "*Corresponding" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9218" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "adequately represent the complexity of real-world lighting, especially outside the realm of natural white-color-tuned lighting. Prior efforts in shadow removal [64, 68], color constancy [6, 15, 16], or color correction [1, 22, 34, 52, 73] address isolated aspects of illumination but fall short of modeling multi-light interactions. Recent attempts [38, 65] introduce variable-intensity lighting datasets but lack chromatic diversity and material-specific reflectance variations. Furthermore, wide access to ambient-lit images outside the scope of scene-rendering software remains a challenge. As a result, learning-based methods often produce artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.472 + ], + "angle": 0, + "content": "In this paper, we address these limitations through two core contributions. First, we introduce CL3AN, the first large-scale dataset capturing high-resolution scenes under multiple colored directional lights paired with ambient-lit references. Each scene triplet includes variations beyond white-balanced lighting, colored lighting, and a diffuse-lit ground truth, enabling systematic study of illumination effects like self-shadows, highlights, glare areas, color spill or saturation, and other material-dependent artifacts. By benchmarking leading image restoration methods [9, 23, 65, 80], we observe that these models frequently suffer from critical shortcomings, such as the conflation of chromatic variations caused by illumination, texture leakage, inconsistent partial normalization, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.718 + ], + "angle": 0, + "content": "We identify the fundamental limitation of existing methods as their inability to systematically disentangle reflectance and illumination—a critical flaw exacerbated by the lack of explicit guidance for separating chromaticity and luminance under multi-colored lighting. Most current approaches [23, 80, 84] rely on single-stream, end-to-end feature modeling, which implicitly maps input images to ambient-normalized counterparts without incorporating physical awareness of light-material interactions. While some recent works attempt to address this by introducing pixel- and frequency-domain supervision [12, 65], more flexible towards reflectance- and chromaticity-information extraction, they remain limited in learning appropriate frequency filters — a challenging task without explicit guidance — resulting in suboptimal performance under the challenging multi-colored lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, from a novel perspective, we propose \\(\\mathbf{RLN}^2\\), a Retinex-inspired framework that introduces parallel, physics-grounded streams to independently model reflectance (material properties) and illumination (lighting geometry/color). We observe that these intrinsic specificities are naturally encoded in the cylindrical representation of the input image, where the Hue (H) and Saturation (S) maps align with the observed chromaticity under the effect of the incoming light, while the Value (V) map provides strong priors for materials-specific reflectance. By embedding these domain-specific clues — such as reflectance information entropy (V), and smoothness constraints (H, S)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "into feature refinement, our framework achieves precise decomposition, effectively eliminating undesirable artifacts while preserving fine-grained details." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.838, + 0.15 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.152, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- The introduction of the CL3AN, a first-of-its-kind dataset enabling the study of ambient lighting normalization under multiple RGB lights;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- The extensive study of a wide range of models, seminal works in various image restoration tasks, when trained and tested on the CL3AN dataset;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- We propose \\(\\mathbf{RLN}^2\\), a robust algorithm built on the principle of simultaneous image reflection and lightness components compensation to recover an ambient lit image, characterized by uniform lighting distribution. \\(\\mathbf{RLN}^2\\) is proposed as a strong baseline for the ambient lighting normalization task, achieving state-of-the-art results on both the public AMBIENT6K dataset and the proposed CL3AN benchmark." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.152, + 0.905, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.655, + 0.393 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.905, + 0.478 + ], + "angle": 0, + "content": "This work expands Ambient Lighting Normalization to multicolor direct lighting, addressing previously unexplored effects appearing in cluttered scenes, and analyzing shadow interactions and highlights appearing under multicolor direct lighting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.905, + 0.569 + ], + "angle": 0, + "content": "Lighting and Color: Color analysis is central to image processing and understanding, with various downstream tasks such as facial recognition [61, 76, 77], or tracking and matching applications [5, 14, 41, 59, 63] relying on the color similarity principle, thus benefiting from increased appearance robustness, especially under challenging lighting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Image Quality Assessment (IQA) methods [3, 79] show a strong correlation between details-abundance and image quality, motivating the focus on professional scene lighting rigs in all expert image photography applications. Given the scarcity of such equipment, alternative strategies [1, 2, 6, 45] are developed around image sensor sensor properties and cross-device constancy [8, 20]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Naturally, registered image stacks either for tasks such as multiple view geometry estimation [17, 90] or different image stitching applications require consistent color and content appearance [4, 11, 42, 60]. In such applications, the effort is focused on consistent lighting in the image acquisition, rather than post-processing correction. In-the-wild applications focusing on widely available amateur images [17, 24, 60] document extensive data curation strategies prior to model estimations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Later applications, given the newly-found popularity of neural 3D modeling, approach scene relighting as part of their rendering pipeline [58, 62]. Alternatively, approaches based on complex generative image models [35, 55] successfully tackle lighting manipulation, but at a high computational cost. Unfortunately, the limited availability of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9219" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.273, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.216, + 0.192, + 0.227 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.281, + 0.093, + 0.508, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.216, + 0.403, + 0.227 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.094, + 0.7, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.216, + 0.614, + 0.227 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.094, + 0.905, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.216, + 0.826, + 0.227 + ], + "angle": 0, + "content": "D" + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.233, + 0.327, + 0.243 + ], + "angle": 0, + "content": "ISTD/WSRD" + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.248, + 0.434, + 0.258 + ], + "angle": 0, + "content": "AMBIENT6K" + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.263, + 0.531, + 0.274 + ], + "angle": 0, + "content": "CL3AN (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.286, + 0.906, + 0.357 + ], + "angle": 0, + "content": "Figure 2. A comparison between the lighting scenarios represented in datasets targeting different applications of Ambient Lighting Normalization (ISTD [68], WSRD[64], or AMBIENT6K [65]). In (A), ambient lighting is used to acquire a uniformly-lit reference image. In (B), direct lighting is used for shadow casting, while constraining consistency to the reference image in shadow-free segments. In (C), more directional lights are added to the direct lighting system. In our proposed setup (D), the direct lighting setup is based on RGB lights, dropping the color consistency constraint, and showcasing complex material-light interactions under various conditions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.482, + 0.432 + ], + "angle": 0, + "content": "real-world ambient lit images, primarily due to the complexity of an ambient lighting setup, explains the difficulty in addressing real-world lighting normalization, even with the obvious availability of large-complexity models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Reliance on natural light for ambient-lit images proved a viable option, but hindered the variety of possible observed degradations, especially in terms of chromaticity shifts. Unfortunately, the large number of factors involved in light occlusion interaction greatly constrains a representative simulation in general settings. While ISTD/ISTD+ [44, 68], SRD [56] or USR [26] show exclusively flat surfaces under the influence of natural light, the influence of object self-shadows was just recently represented in WSRD [64], only possible under a laboratory setup. Recently, multiple source lighting gained in data-level representation. For example, in LSMI [38], the authors target White-Balance correction robustness under natural and artificial lighting. Alternatively, effects characteristic to multiple illuminants were captured in works such as [30, 54], or in the context of flash photography [29]. However, a uniformly-lit ambient normalized image can not be acquired in these settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Alternatively, the introduction of AMBIENT6K [65] coupled multiple light sources of varying intensity and orientation for complex interactions, while also providing an ambient-lit image based on an alternative lighting setup. Unfortunately, the strong connection between the operating parameters set for the direct and ambient lighting systems limits the study to exposure correction for white-aligned scene lighting (see Fig. 2 (A), (B), and (C))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Image Restoration (IR): Joint color and exposure correction, while restoring details lost to occlusion or saturation, defines Ambient Lighting Normalization as an Image Restoration task. Various Convolution Neural Networks (CNNs) [10, 37, 83, 85] became strong baselines for the family of IR tasks. The later introduction of the attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.524 + ], + "angle": 0, + "content": "mechanism [66] in image-based applications [13] enabled consistent performance advantages in the wide IR field. Focusing on different feature fusion techniques through specifically tailored attention mechanisms, based on the IR task particularities, defines the success of multiple models [10, 47, 48, 71, 80, 84] easily becoming benchmark solutions for various IR problems. Separate refinement for image reflectance and lightness under the Retinex model [43] proved particularly effective in low-light image enhancement (LLIE) [9, 72, 78, 88], or image shadow removal [28]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Global range dependencies benefits drive the development of State Space Models (SSM) [21], with promising results in general IR [23]. Alternatively, global knowledge can be leveraged through the integration of alternative frequency image representations in IR solutions. Solutions such as FFTFormer [39], DCFormer [46], SFNet [12], DW-GAN [18], and IFBlend [65] successfully extract priors based on Fourier Transform, Discrete Cosine Transform, or other Wavelet Transform variants." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Finally, diffusion-based iterative recurrent solutions [25] started to mature as benchmark solutions in Image Restoration [50, 69, 75, 89], especially in terms of perceived image quality. However, their characteristic increased computational effort hinders their deployment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.753, + 0.605, + 0.768 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.81 + ], + "angle": 0, + "content": "The core of our work is extending the study of Ambient Lighting Normalization to direct color lighting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 compares the properties of a scene depicted under multi-source color direct lighting, and an equivalent image under white-aligned direct lighting. The goal to recover the ambient-lit reference image. Provided statistics, such as the RGB image histogram, and the Hue/Saturation footprint showcase the complexity of the ALN transformation, espe" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9220" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.485, + 0.351 + ], + "angle": 0, + "content": "Figure 3. A graphical representation of the Ambient Lighting Normalization transformation in the context of the white-aligned lighting, compared to the challenges posed by the introduction of a multicolor direct lighting system. The RGB image histogram, coupled with the Hue-Saturation rough projection is provided as a reference. Note the complexity of the change in image statistics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.483, + 0.456 + ], + "angle": 0, + "content": "cially increasing with the introduction of variable intensity multicolor direct light sources. Naturally, the most notable difference comes from the color hue deviation. Materials adopt the color of incoming light, with the reflected light wavelength depending on the reflectance defining the represented materials, and the incoming light properties." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.622 + ], + "angle": 0, + "content": "Furthermore, the complex geometry of the scene drives light occlusions, with numerous self-shadows at the image level. In this case, given the non-uniform color distribution, the color of the observed shadows depends on the local lighting availability and their light source-defining properties. A successful restoration of the ambient-lit image can not be performed without a full understanding of both scene geometry (for local exposure compensation) and lighting system configuration (for chromaticity uniformization, given the arithmetic of the locally available light radiation given the material properties)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.188, + 0.649 + ], + "angle": 0, + "content": "3.1. CL3AN" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.718 + ], + "angle": 0, + "content": "CL3AN represents an image collection representing Triplets consisting of images acquired under both Color and White Lighting, and Ambient Normalized images, under uniformly distributed diffuse light." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.809 + ], + "angle": 0, + "content": "As introduced in Fig. 2, CL3AN is the only dataset providing equivalent representations of the same scene under white-aligned direct lighting (setups (B), (C)), varying intensity multicolor lighting (setup (D)), while also providing an ambient-lit reference image (setup (A)). This makes CL3AN a first-of-its-kind dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The capturing system relies on a professional photography studio setup, in which a dark room is used to ensure that all image-capturing parameters become trackable and controllable. Two separate scene lighting systems are concurrently developed, optimizing for uniform diffuse light distribution, respectively direct lighting, being triggered only" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "for the acquisition of the corresponding samples. Note that directional lights are fully controllable, so they can be tuned to the white natural light domain, producing images similar to those available in datasets such as AMBIENT6K [65]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.259 + ], + "angle": 0, + "content": "The images are captured with a tripod-mounted and remotely operated Canon R6 MKII camera, at constant focal range. This leads to a perfect pixel alignment between the images characteristic to each triplet. For consistency and traceability, the camera-related parameters are kept fixed during the image-capturing process. All hardware parameters relevant to data acquisition are documented in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.905, + 0.365 + ], + "angle": 0, + "content": "For the direct lighting system, three programmable mobile RGB lights are used to capture scene variants under variable light color, intensity, position, and orientation. The direct consequence of this set of parameters is a nonhomogeneous distribution of the incoming light color across the scene, implying that global color correction can not be used for restoration in this scenario (see Fig. 2 and Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.518 + ], + "angle": 0, + "content": "The diffuse ambient lighting system is composed of five white softbox lights, covering all sides of the cubic acquisition environment, except the base, which is the surface supporting the captured contents. Thus, the possibility for self-shadows in the ambient-lit images is almost fully mitigated, being furthermore enforced through geometry-based object filtering. Calibration over a 24-sample color checker under full white light intensity enables the alignment of both lighting setups, by determining the set of operating parameters producing a similar colors under optimal exposure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.715 + ], + "angle": 0, + "content": "CL3AN is a scene-based dataset, covering 105 cluttered scenes representing an extensive set of materials, both conductors and dielectrics, with varying transparency, color, and surface roughness. The data is available as both RAW and RGB images, at \\(24\\mathrm{MP}\\), the camera sensor resolution. Data splitting is performed based on the represented contents, with 10 scenes used exclusively for testing, and 10 scenes for validation. Both testing and validation splits represent objects unseen during training, but under lighting conditions consistent with the training split. Additional metadata listing the objects within a scene is provided. The training split consists of a number of 3667 samples, with 437 images used for validation and 431 for testing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.596, + 0.741 + ], + "angle": 0, + "content": "3.2. RLN" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Fig. 2 and Fig. 3 show the challenging environment posed by the introduction of multicolor direct lighting in the ALN field. Dramatic changes can be observed when mapping from the input image to the ambient-lit reference, especially in terms of observed chromaticity. Moreover, light occlusion by the many rough surfaces corresponding to the represented objects drives a wide variety of intricate effects, with a strong information loss. Learning a robust ALN mapping requires a robust design, able to disentangle significant information compensation in confined image segments from" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9221" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.079, + 0.904, + 0.151 + ], + "angle": 0, + "content": "
Camera parametersDiffuse LightingDirect Lighting
Exposure time (s)ApertureISOWhite balanceFocal dist. (mm)# Active lightsWhite light temp. (K)White light temp. rangeIntensity# Active lightsWhite light temp. (K)Intensity rangeColor Hue range
1/60f111006400K5056400±5%100%≥ 1640030-100%0°-360°
" + }, + { + "type": "table_caption", + "bbox": [ + 0.097, + 0.155, + 0.9, + 0.17 + ], + "angle": 0, + "content": "Table 1. Dataset capturing parameters regarding the setup of the camera, the diffuse lighting system, and the direct lighting hardware." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.179, + 0.462, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "Figure 4. A graphical representation of the \\(\\mathrm{RLN}^2\\) architecture, including detailed representations for an Encoder (E) - Decoder (D) blocks pair on the Luminance L-branch. The Reflectance R-branch design is similar, being omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.354, + 0.496 + ], + "angle": 0, + "content": "the large magnitude chromaticity shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.602 + ], + "angle": 0, + "content": "Given these observations, problem sub-division is the core of our proposed design \\(\\mathrm{RLN}^2\\), aiming for joint independent compensation in both Image Reflectance and Luminance components. Under the Retinex theory [43], a given image \\(I\\) can be decomposed as a product of two multiplicative terms, representing image illumination or lightness \\(L\\), and the material specific image reflectance \\(R\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.649 + ], + "angle": 0, + "content": "Thus, an image pair \\((I,I^{\\prime})\\) in which \\(I^{\\prime}\\) is the degraded version of the reference image \\(I\\), can be represented by the component set \\(\\{(L,R),(L',R')\\}\\), in line with Eq. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.655, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} I (x, y) = L (x, y) \\cdot R (x, y) \\\\ I ^ {\\prime} (x) = I ^ {\\prime} (x) - R ^ {\\prime} (x) \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.674, + 0.386, + 0.691 + ], + "angle": 0, + "content": "\\[\nI ^ {\\prime} (x, y) = L ^ {\\prime} (x, y) \\cdot R ^ {\\prime} (x, y)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Therefore, our goal is to determine the pair \\((\\overline{L},\\overline{R})\\) in which \\(\\overline{L}\\) restores the details lost due to inadequate lighting, while \\(\\overline{R}\\) aligns the input chromaticity with the white-aligned ambient lit setup." + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.766, + 0.483, + 0.791 + ], + "angle": 0, + "content": "\\[\nL (x, y) = L ^ {\\prime} (x, y) + \\bar {L} (x, y) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.788, + 0.387, + 0.804 + ], + "angle": 0, + "content": "\\[\nR (x, y) = R ^ {\\prime} (x, y) + \\bar {R} (x, y)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fig. 4 provides a graphical representation of the proposed \\(\\mathrm{RLN}^2\\), in which the \\((R', L')\\) components of the input \\(I'\\) are subjected to joint independent compensation, while the alternative HSV image representation information is used for feature refinement. Aligned with the properties of the HSV image representation, which codes color" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.182, + 0.905, + 0.257 + ], + "angle": 0, + "content": "in a cylindrical coordinate system, in which the Value (V) channel is similar to explicit illumination intensity, and the Hue-Saturation (HS) represents color tones, the V channel is used as guidance on the L-branch of our model, while the HS pair drives feature refinement on the R-branch of RLN\\(^2\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.348 + ], + "angle": 0, + "content": "At the encoder level, the information goes through successive decomposition, in which Haar-DWT is used for spectral dissemination, alongside RGB feature downsampling. The dual-domain information will be altered in the feature refinement stage, at the \\( R \\) block level, through the Cross-Domain Feature Fusion Attention module (CDFFA)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Here, the low-frequency information \\(X_{lf}\\), and the RGB extracted features \\(X_{L'}\\), projected to an n-dimensional space, are then subjected to relevance enhancement, through a set of parameters based on dot product vector similarity (see Fig. 5) to features extracted from the guidance component. The V channel controls the estimation of the \\(\\overline{L}\\) residual, while H and S drive the reflectance estimation \\(\\overline{R}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.48, + 0.758, + 0.497 + ], + "angle": 0, + "content": "\\[\nX _ {L ^ {\\prime}} = p r o j _ {n} \\left(X _ {L ^ {\\prime}}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.499, + 0.905, + 0.524 + ], + "angle": 0, + "content": "\\[\nX _ {l f} = \\operatorname {p r o j} _ {n} \\left(X _ {l f}\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.518, + 0.793, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = s i m (X _ {L ^ {\\prime}, i}, X _ {V, i})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.537, + 0.792, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\beta_ {i} = s i m (X _ {l f, i}, X _ {V, i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.626 + ], + "angle": 0, + "content": "The set of similarity coefficients is used to define a relevance probability distribution, then used to enhance the relevant information by filtering outlier features. The enhanced information is further refined before output." + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.638, + 0.715, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\operatorname {s o f t m a x} (\\alpha)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.657, + 0.905, + 0.673 + ], + "angle": 0, + "content": "\\[\n\\beta = \\operatorname {s o f t m a x} (\\beta) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.675, + 0.85, + 0.693 + ], + "angle": 0, + "content": "\\[\nX _ {o u t} = \\left\\{\\alpha_ {i} \\cdot X _ {L ^ {\\prime}, i} \\right\\} _ {i = 0} ^ {n} + \\left\\{\\beta_ {i} \\cdot X _ {l f, i} \\right\\} _ {i = 0} ^ {n}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.825 + ], + "angle": 0, + "content": "At the decoder stage, the refined information is concatenated to the high-frequency spectral information from the corresponding encoder block. Moreover, an ImageNet [40] pretrained ConvNeXt-based feature extractor module [49], provides wide contextual information in feature fusion. A cross-attention module handles information exchange between the wide context information and the dual-domain features refined in the previous stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For a fair comparison, at this level, \\(\\mathrm{RLN}^2\\) benefits from priors identical to [65], even if the implemented strategies are different. After linear projections, irrelevant features are suppressed through Channel Attention [74], being further refined for Reflectance/Luminance compensation [82]." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9222" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.207, + 0.087, + 0.367, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 5. A detailed Reflectance/Luminance Refinement Block (R.Block), including the CDFFA attention module." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.308, + 0.483, + 0.394 + ], + "angle": 0, + "content": "
ModelFusionGuidanceMACs (G.)PSNR↑SSIM↑LPIPS↓
\\( RLN^2 \\)NoneNone20.9519.8920.6890.271
\\( RLN^2 \\)concat.RGB23.6720.0180.7030.233
\\( RLN^2 \\)concat.L*a*b*23.6820.1490.7240.226
\\( RLN^2 \\)concat.HSV23.6820.2160.7310.220
\\( RLN^2 \\)(CDFFA)HSV22.7220.5230.7460.208
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.427 + ], + "angle": 0, + "content": "Table 2. Inner-stage guidance feature fusion ablation study conducted on the testing split of the CL3AN dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Ablative study: Tab. 2 compares the standard RLN\\(^2\\) configuration, as represented in Fig. 4, to three similar variants in which the control mechanism implemented in the middle stage is modified. Reducing the \\(R.\\) block to simple feature refinement by removing any control mechanism shows an important performance deficit, consistent in image restoration fidelity and perceptual quality. Alternatively, using the RGB domain images, passed through a dynamical filter and then split based on the filtered output and its alternative residual shows a slight improvement, at a relatively low induced computational cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Task-aligned image representations, with explicit coding of image luminance and chromaticity, prove rich in relevant information, adding only the image conversion cost to the general computation. Focusing on relevant feature enhancement through the CDFFA introduction further increases model robustness and restoration performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation: \\(\\mathrm{RLN}^2\\) was developed in PyTorch, and the documented experiments were performed on up to three NVIDIA L40s 48 GB VRAM GPUs running a CUDA 12.6 environment. The optimization is split into multiple stages, with the resolution of the training patches gradually increasing as the training procedure progresses. The optimizer used was Adam, with the learning rate \\(lr = 0.0002\\), and the loss optimized was the L1 distance. A cosine learning rate scheduler was used, with two oscillation periods. Optimization deploys gradient clipping, with a standard value of 0.01 under the L1 norm. Further details regarding the training procedure description can be found in the attached supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.6, + 0.106 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.19 + ], + "angle": 0, + "content": "Experimental Setup: Additionally to the introduced CL3AN dataset, we perform experiments on the AMBI-ENT6K image database [65], which consists of 6000 images representing scenes under the effect of up to three white-aligned LED directional lights (see Fig. 3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.281 + ], + "angle": 0, + "content": "As compared metrics, we report restoration fidelity through Peak Signal to Noise Ratio (PSNR), and Structure Similarity Index Measure (SSIM) [70]. To quantify the perceptual quality of the restored images, we deploy the AlexNet-based Learned Perceptual Image Patch Similarity score (LPIPS) [86]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.493 + ], + "angle": 0, + "content": "In terms of compared methods, we chose various types of solutions proposed in the image restoration field, which became seminal work after their introduction. For an extensive benchmark, we combine refined CNN solutions such as MPRNet [83] and HINet [37], with transformer-based architectures like SwinIR [47], NAFNet [10], Uformer [71], Restormer [84], and GRL [80]. Additionally, we consider solutions integrating state-space-models (e.g. MAMBAIR [21]), or solutions optimizing joint Spatial-Frequency entropy such as SFNet [12] and IFBlend [65]. Ultimately, we include Retinexformer [9], representing the class of solutions based on Retinex image decomposition. We train all the aforementioned methods using their official public implementations and settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.493, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Quantitative Results: In Tab. 3, we compare the aforementioned solutions with different variants of the proposed \\(\\mathrm{RLN}^2\\). The different variants represent our proposed models at different complexity levels, such that when comparing to similar existing solutions, advantages become easily identifiable. The computational complexity, as the Multiplications-Additions count (MAC), is reported for a patch of \\(128\\times 128\\) pixels, using ptflops for consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.794 + ], + "angle": 0, + "content": "The naming convention can be easily followed. Firstly, the difference between the RLN\\(^2\\)-S and the RLN\\(^2\\)-L is given by the complexity of the Wide Context Feature Extractor found in the feature fusion module, at the decoder stage, with RLN\\(^2\\)-S using a ConvNeXt-T configuration, while RLN\\(^2\\)-L benefits from a large ConvNeXt-XL model [49]. Secondly, we distinguish between RLN\\(^2\\) variants that have/have not access to the dual image frequency domain representation. Thus RLN\\(^2\\)-S and RLN\\(^2\\)-L drop the frequency domain feature at the encoder, refinement stage, and decoder stage, using cross-attention for fusion, while RLN\\(^2\\)-Sf and RLN\\(^2\\)-Lf use all the available information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.795, + 0.905, + 0.87 + ], + "angle": 0, + "content": "All the compared RLN\\(^2\\) variants are trained under the same conditions, including the optimizer, batch size, data augmentation strategies, optimized objective, and the number of samples seen during training. For AMBIENT6K [65], we adopt their publicly available benchmark." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Reading Tab. 3, it becomes obvious that \\(\\mathrm{RLN}^2\\) is a high-performance solution, improving over similar state-of-the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9223" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.91, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.343, + 0.862, + 0.357 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on scene #1 from test split of the CL3AN dataset. Best viewed in the electronic version." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.36, + 0.908, + 0.642 + ], + "angle": 0, + "content": "
Method nameTypeGeneral informationCL3AN (ours)AMBIENT6K [65]
Restoration TaskPriorMACs (G.)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
unprocessed----10.8370.4470.51813.4030.6520.250
NAFNet[10]Conv.Multi-task IRRGB15.9219.4760.7090.24920.5800.8080.142
MPRNet [83]Conv.Multi-task IRRGB37.2118.4530.6880.29120.9470.8200.129
SFNet [12]Transf.Multi-task IRRGB + Freq.31.2718.3820.6860.29120.5190.8120.141
SwinIR [47]Transf.Multi-task IRRGB37.8116.3860.6430.37220.5280.8170.131
Uformer [71]Transf.Multi-task IRRGB19.3317.5080.6550.31320.7760.8180.131
Restormer [84]Transf.Multi-task IRRGB35.3118.5600.6910.27821.1410.8170.132
HINet [37]Conv.Multi-task IRRGB42.6819.3880.7070.24820.8560.8210.129
IFBlend [65]ConvALNRGB + Freq.26.0120.3700.7200.22821.4430.8190.128
MAMBAIR [23]Transf. + SSMMulti-task IRRGB34.3218.9700.7040.254---
GRL [80]Transf.Multi-task IRRGB2.1618.0890.6720.308---
Retinexformer [9]Transf.LLIERGB4.8618.6490.6830.281---
\\( RLN^2-S \\) (ours)ConvALNRGB3.9519.9920.7180.23621.1810.8150.131
\\( RLN^2-Sf \\) (ours)ConvALNRGB + Freq.4.2420.1280.7300.22321.3330.8190.128
\\( RLN^2-L \\) (ours)ConvALNRGB22.4520.3830.7230.22221.5530.8210.123
\\( RLN^2-Lf \\) (ours)ConvALNRGB + Freq.22.7220.5230.7460.20821.7120.8250.120
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.647, + 0.907, + 0.677 + ], + "angle": 0, + "content": "Table 3. Various RLN\\(^2\\) variants versus current SOTA image restoration models, on the proposed CL3AN dataset test split (1920 × 1440 px.), and AMBIENT6K benchmark [65] (1280 × 960 px. default size). The best and the second best performing models are emphasized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.76 + ], + "angle": 0, + "content": "art solutions at all complexity levels, on both CL3AN and AMBIENT6K [65]. While \\(\\mathrm{RLN}^2\\)-S largely outperforms RetinexFormer [9], at a lower computational cost, \\(\\mathrm{RLN}^2\\)-Sf improves over the perceptual performance of IFBlend [65], with a fraction of its required floating point operations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Against IFBlend [65], a strong recently proposed solution in the ALN field, \\(\\mathrm{RLN}^2\\)-L shows an advantage, especially in joint RGB-Frequency domains processing, with an \\(\\approx 15\\%\\) complexity reduction, while outperforming it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Another important observation is regarding the architectural core of the compared methods. Under both white-aligned and RGB direct lighting, ALN tends to favor solutions showing an extensive receptive field [37, 83], rather than transformers optimized for local dependencies [47, 71," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.907, + 0.806 + ], + "angle": 0, + "content": "84], even when balancing to wider context blocks [80]. However, MPRNet [83] and HINet [37] are larger models, with an additional cost of approx. \\(63\\% - 87\\%\\), compared to the highest complexity RLN²-Lf variant. While state-state models prove computationally ineffective for the level of performance achieved, frequency-domain information access shows a consistent advantage, especially in terms of observed perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results: In Fig. 6, we provide for qualitative evaluation a set of samples representing a cluttered scene under the effect of non-homogeneous RGB direct lighting. Note the complex geometry and the wide RGB spectrum represented at the ambient lit image level. \\(\\mathrm{RLN}^2\\)-Lf shows improved behavior in terms of color normalization (row" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9224" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.09, + 0.196, + 0.102 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.09, + 0.393, + 0.104 + ], + "angle": 0, + "content": "Restormer [84]" + }, + { + "type": "image_caption", + "bbox": [ + 0.463, + 0.089, + 0.551, + 0.104 + ], + "angle": 0, + "content": "IFBlend [65]" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.089, + 0.728, + 0.104 + ], + "angle": 0, + "content": "\\(\\mathrm{RLN}^2\\) -Lf (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.103, + 0.26, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.103, + 0.424, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.103, + 0.588, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.103, + 0.754, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.103, + 0.92, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.201, + 0.259, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.201, + 0.424, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.201, + 0.589, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.201, + 0.754, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.201, + 0.92, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.905, + 0.325 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison challenging samples from the AMBIENT6K testing set [65]. Our \\(\\mathrm{RLN}^2\\) produces improved quality ambient normalized images, with sharper images, and a lower amount of visible artifacts." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.331, + 0.288, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.434, + 0.135, + 0.441 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.331, + 0.481, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.434, + 0.331, + 0.441 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.435, + 0.434, + 0.463, + 0.441 + ], + "angle": 0, + "content": "RLN²-Lf" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.484 + ], + "angle": 0, + "content": "Figure 8. IFBlend[65] and \\(\\mathrm{RLN}^2\\)-Lf on real-world images from LSMI [38] (left), or real product photography samples (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.521 + ], + "angle": 0, + "content": "1), superior restorations in highlights-rich areas, and better high-frequency shadow areas compensation (row 2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.522, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Fig. 7 provides for visual comparison equivalent renderings against the top performing state-of-the-art methods the AMBIENT6K [65] benchmark. Consistently, the restored ambient normalized images are characterized by more natural colors, an increased amount of restored details, and geometrically correct renderings even in areas severely affected by underexposure (second row). Overcoming the challenges specific to non-homogeneous direct color lighting, RLN\\(^2\\) shows consistent performance, confirming the advantages quantified through evaluations provided in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To test \\(\\mathrm{RLN}^2\\) on out-of-domain data, in Fig. 8 we compare IFBlend [65] and our proposed \\(\\mathrm{RLN}^2\\)-Lf on inputs from LSMI [38], or complex lighting scenarios from product photography. Our \\(\\mathrm{RLN}^2\\)-Lf shows increased robustness, with plausible rendered colors under uniform white ambient lighting conditions. Moreover, in Fig. 9 we deploy it on AI-generated images representing various contents in challenging lighting environments. Applying our ALN solution as a pre-processing step in Neural Image Editing proves beneficial for IC-Light [19], with improved layer separation leading to less semantic discontinuities (rows 1, 2, and 4), and improves the rendered results quality in term of shadow casting (row 3) and foreground clarity (row 2)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.332, + 0.71, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.487, + 0.709, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.645, + 0.585, + 0.655 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.645, + 0.698, + 0.655 + ], + "angle": 0, + "content": "\\(\\mathbf{RLN}^2\\) -lfout" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.332, + 0.904, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.645, + 0.799, + 0.654 + ], + "angle": 0, + "content": "w.o. RLN²-Lf" + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.645, + 0.889, + 0.654 + ], + "angle": 0, + "content": "W. RLN²-Lf" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Figure 9. ALN as a preprocessing step before image relighting with foreground condition. Results rendered with ICLight [19] for the input image (col. 3) vs. the \\(\\mathrm{RLN}^2\\) output (col. 4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.634, + 0.723 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose extending the current ALN study to the challenges added by multicolor and multi-intensity RGB direct lighting. We propose CL3AN, a high-resolution image database showcasing a multitude of materials under complex RGB lighting scenarios while providing white-light ambient-lit reference images using a professional photography setup. We introduce RLN², a high-performance ALN algorithm setting a new standard for the ALN task, showing how alternative image representations can provide valuable control signals in feature refinement, or offer natural bypassing options for detail preservation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.456, + 0.901 + ], + "angle": 0, + "content": "deepai.org/machine-learning-model/text2img" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9225" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgments. We thank the reviewers and ACs for their sustained effort and helpful feedback. This research was partially supported by the Alexander von Humboldt Foundation." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.195, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[1] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1397-1406, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.252, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[2] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1981-1990, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.324, + 0.484, + 0.393 + ], + "angle": 0, + "content": "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[4] Alexandre Araujo, Jean Ponce, and Julien Mairal. Towards real-world focus stacking with deep learning. arXiv preprint arXiv:2311.17846, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[5] Monika Bansal, Munish Kumar, and Manish Kumar. 2d object recognition: a comparative analysis of sift, surf and orb feature descriptors. Multimedia Tools and Applications, 80 (12):18839-18857, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[6] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.542, + 0.484, + 0.597 + ], + "angle": 0, + "content": "[7] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[8] Ivaylo Boyadzhiev, Kavita Bala, Sylvain Paris, and Frédo Durand. User-guided white balance for mixed lighting conditions. ACM Trans. Graph., 31(6):200-1, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.644, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[10] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[11] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.195, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[14] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[16] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[17] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, and Marc Pollefeys. Building rome on a cloudless day. In Computer Vision - ECCV 2010, pages 368-381, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[18] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 203-212, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[19] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[22] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5917-5926, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[23] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[24] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF interna" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9226" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tional conference on computer vision, pages 2472-2481, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[27] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0-0, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[28] Yan Huang, Xinchang Lu, Yuhui Quan, Yong Xu, and Hui Ji. Image shadow removal via multi-scale deep retina decomposition. Pattern Recognition, page 111126, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[29] Zhuo Hui, Kalyan Sunkavalli, Sunil Hadap, and Aswin C. Sankaranarayanan. Illuminant spectra-based source separation using flash photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6209-6218, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[30] Zhuo Hui, Ayan Chakrabarti, Kalyan Sunkavalli, and Aswin C Sankaranarayanan. Learning to separate multiple illuminants in a single image. In Computer Vision and Pattern Recognition (CVPR 2019), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[31] Anne Humeau-Heurtier. Texture feature extraction methods: A survey. IEEE access, 7:8975–9000, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[32] Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 536-537, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[33] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[34] Jong Ju Jeon, Jun Young Park, and Il Kyu Eom. Low-light image enhancement using gamma correction prior in mixed color spaces. Pattern Recognition, 146:110001, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[35] Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion. Advances in Neural Information Processing Systems, 37:141129-141152, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[36] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. arXiv preprint arXiv:2211.08089, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[37] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[38] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[39] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[40] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[41] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[42] Wei-Sheng Lai, Orazio Gallo, Jinwei Gu, Deqing Sun, Ming-Hsuan Yang, and Jan Kautz. Video stitching for linear camera arrays. arXiv preprint arXiv:1907.13622, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[43] EDWIN H. LAND. The retina. American Scientist, 52(2): 247-264, 1964. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[44] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[45] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Mimt: Multi-illuminant color constancy via multitask local surface and light color learning. arXiv preprint arXiv:2211.08772, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[46] Xinyu Li, Yanyi Zhang, Jianbo Yuan, Hanlin Lu, and Yibo Zhu. Discrete cosin transformer: Image modeling from frequency domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5468-5478, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[47] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[48] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[49] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[50] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1680-1691, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[51] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Radoslaw Mantiuk, Rafal Mantiuk, Anna Tomaszewska, and Wolfgang Heidrich. Color correction for tone mapping. In Computer graphics forum, pages 193-202. Wiley Online Library, 2009. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9227" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[53] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523-3542, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[54] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[55]Punatawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22646-22657, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[56] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[57] Pramod Rao, Gereon Fox, Abhinitra Meka, Mallikarjun BR, Fangneng Zhan, Tim Weyrich, Bernd Bickel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, et al. Lite2relight: 3d-aware single image portrait relighting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[58] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[60] Richard Szeliski and Richard Szeliski. Image alignment and stitching. Computer Vision: Algorithms and Applications, pages 401-441, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[61] Hironobu Takano, Hiroki Kobayashi, and Kiyomi Nakamura. Rotation invariant iris recognition method adaptive to ambient lighting variation. IEICE transactions on information and systems, 90(6):955-962, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[62] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20762-20772, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[63] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[64] Florin-Alexandru Vasluianu, Tim Seizinger, and Radu Timofte. Wsrd: A novel benchmark for high resolution image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1826-1835, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Florin-Alexandru Vasluianu, Tim Seizinger, Zongwei Wu, Rakesh Ranjan, and Radu Timofte. Towards image ambient" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "lighting normalization. arXiv preprint arXiv:2403.18730, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[66] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[67] Qiyang Wan, Ruiping Wang, and Xilin Chen. Interpretable object recognition by semantic prototype analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 800-809, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[68] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[69] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[70] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[71] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[72] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[73] Ching-Chih Weng, Homer Chen, and Chiou-Shann Fuh. A novel automatic white balance method for digital still cameras. In 2005 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3801-3804. IEEE, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[74] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[75] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.706 + ], + "angle": 0, + "content": "[76] Xudong Xie and Kin-Man Lam. An efficient illumination normalization method for face recognition. Pattern Recognition Letters, 27(6):609-617, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[77] Xiaohua Xie, Wei-Shi Zheng, Jianhuang Lai, Pong C. Yuen, and Ching Y. Suen. Normalization of face illumination based on large-and small-scale features. IEEE Transactions on Image Processing, 20(7):1807-1821, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[78] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. International Journal of Computer Vision, 132(9):3610-3632, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[79] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9228" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[80] Xiaoyu Xiang Denis Demandolx Rakesh Ranjan Radu Timofte Yawei Li, Yuchen Fan and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[81] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[82] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[83] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[85] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[87] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[88] Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed images via robust retina decomposition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[89] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1219-1229, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.796 + ], + "angle": 0, + "content": "[90] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12786-12796, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9229" + } + ] +] \ No newline at end of file diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_origin.pdf b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32b5cf4a9f279e4022fd1326e93a577654db05c9 --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/63a161d7-3eb4-437d-8d2d-2312e022984e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91b0b403cb528d0f62d346828ef7f37f3b1592d3802963527f6a6f20984b5da7 +size 17922087 diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/full.md b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d9bd0a3ecda6aa9c00b8d80949b6b661980a4892 --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/full.md @@ -0,0 +1,392 @@ +# After the Party: Navigating the Mapping From Color to Ambient Lighting + +Florin-Alexandru Vasluianu Tim Seizinger Zongwei Wu* Radu Timofte Computer Vision Lab, CAIDAS & IFI, University of Wurzburg + +{firstname.lastname}@uni-wuerzburg.de + +# Abstract + +Illumination in practical scenarios is inherently complex, involving colored light sources, occlusions, and diverse material interactions that produce intricate reflectance and shading effects. However, existing methods often oversimplify this challenge by assuming a single light source or uniform, white-balanced lighting, leaving many of these complexities unaddressed. In this paper, we introduce CL3AN, the first large-scale, high-resolution dataset of its kind designed to facilitate the restoration of images captured under multiple Colored Light sources to their Ambient-Normalized counterparts. Through benchmarking, we find that leading approaches often produce artifacts—such as illumination inconsistencies, texture leakage, and color distortion—primarily due to their limited ability to precisely disentangle illumination from reflectance. Motivated by this insight, we achieve such a desired decomposition through a novel learning framework that leverages explicit chromaticity-luminance components guidance, drawing inspiration from the principles of the Retinex model. Extensive evaluations on existing benchmarks and our dataset demonstrate the effectiveness of our approach, showcasing enhanced robustness under non-homogeneous color lighting and material-specific reflectance variations, all while maintaining a highly competitive computational cost. The benchmark, codes, and models are available at www.github.com/fvasluianu97/RLN2. + +# 1. Introduction + +The goal of Ambient Lighting Normalization is recovering an ambient-lit image, under uniform lighting distribution, regardless of the lighting scenario characteristic to the depicted contents [65]. The ability to recover ambient-normalized images—free from spatially varying illumination artifacts—is critical for robust visual understanding [33, 51, 53]. In both practical imaging scenarios [27, 32] and the rapidly evolving domain of image generation and + +![](images/75f88847bb339ab0df3037fb024699d37acc278be393a5245022b8346c1a6fe7.jpg) +Figure 1. Effective deployment of Ambient Lighting Normalization (ALN) as a pre-processing step in Neural Image Editing [19]. We compare the effectiveness of our proposed RLN² against state-of-the-art models in fields like Shadow Removal (DES3 [36]), or ALN (IFBlend [65]), trained on settings including single (1) [68] or multiple (3) white-domain-tuned [65] or color light sources. + +synthesis [57, 81], illumination always involves complex interactions of multiple colored light sources, occlusions, and material properties, resulting in intricate shadows, glare, color shifts, and texture distortions. Lighting and material characteristics are the driving factors for explaining the manifestation of such artifacts, thus explaining the need for data-level representativity prior to proposing image restoration solutions aiming for improved degradation robustness or restoration performance. + +While a wide range of applications require illumination-invariant representations [7, 31, 59, 67], existing normalization methods struggle to disentangle these intertwined factors [19, 35, 62, 81], as illustrated in Fig. 1. Even if data contributions further away from the natural light in $\mathrm{STD + }$ [44, 68], or multiple white-light-tuned light sources from AMBIENT6K [65] can improve general effectiveness, model design can further drive robustness. + +This limitation stems primarily from oversimplified assumptions, such as single white light sources [64, 68] or homogeneous color fields [56, 68, 87], which fail to capture the richness of complex illumination. A key technical challenge lies in the lack of comprehensive datasets that + +adequately represent the complexity of real-world lighting, especially outside the realm of natural white-color-tuned lighting. Prior efforts in shadow removal [64, 68], color constancy [6, 15, 16], or color correction [1, 22, 34, 52, 73] address isolated aspects of illumination but fall short of modeling multi-light interactions. Recent attempts [38, 65] introduce variable-intensity lighting datasets but lack chromatic diversity and material-specific reflectance variations. Furthermore, wide access to ambient-lit images outside the scope of scene-rendering software remains a challenge. As a result, learning-based methods often produce artifacts. + +In this paper, we address these limitations through two core contributions. First, we introduce CL3AN, the first large-scale dataset capturing high-resolution scenes under multiple colored directional lights paired with ambient-lit references. Each scene triplet includes variations beyond white-balanced lighting, colored lighting, and a diffuse-lit ground truth, enabling systematic study of illumination effects like self-shadows, highlights, glare areas, color spill or saturation, and other material-dependent artifacts. By benchmarking leading image restoration methods [9, 23, 65, 80], we observe that these models frequently suffer from critical shortcomings, such as the conflation of chromatic variations caused by illumination, texture leakage, inconsistent partial normalization, etc. + +We identify the fundamental limitation of existing methods as their inability to systematically disentangle reflectance and illumination—a critical flaw exacerbated by the lack of explicit guidance for separating chromaticity and luminance under multi-colored lighting. Most current approaches [23, 80, 84] rely on single-stream, end-to-end feature modeling, which implicitly maps input images to ambient-normalized counterparts without incorporating physical awareness of light-material interactions. While some recent works attempt to address this by introducing pixel- and frequency-domain supervision [12, 65], more flexible towards reflectance- and chromaticity-information extraction, they remain limited in learning appropriate frequency filters — a challenging task without explicit guidance — resulting in suboptimal performance under the challenging multi-colored lighting conditions. + +In this paper, from a novel perspective, we propose $\mathbf{RLN}^2$ , a Retinex-inspired framework that introduces parallel, physics-grounded streams to independently model reflectance (material properties) and illumination (lighting geometry/color). We observe that these intrinsic specificities are naturally encoded in the cylindrical representation of the input image, where the Hue (H) and Saturation (S) maps align with the observed chromaticity under the effect of the incoming light, while the Value (V) map provides strong priors for materials-specific reflectance. By embedding these domain-specific clues — such as reflectance information entropy (V), and smoothness constraints (H, S) + +into feature refinement, our framework achieves precise decomposition, effectively eliminating undesirable artifacts while preserving fine-grained details. + +Our contributions can be summarized as follows: + +- The introduction of the CL3AN, a first-of-its-kind dataset enabling the study of ambient lighting normalization under multiple RGB lights; +- The extensive study of a wide range of models, seminal works in various image restoration tasks, when trained and tested on the CL3AN dataset; +- We propose $\mathbf{RLN}^2$ , a robust algorithm built on the principle of simultaneous image reflection and lightness components compensation to recover an ambient lit image, characterized by uniform lighting distribution. $\mathbf{RLN}^2$ is proposed as a strong baseline for the ambient lighting normalization task, achieving state-of-the-art results on both the public AMBIENT6K dataset and the proposed CL3AN benchmark. + +# 2. Related Work + +This work expands Ambient Lighting Normalization to multicolor direct lighting, addressing previously unexplored effects appearing in cluttered scenes, and analyzing shadow interactions and highlights appearing under multicolor direct lighting. + +Lighting and Color: Color analysis is central to image processing and understanding, with various downstream tasks such as facial recognition [61, 76, 77], or tracking and matching applications [5, 14, 41, 59, 63] relying on the color similarity principle, thus benefiting from increased appearance robustness, especially under challenging lighting. + +Image Quality Assessment (IQA) methods [3, 79] show a strong correlation between details-abundance and image quality, motivating the focus on professional scene lighting rigs in all expert image photography applications. Given the scarcity of such equipment, alternative strategies [1, 2, 6, 45] are developed around image sensor sensor properties and cross-device constancy [8, 20]. + +Naturally, registered image stacks either for tasks such as multiple view geometry estimation [17, 90] or different image stitching applications require consistent color and content appearance [4, 11, 42, 60]. In such applications, the effort is focused on consistent lighting in the image acquisition, rather than post-processing correction. In-the-wild applications focusing on widely available amateur images [17, 24, 60] document extensive data curation strategies prior to model estimations. + +Later applications, given the newly-found popularity of neural 3D modeling, approach scene relighting as part of their rendering pipeline [58, 62]. Alternatively, approaches based on complex generative image models [35, 55] successfully tackle lighting manipulation, but at a high computational cost. Unfortunately, the limited availability of + +![](images/5c5d6867ece602091ba1d438df7b6ea0a5a3ef712b75b698daa42d55febdafec.jpg) +A + +![](images/3febe8bc9579f52c5f669791b0ef5d09df51674914218a7419c43e262dee3310.jpg) +B +ISTD/WSRD +AMBIENT6K +CL3AN (ours) + +![](images/1b8ff6d2a50f0fc4d47c05960bfefb803e8858eb4278b104df9aa3795c8f9d2c.jpg) +C +Figure 2. A comparison between the lighting scenarios represented in datasets targeting different applications of Ambient Lighting Normalization (ISTD [68], WSRD[64], or AMBIENT6K [65]). In (A), ambient lighting is used to acquire a uniformly-lit reference image. In (B), direct lighting is used for shadow casting, while constraining consistency to the reference image in shadow-free segments. In (C), more directional lights are added to the direct lighting system. In our proposed setup (D), the direct lighting setup is based on RGB lights, dropping the color consistency constraint, and showcasing complex material-light interactions under various conditions. + +![](images/7af12f05f4f136016eb35f410ddb40aa2bcb29667905c2e376d5b04ee671f0f1.jpg) +D + +real-world ambient lit images, primarily due to the complexity of an ambient lighting setup, explains the difficulty in addressing real-world lighting normalization, even with the obvious availability of large-complexity models. + +Reliance on natural light for ambient-lit images proved a viable option, but hindered the variety of possible observed degradations, especially in terms of chromaticity shifts. Unfortunately, the large number of factors involved in light occlusion interaction greatly constrains a representative simulation in general settings. While ISTD/ISTD+ [44, 68], SRD [56] or USR [26] show exclusively flat surfaces under the influence of natural light, the influence of object self-shadows was just recently represented in WSRD [64], only possible under a laboratory setup. Recently, multiple source lighting gained in data-level representation. For example, in LSMI [38], the authors target White-Balance correction robustness under natural and artificial lighting. Alternatively, effects characteristic to multiple illuminants were captured in works such as [30, 54], or in the context of flash photography [29]. However, a uniformly-lit ambient normalized image can not be acquired in these settings. + +Alternatively, the introduction of AMBIENT6K [65] coupled multiple light sources of varying intensity and orientation for complex interactions, while also providing an ambient-lit image based on an alternative lighting setup. Unfortunately, the strong connection between the operating parameters set for the direct and ambient lighting systems limits the study to exposure correction for white-aligned scene lighting (see Fig. 2 (A), (B), and (C)). + +Image Restoration (IR): Joint color and exposure correction, while restoring details lost to occlusion or saturation, defines Ambient Lighting Normalization as an Image Restoration task. Various Convolution Neural Networks (CNNs) [10, 37, 83, 85] became strong baselines for the family of IR tasks. The later introduction of the attention + +mechanism [66] in image-based applications [13] enabled consistent performance advantages in the wide IR field. Focusing on different feature fusion techniques through specifically tailored attention mechanisms, based on the IR task particularities, defines the success of multiple models [10, 47, 48, 71, 80, 84] easily becoming benchmark solutions for various IR problems. Separate refinement for image reflectance and lightness under the Retinex model [43] proved particularly effective in low-light image enhancement (LLIE) [9, 72, 78, 88], or image shadow removal [28]. + +Global range dependencies benefits drive the development of State Space Models (SSM) [21], with promising results in general IR [23]. Alternatively, global knowledge can be leveraged through the integration of alternative frequency image representations in IR solutions. Solutions such as FFTFormer [39], DCFormer [46], SFNet [12], DW-GAN [18], and IFBlend [65] successfully extract priors based on Fourier Transform, Discrete Cosine Transform, or other Wavelet Transform variants. + +Finally, diffusion-based iterative recurrent solutions [25] started to mature as benchmark solutions in Image Restoration [50, 69, 75, 89], especially in terms of perceived image quality. However, their characteristic increased computational effort hinders their deployment. + +# 3. Method + +The core of our work is extending the study of Ambient Lighting Normalization to direct color lighting. + +Fig. 3 compares the properties of a scene depicted under multi-source color direct lighting, and an equivalent image under white-aligned direct lighting. The goal to recover the ambient-lit reference image. Provided statistics, such as the RGB image histogram, and the Hue/Saturation footprint showcase the complexity of the ALN transformation, espe + +![](images/524aa8bdc7ef33427e115e065022f7192616a5b5f75a203606c166af218e5053.jpg) +Figure 3. A graphical representation of the Ambient Lighting Normalization transformation in the context of the white-aligned lighting, compared to the challenges posed by the introduction of a multicolor direct lighting system. The RGB image histogram, coupled with the Hue-Saturation rough projection is provided as a reference. Note the complexity of the change in image statistics. + +cially increasing with the introduction of variable intensity multicolor direct light sources. Naturally, the most notable difference comes from the color hue deviation. Materials adopt the color of incoming light, with the reflected light wavelength depending on the reflectance defining the represented materials, and the incoming light properties. + +Furthermore, the complex geometry of the scene drives light occlusions, with numerous self-shadows at the image level. In this case, given the non-uniform color distribution, the color of the observed shadows depends on the local lighting availability and their light source-defining properties. A successful restoration of the ambient-lit image can not be performed without a full understanding of both scene geometry (for local exposure compensation) and lighting system configuration (for chromaticity uniformization, given the arithmetic of the locally available light radiation given the material properties). + +# 3.1. CL3AN + +CL3AN represents an image collection representing Triplets consisting of images acquired under both Color and White Lighting, and Ambient Normalized images, under uniformly distributed diffuse light. + +As introduced in Fig. 2, CL3AN is the only dataset providing equivalent representations of the same scene under white-aligned direct lighting (setups (B), (C)), varying intensity multicolor lighting (setup (D)), while also providing an ambient-lit reference image (setup (A)). This makes CL3AN a first-of-its-kind dataset. + +The capturing system relies on a professional photography studio setup, in which a dark room is used to ensure that all image-capturing parameters become trackable and controllable. Two separate scene lighting systems are concurrently developed, optimizing for uniform diffuse light distribution, respectively direct lighting, being triggered only + +for the acquisition of the corresponding samples. Note that directional lights are fully controllable, so they can be tuned to the white natural light domain, producing images similar to those available in datasets such as AMBIENT6K [65]. + +The images are captured with a tripod-mounted and remotely operated Canon R6 MKII camera, at constant focal range. This leads to a perfect pixel alignment between the images characteristic to each triplet. For consistency and traceability, the camera-related parameters are kept fixed during the image-capturing process. All hardware parameters relevant to data acquisition are documented in Tab. 1. + +For the direct lighting system, three programmable mobile RGB lights are used to capture scene variants under variable light color, intensity, position, and orientation. The direct consequence of this set of parameters is a nonhomogeneous distribution of the incoming light color across the scene, implying that global color correction can not be used for restoration in this scenario (see Fig. 2 and Fig. 3). + +The diffuse ambient lighting system is composed of five white softbox lights, covering all sides of the cubic acquisition environment, except the base, which is the surface supporting the captured contents. Thus, the possibility for self-shadows in the ambient-lit images is almost fully mitigated, being furthermore enforced through geometry-based object filtering. Calibration over a 24-sample color checker under full white light intensity enables the alignment of both lighting setups, by determining the set of operating parameters producing a similar colors under optimal exposure. + +CL3AN is a scene-based dataset, covering 105 cluttered scenes representing an extensive set of materials, both conductors and dielectrics, with varying transparency, color, and surface roughness. The data is available as both RAW and RGB images, at $24\mathrm{MP}$ , the camera sensor resolution. Data splitting is performed based on the represented contents, with 10 scenes used exclusively for testing, and 10 scenes for validation. Both testing and validation splits represent objects unseen during training, but under lighting conditions consistent with the training split. Additional metadata listing the objects within a scene is provided. The training split consists of a number of 3667 samples, with 437 images used for validation and 431 for testing. + +# 3.2. RLN + +Fig. 2 and Fig. 3 show the challenging environment posed by the introduction of multicolor direct lighting in the ALN field. Dramatic changes can be observed when mapping from the input image to the ambient-lit reference, especially in terms of observed chromaticity. Moreover, light occlusion by the many rough surfaces corresponding to the represented objects drives a wide variety of intricate effects, with a strong information loss. Learning a robust ALN mapping requires a robust design, able to disentangle significant information compensation in confined image segments from + +
Camera parametersDiffuse LightingDirect Lighting
Exposure time (s)ApertureISOWhite balanceFocal dist. (mm)# Active lightsWhite light temp. (K)White light temp. rangeIntensity# Active lightsWhite light temp. (K)Intensity rangeColor Hue range
1/60f111006400K5056400±5%100%≥ 1640030-100%0°-360°
+ +Table 1. Dataset capturing parameters regarding the setup of the camera, the diffuse lighting system, and the direct lighting hardware. + +![](images/cfaea4fd6f24a885337942fc952b347d0581a2612a8f2b74d9887db7c0d1f8cf.jpg) +Figure 4. A graphical representation of the $\mathrm{RLN}^2$ architecture, including detailed representations for an Encoder (E) - Decoder (D) blocks pair on the Luminance L-branch. The Reflectance R-branch design is similar, being omitted for clarity. + +the large magnitude chromaticity shifts. + +Given these observations, problem sub-division is the core of our proposed design $\mathrm{RLN}^2$ , aiming for joint independent compensation in both Image Reflectance and Luminance components. Under the Retinex theory [43], a given image $I$ can be decomposed as a product of two multiplicative terms, representing image illumination or lightness $L$ , and the material specific image reflectance $R$ . + +Thus, an image pair $(I,I^{\prime})$ in which $I^{\prime}$ is the degraded version of the reference image $I$ , can be represented by the component set $\{(L,R),(L',R')\}$ , in line with Eq. (1). + +$$ +\begin{array}{l} I (x, y) = L (x, y) \cdot R (x, y) \\ I ^ {\prime} (x) = I ^ {\prime} (x) - R ^ {\prime} (x) \end{array} \tag {1} +$$ + +$$ +I ^ {\prime} (x, y) = L ^ {\prime} (x, y) \cdot R ^ {\prime} (x, y) +$$ + +Therefore, our goal is to determine the pair $(\overline{L},\overline{R})$ in which $\overline{L}$ restores the details lost due to inadequate lighting, while $\overline{R}$ aligns the input chromaticity with the white-aligned ambient lit setup. + +$$ +L (x, y) = L ^ {\prime} (x, y) + \bar {L} (x, y) \tag {2} +$$ + +$$ +R (x, y) = R ^ {\prime} (x, y) + \bar {R} (x, y) +$$ + +Fig. 4 provides a graphical representation of the proposed $\mathrm{RLN}^2$ , in which the $(R', L')$ components of the input $I'$ are subjected to joint independent compensation, while the alternative HSV image representation information is used for feature refinement. Aligned with the properties of the HSV image representation, which codes color + +in a cylindrical coordinate system, in which the Value (V) channel is similar to explicit illumination intensity, and the Hue-Saturation (HS) represents color tones, the V channel is used as guidance on the L-branch of our model, while the HS pair drives feature refinement on the R-branch of RLN $^2$ . + +At the encoder level, the information goes through successive decomposition, in which Haar-DWT is used for spectral dissemination, alongside RGB feature downsampling. The dual-domain information will be altered in the feature refinement stage, at the $R$ block level, through the Cross-Domain Feature Fusion Attention module (CDFFA). + +Here, the low-frequency information $X_{lf}$ , and the RGB extracted features $X_{L'}$ , projected to an n-dimensional space, are then subjected to relevance enhancement, through a set of parameters based on dot product vector similarity (see Fig. 5) to features extracted from the guidance component. The V channel controls the estimation of the $\overline{L}$ residual, while H and S drive the reflectance estimation $\overline{R}$ . + +$$ +X _ {L ^ {\prime}} = p r o j _ {n} \left(X _ {L ^ {\prime}}\right) +$$ + +$$ +X _ {l f} = \operatorname {p r o j} _ {n} \left(X _ {l f}\right) \tag {3} +$$ + +$$ +\alpha_ {i} = s i m (X _ {L ^ {\prime}, i}, X _ {V, i}) +$$ + +$$ +\beta_ {i} = s i m (X _ {l f, i}, X _ {V, i}) +$$ + +The set of similarity coefficients is used to define a relevance probability distribution, then used to enhance the relevant information by filtering outlier features. The enhanced information is further refined before output. + +$$ +\alpha = \operatorname {s o f t m a x} (\alpha) +$$ + +$$ +\beta = \operatorname {s o f t m a x} (\beta) \tag {4} +$$ + +$$ +X _ {o u t} = \left\{\alpha_ {i} \cdot X _ {L ^ {\prime}, i} \right\} _ {i = 0} ^ {n} + \left\{\beta_ {i} \cdot X _ {l f, i} \right\} _ {i = 0} ^ {n} +$$ + +At the decoder stage, the refined information is concatenated to the high-frequency spectral information from the corresponding encoder block. Moreover, an ImageNet [40] pretrained ConvNeXt-based feature extractor module [49], provides wide contextual information in feature fusion. A cross-attention module handles information exchange between the wide context information and the dual-domain features refined in the previous stage. + +For a fair comparison, at this level, $\mathrm{RLN}^2$ benefits from priors identical to [65], even if the implemented strategies are different. After linear projections, irrelevant features are suppressed through Channel Attention [74], being further refined for Reflectance/Luminance compensation [82]. + +![](images/5911921078cd2997832e0e46fc94616aae90386352e67c0c7b58314d0470d5c6.jpg) +Figure 5. A detailed Reflectance/Luminance Refinement Block (R.Block), including the CDFFA attention module. + +
ModelFusionGuidanceMACs (G.)PSNR↑SSIM↑LPIPS↓
\( RLN^2 \)NoneNone20.9519.8920.6890.271
\( RLN^2 \)concat.RGB23.6720.0180.7030.233
\( RLN^2 \)concat.L*a*b*23.6820.1490.7240.226
\( RLN^2 \)concat.HSV23.6820.2160.7310.220
\( RLN^2 \)(CDFFA)HSV22.7220.5230.7460.208
+ +Table 2. Inner-stage guidance feature fusion ablation study conducted on the testing split of the CL3AN dataset. + +Ablative study: Tab. 2 compares the standard RLN $^2$ configuration, as represented in Fig. 4, to three similar variants in which the control mechanism implemented in the middle stage is modified. Reducing the $R.$ block to simple feature refinement by removing any control mechanism shows an important performance deficit, consistent in image restoration fidelity and perceptual quality. Alternatively, using the RGB domain images, passed through a dynamical filter and then split based on the filtered output and its alternative residual shows a slight improvement, at a relatively low induced computational cost. + +Task-aligned image representations, with explicit coding of image luminance and chromaticity, prove rich in relevant information, adding only the image conversion cost to the general computation. Focusing on relevant feature enhancement through the CDFFA introduction further increases model robustness and restoration performance. + +Implementation: $\mathrm{RLN}^2$ was developed in PyTorch, and the documented experiments were performed on up to three NVIDIA L40s 48 GB VRAM GPUs running a CUDA 12.6 environment. The optimization is split into multiple stages, with the resolution of the training patches gradually increasing as the training procedure progresses. The optimizer used was Adam, with the learning rate $lr = 0.0002$ , and the loss optimized was the L1 distance. A cosine learning rate scheduler was used, with two oscillation periods. Optimization deploys gradient clipping, with a standard value of 0.01 under the L1 norm. Further details regarding the training procedure description can be found in the attached supplementary material. + +# 4. Results + +Experimental Setup: Additionally to the introduced CL3AN dataset, we perform experiments on the AMBI-ENT6K image database [65], which consists of 6000 images representing scenes under the effect of up to three white-aligned LED directional lights (see Fig. 3). + +As compared metrics, we report restoration fidelity through Peak Signal to Noise Ratio (PSNR), and Structure Similarity Index Measure (SSIM) [70]. To quantify the perceptual quality of the restored images, we deploy the AlexNet-based Learned Perceptual Image Patch Similarity score (LPIPS) [86]. + +In terms of compared methods, we chose various types of solutions proposed in the image restoration field, which became seminal work after their introduction. For an extensive benchmark, we combine refined CNN solutions such as MPRNet [83] and HINet [37], with transformer-based architectures like SwinIR [47], NAFNet [10], Uformer [71], Restormer [84], and GRL [80]. Additionally, we consider solutions integrating state-space-models (e.g. MAMBAIR [21]), or solutions optimizing joint Spatial-Frequency entropy such as SFNet [12] and IFBlend [65]. Ultimately, we include Retinexformer [9], representing the class of solutions based on Retinex image decomposition. We train all the aforementioned methods using their official public implementations and settings. + +Quantitative Results: In Tab. 3, we compare the aforementioned solutions with different variants of the proposed $\mathrm{RLN}^2$ . The different variants represent our proposed models at different complexity levels, such that when comparing to similar existing solutions, advantages become easily identifiable. The computational complexity, as the Multiplications-Additions count (MAC), is reported for a patch of $128\times 128$ pixels, using ptflops for consistency. + +The naming convention can be easily followed. Firstly, the difference between the RLN $^2$ -S and the RLN $^2$ -L is given by the complexity of the Wide Context Feature Extractor found in the feature fusion module, at the decoder stage, with RLN $^2$ -S using a ConvNeXt-T configuration, while RLN $^2$ -L benefits from a large ConvNeXt-XL model [49]. Secondly, we distinguish between RLN $^2$ variants that have/have not access to the dual image frequency domain representation. Thus RLN $^2$ -S and RLN $^2$ -L drop the frequency domain feature at the encoder, refinement stage, and decoder stage, using cross-attention for fusion, while RLN $^2$ -Sf and RLN $^2$ -Lf use all the available information. + +All the compared RLN $^2$ variants are trained under the same conditions, including the optimizer, batch size, data augmentation strategies, optimized objective, and the number of samples seen during training. For AMBIENT6K [65], we adopt their publicly available benchmark. + +Reading Tab. 3, it becomes obvious that $\mathrm{RLN}^2$ is a high-performance solution, improving over similar state-of-the + +![](images/84370e84a46d04b026a9061eed8b7db0076a71406aee594e75b204c73bb9ae3b.jpg) +Figure 6. Qualitative comparison on scene #1 from test split of the CL3AN dataset. Best viewed in the electronic version. + +
Method nameTypeGeneral informationCL3AN (ours)AMBIENT6K [65]
Restoration TaskPriorMACs (G.)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
unprocessed----10.8370.4470.51813.4030.6520.250
NAFNet[10]Conv.Multi-task IRRGB15.9219.4760.7090.24920.5800.8080.142
MPRNet [83]Conv.Multi-task IRRGB37.2118.4530.6880.29120.9470.8200.129
SFNet [12]Transf.Multi-task IRRGB + Freq.31.2718.3820.6860.29120.5190.8120.141
SwinIR [47]Transf.Multi-task IRRGB37.8116.3860.6430.37220.5280.8170.131
Uformer [71]Transf.Multi-task IRRGB19.3317.5080.6550.31320.7760.8180.131
Restormer [84]Transf.Multi-task IRRGB35.3118.5600.6910.27821.1410.8170.132
HINet [37]Conv.Multi-task IRRGB42.6819.3880.7070.24820.8560.8210.129
IFBlend [65]ConvALNRGB + Freq.26.0120.3700.7200.22821.4430.8190.128
MAMBAIR [23]Transf. + SSMMulti-task IRRGB34.3218.9700.7040.254---
GRL [80]Transf.Multi-task IRRGB2.1618.0890.6720.308---
Retinexformer [9]Transf.LLIERGB4.8618.6490.6830.281---
\( RLN^2-S \) (ours)ConvALNRGB3.9519.9920.7180.23621.1810.8150.131
\( RLN^2-Sf \) (ours)ConvALNRGB + Freq.4.2420.1280.7300.22321.3330.8190.128
\( RLN^2-L \) (ours)ConvALNRGB22.4520.3830.7230.22221.5530.8210.123
\( RLN^2-Lf \) (ours)ConvALNRGB + Freq.22.7220.5230.7460.20821.7120.8250.120
+ +Table 3. Various RLN $^2$ variants versus current SOTA image restoration models, on the proposed CL3AN dataset test split (1920 × 1440 px.), and AMBIENT6K benchmark [65] (1280 × 960 px. default size). The best and the second best performing models are emphasized. + +art solutions at all complexity levels, on both CL3AN and AMBIENT6K [65]. While $\mathrm{RLN}^2$ -S largely outperforms RetinexFormer [9], at a lower computational cost, $\mathrm{RLN}^2$ -Sf improves over the perceptual performance of IFBlend [65], with a fraction of its required floating point operations. + +Against IFBlend [65], a strong recently proposed solution in the ALN field, $\mathrm{RLN}^2$ -L shows an advantage, especially in joint RGB-Frequency domains processing, with an $\approx 15\%$ complexity reduction, while outperforming it. + +Another important observation is regarding the architectural core of the compared methods. Under both white-aligned and RGB direct lighting, ALN tends to favor solutions showing an extensive receptive field [37, 83], rather than transformers optimized for local dependencies [47, 71, + +84], even when balancing to wider context blocks [80]. However, MPRNet [83] and HINet [37] are larger models, with an additional cost of approx. $63\% - 87\%$ , compared to the highest complexity RLN²-Lf variant. While state-state models prove computationally ineffective for the level of performance achieved, frequency-domain information access shows a consistent advantage, especially in terms of observed perceptual quality. + +Qualitative Results: In Fig. 6, we provide for qualitative evaluation a set of samples representing a cluttered scene under the effect of non-homogeneous RGB direct lighting. Note the complex geometry and the wide RGB spectrum represented at the ambient lit image level. $\mathrm{RLN}^2$ -Lf shows improved behavior in terms of color normalization (row + +![](images/2d78f011811514e37254f59a4686e5871d6252b64c8b02dde83267fae95fcdb3.jpg) +Input + +![](images/bc79a030fe1552ecff7329773b69dfe881f9d0139781ef004b117c0d5d8ac069.jpg) +Restormer [84] + +![](images/911dcd32925decfa6917cb4c8e710558af1a89e1dc0f849723c425ed092c052d.jpg) +IFBlend [65] + +![](images/b25e940749d3e4d41461e8685f3bf17247e00e4a76f22f180241f0c21170d135.jpg) +$\mathrm{RLN}^2$ -Lf (Ours) + +![](images/ac627ee46216cadc8170a9b557820aa4bdfc9fd2e9efe3d48bbe85fad075985d.jpg) + +![](images/5f7cf4fe9daeea95e6354c8bdd4de511c799801bb7210fa5eb2d38d87e59d36d.jpg) +Figure 7. Qualitative comparison challenging samples from the AMBIENT6K testing set [65]. Our $\mathrm{RLN}^2$ produces improved quality ambient normalized images, with sharper images, and a lower amount of visible artifacts. + +![](images/c564554c3feb8b4576da27f42da1e4ebae64a8ca4bb33a1a0f2b5e4480fa72d0.jpg) + +![](images/7e8b0b6ae3af887a83a546cf328935ef4a65e5088436275c2cb7bb02037e5eb0.jpg) + +![](images/5b530f3ba765428d3f9f6504f6c6cdb178e7796815103c86f340c8a72afdb05f.jpg) + +![](images/a2a4086cf0e714c57920886416500b4edf81d4e0bfc135cb8965d4cc65b3d0f0.jpg) + +![](images/23f5cb000b6f1a367762fdb30f508ecffd5b966eac8bbe21db3bb7ec48320ff5.jpg) +Input + +![](images/8deaa16b60f6fb1936ce7c745081118a29eb0e617197d9366dafae8870ab643f.jpg) +Input +RLN²-Lf +Figure 8. IFBlend[65] and $\mathrm{RLN}^2$ -Lf on real-world images from LSMI [38] (left), or real product photography samples (right). + +1), superior restorations in highlights-rich areas, and better high-frequency shadow areas compensation (row 2). + +Fig. 7 provides for visual comparison equivalent renderings against the top performing state-of-the-art methods the AMBIENT6K [65] benchmark. Consistently, the restored ambient normalized images are characterized by more natural colors, an increased amount of restored details, and geometrically correct renderings even in areas severely affected by underexposure (second row). Overcoming the challenges specific to non-homogeneous direct color lighting, RLN $^2$ shows consistent performance, confirming the advantages quantified through evaluations provided in Tab. 3. + +To test $\mathrm{RLN}^2$ on out-of-domain data, in Fig. 8 we compare IFBlend [65] and our proposed $\mathrm{RLN}^2$ -Lf on inputs from LSMI [38], or complex lighting scenarios from product photography. Our $\mathrm{RLN}^2$ -Lf shows increased robustness, with plausible rendered colors under uniform white ambient lighting conditions. Moreover, in Fig. 9 we deploy it on AI-generated images representing various contents in challenging lighting environments. Applying our ALN solution as a pre-processing step in Neural Image Editing proves beneficial for IC-Light [19], with improved layer separation leading to less semantic discontinuities (rows 1, 2, and 4), and improves the rendered results quality in term of shadow casting (row 3) and foreground clarity (row 2). + +![](images/a5195b0f5809b0f174ee7d2a3c710d956905c7ce436a213f2eb939c6958f370b.jpg) + +![](images/0dc31a86c19e9d7c73ea3a682caddcaf6768d2b8b0a650e7a28ef00c6203f908.jpg) +Input +$\mathbf{RLN}^2$ -lfout + +![](images/76bb35738fc727ad942ba23e1aeceabda5f446a32b30bd3bff75e2e6176527f8.jpg) +w.o. RLN²-Lf +W. RLN²-Lf +Figure 9. ALN as a preprocessing step before image relighting with foreground condition. Results rendered with ICLight [19] for the input image (col. 3) vs. the $\mathrm{RLN}^2$ output (col. 4). + +# 5. Conclusion + +In this work, we propose extending the current ALN study to the challenges added by multicolor and multi-intensity RGB direct lighting. We propose CL3AN, a high-resolution image database showcasing a multitude of materials under complex RGB lighting scenarios while providing white-light ambient-lit reference images using a professional photography setup. We introduce RLN², a high-performance ALN algorithm setting a new standard for the ALN task, showing how alternative image representations can provide valuable control signals in feature refinement, or offer natural bypassing options for detail preservation. + +Acknowledgments. We thank the reviewers and ACs for their sustained effort and helpful feedback. This research was partially supported by the Alexander von Humboldt Foundation. + +# References + +[1] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1397-1406, 2020. 2 +[2] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1981-1990, 2021. 2 +[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 2 +[4] Alexandre Araujo, Jean Ponce, and Julien Mairal. Towards real-world focus stacking with deep learning. arXiv preprint arXiv:2311.17846, 2023. 2 +[5] Monika Bansal, Munish Kumar, and Manish Kumar. 2d object recognition: a comparative analysis of sift, surf and orb feature descriptors. Multimedia Tools and Applications, 80 (12):18839-18857, 2021. 2 +[6] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2 +[7] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 1 +[8] Ivaylo Boyadzhiev, Kavita Bala, Sylvain Paris, and Frédo Durand. User-guided white balance for mixed lighting conditions. ACM Trans. Graph., 31(6):200-1, 2012. 2 +[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 2, 3, 6, 7 +[10] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 3, 6, 7 +[11] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 2 +[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 6, 7 + +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[14] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 2 +[15] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2 +[16] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 2 +[17] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, and Marc Pollefeys. Building rome on a cloudless day. In Computer Vision - ECCV 2010, pages 368-381, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg. 2 +[18] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 203-212, 2021. 3 +[19] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 1, 8 +[20] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2 +[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 6 +[22] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5917-5926, 2023. 2 +[23] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2, 3, 7 +[24] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 2 +[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[26] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF interna + +tional conference on computer vision, pages 2472-2481, 2019.3 +[27] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0-0, 2018. 1 +[28] Yan Huang, Xinchang Lu, Yuhui Quan, Yong Xu, and Hui Ji. Image shadow removal via multi-scale deep retina decomposition. Pattern Recognition, page 111126, 2024. 3 +[29] Zhuo Hui, Kalyan Sunkavalli, Sunil Hadap, and Aswin C. Sankaranarayanan. Illuminant spectra-based source separation using flash photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6209-6218, 2018. 3 +[30] Zhuo Hui, Ayan Chakrabarti, Kalyan Sunkavalli, and Aswin C Sankaranarayanan. Learning to separate multiple illuminants in a single image. In Computer Vision and Pattern Recognition (CVPR 2019), 2019. 3 +[31] Anne Humeau-Heurtier. Texture feature extraction methods: A survey. IEEE access, 7:8975–9000, 2019. 1 +[32] Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 536-537, 2020. 1 +[33] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 1 +[34] Jong Ju Jeon, Jun Young Park, and Il Kyu Eom. Low-light image enhancement using gamma correction prior in mixed color spaces. Pattern Recognition, 146:110001, 2024. 2 +[35] Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion. Advances in Neural Information Processing Systems, 37:141129-141152, 2025. 1, 2 +[36] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. arXiv preprint arXiv:2211.08089, 2022. 1 +[37] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2021. 3, 6, 7 +[38] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 2, 3, 8 +[39] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings + +of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3 +[40] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 5 +[41] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 2 +[42] Wei-Sheng Lai, Orazio Gallo, Jinwei Gu, Deqing Sun, Ming-Hsuan Yang, and Jan Kautz. Video stitching for linear camera arrays. arXiv preprint arXiv:1907.13622, 2019. 2 +[43] EDWIN H. LAND. The retina. American Scientist, 52(2): 247-264, 1964. 3, 5 +[44] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 1, 3 +[45] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Mimt: Multi-illuminant color constancy via multitask local surface and light color learning. arXiv preprint arXiv:2211.08772, 2022. 2 +[46] Xinyu Li, Yanyi Zhang, Jianbo Yuan, Hanlin Lu, and Yibo Zhu. Discrete cosin transformer: Image modeling from frequency domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5468-5478, 2023. 3 +[47] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 3, 6, 7 +[48] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 3 +[49] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 6 +[50] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1680-1691, 2023. 3 +[51] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 1 +[52] Radoslaw Mantiuk, Rafal Mantiuk, Anna Tomaszewska, and Wolfgang Heidrich. Color correction for tone mapping. In Computer graphics forum, pages 193-202. Wiley Online Library, 2009. 2 + +[53] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523-3542, 2021. 1 +[54] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 3 +[55]Punatawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22646-22657, 2023. 2 +[56] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 1, 3 +[57] Pramod Rao, Gereon Fox, Abhinitra Meka, Mallikarjun BR, Fangneng Zhan, Tim Weyrich, Bernd Bickel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, et al. Lite2relight: 3d-aware single image portrait relighting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1 +[58] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 2 +[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2 +[60] Richard Szeliski and Richard Szeliski. Image alignment and stitching. Computer Vision: Algorithms and Applications, pages 401-441, 2022. 2 +[61] Hironobu Takano, Hiroki Kobayashi, and Kiyomi Nakamura. Rotation invariant iris recognition method adaptive to ambient lighting variation. IEICE transactions on information and systems, 90(6):955-962, 2007. 2 +[62] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20762-20772, 2023. 1, 2 +[63] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2 +[64] Florin-Alexandru Vasluianu, Tim Seizinger, and Radu Timofte. Wsrd: A novel benchmark for high resolution image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1826-1835, 2023. 1, 2, 3 +[65] Florin-Alexandru Vasluianu, Tim Seizinger, Zongwei Wu, Rakesh Ranjan, and Radu Timofte. Towards image ambient + +lighting normalization. arXiv preprint arXiv:2403.18730, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[66] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[67] Qiyang Wan, Ruiping Wang, and Xilin Chen. Interpretable object recognition by semantic prototype analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 800-809, 2024. 1 +[68] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 3 +[69] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 3 +[70] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[71] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 3, 6, 7 +[72] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 3 +[73] Ching-Chih Weng, Homer Chen, and Chiou-Shann Fuh. A novel automatic white balance method for digital still cameras. In 2005 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3801-3804. IEEE, 2005. 2 +[74] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5 +[75] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 3 +[76] Xudong Xie and Kin-Man Lam. An efficient illumination normalization method for face recognition. Pattern Recognition Letters, 27(6):609-617, 2006. 2 +[77] Xiaohua Xie, Wei-Shi Zheng, Jianhuang Lai, Pong C. Yuen, and Ching Y. Suen. Normalization of face illumination based on large-and small-scale features. IEEE Transactions on Image Processing, 20(7):1807-1821, 2011. 2 +[78] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. International Journal of Computer Vision, 132(9):3610-3632, 2024. 3 +[79] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 2 + +[80] Xiaoyu Xiang Denis Demandolx Rakesh Ranjan Radu Timofte Yawei Li, Yuchen Fan and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 6, 7 +[81] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 1 +[82] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 5 +[83] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3, 6, 7 +[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3, 6, 7, 8 +[85] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 3 +[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[87] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 1 +[88] Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed images via robust retina decomposition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3 +[89] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1219-1229, 2023. 3 +[90] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12786-12796, 2022. 2 \ No newline at end of file diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/images.zip b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9e4a4d4b78952cc798e5e5491cc8f370825f2541 --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1472e9954b6263a2493bf3ff91b367bfacd33966d7b8b2bd62daa224e079a0d2 +size 930492 diff --git a/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/layout.json b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e1b8cb014caf88904cf434ae243f69cd1b2fbf8b --- /dev/null +++ b/2025/After the Party_ Navigating the Mapping From Color to Ambient Lighting/layout.json @@ -0,0 +1,10315 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "text", + "content": "After the Party: Navigating the Mapping From Color to Ambient Lighting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 143, + 494, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 143, + 494, + 172 + ], + "spans": [ + { + "bbox": [ + 116, + 143, + 494, + 172 + ], + "type": "text", + "content": "Florin-Alexandru Vasluianu Tim Seizinger Zongwei Wu* Radu Timofte Computer Vision Lab, CAIDAS & IFI, University of Wurzburg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 174, + 404, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 174, + 404, + 185 + ], + "spans": [ + { + "bbox": [ + 203, + 174, + 404, + 185 + ], + "type": "text", + "content": "{firstname.lastname}@uni-wuerzburg.de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 548 + ], + "type": "text", + "content": "Illumination in practical scenarios is inherently complex, involving colored light sources, occlusions, and diverse material interactions that produce intricate reflectance and shading effects. However, existing methods often oversimplify this challenge by assuming a single light source or uniform, white-balanced lighting, leaving many of these complexities unaddressed. In this paper, we introduce CL3AN, the first large-scale, high-resolution dataset of its kind designed to facilitate the restoration of images captured under multiple Colored Light sources to their Ambient-Normalized counterparts. Through benchmarking, we find that leading approaches often produce artifacts—such as illumination inconsistencies, texture leakage, and color distortion—primarily due to their limited ability to precisely disentangle illumination from reflectance. Motivated by this insight, we achieve such a desired decomposition through a novel learning framework that leverages explicit chromaticity-luminance components guidance, drawing inspiration from the principles of the Retinex model. Extensive evaluations on existing benchmarks and our dataset demonstrate the effectiveness of our approach, showcasing enhanced robustness under non-homogeneous color lighting and material-specific reflectance variations, all while maintaining a highly competitive computational cost. The benchmark, codes, and models are available at www.github.com/fvasluianu97/RLN2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 135, + 583 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 296, + 688 + ], + "type": "text", + "content": "The goal of Ambient Lighting Normalization is recovering an ambient-lit image, under uniform lighting distribution, regardless of the lighting scenario characteristic to the depicted contents [65]. The ability to recover ambient-normalized images—free from spatially varying illumination artifacts—is critical for robust visual understanding [33, 51, 53]. In both practical imaging scenarios [27, 32] and the rapidly evolving domain of image generation and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 215, + 553, + 360 + ], + "blocks": [ + { + "bbox": [ + 316, + 215, + 553, + 360 + ], + "lines": [ + { + "bbox": [ + 316, + 215, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 215, + 553, + 360 + ], + "type": "image", + "image_path": "75f88847bb339ab0df3037fb024699d37acc278be393a5245022b8346c1a6fe7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 436 + ], + "type": "text", + "content": "Figure 1. Effective deployment of Ambient Lighting Normalization (ALN) as a pre-processing step in Neural Image Editing [19]. We compare the effectiveness of our proposed RLN² against state-of-the-art models in fields like Shadow Removal (DES3 [36]), or ALN (IFBlend [65]), trained on settings including single (1) [68] or multiple (3) white-domain-tuned [65] or color light sources." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 553 + ], + "type": "text", + "content": "synthesis [57, 81], illumination always involves complex interactions of multiple colored light sources, occlusions, and material properties, resulting in intricate shadows, glare, color shifts, and texture distortions. Lighting and material characteristics are the driving factors for explaining the manifestation of such artifacts, thus explaining the need for data-level representativity prior to proposing image restoration solutions aiming for improved degradation robustness or restoration performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 555, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 651 + ], + "type": "text", + "content": "While a wide range of applications require illumination-invariant representations [7, 31, 59, 67], existing normalization methods struggle to disentangle these intertwined factors [19, 35, 62, 81], as illustrated in Fig. 1. Even if data contributions further away from the natural light in " + }, + { + "bbox": [ + 313, + 555, + 556, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{STD + }" + }, + { + "bbox": [ + 313, + 555, + 556, + 651 + ], + "type": "text", + "content": " [44, 68], or multiple white-light-tuned light sources from AMBIENT6K [65] can improve general effectiveness, model design can further drive robustness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "This limitation stems primarily from oversimplified assumptions, such as single white light sources [64, 68] or homogeneous color fields [56, 68, 87], which fail to capture the richness of complex illumination. A key technical challenge lies in the lack of comprehensive datasets that" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 121, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 121, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 121, + 713 + ], + "type": "text", + "content": "*Corresponding" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9218" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "adequately represent the complexity of real-world lighting, especially outside the realm of natural white-color-tuned lighting. Prior efforts in shadow removal [64, 68], color constancy [6, 15, 16], or color correction [1, 22, 34, 52, 73] address isolated aspects of illumination but fall short of modeling multi-light interactions. Recent attempts [38, 65] introduce variable-intensity lighting datasets but lack chromatic diversity and material-specific reflectance variations. Furthermore, wide access to ambient-lit images outside the scope of scene-rendering software remains a challenge. As a result, learning-based methods often produce artifacts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 373 + ], + "type": "text", + "content": "In this paper, we address these limitations through two core contributions. First, we introduce CL3AN, the first large-scale dataset capturing high-resolution scenes under multiple colored directional lights paired with ambient-lit references. Each scene triplet includes variations beyond white-balanced lighting, colored lighting, and a diffuse-lit ground truth, enabling systematic study of illumination effects like self-shadows, highlights, glare areas, color spill or saturation, and other material-dependent artifacts. By benchmarking leading image restoration methods [9, 23, 65, 80], we observe that these models frequently suffer from critical shortcomings, such as the conflation of chromatic variations caused by illumination, texture leakage, inconsistent partial normalization, etc." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "content": "We identify the fundamental limitation of existing methods as their inability to systematically disentangle reflectance and illumination—a critical flaw exacerbated by the lack of explicit guidance for separating chromaticity and luminance under multi-colored lighting. Most current approaches [23, 80, 84] rely on single-stream, end-to-end feature modeling, which implicitly maps input images to ambient-normalized counterparts without incorporating physical awareness of light-material interactions. While some recent works attempt to address this by introducing pixel- and frequency-domain supervision [12, 65], more flexible towards reflectance- and chromaticity-information extraction, they remain limited in learning appropriate frequency filters — a challenging task without explicit guidance — resulting in suboptimal performance under the challenging multi-colored lighting conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "In this paper, from a novel perspective, we propose " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{RLN}^2" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", a Retinex-inspired framework that introduces parallel, physics-grounded streams to independently model reflectance (material properties) and illumination (lighting geometry/color). We observe that these intrinsic specificities are naturally encoded in the cylindrical representation of the input image, where the Hue (H) and Saturation (S) maps align with the observed chromaticity under the effect of the incoming light, while the Value (V) map provides strong priors for materials-specific reflectance. By embedding these domain-specific clues — such as reflectance information entropy (V), and smoothness constraints (H, S)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "into feature refinement, our framework achieves precise decomposition, effectively eliminating undesirable artifacts while preserving fine-grained details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 108, + 512, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 512, + 118 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 512, + 118 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 120, + 553, + 287 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 120, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 156 + ], + "type": "text", + "content": "- The introduction of the CL3AN, a first-of-its-kind dataset enabling the study of ambient lighting normalization under multiple RGB lights;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 191 + ], + "type": "text", + "content": "- The extensive study of a wide range of models, seminal works in various image restoration tasks, when trained and tested on the CL3AN dataset;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "text", + "content": "- We propose " + }, + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{RLN}^2" + }, + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "text", + "content": ", a robust algorithm built on the principle of simultaneous image reflection and lightness components compensation to recover an ambient lit image, characterized by uniform lighting distribution. " + }, + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\mathbf{RLN}^2" + }, + { + "bbox": [ + 313, + 192, + 553, + 287 + ], + "type": "text", + "content": " is proposed as a strong baseline for the ambient lighting normalization task, achieving state-of-the-art results on both the public AMBIENT6K dataset and the proposed CL3AN benchmark." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 400, + 311 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 378 + ], + "type": "text", + "content": "This work expands Ambient Lighting Normalization to multicolor direct lighting, addressing previously unexplored effects appearing in cluttered scenes, and analyzing shadow interactions and highlights appearing under multicolor direct lighting." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 379, + 553, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 553, + 450 + ], + "type": "text", + "content": "Lighting and Color: Color analysis is central to image processing and understanding, with various downstream tasks such as facial recognition [61, 76, 77], or tracking and matching applications [5, 14, 41, 59, 63] relying on the color similarity principle, thus benefiting from increased appearance robustness, especially under challenging lighting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": "Image Quality Assessment (IQA) methods [3, 79] show a strong correlation between details-abundance and image quality, motivating the focus on professional scene lighting rigs in all expert image photography applications. Given the scarcity of such equipment, alternative strategies [1, 2, 6, 45] are developed around image sensor sensor properties and cross-device constancy [8, 20]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 640 + ], + "type": "text", + "content": "Naturally, registered image stacks either for tasks such as multiple view geometry estimation [17, 90] or different image stitching applications require consistent color and content appearance [4, 11, 42, 60]. In such applications, the effort is focused on consistent lighting in the image acquisition, rather than post-processing correction. In-the-wild applications focusing on widely available amateur images [17, 24, 60] document extensive data curation strategies prior to model estimations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Later applications, given the newly-found popularity of neural 3D modeling, approach scene relighting as part of their rendering pipeline [58, 62]. Alternatively, approaches based on complex generative image models [35, 55] successfully tackle lighting manipulation, but at a high computational cost. Unfortunately, the limited availability of" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9219" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 167, + 170 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 167, + 170 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 167, + 170 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 167, + 170 + ], + "type": "image", + "image_path": "5c5d6867ece602091ba1d438df7b6ea0a5a3ef712b75b698daa42d55febdafec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 171, + 117, + 179 + ], + "lines": [ + { + "bbox": [ + 107, + 171, + 117, + 179 + ], + "spans": [ + { + "bbox": [ + 107, + 171, + 117, + 179 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 73, + 310, + 171 + ], + "blocks": [ + { + "bbox": [ + 171, + 73, + 310, + 171 + ], + "lines": [ + { + "bbox": [ + 171, + 73, + 310, + 171 + ], + "spans": [ + { + "bbox": [ + 171, + 73, + 310, + 171 + ], + "type": "image", + "image_path": "3febe8bc9579f52c5f669791b0ef5d09df51674914218a7419c43e262dee3310.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 171, + 246, + 179 + ], + "lines": [ + { + "bbox": [ + 237, + 171, + 246, + 179 + ], + "spans": [ + { + "bbox": [ + 237, + 171, + 246, + 179 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 162, + 184, + 200, + 192 + ], + "lines": [ + { + "bbox": [ + 162, + 184, + 200, + 192 + ], + "spans": [ + { + "bbox": [ + 162, + 184, + 200, + 192 + ], + "type": "text", + "content": "ISTD/WSRD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 196, + 265, + 204 + ], + "lines": [ + { + "bbox": [ + 227, + 196, + 265, + 204 + ], + "spans": [ + { + "bbox": [ + 227, + 196, + 265, + 204 + ], + "type": "text", + "content": "AMBIENT6K" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 208, + 324, + 217 + ], + "lines": [ + { + "bbox": [ + 282, + 208, + 324, + 217 + ], + "spans": [ + { + "bbox": [ + 282, + 208, + 324, + 217 + ], + "type": "text", + "content": "CL3AN (ours)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 314, + 74, + 428, + 171 + ], + "blocks": [ + { + "bbox": [ + 314, + 74, + 428, + 171 + ], + "lines": [ + { + "bbox": [ + 314, + 74, + 428, + 171 + ], + "spans": [ + { + "bbox": [ + 314, + 74, + 428, + 171 + ], + "type": "image", + "image_path": "1b8ff6d2a50f0fc4d47c05960bfefb803e8858eb4278b104df9aa3795c8f9d2c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 171, + 375, + 179 + ], + "lines": [ + { + "bbox": [ + 365, + 171, + 375, + 179 + ], + "spans": [ + { + "bbox": [ + 365, + 171, + 375, + 179 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 226, + 554, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 226, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 226, + 554, + 282 + ], + "type": "text", + "content": "Figure 2. A comparison between the lighting scenarios represented in datasets targeting different applications of Ambient Lighting Normalization (ISTD [68], WSRD[64], or AMBIENT6K [65]). In (A), ambient lighting is used to acquire a uniformly-lit reference image. In (B), direct lighting is used for shadow casting, while constraining consistency to the reference image in shadow-free segments. In (C), more directional lights are added to the direct lighting system. In our proposed setup (D), the direct lighting setup is based on RGB lights, dropping the color consistency constraint, and showcasing complex material-light interactions under various conditions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 430, + 74, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 430, + 74, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 430, + 74, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 430, + 74, + 553, + 171 + ], + "type": "image", + "image_path": "7af12f05f4f136016eb35f410ddb40aa2bcb29667905c2e376d5b04ee671f0f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 171, + 505, + 179 + ], + "lines": [ + { + "bbox": [ + 495, + 171, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 495, + 171, + 505, + 179 + ], + "type": "text", + "content": "D" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 294, + 342 + ], + "type": "text", + "content": "real-world ambient lit images, primarily due to the complexity of an ambient lighting setup, explains the difficulty in addressing real-world lighting normalization, even with the obvious availability of large-complexity models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 342, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 545 + ], + "type": "text", + "content": "Reliance on natural light for ambient-lit images proved a viable option, but hindered the variety of possible observed degradations, especially in terms of chromaticity shifts. Unfortunately, the large number of factors involved in light occlusion interaction greatly constrains a representative simulation in general settings. While ISTD/ISTD+ [44, 68], SRD [56] or USR [26] show exclusively flat surfaces under the influence of natural light, the influence of object self-shadows was just recently represented in WSRD [64], only possible under a laboratory setup. Recently, multiple source lighting gained in data-level representation. For example, in LSMI [38], the authors target White-Balance correction robustness under natural and artificial lighting. Alternatively, effects characteristic to multiple illuminants were captured in works such as [30, 54], or in the context of flash photography [29]. However, a uniformly-lit ambient normalized image can not be acquired in these settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 642 + ], + "type": "text", + "content": "Alternatively, the introduction of AMBIENT6K [65] coupled multiple light sources of varying intensity and orientation for complex interactions, while also providing an ambient-lit image based on an alternative lighting setup. Unfortunately, the strong connection between the operating parameters set for the direct and ambient lighting systems limits the study to exposure correction for white-aligned scene lighting (see Fig. 2 (A), (B), and (C))." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Image Restoration (IR): Joint color and exposure correction, while restoring details lost to occlusion or saturation, defines Ambient Lighting Normalization as an Image Restoration task. Various Convolution Neural Networks (CNNs) [10, 37, 83, 85] became strong baselines for the family of IR tasks. The later introduction of the attention" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 294, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 415 + ], + "type": "text", + "content": "mechanism [66] in image-based applications [13] enabled consistent performance advantages in the wide IR field. Focusing on different feature fusion techniques through specifically tailored attention mechanisms, based on the IR task particularities, defines the success of multiple models [10, 47, 48, 71, 80, 84] easily becoming benchmark solutions for various IR problems. Separate refinement for image reflectance and lightness under the Retinex model [43] proved particularly effective in low-light image enhancement (LLIE) [9, 72, 78, 88], or image shadow removal [28]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 521 + ], + "type": "text", + "content": "Global range dependencies benefits drive the development of State Space Models (SSM) [21], with promising results in general IR [23]. Alternatively, global knowledge can be leveraged through the integration of alternative frequency image representations in IR solutions. Solutions such as FFTFormer [39], DCFormer [46], SFNet [12], DW-GAN [18], and IFBlend [65] successfully extract priors based on Fourier Transform, Discrete Cosine Transform, or other Wavelet Transform variants." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 523, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 583 + ], + "type": "text", + "content": "Finally, diffusion-based iterative recurrent solutions [25] started to mature as benchmark solutions in Image Restoration [50, 69, 75, 89], especially in terms of perceived image quality. However, their characteristic increased computational effort hinders their deployment." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 596, + 370, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 596, + 370, + 608 + ], + "spans": [ + { + "bbox": [ + 314, + 596, + 370, + 608 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 616, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 641 + ], + "type": "text", + "content": "The core of our work is extending the study of Ambient Lighting Normalization to direct color lighting." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Fig. 3 compares the properties of a scene depicted under multi-source color direct lighting, and an equivalent image under white-aligned direct lighting. The goal to recover the ambient-lit reference image. Provided statistics, such as the RGB image histogram, and the Hue/Saturation footprint showcase the complexity of the ALN transformation, espe" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9220" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 294, + 203 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 203 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 203 + ], + "type": "image", + "image_path": "524aa8bdc7ef33427e115e065022f7192616a5b5f75a203606c166af218e5053.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 277 + ], + "type": "text", + "content": "Figure 3. A graphical representation of the Ambient Lighting Normalization transformation in the context of the white-aligned lighting, compared to the challenges posed by the introduction of a multicolor direct lighting system. The RGB image histogram, coupled with the Hue-Saturation rough projection is provided as a reference. Note the complexity of the change in image statistics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 361 + ], + "type": "text", + "content": "cially increasing with the introduction of variable intensity multicolor direct light sources. Naturally, the most notable difference comes from the color hue deviation. Materials adopt the color of incoming light, with the reflected light wavelength depending on the reflectance defining the represented materials, and the incoming light properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 492 + ], + "type": "text", + "content": "Furthermore, the complex geometry of the scene drives light occlusions, with numerous self-shadows at the image level. In this case, given the non-uniform color distribution, the color of the observed shadows depends on the local lighting availability and their light source-defining properties. A successful restoration of the ambient-lit image can not be performed without a full understanding of both scene geometry (for local exposure compensation) and lighting system configuration (for chromaticity uniformization, given the arithmetic of the locally available light radiation given the material properties)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 115, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 115, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 115, + 514 + ], + "type": "text", + "content": "3.1. CL3AN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 568 + ], + "type": "text", + "content": "CL3AN represents an image collection representing Triplets consisting of images acquired under both Color and White Lighting, and Ambient Normalized images, under uniformly distributed diffuse light." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "content": "As introduced in Fig. 2, CL3AN is the only dataset providing equivalent representations of the same scene under white-aligned direct lighting (setups (B), (C)), varying intensity multicolor lighting (setup (D)), while also providing an ambient-lit reference image (setup (A)). This makes CL3AN a first-of-its-kind dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "The capturing system relies on a professional photography studio setup, in which a dark room is used to ensure that all image-capturing parameters become trackable and controllable. Two separate scene lighting systems are concurrently developed, optimizing for uniform diffuse light distribution, respectively direct lighting, being triggered only" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "for the acquisition of the corresponding samples. Note that directional lights are fully controllable, so they can be tuned to the white natural light domain, producing images similar to those available in datasets such as AMBIENT6K [65]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 205 + ], + "type": "text", + "content": "The images are captured with a tripod-mounted and remotely operated Canon R6 MKII camera, at constant focal range. This leads to a perfect pixel alignment between the images characteristic to each triplet. For consistency and traceability, the camera-related parameters are kept fixed during the image-capturing process. All hardware parameters relevant to data acquisition are documented in Tab. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 289 + ], + "type": "text", + "content": "For the direct lighting system, three programmable mobile RGB lights are used to capture scene variants under variable light color, intensity, position, and orientation. The direct consequence of this set of parameters is a nonhomogeneous distribution of the incoming light color across the scene, implying that global color correction can not be used for restoration in this scenario (see Fig. 2 and Fig. 3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 290, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 410 + ], + "type": "text", + "content": "The diffuse ambient lighting system is composed of five white softbox lights, covering all sides of the cubic acquisition environment, except the base, which is the surface supporting the captured contents. Thus, the possibility for self-shadows in the ambient-lit images is almost fully mitigated, being furthermore enforced through geometry-based object filtering. Calibration over a 24-sample color checker under full white light intensity enables the alignment of both lighting setups, by determining the set of operating parameters producing a similar colors under optimal exposure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "text", + "content": "CL3AN is a scene-based dataset, covering 105 cluttered scenes representing an extensive set of materials, both conductors and dielectrics, with varying transparency, color, and surface roughness. The data is available as both RAW and RGB images, at " + }, + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "inline_equation", + "content": "24\\mathrm{MP}" + }, + { + "bbox": [ + 313, + 410, + 554, + 566 + ], + "type": "text", + "content": ", the camera sensor resolution. Data splitting is performed based on the represented contents, with 10 scenes used exclusively for testing, and 10 scenes for validation. Both testing and validation splits represent objects unseen during training, but under lighting conditions consistent with the training split. Additional metadata listing the objects within a scene is provided. The training split consists of a number of 3667 samples, with 437 images used for validation and 431 for testing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 575, + 364, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 364, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 364, + 586 + ], + "type": "text", + "content": "3.2. RLN" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Fig. 2 and Fig. 3 show the challenging environment posed by the introduction of multicolor direct lighting in the ALN field. Dramatic changes can be observed when mapping from the input image to the ambient-lit reference, especially in terms of observed chromaticity. Moreover, light occlusion by the many rough surfaces corresponding to the represented objects drives a wide variety of intricate effects, with a strong information loss. Learning a robust ALN mapping requires a robust design, able to disentangle significant information compensation in confined image segments from" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9221" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 62, + 553, + 119 + ], + "blocks": [ + { + "bbox": [ + 59, + 62, + 553, + 119 + ], + "lines": [ + { + "bbox": [ + 59, + 62, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 59, + 62, + 553, + 119 + ], + "type": "table", + "html": "
Camera parametersDiffuse LightingDirect Lighting
Exposure time (s)ApertureISOWhite balanceFocal dist. (mm)# Active lightsWhite light temp. (K)White light temp. rangeIntensity# Active lightsWhite light temp. (K)Intensity rangeColor Hue range
1/60f111006400K5056400±5%100%≥ 1640030-100%0°-360°
", + "image_path": "d5c4bb563cf260bf05c38f5d3b0cb788cbcce04a4ce5ff303094cb3fbfc36399.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 122, + 550, + 134 + ], + "lines": [ + { + "bbox": [ + 59, + 122, + 550, + 134 + ], + "spans": [ + { + "bbox": [ + 59, + 122, + 550, + 134 + ], + "type": "text", + "content": "Table 1. Dataset capturing parameters regarding the setup of the camera, the diffuse lighting system, and the direct lighting hardware." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 141, + 282, + 331 + ], + "blocks": [ + { + "bbox": [ + 67, + 141, + 282, + 331 + ], + "lines": [ + { + "bbox": [ + 67, + 141, + 282, + 331 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 282, + 331 + ], + "type": "image", + "image_path": "cfaea4fd6f24a885337942fc952b347d0581a2612a8f2b74d9887db7c0d1f8cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 295, + 374 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 374 + ], + "type": "text", + "content": "Figure 4. A graphical representation of the " + }, + { + "bbox": [ + 55, + 331, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 331, + 295, + 374 + ], + "type": "text", + "content": " architecture, including detailed representations for an Encoder (E) - Decoder (D) blocks pair on the Luminance L-branch. The Reflectance R-branch design is similar, being omitted for clarity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 381, + 216, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 216, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 216, + 392 + ], + "type": "text", + "content": "the large magnitude chromaticity shifts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "content": "Given these observations, problem sub-division is the core of our proposed design " + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "content": ", aiming for joint independent compensation in both Image Reflectance and Luminance components. Under the Retinex theory [43], a given image " + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "content": " can be decomposed as a product of two multiplicative terms, representing image illumination or lightness " + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "content": ", and the material specific image reflectance " + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 393, + 295, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "content": "Thus, an image pair " + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "inline_equation", + "content": "(I,I^{\\prime})" + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "content": " is the degraded version of the reference image " + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "content": ", can be represented by the component set " + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\{(L,R),(L',R')\\}" + }, + { + "bbox": [ + 55, + 476, + 295, + 514 + ], + "type": "text", + "content": ", in line with Eq. (1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 518, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 518, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 518, + 295, + 537 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} I (x, y) = L (x, y) \\cdot R (x, y) \\\\ I ^ {\\prime} (x) = I ^ {\\prime} (x) - R ^ {\\prime} (x) \\end{array} \\tag {1}", + "image_path": "125e36c5a745853c15a1e1edcc885550498796b2d732784549126ac6becee8c0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 533, + 236, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 533, + 236, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 533, + 236, + 547 + ], + "type": "interline_equation", + "content": "I ^ {\\prime} (x, y) = L ^ {\\prime} (x, y) \\cdot R ^ {\\prime} (x, y)", + "image_path": "4c0a6937085a87ef07126dee2da511a39875a17dbf2e188472f0022f2391d0cb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": "Therefore, our goal is to determine the pair " + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "(\\overline{L},\\overline{R})" + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\overline{L}" + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": " restores the details lost due to inadequate lighting, while " + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\overline{R}" + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": " aligns the input chromaticity with the white-aligned ambient lit setup." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 606, + 295, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 606, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 113, + 606, + 295, + 626 + ], + "type": "interline_equation", + "content": "L (x, y) = L ^ {\\prime} (x, y) + \\bar {L} (x, y) \\tag {2}", + "image_path": "aa520711c9e9a60f784ee3cef9f59c5a5af044b446bb5fdb97685f50a067a139.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 624, + 236, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 624, + 236, + 636 + ], + "spans": [ + { + "bbox": [ + 115, + 624, + 236, + 636 + ], + "type": "interline_equation", + "content": "R (x, y) = R ^ {\\prime} (x, y) + \\bar {R} (x, y)", + "image_path": "e22ba6b3593fb04ba0dd2f4b20bedb57b1f29538d8f8fbc1d876e2a484cc9cf9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": "Fig. 4 provides a graphical representation of the proposed " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": ", in which the " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "(R', L')" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": " components of the input " + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "inline_equation", + "content": "I'" + }, + { + "bbox": [ + 55, + 642, + 295, + 713 + ], + "type": "text", + "content": " are subjected to joint independent compensation, while the alternative HSV image representation information is used for feature refinement. Aligned with the properties of the HSV image representation, which codes color" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "text", + "content": "in a cylindrical coordinate system, in which the Value (V) channel is similar to explicit illumination intensity, and the Hue-Saturation (HS) represents color tones, the V channel is used as guidance on the L-branch of our model, while the HS pair drives feature refinement on the R-branch of RLN" + }, + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 144, + 553, + 203 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 204, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 275 + ], + "type": "text", + "content": "At the encoder level, the information goes through successive decomposition, in which Haar-DWT is used for spectral dissemination, alongside RGB feature downsampling. The dual-domain information will be altered in the feature refinement stage, at the " + }, + { + "bbox": [ + 313, + 204, + 553, + 275 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 204, + 553, + 275 + ], + "type": "text", + "content": " block level, through the Cross-Domain Feature Fusion Attention module (CDFFA)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "content": "Here, the low-frequency information " + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "inline_equation", + "content": "X_{lf}" + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "content": ", and the RGB extracted features " + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "inline_equation", + "content": "X_{L'}" + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "content": ", projected to an n-dimensional space, are then subjected to relevance enhancement, through a set of parameters based on dot product vector similarity (see Fig. 5) to features extracted from the guidance component. The V channel controls the estimation of the " + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\overline{L}" + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "content": " residual, while H and S drive the reflectance estimation " + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "inline_equation", + "content": "\\overline{R}" + }, + { + "bbox": [ + 313, + 275, + 555, + 360 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 381, + 380, + 463, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 380, + 463, + 393 + ], + "spans": [ + { + "bbox": [ + 381, + 380, + 463, + 393 + ], + "type": "interline_equation", + "content": "X _ {L ^ {\\prime}} = p r o j _ {n} \\left(X _ {L ^ {\\prime}}\\right)", + "image_path": "7dd5c7c232c17aee7e2caeacd8ea4868725e6d134b45a1e8105701ffb54eca26.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 384, + 395, + 553, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 395, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 384, + 395, + 553, + 415 + ], + "type": "interline_equation", + "content": "X _ {l f} = \\operatorname {p r o j} _ {n} \\left(X _ {l f}\\right) \\tag {3}", + "image_path": "f8bf14008c323c3c87f968efa18dec3b595fd610e829d9f43a76240059dc04eb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 391, + 410, + 485, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 410, + 485, + 422 + ], + "spans": [ + { + "bbox": [ + 391, + 410, + 485, + 422 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = s i m (X _ {L ^ {\\prime}, i}, X _ {V, i})", + "image_path": "6c1c76d0de14ac28a6b4d0f4a01090d3427330b6833637b1670c2c6ea5646146.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 391, + 425, + 484, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 425, + 484, + 437 + ], + "spans": [ + { + "bbox": [ + 391, + 425, + 484, + 437 + ], + "type": "interline_equation", + "content": "\\beta_ {i} = s i m (X _ {l f, i}, X _ {V, i})", + "image_path": "d723153ee8a1871aa12e02e430d1b8558e378eadb2488e6dd642ef3993d604b4.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 447, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 495 + ], + "type": "text", + "content": "The set of similarity coefficients is used to define a relevance probability distribution, then used to enhance the relevant information by filtering outlier features. The enhanced information is further refined before output." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 362, + 505, + 437, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 505, + 437, + 518 + ], + "spans": [ + { + "bbox": [ + 362, + 505, + 437, + 518 + ], + "type": "interline_equation", + "content": "\\alpha = \\operatorname {s o f t m a x} (\\alpha)", + "image_path": "4fe4507bac7269b58e5c10518609ade38ce5d040a16ec426b87d5ff3d928b375.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 520, + 553, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 520, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 364, + 520, + 553, + 533 + ], + "type": "interline_equation", + "content": "\\beta = \\operatorname {s o f t m a x} (\\beta) \\tag {4}", + "image_path": "11625b19a7e8e55972474d8d372408416794b35c88e1db9f6484e6c5bc84741b.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 349, + 534, + 520, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 534, + 520, + 548 + ], + "spans": [ + { + "bbox": [ + 349, + 534, + 520, + 548 + ], + "type": "interline_equation", + "content": "X _ {o u t} = \\left\\{\\alpha_ {i} \\cdot X _ {L ^ {\\prime}, i} \\right\\} _ {i = 0} ^ {n} + \\left\\{\\beta_ {i} \\cdot X _ {l f, i} \\right\\} _ {i = 0} ^ {n}", + "image_path": "800d9730a8e28aaec04f77007065a5d1c1843b910838d43213bdc8e0b7d79314.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 653 + ], + "type": "text", + "content": "At the decoder stage, the refined information is concatenated to the high-frequency spectral information from the corresponding encoder block. Moreover, an ImageNet [40] pretrained ConvNeXt-based feature extractor module [49], provides wide contextual information in feature fusion. A cross-attention module handles information exchange between the wide context information and the dual-domain features refined in the previous stage." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "For a fair comparison, at this level, " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " benefits from priors identical to [65], even if the implemented strategies are different. After linear projections, irrelevant features are suppressed through Channel Attention [74], being further refined for Reflectance/Luminance compensation [82]." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9222" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 68, + 224, + 205 + ], + "blocks": [ + { + "bbox": [ + 126, + 68, + 224, + 205 + ], + "lines": [ + { + "bbox": [ + 126, + 68, + 224, + 205 + ], + "spans": [ + { + "bbox": [ + 126, + 68, + 224, + 205 + ], + "type": "image", + "image_path": "5911921078cd2997832e0e46fc94616aae90386352e67c0c7b58314d0470d5c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "type": "text", + "content": "Figure 5. A detailed Reflectance/Luminance Refinement Block (R.Block), including the CDFFA attention module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 243, + 295, + 312 + ], + "blocks": [ + { + "bbox": [ + 56, + 243, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 312 + ], + "type": "table", + "html": "
ModelFusionGuidanceMACs (G.)PSNR↑SSIM↑LPIPS↓
\\( RLN^2 \\)NoneNone20.9519.8920.6890.271
\\( RLN^2 \\)concat.RGB23.6720.0180.7030.233
\\( RLN^2 \\)concat.L*a*b*23.6820.1490.7240.226
\\( RLN^2 \\)concat.HSV23.6820.2160.7310.220
\\( RLN^2 \\)(CDFFA)HSV22.7220.5230.7460.208
", + "image_path": "9afcf01e9b8e49a782ae4b0b8d6045adbeecf0d4a17cc68a30e8e2baf3eb4df9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 315, + 295, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 338 + ], + "type": "text", + "content": "Table 2. Inner-stage guidance feature fusion ablation study conducted on the testing split of the CL3AN dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "text", + "content": "Ablative study: Tab. 2 compares the standard RLN" + }, + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "text", + "content": " configuration, as represented in Fig. 4, to three similar variants in which the control mechanism implemented in the middle stage is modified. Reducing the " + }, + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "inline_equation", + "content": "R." + }, + { + "bbox": [ + 54, + 347, + 295, + 479 + ], + "type": "text", + "content": " block to simple feature refinement by removing any control mechanism shows an important performance deficit, consistent in image restoration fidelity and perceptual quality. Alternatively, using the RGB domain images, passed through a dynamical filter and then split based on the filtered output and its alternative residual shows a slight improvement, at a relatively low induced computational cost." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 555 + ], + "type": "text", + "content": "Task-aligned image representations, with explicit coding of image luminance and chromaticity, prove rich in relevant information, adding only the image conversion cost to the general computation. Focusing on relevant feature enhancement through the CDFFA introduction further increases model robustness and restoration performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "Implementation: " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " was developed in PyTorch, and the documented experiments were performed on up to three NVIDIA L40s 48 GB VRAM GPUs running a CUDA 12.6 environment. The optimization is split into multiple stages, with the resolution of the training patches gradually increasing as the training procedure progresses. The optimizer used was Adam, with the learning rate " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "lr = 0.0002" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", and the loss optimized was the L1 distance. A cosine learning rate scheduler was used, with two oscillation periods. Optimization deploys gradient clipping, with a standard value of 0.01 under the L1 norm. Further details regarding the training procedure description can be found in the attached supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 367, + 83 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 150 + ], + "type": "text", + "content": "Experimental Setup: Additionally to the introduced CL3AN dataset, we perform experiments on the AMBI-ENT6K image database [65], which consists of 6000 images representing scenes under the effect of up to three white-aligned LED directional lights (see Fig. 3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 152, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 222 + ], + "type": "text", + "content": "As compared metrics, we report restoration fidelity through Peak Signal to Noise Ratio (PSNR), and Structure Similarity Index Measure (SSIM) [70]. To quantify the perceptual quality of the restored images, we deploy the AlexNet-based Learned Perceptual Image Patch Similarity score (LPIPS) [86]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 223, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 390 + ], + "type": "text", + "content": "In terms of compared methods, we chose various types of solutions proposed in the image restoration field, which became seminal work after their introduction. For an extensive benchmark, we combine refined CNN solutions such as MPRNet [83] and HINet [37], with transformer-based architectures like SwinIR [47], NAFNet [10], Uformer [71], Restormer [84], and GRL [80]. Additionally, we consider solutions integrating state-space-models (e.g. MAMBAIR [21]), or solutions optimizing joint Spatial-Frequency entropy such as SFNet [12] and IFBlend [65]. Ultimately, we include Retinexformer [9], representing the class of solutions based on Retinex image decomposition. We train all the aforementioned methods using their official public implementations and settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "text", + "content": "Quantitative Results: In Tab. 3, we compare the aforementioned solutions with different variants of the proposed " + }, + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "text", + "content": ". The different variants represent our proposed models at different complexity levels, such that when comparing to similar existing solutions, advantages become easily identifiable. The computational complexity, as the Multiplications-Additions count (MAC), is reported for a patch of " + }, + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 313, + 390, + 553, + 486 + ], + "type": "text", + "content": " pixels, using ptflops for consistency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "The naming convention can be easily followed. Firstly, the difference between the RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-S and the RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-L is given by the complexity of the Wide Context Feature Extractor found in the feature fusion module, at the decoder stage, with RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-S using a ConvNeXt-T configuration, while RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-L benefits from a large ConvNeXt-XL model [49]. Secondly, we distinguish between RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": " variants that have/have not access to the dual image frequency domain representation. Thus RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-S and RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-L drop the frequency domain feature at the encoder, refinement stage, and decoder stage, using cross-attention for fusion, while RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-Sf and RLN" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 486, + 553, + 628 + ], + "type": "text", + "content": "-Lf use all the available information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 689 + ], + "type": "text", + "content": "All the compared RLN" + }, + { + "bbox": [ + 313, + 629, + 553, + 689 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 629, + 553, + 689 + ], + "type": "text", + "content": " variants are trained under the same conditions, including the optimizer, batch size, data augmentation strategies, optimized objective, and the number of samples seen during training. For AMBIENT6K [65], we adopt their publicly available benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Reading Tab. 3, it becomes obvious that " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is a high-performance solution, improving over similar state-of-the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9223" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 556, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 556, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 556, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 556, + 270 + ], + "type": "image", + "image_path": "84370e84a46d04b026a9061eed8b7db0076a71406aee594e75b204c73bb9ae3b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 271, + 527, + 282 + ], + "lines": [ + { + "bbox": [ + 82, + 271, + 527, + 282 + ], + "spans": [ + { + "bbox": [ + 82, + 271, + 527, + 282 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on scene #1 from test split of the CL3AN dataset. Best viewed in the electronic version." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 285, + 555, + 508 + ], + "blocks": [ + { + "bbox": [ + 58, + 285, + 555, + 508 + ], + "lines": [ + { + "bbox": [ + 58, + 285, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 58, + 285, + 555, + 508 + ], + "type": "table", + "html": "
Method nameTypeGeneral informationCL3AN (ours)AMBIENT6K [65]
Restoration TaskPriorMACs (G.)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
unprocessed----10.8370.4470.51813.4030.6520.250
NAFNet[10]Conv.Multi-task IRRGB15.9219.4760.7090.24920.5800.8080.142
MPRNet [83]Conv.Multi-task IRRGB37.2118.4530.6880.29120.9470.8200.129
SFNet [12]Transf.Multi-task IRRGB + Freq.31.2718.3820.6860.29120.5190.8120.141
SwinIR [47]Transf.Multi-task IRRGB37.8116.3860.6430.37220.5280.8170.131
Uformer [71]Transf.Multi-task IRRGB19.3317.5080.6550.31320.7760.8180.131
Restormer [84]Transf.Multi-task IRRGB35.3118.5600.6910.27821.1410.8170.132
HINet [37]Conv.Multi-task IRRGB42.6819.3880.7070.24820.8560.8210.129
IFBlend [65]ConvALNRGB + Freq.26.0120.3700.7200.22821.4430.8190.128
MAMBAIR [23]Transf. + SSMMulti-task IRRGB34.3218.9700.7040.254---
GRL [80]Transf.Multi-task IRRGB2.1618.0890.6720.308---
Retinexformer [9]Transf.LLIERGB4.8618.6490.6830.281---
\\( RLN^2-S \\) (ours)ConvALNRGB3.9519.9920.7180.23621.1810.8150.131
\\( RLN^2-Sf \\) (ours)ConvALNRGB + Freq.4.2420.1280.7300.22321.3330.8190.128
\\( RLN^2-L \\) (ours)ConvALNRGB22.4520.3830.7230.22221.5530.8210.123
\\( RLN^2-Lf \\) (ours)ConvALNRGB + Freq.22.7220.5230.7460.20821.7120.8250.120
", + "image_path": "1cab81132df09f5dc84aec075ea4802c72de1e6f6a3ad73373abad1acde2c1f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 512, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 512, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 555, + 536 + ], + "type": "text", + "content": "Table 3. Various RLN" + }, + { + "bbox": [ + 55, + 512, + 555, + 536 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 512, + 555, + 536 + ], + "type": "text", + "content": " variants versus current SOTA image restoration models, on the proposed CL3AN dataset test split (1920 × 1440 px.), and AMBIENT6K benchmark [65] (1280 × 960 px. default size). The best and the second best performing models are emphasized." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "content": "art solutions at all complexity levels, on both CL3AN and AMBIENT6K [65]. While " + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "content": "-S largely outperforms RetinexFormer [9], at a lower computational cost, " + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 541, + 295, + 601 + ], + "type": "text", + "content": "-Sf improves over the perceptual performance of IFBlend [65], with a fraction of its required floating point operations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "text", + "content": "Against IFBlend [65], a strong recently proposed solution in the ALN field, " + }, + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "text", + "content": "-L shows an advantage, especially in joint RGB-Frequency domains processing, with an " + }, + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\approx 15\\%" + }, + { + "bbox": [ + 55, + 603, + 295, + 652 + ], + "type": "text", + "content": " complexity reduction, while outperforming it." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Another important observation is regarding the architectural core of the compared methods. Under both white-aligned and RGB direct lighting, ALN tends to favor solutions showing an extensive receptive field [37, 83], rather than transformers optimized for local dependencies [47, 71," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 541, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 555, + 638 + ], + "type": "text", + "content": "84], even when balancing to wider context blocks [80]. However, MPRNet [83] and HINet [37] are larger models, with an additional cost of approx. " + }, + { + "bbox": [ + 313, + 541, + 555, + 638 + ], + "type": "inline_equation", + "content": "63\\% - 87\\%" + }, + { + "bbox": [ + 313, + 541, + 555, + 638 + ], + "type": "text", + "content": ", compared to the highest complexity RLN²-Lf variant. While state-state models prove computationally ineffective for the level of performance achieved, frequency-domain information access shows a consistent advantage, especially in terms of observed perceptual quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Qualitative Results: In Fig. 6, we provide for qualitative evaluation a set of samples representing a cluttered scene under the effect of non-homogeneous RGB direct lighting. Note the complex geometry and the wide RGB spectrum represented at the ambient lit image level. " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "-Lf shows improved behavior in terms of color normalization (row" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9224" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 81, + 159, + 156 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 119, + 80 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 119, + 80 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 119, + 80 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 81, + 159, + 156 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 159, + 156 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 159, + 156 + ], + "type": "image", + "image_path": "2d78f011811514e37254f59a4686e5871d6252b64c8b02dde83267fae95fcdb3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 159, + 81, + 259, + 156 + ], + "blocks": [ + { + "bbox": [ + 178, + 71, + 240, + 82 + ], + "lines": [ + { + "bbox": [ + 178, + 71, + 240, + 82 + ], + "spans": [ + { + "bbox": [ + 178, + 71, + 240, + 82 + ], + "type": "text", + "content": "Restormer [84]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 81, + 259, + 156 + ], + "lines": [ + { + "bbox": [ + 159, + 81, + 259, + 156 + ], + "spans": [ + { + "bbox": [ + 159, + 81, + 259, + 156 + ], + "type": "image", + "image_path": "bc79a030fe1552ecff7329773b69dfe881f9d0139781ef004b117c0d5d8ac069.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 260, + 81, + 359, + 156 + ], + "blocks": [ + { + "bbox": [ + 283, + 70, + 337, + 82 + ], + "lines": [ + { + "bbox": [ + 283, + 70, + 337, + 82 + ], + "spans": [ + { + "bbox": [ + 283, + 70, + 337, + 82 + ], + "type": "text", + "content": "IFBlend [65]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 81, + 359, + 156 + ], + "lines": [ + { + "bbox": [ + 260, + 81, + 359, + 156 + ], + "spans": [ + { + "bbox": [ + 260, + 81, + 359, + 156 + ], + "type": "image", + "image_path": "911dcd32925decfa6917cb4c8e710558af1a89e1dc0f849723c425ed092c052d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 360, + 81, + 461, + 156 + ], + "blocks": [ + { + "bbox": [ + 377, + 70, + 445, + 82 + ], + "lines": [ + { + "bbox": [ + 377, + 70, + 445, + 82 + ], + "spans": [ + { + "bbox": [ + 377, + 70, + 445, + 82 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 377, + 70, + 445, + 82 + ], + "type": "text", + "content": " -Lf (Ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 81, + 461, + 156 + ], + "lines": [ + { + "bbox": [ + 360, + 81, + 461, + 156 + ], + "spans": [ + { + "bbox": [ + 360, + 81, + 461, + 156 + ], + "type": "image", + "image_path": "b25e940749d3e4d41461e8685f3bf17247e00e4a76f22f180241f0c21170d135.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 462, + 81, + 563, + 156 + ], + "blocks": [ + { + "bbox": [ + 462, + 81, + 563, + 156 + ], + "lines": [ + { + "bbox": [ + 462, + 81, + 563, + 156 + ], + "spans": [ + { + "bbox": [ + 462, + 81, + 563, + 156 + ], + "type": "image", + "image_path": "ac627ee46216cadc8170a9b557820aa4bdfc9fd2e9efe3d48bbe85fad075985d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 159, + 158, + 233 + ], + "blocks": [ + { + "bbox": [ + 58, + 159, + 158, + 233 + ], + "lines": [ + { + "bbox": [ + 58, + 159, + 158, + 233 + ], + "spans": [ + { + "bbox": [ + 58, + 159, + 158, + 233 + ], + "type": "image", + "image_path": "5f7cf4fe9daeea95e6354c8bdd4de511c799801bb7210fa5eb2d38d87e59d36d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 553, + 257 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison challenging samples from the AMBIENT6K testing set [65]. Our " + }, + { + "bbox": [ + 55, + 235, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 235, + 553, + 257 + ], + "type": "text", + "content": " produces improved quality ambient normalized images, with sharper images, and a lower amount of visible artifacts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 159, + 159, + 259, + 233 + ], + "blocks": [ + { + "bbox": [ + 159, + 159, + 259, + 233 + ], + "lines": [ + { + "bbox": [ + 159, + 159, + 259, + 233 + ], + "spans": [ + { + "bbox": [ + 159, + 159, + 259, + 233 + ], + "type": "image", + "image_path": "c564554c3feb8b4576da27f42da1e4ebae64a8ca4bb33a1a0f2b5e4480fa72d0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 260, + 159, + 360, + 233 + ], + "blocks": [ + { + "bbox": [ + 260, + 159, + 360, + 233 + ], + "lines": [ + { + "bbox": [ + 260, + 159, + 360, + 233 + ], + "spans": [ + { + "bbox": [ + 260, + 159, + 360, + 233 + ], + "type": "image", + "image_path": "7e8b0b6ae3af887a83a546cf328935ef4a65e5088436275c2cb7bb02037e5eb0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 360, + 159, + 461, + 233 + ], + "blocks": [ + { + "bbox": [ + 360, + 159, + 461, + 233 + ], + "lines": [ + { + "bbox": [ + 360, + 159, + 461, + 233 + ], + "spans": [ + { + "bbox": [ + 360, + 159, + 461, + 233 + ], + "type": "image", + "image_path": "5b530f3ba765428d3f9f6504f6c6cdb178e7796815103c86f340c8a72afdb05f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 462, + 159, + 563, + 233 + ], + "blocks": [ + { + "bbox": [ + 462, + 159, + 563, + 233 + ], + "lines": [ + { + "bbox": [ + 462, + 159, + 563, + 233 + ], + "spans": [ + { + "bbox": [ + 462, + 159, + 563, + 233 + ], + "type": "image", + "image_path": "a2a4086cf0e714c57920886416500b4edf81d4e0bfc135cb8965d4cc65b3d0f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 57, + 262, + 176, + 341 + ], + "blocks": [ + { + "bbox": [ + 57, + 262, + 176, + 341 + ], + "lines": [ + { + "bbox": [ + 57, + 262, + 176, + 341 + ], + "spans": [ + { + "bbox": [ + 57, + 262, + 176, + 341 + ], + "type": "image", + "image_path": "23f5cb000b6f1a367762fdb30f508ecffd5b966eac8bbe21db3bb7ec48320ff5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 343, + 82, + 349 + ], + "lines": [ + { + "bbox": [ + 69, + 343, + 82, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 343, + 82, + 349 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 176, + 262, + 294, + 342 + ], + "blocks": [ + { + "bbox": [ + 176, + 262, + 294, + 342 + ], + "lines": [ + { + "bbox": [ + 176, + 262, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 176, + 262, + 294, + 342 + ], + "type": "image", + "image_path": "8deaa16b60f6fb1936ce7c745081118a29eb0e617197d9366dafae8870ab643f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 343, + 202, + 349 + ], + "lines": [ + { + "bbox": [ + 188, + 343, + 202, + 349 + ], + "spans": [ + { + "bbox": [ + 188, + 343, + 202, + 349 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 266, + 343, + 283, + 349 + ], + "lines": [ + { + "bbox": [ + 266, + 343, + 283, + 349 + ], + "spans": [ + { + "bbox": [ + 266, + 343, + 283, + 349 + ], + "type": "text", + "content": "RLN²-Lf" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "type": "text", + "content": "Figure 8. IFBlend[65] and " + }, + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 55, + 360, + 295, + 383 + ], + "type": "text", + "content": "-Lf on real-world images from LSMI [38] (left), or real product photography samples (right)." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 389, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 412 + ], + "type": "text", + "content": "1), superior restorations in highlights-rich areas, and better high-frequency shadow areas compensation (row 2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 413, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 413, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 413, + 295, + 533 + ], + "type": "text", + "content": "Fig. 7 provides for visual comparison equivalent renderings against the top performing state-of-the-art methods the AMBIENT6K [65] benchmark. Consistently, the restored ambient normalized images are characterized by more natural colors, an increased amount of restored details, and geometrically correct renderings even in areas severely affected by underexposure (second row). Overcoming the challenges specific to non-homogeneous direct color lighting, RLN" + }, + { + "bbox": [ + 54, + 413, + 295, + 533 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 413, + 295, + 533 + ], + "type": "text", + "content": " shows consistent performance, confirming the advantages quantified through evaluations provided in Tab. 3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "text", + "content": "To test " + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "text", + "content": " on out-of-domain data, in Fig. 8 we compare IFBlend [65] and our proposed " + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "text", + "content": "-Lf on inputs from LSMI [38], or complex lighting scenarios from product photography. Our " + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 54, + 533, + 295, + 689 + ], + "type": "text", + "content": "-Lf shows increased robustness, with plausible rendered colors under uniform white ambient lighting conditions. Moreover, in Fig. 9 we deploy it on AI-generated images representing various contents in challenging lighting environments. Applying our ALN solution as a pre-processing step in Neural Image Editing proves beneficial for IC-Light [19], with improved layer separation leading to less semantic discontinuities (rows 1, 2, and 4), and improves the rendered results quality in term of shadow casting (row 3) and foreground clarity (row 2)." + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 262, + 434, + 385 + ], + "blocks": [ + { + "bbox": [ + 317, + 262, + 434, + 385 + ], + "lines": [ + { + "bbox": [ + 317, + 262, + 434, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 434, + 385 + ], + "type": "image", + "image_path": "a5195b0f5809b0f174ee7d2a3c710d956905c7ce436a213f2eb939c6958f370b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 317, + 385, + 433, + 506 + ], + "blocks": [ + { + "bbox": [ + 317, + 385, + 433, + 506 + ], + "lines": [ + { + "bbox": [ + 317, + 385, + 433, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 385, + 433, + 506 + ], + "type": "image", + "image_path": "0dc31a86c19e9d7c73ea3a682caddcaf6768d2b8b0a650e7a28ef00c6203f908.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 510, + 358, + 518 + ], + "lines": [ + { + "bbox": [ + 335, + 510, + 358, + 518 + ], + "spans": [ + { + "bbox": [ + 335, + 510, + 358, + 518 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 510, + 427, + 518 + ], + "lines": [ + { + "bbox": [ + 383, + 510, + 427, + 518 + ], + "spans": [ + { + "bbox": [ + 383, + 510, + 427, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{RLN}^2" + }, + { + "bbox": [ + 383, + 510, + 427, + 518 + ], + "type": "text", + "content": " -lfout" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 435, + 262, + 553, + 507 + ], + "blocks": [ + { + "bbox": [ + 435, + 262, + 553, + 507 + ], + "lines": [ + { + "bbox": [ + 435, + 262, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 435, + 262, + 553, + 507 + ], + "type": "image", + "image_path": "76bb35738fc727ad942ba23e1aeceabda5f446a32b30bd3bff75e2e6176527f8.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 510, + 488, + 517 + ], + "lines": [ + { + "bbox": [ + 441, + 510, + 488, + 517 + ], + "spans": [ + { + "bbox": [ + 441, + 510, + 488, + 517 + ], + "type": "text", + "content": "w.o. RLN²-Lf" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 510, + 544, + 517 + ], + "lines": [ + { + "bbox": [ + 503, + 510, + 544, + 517 + ], + "spans": [ + { + "bbox": [ + 503, + 510, + 544, + 517 + ], + "type": "text", + "content": "W. RLN²-Lf" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 523, + 555, + 557 + ], + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 557 + ], + "type": "text", + "content": "Figure 9. ALN as a preprocessing step before image relighting with foreground condition. Results rendered with ICLight [19] for the input image (col. 3) vs. the " + }, + { + "bbox": [ + 313, + 523, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathrm{RLN}^2" + }, + { + "bbox": [ + 313, + 523, + 555, + 557 + ], + "type": "text", + "content": " output (col. 4)." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 555, + 715 + ], + "type": "text", + "content": "In this work, we propose extending the current ALN study to the challenges added by multicolor and multi-intensity RGB direct lighting. We propose CL3AN, a high-resolution image database showcasing a multitude of materials under complex RGB lighting scenarios while providing white-light ambient-lit reference images using a professional photography setup. We introduce RLN², a high-performance ALN algorithm setting a new standard for the ALN task, showing how alternative image representations can provide valuable control signals in feature refinement, or offer natural bypassing options for detail preservation." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 70, + 702, + 279, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 279, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 279, + 713 + ], + "type": "text", + "content": "deepai.org/machine-learning-model/text2img" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9225" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgments. We thank the reviewers and ACs for their sustained effort and helpful feedback. This research was partially supported by the Alexander von Humboldt Foundation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 154, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 154, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 154, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 154, + 296, + 196 + ], + "type": "text", + "content": "[1] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 1397-1406, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "type": "text", + "content": "[2] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1981-1990, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 256, + 296, + 311 + ], + "type": "text", + "content": "[3] Lorenzo Agnolucci, Leonardo Galteri, Marco Bertini, and Alberto Del Bimbo. Arniqa: Learning distortion manifold for image quality assessment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 189-198, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 313, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 346 + ], + "type": "text", + "content": "[4] Alexandre Araujo, Jean Ponce, and Julien Mairal. Towards real-world focus stacking with deep learning. arXiv preprint arXiv:2311.17846, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 296, + 392 + ], + "type": "text", + "content": "[5] Monika Bansal, Munish Kumar, and Manish Kumar. 2d object recognition: a comparative analysis of sift, surf and orb feature descriptors. Multimedia Tools and Applications, 80 (12):18839-18857, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 394, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 427 + ], + "type": "text", + "content": "[6] Jonathan T Barron. Convolutional color constancy. In Proceedings of the IEEE International Conference on Computer Vision, pages 379-387, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 429, + 296, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 429, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 296, + 472 + ], + "type": "text", + "content": "[7] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 475, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 296, + 507 + ], + "type": "text", + "content": "[8] Ivaylo Boyadzhiev, Kavita Bala, Sylvain Paris, and Frédo Durand. User-guided white balance for mixed lighting conditions. ACM Trans. Graph., 31(6):200-1, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 564 + ], + "type": "text", + "content": "[9] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 296, + 599 + ], + "type": "text", + "content": "[10] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. arXiv preprint arXiv:2204.04676, 2022. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 656 + ], + "type": "text", + "content": "[11] Yen-Chi Cheng, Chieh Hubert Lin, Hsin-Ying Lee, Jian Ren, Sergey Tulyakov, and Ming-Hsuan Yang. Inout: Diverse image outpainting via gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11431-11440, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 171 + ], + "type": "text", + "content": "[14] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7157-7173, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[15] Graham D. Finlayson, Steven D. Hordley, and Paul M. Hubel. Color by correlation: A simple, unifying framework for color constancy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(11):1209-1221, 2001. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 240 + ], + "type": "text", + "content": "[16] David H Foster. Color constancy. Vision research, 51(7): 674-700, 2011. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 554, + 308 + ], + "type": "text", + "content": "[17] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, and Marc Pollefeys. Building rome on a cloudless day. In Computer Vision - ECCV 2010, pages 368-381, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 364 + ], + "type": "text", + "content": "[18] Minghan Fu, Huan Liu, Yankun Yu, Jun Chen, and Keyan Wang. Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 203-212, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "text", + "content": "[19] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding instruction-based image editing via multimodal large language models. arXiv preprint arXiv:2309.17102, 2023. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "text", + "content": "[20] Arjan Gijsenij, Rui Lu, and Theo Gevers. Color constancy for multiple light sources. IEEE Transactions on image processing, 21(2):697-707, 2011. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "text", + "content": "[21] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 555, + 533 + ], + "type": "text", + "content": "[22] Julian Jorge Andrade Guerreiro, Mitsuru Nakazawa, and Björn Stenger. Pct-net: Full resolution image harmonization using pixel-wise color transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5917-5926, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 535, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 588 + ], + "type": "text", + "content": "[23] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In European Conference on Computer Vision, pages 222-241. Springer, 2025. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "text", + "content": "[24] Kamal Gupta, Varun Jampani, Carlos Esteves, Abhinav Shrivastava, Ameesh Makadia, Noah Snively, and Abhishek Kar. ASIC: Aligning sparse in-the-wild image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4134-4145, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[25] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF interna" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9226" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "tional conference on computer vision, pages 2472-2481, 2019.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 159 + ], + "type": "text", + "content": "[27] Jie Huang, Pengfei Zhu, Mingrui Geng, Jiewen Ran, Xingguang Zhou, Chen Xing, Pengfei Wan, and Xiangyang Ji. Range scaling global u-net for perceptual image enhancement on mobile devices. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0-0, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 195 + ], + "type": "text", + "content": "[28] Yan Huang, Xinchang Lu, Yuhui Quan, Yong Xu, and Hui Ji. Image shadow removal via multi-scale deep retina decomposition. Pattern Recognition, page 111126, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[29] Zhuo Hui, Kalyan Sunkavalli, Sunil Hadap, and Aswin C. Sankaranarayanan. Illuminant spectra-based source separation using flash photography. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6209-6218, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 294 + ], + "type": "text", + "content": "[30] Zhuo Hui, Ayan Chakrabarti, Kalyan Sunkavalli, and Aswin C Sankaranarayanan. Learning to separate multiple illuminants in a single image. In Computer Vision and Pattern Recognition (CVPR 2019), 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "text", + "content": "[31] Anne Humeau-Heurtier. Texture feature extraction methods: A survey. IEEE access, 7:8975–9000, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[32] Andrey Ignatov, Luc Van Gool, and Radu Timofte. Replacing mobile camera isp with a single deep learning model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 536-537, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[33] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 460 + ], + "type": "text", + "content": "[34] Jong Ju Jeon, Jun Young Park, and Il Kyu Eom. Low-light image enhancement using gamma correction prior in mixed color spaces. Pattern Recognition, 146:110001, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[35] Haian Jin, Yuan Li, Fujun Luan, Yuanbo Xiangli, Sai Bi, Kai Zhang, Zexiang Xu, Jin Sun, and Noah Snavely. Neural gaffer: Relighting any object via diffusion. Advances in Neural Information Processing Systems, 37:141129-141152, 2025. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[36] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. arXiv preprint arXiv:2211.08089, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[37] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2021. 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "text", + "content": "[38] Dongyoung Kim, Jinwoo Kim, Seonghyeon Nam, Dongwoo Lee, Yeonkyung Lee, Nahiup Kang, Hyong-Euk Lee, ByungIn Yoo, Jae-Joon Han, and Seon Joo Kim. Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2410–2419, 2021. 2, 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[39] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "text", + "content": "[40] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, page 1097-1105, Red Hook, NY, USA, 2012. Curran Associates Inc. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 161, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 226 + ], + "type": "text", + "content": "[41] Avisek Lahiri, Vivek Kwatra, Christian Frueh, John Lewis, and Chris Bregler. Lipsync3d: Data-efficient learning of personalized 3d talking faces from video using pose and lighting normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2755-2764, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "text", + "content": "[42] Wei-Sheng Lai, Orazio Gallo, Jinwei Gu, Deqing Sun, Ming-Hsuan Yang, and Jan Kautz. Video stitching for linear camera arrays. arXiv preprint arXiv:1907.13622, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 281 + ], + "type": "text", + "content": "[43] EDWIN H. LAND. The retina. American Scientist, 52(2): 247-264, 1964. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[44] Hieu Le and Dimitris Samaras. Physics-based shadow image decomposition for shadow removal. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(12):9088-9101, 2021. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[45] Shuwei Li, Jikai Wang, Michael S Brown, and Robby T Tan. Mimt: Multi-illuminant color constancy via multitask local surface and light color learning. arXiv preprint arXiv:2211.08772, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[46] Xinyu Li, Yanyi Zhang, Jianbo Yuan, Hanlin Lu, and Yibo Zhu. Discrete cosin transformer: Image modeling from frequency domain. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5468-5478, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[47] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "text", + "content": "[48] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "type": "text", + "content": "[49] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[50] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schön. Refusion: Enabling large-size realistic image restoration with latent-space diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1680-1691, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[51] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[52] Radoslaw Mantiuk, Rafal Mantiuk, Anna Tomaszewska, and Wolfgang Heidrich. Color correction for tone mapping. In Computer graphics forum, pages 193-202. Wiley Online Library, 2009. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9227" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[53] Shervin Minaee, Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3523-3542, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[54] Lukas Murmann, Michael Gharbi, Miika Aittala, and Fredo Durand. A dataset of multi-illumination images in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4080–4089, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[55]Punatawat Ponglertnapakorn, Nontawat Tritrong, and Supasorn Suwajanakorn. Difareli: Diffusion face relighting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22646-22657, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[56] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 341 + ], + "type": "text", + "content": "[57] Pramod Rao, Gereon Fox, Abhinitra Meka, Mallikarjun BR, Fangneng Zhan, Tim Weyrich, Bernd Bickel, Hanspeter Pfister, Wojciech Matusik, Mohamed Elgharib, et al. Lite2relight: 3d-aware single image portrait relighting. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 387 + ], + "type": "text", + "content": "[58] Viktor Rudnev, Mohamed Elgharib, William Smith, Lingjie Liu, Vladislav Golyanik, and Christian Theobalt. Nerf for outdoor scene relighting. In European Conference on Computer Vision, pages 615-631. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[59] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 477 + ], + "type": "text", + "content": "[60] Richard Szeliski and Richard Szeliski. Image alignment and stitching. Computer Vision: Algorithms and Applications, pages 401-441, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[61] Hironobu Takano, Hiroki Kobayashi, and Kiyomi Nakamura. Rotation invariant iris recognition method adaptive to ambient lighting variation. IEICE transactions on information and systems, 90(6):955-962, 2007. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 589 + ], + "type": "text", + "content": "[62] Marco Toschi, Riccardo De Matteo, Riccardo Spezialetti, Daniele De Gregorio, Luigi Di Stefano, and Samuele Salti. Relight my nerf: A dataset for novel view synthesis and relighting of real world objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20762-20772, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[63] Prune Truong, Martin Danelljan, Radu Timofte, and Luc Van Gool. Pdc-net+: Enhanced probabilistic dense correspondence network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10247-10266, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[64] Florin-Alexandru Vasluianu, Tim Seizinger, and Radu Timofte. Wsrd: A novel benchmark for high resolution image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1826-1835, 2023. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[65] Florin-Alexandru Vasluianu, Tim Seizinger, Zongwei Wu, Rakesh Ranjan, and Radu Timofte. Towards image ambient" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "lighting normalization. arXiv preprint arXiv:2403.18730, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 117 + ], + "type": "text", + "content": "[66] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[67] Qiyang Wan, Ruiping Wang, and Xilin Chen. Interpretable object recognition by semantic prototype analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 800-809, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 217 + ], + "type": "text", + "content": "[68] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[69] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[70] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "type": "text", + "content": "[71] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[72] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[73] Ching-Chih Weng, Homer Chen, and Chiou-Shann Fuh. A novel automatic white balance method for digital still cameras. In 2005 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3801-3804. IEEE, 2005. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "type": "text", + "content": "[74] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "text", + "content": "[75] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13095-13105, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 559 + ], + "type": "text", + "content": "[76] Xudong Xie and Kin-Man Lam. An efficient illumination normalization method for face recognition. Pattern Recognition Letters, 27(6):609-617, 2006. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "text", + "content": "[77] Xiaohua Xie, Wei-Shi Zheng, Jianhuang Lai, Pong C. Yuen, and Ching Y. Suen. Normalization of face illumination based on large-and small-scale features. IEEE Transactions on Image Processing, 20(7):1807-1821, 2011. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "text", + "content": "[78] Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retina model for low-light image enhancement. International Journal of Computer Vision, 132(9):3610-3632, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 713 + ], + "type": "text", + "content": "[79] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1191-1200, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9228" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 630 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[80] Xiaoyu Xiang Denis Demandolx Rakesh Ranjan Radu Timofte Yawei Li, Yuchen Fan and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[81] Jae Shin Yoon, Zhixin Shu, Mengwei Ren, Cecilia Zhang, Yannick Hold-Geoffroy, Krishna Kumar Singh, and He Zhang. Generative portrait shadow removal. ACM Transactions on Graphics (TOG), 43(6):1-13, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 205 + ], + "type": "text", + "content": "[82] Weihao Yu and Xinchao Wang. Mambaout: Do we really need mamba for vision? arXiv preprint arXiv:2405.07992, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[83] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14821-14831, 2021. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 328 + ], + "type": "text", + "content": "[84] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[85] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 407 + ], + "type": "text", + "content": "[86] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[87] Xuaner Zhang, Jonathan T Barron, Yun-Ta Tsai, Rohit Pandey, Xiuming Zhang, Ren Ng, and David E Jacobs. Portrait shadow manipulation. ACM Transactions on Graphics (TOG), 39(4):78-1, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 508 + ], + "type": "text", + "content": "[88] Anqi Zhu, Lin Zhang, Ying Shen, Yong Ma, Shengjie Zhao, and Yicong Zhou. Zero-shot restoration of underexposed images via robust retina decomposition. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[89] Yuanzhi Zhu, Kai Zhang, Jingyun Liang, Jiezhang Cao, Bihan Wen, Radu Timofte, and Luc Van Gool. Denoising diffusion models for plug-and-play image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1219-1229, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 630 + ], + "type": "text", + "content": "[90] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9229" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_content_list.json b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..142faf353df47d13a36b621d209bb14ea1e38024 --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_content_list.json @@ -0,0 +1,1590 @@ +[ + { + "type": "text", + "text": "Agreement aware and dissimilarity oriented GLOM", + "text_level": 1, + "bbox": [ + 235, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$\\mathrm{RuZeng^{1}}$ Yan Song $^{1*}$ Yang Zhang $^{2}$ Yanling Hu $^{3}$ Hui Yu $^{4}$", + "bbox": [ + 243, + 178, + 750, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Shanghai for Science and Technology", + "bbox": [ + 292, + 199, + 705, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Shanghai Urban and Rural Construction and Transportation Development Institute", + "bbox": [ + 171, + 215, + 826, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Guangxi Medical University $^{4}$ University of Glasgow", + "bbox": [ + 266, + 234, + 728, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GLOM, an innovative departure from standard deep learning architectures, has been proposed and gained special concern recently due to its good interpretability in representing part-whole relationships in computer vision. However, GLOM faces challenges in achieving agreement and is usually computationally demanding. First, current implementations struggle to produce identical vectors that reliably converge to represent nodes in a parse tree. Second, GLOM is computationally intensive due to the need to maintain equal resolution across all levels. To address these issues, inspired by contrastive learning, we proposed a contrastive agreement enhancer (CAE), which effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby facilitating forming distinct \"islands.\" Furthermore, we introduce a dissimilarity-focused head $(H_{d})$ to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window. The results of comparison experiments indicate that the proposed approach delicately retains informative content and significantly reduces the number of parameters. Additionally, the ablation experiments and visualization results demonstrate that CAE successfully promotes islands of agreement.", + "bbox": [ + 89, + 318, + 483, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 713, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As deep neural networks continue to achieve impressive performance across various tasks, there is increasing focus on model interpretability, particularly in the domain of computer vision [1]. GLOM was proposed to address this need, offering a promising solution to the challenge of representing part-whole hierarchies in neural networks [10]. It integrates the strengths of transformers, neural fields, contrastive representation learning, distillation, and capsules,", + "bbox": [ + 89, + 739, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd4113201db0508a242d8f805c009ab27f570726652f06670d9a7df69b720134.jpg", + "image_caption": [ + "(a) An illustration for the contrastive agreement enhancer" + ], + "image_footnote": [], + "bbox": [ + 522, + 290, + 671, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c9edbc169916ea4572e74b1f9ba96b04b0222cb321f04c95c298c2c72dfa1b66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 291, + 750, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a226053f987e1aec2e39f1505645162ffff982861fea80827c69d069d9718035.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 291, + 898, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad046ef7d996a2d7beca4644516e0a586e9852dfa99ff8b626b01e737035666a.jpg", + "image_caption": [ + "(b) $TV_{n}$ for bottom level" + ], + "image_footnote": [], + "bbox": [ + 522, + 425, + 696, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad3a0e77b2fe2e42d8426e729a6e19ea3dfdf33be3ef2119e73a03d1c2e85437.jpg", + "image_caption": [ + "(c) $TV_{n}$ for top level", + "Figure 1. a): To tackle the challenge in current GLOM implementations, where vectors struggle to form islands representing parts and wholes and fail to grow larger islands at higher levels, we introduce a contrastive agreement enhancer. This approach effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct islands. By adjusting the threshold for distinguishing positive and negative pairs, we ensure increased smoothness at higher levels that indicate the island grows larger. b) and c): Total variance of normalized embeddings $(TV_{n})$ is used as a metric to quantitatively measure smoothness (with lower $TV_{n}$ indicating greater smoothness). Our method enhances smoothness at both the bottom and top levels, with the top level achieving greater smoothness." + ], + "image_footnote": [], + "bbox": [ + 723, + 424, + 897, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "garnering significant attention upon its introduction. At its core, GLOM uses \"islands\"—groups of similar feature vectors dynamically identified as coherent units, such as parts of an object—to form nodes in a parse tree. However, current implementations [6, 8] of GLOM struggle to produce similar vectors that reliably reach agreement to form these islands. Moreover, GLOM models are computationally intensive due to the need to maintain equal resolution across", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: sonya@usst.edu.cn", + "bbox": [ + 107, + 875, + 341, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24351", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "all levels. These limitations inspire the enhancements proposed in this paper, which seek to address the vector agreement problem and improve computational efficiency.", + "bbox": [ + 89, + 90, + 480, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the GLOM system, embeddings at each level are feature vectors that gradually converge to form distinct islands of nearly identical vectors, representing parts like the head, body, and tail, as shown in the right panel of Fig. 1a. To achieve this, GLOM incorporates a regularizer [6] that minimizes the cosine distance between the bottom-up predictions from lower-level embeddings and the top-down predictions from higher-level embeddings at adjacent levels. However, a practical challenge arises due to the need to balance multiple regularizers between various levels, making it difficult to properly manage their interactions. Another option is a regularizer that encourages similarity between embeddings of nearby locations, but this can lead to the collapse of representations. Although contrastive learning, which commonly uses negative samples to prevent collapse [3], could be a solution, it introduces the challenge of producing distinct islands and recognizing objects, as it relies on finding suitable negative samples to differentiate these islands. Additionally, as the islands grow larger at higher levels to represent whole entities like birds, the embeddings should become more similar, resulting in increased smoothness across the feature map. We use total variance as a metric to quantitatively measure this smoothness. As illustrated by the blue line in Fig. 1b and Fig. 1c, we observe that the total variance [22] of GLOM's normalized embeddings is higher at higher levels than at lower levels, which contradicts GLOM's theoretical expectations.", + "bbox": [ + 91, + 140, + 482, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these two issues, we propose a contrastive agreement enhancer (CAE), as shown in Fig. 1a that computes the cosine similarity between two embeddings and determines whether they belong to a positive or negative pair based on a predefined threshold. This mechanism, inspired by contrastive learning, promotes similarity between positive pairs while pushing apart negative pairs. CAE offers several advantages: 1) it can be easily integrated into every level of GLOM; 2) it is not a regularizer, eliminating the need to balance multiple regularizers across levels; 3) it effectively promotes embedding agreement to form distinct islands; and 4) by adjusting the threshold, it ensures increased smoothness at higher levels.", + "bbox": [ + 89, + 550, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another focus of this paper is mitigating the computational challenges in GLOM, which arise from maintaining equal resolution across all levels. Specifically, when implementing training or downstream tasks, different heads are customized and added to the top of GLOM. If these heads receive the entire top-level embeddings as input, the computational burden becomes significant because the top-level resolution is as high as the bottom level. A straightforward solution is to downsample the embeddings, based on the premise that redundancy exists within it. Redundancy is in-", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "deed present at the top level, as the embeddings there have reached agreement and become highly similar. To handle this, we design a dissimilarity-focused head $(H_{d})$ , which is integrated at the top of GLOM. $H_{d}$ similarly to common pooling methods, where a small rectangular window slides across the input feature map. For each window position, $H_{d}$ calculates the similarity of each embedding with the others within the window. The weights for each embedding are then computed based on the negative correlation of these similarities. The output of $H_{d}$ is the weighted sum of the embeddings. The advantages of $H_{d}$ include significantly reducing the parameters and computational load of the heads. Moreover, compared to max pooling, $H_{d}$ preserves more informative content by focusing on embeddings with lower similarity, which tend to contain more useful information—this is an improvement over max pooling, which just discards all but the highest-value embedding.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, the main contributions of this work can be summarized as follows. 1) A new implementation of GLOM is proposed that aligns more closely with GLOM's theoretical framework for forming hierarchies of grouped embeddings to represent nodes in a parse tree. 2) A novel contrastive agreement enhancer is designed to promote agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct \"islands.\" 3) A dissimilarity-focused head is developed to reduce redundancy in the top-level embeddings, preserving more informative content by concentrating on embeddings with lower similarity, which tend to contain more useful information. 4) Quantitative and qualitative analyses show that CAE effectively promotes islands of agreement and $H_{d}$ contributes greatly to improving model efficiency across datasets.", + "bbox": [ + 511, + 348, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 588, + 651, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Capsule networks and GLOM: GLOM provides a more efficient approach to representing part-whole hierarchies than capsule networks (CapsNets) for several reasons [10]. First, CapsNets often face inefficiencies in part representation due to the need to pre-allocate neurons for possible parts at specific locations, leading to underutilized capsules during inference [29, 30]. In contrast, GLOM leverages all neurons to form clusters, facilitating knowledge sharing across locations and improving part representation efficiency. Second, conventional CapsNets require presetting the number of clusters to represent higher-level parts. For instance, studies like [23] and [11] preset the number of capsule types and slice the feature map along the channel direction to form capsules. GLOM, however, allows embeddings to naturally form islands without the need for predefined clusters. Lastly, GLOM eliminates dynamic routing by enabling each part's location to independently construct its own vector representation of the whole, avoiding the need to route information to specific capsules. How", + "bbox": [ + 511, + 613, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24352", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, GLOM is still in its early stages of development and has its own limitations. For example, it primarily functions as an imagery model, with few implementations available to thoroughly evaluate its performance [6, 8]. Current implementations struggle to achieve proper agreement between embeddings, which hampers the formation of islands in GLOM. Additionally, embeddings are set at the same resolution across both low and high levels, leading to redundancy in the model's trainable parameters. This paper focuses on addressing these shortcomings.", + "bbox": [ + 88, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Routing and embedding updating: Routing algorithms are a crucial aspect of CapsNets, intended to replace the pooling algorithms typically used in CNNs by assessing the agreement between capsules [23]. This agreement reflects how much lower-level capsules support the existence of a higher-level capsule, based on the similarity of their predictions [11]. Many works about CapsNets focus on reducing the high computation burden in the capsule routing algorithm. For instance, [27] eliminates the iterations in routing and uses capsule distribution variance to activate high-level capsules. Additionally, [2, 4, 17, 19, 20] introduce efficient attention mechanisms into routing, significantly reducing computational overhead. In contrast, GLOM eliminates the need for routing but depends on similar embeddings to iteratively update and form 'islands' representing parts or whole objects [10]. Both routing and embedding updates aim to use similar capsules/embeddings to represent nodes in a sparse tree, organizing objects into a part-whole hierarchy. A key focus of this paper is the role of embedding similarity, which informs the design of the proposed CAE and $H_{d}$ , each serving different purposes. CAE improves GLOM's island formation by promoting agreement between similar embeddings and disagreement between dissimilar ones, while $H_{d}$ reduces computational overhead at higher levels by selecting embeddings that are less similar to others, suggesting they may contain more valuable information.", + "bbox": [ + 93, + 243, + 485, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 661, + 183, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed model aims to efficiently implement the GLOM architecture while more closely aligning with its theoretical framework for creating hierarchical islands of embeddings that represent nodes in a parse tree, with these islands expanding at higher levels. Additionally, the model is designed to be lightweight, facilitating broader potential applications of the GLOM architecture.", + "bbox": [ + 89, + 686, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, we start with the introduction of the mathematical notation needed to explain the details of the main components of the architecture. Given an input image $I$ , it is spatially transformed into $N = h \\times w$ patches. The $n$ -th patch corresponds to column $C_n(h,w)$ at the spatial location $(h,w)$ , where $n \\in \\{1,\\dots,N\\}$ . For each column, it consists of $K$ -level embeddings $f_{(h,w)}^{k,t}, k \\in \\{1,\\dots,K\\}$ at", + "bbox": [ + 89, + 792, + 482, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "time step $t \\in \\{1, \\dots, T\\}$ with a size of $d$ for each embedding. The subscript $(h, w)$ and superscript $t$ are omitted in subsequent instances for readability. $f^k$ and $f^{k+1}$ are consecutive and represent the part and whole respectively. They are connected by a bottom-up encoder and a top-down decoder which are shared across the spatial dimension.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Patches embedding", + "text_level": 1, + "bbox": [ + 511, + 189, + 699, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline is demonstrated in Fig. 2. Firstly, input image $I$ , with the shape of $H \\times W \\times c$ , is transformed into patch embeddings. Following the recipe provided in [16], a convolutional Tokenizer is used to extract feature maps with the size of $H' \\times W' \\times c'$ . Next, it is reshaped to the size of $(h \\times r) \\times (w \\times r) \\times c'$ , where $r$ is the patch size, $H' = h \\times r$ , and $W' = w \\times r$ . Then, we reshape and permute it with the size of $h \\times w \\times d$ , where $d = r \\times r \\times c'$ . At last, the $n$ $d$ -dimensional features act as the embeddings $f^{1,t}$ at the bottom level in columns $C_n$ at time step $t$ .", + "bbox": [ + 511, + 210, + 906, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Embedding updating", + "text_level": 1, + "bbox": [ + 511, + 369, + 712, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Secondly, we introduce the propagation phase, which governs the interactions between levels and across locations. As shown in Fig. 3, the embedding of a column at location $n$ and level $k$ is updated by integrating multiple sources of information: predictions from lower and higher levels at the same location $n$ using a bottom-up encoder and a top-down decoder, predictions from embeddings at other locations within the same level via self-attention and a contrastive agreement enhancer, and the historical context (i.e., the embedding from the previous iteration). Next, we elaborate on the components of the model that drive this embedding update process.", + "bbox": [ + 511, + 392, + 906, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bottom-up encoder $\\mathcal{N}_{BU}$ : Taking a low-level embedding $f^{k,t}$ as input, the bottom-up encoder $\\mathcal{N}_{BU}$ is used to predict high-level embedding $f^{k+1,t+1}$ with the same size of $d$ . It is constructed by two fully connected layers and shared across $n$ columns.", + "bbox": [ + 511, + 573, + 905, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Top-down decoder $\\mathcal{N}_{TD}$ : Given a high-level embedding $f^{k + 1,t}$ as input, the top-down decoder $\\mathcal{N}_{TD}$ is applied to predict the corresponding low-level embedding $f^{k,t + 1}$ . The intuition is that if the model understands the whole, it can effectively predict its parts. Like the encoder, $\\mathcal{N}_{TD}$ also comprises two fully connected layers.", + "bbox": [ + 511, + 648, + 905, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-attention $\\mathcal{A}_{SA}$ : In GLOM, the interaction between columns is implemented by a simplified self-attention that query, value, and key vector are all the same as input. The aim of self-attention is to produce islands of identical embeddings at a level by making each embedding vector at that level regress towards other similar vectors at nearby locations. $\\mathcal{A}_{SA}$ is computed as:", + "bbox": [ + 511, + 739, + 905, + 844 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 863, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24353", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a2006584f97e92187f33470dade1fc6b10e63664290beee73c81f33c080f6a0.jpg", + "image_caption": [ + "Figure 2. Our model is pre-trained using contrastive learning, followed by supervised training with frozen weights. The input images $I$ are augmented and transformed into patches. These patches are then processed to construct a part-whole hierarchy over $T$ iterations, where embeddings at different levels gradually converge as the levels ascend. Next, the top-level embeddings are fed into $H_{d}$ , which performs downsampling based on weights that are negatively correlated with similarity. In the subsequent step, the embeddings are sent to $H_{1}$ for contrastive pre-training. Finally, the output embeddings are passed to $H_{2}$ for supervised training with frozen weights." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 903, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q, K, V = f^{k,t}$ . It should be noted that $\\mathcal{A}_{SA}$ alone is insufficient for creating part-whole hierarchies, as it lacks the flexibility to adjust the degree of agreement (e.g., it struggles to encourage embeddings to achieve greater consensus at higher levels). To address this, GLOM incorporates additional regularization. In contrast, our model adopts a more effective strategy, drawing inspiration from contrastive learning, to promote agreement among embeddings.", + "bbox": [ + 88, + 332, + 480, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrastive agreement enhancer $\\mathcal{A}_{CAE}$ : the function of our proposed CAE is two-fold: enhancing the agreement between positive embedding pairs; and pushing apart the negative embedding pairs. In contrastive learning, positive sample pairs are generated by applying different augmentations to the same image, while negative sample pairs are derived from images belonging to different classes. Unlike contrastive learning, in $\\mathcal{A}_{CAE}$ , the classification of embedding pairs as positive or negative is based on the similarity between embeddings. To be specific, the similarity is defined as:", + "bbox": [ + 88, + 469, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\frac {f _ {n 1} f _ {n 2} ^ {T}}{\\| f _ {n 1} \\| \\| f _ {n 2} \\|}, n 1, n 2 \\in \\{1, \\dots , N \\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 640, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the size of similarity map $S$ is $N \\times N$ . For positive and negative embedding pairs, we could set different thresholds. Here we take a simple case for example with $S = 0$ as a threshold for both positive and negative embedding pairs. $S > 0$ means that the angle between embeddings $f_{n1}$ and $f_{n2}$ lies within the range $(0^{\\circ}, 90^{\\circ})$ . In this case, $f_{n1}$ and $f_{n2}$ is a positive embedding pair and could belong to the same island. Conversely, $S < 0$ indicates that $f_{n1}$ and $f_{n2}$ may belong to different islands thereby being a negative pair. This process can be described as:", + "bbox": [ + 89, + 681, + 482, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {p} = S \\cdot \\mathbb {I} (S \\geq 0) + (- \\infty) \\cdot \\mathbb {I} (S < 0), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 840, + 482, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {e} = S \\cdot \\mathbb {I} (S < 0), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 859, + 480, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{p}$ and $S_{e}$ represent the similarity map for positive", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and negative embedding pairs respectively. $\\mathbb{I}$ is the indicator function. In the model's bottom level, the threshold is set to 0, while at the top level it is set to $-0.2$ . Next, we compute the embeddings at the time step $t + 1$ :", + "bbox": [ + 511, + 333, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {p} ^ {t + 1} = \\operatorname {s o f t m a x} \\left(S _ {p}\\right) f ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 401, + 903, + 420 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e} ^ {t + 1} = S _ {e} f ^ {t}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 422, + 903, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} _ {C A E} \\left(f ^ {t}\\right) = f _ {p} ^ {t + 1} - f _ {e} ^ {t + 1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 441, + 903, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{p}^{t + 1}$ are the positive embeddings and $f_{e}^{t + 1}$ are the negative embeddings. Compared with $f_{e}^{t + 1}$ , the computation of $f_{p}^{t + 1}$ evolves the softmax function since the main purpose is to facilitate agreement between embeddings and softmax is good at eliminating the effect of negative embeddings ( $S < 0$ ). In addition, to push apart between positive and negative embeddings, we minus $f_{p}^{t + 1}$ by $f_{e}^{t + 1}$ to predict the $f^{t + 1}$ . An illustration for CAE is given in Fig. 1a.", + "bbox": [ + 511, + 467, + 905, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusion: At each time step $t$ , $f^{k,t}$ is computed by the fusion of predictions from $\\mathcal{N}_{BU}, \\mathcal{N}_{TD}, \\mathcal{A}_{SA}$ , and $\\mathcal{A}_{CAE}$ :", + "bbox": [ + 511, + 589, + 905, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {k, t} = w _ {b u} \\mathcal {N} _ {B U} \\left(f ^ {k - 1, t - 1}\\right) + w _ {t d} \\mathcal {N} _ {T D} \\left(f ^ {k + 1, t - 1}\\right) +\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 641, + 877, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {l a} f ^ {k, t - 1} + w _ {a s} \\mathcal {A} _ {S A} \\left(f ^ {k, t - 1}\\right) + w _ {a c} \\mathcal {A} _ {C A E} \\left(f ^ {k, t - 1}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 662, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{bu}, w_{td}, w_{la}, w_{as}, w_{ac}$ are trainable parameters.", + "bbox": [ + 511, + 689, + 880, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dissimilarity-focused head", + "text_level": 1, + "bbox": [ + 511, + 712, + 754, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the aim to relieve the computation overhead introduced by which the embeddings at the top level and the lower level have the same resolution, we proposed a dissimilarity-focused head $H_{d}$ which receives top-level embeddings as input to do downsampling as shown in Fig. 4, and it greatly reduces the number of parameters.", + "bbox": [ + 511, + 734, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$H_{d}$ is inspired by the idea that embeddings at the top level should exhibit agreement, which can result in redundancy between them. To quantify this redundancy, we apply indirect thinking and consider similarity as a suitable metric, since the more similar the embeddings are, the more", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24354", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7ecde751620dfe32aa975dd6a3b8e400ee86c330de3e3a885392efdbf628b949.jpg", + "image_caption": [ + "Figure 3. Architecture of model. The column at location $n$ is updating with a fusion of bottom-up and top-down prediction, history, SA, and CAE." + ], + "image_footnote": [], + "bbox": [ + 109, + 104, + 475, + 380 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cefc780b066a9068d8af1a13d2da4604e5de505f7ee117018bf2aa43786ffe48.jpg", + "image_caption": [ + "Figure 4. An illustration for the dissimilarity-focused head which is designed to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window." + ], + "image_footnote": [], + "bbox": [ + 99, + 469, + 477, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "redundant they become. To be specific, given the top embeddings $f^{K} \\in \\mathbb{R}^{h \\times w \\times d}$ , we first transform them into $p \\times p$ patches each with the shape of $h' \\times w' \\times d$ where $h' = \\frac{h}{p}, w' = \\frac{w}{p}$ . For each patch, similarity map $S'$ with the shape of $(h' \\times w') \\times (h' \\times w')$ is computed using Eq. 2. Next, we compute the weights $W$ for each embedding within the patch:", + "bbox": [ + 89, + 678, + 482, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} = \\sum_ {j = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} e ^ {1 - S _ {i, j} ^ {\\prime}}, i, j \\in \\{1, \\dots , h ^ {\\prime} \\times w ^ {\\prime} \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 796, + 434, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW = \\frac {W}{\\operatorname* {m a x} \\left(W _ {i}\\right)}, W \\in \\mathbb {R} _ {> 0} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 842, + 480, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $1 - S_{i,j}^{\\prime}$ means that embeddings with higher simi", + "bbox": [ + 89, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "larity with others should be assigned with lower weight to downsample. Besides, the weights are summed to measure the overall weight for embedding $f_{i}$ . Finally, the downsampled output is computed by the weighted sum within each patch.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {d} \\left(f ^ {K}\\right) = \\sum_ {i = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} W _ {i} f _ {i}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 174, + 905, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{d}(f^{K})\\in \\mathbb{R}^{p\\times p\\times d}$", + "bbox": [ + 511, + 224, + 696, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 511, + 250, + 617, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training of our model follows the setup in [8], consisting of two steps: 1) a pre-training phase using a supervised contrastive loss function, and 2) a classification training phase using cross-entropy loss.", + "bbox": [ + 511, + 272, + 905, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pre-training, various data augmentation is adopted to each image $I$ within a batch $B$ , producing pairs of image $I_{a}$ and $I_{b}, a, b \\in \\{1, \\dots, 2B\\}$ . Then, the image pairs are fed to our model to produce representation $C_{n,a}$ and $C_{n,b} \\in \\mathbb{R}^d$ , $n \\in \\{1, \\dots, N\\}$ at the top level $K$ . In the following, we reshape them to the shape of $N \\times d$ and send them to a contrastive head $H_{1}$ which is constructed with two fully connected layers. The output is denoted as $O_{a}$ and $O_{b}$ . Finally, a supervised contrastive loss is adopted to pre-train the model:", + "bbox": [ + 511, + 333, + 905, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {1} = - \\log \\frac {e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {b}\\right)}}{\\sum_ {i = 1} ^ {2 B} \\mathbb {I} _ {[ i \\neq a ]} e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {i}\\right)}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 488, + 903, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{I}_{[i\\neq a]}$ is an indicator function that values 0 if $i$ and $a$ belong to the same image class, and 0 otherwise. sim is a cosine similarity function that measures the similarity of normalized input.", + "bbox": [ + 511, + 534, + 905, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the model is pre-trained with contrastive loss, we freeze the network weights. Besides, we add a classification head $H_{2}$ on the top of contrastive head $H_{1}$ . The $H_{2}$ is a linear layer that projects the representation to the final out with the dimension of $C$ . $C$ is the number of classes. The cross-entropy loss is given as:", + "bbox": [ + 511, + 595, + 905, + 685 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {2} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} p _ {i} \\log q _ {i}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 693, + 903, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p_i$ is the true probability distribution for class $i$ and $q_i$ is the predicted probability for class $i$ .", + "bbox": [ + 511, + 741, + 905, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 784, + 637, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments encompass quantitative and qualitative analyses. Quantitatively, we evaluate the formation of islands of agreement using the total variance of normalized embeddings and compare model performance across multiple datasets. An ablation study further examines the contributions of the proposed CAE and $H_{d}$ . For the qualitative", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24355", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2465713049a8580e2e580014a00e072363a4ab0df0165c2d5af07ec9f9ec78c6.jpg", + "image_caption": [ + "(a) C-10 bottom level" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 243, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04ff742b0ddf4cc998367916a00cafd2c2cdb91f16a66c98f9fccf43801614a1.jpg", + "image_caption": [ + "(f) C-10 top level" + ], + "image_footnote": [], + "bbox": [ + 96, + 204, + 243, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2b8f22f3870a2b2e2bf2382ddc70c3c622d6606ed5f31558cf022e59be48cdd.jpg", + "image_caption": [ + "(b) C-100 bottom level" + ], + "image_footnote": [], + "bbox": [ + 261, + 90, + 408, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c587e6392be397c0e822a8dc17d2e7863396a6970310c003ffb98167fcfba4d.jpg", + "image_caption": [ + "(g) C-100 top level" + ], + "image_footnote": [], + "bbox": [ + 261, + 204, + 408, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9134948e8d7e40f4e698c04dc9cb484325eb972329d294a29e64e4eead927ef0.jpg", + "image_caption": [ + "(c) S-NORB bottom level" + ], + "image_footnote": [], + "bbox": [ + 424, + 90, + 571, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/779546a59c5ba37e82e7e54a77847b268bc543898d2f59f229ffd5a667169e66.jpg", + "image_caption": [ + "(h) S-NORB top level", + "Figure 5. The comparison of our model with that of the Agglomerator in terms of $TV_{n}$ at both the bottom and top levels across five datasets. It is evident that the total variance of our model is considerably lower than that of the Agglomerator, indicating that the contrastive attention mechanism effectively improves agreement in representing parts, which is essential for the theory of GLOM. Furthermore, the $TV_{n}$ of our model is greater at the bottom level than at the top level, suggesting that agreement tends to increase as the levels rise." + ], + "image_footnote": [], + "bbox": [ + 424, + 204, + 571, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f011562337a8388a3b410821b75d9bcdce71b696d0a3cbb8ad0c7bc89c40150e.jpg", + "image_caption": [ + "(d) MNIST bottom level" + ], + "image_footnote": [], + "bbox": [ + 589, + 90, + 736, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c058d9ba13e4f98bd6e944a8b31f8f100dbad0143b6a465cbae62cb37cafb810.jpg", + "image_caption": [ + "(i) MNIST top level" + ], + "image_footnote": [], + "bbox": [ + 589, + 204, + 736, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42ec8d94e61af33e2e41839e04789b27dd86d46e26a49aa9a87fde3c7996ba42.jpg", + "image_caption": [ + "(e) F-MNIST bottom level" + ], + "image_footnote": [], + "bbox": [ + 753, + 90, + 900, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/547f56353964ef660722ebbf5a75be6da6dd07c30e1768035e5e7012bbbd09db.jpg", + "image_caption": [ + "(j) F-MNIST top level" + ], + "image_footnote": [], + "bbox": [ + 753, + 204, + 898, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "analysis, we visualize the vector representations of emerging islands across levels. To illustrate CAE's enhancement of agreement intuitively, we depict representations with arrows spanning various locations and levels. Experiments are conducted on the CIFAR-10, CIFAR-100 [13], MNIST [14], FashionMNIST [26], and SmallNORB datasets [15]. The details of the training settings are provided in the supplementary material. The source code is publicly available at: https://github.com/zengru001usst/GLOM", + "bbox": [ + 89, + 407, + 483, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Agreement Evaluation", + "text_level": 1, + "bbox": [ + 89, + 551, + 302, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the effectiveness of the proposed CAE, total variance [22] of the normalized embeddings $(TV_{n})$ , as shown in Eq. 13, is employed as a metric to quantitatively measure the smoothness of embeddings at specific levels within the models. The embeddings are normalized, as a trivial solution may occur where total variance is minimized by making every $f_{i,j}$ infinitesimal. Lower $TV_{n}$ indicates greater smoothness, meaning that more embeddings are similar and achieve a higher level of agreement.", + "bbox": [ + 89, + 573, + 483, + 709 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nT V _ {n} = \\frac {1}{N} \\frac {1}{d} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {d} \\left(f _ {i, j} ^ {\\prime} - \\mu_ {j}\\right) ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 729, + 400, + 773 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} f _ {i, j} ^ {\\prime},\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 776, + 292, + 818 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i, j} ^ {\\prime} = \\frac {f _ {i , j}}{\\sqrt {\\sum_ {j} f _ {i , j} ^ {2}}}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 820, + 482, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results presented in Fig. 5 compare the $TV_{n}$ of our model with that of the Agglomerator at both the bottom and", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "top levels across five datasets. It can be observed that the $TV_{n}$ of our model is significantly lower than that of the Agglomerator, indicating that the CAE effectively enhances agreement in representing parts, which is crucial for the theory of GLOM. Additionally, the $TV_{n}$ of our model is higher at the bottom level compared to the top level, suggesting that agreement increases as the levels ascend. This is consistent with GLOM's framework, where embeddings at lower levels gradually cluster locally to represent parts, while embeddings at higher levels cluster globally to represent the whole. In contrast, the Agglomerator fails to maintain this characteristic.", + "bbox": [ + 511, + 407, + 906, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Classification Results", + "text_level": 1, + "bbox": [ + 511, + 604, + 714, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the efficiency of the proposed model, we report the classification results for each dataset in terms of error percentage and the number of trainable parameters, as shown in Tab. 1. While Capsule Networks (CapsNets), such as [11, 18, 19, 21, 23, 28], perform well on simpler datasets like smallNORB, MNIST, and FashionMNIST, they struggle to scale to more complex datasets with a higher number of classes, such as CIFAR-100, due to inefficiencies in object representation [10]. Convolutional models, such as [9, 24], maintain strong performance across various datasets but fall short in interpretability. Transformer-based and MLP-based methods achieve state-of-the-art performance on more complex datasets, but they lack testing on smaller datasets. As shown in Tab. 1, our model achieves comparable classification error percentages to CapsNets and Agglomerator [8] on simpler datasets, but with significantly fewer parameters, especially compared to Agglomerator. Moreover, our model requires less training time and has a", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24356", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5626814249904f8914640cb98c35e153515adabe102c08999d7cac7e327e760e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRefBackboneErrors %No. params (Millions)Training Arch.
S-NORBMNISTF-MNISTC-10C-100
E-CapsNet[19]Caps2.540.26---0.2GPU
CapsNet[23]2.700.256.3810.682.006.8GPU
Matrix-CapsNet[11]1.400.446.1411.9-0.3GPU
Capsule VB[21]1.600.305.2011.2-0.2GPU
Res-CapsNet[18]0.550.55-7.62-16.4GPU
IAR-CapsNet[28]1.360.77---0.18GPU
ResNet-110[9]Conv-2.105.106.41*27.76*1.7GPU
VGG[24]-0.326.507.74*28.05*20GPU
ViT-L/16[7]Transf---0.85*6.75*632TPU
ConvMLP-L[16]Conv/MLP---1.40*11.40*43TPU
MLP-Mixer-L/16[25]MLP---1.66*-207TPU
Agglomerator[8]Conv/MLP-/Caps0.010.307.4311.1540.9772GPU
Ours-0.010.305.1010.9041.312GPU
", + "bbox": [ + 91, + 88, + 901, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c83b158bd7bb6efc115640c256083f58d477a983ef70f1e384ff37215c80e062.jpg", + "table_caption": [ + "Table 1. Classification results on SmallNORB (S-NORB), MNIST, FashionMNIST (F-MNIST), CIFAR-10 (C-10), and CIFAR-100 (C-100) datasets, demonstrating the efficiency of the proposed model. Results marked with * indicate networks pre-trained on ImageNet." + ], + "table_footnote": [], + "table_body": "
HdCAEErrors (%)No. Param. (Millions)
Baseline××9.172
Model1×√9.072
Model2√×11.62
Model3√√10.92
", + "bbox": [ + 132, + 388, + 444, + 493 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7434e62d9ccdbd13958ce1cf6ba2b49abc5cb552703372c12c2309de0ff6224f.jpg", + "table_caption": [ + "Table 2. Ablation for dissimilarity-focused head $(H_{d})$ and contrastive agreement enhancer (CAE) on CIFAR10." + ], + "table_footnote": [], + "table_body": "
Errors (%)TVnNo. Param. (Millions)
Average pooling11.10.3722
Max pooling11.10.4732
Hd10.90.5442
", + "bbox": [ + 122, + 546, + 452, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of dissimilarity-focused head $(H_{d})$ , average pooling and maxpooling on CIFAR10.", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "much smaller architecture compared to most transformer-based and MLP-based methods.", + "bbox": [ + 89, + 700, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Ablation", + "text_level": 1, + "bbox": [ + 89, + 741, + 294, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To study the contribution of the different components of our model to its performance, we conduct an ablation experiment where we incrementally added the dissimilarity-focused head $(H_{d})$ and contrastive agreement enhancer (CAE) to the baseline model. The baseline model is defined by removing both $H_{d}$ and CAE from our full model. We then compare the classification results and the number of parameters on CIFAR10 to demonstrate the contribution of $H_{d}$ and CAE to model performance.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Tab. 2 indicate that $H_{d}$ can significantly reduce the number of trainable parameters in the model while only causing a slight increase in errors, which suggests the presence of redundancy at the top level. Additionally, when $H_{d}$ is employed, CAM can enhance model performance, as much of the redundancy has been eliminated, allowing CAM to provide valuable supplementary information.", + "bbox": [ + 511, + 392, + 906, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the impact of $H_{d}$ on classification accuracy and $TV_{n}$ , we compare our model against two alternative downsampling strategies: max pooling and average pooling, by substituting $H_{d}$ with these methods. Results in Tab. 3 show that the model with $H_{d}$ achieves a lower error rate and higher $TV_{n}$ than these alternatives under identical downsampling conditions. The higher $TV_{n}$ with $H_{d}$ stems from its emphasis on embeddings with lower similarity, which carry more informative features and enhance classification accuracy.", + "bbox": [ + 511, + 500, + 908, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Ablation", + "text_level": 1, + "bbox": [ + 511, + 664, + 707, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct two qualitative ablation study to demonstrate the effect of the contrastive agreement enhancer.", + "bbox": [ + 511, + 686, + 905, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The emerging island of agreement. To illustrate the emergence of islands across different levels, the embeddings from various $k$ levels of CIFAR-10 are visualized in Fig. 6. When examining the embeddings at the highest level ( $k = 5$ ), those generated by the model with CAE reveal finer details and a more distinct separation between objects and the background. Furthermore, a comparison of embeddings across different levels demonstrates that as the levels increase, the embeddings progressively reach agreement, effectively capturing part-whole hierarchies.", + "bbox": [ + 511, + 718, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The representation visualization using arrows. To provide an intuitive illustration of the enhancement in", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24357", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1cd22bcf16e1eb6cf283642755a31048916e79c227343e383237bf20f3395d3.jpg", + "image_caption": [ + "Figure 6. The emerging island of agreement at various $k$ levels from CIFAR-10. At the top level $(k = 5)$ , the representations generated by the model with CAE display enhanced detail, highlighting a more distinct separation between objects and the background." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 897, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/990584779435dc1e70190d7aef758a33c15837f4712c84560f3234a7e0e55f2b.jpg", + "image_caption": [ + "Figure 7. The representation visualization using arrows for model with and without CAE. The direction and length of each arrow represent the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator." + ], + "image_footnote": [], + "bbox": [ + 106, + 422, + 897, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "agreement introduced by CAE, we visualize the representations using arrows across various locations and levels. The direction and length of each arrow indicate the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator. MNIST is selected for visualization due to its simplicity, minimizing potential interference from complex image backgrounds. As depicted in Fig. 7, the arrows in the model incorporating CAE demonstrate greater consistency, whereas those in the model without CAE appear more disordered and inconsistent.", + "bbox": [ + 89, + 646, + 482, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 828, + 209, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a new implementation of GLOM that aligns more closely with its theoretical framework for forming hierarchies of grouped embeddings to rep-", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "resent nodes in a parse tree. To foster the formation of distinct islands, we have designed a novel contrastive agreement enhancer that encourages agreement between positive embedding pairs while separating negative pairs. To reduce the computational burden of GLOM, we developed a dissimilarity-focused head that preserves more informative content by prioritizing embeddings with lower similarity, which often contain richer information. Both quantitative and qualitative analyses demonstrate that the contrastive agreement enhancer effectively promotes islands of agreement, while the dissimilarity-focused head enhances model efficiency across various datasets. This work significantly improves the practicality of GLOM, and we hope it inspires other researchers to further explore this direction to advance its development.", + "bbox": [ + 511, + 646, + 906, + 872 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24358", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. Publisher Copyright: © 2021 Owner/Author. 1", + "[2] Ran Chen, Hao Shen, Zhong-Qiu Zhao, Yi Yang, and Zhao Zhang. Global routing between capsules. Pattern Recognition, 148:110142, 2024. 3", + "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[4] Jaewoong Choi, Hyun Seo, Suii Im, and Myungjoo Kang. Attention routing between capsules. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 3", + "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1", + "[6] Laura Culp, Sara Sabour, and Geoffrey E Hinton. Testing glom's ability to infer wholes from ambiguous parts. arXiv preprint arXiv:2211.16564, 2022. 1, 2, 3", + "[7] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7", + "[8] Nicola Garau, Niccolò Bisagno, Zeno Sambugaro, and Nicola Conc. Interpretable part-whole hierarchies and conceptual-semantic relationships in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13689-13698, 2022. 1, 3, 5, 6, 7", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7", + "[10] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. Neural Computation, 35(3):413-452, 2023. 1, 2, 3, 6", + "[11] Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing. In International conference on learning representations, 2018. 2, 3, 6, 7", + "[12] Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933. 1", + "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[15] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-104. IEEE, 2004. 6" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jiachen Li, Ali Hassani, Steven Walton, and Humphrey Shi. Convmlp: Hierarchical convolutional mlp's for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6307-6316, 2023. 3, 7", + "[17] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3688-3704, 2022. 3", + "[18] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, pages 1-14, 2024. 6, 7", + "[19] Vittorio Mazzia, Francesco Salvetti, and Marcello Chi-aberge. Efficient-capsnet: Capsule network with self-attention routing. Scientific reports, 11(1):14634, 2021. 3, 6, 7", + "[20] Rita Pucci, Christian Micheloni, and Niki Martinel. Self-attention agreement among capsules. In Proceedings of the iee//cvf international conference on computer vision, pages 272-280, 2021. 3", + "[21] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3749-3756, 2020. 6, 7", + "[22] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 2, 6", + "[23] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 2, 3, 6, 7", + "[24] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6, 7", + "[25] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021. 7", + "[26] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 6", + "[27] Ru Zeng and Yan Song. A fast routing capsule network with improved dense blocks. IEEE Transactions on Industrial Informatics, 18(7):4383-4392, 2021. 3", + "[28] Ru Zeng, Yuzhang Qin, and Yan Song. A non-iterative capsule network with interdependent agreement routing. Expert Systems with Applications, 238:122284, 2024. 6, 7", + "[29] Ru Zeng, Yan Song, and Yuzhang Qin. Spatial attention-based capsule networks with guaranteed group equivariance. IEEE Transactions on Automation Science and Engineering, 2024. 2", + "[30] Ru Zeng, Yan Song, and Yanjiu Zhong. An interpretable unsupervised capsule network via comprehensive contrastive learning and two-stage training. Pattern Recognition, 158: 111059, 2025. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24359", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_model.json b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e4dd866cb7692a6647f2baf360436bdd1cc3cd7 --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_model.json @@ -0,0 +1,2022 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "Agreement aware and dissimilarity oriented GLOM" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.179, + 0.751, + 0.2 + ], + "angle": 0, + "content": "\\(\\mathrm{RuZeng^{1}}\\) Yan Song\\(^{1*}\\) Yang Zhang\\(^{2}\\) Yanling Hu\\(^{3}\\) Hui Yu\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.2, + 0.707, + 0.218 + ], + "angle": 0, + "content": "1University of Shanghai for Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.172, + 0.217, + 0.828, + 0.235 + ], + "angle": 0, + "content": "2Shanghai Urban and Rural Construction and Transportation Development Institute" + }, + { + "type": "text", + "bbox": [ + 0.267, + 0.235, + 0.729, + 0.253 + ], + "angle": 0, + "content": "\\(^{3}\\)Guangxi Medical University \\(^{4}\\)University of Glasgow" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.683 + ], + "angle": 0, + "content": "GLOM, an innovative departure from standard deep learning architectures, has been proposed and gained special concern recently due to its good interpretability in representing part-whole relationships in computer vision. However, GLOM faces challenges in achieving agreement and is usually computationally demanding. First, current implementations struggle to produce identical vectors that reliably converge to represent nodes in a parse tree. Second, GLOM is computationally intensive due to the need to maintain equal resolution across all levels. To address these issues, inspired by contrastive learning, we proposed a contrastive agreement enhancer (CAE), which effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby facilitating forming distinct \"islands.\" Furthermore, we introduce a dissimilarity-focused head \\((H_{d})\\) to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window. The results of comparison experiments indicate that the proposed approach delicately retains informative content and significantly reduces the number of parameters. Additionally, the ablation experiments and visualization results demonstrate that CAE successfully promotes islands of agreement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.714, + 0.222, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.862 + ], + "angle": 0, + "content": "As deep neural networks continue to achieve impressive performance across various tasks, there is increasing focus on model interpretability, particularly in the domain of computer vision [1]. GLOM was proposed to address this need, offering a promising solution to the challenge of representing part-whole hierarchies in neural networks [10]. It integrates the strengths of transformers, neural fields, contrastive representation learning, distillation, and capsules," + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.291, + 0.673, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.292, + 0.75, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.292, + 0.899, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.408, + 0.862, + 0.42 + ], + "angle": 0, + "content": "(a) An illustration for the contrastive agreement enhancer" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.426, + 0.697, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.539, + 0.678, + 0.552 + ], + "angle": 0, + "content": "(b) \\( TV_{n} \\) for bottom level" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.425, + 0.898, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.539, + 0.868, + 0.552 + ], + "angle": 0, + "content": "(c) \\( TV_{n} \\) for top level" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Figure 1. a): To tackle the challenge in current GLOM implementations, where vectors struggle to form islands representing parts and wholes and fail to grow larger islands at higher levels, we introduce a contrastive agreement enhancer. This approach effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct islands. By adjusting the threshold for distinguishing positive and negative pairs, we ensure increased smoothness at higher levels that indicate the island grows larger. b) and c): Total variance of normalized embeddings \\((TV_{n})\\) is used as a metric to quantitatively measure smoothness (with lower \\(TV_{n}\\) indicating greater smoothness). Our method enhances smoothness at both the bottom and top levels, with the top level achieving greater smoothness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "garnering significant attention upon its introduction. At its core, GLOM uses \"islands\"—groups of similar feature vectors dynamically identified as coherent units, such as parts of an object—to form nodes in a parse tree. However, current implementations [6, 8] of GLOM struggle to produce similar vectors that reliably reach agreement to form these islands. Moreover, GLOM models are computationally intensive due to the need to maintain equal resolution across" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.876, + 0.342, + 0.889 + ], + "angle": 0, + "content": "*Corresponding author: sonya@usst.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24351" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.138 + ], + "angle": 0, + "content": "all levels. These limitations inspire the enhancements proposed in this paper, which seek to address the vector agreement problem and improve computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.141, + 0.483, + 0.547 + ], + "angle": 0, + "content": "In the GLOM system, embeddings at each level are feature vectors that gradually converge to form distinct islands of nearly identical vectors, representing parts like the head, body, and tail, as shown in the right panel of Fig. 1a. To achieve this, GLOM incorporates a regularizer [6] that minimizes the cosine distance between the bottom-up predictions from lower-level embeddings and the top-down predictions from higher-level embeddings at adjacent levels. However, a practical challenge arises due to the need to balance multiple regularizers between various levels, making it difficult to properly manage their interactions. Another option is a regularizer that encourages similarity between embeddings of nearby locations, but this can lead to the collapse of representations. Although contrastive learning, which commonly uses negative samples to prevent collapse [3], could be a solution, it introduces the challenge of producing distinct islands and recognizing objects, as it relies on finding suitable negative samples to differentiate these islands. Additionally, as the islands grow larger at higher levels to represent whole entities like birds, the embeddings should become more similar, resulting in increased smoothness across the feature map. We use total variance as a metric to quantitatively measure this smoothness. As illustrated by the blue line in Fig. 1b and Fig. 1c, we observe that the total variance [22] of GLOM's normalized embeddings is higher at higher levels than at lower levels, which contradicts GLOM's theoretical expectations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.746 + ], + "angle": 0, + "content": "To address these two issues, we propose a contrastive agreement enhancer (CAE), as shown in Fig. 1a that computes the cosine similarity between two embeddings and determines whether they belong to a positive or negative pair based on a predefined threshold. This mechanism, inspired by contrastive learning, promotes similarity between positive pairs while pushing apart negative pairs. CAE offers several advantages: 1) it can be easily integrated into every level of GLOM; 2) it is not a regularizer, eliminating the need to balance multiple regularizers across levels; 3) it effectively promotes embedding agreement to form distinct islands; and 4) by adjusting the threshold, it ensures increased smoothness at higher levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Another focus of this paper is mitigating the computational challenges in GLOM, which arise from maintaining equal resolution across all levels. Specifically, when implementing training or downstream tasks, different heads are customized and added to the top of GLOM. If these heads receive the entire top-level embeddings as input, the computational burden becomes significant because the top-level resolution is as high as the bottom level. A straightforward solution is to downsample the embeddings, based on the premise that redundancy exists within it. Redundancy is in-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "deed present at the top level, as the embeddings there have reached agreement and become highly similar. To handle this, we design a dissimilarity-focused head \\((H_{d})\\), which is integrated at the top of GLOM. \\(H_{d}\\) similarly to common pooling methods, where a small rectangular window slides across the input feature map. For each window position, \\(H_{d}\\) calculates the similarity of each embedding with the others within the window. The weights for each embedding are then computed based on the negative correlation of these similarities. The output of \\(H_{d}\\) is the weighted sum of the embeddings. The advantages of \\(H_{d}\\) include significantly reducing the parameters and computational load of the heads. Moreover, compared to max pooling, \\(H_{d}\\) preserves more informative content by focusing on embeddings with lower similarity, which tend to contain more useful information—this is an improvement over max pooling, which just discards all but the highest-value embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Overall, the main contributions of this work can be summarized as follows. 1) A new implementation of GLOM is proposed that aligns more closely with GLOM's theoretical framework for forming hierarchies of grouped embeddings to represent nodes in a parse tree. 2) A novel contrastive agreement enhancer is designed to promote agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct \"islands.\" 3) A dissimilarity-focused head is developed to reduce redundancy in the top-level embeddings, preserving more informative content by concentrating on embeddings with lower similarity, which tend to contain more useful information. 4) Quantitative and qualitative analyses show that CAE effectively promotes islands of agreement and \\( H_{d} \\) contributes greatly to improving model efficiency across datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.589, + 0.652, + 0.605 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Capsule networks and GLOM: GLOM provides a more efficient approach to representing part-whole hierarchies than capsule networks (CapsNets) for several reasons [10]. First, CapsNets often face inefficiencies in part representation due to the need to pre-allocate neurons for possible parts at specific locations, leading to underutilized capsules during inference [29, 30]. In contrast, GLOM leverages all neurons to form clusters, facilitating knowledge sharing across locations and improving part representation efficiency. Second, conventional CapsNets require presetting the number of clusters to represent higher-level parts. For instance, studies like [23] and [11] preset the number of capsule types and slice the feature map along the channel direction to form capsules. GLOM, however, allows embeddings to naturally form islands without the need for predefined clusters. Lastly, GLOM eliminates dynamic routing by enabling each part's location to independently construct its own vector representation of the whole, avoiding the need to route information to specific capsules. How" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24352" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.242 + ], + "angle": 0, + "content": "ever, GLOM is still in its early stages of development and has its own limitations. For example, it primarily functions as an imagery model, with few implementations available to thoroughly evaluate its performance [6, 8]. Current implementations struggle to achieve proper agreement between embeddings, which hampers the formation of islands in GLOM. Additionally, embeddings are set at the same resolution across both low and high levels, leading to redundancy in the model's trainable parameters. This paper focuses on addressing these shortcomings." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.244, + 0.486, + 0.65 + ], + "angle": 0, + "content": "Routing and embedding updating: Routing algorithms are a crucial aspect of CapsNets, intended to replace the pooling algorithms typically used in CNNs by assessing the agreement between capsules [23]. This agreement reflects how much lower-level capsules support the existence of a higher-level capsule, based on the similarity of their predictions [11]. Many works about CapsNets focus on reducing the high computation burden in the capsule routing algorithm. For instance, [27] eliminates the iterations in routing and uses capsule distribution variance to activate high-level capsules. Additionally, [2, 4, 17, 19, 20] introduce efficient attention mechanisms into routing, significantly reducing computational overhead. In contrast, GLOM eliminates the need for routing but depends on similar embeddings to iteratively update and form 'islands' representing parts or whole objects [10]. Both routing and embedding updates aim to use similar capsules/embeddings to represent nodes in a sparse tree, organizing objects into a part-whole hierarchy. A key focus of this paper is the role of embedding similarity, which informs the design of the proposed CAE and \\( H_{d} \\), each serving different purposes. CAE improves GLOM's island formation by promoting agreement between similar embeddings and disagreement between dissimilar ones, while \\( H_{d} \\) reduces computational overhead at higher levels by selecting embeddings that are less similar to others, suggesting they may contain more valuable information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.662, + 0.184, + 0.679 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Our proposed model aims to efficiently implement the GLOM architecture while more closely aligning with its theoretical framework for creating hierarchical islands of embeddings that represent nodes in a parse tree, with these islands expanding at higher levels. Additionally, the model is designed to be lightweight, facilitating broader potential applications of the GLOM architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.906 + ], + "angle": 0, + "content": "Here, we start with the introduction of the mathematical notation needed to explain the details of the main components of the architecture. Given an input image \\(I\\), it is spatially transformed into \\(N = h \\times w\\) patches. The \\(n\\)-th patch corresponds to column \\(C_n(h,w)\\) at the spatial location \\((h,w)\\), where \\(n \\in \\{1,\\dots,N\\}\\). For each column, it consists of \\(K\\)-level embeddings \\(f_{(h,w)}^{k,t}, k \\in \\{1,\\dots,K\\}\\) at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "time step \\( t \\in \\{1, \\dots, T\\} \\) with a size of \\( d \\) for each embedding. The subscript \\( (h, w) \\) and superscript \\( t \\) are omitted in subsequent instances for readability. \\( f^k \\) and \\( f^{k+1} \\) are consecutive and represent the part and whole respectively. They are connected by a bottom-up encoder and a top-down decoder which are shared across the spatial dimension." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.19, + 0.7, + 0.206 + ], + "angle": 0, + "content": "3.1. Patches embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.907, + 0.364 + ], + "angle": 0, + "content": "The pipeline is demonstrated in Fig. 2. Firstly, input image \\(I\\), with the shape of \\(H \\times W \\times c\\), is transformed into patch embeddings. Following the recipe provided in [16], a convolutional Tokenizer is used to extract feature maps with the size of \\(H' \\times W' \\times c'\\). Next, it is reshaped to the size of \\((h \\times r) \\times (w \\times r) \\times c'\\), where \\(r\\) is the patch size, \\(H' = h \\times r\\), and \\(W' = w \\times r\\). Then, we reshape and permute it with the size of \\(h \\times w \\times d\\), where \\(d = r \\times r \\times c'\\). At last, the \\(n\\) \\(d\\)-dimensional features act as the embeddings \\(f^{1,t}\\) at the bottom level in columns \\(C_n\\) at time step \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.714, + 0.387 + ], + "angle": 0, + "content": "3.2. Embedding updating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Secondly, we introduce the propagation phase, which governs the interactions between levels and across locations. As shown in Fig. 3, the embedding of a column at location \\( n \\) and level \\( k \\) is updated by integrating multiple sources of information: predictions from lower and higher levels at the same location \\( n \\) using a bottom-up encoder and a top-down decoder, predictions from embeddings at other locations within the same level via self-attention and a contrastive agreement enhancer, and the historical context (i.e., the embedding from the previous iteration). Next, we elaborate on the components of the model that drive this embedding update process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.906, + 0.648 + ], + "angle": 0, + "content": "Bottom-up encoder \\(\\mathcal{N}_{BU}\\): Taking a low-level embedding \\(f^{k,t}\\) as input, the bottom-up encoder \\(\\mathcal{N}_{BU}\\) is used to predict high-level embedding \\(f^{k+1,t+1}\\) with the same size of \\(d\\). It is constructed by two fully connected layers and shared across \\(n\\) columns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.739 + ], + "angle": 0, + "content": "Top-down decoder \\(\\mathcal{N}_{TD}\\): Given a high-level embedding \\(f^{k + 1,t}\\) as input, the top-down decoder \\(\\mathcal{N}_{TD}\\) is applied to predict the corresponding low-level embedding \\(f^{k,t + 1}\\). The intuition is that if the model understands the whole, it can effectively predict its parts. Like the encoder, \\(\\mathcal{N}_{TD}\\) also comprises two fully connected layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.74, + 0.906, + 0.845 + ], + "angle": 0, + "content": "Self-attention \\(\\mathcal{A}_{SA}\\): In GLOM, the interaction between columns is implemented by a simplified self-attention that query, value, and key vector are all the same as input. The aim of self-attention is to produce islands of identical embeddings at a level by making each embedding vector at that level regress towards other similar vectors at nearby locations. \\(\\mathcal{A}_{SA}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.864, + 0.907, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24353" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.238, + 0.908, + 0.309 + ], + "angle": 0, + "content": "Figure 2. Our model is pre-trained using contrastive learning, followed by supervised training with frozen weights. The input images \\(I\\) are augmented and transformed into patches. These patches are then processed to construct a part-whole hierarchy over \\(T\\) iterations, where embeddings at different levels gradually converge as the levels ascend. Next, the top-level embeddings are fed into \\(H_{d}\\), which performs downsampling based on weights that are negatively correlated with similarity. In the subsequent step, the embeddings are sent to \\(H_{1}\\) for contrastive pre-training. Finally, the output embeddings are passed to \\(H_{2}\\) for supervised training with frozen weights." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.482, + 0.469 + ], + "angle": 0, + "content": "where \\( Q, K, V = f^{k,t} \\). It should be noted that \\( \\mathcal{A}_{SA} \\) alone is insufficient for creating part-whole hierarchies, as it lacks the flexibility to adjust the degree of agreement (e.g., it struggles to encourage embeddings to achieve greater consensus at higher levels). To address this, GLOM incorporates additional regularization. In contrast, our model adopts a more effective strategy, drawing inspiration from contrastive learning, to promote agreement among embeddings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.47, + 0.483, + 0.635 + ], + "angle": 0, + "content": "Contrastive agreement enhancer \\(\\mathcal{A}_{CAE}\\): the function of our proposed CAE is two-fold: enhancing the agreement between positive embedding pairs; and pushing apart the negative embedding pairs. In contrastive learning, positive sample pairs are generated by applying different augmentations to the same image, while negative sample pairs are derived from images belonging to different classes. Unlike contrastive learning, in \\(\\mathcal{A}_{CAE}\\), the classification of embedding pairs as positive or negative is based on the similarity between embeddings. To be specific, the similarity is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.641, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\nS = \\frac {f _ {n 1} f _ {n 2} ^ {T}}{\\| f _ {n 1} \\| \\| f _ {n 2} \\|}, n 1, n 2 \\in \\{1, \\dots , N \\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.833 + ], + "angle": 0, + "content": "where the size of similarity map \\( S \\) is \\( N \\times N \\). For positive and negative embedding pairs, we could set different thresholds. Here we take a simple case for example with \\( S = 0 \\) as a threshold for both positive and negative embedding pairs. \\( S > 0 \\) means that the angle between embeddings \\( f_{n1} \\) and \\( f_{n2} \\) lies within the range \\( (0^{\\circ}, 90^{\\circ}) \\). In this case, \\( f_{n1} \\) and \\( f_{n2} \\) is a positive embedding pair and could belong to the same island. Conversely, \\( S < 0 \\) indicates that \\( f_{n1} \\) and \\( f_{n2} \\) may belong to different islands thereby being a negative pair. This process can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.842, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\nS _ {p} = S \\cdot \\mathbb {I} (S \\geq 0) + (- \\infty) \\cdot \\mathbb {I} (S < 0), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.861, + 0.482, + 0.877 + ], + "angle": 0, + "content": "\\[\nS _ {e} = S \\cdot \\mathbb {I} (S < 0), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(S_{p}\\) and \\(S_{e}\\) represent the similarity map for positive" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.395 + ], + "angle": 0, + "content": "and negative embedding pairs respectively. \\(\\mathbb{I}\\) is the indicator function. In the model's bottom level, the threshold is set to 0, while at the top level it is set to \\(-0.2\\). Next, we compute the embeddings at the time step \\(t + 1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.402, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\nf _ {p} ^ {t + 1} = \\operatorname {s o f t m a x} \\left(S _ {p}\\right) f ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.423, + 0.905, + 0.44 + ], + "angle": 0, + "content": "\\[\nf _ {e} ^ {t + 1} = S _ {e} f ^ {t}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.443, + 0.905, + 0.463 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} _ {C A E} \\left(f ^ {t}\\right) = f _ {p} ^ {t + 1} - f _ {e} ^ {t + 1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.59 + ], + "angle": 0, + "content": "where \\( f_{p}^{t + 1} \\) are the positive embeddings and \\( f_{e}^{t + 1} \\) are the negative embeddings. Compared with \\( f_{e}^{t + 1} \\), the computation of \\( f_{p}^{t + 1} \\) evolves the softmax function since the main purpose is to facilitate agreement between embeddings and softmax is good at eliminating the effect of negative embeddings (\\( S < 0 \\)). In addition, to push apart between positive and negative embeddings, we minus \\( f_{p}^{t + 1} \\) by \\( f_{e}^{t + 1} \\) to predict the \\( f^{t + 1} \\). An illustration for CAE is given in Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Fusion: At each time step \\(t\\), \\(f^{k,t}\\) is computed by the fusion of predictions from \\(\\mathcal{N}_{BU}, \\mathcal{N}_{TD}, \\mathcal{A}_{SA}\\), and \\(\\mathcal{A}_{CAE}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.642, + 0.878, + 0.661 + ], + "angle": 0, + "content": "\\[\nf ^ {k, t} = w _ {b u} \\mathcal {N} _ {B U} \\left(f ^ {k - 1, t - 1}\\right) + w _ {t d} \\mathcal {N} _ {T D} \\left(f ^ {k + 1, t - 1}\\right) +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.663, + 0.905, + 0.682 + ], + "angle": 0, + "content": "\\[\nw _ {l a} f ^ {k, t - 1} + w _ {a s} \\mathcal {A} _ {S A} \\left(f ^ {k, t - 1}\\right) + w _ {a c} \\mathcal {A} _ {C A E} \\left(f ^ {k, t - 1}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.881, + 0.705 + ], + "angle": 0, + "content": "where \\(w_{bu}, w_{td}, w_{la}, w_{as}, w_{ac}\\) are trainable parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.755, + 0.729 + ], + "angle": 0, + "content": "3.3. Dissimilarity-focused head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.825 + ], + "angle": 0, + "content": "With the aim to relieve the computation overhead introduced by which the embeddings at the top level and the lower level have the same resolution, we proposed a dissimilarity-focused head \\( H_{d} \\) which receives top-level embeddings as input to do downsampling as shown in Fig. 4, and it greatly reduces the number of parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "\\(H_{d}\\) is inspired by the idea that embeddings at the top level should exhibit agreement, which can result in redundancy between them. To quantify this redundancy, we apply indirect thinking and consider similarity as a suitable metric, since the more similar the embeddings are, the more" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24354" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.106, + 0.476, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.407, + 0.483, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Architecture of model. The column at location \\( n \\) is updating with a fusion of bottom-up and top-down prediction, history, SA, and CAE." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.47, + 0.478, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Figure 4. An illustration for the dissimilarity-focused head which is designed to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.786 + ], + "angle": 0, + "content": "redundant they become. To be specific, given the top embeddings \\( f^{K} \\in \\mathbb{R}^{h \\times w \\times d} \\), we first transform them into \\( p \\times p \\) patches each with the shape of \\( h' \\times w' \\times d \\) where \\( h' = \\frac{h}{p}, w' = \\frac{w}{p} \\). For each patch, similarity map \\( S' \\) with the shape of \\( (h' \\times w') \\times (h' \\times w') \\) is computed using Eq. 2. Next, we compute the weights \\( W \\) for each embedding within the patch:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.797, + 0.436, + 0.841 + ], + "angle": 0, + "content": "\\[\nW _ {i} = \\sum_ {j = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} e ^ {1 - S _ {i, j} ^ {\\prime}}, i, j \\in \\{1, \\dots , h ^ {\\prime} \\times w ^ {\\prime} \\},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.843, + 0.482, + 0.876 + ], + "angle": 0, + "content": "\\[\nW = \\frac {W}{\\operatorname* {m a x} \\left(W _ {i}\\right)}, W \\in \\mathbb {R} _ {> 0} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "where \\(1 - S_{i,j}^{\\prime}\\) means that embeddings with higher simi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "larity with others should be assigned with lower weight to downsample. Besides, the weights are summed to measure the overall weight for embedding \\( f_{i} \\). Finally, the downsampled output is computed by the weighted sum within each patch." + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.175, + 0.906, + 0.218 + ], + "angle": 0, + "content": "\\[\nH _ {d} \\left(f ^ {K}\\right) = \\sum_ {i = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} W _ {i} f _ {i}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.226, + 0.697, + 0.243 + ], + "angle": 0, + "content": "where \\(H_{d}(f^{K})\\in \\mathbb{R}^{p\\times p\\times d}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.619, + 0.267 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.906, + 0.333 + ], + "angle": 0, + "content": "The training of our model follows the setup in [8], consisting of two steps: 1) a pre-training phase using a supervised contrastive loss function, and 2) a classification training phase using cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.906, + 0.483 + ], + "angle": 0, + "content": "For pre-training, various data augmentation is adopted to each image \\(I\\) within a batch \\(B\\), producing pairs of image \\(I_{a}\\) and \\(I_{b}, a, b \\in \\{1, \\dots, 2B\\}\\). Then, the image pairs are fed to our model to produce representation \\(C_{n,a}\\) and \\(C_{n,b} \\in \\mathbb{R}^d\\), \\(n \\in \\{1, \\dots, N\\}\\) at the top level \\(K\\). In the following, we reshape them to the shape of \\(N \\times d\\) and send them to a contrastive head \\(H_{1}\\) which is constructed with two fully connected layers. The output is denoted as \\(O_{a}\\) and \\(O_{b}\\). Finally, a supervised contrastive loss is adopted to pre-train the model:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.489, + 0.905, + 0.528 + ], + "angle": 0, + "content": "\\[\nL _ {1} = - \\log \\frac {e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {b}\\right)}}{\\sum_ {i = 1} ^ {2 B} \\mathbb {I} _ {[ i \\neq a ]} e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {i}\\right)}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.595 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}_{[i\\neq a]}\\) is an indicator function that values 0 if \\(i\\) and \\(a\\) belong to the same image class, and 0 otherwise. sim is a cosine similarity function that measures the similarity of normalized input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.686 + ], + "angle": 0, + "content": "After the model is pre-trained with contrastive loss, we freeze the network weights. Besides, we add a classification head \\( H_{2} \\) on the top of contrastive head \\( H_{1} \\). The \\( H_{2} \\) is a linear layer that projects the representation to the final out with the dimension of \\( C \\). \\( C \\) is the number of classes. The cross-entropy loss is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.694, + 0.905, + 0.735 + ], + "angle": 0, + "content": "\\[\nL _ {2} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} p _ {i} \\log q _ {i}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.906, + 0.773 + ], + "angle": 0, + "content": "where \\( p_i \\) is the true probability distribution for class \\( i \\) and \\( q_i \\) is the predicted probability for class \\( i \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.638, + 0.802 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our experiments encompass quantitative and qualitative analyses. Quantitatively, we evaluate the formation of islands of agreement using the total variance of normalized embeddings and compare model performance across multiple datasets. An ablation study further examines the contributions of the proposed CAE and \\( H_{d} \\). For the qualitative" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24355" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.245, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.188, + 0.23, + 0.201 + ], + "angle": 0, + "content": "(a) C-10 bottom level" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.205, + 0.245, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.3, + 0.218, + 0.313 + ], + "angle": 0, + "content": "(f) C-10 top level" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.092, + 0.409, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.188, + 0.397, + 0.2 + ], + "angle": 0, + "content": "(b) C-100 bottom level" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.205, + 0.409, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.301, + 0.386, + 0.313 + ], + "angle": 0, + "content": "(g) C-100 top level" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.092, + 0.573, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.431, + 0.188, + 0.567, + 0.2 + ], + "angle": 0, + "content": "(c) S-NORB bottom level" + }, + { + "type": "image", + "bbox": [ + 0.426, + 0.205, + 0.572, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.441, + 0.3, + 0.558, + 0.313 + ], + "angle": 0, + "content": "(h) S-NORB top level" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.092, + 0.738, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.188, + 0.728, + 0.2 + ], + "angle": 0, + "content": "(d) MNIST bottom level" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.205, + 0.737, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.3, + 0.716, + 0.313 + ], + "angle": 0, + "content": "(i) MNIST top level" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.092, + 0.901, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.188, + 0.898, + 0.2 + ], + "angle": 0, + "content": "(e) F-MNIST bottom level" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.205, + 0.9, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.3, + 0.886, + 0.313 + ], + "angle": 0, + "content": "(j) F-MNIST top level" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.907, + 0.382 + ], + "angle": 0, + "content": "Figure 5. The comparison of our model with that of the Agglomerator in terms of \\( TV_{n} \\) at both the bottom and top levels across five datasets. It is evident that the total variance of our model is considerably lower than that of the Agglomerator, indicating that the contrastive attention mechanism effectively improves agreement in representing parts, which is essential for the theory of GLOM. Furthermore, the \\( TV_{n} \\) of our model is greater at the bottom level than at the top level, suggesting that agreement tends to increase as the levels rise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.543 + ], + "angle": 0, + "content": "analysis, we visualize the vector representations of emerging islands across levels. To illustrate CAE's enhancement of agreement intuitively, we depict representations with arrows spanning various locations and levels. Experiments are conducted on the CIFAR-10, CIFAR-100 [13], MNIST [14], FashionMNIST [26], and SmallNORB datasets [15]. The details of the training settings are provided in the supplementary material. The source code is publicly available at: https://github.com/zengru001usst/GLOM" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.303, + 0.567 + ], + "angle": 0, + "content": "4.1. Agreement Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.71 + ], + "angle": 0, + "content": "To verify the effectiveness of the proposed CAE, total variance [22] of the normalized embeddings \\((TV_{n})\\), as shown in Eq. 13, is employed as a metric to quantitatively measure the smoothness of embeddings at specific levels within the models. The embeddings are normalized, as a trivial solution may occur where total variance is minimized by making every \\(f_{i,j}\\) infinitesimal. Lower \\(TV_{n}\\) indicates greater smoothness, meaning that more embeddings are similar and achieve a higher level of agreement." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.731, + 0.401, + 0.774 + ], + "angle": 0, + "content": "\\[\nT V _ {n} = \\frac {1}{N} \\frac {1}{d} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {d} \\left(f _ {i, j} ^ {\\prime} - \\mu_ {j}\\right) ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.777, + 0.294, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\mu_ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} f _ {i, j} ^ {\\prime},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.821, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\nf _ {i, j} ^ {\\prime} = \\frac {f _ {i , j}}{\\sqrt {\\sum_ {j} f _ {i , j} ^ {2}}}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The results presented in Fig. 5 compare the \\( TV_{n} \\) of our model with that of the Agglomerator at both the bottom and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.589 + ], + "angle": 0, + "content": "top levels across five datasets. It can be observed that the \\( TV_{n} \\) of our model is significantly lower than that of the Agglomerator, indicating that the CAE effectively enhances agreement in representing parts, which is crucial for the theory of GLOM. Additionally, the \\( TV_{n} \\) of our model is higher at the bottom level compared to the top level, suggesting that agreement increases as the levels ascend. This is consistent with GLOM's framework, where embeddings at lower levels gradually cluster locally to represent parts, while embeddings at higher levels cluster globally to represent the whole. In contrast, the Agglomerator fails to maintain this characteristic." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.715, + 0.62 + ], + "angle": 0, + "content": "4.2. Classification Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To demonstrate the efficiency of the proposed model, we report the classification results for each dataset in terms of error percentage and the number of trainable parameters, as shown in Tab. 1. While Capsule Networks (CapsNets), such as [11, 18, 19, 21, 23, 28], perform well on simpler datasets like smallNORB, MNIST, and FashionMNIST, they struggle to scale to more complex datasets with a higher number of classes, such as CIFAR-100, due to inefficiencies in object representation [10]. Convolutional models, such as [9, 24], maintain strong performance across various datasets but fall short in interpretability. Transformer-based and MLP-based methods achieve state-of-the-art performance on more complex datasets, but they lack testing on smaller datasets. As shown in Tab. 1, our model achieves comparable classification error percentages to CapsNets and Agglomerator [8] on simpler datasets, but with significantly fewer parameters, especially compared to Agglomerator. Moreover, our model requires less training time and has a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24356" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.902, + 0.328 + ], + "angle": 0, + "content": "
MethodRefBackboneErrors %No. params (Millions)Training Arch.
S-NORBMNISTF-MNISTC-10C-100
E-CapsNet[19]Caps2.540.26---0.2GPU
CapsNet[23]2.700.256.3810.682.006.8GPU
Matrix-CapsNet[11]1.400.446.1411.9-0.3GPU
Capsule VB[21]1.600.305.2011.2-0.2GPU
Res-CapsNet[18]0.550.55-7.62-16.4GPU
IAR-CapsNet[28]1.360.77---0.18GPU
ResNet-110[9]Conv-2.105.106.41*27.76*1.7GPU
VGG[24]-0.326.507.74*28.05*20GPU
ViT-L/16[7]Transf---0.85*6.75*632TPU
ConvMLP-L[16]Conv/MLP---1.40*11.40*43TPU
MLP-Mixer-L/16[25]MLP---1.66*-207TPU
Agglomerator[8]Conv/MLP-/Caps0.010.307.4311.1540.9772GPU
Ours-0.010.305.1010.9041.312GPU
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.339, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Table 1. Classification results on SmallNORB (S-NORB), MNIST, FashionMNIST (F-MNIST), CIFAR-10 (C-10), and CIFAR-100 (C-100) datasets, demonstrating the efficiency of the proposed model. Results marked with * indicate networks pre-trained on ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.133, + 0.39, + 0.445, + 0.494 + ], + "angle": 0, + "content": "
HdCAEErrors (%)No. Param. (Millions)
Baseline××9.172
Model1×√9.072
Model2√×11.62
Model3√√10.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Table 2. Ablation for dissimilarity-focused head \\((H_{d})\\) and contrastive agreement enhancer (CAE) on CIFAR10." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.547, + 0.454, + 0.634 + ], + "angle": 0, + "content": "
Errors (%)TVnNo. Param. (Millions)
Average pooling11.10.3722
Max pooling11.10.4732
Hd10.90.5442
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Table 3. Comparison of dissimilarity-focused head \\((H_{d})\\), average pooling and maxpooling on CIFAR10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.73 + ], + "angle": 0, + "content": "much smaller architecture compared to most transformer-based and MLP-based methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.295, + 0.759 + ], + "angle": 0, + "content": "4.3.Quantitative Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To study the contribution of the different components of our model to its performance, we conduct an ablation experiment where we incrementally added the dissimilarity-focused head \\((H_{d})\\) and contrastive agreement enhancer (CAE) to the baseline model. The baseline model is defined by removing both \\(H_{d}\\) and CAE from our full model. We then compare the classification results and the number of parameters on CIFAR10 to demonstrate the contribution of \\(H_{d}\\) and CAE to model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.499 + ], + "angle": 0, + "content": "The results in Tab. 2 indicate that \\( H_{d} \\) can significantly reduce the number of trainable parameters in the model while only causing a slight increase in errors, which suggests the presence of redundancy at the top level. Additionally, when \\( H_{d} \\) is employed, CAM can enhance model performance, as much of the redundancy has been eliminated, allowing CAM to provide valuable supplementary information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.652 + ], + "angle": 0, + "content": "To evaluate the impact of \\( H_{d} \\) on classification accuracy and \\( TV_{n} \\), we compare our model against two alternative downsampling strategies: max pooling and average pooling, by substituting \\( H_{d} \\) with these methods. Results in Tab. 3 show that the model with \\( H_{d} \\) achieves a lower error rate and higher \\( TV_{n} \\) than these alternatives under identical downsampling conditions. The higher \\( TV_{n} \\) with \\( H_{d} \\) stems from its emphasis on embeddings with lower similarity, which carry more informative features and enhance classification accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.708, + 0.681 + ], + "angle": 0, + "content": "4.4. Qualitative Ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.717 + ], + "angle": 0, + "content": "We also conduct two qualitative ablation study to demonstrate the effect of the contrastive agreement enhancer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.87 + ], + "angle": 0, + "content": "The emerging island of agreement. To illustrate the emergence of islands across different levels, the embeddings from various \\( k \\) levels of CIFAR-10 are visualized in Fig. 6. When examining the embeddings at the highest level (\\( k = 5 \\)), those generated by the model with CAE reveal finer details and a more distinct separation between objects and the background. Furthermore, a comparison of embeddings across different levels demonstrates that as the levels increase, the embeddings progressively reach agreement, effectively capturing part-whole hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The representation visualization using arrows. To provide an intuitive illustration of the enhancement in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.898, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.908, + 0.407 + ], + "angle": 0, + "content": "Figure 6. The emerging island of agreement at various \\(k\\) levels from CIFAR-10. At the top level \\((k = 5)\\), the representations generated by the model with CAE display enhanced detail, highlighting a more distinct separation between objects and the background." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.424, + 0.898, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.592, + 0.908, + 0.621 + ], + "angle": 0, + "content": "Figure 7. The representation visualization using arrows for model with and without CAE. The direction and length of each arrow represent the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.483, + 0.812 + ], + "angle": 0, + "content": "agreement introduced by CAE, we visualize the representations using arrows across various locations and levels. The direction and length of each arrow indicate the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator. MNIST is selected for visualization due to its simplicity, minimizing potential interference from complex image backgrounds. As depicted in Fig. 7, the arrows in the model incorporating CAE demonstrate greater consistency, whereas those in the model without CAE appear more disordered and inconsistent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.21, + 0.845 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this paper, we have proposed a new implementation of GLOM that aligns more closely with its theoretical framework for forming hierarchies of grouped embeddings to rep-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.907, + 0.873 + ], + "angle": 0, + "content": "resent nodes in a parse tree. To foster the formation of distinct islands, we have designed a novel contrastive agreement enhancer that encourages agreement between positive embedding pairs while separating negative pairs. To reduce the computational burden of GLOM, we developed a dissimilarity-focused head that preserves more informative content by prioritizing embeddings with lower similarity, which often contain richer information. Both quantitative and qualitative analyses demonstrate that the contrastive agreement enhancer effectively promotes islands of agreement, while the dissimilarity-focused head enhances model efficiency across various datasets. This work significantly improves the practicality of GLOM, and we hope it inspires other researchers to further explore this direction to advance its development." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24358" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. Publisher Copyright: © 2021 Owner/Author. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[2] Ran Chen, Hao Shen, Zhong-Qiu Zhao, Yi Yang, and Zhao Zhang. Global routing between capsules. Pattern Recognition, 148:110142, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.483, + 0.255 + ], + "angle": 0, + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[4] Jaewoong Choi, Hyun Seo, Suii Im, and Myungjoo Kang. Attention routing between capsules. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[6] Laura Culp, Sara Sabour, and Geoffrey E Hinton. Testing glom's ability to infer wholes from ambiguous parts. arXiv preprint arXiv:2211.16564, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[8] Nicola Garau, Niccolò Bisagno, Zeno Sambugaro, and Nicola Conc. Interpretable part-whole hierarchies and conceptual-semantic relationships in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13689-13698, 2022. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[10] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. Neural Computation, 35(3):413-452, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[11] Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing. In International conference on learning representations, 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[12] Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[15] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-104. IEEE, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.16 + ], + "angle": 0, + "content": "[16] Jiachen Li, Ali Hassani, Steven Walton, and Humphrey Shi. Convmlp: Hierarchical convolutional mlp's for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6307-6316, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[17] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3688-3704, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[18] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, pages 1-14, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.327 + ], + "angle": 0, + "content": "[19] Vittorio Mazzia, Francesco Salvetti, and Marcello Chi-aberge. Efficient-capsnet: Capsule network with self-attention routing. Scientific reports, 11(1):14634, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.906, + 0.383 + ], + "angle": 0, + "content": "[20] Rita Pucci, Christian Micheloni, and Niki Martinel. Self-attention agreement among capsules. In Proceedings of the iee//cvf international conference on computer vision, pages 272-280, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[21] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3749-3756, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[22] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[23] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.565 + ], + "angle": 0, + "content": "[24] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.906, + 0.648 + ], + "angle": 0, + "content": "[25] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[26] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[27] Ru Zeng and Yan Song. A fast routing capsule network with improved dense blocks. IEEE Transactions on Industrial Informatics, 18(7):4383-4392, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[28] Ru Zeng, Yuzhang Qin, and Yan Song. A non-iterative capsule network with interdependent agreement routing. Expert Systems with Applications, 238:122284, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[29] Ru Zeng, Yan Song, and Yuzhang Qin. Spatial attention-based capsule networks with guaranteed group equivariance. IEEE Transactions on Automation Science and Engineering, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[30] Ru Zeng, Yan Song, and Yanjiu Zhong. An interpretable unsupervised capsule network via comprehensive contrastive learning and two-stage training. Pattern Recognition, 158: 111059, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24359" + } + ] +] \ No newline at end of file diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_origin.pdf b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..43d4af9c3993a37aabec4cc76fd0978f93351ae4 --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/bd040c1e-9041-429f-8136-d5caa2846f28_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94e79fc211bcd7fe8faf5af9c3a0f243a32fbceb51328c5b6b999e744c787fa5 +size 3701060 diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/full.md b/2025/Agreement aware and dissimilarity oriented GLOM/full.md new file mode 100644 index 0000000000000000000000000000000000000000..79c9665c59e55dbda391e5d8edbe20eb55999af9 --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/full.md @@ -0,0 +1,321 @@ +# Agreement aware and dissimilarity oriented GLOM + +$\mathrm{RuZeng^{1}}$ Yan Song $^{1*}$ Yang Zhang $^{2}$ Yanling Hu $^{3}$ Hui Yu $^{4}$ + +1University of Shanghai for Science and Technology + +2Shanghai Urban and Rural Construction and Transportation Development Institute + +$^{3}$ Guangxi Medical University $^{4}$ University of Glasgow + +# Abstract + +GLOM, an innovative departure from standard deep learning architectures, has been proposed and gained special concern recently due to its good interpretability in representing part-whole relationships in computer vision. However, GLOM faces challenges in achieving agreement and is usually computationally demanding. First, current implementations struggle to produce identical vectors that reliably converge to represent nodes in a parse tree. Second, GLOM is computationally intensive due to the need to maintain equal resolution across all levels. To address these issues, inspired by contrastive learning, we proposed a contrastive agreement enhancer (CAE), which effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby facilitating forming distinct "islands." Furthermore, we introduce a dissimilarity-focused head $(H_{d})$ to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window. The results of comparison experiments indicate that the proposed approach delicately retains informative content and significantly reduces the number of parameters. Additionally, the ablation experiments and visualization results demonstrate that CAE successfully promotes islands of agreement. + +# 1. Introduction + +As deep neural networks continue to achieve impressive performance across various tasks, there is increasing focus on model interpretability, particularly in the domain of computer vision [1]. GLOM was proposed to address this need, offering a promising solution to the challenge of representing part-whole hierarchies in neural networks [10]. It integrates the strengths of transformers, neural fields, contrastive representation learning, distillation, and capsules, + +![](images/cd4113201db0508a242d8f805c009ab27f570726652f06670d9a7df69b720134.jpg) +(a) An illustration for the contrastive agreement enhancer + +![](images/c9edbc169916ea4572e74b1f9ba96b04b0222cb321f04c95c298c2c72dfa1b66.jpg) + +![](images/a226053f987e1aec2e39f1505645162ffff982861fea80827c69d069d9718035.jpg) + +![](images/ad046ef7d996a2d7beca4644516e0a586e9852dfa99ff8b626b01e737035666a.jpg) +(b) $TV_{n}$ for bottom level + +![](images/ad3a0e77b2fe2e42d8426e729a6e19ea3dfdf33be3ef2119e73a03d1c2e85437.jpg) +(c) $TV_{n}$ for top level +Figure 1. a): To tackle the challenge in current GLOM implementations, where vectors struggle to form islands representing parts and wholes and fail to grow larger islands at higher levels, we introduce a contrastive agreement enhancer. This approach effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct islands. By adjusting the threshold for distinguishing positive and negative pairs, we ensure increased smoothness at higher levels that indicate the island grows larger. b) and c): Total variance of normalized embeddings $(TV_{n})$ is used as a metric to quantitatively measure smoothness (with lower $TV_{n}$ indicating greater smoothness). Our method enhances smoothness at both the bottom and top levels, with the top level achieving greater smoothness. + +garnering significant attention upon its introduction. At its core, GLOM uses "islands"—groups of similar feature vectors dynamically identified as coherent units, such as parts of an object—to form nodes in a parse tree. However, current implementations [6, 8] of GLOM struggle to produce similar vectors that reliably reach agreement to form these islands. Moreover, GLOM models are computationally intensive due to the need to maintain equal resolution across + +all levels. These limitations inspire the enhancements proposed in this paper, which seek to address the vector agreement problem and improve computational efficiency. + +In the GLOM system, embeddings at each level are feature vectors that gradually converge to form distinct islands of nearly identical vectors, representing parts like the head, body, and tail, as shown in the right panel of Fig. 1a. To achieve this, GLOM incorporates a regularizer [6] that minimizes the cosine distance between the bottom-up predictions from lower-level embeddings and the top-down predictions from higher-level embeddings at adjacent levels. However, a practical challenge arises due to the need to balance multiple regularizers between various levels, making it difficult to properly manage their interactions. Another option is a regularizer that encourages similarity between embeddings of nearby locations, but this can lead to the collapse of representations. Although contrastive learning, which commonly uses negative samples to prevent collapse [3], could be a solution, it introduces the challenge of producing distinct islands and recognizing objects, as it relies on finding suitable negative samples to differentiate these islands. Additionally, as the islands grow larger at higher levels to represent whole entities like birds, the embeddings should become more similar, resulting in increased smoothness across the feature map. We use total variance as a metric to quantitatively measure this smoothness. As illustrated by the blue line in Fig. 1b and Fig. 1c, we observe that the total variance [22] of GLOM's normalized embeddings is higher at higher levels than at lower levels, which contradicts GLOM's theoretical expectations. + +To address these two issues, we propose a contrastive agreement enhancer (CAE), as shown in Fig. 1a that computes the cosine similarity between two embeddings and determines whether they belong to a positive or negative pair based on a predefined threshold. This mechanism, inspired by contrastive learning, promotes similarity between positive pairs while pushing apart negative pairs. CAE offers several advantages: 1) it can be easily integrated into every level of GLOM; 2) it is not a regularizer, eliminating the need to balance multiple regularizers across levels; 3) it effectively promotes embedding agreement to form distinct islands; and 4) by adjusting the threshold, it ensures increased smoothness at higher levels. + +Another focus of this paper is mitigating the computational challenges in GLOM, which arise from maintaining equal resolution across all levels. Specifically, when implementing training or downstream tasks, different heads are customized and added to the top of GLOM. If these heads receive the entire top-level embeddings as input, the computational burden becomes significant because the top-level resolution is as high as the bottom level. A straightforward solution is to downsample the embeddings, based on the premise that redundancy exists within it. Redundancy is in- + +deed present at the top level, as the embeddings there have reached agreement and become highly similar. To handle this, we design a dissimilarity-focused head $(H_{d})$ , which is integrated at the top of GLOM. $H_{d}$ similarly to common pooling methods, where a small rectangular window slides across the input feature map. For each window position, $H_{d}$ calculates the similarity of each embedding with the others within the window. The weights for each embedding are then computed based on the negative correlation of these similarities. The output of $H_{d}$ is the weighted sum of the embeddings. The advantages of $H_{d}$ include significantly reducing the parameters and computational load of the heads. Moreover, compared to max pooling, $H_{d}$ preserves more informative content by focusing on embeddings with lower similarity, which tend to contain more useful information—this is an improvement over max pooling, which just discards all but the highest-value embedding. + +Overall, the main contributions of this work can be summarized as follows. 1) A new implementation of GLOM is proposed that aligns more closely with GLOM's theoretical framework for forming hierarchies of grouped embeddings to represent nodes in a parse tree. 2) A novel contrastive agreement enhancer is designed to promote agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct "islands." 3) A dissimilarity-focused head is developed to reduce redundancy in the top-level embeddings, preserving more informative content by concentrating on embeddings with lower similarity, which tend to contain more useful information. 4) Quantitative and qualitative analyses show that CAE effectively promotes islands of agreement and $H_{d}$ contributes greatly to improving model efficiency across datasets. + +# 2. Related work + +Capsule networks and GLOM: GLOM provides a more efficient approach to representing part-whole hierarchies than capsule networks (CapsNets) for several reasons [10]. First, CapsNets often face inefficiencies in part representation due to the need to pre-allocate neurons for possible parts at specific locations, leading to underutilized capsules during inference [29, 30]. In contrast, GLOM leverages all neurons to form clusters, facilitating knowledge sharing across locations and improving part representation efficiency. Second, conventional CapsNets require presetting the number of clusters to represent higher-level parts. For instance, studies like [23] and [11] preset the number of capsule types and slice the feature map along the channel direction to form capsules. GLOM, however, allows embeddings to naturally form islands without the need for predefined clusters. Lastly, GLOM eliminates dynamic routing by enabling each part's location to independently construct its own vector representation of the whole, avoiding the need to route information to specific capsules. How + +ever, GLOM is still in its early stages of development and has its own limitations. For example, it primarily functions as an imagery model, with few implementations available to thoroughly evaluate its performance [6, 8]. Current implementations struggle to achieve proper agreement between embeddings, which hampers the formation of islands in GLOM. Additionally, embeddings are set at the same resolution across both low and high levels, leading to redundancy in the model's trainable parameters. This paper focuses on addressing these shortcomings. + +Routing and embedding updating: Routing algorithms are a crucial aspect of CapsNets, intended to replace the pooling algorithms typically used in CNNs by assessing the agreement between capsules [23]. This agreement reflects how much lower-level capsules support the existence of a higher-level capsule, based on the similarity of their predictions [11]. Many works about CapsNets focus on reducing the high computation burden in the capsule routing algorithm. For instance, [27] eliminates the iterations in routing and uses capsule distribution variance to activate high-level capsules. Additionally, [2, 4, 17, 19, 20] introduce efficient attention mechanisms into routing, significantly reducing computational overhead. In contrast, GLOM eliminates the need for routing but depends on similar embeddings to iteratively update and form 'islands' representing parts or whole objects [10]. Both routing and embedding updates aim to use similar capsules/embeddings to represent nodes in a sparse tree, organizing objects into a part-whole hierarchy. A key focus of this paper is the role of embedding similarity, which informs the design of the proposed CAE and $H_{d}$ , each serving different purposes. CAE improves GLOM's island formation by promoting agreement between similar embeddings and disagreement between dissimilar ones, while $H_{d}$ reduces computational overhead at higher levels by selecting embeddings that are less similar to others, suggesting they may contain more valuable information. + +# 3. Method + +Our proposed model aims to efficiently implement the GLOM architecture while more closely aligning with its theoretical framework for creating hierarchical islands of embeddings that represent nodes in a parse tree, with these islands expanding at higher levels. Additionally, the model is designed to be lightweight, facilitating broader potential applications of the GLOM architecture. + +Here, we start with the introduction of the mathematical notation needed to explain the details of the main components of the architecture. Given an input image $I$ , it is spatially transformed into $N = h \times w$ patches. The $n$ -th patch corresponds to column $C_n(h,w)$ at the spatial location $(h,w)$ , where $n \in \{1,\dots,N\}$ . For each column, it consists of $K$ -level embeddings $f_{(h,w)}^{k,t}, k \in \{1,\dots,K\}$ at + +time step $t \in \{1, \dots, T\}$ with a size of $d$ for each embedding. The subscript $(h, w)$ and superscript $t$ are omitted in subsequent instances for readability. $f^k$ and $f^{k+1}$ are consecutive and represent the part and whole respectively. They are connected by a bottom-up encoder and a top-down decoder which are shared across the spatial dimension. + +# 3.1. Patches embedding + +The pipeline is demonstrated in Fig. 2. Firstly, input image $I$ , with the shape of $H \times W \times c$ , is transformed into patch embeddings. Following the recipe provided in [16], a convolutional Tokenizer is used to extract feature maps with the size of $H' \times W' \times c'$ . Next, it is reshaped to the size of $(h \times r) \times (w \times r) \times c'$ , where $r$ is the patch size, $H' = h \times r$ , and $W' = w \times r$ . Then, we reshape and permute it with the size of $h \times w \times d$ , where $d = r \times r \times c'$ . At last, the $n$ $d$ -dimensional features act as the embeddings $f^{1,t}$ at the bottom level in columns $C_n$ at time step $t$ . + +# 3.2. Embedding updating + +Secondly, we introduce the propagation phase, which governs the interactions between levels and across locations. As shown in Fig. 3, the embedding of a column at location $n$ and level $k$ is updated by integrating multiple sources of information: predictions from lower and higher levels at the same location $n$ using a bottom-up encoder and a top-down decoder, predictions from embeddings at other locations within the same level via self-attention and a contrastive agreement enhancer, and the historical context (i.e., the embedding from the previous iteration). Next, we elaborate on the components of the model that drive this embedding update process. + +Bottom-up encoder $\mathcal{N}_{BU}$ : Taking a low-level embedding $f^{k,t}$ as input, the bottom-up encoder $\mathcal{N}_{BU}$ is used to predict high-level embedding $f^{k+1,t+1}$ with the same size of $d$ . It is constructed by two fully connected layers and shared across $n$ columns. + +Top-down decoder $\mathcal{N}_{TD}$ : Given a high-level embedding $f^{k + 1,t}$ as input, the top-down decoder $\mathcal{N}_{TD}$ is applied to predict the corresponding low-level embedding $f^{k,t + 1}$ . The intuition is that if the model understands the whole, it can effectively predict its parts. Like the encoder, $\mathcal{N}_{TD}$ also comprises two fully connected layers. + +Self-attention $\mathcal{A}_{SA}$ : In GLOM, the interaction between columns is implemented by a simplified self-attention that query, value, and key vector are all the same as input. The aim of self-attention is to produce islands of identical embeddings at a level by making each embedding vector at that level regress towards other similar vectors at nearby locations. $\mathcal{A}_{SA}$ is computed as: + +$$ +\mathcal {A} _ {S A} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V, \tag {1} +$$ + +![](images/6a2006584f97e92187f33470dade1fc6b10e63664290beee73c81f33c080f6a0.jpg) +Figure 2. Our model is pre-trained using contrastive learning, followed by supervised training with frozen weights. The input images $I$ are augmented and transformed into patches. These patches are then processed to construct a part-whole hierarchy over $T$ iterations, where embeddings at different levels gradually converge as the levels ascend. Next, the top-level embeddings are fed into $H_{d}$ , which performs downsampling based on weights that are negatively correlated with similarity. In the subsequent step, the embeddings are sent to $H_{1}$ for contrastive pre-training. Finally, the output embeddings are passed to $H_{2}$ for supervised training with frozen weights. + +where $Q, K, V = f^{k,t}$ . It should be noted that $\mathcal{A}_{SA}$ alone is insufficient for creating part-whole hierarchies, as it lacks the flexibility to adjust the degree of agreement (e.g., it struggles to encourage embeddings to achieve greater consensus at higher levels). To address this, GLOM incorporates additional regularization. In contrast, our model adopts a more effective strategy, drawing inspiration from contrastive learning, to promote agreement among embeddings. + +Contrastive agreement enhancer $\mathcal{A}_{CAE}$ : the function of our proposed CAE is two-fold: enhancing the agreement between positive embedding pairs; and pushing apart the negative embedding pairs. In contrastive learning, positive sample pairs are generated by applying different augmentations to the same image, while negative sample pairs are derived from images belonging to different classes. Unlike contrastive learning, in $\mathcal{A}_{CAE}$ , the classification of embedding pairs as positive or negative is based on the similarity between embeddings. To be specific, the similarity is defined as: + +$$ +S = \frac {f _ {n 1} f _ {n 2} ^ {T}}{\| f _ {n 1} \| \| f _ {n 2} \|}, n 1, n 2 \in \{1, \dots , N \}, \tag {2} +$$ + +where the size of similarity map $S$ is $N \times N$ . For positive and negative embedding pairs, we could set different thresholds. Here we take a simple case for example with $S = 0$ as a threshold for both positive and negative embedding pairs. $S > 0$ means that the angle between embeddings $f_{n1}$ and $f_{n2}$ lies within the range $(0^{\circ}, 90^{\circ})$ . In this case, $f_{n1}$ and $f_{n2}$ is a positive embedding pair and could belong to the same island. Conversely, $S < 0$ indicates that $f_{n1}$ and $f_{n2}$ may belong to different islands thereby being a negative pair. This process can be described as: + +$$ +S _ {p} = S \cdot \mathbb {I} (S \geq 0) + (- \infty) \cdot \mathbb {I} (S < 0), \tag {3} +$$ + +$$ +S _ {e} = S \cdot \mathbb {I} (S < 0), \tag {4} +$$ + +where $S_{p}$ and $S_{e}$ represent the similarity map for positive + +and negative embedding pairs respectively. $\mathbb{I}$ is the indicator function. In the model's bottom level, the threshold is set to 0, while at the top level it is set to $-0.2$ . Next, we compute the embeddings at the time step $t + 1$ : + +$$ +f _ {p} ^ {t + 1} = \operatorname {s o f t m a x} \left(S _ {p}\right) f ^ {t}, \tag {5} +$$ + +$$ +f _ {e} ^ {t + 1} = S _ {e} f ^ {t}, \tag {6} +$$ + +$$ +\mathcal {A} _ {C A E} \left(f ^ {t}\right) = f _ {p} ^ {t + 1} - f _ {e} ^ {t + 1}, \tag {7} +$$ + +where $f_{p}^{t + 1}$ are the positive embeddings and $f_{e}^{t + 1}$ are the negative embeddings. Compared with $f_{e}^{t + 1}$ , the computation of $f_{p}^{t + 1}$ evolves the softmax function since the main purpose is to facilitate agreement between embeddings and softmax is good at eliminating the effect of negative embeddings ( $S < 0$ ). In addition, to push apart between positive and negative embeddings, we minus $f_{p}^{t + 1}$ by $f_{e}^{t + 1}$ to predict the $f^{t + 1}$ . An illustration for CAE is given in Fig. 1a. + +Fusion: At each time step $t$ , $f^{k,t}$ is computed by the fusion of predictions from $\mathcal{N}_{BU}, \mathcal{N}_{TD}, \mathcal{A}_{SA}$ , and $\mathcal{A}_{CAE}$ : + +$$ +f ^ {k, t} = w _ {b u} \mathcal {N} _ {B U} \left(f ^ {k - 1, t - 1}\right) + w _ {t d} \mathcal {N} _ {T D} \left(f ^ {k + 1, t - 1}\right) + +$$ + +$$ +w _ {l a} f ^ {k, t - 1} + w _ {a s} \mathcal {A} _ {S A} \left(f ^ {k, t - 1}\right) + w _ {a c} \mathcal {A} _ {C A E} \left(f ^ {k, t - 1}\right). \tag {8} +$$ + +where $w_{bu}, w_{td}, w_{la}, w_{as}, w_{ac}$ are trainable parameters. + +# 3.3. Dissimilarity-focused head + +With the aim to relieve the computation overhead introduced by which the embeddings at the top level and the lower level have the same resolution, we proposed a dissimilarity-focused head $H_{d}$ which receives top-level embeddings as input to do downsampling as shown in Fig. 4, and it greatly reduces the number of parameters. + +$H_{d}$ is inspired by the idea that embeddings at the top level should exhibit agreement, which can result in redundancy between them. To quantify this redundancy, we apply indirect thinking and consider similarity as a suitable metric, since the more similar the embeddings are, the more + +![](images/7ecde751620dfe32aa975dd6a3b8e400ee86c330de3e3a885392efdbf628b949.jpg) +Figure 3. Architecture of model. The column at location $n$ is updating with a fusion of bottom-up and top-down prediction, history, SA, and CAE. + +![](images/cefc780b066a9068d8af1a13d2da4604e5de505f7ee117018bf2aa43786ffe48.jpg) +Figure 4. An illustration for the dissimilarity-focused head which is designed to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window. + +redundant they become. To be specific, given the top embeddings $f^{K} \in \mathbb{R}^{h \times w \times d}$ , we first transform them into $p \times p$ patches each with the shape of $h' \times w' \times d$ where $h' = \frac{h}{p}, w' = \frac{w}{p}$ . For each patch, similarity map $S'$ with the shape of $(h' \times w') \times (h' \times w')$ is computed using Eq. 2. Next, we compute the weights $W$ for each embedding within the patch: + +$$ +W _ {i} = \sum_ {j = 1} ^ {h ^ {\prime} \times w ^ {\prime}} e ^ {1 - S _ {i, j} ^ {\prime}}, i, j \in \{1, \dots , h ^ {\prime} \times w ^ {\prime} \}, +$$ + +$$ +W = \frac {W}{\operatorname* {m a x} \left(W _ {i}\right)}, W \in \mathbb {R} _ {> 0} ^ {h ^ {\prime} \times w ^ {\prime}}, \tag {9} +$$ + +where $1 - S_{i,j}^{\prime}$ means that embeddings with higher simi + +larity with others should be assigned with lower weight to downsample. Besides, the weights are summed to measure the overall weight for embedding $f_{i}$ . Finally, the downsampled output is computed by the weighted sum within each patch. + +$$ +H _ {d} \left(f ^ {K}\right) = \sum_ {i = 1} ^ {h ^ {\prime} \times w ^ {\prime}} W _ {i} f _ {i}, \tag {10} +$$ + +where $H_{d}(f^{K})\in \mathbb{R}^{p\times p\times d}$ + +# 3.4. Training + +The training of our model follows the setup in [8], consisting of two steps: 1) a pre-training phase using a supervised contrastive loss function, and 2) a classification training phase using cross-entropy loss. + +For pre-training, various data augmentation is adopted to each image $I$ within a batch $B$ , producing pairs of image $I_{a}$ and $I_{b}, a, b \in \{1, \dots, 2B\}$ . Then, the image pairs are fed to our model to produce representation $C_{n,a}$ and $C_{n,b} \in \mathbb{R}^d$ , $n \in \{1, \dots, N\}$ at the top level $K$ . In the following, we reshape them to the shape of $N \times d$ and send them to a contrastive head $H_{1}$ which is constructed with two fully connected layers. The output is denoted as $O_{a}$ and $O_{b}$ . Finally, a supervised contrastive loss is adopted to pre-train the model: + +$$ +L _ {1} = - \log \frac {e ^ {\operatorname {s i m} \left(O _ {a} , O _ {b}\right)}}{\sum_ {i = 1} ^ {2 B} \mathbb {I} _ {[ i \neq a ]} e ^ {\operatorname {s i m} \left(O _ {a} , O _ {i}\right)}}, \tag {11} +$$ + +where $\mathbb{I}_{[i\neq a]}$ is an indicator function that values 0 if $i$ and $a$ belong to the same image class, and 0 otherwise. sim is a cosine similarity function that measures the similarity of normalized input. + +After the model is pre-trained with contrastive loss, we freeze the network weights. Besides, we add a classification head $H_{2}$ on the top of contrastive head $H_{1}$ . The $H_{2}$ is a linear layer that projects the representation to the final out with the dimension of $C$ . $C$ is the number of classes. The cross-entropy loss is given as: + +$$ +L _ {2} = - \frac {1}{C} \sum_ {i = 1} ^ {C} p _ {i} \log q _ {i}, \tag {12} +$$ + +where $p_i$ is the true probability distribution for class $i$ and $q_i$ is the predicted probability for class $i$ . + +# 4. Experiment + +Our experiments encompass quantitative and qualitative analyses. Quantitatively, we evaluate the formation of islands of agreement using the total variance of normalized embeddings and compare model performance across multiple datasets. An ablation study further examines the contributions of the proposed CAE and $H_{d}$ . For the qualitative + +![](images/2465713049a8580e2e580014a00e072363a4ab0df0165c2d5af07ec9f9ec78c6.jpg) +(a) C-10 bottom level + +![](images/04ff742b0ddf4cc998367916a00cafd2c2cdb91f16a66c98f9fccf43801614a1.jpg) +(f) C-10 top level + +![](images/a2b8f22f3870a2b2e2bf2382ddc70c3c622d6606ed5f31558cf022e59be48cdd.jpg) +(b) C-100 bottom level + +![](images/1c587e6392be397c0e822a8dc17d2e7863396a6970310c003ffb98167fcfba4d.jpg) +(g) C-100 top level + +![](images/9134948e8d7e40f4e698c04dc9cb484325eb972329d294a29e64e4eead927ef0.jpg) +(c) S-NORB bottom level + +![](images/779546a59c5ba37e82e7e54a77847b268bc543898d2f59f229ffd5a667169e66.jpg) +(h) S-NORB top level +Figure 5. The comparison of our model with that of the Agglomerator in terms of $TV_{n}$ at both the bottom and top levels across five datasets. It is evident that the total variance of our model is considerably lower than that of the Agglomerator, indicating that the contrastive attention mechanism effectively improves agreement in representing parts, which is essential for the theory of GLOM. Furthermore, the $TV_{n}$ of our model is greater at the bottom level than at the top level, suggesting that agreement tends to increase as the levels rise. + +![](images/f011562337a8388a3b410821b75d9bcdce71b696d0a3cbb8ad0c7bc89c40150e.jpg) +(d) MNIST bottom level + +![](images/c058d9ba13e4f98bd6e944a8b31f8f100dbad0143b6a465cbae62cb37cafb810.jpg) +(i) MNIST top level + +![](images/42ec8d94e61af33e2e41839e04789b27dd86d46e26a49aa9a87fde3c7996ba42.jpg) +(e) F-MNIST bottom level + +![](images/547f56353964ef660722ebbf5a75be6da6dd07c30e1768035e5e7012bbbd09db.jpg) +(j) F-MNIST top level + +analysis, we visualize the vector representations of emerging islands across levels. To illustrate CAE's enhancement of agreement intuitively, we depict representations with arrows spanning various locations and levels. Experiments are conducted on the CIFAR-10, CIFAR-100 [13], MNIST [14], FashionMNIST [26], and SmallNORB datasets [15]. The details of the training settings are provided in the supplementary material. The source code is publicly available at: https://github.com/zengru001usst/GLOM + +# 4.1. Agreement Evaluation + +To verify the effectiveness of the proposed CAE, total variance [22] of the normalized embeddings $(TV_{n})$ , as shown in Eq. 13, is employed as a metric to quantitatively measure the smoothness of embeddings at specific levels within the models. The embeddings are normalized, as a trivial solution may occur where total variance is minimized by making every $f_{i,j}$ infinitesimal. Lower $TV_{n}$ indicates greater smoothness, meaning that more embeddings are similar and achieve a higher level of agreement. + +$$ +T V _ {n} = \frac {1}{N} \frac {1}{d} \sum_ {i = 1} ^ {N} \sum_ {j = 1} ^ {d} \left(f _ {i, j} ^ {\prime} - \mu_ {j}\right) ^ {2}, +$$ + +$$ +\mu_ {j} = \frac {1}{N} \sum_ {i = 1} ^ {N} f _ {i, j} ^ {\prime}, +$$ + +$$ +f _ {i, j} ^ {\prime} = \frac {f _ {i , j}}{\sqrt {\sum_ {j} f _ {i , j} ^ {2}}}. \tag {13} +$$ + +The results presented in Fig. 5 compare the $TV_{n}$ of our model with that of the Agglomerator at both the bottom and + +top levels across five datasets. It can be observed that the $TV_{n}$ of our model is significantly lower than that of the Agglomerator, indicating that the CAE effectively enhances agreement in representing parts, which is crucial for the theory of GLOM. Additionally, the $TV_{n}$ of our model is higher at the bottom level compared to the top level, suggesting that agreement increases as the levels ascend. This is consistent with GLOM's framework, where embeddings at lower levels gradually cluster locally to represent parts, while embeddings at higher levels cluster globally to represent the whole. In contrast, the Agglomerator fails to maintain this characteristic. + +# 4.2. Classification Results + +To demonstrate the efficiency of the proposed model, we report the classification results for each dataset in terms of error percentage and the number of trainable parameters, as shown in Tab. 1. While Capsule Networks (CapsNets), such as [11, 18, 19, 21, 23, 28], perform well on simpler datasets like smallNORB, MNIST, and FashionMNIST, they struggle to scale to more complex datasets with a higher number of classes, such as CIFAR-100, due to inefficiencies in object representation [10]. Convolutional models, such as [9, 24], maintain strong performance across various datasets but fall short in interpretability. Transformer-based and MLP-based methods achieve state-of-the-art performance on more complex datasets, but they lack testing on smaller datasets. As shown in Tab. 1, our model achieves comparable classification error percentages to CapsNets and Agglomerator [8] on simpler datasets, but with significantly fewer parameters, especially compared to Agglomerator. Moreover, our model requires less training time and has a + +
MethodRefBackboneErrors %No. params (Millions)Training Arch.
S-NORBMNISTF-MNISTC-10C-100
E-CapsNet[19]Caps2.540.26---0.2GPU
CapsNet[23]2.700.256.3810.682.006.8GPU
Matrix-CapsNet[11]1.400.446.1411.9-0.3GPU
Capsule VB[21]1.600.305.2011.2-0.2GPU
Res-CapsNet[18]0.550.55-7.62-16.4GPU
IAR-CapsNet[28]1.360.77---0.18GPU
ResNet-110[9]Conv-2.105.106.41*27.76*1.7GPU
VGG[24]-0.326.507.74*28.05*20GPU
ViT-L/16[7]Transf---0.85*6.75*632TPU
ConvMLP-L[16]Conv/MLP---1.40*11.40*43TPU
MLP-Mixer-L/16[25]MLP---1.66*-207TPU
Agglomerator[8]Conv/MLP-/Caps0.010.307.4311.1540.9772GPU
Ours-0.010.305.1010.9041.312GPU
+ +Table 1. Classification results on SmallNORB (S-NORB), MNIST, FashionMNIST (F-MNIST), CIFAR-10 (C-10), and CIFAR-100 (C-100) datasets, demonstrating the efficiency of the proposed model. Results marked with * indicate networks pre-trained on ImageNet. + +
HdCAEErrors (%)No. Param. (Millions)
Baseline××9.172
Model1×√9.072
Model2√×11.62
Model3√√10.92
+ +Table 2. Ablation for dissimilarity-focused head $(H_{d})$ and contrastive agreement enhancer (CAE) on CIFAR10. + +
Errors (%)TVnNo. Param. (Millions)
Average pooling11.10.3722
Max pooling11.10.4732
Hd10.90.5442
+ +Table 3. Comparison of dissimilarity-focused head $(H_{d})$ , average pooling and maxpooling on CIFAR10. + +much smaller architecture compared to most transformer-based and MLP-based methods. + +# 4.3.Quantitative Ablation + +To study the contribution of the different components of our model to its performance, we conduct an ablation experiment where we incrementally added the dissimilarity-focused head $(H_{d})$ and contrastive agreement enhancer (CAE) to the baseline model. The baseline model is defined by removing both $H_{d}$ and CAE from our full model. We then compare the classification results and the number of parameters on CIFAR10 to demonstrate the contribution of $H_{d}$ and CAE to model performance. + +The results in Tab. 2 indicate that $H_{d}$ can significantly reduce the number of trainable parameters in the model while only causing a slight increase in errors, which suggests the presence of redundancy at the top level. Additionally, when $H_{d}$ is employed, CAM can enhance model performance, as much of the redundancy has been eliminated, allowing CAM to provide valuable supplementary information. + +To evaluate the impact of $H_{d}$ on classification accuracy and $TV_{n}$ , we compare our model against two alternative downsampling strategies: max pooling and average pooling, by substituting $H_{d}$ with these methods. Results in Tab. 3 show that the model with $H_{d}$ achieves a lower error rate and higher $TV_{n}$ than these alternatives under identical downsampling conditions. The higher $TV_{n}$ with $H_{d}$ stems from its emphasis on embeddings with lower similarity, which carry more informative features and enhance classification accuracy. + +# 4.4. Qualitative Ablation + +We also conduct two qualitative ablation study to demonstrate the effect of the contrastive agreement enhancer. + +The emerging island of agreement. To illustrate the emergence of islands across different levels, the embeddings from various $k$ levels of CIFAR-10 are visualized in Fig. 6. When examining the embeddings at the highest level ( $k = 5$ ), those generated by the model with CAE reveal finer details and a more distinct separation between objects and the background. Furthermore, a comparison of embeddings across different levels demonstrates that as the levels increase, the embeddings progressively reach agreement, effectively capturing part-whole hierarchies. + +The representation visualization using arrows. To provide an intuitive illustration of the enhancement in + +![](images/a1cd22bcf16e1eb6cf283642755a31048916e79c227343e383237bf20f3395d3.jpg) +Figure 6. The emerging island of agreement at various $k$ levels from CIFAR-10. At the top level $(k = 5)$ , the representations generated by the model with CAE display enhanced detail, highlighting a more distinct separation between objects and the background. + +![](images/990584779435dc1e70190d7aef758a33c15837f4712c84560f3234a7e0e55f2b.jpg) +Figure 7. The representation visualization using arrows for model with and without CAE. The direction and length of each arrow represent the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator. + +agreement introduced by CAE, we visualize the representations using arrows across various locations and levels. The direction and length of each arrow indicate the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator. MNIST is selected for visualization due to its simplicity, minimizing potential interference from complex image backgrounds. As depicted in Fig. 7, the arrows in the model incorporating CAE demonstrate greater consistency, whereas those in the model without CAE appear more disordered and inconsistent. + +# 5. Conclusion + +In this paper, we have proposed a new implementation of GLOM that aligns more closely with its theoretical framework for forming hierarchies of grouped embeddings to rep- + +resent nodes in a parse tree. To foster the formation of distinct islands, we have designed a novel contrastive agreement enhancer that encourages agreement between positive embedding pairs while separating negative pairs. To reduce the computational burden of GLOM, we developed a dissimilarity-focused head that preserves more informative content by prioritizing embeddings with lower similarity, which often contain richer information. Both quantitative and qualitative analyses demonstrate that the contrastive agreement enhancer effectively promotes islands of agreement, while the dissimilarity-focused head enhances model efficiency across various datasets. This work significantly improves the practicality of GLOM, and we hope it inspires other researchers to further explore this direction to advance its development. + +# References + +[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. Publisher Copyright: © 2021 Owner/Author. 1 +[2] Ran Chen, Hao Shen, Zhong-Qiu Zhao, Yi Yang, and Zhao Zhang. Global routing between capsules. Pattern Recognition, 148:110142, 2024. 3 +[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[4] Jaewoong Choi, Hyun Seo, Suii Im, and Myungjoo Kang. Attention routing between capsules. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 3 +[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1 +[6] Laura Culp, Sara Sabour, and Geoffrey E Hinton. Testing glom's ability to infer wholes from ambiguous parts. arXiv preprint arXiv:2211.16564, 2022. 1, 2, 3 +[7] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7 +[8] Nicola Garau, Niccolò Bisagno, Zeno Sambugaro, and Nicola Conc. Interpretable part-whole hierarchies and conceptual-semantic relationships in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13689-13698, 2022. 1, 3, 5, 6, 7 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7 +[10] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. Neural Computation, 35(3):413-452, 2023. 1, 2, 3, 6 +[11] Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing. In International conference on learning representations, 2018. 2, 3, 6, 7 +[12] Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933. 1 +[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[15] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-104. IEEE, 2004. 6 + +[16] Jiachen Li, Ali Hassani, Steven Walton, and Humphrey Shi. Convmlp: Hierarchical convolutional mlp's for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6307-6316, 2023. 3, 7 +[17] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3688-3704, 2022. 3 +[18] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, pages 1-14, 2024. 6, 7 +[19] Vittorio Mazzia, Francesco Salvetti, and Marcello Chi-aberge. Efficient-capsnet: Capsule network with self-attention routing. Scientific reports, 11(1):14634, 2021. 3, 6, 7 +[20] Rita Pucci, Christian Micheloni, and Niki Martinel. Self-attention agreement among capsules. In Proceedings of the iee//cvf international conference on computer vision, pages 272-280, 2021. 3 +[21] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3749-3756, 2020. 6, 7 +[22] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 2, 6 +[23] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 2, 3, 6, 7 +[24] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6, 7 +[25] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021. 7 +[26] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 6 +[27] Ru Zeng and Yan Song. A fast routing capsule network with improved dense blocks. IEEE Transactions on Industrial Informatics, 18(7):4383-4392, 2021. 3 +[28] Ru Zeng, Yuzhang Qin, and Yan Song. A non-iterative capsule network with interdependent agreement routing. Expert Systems with Applications, 238:122284, 2024. 6, 7 +[29] Ru Zeng, Yan Song, and Yuzhang Qin. Spatial attention-based capsule networks with guaranteed group equivariance. IEEE Transactions on Automation Science and Engineering, 2024. 2 +[30] Ru Zeng, Yan Song, and Yanjiu Zhong. An interpretable unsupervised capsule network via comprehensive contrastive learning and two-stage training. Pattern Recognition, 158: 111059, 2025. 2 \ No newline at end of file diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/images.zip b/2025/Agreement aware and dissimilarity oriented GLOM/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..290b961b6096d5c72dff3aaf323c8e1c9d7ce6ee --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96ef2acf7078d6dc7291eac3c983b7e7077bbf47b82fcbbf1f9c511c5a358c75 +size 728333 diff --git a/2025/Agreement aware and dissimilarity oriented GLOM/layout.json b/2025/Agreement aware and dissimilarity oriented GLOM/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..758262ddf65cc0495a6bdfc6f524f5c925738ba7 --- /dev/null +++ b/2025/Agreement aware and dissimilarity oriented GLOM/layout.json @@ -0,0 +1,9669 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 144, + 103, + 467, + 121 + ], + "type": "text", + "content": "Agreement aware and dissimilarity oriented GLOM" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "spans": [ + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "\\mathrm{RuZeng^{1}}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Yan Song" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Yang Zhang" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Yanling Hu" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Hui Yu" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 158, + 432, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 158, + 432, + 172 + ], + "spans": [ + { + "bbox": [ + 179, + 158, + 432, + 172 + ], + "type": "text", + "content": "1University of Shanghai for Science and Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 186 + ], + "type": "text", + "content": "2Shanghai Urban and Rural Construction and Transportation Development Institute" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "spans": [ + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "type": "text", + "content": "Guangxi Medical University " + }, + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 163, + 186, + 446, + 200 + ], + "type": "text", + "content": "University of Glasgow" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 252, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 540 + ], + "type": "text", + "content": "GLOM, an innovative departure from standard deep learning architectures, has been proposed and gained special concern recently due to its good interpretability in representing part-whole relationships in computer vision. However, GLOM faces challenges in achieving agreement and is usually computationally demanding. First, current implementations struggle to produce identical vectors that reliably converge to represent nodes in a parse tree. Second, GLOM is computationally intensive due to the need to maintain equal resolution across all levels. To address these issues, inspired by contrastive learning, we proposed a contrastive agreement enhancer (CAE), which effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby facilitating forming distinct \"islands.\" Furthermore, we introduce a dissimilarity-focused head " + }, + { + "bbox": [ + 55, + 252, + 296, + 540 + ], + "type": "inline_equation", + "content": "(H_{d})" + }, + { + "bbox": [ + 55, + 252, + 296, + 540 + ], + "type": "text", + "content": " to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window. The results of comparison experiments indicate that the proposed approach delicately retains informative content and significantly reduces the number of parameters. Additionally, the ablation experiments and visualization results demonstrate that CAE successfully promotes islands of agreement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 565, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 682 + ], + "type": "text", + "content": "As deep neural networks continue to achieve impressive performance across various tasks, there is increasing focus on model interpretability, particularly in the domain of computer vision [1]. GLOM was proposed to address this need, offering a promising solution to the challenge of representing part-whole hierarchies in neural networks [10]. It integrates the strengths of transformers, neural fields, contrastive representation learning, distillation, and capsules," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 230, + 411, + 317 + ], + "blocks": [ + { + "bbox": [ + 320, + 230, + 411, + 317 + ], + "lines": [ + { + "bbox": [ + 320, + 230, + 411, + 317 + ], + "spans": [ + { + "bbox": [ + 320, + 230, + 411, + 317 + ], + "type": "image", + "image_path": "cd4113201db0508a242d8f805c009ab27f570726652f06670d9a7df69b720134.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 323, + 527, + 332 + ], + "lines": [ + { + "bbox": [ + 342, + 323, + 527, + 332 + ], + "spans": [ + { + "bbox": [ + 342, + 323, + 527, + 332 + ], + "type": "text", + "content": "(a) An illustration for the contrastive agreement enhancer" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 411, + 231, + 459, + 312 + ], + "blocks": [ + { + "bbox": [ + 411, + 231, + 459, + 312 + ], + "lines": [ + { + "bbox": [ + 411, + 231, + 459, + 312 + ], + "spans": [ + { + "bbox": [ + 411, + 231, + 459, + 312 + ], + "type": "image", + "image_path": "c9edbc169916ea4572e74b1f9ba96b04b0222cb321f04c95c298c2c72dfa1b66.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 459, + 231, + 550, + 317 + ], + "blocks": [ + { + "bbox": [ + 459, + 231, + 550, + 317 + ], + "lines": [ + { + "bbox": [ + 459, + 231, + 550, + 317 + ], + "spans": [ + { + "bbox": [ + 459, + 231, + 550, + 317 + ], + "type": "image", + "image_path": "a226053f987e1aec2e39f1505645162ffff982861fea80827c69d069d9718035.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 337, + 426, + 421 + ], + "blocks": [ + { + "bbox": [ + 320, + 337, + 426, + 421 + ], + "lines": [ + { + "bbox": [ + 320, + 337, + 426, + 421 + ], + "spans": [ + { + "bbox": [ + 320, + 337, + 426, + 421 + ], + "type": "image", + "image_path": "ad046ef7d996a2d7beca4644516e0a586e9852dfa99ff8b626b01e737035666a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 426, + 414, + 437 + ], + "lines": [ + { + "bbox": [ + 332, + 426, + 414, + 437 + ], + "spans": [ + { + "bbox": [ + 332, + 426, + 414, + 437 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 332, + 426, + 414, + 437 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 332, + 426, + 414, + 437 + ], + "type": "text", + "content": " for bottom level" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 443, + 336, + 549, + 421 + ], + "blocks": [ + { + "bbox": [ + 443, + 336, + 549, + 421 + ], + "lines": [ + { + "bbox": [ + 443, + 336, + 549, + 421 + ], + "spans": [ + { + "bbox": [ + 443, + 336, + 549, + 421 + ], + "type": "image", + "image_path": "ad3a0e77b2fe2e42d8426e729a6e19ea3dfdf33be3ef2119e73a03d1c2e85437.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 426, + 531, + 437 + ], + "lines": [ + { + "bbox": [ + 461, + 426, + 531, + 437 + ], + "spans": [ + { + "bbox": [ + 461, + 426, + 531, + 437 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 461, + 426, + 531, + 437 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 461, + 426, + 531, + 437 + ], + "type": "text", + "content": " for top level" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "type": "text", + "content": "Figure 1. a): To tackle the challenge in current GLOM implementations, where vectors struggle to form islands representing parts and wholes and fail to grow larger islands at higher levels, we introduce a contrastive agreement enhancer. This approach effectively promotes agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct islands. By adjusting the threshold for distinguishing positive and negative pairs, we ensure increased smoothness at higher levels that indicate the island grows larger. b) and c): Total variance of normalized embeddings " + }, + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "type": "inline_equation", + "content": "(TV_{n})" + }, + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "type": "text", + "content": " is used as a metric to quantitatively measure smoothness (with lower " + }, + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 446, + 555, + 590 + ], + "type": "text", + "content": " indicating greater smoothness). Our method enhances smoothness at both the bottom and top levels, with the top level achieving greater smoothness." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "garnering significant attention upon its introduction. At its core, GLOM uses \"islands\"—groups of similar feature vectors dynamically identified as coherent units, such as parts of an object—to form nodes in a parse tree. However, current implementations [6, 8] of GLOM struggle to produce similar vectors that reliably reach agreement to form these islands. Moreover, GLOM models are computationally intensive due to the need to maintain equal resolution across" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 209, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 209, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 209, + 704 + ], + "type": "text", + "content": "*Corresponding author: sonya@usst.edu.cn" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24351" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 109 + ], + "type": "text", + "content": "all levels. These limitations inspire the enhancements proposed in this paper, which seek to address the vector agreement problem and improve computational efficiency." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 111, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 111, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 56, + 111, + 295, + 433 + ], + "type": "text", + "content": "In the GLOM system, embeddings at each level are feature vectors that gradually converge to form distinct islands of nearly identical vectors, representing parts like the head, body, and tail, as shown in the right panel of Fig. 1a. To achieve this, GLOM incorporates a regularizer [6] that minimizes the cosine distance between the bottom-up predictions from lower-level embeddings and the top-down predictions from higher-level embeddings at adjacent levels. However, a practical challenge arises due to the need to balance multiple regularizers between various levels, making it difficult to properly manage their interactions. Another option is a regularizer that encourages similarity between embeddings of nearby locations, but this can lead to the collapse of representations. Although contrastive learning, which commonly uses negative samples to prevent collapse [3], could be a solution, it introduces the challenge of producing distinct islands and recognizing objects, as it relies on finding suitable negative samples to differentiate these islands. Additionally, as the islands grow larger at higher levels to represent whole entities like birds, the embeddings should become more similar, resulting in increased smoothness across the feature map. We use total variance as a metric to quantitatively measure this smoothness. As illustrated by the blue line in Fig. 1b and Fig. 1c, we observe that the total variance [22] of GLOM's normalized embeddings is higher at higher levels than at lower levels, which contradicts GLOM's theoretical expectations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 590 + ], + "type": "text", + "content": "To address these two issues, we propose a contrastive agreement enhancer (CAE), as shown in Fig. 1a that computes the cosine similarity between two embeddings and determines whether they belong to a positive or negative pair based on a predefined threshold. This mechanism, inspired by contrastive learning, promotes similarity between positive pairs while pushing apart negative pairs. CAE offers several advantages: 1) it can be easily integrated into every level of GLOM; 2) it is not a regularizer, eliminating the need to balance multiple regularizers across levels; 3) it effectively promotes embedding agreement to form distinct islands; and 4) by adjusting the threshold, it ensures increased smoothness at higher levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Another focus of this paper is mitigating the computational challenges in GLOM, which arise from maintaining equal resolution across all levels. Specifically, when implementing training or downstream tasks, different heads are customized and added to the top of GLOM. If these heads receive the entire top-level embeddings as input, the computational burden becomes significant because the top-level resolution is as high as the bottom level. A straightforward solution is to downsample the embeddings, based on the premise that redundancy exists within it. Redundancy is in-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "deed present at the top level, as the embeddings there have reached agreement and become highly similar. To handle this, we design a dissimilarity-focused head " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "(H_{d})" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": ", which is integrated at the top of GLOM. " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": " similarly to common pooling methods, where a small rectangular window slides across the input feature map. For each window position, " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": " calculates the similarity of each embedding with the others within the window. The weights for each embedding are then computed based on the negative correlation of these similarities. The output of " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": " is the weighted sum of the embeddings. The advantages of " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": " include significantly reducing the parameters and computational load of the heads. Moreover, compared to max pooling, " + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": " preserves more informative content by focusing on embeddings with lower similarity, which tend to contain more useful information—this is an improvement over max pooling, which just discards all but the highest-value embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "text", + "content": "Overall, the main contributions of this work can be summarized as follows. 1) A new implementation of GLOM is proposed that aligns more closely with GLOM's theoretical framework for forming hierarchies of grouped embeddings to represent nodes in a parse tree. 2) A novel contrastive agreement enhancer is designed to promote agreement between positive embedding pairs while pushing apart negative pairs, thereby forming distinct \"islands.\" 3) A dissimilarity-focused head is developed to reduce redundancy in the top-level embeddings, preserving more informative content by concentrating on embeddings with lower similarity, which tend to contain more useful information. 4) Quantitative and qualitative analyses show that CAE effectively promotes islands of agreement and " + }, + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "text", + "content": " contributes greatly to improving model efficiency across datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 466, + 399, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 466, + 399, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 466, + 399, + 479 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": "Capsule networks and GLOM: GLOM provides a more efficient approach to representing part-whole hierarchies than capsule networks (CapsNets) for several reasons [10]. First, CapsNets often face inefficiencies in part representation due to the need to pre-allocate neurons for possible parts at specific locations, leading to underutilized capsules during inference [29, 30]. In contrast, GLOM leverages all neurons to form clusters, facilitating knowledge sharing across locations and improving part representation efficiency. Second, conventional CapsNets require presetting the number of clusters to represent higher-level parts. For instance, studies like [23] and [11] preset the number of capsule types and slice the feature map along the channel direction to form capsules. GLOM, however, allows embeddings to naturally form islands without the need for predefined clusters. Lastly, GLOM eliminates dynamic routing by enabling each part's location to independently construct its own vector representation of the whole, avoiding the need to route information to specific capsules. How" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24352" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 191 + ], + "type": "text", + "content": "ever, GLOM is still in its early stages of development and has its own limitations. For example, it primarily functions as an imagery model, with few implementations available to thoroughly evaluate its performance [6, 8]. Current implementations struggle to achieve proper agreement between embeddings, which hampers the formation of islands in GLOM. Additionally, embeddings are set at the same resolution across both low and high levels, leading to redundancy in the model's trainable parameters. This paper focuses on addressing these shortcomings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "text", + "content": "Routing and embedding updating: Routing algorithms are a crucial aspect of CapsNets, intended to replace the pooling algorithms typically used in CNNs by assessing the agreement between capsules [23]. This agreement reflects how much lower-level capsules support the existence of a higher-level capsule, based on the similarity of their predictions [11]. Many works about CapsNets focus on reducing the high computation burden in the capsule routing algorithm. For instance, [27] eliminates the iterations in routing and uses capsule distribution variance to activate high-level capsules. Additionally, [2, 4, 17, 19, 20] introduce efficient attention mechanisms into routing, significantly reducing computational overhead. In contrast, GLOM eliminates the need for routing but depends on similar embeddings to iteratively update and form 'islands' representing parts or whole objects [10]. Both routing and embedding updates aim to use similar capsules/embeddings to represent nodes in a sparse tree, organizing objects into a part-whole hierarchy. A key focus of this paper is the role of embedding similarity, which informs the design of the proposed CAE and " + }, + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "text", + "content": ", each serving different purposes. CAE improves GLOM's island formation by promoting agreement between similar embeddings and disagreement between dissimilar ones, while " + }, + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 57, + 193, + 297, + 514 + ], + "type": "text", + "content": " reduces computational overhead at higher levels by selecting embeddings that are less similar to others, suggesting they may contain more valuable information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 524, + 112, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 112, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 112, + 537 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 628 + ], + "type": "text", + "content": "Our proposed model aims to efficiently implement the GLOM architecture while more closely aligning with its theoretical framework for creating hierarchical islands of embeddings that represent nodes in a parse tree, with these islands expanding at higher levels. Additionally, the model is designed to be lightweight, facilitating broader potential applications of the GLOM architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": "Here, we start with the introduction of the mathematical notation needed to explain the details of the main components of the architecture. Given an input image " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": ", it is spatially transformed into " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "N = h \\times w" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": " patches. The " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": "-th patch corresponds to column " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "C_n(h,w)" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": " at the spatial location " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "n \\in \\{1,\\dots,N\\}" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": ". For each column, it consists of " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": "-level embeddings " + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "inline_equation", + "content": "f_{(h,w)}^{k,t}, k \\in \\{1,\\dots,K\\}" + }, + { + "bbox": [ + 55, + 628, + 295, + 717 + ], + "type": "text", + "content": " at" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "time step " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, T\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " with a size of " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " for each embedding. The subscript " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and superscript " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " are omitted in subsequent instances for readability. " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "f^k" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "f^{k+1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " are consecutive and represent the part and whole respectively. They are connected by a bottom-up encoder and a top-down decoder which are shared across the spatial dimension." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 150, + 428, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 428, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 428, + 163 + ], + "type": "text", + "content": "3.1. Patches embedding" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": "The pipeline is demonstrated in Fig. 2. Firstly, input image " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ", with the shape of " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "H \\times W \\times c" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ", is transformed into patch embeddings. Following the recipe provided in [16], a convolutional Tokenizer is used to extract feature maps with the size of " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "H' \\times W' \\times c'" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ". Next, it is reshaped to the size of " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "(h \\times r) \\times (w \\times r) \\times c'" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": " is the patch size, " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "H' = h \\times r" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "W' = w \\times r" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ". Then, we reshape and permute it with the size of " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "h \\times w \\times d" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "d = r \\times r \\times c'" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": ". At last, the " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": "-dimensional features act as the embeddings " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "f^{1,t}" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": " at the bottom level in columns " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 167, + 555, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 436, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 436, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 436, + 306 + ], + "type": "text", + "content": "3.2. Embedding updating" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "text", + "content": "Secondly, we introduce the propagation phase, which governs the interactions between levels and across locations. As shown in Fig. 3, the embedding of a column at location " + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "text", + "content": " and level " + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "text", + "content": " is updated by integrating multiple sources of information: predictions from lower and higher levels at the same location " + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 311, + 555, + 453 + ], + "type": "text", + "content": " using a bottom-up encoder and a top-down decoder, predictions from embeddings at other locations within the same level via self-attention and a contrastive agreement enhancer, and the historical context (i.e., the embedding from the previous iteration). Next, we elaborate on the components of the model that drive this embedding update process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": "Bottom-up encoder " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{BU}" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": ": Taking a low-level embedding " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "f^{k,t}" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": " as input, the bottom-up encoder " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{BU}" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": " is used to predict high-level embedding " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "f^{k+1,t+1}" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": " with the same size of " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": ". It is constructed by two fully connected layers and shared across " + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 454, + 554, + 513 + ], + "type": "text", + "content": " columns." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": "Top-down decoder " + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{TD}" + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": ": Given a high-level embedding " + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "inline_equation", + "content": "f^{k + 1,t}" + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": " as input, the top-down decoder " + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{TD}" + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": " is applied to predict the corresponding low-level embedding " + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "inline_equation", + "content": "f^{k,t + 1}" + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": ". The intuition is that if the model understands the whole, it can effectively predict its parts. Like the encoder, " + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{TD}" + }, + { + "bbox": [ + 313, + 514, + 554, + 585 + ], + "type": "text", + "content": " also comprises two fully connected layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "text", + "content": "Self-attention " + }, + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{SA}" + }, + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "text", + "content": ": In GLOM, the interaction between columns is implemented by a simplified self-attention that query, value, and key vector are all the same as input. The aim of self-attention is to produce islands of identical embeddings at a level by making each embedding vector at that level regress towards other similar vectors at nearby locations. " + }, + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{SA}" + }, + { + "bbox": [ + 313, + 586, + 554, + 669 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 684, + 555, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 684, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 354, + 684, + 555, + 712 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V, \\tag {1}", + "image_path": "d4c6a529654166948f711a0bdecd44a3fd217265a957213b893c852f77d62251.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24353" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 553, + 175 + ], + "type": "image", + "image_path": "6a2006584f97e92187f33470dade1fc6b10e63664290beee73c81f33c080f6a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": "Figure 2. Our model is pre-trained using contrastive learning, followed by supervised training with frozen weights. The input images " + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": " are augmented and transformed into patches. These patches are then processed to construct a part-whole hierarchy over " + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": " iterations, where embeddings at different levels gradually converge as the levels ascend. Next, the top-level embeddings are fed into " + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": ", which performs downsampling based on weights that are negatively correlated with similarity. In the subsequent step, the embeddings are sent to " + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "inline_equation", + "content": "H_{1}" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": " for contrastive pre-training. Finally, the output embeddings are passed to " + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "inline_equation", + "content": "H_{2}" + }, + { + "bbox": [ + 54, + 188, + 555, + 244 + ], + "type": "text", + "content": " for supervised training with frozen weights." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "inline_equation", + "content": "Q, K, V = f^{k,t}" + }, + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "text", + "content": ". It should be noted that " + }, + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{SA}" + }, + { + "bbox": [ + 54, + 263, + 294, + 371 + ], + "type": "text", + "content": " alone is insufficient for creating part-whole hierarchies, as it lacks the flexibility to adjust the degree of agreement (e.g., it struggles to encourage embeddings to achieve greater consensus at higher levels). To address this, GLOM incorporates additional regularization. In contrast, our model adopts a more effective strategy, drawing inspiration from contrastive learning, to promote agreement among embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "text", + "content": "Contrastive agreement enhancer " + }, + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{CAE}" + }, + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "text", + "content": ": the function of our proposed CAE is two-fold: enhancing the agreement between positive embedding pairs; and pushing apart the negative embedding pairs. In contrastive learning, positive sample pairs are generated by applying different augmentations to the same image, while negative sample pairs are derived from images belonging to different classes. Unlike contrastive learning, in " + }, + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{CAE}" + }, + { + "bbox": [ + 54, + 372, + 295, + 502 + ], + "type": "text", + "content": ", the classification of embedding pairs as positive or negative is based on the similarity between embeddings. To be specific, the similarity is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 507, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 507, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 91, + 507, + 295, + 536 + ], + "type": "interline_equation", + "content": "S = \\frac {f _ {n 1} f _ {n 2} ^ {T}}{\\| f _ {n 1} \\| \\| f _ {n 2} \\|}, n 1, n 2 \\in \\{1, \\dots , N \\}, \\tag {2}", + "image_path": "922cfa0b5b7080efa600565f08066642e80b3607bd1ad60bd382c2c9474af764.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": "where the size of similarity map " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": ". For positive and negative embedding pairs, we could set different thresholds. Here we take a simple case for example with " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "S = 0" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " as a threshold for both positive and negative embedding pairs. " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "S > 0" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " means that the angle between embeddings " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n1}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n2}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " lies within the range " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "(0^{\\circ}, 90^{\\circ})" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": ". In this case, " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n1}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n2}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " is a positive embedding pair and could belong to the same island. Conversely, " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "S < 0" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " indicates that " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n1}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "inline_equation", + "content": "f_{n2}" + }, + { + "bbox": [ + 55, + 540, + 295, + 659 + ], + "type": "text", + "content": " may belong to different islands thereby being a negative pair. This process can be described as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 666, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 666, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 94, + 666, + 295, + 680 + ], + "type": "interline_equation", + "content": "S _ {p} = S \\cdot \\mathbb {I} (S \\geq 0) + (- \\infty) \\cdot \\mathbb {I} (S < 0), \\tag {3}", + "image_path": "0ebcd117c31ee94964e0fa3fa716f133caa98accab0122fd5a9f2281fb7b12bd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 681, + 294, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 681, + 294, + 694 + ], + "spans": [ + { + "bbox": [ + 94, + 681, + 294, + 694 + ], + "type": "interline_equation", + "content": "S _ {e} = S \\cdot \\mathbb {I} (S < 0), \\tag {4}", + "image_path": "659daf843ba314c14f42400e0ca59d590c0a29447a02ec7946dd8abe710b2405.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "S_{p}" + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "S_{e}" + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": " represent the similarity map for positive" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": "and negative embedding pairs respectively. " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": " is the indicator function. In the model's bottom level, the threshold is set to 0, while at the top level it is set to " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "-0.2" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ". Next, we compute the embeddings at the time step " + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 264, + 555, + 312 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 376, + 318, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 318, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 376, + 318, + 553, + 333 + ], + "type": "interline_equation", + "content": "f _ {p} ^ {t + 1} = \\operatorname {s o f t m a x} \\left(S _ {p}\\right) f ^ {t}, \\tag {5}", + "image_path": "99345829cd10fc341a1324075d73224d2cd4658465550fa6a7fc4765d71d24ef.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 335, + 553, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 335, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 377, + 335, + 553, + 348 + ], + "type": "interline_equation", + "content": "f _ {e} ^ {t + 1} = S _ {e} f ^ {t}, \\tag {6}", + "image_path": "6776e8333a08cad4c4b5c1807a8ba39b636cbc491d2a950bd3df1e4f1ec2b527.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 350, + 553, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 350, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 376, + 350, + 553, + 366 + ], + "type": "interline_equation", + "content": "\\mathcal {A} _ {C A E} \\left(f ^ {t}\\right) = f _ {p} ^ {t + 1} - f _ {e} ^ {t + 1}, \\tag {7}", + "image_path": "dc341ee1e54a47f850181b5c6d94778ce2f5d4eb66293b9b216b0df5086235c8.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{p}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": " are the positive embeddings and " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{e}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": " are the negative embeddings. Compared with " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{e}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": ", the computation of " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{p}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": " evolves the softmax function since the main purpose is to facilitate agreement between embeddings and softmax is good at eliminating the effect of negative embeddings (" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "S < 0" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": "). In addition, to push apart between positive and negative embeddings, we minus " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{p}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f_{e}^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": " to predict the " + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "inline_equation", + "content": "f^{t + 1}" + }, + { + "bbox": [ + 313, + 370, + 554, + 467 + ], + "type": "text", + "content": ". An illustration for CAE is given in Fig. 1a." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "content": "Fusion: At each time step " + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "inline_equation", + "content": "f^{k,t}" + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "content": " is computed by the fusion of predictions from " + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{BU}, \\mathcal{N}_{TD}, \\mathcal{A}_{SA}" + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{CAE}" + }, + { + "bbox": [ + 313, + 467, + 554, + 491 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 508, + 537, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 508, + 537, + 523 + ], + "spans": [ + { + "bbox": [ + 321, + 508, + 537, + 523 + ], + "type": "interline_equation", + "content": "f ^ {k, t} = w _ {b u} \\mathcal {N} _ {B U} \\left(f ^ {k - 1, t - 1}\\right) + w _ {t d} \\mathcal {N} _ {T D} \\left(f ^ {k + 1, t - 1}\\right) +", + "image_path": "31f008c6730edaa88b5fc5d885d34a602afd897a4ba46efdfb853da9c6d63ec7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 525, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 525, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 320, + 525, + 553, + 540 + ], + "type": "interline_equation", + "content": "w _ {l a} f ^ {k, t - 1} + w _ {a s} \\mathcal {A} _ {S A} \\left(f ^ {k, t - 1}\\right) + w _ {a c} \\mathcal {A} _ {C A E} \\left(f ^ {k, t - 1}\\right). \\tag {8}", + "image_path": "fa2f713bf70d3c60d6d6b273566da0f33a1d91a7f1fd8ea0faff0b9eceb2d15d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 539, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 539, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 539, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 539, + 558 + ], + "type": "inline_equation", + "content": "w_{bu}, w_{td}, w_{la}, w_{as}, w_{ac}" + }, + { + "bbox": [ + 313, + 546, + 539, + 558 + ], + "type": "text", + "content": " are trainable parameters." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 564, + 462, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 462, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 462, + 577 + ], + "type": "text", + "content": "3.3. Dissimilarity-focused head" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "text", + "content": "With the aim to relieve the computation overhead introduced by which the embeddings at the top level and the lower level have the same resolution, we proposed a dissimilarity-focused head " + }, + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 582, + 554, + 653 + ], + "type": "text", + "content": " which receives top-level embeddings as input to do downsampling as shown in Fig. 4, and it greatly reduces the number of parameters." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": " is inspired by the idea that embeddings at the top level should exhibit agreement, which can result in redundancy between them. To quantify this redundancy, we apply indirect thinking and consider similarity as a suitable metric, since the more similar the embeddings are, the more" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24354" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 83, + 291, + 301 + ], + "blocks": [ + { + "bbox": [ + 67, + 83, + 291, + 301 + ], + "lines": [ + { + "bbox": [ + 67, + 83, + 291, + 301 + ], + "spans": [ + { + "bbox": [ + 67, + 83, + 291, + 301 + ], + "type": "image", + "image_path": "7ecde751620dfe32aa975dd6a3b8e400ee86c330de3e3a885392efdbf628b949.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 322, + 295, + 355 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 295, + 355 + ], + "type": "text", + "content": "Figure 3. Architecture of model. The column at location " + }, + { + "bbox": [ + 55, + 322, + 295, + 355 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 322, + 295, + 355 + ], + "type": "text", + "content": " is updating with a fusion of bottom-up and top-down prediction, history, SA, and CAE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 372, + 292, + 462 + ], + "blocks": [ + { + "bbox": [ + 61, + 372, + 292, + 462 + ], + "lines": [ + { + "bbox": [ + 61, + 372, + 292, + 462 + ], + "spans": [ + { + "bbox": [ + 61, + 372, + 292, + 462 + ], + "type": "image", + "image_path": "cefc780b066a9068d8af1a13d2da4604e5de505f7ee117018bf2aa43786ffe48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 472, + 295, + 517 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 517 + ], + "type": "text", + "content": "Figure 4. An illustration for the dissimilarity-focused head which is designed to reduce redundancy in the top-level embeddings, where embedding weights for downsampling are negatively correlated with similarity within a sliding window." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": "redundant they become. To be specific, given the top embeddings " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "f^{K} \\in \\mathbb{R}^{h \\times w \\times d}" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": ", we first transform them into " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "p \\times p" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": " patches each with the shape of " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "h' \\times w' \\times d" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "h' = \\frac{h}{p}, w' = \\frac{w}{p}" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": ". For each patch, similarity map " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "S'" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": " with the shape of " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "(h' \\times w') \\times (h' \\times w')" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": " is computed using Eq. 2. Next, we compute the weights " + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 537, + 295, + 622 + ], + "type": "text", + "content": " for each embedding within the patch:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 631, + 266, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 631, + 266, + 666 + ], + "spans": [ + { + "bbox": [ + 84, + 631, + 266, + 666 + ], + "type": "interline_equation", + "content": "W _ {i} = \\sum_ {j = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} e ^ {1 - S _ {i, j} ^ {\\prime}}, i, j \\in \\{1, \\dots , h ^ {\\prime} \\times w ^ {\\prime} \\},", + "image_path": "761c3e4919b6a067824ebcdbc4850fd3eff661f30813ede7953d59915e78679c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 667, + 294, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 667, + 294, + 693 + ], + "spans": [ + { + "bbox": [ + 84, + 667, + 294, + 693 + ], + "type": "interline_equation", + "content": "W = \\frac {W}{\\operatorname* {m a x} \\left(W _ {i}\\right)}, W \\in \\mathbb {R} _ {> 0} ^ {h ^ {\\prime} \\times w ^ {\\prime}}, \\tag {9}", + "image_path": "1c00d1336915fe96dbf99d90510b7a684c4765bb0c6876ce245664d188b77462.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 700, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 700, + 295, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 700, + 295, + 715 + ], + "type": "inline_equation", + "content": "1 - S_{i,j}^{\\prime}" + }, + { + "bbox": [ + 55, + 700, + 295, + 715 + ], + "type": "text", + "content": " means that embeddings with higher simi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "larity with others should be assigned with lower weight to downsample. Besides, the weights are summed to measure the overall weight for embedding " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". Finally, the downsampled output is computed by the weighted sum within each patch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 384, + 138, + 554, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 138, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 384, + 138, + 554, + 172 + ], + "type": "interline_equation", + "content": "H _ {d} \\left(f ^ {K}\\right) = \\sum_ {i = 1} ^ {h ^ {\\prime} \\times w ^ {\\prime}} W _ {i} f _ {i}, \\tag {10}", + "image_path": "8de7063854ad18566e518f7d13a4f826b1e09977475eba6fe9662e421483ce3e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 178, + 426, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 426, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 426, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 178, + 426, + 192 + ], + "type": "inline_equation", + "content": "H_{d}(f^{K})\\in \\mathbb{R}^{p\\times p\\times d}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 198, + 378, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 378, + 211 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 378, + 211 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 216, + 554, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 554, + 263 + ], + "type": "text", + "content": "The training of our model follows the setup in [8], consisting of two steps: 1) a pre-training phase using a supervised contrastive loss function, and 2) a classification training phase using cross-entropy loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": "For pre-training, various data augmentation is adopted to each image " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " within a batch " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": ", producing pairs of image " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "I_{b}, a, b \\in \\{1, \\dots, 2B\\}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": ". Then, the image pairs are fed to our model to produce representation " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "C_{n,a}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "C_{n,b} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "n \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " at the top level " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": ". In the following, we reshape them to the shape of " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " and send them to a contrastive head " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "H_{1}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " which is constructed with two fully connected layers. The output is denoted as " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "O_{a}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "inline_equation", + "content": "O_{b}" + }, + { + "bbox": [ + 313, + 264, + 554, + 382 + ], + "type": "text", + "content": ". Finally, a supervised contrastive loss is adopted to pre-train the model:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 359, + 387, + 553, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 387, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 359, + 387, + 553, + 418 + ], + "type": "interline_equation", + "content": "L _ {1} = - \\log \\frac {e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {b}\\right)}}{\\sum_ {i = 1} ^ {2 B} \\mathbb {I} _ {[ i \\neq a ]} e ^ {\\operatorname {s i m} \\left(O _ {a} , O _ {i}\\right)}}, \\tag {11}", + "image_path": "d95fb2f36d18f18f60013d491af6e19ef65577d41f33675733e985f945a9ff2e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{[i\\neq a]}" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": " is an indicator function that values 0 if " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 423, + 554, + 471 + ], + "type": "text", + "content": " belong to the same image class, and 0 otherwise. sim is a cosine similarity function that measures the similarity of normalized input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": "After the model is pre-trained with contrastive loss, we freeze the network weights. Besides, we add a classification head " + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "inline_equation", + "content": "H_{2}" + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": " on the top of contrastive head " + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "inline_equation", + "content": "H_{1}" + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "inline_equation", + "content": "H_{2}" + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": " is a linear layer that projects the representation to the final out with the dimension of " + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 472, + 554, + 543 + ], + "type": "text", + "content": " is the number of classes. The cross-entropy loss is given as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 549, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 549, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 386, + 549, + 553, + 582 + ], + "type": "interline_equation", + "content": "L _ {2} = - \\frac {1}{C} \\sum_ {i = 1} ^ {C} p _ {i} \\log q _ {i}, \\tag {12}", + "image_path": "8e84919c36d195fd9c42ba212431f84e08f6ffb6b890aa826db056215bf86d83.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "content": " is the true probability distribution for class " + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "content": " is the predicted probability for class " + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 587, + 554, + 612 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Our experiments encompass quantitative and qualitative analyses. Quantitatively, we evaluate the formation of islands of agreement using the total variance of normalized embeddings and compare model performance across multiple datasets. An ablation study further examines the contributions of the proposed CAE and " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ". For the qualitative" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24355" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 149, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 149, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 149, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 149, + 144 + ], + "type": "image", + "image_path": "2465713049a8580e2e580014a00e072363a4ab0df0165c2d5af07ec9f9ec78c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 148, + 140, + 159 + ], + "lines": [ + { + "bbox": [ + 69, + 148, + 140, + 159 + ], + "spans": [ + { + "bbox": [ + 69, + 148, + 140, + 159 + ], + "type": "text", + "content": "(a) C-10 bottom level" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 162, + 149, + 232 + ], + "blocks": [ + { + "bbox": [ + 59, + 162, + 149, + 232 + ], + "lines": [ + { + "bbox": [ + 59, + 162, + 149, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 162, + 149, + 232 + ], + "type": "image", + "image_path": "04ff742b0ddf4cc998367916a00cafd2c2cdb91f16a66c98f9fccf43801614a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 237, + 133, + 247 + ], + "lines": [ + { + "bbox": [ + 76, + 237, + 133, + 247 + ], + "spans": [ + { + "bbox": [ + 76, + 237, + 133, + 247 + ], + "type": "text", + "content": "(f) C-10 top level" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 160, + 72, + 250, + 144 + ], + "blocks": [ + { + "bbox": [ + 160, + 72, + 250, + 144 + ], + "lines": [ + { + "bbox": [ + 160, + 72, + 250, + 144 + ], + "spans": [ + { + "bbox": [ + 160, + 72, + 250, + 144 + ], + "type": "image", + "image_path": "a2b8f22f3870a2b2e2bf2382ddc70c3c622d6606ed5f31558cf022e59be48cdd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 148, + 242, + 158 + ], + "lines": [ + { + "bbox": [ + 167, + 148, + 242, + 158 + ], + "spans": [ + { + "bbox": [ + 167, + 148, + 242, + 158 + ], + "type": "text", + "content": "(b) C-100 bottom level" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 160, + 162, + 250, + 232 + ], + "blocks": [ + { + "bbox": [ + 160, + 162, + 250, + 232 + ], + "lines": [ + { + "bbox": [ + 160, + 162, + 250, + 232 + ], + "spans": [ + { + "bbox": [ + 160, + 162, + 250, + 232 + ], + "type": "image", + "image_path": "1c587e6392be397c0e822a8dc17d2e7863396a6970310c003ffb98167fcfba4d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 238, + 236, + 247 + ], + "lines": [ + { + "bbox": [ + 173, + 238, + 236, + 247 + ], + "spans": [ + { + "bbox": [ + 173, + 238, + 236, + 247 + ], + "type": "text", + "content": "(g) C-100 top level" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 260, + 72, + 350, + 144 + ], + "blocks": [ + { + "bbox": [ + 260, + 72, + 350, + 144 + ], + "lines": [ + { + "bbox": [ + 260, + 72, + 350, + 144 + ], + "spans": [ + { + "bbox": [ + 260, + 72, + 350, + 144 + ], + "type": "image", + "image_path": "9134948e8d7e40f4e698c04dc9cb484325eb972329d294a29e64e4eead927ef0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 148, + 347, + 158 + ], + "lines": [ + { + "bbox": [ + 263, + 148, + 347, + 158 + ], + "spans": [ + { + "bbox": [ + 263, + 148, + 347, + 158 + ], + "type": "text", + "content": "(c) S-NORB bottom level" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 260, + 162, + 350, + 232 + ], + "blocks": [ + { + "bbox": [ + 260, + 162, + 350, + 232 + ], + "lines": [ + { + "bbox": [ + 260, + 162, + 350, + 232 + ], + "spans": [ + { + "bbox": [ + 260, + 162, + 350, + 232 + ], + "type": "image", + "image_path": "779546a59c5ba37e82e7e54a77847b268bc543898d2f59f229ffd5a667169e66.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 269, + 237, + 341, + 247 + ], + "lines": [ + { + "bbox": [ + 269, + 237, + 341, + 247 + ], + "spans": [ + { + "bbox": [ + 269, + 237, + 341, + 247 + ], + "type": "text", + "content": "(h) S-NORB top level" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "text", + "content": "Figure 5. The comparison of our model with that of the Agglomerator in terms of " + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "text", + "content": " at both the bottom and top levels across five datasets. It is evident that the total variance of our model is considerably lower than that of the Agglomerator, indicating that the contrastive attention mechanism effectively improves agreement in representing parts, which is essential for the theory of GLOM. Furthermore, the " + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 55, + 257, + 555, + 302 + ], + "type": "text", + "content": " of our model is greater at the bottom level than at the top level, suggesting that agreement tends to increase as the levels rise." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 72, + 451, + 144 + ], + "blocks": [ + { + "bbox": [ + 361, + 72, + 451, + 144 + ], + "lines": [ + { + "bbox": [ + 361, + 72, + 451, + 144 + ], + "spans": [ + { + "bbox": [ + 361, + 72, + 451, + 144 + ], + "type": "image", + "image_path": "f011562337a8388a3b410821b75d9bcdce71b696d0a3cbb8ad0c7bc89c40150e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 148, + 445, + 158 + ], + "lines": [ + { + "bbox": [ + 365, + 148, + 445, + 158 + ], + "spans": [ + { + "bbox": [ + 365, + 148, + 445, + 158 + ], + "type": "text", + "content": "(d) MNIST bottom level" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 361, + 162, + 451, + 232 + ], + "blocks": [ + { + "bbox": [ + 361, + 162, + 451, + 232 + ], + "lines": [ + { + "bbox": [ + 361, + 162, + 451, + 232 + ], + "spans": [ + { + "bbox": [ + 361, + 162, + 451, + 232 + ], + "type": "image", + "image_path": "c058d9ba13e4f98bd6e944a8b31f8f100dbad0143b6a465cbae62cb37cafb810.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 237, + 438, + 247 + ], + "lines": [ + { + "bbox": [ + 373, + 237, + 438, + 247 + ], + "spans": [ + { + "bbox": [ + 373, + 237, + 438, + 247 + ], + "type": "text", + "content": "(i) MNIST top level" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 461, + 72, + 551, + 144 + ], + "blocks": [ + { + "bbox": [ + 461, + 72, + 551, + 144 + ], + "lines": [ + { + "bbox": [ + 461, + 72, + 551, + 144 + ], + "spans": [ + { + "bbox": [ + 461, + 72, + 551, + 144 + ], + "type": "image", + "image_path": "42ec8d94e61af33e2e41839e04789b27dd86d46e26a49aa9a87fde3c7996ba42.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 148, + 549, + 158 + ], + "lines": [ + { + "bbox": [ + 462, + 148, + 549, + 158 + ], + "spans": [ + { + "bbox": [ + 462, + 148, + 549, + 158 + ], + "type": "text", + "content": "(e) F-MNIST bottom level" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 461, + 162, + 550, + 232 + ], + "blocks": [ + { + "bbox": [ + 461, + 162, + 550, + 232 + ], + "lines": [ + { + "bbox": [ + 461, + 162, + 550, + 232 + ], + "spans": [ + { + "bbox": [ + 461, + 162, + 550, + 232 + ], + "type": "image", + "image_path": "547f56353964ef660722ebbf5a75be6da6dd07c30e1768035e5e7012bbbd09db.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 237, + 542, + 247 + ], + "lines": [ + { + "bbox": [ + 469, + 237, + 542, + 247 + ], + "spans": [ + { + "bbox": [ + 469, + 237, + 542, + 247 + ], + "type": "text", + "content": "(j) F-MNIST top level" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": "analysis, we visualize the vector representations of emerging islands across levels. To illustrate CAE's enhancement of agreement intuitively, we depict representations with arrows spanning various locations and levels. Experiments are conducted on the CIFAR-10, CIFAR-100 [13], MNIST [14], FashionMNIST [26], and SmallNORB datasets [15]. The details of the training settings are provided in the supplementary material. The source code is publicly available at: https://github.com/zengru001usst/GLOM" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 437, + 185, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 185, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 185, + 449 + ], + "type": "text", + "content": "4.1. Agreement Evaluation" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "text", + "content": "To verify the effectiveness of the proposed CAE, total variance [22] of the normalized embeddings " + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "inline_equation", + "content": "(TV_{n})" + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "text", + "content": ", as shown in Eq. 13, is employed as a metric to quantitatively measure the smoothness of embeddings at specific levels within the models. The embeddings are normalized, as a trivial solution may occur where total variance is minimized by making every " + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "inline_equation", + "content": "f_{i,j}" + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "text", + "content": " infinitesimal. Lower " + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 55, + 454, + 296, + 562 + ], + "type": "text", + "content": " indicates greater smoothness, meaning that more embeddings are similar and achieve a higher level of agreement." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 578, + 245, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 578, + 245, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 245, + 613 + ], + "type": "interline_equation", + "content": "T V _ {n} = \\frac {1}{N} \\frac {1}{d} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1} ^ {d} \\left(f _ {i, j} ^ {\\prime} - \\mu_ {j}\\right) ^ {2},", + "image_path": "aaef521365ada37b3f4f17d31ce7d7d2bae02ce555f0bb0c5db8a5db179cdd86.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 615, + 179, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 615, + 179, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 179, + 648 + ], + "type": "interline_equation", + "content": "\\mu_ {j} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} f _ {i, j} ^ {\\prime},", + "image_path": "2d7873857c8d9c911b99d6ec02f73e5f4fe238f550e30bb6f745ca98f7f51fcb.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 650, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 650, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 295, + 683 + ], + "type": "interline_equation", + "content": "f _ {i, j} ^ {\\prime} = \\frac {f _ {i , j}}{\\sqrt {\\sum_ {j} f _ {i , j} ^ {2}}}. \\tag {13}", + "image_path": "1fd02db76d4a137d8462939f88bb66dbd0c642ae15add0184c2c437b97ed1112.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "The results presented in Fig. 5 compare the " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " of our model with that of the Agglomerator at both the bottom and" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": "top levels across five datasets. It can be observed that the " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " of our model is significantly lower than that of the Agglomerator, indicating that the CAE effectively enhances agreement in representing parts, which is crucial for the theory of GLOM. Additionally, the " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " of our model is higher at the bottom level compared to the top level, suggesting that agreement increases as the levels ascend. This is consistent with GLOM's framework, where embeddings at lower levels gradually cluster locally to represent parts, while embeddings at higher levels cluster globally to represent the whole. In contrast, the Agglomerator fails to maintain this characteristic." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 479, + 437, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 437, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 437, + 491 + ], + "type": "text", + "content": "4.2. Classification Results" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "To demonstrate the efficiency of the proposed model, we report the classification results for each dataset in terms of error percentage and the number of trainable parameters, as shown in Tab. 1. While Capsule Networks (CapsNets), such as [11, 18, 19, 21, 23, 28], perform well on simpler datasets like smallNORB, MNIST, and FashionMNIST, they struggle to scale to more complex datasets with a higher number of classes, such as CIFAR-100, due to inefficiencies in object representation [10]. Convolutional models, such as [9, 24], maintain strong performance across various datasets but fall short in interpretability. Transformer-based and MLP-based methods achieve state-of-the-art performance on more complex datasets, but they lack testing on smaller datasets. As shown in Tab. 1, our model achieves comparable classification error percentages to CapsNets and Agglomerator [8] on simpler datasets, but with significantly fewer parameters, especially compared to Agglomerator. Moreover, our model requires less training time and has a" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24356" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 552, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 552, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 552, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 552, + 259 + ], + "type": "table", + "html": "
MethodRefBackboneErrors %No. params (Millions)Training Arch.
S-NORBMNISTF-MNISTC-10C-100
E-CapsNet[19]Caps2.540.26---0.2GPU
CapsNet[23]2.700.256.3810.682.006.8GPU
Matrix-CapsNet[11]1.400.446.1411.9-0.3GPU
Capsule VB[21]1.600.305.2011.2-0.2GPU
Res-CapsNet[18]0.550.55-7.62-16.4GPU
IAR-CapsNet[28]1.360.77---0.18GPU
ResNet-110[9]Conv-2.105.106.41*27.76*1.7GPU
VGG[24]-0.326.507.74*28.05*20GPU
ViT-L/16[7]Transf---0.85*6.75*632TPU
ConvMLP-L[16]Conv/MLP---1.40*11.40*43TPU
MLP-Mixer-L/16[25]MLP---1.66*-207TPU
Agglomerator[8]Conv/MLP-/Caps0.010.307.4311.1540.9772GPU
Ours-0.010.305.1010.9041.312GPU
", + "image_path": "5626814249904f8914640cb98c35e153515adabe102c08999d7cac7e327e760e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 81, + 308, + 272, + 391 + ], + "blocks": [ + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 555, + 291 + ], + "type": "text", + "content": "Table 1. Classification results on SmallNORB (S-NORB), MNIST, FashionMNIST (F-MNIST), CIFAR-10 (C-10), and CIFAR-100 (C-100) datasets, demonstrating the efficiency of the proposed model. Results marked with * indicate networks pre-trained on ImageNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 81, + 308, + 272, + 391 + ], + "lines": [ + { + "bbox": [ + 81, + 308, + 272, + 391 + ], + "spans": [ + { + "bbox": [ + 81, + 308, + 272, + 391 + ], + "type": "table", + "html": "
HdCAEErrors (%)No. Param. (Millions)
Baseline××9.172
Model1×√9.072
Model2√×11.62
Model3√√10.92
", + "image_path": "c83b158bd7bb6efc115640c256083f58d477a983ef70f1e384ff37215c80e062.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 75, + 433, + 277, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "type": "text", + "content": "Table 2. Ablation for dissimilarity-focused head " + }, + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "type": "inline_equation", + "content": "(H_{d})" + }, + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "type": "text", + "content": " and contrastive agreement enhancer (CAE) on CIFAR10." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 433, + 277, + 502 + ], + "lines": [ + { + "bbox": [ + 75, + 433, + 277, + 502 + ], + "spans": [ + { + "bbox": [ + 75, + 433, + 277, + 502 + ], + "type": "table", + "html": "
Errors (%)TVnNo. Param. (Millions)
Average pooling11.10.3722
Max pooling11.10.4732
Hd10.90.5442
", + "image_path": "7434e62d9ccdbd13958ce1cf6ba2b49abc5cb552703372c12c2309de0ff6224f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "Table 3. Comparison of dissimilarity-focused head " + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "inline_equation", + "content": "(H_{d})" + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": ", average pooling and maxpooling on CIFAR10." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "much smaller architecture compared to most transformer-based and MLP-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 180, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 180, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 180, + 601 + ], + "type": "text", + "content": "4.3.Quantitative Ablation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "To study the contribution of the different components of our model to its performance, we conduct an ablation experiment where we incrementally added the dissimilarity-focused head " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "(H_{d})" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and contrastive agreement enhancer (CAE) to the baseline model. The baseline model is defined by removing both " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and CAE from our full model. We then compare the classification results and the number of parameters on CIFAR10 to demonstrate the contribution of " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and CAE to model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "content": "The results in Tab. 2 indicate that " + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "content": " can significantly reduce the number of trainable parameters in the model while only causing a slight increase in errors, which suggests the presence of redundancy at the top level. Additionally, when " + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "content": " is employed, CAM can enhance model performance, as much of the redundancy has been eliminated, allowing CAM to provide valuable supplementary information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": "To evaluate the impact of " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " on classification accuracy and " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": ", we compare our model against two alternative downsampling strategies: max pooling and average pooling, by substituting " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " with these methods. Results in Tab. 3 show that the model with " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " achieves a lower error rate and higher " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " than these alternatives under identical downsampling conditions. The higher " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "TV_{n}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "inline_equation", + "content": "H_{d}" + }, + { + "bbox": [ + 313, + 396, + 556, + 516 + ], + "type": "text", + "content": " stems from its emphasis on embeddings with lower similarity, which carry more informative features and enhance classification accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 526, + 433, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 433, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 433, + 539 + ], + "type": "text", + "content": "4.4. Qualitative Ablation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 567 + ], + "type": "text", + "content": "We also conduct two qualitative ablation study to demonstrate the effect of the contrastive agreement enhancer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": "The emerging island of agreement. To illustrate the emergence of islands across different levels, the embeddings from various " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " levels of CIFAR-10 are visualized in Fig. 6. When examining the embeddings at the highest level (" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": "), those generated by the model with CAE reveal finer details and a more distinct separation between objects and the background. Furthermore, a comparison of embeddings across different levels demonstrates that as the levels increase, the embeddings progressively reach agreement, effectively capturing part-whole hierarchies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "The representation visualization using arrows. To provide an intuitive illustration of the enhancement in" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24357" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 549, + 289 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 549, + 289 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 549, + 289 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 549, + 289 + ], + "type": "image", + "image_path": "a1cd22bcf16e1eb6cf283642755a31048916e79c227343e383237bf20f3395d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "type": "text", + "content": "Figure 6. The emerging island of agreement at various " + }, + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "type": "text", + "content": " levels from CIFAR-10. At the top level " + }, + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "type": "inline_equation", + "content": "(k = 5)" + }, + { + "bbox": [ + 55, + 298, + 555, + 322 + ], + "type": "text", + "content": ", the representations generated by the model with CAE display enhanced detail, highlighting a more distinct separation between objects and the background." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 335, + 549, + 457 + ], + "blocks": [ + { + "bbox": [ + 65, + 335, + 549, + 457 + ], + "lines": [ + { + "bbox": [ + 65, + 335, + 549, + 457 + ], + "spans": [ + { + "bbox": [ + 65, + 335, + 549, + 457 + ], + "type": "image", + "image_path": "990584779435dc1e70190d7aef758a33c15837f4712c84560f3234a7e0e55f2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "type": "text", + "content": "Figure 7. The representation visualization using arrows for model with and without CAE. The direction and length of each arrow represent the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 512, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 295, + 643 + ], + "type": "text", + "content": "agreement introduced by CAE, we visualize the representations using arrows across various locations and levels. The direction and length of each arrow indicate the gradient direction and magnitude of an embedding at a specific location, computed via the Sobel operator. MNIST is selected for visualization due to its simplicity, minimizing potential interference from complex image backgrounds. As depicted in Fig. 7, the arrows in the model incorporating CAE demonstrate greater consistency, whereas those in the model without CAE appear more disordered and inconsistent." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 656, + 128, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 128, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 128, + 669 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we have proposed a new implementation of GLOM that aligns more closely with its theoretical framework for forming hierarchies of grouped embeddings to rep-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 512, + 555, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 555, + 691 + ], + "type": "text", + "content": "resent nodes in a parse tree. To foster the formation of distinct islands, we have designed a novel contrastive agreement enhancer that encourages agreement between positive embedding pairs while separating negative pairs. To reduce the computational burden of GLOM, we developed a dissimilarity-focused head that preserves more informative content by prioritizing embeddings with lower similarity, which often contain richer information. Both quantitative and qualitative analyses demonstrate that the contrastive agreement enhancer effectively promotes islands of agreement, while the dissimilarity-focused head enhances model efficiency across various datasets. This work significantly improves the practicality of GLOM, and we hope it inspires other researchers to further explore this direction to advance its development." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24358" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Yann Lecun, and Geoffrey Hinton. Deep learning for ai. Communications of the ACM, 64(7):58-65, 2021. Publisher Copyright: © 2021 Owner/Author. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "text", + "content": "[2] Ran Chen, Hao Shen, Zhong-Qiu Zhao, Yi Yang, and Zhao Zhang. Global routing between capsules. Pattern Recognition, 148:110142, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 295, + 201 + ], + "type": "text", + "content": "[3] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 246 + ], + "type": "text", + "content": "[4] Jaewoong Choi, Hyun Seo, Suii Im, and Myungjoo Kang. Attention routing between capsules. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0-0, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[5] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 113-123, 2019. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 334 + ], + "type": "text", + "content": "[6] Laura Culp, Sara Sabour, and Geoffrey E Hinton. Testing glom's ability to infer wholes from ambiguous parts. arXiv preprint arXiv:2211.16564, 2022. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 434 + ], + "type": "text", + "content": "[8] Nicola Garau, Niccolò Bisagno, Zeno Sambugaro, and Nicola Conc. Interpretable part-whole hierarchies and conceptual-semantic relationships in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13689-13698, 2022. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 512 + ], + "type": "text", + "content": "[10] Geoffrey Hinton. How to represent part-whole hierarchies in a neural network. Neural Computation, 35(3):413-452, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 295, + 546 + ], + "type": "text", + "content": "[11] Geoffrey E Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with em routing. In International conference on learning representations, 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 580 + ], + "type": "text", + "content": "[12] Harold Hotelling. Analysis of a complex of statistical variables into principal components. Journal of educational psychology, 24(6):417, 1933. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 602 + ], + "type": "text", + "content": "[13] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 645 + ], + "type": "text", + "content": "[14] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[15] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., pages II-104. IEEE, 2004. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "text", + "content": "[16] Jiachen Li, Ali Hassani, Steven Walton, and Humphrey Shi. Convmlp: Hierarchical convolutional mlp's for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6307-6316, 2023. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 171 + ], + "type": "text", + "content": "[17] Yi Liu, Dingwen Zhang, Qiang Zhang, and Jungong Han. Part-object relational visual saliency. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(7):3688-3704, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 216 + ], + "type": "text", + "content": "[18] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, pages 1-14, 2024. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 258 + ], + "type": "text", + "content": "[19] Vittorio Mazzia, Francesco Salvetti, and Marcello Chi-aberge. Efficient-capsnet: Capsule network with self-attention routing. Scientific reports, 11(1):14634, 2021. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 554, + 303 + ], + "type": "text", + "content": "[20] Rita Pucci, Christian Micheloni, and Niki Martinel. Self-attention agreement among capsules. In Proceedings of the iee//cvf international conference on computer vision, pages 272-280, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 348 + ], + "type": "text", + "content": "[21] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3749-3756, 2020. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "text", + "content": "[22] Leonid I Rudin, Stanley Osher, and Emad Fatemi. Nonlinear total variation based noise removal algorithms. Physica D: nonlinear phenomena, 60(1-4):259-268, 1992. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[23] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. Advances in neural information processing systems, 30, 2017. 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 554, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 447 + ], + "type": "text", + "content": "[24] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 449, + 554, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 513 + ], + "type": "text", + "content": "[25] Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlx architecture for vision. Advances in neural information processing systems, 34:24261–24272, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 514, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 554, + 557 + ], + "type": "text", + "content": "[26] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "type": "text", + "content": "[27] Ru Zeng and Yan Song. A fast routing capsule network with improved dense blocks. IEEE Transactions on Industrial Informatics, 18(7):4383-4392, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 624 + ], + "type": "text", + "content": "[28] Ru Zeng, Yuzhang Qin, and Yan Song. A non-iterative capsule network with interdependent agreement routing. Expert Systems with Applications, 238:122284, 2024. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 625, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 667 + ], + "type": "text", + "content": "[29] Ru Zeng, Yan Song, and Yuzhang Qin. Spatial attention-based capsule networks with guaranteed group equivariance. IEEE Transactions on Automation Science and Engineering, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[30] Ru Zeng, Yan Song, and Yanjiu Zhong. An interpretable unsupervised capsule network via comprehensive contrastive learning and two-stage training. Pattern Recognition, 158: 111059, 2025. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24359" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_content_list.json b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..94903678e44db2231a774460210c7e0d98d6133e --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_content_list.json @@ -0,0 +1,1839 @@ +[ + { + "type": "text", + "text": "AgroBench: Vision-Language Model Benchmark in Agriculture", + "text_level": 1, + "bbox": [ + 176, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Risa Shinoda $^{1,2,4}$ , Nakamasa Inoue $^{3,4}$ , Hirokatsu Kataoka $^{4,5}$ , Masaki Onishi $^{4}$ , Yoshitaka Ushiku $^{6}$", + "bbox": [ + 120, + 179, + 875, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ The University of Osaka $^{2}$ Kyoto University $^{3}$ Institute of Science Tokyo $^{4}$ National Institute of Advanced Industrial Science and Technology (AIST) $^{5}$ Visual Geometry Group, University of Oxford $^{6}$ OMRON SINIC X", + "bbox": [ + 202, + 214, + 794, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Disease Management", + "text_level": 1, + "bbox": [ + 104, + 304, + 269, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5701473a88fe9a3811a1637fcae31cbd6958bedd62476c53adf9de487a8a6479.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 321, + 187, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What should be done to minimize the spread of the disease shown in the image?", + "bbox": [ + 192, + 323, + 364, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Remove the male buds (bell) immediately using a forked stick.", + "bbox": [ + 192, + 356, + 364, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weed Detection", + "text_level": 1, + "bbox": [ + 372, + 305, + 493, + 318 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df59a31a6a22a927741e16ecb06aa63b545f796fee8dbece99e2200770492df4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 321, + 454, + 386 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What species of weed does the highlighted bounding box represent in the image?", + "bbox": [ + 457, + 324, + 629, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: White goosefoot (Chenopodium album)", + "bbox": [ + 459, + 363, + 611, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Disease Identification", + "text_level": 1, + "bbox": [ + 637, + 305, + 797, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/261f7d4baca2a3ae8dc30f2278b9f5708a2b561e849d34ed807a15d5feb3b5e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 321, + 718, + 387 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What disease is affecting the plant?", + "bbox": [ + 723, + 329, + 895, + 351 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Bacterial canker.", + "bbox": [ + 723, + 369, + 854, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crop Management", + "text_level": 1, + "bbox": [ + 107, + 390, + 246, + 404 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c72cc416234804a15201207f62ef335610b36cdbe555bb2ac6be7f9999a9a69f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 104, + 406, + 187, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: Is the crop shown in the image ready for harvest based on its current condition?", + "bbox": [ + 192, + 409, + 364, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Yes, the leaves and stems have turned yellow, and the seed pods are brown to black.", + "bbox": [ + 192, + 441, + 364, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AGROBENCH", + "text_level": 1, + "bbox": [ + 375, + 412, + 625, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Agronomist Al Benchmark", + "bbox": [ + 401, + 443, + 576, + 455 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pest Identification", + "text_level": 1, + "bbox": [ + 638, + 390, + 772, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32e5cc5c8fcb0136d5eb28ece6926ea4c6e65e4a023f9cfa8b4fb63050430701.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 406, + 718, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What pest is affecting the plant?", + "bbox": [ + 723, + 417, + 895, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Squash bug.", + "bbox": [ + 723, + 454, + 831, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional Methods", + "text_level": 1, + "bbox": [ + 107, + 474, + 254, + 488 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d6aed769e25151ee011dcd88d0d57dfe8be179abeb1afb06b8e009d85285d07.jpg", + "image_caption": [ + "Figure 1. We present AgroBench (Agronomist AI Benchmark) designed to comprehensively evaluate 682 disease categories across 203 agricultural crop types for 7 vision-language question-answer tasks. In the era of larger-scale vision-language models (VLMs), our AgroBench is obviously non-trivial in terms of many more crop and disease categories with all expert annotations for establishing QA benchmarks in the agricultural domain." + ], + "image_footnote": [], + "bbox": [ + 104, + 491, + 187, + 555 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What is one of the main benefits the process shown in the image?", + "bbox": [ + 192, + 494, + 364, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: Weed control and insect pest reduction.", + "bbox": [ + 192, + 532, + 364, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Machine Usage", + "text_level": 1, + "bbox": [ + 372, + 474, + 488, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c540aff0fbaec8558549d07997ba59c11bbca1881dfe4d4199bcf1c0eb4e12a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 494, + 452, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Question: What is the main purpose of using the machine shown in the image?", + "bbox": [ + 457, + 493, + 629, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Answer: To control weeds in the field without relying on chemical herbicides.", + "bbox": [ + 457, + 526, + 630, + 555 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76ed41d7572652f9f13fd383e21a08c29587282abcb0aff33e1fda6e1640a6c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 477, + 669, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Crop categories: 203", + "text_level": 1, + "bbox": [ + 684, + 479, + 838, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d1008e2467a3488f108c5c190325a4d5658c7e1519619c3eeab25ce486430af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 503, + 671, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Disears categories: 684", + "text_level": 1, + "bbox": [ + 684, + 508, + 864, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fea14e2638ce0c66458baa1b6610fb1e1b908b1b496f258f298e950b1bfc08f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 531, + 671, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Expert annotation", + "text_level": 1, + "bbox": [ + 683, + 537, + 820, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 642, + 326, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Precise automated understanding of agricultural tasks such as disease identification is essential for sustainable crop production. Recent advances in vision-language models (VLMs) are expected to further expand the range of agricultural tasks by facilitating human-model interaction through easy, text-based communication. Here, we introduce AgroBench (Agronomist AI Benchmark), a benchmark for evaluating VLM models across seven agricultural topics, covering key areas in agricultural engineering and relevant to real-world farming. Unlike recent agricultural VLM benchmarks, AgroBench is annotated by expert agronomists. Our AgroBench covers a state-of-the-art range of categories, including 203 crop categories and 682 disease categories, to thoroughly evaluate VLM capabilities. In our evaluation on AgroBench, we reveal that", + "bbox": [ + 89, + 675, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VLMs have room for improvement in fine-grained identification tasks. Notably, in weed identification, most opensource VLMs perform close to random. With our wide range of topics and expert-annotated categories, we analyze the types of errors made by VLMs and suggest potential pathways for future VLM development. Our dataset and code are available at https://dahlian00.github.io/AgroBenchPage/.", + "bbox": [ + 511, + 643, + 906, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 643, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Agriculture is a fundamental process for humans to produce crops to live and stay healthy. With the development of computer vision technology, effective and automated management of external crop factors such as diseases and pests has been explored, contributing to stable crop", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7634", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/8cdc7ad134a16e2cb02e1cabaaff0a62a8503b6679393ae6a1855d8a111ee587.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetAnnot.CropWeedDiseasePestImagesQA pairsMain Purpose
Agri-LLaVA [48]GPT-429-109112391k391k (Synthetic)Training
AgroInstruct [4]GPT-41744741270k70k (Synthetic)Training
CDDM [20]GPT-415-60-137k1M (Synthetic)Training
AgroBench (Ours)Expert2031086821343,7454,342 (Expert)Evaluation
", + "bbox": [ + 153, + 88, + 843, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of agricultural vision datasets. Expert refers to 'human expert' in this context. AgroBench provides a comprehensive evaluation framework for assessing VLMs, featuring multiple tasks and a wide range of categories.", + "bbox": [ + 89, + 196, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "production. This includes detecting and classifying undesirable conditions like diseases [1, 10, 13, 34, 38, 41, 47] and pests [3, 49, 53], as well as general plant management tasks such as crop classification [52, 54] and recognition of crop maturity [22, 25, 37, 39], and structure understanding [21, 43].", + "bbox": [ + 88, + 238, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To maintain stable crop production, it is important to recognize a wide range of crop conditions, including undesirable situations such as diseases and pests, and to know how to respond appropriately. A single visual model that can handle various conditions, not only disease and pest detection but also treatment and general crop management, would be highly beneficial. However, most existing approaches use task-specific models. These models usually require large amounts of training images and manual annotations for each task. As a result, farmers often need to use several different models depending on the situation. This increases complexity and makes the overall system less accessible for practical use in agriculture.", + "bbox": [ + 91, + 333, + 482, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For general purpose visual tasks, vision-language models (VLMs) [17, 19, 32, 33, 45, 46, 50] have become widespread recently because they can understand task definitions provided by natural language prompts, covering a wide range of applications without the need for task-specific model training. The recognition ability of VLMs is closely connected to image recognition itself and linked to an object's words, and supports open-vocabulary recognition through web-scale training. Here, zero-shot recognition and few-shot adaptation have also been realized by VLMs with language representations. This ability opens up a wide range of applications; therefore, we believe this can be applied to agricultural scenarios as well. It is worth noting that VLMs offer an easy-to-use interface for the general public, especially the question-answer (QA) and conversation modes. To investigate the range of tasks that VLMs obtained through large-scale vision-language training can cover, recent studies have introduced various benchmark datasets to fully evaluate the VLMs, including tasks such as understanding diagrams and charts [26, 28], and question answering requiring specialized knowledge [7, 55, 56].", + "bbox": [ + 91, + 534, + 482, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, VLM research remains underexplored in agriculture due to a lack of benchmark datasets that include diverse tasks and categories in agriculture. Several pioneer-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing works [20, 48] have adapted open-source VLMs such as LLaVA [18] to the agricultural domain by fine-tuning them on synthetic datasets generated from closed-source VLMs such as GPT-4o. This is recognized as an effective approach because black-box VLMs often possess a certain amount of agricultural knowledge obtained from the Internet. While this helps in generating responses that are generally acceptable to experts, there is almost no way to verify whether the answers are truly correct. Moreover, in the limited evaluation of categories, we are insufficient for fully assessing VLM knowledge; whether they can answer a wide range of types, such as diseases, pests, and weeds. These limitations motivated us to develop a benchmark dataset in order to evaluate the VLM's broad knowledge in the agricultural domain and its applicability as a practical application.", + "bbox": [ + 511, + 238, + 903, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we introduce AgroBench (Agronomist AI Benchmark), a comprehensive benchmark dataset for VLM for the agricultural domain, covering a state-of-the-art range of categories for agriculture-focused benchmark datasets for VLM; 682 disease, 134 pest, 108 weed, and 203 crop categories (Figure 1). We cover not only identification tasks but also crop production and disease management knowledge. Moreover, we include other important topics related to crop management with 98 machine categories and 77 traditional management methods to investigate more about VLM's ability. We carefully selected benchmark tasks from key agricultural engineering research areas, and also the tasks that address challenges faced by farmers in real agricultural scenarios. We employ a multiple-choice format, and all questions are annotated by human agronomist experts, which overcomes the limitations of previous synthetically created datasets. We carefully selected images from Creative Commons-licensed and publicly available datasets, including real farm conditions. Also, under our manual annotation process, unclear images were removed. As shown in Table 1, AgroBench contains data with the most diverse disease and crop variation categories and many more crop/weed annotations.", + "bbox": [ + 511, + 468, + 903, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows;", + "bbox": [ + 531, + 820, + 790, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We developed AgroBench, a benchmark dataset for VLM, to assess their broad agricultural knowledge of VLMs and evaluate their applicability in practical applications.", + "bbox": [ + 511, + 839, + 906, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7635", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In our AgroBench evaluation, VLMs tend to achieve high performance in disease and crop management tasks; however, there is still room for improvement in weed and disease identification.", + "- Our AgroBench enables error analysis by providing broader category annotations, highlighting future directions for model training focus areas." + ], + "bbox": [ + 89, + 90, + 482, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 212, + 230, + 228 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Computer Vision for Agriculture", + "text_level": 1, + "bbox": [ + 89, + 238, + 379, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the rise of computer vision techniques driven by advancements in deep learning, a wide range of agricultural tasks have been explored over the past decade. In particular, dataset construction and benchmarking play an important role in developing models and practical applications in agriculture, while also paving the way for forward-looking advancements. Disease identification is a key research focus across various crops, such as rice [1, 34], tomato [10], cacao [13], and sugarcane [47]. Multicrop datasets have also been created [29, 41]. PlantDoc [41] covers 13 species and 17 classes focused on leaf-based diseases. Plant Village [29] offers 39 classes containing both diseased and healthy leaf categories. In addition to crop disease identification, pest identification [49, 53], weed identification [12, 31, 42] have also been studied.", + "bbox": [ + 89, + 261, + 482, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While most of the agricultural datasets primarily focus on visual data, few computer vision studies in agriculture investigate the potential of multi-modal approaches. The PlantWild dataset [52] includes 56 plant-disease class pairs, which are collected through image search engines. They implement a CLIP-based model and show the possibility of training with combined text and image data. In an instruction-based format, the CDDM [20] has created 16 categories of crops and 60 categories of crop diseases dataset, which generates instructional data using GPT. While these synthetically created datasets explore the agriculture multi-modal models, there remains a lack of datasets for comprehensive multi-modal model evaluation, validated by human experts and covering a wide range of tasks and categories.", + "bbox": [ + 89, + 488, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision-Language Models", + "text_level": 1, + "bbox": [ + 89, + 726, + 316, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Models. Visual models in the computer vision field have been accelerated by language modality with the data resource of sentence-level inputs and web-scale texts. Specifically, CLIP [35] has made a significant contribution in this context, by text and image feature alignment through contrastive learning. CLIP played a key role in introducing more sophisticated visual representation into language explanations e.g., Flamingo [2], BLIP [14, 15], Qwen [5], PaLI [8], LLaVA [17, 18], CogVLM [50], and Emu [44, 45, 51]. Closed-source VLMs achieve state-of-the-art per", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "formance, such as GPT-4o [32] and Gemini Pro [46], which are said to acquire human-level knowledge across diverse fields on the Internet.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Benchmarks.** Along this line, several recent studies introduced benchmark datasets. Especially, the representative examples in terms of universal knowledge benchmarking include MMMU [55], MMMU-Pro [56] and MMStar [6] for multi-modal understanding. These benchmarks contain highly diverse domains (e.g., natural, graph, illustration, medical images) and academic fields (e.g., science, engineering, art, and medical fields) under the tasks of vision language such as question-answering and reasoning. The series of MMMUs have been verified with foundation models like GPT-4V, but they cannot answer the questions perfectly. More specific domain datasets have also been proposed, such as those for Medicine [27, 36], Chart [26, 28, 40], and Video understandings [9, 16, 30], contributing to the accurate evaluation of VLMs and guiding future research directions.", + "bbox": [ + 511, + 136, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. AgroBench", + "text_level": 1, + "bbox": [ + 513, + 391, + 633, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces AgroBench, the first comprehensive benchmark dataset to evaluate VLM models from the perspective of agricultural vision tasks. The benchmark consists of seven tasks covering a wide range of tasks selected from key agricultural engineering research areas, as well as tasks that address real-world challenges faced by farmers in real agriculture scenarios. AgroBench includes 682 disease categories, 134 pest categories, 203 crop categories, and 108 weed categories, representing the largest number of categories in each area to date, to the best of our knowledge.", + "bbox": [ + 511, + 416, + 905, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Benchmark Tasks in Agricultural Scenes", + "text_level": 1, + "bbox": [ + 511, + 575, + 862, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The seven benchmark tasks encompass key research areas in agricultural engineering as well as real-world challenges faced by farmers. To facilitate VLM evaluation, we also provide prompts to address each task in a question-answer format. The details of each task are described below.", + "bbox": [ + 511, + 599, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1) Disease Identification (DID). The DID task aims to accurately diagnose and classify crop diseases. This is a key task in agriculture to protect crop health and maximize yields. Our benchmark provides 1,502 QA pairs that cover 370 disease categories, 160 crop categories and 682 crop-disease combinations. To thoroughly evaluate the capabilities of VLMs, we include four misleading disease labels for each image, featuring diseases with similar symptoms or common diseases affecting the target plant. VLMs are required to diagnose the disease based on the image, considering both symptoms and the crop species. Figure 2a illustrates example images.", + "bbox": [ + 511, + 674, + 906, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) Pest Identification (PID). The PID task aims to identify pests to prevent infestations that can severely impact crop health. Accurate identification reduces economic losses and", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7636", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3d4a2bf5549330284c94ef35053d17fd9ed5e6a77d73d9be524f7b58ef53b0c9.jpg", + "image_caption": [ + "(a) Disease identification (DID)" + ], + "image_footnote": [], + "bbox": [ + 96, + 102, + 898, + 424 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de6f8bc38f4fd53f589354903603a699929451185ee9edf53ae4a78cf7854319.jpg", + "image_caption": [ + "Figure 2. Examples of labeled images for DID, PID, and WID tasks. Our dataset includes 682 crop-disease pairs, 134 pest categories, and 108 weed categories. We prioritized collecting images from real farm settings.", + "Figure 3. Seven benchmark tasks in AgroBench. AgroBench includes multiple topics with a diverse range of categories. The total accuracy is calculated by the average of each task to mitigate the difference in QAs." + ], + "image_footnote": [], + "bbox": [ + 140, + 479, + 434, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "minimizes harm to the environment. Our benchmark provides 544 labeled images that cover 134 pest categories, including insects, mites, and other organisms harmful to plants. For categories where we obtain multiple images, we select multiple insect growth stages whenever possible. To fully evaluate the VLMs, we assign alternative choices that closely resemble or are commonly associated with the target", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "crops. Figure 2b shows examples of labeled pest images.", + "bbox": [ + 511, + 478, + 888, + 493 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3) Weed Identification (WID). The WID task aims to identify weed species. Our benchmark includes 609 images of weeds with ground-truth bounding boxes, covering 108 weed species commonly found in farm fields. We assign bounding boxes because multiple weed types often grow closely together, and we want to clarify which one is the target. Specifically, VLMs are required to identify the weed species within the provided bounding box on the image. Figure 2c shows examples of labeled weed images.", + "4) Crop Management (CMN). Crop management focuses on optimizing farming practices to facilitate crop growth. This involves making decisions on irrigation, fertilization, planting times, and other cultivation practices. Our benchmark provides 411 question-answer pairs for this task. VLMs are required to analyze images of crops and recommend appropriate management strategies by considering factors such as crop health, growth stage, and environmental conditions visible in the images given five answer candidates. In Figure 4a and b, we show two examples of the harvest timing for white and green asparagus, respectively. Our dataset includes complex questions that take into account the differences in their harvest timing.", + "5) Disease Management (DMN). Disease management aims to control and reduce diseases in crops. This involves informed decisions on interventions such as applying pesticides, adopting resistant crop varieties, or modifying cul" + ], + "bbox": [ + 511, + 497, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7637", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4f30009ad05735b9fef67432f58689b7d107f17aac1a3937140a67ac61fae5b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 189, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/076cf1960b742c694f51e06de401e57dcecbd814ee68a1fd2330c0958c7bff6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 90, + 500, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/22f50d4ecc4936e5fbcc49448c773614189727479f39b94e29ed005037bc6335.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 90, + 596, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f25c0cf5da35a4823f8139524b3df03fcc6b1bf6ee618a981dbc71c8454bf43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 90, + 895, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d00e18010cf57c343a3fd5b562a646a52fa3cad0e097a6a3c842e715d6267af.jpg", + "image_caption": [ + "Figure 4. Examples of QA pairs for CMN and DMN tasks. (a) and (b) Crop management QA types for white asparagus and asparagus, respectively. Their difference in the harvest timing affects the answer's difference correctly. (c) and (d) Disease management QA types for the alfalfa bacterial leaf spot with the initial and severe symptoms, respectively. Based on the severity of the symptoms, the annotator changes the answer." + ], + "image_footnote": [], + "bbox": [ + 96, + 167, + 189, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/881a3aff989b5b9ba5381602e7781cc9ab3ba2987b72af4d439156ec73c795cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 167, + 486, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/641deafee5d7f9a560c0bb7676fa3590c8097bd4dbcdd0f33673b761e9db8c78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 167, + 594, + 239 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a729418308079261a220c19eb3ee2ae5a75dd4629d0c7b3e4df566fe914fc51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 167, + 895, + 239 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "tivation practices. Our benchmark provides 569 question-answer pairs covering 141 crop-disease combination categories. Since many diseases share the same management strategies regardless of the crop, we carefully select a diverse range of disease types and management strategies. VLMs are required not only to identify the disease but also to recommend appropriate management strategies based on images of affected crops. In Figure 4c and d show example QAs from the Disease Management tasks. Both images depict Bacterial leaf spot on alfalfa, with c showing the initial stage of the disease and d showing the severe stage. Based on the severity of the disease, we provide different answer options and set different correct answers. The input prompt consists of a question and five answer candidates.", + "6) Machine Usage QA (MQA). Machine usage QA addresses the correct use and choice of agricultural machinery depending on the task and farming conditions. Selecting the appropriate machinery is essential for efficient farming. Our benchmark provides 303 question-answer pairs covering 98 machine categories. Given that many crops share the same machinery (e.g., soil preparation and irrigation), the coverage of this category is comprehensive. VLMs are required to answer questions about machinery operation or select the appropriate machine for a given scenario based on images of machinery or farming conditions.", + "7) Traditional Management (TM). Traditional management methods involve natural and sustainable approaches to farming, such as the use of organic fertilizers, terrace farming, and agroforestry. Recent computer vision studies have not focused on these traditional practices, although many are still used by certain local farmers. Our benchmark includes 404 question-answer pairs, including 77 traditional management practices. VLMs are required to identify the management method or explain its effectiveness, given five answer choices." + ], + "bbox": [ + 88, + 318, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Dataset Construction", + "text_level": 1, + "bbox": [ + 511, + 316, + 714, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Selection. To establish a high-quality benchmark covering a wide range of crops, diseases, and pest categories we initially curated around 50,000 agricultural images from websites supervised by plant pathologists, either where redistribution was permitted or where we obtained redistribution permission. We selected images in real farm settings as much as possible, as shown in Figure 2, to evaluate real-world scenarios. When licensed images for a target category were limited, we used laboratory setting images. For curation, we obtained images along with their corresponding labels. The annotator is one of the authors, who holds a Ph.D. in Agriculture. In the human evaluation conducted during the experiment, other individuals with a Ph.D. or M.S. degree in Agriculture reviewed questions to assess the quality of the QA pairs. The annotator selected images that clearly represented target labels and removed those images with irrelevant labels. For the Machine usage QA and Traditional management, annotators manually created categories based on textbooks and websites, and searched for corresponding images with redistribution licenses. For Weed identification, we use existing dataset [11, 23, 24, 42]. For this Weed identification category, we will provide a simple code to download the data and crop images and assign bounding boxes for Weed Identification, allowing users to work with the existing dataset without redistributing the images. Through these image selection processes, the selected images are high-quality 4,218 representative images. The detailed distribution of tasks and categories for each task is summarized in Figure 3.", + "bbox": [ + 511, + 340, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QA Annotation. For the DMN, CMN, MQA, and TM tasks, all question-answer pairs were created manually, independent of any LLMs or VLMs. Annotators used GPT only for sentence rephrasing, but they were prohibited from using any knowledge from GPT for QA creation. Annotators were allowed to refer to textbooks, academic journals, and other authoritative sources in the field to ensure the accuracy and depth of the dataset. This took around 150 man", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7638", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0c2385a36ec4186adeb5773a47c99b07af88d4cce71892a0f4c830d3e1cfe5ba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDIDDMNPIDWIDCMNMQATMOverall (all)Overall (subset)
Random Choice21.7715.6420.4017.9016.0622.1119.3119.0319.11
Human25.0022.5045.0020.0036.2557.5051.25-36.79
Closed-Source Vision Language Models (VLMs)
GPT-4o mini [33]53.6080.6760.0435.1464.2370.9669.8062.0669.65
GPT-4o [32]64.1887.3577.7644.1775.4382.8482.4373.4579.26
Gemini1.5-Flash [46]55.0679.9670.0450.9064.7278.2273.2767.4568.82
Gemini1.5-Pro [46]62.9281.5574.4555.1771.0582.8477.7272.2469.74
Open-Source Vision Language Models (VLMs)
EMU2Chat [44]42.0148.3343.7523.8140.3937.6247.7740.5333.84
LLaVA-Next-8B [19]45.4772.5843.0130.0554.2656.1157.4651.2857.84
LLaVA-Next-72B [19]54.9580.0049.8126.9866.9266.1170.3859.3164.36
QwenVLM-7B [5]51.2680.4963.9733.1766.4276.2477.4864.1566.41
QwenVLM-72B [5]57.9987.8773.3534.4875.9180.8684.1670.6672.45
CogVLM-19B [50]29.1653.7852.3925.4554.0171.6266.0950.3644.27
LLaVa-7B [18]36.0262.7438.7924.7953.7746.5355.2045.4146.14
LLaVA-13B [18]40.2168.8944.4924.7959.3754.1358.4250.0455.31
", + "bbox": [ + 143, + 88, + 854, + 395 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f7cc85175df4795f87fff46bfb919ea2fa21b7c50ac3d5fd0ba274fce07911ad.jpg", + "table_caption": [ + "Table 2. Results for the seven benchmark tasks with images. We provide results for Random Choice, Human Validation, four closed-source VLMs, and open-source VLMs. Human validation was conducted by 28 people on a subset of 80 samples per task as a reference." + ], + "table_footnote": [], + "table_body": "
DIDDMNPIDWIDCMNMQATMOverall
GPT-4o [32]1.9372.5818.751.0040.3925.0848.2729.71
LLaVA-Next-8B [19]26.1070.3021.8819.7053.7730.3640.3537.49
", + "bbox": [ + 184, + 449, + 815, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Results for the seven benchmark tasks with text only. We additionally evaluate the model without image inputs, and the overall performance is close to random.", + "bbox": [ + 89, + 518, + 906, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hours. We carefully annotate various types of questions. Examples of QAs that require expert knowledge are shown in Figure 4. All questions were carefully created so that image reference is necessary for answering.", + "bbox": [ + 89, + 555, + 482, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Statistics. Following the dataset annotations and selections, our AgroBench comprises seven tasks with 4,342 QA pairs, as shown in Table 1. All tasks comprise a wide range of categories for detailed evaluation. Please refer to the supplementary materials for more details and examples.", + "bbox": [ + 89, + 614, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 718, + 223, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 742, + 297, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Models. We use four closed-source models: GPT-4o [32], GPT-4o mini [33], Gemini1.5-Pro [46], and Gemini1.5-Flash [46]. GPT-4o mini and Gemini1.5-Flash are down-scale versions of GPT-4o and Gemini1.5-Pro, respectively. We use eight open-source models: EMU2Chat [44], LLaVA-Next-8B [19], LLaVA-Next-72B [19], QwenVLM-7B [5], QwenVLM-72B [5], CogVLM-19B [50], LLaVa-7B [18], and LLaVa-13B [18]. For the details of these models, please refer to the supple", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mentary material.", + "bbox": [ + 511, + 556, + 633, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human Results. We also present results from human participants for reference. We surveyed 28 students, each holding at least a bachelor's degree in agriculture, and asked them to answer 20 questions each. This created a test subset of 280 questions, with each question answered by two participants, resulting in a total of 560 responses. We averaged the results per task and reported the accuracy. Each participant was permitted to use a book or translator to look up word meanings but was prohibited from using the internet for searches. If participants were unsure of the answer, they were asked to provide the response they believed to be most accurate.", + "bbox": [ + 511, + 575, + 906, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. Importantly, our dataset evaluation is conducted per task, and overall scores are averaged based on the number of tasks, not the number of QAs. This prevents categories with a large number of evaluations from becoming dominant. We adopt an exact matching approach for our five-option questions. If the model's response matches the option's letter or the answer sentence, we consider it correct. If the model's answer does not match any option, including cases where there is no answer or mul", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7639", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fab5047c0c1479fdb57a3923aafe09d26fc86801887fd7a570cc75d4c1e0152.jpg", + "image_caption": [ + "Figure 5. Results of seven benchmark tasks with Chain of Thought (CoT). Baseline indicates results without CoT. In the one-shot, two-shot, and three-shot settings, we provide one, two, and three CoT examples per task, respectively, to guide the model." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 903, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1d88734ca867c9fda3e9719d9476adcba2160f86d85b76f5a980c6b9e4fdd7b.jpg", + "image_caption": [ + "Figure 6. Error analysis on seven benchmark tasks with GPT-4o. We extract a maximum of 15 errors per task from the zero-shot CoT result. We manually analyze how they conclude the incorrect answer." + ], + "image_footnote": [], + "bbox": [ + 96, + 292, + 898, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tiple answers, we consider it incorrect.", + "bbox": [ + 89, + 531, + 346, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 559, + 230, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we discuss the main results of our AgroBench evaluation performance. We evaluate our AgroBench using eight open-source VLMs and four closed-source VLMs with APIs.", + "bbox": [ + 89, + 580, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Challenges of the AgroBench. We show the main results in Table 2. The most difficult task is Weed Identification (WID), on which most open-source VLMs perform at around the random score. The highest WID accuracy, $55.17\\%$ , was achieved by Gemini 1.5-Pro. This suggests that the knowledge about weeds is not as fully trained as that of crops. All models' Disease Identification (DID) results are lower than their Disease Management (DMN) results. This means that VLM models can gain contextual information, but there is still room for perceptual improvement.", + "bbox": [ + 89, + 643, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Comparison. Overall, closed-source VLMs achieve better results than open-source VLMs and achieve higher performance than humans. GPT-4o model achieved the highest score in overall performance. Among the open-source models, QwenVLM-72B achieves the best result in overall accuracy, which is comparable or even superior to open-source VLMs in some tasks. QwenVLM", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "72B achieves satisfactory results on both identification and question-answering benchmarks on AgroBench.", + "bbox": [ + 511, + 531, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 511, + 574, + 624, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Context Reliance. We further evaluate the results on AgroBench using text-only input to determine whether visual information is necessary to answer the questions. Table 3 presents the experimental results across seven benchmark tasks on AgroBench. With text-only input, performance drops across all models, confirming the reliance on visual information. However, for both models, the DMN, CMN, and TM tasks maintain significantly higher accuracy than random selection. Although we ensured that our questions do not include disease names or appearance-related traits, models tend to infer answers based on estimation. This suggests that many disease types share common management strategies, such as avoiding humidity, preferring cooler temperatures, or pruning infected parts, allowing models to predict the most likely option. A similar pattern is observed in the CMN and TM tasks, where models can make educated guesses based on contextual cues. (See the supplementary material for further detailed examples.)", + "bbox": [ + 511, + 597, + 906, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Chain of Thought. We evaluate the effectiveness of Chain of Thought (CoT) reasoning. In the one-shot setting, we", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7640", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) Lack of Knowledge", + "text_level": 1, + "bbox": [ + 102, + 95, + 263, + 109 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: What disease is affecting the plant?", + "bbox": [ + 104, + 112, + 315, + 125 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0e92e8e14c4de01ed1ee34a40c0381608f2be26e92872a727e1460dc74d5e56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 126, + 254, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-4o: ..., **Sclerotinia blight**: Characterized by white, cottony mycelium and black sclerotia on stems. The image does not show these symptoms. (not enough knowledge) ..., the plant is likely affected by **D. Cylindrocladium black rot**: (This is more common on root)", + "bbox": [ + 258, + 125, + 470, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A: Sclerotinia blight", + "bbox": [ + 104, + 220, + 215, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Perceptual Error", + "text_level": 1, + "bbox": [ + 102, + 239, + 248, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Q: What activity is being performed using the tractor-pulled machine shown in the image?", + "bbox": [ + 104, + 255, + 426, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32076a514abab6329414e314b00c8e580b53bda312b3345459d77b27f5496820.jpg", + "image_caption": [ + "Figure 7. Error examples of GPT-4o. Examples illustrate the two main error types: Lack of Knowledge and Perceptual Error." + ], + "image_footnote": [], + "bbox": [ + 102, + 279, + 264, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GPT-40:The machine seems to be designed to lift plants from the soil, ..., and there are plants being lifted from the ground (There are no harvested ctops)....the activity being performed is: D. Harvesting potato tubers from the soil using a tractor-pulled harvesting machine.", + "bbox": [ + 267, + 277, + 470, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A: Cutting and chopping the stems and leaves using a tractor-pulled leaf chopper.", + "bbox": [ + 106, + 373, + 398, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provide a single CoT reasoning example along with an image and the corresponding prompt for each task to guide the model in answering the question. We select 100 QAs per task for evaluation and use GPT-4 as the VLM. Figure 5 presents the accuracy of CoT reasoning with the baseline results. While CoT achieves slightly higher accuracy, the improvement is not significant compared to the baseline. However, CoT demonstrates effectiveness in certain tasks, such as PID, WID, CMN, and TM. For instance, in PID, the VLM performs step-by-step and careful reasoning (see the supplementary material for more detailed examples). In WID, the most challenging task, CoT provides useful examples that may help the model make predictions. Overall, CoT contributes to accuracy, but we observe performance saturation in the three-shot setting. For further detailed examples, please refer to the supplementary material.", + "bbox": [ + 89, + 441, + 483, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Error Analysis", + "text_level": 1, + "bbox": [ + 89, + 695, + 241, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we analyze the mistake types that occur depending on the agricultural tasks. We bring out up to 15 failure examples per task from the zero-shot CoT results and manually analyze how they reached the incorrect answers.", + "bbox": [ + 89, + 718, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lack of Knowledge (51.92%). This includes cases where VLM can't accurately describe the appearance or relevant knowledge of a choice (e.g., VLM fails to describe disease symptoms or insect characteristics) or lacks context (e.g., VLM doesn't know how to treat diseased crops or manage crops for high yield). Figure 7a shows an example of a Lack of Knowledge case. When the VLM analyzes the correct answer option for Sclerotinia blight, it fails to describe the", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "symptoms of discoloration and wilting. Additionally, its incorrect answer choice is more commonly associated with the root rather than the stem. These errors are based on a Lack of Knowledge, suggesting VLMs need more detailed categories and domain-specific training.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Perceptual Error (32.69%). This indicates that VLM can't pay attention or recognize the answer-related part in the image (e.g., can't recognize the green insect on the leaf), and VLM misunderstands the image, leading to incorrect answers. Figure 7b shows an example of the model hallucinating and misunderstanding the situation. First, the VLM incorrectly identifies the machine as one used for lifting plants from the soil. Then, it describes the scene as if there are harvested crops, even though no harvested crops are present. These errors can be mitigated by enhancing the VLM's perception abilities for domain-specific classes. Additionally, improving general perception capabilities, including reducing hallucinations, can further contribute to VLM performance.", + "bbox": [ + 511, + 167, + 906, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Reasoning Error (7.6%). Reasoning error involves the VLM can describe the options correctly, but can't compare them step by step and conclude the wrong answer. This error is relatively low compared to the existing work [55] since AgroBench requires more specific knowledge and doesn't include reasoning relying on problems (e.g., math).", + "bbox": [ + 511, + 381, + 905, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other Errors (7.79%). For the other errors, we observe Shortcut Error (The VLM can pick up the two candidate options correctly but conclude the answer without comparing the candidates), Double Answer Error (Concluding two answers are correct), Interpretation Misunderstanding (VLM misleading the question and conclude wrong answer), and Reject to Answer (VLM conclude there is no answer).", + "bbox": [ + 511, + 474, + 905, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we develop AgroBench, a comprehensive benchmark dataset for VLMs in the agricultural domain, covering a state-of-the-art range of categories. Our dataset comprises seven benchmark tasks encompassing key research areas in agricultural engineering as well as real-world challenges faced by farmers. AgroBench contributes to agricultural VLM research by addressing the lack of datasets for comprehensive multi-modal model evaluation, validated by human experts.", + "bbox": [ + 511, + 626, + 905, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our evaluation, VLMs exhibit strengths across different tasks. However, in several tasks such as weed identification and disease identification, all models show room for improvement. Our error analysis reveals that most failures are due to a lack of knowledge (51.92%), suggesting that VLMs require more specialized agricultural knowledge. Our dataset will facilitate agricultural VLM research, enabling broad category and task evaluation to support sustainable, automated agriculture.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7641", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 261, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the AIST KAKUSEI Project (FY2024) and JST FOREST Grant Number JPMJFR206F. We would like to thank the Agricultural Administration Division, Department of Agriculture, Hokkaido Government, for providing some of the images. We are also grateful to Daniel Steininger for contributing images from the CropAndWeed dataset to our dataset. We used ABCI 3.0 provided by AIST and AIST Solutions.", + "bbox": [ + 89, + 114, + 483, + 227 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 253, + 187, + 271 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Petchiammal A, Briskline Kiruba S, Murugan D, and Pandarasamy Arjunan. Paddy doctor: A visual image dataset for automated paddy disease classification and benchmarking. IEEE Dataport, 2022. 2, 3", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3", + "[3] Abderraouf Amrani, Ferdous Sohel, Dean Diepeveen, David Murray, and Michael G.K. Jones. Deep learning-based detection of aphid colonies on plants from a reconstructed brassica image dataset. Computers and Electronics in Agriculture, 205:107587, 2023. 2", + "[4] Muhammad Awais, Ali Husain Salem Abdulla Alharthi, Amandeep Kumar, Hisham Cholakkal, and Rao Muhammad Anwer. Agropt: Efficient agricultural vision-language model with expert tuning. arXiv preprint arXiv:2410.08405, 2024. 2", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3, 6", + "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3", + "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao," + ], + "bbox": [ + 99, + 280, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2", + "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Alexander Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. 2023. 3", + "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees G. M. Snoek, and Yuki M. Asano. Tvbench: Redesigning video-language evaluation. 2024. 3", + "[10] Mamta Gehlot, Rakesh Saxena, and Geeta Gandhi. \"tomato-village\": a dataset for end-to-end tomato disease detection in a real-world environment. Multimedia Systems, 29:1-24, 2023. 2, 3", + "[11] Benedikt Geisler. Perrenial plants detection, 2021. 5", + "[12] Sebastian Haug and Jörn Ostermann. A crop/weed field image dataset for the evaluation of computer vision based precision agriculture tasks. In Proc. European Conference on Computer Vision (ECCV), pages 105-116. Springer, 2014. 3", + "[13] Atuhurra Jesse, N'guessan Yves-Roland Douha, and Pabitra Lenka. Image classification for cssvd detection in cacao plants. arXiv preprint 2405.04535, 2024. 2, 3", + "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742. PMLR, 2023. 3", + "[16] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark. In CVPR, 2024. 3", + "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 3", + "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 6", + "[19] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.2,6", + "[20] Xiang Liu, Zhaoxiang Liu, Huan Hu, Zezhou Chen, Kohou Wang, Kai Wang, and Shiguo Lian. A multimodal benchmark dataset and model for crop disease diagnosis. In Proc. European Conference on Computer Vision (ECCV), 2024. 2, 3", + "[21] Xinpeng Liu, Kanyu Xu, Risa Shinoda, Hiroaki Santo, and Fumio Okura. Masks-to-skeleton: Multi-view mask-based" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7642", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tree skeleton extraction with 3d gaussian splatting. Sensors, 25(14), 2025. 2", + "[22] Shenglian Lu, Wenkang Chen, Xin Zhang, and Manoj Karkee. Canopy-attention-yolov4-based immature/mature apple fruit detection on dense-foliage tree architectures for early crop load estimation. Computers and Electronics in Agriculture, 193:106696, 2022. 2", + "[23] Yuzhen Lu. Cottonweeddet3. kaggle, 2022. 5", + "[24] Simon Leminen Madsen, Solvejg Kopp Mathiassen, Mads Dyrmann, Morten Stigaard Laursen, Laura-Carlota Paz, and Rasmus Nyholm Jorgensen. Open Plant Phenotype Database of Common Weeds in Denmark, 2020. 5", + "[25] Abdul Khalique Maitlo, Abdul Aziz, Hassnian Raza, and Neelam Abbas. A novel dataset of guava fruit for grading and classification. Data in Brief, 49:109462, 2023. 2", + "[26] Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, Dublin, Ireland, 2022. Association for Computational Linguistics. 2, 3", + "[27] João Matos, Shan Chen, Siena Placino, Yingya Li, Juan Carlos Climent Pardo, Daphna Idan, Takeshi Tohyama, David Restrepo, Luis F. Nakayama, Jose M. M. Pascual-Leone, Guergana Savova, Hugo Aerts, Leo A. Celi, A. Ian Wong, Danielle S. Bitterman, and Jack Gallifant. Worldmedqa-v: a multilingual, multimodal medical examination dataset for multimodal language models evaluation, 2024. 3", + "[28] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1516-1525, 2020. 2, 3", + "[29] Sharada P. Mohanty, David P. Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in Plant Science, 7, 2016. 3", + "[30] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 3", + "[31] Alex Olsen, Dmitry A Konovalov, Bronson Philippa, Peter Ridd, Jake C Wood, Jamie Johns, Wesley Banks, Benjamin Girgenti, Owen Kenny, James Whinney, et al. Deepweeds: A multiclass weed species image dataset for deep learning. Scientific reports, 9(1):2058, 2019. 3", + "[32] OpenAI. Gpt-4o, 2024. Accessed: 2024-11-13. 2, 3, 6", + "[33] OpenAI. Gpt-4o mini, 2024. Accessed: 2024-11-13. 2, 6", + "[34] Chiranjit Pal, Imon Mukherjee, Sanjay Chatterji, Sanjoy Pratihar, Pabitra Mitra, and Partha Pratim Chakrabarti. Indian rice disease dataset (irdd), 2023. 2, 3", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[36] Corentin Royer, Bjoern Menze, and Anjany Sekuboyina. Multimedieval: A benchmark and a toolkit for evaluating medical vision-language models, 2024. 3" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Rahman Sanya, Ann Lisa Nabiryo, Jeremy Francis Tusubira, Sudi Murindanyi, Andrew Katumba, and Joyce Nakatumba-Nabende. Coffee and cashew nut dataset: A dataset for detection, classification, and yield estimation for machine learning applications. Data in Brief, 52:109952, 2024. 2", + "[38] Nabila Husna Shabrina, Siwi Indarti, Rina Maharani, Dinar Ajeng Kristiyanti, Irmawati, Niki Prastomo, and Tika Adilah M. A novel dataset of potato leaf disease in uncontrolled environment. Data in Brief, 52:109955, 2024. 2", + "[39] Risa Shinoda, Hirokatsu Kataoka, Kensho Hara, and Ryozo Noguchi. Transformer-based ripeness segmentation for tomatoes. Smart Agricultural Technology, 4:100196, 2023. 2", + "[40] Risa Shinoda, Kuniaki Saito, Shohei Tanaka, Tosho Hirasawa, and Yoshitaka Ushiku. Sbs figures: Pre-training figure qa from stage-by-stage synthesized images, 2024. 3", + "[41] Davinder Singh, Naman Jain, Pranjali Jain, Pratik Kayal, Sudhakar Kumawat, and Nipun Batra. Plantdoc: A dataset for visual plant disease detection. In Proc. ACM IKDD CoDS and COMAD, page 249-253, 2020. 2, 3", + "[42] Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, and Verena Widhalm. The cropandweed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In WACV, pages 3729-3738, 2023. 3, 5", + "[43] Daniel Steininger, Julia Simon, Andreas Trondl, and Markus Murschitz. Timbervision: A multi-task dataset and framework for log-component segmentation and tracking in autonomous forestry operations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2", + "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 3, 6", + "[45] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 2, 3", + "[46] Google Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024. 2, 3, 6", + "[47] Sandip Thite, Yogesh Suryawanshi, Kailas Patil, and Prawit Chumchu. Sugarcane leaf dataset: A dataset for disease detection and classification for machine learning applications. Data in Brief, 53:110268, 2024. 2, 3", + "[48] Liqiong Wang, Teng Jin, Jinyu Yang, Ales Leonardis, Fangyi Wang, and Feng Zheng. Agri-llava: Knowledge-infused large multimodal assistant on agricultural pests and diseases. arXiv preprint arXiv:2412.02158, 2024. 2", + "[49] Rujing Wang, Liu Liu, Chengjun Xie, Po Yang, Rui Li, and Man Zhou. Agripest: A large-scale domain-specific benchmark dataset for practical agricultural pest detection in the wild. Sensors, 21(5), 2021. 2, 3", + "[50] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7643", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. In NeurIPS, 2024. 2, 3, 6", + "[51] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3", + "[52] Tianqi Wei, Zhi Chen, Zi Huang, and Xin Yu. Benchmarking in-the-wild multimodal disease recognition and a versatile baseline. In Proc. ACM International Conference on Multimedia (ACMMM), 2024. 2, 3", + "[53] Xiaoping Wu, Chi Zhan, Yukun Lai, Ming-Ming Cheng, and Jufeng Yang. Ip102: A large-scale benchmark dataset for insect pest recognition. In CVPR, pages 8787-8796, 2019. 2, 3", + "[54] Momchil Yordanov, Raphaël d'Andrimont, Laura Martinez-Sanchez, Guido Lemoine, Dominique Fasbender, and Marijn van der Velde. Crop identification using deep learning on lucas crop cover photos. Sensors, 23(14), 2023. 2", + "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 2, 3, 8", + "[56] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 541 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7644", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_model.json b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9d09d00dce30a9d9d7e56554cc8232384c54d5dc --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "AgroBench: Vision-Language Model Benchmark in Agriculture" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.18, + 0.877, + 0.2 + ], + "angle": 0, + "content": "Risa Shinoda\\(^{1,2,4}\\), Nakamasa Inoue\\(^{3,4}\\), Hirokatsu Kataoka\\(^{4,5}\\), Masaki Onishi\\(^{4}\\), Yoshitaka Ushiku\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.215, + 0.795, + 0.272 + ], + "angle": 0, + "content": "\\(^{1}\\)The University of Osaka \\(^{2}\\)Kyoto University \\(^{3}\\)Institute of Science Tokyo \\(^{4}\\)National Institute of Advanced Industrial Science and Technology (AIST) \\(^{5}\\)Visual Geometry Group, University of Oxford \\(^{6}\\)OMRON SINIC X" + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.305, + 0.27, + 0.32 + ], + "angle": 0, + "content": "Disease Management" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.322, + 0.189, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.324, + 0.366, + 0.354 + ], + "angle": 0, + "content": "Question: What should be done to minimize the spread of the disease shown in the image?" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.357, + 0.365, + 0.386 + ], + "angle": 0, + "content": "Answer: Remove the male buds (bell) immediately using a forked stick." + }, + { + "type": "title", + "bbox": [ + 0.373, + 0.306, + 0.494, + 0.319 + ], + "angle": 0, + "content": "Weed Detection" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.322, + 0.455, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.325, + 0.63, + 0.359 + ], + "angle": 0, + "content": "Question: What species of weed does the highlighted bounding box represent in the image?" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.364, + 0.612, + 0.386 + ], + "angle": 0, + "content": "Answer: White goosefoot (Chenopodium album)" + }, + { + "type": "title", + "bbox": [ + 0.638, + 0.306, + 0.798, + 0.32 + ], + "angle": 0, + "content": "Disease Identification" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.322, + 0.72, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.33, + 0.896, + 0.352 + ], + "angle": 0, + "content": "Question: What disease is affecting the plant?" + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.37, + 0.856, + 0.38 + ], + "angle": 0, + "content": "Answer: Bacterial canker." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.391, + 0.248, + 0.405 + ], + "angle": 0, + "content": "Crop Management" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.407, + 0.189, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.41, + 0.365, + 0.438 + ], + "angle": 0, + "content": "Question: Is the crop shown in the image ready for harvest based on its current condition?" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.442, + 0.365, + 0.471 + ], + "angle": 0, + "content": "Answer: Yes, the leaves and stems have turned yellow, and the seed pods are brown to black." + }, + { + "type": "title", + "bbox": [ + 0.376, + 0.413, + 0.626, + 0.442 + ], + "angle": 0, + "content": "AGROBENCH" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.444, + 0.578, + 0.456 + ], + "angle": 0, + "content": "Agronomist Al Benchmark" + }, + { + "type": "title", + "bbox": [ + 0.639, + 0.391, + 0.773, + 0.404 + ], + "angle": 0, + "content": "Pest Identification" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.407, + 0.72, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.418, + 0.896, + 0.442 + ], + "angle": 0, + "content": "Question: What pest is affecting the plant?" + }, + { + "type": "text", + "bbox": [ + 0.724, + 0.455, + 0.832, + 0.467 + ], + "angle": 0, + "content": "Answer: Squash bug." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.476, + 0.256, + 0.489 + ], + "angle": 0, + "content": "Traditional Methods" + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.492, + 0.189, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.495, + 0.365, + 0.53 + ], + "angle": 0, + "content": "Question: What is one of the main benefits the process shown in the image?" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.533, + 0.365, + 0.555 + ], + "angle": 0, + "content": "Answer: Weed control and insect pest reduction." + }, + { + "type": "title", + "bbox": [ + 0.373, + 0.476, + 0.49, + 0.49 + ], + "angle": 0, + "content": "Machine Usage" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.495, + 0.454, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.494, + 0.63, + 0.524 + ], + "angle": 0, + "content": "Question: What is the main purpose of using the machine shown in the image?" + }, + { + "type": "text", + "bbox": [ + 0.459, + 0.527, + 0.631, + 0.556 + ], + "angle": 0, + "content": "Answer: To control weeds in the field without relying on chemical herbicides." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.478, + 0.671, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.685, + 0.481, + 0.839, + 0.496 + ], + "angle": 0, + "content": "Crop categories: 203" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.505, + 0.672, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.685, + 0.509, + 0.866, + 0.524 + ], + "angle": 0, + "content": "Disears categories: 684" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.532, + 0.673, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.684, + 0.538, + 0.821, + 0.552 + ], + "angle": 0, + "content": "Expert annotation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.567, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Figure 1. We present AgroBench (Agronomist AI Benchmark) designed to comprehensively evaluate 682 disease categories across 203 agricultural crop types for 7 vision-language question-answer tasks. In the era of larger-scale vision-language models (VLMs), our AgroBench is obviously non-trivial in terms of many more crop and disease categories with all expert annotations for establishing QA benchmarks in the agricultural domain." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.643, + 0.327, + 0.658 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Precise automated understanding of agricultural tasks such as disease identification is essential for sustainable crop production. Recent advances in vision-language models (VLMs) are expected to further expand the range of agricultural tasks by facilitating human-model interaction through easy, text-based communication. Here, we introduce AgroBench (Agronomist AI Benchmark), a benchmark for evaluating VLM models across seven agricultural topics, covering key areas in agricultural engineering and relevant to real-world farming. Unlike recent agricultural VLM benchmarks, AgroBench is annotated by expert agronomists. Our AgroBench covers a state-of-the-art range of categories, including 203 crop categories and 682 disease categories, to thoroughly evaluate VLM capabilities. In our evaluation on AgroBench, we reveal that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.766 + ], + "angle": 0, + "content": "VLMs have room for improvement in fine-grained identification tasks. Notably, in weed identification, most opensource VLMs perform close to random. With our wide range of topics and expert-annotated categories, we analyze the types of errors made by VLMs and suggest potential pathways for future VLM development. Our dataset and code are available at https://dahlian00.github.io/AgroBenchPage/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.645, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Agriculture is a fundamental process for humans to produce crops to live and stay healthy. With the development of computer vision technology, effective and automated management of external crop factors such as diseases and pests has been explored, contributing to stable crop" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7634" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.154, + 0.089, + 0.844, + 0.187 + ], + "angle": 0, + "content": "
DatasetAnnot.CropWeedDiseasePestImagesQA pairsMain Purpose
Agri-LLaVA [48]GPT-429-109112391k391k (Synthetic)Training
AgroInstruct [4]GPT-41744741270k70k (Synthetic)Training
CDDM [20]GPT-415-60-137k1M (Synthetic)Training
AgroBench (Ours)Expert2031086821343,7454,342 (Expert)Evaluation
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.198, + 0.905, + 0.227 + ], + "angle": 0, + "content": "Table 1. Comparison of agricultural vision datasets. Expert refers to 'human expert' in this context. AgroBench provides a comprehensive evaluation framework for assessing VLMs, featuring multiple tasks and a wide range of categories." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.239, + 0.482, + 0.331 + ], + "angle": 0, + "content": "production. This includes detecting and classifying undesirable conditions like diseases [1, 10, 13, 34, 38, 41, 47] and pests [3, 49, 53], as well as general plant management tasks such as crop classification [52, 54] and recognition of crop maturity [22, 25, 37, 39], and structure understanding [21, 43]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.53 + ], + "angle": 0, + "content": "To maintain stable crop production, it is important to recognize a wide range of crop conditions, including undesirable situations such as diseases and pests, and to know how to respond appropriately. A single visual model that can handle various conditions, not only disease and pest detection but also treatment and general crop management, would be highly beneficial. However, most existing approaches use task-specific models. These models usually require large amounts of training images and manual annotations for each task. As a result, farmers often need to use several different models depending on the situation. This increases complexity and makes the overall system less accessible for practical use in agriculture." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.535, + 0.483, + 0.852 + ], + "angle": 0, + "content": "For general purpose visual tasks, vision-language models (VLMs) [17, 19, 32, 33, 45, 46, 50] have become widespread recently because they can understand task definitions provided by natural language prompts, covering a wide range of applications without the need for task-specific model training. The recognition ability of VLMs is closely connected to image recognition itself and linked to an object's words, and supports open-vocabulary recognition through web-scale training. Here, zero-shot recognition and few-shot adaptation have also been realized by VLMs with language representations. This ability opens up a wide range of applications; therefore, we believe this can be applied to agricultural scenarios as well. It is worth noting that VLMs offer an easy-to-use interface for the general public, especially the question-answer (QA) and conversation modes. To investigate the range of tasks that VLMs obtained through large-scale vision-language training can cover, recent studies have introduced various benchmark datasets to fully evaluate the VLMs, including tasks such as understanding diagrams and charts [26, 28], and question answering requiring specialized knowledge [7, 55, 56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "However, VLM research remains underexplored in agriculture due to a lack of benchmark datasets that include diverse tasks and categories in agriculture. Several pioneer-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.905, + 0.466 + ], + "angle": 0, + "content": "ing works [20, 48] have adapted open-source VLMs such as LLaVA [18] to the agricultural domain by fine-tuning them on synthetic datasets generated from closed-source VLMs such as GPT-4o. This is recognized as an effective approach because black-box VLMs often possess a certain amount of agricultural knowledge obtained from the Internet. While this helps in generating responses that are generally acceptable to experts, there is almost no way to verify whether the answers are truly correct. Moreover, in the limited evaluation of categories, we are insufficient for fully assessing VLM knowledge; whether they can answer a wide range of types, such as diseases, pests, and weeds. These limitations motivated us to develop a benchmark dataset in order to evaluate the VLM's broad knowledge in the agricultural domain and its applicability as a practical application." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.905, + 0.817 + ], + "angle": 0, + "content": "Here, we introduce AgroBench (Agronomist AI Benchmark), a comprehensive benchmark dataset for VLM for the agricultural domain, covering a state-of-the-art range of categories for agriculture-focused benchmark datasets for VLM; 682 disease, 134 pest, 108 weed, and 203 crop categories (Figure 1). We cover not only identification tasks but also crop production and disease management knowledge. Moreover, we include other important topics related to crop management with 98 machine categories and 77 traditional management methods to investigate more about VLM's ability. We carefully selected benchmark tasks from key agricultural engineering research areas, and also the tasks that address challenges faced by farmers in real agricultural scenarios. We employ a multiple-choice format, and all questions are annotated by human agronomist experts, which overcomes the limitations of previous synthetically created datasets. We carefully selected images from Creative Commons-licensed and publicly available datasets, including real farm conditions. Also, under our manual annotation process, unclear images were removed. As shown in Table 1, AgroBench contains data with the most diverse disease and crop variation categories and many more crop/weed annotations." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.821, + 0.792, + 0.836 + ], + "angle": 0, + "content": "Our main contributions are as follows;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.9 + ], + "angle": 0, + "content": "- We developed AgroBench, a benchmark dataset for VLM, to assess their broad agricultural knowledge of VLMs and evaluate their applicability in practical applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7635" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "- In our AgroBench evaluation, VLMs tend to achieve high performance in disease and crop management tasks; however, there is still room for improvement in weed and disease identification." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.152, + 0.483, + 0.198 + ], + "angle": 0, + "content": "- Our AgroBench enables error analysis by providing broader category annotations, highlighting future directions for model training focus areas." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.213, + 0.232, + 0.229 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.38, + 0.256 + ], + "angle": 0, + "content": "2.1. Computer Vision for Agriculture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.488 + ], + "angle": 0, + "content": "With the rise of computer vision techniques driven by advancements in deep learning, a wide range of agricultural tasks have been explored over the past decade. In particular, dataset construction and benchmarking play an important role in developing models and practical applications in agriculture, while also paving the way for forward-looking advancements. Disease identification is a key research focus across various crops, such as rice [1, 34], tomato [10], cacao [13], and sugarcane [47]. Multicrop datasets have also been created [29, 41]. PlantDoc [41] covers 13 species and 17 classes focused on leaf-based diseases. Plant Village [29] offers 39 classes containing both diseased and healthy leaf categories. In addition to crop disease identification, pest identification [49, 53], weed identification [12, 31, 42] have also been studied." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.716 + ], + "angle": 0, + "content": "While most of the agricultural datasets primarily focus on visual data, few computer vision studies in agriculture investigate the potential of multi-modal approaches. The PlantWild dataset [52] includes 56 plant-disease class pairs, which are collected through image search engines. They implement a CLIP-based model and show the possibility of training with combined text and image data. In an instruction-based format, the CDDM [20] has created 16 categories of crops and 60 categories of crop diseases dataset, which generates instructional data using GPT. While these synthetically created datasets explore the agriculture multi-modal models, there remains a lack of datasets for comprehensive multi-modal model evaluation, validated by human experts and covering a wide range of tasks and categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.318, + 0.743 + ], + "angle": 0, + "content": "2.2. Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Models. Visual models in the computer vision field have been accelerated by language modality with the data resource of sentence-level inputs and web-scale texts. Specifically, CLIP [35] has made a significant contribution in this context, by text and image feature alignment through contrastive learning. CLIP played a key role in introducing more sophisticated visual representation into language explanations e.g., Flamingo [2], BLIP [14, 15], Qwen [5], PaLI [8], LLaVA [17, 18], CogVLM [50], and Emu [44, 45, 51]. Closed-source VLMs achieve state-of-the-art per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "formance, such as GPT-4o [32] and Gemini Pro [46], which are said to acquire human-level knowledge across diverse fields on the Internet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.379 + ], + "angle": 0, + "content": "**Benchmarks.** Along this line, several recent studies introduced benchmark datasets. Especially, the representative examples in terms of universal knowledge benchmarking include MMMU [55], MMMU-Pro [56] and MMStar [6] for multi-modal understanding. These benchmarks contain highly diverse domains (e.g., natural, graph, illustration, medical images) and academic fields (e.g., science, engineering, art, and medical fields) under the tasks of vision language such as question-answering and reasoning. The series of MMMUs have been verified with foundation models like GPT-4V, but they cannot answer the questions perfectly. More specific domain datasets have also been proposed, such as those for Medicine [27, 36], Chart [26, 28, 40], and Video understandings [9, 16, 30], contributing to the accurate evaluation of VLMs and guiding future research directions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.392, + 0.634, + 0.409 + ], + "angle": 0, + "content": "3. AgroBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.568 + ], + "angle": 0, + "content": "This section introduces AgroBench, the first comprehensive benchmark dataset to evaluate VLM models from the perspective of agricultural vision tasks. The benchmark consists of seven tasks covering a wide range of tasks selected from key agricultural engineering research areas, as well as tasks that address real-world challenges faced by farmers in real agriculture scenarios. AgroBench includes 682 disease categories, 134 pest categories, 203 crop categories, and 108 weed categories, representing the largest number of categories in each area to date, to the best of our knowledge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.864, + 0.594 + ], + "angle": 0, + "content": "3.1. Benchmark Tasks in Agricultural Scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.906, + 0.674 + ], + "angle": 0, + "content": "The seven benchmark tasks encompass key research areas in agricultural engineering as well as real-world challenges faced by farmers. To facilitate VLM evaluation, we also provide prompts to address each task in a question-answer format. The details of each task are described below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.855 + ], + "angle": 0, + "content": "1) Disease Identification (DID). The DID task aims to accurately diagnose and classify crop diseases. This is a key task in agriculture to protect crop health and maximize yields. Our benchmark provides 1,502 QA pairs that cover 370 disease categories, 160 crop categories and 682 crop-disease combinations. To thoroughly evaluate the capabilities of VLMs, we include four misleading disease labels for each image, featuring diseases with similar symptoms or common diseases affecting the target plant. VLMs are required to diagnose the disease based on the image, considering both symptoms and the crop species. Figure 2a illustrates example images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "2) Pest Identification (PID). The PID task aims to identify pests to prevent infestations that can severely impact crop health. Accurate identification reduces economic losses and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7636" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.092, + 0.248, + 0.102 + ], + "angle": 0, + "content": "(a) Disease identification (DID)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.103, + 0.899, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.438, + 0.904, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Examples of labeled images for DID, PID, and WID tasks. Our dataset includes 682 crop-disease pairs, 134 pest categories, and 108 weed categories. We prioritized collecting images from real farm settings." + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.481, + 0.436, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Figure 3. Seven benchmark tasks in AgroBench. AgroBench includes multiple topics with a diverse range of categories. The total accuracy is calculated by the average of each task to mitigate the difference in QAs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "minimizes harm to the environment. Our benchmark provides 544 labeled images that cover 134 pest categories, including insects, mites, and other organisms harmful to plants. For categories where we obtain multiple images, we select multiple insect growth stages whenever possible. To fully evaluate the VLMs, we assign alternative choices that closely resemble or are commonly associated with the target" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.889, + 0.494 + ], + "angle": 0, + "content": "crops. Figure 2b shows examples of labeled pest images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.635 + ], + "angle": 0, + "content": "3) Weed Identification (WID). The WID task aims to identify weed species. Our benchmark includes 609 images of weeds with ground-truth bounding boxes, covering 108 weed species commonly found in farm fields. We assign bounding boxes because multiple weed types often grow closely together, and we want to clarify which one is the target. Specifically, VLMs are required to identify the weed species within the provided bounding box on the image. Figure 2c shows examples of labeled weed images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.837 + ], + "angle": 0, + "content": "4) Crop Management (CMN). Crop management focuses on optimizing farming practices to facilitate crop growth. This involves making decisions on irrigation, fertilization, planting times, and other cultivation practices. Our benchmark provides 411 question-answer pairs for this task. VLMs are required to analyze images of crops and recommend appropriate management strategies by considering factors such as crop health, growth stage, and environmental conditions visible in the images given five answer candidates. In Figure 4a and b, we show two examples of the harvest timing for white and green asparagus, respectively. Our dataset includes complex questions that take into account the differences in their harvest timing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "5) Disease Management (DMN). Disease management aims to control and reduce diseases in crops. This involves informed decisions on interventions such as applying pesticides, adopting resistant crop varieties, or modifying cul" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.498, + 0.909, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.091, + 0.19, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.092, + 0.501, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.092, + 0.597, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.092, + 0.897, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.168, + 0.19, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.168, + 0.488, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.168, + 0.595, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.168, + 0.897, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Figure 4. Examples of QA pairs for CMN and DMN tasks. (a) and (b) Crop management QA types for white asparagus and asparagus, respectively. Their difference in the harvest timing affects the answer's difference correctly. (c) and (d) Disease management QA types for the alfalfa bacterial leaf spot with the initial and severe symptoms, respectively. Based on the severity of the symptoms, the annotator changes the answer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.485, + 0.531 + ], + "angle": 0, + "content": "tivation practices. Our benchmark provides 569 question-answer pairs covering 141 crop-disease combination categories. Since many diseases share the same management strategies regardless of the crop, we carefully select a diverse range of disease types and management strategies. VLMs are required not only to identify the disease but also to recommend appropriate management strategies based on images of affected crops. In Figure 4c and d show example QAs from the Disease Management tasks. Both images depict Bacterial leaf spot on alfalfa, with c showing the initial stage of the disease and d showing the severe stage. Based on the severity of the disease, we provide different answer options and set different correct answers. The input prompt consists of a question and five answer candidates." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.557, + 0.483, + 0.724 + ], + "angle": 0, + "content": "6) Machine Usage QA (MQA). Machine usage QA addresses the correct use and choice of agricultural machinery depending on the task and farming conditions. Selecting the appropriate machinery is essential for efficient farming. Our benchmark provides 303 question-answer pairs covering 98 machine categories. Given that many crops share the same machinery (e.g., soil preparation and irrigation), the coverage of this category is comprehensive. VLMs are required to answer questions about machinery operation or select the appropriate machine for a given scenario based on images of machinery or farming conditions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.9 + ], + "angle": 0, + "content": "7) Traditional Management (TM). Traditional management methods involve natural and sustainable approaches to farming, such as the use of organic fertilizers, terrace farming, and agroforestry. Recent computer vision studies have not focused on these traditional practices, although many are still used by certain local farmers. Our benchmark includes 404 question-answer pairs, including 77 traditional management practices. VLMs are required to identify the management method or explain its effectiveness, given five answer choices." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.319, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.715, + 0.333 + ], + "angle": 0, + "content": "3.2. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Image Selection. To establish a high-quality benchmark covering a wide range of crops, diseases, and pest categories we initially curated around 50,000 agricultural images from websites supervised by plant pathologists, either where redistribution was permitted or where we obtained redistribution permission. We selected images in real farm settings as much as possible, as shown in Figure 2, to evaluate real-world scenarios. When licensed images for a target category were limited, we used laboratory setting images. For curation, we obtained images along with their corresponding labels. The annotator is one of the authors, who holds a Ph.D. in Agriculture. In the human evaluation conducted during the experiment, other individuals with a Ph.D. or M.S. degree in Agriculture reviewed questions to assess the quality of the QA pairs. The annotator selected images that clearly represented target labels and removed those images with irrelevant labels. For the Machine usage QA and Traditional management, annotators manually created categories based on textbooks and websites, and searched for corresponding images with redistribution licenses. For Weed identification, we use existing dataset [11, 23, 24, 42]. For this Weed identification category, we will provide a simple code to download the data and crop images and assign bounding boxes for Weed Identification, allowing users to work with the existing dataset without redistributing the images. Through these image selection processes, the selected images are high-quality 4,218 representative images. The detailed distribution of tasks and categories for each task is summarized in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "QA Annotation. For the DMN, CMN, MQA, and TM tasks, all question-answer pairs were created manually, independent of any LLMs or VLMs. Annotators used GPT only for sentence rephrasing, but they were prohibited from using any knowledge from GPT for QA creation. Annotators were allowed to refer to textbooks, academic journals, and other authoritative sources in the field to ensure the accuracy and depth of the dataset. This took around 150 man" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7638" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.855, + 0.396 + ], + "angle": 0, + "content": "
ModelDIDDMNPIDWIDCMNMQATMOverall (all)Overall (subset)
Random Choice21.7715.6420.4017.9016.0622.1119.3119.0319.11
Human25.0022.5045.0020.0036.2557.5051.25-36.79
Closed-Source Vision Language Models (VLMs)
GPT-4o mini [33]53.6080.6760.0435.1464.2370.9669.8062.0669.65
GPT-4o [32]64.1887.3577.7644.1775.4382.8482.4373.4579.26
Gemini1.5-Flash [46]55.0679.9670.0450.9064.7278.2273.2767.4568.82
Gemini1.5-Pro [46]62.9281.5574.4555.1771.0582.8477.7272.2469.74
Open-Source Vision Language Models (VLMs)
EMU2Chat [44]42.0148.3343.7523.8140.3937.6247.7740.5333.84
LLaVA-Next-8B [19]45.4772.5843.0130.0554.2656.1157.4651.2857.84
LLaVA-Next-72B [19]54.9580.0049.8126.9866.9266.1170.3859.3164.36
QwenVLM-7B [5]51.2680.4963.9733.1766.4276.2477.4864.1566.41
QwenVLM-72B [5]57.9987.8773.3534.4875.9180.8684.1670.6672.45
CogVLM-19B [50]29.1653.7852.3925.4554.0171.6266.0950.3644.27
LLaVa-7B [18]36.0262.7438.7924.7953.7746.5355.2045.4146.14
LLaVA-13B [18]40.2168.8944.4924.7959.3754.1358.4250.0455.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.408, + 0.905, + 0.437 + ], + "angle": 0, + "content": "Table 2. Results for the seven benchmark tasks with images. We provide results for Random Choice, Human Validation, four closed-source VLMs, and open-source VLMs. Human validation was conducted by 28 people on a subset of 80 samples per task as a reference." + }, + { + "type": "table", + "bbox": [ + 0.185, + 0.45, + 0.816, + 0.508 + ], + "angle": 0, + "content": "
DIDDMNPIDWIDCMNMQATMOverall
GPT-4o [32]1.9372.5818.751.0040.3925.0848.2729.71
LLaVA-Next-8B [19]26.1070.3021.8819.7053.7730.3640.3537.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.519, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Table 3. Results for the seven benchmark tasks with text only. We additionally evaluate the model without image inputs, and the overall performance is close to random." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.616 + ], + "angle": 0, + "content": "hours. We carefully annotate various types of questions. Examples of QAs that require expert knowledge are shown in Figure 4. All questions were carefully created so that image reference is necessary for answering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Dataset Statistics. Following the dataset annotations and selections, our AgroBench comprises seven tasks with 4,342 QA pairs, as shown in Table 1. All tasks comprise a wide range of categories for detailed evaluation. Please refer to the supplementary materials for more details and examples." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.224, + 0.735 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.298, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Baseline Models. We use four closed-source models: GPT-4o [32], GPT-4o mini [33], Gemini1.5-Pro [46], and Gemini1.5-Flash [46]. GPT-4o mini and Gemini1.5-Flash are down-scale versions of GPT-4o and Gemini1.5-Pro, respectively. We use eight open-source models: EMU2Chat [44], LLaVA-Next-8B [19], LLaVA-Next-72B [19], QwenVLM-7B [5], QwenVLM-72B [5], CogVLM-19B [50], LLaVa-7B [18], and LLaVa-13B [18]. For the details of these models, please refer to the supple" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.634, + 0.571 + ], + "angle": 0, + "content": "mentary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Human Results. We also present results from human participants for reference. We surveyed 28 students, each holding at least a bachelor's degree in agriculture, and asked them to answer 20 questions each. This created a test subset of 280 questions, with each question answered by two participants, resulting in a total of 560 responses. We averaged the results per task and reported the accuracy. Each participant was permitted to use a book or translator to look up word meanings but was prohibited from using the internet for searches. If participants were unsure of the answer, they were asked to provide the response they believed to be most accurate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation Protocol. Importantly, our dataset evaluation is conducted per task, and overall scores are averaged based on the number of tasks, not the number of QAs. This prevents categories with a large number of evaluations from becoming dominant. We adopt an exact matching approach for our five-option questions. If the model's response matches the option's letter or the answer sentence, we consider it correct. If the model's answer does not match any option, including cases where there is no answer or mul" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7639" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.904, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Results of seven benchmark tasks with Chain of Thought (CoT). Baseline indicates results without CoT. In the one-shot, two-shot, and three-shot settings, we provide one, two, and three CoT examples per task, respectively, to guide the model." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.294, + 0.9, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Figure 6. Error analysis on seven benchmark tasks with GPT-4o. We extract a maximum of 15 errors per task from the zero-shot CoT result. We manually analyze how they conclude the incorrect answer." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.532, + 0.348, + 0.547 + ], + "angle": 0, + "content": "tiple answers, we consider it incorrect." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.231, + 0.573 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Here, we discuss the main results of our AgroBench evaluation performance. We evaluate our AgroBench using eight open-source VLMs and four closed-source VLMs with APIs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Challenges of the AgroBench. We show the main results in Table 2. The most difficult task is Weed Identification (WID), on which most open-source VLMs perform at around the random score. The highest WID accuracy, \\(55.17\\%\\), was achieved by Gemini 1.5-Pro. This suggests that the knowledge about weeds is not as fully trained as that of crops. All models' Disease Identification (DID) results are lower than their Disease Management (DMN) results. This means that VLM models can gain contextual information, but there is still room for perceptual improvement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Model Comparison. Overall, closed-source VLMs achieve better results than open-source VLMs and achieve higher performance than humans. GPT-4o model achieved the highest score in overall performance. Among the open-source models, QwenVLM-72B achieves the best result in overall accuracy, which is comparable or even superior to open-source VLMs in some tasks. QwenVLM" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.563 + ], + "angle": 0, + "content": "72B achieves satisfactory results on both identification and question-answering benchmarks on AgroBench." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.625, + 0.59 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Context Reliance. We further evaluate the results on AgroBench using text-only input to determine whether visual information is necessary to answer the questions. Table 3 presents the experimental results across seven benchmark tasks on AgroBench. With text-only input, performance drops across all models, confirming the reliance on visual information. However, for both models, the DMN, CMN, and TM tasks maintain significantly higher accuracy than random selection. Although we ensured that our questions do not include disease names or appearance-related traits, models tend to infer answers based on estimation. This suggests that many disease types share common management strategies, such as avoiding humidity, preferring cooler temperatures, or pruning infected parts, allowing models to predict the most likely option. A similar pattern is observed in the CMN and TM tasks, where models can make educated guesses based on contextual cues. (See the supplementary material for further detailed examples.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Chain of Thought. We evaluate the effectiveness of Chain of Thought (CoT) reasoning. In the one-shot setting, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7640" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.096, + 0.264, + 0.111 + ], + "angle": 0, + "content": "(a) Lack of Knowledge" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.113, + 0.316, + 0.125 + ], + "angle": 0, + "content": "Q: What disease is affecting the plant?" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.127, + 0.255, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.126, + 0.472, + 0.219 + ], + "angle": 0, + "content": "GPT-4o: ..., **Sclerotinia blight**: Characterized by white, cottony mycelium and black sclerotia on stems. The image does not show these symptoms. (not enough knowledge) ..., the plant is likely affected by **D. Cylindrocladium black rot**: (This is more common on root)" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.222, + 0.216, + 0.234 + ], + "angle": 0, + "content": "A: Sclerotinia blight" + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.241, + 0.249, + 0.255 + ], + "angle": 0, + "content": "(b) Perceptual Error" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.256, + 0.427, + 0.279 + ], + "angle": 0, + "content": "Q: What activity is being performed using the tractor-pulled machine shown in the image?" + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.28, + 0.266, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.278, + 0.471, + 0.37 + ], + "angle": 0, + "content": "GPT-40:The machine seems to be designed to lift plants from the soil, ..., and there are plants being lifted from the ground (There are no harvested ctops)....the activity being performed is: D. Harvesting potato tubers from the soil using a tractor-pulled harvesting machine." + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.374, + 0.4, + 0.398 + ], + "angle": 0, + "content": "A: Cutting and chopping the stems and leaves using a tractor-pulled leaf chopper." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.405, + 0.483, + 0.434 + ], + "angle": 0, + "content": "Figure 7. Error examples of GPT-4o. Examples illustrate the two main error types: Lack of Knowledge and Perceptual Error." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.685 + ], + "angle": 0, + "content": "provide a single CoT reasoning example along with an image and the corresponding prompt for each task to guide the model in answering the question. We select 100 QAs per task for evaluation and use GPT-4 as the VLM. Figure 5 presents the accuracy of CoT reasoning with the baseline results. While CoT achieves slightly higher accuracy, the improvement is not significant compared to the baseline. However, CoT demonstrates effectiveness in certain tasks, such as PID, WID, CMN, and TM. For instance, in PID, the VLM performs step-by-step and careful reasoning (see the supplementary material for more detailed examples). In WID, the most challenging task, CoT provides useful examples that may help the model make predictions. Overall, CoT contributes to accuracy, but we observe performance saturation in the three-shot setting. For further detailed examples, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.242, + 0.712 + ], + "angle": 0, + "content": "4.4. Error Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Here, we analyze the mistake types that occur depending on the agricultural tasks. We bring out up to 15 failure examples per task from the zero-shot CoT results and manually analyze how they reached the incorrect answers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Lack of Knowledge (51.92%). This includes cases where VLM can't accurately describe the appearance or relevant knowledge of a choice (e.g., VLM fails to describe disease symptoms or insect characteristics) or lacks context (e.g., VLM doesn't know how to treat diseased crops or manage crops for high yield). Figure 7a shows an example of a Lack of Knowledge case. When the VLM analyzes the correct answer option for Sclerotinia blight, it fails to describe the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "symptoms of discoloration and wilting. Additionally, its incorrect answer choice is more commonly associated with the root rather than the stem. These errors are based on a Lack of Knowledge, suggesting VLMs need more detailed categories and domain-specific training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Perceptual Error (32.69%). This indicates that VLM can't pay attention or recognize the answer-related part in the image (e.g., can't recognize the green insect on the leaf), and VLM misunderstands the image, leading to incorrect answers. Figure 7b shows an example of the model hallucinating and misunderstanding the situation. First, the VLM incorrectly identifies the machine as one used for lifting plants from the soil. Then, it describes the scene as if there are harvested crops, even though no harvested crops are present. These errors can be mitigated by enhancing the VLM's perception abilities for domain-specific classes. Additionally, improving general perception capabilities, including reducing hallucinations, can further contribute to VLM performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.906, + 0.473 + ], + "angle": 0, + "content": "Reasoning Error (7.6%). Reasoning error involves the VLM can describe the options correctly, but can't compare them step by step and conclude the wrong answer. This error is relatively low compared to the existing work [55] since AgroBench requires more specific knowledge and doesn't include reasoning relying on problems (e.g., math)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Other Errors (7.79%). For the other errors, we observe Shortcut Error (The VLM can pick up the two candidate options correctly but conclude the answer without comparing the candidates), Double Answer Error (Concluding two answers are correct), Interpretation Misunderstanding (VLM misleading the question and conclude wrong answer), and Reject to Answer (VLM conclude there is no answer)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.6, + 0.634, + 0.616 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.764 + ], + "angle": 0, + "content": "In this paper, we develop AgroBench, a comprehensive benchmark dataset for VLMs in the agricultural domain, covering a state-of-the-art range of categories. Our dataset comprises seven benchmark tasks encompassing key research areas in agricultural engineering as well as real-world challenges faced by farmers. AgroBench contributes to agricultural VLM research by addressing the lack of datasets for comprehensive multi-modal model evaluation, validated by human experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In our evaluation, VLMs exhibit strengths across different tasks. However, in several tasks such as weed identification and disease identification, all models show room for improvement. Our error analysis reveals that most failures are due to a lack of knowledge (51.92%), suggesting that VLMs require more specialized agricultural knowledge. Our dataset will facilitate agricultural VLM research, enabling broad category and task evaluation to support sustainable, automated agriculture." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7641" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.262, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.228 + ], + "angle": 0, + "content": "This work was supported by the AIST KAKUSEI Project (FY2024) and JST FOREST Grant Number JPMJFR206F. We would like to thank the Agricultural Administration Division, Department of Agriculture, Hokkaido Government, for providing some of the images. We are also grateful to Daniel Steininger for contributing images from the CropAndWeed dataset to our dataset. We used ABCI 3.0 provided by AIST and AIST Solutions." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.255, + 0.188, + 0.272 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.281, + 0.484, + 0.336 + ], + "angle": 0, + "content": "[1] Petchiammal A, Briskline Kiruba S, Murugan D, and Pandarasamy Arjunan. Paddy doctor: A visual image dataset for automated paddy disease classification and benchmarking. IEEE Dataport, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.339, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.493, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[3] Abderraouf Amrani, Ferdous Sohel, Dean Diepeveen, David Murray, and Michael G.K. Jones. Deep learning-based detection of aphid colonies on plants from a reconstructed brassica image dataset. Computers and Electronics in Agriculture, 205:107587, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.564, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[4] Muhammad Awais, Ali Husain Salem Abdulla Alharthi, Amandeep Kumar, Hisham Cholakkal, and Rao Muhammad Anwer. Agropt: Efficient agricultural vision-language model with expert tuning. arXiv preprint arXiv:2410.08405, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.634, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao," + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.281, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.136, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Alexander Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.276, + 0.904, + 0.317 + ], + "angle": 0, + "content": "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees G. M. Snoek, and Yuki M. Asano. Tvbench: Redesigning video-language evaluation. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[10] Mamta Gehlot, Rakesh Saxena, and Geeta Gandhi. \"tomato-village\": a dataset for end-to-end tomato disease detection in a real-world environment. Multimedia Systems, 29:1-24, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.861, + 0.389 + ], + "angle": 0, + "content": "[11] Benedikt Geisler. Perrenial plants detection, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[12] Sebastian Haug and Jörn Ostermann. A crop/weed field image dataset for the evaluation of computer vision based precision agriculture tasks. In Proc. European Conference on Computer Vision (ECCV), pages 105-116. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[13] Atuhurra Jesse, N'guessan Yves-Roland Douha, and Pabitra Lenka. Image classification for cssvd detection in cacao plants. arXiv preprint 2405.04535, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[16] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.904, + 0.758 + ], + "angle": 0, + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[19] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.2,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[20] Xiang Liu, Zhaoxiang Liu, Huan Hu, Zezhou Chen, Kohou Wang, Kai Wang, and Shiguo Lian. A multimodal benchmark dataset and model for crop disease diagnosis. In Proc. European Conference on Computer Vision (ECCV), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Xinpeng Liu, Kanyu Xu, Risa Shinoda, Hiroaki Santo, and Fumio Okura. Masks-to-skeleton: Multi-view mask-based" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7642" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tree skeleton extraction with 3d gaussian splatting. Sensors, 25(14), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[22] Shenglian Lu, Wenkang Chen, Xin Zhang, and Manoj Karkee. Canopy-attention-yolov4-based immature/mature apple fruit detection on dense-foliage tree architectures for early crop load estimation. Computers and Electronics in Agriculture, 193:106696, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.398, + 0.206 + ], + "angle": 0, + "content": "[23] Yuzhen Lu. Cottonweeddet3. kaggle, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[24] Simon Leminen Madsen, Solvejg Kopp Mathiassen, Mads Dyrmann, Morten Stigaard Laursen, Laura-Carlota Paz, and Rasmus Nyholm Jorgensen. Open Plant Phenotype Database of Common Weeds in Denmark, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[25] Abdul Khalique Maitlo, Abdul Aziz, Hassnian Raza, and Neelam Abbas. A novel dataset of guava fruit for grading and classification. Data in Brief, 49:109462, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[26] Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, Dublin, Ireland, 2022. Association for Computational Linguistics. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[27] João Matos, Shan Chen, Siena Placino, Yingya Li, Juan Carlos Climent Pardo, Daphna Idan, Takeshi Tohyama, David Restrepo, Luis F. Nakayama, Jose M. M. Pascual-Leone, Guergana Savova, Hugo Aerts, Leo A. Celi, A. Ian Wong, Danielle S. Bitterman, and Jack Gallifant. Worldmedqa-v: a multilingual, multimodal medical examination dataset for multimodal language models evaluation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.531 + ], + "angle": 0, + "content": "[28] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1516-1525, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[29] Sharada P. Mohanty, David P. Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in Plant Science, 7, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[30] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[31] Alex Olsen, Dmitry A Konovalov, Bronson Philippa, Peter Ridd, Jake C Wood, Jamie Johns, Wesley Banks, Benjamin Girgenti, Owen Kenny, James Whinney, et al. Deepweeds: A multiclass weed species image dataset for deep learning. Scientific reports, 9(1):2058, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.45, + 0.716 + ], + "angle": 0, + "content": "[32] OpenAI. Gpt-4o, 2024. Accessed: 2024-11-13. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.466, + 0.731 + ], + "angle": 0, + "content": "[33] OpenAI. Gpt-4o mini, 2024. Accessed: 2024-11-13. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[34] Chiranjit Pal, Imon Mukherjee, Sanjay Chatterji, Sanjoy Pratihar, Pabitra Mitra, and Partha Pratim Chakrabarti. Indian rice disease dataset (irdd), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Corentin Royer, Bjoern Menze, and Anjany Sekuboyina. Multimedieval: A benchmark and a toolkit for evaluating medical vision-language models, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Rahman Sanya, Ann Lisa Nabiryo, Jeremy Francis Tusubira, Sudi Murindanyi, Andrew Katumba, and Joyce Nakatumba-Nabende. Coffee and cashew nut dataset: A dataset for detection, classification, and yield estimation for machine learning applications. Data in Brief, 52:109952, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[38] Nabila Husna Shabrina, Siwi Indarti, Rina Maharani, Dinar Ajeng Kristiyanti, Irmawati, Niki Prastomo, and Tika Adilah M. A novel dataset of potato leaf disease in uncontrolled environment. Data in Brief, 52:109955, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[39] Risa Shinoda, Hirokatsu Kataoka, Kensho Hara, and Ryozo Noguchi. Transformer-based ripeness segmentation for tomatoes. Smart Agricultural Technology, 4:100196, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[40] Risa Shinoda, Kuniaki Saito, Shohei Tanaka, Tosho Hirasawa, and Yoshitaka Ushiku. Sbs figures: Pre-training figure qa from stage-by-stage synthesized images, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[41] Davinder Singh, Naman Jain, Pranjali Jain, Pratik Kayal, Sudhakar Kumawat, and Nipun Batra. Plantdoc: A dataset for visual plant disease detection. In Proc. ACM IKDD CoDS and COMAD, page 249-253, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[42] Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, and Verena Widhalm. The cropandweed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In WACV, pages 3729-3738, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[43] Daniel Steininger, Julia Simon, Andreas Trondl, and Markus Murschitz. Timbervision: A multi-task dataset and framework for log-component segmentation and tracking in autonomous forestry operations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[45] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[46] Google Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[47] Sandip Thite, Yogesh Suryawanshi, Kailas Patil, and Prawit Chumchu. Sugarcane leaf dataset: A dataset for disease detection and classification for machine learning applications. Data in Brief, 53:110268, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[48] Liqiong Wang, Teng Jin, Jinyu Yang, Ales Leonardis, Fangyi Wang, and Feng Zheng. Agri-llava: Knowledge-infused large multimodal assistant on agricultural pests and diseases. arXiv preprint arXiv:2412.02158, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[49] Rujing Wang, Liu Liu, Chengjun Xie, Po Yang, Rui Li, and Man Zhou. Agripest: A large-scale domain-specific benchmark dataset for practical agricultural pest detection in the wild. Sensors, 21(5), 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7643" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. In NeurIPS, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[51] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[52] Tianqi Wei, Zhi Chen, Zi Huang, and Xin Yu. Benchmarking in-the-wild multimodal disease recognition and a versatile baseline. In Proc. ACM International Conference on Multimedia (ACMMM), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[53] Xiaoping Wu, Chi Zhan, Yukun Lai, Ming-Ming Cheng, and Jufeng Yang. Ip102: A large-scale benchmark dataset for insect pest recognition. In CVPR, pages 8787-8796, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[54] Momchil Yordanov, Raphaël d'Andrimont, Laura Martinez-Sanchez, Guido Lemoine, Dominique Fasbender, and Marijn van der Velde. Crop identification using deep learning on lucas crop cover photos. Sensors, 23(14), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[56] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7644" + } + ] +] \ No newline at end of file diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_origin.pdf b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84f2c7b5bebc9cee520b824f1c15e9df570b471f --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/b17f54a3-0198-426c-98d3-db524df55ee2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b607028458f3f3402b7c03bac879c6202fd6351da77eef6a089a0ce16a958194 +size 5823179 diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/full.md b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75a5be836567021e7ab13a7ee3ba9214f53e6545 --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/full.md @@ -0,0 +1,342 @@ +# AgroBench: Vision-Language Model Benchmark in Agriculture + +Risa Shinoda $^{1,2,4}$ , Nakamasa Inoue $^{3,4}$ , Hirokatsu Kataoka $^{4,5}$ , Masaki Onishi $^{4}$ , Yoshitaka Ushiku $^{6}$ + +$^{1}$ The University of Osaka $^{2}$ Kyoto University $^{3}$ Institute of Science Tokyo $^{4}$ National Institute of Advanced Industrial Science and Technology (AIST) $^{5}$ Visual Geometry Group, University of Oxford $^{6}$ OMRON SINIC X + +# Disease Management + +![](images/5701473a88fe9a3811a1637fcae31cbd6958bedd62476c53adf9de487a8a6479.jpg) + +Question: What should be done to minimize the spread of the disease shown in the image? + +Answer: Remove the male buds (bell) immediately using a forked stick. + +# Weed Detection + +![](images/df59a31a6a22a927741e16ecb06aa63b545f796fee8dbece99e2200770492df4.jpg) + +Question: What species of weed does the highlighted bounding box represent in the image? + +Answer: White goosefoot (Chenopodium album) + +# Disease Identification + +![](images/261f7d4baca2a3ae8dc30f2278b9f5708a2b561e849d34ed807a15d5feb3b5e1.jpg) + +Question: What disease is affecting the plant? + +Answer: Bacterial canker. + +# Crop Management + +![](images/c72cc416234804a15201207f62ef335610b36cdbe555bb2ac6be7f9999a9a69f.jpg) + +Question: Is the crop shown in the image ready for harvest based on its current condition? + +Answer: Yes, the leaves and stems have turned yellow, and the seed pods are brown to black. + +# AGROBENCH + +Agronomist Al Benchmark + +# Pest Identification + +![](images/32e5cc5c8fcb0136d5eb28ece6926ea4c6e65e4a023f9cfa8b4fb63050430701.jpg) + +Question: What pest is affecting the plant? + +Answer: Squash bug. + +# Traditional Methods + +![](images/9d6aed769e25151ee011dcd88d0d57dfe8be179abeb1afb06b8e009d85285d07.jpg) +Figure 1. We present AgroBench (Agronomist AI Benchmark) designed to comprehensively evaluate 682 disease categories across 203 agricultural crop types for 7 vision-language question-answer tasks. In the era of larger-scale vision-language models (VLMs), our AgroBench is obviously non-trivial in terms of many more crop and disease categories with all expert annotations for establishing QA benchmarks in the agricultural domain. + +Question: What is one of the main benefits the process shown in the image? + +Answer: Weed control and insect pest reduction. + +# Machine Usage + +![](images/0c540aff0fbaec8558549d07997ba59c11bbca1881dfe4d4199bcf1c0eb4e12a.jpg) + +Question: What is the main purpose of using the machine shown in the image? + +Answer: To control weeds in the field without relying on chemical herbicides. + +![](images/76ed41d7572652f9f13fd383e21a08c29587282abcb0aff33e1fda6e1640a6c9.jpg) + +# Crop categories: 203 + +![](images/2d1008e2467a3488f108c5c190325a4d5658c7e1519619c3eeab25ce486430af.jpg) + +# Disears categories: 684 + +![](images/4fea14e2638ce0c66458baa1b6610fb1e1b908b1b496f258f298e950b1bfc08f.jpg) + +# Expert annotation + +# Abstract + +Precise automated understanding of agricultural tasks such as disease identification is essential for sustainable crop production. Recent advances in vision-language models (VLMs) are expected to further expand the range of agricultural tasks by facilitating human-model interaction through easy, text-based communication. Here, we introduce AgroBench (Agronomist AI Benchmark), a benchmark for evaluating VLM models across seven agricultural topics, covering key areas in agricultural engineering and relevant to real-world farming. Unlike recent agricultural VLM benchmarks, AgroBench is annotated by expert agronomists. Our AgroBench covers a state-of-the-art range of categories, including 203 crop categories and 682 disease categories, to thoroughly evaluate VLM capabilities. In our evaluation on AgroBench, we reveal that + +VLMs have room for improvement in fine-grained identification tasks. Notably, in weed identification, most opensource VLMs perform close to random. With our wide range of topics and expert-annotated categories, we analyze the types of errors made by VLMs and suggest potential pathways for future VLM development. Our dataset and code are available at https://dahlian00.github.io/AgroBenchPage/. + +# 1. Introduction + +Agriculture is a fundamental process for humans to produce crops to live and stay healthy. With the development of computer vision technology, effective and automated management of external crop factors such as diseases and pests has been explored, contributing to stable crop + +
DatasetAnnot.CropWeedDiseasePestImagesQA pairsMain Purpose
Agri-LLaVA [48]GPT-429-109112391k391k (Synthetic)Training
AgroInstruct [4]GPT-41744741270k70k (Synthetic)Training
CDDM [20]GPT-415-60-137k1M (Synthetic)Training
AgroBench (Ours)Expert2031086821343,7454,342 (Expert)Evaluation
+ +Table 1. Comparison of agricultural vision datasets. Expert refers to 'human expert' in this context. AgroBench provides a comprehensive evaluation framework for assessing VLMs, featuring multiple tasks and a wide range of categories. + +production. This includes detecting and classifying undesirable conditions like diseases [1, 10, 13, 34, 38, 41, 47] and pests [3, 49, 53], as well as general plant management tasks such as crop classification [52, 54] and recognition of crop maturity [22, 25, 37, 39], and structure understanding [21, 43]. + +To maintain stable crop production, it is important to recognize a wide range of crop conditions, including undesirable situations such as diseases and pests, and to know how to respond appropriately. A single visual model that can handle various conditions, not only disease and pest detection but also treatment and general crop management, would be highly beneficial. However, most existing approaches use task-specific models. These models usually require large amounts of training images and manual annotations for each task. As a result, farmers often need to use several different models depending on the situation. This increases complexity and makes the overall system less accessible for practical use in agriculture. + +For general purpose visual tasks, vision-language models (VLMs) [17, 19, 32, 33, 45, 46, 50] have become widespread recently because they can understand task definitions provided by natural language prompts, covering a wide range of applications without the need for task-specific model training. The recognition ability of VLMs is closely connected to image recognition itself and linked to an object's words, and supports open-vocabulary recognition through web-scale training. Here, zero-shot recognition and few-shot adaptation have also been realized by VLMs with language representations. This ability opens up a wide range of applications; therefore, we believe this can be applied to agricultural scenarios as well. It is worth noting that VLMs offer an easy-to-use interface for the general public, especially the question-answer (QA) and conversation modes. To investigate the range of tasks that VLMs obtained through large-scale vision-language training can cover, recent studies have introduced various benchmark datasets to fully evaluate the VLMs, including tasks such as understanding diagrams and charts [26, 28], and question answering requiring specialized knowledge [7, 55, 56]. + +However, VLM research remains underexplored in agriculture due to a lack of benchmark datasets that include diverse tasks and categories in agriculture. Several pioneer- + +ing works [20, 48] have adapted open-source VLMs such as LLaVA [18] to the agricultural domain by fine-tuning them on synthetic datasets generated from closed-source VLMs such as GPT-4o. This is recognized as an effective approach because black-box VLMs often possess a certain amount of agricultural knowledge obtained from the Internet. While this helps in generating responses that are generally acceptable to experts, there is almost no way to verify whether the answers are truly correct. Moreover, in the limited evaluation of categories, we are insufficient for fully assessing VLM knowledge; whether they can answer a wide range of types, such as diseases, pests, and weeds. These limitations motivated us to develop a benchmark dataset in order to evaluate the VLM's broad knowledge in the agricultural domain and its applicability as a practical application. + +Here, we introduce AgroBench (Agronomist AI Benchmark), a comprehensive benchmark dataset for VLM for the agricultural domain, covering a state-of-the-art range of categories for agriculture-focused benchmark datasets for VLM; 682 disease, 134 pest, 108 weed, and 203 crop categories (Figure 1). We cover not only identification tasks but also crop production and disease management knowledge. Moreover, we include other important topics related to crop management with 98 machine categories and 77 traditional management methods to investigate more about VLM's ability. We carefully selected benchmark tasks from key agricultural engineering research areas, and also the tasks that address challenges faced by farmers in real agricultural scenarios. We employ a multiple-choice format, and all questions are annotated by human agronomist experts, which overcomes the limitations of previous synthetically created datasets. We carefully selected images from Creative Commons-licensed and publicly available datasets, including real farm conditions. Also, under our manual annotation process, unclear images were removed. As shown in Table 1, AgroBench contains data with the most diverse disease and crop variation categories and many more crop/weed annotations. + +Our main contributions are as follows; + +- We developed AgroBench, a benchmark dataset for VLM, to assess their broad agricultural knowledge of VLMs and evaluate their applicability in practical applications. + +- In our AgroBench evaluation, VLMs tend to achieve high performance in disease and crop management tasks; however, there is still room for improvement in weed and disease identification. +- Our AgroBench enables error analysis by providing broader category annotations, highlighting future directions for model training focus areas. + +# 2. Related Work + +# 2.1. Computer Vision for Agriculture + +With the rise of computer vision techniques driven by advancements in deep learning, a wide range of agricultural tasks have been explored over the past decade. In particular, dataset construction and benchmarking play an important role in developing models and practical applications in agriculture, while also paving the way for forward-looking advancements. Disease identification is a key research focus across various crops, such as rice [1, 34], tomato [10], cacao [13], and sugarcane [47]. Multicrop datasets have also been created [29, 41]. PlantDoc [41] covers 13 species and 17 classes focused on leaf-based diseases. Plant Village [29] offers 39 classes containing both diseased and healthy leaf categories. In addition to crop disease identification, pest identification [49, 53], weed identification [12, 31, 42] have also been studied. + +While most of the agricultural datasets primarily focus on visual data, few computer vision studies in agriculture investigate the potential of multi-modal approaches. The PlantWild dataset [52] includes 56 plant-disease class pairs, which are collected through image search engines. They implement a CLIP-based model and show the possibility of training with combined text and image data. In an instruction-based format, the CDDM [20] has created 16 categories of crops and 60 categories of crop diseases dataset, which generates instructional data using GPT. While these synthetically created datasets explore the agriculture multi-modal models, there remains a lack of datasets for comprehensive multi-modal model evaluation, validated by human experts and covering a wide range of tasks and categories. + +# 2.2. Vision-Language Models + +Models. Visual models in the computer vision field have been accelerated by language modality with the data resource of sentence-level inputs and web-scale texts. Specifically, CLIP [35] has made a significant contribution in this context, by text and image feature alignment through contrastive learning. CLIP played a key role in introducing more sophisticated visual representation into language explanations e.g., Flamingo [2], BLIP [14, 15], Qwen [5], PaLI [8], LLaVA [17, 18], CogVLM [50], and Emu [44, 45, 51]. Closed-source VLMs achieve state-of-the-art per + +formance, such as GPT-4o [32] and Gemini Pro [46], which are said to acquire human-level knowledge across diverse fields on the Internet. + +**Benchmarks.** Along this line, several recent studies introduced benchmark datasets. Especially, the representative examples in terms of universal knowledge benchmarking include MMMU [55], MMMU-Pro [56] and MMStar [6] for multi-modal understanding. These benchmarks contain highly diverse domains (e.g., natural, graph, illustration, medical images) and academic fields (e.g., science, engineering, art, and medical fields) under the tasks of vision language such as question-answering and reasoning. The series of MMMUs have been verified with foundation models like GPT-4V, but they cannot answer the questions perfectly. More specific domain datasets have also been proposed, such as those for Medicine [27, 36], Chart [26, 28, 40], and Video understandings [9, 16, 30], contributing to the accurate evaluation of VLMs and guiding future research directions. + +# 3. AgroBench + +This section introduces AgroBench, the first comprehensive benchmark dataset to evaluate VLM models from the perspective of agricultural vision tasks. The benchmark consists of seven tasks covering a wide range of tasks selected from key agricultural engineering research areas, as well as tasks that address real-world challenges faced by farmers in real agriculture scenarios. AgroBench includes 682 disease categories, 134 pest categories, 203 crop categories, and 108 weed categories, representing the largest number of categories in each area to date, to the best of our knowledge. + +# 3.1. Benchmark Tasks in Agricultural Scenes + +The seven benchmark tasks encompass key research areas in agricultural engineering as well as real-world challenges faced by farmers. To facilitate VLM evaluation, we also provide prompts to address each task in a question-answer format. The details of each task are described below. + +1) Disease Identification (DID). The DID task aims to accurately diagnose and classify crop diseases. This is a key task in agriculture to protect crop health and maximize yields. Our benchmark provides 1,502 QA pairs that cover 370 disease categories, 160 crop categories and 682 crop-disease combinations. To thoroughly evaluate the capabilities of VLMs, we include four misleading disease labels for each image, featuring diseases with similar symptoms or common diseases affecting the target plant. VLMs are required to diagnose the disease based on the image, considering both symptoms and the crop species. Figure 2a illustrates example images. + +2) Pest Identification (PID). The PID task aims to identify pests to prevent infestations that can severely impact crop health. Accurate identification reduces economic losses and + +![](images/3d4a2bf5549330284c94ef35053d17fd9ed5e6a77d73d9be524f7b58ef53b0c9.jpg) +(a) Disease identification (DID) + +![](images/de6f8bc38f4fd53f589354903603a699929451185ee9edf53ae4a78cf7854319.jpg) +Figure 2. Examples of labeled images for DID, PID, and WID tasks. Our dataset includes 682 crop-disease pairs, 134 pest categories, and 108 weed categories. We prioritized collecting images from real farm settings. +Figure 3. Seven benchmark tasks in AgroBench. AgroBench includes multiple topics with a diverse range of categories. The total accuracy is calculated by the average of each task to mitigate the difference in QAs. + +minimizes harm to the environment. Our benchmark provides 544 labeled images that cover 134 pest categories, including insects, mites, and other organisms harmful to plants. For categories where we obtain multiple images, we select multiple insect growth stages whenever possible. To fully evaluate the VLMs, we assign alternative choices that closely resemble or are commonly associated with the target + +crops. Figure 2b shows examples of labeled pest images. + +3) Weed Identification (WID). The WID task aims to identify weed species. Our benchmark includes 609 images of weeds with ground-truth bounding boxes, covering 108 weed species commonly found in farm fields. We assign bounding boxes because multiple weed types often grow closely together, and we want to clarify which one is the target. Specifically, VLMs are required to identify the weed species within the provided bounding box on the image. Figure 2c shows examples of labeled weed images. +4) Crop Management (CMN). Crop management focuses on optimizing farming practices to facilitate crop growth. This involves making decisions on irrigation, fertilization, planting times, and other cultivation practices. Our benchmark provides 411 question-answer pairs for this task. VLMs are required to analyze images of crops and recommend appropriate management strategies by considering factors such as crop health, growth stage, and environmental conditions visible in the images given five answer candidates. In Figure 4a and b, we show two examples of the harvest timing for white and green asparagus, respectively. Our dataset includes complex questions that take into account the differences in their harvest timing. +5) Disease Management (DMN). Disease management aims to control and reduce diseases in crops. This involves informed decisions on interventions such as applying pesticides, adopting resistant crop varieties, or modifying cul + +![](images/4f30009ad05735b9fef67432f58689b7d107f17aac1a3937140a67ac61fae5b7.jpg) + +![](images/076cf1960b742c694f51e06de401e57dcecbd814ee68a1fd2330c0958c7bff6e.jpg) + +![](images/22f50d4ecc4936e5fbcc49448c773614189727479f39b94e29ed005037bc6335.jpg) + +![](images/6f25c0cf5da35a4823f8139524b3df03fcc6b1bf6ee618a981dbc71c8454bf43.jpg) + +![](images/6d00e18010cf57c343a3fd5b562a646a52fa3cad0e097a6a3c842e715d6267af.jpg) +Figure 4. Examples of QA pairs for CMN and DMN tasks. (a) and (b) Crop management QA types for white asparagus and asparagus, respectively. Their difference in the harvest timing affects the answer's difference correctly. (c) and (d) Disease management QA types for the alfalfa bacterial leaf spot with the initial and severe symptoms, respectively. Based on the severity of the symptoms, the annotator changes the answer. + +![](images/881a3aff989b5b9ba5381602e7781cc9ab3ba2987b72af4d439156ec73c795cc.jpg) + +![](images/641deafee5d7f9a560c0bb7676fa3590c8097bd4dbcdd0f33673b761e9db8c78.jpg) + +![](images/8a729418308079261a220c19eb3ee2ae5a75dd4629d0c7b3e4df566fe914fc51.jpg) + +tivation practices. Our benchmark provides 569 question-answer pairs covering 141 crop-disease combination categories. Since many diseases share the same management strategies regardless of the crop, we carefully select a diverse range of disease types and management strategies. VLMs are required not only to identify the disease but also to recommend appropriate management strategies based on images of affected crops. In Figure 4c and d show example QAs from the Disease Management tasks. Both images depict Bacterial leaf spot on alfalfa, with c showing the initial stage of the disease and d showing the severe stage. Based on the severity of the disease, we provide different answer options and set different correct answers. The input prompt consists of a question and five answer candidates. +6) Machine Usage QA (MQA). Machine usage QA addresses the correct use and choice of agricultural machinery depending on the task and farming conditions. Selecting the appropriate machinery is essential for efficient farming. Our benchmark provides 303 question-answer pairs covering 98 machine categories. Given that many crops share the same machinery (e.g., soil preparation and irrigation), the coverage of this category is comprehensive. VLMs are required to answer questions about machinery operation or select the appropriate machine for a given scenario based on images of machinery or farming conditions. +7) Traditional Management (TM). Traditional management methods involve natural and sustainable approaches to farming, such as the use of organic fertilizers, terrace farming, and agroforestry. Recent computer vision studies have not focused on these traditional practices, although many are still used by certain local farmers. Our benchmark includes 404 question-answer pairs, including 77 traditional management practices. VLMs are required to identify the management method or explain its effectiveness, given five answer choices. + +# 3.2. Dataset Construction + +Image Selection. To establish a high-quality benchmark covering a wide range of crops, diseases, and pest categories we initially curated around 50,000 agricultural images from websites supervised by plant pathologists, either where redistribution was permitted or where we obtained redistribution permission. We selected images in real farm settings as much as possible, as shown in Figure 2, to evaluate real-world scenarios. When licensed images for a target category were limited, we used laboratory setting images. For curation, we obtained images along with their corresponding labels. The annotator is one of the authors, who holds a Ph.D. in Agriculture. In the human evaluation conducted during the experiment, other individuals with a Ph.D. or M.S. degree in Agriculture reviewed questions to assess the quality of the QA pairs. The annotator selected images that clearly represented target labels and removed those images with irrelevant labels. For the Machine usage QA and Traditional management, annotators manually created categories based on textbooks and websites, and searched for corresponding images with redistribution licenses. For Weed identification, we use existing dataset [11, 23, 24, 42]. For this Weed identification category, we will provide a simple code to download the data and crop images and assign bounding boxes for Weed Identification, allowing users to work with the existing dataset without redistributing the images. Through these image selection processes, the selected images are high-quality 4,218 representative images. The detailed distribution of tasks and categories for each task is summarized in Figure 3. + +QA Annotation. For the DMN, CMN, MQA, and TM tasks, all question-answer pairs were created manually, independent of any LLMs or VLMs. Annotators used GPT only for sentence rephrasing, but they were prohibited from using any knowledge from GPT for QA creation. Annotators were allowed to refer to textbooks, academic journals, and other authoritative sources in the field to ensure the accuracy and depth of the dataset. This took around 150 man + +
ModelDIDDMNPIDWIDCMNMQATMOverall (all)Overall (subset)
Random Choice21.7715.6420.4017.9016.0622.1119.3119.0319.11
Human25.0022.5045.0020.0036.2557.5051.25-36.79
Closed-Source Vision Language Models (VLMs)
GPT-4o mini [33]53.6080.6760.0435.1464.2370.9669.8062.0669.65
GPT-4o [32]64.1887.3577.7644.1775.4382.8482.4373.4579.26
Gemini1.5-Flash [46]55.0679.9670.0450.9064.7278.2273.2767.4568.82
Gemini1.5-Pro [46]62.9281.5574.4555.1771.0582.8477.7272.2469.74
Open-Source Vision Language Models (VLMs)
EMU2Chat [44]42.0148.3343.7523.8140.3937.6247.7740.5333.84
LLaVA-Next-8B [19]45.4772.5843.0130.0554.2656.1157.4651.2857.84
LLaVA-Next-72B [19]54.9580.0049.8126.9866.9266.1170.3859.3164.36
QwenVLM-7B [5]51.2680.4963.9733.1766.4276.2477.4864.1566.41
QwenVLM-72B [5]57.9987.8773.3534.4875.9180.8684.1670.6672.45
CogVLM-19B [50]29.1653.7852.3925.4554.0171.6266.0950.3644.27
LLaVa-7B [18]36.0262.7438.7924.7953.7746.5355.2045.4146.14
LLaVA-13B [18]40.2168.8944.4924.7959.3754.1358.4250.0455.31
+ +Table 2. Results for the seven benchmark tasks with images. We provide results for Random Choice, Human Validation, four closed-source VLMs, and open-source VLMs. Human validation was conducted by 28 people on a subset of 80 samples per task as a reference. + +
DIDDMNPIDWIDCMNMQATMOverall
GPT-4o [32]1.9372.5818.751.0040.3925.0848.2729.71
LLaVA-Next-8B [19]26.1070.3021.8819.7053.7730.3640.3537.49
+ +Table 3. Results for the seven benchmark tasks with text only. We additionally evaluate the model without image inputs, and the overall performance is close to random. + +hours. We carefully annotate various types of questions. Examples of QAs that require expert knowledge are shown in Figure 4. All questions were carefully created so that image reference is necessary for answering. + +Dataset Statistics. Following the dataset annotations and selections, our AgroBench comprises seven tasks with 4,342 QA pairs, as shown in Table 1. All tasks comprise a wide range of categories for detailed evaluation. Please refer to the supplementary materials for more details and examples. + +# 4. Experiments + +# 4.1. Experimental Settings + +Baseline Models. We use four closed-source models: GPT-4o [32], GPT-4o mini [33], Gemini1.5-Pro [46], and Gemini1.5-Flash [46]. GPT-4o mini and Gemini1.5-Flash are down-scale versions of GPT-4o and Gemini1.5-Pro, respectively. We use eight open-source models: EMU2Chat [44], LLaVA-Next-8B [19], LLaVA-Next-72B [19], QwenVLM-7B [5], QwenVLM-72B [5], CogVLM-19B [50], LLaVa-7B [18], and LLaVa-13B [18]. For the details of these models, please refer to the supple + +mentary material. + +Human Results. We also present results from human participants for reference. We surveyed 28 students, each holding at least a bachelor's degree in agriculture, and asked them to answer 20 questions each. This created a test subset of 280 questions, with each question answered by two participants, resulting in a total of 560 responses. We averaged the results per task and reported the accuracy. Each participant was permitted to use a book or translator to look up word meanings but was prohibited from using the internet for searches. If participants were unsure of the answer, they were asked to provide the response they believed to be most accurate. + +Evaluation Protocol. Importantly, our dataset evaluation is conducted per task, and overall scores are averaged based on the number of tasks, not the number of QAs. This prevents categories with a large number of evaluations from becoming dominant. We adopt an exact matching approach for our five-option questions. If the model's response matches the option's letter or the answer sentence, we consider it correct. If the model's answer does not match any option, including cases where there is no answer or mul + +![](images/2fab5047c0c1479fdb57a3923aafe09d26fc86801887fd7a570cc75d4c1e0152.jpg) +Figure 5. Results of seven benchmark tasks with Chain of Thought (CoT). Baseline indicates results without CoT. In the one-shot, two-shot, and three-shot settings, we provide one, two, and three CoT examples per task, respectively, to guide the model. + +![](images/a1d88734ca867c9fda3e9719d9476adcba2160f86d85b76f5a980c6b9e4fdd7b.jpg) +Figure 6. Error analysis on seven benchmark tasks with GPT-4o. We extract a maximum of 15 errors per task from the zero-shot CoT result. We manually analyze how they conclude the incorrect answer. + +tiple answers, we consider it incorrect. + +# 4.2. Main Results + +Here, we discuss the main results of our AgroBench evaluation performance. We evaluate our AgroBench using eight open-source VLMs and four closed-source VLMs with APIs. + +Challenges of the AgroBench. We show the main results in Table 2. The most difficult task is Weed Identification (WID), on which most open-source VLMs perform at around the random score. The highest WID accuracy, $55.17\%$ , was achieved by Gemini 1.5-Pro. This suggests that the knowledge about weeds is not as fully trained as that of crops. All models' Disease Identification (DID) results are lower than their Disease Management (DMN) results. This means that VLM models can gain contextual information, but there is still room for perceptual improvement. + +Model Comparison. Overall, closed-source VLMs achieve better results than open-source VLMs and achieve higher performance than humans. GPT-4o model achieved the highest score in overall performance. Among the open-source models, QwenVLM-72B achieves the best result in overall accuracy, which is comparable or even superior to open-source VLMs in some tasks. QwenVLM + +72B achieves satisfactory results on both identification and question-answering benchmarks on AgroBench. + +# 4.3. Ablations + +Context Reliance. We further evaluate the results on AgroBench using text-only input to determine whether visual information is necessary to answer the questions. Table 3 presents the experimental results across seven benchmark tasks on AgroBench. With text-only input, performance drops across all models, confirming the reliance on visual information. However, for both models, the DMN, CMN, and TM tasks maintain significantly higher accuracy than random selection. Although we ensured that our questions do not include disease names or appearance-related traits, models tend to infer answers based on estimation. This suggests that many disease types share common management strategies, such as avoiding humidity, preferring cooler temperatures, or pruning infected parts, allowing models to predict the most likely option. A similar pattern is observed in the CMN and TM tasks, where models can make educated guesses based on contextual cues. (See the supplementary material for further detailed examples.) + +Chain of Thought. We evaluate the effectiveness of Chain of Thought (CoT) reasoning. In the one-shot setting, we + +# (a) Lack of Knowledge + +Q: What disease is affecting the plant? + +![](images/c0e92e8e14c4de01ed1ee34a40c0381608f2be26e92872a727e1460dc74d5e56.jpg) + +GPT-4o: ..., **Sclerotinia blight**: Characterized by white, cottony mycelium and black sclerotia on stems. The image does not show these symptoms. (not enough knowledge) ..., the plant is likely affected by **D. Cylindrocladium black rot**: (This is more common on root) + +A: Sclerotinia blight + +# (b) Perceptual Error + +Q: What activity is being performed using the tractor-pulled machine shown in the image? + +![](images/32076a514abab6329414e314b00c8e580b53bda312b3345459d77b27f5496820.jpg) +Figure 7. Error examples of GPT-4o. Examples illustrate the two main error types: Lack of Knowledge and Perceptual Error. + +GPT-40:The machine seems to be designed to lift plants from the soil, ..., and there are plants being lifted from the ground (There are no harvested ctops)....the activity being performed is: D. Harvesting potato tubers from the soil using a tractor-pulled harvesting machine. + +A: Cutting and chopping the stems and leaves using a tractor-pulled leaf chopper. + +provide a single CoT reasoning example along with an image and the corresponding prompt for each task to guide the model in answering the question. We select 100 QAs per task for evaluation and use GPT-4 as the VLM. Figure 5 presents the accuracy of CoT reasoning with the baseline results. While CoT achieves slightly higher accuracy, the improvement is not significant compared to the baseline. However, CoT demonstrates effectiveness in certain tasks, such as PID, WID, CMN, and TM. For instance, in PID, the VLM performs step-by-step and careful reasoning (see the supplementary material for more detailed examples). In WID, the most challenging task, CoT provides useful examples that may help the model make predictions. Overall, CoT contributes to accuracy, but we observe performance saturation in the three-shot setting. For further detailed examples, please refer to the supplementary material. + +# 4.4. Error Analysis + +Here, we analyze the mistake types that occur depending on the agricultural tasks. We bring out up to 15 failure examples per task from the zero-shot CoT results and manually analyze how they reached the incorrect answers. + +Lack of Knowledge (51.92%). This includes cases where VLM can't accurately describe the appearance or relevant knowledge of a choice (e.g., VLM fails to describe disease symptoms or insect characteristics) or lacks context (e.g., VLM doesn't know how to treat diseased crops or manage crops for high yield). Figure 7a shows an example of a Lack of Knowledge case. When the VLM analyzes the correct answer option for Sclerotinia blight, it fails to describe the + +symptoms of discoloration and wilting. Additionally, its incorrect answer choice is more commonly associated with the root rather than the stem. These errors are based on a Lack of Knowledge, suggesting VLMs need more detailed categories and domain-specific training. + +Perceptual Error (32.69%). This indicates that VLM can't pay attention or recognize the answer-related part in the image (e.g., can't recognize the green insect on the leaf), and VLM misunderstands the image, leading to incorrect answers. Figure 7b shows an example of the model hallucinating and misunderstanding the situation. First, the VLM incorrectly identifies the machine as one used for lifting plants from the soil. Then, it describes the scene as if there are harvested crops, even though no harvested crops are present. These errors can be mitigated by enhancing the VLM's perception abilities for domain-specific classes. Additionally, improving general perception capabilities, including reducing hallucinations, can further contribute to VLM performance. + +Reasoning Error (7.6%). Reasoning error involves the VLM can describe the options correctly, but can't compare them step by step and conclude the wrong answer. This error is relatively low compared to the existing work [55] since AgroBench requires more specific knowledge and doesn't include reasoning relying on problems (e.g., math). + +Other Errors (7.79%). For the other errors, we observe Shortcut Error (The VLM can pick up the two candidate options correctly but conclude the answer without comparing the candidates), Double Answer Error (Concluding two answers are correct), Interpretation Misunderstanding (VLM misleading the question and conclude wrong answer), and Reject to Answer (VLM conclude there is no answer). + +# 5. Conclusion + +In this paper, we develop AgroBench, a comprehensive benchmark dataset for VLMs in the agricultural domain, covering a state-of-the-art range of categories. Our dataset comprises seven benchmark tasks encompassing key research areas in agricultural engineering as well as real-world challenges faced by farmers. AgroBench contributes to agricultural VLM research by addressing the lack of datasets for comprehensive multi-modal model evaluation, validated by human experts. + +In our evaluation, VLMs exhibit strengths across different tasks. However, in several tasks such as weed identification and disease identification, all models show room for improvement. Our error analysis reveals that most failures are due to a lack of knowledge (51.92%), suggesting that VLMs require more specialized agricultural knowledge. Our dataset will facilitate agricultural VLM research, enabling broad category and task evaluation to support sustainable, automated agriculture. + +# 6. Acknowledgment + +This work was supported by the AIST KAKUSEI Project (FY2024) and JST FOREST Grant Number JPMJFR206F. We would like to thank the Agricultural Administration Division, Department of Agriculture, Hokkaido Government, for providing some of the images. We are also grateful to Daniel Steininger for contributing images from the CropAndWeed dataset to our dataset. We used ABCI 3.0 provided by AIST and AIST Solutions. + +# References + +[1] Petchiammal A, Briskline Kiruba S, Murugan D, and Pandarasamy Arjunan. Paddy doctor: A visual image dataset for automated paddy disease classification and benchmarking. IEEE Dataport, 2022. 2, 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3 +[3] Abderraouf Amrani, Ferdous Sohel, Dean Diepeveen, David Murray, and Michael G.K. Jones. Deep learning-based detection of aphid colonies on plants from a reconstructed brassica image dataset. Computers and Electronics in Agriculture, 205:107587, 2023. 2 +[4] Muhammad Awais, Ali Husain Salem Abdulla Alharthi, Amandeep Kumar, Hisham Cholakkal, and Rao Muhammad Anwer. Agropt: Efficient agricultural vision-language model with expert tuning. arXiv preprint arXiv:2410.08405, 2024. 2 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3, 6 +[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3 +[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, + +Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2 +[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Alexander Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. 2023. 3 +[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees G. M. Snoek, and Yuki M. Asano. Tvbench: Redesigning video-language evaluation. 2024. 3 +[10] Mamta Gehlot, Rakesh Saxena, and Geeta Gandhi. "tomato-village": a dataset for end-to-end tomato disease detection in a real-world environment. Multimedia Systems, 29:1-24, 2023. 2, 3 +[11] Benedikt Geisler. Perrenial plants detection, 2021. 5 +[12] Sebastian Haug and Jörn Ostermann. A crop/weed field image dataset for the evaluation of computer vision based precision agriculture tasks. In Proc. European Conference on Computer Vision (ECCV), pages 105-116. Springer, 2014. 3 +[13] Atuhurra Jesse, N'guessan Yves-Roland Douha, and Pabitra Lenka. Image classification for cssvd detection in cacao plants. arXiv preprint 2405.04535, 2024. 2, 3 +[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742. PMLR, 2023. 3 +[16] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark. In CVPR, 2024. 3 +[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 3 +[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 6 +[19] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.2,6 +[20] Xiang Liu, Zhaoxiang Liu, Huan Hu, Zezhou Chen, Kohou Wang, Kai Wang, and Shiguo Lian. A multimodal benchmark dataset and model for crop disease diagnosis. In Proc. European Conference on Computer Vision (ECCV), 2024. 2, 3 +[21] Xinpeng Liu, Kanyu Xu, Risa Shinoda, Hiroaki Santo, and Fumio Okura. Masks-to-skeleton: Multi-view mask-based + +tree skeleton extraction with 3d gaussian splatting. Sensors, 25(14), 2025. 2 +[22] Shenglian Lu, Wenkang Chen, Xin Zhang, and Manoj Karkee. Canopy-attention-yolov4-based immature/mature apple fruit detection on dense-foliage tree architectures for early crop load estimation. Computers and Electronics in Agriculture, 193:106696, 2022. 2 +[23] Yuzhen Lu. Cottonweeddet3. kaggle, 2022. 5 +[24] Simon Leminen Madsen, Solvejg Kopp Mathiassen, Mads Dyrmann, Morten Stigaard Laursen, Laura-Carlota Paz, and Rasmus Nyholm Jorgensen. Open Plant Phenotype Database of Common Weeds in Denmark, 2020. 5 +[25] Abdul Khalique Maitlo, Abdul Aziz, Hassnian Raza, and Neelam Abbas. A novel dataset of guava fruit for grading and classification. Data in Brief, 49:109462, 2023. 2 +[26] Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, Dublin, Ireland, 2022. Association for Computational Linguistics. 2, 3 +[27] João Matos, Shan Chen, Siena Placino, Yingya Li, Juan Carlos Climent Pardo, Daphna Idan, Takeshi Tohyama, David Restrepo, Luis F. Nakayama, Jose M. M. Pascual-Leone, Guergana Savova, Hugo Aerts, Leo A. Celi, A. Ian Wong, Danielle S. Bitterman, and Jack Gallifant. Worldmedqa-v: a multilingual, multimodal medical examination dataset for multimodal language models evaluation, 2024. 3 +[28] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1516-1525, 2020. 2, 3 +[29] Sharada P. Mohanty, David P. Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in Plant Science, 7, 2016. 3 +[30] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 3 +[31] Alex Olsen, Dmitry A Konovalov, Bronson Philippa, Peter Ridd, Jake C Wood, Jamie Johns, Wesley Banks, Benjamin Girgenti, Owen Kenny, James Whinney, et al. Deepweeds: A multiclass weed species image dataset for deep learning. Scientific reports, 9(1):2058, 2019. 3 +[32] OpenAI. Gpt-4o, 2024. Accessed: 2024-11-13. 2, 3, 6 +[33] OpenAI. Gpt-4o mini, 2024. Accessed: 2024-11-13. 2, 6 +[34] Chiranjit Pal, Imon Mukherjee, Sanjay Chatterji, Sanjoy Pratihar, Pabitra Mitra, and Partha Pratim Chakrabarti. Indian rice disease dataset (irdd), 2023. 2, 3 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[36] Corentin Royer, Bjoern Menze, and Anjany Sekuboyina. Multimedieval: A benchmark and a toolkit for evaluating medical vision-language models, 2024. 3 + +[37] Rahman Sanya, Ann Lisa Nabiryo, Jeremy Francis Tusubira, Sudi Murindanyi, Andrew Katumba, and Joyce Nakatumba-Nabende. Coffee and cashew nut dataset: A dataset for detection, classification, and yield estimation for machine learning applications. Data in Brief, 52:109952, 2024. 2 +[38] Nabila Husna Shabrina, Siwi Indarti, Rina Maharani, Dinar Ajeng Kristiyanti, Irmawati, Niki Prastomo, and Tika Adilah M. A novel dataset of potato leaf disease in uncontrolled environment. Data in Brief, 52:109955, 2024. 2 +[39] Risa Shinoda, Hirokatsu Kataoka, Kensho Hara, and Ryozo Noguchi. Transformer-based ripeness segmentation for tomatoes. Smart Agricultural Technology, 4:100196, 2023. 2 +[40] Risa Shinoda, Kuniaki Saito, Shohei Tanaka, Tosho Hirasawa, and Yoshitaka Ushiku. Sbs figures: Pre-training figure qa from stage-by-stage synthesized images, 2024. 3 +[41] Davinder Singh, Naman Jain, Pranjali Jain, Pratik Kayal, Sudhakar Kumawat, and Nipun Batra. Plantdoc: A dataset for visual plant disease detection. In Proc. ACM IKDD CoDS and COMAD, page 249-253, 2020. 2, 3 +[42] Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, and Verena Widhalm. The cropandweed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In WACV, pages 3729-3738, 2023. 3, 5 +[43] Daniel Steininger, Julia Simon, Andreas Trondl, and Markus Murschitz. Timbervision: A multi-task dataset and framework for log-component segmentation and tracking in autonomous forestry operations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2 +[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 3, 6 +[45] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 2, 3 +[46] Google Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024. 2, 3, 6 +[47] Sandip Thite, Yogesh Suryawanshi, Kailas Patil, and Prawit Chumchu. Sugarcane leaf dataset: A dataset for disease detection and classification for machine learning applications. Data in Brief, 53:110268, 2024. 2, 3 +[48] Liqiong Wang, Teng Jin, Jinyu Yang, Ales Leonardis, Fangyi Wang, and Feng Zheng. Agri-llava: Knowledge-infused large multimodal assistant on agricultural pests and diseases. arXiv preprint arXiv:2412.02158, 2024. 2 +[49] Rujing Wang, Liu Liu, Chengjun Xie, Po Yang, Rui Li, and Man Zhou. Agripest: A large-scale domain-specific benchmark dataset for practical agricultural pest detection in the wild. Sensors, 21(5), 2021. 2, 3 +[50] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming + +Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. In NeurIPS, 2024. 2, 3, 6 +[51] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 +[52] Tianqi Wei, Zhi Chen, Zi Huang, and Xin Yu. Benchmarking in-the-wild multimodal disease recognition and a versatile baseline. In Proc. ACM International Conference on Multimedia (ACMMM), 2024. 2, 3 +[53] Xiaoping Wu, Chi Zhan, Yukun Lai, Ming-Ming Cheng, and Jufeng Yang. Ip102: A large-scale benchmark dataset for insect pest recognition. In CVPR, pages 8787-8796, 2019. 2, 3 +[54] Momchil Yordanov, Raphaël d'Andrimont, Laura Martinez-Sanchez, Guido Lemoine, Dominique Fasbender, and Marijn van der Velde. Crop identification using deep learning on lucas crop cover photos. Sensors, 23(14), 2023. 2 +[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 2, 3, 8 +[56] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 2, 3 \ No newline at end of file diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/images.zip b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..670b45f46b66979d24669dfb3a2b30fee809ac63 --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8949df8191dff455dd836e4de12f3950c6c549b2581620a712626dc0ef8dbc85 +size 831511 diff --git a/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/layout.json b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..435945a6055bb7401c51a7c95e714b78d27a078e --- /dev/null +++ b/2025/AgroBench_ Vision-Language Model Benchmark in Agriculture/layout.json @@ -0,0 +1,7980 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "text", + "content": "AgroBench: Vision-Language Model Benchmark in Agriculture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "spans": [ + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "content": "Risa Shinoda" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "content": ", Nakamasa Inoue" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "content": ", Hirokatsu Kataoka" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "inline_equation", + "content": "^{4,5}" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "content": ", Masaki Onishi" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "text", + "content": ", Yoshitaka Ushiku" + }, + { + "bbox": [ + 74, + 142, + 536, + 158 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "spans": [ + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "The University of Osaka " + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "Kyoto University " + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "Institute of Science Tokyo " + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "National Institute of Advanced Industrial Science and Technology (AIST) " + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "Visual Geometry Group, University of Oxford " + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 124, + 170, + 486, + 215 + ], + "type": "text", + "content": "OMRON SINIC X" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 241, + 165, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 241, + 165, + 253 + ], + "spans": [ + { + "bbox": [ + 64, + 241, + 165, + 253 + ], + "type": "text", + "content": "Disease Management" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 64, + 255, + 115, + 306 + ], + "blocks": [ + { + "bbox": [ + 64, + 255, + 115, + 306 + ], + "lines": [ + { + "bbox": [ + 64, + 255, + 115, + 306 + ], + "spans": [ + { + "bbox": [ + 64, + 255, + 115, + 306 + ], + "type": "image", + "image_path": "5701473a88fe9a3811a1637fcae31cbd6958bedd62476c53adf9de487a8a6479.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 256, + 223, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 256, + 223, + 280 + ], + "spans": [ + { + "bbox": [ + 118, + 256, + 223, + 280 + ], + "type": "text", + "content": "Question: What should be done to minimize the spread of the disease shown in the image?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 282, + 223, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 282, + 223, + 305 + ], + "spans": [ + { + "bbox": [ + 118, + 282, + 223, + 305 + ], + "type": "text", + "content": "Answer: Remove the male buds (bell) immediately using a forked stick." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 228, + 242, + 302, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 242, + 302, + 252 + ], + "spans": [ + { + "bbox": [ + 228, + 242, + 302, + 252 + ], + "type": "text", + "content": "Weed Detection" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 227, + 255, + 278, + 306 + ], + "blocks": [ + { + "bbox": [ + 227, + 255, + 278, + 306 + ], + "lines": [ + { + "bbox": [ + 227, + 255, + 278, + 306 + ], + "spans": [ + { + "bbox": [ + 227, + 255, + 278, + 306 + ], + "type": "image", + "image_path": "df59a31a6a22a927741e16ecb06aa63b545f796fee8dbece99e2200770492df4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 280, + 257, + 385, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 257, + 385, + 284 + ], + "spans": [ + { + "bbox": [ + 280, + 257, + 385, + 284 + ], + "type": "text", + "content": "Question: What species of weed does the highlighted bounding box represent in the image?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 281, + 288, + 374, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 288, + 374, + 305 + ], + "spans": [ + { + "bbox": [ + 281, + 288, + 374, + 305 + ], + "type": "text", + "content": "Answer: White goosefoot (Chenopodium album)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 390, + 242, + 488, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 242, + 488, + 253 + ], + "spans": [ + { + "bbox": [ + 390, + 242, + 488, + 253 + ], + "type": "text", + "content": "Disease Identification" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 389, + 255, + 440, + 307 + ], + "blocks": [ + { + "bbox": [ + 389, + 255, + 440, + 307 + ], + "lines": [ + { + "bbox": [ + 389, + 255, + 440, + 307 + ], + "spans": [ + { + "bbox": [ + 389, + 255, + 440, + 307 + ], + "type": "image", + "image_path": "261f7d4baca2a3ae8dc30f2278b9f5708a2b561e849d34ed807a15d5feb3b5e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 443, + 261, + 548, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 261, + 548, + 278 + ], + "spans": [ + { + "bbox": [ + 443, + 261, + 548, + 278 + ], + "type": "text", + "content": "Question: What disease is affecting the plant?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 443, + 293, + 523, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 293, + 523, + 300 + ], + "spans": [ + { + "bbox": [ + 443, + 293, + 523, + 300 + ], + "type": "text", + "content": "Answer: Bacterial canker." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 66, + 309, + 151, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 309, + 151, + 320 + ], + "spans": [ + { + "bbox": [ + 66, + 309, + 151, + 320 + ], + "type": "text", + "content": "Crop Management" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 64, + 322, + 115, + 374 + ], + "blocks": [ + { + "bbox": [ + 64, + 322, + 115, + 374 + ], + "lines": [ + { + "bbox": [ + 64, + 322, + 115, + 374 + ], + "spans": [ + { + "bbox": [ + 64, + 322, + 115, + 374 + ], + "type": "image", + "image_path": "c72cc416234804a15201207f62ef335610b36cdbe555bb2ac6be7f9999a9a69f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 118, + 324, + 223, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 324, + 223, + 346 + ], + "spans": [ + { + "bbox": [ + 118, + 324, + 223, + 346 + ], + "type": "text", + "content": "Question: Is the crop shown in the image ready for harvest based on its current condition?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 118, + 350, + 223, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 350, + 223, + 373 + ], + "spans": [ + { + "bbox": [ + 118, + 350, + 223, + 373 + ], + "type": "text", + "content": "Answer: Yes, the leaves and stems have turned yellow, and the seed pods are brown to black." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 230, + 327, + 383, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 327, + 383, + 350 + ], + "spans": [ + { + "bbox": [ + 230, + 327, + 383, + 350 + ], + "type": "text", + "content": "AGROBENCH" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 246, + 351, + 353, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 351, + 353, + 361 + ], + "spans": [ + { + "bbox": [ + 246, + 351, + 353, + 361 + ], + "type": "text", + "content": "Agronomist Al Benchmark" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 391, + 309, + 473, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 309, + 473, + 319 + ], + "spans": [ + { + "bbox": [ + 391, + 309, + 473, + 319 + ], + "type": "text", + "content": "Pest Identification" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 389, + 322, + 440, + 374 + ], + "blocks": [ + { + "bbox": [ + 389, + 322, + 440, + 374 + ], + "lines": [ + { + "bbox": [ + 389, + 322, + 440, + 374 + ], + "spans": [ + { + "bbox": [ + 389, + 322, + 440, + 374 + ], + "type": "image", + "image_path": "32e5cc5c8fcb0136d5eb28ece6926ea4c6e65e4a023f9cfa8b4fb63050430701.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 443, + 331, + 548, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 331, + 548, + 350 + ], + "spans": [ + { + "bbox": [ + 443, + 331, + 548, + 350 + ], + "type": "text", + "content": "Question: What pest is affecting the plant?" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 443, + 360, + 509, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 360, + 509, + 369 + ], + "spans": [ + { + "bbox": [ + 443, + 360, + 509, + 369 + ], + "type": "text", + "content": "Answer: Squash bug." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 66, + 376, + 156, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 376, + 156, + 387 + ], + "spans": [ + { + "bbox": [ + 66, + 376, + 156, + 387 + ], + "type": "text", + "content": "Traditional Methods" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 64, + 389, + 115, + 440 + ], + "blocks": [ + { + "bbox": [ + 64, + 389, + 115, + 440 + ], + "lines": [ + { + "bbox": [ + 64, + 389, + 115, + 440 + ], + "spans": [ + { + "bbox": [ + 64, + 389, + 115, + 440 + ], + "type": "image", + "image_path": "9d6aed769e25151ee011dcd88d0d57dfe8be179abeb1afb06b8e009d85285d07.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. We present AgroBench (Agronomist AI Benchmark) designed to comprehensively evaluate 682 disease categories across 203 agricultural crop types for 7 vision-language question-answer tasks. In the era of larger-scale vision-language models (VLMs), our AgroBench is obviously non-trivial in terms of many more crop and disease categories with all expert annotations for establishing QA benchmarks in the agricultural domain." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 118, + 392, + 223, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 392, + 223, + 419 + ], + "spans": [ + { + "bbox": [ + 118, + 392, + 223, + 419 + ], + "type": "text", + "content": "Question: What is one of the main benefits the process shown in the image?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 118, + 422, + 223, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 422, + 223, + 439 + ], + "spans": [ + { + "bbox": [ + 118, + 422, + 223, + 439 + ], + "type": "text", + "content": "Answer: Weed control and insect pest reduction." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 228, + 376, + 299, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 376, + 299, + 388 + ], + "spans": [ + { + "bbox": [ + 228, + 376, + 299, + 388 + ], + "type": "text", + "content": "Machine Usage" + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 227, + 392, + 277, + 441 + ], + "blocks": [ + { + "bbox": [ + 227, + 392, + 277, + 441 + ], + "lines": [ + { + "bbox": [ + 227, + 392, + 277, + 441 + ], + "spans": [ + { + "bbox": [ + 227, + 392, + 277, + 441 + ], + "type": "image", + "image_path": "0c540aff0fbaec8558549d07997ba59c11bbca1881dfe4d4199bcf1c0eb4e12a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 280, + 391, + 385, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 391, + 385, + 415 + ], + "spans": [ + { + "bbox": [ + 280, + 391, + 385, + 415 + ], + "type": "text", + "content": "Question: What is the main purpose of using the machine shown in the image?" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 280, + 417, + 386, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 417, + 386, + 440 + ], + "spans": [ + { + "bbox": [ + 280, + 417, + 386, + 440 + ], + "type": "text", + "content": "Answer: To control weeds in the field without relying on chemical herbicides." + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 395, + 378, + 410, + 392 + ], + "blocks": [ + { + "bbox": [ + 395, + 378, + 410, + 392 + ], + "lines": [ + { + "bbox": [ + 395, + 378, + 410, + 392 + ], + "spans": [ + { + "bbox": [ + 395, + 378, + 410, + 392 + ], + "type": "image", + "image_path": "76ed41d7572652f9f13fd383e21a08c29587282abcb0aff33e1fda6e1640a6c9.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 419, + 380, + 513, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 380, + 513, + 392 + ], + "spans": [ + { + "bbox": [ + 419, + 380, + 513, + 392 + ], + "type": "text", + "content": "Crop categories: 203" + } + ] + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 395, + 399, + 411, + 415 + ], + "blocks": [ + { + "bbox": [ + 395, + 399, + 411, + 415 + ], + "lines": [ + { + "bbox": [ + 395, + 399, + 411, + 415 + ], + "spans": [ + { + "bbox": [ + 395, + 399, + 411, + 415 + ], + "type": "image", + "image_path": "2d1008e2467a3488f108c5c190325a4d5658c7e1519619c3eeab25ce486430af.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 419, + 403, + 529, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 403, + 529, + 415 + ], + "spans": [ + { + "bbox": [ + 419, + 403, + 529, + 415 + ], + "type": "text", + "content": "Disears categories: 684" + } + ] + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 395, + 421, + 411, + 437 + ], + "blocks": [ + { + "bbox": [ + 395, + 421, + 411, + 437 + ], + "lines": [ + { + "bbox": [ + 395, + 421, + 411, + 437 + ], + "spans": [ + { + "bbox": [ + 395, + 421, + 411, + 437 + ], + "type": "image", + "image_path": "4fea14e2638ce0c66458baa1b6610fb1e1b908b1b496f258f298e950b1bfc08f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 418, + 426, + 502, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 418, + 426, + 502, + 437 + ], + "spans": [ + { + "bbox": [ + 418, + 426, + 502, + 437 + ], + "type": "text", + "content": "Expert annotation" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 152, + 509, + 200, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 509, + 200, + 521 + ], + "spans": [ + { + "bbox": [ + 152, + 509, + 200, + 521 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 55, + 535, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 297, + 715 + ], + "type": "text", + "content": "Precise automated understanding of agricultural tasks such as disease identification is essential for sustainable crop production. Recent advances in vision-language models (VLMs) are expected to further expand the range of agricultural tasks by facilitating human-model interaction through easy, text-based communication. Here, we introduce AgroBench (Agronomist AI Benchmark), a benchmark for evaluating VLM models across seven agricultural topics, covering key areas in agricultural engineering and relevant to real-world farming. Unlike recent agricultural VLM benchmarks, AgroBench is annotated by expert agronomists. Our AgroBench covers a state-of-the-art range of categories, including 203 crop categories and 682 disease categories, to thoroughly evaluate VLM capabilities. In our evaluation on AgroBench, we reveal that" + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 606 + ], + "type": "text", + "content": "VLMs have room for improvement in fine-grained identification tasks. Notably, in weed identification, most opensource VLMs perform close to random. With our wide range of topics and expert-annotated categories, we analyze the types of errors made by VLMs and suggest potential pathways for future VLM development. Our dataset and code are available at https://dahlian00.github.io/AgroBenchPage/." + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Agriculture is a fundamental process for humans to produce crops to live and stay healthy. With the development of computer vision technology, effective and automated management of external crop factors such as diseases and pests has been explored, contributing to stable crop" + } + ] + } + ], + "index": 46 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7634" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 70, + 516, + 148 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 516, + 148 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 516, + 148 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 516, + 148 + ], + "type": "table", + "html": "
DatasetAnnot.CropWeedDiseasePestImagesQA pairsMain Purpose
Agri-LLaVA [48]GPT-429-109112391k391k (Synthetic)Training
AgroInstruct [4]GPT-41744741270k70k (Synthetic)Training
CDDM [20]GPT-415-60-137k1M (Synthetic)Training
AgroBench (Ours)Expert2031086821343,7454,342 (Expert)Evaluation
", + "image_path": "8cdc7ad134a16e2cb02e1cabaaff0a62a8503b6679393ae6a1855d8a111ee587.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 553, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 553, + 179 + ], + "type": "text", + "content": "Table 1. Comparison of agricultural vision datasets. Expert refers to 'human expert' in this context. AgroBench provides a comprehensive evaluation framework for assessing VLMs, featuring multiple tasks and a wide range of categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 189, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 189, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 54, + 189, + 294, + 262 + ], + "type": "text", + "content": "production. This includes detecting and classifying undesirable conditions like diseases [1, 10, 13, 34, 38, 41, 47] and pests [3, 49, 53], as well as general plant management tasks such as crop classification [52, 54] and recognition of crop maturity [22, 25, 37, 39], and structure understanding [21, 43]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 295, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 419 + ], + "type": "text", + "content": "To maintain stable crop production, it is important to recognize a wide range of crop conditions, including undesirable situations such as diseases and pests, and to know how to respond appropriately. A single visual model that can handle various conditions, not only disease and pest detection but also treatment and general crop management, would be highly beneficial. However, most existing approaches use task-specific models. These models usually require large amounts of training images and manual annotations for each task. As a result, farmers often need to use several different models depending on the situation. This increases complexity and makes the overall system less accessible for practical use in agriculture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 423, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 295, + 674 + ], + "type": "text", + "content": "For general purpose visual tasks, vision-language models (VLMs) [17, 19, 32, 33, 45, 46, 50] have become widespread recently because they can understand task definitions provided by natural language prompts, covering a wide range of applications without the need for task-specific model training. The recognition ability of VLMs is closely connected to image recognition itself and linked to an object's words, and supports open-vocabulary recognition through web-scale training. Here, zero-shot recognition and few-shot adaptation have also been realized by VLMs with language representations. This ability opens up a wide range of applications; therefore, we believe this can be applied to agricultural scenarios as well. It is worth noting that VLMs offer an easy-to-use interface for the general public, especially the question-answer (QA) and conversation modes. To investigate the range of tasks that VLMs obtained through large-scale vision-language training can cover, recent studies have introduced various benchmark datasets to fully evaluate the VLMs, including tasks such as understanding diagrams and charts [26, 28], and question answering requiring specialized knowledge [7, 55, 56]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "However, VLM research remains underexplored in agriculture due to a lack of benchmark datasets that include diverse tasks and categories in agriculture. Several pioneer-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 189, + 553, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 553, + 369 + ], + "type": "text", + "content": "ing works [20, 48] have adapted open-source VLMs such as LLaVA [18] to the agricultural domain by fine-tuning them on synthetic datasets generated from closed-source VLMs such as GPT-4o. This is recognized as an effective approach because black-box VLMs often possess a certain amount of agricultural knowledge obtained from the Internet. While this helps in generating responses that are generally acceptable to experts, there is almost no way to verify whether the answers are truly correct. Moreover, in the limited evaluation of categories, we are insufficient for fully assessing VLM knowledge; whether they can answer a wide range of types, such as diseases, pests, and weeds. These limitations motivated us to develop a benchmark dataset in order to evaluate the VLM's broad knowledge in the agricultural domain and its applicability as a practical application." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 553, + 647 + ], + "type": "text", + "content": "Here, we introduce AgroBench (Agronomist AI Benchmark), a comprehensive benchmark dataset for VLM for the agricultural domain, covering a state-of-the-art range of categories for agriculture-focused benchmark datasets for VLM; 682 disease, 134 pest, 108 weed, and 203 crop categories (Figure 1). We cover not only identification tasks but also crop production and disease management knowledge. Moreover, we include other important topics related to crop management with 98 machine categories and 77 traditional management methods to investigate more about VLM's ability. We carefully selected benchmark tasks from key agricultural engineering research areas, and also the tasks that address challenges faced by farmers in real agricultural scenarios. We employ a multiple-choice format, and all questions are annotated by human agronomist experts, which overcomes the limitations of previous synthetically created datasets. We carefully selected images from Creative Commons-licensed and publicly available datasets, including real farm conditions. Also, under our manual annotation process, unclear images were removed. As shown in Table 1, AgroBench contains data with the most diverse disease and crop variation categories and many more crop/weed annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 650, + 484, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 650, + 484, + 662 + ], + "spans": [ + { + "bbox": [ + 325, + 650, + 484, + 662 + ], + "type": "text", + "content": "Our main contributions are as follows;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 712 + ], + "type": "text", + "content": "- We developed AgroBench, a benchmark dataset for VLM, to assess their broad agricultural knowledge of VLMs and evaluate their applicability in practical applications." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7635" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 156 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "- In our AgroBench evaluation, VLMs tend to achieve high performance in disease and crop management tasks; however, there is still room for improvement in weed and disease identification." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 295, + 156 + ], + "type": "text", + "content": "- Our AgroBench enables error analysis by providing broader category annotations, highlighting future directions for model training focus areas." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 168, + 141, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 141, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 141, + 181 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 189, + 232, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 232, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 232, + 202 + ], + "type": "text", + "content": "2.1. Computer Vision for Agriculture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 207, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 386 + ], + "type": "text", + "content": "With the rise of computer vision techniques driven by advancements in deep learning, a wide range of agricultural tasks have been explored over the past decade. In particular, dataset construction and benchmarking play an important role in developing models and practical applications in agriculture, while also paving the way for forward-looking advancements. Disease identification is a key research focus across various crops, such as rice [1, 34], tomato [10], cacao [13], and sugarcane [47]. Multicrop datasets have also been created [29, 41]. PlantDoc [41] covers 13 species and 17 classes focused on leaf-based diseases. Plant Village [29] offers 39 classes containing both diseased and healthy leaf categories. In addition to crop disease identification, pest identification [49, 53], weed identification [12, 31, 42] have also been studied." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 387, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 567 + ], + "type": "text", + "content": "While most of the agricultural datasets primarily focus on visual data, few computer vision studies in agriculture investigate the potential of multi-modal approaches. The PlantWild dataset [52] includes 56 plant-disease class pairs, which are collected through image search engines. They implement a CLIP-based model and show the possibility of training with combined text and image data. In an instruction-based format, the CDDM [20] has created 16 categories of crops and 60 categories of crop diseases dataset, which generates instructional data using GPT. While these synthetically created datasets explore the agriculture multi-modal models, there remains a lack of datasets for comprehensive multi-modal model evaluation, validated by human experts and covering a wide range of tasks and categories." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 194, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 194, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 194, + 588 + ], + "type": "text", + "content": "2.2. Vision-Language Models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Models. Visual models in the computer vision field have been accelerated by language modality with the data resource of sentence-level inputs and web-scale texts. Specifically, CLIP [35] has made a significant contribution in this context, by text and image feature alignment through contrastive learning. CLIP played a key role in introducing more sophisticated visual representation into language explanations e.g., Flamingo [2], BLIP [14, 15], Qwen [5], PaLI [8], LLaVA [17, 18], CogVLM [50], and Emu [44, 45, 51]. Closed-source VLMs achieve state-of-the-art per" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "formance, such as GPT-4o [32] and Gemini Pro [46], which are said to acquire human-level knowledge across diverse fields on the Internet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 300 + ], + "type": "text", + "content": "**Benchmarks.** Along this line, several recent studies introduced benchmark datasets. Especially, the representative examples in terms of universal knowledge benchmarking include MMMU [55], MMMU-Pro [56] and MMStar [6] for multi-modal understanding. These benchmarks contain highly diverse domains (e.g., natural, graph, illustration, medical images) and academic fields (e.g., science, engineering, art, and medical fields) under the tasks of vision language such as question-answering and reasoning. The series of MMMUs have been verified with foundation models like GPT-4V, but they cannot answer the questions perfectly. More specific domain datasets have also been proposed, such as those for Medicine [27, 36], Chart [26, 28, 40], and Video understandings [9, 16, 30], contributing to the accurate evaluation of VLMs and guiding future research directions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 310, + 388, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 310, + 388, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 388, + 323 + ], + "type": "text", + "content": "3. AgroBench" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 449 + ], + "type": "text", + "content": "This section introduces AgroBench, the first comprehensive benchmark dataset to evaluate VLM models from the perspective of agricultural vision tasks. The benchmark consists of seven tasks covering a wide range of tasks selected from key agricultural engineering research areas, as well as tasks that address real-world challenges faced by farmers in real agriculture scenarios. AgroBench includes 682 disease categories, 134 pest categories, 203 crop categories, and 108 weed categories, representing the largest number of categories in each area to date, to the best of our knowledge." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 528, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 528, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 528, + 470 + ], + "type": "text", + "content": "3.1. Benchmark Tasks in Agricultural Scenes" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 533 + ], + "type": "text", + "content": "The seven benchmark tasks encompass key research areas in agricultural engineering as well as real-world challenges faced by farmers. To facilitate VLM evaluation, we also provide prompts to address each task in a question-answer format. The details of each task are described below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": "1) Disease Identification (DID). The DID task aims to accurately diagnose and classify crop diseases. This is a key task in agriculture to protect crop health and maximize yields. Our benchmark provides 1,502 QA pairs that cover 370 disease categories, 160 crop categories and 682 crop-disease combinations. To thoroughly evaluate the capabilities of VLMs, we include four misleading disease labels for each image, featuring diseases with similar symptoms or common diseases affecting the target plant. VLMs are required to diagnose the disease based on the image, considering both symptoms and the crop species. Figure 2a illustrates example images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "2) Pest Identification (PID). The PID task aims to identify pests to prevent infestations that can severely impact crop health. Accurate identification reduces economic losses and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7636" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 81, + 550, + 336 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 151, + 80 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 151, + 80 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 151, + 80 + ], + "type": "text", + "content": "(a) Disease identification (DID)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 81, + 550, + 336 + ], + "lines": [ + { + "bbox": [ + 59, + 81, + 550, + 336 + ], + "spans": [ + { + "bbox": [ + 59, + 81, + 550, + 336 + ], + "type": "image", + "image_path": "3d4a2bf5549330284c94ef35053d17fd9ed5e6a77d73d9be524f7b58ef53b0c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 380, + 266, + 563 + ], + "blocks": [ + { + "bbox": [ + 55, + 346, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 553, + 369 + ], + "type": "text", + "content": "Figure 2. Examples of labeled images for DID, PID, and WID tasks. Our dataset includes 682 crop-disease pairs, 134 pest categories, and 108 weed categories. We prioritized collecting images from real farm settings." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 380, + 266, + 563 + ], + "lines": [ + { + "bbox": [ + 86, + 380, + 266, + 563 + ], + "spans": [ + { + "bbox": [ + 86, + 380, + 266, + 563 + ], + "type": "image", + "image_path": "de6f8bc38f4fd53f589354903603a699929451185ee9edf53ae4a78cf7854319.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 575, + 295, + 620 + ], + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 620 + ], + "type": "text", + "content": "Figure 3. Seven benchmark tasks in AgroBench. AgroBench includes multiple topics with a diverse range of categories. The total accuracy is calculated by the average of each task to mitigate the difference in QAs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "minimizes harm to the environment. Our benchmark provides 544 labeled images that cover 134 pest categories, including insects, mites, and other organisms harmful to plants. For categories where we obtain multiple images, we select multiple insect growth stages whenever possible. To fully evaluate the VLMs, we assign alternative choices that closely resemble or are commonly associated with the target" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 379, + 544, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 544, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 544, + 391 + ], + "type": "text", + "content": "crops. Figure 2b shows examples of labeled pest images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 394, + 556, + 715 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 502 + ], + "type": "text", + "content": "3) Weed Identification (WID). The WID task aims to identify weed species. Our benchmark includes 609 images of weeds with ground-truth bounding boxes, covering 108 weed species commonly found in farm fields. We assign bounding boxes because multiple weed types often grow closely together, and we want to clarify which one is the target. Specifically, VLMs are required to identify the weed species within the provided bounding box on the image. Figure 2c shows examples of labeled weed images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 662 + ], + "type": "text", + "content": "4) Crop Management (CMN). Crop management focuses on optimizing farming practices to facilitate crop growth. This involves making decisions on irrigation, fertilization, planting times, and other cultivation practices. Our benchmark provides 411 question-answer pairs for this task. VLMs are required to analyze images of crops and recommend appropriate management strategies by considering factors such as crop health, growth stage, and environmental conditions visible in the images given five answer candidates. In Figure 4a and b, we show two examples of the harvest timing for white and green asparagus, respectively. Our dataset includes complex questions that take into account the differences in their harvest timing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "5) Disease Management (DMN). Disease management aims to control and reduce diseases in crops. This involves informed decisions on interventions such as applying pesticides, adopting resistant crop varieties, or modifying cul" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7637" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 116, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 116, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 116, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 116, + 131 + ], + "type": "image", + "image_path": "4f30009ad05735b9fef67432f58689b7d107f17aac1a3937140a67ac61fae5b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 72, + 306, + 131 + ], + "blocks": [ + { + "bbox": [ + 117, + 72, + 306, + 131 + ], + "lines": [ + { + "bbox": [ + 117, + 72, + 306, + 131 + ], + "spans": [ + { + "bbox": [ + 117, + 72, + 306, + 131 + ], + "type": "image", + "image_path": "076cf1960b742c694f51e06de401e57dcecbd814ee68a1fd2330c0958c7bff6e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 308, + 72, + 365, + 131 + ], + "blocks": [ + { + "bbox": [ + 308, + 72, + 365, + 131 + ], + "lines": [ + { + "bbox": [ + 308, + 72, + 365, + 131 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 365, + 131 + ], + "type": "image", + "image_path": "22f50d4ecc4936e5fbcc49448c773614189727479f39b94e29ed005037bc6335.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 366, + 72, + 548, + 131 + ], + "blocks": [ + { + "bbox": [ + 366, + 72, + 548, + 131 + ], + "lines": [ + { + "bbox": [ + 366, + 72, + 548, + 131 + ], + "spans": [ + { + "bbox": [ + 366, + 72, + 548, + 131 + ], + "type": "image", + "image_path": "6f25c0cf5da35a4823f8139524b3df03fcc6b1bf6ee618a981dbc71c8454bf43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 133, + 116, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 133, + 116, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 133, + 116, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 133, + 116, + 190 + ], + "type": "image", + "image_path": "6d00e18010cf57c343a3fd5b562a646a52fa3cad0e097a6a3c842e715d6267af.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 242 + ], + "type": "text", + "content": "Figure 4. Examples of QA pairs for CMN and DMN tasks. (a) and (b) Crop management QA types for white asparagus and asparagus, respectively. Their difference in the harvest timing affects the answer's difference correctly. (c) and (d) Disease management QA types for the alfalfa bacterial leaf spot with the initial and severe symptoms, respectively. Based on the severity of the symptoms, the annotator changes the answer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 117, + 133, + 298, + 190 + ], + "blocks": [ + { + "bbox": [ + 117, + 133, + 298, + 190 + ], + "lines": [ + { + "bbox": [ + 117, + 133, + 298, + 190 + ], + "spans": [ + { + "bbox": [ + 117, + 133, + 298, + 190 + ], + "type": "image", + "image_path": "881a3aff989b5b9ba5381602e7781cc9ab3ba2987b72af4d439156ec73c795cc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 133, + 364, + 190 + ], + "blocks": [ + { + "bbox": [ + 309, + 133, + 364, + 190 + ], + "lines": [ + { + "bbox": [ + 309, + 133, + 364, + 190 + ], + "spans": [ + { + "bbox": [ + 309, + 133, + 364, + 190 + ], + "type": "image", + "image_path": "641deafee5d7f9a560c0bb7676fa3590c8097bd4dbcdd0f33673b761e9db8c78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 365, + 133, + 548, + 190 + ], + "blocks": [ + { + "bbox": [ + 365, + 133, + 548, + 190 + ], + "lines": [ + { + "bbox": [ + 365, + 133, + 548, + 190 + ], + "spans": [ + { + "bbox": [ + 365, + 133, + 548, + 190 + ], + "type": "image", + "image_path": "8a729418308079261a220c19eb3ee2ae5a75dd4629d0c7b3e4df566fe914fc51.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 252, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 54, + 252, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 296, + 420 + ], + "type": "text", + "content": "tivation practices. Our benchmark provides 569 question-answer pairs covering 141 crop-disease combination categories. Since many diseases share the same management strategies regardless of the crop, we carefully select a diverse range of disease types and management strategies. VLMs are required not only to identify the disease but also to recommend appropriate management strategies based on images of affected crops. In Figure 4c and d show example QAs from the Disease Management tasks. Both images depict Bacterial leaf spot on alfalfa, with c showing the initial stage of the disease and d showing the severe stage. Based on the severity of the disease, we provide different answer options and set different correct answers. The input prompt consists of a question and five answer candidates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 441, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 441, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 54, + 441, + 295, + 573 + ], + "type": "text", + "content": "6) Machine Usage QA (MQA). Machine usage QA addresses the correct use and choice of agricultural machinery depending on the task and farming conditions. Selecting the appropriate machinery is essential for efficient farming. Our benchmark provides 303 question-answer pairs covering 98 machine categories. Given that many crops share the same machinery (e.g., soil preparation and irrigation), the coverage of this category is comprehensive. VLMs are required to answer questions about machinery operation or select the appropriate machine for a given scenario based on images of machinery or farming conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 594, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 712 + ], + "type": "text", + "content": "7) Traditional Management (TM). Traditional management methods involve natural and sustainable approaches to farming, such as the use of organic fertilizers, terrace farming, and agroforestry. Recent computer vision studies have not focused on these traditional practices, although many are still used by certain local farmers. Our benchmark includes 404 question-answer pairs, including 77 traditional management practices. VLMs are required to identify the management method or explain its effectiveness, given five answer choices." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 251, + 437, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 437, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 437, + 263 + ], + "type": "text", + "content": "3.2. Dataset Construction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 617 + ], + "type": "text", + "content": "Image Selection. To establish a high-quality benchmark covering a wide range of crops, diseases, and pest categories we initially curated around 50,000 agricultural images from websites supervised by plant pathologists, either where redistribution was permitted or where we obtained redistribution permission. We selected images in real farm settings as much as possible, as shown in Figure 2, to evaluate real-world scenarios. When licensed images for a target category were limited, we used laboratory setting images. For curation, we obtained images along with their corresponding labels. The annotator is one of the authors, who holds a Ph.D. in Agriculture. In the human evaluation conducted during the experiment, other individuals with a Ph.D. or M.S. degree in Agriculture reviewed questions to assess the quality of the QA pairs. The annotator selected images that clearly represented target labels and removed those images with irrelevant labels. For the Machine usage QA and Traditional management, annotators manually created categories based on textbooks and websites, and searched for corresponding images with redistribution licenses. For Weed identification, we use existing dataset [11, 23, 24, 42]. For this Weed identification category, we will provide a simple code to download the data and crop images and assign bounding boxes for Weed Identification, allowing users to work with the existing dataset without redistributing the images. Through these image selection processes, the selected images are high-quality 4,218 representative images. The detailed distribution of tasks and categories for each task is summarized in Figure 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "QA Annotation. For the DMN, CMN, MQA, and TM tasks, all question-answer pairs were created manually, independent of any LLMs or VLMs. Annotators used GPT only for sentence rephrasing, but they were prohibited from using any knowledge from GPT for QA creation. Annotators were allowed to refer to textbooks, academic journals, and other authoritative sources in the field to ensure the accuracy and depth of the dataset. This took around 150 man" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7638" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 523, + 313 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 523, + 313 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 523, + 313 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 523, + 313 + ], + "type": "table", + "html": "
ModelDIDDMNPIDWIDCMNMQATMOverall (all)Overall (subset)
Random Choice21.7715.6420.4017.9016.0622.1119.3119.0319.11
Human25.0022.5045.0020.0036.2557.5051.25-36.79
Closed-Source Vision Language Models (VLMs)
GPT-4o mini [33]53.6080.6760.0435.1464.2370.9669.8062.0669.65
GPT-4o [32]64.1887.3577.7644.1775.4382.8482.4373.4579.26
Gemini1.5-Flash [46]55.0679.9670.0450.9064.7278.2273.2767.4568.82
Gemini1.5-Pro [46]62.9281.5574.4555.1771.0582.8477.7272.2469.74
Open-Source Vision Language Models (VLMs)
EMU2Chat [44]42.0148.3343.7523.8140.3937.6247.7740.5333.84
LLaVA-Next-8B [19]45.4772.5843.0130.0554.2656.1157.4651.2857.84
LLaVA-Next-72B [19]54.9580.0049.8126.9866.9266.1170.3859.3164.36
QwenVLM-7B [5]51.2680.4963.9733.1766.4276.2477.4864.1566.41
QwenVLM-72B [5]57.9987.8773.3534.4875.9180.8684.1670.6672.45
CogVLM-19B [50]29.1653.7852.3925.4554.0171.6266.0950.3644.27
LLaVa-7B [18]36.0262.7438.7924.7953.7746.5355.2045.4146.14
LLaVA-13B [18]40.2168.8944.4924.7959.3754.1358.4250.0455.31
", + "image_path": "0c2385a36ec4186adeb5773a47c99b07af88d4cce71892a0f4c830d3e1cfe5ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 113, + 356, + 499, + 402 + ], + "blocks": [ + { + "bbox": [ + 55, + 323, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 553, + 346 + ], + "type": "text", + "content": "Table 2. Results for the seven benchmark tasks with images. We provide results for Random Choice, Human Validation, four closed-source VLMs, and open-source VLMs. Human validation was conducted by 28 people on a subset of 80 samples per task as a reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 113, + 356, + 499, + 402 + ], + "lines": [ + { + "bbox": [ + 113, + 356, + 499, + 402 + ], + "spans": [ + { + "bbox": [ + 113, + 356, + 499, + 402 + ], + "type": "table", + "html": "
DIDDMNPIDWIDCMNMQATMOverall
GPT-4o [32]1.9372.5818.751.0040.3925.0848.2729.71
LLaVA-Next-8B [19]26.1070.3021.8819.7053.7730.3640.3537.49
", + "image_path": "f7cc85175df4795f87fff46bfb919ea2fa21b7c50ac3d5fd0ba274fce07911ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 555, + 433 + ], + "type": "text", + "content": "Table 3. Results for the seven benchmark tasks with text only. We additionally evaluate the model without image inputs, and the overall performance is close to random." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 440, + 295, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 487 + ], + "type": "text", + "content": "hours. We carefully annotate various types of questions. Examples of QAs that require expert knowledge are shown in Figure 4. All questions were carefully created so that image reference is necessary for answering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 559 + ], + "type": "text", + "content": "Dataset Statistics. Following the dataset annotations and selections, our AgroBench comprises seven tasks with 4,342 QA pairs, as shown in Table 1. All tasks comprise a wide range of categories for detailed evaluation. Please refer to the supplementary materials for more details and examples." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Baseline Models. We use four closed-source models: GPT-4o [32], GPT-4o mini [33], Gemini1.5-Pro [46], and Gemini1.5-Flash [46]. GPT-4o mini and Gemini1.5-Flash are down-scale versions of GPT-4o and Gemini1.5-Pro, respectively. We use eight open-source models: EMU2Chat [44], LLaVA-Next-8B [19], LLaVA-Next-72B [19], QwenVLM-7B [5], QwenVLM-72B [5], CogVLM-19B [50], LLaVa-7B [18], and LLaVa-13B [18]. For the details of these models, please refer to the supple" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 388, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 388, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 388, + 452 + ], + "type": "text", + "content": "mentary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 599 + ], + "type": "text", + "content": "Human Results. We also present results from human participants for reference. We surveyed 28 students, each holding at least a bachelor's degree in agriculture, and asked them to answer 20 questions each. This created a test subset of 280 questions, with each question answered by two participants, resulting in a total of 560 responses. We averaged the results per task and reported the accuracy. Each participant was permitted to use a book or translator to look up word meanings but was prohibited from using the internet for searches. If participants were unsure of the answer, they were asked to provide the response they believed to be most accurate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Evaluation Protocol. Importantly, our dataset evaluation is conducted per task, and overall scores are averaged based on the number of tasks, not the number of QAs. This prevents categories with a large number of evaluations from becoming dominant. We adopt an exact matching approach for our five-option questions. If the model's response matches the option's letter or the answer sentence, we consider it correct. If the model's answer does not match any option, including cases where there is no answer or mul" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7639" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 553, + 185 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 553, + 185 + ], + "type": "image", + "image_path": "2fab5047c0c1479fdb57a3923aafe09d26fc86801887fd7a570cc75d4c1e0152.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 555, + 220 + ], + "type": "text", + "content": "Figure 5. Results of seven benchmark tasks with Chain of Thought (CoT). Baseline indicates results without CoT. In the one-shot, two-shot, and three-shot settings, we provide one, two, and three CoT examples per task, respectively, to guide the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 232, + 550, + 381 + ], + "blocks": [ + { + "bbox": [ + 59, + 232, + 550, + 381 + ], + "lines": [ + { + "bbox": [ + 59, + 232, + 550, + 381 + ], + "spans": [ + { + "bbox": [ + 59, + 232, + 550, + 381 + ], + "type": "image", + "image_path": "a1d88734ca867c9fda3e9719d9476adcba2160f86d85b76f5a980c6b9e4fdd7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 6. Error analysis on seven benchmark tasks with GPT-4o. We extract a maximum of 15 errors per task from the zero-shot CoT result. We manually analyze how they conclude the incorrect answer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 212, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 212, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 212, + 433 + ], + "type": "text", + "content": "tiple answers, we consider it incorrect." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 443, + 141, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 141, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 141, + 453 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "type": "text", + "content": "Here, we discuss the main results of our AgroBench evaluation performance. We evaluate our AgroBench using eight open-source VLMs and four closed-source VLMs with APIs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 628 + ], + "type": "text", + "content": "Challenges of the AgroBench. We show the main results in Table 2. The most difficult task is Weed Identification (WID), on which most open-source VLMs perform at around the random score. The highest WID accuracy, " + }, + { + "bbox": [ + 55, + 510, + 295, + 628 + ], + "type": "inline_equation", + "content": "55.17\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 628 + ], + "type": "text", + "content": ", was achieved by Gemini 1.5-Pro. This suggests that the knowledge about weeds is not as fully trained as that of crops. All models' Disease Identification (DID) results are lower than their Disease Management (DMN) results. This means that VLM models can gain contextual information, but there is still room for perceptual improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Model Comparison. Overall, closed-source VLMs achieve better results than open-source VLMs and achieve higher performance than humans. GPT-4o model achieved the highest score in overall performance. Among the open-source models, QwenVLM-72B achieves the best result in overall accuracy, which is comparable or even superior to open-source VLMs in some tasks. QwenVLM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 445 + ], + "type": "text", + "content": "72B achieves satisfactory results on both identification and question-answering benchmarks on AgroBench." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 455, + 382, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 382, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 382, + 467 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 473, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 688 + ], + "type": "text", + "content": "Context Reliance. We further evaluate the results on AgroBench using text-only input to determine whether visual information is necessary to answer the questions. Table 3 presents the experimental results across seven benchmark tasks on AgroBench. With text-only input, performance drops across all models, confirming the reliance on visual information. However, for both models, the DMN, CMN, and TM tasks maintain significantly higher accuracy than random selection. Although we ensured that our questions do not include disease names or appearance-related traits, models tend to infer answers based on estimation. This suggests that many disease types share common management strategies, such as avoiding humidity, preferring cooler temperatures, or pruning infected parts, allowing models to predict the most likely option. A similar pattern is observed in the CMN and TM tasks, where models can make educated guesses based on contextual cues. (See the supplementary material for further detailed examples.)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Chain of Thought. We evaluate the effectiveness of Chain of Thought (CoT) reasoning. In the one-shot setting, we" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 76, + 161, + 87 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 76, + 161, + 87 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 161, + 87 + ], + "type": "text", + "content": "(a) Lack of Knowledge" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 89, + 193, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 89, + 193, + 99 + ], + "spans": [ + { + "bbox": [ + 64, + 89, + 193, + 99 + ], + "type": "text", + "content": "Q: What disease is affecting the plant?" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 100, + 156, + 173 + ], + "blocks": [ + { + "bbox": [ + 63, + 100, + 156, + 173 + ], + "lines": [ + { + "bbox": [ + 63, + 100, + 156, + 173 + ], + "spans": [ + { + "bbox": [ + 63, + 100, + 156, + 173 + ], + "type": "image", + "image_path": "c0e92e8e14c4de01ed1ee34a40c0381608f2be26e92872a727e1460dc74d5e56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 158, + 99, + 288, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 99, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 158, + 99, + 288, + 173 + ], + "type": "text", + "content": "GPT-4o: ..., **Sclerotinia blight**: Characterized by white, cottony mycelium and black sclerotia on stems. The image does not show these symptoms. (not enough knowledge) ..., the plant is likely affected by **D. Cylindrocladium black rot**: (This is more common on root)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 175, + 132, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 175, + 132, + 185 + ], + "spans": [ + { + "bbox": [ + 64, + 175, + 132, + 185 + ], + "type": "text", + "content": "A: Sclerotinia blight" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 190, + 152, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 190, + 152, + 201 + ], + "spans": [ + { + "bbox": [ + 63, + 190, + 152, + 201 + ], + "type": "text", + "content": "(b) Perceptual Error" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 202, + 261, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 202, + 261, + 220 + ], + "spans": [ + { + "bbox": [ + 64, + 202, + 261, + 220 + ], + "type": "text", + "content": "Q: What activity is being performed using the tractor-pulled machine shown in the image?" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 221, + 162, + 291 + ], + "blocks": [ + { + "bbox": [ + 63, + 221, + 162, + 291 + ], + "lines": [ + { + "bbox": [ + 63, + 221, + 162, + 291 + ], + "spans": [ + { + "bbox": [ + 63, + 221, + 162, + 291 + ], + "type": "image", + "image_path": "32076a514abab6329414e314b00c8e580b53bda312b3345459d77b27f5496820.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 295, + 343 + ], + "type": "text", + "content": "Figure 7. Error examples of GPT-4o. Examples illustrate the two main error types: Lack of Knowledge and Perceptual Error." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 164, + 220, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 220, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 164, + 220, + 288, + 293 + ], + "type": "text", + "content": "GPT-40:The machine seems to be designed to lift plants from the soil, ..., and there are plants being lifted from the ground (There are no harvested ctops)....the activity being performed is: D. Harvesting potato tubers from the soil using a tractor-pulled harvesting machine." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 296, + 244, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 296, + 244, + 315 + ], + "spans": [ + { + "bbox": [ + 65, + 296, + 244, + 315 + ], + "type": "text", + "content": "A: Cutting and chopping the stems and leaves using a tractor-pulled leaf chopper." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 542 + ], + "type": "text", + "content": "provide a single CoT reasoning example along with an image and the corresponding prompt for each task to guide the model in answering the question. We select 100 QAs per task for evaluation and use GPT-4 as the VLM. Figure 5 presents the accuracy of CoT reasoning with the baseline results. While CoT achieves slightly higher accuracy, the improvement is not significant compared to the baseline. However, CoT demonstrates effectiveness in certain tasks, such as PID, WID, CMN, and TM. For instance, in PID, the VLM performs step-by-step and careful reasoning (see the supplementary material for more detailed examples). In WID, the most challenging task, CoT provides useful examples that may help the model make predictions. Overall, CoT contributes to accuracy, but we observe performance saturation in the three-shot setting. For further detailed examples, please refer to the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 551, + 148, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 148, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 148, + 563 + ], + "type": "text", + "content": "4.4. Error Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "type": "text", + "content": "Here, we analyze the mistake types that occur depending on the agricultural tasks. We bring out up to 15 failure examples per task from the zero-shot CoT results and manually analyze how they reached the incorrect answers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Lack of Knowledge (51.92%). This includes cases where VLM can't accurately describe the appearance or relevant knowledge of a choice (e.g., VLM fails to describe disease symptoms or insect characteristics) or lacks context (e.g., VLM doesn't know how to treat diseased crops or manage crops for high yield). Figure 7a shows an example of a Lack of Knowledge case. When the VLM analyzes the correct answer option for Sclerotinia blight, it fails to describe the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "symptoms of discoloration and wilting. Additionally, its incorrect answer choice is more commonly associated with the root rather than the stem. These errors are based on a Lack of Knowledge, suggesting VLMs need more detailed categories and domain-specific training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 133, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 300 + ], + "type": "text", + "content": "Perceptual Error (32.69%). This indicates that VLM can't pay attention or recognize the answer-related part in the image (e.g., can't recognize the green insect on the leaf), and VLM misunderstands the image, leading to incorrect answers. Figure 7b shows an example of the model hallucinating and misunderstanding the situation. First, the VLM incorrectly identifies the machine as one used for lifting plants from the soil. Then, it describes the scene as if there are harvested crops, even though no harvested crops are present. These errors can be mitigated by enhancing the VLM's perception abilities for domain-specific classes. Additionally, improving general perception capabilities, including reducing hallucinations, can further contribute to VLM performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 302, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 554, + 374 + ], + "type": "text", + "content": "Reasoning Error (7.6%). Reasoning error involves the VLM can describe the options correctly, but can't compare them step by step and conclude the wrong answer. This error is relatively low compared to the existing work [55] since AgroBench requires more specific knowledge and doesn't include reasoning relying on problems (e.g., math)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": "Other Errors (7.79%). For the other errors, we observe Shortcut Error (The VLM can pick up the two candidate options correctly but conclude the answer without comparing the candidates), Double Answer Error (Concluding two answers are correct), Interpretation Misunderstanding (VLM misleading the question and conclude wrong answer), and Reject to Answer (VLM conclude there is no answer)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 496, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 605 + ], + "type": "text", + "content": "In this paper, we develop AgroBench, a comprehensive benchmark dataset for VLMs in the agricultural domain, covering a state-of-the-art range of categories. Our dataset comprises seven benchmark tasks encompassing key research areas in agricultural engineering as well as real-world challenges faced by farmers. AgroBench contributes to agricultural VLM research by addressing the lack of datasets for comprehensive multi-modal model evaluation, validated by human experts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "In our evaluation, VLMs exhibit strengths across different tasks. However, in several tasks such as weed identification and disease identification, all models show room for improvement. Our error analysis reveals that most failures are due to a lack of knowledge (51.92%), suggesting that VLMs require more specialized agricultural knowledge. Our dataset will facilitate agricultural VLM research, enabling broad category and task evaluation to support sustainable, automated agriculture." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7641" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 160, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "content": "This work was supported by the AIST KAKUSEI Project (FY2024) and JST FOREST Grant Number JPMJFR206F. We would like to thank the Agricultural Administration Division, Department of Agriculture, Hokkaido Government, for providing some of the images. We are also grateful to Daniel Steininger for contributing images from the CropAndWeed dataset to our dataset. We used ABCI 3.0 provided by AIST and AIST Solutions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 115, + 215 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 222, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 266 + ], + "type": "text", + "content": "[1] Petchiammal A, Briskline Kiruba S, Murugan D, and Pandarasamy Arjunan. Paddy doctor: A visual image dataset for automated paddy disease classification and benchmarking. IEEE Dataport, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 268, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 268, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 268, + 296, + 388 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Proc. Annual Conference on Neural Information Processing Systems (NeurIPS), pages 23716-23736, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 390, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 390, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 390, + 295, + 443 + ], + "type": "text", + "content": "[3] Abderraouf Amrani, Ferdous Sohel, Dean Diepeveen, David Murray, and Michael G.K. Jones. Deep learning-based detection of aphid colonies on plants from a reconstructed brassica image dataset. Computers and Electronics in Agriculture, 205:107587, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 446, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 446, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 61, + 446, + 294, + 499 + ], + "type": "text", + "content": "[4] Muhammad Awais, Ali Husain Salem Abdulla Alharthi, Amandeep Kumar, Hisham Cholakkal, and Rao Muhammad Anwer. Agropt: Efficient agricultural vision-language model with expert tuning. arXiv preprint arXiv:2410.08405, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 502, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 502, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 61, + 502, + 295, + 633 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou, and Tianhang Zhu. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 294, + 689 + ], + "type": "text", + "content": "[6] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "text", + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao," + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 107, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 217 + ], + "type": "text", + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Alexander Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, and Radu Soricut. Pali: A jointly-scaled multilingual language-image model. 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 218, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 218, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 218, + 553, + 251 + ], + "type": "text", + "content": "[9] Daniel Cores, Michael Dorkenwald, Manuel Mucientes, Cees G. M. Snoek, and Yuki M. Asano. Tvbench: Redesigning video-language evaluation. 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "text", + "content": "[10] Mamta Gehlot, Rakesh Saxena, and Geeta Gandhi. \"tomato-village\": a dataset for end-to-end tomato disease detection in a real-world environment. Multimedia Systems, 29:1-24, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 297, + 526, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 526, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 526, + 308 + ], + "type": "text", + "content": "[11] Benedikt Geisler. Perrenial plants detection, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[12] Sebastian Haug and Jörn Ostermann. A crop/weed field image dataset for the evaluation of computer vision based precision agriculture tasks. In Proc. European Conference on Computer Vision (ECCV), pages 105-116. Springer, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[13] Atuhurra Jesse, N'guessan Yves-Roland Douha, and Pabitra Lenka. Image classification for cssvd detection in cacao plants. arXiv preprint 2405.04535, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[14] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[15] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning, pages 19730–19742. PMLR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 543 + ], + "type": "text", + "content": "[16] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, Limin Wang, and Yu Qiao. Mybench: A comprehensive multimodal video understanding benchmark. In CVPR, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 545, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 577 + ], + "type": "text", + "content": "[17] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv:2310.03744, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 579, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 600 + ], + "type": "text", + "content": "[18] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 634 + ], + "type": "text", + "content": "[19] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.2,6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[20] Xiang Liu, Zhaoxiang Liu, Huan Hu, Zezhou Chen, Kohou Wang, Kai Wang, and Shiguo Lian. A multimodal benchmark dataset and model for crop disease diagnosis. In Proc. European Conference on Computer Vision (ECCV), 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[21] Xinpeng Liu, Kanyu Xu, Risa Shinoda, Hiroaki Santo, and Fumio Okura. Masks-to-skeleton: Multi-view mask-based" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7642" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "tree skeleton extraction with 3d gaussian splatting. Sensors, 25(14), 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[22] Shenglian Lu, Wenkang Chen, Xin Zhang, and Manoj Karkee. Canopy-attention-yolov4-based immature/mature apple fruit detection on dense-foliage tree architectures for early crop load estimation. Computers and Electronics in Agriculture, 193:106696, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 243, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 243, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 243, + 163 + ], + "type": "text", + "content": "[23] Yuzhen Lu. Cottonweeddet3. kaggle, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "text", + "content": "[24] Simon Leminen Madsen, Solvejg Kopp Mathiassen, Mads Dyrmann, Morten Stigaard Laursen, Laura-Carlota Paz, and Rasmus Nyholm Jorgensen. Open Plant Phenotype Database of Common Weeds in Denmark, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "type": "text", + "content": "[25] Abdul Khalique Maitlo, Abdul Aziz, Hassnian Raza, and Neelam Abbas. A novel dataset of guava fruit for grading and classification. Data in Brief, 49:109462, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 308 + ], + "type": "text", + "content": "[26] Ahmed Masry, Do Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, Dublin, Ireland, 2022. Association for Computational Linguistics. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 386 + ], + "type": "text", + "content": "[27] João Matos, Shan Chen, Siena Placino, Yingya Li, Juan Carlos Climent Pardo, Daphna Idan, Takeshi Tohyama, David Restrepo, Luis F. Nakayama, Jose M. M. Pascual-Leone, Guergana Savova, Hugo Aerts, Leo A. Celi, A. Ian Wong, Danielle S. Bitterman, and Jack Gallifant. Worldmedqa-v: a multilingual, multimodal medical examination dataset for multimodal language models evaluation, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 420 + ], + "type": "text", + "content": "[28] Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1516-1525, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 453 + ], + "type": "text", + "content": "[29] Sharada P. Mohanty, David P. Hughes, and Marcel Salathé. Using deep learning for image-based plant disease detection. Frontiers in Plant Science, 7, 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "text", + "content": "[30] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "text", + "content": "[31] Alex Olsen, Dmitry A Konovalov, Bronson Philippa, Peter Ridd, Jake C Wood, Jamie Johns, Wesley Banks, Benjamin Girgenti, Owen Kenny, James Whinney, et al. Deepweeds: A multiclass weed species image dataset for deep learning. Scientific reports, 9(1):2058, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 275, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 275, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 275, + 567 + ], + "type": "text", + "content": "[32] OpenAI. Gpt-4o, 2024. Accessed: 2024-11-13. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 568, + 285, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 285, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 285, + 578 + ], + "type": "text", + "content": "[33] OpenAI. Gpt-4o mini, 2024. Accessed: 2024-11-13. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 612 + ], + "type": "text", + "content": "[34] Chiranjit Pal, Imon Mukherjee, Sanjay Chatterji, Sanjoy Pratihar, Pabitra Mitra, and Partha Pratim Chakrabarti. Indian rice disease dataset (irdd), 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 677 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[36] Corentin Royer, Bjoern Menze, and Anjany Sekuboyina. Multimedieval: A benchmark and a toolkit for evaluating medical vision-language models, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[37] Rahman Sanya, Ann Lisa Nabiryo, Jeremy Francis Tusubira, Sudi Murindanyi, Andrew Katumba, and Joyce Nakatumba-Nabende. Coffee and cashew nut dataset: A dataset for detection, classification, and yield estimation for machine learning applications. Data in Brief, 52:109952, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[38] Nabila Husna Shabrina, Siwi Indarti, Rina Maharani, Dinar Ajeng Kristiyanti, Irmawati, Niki Prastomo, and Tika Adilah M. A novel dataset of potato leaf disease in uncontrolled environment. Data in Brief, 52:109955, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[39] Risa Shinoda, Hirokatsu Kataoka, Kensho Hara, and Ryozo Noguchi. Transformer-based ripeness segmentation for tomatoes. Smart Agricultural Technology, 4:100196, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "text", + "content": "[40] Risa Shinoda, Kuniaki Saito, Shohei Tanaka, Tosho Hirasawa, and Yoshitaka Ushiku. Sbs figures: Pre-training figure qa from stage-by-stage synthesized images, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[41] Davinder Singh, Naman Jain, Pranjali Jain, Pratik Kayal, Sudhakar Kumawat, and Nipun Batra. Plantdoc: A dataset for visual plant disease detection. In Proc. ACM IKDD CoDS and COMAD, page 249-253, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 342 + ], + "type": "text", + "content": "[42] Daniel Steininger, Andreas Trondl, Gerardus Croonen, Julia Simon, and Verena Widhalm. The cropandweed dataset: A multi-modal learning approach for efficient crop and weed manipulation. In WACV, pages 3729-3738, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 408 + ], + "type": "text", + "content": "[43] Daniel Steininger, Julia Simon, Andreas Trondl, and Markus Murschitz. Timbervision: A multi-task dataset and framework for log-component segmentation and tracking in autonomous forestry operations. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. arXiv preprint arXiv:2312.13286, 2023. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[45] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "text", + "content": "[46] Google Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context, 2024. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[47] Sandip Thite, Yogesh Suryawanshi, Kailas Patil, and Prawit Chumchu. Sugarcane leaf dataset: A dataset for disease detection and classification for machine learning applications. Data in Brief, 53:110268, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[48] Liqiong Wang, Teng Jin, Jinyu Yang, Ales Leonardis, Fangyi Wang, and Feng Zheng. Agri-llava: Knowledge-infused large multimodal assistant on agricultural pests and diseases. arXiv preprint arXiv:2412.02158, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[49] Rujing Wang, Liu Liu, Chengjun Xie, Po Yang, Rui Li, and Man Zhou. Agripest: A large-scale domain-specific benchmark dataset for practical agricultural pest detection in the wild. Sensors, 21(5), 2021. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[50] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7643" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 429 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models. In NeurIPS, 2024. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[51] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[52] Tianqi Wei, Zhi Chen, Zi Huang, and Xin Yu. Benchmarking in-the-wild multimodal disease recognition and a versatile baseline. In Proc. ACM International Conference on Multimedia (ACMMM), 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[53] Xiaoping Wu, Chi Zhan, Yukun Lai, Ming-Ming Cheng, and Jufeng Yang. Ip102: A large-scale benchmark dataset for insect pest recognition. In CVPR, pages 8787-8796, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[54] Momchil Yordanov, Raphaël d'Andrimont, Laura Martinez-Sanchez, Guido Lemoine, Dominique Fasbender, and Marijn van der Velde. Crop identification using deep learning on lucas crop cover photos. Sensors, 23(14), 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 362 + ], + "type": "text", + "content": "[55] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Ren-liang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In CVPR, pages 9556-9567, 2024. 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 429 + ], + "type": "text", + "content": "[56] Xiang Yue, Tianyu Zheng, Yuansheng Ni, Yubo Wang, Kai Zhang, Shengbang Tong, Yuxuan Sun, Botao Yu, Ge Zhang, Huan Sun, Yu Su, Wenhu Chen, and Graham Neubig. Mmmu-pro: A more robust multi-discipline multimodal understanding benchmark. arXiv preprint arXiv:2409.02813, 2024. 2, 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7644" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_content_list.json b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9011e589b0a4eb7f277934624c017367c8b8616 --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_content_list.json @@ -0,0 +1,1389 @@ +[ + { + "type": "text", + "text": "AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference", + "text_level": 1, + "bbox": [ + 117, + 128, + 852, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie\\*, Hao Wang Alibaba Group", + "bbox": [ + 223, + 203, + 743, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhouwan.hk, zh372956, bochen.wbc, yx150449, xiezhen.xz, qiao.wh@alibaba-inc.com", + "bbox": [ + 133, + 241, + 834, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. We introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only $10\\%$ of visual KV cache, thereby reducing decoding latency by $29\\%$ to $66\\%$ across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches.", + "bbox": [ + 75, + 321, + 472, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 805, + 209, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past few years have witnessed remarkable advancements in Large Visual-Language Models (LVLMs) [1, 6, 16, 22, 23, 32, 35, 36], both in research and prac", + "bbox": [ + 76, + 830, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a467d1151f33fa4cc67bc19015d697066262c2a76b64e4fdc190bf02c224b4ea.jpg", + "image_caption": [ + "Figure 1. Illustration of our motivation. (a) An example of attention maps during the decoding process of LLaVA-OV-7B [16]. (b) Partial visual attention maps obtained using all text tokens as the observation window. (c) Using the last 16 text tokens as the observation window. (d) Using the last 16 visual tokens as the observation window, demonstrating more consistent attention distribution." + ], + "image_footnote": [], + "bbox": [ + 501, + 289, + 686, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/47b792cdd7f3154c8287184933cc516431f3b598ea0807c48f077f6b9332380e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 289, + 890, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tical applications. Although LVLMs exhibit increasingly advanced visual processing capabilities, such as handling high-resolution, multiple images, and video sequences, these improvements entail substantial computational costs. The exponential increase in visual tokens, coupled with the demands of long-context generation tasks, results in unsustainable memory overhead from key-value (KV) cache storage [30]. This not only significantly increases GPU memory consumption but also severely degrades computational efficiency due to heightened memory bandwidth contention.", + "bbox": [ + 496, + 563, + 890, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches [5, 13, 14, 19, 39, 42] aim to prune visual tokens during inference, which can be categorized into token pruning and KV cache compression. Token pruning [11, 13, 14, 18, 20, 39] involves reducing visual tokens during the prefetch phase, which decreases both the number of tokens processed in subsequent transformer layers and the tokens stored in the KV cache. Although it significantly enhances inference speed, the aggressive elimination of visual tokens during prefetch leads to a substantial loss of visual information, severely degrading model performance. KV cache compression [19, 21, 25, 34, 42] reduces token solely during the decoding phase by pruning stored data in", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 886, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23958", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the KV cache. This technique has been extensively validated in LLMs. Since all tokens undergo a complete forward pass and the causal attention mechanism in LLMs differentiates the importance distribution among tokens, selectively deleting certain tokens has minimal impact on model performance. These methods typically assess the importance of visual tokens by evaluating their strength or sparsity within the attention maps computed under the causal attention mechanism. Given the strict unidirectionality of causal attention, using cross-modality attention scores between instruction tokens and visual tokens provides a more comprehensive evaluation of visual token importance than using full attention scores. Moreover, we observe significant variations in the distribution of visual token strength across different layers, indicating that uniformly allocating the budget is not the optimal strategy.", + "bbox": [ + 80, + 90, + 467, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose AirCache, a novel KV cache compression method designed for efficient LVLMs inference. AirCache comprises two main components: visual token importance scoring and layerwise KV cache budget allocation. To effectively evaluate the importance of visual tokens in the KV cache, we introduce an elite observation window that selects critical instruction tokens by leveraging self-attention scores among instruction tokens rather than using all or continuous local instruction tokens. As illustrated in Figure 1, compared to mainstream methods that assess visual token importance based on observation windows guided by all or partial text instruction tokens, the elite observation window demonstrates stronger consistency. Most text tokens within the elite observation window tend to provide more similar evaluations for the same visual token, thereby enhancing the effectiveness and stability of subsequent voting-based ranking. We further propose to quantify the compression budget for different layers from two perspectives. The first aspect is the emphasis a layer places on visual information, measured by the sum of attention allocated to all visual tokens, as well as the strength of importance score distribution. Additionally, analysis of attention distributions across layers reveals a distinct head effect, where a few visual tokens receive high importance scores, while the majority exhibit mediocre and average importance characteristics. Our experimental results show that preserving only $10\\%$ of the important visual KV cache results in less than $1\\%$ average model performance degradation, corroborating the significant head effect in the importance distribution of visual tokens. Combining these two aspects, we allocate the budget for visual KV cache based on both the strength and skewness of importance scores distributions.", + "bbox": [ + 80, + 337, + 467, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our primary contributions include: (1) We conduct a comprehensive and detailed exploration of the differences in inter-modal relevancy patterns based on attention interaction in evaluating visual KV cache importance. We in", + "bbox": [ + 80, + 840, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "troduce an elite observation window with carefully selected key textural instruction tokens to achieve more effective and stable importance evaluations. (2) We propose to quantify the differentiated compression budget allocation across different layers using the strength and skewness of the importance scores distribution, based on the prominence of visual information and the efficacy of attention allocation. (3) Experimental results on widely used LVLMs and benchmark datasets demonstrate that the proposed method achieves superior performance compared to other existing methods.", + "bbox": [ + 503, + 90, + 888, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 256, + 638, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Token Pruning. Visual token pruning [4, 5, 12, 13, 20, 39] accelerates model inference by reducing visual tokens during the prefill phase. BLIP-2 [17] and Qwen-VL [1] employ cross-attention with learnable query embeddings to aggregate input visual tokens, while HoneyBee [4] and MobileVLM [7] use convolutional neural networks. These approaches necessitate model retraining, which significantly increases migration costs. In contrast, methods like [5, 39] introduce no extra parameters but leverage attention maps to score and evict visual tokens based on importance. IVTP [13] implements two-stage pruning by deleting tokens in both Vision Transformer (ViT) and LLM, enabling text-guided pruning through instruction integration. However, it is important to note that visual token pruning inevitably leads to visual information loss by discarding tokens prior to or within early layers, thereby substantially degrading model performance.", + "bbox": [ + 503, + 281, + 888, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "KV cache compression. KV cache strategies accelerate inference by storing the key and value states of previous tokens, thereby avoiding redundant computations during decoding. However, as the number of tokens increases, the pressure on memory and bandwidth also rises, limiting the model's application in long content understanding or generation. To address this challenge, mainstream methods primarily focus on two directions: intra-layer KV cache pruning [15, 29, 40] and inter-layer compression budget allocation [21, 38]. H2O [42] and SnapKV [19] propose evicting tokens in the KV cache using global and local cumulative attention scores, while PyramidKV [3] constructs a pyramid-like hierarchical budget allocation to achieve more refined KV cache compression management. The input to LVLMs is a cross-modal construct primarily composed of visual tokens with auxiliary text tokens, necessitating additional consideration of the impact of different modality differences when extending KV cache compression to LVLMs. Elastic [25] merges non-critical tokens with important tokens, PrefixKV [34] introduces an adaptive layer-wise KV retention recipe based on binary search for maximal preservation of contextual information. VL-Cache [33] utilizes visual token sparsity for layer-wise budget allocation and token eviction through dedicated scoring policies.", + "bbox": [ + 503, + 539, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23959", + "bbox": [ + 478, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d04ec30050357743f6fe7d5491a37dd44d7667b66434e29671518682c6845f27.jpg", + "image_caption": [ + "Figure 2. The left panel illustrates the inference workflow of standard LVLMs. The proposed method integrates during the KV cache storage process following the prefetch stage, maintaining compatibility with mainstream LVLMs architecture. The right panel presents the overview of AirCache. We first employ self-attention across the instruction text tokens to identify the key text tokens, establishing an elite observation window for assessing visual token significance. We further reallocate the compression budgets between layers based on the strength and skewness of the importance distribution. Finally, pruning is applied to the KV cache according to the importance ranking of visual tokens in the KV Cache and the compression budget to obtain the final compressed KV cache." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 282, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/85ae4793679c26c7c20cf9eff09e7219b954eec717ea08769c7320fda8a0cb43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 433, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee244de9c3edf0a2bba9c6b1f91b7beeb1466967113cd9d87b271b7410bf3e39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 88, + 720, + 258 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4a3d2980077e8266a1e24e7c36890548bd4d3c7be20bcb53ec7916cf964c7611.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 88, + 890, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 358, + 210, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 382, + 207, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to most LLMs, the inference process of LVLMs is primarily divided into prompt prefill and token decoding. Prompt Prefill. Given the hidden states $X \\in \\mathbb{R}^{N \\times D}$ of the input prompt, where $N$ is the total number of visual and text prompt tokens, and $D$ denotes hidden dimension. We simplify the expression by omitting the indices of the head and layer of the hidden states. The query, key, and value states of the attention block can be represented as:", + "bbox": [ + 75, + 406, + 470, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {Q}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {K}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {V}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 532, + 468, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V \\in \\mathbb{R}^{D \\times D}$ are learnable projection matrices. Following the completion of the attention interaction, the key and value are stored in the KV cache to support subsequent token generation.", + "bbox": [ + 75, + 554, + 468, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token Decoding. In this stage, the model efficiently generates new tokens by iteratively utilizing and updating the KV cache. Specifically, during the $i$ th iteration, we only need to compute the key and value associated with the newly generated token $\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}$ . By employing cache indexing, the model avoids recomputing the key and value for tokens that have already been processed. Subsequently, the KV cache is updated with the newly computed key and value for the latest token:", + "bbox": [ + 75, + 616, + 468, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = \\operatorname {C o n c a t} \\left(\\mathbf {K}, \\mathbf {x} _ {i} \\mathbf {W} _ {K}\\right), \\mathbf {V} = \\operatorname {C o n c a t} \\left(\\mathbf {V}, \\mathbf {x} _ {i} \\mathbf {W} _ {V}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 758, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although the KV cache mitigates the issue of redundant computation during the decoding process, its characteristic of linear growth in computational requirements relative to the length of the input sequence poses challenges regarding latency and memory usage, especially in scenarios involving longer inputs or outputs. This issue is further exacerbated as recent LVLMs tend to prioritize high-resolution, high-frame-rate visual inputs.", + "bbox": [ + 75, + 780, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 359, + 609, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The AirCache is primarily divided into two components. The first component evaluates the significance of visual tokens based on the elite observation window. Section 3.3 details how to optimize the relevancy between visual and language modalities to guide the compression of the visual segment in the KV cache. The second component quantifies the KV cache budget requirements across various layers by analyzing the overall strength and skewness of the importance score distribution, thereby facilitating dynamic layerwise KV cache budget allocation, as elaborated in Section 3.4. The architecture and algorithm underlying the method are illustrated in Figure 2 and Algorithm 1.", + "bbox": [ + 496, + 382, + 890, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Elite Observation Window", + "text_level": 1, + "bbox": [ + 500, + 573, + 740, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In LVLMs, the presence of multiple modalities complicates the selection of observation windows, suggesting that adhering to the relevant practices established in LLMs may not be optimal. For instance, as illustrated in Figure 1, the visual token importance measures derived from the continuous last segment of the prompt exhibit poor consistency. This variability indicates that different text tokens within the observation window yield significantly diverse visual tokens importance distributions. Consequently, the hit rate statistics derived from the voting mechanism are prone to substantial noise, and the disparities among modalities may further amplify this noise interference. Based on this consideration, we propose refining the observation windows by incorporating carefully selected text tokens. By utilizing the attention hit rate within the text modality, we aim to enhance the consistency of visual token evaluation, thereby improving the performance of KV cache compression. To this end, we reorganize the hidden states of the input prompt by:", + "bbox": [ + 496, + 597, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} = \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {v}, \\mathbf {X} _ {t}\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t}\\right) \\times D}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 883, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23960", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92ec940688a6a44686f2e7900d795bace6e62108207312c001632605d6c14ed2.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 138, + 87, + 380, + 227 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4ee76410aaa113aec42e8a0152bcaa354c004f2cf32332b0159a0b85871e011a.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) The strength of the distribution of visual token importance scores across different layers, which is defined as the sum of all visual token importance scores. (b) Performance with selecting only one visual token each time based on the sorted results of visual token importance scores, the index of 0 indicates the setting of removing all visual tokens. (c) The distribution of visual token importance scores across different layers, with only selected layers displayed for clarity. All results are derived from LLaVA-OV-7B [16] on the ChatQA [26]." + ], + "image_footnote": [], + "bbox": [ + 380, + 88, + 609, + 227 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e20ca144141a19854a637bca3fc3640d2a7b0c2a106349a95f6a1de0bcdbc7a7.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 611, + 88, + 828, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_v$ and $\\mathbf{X}_t$ represent the hidden states corresponding to visual tokens and instruction text tokens, respectively. $N_v$ and $N_t$ denote the respective quantities of these tokens, satisfying the condition $N_v + N_t = N$ . For convenience, the system prompts are omitted. We compute the text unimodal attention matrix composed of instruction text tokens as:", + "bbox": [ + 75, + 313, + 467, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {t t} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\mathbf {K} _ {t} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t} \\times N _ {t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 411, + 468, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_t$ and $\\mathbf{K}_t$ denote the query and key of the text tokens, respectively. Considering the characteristics of causal attention, we select the last token of the instruction text as reference and filter out key text tokens that receive high attention scores based on their interactions with other text tokens. Consistent with [42], we employ a relevance threshold to regulate the sparsity of key text tokens:", + "bbox": [ + 75, + 450, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk = \\{j \\mid \\mathbf {A} _ {t t} [ N _ {t} - 1, j ] \\geq \\alpha \\cdot \\max \\mathbf {A} _ {t t} [ N _ {t} - 1,: ] \\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 564, + 468, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha \\in [0,1]$ represents the relevance threshold, and $j\\in \\{0,1,\\dots ,N_t - 1\\}$ is the index of the text token. When $\\alpha$ is equal to 0, all text tokens are retained, and when $\\alpha$ is equal to 1, only the text token with the highest score is preserved. We redefine the key for the aggregated visual tokens and key text tokens as $\\mathbf{K}_{vt_k}$ , along with the query corresponding to the key text tokens $\\mathbf{Q}_{t_k}$ , as follows:", + "bbox": [ + 75, + 588, + 468, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t _ {k}} = \\mathbf {Q} _ {t} [ k ] \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times D}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 700, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {v t k} = \\operatorname {C o n c a t} \\left(\\mathbf {K} _ {v}, \\mathbf {K} _ {t} [ k ]\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t _ {k}}\\right) \\times D},\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 718, + 416, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{K}_v$ represents the key of the visual token, and $N_{t_k}$ denotes the number of selected key text tokens. Furthermore, the attention matrix between the visual tokens and the key text tokens can be computed as follows:", + "bbox": [ + 75, + 741, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {v t _ {k}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t _ {k}} \\mathbf {K} _ {v t _ {k}} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times \\left(N _ {v} + N _ {t _ {k}}\\right)}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 808, + 468, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The elite observation window, comprising only these key text tokens, offers a more comprehensive and focused semantic representation compared to those formed by all text", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tokens or fixed continuous sequences. It provides enhanced stability and accuracy in assessing the importance of visual tokens. The final importance scores of the visual tokens can be obtained by performing average pooling of the attention matrix along the text dimension:", + "bbox": [ + 498, + 313, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} _ {v} = \\frac {1}{N _ {t _ {k}}} \\sum_ {j = 0} ^ {N _ {t _ {k}} - 1} \\mathbf {A} _ {v t _ {k}} [ j,: N _ {v} ]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 401, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By sorting the importance scores, the KV cache of visual tokens with lower rankings can be discarded in accordance with the compression budget. Furthermore, the elite observation window we employ exhibits reduced computational complexity, and the additional overhead of key text token selection is negligible in visual token-heavy LVLMs.", + "bbox": [ + 498, + 458, + 890, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Layer-wise KV Cache Budget Allocation", + "text_level": 1, + "bbox": [ + 500, + 559, + 846, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further investigate the characteristics of the visual token importance score distribution based on the elite observation window across various layers. Specifically, we consider the sum of the importance scores of all visual tokens as an indicator of the layer's attention to visual information. As illustrated in Figure 3(a), there is a significant disparity in the attention given to visual tokens across different layers. This observation highlights the necessity of performing layer-wise token budget allocation, where layers dedicating more attention to visual information merit a larger token budget. As shown in Figure 3(b), we select one visual token at a time based on the importance scores. It can be observed that only approximately $10\\%$ of all visual tokens have a significantly positive impact on the final result. This suggests a head effect in the model's attention distribution over visual tokens, with a small subset receiving higher attention, while the majority have lower and more balanced scores. Extending to the importance distribution across different layers in Figure 3(c), it shows significant differences in importance distribution among the layers, particularly regarding the head effect of visual tokens. Consequently,", + "bbox": [ + 496, + 583, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23961", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we aim to allocate a larger token budget to the layers that demonstrate a pronounced head effect in their visual tokens importance distribution. This indicates a more accurate and nuanced understanding of the visual content by those layers.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As aforementioned, we quantify the allocation of visual token budget across different layers from two perspectives: the strength and skewness of the importance distribution. The strength of the importance distribution is calculated as:", + "bbox": [ + 75, + 152, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} = \\sum_ {i = 0} ^ {N _ {v} - 1} \\mathbf {I} _ {v} [ i ]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 222, + 468, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We exclude the attention between text tokens and instead aggregate the attention scores between text tokens and all visual tokens to quantify the importance of visual information in each layer. The greater the importance, the more budget is allocated. Meanwhile, we use the skewness of importance distribution to assess how effectively each layer aligns with the distribution characteristics of attention to visual tokens:", + "bbox": [ + 75, + 268, + 470, + 387 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {k} = \\frac {N _ {v}}{\\left(N _ {v} - 1\\right) \\left(N _ {v} - 2\\right)} \\sum_ {i = 1} ^ {N _ {v}} \\left(\\frac {\\mathbf {I} _ {v} [ i ] - \\mu_ {I _ {v}}}{\\sigma_ {I _ {v}}}\\right) ^ {3}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 395, + 468, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{I_v}$ and $\\sigma_{I_v}$ represent the mean and standard deviation of the importance as well as attention scores distribution of visual tokens, which essentially reflect the distribution of attention scores. Ultimately, the reallocated token budget integrates these factors into the original budget $r$ :", + "bbox": [ + 75, + 444, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {r} = \\frac {1}{2} \\left(s _ {t} ^ {\\prime} + s _ {k} ^ {\\prime}\\right) r. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 518, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s_t'$ and $s_k'$ represent the normalized strength and skewness across different layers, respectively.", + "bbox": [ + 75, + 551, + 468, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 595, + 207, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 76, + 619, + 158, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. The selection of LVLMs is primarily based on two criteria: model architecture and parameter size. In terms of architecture, the objective is to validate the effectiveness of the proposed method across various architectures, including LLaVA-OV-7B [16], InternVL2-8B [6], and Qwen2-VL-7B [35]. These models exhibit significant differences in visual token encoding, ViT, and LLM. For example, both LLaVA-OV-7B [16] and InternVL2-8B [6] utilize the AnyRes strategy with distinct ViT to increase the number of visual tokens, while Qwen2-VL-7B [35] supports native resolution with a large quantity of visual tokens. In terms of parameter size, models of InternVL2 series are selected for their diverse range of parameter scales, including InternVL2-1B [6], InternVL2-4B [6], and InternVL2-26B [6]. Due to space constraints, results for this section are provided in the appendix. Additionally, LLaVA-Video-7B-Qwen2 [41] serves as the base model for the video tasks evaluation.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 AirCache for KV Cache Compression." + ], + "code_body": "1: Input: Total cache budget $r$ , key, query, value of text prompt tokens $K_{t}$ , $Q_{t}$ , $V_{t} \\in \\mathbb{R}^{N_{t} \\times D}$ , key, query, value of visual prompt tokens $K_{v}$ , $Q_{v}$ , $V_{v} \\in \\mathbb{R}^{N_{v} \\times D}$ , number of layers L. \n2: for layer $l \\gets 0$ to $N - 1$ do \n3: $A_{tt} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t} \\mathbf{K}_{t}^{T}}{\\sqrt{D}}\\right)$ \n4: Obtain the index of key text tokens $k$ using Eq. (5). \n5: $Q_{t_{k}}$ , $K_{vt_{k}} \\gets Q_{t}[k]$ , $\\text{Concat}(K_{v}, K_{t}[k])$ \n6: $\\mathbf{A}_{vt_{k}} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t_{k}} \\mathbf{K}_{vt_{k}}^{T}}{\\sqrt{D}}\\right)$ \n7: $\\mathbf{I}_{v} \\gets \\frac{1}{N_{t_{k}}} \\sum_{j=0}^{N_{t_{k}}-1} \\mathbf{A}_{vt_{k}}[j,: N_{v}] \\triangleright$ Importance Scores \n8: $s_{t} \\gets \\sum_{i=0}^{N_{v}-1} \\mathbf{I}_{v}[i]$ \n9: $s_{k} \\gets \\frac{N_{v}}{(N_{v}-1)(N_{v}-2)} \\sum_{i=1}^{N_{v}}\\left(\\frac{\\mathbf{I}_{v}[i]-\\mu_{I_{v}}}{\\sigma_{I_{v}}}\\right)^{3} \\triangleright$ Skewness \n10: $C^{l} \\gets (\\{(K_{v}, K_{t}), (V_{v}, V_{t}), (\\mathbf{I}_{v}, s_{t}, s_{k})\\})$ \n11: end for \n12: for layer $l \\gets 0$ to $N - 1$ do \n13: $K_{v}, K_{t}, V_{v}, V_{t}, \\mathbf{I}_{v}, s_{t}, s_{k} \\gets C^{l}$ \n14: $s_{t}', s_{k}' \\gets s_{t}, s_{k} \\triangleright$ Normalized Across Different Layers \n15: $\\hat{r} = \\frac{1}{2}(s_{t}' + s_{k}') r \\quad \\triangleright$ Obtain Layer-wise Budget \n16: $K' \\gets \\text{Concat}(K_{v}(\\mathbf{I}[N_{v} \\times \\hat{r}]), K_{t}) \\quad \\triangleright$ Compress Key \n17: $V' \\gets \\text{Concat}(V_{v}(\\mathbf{I}[V_{v} \\times \\hat{r}]), V_{t}) \\quad \\triangleright$ Compress Value \n18: $C^{l} \\gets (\\{K', V'\\})$ \n19: end for \n20: Return: $C \\triangleright$ Compressed KV Cache", + "bbox": [ + 504, + 107, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. The performance is primarily evaluated across four types of datasets: 1) Visual Question Answering (VQA) datasets, which include ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31]. Since KV cache compression mainly affects the decoding stage, VQA datasets that require longer output tokens are preferred for evaluating the method's efficacy. The similarity between generated and reference answers is typically measured by the Average Normalized Levenshtein Similarity (ANLS) metric [2]. 2) Multiple-choice question datasets, such as MMBench-EN [24], involves selecting the correct option from several choices, with accuracy serving as the ultimate metric. 3) Judgment datasets like MME [10], also utilize judgment accuracy as the evaluation criterion. 4) Video VQA datasets, such as MMBench-Video [9], measure the alignment between generated and reference answers using metrics provided by a large model evaluator.", + "bbox": [ + 496, + 491, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. The comparison focuses on two main categories of methods. The first category encompasses classic and prominent KV cache compression techniques in LLMs, such as H2O [42], and SnapKV [19]. The second category comprises methods specifically optimized for LVLMs, namely PrefixKV [34] and Elastic Cache [25]. H2O [42] employs a full-range approach, while SnapKV [19] adopts a window-based strategy using textual tokens as the observation window for evaluating token importance. PrefixKV [34] utilizes the Lorenz curve to characterize the impor", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23962", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/de445ad5050c47fbf0f5c463223655330d1d295cf7fddf2eea61b2c40b3f047c.jpg", + "table_caption": [ + "Table 1. The comparison of the KV cache compression methods on multimodal VQA benchmarks. The best result is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
ModelsMethodsChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
50%10%5%1%50%10%5%1%50%10%5%1%50%10%5%1%
LLaVA-OV-7B [16]Full80.380.380.380.366.166.166.166.187.087.087.087.076.076.076.076.0
H2O [42]79.877.474.071.065.459.254.252.086.474.260.855.375.570.167.960.4
Elastic [25]80.077.976.571.065.360.557.252.086.674.260.855.375.572.068.760.2
PrefixKV [34]80.178.276.870.965.461.158.051.986.680.574.755.475.872.770.060.2
SnapKV [19]80.179.378.372.966.164.263.057.886.984.480.664.175.873.469.558.2
Ours80.479.979.476.466.065.764.662.586.885.583.373.276.075.373.467.1
InternVL2-8B [6]Full82.282.282.282.273.073.073.073.091.091.091.091.077.777.777.777.7
H2O [42]81.981.679.376.872.370.665.161.290.579.470.348.977.170.066.059.6
Elastic [25]81.881.079.476.872.367.065.160.890.579.470.353.577.070.266.059.0
PrefixKV [34]81.980.779.376.772.367.465.361.090.679.570.255.177.170.366.258.6
SnapKV [19]82.280.479.575.072.872.368.464.990.990.183.571.177.473.971.061.7
Ours82.281.781.078.473.272.671.868.291.090.088.578.577.577.075.468.6
Qwen2-VL-7B [35]Full83.283.283.283.276.276.276.276.293.893.893.893.884.484.484.484.4
H2O [42]83.278.975.066.475.465.761.055.190.582.370.748.984.275.668.254.5
Elastic [25]83.279.775.468.075.567.162.355.593.583.071.949.284.276.768.854.4
PrefixKV [34]83.279.475.968.075.767.962.755.793.783.772.849.884.276.969.454.3
SnapKV [19]83.381.680.069.176.374.973.066.293.787.287.171.284.383.174.660.2
Ours83.482.380.875.976.275.274.570.793.892.990.978.384.483.480.968.6
", + "bbox": [ + 80, + 101, + 888, + 340 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/67059182a9c3795f7c041d3c6fb5af57a8f1be7515d9c9e53eeb49e8919b85b7.jpg", + "table_caption": [ + "Table 2. Performance comparison on other types datasets. Best-performing results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodsMMBench-EN [24]MME [10]MMBench-Video [9]
50%10%1%50%10%1%50%10%1%
Full82.382.382.31585158515851.811.811.81
H2O [42]0.820.820.821585158515851.661.631.47
Elastic [25]0.820.820.821585158515851.681.621.51
PrefixKV [34]0.820.820.821585158515851.721.681.50
SnapKV [19]0.820.820.821585158515851.751.701.55
Ours0.820.820.821585158515851.801.781.67
", + "bbox": [ + 80, + 377, + 465, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tance distribution across different layers, guided by the retained important KV matrix. Elastic Cache [25] employs a fixed-point elimination strategy for the token importance with cache merging. All KV cache compression methods are applied exclusively to the visual tokens, as experiments have demonstrated that it offers greater stability and better performance compared to compressing the entire tokens.", + "bbox": [ + 75, + 474, + 467, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. All evaluation results are derived from VLMEvalKit [8], which is an open-source evaluation toolkit for LVLMs. The relevance threshold in Equation 5 is set to 0.9. The referee model for MMBench-Video [9] is Qwen2.5-72B [37]. The proposed method performs a one-time KV cache compression only after the prefetch stage is completed. Specifically, once the prefetch ends for each layer, the complete KV cache is fully saved. After computing all layers, we then calculate the budget for each layer and compress the KV cache accordingly. All experiments are conducted on $8 \\times \\mathrm{A}100 - 80\\mathrm{G}$ GPUs.", + "bbox": [ + 75, + 580, + 467, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 76, + 756, + 215, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report the results in four sections: VQA datasets, other types of datasets including multiple-choice questions, judgment, and video VQA, inference performance, and the comparison with token pruning methods.", + "bbox": [ + 75, + 779, + 467, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VQA datasets. Table 1 presents the comparative results on the ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31] datasets. These datasets encompass tasks such as conventional image question answering, image in-", + "bbox": [ + 75, + 840, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "formation extraction, document image understanding, and text recognition within images. It can be observed that the proposed method achieves the best results under most reduction ratios. When $50\\%$ of visual KV cache is pruned, our method performs almost equivalently to the full KV cache, indicating near lossless performance. Even when reduced to only $10\\%$ of visual KV cache, the performance gap with the full cache is limited to approximately $1\\%$ . As the token compression ratio increases, the advantage of the proposed method becomes more pronounced. For instance, when only $1\\%$ of visual tokens are retained, we outperform SnapKV [19] by an average of $6.5\\%$ across four VQA datasets with the LLaVA-OV-7B [16]. Moreover, the performance differences across different types of VQA datasets indicate that image documents or text understanding, which require higher local visual perception, have stricter demands on KV cache compression than general image understanding. The proposed method demonstrates a more significant advantage over other methods with these datasets, providing strong evidence of its superiority. Furthermore, the above conclusions also apply to different LVLMs, we achieve a significant advantage with InternVL2-8B [6] and Qwen2-VL-7B [35].", + "bbox": [ + 496, + 354, + 890, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other types datasets. Table 2 presents the results of various methods on other types of multimodal datasets. For datasets involving multiple-choice questions and judgment types, such as MMBench-EN [24] and MME [10], the final result heavily depends on the accuracy of decoding the first token. Since the KV cache only influences the tokens decoded after the initial token, all methods achieve the same performance as the full KV cache for these types of datasets. The results from the MMBench-Video [9] dataset indicate that our proposed method achieves superior performance, further demonstrating its generality and superiority.", + "bbox": [ + 496, + 702, + 890, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference efficiency. We maintain a constant output token count of 512 and compare the full cache with the proposed", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23963", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69be38c234138381b15646a272f71b64cb8609a775fef255c41d713bcb88f9fa.jpg", + "table_caption": [ + "Table 3. Quantitative results on inference latency and throughput. The number of tokens output is consistently set to 512." + ], + "table_footnote": [], + "table_body": "
Batch SizePrompt LengthPrefill Latency (s)Decoding Latency (s)Throughput (token/s)
FullOursFull50%10%Full50%10%
82k1.61.8+12.5%11.69.4-19.0%8.2-29.3%353436+23.5%500+41.6%
8k4.95.3+8.2%15.011.9-20.7%9.4-37.3%273344+26.0%436+59.7%
16k11.311.9+5.3%21.715.0-30.9%10.4-47.5%189273+44.4%394+108.5%
32k25.226.4+4.8%35.922.0-38.7%12.3-65.7%114186+63.2%333+192.1%
162k2.32.5+8.7%14.411.2-22.2%9.7-32.6%569731+28.5%845+48.5%
8k9.49.8+4.3%22.615.6-31.0%11.8-47.8%362525+45.0%694+91.7%
16k20.321.4+5.4%36.923.0-37.7%12.8-65.3%222356+60.4%640+188.3%
", + "bbox": [ + 120, + 99, + 846, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c56ab96b1ed2502d548accd666bcd63d7b9f2ef120bab02e0dfb50db6573ecd7.jpg", + "table_caption": [ + "Table 4. Performance comparison with token pruning methods under various compression ratio. The prefetch and decoding latency are measured under the settings of batch size of 16, total input token of 8k, and the compression ratio of $10\\%$ ." + ], + "table_footnote": [], + "table_body": "
MethodsChatQA [26]MMBench-EN [24]MME [10]MMBench-Video [9]Latency (s)
50%10%1%50%10%1%50%10%1%50%10%1%PrefillDecoding
Full80.380.380.382.382.382.31585158515851.811.811.819.422.6
FastV [5]73.647.716.981.474.833.6157413787861.691.361.025.411.7
FasterVLM [39]75.550.118.881.477.026.5155914417311.711.421.175.211.3
IVTP [13]74.955.822.581.777.536.2156314698491.701.471.245.812.6
Ours80.479.976.482.382.382.31585158515851.791.721.609.811.8
", + "bbox": [ + 81, + 267, + 888, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AirCache across different batch sizes and input token counts in terms of prefetch latency, decoding latency, and throughput. To facilitate testing, the text instruction token count is set to 64. As shown in Table 3, the proposed method significantly improves inference decoding speed under various inputs compared to the full cache. When compressing $50\\%$ of the visual KV cache, the proposed method minimally impacts model performance and prefetch latency, while reducing decoding latency by $19\\%$ to $39\\%$ and increasing model throughput by $24\\%$ to $63\\%$ . When compressing $10\\%$ of the visual KV cache, the proposed method slightly decreases model performance by less than $1\\%$ , while further reducing decoding latency by $29\\%$ to $66\\%$ and increasing model throughput by $42\\%$ to $192\\%$ . These results demonstrate that our proposed method can maintain strong model performance while ensuring significant decoding acceleration, indicating the effectiveness of visual token importance assessment and hierarchical budget allocation.", + "bbox": [ + 75, + 383, + 472, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared with token pruning methods. In addition to KV cache compression, another important method for accelerating LVLMs inference is to prune visual tokens before the prefetch stage. Table 4 compares the proposed method with these type of acceleration methods, including FastV [5], FasterVLM [39], and IVTP [13]. A notable characteristic is that the token pruning methods exhibit acceleration effects during the prefetch stage as they preemptively prune visual tokens. In contrast, all KV cache compression methods, including the proposed approach, demonstrate prefetch stage inference speeds comparable to those of the full cache. However, comparisons across various evaluation sets clearly show that the proposed method achieves an absolute advantage, particularly on VQA datasets, even when retaining only $1\\%$ of the visual tokens. In contrast, token pruning methods experience a significant performance", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74c4ae42c0e71a4641f6fdd8b5511b64cb9472d3090393b50fe13841f2c6de57.jpg", + "table_caption": [ + "Table 5. Ablation results on the components of the AirCache." + ], + "table_footnote": [], + "table_body": "
SettingChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
Continuous window (16)70.456.661.355.9
Continuous window (32)72.957.864.158.2
All text tokens72.258.465.757.0
Visual window (32)68.855.159.253.7
Average allocation72.257.569.962.4
Pyramid allocation69.654.955.852.6
Only w/ st74.259.871.164.9
Only w/ sk74.761.471.963.6
Ours76.462.573.267.1
", + "bbox": [ + 503, + 393, + 888, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "decline under these conditions. The prefetch stage facilitates sufficient interaction between visual and textual information, as the attention mechanism aggregates visual information to key text tokens. The proposed method effectively preserves these key text tokens during KV cache token pruning. Consequently, even when only $1\\%$ of visual tokens are retained, the model maintains good inference performance. In contrast, most token pruning methods do not involve cross-modal interaction, resulting in direct visual information loss that cannot be transferred to the subsequent decoding stage in any form. This leads to insufficient and inaccurate visual guidance for generating, significantly degrading performance.", + "bbox": [ + 496, + 515, + 890, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 724, + 663, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we evaluate our method under various settings to verify the proposed modules from several perspectives. The base model is LLaVA-OV-7B [16].", + "bbox": [ + 496, + 750, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Elite Observation Window. We introduce three common types of observation windows to assess the effectiveness of the proposed elite observation window. The first type employs a sequence of continuous text tokens as the observation window for evaluating the significance of visual tokens. The second type utilizes all text tokens, whereas the third employs the continuous visual token from the last visual", + "bbox": [ + 496, + 795, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23964", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/81388583dc675603e038cc565e5f077bb61a7cfe0dbe4b6bb9238d87c9a0adbc.jpg", + "table_caption": [ + "Table 6. Comparison results with different compression audiences, where * indicates uniform compression without distinguishing between visual tokens and text tokens." + ], + "table_footnote": [], + "table_body": "
Models50%10%5%1%
H2O* [42]74.6/407s46.5/559s32.0/733s27.8/1048s
H2O [42]79.8/323s77.4/251s74.0/239s71.0/215s
Ours*75.8/374s53.4/530s39.7/688s30.1/963s
Ours80.4/306s79.9/235s79.4/221s76.4/207s
", + "bbox": [ + 81, + 126, + 467, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/55cd4e0812dc1f5eebbe52655f686b89d3a96dd919ed450d52b2921ceda4c086.jpg", + "image_caption": [ + "Figure 4. Comparison on ChatQA [26] and TextVQA [31] by retaining only one visual token, which is selected based on the sorting of visual token importance scores using different methods." + ], + "image_footnote": [], + "bbox": [ + 78, + 213, + 468, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "segment as a reference. As shown in the first column of Table 5, the proposed elite observation window yields the best performance compared to the other three aforementioned approaches. The method using only visual tokens performs significantly worse due to the absence of textual instruction guidance, highlighting the importance of activating cross-modal associations in KV cache compression. Additionally, compared to methods that directly use all or portions of continuous text token sequences without considering variations in visual information perception among text tokens, the proposed method selects key text tokens to enhance the consistency of the observation window. This selection ensures greater stability and accuracy in evaluating the significance of visual tokens.", + "bbox": [ + 75, + 398, + 467, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Layer-wise Token Budget Allocation Module. The second column of Table 5 presents the comparison using various budget allocation strategies across layers. Aside from the basic equal allocation, we also incorporate the strategy from PyramidKV [3], which allocates budgets in a descending ratio from shallow to deep layers. The results demonstrate that the proposed budget allocation, based on the strength and skewness of importance scores, delivers superior outcomes. Although PyramidKV [3] perform well in LLMs, its suboptimal results in LVLMs highlight the distinct characteristics of multimodal models and their heightened sensitivity to hierarchical budget allocation. An inappropriate allocation strategy is more likely to adversely affect the final results. We also investigate the separate use of strength and skewness for layer-wise budget quantification, revealing that their combination produces the best results.", + "bbox": [ + 75, + 611, + 467, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visual Token Importance Evaluation. Figure 4 visually demonstrates the comparative effectiveness of various methods in evaluating the importance of visual tokens. To", + "bbox": [ + 76, + 854, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "directly assess the efficacy of these methods in ranking visual token importance, we removed all non-essential modules and selected only one visual token at a time from the ranked sorting results, ordered from high to low importance. Theoretically, as the importance ranking of tokens declines, the model's performance is also expected to decrease. The graph clearly shows that the proposed method aligns more closely with the theoretical trend, where the visual token ranked as TOP1 achieves the highest metric, followed by a continuous and smooth decline. Notably, after around $10\\%$ , the performance of the visual tokens approximates that of the case with all visual tokens removed. This indicates significant redundancy among visual tokens in the KV cache of LVLMs. The results in Table 1, which demonstrate that retaining over $10\\%$ of the visual tokens limits performance decline to no more than $1\\%$ , further corroborate this observation. In contrast, the performance derived from other methods exhibit significant noise, indicating instability and deficiencies in their importance assessment.", + "bbox": [ + 496, + 90, + 890, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unified Compression or Vision-only Compression. Given that the KV cache in LVLMs encompasses both visual and textual modalities, in contrast to the single textual modality of LLMs, we compared Unified Compression, which does not differentiate between modal differences during KV cache compression, with Vision-only Compression, which compresses only visual tokens. As shown in Table 6, the results indicate that unified compression performs significantly worse than vision-only compression in terms of both inference performance and speed. Moreover, as the compression ratio increases, the inference time for unified compression actually continues to rise. The primary reason is that unified compression disrupts the effective representation of textual instructions, leading the model to produce incorrect outputs and exhibit a tendency for repetition, which in turn increases inference time.", + "bbox": [ + 496, + 378, + 890, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 633, + 617, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose Activating Inter-modal Relevancy (AirCache), a novel KV cache compression for efficient LVLMs inference. By leveraging inter-modal relevancy between textual and visual tokens to guide KV cache eviction, we identified significant differences in the attention distribution of visual tokens corresponding to different textual tokens. Based on this finding, we propose an elite observation window with refined key text tokens, which provides a more stable and effective assessment of visual token importance. Furthermore, by exploring the diversity of importance score distribution across different layers, we propose to achieve comprehensive quantification in terms of distribution strength and skewness for adaptive layer-wise compression budget allocation. Extensive experiments across multiple benchmarks and LVLMs demonstrate the effectiveness of AirCache in KV cache compression.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23965", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2", + "[2] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5", + "[3] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information funneling. arXiv preprint arXiv:2406.02069, 2024. 2, 8", + "[4] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2", + "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 1, 2, 7", + "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 5, 6", + "[7] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 2", + "[8] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 6", + "[9] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 5, 6, 7", + "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 5, 6, 7", + "[11] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2", + "[13] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2024. 1, 2, 7", + "[14] Chaoya Jiang, Haiyang Xu, Wei Ye, Qinghao Ye, Chen-liang Li, Ming Yan, Bin Bi, Shikun Zhang, Fei Huang, and Songfang Huang. Bus: Efficient and effective vision-language pre-training with bottom-up patch summarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2900–2910, 2023. 1", + "[15] Jang-Hyun Kim, Junyoung Yeom, Sangwoo Yun, and Hyun Oh Song. Compressed context memory for online language model interaction. arXiv preprint arXiv:2312.03414, 2023. 2", + "[16] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 4, 5, 6, 7", + "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2", + "[18] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1", + "[19] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: Llm knows what you are looking for before generation. arXiv preprint arXiv:2404.14469, 2024. 1, 2, 5, 6", + "[20] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 2", + "[21] Akide Liu, Jing Liu, Zizheng Pan, Yefei He, Reza Haffari, and Bohan Zhuang. Minicache: Kv cache compression in depth dimension for large language models. Advances in Neural Information Processing Systems, 37:139997-140031, 2025. 1, 2", + "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1", + "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23966", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 1, 2, 5, 6", + "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 4, 5, 6, 7, 8", + "[27] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5, 6, 7", + "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5, 6, 7", + "[29] Piotr Nawrot, Adrian Łancucki, Marcin Chochowski, David Tarjan, and Edoardo M Ponti. Dynamic memory compression: Retrofitting llms for accelerated inference. arXiv preprint arXiv:2403.09636, 2024. 2", + "[30] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 1", + "[31] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 7, 8", + "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2025. 1", + "[33] Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, and Panpan Xu. VI-cache: Sparsity and modality-aware kv cache compression for vision-language model inference acceleration. arXiv preprint arXiv:2410.23317, 2024. 2", + "[34] Ao Wang, Hui Chen, Jianchao Tan, Kefeng Zhang, Xunliang Cai, Zijia Lin, Jungong Han, and Guiguang Ding. Prefixkv: Adaptive prefix kv cache is what vision instruction-following models need for efficient generation. arXiv preprint arXiv:2412.03409, 2024. 1, 2, 5, 6", + "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6", + "[36] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6", + "[38] Yifei Yang, Zouying Cao, Qiguang Chen, Libo Qin, Dongjie Yang, Hai Zhao, and Zhi Chen. Kvsharer: Efficient inference via layer-wise dissimilar kv cache sharing. arXiv preprint arXiv:2410.18517, 2024. 2", + "[39] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 1, 2, 7", + "[40] Yuxin Zhang, Yuxuan Du, Gen Luo, Yunshan Zhong, Zhenyu Zhang, Shiwei Liu, and Rongrong Ji. Cam: Cache merging for memory-efficient llms inference. In *Forty-first International Conference on Machine Learning*, 2024. 2", + "[41] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5", + "[42] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 4, 5, 6, 8" + ], + "bbox": [ + 501, + 92, + 890, + 455 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23967", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_model.json b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..798f7742902137b6e3591d30b1b84ddb1c9680c0 --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_model.json @@ -0,0 +1,1881 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.853, + 0.177 + ], + "angle": 0, + "content": "AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.204, + 0.745, + 0.24 + ], + "angle": 0, + "content": "Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie\\*, Hao Wang Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.242, + 0.835, + 0.256 + ], + "angle": 0, + "content": "zhouwan.hk, zh372956, bochen.wbc, yx150449, xiezhen.xz, qiao.wh@alibaba-inc.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.779 + ], + "angle": 0, + "content": "Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. We introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only \\(10\\%\\) of visual KV cache, thereby reducing decoding latency by \\(29\\%\\) to \\(66\\%\\) across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.806, + 0.21, + 0.822 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.879 + ], + "angle": 0, + "content": "The past few years have witnessed remarkable advancements in Large Visual-Language Models (LVLMs) [1, 6, 16, 22, 23, 32, 35, 36], both in research and prac" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.29, + 0.687, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.29, + 0.892, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.446, + 0.895, + 0.544 + ], + "angle": 0, + "content": "Figure 1. Illustration of our motivation. (a) An example of attention maps during the decoding process of LLaVA-OV-7B [16]. (b) Partial visual attention maps obtained using all text tokens as the observation window. (c) Using the last 16 text tokens as the observation window. (d) Using the last 16 visual tokens as the observation window, demonstrating more consistent attention distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.564, + 0.892, + 0.716 + ], + "angle": 0, + "content": "tical applications. Although LVLMs exhibit increasingly advanced visual processing capabilities, such as handling high-resolution, multiple images, and video sequences, these improvements entail substantial computational costs. The exponential increase in visual tokens, coupled with the demands of long-context generation tasks, results in unsustainable memory overhead from key-value (KV) cache storage [30]. This not only significantly increases GPU memory consumption but also severely degrades computational efficiency due to heightened memory bandwidth contention." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Recent approaches [5, 13, 14, 19, 39, 42] aim to prune visual tokens during inference, which can be categorized into token pruning and KV cache compression. Token pruning [11, 13, 14, 18, 20, 39] involves reducing visual tokens during the prefetch phase, which decreases both the number of tokens processed in subsequent transformer layers and the tokens stored in the KV cache. Although it significantly enhances inference speed, the aggressive elimination of visual tokens during prefetch leads to a substantial loss of visual information, severely degrading model performance. KV cache compression [19, 21, 25, 34, 42] reduces token solely during the decoding phase by pruning stored data in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.228, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23958" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.333 + ], + "angle": 0, + "content": "the KV cache. This technique has been extensively validated in LLMs. Since all tokens undergo a complete forward pass and the causal attention mechanism in LLMs differentiates the importance distribution among tokens, selectively deleting certain tokens has minimal impact on model performance. These methods typically assess the importance of visual tokens by evaluating their strength or sparsity within the attention maps computed under the causal attention mechanism. Given the strict unidirectionality of causal attention, using cross-modality attention scores between instruction tokens and visual tokens provides a more comprehensive evaluation of visual token importance than using full attention scores. Moreover, we observe significant variations in the distribution of visual token strength across different layers, indicating that uniformly allocating the budget is not the optimal strategy." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.338, + 0.468, + 0.835 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose AirCache, a novel KV cache compression method designed for efficient LVLMs inference. AirCache comprises two main components: visual token importance scoring and layerwise KV cache budget allocation. To effectively evaluate the importance of visual tokens in the KV cache, we introduce an elite observation window that selects critical instruction tokens by leveraging self-attention scores among instruction tokens rather than using all or continuous local instruction tokens. As illustrated in Figure 1, compared to mainstream methods that assess visual token importance based on observation windows guided by all or partial text instruction tokens, the elite observation window demonstrates stronger consistency. Most text tokens within the elite observation window tend to provide more similar evaluations for the same visual token, thereby enhancing the effectiveness and stability of subsequent voting-based ranking. We further propose to quantify the compression budget for different layers from two perspectives. The first aspect is the emphasis a layer places on visual information, measured by the sum of attention allocated to all visual tokens, as well as the strength of importance score distribution. Additionally, analysis of attention distributions across layers reveals a distinct head effect, where a few visual tokens receive high importance scores, while the majority exhibit mediocre and average importance characteristics. Our experimental results show that preserving only \\(10\\%\\) of the important visual KV cache results in less than \\(1\\%\\) average model performance degradation, corroborating the significant head effect in the importance distribution of visual tokens. Combining these two aspects, we allocate the budget for visual KV cache based on both the strength and skewness of importance scores distributions." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.841, + 0.468, + 0.899 + ], + "angle": 0, + "content": "Our primary contributions include: (1) We conduct a comprehensive and detailed exploration of the differences in inter-modal relevancy patterns based on attention interaction in evaluating visual KV cache importance. We in" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.242 + ], + "angle": 0, + "content": "troduce an elite observation window with carefully selected key textural instruction tokens to achieve more effective and stable importance evaluations. (2) We propose to quantify the differentiated compression budget allocation across different layers using the strength and skewness of the importance scores distribution, based on the prominence of visual information and the efficacy of attention allocation. (3) Experimental results on widely used LVLMs and benchmark datasets demonstrate that the proposed method achieves superior performance compared to other existing methods." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.257, + 0.64, + 0.272 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.282, + 0.89, + 0.537 + ], + "angle": 0, + "content": "Visual Token Pruning. Visual token pruning [4, 5, 12, 13, 20, 39] accelerates model inference by reducing visual tokens during the prefill phase. BLIP-2 [17] and Qwen-VL [1] employ cross-attention with learnable query embeddings to aggregate input visual tokens, while HoneyBee [4] and MobileVLM [7] use convolutional neural networks. These approaches necessitate model retraining, which significantly increases migration costs. In contrast, methods like [5, 39] introduce no extra parameters but leverage attention maps to score and evict visual tokens based on importance. IVTP [13] implements two-stage pruning by deleting tokens in both Vision Transformer (ViT) and LLM, enabling text-guided pruning through instruction integration. However, it is important to note that visual token pruning inevitably leads to visual information loss by discarding tokens prior to or within early layers, thereby substantially degrading model performance." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.54, + 0.89, + 0.9 + ], + "angle": 0, + "content": "KV cache compression. KV cache strategies accelerate inference by storing the key and value states of previous tokens, thereby avoiding redundant computations during decoding. However, as the number of tokens increases, the pressure on memory and bandwidth also rises, limiting the model's application in long content understanding or generation. To address this challenge, mainstream methods primarily focus on two directions: intra-layer KV cache pruning [15, 29, 40] and inter-layer compression budget allocation [21, 38]. H2O [42] and SnapKV [19] propose evicting tokens in the KV cache using global and local cumulative attention scores, while PyramidKV [3] constructs a pyramid-like hierarchical budget allocation to achieve more refined KV cache compression management. The input to LVLMs is a cross-modal construct primarily composed of visual tokens with auxiliary text tokens, necessitating additional consideration of the impact of different modality differences when extending KV cache compression to LVLMs. Elastic [25] merges non-critical tokens with important tokens, PrefixKV [34] introduces an adaptive layer-wise KV retention recipe based on binary search for maximal preservation of contextual information. VL-Cache [33] utilizes visual token sparsity for layer-wise budget allocation and token eviction through dedicated scoring policies." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23959" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.283, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.434, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.089, + 0.721, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.089, + 0.891, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.264, + 0.894, + 0.349 + ], + "angle": 0, + "content": "Figure 2. The left panel illustrates the inference workflow of standard LVLMs. The proposed method integrates during the KV cache storage process following the prefetch stage, maintaining compatibility with mainstream LVLMs architecture. The right panel presents the overview of AirCache. We first employ self-attention across the instruction text tokens to identify the key text tokens, establishing an elite observation window for assessing visual token significance. We further reallocate the compression budgets between layers based on the strength and skewness of the importance distribution. Finally, pruning is applied to the KV cache according to the importance ranking of visual tokens in the KV Cache and the compression budget to obtain the final compressed KV cache." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.359, + 0.212, + 0.376 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.383, + 0.208, + 0.4 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.471, + 0.528 + ], + "angle": 0, + "content": "Similar to most LLMs, the inference process of LVLMs is primarily divided into prompt prefill and token decoding. Prompt Prefill. Given the hidden states \\( X \\in \\mathbb{R}^{N \\times D} \\) of the input prompt, where \\( N \\) is the total number of visual and text prompt tokens, and \\( D \\) denotes hidden dimension. We simplify the expression by omitting the indices of the head and layer of the hidden states. The query, key, and value states of the attention block can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.534, + 0.47, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {Q}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {K}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {V}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.555, + 0.47, + 0.616 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V \\in \\mathbb{R}^{D \\times D}\\) are learnable projection matrices. Following the completion of the attention interaction, the key and value are stored in the KV cache to support subsequent token generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.47, + 0.751 + ], + "angle": 0, + "content": "Token Decoding. In this stage, the model efficiently generates new tokens by iteratively utilizing and updating the KV cache. Specifically, during the \\(i\\)th iteration, we only need to compute the key and value associated with the newly generated token \\(\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}\\). By employing cache indexing, the model avoids recomputing the key and value for tokens that have already been processed. Subsequently, the KV cache is updated with the newly computed key and value for the latest token:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.759, + 0.469, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = \\operatorname {C o n c a t} \\left(\\mathbf {K}, \\mathbf {x} _ {i} \\mathbf {W} _ {K}\\right), \\mathbf {V} = \\operatorname {C o n c a t} \\left(\\mathbf {V}, \\mathbf {x} _ {i} \\mathbf {W} _ {V}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Although the KV cache mitigates the issue of redundant computation during the decoding process, its characteristic of linear growth in computational requirements relative to the length of the input sequence poses challenges regarding latency and memory usage, especially in scenarios involving longer inputs or outputs. This issue is further exacerbated as recent LVLMs tend to prioritize high-resolution, high-frame-rate visual inputs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.36, + 0.611, + 0.374 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.565 + ], + "angle": 0, + "content": "The AirCache is primarily divided into two components. The first component evaluates the significance of visual tokens based on the elite observation window. Section 3.3 details how to optimize the relevancy between visual and language modalities to guide the compression of the visual segment in the KV cache. The second component quantifies the KV cache budget requirements across various layers by analyzing the overall strength and skewness of the importance score distribution, thereby facilitating dynamic layerwise KV cache budget allocation, as elaborated in Section 3.4. The architecture and algorithm underlying the method are illustrated in Figure 2 and Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.741, + 0.589 + ], + "angle": 0, + "content": "3.3. Elite Observation Window" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.871 + ], + "angle": 0, + "content": "In LVLMs, the presence of multiple modalities complicates the selection of observation windows, suggesting that adhering to the relevant practices established in LLMs may not be optimal. For instance, as illustrated in Figure 1, the visual token importance measures derived from the continuous last segment of the prompt exhibit poor consistency. This variability indicates that different text tokens within the observation window yield significantly diverse visual tokens importance distributions. Consequently, the hit rate statistics derived from the voting mechanism are prone to substantial noise, and the disparities among modalities may further amplify this noise interference. Based on this consideration, we propose refining the observation windows by incorporating carefully selected text tokens. By utilizing the attention hit rate within the text modality, we aim to enhance the consistency of visual token evaluation, thereby improving the performance of KV cache compression. To this end, we reorganize the hidden states of the input prompt by:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.884, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} = \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {v}, \\mathbf {X} _ {t}\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t}\\right) \\times D}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23960" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.088, + 0.382, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.23, + 0.269, + 0.242 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.089, + 0.61, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.231, + 0.506, + 0.242 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.089, + 0.83, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.231, + 0.73, + 0.242 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.246, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Figure 3. (a) The strength of the distribution of visual token importance scores across different layers, which is defined as the sum of all visual token importance scores. (b) Performance with selecting only one visual token each time based on the sorted results of visual token importance scores, the index of 0 indicates the setting of removing all visual tokens. (c) The distribution of visual token importance scores across different layers, with only selected layers displayed for clarity. All results are derived from LLaVA-OV-7B [16] on the ChatQA [26]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.468, + 0.404 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_v\\) and \\(\\mathbf{X}_t\\) represent the hidden states corresponding to visual tokens and instruction text tokens, respectively. \\(N_v\\) and \\(N_t\\) denote the respective quantities of these tokens, satisfying the condition \\(N_v + N_t = N\\). For convenience, the system prompts are omitted. We compute the text unimodal attention matrix composed of instruction text tokens as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.412, + 0.469, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {t t} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\mathbf {K} _ {t} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t} \\times N _ {t}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.452, + 0.469, + 0.557 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_t\\) and \\(\\mathbf{K}_t\\) denote the query and key of the text tokens, respectively. Considering the characteristics of causal attention, we select the last token of the instruction text as reference and filter out key text tokens that receive high attention scores based on their interactions with other text tokens. Consistent with [42], we employ a relevance threshold to regulate the sparsity of key text tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.565, + 0.469, + 0.582 + ], + "angle": 0, + "content": "\\[\nk = \\{j \\mid \\mathbf {A} _ {t t} [ N _ {t} - 1, j ] \\geq \\alpha \\cdot \\max \\mathbf {A} _ {t t} [ N _ {t} - 1,: ] \\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.695 + ], + "angle": 0, + "content": "where \\(\\alpha \\in [0,1]\\) represents the relevance threshold, and \\(j\\in \\{0,1,\\dots ,N_t - 1\\}\\) is the index of the text token. When \\(\\alpha\\) is equal to 0, all text tokens are retained, and when \\(\\alpha\\) is equal to 1, only the text token with the highest score is preserved. We redefine the key for the aggregated visual tokens and key text tokens as \\(\\mathbf{K}_{vt_k}\\), along with the query corresponding to the key text tokens \\(\\mathbf{Q}_{t_k}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.702, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t _ {k}} = \\mathbf {Q} _ {t} [ k ] \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times D}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.719, + 0.418, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {v t k} = \\operatorname {C o n c a t} \\left(\\mathbf {K} _ {v}, \\mathbf {K} _ {t} [ k ]\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t _ {k}}\\right) \\times D},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.469, + 0.802 + ], + "angle": 0, + "content": "where \\(\\mathbf{K}_v\\) represents the key of the visual token, and \\(N_{t_k}\\) denotes the number of selected key text tokens. Furthermore, the attention matrix between the visual tokens and the key text tokens can be computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.809, + 0.469, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {v t _ {k}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t _ {k}} \\mathbf {K} _ {v t _ {k}} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times \\left(N _ {v} + N _ {t _ {k}}\\right)}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The elite observation window, comprising only these key text tokens, offers a more comprehensive and focused semantic representation compared to those formed by all text" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.314, + 0.892, + 0.389 + ], + "angle": 0, + "content": "tokens or fixed continuous sequences. It provides enhanced stability and accuracy in assessing the importance of visual tokens. The final importance scores of the visual tokens can be obtained by performing average pooling of the attention matrix along the text dimension:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.402, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} _ {v} = \\frac {1}{N _ {t _ {k}}} \\sum_ {j = 0} ^ {N _ {t _ {k}} - 1} \\mathbf {A} _ {v t _ {k}} [ j,: N _ {v} ]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.459, + 0.892, + 0.549 + ], + "angle": 0, + "content": "By sorting the importance scores, the KV cache of visual tokens with lower rankings can be discarded in accordance with the compression budget. Furthermore, the elite observation window we employ exhibits reduced computational complexity, and the additional overhead of key text token selection is negligible in visual token-heavy LVLMs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.56, + 0.847, + 0.576 + ], + "angle": 0, + "content": "3.4. Layer-wise KV Cache Budget Allocation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We further investigate the characteristics of the visual token importance score distribution based on the elite observation window across various layers. Specifically, we consider the sum of the importance scores of all visual tokens as an indicator of the layer's attention to visual information. As illustrated in Figure 3(a), there is a significant disparity in the attention given to visual tokens across different layers. This observation highlights the necessity of performing layer-wise token budget allocation, where layers dedicating more attention to visual information merit a larger token budget. As shown in Figure 3(b), we select one visual token at a time based on the importance scores. It can be observed that only approximately \\(10\\%\\) of all visual tokens have a significantly positive impact on the final result. This suggests a head effect in the model's attention distribution over visual tokens, with a small subset receiving higher attention, while the majority have lower and more balanced scores. Extending to the importance distribution across different layers in Figure 3(c), it shows significant differences in importance distribution among the layers, particularly regarding the head effect of visual tokens. Consequently," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "we aim to allocate a larger token budget to the layers that demonstrate a pronounced head effect in their visual tokens importance distribution. This indicates a more accurate and nuanced understanding of the visual content by those layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.469, + 0.213 + ], + "angle": 0, + "content": "As aforementioned, we quantify the allocation of visual token budget across different layers from two perspectives: the strength and skewness of the importance distribution. The strength of the importance distribution is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.223, + 0.47, + 0.265 + ], + "angle": 0, + "content": "\\[\ns _ {t} = \\sum_ {i = 0} ^ {N _ {v} - 1} \\mathbf {I} _ {v} [ i ]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.471, + 0.388 + ], + "angle": 0, + "content": "We exclude the attention between text tokens and instead aggregate the attention scores between text tokens and all visual tokens to quantify the importance of visual information in each layer. The greater the importance, the more budget is allocated. Meanwhile, we use the skewness of importance distribution to assess how effectively each layer aligns with the distribution characteristics of attention to visual tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.396, + 0.47, + 0.438 + ], + "angle": 0, + "content": "\\[\ns _ {k} = \\frac {N _ {v}}{\\left(N _ {v} - 1\\right) \\left(N _ {v} - 2\\right)} \\sum_ {i = 1} ^ {N _ {v}} \\left(\\frac {\\mathbf {I} _ {v} [ i ] - \\mu_ {I _ {v}}}{\\sigma_ {I _ {v}}}\\right) ^ {3}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.445, + 0.469, + 0.52 + ], + "angle": 0, + "content": "where \\(\\mu_{I_v}\\) and \\(\\sigma_{I_v}\\) represent the mean and standard deviation of the importance as well as attention scores distribution of visual tokens, which essentially reflect the distribution of attention scores. Ultimately, the reallocated token budget integrates these factors into the original budget \\(r\\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.52, + 0.469, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\hat {r} = \\frac {1}{2} \\left(s _ {t} ^ {\\prime} + s _ {k} ^ {\\prime}\\right) r. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.469, + 0.585 + ], + "angle": 0, + "content": "where \\( s_t' \\) and \\( s_k' \\) represent the normalized strength and skewness across different layers, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.596, + 0.209, + 0.614 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.159, + 0.638 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Models. The selection of LVLMs is primarily based on two criteria: model architecture and parameter size. In terms of architecture, the objective is to validate the effectiveness of the proposed method across various architectures, including LLaVA-OV-7B [16], InternVL2-8B [6], and Qwen2-VL-7B [35]. These models exhibit significant differences in visual token encoding, ViT, and LLM. For example, both LLaVA-OV-7B [16] and InternVL2-8B [6] utilize the AnyRes strategy with distinct ViT to increase the number of visual tokens, while Qwen2-VL-7B [35] supports native resolution with a large quantity of visual tokens. In terms of parameter size, models of InternVL2 series are selected for their diverse range of parameter scales, including InternVL2-1B [6], InternVL2-4B [6], and InternVL2-26B [6]. Due to space constraints, results for this section are provided in the appendix. Additionally, LLaVA-Video-7B-Qwen2 [41] serves as the base model for the video tasks evaluation." + }, + { + "type": "code_caption", + "bbox": [ + 0.502, + 0.092, + 0.845, + 0.106 + ], + "angle": 0, + "content": "Algorithm 1 AirCache for KV Cache Compression." + }, + { + "type": "algorithm", + "bbox": [ + 0.506, + 0.108, + 0.892, + 0.462 + ], + "angle": 0, + "content": "1: Input: Total cache budget \\(r\\), key, query, value of text prompt tokens \\(K_{t}\\), \\(Q_{t}\\), \\(V_{t} \\in \\mathbb{R}^{N_{t} \\times D}\\), key, query, value of visual prompt tokens \\(K_{v}\\), \\(Q_{v}\\), \\(V_{v} \\in \\mathbb{R}^{N_{v} \\times D}\\), number of layers L. \n2: for layer \\(l \\gets 0\\) to \\(N - 1\\) do \n3: \\(A_{tt} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t} \\mathbf{K}_{t}^{T}}{\\sqrt{D}}\\right)\\) \n4: Obtain the index of key text tokens \\(k\\) using Eq. (5). \n5: \\(Q_{t_{k}}\\), \\(K_{vt_{k}} \\gets Q_{t}[k]\\), \\(\\text{Concat}(K_{v}, K_{t}[k])\\) \n6: \\(\\mathbf{A}_{vt_{k}} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t_{k}} \\mathbf{K}_{vt_{k}}^{T}}{\\sqrt{D}}\\right)\\) \n7: \\(\\mathbf{I}_{v} \\gets \\frac{1}{N_{t_{k}}} \\sum_{j=0}^{N_{t_{k}}-1} \\mathbf{A}_{vt_{k}}[j,: N_{v}] \\triangleright\\) Importance Scores \n8: \\(s_{t} \\gets \\sum_{i=0}^{N_{v}-1} \\mathbf{I}_{v}[i]\\) \n9: \\(s_{k} \\gets \\frac{N_{v}}{(N_{v}-1)(N_{v}-2)} \\sum_{i=1}^{N_{v}}\\left(\\frac{\\mathbf{I}_{v}[i]-\\mu_{I_{v}}}{\\sigma_{I_{v}}}\\right)^{3} \\triangleright\\) Skewness \n10: \\(C^{l} \\gets (\\{(K_{v}, K_{t}), (V_{v}, V_{t}), (\\mathbf{I}_{v}, s_{t}, s_{k})\\})\\) \n11: end for \n12: for layer \\(l \\gets 0\\) to \\(N - 1\\) do \n13: \\(K_{v}, K_{t}, V_{v}, V_{t}, \\mathbf{I}_{v}, s_{t}, s_{k} \\gets C^{l}\\) \n14: \\(s_{t}', s_{k}' \\gets s_{t}, s_{k} \\triangleright\\) Normalized Across Different Layers \n15: \\(\\hat{r} = \\frac{1}{2}(s_{t}' + s_{k}') r \\quad \\triangleright\\) Obtain Layer-wise Budget \n16: \\(K' \\gets \\text{Concat}(K_{v}(\\mathbf{I}[N_{v} \\times \\hat{r}]), K_{t}) \\quad \\triangleright\\) Compress Key \n17: \\(V' \\gets \\text{Concat}(V_{v}(\\mathbf{I}[V_{v} \\times \\hat{r}]), V_{t}) \\quad \\triangleright\\) Compress Value \n18: \\(C^{l} \\gets (\\{K', V'\\})\\) \n19: end for \n20: Return: \\(C \\triangleright\\) Compressed KV Cache" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Datasets. The performance is primarily evaluated across four types of datasets: 1) Visual Question Answering (VQA) datasets, which include ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31]. Since KV cache compression mainly affects the decoding stage, VQA datasets that require longer output tokens are preferred for evaluating the method's efficacy. The similarity between generated and reference answers is typically measured by the Average Normalized Levenshtein Similarity (ANLS) metric [2]. 2) Multiple-choice question datasets, such as MMBench-EN [24], involves selecting the correct option from several choices, with accuracy serving as the ultimate metric. 3) Judgment datasets like MME [10], also utilize judgment accuracy as the evaluation criterion. 4) Video VQA datasets, such as MMBench-Video [9], measure the alignment between generated and reference answers using metrics provided by a large model evaluator." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Baselines. The comparison focuses on two main categories of methods. The first category encompasses classic and prominent KV cache compression techniques in LLMs, such as H2O [42], and SnapKV [19]. The second category comprises methods specifically optimized for LVLMs, namely PrefixKV [34] and Elastic Cache [25]. H2O [42] employs a full-range approach, while SnapKV [19] adopts a window-based strategy using textual tokens as the observation window for evaluating token importance. PrefixKV [34] utilizes the Lorenz curve to characterize the impor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23962" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.088, + 0.885, + 0.101 + ], + "angle": 0, + "content": "Table 1. The comparison of the KV cache compression methods on multimodal VQA benchmarks. The best result is highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.102, + 0.89, + 0.342 + ], + "angle": 0, + "content": "
ModelsMethodsChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
50%10%5%1%50%10%5%1%50%10%5%1%50%10%5%1%
LLaVA-OV-7B [16]Full80.380.380.380.366.166.166.166.187.087.087.087.076.076.076.076.0
H2O [42]79.877.474.071.065.459.254.252.086.474.260.855.375.570.167.960.4
Elastic [25]80.077.976.571.065.360.557.252.086.674.260.855.375.572.068.760.2
PrefixKV [34]80.178.276.870.965.461.158.051.986.680.574.755.475.872.770.060.2
SnapKV [19]80.179.378.372.966.164.263.057.886.984.480.664.175.873.469.558.2
Ours80.479.979.476.466.065.764.662.586.885.583.373.276.075.373.467.1
InternVL2-8B [6]Full82.282.282.282.273.073.073.073.091.091.091.091.077.777.777.777.7
H2O [42]81.981.679.376.872.370.665.161.290.579.470.348.977.170.066.059.6
Elastic [25]81.881.079.476.872.367.065.160.890.579.470.353.577.070.266.059.0
PrefixKV [34]81.980.779.376.772.367.465.361.090.679.570.255.177.170.366.258.6
SnapKV [19]82.280.479.575.072.872.368.464.990.990.183.571.177.473.971.061.7
Ours82.281.781.078.473.272.671.868.291.090.088.578.577.577.075.468.6
Qwen2-VL-7B [35]Full83.283.283.283.276.276.276.276.293.893.893.893.884.484.484.484.4
H2O [42]83.278.975.066.475.465.761.055.190.582.370.748.984.275.668.254.5
Elastic [25]83.279.775.468.075.567.162.355.593.583.071.949.284.276.768.854.4
PrefixKV [34]83.279.475.968.075.767.962.755.793.783.772.849.884.276.969.454.3
SnapKV [19]83.381.680.069.176.374.973.066.293.787.287.171.284.383.174.660.2
Ours83.482.380.875.976.275.274.570.793.892.990.978.384.483.480.968.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.352, + 0.468, + 0.378 + ], + "angle": 0, + "content": "Table 2. Performance comparison on other types datasets. Best-performing results are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.378, + 0.466, + 0.463 + ], + "angle": 0, + "content": "
MethodsMMBench-EN [24]MME [10]MMBench-Video [9]
50%10%1%50%10%1%50%10%1%
Full82.382.382.31585158515851.811.811.81
H2O [42]0.820.820.821585158515851.661.631.47
Elastic [25]0.820.820.821585158515851.681.621.51
PrefixKV [34]0.820.820.821585158515851.721.681.50
SnapKV [19]0.820.820.821585158515851.751.701.55
Ours0.820.820.821585158515851.801.781.67
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.581 + ], + "angle": 0, + "content": "tance distribution across different layers, guided by the retained important KV matrix. Elastic Cache [25] employs a fixed-point elimination strategy for the token importance with cache merging. All KV cache compression methods are applied exclusively to the visual tokens, as experiments have demonstrated that it offers greater stability and better performance compared to compressing the entire tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.747 + ], + "angle": 0, + "content": "Implementation details. All evaluation results are derived from VLMEvalKit [8], which is an open-source evaluation toolkit for LVLMs. The relevance threshold in Equation 5 is set to 0.9. The referee model for MMBench-Video [9] is Qwen2.5-72B [37]. The proposed method performs a one-time KV cache compression only after the prefetch stage is completed. Specifically, once the prefetch ends for each layer, the complete KV cache is fully saved. After computing all layers, we then calculate the budget for each layer and compress the KV cache accordingly. All experiments are conducted on \\(8 \\times \\mathrm{A}100 - 80\\mathrm{G}\\) GPUs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.757, + 0.216, + 0.771 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.841 + ], + "angle": 0, + "content": "We report the results in four sections: VQA datasets, other types of datasets including multiple-choice questions, judgment, and video VQA, inference performance, and the comparison with token pruning methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.468, + 0.901 + ], + "angle": 0, + "content": "VQA datasets. Table 1 presents the comparative results on the ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31] datasets. These datasets encompass tasks such as conventional image question answering, image in-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.356, + 0.892, + 0.702 + ], + "angle": 0, + "content": "formation extraction, document image understanding, and text recognition within images. It can be observed that the proposed method achieves the best results under most reduction ratios. When \\(50\\%\\) of visual KV cache is pruned, our method performs almost equivalently to the full KV cache, indicating near lossless performance. Even when reduced to only \\(10\\%\\) of visual KV cache, the performance gap with the full cache is limited to approximately \\(1\\%\\). As the token compression ratio increases, the advantage of the proposed method becomes more pronounced. For instance, when only \\(1\\%\\) of visual tokens are retained, we outperform SnapKV [19] by an average of \\(6.5\\%\\) across four VQA datasets with the LLaVA-OV-7B [16]. Moreover, the performance differences across different types of VQA datasets indicate that image documents or text understanding, which require higher local visual perception, have stricter demands on KV cache compression than general image understanding. The proposed method demonstrates a more significant advantage over other methods with these datasets, providing strong evidence of its superiority. Furthermore, the above conclusions also apply to different LVLMs, we achieve a significant advantage with InternVL2-8B [6] and Qwen2-VL-7B [35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Other types datasets. Table 2 presents the results of various methods on other types of multimodal datasets. For datasets involving multiple-choice questions and judgment types, such as MMBench-EN [24] and MME [10], the final result heavily depends on the accuracy of decoding the first token. Since the KV cache only influences the tokens decoded after the initial token, all methods achieve the same performance as the full KV cache for these types of datasets. The results from the MMBench-Video [9] dataset indicate that our proposed method achieves superior performance, further demonstrating its generality and superiority." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inference efficiency. We maintain a constant output token count of 512 and compare the full cache with the proposed" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23963" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.088, + 0.842, + 0.101 + ], + "angle": 0, + "content": "Table 3. Quantitative results on inference latency and throughput. The number of tokens output is consistently set to 512." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.101, + 0.847, + 0.23 + ], + "angle": 0, + "content": "
Batch SizePrompt LengthPrefill Latency (s)Decoding Latency (s)Throughput (token/s)
FullOursFull50%10%Full50%10%
82k1.61.8+12.5%11.69.4-19.0%8.2-29.3%353436+23.5%500+41.6%
8k4.95.3+8.2%15.011.9-20.7%9.4-37.3%273344+26.0%436+59.7%
16k11.311.9+5.3%21.715.0-30.9%10.4-47.5%189273+44.4%394+108.5%
32k25.226.4+4.8%35.922.0-38.7%12.3-65.7%114186+63.2%333+192.1%
162k2.32.5+8.7%14.411.2-22.2%9.7-32.6%569731+28.5%845+48.5%
8k9.49.8+4.3%22.615.6-31.0%11.8-47.8%362525+45.0%694+91.7%
16k20.321.4+5.4%36.923.0-37.7%12.8-65.3%222356+60.4%640+188.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.241, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Table 4. Performance comparison with token pruning methods under various compression ratio. The prefetch and decoding latency are measured under the settings of batch size of 16, total input token of 8k, and the compression ratio of \\(10\\%\\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.268, + 0.89, + 0.372 + ], + "angle": 0, + "content": "
MethodsChatQA [26]MMBench-EN [24]MME [10]MMBench-Video [9]Latency (s)
50%10%1%50%10%1%50%10%1%50%10%1%PrefillDecoding
Full80.380.380.382.382.382.31585158515851.811.811.819.422.6
FastV [5]73.647.716.981.474.833.6157413787861.691.361.025.411.7
FasterVLM [39]75.550.118.881.477.026.5155914417311.711.421.175.211.3
IVTP [13]74.955.822.581.777.536.2156314698491.701.471.245.812.6
Ours80.479.976.482.382.382.31585158515851.791.721.609.811.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.473, + 0.656 + ], + "angle": 0, + "content": "AirCache across different batch sizes and input token counts in terms of prefetch latency, decoding latency, and throughput. To facilitate testing, the text instruction token count is set to 64. As shown in Table 3, the proposed method significantly improves inference decoding speed under various inputs compared to the full cache. When compressing \\(50\\%\\) of the visual KV cache, the proposed method minimally impacts model performance and prefetch latency, while reducing decoding latency by \\(19\\%\\) to \\(39\\%\\) and increasing model throughput by \\(24\\%\\) to \\(63\\%\\). When compressing \\(10\\%\\) of the visual KV cache, the proposed method slightly decreases model performance by less than \\(1\\%\\), while further reducing decoding latency by \\(29\\%\\) to \\(66\\%\\) and increasing model throughput by \\(42\\%\\) to \\(192\\%\\). These results demonstrate that our proposed method can maintain strong model performance while ensuring significant decoding acceleration, indicating the effectiveness of visual token importance assessment and hierarchical budget allocation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Compared with token pruning methods. In addition to KV cache compression, another important method for accelerating LVLMs inference is to prune visual tokens before the prefetch stage. Table 4 compares the proposed method with these type of acceleration methods, including FastV [5], FasterVLM [39], and IVTP [13]. A notable characteristic is that the token pruning methods exhibit acceleration effects during the prefetch stage as they preemptively prune visual tokens. In contrast, all KV cache compression methods, including the proposed approach, demonstrate prefetch stage inference speeds comparable to those of the full cache. However, comparisons across various evaluation sets clearly show that the proposed method achieves an absolute advantage, particularly on VQA datasets, even when retaining only \\(1\\%\\) of the visual tokens. In contrast, token pruning methods experience a significant performance" + }, + { + "type": "table_caption", + "bbox": [ + 0.501, + 0.381, + 0.865, + 0.394 + ], + "angle": 0, + "content": "Table 5. Ablation results on the components of the AirCache." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.394, + 0.89, + 0.501 + ], + "angle": 0, + "content": "
SettingChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
Continuous window (16)70.456.661.355.9
Continuous window (32)72.957.864.158.2
All text tokens72.258.465.757.0
Visual window (32)68.855.159.253.7
Average allocation72.257.569.962.4
Pyramid allocation69.654.955.852.6
Only w/ st74.259.871.164.9
Only w/ sk74.761.471.963.6
Ours76.462.573.267.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.713 + ], + "angle": 0, + "content": "decline under these conditions. The prefetch stage facilitates sufficient interaction between visual and textual information, as the attention mechanism aggregates visual information to key text tokens. The proposed method effectively preserves these key text tokens during KV cache token pruning. Consequently, even when only \\(1\\%\\) of visual tokens are retained, the model maintains good inference performance. In contrast, most token pruning methods do not involve cross-modal interaction, resulting in direct visual information loss that cannot be transferred to the subsequent decoding stage in any form. This leads to insufficient and inaccurate visual guidance for generating, significantly degrading performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.665, + 0.74 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.794 + ], + "angle": 0, + "content": "In this section, we evaluate our method under various settings to verify the proposed modules from several perspectives. The base model is LLaVA-OV-7B [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Elite Observation Window. We introduce three common types of observation windows to assess the effectiveness of the proposed elite observation window. The first type employs a sequence of continuous text tokens as the observation window for evaluating the significance of visual tokens. The second type utilizes all text tokens, whereas the third employs the continuous visual token from the last visual" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23964" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.088, + 0.468, + 0.128 + ], + "angle": 0, + "content": "Table 6. Comparison results with different compression audiences, where * indicates uniform compression without distinguishing between visual tokens and text tokens." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.127, + 0.468, + 0.197 + ], + "angle": 0, + "content": "
Models50%10%5%1%
H2O* [42]74.6/407s46.5/559s32.0/733s27.8/1048s
H2O [42]79.8/323s77.4/251s74.0/239s71.0/215s
Ours*75.8/374s53.4/530s39.7/688s30.1/963s
Ours80.4/306s79.9/235s79.4/221s76.4/207s
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.214, + 0.47, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.468, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Comparison on ChatQA [26] and TextVQA [31] by retaining only one visual token, which is selected based on the sorting of visual token importance scores using different methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.468, + 0.609 + ], + "angle": 0, + "content": "segment as a reference. As shown in the first column of Table 5, the proposed elite observation window yields the best performance compared to the other three aforementioned approaches. The method using only visual tokens performs significantly worse due to the absence of textual instruction guidance, highlighting the importance of activating cross-modal associations in KV cache compression. Additionally, compared to methods that directly use all or portions of continuous text token sequences without considering variations in visual information perception among text tokens, the proposed method selects key text tokens to enhance the consistency of the observation window. This selection ensures greater stability and accuracy in evaluating the significance of visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.468, + 0.852 + ], + "angle": 0, + "content": "Layer-wise Token Budget Allocation Module. The second column of Table 5 presents the comparison using various budget allocation strategies across layers. Aside from the basic equal allocation, we also incorporate the strategy from PyramidKV [3], which allocates budgets in a descending ratio from shallow to deep layers. The results demonstrate that the proposed budget allocation, based on the strength and skewness of importance scores, delivers superior outcomes. Although PyramidKV [3] perform well in LLMs, its suboptimal results in LVLMs highlight the distinct characteristics of multimodal models and their heightened sensitivity to hierarchical budget allocation. An inappropriate allocation strategy is more likely to adversely affect the final results. We also investigate the separate use of strength and skewness for layer-wise budget quantification, revealing that their combination produces the best results." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Visual Token Importance Evaluation. Figure 4 visually demonstrates the comparative effectiveness of various methods in evaluating the importance of visual tokens. To" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.378 + ], + "angle": 0, + "content": "directly assess the efficacy of these methods in ranking visual token importance, we removed all non-essential modules and selected only one visual token at a time from the ranked sorting results, ordered from high to low importance. Theoretically, as the importance ranking of tokens declines, the model's performance is also expected to decrease. The graph clearly shows that the proposed method aligns more closely with the theoretical trend, where the visual token ranked as TOP1 achieves the highest metric, followed by a continuous and smooth decline. Notably, after around \\(10\\%\\), the performance of the visual tokens approximates that of the case with all visual tokens removed. This indicates significant redundancy among visual tokens in the KV cache of LVLMs. The results in Table 1, which demonstrate that retaining over \\(10\\%\\) of the visual tokens limits performance decline to no more than \\(1\\%\\), further corroborate this observation. In contrast, the performance derived from other methods exhibit significant noise, indicating instability and deficiencies in their importance assessment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.619 + ], + "angle": 0, + "content": "Unified Compression or Vision-only Compression. Given that the KV cache in LVLMs encompasses both visual and textual modalities, in contrast to the single textual modality of LLMs, we compared Unified Compression, which does not differentiate between modal differences during KV cache compression, with Vision-only Compression, which compresses only visual tokens. As shown in Table 6, the results indicate that unified compression performs significantly worse than vision-only compression in terms of both inference performance and speed. Moreover, as the compression ratio increases, the inference time for unified compression actually continues to rise. The primary reason is that unified compression disrupts the effective representation of textual instructions, leading the model to produce incorrect outputs and exhibit a tendency for repetition, which in turn increases inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.634, + 0.618, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose Activating Inter-modal Relevancy (AirCache), a novel KV cache compression for efficient LVLMs inference. By leveraging inter-modal relevancy between textual and visual tokens to guide KV cache eviction, we identified significant differences in the attention distribution of visual tokens corresponding to different textual tokens. Based on this finding, we propose an elite observation window with refined key text tokens, which provides a more stable and effective assessment of visual token importance. Furthermore, by exploring the diversity of importance score distribution across different layers, we propose to achieve comprehensive quantification in terms of distribution strength and skewness for adaptive layer-wise compression budget allocation. Extensive experiments across multiple benchmarks and LVLMs demonstrate the effectiveness of AirCache in KV cache compression." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23965" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.183 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.255 + ], + "angle": 0, + "content": "[2] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information funneling. arXiv preprint arXiv:2406.02069, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[4] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.471, + 0.466 + ], + "angle": 0, + "content": "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.47, + 0.62 + ], + "angle": 0, + "content": "[7] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.621, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[8] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[9] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[11] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[12] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[13] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[14] Chaoya Jiang, Haiyang Xu, Wei Ye, Qinghao Ye, Chen-liang Li, Ming Yan, Bin Bi, Shikun Zhang, Fei Huang, and Songfang Huang. Bus: Efficient and effective vision-language pre-training with bottom-up patch summarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2900–2910, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[15] Jang-Hyun Kim, Junyoung Yeom, Sangwoo Yun, and Hyun Oh Song. Compressed context memory for online language model interaction. arXiv preprint arXiv:2312.03414, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[16] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[18] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[19] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: Llm knows what you are looking for before generation. arXiv preprint arXiv:2404.14469, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[20] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[21] Akide Liu, Jing Liu, Zizheng Pan, Yefei He, Reza Haffari, and Bohan Zhuang. Minicache: Kv cache compression in depth dimension for large language models. Advances in Neural Information Processing Systems, 37:139997-140031, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23966" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[25] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[27] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[29] Piotr Nawrot, Adrian Łancucki, Marcin Chochowski, David Tarjan, and Edoardo M Ponti. Dynamic memory compression: Retrofitting llms for accelerated inference. arXiv preprint arXiv:2403.09636, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[30] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.477, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[31] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.549, + 0.469, + 0.63 + ], + "angle": 0, + "content": "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.633, + 0.469, + 0.688 + ], + "angle": 0, + "content": "[33] Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, and Panpan Xu. VI-cache: Sparsity and modality-aware kv cache compression for vision-language model inference acceleration. arXiv preprint arXiv:2410.23317, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.69, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[34] Ao Wang, Hui Chen, Jianchao Tan, Kefeng Zhang, Xunliang Cai, Zijia Lin, Jungong Han, and Guiguang Ding. Prefixkv: Adaptive prefix kv cache is what vision instruction-following models need for efficient generation. arXiv preprint arXiv:2412.03409, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.146 + ], + "angle": 0, + "content": "[37] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[38] Yifei Yang, Zouying Cao, Qiguang Chen, Libo Qin, Dongjie Yang, Hai Zhao, and Zhi Chen. Kvsharer: Efficient inference via layer-wise dissimilar kv cache sharing. arXiv preprint arXiv:2410.18517, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[39] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[40] Yuxin Zhang, Yuxuan Du, Gen Luo, Yunshan Zhong, Zhenyu Zhang, Shiwei Liu, and Rongrong Ji. Cam: Cache merging for memory-efficient llms inference. In *Forty-first International Conference on Machine Learning*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[41] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.457 + ], + "angle": 0, + "content": "[42] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 4, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "23967" + } + ] +] \ No newline at end of file diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_origin.pdf b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0cda351834821205bd2c05a1226e7551f2cbe21a --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/7536831b-10b6-4fe8-b035-0c2896718fa3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e3e179adb9d075d779cde2c6d03f11f47d80d6f8c6fd84f4acc3538fe113c1a +size 2348143 diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/full.md b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/full.md new file mode 100644 index 0000000000000000000000000000000000000000..231e8a0651125c957dd910fd4de55c78a3fa0705 --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/full.md @@ -0,0 +1,300 @@ +# AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference + +Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie\*, Hao Wang Alibaba Group + +zhouwan.hk, zh372956, bochen.wbc, yx150449, xiezhen.xz, qiao.wh@alibaba-inc.com + +# Abstract + +Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. We introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only $10\%$ of visual KV cache, thereby reducing decoding latency by $29\%$ to $66\%$ across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches. + +# 1. Introduction + +The past few years have witnessed remarkable advancements in Large Visual-Language Models (LVLMs) [1, 6, 16, 22, 23, 32, 35, 36], both in research and prac + +![](images/a467d1151f33fa4cc67bc19015d697066262c2a76b64e4fdc190bf02c224b4ea.jpg) +Figure 1. Illustration of our motivation. (a) An example of attention maps during the decoding process of LLaVA-OV-7B [16]. (b) Partial visual attention maps obtained using all text tokens as the observation window. (c) Using the last 16 text tokens as the observation window. (d) Using the last 16 visual tokens as the observation window, demonstrating more consistent attention distribution. + +![](images/47b792cdd7f3154c8287184933cc516431f3b598ea0807c48f077f6b9332380e.jpg) + +tical applications. Although LVLMs exhibit increasingly advanced visual processing capabilities, such as handling high-resolution, multiple images, and video sequences, these improvements entail substantial computational costs. The exponential increase in visual tokens, coupled with the demands of long-context generation tasks, results in unsustainable memory overhead from key-value (KV) cache storage [30]. This not only significantly increases GPU memory consumption but also severely degrades computational efficiency due to heightened memory bandwidth contention. + +Recent approaches [5, 13, 14, 19, 39, 42] aim to prune visual tokens during inference, which can be categorized into token pruning and KV cache compression. Token pruning [11, 13, 14, 18, 20, 39] involves reducing visual tokens during the prefetch phase, which decreases both the number of tokens processed in subsequent transformer layers and the tokens stored in the KV cache. Although it significantly enhances inference speed, the aggressive elimination of visual tokens during prefetch leads to a substantial loss of visual information, severely degrading model performance. KV cache compression [19, 21, 25, 34, 42] reduces token solely during the decoding phase by pruning stored data in + +the KV cache. This technique has been extensively validated in LLMs. Since all tokens undergo a complete forward pass and the causal attention mechanism in LLMs differentiates the importance distribution among tokens, selectively deleting certain tokens has minimal impact on model performance. These methods typically assess the importance of visual tokens by evaluating their strength or sparsity within the attention maps computed under the causal attention mechanism. Given the strict unidirectionality of causal attention, using cross-modality attention scores between instruction tokens and visual tokens provides a more comprehensive evaluation of visual token importance than using full attention scores. Moreover, we observe significant variations in the distribution of visual token strength across different layers, indicating that uniformly allocating the budget is not the optimal strategy. + +To address the aforementioned issues, we propose AirCache, a novel KV cache compression method designed for efficient LVLMs inference. AirCache comprises two main components: visual token importance scoring and layerwise KV cache budget allocation. To effectively evaluate the importance of visual tokens in the KV cache, we introduce an elite observation window that selects critical instruction tokens by leveraging self-attention scores among instruction tokens rather than using all or continuous local instruction tokens. As illustrated in Figure 1, compared to mainstream methods that assess visual token importance based on observation windows guided by all or partial text instruction tokens, the elite observation window demonstrates stronger consistency. Most text tokens within the elite observation window tend to provide more similar evaluations for the same visual token, thereby enhancing the effectiveness and stability of subsequent voting-based ranking. We further propose to quantify the compression budget for different layers from two perspectives. The first aspect is the emphasis a layer places on visual information, measured by the sum of attention allocated to all visual tokens, as well as the strength of importance score distribution. Additionally, analysis of attention distributions across layers reveals a distinct head effect, where a few visual tokens receive high importance scores, while the majority exhibit mediocre and average importance characteristics. Our experimental results show that preserving only $10\%$ of the important visual KV cache results in less than $1\%$ average model performance degradation, corroborating the significant head effect in the importance distribution of visual tokens. Combining these two aspects, we allocate the budget for visual KV cache based on both the strength and skewness of importance scores distributions. + +Our primary contributions include: (1) We conduct a comprehensive and detailed exploration of the differences in inter-modal relevancy patterns based on attention interaction in evaluating visual KV cache importance. We in + +troduce an elite observation window with carefully selected key textural instruction tokens to achieve more effective and stable importance evaluations. (2) We propose to quantify the differentiated compression budget allocation across different layers using the strength and skewness of the importance scores distribution, based on the prominence of visual information and the efficacy of attention allocation. (3) Experimental results on widely used LVLMs and benchmark datasets demonstrate that the proposed method achieves superior performance compared to other existing methods. + +# 2. Related Work + +Visual Token Pruning. Visual token pruning [4, 5, 12, 13, 20, 39] accelerates model inference by reducing visual tokens during the prefill phase. BLIP-2 [17] and Qwen-VL [1] employ cross-attention with learnable query embeddings to aggregate input visual tokens, while HoneyBee [4] and MobileVLM [7] use convolutional neural networks. These approaches necessitate model retraining, which significantly increases migration costs. In contrast, methods like [5, 39] introduce no extra parameters but leverage attention maps to score and evict visual tokens based on importance. IVTP [13] implements two-stage pruning by deleting tokens in both Vision Transformer (ViT) and LLM, enabling text-guided pruning through instruction integration. However, it is important to note that visual token pruning inevitably leads to visual information loss by discarding tokens prior to or within early layers, thereby substantially degrading model performance. + +KV cache compression. KV cache strategies accelerate inference by storing the key and value states of previous tokens, thereby avoiding redundant computations during decoding. However, as the number of tokens increases, the pressure on memory and bandwidth also rises, limiting the model's application in long content understanding or generation. To address this challenge, mainstream methods primarily focus on two directions: intra-layer KV cache pruning [15, 29, 40] and inter-layer compression budget allocation [21, 38]. H2O [42] and SnapKV [19] propose evicting tokens in the KV cache using global and local cumulative attention scores, while PyramidKV [3] constructs a pyramid-like hierarchical budget allocation to achieve more refined KV cache compression management. The input to LVLMs is a cross-modal construct primarily composed of visual tokens with auxiliary text tokens, necessitating additional consideration of the impact of different modality differences when extending KV cache compression to LVLMs. Elastic [25] merges non-critical tokens with important tokens, PrefixKV [34] introduces an adaptive layer-wise KV retention recipe based on binary search for maximal preservation of contextual information. VL-Cache [33] utilizes visual token sparsity for layer-wise budget allocation and token eviction through dedicated scoring policies. + +![](images/d04ec30050357743f6fe7d5491a37dd44d7667b66434e29671518682c6845f27.jpg) +Figure 2. The left panel illustrates the inference workflow of standard LVLMs. The proposed method integrates during the KV cache storage process following the prefetch stage, maintaining compatibility with mainstream LVLMs architecture. The right panel presents the overview of AirCache. We first employ self-attention across the instruction text tokens to identify the key text tokens, establishing an elite observation window for assessing visual token significance. We further reallocate the compression budgets between layers based on the strength and skewness of the importance distribution. Finally, pruning is applied to the KV cache according to the importance ranking of visual tokens in the KV Cache and the compression budget to obtain the final compressed KV cache. + +![](images/85ae4793679c26c7c20cf9eff09e7219b954eec717ea08769c7320fda8a0cb43.jpg) + +![](images/ee244de9c3edf0a2bba9c6b1f91b7beeb1466967113cd9d87b271b7410bf3e39.jpg) + +![](images/4a3d2980077e8266a1e24e7c36890548bd4d3c7be20bcb53ec7916cf964c7611.jpg) + +# 3. Methodology + +# 3.1. Preliminary + +Similar to most LLMs, the inference process of LVLMs is primarily divided into prompt prefill and token decoding. Prompt Prefill. Given the hidden states $X \in \mathbb{R}^{N \times D}$ of the input prompt, where $N$ is the total number of visual and text prompt tokens, and $D$ denotes hidden dimension. We simplify the expression by omitting the indices of the head and layer of the hidden states. The query, key, and value states of the attention block can be represented as: + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} _ {Q}, \quad \mathbf {K} = \mathbf {X} \mathbf {W} _ {K}, \quad \mathbf {V} = \mathbf {X} \mathbf {W} _ {V}, \tag {1} +$$ + +where $\mathbf{W}_Q, \mathbf{W}_K, \mathbf{W}_V \in \mathbb{R}^{D \times D}$ are learnable projection matrices. Following the completion of the attention interaction, the key and value are stored in the KV cache to support subsequent token generation. + +Token Decoding. In this stage, the model efficiently generates new tokens by iteratively utilizing and updating the KV cache. Specifically, during the $i$ th iteration, we only need to compute the key and value associated with the newly generated token $\mathbf{x}_i \in \mathbb{R}^{1 \times D}$ . By employing cache indexing, the model avoids recomputing the key and value for tokens that have already been processed. Subsequently, the KV cache is updated with the newly computed key and value for the latest token: + +$$ +\mathbf {K} = \operatorname {C o n c a t} \left(\mathbf {K}, \mathbf {x} _ {i} \mathbf {W} _ {K}\right), \mathbf {V} = \operatorname {C o n c a t} \left(\mathbf {V}, \mathbf {x} _ {i} \mathbf {W} _ {V}\right). \tag {2} +$$ + +Although the KV cache mitigates the issue of redundant computation during the decoding process, its characteristic of linear growth in computational requirements relative to the length of the input sequence poses challenges regarding latency and memory usage, especially in scenarios involving longer inputs or outputs. This issue is further exacerbated as recent LVLMs tend to prioritize high-resolution, high-frame-rate visual inputs. + +# 3.2. Overview + +The AirCache is primarily divided into two components. The first component evaluates the significance of visual tokens based on the elite observation window. Section 3.3 details how to optimize the relevancy between visual and language modalities to guide the compression of the visual segment in the KV cache. The second component quantifies the KV cache budget requirements across various layers by analyzing the overall strength and skewness of the importance score distribution, thereby facilitating dynamic layerwise KV cache budget allocation, as elaborated in Section 3.4. The architecture and algorithm underlying the method are illustrated in Figure 2 and Algorithm 1. + +# 3.3. Elite Observation Window + +In LVLMs, the presence of multiple modalities complicates the selection of observation windows, suggesting that adhering to the relevant practices established in LLMs may not be optimal. For instance, as illustrated in Figure 1, the visual token importance measures derived from the continuous last segment of the prompt exhibit poor consistency. This variability indicates that different text tokens within the observation window yield significantly diverse visual tokens importance distributions. Consequently, the hit rate statistics derived from the voting mechanism are prone to substantial noise, and the disparities among modalities may further amplify this noise interference. Based on this consideration, we propose refining the observation windows by incorporating carefully selected text tokens. By utilizing the attention hit rate within the text modality, we aim to enhance the consistency of visual token evaluation, thereby improving the performance of KV cache compression. To this end, we reorganize the hidden states of the input prompt by: + +$$ +\mathbf {X} = \operatorname {C o n c a t} \left(\mathbf {X} _ {v}, \mathbf {X} _ {t}\right) \in \mathbb {R} ^ {\left(N _ {v} + N _ {t}\right) \times D}, \tag {3} +$$ + +![](images/92ec940688a6a44686f2e7900d795bace6e62108207312c001632605d6c14ed2.jpg) +(a) + +![](images/4ee76410aaa113aec42e8a0152bcaa354c004f2cf32332b0159a0b85871e011a.jpg) +(b) +Figure 3. (a) The strength of the distribution of visual token importance scores across different layers, which is defined as the sum of all visual token importance scores. (b) Performance with selecting only one visual token each time based on the sorted results of visual token importance scores, the index of 0 indicates the setting of removing all visual tokens. (c) The distribution of visual token importance scores across different layers, with only selected layers displayed for clarity. All results are derived from LLaVA-OV-7B [16] on the ChatQA [26]. + +![](images/e20ca144141a19854a637bca3fc3640d2a7b0c2a106349a95f6a1de0bcdbc7a7.jpg) +(c) + +where $\mathbf{X}_v$ and $\mathbf{X}_t$ represent the hidden states corresponding to visual tokens and instruction text tokens, respectively. $N_v$ and $N_t$ denote the respective quantities of these tokens, satisfying the condition $N_v + N_t = N$ . For convenience, the system prompts are omitted. We compute the text unimodal attention matrix composed of instruction text tokens as: + +$$ +\mathbf {A} _ {t t} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {t} \mathbf {K} _ {t} ^ {T}}{\sqrt {D}}\right) \in \mathbb {R} ^ {N _ {t} \times N _ {t}}, \tag {4} +$$ + +where $\mathbf{Q}_t$ and $\mathbf{K}_t$ denote the query and key of the text tokens, respectively. Considering the characteristics of causal attention, we select the last token of the instruction text as reference and filter out key text tokens that receive high attention scores based on their interactions with other text tokens. Consistent with [42], we employ a relevance threshold to regulate the sparsity of key text tokens: + +$$ +k = \{j \mid \mathbf {A} _ {t t} [ N _ {t} - 1, j ] \geq \alpha \cdot \max \mathbf {A} _ {t t} [ N _ {t} - 1,: ] \}, \tag {5} +$$ + +where $\alpha \in [0,1]$ represents the relevance threshold, and $j\in \{0,1,\dots ,N_t - 1\}$ is the index of the text token. When $\alpha$ is equal to 0, all text tokens are retained, and when $\alpha$ is equal to 1, only the text token with the highest score is preserved. We redefine the key for the aggregated visual tokens and key text tokens as $\mathbf{K}_{vt_k}$ , along with the query corresponding to the key text tokens $\mathbf{Q}_{t_k}$ , as follows: + +$$ +\mathbf {Q} _ {t _ {k}} = \mathbf {Q} _ {t} [ k ] \in \mathbb {R} ^ {N _ {t _ {k}} \times D}, \tag {6} +$$ + +$$ +\mathbf {K} _ {v t k} = \operatorname {C o n c a t} \left(\mathbf {K} _ {v}, \mathbf {K} _ {t} [ k ]\right) \in \mathbb {R} ^ {\left(N _ {v} + N _ {t _ {k}}\right) \times D}, +$$ + +where $\mathbf{K}_v$ represents the key of the visual token, and $N_{t_k}$ denotes the number of selected key text tokens. Furthermore, the attention matrix between the visual tokens and the key text tokens can be computed as follows: + +$$ +\mathbf {A} _ {v t _ {k}} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {t _ {k}} \mathbf {K} _ {v t _ {k}} ^ {T}}{\sqrt {D}}\right) \in \mathbb {R} ^ {N _ {t _ {k}} \times \left(N _ {v} + N _ {t _ {k}}\right)}. \tag {7} +$$ + +The elite observation window, comprising only these key text tokens, offers a more comprehensive and focused semantic representation compared to those formed by all text + +tokens or fixed continuous sequences. It provides enhanced stability and accuracy in assessing the importance of visual tokens. The final importance scores of the visual tokens can be obtained by performing average pooling of the attention matrix along the text dimension: + +$$ +\mathbf {I} _ {v} = \frac {1}{N _ {t _ {k}}} \sum_ {j = 0} ^ {N _ {t _ {k}} - 1} \mathbf {A} _ {v t _ {k}} [ j,: N _ {v} ]. \tag {8} +$$ + +By sorting the importance scores, the KV cache of visual tokens with lower rankings can be discarded in accordance with the compression budget. Furthermore, the elite observation window we employ exhibits reduced computational complexity, and the additional overhead of key text token selection is negligible in visual token-heavy LVLMs. + +# 3.4. Layer-wise KV Cache Budget Allocation + +We further investigate the characteristics of the visual token importance score distribution based on the elite observation window across various layers. Specifically, we consider the sum of the importance scores of all visual tokens as an indicator of the layer's attention to visual information. As illustrated in Figure 3(a), there is a significant disparity in the attention given to visual tokens across different layers. This observation highlights the necessity of performing layer-wise token budget allocation, where layers dedicating more attention to visual information merit a larger token budget. As shown in Figure 3(b), we select one visual token at a time based on the importance scores. It can be observed that only approximately $10\%$ of all visual tokens have a significantly positive impact on the final result. This suggests a head effect in the model's attention distribution over visual tokens, with a small subset receiving higher attention, while the majority have lower and more balanced scores. Extending to the importance distribution across different layers in Figure 3(c), it shows significant differences in importance distribution among the layers, particularly regarding the head effect of visual tokens. Consequently, + +we aim to allocate a larger token budget to the layers that demonstrate a pronounced head effect in their visual tokens importance distribution. This indicates a more accurate and nuanced understanding of the visual content by those layers. + +As aforementioned, we quantify the allocation of visual token budget across different layers from two perspectives: the strength and skewness of the importance distribution. The strength of the importance distribution is calculated as: + +$$ +s _ {t} = \sum_ {i = 0} ^ {N _ {v} - 1} \mathbf {I} _ {v} [ i ]. \tag {9} +$$ + +We exclude the attention between text tokens and instead aggregate the attention scores between text tokens and all visual tokens to quantify the importance of visual information in each layer. The greater the importance, the more budget is allocated. Meanwhile, we use the skewness of importance distribution to assess how effectively each layer aligns with the distribution characteristics of attention to visual tokens: + +$$ +s _ {k} = \frac {N _ {v}}{\left(N _ {v} - 1\right) \left(N _ {v} - 2\right)} \sum_ {i = 1} ^ {N _ {v}} \left(\frac {\mathbf {I} _ {v} [ i ] - \mu_ {I _ {v}}}{\sigma_ {I _ {v}}}\right) ^ {3}, \tag {10} +$$ + +where $\mu_{I_v}$ and $\sigma_{I_v}$ represent the mean and standard deviation of the importance as well as attention scores distribution of visual tokens, which essentially reflect the distribution of attention scores. Ultimately, the reallocated token budget integrates these factors into the original budget $r$ : + +$$ +\hat {r} = \frac {1}{2} \left(s _ {t} ^ {\prime} + s _ {k} ^ {\prime}\right) r. \tag {11} +$$ + +where $s_t'$ and $s_k'$ represent the normalized strength and skewness across different layers, respectively. + +# 4. Experiments + +# 4.1. Setup + +Models. The selection of LVLMs is primarily based on two criteria: model architecture and parameter size. In terms of architecture, the objective is to validate the effectiveness of the proposed method across various architectures, including LLaVA-OV-7B [16], InternVL2-8B [6], and Qwen2-VL-7B [35]. These models exhibit significant differences in visual token encoding, ViT, and LLM. For example, both LLaVA-OV-7B [16] and InternVL2-8B [6] utilize the AnyRes strategy with distinct ViT to increase the number of visual tokens, while Qwen2-VL-7B [35] supports native resolution with a large quantity of visual tokens. In terms of parameter size, models of InternVL2 series are selected for their diverse range of parameter scales, including InternVL2-1B [6], InternVL2-4B [6], and InternVL2-26B [6]. Due to space constraints, results for this section are provided in the appendix. Additionally, LLaVA-Video-7B-Qwen2 [41] serves as the base model for the video tasks evaluation. + +Algorithm 1 AirCache for KV Cache Compression. +1: Input: Total cache budget $r$ , key, query, value of text prompt tokens $K_{t}$ , $Q_{t}$ , $V_{t} \in \mathbb{R}^{N_{t} \times D}$ , key, query, value of visual prompt tokens $K_{v}$ , $Q_{v}$ , $V_{v} \in \mathbb{R}^{N_{v} \times D}$ , number of layers L. +2: for layer $l \gets 0$ to $N - 1$ do +3: $A_{tt} \gets \text{Softmax}\left(\frac{\mathbf{Q}_{t} \mathbf{K}_{t}^{T}}{\sqrt{D}}\right)$ +4: Obtain the index of key text tokens $k$ using Eq. (5). +5: $Q_{t_{k}}$ , $K_{vt_{k}} \gets Q_{t}[k]$ , $\text{Concat}(K_{v}, K_{t}[k])$ +6: $\mathbf{A}_{vt_{k}} \gets \text{Softmax}\left(\frac{\mathbf{Q}_{t_{k}} \mathbf{K}_{vt_{k}}^{T}}{\sqrt{D}}\right)$ +7: $\mathbf{I}_{v} \gets \frac{1}{N_{t_{k}}} \sum_{j=0}^{N_{t_{k}}-1} \mathbf{A}_{vt_{k}}[j,: N_{v}] \triangleright$ Importance Scores +8: $s_{t} \gets \sum_{i=0}^{N_{v}-1} \mathbf{I}_{v}[i]$ +9: $s_{k} \gets \frac{N_{v}}{(N_{v}-1)(N_{v}-2)} \sum_{i=1}^{N_{v}}\left(\frac{\mathbf{I}_{v}[i]-\mu_{I_{v}}}{\sigma_{I_{v}}}\right)^{3} \triangleright$ Skewness +10: $C^{l} \gets (\{(K_{v}, K_{t}), (V_{v}, V_{t}), (\mathbf{I}_{v}, s_{t}, s_{k})\})$ +11: end for +12: for layer $l \gets 0$ to $N - 1$ do +13: $K_{v}, K_{t}, V_{v}, V_{t}, \mathbf{I}_{v}, s_{t}, s_{k} \gets C^{l}$ +14: $s_{t}', s_{k}' \gets s_{t}, s_{k} \triangleright$ Normalized Across Different Layers +15: $\hat{r} = \frac{1}{2}(s_{t}' + s_{k}') r \quad \triangleright$ Obtain Layer-wise Budget +16: $K' \gets \text{Concat}(K_{v}(\mathbf{I}[N_{v} \times \hat{r}]), K_{t}) \quad \triangleright$ Compress Key +17: $V' \gets \text{Concat}(V_{v}(\mathbf{I}[V_{v} \times \hat{r}]), V_{t}) \quad \triangleright$ Compress Value +18: $C^{l} \gets (\{K', V'\})$ +19: end for +20: Return: $C \triangleright$ Compressed KV Cache + +Datasets. The performance is primarily evaluated across four types of datasets: 1) Visual Question Answering (VQA) datasets, which include ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31]. Since KV cache compression mainly affects the decoding stage, VQA datasets that require longer output tokens are preferred for evaluating the method's efficacy. The similarity between generated and reference answers is typically measured by the Average Normalized Levenshtein Similarity (ANLS) metric [2]. 2) Multiple-choice question datasets, such as MMBench-EN [24], involves selecting the correct option from several choices, with accuracy serving as the ultimate metric. 3) Judgment datasets like MME [10], also utilize judgment accuracy as the evaluation criterion. 4) Video VQA datasets, such as MMBench-Video [9], measure the alignment between generated and reference answers using metrics provided by a large model evaluator. + +Baselines. The comparison focuses on two main categories of methods. The first category encompasses classic and prominent KV cache compression techniques in LLMs, such as H2O [42], and SnapKV [19]. The second category comprises methods specifically optimized for LVLMs, namely PrefixKV [34] and Elastic Cache [25]. H2O [42] employs a full-range approach, while SnapKV [19] adopts a window-based strategy using textual tokens as the observation window for evaluating token importance. PrefixKV [34] utilizes the Lorenz curve to characterize the impor + +Table 1. The comparison of the KV cache compression methods on multimodal VQA benchmarks. The best result is highlighted in bold. + +
ModelsMethodsChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
50%10%5%1%50%10%5%1%50%10%5%1%50%10%5%1%
LLaVA-OV-7B [16]Full80.380.380.380.366.166.166.166.187.087.087.087.076.076.076.076.0
H2O [42]79.877.474.071.065.459.254.252.086.474.260.855.375.570.167.960.4
Elastic [25]80.077.976.571.065.360.557.252.086.674.260.855.375.572.068.760.2
PrefixKV [34]80.178.276.870.965.461.158.051.986.680.574.755.475.872.770.060.2
SnapKV [19]80.179.378.372.966.164.263.057.886.984.480.664.175.873.469.558.2
Ours80.479.979.476.466.065.764.662.586.885.583.373.276.075.373.467.1
InternVL2-8B [6]Full82.282.282.282.273.073.073.073.091.091.091.091.077.777.777.777.7
H2O [42]81.981.679.376.872.370.665.161.290.579.470.348.977.170.066.059.6
Elastic [25]81.881.079.476.872.367.065.160.890.579.470.353.577.070.266.059.0
PrefixKV [34]81.980.779.376.772.367.465.361.090.679.570.255.177.170.366.258.6
SnapKV [19]82.280.479.575.072.872.368.464.990.990.183.571.177.473.971.061.7
Ours82.281.781.078.473.272.671.868.291.090.088.578.577.577.075.468.6
Qwen2-VL-7B [35]Full83.283.283.283.276.276.276.276.293.893.893.893.884.484.484.484.4
H2O [42]83.278.975.066.475.465.761.055.190.582.370.748.984.275.668.254.5
Elastic [25]83.279.775.468.075.567.162.355.593.583.071.949.284.276.768.854.4
PrefixKV [34]83.279.475.968.075.767.962.755.793.783.772.849.884.276.969.454.3
SnapKV [19]83.381.680.069.176.374.973.066.293.787.287.171.284.383.174.660.2
Ours83.482.380.875.976.275.274.570.793.892.990.978.384.483.480.968.6
+ +Table 2. Performance comparison on other types datasets. Best-performing results are highlighted in bold. + +
MethodsMMBench-EN [24]MME [10]MMBench-Video [9]
50%10%1%50%10%1%50%10%1%
Full82.382.382.31585158515851.811.811.81
H2O [42]0.820.820.821585158515851.661.631.47
Elastic [25]0.820.820.821585158515851.681.621.51
PrefixKV [34]0.820.820.821585158515851.721.681.50
SnapKV [19]0.820.820.821585158515851.751.701.55
Ours0.820.820.821585158515851.801.781.67
+ +tance distribution across different layers, guided by the retained important KV matrix. Elastic Cache [25] employs a fixed-point elimination strategy for the token importance with cache merging. All KV cache compression methods are applied exclusively to the visual tokens, as experiments have demonstrated that it offers greater stability and better performance compared to compressing the entire tokens. + +Implementation details. All evaluation results are derived from VLMEvalKit [8], which is an open-source evaluation toolkit for LVLMs. The relevance threshold in Equation 5 is set to 0.9. The referee model for MMBench-Video [9] is Qwen2.5-72B [37]. The proposed method performs a one-time KV cache compression only after the prefetch stage is completed. Specifically, once the prefetch ends for each layer, the complete KV cache is fully saved. After computing all layers, we then calculate the budget for each layer and compress the KV cache accordingly. All experiments are conducted on $8 \times \mathrm{A}100 - 80\mathrm{G}$ GPUs. + +# 4.2. Main Results + +We report the results in four sections: VQA datasets, other types of datasets including multiple-choice questions, judgment, and video VQA, inference performance, and the comparison with token pruning methods. + +VQA datasets. Table 1 presents the comparative results on the ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31] datasets. These datasets encompass tasks such as conventional image question answering, image in- + +formation extraction, document image understanding, and text recognition within images. It can be observed that the proposed method achieves the best results under most reduction ratios. When $50\%$ of visual KV cache is pruned, our method performs almost equivalently to the full KV cache, indicating near lossless performance. Even when reduced to only $10\%$ of visual KV cache, the performance gap with the full cache is limited to approximately $1\%$ . As the token compression ratio increases, the advantage of the proposed method becomes more pronounced. For instance, when only $1\%$ of visual tokens are retained, we outperform SnapKV [19] by an average of $6.5\%$ across four VQA datasets with the LLaVA-OV-7B [16]. Moreover, the performance differences across different types of VQA datasets indicate that image documents or text understanding, which require higher local visual perception, have stricter demands on KV cache compression than general image understanding. The proposed method demonstrates a more significant advantage over other methods with these datasets, providing strong evidence of its superiority. Furthermore, the above conclusions also apply to different LVLMs, we achieve a significant advantage with InternVL2-8B [6] and Qwen2-VL-7B [35]. + +Other types datasets. Table 2 presents the results of various methods on other types of multimodal datasets. For datasets involving multiple-choice questions and judgment types, such as MMBench-EN [24] and MME [10], the final result heavily depends on the accuracy of decoding the first token. Since the KV cache only influences the tokens decoded after the initial token, all methods achieve the same performance as the full KV cache for these types of datasets. The results from the MMBench-Video [9] dataset indicate that our proposed method achieves superior performance, further demonstrating its generality and superiority. + +Inference efficiency. We maintain a constant output token count of 512 and compare the full cache with the proposed + +Table 3. Quantitative results on inference latency and throughput. The number of tokens output is consistently set to 512. + +
Batch SizePrompt LengthPrefill Latency (s)Decoding Latency (s)Throughput (token/s)
FullOursFull50%10%Full50%10%
82k1.61.8+12.5%11.69.4-19.0%8.2-29.3%353436+23.5%500+41.6%
8k4.95.3+8.2%15.011.9-20.7%9.4-37.3%273344+26.0%436+59.7%
16k11.311.9+5.3%21.715.0-30.9%10.4-47.5%189273+44.4%394+108.5%
32k25.226.4+4.8%35.922.0-38.7%12.3-65.7%114186+63.2%333+192.1%
162k2.32.5+8.7%14.411.2-22.2%9.7-32.6%569731+28.5%845+48.5%
8k9.49.8+4.3%22.615.6-31.0%11.8-47.8%362525+45.0%694+91.7%
16k20.321.4+5.4%36.923.0-37.7%12.8-65.3%222356+60.4%640+188.3%
+ +Table 4. Performance comparison with token pruning methods under various compression ratio. The prefetch and decoding latency are measured under the settings of batch size of 16, total input token of 8k, and the compression ratio of $10\%$ . + +
MethodsChatQA [26]MMBench-EN [24]MME [10]MMBench-Video [9]Latency (s)
50%10%1%50%10%1%50%10%1%50%10%1%PrefillDecoding
Full80.380.380.382.382.382.31585158515851.811.811.819.422.6
FastV [5]73.647.716.981.474.833.6157413787861.691.361.025.411.7
FasterVLM [39]75.550.118.881.477.026.5155914417311.711.421.175.211.3
IVTP [13]74.955.822.581.777.536.2156314698491.701.471.245.812.6
Ours80.479.976.482.382.382.31585158515851.791.721.609.811.8
+ +AirCache across different batch sizes and input token counts in terms of prefetch latency, decoding latency, and throughput. To facilitate testing, the text instruction token count is set to 64. As shown in Table 3, the proposed method significantly improves inference decoding speed under various inputs compared to the full cache. When compressing $50\%$ of the visual KV cache, the proposed method minimally impacts model performance and prefetch latency, while reducing decoding latency by $19\%$ to $39\%$ and increasing model throughput by $24\%$ to $63\%$ . When compressing $10\%$ of the visual KV cache, the proposed method slightly decreases model performance by less than $1\%$ , while further reducing decoding latency by $29\%$ to $66\%$ and increasing model throughput by $42\%$ to $192\%$ . These results demonstrate that our proposed method can maintain strong model performance while ensuring significant decoding acceleration, indicating the effectiveness of visual token importance assessment and hierarchical budget allocation. + +Compared with token pruning methods. In addition to KV cache compression, another important method for accelerating LVLMs inference is to prune visual tokens before the prefetch stage. Table 4 compares the proposed method with these type of acceleration methods, including FastV [5], FasterVLM [39], and IVTP [13]. A notable characteristic is that the token pruning methods exhibit acceleration effects during the prefetch stage as they preemptively prune visual tokens. In contrast, all KV cache compression methods, including the proposed approach, demonstrate prefetch stage inference speeds comparable to those of the full cache. However, comparisons across various evaluation sets clearly show that the proposed method achieves an absolute advantage, particularly on VQA datasets, even when retaining only $1\%$ of the visual tokens. In contrast, token pruning methods experience a significant performance + +Table 5. Ablation results on the components of the AirCache. + +
SettingChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
Continuous window (16)70.456.661.355.9
Continuous window (32)72.957.864.158.2
All text tokens72.258.465.757.0
Visual window (32)68.855.159.253.7
Average allocation72.257.569.962.4
Pyramid allocation69.654.955.852.6
Only w/ st74.259.871.164.9
Only w/ sk74.761.471.963.6
Ours76.462.573.267.1
+ +decline under these conditions. The prefetch stage facilitates sufficient interaction between visual and textual information, as the attention mechanism aggregates visual information to key text tokens. The proposed method effectively preserves these key text tokens during KV cache token pruning. Consequently, even when only $1\%$ of visual tokens are retained, the model maintains good inference performance. In contrast, most token pruning methods do not involve cross-modal interaction, resulting in direct visual information loss that cannot be transferred to the subsequent decoding stage in any form. This leads to insufficient and inaccurate visual guidance for generating, significantly degrading performance. + +# 4.3. Ablation Studies + +In this section, we evaluate our method under various settings to verify the proposed modules from several perspectives. The base model is LLaVA-OV-7B [16]. + +Elite Observation Window. We introduce three common types of observation windows to assess the effectiveness of the proposed elite observation window. The first type employs a sequence of continuous text tokens as the observation window for evaluating the significance of visual tokens. The second type utilizes all text tokens, whereas the third employs the continuous visual token from the last visual + +Table 6. Comparison results with different compression audiences, where * indicates uniform compression without distinguishing between visual tokens and text tokens. + +
Models50%10%5%1%
H2O* [42]74.6/407s46.5/559s32.0/733s27.8/1048s
H2O [42]79.8/323s77.4/251s74.0/239s71.0/215s
Ours*75.8/374s53.4/530s39.7/688s30.1/963s
Ours80.4/306s79.9/235s79.4/221s76.4/207s
+ +![](images/55cd4e0812dc1f5eebbe52655f686b89d3a96dd919ed450d52b2921ceda4c086.jpg) +Figure 4. Comparison on ChatQA [26] and TextVQA [31] by retaining only one visual token, which is selected based on the sorting of visual token importance scores using different methods. + +segment as a reference. As shown in the first column of Table 5, the proposed elite observation window yields the best performance compared to the other three aforementioned approaches. The method using only visual tokens performs significantly worse due to the absence of textual instruction guidance, highlighting the importance of activating cross-modal associations in KV cache compression. Additionally, compared to methods that directly use all or portions of continuous text token sequences without considering variations in visual information perception among text tokens, the proposed method selects key text tokens to enhance the consistency of the observation window. This selection ensures greater stability and accuracy in evaluating the significance of visual tokens. + +Layer-wise Token Budget Allocation Module. The second column of Table 5 presents the comparison using various budget allocation strategies across layers. Aside from the basic equal allocation, we also incorporate the strategy from PyramidKV [3], which allocates budgets in a descending ratio from shallow to deep layers. The results demonstrate that the proposed budget allocation, based on the strength and skewness of importance scores, delivers superior outcomes. Although PyramidKV [3] perform well in LLMs, its suboptimal results in LVLMs highlight the distinct characteristics of multimodal models and their heightened sensitivity to hierarchical budget allocation. An inappropriate allocation strategy is more likely to adversely affect the final results. We also investigate the separate use of strength and skewness for layer-wise budget quantification, revealing that their combination produces the best results. + +Visual Token Importance Evaluation. Figure 4 visually demonstrates the comparative effectiveness of various methods in evaluating the importance of visual tokens. To + +directly assess the efficacy of these methods in ranking visual token importance, we removed all non-essential modules and selected only one visual token at a time from the ranked sorting results, ordered from high to low importance. Theoretically, as the importance ranking of tokens declines, the model's performance is also expected to decrease. The graph clearly shows that the proposed method aligns more closely with the theoretical trend, where the visual token ranked as TOP1 achieves the highest metric, followed by a continuous and smooth decline. Notably, after around $10\%$ , the performance of the visual tokens approximates that of the case with all visual tokens removed. This indicates significant redundancy among visual tokens in the KV cache of LVLMs. The results in Table 1, which demonstrate that retaining over $10\%$ of the visual tokens limits performance decline to no more than $1\%$ , further corroborate this observation. In contrast, the performance derived from other methods exhibit significant noise, indicating instability and deficiencies in their importance assessment. + +Unified Compression or Vision-only Compression. Given that the KV cache in LVLMs encompasses both visual and textual modalities, in contrast to the single textual modality of LLMs, we compared Unified Compression, which does not differentiate between modal differences during KV cache compression, with Vision-only Compression, which compresses only visual tokens. As shown in Table 6, the results indicate that unified compression performs significantly worse than vision-only compression in terms of both inference performance and speed. Moreover, as the compression ratio increases, the inference time for unified compression actually continues to rise. The primary reason is that unified compression disrupts the effective representation of textual instructions, leading the model to produce incorrect outputs and exhibit a tendency for repetition, which in turn increases inference time. + +# 5. Conclusion + +In this work, we propose Activating Inter-modal Relevancy (AirCache), a novel KV cache compression for efficient LVLMs inference. By leveraging inter-modal relevancy between textual and visual tokens to guide KV cache eviction, we identified significant differences in the attention distribution of visual tokens corresponding to different textual tokens. Based on this finding, we propose an elite observation window with refined key text tokens, which provides a more stable and effective assessment of visual token importance. Furthermore, by exploring the diversity of importance score distribution across different layers, we propose to achieve comprehensive quantification in terms of distribution strength and skewness for adaptive layer-wise compression budget allocation. Extensive experiments across multiple benchmarks and LVLMs demonstrate the effectiveness of AirCache in KV cache compression. + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2 +[2] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5 +[3] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information funneling. arXiv preprint arXiv:2406.02069, 2024. 2, 8 +[4] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2 +[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 1, 2, 7 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 5, 6 +[7] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 2 +[8] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 6 +[9] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 5, 6, 7 +[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 5, 6, 7 +[11] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024. 1 + +[12] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2 +[13] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2024. 1, 2, 7 +[14] Chaoya Jiang, Haiyang Xu, Wei Ye, Qinghao Ye, Chen-liang Li, Ming Yan, Bin Bi, Shikun Zhang, Fei Huang, and Songfang Huang. Bus: Efficient and effective vision-language pre-training with bottom-up patch summarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2900–2910, 2023. 1 +[15] Jang-Hyun Kim, Junyoung Yeom, Sangwoo Yun, and Hyun Oh Song. Compressed context memory for online language model interaction. arXiv preprint arXiv:2312.03414, 2023. 2 +[16] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 4, 5, 6, 7 +[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[18] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1 +[19] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: Llm knows what you are looking for before generation. arXiv preprint arXiv:2404.14469, 2024. 1, 2, 5, 6 +[20] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 2 +[21] Akide Liu, Jing Liu, Zizheng Pan, Yefei He, Reza Haffari, and Bohan Zhuang. Minicache: Kv cache compression in depth dimension for large language models. Advances in Neural Information Processing Systems, 37:139997-140031, 2025. 1, 2 +[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1 +[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 7 + +[25] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 1, 2, 5, 6 +[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 4, 5, 6, 7, 8 +[27] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5, 6, 7 +[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5, 6, 7 +[29] Piotr Nawrot, Adrian Łancucki, Marcin Chochowski, David Tarjan, and Edoardo M Ponti. Dynamic memory compression: Retrofitting llms for accelerated inference. arXiv preprint arXiv:2403.09636, 2024. 2 +[30] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 1 +[31] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 7, 8 +[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2025. 1 +[33] Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, and Panpan Xu. VI-cache: Sparsity and modality-aware kv cache compression for vision-language model inference acceleration. arXiv preprint arXiv:2410.23317, 2024. 2 +[34] Ao Wang, Hui Chen, Jianchao Tan, Kefeng Zhang, Xunliang Cai, Zijia Lin, Jungong Han, and Guiguang Ding. Prefixkv: Adaptive prefix kv cache is what vision instruction-following models need for efficient generation. arXiv preprint arXiv:2412.03409, 2024. 1, 2, 5, 6 +[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6 +[36] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 1 + +[37] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6 +[38] Yifei Yang, Zouying Cao, Qiguang Chen, Libo Qin, Dongjie Yang, Hai Zhao, and Zhi Chen. Kvsharer: Efficient inference via layer-wise dissimilar kv cache sharing. arXiv preprint arXiv:2410.18517, 2024. 2 +[39] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 1, 2, 7 +[40] Yuxin Zhang, Yuxuan Du, Gen Luo, Yunshan Zhong, Zhenyu Zhang, Shiwei Liu, and Rongrong Ji. Cam: Cache merging for memory-efficient llms inference. In *Forty-first International Conference on Machine Learning*, 2024. 2 +[41] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5 +[42] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 4, 5, 6, 8 \ No newline at end of file diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/images.zip b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..73073edd2f2b6132f103c0afb373a0c7e7c9357c --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89b7ce12eabc13b4cf4c8553955a87666adc4c17beeff1b5e11155ad1962ab1b +size 583998 diff --git a/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/layout.json b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a03f0bc2ca777c7c9f71c321ffd449f9b3e274c --- /dev/null +++ b/2025/AirCache_ Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference/layout.json @@ -0,0 +1,7500 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 522, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 522, + 140 + ], + "type": "text", + "content": "AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 161, + 455, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 161, + 455, + 190 + ], + "spans": [ + { + "bbox": [ + 137, + 161, + 455, + 190 + ], + "type": "text", + "content": "Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie\\*, Hao Wang Alibaba Group" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 191, + 511, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 191, + 511, + 202 + ], + "spans": [ + { + "bbox": [ + 82, + 191, + 511, + 202 + ], + "type": "text", + "content": "zhouwan.hk, zh372956, bochen.wbc, yx150449, xiezhen.xz, qiao.wh@alibaba-inc.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "text", + "content": "Recent advancements in Large Visual Language Models (LVLMs) have gained significant attention due to their remarkable reasoning capabilities and proficiency in generalization. However, processing a large number of visual tokens and generating long-context outputs impose substantial computational overhead, leading to excessive demands for key-value (KV) cache. To address this critical bottleneck, we propose AirCache, a novel KV cache compression method aimed at accelerating LVLMs inference. This work systematically investigates the correlations between visual and textual tokens within the attention mechanisms of LVLMs. Our empirical analysis reveals considerable redundancy in cached visual tokens, wherein strategically eliminating these tokens preserves model performance while significantly accelerating context generation. We introduce an elite observation window for assessing the importance of visual components in the KV cache, focusing on stable inter-modal relevancy modeling with enhanced multi-perspective consistency. Additionally, we develop an adaptive layer-wise budget allocation strategy that capitalizes on the strength and skewness of token importance distribution, showcasing superior efficiency compared to uniform allocation. Comprehensive evaluations across multiple LVLMs and benchmarks demonstrate that our method achieves comparable performance to the full cache while retaining only " + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "text", + "content": " of visual KV cache, thereby reducing decoding latency by " + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 46, + 255, + 289, + 616 + ], + "type": "text", + "content": " across various batch size and prompt length of inputs. Notably, as cache retention rates decrease, our method exhibits increasing performance advantages over existing approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 638, + 128, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 696 + ], + "type": "text", + "content": "The past few years have witnessed remarkable advancements in Large Visual-Language Models (LVLMs) [1, 6, 16, 22, 23, 32, 35, 36], both in research and prac" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 229, + 420, + 346 + ], + "blocks": [ + { + "bbox": [ + 307, + 229, + 420, + 346 + ], + "lines": [ + { + "bbox": [ + 307, + 229, + 420, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 420, + 346 + ], + "type": "image", + "image_path": "a467d1151f33fa4cc67bc19015d697066262c2a76b64e4fdc190bf02c224b4ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 353, + 547, + 430 + ], + "lines": [ + { + "bbox": [ + 305, + 353, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 547, + 430 + ], + "type": "text", + "content": "Figure 1. Illustration of our motivation. (a) An example of attention maps during the decoding process of LLaVA-OV-7B [16]. (b) Partial visual attention maps obtained using all text tokens as the observation window. (c) Using the last 16 text tokens as the observation window. (d) Using the last 16 visual tokens as the observation window, demonstrating more consistent attention distribution." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 422, + 229, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 422, + 229, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 422, + 229, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 422, + 229, + 545, + 350 + ], + "type": "image", + "image_path": "47b792cdd7f3154c8287184933cc516431f3b598ea0807c48f077f6b9332380e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 446, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 446, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 446, + 545, + 567 + ], + "type": "text", + "content": "tical applications. Although LVLMs exhibit increasingly advanced visual processing capabilities, such as handling high-resolution, multiple images, and video sequences, these improvements entail substantial computational costs. The exponential increase in visual tokens, coupled with the demands of long-context generation tasks, results in unsustainable memory overhead from key-value (KV) cache storage [30]. This not only significantly increases GPU memory consumption but also severely degrades computational efficiency due to heightened memory bandwidth contention." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Recent approaches [5, 13, 14, 19, 39, 42] aim to prune visual tokens during inference, which can be categorized into token pruning and KV cache compression. Token pruning [11, 13, 14, 18, 20, 39] involves reducing visual tokens during the prefetch phase, which decreases both the number of tokens processed in subsequent transformer layers and the tokens stored in the KV cache. Although it significantly enhances inference speed, the aggressive elimination of visual tokens during prefetch leads to a substantial loss of visual information, severely degrading model performance. KV cache compression [19, 21, 25, 34, 42] reduces token solely during the decoding phase by pruning stored data in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 139, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23958" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "type": "text", + "content": "the KV cache. This technique has been extensively validated in LLMs. Since all tokens undergo a complete forward pass and the causal attention mechanism in LLMs differentiates the importance distribution among tokens, selectively deleting certain tokens has minimal impact on model performance. These methods typically assess the importance of visual tokens by evaluating their strength or sparsity within the attention maps computed under the causal attention mechanism. Given the strict unidirectionality of causal attention, using cross-modality attention scores between instruction tokens and visual tokens provides a more comprehensive evaluation of visual token importance than using full attention scores. Moreover, we observe significant variations in the distribution of visual token strength across different layers, indicating that uniformly allocating the budget is not the optimal strategy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "spans": [ + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose AirCache, a novel KV cache compression method designed for efficient LVLMs inference. AirCache comprises two main components: visual token importance scoring and layerwise KV cache budget allocation. To effectively evaluate the importance of visual tokens in the KV cache, we introduce an elite observation window that selects critical instruction tokens by leveraging self-attention scores among instruction tokens rather than using all or continuous local instruction tokens. As illustrated in Figure 1, compared to mainstream methods that assess visual token importance based on observation windows guided by all or partial text instruction tokens, the elite observation window demonstrates stronger consistency. Most text tokens within the elite observation window tend to provide more similar evaluations for the same visual token, thereby enhancing the effectiveness and stability of subsequent voting-based ranking. We further propose to quantify the compression budget for different layers from two perspectives. The first aspect is the emphasis a layer places on visual information, measured by the sum of attention allocated to all visual tokens, as well as the strength of importance score distribution. Additionally, analysis of attention distributions across layers reveals a distinct head effect, where a few visual tokens receive high importance scores, while the majority exhibit mediocre and average importance characteristics. Our experimental results show that preserving only " + }, + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "text", + "content": " of the important visual KV cache results in less than " + }, + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 49, + 267, + 286, + 661 + ], + "type": "text", + "content": " average model performance degradation, corroborating the significant head effect in the importance distribution of visual tokens. Combining these two aspects, we allocate the budget for visual KV cache based on both the strength and skewness of importance scores distributions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 666, + 286, + 712 + ], + "type": "text", + "content": "Our primary contributions include: (1) We conduct a comprehensive and detailed exploration of the differences in inter-modal relevancy patterns based on attention interaction in evaluating visual KV cache importance. We in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 308, + 72, + 544, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 191 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 191 + ], + "type": "text", + "content": "troduce an elite observation window with carefully selected key textural instruction tokens to achieve more effective and stable importance evaluations. (2) We propose to quantify the differentiated compression budget allocation across different layers using the strength and skewness of the importance scores distribution, based on the prominence of visual information and the efficacy of attention allocation. (3) Experimental results on widely used LVLMs and benchmark datasets demonstrate that the proposed method achieves superior performance compared to other existing methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 308, + 203, + 391, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 203, + 391, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 203, + 391, + 215 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 223, + 544, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 223, + 544, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 223, + 544, + 425 + ], + "type": "text", + "content": "Visual Token Pruning. Visual token pruning [4, 5, 12, 13, 20, 39] accelerates model inference by reducing visual tokens during the prefill phase. BLIP-2 [17] and Qwen-VL [1] employ cross-attention with learnable query embeddings to aggregate input visual tokens, while HoneyBee [4] and MobileVLM [7] use convolutional neural networks. These approaches necessitate model retraining, which significantly increases migration costs. In contrast, methods like [5, 39] introduce no extra parameters but leverage attention maps to score and evict visual tokens based on importance. IVTP [13] implements two-stage pruning by deleting tokens in both Vision Transformer (ViT) and LLM, enabling text-guided pruning through instruction integration. However, it is important to note that visual token pruning inevitably leads to visual information loss by discarding tokens prior to or within early layers, thereby substantially degrading model performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 427, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 544, + 712 + ], + "type": "text", + "content": "KV cache compression. KV cache strategies accelerate inference by storing the key and value states of previous tokens, thereby avoiding redundant computations during decoding. However, as the number of tokens increases, the pressure on memory and bandwidth also rises, limiting the model's application in long content understanding or generation. To address this challenge, mainstream methods primarily focus on two directions: intra-layer KV cache pruning [15, 29, 40] and inter-layer compression budget allocation [21, 38]. H2O [42] and SnapKV [19] propose evicting tokens in the KV cache using global and local cumulative attention scores, while PyramidKV [3] constructs a pyramid-like hierarchical budget allocation to achieve more refined KV cache compression management. The input to LVLMs is a cross-modal construct primarily composed of visual tokens with auxiliary text tokens, necessitating additional consideration of the impact of different modality differences when extending KV cache compression to LVLMs. Elastic [25] merges non-critical tokens with important tokens, PrefixKV [34] introduces an adaptive layer-wise KV retention recipe based on binary search for maximal preservation of contextual information. VL-Cache [33] utilizes visual token sparsity for layer-wise budget allocation and token eviction through dedicated scoring policies." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 757 + ], + "type": "text", + "content": "23959" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 173, + 205 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 173, + 205 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 173, + 205 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 173, + 205 + ], + "type": "image", + "image_path": "d04ec30050357743f6fe7d5491a37dd44d7667b66434e29671518682c6845f27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 276 + ], + "type": "text", + "content": "Figure 2. The left panel illustrates the inference workflow of standard LVLMs. The proposed method integrates during the KV cache storage process following the prefetch stage, maintaining compatibility with mainstream LVLMs architecture. The right panel presents the overview of AirCache. We first employ self-attention across the instruction text tokens to identify the key text tokens, establishing an elite observation window for assessing visual token significance. We further reallocate the compression budgets between layers based on the strength and skewness of the importance distribution. Finally, pruning is applied to the KV cache according to the importance ranking of visual tokens in the KV Cache and the compression budget to obtain the final compressed KV cache." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 265, + 205 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 265, + 205 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 265, + 205 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 265, + 205 + ], + "type": "image", + "image_path": "85ae4793679c26c7c20cf9eff09e7219b954eec717ea08769c7320fda8a0cb43.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 266, + 70, + 441, + 205 + ], + "blocks": [ + { + "bbox": [ + 266, + 70, + 441, + 205 + ], + "lines": [ + { + "bbox": [ + 266, + 70, + 441, + 205 + ], + "spans": [ + { + "bbox": [ + 266, + 70, + 441, + 205 + ], + "type": "image", + "image_path": "ee244de9c3edf0a2bba9c6b1f91b7beeb1466967113cd9d87b271b7410bf3e39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 441, + 70, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 441, + 70, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 441, + 70, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 441, + 70, + 545, + 205 + ], + "type": "image", + "image_path": "4a3d2980077e8266a1e24e7c36890548bd4d3c7be20bcb53ec7916cf964c7611.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 284, + 129, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 284, + 129, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 284, + 129, + 297 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 303, + 127, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 127, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 127, + 316 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "text", + "content": "Similar to most LLMs, the inference process of LVLMs is primarily divided into prompt prefill and token decoding. Prompt Prefill. Given the hidden states " + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "text", + "content": " of the input prompt, where " + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "text", + "content": " is the total number of visual and text prompt tokens, and " + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 322, + 288, + 418 + ], + "type": "text", + "content": " denotes hidden dimension. We simplify the expression by omitting the indices of the head and layer of the hidden states. The query, key, and value states of the attention block can be represented as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 422, + 287, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 422, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 74, + 422, + 287, + 436 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {Q}, \\quad \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {K}, \\quad \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {V}, \\tag {1}", + "image_path": "999e70259ea8f155a69b9d6a331e7d0ee3df971be025c35f0cd25380618ab84c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_Q, \\mathbf{W}_K, \\mathbf{W}_V \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 46, + 439, + 287, + 487 + ], + "type": "text", + "content": " are learnable projection matrices. Following the completion of the attention interaction, the key and value are stored in the KV cache to support subsequent token generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "text", + "content": "Token Decoding. In this stage, the model efficiently generates new tokens by iteratively utilizing and updating the KV cache. Specifically, during the " + }, + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "text", + "content": "th iteration, we only need to compute the key and value associated with the newly generated token " + }, + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 46, + 488, + 287, + 594 + ], + "type": "text", + "content": ". By employing cache indexing, the model avoids recomputing the key and value for tokens that have already been processed. Subsequently, the KV cache is updated with the newly computed key and value for the latest token:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 601, + 287, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 287, + 613 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = \\operatorname {C o n c a t} \\left(\\mathbf {K}, \\mathbf {x} _ {i} \\mathbf {W} _ {K}\\right), \\mathbf {V} = \\operatorname {C o n c a t} \\left(\\mathbf {V}, \\mathbf {x} _ {i} \\mathbf {W} _ {V}\\right). \\tag {2}", + "image_path": "096e7c9f5bea0e0be6863870330c70a060b93255374a73d0134cafeb8ba19e39.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 713 + ], + "type": "text", + "content": "Although the KV cache mitigates the issue of redundant computation during the decoding process, its characteristic of linear growth in computational requirements relative to the length of the input sequence poses challenges regarding latency and memory usage, especially in scenarios involving longer inputs or outputs. This issue is further exacerbated as recent LVLMs tend to prioritize high-resolution, high-frame-rate visual inputs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 285, + 373, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 285, + 373, + 296 + ], + "spans": [ + { + "bbox": [ + 306, + 285, + 373, + 296 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 303, + 545, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 447 + ], + "type": "text", + "content": "The AirCache is primarily divided into two components. The first component evaluates the significance of visual tokens based on the elite observation window. Section 3.3 details how to optimize the relevancy between visual and language modalities to guide the compression of the visual segment in the KV cache. The second component quantifies the KV cache budget requirements across various layers by analyzing the overall strength and skewness of the importance score distribution, thereby facilitating dynamic layerwise KV cache budget allocation, as elaborated in Section 3.4. The architecture and algorithm underlying the method are illustrated in Figure 2 and Algorithm 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 454, + 453, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 453, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 453, + 466 + ], + "type": "text", + "content": "3.3. Elite Observation Window" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 689 + ], + "type": "text", + "content": "In LVLMs, the presence of multiple modalities complicates the selection of observation windows, suggesting that adhering to the relevant practices established in LLMs may not be optimal. For instance, as illustrated in Figure 1, the visual token importance measures derived from the continuous last segment of the prompt exhibit poor consistency. This variability indicates that different text tokens within the observation window yield significantly diverse visual tokens importance distributions. Consequently, the hit rate statistics derived from the voting mechanism are prone to substantial noise, and the disparities among modalities may further amplify this noise interference. Based on this consideration, we propose refining the observation windows by incorporating carefully selected text tokens. By utilizing the attention hit rate within the text modality, we aim to enhance the consistency of visual token evaluation, thereby improving the performance of KV cache compression. To this end, we reorganize the hidden states of the input prompt by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 344, + 700, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 700, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 344, + 700, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {X} = \\operatorname {C o n c a t} \\left(\\mathbf {X} _ {v}, \\mathbf {X} _ {t}\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t}\\right) \\times D}, \\tag {3}", + "image_path": "c3ce400666502f46c6f18dd461092f63e131f6acb9b6ba892fc159b52c309deb.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23960" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 69, + 233, + 180 + ], + "blocks": [ + { + "bbox": [ + 85, + 69, + 233, + 180 + ], + "lines": [ + { + "bbox": [ + 85, + 69, + 233, + 180 + ], + "spans": [ + { + "bbox": [ + 85, + 69, + 233, + 180 + ], + "type": "image", + "image_path": "92ec940688a6a44686f2e7900d795bace6e62108207312c001632605d6c14ed2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 182, + 164, + 191 + ], + "lines": [ + { + "bbox": [ + 154, + 182, + 164, + 191 + ], + "spans": [ + { + "bbox": [ + 154, + 182, + 164, + 191 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 233, + 70, + 373, + 180 + ], + "blocks": [ + { + "bbox": [ + 233, + 70, + 373, + 180 + ], + "lines": [ + { + "bbox": [ + 233, + 70, + 373, + 180 + ], + "spans": [ + { + "bbox": [ + 233, + 70, + 373, + 180 + ], + "type": "image", + "image_path": "4ee76410aaa113aec42e8a0152bcaa354c004f2cf32332b0159a0b85871e011a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 182, + 309, + 191 + ], + "lines": [ + { + "bbox": [ + 298, + 182, + 309, + 191 + ], + "spans": [ + { + "bbox": [ + 298, + 182, + 309, + 191 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 194, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 545, + 239 + ], + "type": "text", + "content": "Figure 3. (a) The strength of the distribution of visual token importance scores across different layers, which is defined as the sum of all visual token importance scores. (b) Performance with selecting only one visual token each time based on the sorted results of visual token importance scores, the index of 0 indicates the setting of removing all visual tokens. (c) The distribution of visual token importance scores across different layers, with only selected layers displayed for clarity. All results are derived from LLaVA-OV-7B [16] on the ChatQA [26]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 374, + 70, + 507, + 174 + ], + "blocks": [ + { + "bbox": [ + 374, + 70, + 507, + 174 + ], + "lines": [ + { + "bbox": [ + 374, + 70, + 507, + 174 + ], + "spans": [ + { + "bbox": [ + 374, + 70, + 507, + 174 + ], + "type": "image", + "image_path": "e20ca144141a19854a637bca3fc3640d2a7b0c2a106349a95f6a1de0bcdbc7a7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 182, + 446, + 191 + ], + "lines": [ + { + "bbox": [ + 435, + 182, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 435, + 182, + 446, + 191 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_v" + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_t" + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": " represent the hidden states corresponding to visual tokens and instruction text tokens, respectively. " + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "inline_equation", + "content": "N_v" + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "inline_equation", + "content": "N_t" + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": " denote the respective quantities of these tokens, satisfying the condition " + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "inline_equation", + "content": "N_v + N_t = N" + }, + { + "bbox": [ + 46, + 248, + 286, + 319 + ], + "type": "text", + "content": ". For convenience, the system prompts are omitted. We compute the text unimodal attention matrix composed of instruction text tokens as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 86, + 326, + 287, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 326, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 86, + 326, + 287, + 353 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {t t} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t} \\mathbf {K} _ {t} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t} \\times N _ {t}}, \\tag {4}", + "image_path": "5986d2bcab7a30fa6471170273c32880b5883305e46bcc276f3404a0bdbb7a32.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_t" + }, + { + "bbox": [ + 46, + 357, + 287, + 441 + ], + "type": "text", + "content": " denote the query and key of the text tokens, respectively. Considering the characteristics of causal attention, we select the last token of the instruction text as reference and filter out key text tokens that receive high attention scores based on their interactions with other text tokens. Consistent with [42], we employ a relevance threshold to regulate the sparsity of key text tokens:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 447, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 460 + ], + "type": "interline_equation", + "content": "k = \\{j \\mid \\mathbf {A} _ {t t} [ N _ {t} - 1, j ] \\geq \\alpha \\cdot \\max \\mathbf {A} _ {t t} [ N _ {t} - 1,: ] \\}, \\tag {5}", + "image_path": "d54962da0968e7776c2dd0642e63d3ba1b0aa0c64a64c8e6fc3abbd1877871fc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,1]" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": " represents the relevance threshold, and " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "j\\in \\{0,1,\\dots ,N_t - 1\\}" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": " is the index of the text token. When " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": " is equal to 0, all text tokens are retained, and when " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": " is equal to 1, only the text token with the highest score is preserved. We redefine the key for the aggregated visual tokens and key text tokens as " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{vt_k}" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": ", along with the query corresponding to the key text tokens " + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{t_k}" + }, + { + "bbox": [ + 46, + 466, + 287, + 550 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 113, + 555, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 555, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 113, + 555, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t _ {k}} = \\mathbf {Q} _ {t} [ k ] \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times D}, \\tag {6}", + "image_path": "a3cd576ecf0ac5fe737180431e961e5a2a1cc871a3a53398e190afd238333d18.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 569, + 255, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 569, + 255, + 582 + ], + "spans": [ + { + "bbox": [ + 67, + 569, + 255, + 582 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {v t k} = \\operatorname {C o n c a t} \\left(\\mathbf {K} _ {v}, \\mathbf {K} _ {t} [ k ]\\right) \\in \\mathbb {R} ^ {\\left(N _ {v} + N _ {t _ {k}}\\right) \\times D},", + "image_path": "18ca4a60dd80957a1310c0445cdf6ecb16c5f96cf40d05d9015ac47c7213813d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_v" + }, + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "text", + "content": " represents the key of the visual token, and " + }, + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "inline_equation", + "content": "N_{t_k}" + }, + { + "bbox": [ + 46, + 587, + 287, + 635 + ], + "type": "text", + "content": " denotes the number of selected key text tokens. Furthermore, the attention matrix between the visual tokens and the key text tokens can be computed as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 640, + 287, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 287, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 287, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {v t _ {k}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {t _ {k}} \\mathbf {K} _ {v t _ {k}} ^ {T}}{\\sqrt {D}}\\right) \\in \\mathbb {R} ^ {N _ {t _ {k}} \\times \\left(N _ {v} + N _ {t _ {k}}\\right)}. \\tag {7}", + "image_path": "3be13edafdf79a6e5b36b67672122495bc91874e74de8ecc7f583564fc2d0a94.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The elite observation window, comprising only these key text tokens, offers a more comprehensive and focused semantic representation compared to those formed by all text" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 248, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 248, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 248, + 545, + 308 + ], + "type": "text", + "content": "tokens or fixed continuous sequences. It provides enhanced stability and accuracy in assessing the importance of visual tokens. The final importance scores of the visual tokens can be obtained by performing average pooling of the attention matrix along the text dimension:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 318, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 318, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 361, + 318, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\mathbf {I} _ {v} = \\frac {1}{N _ {t _ {k}}} \\sum_ {j = 0} ^ {N _ {t _ {k}} - 1} \\mathbf {A} _ {v t _ {k}} [ j,: N _ {v} ]. \\tag {8}", + "image_path": "3de3ffeda61305f82cdc32c3902b7fb579d80822cf16d22c7958b4c749542a62.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 363, + 545, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 545, + 434 + ], + "type": "text", + "content": "By sorting the importance scores, the KV cache of visual tokens with lower rankings can be discarded in accordance with the compression budget. Furthermore, the elite observation window we employ exhibits reduced computational complexity, and the additional overhead of key text token selection is negligible in visual token-heavy LVLMs." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 443, + 518, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 443, + 518, + 456 + ], + "spans": [ + { + "bbox": [ + 306, + 443, + 518, + 456 + ], + "type": "text", + "content": "3.4. Layer-wise KV Cache Budget Allocation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": "We further investigate the characteristics of the visual token importance score distribution based on the elite observation window across various layers. Specifically, we consider the sum of the importance scores of all visual tokens as an indicator of the layer's attention to visual information. As illustrated in Figure 3(a), there is a significant disparity in the attention given to visual tokens across different layers. This observation highlights the necessity of performing layer-wise token budget allocation, where layers dedicating more attention to visual information merit a larger token budget. As shown in Figure 3(b), we select one visual token at a time based on the importance scores. It can be observed that only approximately " + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 462, + 545, + 713 + ], + "type": "text", + "content": " of all visual tokens have a significantly positive impact on the final result. This suggests a head effect in the model's attention distribution over visual tokens, with a small subset receiving higher attention, while the majority have lower and more balanced scores. Extending to the importance distribution across different layers in Figure 3(c), it shows significant differences in importance distribution among the layers, particularly regarding the head effect of visual tokens. Consequently," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23961" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "we aim to allocate a larger token budget to the layers that demonstrate a pronounced head effect in their visual tokens importance distribution. This indicates a more accurate and nuanced understanding of the visual content by those layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 287, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 168 + ], + "type": "text", + "content": "As aforementioned, we quantify the allocation of visual token budget across different layers from two perspectives: the strength and skewness of the importance distribution. The strength of the importance distribution is calculated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 176, + 287, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 176, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 132, + 176, + 287, + 209 + ], + "type": "interline_equation", + "content": "s _ {t} = \\sum_ {i = 0} ^ {N _ {v} - 1} \\mathbf {I} _ {v} [ i ]. \\tag {9}", + "image_path": "776a6181bf48cc30c9da4498b8f43fed49ec812dd8667c6a180a12030c1bef23.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 213, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 288, + 307 + ], + "type": "text", + "content": "We exclude the attention between text tokens and instead aggregate the attention scores between text tokens and all visual tokens to quantify the importance of visual information in each layer. The greater the importance, the more budget is allocated. Meanwhile, we use the skewness of importance distribution to assess how effectively each layer aligns with the distribution characteristics of attention to visual tokens:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 313, + 287, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 287, + 346 + ], + "type": "interline_equation", + "content": "s _ {k} = \\frac {N _ {v}}{\\left(N _ {v} - 1\\right) \\left(N _ {v} - 2\\right)} \\sum_ {i = 1} ^ {N _ {v}} \\left(\\frac {\\mathbf {I} _ {v} [ i ] - \\mu_ {I _ {v}}}{\\sigma_ {I _ {v}}}\\right) ^ {3}, \\tag {10}", + "image_path": "36465846eb22f0c51b691af44cf9a3c3473e9e2afe2d4c32d9fd49a9c4fc8226.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mu_{I_v}" + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\sigma_{I_v}" + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "text", + "content": " represent the mean and standard deviation of the importance as well as attention scores distribution of visual tokens, which essentially reflect the distribution of attention scores. Ultimately, the reallocated token budget integrates these factors into the original budget " + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 352, + 287, + 411 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 411, + 287, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 411, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 129, + 411, + 287, + 433 + ], + "type": "interline_equation", + "content": "\\hat {r} = \\frac {1}{2} \\left(s _ {t} ^ {\\prime} + s _ {k} ^ {\\prime}\\right) r. \\tag {11}", + "image_path": "dc077f4e2210cbef5012fe3253940f814e1e564eebfe8a72baa78b4d75178f11.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "inline_equation", + "content": "s_k'" + }, + { + "bbox": [ + 46, + 437, + 287, + 463 + ], + "type": "text", + "content": " represent the normalized strength and skewness across different layers, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 472, + 127, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 127, + 486 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 127, + 486 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 491, + 97, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 97, + 505 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 97, + 505 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Models. The selection of LVLMs is primarily based on two criteria: model architecture and parameter size. In terms of architecture, the objective is to validate the effectiveness of the proposed method across various architectures, including LLaVA-OV-7B [16], InternVL2-8B [6], and Qwen2-VL-7B [35]. These models exhibit significant differences in visual token encoding, ViT, and LLM. For example, both LLaVA-OV-7B [16] and InternVL2-8B [6] utilize the AnyRes strategy with distinct ViT to increase the number of visual tokens, while Qwen2-VL-7B [35] supports native resolution with a large quantity of visual tokens. In terms of parameter size, models of InternVL2 series are selected for their diverse range of parameter scales, including InternVL2-1B [6], InternVL2-4B [6], and InternVL2-26B [6]. Due to space constraints, results for this section are provided in the appendix. Additionally, LLaVA-Video-7B-Qwen2 [41] serves as the base model for the video tasks evaluation." + } + ] + } + ], + "index": 10 + }, + { + "type": "code", + "bbox": [ + 309, + 85, + 545, + 365 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 517, + 83 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 517, + 83 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 517, + 83 + ], + "type": "text", + "content": "Algorithm 1 AirCache for KV Cache Compression." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "lines": [ + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": "1: Input: Total cache budget " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", key, query, value of text prompt tokens " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "V_{t} \\in \\mathbb{R}^{N_{t} \\times D}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", key, query, value of visual prompt tokens " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "K_{v}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "Q_{v}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "V_{v} \\in \\mathbb{R}^{N_{v} \\times D}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", number of layers L. \n2: for layer " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "l \\gets 0" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "A_{tt} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t} \\mathbf{K}_{t}^{T}}{\\sqrt{D}}\\right)" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n4: Obtain the index of key text tokens " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " using Eq. (5). \n5: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "Q_{t_{k}}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "K_{vt_{k}} \\gets Q_{t}[k]" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\text{Concat}(K_{v}, K_{t}[k])" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{vt_{k}} \\gets \\text{Softmax}\\left(\\frac{\\mathbf{Q}_{t_{k}} \\mathbf{K}_{vt_{k}}^{T}}{\\sqrt{D}}\\right)" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{v} \\gets \\frac{1}{N_{t_{k}}} \\sum_{j=0}^{N_{t_{k}}-1} \\mathbf{A}_{vt_{k}}[j,: N_{v}] \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Importance Scores \n8: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "s_{t} \\gets \\sum_{i=0}^{N_{v}-1} \\mathbf{I}_{v}[i]" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "s_{k} \\gets \\frac{N_{v}}{(N_{v}-1)(N_{v}-2)} \\sum_{i=1}^{N_{v}}\\left(\\frac{\\mathbf{I}_{v}[i]-\\mu_{I_{v}}}{\\sigma_{I_{v}}}\\right)^{3} \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Skewness \n10: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "C^{l} \\gets (\\{(K_{v}, K_{t}), (V_{v}, V_{t}), (\\mathbf{I}_{v}, s_{t}, s_{k})\\})" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n11: end for \n12: for layer " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "l \\gets 0" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " do \n13: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "K_{v}, K_{t}, V_{v}, V_{t}, \\mathbf{I}_{v}, s_{t}, s_{k} \\gets C^{l}" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n14: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "s_{t}', s_{k}' \\gets s_{t}, s_{k} \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Normalized Across Different Layers \n15: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\hat{r} = \\frac{1}{2}(s_{t}' + s_{k}') r \\quad \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Obtain Layer-wise Budget \n16: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "K' \\gets \\text{Concat}(K_{v}(\\mathbf{I}[N_{v} \\times \\hat{r}]), K_{t}) \\quad \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Compress Key \n17: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "V' \\gets \\text{Concat}(V_{v}(\\mathbf{I}[V_{v} \\times \\hat{r}]), V_{t}) \\quad \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Compress Value \n18: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "C^{l} \\gets (\\{K', V'\\})" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " \n19: end for \n20: Return: " + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "inline_equation", + "content": "C \\triangleright" + }, + { + "bbox": [ + 309, + 85, + 545, + 365 + ], + "type": "text", + "content": " Compressed KV Cache" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "algorithm" + }, + { + "bbox": [ + 304, + 389, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 593 + ], + "type": "text", + "content": "Datasets. The performance is primarily evaluated across four types of datasets: 1) Visual Question Answering (VQA) datasets, which include ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31]. Since KV cache compression mainly affects the decoding stage, VQA datasets that require longer output tokens are preferred for evaluating the method's efficacy. The similarity between generated and reference answers is typically measured by the Average Normalized Levenshtein Similarity (ANLS) metric [2]. 2) Multiple-choice question datasets, such as MMBench-EN [24], involves selecting the correct option from several choices, with accuracy serving as the ultimate metric. 3) Judgment datasets like MME [10], also utilize judgment accuracy as the evaluation criterion. 4) Video VQA datasets, such as MMBench-Video [9], measure the alignment between generated and reference answers using metrics provided by a large model evaluator." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Baselines. The comparison focuses on two main categories of methods. The first category encompasses classic and prominent KV cache compression techniques in LLMs, such as H2O [42], and SnapKV [19]. The second category comprises methods specifically optimized for LVLMs, namely PrefixKV [34] and Elastic Cache [25]. H2O [42] employs a full-range approach, while SnapKV [19] adopts a window-based strategy using textual tokens as the observation window for evaluating token importance. PrefixKV [34] utilizes the Lorenz curve to characterize the impor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23962" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 80, + 544, + 270 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 541, + 79 + ], + "type": "text", + "content": "Table 1. The comparison of the KV cache compression methods on multimodal VQA benchmarks. The best result is highlighted in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 80, + 544, + 270 + ], + "lines": [ + { + "bbox": [ + 49, + 80, + 544, + 270 + ], + "spans": [ + { + "bbox": [ + 49, + 80, + 544, + 270 + ], + "type": "table", + "html": "
ModelsMethodsChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
50%10%5%1%50%10%5%1%50%10%5%1%50%10%5%1%
LLaVA-OV-7B [16]Full80.380.380.380.366.166.166.166.187.087.087.087.076.076.076.076.0
H2O [42]79.877.474.071.065.459.254.252.086.474.260.855.375.570.167.960.4
Elastic [25]80.077.976.571.065.360.557.252.086.674.260.855.375.572.068.760.2
PrefixKV [34]80.178.276.870.965.461.158.051.986.680.574.755.475.872.770.060.2
SnapKV [19]80.179.378.372.966.164.263.057.886.984.480.664.175.873.469.558.2
Ours80.479.979.476.466.065.764.662.586.885.583.373.276.075.373.467.1
InternVL2-8B [6]Full82.282.282.282.273.073.073.073.091.091.091.091.077.777.777.777.7
H2O [42]81.981.679.376.872.370.665.161.290.579.470.348.977.170.066.059.6
Elastic [25]81.881.079.476.872.367.065.160.890.579.470.353.577.070.266.059.0
PrefixKV [34]81.980.779.376.772.367.465.361.090.679.570.255.177.170.366.258.6
SnapKV [19]82.280.479.575.072.872.368.464.990.990.183.571.177.473.971.061.7
Ours82.281.781.078.473.272.671.868.291.090.088.578.577.577.075.468.6
Qwen2-VL-7B [35]Full83.283.283.283.276.276.276.276.293.893.893.893.884.484.484.484.4
H2O [42]83.278.975.066.475.465.761.055.190.582.370.748.984.275.668.254.5
Elastic [25]83.279.775.468.075.567.162.355.593.583.071.949.284.276.768.854.4
PrefixKV [34]83.279.475.968.075.767.962.755.793.783.772.849.884.276.969.454.3
SnapKV [19]83.381.680.069.176.374.973.066.293.787.287.171.284.383.174.660.2
Ours83.482.380.875.976.275.274.570.793.892.990.978.384.483.480.968.6
", + "image_path": "de445ad5050c47fbf0f5c463223655330d1d295cf7fddf2eea61b2c40b3f047c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 299, + 285, + 366 + ], + "blocks": [ + { + "bbox": [ + 47, + 278, + 286, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 278, + 286, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 286, + 299 + ], + "type": "text", + "content": "Table 2. Performance comparison on other types datasets. Best-performing results are highlighted in bold." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 299, + 285, + 366 + ], + "lines": [ + { + "bbox": [ + 49, + 299, + 285, + 366 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 285, + 366 + ], + "type": "table", + "html": "
MethodsMMBench-EN [24]MME [10]MMBench-Video [9]
50%10%1%50%10%1%50%10%1%
Full82.382.382.31585158515851.811.811.81
H2O [42]0.820.820.821585158515851.661.631.47
Elastic [25]0.820.820.821585158515851.681.621.51
PrefixKV [34]0.820.820.821585158515851.721.681.50
SnapKV [19]0.820.820.821585158515851.751.701.55
Ours0.820.820.821585158515851.801.781.67
", + "image_path": "67059182a9c3795f7c041d3c6fb5af57a8f1be7515d9c9e53eeb49e8919b85b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 460 + ], + "type": "text", + "content": "tance distribution across different layers, guided by the retained important KV matrix. Elastic Cache [25] employs a fixed-point elimination strategy for the token importance with cache merging. All KV cache compression methods are applied exclusively to the visual tokens, as experiments have demonstrated that it offers greater stability and better performance compared to compressing the entire tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 286, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 591 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 591 + ], + "type": "text", + "content": "Implementation details. All evaluation results are derived from VLMEvalKit [8], which is an open-source evaluation toolkit for LVLMs. The relevance threshold in Equation 5 is set to 0.9. The referee model for MMBench-Video [9] is Qwen2.5-72B [37]. The proposed method performs a one-time KV cache compression only after the prefetch stage is completed. Specifically, once the prefetch ends for each layer, the complete KV cache is fully saved. After computing all layers, we then calculate the budget for each layer and compress the KV cache accordingly. All experiments are conducted on " + }, + { + "bbox": [ + 46, + 460, + 286, + 591 + ], + "type": "inline_equation", + "content": "8 \\times \\mathrm{A}100 - 80\\mathrm{G}" + }, + { + "bbox": [ + 46, + 460, + 286, + 591 + ], + "type": "text", + "content": " GPUs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 599, + 132, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 599, + 132, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 599, + 132, + 610 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 666 + ], + "type": "text", + "content": "We report the results in four sections: VQA datasets, other types of datasets including multiple-choice questions, judgment, and video VQA, inference performance, and the comparison with token pruning methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 286, + 713 + ], + "type": "text", + "content": "VQA datasets. Table 1 presents the comparative results on the ChatQA [26], InfoVQA [27], DocVQA [28], and TextVQA [31] datasets. These datasets encompass tasks such as conventional image question answering, image in-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": "formation extraction, document image understanding, and text recognition within images. It can be observed that the proposed method achieves the best results under most reduction ratios. When " + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": " of visual KV cache is pruned, our method performs almost equivalently to the full KV cache, indicating near lossless performance. Even when reduced to only " + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": " of visual KV cache, the performance gap with the full cache is limited to approximately " + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": ". As the token compression ratio increases, the advantage of the proposed method becomes more pronounced. For instance, when only " + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": " of visual tokens are retained, we outperform SnapKV [19] by an average of " + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 555 + ], + "type": "text", + "content": " across four VQA datasets with the LLaVA-OV-7B [16]. Moreover, the performance differences across different types of VQA datasets indicate that image documents or text understanding, which require higher local visual perception, have stricter demands on KV cache compression than general image understanding. The proposed method demonstrates a more significant advantage over other methods with these datasets, providing strong evidence of its superiority. Furthermore, the above conclusions also apply to different LVLMs, we achieve a significant advantage with InternVL2-8B [6] and Qwen2-VL-7B [35]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 556, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 689 + ], + "type": "text", + "content": "Other types datasets. Table 2 presents the results of various methods on other types of multimodal datasets. For datasets involving multiple-choice questions and judgment types, such as MMBench-EN [24] and MME [10], the final result heavily depends on the accuracy of decoding the first token. Since the KV cache only influences the tokens decoded after the initial token, all methods achieve the same performance as the full KV cache for these types of datasets. The results from the MMBench-Video [9] dataset indicate that our proposed method achieves superior performance, further demonstrating its generality and superiority." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Inference efficiency. We maintain a constant output token count of 512 and compare the full cache with the proposed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23963" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 79, + 518, + 182 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 515, + 79 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 515, + 79 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 515, + 79 + ], + "type": "text", + "content": "Table 3. Quantitative results on inference latency and throughput. The number of tokens output is consistently set to 512." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 79, + 518, + 182 + ], + "lines": [ + { + "bbox": [ + 74, + 79, + 518, + 182 + ], + "spans": [ + { + "bbox": [ + 74, + 79, + 518, + 182 + ], + "type": "table", + "html": "
Batch SizePrompt LengthPrefill Latency (s)Decoding Latency (s)Throughput (token/s)
FullOursFull50%10%Full50%10%
82k1.61.8+12.5%11.69.4-19.0%8.2-29.3%353436+23.5%500+41.6%
8k4.95.3+8.2%15.011.9-20.7%9.4-37.3%273344+26.0%436+59.7%
16k11.311.9+5.3%21.715.0-30.9%10.4-47.5%189273+44.4%394+108.5%
32k25.226.4+4.8%35.922.0-38.7%12.3-65.7%114186+63.2%333+192.1%
162k2.32.5+8.7%14.411.2-22.2%9.7-32.6%569731+28.5%845+48.5%
8k9.49.8+4.3%22.615.6-31.0%11.8-47.8%362525+45.0%694+91.7%
16k20.321.4+5.4%36.923.0-37.7%12.8-65.3%222356+60.4%640+188.3%
", + "image_path": "69be38c234138381b15646a272f71b64cb8609a775fef255c41d713bcb88f9fa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 212, + 544, + 294 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 545, + 212 + ], + "type": "text", + "content": "Table 4. Performance comparison with token pruning methods under various compression ratio. The prefetch and decoding latency are measured under the settings of batch size of 16, total input token of 8k, and the compression ratio of " + }, + { + "bbox": [ + 47, + 190, + 545, + 212 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 47, + 190, + 545, + 212 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 212, + 544, + 294 + ], + "lines": [ + { + "bbox": [ + 50, + 212, + 544, + 294 + ], + "spans": [ + { + "bbox": [ + 50, + 212, + 544, + 294 + ], + "type": "table", + "html": "
MethodsChatQA [26]MMBench-EN [24]MME [10]MMBench-Video [9]Latency (s)
50%10%1%50%10%1%50%10%1%50%10%1%PrefillDecoding
Full80.380.380.382.382.382.31585158515851.811.811.819.422.6
FastV [5]73.647.716.981.474.833.6157413787861.691.361.025.411.7
FasterVLM [39]75.550.118.881.477.026.5155914417311.711.421.175.211.3
IVTP [13]74.955.822.581.777.536.2156314698491.701.471.245.812.6
Ours80.479.976.482.382.382.31585158515851.791.721.609.811.8
", + "image_path": "c56ab96b1ed2502d548accd666bcd63d7b9f2ef120bab02e0dfb50db6573ecd7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": "AirCache across different batch sizes and input token counts in terms of prefetch latency, decoding latency, and throughput. To facilitate testing, the text instruction token count is set to 64. As shown in Table 3, the proposed method significantly improves inference decoding speed under various inputs compared to the full cache. When compressing " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " of the visual KV cache, the proposed method minimally impacts model performance and prefetch latency, while reducing decoding latency by " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "39\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " and increasing model throughput by " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "63\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": ". When compressing " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " of the visual KV cache, the proposed method slightly decreases model performance by less than " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": ", while further reducing decoding latency by " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " and increasing model throughput by " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "inline_equation", + "content": "192\\%" + }, + { + "bbox": [ + 46, + 304, + 289, + 519 + ], + "type": "text", + "content": ". These results demonstrate that our proposed method can maintain strong model performance while ensuring significant decoding acceleration, indicating the effectiveness of visual token importance assessment and hierarchical budget allocation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "Compared with token pruning methods. In addition to KV cache compression, another important method for accelerating LVLMs inference is to prune visual tokens before the prefetch stage. Table 4 compares the proposed method with these type of acceleration methods, including FastV [5], FasterVLM [39], and IVTP [13]. A notable characteristic is that the token pruning methods exhibit acceleration effects during the prefetch stage as they preemptively prune visual tokens. In contrast, all KV cache compression methods, including the proposed approach, demonstrate prefetch stage inference speeds comparable to those of the full cache. However, comparisons across various evaluation sets clearly show that the proposed method achieves an absolute advantage, particularly on VQA datasets, even when retaining only " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " of the visual tokens. In contrast, token pruning methods experience a significant performance" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 308, + 312, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 306, + 301, + 529, + 312 + ], + "lines": [ + { + "bbox": [ + 306, + 301, + 529, + 312 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 529, + 312 + ], + "type": "text", + "content": "Table 5. Ablation results on the components of the AirCache." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 312, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 308, + 312, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 544, + 396 + ], + "type": "table", + "html": "
SettingChatQA [26]InfoVQA [27]DocVQA [28]TextVQA [31]
Continuous window (16)70.456.661.355.9
Continuous window (32)72.957.864.158.2
All text tokens72.258.465.757.0
Visual window (32)68.855.159.253.7
Average allocation72.257.569.962.4
Pyramid allocation69.654.955.852.6
Only w/ st74.259.871.164.9
Only w/ sk74.761.471.963.6
Ours76.462.573.267.1
", + "image_path": "74c4ae42c0e71a4641f6fdd8b5511b64cb9472d3090393b50fe13841f2c6de57.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": "decline under these conditions. The prefetch stage facilitates sufficient interaction between visual and textual information, as the attention mechanism aggregates visual information to key text tokens. The proposed method effectively preserves these key text tokens during KV cache token pruning. Consequently, even when only " + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 408, + 545, + 564 + ], + "type": "text", + "content": " of visual tokens are retained, the model maintains good inference performance. In contrast, most token pruning methods do not involve cross-modal interaction, resulting in direct visual information loss that cannot be transferred to the subsequent decoding stage in any form. This leads to insufficient and inaccurate visual guidance for generating, significantly degrading performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 406, + 586 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 594, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 628 + ], + "type": "text", + "content": "In this section, we evaluate our method under various settings to verify the proposed modules from several perspectives. The base model is LLaVA-OV-7B [16]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 714 + ], + "type": "text", + "content": "Elite Observation Window. We introduce three common types of observation windows to assess the effectiveness of the proposed elite observation window. The first type employs a sequence of continuous text tokens as the observation window for evaluating the significance of visual tokens. The second type utilizes all text tokens, whereas the third employs the continuous visual token from the last visual" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23964" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 100, + 286, + 156 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 286, + 101 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 286, + 101 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 286, + 101 + ], + "type": "text", + "content": "Table 6. Comparison results with different compression audiences, where * indicates uniform compression without distinguishing between visual tokens and text tokens." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 100, + 286, + 156 + ], + "lines": [ + { + "bbox": [ + 50, + 100, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 50, + 100, + 286, + 156 + ], + "type": "table", + "html": "
Models50%10%5%1%
H2O* [42]74.6/407s46.5/559s32.0/733s27.8/1048s
H2O [42]79.8/323s77.4/251s74.0/239s71.0/215s
Ours*75.8/374s53.4/530s39.7/688s30.1/963s
Ours80.4/306s79.9/235s79.4/221s76.4/207s
", + "image_path": "81388583dc675603e038cc565e5f077bb61a7cfe0dbe4b6bb9238d87c9a0adbc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 169, + 287, + 266 + ], + "blocks": [ + { + "bbox": [ + 48, + 169, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 48, + 169, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 48, + 169, + 287, + 266 + ], + "type": "image", + "image_path": "55cd4e0812dc1f5eebbe52655f686b89d3a96dd919ed450d52b2921ceda4c086.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 269, + 286, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 286, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 286, + 301 + ], + "type": "text", + "content": "Figure 4. Comparison on ChatQA [26] and TextVQA [31] by retaining only one visual token, which is selected based on the sorting of visual token importance scores using different methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 286, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 286, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 286, + 482 + ], + "type": "text", + "content": "segment as a reference. As shown in the first column of Table 5, the proposed elite observation window yields the best performance compared to the other three aforementioned approaches. The method using only visual tokens performs significantly worse due to the absence of textual instruction guidance, highlighting the importance of activating cross-modal associations in KV cache compression. Additionally, compared to methods that directly use all or portions of continuous text token sequences without considering variations in visual information perception among text tokens, the proposed method selects key text tokens to enhance the consistency of the observation window. This selection ensures greater stability and accuracy in evaluating the significance of visual tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 286, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 286, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 286, + 674 + ], + "type": "text", + "content": "Layer-wise Token Budget Allocation Module. The second column of Table 5 presents the comparison using various budget allocation strategies across layers. Aside from the basic equal allocation, we also incorporate the strategy from PyramidKV [3], which allocates budgets in a descending ratio from shallow to deep layers. The results demonstrate that the proposed budget allocation, based on the strength and skewness of importance scores, delivers superior outcomes. Although PyramidKV [3] perform well in LLMs, its suboptimal results in LVLMs highlight the distinct characteristics of multimodal models and their heightened sensitivity to hierarchical budget allocation. An inappropriate allocation strategy is more likely to adversely affect the final results. We also investigate the separate use of strength and skewness for layer-wise budget quantification, revealing that their combination produces the best results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 712 + ], + "type": "text", + "content": "Visual Token Importance Evaluation. Figure 4 visually demonstrates the comparative effectiveness of various methods in evaluating the importance of visual tokens. To" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": "directly assess the efficacy of these methods in ranking visual token importance, we removed all non-essential modules and selected only one visual token at a time from the ranked sorting results, ordered from high to low importance. Theoretically, as the importance ranking of tokens declines, the model's performance is also expected to decrease. The graph clearly shows that the proposed method aligns more closely with the theoretical trend, where the visual token ranked as TOP1 achieves the highest metric, followed by a continuous and smooth decline. Notably, after around " + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": ", the performance of the visual tokens approximates that of the case with all visual tokens removed. This indicates significant redundancy among visual tokens in the KV cache of LVLMs. The results in Table 1, which demonstrate that retaining over " + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": " of the visual tokens limits performance decline to no more than " + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 299 + ], + "type": "text", + "content": ", further corroborate this observation. In contrast, the performance derived from other methods exhibit significant noise, indicating instability and deficiencies in their importance assessment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 490 + ], + "type": "text", + "content": "Unified Compression or Vision-only Compression. Given that the KV cache in LVLMs encompasses both visual and textual modalities, in contrast to the single textual modality of LLMs, we compared Unified Compression, which does not differentiate between modal differences during KV cache compression, with Vision-only Compression, which compresses only visual tokens. As shown in Table 6, the results indicate that unified compression performs significantly worse than vision-only compression in terms of both inference performance and speed. Moreover, as the compression ratio increases, the inference time for unified compression actually continues to rise. The primary reason is that unified compression disrupts the effective representation of textual instructions, leading the model to produce incorrect outputs and exhibit a tendency for repetition, which in turn increases inference time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "spans": [ + { + "bbox": [ + 306, + 502, + 378, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "In this work, we propose Activating Inter-modal Relevancy (AirCache), a novel KV cache compression for efficient LVLMs inference. By leveraging inter-modal relevancy between textual and visual tokens to guide KV cache eviction, we identified significant differences in the attention distribution of visual tokens corresponding to different textual tokens. Based on this finding, we propose an elite observation window with refined key text tokens, which provides a more stable and effective assessment of visual token importance. Furthermore, by exploring the diversity of importance score distribution across different layers, we propose to achieve comprehensive quantification in terms of distribution strength and skewness for adaptive layer-wise compression budget allocation. Extensive experiments across multiple benchmarks and LVLMs demonstrate the effectiveness of AirCache in KV cache compression." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23965" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 144 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 201 + ], + "type": "text", + "content": "[2] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4291-4301, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Zefan Cai, Yichi Zhang, Bofei Gao, Yuliang Liu, Tianyu Liu, Keming Lu, Wayne Xiong, Yue Dong, Baobao Chang, Junjie Hu, et al. Pyramidkv: Dynamic kv cache compression based on pyramidal information funneling. arXiv preprint arXiv:2406.02069, 2024. 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 312 + ], + "type": "text", + "content": "[4] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 288, + 369 + ], + "type": "text", + "content": "[5] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 1, 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 369, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 435 + ], + "type": "text", + "content": "[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 287, + 491 + ], + "type": "text", + "content": "[7] Xiangxiang Chu, Limeng Qiao, Xinyang Lin, Shuang Xu, Yang Yang, Yiming Hu, Fei Wei, Xinyu Zhang, Bo Zhang, Xiaolin Wei, et al. Mobilevm: A fast, reproducible and strong vision language assistant for mobile devices. arXiv preprint arXiv:2312.16886, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 287, + 557 + ], + "type": "text", + "content": "[8] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 558, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 287, + 602 + ], + "type": "text", + "content": "[9] Xinyu Fang, Kangrui Mao, Haodong Duan, Xiangyu Zhao, Yining Li, Dahua Lin, and Kai Chen. Mmbench-video: A long-form multi-shot benchmark for holistic video understanding. arXiv preprint arXiv:2406.14515, 2024. 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 657 + ], + "type": "text", + "content": "[10] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[11] Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, and Bohan Zhuang. Zipvl: Efficient large vision-language models with dynamic token sparsification and kv cache compression. arXiv preprint arXiv:2410.08584, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[12] Wenbo Hu, Zi-Yi Dou, Liunian Harold Li, Anita Kamath, Nanyun Peng, and Kai-Wei Chang. Matryoshka query transformer for large vision-language models. arXiv preprint arXiv:2405.19315, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 163 + ], + "type": "text", + "content": "[13] Kai Huang, Hao Zou, Ye Xi, BoChen Wang, Zhen Xie, and Liang Yu. Ivtp: Instruction-guided visual token pruning for large vision-language models. In European Conference on Computer Vision, pages 214-230. Springer, 2024. 1, 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 229 + ], + "type": "text", + "content": "[14] Chaoya Jiang, Haiyang Xu, Wei Ye, Qinghao Ye, Chen-liang Li, Ming Yan, Bin Bi, Shikun Zhang, Fei Huang, and Songfang Huang. Bus: Efficient and effective vision-language pre-training with bottom-up patch summarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2900–2910, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 273 + ], + "type": "text", + "content": "[15] Jang-Hyun Kim, Junyoung Yeom, Sangwoo Yun, and Hyun Oh Song. Compressed context memory for online language model interaction. arXiv preprint arXiv:2312.03414, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 319 + ], + "type": "text", + "content": "[16] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 308, + 320, + 545, + 374 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 420 + ], + "type": "text", + "content": "[18] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 475 + ], + "type": "text", + "content": "[19] Yuhong Li, Yingbing Huang, Bowen Yang, Bharat Venkitesh, Acyr Locatelli, Hanchen Ye, Tianle Cai, Patrick Lewis, and Deming Chen. Snapkv: Llm knows what you are looking for before generation. arXiv preprint arXiv:2404.14469, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 521 + ], + "type": "text", + "content": "[20] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 578 + ], + "type": "text", + "content": "[21] Akide Liu, Jing Liu, Zizheng Pan, Yefei He, Reza Haffari, and Bohan Zhuang. Minicache: Kv cache compression in depth dimension for large language models. Advances in Neural Information Processing Systems, 37:139997-140031, 2025. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "text", + "content": "[22] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[24] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23966" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[25] Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, and Jiwen Lu. Efficient inference of vision instruction-following models with elastic cache. In European Conference on Computer Vision, pages 54-69. Springer, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 129, + 287, + 173 + ], + "type": "text", + "content": "[26] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 227 + ], + "type": "text", + "content": "[27] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "text", + "content": "[28] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "text", + "content": "[29] Piotr Nawrot, Adrian Łancucki, Marcin Chochowski, David Tarjan, and Edoardo M Ponti. Dynamic memory compression: Retrofitting llms for accelerated inference. arXiv preprint arXiv:2403.09636, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 322, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 287, + 374 + ], + "type": "text", + "content": "[30] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Jonathan Heek, Kefan Xiao, Shivani Agrawal, and Jeff Dean. Efficiently scaling transformer inference. Proceedings of Machine Learning and Systems, 5:606-624, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 49, + 377, + 287, + 432 + ], + "type": "text", + "content": "[31] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 49, + 434, + 287, + 498 + ], + "type": "text", + "content": "[32] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 49, + 501, + 287, + 544 + ], + "type": "text", + "content": "[33] Dezhan Tu, Danylo Vashchilenko, Yuzhe Lu, and Panpan Xu. VI-cache: Sparsity and modality-aware kv cache compression for vision-language model inference acceleration. arXiv preprint arXiv:2410.23317, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 49, + 546, + 287, + 601 + ], + "type": "text", + "content": "[34] Ao Wang, Hui Chen, Jianchao Tan, Kefeng Zhang, Xunliang Cai, Zijia Lin, Jungong Han, and Guiguang Ding. Prefixkv: Adaptive prefix kv cache is what vision instruction-following models need for efficient generation. arXiv preprint arXiv:2412.03409, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 49, + 602, + 287, + 656 + ], + "type": "text", + "content": "[35] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[36] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 361 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 115 + ], + "type": "text", + "content": "[37] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 161 + ], + "type": "text", + "content": "[38] Yifei Yang, Zouying Cao, Qiguang Chen, Libo Qin, Dongjie Yang, Hai Zhao, and Zhi Chen. Kvsharer: Efficient inference via layer-wise dissimilar kv cache sharing. arXiv preprint arXiv:2410.18517, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[39] Qizhe Zhang, Aosong Cheng, Ming Lu, Zhiyong Zhuo, Minqi Wang, Jiajun Cao, Shaobo Guo, Qi She, and Shang-hang Zhang. [cls] attention is all you need for training-free visual token pruning: Make vlm inference faster. arXiv preprint arXiv:2412.01818, 2024. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 262 + ], + "type": "text", + "content": "[40] Yuxin Zhang, Yuxuan Du, Gen Luo, Yunshan Zhong, Zhenyu Zhang, Shiwei Liu, and Rongrong Ji. Cam: Cache merging for memory-efficient llms inference. In *Forty-first International Conference on Machine Learning*, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 296 + ], + "type": "text", + "content": "[41] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 297, + 545, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 361 + ], + "type": "text", + "content": "[42] Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23967" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_content_list.json b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c118ee6920b9b80a0ca9013efe82db4676754a6d --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_content_list.json @@ -0,0 +1,1530 @@ +[ + { + "type": "text", + "text": "Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Congyi Fan $^{1}$ , Jian Guan $^{1*}$ , Xuanjia Zhao $^{1}$ , Dongli Xu $^{2}$ , Youtian Lin $^{3}$ , Tong Ye $^{1}$ , Pengming Feng $^{4}$ , Haiwei Pan $^{1}$", + "bbox": [ + 269, + 202, + 727, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ College of Computer Science and Technology, Harbin Engineering University, Harbin, China $^{2}$ Shanghai Academy of Artificial Intelligence for Science, Shanghai, China", + "bbox": [ + 127, + 255, + 870, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Intelligence Science and Technology, Nanjing University, Nanjing, China $^{4}$ State Key Laboratory of Space-Ground Integrated Information Technology, Beijing, China", + "bbox": [ + 137, + 291, + 859, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatically generating natural, diverse, and rhythmic human dance movements driven by music is vital for virtual reality and film industries. However, generating dance that naturally follows music remains a challenge, as existing methods lack proper beat alignment and exhibit unnatural motion dynamics. In this paper, we propose Danceba, a novel framework that leverages gating mechanism to enhance rhythm-aware feature representation for music-driven dance generation, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, we introduce Phase-Based Rhythm Extraction (PRE) to precisely extract rhythmic information from musical phase data, capitalizing on the intrinsic periodicity and temporal structures of music. Additionally, we propose Temporal-Gated Causal Attention (TGCA) to focus on global rhythmic features, ensuring that dance movements closely follow the musical rhythm. We also introduce Parallel Mamba Motion Modeling (PMMM) architecture to separately model upper and lower body motions along with musical features, thereby improving the naturalness and diversity of generated dance movements. Extensive experiments confirm that Danceba outperforms state-of-the-art methods, achieving improved rhythmic alignment and motion diversity. Project page: https://danceba.github.io/.", + "bbox": [ + 88, + 393, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 787, + 220, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Music-driven dance generation aims to automatically synthesize realistic and expressive 3D dance movements synchronized precisely with a given musical sequence, enabling natural animation of 3D character models. Due to", + "bbox": [ + 89, + 813, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "its broad applications in gaming, virtual reality, film production, and interactive entertainment [14, 34, 38, 43], this task has attracted considerable attention. However, generating dances that accurately reflect musical rhythm while preserving natural and diverse movement dynamics remains a significant challenge, as it requires effectively capturing and utilizing complex temporal relationships between music and motion.", + "bbox": [ + 511, + 362, + 905, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, generative models have demonstrated remarkable success across various tasks, such as text and image generation [9, 29]. Similarly, these generative approaches have also been applied to music-driven dance generation tasks [11, 19, 20, 27, 30, 31, 33, 36, 42, 44]. Among these, Bailando [30] introduces a cross-conditional causal attention mechanism to align musical cues with corresponding dance poses, significantly improving the quality of generated dance movements. However, its attention mechanism primarily captures local correspondence between music and poses, failing to sufficiently utilize global rhythmic characteristics within music, thus limiting the expressiveness and rhythmic coherence of the generated dances.", + "bbox": [ + 511, + 488, + 908, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Follow-up studies [11, 31, 36, 44] have extended Bailando's design to further improve dance generation performance. For instance, Enhancing-Bailando [11] employs a pretrained audio encoder (MERT) [21] to obtain richer musical feature representations. However, this encoder primarily captures semantic information from general audio data rather than explicitly modeling rhythmic characteristics, limiting its effectiveness in precise beat-dance synchronization. Moreover, Bailando and related methods typically concatenate three distinct input modalities, i.e., upper-body pose, lower-body pose, and musical features, and rely heavily on cross-conditional causal attention to align these inputs. Nevertheless, this attention mechanism primarily aligns pose and music at a local feature level, but struggles", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: j.guan@hrbeu.edu.cn", + "bbox": [ + 107, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to capture continuous motion dynamics, limiting the naturalness and diversity of generated dance movements.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further tackle the rhythm alignment issue, Beat-It [42] explicitly introduces beat synchronization via a beat distance predictor. However, such an explicit synchronization approach imposes rigid control constraints and strong regularization during generation, significantly limiting the diversity and expressiveness of generated dances [27]. Thus, a more balanced approach capable of achieving precise rhythmic alignment without sacrificing motion naturalness and diversity is highly desirable.", + "bbox": [ + 88, + 121, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel framework equipped with a gating-enhanced rhythm-aware feature representation for music-driven dance generation, i.e., Danceba, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, Danceba introduces Phase-Based Rhythm Extraction (PRE), a rhythm-aware feature extraction module that explicitly captures rhythmic characteristics from musical phase information. Leveraging the temporal phase information intrinsic to rhythm perception, PRE effectively separates rhythmic from semantic music features, significantly improving the rhythmic alignment between generated dance sequences and musical inputs.", + "bbox": [ + 89, + 258, + 482, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further enhance beat-dance alignment using the rhythm-aware features extracted by PRE, it is essential to establish a stronger contextual association between musical rhythm and dance movements. To achieve this, we present a Temporal-Gated Causal Attention (TGCA) mechanism, enhanced by a gating strategy to explicitly prioritize and leverage global rhythmic structures within cross-conditional causal attention. By integrating gating mechanism, TGCA effectively strengthens the contextual connection between music rhythm and generated movements, ensuring that generated dance sequences accurately reflect the underlying musical rhythm and exhibit greater coherence and expressive diversity.", + "bbox": [ + 89, + 439, + 482, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to overcome the limitations of traditional attention-based approaches in motion modeling, we introduce Parallel Mamba Motion Modeling (PMMM), inspired by recent advances demonstrating that Mamba architectures achieve state-of-the-art performance in modeling complex human motion sequences [16, 39, 41]. Our parallel Mamba architecture separately models upper and lower-body dance sequences, with input features enhanced by Temporal-Gated Causal Attention (TGCA), which incorporates rhythm-aware representations alongside distinct upper and lower-body pose features. This design effectively captures diverse and nuanced motion characteristics. By leveraging Mamba's superior sequential modeling capability, our method significantly improves the expressiveness, diversity, and temporal coherence of generated dance movements. Fi", + "bbox": [ + 89, + 637, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nally, an additional TGCA module is applied after PMMM to further reinforce rhythmic alignment, ensuring that the generated dances precisely reflect the musical rhythm.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments show that our method improves $\\mathrm{FID}_k$ by $48.68\\%$ , $\\mathrm{Div}_k$ by $7.0\\%$ , $\\mathrm{Div}_g$ by $16.3\\%$ , and BAS by $12.0\\%$ over the second-best results, while maintaining a competitive $\\mathrm{FID}_g$ of 11.90. This significantly surpasses state-of-the-art methods, producing dance movements that are more natural, diverse, and better synchronized with musical rhythm.", + "bbox": [ + 511, + 137, + 906, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 258, + 653, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Music-Driven Dance Generation", + "text_level": 1, + "bbox": [ + 511, + 284, + 797, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Music-driven dance generation focuses on automatically synthesizing 3D dance movements from musical input, creating temporally coherent pose sequences for animating 3D character models. It has attracted increasing attention in both the computer vision and virtual avatar communities [1, 13] due to its applications in entertainment, gaming, and virtual reality. Early approaches, including [10, 12, 15, 17, 18, 32, 45], investigated diverse methodologies for this task. The study in [17] utilized a standard Transformer [35] with a late-fusion module to predict subsequent dance steps. The FACT framework [18] introduced the AIST++ dataset and employed a cross-modal Transformer with full attention mechanisms for music-to-dance generation. However, these initial methods neglect the rhythm features of music and struggle to effectively integrate musical and motion features, leading to suboptimal generation quality.", + "bbox": [ + 511, + 306, + 903, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bailando [30] reformulated dance generation as a next-step pose prediction problem using cross-conditional causal attention, achieving high-quality music-driven dance synthesis. Subsequent works [11, 31, 44] extended this framework: Bailando++ [31] enhanced pattern recognition with musical context encoding, while Enhancing-Bailando [11] incorporated a pre-trained MERT audio encoder [21] for improved feature extraction. However, these methods often overlook rhythmic structures or rely on generic audio features. Beat-It [42] introduced explicit beat alignment via a conditional diffusion model [9, 25], ensuring precise synchronization. Yet, its reliance on rigid beat control signals and strong regularization may constrain pose diversity [27]. In contrast, we emphasize rhythm-aware feature representation, achieving superior rhythmic alignment without sacrificing movement diversity.", + "bbox": [ + 511, + 564, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Mamba in Human Motion Generation", + "text_level": 1, + "bbox": [ + 511, + 816, + 843, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The integration of Mamba [7] and Mamba2 [2] into human motion generation represents a pivotal advancement, leveraging their efficient sequence modeling capabilities. Within human motion generation, Mamba surpasses con", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Note that, the terms \"beat\" and \"rhythm\" are used interchangeably in music processing following DiffDance [27] and Bailando++ [31].", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13194", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/65a682da204142f85eba68eb6ffe1107f11afd402c702d7ee96ae4dd7207e8a0.jpg", + "image_caption": [ + "Figure 1. The overall framework of Danceba consists of three core modules: Phase-based Rhythm Extraction (PRE), Temporal-Gated Causal Attention (TGCA), and Parallel Mamba Motion Modeling (PMMM). PRE precisely extracts rhythm-aware features from musical phase information, providing accurate rhythmic signals. These rhythm-enhanced music features are fused with pose embeddings (upper-body and lower-body poses) and processed through the TGCA module, which utilizes gating mechanism to reinforce rhythmic sensitivity and align dance movements accurately with music beats. Parallel Mamba Motion Modeling separately models upper and lower body motion sequences, effectively capturing distinct dance dynamics aligned with rhythm-aware music features, thus significantly enhancing the naturalness, diversity, and temporal coherence of the generated dance movements." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 908, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ventional Transformers in performance, as demonstrated by Motion Mamba [41], which highlights its efficacy for extended motion sequence synthesis. Furthermore, InfiniMotion [40] utilizes Mamba's hierarchical and bidirectional temporal processing to generate continuous, smooth motion sequences of indefinite length. Recent approaches including SMCD [28], Light-T2M [39], FTMoMamba [16], and MMD [4] have further incorporated Mamba architectures, advancing the field through enhanced efficiency and high-fidelity motion modeling. Building upon recent advancements, Mamba has achieved state-of-the-art performance in human motion modeling [16, 39, 41], motivating our selection of Mamba over self-attention mechanisms for information exchange within dance sequences, which significantly enhances the diversity of generated dance motions.", + "bbox": [ + 88, + 455, + 485, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 705, + 264, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our proposed Danceba framework for music-driven 3D dance generation. By leveraging gating-enhanced rhythm-aware features, Danceba ensures precise alignment between dance poses and music. The overall framework is illustrated in Figure 1, and its components are detailed as follows: Section 3.1 outlines the preliminary framework and key notations. Specifically, we introduce phase-based rhythm feature extraction (Section 3.2), global beat-dance alignment via temporal-gated causal attention (Section 3.3), and Mamba-based parallel motion modeling (Section 3.4).", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 455, + 643, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a dance sequence $\\mathbf{P} \\in \\mathbb{R}^{T \\times (J \\times 3)}$ , where $T$ is the time length and $J$ is joint amount. Using two pretrained Pose VQ-VAEs [30] $\\mathcal{F}_{VAE}^u$ and $\\mathcal{F}_{VAE}^l$ with codebook $\\mathcal{Z}^u$ and $\\mathcal{Z}^l$ for the upper and lower half bodies, it separately encodes upper and lower body movements into compositional code pairs $\\mathbf{p} = [\\mathbf{p}^u, \\mathbf{p}^l]$ :", + "bbox": [ + 511, + 476, + 906, + 570 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {u} = \\mathcal {F} _ {V A E} ^ {u} (\\mathbf {P}), \\mathbf {p} ^ {l} = \\mathcal {F} _ {V A E} ^ {l} (\\mathbf {P}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 574, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p}^u\\in (\\mathcal{Z}^u)^{T'}$ $\\mathbf{p}^l\\in (\\mathcal{Z}^l)^{T'}$ .Here $T^{\\prime} = T / \\lambda$ $\\lambda$ is the temporal down-sampling rate. The input music feature $\\mathbf{m}\\in \\mathbb{R}^{T'\\times D_m}$ , where $D_{m}$ is the channel dimension of the input music features. Then we embed music features $\\mathbf{m}$ upper $p^u$ and lower $p^l$ pose codes to learnable features with three separate linear layers: $\\mathbf{X}_{\\mathrm{m}} = \\mathrm{Linear}(\\mathbf{m}),\\mathbf{X}_{\\mathrm{u}} =$ Linear $(\\mathbf{p}^{u})$ and $\\mathbf{X}_1 = \\mathrm{Linear}(\\mathbf{p}^l)$ , where $\\mathbf{X}_{\\mathrm{m}}\\in \\mathbb{R}^{T^{\\prime}\\times D}$ $\\mathbf{X}_{\\mathrm{u}}\\in \\mathbb{R}^{T^{\\prime}\\times D}$ $\\mathbf{X}_1\\in \\mathbb{R}^{T^{\\prime}\\times D}$ and D is the channel dimension of the features.", + "bbox": [ + 511, + 598, + 908, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The dance generation task is reframed as selecting the most probable future pose code from the codebook $\\mathcal{Z}$ , conditioned on the music and prior movements. Since the upper and lower bodies are modeled separately, maintaining coherence and avoiding asynchrony (e.g., opposing directions) requires cross-conditioned prediction, leveraging mutual information between existing movements:", + "bbox": [ + 511, + 734, + 908, + 843 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {t} ^ {u} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {u} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 848, + 906, + 871 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} _ {t} ^ {l} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {l} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 873, + 906, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At each time step $t$ , MotionGPT [30] estimates the probabilities of pose codes $\\mathbf{z}_i \\in \\mathbb{Z}$ and selects the one with the most probable pose codes as the predicted upper pose $\\hat{p}_t^u$ and predicted lower pose $\\hat{p}_t^l$ .", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the input feature $\\mathbf{X}_i\\in \\mathbb{R}^{(3\\times T')\\times D}$ are concatenated by upper body motion feature, lower body motion feature and music feature as follows: $\\mathbf{X}_i =$ $\\mathrm{Concat}(\\mathbf{X}_{\\mathrm{m}},\\mathbf{X}_{\\mathrm{u}},\\mathbf{X}_{\\mathrm{l}})$ . Subsequently, $\\mathbf{X}_i$ is passed through successive Autoregressive Transformer layers $\\mathcal{F}_{AR}$ , followed by a linear transformation and a softmax layer, generates the probability distributions for predicting body pose collectively:", + "bbox": [ + 89, + 151, + 482, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {a} ^ {h} = \\operatorname {S o f t m a x} (\\operatorname {L i n e a r} \\left(\\mathcal {F} _ {A R} \\left(\\mathbf {X} _ {i}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 282, + 482, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{a}^h = [\\mathbf{a}^u,\\mathbf{a}^l]$ , $\\mathbf{a}^u$ denotes that of upper body and $\\mathbf{a}^l$ denotes that of lower body.", + "bbox": [ + 89, + 309, + 480, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the model is optimized via supervised training with the cross-entropy loss on action probability $\\mathbf{a}_t^h$ at each time step $t$ and for both upper body and lower body:", + "bbox": [ + 89, + 340, + 482, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C E} = \\frac {1}{T ^ {\\prime}} \\sum_ {t = 0} ^ {T ^ {\\prime} - 1} \\sum_ {h = u, l} \\text {C r o s s E n t r o p y} \\left(\\mathrm {a} _ {t} ^ {h}, p _ {t + 1} ^ {h}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 397, + 482, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In inference, the MotionGPT predicts pose codes according to the initial pose code and the entire music, followed by Pose VQ-VAE decoders generating a new dance.", + "bbox": [ + 89, + 450, + 482, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Phase-Based Rhythm Feature Extraction", + "text_level": 1, + "bbox": [ + 89, + 505, + 441, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance rhythmic feature extraction in music-driven dance generation, we propose Phase-Based Rhythm Extraction (PRE), which decouples rhythm from musical semantics by leveraging the inherent periodicity and temporal positioning of phase representations. The architecture of PRE is shown in the green part of Figure 1. Unlike existing methods that entangle rhythm with melody and harmony, PRE enables independent and precise rhythm modeling as shown in Figure 2, forming a solid foundation for gating-enhanced rhythm-aware mechanism.", + "bbox": [ + 89, + 527, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we adopt the Short-Time Fourier Transform (STFT) instead of the conventional Fourier Transform (FT), since the non-stationary nature of music and the sliding-window approach of STFT effectively captures sudden tempo shifts and complex rhythmic patterns [37]. Specifically, we first perform STFT on the input music feature to extract phase angles:", + "bbox": [ + 89, + 678, + 482, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S} _ {m} = \\operatorname {S T F T} (\\mathbf {m}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 794, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\varphi} = \\operatorname {A n g l e} \\left(\\mathbf {S} _ {m}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 813, + 352, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{S}_m$ represents the complex spectrogram obtained via STFT, $\\mathrm{Angle}(\\cdot)$ computes the phase angle of a complex-valued input and $\\varphi^{\\prime}$ denotes the phase angles extracted from $\\mathbf{S}_m$ . By taking the angle of the STFT output, we obtain the", + "bbox": [ + 89, + 840, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4516cdb0664827a71e3752b6f55f139686de0bdcbbea09f1563fc9974db24f4f.jpg", + "image_caption": [ + "Figure 2. Visualizing the input music features and phase features. We notice that only a small portion of the input music features contains meaningful musical information, i.e., rhythm-poor features. And after a short-time Fourier transform (STFT), each corresponding token has a rich periodic Phase Feature, i.e., rhythm-rich features." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 903, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "music phase data $\\varphi^{\\prime}\\in \\mathbb{C}^{T_{\\varphi}\\times D_{\\varphi}}$ , where $T_{\\varphi}$ represents the number of time frames obtained after the STFT transformation, and $D_{\\varphi}$ is the channel dimension of the $\\varphi^{\\prime}$", + "bbox": [ + 511, + 344, + 903, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The extracted phase angles $\\varphi^{\\prime}$ inherently reflect temporal shifts and periodic structures of musical rhythms: low-frequency phase variations correspond to broader rhythmic patterns, while high-frequency variations capture finer rhythmic details. We also apply a center-cropping strategy, selectively focusing on the most informative phase regions:", + "bbox": [ + 511, + 391, + 903, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi = \\operatorname {C e n t e r C r o p} \\left(\\varphi^ {\\prime}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 494, + 903, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varphi \\in \\mathbb{C}^{T' \\times D_{\\varphi}}$ has the same temporal length $T'$ , addressing the potential temporal misalignment and preserving critical rhythmic cues.", + "bbox": [ + 511, + 516, + 903, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the phase features through a linear transformation, enhancing phase features representing musical rhythms. The process is shown in the following equation:", + "bbox": [ + 511, + 561, + 903, + 607 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\varphi} = \\operatorname {R e L U} (\\mathrm {B N} (\\operatorname {L i n e a r} (\\varphi))), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 616, + 903, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the embedded phase features $\\mathbf{X}_{\\varphi} \\in \\mathbb{C}^{T' \\times D}$ encode rich rhythmic information, providing a clearer and more structured representation of rhythmic and musical elements. Finally, we fuse the input feature $\\mathbf{X}_i$ with the rhythmic features $\\mathbf{X}_{\\varphi}$ by concatenating them: $\\mathbf{X}_{\\gamma} = \\mathrm{Concat}(\\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi})$ . Then, we sum the rhythm-enhanced features with the original music, upper body, and lower body motion features: $\\mathbf{X} = \\mathbf{X}_f + \\mathbf{X}_{\\gamma}$ , where $\\mathbf{X}$ represents the rhythmically enhanced features.", + "bbox": [ + 511, + 642, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Global Beat Attention via Temporal Gating", + "text_level": 1, + "bbox": [ + 511, + 787, + 880, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging these rhythm-rich features, we next investigate how to integrate them into the alignment process between music and dance. However, prior methods [11, 30, 31, 36, 44] relying on cross-conditional causal attention mechanism (denoted as $\\mathrm{C}^3\\mathrm{Attention}$ ) exhibit a critical misalignment issue. As shown in Figure 3 (a), we visualize the", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13196", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cross-conditional causal attention mechanism and observe that the next predicted token lacks direct control signals from historical tokens, failing to establish a stable global beat attention. This misalignment leads to error accumulation, which we identify as a key factor behind the unnatural dance generation in prior work.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaec8bec3d5f539fab42a024e18362868b0f88f6dc153345758089dcf623cdb9.jpg", + "image_caption": [ + "(a) Cross-Conditional Causal Attention" + ], + "image_footnote": [], + "bbox": [ + 94, + 194, + 256, + 318 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/964265c1bd018d27bf9512a68d8abf13ad6ad772252f81b68fc9c55ede5bf566.jpg", + "image_caption": [ + "(b) Temporal-Gated Causal Attention (TGCA)", + "Figure 3. Heat map of attention. The two displayed images show the results learned by Cross-Conditional Causal Attention (a) and Temporal-Gated Causal Attention (b) under the same piece of music. The brighter the color, the higher the attention; the darker the color, the lower the attention. TGCA the next token has a clear control signal, preventing overly random dance movements and achieving enhanced rhythm-aware feature representation." + ], + "image_footnote": [], + "bbox": [ + 272, + 194, + 480, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on this, we propose Temporal Gated Causal Attention (TGCA) to enhance global rhythm feature attention, improving music-dance alignment and ensuring the next token is guided by explicit historical information, preventing random action generation. As illustrated in the blue section of Figure 1, the rhythmically enhanced features $\\mathbf{X}$ are passed through TGCA module, which can be expressed as the following:", + "bbox": [ + 89, + 465, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G a t i n g} (\\mathbf {X}) = \\operatorname {S i L U} (\\operatorname {L i n e a r} (\\mathbf {X})). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 595, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final output is obtained by the element-wise multiplication of cross-conditional causal attention $\\mathrm{C}^3$ Attention and Gating, as shown in the following equation:", + "bbox": [ + 89, + 621, + 483, + 666 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {X} _ {a t t n} = \\operatorname {T G C A} (\\mathbf {X}) \\\\ = \\mathrm {C} ^ {3} \\text {A t t e n t i o n} (\\mathbf {X}) \\odot \\operatorname {G a t i n g} (\\mathbf {X}). \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 675, + 480, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3 (b), our proposed gated causal attention mechanism allows historical tokens to better predict the next token, implying an enhanced influence of global rhythmic patterns on dance generation.", + "bbox": [ + 89, + 719, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mamba-Based Parallel Motion Modeling", + "text_level": 1, + "bbox": [ + 89, + 787, + 441, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After TGCA enhances dance movements in $\\mathbf{X}_{attn}$ with global rhythmic perception, we looking into dance modeling for generation. Dance generation faces two key challenges: (1) Dance is a complex spatial-temporal sequence encoding trajectory, posture, and velocity over time, requiring a model capable of capturing both local transitions and", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "global rhythmic variations. (2) The upper and lower body exhibit distinct movement patterns and speeds, necessitating separate processing to maintain coherence and diversity. To address these issues, we propose a parallel Mamba motion modeling architecture that separately models music, upper-body, and lower-body movements while leveraging state space modeling to ensure fluid and diverse motion generation.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, to ensure each action is primarily influenced by relevant preceding movements, we introduce a GateMlp in the Mamba block, as shown in Figure 1. This mechanism selectively retains essential motion features while filtering out less relevant information, thereby enhancing the coherence and expressiveness of the generated dance sequences. The process is formulated as follows:", + "bbox": [ + 511, + 212, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {m b} ^ {\\prime} = \\operatorname {M a m b a} (\\operatorname {R M S N o r m} (\\mathbf {X} _ {a t t n})) + \\mathbf {X} _ {a t t n}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 329, + 903, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {m b} = \\operatorname {G a t e M l p} \\left(\\operatorname {R M S N o r m} \\left(\\mathbf {X} _ {m b} ^ {\\prime}\\right)\\right) + \\mathbf {X} _ {m b} ^ {\\prime}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 354, + 903, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gating mechanism dynamically regulates the frequency, speed, and amplitude of movements in response to musical beats, ensuring that dance sequences remain synchronized with the music. Finally, three parallel Mamba architectures model upper- and lower-body dance sequences separately, therefore ensure the generated dance movements are expressive, diverse, and temporally coherent.", + "bbox": [ + 511, + 385, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 505, + 645, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 531, + 702, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset: We follow prior works [27, 30, 36] and perform all experiments on the AIST++ dataset [18], the most widely used benchmark for music-driven dance generation that avoids observable biases (e.g., offensive postures). AIST++ consists of 991 high-quality 3D pose sequences recorded at 60 FPS in skinned multi-person linear (SMPL) format [22], with 951 sequences designated for training and 40 reserved for evaluation. Following [27, 30, 36], we generate 40 pieces of dance sequences on the test set of AIST++ dataset, and sample the generated dance sequence with length of 20 seconds for performance evaluation.", + "bbox": [ + 511, + 551, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: We follow previous studies [18, 30] to quantitatively evaluate the generated samples in three key aspects: quality, diversity, and beat-dance alignment. Prior to evaluation, it is necessary to extract kinetic features [26] and geometric features [24] from both the ground truth and the generated samples via a specific toolbox [6]. For dance quality, we compute two types of the Fréchet Inception Distance (FID) [8], including $\\mathrm{FID}_k$ based on kinetic features and $\\mathrm{FID}_g$ based on geometric features. Lower $\\mathrm{FID}_k$ indicates that the generated dance has more similar kinetic features to the original dance, meaning the generated dances are more natural and smooth. Lower $\\mathrm{FID}_g$ indicates that", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/45bfca63d204bd684bec9d822b2e77fec636aa7a3dc03d40cfe1ecd3f120766c.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods on the AIST++ dataset. Underlining indicates the best performance among existing methods. Blue indicates results that surpass the best existing method. ↓ indicates that lower values are better, while ↑ indicates that higher values are better. * represents the dances generated by \"Li et al.\" that exhibit significant jitteriness, resulting in extremely high velocity variation, as also noted in [18]. † indicates the reproduced results obtained using Bailando++ official code and published checkpoints. ‡ denotes the results that are averaged over five independent training runs." + ], + "table_footnote": [], + "table_body": "
METHODVENUEMOTION QUALITYMOTION DIVERSITYBEAT ALIGN SCORE↑
FIDk↓FIDg↓DIVk↑DIVg↑
GROUND TRUTH-17.1010.608.197.450.2374
Li et al. [17]ARXIV 202086.4343.46*6.853.320.1607
DANCERUVOLUTION [12]ICLR 202173.4225.923.524.870.1950
DANCENET [45]TOMM 202269.1825.492.862.850.1430
FACT [18]ICCV 202135.3522.115.946.180.2209
BAILANDO [30]CVPR 202228.169.627.836.340.2332
EDGE [33]CVPR 202342.1622.123.964.610.2334
DIFFDANCE [27]ACMMM 202324.0920.686.022.890.2418
†BAILANDO++ [31]TPAMI 202322.7411.587.946.340.2263
E3D2 [36]AAAI 202426.258.947.966.490.2232
LODGGE [20]CVPR 202437.0918.795.584.850.2423
OURS (BEST)11.67(-11.07)11.908.52(+0.58)7.55(+1.21)0.2714(+2.91%)
‡OURS (AVERAGE)15.48(-7.26)13.577.857.49(+1.15)0.2779(+3.56%)
", + "bbox": [ + 98, + 170, + 897, + 390 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/671396d25209960c5a5cb3ed3dce7add1dd04ea85c7e68ad8cb7e095eabf6218.jpg", + "image_caption": [ + "Figure 4. Dance examples generated by our proposed method on various types of music. The demo videos can be found in the supplementary material. The character is sourced from Mixamo [23]." + ], + "image_footnote": [], + "bbox": [ + 133, + 400, + 864, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the generated dance has more similar geometric features, reflecting the quality of choreography. For dance diversity, we employ metrics $\\mathrm{Div}_k$ and $\\mathrm{Div}_g$ [18], which denote the distance of kinetic and distance of geometric features, respectively. Higher $\\mathrm{Div}_k$ and $\\mathrm{Div}_g$ signify the greater amplitude and diversity of motion, respectively. Finally, to evaluate beat-dance alignment, we use the Beat Align Score (BAS) [18], which quantifies the average distance between a music beat and its nearest dance beat. Higher BAS indicates greater degree of beat-dance alignment.", + "bbox": [ + 89, + 743, + 485, + 895 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 743, + 880, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the superiority of our method, we conduct comparative experiments with the state-of-the-art methods [12, 17, 18, 20, 27, 30, 31, 33, 36, 45], as well as the ground truth. Following [27, 30, 36], we generate 40 pieces of dance sequences on the AIST++ test set, and sample the generated dance sequence with the length of 20 seconds to compute the metrics mentioned above. Among the compared methods, \"Li et al.\" [17],", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13198", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DanceNet [45] and FACT [18] use the results from the AIST++ benchmark [18], while the results of DanceRevolution [12] are derived from its reproduction as presented in Bailando [30]. EDGE [33] utilizes the reproduced results from the Lodge [20]. The Bailando, DiffDance [27] and Lodge methods use the experimental results from their respective publications.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The comparative results are presented in Table 1. Our method achieves state-of-the-art performance, significantly outperforming existing methods in terms of $\\mathrm{FID}_k$ , $\\mathrm{Div}_g$ , $\\mathrm{Div}_k$ , and Beat Align Score (BAS), achieving the best results. Additionally, our method ranks third on the $\\mathrm{FID}_g$ metric. The generated dance examples of our proposed Danceba on various types of music are shown in Figure 4, and the demo videos are provided in the supplementary files.", + "bbox": [ + 89, + 198, + 480, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Quality Metrics: As illustrated in Table 1, our method outperforms existing methods in overall quality performance. To be specific, our method improves on $\\mathrm{FID}_k$ by $48.68\\%$ (11.67 vs. 22.74), while exhibits competitive edge of $\\mathrm{FID}_g$ . Further analysis of the employed quality metrics reveals distinct roles for the two metrics types. The metric $\\mathrm{FID}_k$ , defined based on velocity and energy, primarily captures the kinetic characteristics of the generated dance, while $\\mathrm{FID}_g$ , derived from multiple manually designed motion templates, emphasizes the geometric structural properties of the dance. Thus, our method demonstrates superior performance in modeling the kinetic aspects of sequential motions, as evidenced by its exceptional $\\mathrm{FID}_k$ scores. Although it exhibits a modest gap in $\\mathrm{FID}_g$ (11.90 vs. 11.58) compared to the state-of-the-art Bailando++ [31], it achieves a $19.0\\%$ higher $\\mathrm{Div}_g$ (7.55 vs. 6.34), indicating improved motion diversity with comparable generative quality. Despite this minor trade-off, our method outperforms existing approaches overall, achieving a robust balance of efficiency and quality across a broader range of metrics.", + "bbox": [ + 89, + 335, + 482, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Diversity Metrics: Regarding diversity evaluation, Danceba achieves the best diversity performance, which improves $\\mathrm{Div}_k$ by $7.0\\%$ (8.52 vs. 7.96) and even slightly outperforms the Ground Truth (8.52 vs. 8.19). Meanwhile, our method improves $\\mathrm{Div}_g$ by $16.3\\%$ (7.55 vs. 6.49) and also slightly surpasses the Ground Truth (7.55 vs. 7.45), as illustrated in Table 1. For a visual comparison, we also present a comparison with the state-of-the-art method Bailando, as illustrated in Figure 5. It can be observed that Bailando exhibits relatively small movement amplitudes, with some dance movements missing and the actions appearing stiff. In contrast, our method shows larger movement amplitudes and movement diversity.", + "bbox": [ + 89, + 642, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beat-dance Alignment Metrics: Danceba achieves the best performance in beat-dance alignment, as quantified by the BAS metric. Specifically, it improves BAS by $12.0\\%$ (0.2714 vs. 0.2423) compared to the previous best re", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e51635a02850c5a5e6e05580b01b05b7d4c0f246b64a550d94bf06e9342243d.jpg", + "image_caption": [ + "Figure 5. Comparison with the state-of-the-art method Bailando [30]. Visual comparisons with Bailando can be found in the supplementary video." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sult and surpasses the ground truth by $14.3\\%$ (0.2714 vs. 0.2374), as shown in Table 1. Figure 6 further visualizes the alignment between music beats and motion beats of the generated dance compared to Bailando++ [31]. The results show that Danceba achieves a significantly smaller beat distance, demonstrating its ability to generate dance sequences that more accurately follow musical rhythms.", + "bbox": [ + 511, + 345, + 903, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddfe8af1c30e219ea492b3880cb58289e2af562b3847287755c8ed37fd4fdaec.jpg", + "image_caption": [ + "Figure 6. Beats alignment visualization, where the horizontal axis shows frame indices of beat events. Comparing Danceba with Balando++ [31], we can find that the distance between motion beats and music beats generated by our method is smaller. This indicating that Danceba performs better in terms of rhythmic alignment." + ], + "image_footnote": [], + "bbox": [ + 521, + 465, + 705, + 556 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/424c83aebeab355a4a9655b97400f17d55a07a4764b95ef6bd8ee43b7ab8b947.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 465, + 895, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 651, + 678, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies to evaluate the contributions of the key modules (i.e., PRE, TGCA, and PMMM) in Danceba, where \"w/o\" denotes the removal of a specific module. \"w/o PMMM\" refers to the setting where PMMM is replaced with TGCA layers to ensure fairness. The results are presented in Table 2. Additionally, we assess our method using a single Mamba structure instead of the parallel Mamba structure for motion modeling, referred to as Danceba-Single, with results shown in Table 3.", + "bbox": [ + 511, + 674, + 903, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Note that, all ablation results are averaged over five independent training runs to ensure the credibility and robustness of our results. The motion quality score (FID) is computed as the average of $\\mathrm{FID}_k$ and $\\mathrm{FID}_g$ , while the motion diversity score (Div) is the average of $\\mathrm{Div}_k$ and $\\mathrm{Div}_g$ .", + "bbox": [ + 511, + 810, + 903, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Phase-Based Rhythm Extraction: To validate the effec", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13199", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6bf13e4f27a44c31d67508035a0d8f6731cf8be5ef6d9daf389f96d2c62a314.jpg", + "table_caption": [ + "Table 2. Ablation study of three key modules (i.e., PRE, TGCA, and PMMM), and \"w/o\" indicates Danceba without specified module. Visual comparisons with Bailando can be found in the supplementary video." + ], + "table_footnote": [], + "table_body": "
METHODFID ↓DIV ↑BAS ↑
GROUND TRUTH13.857.820.2374
w/o TGCA33.226.74-
w/o PMMM25.245.78-
w/o PRE20.626.530.2474
DANCEBA14.537.670.2779
", + "bbox": [ + 117, + 148, + 450, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tiveness of PRE, we replace the rhythm feature extraction module with a learnable linear layer as used in previous works [11, 30, 31, 36, 44]. Consequently, the motion quality (i.e., FID), motion diversity (i.e., Div), and beat alignment score (i.e., BAS) decrease by 6.09 (41.91%), 1.14 (14.86%), and 0.0305 (10.96%), respectively, as illustrated in Table 2. Specifically, this performance drop highlights the contribution of PPE, which effectively captures the critical rhythm features of music, a crucial aspect that previous methods have largely overlooked.", + "bbox": [ + 88, + 277, + 480, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal-Gated Causal Attention: As shown in Table 2, we validate the effectiveness of the proposed TGCA by training a Danceba variant that replaces TGCA with conventional causal cross-conditional attention. The results demonstrate that the full Danceba model outperforms this variant in both FID and Div, highlighting TGCA's ability to effectively leverage global rhythmic features extracted by PRE. This significantly enhances the contextual association between music and dance movements, resulting in dance sequences that not only accurately reflect musical rhythm but also exhibit greater coherence and expressive diversity.", + "bbox": [ + 89, + 430, + 482, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parallel Mamba Motion Modeling: To evaluate the effectiveness of Parallel Mamba Motion Modeling, we train a model without this module. As shown in Table 2, Danceba outperforms its variant without using PMMM in terms of both FID and Div, demonstrating the importance of parallel Mamba modeling. These results validate Mamba's superior sequential modeling capability, which is particularly well-suited for dance generation as it requires both fine-grained inter-modal interactions and long-sequence modeling.", + "bbox": [ + 89, + 597, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further assess the advantage of using a parallel Mamba architecture instead of a single Mamba architecture for motion modeling, we compare Danceba with Danceba-Single. Unlike Danceba-Single, which treats the body as a unified entity and disregards the distinction between upper and lower-body movements, PMMM explicitly models their distinct motion dynamics. Additionally, PMMM integrates rhythm-aware representation alongside these distinct pose features, enhancing beat-dance alignment with the musical rhythm. As shown in Table 3, Danceba significantly improves expressiveness and temporal coherence, with $\\mathrm{FID}_k$", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0dd0c8845ca5b1e932339cc43bb1f2a2caa369b4747c4fb371e1ea5fa9c52515.jpg", + "table_caption": [ + "Table 3. Ablation study on parallel Mamba motion modeling, where Danceba-Single denotes our method using single Mamba architecture." + ], + "table_footnote": [], + "table_body": "
METHODFIDk ↓FIDg ↑
GROUND TRUTH17.1010.60
DANCEBA-SINGLE82.5083.97
DANCEBA15.48(-67.02)13.57(-60.40)
", + "bbox": [ + 535, + 133, + 877, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and $\\mathrm{FID}_g$ scores improving by 67.02 and 60.40, respectively.", + "bbox": [ + 511, + 227, + 903, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 511, + 279, + 635, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although Danceba achieves superior dance generation quality compared to state-of-the-art methods, it has two key limitations: (1) Music feature encoding: Danceba uses a simple linear layer instead of a pre-trained audio model (e.g., Jukebox [3], MERT [21], or CLAP [5]). While computationally efficient, this design may fail to capture nuanced musical structures, affecting rhythm-motion synchronization. Future work could explore pre-trained music encoders to enhance feature extraction. (2) 3D pose quantization: Danceba adopts the Pose VQ-VAE framework from Bailando [30] for fair comparison with prior SOTA methods [30, 31, 36]. However, it has a limited encoding space, which may limit motion diversity. Such a trade-off between quantization efficiency and expressiveness could hinder the preservation of fine-grained motion details. Exploring hierarchical quantization or adaptive codebook learning could further improve motion quality.", + "bbox": [ + 511, + 306, + 906, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 585, + 633, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has presented a novel framework Danceba for music-driven dance generation that enhances rhythm-aware feature representation and motion modeling. By introducing phase-based rhythm extraction and temporal-gated causal attention, our approach ensures precise beat-dance synchronization and effectively integrates rhythmic structures into dance movements. Additionally, parallel Mamba motion modeling separately captures the distinct motion dynamics of the upper and lower body while incorporating rhythm-aware representations, improving the naturalness and diversity of the generated dance. Extensive experiments demonstrate that Danceba significantly outperforms state-of-the-art methods, generating dance movements that are more rhythmic, natural, and diverse while maintaining strong alignment with musical beats. Future work could integrate pre-trained audio encoders for richer music representation, explore advanced quantization to enhance motion expressiveness and diversity, or extend to style-conditioned motion synthesis for broader applicability.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13200", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the Fundamental Research Funds for the Central Universities (Grant No. 3072025YY0601), and the Beijing Nova Program (Grant No. 20230484261).", + "bbox": [ + 89, + 116, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 194, + 187, + 210 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Trans. Graph. (TOG), 2023. 2", + "[2] Tri Dao and Albert Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. In Proc. Int. Conf. Mach. Learn. (ICML), 2024. 2", + "[3] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020.8", + "[4] Kun Dong, Jian Xue, Zehai Niu, Xing Lan, Ke Lu, Qingyuan Liu, and Xiaoyu Qin. Realistic full-body motion generation from sparse tracking with state space model. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2024. 3", + "[5] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP learning audio concepts from natural language supervision. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 8", + "[6] Deepak Gopinath and Jungdam Won. Fairmotion - tools to load, process and visualize motion capture data. *GibHub*, 2020. 5", + "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2", + "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 5", + "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020. 1, 2", + "[10] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph. (TOG), 2016. 2", + "[11] Qiaochu Huang, Xu He, Boshi Tang, Haolin Zhuang, Liyang Chen, Shuochen Gao, Zhiyong Wu, Haozhi Huang, and Helen Meng. Enhancing expressiveness in dance generation via integrating frequency and music style information. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2024. 1, 2, 4, 8", + "[12] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In Proc. Int. Conf. Learn. Represent. (ICLR), 2021. 2, 6, 7" + ], + "bbox": [ + 93, + 220, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jinwoo Kim, Heeseok Oh, Seongjean Kim, Hoseok Tong, and Sanghoon Lee. A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 2", + "[14] Matthew Kyan, Guoyu Sun, Haiyan Li, Ling Zhong, Paisarn Muneesawang, Nan Dong, Bruce Elder, and Ling Guan. An approach to ballet dance training through MS Kinect and visualization in a CAVE virtual reality environment. ACM Trans. Intell. Syst. Technol. (TIST), 2015. 1", + "[15] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2019. 2", + "[16] Chengjian Li, Xiangbo Shu, Qiongjie Cui, Yazhou Yao, and Jinhui Tang. FTMoMamba: Motion generation with frequency and text state space models. arXiv preprint arXiv:2411.17532, 2024. 2, 3", + "[17] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with Transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6", + "[18] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3D dance generation with AIST++. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021. 2, 5, 6, 7", + "[19] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3D full body dance generation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1", + "[20] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024. 1, 6, 7", + "[21] Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Chenghao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, et al. MERT: Acoustic music understanding model with large-scale self-supervised training. In Proc. Int. Conf. Learn. Represent. (ICLR), 2024. 1, 2, 8", + "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 2015. 5", + "[23] Mixamo. https://www MIXamo.com/.6", + "[24] Meinard Müller, Tido Röder, and Michael Clausen. Efficient content-based retrieval of motion capture data. In Proc. ACM Conf. SIGGRAPH, 2005. 5", + "[25] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[26] Kensuke Onuma, Christos Faloutsos, and Jessica K. Hodgins. Fmdistance: A fast and effective distance function for motion capture data. In Eurographics, 2008. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13201", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Qiaosong Qi, Le Zhuo, Aixi Zhang, Yue Liao, Fei Fang, Si Liu, and Shuicheng Yan. DiffDance: Cascaded human motion diffusion model for dance generation. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2023. 1, 2, 5, 6, 7", + "[28] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. SMCD: High realism motion style transfer via Mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 3", + "[29] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1", + "[30] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3D dance generation by actor-critic GPT with choreographic memory. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[31] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3D dance GPT with choreographic memory. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1, 2, 4, 6, 7, 8", + "[32] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Trans. Multimedia (TMM), 2020. 2", + "[33] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1, 6, 7", + "[34] Shuhei Tsuchida. Dance information processing: Computational approaches for assisting dance composition. New Gener. Comput., 2024. 1", + "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 2", + "[36] Zilin Wang, Haolin Zhuang, Lu Li, Yinmin Zhang, Junjie Zhong, Jun Chen, Yu Yang, Boshi Tang, and Zhiyong Wu. Explore 3D dance generation via reward model from automatically-ranked demonstrations. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2024. 1, 4, 5, 6, 8", + "[37] Simon Welker, Julius Richter, and Timo Gerkmann. Speech enhancement with score-based generative models in the complex STFT domain. arXiv preprint arXiv:2203.17004, 2022. 4", + "[38] Shuo Yan, Gangyi Ding, Zheng Guan, Ningxiao Sun, Hong-song Li, and Longfei Zhang. OutsideMe: Augmenting dancer's external self-image by using a mixed reality system. In Proc. Annu. ACM Conf. Extended Abstr. Hum. Factors Comput. Syst. (CHI), 2015. 1", + "[39] Ling-An Zeng, Guohong Huang, Gaojie Wu, and Wei-Shi Zheng. Light-t2m: A lightweight and fast model for text-to-motion generation. arXiv preprint arXiv:2412.11193, 2024. 2, 3", + "[40] Zeyu Zhang, Akide Liu, Qi Chen, Feng chen, Reid Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. InfiniMotion: Mamba boosts memory in Transformer for arbitrary long motion generation. arXiv preprint arXiv:2407.10061, 2024.3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion Mamba: Efficient and long sequence motion generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2025. 2, 3", + "[42] Chenxi Zheng, Jing Qin, and Shengfeng He. Beat-It: Beatsynchronized multi-condition 3D dance generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2", + "[43] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1", + "[44] Haolin Zhuang, Shun Lei, Long Xiao, Weiqin Li, Liyang Chen, Sicheng Yang, Zhiyong Wu, Shiyan Kang, and Helen Meng. GTN-Bailando: Genre consistent long-term 3D dance generation based on pre-trained genre token network. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 1, 2, 4, 8", + "[45] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2Dance: DanceNet for music-driven dance generation. ACM Trans. Multimedia Comput. Commun. Appl. (TOMM), 2022. 2, 6, 7" + ], + "bbox": [ + 514, + 92, + 903, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13202", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_model.json b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbfb80151d1c4f24fdda169d7d3e0d41ae98332b --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.177 + ], + "angle": 0, + "content": "Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.203, + 0.728, + 0.24 + ], + "angle": 0, + "content": "Congyi Fan\\(^{1}\\), Jian Guan\\(^{1*}\\), Xuanjia Zhao\\(^{1}\\), Dongli Xu\\(^{2}\\), Youtian Lin\\(^{3}\\), Tong Ye\\(^{1}\\), Pengming Feng\\(^{4}\\), Haiwei Pan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.256, + 0.871, + 0.292 + ], + "angle": 0, + "content": "\\(^{1}\\)College of Computer Science and Technology, Harbin Engineering University, Harbin, China \\(^{2}\\)Shanghai Academy of Artificial Intelligence for Science, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.292, + 0.86, + 0.328 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Intelligence Science and Technology, Nanjing University, Nanjing, China \\(^{4}\\)State Key Laboratory of Space-Ground Integrated Information Technology, Beijing, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Automatically generating natural, diverse, and rhythmic human dance movements driven by music is vital for virtual reality and film industries. However, generating dance that naturally follows music remains a challenge, as existing methods lack proper beat alignment and exhibit unnatural motion dynamics. In this paper, we propose Danceba, a novel framework that leverages gating mechanism to enhance rhythm-aware feature representation for music-driven dance generation, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, we introduce Phase-Based Rhythm Extraction (PRE) to precisely extract rhythmic information from musical phase data, capitalizing on the intrinsic periodicity and temporal structures of music. Additionally, we propose Temporal-Gated Causal Attention (TGCA) to focus on global rhythmic features, ensuring that dance movements closely follow the musical rhythm. We also introduce Parallel Mamba Motion Modeling (PMMM) architecture to separately model upper and lower body motions along with musical features, thereby improving the naturalness and diversity of generated dance movements. Extensive experiments confirm that Danceba outperforms state-of-the-art methods, achieving improved rhythmic alignment and motion diversity. Project page: https://danceba.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.222, + 0.804 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.814, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Music-driven dance generation aims to automatically synthesize realistic and expressive 3D dance movements synchronized precisely with a given musical sequence, enabling natural animation of 3D character models. Due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.483 + ], + "angle": 0, + "content": "its broad applications in gaming, virtual reality, film production, and interactive entertainment [14, 34, 38, 43], this task has attracted considerable attention. However, generating dances that accurately reflect musical rhythm while preserving natural and diverse movement dynamics remains a significant challenge, as it requires effectively capturing and utilizing complex temporal relationships between music and motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Recently, generative models have demonstrated remarkable success across various tasks, such as text and image generation [9, 29]. Similarly, these generative approaches have also been applied to music-driven dance generation tasks [11, 19, 20, 27, 30, 31, 33, 36, 42, 44]. Among these, Bailando [30] introduces a cross-conditional causal attention mechanism to align musical cues with corresponding dance poses, significantly improving the quality of generated dance movements. However, its attention mechanism primarily captures local correspondence between music and poses, failing to sufficiently utilize global rhythmic characteristics within music, thus limiting the expressiveness and rhythmic coherence of the generated dances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Follow-up studies [11, 31, 36, 44] have extended Bailando's design to further improve dance generation performance. For instance, Enhancing-Bailando [11] employs a pretrained audio encoder (MERT) [21] to obtain richer musical feature representations. However, this encoder primarily captures semantic information from general audio data rather than explicitly modeling rhythmic characteristics, limiting its effectiveness in precise beat-dance synchronization. Moreover, Bailando and related methods typically concatenate three distinct input modalities, i.e., upper-body pose, lower-body pose, and musical features, and rely heavily on cross-conditional causal attention to align these inputs. Nevertheless, this attention mechanism primarily aligns pose and music at a local feature level, but struggles" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.355, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: j.guan@hrbeu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13193" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "to capture continuous motion dynamics, limiting the naturalness and diversity of generated dance movements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.482, + 0.258 + ], + "angle": 0, + "content": "To further tackle the rhythm alignment issue, Beat-It [42] explicitly introduces beat synchronization via a beat distance predictor. However, such an explicit synchronization approach imposes rigid control constraints and strong regularization during generation, significantly limiting the diversity and expressiveness of generated dances [27]. Thus, a more balanced approach capable of achieving precise rhythmic alignment without sacrificing motion naturalness and diversity is highly desirable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.44 + ], + "angle": 0, + "content": "To this end, we propose a novel framework equipped with a gating-enhanced rhythm-aware feature representation for music-driven dance generation, i.e., Danceba, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, Danceba introduces Phase-Based Rhythm Extraction (PRE), a rhythm-aware feature extraction module that explicitly captures rhythmic characteristics from musical phase information. Leveraging the temporal phase information intrinsic to rhythm perception, PRE effectively separates rhythmic from semantic music features, significantly improving the rhythmic alignment between generated dance sequences and musical inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.637 + ], + "angle": 0, + "content": "To further enhance beat-dance alignment using the rhythm-aware features extracted by PRE, it is essential to establish a stronger contextual association between musical rhythm and dance movements. To achieve this, we present a Temporal-Gated Causal Attention (TGCA) mechanism, enhanced by a gating strategy to explicitly prioritize and leverage global rhythmic structures within cross-conditional causal attention. By integrating gating mechanism, TGCA effectively strengthens the contextual connection between music rhythm and generated movements, ensuring that generated dance sequences accurately reflect the underlying musical rhythm and exhibit greater coherence and expressive diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Furthermore, to overcome the limitations of traditional attention-based approaches in motion modeling, we introduce Parallel Mamba Motion Modeling (PMMM), inspired by recent advances demonstrating that Mamba architectures achieve state-of-the-art performance in modeling complex human motion sequences [16, 39, 41]. Our parallel Mamba architecture separately models upper and lower-body dance sequences, with input features enhanced by Temporal-Gated Causal Attention (TGCA), which incorporates rhythm-aware representations alongside distinct upper and lower-body pose features. This design effectively captures diverse and nuanced motion characteristics. By leveraging Mamba's superior sequential modeling capability, our method significantly improves the expressiveness, diversity, and temporal coherence of generated dance movements. Fi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "nally, an additional TGCA module is applied after PMMM to further reinforce rhythmic alignment, ensuring that the generated dances precisely reflect the musical rhythm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.243 + ], + "angle": 0, + "content": "Extensive experiments show that our method improves \\(\\mathrm{FID}_k\\) by \\(48.68\\%\\), \\(\\mathrm{Div}_k\\) by \\(7.0\\%\\), \\(\\mathrm{Div}_g\\) by \\(16.3\\%\\), and BAS by \\(12.0\\%\\) over the second-best results, while maintaining a competitive \\(\\mathrm{FID}_g\\) of 11.90. This significantly surpasses state-of-the-art methods, producing dance movements that are more natural, diverse, and better synchronized with musical rhythm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.259, + 0.655, + 0.274 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.285, + 0.799, + 0.3 + ], + "angle": 0, + "content": "2.1. Music-Driven Dance Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Music-driven dance generation focuses on automatically synthesizing 3D dance movements from musical input, creating temporally coherent pose sequences for animating 3D character models. It has attracted increasing attention in both the computer vision and virtual avatar communities [1, 13] due to its applications in entertainment, gaming, and virtual reality. Early approaches, including [10, 12, 15, 17, 18, 32, 45], investigated diverse methodologies for this task. The study in [17] utilized a standard Transformer [35] with a late-fusion module to predict subsequent dance steps. The FACT framework [18] introduced the AIST++ dataset and employed a cross-modal Transformer with full attention mechanisms for music-to-dance generation. However, these initial methods neglect the rhythm features of music and struggle to effectively integrate musical and motion features, leading to suboptimal generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Bailando [30] reformulated dance generation as a next-step pose prediction problem using cross-conditional causal attention, achieving high-quality music-driven dance synthesis. Subsequent works [11, 31, 44] extended this framework: Bailando++ [31] enhanced pattern recognition with musical context encoding, while Enhancing-Bailando [11] incorporated a pre-trained MERT audio encoder [21] for improved feature extraction. However, these methods often overlook rhythmic structures or rely on generic audio features. Beat-It [42] introduced explicit beat alignment via a conditional diffusion model [9, 25], ensuring precise synchronization. Yet, its reliance on rigid beat control signals and strong regularization may constrain pose diversity [27]. In contrast, we emphasize rhythm-aware feature representation, achieving superior rhythmic alignment without sacrificing movement diversity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.844, + 0.833 + ], + "angle": 0, + "content": "2.2. Mamba in Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The integration of Mamba [7] and Mamba2 [2] into human motion generation represents a pivotal advancement, leveraging their efficient sequence modeling capabilities. Within human motion generation, Mamba surpasses con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Note that, the terms \"beat\" and \"rhythm\" are used interchangeably in music processing following DiffDance [27] and Bailando++ [31]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13194" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.91, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.35, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 1. The overall framework of Danceba consists of three core modules: Phase-based Rhythm Extraction (PRE), Temporal-Gated Causal Attention (TGCA), and Parallel Mamba Motion Modeling (PMMM). PRE precisely extracts rhythm-aware features from musical phase information, providing accurate rhythmic signals. These rhythm-enhanced music features are fused with pose embeddings (upper-body and lower-body poses) and processed through the TGCA module, which utilizes gating mechanism to reinforce rhythmic sensitivity and align dance movements accurately with music beats. Parallel Mamba Motion Modeling separately models upper and lower body motion sequences, effectively capturing distinct dance dynamics aligned with rhythm-aware music features, thus significantly enhancing the naturalness, diversity, and temporal coherence of the generated dance movements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.486, + 0.685 + ], + "angle": 0, + "content": "ventional Transformers in performance, as demonstrated by Motion Mamba [41], which highlights its efficacy for extended motion sequence synthesis. Furthermore, InfiniMotion [40] utilizes Mamba's hierarchical and bidirectional temporal processing to generate continuous, smooth motion sequences of indefinite length. Recent approaches including SMCD [28], Light-T2M [39], FTMoMamba [16], and MMD [4] have further incorporated Mamba architectures, advancing the field through enhanced efficiency and high-fidelity motion modeling. Building upon recent advancements, Mamba has achieved state-of-the-art performance in human motion modeling [16, 39, 41], motivating our selection of Mamba over self-attention mechanisms for information exchange within dance sequences, which significantly enhances the diversity of generated dance motions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.707, + 0.265, + 0.723 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.901 + ], + "angle": 0, + "content": "This section presents our proposed Danceba framework for music-driven 3D dance generation. By leveraging gating-enhanced rhythm-aware features, Danceba ensures precise alignment between dance poses and music. The overall framework is illustrated in Figure 1, and its components are detailed as follows: Section 3.1 outlines the preliminary framework and key notations. Specifically, we introduce phase-based rhythm feature extraction (Section 3.2), global beat-dance alignment via temporal-gated causal attention (Section 3.3), and Mamba-based parallel motion modeling (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.456, + 0.644, + 0.472 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Given a dance sequence \\(\\mathbf{P} \\in \\mathbb{R}^{T \\times (J \\times 3)}\\), where \\(T\\) is the time length and \\(J\\) is joint amount. Using two pretrained Pose VQ-VAEs [30] \\(\\mathcal{F}_{VAE}^u\\) and \\(\\mathcal{F}_{VAE}^l\\) with codebook \\(\\mathcal{Z}^u\\) and \\(\\mathcal{Z}^l\\) for the upper and lower half bodies, it separately encodes upper and lower body movements into compositional code pairs \\(\\mathbf{p} = [\\mathbf{p}^u, \\mathbf{p}^l]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.575, + 0.907, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {u} = \\mathcal {F} _ {V A E} ^ {u} (\\mathbf {P}), \\mathbf {p} ^ {l} = \\mathcal {F} _ {V A E} ^ {l} (\\mathbf {P}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}^u\\in (\\mathcal{Z}^u)^{T'}\\) \\(\\mathbf{p}^l\\in (\\mathcal{Z}^l)^{T'}\\) .Here \\(T^{\\prime} = T / \\lambda\\) \\(\\lambda\\) is the temporal down-sampling rate. The input music feature \\(\\mathbf{m}\\in \\mathbb{R}^{T'\\times D_m}\\) , where \\(D_{m}\\) is the channel dimension of the input music features. Then we embed music features \\(\\mathbf{m}\\) upper \\(p^u\\) and lower \\(p^l\\) pose codes to learnable features with three separate linear layers: \\(\\mathbf{X}_{\\mathrm{m}} = \\mathrm{Linear}(\\mathbf{m}),\\mathbf{X}_{\\mathrm{u}} =\\) Linear \\((\\mathbf{p}^{u})\\) and \\(\\mathbf{X}_1 = \\mathrm{Linear}(\\mathbf{p}^l)\\) , where \\(\\mathbf{X}_{\\mathrm{m}}\\in \\mathbb{R}^{T^{\\prime}\\times D}\\) \\(\\mathbf{X}_{\\mathrm{u}}\\in \\mathbb{R}^{T^{\\prime}\\times D}\\) \\(\\mathbf{X}_1\\in \\mathbb{R}^{T^{\\prime}\\times D}\\) and D is the channel dimension of the features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.909, + 0.844 + ], + "angle": 0, + "content": "The dance generation task is reframed as selecting the most probable future pose code from the codebook \\(\\mathcal{Z}\\), conditioned on the music and prior movements. Since the upper and lower bodies are modeled separately, maintaining coherence and avoiding asynchrony (e.g., opposing directions) requires cross-conditioned prediction, leveraging mutual information between existing movements:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.849, + 0.907, + 0.872 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {t} ^ {u} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {u} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.874, + 0.907, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\hat {p} _ {t} ^ {l} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {l} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right). \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "At each time step \\(t\\), MotionGPT [30] estimates the probabilities of pose codes \\(\\mathbf{z}_i \\in \\mathbb{Z}\\) and selects the one with the most probable pose codes as the predicted upper pose \\(\\hat{p}_t^u\\) and predicted lower pose \\(\\hat{p}_t^l\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Specifically, the input feature \\(\\mathbf{X}_i\\in \\mathbb{R}^{(3\\times T')\\times D}\\) are concatenated by upper body motion feature, lower body motion feature and music feature as follows: \\(\\mathbf{X}_i =\\) \\(\\mathrm{Concat}(\\mathbf{X}_{\\mathrm{m}},\\mathbf{X}_{\\mathrm{u}},\\mathbf{X}_{\\mathrm{l}})\\). Subsequently, \\(\\mathbf{X}_i\\) is passed through successive Autoregressive Transformer layers \\(\\mathcal{F}_{AR}\\), followed by a linear transformation and a softmax layer, generates the probability distributions for predicting body pose collectively:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.283, + 0.483, + 0.3 + ], + "angle": 0, + "content": "\\[\n\\mathbf {a} ^ {h} = \\operatorname {S o f t m a x} (\\operatorname {L i n e a r} \\left(\\mathcal {F} _ {A R} \\left(\\mathbf {X} _ {i}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.482, + 0.341 + ], + "angle": 0, + "content": "where \\(\\mathbf{a}^h = [\\mathbf{a}^u,\\mathbf{a}^l]\\), \\(\\mathbf{a}^u\\) denotes that of upper body and \\(\\mathbf{a}^l\\) denotes that of lower body." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.483, + 0.387 + ], + "angle": 0, + "content": "Finally, the model is optimized via supervised training with the cross-entropy loss on action probability \\( \\mathbf{a}_t^h \\) at each time step \\( t \\) and for both upper body and lower body:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.398, + 0.483, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C E} = \\frac {1}{T ^ {\\prime}} \\sum_ {t = 0} ^ {T ^ {\\prime} - 1} \\sum_ {h = u, l} \\text {C r o s s E n t r o p y} \\left(\\mathrm {a} _ {t} ^ {h}, p _ {t + 1} ^ {h}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.497 + ], + "angle": 0, + "content": "In inference, the MotionGPT predicts pose codes according to the initial pose code and the entire music, followed by Pose VQ-VAE decoders generating a new dance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.506, + 0.442, + 0.521 + ], + "angle": 0, + "content": "3.2. Phase-Based Rhythm Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.483, + 0.678 + ], + "angle": 0, + "content": "To enhance rhythmic feature extraction in music-driven dance generation, we propose Phase-Based Rhythm Extraction (PRE), which decouples rhythm from musical semantics by leveraging the inherent periodicity and temporal positioning of phase representations. The architecture of PRE is shown in the green part of Figure 1. Unlike existing methods that entangle rhythm with melody and harmony, PRE enables independent and precise rhythm modeling as shown in Figure 2, forming a solid foundation for gating-enhanced rhythm-aware mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.785 + ], + "angle": 0, + "content": "Specifically, we adopt the Short-Time Fourier Transform (STFT) instead of the conventional Fourier Transform (FT), since the non-stationary nature of music and the sliding-window approach of STFT effectively captures sudden tempo shifts and complex rhythmic patterns [37]. Specifically, we first perform STFT on the input music feature to extract phase angles:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.795, + 0.483, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S} _ {m} = \\operatorname {S T F T} (\\mathbf {m}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.814, + 0.354, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\varphi} = \\operatorname {A n g l e} \\left(\\mathbf {S} _ {m}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{S}_m\\) represents the complex spectrogram obtained via STFT, \\(\\mathrm{Angle}(\\cdot)\\) computes the phase angle of a complex-valued input and \\(\\varphi^{\\prime}\\) denotes the phase angles extracted from \\(\\mathbf{S}_m\\). By taking the angle of the STFT output, we obtain the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.253, + 0.905, + 0.335 + ], + "angle": 0, + "content": "Figure 2. Visualizing the input music features and phase features. We notice that only a small portion of the input music features contains meaningful musical information, i.e., rhythm-poor features. And after a short-time Fourier transform (STFT), each corresponding token has a rich periodic Phase Feature, i.e., rhythm-rich features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.905, + 0.392 + ], + "angle": 0, + "content": "music phase data \\(\\varphi^{\\prime}\\in \\mathbb{C}^{T_{\\varphi}\\times D_{\\varphi}}\\) , where \\(T_{\\varphi}\\) represents the number of time frames obtained after the STFT transformation, and \\(D_{\\varphi}\\) is the channel dimension of the \\(\\varphi^{\\prime}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.905, + 0.482 + ], + "angle": 0, + "content": "The extracted phase angles \\(\\varphi^{\\prime}\\) inherently reflect temporal shifts and periodic structures of musical rhythms: low-frequency phase variations correspond to broader rhythmic patterns, while high-frequency variations capture finer rhythmic details. We also apply a center-cropping strategy, selectively focusing on the most informative phase regions:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.496, + 0.905, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\varphi = \\operatorname {C e n t e r C r o p} \\left(\\varphi^ {\\prime}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.905, + 0.562 + ], + "angle": 0, + "content": "where \\(\\varphi \\in \\mathbb{C}^{T' \\times D_{\\varphi}}\\) has the same temporal length \\(T'\\), addressing the potential temporal misalignment and preserving critical rhythmic cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.905, + 0.608 + ], + "angle": 0, + "content": "Subsequently, the phase features through a linear transformation, enhancing phase features representing musical rhythms. The process is shown in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.617, + 0.905, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\varphi} = \\operatorname {R e L U} (\\mathrm {B N} (\\operatorname {L i n e a r} (\\varphi))), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "where the embedded phase features \\(\\mathbf{X}_{\\varphi} \\in \\mathbb{C}^{T' \\times D}\\) encode rich rhythmic information, providing a clearer and more structured representation of rhythmic and musical elements. Finally, we fuse the input feature \\(\\mathbf{X}_i\\) with the rhythmic features \\(\\mathbf{X}_{\\varphi}\\) by concatenating them: \\(\\mathbf{X}_{\\gamma} = \\mathrm{Concat}(\\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi})\\). Then, we sum the rhythm-enhanced features with the original music, upper body, and lower body motion features: \\(\\mathbf{X} = \\mathbf{X}_f + \\mathbf{X}_{\\gamma}\\), where \\(\\mathbf{X}\\) represents the rhythmically enhanced features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.882, + 0.805 + ], + "angle": 0, + "content": "3.3. Global Beat Attention via Temporal Gating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Leveraging these rhythm-rich features, we next investigate how to integrate them into the alignment process between music and dance. However, prior methods [11, 30, 31, 36, 44] relying on cross-conditional causal attention mechanism (denoted as \\(\\mathrm{C}^3\\mathrm{Attention}\\)) exhibit a critical misalignment issue. As shown in Figure 3 (a), we visualize the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "cross-conditional causal attention mechanism and observe that the next predicted token lacks direct control signals from historical tokens, failing to establish a stable global beat attention. This misalignment leads to error accumulation, which we identify as a key factor behind the unnatural dance generation in prior work." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.195, + 0.257, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.321, + 0.25, + 0.342 + ], + "angle": 0, + "content": "(a) Cross-Conditional Causal Attention" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.195, + 0.481, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.321, + 0.446, + 0.342 + ], + "angle": 0, + "content": "(b) Temporal-Gated Causal Attention (TGCA)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.452 + ], + "angle": 0, + "content": "Figure 3. Heat map of attention. The two displayed images show the results learned by Cross-Conditional Causal Attention (a) and Temporal-Gated Causal Attention (b) under the same piece of music. The brighter the color, the higher the attention; the darker the color, the lower the attention. TGCA the next token has a clear control signal, preventing overly random dance movements and achieving enhanced rhythm-aware feature representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Building on this, we propose Temporal Gated Causal Attention (TGCA) to enhance global rhythm feature attention, improving music-dance alignment and ensuring the next token is guided by explicit historical information, preventing random action generation. As illustrated in the blue section of Figure 1, the rhythmically enhanced features \\(\\mathbf{X}\\) are passed through TGCA module, which can be expressed as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.597, + 0.483, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G a t i n g} (\\mathbf {X}) = \\operatorname {S i L U} (\\operatorname {L i n e a r} (\\mathbf {X})). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.667 + ], + "angle": 0, + "content": "The final output is obtained by the element-wise multiplication of cross-conditional causal attention \\( \\mathrm{C}^3 \\) Attention and Gating, as shown in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.676, + 0.482, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {X} _ {a t t n} = \\operatorname {T G C A} (\\mathbf {X}) \\\\ = \\mathrm {C} ^ {3} \\text {A t t e n t i o n} (\\mathbf {X}) \\odot \\operatorname {G a t i n g} (\\mathbf {X}). \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.781 + ], + "angle": 0, + "content": "As shown in Figure 3 (b), our proposed gated causal attention mechanism allows historical tokens to better predict the next token, implying an enhanced influence of global rhythmic patterns on dance generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.442, + 0.805 + ], + "angle": 0, + "content": "3.4. Mamba-Based Parallel Motion Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "After TGCA enhances dance movements in \\(\\mathbf{X}_{attn}\\) with global rhythmic perception, we looking into dance modeling for generation. Dance generation faces two key challenges: (1) Dance is a complex spatial-temporal sequence encoding trajectory, posture, and velocity over time, requiring a model capable of capturing both local transitions and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "global rhythmic variations. (2) The upper and lower body exhibit distinct movement patterns and speeds, necessitating separate processing to maintain coherence and diversity. To address these issues, we propose a parallel Mamba motion modeling architecture that separately models music, upper-body, and lower-body movements while leveraging state space modeling to ensure fluid and diverse motion generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Specifically, to ensure each action is primarily influenced by relevant preceding movements, we introduce a GateMlp in the Mamba block, as shown in Figure 1. This mechanism selectively retains essential motion features while filtering out less relevant information, thereby enhancing the coherence and expressiveness of the generated dance sequences. The process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.33, + 0.905, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {m b} ^ {\\prime} = \\operatorname {M a m b a} (\\operatorname {R M S N o r m} (\\mathbf {X} _ {a t t n})) + \\mathbf {X} _ {a t t n}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.356, + 0.905, + 0.374 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {m b} = \\operatorname {G a t e M l p} \\left(\\operatorname {R M S N o r m} \\left(\\mathbf {X} _ {m b} ^ {\\prime}\\right)\\right) + \\mathbf {X} _ {m b} ^ {\\prime}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.491 + ], + "angle": 0, + "content": "The gating mechanism dynamically regulates the frequency, speed, and amplitude of movements in response to musical beats, ensuring that dance sequences remain synchronized with the music. Finally, three parallel Mamba architectures model upper- and lower-body dance sequences separately, therefore ensure the generated dance movements are expressive, diverse, and temporally coherent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.646, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.704, + 0.548 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Dataset: We follow prior works [27, 30, 36] and perform all experiments on the AIST++ dataset [18], the most widely used benchmark for music-driven dance generation that avoids observable biases (e.g., offensive postures). AIST++ consists of 991 high-quality 3D pose sequences recorded at 60 FPS in skinned multi-person linear (SMPL) format [22], with 951 sequences designated for training and 40 reserved for evaluation. Following [27, 30, 36], we generate 40 pieces of dance sequences on the test set of AIST++ dataset, and sample the generated dance sequence with length of 20 seconds for performance evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics: We follow previous studies [18, 30] to quantitatively evaluate the generated samples in three key aspects: quality, diversity, and beat-dance alignment. Prior to evaluation, it is necessary to extract kinetic features [26] and geometric features [24] from both the ground truth and the generated samples via a specific toolbox [6]. For dance quality, we compute two types of the Fréchet Inception Distance (FID) [8], including \\(\\mathrm{FID}_k\\) based on kinetic features and \\(\\mathrm{FID}_g\\) based on geometric features. Lower \\(\\mathrm{FID}_k\\) indicates that the generated dance has more similar kinetic features to the original dance, meaning the generated dances are more natural and smooth. Lower \\(\\mathrm{FID}_g\\) indicates that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13197" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on the AIST++ dataset. Underlining indicates the best performance among existing methods. Blue indicates results that surpass the best existing method. ↓ indicates that lower values are better, while ↑ indicates that higher values are better. * represents the dances generated by \"Li et al.\" that exhibit significant jitteriness, resulting in extremely high velocity variation, as also noted in [18]. † indicates the reproduced results obtained using Bailando++ official code and published checkpoints. ‡ denotes the results that are averaged over five independent training runs." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.171, + 0.898, + 0.391 + ], + "angle": 0, + "content": "
METHODVENUEMOTION QUALITYMOTION DIVERSITYBEAT ALIGN SCORE↑
FIDk↓FIDg↓DIVk↑DIVg↑
GROUND TRUTH-17.1010.608.197.450.2374
Li et al. [17]ARXIV 202086.4343.46*6.853.320.1607
DANCERUVOLUTION [12]ICLR 202173.4225.923.524.870.1950
DANCENET [45]TOMM 202269.1825.492.862.850.1430
FACT [18]ICCV 202135.3522.115.946.180.2209
BAILANDO [30]CVPR 202228.169.627.836.340.2332
EDGE [33]CVPR 202342.1622.123.964.610.2334
DIFFDANCE [27]ACMMM 202324.0920.686.022.890.2418
†BAILANDO++ [31]TPAMI 202322.7411.587.946.340.2263
E3D2 [36]AAAI 202426.258.947.966.490.2232
LODGGE [20]CVPR 202437.0918.795.584.850.2423
OURS (BEST)11.67(-11.07)11.908.52(+0.58)7.55(+1.21)0.2714(+2.91%)
‡OURS (AVERAGE)15.48(-7.26)13.577.857.49(+1.15)0.2779(+3.56%)
" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.401, + 0.866, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.699, + 0.908, + 0.729 + ], + "angle": 0, + "content": "Figure 4. Dance examples generated by our proposed method on various types of music. The demo videos can be found in the supplementary material. The character is sourced from Mixamo [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.486, + 0.896 + ], + "angle": 0, + "content": "the generated dance has more similar geometric features, reflecting the quality of choreography. For dance diversity, we employ metrics \\(\\mathrm{Div}_k\\) and \\(\\mathrm{Div}_g\\) [18], which denote the distance of kinetic and distance of geometric features, respectively. Higher \\(\\mathrm{Div}_k\\) and \\(\\mathrm{Div}_g\\) signify the greater amplitude and diversity of motion, respectively. Finally, to evaluate beat-dance alignment, we use the Beat Align Score (BAS) [18], which quantifies the average distance between a music beat and its nearest dance beat. Higher BAS indicates greater degree of beat-dance alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.744, + 0.882, + 0.76 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To validate the superiority of our method, we conduct comparative experiments with the state-of-the-art methods [12, 17, 18, 20, 27, 30, 31, 33, 36, 45], as well as the ground truth. Following [27, 30, 36], we generate 40 pieces of dance sequences on the AIST++ test set, and sample the generated dance sequence with the length of 20 seconds to compute the metrics mentioned above. Among the compared methods, \"Li et al.\" [17]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.198 + ], + "angle": 0, + "content": "DanceNet [45] and FACT [18] use the results from the AIST++ benchmark [18], while the results of DanceRevolution [12] are derived from its reproduction as presented in Bailando [30]. EDGE [33] utilizes the reproduced results from the Lodge [20]. The Bailando, DiffDance [27] and Lodge methods use the experimental results from their respective publications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.336 + ], + "angle": 0, + "content": "The comparative results are presented in Table 1. Our method achieves state-of-the-art performance, significantly outperforming existing methods in terms of \\(\\mathrm{FID}_k\\), \\(\\mathrm{Div}_g\\), \\(\\mathrm{Div}_k\\), and Beat Align Score (BAS), achieving the best results. Additionally, our method ranks third on the \\(\\mathrm{FID}_g\\) metric. The generated dance examples of our proposed Danceba on various types of music are shown in Figure 4, and the demo videos are provided in the supplementary files." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.64 + ], + "angle": 0, + "content": "Motion Quality Metrics: As illustrated in Table 1, our method outperforms existing methods in overall quality performance. To be specific, our method improves on \\(\\mathrm{FID}_k\\) by \\(48.68\\%\\) (11.67 vs. 22.74), while exhibits competitive edge of \\(\\mathrm{FID}_g\\). Further analysis of the employed quality metrics reveals distinct roles for the two metrics types. The metric \\(\\mathrm{FID}_k\\), defined based on velocity and energy, primarily captures the kinetic characteristics of the generated dance, while \\(\\mathrm{FID}_g\\), derived from multiple manually designed motion templates, emphasizes the geometric structural properties of the dance. Thus, our method demonstrates superior performance in modeling the kinetic aspects of sequential motions, as evidenced by its exceptional \\(\\mathrm{FID}_k\\) scores. Although it exhibits a modest gap in \\(\\mathrm{FID}_g\\) (11.90 vs. 11.58) compared to the state-of-the-art Bailando++ [31], it achieves a \\(19.0\\%\\) higher \\(\\mathrm{Div}_g\\) (7.55 vs. 6.34), indicating improved motion diversity with comparable generative quality. Despite this minor trade-off, our method outperforms existing approaches overall, achieving a robust balance of efficiency and quality across a broader range of metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Motion Diversity Metrics: Regarding diversity evaluation, Danceba achieves the best diversity performance, which improves \\(\\mathrm{Div}_k\\) by \\(7.0\\%\\) (8.52 vs. 7.96) and even slightly outperforms the Ground Truth (8.52 vs. 8.19). Meanwhile, our method improves \\(\\mathrm{Div}_g\\) by \\(16.3\\%\\) (7.55 vs. 6.49) and also slightly surpasses the Ground Truth (7.55 vs. 7.45), as illustrated in Table 1. For a visual comparison, we also present a comparison with the state-of-the-art method Bailando, as illustrated in Figure 5. It can be observed that Bailando exhibits relatively small movement amplitudes, with some dance movements missing and the actions appearing stiff. In contrast, our method shows larger movement amplitudes and movement diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Beat-dance Alignment Metrics: Danceba achieves the best performance in beat-dance alignment, as quantified by the BAS metric. Specifically, it improves BAS by \\(12.0\\%\\) (0.2714 vs. 0.2423) compared to the previous best re" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.294, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Comparison with the state-of-the-art method Bailando [30]. Visual comparisons with Bailando can be found in the supplementary video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.905, + 0.452 + ], + "angle": 0, + "content": "sult and surpasses the ground truth by \\(14.3\\%\\) (0.2714 vs. 0.2374), as shown in Table 1. Figure 6 further visualizes the alignment between music beats and motion beats of the generated dance compared to Bailando++ [31]. The results show that Danceba achieves a significantly smaller beat distance, demonstrating its ability to generate dance sequences that more accurately follow musical rhythms." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.467, + 0.707, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.467, + 0.897, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.571, + 0.905, + 0.641 + ], + "angle": 0, + "content": "Figure 6. Beats alignment visualization, where the horizontal axis shows frame indices of beat events. Comparing Danceba with Balando++ [31], we can find that the distance between motion beats and music beats generated by our method is smaller. This indicating that Danceba performs better in terms of rhythmic alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.679, + 0.667 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.81 + ], + "angle": 0, + "content": "We conduct ablation studies to evaluate the contributions of the key modules (i.e., PRE, TGCA, and PMMM) in Danceba, where \"w/o\" denotes the removal of a specific module. \"w/o PMMM\" refers to the setting where PMMM is replaced with TGCA layers to ensure fairness. The results are presented in Table 2. Additionally, we assess our method using a single Mamba structure instead of the parallel Mamba structure for motion modeling, referred to as Danceba-Single, with results shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.886 + ], + "angle": 0, + "content": "Note that, all ablation results are averaged over five independent training runs to ensure the credibility and robustness of our results. The motion quality score (FID) is computed as the average of \\(\\mathrm{FID}_k\\) and \\(\\mathrm{FID}_g\\), while the motion diversity score (Div) is the average of \\(\\mathrm{Div}_k\\) and \\(\\mathrm{Div}_g\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Phase-Based Rhythm Extraction: To validate the effec" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13199" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 2. Ablation study of three key modules (i.e., PRE, TGCA, and PMMM), and \"w/o\" indicates Danceba without specified module. Visual comparisons with Bailando can be found in the supplementary video." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.149, + 0.451, + 0.251 + ], + "angle": 0, + "content": "
METHODFID ↓DIV ↑BAS ↑
GROUND TRUTH13.857.820.2374
w/o TGCA33.226.74-
w/o PMMM25.245.78-
w/o PRE20.626.530.2474
DANCEBA14.537.670.2779
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.482, + 0.429 + ], + "angle": 0, + "content": "tiveness of PRE, we replace the rhythm feature extraction module with a learnable linear layer as used in previous works [11, 30, 31, 36, 44]. Consequently, the motion quality (i.e., FID), motion diversity (i.e., Div), and beat alignment score (i.e., BAS) decrease by 6.09 (41.91%), 1.14 (14.86%), and 0.0305 (10.96%), respectively, as illustrated in Table 2. Specifically, this performance drop highlights the contribution of PPE, which effectively captures the critical rhythm features of music, a crucial aspect that previous methods have largely overlooked." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Temporal-Gated Causal Attention: As shown in Table 2, we validate the effectiveness of the proposed TGCA by training a Danceba variant that replaces TGCA with conventional causal cross-conditional attention. The results demonstrate that the full Danceba model outperforms this variant in both FID and Div, highlighting TGCA's ability to effectively leverage global rhythmic features extracted by PRE. This significantly enhances the contextual association between music and dance movements, resulting in dance sequences that not only accurately reflect musical rhythm but also exhibit greater coherence and expressive diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Parallel Mamba Motion Modeling: To evaluate the effectiveness of Parallel Mamba Motion Modeling, we train a model without this module. As shown in Table 2, Danceba outperforms its variant without using PMMM in terms of both FID and Div, demonstrating the importance of parallel Mamba modeling. These results validate Mamba's superior sequential modeling capability, which is particularly well-suited for dance generation as it requires both fine-grained inter-modal interactions and long-sequence modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To further assess the advantage of using a parallel Mamba architecture instead of a single Mamba architecture for motion modeling, we compare Danceba with Danceba-Single. Unlike Danceba-Single, which treats the body as a unified entity and disregards the distinction between upper and lower-body movements, PMMM explicitly models their distinct motion dynamics. Additionally, PMMM integrates rhythm-aware representation alongside these distinct pose features, enhancing beat-dance alignment with the musical rhythm. As shown in Table 3, Danceba significantly improves expressiveness and temporal coherence, with \\(\\mathrm{FID}_k\\)" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.131 + ], + "angle": 0, + "content": "Table 3. Ablation study on parallel Mamba motion modeling, where Danceba-Single denotes our method using single Mamba architecture." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.135, + 0.878, + 0.212 + ], + "angle": 0, + "content": "
METHODFIDk ↓FIDg ↑
GROUND TRUTH17.1010.60
DANCEBA-SINGLE82.5083.97
DANCEBA15.48(-67.02)13.57(-60.40)
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.26 + ], + "angle": 0, + "content": "and \\(\\mathrm{FID}_g\\) scores improving by 67.02 and 60.40, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.28, + 0.637, + 0.296 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.566 + ], + "angle": 0, + "content": "Although Danceba achieves superior dance generation quality compared to state-of-the-art methods, it has two key limitations: (1) Music feature encoding: Danceba uses a simple linear layer instead of a pre-trained audio model (e.g., Jukebox [3], MERT [21], or CLAP [5]). While computationally efficient, this design may fail to capture nuanced musical structures, affecting rhythm-motion synchronization. Future work could explore pre-trained music encoders to enhance feature extraction. (2) 3D pose quantization: Danceba adopts the Pose VQ-VAE framework from Bailando [30] for fair comparison with prior SOTA methods [30, 31, 36]. However, it has a limited encoding space, which may limit motion diversity. Such a trade-off between quantization efficiency and expressiveness could hinder the preservation of fine-grained motion details. Exploring hierarchical quantization or adaptive codebook learning could further improve motion quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.634, + 0.603 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper has presented a novel framework Danceba for music-driven dance generation that enhances rhythm-aware feature representation and motion modeling. By introducing phase-based rhythm extraction and temporal-gated causal attention, our approach ensures precise beat-dance synchronization and effectively integrates rhythmic structures into dance movements. Additionally, parallel Mamba motion modeling separately captures the distinct motion dynamics of the upper and lower body while incorporating rhythm-aware representations, improving the naturalness and diversity of the generated dance. Extensive experiments demonstrate that Danceba significantly outperforms state-of-the-art methods, generating dance movements that are more rhythmic, natural, and diverse while maintaining strong alignment with musical beats. Future work could integrate pre-trained audio encoders for richer music representation, explore advanced quantization to enhance motion expressiveness and diversity, or extend to style-conditioned motion synthesis for broader applicability." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13200" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.486, + 0.178 + ], + "angle": 0, + "content": "This work was partly supported by the Fundamental Research Funds for the Central Universities (Grant No. 3072025YY0601), and the Beijing Nova Program (Grant No. 20230484261)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.195, + 0.188, + 0.211 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Trans. Graph. (TOG), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[2] Tri Dao and Albert Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. In Proc. Int. Conf. Mach. Learn. (ICML), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.336, + 0.483, + 0.391 + ], + "angle": 0, + "content": "[3] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.394, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[4] Kun Dong, Jian Xue, Zehai Niu, Xing Lan, Ke Lu, Qingyuan Liu, and Xiaoyu Qin. Realistic full-body motion generation from sparse tracking with state space model. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[5] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP learning audio concepts from natural language supervision. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[6] Deepak Gopinath and Jungdam Won. Fairmotion - tools to load, process and visualize motion capture data. *GibHub*, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.599, + 0.482, + 0.667 + ], + "angle": 0, + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.671, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph. (TOG), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[11] Qiaochu Huang, Xu He, Boshi Tang, Haolin Zhuang, Liyang Chen, Shuochen Gao, Zhiyong Wu, Haozhi Huang, and Helen Meng. Enhancing expressiveness in dance generation via integrating frequency and music style information. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2024. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In Proc. Int. Conf. Learn. Represent. (ICLR), 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Jinwoo Kim, Heeseok Oh, Seongjean Kim, Hoseok Tong, and Sanghoon Lee. A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Matthew Kyan, Guoyu Sun, Haiyan Li, Ling Zhong, Paisarn Muneesawang, Nan Dong, Bruce Elder, and Ling Guan. An approach to ballet dance training through MS Kinect and visualization in a CAVE virtual reality environment. ACM Trans. Intell. Syst. Technol. (TIST), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[15] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[16] Chengjian Li, Xiangbo Shu, Qiongjie Cui, Yazhou Yao, and Jinhui Tang. FTMoMamba: Motion generation with frequency and text state space models. arXiv preprint arXiv:2411.17532, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[17] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with Transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[18] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3D dance generation with AIST++. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[19] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3D full body dance generation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[20] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[21] Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Chenghao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, et al. MERT: Acoustic music understanding model with large-scale self-supervised training. In Proc. Int. Conf. Learn. Represent. (ICLR), 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.829, + 0.747 + ], + "angle": 0, + "content": "[23] Mixamo. https://www MIXamo.com/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[24] Meinard Müller, Tido Röder, and Michael Clausen. Efficient content-based retrieval of motion capture data. In Proc. ACM Conf. SIGGRAPH, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[25] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Kensuke Onuma, Christos Faloutsos, and Jessica K. Hodgins. Fmdistance: A fast and effective distance function for motion capture data. In Eurographics, 2008. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Qiaosong Qi, Le Zhuo, Aixi Zhang, Yue Liao, Fei Fang, Si Liu, and Shuicheng Yan. DiffDance: Cascaded human motion diffusion model for dance generation. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[28] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. SMCD: High realism motion style transfer via Mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[30] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3D dance generation by actor-critic GPT with choreographic memory. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[31] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3D dance GPT with choreographic memory. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[32] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Trans. Multimedia (TMM), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[34] Shuhei Tsuchida. Dance information processing: Computational approaches for assisting dance composition. New Gener. Comput., 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.582, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[36] Zilin Wang, Haolin Zhuang, Lu Li, Yinmin Zhang, Junjie Zhong, Jun Chen, Yu Yang, Boshi Tang, and Zhiyong Wu. Explore 3D dance generation via reward model from automatically-ranked demonstrations. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2024. 1, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[37] Simon Welker, Julius Richter, and Timo Gerkmann. Speech enhancement with score-based generative models in the complex STFT domain. arXiv preprint arXiv:2203.17004, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[38] Shuo Yan, Gangyi Ding, Zheng Guan, Ningxiao Sun, Hong-song Li, and Longfei Zhang. OutsideMe: Augmenting dancer's external self-image by using a mixed reality system. In Proc. Annu. ACM Conf. Extended Abstr. Hum. Factors Comput. Syst. (CHI), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[39] Ling-An Zeng, Guohong Huang, Gaojie Wu, and Wei-Shi Zheng. Light-t2m: A lightweight and fast model for text-to-motion generation. arXiv preprint arXiv:2412.11193, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Zeyu Zhang, Akide Liu, Qi Chen, Feng chen, Reid Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. InfiniMotion: Mamba boosts memory in Transformer for arbitrary long motion generation. arXiv preprint arXiv:2407.10061, 2024.3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion Mamba: Efficient and long sequence motion generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[42] Chenxi Zheng, Jing Qin, and Shengfeng He. Beat-It: Beatsynchronized multi-condition 3D dance generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[44] Haolin Zhuang, Shun Lei, Long Xiao, Weiqin Li, Liyang Chen, Sicheng Yang, Zhiyong Wu, Shiyan Kang, and Helen Meng. GTN-Bailando: Genre consistent long-term 3D dance generation based on pre-trained genre token network. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[45] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2Dance: DanceNet for music-driven dance generation. ACM Trans. Multimedia Comput. Commun. Appl. (TOMM), 2022. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13202" + } + ] +] \ No newline at end of file diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_origin.pdf b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b6b9936a6b33723d0508664e157f902fdacdac2e --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/bae0438d-2c16-45f8-9ffb-98e7eef3786c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a99116a56d0ad3a7ca0ff0a381b7c00248a4ea8b5748dc0ebf19a6bb9956db9 +size 4278810 diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/full.md b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1b375bde84ef2a1dcde3fde70eaffb7d4d2eced0 --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/full.md @@ -0,0 +1,306 @@ +# Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation + +Congyi Fan $^{1}$ , Jian Guan $^{1*}$ , Xuanjia Zhao $^{1}$ , Dongli Xu $^{2}$ , Youtian Lin $^{3}$ , Tong Ye $^{1}$ , Pengming Feng $^{4}$ , Haiwei Pan $^{1}$ + +$^{1}$ College of Computer Science and Technology, Harbin Engineering University, Harbin, China $^{2}$ Shanghai Academy of Artificial Intelligence for Science, Shanghai, China + +$^{3}$ School of Intelligence Science and Technology, Nanjing University, Nanjing, China $^{4}$ State Key Laboratory of Space-Ground Integrated Information Technology, Beijing, China + +# Abstract + +Automatically generating natural, diverse, and rhythmic human dance movements driven by music is vital for virtual reality and film industries. However, generating dance that naturally follows music remains a challenge, as existing methods lack proper beat alignment and exhibit unnatural motion dynamics. In this paper, we propose Danceba, a novel framework that leverages gating mechanism to enhance rhythm-aware feature representation for music-driven dance generation, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, we introduce Phase-Based Rhythm Extraction (PRE) to precisely extract rhythmic information from musical phase data, capitalizing on the intrinsic periodicity and temporal structures of music. Additionally, we propose Temporal-Gated Causal Attention (TGCA) to focus on global rhythmic features, ensuring that dance movements closely follow the musical rhythm. We also introduce Parallel Mamba Motion Modeling (PMMM) architecture to separately model upper and lower body motions along with musical features, thereby improving the naturalness and diversity of generated dance movements. Extensive experiments confirm that Danceba outperforms state-of-the-art methods, achieving improved rhythmic alignment and motion diversity. Project page: https://danceba.github.io/. + +# 1. Introduction + +Music-driven dance generation aims to automatically synthesize realistic and expressive 3D dance movements synchronized precisely with a given musical sequence, enabling natural animation of 3D character models. Due to + +its broad applications in gaming, virtual reality, film production, and interactive entertainment [14, 34, 38, 43], this task has attracted considerable attention. However, generating dances that accurately reflect musical rhythm while preserving natural and diverse movement dynamics remains a significant challenge, as it requires effectively capturing and utilizing complex temporal relationships between music and motion. + +Recently, generative models have demonstrated remarkable success across various tasks, such as text and image generation [9, 29]. Similarly, these generative approaches have also been applied to music-driven dance generation tasks [11, 19, 20, 27, 30, 31, 33, 36, 42, 44]. Among these, Bailando [30] introduces a cross-conditional causal attention mechanism to align musical cues with corresponding dance poses, significantly improving the quality of generated dance movements. However, its attention mechanism primarily captures local correspondence between music and poses, failing to sufficiently utilize global rhythmic characteristics within music, thus limiting the expressiveness and rhythmic coherence of the generated dances. + +Follow-up studies [11, 31, 36, 44] have extended Bailando's design to further improve dance generation performance. For instance, Enhancing-Bailando [11] employs a pretrained audio encoder (MERT) [21] to obtain richer musical feature representations. However, this encoder primarily captures semantic information from general audio data rather than explicitly modeling rhythmic characteristics, limiting its effectiveness in precise beat-dance synchronization. Moreover, Bailando and related methods typically concatenate three distinct input modalities, i.e., upper-body pose, lower-body pose, and musical features, and rely heavily on cross-conditional causal attention to align these inputs. Nevertheless, this attention mechanism primarily aligns pose and music at a local feature level, but struggles + +to capture continuous motion dynamics, limiting the naturalness and diversity of generated dance movements. + +To further tackle the rhythm alignment issue, Beat-It [42] explicitly introduces beat synchronization via a beat distance predictor. However, such an explicit synchronization approach imposes rigid control constraints and strong regularization during generation, significantly limiting the diversity and expressiveness of generated dances [27]. Thus, a more balanced approach capable of achieving precise rhythmic alignment without sacrificing motion naturalness and diversity is highly desirable. + +To this end, we propose a novel framework equipped with a gating-enhanced rhythm-aware feature representation for music-driven dance generation, i.e., Danceba, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, Danceba introduces Phase-Based Rhythm Extraction (PRE), a rhythm-aware feature extraction module that explicitly captures rhythmic characteristics from musical phase information. Leveraging the temporal phase information intrinsic to rhythm perception, PRE effectively separates rhythmic from semantic music features, significantly improving the rhythmic alignment between generated dance sequences and musical inputs. + +To further enhance beat-dance alignment using the rhythm-aware features extracted by PRE, it is essential to establish a stronger contextual association between musical rhythm and dance movements. To achieve this, we present a Temporal-Gated Causal Attention (TGCA) mechanism, enhanced by a gating strategy to explicitly prioritize and leverage global rhythmic structures within cross-conditional causal attention. By integrating gating mechanism, TGCA effectively strengthens the contextual connection between music rhythm and generated movements, ensuring that generated dance sequences accurately reflect the underlying musical rhythm and exhibit greater coherence and expressive diversity. + +Furthermore, to overcome the limitations of traditional attention-based approaches in motion modeling, we introduce Parallel Mamba Motion Modeling (PMMM), inspired by recent advances demonstrating that Mamba architectures achieve state-of-the-art performance in modeling complex human motion sequences [16, 39, 41]. Our parallel Mamba architecture separately models upper and lower-body dance sequences, with input features enhanced by Temporal-Gated Causal Attention (TGCA), which incorporates rhythm-aware representations alongside distinct upper and lower-body pose features. This design effectively captures diverse and nuanced motion characteristics. By leveraging Mamba's superior sequential modeling capability, our method significantly improves the expressiveness, diversity, and temporal coherence of generated dance movements. Fi + +nally, an additional TGCA module is applied after PMMM to further reinforce rhythmic alignment, ensuring that the generated dances precisely reflect the musical rhythm. + +Extensive experiments show that our method improves $\mathrm{FID}_k$ by $48.68\%$ , $\mathrm{Div}_k$ by $7.0\%$ , $\mathrm{Div}_g$ by $16.3\%$ , and BAS by $12.0\%$ over the second-best results, while maintaining a competitive $\mathrm{FID}_g$ of 11.90. This significantly surpasses state-of-the-art methods, producing dance movements that are more natural, diverse, and better synchronized with musical rhythm. + +# 2. Related Work + +# 2.1. Music-Driven Dance Generation + +Music-driven dance generation focuses on automatically synthesizing 3D dance movements from musical input, creating temporally coherent pose sequences for animating 3D character models. It has attracted increasing attention in both the computer vision and virtual avatar communities [1, 13] due to its applications in entertainment, gaming, and virtual reality. Early approaches, including [10, 12, 15, 17, 18, 32, 45], investigated diverse methodologies for this task. The study in [17] utilized a standard Transformer [35] with a late-fusion module to predict subsequent dance steps. The FACT framework [18] introduced the AIST++ dataset and employed a cross-modal Transformer with full attention mechanisms for music-to-dance generation. However, these initial methods neglect the rhythm features of music and struggle to effectively integrate musical and motion features, leading to suboptimal generation quality. + +Bailando [30] reformulated dance generation as a next-step pose prediction problem using cross-conditional causal attention, achieving high-quality music-driven dance synthesis. Subsequent works [11, 31, 44] extended this framework: Bailando++ [31] enhanced pattern recognition with musical context encoding, while Enhancing-Bailando [11] incorporated a pre-trained MERT audio encoder [21] for improved feature extraction. However, these methods often overlook rhythmic structures or rely on generic audio features. Beat-It [42] introduced explicit beat alignment via a conditional diffusion model [9, 25], ensuring precise synchronization. Yet, its reliance on rigid beat control signals and strong regularization may constrain pose diversity [27]. In contrast, we emphasize rhythm-aware feature representation, achieving superior rhythmic alignment without sacrificing movement diversity. + +# 2.2. Mamba in Human Motion Generation + +The integration of Mamba [7] and Mamba2 [2] into human motion generation represents a pivotal advancement, leveraging their efficient sequence modeling capabilities. Within human motion generation, Mamba surpasses con + +![](images/65a682da204142f85eba68eb6ffe1107f11afd402c702d7ee96ae4dd7207e8a0.jpg) +Figure 1. The overall framework of Danceba consists of three core modules: Phase-based Rhythm Extraction (PRE), Temporal-Gated Causal Attention (TGCA), and Parallel Mamba Motion Modeling (PMMM). PRE precisely extracts rhythm-aware features from musical phase information, providing accurate rhythmic signals. These rhythm-enhanced music features are fused with pose embeddings (upper-body and lower-body poses) and processed through the TGCA module, which utilizes gating mechanism to reinforce rhythmic sensitivity and align dance movements accurately with music beats. Parallel Mamba Motion Modeling separately models upper and lower body motion sequences, effectively capturing distinct dance dynamics aligned with rhythm-aware music features, thus significantly enhancing the naturalness, diversity, and temporal coherence of the generated dance movements. + +ventional Transformers in performance, as demonstrated by Motion Mamba [41], which highlights its efficacy for extended motion sequence synthesis. Furthermore, InfiniMotion [40] utilizes Mamba's hierarchical and bidirectional temporal processing to generate continuous, smooth motion sequences of indefinite length. Recent approaches including SMCD [28], Light-T2M [39], FTMoMamba [16], and MMD [4] have further incorporated Mamba architectures, advancing the field through enhanced efficiency and high-fidelity motion modeling. Building upon recent advancements, Mamba has achieved state-of-the-art performance in human motion modeling [16, 39, 41], motivating our selection of Mamba over self-attention mechanisms for information exchange within dance sequences, which significantly enhances the diversity of generated dance motions. + +# 3. Proposed Method + +This section presents our proposed Danceba framework for music-driven 3D dance generation. By leveraging gating-enhanced rhythm-aware features, Danceba ensures precise alignment between dance poses and music. The overall framework is illustrated in Figure 1, and its components are detailed as follows: Section 3.1 outlines the preliminary framework and key notations. Specifically, we introduce phase-based rhythm feature extraction (Section 3.2), global beat-dance alignment via temporal-gated causal attention (Section 3.3), and Mamba-based parallel motion modeling (Section 3.4). + +# 3.1. Preliminary + +Given a dance sequence $\mathbf{P} \in \mathbb{R}^{T \times (J \times 3)}$ , where $T$ is the time length and $J$ is joint amount. Using two pretrained Pose VQ-VAEs [30] $\mathcal{F}_{VAE}^u$ and $\mathcal{F}_{VAE}^l$ with codebook $\mathcal{Z}^u$ and $\mathcal{Z}^l$ for the upper and lower half bodies, it separately encodes upper and lower body movements into compositional code pairs $\mathbf{p} = [\mathbf{p}^u, \mathbf{p}^l]$ : + +$$ +\mathbf {p} ^ {u} = \mathcal {F} _ {V A E} ^ {u} (\mathbf {P}), \mathbf {p} ^ {l} = \mathcal {F} _ {V A E} ^ {l} (\mathbf {P}), \tag {1} +$$ + +where $\mathbf{p}^u\in (\mathcal{Z}^u)^{T'}$ $\mathbf{p}^l\in (\mathcal{Z}^l)^{T'}$ .Here $T^{\prime} = T / \lambda$ $\lambda$ is the temporal down-sampling rate. The input music feature $\mathbf{m}\in \mathbb{R}^{T'\times D_m}$ , where $D_{m}$ is the channel dimension of the input music features. Then we embed music features $\mathbf{m}$ upper $p^u$ and lower $p^l$ pose codes to learnable features with three separate linear layers: $\mathbf{X}_{\mathrm{m}} = \mathrm{Linear}(\mathbf{m}),\mathbf{X}_{\mathrm{u}} =$ Linear $(\mathbf{p}^{u})$ and $\mathbf{X}_1 = \mathrm{Linear}(\mathbf{p}^l)$ , where $\mathbf{X}_{\mathrm{m}}\in \mathbb{R}^{T^{\prime}\times D}$ $\mathbf{X}_{\mathrm{u}}\in \mathbb{R}^{T^{\prime}\times D}$ $\mathbf{X}_1\in \mathbb{R}^{T^{\prime}\times D}$ and D is the channel dimension of the features. + +The dance generation task is reframed as selecting the most probable future pose code from the codebook $\mathcal{Z}$ , conditioned on the music and prior movements. Since the upper and lower bodies are modeled separately, maintaining coherence and avoiding asynchrony (e.g., opposing directions) requires cross-conditioned prediction, leveraging mutual information between existing movements: + +$$ +\hat {p} _ {t} ^ {u} = \arg \max _ {k} \mathbb {P} \left(\mathbf {z} _ {k} ^ {u} \mid \mathbf {m} _ {1 \dots t}, p _ {0 \dots t - 1} ^ {u}, p _ {0 \dots t - 1} ^ {l}\right), \tag {2} +$$ + +$$ +\hat {p} _ {t} ^ {l} = \arg \max _ {k} \mathbb {P} \left(\mathbf {z} _ {k} ^ {l} \mid \mathbf {m} _ {1 \dots t}, p _ {0 \dots t - 1} ^ {u}, p _ {0 \dots t - 1} ^ {l}\right). \tag {3} +$$ + +At each time step $t$ , MotionGPT [30] estimates the probabilities of pose codes $\mathbf{z}_i \in \mathbb{Z}$ and selects the one with the most probable pose codes as the predicted upper pose $\hat{p}_t^u$ and predicted lower pose $\hat{p}_t^l$ . + +Specifically, the input feature $\mathbf{X}_i\in \mathbb{R}^{(3\times T')\times D}$ are concatenated by upper body motion feature, lower body motion feature and music feature as follows: $\mathbf{X}_i =$ $\mathrm{Concat}(\mathbf{X}_{\mathrm{m}},\mathbf{X}_{\mathrm{u}},\mathbf{X}_{\mathrm{l}})$ . Subsequently, $\mathbf{X}_i$ is passed through successive Autoregressive Transformer layers $\mathcal{F}_{AR}$ , followed by a linear transformation and a softmax layer, generates the probability distributions for predicting body pose collectively: + +$$ +\mathbf {a} ^ {h} = \operatorname {S o f t m a x} (\operatorname {L i n e a r} \left(\mathcal {F} _ {A R} \left(\mathbf {X} _ {i}\right)\right), \tag {4} +$$ + +where $\mathbf{a}^h = [\mathbf{a}^u,\mathbf{a}^l]$ , $\mathbf{a}^u$ denotes that of upper body and $\mathbf{a}^l$ denotes that of lower body. + +Finally, the model is optimized via supervised training with the cross-entropy loss on action probability $\mathbf{a}_t^h$ at each time step $t$ and for both upper body and lower body: + +$$ +\mathcal {L} _ {C E} = \frac {1}{T ^ {\prime}} \sum_ {t = 0} ^ {T ^ {\prime} - 1} \sum_ {h = u, l} \text {C r o s s E n t r o p y} \left(\mathrm {a} _ {t} ^ {h}, p _ {t + 1} ^ {h}\right). \tag {5} +$$ + +In inference, the MotionGPT predicts pose codes according to the initial pose code and the entire music, followed by Pose VQ-VAE decoders generating a new dance. + +# 3.2. Phase-Based Rhythm Feature Extraction + +To enhance rhythmic feature extraction in music-driven dance generation, we propose Phase-Based Rhythm Extraction (PRE), which decouples rhythm from musical semantics by leveraging the inherent periodicity and temporal positioning of phase representations. The architecture of PRE is shown in the green part of Figure 1. Unlike existing methods that entangle rhythm with melody and harmony, PRE enables independent and precise rhythm modeling as shown in Figure 2, forming a solid foundation for gating-enhanced rhythm-aware mechanism. + +Specifically, we adopt the Short-Time Fourier Transform (STFT) instead of the conventional Fourier Transform (FT), since the non-stationary nature of music and the sliding-window approach of STFT effectively captures sudden tempo shifts and complex rhythmic patterns [37]. Specifically, we first perform STFT on the input music feature to extract phase angles: + +$$ +\mathbf {S} _ {m} = \operatorname {S T F T} (\mathbf {m}), \tag {6} +$$ + +$$ +\boldsymbol {\varphi} = \operatorname {A n g l e} \left(\mathbf {S} _ {m}\right), +$$ + +where $\mathbf{S}_m$ represents the complex spectrogram obtained via STFT, $\mathrm{Angle}(\cdot)$ computes the phase angle of a complex-valued input and $\varphi^{\prime}$ denotes the phase angles extracted from $\mathbf{S}_m$ . By taking the angle of the STFT output, we obtain the + +![](images/4516cdb0664827a71e3752b6f55f139686de0bdcbbea09f1563fc9974db24f4f.jpg) +Figure 2. Visualizing the input music features and phase features. We notice that only a small portion of the input music features contains meaningful musical information, i.e., rhythm-poor features. And after a short-time Fourier transform (STFT), each corresponding token has a rich periodic Phase Feature, i.e., rhythm-rich features. + +music phase data $\varphi^{\prime}\in \mathbb{C}^{T_{\varphi}\times D_{\varphi}}$ , where $T_{\varphi}$ represents the number of time frames obtained after the STFT transformation, and $D_{\varphi}$ is the channel dimension of the $\varphi^{\prime}$ + +The extracted phase angles $\varphi^{\prime}$ inherently reflect temporal shifts and periodic structures of musical rhythms: low-frequency phase variations correspond to broader rhythmic patterns, while high-frequency variations capture finer rhythmic details. We also apply a center-cropping strategy, selectively focusing on the most informative phase regions: + +$$ +\varphi = \operatorname {C e n t e r C r o p} \left(\varphi^ {\prime}\right), \tag {7} +$$ + +where $\varphi \in \mathbb{C}^{T' \times D_{\varphi}}$ has the same temporal length $T'$ , addressing the potential temporal misalignment and preserving critical rhythmic cues. + +Subsequently, the phase features through a linear transformation, enhancing phase features representing musical rhythms. The process is shown in the following equation: + +$$ +\mathbf {X} _ {\varphi} = \operatorname {R e L U} (\mathrm {B N} (\operatorname {L i n e a r} (\varphi))), \tag {8} +$$ + +where the embedded phase features $\mathbf{X}_{\varphi} \in \mathbb{C}^{T' \times D}$ encode rich rhythmic information, providing a clearer and more structured representation of rhythmic and musical elements. Finally, we fuse the input feature $\mathbf{X}_i$ with the rhythmic features $\mathbf{X}_{\varphi}$ by concatenating them: $\mathbf{X}_{\gamma} = \mathrm{Concat}(\mathbf{X}_{\varphi}, \mathbf{X}_{\varphi}, \mathbf{X}_{\varphi})$ . Then, we sum the rhythm-enhanced features with the original music, upper body, and lower body motion features: $\mathbf{X} = \mathbf{X}_f + \mathbf{X}_{\gamma}$ , where $\mathbf{X}$ represents the rhythmically enhanced features. + +# 3.3. Global Beat Attention via Temporal Gating + +Leveraging these rhythm-rich features, we next investigate how to integrate them into the alignment process between music and dance. However, prior methods [11, 30, 31, 36, 44] relying on cross-conditional causal attention mechanism (denoted as $\mathrm{C}^3\mathrm{Attention}$ ) exhibit a critical misalignment issue. As shown in Figure 3 (a), we visualize the + +cross-conditional causal attention mechanism and observe that the next predicted token lacks direct control signals from historical tokens, failing to establish a stable global beat attention. This misalignment leads to error accumulation, which we identify as a key factor behind the unnatural dance generation in prior work. + +![](images/eaec8bec3d5f539fab42a024e18362868b0f88f6dc153345758089dcf623cdb9.jpg) +(a) Cross-Conditional Causal Attention + +![](images/964265c1bd018d27bf9512a68d8abf13ad6ad772252f81b68fc9c55ede5bf566.jpg) +(b) Temporal-Gated Causal Attention (TGCA) +Figure 3. Heat map of attention. The two displayed images show the results learned by Cross-Conditional Causal Attention (a) and Temporal-Gated Causal Attention (b) under the same piece of music. The brighter the color, the higher the attention; the darker the color, the lower the attention. TGCA the next token has a clear control signal, preventing overly random dance movements and achieving enhanced rhythm-aware feature representation. + +Building on this, we propose Temporal Gated Causal Attention (TGCA) to enhance global rhythm feature attention, improving music-dance alignment and ensuring the next token is guided by explicit historical information, preventing random action generation. As illustrated in the blue section of Figure 1, the rhythmically enhanced features $\mathbf{X}$ are passed through TGCA module, which can be expressed as the following: + +$$ +\operatorname {G a t i n g} (\mathbf {X}) = \operatorname {S i L U} (\operatorname {L i n e a r} (\mathbf {X})). \tag {9} +$$ + +The final output is obtained by the element-wise multiplication of cross-conditional causal attention $\mathrm{C}^3$ Attention and Gating, as shown in the following equation: + +$$ +\begin{array}{l} \mathbf {X} _ {a t t n} = \operatorname {T G C A} (\mathbf {X}) \\ = \mathrm {C} ^ {3} \text {A t t e n t i o n} (\mathbf {X}) \odot \operatorname {G a t i n g} (\mathbf {X}). \tag {10} \\ \end{array} +$$ + +As shown in Figure 3 (b), our proposed gated causal attention mechanism allows historical tokens to better predict the next token, implying an enhanced influence of global rhythmic patterns on dance generation. + +# 3.4. Mamba-Based Parallel Motion Modeling + +After TGCA enhances dance movements in $\mathbf{X}_{attn}$ with global rhythmic perception, we looking into dance modeling for generation. Dance generation faces two key challenges: (1) Dance is a complex spatial-temporal sequence encoding trajectory, posture, and velocity over time, requiring a model capable of capturing both local transitions and + +global rhythmic variations. (2) The upper and lower body exhibit distinct movement patterns and speeds, necessitating separate processing to maintain coherence and diversity. To address these issues, we propose a parallel Mamba motion modeling architecture that separately models music, upper-body, and lower-body movements while leveraging state space modeling to ensure fluid and diverse motion generation. + +Specifically, to ensure each action is primarily influenced by relevant preceding movements, we introduce a GateMlp in the Mamba block, as shown in Figure 1. This mechanism selectively retains essential motion features while filtering out less relevant information, thereby enhancing the coherence and expressiveness of the generated dance sequences. The process is formulated as follows: + +$$ +\mathbf {X} _ {m b} ^ {\prime} = \operatorname {M a m b a} (\operatorname {R M S N o r m} (\mathbf {X} _ {a t t n})) + \mathbf {X} _ {a t t n}, \tag {11} +$$ + +$$ +\mathbf {X} _ {m b} = \operatorname {G a t e M l p} \left(\operatorname {R M S N o r m} \left(\mathbf {X} _ {m b} ^ {\prime}\right)\right) + \mathbf {X} _ {m b} ^ {\prime}. \tag {12} +$$ + +The gating mechanism dynamically regulates the frequency, speed, and amplitude of movements in response to musical beats, ensuring that dance sequences remain synchronized with the music. Finally, three parallel Mamba architectures model upper- and lower-body dance sequences separately, therefore ensure the generated dance movements are expressive, diverse, and temporally coherent. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset: We follow prior works [27, 30, 36] and perform all experiments on the AIST++ dataset [18], the most widely used benchmark for music-driven dance generation that avoids observable biases (e.g., offensive postures). AIST++ consists of 991 high-quality 3D pose sequences recorded at 60 FPS in skinned multi-person linear (SMPL) format [22], with 951 sequences designated for training and 40 reserved for evaluation. Following [27, 30, 36], we generate 40 pieces of dance sequences on the test set of AIST++ dataset, and sample the generated dance sequence with length of 20 seconds for performance evaluation. + +Evaluation Metrics: We follow previous studies [18, 30] to quantitatively evaluate the generated samples in three key aspects: quality, diversity, and beat-dance alignment. Prior to evaluation, it is necessary to extract kinetic features [26] and geometric features [24] from both the ground truth and the generated samples via a specific toolbox [6]. For dance quality, we compute two types of the Fréchet Inception Distance (FID) [8], including $\mathrm{FID}_k$ based on kinetic features and $\mathrm{FID}_g$ based on geometric features. Lower $\mathrm{FID}_k$ indicates that the generated dance has more similar kinetic features to the original dance, meaning the generated dances are more natural and smooth. Lower $\mathrm{FID}_g$ indicates that + +Table 1. Comparison with state-of-the-art methods on the AIST++ dataset. Underlining indicates the best performance among existing methods. Blue indicates results that surpass the best existing method. ↓ indicates that lower values are better, while ↑ indicates that higher values are better. * represents the dances generated by "Li et al." that exhibit significant jitteriness, resulting in extremely high velocity variation, as also noted in [18]. † indicates the reproduced results obtained using Bailando++ official code and published checkpoints. ‡ denotes the results that are averaged over five independent training runs. + +
METHODVENUEMOTION QUALITYMOTION DIVERSITYBEAT ALIGN SCORE↑
FIDk↓FIDg↓DIVk↑DIVg↑
GROUND TRUTH-17.1010.608.197.450.2374
Li et al. [17]ARXIV 202086.4343.46*6.853.320.1607
DANCERUVOLUTION [12]ICLR 202173.4225.923.524.870.1950
DANCENET [45]TOMM 202269.1825.492.862.850.1430
FACT [18]ICCV 202135.3522.115.946.180.2209
BAILANDO [30]CVPR 202228.169.627.836.340.2332
EDGE [33]CVPR 202342.1622.123.964.610.2334
DIFFDANCE [27]ACMMM 202324.0920.686.022.890.2418
†BAILANDO++ [31]TPAMI 202322.7411.587.946.340.2263
E3D2 [36]AAAI 202426.258.947.966.490.2232
LODGGE [20]CVPR 202437.0918.795.584.850.2423
OURS (BEST)11.67(-11.07)11.908.52(+0.58)7.55(+1.21)0.2714(+2.91%)
‡OURS (AVERAGE)15.48(-7.26)13.577.857.49(+1.15)0.2779(+3.56%)
+ +![](images/671396d25209960c5a5cb3ed3dce7add1dd04ea85c7e68ad8cb7e095eabf6218.jpg) +Figure 4. Dance examples generated by our proposed method on various types of music. The demo videos can be found in the supplementary material. The character is sourced from Mixamo [23]. + +the generated dance has more similar geometric features, reflecting the quality of choreography. For dance diversity, we employ metrics $\mathrm{Div}_k$ and $\mathrm{Div}_g$ [18], which denote the distance of kinetic and distance of geometric features, respectively. Higher $\mathrm{Div}_k$ and $\mathrm{Div}_g$ signify the greater amplitude and diversity of motion, respectively. Finally, to evaluate beat-dance alignment, we use the Beat Align Score (BAS) [18], which quantifies the average distance between a music beat and its nearest dance beat. Higher BAS indicates greater degree of beat-dance alignment. + +# 4.2. Comparison with State-of-the-Art Methods + +To validate the superiority of our method, we conduct comparative experiments with the state-of-the-art methods [12, 17, 18, 20, 27, 30, 31, 33, 36, 45], as well as the ground truth. Following [27, 30, 36], we generate 40 pieces of dance sequences on the AIST++ test set, and sample the generated dance sequence with the length of 20 seconds to compute the metrics mentioned above. Among the compared methods, "Li et al." [17], + +DanceNet [45] and FACT [18] use the results from the AIST++ benchmark [18], while the results of DanceRevolution [12] are derived from its reproduction as presented in Bailando [30]. EDGE [33] utilizes the reproduced results from the Lodge [20]. The Bailando, DiffDance [27] and Lodge methods use the experimental results from their respective publications. + +The comparative results are presented in Table 1. Our method achieves state-of-the-art performance, significantly outperforming existing methods in terms of $\mathrm{FID}_k$ , $\mathrm{Div}_g$ , $\mathrm{Div}_k$ , and Beat Align Score (BAS), achieving the best results. Additionally, our method ranks third on the $\mathrm{FID}_g$ metric. The generated dance examples of our proposed Danceba on various types of music are shown in Figure 4, and the demo videos are provided in the supplementary files. + +Motion Quality Metrics: As illustrated in Table 1, our method outperforms existing methods in overall quality performance. To be specific, our method improves on $\mathrm{FID}_k$ by $48.68\%$ (11.67 vs. 22.74), while exhibits competitive edge of $\mathrm{FID}_g$ . Further analysis of the employed quality metrics reveals distinct roles for the two metrics types. The metric $\mathrm{FID}_k$ , defined based on velocity and energy, primarily captures the kinetic characteristics of the generated dance, while $\mathrm{FID}_g$ , derived from multiple manually designed motion templates, emphasizes the geometric structural properties of the dance. Thus, our method demonstrates superior performance in modeling the kinetic aspects of sequential motions, as evidenced by its exceptional $\mathrm{FID}_k$ scores. Although it exhibits a modest gap in $\mathrm{FID}_g$ (11.90 vs. 11.58) compared to the state-of-the-art Bailando++ [31], it achieves a $19.0\%$ higher $\mathrm{Div}_g$ (7.55 vs. 6.34), indicating improved motion diversity with comparable generative quality. Despite this minor trade-off, our method outperforms existing approaches overall, achieving a robust balance of efficiency and quality across a broader range of metrics. + +Motion Diversity Metrics: Regarding diversity evaluation, Danceba achieves the best diversity performance, which improves $\mathrm{Div}_k$ by $7.0\%$ (8.52 vs. 7.96) and even slightly outperforms the Ground Truth (8.52 vs. 8.19). Meanwhile, our method improves $\mathrm{Div}_g$ by $16.3\%$ (7.55 vs. 6.49) and also slightly surpasses the Ground Truth (7.55 vs. 7.45), as illustrated in Table 1. For a visual comparison, we also present a comparison with the state-of-the-art method Bailando, as illustrated in Figure 5. It can be observed that Bailando exhibits relatively small movement amplitudes, with some dance movements missing and the actions appearing stiff. In contrast, our method shows larger movement amplitudes and movement diversity. + +Beat-dance Alignment Metrics: Danceba achieves the best performance in beat-dance alignment, as quantified by the BAS metric. Specifically, it improves BAS by $12.0\%$ (0.2714 vs. 0.2423) compared to the previous best re + +![](images/3e51635a02850c5a5e6e05580b01b05b7d4c0f246b64a550d94bf06e9342243d.jpg) +Figure 5. Comparison with the state-of-the-art method Bailando [30]. Visual comparisons with Bailando can be found in the supplementary video. + +sult and surpasses the ground truth by $14.3\%$ (0.2714 vs. 0.2374), as shown in Table 1. Figure 6 further visualizes the alignment between music beats and motion beats of the generated dance compared to Bailando++ [31]. The results show that Danceba achieves a significantly smaller beat distance, demonstrating its ability to generate dance sequences that more accurately follow musical rhythms. + +![](images/ddfe8af1c30e219ea492b3880cb58289e2af562b3847287755c8ed37fd4fdaec.jpg) +Figure 6. Beats alignment visualization, where the horizontal axis shows frame indices of beat events. Comparing Danceba with Balando++ [31], we can find that the distance between motion beats and music beats generated by our method is smaller. This indicating that Danceba performs better in terms of rhythmic alignment. + +![](images/424c83aebeab355a4a9655b97400f17d55a07a4764b95ef6bd8ee43b7ab8b947.jpg) + +# 4.3. Ablation Studies + +We conduct ablation studies to evaluate the contributions of the key modules (i.e., PRE, TGCA, and PMMM) in Danceba, where "w/o" denotes the removal of a specific module. "w/o PMMM" refers to the setting where PMMM is replaced with TGCA layers to ensure fairness. The results are presented in Table 2. Additionally, we assess our method using a single Mamba structure instead of the parallel Mamba structure for motion modeling, referred to as Danceba-Single, with results shown in Table 3. + +Note that, all ablation results are averaged over five independent training runs to ensure the credibility and robustness of our results. The motion quality score (FID) is computed as the average of $\mathrm{FID}_k$ and $\mathrm{FID}_g$ , while the motion diversity score (Div) is the average of $\mathrm{Div}_k$ and $\mathrm{Div}_g$ . + +Phase-Based Rhythm Extraction: To validate the effec + +Table 2. Ablation study of three key modules (i.e., PRE, TGCA, and PMMM), and "w/o" indicates Danceba without specified module. Visual comparisons with Bailando can be found in the supplementary video. + +
METHODFID ↓DIV ↑BAS ↑
GROUND TRUTH13.857.820.2374
w/o TGCA33.226.74-
w/o PMMM25.245.78-
w/o PRE20.626.530.2474
DANCEBA14.537.670.2779
+ +tiveness of PRE, we replace the rhythm feature extraction module with a learnable linear layer as used in previous works [11, 30, 31, 36, 44]. Consequently, the motion quality (i.e., FID), motion diversity (i.e., Div), and beat alignment score (i.e., BAS) decrease by 6.09 (41.91%), 1.14 (14.86%), and 0.0305 (10.96%), respectively, as illustrated in Table 2. Specifically, this performance drop highlights the contribution of PPE, which effectively captures the critical rhythm features of music, a crucial aspect that previous methods have largely overlooked. + +Temporal-Gated Causal Attention: As shown in Table 2, we validate the effectiveness of the proposed TGCA by training a Danceba variant that replaces TGCA with conventional causal cross-conditional attention. The results demonstrate that the full Danceba model outperforms this variant in both FID and Div, highlighting TGCA's ability to effectively leverage global rhythmic features extracted by PRE. This significantly enhances the contextual association between music and dance movements, resulting in dance sequences that not only accurately reflect musical rhythm but also exhibit greater coherence and expressive diversity. + +Parallel Mamba Motion Modeling: To evaluate the effectiveness of Parallel Mamba Motion Modeling, we train a model without this module. As shown in Table 2, Danceba outperforms its variant without using PMMM in terms of both FID and Div, demonstrating the importance of parallel Mamba modeling. These results validate Mamba's superior sequential modeling capability, which is particularly well-suited for dance generation as it requires both fine-grained inter-modal interactions and long-sequence modeling. + +To further assess the advantage of using a parallel Mamba architecture instead of a single Mamba architecture for motion modeling, we compare Danceba with Danceba-Single. Unlike Danceba-Single, which treats the body as a unified entity and disregards the distinction between upper and lower-body movements, PMMM explicitly models their distinct motion dynamics. Additionally, PMMM integrates rhythm-aware representation alongside these distinct pose features, enhancing beat-dance alignment with the musical rhythm. As shown in Table 3, Danceba significantly improves expressiveness and temporal coherence, with $\mathrm{FID}_k$ + +Table 3. Ablation study on parallel Mamba motion modeling, where Danceba-Single denotes our method using single Mamba architecture. + +
METHODFIDk ↓FIDg ↑
GROUND TRUTH17.1010.60
DANCEBA-SINGLE82.5083.97
DANCEBA15.48(-67.02)13.57(-60.40)
+ +and $\mathrm{FID}_g$ scores improving by 67.02 and 60.40, respectively. + +# 5. Limitations + +Although Danceba achieves superior dance generation quality compared to state-of-the-art methods, it has two key limitations: (1) Music feature encoding: Danceba uses a simple linear layer instead of a pre-trained audio model (e.g., Jukebox [3], MERT [21], or CLAP [5]). While computationally efficient, this design may fail to capture nuanced musical structures, affecting rhythm-motion synchronization. Future work could explore pre-trained music encoders to enhance feature extraction. (2) 3D pose quantization: Danceba adopts the Pose VQ-VAE framework from Bailando [30] for fair comparison with prior SOTA methods [30, 31, 36]. However, it has a limited encoding space, which may limit motion diversity. Such a trade-off between quantization efficiency and expressiveness could hinder the preservation of fine-grained motion details. Exploring hierarchical quantization or adaptive codebook learning could further improve motion quality. + +# 6. Conclusion + +This paper has presented a novel framework Danceba for music-driven dance generation that enhances rhythm-aware feature representation and motion modeling. By introducing phase-based rhythm extraction and temporal-gated causal attention, our approach ensures precise beat-dance synchronization and effectively integrates rhythmic structures into dance movements. Additionally, parallel Mamba motion modeling separately captures the distinct motion dynamics of the upper and lower body while incorporating rhythm-aware representations, improving the naturalness and diversity of the generated dance. Extensive experiments demonstrate that Danceba significantly outperforms state-of-the-art methods, generating dance movements that are more rhythmic, natural, and diverse while maintaining strong alignment with musical beats. Future work could integrate pre-trained audio encoders for richer music representation, explore advanced quantization to enhance motion expressiveness and diversity, or extend to style-conditioned motion synthesis for broader applicability. + +# Acknowledgments + +This work was partly supported by the Fundamental Research Funds for the Central Universities (Grant No. 3072025YY0601), and the Beijing Nova Program (Grant No. 20230484261). + +# References + +[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Trans. Graph. (TOG), 2023. 2 +[2] Tri Dao and Albert Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. In Proc. Int. Conf. Mach. Learn. (ICML), 2024. 2 +[3] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020.8 +[4] Kun Dong, Jian Xue, Zehai Niu, Xing Lan, Ke Lu, Qingyuan Liu, and Xiaoyu Qin. Realistic full-body motion generation from sparse tracking with state space model. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2024. 3 +[5] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP learning audio concepts from natural language supervision. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 8 +[6] Deepak Gopinath and Jungdam Won. Fairmotion - tools to load, process and visualize motion capture data. *GibHub*, 2020. 5 +[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2 +[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 5 +[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020. 1, 2 +[10] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph. (TOG), 2016. 2 +[11] Qiaochu Huang, Xu He, Boshi Tang, Haolin Zhuang, Liyang Chen, Shuochen Gao, Zhiyong Wu, Haozhi Huang, and Helen Meng. Enhancing expressiveness in dance generation via integrating frequency and music style information. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2024. 1, 2, 4, 8 +[12] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In Proc. Int. Conf. Learn. Represent. (ICLR), 2021. 2, 6, 7 + +[13] Jinwoo Kim, Heeseok Oh, Seongjean Kim, Hoseok Tong, and Sanghoon Lee. A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 2 +[14] Matthew Kyan, Guoyu Sun, Haiyan Li, Ling Zhong, Paisarn Muneesawang, Nan Dong, Bruce Elder, and Ling Guan. An approach to ballet dance training through MS Kinect and visualization in a CAVE virtual reality environment. ACM Trans. Intell. Syst. Technol. (TIST), 2015. 1 +[15] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2019. 2 +[16] Chengjian Li, Xiangbo Shu, Qiongjie Cui, Yazhou Yao, and Jinhui Tang. FTMoMamba: Motion generation with frequency and text state space models. arXiv preprint arXiv:2411.17532, 2024. 2, 3 +[17] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with Transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6 +[18] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3D dance generation with AIST++. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021. 2, 5, 6, 7 +[19] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3D full body dance generation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1 +[20] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024. 1, 6, 7 +[21] Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Chenghao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, et al. MERT: Acoustic music understanding model with large-scale self-supervised training. In Proc. Int. Conf. Learn. Represent. (ICLR), 2024. 1, 2, 8 +[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 2015. 5 +[23] Mixamo. https://www MIXamo.com/.6 +[24] Meinard Müller, Tido Röder, and Michael Clausen. Efficient content-based retrieval of motion capture data. In Proc. ACM Conf. SIGGRAPH, 2005. 5 +[25] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[26] Kensuke Onuma, Christos Faloutsos, and Jessica K. Hodgins. Fmdistance: A fast and effective distance function for motion capture data. In Eurographics, 2008. 5 + +[27] Qiaosong Qi, Le Zhuo, Aixi Zhang, Yue Liao, Fei Fang, Si Liu, and Shuicheng Yan. DiffDance: Cascaded human motion diffusion model for dance generation. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2023. 1, 2, 5, 6, 7 +[28] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. SMCD: High realism motion style transfer via Mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 3 +[29] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1 +[30] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3D dance generation by actor-critic GPT with choreographic memory. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[31] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3D dance GPT with choreographic memory. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1, 2, 4, 6, 7, 8 +[32] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Trans. Multimedia (TMM), 2020. 2 +[33] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1, 6, 7 +[34] Shuhei Tsuchida. Dance information processing: Computational approaches for assisting dance composition. New Gener. Comput., 2024. 1 +[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 2 +[36] Zilin Wang, Haolin Zhuang, Lu Li, Yinmin Zhang, Junjie Zhong, Jun Chen, Yu Yang, Boshi Tang, and Zhiyong Wu. Explore 3D dance generation via reward model from automatically-ranked demonstrations. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2024. 1, 4, 5, 6, 8 +[37] Simon Welker, Julius Richter, and Timo Gerkmann. Speech enhancement with score-based generative models in the complex STFT domain. arXiv preprint arXiv:2203.17004, 2022. 4 +[38] Shuo Yan, Gangyi Ding, Zheng Guan, Ningxiao Sun, Hong-song Li, and Longfei Zhang. OutsideMe: Augmenting dancer's external self-image by using a mixed reality system. In Proc. Annu. ACM Conf. Extended Abstr. Hum. Factors Comput. Syst. (CHI), 2015. 1 +[39] Ling-An Zeng, Guohong Huang, Gaojie Wu, and Wei-Shi Zheng. Light-t2m: A lightweight and fast model for text-to-motion generation. arXiv preprint arXiv:2412.11193, 2024. 2, 3 +[40] Zeyu Zhang, Akide Liu, Qi Chen, Feng chen, Reid Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. InfiniMotion: Mamba boosts memory in Transformer for arbitrary long motion generation. arXiv preprint arXiv:2407.10061, 2024.3 + +[41] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion Mamba: Efficient and long sequence motion generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2025. 2, 3 +[42] Chenxi Zheng, Jing Qin, and Shengfeng He. Beat-It: Beatsynchronized multi-condition 3D dance generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2 +[43] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1 +[44] Haolin Zhuang, Shun Lei, Long Xiao, Weiqin Li, Liyang Chen, Sicheng Yang, Zhiyong Wu, Shiyan Kang, and Helen Meng. GTN-Bailando: Genre consistent long-term 3D dance generation based on pre-trained genre token network. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 1, 2, 4, 8 +[45] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2Dance: DanceNet for music-driven dance generation. ACM Trans. Multimedia Comput. Commun. Appl. (TOMM), 2022. 2, 6, 7 \ No newline at end of file diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/images.zip b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c0328fd3dbbb6b1ff8dbcb233ad680649e22ce6e --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ca9f8e9af5cd2e3e49aeb2a6a3074244fb3f7507bc4f171dd3f4f085d457dd5 +size 532635 diff --git a/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/layout.json b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a9a08c75d12d6a7c728b383fb19233f026042a76 --- /dev/null +++ b/2025/Align Your Rhythm_ Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation/layout.json @@ -0,0 +1,8367 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 140 + ], + "type": "text", + "content": "Align Your Rhythm: Generating Highly Aligned Dance Poses with Gating-Enhanced Rhythm-Aware Feature Representation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "spans": [ + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": "Congyi Fan" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Jian Guan" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Xuanjia Zhao" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Dongli Xu" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Youtian Lin" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Tong Ye" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Pengming Feng" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "text", + "content": ", Haiwei Pan" + }, + { + "bbox": [ + 165, + 160, + 445, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "type": "text", + "content": "College of Computer Science and Technology, Harbin Engineering University, Harbin, China " + }, + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 202, + 533, + 231 + ], + "type": "text", + "content": "Shanghai Academy of Artificial Intelligence for Science, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "spans": [ + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "type": "text", + "content": "School of Intelligence Science and Technology, Nanjing University, Nanjing, China " + }, + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 84, + 231, + 526, + 259 + ], + "type": "text", + "content": "State Key Laboratory of Space-Ground Integrated Information Technology, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 312, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 297, + 601 + ], + "type": "text", + "content": "Automatically generating natural, diverse, and rhythmic human dance movements driven by music is vital for virtual reality and film industries. However, generating dance that naturally follows music remains a challenge, as existing methods lack proper beat alignment and exhibit unnatural motion dynamics. In this paper, we propose Danceba, a novel framework that leverages gating mechanism to enhance rhythm-aware feature representation for music-driven dance generation, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, we introduce Phase-Based Rhythm Extraction (PRE) to precisely extract rhythmic information from musical phase data, capitalizing on the intrinsic periodicity and temporal structures of music. Additionally, we propose Temporal-Gated Causal Attention (TGCA) to focus on global rhythmic features, ensuring that dance movements closely follow the musical rhythm. We also introduce Parallel Mamba Motion Modeling (PMMM) architecture to separately model upper and lower body motions along with musical features, thereby improving the naturalness and diversity of generated dance movements. Extensive experiments confirm that Danceba outperforms state-of-the-art methods, achieving improved rhythmic alignment and motion diversity. Project page: https://danceba.github.io/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 135, + 636 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 296, + 693 + ], + "type": "text", + "content": "Music-driven dance generation aims to automatically synthesize realistic and expressive 3D dance movements synchronized precisely with a given musical sequence, enabling natural animation of 3D character models. Due to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 287, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 382 + ], + "type": "text", + "content": "its broad applications in gaming, virtual reality, film production, and interactive entertainment [14, 34, 38, 43], this task has attracted considerable attention. However, generating dances that accurately reflect musical rhythm while preserving natural and diverse movement dynamics remains a significant challenge, as it requires effectively capturing and utilizing complex temporal relationships between music and motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 542 + ], + "type": "text", + "content": "Recently, generative models have demonstrated remarkable success across various tasks, such as text and image generation [9, 29]. Similarly, these generative approaches have also been applied to music-driven dance generation tasks [11, 19, 20, 27, 30, 31, 33, 36, 42, 44]. Among these, Bailando [30] introduces a cross-conditional causal attention mechanism to align musical cues with corresponding dance poses, significantly improving the quality of generated dance movements. However, its attention mechanism primarily captures local correspondence between music and poses, failing to sufficiently utilize global rhythmic characteristics within music, thus limiting the expressiveness and rhythmic coherence of the generated dances." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Follow-up studies [11, 31, 36, 44] have extended Bailando's design to further improve dance generation performance. For instance, Enhancing-Bailando [11] employs a pretrained audio encoder (MERT) [21] to obtain richer musical feature representations. However, this encoder primarily captures semantic information from general audio data rather than explicitly modeling rhythmic characteristics, limiting its effectiveness in precise beat-dance synchronization. Moreover, Bailando and related methods typically concatenate three distinct input modalities, i.e., upper-body pose, lower-body pose, and musical features, and rely heavily on cross-conditional causal attention to align these inputs. Nevertheless, this attention mechanism primarily aligns pose and music at a local feature level, but struggles" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "text", + "content": "*Corresponding author: j.guan@hrbeu.edu.cn" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13193" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "to capture continuous motion dynamics, limiting the naturalness and diversity of generated dance movements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 294, + 204 + ], + "type": "text", + "content": "To further tackle the rhythm alignment issue, Beat-It [42] explicitly introduces beat synchronization via a beat distance predictor. However, such an explicit synchronization approach imposes rigid control constraints and strong regularization during generation, significantly limiting the diversity and expressiveness of generated dances [27]. Thus, a more balanced approach capable of achieving precise rhythmic alignment without sacrificing motion naturalness and diversity is highly desirable." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 348 + ], + "type": "text", + "content": "To this end, we propose a novel framework equipped with a gating-enhanced rhythm-aware feature representation for music-driven dance generation, i.e., Danceba, which achieves highly aligned dance poses with enhanced rhythmic sensitivity. Specifically, Danceba introduces Phase-Based Rhythm Extraction (PRE), a rhythm-aware feature extraction module that explicitly captures rhythmic characteristics from musical phase information. Leveraging the temporal phase information intrinsic to rhythm perception, PRE effectively separates rhythmic from semantic music features, significantly improving the rhythmic alignment between generated dance sequences and musical inputs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 504 + ], + "type": "text", + "content": "To further enhance beat-dance alignment using the rhythm-aware features extracted by PRE, it is essential to establish a stronger contextual association between musical rhythm and dance movements. To achieve this, we present a Temporal-Gated Causal Attention (TGCA) mechanism, enhanced by a gating strategy to explicitly prioritize and leverage global rhythmic structures within cross-conditional causal attention. By integrating gating mechanism, TGCA effectively strengthens the contextual connection between music rhythm and generated movements, ensuring that generated dance sequences accurately reflect the underlying musical rhythm and exhibit greater coherence and expressive diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 505, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 684 + ], + "type": "text", + "content": "Furthermore, to overcome the limitations of traditional attention-based approaches in motion modeling, we introduce Parallel Mamba Motion Modeling (PMMM), inspired by recent advances demonstrating that Mamba architectures achieve state-of-the-art performance in modeling complex human motion sequences [16, 39, 41]. Our parallel Mamba architecture separately models upper and lower-body dance sequences, with input features enhanced by Temporal-Gated Causal Attention (TGCA), which incorporates rhythm-aware representations alongside distinct upper and lower-body pose features. This design effectively captures diverse and nuanced motion characteristics. By leveraging Mamba's superior sequential modeling capability, our method significantly improves the expressiveness, diversity, and temporal coherence of generated dance movements. Fi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "nally, an additional TGCA module is applied after PMMM to further reinforce rhythmic alignment, ensuring that the generated dances precisely reflect the musical rhythm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": "Extensive experiments show that our method improves " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "48.68\\%" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": ", and BAS by " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": " over the second-best results, while maintaining a competitive " + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 313, + 109, + 555, + 192 + ], + "type": "text", + "content": " of 11.90. This significantly surpasses state-of-the-art methods, producing dance movements that are more natural, diverse, and better synchronized with musical rhythm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 400, + 217 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 225, + 488, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 488, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 488, + 237 + ], + "type": "text", + "content": "2.1. Music-Driven Dance Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 243, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 447 + ], + "type": "text", + "content": "Music-driven dance generation focuses on automatically synthesizing 3D dance movements from musical input, creating temporally coherent pose sequences for animating 3D character models. It has attracted increasing attention in both the computer vision and virtual avatar communities [1, 13] due to its applications in entertainment, gaming, and virtual reality. Early approaches, including [10, 12, 15, 17, 18, 32, 45], investigated diverse methodologies for this task. The study in [17] utilized a standard Transformer [35] with a late-fusion module to predict subsequent dance steps. The FACT framework [18] introduced the AIST++ dataset and employed a cross-modal Transformer with full attention mechanisms for music-to-dance generation. However, these initial methods neglect the rhythm features of music and struggle to effectively integrate musical and motion features, leading to suboptimal generation quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 447, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 639 + ], + "type": "text", + "content": "Bailando [30] reformulated dance generation as a next-step pose prediction problem using cross-conditional causal attention, achieving high-quality music-driven dance synthesis. Subsequent works [11, 31, 44] extended this framework: Bailando++ [31] enhanced pattern recognition with musical context encoding, while Enhancing-Bailando [11] incorporated a pre-trained MERT audio encoder [21] for improved feature extraction. However, these methods often overlook rhythmic structures or rely on generic audio features. Beat-It [42] introduced explicit beat alignment via a conditional diffusion model [9, 25], ensuring precise synchronization. Yet, its reliance on rigid beat control signals and strong regularization may constrain pose diversity [27]. In contrast, we emphasize rhythm-aware feature representation, achieving superior rhythmic alignment without sacrificing movement diversity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 516, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 516, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 516, + 659 + ], + "type": "text", + "content": "2.2. Mamba in Human Motion Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "The integration of Mamba [7] and Mamba2 [2] into human motion generation represents a pivotal advancement, leveraging their efficient sequence modeling capabilities. Within human motion generation, Mamba surpasses con" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Note that, the terms \"beat\" and \"rhythm\" are used interchangeably in music processing following DiffDance [27] and Bailando++ [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13194" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 556, + 268 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 556, + 268 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 556, + 268 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 556, + 268 + ], + "type": "image", + "image_path": "65a682da204142f85eba68eb6ffe1107f11afd402c702d7ee96ae4dd7207e8a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 354 + ], + "type": "text", + "content": "Figure 1. The overall framework of Danceba consists of three core modules: Phase-based Rhythm Extraction (PRE), Temporal-Gated Causal Attention (TGCA), and Parallel Mamba Motion Modeling (PMMM). PRE precisely extracts rhythm-aware features from musical phase information, providing accurate rhythmic signals. These rhythm-enhanced music features are fused with pose embeddings (upper-body and lower-body poses) and processed through the TGCA module, which utilizes gating mechanism to reinforce rhythmic sensitivity and align dance movements accurately with music beats. Parallel Mamba Motion Modeling separately models upper and lower body motion sequences, effectively capturing distinct dance dynamics aligned with rhythm-aware music features, thus significantly enhancing the naturalness, diversity, and temporal coherence of the generated dance movements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 361, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 297, + 542 + ], + "type": "text", + "content": "ventional Transformers in performance, as demonstrated by Motion Mamba [41], which highlights its efficacy for extended motion sequence synthesis. Furthermore, InfiniMotion [40] utilizes Mamba's hierarchical and bidirectional temporal processing to generate continuous, smooth motion sequences of indefinite length. Recent approaches including SMCD [28], Light-T2M [39], FTMoMamba [16], and MMD [4] have further incorporated Mamba architectures, advancing the field through enhanced efficiency and high-fidelity motion modeling. Building upon recent advancements, Mamba has achieved state-of-the-art performance in human motion modeling [16, 39, 41], motivating our selection of Mamba over self-attention mechanisms for information exchange within dance sequences, which significantly enhances the diversity of generated dance motions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 559, + 162, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 162, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 162, + 572 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "This section presents our proposed Danceba framework for music-driven 3D dance generation. By leveraging gating-enhanced rhythm-aware features, Danceba ensures precise alignment between dance poses and music. The overall framework is illustrated in Figure 1, and its components are detailed as follows: Section 3.1 outlines the preliminary framework and key notations. Specifically, we introduce phase-based rhythm feature extraction (Section 3.2), global beat-dance alignment via temporal-gated causal attention (Section 3.3), and Mamba-based parallel motion modeling (Section 3.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 361, + 394, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 394, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 394, + 373 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": "Given a dance sequence " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{T \\times (J \\times 3)}" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " is the time length and " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " is joint amount. Using two pretrained Pose VQ-VAEs [30] " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{VAE}^u" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{VAE}^l" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " with codebook " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^u" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^l" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": " for the upper and lower half bodies, it separately encodes upper and lower body movements into compositional code pairs " + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [\\mathbf{p}^u, \\mathbf{p}^l]" + }, + { + "bbox": [ + 313, + 377, + 555, + 452 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 362, + 455, + 555, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 455, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 362, + 455, + 555, + 471 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {u} = \\mathcal {F} _ {V A E} ^ {u} (\\mathbf {P}), \\mathbf {p} ^ {l} = \\mathcal {F} _ {V A E} ^ {l} (\\mathbf {P}), \\tag {1}", + "image_path": "347b447094fe5cf214ed431b2d559c9e6f07eded4bc88fc529033cc4ddbb9c78.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^u\\in (\\mathcal{Z}^u)^{T'}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{p}^l\\in (\\mathcal{Z}^l)^{T'}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " .Here " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "T^{\\prime} = T / \\lambda" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " is the temporal down-sampling rate. The input music feature " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{m}\\in \\mathbb{R}^{T'\\times D_m}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " is the channel dimension of the input music features. Then we embed music features " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " upper " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "p^u" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " and lower " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "p^l" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " pose codes to learnable features with three separate linear layers: " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{m}} = \\mathrm{Linear}(\\mathbf{m}),\\mathbf{X}_{\\mathrm{u}} =" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " Linear " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}^{u})" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1 = \\mathrm{Linear}(\\mathbf{p}^l)" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{m}}\\in \\mathbb{R}^{T^{\\prime}\\times D}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{u}}\\in \\mathbb{R}^{T^{\\prime}\\times D}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1\\in \\mathbb{R}^{T^{\\prime}\\times D}" + }, + { + "bbox": [ + 313, + 474, + 556, + 581 + ], + "type": "text", + "content": " and D is the channel dimension of the features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 668 + ], + "type": "text", + "content": "The dance generation task is reframed as selecting the most probable future pose code from the codebook " + }, + { + "bbox": [ + 313, + 582, + 556, + 668 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 582, + 556, + 668 + ], + "type": "text", + "content": ", conditioned on the music and prior movements. Since the upper and lower bodies are modeled separately, maintaining coherence and avoiding asynchrony (e.g., opposing directions) requires cross-conditioned prediction, leveraging mutual information between existing movements:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 672, + 555, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 672, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 329, + 672, + 555, + 690 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {t} ^ {u} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {u} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right), \\tag {2}", + "image_path": "aaf5de697c89bdd795fbef2230bf0cd983a832cf9425a01b0f02684a1666bcbd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 332, + 692, + 555, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 692, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 332, + 692, + 555, + 711 + ], + "type": "interline_equation", + "content": "\\hat {p} _ {t} ^ {l} = \\arg \\max _ {k} \\mathbb {P} \\left(\\mathbf {z} _ {k} ^ {l} \\mid \\mathbf {m} _ {1 \\dots t}, p _ {0 \\dots t - 1} ^ {u}, p _ {0 \\dots t - 1} ^ {l}\\right). \\tag {3}", + "image_path": "2c348a74dac5e34d49259316fa9de14178e0f6a02a46312b6ec393cfc3fc6e20.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13195" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "At each time step " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ", MotionGPT [30] estimates the probabilities of pose codes " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i \\in \\mathbb{Z}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and selects the one with the most probable pose codes as the predicted upper pose " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^u" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and predicted lower pose " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\hat{p}_t^l" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "content": "Specifically, the input feature " + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i\\in \\mathbb{R}^{(3\\times T')\\times D}" + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "content": " are concatenated by upper body motion feature, lower body motion feature and music feature as follows: " + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i =" + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\mathbf{X}_{\\mathrm{m}},\\mathbf{X}_{\\mathrm{u}},\\mathbf{X}_{\\mathrm{l}})" + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "content": ". Subsequently, " + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "content": " is passed through successive Autoregressive Transformer layers " + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{AR}" + }, + { + "bbox": [ + 55, + 120, + 295, + 216 + ], + "type": "text", + "content": ", followed by a linear transformation and a softmax layer, generates the probability distributions for predicting body pose collectively:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 224, + 295, + 237 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 224, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 224, + 295, + 237 + ], + "type": "interline_equation", + "content": "\\mathbf {a} ^ {h} = \\operatorname {S o f t m a x} (\\operatorname {L i n e a r} \\left(\\mathcal {F} _ {A R} \\left(\\mathbf {X} _ {i}\\right)\\right), \\tag {4}", + "image_path": "db77bcb96e746ba5852308997876d522b6518d2982311e8c5856154712a72f77.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^h = [\\mathbf{a}^u,\\mathbf{a}^l]" + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^u" + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "text", + "content": " denotes that of upper body and " + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^l" + }, + { + "bbox": [ + 55, + 245, + 294, + 270 + ], + "type": "text", + "content": " denotes that of lower body." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "text", + "content": "Finally, the model is optimized via supervised training with the cross-entropy loss on action probability " + }, + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t^h" + }, + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "text", + "content": " at each time step " + }, + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 270, + 295, + 306 + ], + "type": "text", + "content": " and for both upper body and lower body:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 315, + 295, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 315, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 69, + 315, + 295, + 349 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C E} = \\frac {1}{T ^ {\\prime}} \\sum_ {t = 0} ^ {T ^ {\\prime} - 1} \\sum_ {h = u, l} \\text {C r o s s E n t r o p y} \\left(\\mathrm {a} _ {t} ^ {h}, p _ {t + 1} ^ {h}\\right). \\tag {5}", + "image_path": "e8a36057c0912227503c52f03fae8bac957d1f70a6836cb8433760b007af4d62.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 393 + ], + "type": "text", + "content": "In inference, the MotionGPT predicts pose codes according to the initial pose code and the entire music, followed by Pose VQ-VAE decoders generating a new dance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 400, + 270, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 270, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 270, + 412 + ], + "type": "text", + "content": "3.2. Phase-Based Rhythm Feature Extraction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 418, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 536 + ], + "type": "text", + "content": "To enhance rhythmic feature extraction in music-driven dance generation, we propose Phase-Based Rhythm Extraction (PRE), which decouples rhythm from musical semantics by leveraging the inherent periodicity and temporal positioning of phase representations. The architecture of PRE is shown in the green part of Figure 1. Unlike existing methods that entangle rhythm with melody and harmony, PRE enables independent and precise rhythm modeling as shown in Figure 2, forming a solid foundation for gating-enhanced rhythm-aware mechanism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 537, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 621 + ], + "type": "text", + "content": "Specifically, we adopt the Short-Time Fourier Transform (STFT) instead of the conventional Fourier Transform (FT), since the non-stationary nature of music and the sliding-window approach of STFT effectively captures sudden tempo shifts and complex rhythmic patterns [37]. Specifically, we first perform STFT on the input music feature to extract phase angles:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 132, + 629, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 629, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 132, + 629, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\mathbf {S} _ {m} = \\operatorname {S T F T} (\\mathbf {m}), \\tag {6}", + "image_path": "887a2d2da1ae71bdaabc4404cb0ddab71ee36d5c48b8c54c87e77cf2c591d16f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 644, + 216, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 644, + 216, + 658 + ], + "spans": [ + { + "bbox": [ + 141, + 644, + 216, + 658 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\varphi} = \\operatorname {A n g l e} \\left(\\mathbf {S} _ {m}\\right),", + "image_path": "c26c906a1fa60e0a6b80021efb263d781a0c65d05f4232aa2691c2863805d258.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_m" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": " represents the complex spectrogram obtained via STFT, " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Angle}(\\cdot)" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": " computes the phase angle of a complex-valued input and " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\varphi^{\\prime}" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": " denotes the phase angles extracted from " + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_m" + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": ". By taking the angle of the STFT output, we obtain the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 191 + ], + "type": "image", + "image_path": "4516cdb0664827a71e3752b6f55f139686de0bdcbbea09f1563fc9974db24f4f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 200, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 553, + 265 + ], + "type": "text", + "content": "Figure 2. Visualizing the input music features and phase features. We notice that only a small portion of the input music features contains meaningful musical information, i.e., rhythm-poor features. And after a short-time Fourier transform (STFT), each corresponding token has a rich periodic Phase Feature, i.e., rhythm-rich features." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "text", + "content": "music phase data " + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "inline_equation", + "content": "\\varphi^{\\prime}\\in \\mathbb{C}^{T_{\\varphi}\\times D_{\\varphi}}" + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "inline_equation", + "content": "T_{\\varphi}" + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "text", + "content": " represents the number of time frames obtained after the STFT transformation, and " + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "inline_equation", + "content": "D_{\\varphi}" + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "text", + "content": " is the channel dimension of the " + }, + { + "bbox": [ + 313, + 273, + 553, + 310 + ], + "type": "inline_equation", + "content": "\\varphi^{\\prime}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 310, + 553, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 553, + 381 + ], + "type": "text", + "content": "The extracted phase angles " + }, + { + "bbox": [ + 313, + 310, + 553, + 381 + ], + "type": "inline_equation", + "content": "\\varphi^{\\prime}" + }, + { + "bbox": [ + 313, + 310, + 553, + 381 + ], + "type": "text", + "content": " inherently reflect temporal shifts and periodic structures of musical rhythms: low-frequency phase variations correspond to broader rhythmic patterns, while high-frequency variations capture finer rhythmic details. We also apply a center-cropping strategy, selectively focusing on the most informative phase regions:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 392, + 553, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 392, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 388, + 392, + 553, + 406 + ], + "type": "interline_equation", + "content": "\\varphi = \\operatorname {C e n t e r C r o p} \\left(\\varphi^ {\\prime}\\right), \\tag {7}", + "image_path": "c401034b8cffe3034072c766afd2fd12ff90cffc94e0e222036f15090c5cf2d4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\varphi \\in \\mathbb{C}^{T' \\times D_{\\varphi}}" + }, + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "text", + "content": " has the same temporal length " + }, + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 313, + 409, + 553, + 445 + ], + "type": "text", + "content": ", addressing the potential temporal misalignment and preserving critical rhythmic cues." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 445, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 481 + ], + "type": "text", + "content": "Subsequently, the phase features through a linear transformation, enhancing phase features representing musical rhythms. The process is shown in the following equation:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 488, + 553, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 488, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 367, + 488, + 553, + 502 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\varphi} = \\operatorname {R e L U} (\\mathrm {B N} (\\operatorname {L i n e a r} (\\varphi))), \\tag {8}", + "image_path": "7809cafb8eb2647058368d8167b5ddf3761e9a8141f4848d90dabcb11245ad96.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "where the embedded phase features " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\varphi} \\in \\mathbb{C}^{T' \\times D}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " encode rich rhythmic information, providing a clearer and more structured representation of rhythmic and musical elements. Finally, we fuse the input feature " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " with the rhythmic features " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\varphi}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " by concatenating them: " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\gamma} = \\mathrm{Concat}(\\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi}, \\mathbf{X}_{\\varphi})" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": ". Then, we sum the rhythm-enhanced features with the original music, upper body, and lower body motion features: " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\mathbf{X}_f + \\mathbf{X}_{\\gamma}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " represents the rhythmically enhanced features." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 624, + 539, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 539, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 539, + 637 + ], + "type": "text", + "content": "3.3. Global Beat Attention via Temporal Gating" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Leveraging these rhythm-rich features, we next investigate how to integrate them into the alignment process between music and dance. However, prior methods [11, 30, 31, 36, 44] relying on cross-conditional causal attention mechanism (denoted as " + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^3\\mathrm{Attention}" + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": ") exhibit a critical misalignment issue. As shown in Figure 3 (a), we visualize the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13196" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "cross-conditional causal attention mechanism and observe that the next predicted token lacks direct control signals from historical tokens, failing to establish a stable global beat attention. This misalignment leads to error accumulation, which we identify as a key factor behind the unnatural dance generation in prior work." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 154, + 157, + 252 + ], + "blocks": [ + { + "bbox": [ + 58, + 154, + 157, + 252 + ], + "lines": [ + { + "bbox": [ + 58, + 154, + 157, + 252 + ], + "spans": [ + { + "bbox": [ + 58, + 154, + 157, + 252 + ], + "type": "image", + "image_path": "eaec8bec3d5f539fab42a024e18362868b0f88f6dc153345758089dcf623cdb9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 254, + 153, + 270 + ], + "lines": [ + { + "bbox": [ + 63, + 254, + 153, + 270 + ], + "spans": [ + { + "bbox": [ + 63, + 254, + 153, + 270 + ], + "type": "text", + "content": "(a) Cross-Conditional Causal Attention" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 154, + 294, + 253 + ], + "blocks": [ + { + "bbox": [ + 167, + 154, + 294, + 253 + ], + "lines": [ + { + "bbox": [ + 167, + 154, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 167, + 154, + 294, + 253 + ], + "type": "image", + "image_path": "964265c1bd018d27bf9512a68d8abf13ad6ad772252f81b68fc9c55ede5bf566.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 167, + 254, + 272, + 270 + ], + "lines": [ + { + "bbox": [ + 167, + 254, + 272, + 270 + ], + "spans": [ + { + "bbox": [ + 167, + 254, + 272, + 270 + ], + "type": "text", + "content": "(b) Temporal-Gated Causal Attention (TGCA)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 280, + 296, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 357 + ], + "type": "text", + "content": "Figure 3. Heat map of attention. The two displayed images show the results learned by Cross-Conditional Causal Attention (a) and Temporal-Gated Causal Attention (b) under the same piece of music. The brighter the color, the higher the attention; the darker the color, the lower the attention. TGCA the next token has a clear control signal, preventing overly random dance movements and achieving enhanced rhythm-aware feature representation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 369, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 464 + ], + "type": "text", + "content": "Building on this, we propose Temporal Gated Causal Attention (TGCA) to enhance global rhythm feature attention, improving music-dance alignment and ensuring the next token is guided by explicit historical information, preventing random action generation. As illustrated in the blue section of Figure 1, the rhythmically enhanced features " + }, + { + "bbox": [ + 55, + 369, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 369, + 296, + 464 + ], + "type": "text", + "content": " are passed through TGCA module, which can be expressed as the following:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 472, + 295, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 472, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 107, + 472, + 295, + 485 + ], + "type": "interline_equation", + "content": "\\operatorname {G a t i n g} (\\mathbf {X}) = \\operatorname {S i L U} (\\operatorname {L i n e a r} (\\mathbf {X})). \\tag {9}", + "image_path": "2e400d129961f80e436b4773e72e89f6e85cc77a62c1af5e8389ef8c5a92066e.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 492, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 528 + ], + "type": "text", + "content": "The final output is obtained by the element-wise multiplication of cross-conditional causal attention " + }, + { + "bbox": [ + 55, + 492, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^3" + }, + { + "bbox": [ + 55, + 492, + 296, + 528 + ], + "type": "text", + "content": " Attention and Gating, as shown in the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 535, + 294, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 535, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 108, + 535, + 294, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {X} _ {a t t n} = \\operatorname {T G C A} (\\mathbf {X}) \\\\ = \\mathrm {C} ^ {3} \\text {A t t e n t i o n} (\\mathbf {X}) \\odot \\operatorname {G a t i n g} (\\mathbf {X}). \\tag {10} \\\\ \\end{array}", + "image_path": "d248a1197e0efe1ec4ce9a65f077ae74288983ebd0ecafaedb9497e992949366.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "content": "As shown in Figure 3 (b), our proposed gated causal attention mechanism allows historical tokens to better predict the next token, implying an enhanced influence of global rhythmic patterns on dance generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 270, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 270, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 270, + 637 + ], + "type": "text", + "content": "3.4. Mamba-Based Parallel Motion Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "After TGCA enhances dance movements in " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{attn}" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " with global rhythmic perception, we looking into dance modeling for generation. Dance generation faces two key challenges: (1) Dance is a complex spatial-temporal sequence encoding trajectory, posture, and velocity over time, requiring a model capable of capturing both local transitions and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "global rhythmic variations. (2) The upper and lower body exhibit distinct movement patterns and speeds, necessitating separate processing to maintain coherence and diversity. To address these issues, we propose a parallel Mamba motion modeling architecture that separately models music, upper-body, and lower-body movements while leveraging state space modeling to ensure fluid and diverse motion generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 251 + ], + "type": "text", + "content": "Specifically, to ensure each action is primarily influenced by relevant preceding movements, we introduce a GateMlp in the Mamba block, as shown in Figure 1. This mechanism selectively retains essential motion features while filtering out less relevant information, thereby enhancing the coherence and expressiveness of the generated dance sequences. The process is formulated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 330, + 261, + 553, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 261, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 330, + 261, + 553, + 275 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {m b} ^ {\\prime} = \\operatorname {M a m b a} (\\operatorname {R M S N o r m} (\\mathbf {X} _ {a t t n})) + \\mathbf {X} _ {a t t n}, \\tag {11}", + "image_path": "316bc2cda0f7a019c5d26a3b63f9bdacea3acd83cd56c9fcac3aca6a1fb533d6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 333, + 281, + 553, + 296 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 281, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 333, + 281, + 553, + 296 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {m b} = \\operatorname {G a t e M l p} \\left(\\operatorname {R M S N o r m} \\left(\\mathbf {X} _ {m b} ^ {\\prime}\\right)\\right) + \\mathbf {X} _ {m b} ^ {\\prime}. \\tag {12}", + "image_path": "c82def2246b9e9fa76f9b293ddd293478cb3df044e76464c068b8b04a6c2df7e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 388 + ], + "type": "text", + "content": "The gating mechanism dynamically regulates the frequency, speed, and amplitude of movements in response to musical beats, ensuring that dance sequences remain synchronized with the music. Finally, three parallel Mamba architectures model upper- and lower-body dance sequences separately, therefore ensure the generated dance movements are expressive, diverse, and temporally coherent." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 400, + 395, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 395, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 395, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 430, + 434 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": "Dataset: We follow prior works [27, 30, 36] and perform all experiments on the AIST++ dataset [18], the most widely used benchmark for music-driven dance generation that avoids observable biases (e.g., offensive postures). AIST++ consists of 991 high-quality 3D pose sequences recorded at 60 FPS in skinned multi-person linear (SMPL) format [22], with 951 sequences designated for training and 40 reserved for evaluation. Following [27, 30, 36], we generate 40 pieces of dance sequences on the test set of AIST++ dataset, and sample the generated dance sequence with length of 20 seconds for performance evaluation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Evaluation Metrics: We follow previous studies [18, 30] to quantitatively evaluate the generated samples in three key aspects: quality, diversity, and beat-dance alignment. Prior to evaluation, it is necessary to extract kinetic features [26] and geometric features [24] from both the ground truth and the generated samples via a specific toolbox [6]. For dance quality, we compute two types of the Fréchet Inception Distance (FID) [8], including " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " based on kinetic features and " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " based on geometric features. Lower " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " indicates that the generated dance has more similar kinetic features to the original dance, meaning the generated dances are more natural and smooth. Lower " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " indicates that" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13197" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 135, + 549, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 126 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on the AIST++ dataset. Underlining indicates the best performance among existing methods. Blue indicates results that surpass the best existing method. ↓ indicates that lower values are better, while ↑ indicates that higher values are better. * represents the dances generated by \"Li et al.\" that exhibit significant jitteriness, resulting in extremely high velocity variation, as also noted in [18]. † indicates the reproduced results obtained using Bailando++ official code and published checkpoints. ‡ denotes the results that are averaged over five independent training runs." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 135, + 549, + 309 + ], + "lines": [ + { + "bbox": [ + 60, + 135, + 549, + 309 + ], + "spans": [ + { + "bbox": [ + 60, + 135, + 549, + 309 + ], + "type": "table", + "html": "
METHODVENUEMOTION QUALITYMOTION DIVERSITYBEAT ALIGN SCORE↑
FIDk↓FIDg↓DIVk↑DIVg↑
GROUND TRUTH-17.1010.608.197.450.2374
Li et al. [17]ARXIV 202086.4343.46*6.853.320.1607
DANCERUVOLUTION [12]ICLR 202173.4225.923.524.870.1950
DANCENET [45]TOMM 202269.1825.492.862.850.1430
FACT [18]ICCV 202135.3522.115.946.180.2209
BAILANDO [30]CVPR 202228.169.627.836.340.2332
EDGE [33]CVPR 202342.1622.123.964.610.2334
DIFFDANCE [27]ACMMM 202324.0920.686.022.890.2418
†BAILANDO++ [31]TPAMI 202322.7411.587.946.340.2263
E3D2 [36]AAAI 202426.258.947.966.490.2232
LODGGE [20]CVPR 202437.0918.795.584.850.2423
OURS (BEST)11.67(-11.07)11.908.52(+0.58)7.55(+1.21)0.2714(+2.91%)
‡OURS (AVERAGE)15.48(-7.26)13.577.857.49(+1.15)0.2779(+3.56%)
", + "image_path": "45bfca63d204bd684bec9d822b2e77fec636aa7a3dc03d40cfe1ecd3f120766c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 317, + 529, + 545 + ], + "blocks": [ + { + "bbox": [ + 82, + 317, + 529, + 545 + ], + "lines": [ + { + "bbox": [ + 82, + 317, + 529, + 545 + ], + "spans": [ + { + "bbox": [ + 82, + 317, + 529, + 545 + ], + "type": "image", + "image_path": "671396d25209960c5a5cb3ed3dce7add1dd04ea85c7e68ad8cb7e095eabf6218.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 553, + 555, + 577 + ], + "lines": [ + { + "bbox": [ + 55, + 553, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 555, + 577 + ], + "type": "text", + "content": "Figure 4. Dance examples generated by our proposed method on various types of music. The demo videos can be found in the supplementary material. The character is sourced from Mixamo [23]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "content": "the generated dance has more similar geometric features, reflecting the quality of choreography. For dance diversity, we employ metrics " + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "content": " [18], which denote the distance of kinetic and distance of geometric features, respectively. Higher " + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 55, + 589, + 297, + 709 + ], + "type": "text", + "content": " signify the greater amplitude and diversity of motion, respectively. Finally, to evaluate beat-dance alignment, we use the Beat Align Score (BAS) [18], which quantifies the average distance between a music beat and its nearest dance beat. Higher BAS indicates greater degree of beat-dance alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 589, + 539, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 539, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 539, + 601 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "To validate the superiority of our method, we conduct comparative experiments with the state-of-the-art methods [12, 17, 18, 20, 27, 30, 31, 33, 36, 45], as well as the ground truth. Following [27, 30, 36], we generate 40 pieces of dance sequences on the AIST++ test set, and sample the generated dance sequence with the length of 20 seconds to compute the metrics mentioned above. Among the compared methods, \"Li et al.\" [17]," + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13198" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "DanceNet [45] and FACT [18] use the results from the AIST++ benchmark [18], while the results of DanceRevolution [12] are derived from its reproduction as presented in Bailando [30]. EDGE [33] utilizes the reproduced results from the Lodge [20]. The Bailando, DiffDance [27] and Lodge methods use the experimental results from their respective publications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": "The comparative results are presented in Table 1. Our method achieves state-of-the-art performance, significantly outperforming existing methods in terms of " + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": ", and Beat Align Score (BAS), achieving the best results. Additionally, our method ranks third on the " + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 55, + 157, + 294, + 266 + ], + "type": "text", + "content": " metric. The generated dance examples of our proposed Danceba on various types of music are shown in Figure 4, and the demo videos are provided in the supplementary files." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": "Motion Quality Metrics: As illustrated in Table 1, our method outperforms existing methods in overall quality performance. To be specific, our method improves on " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "48.68\\%" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " (11.67 vs. 22.74), while exhibits competitive edge of " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": ". Further analysis of the employed quality metrics reveals distinct roles for the two metrics types. The metric " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": ", defined based on velocity and energy, primarily captures the kinetic characteristics of the generated dance, while " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": ", derived from multiple manually designed motion templates, emphasizes the geometric structural properties of the dance. Thus, our method demonstrates superior performance in modeling the kinetic aspects of sequential motions, as evidenced by its exceptional " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " scores. Although it exhibits a modest gap in " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " (11.90 vs. 11.58) compared to the state-of-the-art Bailando++ [31], it achieves a " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "19.0\\%" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " higher " + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 55, + 266, + 295, + 506 + ], + "type": "text", + "content": " (7.55 vs. 6.34), indicating improved motion diversity with comparable generative quality. Despite this minor trade-off, our method outperforms existing approaches overall, achieving a robust balance of efficiency and quality across a broader range of metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": "Motion Diversity Metrics: Regarding diversity evaluation, Danceba achieves the best diversity performance, which improves " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": " (8.52 vs. 7.96) and even slightly outperforms the Ground Truth (8.52 vs. 8.19). Meanwhile, our method improves " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "inline_equation", + "content": "16.3\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 664 + ], + "type": "text", + "content": " (7.55 vs. 6.49) and also slightly surpasses the Ground Truth (7.55 vs. 7.45), as illustrated in Table 1. For a visual comparison, we also present a comparison with the state-of-the-art method Bailando, as illustrated in Figure 5. It can be observed that Bailando exhibits relatively small movement amplitudes, with some dance movements missing and the actions appearing stiff. In contrast, our method shows larger movement amplitudes and movement diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Beat-dance Alignment Metrics: Danceba achieves the best performance in beat-dance alignment, as quantified by the BAS metric. Specifically, it improves BAS by " + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": " (0.2714 vs. 0.2423) compared to the previous best re" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 224 + ], + "type": "image", + "image_path": "3e51635a02850c5a5e6e05580b01b05b7d4c0f246b64a550d94bf06e9342243d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 232, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 266 + ], + "type": "text", + "content": "Figure 5. Comparison with the state-of-the-art method Bailando [30]. Visual comparisons with Bailando can be found in the supplementary video." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 274, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 553, + 357 + ], + "type": "text", + "content": "sult and surpasses the ground truth by " + }, + { + "bbox": [ + 313, + 274, + 553, + 357 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 313, + 274, + 553, + 357 + ], + "type": "text", + "content": " (0.2714 vs. 0.2374), as shown in Table 1. Figure 6 further visualizes the alignment between music beats and motion beats of the generated dance compared to Bailando++ [31]. The results show that Danceba achieves a significantly smaller beat distance, demonstrating its ability to generate dance sequences that more accurately follow musical rhythms." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 369, + 432, + 441 + ], + "blocks": [ + { + "bbox": [ + 319, + 369, + 432, + 441 + ], + "lines": [ + { + "bbox": [ + 319, + 369, + 432, + 441 + ], + "spans": [ + { + "bbox": [ + 319, + 369, + 432, + 441 + ], + "type": "image", + "image_path": "ddfe8af1c30e219ea492b3880cb58289e2af562b3847287755c8ed37fd4fdaec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 452, + 553, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 452, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 553, + 507 + ], + "type": "text", + "content": "Figure 6. Beats alignment visualization, where the horizontal axis shows frame indices of beat events. Comparing Danceba with Balando++ [31], we can find that the distance between motion beats and music beats generated by our method is smaller. This indicating that Danceba performs better in terms of rhythmic alignment." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 443, + 369, + 548, + 441 + ], + "blocks": [ + { + "bbox": [ + 443, + 369, + 548, + 441 + ], + "lines": [ + { + "bbox": [ + 443, + 369, + 548, + 441 + ], + "spans": [ + { + "bbox": [ + 443, + 369, + 548, + 441 + ], + "type": "image", + "image_path": "424c83aebeab355a4a9655b97400f17d55a07a4764b95ef6bd8ee43b7ab8b947.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 516, + 415, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 415, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 415, + 528 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 641 + ], + "type": "text", + "content": "We conduct ablation studies to evaluate the contributions of the key modules (i.e., PRE, TGCA, and PMMM) in Danceba, where \"w/o\" denotes the removal of a specific module. \"w/o PMMM\" refers to the setting where PMMM is replaced with TGCA layers to ensure fairness. The results are presented in Table 2. Additionally, we assess our method using a single Mamba structure instead of the parallel Mamba structure for motion modeling, referred to as Danceba-Single, with results shown in Table 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": "Note that, all ablation results are averaged over five independent training runs to ensure the credibility and robustness of our results. The motion quality score (FID) is computed as the average of " + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": ", while the motion diversity score (Div) is the average of " + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_k" + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{Div}_g" + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Phase-Based Rhythm Extraction: To validate the effec" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13199" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 118, + 276, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 2. Ablation study of three key modules (i.e., PRE, TGCA, and PMMM), and \"w/o\" indicates Danceba without specified module. Visual comparisons with Bailando can be found in the supplementary video." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 118, + 276, + 198 + ], + "lines": [ + { + "bbox": [ + 72, + 118, + 276, + 198 + ], + "spans": [ + { + "bbox": [ + 72, + 118, + 276, + 198 + ], + "type": "table", + "html": "
METHODFID ↓DIV ↑BAS ↑
GROUND TRUTH13.857.820.2374
w/o TGCA33.226.74-
w/o PMMM25.245.78-
w/o PRE20.626.530.2474
DANCEBA14.537.670.2779
", + "image_path": "c6bf13e4f27a44c31d67508035a0d8f6731cf8be5ef6d9daf389f96d2c62a314.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 220, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 294, + 339 + ], + "type": "text", + "content": "tiveness of PRE, we replace the rhythm feature extraction module with a learnable linear layer as used in previous works [11, 30, 31, 36, 44]. Consequently, the motion quality (i.e., FID), motion diversity (i.e., Div), and beat alignment score (i.e., BAS) decrease by 6.09 (41.91%), 1.14 (14.86%), and 0.0305 (10.96%), respectively, as illustrated in Table 2. Specifically, this performance drop highlights the contribution of PPE, which effectively captures the critical rhythm features of music, a crucial aspect that previous methods have largely overlooked." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 472 + ], + "type": "text", + "content": "Temporal-Gated Causal Attention: As shown in Table 2, we validate the effectiveness of the proposed TGCA by training a Danceba variant that replaces TGCA with conventional causal cross-conditional attention. The results demonstrate that the full Danceba model outperforms this variant in both FID and Div, highlighting TGCA's ability to effectively leverage global rhythmic features extracted by PRE. This significantly enhances the contextual association between music and dance movements, resulting in dance sequences that not only accurately reflect musical rhythm but also exhibit greater coherence and expressive diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 581 + ], + "type": "text", + "content": "Parallel Mamba Motion Modeling: To evaluate the effectiveness of Parallel Mamba Motion Modeling, we train a model without this module. As shown in Table 2, Danceba outperforms its variant without using PMMM in terms of both FID and Div, demonstrating the importance of parallel Mamba modeling. These results validate Mamba's superior sequential modeling capability, which is particularly well-suited for dance generation as it requires both fine-grained inter-modal interactions and long-sequence modeling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "To further assess the advantage of using a parallel Mamba architecture instead of a single Mamba architecture for motion modeling, we compare Danceba with Danceba-Single. Unlike Danceba-Single, which treats the body as a unified entity and disregards the distinction between upper and lower-body movements, PMMM explicitly models their distinct motion dynamics. Additionally, PMMM integrates rhythm-aware representation alongside these distinct pose features, enhancing beat-dance alignment with the musical rhythm. As shown in Table 3, Danceba significantly improves expressiveness and temporal coherence, with " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_k" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 106, + 537, + 167 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 103 + ], + "type": "text", + "content": "Table 3. Ablation study on parallel Mamba motion modeling, where Danceba-Single denotes our method using single Mamba architecture." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 106, + 537, + 167 + ], + "lines": [ + { + "bbox": [ + 328, + 106, + 537, + 167 + ], + "spans": [ + { + "bbox": [ + 328, + 106, + 537, + 167 + ], + "type": "table", + "html": "
METHODFIDk ↓FIDg ↑
GROUND TRUTH17.1010.60
DANCEBA-SINGLE82.5083.97
DANCEBA15.48(-67.02)13.57(-60.40)
", + "image_path": "0dd0c8845ca5b1e932339cc43bb1f2a2caa369b4747c4fb371e1ea5fa9c52515.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 205 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 180, + 553, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_g" + }, + { + "bbox": [ + 313, + 180, + 553, + 205 + ], + "type": "text", + "content": " scores improving by 67.02 and 60.40, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 221, + 389, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 389, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 389, + 234 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 448 + ], + "type": "text", + "content": "Although Danceba achieves superior dance generation quality compared to state-of-the-art methods, it has two key limitations: (1) Music feature encoding: Danceba uses a simple linear layer instead of a pre-trained audio model (e.g., Jukebox [3], MERT [21], or CLAP [5]). While computationally efficient, this design may fail to capture nuanced musical structures, affecting rhythm-motion synchronization. Future work could explore pre-trained music encoders to enhance feature extraction. (2) 3D pose quantization: Danceba adopts the Pose VQ-VAE framework from Bailando [30] for fair comparison with prior SOTA methods [30, 31, 36]. However, it has a limited encoding space, which may limit motion diversity. Such a trade-off between quantization efficiency and expressiveness could hinder the preservation of fine-grained motion details. Exploring hierarchical quantization or adaptive codebook learning could further improve motion quality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "This paper has presented a novel framework Danceba for music-driven dance generation that enhances rhythm-aware feature representation and motion modeling. By introducing phase-based rhythm extraction and temporal-gated causal attention, our approach ensures precise beat-dance synchronization and effectively integrates rhythmic structures into dance movements. Additionally, parallel Mamba motion modeling separately captures the distinct motion dynamics of the upper and lower body while incorporating rhythm-aware representations, improving the naturalness and diversity of the generated dance. Extensive experiments demonstrate that Danceba significantly outperforms state-of-the-art methods, generating dance movements that are more rhythmic, natural, and diverse while maintaining strong alignment with musical beats. Future work could integrate pre-trained audio encoders for richer music representation, explore advanced quantization to enhance motion expressiveness and diversity, or extend to style-conditioned motion synthesis for broader applicability." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13200" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 140 + ], + "type": "text", + "content": "This work was partly supported by the Fundamental Research Funds for the Central Universities (Grant No. 3072025YY0601), and the Beijing Nova Program (Grant No. 20230484261)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 115, + 167 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 294, + 218 + ], + "type": "text", + "content": "[1] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Trans. Graph. (TOG), 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 294, + 262 + ], + "type": "text", + "content": "[2] Tri Dao and Albert Gu. Transformers are SSMs: Generalized models and efficient algorithms through structured state space duality. In Proc. Int. Conf. Mach. Learn. (ICML), 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 266, + 295, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 266, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 62, + 266, + 295, + 309 + ], + "type": "text", + "content": "[3] Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020.8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 312, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 295, + 356 + ], + "type": "text", + "content": "[4] Kun Dong, Jian Xue, Zehai Niu, Xing Lan, Ke Lu, Qingyuan Liu, and Xiaoyu Qin. Realistic full-body motion generation from sparse tracking with state space model. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 402 + ], + "type": "text", + "content": "[5] Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP learning audio concepts from natural language supervision. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 436 + ], + "type": "text", + "content": "[6] Deepak Gopinath and Jungdam Won. Fairmotion - tools to load, process and visualize motion capture data. *GibHub*, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 439, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 294, + 472 + ], + "type": "text", + "content": "[7] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 474, + 294, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 474, + 294, + 528 + ], + "spans": [ + { + "bbox": [ + 62, + 474, + 294, + 528 + ], + "type": "text", + "content": "[8] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 531, + 294, + 563 + ], + "type": "text", + "content": "[9] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "type": "text", + "content": "[10] Daniel Holden, Jun Saito, and Taku Komura. A deep learning framework for character motion synthesis and editing. ACM Trans. Graph. (TOG), 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 666 + ], + "type": "text", + "content": "[11] Qiaochu Huang, Xu He, Boshi Tang, Haolin Zhuang, Liyang Chen, Shuochen Gao, Zhiyong Wu, Haozhi Huang, and Helen Meng. Enhancing expressiveness in dance generation via integrating frequency and music style information. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2024. 1, 2, 4, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[12] Ruozhi Huang, Huang Hu, Wei Wu, Kei Sawada, Mi Zhang, and Daxin Jiang. Dance revolution: Long-term dance generation with music via curriculum learning. In Proc. Int. Conf. Learn. Represent. (ICLR), 2021. 2, 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Jinwoo Kim, Heeseok Oh, Seongjean Kim, Hoseok Tong, and Sanghoon Lee. A brand new dance partner: Music-conditioned pluralistic dancing controlled by multiple dance genres. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[14] Matthew Kyan, Guoyu Sun, Haiyan Li, Ling Zhong, Paisarn Muneesawang, Nan Dong, Bruce Elder, and Ling Guan. An approach to ballet dance training through MS Kinect and visualization in a CAVE virtual reality environment. ACM Trans. Intell. Syst. Technol. (TIST), 2015. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "text", + "content": "[15] Hsin-Ying Lee, Xiaodong Yang, Ming-Yu Liu, Ting-Chun Wang, Yu-Ding Lu, Ming-Hsuan Yang, and Jan Kautz. Dancing to music. In Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "text", + "content": "[16] Chengjian Li, Xiangbo Shu, Qiongjie Cui, Yazhou Yao, and Jinhui Tang. FTMoMamba: Motion generation with frequency and text state space models. arXiv preprint arXiv:2411.17532, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 315 + ], + "type": "text", + "content": "[17] Jiaman Li, Yihang Yin, Hang Chu, Yi Zhou, Tingwu Wang, Sanja Fidler, and Hao Li. Learning to generate diverse dance motions with Transformer. arXiv preprint arXiv:2008.08171, 2020. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 360 + ], + "type": "text", + "content": "[18] Ruilong Li, Shan Yang, David A Ross, and Angjoo Kanazawa. Ai choreographer: Music conditioned 3D dance generation with AIST++. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 415 + ], + "type": "text", + "content": "[19] Ronghui Li, Junfan Zhao, Yachao Zhang, Mingyang Su, Zeping Ren, Han Zhang, Yansong Tang, and Xiu Li. Finedance: A fine-grained choreography dataset for 3D full body dance generation. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 480 + ], + "type": "text", + "content": "[20] Ronghui Li, YuXiang Zhang, Yachao Zhang, Hongwen Zhang, Jie Guo, Yan Zhang, Yebin Liu, and Xiu Li. Lodge: A coarse to fine diffusion network for long dance generation guided by the characteristic dance primitives. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2024. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 482, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 547 + ], + "type": "text", + "content": "[21] Yizhi Li, Ruibin Yuan, Ge Zhang, Yinghao Ma, Xingran Chen, Hanzhi Yin, Chenghao Xiao, Chenghua Lin, Anton Ragni, Emmanouil Benetos, et al. MERT: Acoustic music understanding model with large-scale self-supervised training. In Proc. Int. Conf. Learn. Represent. (ICLR), 2024. 1, 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "text", + "content": "[22] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Trans. Graph. (TOG), 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 582, + 507, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 507, + 591 + ], + "type": "text", + "content": "[23] Mixamo. https://www MIXamo.com/.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 624 + ], + "type": "text", + "content": "[24] Meinard Müller, Tido Röder, and Michael Clausen. Efficient content-based retrieval of motion capture data. In Proc. ACM Conf. SIGGRAPH, 2005. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "text", + "content": "[25] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Kensuke Onuma, Christos Faloutsos, and Jessica K. Hodgins. Fmdistance: A fast and effective distance function for motion capture data. In Eurographics, 2008. 5" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13201" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Qiaosong Qi, Le Zhuo, Aixi Zhang, Yue Liao, Fei Fang, Si Liu, and Shuicheng Yan. DiffDance: Cascaded human motion diffusion model for dance generation. In Proc. ACM Int. Conf. Multimedia (ACMMM), 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[28] Ziyun Qian, Zeyu Xiao, Zhenyi Wu, Dingkang Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, Dongliang Kou, and Lihua Zhang. SMCD: High realism motion style transfer via Mamba-based diffusion. arXiv preprint arXiv:2405.02844, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[29] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[30] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando: 3D dance generation by actor-critic GPT with choreographic memory. In Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[31] Li Siyao, Weijiang Yu, Tianpei Gu, Chunze Lin, Quan Wang, Chen Qian, Chen Change Loy, and Ziwei Liu. Bailando++: 3D dance GPT with choreographic memory. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "text", + "content": "[32] Guofei Sun, Yongkang Wong, Zhiyong Cheng, Mohan S Kankanhalli, Weidong Geng, and Xiangdong Li. Deepdance: music-to-dance motion choreography with adversarial learning. IEEE Trans. Multimedia (TMM), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 381 + ], + "type": "text", + "content": "[33] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), 2023. 1, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[34] Shuhei Tsuchida. Dance information processing: Computational approaches for assisting dance composition. New Gener. Comput., 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 459 + ], + "type": "text", + "content": "[35] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[36] Zilin Wang, Haolin Zhuang, Lu Li, Yinmin Zhang, Junjie Zhong, Jun Chen, Yu Yang, Boshi Tang, and Zhiyong Wu. Explore 3D dance generation via reward model from automatically-ranked demonstrations. In Proc. AAAI Conf. Artif. Intell. (AAAI), 2024. 1, 4, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "type": "text", + "content": "[37] Simon Welker, Julius Richter, and Timo Gerkmann. Speech enhancement with score-based generative models in the complex STFT domain. arXiv preprint arXiv:2203.17004, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[38] Shuo Yan, Gangyi Ding, Zheng Guan, Ningxiao Sun, Hong-song Li, and Longfei Zhang. OutsideMe: Augmenting dancer's external self-image by using a mixed reality system. In Proc. Annu. ACM Conf. Extended Abstr. Hum. Factors Comput. Syst. (CHI), 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[39] Ling-An Zeng, Guohong Huang, Gaojie Wu, and Wei-Shi Zheng. Light-t2m: A lightweight and fast model for text-to-motion generation. arXiv preprint arXiv:2412.11193, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Zeyu Zhang, Akide Liu, Qi Chen, Feng chen, Reid Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. InfiniMotion: Mamba boosts memory in Transformer for arbitrary long motion generation. arXiv preprint arXiv:2407.10061, 2024.3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 307 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Zeyu Zhang, Akide Liu, Ian Reid, Richard Hartley, Bohan Zhuang, and Hao Tang. Motion Mamba: Efficient and long sequence motion generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2025. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[42] Chenxi Zheng, Jing Qin, and Shengfeng He. Beat-It: Beatsynchronized multi-condition 3D dance generation. In Proc. Springer Eur. Conf. Comput. Vis. (ECCV), 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 195 + ], + "type": "text", + "content": "[43] Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 315, + 197, + 553, + 261 + ], + "type": "text", + "content": "[44] Haolin Zhuang, Shun Lei, Long Xiao, Weiqin Li, Liyang Chen, Sicheng Yang, Zhiyong Wu, Shiyan Kang, and Helen Meng. GTN-Bailando: Genre consistent long-term 3D dance generation based on pre-trained genre token network. In Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP), 2023. 1, 2, 4, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 315, + 263, + 553, + 307 + ], + "type": "text", + "content": "[45] Wenlin Zhuang, Congyi Wang, Jinxiang Chai, Yangang Wang, Ming Shao, and Siyu Xia. Music2Dance: DanceNet for music-driven dance generation. ACM Trans. Multimedia Comput. Commun. Appl. (TOMM), 2022. 2, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13202" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_content_list.json b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fb7f6a324792176f9b8b2f3ad13b49bf18549374 --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_content_list.json @@ -0,0 +1,1727 @@ +[ + { + "type": "text", + "text": "AlignDiff: Learning Physically-Grounded Camera Alignment via Diffusion", + "text_level": 1, + "bbox": [ + 120, + 128, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liyue Xie $^{1*}$ Jiancong Guo $^{2}$ Ozan Cakmakci $^{2}$ Andre Araujo $^{3}$ Laszlo A. Jeni $^{1}$ Zhiheng Jia $^{2}$", + "bbox": [ + 227, + 179, + 767, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ Google $^{3}$ Google DeepMind", + "bbox": [ + 238, + 232, + 753, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate camera calibration is a fundamental task for 3D perception, especially when dealing with real-world, in-the-wild environments where complex optical distortions are common. Existing methods often rely on pre-rectified images or calibration patterns, which limits their applicability and flexibility. In this work, we introduce a novel framework that addresses these challenges by jointly modeling camera intrinsic and extrinsic parameters using a generic ray camera model. Unlike previous approaches, AlignDiff shifts focus from semantic to geometric features, enabling more accurate modeling of local distortions. We propose AlignDiff, a diffusion model conditioned on geometric priors, enabling the simultaneous estimation of camera distortions and scene geometry. To enhance distortion prediction, we incorporate edge-aware attention, focusing the model on geometric features around image edges, rather than semantic content. Furthermore, to enhance generalizability to real-world captures, we incorporate a large database of ray-traced lenses containing over three thousand samples. This database characterizes the distortion inherent in a diverse variety of lens forms. Our experiments demonstrate that the proposed method significantly reduces the angular error of estimated ray bundles by $\\sim 8.2^{\\circ}$ and overall calibration accuracy, outperforming existing approaches on challenging, real-world datasets.", + "bbox": [ + 89, + 318, + 485, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 727, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate camera calibration, involving the estimation of intrinsic parameters such as focal length and lens distortions and extrinsic parameters such as camera pose, is essential for robust 3D perception in real-world environments. However, prevailing methods typically address these components independently, focusing on either intrinsic calibration or pose estimation while relying on simplified camera models that are insufficient to capture complex real-world opti", + "bbox": [ + 89, + 753, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6801c14600fa42551e31acef456bcafdd21bf9c38d3cc60822f075076ce47f36.jpg", + "image_caption": [ + "Figure 1. AlignDiff is proposed to address common image geometric aberrations with a unified ray camera representation while jointly recovering the camera extrinsics. With groundings on physical camera lens designs, as well as the disassociation of geometric cues from semantic features, it demonstrates an ability to generalize to real video sequences." + ], + "image_footnote": [], + "bbox": [ + 535, + 287, + 885, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal aberrations. This separation is fundamentally limiting, as intrinsic and extrinsic parameters are closely interdependent; inaccuracies in modeling lens distortions propagate directly to pose estimation errors. Consequently, joint optimization of intrinsic and extrinsic parameters is critical to ensure reliable calibration under unconstrained conditions.", + "bbox": [ + 511, + 521, + 906, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent diffusion-based and transformer-powered models, such as PoseDiffusion [55], RayDiffusion [63], and DiffusionSfM [1] have improved calibration accuracy and zero-shot generalizability by conditioning on image features extracted from vision encoders. However, these methods typically focus on high-level, semantic image features rather than on structural cues critical for modeling optical aberrations that directly impact geometry. This is particularly limiting because accurate intrinsic calibration is essential for precise extrinsic calibration: aberrations in the intrinsic parameters can significantly affect the accuracy of extrinsic parameters, making the two tightly interdependent.", + "bbox": [ + 511, + 613, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose AlignDiff, shown Figure 1, a diffusion-based calibration framework, to learn the fine-grained ray profile from video sequences in world space, with a conditioning strategy that better captures fine-grained optical distortions. To shift focus from semantic to structural features, we condition the diffusion model with line embed", + "bbox": [ + 511, + 810, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Google.", + "bbox": [ + 107, + 886, + 341, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26901", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1869ee0aa8bc9062b2afa671dd171f8fd9fc04f9e99abe03d1d13e436662b775.jpg", + "image_caption": [ + "Figure 2. AlignDiff Architecture. We promote learning camera ray profiles in three main steps: geometric cue conditioning from line features, edge-aware attention, and physical camera groundings." + ], + "image_footnote": [], + "bbox": [ + 171, + 90, + 830, + 338 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dings from a line detection network. This approach emphasizes the geometric structures, helping the model prioritize optical aberrations over object content. In addition, we further introduce edge-focused attention to further enhance the model's sensitivity to regions around edges, where aberrations often manifest most prominently.", + "bbox": [ + 88, + 405, + 480, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An additional limitation of prior models is their reliance on simulated optical aberrations derived from predefined models, which often lack the subtlety of real-world aberrations, such as localized or asymmetric profiles. To address this, our framework incorporates real optical profiles from actual lens designs, grounding the model in authentic optical characteristics and boosting its generalizability.", + "bbox": [ + 88, + 496, + 482, + 601 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our framework builds upon existing diffusion-based calibration techniques by incorporating targeted conditioning strategies and real-world optical data, providing a robust solution that accurately captures complex optical aberrations. By effectively addressing intrinsic aberrations, our model lays a solid foundation for accurate extrinsic calibration, making it highly applicable for in-the-wild scenarios requiring precise camera parameters.", + "bbox": [ + 88, + 602, + 482, + 722 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To our knowledge, AlignDiff is the first unified diffusion-based approach addressing complex optical aberrations in joint intrinsic and extrinsic calibration.", + "- By conditioning the diffusion model on line embeddings and incorporating edge-focused attention, our method prioritizes structural over semantic features, achieving generalization to diverse, real-world environments without requiring extensive retraining.", + "- Our approach incorporates authentic optical profiles from actual lens designs, grounding the model in real-world aberrations and improving generalization to natural images." + ], + "bbox": [ + 89, + 723, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 402, + 663, + 419 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Aberration representation for in-the-wild images. Traditional camera calibration methods model intrinsic characteristics, such as focal length and camera center, alongside distortion parameters that describe image geometric aberrations. Early deep learning-based methods [3, 9, 15, 45, 57, 59, 60] enabled calibration from images. Recent diffusion based approaches like PoseDiffusion [55], extend from prior approaches with improved zero-shot performance on wild captures. While effective for rectified images, these methods are limited in their ability to account for local optical aberrations resulting from factors like wear, temperature fluctuations, and lens designs.", + "bbox": [ + 511, + 430, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generic ray camera models were introduced to better capture local geometric distortions [43, 52] by modeling ray bundles across image patches, offering a more detailed representation of pixel deviations compared to the pinhole model. While accurate, they require denser calibration patterns, especially when optimized using methods like PnP [64]. Several recent work [15, 63, 66] adopted the ray representation in deep learning and proved its feasibility. Yet, these prior works remain largely restricted to the pinhole model and did not exploit the expressiveness of ray representation for aberration modeling. The recent blind camera undistortion line of work [29, 32] infers geometric displacements from a single image, often as a preprocessing step to provide rectified images.", + "bbox": [ + 511, + 612, + 908, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work advances these approaches by introducing a unified ray-based framework that models both pinhole projections and local distortions. We directly estimate the ray profile and camera aberrations from a raw monocular sequence, enhancing generalizability in diverse real-world", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26902", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conditions.", + "bbox": [ + 89, + 90, + 166, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Joint Intrinsic and Extrinsic Calibration Classical camera calibration methods often assume a predefined camera model for each type of aberration, relying on calibration patterns and multi-view geometric constraints to estimate camera parameters from a sequence of images [13, 24-26, 33, 54, 55]. Early approaches solved joint intrinsic and extrinsic calibration through point matching across views [2, 3, 11, 22, 38, 49, 65], with refinements summarized in [31]. Structure-from-motion (SfM) techniques later optimized reprojection loss to refine parameters further.", + "bbox": [ + 89, + 106, + 483, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For in-the-wild images, Hold-Geoffroy et al. [17] leveraged DenseNet [18] to estimate parameters such as horizon angle and vertical field of view. Other methods [20, 24] decouple parameter estimation from direct regression, and with integrated semantic or geometric guidance [6, 8, 21, 46]. Further advancements have shown that models pretrained with geometric understanding can act as better feature encoders [5, 15, 23, 40, 42, 45, 51, 53, 61]. Recent methods, including PoseDiffusion [55], DuSt3R [57], RayDiffusion [63] use diffusion and transformer to recover camera parameters or ray profiles directly from images, though they still assume fixed distortion models, limiting generalization to diverse real-world aberrations [31, 32]. The camera parameters are recovered by conditioning on local image features to capture local and cross-view geometric associations. However, these methods can yield residual errors around object contours due to reliance on semantic rather than purely geometric cues [10, 12, 27].", + "bbox": [ + 89, + 257, + 483, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach extends the ray representation without a fixed distortion model, disentangling semantic and geometric cues to reduce contour errors and enhance calibration accuracy. By integrating undistortion and calibration in a unified framework, we increase flexibility and applicability for real-world captures.", + "bbox": [ + 89, + 529, + 483, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. AlignDiff", + "text_level": 1, + "bbox": [ + 89, + 631, + 194, + 648 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/99a9178ef9bf3769274b010533bfbb83eb39affe781463e199b1868f19510f13.jpg", + "image_caption": [ + "Figure 3. The latents are reweighted through edge-attention that aggregates a learned mask to the Query feature to promote information along edges where the geometric cues are more prominent. Following, the multi-view attention further captures features from different views." + ], + "image_footnote": [], + "bbox": [ + 91, + 667, + 209, + 789 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f254a720e0a4f370a862b472a17198001356c9e908487e19dd9fa2be351b9bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 678, + 480, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to recover generic cameras represented as ray", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bundles in world space from a set of $N$ input images $\\{\\mathcal{I}_1,\\mathcal{I}_2,\\mathcal{I}_3,\\dots ,\\mathcal{I}_N\\}$ . Our approach, as presented in Figure 2, captures fine-grained geometric distortions while maintaining compatibility with classical parameterized camera models. This formulation is conceptually similar to works like [63, 66], but extends the ray camera model to account for local distortions (Sec. 3.1). In contrast, previous methods assume a pinhole camera model, neglecting distortions, even though they predict rays for local image patches. We adopt a Diffusion Transformer as the base architecture while enhancing this architecture by introducing geometric cues that guide ray directions based on local geometry, disentangling image perceptual features from distortions (Sec. 3.2). From the predicted rays, we propose DistortionNet to recover the underlying lens geometric aberration that describes the ray profile's deviation from a perspective pinhole camera with the same focal length and field of view (Sec. 3.3). We further discuss the selected optimization objectives in Sec. 3.4.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generic Ray Camera Representation", + "text_level": 1, + "bbox": [ + 511, + 385, + 833, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective of camera calibration is to recover optimal parameters that describe the projection of 3D world points $\\mathcal{X} \\in \\mathbb{R}^3$ to 2D sensor locations $\\pmb{x} \\in \\mathbb{R}^2$ . Our method models this with a ray-based camera model, where each unit of ray $\\pmb{r}_i \\in \\mathbb{R}^6$ is stored on a grid centered on image patches. Rays originate from the camera center $\\pmb{r}_o \\in \\mathbb{R}^3$ , with directions $\\pmb{r}_d \\in \\mathbb{R}^3$ derived by lifting 2D grid points to the camera frame using:", + "bbox": [ + 511, + 407, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {\\boldsymbol {d}} = \\boldsymbol {K} ^ {- 1} \\mathcal {D} _ {\\zeta} (x), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 532, + 903, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{D}_{\\zeta}$ is the aberration function parameterized by $\\zeta$ , and $\\pmb{K}$ is the intrinsic matrix, accounting for geometric aberrations. The ray bundle, transformed to world coordinates by rotation $\\pmb{R}$ and translation $\\pmb{t}$ ,", + "bbox": [ + 511, + 556, + 905, + 614 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {w} = \\boldsymbol {R} ^ {T} \\boldsymbol {r} + \\boldsymbol {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 621, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "represents aberrated cameras in world coordinates, allowing holistic downstream recovery of rays, camera extrinsics, and optical aberrations.", + "bbox": [ + 511, + 643, + 905, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Off-the-shelf image quality is completely characterized by the modulation transfer function (MTF) and distortion. The MTF performance includes all monochromatic and chromatic aberrations. Typically, the driving residual aberration in high-quality off-the-shelf lenses is optical distortion $(W_{311}$ in the wavefront to third order[47]), therefore, in this paper, our ray trace models focus on extracting the optical distortion characteristics from our dataset of lenses.", + "bbox": [ + 511, + 690, + 905, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Geometry-controlled ray diffuser", + "text_level": 1, + "bbox": [ + 511, + 818, + 805, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train a multi-view diffusion model $\\mathcal{M}_{\\phi}$ that takes multiple images of a 3D scene as input and generates the corresponding output camera ray bundles given their geometric aberration cues. Specifically, given $N$ conditional views", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "containing the images and their corresponding geometric aberration latents $\\mathcal{Z}^g$ from a line-segment detection network, the model learns to capture the joint distribution of $N$ target ray bundles $\\mathcal{R}^{tgt} = \\{\\mathcal{R}_1,\\mathcal{R}_2,\\mathcal{R}_3,\\dots ,\\mathcal{R}_N\\}^{tgt}$ from noised ray bundles $\\mathcal{R}^\\epsilon$ :", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathcal {R} ^ {t g t} | \\mathcal {I}, \\mathcal {Z} ^ {g}, \\mathcal {R} ^ {\\epsilon}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 179, + 482, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion model architecture. Diffusion models approximate the data manifold by learning to invert a diffusion process from data to a presumed distribution through a denoising process. Our adopted Denoising Diffusion Probabilistic Model (DDPM) specifically defines the noise distribution to be Gaussian and transitions to the next noising step:", + "bbox": [ + 89, + 208, + 483, + 299 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbb {I}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 311, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{\\beta_{1},\\beta_{2},\\dots,\\beta_{T}\\}$ are variances within the T step nois-ing schedule, and $\\mathbb{I}$ is the identity matrix. Similar to prior works [35, 55, 63], we adopt a DiT [16] as the backbone of the diffusion model, receiving image latent embeddings conditioned on aberration cues. Given a set of sequential images and target ray bundles, the model encodes each image into a latent representation $\\mathcal{Z}_i^c$ through an image feature encoder TIPS [34], as well as a conditioning branch comprised of multiple ResNet [14] blocks. The image latent feature is concatenated to a noised ray bundle of the same dimension. Then, the diffusion model is trained to estimate the joint distribution of the latent-conditioned ray bundle given the conditioning geometric cues. We initialize the model from a DiT model trained for 3D shape generation, with an input resolution of $448\\times 448\\times 3$ . We directly inflate the latent space of the original DiT to connect with the concatenated features of noised rays and image latents, while inheriting the rest of the trained model parameters to leverage the perceptual knowledge.", + "bbox": [ + 89, + 340, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Utilizing geometric cues. The conditioning branch is designed to extract guidance information that describes geometric aberration cues and align it with the denoising features. The conditioning branch is a trained line segment detection network [28]. We observe that the controls maintain a high level of consistency with the denoising features and eliminate the need to be inserted at multiple stages, as also indicated in [41]. We thus integrate the controls at a single middle block into the denoising model by adding them to the denoising features after the cross-normalization. It can serve as a plug-and-play guidance module that provides geometric aberration information about the conditioning views, such that the diffuser model is aware of the levels and types of aberration from geometric cues.", + "bbox": [ + 89, + 628, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Edge Attention. We introduce Edge Attention responsible for parsing features along the edges, as shown in Figure 3, which highlights the edge-aware attention mechanism designed to prioritize geometric cues along image edges. In", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "particular, edge attention takes the input embedding of an image sequence $\\mathcal{Z}^c\\in \\mathbb{R}^{N\\times H\\times W\\times C}$ , allowing it to perform self-attention operations across the edges. We apply a patchification operator with patch size $p\\times p$ to generate patch tokens $\\pmb {t}\\in \\mathbb{R}^{L\\times 3}$ with $L = (N\\times H\\times W / p^2)$ denoting the length of patched tokens. To enable the model to disentangle the geometric information from the semantics, we integrate the edge information $e_{i\\dots N}$ for guidance into the self-attention mechanism:", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {(Q + e ^ {T}) K}{\\sqrt {H}} V) \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 234, + 906, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ is the dimension size of each head. The edge embedding is designed as a soft weighting to highlight the features along the geometrically prominent areas. The image edges are first extracted from the original image sequences using a Canny Edge detector. We then patchify them and apply a 3D convolution, mapping the embedding as $e_i \\in \\mathbb{R}^L$ . Ultimately, we obtain the aggregated features $\\hat{\\mathcal{Z}}_i$ , and we employ a feedforward network as in the original DiT before proceeding to the subsequent blocks.", + "bbox": [ + 511, + 277, + 906, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Aberration profiling with DistortionNet", + "text_level": 1, + "bbox": [ + 511, + 422, + 854, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt warping flow as a unified representation to model geometric aberrations of the image compared to ideal pin-hole cameras, similar to prior blind camera undistortion frameworks [29, 32]. To query the aberration profile, we connect the denoised rays to a network devised to estimate deviations of predicted ray bundles to that of a pinhole camera. The DistortNet $\\mathcal{M}_{\\theta}$ , with a backbone of MAE-Tiny [56], learns the mapping from the predicted set of ray bundles $\\mathcal{R}^{tgt}$ to a set of warping flow maps $\\mathcal{F}^{tgt}$ . The learned flow maps can be applied to the image sequences for image undistortion.", + "bbox": [ + 511, + 444, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Optimization objectives", + "text_level": 1, + "bbox": [ + 511, + 619, + 732, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe the optimization objectives for estimating ray bundles and extracting the distortion map from $N$ images $\\{\\mathcal{I}_1,\\mathcal{I}_2,\\dots ,\\mathcal{I}_N\\}$ . Estimating distorted ray bundles from in-the-wild images is under-constrained and affected by uncertainties from limited overlap and artifacts (e.g., motion blur, lighting changes). To address this, we train the network with a denoising process that reverses noised samples to match the reference:", + "bbox": [ + 511, + 641, + 905, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d e n o i s e} (\\phi) = \\mathbb {E} _ {t, \\boldsymbol {r} _ {o}, \\epsilon} | | \\boldsymbol {r} _ {o} - \\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t) | | ^ {2}, \\qquad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 772, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{r}_o$ is the original ray bundle, $\\mathcal{M}_{\\phi}$ is the denoising model, and $\\pmb{r}_t$ is the sample with Gaussian noise at time $t$ :", + "bbox": [ + 511, + 801, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {r} _ {o} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 842, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,\\mathbb{I})$ . The denoising reverts to a regressive loss, allowing additional geometric constraints.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26904", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80772bc658c3b1f22eef18527756e6b4c1bb5b8e5429969b9ebd6718b325e82f.jpg", + "image_caption": [ + "Figure 4. Recovered aberration profile and undistorted images. From denoised rays in world space, the DistortNet estimates the aberration pattern represented in warping flow. The undistorted images maintain coherent structural distribution compared to aberration-free images." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 500, + 368 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6f44db766740fd90dc34d63f2cbc7e9f77e4d15e9859debead5069d3ffc21549.jpg", + "image_caption": [ + "(e) predicted ray maps with and without edge conditioning" + ], + "image_footnote": [], + "bbox": [ + 524, + 108, + 883, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recovering local camera ray profiles from an image sequence is susceptible to rotational ambiguities, leading to errors in ray bundle and pose estimates. We introduce a ray directional loss to measure angular errors in ray bundles:", + "bbox": [ + 89, + 424, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a n g u l a r}} (\\phi) = \\cos^ {- 1} \\left(\\frac {r _ {o} \\cdot \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right)}{\\left| r _ {o} \\right| \\left| \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right) \\right|}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 489, + 482, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DistortNet is trained to regress backward flow, representing deviations from perspective ray bundles, by comparing each predicted flow map $\\pmb{f}_o$ to the reference set $\\mathcal{F}^{tgt}$ :", + "bbox": [ + 89, + 526, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t o r t}} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {f} _ {o}, \\boldsymbol {r} _ {t}} | | \\boldsymbol {f} _ {o} - \\mathcal {M} _ {\\theta} \\left(\\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t)\\right) | | ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 577, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This way, we recover the camera distortion maps at each image pixel, which can be used to directly unwarp the images and their associated rays.", + "bbox": [ + 89, + 599, + 483, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Physical aberration groundings", + "text_level": 1, + "bbox": [ + 89, + 652, + 369, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the largest available lens database to ray trace and extract distortion characteristics from about 3000 optical systems. We wrote custom scripts that extracted each lens prescription from the LensView database and used commercial raytracing software to extract the optical distortion maps. These systems include camera lenses, lithography lenses, and freeform optics, as shown in Fig 5. These systems produce a diverse set of optical distortion functions that map the object to the image. Detailed description on the lens dataset can be found in Appendix A.", + "bbox": [ + 89, + 674, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since images within each dataset are usually captured with similar equipment sharing identical camera settings, to learn different camera modalities, we augment the sequences upon loading with sampled camera distortion profiles representing shear, barrel, fisheye, and pincushion. We", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/79469f2eb5ef24c798a707a056b8f0eda675c269adeb76394a2e29d0f832e4ca.jpg", + "image_caption": [ + "Figure 5. Ray-traced lens designs are utilized, enabling the accurate simulation of geometric aberrations. The aberrations are encoded as local geometric distortion [19], with a percentage deviation of pixels from their ideal positions on a regular grid." + ], + "image_footnote": [], + "bbox": [ + 519, + 421, + 903, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "further include a set of diverse optics dataset comprised of 3187 patented lens profiles from LensView [4]. These profiles span a wide range of professional and industrial lenses. Figure 5 shows examples from the grounding lenses, along with the distribution of the distortion with respect to the field of view (FOV), Numerical Aperture, and F-number. This augmentation introduces groundings in intrinsic and distortion profiles, addressing the scarcity of camera profiles within the datasets. We assess the out-of-distribution distortions not present in the lens set through experiments on the Aria Digital Twins dataset in Table 1.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26905", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 89, + 223, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first provide the implementation details of the proposed framework and then validate on two real world datasets with different camera modalities. Our method outperforms baseline approaches in both aberrated and non-aberrated images, achieving state-of-the-art results in quantitative and qualitative evaluations.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 213, + 307, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a pre-trained TIPS [34] network as the image feature extractor, which gives high-quality, fine-grained image embeddings. We adopted a DiT [16] as the base structure of the denoising diffuser, and augmented the attention blocks with the proposed edge conditioning attention. We train our diffusion model with $T = 100$ timesteps, with the training taking roughly 3 days on 8 H100 GPUs.", + "bbox": [ + 89, + 234, + 483, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following prior works [33, 55, 63], we use the first camera as the coordinate anchor to define the scene. The scene is rescaled such that the first camera has a unit from translation and is rotated to grant the first camera identity rotation. In this way, all the following predicted cameras would use the first camera as a reference to express their relative orientations.", + "bbox": [ + 89, + 342, + 483, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 455, + 370, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We chose CO3D [44] as the primary training dataset for the generation of ray maps. The dataset consists of roughly $37\\mathrm{k}360^{\\circ}$ videos of common household objects from 51 MS-COCO categories. The dataset is annotated with COLMAP [49, 50] camera poses and intrinsics, with the frames undistorted with the estimated coefficients. Each video spans over an average 200 frames. The dataset provides a diverse profile of common objects sequences, making it suitable for training the network. In order to enhance the variety of training scenarios, we add MegaDepth [30] and TartanAir [58] as our additional training data.", + "bbox": [ + 89, + 477, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Secondly, we evaluate the inference performance in Aria Digital Twins [39] that comprises 200 sequences and 400 min of videos that capture the daily activities of the common household using egocentric cameras. Its camera trajectories were provided by Optitrack [36] system and IMU sensor data on the Aria glass, further optimized by matching the two sources of estimated trajectories. The camera calibration parameters are provided by fitting a Kannala Brandt and Fisheye radial-tangential thin film model with the estimated trajectories. We use the same division as in [39], with a subset of 50 tests. With the two chosen datasets, we showcase the model's capabilities of estimating the ray maps from both common household videos and its ability to generalize to recent AR egocentric captures.", + "bbox": [ + 89, + 645, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. We evaluated spare view predictions for recovered poses and the accuracy of the camera ray profile compared to the reference distorted camera, without", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2aa061e1f3b1d279c41b06c9a76d815d5bb179c54f354d027f7ab109864294f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 92, + 622, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6ebdda4b96ca1c4b584103dbcc787625ab5aa7337ad65b3365bbc88a63bbb14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 154, + 622, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/90b6e6e4dac0647685df9de18e11ffe7110160481c8247dc550ae8b6c72ce955.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 540, + 214, + 622, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d527e0c82e6570af51fbfee17fdafdfddd9a5c8536fb6236c4b57216eb38efed.jpg", + "image_caption": [ + "input" + ], + "image_footnote": [], + "bbox": [ + 540, + 275, + 622, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ede4262645e380177b8da2ff7b63094b8c22e805b43a4e1426c87a8ddd8d80a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 92, + 707, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ad7d1ab580bffbbf343647862cc2175044c73d68d2d9c11699f34174740c7ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 154, + 707, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e51077ac9c4b3784915e143f89019b9790fe248a62e9d6de4e2124e051e1b08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 214, + 707, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25334b9a2a1f07b39dabe9728ac522e5c76bd6915cc01846bcaf37003e91682c.jpg", + "image_caption": [ + "" + ], + "image_footnote": [], + "bbox": [ + 625, + 275, + 707, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df345b5980c9ea2a97fe90878b16e2a8b1408ad85e97b9858c57fa64f5128f5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 92, + 794, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/daa4e4e443d80cb87179a97255ef9abac301fa224329e32ba5aa123c0f4fd8e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 154, + 794, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67afd2ed9f2666931e534bcb5e0d679dfc7395fe32dccaee519820340a0a5000.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 214, + 794, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1714555cb57e0eb57c3f0a8daa42c6169172adf96f22e1f1159fa2d08e453008.jpg", + "image_caption": [ + "RayDiffusion" + ], + "image_footnote": [], + "bbox": [ + 710, + 275, + 794, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d56b51ac1b33107d8cacba30d8f3ab511f6d490fcdd20b13aa20c0497267eacc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 92, + 880, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8648f92ffdc827cc0d64d867bf936874ac44d03cb5594ab8ef936f87ff0a11a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 154, + 880, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06db193a3419859fdc0b76a010df49e7ad7e14274bb93df0fabeb7376c1d6ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 214, + 880, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43685977d291fe60d2696ffe701511f5771cf0d9f3d4c5c25e6a605030aad3d3.jpg", + "image_caption": [ + "ours", + "Figure 6. Generalization to out-of-distribution camera from Aria Digital Twins. Our approach produces better results compared to [63]. See supplementary for quantitative inference results." + ], + "image_footnote": [], + "bbox": [ + 797, + 275, + 880, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "employing the DistortionNet. In accordance with [63], we randomly sampled $N$ images for an evaluation of $N$ . Each reported accuracy is averaged over 5 runs to reduce stochasticity.", + "bbox": [ + 511, + 415, + 905, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Rotation accuracy. We first compute the relative rotations between each pair of cameras for both predicted and ground-truth poses. The errors are computed for the pairwise relative rotations and reported for the percentage of deviations less than $15^{\\circ}$ .", + "bbox": [ + 511, + 476, + 905, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera center accuracy. Using the fitted ground truth and predicted poses, we align the cameras and compute the distance from the predicted camera center to the ground truth correspondence. The fraction of distances within $10\\%$ of the scene scale is reported.", + "bbox": [ + 511, + 551, + 905, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Angular error. With the recovered ray bundles, we compute the distance of the predicted rays to the ground truth correspondences. The mean angular error in degrees is reported.", + "bbox": [ + 511, + 627, + 905, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation", + "text_level": 1, + "bbox": [ + 513, + 681, + 635, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline comparisons. In Table 1, we present the mean ray angular error measurements for the distorted CO3D dataset, demonstrating the effectiveness of our approach compared to existing methods. Our results indicate that our method excels at extracting precise, arbitrary ray profiles even in the absence of a prior camera model. This capability is particularly significant given the challenges of distorted sequences. Notably, using the same ray representation as RayDiffusion, our approach consistently surpasses its performance, where the diffusion model depends solely on image-based features. The experiments reveal that our method achieves a more robust disentanglement from local image content, effectively isolating structural cues from object-specific fea", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26906", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f162fb3b57bafaa3368c34c51eda88e07804ca8f8dfe81df833ace40efe11214.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of camera angular error on CO3D and Aria Digital Twins datasets. Averaged angular error (degrees) across varying numbers of input images." + ], + "table_footnote": [], + "table_body": "
CO3D – Seen Categories
Method2345678
RayRegression [63]18.420.621.721.920.323.023.7
RayDiffusion [63]9.69.710.28.79.411.913.2
AlignDiff (Ours)3.43.12.83.23.03.64.1
CO3D – Unseen Categories
RayRegression [63]24.122.521.624.823.925.126.2
RayDiffusion [63]11.410.814.112.315.816.619.4
AlignDiff (Ours)4.65.25.85.76.46.88.2
Aria Digital Twins
RayRegression [63]34.638.236.540.841.240.943.7
RayDiffusion [63]28.429.134.737.240.442.345.6
AlignDiff (Ours)15.413.814.220.621.324.624.9
", + "bbox": [ + 93, + 142, + 480, + 319 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7925d4679395274aef488008009416dbddd8ffedddb44391b9781b722d4deeb5.jpg", + "table_caption": [ + "Table 2. Camera Rotation and Camera Center Accuracy on Distorted CO3D dataset. Experiments using geometrically aberrated videos, comparing recent methods against AlignDiff." + ], + "table_footnote": [], + "table_body": "
Rotation @ 15°
number of Images2345678
Seen ClsRelPose [62]62.024.034.024.025.033.032.0
RelPose++ [33]72.573.473.674.775.076.075.7
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]74.080.085.382.484.882.586.6
AlignDiff (Ours)90.691.191.391.892.492.793.1
Unseen ClsRelPose [62]61.027.037.026.030.026.025.0
RelPose++ [33]61.460.863.064.165.765.765.4
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]69.972.475.376.276.477.178.5
AlignDiff (Ours)83.684.884.985.285.786.186.1
Camera Center @ 0.1
Seen ClsRelPose++ [33]10085.279.074.370.568.666.0
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10073.472.571.171.070.670.2
AlignDiff (Ours)10092.590.788.687.285.884.6
Unseen ClsRelPose++ [33]10066.457.051.348.944.544.1
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10074.871.670.267.366.762.5
AlignDiff (Ours)10086.384.280.575.273.070.1
", + "bbox": [ + 91, + 388, + 500, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tures. This disentanglement minimizes residual errors tied to object appearance, allowing structural details to be more accurately emphasized.", + "bbox": [ + 89, + 657, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we report metrics for camera rotation accuracy and camera center accuracy, with aberrated images in Table 2 and rectified images in Table 3. Our approach consistently outperforms existing methods in estimating both camera position and orientation under aberrated conditions, confirming its resilience to challenging visual distortions. For rectified images, our conditioning method does not simply maintain performance but rather enhances it by recognizing and compensating for residual aberrations. This capability allows our approach to improve predictions even on images with minimal distortion, achieving a high level of precision and adaptability in handling diverse image qualities.", + "bbox": [ + 89, + 704, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f389019e5bd57d70edf16e8e9ae96e1830833582183a725c924688e7ccaed36.jpg", + "table_caption": [ + "Table 3. Camera Rotation and Camera Center Accuracy on Rectified CO3D dataset. AlignDiff identifies subtle local distortions, achieving improvements over previous methods." + ], + "table_footnote": [], + "table_body": "
Rotation @ 15°
number of Images2345678
Seen ClsCOLMAP [48]30.728.426.526.827.028.130.6
RelPose [62]56.056.557.057.257.257.357.2
PoseDiffusion [55]75.776.476.877.478.078.778.8
RelPose++ [33]81.882.884.184.784.985.385.5
RayRegression [63]88.888.788.789.089.489.389.2
RayDiffusion [63]91.892.492.692.993.193.393.3
AlignDiff (Ours)92.692.392.793.093.693.493.5
Unseen ClsCOLMAP [48]34.531.831.031.732.735.038.5
RelPose [62]48.647.548.148.348.448.448.3
PoseDiffusion [55]63.264.264.265.766.267.067.7
RelPose++ [33]69.871.171.972.873.874.474.9
RayRegression [63]79.079.680.681.481.381.981.9
RayDiffusion [63]83.585.686.386.987.287.588.1
AlignDiff (Ours)86.386.687.487.788.588.789.2
Camera Center @ 0.1
Seen ClsCOLMAP [48]10034.523.818.915.614.515.0
RelPose [62]10076.566.962.459.458.056.5
PoseDiffusion [55]10077.569.765.963.762.861.9
RelPose++ [33]10085.078.074.271.970.368.8
RayRegression [63]10091.785.782.179.877.976.2
RayDiffusion [63]10094.290.587.886.285.084.1
AlignDiff (Ours)10093.791.887.587.185.283.9
Unseen ClsCOLMAP [48]10036.025.520.017.917.619.1
PoseDiffusion [55]10063.650.545.743.041.239.9
RelPose++ [33]10070.658.853.450.447.846.6
RayRegression [63]10083.775.670.867.465.363.9
RayDiffusion [63]10087.781.177.074.172.471.4
AlignDiff (Ours)10088.281.079.377.572.671.0
", + "bbox": [ + 516, + 142, + 903, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e19b2f7d78cfd44b9b4586f1b39099ba916ee72824fc9b522b81963c63aea37.jpg", + "table_caption": [ + "Table 4. Camera Rotation and Camera Center Accuracy on Aria Digital Twins dataset." + ], + "table_footnote": [], + "table_body": "
Rotation @ 15°
Number of Images2345678
Unseen ClsRelPose [62]59.632.738.031.228.627.426.0
RelPose++ [33]62.861.362.662.863.264.964.8
PoseDiffusion [55]75.474.874.175.875.275.976.2
RayDiffusion [63]75.276.378.979.279.680.080.2
AlignDiff (Ours)81.482.782.884.384.184.685.5
Camera Center @ 0.1
Unseen ClsRelPose++ [33]10070.356.152.848.646.444.7
PoseDiffusion [55]10071.857.454.651.852.248.5
RayDiffusion [63]10075.969.261.860.455.651.1
AlignDiff (Ours)10078.370.064.462.758.152.9
", + "bbox": [ + 517, + 529, + 903, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis of recovered cameras. We compare the generated cameras with the ground truth cameras in world space, as shown in Figure 7. The predicted ray moments and angular error from ground truth ray directions are also shown. Remarkably, the generated cameras exhibit little rotational difference from the reference cameras. While there are observed to have noticeable translation errors, we argue that this is due to scale ambiguities of the captures. The visualized ray moments present a smooth profile with very low angular errors. We further compare the Aria Digital Twins dataset for the inferred rays in vector form, shown in Figure 6. These further demonstrate that our approach can achieve robust and accurate ray profiles through camera", + "bbox": [ + 511, + 704, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26907", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/27297decd2a4d5fd8c36c314c8ed606f4d55262b7582623820a757ff1f646760.jpg", + "table_caption": [ + "Table 5. Ablated framework design evaluated with averaged ray angular error." + ], + "table_footnote": [], + "table_body": "
Methods2345678
Baseline8.89.510.49.212.412.713.9
AlignDiff (w/o angular loss)5.86.18.87.39.611.314.7
AlignDiff (w/o line conditioning)6.78.49.89.910.310.112.8
AlignDiff (w/o edge-attention)5.37.17.49.210.811.310.5
AlignDiff (w/o optics grounding)5.26.96.87.78.49.111.3
AlignDiff4.65.25.85.76.46.88.2
", + "bbox": [ + 91, + 121, + 480, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5734de060f1eca43989a6d99eaf39453160d32b0bf9e2e108a8ecbb046e1c69b.jpg", + "table_caption": [ + "Table 6. Evaluations on different image latent encoder choices." + ], + "table_footnote": [], + "table_body": "
Methods2345678
DINOv2-s/14 [37]8.28.18.610.49.810.112.5
DINOv2-g/14 [37]4.56.46.36.66.17.89.6
TIPS-g/14 [34]4.65.25.85.76.46.88.2
", + "bbox": [ + 91, + 237, + 500, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c8344e3c869e1a55aad0161f1562c8805c73c421908518388b95933041fd532e.jpg", + "image_caption": [ + "Figure 7. Recovered cameras from aberrated image sequences. Predicted and expected cameras are visualized in solid and dashed lines, respectively. The cameras align with the rotational orientation of the expected counterparts. The ray moment and residual errors show low levels of deviation from reference rays." + ], + "image_footnote": [], + "bbox": [ + 93, + 305, + 480, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "conditioning. Detailed quantitative evaluation on the Aria Digital Twins dataset are included in Appendix D.", + "bbox": [ + 89, + 595, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Undistortion using predicted flow. We conduct image undistortion with the predicted aberration map, as shown in Figure 4. Since we adopt a unified representation for aberrations, the radial fisheye camera model's vignetting effect leads to mapping artifacts. The predicted aberration map overall correctly captures the image distortions and reliably remaps the images regardless of the aberration types.", + "bbox": [ + 89, + 627, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study. We perform an ablation study on CO3D data to evaluate the impact of each conditioning module, enhanced geometric guidance, and additional losses, as shown in Table 5. For each comparison, we remove one conditioning module from the framework. The results show that all modules significantly improve performance, and latent conditioning of the line segment has the greatest effect. Figure 4 visualizes the ray residuals with and without aberration conditioning, further validating its effectiveness. Additionally, we test the angular loss, finding that our localized ray angular loss yields notable improvements.", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our multiview diffusion approach implicitly incorporates temporal consistency by explicit temporal losses, we chose to handle temporal consistency as a post-processing step to keep the model lightweight and directly applicable in real-world deployments. This approach reduces computational overhead while still providing meaningful consistency improvements, leaving more complex temporal modeling for future work. This is based on the premise that camera profiles remain generally consistent over short periods, barring changes due to factors like camera sensor temperature fluctuations. To test this, we implemented a post-processing step that rejects low-confidence pose estimations, which reduced the average ray angular error from $8.20^{\\circ}$ to $8.06^{\\circ}$ over 8 frames. Furthermore, enforcing temporal consistency on local camera ray profiles using reprojection mean squared error loss yielded an average reduction of $0.06^{\\circ}$ in angular error across the frames. We plan to integrate these enhancements, along with outlier rejection and a frame selection mechanism, into our future work.", + "bbox": [ + 511, + 90, + 903, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluations on image representation models. The image sequence encoder is designed to generate dense global embeddings that capture perceptual, geometric, and semantic representations. Among the available models, DINOv2 and TIPS are well suited for extracting high-quality dense features. We conduct an ablation study to evaluate these models' effectiveness in querying camera ray profiles. As shown in Table 6, the TIPS-g/14 model achieves the lowest angular errors on CO3D, with DINOv2-g/14 as a close competitor.", + "bbox": [ + 511, + 393, + 906, + 544 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 556, + 633, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce AlignDiff, a unified calibration framework for real-world optical aberrations, enabling the recovery of camera ray bundles, extrinsics, and aberration profiles from multi-view image sequences using simple conditioning techniques. First, we propose structural conditioning to separate image geometry from semantic features, promoting emphasis on structural cues for accurate ray profile estimation. Second, we aggregate features along edges, ensuring that high-quality image embeddings are paired with effective guidance to avoid semantic interference. Finally, we incorporate real-world lens profiles into training, grounding AlignDiff in actual camera designs. These contributions collectively enhance generalization and accuracy across diverse distortion profiles, as demonstrated by consistent improvements over strong baselines.", + "bbox": [ + 511, + 582, + 903, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While AlignDiff performs well in controlled indoor settings, ray estimation may become unstable in arbitrary outdoor scenes due to scale ambiguity. To enhance robustness in such cases, integrating geometric priors like optical flow or combining traditional camera parameter estimation with network inference may be promising for future directions.", + "bbox": [ + 511, + 809, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26908", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Diffusionsfm: Predicting structure and motion via ray origin and endpoint diffusion, 2024. [Online; accessed 2025-03-06]. 1", + "[2] Tommaso Cavallari, Stuart Golodetz, Nicholas A. Lord, Julien Valentin, Victor A. Prisacariu, Luigi Di Stefano, and Philip H.S. Torr. Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2465-2477, 2020. 3", + "[3] Shuai Chen, Xinghui Li, Zirui Wang, and Victor Adrian Prisacariu. DFNet: Enhance Absolute Pose Regression with Direct Feature Matching. 2022. 2, 3", + "[4] Lambda Research Corp. Lensview optical design database. [Online; accessed 2025-03-07]. 5", + "[5] Andrei Cramariuc, Aleksandar Petrov, Rohit Suri, Mayank Mittal, Roland Siegwart, and Cesar Cadena. Learning camera miscalibration detection. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 4997-5003, 2020. 3", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-Supervised Interest Point Detection and Description. 2017. 3", + "[7] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 1", + "[8] Satoshi Fujimoto and Nobutomo Matsunaga. Deep feature-based rgb-d odometry using superpoint and superglue. *Proceeding Computer Science*, 227:1127–1134, 2023. 3", + "[9] Fabrizio Gentile, Crescenzo Tortora, Giovanni Covone, Léon V. E. Koopmans, Rui Li, Laura Leuzzi, and Nicola R. Napolitano. *LeMoN: Lens Modelling with Neural networks* - I. Automated modelling of strong gravitational lenses with Bayesian Neural Networks. 2022. 2", + "[10] Behnam Gholami, Mostafa El-Khamy, and Kee-Bong Song. Latent feature disentanglement for visual domain generalization. IEEE Transactions on Image Processing, 2023. 3", + "[11] Diandian Guo, Deng-Ping Fan, Tongyu Lu, Christos Sakaridis, and Luc Van Gool. Vanishing-point-guided video semantic segmentation of driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3544-3553, 2024. 3", + "[12] Lanqing Guo, Renjie Wan, Wenhan Yang, Alex C Kot, and Bihan Wen. Cross-image disentanglement for low-light enhancement in real world. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2550-2563, 2023. 3", + "[13] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xiankang He, Guangkai Xu, Bo Zhang, Hao Chen, Ying Cui, and Dongyan Guo. DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation. 2024. 2, 3", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Technical report. 4, 6", + "[17] Yannick Hold-Geoffroy, Dominique Piche-Meunier, Kalyan Sunkavalli, Jean-Charles Bazin, François Rameau, and Jean-François Lalonde. A perceptual measure for deep single image camera and lens calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10603-10614, 2023. 3", + "[18] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 3", + "[19] ISO 17850:2015(en). Photography — Digital cameras — Geometric distortion (GD) measurements. Standard, International Organization for Standardization, 2015. 5", + "[20] Y. Y. Jau, R. Zhu, H. Su, and M. Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4950-4957, 2020. 3", + "[21] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable Feature Matching with Foundation Model Guidance. Technical report. 3", + "[22] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-View Object Reconstruction with Unknown Categories and Camera Poses. 2022. 3", + "[23] Yang Jiao, Trac D. Tran, and Guangming Shi. Effiscene: Efficient per-pixel rigidity inference for unsupervised joint learning of optical flow, depth, camera pose and motion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5538-5547, 2021. 3", + "[24] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F. Fouhey. Perspective fields for single image camera calibration. In CVPR, 2023. 3", + "[25] Alex Kendall, Matthew Koichi Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. 2015 IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015.", + "[26] Julius Kümmerle and Tilman Kühner. Unified intrinsic and extrinsic camera and lidar calibration under uncertainties. In 2020 IEEE International conference on Robotics and Automation (ICRA), pages 6028-6034. IEEE, 2020. 3", + "[27] Jinwoo Lee, Hyunsung Go, Hyunjoon Lee, Sunghyun Cho, Minhyuk Sung, and Junho Kim. Ctrl-c: Camera calibration transformer with line-classification. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16208-16217, 2021. 3", + "[28] Hao Li, Huai Yu, Jinwang Wang, Wen Yang, Lei Yu, and Sebastian Scherer. Ulsd: Unified line segment detection across pinhole, fisheye, and spherical cameras. ISPRS Journal of Photogrammetry and Remote Sensing, 178:187-202, 2021." + ], + "bbox": [ + 516, + 92, + 906, + 886 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4", + "bbox": [ + 547, + 888, + 557, + 897 + ], + "page_idx": 8 + }, + { + "type": "footer", + "text": "26909", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4855-4864, 2019. 2, 4", + "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[31] Kang Liao, Lang Nie, Shujuan Huang, Chunyu Lin, Jing Zhang, Yao Zhao, Moncef Gabbouj, and Dacheng Tao. Deep learning for camera calibration and beyond: A survey. arXiv preprint arXiv:2303.10559, 2023. 3", + "[32] Kang Liao, Zongsheng Yue, Zhonghua Wu, and Chen Change Loy. Mowa: Multiple-in-one image warping model. ArXiv, abs/2404.10716, 2024. 2, 3, 4", + "[33] Amy Lin, Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. RelPose++: Recovering 6D Poses from Sparse-view Observations. 2023. 3, 6, 7", + "[34] Kevis-Kokitsi Maninis, Kaifeng Chen, Soham Ghosh, Arjun Karpur, Koert Chen, Ye Xia, Bingyi Cao, Daniel Salz, Guangxing Han, Jan Dlabal, et al. Tips: Text-image pretraining with spatial awareness. arXiv preprint arXiv:2410.16512, 2024. 4, 6, 8", + "[35] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In Advances in Neural Information Processing Systems, pages 67960–67971. Curran Associates, Inc., 2023. 4", + "[36] Inc. DBA OptiTrack NaturalPoint. Optitrack system, 2024. 6", + "[37] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 8", + "[38] Linfei Pan, Marc Pollefeys, and Viktor Larsson. Camera pose estimation using implicit distortion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12819-12828, 2022. 3", + "[39] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 6", + "[40] Chethan M. Parameshwara, Gokul Hari, Cornelia Fermüller, Nitin J. Sanket, and Yiannis Aloimonos. Diffposenet: Direct differentiable camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6845-6854, 2022. 3", + "[41] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. ControlNeXt: Powerful and Efficient Control for Image and Video Generation. 2024. 4" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Martins, and Erickson R. Nascimento. XFeat: Accelerated Features for Lightweight Image Matching. 2024. 3", + "[43] Srikumar Ramalingam and Peter Sturm. A Unifying Model for Camera Calibration: A Unifying Model for Camera Calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence: A Unifying Model for Camera Calibration. (7): 1309-1319, 2017. 2", + "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction. Technical report. 6", + "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. 2018. 2, 3", + "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3", + "[47] José Sasán. Introduction to aberrations in optical imaging systems. Cambridge University Press, 2012. 3", + "[48] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. Technical report. 7", + "[49] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 6", + "[50] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6", + "[51] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In arXiv, 2024. 3", + "[52] Peter Sturm and Srikumar Ramalingam. A Generic Concept for Camera Calibration. pages 1-13, 2004. 2", + "[53] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement. Technical report. 3", + "[54] Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys. GeoCalib: Single-image Calibration with Geometric Optimization. In ECCV, 2024. 3", + "[55] Jianyuan Wang, Christian Rupprecht, and David Novotny. PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment. 2023. 1, 2, 3, 4, 6, 7", + "[56] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 4", + "[57] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. 2023. 2, 3", + "[58] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. 2020. 6" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26910", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Scott Workman, Connor Greenwell, Menghua Zhai, Ryan Baltenberger, and Nathan Jacobs. Deepfocal: A method for direct focal length estimation. In 2015 IEEE International Conference on Image Processing (ICIP), pages 1369-1373, 2015. 2", + "[60] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Innovating Real Fisheye Image Correction with Dual Diffusion Architecture. Technical report. 2", + "[61] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3", + "[62] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. page 592-611, Berlin, Heidelberg, 2022. Springer-Verlag. 7", + "[63] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as Rays: Pose Estimation via Ray Diffusion. 2024. 1, 2, 3, 4, 6, 7", + "[64] Yinqiang Zheng, Yubin Kuang, Shigeki Sugimoto, Kalle Åström, and Masatoshi Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In 2013 IEEE International Conference on Computer Vision, pages 2344-2351, 2013. 2", + "[65] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[66] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a Wild Camera: In-the-Wild Monocular Camera Calibration. 2023. 2, 3" + ], + "bbox": [ + 91, + 92, + 480, + 542 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26911", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_model.json b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_model.json new file mode 100644 index 0000000000000000000000000000000000000000..06989ef3c7e06d956ec7b0c042ace445824d2467 --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_model.json @@ -0,0 +1,2499 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.877, + 0.153 + ], + "angle": 0, + "content": "AlignDiff: Learning Physically-Grounded Camera Alignment via Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.18, + 0.768, + 0.218 + ], + "angle": 0, + "content": "Liyue Xie\\(^{1*}\\) Jiancong Guo\\(^{2}\\) Ozan Cakmakci\\(^{2}\\) Andre Araujo\\(^{3}\\) Laszlo A. Jeni\\(^{1}\\) Zhiheng Jia\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.233, + 0.754, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\) Carnegie Mellon University \\(^{2}\\) Google \\(^{3}\\) Google DeepMind" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.699 + ], + "angle": 0, + "content": "Accurate camera calibration is a fundamental task for 3D perception, especially when dealing with real-world, in-the-wild environments where complex optical distortions are common. Existing methods often rely on pre-rectified images or calibration patterns, which limits their applicability and flexibility. In this work, we introduce a novel framework that addresses these challenges by jointly modeling camera intrinsic and extrinsic parameters using a generic ray camera model. Unlike previous approaches, AlignDiff shifts focus from semantic to geometric features, enabling more accurate modeling of local distortions. We propose AlignDiff, a diffusion model conditioned on geometric priors, enabling the simultaneous estimation of camera distortions and scene geometry. To enhance distortion prediction, we incorporate edge-aware attention, focusing the model on geometric features around image edges, rather than semantic content. Furthermore, to enhance generalizability to real-world captures, we incorporate a large database of ray-traced lenses containing over three thousand samples. This database characterizes the distortion inherent in a diverse variety of lens forms. Our experiments demonstrate that the proposed method significantly reduces the angular error of estimated ray bundles by \\(\\sim 8.2^{\\circ}\\) and overall calibration accuracy, outperforming existing approaches on challenging, real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.728, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Accurate camera calibration, involving the estimation of intrinsic parameters such as focal length and lens distortions and extrinsic parameters such as camera pose, is essential for robust 3D perception in real-world environments. However, prevailing methods typically address these components independently, focusing on either intrinsic calibration or pose estimation while relying on simplified camera models that are insufficient to capture complex real-world opti" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.288, + 0.887, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Figure 1. AlignDiff is proposed to address common image geometric aberrations with a unified ray camera representation while jointly recovering the camera extrinsics. With groundings on physical camera lens designs, as well as the disassociation of geometric cues from semantic features, it demonstrates an ability to generalize to real video sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.613 + ], + "angle": 0, + "content": "cal aberrations. This separation is fundamentally limiting, as intrinsic and extrinsic parameters are closely interdependent; inaccuracies in modeling lens distortions propagate directly to pose estimation errors. Consequently, joint optimization of intrinsic and extrinsic parameters is critical to ensure reliable calibration under unconstrained conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Recent diffusion-based and transformer-powered models, such as PoseDiffusion [55], RayDiffusion [63], and DiffusionSfM [1] have improved calibration accuracy and zero-shot generalizability by conditioning on image features extracted from vision encoders. However, these methods typically focus on high-level, semantic image features rather than on structural cues critical for modeling optical aberrations that directly impact geometry. This is particularly limiting because accurate intrinsic calibration is essential for precise extrinsic calibration: aberrations in the intrinsic parameters can significantly affect the accuracy of extrinsic parameters, making the two tightly interdependent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.9 + ], + "angle": 0, + "content": "We propose AlignDiff, shown Figure 1, a diffusion-based calibration framework, to learn the fine-grained ray profile from video sequences in world space, with a conditioning strategy that better captures fine-grained optical distortions. To shift focus from semantic to structural features, we condition the diffusion model with line embed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.343, + 0.901 + ], + "angle": 0, + "content": "*Work done during an internship at Google." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26901" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.172, + 0.092, + 0.831, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.351, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. AlignDiff Architecture. We promote learning camera ray profiles in three main steps: geometric cue conditioning from line features, edge-aware attention, and physical camera groundings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.406, + 0.482, + 0.496 + ], + "angle": 0, + "content": "dings from a line detection network. This approach emphasizes the geometric structures, helping the model prioritize optical aberrations over object content. In addition, we further introduce edge-focused attention to further enhance the model's sensitivity to regions around edges, where aberrations often manifest most prominently." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.497, + 0.483, + 0.602 + ], + "angle": 0, + "content": "An additional limitation of prior models is their reliance on simulated optical aberrations derived from predefined models, which often lack the subtlety of real-world aberrations, such as localized or asymmetric profiles. To address this, our framework incorporates real optical profiles from actual lens designs, grounding the model in authentic optical characteristics and boosting its generalizability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.603, + 0.483, + 0.723 + ], + "angle": 0, + "content": "In summary, our framework builds upon existing diffusion-based calibration techniques by incorporating targeted conditioning strategies and real-world optical data, providing a robust solution that accurately captures complex optical aberrations. By effectively addressing intrinsic aberrations, our model lays a solid foundation for accurate extrinsic calibration, making it highly applicable for in-the-wild scenarios requiring precise camera parameters." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.724, + 0.483, + 0.768 + ], + "angle": 0, + "content": "- To our knowledge, AlignDiff is the first unified diffusion-based approach addressing complex optical aberrations in joint intrinsic and extrinsic calibration." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.767, + 0.483, + 0.842 + ], + "angle": 0, + "content": "- By conditioning the diffusion model on line embeddings and incorporating edge-focused attention, our method prioritizes structural over semantic features, achieving generalization to diverse, real-world environments without requiring extensive retraining." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Our approach incorporates authentic optical profiles from actual lens designs, grounding the model in real-world aberrations and improving generalization to natural images." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.724, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.404, + 0.665, + 0.42 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Aberration representation for in-the-wild images. Traditional camera calibration methods model intrinsic characteristics, such as focal length and camera center, alongside distortion parameters that describe image geometric aberrations. Early deep learning-based methods [3, 9, 15, 45, 57, 59, 60] enabled calibration from images. Recent diffusion based approaches like PoseDiffusion [55], extend from prior approaches with improved zero-shot performance on wild captures. While effective for rectified images, these methods are limited in their ability to account for local optical aberrations resulting from factors like wear, temperature fluctuations, and lens designs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Generic ray camera models were introduced to better capture local geometric distortions [43, 52] by modeling ray bundles across image patches, offering a more detailed representation of pixel deviations compared to the pinhole model. While accurate, they require denser calibration patterns, especially when optimized using methods like PnP [64]. Several recent work [15, 63, 66] adopted the ray representation in deep learning and proved its feasibility. Yet, these prior works remain largely restricted to the pinhole model and did not exploit the expressiveness of ray representation for aberration modeling. The recent blind camera undistortion line of work [29, 32] infers geometric displacements from a single image, often as a preprocessing step to provide rectified images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our work advances these approaches by introducing a unified ray-based framework that models both pinhole projections and local distortions. We directly estimate the ray profile and camera aberrations from a raw monocular sequence, enhancing generalizability in diverse real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.168, + 0.105 + ], + "angle": 0, + "content": "conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.257 + ], + "angle": 0, + "content": "Joint Intrinsic and Extrinsic Calibration Classical camera calibration methods often assume a predefined camera model for each type of aberration, relying on calibration patterns and multi-view geometric constraints to estimate camera parameters from a sequence of images [13, 24-26, 33, 54, 55]. Early approaches solved joint intrinsic and extrinsic calibration through point matching across views [2, 3, 11, 22, 38, 49, 65], with refinements summarized in [31]. Structure-from-motion (SfM) techniques later optimized reprojection loss to refine parameters further." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.529 + ], + "angle": 0, + "content": "For in-the-wild images, Hold-Geoffroy et al. [17] leveraged DenseNet [18] to estimate parameters such as horizon angle and vertical field of view. Other methods [20, 24] decouple parameter estimation from direct regression, and with integrated semantic or geometric guidance [6, 8, 21, 46]. Further advancements have shown that models pretrained with geometric understanding can act as better feature encoders [5, 15, 23, 40, 42, 45, 51, 53, 61]. Recent methods, including PoseDiffusion [55], DuSt3R [57], RayDiffusion [63] use diffusion and transformer to recover camera parameters or ray profiles directly from images, though they still assume fixed distortion models, limiting generalization to diverse real-world aberrations [31, 32]. The camera parameters are recovered by conditioning on local image features to capture local and cross-view geometric associations. However, these methods can yield residual errors around object contours due to reliance on semantic rather than purely geometric cues [10, 12, 27]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Our approach extends the ray representation without a fixed distortion model, disentangling semantic and geometric cues to reduce contour errors and enhance calibration accuracy. By integrating undistortion and calibration in a unified framework, we increase flexibility and applicability for real-world captures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.632, + 0.196, + 0.65 + ], + "angle": 0, + "content": "3. AlignDiff" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.669, + 0.21, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.213, + 0.679, + 0.482, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.801, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Figure 3. The latents are reweighted through edge-attention that aggregates a learned mask to the Query feature to promote information along edges where the geometric cues are more prominent. Following, the multi-view attention further captures features from different views." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We aim to recover generic cameras represented as ray" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "bundles in world space from a set of \\(N\\) input images \\(\\{\\mathcal{I}_1,\\mathcal{I}_2,\\mathcal{I}_3,\\dots ,\\mathcal{I}_N\\}\\). Our approach, as presented in Figure 2, captures fine-grained geometric distortions while maintaining compatibility with classical parameterized camera models. This formulation is conceptually similar to works like [63, 66], but extends the ray camera model to account for local distortions (Sec. 3.1). In contrast, previous methods assume a pinhole camera model, neglecting distortions, even though they predict rays for local image patches. We adopt a Diffusion Transformer as the base architecture while enhancing this architecture by introducing geometric cues that guide ray directions based on local geometry, disentangling image perceptual features from distortions (Sec. 3.2). From the predicted rays, we propose DistortionNet to recover the underlying lens geometric aberration that describes the ray profile's deviation from a perspective pinhole camera with the same focal length and field of view (Sec. 3.3). We further discuss the selected optimization objectives in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.386, + 0.834, + 0.402 + ], + "angle": 0, + "content": "3.1. Generic Ray Camera Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.528 + ], + "angle": 0, + "content": "The objective of camera calibration is to recover optimal parameters that describe the projection of 3D world points \\(\\mathcal{X} \\in \\mathbb{R}^3\\) to 2D sensor locations \\(\\pmb{x} \\in \\mathbb{R}^2\\). Our method models this with a ray-based camera model, where each unit of ray \\(\\pmb{r}_i \\in \\mathbb{R}^6\\) is stored on a grid centered on image patches. Rays originate from the camera center \\(\\pmb{r}_o \\in \\mathbb{R}^3\\), with directions \\(\\pmb{r}_d \\in \\mathbb{R}^3\\) derived by lifting 2D grid points to the camera frame using:" + }, + { + "type": "equation", + "bbox": [ + 0.646, + 0.534, + 0.905, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {\\boldsymbol {d}} = \\boldsymbol {K} ^ {- 1} \\mathcal {D} _ {\\zeta} (x), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.906, + 0.616 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}_{\\zeta}\\) is the aberration function parameterized by \\(\\zeta\\), and \\(\\pmb{K}\\) is the intrinsic matrix, accounting for geometric aberrations. The ray bundle, transformed to world coordinates by rotation \\(\\pmb{R}\\) and translation \\(\\pmb{t}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.622, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {w} = \\boldsymbol {R} ^ {T} \\boldsymbol {r} + \\boldsymbol {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.69 + ], + "angle": 0, + "content": "represents aberrated cameras in world coordinates, allowing holistic downstream recovery of rays, camera extrinsics, and optical aberrations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.906, + 0.811 + ], + "angle": 0, + "content": "Off-the-shelf image quality is completely characterized by the modulation transfer function (MTF) and distortion. The MTF performance includes all monochromatic and chromatic aberrations. Typically, the driving residual aberration in high-quality off-the-shelf lenses is optical distortion \\((W_{311}\\) in the wavefront to third order[47]), therefore, in this paper, our ray trace models focus on extracting the optical distortion characteristics from our dataset of lenses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.806, + 0.835 + ], + "angle": 0, + "content": "3.2. Geometry-controlled ray diffuser" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We train a multi-view diffusion model \\(\\mathcal{M}_{\\phi}\\) that takes multiple images of a 3D scene as input and generates the corresponding output camera ray bundles given their geometric aberration cues. Specifically, given \\(N\\) conditional views" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "containing the images and their corresponding geometric aberration latents \\(\\mathcal{Z}^g\\) from a line-segment detection network, the model learns to capture the joint distribution of \\(N\\) target ray bundles \\(\\mathcal{R}^{tgt} = \\{\\mathcal{R}_1,\\mathcal{R}_2,\\mathcal{R}_3,\\dots ,\\mathcal{R}_N\\}^{tgt}\\) from noised ray bundles \\(\\mathcal{R}^\\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.18, + 0.483, + 0.198 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathcal {R} ^ {t g t} | \\mathcal {I}, \\mathcal {Z} ^ {g}, \\mathcal {R} ^ {\\epsilon}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.484, + 0.3 + ], + "angle": 0, + "content": "Diffusion model architecture. Diffusion models approximate the data manifold by learning to invert a diffusion process from data to a presumed distribution through a denoising process. Our adopted Denoising Diffusion Probabilistic Model (DDPM) specifically defines the noise distribution to be Gaussian and transitions to the next noising step:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.312, + 0.483, + 0.332 + ], + "angle": 0, + "content": "\\[\nq (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbb {I}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.629 + ], + "angle": 0, + "content": "where \\(\\{\\beta_{1},\\beta_{2},\\dots,\\beta_{T}\\}\\) are variances within the T step nois-ing schedule, and \\(\\mathbb{I}\\) is the identity matrix. Similar to prior works [35, 55, 63], we adopt a DiT [16] as the backbone of the diffusion model, receiving image latent embeddings conditioned on aberration cues. Given a set of sequential images and target ray bundles, the model encodes each image into a latent representation \\(\\mathcal{Z}_i^c\\) through an image feature encoder TIPS [34], as well as a conditioning branch comprised of multiple ResNet [14] blocks. The image latent feature is concatenated to a noised ray bundle of the same dimension. Then, the diffusion model is trained to estimate the joint distribution of the latent-conditioned ray bundle given the conditioning geometric cues. We initialize the model from a DiT model trained for 3D shape generation, with an input resolution of \\(448\\times 448\\times 3\\). We directly inflate the latent space of the original DiT to connect with the concatenated features of noised rays and image latents, while inheriting the rest of the trained model parameters to leverage the perceptual knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Utilizing geometric cues. The conditioning branch is designed to extract guidance information that describes geometric aberration cues and align it with the denoising features. The conditioning branch is a trained line segment detection network [28]. We observe that the controls maintain a high level of consistency with the denoising features and eliminate the need to be inserted at multiple stages, as also indicated in [41]. We thus integrate the controls at a single middle block into the denoising model by adding them to the denoising features after the cross-normalization. It can serve as a plug-and-play guidance module that provides geometric aberration information about the conditioning views, such that the diffuser model is aware of the levels and types of aberration from geometric cues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Edge Attention. We introduce Edge Attention responsible for parsing features along the edges, as shown in Figure 3, which highlights the edge-aware attention mechanism designed to prioritize geometric cues along image edges. In" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "particular, edge attention takes the input embedding of an image sequence \\(\\mathcal{Z}^c\\in \\mathbb{R}^{N\\times H\\times W\\times C}\\), allowing it to perform self-attention operations across the edges. We apply a patchification operator with patch size \\(p\\times p\\) to generate patch tokens \\(\\pmb {t}\\in \\mathbb{R}^{L\\times 3}\\) with \\(L = (N\\times H\\times W / p^2)\\) denoting the length of patched tokens. To enable the model to disentangle the geometric information from the semantics, we integrate the edge information \\(e_{i\\dots N}\\) for guidance into the self-attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.235, + 0.907, + 0.27 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {(Q + e ^ {T}) K}{\\sqrt {H}} V) \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.415 + ], + "angle": 0, + "content": "where \\(H\\) is the dimension size of each head. The edge embedding is designed as a soft weighting to highlight the features along the geometrically prominent areas. The image edges are first extracted from the original image sequences using a Canny Edge detector. We then patchify them and apply a 3D convolution, mapping the embedding as \\(e_i \\in \\mathbb{R}^L\\). Ultimately, we obtain the aggregated features \\(\\hat{\\mathcal{Z}}_i\\), and we employ a feedforward network as in the original DiT before proceeding to the subsequent blocks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.855, + 0.439 + ], + "angle": 0, + "content": "3.3. Aberration profiling with DistortionNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.61 + ], + "angle": 0, + "content": "We adopt warping flow as a unified representation to model geometric aberrations of the image compared to ideal pin-hole cameras, similar to prior blind camera undistortion frameworks [29, 32]. To query the aberration profile, we connect the denoised rays to a network devised to estimate deviations of predicted ray bundles to that of a pinhole camera. The DistortNet \\(\\mathcal{M}_{\\theta}\\), with a backbone of MAE-Tiny [56], learns the mapping from the predicted set of ray bundles \\(\\mathcal{R}^{tgt}\\) to a set of warping flow maps \\(\\mathcal{F}^{tgt}\\). The learned flow maps can be applied to the image sequences for image undistortion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.733, + 0.637 + ], + "angle": 0, + "content": "3.4. Optimization objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.763 + ], + "angle": 0, + "content": "We describe the optimization objectives for estimating ray bundles and extracting the distortion map from \\(N\\) images \\(\\{\\mathcal{I}_1,\\mathcal{I}_2,\\dots ,\\mathcal{I}_N\\}\\). Estimating distorted ray bundles from in-the-wild images is under-constrained and affected by uncertainties from limited overlap and artifacts (e.g., motion blur, lighting changes). To address this, we train the network with a denoising process that reverses noised samples to match the reference:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.773, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d e n o i s e} (\\phi) = \\mathbb {E} _ {t, \\boldsymbol {r} _ {o}, \\epsilon} | | \\boldsymbol {r} _ {o} - \\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t) | | ^ {2}, \\qquad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.802, + 0.906, + 0.832 + ], + "angle": 0, + "content": "where \\(\\pmb{r}_o\\) is the original ray bundle, \\(\\mathcal{M}_{\\phi}\\) is the denoising model, and \\(\\pmb{r}_t\\) is the sample with Gaussian noise at time \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.843, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {r} _ {o} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbb{I})\\). The denoising reverts to a regressive loss, allowing additional geometric constraints." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26904" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.5, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.109, + 0.885, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.354, + 0.88, + 0.37 + ], + "angle": 0, + "content": "(e) predicted ray maps with and without edge conditioning" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.382, + 0.907, + 0.411 + ], + "angle": 0, + "content": "Figure 4. Recovered aberration profile and undistorted images. From denoised rays in world space, the DistortNet estimates the aberration pattern represented in warping flow. The undistorted images maintain coherent structural distribution compared to aberration-free images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Recovering local camera ray profiles from an image sequence is susceptible to rotational ambiguities, leading to errors in ray bundle and pose estimates. We introduce a ray directional loss to measure angular errors in ray bundles:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.49, + 0.483, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a n g u l a r}} (\\phi) = \\cos^ {- 1} \\left(\\frac {r _ {o} \\cdot \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right)}{\\left| r _ {o} \\right| \\left| \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right) \\right|}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.573 + ], + "angle": 0, + "content": "DistortNet is trained to regress backward flow, representing deviations from perspective ray bundles, by comparing each predicted flow map \\( \\pmb{f}_o \\) to the reference set \\( \\mathcal{F}^{tgt} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.578, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t o r t}} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {f} _ {o}, \\boldsymbol {r} _ {t}} | | \\boldsymbol {f} _ {o} - \\mathcal {M} _ {\\theta} \\left(\\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t)\\right) | | ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.645 + ], + "angle": 0, + "content": "This way, we recover the camera distortion maps at each image pixel, which can be used to directly unwarp the images and their associated rays." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.37, + 0.669 + ], + "angle": 0, + "content": "3.5. Physical aberration groundings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We use the largest available lens database to ray trace and extract distortion characteristics from about 3000 optical systems. We wrote custom scripts that extracted each lens prescription from the LensView database and used commercial raytracing software to extract the optical distortion maps. These systems include camera lenses, lithography lenses, and freeform optics, as shown in Fig 5. These systems produce a diverse set of optical distortion functions that map the object to the image. Detailed description on the lens dataset can be found in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Since images within each dataset are usually captured with similar equipment sharing identical camera settings, to learn different camera modalities, we augment the sequences upon loading with sampled camera distortion profiles representing shear, barrel, fisheye, and pincushion. We" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.422, + 0.905, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.619, + 0.907, + 0.676 + ], + "angle": 0, + "content": "Figure 5. Ray-traced lens designs are utilized, enabling the accurate simulation of geometric aberrations. The aberrations are encoded as local geometric distortion [19], with a percentage deviation of pixels from their ideal positions on a regular grid." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "further include a set of diverse optics dataset comprised of 3187 patented lens profiles from LensView [4]. These profiles span a wide range of professional and industrial lenses. Figure 5 shows examples from the grounding lenses, along with the distribution of the distortion with respect to the field of view (FOV), Numerical Aperture, and F-number. This augmentation introduces groundings in intrinsic and distortion profiles, addressing the scarcity of camera profiles within the datasets. We assess the out-of-distribution distortions not present in the lens set through experiments on the Aria Digital Twins dataset in Table 1." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26905" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.224, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "In this section, we first provide the implementation details of the proposed framework and then validate on two real world datasets with different camera modalities. Our method outperforms baseline approaches in both aberrated and non-aberrated images, achieving state-of-the-art results in quantitative and qualitative evaluations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.214, + 0.308, + 0.23 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.484, + 0.341 + ], + "angle": 0, + "content": "We use a pre-trained TIPS [34] network as the image feature extractor, which gives high-quality, fine-grained image embeddings. We adopted a DiT [16] as the base structure of the denoising diffuser, and augmented the attention blocks with the proposed edge conditioning attention. We train our diffusion model with \\( T = 100 \\) timesteps, with the training taking roughly 3 days on 8 H100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Following prior works [33, 55, 63], we use the first camera as the coordinate anchor to define the scene. The scene is rescaled such that the first camera has a unit from translation and is rotated to grant the first camera identity rotation. In this way, all the following predicted cameras would use the first camera as a reference to express their relative orientations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.371, + 0.471 + ], + "angle": 0, + "content": "4.2. Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Datasets. We chose CO3D [44] as the primary training dataset for the generation of ray maps. The dataset consists of roughly \\(37\\mathrm{k}360^{\\circ}\\) videos of common household objects from 51 MS-COCO categories. The dataset is annotated with COLMAP [49, 50] camera poses and intrinsics, with the frames undistorted with the estimated coefficients. Each video spans over an average 200 frames. The dataset provides a diverse profile of common objects sequences, making it suitable for training the network. In order to enhance the variety of training scenarios, we add MegaDepth [30] and TartanAir [58] as our additional training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Secondly, we evaluate the inference performance in Aria Digital Twins [39] that comprises 200 sequences and 400 min of videos that capture the daily activities of the common household using egocentric cameras. Its camera trajectories were provided by Optitrack [36] system and IMU sensor data on the Aria glass, further optimized by matching the two sources of estimated trajectories. The camera calibration parameters are provided by fitting a Kannala Brandt and Fisheye radial-tangential thin film model with the estimated trajectories. We use the same division as in [39], with a subset of 50 tests. With the two chosen datasets, we showcase the model's capabilities of estimating the ray maps from both common household videos and its ability to generalize to recent AR egocentric captures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocol. We evaluated spare view predictions for recovered poses and the accuracy of the camera ray profile compared to the reference distorted camera, without" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.093, + 0.623, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.155, + 0.623, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.215, + 0.623, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.276, + 0.623, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.337, + 0.597, + 0.349 + ], + "angle": 0, + "content": "input" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.093, + 0.709, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.155, + 0.709, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.215, + 0.709, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.276, + 0.709, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.337, + 0.709, + 0.349 + ], + "angle": 0, + "content": "" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.093, + 0.795, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.155, + 0.795, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.215, + 0.795, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.276, + 0.795, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.337, + 0.794, + 0.349 + ], + "angle": 0, + "content": "RayDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.093, + 0.882, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.155, + 0.882, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.215, + 0.882, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.276, + 0.882, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.337, + 0.849, + 0.348 + ], + "angle": 0, + "content": "ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.36, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Figure 6. Generalization to out-of-distribution camera from Aria Digital Twins. Our approach produces better results compared to [63]. See supplementary for quantitative inference results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.906, + 0.476 + ], + "angle": 0, + "content": "employing the DistortionNet. In accordance with [63], we randomly sampled \\(N\\) images for an evaluation of \\(N\\). Each reported accuracy is averaged over 5 runs to reduce stochasticity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Rotation accuracy. We first compute the relative rotations between each pair of cameras for both predicted and ground-truth poses. The errors are computed for the pairwise relative rotations and reported for the percentage of deviations less than \\(15^{\\circ}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.627 + ], + "angle": 0, + "content": "Camera center accuracy. Using the fitted ground truth and predicted poses, we align the cameras and compute the distance from the predicted camera center to the ground truth correspondence. The fraction of distances within \\(10\\%\\) of the scene scale is reported." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Angular error. With the recovered ray bundles, we compute the distance of the predicted rays to the ground truth correspondences. The mean angular error in degrees is reported." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.683, + 0.636, + 0.698 + ], + "angle": 0, + "content": "4.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Baseline comparisons. In Table 1, we present the mean ray angular error measurements for the distorted CO3D dataset, demonstrating the effectiveness of our approach compared to existing methods. Our results indicate that our method excels at extracting precise, arbitrary ray profiles even in the absence of a prior camera model. This capability is particularly significant given the challenges of distorted sequences. Notably, using the same ray representation as RayDiffusion, our approach consistently surpasses its performance, where the diffusion model depends solely on image-based features. The experiments reveal that our method achieves a more robust disentanglement from local image content, effectively isolating structural cues from object-specific fea" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26906" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.132 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of camera angular error on CO3D and Aria Digital Twins datasets. Averaged angular error (degrees) across varying numbers of input images." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.143, + 0.481, + 0.32 + ], + "angle": 0, + "content": "
CO3D – Seen Categories
Method2345678
RayRegression [63]18.420.621.721.920.323.023.7
RayDiffusion [63]9.69.710.28.79.411.913.2
AlignDiff (Ours)3.43.12.83.23.03.64.1
CO3D – Unseen Categories
RayRegression [63]24.122.521.624.823.925.126.2
RayDiffusion [63]11.410.814.112.315.816.619.4
AlignDiff (Ours)4.65.25.85.76.46.88.2
Aria Digital Twins
RayRegression [63]34.638.236.540.841.240.943.7
RayDiffusion [63]28.429.134.737.240.442.345.6
AlignDiff (Ours)15.413.814.220.621.324.624.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.337, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Table 2. Camera Rotation and Camera Center Accuracy on Distorted CO3D dataset. Experiments using geometrically aberrated videos, comparing recent methods against AlignDiff." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.39, + 0.501, + 0.64 + ], + "angle": 0, + "content": "
Rotation @ 15°
number of Images2345678
Seen ClsRelPose [62]62.024.034.024.025.033.032.0
RelPose++ [33]72.573.473.674.775.076.075.7
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]74.080.085.382.484.882.586.6
AlignDiff (Ours)90.691.191.391.892.492.793.1
Unseen ClsRelPose [62]61.027.037.026.030.026.025.0
RelPose++ [33]61.460.863.064.165.765.765.4
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]69.972.475.376.276.477.178.5
AlignDiff (Ours)83.684.884.985.285.786.186.1
Camera Center @ 0.1
Seen ClsRelPose++ [33]10085.279.074.370.568.666.0
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10073.472.571.171.070.670.2
AlignDiff (Ours)10092.590.788.687.285.884.6
Unseen ClsRelPose++ [33]10066.457.051.348.944.544.1
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10074.871.670.267.366.762.5
AlignDiff (Ours)10086.384.280.575.273.070.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.703 + ], + "angle": 0, + "content": "tures. This disentanglement minimizes residual errors tied to object appearance, allowing structural details to be more accurately emphasized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Additionally, we report metrics for camera rotation accuracy and camera center accuracy, with aberrated images in Table 2 and rectified images in Table 3. Our approach consistently outperforms existing methods in estimating both camera position and orientation under aberrated conditions, confirming its resilience to challenging visual distortions. For rectified images, our conditioning method does not simply maintain performance but rather enhances it by recognizing and compensating for residual aberrations. This capability allows our approach to improve predictions even on images with minimal distortion, achieving a high level of precision and adaptability in handling diverse image qualities." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 3. Camera Rotation and Camera Center Accuracy on Rectified CO3D dataset. AlignDiff identifies subtle local distortions, achieving improvements over previous methods." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.143, + 0.905, + 0.471 + ], + "angle": 0, + "content": "
Rotation @ 15°
number of Images2345678
Seen ClsCOLMAP [48]30.728.426.526.827.028.130.6
RelPose [62]56.056.557.057.257.257.357.2
PoseDiffusion [55]75.776.476.877.478.078.778.8
RelPose++ [33]81.882.884.184.784.985.385.5
RayRegression [63]88.888.788.789.089.489.389.2
RayDiffusion [63]91.892.492.692.993.193.393.3
AlignDiff (Ours)92.692.392.793.093.693.493.5
Unseen ClsCOLMAP [48]34.531.831.031.732.735.038.5
RelPose [62]48.647.548.148.348.448.448.3
PoseDiffusion [55]63.264.264.265.766.267.067.7
RelPose++ [33]69.871.171.972.873.874.474.9
RayRegression [63]79.079.680.681.481.381.981.9
RayDiffusion [63]83.585.686.386.987.287.588.1
AlignDiff (Ours)86.386.687.487.788.588.789.2
Camera Center @ 0.1
Seen ClsCOLMAP [48]10034.523.818.915.614.515.0
RelPose [62]10076.566.962.459.458.056.5
PoseDiffusion [55]10077.569.765.963.762.861.9
RelPose++ [33]10085.078.074.271.970.368.8
RayRegression [63]10091.785.782.179.877.976.2
RayDiffusion [63]10094.290.587.886.285.084.1
AlignDiff (Ours)10093.791.887.587.185.283.9
Unseen ClsCOLMAP [48]10036.025.520.017.917.619.1
PoseDiffusion [55]10063.650.545.743.041.239.9
RelPose++ [33]10070.658.853.450.447.846.6
RayRegression [63]10083.775.670.867.465.363.9
RayDiffusion [63]10087.781.177.074.172.471.4
AlignDiff (Ours)10088.281.079.377.572.671.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Table 4. Camera Rotation and Camera Center Accuracy on Aria Digital Twins dataset." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.53, + 0.905, + 0.684 + ], + "angle": 0, + "content": "
Rotation @ 15°
Number of Images2345678
Unseen ClsRelPose [62]59.632.738.031.228.627.426.0
RelPose++ [33]62.861.362.662.863.264.964.8
PoseDiffusion [55]75.474.874.175.875.275.976.2
RayDiffusion [63]75.276.378.979.279.680.080.2
AlignDiff (Ours)81.482.782.884.384.184.685.5
Camera Center @ 0.1
Unseen ClsRelPose++ [33]10070.356.152.848.646.444.7
PoseDiffusion [55]10071.857.454.651.852.248.5
RayDiffusion [63]10075.969.261.860.455.651.1
AlignDiff (Ours)10078.370.064.462.758.152.9
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Qualitative analysis of recovered cameras. We compare the generated cameras with the ground truth cameras in world space, as shown in Figure 7. The predicted ray moments and angular error from ground truth ray directions are also shown. Remarkably, the generated cameras exhibit little rotational difference from the reference cameras. While there are observed to have noticeable translation errors, we argue that this is due to scale ambiguities of the captures. The visualized ray moments present a smooth profile with very low angular errors. We further compare the Aria Digital Twins dataset for the inferred rays in vector form, shown in Figure 6. These further demonstrate that our approach can achieve robust and accurate ray profiles through camera" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26907" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.118 + ], + "angle": 0, + "content": "Table 5. Ablated framework design evaluated with averaged ray angular error." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.122, + 0.482, + 0.204 + ], + "angle": 0, + "content": "
Methods2345678
Baseline8.89.510.49.212.412.713.9
AlignDiff (w/o angular loss)5.86.18.87.39.611.314.7
AlignDiff (w/o line conditioning)6.78.49.89.910.310.112.8
AlignDiff (w/o edge-attention)5.37.17.49.210.811.310.5
AlignDiff (w/o optics grounding)5.26.96.87.78.49.111.3
AlignDiff4.65.25.85.76.46.88.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.221, + 0.473, + 0.235 + ], + "angle": 0, + "content": "Table 6. Evaluations on different image latent encoder choices." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.238, + 0.501, + 0.29 + ], + "angle": 0, + "content": "
Methods2345678
DINOv2-s/14 [37]8.28.18.610.49.810.112.5
DINOv2-g/14 [37]4.56.46.36.66.17.89.6
TIPS-g/14 [34]4.65.25.85.76.46.88.2
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.306, + 0.482, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Figure 7. Recovered cameras from aberrated image sequences. Predicted and expected cameras are visualized in solid and dashed lines, respectively. The cameras align with the rotational orientation of the expected counterparts. The ray moment and residual errors show low levels of deviation from reference rays." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.627 + ], + "angle": 0, + "content": "conditioning. Detailed quantitative evaluation on the Aria Digital Twins dataset are included in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Undistortion using predicted flow. We conduct image undistortion with the predicted aberration map, as shown in Figure 4. Since we adopt a unified representation for aberrations, the radial fisheye camera model's vignetting effect leads to mapping artifacts. The predicted aberration map overall correctly captures the image distortions and reliably remaps the images regardless of the aberration types." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation study. We perform an ablation study on CO3D data to evaluate the impact of each conditioning module, enhanced geometric guidance, and additional losses, as shown in Table 5. For each comparison, we remove one conditioning module from the framework. The results show that all modules significantly improve performance, and latent conditioning of the line segment has the greatest effect. Figure 4 visualizes the ray residuals with and without aberration conditioning, further validating its effectiveness. Additionally, we test the angular loss, finding that our localized ray angular loss yields notable improvements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.392 + ], + "angle": 0, + "content": "While our multiview diffusion approach implicitly incorporates temporal consistency by explicit temporal losses, we chose to handle temporal consistency as a post-processing step to keep the model lightweight and directly applicable in real-world deployments. This approach reduces computational overhead while still providing meaningful consistency improvements, leaving more complex temporal modeling for future work. This is based on the premise that camera profiles remain generally consistent over short periods, barring changes due to factors like camera sensor temperature fluctuations. To test this, we implemented a post-processing step that rejects low-confidence pose estimations, which reduced the average ray angular error from \\(8.20^{\\circ}\\) to \\(8.06^{\\circ}\\) over 8 frames. Furthermore, enforcing temporal consistency on local camera ray profiles using reprojection mean squared error loss yielded an average reduction of \\(0.06^{\\circ}\\) in angular error across the frames. We plan to integrate these enhancements, along with outlier rejection and a frame selection mechanism, into our future work." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Evaluations on image representation models. The image sequence encoder is designed to generate dense global embeddings that capture perceptual, geometric, and semantic representations. Among the available models, DINOv2 and TIPS are well suited for extracting high-quality dense features. We conduct an ablation study to evaluate these models' effectiveness in querying camera ray profiles. As shown in Table 6, the TIPS-g/14 model achieves the lowest angular errors on CO3D, with DINOv2-g/14 as a close competitor." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.558, + 0.634, + 0.573 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.905, + 0.809 + ], + "angle": 0, + "content": "We introduce AlignDiff, a unified calibration framework for real-world optical aberrations, enabling the recovery of camera ray bundles, extrinsics, and aberration profiles from multi-view image sequences using simple conditioning techniques. First, we propose structural conditioning to separate image geometry from semantic features, promoting emphasis on structural cues for accurate ray profile estimation. Second, we aggregate features along edges, ensuring that high-quality image embeddings are paired with effective guidance to avoid semantic interference. Finally, we incorporate real-world lens profiles into training, grounding AlignDiff in actual camera designs. These contributions collectively enhance generalization and accuracy across diverse distortion profiles, as demonstrated by consistent improvements over strong baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.9 + ], + "angle": 0, + "content": "While AlignDiff performs well in controlled indoor settings, ray estimation may become unstable in arbitrary outdoor scenes due to scale ambiguity. To enhance robustness in such cases, integrating geometric priors like optical flow or combining traditional camera parameter estimation with network inference may be promising for future directions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Diffusionsfm: Predicting structure and motion via ray origin and endpoint diffusion, 2024. [Online; accessed 2025-03-06]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Tommaso Cavallari, Stuart Golodetz, Nicholas A. Lord, Julien Valentin, Victor A. Prisacariu, Luigi Di Stefano, and Philip H.S. Torr. Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2465-2477, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Shuai Chen, Xinghui Li, Zirui Wang, and Victor Adrian Prisacariu. DFNet: Enhance Absolute Pose Regression with Direct Feature Matching. 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[4] Lambda Research Corp. Lensview optical design database. [Online; accessed 2025-03-07]. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[5] Andrei Cramariuc, Aleksandar Petrov, Rohit Suri, Mayank Mittal, Roland Siegwart, and Cesar Cadena. Learning camera miscalibration detection. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 4997-5003, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-Supervised Interest Point Detection and Description. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[7] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[8] Satoshi Fujimoto and Nobutomo Matsunaga. Deep feature-based rgb-d odometry using superpoint and superglue. *Proceeding Computer Science*, 227:1127–1134, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[9] Fabrizio Gentile, Crescenzo Tortora, Giovanni Covone, Léon V. E. Koopmans, Rui Li, Laura Leuzzi, and Nicola R. Napolitano. *LeMoN: Lens Modelling with Neural networks* - I. Automated modelling of strong gravitational lenses with Bayesian Neural Networks. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[10] Behnam Gholami, Mostafa El-Khamy, and Kee-Bong Song. Latent feature disentanglement for visual domain generalization. IEEE Transactions on Image Processing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[11] Diandian Guo, Deng-Ping Fan, Tongyu Lu, Christos Sakaridis, and Luc Van Gool. Vanishing-point-guided video semantic segmentation of driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3544-3553, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Lanqing Guo, Renjie Wan, Wenhan Yang, Alex C Kot, and Bihan Wen. Cross-image disentanglement for low-light enhancement in real world. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2550-2563, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[13] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Xiankang He, Guangkai Xu, Bo Zhang, Hao Chen, Ying Cui, and Dongyan Guo. DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation. 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Technical report. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[17] Yannick Hold-Geoffroy, Dominique Piche-Meunier, Kalyan Sunkavalli, Jean-Charles Bazin, François Rameau, and Jean-François Lalonde. A perceptual measure for deep single image camera and lens calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10603-10614, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[18] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[19] ISO 17850:2015(en). Photography — Digital cameras — Geometric distortion (GD) measurements. Standard, International Organization for Standardization, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[20] Y. Y. Jau, R. Zhu, H. Su, and M. Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4950-4957, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.904, + 0.469 + ], + "angle": 0, + "content": "[21] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable Feature Matching with Foundation Model Guidance. Technical report. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.904, + 0.51 + ], + "angle": 0, + "content": "[22] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-View Object Reconstruction with Unknown Categories and Camera Poses. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[23] Yang Jiao, Trac D. Tran, and Guangming Shi. Effiscene: Efficient per-pixel rigidity inference for unsupervised joint learning of optical flow, depth, camera pose and motion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5538-5547, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[24] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F. Fouhey. Perspective fields for single image camera calibration. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[25] Alex Kendall, Matthew Koichi Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. 2015 IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[26] Julius Kümmerle and Tilman Kühner. Unified intrinsic and extrinsic camera and lidar calibration under uncertainties. In 2020 IEEE International conference on Robotics and Automation (ICRA), pages 6028-6034. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[27] Jinwoo Lee, Hyunsung Go, Hyunjoon Lee, Sunghyun Cho, Minhyuk Sung, and Junho Kim. Ctrl-c: Camera calibration transformer with line-classification. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16208-16217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.887 + ], + "angle": 0, + "content": "[28] Hao Li, Huai Yu, Jinwang Wang, Wen Yang, Lei Yu, and Sebastian Scherer. Ulsd: Unified line segment detection across pinhole, fisheye, and spherical cameras. ISPRS Journal of Photogrammetry and Remote Sensing, 178:187-202, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.549, + 0.889, + 0.558, + 0.898 + ], + "angle": 0, + "content": "4" + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26909" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[29] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4855-4864, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[31] Kang Liao, Lang Nie, Shujuan Huang, Chunyu Lin, Jing Zhang, Yao Zhao, Moncef Gabbouj, and Dacheng Tao. Deep learning for camera calibration and beyond: A survey. arXiv preprint arXiv:2303.10559, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[32] Kang Liao, Zongsheng Yue, Zhonghua Wu, and Chen Change Loy. Mowa: Multiple-in-one image warping model. ArXiv, abs/2404.10716, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[33] Amy Lin, Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. RelPose++: Recovering 6D Poses from Sparse-view Observations. 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[34] Kevis-Kokitsi Maninis, Kaifeng Chen, Soham Ghosh, Arjun Karpur, Koert Chen, Ye Xia, Bingyi Cao, Daniel Salz, Guangxing Han, Jan Dlabal, et al. Tips: Text-image pretraining with spatial awareness. arXiv preprint arXiv:2410.16512, 2024. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[35] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In Advances in Neural Information Processing Systems, pages 67960–67971. Curran Associates, Inc., 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.518 + ], + "angle": 0, + "content": "[36] Inc. DBA OptiTrack NaturalPoint. Optitrack system, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[38] Linfei Pan, Marc Pollefeys, and Viktor Larsson. Camera pose estimation using implicit distortion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12819-12828, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[39] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[40] Chethan M. Parameshwara, Gokul Hari, Cornelia Fermüller, Nitin J. Sanket, and Yiannis Aloimonos. Diffposenet: Direct differentiable camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6845-6854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. ControlNeXt: Powerful and Efficient Control for Image and Video Generation. 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[42] Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Martins, and Erickson R. Nascimento. XFeat: Accelerated Features for Lightweight Image Matching. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[43] Srikumar Ramalingam and Peter Sturm. A Unifying Model for Camera Calibration: A Unifying Model for Camera Calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence: A Unifying Model for Camera Calibration. (7): 1309-1319, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction. Technical report. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] José Sasán. Introduction to aberrations in optical imaging systems. Cambridge University Press, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[48] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. Technical report. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[49] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[50] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[51] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[52] Peter Sturm and Srikumar Ramalingam. A Generic Concept for Camera Calibration. pages 1-13, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[53] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement. Technical report. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[54] Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys. GeoCalib: Single-image Calibration with Geometric Optimization. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[55] Jianyuan Wang, Christian Rupprecht, and David Novotny. PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment. 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[56] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[57] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[58] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[59] Scott Workman, Connor Greenwell, Menghua Zhai, Ryan Baltenberger, and Nathan Jacobs. Deepfocal: A method for direct focal length estimation. In 2015 IEEE International Conference on Image Processing (ICIP), pages 1369-1373, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[60] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Innovating Real Fisheye Image Correction with Dual Diffusion Architecture. Technical report. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[61] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[62] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. page 592-611, Berlin, Heidelberg, 2022. Springer-Verlag. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[63] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as Rays: Pose Estimation via Ray Diffusion. 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[64] Yinqiang Zheng, Yubin Kuang, Shigeki Sugimoto, Kalle Åström, and Masatoshi Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In 2013 IEEE International Conference on Computer Vision, pages 2344-2351, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[65] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[66] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a Wild Camera: In-the-Wild Monocular Camera Calibration. 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26911" + } + ] +] \ No newline at end of file diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_origin.pdf b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab163b74b0e502812f40b97e4c728e55a1c16235 --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/6380161a-ecce-4164-bb4a-61e264b72668_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8a0f1a0fc2038d68e9aa4f371e439ea7be7053fc6deaf5750b19d59e5b6050b +size 9756932 diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/full.md b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa9004ac41568961182bcd4cbd6d33e6d66e01e2 --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/full.md @@ -0,0 +1,349 @@ +# AlignDiff: Learning Physically-Grounded Camera Alignment via Diffusion + +Liyue Xie $^{1*}$ Jiancong Guo $^{2}$ Ozan Cakmakci $^{2}$ Andre Araujo $^{3}$ Laszlo A. Jeni $^{1}$ Zhiheng Jia $^{2}$ + +$^{1}$ Carnegie Mellon University $^{2}$ Google $^{3}$ Google DeepMind + +# Abstract + +Accurate camera calibration is a fundamental task for 3D perception, especially when dealing with real-world, in-the-wild environments where complex optical distortions are common. Existing methods often rely on pre-rectified images or calibration patterns, which limits their applicability and flexibility. In this work, we introduce a novel framework that addresses these challenges by jointly modeling camera intrinsic and extrinsic parameters using a generic ray camera model. Unlike previous approaches, AlignDiff shifts focus from semantic to geometric features, enabling more accurate modeling of local distortions. We propose AlignDiff, a diffusion model conditioned on geometric priors, enabling the simultaneous estimation of camera distortions and scene geometry. To enhance distortion prediction, we incorporate edge-aware attention, focusing the model on geometric features around image edges, rather than semantic content. Furthermore, to enhance generalizability to real-world captures, we incorporate a large database of ray-traced lenses containing over three thousand samples. This database characterizes the distortion inherent in a diverse variety of lens forms. Our experiments demonstrate that the proposed method significantly reduces the angular error of estimated ray bundles by $\sim 8.2^{\circ}$ and overall calibration accuracy, outperforming existing approaches on challenging, real-world datasets. + +# 1. Introduction + +Accurate camera calibration, involving the estimation of intrinsic parameters such as focal length and lens distortions and extrinsic parameters such as camera pose, is essential for robust 3D perception in real-world environments. However, prevailing methods typically address these components independently, focusing on either intrinsic calibration or pose estimation while relying on simplified camera models that are insufficient to capture complex real-world opti + +![](images/6801c14600fa42551e31acef456bcafdd21bf9c38d3cc60822f075076ce47f36.jpg) +Figure 1. AlignDiff is proposed to address common image geometric aberrations with a unified ray camera representation while jointly recovering the camera extrinsics. With groundings on physical camera lens designs, as well as the disassociation of geometric cues from semantic features, it demonstrates an ability to generalize to real video sequences. + +cal aberrations. This separation is fundamentally limiting, as intrinsic and extrinsic parameters are closely interdependent; inaccuracies in modeling lens distortions propagate directly to pose estimation errors. Consequently, joint optimization of intrinsic and extrinsic parameters is critical to ensure reliable calibration under unconstrained conditions. + +Recent diffusion-based and transformer-powered models, such as PoseDiffusion [55], RayDiffusion [63], and DiffusionSfM [1] have improved calibration accuracy and zero-shot generalizability by conditioning on image features extracted from vision encoders. However, these methods typically focus on high-level, semantic image features rather than on structural cues critical for modeling optical aberrations that directly impact geometry. This is particularly limiting because accurate intrinsic calibration is essential for precise extrinsic calibration: aberrations in the intrinsic parameters can significantly affect the accuracy of extrinsic parameters, making the two tightly interdependent. + +We propose AlignDiff, shown Figure 1, a diffusion-based calibration framework, to learn the fine-grained ray profile from video sequences in world space, with a conditioning strategy that better captures fine-grained optical distortions. To shift focus from semantic to structural features, we condition the diffusion model with line embed + +![](images/1869ee0aa8bc9062b2afa671dd171f8fd9fc04f9e99abe03d1d13e436662b775.jpg) +Figure 2. AlignDiff Architecture. We promote learning camera ray profiles in three main steps: geometric cue conditioning from line features, edge-aware attention, and physical camera groundings. + +dings from a line detection network. This approach emphasizes the geometric structures, helping the model prioritize optical aberrations over object content. In addition, we further introduce edge-focused attention to further enhance the model's sensitivity to regions around edges, where aberrations often manifest most prominently. + +An additional limitation of prior models is their reliance on simulated optical aberrations derived from predefined models, which often lack the subtlety of real-world aberrations, such as localized or asymmetric profiles. To address this, our framework incorporates real optical profiles from actual lens designs, grounding the model in authentic optical characteristics and boosting its generalizability. + +In summary, our framework builds upon existing diffusion-based calibration techniques by incorporating targeted conditioning strategies and real-world optical data, providing a robust solution that accurately captures complex optical aberrations. By effectively addressing intrinsic aberrations, our model lays a solid foundation for accurate extrinsic calibration, making it highly applicable for in-the-wild scenarios requiring precise camera parameters. + +- To our knowledge, AlignDiff is the first unified diffusion-based approach addressing complex optical aberrations in joint intrinsic and extrinsic calibration. +- By conditioning the diffusion model on line embeddings and incorporating edge-focused attention, our method prioritizes structural over semantic features, achieving generalization to diverse, real-world environments without requiring extensive retraining. +- Our approach incorporates authentic optical profiles from actual lens designs, grounding the model in real-world aberrations and improving generalization to natural images. + +# 2. Related Works + +Aberration representation for in-the-wild images. Traditional camera calibration methods model intrinsic characteristics, such as focal length and camera center, alongside distortion parameters that describe image geometric aberrations. Early deep learning-based methods [3, 9, 15, 45, 57, 59, 60] enabled calibration from images. Recent diffusion based approaches like PoseDiffusion [55], extend from prior approaches with improved zero-shot performance on wild captures. While effective for rectified images, these methods are limited in their ability to account for local optical aberrations resulting from factors like wear, temperature fluctuations, and lens designs. + +Generic ray camera models were introduced to better capture local geometric distortions [43, 52] by modeling ray bundles across image patches, offering a more detailed representation of pixel deviations compared to the pinhole model. While accurate, they require denser calibration patterns, especially when optimized using methods like PnP [64]. Several recent work [15, 63, 66] adopted the ray representation in deep learning and proved its feasibility. Yet, these prior works remain largely restricted to the pinhole model and did not exploit the expressiveness of ray representation for aberration modeling. The recent blind camera undistortion line of work [29, 32] infers geometric displacements from a single image, often as a preprocessing step to provide rectified images. + +Our work advances these approaches by introducing a unified ray-based framework that models both pinhole projections and local distortions. We directly estimate the ray profile and camera aberrations from a raw monocular sequence, enhancing generalizability in diverse real-world + +conditions. + +Joint Intrinsic and Extrinsic Calibration Classical camera calibration methods often assume a predefined camera model for each type of aberration, relying on calibration patterns and multi-view geometric constraints to estimate camera parameters from a sequence of images [13, 24-26, 33, 54, 55]. Early approaches solved joint intrinsic and extrinsic calibration through point matching across views [2, 3, 11, 22, 38, 49, 65], with refinements summarized in [31]. Structure-from-motion (SfM) techniques later optimized reprojection loss to refine parameters further. + +For in-the-wild images, Hold-Geoffroy et al. [17] leveraged DenseNet [18] to estimate parameters such as horizon angle and vertical field of view. Other methods [20, 24] decouple parameter estimation from direct regression, and with integrated semantic or geometric guidance [6, 8, 21, 46]. Further advancements have shown that models pretrained with geometric understanding can act as better feature encoders [5, 15, 23, 40, 42, 45, 51, 53, 61]. Recent methods, including PoseDiffusion [55], DuSt3R [57], RayDiffusion [63] use diffusion and transformer to recover camera parameters or ray profiles directly from images, though they still assume fixed distortion models, limiting generalization to diverse real-world aberrations [31, 32]. The camera parameters are recovered by conditioning on local image features to capture local and cross-view geometric associations. However, these methods can yield residual errors around object contours due to reliance on semantic rather than purely geometric cues [10, 12, 27]. + +Our approach extends the ray representation without a fixed distortion model, disentangling semantic and geometric cues to reduce contour errors and enhance calibration accuracy. By integrating undistortion and calibration in a unified framework, we increase flexibility and applicability for real-world captures. + +# 3. AlignDiff + +![](images/99a9178ef9bf3769274b010533bfbb83eb39affe781463e199b1868f19510f13.jpg) +Figure 3. The latents are reweighted through edge-attention that aggregates a learned mask to the Query feature to promote information along edges where the geometric cues are more prominent. Following, the multi-view attention further captures features from different views. + +![](images/0f254a720e0a4f370a862b472a17198001356c9e908487e19dd9fa2be351b9bc.jpg) + +We aim to recover generic cameras represented as ray + +bundles in world space from a set of $N$ input images $\{\mathcal{I}_1,\mathcal{I}_2,\mathcal{I}_3,\dots ,\mathcal{I}_N\}$ . Our approach, as presented in Figure 2, captures fine-grained geometric distortions while maintaining compatibility with classical parameterized camera models. This formulation is conceptually similar to works like [63, 66], but extends the ray camera model to account for local distortions (Sec. 3.1). In contrast, previous methods assume a pinhole camera model, neglecting distortions, even though they predict rays for local image patches. We adopt a Diffusion Transformer as the base architecture while enhancing this architecture by introducing geometric cues that guide ray directions based on local geometry, disentangling image perceptual features from distortions (Sec. 3.2). From the predicted rays, we propose DistortionNet to recover the underlying lens geometric aberration that describes the ray profile's deviation from a perspective pinhole camera with the same focal length and field of view (Sec. 3.3). We further discuss the selected optimization objectives in Sec. 3.4. + +# 3.1. Generic Ray Camera Representation + +The objective of camera calibration is to recover optimal parameters that describe the projection of 3D world points $\mathcal{X} \in \mathbb{R}^3$ to 2D sensor locations $\pmb{x} \in \mathbb{R}^2$ . Our method models this with a ray-based camera model, where each unit of ray $\pmb{r}_i \in \mathbb{R}^6$ is stored on a grid centered on image patches. Rays originate from the camera center $\pmb{r}_o \in \mathbb{R}^3$ , with directions $\pmb{r}_d \in \mathbb{R}^3$ derived by lifting 2D grid points to the camera frame using: + +$$ +\boldsymbol {r} _ {\boldsymbol {d}} = \boldsymbol {K} ^ {- 1} \mathcal {D} _ {\zeta} (x), \tag {1} +$$ + +where $\mathcal{D}_{\zeta}$ is the aberration function parameterized by $\zeta$ , and $\pmb{K}$ is the intrinsic matrix, accounting for geometric aberrations. The ray bundle, transformed to world coordinates by rotation $\pmb{R}$ and translation $\pmb{t}$ , + +$$ +\boldsymbol {r} _ {w} = \boldsymbol {R} ^ {T} \boldsymbol {r} + \boldsymbol {t}, \tag {2} +$$ + +represents aberrated cameras in world coordinates, allowing holistic downstream recovery of rays, camera extrinsics, and optical aberrations. + +Off-the-shelf image quality is completely characterized by the modulation transfer function (MTF) and distortion. The MTF performance includes all monochromatic and chromatic aberrations. Typically, the driving residual aberration in high-quality off-the-shelf lenses is optical distortion $(W_{311}$ in the wavefront to third order[47]), therefore, in this paper, our ray trace models focus on extracting the optical distortion characteristics from our dataset of lenses. + +# 3.2. Geometry-controlled ray diffuser + +We train a multi-view diffusion model $\mathcal{M}_{\phi}$ that takes multiple images of a 3D scene as input and generates the corresponding output camera ray bundles given their geometric aberration cues. Specifically, given $N$ conditional views + +containing the images and their corresponding geometric aberration latents $\mathcal{Z}^g$ from a line-segment detection network, the model learns to capture the joint distribution of $N$ target ray bundles $\mathcal{R}^{tgt} = \{\mathcal{R}_1,\mathcal{R}_2,\mathcal{R}_3,\dots ,\mathcal{R}_N\}^{tgt}$ from noised ray bundles $\mathcal{R}^\epsilon$ : + +$$ +p \left(\mathcal {R} ^ {t g t} | \mathcal {I}, \mathcal {Z} ^ {g}, \mathcal {R} ^ {\epsilon}\right). \tag {3} +$$ + +Diffusion model architecture. Diffusion models approximate the data manifold by learning to invert a diffusion process from data to a presumed distribution through a denoising process. Our adopted Denoising Diffusion Probabilistic Model (DDPM) specifically defines the noise distribution to be Gaussian and transitions to the next noising step: + +$$ +q (x _ {t} | x _ {t - 1}) = \mathcal {N} (x _ {t}; \sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathbb {I}), \tag {4} +$$ + +where $\{\beta_{1},\beta_{2},\dots,\beta_{T}\}$ are variances within the T step nois-ing schedule, and $\mathbb{I}$ is the identity matrix. Similar to prior works [35, 55, 63], we adopt a DiT [16] as the backbone of the diffusion model, receiving image latent embeddings conditioned on aberration cues. Given a set of sequential images and target ray bundles, the model encodes each image into a latent representation $\mathcal{Z}_i^c$ through an image feature encoder TIPS [34], as well as a conditioning branch comprised of multiple ResNet [14] blocks. The image latent feature is concatenated to a noised ray bundle of the same dimension. Then, the diffusion model is trained to estimate the joint distribution of the latent-conditioned ray bundle given the conditioning geometric cues. We initialize the model from a DiT model trained for 3D shape generation, with an input resolution of $448\times 448\times 3$ . We directly inflate the latent space of the original DiT to connect with the concatenated features of noised rays and image latents, while inheriting the rest of the trained model parameters to leverage the perceptual knowledge. + +Utilizing geometric cues. The conditioning branch is designed to extract guidance information that describes geometric aberration cues and align it with the denoising features. The conditioning branch is a trained line segment detection network [28]. We observe that the controls maintain a high level of consistency with the denoising features and eliminate the need to be inserted at multiple stages, as also indicated in [41]. We thus integrate the controls at a single middle block into the denoising model by adding them to the denoising features after the cross-normalization. It can serve as a plug-and-play guidance module that provides geometric aberration information about the conditioning views, such that the diffuser model is aware of the levels and types of aberration from geometric cues. + +Edge Attention. We introduce Edge Attention responsible for parsing features along the edges, as shown in Figure 3, which highlights the edge-aware attention mechanism designed to prioritize geometric cues along image edges. In + +particular, edge attention takes the input embedding of an image sequence $\mathcal{Z}^c\in \mathbb{R}^{N\times H\times W\times C}$ , allowing it to perform self-attention operations across the edges. We apply a patchification operator with patch size $p\times p$ to generate patch tokens $\pmb {t}\in \mathbb{R}^{L\times 3}$ with $L = (N\times H\times W / p^2)$ denoting the length of patched tokens. To enable the model to disentangle the geometric information from the semantics, we integrate the edge information $e_{i\dots N}$ for guidance into the self-attention mechanism: + +$$ +\operatorname {A t t e n t i o n} (Q, K, V) = \operatorname {S o f t m a x} (\frac {(Q + e ^ {T}) K}{\sqrt {H}} V) \quad (5) +$$ + +where $H$ is the dimension size of each head. The edge embedding is designed as a soft weighting to highlight the features along the geometrically prominent areas. The image edges are first extracted from the original image sequences using a Canny Edge detector. We then patchify them and apply a 3D convolution, mapping the embedding as $e_i \in \mathbb{R}^L$ . Ultimately, we obtain the aggregated features $\hat{\mathcal{Z}}_i$ , and we employ a feedforward network as in the original DiT before proceeding to the subsequent blocks. + +# 3.3. Aberration profiling with DistortionNet + +We adopt warping flow as a unified representation to model geometric aberrations of the image compared to ideal pin-hole cameras, similar to prior blind camera undistortion frameworks [29, 32]. To query the aberration profile, we connect the denoised rays to a network devised to estimate deviations of predicted ray bundles to that of a pinhole camera. The DistortNet $\mathcal{M}_{\theta}$ , with a backbone of MAE-Tiny [56], learns the mapping from the predicted set of ray bundles $\mathcal{R}^{tgt}$ to a set of warping flow maps $\mathcal{F}^{tgt}$ . The learned flow maps can be applied to the image sequences for image undistortion. + +# 3.4. Optimization objectives + +We describe the optimization objectives for estimating ray bundles and extracting the distortion map from $N$ images $\{\mathcal{I}_1,\mathcal{I}_2,\dots ,\mathcal{I}_N\}$ . Estimating distorted ray bundles from in-the-wild images is under-constrained and affected by uncertainties from limited overlap and artifacts (e.g., motion blur, lighting changes). To address this, we train the network with a denoising process that reverses noised samples to match the reference: + +$$ +\mathcal {L} _ {d e n o i s e} (\phi) = \mathbb {E} _ {t, \boldsymbol {r} _ {o}, \epsilon} | | \boldsymbol {r} _ {o} - \mathcal {M} _ {\phi} (\boldsymbol {r} _ {t}, t) | | ^ {2}, \qquad (6) +$$ + +where $\pmb{r}_o$ is the original ray bundle, $\mathcal{M}_{\phi}$ is the denoising model, and $\pmb{r}_t$ is the sample with Gaussian noise at time $t$ : + +$$ +\boldsymbol {r} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {r} _ {o} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {7} +$$ + +where $\epsilon \sim \mathcal{N}(0,\mathbb{I})$ . The denoising reverts to a regressive loss, allowing additional geometric constraints. + +![](images/80772bc658c3b1f22eef18527756e6b4c1bb5b8e5429969b9ebd6718b325e82f.jpg) +Figure 4. Recovered aberration profile and undistorted images. From denoised rays in world space, the DistortNet estimates the aberration pattern represented in warping flow. The undistorted images maintain coherent structural distribution compared to aberration-free images. + +![](images/6f44db766740fd90dc34d63f2cbc7e9f77e4d15e9859debead5069d3ffc21549.jpg) +(e) predicted ray maps with and without edge conditioning + +Recovering local camera ray profiles from an image sequence is susceptible to rotational ambiguities, leading to errors in ray bundle and pose estimates. We introduce a ray directional loss to measure angular errors in ray bundles: + +$$ +\mathcal {L} _ {\text {a n g u l a r}} (\phi) = \cos^ {- 1} \left(\frac {r _ {o} \cdot \mathcal {M} _ {\phi} \left(\boldsymbol {r} _ {t} , t\right)}{\left| r _ {o} \right| \left| \mathcal {M} _ {\phi} \left(\boldsymbol {r} _ {t} , t\right) \right|}\right), \tag {8} +$$ + +DistortNet is trained to regress backward flow, representing deviations from perspective ray bundles, by comparing each predicted flow map $\pmb{f}_o$ to the reference set $\mathcal{F}^{tgt}$ : + +$$ +\mathcal {L} _ {\text {d i s t o r t}} (\theta) = \mathbb {E} _ {t, \boldsymbol {f} _ {o}, \boldsymbol {r} _ {t}} | | \boldsymbol {f} _ {o} - \mathcal {M} _ {\theta} \left(\mathcal {M} _ {\phi} (\boldsymbol {r} _ {t}, t)\right) | | ^ {2}. \tag {9} +$$ + +This way, we recover the camera distortion maps at each image pixel, which can be used to directly unwarp the images and their associated rays. + +# 3.5. Physical aberration groundings + +We use the largest available lens database to ray trace and extract distortion characteristics from about 3000 optical systems. We wrote custom scripts that extracted each lens prescription from the LensView database and used commercial raytracing software to extract the optical distortion maps. These systems include camera lenses, lithography lenses, and freeform optics, as shown in Fig 5. These systems produce a diverse set of optical distortion functions that map the object to the image. Detailed description on the lens dataset can be found in Appendix A. + +Since images within each dataset are usually captured with similar equipment sharing identical camera settings, to learn different camera modalities, we augment the sequences upon loading with sampled camera distortion profiles representing shear, barrel, fisheye, and pincushion. We + +![](images/79469f2eb5ef24c798a707a056b8f0eda675c269adeb76394a2e29d0f832e4ca.jpg) +Figure 5. Ray-traced lens designs are utilized, enabling the accurate simulation of geometric aberrations. The aberrations are encoded as local geometric distortion [19], with a percentage deviation of pixels from their ideal positions on a regular grid. + +further include a set of diverse optics dataset comprised of 3187 patented lens profiles from LensView [4]. These profiles span a wide range of professional and industrial lenses. Figure 5 shows examples from the grounding lenses, along with the distribution of the distortion with respect to the field of view (FOV), Numerical Aperture, and F-number. This augmentation introduces groundings in intrinsic and distortion profiles, addressing the scarcity of camera profiles within the datasets. We assess the out-of-distribution distortions not present in the lens set through experiments on the Aria Digital Twins dataset in Table 1. + +# 4. Experiments + +In this section, we first provide the implementation details of the proposed framework and then validate on two real world datasets with different camera modalities. Our method outperforms baseline approaches in both aberrated and non-aberrated images, achieving state-of-the-art results in quantitative and qualitative evaluations. + +# 4.1. Implementation Details + +We use a pre-trained TIPS [34] network as the image feature extractor, which gives high-quality, fine-grained image embeddings. We adopted a DiT [16] as the base structure of the denoising diffuser, and augmented the attention blocks with the proposed edge conditioning attention. We train our diffusion model with $T = 100$ timesteps, with the training taking roughly 3 days on 8 H100 GPUs. + +Following prior works [33, 55, 63], we use the first camera as the coordinate anchor to define the scene. The scene is rescaled such that the first camera has a unit from translation and is rotated to grant the first camera identity rotation. In this way, all the following predicted cameras would use the first camera as a reference to express their relative orientations. + +# 4.2. Dataset and Evaluation Metrics + +Datasets. We chose CO3D [44] as the primary training dataset for the generation of ray maps. The dataset consists of roughly $37\mathrm{k}360^{\circ}$ videos of common household objects from 51 MS-COCO categories. The dataset is annotated with COLMAP [49, 50] camera poses and intrinsics, with the frames undistorted with the estimated coefficients. Each video spans over an average 200 frames. The dataset provides a diverse profile of common objects sequences, making it suitable for training the network. In order to enhance the variety of training scenarios, we add MegaDepth [30] and TartanAir [58] as our additional training data. + +Secondly, we evaluate the inference performance in Aria Digital Twins [39] that comprises 200 sequences and 400 min of videos that capture the daily activities of the common household using egocentric cameras. Its camera trajectories were provided by Optitrack [36] system and IMU sensor data on the Aria glass, further optimized by matching the two sources of estimated trajectories. The camera calibration parameters are provided by fitting a Kannala Brandt and Fisheye radial-tangential thin film model with the estimated trajectories. We use the same division as in [39], with a subset of 50 tests. With the two chosen datasets, we showcase the model's capabilities of estimating the ray maps from both common household videos and its ability to generalize to recent AR egocentric captures. + +Evaluation protocol. We evaluated spare view predictions for recovered poses and the accuracy of the camera ray profile compared to the reference distorted camera, without + +![](images/2aa061e1f3b1d279c41b06c9a76d815d5bb179c54f354d027f7ab109864294f7.jpg) + +![](images/c6ebdda4b96ca1c4b584103dbcc787625ab5aa7337ad65b3365bbc88a63bbb14.jpg) + +![](images/90b6e6e4dac0647685df9de18e11ffe7110160481c8247dc550ae8b6c72ce955.jpg) + +![](images/d527e0c82e6570af51fbfee17fdafdfddd9a5c8536fb6236c4b57216eb38efed.jpg) +input + +![](images/0ede4262645e380177b8da2ff7b63094b8c22e805b43a4e1426c87a8ddd8d80a.jpg) + +![](images/4ad7d1ab580bffbbf343647862cc2175044c73d68d2d9c11699f34174740c7ad.jpg) + +![](images/4e51077ac9c4b3784915e143f89019b9790fe248a62e9d6de4e2124e051e1b08.jpg) + +![](images/25334b9a2a1f07b39dabe9728ac522e5c76bd6915cc01846bcaf37003e91682c.jpg) + +![](images/df345b5980c9ea2a97fe90878b16e2a8b1408ad85e97b9858c57fa64f5128f5a.jpg) + +![](images/daa4e4e443d80cb87179a97255ef9abac301fa224329e32ba5aa123c0f4fd8e3.jpg) + +![](images/67afd2ed9f2666931e534bcb5e0d679dfc7395fe32dccaee519820340a0a5000.jpg) + +![](images/1714555cb57e0eb57c3f0a8daa42c6169172adf96f22e1f1159fa2d08e453008.jpg) +RayDiffusion + +![](images/d56b51ac1b33107d8cacba30d8f3ab511f6d490fcdd20b13aa20c0497267eacc.jpg) + +![](images/c8648f92ffdc827cc0d64d867bf936874ac44d03cb5594ab8ef936f87ff0a11a.jpg) + +![](images/06db193a3419859fdc0b76a010df49e7ad7e14274bb93df0fabeb7376c1d6ed5.jpg) + +![](images/43685977d291fe60d2696ffe701511f5771cf0d9f3d4c5c25e6a605030aad3d3.jpg) +ours +Figure 6. Generalization to out-of-distribution camera from Aria Digital Twins. Our approach produces better results compared to [63]. See supplementary for quantitative inference results. + +employing the DistortionNet. In accordance with [63], we randomly sampled $N$ images for an evaluation of $N$ . Each reported accuracy is averaged over 5 runs to reduce stochasticity. + +Rotation accuracy. We first compute the relative rotations between each pair of cameras for both predicted and ground-truth poses. The errors are computed for the pairwise relative rotations and reported for the percentage of deviations less than $15^{\circ}$ . + +Camera center accuracy. Using the fitted ground truth and predicted poses, we align the cameras and compute the distance from the predicted camera center to the ground truth correspondence. The fraction of distances within $10\%$ of the scene scale is reported. + +Angular error. With the recovered ray bundles, we compute the distance of the predicted rays to the ground truth correspondences. The mean angular error in degrees is reported. + +# 4.3. Evaluation + +Baseline comparisons. In Table 1, we present the mean ray angular error measurements for the distorted CO3D dataset, demonstrating the effectiveness of our approach compared to existing methods. Our results indicate that our method excels at extracting precise, arbitrary ray profiles even in the absence of a prior camera model. This capability is particularly significant given the challenges of distorted sequences. Notably, using the same ray representation as RayDiffusion, our approach consistently surpasses its performance, where the diffusion model depends solely on image-based features. The experiments reveal that our method achieves a more robust disentanglement from local image content, effectively isolating structural cues from object-specific fea + +Table 1. Quantitative evaluation of camera angular error on CO3D and Aria Digital Twins datasets. Averaged angular error (degrees) across varying numbers of input images. + +
CO3D – Seen Categories
Method2345678
RayRegression [63]18.420.621.721.920.323.023.7
RayDiffusion [63]9.69.710.28.79.411.913.2
AlignDiff (Ours)3.43.12.83.23.03.64.1
CO3D – Unseen Categories
RayRegression [63]24.122.521.624.823.925.126.2
RayDiffusion [63]11.410.814.112.315.816.619.4
AlignDiff (Ours)4.65.25.85.76.46.88.2
Aria Digital Twins
RayRegression [63]34.638.236.540.841.240.943.7
RayDiffusion [63]28.429.134.737.240.442.345.6
AlignDiff (Ours)15.413.814.220.621.324.624.9
+ +Table 2. Camera Rotation and Camera Center Accuracy on Distorted CO3D dataset. Experiments using geometrically aberrated videos, comparing recent methods against AlignDiff. + +
Rotation @ 15°
number of Images2345678
Seen ClsRelPose [62]62.024.034.024.025.033.032.0
RelPose++ [33]72.573.473.674.775.076.075.7
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]74.080.085.382.484.882.586.6
AlignDiff (Ours)90.691.191.391.892.492.793.1
Unseen ClsRelPose [62]61.027.037.026.030.026.025.0
RelPose++ [33]61.460.863.064.165.765.765.4
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]69.972.475.376.276.477.178.5
AlignDiff (Ours)83.684.884.985.285.786.186.1
Camera Center @ 0.1
Seen ClsRelPose++ [33]10085.279.074.370.568.666.0
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10073.472.571.171.070.670.2
AlignDiff (Ours)10092.590.788.687.285.884.6
Unseen ClsRelPose++ [33]10066.457.051.348.944.544.1
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10074.871.670.267.366.762.5
AlignDiff (Ours)10086.384.280.575.273.070.1
+ +tures. This disentanglement minimizes residual errors tied to object appearance, allowing structural details to be more accurately emphasized. + +Additionally, we report metrics for camera rotation accuracy and camera center accuracy, with aberrated images in Table 2 and rectified images in Table 3. Our approach consistently outperforms existing methods in estimating both camera position and orientation under aberrated conditions, confirming its resilience to challenging visual distortions. For rectified images, our conditioning method does not simply maintain performance but rather enhances it by recognizing and compensating for residual aberrations. This capability allows our approach to improve predictions even on images with minimal distortion, achieving a high level of precision and adaptability in handling diverse image qualities. + +Table 3. Camera Rotation and Camera Center Accuracy on Rectified CO3D dataset. AlignDiff identifies subtle local distortions, achieving improvements over previous methods. + +
Rotation @ 15°
number of Images2345678
Seen ClsCOLMAP [48]30.728.426.526.827.028.130.6
RelPose [62]56.056.557.057.257.257.357.2
PoseDiffusion [55]75.776.476.877.478.078.778.8
RelPose++ [33]81.882.884.184.784.985.385.5
RayRegression [63]88.888.788.789.089.489.389.2
RayDiffusion [63]91.892.492.692.993.193.393.3
AlignDiff (Ours)92.692.392.793.093.693.493.5
Unseen ClsCOLMAP [48]34.531.831.031.732.735.038.5
RelPose [62]48.647.548.148.348.448.448.3
PoseDiffusion [55]63.264.264.265.766.267.067.7
RelPose++ [33]69.871.171.972.873.874.474.9
RayRegression [63]79.079.680.681.481.381.981.9
RayDiffusion [63]83.585.686.386.987.287.588.1
AlignDiff (Ours)86.386.687.487.788.588.789.2
Camera Center @ 0.1
Seen ClsCOLMAP [48]10034.523.818.915.614.515.0
RelPose [62]10076.566.962.459.458.056.5
PoseDiffusion [55]10077.569.765.963.762.861.9
RelPose++ [33]10085.078.074.271.970.368.8
RayRegression [63]10091.785.782.179.877.976.2
RayDiffusion [63]10094.290.587.886.285.084.1
AlignDiff (Ours)10093.791.887.587.185.283.9
Unseen ClsCOLMAP [48]10036.025.520.017.917.619.1
PoseDiffusion [55]10063.650.545.743.041.239.9
RelPose++ [33]10070.658.853.450.447.846.6
RayRegression [63]10083.775.670.867.465.363.9
RayDiffusion [63]10087.781.177.074.172.471.4
AlignDiff (Ours)10088.281.079.377.572.671.0
+ +Table 4. Camera Rotation and Camera Center Accuracy on Aria Digital Twins dataset. + +
Rotation @ 15°
Number of Images2345678
Unseen ClsRelPose [62]59.632.738.031.228.627.426.0
RelPose++ [33]62.861.362.662.863.264.964.8
PoseDiffusion [55]75.474.874.175.875.275.976.2
RayDiffusion [63]75.276.378.979.279.680.080.2
AlignDiff (Ours)81.482.782.884.384.184.685.5
Camera Center @ 0.1
Unseen ClsRelPose++ [33]10070.356.152.848.646.444.7
PoseDiffusion [55]10071.857.454.651.852.248.5
RayDiffusion [63]10075.969.261.860.455.651.1
AlignDiff (Ours)10078.370.064.462.758.152.9
+ +Qualitative analysis of recovered cameras. We compare the generated cameras with the ground truth cameras in world space, as shown in Figure 7. The predicted ray moments and angular error from ground truth ray directions are also shown. Remarkably, the generated cameras exhibit little rotational difference from the reference cameras. While there are observed to have noticeable translation errors, we argue that this is due to scale ambiguities of the captures. The visualized ray moments present a smooth profile with very low angular errors. We further compare the Aria Digital Twins dataset for the inferred rays in vector form, shown in Figure 6. These further demonstrate that our approach can achieve robust and accurate ray profiles through camera + +Table 5. Ablated framework design evaluated with averaged ray angular error. + +
Methods2345678
Baseline8.89.510.49.212.412.713.9
AlignDiff (w/o angular loss)5.86.18.87.39.611.314.7
AlignDiff (w/o line conditioning)6.78.49.89.910.310.112.8
AlignDiff (w/o edge-attention)5.37.17.49.210.811.310.5
AlignDiff (w/o optics grounding)5.26.96.87.78.49.111.3
AlignDiff4.65.25.85.76.46.88.2
+ +Table 6. Evaluations on different image latent encoder choices. + +
Methods2345678
DINOv2-s/14 [37]8.28.18.610.49.810.112.5
DINOv2-g/14 [37]4.56.46.36.66.17.89.6
TIPS-g/14 [34]4.65.25.85.76.46.88.2
+ +![](images/c8344e3c869e1a55aad0161f1562c8805c73c421908518388b95933041fd532e.jpg) +Figure 7. Recovered cameras from aberrated image sequences. Predicted and expected cameras are visualized in solid and dashed lines, respectively. The cameras align with the rotational orientation of the expected counterparts. The ray moment and residual errors show low levels of deviation from reference rays. + +conditioning. Detailed quantitative evaluation on the Aria Digital Twins dataset are included in Appendix D. + +Undistortion using predicted flow. We conduct image undistortion with the predicted aberration map, as shown in Figure 4. Since we adopt a unified representation for aberrations, the radial fisheye camera model's vignetting effect leads to mapping artifacts. The predicted aberration map overall correctly captures the image distortions and reliably remaps the images regardless of the aberration types. + +Ablation study. We perform an ablation study on CO3D data to evaluate the impact of each conditioning module, enhanced geometric guidance, and additional losses, as shown in Table 5. For each comparison, we remove one conditioning module from the framework. The results show that all modules significantly improve performance, and latent conditioning of the line segment has the greatest effect. Figure 4 visualizes the ray residuals with and without aberration conditioning, further validating its effectiveness. Additionally, we test the angular loss, finding that our localized ray angular loss yields notable improvements. + +While our multiview diffusion approach implicitly incorporates temporal consistency by explicit temporal losses, we chose to handle temporal consistency as a post-processing step to keep the model lightweight and directly applicable in real-world deployments. This approach reduces computational overhead while still providing meaningful consistency improvements, leaving more complex temporal modeling for future work. This is based on the premise that camera profiles remain generally consistent over short periods, barring changes due to factors like camera sensor temperature fluctuations. To test this, we implemented a post-processing step that rejects low-confidence pose estimations, which reduced the average ray angular error from $8.20^{\circ}$ to $8.06^{\circ}$ over 8 frames. Furthermore, enforcing temporal consistency on local camera ray profiles using reprojection mean squared error loss yielded an average reduction of $0.06^{\circ}$ in angular error across the frames. We plan to integrate these enhancements, along with outlier rejection and a frame selection mechanism, into our future work. + +Evaluations on image representation models. The image sequence encoder is designed to generate dense global embeddings that capture perceptual, geometric, and semantic representations. Among the available models, DINOv2 and TIPS are well suited for extracting high-quality dense features. We conduct an ablation study to evaluate these models' effectiveness in querying camera ray profiles. As shown in Table 6, the TIPS-g/14 model achieves the lowest angular errors on CO3D, with DINOv2-g/14 as a close competitor. + +# 5. Conclusion + +We introduce AlignDiff, a unified calibration framework for real-world optical aberrations, enabling the recovery of camera ray bundles, extrinsics, and aberration profiles from multi-view image sequences using simple conditioning techniques. First, we propose structural conditioning to separate image geometry from semantic features, promoting emphasis on structural cues for accurate ray profile estimation. Second, we aggregate features along edges, ensuring that high-quality image embeddings are paired with effective guidance to avoid semantic interference. Finally, we incorporate real-world lens profiles into training, grounding AlignDiff in actual camera designs. These contributions collectively enhance generalization and accuracy across diverse distortion profiles, as demonstrated by consistent improvements over strong baselines. + +While AlignDiff performs well in controlled indoor settings, ray estimation may become unstable in arbitrary outdoor scenes due to scale ambiguity. To enhance robustness in such cases, integrating geometric priors like optical flow or combining traditional camera parameter estimation with network inference may be promising for future directions. + +# References + +[1] Diffusionsfm: Predicting structure and motion via ray origin and endpoint diffusion, 2024. [Online; accessed 2025-03-06]. 1 +[2] Tommaso Cavallari, Stuart Golodetz, Nicholas A. Lord, Julien Valentin, Victor A. Prisacariu, Luigi Di Stefano, and Philip H.S. Torr. Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2465-2477, 2020. 3 +[3] Shuai Chen, Xinghui Li, Zirui Wang, and Victor Adrian Prisacariu. DFNet: Enhance Absolute Pose Regression with Direct Feature Matching. 2022. 2, 3 +[4] Lambda Research Corp. Lensview optical design database. [Online; accessed 2025-03-07]. 5 +[5] Andrei Cramariuc, Aleksandar Petrov, Rohit Suri, Mayank Mittal, Roland Siegwart, and Cesar Cadena. Learning camera miscalibration detection. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 4997-5003, 2020. 3 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-Supervised Interest Point Detection and Description. 2017. 3 +[7] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 1 +[8] Satoshi Fujimoto and Nobutomo Matsunaga. Deep feature-based rgb-d odometry using superpoint and superglue. *Proceeding Computer Science*, 227:1127–1134, 2023. 3 +[9] Fabrizio Gentile, Crescenzo Tortora, Giovanni Covone, Léon V. E. Koopmans, Rui Li, Laura Leuzzi, and Nicola R. Napolitano. *LeMoN: Lens Modelling with Neural networks* - I. Automated modelling of strong gravitational lenses with Bayesian Neural Networks. 2022. 2 +[10] Behnam Gholami, Mostafa El-Khamy, and Kee-Bong Song. Latent feature disentanglement for visual domain generalization. IEEE Transactions on Image Processing, 2023. 3 +[11] Diandian Guo, Deng-Ping Fan, Tongyu Lu, Christos Sakaridis, and Luc Van Gool. Vanishing-point-guided video semantic segmentation of driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3544-3553, 2024. 3 +[12] Lanqing Guo, Renjie Wan, Wenhan Yang, Alex C Kot, and Bihan Wen. Cross-image disentanglement for low-light enhancement in real world. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2550-2563, 2023. 3 +[13] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4 + +[15] Xiankang He, Guangkai Xu, Bo Zhang, Hao Chen, Ying Cui, and Dongyan Guo. DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation. 2024. 2, 3 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Technical report. 4, 6 +[17] Yannick Hold-Geoffroy, Dominique Piche-Meunier, Kalyan Sunkavalli, Jean-Charles Bazin, François Rameau, and Jean-François Lalonde. A perceptual measure for deep single image camera and lens calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10603-10614, 2023. 3 +[18] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 3 +[19] ISO 17850:2015(en). Photography — Digital cameras — Geometric distortion (GD) measurements. Standard, International Organization for Standardization, 2015. 5 +[20] Y. Y. Jau, R. Zhu, H. Su, and M. Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4950-4957, 2020. 3 +[21] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable Feature Matching with Foundation Model Guidance. Technical report. 3 +[22] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-View Object Reconstruction with Unknown Categories and Camera Poses. 2022. 3 +[23] Yang Jiao, Trac D. Tran, and Guangming Shi. Effiscene: Efficient per-pixel rigidity inference for unsupervised joint learning of optical flow, depth, camera pose and motion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5538-5547, 2021. 3 +[24] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F. Fouhey. Perspective fields for single image camera calibration. In CVPR, 2023. 3 +[25] Alex Kendall, Matthew Koichi Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. 2015 IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015. +[26] Julius Kümmerle and Tilman Kühner. Unified intrinsic and extrinsic camera and lidar calibration under uncertainties. In 2020 IEEE International conference on Robotics and Automation (ICRA), pages 6028-6034. IEEE, 2020. 3 +[27] Jinwoo Lee, Hyunsung Go, Hyunjoon Lee, Sunghyun Cho, Minhyuk Sung, and Junho Kim. Ctrl-c: Camera calibration transformer with line-classification. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16208-16217, 2021. 3 +[28] Hao Li, Huai Yu, Jinwang Wang, Wen Yang, Lei Yu, and Sebastian Scherer. Ulsd: Unified line segment detection across pinhole, fisheye, and spherical cameras. ISPRS Journal of Photogrammetry and Remote Sensing, 178:187-202, 2021. + +[29] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4855-4864, 2019. 2, 4 +[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[31] Kang Liao, Lang Nie, Shujuan Huang, Chunyu Lin, Jing Zhang, Yao Zhao, Moncef Gabbouj, and Dacheng Tao. Deep learning for camera calibration and beyond: A survey. arXiv preprint arXiv:2303.10559, 2023. 3 +[32] Kang Liao, Zongsheng Yue, Zhonghua Wu, and Chen Change Loy. Mowa: Multiple-in-one image warping model. ArXiv, abs/2404.10716, 2024. 2, 3, 4 +[33] Amy Lin, Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. RelPose++: Recovering 6D Poses from Sparse-view Observations. 2023. 3, 6, 7 +[34] Kevis-Kokitsi Maninis, Kaifeng Chen, Soham Ghosh, Arjun Karpur, Koert Chen, Ye Xia, Bingyi Cao, Daniel Salz, Guangxing Han, Jan Dlabal, et al. Tips: Text-image pretraining with spatial awareness. arXiv preprint arXiv:2410.16512, 2024. 4, 6, 8 +[35] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In Advances in Neural Information Processing Systems, pages 67960–67971. Curran Associates, Inc., 2023. 4 +[36] Inc. DBA OptiTrack NaturalPoint. Optitrack system, 2024. 6 +[37] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 8 +[38] Linfei Pan, Marc Pollefeys, and Viktor Larsson. Camera pose estimation using implicit distortion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12819-12828, 2022. 3 +[39] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 6 +[40] Chethan M. Parameshwara, Gokul Hari, Cornelia Fermüller, Nitin J. Sanket, and Yiannis Aloimonos. Diffposenet: Direct differentiable camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6845-6854, 2022. 3 +[41] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. ControlNeXt: Powerful and Efficient Control for Image and Video Generation. 2024. 4 + +[42] Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Martins, and Erickson R. Nascimento. XFeat: Accelerated Features for Lightweight Image Matching. 2024. 3 +[43] Srikumar Ramalingam and Peter Sturm. A Unifying Model for Camera Calibration: A Unifying Model for Camera Calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence: A Unifying Model for Camera Calibration. (7): 1309-1319, 2017. 2 +[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction. Technical report. 6 +[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. 2018. 2, 3 +[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3 +[47] José Sasán. Introduction to aberrations in optical imaging systems. Cambridge University Press, 2012. 3 +[48] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. Technical report. 7 +[49] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 6 +[50] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6 +[51] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In arXiv, 2024. 3 +[52] Peter Sturm and Srikumar Ramalingam. A Generic Concept for Camera Calibration. pages 1-13, 2004. 2 +[53] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement. Technical report. 3 +[54] Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys. GeoCalib: Single-image Calibration with Geometric Optimization. In ECCV, 2024. 3 +[55] Jianyuan Wang, Christian Rupprecht, and David Novotny. PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment. 2023. 1, 2, 3, 4, 6, 7 +[56] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 4 +[57] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. 2023. 2, 3 +[58] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. 2020. 6 + +[59] Scott Workman, Connor Greenwell, Menghua Zhai, Ryan Baltenberger, and Nathan Jacobs. Deepfocal: A method for direct focal length estimation. In 2015 IEEE International Conference on Image Processing (ICIP), pages 1369-1373, 2015. 2 +[60] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Innovating Real Fisheye Image Correction with Dual Diffusion Architecture. Technical report. 2 +[61] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 +[62] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. page 592-611, Berlin, Heidelberg, 2022. Springer-Verlag. 7 +[63] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as Rays: Pose Estimation via Ray Diffusion. 2024. 1, 2, 3, 4, 6, 7 +[64] Yinqiang Zheng, Yubin Kuang, Shigeki Sugimoto, Kalle Åström, and Masatoshi Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In 2013 IEEE International Conference on Computer Vision, pages 2344-2351, 2013. 2 +[65] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[66] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a Wild Camera: In-the-Wild Monocular Camera Calibration. 2023. 2, 3 \ No newline at end of file diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/images.zip b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..414934614a23633bfe26577d0c5d953ebb1eee5b --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba63cc1313caf4a91fe247775394d09353b2fdc3976431dca99339c114fb97a0 +size 729950 diff --git a/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/layout.json b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0304442523f8142831f2e49da6e166747cafc29c --- /dev/null +++ b/2025/AlignDiff_ Learning Physically-Grounded Camera Alignment via Diffusion/layout.json @@ -0,0 +1,9033 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 536, + 121 + ], + "type": "text", + "content": "AlignDiff: Learning Physically-Grounded Camera Alignment via Diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "spans": [ + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": "Liyue Xie" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": " Jiancong Guo" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": " Ozan Cakmakci" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": " Andre Araujo" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": " Laszlo A. Jeni" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "text", + "content": " Zhiheng Jia" + }, + { + "bbox": [ + 139, + 142, + 470, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "spans": [ + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "text", + "content": " Carnegie Mellon University " + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "text", + "content": " Google " + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 184, + 461, + 200 + ], + "type": "text", + "content": " Google DeepMind" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 553 + ], + "type": "text", + "content": "Accurate camera calibration is a fundamental task for 3D perception, especially when dealing with real-world, in-the-wild environments where complex optical distortions are common. Existing methods often rely on pre-rectified images or calibration patterns, which limits their applicability and flexibility. In this work, we introduce a novel framework that addresses these challenges by jointly modeling camera intrinsic and extrinsic parameters using a generic ray camera model. Unlike previous approaches, AlignDiff shifts focus from semantic to geometric features, enabling more accurate modeling of local distortions. We propose AlignDiff, a diffusion model conditioned on geometric priors, enabling the simultaneous estimation of camera distortions and scene geometry. To enhance distortion prediction, we incorporate edge-aware attention, focusing the model on geometric features around image edges, rather than semantic content. Furthermore, to enhance generalizability to real-world captures, we incorporate a large database of ray-traced lenses containing over three thousand samples. This database characterizes the distortion inherent in a diverse variety of lens forms. Our experiments demonstrate that the proposed method significantly reduces the angular error of estimated ray bundles by " + }, + { + "bbox": [ + 55, + 252, + 297, + 553 + ], + "type": "inline_equation", + "content": "\\sim 8.2^{\\circ}" + }, + { + "bbox": [ + 55, + 252, + 297, + 553 + ], + "type": "text", + "content": " and overall calibration accuracy, outperforming existing approaches on challenging, real-world datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": "Accurate camera calibration, involving the estimation of intrinsic parameters such as focal length and lens distortions and extrinsic parameters such as camera pose, is essential for robust 3D perception in real-world environments. However, prevailing methods typically address these components independently, focusing on either intrinsic calibration or pose estimation while relying on simplified camera models that are insufficient to capture complex real-world opti" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 228, + 542, + 334 + ], + "blocks": [ + { + "bbox": [ + 328, + 228, + 542, + 334 + ], + "lines": [ + { + "bbox": [ + 328, + 228, + 542, + 334 + ], + "spans": [ + { + "bbox": [ + 328, + 228, + 542, + 334 + ], + "type": "image", + "image_path": "6801c14600fa42551e31acef456bcafdd21bf9c38d3cc60822f075076ce47f36.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 340, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 407 + ], + "type": "text", + "content": "Figure 1. AlignDiff is proposed to address common image geometric aberrations with a unified ray camera representation while jointly recovering the camera extrinsics. With groundings on physical camera lens designs, as well as the disassociation of geometric cues from semantic features, it demonstrates an ability to generalize to real video sequences." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 485 + ], + "type": "text", + "content": "cal aberrations. This separation is fundamentally limiting, as intrinsic and extrinsic parameters are closely interdependent; inaccuracies in modeling lens distortions propagate directly to pose estimation errors. Consequently, joint optimization of intrinsic and extrinsic parameters is critical to ensure reliable calibration under unconstrained conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 640 + ], + "type": "text", + "content": "Recent diffusion-based and transformer-powered models, such as PoseDiffusion [55], RayDiffusion [63], and DiffusionSfM [1] have improved calibration accuracy and zero-shot generalizability by conditioning on image features extracted from vision encoders. However, these methods typically focus on high-level, semantic image features rather than on structural cues critical for modeling optical aberrations that directly impact geometry. This is particularly limiting because accurate intrinsic calibration is essential for precise extrinsic calibration: aberrations in the intrinsic parameters can significantly affect the accuracy of extrinsic parameters, making the two tightly interdependent." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 712 + ], + "type": "text", + "content": "We propose AlignDiff, shown Figure 1, a diffusion-based calibration framework, to learn the fine-grained ray profile from video sequences in world space, with a conditioning strategy that better captures fine-grained optical distortions. To shift focus from semantic to structural features, we condition the diffusion model with line embed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 209, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 209, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 209, + 713 + ], + "type": "text", + "content": "*Work done during an internship at Google." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26901" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 72, + 508, + 268 + ], + "blocks": [ + { + "bbox": [ + 105, + 72, + 508, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 508, + 268 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 508, + 268 + ], + "type": "image", + "image_path": "1869ee0aa8bc9062b2afa671dd171f8fd9fc04f9e99abe03d1d13e436662b775.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 277, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 277, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. AlignDiff Architecture. We promote learning camera ray profiles in three main steps: geometric cue conditioning from line features, edge-aware attention, and physical camera groundings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 321, + 294, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 321, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 321, + 294, + 392 + ], + "type": "text", + "content": "dings from a line detection network. This approach emphasizes the geometric structures, helping the model prioritize optical aberrations over object content. In addition, we further introduce edge-focused attention to further enhance the model's sensitivity to regions around edges, where aberrations often manifest most prominently." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 393, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 295, + 476 + ], + "type": "text", + "content": "An additional limitation of prior models is their reliance on simulated optical aberrations derived from predefined models, which often lack the subtlety of real-world aberrations, such as localized or asymmetric profiles. To address this, our framework incorporates real optical profiles from actual lens designs, grounding the model in authentic optical characteristics and boosting its generalizability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 477, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 477, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 54, + 477, + 295, + 572 + ], + "type": "text", + "content": "In summary, our framework builds upon existing diffusion-based calibration techniques by incorporating targeted conditioning strategies and real-world optical data, providing a robust solution that accurately captures complex optical aberrations. By effectively addressing intrinsic aberrations, our model lays a solid foundation for accurate extrinsic calibration, making it highly applicable for in-the-wild scenarios requiring precise camera parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 573, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 295, + 608 + ], + "type": "text", + "content": "- To our knowledge, AlignDiff is the first unified diffusion-based approach addressing complex optical aberrations in joint intrinsic and extrinsic calibration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 607, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 666 + ], + "type": "text", + "content": "- By conditioning the diffusion model on line embeddings and incorporating edge-focused attention, our method prioritizes structural over semantic features, achieving generalization to diverse, real-world environments without requiring extensive retraining." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "- Our approach incorporates authentic optical profiles from actual lens designs, grounding the model in real-world aberrations and improving generalization to natural images." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 319, + 406, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 406, + 332 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 406, + 332 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 341, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 484 + ], + "type": "text", + "content": "Aberration representation for in-the-wild images. Traditional camera calibration methods model intrinsic characteristics, such as focal length and camera center, alongside distortion parameters that describe image geometric aberrations. Early deep learning-based methods [3, 9, 15, 45, 57, 59, 60] enabled calibration from images. Recent diffusion based approaches like PoseDiffusion [55], extend from prior approaches with improved zero-shot performance on wild captures. While effective for rectified images, these methods are limited in their ability to account for local optical aberrations resulting from factors like wear, temperature fluctuations, and lens designs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 653 + ], + "type": "text", + "content": "Generic ray camera models were introduced to better capture local geometric distortions [43, 52] by modeling ray bundles across image patches, offering a more detailed representation of pixel deviations compared to the pinhole model. While accurate, they require denser calibration patterns, especially when optimized using methods like PnP [64]. Several recent work [15, 63, 66] adopted the ray representation in deep learning and proved its feasibility. Yet, these prior works remain largely restricted to the pinhole model and did not exploit the expressiveness of ray representation for aberration modeling. The recent blind camera undistortion line of work [29, 32] infers geometric displacements from a single image, often as a preprocessing step to provide rectified images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Our work advances these approaches by introducing a unified ray-based framework that models both pinhole projections and local distortions. We directly estimate the ray profile and camera aberrations from a raw monocular sequence, enhancing generalizability in diverse real-world" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26902" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 102, + 83 + ], + "type": "text", + "content": "conditions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 203 + ], + "type": "text", + "content": "Joint Intrinsic and Extrinsic Calibration Classical camera calibration methods often assume a predefined camera model for each type of aberration, relying on calibration patterns and multi-view geometric constraints to estimate camera parameters from a sequence of images [13, 24-26, 33, 54, 55]. Early approaches solved joint intrinsic and extrinsic calibration through point matching across views [2, 3, 11, 22, 38, 49, 65], with refinements summarized in [31]. Structure-from-motion (SfM) techniques later optimized reprojection loss to refine parameters further." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 418 + ], + "type": "text", + "content": "For in-the-wild images, Hold-Geoffroy et al. [17] leveraged DenseNet [18] to estimate parameters such as horizon angle and vertical field of view. Other methods [20, 24] decouple parameter estimation from direct regression, and with integrated semantic or geometric guidance [6, 8, 21, 46]. Further advancements have shown that models pretrained with geometric understanding can act as better feature encoders [5, 15, 23, 40, 42, 45, 51, 53, 61]. Recent methods, including PoseDiffusion [55], DuSt3R [57], RayDiffusion [63] use diffusion and transformer to recover camera parameters or ray profiles directly from images, though they still assume fixed distortion models, limiting generalization to diverse real-world aberrations [31, 32]. The camera parameters are recovered by conditioning on local image features to capture local and cross-view geometric associations. However, these methods can yield residual errors around object contours due to reliance on semantic rather than purely geometric cues [10, 12, 27]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 491 + ], + "type": "text", + "content": "Our approach extends the ray representation without a fixed distortion model, disentangling semantic and geometric cues to reduce contour errors and enhance calibration accuracy. By integrating undistortion and calibration in a unified framework, we increase flexibility and applicability for real-world captures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 500, + 119, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 119, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 119, + 514 + ], + "type": "text", + "content": "3. AlignDiff" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 529, + 128, + 625 + ], + "blocks": [ + { + "bbox": [ + 56, + 529, + 128, + 625 + ], + "lines": [ + { + "bbox": [ + 56, + 529, + 128, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 128, + 625 + ], + "type": "image", + "image_path": "99a9178ef9bf3769274b010533bfbb83eb39affe781463e199b1868f19510f13.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 634, + 295, + 689 + ], + "lines": [ + { + "bbox": [ + 55, + 634, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 295, + 689 + ], + "type": "text", + "content": "Figure 3. The latents are reweighted through edge-attention that aggregates a learned mask to the Query feature to promote information along edges where the geometric cues are more prominent. Following, the multi-view attention further captures features from different views." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 130, + 537, + 294, + 625 + ], + "blocks": [ + { + "bbox": [ + 130, + 537, + 294, + 625 + ], + "lines": [ + { + "bbox": [ + 130, + 537, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 130, + 537, + 294, + 625 + ], + "type": "image", + "image_path": "0f254a720e0a4f370a862b472a17198001356c9e908487e19dd9fa2be351b9bc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "We aim to recover generic cameras represented as ray" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "bundles in world space from a set of " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " input images " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_1,\\mathcal{I}_2,\\mathcal{I}_3,\\dots ,\\mathcal{I}_N\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": ". Our approach, as presented in Figure 2, captures fine-grained geometric distortions while maintaining compatibility with classical parameterized camera models. This formulation is conceptually similar to works like [63, 66], but extends the ray camera model to account for local distortions (Sec. 3.1). In contrast, previous methods assume a pinhole camera model, neglecting distortions, even though they predict rays for local image patches. We adopt a Diffusion Transformer as the base architecture while enhancing this architecture by introducing geometric cues that guide ray directions based on local geometry, disentangling image perceptual features from distortions (Sec. 3.2). From the predicted rays, we propose DistortionNet to recover the underlying lens geometric aberration that describes the ray profile's deviation from a perspective pinhole camera with the same focal length and field of view (Sec. 3.3). We further discuss the selected optimization objectives in Sec. 3.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 510, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 510, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 510, + 318 + ], + "type": "text", + "content": "3.1. Generic Ray Camera Representation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": "The objective of camera calibration is to recover optimal parameters that describe the projection of 3D world points " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": " to 2D sensor locations " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ". Our method models this with a ray-based camera model, where each unit of ray " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{r}_i \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": " is stored on a grid centered on image patches. Rays originate from the camera center " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{r}_o \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", with directions " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\pmb{r}_d \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": " derived by lifting 2D grid points to the camera frame using:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 395, + 422, + 553, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 422, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 395, + 422, + 553, + 436 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {\\boldsymbol {d}} = \\boldsymbol {K} ^ {- 1} \\mathcal {D} _ {\\zeta} (x), \\tag {1}", + "image_path": "0ca42599536dbd9ed5020606a90483162309e50fbb54bad6c9a81cd7b0fb701b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\zeta}" + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": " is the aberration function parameterized by " + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{K}" + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": " is the intrinsic matrix, accounting for geometric aberrations. The ray bundle, transformed to world coordinates by rotation " + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\pmb{t}" + }, + { + "bbox": [ + 313, + 441, + 554, + 487 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 399, + 492, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 492, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 399, + 492, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {w} = \\boldsymbol {R} ^ {T} \\boldsymbol {r} + \\boldsymbol {t}, \\tag {2}", + "image_path": "7d84172bfd7622671017ac4b5a93c01a76a73bd9c1103015cd44250b143e3f76.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 546 + ], + "type": "text", + "content": "represents aberrated cameras in world coordinates, allowing holistic downstream recovery of rays, camera extrinsics, and optical aberrations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 547, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 554, + 642 + ], + "type": "text", + "content": "Off-the-shelf image quality is completely characterized by the modulation transfer function (MTF) and distortion. The MTF performance includes all monochromatic and chromatic aberrations. Typically, the driving residual aberration in high-quality off-the-shelf lenses is optical distortion " + }, + { + "bbox": [ + 313, + 547, + 554, + 642 + ], + "type": "inline_equation", + "content": "(W_{311}" + }, + { + "bbox": [ + 313, + 547, + 554, + 642 + ], + "type": "text", + "content": " in the wavefront to third order[47]), therefore, in this paper, our ray trace models focus on extracting the optical distortion characteristics from our dataset of lenses." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 648, + 493, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 493, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 493, + 661 + ], + "type": "text", + "content": "3.2. Geometry-controlled ray diffuser" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "We train a multi-view diffusion model " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\phi}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " that takes multiple images of a 3D scene as input and generates the corresponding output camera ray bundles given their geometric aberration cues. Specifically, given " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " conditional views" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26903" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "containing the images and their corresponding geometric aberration latents " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^g" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " from a line-segment detection network, the model learns to capture the joint distribution of " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " target ray bundles " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{tgt} = \\{\\mathcal{R}_1,\\mathcal{R}_2,\\mathcal{R}_3,\\dots ,\\mathcal{R}_N\\}^{tgt}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " from noised ray bundles " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^\\epsilon" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 134, + 142, + 295, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 142, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 134, + 142, + 295, + 156 + ], + "type": "interline_equation", + "content": "p \\left(\\mathcal {R} ^ {t g t} | \\mathcal {I}, \\mathcal {Z} ^ {g}, \\mathcal {R} ^ {\\epsilon}\\right). \\tag {3}", + "image_path": "2506f8aebac1bbb7cee5c87039c6427271fd5076e629f469a7277868ef6f65e8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 165, + 296, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 237 + ], + "type": "text", + "content": "Diffusion model architecture. Diffusion models approximate the data manifold by learning to invert a diffusion process from data to a presumed distribution through a denoising process. Our adopted Denoising Diffusion Probabilistic Model (DDPM) specifically defines the noise distribution to be Gaussian and transitions to the next noising step:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 247, + 295, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 247, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 91, + 247, + 295, + 262 + ], + "type": "interline_equation", + "content": "q (x _ {t} | x _ {t - 1}) = \\mathcal {N} (x _ {t}; \\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbb {I}), \\tag {4}", + "image_path": "ed34190cdc3cbbbe77841e1654172e1c34dd04c6f6a505eeecb8ee97dd168e4d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\{\\beta_{1},\\beta_{2},\\dots,\\beta_{T}\\}" + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": " are variances within the T step nois-ing schedule, and " + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": " is the identity matrix. Similar to prior works [35, 55, 63], we adopt a DiT [16] as the backbone of the diffusion model, receiving image latent embeddings conditioned on aberration cues. Given a set of sequential images and target ray bundles, the model encodes each image into a latent representation " + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_i^c" + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": " through an image feature encoder TIPS [34], as well as a conditioning branch comprised of multiple ResNet [14] blocks. The image latent feature is concatenated to a noised ray bundle of the same dimension. Then, the diffusion model is trained to estimate the joint distribution of the latent-conditioned ray bundle given the conditioning geometric cues. We initialize the model from a DiT model trained for 3D shape generation, with an input resolution of " + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "inline_equation", + "content": "448\\times 448\\times 3" + }, + { + "bbox": [ + 55, + 270, + 296, + 498 + ], + "type": "text", + "content": ". We directly inflate the latent space of the original DiT to connect with the concatenated features of noised rays and image latents, while inheriting the rest of the trained model parameters to leverage the perceptual knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "Utilizing geometric cues. The conditioning branch is designed to extract guidance information that describes geometric aberration cues and align it with the denoising features. The conditioning branch is a trained line segment detection network [28]. We observe that the controls maintain a high level of consistency with the denoising features and eliminate the need to be inserted at multiple stages, as also indicated in [41]. We thus integrate the controls at a single middle block into the denoising model by adding them to the denoising features after the cross-normalization. It can serve as a plug-and-play guidance module that provides geometric aberration information about the conditioning views, such that the diffuser model is aware of the levels and types of aberration from geometric cues." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Edge Attention. We introduce Edge Attention responsible for parsing features along the edges, as shown in Figure 3, which highlights the edge-aware attention mechanism designed to prioritize geometric cues along image edges. In" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "particular, edge attention takes the input embedding of an image sequence " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}^c\\in \\mathbb{R}^{N\\times H\\times W\\times C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", allowing it to perform self-attention operations across the edges. We apply a patchification operator with patch size " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "p\\times p" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " to generate patch tokens " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb {t}\\in \\mathbb{R}^{L\\times 3}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "L = (N\\times H\\times W / p^2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " denoting the length of patched tokens. To enable the model to disentangle the geometric information from the semantics, we integrate the edge information " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "e_{i\\dots N}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " for guidance into the self-attention mechanism:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 186, + 555, + 213 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 186, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 324, + 186, + 555, + 213 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t e n t i o n} (Q, K, V) = \\operatorname {S o f t m a x} (\\frac {(Q + e ^ {T}) K}{\\sqrt {H}} V) \\quad (5)", + "image_path": "65a935fd066cfb164cf41b879979620fc10118868e0be3012fa56ec3229c2b78.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "text", + "content": " is the dimension size of each head. The edge embedding is designed as a soft weighting to highlight the features along the geometrically prominent areas. The image edges are first extracted from the original image sequences using a Canny Edge detector. We then patchify them and apply a 3D convolution, mapping the embedding as " + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "inline_equation", + "content": "e_i \\in \\mathbb{R}^L" + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "text", + "content": ". Ultimately, we obtain the aggregated features " + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Z}}_i" + }, + { + "bbox": [ + 313, + 220, + 555, + 328 + ], + "type": "text", + "content": ", and we employ a feedforward network as in the original DiT before proceeding to the subsequent blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 523, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 523, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 523, + 347 + ], + "type": "text", + "content": "3.3. Aberration profiling with DistortionNet" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": "We adopt warping flow as a unified representation to model geometric aberrations of the image compared to ideal pin-hole cameras, similar to prior blind camera undistortion frameworks [29, 32]. To query the aberration profile, we connect the denoised rays to a network devised to estimate deviations of predicted ray bundles to that of a pinhole camera. The DistortNet " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\theta}" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": ", with a backbone of MAE-Tiny [56], learns the mapping from the predicted set of ray bundles " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{tgt}" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": " to a set of warping flow maps " + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{tgt}" + }, + { + "bbox": [ + 313, + 352, + 555, + 483 + ], + "type": "text", + "content": ". The learned flow maps can be applied to the image sequences for image undistortion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 448, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 448, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 448, + 504 + ], + "type": "text", + "content": "3.4. Optimization objectives" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "text", + "content": "We describe the optimization objectives for estimating ray bundles and extracting the distortion map from " + }, + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_1,\\mathcal{I}_2,\\dots ,\\mathcal{I}_N\\}" + }, + { + "bbox": [ + 313, + 508, + 554, + 604 + ], + "type": "text", + "content": ". Estimating distorted ray bundles from in-the-wild images is under-constrained and affected by uncertainties from limited overlap and artifacts (e.g., motion blur, lighting changes). To address this, we train the network with a denoising process that reverses noised samples to match the reference:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 612, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 612, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 345, + 612, + 555, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d e n o i s e} (\\phi) = \\mathbb {E} _ {t, \\boldsymbol {r} _ {o}, \\epsilon} | | \\boldsymbol {r} _ {o} - \\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t) | | ^ {2}, \\qquad (6)", + "image_path": "ea76935b6c67eb2d791cd75e20f20f04b941ff06754fbc1c72ecb9408497849d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{r}_o" + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": " is the original ray bundle, " + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\phi}" + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": " is the denoising model, and " + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{r}_t" + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": " is the sample with Gaussian noise at time " + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 635, + 554, + 658 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 380, + 667, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 667, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 380, + 667, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {r} _ {o} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {7}", + "image_path": "5246f9fe735d62248d73f386572f3cc11ca49a8e308eb15728041e2e6b82a0d9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbb{I})" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": ". The denoising reverts to a regressive loss, allowing additional geometric constraints." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26904" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 306, + 292 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 306, + 292 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 306, + 292 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 306, + 292 + ], + "type": "image", + "image_path": "80772bc658c3b1f22eef18527756e6b4c1bb5b8e5429969b9ebd6718b325e82f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 302, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 55, + 302, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 555, + 325 + ], + "type": "text", + "content": "Figure 4. Recovered aberration profile and undistorted images. From denoised rays in world space, the DistortNet estimates the aberration pattern represented in warping flow. The undistorted images maintain coherent structural distribution compared to aberration-free images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 321, + 86, + 541, + 277 + ], + "blocks": [ + { + "bbox": [ + 321, + 86, + 541, + 277 + ], + "lines": [ + { + "bbox": [ + 321, + 86, + 541, + 277 + ], + "spans": [ + { + "bbox": [ + 321, + 86, + 541, + 277 + ], + "type": "image", + "image_path": "6f44db766740fd90dc34d63f2cbc7e9f77e4d15e9859debead5069d3ffc21549.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 280, + 538, + 293 + ], + "lines": [ + { + "bbox": [ + 333, + 280, + 538, + 293 + ], + "spans": [ + { + "bbox": [ + 333, + 280, + 538, + 293 + ], + "type": "text", + "content": "(e) predicted ray maps with and without edge conditioning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "type": "text", + "content": "Recovering local camera ray profiles from an image sequence is susceptible to rotational ambiguities, leading to errors in ray bundle and pose estimates. We introduce a ray directional loss to measure angular errors in ray bundles:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 388, + 295, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 388, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 91, + 388, + 295, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a n g u l a r}} (\\phi) = \\cos^ {- 1} \\left(\\frac {r _ {o} \\cdot \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right)}{\\left| r _ {o} \\right| \\left| \\mathcal {M} _ {\\phi} \\left(\\boldsymbol {r} _ {t} , t\\right) \\right|}\\right), \\tag {8}", + "image_path": "a81b29a8a0d58420385f7bbae04440f11cd445d5cf8b60cff9bbf884e7859866.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "text", + "content": "DistortNet is trained to regress backward flow, representing deviations from perspective ray bundles, by comparing each predicted flow map " + }, + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\pmb{f}_o" + }, + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "text", + "content": " to the reference set " + }, + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{tgt}" + }, + { + "bbox": [ + 55, + 417, + 296, + 453 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 457, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 295, + 472 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t o r t}} (\\theta) = \\mathbb {E} _ {t, \\boldsymbol {f} _ {o}, \\boldsymbol {r} _ {t}} | | \\boldsymbol {f} _ {o} - \\mathcal {M} _ {\\theta} \\left(\\mathcal {M} _ {\\phi} (\\boldsymbol {r} _ {t}, t)\\right) | | ^ {2}. \\tag {9}", + "image_path": "763212c5a4327e836eeb7327721569fe4cd216773ab4c66e1ca8afd1726726d1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "content": "This way, we recover the camera distortion maps at each image pixel, which can be used to directly unwarp the images and their associated rays." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 226, + 529 + ], + "type": "text", + "content": "3.5. Physical aberration groundings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 653 + ], + "type": "text", + "content": "We use the largest available lens database to ray trace and extract distortion characteristics from about 3000 optical systems. We wrote custom scripts that extracted each lens prescription from the LensView database and used commercial raytracing software to extract the optical distortion maps. These systems include camera lenses, lithography lenses, and freeform optics, as shown in Fig 5. These systems produce a diverse set of optical distortion functions that map the object to the image. Detailed description on the lens dataset can be found in Appendix A." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Since images within each dataset are usually captured with similar equipment sharing identical camera settings, to learn different camera modalities, we augment the sequences upon loading with sampled camera distortion profiles representing shear, barrel, fisheye, and pincushion. We" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 334, + 553, + 482 + ], + "blocks": [ + { + "bbox": [ + 318, + 334, + 553, + 482 + ], + "lines": [ + { + "bbox": [ + 318, + 334, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 318, + 334, + 553, + 482 + ], + "type": "image", + "image_path": "79469f2eb5ef24c798a707a056b8f0eda675c269adeb76394a2e29d0f832e4ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 490, + 555, + 535 + ], + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 535 + ], + "type": "text", + "content": "Figure 5. Ray-traced lens designs are utilized, enabling the accurate simulation of geometric aberrations. The aberrations are encoded as local geometric distortion [19], with a percentage deviation of pixels from their ideal positions on a regular grid." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "further include a set of diverse optics dataset comprised of 3187 patented lens profiles from LensView [4]. These profiles span a wide range of professional and industrial lenses. Figure 5 shows examples from the grounding lenses, along with the distribution of the distortion with respect to the field of view (FOV), Numerical Aperture, and F-number. This augmentation introduces groundings in intrinsic and distortion profiles, addressing the scarcity of camera profiles within the datasets. We assess the out-of-distribution distortions not present in the lens set through experiments on the Aria Digital Twins dataset in Table 1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26905" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 137, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "In this section, we first provide the implementation details of the proposed framework and then validate on two real world datasets with different camera modalities. Our method outperforms baseline approaches in both aberrated and non-aberrated images, achieving state-of-the-art results in quantitative and qualitative evaluations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 188, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 188, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 188, + 182 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 186, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 270 + ], + "type": "text", + "content": "We use a pre-trained TIPS [34] network as the image feature extractor, which gives high-quality, fine-grained image embeddings. We adopted a DiT [16] as the base structure of the denoising diffuser, and augmented the attention blocks with the proposed edge conditioning attention. We train our diffusion model with " + }, + { + "bbox": [ + 55, + 186, + 296, + 270 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 55, + 186, + 296, + 270 + ], + "type": "text", + "content": " timesteps, with the training taking roughly 3 days on 8 H100 GPUs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 271, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 354 + ], + "type": "text", + "content": "Following prior works [33, 55, 63], we use the first camera as the coordinate anchor to define the scene. The scene is rescaled such that the first camera has a unit from translation and is rotated to grant the first camera identity rotation. In this way, all the following predicted cameras would use the first camera as a reference to express their relative orientations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 227, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 227, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 227, + 373 + ], + "type": "text", + "content": "4.2. Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 510 + ], + "type": "text", + "content": "Datasets. We chose CO3D [44] as the primary training dataset for the generation of ray maps. The dataset consists of roughly " + }, + { + "bbox": [ + 55, + 378, + 295, + 510 + ], + "type": "inline_equation", + "content": "37\\mathrm{k}360^{\\circ}" + }, + { + "bbox": [ + 55, + 378, + 295, + 510 + ], + "type": "text", + "content": " videos of common household objects from 51 MS-COCO categories. The dataset is annotated with COLMAP [49, 50] camera poses and intrinsics, with the frames undistorted with the estimated coefficients. Each video spans over an average 200 frames. The dataset provides a diverse profile of common objects sequences, making it suitable for training the network. In order to enhance the variety of training scenarios, we add MegaDepth [30] and TartanAir [58] as our additional training data." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 677 + ], + "type": "text", + "content": "Secondly, we evaluate the inference performance in Aria Digital Twins [39] that comprises 200 sequences and 400 min of videos that capture the daily activities of the common household using egocentric cameras. Its camera trajectories were provided by Optitrack [36] system and IMU sensor data on the Aria glass, further optimized by matching the two sources of estimated trajectories. The camera calibration parameters are provided by fitting a Kannala Brandt and Fisheye radial-tangential thin film model with the estimated trajectories. We use the same division as in [39], with a subset of 50 tests. With the two chosen datasets, we showcase the model's capabilities of estimating the ray maps from both common household videos and its ability to generalize to recent AR egocentric captures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Evaluation protocol. We evaluated spare view predictions for recovered poses and the accuracy of the camera ray profile compared to the reference distorted camera, without" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 73, + 381, + 121 + ], + "blocks": [ + { + "bbox": [ + 331, + 73, + 381, + 121 + ], + "lines": [ + { + "bbox": [ + 331, + 73, + 381, + 121 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 381, + 121 + ], + "type": "image", + "image_path": "2aa061e1f3b1d279c41b06c9a76d815d5bb179c54f354d027f7ab109864294f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 122, + 381, + 168 + ], + "blocks": [ + { + "bbox": [ + 331, + 122, + 381, + 168 + ], + "lines": [ + { + "bbox": [ + 331, + 122, + 381, + 168 + ], + "spans": [ + { + "bbox": [ + 331, + 122, + 381, + 168 + ], + "type": "image", + "image_path": "c6ebdda4b96ca1c4b584103dbcc787625ab5aa7337ad65b3365bbc88a63bbb14.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 331, + 170, + 381, + 217 + ], + "blocks": [ + { + "bbox": [ + 331, + 170, + 381, + 217 + ], + "lines": [ + { + "bbox": [ + 331, + 170, + 381, + 217 + ], + "spans": [ + { + "bbox": [ + 331, + 170, + 381, + 217 + ], + "type": "image", + "image_path": "90b6e6e4dac0647685df9de18e11ffe7110160481c8247dc550ae8b6c72ce955.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 331, + 218, + 381, + 264 + ], + "blocks": [ + { + "bbox": [ + 331, + 218, + 381, + 264 + ], + "lines": [ + { + "bbox": [ + 331, + 218, + 381, + 264 + ], + "spans": [ + { + "bbox": [ + 331, + 218, + 381, + 264 + ], + "type": "image", + "image_path": "d527e0c82e6570af51fbfee17fdafdfddd9a5c8536fb6236c4b57216eb38efed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 266, + 365, + 276 + ], + "lines": [ + { + "bbox": [ + 346, + 266, + 365, + 276 + ], + "spans": [ + { + "bbox": [ + 346, + 266, + 365, + 276 + ], + "type": "text", + "content": "input" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 383, + 73, + 433, + 121 + ], + "blocks": [ + { + "bbox": [ + 383, + 73, + 433, + 121 + ], + "lines": [ + { + "bbox": [ + 383, + 73, + 433, + 121 + ], + "spans": [ + { + "bbox": [ + 383, + 73, + 433, + 121 + ], + "type": "image", + "image_path": "0ede4262645e380177b8da2ff7b63094b8c22e805b43a4e1426c87a8ddd8d80a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 383, + 122, + 433, + 168 + ], + "blocks": [ + { + "bbox": [ + 383, + 122, + 433, + 168 + ], + "lines": [ + { + "bbox": [ + 383, + 122, + 433, + 168 + ], + "spans": [ + { + "bbox": [ + 383, + 122, + 433, + 168 + ], + "type": "image", + "image_path": "4ad7d1ab580bffbbf343647862cc2175044c73d68d2d9c11699f34174740c7ad.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 170, + 433, + 217 + ], + "blocks": [ + { + "bbox": [ + 383, + 170, + 433, + 217 + ], + "lines": [ + { + "bbox": [ + 383, + 170, + 433, + 217 + ], + "spans": [ + { + "bbox": [ + 383, + 170, + 433, + 217 + ], + "type": "image", + "image_path": "4e51077ac9c4b3784915e143f89019b9790fe248a62e9d6de4e2124e051e1b08.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 383, + 218, + 433, + 264 + ], + "blocks": [ + { + "bbox": [ + 383, + 218, + 433, + 264 + ], + "lines": [ + { + "bbox": [ + 383, + 218, + 433, + 264 + ], + "spans": [ + { + "bbox": [ + 383, + 218, + 433, + 264 + ], + "type": "image", + "image_path": "25334b9a2a1f07b39dabe9728ac522e5c76bd6915cc01846bcaf37003e91682c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 266, + 433, + 276 + ], + "lines": [ + { + "bbox": [ + 420, + 266, + 433, + 276 + ], + "spans": [ + { + "bbox": [ + 420, + 266, + 433, + 276 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 435, + 73, + 486, + 121 + ], + "blocks": [ + { + "bbox": [ + 435, + 73, + 486, + 121 + ], + "lines": [ + { + "bbox": [ + 435, + 73, + 486, + 121 + ], + "spans": [ + { + "bbox": [ + 435, + 73, + 486, + 121 + ], + "type": "image", + "image_path": "df345b5980c9ea2a97fe90878b16e2a8b1408ad85e97b9858c57fa64f5128f5a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 435, + 122, + 486, + 168 + ], + "blocks": [ + { + "bbox": [ + 435, + 122, + 486, + 168 + ], + "lines": [ + { + "bbox": [ + 435, + 122, + 486, + 168 + ], + "spans": [ + { + "bbox": [ + 435, + 122, + 486, + 168 + ], + "type": "image", + "image_path": "daa4e4e443d80cb87179a97255ef9abac301fa224329e32ba5aa123c0f4fd8e3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 435, + 170, + 486, + 217 + ], + "blocks": [ + { + "bbox": [ + 435, + 170, + 486, + 217 + ], + "lines": [ + { + "bbox": [ + 435, + 170, + 486, + 217 + ], + "spans": [ + { + "bbox": [ + 435, + 170, + 486, + 217 + ], + "type": "image", + "image_path": "67afd2ed9f2666931e534bcb5e0d679dfc7395fe32dccaee519820340a0a5000.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 435, + 218, + 486, + 264 + ], + "blocks": [ + { + "bbox": [ + 435, + 218, + 486, + 264 + ], + "lines": [ + { + "bbox": [ + 435, + 218, + 486, + 264 + ], + "spans": [ + { + "bbox": [ + 435, + 218, + 486, + 264 + ], + "type": "image", + "image_path": "1714555cb57e0eb57c3f0a8daa42c6169172adf96f22e1f1159fa2d08e453008.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 266, + 485, + 276 + ], + "lines": [ + { + "bbox": [ + 437, + 266, + 485, + 276 + ], + "spans": [ + { + "bbox": [ + 437, + 266, + 485, + 276 + ], + "type": "text", + "content": "RayDiffusion" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 488, + 73, + 539, + 121 + ], + "blocks": [ + { + "bbox": [ + 488, + 73, + 539, + 121 + ], + "lines": [ + { + "bbox": [ + 488, + 73, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 488, + 73, + 539, + 121 + ], + "type": "image", + "image_path": "d56b51ac1b33107d8cacba30d8f3ab511f6d490fcdd20b13aa20c0497267eacc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 488, + 122, + 539, + 168 + ], + "blocks": [ + { + "bbox": [ + 488, + 122, + 539, + 168 + ], + "lines": [ + { + "bbox": [ + 488, + 122, + 539, + 168 + ], + "spans": [ + { + "bbox": [ + 488, + 122, + 539, + 168 + ], + "type": "image", + "image_path": "c8648f92ffdc827cc0d64d867bf936874ac44d03cb5594ab8ef936f87ff0a11a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 488, + 170, + 539, + 217 + ], + "blocks": [ + { + "bbox": [ + 488, + 170, + 539, + 217 + ], + "lines": [ + { + "bbox": [ + 488, + 170, + 539, + 217 + ], + "spans": [ + { + "bbox": [ + 488, + 170, + 539, + 217 + ], + "type": "image", + "image_path": "06db193a3419859fdc0b76a010df49e7ad7e14274bb93df0fabeb7376c1d6ed5.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 488, + 218, + 539, + 264 + ], + "blocks": [ + { + "bbox": [ + 488, + 218, + 539, + 264 + ], + "lines": [ + { + "bbox": [ + 488, + 218, + 539, + 264 + ], + "spans": [ + { + "bbox": [ + 488, + 218, + 539, + 264 + ], + "type": "image", + "image_path": "43685977d291fe60d2696ffe701511f5771cf0d9f3d4c5c25e6a605030aad3d3.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 266, + 519, + 275 + ], + "lines": [ + { + "bbox": [ + 503, + 266, + 519, + 275 + ], + "spans": [ + { + "bbox": [ + 503, + 266, + 519, + 275 + ], + "type": "text", + "content": "ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 319 + ], + "type": "text", + "content": "Figure 6. Generalization to out-of-distribution camera from Aria Digital Twins. Our approach produces better results compared to [63]. See supplementary for quantitative inference results." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "text", + "content": "employing the DistortionNet. In accordance with [63], we randomly sampled " + }, + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "text", + "content": " images for an evaluation of " + }, + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 329, + 554, + 376 + ], + "type": "text", + "content": ". Each reported accuracy is averaged over 5 runs to reduce stochasticity." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": "Rotation accuracy. We first compute the relative rotations between each pair of cameras for both predicted and ground-truth poses. The errors are computed for the pairwise relative rotations and reported for the percentage of deviations less than " + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 313, + 377, + 554, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": "Camera center accuracy. Using the fitted ground truth and predicted poses, we align the cameras and compute the distance from the predicted camera center to the ground truth correspondence. The fraction of distances within " + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 437, + 554, + 496 + ], + "type": "text", + "content": " of the scene scale is reported." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 497, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 533 + ], + "type": "text", + "content": "Angular error. With the recovered ray bundles, we compute the distance of the predicted rays to the ground truth correspondences. The mean angular error in degrees is reported." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 540, + 389, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 389, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 389, + 552 + ], + "type": "text", + "content": "4.3. Evaluation" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Baseline comparisons. In Table 1, we present the mean ray angular error measurements for the distorted CO3D dataset, demonstrating the effectiveness of our approach compared to existing methods. Our results indicate that our method excels at extracting precise, arbitrary ray profiles even in the absence of a prior camera model. This capability is particularly significant given the challenges of distorted sequences. Notably, using the same ray representation as RayDiffusion, our approach consistently surpasses its performance, where the diffusion model depends solely on image-based features. The experiments reveal that our method achieves a more robust disentanglement from local image content, effectively isolating structural cues from object-specific fea" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26906" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 113, + 294, + 253 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 104 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 104 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 104 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of camera angular error on CO3D and Aria Digital Twins datasets. Averaged angular error (degrees) across varying numbers of input images." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 113, + 294, + 253 + ], + "lines": [ + { + "bbox": [ + 57, + 113, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 113, + 294, + 253 + ], + "type": "table", + "html": "
CO3D – Seen Categories
Method2345678
RayRegression [63]18.420.621.721.920.323.023.7
RayDiffusion [63]9.69.710.28.79.411.913.2
AlignDiff (Ours)3.43.12.83.23.03.64.1
CO3D – Unseen Categories
RayRegression [63]24.122.521.624.823.925.126.2
RayDiffusion [63]11.410.814.112.315.816.619.4
AlignDiff (Ours)4.65.25.85.76.46.88.2
Aria Digital Twins
RayRegression [63]34.638.236.540.841.240.943.7
RayDiffusion [63]28.429.134.737.240.442.345.6
AlignDiff (Ours)15.413.814.220.621.324.624.9
", + "image_path": "f162fb3b57bafaa3368c34c51eda88e07804ca8f8dfe81df833ace40efe11214.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 308, + 306, + 506 + ], + "blocks": [ + { + "bbox": [ + 56, + 266, + 295, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 300 + ], + "type": "text", + "content": "Table 2. Camera Rotation and Camera Center Accuracy on Distorted CO3D dataset. Experiments using geometrically aberrated videos, comparing recent methods against AlignDiff." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 308, + 306, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 306, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 306, + 506 + ], + "type": "table", + "html": "
Rotation @ 15°
number of Images2345678
Seen ClsRelPose [62]62.024.034.024.025.033.032.0
RelPose++ [33]72.573.473.674.775.076.075.7
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]74.080.085.382.484.882.586.6
AlignDiff (Ours)90.691.191.391.892.492.793.1
Unseen ClsRelPose [62]61.027.037.026.030.026.025.0
RelPose++ [33]61.460.863.064.165.765.765.4
PoseDiffusion [55]74.574.974.474.775.175.476.0
RayDiffusion [63]69.972.475.376.276.477.178.5
AlignDiff (Ours)83.684.884.985.285.786.186.1
Camera Center @ 0.1
Seen ClsRelPose++ [33]10085.279.074.370.568.666.0
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10073.472.571.171.070.670.2
AlignDiff (Ours)10092.590.788.687.285.884.6
Unseen ClsRelPose++ [33]10066.457.051.348.944.544.1
PoseDiffusion [55]10072.356.753.652.457.152.1
RayDiffusion [63]10074.871.670.267.366.762.5
AlignDiff (Ours)10086.384.280.575.273.070.1
", + "image_path": "7925d4679395274aef488008009416dbddd8ffedddb44391b9781b722d4deeb5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 556 + ], + "type": "text", + "content": "tures. This disentanglement minimizes residual errors tied to object appearance, allowing structural details to be more accurately emphasized." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 712 + ], + "type": "text", + "content": "Additionally, we report metrics for camera rotation accuracy and camera center accuracy, with aberrated images in Table 2 and rectified images in Table 3. Our approach consistently outperforms existing methods in estimating both camera position and orientation under aberrated conditions, confirming its resilience to challenging visual distortions. For rectified images, our conditioning method does not simply maintain performance but rather enhances it by recognizing and compensating for residual aberrations. This capability allows our approach to improve predictions even on images with minimal distortion, achieving a high level of precision and adaptability in handling diverse image qualities." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 113, + 553, + 373 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 3. Camera Rotation and Camera Center Accuracy on Rectified CO3D dataset. AlignDiff identifies subtle local distortions, achieving improvements over previous methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 113, + 553, + 373 + ], + "lines": [ + { + "bbox": [ + 316, + 113, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 113, + 553, + 373 + ], + "type": "table", + "html": "
Rotation @ 15°
number of Images2345678
Seen ClsCOLMAP [48]30.728.426.526.827.028.130.6
RelPose [62]56.056.557.057.257.257.357.2
PoseDiffusion [55]75.776.476.877.478.078.778.8
RelPose++ [33]81.882.884.184.784.985.385.5
RayRegression [63]88.888.788.789.089.489.389.2
RayDiffusion [63]91.892.492.692.993.193.393.3
AlignDiff (Ours)92.692.392.793.093.693.493.5
Unseen ClsCOLMAP [48]34.531.831.031.732.735.038.5
RelPose [62]48.647.548.148.348.448.448.3
PoseDiffusion [55]63.264.264.265.766.267.067.7
RelPose++ [33]69.871.171.972.873.874.474.9
RayRegression [63]79.079.680.681.481.381.981.9
RayDiffusion [63]83.585.686.386.987.287.588.1
AlignDiff (Ours)86.386.687.487.788.588.789.2
Camera Center @ 0.1
Seen ClsCOLMAP [48]10034.523.818.915.614.515.0
RelPose [62]10076.566.962.459.458.056.5
PoseDiffusion [55]10077.569.765.963.762.861.9
RelPose++ [33]10085.078.074.271.970.368.8
RayRegression [63]10091.785.782.179.877.976.2
RayDiffusion [63]10094.290.587.886.285.084.1
AlignDiff (Ours)10093.791.887.587.185.283.9
Unseen ClsCOLMAP [48]10036.025.520.017.917.619.1
PoseDiffusion [55]10063.650.545.743.041.239.9
RelPose++ [33]10070.658.853.450.447.846.6
RayRegression [63]10083.775.670.867.465.363.9
RayDiffusion [63]10087.781.177.074.172.471.4
AlignDiff (Ours)10088.281.079.377.572.671.0
", + "image_path": "5f389019e5bd57d70edf16e8e9ae96e1830833582183a725c924688e7ccaed36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 419, + 553, + 541 + ], + "blocks": [ + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 412 + ], + "type": "text", + "content": "Table 4. Camera Rotation and Camera Center Accuracy on Aria Digital Twins dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 419, + 553, + 541 + ], + "lines": [ + { + "bbox": [ + 317, + 419, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 317, + 419, + 553, + 541 + ], + "type": "table", + "html": "
Rotation @ 15°
Number of Images2345678
Unseen ClsRelPose [62]59.632.738.031.228.627.426.0
RelPose++ [33]62.861.362.662.863.264.964.8
PoseDiffusion [55]75.474.874.175.875.275.976.2
RayDiffusion [63]75.276.378.979.279.680.080.2
AlignDiff (Ours)81.482.782.884.384.184.685.5
Camera Center @ 0.1
Unseen ClsRelPose++ [33]10070.356.152.848.646.444.7
PoseDiffusion [55]10071.857.454.651.852.248.5
RayDiffusion [63]10075.969.261.860.455.651.1
AlignDiff (Ours)10078.370.064.462.758.152.9
", + "image_path": "6e19b2f7d78cfd44b9b4586f1b39099ba916ee72824fc9b522b81963c63aea37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 715 + ], + "type": "text", + "content": "Qualitative analysis of recovered cameras. We compare the generated cameras with the ground truth cameras in world space, as shown in Figure 7. The predicted ray moments and angular error from ground truth ray directions are also shown. Remarkably, the generated cameras exhibit little rotational difference from the reference cameras. While there are observed to have noticeable translation errors, we argue that this is due to scale ambiguities of the captures. The visualized ray moments present a smooth profile with very low angular errors. We further compare the Aria Digital Twins dataset for the inferred rays in vector form, shown in Figure 6. These further demonstrate that our approach can achieve robust and accurate ray profiles through camera" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26907" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 96, + 294, + 161 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 93 + ], + "type": "text", + "content": "Table 5. Ablated framework design evaluated with averaged ray angular error." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 96, + 294, + 161 + ], + "lines": [ + { + "bbox": [ + 56, + 96, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 294, + 161 + ], + "type": "table", + "html": "
Methods2345678
Baseline8.89.510.49.212.412.713.9
AlignDiff (w/o angular loss)5.86.18.87.39.611.314.7
AlignDiff (w/o line conditioning)6.78.49.89.910.310.112.8
AlignDiff (w/o edge-attention)5.37.17.49.210.811.310.5
AlignDiff (w/o optics grounding)5.26.96.87.78.49.111.3
AlignDiff4.65.25.85.76.46.88.2
", + "image_path": "27297decd2a4d5fd8c36c314c8ed606f4d55262b7582623820a757ff1f646760.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 188, + 306, + 229 + ], + "blocks": [ + { + "bbox": [ + 61, + 175, + 289, + 186 + ], + "lines": [ + { + "bbox": [ + 61, + 175, + 289, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 175, + 289, + 186 + ], + "type": "text", + "content": "Table 6. Evaluations on different image latent encoder choices." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 188, + 306, + 229 + ], + "lines": [ + { + "bbox": [ + 56, + 188, + 306, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 306, + 229 + ], + "type": "table", + "html": "
Methods2345678
DINOv2-s/14 [37]8.28.18.610.49.810.112.5
DINOv2-g/14 [37]4.56.46.36.66.17.89.6
TIPS-g/14 [34]4.65.25.85.76.46.88.2
", + "image_path": "5734de060f1eca43989a6d99eaf39453160d32b0bf9e2e108a8ecbb046e1c69b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 242, + 294, + 395 + ], + "blocks": [ + { + "bbox": [ + 57, + 242, + 294, + 395 + ], + "lines": [ + { + "bbox": [ + 57, + 242, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 57, + 242, + 294, + 395 + ], + "type": "image", + "image_path": "c8344e3c869e1a55aad0161f1562c8805c73c421908518388b95933041fd532e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 460 + ], + "type": "text", + "content": "Figure 7. Recovered cameras from aberrated image sequences. Predicted and expected cameras are visualized in solid and dashed lines, respectively. The cameras align with the rotational orientation of the expected counterparts. The ray moment and residual errors show low levels of deviation from reference rays." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 496 + ], + "type": "text", + "content": "conditioning. Detailed quantitative evaluation on the Aria Digital Twins dataset are included in Appendix D." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 581 + ], + "type": "text", + "content": "Undistortion using predicted flow. We conduct image undistortion with the predicted aberration map, as shown in Figure 4. Since we adopt a unified representation for aberrations, the radial fisheye camera model's vignetting effect leads to mapping artifacts. The predicted aberration map overall correctly captures the image distortions and reliably remaps the images regardless of the aberration types." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Ablation study. We perform an ablation study on CO3D data to evaluate the impact of each conditioning module, enhanced geometric guidance, and additional losses, as shown in Table 5. For each comparison, we remove one conditioning module from the framework. The results show that all modules significantly improve performance, and latent conditioning of the line segment has the greatest effect. Figure 4 visualizes the ray residuals with and without aberration conditioning, further validating its effectiveness. Additionally, we test the angular loss, finding that our localized ray angular loss yields notable improvements." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": "While our multiview diffusion approach implicitly incorporates temporal consistency by explicit temporal losses, we chose to handle temporal consistency as a post-processing step to keep the model lightweight and directly applicable in real-world deployments. This approach reduces computational overhead while still providing meaningful consistency improvements, leaving more complex temporal modeling for future work. This is based on the premise that camera profiles remain generally consistent over short periods, barring changes due to factors like camera sensor temperature fluctuations. To test this, we implemented a post-processing step that rejects low-confidence pose estimations, which reduced the average ray angular error from " + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "inline_equation", + "content": "8.20^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "inline_equation", + "content": "8.06^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": " over 8 frames. Furthermore, enforcing temporal consistency on local camera ray profiles using reprojection mean squared error loss yielded an average reduction of " + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "inline_equation", + "content": "0.06^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": " in angular error across the frames. We plan to integrate these enhancements, along with outlier rejection and a frame selection mechanism, into our future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 431 + ], + "type": "text", + "content": "Evaluations on image representation models. The image sequence encoder is designed to generate dense global embeddings that capture perceptual, geometric, and semantic representations. Among the available models, DINOv2 and TIPS are well suited for extracting high-quality dense features. We conduct an ablation study to evaluate these models' effectiveness in querying camera ray profiles. As shown in Table 6, the TIPS-g/14 model achieves the lowest angular errors on CO3D, with DINOv2-g/14 as a close competitor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 388, + 453 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 553, + 640 + ], + "type": "text", + "content": "We introduce AlignDiff, a unified calibration framework for real-world optical aberrations, enabling the recovery of camera ray bundles, extrinsics, and aberration profiles from multi-view image sequences using simple conditioning techniques. First, we propose structural conditioning to separate image geometry from semantic features, promoting emphasis on structural cues for accurate ray profile estimation. Second, we aggregate features along edges, ensuring that high-quality image embeddings are paired with effective guidance to avoid semantic interference. Finally, we incorporate real-world lens profiles into training, grounding AlignDiff in actual camera designs. These contributions collectively enhance generalization and accuracy across diverse distortion profiles, as demonstrated by consistent improvements over strong baselines." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "type": "text", + "content": "While AlignDiff performs well in controlled indoor settings, ray estimation may become unstable in arbitrary outdoor scenes due to scale ambiguity. To enhance robustness in such cases, integrating geometric priors like optical flow or combining traditional camera parameter estimation with network inference may be promising for future directions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26908" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Diffusionsfm: Predicting structure and motion via ray origin and endpoint diffusion, 2024. [Online; accessed 2025-03-06]. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "text", + "content": "[2] Tommaso Cavallari, Stuart Golodetz, Nicholas A. Lord, Julien Valentin, Victor A. Prisacariu, Luigi Di Stefano, and Philip H.S. Torr. Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2465-2477, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "text", + "content": "[3] Shuai Chen, Xinghui Li, Zirui Wang, and Victor Adrian Prisacariu. DFNet: Enhance Absolute Pose Regression with Direct Feature Matching. 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 249 + ], + "type": "text", + "content": "[4] Lambda Research Corp. Lensview optical design database. [Online; accessed 2025-03-07]. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 304 + ], + "type": "text", + "content": "[5] Andrei Cramariuc, Aleksandar Petrov, Rohit Suri, Mayank Mittal, Roland Siegwart, and Cesar Cadena. Learning camera miscalibration detection. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 4997-5003, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 339 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-Supervised Interest Point Detection and Description. 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 396 + ], + "type": "text", + "content": "[7] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "text", + "content": "[8] Satoshi Fujimoto and Nobutomo Matsunaga. Deep feature-based rgb-d odometry using superpoint and superglue. *Proceeding Computer Science*, 227:1127–1134, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 486 + ], + "type": "text", + "content": "[9] Fabrizio Gentile, Crescenzo Tortora, Giovanni Covone, Léon V. E. Koopmans, Rui Li, Laura Leuzzi, and Nicola R. Napolitano. *LeMoN: Lens Modelling with Neural networks* - I. Automated modelling of strong gravitational lenses with Bayesian Neural Networks. 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 521 + ], + "type": "text", + "content": "[10] Behnam Gholami, Mostafa El-Khamy, and Kee-Bong Song. Latent feature disentanglement for visual domain generalization. IEEE Transactions on Image Processing, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 295, + 577 + ], + "type": "text", + "content": "[11] Diandian Guo, Deng-Ping Fan, Tongyu Lu, Christos Sakaridis, and Luc Van Gool. Vanishing-point-guided video semantic segmentation of driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3544-3553, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[12] Lanqing Guo, Renjie Wan, Wenhan Yang, Alex C Kot, and Bihan Wen. Cross-image disentanglement for low-light enhancement in real world. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2550-2563, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 667 + ], + "type": "text", + "content": "[13] Annika Hagemann, Moritz Knorr, and Christoph Stiller. Deep geometry-aware camera self-calibration from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3438-3448, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 702 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Xiankang He, Guangkai Xu, Bo Zhang, Hao Chen, Ying Cui, and Dongyan Guo. DiffCalib: Reformulating Monocular Camera Calibration as Diffusion-Based Dense Incident Map Generation. 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. Technical report. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 204 + ], + "type": "text", + "content": "[17] Yannick Hold-Geoffroy, Dominique Piche-Meunier, Kalyan Sunkavalli, Jean-Charles Bazin, François Rameau, and Jean-François Lalonde. A perceptual measure for deep single image camera and lens calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(9):10603-10614, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[18] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 282 + ], + "type": "text", + "content": "[19] ISO 17850:2015(en). Photography — Digital cameras — Geometric distortion (GD) measurements. Standard, International Organization for Standardization, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[20] Y. Y. Jau, R. Zhu, H. Su, and M. Chandraker. Deep keypoint-based camera pose estimation with geometric constraints. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4950-4957, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[21] Hanwen Jiang, Arjun Karpur, Bingyi Cao, Qixing Huang, and André Araujo. OmniGlue: Generalizable Feature Matching with Foundation Model Guidance. Technical report. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "type": "text", + "content": "[22] Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-View Object Reconstruction with Unknown Categories and Camera Poses. 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 469 + ], + "type": "text", + "content": "[23] Yang Jiao, Trac D. Tran, and Guangming Shi. Effiscene: Efficient per-pixel rigidity inference for unsupervised joint learning of optical flow, depth, camera pose and motion segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5538-5547, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[24] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Matzen, Matthew Sticha, and David F. Fouhey. Perspective fields for single image camera calibration. In CVPR, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 558 + ], + "type": "text", + "content": "[25] Alex Kendall, Matthew Koichi Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. 2015 IEEE International Conference on Computer Vision (ICCV), pages 2938-2946, 2015." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[26] Julius Kümmerle and Tilman Kühner. Unified intrinsic and extrinsic camera and lidar calibration under uncertainties. In 2020 IEEE International conference on Robotics and Automation (ICRA), pages 6028-6034. IEEE, 2020. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "text", + "content": "[27] Jinwoo Lee, Hyunsung Go, Hyunjoon Lee, Sunghyun Cho, Minhyuk Sung, and Junho Kim. Ctrl-c: Camera calibration transformer with line-classification. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 16208-16217, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 702 + ], + "type": "text", + "content": "[28] Hao Li, Huai Yu, Jinwang Wang, Wen Yang, Lei Yu, and Sebastian Scherer. Ulsd: Unified line segment detection across pinhole, fisheye, and spherical cameras. ISPRS Journal of Photogrammetry and Remote Sensing, 178:187-202, 2021." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 335, + 704, + 341, + 711 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 704, + 341, + 711 + ], + "spans": [ + { + "bbox": [ + 335, + 704, + 341, + 711 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26909" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[29] Xiaoyu Li, Bo Zhang, Pedro V Sander, and Jing Liao. Blind geometric distortion correction on images through deep learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4855-4864, 2019. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 162 + ], + "type": "text", + "content": "[30] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[31] Kang Liao, Lang Nie, Shujuan Huang, Chunyu Lin, Jing Zhang, Yao Zhao, Moncef Gabbouj, and Dacheng Tao. Deep learning for camera calibration and beyond: A survey. arXiv preprint arXiv:2303.10559, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 241 + ], + "type": "text", + "content": "[32] Kang Liao, Zongsheng Yue, Zhonghua Wu, and Chen Change Loy. Mowa: Multiple-in-one image warping model. ArXiv, abs/2404.10716, 2024. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 275 + ], + "type": "text", + "content": "[33] Amy Lin, Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. RelPose++: Recovering 6D Poses from Sparse-view Observations. 2023. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "text", + "content": "[34] Kevis-Kokitsi Maninis, Kaifeng Chen, Soham Ghosh, Arjun Karpur, Koert Chen, Ye Xia, Bingyi Cao, Daniel Salz, Guangxing Han, Jan Dlabal, et al. Tips: Text-image pretraining with spatial awareness. arXiv preprint arXiv:2410.16512, 2024. 4, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "text", + "content": "[35] Shentong Mo, Enze Xie, Ruihang Chu, Lanqing Hong, Matthias Niessner, and Zhenguo Li. Dit-3d: Exploring plain diffusion transformers for 3d shape generation. In Advances in Neural Information Processing Systems, pages 67960–67971. Curran Associates, Inc., 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 295, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 410 + ], + "type": "text", + "content": "[36] Inc. DBA OptiTrack NaturalPoint. Optitrack system, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 412, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 295, + 510 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "text", + "content": "[38] Linfei Pan, Marc Pollefeys, and Viktor Larsson. Camera pose estimation using implicit distortion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12819-12828, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 624 + ], + "type": "text", + "content": "[39] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Yuheng Carl Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20133-20143, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 679 + ], + "type": "text", + "content": "[40] Chethan M. Parameshwara, Gokul Hari, Cornelia Fermüller, Nitin J. Sanket, and Yiannis Aloimonos. Diffposenet: Direct differentiable camera pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6845-6854, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[41] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. ControlNeXt: Powerful and Efficient Control for Image and Video Generation. 2024. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[42] Guilherme Potje, Felipe Cadar, Andre Araujo, Renato Martins, and Erickson R. Nascimento. XFeat: Accelerated Features for Lightweight Image Matching. 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[43] Srikumar Ramalingam and Peter Sturm. A Unifying Model for Camera Calibration: A Unifying Model for Camera Calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence: A Unifying Model for Camera Calibration. (7): 1309-1319, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[44] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction. Technical report. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[45] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From Coarse to Fine: Robust Hierarchical Localization at Large Scale. 2018. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 296 + ], + "type": "text", + "content": "[46] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 319 + ], + "type": "text", + "content": "[47] José Sasán. Introduction to aberrations in optical imaging systems. Cambridge University Press, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 342 + ], + "type": "text", + "content": "[48] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. Technical report. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[49] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[50] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[51] Cameron Smith, David Charatan, Ayush Tewari, and Vincent Sitzmann. Flowmap: High-quality camera poses, intrinsics, and depth via gradient descent. In arXiv, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 477 + ], + "type": "text", + "content": "[52] Peter Sturm and Srikumar Ramalingam. A Generic Concept for Camera Calibration. pages 1-13, 2004. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[53] Gabriele Trivigno, Carlo Masone, Barbara Caputo, and Torsten Sattler. The Unreasonable Effectiveness of Pre-Trained Features for Camera Pose Refinement. Technical report. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[54] Alexander Veicht, Paul-Edouard Sarlin, Philipp Lindenberger, and Marc Pollefeys. GeoCalib: Single-image Calibration with Geometric Optimization. In ECCV, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 590 + ], + "type": "text", + "content": "[55] Jianyuan Wang, Christian Rupprecht, and David Novotny. PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment. 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "text", + "content": "[56] Shaoru Wang, Jin Gao, Zeming Li, Xiaoqin Zhang, and Weiming Hu. A closer look at self-supervised lightweight vision transformers. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 668 + ], + "type": "text", + "content": "[57] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3D Vision Made Easy. 2023. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[58] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. 2020. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26910" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 430 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[59] Scott Workman, Connor Greenwell, Menghua Zhai, Ryan Baltenberger, and Nathan Jacobs. Deepfocal: A method for direct focal length estimation. In 2015 IEEE International Conference on Image Processing (ICIP), pages 1369-1373, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 162 + ], + "type": "text", + "content": "[60] Shangrong Yang, Chunyu Lin, Kang Liao, and Yao Zhao. Innovating Real Fisheye Image Correction with Dual Diffusion Architecture. Technical report. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[61] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[62] Jason Y. Zhang, Deva Ramanan, and Shubham Tulsiani. Relapse: Predicting probabilistic relative rotation for single objects in the wild. page 592-611, Berlin, Heidelberg, 2022. Springer-Verlag. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[63] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as Rays: Pose Estimation via Ray Diffusion. 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[64] Yinqiang Zheng, Yubin Kuang, Shigeki Sugimoto, Kalle Åström, and Masatoshi Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In 2013 IEEE International Conference on Computer Vision, pages 2344-2351, 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[65] Yichao Zhou, Haozhi Qi, Jingwei Huang, and Yi Ma. Neurvps: Neural vanishing point scanning via conic convolution. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 430 + ], + "type": "text", + "content": "[66] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a Wild Camera: In-the-Wild Monocular Camera Calibration. 2023. 2, 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26911" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_content_list.json b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e0a1cde9450a93bfa3aac443a5cc7b9b43fcd244 --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_content_list.json @@ -0,0 +1,1515 @@ +[ + { + "type": "text", + "text": "AlignGuard: Scalable Safety Alignment for Text-to-Image Generation", + "text_level": 1, + "bbox": [ + 145, + 130, + 852, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Runtao Liu $^{1*}$ I Chieh Chen $^{1*}$ Jindong Gu $^{2}$ Jipeng Zhang $^{1}$ Renjie Pi $^{1}$", + "bbox": [ + 196, + 179, + 799, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qifeng Chen $^{1}$ Philip Torr $^{2}$ Ashkan Khakzar $^{2}$ Fabio Pizzati $^{2,3}$", + "bbox": [ + 228, + 198, + 766, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology $^{2}$ University of Oxford $^{3}$ MBZUAI", + "bbox": [ + 143, + 215, + 854, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rliuay,icchen}@connect.ust.hk", + "bbox": [ + 367, + 236, + 630, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://alignguard.github.io/", + "bbox": [ + 323, + 252, + 666, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) models are widespread, but their limited safety guardrails expose end users to harmful content and potentially allow for model misuse. Current safety measures are typically limited to text-based filtering or concept removal strategies, able to remove just a few concepts from the model's generative capabilities. In this work, we introduce AlignGuard, a method for safety alignment of T2I models. We enable the application of Direct Preference Optimization (DPO) for safety purposes in T2I models by synthetically generating a dataset of harmful and safe image-text pairs, which we call CoProV2. Using a custom DPO strategy and this dataset, we train safety experts, in the form of low-rank adaptation (LoRA) matrices, able to guide the generation process away from specific safety-related concepts. Then, we merge the experts into a single LoRA using a novel merging strategy for optimal scaling performance. This expert-based approach enables scalability, allowing us to remove $7 \\times$ more harmful concepts from T2I models compared to baselines. AlignGuard consistently outperforms the state-of-the-art on many benchmarks and establishes new practices for safety alignment in T2I networks.", + "bbox": [ + 89, + 335, + 483, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Warning: this paper includes potentially offensive content.", + "bbox": [ + 94, + 656, + 480, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 698, + 220, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) models have advanced rapidly in recent years, becoming key tools in content creation and entertainment, used by both professionals and hobbyists [48, 49]. Such an impressive growth in image synthesis capabilities, in some cases reaching indistinguishable realism from real samples, has generated concerns about the risks of releasing such models [9, 14, 15]. Indeed, T2I may generate unsettling or harmful content, and expose people, particularly minors or those from sensitive cultural backgrounds, to inappropriate outputs. Moreover, malicious users could", + "bbox": [ + 89, + 724, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57319123dad845f0a64483683363f391a1f5b085094ae51688fac22b17f89d24.jpg", + "image_caption": [ + "Figure 1. Safety alignment for T2I. T2I models released without safety alignment risk to be misused (top). We propose Align-Guard, a scalable safety alignment framework for T2I models supporting the mass removal of harmful concepts (middle). We allow for scalability by training safety experts focusing on separate categories such as \"Hate\", \"Sexual\", \"Violence\", etc. We then merge the experts with a novel strategy. By doing so, we obtain safety-aligned models, mitigating unsafe content generation (bottom)." + ], + "image_footnote": [], + "bbox": [ + 519, + 301, + 906, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "easily generate, among others, violent pictures, or sexually-explicit deepfakes, potentially causing misinformation or harassing specific users. Regardless of these risks, there has been little effort in proposing effective safety alignment techniques, i.e. strategies for limiting harmful outputs, for T2I models. Most of the current approaches are focused on simple input filtering strategies, such as the analysis of the input prompt, or detectors on output images [2]. However, these strategies are easy to circumvent [46], or they can be deactivated if models are openly released. Conversely, large language models (LLMs) are subject to rigorous safety alignment procedures before being deployed [13]. It is our objective to bridge this gap.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 109, + 887, + 225, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17024", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For preventing harmful outputs, one straightforward approach might seem to be training on curated datasets excluding unsafe images. However, this is nontrivial. Large T2I trainings usually use vast web-crawled datasets, such as LAION-5B [53]. Despite filtering efforts, these datasets contain unsafe or illegal content, making full control nearly impossible. More realistically, we could limit the capability of the T2I to generate harmful concepts post-training. Recent works explored methods to delete specific concepts from pre-trained models, with promising results [16, 17, 29, 34]. Yet, these methods limitedly scale. While they can remove a limited number of concepts, deleting hundreds strongly degrades the model's generative capacity [17, 34].", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce AlignGuard, a scalable method for safety alignment of T2I models. Our idea is to remove a large number of concepts from T2I models, by exploiting an ensemble of safety experts trained with Direct Preference Optimization (DPO) on safety-oriented synthetic data. By doing so, we are the first to propose a scalable safety alignment method for T2I, allowing for a safe deployment of trained T2I models. Considering the lack of data for DPO-based safety alignment available for image generation, we introduce CoProV2, a dataset encompassing a broad array of unsafe concepts. CoProV2 generation is fully automatic, including paired images with associated prompts: one harmful image with its corresponding textual description, and one safe image and prompt with similar content and structure. This setup enables the application of DPO for aligning T2I models effectively. In practice, we train Low-Rank Adaptation (LoRA) [24] matrices. Each LoRA serves as a safety expert trained to prevent the generation of specific harmful content, i.e. content related to different categories such as \"Hate,\" \"Sexual,\" \"Violence,\" among others. The LoRAs guide the vision representation of the pre-trained T2I model towards safe outputs during the diffusion process.", + "bbox": [ + 91, + 291, + 483, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "After training the safety experts, we use a new algorithm called Co-Merge to merge all safety LoRA experts into one. This mitigates interferences among different experts, allowing for a significant increase in scalability. As shown in Figure 1, through joint expert training and subsequent merging, we can train on images encompassing 723 harmful concepts—approximately seven times more than existing approaches [17]. In summary, AlignGuard enables safety alignment of T2I models at scale, preventing the generation of inappropriate outputs without impacting generative capabilities on safe prompts. Our contributions are:", + "bbox": [ + 89, + 641, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose AlignGuard, the first scalable approach for T2I models specifically targeting safety alignment.", + "2. We introduce a training method based on expert models, and a novel merging strategy based on weight activation frequency.", + "3. For the training of AlignGuard, we craft a new dataset" + ], + "bbox": [ + 91, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for safety alignment in T2I, coined CoProV2, that we will release as open source.", + "bbox": [ + 532, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4. We validate AlignGuard on popular T2I models and multiple benchmarks, demonstrating its effectiveness in many setups.", + "bbox": [ + 513, + 121, + 903, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 181, + 653, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Content filtering. Often, closed-source commercial T2I use prompt blacklists, LLM preprocessing, and image analysis for content filtering [1, 3, 4]. In particular, some use LLM-based prompt analysis [35] to detect harmful inputs, even with ad-hoc models [27]. Recently, Latent Guard [32] proposed a latent-based blacklist within text encoders in T2I models. Some works rely instead on the analysis of generated images [2], where an NSFW classifier is applied to generated images. Instead, others [40] use inpainting to mask potentially unsafe content. These approaches can easily be deactivated if the T2I models' weights are available.", + "bbox": [ + 511, + 207, + 906, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept removal in T2I. Some have explored removing the capability of generating concepts in T2I, like SLD [52], using classifier-free guidance [22] to steer generation away from undesirable outputs. Similarly, Li et al. [30] identify interpretable directions that can be used for safety steering. Alternatively, many fine-tune the T2I. While seminal works finetune the entire model [29], some focus on specific components, such as attentions [17, 38, 62], specific neurons [10], or the textual encoder [43]. There has been a recent interest in mass removal of concepts from T2I models for safety purposes [17, 34]. While MACE [34] uses a similar strategy as ours, they make use of segmentation masks, constraining the erasable concepts at spatially-defined ones. Moreover, all available approaches [17, 34] degrade performance with more than 100 erased concepts.", + "bbox": [ + 511, + 382, + 908, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Model merging. Model merging, i.e. combining multiple models into one model, has gained attention [12, 61]. Weighted averaging methods are commonly used to enhance performance [18, 26, 57], especially in scenarios where only model weights are accessible [28]. Beyond these, advanced techniques have emerged, improving over basic averaging [6, 36, 39]. For example, TIES [58] resolves operator conflicts to improve merging, at the cost of hyperparameter tuning. Merging has been explored for safety in LLMs [8, 19] and for concept removal in MACE [34], but only using expensive optimization procedures. Existing LoRA merging for T2I focus on single subjects [54, 63]. Instead, we propose a cheap and effective merging strategy, focusing on broad safety categories.", + "bbox": [ + 511, + 617, + 908, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 844, + 651, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use preference optimization algorithms such as DPO [56] to perform scalable safety alignment of T2I mod", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17025", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els. Here, we revise the fundamental concepts to allow for the interpretation of our method.", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-to-image diffusion models", + "text_level": 1, + "bbox": [ + 89, + 131, + 366, + 147 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models allow for image generation by iteratively denoising gaussian noise with a network $\\epsilon$ for $t\\in [0,T]$ iterations [23]. In particular, T2I diffusion models include natural language conditioning, so the desired output image can be described with text. We now briefly introduce the training procedure for a T2I model. Let us assume an input pair $(x,p)\\sim \\mathcal{D}_{\\mathrm{train}}$ sampled from a training dataset $\\mathcal{D}_{\\mathrm{train}}$ of images $x$ and paired textual description $p$ . The network $\\epsilon$ is trained by estimating the ground truth noise $\\tilde{\\epsilon}_t$ injected on $x$ for a random $t\\in [0,T]$ . We define the input image with the addition of noise as $x_{t}$ . Hence, the training loss is", + "bbox": [ + 89, + 152, + 483, + 320 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} (\\epsilon , x, p) = | | \\tilde {\\epsilon} _ {t} - \\epsilon \\left(x _ {t}, p\\right) | |. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 332, + 482, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Eq. (1), $\\epsilon (\\cdot)$ is the denoising operation. The network weights $\\theta^{*}$ are optimized by minimizing the following:", + "bbox": [ + 89, + 361, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, p) \\sim \\mathcal {D} _ {\\text {t r a i n}}} \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\theta}, x, p\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 406, + 482, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ refers to the denoising network with a set of weights $\\theta$ . During inference, sampled Gaussian noise is iteratively processed for $t \\in [0, T]$ with $\\epsilon_{\\theta^*}$ following specific scheduling policies, ultimately allowing image synthesis. We refer to [23] for additional details.", + "bbox": [ + 89, + 438, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. DPO for diffusion models", + "text_level": 1, + "bbox": [ + 89, + 523, + 323, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DPO is a technique for preference alignment initially developed for LLMs [41, 45] and recently extended to diffusion models [33, 56]. The core idea is to benefit from pairwise preferences obtained by labeling. Let us assume a dataset $\\mathcal{D}_{\\mathrm{DPO}}$ of paired images and textual descriptions $(x^{+}, x^{-}, p)$ . For a given description $p$ , $x^{+}$ is a image that humans indicated as preferred output with respect to $x^{-}$ . The intuition of DPO is to increase the likelihood to generate the preferred output $x^{+}$ , while discouraging the generation of $x^{-}$ . This translates into the following loss:", + "bbox": [ + 89, + 545, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon , x ^ {+}, x ^ {-}, p\\right) = - \\log \\sigma (- \\beta ( \\\\ \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {+}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {+}, p\\right)\\right) \\tag {3} \\\\ - \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {-}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {-}, p\\right)\\right)\\left. \\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 708, + 480, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\mathrm{ref}}$ is a reference pre-trained network, typically resulting from a previous optimization of Eq. 2. Also, $\\sigma$ is the sigmoid operation and $\\beta$ is a weighting constant [56]. Finally, one could optimize as:", + "bbox": [ + 89, + 777, + 483, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\mathrm {D P O}} ^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x ^ {+}, x ^ {-}, p) \\sim \\mathcal {D} _ {\\mathrm {D P O}}} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {+}, x ^ {-}, p\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 849, + 483, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For further details, we refer to the original paper [56].", + "bbox": [ + 93, + 885, + 450, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5fdb24e42ff8224b830deaba3abed6b578568830b836ac19886aeb974c6385d9.jpg", + "image_caption": [ + "Figure 2. Dataset generation. For each unsafe concept in different categories, we generate an corresponding prompts with an LLM. We generated paired safe prompts using an LLM, minimizing semantic differences. Then, we use the T2I we intend to align to generate corresponding images for both prompts." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 924, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 371, + 604, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to align T2I models inspired by practices in language model alignment [13]. We avoid expensive human annotations by generating safety-oriented preference data in Section 4.1. We also propose a training procedure based on expert networks (Section 4.2) and subsequent merging (Section 4.3). After training, AlignGuard requires no additional computation at inference (more details in supplement).", + "bbox": [ + 511, + 397, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. DPO for safety alignment in T2I", + "text_level": 1, + "bbox": [ + 511, + 513, + 797, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to use DPO as a strategy to perform safety alignment on pretrained T2I models. Assuming an unsafe textual input describing a sexually explicit, shocking, or violent scene, we would like the diffusion model to ignore the unsafe requested traits, and generate a safe image, i.e. not including any visually disturbing trait. We achieve this by generating automatically {unsafe, safe} image pairs for DPO training, obtained by propting the T2I with LLM-manipulated text. In a nutshell, we can discourage the generation of unsafe images, having as preferred DPO image a visually-close safe sample, for an unsafe input T2I prompt.", + "bbox": [ + 511, + 537, + 906, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start by generating a set of unsafe concepts $\\mathcal{C}$ with an LLM. Each $c\\in \\mathcal{C}$ is a keyword representing unsafe elements that may be included in an image, such as \"homicide\", \"nude\", etc. Then, we follow the prompt synthesis procedure of [32], generating both unsafe and safe prompts for image generators from concepts, enforcing minimal differences. Specifically, we sample an unsafe prompt $p^{\\mathrm{U}}$ with an LLM to create text visually describing a scene using the input concept $c$ . Then, we further process the generated $p^{\\mathrm{U}}$ , prompting an LLM to remove any reference to the input unsafe concept $c$ . This transforms the prompt into a safe $p^{\\mathrm{S}}$ , while minimizing semantic changes. Differently from [32], we process both prompts with a pretrained", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17026", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6c3051da9780dc1e8aa531b8cf54a27aea7d983ce297ae0892e469b75baa53b5.jpg", + "image_caption": [ + "Figure 3. Expert training and merging. First, we use the previously generated prompts and images to train LoRA experts on specific safety categories (left), exploiting our DPO-based losses. Then, we merge all the safety experts with Co-Merge (right). This allows to achieve general safety experts that produce safe outputs for a generic unsafe input prompt in any category." + ], + "image_footnote": [], + "bbox": [ + 96, + 95, + 598, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3bc2b8c155eeeec4e24869edca388100096047370f005f4e1587af59a980998b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 95, + 898, + 268 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b771cc2f950c55918ffe042517d06b9196474e654b34b1a768f3d99b12d54e30.jpg", + "image_caption": [ + "Figure 4. Merging experts with Co-Merge. (left) Assuming LoRA experts with the same architecture, we analyze which expert has the highest activation for each weights for all inputs. (right) Then, we obtained the merged weights from multiple experts, merging only the most active weights per expert." + ], + "image_footnote": [], + "bbox": [ + 93, + 354, + 486, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diffusion model, obtaining images $\\{x^{\\mathrm{U}}, x^{\\mathrm{S}}\\}$ , derived from $p^{\\mathrm{U}}$ and $p^{\\mathrm{S}}$ , respectively. We show this in Figure 2.", + "bbox": [ + 89, + 571, + 482, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we can automatically create image preference pairs, assuming that the preferred output in safety alignment would always be a similar safe image for an unsafe prompt. Indeed, by repeating the process multiple times, we can construct a dataset $\\mathcal{D}_{\\mathrm{safety}}$ consisting of tuples including the sampled prompts and corresponding generated images, i.e. $(x^{\\mathrm{S}}, x^{\\mathrm{U}}, p^{\\mathrm{S}}, p^{\\mathrm{U}})$ . Then, we use DPO to discourage the generation of unsafe outputs for an unsafe prompt $p^{\\mathrm{U}}$ , and promote the generation of the paired safe image $x^{\\mathrm{S}}$ instead. To achieve this, we set $x^{\\mathrm{S}}$ as preferred output for $p^{\\mathrm{U}}$ , discouraging the generation of $x^{\\mathrm{U}}$ . The semantic similarity of $p^{\\mathrm{U}}$ and $p^{\\mathrm{S}}$ will encourage the DPO training to focus exclusively on the visual traits making $x^{\\mathrm{U}}$ unsafe. We define a first loss term:", + "bbox": [ + 89, + 619, + 483, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {U}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 844, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While effective in removing the unsafe concepts $c$ from the generated images, this risks to reduce the generative capa", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bilities of the network, since we are encouraging to generate images that do not respect the input prompt. To prevent this, we also use $x^{\\mathrm{S}}$ as preferred generation for the input prompt $p^{\\mathrm{S}}$ , while penalizing on $x^{\\mathrm{U}}$ . This writes:", + "bbox": [ + 511, + 358, + 906, + 419 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {S}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 429, + 906, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we modify Eq. (4) proposing our training objective:", + "bbox": [ + 511, + 460, + 905, + 476 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {\\text {A l i g n G u a r d}} ^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x ^ {\\mathrm {s}}, x ^ {\\mathrm {U}}, p ^ {\\mathrm {s}}, p ^ {\\mathrm {U}}) \\sim \\mathcal {D} _ {\\text {s a f e t y}}} \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {c o n}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 501, + 905, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This objective allows for a stable safety-oriented finetuning of the T2I model for a large number of concepts $c \\in \\mathcal{C}$ without impacting the generative capabilities of the T2I model.", + "bbox": [ + 511, + 536, + 905, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Improving scaling with safety experts", + "text_level": 1, + "bbox": [ + 511, + 590, + 836, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training with our proposed strategy in Section 4.1 enables us to scale the number of concepts used for safety alignment, while keeping generative capabilities intact. However, by increasing the number of concepts in $\\mathcal{C}$ , we noticed a decrease in performance. We suppose this is due to the complexity of the task: in safety alignment there is a need of a high contextual understanding of the generated scenes. As an example, an image of a kitchen including knives is a safe output, while a scene of harassment with the usage of a knife is an undesired unsafe outcome of the image synthesis process, that we aim to prevent. As a solution, we propose to decompose the safety alignment on separate categories, such as \"violence\", \"sexual content\", \"harassment\", and others. Our intuition is that by limiting the variability of the generated images, it would be easier to identify visual patterns that would make a scene \"violent\", or \"sexually explicit\". Hence, we aim to train safety experts, each focusing on one category only.", + "bbox": [ + 511, + 613, + 906, + 887 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17027", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Co-Merge \n1: Input: Safety expert LoRAs $\\{L_1, L_2, \\ldots, L_N\\}$ , unsafe prompts $\\{p_1, p_2, \\ldots, p_K\\}$ \n2: Output: Merged LoRA $L_{\\text{merged}}$ \n3: Init a count matrix $C$ of $J \\times N$ elements with zeros. \n4: for each prompt $p_k$ , $k \\in [1, K]$ do \n5: for each neuron index $j$ , $j \\in [1, J]$ do \n6: $i_{\\text{max}} := \\arg \\max_i |L_i^j(p_k)|$ , $i \\in [1, N]$ \n7: Increment count $C[j, i_{\\text{max}}]$ \n8: end for \n9: end for \n10: for each neuron index $j$ , $j \\in [1, J]$ do \n11: $m := \\arg \\max_i C[j, i]$ , $i \\in [1, N]$ \n12: $L_{\\text{merged}}^j := L_m^j$ \n13: end for \n14: Return $L_{\\text{merged}}$", + "bbox": [ + 91, + 89, + 483, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, we identified $N$ safety-related concept categories, each linked to a generated concept set $\\{\\mathcal{C}^1,\\dots,\\mathcal{C}^N\\}$ . Each set is created by prompting a large language model to generate concepts within a given unsafe category. We then generated the corresponding safety datasets $\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}$ following the approach in Section 4.1. Rather than fine-tuning the entire model's weights, $\\theta$ , to create the expert networks, we used low-rank adaptation (LoRA) [24], updating only a minimal subset of weights and as such saving significant computational costs. Ultimately, we apply Equation 7 for $\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}$ , obtaining a set of LoRAs $\\{L^1,\\dots,L^N\\}$ , as shown in Figure 3. At inference time, the obtained LoRAs can be applied to the pretrained T2I model to enable safe image generation.", + "bbox": [ + 89, + 367, + 483, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Experts merging", + "text_level": 1, + "bbox": [ + 89, + 590, + 256, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prevent the generation of content across multiple safety-related categories, additional considerations are necessary. Indeed, loading multiple LoRAs into a model can lead to interference [25], reducing overall performance in safety-oriented image generation. Alternatively, multiple experts could be run independently, at the cost of multiplying inference costs. To mitigate these issues, we aim to merge the trained LoRAs into a single one, as visualized in Figure 3, without sacrificing performance.", + "bbox": [ + 88, + 613, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a novel data-based strategy, called CoMerge, to merge multiple safety experts, specifically designed around our use case. Our intuition is that by decomposing safety alignment in multiple categories (Section 4.2), trained LoRAs sharing the same architecture would encode information about category-specific alignment in different neurons. In other words, we assume that in the T2I model, the weights responsible for generating violent content (i.e. including, for instance, blood) will be different from those generating sexually-explicit content", + "bbox": [ + 88, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c843468680b73d2ef2639073d1a00ea04a13fc2a9c79b526f9220f2d412f698c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetImage # of Prompts # of Categories # of ConceptsIP
COCO [31]✓3,000N/AN/A0.06
I2P [51]✗4,7037N/A0.36
UD [44]✗9325N/A0.47
CoPro [32]✗56,52677230.23
CoProV2 (ours)✓23,69077230.43
", + "bbox": [ + 521, + 89, + 903, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Datasets comparison. Our LLM-generated dataset, CoProV2, achieves comparable Inappropriate Probability (IP) to human-crafted datasets (UD [44], I2P [51]) and offers similar scale to CoPro [32]. COCO [31], exhibiting a low IP, is used as a benchmark for image generation with safe prompts as input.", + "bbox": [ + 511, + 185, + 906, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(e.g. naked skin). LoRAs work as updates for the model weights [24], so, for concepts associated to different neurons, they will encode related information in different parts of the same architecture. Then, our idea is to isolate the neurons inside the trained safety experts that activate the most for unsafe input prompts, and construct a merged expert as a LoRA $L_{\\text{merged}}$ including the parameters of different experts associated to the most significant activations only.", + "bbox": [ + 511, + 263, + 906, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We start by randomly sampling $K$ unsafe prompts from the training set of $\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}$ , equally distributed across each category. We then process all prompts with the T2I model, where we load all trained experts $\\{L_1,\\dots,L_N\\}$ one at the time. Each expert LoRA is a multi-layer perceptron including $J$ neurons. For each prompt $p_k^U$ , where $k\\in [1,K]$ , and each expert model $L_{i}$ , where $i\\in [1,N]$ , we record the absolute value of the activation from each neuron $L_{i}^{j}$ , where $j\\in [1,J]$ . We denoted the absolute value as $|L_i^j (p_k^U)|$ . This provides a measure of the response for each neuron to that prompt. These prompts are uniformly sampled from distinct safety categories, so different neurons of the same index $j$ across the experts will exhibit different activations in response to different prompts. To create the merged $L_{\\mathrm{merged}}$ , we identify the neurons with the highest activation frequencies across experts. Specifically, for each neuron $j$ in the merged expert $L_{\\mathrm{merged}}$ , we select the neuron from the original set of experts $\\{L_1,\\dots,L_N\\}$ that has the highest activation frequency count across the $K$ prompts. We summarize this in Algorithm 1, where the $p^U$ superscript is omitted for clarity. Figure 4 illustrates Co-Merge with $K = 10$ , describing how the frequency count leads to the network weight merging. Note that merging $L_{\\mathrm{merged}}$ with the original T2I weights makes safety alignment difficulty reversible, a beneficial practice for T2I releases.", + "bbox": [ + 511, + 386, + 908, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 771, + 645, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first introduce our setup (Section 5.1), and compare with baselines in Section 5.2. We then analyze the properties of AlignGuard (Section 5.3) and show ablations (Section 5.4).", + "bbox": [ + 511, + 794, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 848, + 700, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We use AlignGuard to align three models for T2I, namely Stable Diffusion (SD) v1.5/v2.1 [48] and", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17028", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e7e2e05e479fd3cc956c447ce19f6d02fd4fc1b209c090d377ec345632a17b3c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIP↓FID ↓ COCOCLIP ↑
CoProV2I2PUD
SD1.5No alignment0.510.360.5269.7733.52
SLD [51]0.270.190.3071.4532.24
ESD-u [16]0.220.250.2172.9829.61
UCE [17]0.330.300.3872.0132.01
AlignGuard0.070.110.1670.9632.32
SD2.1No alignment0.510.350.5570.4534.99
AlignGuard0.120.120.1771.9934.13
SDXLNo alignment0.490.310.4768.9035.60
AlignGuard0.090.080.1175.2034.67
", + "bbox": [ + 91, + 88, + 483, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Benchmarks. AlignGuard achieves best performance both in generated images alignment (IP) and image quality (FID, CLIPScore), with two T2I models and against 3 methods for SD v1.5. Note that we use CoProV2 only for training, hence I2P and UD are out-of-distribution. Yet, AlignGuard allows a robust safety alignment. Best is bold, second is underlined.", + "bbox": [ + 89, + 267, + 482, + 351 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SDXL [42]. For all, we compare the original versions released on HuggingFace, and our finetuned version with AlignGuard. Moreover, we compare with recent methods based on SD v1.5 for safe image generation (SLD [51]) and concept erasure (ESD [16], UCE [17]). In particular, for ESD, we follow the paper indications and use ESD-u [16], a version of ESD for broad concept removal. We perform our analysis and ablations on SD v1.5.", + "bbox": [ + 89, + 364, + 483, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate the effectiveness of safety alignment, we use the inappropriate probability metric (IP) from SLD [51], measuring the ratio of unsafe contents generated by T2I with unsafe prompts. The unsafe content detection is performed by using classification results of Q16 [50] and NudeNet [47]. For evaluating image quality, we use FID [21], and CLIPScore [20] for text-image alignment.", + "bbox": [ + 89, + 496, + 483, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For training of AlignGuard, we generate a new dataset following Section 4.1, named CoProV2 (Concepts and Prompts). We motivate this choice by observing a limitation in the CoPro dataset [32], i.e. the limited IP measured on images generated with the prompts of the dataset. This suggests that many prompts in CoPro are not leading to unsafe generations. This is due to the limited performance of the LLM building CoPro. We start from the same $\\mathcal{C}$ as CoPro: 723 harmful concepts across 7 categories (Hate, Harassment, Violence, Self-Harm, Sexual, Shocking, Illegal). We re-generate the unsafe/safe prompts pairs with a new LLM prompt reported in the supplement. We use Mistral-Nemo-Instruct [37] for the prompt synthesis. We generate 23,690 pairs of safe/unsafe prompts, and we use 15,690/8,000 pairs for training/ testing. For testing only, we use I2P [7] and Unsafe Diffusion (UD) [44], including 4703/932 human-designed unsafe prompts on 7/5 categories, respectively. Following common practices [16, 48, 51] we use COCO [31], with 3000 safe", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "captions and corresponding images, for evaluating the performance of the T2I model in image generation after alignment. We compare the datasets quantitatively in Table 1.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We train AlignGuard with batch-size 8, accumulating over 16 steps. We use AdamW with a learning rate of $10^{-5}$ for training each LoRA with rank 4. For Co-Merge, we set $K = 100$ unsafe prompts. Each LoRA takes $\\sim 2$ hours to train for SD and $\\sim 18$ hours for SDXL on a single Nvidia 5880 Ada GPU for 2000 steps. We normalize the computation needed for training AlignGuard and baselines. More details are in the supplement.", + "bbox": [ + 511, + 141, + 906, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmarks", + "text_level": 1, + "bbox": [ + 511, + 268, + 648, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation. In Table 2, we compare with alignment baselines for IP, FID, and CLIPScore. For fairness, all baselines were re-trained on CoProV2 training set and evaluated using the CoProV2 test set, I2P, and UD. Note that I2P and UD serve as out-of-distribution evaluations as they are crafted by human users in real deployment scenarios, differing in both prompt format and content from CoProV2. Initially, we trained the baselines using the complete $\\mathcal{C}$ set of CoProV2 with Stable Diffusion v1.5 as the backbone. This involved using $\\mathcal{C}$ to identify negative embeddings in SLD and excluding all $\\mathcal{C}$ concepts for ESD-u and UCE. All baselines collapsed due to the high number of concepts in CoProV2, aligning with the limited scaling performance noted in previous works [16, 17]. These results are included in the supplement. For successful training comparisons, we used the 7 category names of CoProV2 as removable concepts for the baselines, while we use the full prompts and images in CoProV2 for AlignGuard. We trained our expert LoRAs on samples generated with all $\\mathcal{C}$ , one per category, and merge them with Co-Merge. As shown in the table, we outperform considerably all baselines in every metric. Notably, our IP for SD v1.5 on CoProV2 (0.07) is much lower than the next best (ESD-u, 0.22). Additionally, AlignGuard demonstrates superior generalization, achieving better IP scores on unseen I2P (0.08) and UD (0.16). Our DPO-based strategy effectively maintains Stable Diffusion's generation capabilities preventing forgetting of concepts unrelated to safety. Indeed we report nearing the performance of the Baseline method without safety alignment in FID (70.96 vs 69.77) and CLIPScore (32.32 vs 33.52) on COCO captions. We also applied AlignGuard to SD v2.1 and SDXL to test the generalization of our method to a different diffusion model. Our results are consistent with different backbones like SD v2.1 and SDXL, achieving performance comparable to our aligned SD v1.5 with IP 0.09 and 0.12 on CoProV2. This suggests that AlignGuard can be used for the safety alignment of arbitrary T2I models.", + "bbox": [ + 511, + 291, + 906, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation. We show in Figure 5 a selection of samples advocating the effectiveness of AlignGuard. We", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17029", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21c1d789409b3ec5f0c56b1782d9d048cef3fc9b545158863582e81dabbb3421.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison. Compared to non-aligned baseline models, AlignGuard allows to synthesize safe images for unsafe input prompts. Please note the layout similarity between the unsafe and safe outputs: thanks to our training, only the harmful image traits are removed from the generated images. Concepts in ⟨brackets⟩. Prompts are shortened, for full ones see the supplement." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 906, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sample prompts from CoProV2. For SD v1.5 and SDXL (SD v2.1 in the supplement), we prevent the generation of unsafe content across 7 categories. Please also note how the quality of the generated images, in both models aligned with AlignGuard, is not impacted by our alignment procedure. Also, let us highlight that although the model refuses to follow unsafe prompts, the generated images preserve similar traits and global layout, e.g. the pose of the man in the \"Shocking\" column for SDXL. This property results from training on safe/unsafe pairs.", + "bbox": [ + 88, + 546, + 482, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Properties", + "text_level": 1, + "bbox": [ + 89, + 710, + 207, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experts performance. For AlignGuard, we train an expert for each category in CoProV2. We now examine the effectiveness of each expert LoRA across all categories in the CoProV2 test set in Table 3. The results show that training and evaluating on the same category generally leads to good performance, as expected. Interestingly though, applying one expert LoRA consistently improves IP across all categories and often surpasses the expert trained specifically for that category. We attribute this to the interaction of multiple concepts across categories. While the concepts differ, the visual features of generated", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images may share similarities. For example, scenes in the \"Violence\" or \"Self-harm\" categories often depict blood, while the \"Hate\" category may include sexual slurs, which could explain the strong performance on \"Sexual\" data. Ultimately, this suggests that for safety alignment tasks, AlignGuard is robust to the concept selection. However, experts trained on different categories still encode distinct concepts. This becomes evident when merging all LoRAs with Co-Merge (All - Ours), since we achieve the best performance across all categories. We also include a baseline where a single LoRA is trained on prompts from all categories (All - single), which shows suboptimal performance. Here, the lack of specialized expert models results in less effective training, justifying our design.", + "bbox": [ + 511, + 546, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adversarial attacks robustness. To understand if harmful concepts are effectively removed from the T2I model capabilities, we perform an additional experiment based on text-based adversarial attacks. Those allow to optimize text apparently innocuous, but leading to unsafe generation. If AlignGuard is able to prevent harmful generations in this setup, it means that our alignment is effective even if harmful concepts are not explicitly present in the input prompt.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17030", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ac6497bcc8fc3a8317cebcb6fbd562db60eb152fc5469ee7f853e355ec1bd634.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ExpertIP on specific category ↓Avg.
HateHarass.ViolenceSelf-harmSexualShockingIllegal
No alignment0.490.480.540.590.540.520.440.51
Hate0.110.150.190.160.110.170.110.14
Harass.0.170.160.200.230.190.160.150.18
Violence0.180.150.160.190.180.170.160.17
Self-Harm0.210.240.260.250.220.240.230.23
Sexual0.180.190.250.250.160.210.190.20
Shocking0.170.180.220.200.150.190.160.18
Illegal0.170.140.190.200.200.210.120.18
All - Single0.170.210.220.220.160.210.180.16
All - Ours0.060.060.090.070.070.080.040.07
", + "bbox": [ + 94, + 88, + 483, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/87a38a19d3f9a8dc3ddc5d4aeca2664ced9bdc06c9b24560665987e6fc16ad5e.jpg", + "table_caption": [ + "Table 3. Effectiveness of merging. While training a single safety expert across all data (All - single), IP performance are lower or comparable to single experts (previous rows). Instead, by merging safety experts (All - ours) we considerably improve results." + ], + "table_footnote": [], + "table_body": "
MethodMMA[59]Ring-A-Bell[55]SneakyPrompt[60]P4D[11]
No align.0.430.670.570.48
ESD-u [16]0.220.430.210.23
AlignGuard0.080.120.090.07
", + "bbox": [ + 94, + 387, + 483, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use 4 state-of-the-art attack methods listed in Table 4 to optimize seemingly innocuous prompts, starting from Co-ProV2 test prompts. We test alignment with AlignGuard and ESD-u [16]. The results advocate for AlignGuard's high robustness to adversarial attacks, proof of the effectiveness of alignment. In particular, for Ring-A-Bell [55], AlignGuard maintains an IP of 0.12, even for a particularly successful attack able to raise the IP of the baseline SD v1.5 to 0.67.", + "bbox": [ + 89, + 537, + 483, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 681, + 251, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DPO strategy. In AlignGuard training, several options exist for selecting preferred samples for DPO. Indeed, by selecting a paired safe image as a safe output for an unsafe prompt, we depart slightly from common practices in LLM alignment, where the preferred answer in presence of an unsafe input is a refusal [13]. We here investigate this choice by mimicking refusals supervision in language models. To do so, we create two different alternative images that we use as positive samples for DPO in presence of an unsafe input: (1) a Black Image, and (2) an image constructed by applying a Warning Sign on top of an unsafe output. Examples are in the supplement. From results in Table 5a, we evince that our strategy exploiting a paired safe image is perform-", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/862b35c8486c81b2e7f3a9e91f1b6b809f183636bf8eea800b90dff9d0f6ef68.jpg", + "table_caption": [ + "Table 4. Resistance to adversarial attacks. We evaluate with 4 adversarial attacks methods the performance of AlignGuard and the best baseline, ESD-u, in terms of IP. For a wide range of attacks, we are able to outperform the baselines, advocating for the effectiveness of our scalable concept removal strategy." + ], + "table_footnote": [], + "table_body": "
DPO strategyIP ↓FID ↓CLIP ↑
Black image0.2185.3831.22
Warning sign0.1670.0133.05
w/o Lcon0.1674.8230.78
Paired safe image (ours)0.0770.9632.32
", + "bbox": [ + 517, + 88, + 903, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/80274c555d5d57f9750a88f4526d923cd27eb4cce6d79cd0efb636582f4b3bd0.jpg", + "table_caption": [ + "(a) DPO strategy ablation" + ], + "table_footnote": [], + "table_body": "
MergingIP ↓FID ↓CLIP ↑DataIP ↓FID ↓CLIP ↑
Git Re-Basin [5]0.3069.9833.4010%0.1972.8727.30
Task Vectors [26]0.25179.604.8025%0.1671.9127.34
Model Soups [57]0.1469.8333.4050%0.1772.7627.20
TIES [58]0.0979.7928.85100%0.0770.9632.32
Co-Merge0.0770.9632.32
", + "bbox": [ + 517, + 189, + 903, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Merging method ablation", + "bbox": [ + 545, + 270, + 699, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(c) Scaling performance.", + "bbox": [ + 751, + 270, + 883, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies. We check the effects of alternative strategies for DPO, proving that our approach is best (a). Co-Merge is also the best merging strategy compared to baselines (b). Finally, we verify that scaling data improve our performance (c).", + "bbox": [ + 511, + 294, + 906, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing significantly better, justifying the different design of the DPO positive between T2I and LLM alignment. Also, we compared in the same table with a setup excluding $\\mathcal{L}_{con}$ , i.e. training without considering generative capabilities preservation among safe samples. This results in lower FID and higher CLIP scores, highlighting the importance of $\\mathcal{L}_{con}$ to preserve generative capabilities in presence of safe inputs.", + "bbox": [ + 511, + 367, + 906, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Merging strategy. We evaluate our merging method against four baselines: (1) Git Re-Basin, (2) Task Vectors, (3) Model Soups [57] and (4) TIES [58]. We report results in Table 5b, showing that Co-Merge consistently outperform these baselines. Let us highlight that baselines do not use data for performing merging, leading to suboptimal results and at the cost of tedious hyperparameter tuning. By performing a data-aware merging using unsafe prompts, instead, we are able to optimally balance the contributions of each expert with minimal effort and computational load.", + "bbox": [ + 511, + 479, + 908, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scaling performance. We evaluate the importance of the scale of data in Table 5c. We subsample CoProV2 with different percentages and retrain SD v1.5 with AlignGuard. For datasets increasingly smaller, we observe an expected decrease in performance. With $100\\%$ of the data, we perform best, proving that AlignGuard benefits the most from large datasets and paving the way for larger trainings.", + "bbox": [ + 511, + 635, + 908, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced AlignGuard, the first approach dedicated to safety alignment for T2I models at scale. By achieving safety alignment on a broad range of concepts, we allow for safe generations in the presence of harmful inputs. AlignGuard enables text-to-image models to generate outputs that are not only of high quality but also aligned with safety and ethical guidelines. We hope our work will inspire future practices in T2I deployment.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the Research Grant Council of the Hong Kong Special Administrative Region under grant number 16212623. FP is funded by KAUST (Grant DFR07910). AK, JG, and PT are supported by UKRI grant: Turing AI Fellowship EP/W002981/1, and by the Royal Academy of Engineering under the Research Chair and Senior Research Fellowships scheme. The authors thank Alexander Pondaven and David Romero for their kind proofreading.", + "bbox": [ + 89, + 114, + 485, + 241 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 268, + 187, + 284 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] The complete list of banned words in midjourney you need to know, 2022. Link. 2", + "[2] Diffusers: State-of-the-art diffusion models, 2022. Link. 1, 2", + "[3] Dall-e 3 system card, 2023. Link. 2", + "[4] Leonardo ai content moderation filter: Everything you need to know, 2023. Link. 2", + "[5] Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 8", + "[6] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2", + "[7] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. TMLR, 2024. 6", + "[8] Rishabh Bhardwaj, Do Duc Anh, and Soujanya Poria. Language models are a hero simpson! safety re-alignment of fine-tuned language models through task arithmetic. In ACL, 2024. 2", + "[9] Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. Typology of risks of generative text-to-image models. In Conference on AI, Ethics, and Society, 2023. 1", + "[10] Ruchika Chavhan, Da Li, and Timothy Hospedales. Conceptprune: Concept editing in diffusion models via skilled neuron pruning. arXiv preprint arXiv:2405.19237, 2024. 2", + "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. In ICML, 2024. 8", + "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2", + "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3, 8", + "[14] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, et al. Near to mid-term risks and opportunities of open source generative ai. In ICML, 2024. 1" + ], + "bbox": [ + 93, + 294, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Aaron Purewal, Csaba Botos, et al. Risks and opportunities of open-source generative ai. arXiv preprint arXiv:2405.08597, 2024. 1", + "[16] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In ICCV, 2023. 2, 6, 8", + "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In WACV, 2024. 2, 6", + "[18] Vipul Gupta, Santiago Akle Serrano, and Dennis DeCoste. Stochastic weight averaging in parallel: Large-batch training that generalizes well. In ICLR, 2020. 2", + "[19] Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio Pizzati, Philip Torr, Adel Bibi, Bernard Ghanem, and Mete Ozay. Model merging and safety alignment: One bad model spoils the bunch. In EMNLP Findings, 2024. 2", + "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 6", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6", + "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshops, 2022. 2", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 5", + "[25] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. In $COLM$ , 2024. 5", + "[26] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In ICLR, 2022. 2, 8", + "[27] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 2", + "[28] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In ICLR, 2023. 2", + "[29] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2", + "[30] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In CVPR, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17032", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6", + "[32] Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, and Fabio Pizzati. Latent guard: a safety framework for text-to-image generation. In ECCV, 2024. 2, 3, 5, 6", + "[33] Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. Videodpo: Omni-preference alignment for video diffusion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8009-8019, 2025. 3", + "[34] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In CVPR, 2024. 2", + "[35] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A holistic approach to undesired content detection in the real world. In AAAI, 2023. 2", + "[36] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 2022. 2", + "[37] Mistral. Mistral nemo, 2024. Link. 6", + "[38] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In ICCV, 2023. 2", + "[39] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. NeurIPS, 36, 2024. 2", + "[40] Seongbeom Park, Suhong Moon, Seunghyun Park, and Jinkyu Kim. Localization and manipulation of immoral visual cues for safe text-to-image generation. In WACV, 2024. 2", + "[41] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2024. 3", + "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6", + "[43] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Removing nsfw concepts from vision-and-language models for text-to-image retrieval and generation. In ECCV, 2024. 2", + "[44] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. In Conference on Computer and Communications Security, 2023. 5, 6", + "[45] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3", + "[46] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. In NeurIPS workshops, 2022. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Github Repo. Nudenet, 2022. Link. 6", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 5, 6", + "[49] Téo Sanchez. Examining the text-to-image community of practice: Why and how do people prompt generative ais? In Conference on Creativity and Cognition, 2023. 1", + "[50] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In FAccT, 2022. 6", + "[51] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 5, 6", + "[52] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 2", + "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2", + "[54] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV. Springer, 2025. 2", + "[55] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In ICLR, 2024. 8", + "[56] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 2, 3", + "[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, 2022. 2, 8", + "[58] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. In NeurIPS, 2024. 2, 8", + "[59] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. In CVPR, 2024. 8", + "[60] Yuchen Yang, Bo Hui, Haolin Yuan, Neil Gong, and Yinzhi Cao. Sneakyprompt: Jailbreaking text-to-image generative models. In Symposium on Security and Privacy, 2024. 8", + "[61] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In ICML, 2024. 2", + "[62] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In CVPR, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[63] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2", + "bbox": [ + 91, + 90, + 482, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17034", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_model.json b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0fbeba7dadf317a64c909d785bd575cb49f3204f --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.853, + 0.154 + ], + "angle": 0, + "content": "AlignGuard: Scalable Safety Alignment for Text-to-Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.18, + 0.8, + 0.199 + ], + "angle": 0, + "content": "Runtao Liu\\(^{1*}\\) I Chieh Chen\\(^{1*}\\) Jindong Gu\\(^{2}\\) Jipeng Zhang\\(^{1}\\) Renjie Pi\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.199, + 0.767, + 0.216 + ], + "angle": 0, + "content": "Qifeng Chen\\(^{1}\\) Philip Torr\\(^{2}\\) Ashkan Khakzar\\(^{2}\\) Fabio Pizzati\\(^{2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.216, + 0.855, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Hong Kong University of Science and Technology \\(^{2}\\)University of Oxford \\(^{3}\\)MBZUAI" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.237, + 0.632, + 0.251 + ], + "angle": 0, + "content": "{rliuay,icchen}@connect.ust.hk" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.253, + 0.667, + 0.27 + ], + "angle": 0, + "content": "https://alignguard.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.655 + ], + "angle": 0, + "content": "Text-to-image (T2I) models are widespread, but their limited safety guardrails expose end users to harmful content and potentially allow for model misuse. Current safety measures are typically limited to text-based filtering or concept removal strategies, able to remove just a few concepts from the model's generative capabilities. In this work, we introduce AlignGuard, a method for safety alignment of T2I models. We enable the application of Direct Preference Optimization (DPO) for safety purposes in T2I models by synthetically generating a dataset of harmful and safe image-text pairs, which we call CoProV2. Using a custom DPO strategy and this dataset, we train safety experts, in the form of low-rank adaptation (LoRA) matrices, able to guide the generation process away from specific safety-related concepts. Then, we merge the experts into a single LoRA using a novel merging strategy for optimal scaling performance. This expert-based approach enables scalability, allowing us to remove \\(7 \\times\\) more harmful concepts from T2I models compared to baselines. AlignGuard consistently outperforms the state-of-the-art on many benchmarks and establishes new practices for safety alignment in T2I networks." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.657, + 0.482, + 0.673 + ], + "angle": 0, + "content": "Warning: this paper includes potentially offensive content." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.699, + 0.222, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Text-to-image (T2I) models have advanced rapidly in recent years, becoming key tools in content creation and entertainment, used by both professionals and hobbyists [48, 49]. Such an impressive growth in image synthesis capabilities, in some cases reaching indistinguishable realism from real samples, has generated concerns about the risks of releasing such models [9, 14, 15]. Indeed, T2I may generate unsettling or harmful content, and expose people, particularly minors or those from sensitive cultural backgrounds, to inappropriate outputs. Moreover, malicious users could" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.302, + 0.907, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.67 + ], + "angle": 0, + "content": "Figure 1. Safety alignment for T2I. T2I models released without safety alignment risk to be misused (top). We propose Align-Guard, a scalable safety alignment framework for T2I models supporting the mass removal of harmful concepts (middle). We allow for scalability by training safety experts focusing on separate categories such as \"Hate\", \"Sexual\", \"Violence\", etc. We then merge the experts with a novel strategy. By doing so, we obtain safety-aligned models, mitigating unsafe content generation (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "easily generate, among others, violent pictures, or sexually-explicit deepfakes, potentially causing misinformation or harassing specific users. Regardless of these risks, there has been little effort in proposing effective safety alignment techniques, i.e. strategies for limiting harmful outputs, for T2I models. Most of the current approaches are focused on simple input filtering strategies, such as the analysis of the input prompt, or detectors on output images [2]. However, these strategies are easy to circumvent [46], or they can be deactivated if models are openly released. Conversely, large language models (LLMs) are subject to rigorous safety alignment procedures before being deployed [13]. It is our objective to bridge this gap." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.227, + 0.9 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17024" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": "For preventing harmful outputs, one straightforward approach might seem to be training on curated datasets excluding unsafe images. However, this is nontrivial. Large T2I trainings usually use vast web-crawled datasets, such as LAION-5B [53]. Despite filtering efforts, these datasets contain unsafe or illegal content, making full control nearly impossible. More realistically, we could limit the capability of the T2I to generate harmful concepts post-training. Recent works explored methods to delete specific concepts from pre-trained models, with promising results [16, 17, 29, 34]. Yet, these methods limitedly scale. While they can remove a limited number of concepts, deleting hundreds strongly degrades the model's generative capacity [17, 34]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.638 + ], + "angle": 0, + "content": "To address these challenges, we introduce AlignGuard, a scalable method for safety alignment of T2I models. Our idea is to remove a large number of concepts from T2I models, by exploiting an ensemble of safety experts trained with Direct Preference Optimization (DPO) on safety-oriented synthetic data. By doing so, we are the first to propose a scalable safety alignment method for T2I, allowing for a safe deployment of trained T2I models. Considering the lack of data for DPO-based safety alignment available for image generation, we introduce CoProV2, a dataset encompassing a broad array of unsafe concepts. CoProV2 generation is fully automatic, including paired images with associated prompts: one harmful image with its corresponding textual description, and one safe image and prompt with similar content and structure. This setup enables the application of DPO for aligning T2I models effectively. In practice, we train Low-Rank Adaptation (LoRA) [24] matrices. Each LoRA serves as a safety expert trained to prevent the generation of specific harmful content, i.e. content related to different categories such as \"Hate,\" \"Sexual,\" \"Violence,\" among others. The LoRAs guide the vision representation of the pre-trained T2I model towards safe outputs during the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.808 + ], + "angle": 0, + "content": "After training the safety experts, we use a new algorithm called Co-Merge to merge all safety LoRA experts into one. This mitigates interferences among different experts, allowing for a significant increase in scalability. As shown in Figure 1, through joint expert training and subsequent merging, we can train on images encompassing 723 harmful concepts—approximately seven times more than existing approaches [17]. In summary, AlignGuard enables safety alignment of T2I models at scale, preventing the generation of inappropriate outputs without impacting generative capabilities on safe prompts. Our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.841 + ], + "angle": 0, + "content": "1. We propose AlignGuard, the first scalable approach for T2I models specifically targeting safety alignment." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.841, + 0.483, + 0.886 + ], + "angle": 0, + "content": "2. We introduce a training method based on expert models, and a novel merging strategy based on weight activation frequency." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3. For the training of AlignGuard, we craft a new dataset" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "for safety alignment in T2I, coined CoProV2, that we will release as open source." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.167 + ], + "angle": 0, + "content": "4. We validate AlignGuard on popular T2I models and multiple benchmarks, demonstrating its effectiveness in many setups." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.183, + 0.655, + 0.198 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.907, + 0.375 + ], + "angle": 0, + "content": "Content filtering. Often, closed-source commercial T2I use prompt blacklists, LLM preprocessing, and image analysis for content filtering [1, 3, 4]. In particular, some use LLM-based prompt analysis [35] to detect harmful inputs, even with ad-hoc models [27]. Recently, Latent Guard [32] proposed a latent-based blacklist within text encoders in T2I models. Some works rely instead on the analysis of generated images [2], where an NSFW classifier is applied to generated images. Instead, others [40] use inpainting to mask potentially unsafe content. These approaches can easily be deactivated if the T2I models' weights are available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.61 + ], + "angle": 0, + "content": "Concept removal in T2I. Some have explored removing the capability of generating concepts in T2I, like SLD [52], using classifier-free guidance [22] to steer generation away from undesirable outputs. Similarly, Li et al. [30] identify interpretable directions that can be used for safety steering. Alternatively, many fine-tune the T2I. While seminal works finetune the entire model [29], some focus on specific components, such as attentions [17, 38, 62], specific neurons [10], or the textual encoder [43]. There has been a recent interest in mass removal of concepts from T2I models for safety purposes [17, 34]. While MACE [34] uses a similar strategy as ours, they make use of segmentation masks, constraining the erasable concepts at spatially-defined ones. Moreover, all available approaches [17, 34] degrade performance with more than 100 erased concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.831 + ], + "angle": 0, + "content": "Model merging. Model merging, i.e. combining multiple models into one model, has gained attention [12, 61]. Weighted averaging methods are commonly used to enhance performance [18, 26, 57], especially in scenarios where only model weights are accessible [28]. Beyond these, advanced techniques have emerged, improving over basic averaging [6, 36, 39]. For example, TIES [58] resolves operator conflicts to improve merging, at the cost of hyperparameter tuning. Merging has been explored for safety in LLMs [8, 19] and for concept removal in MACE [34], but only using expensive optimization procedures. Existing LoRA merging for T2I focus on single subjects [54, 63]. Instead, we propose a cheap and effective merging strategy, focusing on broad safety categories." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.652, + 0.86 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We use preference optimization algorithms such as DPO [56] to perform scalable safety alignment of T2I mod" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17025" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "els. Here, we revise the fundamental concepts to allow for the interpretation of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.132, + 0.367, + 0.148 + ], + "angle": 0, + "content": "3.1. Text-to-image diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.485, + 0.321 + ], + "angle": 0, + "content": "Diffusion models allow for image generation by iteratively denoising gaussian noise with a network \\(\\epsilon\\) for \\(t\\in [0,T]\\) iterations [23]. In particular, T2I diffusion models include natural language conditioning, so the desired output image can be described with text. We now briefly introduce the training procedure for a T2I model. Let us assume an input pair \\((x,p)\\sim \\mathcal{D}_{\\mathrm{train}}\\) sampled from a training dataset \\(\\mathcal{D}_{\\mathrm{train}}\\) of images \\(x\\) and paired textual description \\(p\\). The network \\(\\epsilon\\) is trained by estimating the ground truth noise \\(\\tilde{\\epsilon}_t\\) injected on \\(x\\) for a random \\(t\\in [0,T]\\). We define the input image with the addition of noise as \\(x_{t}\\). Hence, the training loss is" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.333, + 0.483, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} (\\epsilon , x, p) = | | \\tilde {\\epsilon} _ {t} - \\epsilon \\left(x _ {t}, p\\right) | |. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.394 + ], + "angle": 0, + "content": "In Eq. (1), \\(\\epsilon (\\cdot)\\) is the denoising operation. The network weights \\(\\theta^{*}\\) are optimized by minimizing the following:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.407, + 0.483, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, p) \\sim \\mathcal {D} _ {\\text {t r a i n}}} \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\theta}, x, p\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.485, + 0.514 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) refers to the denoising network with a set of weights \\(\\theta\\). During inference, sampled Gaussian noise is iteratively processed for \\(t \\in [0, T]\\) with \\(\\epsilon_{\\theta^*}\\) following specific scheduling policies, ultimately allowing image synthesis. We refer to [23] for additional details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.325, + 0.538 + ], + "angle": 0, + "content": "3.2. DPO for diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.697 + ], + "angle": 0, + "content": "DPO is a technique for preference alignment initially developed for LLMs [41, 45] and recently extended to diffusion models [33, 56]. The core idea is to benefit from pairwise preferences obtained by labeling. Let us assume a dataset \\(\\mathcal{D}_{\\mathrm{DPO}}\\) of paired images and textual descriptions \\((x^{+}, x^{-}, p)\\). For a given description \\(p\\), \\(x^{+}\\) is a image that humans indicated as preferred output with respect to \\(x^{-}\\). The intuition of DPO is to increase the likelihood to generate the preferred output \\(x^{+}\\), while discouraging the generation of \\(x^{-}\\). This translates into the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.709, + 0.482, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon , x ^ {+}, x ^ {-}, p\\right) = - \\log \\sigma (- \\beta ( \\\\ \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {+}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {+}, p\\right)\\right) \\tag {3} \\\\ - \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {-}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {-}, p\\right)\\right)\\left. \\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.839 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\mathrm{ref}}\\) is a reference pre-trained network, typically resulting from a previous optimization of Eq. 2. Also, \\(\\sigma\\) is the sigmoid operation and \\(\\beta\\) is a weighting constant [56]. Finally, one could optimize as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.85, + 0.484, + 0.875 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\mathrm {D P O}} ^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x ^ {+}, x ^ {-}, p) \\sim \\mathcal {D} _ {\\mathrm {D P O}}} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {+}, x ^ {-}, p\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.886, + 0.452, + 0.902 + ], + "angle": 0, + "content": "For further details, we refer to the original paper [56]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.925, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Dataset generation. For each unsafe concept in different categories, we generate an corresponding prompts with an LLM. We generated paired safe prompts using an LLM, minimizing semantic differences. Then, we use the T2I we intend to align to generate corresponding images for both prompts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.372, + 0.606, + 0.387 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.504 + ], + "angle": 0, + "content": "We aim to align T2I models inspired by practices in language model alignment [13]. We avoid expensive human annotations by generating safety-oriented preference data in Section 4.1. We also propose a training procedure based on expert networks (Section 4.2) and subsequent merging (Section 4.3). After training, AlignGuard requires no additional computation at inference (more details in supplement)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.799, + 0.532 + ], + "angle": 0, + "content": "4.1. DPO for safety alignment in T2I" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.704 + ], + "angle": 0, + "content": "We aim to use DPO as a strategy to perform safety alignment on pretrained T2I models. Assuming an unsafe textual input describing a sexually explicit, shocking, or violent scene, we would like the diffusion model to ignore the unsafe requested traits, and generate a safe image, i.e. not including any visually disturbing trait. We achieve this by generating automatically {unsafe, safe} image pairs for DPO training, obtained by propting the T2I with LLM-manipulated text. In a nutshell, we can discourage the generation of unsafe images, having as preferred DPO image a visually-close safe sample, for an unsafe input T2I prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We start by generating a set of unsafe concepts \\(\\mathcal{C}\\) with an LLM. Each \\(c\\in \\mathcal{C}\\) is a keyword representing unsafe elements that may be included in an image, such as \"homicide\", \"nude\", etc. Then, we follow the prompt synthesis procedure of [32], generating both unsafe and safe prompts for image generators from concepts, enforcing minimal differences. Specifically, we sample an unsafe prompt \\(p^{\\mathrm{U}}\\) with an LLM to create text visually describing a scene using the input concept \\(c\\). Then, we further process the generated \\(p^{\\mathrm{U}}\\), prompting an LLM to remove any reference to the input unsafe concept \\(c\\). This transforms the prompt into a safe \\(p^{\\mathrm{S}}\\), while minimizing semantic changes. Differently from [32], we process both prompts with a pretrained" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17026" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.096, + 0.599, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.097, + 0.9, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.289, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 3. Expert training and merging. First, we use the previously generated prompts and images to train LoRA experts on specific safety categories (left), exploiting our DPO-based losses. Then, we merge all the safety experts with Co-Merge (right). This allows to achieve general safety experts that produce safe outputs for a generic unsafe input prompt in any category." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.356, + 0.488, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.548 + ], + "angle": 0, + "content": "Figure 4. Merging experts with Co-Merge. (left) Assuming LoRA experts with the same architecture, we analyze which expert has the highest activation for each weights for all inputs. (right) Then, we obtained the merged weights from multiple experts, merging only the most active weights per expert." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.606 + ], + "angle": 0, + "content": "diffusion model, obtaining images \\(\\{x^{\\mathrm{U}}, x^{\\mathrm{S}}\\}\\), derived from \\(p^{\\mathrm{U}}\\) and \\(p^{\\mathrm{S}}\\), respectively. We show this in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.831 + ], + "angle": 0, + "content": "In this way, we can automatically create image preference pairs, assuming that the preferred output in safety alignment would always be a similar safe image for an unsafe prompt. Indeed, by repeating the process multiple times, we can construct a dataset \\(\\mathcal{D}_{\\mathrm{safety}}\\) consisting of tuples including the sampled prompts and corresponding generated images, i.e. \\((x^{\\mathrm{S}}, x^{\\mathrm{U}}, p^{\\mathrm{S}}, p^{\\mathrm{U}})\\). Then, we use DPO to discourage the generation of unsafe outputs for an unsafe prompt \\(p^{\\mathrm{U}}\\), and promote the generation of the paired safe image \\(x^{\\mathrm{S}}\\) instead. To achieve this, we set \\(x^{\\mathrm{S}}\\) as preferred output for \\(p^{\\mathrm{U}}\\), discouraging the generation of \\(x^{\\mathrm{U}}\\). The semantic similarity of \\(p^{\\mathrm{U}}\\) and \\(p^{\\mathrm{S}}\\) will encourage the DPO training to focus exclusively on the visual traits making \\(x^{\\mathrm{U}}\\) unsafe. We define a first loss term:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.845, + 0.483, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {U}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "While effective in removing the unsafe concepts \\( c \\) from the generated images, this risks to reduce the generative capa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.42 + ], + "angle": 0, + "content": "bilities of the network, since we are encouraging to generate images that do not respect the input prompt. To prevent this, we also use \\( x^{\\mathrm{S}} \\) as preferred generation for the input prompt \\( p^{\\mathrm{S}} \\), while penalizing on \\( x^{\\mathrm{U}} \\). This writes:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.43, + 0.907, + 0.45 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {S}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.477 + ], + "angle": 0, + "content": "Finally, we modify Eq. (4) proposing our training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.502, + 0.906, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\theta_ {\\text {A l i g n G u a r d}} ^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x ^ {\\mathrm {s}}, x ^ {\\mathrm {U}}, p ^ {\\mathrm {s}}, p ^ {\\mathrm {U}}) \\sim \\mathcal {D} _ {\\text {s a f e t y}}} \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {c o n}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.583 + ], + "angle": 0, + "content": "This objective allows for a stable safety-oriented finetuning of the T2I model for a large number of concepts \\( c \\in \\mathcal{C} \\) without impacting the generative capabilities of the T2I model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.838, + 0.608 + ], + "angle": 0, + "content": "4.2. Improving scaling with safety experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.888 + ], + "angle": 0, + "content": "Training with our proposed strategy in Section 4.1 enables us to scale the number of concepts used for safety alignment, while keeping generative capabilities intact. However, by increasing the number of concepts in \\(\\mathcal{C}\\), we noticed a decrease in performance. We suppose this is due to the complexity of the task: in safety alignment there is a need of a high contextual understanding of the generated scenes. As an example, an image of a kitchen including knives is a safe output, while a scene of harassment with the usage of a knife is an undesired unsafe outcome of the image synthesis process, that we aim to prevent. As a solution, we propose to decompose the safety alignment on separate categories, such as \"violence\", \"sexual content\", \"harassment\", and others. Our intuition is that by limiting the variability of the generated images, it would be easier to identify visual patterns that would make a scene \"violent\", or \"sexually explicit\". Hence, we aim to train safety experts, each focusing on one category only." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17027" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.093, + 0.09, + 0.484, + 0.341 + ], + "angle": 0, + "content": "Algorithm 1 Co-Merge \n1: Input: Safety expert LoRAs \\(\\{L_1, L_2, \\ldots, L_N\\}\\), unsafe prompts \\(\\{p_1, p_2, \\ldots, p_K\\}\\) \n2: Output: Merged LoRA \\(L_{\\text{merged}}\\) \n3: Init a count matrix \\(C\\) of \\(J \\times N\\) elements with zeros. \n4: for each prompt \\(p_k\\), \\(k \\in [1, K]\\) do \n5: for each neuron index \\(j\\), \\(j \\in [1, J]\\) do \n6: \\(i_{\\text{max}} := \\arg \\max_i |L_i^j(p_k)|\\), \\(i \\in [1, N]\\) \n7: Increment count \\(C[j, i_{\\text{max}}]\\) \n8: end for \n9: end for \n10: for each neuron index \\(j\\), \\(j \\in [1, J]\\) do \n11: \\(m := \\arg \\max_i C[j, i]\\), \\(i \\in [1, N]\\) \n12: \\(L_{\\text{merged}}^j := L_m^j\\) \n13: end for \n14: Return \\(L_{\\text{merged}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.584 + ], + "angle": 0, + "content": "In practice, we identified \\(N\\) safety-related concept categories, each linked to a generated concept set \\(\\{\\mathcal{C}^1,\\dots,\\mathcal{C}^N\\}\\). Each set is created by prompting a large language model to generate concepts within a given unsafe category. We then generated the corresponding safety datasets \\(\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}\\) following the approach in Section 4.1. Rather than fine-tuning the entire model's weights, \\(\\theta\\), to create the expert networks, we used low-rank adaptation (LoRA) [24], updating only a minimal subset of weights and as such saving significant computational costs. Ultimately, we apply Equation 7 for \\(\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}\\), obtaining a set of LoRAs \\(\\{L^1,\\dots,L^N\\}\\), as shown in Figure 3. At inference time, the obtained LoRAs can be applied to the pretrained T2I model to enable safe image generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.592, + 0.258, + 0.608 + ], + "angle": 0, + "content": "4.3. Experts merging" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.749 + ], + "angle": 0, + "content": "To prevent the generation of content across multiple safety-related categories, additional considerations are necessary. Indeed, loading multiple LoRAs into a model can lead to interference [25], reducing overall performance in safety-oriented image generation. Alternatively, multiple experts could be run independently, at the cost of multiplying inference costs. To mitigate these issues, we aim to merge the trained LoRAs into a single one, as visualized in Figure 3, without sacrificing performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We propose a novel data-based strategy, called CoMerge, to merge multiple safety experts, specifically designed around our use case. Our intuition is that by decomposing safety alignment in multiple categories (Section 4.2), trained LoRAs sharing the same architecture would encode information about category-specific alignment in different neurons. In other words, we assume that in the T2I model, the weights responsible for generating violent content (i.e. including, for instance, blood) will be different from those generating sexually-explicit content" + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.09, + 0.905, + 0.177 + ], + "angle": 0, + "content": "
DatasetImage # of Prompts # of Categories # of ConceptsIP
COCO [31]✓3,000N/AN/A0.06
I2P [51]✗4,7037N/A0.36
UD [44]✗9325N/A0.47
CoPro [32]✗56,52677230.23
CoProV2 (ours)✓23,69077230.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.186, + 0.907, + 0.257 + ], + "angle": 0, + "content": "Table 1. Datasets comparison. Our LLM-generated dataset, CoProV2, achieves comparable Inappropriate Probability (IP) to human-crafted datasets (UD [44], I2P [51]) and offers similar scale to CoPro [32]. COCO [31], exhibiting a low IP, is used as a benchmark for image generation with safe prompts as input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.386 + ], + "angle": 0, + "content": "(e.g. naked skin). LoRAs work as updates for the model weights [24], so, for concepts associated to different neurons, they will encode related information in different parts of the same architecture. Then, our idea is to isolate the neurons inside the trained safety experts that activate the most for unsafe input prompts, and construct a merged expert as a LoRA \\( L_{\\text{merged}} \\) including the parameters of different experts associated to the most significant activations only." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.909, + 0.766 + ], + "angle": 0, + "content": "We start by randomly sampling \\(K\\) unsafe prompts from the training set of \\(\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}\\), equally distributed across each category. We then process all prompts with the T2I model, where we load all trained experts \\(\\{L_1,\\dots,L_N\\}\\) one at the time. Each expert LoRA is a multi-layer perceptron including \\(J\\) neurons. For each prompt \\(p_k^U\\), where \\(k\\in [1,K]\\), and each expert model \\(L_{i}\\), where \\(i\\in [1,N]\\), we record the absolute value of the activation from each neuron \\(L_{i}^{j}\\), where \\(j\\in [1,J]\\). We denoted the absolute value as \\(|L_i^j (p_k^U)|\\). This provides a measure of the response for each neuron to that prompt. These prompts are uniformly sampled from distinct safety categories, so different neurons of the same index \\(j\\) across the experts will exhibit different activations in response to different prompts. To create the merged \\(L_{\\mathrm{merged}}\\), we identify the neurons with the highest activation frequencies across experts. Specifically, for each neuron \\(j\\) in the merged expert \\(L_{\\mathrm{merged}}\\), we select the neuron from the original set of experts \\(\\{L_1,\\dots,L_N\\}\\) that has the highest activation frequency count across the \\(K\\) prompts. We summarize this in Algorithm 1, where the \\(p^U\\) superscript is omitted for clarity. Figure 4 illustrates Co-Merge with \\(K = 10\\), describing how the frequency count leads to the network weight merging. Note that merging \\(L_{\\mathrm{merged}}\\) with the original T2I weights makes safety alignment difficulty reversible, a beneficial practice for T2I releases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.646, + 0.789 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.84 + ], + "angle": 0, + "content": "We first introduce our setup (Section 5.1), and compare with baselines in Section 5.2. We then analyze the properties of AlignGuard (Section 5.3) and show ablations (Section 5.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.702, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Baselines. We use AlignGuard to align three models for T2I, namely Stable Diffusion (SD) v1.5/v2.1 [48] and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17028" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.259 + ], + "angle": 0, + "content": "
MethodIP↓FID ↓ COCOCLIP ↑
CoProV2I2PUD
SD1.5No alignment0.510.360.5269.7733.52
SLD [51]0.270.190.3071.4532.24
ESD-u [16]0.220.250.2172.9829.61
UCE [17]0.330.300.3872.0132.01
AlignGuard0.070.110.1670.9632.32
SD2.1No alignment0.510.350.5570.4534.99
AlignGuard0.120.120.1771.9934.13
SDXLNo alignment0.490.310.4768.9035.60
AlignGuard0.090.080.1175.2034.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Table 2. Benchmarks. AlignGuard achieves best performance both in generated images alignment (IP) and image quality (FID, CLIPScore), with two T2I models and against 3 methods for SD v1.5. Note that we use CoProV2 only for training, hence I2P and UD are out-of-distribution. Yet, AlignGuard allows a robust safety alignment. Best is bold, second is underlined." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.487 + ], + "angle": 0, + "content": "SDXL [42]. For all, we compare the original versions released on HuggingFace, and our finetuned version with AlignGuard. Moreover, we compare with recent methods based on SD v1.5 for safe image generation (SLD [51]) and concept erasure (ESD [16], UCE [17]). In particular, for ESD, we follow the paper indications and use ESD-u [16], a version of ESD for broad concept removal. We perform our analysis and ablations on SD v1.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.604 + ], + "angle": 0, + "content": "Metrics. To evaluate the effectiveness of safety alignment, we use the inappropriate probability metric (IP) from SLD [51], measuring the ratio of unsafe contents generated by T2I with unsafe prompts. The unsafe content detection is performed by using classification results of Q16 [50] and NudeNet [47]. For evaluating image quality, we use FID [21], and CLIPScore [20] for text-image alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. For training of AlignGuard, we generate a new dataset following Section 4.1, named CoProV2 (Concepts and Prompts). We motivate this choice by observing a limitation in the CoPro dataset [32], i.e. the limited IP measured on images generated with the prompts of the dataset. This suggests that many prompts in CoPro are not leading to unsafe generations. This is due to the limited performance of the LLM building CoPro. We start from the same \\(\\mathcal{C}\\) as CoPro: 723 harmful concepts across 7 categories (Hate, Harassment, Violence, Self-Harm, Sexual, Shocking, Illegal). We re-generate the unsafe/safe prompts pairs with a new LLM prompt reported in the supplement. We use Mistral-Nemo-Instruct [37] for the prompt synthesis. We generate 23,690 pairs of safe/unsafe prompts, and we use 15,690/8,000 pairs for training/ testing. For testing only, we use I2P [7] and Unsafe Diffusion (UD) [44], including 4703/932 human-designed unsafe prompts on 7/5 categories, respectively. Following common practices [16, 48, 51] we use COCO [31], with 3000 safe" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "captions and corresponding images, for evaluating the performance of the T2I model in image generation after alignment. We compare the datasets quantitatively in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.142, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Implementation details. We train AlignGuard with batch-size 8, accumulating over 16 steps. We use AdamW with a learning rate of \\(10^{-5}\\) for training each LoRA with rank 4. For Co-Merge, we set \\(K = 100\\) unsafe prompts. Each LoRA takes \\(\\sim 2\\) hours to train for SD and \\(\\sim 18\\) hours for SDXL on a single Nvidia 5880 Ada GPU for 2000 steps. We normalize the computation needed for training AlignGuard and baselines. More details are in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.27, + 0.649, + 0.285 + ], + "angle": 0, + "content": "5.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.908, + 0.868 + ], + "angle": 0, + "content": "Quantitative evaluation. In Table 2, we compare with alignment baselines for IP, FID, and CLIPScore. For fairness, all baselines were re-trained on CoProV2 training set and evaluated using the CoProV2 test set, I2P, and UD. Note that I2P and UD serve as out-of-distribution evaluations as they are crafted by human users in real deployment scenarios, differing in both prompt format and content from CoProV2. Initially, we trained the baselines using the complete \\(\\mathcal{C}\\) set of CoProV2 with Stable Diffusion v1.5 as the backbone. This involved using \\(\\mathcal{C}\\) to identify negative embeddings in SLD and excluding all \\(\\mathcal{C}\\) concepts for ESD-u and UCE. All baselines collapsed due to the high number of concepts in CoProV2, aligning with the limited scaling performance noted in previous works [16, 17]. These results are included in the supplement. For successful training comparisons, we used the 7 category names of CoProV2 as removable concepts for the baselines, while we use the full prompts and images in CoProV2 for AlignGuard. We trained our expert LoRAs on samples generated with all \\(\\mathcal{C}\\), one per category, and merge them with Co-Merge. As shown in the table, we outperform considerably all baselines in every metric. Notably, our IP for SD v1.5 on CoProV2 (0.07) is much lower than the next best (ESD-u, 0.22). Additionally, AlignGuard demonstrates superior generalization, achieving better IP scores on unseen I2P (0.08) and UD (0.16). Our DPO-based strategy effectively maintains Stable Diffusion's generation capabilities preventing forgetting of concepts unrelated to safety. Indeed we report nearing the performance of the Baseline method without safety alignment in FID (70.96 vs 69.77) and CLIPScore (32.32 vs 33.52) on COCO captions. We also applied AlignGuard to SD v2.1 and SDXL to test the generalization of our method to a different diffusion model. Our results are consistent with different backbones like SD v2.1 and SDXL, achieving performance comparable to our aligned SD v1.5 with IP 0.09 and 0.12 on CoProV2. This suggests that AlignGuard can be used for the safety alignment of arbitrary T2I models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative evaluation. We show in Figure 5 a selection of samples advocating the effectiveness of AlignGuard. We" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17029" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.907, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.491, + 0.907, + 0.534 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison. Compared to non-aligned baseline models, AlignGuard allows to synthesize safe images for unsafe input prompts. Please note the layout similarity between the unsafe and safe outputs: thanks to our training, only the harmful image traits are removed from the generated images. Concepts in ⟨brackets⟩. Prompts are shortened, for full ones see the supplement." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.699 + ], + "angle": 0, + "content": "sample prompts from CoProV2. For SD v1.5 and SDXL (SD v2.1 in the supplement), we prevent the generation of unsafe content across 7 categories. Please also note how the quality of the generated images, in both models aligned with AlignGuard, is not impacted by our alignment procedure. Also, let us highlight that although the model refuses to follow unsafe prompts, the generated images preserve similar traits and global layout, e.g. the pose of the man in the \"Shocking\" column for SDXL. This property results from training on safe/unsafe pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.208, + 0.728 + ], + "angle": 0, + "content": "5.3. Properties" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Experts performance. For AlignGuard, we train an expert for each category in CoProV2. We now examine the effectiveness of each expert LoRA across all categories in the CoProV2 test set in Table 3. The results show that training and evaluating on the same category generally leads to good performance, as expected. Interestingly though, applying one expert LoRA consistently improves IP across all categories and often surpasses the expert trained specifically for that category. We attribute this to the interaction of multiple concepts across categories. While the concepts differ, the visual features of generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.76 + ], + "angle": 0, + "content": "images may share similarities. For example, scenes in the \"Violence\" or \"Self-harm\" categories often depict blood, while the \"Hate\" category may include sexual slurs, which could explain the strong performance on \"Sexual\" data. Ultimately, this suggests that for safety alignment tasks, AlignGuard is robust to the concept selection. However, experts trained on different categories still encode distinct concepts. This becomes evident when merging all LoRAs with Co-Merge (All - Ours), since we achieve the best performance across all categories. We also include a baseline where a single LoRA is trained on prompts from all categories (All - single), which shows suboptimal performance. Here, the lack of specialized expert models results in less effective training, justifying our design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Adversarial attacks robustness. To understand if harmful concepts are effectively removed from the T2I model capabilities, we perform an additional experiment based on text-based adversarial attacks. Those allow to optimize text apparently innocuous, but leading to unsafe generation. If AlignGuard is able to prevent harmful generations in this setup, it means that our alignment is effective even if harmful concepts are not explicitly present in the input prompt." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17030" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.485, + 0.308 + ], + "angle": 0, + "content": "
ExpertIP on specific category ↓Avg.
HateHarass.ViolenceSelf-harmSexualShockingIllegal
No alignment0.490.480.540.590.540.520.440.51
Hate0.110.150.190.160.110.170.110.14
Harass.0.170.160.200.230.190.160.150.18
Violence0.180.150.160.190.180.170.160.17
Self-Harm0.210.240.260.250.220.240.230.23
Sexual0.180.190.250.250.160.210.190.20
Shocking0.170.180.220.200.150.190.160.18
Illegal0.170.140.190.200.200.210.120.18
All - Single0.170.210.220.220.160.210.180.16
All - Ours0.060.060.090.070.070.080.040.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.377 + ], + "angle": 0, + "content": "Table 3. Effectiveness of merging. While training a single safety expert across all data (All - single), IP performance are lower or comparable to single experts (previous rows). Instead, by merging safety experts (All - ours) we considerably improve results." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.388, + 0.485, + 0.445 + ], + "angle": 0, + "content": "
MethodMMA[59]Ring-A-Bell[55]SneakyPrompt[60]P4D[11]
No align.0.430.670.570.48
ESD-u [16]0.220.430.210.23
AlignGuard0.080.120.090.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Table 4. Resistance to adversarial attacks. We evaluate with 4 adversarial attacks methods the performance of AlignGuard and the best baseline, ESD-u, in terms of IP. For a wide range of attacks, we are able to outperform the baselines, advocating for the effectiveness of our scalable concept removal strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.673 + ], + "angle": 0, + "content": "We use 4 state-of-the-art attack methods listed in Table 4 to optimize seemingly innocuous prompts, starting from Co-ProV2 test prompts. We test alignment with AlignGuard and ESD-u [16]. The results advocate for AlignGuard's high robustness to adversarial attacks, proof of the effectiveness of alignment. In particular, for Ring-A-Bell [55], AlignGuard maintains an IP of 0.12, even for a particularly successful attack able to raise the IP of the baseline SD v1.5 to 0.67." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.253, + 0.697 + ], + "angle": 0, + "content": "5.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "DPO strategy. In AlignGuard training, several options exist for selecting preferred samples for DPO. Indeed, by selecting a paired safe image as a safe output for an unsafe prompt, we depart slightly from common practices in LLM alignment, where the preferred answer in presence of an unsafe input is a refusal [13]. We here investigate this choice by mimicking refusals supervision in language models. To do so, we create two different alternative images that we use as positive samples for DPO in presence of an unsafe input: (1) a Black Image, and (2) an image constructed by applying a Warning Sign on top of an unsafe output. Examples are in the supplement. From results in Table 5a, we evince that our strategy exploiting a paired safe image is perform-" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.174 + ], + "angle": 0, + "content": "
DPO strategyIP ↓FID ↓CLIP ↑
Black image0.2185.3831.22
Warning sign0.1670.0133.05
w/o Lcon0.1674.8230.78
Paired safe image (ours)0.0770.9632.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.642, + 0.176, + 0.779, + 0.189 + ], + "angle": 0, + "content": "(a) DPO strategy ablation" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.27 + ], + "angle": 0, + "content": "
MergingIP ↓FID ↓CLIP ↑DataIP ↓FID ↓CLIP ↑
Git Re-Basin [5]0.3069.9833.4010%0.1972.8727.30
Task Vectors [26]0.25179.604.8025%0.1671.9127.34
Model Soups [57]0.1469.8333.4050%0.1772.7627.20
TIES [58]0.0979.7928.85100%0.0770.9632.32
Co-Merge0.0770.9632.32
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.271, + 0.7, + 0.284 + ], + "angle": 0, + "content": "(b) Merging method ablation" + }, + { + "type": "table_caption", + "bbox": [ + 0.752, + 0.271, + 0.884, + 0.284 + ], + "angle": 0, + "content": "(c) Scaling performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Table 5. Ablation studies. We check the effects of alternative strategies for DPO, proving that our approach is best (a). Co-Merge is also the best merging strategy compared to baselines (b). Finally, we verify that scaling data improve our performance (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.476 + ], + "angle": 0, + "content": "ing significantly better, justifying the different design of the DPO positive between T2I and LLM alignment. Also, we compared in the same table with a setup excluding \\(\\mathcal{L}_{con}\\), i.e. training without considering generative capabilities preservation among safe samples. This results in lower FID and higher CLIP scores, highlighting the importance of \\(\\mathcal{L}_{con}\\) to preserve generative capabilities in presence of safe inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.632 + ], + "angle": 0, + "content": "Merging strategy. We evaluate our merging method against four baselines: (1) Git Re-Basin, (2) Task Vectors, (3) Model Soups [57] and (4) TIES [58]. We report results in Table 5b, showing that Co-Merge consistently outperform these baselines. Let us highlight that baselines do not use data for performing merging, leading to suboptimal results and at the cost of tedious hyperparameter tuning. By performing a data-aware merging using unsafe prompts, instead, we are able to optimally balance the contributions of each expert with minimal effort and computational load." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.909, + 0.743 + ], + "angle": 0, + "content": "Scaling performance. We evaluate the importance of the scale of data in Table 5c. We subsample CoProV2 with different percentages and retrain SD v1.5 with AlignGuard. For datasets increasingly smaller, we observe an expected decrease in performance. With \\(100\\%\\) of the data, we perform best, proving that AlignGuard benefits the most from large datasets and paving the way for larger trainings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We introduced AlignGuard, the first approach dedicated to safety alignment for T2I models at scale. By achieving safety alignment on a broad range of concepts, we allow for safe generations in the presence of harmful inputs. AlignGuard enables text-to-image models to generate outputs that are not only of high quality but also aligned with safety and ethical guidelines. We hope our work will inspire future practices in T2I deployment." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17031" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.242 + ], + "angle": 0, + "content": "This research was supported by the Research Grant Council of the Hong Kong Special Administrative Region under grant number 16212623. FP is funded by KAUST (Grant DFR07910). AK, JG, and PT are supported by UKRI grant: Turing AI Fellowship EP/W002981/1, and by the Royal Academy of Engineering under the Research Chair and Senior Research Fellowships scheme. The authors thank Alexander Pondaven and David Romero for their kind proofreading." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.269, + 0.188, + 0.285 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.295, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[1] The complete list of banned words in midjourney you need to know, 2022. Link. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.325, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[2] Diffusers: State-of-the-art diffusion models, 2022. Link. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.34, + 0.368 + ], + "angle": 0, + "content": "[3] Dall-e 3 system card, 2023. Link. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.37, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[4] Leonardo ai content moderation filter: Everything you need to know, 2023. Link. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[5] Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.482, + 0.484 + ], + "angle": 0, + "content": "[6] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[7] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. TMLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.57 + ], + "angle": 0, + "content": "[8] Rishabh Bhardwaj, Do Duc Anh, and Soujanya Poria. Language models are a hero simpson! safety re-alignment of fine-tuned language models through task arithmetic. In ACL, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[9] Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. Typology of risks of generative text-to-image models. In Conference on AI, Ethics, and Society, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[10] Ruchika Chavhan, Da Li, and Timothy Hospedales. Conceptprune: Concept editing in diffusion models via skilled neuron pruning. arXiv preprint arXiv:2405.19237, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. In ICML, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, et al. Near to mid-term risks and opportunities of open source generative ai. In ICML, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.295, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[15] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Aaron Purewal, Csaba Botos, et al. Risks and opportunities of open-source generative ai. arXiv preprint arXiv:2405.08597, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[16] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In ICCV, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In WACV, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.252, + 0.904, + 0.293 + ], + "angle": 0, + "content": "[18] Vipul Gupta, Santiago Akle Serrano, and Dennis DeCoste. Stochastic weight averaging in parallel: Large-batch training that generalizes well. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[19] Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio Pizzati, Philip Torr, Adel Bibi, Bernard Ghanem, and Mete Ozay. Model merging and safety alignment: One bad model spoils the bunch. In EMNLP Findings, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.904, + 0.394 + ], + "angle": 0, + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.454, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshops, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.514, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.904, + 0.625 + ], + "angle": 0, + "content": "[25] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. In \\(COLM\\), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[26] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In ICLR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[27] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[28] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17032" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[32] Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, and Fabio Pizzati. Latent guard: a safety framework for text-to-image generation. In ECCV, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[33] Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. Videodpo: Omni-preference alignment for video diffusion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8009-8019, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[34] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[35] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A holistic approach to undesired content detection in the real world. In AAAI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[36] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.349, + 0.412 + ], + "angle": 0, + "content": "[37] Mistral. Mistral nemo, 2024. Link. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[38] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.497 + ], + "angle": 0, + "content": "[39] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. NeurIPS, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[40] Seongbeom Park, Suhong Moon, Seunghyun Park, and Jinkyu Kim. Localization and manipulation of immoral visual cues for safe text-to-image generation. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[41] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[43] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Removing nsfw concepts from vision-and-language models for text-to-image retrieval and generation. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[44] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. In Conference on Computer and Communications Security, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[45] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. In NeurIPS workshops, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.775, + 0.106 + ], + "angle": 0, + "content": "[47] Github Repo. Nudenet, 2022. Link. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.108, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[49] Téo Sanchez. Examining the text-to-image community of practice: Why and how do people prompt generative ais? In Conference on Creativity and Cognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[50] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In FAccT, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[51] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[52] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[54] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[55] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In ICLR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[56] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[58] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. In NeurIPS, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[59] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[60] Yuchen Yang, Bo Hui, Haolin Yuan, Neil Gong, and Yinzhi Cao. Sneakyprompt: Jailbreaking text-to-image generative models. In Symposium on Security and Privacy, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[61] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[62] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In CVPR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17033" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[63] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17034" + } + ] +] \ No newline at end of file diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_origin.pdf b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a9f4d8d72b0a5343b48d1d29236b2bf29bbf43e --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/821c3c22-a165-4d20-a058-8dccd9e55058_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54538a2e763359890bb457d05b4b731c6cf4e59dd665039233d7413e719172c7 +size 34087227 diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/full.md b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e82abb4e0f8ae0478c17eef4afcb69e8c6b2506 --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/full.md @@ -0,0 +1,318 @@ +# AlignGuard: Scalable Safety Alignment for Text-to-Image Generation + +Runtao Liu $^{1*}$ I Chieh Chen $^{1*}$ Jindong Gu $^{2}$ Jipeng Zhang $^{1}$ Renjie Pi $^{1}$ + +Qifeng Chen $^{1}$ Philip Torr $^{2}$ Ashkan Khakzar $^{2}$ Fabio Pizzati $^{2,3}$ + +$^{1}$ Hong Kong University of Science and Technology $^{2}$ University of Oxford $^{3}$ MBZUAI + +{rliuay,icchen}@connect.ust.hk + +https://alignguard.github.io/ + +# Abstract + +Text-to-image (T2I) models are widespread, but their limited safety guardrails expose end users to harmful content and potentially allow for model misuse. Current safety measures are typically limited to text-based filtering or concept removal strategies, able to remove just a few concepts from the model's generative capabilities. In this work, we introduce AlignGuard, a method for safety alignment of T2I models. We enable the application of Direct Preference Optimization (DPO) for safety purposes in T2I models by synthetically generating a dataset of harmful and safe image-text pairs, which we call CoProV2. Using a custom DPO strategy and this dataset, we train safety experts, in the form of low-rank adaptation (LoRA) matrices, able to guide the generation process away from specific safety-related concepts. Then, we merge the experts into a single LoRA using a novel merging strategy for optimal scaling performance. This expert-based approach enables scalability, allowing us to remove $7 \times$ more harmful concepts from T2I models compared to baselines. AlignGuard consistently outperforms the state-of-the-art on many benchmarks and establishes new practices for safety alignment in T2I networks. + +Warning: this paper includes potentially offensive content. + +# 1. Introduction + +Text-to-image (T2I) models have advanced rapidly in recent years, becoming key tools in content creation and entertainment, used by both professionals and hobbyists [48, 49]. Such an impressive growth in image synthesis capabilities, in some cases reaching indistinguishable realism from real samples, has generated concerns about the risks of releasing such models [9, 14, 15]. Indeed, T2I may generate unsettling or harmful content, and expose people, particularly minors or those from sensitive cultural backgrounds, to inappropriate outputs. Moreover, malicious users could + +![](images/57319123dad845f0a64483683363f391a1f5b085094ae51688fac22b17f89d24.jpg) +Figure 1. Safety alignment for T2I. T2I models released without safety alignment risk to be misused (top). We propose Align-Guard, a scalable safety alignment framework for T2I models supporting the mass removal of harmful concepts (middle). We allow for scalability by training safety experts focusing on separate categories such as "Hate", "Sexual", "Violence", etc. We then merge the experts with a novel strategy. By doing so, we obtain safety-aligned models, mitigating unsafe content generation (bottom). + +easily generate, among others, violent pictures, or sexually-explicit deepfakes, potentially causing misinformation or harassing specific users. Regardless of these risks, there has been little effort in proposing effective safety alignment techniques, i.e. strategies for limiting harmful outputs, for T2I models. Most of the current approaches are focused on simple input filtering strategies, such as the analysis of the input prompt, or detectors on output images [2]. However, these strategies are easy to circumvent [46], or they can be deactivated if models are openly released. Conversely, large language models (LLMs) are subject to rigorous safety alignment procedures before being deployed [13]. It is our objective to bridge this gap. + +For preventing harmful outputs, one straightforward approach might seem to be training on curated datasets excluding unsafe images. However, this is nontrivial. Large T2I trainings usually use vast web-crawled datasets, such as LAION-5B [53]. Despite filtering efforts, these datasets contain unsafe or illegal content, making full control nearly impossible. More realistically, we could limit the capability of the T2I to generate harmful concepts post-training. Recent works explored methods to delete specific concepts from pre-trained models, with promising results [16, 17, 29, 34]. Yet, these methods limitedly scale. While they can remove a limited number of concepts, deleting hundreds strongly degrades the model's generative capacity [17, 34]. + +To address these challenges, we introduce AlignGuard, a scalable method for safety alignment of T2I models. Our idea is to remove a large number of concepts from T2I models, by exploiting an ensemble of safety experts trained with Direct Preference Optimization (DPO) on safety-oriented synthetic data. By doing so, we are the first to propose a scalable safety alignment method for T2I, allowing for a safe deployment of trained T2I models. Considering the lack of data for DPO-based safety alignment available for image generation, we introduce CoProV2, a dataset encompassing a broad array of unsafe concepts. CoProV2 generation is fully automatic, including paired images with associated prompts: one harmful image with its corresponding textual description, and one safe image and prompt with similar content and structure. This setup enables the application of DPO for aligning T2I models effectively. In practice, we train Low-Rank Adaptation (LoRA) [24] matrices. Each LoRA serves as a safety expert trained to prevent the generation of specific harmful content, i.e. content related to different categories such as "Hate," "Sexual," "Violence," among others. The LoRAs guide the vision representation of the pre-trained T2I model towards safe outputs during the diffusion process. + +After training the safety experts, we use a new algorithm called Co-Merge to merge all safety LoRA experts into one. This mitigates interferences among different experts, allowing for a significant increase in scalability. As shown in Figure 1, through joint expert training and subsequent merging, we can train on images encompassing 723 harmful concepts—approximately seven times more than existing approaches [17]. In summary, AlignGuard enables safety alignment of T2I models at scale, preventing the generation of inappropriate outputs without impacting generative capabilities on safe prompts. Our contributions are: + +1. We propose AlignGuard, the first scalable approach for T2I models specifically targeting safety alignment. +2. We introduce a training method based on expert models, and a novel merging strategy based on weight activation frequency. +3. For the training of AlignGuard, we craft a new dataset + +for safety alignment in T2I, coined CoProV2, that we will release as open source. + +4. We validate AlignGuard on popular T2I models and multiple benchmarks, demonstrating its effectiveness in many setups. + +# 2. Related Work + +Content filtering. Often, closed-source commercial T2I use prompt blacklists, LLM preprocessing, and image analysis for content filtering [1, 3, 4]. In particular, some use LLM-based prompt analysis [35] to detect harmful inputs, even with ad-hoc models [27]. Recently, Latent Guard [32] proposed a latent-based blacklist within text encoders in T2I models. Some works rely instead on the analysis of generated images [2], where an NSFW classifier is applied to generated images. Instead, others [40] use inpainting to mask potentially unsafe content. These approaches can easily be deactivated if the T2I models' weights are available. + +Concept removal in T2I. Some have explored removing the capability of generating concepts in T2I, like SLD [52], using classifier-free guidance [22] to steer generation away from undesirable outputs. Similarly, Li et al. [30] identify interpretable directions that can be used for safety steering. Alternatively, many fine-tune the T2I. While seminal works finetune the entire model [29], some focus on specific components, such as attentions [17, 38, 62], specific neurons [10], or the textual encoder [43]. There has been a recent interest in mass removal of concepts from T2I models for safety purposes [17, 34]. While MACE [34] uses a similar strategy as ours, they make use of segmentation masks, constraining the erasable concepts at spatially-defined ones. Moreover, all available approaches [17, 34] degrade performance with more than 100 erased concepts. + +Model merging. Model merging, i.e. combining multiple models into one model, has gained attention [12, 61]. Weighted averaging methods are commonly used to enhance performance [18, 26, 57], especially in scenarios where only model weights are accessible [28]. Beyond these, advanced techniques have emerged, improving over basic averaging [6, 36, 39]. For example, TIES [58] resolves operator conflicts to improve merging, at the cost of hyperparameter tuning. Merging has been explored for safety in LLMs [8, 19] and for concept removal in MACE [34], but only using expensive optimization procedures. Existing LoRA merging for T2I focus on single subjects [54, 63]. Instead, we propose a cheap and effective merging strategy, focusing on broad safety categories. + +# 3. Preliminaries + +We use preference optimization algorithms such as DPO [56] to perform scalable safety alignment of T2I mod + +els. Here, we revise the fundamental concepts to allow for the interpretation of our method. + +# 3.1. Text-to-image diffusion models + +Diffusion models allow for image generation by iteratively denoising gaussian noise with a network $\epsilon$ for $t\in [0,T]$ iterations [23]. In particular, T2I diffusion models include natural language conditioning, so the desired output image can be described with text. We now briefly introduce the training procedure for a T2I model. Let us assume an input pair $(x,p)\sim \mathcal{D}_{\mathrm{train}}$ sampled from a training dataset $\mathcal{D}_{\mathrm{train}}$ of images $x$ and paired textual description $p$ . The network $\epsilon$ is trained by estimating the ground truth noise $\tilde{\epsilon}_t$ injected on $x$ for a random $t\in [0,T]$ . We define the input image with the addition of noise as $x_{t}$ . Hence, the training loss is + +$$ +\mathcal {L} _ {\text {d i f f}} (\epsilon , x, p) = | | \tilde {\epsilon} _ {t} - \epsilon \left(x _ {t}, p\right) | |. \tag {1} +$$ + +In Eq. (1), $\epsilon (\cdot)$ is the denoising operation. The network weights $\theta^{*}$ are optimized by minimizing the following: + +$$ +\theta^ {*} = \underset {\theta} {\arg \min } \mathbb {E} _ {(x, p) \sim \mathcal {D} _ {\text {t r a i n}}} \left(\mathcal {L} _ {\text {d i f f}} \left(\epsilon_ {\theta}, x, p\right)\right), \tag {2} +$$ + +where $\epsilon_{\theta}$ refers to the denoising network with a set of weights $\theta$ . During inference, sampled Gaussian noise is iteratively processed for $t \in [0, T]$ with $\epsilon_{\theta^*}$ following specific scheduling policies, ultimately allowing image synthesis. We refer to [23] for additional details. + +# 3.2. DPO for diffusion models + +DPO is a technique for preference alignment initially developed for LLMs [41, 45] and recently extended to diffusion models [33, 56]. The core idea is to benefit from pairwise preferences obtained by labeling. Let us assume a dataset $\mathcal{D}_{\mathrm{DPO}}$ of paired images and textual descriptions $(x^{+}, x^{-}, p)$ . For a given description $p$ , $x^{+}$ is a image that humans indicated as preferred output with respect to $x^{-}$ . The intuition of DPO is to increase the likelihood to generate the preferred output $x^{+}$ , while discouraging the generation of $x^{-}$ . This translates into the following loss: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {D P O}} \left(\epsilon , x ^ {+}, x ^ {-}, p\right) = - \log \sigma (- \beta ( \\ \left(\mathcal {L} _ {\text {d i f f}} \left(\epsilon , x ^ {+}, p\right) - \mathcal {L} _ {\text {d i f f}} \left(\epsilon_ {\text {r e f}}, x ^ {+}, p\right)\right) \tag {3} \\ - \left(\mathcal {L} _ {\text {d i f f}} \left(\epsilon , x ^ {-}, p\right) - \mathcal {L} _ {\text {d i f f}} \left(\epsilon_ {\text {r e f}}, x ^ {-}, p\right)\right)\left. \right), \\ \end{array} +$$ + +where $\epsilon_{\mathrm{ref}}$ is a reference pre-trained network, typically resulting from a previous optimization of Eq. 2. Also, $\sigma$ is the sigmoid operation and $\beta$ is a weighting constant [56]. Finally, one could optimize as: + +$$ +\theta_ {\mathrm {D P O}} ^ {*} = \underset {\theta} {\arg \min } \mathbb {E} _ {(x ^ {+}, x ^ {-}, p) \sim \mathcal {D} _ {\mathrm {D P O}}} \mathcal {L} _ {\mathrm {D P O}} \left(\epsilon_ {\theta}, x ^ {+}, x ^ {-}, p\right). \tag {4} +$$ + +For further details, we refer to the original paper [56]. + +![](images/5fdb24e42ff8224b830deaba3abed6b578568830b836ac19886aeb974c6385d9.jpg) +Figure 2. Dataset generation. For each unsafe concept in different categories, we generate an corresponding prompts with an LLM. We generated paired safe prompts using an LLM, minimizing semantic differences. Then, we use the T2I we intend to align to generate corresponding images for both prompts. + +# 4. Method + +We aim to align T2I models inspired by practices in language model alignment [13]. We avoid expensive human annotations by generating safety-oriented preference data in Section 4.1. We also propose a training procedure based on expert networks (Section 4.2) and subsequent merging (Section 4.3). After training, AlignGuard requires no additional computation at inference (more details in supplement). + +# 4.1. DPO for safety alignment in T2I + +We aim to use DPO as a strategy to perform safety alignment on pretrained T2I models. Assuming an unsafe textual input describing a sexually explicit, shocking, or violent scene, we would like the diffusion model to ignore the unsafe requested traits, and generate a safe image, i.e. not including any visually disturbing trait. We achieve this by generating automatically {unsafe, safe} image pairs for DPO training, obtained by propting the T2I with LLM-manipulated text. In a nutshell, we can discourage the generation of unsafe images, having as preferred DPO image a visually-close safe sample, for an unsafe input T2I prompt. + +We start by generating a set of unsafe concepts $\mathcal{C}$ with an LLM. Each $c\in \mathcal{C}$ is a keyword representing unsafe elements that may be included in an image, such as "homicide", "nude", etc. Then, we follow the prompt synthesis procedure of [32], generating both unsafe and safe prompts for image generators from concepts, enforcing minimal differences. Specifically, we sample an unsafe prompt $p^{\mathrm{U}}$ with an LLM to create text visually describing a scene using the input concept $c$ . Then, we further process the generated $p^{\mathrm{U}}$ , prompting an LLM to remove any reference to the input unsafe concept $c$ . This transforms the prompt into a safe $p^{\mathrm{S}}$ , while minimizing semantic changes. Differently from [32], we process both prompts with a pretrained + +![](images/6c3051da9780dc1e8aa531b8cf54a27aea7d983ce297ae0892e469b75baa53b5.jpg) +Figure 3. Expert training and merging. First, we use the previously generated prompts and images to train LoRA experts on specific safety categories (left), exploiting our DPO-based losses. Then, we merge all the safety experts with Co-Merge (right). This allows to achieve general safety experts that produce safe outputs for a generic unsafe input prompt in any category. + +![](images/3bc2b8c155eeeec4e24869edca388100096047370f005f4e1587af59a980998b.jpg) + +![](images/b771cc2f950c55918ffe042517d06b9196474e654b34b1a768f3d99b12d54e30.jpg) +Figure 4. Merging experts with Co-Merge. (left) Assuming LoRA experts with the same architecture, we analyze which expert has the highest activation for each weights for all inputs. (right) Then, we obtained the merged weights from multiple experts, merging only the most active weights per expert. + +diffusion model, obtaining images $\{x^{\mathrm{U}}, x^{\mathrm{S}}\}$ , derived from $p^{\mathrm{U}}$ and $p^{\mathrm{S}}$ , respectively. We show this in Figure 2. + +In this way, we can automatically create image preference pairs, assuming that the preferred output in safety alignment would always be a similar safe image for an unsafe prompt. Indeed, by repeating the process multiple times, we can construct a dataset $\mathcal{D}_{\mathrm{safety}}$ consisting of tuples including the sampled prompts and corresponding generated images, i.e. $(x^{\mathrm{S}}, x^{\mathrm{U}}, p^{\mathrm{S}}, p^{\mathrm{U}})$ . Then, we use DPO to discourage the generation of unsafe outputs for an unsafe prompt $p^{\mathrm{U}}$ , and promote the generation of the paired safe image $x^{\mathrm{S}}$ instead. To achieve this, we set $x^{\mathrm{S}}$ as preferred output for $p^{\mathrm{U}}$ , discouraging the generation of $x^{\mathrm{U}}$ . The semantic similarity of $p^{\mathrm{U}}$ and $p^{\mathrm{S}}$ will encourage the DPO training to focus exclusively on the visual traits making $x^{\mathrm{U}}$ unsafe. We define a first loss term: + +$$ +\mathcal {L} _ {\text {a l i g n}} = \mathcal {L} _ {\mathrm {D P O}} \left(\epsilon_ {\theta}, x ^ {S}, x ^ {U}, p ^ {U}\right). \tag {5} +$$ + +While effective in removing the unsafe concepts $c$ from the generated images, this risks to reduce the generative capa + +bilities of the network, since we are encouraging to generate images that do not respect the input prompt. To prevent this, we also use $x^{\mathrm{S}}$ as preferred generation for the input prompt $p^{\mathrm{S}}$ , while penalizing on $x^{\mathrm{U}}$ . This writes: + +$$ +\mathcal {L} _ {\mathrm {c o n}} = \mathcal {L} _ {\mathrm {D P O}} \left(\epsilon_ {\theta}, x ^ {S}, x ^ {U}, p ^ {S}\right). \tag {6} +$$ + +Finally, we modify Eq. (4) proposing our training objective: + +$$ +\theta_ {\text {A l i g n G u a r d}} ^ {*} = \arg \min _ {\theta} \mathbb {E} _ {(x ^ {\mathrm {s}}, x ^ {\mathrm {U}}, p ^ {\mathrm {s}}, p ^ {\mathrm {U}}) \sim \mathcal {D} _ {\text {s a f e t y}}} \mathcal {L} _ {\text {a l i g n}} + \mathcal {L} _ {\text {c o n}}. \tag {7} +$$ + +This objective allows for a stable safety-oriented finetuning of the T2I model for a large number of concepts $c \in \mathcal{C}$ without impacting the generative capabilities of the T2I model. + +# 4.2. Improving scaling with safety experts + +Training with our proposed strategy in Section 4.1 enables us to scale the number of concepts used for safety alignment, while keeping generative capabilities intact. However, by increasing the number of concepts in $\mathcal{C}$ , we noticed a decrease in performance. We suppose this is due to the complexity of the task: in safety alignment there is a need of a high contextual understanding of the generated scenes. As an example, an image of a kitchen including knives is a safe output, while a scene of harassment with the usage of a knife is an undesired unsafe outcome of the image synthesis process, that we aim to prevent. As a solution, we propose to decompose the safety alignment on separate categories, such as "violence", "sexual content", "harassment", and others. Our intuition is that by limiting the variability of the generated images, it would be easier to identify visual patterns that would make a scene "violent", or "sexually explicit". Hence, we aim to train safety experts, each focusing on one category only. + +Algorithm 1 Co-Merge +1: Input: Safety expert LoRAs $\{L_1, L_2, \ldots, L_N\}$ , unsafe prompts $\{p_1, p_2, \ldots, p_K\}$ +2: Output: Merged LoRA $L_{\text{merged}}$ +3: Init a count matrix $C$ of $J \times N$ elements with zeros. +4: for each prompt $p_k$ , $k \in [1, K]$ do +5: for each neuron index $j$ , $j \in [1, J]$ do +6: $i_{\text{max}} := \arg \max_i |L_i^j(p_k)|$ , $i \in [1, N]$ +7: Increment count $C[j, i_{\text{max}}]$ +8: end for +9: end for +10: for each neuron index $j$ , $j \in [1, J]$ do +11: $m := \arg \max_i C[j, i]$ , $i \in [1, N]$ +12: $L_{\text{merged}}^j := L_m^j$ +13: end for +14: Return $L_{\text{merged}}$ + +In practice, we identified $N$ safety-related concept categories, each linked to a generated concept set $\{\mathcal{C}^1,\dots,\mathcal{C}^N\}$ . Each set is created by prompting a large language model to generate concepts within a given unsafe category. We then generated the corresponding safety datasets $\{\mathcal{D}_{\mathrm{safety}}^1,\dots,\mathcal{D}_{\mathrm{safety}}^N\}$ following the approach in Section 4.1. Rather than fine-tuning the entire model's weights, $\theta$ , to create the expert networks, we used low-rank adaptation (LoRA) [24], updating only a minimal subset of weights and as such saving significant computational costs. Ultimately, we apply Equation 7 for $\{\mathcal{D}_{\mathrm{safety}}^1,\dots,\mathcal{D}_{\mathrm{safety}}^N\}$ , obtaining a set of LoRAs $\{L^1,\dots,L^N\}$ , as shown in Figure 3. At inference time, the obtained LoRAs can be applied to the pretrained T2I model to enable safe image generation. + +# 4.3. Experts merging + +To prevent the generation of content across multiple safety-related categories, additional considerations are necessary. Indeed, loading multiple LoRAs into a model can lead to interference [25], reducing overall performance in safety-oriented image generation. Alternatively, multiple experts could be run independently, at the cost of multiplying inference costs. To mitigate these issues, we aim to merge the trained LoRAs into a single one, as visualized in Figure 3, without sacrificing performance. + +We propose a novel data-based strategy, called CoMerge, to merge multiple safety experts, specifically designed around our use case. Our intuition is that by decomposing safety alignment in multiple categories (Section 4.2), trained LoRAs sharing the same architecture would encode information about category-specific alignment in different neurons. In other words, we assume that in the T2I model, the weights responsible for generating violent content (i.e. including, for instance, blood) will be different from those generating sexually-explicit content + +
DatasetImage # of Prompts # of Categories # of ConceptsIP
COCO [31]✓3,000N/AN/A0.06
I2P [51]✗4,7037N/A0.36
UD [44]✗9325N/A0.47
CoPro [32]✗56,52677230.23
CoProV2 (ours)✓23,69077230.43
+ +Table 1. Datasets comparison. Our LLM-generated dataset, CoProV2, achieves comparable Inappropriate Probability (IP) to human-crafted datasets (UD [44], I2P [51]) and offers similar scale to CoPro [32]. COCO [31], exhibiting a low IP, is used as a benchmark for image generation with safe prompts as input. + +(e.g. naked skin). LoRAs work as updates for the model weights [24], so, for concepts associated to different neurons, they will encode related information in different parts of the same architecture. Then, our idea is to isolate the neurons inside the trained safety experts that activate the most for unsafe input prompts, and construct a merged expert as a LoRA $L_{\text{merged}}$ including the parameters of different experts associated to the most significant activations only. + +We start by randomly sampling $K$ unsafe prompts from the training set of $\{\mathcal{D}_{\mathrm{safety}}^1,\dots,\mathcal{D}_{\mathrm{safety}}^N\}$ , equally distributed across each category. We then process all prompts with the T2I model, where we load all trained experts $\{L_1,\dots,L_N\}$ one at the time. Each expert LoRA is a multi-layer perceptron including $J$ neurons. For each prompt $p_k^U$ , where $k\in [1,K]$ , and each expert model $L_{i}$ , where $i\in [1,N]$ , we record the absolute value of the activation from each neuron $L_{i}^{j}$ , where $j\in [1,J]$ . We denoted the absolute value as $|L_i^j (p_k^U)|$ . This provides a measure of the response for each neuron to that prompt. These prompts are uniformly sampled from distinct safety categories, so different neurons of the same index $j$ across the experts will exhibit different activations in response to different prompts. To create the merged $L_{\mathrm{merged}}$ , we identify the neurons with the highest activation frequencies across experts. Specifically, for each neuron $j$ in the merged expert $L_{\mathrm{merged}}$ , we select the neuron from the original set of experts $\{L_1,\dots,L_N\}$ that has the highest activation frequency count across the $K$ prompts. We summarize this in Algorithm 1, where the $p^U$ superscript is omitted for clarity. Figure 4 illustrates Co-Merge with $K = 10$ , describing how the frequency count leads to the network weight merging. Note that merging $L_{\mathrm{merged}}$ with the original T2I weights makes safety alignment difficulty reversible, a beneficial practice for T2I releases. + +# 5. Experiments + +We first introduce our setup (Section 5.1), and compare with baselines in Section 5.2. We then analyze the properties of AlignGuard (Section 5.3) and show ablations (Section 5.4). + +# 5.1. Experimental setup + +Baselines. We use AlignGuard to align three models for T2I, namely Stable Diffusion (SD) v1.5/v2.1 [48] and + +
MethodIP↓FID ↓ COCOCLIP ↑
CoProV2I2PUD
SD1.5No alignment0.510.360.5269.7733.52
SLD [51]0.270.190.3071.4532.24
ESD-u [16]0.220.250.2172.9829.61
UCE [17]0.330.300.3872.0132.01
AlignGuard0.070.110.1670.9632.32
SD2.1No alignment0.510.350.5570.4534.99
AlignGuard0.120.120.1771.9934.13
SDXLNo alignment0.490.310.4768.9035.60
AlignGuard0.090.080.1175.2034.67
+ +Table 2. Benchmarks. AlignGuard achieves best performance both in generated images alignment (IP) and image quality (FID, CLIPScore), with two T2I models and against 3 methods for SD v1.5. Note that we use CoProV2 only for training, hence I2P and UD are out-of-distribution. Yet, AlignGuard allows a robust safety alignment. Best is bold, second is underlined. + +SDXL [42]. For all, we compare the original versions released on HuggingFace, and our finetuned version with AlignGuard. Moreover, we compare with recent methods based on SD v1.5 for safe image generation (SLD [51]) and concept erasure (ESD [16], UCE [17]). In particular, for ESD, we follow the paper indications and use ESD-u [16], a version of ESD for broad concept removal. We perform our analysis and ablations on SD v1.5. + +Metrics. To evaluate the effectiveness of safety alignment, we use the inappropriate probability metric (IP) from SLD [51], measuring the ratio of unsafe contents generated by T2I with unsafe prompts. The unsafe content detection is performed by using classification results of Q16 [50] and NudeNet [47]. For evaluating image quality, we use FID [21], and CLIPScore [20] for text-image alignment. + +Datasets. For training of AlignGuard, we generate a new dataset following Section 4.1, named CoProV2 (Concepts and Prompts). We motivate this choice by observing a limitation in the CoPro dataset [32], i.e. the limited IP measured on images generated with the prompts of the dataset. This suggests that many prompts in CoPro are not leading to unsafe generations. This is due to the limited performance of the LLM building CoPro. We start from the same $\mathcal{C}$ as CoPro: 723 harmful concepts across 7 categories (Hate, Harassment, Violence, Self-Harm, Sexual, Shocking, Illegal). We re-generate the unsafe/safe prompts pairs with a new LLM prompt reported in the supplement. We use Mistral-Nemo-Instruct [37] for the prompt synthesis. We generate 23,690 pairs of safe/unsafe prompts, and we use 15,690/8,000 pairs for training/ testing. For testing only, we use I2P [7] and Unsafe Diffusion (UD) [44], including 4703/932 human-designed unsafe prompts on 7/5 categories, respectively. Following common practices [16, 48, 51] we use COCO [31], with 3000 safe + +captions and corresponding images, for evaluating the performance of the T2I model in image generation after alignment. We compare the datasets quantitatively in Table 1. + +Implementation details. We train AlignGuard with batch-size 8, accumulating over 16 steps. We use AdamW with a learning rate of $10^{-5}$ for training each LoRA with rank 4. For Co-Merge, we set $K = 100$ unsafe prompts. Each LoRA takes $\sim 2$ hours to train for SD and $\sim 18$ hours for SDXL on a single Nvidia 5880 Ada GPU for 2000 steps. We normalize the computation needed for training AlignGuard and baselines. More details are in the supplement. + +# 5.2. Benchmarks + +Quantitative evaluation. In Table 2, we compare with alignment baselines for IP, FID, and CLIPScore. For fairness, all baselines were re-trained on CoProV2 training set and evaluated using the CoProV2 test set, I2P, and UD. Note that I2P and UD serve as out-of-distribution evaluations as they are crafted by human users in real deployment scenarios, differing in both prompt format and content from CoProV2. Initially, we trained the baselines using the complete $\mathcal{C}$ set of CoProV2 with Stable Diffusion v1.5 as the backbone. This involved using $\mathcal{C}$ to identify negative embeddings in SLD and excluding all $\mathcal{C}$ concepts for ESD-u and UCE. All baselines collapsed due to the high number of concepts in CoProV2, aligning with the limited scaling performance noted in previous works [16, 17]. These results are included in the supplement. For successful training comparisons, we used the 7 category names of CoProV2 as removable concepts for the baselines, while we use the full prompts and images in CoProV2 for AlignGuard. We trained our expert LoRAs on samples generated with all $\mathcal{C}$ , one per category, and merge them with Co-Merge. As shown in the table, we outperform considerably all baselines in every metric. Notably, our IP for SD v1.5 on CoProV2 (0.07) is much lower than the next best (ESD-u, 0.22). Additionally, AlignGuard demonstrates superior generalization, achieving better IP scores on unseen I2P (0.08) and UD (0.16). Our DPO-based strategy effectively maintains Stable Diffusion's generation capabilities preventing forgetting of concepts unrelated to safety. Indeed we report nearing the performance of the Baseline method without safety alignment in FID (70.96 vs 69.77) and CLIPScore (32.32 vs 33.52) on COCO captions. We also applied AlignGuard to SD v2.1 and SDXL to test the generalization of our method to a different diffusion model. Our results are consistent with different backbones like SD v2.1 and SDXL, achieving performance comparable to our aligned SD v1.5 with IP 0.09 and 0.12 on CoProV2. This suggests that AlignGuard can be used for the safety alignment of arbitrary T2I models. + +Qualitative evaluation. We show in Figure 5 a selection of samples advocating the effectiveness of AlignGuard. We + +![](images/21c1d789409b3ec5f0c56b1782d9d048cef3fc9b545158863582e81dabbb3421.jpg) +Figure 5. Qualitative comparison. Compared to non-aligned baseline models, AlignGuard allows to synthesize safe images for unsafe input prompts. Please note the layout similarity between the unsafe and safe outputs: thanks to our training, only the harmful image traits are removed from the generated images. Concepts in ⟨brackets⟩. Prompts are shortened, for full ones see the supplement. + +sample prompts from CoProV2. For SD v1.5 and SDXL (SD v2.1 in the supplement), we prevent the generation of unsafe content across 7 categories. Please also note how the quality of the generated images, in both models aligned with AlignGuard, is not impacted by our alignment procedure. Also, let us highlight that although the model refuses to follow unsafe prompts, the generated images preserve similar traits and global layout, e.g. the pose of the man in the "Shocking" column for SDXL. This property results from training on safe/unsafe pairs. + +# 5.3. Properties + +Experts performance. For AlignGuard, we train an expert for each category in CoProV2. We now examine the effectiveness of each expert LoRA across all categories in the CoProV2 test set in Table 3. The results show that training and evaluating on the same category generally leads to good performance, as expected. Interestingly though, applying one expert LoRA consistently improves IP across all categories and often surpasses the expert trained specifically for that category. We attribute this to the interaction of multiple concepts across categories. While the concepts differ, the visual features of generated + +images may share similarities. For example, scenes in the "Violence" or "Self-harm" categories often depict blood, while the "Hate" category may include sexual slurs, which could explain the strong performance on "Sexual" data. Ultimately, this suggests that for safety alignment tasks, AlignGuard is robust to the concept selection. However, experts trained on different categories still encode distinct concepts. This becomes evident when merging all LoRAs with Co-Merge (All - Ours), since we achieve the best performance across all categories. We also include a baseline where a single LoRA is trained on prompts from all categories (All - single), which shows suboptimal performance. Here, the lack of specialized expert models results in less effective training, justifying our design. + +Adversarial attacks robustness. To understand if harmful concepts are effectively removed from the T2I model capabilities, we perform an additional experiment based on text-based adversarial attacks. Those allow to optimize text apparently innocuous, but leading to unsafe generation. If AlignGuard is able to prevent harmful generations in this setup, it means that our alignment is effective even if harmful concepts are not explicitly present in the input prompt. + +
ExpertIP on specific category ↓Avg.
HateHarass.ViolenceSelf-harmSexualShockingIllegal
No alignment0.490.480.540.590.540.520.440.51
Hate0.110.150.190.160.110.170.110.14
Harass.0.170.160.200.230.190.160.150.18
Violence0.180.150.160.190.180.170.160.17
Self-Harm0.210.240.260.250.220.240.230.23
Sexual0.180.190.250.250.160.210.190.20
Shocking0.170.180.220.200.150.190.160.18
Illegal0.170.140.190.200.200.210.120.18
All - Single0.170.210.220.220.160.210.180.16
All - Ours0.060.060.090.070.070.080.040.07
+ +Table 3. Effectiveness of merging. While training a single safety expert across all data (All - single), IP performance are lower or comparable to single experts (previous rows). Instead, by merging safety experts (All - ours) we considerably improve results. + +
MethodMMA[59]Ring-A-Bell[55]SneakyPrompt[60]P4D[11]
No align.0.430.670.570.48
ESD-u [16]0.220.430.210.23
AlignGuard0.080.120.090.07
+ +We use 4 state-of-the-art attack methods listed in Table 4 to optimize seemingly innocuous prompts, starting from Co-ProV2 test prompts. We test alignment with AlignGuard and ESD-u [16]. The results advocate for AlignGuard's high robustness to adversarial attacks, proof of the effectiveness of alignment. In particular, for Ring-A-Bell [55], AlignGuard maintains an IP of 0.12, even for a particularly successful attack able to raise the IP of the baseline SD v1.5 to 0.67. + +# 5.4. Ablation studies + +DPO strategy. In AlignGuard training, several options exist for selecting preferred samples for DPO. Indeed, by selecting a paired safe image as a safe output for an unsafe prompt, we depart slightly from common practices in LLM alignment, where the preferred answer in presence of an unsafe input is a refusal [13]. We here investigate this choice by mimicking refusals supervision in language models. To do so, we create two different alternative images that we use as positive samples for DPO in presence of an unsafe input: (1) a Black Image, and (2) an image constructed by applying a Warning Sign on top of an unsafe output. Examples are in the supplement. From results in Table 5a, we evince that our strategy exploiting a paired safe image is perform- + +Table 4. Resistance to adversarial attacks. We evaluate with 4 adversarial attacks methods the performance of AlignGuard and the best baseline, ESD-u, in terms of IP. For a wide range of attacks, we are able to outperform the baselines, advocating for the effectiveness of our scalable concept removal strategy. + +
DPO strategyIP ↓FID ↓CLIP ↑
Black image0.2185.3831.22
Warning sign0.1670.0133.05
w/o Lcon0.1674.8230.78
Paired safe image (ours)0.0770.9632.32
+ +(a) DPO strategy ablation + +
MergingIP ↓FID ↓CLIP ↑DataIP ↓FID ↓CLIP ↑
Git Re-Basin [5]0.3069.9833.4010%0.1972.8727.30
Task Vectors [26]0.25179.604.8025%0.1671.9127.34
Model Soups [57]0.1469.8333.4050%0.1772.7627.20
TIES [58]0.0979.7928.85100%0.0770.9632.32
Co-Merge0.0770.9632.32
+ +(b) Merging method ablation + +(c) Scaling performance. + +Table 5. Ablation studies. We check the effects of alternative strategies for DPO, proving that our approach is best (a). Co-Merge is also the best merging strategy compared to baselines (b). Finally, we verify that scaling data improve our performance (c). + +ing significantly better, justifying the different design of the DPO positive between T2I and LLM alignment. Also, we compared in the same table with a setup excluding $\mathcal{L}_{con}$ , i.e. training without considering generative capabilities preservation among safe samples. This results in lower FID and higher CLIP scores, highlighting the importance of $\mathcal{L}_{con}$ to preserve generative capabilities in presence of safe inputs. + +Merging strategy. We evaluate our merging method against four baselines: (1) Git Re-Basin, (2) Task Vectors, (3) Model Soups [57] and (4) TIES [58]. We report results in Table 5b, showing that Co-Merge consistently outperform these baselines. Let us highlight that baselines do not use data for performing merging, leading to suboptimal results and at the cost of tedious hyperparameter tuning. By performing a data-aware merging using unsafe prompts, instead, we are able to optimally balance the contributions of each expert with minimal effort and computational load. + +Scaling performance. We evaluate the importance of the scale of data in Table 5c. We subsample CoProV2 with different percentages and retrain SD v1.5 with AlignGuard. For datasets increasingly smaller, we observe an expected decrease in performance. With $100\%$ of the data, we perform best, proving that AlignGuard benefits the most from large datasets and paving the way for larger trainings. + +# 6. Conclusion + +We introduced AlignGuard, the first approach dedicated to safety alignment for T2I models at scale. By achieving safety alignment on a broad range of concepts, we allow for safe generations in the presence of harmful inputs. AlignGuard enables text-to-image models to generate outputs that are not only of high quality but also aligned with safety and ethical guidelines. We hope our work will inspire future practices in T2I deployment. + +# Acknowledgements + +This research was supported by the Research Grant Council of the Hong Kong Special Administrative Region under grant number 16212623. FP is funded by KAUST (Grant DFR07910). AK, JG, and PT are supported by UKRI grant: Turing AI Fellowship EP/W002981/1, and by the Royal Academy of Engineering under the Research Chair and Senior Research Fellowships scheme. The authors thank Alexander Pondaven and David Romero for their kind proofreading. + +# References + +[1] The complete list of banned words in midjourney you need to know, 2022. Link. 2 +[2] Diffusers: State-of-the-art diffusion models, 2022. Link. 1, 2 +[3] Dall-e 3 system card, 2023. Link. 2 +[4] Leonardo ai content moderation filter: Everything you need to know, 2023. Link. 2 +[5] Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 8 +[6] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2 +[7] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. TMLR, 2024. 6 +[8] Rishabh Bhardwaj, Do Duc Anh, and Soujanya Poria. Language models are a hero simpson! safety re-alignment of fine-tuned language models through task arithmetic. In ACL, 2024. 2 +[9] Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. Typology of risks of generative text-to-image models. In Conference on AI, Ethics, and Society, 2023. 1 +[10] Ruchika Chavhan, Da Li, and Timothy Hospedales. Conceptprune: Concept editing in diffusion models via skilled neuron pruning. arXiv preprint arXiv:2405.19237, 2024. 2 +[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. In ICML, 2024. 8 +[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3, 8 +[14] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, et al. Near to mid-term risks and opportunities of open source generative ai. In ICML, 2024. 1 + +[15] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Aaron Purewal, Csaba Botos, et al. Risks and opportunities of open-source generative ai. arXiv preprint arXiv:2405.08597, 2024. 1 +[16] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In ICCV, 2023. 2, 6, 8 +[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In WACV, 2024. 2, 6 +[18] Vipul Gupta, Santiago Akle Serrano, and Dennis DeCoste. Stochastic weight averaging in parallel: Large-batch training that generalizes well. In ICLR, 2020. 2 +[19] Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio Pizzati, Philip Torr, Adel Bibi, Bernard Ghanem, and Mete Ozay. Model merging and safety alignment: One bad model spoils the bunch. In EMNLP Findings, 2024. 2 +[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 6 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6 +[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshops, 2022. 2 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 5 +[25] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. In $COLM$ , 2024. 5 +[26] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In ICLR, 2022. 2, 8 +[27] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 2 +[28] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In ICLR, 2023. 2 +[29] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2 +[30] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In CVPR, 2024. 2 + +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6 +[32] Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, and Fabio Pizzati. Latent guard: a safety framework for text-to-image generation. In ECCV, 2024. 2, 3, 5, 6 +[33] Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. Videodpo: Omni-preference alignment for video diffusion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8009-8019, 2025. 3 +[34] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In CVPR, 2024. 2 +[35] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A holistic approach to undesired content detection in the real world. In AAAI, 2023. 2 +[36] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 2022. 2 +[37] Mistral. Mistral nemo, 2024. Link. 6 +[38] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In ICCV, 2023. 2 +[39] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. NeurIPS, 36, 2024. 2 +[40] Seongbeom Park, Suhong Moon, Seunghyun Park, and Jinkyu Kim. Localization and manipulation of immoral visual cues for safe text-to-image generation. In WACV, 2024. 2 +[41] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2024. 3 +[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6 +[43] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Removing nsfw concepts from vision-and-language models for text-to-image retrieval and generation. In ECCV, 2024. 2 +[44] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. In Conference on Computer and Communications Security, 2023. 5, 6 +[45] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3 +[46] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. In NeurIPS workshops, 2022. 1 + +[47] Github Repo. Nudenet, 2022. Link. 6 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 5, 6 +[49] Téo Sanchez. Examining the text-to-image community of practice: Why and how do people prompt generative ais? In Conference on Creativity and Cognition, 2023. 1 +[50] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In FAccT, 2022. 6 +[51] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 5, 6 +[52] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 2 +[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[54] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV. Springer, 2025. 2 +[55] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In ICLR, 2024. 8 +[56] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 2, 3 +[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, 2022. 2, 8 +[58] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. In NeurIPS, 2024. 2, 8 +[59] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. In CVPR, 2024. 8 +[60] Yuchen Yang, Bo Hui, Haolin Yuan, Neil Gong, and Yinzhi Cao. Sneakyprompt: Jailbreaking text-to-image generative models. In Symposium on Security and Privacy, 2024. 8 +[61] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In ICML, 2024. 2 +[62] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In CVPR, 2024. 2 + +[63] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2 \ No newline at end of file diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/images.zip b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8cb3c5f7652434dced29079b9ae2b85f7c908d1 --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d259d3ba27863bdc7aa66284fe17e8fde7e08052aa6457e8b97295b481615992 +size 677851 diff --git a/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/layout.json b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..512da12aba6ab4e84b74186ea9a59c8bfaab9a56 --- /dev/null +++ b/2025/AlignGuard_ Scalable Safety Alignment for Text-to-Image Generation/layout.json @@ -0,0 +1,9143 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 522, + 121 + ], + "type": "text", + "content": "AlignGuard: Scalable Safety Alignment for Text-to-Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "content": "Runtao Liu" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "content": " I Chieh Chen" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "content": " Jindong Gu" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "content": " Jipeng Zhang" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "text", + "content": " Renjie Pi" + }, + { + "bbox": [ + 120, + 142, + 489, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "text", + "content": "Qifeng Chen" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "text", + "content": " Philip Torr" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "text", + "content": " Ashkan Khakzar" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "text", + "content": " Fabio Pizzati" + }, + { + "bbox": [ + 140, + 157, + 469, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "spans": [ + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "text", + "content": "University of Oxford " + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 171, + 523, + 186 + ], + "type": "text", + "content": "MBZUAI" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 225, + 187, + 386, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 187, + 386, + 198 + ], + "spans": [ + { + "bbox": [ + 225, + 187, + 386, + 198 + ], + "type": "text", + "content": "{rliuay,icchen}@connect.ust.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 198, + 200, + 408, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 200, + 408, + 213 + ], + "spans": [ + { + "bbox": [ + 198, + 200, + 408, + 213 + ], + "type": "text", + "content": "https://alignguard.github.io/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 266, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 518 + ], + "type": "text", + "content": "Text-to-image (T2I) models are widespread, but their limited safety guardrails expose end users to harmful content and potentially allow for model misuse. Current safety measures are typically limited to text-based filtering or concept removal strategies, able to remove just a few concepts from the model's generative capabilities. In this work, we introduce AlignGuard, a method for safety alignment of T2I models. We enable the application of Direct Preference Optimization (DPO) for safety purposes in T2I models by synthetically generating a dataset of harmful and safe image-text pairs, which we call CoProV2. Using a custom DPO strategy and this dataset, we train safety experts, in the form of low-rank adaptation (LoRA) matrices, able to guide the generation process away from specific safety-related concepts. Then, we merge the experts into a single LoRA using a novel merging strategy for optimal scaling performance. This expert-based approach enables scalability, allowing us to remove " + }, + { + "bbox": [ + 55, + 266, + 296, + 518 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 55, + 266, + 296, + 518 + ], + "type": "text", + "content": " more harmful concepts from T2I models compared to baselines. AlignGuard consistently outperforms the state-of-the-art on many benchmarks and establishes new practices for safety alignment in T2I networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 520, + 294, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 520, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 58, + 520, + 294, + 533 + ], + "type": "text", + "content": "Warning: this paper includes potentially offensive content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 135, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 694 + ], + "type": "text", + "content": "Text-to-image (T2I) models have advanced rapidly in recent years, becoming key tools in content creation and entertainment, used by both professionals and hobbyists [48, 49]. Such an impressive growth in image synthesis capabilities, in some cases reaching indistinguishable realism from real samples, has generated concerns about the risks of releasing such models [9, 14, 15]. Indeed, T2I may generate unsettling or harmful content, and expose people, particularly minors or those from sensitive cultural backgrounds, to inappropriate outputs. Moreover, malicious users could" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 239, + 555, + 432 + ], + "blocks": [ + { + "bbox": [ + 318, + 239, + 555, + 432 + ], + "lines": [ + { + "bbox": [ + 318, + 239, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 318, + 239, + 555, + 432 + ], + "type": "image", + "image_path": "57319123dad845f0a64483683363f391a1f5b085094ae51688fac22b17f89d24.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 530 + ], + "type": "text", + "content": "Figure 1. Safety alignment for T2I. T2I models released without safety alignment risk to be misused (top). We propose Align-Guard, a scalable safety alignment framework for T2I models supporting the mass removal of harmful concepts (middle). We allow for scalability by training safety experts focusing on separate categories such as \"Hate\", \"Sexual\", \"Violence\", etc. We then merge the experts with a novel strategy. By doing so, we obtain safety-aligned models, mitigating unsafe content generation (bottom)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "easily generate, among others, violent pictures, or sexually-explicit deepfakes, potentially causing misinformation or harassing specific users. Regardless of these risks, there has been little effort in proposing effective safety alignment techniques, i.e. strategies for limiting harmful outputs, for T2I models. Most of the current approaches are focused on simple input filtering strategies, such as the analysis of the input prompt, or detectors on output images [2]. However, these strategies are easy to circumvent [46], or they can be deactivated if models are openly released. Conversely, large language models (LLMs) are subject to rigorous safety alignment procedures before being deployed [13]. It is our objective to bridge this gap." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 138, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 138, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 138, + 712 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17024" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "For preventing harmful outputs, one straightforward approach might seem to be training on curated datasets excluding unsafe images. However, this is nontrivial. Large T2I trainings usually use vast web-crawled datasets, such as LAION-5B [53]. Despite filtering efforts, these datasets contain unsafe or illegal content, making full control nearly impossible. More realistically, we could limit the capability of the T2I to generate harmful concepts post-training. Recent works explored methods to delete specific concepts from pre-trained models, with promising results [16, 17, 29, 34]. Yet, these methods limitedly scale. While they can remove a limited number of concepts, deleting hundreds strongly degrades the model's generative capacity [17, 34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 505 + ], + "type": "text", + "content": "To address these challenges, we introduce AlignGuard, a scalable method for safety alignment of T2I models. Our idea is to remove a large number of concepts from T2I models, by exploiting an ensemble of safety experts trained with Direct Preference Optimization (DPO) on safety-oriented synthetic data. By doing so, we are the first to propose a scalable safety alignment method for T2I, allowing for a safe deployment of trained T2I models. Considering the lack of data for DPO-based safety alignment available for image generation, we introduce CoProV2, a dataset encompassing a broad array of unsafe concepts. CoProV2 generation is fully automatic, including paired images with associated prompts: one harmful image with its corresponding textual description, and one safe image and prompt with similar content and structure. This setup enables the application of DPO for aligning T2I models effectively. In practice, we train Low-Rank Adaptation (LoRA) [24] matrices. Each LoRA serves as a safety expert trained to prevent the generation of specific harmful content, i.e. content related to different categories such as \"Hate,\" \"Sexual,\" \"Violence,\" among others. The LoRAs guide the vision representation of the pre-trained T2I model towards safe outputs during the diffusion process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 508, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 639 + ], + "type": "text", + "content": "After training the safety experts, we use a new algorithm called Co-Merge to merge all safety LoRA experts into one. This mitigates interferences among different experts, allowing for a significant increase in scalability. As shown in Figure 1, through joint expert training and subsequent merging, we can train on images encompassing 723 harmful concepts—approximately seven times more than existing approaches [17]. In summary, AlignGuard enables safety alignment of T2I models at scale, preventing the generation of inappropriate outputs without impacting generative capabilities on safe prompts. Our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 642, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 295, + 666 + ], + "type": "text", + "content": "1. We propose AlignGuard, the first scalable approach for T2I models specifically targeting safety alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 666, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 295, + 701 + ], + "type": "text", + "content": "2. We introduce a training method based on expert models, and a novel merging strategy based on weight activation frequency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "content": "3. For the training of AlignGuard, we craft a new dataset" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 95 + ], + "type": "text", + "content": "for safety alignment in T2I, coined CoProV2, that we will release as open source." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 96, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 132 + ], + "type": "text", + "content": "4. We validate AlignGuard on popular T2I models and multiple benchmarks, demonstrating its effectiveness in many setups." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 144, + 400, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 400, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 400, + 156 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 164, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 297 + ], + "type": "text", + "content": "Content filtering. Often, closed-source commercial T2I use prompt blacklists, LLM preprocessing, and image analysis for content filtering [1, 3, 4]. In particular, some use LLM-based prompt analysis [35] to detect harmful inputs, even with ad-hoc models [27]. Recently, Latent Guard [32] proposed a latent-based blacklist within text encoders in T2I models. Some works rely instead on the analysis of generated images [2], where an NSFW classifier is applied to generated images. Instead, others [40] use inpainting to mask potentially unsafe content. These approaches can easily be deactivated if the T2I models' weights are available." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 303, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 483 + ], + "type": "text", + "content": "Concept removal in T2I. Some have explored removing the capability of generating concepts in T2I, like SLD [52], using classifier-free guidance [22] to steer generation away from undesirable outputs. Similarly, Li et al. [30] identify interpretable directions that can be used for safety steering. Alternatively, many fine-tune the T2I. While seminal works finetune the entire model [29], some focus on specific components, such as attentions [17, 38, 62], specific neurons [10], or the textual encoder [43]. There has been a recent interest in mass removal of concepts from T2I models for safety purposes [17, 34]. While MACE [34] uses a similar strategy as ours, they make use of segmentation masks, constraining the erasable concepts at spatially-defined ones. Moreover, all available approaches [17, 34] degrade performance with more than 100 erased concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 489, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 658 + ], + "type": "text", + "content": "Model merging. Model merging, i.e. combining multiple models into one model, has gained attention [12, 61]. Weighted averaging methods are commonly used to enhance performance [18, 26, 57], especially in scenarios where only model weights are accessible [28]. Beyond these, advanced techniques have emerged, improving over basic averaging [6, 36, 39]. For example, TIES [58] resolves operator conflicts to improve merging, at the cost of hyperparameter tuning. Merging has been explored for safety in LLMs [8, 19] and for concept removal in MACE [34], but only using expensive optimization procedures. Existing LoRA merging for T2I focus on single subjects [54, 63]. Instead, we propose a cheap and effective merging strategy, focusing on broad safety categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 669, + 399, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 399, + 681 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 399, + 681 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We use preference optimization algorithms such as DPO [56] to perform scalable safety alignment of T2I mod" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17025" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "els. Here, we revise the fundamental concepts to allow for the interpretation of our method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 224, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 224, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 224, + 117 + ], + "type": "text", + "content": "3.1. Text-to-image diffusion models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": "Diffusion models allow for image generation by iteratively denoising gaussian noise with a network " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " iterations [23]. In particular, T2I diffusion models include natural language conditioning, so the desired output image can be described with text. We now briefly introduce the training procedure for a T2I model. Let us assume an input pair " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "(x,p)\\sim \\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " sampled from a training dataset " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}}" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " of images " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " and paired textual description " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": ". The network " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " is trained by estimating the ground truth noise " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\tilde{\\epsilon}_t" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " injected on " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": " for a random " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": ". We define the input image with the addition of noise as " + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 121, + 296, + 254 + ], + "type": "text", + "content": ". Hence, the training loss is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 263, + 295, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 263, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 110, + 263, + 295, + 277 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} (\\epsilon , x, p) = | | \\tilde {\\epsilon} _ {t} - \\epsilon \\left(x _ {t}, p\\right) | |. \\tag {1}", + "image_path": "a2e9391457ec7708ac32bc148d0754b7787a9e1b2a5be26d6a8d0609dcba20c2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "text", + "content": "In Eq. (1), " + }, + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\epsilon (\\cdot)" + }, + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "text", + "content": " is the denoising operation. The network weights " + }, + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 286, + 296, + 312 + ], + "type": "text", + "content": " are optimized by minimizing the following:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 322, + 295, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 322, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 89, + 322, + 295, + 341 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x, p) \\sim \\mathcal {D} _ {\\text {t r a i n}}} \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\theta}, x, p\\right)\\right), \\tag {2}", + "image_path": "241cb138d487083b1fc965eb9c7d1ead9dd8411e47c28d92c0a7ead8d1e09072.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": " refers to the denoising network with a set of weights " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ". During inference, sampled Gaussian noise is iteratively processed for " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta^*}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": " following specific scheduling policies, ultimately allowing image synthesis. We refer to [23] for additional details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 198, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 198, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 198, + 426 + ], + "type": "text", + "content": "3.2. DPO for diffusion models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": "DPO is a technique for preference alignment initially developed for LLMs [41, 45] and recently extended to diffusion models [33, 56]. The core idea is to benefit from pairwise preferences obtained by labeling. Let us assume a dataset " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{DPO}}" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": " of paired images and textual descriptions " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "(x^{+}, x^{-}, p)" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": ". For a given description " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": " is a image that humans indicated as preferred output with respect to " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "x^{-}" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": ". The intuition of DPO is to increase the likelihood to generate the preferred output " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "x^{+}" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": ", while discouraging the generation of " + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "inline_equation", + "content": "x^{-}" + }, + { + "bbox": [ + 55, + 432, + 296, + 552 + ], + "type": "text", + "content": ". This translates into the following loss:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 561, + 294, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 561, + 294, + 607 + ], + "spans": [ + { + "bbox": [ + 88, + 561, + 294, + 607 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon , x ^ {+}, x ^ {-}, p\\right) = - \\log \\sigma (- \\beta ( \\\\ \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {+}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {+}, p\\right)\\right) \\tag {3} \\\\ - \\left(\\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon , x ^ {-}, p\\right) - \\mathcal {L} _ {\\text {d i f f}} \\left(\\epsilon_ {\\text {r e f}}, x ^ {-}, p\\right)\\right)\\left. \\right), \\\\ \\end{array}", + "image_path": "c91b84017c06b3613f772743c79a751c361b0dcffc296b0040010a142a302053.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "content": " is a reference pre-trained network, typically resulting from a previous optimization of Eq. 2. Also, " + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "content": " is the sigmoid operation and " + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 616, + 296, + 664 + ], + "type": "text", + "content": " is a weighting constant [56]. Finally, one could optimize as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 673, + 296, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 673, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 56, + 673, + 296, + 693 + ], + "type": "interline_equation", + "content": "\\theta_ {\\mathrm {D P O}} ^ {*} = \\underset {\\theta} {\\arg \\min } \\mathbb {E} _ {(x ^ {+}, x ^ {-}, p) \\sim \\mathcal {D} _ {\\mathrm {D P O}}} \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {+}, x ^ {-}, p\\right). \\tag {4}", + "image_path": "a2a359ae4d8158d3a8c3b46dc4190740b81f9627e207cda9b1e2f5e5e2e12c5f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 701, + 276, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 701, + 276, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 701, + 276, + 714 + ], + "type": "text", + "content": "For further details, we refer to the original paper [56]." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 566, + 207 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 566, + 207 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 566, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 566, + 207 + ], + "type": "image", + "image_path": "5fdb24e42ff8224b830deaba3abed6b578568830b836ac19886aeb974c6385d9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 218, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 275 + ], + "type": "text", + "content": "Figure 2. Dataset generation. For each unsafe concept in different categories, we generate an corresponding prompts with an LLM. We generated paired safe prompts using an LLM, minimizing semantic differences. Then, we use the T2I we intend to align to generate corresponding images for both prompts." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 294, + 370, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 294, + 370, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 294, + 370, + 306 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 399 + ], + "type": "text", + "content": "We aim to align T2I models inspired by practices in language model alignment [13]. We avoid expensive human annotations by generating safety-oriented preference data in Section 4.1. We also propose a training procedure based on expert networks (Section 4.2) and subsequent merging (Section 4.3). After training, AlignGuard requires no additional computation at inference (more details in supplement)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 407, + 488, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 488, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 488, + 421 + ], + "type": "text", + "content": "4.1. DPO for safety alignment in T2I" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 557 + ], + "type": "text", + "content": "We aim to use DPO as a strategy to perform safety alignment on pretrained T2I models. Assuming an unsafe textual input describing a sexually explicit, shocking, or violent scene, we would like the diffusion model to ignore the unsafe requested traits, and generate a safe image, i.e. not including any visually disturbing trait. We achieve this by generating automatically {unsafe, safe} image pairs for DPO training, obtained by propting the T2I with LLM-manipulated text. In a nutshell, we can discourage the generation of unsafe images, having as preferred DPO image a visually-close safe sample, for an unsafe input T2I prompt." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "We start by generating a set of unsafe concepts " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " with an LLM. Each " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " is a keyword representing unsafe elements that may be included in an image, such as \"homicide\", \"nude\", etc. Then, we follow the prompt synthesis procedure of [32], generating both unsafe and safe prompts for image generators from concepts, enforcing minimal differences. Specifically, we sample an unsafe prompt " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " with an LLM to create text visually describing a scene using the input concept " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ". Then, we further process the generated " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ", prompting an LLM to remove any reference to the input unsafe concept " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ". This transforms the prompt into a safe " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ", while minimizing semantic changes. Differently from [32], we process both prompts with a pretrained" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17026" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 76, + 366, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 76, + 366, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 76, + 366, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 76, + 366, + 220 + ], + "type": "image", + "image_path": "6c3051da9780dc1e8aa531b8cf54a27aea7d983ce297ae0892e469b75baa53b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 264 + ], + "type": "text", + "content": "Figure 3. Expert training and merging. First, we use the previously generated prompts and images to train LoRA experts on specific safety categories (left), exploiting our DPO-based losses. Then, we merge all the safety experts with Co-Merge (right). This allows to achieve general safety experts that produce safe outputs for a generic unsafe input prompt in any category." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 374, + 76, + 550, + 213 + ], + "blocks": [ + { + "bbox": [ + 374, + 76, + 550, + 213 + ], + "lines": [ + { + "bbox": [ + 374, + 76, + 550, + 213 + ], + "spans": [ + { + "bbox": [ + 374, + 76, + 550, + 213 + ], + "type": "image", + "image_path": "3bc2b8c155eeeec4e24869edca388100096047370f005f4e1587af59a980998b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 281, + 298, + 369 + ], + "blocks": [ + { + "bbox": [ + 57, + 281, + 298, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 281, + 298, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 281, + 298, + 369 + ], + "type": "image", + "image_path": "b771cc2f950c55918ffe042517d06b9196474e654b34b1a768f3d99b12d54e30.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 295, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 434 + ], + "type": "text", + "content": "Figure 4. Merging experts with Co-Merge. (left) Assuming LoRA experts with the same architecture, we analyze which expert has the highest activation for each weights for all inputs. (right) Then, we obtained the merged weights from multiple experts, merging only the most active weights per expert." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "text", + "content": "diffusion model, obtaining images " + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\{x^{\\mathrm{U}}, x^{\\mathrm{S}}\\}" + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 453, + 295, + 479 + ], + "type": "text", + "content": ", respectively. We show this in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": "In this way, we can automatically create image preference pairs, assuming that the preferred output in safety alignment would always be a similar safe image for an unsafe prompt. Indeed, by repeating the process multiple times, we can construct a dataset " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{safety}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " consisting of tuples including the sampled prompts and corresponding generated images, i.e. " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "(x^{\\mathrm{S}}, x^{\\mathrm{U}}, p^{\\mathrm{S}}, p^{\\mathrm{U}})" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": ". Then, we use DPO to discourage the generation of unsafe outputs for an unsafe prompt " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": ", and promote the generation of the paired safe image " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " instead. To achieve this, we set " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " as preferred output for " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": ", discouraging the generation of " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": ". The semantic similarity of " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " will encourage the DPO training to focus exclusively on the visual traits making " + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{U}}" + }, + { + "bbox": [ + 55, + 491, + 296, + 658 + ], + "type": "text", + "content": " unsafe. We define a first loss term:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 111, + 669, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {U}\\right). \\tag {5}", + "image_path": "14a93a42c1b1974887c8682036bc94c64272866fb6d8e5910e120f7edd70637c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "While effective in removing the unsafe concepts " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " from the generated images, this risks to reduce the generative capa" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "text", + "content": "bilities of the network, since we are encouraging to generate images that do not respect the input prompt. To prevent this, we also use " + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "text", + "content": " as preferred generation for the input prompt " + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "inline_equation", + "content": "p^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "text", + "content": ", while penalizing on " + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "inline_equation", + "content": "x^{\\mathrm{U}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 332 + ], + "type": "text", + "content": ". This writes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 373, + 340, + 555, + 356 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 340, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 373, + 340, + 555, + 356 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n}} = \\mathcal {L} _ {\\mathrm {D P O}} \\left(\\epsilon_ {\\theta}, x ^ {S}, x ^ {U}, p ^ {S}\\right). \\tag {6}", + "image_path": "a9dc53e6a6158f68b9bdf9aaa706bd371b3e025e882517c992b96e7f4cbe2c80.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 554, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 377 + ], + "type": "text", + "content": "Finally, we modify Eq. (4) proposing our training objective:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 397, + 554, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 397, + 554, + 416 + ], + "spans": [ + { + "bbox": [ + 312, + 397, + 554, + 416 + ], + "type": "interline_equation", + "content": "\\theta_ {\\text {A l i g n G u a r d}} ^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {(x ^ {\\mathrm {s}}, x ^ {\\mathrm {U}}, p ^ {\\mathrm {s}}, p ^ {\\mathrm {U}}) \\sim \\mathcal {D} _ {\\text {s a f e t y}}} \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {c o n}}. \\tag {7}", + "image_path": "65e77607eed8cf193d1220ba1208cea6d2ef74999f6d0673556e6dd69a885a11.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 461 + ], + "type": "text", + "content": "This objective allows for a stable safety-oriented finetuning of the T2I model for a large number of concepts " + }, + { + "bbox": [ + 313, + 425, + 554, + 461 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 425, + 554, + 461 + ], + "type": "text", + "content": " without impacting the generative capabilities of the T2I model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 512, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 512, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 512, + 481 + ], + "type": "text", + "content": "4.2. Improving scaling with safety experts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 486, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 703 + ], + "type": "text", + "content": "Training with our proposed strategy in Section 4.1 enables us to scale the number of concepts used for safety alignment, while keeping generative capabilities intact. However, by increasing the number of concepts in " + }, + { + "bbox": [ + 313, + 486, + 555, + 703 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 486, + 555, + 703 + ], + "type": "text", + "content": ", we noticed a decrease in performance. We suppose this is due to the complexity of the task: in safety alignment there is a need of a high contextual understanding of the generated scenes. As an example, an image of a kitchen including knives is a safe output, while a scene of harassment with the usage of a knife is an undesired unsafe outcome of the image synthesis process, that we aim to prevent. As a solution, we propose to decompose the safety alignment on separate categories, such as \"violence\", \"sexual content\", \"harassment\", and others. Our intuition is that by limiting the variability of the generated images, it would be easier to identify visual patterns that would make a scene \"violent\", or \"sexually explicit\". Hence, we aim to train safety experts, each focusing on one category only." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17027" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 71, + 296, + 270 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": "Algorithm 1 Co-Merge \n1: Input: Safety expert LoRAs " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\{L_1, L_2, \\ldots, L_N\\}" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", unsafe prompts " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "\\{p_1, p_2, \\ldots, p_K\\}" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n2: Output: Merged LoRA " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "L_{\\text{merged}}" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n3: Init a count matrix " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "J \\times N" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " elements with zeros. \n4: for each prompt " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "k \\in [1, K]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " do \n5: for each neuron index " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "j \\in [1, J]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " do \n6: " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "i_{\\text{max}} := \\arg \\max_i |L_i^j(p_k)|" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "i \\in [1, N]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n7: Increment count " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "C[j, i_{\\text{max}}]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n8: end for \n9: end for \n10: for each neuron index " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "j \\in [1, J]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " do \n11: " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "m := \\arg \\max_i C[j, i]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "i \\in [1, N]" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "L_{\\text{merged}}^j := L_m^j" + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "text", + "content": " \n13: end for \n14: Return " + }, + { + "bbox": [ + 56, + 71, + 296, + 270 + ], + "type": "inline_equation", + "content": "L_{\\text{merged}}" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": "In practice, we identified " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": " safety-related concept categories, each linked to a generated concept set " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}^1,\\dots,\\mathcal{C}^N\\}" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": ". Each set is created by prompting a large language model to generate concepts within a given unsafe category. We then generated the corresponding safety datasets " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": " following the approach in Section 4.1. Rather than fine-tuning the entire model's weights, " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": ", to create the expert networks, we used low-rank adaptation (LoRA) [24], updating only a minimal subset of weights and as such saving significant computational costs. Ultimately, we apply Equation 7 for " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": ", obtaining a set of LoRAs " + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\{L^1,\\dots,L^N\\}" + }, + { + "bbox": [ + 55, + 291, + 296, + 462 + ], + "type": "text", + "content": ", as shown in Figure 3. At inference time, the obtained LoRAs can be applied to the pretrained T2I model to enable safe image generation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 468, + 157, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 157, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 157, + 481 + ], + "type": "text", + "content": "4.3. Experts merging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 593 + ], + "type": "text", + "content": "To prevent the generation of content across multiple safety-related categories, additional considerations are necessary. Indeed, loading multiple LoRAs into a model can lead to interference [25], reducing overall performance in safety-oriented image generation. Alternatively, multiple experts could be run independently, at the cost of multiplying inference costs. To mitigate these issues, we aim to merge the trained LoRAs into a single one, as visualized in Figure 3, without sacrificing performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 296, + 714 + ], + "type": "text", + "content": "We propose a novel data-based strategy, called CoMerge, to merge multiple safety experts, specifically designed around our use case. Our intuition is that by decomposing safety alignment in multiple categories (Section 4.2), trained LoRAs sharing the same architecture would encode information about category-specific alignment in different neurons. In other words, we assume that in the T2I model, the weights responsible for generating violent content (i.e. including, for instance, blood) will be different from those generating sexually-explicit content" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 319, + 71, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 553, + 140 + ], + "type": "table", + "html": "
DatasetImage # of Prompts # of Categories # of ConceptsIP
COCO [31]✓3,000N/AN/A0.06
I2P [51]✗4,7037N/A0.36
UD [44]✗9325N/A0.47
CoPro [32]✗56,52677230.23
CoProV2 (ours)✓23,69077230.43
", + "image_path": "c843468680b73d2ef2639073d1a00ea04a13fc2a9c79b526f9220f2d412f698c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 147, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 313, + 147, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 555, + 203 + ], + "type": "text", + "content": "Table 1. Datasets comparison. Our LLM-generated dataset, CoProV2, achieves comparable Inappropriate Probability (IP) to human-crafted datasets (UD [44], I2P [51]) and offers similar scale to CoPro [32]. COCO [31], exhibiting a low IP, is used as a benchmark for image generation with safe prompts as input." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 209, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 305 + ], + "type": "text", + "content": "(e.g. naked skin). LoRAs work as updates for the model weights [24], so, for concepts associated to different neurons, they will encode related information in different parts of the same architecture. Then, our idea is to isolate the neurons inside the trained safety experts that activate the most for unsafe input prompts, and construct a merged expert as a LoRA " + }, + { + "bbox": [ + 313, + 209, + 555, + 305 + ], + "type": "inline_equation", + "content": "L_{\\text{merged}}" + }, + { + "bbox": [ + 313, + 209, + 555, + 305 + ], + "type": "text", + "content": " including the parameters of different experts associated to the most significant activations only." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": "We start by randomly sampling " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " unsafe prompts from the training set of " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\mathrm{safety}}^1,\\dots,\\mathcal{D}_{\\mathrm{safety}}^N\\}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", equally distributed across each category. We then process all prompts with the T2I model, where we load all trained experts " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\{L_1,\\dots,L_N\\}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " one at the time. Each expert LoRA is a multi-layer perceptron including " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " neurons. For each prompt " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "p_k^U" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "k\\in [1,K]" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", and each expert model " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "i\\in [1,N]" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", we record the absolute value of the activation from each neuron " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "L_{i}^{j}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "j\\in [1,J]" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ". We denoted the absolute value as " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "|L_i^j (p_k^U)|" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ". This provides a measure of the response for each neuron to that prompt. These prompts are uniformly sampled from distinct safety categories, so different neurons of the same index " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " across the experts will exhibit different activations in response to different prompts. To create the merged " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{merged}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", we identify the neurons with the highest activation frequencies across experts. Specifically, for each neuron " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " in the merged expert " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{merged}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", we select the neuron from the original set of experts " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\{L_1,\\dots,L_N\\}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " that has the highest activation frequency count across the " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " prompts. We summarize this in Algorithm 1, where the " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "p^U" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " superscript is omitted for clarity. Figure 4 illustrates Co-Merge with " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "K = 10" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": ", describing how the frequency count leads to the network weight merging. Note that merging " + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{merged}}" + }, + { + "bbox": [ + 313, + 306, + 556, + 606 + ], + "type": "text", + "content": " with the original T2I weights makes safety alignment difficulty reversible, a beneficial practice for T2I releases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 395, + 624 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 665 + ], + "type": "text", + "content": "We first introduce our setup (Section 5.1), and compare with baselines in Section 5.2. We then analyze the properties of AlignGuard (Section 5.3) and show ablations (Section 5.4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 429, + 685 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Baselines. We use AlignGuard to align three models for T2I, namely Stable Diffusion (SD) v1.5/v2.1 [48] and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17028" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 205 + ], + "type": "table", + "html": "
MethodIP↓FID ↓ COCOCLIP ↑
CoProV2I2PUD
SD1.5No alignment0.510.360.5269.7733.52
SLD [51]0.270.190.3071.4532.24
ESD-u [16]0.220.250.2172.9829.61
UCE [17]0.330.300.3872.0132.01
AlignGuard0.070.110.1670.9632.32
SD2.1No alignment0.510.350.5570.4534.99
AlignGuard0.120.120.1771.9934.13
SDXLNo alignment0.490.310.4768.9035.60
AlignGuard0.090.080.1175.2034.67
", + "image_path": "e7e2e05e479fd3cc956c447ce19f6d02fd4fc1b209c090d377ec345632a17b3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 278 + ], + "type": "text", + "content": "Table 2. Benchmarks. AlignGuard achieves best performance both in generated images alignment (IP) and image quality (FID, CLIPScore), with two T2I models and against 3 methods for SD v1.5. Note that we use CoProV2 only for training, hence I2P and UD are out-of-distribution. Yet, AlignGuard allows a robust safety alignment. Best is bold, second is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 385 + ], + "type": "text", + "content": "SDXL [42]. For all, we compare the original versions released on HuggingFace, and our finetuned version with AlignGuard. Moreover, we compare with recent methods based on SD v1.5 for safe image generation (SLD [51]) and concept erasure (ESD [16], UCE [17]). In particular, for ESD, we follow the paper indications and use ESD-u [16], a version of ESD for broad concept removal. We perform our analysis and ablations on SD v1.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 393, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 478 + ], + "type": "text", + "content": "Metrics. To evaluate the effectiveness of safety alignment, we use the inappropriate probability metric (IP) from SLD [51], measuring the ratio of unsafe contents generated by T2I with unsafe prompts. The unsafe content detection is performed by using classification results of Q16 [50] and NudeNet [47]. For evaluating image quality, we use FID [21], and CLIPScore [20] for text-image alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Datasets. For training of AlignGuard, we generate a new dataset following Section 4.1, named CoProV2 (Concepts and Prompts). We motivate this choice by observing a limitation in the CoPro dataset [32], i.e. the limited IP measured on images generated with the prompts of the dataset. This suggests that many prompts in CoPro are not leading to unsafe generations. This is due to the limited performance of the LLM building CoPro. We start from the same " + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": " as CoPro: 723 harmful concepts across 7 categories (Hate, Harassment, Violence, Self-Harm, Sexual, Shocking, Illegal). We re-generate the unsafe/safe prompts pairs with a new LLM prompt reported in the supplement. We use Mistral-Nemo-Instruct [37] for the prompt synthesis. We generate 23,690 pairs of safe/unsafe prompts, and we use 15,690/8,000 pairs for training/ testing. For testing only, we use I2P [7] and Unsafe Diffusion (UD) [44], including 4703/932 human-designed unsafe prompts on 7/5 categories, respectively. Following common practices [16, 48, 51] we use COCO [31], with 3000 safe" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "captions and corresponding images, for evaluating the performance of the T2I model in image generation after alignment. We compare the datasets quantitatively in Table 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "content": "Implementation details. We train AlignGuard with batch-size 8, accumulating over 16 steps. We use AdamW with a learning rate of " + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "content": " for training each LoRA with rank 4. For Co-Merge, we set " + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "inline_equation", + "content": "K = 100" + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "content": " unsafe prompts. Each LoRA takes " + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "inline_equation", + "content": "\\sim 2" + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "content": " hours to train for SD and " + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "inline_equation", + "content": "\\sim 18" + }, + { + "bbox": [ + 313, + 112, + 555, + 209 + ], + "type": "text", + "content": " hours for SDXL on a single Nvidia 5880 Ada GPU for 2000 steps. We normalize the computation needed for training AlignGuard and baselines. More details are in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 213, + 397, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 397, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 397, + 225 + ], + "type": "text", + "content": "5.2. Benchmarks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "content": "Quantitative evaluation. In Table 2, we compare with alignment baselines for IP, FID, and CLIPScore. For fairness, all baselines were re-trained on CoProV2 training set and evaluated using the CoProV2 test set, I2P, and UD. Note that I2P and UD serve as out-of-distribution evaluations as they are crafted by human users in real deployment scenarios, differing in both prompt format and content from CoProV2. Initially, we trained the baselines using the complete " + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "content": " set of CoProV2 with Stable Diffusion v1.5 as the backbone. This involved using " + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "content": " to identify negative embeddings in SLD and excluding all " + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "content": " concepts for ESD-u and UCE. All baselines collapsed due to the high number of concepts in CoProV2, aligning with the limited scaling performance noted in previous works [16, 17]. These results are included in the supplement. For successful training comparisons, we used the 7 category names of CoProV2 as removable concepts for the baselines, while we use the full prompts and images in CoProV2 for AlignGuard. We trained our expert LoRAs on samples generated with all " + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 231, + 555, + 687 + ], + "type": "text", + "content": ", one per category, and merge them with Co-Merge. As shown in the table, we outperform considerably all baselines in every metric. Notably, our IP for SD v1.5 on CoProV2 (0.07) is much lower than the next best (ESD-u, 0.22). Additionally, AlignGuard demonstrates superior generalization, achieving better IP scores on unseen I2P (0.08) and UD (0.16). Our DPO-based strategy effectively maintains Stable Diffusion's generation capabilities preventing forgetting of concepts unrelated to safety. Indeed we report nearing the performance of the Baseline method without safety alignment in FID (70.96 vs 69.77) and CLIPScore (32.32 vs 33.52) on COCO captions. We also applied AlignGuard to SD v2.1 and SDXL to test the generalization of our method to a different diffusion model. Our results are consistent with different backbones like SD v2.1 and SDXL, achieving performance comparable to our aligned SD v1.5 with IP 0.09 and 0.12 on CoProV2. This suggests that AlignGuard can be used for the safety alignment of arbitrary T2I models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Qualitative evaluation. We show in Figure 5 a selection of samples advocating the effectiveness of AlignGuard. We" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17029" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 555, + 378 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 378 + ], + "type": "image", + "image_path": "21c1d789409b3ec5f0c56b1782d9d048cef3fc9b545158863582e81dabbb3421.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 555, + 422 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison. Compared to non-aligned baseline models, AlignGuard allows to synthesize safe images for unsafe input prompts. Please note the layout similarity between the unsafe and safe outputs: thanks to our training, only the harmful image traits are removed from the generated images. Concepts in ⟨brackets⟩. Prompts are shortened, for full ones see the supplement." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 433, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 553 + ], + "type": "text", + "content": "sample prompts from CoProV2. For SD v1.5 and SDXL (SD v2.1 in the supplement), we prevent the generation of unsafe content across 7 categories. Please also note how the quality of the generated images, in both models aligned with AlignGuard, is not impacted by our alignment procedure. Also, let us highlight that although the model refuses to follow unsafe prompts, the generated images preserve similar traits and global layout, e.g. the pose of the man in the \"Shocking\" column for SDXL. This property results from training on safe/unsafe pairs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 127, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 127, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 127, + 576 + ], + "type": "text", + "content": "5.3. Properties" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Experts performance. For AlignGuard, we train an expert for each category in CoProV2. We now examine the effectiveness of each expert LoRA across all categories in the CoProV2 test set in Table 3. The results show that training and evaluating on the same category generally leads to good performance, as expected. Interestingly though, applying one expert LoRA consistently improves IP across all categories and often surpasses the expert trained specifically for that category. We attribute this to the interaction of multiple concepts across categories. While the concepts differ, the visual features of generated" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 433, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 601 + ], + "type": "text", + "content": "images may share similarities. For example, scenes in the \"Violence\" or \"Self-harm\" categories often depict blood, while the \"Hate\" category may include sexual slurs, which could explain the strong performance on \"Sexual\" data. Ultimately, this suggests that for safety alignment tasks, AlignGuard is robust to the concept selection. However, experts trained on different categories still encode distinct concepts. This becomes evident when merging all LoRAs with Co-Merge (All - Ours), since we achieve the best performance across all categories. We also include a baseline where a single LoRA is trained on prompts from all categories (All - single), which shows suboptimal performance. Here, the lack of specialized expert models results in less effective training, justifying our design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Adversarial attacks robustness. To understand if harmful concepts are effectively removed from the T2I model capabilities, we perform an additional experiment based on text-based adversarial attacks. Those allow to optimize text apparently innocuous, but leading to unsafe generation. If AlignGuard is able to prevent harmful generations in this setup, it means that our alignment is effective even if harmful concepts are not explicitly present in the input prompt." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17030" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 296, + 243 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 296, + 243 + ], + "type": "table", + "html": "
ExpertIP on specific category ↓Avg.
HateHarass.ViolenceSelf-harmSexualShockingIllegal
No alignment0.490.480.540.590.540.520.440.51
Hate0.110.150.190.160.110.170.110.14
Harass.0.170.160.200.230.190.160.150.18
Violence0.180.150.160.190.180.170.160.17
Self-Harm0.210.240.260.250.220.240.230.23
Sexual0.180.190.250.250.160.210.190.20
Shocking0.170.180.220.200.150.190.160.18
Illegal0.170.140.190.200.200.210.120.18
All - Single0.170.210.220.220.160.210.180.16
All - Ours0.060.060.090.070.070.080.040.07
", + "image_path": "ac6497bcc8fc3a8317cebcb6fbd562db60eb152fc5469ee7f853e355ec1bd634.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 307, + 296, + 352 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 295, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 298 + ], + "type": "text", + "content": "Table 3. Effectiveness of merging. While training a single safety expert across all data (All - single), IP performance are lower or comparable to single experts (previous rows). Instead, by merging safety experts (All - ours) we considerably improve results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 307, + 296, + 352 + ], + "lines": [ + { + "bbox": [ + 58, + 307, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 58, + 307, + 296, + 352 + ], + "type": "table", + "html": "
MethodMMA[59]Ring-A-Bell[55]SneakyPrompt[60]P4D[11]
No align.0.430.670.570.48
ESD-u [16]0.220.430.210.23
AlignGuard0.080.120.090.07
", + "image_path": "87a38a19d3f9a8dc3ddc5d4aeca2664ced9bdc06c9b24560665987e6fc16ad5e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "content": "We use 4 state-of-the-art attack methods listed in Table 4 to optimize seemingly innocuous prompts, starting from Co-ProV2 test prompts. We test alignment with AlignGuard and ESD-u [16]. The results advocate for AlignGuard's high robustness to adversarial attacks, proof of the effectiveness of alignment. In particular, for Ring-A-Bell [55], AlignGuard maintains an IP of 0.12, even for a particularly successful attack able to raise the IP of the baseline SD v1.5 to 0.67." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 154, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 154, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 154, + 552 + ], + "type": "text", + "content": "5.4. Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "DPO strategy. In AlignGuard training, several options exist for selecting preferred samples for DPO. Indeed, by selecting a paired safe image as a safe output for an unsafe prompt, we depart slightly from common practices in LLM alignment, where the preferred answer in presence of an unsafe input is a refusal [13]. We here investigate this choice by mimicking refusals supervision in language models. To do so, we create two different alternative images that we use as positive samples for DPO in presence of an unsafe input: (1) a Black Image, and (2) an image constructed by applying a Warning Sign on top of an unsafe output. Examples are in the supplement. From results in Table 5a, we evince that our strategy exploiting a paired safe image is perform-" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 137 + ], + "blocks": [ + { + "bbox": [ + 55, + 360, + 296, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 416 + ], + "type": "text", + "content": "Table 4. Resistance to adversarial attacks. We evaluate with 4 adversarial attacks methods the performance of AlignGuard and the best baseline, ESD-u, in terms of IP. For a wide range of attacks, we are able to outperform the baselines, advocating for the effectiveness of our scalable concept removal strategy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 137 + ], + "type": "table", + "html": "
DPO strategyIP ↓FID ↓CLIP ↑
Black image0.2185.3831.22
Warning sign0.1670.0133.05
w/o Lcon0.1674.8230.78
Paired safe image (ours)0.0770.9632.32
", + "image_path": "862b35c8486c81b2e7f3a9e91f1b6b809f183636bf8eea800b90dff9d0f6ef68.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 150, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 392, + 139, + 476, + 149 + ], + "lines": [ + { + "bbox": [ + 392, + 139, + 476, + 149 + ], + "spans": [ + { + "bbox": [ + 392, + 139, + 476, + 149 + ], + "type": "text", + "content": "(a) DPO strategy ablation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 150, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 213 + ], + "type": "table", + "html": "
MergingIP ↓FID ↓CLIP ↑DataIP ↓FID ↓CLIP ↑
Git Re-Basin [5]0.3069.9833.4010%0.1972.8727.30
Task Vectors [26]0.25179.604.8025%0.1671.9127.34
Model Soups [57]0.1469.8333.4050%0.1772.7627.20
TIES [58]0.0979.7928.85100%0.0770.9632.32
Co-Merge0.0770.9632.32
", + "image_path": "80274c555d5d57f9750a88f4526d923cd27eb4cce6d79cd0efb636582f4b3bd0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 334, + 214, + 428, + 224 + ], + "lines": [ + { + "bbox": [ + 334, + 214, + 428, + 224 + ], + "spans": [ + { + "bbox": [ + 334, + 214, + 428, + 224 + ], + "type": "text", + "content": "(b) Merging method ablation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 460, + 214, + 541, + 224 + ], + "lines": [ + { + "bbox": [ + 460, + 214, + 541, + 224 + ], + "spans": [ + { + "bbox": [ + 460, + 214, + 541, + 224 + ], + "type": "text", + "content": "(c) Scaling performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 233, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 278 + ], + "type": "text", + "content": "Table 5. Ablation studies. We check the effects of alternative strategies for DPO, proving that our approach is best (a). Co-Merge is also the best merging strategy compared to baselines (b). Finally, we verify that scaling data improve our performance (c)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "text", + "content": "ing significantly better, justifying the different design of the DPO positive between T2I and LLM alignment. Also, we compared in the same table with a setup excluding " + }, + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "text", + "content": ", i.e. training without considering generative capabilities preservation among safe samples. This results in lower FID and higher CLIP scores, highlighting the importance of " + }, + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{con}" + }, + { + "bbox": [ + 313, + 291, + 555, + 376 + ], + "type": "text", + "content": " to preserve generative capabilities in presence of safe inputs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 380, + 556, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 500 + ], + "type": "text", + "content": "Merging strategy. We evaluate our merging method against four baselines: (1) Git Re-Basin, (2) Task Vectors, (3) Model Soups [57] and (4) TIES [58]. We report results in Table 5b, showing that Co-Merge consistently outperform these baselines. Let us highlight that baselines do not use data for performing merging, leading to suboptimal results and at the cost of tedious hyperparameter tuning. By performing a data-aware merging using unsafe prompts, instead, we are able to optimally balance the contributions of each expert with minimal effort and computational load." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 503, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 556, + 588 + ], + "type": "text", + "content": "Scaling performance. We evaluate the importance of the scale of data in Table 5c. We subsample CoProV2 with different percentages and retrain SD v1.5 with AlignGuard. For datasets increasingly smaller, we observe an expected decrease in performance. With " + }, + { + "bbox": [ + 313, + 503, + 556, + 588 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 503, + 556, + 588 + ], + "type": "text", + "content": " of the data, we perform best, proving that AlignGuard benefits the most from large datasets and paving the way for larger trainings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "We introduced AlignGuard, the first approach dedicated to safety alignment for T2I models at scale. By achieving safety alignment on a broad range of concepts, we allow for safe generations in the presence of harmful inputs. AlignGuard enables text-to-image models to generate outputs that are not only of high quality but also aligned with safety and ethical guidelines. We hope our work will inspire future practices in T2I deployment." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17031" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 191 + ], + "type": "text", + "content": "This research was supported by the Research Grant Council of the Hong Kong Special Administrative Region under grant number 16212623. FP is funded by KAUST (Grant DFR07910). AK, JG, and PT are supported by UKRI grant: Turing AI Fellowship EP/W002981/1, and by the Royal Academy of Engineering under the Research Chair and Senior Research Fellowships scheme. The authors thank Alexander Pondaven and David Romero for their kind proofreading." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 213, + 115, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 115, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 115, + 225 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 233, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 233, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 294, + 255 + ], + "type": "text", + "content": "[1] The complete list of banned words in midjourney you need to know, 2022. Link. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 257, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 257, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 294, + 278 + ], + "type": "text", + "content": "[2] Diffusers: State-of-the-art diffusion models, 2022. Link. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 280, + 208, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 208, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 208, + 291 + ], + "type": "text", + "content": "[3] Dall-e 3 system card, 2023. Link. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 293, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 295, + 314 + ], + "type": "text", + "content": "[4] Leonardo ai content moderation filter: Everything you need to know, 2023. Link. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 316, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 349 + ], + "type": "text", + "content": "[5] Samuel K Ainsworth, Jonathan Hayase, and Siddhartha Srinivasa. Git re-basin: Merging models modulo permutation symmetries. arXiv preprint arXiv:2209.04836, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 350, + 294, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 294, + 383 + ], + "type": "text", + "content": "[6] Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes. arXiv preprint arXiv:2403.13187, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 294, + 407 + ], + "type": "text", + "content": "[7] Leonard Bereska and Efstratios Gavves. Mechanistic interpretability for ai safety-a review. TMLR, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 409, + 295, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 451 + ], + "type": "text", + "content": "[8] Rishabh Bhardwaj, Do Duc Anh, and Soujanya Poria. Language models are a hero simpson! safety re-alignment of fine-tuned language models through task arithmetic. In ACL, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 486 + ], + "type": "text", + "content": "[9] Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. Typology of risks of generative text-to-image models. In Conference on AI, Ethics, and Society, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 521 + ], + "type": "text", + "content": "[10] Ruchika Chavhan, Da Li, and Timothy Hospedales. Conceptprune: Concept editing in diffusion models via skilled neuron pruning. arXiv preprint arXiv:2405.19237, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "text", + "content": "[11] Zhi-Yi Chin, Chieh-Ming Jiang, Ching-Chun Huang, PinYu Chen, and Wei-Chen Chiu. Prompting4debugging: Redteaming text-to-image diffusion models by finding problematic prompts. In ICML, 2024. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 568, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 600 + ], + "type": "text", + "content": "[12] Leshem Choshen, Elad Venezian, Noam Slonim, and Yoav Katz. Fusing finetuned models for better pretraining. arXiv preprint arXiv:2204.03044, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 656 + ], + "type": "text", + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, et al. Near to mid-term risks and opportunities of open source generative ai. In ICML, 2024. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[15] Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Aaron Purewal, Csaba Botos, et al. Risks and opportunities of open-source generative ai. arXiv preprint arXiv:2405.08597, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[16] Rohit Gandikota, Joanna Materzynska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In ICCV, 2023. 2, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 196 + ], + "type": "text", + "content": "[17] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In WACV, 2024. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 199, + 553, + 232 + ], + "type": "text", + "content": "[18] Vipul Gupta, Santiago Akle Serrano, and Dennis DeCoste. Stochastic weight averaging in parallel: Large-batch training that generalizes well. In ICLR, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 234, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 277 + ], + "type": "text", + "content": "[19] Hasan Abed Al Kader Hammoud, Umberto Michieli, Fabio Pizzati, Philip Torr, Adel Bibi, Bernard Ghanem, and Mete Ozay. Model merging and safety alignment: One bad model spoils the bunch. In EMNLP Findings, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 312 + ], + "type": "text", + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 357 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 359, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 359, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 359, + 553, + 381 + ], + "type": "text", + "content": "[22] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshops, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 404 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 407, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 449 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 453, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 495 + ], + "type": "text", + "content": "[25] Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. In " + }, + { + "bbox": [ + 316, + 453, + 553, + 495 + ], + "type": "inline_equation", + "content": "COLM" + }, + { + "bbox": [ + 316, + 453, + 553, + 495 + ], + "type": "text", + "content": ", 2024. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[26] Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. In ICLR, 2022. 2, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 598 + ], + "type": "text", + "content": "[27] Hakan Inan, Kartikeya Upasani, Jianfeng Chi, Rashi Rungta, Krithika Iyer, Yuning Mao, Michael Tontchev, Qing Hu, Brian Fuller, Davide Testuggine, et al. Llama guard: Lmbased input-output safeguard for human-ai conversations. arXiv preprint arXiv:2312.06674, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "type": "text", + "content": "[28] Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. Dataless knowledge fusion by merging weights of language models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[29] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17032" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[32] Runtao Liu, Ashkan Khakzar, Jindong Gu, Qifeng Chen, Philip Torr, and Fabio Pizzati. Latent guard: a safety framework for text-to-image generation. In ECCV, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[33] Runtao Liu, Haoyu Wu, Ziqiang Zheng, Chen Wei, Yingqing He, Renjie Pi, and Qifeng Chen. Videodpo: Omni-preference alignment for video diffusion generation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8009-8019, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "text", + "content": "[34] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[35] Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. A holistic approach to undesired content detection in the real world. In AAAI, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "text", + "content": "[36] Michael S Matena and Colin A Raffel. Merging models with fisher-weighted averaging. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 213, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 213, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 213, + 326 + ], + "type": "text", + "content": "[37] Mistral. Mistral nemo, 2024. Link. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[38] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In ICCV, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 294, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 393 + ], + "type": "text", + "content": "[39] Guillermo Ortiz-Jimenez, Alessandro Favero, and Pascal Frossard. Task arithmetic in the tangent space: Improved editing of pre-trained models. NeurIPS, 36, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 436 + ], + "type": "text", + "content": "[40] Seongbeom Park, Suhong Moon, Seunghyun Park, and Jinkyu Kim. Localization and manipulation of immoral visual cues for safe text-to-image generation. In WACV, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 492 + ], + "type": "text", + "content": "[41] Renjie Pi, Tianyang Han, Wei Xiong, Jipeng Zhang, Runtao Liu, Rui Pan, and Tong Zhang. Strengthening multimodal large language model with bootstrapped preference optimization. In European Conference on Computer Vision, pages 382-398. Springer, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[43] Samuele Poppi, Tobia Poppi, Federico Cocchi, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Removing nsfw concepts from vision-and-language models for text-to-image retrieval and generation. In ECCV, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[44] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe diffusion: On the generation of unsafe images and hateful memes from text-to-image models. In Conference on Computer and Communications Security, 2023. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "text", + "content": "[45] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[46] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. In NeurIPS workshops, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 474, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 474, + 83 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 474, + 83 + ], + "type": "text", + "content": "[47] Github Repo. Nudenet, 2022. Link. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 85, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 85, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 85, + 553, + 118 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 152 + ], + "type": "text", + "content": "[49] Téo Sanchez. Examining the text-to-image community of practice: Why and how do people prompt generative ais? In Conference on Creativity and Cognition, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 196 + ], + "type": "text", + "content": "[50] Patrick Schramowski, Christopher Tauchmann, and Kristian Kersting. Can machines help us answering question 16 in datasheets, and in turn reflecting on inappropriate content? In FAccT, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[51] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "text", + "content": "[52] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "text", + "content": "[53] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[54] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. In ECCV. Springer, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 420 + ], + "type": "text", + "content": "[55] Yu-Lin Tsai, Chia-Yi Hsu, Chulin Xie, Chih-Hsun Lin, Jia-You Chen, Bo Li, Pin-Yu Chen, Chia-Mu Yu, and Chun-Ying Huang. Ring-a-bell! how reliable are concept removal methods for diffusion models? In ICLR, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 475 + ], + "type": "text", + "content": "[56] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 543 + ], + "type": "text", + "content": "[57] Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In ICML, 2022. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 577 + ], + "type": "text", + "content": "[58] Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merging models. In NeurIPS, 2024. 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "text", + "content": "[59] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Tsung-Yi Ho, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. In CVPR, 2024. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[60] Yuchen Yang, Bo Hui, Haolin Yuan, Neil Gong, and Yinzhi Cao. Sneakyprompt: Jailbreaking text-to-image generative models. In Symposium on Security and Privacy, 2024. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[61] Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In ICML, 2024. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[62] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17033" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "text", + "content": "[63] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. arXiv preprint arXiv:2402.16843, 2024. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17034" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_content_list.json b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..732795307f9974ca0fc4469e6acf0ef0f26414e7 --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_content_list.json @@ -0,0 +1,1597 @@ +[ + { + "type": "text", + "text": "Aligning Constraint Generation with Design Intent in Parametric CAD", + "text_level": 1, + "bbox": [ + 140, + 130, + 856, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Evan Casey Tianyu Zhang Joseph George Lambourne", + "bbox": [ + 109, + 180, + 385, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shu Ishida John Roger Thompson Amir Khasahmadi Pradeep Kumar Jayaraman Karl D.D. Willis", + "bbox": [ + 393, + 180, + 885, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/226e5492028c66d41e667c9d13e95612edc325a53548ccdbf76435d3b0d4da3a.jpg", + "image_caption": [ + "Figure 1. Left: a constraint solver is used to score model generated constraints A, B, C $\\sim$ $\\pi_{\\mathrm{base}}$ (and $A', B', C' \\sim \\pi_{\\mathrm{aligned}}$ ). Starting with the base model $\\pi_{\\mathrm{base}}$ , we post-train an aligned model $\\pi_{\\mathrm{aligned}}$ from this feedback. Right: Blue lines show original primitives and gray lines show geometric distortion when dimensions vary. The aligned model $\\pi_{\\mathrm{aligned}}$ produces fully-constrained sketches that preserve relative geometric relationships, whereas the base model $\\pi_{\\mathrm{base}}$ produces under-constrained sketches that may distort the geometry in unintended ways." + ], + "image_footnote": [], + "bbox": [ + 93, + 268, + 369, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/035129874c0020b4802ee29261be9676517a00e031d9f0265e15dae5b6b92f36.jpg", + "image_caption": [ + "Autodesk Research" + ], + "image_footnote": [], + "bbox": [ + 406, + 268, + 906, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 549, + 326, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We adapt alignment techniques from reasoning LLMs to the task of generating engineering sketch constraints found in computer-aided design (CAD) models. Engineering sketches consist of geometric primitives (e.g. points, lines) connected by constraints (e.g. perpendicular, tangent) that define the relationships between them. For a design to be easily editable, the constraints must effectively capture design intent, ensuring the geometry updates predictably when parameters change. Although current approaches can generate CAD designs, an open challenge remains to align model outputs with design intent, we label this problem 'design alignment'. A critical first step towards aligning generative CAD models is to generate constraints which fully-constrain all geometric primitives, without over-constraining or distorting sketch geometry. Using alignment techniques to train an existing constraint generation model with feedback from a constraint solver, we are able to fully-constrain $93\\%$ of sketches compared to $34\\%$ when using a naive supervised fine-tuning (SFT) baseline and only $8.9\\%$ without SFT. Our approach can be applied to any existing constraint generation model and sets the", + "bbox": [ + 88, + 583, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "stage for further research bridging alignment strategies between the language and design domains. Additional results can be found at https://autodeskailab.github.io/aligning-constraint-generation/.", + "bbox": [ + 511, + 551, + 906, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 631, + 645, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A central challenge in artificial intelligence (AI) is alignment: ensuring that AI systems produce outputs that adhere to human goals and expectations [11, 26, 30]. Although alignment of language models has been researched extensively [26, 28, 35], the application of alignment techniques to parametric design problems has yet to be studied. The use of AI in this discipline covers a broad range of areas, ranging from floor-plan layout [24, 34], to engineering design problems [9, 29, 41], to 3D generation [44]. Design problems are often visual in nature and incorporate other functional requirements, making them unique when compared with language model alignment. In this paper, we establish the problem of design alignment and demonstrate how this expansive problem can be made tractable by adapting techniques from the alignment literature into a new context.", + "bbox": [ + 509, + 657, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In language models, alignment is achieved by incorpo", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8613", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1589a5d6667317960fd546a4aa2613fd48726f40a6ef91ac873b37f59b3bc62.jpg", + "image_caption": [ + "Figure 2. An illustration of design intent in CAD modeling. The bottom sketch maintains symmetry after modifying a dimension due to properly applied constraints, while the top sketch, lacking adequate constraints, becomes asymmetrical and distorted." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 480, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rating feedback to generate coherent, contextually appropriate responses. Similarly, with parametric CAD modeling, AI tools must be aligned with a designer's intent by maintaining the underlying structural relationships to produce outputs that are both meaningful and functional. Otey et al. [25] define design intent as \"a CAD model's anticipated behavior when altered,\" while Martin [23] characterizes it as \"relationships between objects, so that a change to one propagates automatically to others.\" This means that modifications of a design by an AI system should yield outcomes where the established design relationships remain intact. To that end, we define design alignment as the application of generative modeling alignment techniques to produce outcomes that maintain design intent.", + "bbox": [ + 88, + 321, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The realization of AI systems that observe and maintain design intent has broad implications for the manufacturing and construction industries. Almost every manufactured object or structure begins as a CAD model. At the core of parametric CAD modeling are 2D engineering sketches, which can be extruded or revolved to generate 3D models. Engineering sketches are composed of geometric primitives, such as points, lines, and circles, that are organized using constraints and dimensions [7]. These constraints1 define geometric rules, including equality, perpendicularity, and radial or linear dimensions, which collectively shape the final layout of the sketch. When applied correctly, they enable efficient modifications while preserving the original design intent.", + "bbox": [ + 89, + 534, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 2 illustrates the impact of constraint quality: a poorly constrained sketch loses symmetry when a dimension is changed, whereas a well-constrained sketch preserves its intended relationships. This underscores the important role of constraints in maintaining design intent, and the need for AI systems that can align with this intent encoded in designs. We focus on the problem of sketch constraint generation [32] to demonstrate the adaption of align-", + "bbox": [ + 89, + 744, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment techniques to a design problem. Using an existing sketch constraint generation model Vitruvion [33], we align the model with algorithms that learn from feedback (Direct Preference Optimization [28], Expert Iteration [2, 36], RLOO [1], ReMax [22] and Group Relative Policy Optimization [35]) using the sketch constraint solver in Autodesk Fusion [3] as the learning signal. We optimize the models to remove all degrees of freedom in the sketches to become 'fully-constrained' [5], without causing sketches to be distorted, over-constrained or unsolvable. We further define these conditions in Section 3.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, this is the first instance of alignment methods being successfully applied to a parametric CAD design task; representing an important step forward for AI-assisted design tools. We present the following contributions:", + "bbox": [ + 511, + 258, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We establish the problem of design alignment, in the context of generative CAD models, as a critical component of AI-assisted CAD tools. We focus on the necessary first step of alignment for engineering sketches.", + "- We introduce a post-training strategy for a sketch constraint generation model using feedback from a sketch constraint solver. We define novel metrics and reward functions that directly optimize a base model for improved alignment.", + "- We conduct extensive experiments and demonstrate alignment techniques that fully-constrain $93\\%$ of sketches compared to $34\\%$ when using a naive supervised fine-tuning (SFT) baseline and only $8.9\\%$ without alignment. We posit that our approach is broadly applicable to other design tasks that require compilation of elements with rule-based algorithms." + ], + "bbox": [ + 511, + 338, + 903, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 595, + 653, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Engineering Sketches Engineering sketches form the 2D basis for 3D CAD models used to design mechanical parts for manufacturing. The availability of engineering sketch datasets [13, 32, 41] has enabled the development of generative models [13, 27, 33, 40] that can predict sketch geometry and/or the underlying constraints and dimensions that encode design intent. These Transformer-based [37] approaches create geometry by autoregressively generating tokens representing points and curves, then add constraints by referencing this geometry using Pointer Networks [38]. More recent approaches leverage image-based guidance [19, 42] or large language models (LLM) [18] in the constraint prediction task. Yang and Pan [45] learn to group together recurring patterns of geometric and constraint entities within a sketch, effectively discovering latent design concepts. However, none of these approaches explicitly optimize for preserving design intent – as a result, generated sketches may require additional manual refinement", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1 Throughout this paper, the term \"constraints\" is used in a broad sense to include both constraints (e.g., parallel) and dimensions (e.g., diameter).", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8614", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to capture the designer's intent.", + "bbox": [ + 89, + 90, + 300, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work builds upon these foundations by explicitly incorporating design intent as a post-training process. Instead of merely modeling the ground truth data, our method learns from constraint solver feedback, ensuring that generated sketches are geometrically plausible and structurally well-constrained. By doing so, we enable data-driven generation that aligns with design intent.", + "bbox": [ + 89, + 109, + 485, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Design Alignment Beyond the language domain, alignment techniques have been used to improve and align image generation. Lee et al. [21] propose fine-tuning diffusion-based text-to-image models using human feedback, significantly improving alignment between textual prompts and generated visuals. Similarly, ImageReward [43] uses a learned reward model trained on human preference data, which guides the diffusion model fine-tuning toward images preferred by human evaluators. Extending this idea, Black et al. [6] reframes image generation as a sequential RL task, introducing Denoising Diffusion Policy Optimization (DDPO) to optimize complex user-defined objectives directly for alignment without explicit human annotation. Few works have applied alignment techniques in the design domain. GearFormer [12] used differentiable sampling to enforce preferences in the solutions for mechanical configuration design problems, however, this approach does not work with rewards that require blackbox solvers in the loop. In concurrent work, e-SimFT [10] used preference data obtained from physics simulations to enhance the exploration of the Pareto front in a multi-preference setting, improving the solutions generated with GearFormer.", + "bbox": [ + 91, + 223, + 485, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning LLMs with RL Reinforcement Learning from Human Feedback (RLHF) has emerged as a cornerstone approach for aligning large language models (LLMs) with human preferences. In RLHF [4, 11, 26, 31], a learned reward model is usually trained to capture human preferences and is provides the learning signal that the policy is trained on. Other methods such as Direct Preference Optimization (DPO) [28] and Reinforced Self-Training (ReST) [36] use a simpler approach of optimizing the model directly from model generated data without the use of a learned reward model. While DPO learns from ranked pairs of generated model outputs from human annotators, ReST uses rejection sampling to remove incorrect generations and trains the model standard cross-entropy loss on the correct samples. In this paper, we refer to the approach of using fine-tuning on high return responses as Expert Iteration (ExIt). We broadly refer to algorithms that learn from ranked/filtered model outputs (such as DPO and Expert Iteration) as Preference Optimization (PO).", + "bbox": [ + 89, + 564, + 483, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recently, a large body of work has focused on the task of teaching LLMs to solve reasoning tasks with reinforcement learning [14, 15, 20, 36], such as math and cod", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing, which can be checked with rule-based systems. Specifically we are inspired by approaches which forego the use of a learned reward model and directly learn from verifiable rewards. We build off of several approaches that have shown success when applied to reasoning LLMs – these include Group Relative Policy Optimization (GRPO) [35], ReMax [22], and Reinforce Leave-One-Out (RLOO) [1]. Both GRPO and RLOO estimate the baseline via the average reward of multiple sampled outputs instead of learned value model but differ in how they apply the KL divergence penalty, advantage normalization, and PPO-style reward clipping. ReMax [22] also obviates the need for a learned value model but instead estimates the baseline from the argmax result (greedy sampling).", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we adapt the aforementioned post-training methods—DPO, Expert Iteration, RLOO, GRPO, ReMax, for use in aligning a constraint generation model using feedback from a constraint solver. In Section 4 we provide additional details on the ranking/filtering criteria for the Preference Optimization (PO) algorithms and the reward design for the RL algorithms (GRPO, RLOO, ReMax).", + "bbox": [ + 511, + 301, + 908, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem", + "text_level": 1, + "bbox": [ + 511, + 422, + 609, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sketch constraining is a fundamental component of parametric CAD modeling, where geometric relationships define the structure and behavior of the sketch. Applying constraints ensures stability and editability, allowing for parametric modifications that align with the design intent. Automating constraint generation requires producing a valid and efficient set of constraints that fully define a given sketch while avoiding unnecessary redundancy or conflicts.", + "bbox": [ + 511, + 446, + 906, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sketch constraining problem can be formulated as a sequence modeling task similar to natural language generation, where constraints are predicted autoregressively. Given the sketch geometry as input, the model generates a sequence of constraints in the order they will be applied. Tokens in the sequence represent either a constraint (e.g., coincident, parallel, perpendicular), a dimension (e.g., horizontal, vertical, radial), or a pointer to one of the input geometric entities [38].", + "bbox": [ + 511, + 568, + 908, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In parametric CAD, sketch geometry is modified using a constraint solver. The updated sketch respects any present constraints while moving the geometry to reflect changes to the dimension parameters. Unlike natural language, which is inherently sequential and follows flexible grammar rules, sketch constraints must adhere to strict geometric principles to ensure structural validity. A set of constraints may be incorrect for a variety of reasons: they can reference the wrong primitives for the constraint type, be redundant, specify inconsistent geometric relationships, or cause unexpected geometric distortions.", + "bbox": [ + 511, + 704, + 908, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formally define five conditions describing the state of a sketch after applying constraints. These conditions are", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8615", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/db5bcb354e67d8d494bbb56eb5029dd1a66a5904cf42954d457e02fc405b279b.jpg", + "image_caption": [ + "Figure 3. Comparing different outcomes when a designer updates a sketch parameter on the constrained sketch. First row: an underconstrained sketch only preserves a subset of the geometric relationships. Second row: the sketch is unstable, adding a coincident constraint flattens the geometry of the sketch. Third row: the sketch is over-constrained, causing the sketch to be uneditable. Fourth row: all geometric relationships are maintained after the parameter is updated." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "not mutually exclusive. A sketch may satisfy one or more conditions depending on the applied constraints. Figure 3 provides illustrative examples of each condition.", + "bbox": [ + 89, + 525, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Under-constrained (UC) A sketch containing primitives retains some unconstrained degrees of freedom, resulting in incomplete specification of their positions or dimensions.", + "bbox": [ + 89, + 575, + 483, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fully-constrained (FC) A sketch in which all primitives have their degrees of freedom completely determined, removing positional or dimensional ambiguities.", + "bbox": [ + 89, + 627, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Over-constrained (OC) A sketch primitives have more constraints applied than degrees of freedom, potentially leading to conflicts. Note some over-constrained sketches remain solvable if the constraints are consistent and do not conflict with each other [7].", + "bbox": [ + 89, + 678, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Not solvable A sketch that cannot achieve a valid solution due to contradictory or redundant constraints, leading to an impossible or conflicting geometry.", + "bbox": [ + 89, + 758, + 483, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stability We discretize the sketch plane into a grid and classify a sketch as unstable if the positions of primitives after constraint solving shift into different cells. The number of cells (bins) on each axis determines the sensitivity of this measurement.", + "bbox": [ + 89, + 809, + 483, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our objective is to align the model toward generating", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "constraint sets that yield fully-constrained sketches while minimizing cases of under-constrained, over-constrained, not solvable, or instability. This is a necessary prerequisite toward the ultimate goal of generating constraints that preserve the original sketch design intent when the designer varies the geometric parameters. A formal definition of FC, UC, and OC is described in [16, 17] and in the appendix.", + "bbox": [ + 511, + 90, + 905, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 210, + 604, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we outline the post-training techniques used for aligning a constraint generation model with feedback from a constraint solver. The approaches are grouped into three categories: supervised learning methods, preference-based optimization methods, and RL methods. Figure 4 illustrates the high-level workflow of this work.", + "bbox": [ + 511, + 237, + 905, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Constraint Solver", + "text_level": 1, + "bbox": [ + 511, + 338, + 686, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate generated constraints using the commercially available constraint solver in Autodesk Fusion. This solver is treated as a black box, from which we retrieve sketch conditions as defined in Section 3. A general algorithmic approach commonly used in constraint solvers is described in [8]. The solver outputs the fully-constrained status for each entity if the sketch is solvable. Otherwise it classifies the sketch as over-constrained or unsolvable. The solver adjusts geometry to resolve all constraints which we compare with the original sketch to determine if the constraints caused geometry distortion (instability). On average, it takes 0.1-0.2 seconds to per solve, although complex cases may take tens of seconds. Any sketch taking longer than two seconds to solve is automatically deemed unsolvable.", + "bbox": [ + 511, + 361, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Supervised Learning", + "text_level": 1, + "bbox": [ + 511, + 583, + 710, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We pre-train Vitruvion [33] as our base model using the same procedure described in their paper with a next-token prediction objective. Given an input sequence of geometric primitives, the model is trained to predict the next correct constraint or dimension based on the ground truth data. Further details about our implementation of the Vitruvion architecture and training procedure are provided in the appendix.", + "bbox": [ + 511, + 604, + 905, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the majority of the ground truth data contains under-constrained or over-constrained sketches, we additionally perform supervised fine-tuning (SFT). During SFT, the training data is limited to sketches verified by the constraint solver as solvable, fully-constrained, stable, and free of over-constrained conditions, ensuring the model explicitly learns from ideal examples.", + "bbox": [ + 511, + 727, + 905, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Preference-Based Optimization (PO)", + "text_level": 1, + "bbox": [ + 511, + 843, + 831, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Expert Iteration (ExIt) alternates between expert improvement and policy distillation. Following [15, 36],", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8616", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/44f4104c4891fa907efad26cb8bd86f2e9232912c65e20c4d71d3ef6017c8820.jpg", + "image_caption": [ + "Figure 4. Illustration of the proposed alignment workflow for constraint generation models. Left: A base constraint-generation model is first fine-tuned using supervised fine-tuning (SFT). Middle: Generated constraint sequences $C_i$ are evaluated using a constraint solver, which provides feedback on sketch stability and fully- and over-constrained statuses, forming the quality score vector $S_i$ . Right: Two groups of alignment methods leveraging solver feedback: preference-based optimization (PO) and reinforcement learning (RL). PO uses $S_i$ to construct training data, iteratively improving constraint prediction quality. RL methods assign per-constraint and per-sequence rewards to $C_i$ based on solver feedback and $S_i$ to incentivize the generation of constraints, making sketches stable and fully-constrained." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we use temperature sampling combined with rejection sampling to generate high-quality candidate constraint sequences. During the exploration step we initialize the policy model $\\pi_{\\theta_{t = 0}}$ from the SFT model and sample $K = 8$ candidate constraint sequences $\\tau$ at temperature $T = 1.0$ for each sketch query $q$ in the initial training set $\\mathcal{D}$ . A new training dataset $\\mathcal{D}_{i\\in N}^{*}$ is constructed by discarding sequences that are under-constrained, over-constrained, or unsolvable solutions. This process is repeated $N = 2$ times over the dataset, and the policy is trained using cross-entropy loss:", + "bbox": [ + 89, + 441, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(q, \\tau) \\sim \\mathcal {D} _ {i \\in N} ^ {*}} [ \\log \\pi_ {\\theta_ {t}} (\\tau | q) ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 607, + 482, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi_{\\theta_t}$ is updated after the distillation phase of every training round.", + "bbox": [ + 89, + 628, + 482, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Direct Preference Optimization (DPO) learns from pairwise preference data, approximating an implicit reward via a reparameterized Bradley-Terry model [28]. Similar to ExIt, we construct the training dataset by sampling $K = 8$ constraint sequence completions $\\tau$ for each sketch query $q$ from the policy model $\\pi_{\\theta_t}$ at temperature $T = 1.0$ . Pairs $(\\tau_w, \\tau_l)$ are ranked based on the differences in the percentage of fully-constrained curves between $\\tau_w$ (preferred sequence), and $\\tau_l$ (non-preferred sequence). The optimization objective is:", + "bbox": [ + 89, + 664, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {(q, \\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {w} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {w} | q)} - \\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {l} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {l} | q)}\\right) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 824, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyperparameter, and $\\sigma$ is the logistic function. This formulation ensures the learned policy $\\pi_{\\theta}$ aligns with", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the ranking preference of fully-constrained sketches while maintaining proximity to the reference policy $\\pi_{\\theta_r}$ .", + "bbox": [ + 511, + 441, + 906, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We initialize $\\pi_{\\theta_{t = 0}}$ from the SFT model and repeat the entire process $N = 2$ times, updating $\\pi_{\\theta_t}$ with the new policy after every training iteration. Additional hyperparameters and ranking details can be found in the appendix.", + "bbox": [ + 511, + 472, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Reinforcement Learning (RL)", + "text_level": 1, + "bbox": [ + 511, + 542, + 779, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4.1. Reward design", + "text_level": 1, + "bbox": [ + 511, + 564, + 663, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike natural language tasks, where human preferences are ambiguous and ill-defined, the stability and solvability of sketch constraints can be verified, making it compatible with RL methods that directly optimize for mechanically defined rewards without a learned preference model. We define the rewards used for RL as follows:", + "bbox": [ + 511, + 583, + 905, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Rewards for valid constraint sequence $\\tau$ :", + "bbox": [ + 513, + 676, + 795, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$r_{\\mathrm{curves}}(\\tau)$ : $\\%$ of fully-constrained curves over all curves, $r_{\\mathrm{points}}(\\tau)$ : $\\%$ of fully-constrained points over all points, $r_{\\mathrm{unstable}}$ : penalty for unstable sketches,", + "bbox": [ + 524, + 696, + 901, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Rewards for invalid constraint sequence:", + "bbox": [ + 513, + 758, + 795, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$r_{\\mathrm{NS}}$ : penalty for not solvable sketches,", + "bbox": [ + 526, + 779, + 785, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$r_{\\mathrm{OC}}$ : penalty for over-constrained sketches,", + "bbox": [ + 526, + 799, + 816, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$r_{\\mathbf{F}}$ : penalty for sketches resulting in other failures.", + "bbox": [ + 526, + 820, + 862, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall sequence-wise reward $R(\\tau)$ is the sum of $r_{\\mathrm{curves}}(\\tau), r_{\\mathrm{points}}(\\tau)$ , and conditionally $r_{\\mathrm{unstable}}$ for valid sketches, and either $r_{\\mathrm{NS}}, r_{\\mathrm{OC}}$ or $r_{\\mathrm{F}}$ for invalid sketches according to the failure mode.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8617", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We additionally define a constraint-wise penalty to provide granular feedback on cases where the constraint sequence causes sketches to be over-constrained or fully-constrained. A constraint solver iteratively attempts to add each generated constraint one-by-one, dropping any problematic constraints that caused the sketch to be over-constrained or not-solvable. In training, we add a constant of -1 loss penalty directly to the per-token log likelihood loss for the problematic constraints.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.2. RL fine-tuning formulation", + "text_level": 1, + "bbox": [ + 89, + 236, + 326, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We formulate constraint generation fine-tuning as follows; given a dataset of sketch queries $D = \\{q_i\\}_{i=1}^N$ and reward function $R(\\tau)$ using the constraint solver and reward design in Section 4.4.1, learn a policy $\\pi_{\\theta}(\\tau|q)$ that generates a sequence of constraints $\\tau$ for sketch $q$ , such that it maximizes the expected rewards $\\mathbb{E}_{q_i \\sim D, \\tau_i \\sim \\pi_{\\theta}(\\cdot | q_i)}[R(\\tau_i)]$ .", + "bbox": [ + 89, + 253, + 483, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.3. Policy gradient methods", + "text_level": 1, + "bbox": [ + 89, + 354, + 305, + 368 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For RLHF which uses a pre-trained policy, not all the complexity of RL algorithms is necessary. This allows the algorithms to be simplified and the number of learnable components to be reduced, contributing to performance improvement. We considered three policy gradient algorithms: ReMax [22], RLOO [1], and GRPO [35]. Unlike PPO [31], which treats each token generation as an action, these algorithms treat generation of a sequence as a single action and adopt a REINFORCE with baselines approach [39]. We apply these to optimize the constraint generation policy.", + "bbox": [ + 89, + 373, + 483, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ReMax [22] uses the rewards corresponding to sequences generated by a greedy (argmax) policy as a baseline to normalize the rewards of sequences sampled from the policy.", + "bbox": [ + 89, + 529, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Considering sequence $\\tau$ sampled from policy $\\pi_{\\theta}(\\tau |q)$ , sequence $\\tau^{*} = \\operatorname{argmax}_{\\tau}\\pi_{\\theta}(\\tau |q)$ greedily sampled by taking an argmax of the policy, and corresponding rewards $r$ and $r^{*}$ , the policy gradient objective for ReMax is:", + "bbox": [ + 89, + 575, + 483, + 637 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\tau \\sim \\pi} {\\mathbb {E}} \\left[ \\left(r - r ^ {*}\\right) \\nabla \\log \\pi_ {\\theta} (\\tau | q) \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 651, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "REINFORCE-Leave-One-Out (RLOO) [1] samples $G$ number of constraint sequences for every sketch query. The baseline for each sample is evaluated as the mean of the rewards for all other samples in the group.", + "bbox": [ + 89, + 686, + 483, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $G$ number of sequences $\\{\\tau_g\\}_{g = 1}^G$ sampled from policy $\\pi_{\\theta}(\\tau |q)$ for a given sketch query $q$ , the policy gradient objective of RLOO is:", + "bbox": [ + 89, + 747, + 483, + 792 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\{\\tau \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\frac {1}{G} \\sum_ {g = 1} ^ {G} \\left[ \\left(r _ {g} - \\operatorname {m e a n} \\left(\\left\\{r _ {i} \\right\\} _ {i \\neq g} ^ {G}\\right)\\right) \\nabla \\log \\pi_ {\\theta} \\left(\\tau_ {g} \\mid q _ {g}\\right) \\right] \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 814, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Group Relative Policy Optimization (GRPO) [35] uses group-based baseline estimation, like RLOO, but the mean", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is taken over all reward samples in the group. It uses a clipped policy optimization objective similarly to PPO [31], as well as a low-variance KL regularization term.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For $G$ number of sequences $\\{\\tau_g\\}_{g = 1}^G$ sampled from the reference policy $\\pi_{\\theta_{\\mathrm{r}}}(\\tau |q)$ for a given sketch query $q$ , letting $\\rho_{g} = \\frac{\\nabla\\pi_{\\theta}(\\tau_{g}|q)}{\\pi_{\\theta_{\\mathrm{r}}}(\\tau_{g}|q)}$ , the optimization objective of GRPO is:", + "bbox": [ + 511, + 135, + 903, + 188 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\{\\tau_ {g} \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\min \\left(\\rho_ {g} A _ {g}, \\operatorname {c l i p} \\left(\\rho_ {g}, 1 - \\epsilon , 1 + \\epsilon\\right) A _ {g}\\right) - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} | | \\pi_ {\\theta_ {\\mathrm {r}}}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 194, + 916, + 223 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} \\| \\pi_ {\\theta_ {\\mathrm {r}}}\\right) = \\frac {1}{\\rho_ {g}} + \\log \\rho_ {g} - 1, \\quad A _ {g} = \\frac {r _ {g} - \\operatorname* {m e a n} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 226, + 911, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon$ and $\\beta$ are hyper-parameters for the clipped policy optimization and KL regularization terms, respectively.", + "bbox": [ + 511, + 281, + 903, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ReMax and RLOO, we also added a small KL penalty term to the rewards to discourage divergence from the reference policy. GRPO applies group-normalization on the advantage, which we also applied in RLOO. For ReMax, we batch-normalized the advantage. For all algorithms, $\\pi_{\\theta}$ is initialized from the SFT model. Further implementation details of the algorithms can be found in the appendix.", + "bbox": [ + 511, + 311, + 905, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 430, + 643, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 513, + 455, + 607, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our models on SketchGraphs [32], a large-scale dataset of CAD sketches created in Onshape. SketchGraphs captures real-world parametric modeling workflows, providing geometry and constraint construction operations from actual design steps. However, the dataset was not originally designed for direct constraint inference; only $8.27\\%$ of its sketches are fully-constrained, making it imperfect for training the constraint generation model.", + "bbox": [ + 511, + 477, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For computational feasibility, we deduplicate and filter sketches, retaining only those with at most 16 geometric primitives and 64 constraints, yielding a dataset of 2.8 million unique sketches. Certain constraint types, such as symmetry, are excluded to simplify the learning task, ensuring the focus remains on constraints most relevant to engineering design. We also convert Onshape sketches into the Fusion sketch format to utilize the solver in Fusion. Additional details are provided in the appendix.", + "bbox": [ + 511, + 598, + 905, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Another challenge lies in how SketchGraphs positions primitives. Rather than being placed in valid, constraintsatisfying layouts, primitives often have arbitrary coordinates that do not reflect a solved state. We therefore preprocess the dataset using Fusion to resolve each sketch's primitives according to its constraints, ensuring that geometry and constraints match before training.", + "bbox": [ + 511, + 734, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 848, + 707, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we list results comparing the performance of each alignment method with respect to the five sketch condi", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8618", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4fc8240b5e47ffd1720e8ee0bab33703655e8ed51b962dd96783629da456c48.jpg", + "table_caption": [ + "Table 1. Sketch constraint generation results for Fully Constrained (FC), Under Constrained (UC), Over Constrained (OC), not solvable, and stability for the base model, SFT model, and aligned models. Results are computed over 8 samples per sketch with a temperature of 1.0. Numbers following $\\pm$ indicate the standard deviation." + ], + "table_footnote": [], + "table_body": "
Model% FC ↑% UC ↓% OC ↓% Not solvable ↓% Stable (bins=4) ↑
Vitruvion (base)8.87 ±0.0971.38 ±0.2016.83 ±0.173.05 ±0.0392.15 ±0.06
SFT34.24 ±0.0946.61 ±0.1515.30 ±0.083.85 ±0.0392.48 ±0.05
Iterative DPO64.91 ±0.1114.97 ±0.1312.47 ±0.097.64 ±0.0787.63 ±0.09
Expert Iteration71.70 ±0.1313.38 ±0.137.25 ±0.067.67 ±0.0785.50 ±0.10
ReMax79.84 ±0.0915.86 ±0.071.49 ±0.012.82 ±0.0275.77 ±0.05
RLOO93.05 ±0.033.55 ±0.022.15 ±0.011.25 ±0.0189.16 ±0.02
GRPO91.59 ±0.034.18 ±0.031.94 ±0.012.28 ±0.0288.28 ±0.03
", + "bbox": [ + 179, + 142, + 816, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tions described in Section 3. The base model is able to fully-constrain sketches only $8.87\\%$ of the time, consistent with the dataset distribution where only $8.27\\%$ of sketches are fully constrained. We find that RLOO and GRPO perform similarly, giving the best performance at fully-constraining sketches $93.05\\%$ and $91.59\\%$ of the time, respectively. They have the lowest indices of over-constrained or unsolvable results and maintain stability rates that are on par with other methods.", + "bbox": [ + 88, + 294, + 480, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Iterative DPO and ExIt significantly improve upon the base and SFT models but still fall short of the performance achieved by policy gradient-based RL methods. We attribute this gap to the online nature of policy gradient-based RL, which continuously refines the policy through feedback while actively exploring a broader range of solutions. In contrast, Iterative DPO and ExIt are offline methods and rely on predefined ranking and filtering signals to generate training data, which limits their ability to explore the solution space. The superior performance of online RL underscores its advantage in directly optimizing the shaped rewards from the constraint solver.", + "bbox": [ + 89, + 431, + 482, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 demonstrates the evaluation result of our methods in a few-shot inference setting, where the model has $K$ attempts to generate a fully-constrained and stable output. This scenario reflects real-world use cases where the", + "bbox": [ + 89, + 614, + 482, + 676 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/265f2da3f9858e40fad653ace59fb9346d9fc7652f08445447b755893b6b0a62.jpg", + "table_caption": [ + "Table 2. Sketch constraint generation results for Pass@1 and Pass@8 across the post-training algorithms. We define a successful result as fully-constrained, not over-constrained, solvable, and stable at 4 bins. Results are generated with temperature of 1.0." + ], + "table_footnote": [], + "table_body": "
ModelPass@1Pass@8
Vitruvion (base)8.5320.47
SFT33.3242.62
Iterative DPO59.3868.32
Expert Iteration64.0972.56
ReMax62.7465.89
RLOO83.5784.96
GRPO81.4983.42
", + "bbox": [ + 158, + 768, + 411, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "goal is to maximize the likelihood of producing an acceptable solution within a fixed inference budget of $K$ samples (Pass@ $K$ ). We find that while the RL-based methods still have the overall highest performance, increasing the number of samples $K$ has a comparatively small impact on performance compared to the other methods. Additional results and analysis on the impact of sampling parameters such as temperature and top- $p$ are available in the appendix.", + "bbox": [ + 511, + 294, + 906, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 425, + 696, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We randomly select sketches from the test set and show the results generated by different alignment methods in Figure 5. Curves are colored black when constrained and blue when not. Methods leveraging alignment techniques demonstrate a promising trend towards generating fully-constrained sketches, whereas the base Vitruvion and SFT models typically leave sketches under-constrained. However, across different alignment algorithms, we observe substantial variance in the degree of geometric distortion introduced by the aligned models.", + "bbox": [ + 511, + 446, + 906, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, columns A and B depict simple sketches mainly composed of horizontal and vertical lines, for which all solver-feedback methods consistently yield fully-constrained, stable results. In contrast, Column C presents a challenging sketch due to the absence of appropriate constraints for oblique lines. Achieving stability in this scenario typically requires many dimensions but few constraints; however, models are optimized to generate more constraints to align with parametric CAD design principles.", + "bbox": [ + 511, + 598, + 908, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sketches in columns D through H include arcs. Column E is particularly notable as all solver-feedback methods achieve a fully-constrained condition, yet only RLOO produces visually stable results. Further examination reveals that distortions induced by other models were still classified as stable due to our bin size ( $bins = 4$ ). Column G presents unique challenges with an isolated point, causing confusion in the model and demanding extensive use of non-horizontal/vertical constraints to make the sketch fully-constrained and stable.", + "bbox": [ + 511, + 734, + 908, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our results indicate that models more easily fully-", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8619", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0c5fa2a1e2d2718102b4c6d50fcd713bcc34b8cd20e678f92ea061468739bf1.jpg", + "image_caption": [ + "Figure 5. Visual comparison of solved sketches across the aforementioned baseline and post-training methods. Curves are colored black when fully-constrained and blue when not. Overlaid red lines are the original input sketch for reference." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 903, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "constrain sketches consisting primarily of horizontal and vertical lines without distortion. In contrast, sketches involving oblique lines increase the challenge of maintaining stability, and arcs further complicate achieving fully-constrained status. Among the tested methods, RLOO and GRPO exhibit the strongest overall performance in achieving fully-constrained status and geometric stability.", + "bbox": [ + 89, + 541, + 482, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Upon closer inspection of the generated constraints for the RL-based methods, we find that these models learn to \"reward hack\" by adding far more dimensions than constraints. This allows the model to avoid the instability penalty because dimensions cannot cause geometry distortion. In order to counteract this behavior, we added an additional reward penalty on the ratio of constraints versus dimensions which fixes the over-dimensioning problem at the cost of reducing stability and fully-constrained status. Additional details are provided in the appendix.", + "bbox": [ + 89, + 652, + 482, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further evaluate the efficacy of each method, we conduct a forced-choice perceptual study with professional CAD designers. We show them paired images of the same sketch with different constraints applied from each method, to understand which set of constraints they prefer. We find that the preference-based optimization methods (ExIT,", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "DPO) and RLOO with the over-dimensioning penalty consistently outperform the SFT and vanilla RLOO results. Results of this study are provided in the appendix.", + "bbox": [ + 511, + 541, + 906, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 603, + 635, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments focus on sketches of moderate complexity, with a maximum of 16 geometric primitives and 64 constraints. Evaluating performance on more complex sketches requires further exploration. Additionally, our alignment approach only uses feedback signals from a constraint solver to represent design intent. Incorporating subjective human preferences and explicit design intent into alignment objectives remains a promising area for future work.", + "bbox": [ + 511, + 630, + 906, + 751 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 768, + 633, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrated the adaption of alignment strategies from language modeling to preserve design intent in parametric CAD sketches. By using feedback from a constraint solver as a learning signal, we show the feasibility and value of alignment in parametric CAD tasks. In doing so, we pave the way for future AI-assisted design tools that incorporate design alignment.", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8620", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank William McCarthy for his valuable contribution to the human evaluation experiments, including setup, execution, and analysis. His efforts were essential in validating the effectiveness of our proposed methods from a user-centric perspective.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arash Ahmadian, Chris Cremer, Matthias Galle, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, and Sara Hooker. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. arXiv preprint arXiv:2402.14740, 2024. 2, 3, 6", + "[2] Thomas W. Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In *Neural Information Processing Systems*, 2017. 2", + "[3] Autodesk. Sketches in Fusion, 2014. 2", + "[4] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3", + "[5] Bernhard Bettig and Christoph M. Hoffmann. Geometric constraint solving in parametric computer-aided design. Journal of Computing and Information Science in Engineering, 11(2):021001, 2011. 2", + "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3", + "[7] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-Aided Design, 27(6):487-501, 1995. 2, 4", + "[8] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-aided design, 27(6):487-501, 1995. 4", + "[9] Wei Chen, Kevin Chiu, and Mark D Fuge. Airfoil design parameterization and optimization using bezier generative adversarial networks. AIAA journal, 58(11):4723-4735, 2020. 1", + "[10] Hyunmin Cheong, Mohammad Mehdi Ataei, Amir Hosein Khasahmadi, and Pradeep Kumar Jayaraman. e-simft: Alignment of generative models with simulation feedback for pareto-front design exploration. arXiv preprint arXiv:2502.02628, 2025.3", + "[11] Paul F. Christiano, Jan Leike, Tom B. Brown, et al. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 1, 3", + "[12] Yasaman Etesam, Hyunmin Cheong, Mohammad Mehdi Ataei, and Pradeep Kumar Jayaraman. Deep generative model for mechanical system configuration design. arXiv preprint arXiv:2409.06016, 2024. 3", + "[13] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. NeurIPS, 34:5885-5897, 2021. 2", + "[14] Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code" + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "llms in execution feedback with reinforcement learning. arXiv preprint arXiv:2410.02089, 2024.3", + "[15] Alex Havrilla, Yuqing Du, Sharath Chandra Rarparthy, Christoforos Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, Sainbayar Sukhbaatar, and Roberta Raileanu. Teaching large language models to reason with reinforcement learning. arXiv preprint arXiv:2403.04642, 2024. 3, 4", + "[16] Christoph M Hoffmann and Robert Joan-Arinyo. Symbolic constraints in constructive geometric constraint solving. Journal of Symbolic Computation, 23(2-3):287-299, 1997. 4", + "[17] Christoph M Hoffmann and Robert Joan-Arinyo. A brief on constraint solving. Computer-Aided Design and Applications, 2(5):655-663, 2005. 4", + "[18] Benjamin T Jones, Felix Hahnlein, Zihan Zhang, Maaz Ahmad, Vladimir Kim, and Adriana Schulz. A solver-aided hierarchical language for ltm-driven cad design. arXiv preprint arXiv:2502.09819, 2025. 2", + "[19] Ahmet Serdar Karadeniz, Dimitrios Mallis, Nesryne Mejri, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Davinci: A single-stage architecture for constrained cad sketch inference. arXiv preprint arXiv:2410.22857, 2024. 2", + "[20] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\\”ulu 3: Pushing frontiers in open language model posttraining. arXiv preprint arXiv:2411.15124, 2024. 3", + "[21] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3", + "[22] Ziniu Li, Tian Xu, Yushun Zhang, Zhihang Lin, Yang Yu, Ruoyu Sun, and Zhi-Quan Luo. Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models. In ICML, 2024. 2, 3, 6", + "[23] Dave Martin. What is design intent?, 2023. Accessed: January 19, 2023. 2", + "[24] Nelson Nauata, Kai-Hung Chang, Chin-Yi Cheng, Greg Mori, and Yasutaka Furukawa. House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In ECCV, pages 162-177. Springer, 2020. 1", + "[25] Jeffrey Otey, Pedro Company, Manuel Contrero, and Jorge D. Camba. Revisiting the design intent concept in the context of mechanical cad education. Computer-Aided Design and Applications, 15(1):47-60, 2018. 2", + "[26] Long Ouyang, Jeff Wu, Xu Jiang, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 3", + "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. NeurIPS, 34:5077-5088, 2021. 2", + "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8621", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "preference optimization: Your language model is secretly a reward model. NeurIPS, 36:53728-53741, 2023. 1, 2, 3, 5", + "[29] Lyle Regenwetter, Amin Heyrani Nobari, and Faziz Ahmed. Deep generative models in engineering design: A review. Journal of Mechanical Design, 144(7):071704, 2022. 1", + "[30] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 3rd edition, 2010. 1", + "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 6", + "[32] Ari Seff, Yaniv Ovadia, Wenda Zhou, and Ryan P. Adams. SketchGraphs: A large-scale dataset for modeling relational geometry in computer-aided design. In ICML 2020 Workshop on Object-Oriented Learning, 2020. 2, 6", + "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In ICLR, 2021. 2, 4", + "[34] Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa. Housediffusion: Vector floorplan generation via a diffusion model with discrete and continuous denoising. In CVPR, pages 5466-5475, 2023. 1", + "[35] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 2, 3, 6", + "[36] Avi Singh, John D Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alexander A Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Fathy Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura A Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research, 2024. 2, 3, 4", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2", + "[38] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, page 2692-2700, Cambridge, MA, USA, 2015. MIT Press. 2, 3", + "[39] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 6", + "[40] Karl DD Willis, Pradeep Kumar Jayaraman, Joseph G Lambourne, Hang Chu, and Yewen Pu. Engineering sketch generation for computer-aided design. In CVPRW, pages 2105–2114, 2021. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM TOG, 40(4), 2021. 1, 2", + "[42] Sifan Wu, Amir Hosein Khasahmadi, Mor Katz, Pradeep Kumar Jayaraman, Yewen Pu, Karl Willis, and Bang Liu. Cadvlm: Bridging language and vision in the generation of parametric cad sketches. In ECCV, pages 368-384. Springer, 2024. 2", + "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3", + "[44] Xiang Xu, Pradeep Kumar Jayaraman, Joseph G Lambourne, Karl DD Willis, and Yasutaka Furukawa. Hierarchical neural coding for controllable cad model generation. In ICML, pages 38443-38461, 2023. 1", + "[45] Yuezhi Yang and Hao Pan. Discovering design concepts for cad sketches. NeurIPS, 35:28803-28814, 2022. 2" + ], + "bbox": [ + 514, + 90, + 903, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8622", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_model.json b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d82fd9988ab2acb0437f9b130987cb20fa084323 --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.857, + 0.153 + ], + "angle": 0, + "content": "Aligning Constraint Generation with Design Intent in Parametric CAD" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.181, + 0.386, + 0.218 + ], + "angle": 0, + "content": "Evan Casey Tianyu Zhang Joseph George Lambourne" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.181, + 0.886, + 0.217 + ], + "angle": 0, + "content": "Shu Ishida John Roger Thompson Amir Khasahmadi Pradeep Kumar Jayaraman Karl D.D. Willis" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.234, + 0.575, + 0.251 + ], + "angle": 0, + "content": "Autodesk Research" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.27, + 0.37, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.408, + 0.27, + 0.907, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.481, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Left: a constraint solver is used to score model generated constraints A, B, C \\(\\sim\\) \\(\\pi_{\\mathrm{base}}\\) (and \\(A', B', C' \\sim \\pi_{\\mathrm{aligned}}\\)). Starting with the base model \\(\\pi_{\\mathrm{base}}\\), we post-train an aligned model \\(\\pi_{\\mathrm{aligned}}\\) from this feedback. Right: Blue lines show original primitives and gray lines show geometric distortion when dimensions vary. The aligned model \\(\\pi_{\\mathrm{aligned}}\\) produces fully-constrained sketches that preserve relative geometric relationships, whereas the base model \\(\\pi_{\\mathrm{base}}\\) produces under-constrained sketches that may distort the geometry in unintended ways." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.55, + 0.327, + 0.566 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We adapt alignment techniques from reasoning LLMs to the task of generating engineering sketch constraints found in computer-aided design (CAD) models. Engineering sketches consist of geometric primitives (e.g. points, lines) connected by constraints (e.g. perpendicular, tangent) that define the relationships between them. For a design to be easily editable, the constraints must effectively capture design intent, ensuring the geometry updates predictably when parameters change. Although current approaches can generate CAD designs, an open challenge remains to align model outputs with design intent, we label this problem 'design alignment'. A critical first step towards aligning generative CAD models is to generate constraints which fully-constrain all geometric primitives, without over-constraining or distorting sketch geometry. Using alignment techniques to train an existing constraint generation model with feedback from a constraint solver, we are able to fully-constrain \\(93\\%\\) of sketches compared to \\(34\\%\\) when using a naive supervised fine-tuning (SFT) baseline and only \\(8.9\\%\\) without SFT. Our approach can be applied to any existing constraint generation model and sets the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.907, + 0.613 + ], + "angle": 0, + "content": "stage for further research bridging alignment strategies between the language and design domains. Additional results can be found at https://autodeskailab.github.io/aligning-constraint-generation/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.632, + 0.646, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.658, + 0.907, + 0.884 + ], + "angle": 0, + "content": "A central challenge in artificial intelligence (AI) is alignment: ensuring that AI systems produce outputs that adhere to human goals and expectations [11, 26, 30]. Although alignment of language models has been researched extensively [26, 28, 35], the application of alignment techniques to parametric design problems has yet to be studied. The use of AI in this discipline covers a broad range of areas, ranging from floor-plan layout [24, 34], to engineering design problems [9, 29, 41], to 3D generation [44]. Design problems are often visual in nature and incorporate other functional requirements, making them unique when compared with language model alignment. In this paper, we establish the problem of design alignment and demonstrate how this expansive problem can be made tractable by adapting techniques from the alignment literature into a new context." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In language models, alignment is achieved by incorpo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8613" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.254, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 2. An illustration of design intent in CAD modeling. The bottom sketch maintains symmetry after modifying a dimension due to properly applied constraints, while the top sketch, lacking adequate constraints, becomes asymmetrical and distorted." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.483, + 0.534 + ], + "angle": 0, + "content": "rating feedback to generate coherent, contextually appropriate responses. Similarly, with parametric CAD modeling, AI tools must be aligned with a designer's intent by maintaining the underlying structural relationships to produce outputs that are both meaningful and functional. Otey et al. [25] define design intent as \"a CAD model's anticipated behavior when altered,\" while Martin [23] characterizes it as \"relationships between objects, so that a change to one propagates automatically to others.\" This means that modifications of a design by an AI system should yield outcomes where the established design relationships remain intact. To that end, we define design alignment as the application of generative modeling alignment techniques to produce outcomes that maintain design intent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.746 + ], + "angle": 0, + "content": "The realization of AI systems that observe and maintain design intent has broad implications for the manufacturing and construction industries. Almost every manufactured object or structure begins as a CAD model. At the core of parametric CAD modeling are 2D engineering sketches, which can be extruded or revolved to generate 3D models. Engineering sketches are composed of geometric primitives, such as points, lines, and circles, that are organized using constraints and dimensions [7]. These constraints1 define geometric rules, including equality, perpendicularity, and radial or linear dimensions, which collectively shape the final layout of the sketch. When applied correctly, they enable efficient modifications while preserving the original design intent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Figure 2 illustrates the impact of constraint quality: a poorly constrained sketch loses symmetry when a dimension is changed, whereas a well-constrained sketch preserves its intended relationships. This underscores the important role of constraints in maintaining design intent, and the need for AI systems that can align with this intent encoded in designs. We focus on the problem of sketch constraint generation [32] to demonstrate the adaption of align-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "ment techniques to a design problem. Using an existing sketch constraint generation model Vitruvion [33], we align the model with algorithms that learn from feedback (Direct Preference Optimization [28], Expert Iteration [2, 36], RLOO [1], ReMax [22] and Group Relative Policy Optimization [35]) using the sketch constraint solver in Autodesk Fusion [3] as the learning signal. We optimize the models to remove all degrees of freedom in the sketches to become 'fully-constrained' [5], without causing sketches to be distorted, over-constrained or unsolvable. We further define these conditions in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.333 + ], + "angle": 0, + "content": "To the best of our knowledge, this is the first instance of alignment methods being successfully applied to a parametric CAD design task; representing an important step forward for AI-assisted design tools. We present the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.339, + 0.904, + 0.4 + ], + "angle": 0, + "content": "- We establish the problem of design alignment, in the context of generative CAD models, as a critical component of AI-assisted CAD tools. We focus on the necessary first step of alignment for engineering sketches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.905, + 0.475 + ], + "angle": 0, + "content": "- We introduce a post-training strategy for a sketch constraint generation model using feedback from a sketch constraint solver. We define novel metrics and reward functions that directly optimize a base model for improved alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.905, + 0.581 + ], + "angle": 0, + "content": "- We conduct extensive experiments and demonstrate alignment techniques that fully-constrain \\(93\\%\\) of sketches compared to \\(34\\%\\) when using a naive supervised fine-tuning (SFT) baseline and only \\(8.9\\%\\) without alignment. We posit that our approach is broadly applicable to other design tasks that require compilation of elements with rule-based algorithms." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.339, + 0.905, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.597, + 0.655, + 0.614 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Engineering Sketches Engineering sketches form the 2D basis for 3D CAD models used to design mechanical parts for manufacturing. The availability of engineering sketch datasets [13, 32, 41] has enabled the development of generative models [13, 27, 33, 40] that can predict sketch geometry and/or the underlying constraints and dimensions that encode design intent. These Transformer-based [37] approaches create geometry by autoregressively generating tokens representing points and curves, then add constraints by referencing this geometry using Pointer Networks [38]. More recent approaches leverage image-based guidance [19, 42] or large language models (LLM) [18] in the constraint prediction task. Yang and Pan [45] learn to group together recurring patterns of geometric and constraint entities within a sketch, effectively discovering latent design concepts. However, none of these approaches explicitly optimize for preserving design intent – as a result, generated sketches may require additional manual refinement" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1 Throughout this paper, the term \"constraints\" is used in a broad sense to include both constraints (e.g., parallel) and dimensions (e.g., diameter)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8614" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.302, + 0.108 + ], + "angle": 0, + "content": "to capture the designer's intent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.486, + 0.218 + ], + "angle": 0, + "content": "Our work builds upon these foundations by explicitly incorporating design intent as a post-training process. Instead of merely modeling the ground truth data, our method learns from constraint solver feedback, ensuring that generated sketches are geometrically plausible and structurally well-constrained. By doing so, we enable data-driven generation that aligns with design intent." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.224, + 0.486, + 0.558 + ], + "angle": 0, + "content": "Design Alignment Beyond the language domain, alignment techniques have been used to improve and align image generation. Lee et al. [21] propose fine-tuning diffusion-based text-to-image models using human feedback, significantly improving alignment between textual prompts and generated visuals. Similarly, ImageReward [43] uses a learned reward model trained on human preference data, which guides the diffusion model fine-tuning toward images preferred by human evaluators. Extending this idea, Black et al. [6] reframes image generation as a sequential RL task, introducing Denoising Diffusion Policy Optimization (DDPO) to optimize complex user-defined objectives directly for alignment without explicit human annotation. Few works have applied alignment techniques in the design domain. GearFormer [12] used differentiable sampling to enforce preferences in the solutions for mechanical configuration design problems, however, this approach does not work with rewards that require blackbox solvers in the loop. In concurrent work, e-SimFT [10] used preference data obtained from physics simulations to enhance the exploration of the Pareto front in a multi-preference setting, improving the solutions generated with GearFormer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.485, + 0.853 + ], + "angle": 0, + "content": "Fine-tuning LLMs with RL Reinforcement Learning from Human Feedback (RLHF) has emerged as a cornerstone approach for aligning large language models (LLMs) with human preferences. In RLHF [4, 11, 26, 31], a learned reward model is usually trained to capture human preferences and is provides the learning signal that the policy is trained on. Other methods such as Direct Preference Optimization (DPO) [28] and Reinforced Self-Training (ReST) [36] use a simpler approach of optimizing the model directly from model generated data without the use of a learned reward model. While DPO learns from ranked pairs of generated model outputs from human annotators, ReST uses rejection sampling to remove incorrect generations and trains the model standard cross-entropy loss on the correct samples. In this paper, we refer to the approach of using fine-tuning on high return responses as Expert Iteration (ExIt). We broadly refer to algorithms that learn from ranked/filtered model outputs (such as DPO and Expert Iteration) as Preference Optimization (PO)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "More recently, a large body of work has focused on the task of teaching LLMs to solve reasoning tasks with reinforcement learning [14, 15, 20, 36], such as math and cod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "ing, which can be checked with rule-based systems. Specifically we are inspired by approaches which forego the use of a learned reward model and directly learn from verifiable rewards. We build off of several approaches that have shown success when applied to reasoning LLMs – these include Group Relative Policy Optimization (GRPO) [35], ReMax [22], and Reinforce Leave-One-Out (RLOO) [1]. Both GRPO and RLOO estimate the baseline via the average reward of multiple sampled outputs instead of learned value model but differ in how they apply the KL divergence penalty, advantage normalization, and PPO-style reward clipping. ReMax [22] also obviates the need for a learned value model but instead estimates the baseline from the argmax result (greedy sampling)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.41 + ], + "angle": 0, + "content": "In this paper, we adapt the aforementioned post-training methods—DPO, Expert Iteration, RLOO, GRPO, ReMax, for use in aligning a constraint generation model using feedback from a constraint solver. In Section 4 we provide additional details on the ranking/filtering criteria for the Preference Optimization (PO) algorithms and the reward design for the RL algorithms (GRPO, RLOO, ReMax)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.611, + 0.438 + ], + "angle": 0, + "content": "3. Problem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Sketch constraining is a fundamental component of parametric CAD modeling, where geometric relationships define the structure and behavior of the sketch. Applying constraints ensures stability and editability, allowing for parametric modifications that align with the design intent. Automating constraint generation requires producing a valid and efficient set of constraints that fully define a given sketch while avoiding unnecessary redundancy or conflicts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.704 + ], + "angle": 0, + "content": "The sketch constraining problem can be formulated as a sequence modeling task similar to natural language generation, where constraints are predicted autoregressively. Given the sketch geometry as input, the model generates a sequence of constraints in the order they will be applied. Tokens in the sequence represent either a constraint (e.g., coincident, parallel, perpendicular), a dimension (e.g., horizontal, vertical, radial), or a pointer to one of the input geometric entities [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.87 + ], + "angle": 0, + "content": "In parametric CAD, sketch geometry is modified using a constraint solver. The updated sketch respects any present constraints while moving the geometry to reflect changes to the dimension parameters. Unlike natural language, which is inherently sequential and follows flexible grammar rules, sketch constraints must adhere to strict geometric principles to ensure structural validity. A set of constraints may be incorrect for a variety of reasons: they can reference the wrong primitives for the constraint type, be redundant, specify inconsistent geometric relationships, or cause unexpected geometric distortions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We formally define five conditions describing the state of a sketch after applying constraints. These conditions are" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8615" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Figure 3. Comparing different outcomes when a designer updates a sketch parameter on the constrained sketch. First row: an underconstrained sketch only preserves a subset of the geometric relationships. Second row: the sketch is unstable, adding a coincident constraint flattens the geometry of the sketch. Third row: the sketch is over-constrained, causing the sketch to be uneditable. Fourth row: all geometric relationships are maintained after the parameter is updated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.573 + ], + "angle": 0, + "content": "not mutually exclusive. A sketch may satisfy one or more conditions depending on the applied constraints. Figure 3 provides illustrative examples of each condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Under-constrained (UC) A sketch containing primitives retains some unconstrained degrees of freedom, resulting in incomplete specification of their positions or dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Fully-constrained (FC) A sketch in which all primitives have their degrees of freedom completely determined, removing positional or dimensional ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Over-constrained (OC) A sketch primitives have more constraints applied than degrees of freedom, potentially leading to conflicts. Note some over-constrained sketches remain solvable if the constraints are consistent and do not conflict with each other [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.484, + 0.806 + ], + "angle": 0, + "content": "Not solvable A sketch that cannot achieve a valid solution due to contradictory or redundant constraints, leading to an impossible or conflicting geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Stability We discretize the sketch plane into a grid and classify a sketch as unstable if the positions of primitives after constraint solving shift into different cells. The number of cells (bins) on each axis determines the sensitivity of this measurement." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our objective is to align the model toward generating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.199 + ], + "angle": 0, + "content": "constraint sets that yield fully-constrained sketches while minimizing cases of under-constrained, over-constrained, not solvable, or instability. This is a necessary prerequisite toward the ultimate goal of generating constraints that preserve the original sketch design intent when the designer varies the geometric parameters. A formal definition of FC, UC, and OC is described in [16, 17] and in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.212, + 0.605, + 0.227 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.906, + 0.329 + ], + "angle": 0, + "content": "In this section, we outline the post-training techniques used for aligning a constraint generation model with feedback from a constraint solver. The approaches are grouped into three categories: supervised learning methods, preference-based optimization methods, and RL methods. Figure 4 illustrates the high-level workflow of this work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.687, + 0.353 + ], + "angle": 0, + "content": "4.1. Constraint Solver" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.907, + 0.573 + ], + "angle": 0, + "content": "We evaluate generated constraints using the commercially available constraint solver in Autodesk Fusion. This solver is treated as a black box, from which we retrieve sketch conditions as defined in Section 3. A general algorithmic approach commonly used in constraint solvers is described in [8]. The solver outputs the fully-constrained status for each entity if the sketch is solvable. Otherwise it classifies the sketch as over-constrained or unsolvable. The solver adjusts geometry to resolve all constraints which we compare with the original sketch to determine if the constraints caused geometry distortion (instability). On average, it takes 0.1-0.2 seconds to per solve, although complex cases may take tens of seconds. Any sketch taking longer than two seconds to solve is automatically deemed unsolvable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.584, + 0.712, + 0.6 + ], + "angle": 0, + "content": "4.2. Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.727 + ], + "angle": 0, + "content": "We pre-train Vitruvion [33] as our base model using the same procedure described in their paper with a next-token prediction objective. Given an input sequence of geometric primitives, the model is trained to predict the next correct constraint or dimension based on the ground truth data. Further details about our implementation of the Vitruvion architecture and training procedure are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.833 + ], + "angle": 0, + "content": "Since the majority of the ground truth data contains under-constrained or over-constrained sketches, we additionally perform supervised fine-tuning (SFT). During SFT, the training data is limited to sketches verified by the constraint solver as solvable, fully-constrained, stable, and free of over-constrained conditions, ensuring the model explicitly learns from ideal examples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.844, + 0.833, + 0.86 + ], + "angle": 0, + "content": "4.3. Preference-Based Optimization (PO)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Expert Iteration (ExIt) alternates between expert improvement and policy distillation. Following [15, 36]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8616" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.345, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Illustration of the proposed alignment workflow for constraint generation models. Left: A base constraint-generation model is first fine-tuned using supervised fine-tuning (SFT). Middle: Generated constraint sequences \\( C_i \\) are evaluated using a constraint solver, which provides feedback on sketch stability and fully- and over-constrained statuses, forming the quality score vector \\( S_i \\). Right: Two groups of alignment methods leveraging solver feedback: preference-based optimization (PO) and reinforcement learning (RL). PO uses \\( S_i \\) to construct training data, iteratively improving constraint prediction quality. RL methods assign per-constraint and per-sequence rewards to \\( C_i \\) based on solver feedback and \\( S_i \\) to incentivize the generation of constraints, making sketches stable and fully-constrained." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.595 + ], + "angle": 0, + "content": "we use temperature sampling combined with rejection sampling to generate high-quality candidate constraint sequences. During the exploration step we initialize the policy model \\(\\pi_{\\theta_{t = 0}}\\) from the SFT model and sample \\(K = 8\\) candidate constraint sequences \\(\\tau\\) at temperature \\(T = 1.0\\) for each sketch query \\(q\\) in the initial training set \\(\\mathcal{D}\\). A new training dataset \\(\\mathcal{D}_{i\\in N}^{*}\\) is constructed by discarding sequences that are under-constrained, over-constrained, or unsolvable solutions. This process is repeated \\(N = 2\\) times over the dataset, and the policy is trained using cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.608, + 0.483, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(q, \\tau) \\sim \\mathcal {D} _ {i \\in N} ^ {*}} [ \\log \\pi_ {\\theta_ {t}} (\\tau | q) ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.661 + ], + "angle": 0, + "content": "where \\(\\pi_{\\theta_t}\\) is updated after the distillation phase of every training round." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.817 + ], + "angle": 0, + "content": "Direct Preference Optimization (DPO) learns from pairwise preference data, approximating an implicit reward via a reparameterized Bradley-Terry model [28]. Similar to ExIt, we construct the training dataset by sampling \\( K = 8 \\) constraint sequence completions \\( \\tau \\) for each sketch query \\( q \\) from the policy model \\( \\pi_{\\theta_t} \\) at temperature \\( T = 1.0 \\). Pairs \\( (\\tau_w, \\tau_l) \\) are ranked based on the differences in the percentage of fully-constrained curves between \\( \\tau_w \\) (preferred sequence), and \\( \\tau_l \\) (non-preferred sequence). The optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.825, + 0.484, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {(q, \\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {w} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {w} | q)} - \\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {l} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {l} | q)}\\right) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyperparameter, and \\(\\sigma\\) is the logistic function. This formulation ensures the learned policy \\(\\pi_{\\theta}\\) aligns with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.907, + 0.473 + ], + "angle": 0, + "content": "the ranking preference of fully-constrained sketches while maintaining proximity to the reference policy \\(\\pi_{\\theta_r}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.534 + ], + "angle": 0, + "content": "We initialize \\(\\pi_{\\theta_{t = 0}}\\) from the SFT model and repeat the entire process \\(N = 2\\) times, updating \\(\\pi_{\\theta_t}\\) with the new policy after every training iteration. Additional hyperparameters and ranking details can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.78, + 0.559 + ], + "angle": 0, + "content": "4.4. Reinforcement Learning (RL)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.665, + 0.58 + ], + "angle": 0, + "content": "4.4.1. Reward design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Unlike natural language tasks, where human preferences are ambiguous and ill-defined, the stability and solvability of sketch constraints can be verified, making it compatible with RL methods that directly optimize for mechanically defined rewards without a learned preference model. We define the rewards used for RL as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.678, + 0.797, + 0.693 + ], + "angle": 0, + "content": "- Rewards for valid constraint sequence \\(\\tau\\):" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.697, + 0.903, + 0.755 + ], + "angle": 0, + "content": "\\(r_{\\mathrm{curves}}(\\tau)\\) : \\(\\%\\) of fully-constrained curves over all curves, \\(r_{\\mathrm{points}}(\\tau)\\) : \\(\\%\\) of fully-constrained points over all points, \\(r_{\\mathrm{unstable}}\\) : penalty for unstable sketches," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.759, + 0.797, + 0.774 + ], + "angle": 0, + "content": "- Rewards for invalid constraint sequence:" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.78, + 0.786, + 0.795 + ], + "angle": 0, + "content": "\\(r_{\\mathrm{NS}}\\) : penalty for not solvable sketches," + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.8, + 0.818, + 0.815 + ], + "angle": 0, + "content": "\\(r_{\\mathrm{OC}}\\) : penalty for over-constrained sketches," + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.821, + 0.863, + 0.836 + ], + "angle": 0, + "content": "\\(r_{\\mathbf{F}}\\) : penalty for sketches resulting in other failures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The overall sequence-wise reward \\( R(\\tau) \\) is the sum of \\( r_{\\mathrm{curves}}(\\tau), r_{\\mathrm{points}}(\\tau) \\), and conditionally \\( r_{\\mathrm{unstable}} \\) for valid sketches, and either \\( r_{\\mathrm{NS}}, r_{\\mathrm{OC}} \\) or \\( r_{\\mathrm{F}} \\) for invalid sketches according to the failure mode." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8617" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "We additionally define a constraint-wise penalty to provide granular feedback on cases where the constraint sequence causes sketches to be over-constrained or fully-constrained. A constraint solver iteratively attempts to add each generated constraint one-by-one, dropping any problematic constraints that caused the sketch to be over-constrained or not-solvable. In training, we add a constant of -1 loss penalty directly to the per-token log likelihood loss for the problematic constraints." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.237, + 0.328, + 0.251 + ], + "angle": 0, + "content": "4.4.2. RL fine-tuning formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.484, + 0.349 + ], + "angle": 0, + "content": "We formulate constraint generation fine-tuning as follows; given a dataset of sketch queries \\( D = \\{q_i\\}_{i=1}^N \\) and reward function \\( R(\\tau) \\) using the constraint solver and reward design in Section 4.4.1, learn a policy \\( \\pi_{\\theta}(\\tau|q) \\) that generates a sequence of constraints \\( \\tau \\) for sketch \\( q \\), such that it maximizes the expected rewards \\( \\mathbb{E}_{q_i \\sim D, \\tau_i \\sim \\pi_{\\theta}(\\cdot | q_i)}[R(\\tau_i)] \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.306, + 0.369 + ], + "angle": 0, + "content": "4.4.3. Policy gradient methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.526 + ], + "angle": 0, + "content": "For RLHF which uses a pre-trained policy, not all the complexity of RL algorithms is necessary. This allows the algorithms to be simplified and the number of learnable components to be reduced, contributing to performance improvement. We considered three policy gradient algorithms: ReMax [22], RLOO [1], and GRPO [35]. Unlike PPO [31], which treats each token generation as an action, these algorithms treat generation of a sequence as a single action and adopt a REINFORCE with baselines approach [39]. We apply these to optimize the constraint generation policy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.576 + ], + "angle": 0, + "content": "ReMax [22] uses the rewards corresponding to sequences generated by a greedy (argmax) policy as a baseline to normalize the rewards of sequences sampled from the policy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.638 + ], + "angle": 0, + "content": "Considering sequence \\(\\tau\\) sampled from policy \\(\\pi_{\\theta}(\\tau |q)\\), sequence \\(\\tau^{*} = \\operatorname{argmax}_{\\tau}\\pi_{\\theta}(\\tau |q)\\) greedily sampled by taking an argmax of the policy, and corresponding rewards \\(r\\) and \\(r^{*}\\), the policy gradient objective for ReMax is:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.652, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\underset {\\tau \\sim \\pi} {\\mathbb {E}} \\left[ \\left(r - r ^ {*}\\right) \\nabla \\log \\pi_ {\\theta} (\\tau | q) \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.747 + ], + "angle": 0, + "content": "REINFORCE-Leave-One-Out (RLOO) [1] samples \\(G\\) number of constraint sequences for every sketch query. The baseline for each sample is evaluated as the mean of the rewards for all other samples in the group." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.484, + 0.793 + ], + "angle": 0, + "content": "For \\(G\\) number of sequences \\(\\{\\tau_g\\}_{g = 1}^G\\) sampled from policy \\(\\pi_{\\theta}(\\tau |q)\\) for a given sketch query \\(q\\), the policy gradient objective of RLOO is:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.815, + 0.483, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\underset {\\{\\tau \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\frac {1}{G} \\sum_ {g = 1} ^ {G} \\left[ \\left(r _ {g} - \\operatorname {m e a n} \\left(\\left\\{r _ {i} \\right\\} _ {i \\neq g} ^ {G}\\right)\\right) \\nabla \\log \\pi_ {\\theta} \\left(\\tau_ {g} \\mid q _ {g}\\right) \\right] \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Group Relative Policy Optimization (GRPO) [35] uses group-based baseline estimation, like RLOO, but the mean" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "is taken over all reward samples in the group. It uses a clipped policy optimization objective similarly to PPO [31], as well as a low-variance KL regularization term." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "For \\(G\\) number of sequences \\(\\{\\tau_g\\}_{g = 1}^G\\) sampled from the reference policy \\(\\pi_{\\theta_{\\mathrm{r}}}(\\tau |q)\\) for a given sketch query \\(q\\), letting \\(\\rho_{g} = \\frac{\\nabla\\pi_{\\theta}(\\tau_{g}|q)}{\\pi_{\\theta_{\\mathrm{r}}}(\\tau_{g}|q)}\\), the optimization objective of GRPO is:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.195, + 0.917, + 0.224 + ], + "angle": 0, + "content": "\\[\n\\underset {\\{\\tau_ {g} \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\min \\left(\\rho_ {g} A _ {g}, \\operatorname {c l i p} \\left(\\rho_ {g}, 1 - \\epsilon , 1 + \\epsilon\\right) A _ {g}\\right) - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} | | \\pi_ {\\theta_ {\\mathrm {r}}}\\right) \\right],\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.227, + 0.913, + 0.274 + ], + "angle": 0, + "content": "\\[\n\\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} \\| \\pi_ {\\theta_ {\\mathrm {r}}}\\right) = \\frac {1}{\\rho_ {g}} + \\log \\rho_ {g} - 1, \\quad A _ {g} = \\frac {r _ {g} - \\operatorname* {m e a n} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.905, + 0.312 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) and \\(\\beta\\) are hyper-parameters for the clipped policy optimization and KL regularization terms, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.906, + 0.419 + ], + "angle": 0, + "content": "For ReMax and RLOO, we also added a small KL penalty term to the rewards to discourage divergence from the reference policy. GRPO applies group-normalization on the advantage, which we also applied in RLOO. For ReMax, we batch-normalized the advantage. For all algorithms, \\(\\pi_{\\theta}\\) is initialized from the SFT model. Further implementation details of the algorithms can be found in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.431, + 0.645, + 0.448 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.456, + 0.609, + 0.47 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.598 + ], + "angle": 0, + "content": "We train our models on SketchGraphs [32], a large-scale dataset of CAD sketches created in Onshape. SketchGraphs captures real-world parametric modeling workflows, providing geometry and constraint construction operations from actual design steps. However, the dataset was not originally designed for direct constraint inference; only \\(8.27\\%\\) of its sketches are fully-constrained, making it imperfect for training the constraint generation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.734 + ], + "angle": 0, + "content": "For computational feasibility, we deduplicate and filter sketches, retaining only those with at most 16 geometric primitives and 64 constraints, yielding a dataset of 2.8 million unique sketches. Certain constraint types, such as symmetry, are excluded to simplify the learning task, ensuring the focus remains on constraints most relevant to engineering design. We also convert Onshape sketches into the Fusion sketch format to utilize the solver in Fusion. Additional details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Another challenge lies in how SketchGraphs positions primitives. Rather than being placed in valid, constraintsatisfying layouts, primitives often have arbitrary coordinates that do not reflect a solved state. We therefore preprocess the dataset using Fusion to resolve each sketch's primitives according to its constraints, ensuring that geometry and constraints match before training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.708, + 0.864 + ], + "angle": 0, + "content": "5.2.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In Table 1, we list results comparing the performance of each alignment method with respect to the five sketch condi" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8618" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Sketch constraint generation results for Fully Constrained (FC), Under Constrained (UC), Over Constrained (OC), not solvable, and stability for the base model, SFT model, and aligned models. Results are computed over 8 samples per sketch with a temperature of 1.0. Numbers following \\(\\pm\\) indicate the standard deviation." + }, + { + "type": "table", + "bbox": [ + 0.18, + 0.143, + 0.817, + 0.272 + ], + "angle": 0, + "content": "
Model% FC ↑% UC ↓% OC ↓% Not solvable ↓% Stable (bins=4) ↑
Vitruvion (base)8.87 ±0.0971.38 ±0.2016.83 ±0.173.05 ±0.0392.15 ±0.06
SFT34.24 ±0.0946.61 ±0.1515.30 ±0.083.85 ±0.0392.48 ±0.05
Iterative DPO64.91 ±0.1114.97 ±0.1312.47 ±0.097.64 ±0.0787.63 ±0.09
Expert Iteration71.70 ±0.1313.38 ±0.137.25 ±0.067.67 ±0.0785.50 ±0.10
ReMax79.84 ±0.0915.86 ±0.071.49 ±0.012.82 ±0.0275.77 ±0.05
RLOO93.05 ±0.033.55 ±0.022.15 ±0.011.25 ±0.0189.16 ±0.02
GRPO91.59 ±0.034.18 ±0.031.94 ±0.012.28 ±0.0288.28 ±0.03
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.295, + 0.482, + 0.431 + ], + "angle": 0, + "content": "tions described in Section 3. The base model is able to fully-constrain sketches only \\(8.87\\%\\) of the time, consistent with the dataset distribution where only \\(8.27\\%\\) of sketches are fully constrained. We find that RLOO and GRPO perform similarly, giving the best performance at fully-constraining sketches \\(93.05\\%\\) and \\(91.59\\%\\) of the time, respectively. They have the lowest indices of over-constrained or unsolvable results and maintain stability rates that are on par with other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Iterative DPO and ExIt significantly improve upon the base and SFT models but still fall short of the performance achieved by policy gradient-based RL methods. We attribute this gap to the online nature of policy gradient-based RL, which continuously refines the policy through feedback while actively exploring a broader range of solutions. In contrast, Iterative DPO and ExIt are offline methods and rely on predefined ranking and filtering signals to generate training data, which limits their ability to explore the solution space. The superior performance of online RL underscores its advantage in directly optimizing the shaped rewards from the constraint solver." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 2 demonstrates the evaluation result of our methods in a few-shot inference setting, where the model has \\( K \\) attempts to generate a fully-constrained and stable output. This scenario reflects real-world use cases where the" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Table 2. Sketch constraint generation results for Pass@1 and Pass@8 across the post-training algorithms. We define a successful result as fully-constrained, not over-constrained, solvable, and stable at 4 bins. Results are generated with temperature of 1.0." + }, + { + "type": "table", + "bbox": [ + 0.16, + 0.77, + 0.413, + 0.899 + ], + "angle": 0, + "content": "
ModelPass@1Pass@8
Vitruvion (base)8.5320.47
SFT33.3242.62
Iterative DPO59.3868.32
Expert Iteration64.0972.56
ReMax62.7465.89
RLOO83.5784.96
GRPO81.4983.42
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.419 + ], + "angle": 0, + "content": "goal is to maximize the likelihood of producing an acceptable solution within a fixed inference budget of \\( K \\) samples (Pass@ \\( K \\)). We find that while the RL-based methods still have the overall highest performance, increasing the number of samples \\( K \\) has a comparatively small impact on performance compared to the other methods. Additional results and analysis on the impact of sampling parameters such as temperature and top- \\( p \\) are available in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.697, + 0.442 + ], + "angle": 0, + "content": "5.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.599 + ], + "angle": 0, + "content": "We randomly select sketches from the test set and show the results generated by different alignment methods in Figure 5. Curves are colored black when constrained and blue when not. Methods leveraging alignment techniques demonstrate a promising trend towards generating fully-constrained sketches, whereas the base Vitruvion and SFT models typically leave sketches under-constrained. However, across different alignment algorithms, we observe substantial variance in the degree of geometric distortion introduced by the aligned models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Specifically, columns A and B depict simple sketches mainly composed of horizontal and vertical lines, for which all solver-feedback methods consistently yield fully-constrained, stable results. In contrast, Column C presents a challenging sketch due to the absence of appropriate constraints for oblique lines. Achieving stability in this scenario typically requires many dimensions but few constraints; however, models are optimized to generate more constraints to align with parametric CAD design principles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Sketches in columns D through H include arcs. Column E is particularly notable as all solver-feedback methods achieve a fully-constrained condition, yet only RLOO produces visually stable results. Further examination reveals that distortions induced by other models were still classified as stable due to our bin size (\\( bins = 4 \\)). Column G presents unique challenges with an isolated point, causing confusion in the model and demanding extensive use of non-horizontal/vertical constraints to make the sketch fully-constrained and stable." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Our results indicate that models more easily fully-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.904, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.908, + 0.528 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of solved sketches across the aforementioned baseline and post-training methods. Curves are colored black when fully-constrained and blue when not. Overlaid red lines are the original input sketch for reference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.483, + 0.648 + ], + "angle": 0, + "content": "constrain sketches consisting primarily of horizontal and vertical lines without distortion. In contrast, sketches involving oblique lines increase the challenge of maintaining stability, and arcs further complicate achieving fully-constrained status. Among the tested methods, RLOO and GRPO exhibit the strongest overall performance in achieving fully-constrained status and geometric stability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Upon closer inspection of the generated constraints for the RL-based methods, we find that these models learn to \"reward hack\" by adding far more dimensions than constraints. This allows the model to avoid the instability penalty because dimensions cannot cause geometry distortion. In order to counteract this behavior, we added an additional reward penalty on the ratio of constraints versus dimensions which fixes the over-dimensioning problem at the cost of reducing stability and fully-constrained status. Additional details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To further evaluate the efficacy of each method, we conduct a forced-choice perceptual study with professional CAD designers. We show them paired images of the same sketch with different constraints applied from each method, to understand which set of constraints they prefer. We find that the preference-based optimization methods (ExIT," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.907, + 0.587 + ], + "angle": 0, + "content": "DPO) and RLOO with the over-dimensioning penalty consistently outperform the SFT and vanilla RLOO results. Results of this study are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.604, + 0.637, + 0.62 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.752 + ], + "angle": 0, + "content": "Our experiments focus on sketches of moderate complexity, with a maximum of 16 geometric primitives and 64 constraints. Evaluating performance on more complex sketches requires further exploration. Additionally, our alignment approach only uses feedback signals from a constraint solver to represent design intent. Incorporating subjective human preferences and explicit design intent into alignment objectives remains a promising area for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.634, + 0.784 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We demonstrated the adaption of alignment strategies from language modeling to preserve design intent in parametric CAD sketches. By using feedback from a constraint solver as a learning signal, we show the feasibility and value of alignment in parametric CAD tasks. In doing so, we pave the way for future AI-assisted design tools that incorporate design alignment." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8620" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "We thank William McCarthy for his valuable contribution to the human evaluation experiments, including setup, execution, and analysis. His efforts were essential in validating the effectiveness of our proposed methods from a user-centric perspective." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[1] Arash Ahmadian, Chris Cremer, Matthias Galle, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, and Sara Hooker. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. arXiv preprint arXiv:2402.14740, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Thomas W. Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In *Neural Information Processing Systems*, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.357, + 0.354 + ], + "angle": 0, + "content": "[3] Autodesk. Sketches in Fusion, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[4] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[5] Bernhard Bettig and Christoph M. Hoffmann. Geometric constraint solving in parametric computer-aided design. Journal of Computing and Information Science in Engineering, 11(2):021001, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[7] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-Aided Design, 27(6):487-501, 1995. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[8] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-aided design, 27(6):487-501, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Wei Chen, Kevin Chiu, and Mark D Fuge. Airfoil design parameterization and optimization using bezier generative adversarial networks. AIAA journal, 58(11):4723-4735, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Hyunmin Cheong, Mohammad Mehdi Ataei, Amir Hosein Khasahmadi, and Pradeep Kumar Jayaraman. e-simft: Alignment of generative models with simulation feedback for pareto-front design exploration. arXiv preprint arXiv:2502.02628, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[11] Paul F. Christiano, Jan Leike, Tom B. Brown, et al. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[12] Yasaman Etesam, Hyunmin Cheong, Mohammad Mehdi Ataei, and Pradeep Kumar Jayaraman. Deep generative model for mechanical system configuration design. arXiv preprint arXiv:2409.06016, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. NeurIPS, 34:5885-5897, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "llms in execution feedback with reinforcement learning. arXiv preprint arXiv:2410.02089, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[15] Alex Havrilla, Yuqing Du, Sharath Chandra Rarparthy, Christoforos Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, Sainbayar Sukhbaatar, and Roberta Raileanu. Teaching large language models to reason with reinforcement learning. arXiv preprint arXiv:2403.04642, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[16] Christoph M Hoffmann and Robert Joan-Arinyo. Symbolic constraints in constructive geometric constraint solving. Journal of Symbolic Computation, 23(2-3):287-299, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[17] Christoph M Hoffmann and Robert Joan-Arinyo. A brief on constraint solving. Computer-Aided Design and Applications, 2(5):655-663, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[18] Benjamin T Jones, Felix Hahnlein, Zihan Zhang, Maaz Ahmad, Vladimir Kim, and Adriana Schulz. A solver-aided hierarchical language for ltm-driven cad design. arXiv preprint arXiv:2502.09819, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[19] Ahmet Serdar Karadeniz, Dimitrios Mallis, Nesryne Mejri, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Davinci: A single-stage architecture for constrained cad sketch inference. arXiv preprint arXiv:2410.22857, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[20] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\\”ulu 3: Pushing frontiers in open language model posttraining. arXiv preprint arXiv:2411.15124, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[21] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[22] Ziniu Li, Tian Xu, Yushun Zhang, Zhihang Lin, Yang Yu, Ruoyu Sun, and Zhi-Quan Luo. Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models. In ICML, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[23] Dave Martin. What is design intent?, 2023. Accessed: January 19, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[24] Nelson Nauata, Kai-Hung Chang, Chin-Yi Cheng, Greg Mori, and Yasutaka Furukawa. House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In ECCV, pages 162-177. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[25] Jeffrey Otey, Pedro Company, Manuel Contrero, and Jorge D. Camba. Revisiting the design intent concept in the context of mechanical cad education. Computer-Aided Design and Applications, 15(1):47-60, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[26] Long Ouyang, Jeff Wu, Xu Jiang, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. NeurIPS, 34:5077-5088, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8621" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "preference optimization: Your language model is secretly a reward model. NeurIPS, 36:53728-53741, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[29] Lyle Regenwetter, Amin Heyrani Nobari, and Faziz Ahmed. Deep generative models in engineering design: A review. Journal of Mechanical Design, 144(7):071704, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[30] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 3rd edition, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[32] Ari Seff, Yaniv Ovadia, Wenda Zhou, and Ryan P. Adams. SketchGraphs: A large-scale dataset for modeling relational geometry in computer-aided design. In ICML 2020 Workshop on Object-Oriented Learning, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In ICLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.337, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[34] Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa. Housediffusion: Vector floorplan generation via a diffusion model with discrete and continuous denoising. In CVPR, pages 5466-5475, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.395, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[35] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.466, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[36] Avi Singh, John D Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alexander A Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Fathy Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura A Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[38] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, page 2692-2700, Cambridge, MA, USA, 2015. MIT Press. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[39] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Karl DD Willis, Pradeep Kumar Jayaraman, Joseph G Lambourne, Hang Chu, and Yewen Pu. Engineering sketch generation for computer-aided design. In CVPRW, pages 2105–2114, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM TOG, 40(4), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[42] Sifan Wu, Amir Hosein Khasahmadi, Mor Katz, Pradeep Kumar Jayaraman, Yewen Pu, Karl Willis, and Bang Liu. Cadvlm: Bridging language and vision in the generation of parametric cad sketches. In ECCV, pages 368-384. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[44] Xiang Xu, Pradeep Kumar Jayaraman, Joseph G Lambourne, Karl DD Willis, and Yasutaka Furukawa. Hierarchical neural coding for controllable cad model generation. In ICML, pages 38443-38461, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.361, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[45] Yuezhi Yang and Hao Pan. Discovering design concepts for cad sketches. NeurIPS, 35:28803-28814, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8622" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_origin.pdf b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f688cc71c9542b3599654dd8e89ee8cb3486287d --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/cc603c3c-bad9-4030-b99c-93bffb00f1b0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcb484d9d285f08d68ac8c840b5f38f16f8dc4a2d4f62c71762de1a30517f0c9 +size 3533021 diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/full.md b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d4dbcd927ec5da6ee65a6c10610cce8dd8c6646 --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/full.md @@ -0,0 +1,308 @@ +# Aligning Constraint Generation with Design Intent in Parametric CAD + +Evan Casey Tianyu Zhang Joseph George Lambourne + +Shu Ishida John Roger Thompson Amir Khasahmadi Pradeep Kumar Jayaraman Karl D.D. Willis + +![](images/226e5492028c66d41e667c9d13e95612edc325a53548ccdbf76435d3b0d4da3a.jpg) +Figure 1. Left: a constraint solver is used to score model generated constraints A, B, C $\sim$ $\pi_{\mathrm{base}}$ (and $A', B', C' \sim \pi_{\mathrm{aligned}}$ ). Starting with the base model $\pi_{\mathrm{base}}$ , we post-train an aligned model $\pi_{\mathrm{aligned}}$ from this feedback. Right: Blue lines show original primitives and gray lines show geometric distortion when dimensions vary. The aligned model $\pi_{\mathrm{aligned}}$ produces fully-constrained sketches that preserve relative geometric relationships, whereas the base model $\pi_{\mathrm{base}}$ produces under-constrained sketches that may distort the geometry in unintended ways. + +![](images/035129874c0020b4802ee29261be9676517a00e031d9f0265e15dae5b6b92f36.jpg) +Autodesk Research + +# Abstract + +We adapt alignment techniques from reasoning LLMs to the task of generating engineering sketch constraints found in computer-aided design (CAD) models. Engineering sketches consist of geometric primitives (e.g. points, lines) connected by constraints (e.g. perpendicular, tangent) that define the relationships between them. For a design to be easily editable, the constraints must effectively capture design intent, ensuring the geometry updates predictably when parameters change. Although current approaches can generate CAD designs, an open challenge remains to align model outputs with design intent, we label this problem 'design alignment'. A critical first step towards aligning generative CAD models is to generate constraints which fully-constrain all geometric primitives, without over-constraining or distorting sketch geometry. Using alignment techniques to train an existing constraint generation model with feedback from a constraint solver, we are able to fully-constrain $93\%$ of sketches compared to $34\%$ when using a naive supervised fine-tuning (SFT) baseline and only $8.9\%$ without SFT. Our approach can be applied to any existing constraint generation model and sets the + +stage for further research bridging alignment strategies between the language and design domains. Additional results can be found at https://autodeskailab.github.io/aligning-constraint-generation/. + +# 1. Introduction + +A central challenge in artificial intelligence (AI) is alignment: ensuring that AI systems produce outputs that adhere to human goals and expectations [11, 26, 30]. Although alignment of language models has been researched extensively [26, 28, 35], the application of alignment techniques to parametric design problems has yet to be studied. The use of AI in this discipline covers a broad range of areas, ranging from floor-plan layout [24, 34], to engineering design problems [9, 29, 41], to 3D generation [44]. Design problems are often visual in nature and incorporate other functional requirements, making them unique when compared with language model alignment. In this paper, we establish the problem of design alignment and demonstrate how this expansive problem can be made tractable by adapting techniques from the alignment literature into a new context. + +In language models, alignment is achieved by incorpo + +![](images/c1589a5d6667317960fd546a4aa2613fd48726f40a6ef91ac873b37f59b3bc62.jpg) +Figure 2. An illustration of design intent in CAD modeling. The bottom sketch maintains symmetry after modifying a dimension due to properly applied constraints, while the top sketch, lacking adequate constraints, becomes asymmetrical and distorted. + +rating feedback to generate coherent, contextually appropriate responses. Similarly, with parametric CAD modeling, AI tools must be aligned with a designer's intent by maintaining the underlying structural relationships to produce outputs that are both meaningful and functional. Otey et al. [25] define design intent as "a CAD model's anticipated behavior when altered," while Martin [23] characterizes it as "relationships between objects, so that a change to one propagates automatically to others." This means that modifications of a design by an AI system should yield outcomes where the established design relationships remain intact. To that end, we define design alignment as the application of generative modeling alignment techniques to produce outcomes that maintain design intent. + +The realization of AI systems that observe and maintain design intent has broad implications for the manufacturing and construction industries. Almost every manufactured object or structure begins as a CAD model. At the core of parametric CAD modeling are 2D engineering sketches, which can be extruded or revolved to generate 3D models. Engineering sketches are composed of geometric primitives, such as points, lines, and circles, that are organized using constraints and dimensions [7]. These constraints1 define geometric rules, including equality, perpendicularity, and radial or linear dimensions, which collectively shape the final layout of the sketch. When applied correctly, they enable efficient modifications while preserving the original design intent. + +Figure 2 illustrates the impact of constraint quality: a poorly constrained sketch loses symmetry when a dimension is changed, whereas a well-constrained sketch preserves its intended relationships. This underscores the important role of constraints in maintaining design intent, and the need for AI systems that can align with this intent encoded in designs. We focus on the problem of sketch constraint generation [32] to demonstrate the adaption of align- + +ment techniques to a design problem. Using an existing sketch constraint generation model Vitruvion [33], we align the model with algorithms that learn from feedback (Direct Preference Optimization [28], Expert Iteration [2, 36], RLOO [1], ReMax [22] and Group Relative Policy Optimization [35]) using the sketch constraint solver in Autodesk Fusion [3] as the learning signal. We optimize the models to remove all degrees of freedom in the sketches to become 'fully-constrained' [5], without causing sketches to be distorted, over-constrained or unsolvable. We further define these conditions in Section 3. + +To the best of our knowledge, this is the first instance of alignment methods being successfully applied to a parametric CAD design task; representing an important step forward for AI-assisted design tools. We present the following contributions: + +- We establish the problem of design alignment, in the context of generative CAD models, as a critical component of AI-assisted CAD tools. We focus on the necessary first step of alignment for engineering sketches. +- We introduce a post-training strategy for a sketch constraint generation model using feedback from a sketch constraint solver. We define novel metrics and reward functions that directly optimize a base model for improved alignment. +- We conduct extensive experiments and demonstrate alignment techniques that fully-constrain $93\%$ of sketches compared to $34\%$ when using a naive supervised fine-tuning (SFT) baseline and only $8.9\%$ without alignment. We posit that our approach is broadly applicable to other design tasks that require compilation of elements with rule-based algorithms. + +# 2. Related Work + +Engineering Sketches Engineering sketches form the 2D basis for 3D CAD models used to design mechanical parts for manufacturing. The availability of engineering sketch datasets [13, 32, 41] has enabled the development of generative models [13, 27, 33, 40] that can predict sketch geometry and/or the underlying constraints and dimensions that encode design intent. These Transformer-based [37] approaches create geometry by autoregressively generating tokens representing points and curves, then add constraints by referencing this geometry using Pointer Networks [38]. More recent approaches leverage image-based guidance [19, 42] or large language models (LLM) [18] in the constraint prediction task. Yang and Pan [45] learn to group together recurring patterns of geometric and constraint entities within a sketch, effectively discovering latent design concepts. However, none of these approaches explicitly optimize for preserving design intent – as a result, generated sketches may require additional manual refinement + +to capture the designer's intent. + +Our work builds upon these foundations by explicitly incorporating design intent as a post-training process. Instead of merely modeling the ground truth data, our method learns from constraint solver feedback, ensuring that generated sketches are geometrically plausible and structurally well-constrained. By doing so, we enable data-driven generation that aligns with design intent. + +Design Alignment Beyond the language domain, alignment techniques have been used to improve and align image generation. Lee et al. [21] propose fine-tuning diffusion-based text-to-image models using human feedback, significantly improving alignment between textual prompts and generated visuals. Similarly, ImageReward [43] uses a learned reward model trained on human preference data, which guides the diffusion model fine-tuning toward images preferred by human evaluators. Extending this idea, Black et al. [6] reframes image generation as a sequential RL task, introducing Denoising Diffusion Policy Optimization (DDPO) to optimize complex user-defined objectives directly for alignment without explicit human annotation. Few works have applied alignment techniques in the design domain. GearFormer [12] used differentiable sampling to enforce preferences in the solutions for mechanical configuration design problems, however, this approach does not work with rewards that require blackbox solvers in the loop. In concurrent work, e-SimFT [10] used preference data obtained from physics simulations to enhance the exploration of the Pareto front in a multi-preference setting, improving the solutions generated with GearFormer. + +Fine-tuning LLMs with RL Reinforcement Learning from Human Feedback (RLHF) has emerged as a cornerstone approach for aligning large language models (LLMs) with human preferences. In RLHF [4, 11, 26, 31], a learned reward model is usually trained to capture human preferences and is provides the learning signal that the policy is trained on. Other methods such as Direct Preference Optimization (DPO) [28] and Reinforced Self-Training (ReST) [36] use a simpler approach of optimizing the model directly from model generated data without the use of a learned reward model. While DPO learns from ranked pairs of generated model outputs from human annotators, ReST uses rejection sampling to remove incorrect generations and trains the model standard cross-entropy loss on the correct samples. In this paper, we refer to the approach of using fine-tuning on high return responses as Expert Iteration (ExIt). We broadly refer to algorithms that learn from ranked/filtered model outputs (such as DPO and Expert Iteration) as Preference Optimization (PO). + +More recently, a large body of work has focused on the task of teaching LLMs to solve reasoning tasks with reinforcement learning [14, 15, 20, 36], such as math and cod + +ing, which can be checked with rule-based systems. Specifically we are inspired by approaches which forego the use of a learned reward model and directly learn from verifiable rewards. We build off of several approaches that have shown success when applied to reasoning LLMs – these include Group Relative Policy Optimization (GRPO) [35], ReMax [22], and Reinforce Leave-One-Out (RLOO) [1]. Both GRPO and RLOO estimate the baseline via the average reward of multiple sampled outputs instead of learned value model but differ in how they apply the KL divergence penalty, advantage normalization, and PPO-style reward clipping. ReMax [22] also obviates the need for a learned value model but instead estimates the baseline from the argmax result (greedy sampling). + +In this paper, we adapt the aforementioned post-training methods—DPO, Expert Iteration, RLOO, GRPO, ReMax, for use in aligning a constraint generation model using feedback from a constraint solver. In Section 4 we provide additional details on the ranking/filtering criteria for the Preference Optimization (PO) algorithms and the reward design for the RL algorithms (GRPO, RLOO, ReMax). + +# 3. Problem + +Sketch constraining is a fundamental component of parametric CAD modeling, where geometric relationships define the structure and behavior of the sketch. Applying constraints ensures stability and editability, allowing for parametric modifications that align with the design intent. Automating constraint generation requires producing a valid and efficient set of constraints that fully define a given sketch while avoiding unnecessary redundancy or conflicts. + +The sketch constraining problem can be formulated as a sequence modeling task similar to natural language generation, where constraints are predicted autoregressively. Given the sketch geometry as input, the model generates a sequence of constraints in the order they will be applied. Tokens in the sequence represent either a constraint (e.g., coincident, parallel, perpendicular), a dimension (e.g., horizontal, vertical, radial), or a pointer to one of the input geometric entities [38]. + +In parametric CAD, sketch geometry is modified using a constraint solver. The updated sketch respects any present constraints while moving the geometry to reflect changes to the dimension parameters. Unlike natural language, which is inherently sequential and follows flexible grammar rules, sketch constraints must adhere to strict geometric principles to ensure structural validity. A set of constraints may be incorrect for a variety of reasons: they can reference the wrong primitives for the constraint type, be redundant, specify inconsistent geometric relationships, or cause unexpected geometric distortions. + +We formally define five conditions describing the state of a sketch after applying constraints. These conditions are + +![](images/db5bcb354e67d8d494bbb56eb5029dd1a66a5904cf42954d457e02fc405b279b.jpg) +Figure 3. Comparing different outcomes when a designer updates a sketch parameter on the constrained sketch. First row: an underconstrained sketch only preserves a subset of the geometric relationships. Second row: the sketch is unstable, adding a coincident constraint flattens the geometry of the sketch. Third row: the sketch is over-constrained, causing the sketch to be uneditable. Fourth row: all geometric relationships are maintained after the parameter is updated. + +not mutually exclusive. A sketch may satisfy one or more conditions depending on the applied constraints. Figure 3 provides illustrative examples of each condition. + +Under-constrained (UC) A sketch containing primitives retains some unconstrained degrees of freedom, resulting in incomplete specification of their positions or dimensions. + +Fully-constrained (FC) A sketch in which all primitives have their degrees of freedom completely determined, removing positional or dimensional ambiguities. + +Over-constrained (OC) A sketch primitives have more constraints applied than degrees of freedom, potentially leading to conflicts. Note some over-constrained sketches remain solvable if the constraints are consistent and do not conflict with each other [7]. + +Not solvable A sketch that cannot achieve a valid solution due to contradictory or redundant constraints, leading to an impossible or conflicting geometry. + +Stability We discretize the sketch plane into a grid and classify a sketch as unstable if the positions of primitives after constraint solving shift into different cells. The number of cells (bins) on each axis determines the sensitivity of this measurement. + +Our objective is to align the model toward generating + +constraint sets that yield fully-constrained sketches while minimizing cases of under-constrained, over-constrained, not solvable, or instability. This is a necessary prerequisite toward the ultimate goal of generating constraints that preserve the original sketch design intent when the designer varies the geometric parameters. A formal definition of FC, UC, and OC is described in [16, 17] and in the appendix. + +# 4. Method + +In this section, we outline the post-training techniques used for aligning a constraint generation model with feedback from a constraint solver. The approaches are grouped into three categories: supervised learning methods, preference-based optimization methods, and RL methods. Figure 4 illustrates the high-level workflow of this work. + +# 4.1. Constraint Solver + +We evaluate generated constraints using the commercially available constraint solver in Autodesk Fusion. This solver is treated as a black box, from which we retrieve sketch conditions as defined in Section 3. A general algorithmic approach commonly used in constraint solvers is described in [8]. The solver outputs the fully-constrained status for each entity if the sketch is solvable. Otherwise it classifies the sketch as over-constrained or unsolvable. The solver adjusts geometry to resolve all constraints which we compare with the original sketch to determine if the constraints caused geometry distortion (instability). On average, it takes 0.1-0.2 seconds to per solve, although complex cases may take tens of seconds. Any sketch taking longer than two seconds to solve is automatically deemed unsolvable. + +# 4.2. Supervised Learning + +We pre-train Vitruvion [33] as our base model using the same procedure described in their paper with a next-token prediction objective. Given an input sequence of geometric primitives, the model is trained to predict the next correct constraint or dimension based on the ground truth data. Further details about our implementation of the Vitruvion architecture and training procedure are provided in the appendix. + +Since the majority of the ground truth data contains under-constrained or over-constrained sketches, we additionally perform supervised fine-tuning (SFT). During SFT, the training data is limited to sketches verified by the constraint solver as solvable, fully-constrained, stable, and free of over-constrained conditions, ensuring the model explicitly learns from ideal examples. + +# 4.3. Preference-Based Optimization (PO) + +Expert Iteration (ExIt) alternates between expert improvement and policy distillation. Following [15, 36], + +![](images/44f4104c4891fa907efad26cb8bd86f2e9232912c65e20c4d71d3ef6017c8820.jpg) +Figure 4. Illustration of the proposed alignment workflow for constraint generation models. Left: A base constraint-generation model is first fine-tuned using supervised fine-tuning (SFT). Middle: Generated constraint sequences $C_i$ are evaluated using a constraint solver, which provides feedback on sketch stability and fully- and over-constrained statuses, forming the quality score vector $S_i$ . Right: Two groups of alignment methods leveraging solver feedback: preference-based optimization (PO) and reinforcement learning (RL). PO uses $S_i$ to construct training data, iteratively improving constraint prediction quality. RL methods assign per-constraint and per-sequence rewards to $C_i$ based on solver feedback and $S_i$ to incentivize the generation of constraints, making sketches stable and fully-constrained. + +we use temperature sampling combined with rejection sampling to generate high-quality candidate constraint sequences. During the exploration step we initialize the policy model $\pi_{\theta_{t = 0}}$ from the SFT model and sample $K = 8$ candidate constraint sequences $\tau$ at temperature $T = 1.0$ for each sketch query $q$ in the initial training set $\mathcal{D}$ . A new training dataset $\mathcal{D}_{i\in N}^{*}$ is constructed by discarding sequences that are under-constrained, over-constrained, or unsolvable solutions. This process is repeated $N = 2$ times over the dataset, and the policy is trained using cross-entropy loss: + +$$ +\mathbb {E} _ {(q, \tau) \sim \mathcal {D} _ {i \in N} ^ {*}} [ \log \pi_ {\theta_ {t}} (\tau | q) ] \tag {1} +$$ + +where $\pi_{\theta_t}$ is updated after the distillation phase of every training round. + +Direct Preference Optimization (DPO) learns from pairwise preference data, approximating an implicit reward via a reparameterized Bradley-Terry model [28]. Similar to ExIt, we construct the training dataset by sampling $K = 8$ constraint sequence completions $\tau$ for each sketch query $q$ from the policy model $\pi_{\theta_t}$ at temperature $T = 1.0$ . Pairs $(\tau_w, \tau_l)$ are ranked based on the differences in the percentage of fully-constrained curves between $\tau_w$ (preferred sequence), and $\tau_l$ (non-preferred sequence). The optimization objective is: + +$$ +\mathbb {E} _ {(q, \tau_ {w}, \tau_ {l}) \sim \mathcal {D}} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta_ {t}} (\tau_ {w} | q)}{\pi_ {\theta_ {r}} (\tau_ {w} | q)} - \beta \log \frac {\pi_ {\theta_ {t}} (\tau_ {l} | q)}{\pi_ {\theta_ {r}} (\tau_ {l} | q)}\right) \right] \tag {2} +$$ + +where $\beta$ is a hyperparameter, and $\sigma$ is the logistic function. This formulation ensures the learned policy $\pi_{\theta}$ aligns with + +the ranking preference of fully-constrained sketches while maintaining proximity to the reference policy $\pi_{\theta_r}$ . + +We initialize $\pi_{\theta_{t = 0}}$ from the SFT model and repeat the entire process $N = 2$ times, updating $\pi_{\theta_t}$ with the new policy after every training iteration. Additional hyperparameters and ranking details can be found in the appendix. + +# 4.4. Reinforcement Learning (RL) + +# 4.4.1. Reward design + +Unlike natural language tasks, where human preferences are ambiguous and ill-defined, the stability and solvability of sketch constraints can be verified, making it compatible with RL methods that directly optimize for mechanically defined rewards without a learned preference model. We define the rewards used for RL as follows: + +- Rewards for valid constraint sequence $\tau$ : + +$r_{\mathrm{curves}}(\tau)$ : $\%$ of fully-constrained curves over all curves, $r_{\mathrm{points}}(\tau)$ : $\%$ of fully-constrained points over all points, $r_{\mathrm{unstable}}$ : penalty for unstable sketches, + +- Rewards for invalid constraint sequence: + +$r_{\mathrm{NS}}$ : penalty for not solvable sketches, + +$r_{\mathrm{OC}}$ : penalty for over-constrained sketches, + +$r_{\mathbf{F}}$ : penalty for sketches resulting in other failures. + +The overall sequence-wise reward $R(\tau)$ is the sum of $r_{\mathrm{curves}}(\tau), r_{\mathrm{points}}(\tau)$ , and conditionally $r_{\mathrm{unstable}}$ for valid sketches, and either $r_{\mathrm{NS}}, r_{\mathrm{OC}}$ or $r_{\mathrm{F}}$ for invalid sketches according to the failure mode. + +We additionally define a constraint-wise penalty to provide granular feedback on cases where the constraint sequence causes sketches to be over-constrained or fully-constrained. A constraint solver iteratively attempts to add each generated constraint one-by-one, dropping any problematic constraints that caused the sketch to be over-constrained or not-solvable. In training, we add a constant of -1 loss penalty directly to the per-token log likelihood loss for the problematic constraints. + +# 4.4.2. RL fine-tuning formulation + +We formulate constraint generation fine-tuning as follows; given a dataset of sketch queries $D = \{q_i\}_{i=1}^N$ and reward function $R(\tau)$ using the constraint solver and reward design in Section 4.4.1, learn a policy $\pi_{\theta}(\tau|q)$ that generates a sequence of constraints $\tau$ for sketch $q$ , such that it maximizes the expected rewards $\mathbb{E}_{q_i \sim D, \tau_i \sim \pi_{\theta}(\cdot | q_i)}[R(\tau_i)]$ . + +# 4.4.3. Policy gradient methods + +For RLHF which uses a pre-trained policy, not all the complexity of RL algorithms is necessary. This allows the algorithms to be simplified and the number of learnable components to be reduced, contributing to performance improvement. We considered three policy gradient algorithms: ReMax [22], RLOO [1], and GRPO [35]. Unlike PPO [31], which treats each token generation as an action, these algorithms treat generation of a sequence as a single action and adopt a REINFORCE with baselines approach [39]. We apply these to optimize the constraint generation policy. + +ReMax [22] uses the rewards corresponding to sequences generated by a greedy (argmax) policy as a baseline to normalize the rewards of sequences sampled from the policy. + +Considering sequence $\tau$ sampled from policy $\pi_{\theta}(\tau |q)$ , sequence $\tau^{*} = \operatorname{argmax}_{\tau}\pi_{\theta}(\tau |q)$ greedily sampled by taking an argmax of the policy, and corresponding rewards $r$ and $r^{*}$ , the policy gradient objective for ReMax is: + +$$ +\underset {\tau \sim \pi} {\mathbb {E}} \left[ \left(r - r ^ {*}\right) \nabla \log \pi_ {\theta} (\tau | q) \right]. \tag {3} +$$ + +REINFORCE-Leave-One-Out (RLOO) [1] samples $G$ number of constraint sequences for every sketch query. The baseline for each sample is evaluated as the mean of the rewards for all other samples in the group. + +For $G$ number of sequences $\{\tau_g\}_{g = 1}^G$ sampled from policy $\pi_{\theta}(\tau |q)$ for a given sketch query $q$ , the policy gradient objective of RLOO is: + +$$ +\underset {\{\tau \} \sim \pi} {\mathbb {E}} \left[ \frac {1}{G} \sum_ {g = 1} ^ {G} \left[ \left(r _ {g} - \operatorname {m e a n} \left(\left\{r _ {i} \right\} _ {i \neq g} ^ {G}\right)\right) \nabla \log \pi_ {\theta} \left(\tau_ {g} \mid q _ {g}\right) \right] \right]. \tag {4} +$$ + +Group Relative Policy Optimization (GRPO) [35] uses group-based baseline estimation, like RLOO, but the mean + +is taken over all reward samples in the group. It uses a clipped policy optimization objective similarly to PPO [31], as well as a low-variance KL regularization term. + +For $G$ number of sequences $\{\tau_g\}_{g = 1}^G$ sampled from the reference policy $\pi_{\theta_{\mathrm{r}}}(\tau |q)$ for a given sketch query $q$ , letting $\rho_{g} = \frac{\nabla\pi_{\theta}(\tau_{g}|q)}{\pi_{\theta_{\mathrm{r}}}(\tau_{g}|q)}$ , the optimization objective of GRPO is: + +$$ +\underset {\{\tau_ {g} \} \sim \pi} {\mathbb {E}} \left[ \min \left(\rho_ {g} A _ {g}, \operatorname {c l i p} \left(\rho_ {g}, 1 - \epsilon , 1 + \epsilon\right) A _ {g}\right) - \beta \mathbb {D} _ {\mathrm {K L}} \left(\pi_ {\theta} | | \pi_ {\theta_ {\mathrm {r}}}\right) \right], +$$ + +$$ +\mathbb {D} _ {\mathrm {K L}} \left(\pi_ {\theta} \| \pi_ {\theta_ {\mathrm {r}}}\right) = \frac {1}{\rho_ {g}} + \log \rho_ {g} - 1, \quad A _ {g} = \frac {r _ {g} - \operatorname* {m e a n} \left(\left\{r _ {g} \right\} _ {g = 1} ^ {G}\right)}{\operatorname {s t d} \left(\left\{r _ {g} \right\} _ {g = 1} ^ {G}\right)}, \tag {5} +$$ + +where $\epsilon$ and $\beta$ are hyper-parameters for the clipped policy optimization and KL regularization terms, respectively. + +For ReMax and RLOO, we also added a small KL penalty term to the rewards to discourage divergence from the reference policy. GRPO applies group-normalization on the advantage, which we also applied in RLOO. For ReMax, we batch-normalized the advantage. For all algorithms, $\pi_{\theta}$ is initialized from the SFT model. Further implementation details of the algorithms can be found in the appendix. + +# 5. Experiments + +# 5.1. Dataset + +We train our models on SketchGraphs [32], a large-scale dataset of CAD sketches created in Onshape. SketchGraphs captures real-world parametric modeling workflows, providing geometry and constraint construction operations from actual design steps. However, the dataset was not originally designed for direct constraint inference; only $8.27\%$ of its sketches are fully-constrained, making it imperfect for training the constraint generation model. + +For computational feasibility, we deduplicate and filter sketches, retaining only those with at most 16 geometric primitives and 64 constraints, yielding a dataset of 2.8 million unique sketches. Certain constraint types, such as symmetry, are excluded to simplify the learning task, ensuring the focus remains on constraints most relevant to engineering design. We also convert Onshape sketches into the Fusion sketch format to utilize the solver in Fusion. Additional details are provided in the appendix. + +Another challenge lies in how SketchGraphs positions primitives. Rather than being placed in valid, constraintsatisfying layouts, primitives often have arbitrary coordinates that do not reflect a solved state. We therefore preprocess the dataset using Fusion to resolve each sketch's primitives according to its constraints, ensuring that geometry and constraints match before training. + +# 5.2.Quantitative Results + +In Table 1, we list results comparing the performance of each alignment method with respect to the five sketch condi + +Table 1. Sketch constraint generation results for Fully Constrained (FC), Under Constrained (UC), Over Constrained (OC), not solvable, and stability for the base model, SFT model, and aligned models. Results are computed over 8 samples per sketch with a temperature of 1.0. Numbers following $\pm$ indicate the standard deviation. + +
Model% FC ↑% UC ↓% OC ↓% Not solvable ↓% Stable (bins=4) ↑
Vitruvion (base)8.87 ±0.0971.38 ±0.2016.83 ±0.173.05 ±0.0392.15 ±0.06
SFT34.24 ±0.0946.61 ±0.1515.30 ±0.083.85 ±0.0392.48 ±0.05
Iterative DPO64.91 ±0.1114.97 ±0.1312.47 ±0.097.64 ±0.0787.63 ±0.09
Expert Iteration71.70 ±0.1313.38 ±0.137.25 ±0.067.67 ±0.0785.50 ±0.10
ReMax79.84 ±0.0915.86 ±0.071.49 ±0.012.82 ±0.0275.77 ±0.05
RLOO93.05 ±0.033.55 ±0.022.15 ±0.011.25 ±0.0189.16 ±0.02
GRPO91.59 ±0.034.18 ±0.031.94 ±0.012.28 ±0.0288.28 ±0.03
+ +tions described in Section 3. The base model is able to fully-constrain sketches only $8.87\%$ of the time, consistent with the dataset distribution where only $8.27\%$ of sketches are fully constrained. We find that RLOO and GRPO perform similarly, giving the best performance at fully-constraining sketches $93.05\%$ and $91.59\%$ of the time, respectively. They have the lowest indices of over-constrained or unsolvable results and maintain stability rates that are on par with other methods. + +Iterative DPO and ExIt significantly improve upon the base and SFT models but still fall short of the performance achieved by policy gradient-based RL methods. We attribute this gap to the online nature of policy gradient-based RL, which continuously refines the policy through feedback while actively exploring a broader range of solutions. In contrast, Iterative DPO and ExIt are offline methods and rely on predefined ranking and filtering signals to generate training data, which limits their ability to explore the solution space. The superior performance of online RL underscores its advantage in directly optimizing the shaped rewards from the constraint solver. + +Table 2 demonstrates the evaluation result of our methods in a few-shot inference setting, where the model has $K$ attempts to generate a fully-constrained and stable output. This scenario reflects real-world use cases where the + +Table 2. Sketch constraint generation results for Pass@1 and Pass@8 across the post-training algorithms. We define a successful result as fully-constrained, not over-constrained, solvable, and stable at 4 bins. Results are generated with temperature of 1.0. + +
ModelPass@1Pass@8
Vitruvion (base)8.5320.47
SFT33.3242.62
Iterative DPO59.3868.32
Expert Iteration64.0972.56
ReMax62.7465.89
RLOO83.5784.96
GRPO81.4983.42
+ +goal is to maximize the likelihood of producing an acceptable solution within a fixed inference budget of $K$ samples (Pass@ $K$ ). We find that while the RL-based methods still have the overall highest performance, increasing the number of samples $K$ has a comparatively small impact on performance compared to the other methods. Additional results and analysis on the impact of sampling parameters such as temperature and top- $p$ are available in the appendix. + +# 5.3. Qualitative Results + +We randomly select sketches from the test set and show the results generated by different alignment methods in Figure 5. Curves are colored black when constrained and blue when not. Methods leveraging alignment techniques demonstrate a promising trend towards generating fully-constrained sketches, whereas the base Vitruvion and SFT models typically leave sketches under-constrained. However, across different alignment algorithms, we observe substantial variance in the degree of geometric distortion introduced by the aligned models. + +Specifically, columns A and B depict simple sketches mainly composed of horizontal and vertical lines, for which all solver-feedback methods consistently yield fully-constrained, stable results. In contrast, Column C presents a challenging sketch due to the absence of appropriate constraints for oblique lines. Achieving stability in this scenario typically requires many dimensions but few constraints; however, models are optimized to generate more constraints to align with parametric CAD design principles. + +Sketches in columns D through H include arcs. Column E is particularly notable as all solver-feedback methods achieve a fully-constrained condition, yet only RLOO produces visually stable results. Further examination reveals that distortions induced by other models were still classified as stable due to our bin size ( $bins = 4$ ). Column G presents unique challenges with an isolated point, causing confusion in the model and demanding extensive use of non-horizontal/vertical constraints to make the sketch fully-constrained and stable. + +Our results indicate that models more easily fully- + +![](images/a0c5fa2a1e2d2718102b4c6d50fcd713bcc34b8cd20e678f92ea061468739bf1.jpg) +Figure 5. Visual comparison of solved sketches across the aforementioned baseline and post-training methods. Curves are colored black when fully-constrained and blue when not. Overlaid red lines are the original input sketch for reference. + +constrain sketches consisting primarily of horizontal and vertical lines without distortion. In contrast, sketches involving oblique lines increase the challenge of maintaining stability, and arcs further complicate achieving fully-constrained status. Among the tested methods, RLOO and GRPO exhibit the strongest overall performance in achieving fully-constrained status and geometric stability. + +Upon closer inspection of the generated constraints for the RL-based methods, we find that these models learn to "reward hack" by adding far more dimensions than constraints. This allows the model to avoid the instability penalty because dimensions cannot cause geometry distortion. In order to counteract this behavior, we added an additional reward penalty on the ratio of constraints versus dimensions which fixes the over-dimensioning problem at the cost of reducing stability and fully-constrained status. Additional details are provided in the appendix. + +To further evaluate the efficacy of each method, we conduct a forced-choice perceptual study with professional CAD designers. We show them paired images of the same sketch with different constraints applied from each method, to understand which set of constraints they prefer. We find that the preference-based optimization methods (ExIT, + +DPO) and RLOO with the over-dimensioning penalty consistently outperform the SFT and vanilla RLOO results. Results of this study are provided in the appendix. + +# 6. Limitations + +Our experiments focus on sketches of moderate complexity, with a maximum of 16 geometric primitives and 64 constraints. Evaluating performance on more complex sketches requires further exploration. Additionally, our alignment approach only uses feedback signals from a constraint solver to represent design intent. Incorporating subjective human preferences and explicit design intent into alignment objectives remains a promising area for future work. + +# 7. Conclusion + +We demonstrated the adaption of alignment strategies from language modeling to preserve design intent in parametric CAD sketches. By using feedback from a constraint solver as a learning signal, we show the feasibility and value of alignment in parametric CAD tasks. In doing so, we pave the way for future AI-assisted design tools that incorporate design alignment. + +# Acknowledgement + +We thank William McCarthy for his valuable contribution to the human evaluation experiments, including setup, execution, and analysis. His efforts were essential in validating the effectiveness of our proposed methods from a user-centric perspective. + +# References + +[1] Arash Ahmadian, Chris Cremer, Matthias Galle, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, and Sara Hooker. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. arXiv preprint arXiv:2402.14740, 2024. 2, 3, 6 +[2] Thomas W. Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In *Neural Information Processing Systems*, 2017. 2 +[3] Autodesk. Sketches in Fusion, 2014. 2 +[4] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3 +[5] Bernhard Bettig and Christoph M. Hoffmann. Geometric constraint solving in parametric computer-aided design. Journal of Computing and Information Science in Engineering, 11(2):021001, 2011. 2 +[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3 +[7] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-Aided Design, 27(6):487-501, 1995. 2, 4 +[8] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-aided design, 27(6):487-501, 1995. 4 +[9] Wei Chen, Kevin Chiu, and Mark D Fuge. Airfoil design parameterization and optimization using bezier generative adversarial networks. AIAA journal, 58(11):4723-4735, 2020. 1 +[10] Hyunmin Cheong, Mohammad Mehdi Ataei, Amir Hosein Khasahmadi, and Pradeep Kumar Jayaraman. e-simft: Alignment of generative models with simulation feedback for pareto-front design exploration. arXiv preprint arXiv:2502.02628, 2025.3 +[11] Paul F. Christiano, Jan Leike, Tom B. Brown, et al. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 1, 3 +[12] Yasaman Etesam, Hyunmin Cheong, Mohammad Mehdi Ataei, and Pradeep Kumar Jayaraman. Deep generative model for mechanical system configuration design. arXiv preprint arXiv:2409.06016, 2024. 3 +[13] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. NeurIPS, 34:5885-5897, 2021. 2 +[14] Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code + +llms in execution feedback with reinforcement learning. arXiv preprint arXiv:2410.02089, 2024.3 +[15] Alex Havrilla, Yuqing Du, Sharath Chandra Rarparthy, Christoforos Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, Sainbayar Sukhbaatar, and Roberta Raileanu. Teaching large language models to reason with reinforcement learning. arXiv preprint arXiv:2403.04642, 2024. 3, 4 +[16] Christoph M Hoffmann and Robert Joan-Arinyo. Symbolic constraints in constructive geometric constraint solving. Journal of Symbolic Computation, 23(2-3):287-299, 1997. 4 +[17] Christoph M Hoffmann and Robert Joan-Arinyo. A brief on constraint solving. Computer-Aided Design and Applications, 2(5):655-663, 2005. 4 +[18] Benjamin T Jones, Felix Hahnlein, Zihan Zhang, Maaz Ahmad, Vladimir Kim, and Adriana Schulz. A solver-aided hierarchical language for ltm-driven cad design. arXiv preprint arXiv:2502.09819, 2025. 2 +[19] Ahmet Serdar Karadeniz, Dimitrios Mallis, Nesryne Mejri, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Davinci: A single-stage architecture for constrained cad sketch inference. arXiv preprint arXiv:2410.22857, 2024. 2 +[20] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\”ulu 3: Pushing frontiers in open language model posttraining. arXiv preprint arXiv:2411.15124, 2024. 3 +[21] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3 +[22] Ziniu Li, Tian Xu, Yushun Zhang, Zhihang Lin, Yang Yu, Ruoyu Sun, and Zhi-Quan Luo. Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models. In ICML, 2024. 2, 3, 6 +[23] Dave Martin. What is design intent?, 2023. Accessed: January 19, 2023. 2 +[24] Nelson Nauata, Kai-Hung Chang, Chin-Yi Cheng, Greg Mori, and Yasutaka Furukawa. House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In ECCV, pages 162-177. Springer, 2020. 1 +[25] Jeffrey Otey, Pedro Company, Manuel Contrero, and Jorge D. Camba. Revisiting the design intent concept in the context of mechanical cad education. Computer-Aided Design and Applications, 15(1):47-60, 2018. 2 +[26] Long Ouyang, Jeff Wu, Xu Jiang, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 3 +[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. NeurIPS, 34:5077-5088, 2021. 2 +[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct + +preference optimization: Your language model is secretly a reward model. NeurIPS, 36:53728-53741, 2023. 1, 2, 3, 5 +[29] Lyle Regenwetter, Amin Heyrani Nobari, and Faziz Ahmed. Deep generative models in engineering design: A review. Journal of Mechanical Design, 144(7):071704, 2022. 1 +[30] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 3rd edition, 2010. 1 +[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 6 +[32] Ari Seff, Yaniv Ovadia, Wenda Zhou, and Ryan P. Adams. SketchGraphs: A large-scale dataset for modeling relational geometry in computer-aided design. In ICML 2020 Workshop on Object-Oriented Learning, 2020. 2, 6 +[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In ICLR, 2021. 2, 4 +[34] Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa. Housediffusion: Vector floorplan generation via a diffusion model with discrete and continuous denoising. In CVPR, pages 5466-5475, 2023. 1 +[35] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 2, 3, 6 +[36] Avi Singh, John D Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alexander A Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Fathy Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura A Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research, 2024. 2, 3, 4 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2 +[38] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, page 2692-2700, Cambridge, MA, USA, 2015. MIT Press. 2, 3 +[39] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 6 +[40] Karl DD Willis, Pradeep Kumar Jayaraman, Joseph G Lambourne, Hang Chu, and Yewen Pu. Engineering sketch generation for computer-aided design. In CVPRW, pages 2105–2114, 2021. 2 + +[41] Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM TOG, 40(4), 2021. 1, 2 +[42] Sifan Wu, Amir Hosein Khasahmadi, Mor Katz, Pradeep Kumar Jayaraman, Yewen Pu, Karl Willis, and Bang Liu. Cadvlm: Bridging language and vision in the generation of parametric cad sketches. In ECCV, pages 368-384. Springer, 2024. 2 +[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3 +[44] Xiang Xu, Pradeep Kumar Jayaraman, Joseph G Lambourne, Karl DD Willis, and Yasutaka Furukawa. Hierarchical neural coding for controllable cad model generation. In ICML, pages 38443-38461, 2023. 1 +[45] Yuezhi Yang and Hao Pan. Discovering design concepts for cad sketches. NeurIPS, 35:28803-28814, 2022. 2 \ No newline at end of file diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/images.zip b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f3c0969f709f3d3abab983ba43bbcfd9c0c4f7cd --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4d21d0f41eed703b4ed69b52374da6262390cfdffe1968aab84497d920a9ab4 +size 509337 diff --git a/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/layout.json b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d3b1fd1dd64dd4da15ca065d6b4b8d78d2c6050d --- /dev/null +++ b/2025/Aligning Constraint Generation with Design Intent in Parametric CAD/layout.json @@ -0,0 +1,8251 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 524, + 121 + ], + "type": "text", + "content": "Aligning Constraint Generation with Design Intent in Parametric CAD" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 143, + 236, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 143, + 236, + 172 + ], + "spans": [ + { + "bbox": [ + 67, + 143, + 236, + 172 + ], + "type": "text", + "content": "Evan Casey Tianyu Zhang Joseph George Lambourne" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 143, + 542, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 143, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 241, + 143, + 542, + 171 + ], + "type": "text", + "content": "Shu Ishida John Roger Thompson Amir Khasahmadi Pradeep Kumar Jayaraman Karl D.D. Willis" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 213, + 226, + 369 + ], + "blocks": [ + { + "bbox": [ + 57, + 213, + 226, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 213, + 226, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 213, + 226, + 369 + ], + "type": "image", + "image_path": "226e5492028c66d41e667c9d13e95612edc325a53548ccdbf76435d3b0d4da3a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": "Figure 1. Left: a constraint solver is used to score model generated constraints A, B, C " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{base}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": " (and " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "A', B', C' \\sim \\pi_{\\mathrm{aligned}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": "). Starting with the base model " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{base}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": ", we post-train an aligned model " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{aligned}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": " from this feedback. Right: Blue lines show original primitives and gray lines show geometric distortion when dimensions vary. The aligned model " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{aligned}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": " produces fully-constrained sketches that preserve relative geometric relationships, whereas the base model " + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{base}}" + }, + { + "bbox": [ + 54, + 380, + 555, + 427 + ], + "type": "text", + "content": " produces under-constrained sketches that may distort the geometry in unintended ways." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 249, + 213, + 555, + 365 + ], + "blocks": [ + { + "bbox": [ + 255, + 185, + 351, + 198 + ], + "lines": [ + { + "bbox": [ + 255, + 185, + 351, + 198 + ], + "spans": [ + { + "bbox": [ + 255, + 185, + 351, + 198 + ], + "type": "text", + "content": "Autodesk Research" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 249, + 213, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 249, + 213, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 249, + 213, + 555, + 365 + ], + "type": "image", + "image_path": "035129874c0020b4802ee29261be9676517a00e031d9f0265e15dae5b6b92f36.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "spans": [ + { + "bbox": [ + 151, + 435, + 200, + 448 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": "We adapt alignment techniques from reasoning LLMs to the task of generating engineering sketch constraints found in computer-aided design (CAD) models. Engineering sketches consist of geometric primitives (e.g. points, lines) connected by constraints (e.g. perpendicular, tangent) that define the relationships between them. For a design to be easily editable, the constraints must effectively capture design intent, ensuring the geometry updates predictably when parameters change. Although current approaches can generate CAD designs, an open challenge remains to align model outputs with design intent, we label this problem 'design alignment'. A critical first step towards aligning generative CAD models is to generate constraints which fully-constrain all geometric primitives, without over-constraining or distorting sketch geometry. Using alignment techniques to train an existing constraint generation model with feedback from a constraint solver, we are able to fully-constrain " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " of sketches compared to " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " when using a naive supervised fine-tuning (SFT) baseline and only " + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 54, + 462, + 297, + 714 + ], + "type": "text", + "content": " without SFT. Our approach can be applied to any existing constraint generation model and sets the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 485 + ], + "type": "text", + "content": "stage for further research bridging alignment strategies between the language and design domains. Additional results can be found at https://autodeskailab.github.io/aligning-constraint-generation/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 395, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "type": "text", + "content": "A central challenge in artificial intelligence (AI) is alignment: ensuring that AI systems produce outputs that adhere to human goals and expectations [11, 26, 30]. Although alignment of language models has been researched extensively [26, 28, 35], the application of alignment techniques to parametric design problems has yet to be studied. The use of AI in this discipline covers a broad range of areas, ranging from floor-plan layout [24, 34], to engineering design problems [9, 29, 41], to 3D generation [44]. Design problems are often visual in nature and incorporate other functional requirements, making them unique when compared with language model alignment. In this paper, we establish the problem of design alignment and demonstrate how this expansive problem can be made tractable by adapting techniques from the alignment literature into a new context." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "In language models, alignment is achieved by incorpo" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8613" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 194 + ], + "type": "image", + "image_path": "c1589a5d6667317960fd546a4aa2613fd48726f40a6ef91ac873b37f59b3bc62.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "type": "text", + "content": "Figure 2. An illustration of design intent in CAD modeling. The bottom sketch maintains symmetry after modifying a dimension due to properly applied constraints, while the top sketch, lacking adequate constraints, becomes asymmetrical and distorted." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 255, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 295, + 422 + ], + "type": "text", + "content": "rating feedback to generate coherent, contextually appropriate responses. Similarly, with parametric CAD modeling, AI tools must be aligned with a designer's intent by maintaining the underlying structural relationships to produce outputs that are both meaningful and functional. Otey et al. [25] define design intent as \"a CAD model's anticipated behavior when altered,\" while Martin [23] characterizes it as \"relationships between objects, so that a change to one propagates automatically to others.\" This means that modifications of a design by an AI system should yield outcomes where the established design relationships remain intact. To that end, we define design alignment as the application of generative modeling alignment techniques to produce outcomes that maintain design intent." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 590 + ], + "type": "text", + "content": "The realization of AI systems that observe and maintain design intent has broad implications for the manufacturing and construction industries. Almost every manufactured object or structure begins as a CAD model. At the core of parametric CAD modeling are 2D engineering sketches, which can be extruded or revolved to generate 3D models. Engineering sketches are composed of geometric primitives, such as points, lines, and circles, that are organized using constraints and dimensions [7]. These constraints1 define geometric rules, including equality, perpendicularity, and radial or linear dimensions, which collectively shape the final layout of the sketch. When applied correctly, they enable efficient modifications while preserving the original design intent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 687 + ], + "type": "text", + "content": "Figure 2 illustrates the impact of constraint quality: a poorly constrained sketch loses symmetry when a dimension is changed, whereas a well-constrained sketch preserves its intended relationships. This underscores the important role of constraints in maintaining design intent, and the need for AI systems that can align with this intent encoded in designs. We focus on the problem of sketch constraint generation [32] to demonstrate the adaption of align-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "ment techniques to a design problem. Using an existing sketch constraint generation model Vitruvion [33], we align the model with algorithms that learn from feedback (Direct Preference Optimization [28], Expert Iteration [2, 36], RLOO [1], ReMax [22] and Group Relative Policy Optimization [35]) using the sketch constraint solver in Autodesk Fusion [3] as the learning signal. We optimize the models to remove all degrees of freedom in the sketches to become 'fully-constrained' [5], without causing sketches to be distorted, over-constrained or unsolvable. We further define these conditions in Section 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 263 + ], + "type": "text", + "content": "To the best of our knowledge, this is the first instance of alignment methods being successfully applied to a parametric CAD design task; representing an important step forward for AI-assisted design tools. We present the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 268, + 553, + 460 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 268, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 316 + ], + "type": "text", + "content": "- We establish the problem of design alignment, in the context of generative CAD models, as a critical component of AI-assisted CAD tools. We focus on the necessary first step of alignment for engineering sketches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "type": "text", + "content": "- We introduce a post-training strategy for a sketch constraint generation model using feedback from a sketch constraint solver. We define novel metrics and reward functions that directly optimize a base model for improved alignment." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": "- We conduct extensive experiments and demonstrate alignment techniques that fully-constrain " + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": " of sketches compared to " + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "inline_equation", + "content": "34\\%" + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": " when using a naive supervised fine-tuning (SFT) baseline and only " + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 313, + 376, + 553, + 460 + ], + "type": "text", + "content": " without alignment. We posit that our approach is broadly applicable to other design tasks that require compilation of elements with rule-based algorithms." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 472, + 400, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 400, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 400, + 486 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "Engineering Sketches Engineering sketches form the 2D basis for 3D CAD models used to design mechanical parts for manufacturing. The availability of engineering sketch datasets [13, 32, 41] has enabled the development of generative models [13, 27, 33, 40] that can predict sketch geometry and/or the underlying constraints and dimensions that encode design intent. These Transformer-based [37] approaches create geometry by autoregressively generating tokens representing points and curves, then add constraints by referencing this geometry using Pointer Networks [38]. More recent approaches leverage image-based guidance [19, 42] or large language models (LLM) [18] in the constraint prediction task. Yang and Pan [45] learn to group together recurring patterns of geometric and constraint entities within a sketch, effectively discovering latent design concepts. However, none of these approaches explicitly optimize for preserving design intent – as a result, generated sketches may require additional manual refinement" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1 Throughout this paper, the term \"constraints\" is used in a broad sense to include both constraints (e.g., parallel) and dimensions (e.g., diameter)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8614" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 184, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 184, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 184, + 85 + ], + "type": "text", + "content": "to capture the designer's intent." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 172 + ], + "type": "text", + "content": "Our work builds upon these foundations by explicitly incorporating design intent as a post-training process. Instead of merely modeling the ground truth data, our method learns from constraint solver feedback, ensuring that generated sketches are geometrically plausible and structurally well-constrained. By doing so, we enable data-driven generation that aligns with design intent." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 177, + 297, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 297, + 441 + ], + "type": "text", + "content": "Design Alignment Beyond the language domain, alignment techniques have been used to improve and align image generation. Lee et al. [21] propose fine-tuning diffusion-based text-to-image models using human feedback, significantly improving alignment between textual prompts and generated visuals. Similarly, ImageReward [43] uses a learned reward model trained on human preference data, which guides the diffusion model fine-tuning toward images preferred by human evaluators. Extending this idea, Black et al. [6] reframes image generation as a sequential RL task, introducing Denoising Diffusion Policy Optimization (DDPO) to optimize complex user-defined objectives directly for alignment without explicit human annotation. Few works have applied alignment techniques in the design domain. GearFormer [12] used differentiable sampling to enforce preferences in the solutions for mechanical configuration design problems, however, this approach does not work with rewards that require blackbox solvers in the loop. In concurrent work, e-SimFT [10] used preference data obtained from physics simulations to enhance the exploration of the Pareto front in a multi-preference setting, improving the solutions generated with GearFormer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 296, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 675 + ], + "type": "text", + "content": "Fine-tuning LLMs with RL Reinforcement Learning from Human Feedback (RLHF) has emerged as a cornerstone approach for aligning large language models (LLMs) with human preferences. In RLHF [4, 11, 26, 31], a learned reward model is usually trained to capture human preferences and is provides the learning signal that the policy is trained on. Other methods such as Direct Preference Optimization (DPO) [28] and Reinforced Self-Training (ReST) [36] use a simpler approach of optimizing the model directly from model generated data without the use of a learned reward model. While DPO learns from ranked pairs of generated model outputs from human annotators, ReST uses rejection sampling to remove incorrect generations and trains the model standard cross-entropy loss on the correct samples. In this paper, we refer to the approach of using fine-tuning on high return responses as Expert Iteration (ExIt). We broadly refer to algorithms that learn from ranked/filtered model outputs (such as DPO and Expert Iteration) as Preference Optimization (PO)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "More recently, a large body of work has focused on the task of teaching LLMs to solve reasoning tasks with reinforcement learning [14, 15, 20, 36], such as math and cod" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "ing, which can be checked with rule-based systems. Specifically we are inspired by approaches which forego the use of a learned reward model and directly learn from verifiable rewards. We build off of several approaches that have shown success when applied to reasoning LLMs – these include Group Relative Policy Optimization (GRPO) [35], ReMax [22], and Reinforce Leave-One-Out (RLOO) [1]. Both GRPO and RLOO estimate the baseline via the average reward of multiple sampled outputs instead of learned value model but differ in how they apply the KL divergence penalty, advantage normalization, and PPO-style reward clipping. ReMax [22] also obviates the need for a learned value model but instead estimates the baseline from the argmax result (greedy sampling)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 556, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 324 + ], + "type": "text", + "content": "In this paper, we adapt the aforementioned post-training methods—DPO, Expert Iteration, RLOO, GRPO, ReMax, for use in aligning a constraint generation model using feedback from a constraint solver. In Section 4 we provide additional details on the ranking/filtering criteria for the Preference Optimization (PO) algorithms and the reward design for the RL algorithms (GRPO, RLOO, ReMax)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 335, + 373, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 373, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 373, + 346 + ], + "type": "text", + "content": "3. Problem" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 450 + ], + "type": "text", + "content": "Sketch constraining is a fundamental component of parametric CAD modeling, where geometric relationships define the structure and behavior of the sketch. Applying constraints ensures stability and editability, allowing for parametric modifications that align with the design intent. Automating constraint generation requires producing a valid and efficient set of constraints that fully define a given sketch while avoiding unnecessary redundancy or conflicts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 556, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 557 + ], + "type": "text", + "content": "The sketch constraining problem can be formulated as a sequence modeling task similar to natural language generation, where constraints are predicted autoregressively. Given the sketch geometry as input, the model generates a sequence of constraints in the order they will be applied. Tokens in the sequence represent either a constraint (e.g., coincident, parallel, perpendicular), a dimension (e.g., horizontal, vertical, radial), or a pointer to one of the input geometric entities [38]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 689 + ], + "type": "text", + "content": "In parametric CAD, sketch geometry is modified using a constraint solver. The updated sketch respects any present constraints while moving the geometry to reflect changes to the dimension parameters. Unlike natural language, which is inherently sequential and follows flexible grammar rules, sketch constraints must adhere to strict geometric principles to ensure structural validity. A set of constraints may be incorrect for a variety of reasons: they can reference the wrong primitives for the constraint type, be redundant, specify inconsistent geometric relationships, or cause unexpected geometric distortions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "We formally define five conditions describing the state of a sketch after applying constraints. These conditions are" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8615" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 308 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 308 + ], + "type": "image", + "image_path": "db5bcb354e67d8d494bbb56eb5029dd1a66a5904cf42954d457e02fc405b279b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 316, + 296, + 406 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 406 + ], + "type": "text", + "content": "Figure 3. Comparing different outcomes when a designer updates a sketch parameter on the constrained sketch. First row: an underconstrained sketch only preserves a subset of the geometric relationships. Second row: the sketch is unstable, adding a coincident constraint flattens the geometry of the sketch. Third row: the sketch is over-constrained, causing the sketch to be uneditable. Fourth row: all geometric relationships are maintained after the parameter is updated." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 416, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 453 + ], + "type": "text", + "content": "not mutually exclusive. A sketch may satisfy one or more conditions depending on the applied constraints. Figure 3 provides illustrative examples of each condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 493 + ], + "type": "text", + "content": "Under-constrained (UC) A sketch containing primitives retains some unconstrained degrees of freedom, resulting in incomplete specification of their positions or dimensions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 534 + ], + "type": "text", + "content": "Fully-constrained (FC) A sketch in which all primitives have their degrees of freedom completely determined, removing positional or dimensional ambiguities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 537, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 296, + 597 + ], + "type": "text", + "content": "Over-constrained (OC) A sketch primitives have more constraints applied than degrees of freedom, potentially leading to conflicts. Note some over-constrained sketches remain solvable if the constraints are consistent and do not conflict with each other [7]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 638 + ], + "type": "text", + "content": "Not solvable A sketch that cannot achieve a valid solution due to contradictory or redundant constraints, leading to an impossible or conflicting geometry." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 700 + ], + "type": "text", + "content": "Stability We discretize the sketch plane into a grid and classify a sketch as unstable if the positions of primitives after constraint solving shift into different cells. The number of cells (bins) on each axis determines the sensitivity of this measurement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Our objective is to align the model toward generating" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 157 + ], + "type": "text", + "content": "constraint sets that yield fully-constrained sketches while minimizing cases of under-constrained, over-constrained, not solvable, or instability. This is a necessary prerequisite toward the ultimate goal of generating constraints that preserve the original sketch design intent when the designer varies the geometric parameters. A formal definition of FC, UC, and OC is described in [16, 17] and in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 167, + 370, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 167, + 370, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 167, + 370, + 179 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 188, + 554, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 554, + 260 + ], + "type": "text", + "content": "In this section, we outline the post-training techniques used for aligning a constraint generation model with feedback from a constraint solver. The approaches are grouped into three categories: supervised learning methods, preference-based optimization methods, and RL methods. Figure 4 illustrates the high-level workflow of this work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 268, + 420, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 420, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 420, + 279 + ], + "type": "text", + "content": "4.1. Constraint Solver" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 286, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 555, + 453 + ], + "type": "text", + "content": "We evaluate generated constraints using the commercially available constraint solver in Autodesk Fusion. This solver is treated as a black box, from which we retrieve sketch conditions as defined in Section 3. A general algorithmic approach commonly used in constraint solvers is described in [8]. The solver outputs the fully-constrained status for each entity if the sketch is solvable. Otherwise it classifies the sketch as over-constrained or unsolvable. The solver adjusts geometry to resolve all constraints which we compare with the original sketch to determine if the constraints caused geometry distortion (instability). On average, it takes 0.1-0.2 seconds to per solve, although complex cases may take tens of seconds. Any sketch taking longer than two seconds to solve is automatically deemed unsolvable." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 435, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 435, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 435, + 475 + ], + "type": "text", + "content": "4.2. Supervised Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 575 + ], + "type": "text", + "content": "We pre-train Vitruvion [33] as our base model using the same procedure described in their paper with a next-token prediction objective. Given an input sequence of geometric primitives, the model is trained to predict the next correct constraint or dimension based on the ground truth data. Further details about our implementation of the Vitruvion architecture and training procedure are provided in the appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 576, + 554, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 659 + ], + "type": "text", + "content": "Since the majority of the ground truth data contains under-constrained or over-constrained sketches, we additionally perform supervised fine-tuning (SFT). During SFT, the training data is limited to sketches verified by the constraint solver as solvable, fully-constrained, stable, and free of over-constrained conditions, ensuring the model explicitly learns from ideal examples." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 668, + 509, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 668, + 509, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 668, + 509, + 681 + ], + "type": "text", + "content": "4.3. Preference-Based Optimization (PO)" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Expert Iteration (ExIt) alternates between expert improvement and policy distillation. Following [15, 36]," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8616" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 265 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 265 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 265 + ], + "type": "image", + "image_path": "44f4104c4891fa907efad26cb8bd86f2e9232912c65e20c4d71d3ef6017c8820.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Illustration of the proposed alignment workflow for constraint generation models. Left: A base constraint-generation model is first fine-tuned using supervised fine-tuning (SFT). Middle: Generated constraint sequences " + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": " are evaluated using a constraint solver, which provides feedback on sketch stability and fully- and over-constrained statuses, forming the quality score vector " + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": ". Right: Two groups of alignment methods leveraging solver feedback: preference-based optimization (PO) and reinforcement learning (RL). PO uses " + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": " to construct training data, iteratively improving constraint prediction quality. RL methods assign per-constraint and per-sequence rewards to " + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": " based on solver feedback and " + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 55, + 273, + 555, + 340 + ], + "type": "text", + "content": " to incentivize the generation of constraints, making sketches stable and fully-constrained." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": "we use temperature sampling combined with rejection sampling to generate high-quality candidate constraint sequences. During the exploration step we initialize the policy model " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_{t = 0}}" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " from the SFT model and sample " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " candidate constraint sequences " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " at temperature " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "T = 1.0" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " for each sketch query " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " in the initial training set " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": ". A new training dataset " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{i\\in N}^{*}" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " is constructed by discarding sequences that are under-constrained, over-constrained, or unsolvable solutions. This process is repeated " + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 55, + 350, + 296, + 471 + ], + "type": "text", + "content": " times over the dataset, and the policy is trained using cross-entropy loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 481, + 295, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 481, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 120, + 481, + 295, + 496 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(q, \\tau) \\sim \\mathcal {D} _ {i \\in N} ^ {*}} [ \\log \\pi_ {\\theta_ {t}} (\\tau | q) ] \\tag {1}", + "image_path": "530803efd42dbcfe5f0dfdcf17d0a4979602406d0b702385b5640e27b2bdcc66.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_t}" + }, + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "text", + "content": " is updated after the distillation phase of every training round." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": "Direct Preference Optimization (DPO) learns from pairwise preference data, approximating an implicit reward via a reparameterized Bradley-Terry model [28]. Similar to ExIt, we construct the training dataset by sampling " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " constraint sequence completions " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " for each sketch query " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " from the policy model " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_t}" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " at temperature " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "T = 1.0" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": ". Pairs " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "(\\tau_w, \\tau_l)" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " are ranked based on the differences in the percentage of fully-constrained curves between " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\tau_w" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " (preferred sequence), and " + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\tau_l" + }, + { + "bbox": [ + 55, + 526, + 296, + 647 + ], + "type": "text", + "content": " (non-preferred sequence). The optimization objective is:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {(q, \\tau_ {w}, \\tau_ {l}) \\sim \\mathcal {D}} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {w} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {w} | q)} - \\beta \\log \\frac {\\pi_ {\\theta_ {t}} (\\tau_ {l} | q)}{\\pi_ {\\theta_ {r}} (\\tau_ {l} | q)}\\right) \\right] \\tag {2}", + "image_path": "7465499b7df10729ee7c3e6c208da66ce3cd386f6c385a96aebe53fcf73e5f06.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is a hyperparameter, and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is the logistic function. This formulation ensures the learned policy " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " aligns with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "text", + "content": "the ranking preference of fully-constrained sketches while maintaining proximity to the reference policy " + }, + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_r}" + }, + { + "bbox": [ + 313, + 350, + 555, + 374 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": "We initialize " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_{t = 0}}" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " from the SFT model and repeat the entire process " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "N = 2" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " times, updating " + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_t}" + }, + { + "bbox": [ + 313, + 374, + 554, + 422 + ], + "type": "text", + "content": " with the new policy after every training iteration. Additional hyperparameters and ranking details can be found in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 430, + 477, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 477, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 477, + 442 + ], + "type": "text", + "content": "4.4. Reinforcement Learning (RL)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 406, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 406, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 406, + 459 + ], + "type": "text", + "content": "4.4.1. Reward design" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 533 + ], + "type": "text", + "content": "Unlike natural language tasks, where human preferences are ambiguous and ill-defined, the stability and solvability of sketch constraints can be verified, making it compatible with RL methods that directly optimize for mechanically defined rewards without a learned preference model. We define the rewards used for RL as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 536, + 487, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 487, + 548 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 487, + 548 + ], + "type": "text", + "content": "- Rewards for valid constraint sequence " + }, + { + "bbox": [ + 314, + 536, + 487, + 548 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 314, + 536, + 487, + 548 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "spans": [ + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{curves}}(\\tau)" + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "content": " : " + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "content": " of fully-constrained curves over all curves, " + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{points}}(\\tau)" + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "content": " : " + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "content": " of fully-constrained points over all points, " + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{unstable}}" + }, + { + "bbox": [ + 321, + 552, + 552, + 597 + ], + "type": "text", + "content": " : penalty for unstable sketches," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 601, + 487, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 601, + 487, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 601, + 487, + 613 + ], + "type": "text", + "content": "- Rewards for invalid constraint sequence:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 617, + 481, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 617, + 481, + 629 + ], + "spans": [ + { + "bbox": [ + 322, + 617, + 481, + 629 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{NS}}" + }, + { + "bbox": [ + 322, + 617, + 481, + 629 + ], + "type": "text", + "content": " : penalty for not solvable sketches," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 322, + 633, + 500, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 633, + 500, + 645 + ], + "spans": [ + { + "bbox": [ + 322, + 633, + 500, + 645 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{OC}}" + }, + { + "bbox": [ + 322, + 633, + 500, + 645 + ], + "type": "text", + "content": " : penalty for over-constrained sketches," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 322, + 650, + 528, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 650, + 528, + 662 + ], + "spans": [ + { + "bbox": [ + 322, + 650, + 528, + 662 + ], + "type": "inline_equation", + "content": "r_{\\mathbf{F}}" + }, + { + "bbox": [ + 322, + 650, + 528, + 662 + ], + "type": "text", + "content": " : penalty for sketches resulting in other failures." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "The overall sequence-wise reward " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "R(\\tau)" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " is the sum of " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{curves}}(\\tau), r_{\\mathrm{points}}(\\tau)" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ", and conditionally " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{unstable}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " for valid sketches, and either " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{NS}}, r_{\\mathrm{OC}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " for invalid sketches according to the failure mode." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8617" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "We additionally define a constraint-wise penalty to provide granular feedback on cases where the constraint sequence causes sketches to be over-constrained or fully-constrained. A constraint solver iteratively attempts to add each generated constraint one-by-one, dropping any problematic constraints that caused the sketch to be over-constrained or not-solvable. In training, we add a constant of -1 loss penalty directly to the per-token log likelihood loss for the problematic constraints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 187, + 200, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 200, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 200, + 198 + ], + "type": "text", + "content": "4.4.2. RL fine-tuning formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": "We formulate constraint generation fine-tuning as follows; given a dataset of sketch queries " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "D = \\{q_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": " and reward function " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "R(\\tau)" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": " using the constraint solver and reward design in Section 4.4.1, learn a policy " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\tau|q)" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": " that generates a sequence of constraints " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": " for sketch " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": ", such that it maximizes the expected rewards " + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{q_i \\sim D, \\tau_i \\sim \\pi_{\\theta}(\\cdot | q_i)}[R(\\tau_i)]" + }, + { + "bbox": [ + 55, + 201, + 296, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 187, + 292 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 187, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 187, + 292 + ], + "type": "text", + "content": "4.4.3. Policy gradient methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 416 + ], + "type": "text", + "content": "For RLHF which uses a pre-trained policy, not all the complexity of RL algorithms is necessary. This allows the algorithms to be simplified and the number of learnable components to be reduced, contributing to performance improvement. We considered three policy gradient algorithms: ReMax [22], RLOO [1], and GRPO [35]. Unlike PPO [31], which treats each token generation as an action, these algorithms treat generation of a sequence as a single action and adopt a REINFORCE with baselines approach [39]. We apply these to optimize the constraint generation policy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 456 + ], + "type": "text", + "content": "ReMax [22] uses the rewards corresponding to sequences generated by a greedy (argmax) policy as a baseline to normalize the rewards of sequences sampled from the policy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": "Considering sequence " + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": " sampled from policy " + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\tau |q)" + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": ", sequence " + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "inline_equation", + "content": "\\tau^{*} = \\operatorname{argmax}_{\\tau}\\pi_{\\theta}(\\tau |q)" + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": " greedily sampled by taking an argmax of the policy, and corresponding rewards " + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "inline_equation", + "content": "r^{*}" + }, + { + "bbox": [ + 55, + 456, + 296, + 505 + ], + "type": "text", + "content": ", the policy gradient objective for ReMax is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 516, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 516, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 116, + 516, + 295, + 534 + ], + "type": "interline_equation", + "content": "\\underset {\\tau \\sim \\pi} {\\mathbb {E}} \\left[ \\left(r - r ^ {*}\\right) \\nabla \\log \\pi_ {\\theta} (\\tau | q) \\right]. \\tag {3}", + "image_path": "83be57720536f4cd38d1dc4e3f4ab53e10931c1d0638df83dd875ff788e3a035.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "text", + "content": "REINFORCE-Leave-One-Out (RLOO) [1] samples " + }, + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 544, + 296, + 591 + ], + "type": "text", + "content": " number of constraint sequences for every sketch query. The baseline for each sample is evaluated as the mean of the rewards for all other samples in the group." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "content": " number of sequences " + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\{\\tau_g\\}_{g = 1}^G" + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "content": " sampled from policy " + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\tau |q)" + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "content": " for a given sketch query " + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 592, + 296, + 628 + ], + "type": "text", + "content": ", the policy gradient objective of RLOO is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 64, + 645, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 645, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 64, + 645, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\underset {\\{\\tau \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\frac {1}{G} \\sum_ {g = 1} ^ {G} \\left[ \\left(r _ {g} - \\operatorname {m e a n} \\left(\\left\\{r _ {i} \\right\\} _ {i \\neq g} ^ {G}\\right)\\right) \\nabla \\log \\pi_ {\\theta} \\left(\\tau_ {g} \\mid q _ {g}\\right) \\right] \\right]. \\tag {4}", + "image_path": "8a480a8c7fd4bb81a05275035e313a7a0fb6a2d3004f851189af65f1851f53ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Group Relative Policy Optimization (GRPO) [35] uses group-based baseline estimation, like RLOO, but the mean" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "is taken over all reward samples in the group. It uses a clipped policy optimization objective similarly to PPO [31], as well as a low-variance KL regularization term." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": " number of sequences " + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\{\\tau_g\\}_{g = 1}^G" + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": " sampled from the reference policy " + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta_{\\mathrm{r}}}(\\tau |q)" + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": " for a given sketch query " + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": ", letting " + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\rho_{g} = \\frac{\\nabla\\pi_{\\theta}(\\tau_{g}|q)}{\\pi_{\\theta_{\\mathrm{r}}}(\\tau_{g}|q)}" + }, + { + "bbox": [ + 313, + 107, + 553, + 149 + ], + "type": "text", + "content": ", the optimization objective of GRPO is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 154, + 561, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 154, + 561, + 177 + ], + "spans": [ + { + "bbox": [ + 315, + 154, + 561, + 177 + ], + "type": "interline_equation", + "content": "\\underset {\\{\\tau_ {g} \\} \\sim \\pi} {\\mathbb {E}} \\left[ \\min \\left(\\rho_ {g} A _ {g}, \\operatorname {c l i p} \\left(\\rho_ {g}, 1 - \\epsilon , 1 + \\epsilon\\right) A _ {g}\\right) - \\beta \\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} | | \\pi_ {\\theta_ {\\mathrm {r}}}\\right) \\right],", + "image_path": "a51bbb5b3fbf3aca4a09f4be0da6c8e929937d06a7386b46125ecc90f9205a0a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 179, + 558, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 179, + 558, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 179, + 558, + 217 + ], + "type": "interline_equation", + "content": "\\mathbb {D} _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} \\| \\pi_ {\\theta_ {\\mathrm {r}}}\\right) = \\frac {1}{\\rho_ {g}} + \\log \\rho_ {g} - 1, \\quad A _ {g} = \\frac {r _ {g} - \\operatorname* {m e a n} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}{\\operatorname {s t d} \\left(\\left\\{r _ {g} \\right\\} _ {g = 1} ^ {G}\\right)}, \\tag {5}", + "image_path": "5a53e308fae54ec10b1efe61e5ca86907ea7b35548849845bceb89b8507c7560.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 223, + 553, + 247 + ], + "type": "text", + "content": " are hyper-parameters for the clipped policy optimization and KL regularization terms, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "content": "For ReMax and RLOO, we also added a small KL penalty term to the rewards to discourage divergence from the reference policy. GRPO applies group-normalization on the advantage, which we also applied in RLOO. For ReMax, we batch-normalized the advantage. For all algorithms, " + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 247, + 554, + 331 + ], + "type": "text", + "content": " is initialized from the SFT model. Further implementation details of the algorithms can be found in the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 394, + 354 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 361, + 372, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 372, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 372, + 372 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": "We train our models on SketchGraphs [32], a large-scale dataset of CAD sketches created in Onshape. SketchGraphs captures real-world parametric modeling workflows, providing geometry and constraint construction operations from actual design steps. However, the dataset was not originally designed for direct constraint inference; only " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "8.27\\%" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": " of its sketches are fully-constrained, making it imperfect for training the constraint generation model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 474, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 581 + ], + "type": "text", + "content": "For computational feasibility, we deduplicate and filter sketches, retaining only those with at most 16 geometric primitives and 64 constraints, yielding a dataset of 2.8 million unique sketches. Certain constraint types, such as symmetry, are excluded to simplify the learning task, ensuring the focus remains on constraints most relevant to engineering design. We also convert Onshape sketches into the Fusion sketch format to utilize the solver in Fusion. Additional details are provided in the appendix." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 665 + ], + "type": "text", + "content": "Another challenge lies in how SketchGraphs positions primitives. Rather than being placed in valid, constraintsatisfying layouts, primitives often have arbitrary coordinates that do not reflect a solved state. We therefore preprocess the dataset using Fusion to resolve each sketch's primitives according to its constraints, ensuring that geometry and constraints match before training." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 672, + 433, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 433, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 433, + 684 + ], + "type": "text", + "content": "5.2.Quantitative Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "In Table 1, we list results comparing the performance of each alignment method with respect to the five sketch condi" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8618" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 113, + 500, + 215 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Sketch constraint generation results for Fully Constrained (FC), Under Constrained (UC), Over Constrained (OC), not solvable, and stability for the base model, SFT model, and aligned models. Results are computed over 8 samples per sketch with a temperature of 1.0. Numbers following " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " indicate the standard deviation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 110, + 113, + 500, + 215 + ], + "lines": [ + { + "bbox": [ + 110, + 113, + 500, + 215 + ], + "spans": [ + { + "bbox": [ + 110, + 113, + 500, + 215 + ], + "type": "table", + "html": "
Model% FC ↑% UC ↓% OC ↓% Not solvable ↓% Stable (bins=4) ↑
Vitruvion (base)8.87 ±0.0971.38 ±0.2016.83 ±0.173.05 ±0.0392.15 ±0.06
SFT34.24 ±0.0946.61 ±0.1515.30 ±0.083.85 ±0.0392.48 ±0.05
Iterative DPO64.91 ±0.1114.97 ±0.1312.47 ±0.097.64 ±0.0787.63 ±0.09
Expert Iteration71.70 ±0.1313.38 ±0.137.25 ±0.067.67 ±0.0785.50 ±0.10
ReMax79.84 ±0.0915.86 ±0.071.49 ±0.012.82 ±0.0275.77 ±0.05
RLOO93.05 ±0.033.55 ±0.022.15 ±0.011.25 ±0.0189.16 ±0.02
GRPO91.59 ±0.034.18 ±0.031.94 ±0.012.28 ±0.0288.28 ±0.03
", + "image_path": "a4fc8240b5e47ffd1720e8ee0bab33703655e8ed51b962dd96783629da456c48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "content": "tions described in Section 3. The base model is able to fully-constrain sketches only " + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "inline_equation", + "content": "8.87\\%" + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "content": " of the time, consistent with the dataset distribution where only " + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "inline_equation", + "content": "8.27\\%" + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "content": " of sketches are fully constrained. We find that RLOO and GRPO perform similarly, giving the best performance at fully-constraining sketches " + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "inline_equation", + "content": "93.05\\%" + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "inline_equation", + "content": "91.59\\%" + }, + { + "bbox": [ + 54, + 233, + 294, + 341 + ], + "type": "text", + "content": " of the time, respectively. They have the lowest indices of over-constrained or unsolvable results and maintain stability rates that are on par with other methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": "Iterative DPO and ExIt significantly improve upon the base and SFT models but still fall short of the performance achieved by policy gradient-based RL methods. We attribute this gap to the online nature of policy gradient-based RL, which continuously refines the policy through feedback while actively exploring a broader range of solutions. In contrast, Iterative DPO and ExIt are offline methods and rely on predefined ranking and filtering signals to generate training data, which limits their ability to explore the solution space. The superior performance of online RL underscores its advantage in directly optimizing the shaped rewards from the constraint solver." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": "Table 2 demonstrates the evaluation result of our methods in a few-shot inference setting, where the model has " + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 487, + 295, + 536 + ], + "type": "text", + "content": " attempts to generate a fully-constrained and stable output. This scenario reflects real-world use cases where the" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 97, + 609, + 252, + 712 + ], + "blocks": [ + { + "bbox": [ + 55, + 557, + 295, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 601 + ], + "type": "text", + "content": "Table 2. Sketch constraint generation results for Pass@1 and Pass@8 across the post-training algorithms. We define a successful result as fully-constrained, not over-constrained, solvable, and stable at 4 bins. Results are generated with temperature of 1.0." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 609, + 252, + 712 + ], + "lines": [ + { + "bbox": [ + 97, + 609, + 252, + 712 + ], + "spans": [ + { + "bbox": [ + 97, + 609, + 252, + 712 + ], + "type": "table", + "html": "
ModelPass@1Pass@8
Vitruvion (base)8.5320.47
SFT33.3242.62
Iterative DPO59.3868.32
Expert Iteration64.0972.56
ReMax62.7465.89
RLOO83.5784.96
GRPO81.4983.42
", + "image_path": "265f2da3f9858e40fad653ace59fb9346d9fc7652f08445447b755893b6b0a62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "content": "goal is to maximize the likelihood of producing an acceptable solution within a fixed inference budget of " + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "content": " samples (Pass@ " + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "content": "). We find that while the RL-based methods still have the overall highest performance, increasing the number of samples " + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "content": " has a comparatively small impact on performance compared to the other methods. Additional results and analysis on the impact of sampling parameters such as temperature and top- " + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 233, + 555, + 331 + ], + "type": "text", + "content": " are available in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 337, + 426, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 426, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 426, + 350 + ], + "type": "text", + "content": "5.3. Qualitative Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "type": "text", + "content": "We randomly select sketches from the test set and show the results generated by different alignment methods in Figure 5. Curves are colored black when constrained and blue when not. Methods leveraging alignment techniques demonstrate a promising trend towards generating fully-constrained sketches, whereas the base Vitruvion and SFT models typically leave sketches under-constrained. However, across different alignment algorithms, we observe substantial variance in the degree of geometric distortion introduced by the aligned models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "type": "text", + "content": "Specifically, columns A and B depict simple sketches mainly composed of horizontal and vertical lines, for which all solver-feedback methods consistently yield fully-constrained, stable results. In contrast, Column C presents a challenging sketch due to the absence of appropriate constraints for oblique lines. Achieving stability in this scenario typically requires many dimensions but few constraints; however, models are optimized to generate more constraints to align with parametric CAD design principles." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "content": "Sketches in columns D through H include arcs. Column E is particularly notable as all solver-feedback methods achieve a fully-constrained condition, yet only RLOO produces visually stable results. Further examination reveals that distortions induced by other models were still classified as stable due to our bin size (" + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "inline_equation", + "content": "bins = 4" + }, + { + "bbox": [ + 313, + 582, + 556, + 700 + ], + "type": "text", + "content": "). Column G presents unique challenges with an isolated point, causing confusion in the model and demanding extensive use of non-horizontal/vertical constraints to make the sketch fully-constrained and stable." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Our results indicate that models more easily fully-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8619" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 553, + 386 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 553, + 386 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 553, + 386 + ], + "type": "image", + "image_path": "a0c5fa2a1e2d2718102b4c6d50fcd713bcc34b8cd20e678f92ea061468739bf1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 395, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 555, + 418 + ], + "type": "text", + "content": "Figure 5. Visual comparison of solved sketches across the aforementioned baseline and post-training methods. Curves are colored black when fully-constrained and blue when not. Overlaid red lines are the original input sketch for reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 429, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 295, + 513 + ], + "type": "text", + "content": "constrain sketches consisting primarily of horizontal and vertical lines without distortion. In contrast, sketches involving oblique lines increase the challenge of maintaining stability, and arcs further complicate achieving fully-constrained status. Among the tested methods, RLOO and GRPO exhibit the strongest overall performance in achieving fully-constrained status and geometric stability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 636 + ], + "type": "text", + "content": "Upon closer inspection of the generated constraints for the RL-based methods, we find that these models learn to \"reward hack\" by adding far more dimensions than constraints. This allows the model to avoid the instability penalty because dimensions cannot cause geometry distortion. In order to counteract this behavior, we added an additional reward penalty on the ratio of constraints versus dimensions which fixes the over-dimensioning problem at the cost of reducing stability and fully-constrained status. Additional details are provided in the appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "To further evaluate the efficacy of each method, we conduct a forced-choice perceptual study with professional CAD designers. We show them paired images of the same sketch with different constraints applied from each method, to understand which set of constraints they prefer. We find that the preference-based optimization methods (ExIT," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 429, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 464 + ], + "type": "text", + "content": "DPO) and RLOO with the over-dimensioning penalty consistently outperform the SFT and vanilla RLOO results. Results of this study are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 389, + 491 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 499, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 595 + ], + "type": "text", + "content": "Our experiments focus on sketches of moderate complexity, with a maximum of 16 geometric primitives and 64 constraints. Evaluating performance on more complex sketches requires further exploration. Additionally, our alignment approach only uses feedback signals from a constraint solver to represent design intent. Incorporating subjective human preferences and explicit design intent into alignment objectives remains a promising area for future work." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 609, + 388, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 388, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 388, + 620 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "We demonstrated the adaption of alignment strategies from language modeling to preserve design intent in parametric CAD sketches. By using feedback from a constraint solver as a learning signal, we show the feasibility and value of alignment in parametric CAD tasks. In doing so, we pave the way for future AI-assisted design tools that incorporate design alignment." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8620" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "We thank William McCarthy for his valuable contribution to the human evaluation experiments, including setup, execution, and analysis. His efforts were essential in validating the effectiveness of our proposed methods from a user-centric perspective." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 236 + ], + "type": "text", + "content": "[1] Arash Ahmadian, Chris Cremer, Matthias Galle, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, and Sara Hooker. Back to basics: Revisiting reinforce style optimization for learning from human feedback in llms. arXiv preprint arXiv:2402.14740, 2024. 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 269 + ], + "type": "text", + "content": "[2] Thomas W. Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In *Neural Information Processing Systems*, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 218, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 218, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 218, + 280 + ], + "type": "text", + "content": "[3] Autodesk. Sketches in Fusion, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 335 + ], + "type": "text", + "content": "[4] Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, John Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "type": "text", + "content": "[5] Bernhard Bettig and Christoph M. Hoffmann. Geometric constraint solving in parametric computer-aided design. Journal of Computing and Information Science in Engineering, 11(2):021001, 2011. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 413 + ], + "type": "text", + "content": "[6] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 446 + ], + "type": "text", + "content": "[7] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-Aided Design, 27(6):487-501, 1995. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 447, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 480 + ], + "type": "text", + "content": "[8] William Bouma, Ioannis Fudos, Christoph Hoffmann, Ji-zhen Cai, and Robert Paige. Geometric constraint solver. Computer-aided design, 27(6):487-501, 1995. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 523 + ], + "type": "text", + "content": "[9] Wei Chen, Kevin Chiu, and Mark D Fuge. Airfoil design parameterization and optimization using bezier generative adversarial networks. AIAA journal, 58(11):4723-4735, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "text", + "content": "[10] Hyunmin Cheong, Mohammad Mehdi Ataei, Amir Hosein Khasahmadi, and Pradeep Kumar Jayaraman. e-simft: Alignment of generative models with simulation feedback for pareto-front design exploration. arXiv preprint arXiv:2502.02628, 2025.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 613 + ], + "type": "text", + "content": "[11] Paul F. Christiano, Jan Leike, Tom B. Brown, et al. Deep reinforcement learning from human preferences. In NeurIPS, 2017. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "text", + "content": "[12] Yasaman Etesam, Hyunmin Cheong, Mohammad Mehdi Ataei, and Pradeep Kumar Jayaraman. Deep generative model for mechanical system configuration design. arXiv preprint arXiv:2409.06016, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "text", + "content": "[13] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. NeurIPS, 34:5885-5897, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Jonas Gehring, Kunhao Zheng, Jade Copet, Vegard Mella, Taco Cohen, and Gabriel Synnaeve. Rlef: Grounding code" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "llms in execution feedback with reinforcement learning. arXiv preprint arXiv:2410.02089, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 160 + ], + "type": "text", + "content": "[15] Alex Havrilla, Yuqing Du, Sharath Chandra Rarparthy, Christoforos Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, Sainbayar Sukhbaatar, and Roberta Raileanu. Teaching large language models to reason with reinforcement learning. arXiv preprint arXiv:2403.04642, 2024. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 204 + ], + "type": "text", + "content": "[16] Christoph M Hoffmann and Robert Joan-Arinyo. Symbolic constraints in constructive geometric constraint solving. Journal of Symbolic Computation, 23(2-3):287-299, 1997. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 205, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 237 + ], + "type": "text", + "content": "[17] Christoph M Hoffmann and Robert Joan-Arinyo. A brief on constraint solving. Computer-Aided Design and Applications, 2(5):655-663, 2005. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[18] Benjamin T Jones, Felix Hahnlein, Zihan Zhang, Maaz Ahmad, Vladimir Kim, and Adriana Schulz. A solver-aided hierarchical language for ltm-driven cad design. arXiv preprint arXiv:2502.09819, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 282, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 335 + ], + "type": "text", + "content": "[19] Ahmet Serdar Karadeniz, Dimitrios Mallis, Nesryne Mejri, Kseniya Cherenkova, Anis Kacem, and Djamila Aouada. Davinci: A single-stage architecture for constrained cad sketch inference. arXiv preprint arXiv:2410.22857, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "type": "text", + "content": "[20] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\\”ulu 3: Pushing frontiers in open language model posttraining. arXiv preprint arXiv:2411.15124, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 393, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 447 + ], + "type": "text", + "content": "[21] Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text-to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[22] Ziniu Li, Tian Xu, Yushun Zhang, Zhihang Lin, Yang Yu, Ruoyu Sun, and Zhi-Quan Luo. Remax: A simple, effective, and efficient reinforcement learning method for aligning large language models. In ICML, 2024. 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 514 + ], + "type": "text", + "content": "[23] Dave Martin. What is design intent?, 2023. Accessed: January 19, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[24] Nelson Nauata, Kai-Hung Chang, Chin-Yi Cheng, Greg Mori, and Yasutaka Furukawa. House-gan: Relational generative adversarial networks for graph-constrained house layout generation. In ECCV, pages 162-177. Springer, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[25] Jeffrey Otey, Pedro Company, Manuel Contrero, and Jorge D. Camba. Revisiting the design intent concept in the context of mechanical cad education. Computer-Aided Design and Applications, 15(1):47-60, 2018. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 646 + ], + "type": "text", + "content": "[26] Long Ouyang, Jeff Wu, Xu Jiang, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[27] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. Sketchgen: Generating constrained cad sketches. NeurIPS, 34:5077-5088, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8621" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "preference optimization: Your language model is secretly a reward model. NeurIPS, 36:53728-53741, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[29] Lyle Regenwetter, Amin Heyrani Nobari, and Faziz Ahmed. Deep generative models in engineering design: A review. Journal of Mechanical Design, 144(7):071704, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "type": "text", + "content": "[30] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 3rd edition, 2010. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 186 + ], + "type": "text", + "content": "[31] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "type": "text", + "content": "[32] Ari Seff, Yaniv Ovadia, Wenda Zhou, and Ryan P. Adams. SketchGraphs: A large-scale dataset for modeling relational geometry in computer-aided design. In ICML 2020 Workshop on Object-Oriented Learning, 2020. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 265 + ], + "type": "text", + "content": "[33] Ari Seff, Wenda Zhou, Nick Richardson, and Ryan P Adams. Vitruvion: A generative model of parametric cad sketches. In ICLR, 2021. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 266, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 311 + ], + "type": "text", + "content": "[34] Mohammad Amin Shabani, Sepidehsadat Hosseini, and Yasutaka Furukawa. Housediffusion: Vector floorplan generation via a diffusion model with discrete and continuous denoising. In CVPR, pages 5466-5475, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 312, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 294, + 367 + ], + "type": "text", + "content": "[35] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 369, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 294, + 521 + ], + "type": "text", + "content": "[36] Avi Singh, John D Co-Reyes, Rishabh Agarwal, Ankesh Anand, Piyush Patil, Xavier Garcia, Peter J Liu, James Harrison, Jaehoon Lee, Kelvin Xu, Aaron T Parisi, Abhishek Kumar, Alexander A Alemi, Alex Rizkowsky, Azade Nova, Ben Adlam, Bernd Bohnet, Gamaleldin Fathy Elsayed, Hanie Sedghi, Igor Mordatch, Isabelle Simpson, Izzeddin Gur, Jasper Snoek, Jeffrey Pennington, Jiri Hron, Kathleen Kenealy, Kevin Swersky, Kshitteej Mahajan, Laura A Culp, Lechao Xiao, Maxwell Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Transactions on Machine Learning Research, 2024. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 555 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[38] Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume 2, page 2692-2700, Cambridge, MA, USA, 2015. MIT Press. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 668 + ], + "type": "text", + "content": "[39] Lex Weaver and Nigel Tao. The optimal reward baseline for gradient-based reinforcement learning. In Proceedings of the Seventeenth Conference on Uncertainty in Artificial Intelligence, page 538-545, San Francisco, CA, USA, 2001. Morgan Kaufmann Publishers Inc. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Karl DD Willis, Pradeep Kumar Jayaraman, Joseph G Lambourne, Hang Chu, and Yewen Pu. Engineering sketch generation for computer-aided design. In CVPRW, pages 2105–2114, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 307 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "text", + "content": "[41] Karl D. D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, and Wojciech Matusik. Fusion 360 gallery: A dataset and environment for programmatic cad construction from human design sequences. ACM TOG, 40(4), 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "type": "text", + "content": "[42] Sifan Wu, Amir Hosein Khasahmadi, Mor Katz, Pradeep Kumar Jayaraman, Yewen Pu, Karl Willis, and Bang Liu. Cadvlm: Bridging language and vision in the generation of parametric cad sketches. In ECCV, pages 368-384. Springer, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[43] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36:15903-15935, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "text", + "content": "[44] Xiang Xu, Pradeep Kumar Jayaraman, Joseph G Lambourne, Karl DD Willis, and Yasutaka Furukawa. Hierarchical neural coding for controllable cad model generation. In ICML, pages 38443-38461, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 285, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 553, + 307 + ], + "type": "text", + "content": "[45] Yuezhi Yang and Hao Pan. Discovering design concepts for cad sketches. NeurIPS, 35:28803-28814, 2022. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8622" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_content_list.json b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9298b6b53b5eaa72e9aa84820658ccb1ea0511e5 --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_content_list.json @@ -0,0 +1,1737 @@ +[ + { + "type": "text", + "text": "Aligning Effective Tokens with Video Anomaly in Large Language Models", + "text_level": 1, + "bbox": [ + 125, + 130, + 872, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingxian Chen $^{1*}$ Jiahui Liu $^{1*}$ Ruidi Fan $^{1}$ Yanwei Li $^{2}$ Chirui Chang $^{1}$ Shizhen Zhao $^{1}$ \nWilton W.T.Fok $^{1}$ Xiaojuan Qi $^{1\\dagger}$ Yik-Chung Wu $^{1\\dagger}$ $^{1}$ The University of Hong Kong $^{2}$ The Chinese University of Hong Kong \n{chenyx, liujh, xjqi, ycuw}@eee.hku.hk \n*equal contributions †corresponding authors", + "bbox": [ + 142, + 178, + 849, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding abnormal events in videos is a vital and challenging task that has garnered significant attention in a wide range of applications. Although current video understanding Multi-modal Large Language Models (MLLMs) are capable of analyzing general videos, they often struggle to handle anomalies due to the spatial and temporal sparsity of abnormal events, where the redundant information always leads to suboptimal outcomes. To address these challenges, exploiting the representation and generalization capabilities of Vison Language Models (VLMs) and Large Language Models (LLMs), we propose VA-GPT, a novel MLLM designed for summarizing and localizing abnormal events in various videos. Our approach efficiently aligns effective tokens between visual encoders and LLMs through two key proposed modules: Spatial Effective Token Selection (SETS) and Temporal Effective Token Generation (TETG). These modules enable our model to effectively capture and analyze both spatial and temporal information associated with abnormal events, resulting in more accurate responses and interactions. Furthermore, we construct an instruction-following dataset specifically for fine-tuning video-anomaly-aware MLLMs, and introduce a cross-domain evaluation benchmark based on XD-Violence dataset. Our proposed method outperforms existing state-of-the-art methods on various benchmarks.", + "bbox": [ + 88, + 335, + 485, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 724, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Detecting and summarizing abnormal events in videos is critical and challenging, and it has garnered considerable attention across multiple research domains and real-world applications, such as security monitoring, video analysis, and crime detection.", + "bbox": [ + 89, + 750, + 483, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although many traditional methods [4, 16, 37, 43, 47, 73, 86] have been widely explored for video anomaly detection, they exhibit substantial limitations in their effectiveness [22, 51, 60, 66, 71, 84]. These limitations manifest in two aspects: 1) Traditional video anomaly detection", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa9e6ed126b55907fe3f366578df2edc7f5d3609e862de6448b78628b49de336.jpg", + "image_caption": [ + "Figure 1. Baseline video understanding MLLM feeds forward every visual token (yellow squares) equally to participate in fine-tuning and inference (top row). Different from it, our method focuses on the effective area (unobstructed area in medium video frames) in each frame and select the Spatial Effective Tokens (orange squares) for the LLM (see Section 3.2) (filtered tokens are shown as gray squares). At the same time, we generate anomaly-aware Temporal Effective Tokens (green squares) (see Section 3.3) based on the assigned anomaly scores (denoted as $s$ ) of each frame from a pretrained classifier for better temporal localization of anomalies." + ], + "image_footnote": [], + "bbox": [ + 519, + 306, + 898, + 506 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods [6, 13, 57, 63, 64, 86] essentially approach the task as a closed-set detection and classification problem, inherently limiting their ability to achieve a comprehensive understanding and interpretation of anomalies; 2) These methods [2, 19, 23, 65, 68, 78] are restricted by a limited vocabulary, making it difficult for them to handle unseen or novel situations effectively.", + "bbox": [ + 511, + 672, + 908, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements [1, 28, 34, 36, 45, 58] in Vision Language Models (VLMs) and Large Language Models (LLMs) have demonstrated remarkable capabilities in scene understanding and comprehensive analysis. Multimodal Large Language Models (MLLMs), particularly those designed for video understanding [27, 29, 31, 39, 42, 74], have achieved significant progress in general video analysis tasks. However, while these models exhibit strong performance", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22695", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in general video understanding, they fall short in accurately detecting and interpreting anomalies.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For mitigating the above challenges, some works [12, 55, 67, 75, 79] proposed anomaly-aware video MLLMs to better understand the anomalies in videos. Although these models work well for detecting obvious abnormal events, such as fighting or fire, they typically struggle to effectively align abnormal regions with relevant captions which requires addressing spatial redundancy, and accurately identifying abnormal time intervals by mitigating temporal redundancy. This is because these methods treat all latent tokens with equal priority across spatial and temporal dimensions. This leads to performance degradation caused by redundant tokens unrelated to anomalies. However, in most cases, only small regions within a few frames contain the essential information help to identify an anomaly (as shown in Figure 1). Thus, we explore: How can multimodal architectures evolve selective token generation and processing mechanisms to dynamically prioritize anomaly-salient information while preserving comprehensive scene understanding capabilities?", + "bbox": [ + 91, + 125, + 485, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose a new model named VA-GPT for analyzing various Videos for Abnormal events by aligning effective and accurate tokens with LLMs across both spatial and temporal dimensions. VA-GPT integrates two key components to identify effective visual tokens for alignment while eliminating redundant tokens that could hinder anomaly analysis and distract model from extracting useful information: 1) we develop the Spatial Effective Token Selection (SETS) module for identifying tokens corresponding to regions with challenges for aligning them with LLMs, while filtering out tokens associated with minor dynamics to remove redundancy. This is because we find that abnormal events often result in different visual changes and variations in local areas (see Figure 1); and 2) we propose the Temporal Effective Token Generation (TETG) module which employs a lightweight pre-trained classifier to assign a confidence score to each frame indicating the possibility of containing abnormal events. Then TETG generates efficient tokens with prior knowledge of the temporal information of abnormal events directly in the language space as additional input to the LLMs, effectively enhancing the model's temporal reasoning and understanding abilities about abnormal events.", + "bbox": [ + 91, + 398, + 483, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, beyond conventional benchmarks (domain benchmark), we establish a new cross-domain evaluation protocol that systematically evaluates model robustness with domain shifts. Based on a novel video dataset, XD-Violence [64], we design comprehensive QAs about abnormal events which include different visual contents from our training data and integrate it as a new cross-domain benchmark. Meanwhile, we design temporal-information-oriented QAs on both in- and cross- domain benchmarks for evaluating temporal localization abilities. Comprehensive", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "experiments demonstrate VA-GPT's superiority, achieving state-of-the-art performance in both in-domain anomaly localization and cross-domain generalization scenarios.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 532, + 137, + 872, + 150 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose VA-GPT, a video-anomaly-aware MLLM for detecting and summarizing anomalies in various videos, which introduces the MLLM to the specific domain of video anomaly understanding.", + "- We introduce the SETS and TETG, which enable our MLLM to effectively capture both spatial and temporal information in video sequences, resulting in accurate understanding and localization of abnormal events. Meanwhile, we propose a new instruct-following dataset for video anomaly analysis and a comprehensive cross-domain evaluation benchmark for better evaluating the generalization abilities of MLLMs on video anomalies.", + "- Our extensive experiments demonstrate that our method outperforms existing state-of-the-art methods in various benchmarks, highlighting its effectiveness and potential for practical applications in video anomaly understanding." + ], + "bbox": [ + 511, + 155, + 906, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 412, + 653, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large Language Models (LLMs)", + "text_level": 1, + "bbox": [ + 511, + 436, + 800, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The domain of Natural Language Processing (NLP) has experienced significant progress, particularly with the emergence of Large Language Models (LLMs). The introduction of the Transformer architecture [3, 18, 59] is a critical turning point, followed by other influential language models [3, 10, 25, 80] that exhibited remarkable proficiency. Generative Pre-trained Transformers (GPT) [49] brought about a revolution in NLP by employing auto-regressive prediction, establishing itself as a powerful language modeling approach. More recent groundbreaking contributions, such as ChatGPT [45], GPT-4 [1], LaMDA [56] and LLaMA [58], have expanded the horizon even further. These models, trained on extensive textual data, display extraordinary performance in intricate linguistic tasks.", + "bbox": [ + 511, + 459, + 908, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision Language Models (VLMs)", + "text_level": 1, + "bbox": [ + 511, + 681, + 805, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Progress in the fields of computer vision and natural language processing has given rise to the development of vision-language models (VLMs) [14, 21, 33, 34, 50, 61, 69]. These models combine visual and linguistic systems to facilitate cross-modal comprehension and reasoning. Notable examples include CLIP [50] which pairs BERT [10] with ViT [11]; BLIP-2 [28] incorporating Vision Transformer features into Flan-T5 [8]; MiniGPT4 [85], connecting BLIP-2 with Vicuna [28, 48]; PandaGPT [53], bridging ImageBind [15] with Vicuna. These models excel in tasks like image classification, captioning, and object detection [24, 53, 83]. Recent developments in vision-language models have extended into video processing with models", + "bbox": [ + 511, + 704, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f228863632d5c7156edebc26db7fe2beb86e0b1ee0ef22bd07d348a1bed38767.jpg", + "image_caption": [ + "Figure 2. Detailed illustration of our proposed model. When a video is fed into the model, patch embeddings $\\square$ and class embeddings (c.ebd) $\\square$ are extracted from all frames. 1) Based on the difference in patch embeddings between current frame and its neighbour frame, we can get a filter mask $\\square$ to filter out unimportant visual tokens (dashed square $\\square$ ) from current frame's visual tokens $\\square$ , thereby selecting Spatial Effective Tokens $\\square$ that are compressed with a projector with pooling into aligned content token $\\square$ for each frame, meanwhile take attention with text input from users for resulting aligned context token $\\square$ for each frame; 2) Based on class embeddings (c.ebd) $\\square$ of all frames, we use a pre-trained Anomaly-aware Classifier to localize the time period of abnormal events, thereby generating Temporal Effective Tokens $\\square$ to feed forward into the LLM. All of the resulting aligned tokens $\\square \\square \\square$ are fed into the LLM for reasoning and inference of the whole model." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "like Video-Chat [29], Video-ChatGPT [42], Otter [26], Valley [39], mPLUG [27], Video-LLaMA [74], and LLaMAVID [31]. These systems enable interactive video querying, enhance comprehension through audio-visual-text alignment, and support comprehensive video analysis. In this paper, we leverage VLMs and LLMs to develop a novel approach for video anomaly understanding.", + "bbox": [ + 88, + 440, + 485, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Video Anomaly Understanding (VAU)", + "text_level": 1, + "bbox": [ + 89, + 560, + 416, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotating surveillance video frames is labor-intensive, prompting researchers to explore alternatives: one-class learning [72], unsupervised anomaly detection without annotations [4, 5, 16, 20, 37, 38, 47], or weakly supervised methods using only video-level annotations [6, 13, 30, 57, 60, 62, 63, 71, 76]. In one-class learning, Luo et al. developed a ConvLSTM network for normal segment learning [40]. Several researchers employed Auto-Encoder networks to reconstruct normal frame features [5, 20, 81], while others implemented memory mechanisms [4, 16, 37, 47] or meta-learning [38] to enhance generalizability. For weakly supervised learning, Tian [57] sed multiple-instance learning to localize anomalous clips. Zhong et al. uutilized graph convolution networks, though with limited generalization capability. [82]. To address this, Ramachandra et al. developed a Siamese network for normal feature learning. Wan et al. and Zaheer et al. [60, 71]proposed clustering-based frameworks for anomaly identification. Recent studies have introduced new architectures for spatial-temporal feature ensemble learning [6, 13, 30, 57, 62, 63, 76]. However, these methods merely supply anomaly scores during inference,", + "bbox": [ + 91, + 583, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "necessitating empirical threshold establishment on test sets to differentiate abnormal events. Recent research has begun exploring MLLMs to enhance models' capabilities in identifying and describing anomalies [12, 55, 67, 75, 79].", + "bbox": [ + 511, + 441, + 908, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 518, + 604, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 544, + 624, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task. Video anomaly understanding MLLMs aims to determine whether an input video contains abnormal events, meanwhile describing and interacting with the temporal localization and the entire process of the detected abnormal events (if has). We train the model with an instruct-following dataset based on abnormal videos [54], so that the model can better align the tokens between visual encoders and MLLMs for presenting and generalizing information about abnormal events.", + "bbox": [ + 511, + 566, + 908, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pipeline. Considering the video understanding MLLM framework as shown in Figure 2, taking a video (contains $T$ frames) as input, a frozen ViT-based [11] visual encoder (CLIP [52]) extracts visual tokens $\\mathbf{X}^t$ from each video frame $\\mathbf{V}^t (t = 1,\\dots,T)$ . For $\\mathbf{X}^t = \\{\\mathbf{x}_i^t\\}_{i = 1,\\dots ,N}$ of the current frame, there are $N$ visual tokens corresponding to equal amounts of image patches. Modality alignment converts the processed visual tokens $\\mathbf{X}^t$ into the semantic space of LLMs. At the same time, text prompts are processed and encoded as text tokens into the same semantic space and serve as a part of input to LLMs. Our key design on models consists of (1) selecting Spatial Effective Tokens $\\mathbf{X}^{*t}$ (SET) from $\\mathbf{X}^t$ for each frame participating in fine-tuning and inference", + "bbox": [ + 511, + 704, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22697", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "instead of $\\mathbf{X}^t$ (see Section 3.2); and (2) generating Temporal Effective Tokens $\\mathbf{S}^{*t}$ (TET) as anomaly-aware temporal priors, participating in inference to facilitate the temporal localization of abnormal events for LLMs (see Section 3.3). In addition, we produce high-quality instruct-following data on abnormal videos and develop a training strategy for it to maximize the effectiveness of our proposed method (see Section 3.4).", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spatial Effective Token Selection (SETS)", + "text_level": 1, + "bbox": [ + 89, + 219, + 437, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In classical video classification tasks, context and relationships are critical. However, in our MLLM setting, beyond leveraging contextual information, the most crucial problem is aligning the visual and language modalities. Therefore, the key aspect of our design is to extract useful information for effectively aligning visual tokens with the LLM. Since text captions primarily describe anomaly events, which occupy only a small portion of the entire video, aligning all visual patterns with text tokens would be unreasonable and computationally heavy. Thus, we are the first to propose a novel token selection method SETS to achieve efficient and effective alignment.", + "bbox": [ + 89, + 241, + 483, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-frame Difference. For a video, we believe that areas with large changes in adjacent frames are more worthy of attention. As illustrated in Figure 2, for each frame $\\mathbf{V}^t$ of the video, we can regard its previous frame $\\mathbf{V}^{t-1}$ as the reference frame for investigating the difference between current timestamp and previous timestamp. Employing DINOV2 [46] as the feature extractor, denoted as $\\mathcal{F}_{\\mathrm{E}}$ , we can extract patch embeddings:", + "bbox": [ + 89, + 422, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1} = \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t}), \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t - 1}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 550, + 482, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}^t, \\mathbf{F}^{t-1} \\in \\mathbb{R}^{N \\times C}$ are the extracted embeddings ( $N$ indicates the number of image patches and $C$ indicates the channels). Thanks to the distinction and stability of the extracted features, we calculate their patch-wise distances as the inter-frame difference map of the current frame:", + "bbox": [ + 89, + 575, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {D} ^ {t} = \\operatorname {d i s} \\left(\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 659, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $dis(\\cdot)$ indicates Manhattan distance [17] and $\\mathbf{D}^t\\in$ $\\mathbb{R}^N$ indicates the distances between corresponding patch pairs in neighbour frames.", + "bbox": [ + 89, + 684, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Select Spatial Effective Tokens. According to the interframe difference map $\\mathbf{D}^t$ , we can set up a vector $\\mathbf{M}^t = [m_1^t, m_2^t, \\dots, m_N^t]$ to record the difference of each patch, where top $K$ ratio of elements with the largest distance are assigned with the value of 1, and the rest are assigned as 0. Thus we get a mask for filtering and updating the visual tokens as:", + "bbox": [ + 89, + 729, + 483, + 834 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {* t} = \\left\\{\\mathbf {x} _ {i} ^ {t} \\mid m _ {i} ^ {t} = 1, m _ {i} ^ {t} \\in \\mathbf {M} ^ {t} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 843, + 482, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{X}^{*t}$ contains the selected Spatial Effective Tokens (SET) which are fed into subsequent processing instead of", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{X}^t$ as shown in Figure 2. SET can efficiently isolate the regions highly related to the abnormal events to participate in both fine-tuning and inference.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Temporal Effective Token Generation (TETG)", + "text_level": 1, + "bbox": [ + 511, + 143, + 905, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Anomaly-aware Classifier. We design a simple but effective MLP $\\mathcal{F}_{\\mathrm{A}}$ for learning whether each frame is related to an abnormal event. For the class embeddings (denoted as $\\mathbf{z}$ ) extracted from the feature encoder, we can split them based on training video caption into normal and anomaly embeddings, denoted as $\\mathbf{z}^{\\mathrm{n}}$ and $\\mathbf{z}^{\\mathrm{a}}$ , respectively. Thus we can optimize $\\mathcal{F}_{\\mathrm{A}}$ using a binary classification loss:", + "bbox": [ + 511, + 166, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {n}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {a}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 280, + 906, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The anomaly-aware classifier predicts whether each frame is related to anomalies in a video, which can bring additional important prior knowledge to LLMs at a very low cost to facilitate its inference.", + "bbox": [ + 511, + 311, + 905, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generate Temporal Effective Tokens. Since the information drawn from the anomaly-aware classifier is explicit, we can easily project it to LLMs' text token space through natural languages. Based on the prediction results of the anomaly-aware classifier, we select the first and last frames' timestamps with high confidence of containing abnormal events, denoted as $$ and $$ , respectively. Then we tokenize them with a template as: \"Known common crime types are: 'Shooting', 'Arson', 'Arrest', ... There is one of the crime types occurring from $$ to $$ \", resulting Temporal Effective Tokens (TET) in the text token space of LLMs. During inference, with the well-trained lightweight anomaly-aware classifier, TET is used as an additional input to participate in the forward process of the LLM to provide prior knowledge about abnormal events temporally (as shown in Figure 2).", + "bbox": [ + 511, + 371, + 908, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training Strategy", + "text_level": 1, + "bbox": [ + 511, + 652, + 686, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For modality alignment and instruction tuning, we follow the baseline [31] to ensure visual features are well aligned with the language space. In this work, the training strategy can be divided into two stages: 1) Stage One: Fine-tuning with anomaly video data, and 2) Stage Two: Aligning Spatial Effective Tokens with LLMs.", + "bbox": [ + 511, + 674, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-tuning with Video Anomaly Data. For enhancing the abnormal scene understanding of LLMs, we construct the UCF-crime [70] Question-Answer pairs for fine-tuning. We also mix different instruction pairs from various sources [32], including text conversations, single/multi-turn visual Question-Answer pairs, and video Question-Answer pairs. Different formats for text, image, and video inputs are adopted, and the image token is randomly inserted at the beginning or end of user input during training. All modules,", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22698", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/63dbad57a8139d18d76768fc40e64105b4d5de973aebc253480a5f905b3b2ac8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLMIn-domainCross-domain
Total Acc.(%)Temporal Acc.(%)Total Acc.(%)Temporal Acc.(%)
Video-Chat [29]Vicuna-7B22.41---
Video-ChatGPT [42]Vicuna-7B24.1328.5124.0029.10
Otter [26]LLaMa-7B22.4122.1725.2023.80
Valley [39]Vicuna-7B20.3414.4821.0020.20
mPLUG [27]LLaMa-7B22.76---
Video-LLaMA2 [7]Vicuna-7B21.3826.6224.2023.00
Hawkeye [79]LLaVA-7B28.6030.0025.3028.50
LLaMA-VID [31]Vicuna-7B14.8326.7018.8023.60
VA-GPT (Ours)Vicuna-7B30.6935.0026.2031.02
", + "bbox": [ + 98, + 88, + 898, + 275 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d16fbeb61164e4c66f1dc5b2c7acd2a449c9c6cb47ba56af06a4f246a2ae9dbd.jpg", + "table_caption": [ + "Table 1. Comparing on in-domain (UCF-Crime [54]) and the proposed cross-domain (XD-Violence [64]) benchmarks, our method significantly outperforms other models and achieve the state-of-the-art performance (accuracy is detonated as Acc.) on anomaly video understanding and temporal localization with LLMs (Best results are shown in bold)." + ], + "table_footnote": [], + "table_body": "
MethodDescriptionCausesEffect
A-Guardian [12]79.6558.9250.64
Ours80.8359.5551.08
", + "bbox": [ + 93, + 345, + 478, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. We evaluate our method on the MMEval [12] benchmark and show that our method outperforms the related previous method in different aspects.", + "bbox": [ + 89, + 404, + 482, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "except the frozen visual encoder, are optimized in this stage. After fine-tuning, the LLM has a preferred perception of anomalies, thus ensuring the effectiveness of the temporal effective tokens (see Section 3.3) during inference. More dataset details are illustrated in Section 4.", + "bbox": [ + 88, + 454, + 482, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Aligning Spatial Effective Tokens with LLMs. For abnormal video scenes, most areas would not be aligned with languages. Therefore, we implement an additional fine-tuning step. This step involves utilizing Spatial Effective Tokens (see Section 3.2) derived from each video frame within the UCF-Crime dataset. By incorporating these tokens, we aim to provide the model with a more refined understanding of the spatial context of anomalies. It also brings efficient optimization, and the alignment here is only designed for very short-term fine-tuning, which can greatly improve the model's ability to detect and understand anomalies.", + "bbox": [ + 88, + 529, + 482, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We fine-tune our model on our proposed instruct-following format [34] training dataset including 4077 videos and 7730 images based on UCF-Crime [54] dataset. We evaluate the models on two anomaly video understanding benchmarks: UCF-Crime [54] for in-domain evaluation and a proposed benchmark designed based on XD-Violence [64] dataset for cross-domain evaluation, respectively. More details are shown in the supplementary.", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks and Metrics. To evaluate the ability to review videos and identify anomalies, we utilize a video anomaly understanding evaluation from Video-Bench [44] to assess the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "temporal comprehensive ability, which contains nature language based Question-Answer pairs from UCF-Crime [54] dataset. Meanwhile, in order to evaluate the model's cross-domain video anomaly understanding ability, we contribute Question-Answer pairs as an extra benchmark based on the XD-Violence dataset. These Question-Answer pairs encompass four options, with each presenting the anomaly category and the respective time intervals during which the anomalies transpire. For each benchmark, we design different sets of questions for two evaluations: one is an overall evaluation of abnormal event detection and understanding, and the other is a special evaluation focusing on temporal localization ability, measured by question answering accuracy (denoted as Total Acc. and Temporal Acc., respectively, higher is better).", + "bbox": [ + 511, + 349, + 906, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For network structure, we incorporate the pre-trained CLIP [52] and DINOv2 [46] as the visual encoder and Qformer [9] as the text decoder. We follow [31] to freeze the encoder during the modality alignment and to optimize the trainable parameters using anomaly videos and instructions for instruction tuning. During the training process, our model utilizes PyTorch on four NVIDIA A100 GPUs. We employ the AdamW optimizer with a batch size of 64. The learning schedule is set to cosine decay, with a learning rate of 2e-5 and a total of 1 epoch.", + "bbox": [ + 511, + 561, + 908, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 726, + 651, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on In-domain Dataset. We first evaluate our method on the in-domain dataset, where the test set belongs to the same style and recording mode as the data used for training in Section 3.4. As shown in Table 1, compared with the baseline [31], with fewer visual embedding tokens and temporal effective tokens, our method brings more than double the performance improvement on total accuracy, also brings a significant increase in temporal localization. Driven by our proposed training strategy and designed effective tokens, more pure and effective visual-semantic information", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22699", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c28d2ffd063a24261cf9b218ac80213ecd9ec39446deca8bbf49b74f1aff9a22.jpg", + "image_caption": [ + "A 11-seconds video about Road Accident" + ], + "image_footnote": [], + "bbox": [ + 96, + 112, + 344, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1001798447d95625b5f868e07fe148bd0eb3bc8938ed35a3ffceb4d3bd70846f.jpg", + "image_caption": [ + "Describe the video in detail." + ], + "image_footnote": [], + "bbox": [ + 98, + 161, + 127, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/056f08cac14508ecd68c98aba133246a4e9cbbf45a565da7b9e49c85f56adf5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 193, + 341, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video shows a group of people walking down a sidewalk and suddenly a car appears driving very fast towards them. Some people fall onto the ground as the car crashes into them.", + "bbox": [ + 99, + 215, + 341, + 270 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66c32a78cc0d4870d9f3beb41b7cd1fb6ca333003dd36e684f33d37dbe85b790.jpg", + "image_caption": [ + "Are there any anomalies in the video?" + ], + "image_footnote": [], + "bbox": [ + 98, + 279, + 129, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d3ea04034047572675ca6b14e656d1bcacbb607a06b50d88a8527a399aecb73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 316, + 341, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Yes, there is a car accident in the video. It shows a car flipping over and people running away from the scene.", + "bbox": [ + 99, + 340, + 325, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6c5b72592bb6bc56ecb6bc7fe41d3e5133aab0994e1c33b530bd1896d516c22.jpg", + "image_caption": [ + "A 1.5-minutes video about Explosion" + ], + "image_footnote": [], + "bbox": [ + 375, + 111, + 488, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/970041b38aaa198a4b4e4a61b179813fd7c428ddc02e2a520ddcfecc50fe793e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 109, + 620, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ac0a8b3f002a356b2acab33ec6893d226481b388a5b68f893d9283b14a68ebf.jpg", + "image_caption": [ + "What happened in this video?" + ], + "image_footnote": [], + "bbox": [ + 375, + 160, + 406, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18afc59029baa429109e6ee3c3a4863d4b1aa8e6550e03eef4f5628e75c96e61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 193, + 620, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video shows a large group of trees in the woods that were burnt down in a fire. The fire was so intense that it caused the trees to fall over and burn in different directions. The video also shows a large plume of smoke coming from the forest, indicating that the fire was still active.", + "bbox": [ + 379, + 215, + 622, + 291 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/092f3be4f298fd70a9ae29120af13424adc5c052750c3a4fca4c6702d5122a2d.jpg", + "image_caption": [ + "Describe the anomaly present in the video." + ], + "image_footnote": [], + "bbox": [ + 375, + 300, + 406, + 319 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a87329e5ab5445c48b1104ae5214e13e63a1412d4e179df459760a44802e052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 338, + 620, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video shows a large explosion in the woods, which is an anomaly.", + "bbox": [ + 379, + 361, + 602, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b06b80aaaafc591b0b549b12a8e7b126478728542db979e05c7e713d1569375.jpg", + "image_caption": [ + "A 32-minutes normal video" + ], + "image_footnote": [], + "bbox": [ + 651, + 112, + 763, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae089a3d114f6671e9462e8a396ea8dfb86f3627f8df15504542da22b21da066.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 782, + 112, + 895, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94ed00ca72b5bc96a30d678acc08ac7002cea474452340af29cc312023a60214.jpg", + "image_caption": [ + "Describe the video's main features in detail." + ], + "image_footnote": [], + "bbox": [ + 651, + 161, + 681, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b391f11031e1277eaafab436f151593ef7d2a011e0479513bf232f84f3448ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 870, + 194, + 895, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The video features a group of workers at a bank who are interacting with customers. The workers are seen helping customers with various transactions, including deposits, withdrawals, and account inquiries. They are dressed in uniforms and are sitting behind a counter. The customers are also seen talking to the workers and using their cell phones. The video captures a typical day at a bank and the interactions between the workers and customers.", + "bbox": [ + 651, + 215, + 893, + 334 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a08efa9fa111c440e2d52373bf93f6b8b321a8fde23740f29e34245ac33e1b50.jpg", + "image_caption": [ + "Is this video normal or abnormal?" + ], + "image_footnote": [], + "bbox": [ + 651, + 338, + 681, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3dfcba1ea0d44f79abec27da9a772b6ba79f20a00ad9fd57b7eb5043500c8b8b.jpg", + "image_caption": [ + "Normal.", + "Figure 3. Qualitative results in Question-Answer diagrams, the red circles in the figures correspond to the bold text in the answers. From short video of only a dozen seconds to medium video of longer than one minute and long video of about half an hour, our model can reason well and understand the content." + ], + "image_footnote": [], + "bbox": [ + 869, + 352, + 895, + 371 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/efaa98168085d879e57dfff2d1db13d36ead6f6b3d8648e8d894e936080c2982.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBaselineStage One Fine-tuningStage Two Fine-tuning
w/o Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Both
Total Acc. (%)↑14.8324.8323.7925.1225.8626.1027.5029.3128.9630.69
Temporal Acc. (%)↑26.7027.2027.7628.8129.6830.0230.7731.6033.5835.00
", + "bbox": [ + 96, + 452, + 898, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on Spatial Effective Token Selection (SETS), Temporal Effective Tokens Generation (TETG), and progressive training strategies. At different model training stages, starting from the baseline (w/o fine-tuning, w/o both), we compare the performance of only using SETs (w.SETS), only using TETG (w.TETG), and using both (w.Both) on the UCF-Crime benchmark. Stage One: Only anomaly video fine-tuning. Stage Two: Anomaly video fine-tuning + Fine-tuning with SETs (Best results are shown in bold).", + "bbox": [ + 89, + 547, + 906, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of abnormal events is efficiently aligned with LLMs and exhibits powerful anomaly video understanding capabilities. At the same time, we conduct fair comparisons with existing video understanding models [26, 27, 29, 39, 42, 74] (see Table 1), and we demonstrate competitive performance. It is worth noting that we use the fewest tokens among all methods to achieve the state-of-the-art results on both total and temporal accuracies.", + "bbox": [ + 89, + 619, + 483, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on Cross-domain Dataset. For evaluating the robustness and generalization of the models, we additionally design a cross-domain benchmark. We conduct a fair comparison of our method with the baseline [31] and the existing in-domain methods on the proposed cross-domain benchmark. The results presented in Table 1 showcase a substantial performance improvement over existing methods on the cross-domain dataset, underscoring the exceptional generalization and temporal localization capabilities of our methodology. This clear superiority in performance serves", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as a compelling validation of the robustness and adaptability of our approach across diverse domains.", + "bbox": [ + 511, + 619, + 906, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Interaction with the Model. We take some interactions with our well-trained model for better evaluation. As shown in Figure 3, we demonstrate the performance of our model in addressing various video anomaly understanding challenges. To evaluate the model's effectiveness, we select videos of different durations: short (0 to 1 minute), medium (1 to 30 minutes), and long (over 30 minutes). This variety allows us to thoroughly assess the model's capabilities in handling diverse anomaly understanding scenarios. In the Road Accident video (left side in Figure 3), our method successfully identifies a car driving at high speed and detects people falling, even in low-resolution footage. For the Explosion video (middle in Figure 3), the model accurately predicts the scene and the anomaly in a medium-length video depicting an explosion. In a normal video exceeding 30 minutes (right side in Figure 3), we demonstrate the model's ability", + "bbox": [ + 511, + 659, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22700", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9f5b98fd8a4499b4e47042dca1e5ab8ef21227ee8952f4396f4f5ec147a0fb9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLLMFine-tunedTotal Acc.(%)↑
Video-ChatGPTVicuna-7B-24.13
✓26.23
LLaMA-VID (Baseline)Vicuna-7B-14.83
✓23.10
OursVicuna-7B-25.12
✓30.69
", + "bbox": [ + 94, + 89, + 482, + 220 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to focus on both global and local information by asking it to summarize the content.", + "bbox": [ + 89, + 280, + 482, + 309 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on Other Benchmark. We additionally compare our method on another benchmark (MMEval [12]) about anomaly video understanding with LLMs from different aspects. We follow a fair evaluation on our proposed model and obtain the quantity results as shown in Table 2, which shows the superiority of our method.", + "bbox": [ + 89, + 311, + 483, + 401 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 410, + 254, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct extensive ablation studies to validate the effectiveness of the key components in our method: Spatial Effective Token Selection (SETS, Section 3.2) and Temporal Effective Tokens Generation (TETG, Section 3.3) on progressive training strategies in Section 3.4.", + "bbox": [ + 89, + 431, + 483, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-tuning Stages. The utilization of our high-quality UCF instruct-following data has proven to enhance the model's performance. Fine-tuning with this dataset has effectively contributed to a notable accuracy compared with the baseline. As evidenced in Table 3, with our model designing (both SETS and TETG), the total accuracy for anomaly detection only achieves $25.12\\%$ without any fine-tuning (denoted as Baseline). With anomaly video fine-tuning (denoted as Stage One Fine-tuning), the accuracy increases to $27.5\\%$ . Furthermore, an efficient tuning with SETS (denoted as Stage Two Fine-tuning) can achieve our final performance of $30.69\\%$ total accuracy. Temporal accuracy also shows similar increasing patterns with the scaling tuning stages.", + "bbox": [ + 89, + 508, + 483, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Fine-tuning Data. For a fair comparison, we fine-tune some high-performance comparison models [31, 42] and compare them with our proposed UCF instruct-following data. As shown in Table 4, the performance of these comparison models has increased after fine-tuning, which proves the effectiveness of our proposed data. However, their performance is still not as good as our method, which proves the effectiveness of our proposed model.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of SETS. Our proposed SETS demonstrates efficiency in extracting useful abnormal information, leading to performance enhancement. As illustrated in Table 3, with the SETS, the accuracy reaches $24.83\\%$ , $25.86\\%$ and $29.31\\%$ without fine-tuning, anomaly video fine-tuning and", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af1806027d3eddbfeb7042ece1355b1564e564d5ec9f076fa3833c0042f63b88.jpg", + "table_caption": [ + "Table 4. We fine-tune comparison models with our proposed UCF instruct-following data and evaluate the performance of these models before and after fine-tuning on the UCF benchmark." + ], + "table_footnote": [], + "table_body": "
# Sample Rate K0.10.30.50.70.9
Total Acc.(%)↑23.6124.8330.6928.6727.27
Temporal Acc.(%)↑29.0329.9335.0031.2331.03
", + "bbox": [ + 517, + 89, + 905, + 150 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. For the sample rate of tokens, $K$ ratios in SETS, we sample the patch tokens with ordered distance at a fixed sample rate (0.5). The ablation indicates too large sampling rates cause too much context noise, and too small sampling rates lose visual information.", + "bbox": [ + 511, + 157, + 906, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fine-tuning with SETS, respectively, which far exceed the accuracy of the baseline. Its intuitive mechanism for information filtering can be further analyzed with reference to Figure 4. The initial video is often cluttered with irrelevant and deceptive data. For example, in Figure 4, case one illustrates a scenario where the overall structure is complex, yet only a small segment requires attention. The SETS effectively filters out the dynamic features that do not require attention. Similarly, as in case two, the abnormal area is quite small. Our proposed SETS mechanism effectively filters out redundant and irrelevant information, significantly enhancing the model's ability to accurately pinpoint and recognize abnormal situations.", + "bbox": [ + 511, + 232, + 906, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct the ablation studies on $K$ ratios about SETS as shown in Tabel 5. Too small or too large $K$ ratios will cause performance degradation in both total and temporal accuracy. If $K$ ratios is too small, redundant information will affect the effectiveness of aligning abnormal event information with corresponding captions. In contrast, if $K$ ratios is too large, some important areas will be filtered out, resulting in information loss and suboptimal performance.", + "bbox": [ + 511, + 430, + 908, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of TETG. Our proposed TETG generates tokens directly in the text token space of LLMs as priors for each video, offering robust priors on the temporal aspects of abnormal events. The provision results in performance improvement without the necessity for fine-tuning. As shown in Table 3, the accuracy rises from $14.83\\%$ to $23.79\\%$ . Fine-tuning with anomaly video and SETS, the results even achieve $26.10\\%$ and $30.69\\%$ independently, which manifests the effectiveness of TETG. Besides, the integration of SETS and TETG highlights the importance of leveraging spatial and temporal information effectively in anomaly detection systems, boosting the results to $25.15\\%$ , $27.50\\%$ and $30.69\\%$ , respectively.", + "bbox": [ + 511, + 553, + 908, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 768, + 627, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key tokens play key roles. To the best of our knowledge, we are the first to explore how to assign different learnable knowledge to different tokens for better alignment with LLMs on visual contents, thus promoting video anomaly detection and understanding (see Table 1 and Figure 3). We assign the most effective roles to different tokens in both spatial and temporal dimensions, enabling the model to handle", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f790a42a12bb676fc8dee77ed33d810b32d38082bda8637bd4f07397f2bc34e0.jpg", + "image_caption": [ + "Figure 4. Visualization of the initial videos and our masked results. These two cases illustrate road accident scenarios: one occurring in a bustling street and the other in an empty suburb. Our SETS effectively filters redundant and irrelevant regions (with black patch-level masks)." + ], + "image_footnote": [], + "bbox": [ + 94, + 92, + 897, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "various tokens more efficiently. The video contains abundant but redundant information. Our proposed SETS and TETG effectively compress the spatial and temporal information of abnormal events respectively, and utilize the existing alignment mechanism of MLLMs at a very low cost to participate in LLMs' reasoning (see Table 3). Our exploration inspires more representation learning of MLLMs to facilitate downstream tasks.", + "bbox": [ + 88, + 452, + 485, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data matters a lot. We construct instruct-following data for anomaly videos, containing approximately 4,000 videos, which is significantly less than the amount of baseline fine-tuning video data (e.g., over 90k videos for fine-tuning in baseline model [31]). We still achieve promising performance on both in-domain and cross-domain benchmarks (see Table 1). This relies on the high-quality Question-Answer pairs in our instruct-following data. Meanwhile, SETS also improves data quality during fine-tuning: visual areas irrelevant to Question-Answer pairs are filtered out (see Figure 4), which allows for significant performance improvements in the second stage of fine-tuning (see Section 3.4) with very few steps (less than 150 iterations).", + "bbox": [ + 88, + 578, + 486, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader impacts. Video anomaly understanding has far-reaching implications across various sectors, including security, healthcare, industrial safety, and so on. By enhancing the ability to automatically identify and respond to unusual or suspicious activities in real-time, LLMs can significantly improve public safety, crime prevention, patient monitoring, hazard detection, loss prevention, traffic management, and urban planning. These systems offer substantial benefits in", + "bbox": [ + 89, + 780, + 486, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "terms of operational efficiency and safety.", + "bbox": [ + 511, + 453, + 792, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our model adeptly portrays the occurrence, type, and area of video abnormal events, it still faces challenges in detecting and describing certain complex scenes. Our strategy represents an early successful validation and investigation of large models for video anomaly identification and localization. Consequently, our method possesses significant potential for enhancement in recognizing diverse abnormal video scenes. These insights motivate us to continue pursuing more powerful and efficient video anomaly understanding technologies in the future, aiming to address more challenges in the real world [35, 41, 77].", + "bbox": [ + 511, + 469, + 908, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 648, + 640, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel MLLM for understanding anomalies in videos with LLMs by aligning effective tokens in both temporal and spatial space. The proposed method includes Spatial Effective Token Selection (SETS) for identifying abnormal events in small areas of large scenes and Temporal Effective Tokens Generation (TETG) for addressing the sparseness of abnormal events in video time sequences. We also develop instruct-following data of video anomaly detection to fine-tune the model. Besides, evaluation on the video anomaly understanding benchmark and a proposed cross-domain benchmark demonstrates the effectiveness of the proposed method. It further presents a promising approach for video anomaly understanding using MLLMs, showcasing the potential of effective tokens for enhancing video understanding tasks.", + "bbox": [ + 511, + 672, + 910, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22702", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No.27209621), General Research Fund Scheme (Grant No.17202422, 17212923, 17215025), Theme-based Research (Grant No.T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust.", + "bbox": [ + 89, + 113, + 485, + 255 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 282, + 187, + 297 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2", + "[2] Borislav Antić and Björn Ommer. Video parsing for abnormality detection. In 2011 International Conference on Computer Vision, pages 2415-2422, 2011. 1", + "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2", + "[4] Ruichu Cai, Hao Zhang, Wen Liu, Shenghua Gao, and Zhifeng Hao. Appearance-motion memory consistency network for video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):938-946, 2021. 1, 3", + "[5] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In Computer Vision - ECCV 2020, 16th European Conference, pages 329-345, 2022. 3", + "[6] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. AAAI2023, 2022. 1, 3", + "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5", + "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai," + ], + "bbox": [ + 99, + 306, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 2", + "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3", + "[12] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, Dajiu Huang, Jing Feng, Linli Chen, Can Zhang, Xuhuan Li, Hao Zhang, Jianhang Chen, Qimei Cui, and Xiaofeng Tao. Uncovering what, why and how: A comprehensive benchmark for causation understanding of video anomaly, 2024. 2, 3, 5, 7", + "[13] JiaChang Feng, FaTing Hong, and WeiShi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3", + "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, Rongrong Ji, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining, 2022. 2", + "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all, 2023. 2", + "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In IEEE International Conference on Computer Vision (ICCV), 2019. 1, 3", + "[17] Rodolfo Gonzalez and Sandra Palais. A path-building procedure for iterative circuit computers. Technical report, 1962. 4", + "[18] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, Z. Yang, Y. Zhang, and D. Tao. A survey on vision transformer. IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1-1, 2020. 2", + "[19] Mahmudul Hasan, Jonghyun Choi, jan Neumann, Amit K Roy-Chowdhury, and Larry Davis. Learning temporal regularity in video sequences. In Proceedings of IEEE Computer Vision and Pattern Recognition, 2016. 1", + "[20] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. 2", + "[22] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection, 2023. 1", + "[23] Louis Kratz and Ko Nishino. Anomaly detection in extremely crowded scenes using spatio-temporal motion pattern models. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1446-1453, 2009. 1", + "[24] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models, 2023. 2", + "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension, 2019. 2", + "[26] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning, 2023. 3, 5, 6", + "[27] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections, 2022. 1, 3, 5, 6", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 1, 2", + "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 1, 3, 5, 6", + "[30] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisecquence learning with transformer for weakly supervised video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395-1403, 2022. 3", + "[31] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models, 2023. 1, 3, 4, 5, 6, 7, 8", + "[32] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2023. 4", + "[33] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jinfa Huang, Junwu Zhang, Munan Ning, and Li Yuan. Moellava: Mixture of experts for large vision-language models, 2024. 2", + "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5", + "[35] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 8", + "[36] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 1", + "[37] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1, 3", + "[38] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In European Conference on Computer Vision, 2020. 3", + "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability, 2023. 1, 3, 5, 6", + "[40] Weixin Luo, Wen Liu, and Shenghua Gao. Remembering history with convolutional LSTM for anomaly detection. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 439-444, 2017. 3", + "[41] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 8", + "[42] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 1, 3, 5, 6, 7", + "[43] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. CoRR, abs/1908.06351, 2019. 1", + "[44] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 5", + "[45] OpenAI. Introducing chatgpt. 2022. 1, 2", + "[46] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 4, 5", + "[47] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14372-14381, 2020. 1, 3", + "[48] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4, 2023. 2", + "[49] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22704", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1", + "[52] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, 2022. 3, 5", + "[53] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all, 2023. 2", + "[54] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 5", + "[55] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Ying-Cong Chen. Hawk: Learning to understand open-world video anomalies, 2024. 2, 3", + "[56] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog applications, 2022. 2", + "[57] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4975-4986, 2021. 1, 3", + "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1, 2", + "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2", + "[60] Boyang Wan, Yuming Fang, Xue Xia, and Jiajie Mei. Weakly supervised video anomaly detection via center-guided discriminative learning. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2020. 1, 3" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks, 2023. 2", + "[62] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3", + "[63] Peng Wu and Jing Liu. Learning causal temporal relation and feature discrimination for anomaly detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. 1, 3", + "[64] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5", + "[65] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection, 2023. 1", + "[66] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection, 2024. 1", + "[67] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models, 2024. 2, 3", + "[68] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detection, 2024. 1", + "[69] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training, 2021. 2", + "[70] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 4", + "[71] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision (ECCV), pages 358-376. Springer, 2020. 1, 3", + "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection, 2022. 3", + "[73] Muhammad Zaigham Zaheer, Jin-Ha Lee, Marcella Astrid, Arif Mahmood, and Seung-Ik Lee. Cleaning label noise with clusters for minimally supervised anomaly detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, June 2020. 1", + "[74] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 1, 3, 6", + "[75] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22705", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity, 2024. 2, 3", + "[76] Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4030-4034, 2019. 3", + "[77] Jiaqi Zhang, Yan Hu, Xiaojuan Qi, Ting Meng, Lihui Wang, Huazhu Fu, Mingming Yang, and Jiang Liu. Polar eyeball shape net for 3d posterior ocular shape representation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 180-190. Springer, 2023. 8", + "[78] Bin Zhao, Fei-Fei Li, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 1", + "[79] Jianing Zhao, Jingjing Wang, Yujie Jin, Jiamin Luo, and Guodong Zhou. Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene-enhanced video large language model. In Proceedings of ACM MM 2024, 2024. 2, 3, 5", + "[80] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2023. 2", + "[81] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In Proceedings of the 25th ACM International Conference on Multimedia, page 1933-1941, New York, NY, USA, 2017. Association for Computing Machinery. 3", + "[82] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[83] Lanfeng Zhong, Xin Liao, Shaoting Zhang, Xiaofan Zhang, and Guotai Wang. Vlm-cpl: Consensus pseudo labels from vision-language models for human annotation-free pathological image classification, 2024. 2", + "[84] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection, 2023. 1", + "[85] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2", + "[86] Yi Zhu and Shawn D. Newsam. Motion-aware feature for improved video anomaly detection. In *British Machine Vision Conference (BMVC)*, 2019. 1" + ], + "bbox": [ + 91, + 90, + 482, + 810 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22706", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_model.json b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9bdbab352c2cf4457e3122aa30545b25ac7e94 --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_model.json @@ -0,0 +1,2732 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.131, + 0.873, + 0.153 + ], + "angle": 0, + "content": "Aligning Effective Tokens with Video Anomaly in Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.179, + 0.851, + 0.273 + ], + "angle": 0, + "content": "Yingxian Chen\\(^{1*}\\) Jiahui Liu\\(^{1*}\\) Ruidi Fan\\(^{1}\\) Yanwei Li\\(^{2}\\) Chirui Chang\\(^{1}\\) Shizhen Zhao\\(^{1}\\) \nWilton W.T.Fok\\(^{1}\\) Xiaojuan Qi\\(^{1\\dagger}\\) Yik-Chung Wu\\(^{1\\dagger}\\) \n\\(^{1}\\)The University of Hong Kong \\(^{2}\\)The Chinese University of Hong Kong \n{chenyx, liujh, xjqi, ycuw}@eee.hku.hk \n*equal contributions †corresponding authors" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.486, + 0.699 + ], + "angle": 0, + "content": "Understanding abnormal events in videos is a vital and challenging task that has garnered significant attention in a wide range of applications. Although current video understanding Multi-modal Large Language Models (MLLMs) are capable of analyzing general videos, they often struggle to handle anomalies due to the spatial and temporal sparsity of abnormal events, where the redundant information always leads to suboptimal outcomes. To address these challenges, exploiting the representation and generalization capabilities of Vison Language Models (VLMs) and Large Language Models (LLMs), we propose VA-GPT, a novel MLLM designed for summarizing and localizing abnormal events in various videos. Our approach efficiently aligns effective tokens between visual encoders and LLMs through two key proposed modules: Spatial Effective Token Selection (SETS) and Temporal Effective Token Generation (TETG). These modules enable our model to effectively capture and analyze both spatial and temporal information associated with abnormal events, resulting in more accurate responses and interactions. Furthermore, we construct an instruction-following dataset specifically for fine-tuning video-anomaly-aware MLLMs, and introduce a cross-domain evaluation benchmark based on XD-Violence dataset. Our proposed method outperforms existing state-of-the-art methods on various benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Detecting and summarizing abnormal events in videos is critical and challenging, and it has garnered considerable attention across multiple research domains and real-world applications, such as security monitoring, video analysis, and crime detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Although many traditional methods [4, 16, 37, 43, 47, 73, 86] have been widely explored for video anomaly detection, they exhibit substantial limitations in their effectiveness [22, 51, 60, 66, 71, 84]. These limitations manifest in two aspects: 1) Traditional video anomaly detection" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.307, + 0.9, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.662 + ], + "angle": 0, + "content": "Figure 1. Baseline video understanding MLLM feeds forward every visual token (yellow squares) equally to participate in fine-tuning and inference (top row). Different from it, our method focuses on the effective area (unobstructed area in medium video frames) in each frame and select the Spatial Effective Tokens (orange squares) for the LLM (see Section 3.2) (filtered tokens are shown as gray squares). At the same time, we generate anomaly-aware Temporal Effective Tokens (green squares) (see Section 3.3) based on the assigned anomaly scores (denoted as \\( s \\)) of each frame from a pretrained classifier for better temporal localization of anomalies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.78 + ], + "angle": 0, + "content": "methods [6, 13, 57, 63, 64, 86] essentially approach the task as a closed-set detection and classification problem, inherently limiting their ability to achieve a comprehensive understanding and interpretation of anomalies; 2) These methods [2, 19, 23, 65, 68, 78] are restricted by a limited vocabulary, making it difficult for them to handle unseen or novel situations effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Recent advancements [1, 28, 34, 36, 45, 58] in Vision Language Models (VLMs) and Large Language Models (LLMs) have demonstrated remarkable capabilities in scene understanding and comprehensive analysis. Multimodal Large Language Models (MLLMs), particularly those designed for video understanding [27, 29, 31, 39, 42, 74], have achieved significant progress in general video analysis tasks. However, while these models exhibit strong performance" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22695" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "in general video understanding, they fall short in accurately detecting and interpreting anomalies." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.125, + 0.486, + 0.398 + ], + "angle": 0, + "content": "For mitigating the above challenges, some works [12, 55, 67, 75, 79] proposed anomaly-aware video MLLMs to better understand the anomalies in videos. Although these models work well for detecting obvious abnormal events, such as fighting or fire, they typically struggle to effectively align abnormal regions with relevant captions which requires addressing spatial redundancy, and accurately identifying abnormal time intervals by mitigating temporal redundancy. This is because these methods treat all latent tokens with equal priority across spatial and temporal dimensions. This leads to performance degradation caused by redundant tokens unrelated to anomalies. However, in most cases, only small regions within a few frames contain the essential information help to identify an anomaly (as shown in Figure 1). Thus, we explore: How can multimodal architectures evolve selective token generation and processing mechanisms to dynamically prioritize anomaly-salient information while preserving comprehensive scene understanding capabilities?" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.4, + 0.485, + 0.747 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose a new model named VA-GPT for analyzing various Videos for Abnormal events by aligning effective and accurate tokens with LLMs across both spatial and temporal dimensions. VA-GPT integrates two key components to identify effective visual tokens for alignment while eliminating redundant tokens that could hinder anomaly analysis and distract model from extracting useful information: 1) we develop the Spatial Effective Token Selection (SETS) module for identifying tokens corresponding to regions with challenges for aligning them with LLMs, while filtering out tokens associated with minor dynamics to remove redundancy. This is because we find that abnormal events often result in different visual changes and variations in local areas (see Figure 1); and 2) we propose the Temporal Effective Token Generation (TETG) module which employs a lightweight pre-trained classifier to assign a confidence score to each frame indicating the possibility of containing abnormal events. Then TETG generates efficient tokens with prior knowledge of the temporal information of abnormal events directly in the language space as additional input to the LLMs, effectively enhancing the model's temporal reasoning and understanding abilities about abnormal events." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Furthermore, beyond conventional benchmarks (domain benchmark), we establish a new cross-domain evaluation protocol that systematically evaluates model robustness with domain shifts. Based on a novel video dataset, XD-Violence [64], we design comprehensive QAs about abnormal events which include different visual contents from our training data and integrate it as a new cross-domain benchmark. Meanwhile, we design temporal-information-oriented QAs on both in- and cross- domain benchmarks for evaluating temporal localization abilities. Comprehensive" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.137 + ], + "angle": 0, + "content": "experiments demonstrate VA-GPT's superiority, achieving state-of-the-art performance in both in-domain anomaly localization and cross-domain generalization scenarios." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.138, + 0.874, + 0.151 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.156, + 0.907, + 0.216 + ], + "angle": 0, + "content": "- We propose VA-GPT, a video-anomaly-aware MLLM for detecting and summarizing anomalies in various videos, which introduces the MLLM to the specific domain of video anomaly understanding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.907, + 0.336 + ], + "angle": 0, + "content": "- We introduce the SETS and TETG, which enable our MLLM to effectively capture both spatial and temporal information in video sequences, resulting in accurate understanding and localization of abnormal events. Meanwhile, we propose a new instruct-following dataset for video anomaly analysis and a comprehensive cross-domain evaluation benchmark for better evaluating the generalization abilities of MLLMs on video anomalies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.907, + 0.398 + ], + "angle": 0, + "content": "- Our extensive experiments demonstrate that our method outperforms existing state-of-the-art methods in various benchmarks, highlighting its effectiveness and potential for practical applications in video anomaly understanding." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.156, + 0.907, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.413, + 0.655, + 0.428 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.802, + 0.454 + ], + "angle": 0, + "content": "2.1. Large Language Models (LLMs)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.909, + 0.672 + ], + "angle": 0, + "content": "The domain of Natural Language Processing (NLP) has experienced significant progress, particularly with the emergence of Large Language Models (LLMs). The introduction of the Transformer architecture [3, 18, 59] is a critical turning point, followed by other influential language models [3, 10, 25, 80] that exhibited remarkable proficiency. Generative Pre-trained Transformers (GPT) [49] brought about a revolution in NLP by employing auto-regressive prediction, establishing itself as a powerful language modeling approach. More recent groundbreaking contributions, such as ChatGPT [45], GPT-4 [1], LaMDA [56] and LLaMA [58], have expanded the horizon even further. These models, trained on extensive textual data, display extraordinary performance in intricate linguistic tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.806, + 0.699 + ], + "angle": 0, + "content": "2.2. Vision Language Models (VLMs)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Progress in the fields of computer vision and natural language processing has given rise to the development of vision-language models (VLMs) [14, 21, 33, 34, 50, 61, 69]. These models combine visual and linguistic systems to facilitate cross-modal comprehension and reasoning. Notable examples include CLIP [50] which pairs BERT [10] with ViT [11]; BLIP-2 [28] incorporating Vision Transformer features into Flan-T5 [8]; MiniGPT4 [85], connecting BLIP-2 with Vicuna [28, 48]; PandaGPT [53], bridging ImageBind [15] with Vicuna. These models excel in tasks like image classification, captioning, and object detection [24, 53, 83]. Recent developments in vision-language models have extended into video processing with models" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22696" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.318, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Detailed illustration of our proposed model. When a video is fed into the model, patch embeddings \\(\\square\\) and class embeddings (c.ebd) \\(\\square\\) are extracted from all frames. 1) Based on the difference in patch embeddings between current frame and its neighbour frame, we can get a filter mask \\(\\square\\) to filter out unimportant visual tokens (dashed square \\(\\square\\)) from current frame's visual tokens \\(\\square\\), thereby selecting Spatial Effective Tokens \\(\\square\\) that are compressed with a projector with pooling into aligned content token \\(\\square\\) for each frame, meanwhile take attention with text input from users for resulting aligned context token \\(\\square\\) for each frame; 2) Based on class embeddings (c.ebd) \\(\\square\\) of all frames, we use a pre-trained Anomaly-aware Classifier to localize the time period of abnormal events, thereby generating Temporal Effective Tokens \\(\\square\\) to feed forward into the LLM. All of the resulting aligned tokens \\(\\square \\square \\square\\) are fed into the LLM for reasoning and inference of the whole model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.486, + 0.548 + ], + "angle": 0, + "content": "like Video-Chat [29], Video-ChatGPT [42], Otter [26], Valley [39], mPLUG [27], Video-LLaMA [74], and LLaMAVID [31]. These systems enable interactive video querying, enhance comprehension through audio-visual-text alignment, and support comprehensive video analysis. In this paper, we leverage VLMs and LLMs to develop a novel approach for video anomaly understanding." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.561, + 0.418, + 0.577 + ], + "angle": 0, + "content": "2.3. Video Anomaly Understanding (VAU)" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.584, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Annotating surveillance video frames is labor-intensive, prompting researchers to explore alternatives: one-class learning [72], unsupervised anomaly detection without annotations [4, 5, 16, 20, 37, 38, 47], or weakly supervised methods using only video-level annotations [6, 13, 30, 57, 60, 62, 63, 71, 76]. In one-class learning, Luo et al. developed a ConvLSTM network for normal segment learning [40]. Several researchers employed Auto-Encoder networks to reconstruct normal frame features [5, 20, 81], while others implemented memory mechanisms [4, 16, 37, 47] or meta-learning [38] to enhance generalizability. For weakly supervised learning, Tian [57] sed multiple-instance learning to localize anomalous clips. Zhong et al. uutilized graph convolution networks, though with limited generalization capability. [82]. To address this, Ramachandra et al. developed a Siamese network for normal feature learning. Wan et al. and Zaheer et al. [60, 71]proposed clustering-based frameworks for anomaly identification. Recent studies have introduced new architectures for spatial-temporal feature ensemble learning [6, 13, 30, 57, 62, 63, 76]. However, these methods merely supply anomaly scores during inference," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.504 + ], + "angle": 0, + "content": "necessitating empirical threshold establishment on test sets to differentiate abnormal events. Recent research has begun exploring MLLMs to enhance models' capabilities in identifying and describing anomalies [12, 55, 67, 75, 79]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.605, + 0.535 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.625, + 0.559 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.702 + ], + "angle": 0, + "content": "Task. Video anomaly understanding MLLMs aims to determine whether an input video contains abnormal events, meanwhile describing and interacting with the temporal localization and the entire process of the detected abnormal events (if has). We train the model with an instruct-following dataset based on abnormal videos [54], so that the model can better align the tokens between visual encoders and MLLMs for presenting and generalizing information about abnormal events." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Pipeline. Considering the video understanding MLLM framework as shown in Figure 2, taking a video (contains \\(T\\) frames) as input, a frozen ViT-based [11] visual encoder (CLIP [52]) extracts visual tokens \\(\\mathbf{X}^t\\) from each video frame \\(\\mathbf{V}^t (t = 1,\\dots,T)\\). For \\(\\mathbf{X}^t = \\{\\mathbf{x}_i^t\\}_{i = 1,\\dots ,N}\\) of the current frame, there are \\(N\\) visual tokens corresponding to equal amounts of image patches. Modality alignment converts the processed visual tokens \\(\\mathbf{X}^t\\) into the semantic space of LLMs. At the same time, text prompts are processed and encoded as text tokens into the same semantic space and serve as a part of input to LLMs. Our key design on models consists of (1) selecting Spatial Effective Tokens \\(\\mathbf{X}^{*t}\\) (SET) from \\(\\mathbf{X}^t\\) for each frame participating in fine-tuning and inference" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22697" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.212 + ], + "angle": 0, + "content": "instead of \\(\\mathbf{X}^t\\) (see Section 3.2); and (2) generating Temporal Effective Tokens \\(\\mathbf{S}^{*t}\\) (TET) as anomaly-aware temporal priors, participating in inference to facilitate the temporal localization of abnormal events for LLMs (see Section 3.3). In addition, we produce high-quality instruct-following data on abnormal videos and develop a training strategy for it to maximize the effectiveness of our proposed method (see Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.22, + 0.438, + 0.236 + ], + "angle": 0, + "content": "3.2. Spatial Effective Token Selection (SETS)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.422 + ], + "angle": 0, + "content": "In classical video classification tasks, context and relationships are critical. However, in our MLLM setting, beyond leveraging contextual information, the most crucial problem is aligning the visual and language modalities. Therefore, the key aspect of our design is to extract useful information for effectively aligning visual tokens with the LLM. Since text captions primarily describe anomaly events, which occupy only a small portion of the entire video, aligning all visual patterns with text tokens would be unreasonable and computationally heavy. Thus, we are the first to propose a novel token selection method SETS to achieve efficient and effective alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Inter-frame Difference. For a video, we believe that areas with large changes in adjacent frames are more worthy of attention. As illustrated in Figure 2, for each frame \\(\\mathbf{V}^t\\) of the video, we can regard its previous frame \\(\\mathbf{V}^{t-1}\\) as the reference frame for investigating the difference between current timestamp and previous timestamp. Employing DINOV2 [46] as the feature extractor, denoted as \\(\\mathcal{F}_{\\mathrm{E}}\\), we can extract patch embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.551, + 0.483, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1} = \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t}), \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t - 1}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.653 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}^t, \\mathbf{F}^{t-1} \\in \\mathbb{R}^{N \\times C}\\) are the extracted embeddings (\\(N\\) indicates the number of image patches and \\(C\\) indicates the channels). Thanks to the distinction and stability of the extracted features, we calculate their patch-wise distances as the inter-frame difference map of the current frame:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.66, + 0.483, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {D} ^ {t} = \\operatorname {d i s} \\left(\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.483, + 0.731 + ], + "angle": 0, + "content": "where \\(dis(\\cdot)\\) indicates Manhattan distance [17] and \\(\\mathbf{D}^t\\in\\) \\(\\mathbb{R}^N\\) indicates the distances between corresponding patch pairs in neighbour frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.484, + 0.835 + ], + "angle": 0, + "content": "Select Spatial Effective Tokens. According to the interframe difference map \\(\\mathbf{D}^t\\), we can set up a vector \\(\\mathbf{M}^t = [m_1^t, m_2^t, \\dots, m_N^t]\\) to record the difference of each patch, where top \\(K\\) ratio of elements with the largest distance are assigned with the value of 1, and the rest are assigned as 0. Thus we get a mask for filtering and updating the visual tokens as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.844, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {* t} = \\left\\{\\mathbf {x} _ {i} ^ {t} \\mid m _ {i} ^ {t} = 1, m _ {i} ^ {t} \\in \\mathbf {M} ^ {t} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}^{*t}\\) contains the selected Spatial Effective Tokens (SET) which are fed into subsequent processing instead of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.906, + 0.137 + ], + "angle": 0, + "content": "\\(\\mathbf{X}^t\\) as shown in Figure 2. SET can efficiently isolate the regions highly related to the abnormal events to participate in both fine-tuning and inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.145, + 0.906, + 0.162 + ], + "angle": 0, + "content": "3.3. Temporal Effective Token Generation (TETG)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Anomaly-aware Classifier. We design a simple but effective MLP \\(\\mathcal{F}_{\\mathrm{A}}\\) for learning whether each frame is related to an abnormal event. For the class embeddings (denoted as \\(\\mathbf{z}\\)) extracted from the feature encoder, we can split them based on training video caption into normal and anomaly embeddings, denoted as \\(\\mathbf{z}^{\\mathrm{n}}\\) and \\(\\mathbf{z}^{\\mathrm{a}}\\), respectively. Thus we can optimize \\(\\mathcal{F}_{\\mathrm{A}}\\) using a binary classification loss:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.281, + 0.907, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {n}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {a}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.906, + 0.37 + ], + "angle": 0, + "content": "The anomaly-aware classifier predicts whether each frame is related to anomalies in a video, which can bring additional important prior knowledge to LLMs at a very low cost to facilitate its inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.372, + 0.909, + 0.646 + ], + "angle": 0, + "content": "Generate Temporal Effective Tokens. Since the information drawn from the anomaly-aware classifier is explicit, we can easily project it to LLMs' text token space through natural languages. Based on the prediction results of the anomaly-aware classifier, we select the first and last frames' timestamps with high confidence of containing abnormal events, denoted as \\(\\) and \\(\\), respectively. Then we tokenize them with a template as: \"Known common crime types are: 'Shooting', 'Arson', 'Arrest', ... There is one of the crime types occurring from \\(\\) to \\(\\)\", resulting Temporal Effective Tokens (TET) in the text token space of LLMs. During inference, with the well-trained lightweight anomaly-aware classifier, TET is used as an additional input to participate in the forward process of the LLM to provide prior knowledge about abnormal events temporally (as shown in Figure 2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.653, + 0.687, + 0.669 + ], + "angle": 0, + "content": "3.4. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.764 + ], + "angle": 0, + "content": "For modality alignment and instruction tuning, we follow the baseline [31] to ensure visual features are well aligned with the language space. In this work, the training strategy can be divided into two stages: 1) Stage One: Fine-tuning with anomaly video data, and 2) Stage Two: Aligning Spatial Effective Tokens with LLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Fine-tuning with Video Anomaly Data. For enhancing the abnormal scene understanding of LLMs, we construct the UCF-crime [70] Question-Answer pairs for fine-tuning. We also mix different instruction pairs from various sources [32], including text conversations, single/multi-turn visual Question-Answer pairs, and video Question-Answer pairs. Different formats for text, image, and video inputs are adopted, and the image token is randomly inserted at the beginning or end of user input during training. All modules," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22698" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.9, + 0.276 + ], + "angle": 0, + "content": "
MethodLLMIn-domainCross-domain
Total Acc.(%)Temporal Acc.(%)Total Acc.(%)Temporal Acc.(%)
Video-Chat [29]Vicuna-7B22.41---
Video-ChatGPT [42]Vicuna-7B24.1328.5124.0029.10
Otter [26]LLaMa-7B22.4122.1725.2023.80
Valley [39]Vicuna-7B20.3414.4821.0020.20
mPLUG [27]LLaMa-7B22.76---
Video-LLaMA2 [7]Vicuna-7B21.3826.6224.2023.00
Hawkeye [79]LLaVA-7B28.6030.0025.3028.50
LLaMA-VID [31]Vicuna-7B14.8326.7018.8023.60
VA-GPT (Ours)Vicuna-7B30.6935.0026.2031.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.282, + 0.907, + 0.325 + ], + "angle": 0, + "content": "Table 1. Comparing on in-domain (UCF-Crime [54]) and the proposed cross-domain (XD-Violence [64]) benchmarks, our method significantly outperforms other models and achieve the state-of-the-art performance (accuracy is detonated as Acc.) on anomaly video understanding and temporal localization with LLMs (Best results are shown in bold)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.347, + 0.48, + 0.402 + ], + "angle": 0, + "content": "
MethodDescriptionCausesEffect
A-Guardian [12]79.6558.9250.64
Ours80.8359.5551.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.405, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Table 2. We evaluate our method on the MMEval [12] benchmark and show that our method outperforms the related previous method in different aspects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.483, + 0.529 + ], + "angle": 0, + "content": "except the frozen visual encoder, are optimized in this stage. After fine-tuning, the LLM has a preferred perception of anomalies, thus ensuring the effectiveness of the temporal effective tokens (see Section 3.3) during inference. More dataset details are illustrated in Section 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Aligning Spatial Effective Tokens with LLMs. For abnormal video scenes, most areas would not be aligned with languages. Therefore, we implement an additional fine-tuning step. This step involves utilizing Spatial Effective Tokens (see Section 3.2) derived from each video frame within the UCF-Crime dataset. By incorporating these tokens, we aim to provide the model with a more refined understanding of the spatial context of anomalies. It also brings efficient optimization, and the alignment here is only designed for very short-term fine-tuning, which can greatly improve the model's ability to detect and understand anomalies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.726 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Datasets. We fine-tune our model on our proposed instruct-following format [34] training dataset including 4077 videos and 7730 images based on UCF-Crime [54] dataset. We evaluate the models on two anomaly video understanding benchmarks: UCF-Crime [54] for in-domain evaluation and a proposed benchmark designed based on XD-Violence [64] dataset for cross-domain evaluation, respectively. More details are shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Benchmarks and Metrics. To evaluate the ability to review videos and identify anomalies, we utilize a video anomaly understanding evaluation from Video-Bench [44] to assess the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.908, + 0.561 + ], + "angle": 0, + "content": "temporal comprehensive ability, which contains nature language based Question-Answer pairs from UCF-Crime [54] dataset. Meanwhile, in order to evaluate the model's cross-domain video anomaly understanding ability, we contribute Question-Answer pairs as an extra benchmark based on the XD-Violence dataset. These Question-Answer pairs encompass four options, with each presenting the anomaly category and the respective time intervals during which the anomalies transpire. For each benchmark, we design different sets of questions for two evaluations: one is an overall evaluation of abnormal event detection and understanding, and the other is a special evaluation focusing on temporal localization ability, measured by question answering accuracy (denoted as Total Acc. and Temporal Acc., respectively, higher is better)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Implementation Details. For network structure, we incorporate the pre-trained CLIP [52] and DINOv2 [46] as the visual encoder and Qformer [9] as the text decoder. We follow [31] to freeze the encoder during the modality alignment and to optimize the trainable parameters using anomaly videos and instructions for instruction tuning. During the training process, our model utilizes PyTorch on four NVIDIA A100 GPUs. We employ the AdamW optimizer with a batch size of 64. The learning schedule is set to cosine decay, with a learning rate of 2e-5 and a total of 1 epoch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.652, + 0.741 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Results on In-domain Dataset. We first evaluate our method on the in-domain dataset, where the test set belongs to the same style and recording mode as the data used for training in Section 3.4. As shown in Table 1, compared with the baseline [31], with fewer visual embedding tokens and temporal effective tokens, our method brings more than double the performance improvement on total accuracy, also brings a significant increase in temporal localization. Driven by our proposed training strategy and designed effective tokens, more pure and effective visual-semantic information" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22699" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.094, + 0.338, + 0.105 + ], + "angle": 0, + "content": "A 11-seconds video about Road Accident" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.113, + 0.345, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.162, + 0.129, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.183, + 0.245, + 0.193 + ], + "angle": 0, + "content": "Describe the video in detail." + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.194, + 0.343, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.217, + 0.343, + 0.271 + ], + "angle": 0, + "content": "The video shows a group of people walking down a sidewalk and suddenly a car appears driving very fast towards them. Some people fall onto the ground as the car crashes into them." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.28, + 0.13, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.301, + 0.299, + 0.311 + ], + "angle": 0, + "content": "Are there any anomalies in the video?" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.317, + 0.343, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.342, + 0.326, + 0.375 + ], + "angle": 0, + "content": "Yes, there is a car accident in the video. It shows a car flipping over and people running away from the scene." + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.093, + 0.606, + 0.106 + ], + "angle": 0, + "content": "A 1.5-minutes video about Explosion" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.112, + 0.489, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.111, + 0.622, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.161, + 0.408, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.183, + 0.535, + 0.193 + ], + "angle": 0, + "content": "What happened in this video?" + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.194, + 0.622, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.217, + 0.623, + 0.292 + ], + "angle": 0, + "content": "The video shows a large group of trees in the woods that were burnt down in a fire. The fire was so intense that it caused the trees to fall over and burn in different directions. The video also shows a large plume of smoke coming from the forest, indicating that the fire was still active." + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.301, + 0.408, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.321, + 0.602, + 0.332 + ], + "angle": 0, + "content": "Describe the anomaly present in the video." + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.339, + 0.622, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.362, + 0.603, + 0.384 + ], + "angle": 0, + "content": "The video shows a large explosion in the woods, which is an anomaly." + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.094, + 0.854, + 0.105 + ], + "angle": 0, + "content": "A 32-minutes normal video" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.113, + 0.764, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.784, + 0.113, + 0.896, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.162, + 0.683, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.183, + 0.88, + 0.193 + ], + "angle": 0, + "content": "Describe the video's main features in detail." + }, + { + "type": "image", + "bbox": [ + 0.872, + 0.195, + 0.896, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.653, + 0.217, + 0.894, + 0.335 + ], + "angle": 0, + "content": "The video features a group of workers at a bank who are interacting with customers. The workers are seen helping customers with various transactions, including deposits, withdrawals, and account inquiries. They are dressed in uniforms and are sitting behind a counter. The customers are also seen talking to the workers and using their cell phones. The video captures a typical day at a bank and the interactions between the workers and customers." + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.339, + 0.683, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.358, + 0.829, + 0.368 + ], + "angle": 0, + "content": "Is this video normal or abnormal?" + }, + { + "type": "image", + "bbox": [ + 0.87, + 0.353, + 0.897, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.848, + 0.375, + 0.894, + 0.385 + ], + "angle": 0, + "content": "Normal." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.398, + 0.907, + 0.439 + ], + "angle": 0, + "content": "Figure 3. Qualitative results in Question-Answer diagrams, the red circles in the figures correspond to the bold text in the answers. From short video of only a dozen seconds to medium video of longer than one minute and long video of about half an hour, our model can reason well and understand the content." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.453, + 0.9, + 0.54 + ], + "angle": 0, + "content": "
MethodBaselineStage One Fine-tuningStage Two Fine-tuning
w/o Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Both
Total Acc. (%)↑14.8324.8323.7925.1225.8626.1027.5029.3128.9630.69
Temporal Acc. (%)↑26.7027.2027.7628.8129.6830.0230.7731.6033.5835.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.549, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Table 3. Ablation studies on Spatial Effective Token Selection (SETS), Temporal Effective Tokens Generation (TETG), and progressive training strategies. At different model training stages, starting from the baseline (w/o fine-tuning, w/o both), we compare the performance of only using SETs (w.SETS), only using TETG (w.TETG), and using both (w.Both) on the UCF-Crime benchmark. Stage One: Only anomaly video fine-tuning. Stage Two: Anomaly video fine-tuning + Fine-tuning with SETs (Best results are shown in bold)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.741 + ], + "angle": 0, + "content": "of abnormal events is efficiently aligned with LLMs and exhibits powerful anomaly video understanding capabilities. At the same time, we conduct fair comparisons with existing video understanding models [26, 27, 29, 39, 42, 74] (see Table 1), and we demonstrate competitive performance. It is worth noting that we use the fewest tokens among all methods to achieve the state-of-the-art results on both total and temporal accuracies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results on Cross-domain Dataset. For evaluating the robustness and generalization of the models, we additionally design a cross-domain benchmark. We conduct a fair comparison of our method with the baseline [31] and the existing in-domain methods on the proposed cross-domain benchmark. The results presented in Table 1 showcase a substantial performance improvement over existing methods on the cross-domain dataset, underscoring the exceptional generalization and temporal localization capabilities of our methodology. This clear superiority in performance serves" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.907, + 0.651 + ], + "angle": 0, + "content": "as a compelling validation of the robustness and adaptability of our approach across diverse domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Interaction with the Model. We take some interactions with our well-trained model for better evaluation. As shown in Figure 3, we demonstrate the performance of our model in addressing various video anomaly understanding challenges. To evaluate the model's effectiveness, we select videos of different durations: short (0 to 1 minute), medium (1 to 30 minutes), and long (over 30 minutes). This variety allows us to thoroughly assess the model's capabilities in handling diverse anomaly understanding scenarios. In the Road Accident video (left side in Figure 3), our method successfully identifies a car driving at high speed and detects people falling, even in low-resolution footage. For the Explosion video (middle in Figure 3), the model accurately predicts the scene and the anomaly in a medium-length video depicting an explosion. In a normal video exceeding 30 minutes (right side in Figure 3), we demonstrate the model's ability" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22700" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.483, + 0.221 + ], + "angle": 0, + "content": "
MethodLLMFine-tunedTotal Acc.(%)↑
Video-ChatGPTVicuna-7B-24.13
✓26.23
LLaMA-VID (Baseline)Vicuna-7B-14.83
✓23.10
OursVicuna-7B-25.12
✓30.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.227, + 0.485, + 0.268 + ], + "angle": 0, + "content": "Table 4. We fine-tune comparison models with our proposed UCF instruct-following data and evaluate the performance of these models before and after fine-tuning on the UCF benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.31 + ], + "angle": 0, + "content": "to focus on both global and local information by asking it to summarize the content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.402 + ], + "angle": 0, + "content": "Comparison on Other Benchmark. We additionally compare our method on another benchmark (MMEval [12]) about anomaly video understanding with LLMs from different aspects. We follow a fair evaluation on our proposed model and obtain the quantity results as shown in Table 2, which shows the superiority of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.255, + 0.426 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.509 + ], + "angle": 0, + "content": "We conduct extensive ablation studies to validate the effectiveness of the key components in our method: Spatial Effective Token Selection (SETS, Section 3.2) and Temporal Effective Tokens Generation (TETG, Section 3.3) on progressive training strategies in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Fine-tuning Stages. The utilization of our high-quality UCF instruct-following data has proven to enhance the model's performance. Fine-tuning with this dataset has effectively contributed to a notable accuracy compared with the baseline. As evidenced in Table 3, with our model designing (both SETS and TETG), the total accuracy for anomaly detection only achieves \\(25.12\\%\\) without any fine-tuning (denoted as Baseline). With anomaly video fine-tuning (denoted as Stage One Fine-tuning), the accuracy increases to \\(27.5\\%\\). Furthermore, an efficient tuning with SETS (denoted as Stage Two Fine-tuning) can achieve our final performance of \\(30.69\\%\\) total accuracy. Temporal accuracy also shows similar increasing patterns with the scaling tuning stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Effectiveness of Fine-tuning Data. For a fair comparison, we fine-tune some high-performance comparison models [31, 42] and compare them with our proposed UCF instruct-following data. As shown in Table 4, the performance of these comparison models has increased after fine-tuning, which proves the effectiveness of our proposed data. However, their performance is still not as good as our method, which proves the effectiveness of our proposed model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of SETS. Our proposed SETS demonstrates efficiency in extracting useful abnormal information, leading to performance enhancement. As illustrated in Table 3, with the SETS, the accuracy reaches \\(24.83\\%\\), \\(25.86\\%\\) and \\(29.31\\%\\) without fine-tuning, anomaly video fine-tuning and" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.906, + 0.151 + ], + "angle": 0, + "content": "
# Sample Rate K0.10.30.50.70.9
Total Acc.(%)↑23.6124.8330.6928.6727.27
Temporal Acc.(%)↑29.0329.9335.0031.2331.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.159, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Table 5. For the sample rate of tokens, \\( K \\) ratios in SETS, we sample the patch tokens with ordered distance at a fixed sample rate (0.5). The ablation indicates too large sampling rates cause too much context noise, and too small sampling rates lose visual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.908, + 0.43 + ], + "angle": 0, + "content": "fine-tuning with SETS, respectively, which far exceed the accuracy of the baseline. Its intuitive mechanism for information filtering can be further analyzed with reference to Figure 4. The initial video is often cluttered with irrelevant and deceptive data. For example, in Figure 4, case one illustrates a scenario where the overall structure is complex, yet only a small segment requires attention. The SETS effectively filters out the dynamic features that do not require attention. Similarly, as in case two, the abnormal area is quite small. Our proposed SETS mechanism effectively filters out redundant and irrelevant information, significantly enhancing the model's ability to accurately pinpoint and recognize abnormal situations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.553 + ], + "angle": 0, + "content": "We also conduct the ablation studies on \\( K \\) ratios about SETS as shown in Tabel 5. Too small or too large \\( K \\) ratios will cause performance degradation in both total and temporal accuracy. If \\( K \\) ratios is too small, redundant information will affect the effectiveness of aligning abnormal event information with corresponding captions. In contrast, if \\( K \\) ratios is too large, some important areas will be filtered out, resulting in information loss and suboptimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.751 + ], + "angle": 0, + "content": "Effectiveness of TETG. Our proposed TETG generates tokens directly in the text token space of LLMs as priors for each video, offering robust priors on the temporal aspects of abnormal events. The provision results in performance improvement without the necessity for fine-tuning. As shown in Table 3, the accuracy rises from \\(14.83\\%\\) to \\(23.79\\%\\). Fine-tuning with anomaly video and SETS, the results even achieve \\(26.10\\%\\) and \\(30.69\\%\\) independently, which manifests the effectiveness of TETG. Besides, the integration of SETS and TETG highlights the importance of leveraging spatial and temporal information effectively in anomaly detection systems, boosting the results to \\(25.15\\%\\), \\(27.50\\%\\) and \\(30.69\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.628, + 0.784 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Key tokens play key roles. To the best of our knowledge, we are the first to explore how to assign different learnable knowledge to different tokens for better alignment with LLMs on visual contents, thus promoting video anomaly detection and understanding (see Table 1 and Figure 3). We assign the most effective roles to different tokens in both spatial and temporal dimensions, enabling the model to handle" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.093, + 0.898, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.908, + 0.438 + ], + "angle": 0, + "content": "Figure 4. Visualization of the initial videos and our masked results. These two cases illustrate road accident scenarios: one occurring in a bustling street and the other in an empty suburb. Our SETS effectively filters redundant and irrelevant regions (with black patch-level masks)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.453, + 0.486, + 0.573 + ], + "angle": 0, + "content": "various tokens more efficiently. The video contains abundant but redundant information. Our proposed SETS and TETG effectively compress the spatial and temporal information of abnormal events respectively, and utilize the existing alignment mechanism of MLLMs at a very low cost to participate in LLMs' reasoning (see Table 3). Our exploration inspires more representation learning of MLLMs to facilitate downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.579, + 0.487, + 0.776 + ], + "angle": 0, + "content": "Data matters a lot. We construct instruct-following data for anomaly videos, containing approximately 4,000 videos, which is significantly less than the amount of baseline fine-tuning video data (e.g., over 90k videos for fine-tuning in baseline model [31]). We still achieve promising performance on both in-domain and cross-domain benchmarks (see Table 1). This relies on the high-quality Question-Answer pairs in our instruct-following data. Meanwhile, SETS also improves data quality during fine-tuning: visual areas irrelevant to Question-Answer pairs are filtered out (see Figure 4), which allows for significant performance improvements in the second stage of fine-tuning (see Section 3.4) with very few steps (less than 150 iterations)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Broader impacts. Video anomaly understanding has far-reaching implications across various sectors, including security, healthcare, industrial safety, and so on. By enhancing the ability to automatically identify and respond to unusual or suspicious activities in real-time, LLMs can significantly improve public safety, crime prevention, patient monitoring, hazard detection, loss prevention, traffic management, and urban planning. These systems offer substantial benefits in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.454, + 0.793, + 0.469 + ], + "angle": 0, + "content": "terms of operational efficiency and safety." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.636 + ], + "angle": 0, + "content": "Limitations. Although our model adeptly portrays the occurrence, type, and area of video abnormal events, it still faces challenges in detecting and describing certain complex scenes. Our strategy represents an early successful validation and investigation of large models for video anomaly identification and localization. Consequently, our method possesses significant potential for enhancement in recognizing diverse abnormal video scenes. These insights motivate us to continue pursuing more powerful and efficient video anomaly understanding technologies in the future, aiming to address more challenges in the real world [35, 41, 77]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.642, + 0.665 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.911, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel MLLM for understanding anomalies in videos with LLMs by aligning effective tokens in both temporal and spatial space. The proposed method includes Spatial Effective Token Selection (SETS) for identifying abnormal events in small areas of large scenes and Temporal Effective Tokens Generation (TETG) for addressing the sparseness of abnormal events in video time sequences. We also develop instruct-following data of video anomaly detection to fine-tune the model. Besides, evaluation on the video anomaly understanding benchmark and a proposed cross-domain benchmark demonstrates the effectiveness of the proposed method. It further presents a promising approach for video anomaly understanding using MLLMs, showcasing the potential of effective tokens for enhancing video understanding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22702" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.256 + ], + "angle": 0, + "content": "This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No.27209621), General Research Fund Scheme (Grant No.17202422, 17212923, 17215025), Theme-based Research (Grant No.T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.283, + 0.188, + 0.299 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.308, + 0.484, + 0.365 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.366, + 0.484, + 0.407 + ], + "angle": 0, + "content": "[2] Borislav Antić and Björn Ommer. Video parsing for abnormality detection. In 2011 International Conference on Computer Vision, pages 2415-2422, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.485, + 0.546 + ], + "angle": 0, + "content": "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[4] Ruichu Cai, Hao Zhang, Wen Liu, Shenghua Gao, and Zhifeng Hao. Appearance-motion memory consistency network for video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):938-946, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[5] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In Computer Vision - ECCV 2020, 16th European Conference, pages 329-345, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[6] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. AAAI2023, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.308, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.137, + 0.908, + 0.192 + ], + "angle": 0, + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.908, + 0.319 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.908, + 0.417 + ], + "angle": 0, + "content": "[12] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, Dajiu Huang, Jing Feng, Linli Chen, Can Zhang, Xuhuan Li, Hao Zhang, Jianhang Chen, Qimei Cui, and Xiaofeng Tao. Uncovering what, why and how: A comprehensive benchmark for causation understanding of video anomaly, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[13] JiaChang Feng, FaTing Hong, and WeiShi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, Rongrong Ji, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.908, + 0.673 + ], + "angle": 0, + "content": "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In IEEE International Conference on Computer Vision (ICCV), 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.908, + 0.715 + ], + "angle": 0, + "content": "[17] Rodolfo Gonzalez and Sandra Palais. A path-building procedure for iterative circuit computers. Technical report, 1962. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.908, + 0.774 + ], + "angle": 0, + "content": "[18] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, Z. Yang, Y. Zhang, and D. Tao. A survey on vision transformer. IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.908, + 0.83 + ], + "angle": 0, + "content": "[19] Mahmudul Hasan, Jonghyun Choi, jan Neumann, Amit K Roy-Chowdhury, and Larry Davis. Learning temporal regularity in video sequences. In Proceedings of IEEE Computer Vision and Pattern Recognition, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[20] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22703" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.19 + ], + "angle": 0, + "content": "[22] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[23] Louis Kratz and Ko Nishino. Anomaly detection in extremely crowded scenes using spatio-temporal motion pattern models. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1446-1453, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[24] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[26] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning, 2023. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[27] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[30] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisecquence learning with transformer for weakly supervised video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395-1403, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[31] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models, 2023. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[32] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[33] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jinfa Huang, Junwu Zhang, Munan Ning, and Li Yuan. Moellava: Mixture of experts for large vision-language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[35] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[37] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In European Conference on Computer Vision, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[40] Weixin Luo, Wen Liu, and Shenghua Gao. Remembering history with convolutional LSTM for anomaly detection. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 439-444, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[41] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.473 + ], + "angle": 0, + "content": "[42] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[43] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. CoRR, abs/1908.06351, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[44] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.792, + 0.589 + ], + "angle": 0, + "content": "[45] OpenAI. Introducing chatgpt. 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.907, + 0.714 + ], + "angle": 0, + "content": "[46] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[47] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14372-14381, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[48] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[49] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "[51] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.207 + ], + "angle": 0, + "content": "[52] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[53] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[54] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.309, + 0.484, + 0.364 + ], + "angle": 0, + "content": "[55] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Ying-Cong Chen. Hawk: Learning to understand open-world video anomalies, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.366, + 0.485, + 0.617 + ], + "angle": 0, + "content": "[56] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog applications, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.485, + 0.7 + ], + "angle": 0, + "content": "[57] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4975-4986, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.485, + 0.785 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.485, + 0.844 + ], + "angle": 0, + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.485, + 0.901 + ], + "angle": 0, + "content": "[60] Boyang Wan, Yuming Fang, Xue Xia, and Jiajie Mei. Weakly supervised video anomaly detection via center-guided discriminative learning. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[62] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.908, + 0.262 + ], + "angle": 0, + "content": "[63] Peng Wu and Jing Liu. Learning causal temporal relation and feature discrimination for anomaly detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[64] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[65] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[66] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[67] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[68] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detection, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.575 + ], + "angle": 0, + "content": "[69] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[70] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[71] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision (ECCV), pages 358-376. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.907, + 0.758 + ], + "angle": 0, + "content": "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[73] Muhammad Zaigham Zaheer, Jin-Ha Lee, Marcella Astrid, Arif Mahmood, and Seung-Ik Lee. Cleaning label noise with clusters for minimally supervised anomaly detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, June 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[74] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[75] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22705" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[76] Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4030-4034, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[77] Jiaqi Zhang, Yan Hu, Xiaojuan Qi, Ting Meng, Lihui Wang, Huazhu Fu, Mingming Yang, and Jiang Liu. Polar eyeball shape net for 3d posterior ocular shape representation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 180-190. Springer, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[78] Bin Zhao, Fei-Fei Li, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[79] Jianing Zhao, Jingjing Wang, Yujie Jin, Jiamin Luo, and Guodong Zhou. Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene-enhanced video large language model. In Proceedings of ACM MM 2024, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[80] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.483, + 0.555 + ], + "angle": 0, + "content": "[81] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In Proceedings of the 25th ACM International Conference on Multimedia, page 1933-1941, New York, NY, USA, 2017. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[82] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.483, + 0.683 + ], + "angle": 0, + "content": "[83] Lanfeng Zhong, Xin Liao, Shaoting Zhang, Xiaofan Zhang, and Guotai Wang. Vlm-cpl: Consensus pseudo labels from vision-language models for human annotation-free pathological image classification, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.483, + 0.726 + ], + "angle": 0, + "content": "[84] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[85] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.483, + 0.811 + ], + "angle": 0, + "content": "[86] Yi Zhu and Shawn D. Newsam. Motion-aware feature for improved video anomaly detection. In *British Machine Vision Conference (BMVC)*, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22706" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_origin.pdf b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88c02a0ef2905c1989b393d8cb14a88faf2db64d --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/8e282f6a-9c9d-4443-a523-c2f4bcdbbda6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51370e16d9d88b006dc2824764ae59785db9d504d3a12ac1b1d705085271f13d +size 1645389 diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/full.md b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..574a51c2cc8db793a51e0ecffc55977c742ed59c --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/full.md @@ -0,0 +1,360 @@ +# Aligning Effective Tokens with Video Anomaly in Large Language Models + +Yingxian Chen $^{1*}$ Jiahui Liu $^{1*}$ Ruidi Fan $^{1}$ Yanwei Li $^{2}$ Chirui Chang $^{1}$ Shizhen Zhao $^{1}$ +Wilton W.T.Fok $^{1}$ Xiaojuan Qi $^{1\dagger}$ Yik-Chung Wu $^{1\dagger}$ $^{1}$ The University of Hong Kong $^{2}$ The Chinese University of Hong Kong +{chenyx, liujh, xjqi, ycuw}@eee.hku.hk +*equal contributions †corresponding authors + +# Abstract + +Understanding abnormal events in videos is a vital and challenging task that has garnered significant attention in a wide range of applications. Although current video understanding Multi-modal Large Language Models (MLLMs) are capable of analyzing general videos, they often struggle to handle anomalies due to the spatial and temporal sparsity of abnormal events, where the redundant information always leads to suboptimal outcomes. To address these challenges, exploiting the representation and generalization capabilities of Vison Language Models (VLMs) and Large Language Models (LLMs), we propose VA-GPT, a novel MLLM designed for summarizing and localizing abnormal events in various videos. Our approach efficiently aligns effective tokens between visual encoders and LLMs through two key proposed modules: Spatial Effective Token Selection (SETS) and Temporal Effective Token Generation (TETG). These modules enable our model to effectively capture and analyze both spatial and temporal information associated with abnormal events, resulting in more accurate responses and interactions. Furthermore, we construct an instruction-following dataset specifically for fine-tuning video-anomaly-aware MLLMs, and introduce a cross-domain evaluation benchmark based on XD-Violence dataset. Our proposed method outperforms existing state-of-the-art methods on various benchmarks. + +# 1. Introduction + +Detecting and summarizing abnormal events in videos is critical and challenging, and it has garnered considerable attention across multiple research domains and real-world applications, such as security monitoring, video analysis, and crime detection. + +Although many traditional methods [4, 16, 37, 43, 47, 73, 86] have been widely explored for video anomaly detection, they exhibit substantial limitations in their effectiveness [22, 51, 60, 66, 71, 84]. These limitations manifest in two aspects: 1) Traditional video anomaly detection + +![](images/aa9e6ed126b55907fe3f366578df2edc7f5d3609e862de6448b78628b49de336.jpg) +Figure 1. Baseline video understanding MLLM feeds forward every visual token (yellow squares) equally to participate in fine-tuning and inference (top row). Different from it, our method focuses on the effective area (unobstructed area in medium video frames) in each frame and select the Spatial Effective Tokens (orange squares) for the LLM (see Section 3.2) (filtered tokens are shown as gray squares). At the same time, we generate anomaly-aware Temporal Effective Tokens (green squares) (see Section 3.3) based on the assigned anomaly scores (denoted as $s$ ) of each frame from a pretrained classifier for better temporal localization of anomalies. + +methods [6, 13, 57, 63, 64, 86] essentially approach the task as a closed-set detection and classification problem, inherently limiting their ability to achieve a comprehensive understanding and interpretation of anomalies; 2) These methods [2, 19, 23, 65, 68, 78] are restricted by a limited vocabulary, making it difficult for them to handle unseen or novel situations effectively. + +Recent advancements [1, 28, 34, 36, 45, 58] in Vision Language Models (VLMs) and Large Language Models (LLMs) have demonstrated remarkable capabilities in scene understanding and comprehensive analysis. Multimodal Large Language Models (MLLMs), particularly those designed for video understanding [27, 29, 31, 39, 42, 74], have achieved significant progress in general video analysis tasks. However, while these models exhibit strong performance + +in general video understanding, they fall short in accurately detecting and interpreting anomalies. + +For mitigating the above challenges, some works [12, 55, 67, 75, 79] proposed anomaly-aware video MLLMs to better understand the anomalies in videos. Although these models work well for detecting obvious abnormal events, such as fighting or fire, they typically struggle to effectively align abnormal regions with relevant captions which requires addressing spatial redundancy, and accurately identifying abnormal time intervals by mitigating temporal redundancy. This is because these methods treat all latent tokens with equal priority across spatial and temporal dimensions. This leads to performance degradation caused by redundant tokens unrelated to anomalies. However, in most cases, only small regions within a few frames contain the essential information help to identify an anomaly (as shown in Figure 1). Thus, we explore: How can multimodal architectures evolve selective token generation and processing mechanisms to dynamically prioritize anomaly-salient information while preserving comprehensive scene understanding capabilities? + +To address the aforementioned issues, we propose a new model named VA-GPT for analyzing various Videos for Abnormal events by aligning effective and accurate tokens with LLMs across both spatial and temporal dimensions. VA-GPT integrates two key components to identify effective visual tokens for alignment while eliminating redundant tokens that could hinder anomaly analysis and distract model from extracting useful information: 1) we develop the Spatial Effective Token Selection (SETS) module for identifying tokens corresponding to regions with challenges for aligning them with LLMs, while filtering out tokens associated with minor dynamics to remove redundancy. This is because we find that abnormal events often result in different visual changes and variations in local areas (see Figure 1); and 2) we propose the Temporal Effective Token Generation (TETG) module which employs a lightweight pre-trained classifier to assign a confidence score to each frame indicating the possibility of containing abnormal events. Then TETG generates efficient tokens with prior knowledge of the temporal information of abnormal events directly in the language space as additional input to the LLMs, effectively enhancing the model's temporal reasoning and understanding abilities about abnormal events. + +Furthermore, beyond conventional benchmarks (domain benchmark), we establish a new cross-domain evaluation protocol that systematically evaluates model robustness with domain shifts. Based on a novel video dataset, XD-Violence [64], we design comprehensive QAs about abnormal events which include different visual contents from our training data and integrate it as a new cross-domain benchmark. Meanwhile, we design temporal-information-oriented QAs on both in- and cross- domain benchmarks for evaluating temporal localization abilities. Comprehensive + +experiments demonstrate VA-GPT's superiority, achieving state-of-the-art performance in both in-domain anomaly localization and cross-domain generalization scenarios. + +The main contributions are summarized as follows: + +- We propose VA-GPT, a video-anomaly-aware MLLM for detecting and summarizing anomalies in various videos, which introduces the MLLM to the specific domain of video anomaly understanding. +- We introduce the SETS and TETG, which enable our MLLM to effectively capture both spatial and temporal information in video sequences, resulting in accurate understanding and localization of abnormal events. Meanwhile, we propose a new instruct-following dataset for video anomaly analysis and a comprehensive cross-domain evaluation benchmark for better evaluating the generalization abilities of MLLMs on video anomalies. +- Our extensive experiments demonstrate that our method outperforms existing state-of-the-art methods in various benchmarks, highlighting its effectiveness and potential for practical applications in video anomaly understanding. + +# 2. Related Work + +# 2.1. Large Language Models (LLMs) + +The domain of Natural Language Processing (NLP) has experienced significant progress, particularly with the emergence of Large Language Models (LLMs). The introduction of the Transformer architecture [3, 18, 59] is a critical turning point, followed by other influential language models [3, 10, 25, 80] that exhibited remarkable proficiency. Generative Pre-trained Transformers (GPT) [49] brought about a revolution in NLP by employing auto-regressive prediction, establishing itself as a powerful language modeling approach. More recent groundbreaking contributions, such as ChatGPT [45], GPT-4 [1], LaMDA [56] and LLaMA [58], have expanded the horizon even further. These models, trained on extensive textual data, display extraordinary performance in intricate linguistic tasks. + +# 2.2. Vision Language Models (VLMs) + +Progress in the fields of computer vision and natural language processing has given rise to the development of vision-language models (VLMs) [14, 21, 33, 34, 50, 61, 69]. These models combine visual and linguistic systems to facilitate cross-modal comprehension and reasoning. Notable examples include CLIP [50] which pairs BERT [10] with ViT [11]; BLIP-2 [28] incorporating Vision Transformer features into Flan-T5 [8]; MiniGPT4 [85], connecting BLIP-2 with Vicuna [28, 48]; PandaGPT [53], bridging ImageBind [15] with Vicuna. These models excel in tasks like image classification, captioning, and object detection [24, 53, 83]. Recent developments in vision-language models have extended into video processing with models + +![](images/f228863632d5c7156edebc26db7fe2beb86e0b1ee0ef22bd07d348a1bed38767.jpg) +Figure 2. Detailed illustration of our proposed model. When a video is fed into the model, patch embeddings $\square$ and class embeddings (c.ebd) $\square$ are extracted from all frames. 1) Based on the difference in patch embeddings between current frame and its neighbour frame, we can get a filter mask $\square$ to filter out unimportant visual tokens (dashed square $\square$ ) from current frame's visual tokens $\square$ , thereby selecting Spatial Effective Tokens $\square$ that are compressed with a projector with pooling into aligned content token $\square$ for each frame, meanwhile take attention with text input from users for resulting aligned context token $\square$ for each frame; 2) Based on class embeddings (c.ebd) $\square$ of all frames, we use a pre-trained Anomaly-aware Classifier to localize the time period of abnormal events, thereby generating Temporal Effective Tokens $\square$ to feed forward into the LLM. All of the resulting aligned tokens $\square \square \square$ are fed into the LLM for reasoning and inference of the whole model. + +like Video-Chat [29], Video-ChatGPT [42], Otter [26], Valley [39], mPLUG [27], Video-LLaMA [74], and LLaMAVID [31]. These systems enable interactive video querying, enhance comprehension through audio-visual-text alignment, and support comprehensive video analysis. In this paper, we leverage VLMs and LLMs to develop a novel approach for video anomaly understanding. + +# 2.3. Video Anomaly Understanding (VAU) + +Annotating surveillance video frames is labor-intensive, prompting researchers to explore alternatives: one-class learning [72], unsupervised anomaly detection without annotations [4, 5, 16, 20, 37, 38, 47], or weakly supervised methods using only video-level annotations [6, 13, 30, 57, 60, 62, 63, 71, 76]. In one-class learning, Luo et al. developed a ConvLSTM network for normal segment learning [40]. Several researchers employed Auto-Encoder networks to reconstruct normal frame features [5, 20, 81], while others implemented memory mechanisms [4, 16, 37, 47] or meta-learning [38] to enhance generalizability. For weakly supervised learning, Tian [57] sed multiple-instance learning to localize anomalous clips. Zhong et al. uutilized graph convolution networks, though with limited generalization capability. [82]. To address this, Ramachandra et al. developed a Siamese network for normal feature learning. Wan et al. and Zaheer et al. [60, 71]proposed clustering-based frameworks for anomaly identification. Recent studies have introduced new architectures for spatial-temporal feature ensemble learning [6, 13, 30, 57, 62, 63, 76]. However, these methods merely supply anomaly scores during inference, + +necessitating empirical threshold establishment on test sets to differentiate abnormal events. Recent research has begun exploring MLLMs to enhance models' capabilities in identifying and describing anomalies [12, 55, 67, 75, 79]. + +# 3. Method + +# 3.1. Overview + +Task. Video anomaly understanding MLLMs aims to determine whether an input video contains abnormal events, meanwhile describing and interacting with the temporal localization and the entire process of the detected abnormal events (if has). We train the model with an instruct-following dataset based on abnormal videos [54], so that the model can better align the tokens between visual encoders and MLLMs for presenting and generalizing information about abnormal events. + +Pipeline. Considering the video understanding MLLM framework as shown in Figure 2, taking a video (contains $T$ frames) as input, a frozen ViT-based [11] visual encoder (CLIP [52]) extracts visual tokens $\mathbf{X}^t$ from each video frame $\mathbf{V}^t (t = 1,\dots,T)$ . For $\mathbf{X}^t = \{\mathbf{x}_i^t\}_{i = 1,\dots ,N}$ of the current frame, there are $N$ visual tokens corresponding to equal amounts of image patches. Modality alignment converts the processed visual tokens $\mathbf{X}^t$ into the semantic space of LLMs. At the same time, text prompts are processed and encoded as text tokens into the same semantic space and serve as a part of input to LLMs. Our key design on models consists of (1) selecting Spatial Effective Tokens $\mathbf{X}^{*t}$ (SET) from $\mathbf{X}^t$ for each frame participating in fine-tuning and inference + +instead of $\mathbf{X}^t$ (see Section 3.2); and (2) generating Temporal Effective Tokens $\mathbf{S}^{*t}$ (TET) as anomaly-aware temporal priors, participating in inference to facilitate the temporal localization of abnormal events for LLMs (see Section 3.3). In addition, we produce high-quality instruct-following data on abnormal videos and develop a training strategy for it to maximize the effectiveness of our proposed method (see Section 3.4). + +# 3.2. Spatial Effective Token Selection (SETS) + +In classical video classification tasks, context and relationships are critical. However, in our MLLM setting, beyond leveraging contextual information, the most crucial problem is aligning the visual and language modalities. Therefore, the key aspect of our design is to extract useful information for effectively aligning visual tokens with the LLM. Since text captions primarily describe anomaly events, which occupy only a small portion of the entire video, aligning all visual patterns with text tokens would be unreasonable and computationally heavy. Thus, we are the first to propose a novel token selection method SETS to achieve efficient and effective alignment. + +Inter-frame Difference. For a video, we believe that areas with large changes in adjacent frames are more worthy of attention. As illustrated in Figure 2, for each frame $\mathbf{V}^t$ of the video, we can regard its previous frame $\mathbf{V}^{t-1}$ as the reference frame for investigating the difference between current timestamp and previous timestamp. Employing DINOV2 [46] as the feature extractor, denoted as $\mathcal{F}_{\mathrm{E}}$ , we can extract patch embeddings: + +$$ +\mathbf {F} ^ {t}, \mathbf {F} ^ {t - 1} = \mathcal {F} _ {\mathrm {E}} (\mathbf {V} ^ {t}), \mathcal {F} _ {\mathrm {E}} (\mathbf {V} ^ {t - 1}), \tag {1} +$$ + +where $\mathbf{F}^t, \mathbf{F}^{t-1} \in \mathbb{R}^{N \times C}$ are the extracted embeddings ( $N$ indicates the number of image patches and $C$ indicates the channels). Thanks to the distinction and stability of the extracted features, we calculate their patch-wise distances as the inter-frame difference map of the current frame: + +$$ +\mathbf {D} ^ {t} = \operatorname {d i s} \left(\mathbf {F} ^ {t}, \mathbf {F} ^ {t - 1}\right), \tag {2} +$$ + +where $dis(\cdot)$ indicates Manhattan distance [17] and $\mathbf{D}^t\in$ $\mathbb{R}^N$ indicates the distances between corresponding patch pairs in neighbour frames. + +Select Spatial Effective Tokens. According to the interframe difference map $\mathbf{D}^t$ , we can set up a vector $\mathbf{M}^t = [m_1^t, m_2^t, \dots, m_N^t]$ to record the difference of each patch, where top $K$ ratio of elements with the largest distance are assigned with the value of 1, and the rest are assigned as 0. Thus we get a mask for filtering and updating the visual tokens as: + +$$ +\mathbf {X} ^ {* t} = \left\{\mathbf {x} _ {i} ^ {t} \mid m _ {i} ^ {t} = 1, m _ {i} ^ {t} \in \mathbf {M} ^ {t} \right\}, \tag {3} +$$ + +where $\mathbf{X}^{*t}$ contains the selected Spatial Effective Tokens (SET) which are fed into subsequent processing instead of + +$\mathbf{X}^t$ as shown in Figure 2. SET can efficiently isolate the regions highly related to the abnormal events to participate in both fine-tuning and inference. + +# 3.3. Temporal Effective Token Generation (TETG) + +Anomaly-aware Classifier. We design a simple but effective MLP $\mathcal{F}_{\mathrm{A}}$ for learning whether each frame is related to an abnormal event. For the class embeddings (denoted as $\mathbf{z}$ ) extracted from the feature encoder, we can split them based on training video caption into normal and anomaly embeddings, denoted as $\mathbf{z}^{\mathrm{n}}$ and $\mathbf{z}^{\mathrm{a}}$ , respectively. Thus we can optimize $\mathcal{F}_{\mathrm{A}}$ using a binary classification loss: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {z} \sim \mathbf {Z} ^ {\mathrm {n}}} \left[ - \log \frac {1}{1 + \exp^ {- \mathcal {F} _ {\mathrm {A}} (\mathbf {z})}} \right] + \mathbb {E} _ {\mathbf {z} \sim \mathbf {Z} ^ {\mathrm {a}}} \left[ - \log \frac {\exp^ {- \mathcal {F} _ {\mathrm {A}} (\mathbf {z})}}{1 + \exp^ {- \mathcal {F} _ {\mathrm {A}} (\mathbf {z})}} \right]. \tag {4} +$$ + +The anomaly-aware classifier predicts whether each frame is related to anomalies in a video, which can bring additional important prior knowledge to LLMs at a very low cost to facilitate its inference. + +Generate Temporal Effective Tokens. Since the information drawn from the anomaly-aware classifier is explicit, we can easily project it to LLMs' text token space through natural languages. Based on the prediction results of the anomaly-aware classifier, we select the first and last frames' timestamps with high confidence of containing abnormal events, denoted as $$ and $$ , respectively. Then we tokenize them with a template as: "Known common crime types are: 'Shooting', 'Arson', 'Arrest', ... There is one of the crime types occurring from $$ to $$ ", resulting Temporal Effective Tokens (TET) in the text token space of LLMs. During inference, with the well-trained lightweight anomaly-aware classifier, TET is used as an additional input to participate in the forward process of the LLM to provide prior knowledge about abnormal events temporally (as shown in Figure 2). + +# 3.4. Training Strategy + +For modality alignment and instruction tuning, we follow the baseline [31] to ensure visual features are well aligned with the language space. In this work, the training strategy can be divided into two stages: 1) Stage One: Fine-tuning with anomaly video data, and 2) Stage Two: Aligning Spatial Effective Tokens with LLMs. + +Fine-tuning with Video Anomaly Data. For enhancing the abnormal scene understanding of LLMs, we construct the UCF-crime [70] Question-Answer pairs for fine-tuning. We also mix different instruction pairs from various sources [32], including text conversations, single/multi-turn visual Question-Answer pairs, and video Question-Answer pairs. Different formats for text, image, and video inputs are adopted, and the image token is randomly inserted at the beginning or end of user input during training. All modules, + +
MethodLLMIn-domainCross-domain
Total Acc.(%)Temporal Acc.(%)Total Acc.(%)Temporal Acc.(%)
Video-Chat [29]Vicuna-7B22.41---
Video-ChatGPT [42]Vicuna-7B24.1328.5124.0029.10
Otter [26]LLaMa-7B22.4122.1725.2023.80
Valley [39]Vicuna-7B20.3414.4821.0020.20
mPLUG [27]LLaMa-7B22.76---
Video-LLaMA2 [7]Vicuna-7B21.3826.6224.2023.00
Hawkeye [79]LLaVA-7B28.6030.0025.3028.50
LLaMA-VID [31]Vicuna-7B14.8326.7018.8023.60
VA-GPT (Ours)Vicuna-7B30.6935.0026.2031.02
+ +Table 1. Comparing on in-domain (UCF-Crime [54]) and the proposed cross-domain (XD-Violence [64]) benchmarks, our method significantly outperforms other models and achieve the state-of-the-art performance (accuracy is detonated as Acc.) on anomaly video understanding and temporal localization with LLMs (Best results are shown in bold). + +
MethodDescriptionCausesEffect
A-Guardian [12]79.6558.9250.64
Ours80.8359.5551.08
+ +Table 2. We evaluate our method on the MMEval [12] benchmark and show that our method outperforms the related previous method in different aspects. + +except the frozen visual encoder, are optimized in this stage. After fine-tuning, the LLM has a preferred perception of anomalies, thus ensuring the effectiveness of the temporal effective tokens (see Section 3.3) during inference. More dataset details are illustrated in Section 4. + +Aligning Spatial Effective Tokens with LLMs. For abnormal video scenes, most areas would not be aligned with languages. Therefore, we implement an additional fine-tuning step. This step involves utilizing Spatial Effective Tokens (see Section 3.2) derived from each video frame within the UCF-Crime dataset. By incorporating these tokens, we aim to provide the model with a more refined understanding of the spatial context of anomalies. It also brings efficient optimization, and the alignment here is only designed for very short-term fine-tuning, which can greatly improve the model's ability to detect and understand anomalies. + +# 4. Experiments + +Datasets. We fine-tune our model on our proposed instruct-following format [34] training dataset including 4077 videos and 7730 images based on UCF-Crime [54] dataset. We evaluate the models on two anomaly video understanding benchmarks: UCF-Crime [54] for in-domain evaluation and a proposed benchmark designed based on XD-Violence [64] dataset for cross-domain evaluation, respectively. More details are shown in the supplementary. + +Benchmarks and Metrics. To evaluate the ability to review videos and identify anomalies, we utilize a video anomaly understanding evaluation from Video-Bench [44] to assess the + +temporal comprehensive ability, which contains nature language based Question-Answer pairs from UCF-Crime [54] dataset. Meanwhile, in order to evaluate the model's cross-domain video anomaly understanding ability, we contribute Question-Answer pairs as an extra benchmark based on the XD-Violence dataset. These Question-Answer pairs encompass four options, with each presenting the anomaly category and the respective time intervals during which the anomalies transpire. For each benchmark, we design different sets of questions for two evaluations: one is an overall evaluation of abnormal event detection and understanding, and the other is a special evaluation focusing on temporal localization ability, measured by question answering accuracy (denoted as Total Acc. and Temporal Acc., respectively, higher is better). + +Implementation Details. For network structure, we incorporate the pre-trained CLIP [52] and DINOv2 [46] as the visual encoder and Qformer [9] as the text decoder. We follow [31] to freeze the encoder during the modality alignment and to optimize the trainable parameters using anomaly videos and instructions for instruction tuning. During the training process, our model utilizes PyTorch on four NVIDIA A100 GPUs. We employ the AdamW optimizer with a batch size of 64. The learning schedule is set to cosine decay, with a learning rate of 2e-5 and a total of 1 epoch. + +# 4.1. Main Results + +Results on In-domain Dataset. We first evaluate our method on the in-domain dataset, where the test set belongs to the same style and recording mode as the data used for training in Section 3.4. As shown in Table 1, compared with the baseline [31], with fewer visual embedding tokens and temporal effective tokens, our method brings more than double the performance improvement on total accuracy, also brings a significant increase in temporal localization. Driven by our proposed training strategy and designed effective tokens, more pure and effective visual-semantic information + +![](images/c28d2ffd063a24261cf9b218ac80213ecd9ec39446deca8bbf49b74f1aff9a22.jpg) +A 11-seconds video about Road Accident + +![](images/1001798447d95625b5f868e07fe148bd0eb3bc8938ed35a3ffceb4d3bd70846f.jpg) +Describe the video in detail. + +![](images/056f08cac14508ecd68c98aba133246a4e9cbbf45a565da7b9e49c85f56adf5e.jpg) + +The video shows a group of people walking down a sidewalk and suddenly a car appears driving very fast towards them. Some people fall onto the ground as the car crashes into them. + +![](images/66c32a78cc0d4870d9f3beb41b7cd1fb6ca333003dd36e684f33d37dbe85b790.jpg) +Are there any anomalies in the video? + +![](images/7d3ea04034047572675ca6b14e656d1bcacbb607a06b50d88a8527a399aecb73.jpg) + +Yes, there is a car accident in the video. It shows a car flipping over and people running away from the scene. + +![](images/b6c5b72592bb6bc56ecb6bc7fe41d3e5133aab0994e1c33b530bd1896d516c22.jpg) +A 1.5-minutes video about Explosion + +![](images/970041b38aaa198a4b4e4a61b179813fd7c428ddc02e2a520ddcfecc50fe793e.jpg) + +![](images/7ac0a8b3f002a356b2acab33ec6893d226481b388a5b68f893d9283b14a68ebf.jpg) +What happened in this video? + +![](images/18afc59029baa429109e6ee3c3a4863d4b1aa8e6550e03eef4f5628e75c96e61.jpg) + +The video shows a large group of trees in the woods that were burnt down in a fire. The fire was so intense that it caused the trees to fall over and burn in different directions. The video also shows a large plume of smoke coming from the forest, indicating that the fire was still active. + +![](images/092f3be4f298fd70a9ae29120af13424adc5c052750c3a4fca4c6702d5122a2d.jpg) +Describe the anomaly present in the video. + +![](images/3a87329e5ab5445c48b1104ae5214e13e63a1412d4e179df459760a44802e052.jpg) + +The video shows a large explosion in the woods, which is an anomaly. + +![](images/7b06b80aaaafc591b0b549b12a8e7b126478728542db979e05c7e713d1569375.jpg) +A 32-minutes normal video + +![](images/ae089a3d114f6671e9462e8a396ea8dfb86f3627f8df15504542da22b21da066.jpg) + +![](images/94ed00ca72b5bc96a30d678acc08ac7002cea474452340af29cc312023a60214.jpg) +Describe the video's main features in detail. + +![](images/0b391f11031e1277eaafab436f151593ef7d2a011e0479513bf232f84f3448ea.jpg) + +The video features a group of workers at a bank who are interacting with customers. The workers are seen helping customers with various transactions, including deposits, withdrawals, and account inquiries. They are dressed in uniforms and are sitting behind a counter. The customers are also seen talking to the workers and using their cell phones. The video captures a typical day at a bank and the interactions between the workers and customers. + +![](images/a08efa9fa111c440e2d52373bf93f6b8b321a8fde23740f29e34245ac33e1b50.jpg) +Is this video normal or abnormal? + +![](images/3dfcba1ea0d44f79abec27da9a772b6ba79f20a00ad9fd57b7eb5043500c8b8b.jpg) +Normal. +Figure 3. Qualitative results in Question-Answer diagrams, the red circles in the figures correspond to the bold text in the answers. From short video of only a dozen seconds to medium video of longer than one minute and long video of about half an hour, our model can reason well and understand the content. + +
MethodBaselineStage One Fine-tuningStage Two Fine-tuning
w/o Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Both
Total Acc. (%)↑14.8324.8323.7925.1225.8626.1027.5029.3128.9630.69
Temporal Acc. (%)↑26.7027.2027.7628.8129.6830.0230.7731.6033.5835.00
+ +Table 3. Ablation studies on Spatial Effective Token Selection (SETS), Temporal Effective Tokens Generation (TETG), and progressive training strategies. At different model training stages, starting from the baseline (w/o fine-tuning, w/o both), we compare the performance of only using SETs (w.SETS), only using TETG (w.TETG), and using both (w.Both) on the UCF-Crime benchmark. Stage One: Only anomaly video fine-tuning. Stage Two: Anomaly video fine-tuning + Fine-tuning with SETs (Best results are shown in bold). + +of abnormal events is efficiently aligned with LLMs and exhibits powerful anomaly video understanding capabilities. At the same time, we conduct fair comparisons with existing video understanding models [26, 27, 29, 39, 42, 74] (see Table 1), and we demonstrate competitive performance. It is worth noting that we use the fewest tokens among all methods to achieve the state-of-the-art results on both total and temporal accuracies. + +Results on Cross-domain Dataset. For evaluating the robustness and generalization of the models, we additionally design a cross-domain benchmark. We conduct a fair comparison of our method with the baseline [31] and the existing in-domain methods on the proposed cross-domain benchmark. The results presented in Table 1 showcase a substantial performance improvement over existing methods on the cross-domain dataset, underscoring the exceptional generalization and temporal localization capabilities of our methodology. This clear superiority in performance serves + +as a compelling validation of the robustness and adaptability of our approach across diverse domains. + +Interaction with the Model. We take some interactions with our well-trained model for better evaluation. As shown in Figure 3, we demonstrate the performance of our model in addressing various video anomaly understanding challenges. To evaluate the model's effectiveness, we select videos of different durations: short (0 to 1 minute), medium (1 to 30 minutes), and long (over 30 minutes). This variety allows us to thoroughly assess the model's capabilities in handling diverse anomaly understanding scenarios. In the Road Accident video (left side in Figure 3), our method successfully identifies a car driving at high speed and detects people falling, even in low-resolution footage. For the Explosion video (middle in Figure 3), the model accurately predicts the scene and the anomaly in a medium-length video depicting an explosion. In a normal video exceeding 30 minutes (right side in Figure 3), we demonstrate the model's ability + +
MethodLLMFine-tunedTotal Acc.(%)↑
Video-ChatGPTVicuna-7B-24.13
✓26.23
LLaMA-VID (Baseline)Vicuna-7B-14.83
✓23.10
OursVicuna-7B-25.12
✓30.69
+ +to focus on both global and local information by asking it to summarize the content. + +Comparison on Other Benchmark. We additionally compare our method on another benchmark (MMEval [12]) about anomaly video understanding with LLMs from different aspects. We follow a fair evaluation on our proposed model and obtain the quantity results as shown in Table 2, which shows the superiority of our method. + +# 4.2. Ablation Studies + +We conduct extensive ablation studies to validate the effectiveness of the key components in our method: Spatial Effective Token Selection (SETS, Section 3.2) and Temporal Effective Tokens Generation (TETG, Section 3.3) on progressive training strategies in Section 3.4. + +Fine-tuning Stages. The utilization of our high-quality UCF instruct-following data has proven to enhance the model's performance. Fine-tuning with this dataset has effectively contributed to a notable accuracy compared with the baseline. As evidenced in Table 3, with our model designing (both SETS and TETG), the total accuracy for anomaly detection only achieves $25.12\%$ without any fine-tuning (denoted as Baseline). With anomaly video fine-tuning (denoted as Stage One Fine-tuning), the accuracy increases to $27.5\%$ . Furthermore, an efficient tuning with SETS (denoted as Stage Two Fine-tuning) can achieve our final performance of $30.69\%$ total accuracy. Temporal accuracy also shows similar increasing patterns with the scaling tuning stages. + +Effectiveness of Fine-tuning Data. For a fair comparison, we fine-tune some high-performance comparison models [31, 42] and compare them with our proposed UCF instruct-following data. As shown in Table 4, the performance of these comparison models has increased after fine-tuning, which proves the effectiveness of our proposed data. However, their performance is still not as good as our method, which proves the effectiveness of our proposed model. + +Effectiveness of SETS. Our proposed SETS demonstrates efficiency in extracting useful abnormal information, leading to performance enhancement. As illustrated in Table 3, with the SETS, the accuracy reaches $24.83\%$ , $25.86\%$ and $29.31\%$ without fine-tuning, anomaly video fine-tuning and + +Table 4. We fine-tune comparison models with our proposed UCF instruct-following data and evaluate the performance of these models before and after fine-tuning on the UCF benchmark. + +
# Sample Rate K0.10.30.50.70.9
Total Acc.(%)↑23.6124.8330.6928.6727.27
Temporal Acc.(%)↑29.0329.9335.0031.2331.03
+ +Table 5. For the sample rate of tokens, $K$ ratios in SETS, we sample the patch tokens with ordered distance at a fixed sample rate (0.5). The ablation indicates too large sampling rates cause too much context noise, and too small sampling rates lose visual information. + +fine-tuning with SETS, respectively, which far exceed the accuracy of the baseline. Its intuitive mechanism for information filtering can be further analyzed with reference to Figure 4. The initial video is often cluttered with irrelevant and deceptive data. For example, in Figure 4, case one illustrates a scenario where the overall structure is complex, yet only a small segment requires attention. The SETS effectively filters out the dynamic features that do not require attention. Similarly, as in case two, the abnormal area is quite small. Our proposed SETS mechanism effectively filters out redundant and irrelevant information, significantly enhancing the model's ability to accurately pinpoint and recognize abnormal situations. + +We also conduct the ablation studies on $K$ ratios about SETS as shown in Tabel 5. Too small or too large $K$ ratios will cause performance degradation in both total and temporal accuracy. If $K$ ratios is too small, redundant information will affect the effectiveness of aligning abnormal event information with corresponding captions. In contrast, if $K$ ratios is too large, some important areas will be filtered out, resulting in information loss and suboptimal performance. + +Effectiveness of TETG. Our proposed TETG generates tokens directly in the text token space of LLMs as priors for each video, offering robust priors on the temporal aspects of abnormal events. The provision results in performance improvement without the necessity for fine-tuning. As shown in Table 3, the accuracy rises from $14.83\%$ to $23.79\%$ . Fine-tuning with anomaly video and SETS, the results even achieve $26.10\%$ and $30.69\%$ independently, which manifests the effectiveness of TETG. Besides, the integration of SETS and TETG highlights the importance of leveraging spatial and temporal information effectively in anomaly detection systems, boosting the results to $25.15\%$ , $27.50\%$ and $30.69\%$ , respectively. + +# 5. Discussion + +Key tokens play key roles. To the best of our knowledge, we are the first to explore how to assign different learnable knowledge to different tokens for better alignment with LLMs on visual contents, thus promoting video anomaly detection and understanding (see Table 1 and Figure 3). We assign the most effective roles to different tokens in both spatial and temporal dimensions, enabling the model to handle + +![](images/f790a42a12bb676fc8dee77ed33d810b32d38082bda8637bd4f07397f2bc34e0.jpg) +Figure 4. Visualization of the initial videos and our masked results. These two cases illustrate road accident scenarios: one occurring in a bustling street and the other in an empty suburb. Our SETS effectively filters redundant and irrelevant regions (with black patch-level masks). + +various tokens more efficiently. The video contains abundant but redundant information. Our proposed SETS and TETG effectively compress the spatial and temporal information of abnormal events respectively, and utilize the existing alignment mechanism of MLLMs at a very low cost to participate in LLMs' reasoning (see Table 3). Our exploration inspires more representation learning of MLLMs to facilitate downstream tasks. + +Data matters a lot. We construct instruct-following data for anomaly videos, containing approximately 4,000 videos, which is significantly less than the amount of baseline fine-tuning video data (e.g., over 90k videos for fine-tuning in baseline model [31]). We still achieve promising performance on both in-domain and cross-domain benchmarks (see Table 1). This relies on the high-quality Question-Answer pairs in our instruct-following data. Meanwhile, SETS also improves data quality during fine-tuning: visual areas irrelevant to Question-Answer pairs are filtered out (see Figure 4), which allows for significant performance improvements in the second stage of fine-tuning (see Section 3.4) with very few steps (less than 150 iterations). + +Broader impacts. Video anomaly understanding has far-reaching implications across various sectors, including security, healthcare, industrial safety, and so on. By enhancing the ability to automatically identify and respond to unusual or suspicious activities in real-time, LLMs can significantly improve public safety, crime prevention, patient monitoring, hazard detection, loss prevention, traffic management, and urban planning. These systems offer substantial benefits in + +terms of operational efficiency and safety. + +Limitations. Although our model adeptly portrays the occurrence, type, and area of video abnormal events, it still faces challenges in detecting and describing certain complex scenes. Our strategy represents an early successful validation and investigation of large models for video anomaly identification and localization. Consequently, our method possesses significant potential for enhancement in recognizing diverse abnormal video scenes. These insights motivate us to continue pursuing more powerful and efficient video anomaly understanding technologies in the future, aiming to address more challenges in the real world [35, 41, 77]. + +# 6. Conclusions + +In this paper, we propose a novel MLLM for understanding anomalies in videos with LLMs by aligning effective tokens in both temporal and spatial space. The proposed method includes Spatial Effective Token Selection (SETS) for identifying abnormal events in small areas of large scenes and Temporal Effective Tokens Generation (TETG) for addressing the sparseness of abnormal events in video time sequences. We also develop instruct-following data of video anomaly detection to fine-tune the model. Besides, evaluation on the video anomaly understanding benchmark and a proposed cross-domain benchmark demonstrates the effectiveness of the proposed method. It further presents a promising approach for video anomaly understanding using MLLMs, showcasing the potential of effective tokens for enhancing video understanding tasks. + +# Acknowledgements + +This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No.27209621), General Research Fund Scheme (Grant No.17202422, 17212923, 17215025), Theme-based Research (Grant No.T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2 +[2] Borislav Antić and Björn Ommer. Video parsing for abnormality detection. In 2011 International Conference on Computer Vision, pages 2415-2422, 2011. 1 +[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2 +[4] Ruichu Cai, Hao Zhang, Wen Liu, Shenghua Gao, and Zhifeng Hao. Appearance-motion memory consistency network for video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):938-946, 2021. 1, 3 +[5] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In Computer Vision - ECCV 2020, 16th European Conference, pages 329-345, 2022. 3 +[6] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. AAAI2023, 2022. 1, 3 +[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5 +[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, + +Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 2 +[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3 +[12] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, Dajiu Huang, Jing Feng, Linli Chen, Can Zhang, Xuhuan Li, Hao Zhang, Jianhang Chen, Qimei Cui, and Xiaofeng Tao. Uncovering what, why and how: A comprehensive benchmark for causation understanding of video anomaly, 2024. 2, 3, 5, 7 +[13] JiaChang Feng, FaTing Hong, and WeiShi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3 +[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, Rongrong Ji, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining, 2022. 2 +[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all, 2023. 2 +[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In IEEE International Conference on Computer Vision (ICCV), 2019. 1, 3 +[17] Rodolfo Gonzalez and Sandra Palais. A path-building procedure for iterative circuit computers. Technical report, 1962. 4 +[18] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, Z. Yang, Y. Zhang, and D. Tao. A survey on vision transformer. IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1-1, 2020. 2 +[19] Mahmudul Hasan, Jonghyun Choi, jan Neumann, Amit K Roy-Chowdhury, and Larry Davis. Learning temporal regularity in video sequences. In Proceedings of IEEE Computer Vision and Pattern Recognition, 2016. 1 +[20] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 + +[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. 2 +[22] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection, 2023. 1 +[23] Louis Kratz and Ko Nishino. Anomaly detection in extremely crowded scenes using spatio-temporal motion pattern models. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1446-1453, 2009. 1 +[24] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models, 2023. 2 +[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension, 2019. 2 +[26] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning, 2023. 3, 5, 6 +[27] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections, 2022. 1, 3, 5, 6 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 1, 2 +[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 1, 3, 5, 6 +[30] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisecquence learning with transformer for weakly supervised video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395-1403, 2022. 3 +[31] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models, 2023. 1, 3, 4, 5, 6, 7, 8 +[32] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2023. 4 +[33] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jinfa Huang, Junwu Zhang, Munan Ning, and Li Yuan. Moellava: Mixture of experts for large vision-language models, 2024. 2 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5 +[35] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 8 +[36] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation + +models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 1 +[37] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1, 3 +[38] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In European Conference on Computer Vision, 2020. 3 +[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability, 2023. 1, 3, 5, 6 +[40] Weixin Luo, Wen Liu, and Shenghua Gao. Remembering history with convolutional LSTM for anomaly detection. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 439-444, 2017. 3 +[41] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 8 +[42] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 1, 3, 5, 6, 7 +[43] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. CoRR, abs/1908.06351, 2019. 1 +[44] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 5 +[45] OpenAI. Introducing chatgpt. 2022. 1, 2 +[46] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 4, 5 +[47] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14372-14381, 2020. 1, 3 +[48] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4, 2023. 2 +[49] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2 + +[51] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1 +[52] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, 2022. 3, 5 +[53] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all, 2023. 2 +[54] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 5 +[55] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Ying-Cong Chen. Hawk: Learning to understand open-world video anomalies, 2024. 2, 3 +[56] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog applications, 2022. 2 +[57] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4975-4986, 2021. 1, 3 +[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1, 2 +[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2 +[60] Boyang Wan, Yuming Fang, Xue Xia, and Jiajie Mei. Weakly supervised video anomaly detection via center-guided discriminative learning. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2020. 1, 3 + +[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks, 2023. 2 +[62] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3 +[63] Peng Wu and Jing Liu. Learning causal temporal relation and feature discrimination for anomaly detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. 1, 3 +[64] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5 +[65] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection, 2023. 1 +[66] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection, 2024. 1 +[67] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models, 2024. 2, 3 +[68] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detection, 2024. 1 +[69] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training, 2021. 2 +[70] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 4 +[71] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision (ECCV), pages 358-376. Springer, 2020. 1, 3 +[72] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection, 2022. 3 +[73] Muhammad Zaigham Zaheer, Jin-Ha Lee, Marcella Astrid, Arif Mahmood, and Seung-Ik Lee. Cleaning label noise with clusters for minimally supervised anomaly detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, June 2020. 1 +[74] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 1, 3, 6 +[75] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and + +Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity, 2024. 2, 3 +[76] Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4030-4034, 2019. 3 +[77] Jiaqi Zhang, Yan Hu, Xiaojuan Qi, Ting Meng, Lihui Wang, Huazhu Fu, Mingming Yang, and Jiang Liu. Polar eyeball shape net for 3d posterior ocular shape representation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 180-190. Springer, 2023. 8 +[78] Bin Zhao, Fei-Fei Li, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 1 +[79] Jianing Zhao, Jingjing Wang, Yujie Jin, Jiamin Luo, and Guodong Zhou. Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene-enhanced video large language model. In Proceedings of ACM MM 2024, 2024. 2, 3, 5 +[80] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2023. 2 +[81] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In Proceedings of the 25th ACM International Conference on Multimedia, page 1933-1941, New York, NY, USA, 2017. Association for Computing Machinery. 3 +[82] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[83] Lanfeng Zhong, Xin Liao, Shaoting Zhang, Xiaofan Zhang, and Guotai Wang. Vlm-cpl: Consensus pseudo labels from vision-language models for human annotation-free pathological image classification, 2024. 2 +[84] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection, 2023. 1 +[85] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2 +[86] Yi Zhu and Shawn D. Newsam. Motion-aware feature for improved video anomaly detection. In *British Machine Vision Conference (BMVC)*, 2019. 1 \ No newline at end of file diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/images.zip b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..20f841dfe7b2dfeb118ee6bcb6453856a0e5cdf8 --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad0a148f8f1d910413b9ebf0cf935dc84f9809a729e1dda192e5e1353b7e391a +size 510462 diff --git a/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/layout.json b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b8a6b75c96a87f987dfd82eafc9afb7be8047f8f --- /dev/null +++ b/2025/Aligning Effective Tokens with Video Anomaly in Large Language Models/layout.json @@ -0,0 +1,9863 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 534, + 121 + ], + "type": "text", + "content": "Aligning Effective Tokens with Video Anomaly in Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "spans": [ + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": "Yingxian Chen" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Jiahui Liu" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Ruidi Fan" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Yanwei Li" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Chirui Chang" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Shizhen Zhao" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " \nWilton W.T.Fok" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Xiaojuan Qi" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": " Yik-Chung Wu" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 141, + 520, + 216 + ], + "type": "text", + "content": "The Chinese University of Hong Kong \n{chenyx, liujh, xjqi, ycuw}@eee.hku.hk \n*equal contributions †corresponding authors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 266, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 553 + ], + "type": "text", + "content": "Understanding abnormal events in videos is a vital and challenging task that has garnered significant attention in a wide range of applications. Although current video understanding Multi-modal Large Language Models (MLLMs) are capable of analyzing general videos, they often struggle to handle anomalies due to the spatial and temporal sparsity of abnormal events, where the redundant information always leads to suboptimal outcomes. To address these challenges, exploiting the representation and generalization capabilities of Vison Language Models (VLMs) and Large Language Models (LLMs), we propose VA-GPT, a novel MLLM designed for summarizing and localizing abnormal events in various videos. Our approach efficiently aligns effective tokens between visual encoders and LLMs through two key proposed modules: Spatial Effective Token Selection (SETS) and Temporal Effective Token Generation (TETG). These modules enable our model to effectively capture and analyze both spatial and temporal information associated with abnormal events, resulting in more accurate responses and interactions. Furthermore, we construct an instruction-following dataset specifically for fine-tuning video-anomaly-aware MLLMs, and introduce a cross-domain evaluation benchmark based on XD-Violence dataset. Our proposed method outperforms existing state-of-the-art methods on various benchmarks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 653 + ], + "type": "text", + "content": "Detecting and summarizing abnormal events in videos is critical and challenging, and it has garnered considerable attention across multiple research domains and real-world applications, such as security monitoring, video analysis, and crime detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Although many traditional methods [4, 16, 37, 43, 47, 73, 86] have been widely explored for video anomaly detection, they exhibit substantial limitations in their effectiveness [22, 51, 60, 66, 71, 84]. These limitations manifest in two aspects: 1) Traditional video anomaly detection" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 243, + 550, + 401 + ], + "blocks": [ + { + "bbox": [ + 318, + 243, + 550, + 401 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 550, + 401 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 550, + 401 + ], + "type": "image", + "image_path": "aa9e6ed126b55907fe3f366578df2edc7f5d3609e862de6448b78628b49de336.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 414, + 555, + 524 + ], + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 524 + ], + "type": "text", + "content": "Figure 1. Baseline video understanding MLLM feeds forward every visual token (yellow squares) equally to participate in fine-tuning and inference (top row). Different from it, our method focuses on the effective area (unobstructed area in medium video frames) in each frame and select the Spatial Effective Tokens (orange squares) for the LLM (see Section 3.2) (filtered tokens are shown as gray squares). At the same time, we generate anomaly-aware Temporal Effective Tokens (green squares) (see Section 3.3) based on the assigned anomaly scores (denoted as " + }, + { + "bbox": [ + 313, + 414, + 555, + 524 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 414, + 555, + 524 + ], + "type": "text", + "content": ") of each frame from a pretrained classifier for better temporal localization of anomalies." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 617 + ], + "type": "text", + "content": "methods [6, 13, 57, 63, 64, 86] essentially approach the task as a closed-set detection and classification problem, inherently limiting their ability to achieve a comprehensive understanding and interpretation of anomalies; 2) These methods [2, 19, 23, 65, 68, 78] are restricted by a limited vocabulary, making it difficult for them to handle unseen or novel situations effectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Recent advancements [1, 28, 34, 36, 45, 58] in Vision Language Models (VLMs) and Large Language Models (LLMs) have demonstrated remarkable capabilities in scene understanding and comprehensive analysis. Multimodal Large Language Models (MLLMs), particularly those designed for video understanding [27, 29, 31, 39, 42, 74], have achieved significant progress in general video analysis tasks. However, while these models exhibit strong performance" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22695" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "in general video understanding, they fall short in accurately detecting and interpreting anomalies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 297, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 297, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 297, + 315 + ], + "type": "text", + "content": "For mitigating the above challenges, some works [12, 55, 67, 75, 79] proposed anomaly-aware video MLLMs to better understand the anomalies in videos. Although these models work well for detecting obvious abnormal events, such as fighting or fire, they typically struggle to effectively align abnormal regions with relevant captions which requires addressing spatial redundancy, and accurately identifying abnormal time intervals by mitigating temporal redundancy. This is because these methods treat all latent tokens with equal priority across spatial and temporal dimensions. This leads to performance degradation caused by redundant tokens unrelated to anomalies. However, in most cases, only small regions within a few frames contain the essential information help to identify an anomaly (as shown in Figure 1). Thus, we explore: How can multimodal architectures evolve selective token generation and processing mechanisms to dynamically prioritize anomaly-salient information while preserving comprehensive scene understanding capabilities?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 316, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 591 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose a new model named VA-GPT for analyzing various Videos for Abnormal events by aligning effective and accurate tokens with LLMs across both spatial and temporal dimensions. VA-GPT integrates two key components to identify effective visual tokens for alignment while eliminating redundant tokens that could hinder anomaly analysis and distract model from extracting useful information: 1) we develop the Spatial Effective Token Selection (SETS) module for identifying tokens corresponding to regions with challenges for aligning them with LLMs, while filtering out tokens associated with minor dynamics to remove redundancy. This is because we find that abnormal events often result in different visual changes and variations in local areas (see Figure 1); and 2) we propose the Temporal Effective Token Generation (TETG) module which employs a lightweight pre-trained classifier to assign a confidence score to each frame indicating the possibility of containing abnormal events. Then TETG generates efficient tokens with prior knowledge of the temporal information of abnormal events directly in the language space as additional input to the LLMs, effectively enhancing the model's temporal reasoning and understanding abilities about abnormal events." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Furthermore, beyond conventional benchmarks (domain benchmark), we establish a new cross-domain evaluation protocol that systematically evaluates model robustness with domain shifts. Based on a novel video dataset, XD-Violence [64], we design comprehensive QAs about abnormal events which include different visual contents from our training data and integrate it as a new cross-domain benchmark. Meanwhile, we design temporal-information-oriented QAs on both in- and cross- domain benchmarks for evaluating temporal localization abilities. Comprehensive" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "experiments demonstrate VA-GPT's superiority, achieving state-of-the-art performance in both in-domain anomaly localization and cross-domain generalization scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 109, + 534, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 534, + 119 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 534, + 119 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 123, + 555, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 123, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 123, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 314, + 123, + 555, + 171 + ], + "type": "text", + "content": "- We propose VA-GPT, a video-anomaly-aware MLLM for detecting and summarizing anomalies in various videos, which introduces the MLLM to the specific domain of video anomaly understanding." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 266 + ], + "type": "text", + "content": "- We introduce the SETS and TETG, which enable our MLLM to effectively capture both spatial and temporal information in video sequences, resulting in accurate understanding and localization of abnormal events. Meanwhile, we propose a new instruct-following dataset for video anomaly analysis and a comprehensive cross-domain evaluation benchmark for better evaluating the generalization abilities of MLLMs on video anomalies." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 267, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 315 + ], + "type": "text", + "content": "- Our extensive experiments demonstrate that our method outperforms existing state-of-the-art methods in various benchmarks, highlighting its effectiveness and potential for practical applications in video anomaly understanding." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 400, + 338 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 490, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 490, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 490, + 359 + ], + "type": "text", + "content": "2.1. Large Language Models (LLMs)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 556, + 532 + ], + "type": "text", + "content": "The domain of Natural Language Processing (NLP) has experienced significant progress, particularly with the emergence of Large Language Models (LLMs). The introduction of the Transformer architecture [3, 18, 59] is a critical turning point, followed by other influential language models [3, 10, 25, 80] that exhibited remarkable proficiency. Generative Pre-trained Transformers (GPT) [49] brought about a revolution in NLP by employing auto-regressive prediction, establishing itself as a powerful language modeling approach. More recent groundbreaking contributions, such as ChatGPT [45], GPT-4 [1], LaMDA [56] and LLaMA [58], have expanded the horizon even further. These models, trained on extensive textual data, display extraordinary performance in intricate linguistic tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 493, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 493, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 493, + 553 + ], + "type": "text", + "content": "2.2. Vision Language Models (VLMs)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 715 + ], + "type": "text", + "content": "Progress in the fields of computer vision and natural language processing has given rise to the development of vision-language models (VLMs) [14, 21, 33, 34, 50, 61, 69]. These models combine visual and linguistic systems to facilitate cross-modal comprehension and reasoning. Notable examples include CLIP [50] which pairs BERT [10] with ViT [11]; BLIP-2 [28] incorporating Vision Transformer features into Flan-T5 [8]; MiniGPT4 [85], connecting BLIP-2 with Vicuna [28, 48]; PandaGPT [53], bridging ImageBind [15] with Vicuna. These models excel in tasks like image classification, captioning, and object detection [24, 53, 83]. Recent developments in vision-language models have extended into video processing with models" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22696" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 242 + ], + "type": "image", + "image_path": "f228863632d5c7156edebc26db7fe2beb86e0b1ee0ef22bd07d348a1bed38767.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Detailed illustration of our proposed model. When a video is fed into the model, patch embeddings " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " and class embeddings (c.ebd) " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " are extracted from all frames. 1) Based on the difference in patch embeddings between current frame and its neighbour frame, we can get a filter mask " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " to filter out unimportant visual tokens (dashed square " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": ") from current frame's visual tokens " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": ", thereby selecting Spatial Effective Tokens " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " that are compressed with a projector with pooling into aligned content token " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " for each frame, meanwhile take attention with text input from users for resulting aligned context token " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " for each frame; 2) Based on class embeddings (c.ebd) " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " of all frames, we use a pre-trained Anomaly-aware Classifier to localize the time period of abnormal events, thereby generating Temporal Effective Tokens " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " to feed forward into the LLM. All of the resulting aligned tokens " + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\square \\square \\square" + }, + { + "bbox": [ + 54, + 251, + 555, + 329 + ], + "type": "text", + "content": " are fed into the LLM for reasoning and inference of the whole model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 349, + 297, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 297, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 297, + 434 + ], + "type": "text", + "content": "like Video-Chat [29], Video-ChatGPT [42], Otter [26], Valley [39], mPLUG [27], Video-LLaMA [74], and LLaMAVID [31]. These systems enable interactive video querying, enhance comprehension through audio-visual-text alignment, and support comprehensive video analysis. In this paper, we leverage VLMs and LLMs to develop a novel approach for video anomaly understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 444, + 255, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 255, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 255, + 456 + ], + "type": "text", + "content": "2.3. Video Anomaly Understanding (VAU)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 462, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 297, + 714 + ], + "type": "text", + "content": "Annotating surveillance video frames is labor-intensive, prompting researchers to explore alternatives: one-class learning [72], unsupervised anomaly detection without annotations [4, 5, 16, 20, 37, 38, 47], or weakly supervised methods using only video-level annotations [6, 13, 30, 57, 60, 62, 63, 71, 76]. In one-class learning, Luo et al. developed a ConvLSTM network for normal segment learning [40]. Several researchers employed Auto-Encoder networks to reconstruct normal frame features [5, 20, 81], while others implemented memory mechanisms [4, 16, 37, 47] or meta-learning [38] to enhance generalizability. For weakly supervised learning, Tian [57] sed multiple-instance learning to localize anomalous clips. Zhong et al. uutilized graph convolution networks, though with limited generalization capability. [82]. To address this, Ramachandra et al. developed a Siamese network for normal feature learning. Wan et al. and Zaheer et al. [60, 71]proposed clustering-based frameworks for anomaly identification. Recent studies have introduced new architectures for spatial-temporal feature ensemble learning [6, 13, 30, 57, 62, 63, 76]. However, these methods merely supply anomaly scores during inference," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 350, + 556, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 399 + ], + "type": "text", + "content": "necessitating empirical threshold establishment on test sets to differentiate abnormal events. Recent research has begun exploring MLLMs to enhance models' capabilities in identifying and describing anomalies [12, 55, 67, 75, 79]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 411, + 370, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 370, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 370, + 423 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 382, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 382, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 382, + 442 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "type": "text", + "content": "Task. Video anomaly understanding MLLMs aims to determine whether an input video contains abnormal events, meanwhile describing and interacting with the temporal localization and the entire process of the detected abnormal events (if has). We train the model with an instruct-following dataset based on abnormal videos [54], so that the model can better align the tokens between visual encoders and MLLMs for presenting and generalizing information about abnormal events." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": "Pipeline. Considering the video understanding MLLM framework as shown in Figure 2, taking a video (contains " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " frames) as input, a frozen ViT-based [11] visual encoder (CLIP [52]) extracts visual tokens " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " from each video frame " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t (t = 1,\\dots,T)" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t = \\{\\mathbf{x}_i^t\\}_{i = 1,\\dots ,N}" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " of the current frame, there are " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " visual tokens corresponding to equal amounts of image patches. Modality alignment converts the processed visual tokens " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " into the semantic space of LLMs. At the same time, text prompts are processed and encoded as text tokens into the same semantic space and serve as a part of input to LLMs. Our key design on models consists of (1) selecting Spatial Effective Tokens " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{*t}" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " (SET) from " + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 558, + 557, + 714 + ], + "type": "text", + "content": " for each frame participating in fine-tuning and inference" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22697" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "instead of " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " (see Section 3.2); and (2) generating Temporal Effective Tokens " + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^{*t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": " (TET) as anomaly-aware temporal priors, participating in inference to facilitate the temporal localization of abnormal events for LLMs (see Section 3.3). In addition, we produce high-quality instruct-following data on abnormal videos and develop a training strategy for it to maximize the effectiveness of our proposed method (see Section 3.4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 174, + 268, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 268, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 268, + 186 + ], + "type": "text", + "content": "3.2. Spatial Effective Token Selection (SETS)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 191, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 334 + ], + "type": "text", + "content": "In classical video classification tasks, context and relationships are critical. However, in our MLLM setting, beyond leveraging contextual information, the most crucial problem is aligning the visual and language modalities. Therefore, the key aspect of our design is to extract useful information for effectively aligning visual tokens with the LLM. Since text captions primarily describe anomaly events, which occupy only a small portion of the entire video, aligning all visual patterns with text tokens would be unreasonable and computationally heavy. Thus, we are the first to propose a novel token selection method SETS to achieve efficient and effective alignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": "Inter-frame Difference. For a video, we believe that areas with large changes in adjacent frames are more worthy of attention. As illustrated in Figure 2, for each frame " + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^t" + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": " of the video, we can regard its previous frame " + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{t-1}" + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": " as the reference frame for investigating the difference between current timestamp and previous timestamp. Employing DINOV2 [46] as the feature extractor, denoted as " + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 335, + 296, + 430 + ], + "type": "text", + "content": ", we can extract patch embeddings:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 436, + 295, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 436, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 111, + 436, + 295, + 451 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1} = \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t}), \\mathcal {F} _ {\\mathrm {E}} (\\mathbf {V} ^ {t - 1}), \\tag {1}", + "image_path": "beeb194a56b141635a8d57fff4c9a696c2f0baa4c16e3a425c8f4397d46bfe87.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^t, \\mathbf{F}^{t-1} \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "text", + "content": " are the extracted embeddings (" + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "text", + "content": " indicates the number of image patches and " + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 456, + 296, + 517 + ], + "type": "text", + "content": " indicates the channels). Thanks to the distinction and stability of the extracted features, we calculate their patch-wise distances as the inter-frame difference map of the current frame:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 522, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 522, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 132, + 522, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {D} ^ {t} = \\operatorname {d i s} \\left(\\mathbf {F} ^ {t}, \\mathbf {F} ^ {t - 1}\\right), \\tag {2}", + "image_path": "f0e55b669049bab5c96c3a10df0b4b017a193f49506b126b4179af3f1359c7ee.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "inline_equation", + "content": "dis(\\cdot)" + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "text", + "content": " indicates Manhattan distance [17] and " + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^t\\in" + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 542, + 295, + 578 + ], + "type": "text", + "content": " indicates the distances between corresponding patch pairs in neighbour frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "text", + "content": "Select Spatial Effective Tokens. According to the interframe difference map " + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^t" + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "text", + "content": ", we can set up a vector " + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^t = [m_1^t, m_2^t, \\dots, m_N^t]" + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "text", + "content": " to record the difference of each patch, where top " + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 578, + 296, + 661 + ], + "type": "text", + "content": " ratio of elements with the largest distance are assigned with the value of 1, and the rest are assigned as 0. Thus we get a mask for filtering and updating the visual tokens as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 668, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 668, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 107, + 668, + 295, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {* t} = \\left\\{\\mathbf {x} _ {i} ^ {t} \\mid m _ {i} ^ {t} = 1, m _ {i} ^ {t} \\in \\mathbf {M} ^ {t} \\right\\}, \\tag {3}", + "image_path": "1d7b7a0f9f3c10dfe4926bcb6c8387c9e96c4b376e35b63b0ca5572d8536545d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{*t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " contains the selected Spatial Effective Tokens (SET) which are fed into subsequent processing instead of" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^t" + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": " as shown in Figure 2. SET can efficiently isolate the regions highly related to the abnormal events to participate in both fine-tuning and inference." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 114, + 554, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 554, + 128 + ], + "type": "text", + "content": "3.3. Temporal Effective Token Generation (TETG)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": "Anomaly-aware Classifier. We design a simple but effective MLP " + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": " for learning whether each frame is related to an abnormal event. For the class embeddings (denoted as " + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": ") extracted from the feature encoder, we can split them based on training video caption into normal and anomaly embeddings, denoted as " + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": ", respectively. Thus we can optimize " + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{A}}" + }, + { + "bbox": [ + 313, + 132, + 555, + 216 + ], + "type": "text", + "content": " using a binary classification loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 222, + 555, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 222, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 321, + 222, + 555, + 242 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {n}}} \\left[ - \\log \\frac {1}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right] + \\mathbb {E} _ {\\mathbf {z} \\sim \\mathbf {Z} ^ {\\mathrm {a}}} \\left[ - \\log \\frac {\\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}}{1 + \\exp^ {- \\mathcal {F} _ {\\mathrm {A}} (\\mathbf {z})}} \\right]. \\tag {4}", + "image_path": "295bcbf45fd343780492948d03b867f2e67e0b35c49f5910735c5eb0b7d65093.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 247, + 554, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 293 + ], + "type": "text", + "content": "The anomaly-aware classifier predicts whether each frame is related to anomalies in a video, which can bring additional important prior knowledge to LLMs at a very low cost to facilitate its inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "content": "Generate Temporal Effective Tokens. Since the information drawn from the anomaly-aware classifier is explicit, we can easily project it to LLMs' text token space through natural languages. Based on the prediction results of the anomaly-aware classifier, we select the first and last frames' timestamps with high confidence of containing abnormal events, denoted as " + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "content": ", respectively. Then we tokenize them with a template as: \"Known common crime types are: 'Shooting', 'Arson', 'Arrest', ... There is one of the crime types occurring from " + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "inline_equation", + "content": "" + }, + { + "bbox": [ + 313, + 294, + 556, + 511 + ], + "type": "text", + "content": "\", resulting Temporal Effective Tokens (TET) in the text token space of LLMs. During inference, with the well-trained lightweight anomaly-aware classifier, TET is used as an additional input to participate in the forward process of the LLM to provide prior knowledge about abnormal events temporally (as shown in Figure 2)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 517, + 420, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 420, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 420, + 529 + ], + "type": "text", + "content": "3.4. Training Strategy" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 605 + ], + "type": "text", + "content": "For modality alignment and instruction tuning, we follow the baseline [31] to ensure visual features are well aligned with the language space. In this work, the training strategy can be divided into two stages: 1) Stage One: Fine-tuning with anomaly video data, and 2) Stage Two: Aligning Spatial Effective Tokens with LLMs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Fine-tuning with Video Anomaly Data. For enhancing the abnormal scene understanding of LLMs, we construct the UCF-crime [70] Question-Answer pairs for fine-tuning. We also mix different instruction pairs from various sources [32], including text conversations, single/multi-turn visual Question-Answer pairs, and video Question-Answer pairs. Different formats for text, image, and video inputs are adopted, and the image token is randomly inserted at the beginning or end of user input during training. All modules," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22698" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 550, + 218 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 550, + 218 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 550, + 218 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 550, + 218 + ], + "type": "table", + "html": "
MethodLLMIn-domainCross-domain
Total Acc.(%)Temporal Acc.(%)Total Acc.(%)Temporal Acc.(%)
Video-Chat [29]Vicuna-7B22.41---
Video-ChatGPT [42]Vicuna-7B24.1328.5124.0029.10
Otter [26]LLaMa-7B22.4122.1725.2023.80
Valley [39]Vicuna-7B20.3414.4821.0020.20
mPLUG [27]LLaMa-7B22.76---
Video-LLaMA2 [7]Vicuna-7B21.3826.6224.2023.00
Hawkeye [79]LLaVA-7B28.6030.0025.3028.50
LLaMA-VID [31]Vicuna-7B14.8326.7018.8023.60
VA-GPT (Ours)Vicuna-7B30.6935.0026.2031.02
", + "image_path": "63dbad57a8139d18d76768fc40e64105b4d5de973aebc253480a5f905b3b2ac8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 274, + 293, + 318 + ], + "blocks": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 257 + ], + "type": "text", + "content": "Table 1. Comparing on in-domain (UCF-Crime [54]) and the proposed cross-domain (XD-Violence [64]) benchmarks, our method significantly outperforms other models and achieve the state-of-the-art performance (accuracy is detonated as Acc.) on anomaly video understanding and temporal localization with LLMs (Best results are shown in bold)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 274, + 293, + 318 + ], + "lines": [ + { + "bbox": [ + 57, + 274, + 293, + 318 + ], + "spans": [ + { + "bbox": [ + 57, + 274, + 293, + 318 + ], + "type": "table", + "html": "
MethodDescriptionCausesEffect
A-Guardian [12]79.6558.9250.64
Ours80.8359.5551.08
", + "image_path": "d16fbeb61164e4c66f1dc5b2c7acd2a449c9c6cb47ba56af06a4f246a2ae9dbd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 320, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 320, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 295, + 354 + ], + "type": "text", + "content": "Table 2. We evaluate our method on the MMEval [12] benchmark and show that our method outperforms the related previous method in different aspects." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 360, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 295, + 418 + ], + "type": "text", + "content": "except the frozen visual encoder, are optimized in this stage. After fine-tuning, the LLM has a preferred perception of anomalies, thus ensuring the effectiveness of the temporal effective tokens (see Section 3.3) during inference. More dataset details are illustrated in Section 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 419, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 295, + 551 + ], + "type": "text", + "content": "Aligning Spatial Effective Tokens with LLMs. For abnormal video scenes, most areas would not be aligned with languages. Therefore, we implement an additional fine-tuning step. This step involves utilizing Spatial Effective Tokens (see Section 3.2) derived from each video frame within the UCF-Crime dataset. By incorporating these tokens, we aim to provide the model with a more refined understanding of the spatial context of anomalies. It also brings efficient optimization, and the alignment here is only designed for very short-term fine-tuning, which can greatly improve the model's ability to detect and understand anomalies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "Datasets. We fine-tune our model on our proposed instruct-following format [34] training dataset including 4077 videos and 7730 images based on UCF-Crime [54] dataset. We evaluate the models on two anomaly video understanding benchmarks: UCF-Crime [54] for in-domain evaluation and a proposed benchmark designed based on XD-Violence [64] dataset for cross-domain evaluation, respectively. More details are shown in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Benchmarks and Metrics. To evaluate the ability to review videos and identify anomalies, we utilize a video anomaly understanding evaluation from Video-Bench [44] to assess the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 555, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 444 + ], + "type": "text", + "content": "temporal comprehensive ability, which contains nature language based Question-Answer pairs from UCF-Crime [54] dataset. Meanwhile, in order to evaluate the model's cross-domain video anomaly understanding ability, we contribute Question-Answer pairs as an extra benchmark based on the XD-Violence dataset. These Question-Answer pairs encompass four options, with each presenting the anomaly category and the respective time intervals during which the anomalies transpire. For each benchmark, we design different sets of questions for two evaluations: one is an overall evaluation of abnormal event detection and understanding, and the other is a special evaluation focusing on temporal localization ability, measured by question answering accuracy (denoted as Total Acc. and Temporal Acc., respectively, higher is better)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 564 + ], + "type": "text", + "content": "Implementation Details. For network structure, we incorporate the pre-trained CLIP [52] and DINOv2 [46] as the visual encoder and Qformer [9] as the text decoder. We follow [31] to freeze the encoder during the modality alignment and to optimize the trainable parameters using anomaly videos and instructions for instruction tuning. During the training process, our model utilizes PyTorch on four NVIDIA A100 GPUs. We employ the AdamW optimizer with a batch size of 64. The learning schedule is set to cosine decay, with a learning rate of 2e-5 and a total of 1 epoch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 575, + 399, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 399, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 399, + 586 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Results on In-domain Dataset. We first evaluate our method on the in-domain dataset, where the test set belongs to the same style and recording mode as the data used for training in Section 3.4. As shown in Table 1, compared with the baseline [31], with fewer visual embedding tokens and temporal effective tokens, our method brings more than double the performance improvement on total accuracy, also brings a significant increase in temporal localization. Driven by our proposed training strategy and designed effective tokens, more pure and effective visual-semantic information" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22699" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 89, + 211, + 125 + ], + "blocks": [ + { + "bbox": [ + 64, + 74, + 206, + 83 + ], + "lines": [ + { + "bbox": [ + 64, + 74, + 206, + 83 + ], + "spans": [ + { + "bbox": [ + 64, + 74, + 206, + 83 + ], + "type": "text", + "content": "A 11-seconds video about Road Accident" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 89, + 211, + 125 + ], + "lines": [ + { + "bbox": [ + 59, + 89, + 211, + 125 + ], + "spans": [ + { + "bbox": [ + 59, + 89, + 211, + 125 + ], + "type": "image", + "image_path": "c28d2ffd063a24261cf9b218ac80213ecd9ec39446deca8bbf49b74f1aff9a22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 128, + 78, + 142 + ], + "blocks": [ + { + "bbox": [ + 60, + 128, + 78, + 142 + ], + "lines": [ + { + "bbox": [ + 60, + 128, + 78, + 142 + ], + "spans": [ + { + "bbox": [ + 60, + 128, + 78, + 142 + ], + "type": "image", + "image_path": "1001798447d95625b5f868e07fe148bd0eb3bc8938ed35a3ffceb4d3bd70846f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 144, + 149, + 152 + ], + "lines": [ + { + "bbox": [ + 62, + 144, + 149, + 152 + ], + "spans": [ + { + "bbox": [ + 62, + 144, + 149, + 152 + ], + "type": "text", + "content": "Describe the video in detail." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 153, + 209, + 167 + ], + "blocks": [ + { + "bbox": [ + 194, + 153, + 209, + 167 + ], + "lines": [ + { + "bbox": [ + 194, + 153, + 209, + 167 + ], + "spans": [ + { + "bbox": [ + 194, + 153, + 209, + 167 + ], + "type": "image", + "image_path": "056f08cac14508ecd68c98aba133246a4e9cbbf45a565da7b9e49c85f56adf5e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 171, + 209, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 209, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 209, + 214 + ], + "type": "text", + "content": "The video shows a group of people walking down a sidewalk and suddenly a car appears driving very fast towards them. Some people fall onto the ground as the car crashes into them." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 221, + 79, + 236 + ], + "blocks": [ + { + "bbox": [ + 60, + 221, + 79, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 221, + 79, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 221, + 79, + 236 + ], + "type": "image", + "image_path": "66c32a78cc0d4870d9f3beb41b7cd1fb6ca333003dd36e684f33d37dbe85b790.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 238, + 182, + 246 + ], + "lines": [ + { + "bbox": [ + 62, + 238, + 182, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 182, + 246 + ], + "type": "text", + "content": "Are there any anomalies in the video?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 194, + 251, + 209, + 265 + ], + "blocks": [ + { + "bbox": [ + 194, + 251, + 209, + 265 + ], + "lines": [ + { + "bbox": [ + 194, + 251, + 209, + 265 + ], + "spans": [ + { + "bbox": [ + 194, + 251, + 209, + 265 + ], + "type": "image", + "image_path": "7d3ea04034047572675ca6b14e656d1bcacbb607a06b50d88a8527a399aecb73.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 270, + 199, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 199, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 199, + 297 + ], + "type": "text", + "content": "Yes, there is a car accident in the video. It shows a car flipping over and people running away from the scene." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 230, + 88, + 299, + 125 + ], + "blocks": [ + { + "bbox": [ + 241, + 73, + 370, + 83 + ], + "lines": [ + { + "bbox": [ + 241, + 73, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 241, + 73, + 370, + 83 + ], + "type": "text", + "content": "A 1.5-minutes video about Explosion" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 230, + 88, + 299, + 125 + ], + "lines": [ + { + "bbox": [ + 230, + 88, + 299, + 125 + ], + "spans": [ + { + "bbox": [ + 230, + 88, + 299, + 125 + ], + "type": "image", + "image_path": "b6c5b72592bb6bc56ecb6bc7fe41d3e5133aab0994e1c33b530bd1896d516c22.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 301, + 87, + 380, + 125 + ], + "blocks": [ + { + "bbox": [ + 301, + 87, + 380, + 125 + ], + "lines": [ + { + "bbox": [ + 301, + 87, + 380, + 125 + ], + "spans": [ + { + "bbox": [ + 301, + 87, + 380, + 125 + ], + "type": "image", + "image_path": "970041b38aaa198a4b4e4a61b179813fd7c428ddc02e2a520ddcfecc50fe793e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 230, + 127, + 249, + 142 + ], + "blocks": [ + { + "bbox": [ + 230, + 127, + 249, + 142 + ], + "lines": [ + { + "bbox": [ + 230, + 127, + 249, + 142 + ], + "spans": [ + { + "bbox": [ + 230, + 127, + 249, + 142 + ], + "type": "image", + "image_path": "7ac0a8b3f002a356b2acab33ec6893d226481b388a5b68f893d9283b14a68ebf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 144, + 327, + 152 + ], + "lines": [ + { + "bbox": [ + 233, + 144, + 327, + 152 + ], + "spans": [ + { + "bbox": [ + 233, + 144, + 327, + 152 + ], + "type": "text", + "content": "What happened in this video?" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 153, + 380, + 167 + ], + "blocks": [ + { + "bbox": [ + 364, + 153, + 380, + 167 + ], + "lines": [ + { + "bbox": [ + 364, + 153, + 380, + 167 + ], + "spans": [ + { + "bbox": [ + 364, + 153, + 380, + 167 + ], + "type": "image", + "image_path": "18afc59029baa429109e6ee3c3a4863d4b1aa8e6550e03eef4f5628e75c96e61.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 232, + 171, + 381, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 171, + 381, + 231 + ], + "spans": [ + { + "bbox": [ + 232, + 171, + 381, + 231 + ], + "type": "text", + "content": "The video shows a large group of trees in the woods that were burnt down in a fire. The fire was so intense that it caused the trees to fall over and burn in different directions. The video also shows a large plume of smoke coming from the forest, indicating that the fire was still active." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 230, + 238, + 249, + 253 + ], + "blocks": [ + { + "bbox": [ + 230, + 238, + 249, + 253 + ], + "lines": [ + { + "bbox": [ + 230, + 238, + 249, + 253 + ], + "spans": [ + { + "bbox": [ + 230, + 238, + 249, + 253 + ], + "type": "image", + "image_path": "092f3be4f298fd70a9ae29120af13424adc5c052750c3a4fca4c6702d5122a2d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 254, + 368, + 262 + ], + "lines": [ + { + "bbox": [ + 233, + 254, + 368, + 262 + ], + "spans": [ + { + "bbox": [ + 233, + 254, + 368, + 262 + ], + "type": "text", + "content": "Describe the anomaly present in the video." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 364, + 268, + 380, + 283 + ], + "blocks": [ + { + "bbox": [ + 364, + 268, + 380, + 283 + ], + "lines": [ + { + "bbox": [ + 364, + 268, + 380, + 283 + ], + "spans": [ + { + "bbox": [ + 364, + 268, + 380, + 283 + ], + "type": "image", + "image_path": "3a87329e5ab5445c48b1104ae5214e13e63a1412d4e179df459760a44802e052.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 232, + 286, + 369, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 286, + 369, + 304 + ], + "spans": [ + { + "bbox": [ + 232, + 286, + 369, + 304 + ], + "type": "text", + "content": "The video shows a large explosion in the woods, which is an anomaly." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 399, + 89, + 467, + 124 + ], + "blocks": [ + { + "bbox": [ + 427, + 74, + 522, + 83 + ], + "lines": [ + { + "bbox": [ + 427, + 74, + 522, + 83 + ], + "spans": [ + { + "bbox": [ + 427, + 74, + 522, + 83 + ], + "type": "text", + "content": "A 32-minutes normal video" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 89, + 467, + 124 + ], + "lines": [ + { + "bbox": [ + 399, + 89, + 467, + 124 + ], + "spans": [ + { + "bbox": [ + 399, + 89, + 467, + 124 + ], + "type": "image", + "image_path": "7b06b80aaaafc591b0b549b12a8e7b126478728542db979e05c7e713d1569375.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 479, + 89, + 548, + 124 + ], + "blocks": [ + { + "bbox": [ + 479, + 89, + 548, + 124 + ], + "lines": [ + { + "bbox": [ + 479, + 89, + 548, + 124 + ], + "spans": [ + { + "bbox": [ + 479, + 89, + 548, + 124 + ], + "type": "image", + "image_path": "ae089a3d114f6671e9462e8a396ea8dfb86f3627f8df15504542da22b21da066.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 399, + 128, + 417, + 142 + ], + "blocks": [ + { + "bbox": [ + 399, + 128, + 417, + 142 + ], + "lines": [ + { + "bbox": [ + 399, + 128, + 417, + 142 + ], + "spans": [ + { + "bbox": [ + 399, + 128, + 417, + 142 + ], + "type": "image", + "image_path": "94ed00ca72b5bc96a30d678acc08ac7002cea474452340af29cc312023a60214.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 144, + 538, + 152 + ], + "lines": [ + { + "bbox": [ + 400, + 144, + 538, + 152 + ], + "spans": [ + { + "bbox": [ + 400, + 144, + 538, + 152 + ], + "type": "text", + "content": "Describe the video's main features in detail." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 533, + 154, + 548, + 168 + ], + "blocks": [ + { + "bbox": [ + 533, + 154, + 548, + 168 + ], + "lines": [ + { + "bbox": [ + 533, + 154, + 548, + 168 + ], + "spans": [ + { + "bbox": [ + 533, + 154, + 548, + 168 + ], + "type": "image", + "image_path": "0b391f11031e1277eaafab436f151593ef7d2a011e0479513bf232f84f3448ea.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 399, + 171, + 547, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 171, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 399, + 171, + 547, + 265 + ], + "type": "text", + "content": "The video features a group of workers at a bank who are interacting with customers. The workers are seen helping customers with various transactions, including deposits, withdrawals, and account inquiries. They are dressed in uniforms and are sitting behind a counter. The customers are also seen talking to the workers and using their cell phones. The video captures a typical day at a bank and the interactions between the workers and customers." + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 399, + 268, + 417, + 281 + ], + "blocks": [ + { + "bbox": [ + 399, + 268, + 417, + 281 + ], + "lines": [ + { + "bbox": [ + 399, + 268, + 417, + 281 + ], + "spans": [ + { + "bbox": [ + 399, + 268, + 417, + 281 + ], + "type": "image", + "image_path": "a08efa9fa111c440e2d52373bf93f6b8b321a8fde23740f29e34245ac33e1b50.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 283, + 507, + 291 + ], + "lines": [ + { + "bbox": [ + 400, + 283, + 507, + 291 + ], + "spans": [ + { + "bbox": [ + 400, + 283, + 507, + 291 + ], + "type": "text", + "content": "Is this video normal or abnormal?" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 532, + 279, + 548, + 294 + ], + "blocks": [ + { + "bbox": [ + 532, + 279, + 548, + 294 + ], + "lines": [ + { + "bbox": [ + 532, + 279, + 548, + 294 + ], + "spans": [ + { + "bbox": [ + 532, + 279, + 548, + 294 + ], + "type": "image", + "image_path": "3dfcba1ea0d44f79abec27da9a772b6ba79f20a00ad9fd57b7eb5043500c8b8b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 297, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 518, + 297, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 518, + 297, + 547, + 304 + ], + "type": "text", + "content": "Normal." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 315, + 555, + 347 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 555, + 347 + ], + "type": "text", + "content": "Figure 3. Qualitative results in Question-Answer diagrams, the red circles in the figures correspond to the bold text in the answers. From short video of only a dozen seconds to medium video of longer than one minute and long video of about half an hour, our model can reason well and understand the content." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 59, + 358, + 550, + 427 + ], + "blocks": [ + { + "bbox": [ + 59, + 358, + 550, + 427 + ], + "lines": [ + { + "bbox": [ + 59, + 358, + 550, + 427 + ], + "spans": [ + { + "bbox": [ + 59, + 358, + 550, + 427 + ], + "type": "table", + "html": "
MethodBaselineStage One Fine-tuningStage Two Fine-tuning
w/o Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Bothw.SETSw.TETGw.Both
Total Acc. (%)↑14.8324.8323.7925.1225.8626.1027.5029.3128.9630.69
Temporal Acc. (%)↑26.7027.2027.7628.8129.6830.0230.7731.6033.5835.00
", + "image_path": "efaa98168085d879e57dfff2d1db13d36ead6f6b3d8648e8d894e936080c2982.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 555, + 478 + ], + "type": "text", + "content": "Table 3. Ablation studies on Spatial Effective Token Selection (SETS), Temporal Effective Tokens Generation (TETG), and progressive training strategies. At different model training stages, starting from the baseline (w/o fine-tuning, w/o both), we compare the performance of only using SETs (w.SETS), only using TETG (w.TETG), and using both (w.Both) on the UCF-Crime benchmark. Stage One: Only anomaly video fine-tuning. Stage Two: Anomaly video fine-tuning + Fine-tuning with SETs (Best results are shown in bold)." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 491, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 586 + ], + "type": "text", + "content": "of abnormal events is efficiently aligned with LLMs and exhibits powerful anomaly video understanding capabilities. At the same time, we conduct fair comparisons with existing video understanding models [26, 27, 29, 39, 42, 74] (see Table 1), and we demonstrate competitive performance. It is worth noting that we use the fewest tokens among all methods to achieve the state-of-the-art results on both total and temporal accuracies." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Results on Cross-domain Dataset. For evaluating the robustness and generalization of the models, we additionally design a cross-domain benchmark. We conduct a fair comparison of our method with the baseline [31] and the existing in-domain methods on the proposed cross-domain benchmark. The results presented in Table 1 showcase a substantial performance improvement over existing methods on the cross-domain dataset, underscoring the exceptional generalization and temporal localization capabilities of our methodology. This clear superiority in performance serves" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 491, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 515 + ], + "type": "text", + "content": "as a compelling validation of the robustness and adaptability of our approach across diverse domains." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 713 + ], + "type": "text", + "content": "Interaction with the Model. We take some interactions with our well-trained model for better evaluation. As shown in Figure 3, we demonstrate the performance of our model in addressing various video anomaly understanding challenges. To evaluate the model's effectiveness, we select videos of different durations: short (0 to 1 minute), medium (1 to 30 minutes), and long (over 30 minutes). This variety allows us to thoroughly assess the model's capabilities in handling diverse anomaly understanding scenarios. In the Road Accident video (left side in Figure 3), our method successfully identifies a car driving at high speed and detects people falling, even in low-resolution footage. For the Explosion video (middle in Figure 3), the model accurately predicts the scene and the anomaly in a medium-length video depicting an explosion. In a normal video exceeding 30 minutes (right side in Figure 3), we demonstrate the model's ability" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22700" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 295, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 295, + 175 + ], + "type": "table", + "html": "
MethodLLMFine-tunedTotal Acc.(%)↑
Video-ChatGPTVicuna-7B-24.13
✓26.23
LLaMA-VID (Baseline)Vicuna-7B-14.83
✓23.10
OursVicuna-7B-25.12
✓30.69
", + "image_path": "9f5b98fd8a4499b4e47042dca1e5ab8ef21227ee8952f4396f4f5ec147a0fb9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 295, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 245 + ], + "type": "text", + "content": "to focus on both global and local information by asking it to summarize the content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 318 + ], + "type": "text", + "content": "Comparison on Other Benchmark. We additionally compare our method on another benchmark (MMEval [12]) about anomaly video understanding with LLMs from different aspects. We follow a fair evaluation on our proposed model and obtain the quantity results as shown in Table 2, which shows the superiority of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 325, + 156, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 156, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 156, + 337 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "type": "text", + "content": "We conduct extensive ablation studies to validate the effectiveness of the key components in our method: Spatial Effective Token Selection (SETS, Section 3.2) and Temporal Effective Tokens Generation (TETG, Section 3.3) on progressive training strategies in Section 3.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": "Fine-tuning Stages. The utilization of our high-quality UCF instruct-following data has proven to enhance the model's performance. Fine-tuning with this dataset has effectively contributed to a notable accuracy compared with the baseline. As evidenced in Table 3, with our model designing (both SETS and TETG), the total accuracy for anomaly detection only achieves " + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "inline_equation", + "content": "25.12\\%" + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": " without any fine-tuning (denoted as Baseline). With anomaly video fine-tuning (denoted as Stage One Fine-tuning), the accuracy increases to " + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "inline_equation", + "content": "27.5\\%" + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": ". Furthermore, an efficient tuning with SETS (denoted as Stage Two Fine-tuning) can achieve our final performance of " + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "inline_equation", + "content": "30.69\\%" + }, + { + "bbox": [ + 55, + 403, + 296, + 558 + ], + "type": "text", + "content": " total accuracy. Temporal accuracy also shows similar increasing patterns with the scaling tuning stages." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "Effectiveness of Fine-tuning Data. For a fair comparison, we fine-tune some high-performance comparison models [31, 42] and compare them with our proposed UCF instruct-following data. As shown in Table 4, the performance of these comparison models has increased after fine-tuning, which proves the effectiveness of our proposed data. However, their performance is still not as good as our method, which proves the effectiveness of our proposed model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of SETS. Our proposed SETS demonstrates efficiency in extracting useful abnormal information, leading to performance enhancement. As illustrated in Table 3, with the SETS, the accuracy reaches " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "24.83\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "25.86\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "29.31\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " without fine-tuning, anomaly video fine-tuning and" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 554, + 119 + ], + "blocks": [ + { + "bbox": [ + 55, + 179, + 296, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 296, + 212 + ], + "type": "text", + "content": "Table 4. We fine-tune comparison models with our proposed UCF instruct-following data and evaluate the performance of these models before and after fine-tuning on the UCF benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 71, + 554, + 119 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 554, + 119 + ], + "type": "table", + "html": "
# Sample Rate K0.10.30.50.70.9
Total Acc.(%)↑23.6124.8330.6928.6727.27
Temporal Acc.(%)↑29.0329.9335.0031.2331.03
", + "image_path": "af1806027d3eddbfeb7042ece1355b1564e564d5ec9f076fa3833c0042f63b88.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 125, + 555, + 180 + ], + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 180 + ], + "type": "text", + "content": "Table 5. For the sample rate of tokens, " + }, + { + "bbox": [ + 313, + 125, + 555, + 180 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 125, + 555, + 180 + ], + "type": "text", + "content": " ratios in SETS, we sample the patch tokens with ordered distance at a fixed sample rate (0.5). The ablation indicates too large sampling rates cause too much context noise, and too small sampling rates lose visual information." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 340 + ], + "type": "text", + "content": "fine-tuning with SETS, respectively, which far exceed the accuracy of the baseline. Its intuitive mechanism for information filtering can be further analyzed with reference to Figure 4. The initial video is often cluttered with irrelevant and deceptive data. For example, in Figure 4, case one illustrates a scenario where the overall structure is complex, yet only a small segment requires attention. The SETS effectively filters out the dynamic features that do not require attention. Similarly, as in case two, the abnormal area is quite small. Our proposed SETS mechanism effectively filters out redundant and irrelevant information, significantly enhancing the model's ability to accurately pinpoint and recognize abnormal situations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": "We also conduct the ablation studies on " + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": " ratios about SETS as shown in Tabel 5. Too small or too large " + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": " ratios will cause performance degradation in both total and temporal accuracy. If " + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": " ratios is too small, redundant information will affect the effectiveness of aligning abnormal event information with corresponding captions. In contrast, if " + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 341, + 556, + 437 + ], + "type": "text", + "content": " ratios is too large, some important areas will be filtered out, resulting in information loss and suboptimal performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": "Effectiveness of TETG. Our proposed TETG generates tokens directly in the text token space of LLMs as priors for each video, offering robust priors on the temporal aspects of abnormal events. The provision results in performance improvement without the necessity for fine-tuning. As shown in Table 3, the accuracy rises from " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "14.83\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "23.79\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": ". Fine-tuning with anomaly video and SETS, the results even achieve " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "26.10\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "30.69\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": " independently, which manifests the effectiveness of TETG. Besides, the integration of SETS and TETG highlights the importance of leveraging spatial and temporal information effectively in anomaly detection systems, boosting the results to " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "25.15\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "27.50\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "inline_equation", + "content": "30.69\\%" + }, + { + "bbox": [ + 313, + 438, + 556, + 594 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 609, + 384, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 384, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 384, + 620 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Key tokens play key roles. To the best of our knowledge, we are the first to explore how to assign different learnable knowledge to different tokens for better alignment with LLMs on visual contents, thus promoting video anomaly detection and understanding (see Table 1 and Figure 3). We assign the most effective roles to different tokens in both spatial and temporal dimensions, enabling the model to handle" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22701" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 73, + 549, + 315 + ], + "blocks": [ + { + "bbox": [ + 58, + 73, + 549, + 315 + ], + "lines": [ + { + "bbox": [ + 58, + 73, + 549, + 315 + ], + "spans": [ + { + "bbox": [ + 58, + 73, + 549, + 315 + ], + "type": "image", + "image_path": "f790a42a12bb676fc8dee77ed33d810b32d38082bda8637bd4f07397f2bc34e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 555, + 346 + ], + "type": "text", + "content": "Figure 4. Visualization of the initial videos and our masked results. These two cases illustrate road accident scenarios: one occurring in a bustling street and the other in an empty suburb. Our SETS effectively filters redundant and irrelevant regions (with black patch-level masks)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 358, + 297, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 358, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 297, + 453 + ], + "type": "text", + "content": "various tokens more efficiently. The video contains abundant but redundant information. Our proposed SETS and TETG effectively compress the spatial and temporal information of abnormal events respectively, and utilize the existing alignment mechanism of MLLMs at a very low cost to participate in LLMs' reasoning (see Table 3). Our exploration inspires more representation learning of MLLMs to facilitate downstream tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 458, + 298, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 458, + 298, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 458, + 298, + 614 + ], + "type": "text", + "content": "Data matters a lot. We construct instruct-following data for anomaly videos, containing approximately 4,000 videos, which is significantly less than the amount of baseline fine-tuning video data (e.g., over 90k videos for fine-tuning in baseline model [31]). We still achieve promising performance on both in-domain and cross-domain benchmarks (see Table 1). This relies on the high-quality Question-Answer pairs in our instruct-following data. Meanwhile, SETS also improves data quality during fine-tuning: visual areas irrelevant to Question-Answer pairs are filtered out (see Figure 4), which allows for significant performance improvements in the second stage of fine-tuning (see Section 3.4) with very few steps (less than 150 iterations)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 298, + 715 + ], + "type": "text", + "content": "Broader impacts. Video anomaly understanding has far-reaching implications across various sectors, including security, healthcare, industrial safety, and so on. By enhancing the ability to automatically identify and respond to unusual or suspicious activities in real-time, LLMs can significantly improve public safety, crime prevention, patient monitoring, hazard detection, loss prevention, traffic management, and urban planning. These systems offer substantial benefits in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 359, + 485, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 485, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 485, + 371 + ], + "type": "text", + "content": "terms of operational efficiency and safety." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 372, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 503 + ], + "type": "text", + "content": "Limitations. Although our model adeptly portrays the occurrence, type, and area of video abnormal events, it still faces challenges in detecting and describing certain complex scenes. Our strategy represents an early successful validation and investigation of large models for video anomaly identification and localization. Consequently, our method possesses significant potential for enhancement in recognizing diverse abnormal video scenes. These insights motivate us to continue pursuing more powerful and efficient video anomaly understanding technologies in the future, aiming to address more challenges in the real world [35, 41, 77]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 557, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel MLLM for understanding anomalies in videos with LLMs by aligning effective tokens in both temporal and spatial space. The proposed method includes Spatial Effective Token Selection (SETS) for identifying abnormal events in small areas of large scenes and Temporal Effective Tokens Generation (TETG) for addressing the sparseness of abnormal events in video time sequences. We also develop instruct-following data of video anomaly detection to fine-tune the model. Besides, evaluation on the video anomaly understanding benchmark and a proposed cross-domain benchmark demonstrates the effectiveness of the proposed method. It further presents a promising approach for video anomaly understanding using MLLMs, showcasing the potential of effective tokens for enhancing video understanding tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22702" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 202 + ], + "type": "text", + "content": "This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No.27209621), General Research Fund Scheme (Grant No.17202422, 17212923, 17215025), Theme-based Research (Grant No.T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 224, + 115, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 224, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 115, + 236 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 243, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 243, + 296, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 289 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 289, + 296, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 296, + 322 + ], + "type": "text", + "content": "[2] Borislav Antić and Björn Ommer. Video parsing for abnormality detection. In 2011 International Conference on Computer Vision, pages 2415-2422, 2011. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 296, + 432 + ], + "type": "text", + "content": "[3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. CoRR, abs/2005.14165, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "text", + "content": "[4] Ruichu Cai, Hao Zhang, Wen Liu, Shenghua Gao, and Zhifeng Hao. Appearance-motion memory consistency network for video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):938-946, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 490, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 534 + ], + "type": "text", + "content": "[5] Yunpeng Chang, Zhigang Tu, Wei Xie, and Junsong Yuan. Clustering driven deep autoencoder for video anomaly detection. In Computer Vision - ECCV 2020, 16th European Conference, pages 329-345, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "text", + "content": "[6] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton Fok, Xiaojuan Qi, and Yik-Chung Wu. Mgfn: Magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. AAAI2023, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "text", + "content": "[7] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 637, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 296, + 713 + ], + "type": "text", + "content": "[8] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai," + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. Scaling instruction-finetuned language models, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 108, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 108, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 555, + 152 + ], + "type": "text", + "content": "[9] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 252 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 254, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 330 + ], + "type": "text", + "content": "[12] Hang Du, Sicheng Zhang, Binzhu Xie, Guoshun Nan, Jiayang Zhang, Junrui Xu, Hangyu Liu, Sicong Leng, Jiangming Liu, Hehe Fan, Dajiu Huang, Jing Feng, Linli Chen, Can Zhang, Xuhuan Li, Hao Zhang, Jianhang Chen, Qimei Cui, and Xiaofeng Tao. Uncovering what, why and how: A comprehensive benchmark for causation understanding of video anomaly, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 376 + ], + "type": "text", + "content": "[13] JiaChang Feng, FaTing Hong, and WeiShi Zheng. MIST: multiple instance self-training framework for video anomaly detection. In 2021 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 420 + ], + "type": "text", + "content": "[14] Yuting Gao, Jinfeng Liu, Zihan Xu, Jun Zhang, Ke Li, Rongrong Ji, and Chunhua Shen. Pyramidclip: Hierarchical feature alignment for vision-language model pretraining, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 555, + 465 + ], + "type": "text", + "content": "[15] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 533 + ], + "type": "text", + "content": "[16] Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In IEEE International Conference on Computer Vision (ICCV), 2019. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 535, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 566 + ], + "type": "text", + "content": "[17] Rodolfo Gonzalez and Sandra Palais. A path-building procedure for iterative circuit computers. Technical report, 1962. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 568, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 555, + 613 + ], + "type": "text", + "content": "[18] K. Han, Y. Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y. Tang, A. Xiao, C. Xu, Y. Xu, Z. Yang, Y. Zhang, and D. Tao. A survey on vision transformer. IEEE Transactions on Pattern Analysis & Machine Intelligence, pages 1-1, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 614, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 657 + ], + "type": "text", + "content": "[19] Mahmudul Hasan, Jonghyun Choi, jan Neumann, Amit K Roy-Chowdhury, and Larry Davis. Learning temporal regularity in video sequences. In Proceedings of IEEE Computer Vision and Pattern Recognition, 2016. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[20] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22703" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[21] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 150 + ], + "type": "text", + "content": "[22] Hyekang Kevin Joo, Khoa Vo, Kashu Yamazaki, and Ngan Le. Clip-tsa: Clip-assisted temporal self-attention for weakly-supervised video anomaly detection, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "text", + "content": "[23] Louis Kratz and Ko Nishino. Anomaly detection in extremely crowded scenes using spatio-temporal motion pattern models. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1446-1453, 2009. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "text", + "content": "[24] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 285 + ], + "type": "text", + "content": "[25] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 319 + ], + "type": "text", + "content": "[26] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multi-modal model with in-context instruction tuning, 2023. 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "type": "text", + "content": "[27] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, Hehong Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mplug: Effective and efficient vision-language learning by cross-modal skip-connections, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 409 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "text", + "content": "[29] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding, 2024. 1, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "text", + "content": "[30] Shuo Li, Fang Liu, and Licheng Jiao. Self-training multisecquence learning with transformer for weakly supervised video anomaly detection. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):1395-1403, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 520 + ], + "type": "text", + "content": "[31] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models, 2023. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "text", + "content": "[32] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 599 + ], + "type": "text", + "content": "[33] Bin Lin, Zhenyu Tang, Yang Ye, Jiaxi Cui, Bin Zhu, Peng Jin, Jinfa Huang, Junwu Zhang, Munan Ning, and Li Yuan. Moellava: Mixture of experts for large vision-language models, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 624 + ], + "type": "text", + "content": "[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 689 + ], + "type": "text", + "content": "[35] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[36] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[37] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE International Conference on Computer Vision, 2021. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[38] Yiwei Lu, Frank Yu, Mahesh Kumar Krishna Reddy, and Yang Wang. Few-shot scene-adaptive anomaly detection. In European Conference on Computer Vision, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 229 + ], + "type": "text", + "content": "[39] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[40] Weixin Luo, Wen Liu, and Shenghua Gao. Remembering history with convolutional LSTM for anomaly detection. In 2017 IEEE International Conference on Multimedia and Expo (ICME), pages 439-444, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[41] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 374 + ], + "type": "text", + "content": "[42] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[43] Trong-Nguyen Nguyen and Jean Meunier. Anomaly detection in video sequence with appearance-motion correspondence. CoRR, abs/1908.06351, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[44] Munan Ning, Bin Zhu, Yujia Xie, Bin Lin, Jiaxi Cui, Lu Yuan, Dongdong Chen, and Li Yuan. Video-bench: A comprehensive benchmark and toolkit for evaluating video-based large language models, 2023. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 484, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 484, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 484, + 466 + ], + "type": "text", + "content": "[45] OpenAI. Introducing chatgpt. 2022. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 468, + 555, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 565 + ], + "type": "text", + "content": "[46] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 611 + ], + "type": "text", + "content": "[47] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14372-14381, 2020. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 635 + ], + "type": "text", + "content": "[48] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 657 + ], + "type": "text", + "content": "[49] Alec Radford and Karthik Narasimhan. Improving language understanding by generative pre-training. 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22704" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "text", + "content": "[51] Bharathkumar Ramachandra, Michael J. Jones, and Ranga Raju Vatsavai. Learning a distance function with a siamese network to localize anomalies in videos. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 163 + ], + "type": "text", + "content": "[52] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, and Marco Fumero. Clip-forge: Towards zero-shot text-to-shape generation. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 196 + ], + "type": "text", + "content": "[53] Yixuan Su, Tian Lan, Huayang Li, Jialu Xu, Yan Wang, and Deng Cai. Pandagpt: One model to instruction-follow them all, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 243 + ], + "type": "text", + "content": "[54] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 244, + 296, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 296, + 288 + ], + "type": "text", + "content": "[55] Jiaqi Tang, Hao Lu, Ruizheng Wu, Xiaogang Xu, Ke Ma, Cheng Fang, Bin Guo, Jiangbo Lu, Qifeng Chen, and Ying-Cong Chen. Hawk: Learning to understand open-world video anomalies, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 289, + 296, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 296, + 488 + ], + "type": "text", + "content": "[56] Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin, James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Vincent Zhao, Yanqi Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Pranesh Srinivasan, Laichee Man, Kathleen Meier-Hellstern, Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. Lamda: Language models for dialog applications, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 489, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 554 + ], + "type": "text", + "content": "[57] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W. Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4975-4986, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 556, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 621 + ], + "type": "text", + "content": "[58] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "text", + "content": "[59] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[60] Boyang Wan, Yuming Fang, Xue Xia, and Jiajie Mei. Weakly supervised video anomaly detection via center-guided discriminative learning. In Proceedings of the IEEE International Conference on Multimedia and Expo, 2020. 1, 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 117 + ], + "type": "text", + "content": "[61] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, and Jifeng Dai. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 173 + ], + "type": "text", + "content": "[62] Jie Wu, Wei Zhang, Guanbin Li, Wenhao Wu, Xiao Tan, Yingying Li, Errui Ding, and Liang Lin. Weakly-supervised spatio-temporal anomaly detection in surveillance video. In the Thirtieth International Joint Conference on Artificial Intelligence, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 207 + ], + "type": "text", + "content": "[63] Peng Wu and Jing Liu. Learning causal temporal relation and feature discrimination for anomaly detection. IEEE Transactions on Image Processing, 30:3513-3527, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "text", + "content": "[64] Peng Wu, jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak supervision. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 263, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 308 + ], + "type": "text", + "content": "[65] Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 309, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 342 + ], + "type": "text", + "content": "[66] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 376 + ], + "type": "text", + "content": "[67] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "type": "text", + "content": "[68] Zhiwei Yang, Jing Liu, and Peng Wu. Text prompt with normality guidance for weakly supervised video anomaly detection, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 554, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 455 + ], + "type": "text", + "content": "[69] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "text", + "content": "[70] Tongtong Yuan, Xuange Zhang, Kun Liu, Bo Liu, Chen Chen, Jian Jin, and Zhenzhen Jiao. Towards surveillance video-and-language understanding: New dataset, baselines, and challenges, 2023. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 555 + ], + "type": "text", + "content": "[71] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision (ECCV), pages 358-376. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 600 + ], + "type": "text", + "content": "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Muhammad Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 656 + ], + "type": "text", + "content": "[73] Muhammad Zaigham Zaheer, Jin-Ha Lee, Marcella Astrid, Arif Mahmood, and Seung-Ik Lee. Cleaning label noise with clusters for minimally supervised anomaly detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, June 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 690 + ], + "type": "text", + "content": "[74] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding, 2023. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[75] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22705" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 642 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[76] Jiangong Zhang, Laiyun Qing, and Jun Miao. Temporal convolutional network with complementary inner bag loss for weakly supervised anomaly detection. In 2019 IEEE International Conference on Image Processing (ICIP), pages 4030-4034, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "text", + "content": "[77] Jiaqi Zhang, Yan Hu, Xiaojuan Qi, Ting Meng, Lihui Wang, Huazhu Fu, Mingming Yang, and Jiang Liu. Polar eyeball shape net for 3d posterior ocular shape representation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 180-190. Springer, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[78] Bin Zhao, Fei-Fei Li, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. In CVPR 2011, pages 3313–3320, 2011. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 306 + ], + "type": "text", + "content": "[79] Jianing Zhao, Jingjing Wang, Yujie Jin, Jiamin Luo, and Guodong Zhou. Hawkeye: Discovering and grounding implicit anomalous sentiment in recon-videos via scene-enhanced video large language model. In Proceedings of ACM MM 2024, 2024. 2, 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 373 + ], + "type": "text", + "content": "[80] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 295, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 295, + 439 + ], + "type": "text", + "content": "[81] Yiru Zhao, Bing Deng, Chen Shen, Yao Liu, Hongtao Lu, and Xian-Sheng Hua. Spatio-temporal autoencoder for video anomaly detection. In Proceedings of the 25th ACM International Conference on Multimedia, page 1933-1941, New York, NY, USA, 2017. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 441, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 295, + 496 + ], + "type": "text", + "content": "[82] Jia-Xing Zhong, Nannan Li, Weijie Kong, Shan Liu, Thomas H. Li, and Ge Li. Graph convolutional label noise cleaner: Train a plug-and-play action classifier for anomaly detection. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 498, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 540 + ], + "type": "text", + "content": "[83] Lanfeng Zhong, Xin Liao, Shaoting Zhang, Xiaofan Zhang, and Guotai Wang. Vlm-cpl: Consensus pseudo labels from vision-language models for human annotation-free pathological image classification, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 295, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 295, + 574 + ], + "type": "text", + "content": "[84] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video anomaly detection, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 295, + 609 + ], + "type": "text", + "content": "[85] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 295, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 295, + 642 + ], + "type": "text", + "content": "[86] Yi Zhu and Shawn D. Newsam. Motion-aware feature for improved video anomaly detection. In *British Machine Vision Conference (BMVC)*, 2019. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22706" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_content_list.json b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d36ea052d34901bc4423123b4f5627090de9f8ce --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_content_list.json @@ -0,0 +1,2174 @@ +[ + { + "type": "text", + "text": "Aligning Global Semantics and Local Textures in Generative Video Enhancement*", + "text_level": 1, + "bbox": [ + 91, + 130, + 913, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhikai Chen $^{1}$ , Fuchen Long $^{2,\\dagger}$ , Zhaofan Qiu $^{2}$ , Ting Yao $^{2}$ , Wengang Zhou $^{1}$ , Jiebo Luo $^{3}$ , and Tao Mei $^{2}$", + "bbox": [ + 181, + 181, + 844, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China $^{2}$ HiDream.ai Inc. $^{3}$ University of Rochester, Rochester, NY USA", + "bbox": [ + 101, + 199, + 926, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "czk654@mail.ustc.edu.cn,{longfuchen, qiuzhaofan, tiyao}@hidream.ai", + "bbox": [ + 225, + 236, + 808, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhwg@ustc.edu.cn, jluo@cs.ROchester.edu, tmei@hidream.ai", + "bbox": [ + 271, + 255, + 761, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in video generation have demonstrated the utility of powerful diffusion models. One important direction among them is to enhance the visual quality of the AI-synthesized videos for artistic creation. Nevertheless, solely relying on the knowledge embedded in the pretrained video diffusion models might limit the generalization ability of local details (e.g., texture). In this paper, we address this issue by exploring the visual cues from a high-quality (HQ) image reference to facilitate visual details generation in video enhancement. We present GenVE, a new recipe of generative video enhancement framework that pursues the semantic and texture alignment between HQ image reference and denoised video in diffusion. Technically, GenVE first leverages an image diffusion model to magnify a key frame of the input video to attain a semantics-aligned HQ image reference. Then, a video controller is integrated into 3D-UNet to capture patch-level texture of the image reference to enhance fine-grained details generation at the corresponding region of low-quality (LQ) video. Moreover, a series of conditioning augmentation strategies are implemented for effective model training and algorithm robustness. Extensive experiments conducted on the public YouHQ40 and VideoLQ, as well as self-built AIGC-Vid dataset, quantitatively and qualitatively demonstrate the efficacy of our GenVE over the state-of-the-art video enhancement approaches. Source code is available at https://github.com/HiDream-ai/GenVE.", + "bbox": [ + 89, + 335, + 483, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 763, + 220, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of video diffusion models [18, 39], there is a growing demand to elevate the visual quality of the synthesized video contents. Video enhancement, as an emerging compute vision topic distinct from traditional video super-resolution, aims to refine the low", + "bbox": [ + 89, + 789, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d20047f8fd5ed59d7b05e491665c1a6bc4afdd5a168baa24ea06a5574c265e54.jpg", + "image_caption": [ + "Figure 1. An illustration of visual alignment in our generative video enhancement (GenVE), including the global semantic alignment between LQ video and image reference, and the local texture alignment between image reference and HQ video." + ], + "image_footnote": [], + "bbox": [ + 522, + 304, + 906, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality (LQ) video to achieve better appearance and motion quality, adding more innovative details and endowing smoother temporal dynamics to the input video. Several studies [14, 39, 50, 55] have tackled this issue by treating it as a conditioning video diffusion paradigm. The prior knowledge of the input LQ video is exploited to regulate the video denoising procedure to produce the high-quality (HQ) counterpart. Despite having impressive results in the realm of video enhancement, such recipe still faces the challenge of generating fine-grained details (e.g., texture) due to the lack of some content prior knowledge.", + "bbox": [ + 511, + 613, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instead of solely relying on the input LQ video as the condition, we consider involving an additional knowledge connection, i.e., high-quality image reference, to bridge the gap of fine-grained details between LQ and HQ videos. One intuition behind this is that the visual quality of image is usually higher than that of video with the same resolution, which is caused by degradation brought by temporal dynamics (e.g., motion blur). Technically, we formulate our", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was performed at HiDream.ai.", + "bbox": [ + 107, + 875, + 359, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Co-corresponding author.", + "bbox": [ + 109, + 887, + 251, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reference-guided video enhancement in diffusion paradigm from two perspectives: global semantic alignment and local texture alignment. As depicted in the upper part of Figure 1, global semantic alignment is achieved by transferring the information of holistic layout (e.g., the relative position of cat and cloth) of a key frame to the HQ image reference via diffusion-based image up-scaling. Next, we develop a video controller in 3D-UNet to calibrate the video patch denoising with the guidance of local texture details (e.g., the cat eye and nose) in the obtained image reference. The learning of patch-level correspondences between denoised videos and image reference are executed by a cross-attention mechanism, boosting the fine-grained contextual interaction. Moreover, we investigate a set of augmentation approaches to enhance the robust learning of our reference-guided video enhancement model.", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To materialize our idea, we present GenVE to carry out semantic and texture alignment for generative video enhancement. Technically, GenVE first utilizes an image diffusion model to create a high-quality (HQ) image reference based on a key frame of input low-quality (LQ) video through forward-backward diffusion process. In each step of backward diffusion, the image reference denoising is regulated by a weighted summation of LQ key frame features to achieve global semantic alignment. Then, both of the image reference and LQ video are fed into a video controller with the newly-devised local-aware cross-attention (LCA) to attain local texture alignment between image reference and denoised video in diffusion. In LCA, the feature patch of LQ video acts as the query while that of image reference serves as the key and value for computation. A local attention bias is further exploited to impose the effects of neighboring patches around each patch in LQ video. Furthermore, GenVE involves a series of data augmentation strategies on input conditions (i.e., LQ video and image reference) to strengthen the model robustness during training.", + "bbox": [ + 91, + 332, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contribution of this paper is the proposed GenVE to explore semantic and texture alignment between image reference and denoised video in diffusion. The solution also leads to the elegant views of how to maintain the semantic correspondence in high-quality image reference synthesis, and how to leverage the texture information in the reference to benefit local detail generation in video denoising. Extensive experiments on three benchmarks verify the superiority of GenVE for video quality improvement.", + "bbox": [ + 89, + 635, + 480, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional video super-resolution. Traditional video super-resolution (VSR) aims to up-scale video frames while keeping the fidelity, which could be categorized into two groups, i.e., sliding window-based methods and recurrent methods. Sliding window-based techniques [2, 23, 24, 42, 43, 46] rely on CNNs [28] or Transformers [26, 27] which", + "bbox": [ + 89, + 809, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "aggregate the information of a sequence of low-resolution (LR) frames to predict center high-resolution (HR) frame. One representative is VRT [24] that plugs the temporal mutual attention block into transformer backbone to facilitate motion estimation, feature alignment and fusion. In contrast to aggregate information from adjacent frames, recurrent approaches [5, 6, 21, 25, 34, 47] utilize a hidden state to sequentially propagate information from all previous frames to the current frame, benefiting the frame restoration. Despite the great performance on magnifying original visual contents in the input LR video, such regression paradigm still faces the generalization problem to produce novel fine-grained details and improve the video aesthetic quality.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reference-based image super-resolution. There has been a series of investigation [35, 41, 44, 45, 48, 51, 54] to leverage the reference for image super-resolution (ISR). One typical way [48, 54] is to refine low-resolution (LR) image with the texture information in reference image via global visual alignment. For instance, Landmark [48] aligns the reference image with the LR image through optimizing a global registration problem and exploits the texture for visual enhancement. To fully mine the correspondence between reference image and input LR image, Zheng et al. [54] introduce cross-scale warping in CNNs to fuse features in different scales. Another direction for reference-based ISR is conducted in patch level. Recent advance TTSR [45] adopts a texture transformer to search active reference patches via cross-attention for image detail refinement. In this paper, we leverage the reference-based video enhancement recipe, and further delve into the context alignment between LQ and HQ videos from both semantic and texture perspectives.", + "bbox": [ + 511, + 289, + 903, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion-based video enhancement. Significant progress of video synthesis achieved by diffusion models [3, 9-13, 15, 29, 30, 32, 52, 53] encourage the development on video enhancement. Distinct from traditional video super-resolution that focuses on fidelity preserving, diffusion-based video enhancement approaches [14, 39, 50, 55] can imagine fancy details for appearance improvement. Most of them leverage the knowledge prior embedded in the pre-trained image diffusion models and formulate the problem as conditional video diffusion. For example, Zhou et al. [55] freeze the pre-trained spatial layers of the image diffusion model [33] and train the temporal latent propagation module for video enhancement. In comparison, LaVie [39] jointly trains a spatial-temporal video diffusion network and attains better temporal consistency. Later in VEnhancer [14], ControlNet [49] is further exploited to learn videow-aware conditional features to enhance the fine-grained object synthesis in video diffusion. Nevertheless, solely depending on the knowledge in pre-trained diffusion models could limit the generalization for local details (e.g., texture).", + "bbox": [ + 511, + 580, + 903, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our work focuses on video enhancement", + "bbox": [ + 531, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17088", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da8cbe43d01ae59b7ae4f6a5f63444e613c20494c9edc93d947b71f0267c63dd.jpg", + "image_caption": [ + "Figure 2. An overview of our GenVE. The input LQ video is first up-sampled to the target resolution via bicubic interpolation. The up-sampled video is compressed as the video latent $Z$ and then fed into the image diffuser to synthesize an image reference latent $y$ . Next, the latents $y$ and $Z$ are converted to the augmented image reference latent $Y_{n}$ and video latent $Z_{n}$ as the conditional signals. Concurrently, the up-sampled video latent $Z$ is noised by the Gaussian vector $\\epsilon$ to obtain the noisy video latent $Z_{T_1}$ . Hence, the video controller exploits the two augmented latents $Y_{n}$ and $Z_{n}$ , and the noisy video latent $Z_{T_1}$ to learn a visual guidance for diffusion calibration. Finally, the HQ video latent $Z_{0}$ is synthesized from $Z_{T_1}$ via video diffuser based on the learnt visual guidance." + ], + "image_footnote": [], + "bbox": [ + 93, + 78, + 619, + 263 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5d88722cfd29ec5b0a7c2b5f7821c9d036bc964add3a687589fc7bd10663d95e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 82, + 903, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "using image reference in diffusion models. The proposal of GenVE contributes by investigating not only how to obtain semantics-aligned image reference through diffusion forward-backward calibration, but also how to benefit patch-level texture details synthesis in video denoising.", + "bbox": [ + 88, + 359, + 483, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 89, + 449, + 199, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our newly-minted GenVE, pursuing global semantics alignment and local texture alignment between LQ and HQ videos for video enhancement. Figure 2 depicts an overview of this pipeline. Given a LQ video, we up-sample it through bicubic interpolation, and then encode it to the latent representation via VAE. Then, the image diffuser is exploited to refine a key frame from the up-sampled video to a semantics-aligned high-quality image reference. After that, the latents of image reference and up-sampled video are transformed by a series of conditioning augmentation strategies. The video controller performs local-aware cross-attention (LCA) between the feature patches of the augmented video frame latent and image reference latent for local texture alignment. These features are passed through cascaded 3D encoder blocks to learn the visual guidance. Finally, the video diffuser exploits the learnt visual guidance to calibrate the denoising procedure for video enhancement.", + "bbox": [ + 88, + 474, + 482, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: Video Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 757, + 423, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video diffusion models [1, 18, 19, 50] typically employ a 2D-VAE as the video encoder and a 3D-UNet as the latent denoiser. Given a video with $I$ frames, the VAE first encodes it into the latent code $Z = \\{z^i\\}_{i=1}^I$ frame-by-frame. Each feature point in the latent denotes a local patch in the original frame. To conduct video refinement using diffusion models, we follow the natural solution and implement forward and backward diffusion procedures on the input video.", + "bbox": [ + 88, + 779, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the forward process, the Gaussian noise is gradually added to the video latent code $Z$ with time step $T$ , formulated by a Markov chain. The noised latent $Z_{t}$ at each forward diffusion step $t$ is thus formulated as:", + "bbox": [ + 511, + 359, + 906, + 419 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t} = \\alpha_ {t} Z + \\sigma_ {t} \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 429, + 903, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ is the Gaussian noise and $\\alpha_{t},\\sigma_{t}$ specify the noise schedule of DDPM [17], leading the proportion of noise to gradually increase over the steps $t$ .", + "bbox": [ + 511, + 455, + 906, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the backward process, the data reconstruction for refinement eliminates the added noise in the forward diffusion process, which can start from a noise sample $Z_{t}$ at arbitrary time step $t$ . Each backward diffusion step is a denoising process formulated by:", + "bbox": [ + 511, + 501, + 905, + 577 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {t - 1} = \\left(Z _ {t} - \\tilde {\\sigma} _ {t} \\epsilon_ {t}\\right) / \\tilde {\\alpha} _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 587, + 903, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\sigma}_t$ , $\\tilde{\\alpha}_t$ are the constant hyper-parameters related to time step $t$ based on the noise scheduler [16, 17], and $\\epsilon_t = \\epsilon_\\theta(Z_t)$ is the estimated noise vector using the denoising network $\\epsilon_\\theta$ . The HQ video latent $Z_0 = \\{z_0^i\\}_{i=1}^I$ is thus iteratively denoised by the backward diffusion from the noised LQ video latent $Z_t$ to improve the video quality.", + "bbox": [ + 511, + 613, + 905, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Global Semantic Alignment", + "text_level": 1, + "bbox": [ + 511, + 712, + 764, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reference-based super-resolution [45, 51] usually exploits a semantic-related HQ image reference to guide LQ enhancement. Nevertheless, the layouts are not required to be matched between each other, leading to potential spatial misalignment of the objects. Such misalignment could limit the enhancement quality of local details. To alleviate this, we propose to generate a semantics-aligned HQ image reference for reference-based video enhancement. An image diffuser first leverages the original noisy image latent to refine denoising procedure for HQ image reference generation. The generated image reference will be further em", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17089", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ployed to transfer the patch-level texture information into LQ video frames, guiding the final video enhancement.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantics-Aligning Image Diffuser. Given the latent $z$ of a key frame from the input video latent $Z = \\{z^i\\}_{i=1}^I$ , we first add the Gaussian noise with time step $T_0$ on $z$ . Then, an image diffusion model (i.e., image diffuser in Figure 2) performs backward diffusion on the noisy frame latent $z_{T_0}$ for image refinement. At each denoising step $t$ , two intermediate latents are prepared in advance. The first is the denoised frame latent $z_t$ computed through the backward diffusion, and the other is a noisy frame latent $\\tilde{z}_t$ with time steps $t$ . We refine the denoised frame latent $z_t$ in each denoising step by leveraging the component from $\\tilde{z}_t$ through weighted summation as:", + "bbox": [ + 89, + 121, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {t} = \\alpha_ {t} z + \\sigma_ {t} \\epsilon , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 313, + 482, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = (1 - \\gamma_ {t}) z _ {t} + \\gamma_ {t} \\tilde {z} _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 333, + 364, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and $\\alpha_{t},\\sigma_{t}$ are the constant hyperparameters as mentioned in Eq.(1). Note that the parameter $\\gamma_{t}$ is defined by a cosine function on $t$ , decreasing from 0.5 to 0 monotonically with $t$ from $T$ to 0. In this way, the image diffuser introduces a substantial amount of visual content from the original frame at the early denoising phase, exactly where the latent features have rich high-level visual semantics. This denoising mechanism encourages the denoised HQ image reference to align with visual semantics of the original LQ frame, while maintaining low-level texture details as more as possible.", + "bbox": [ + 89, + 356, + 483, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Through iterative latent denoising of image refinement, the image diffuser successfully derives the image reference latent $y$ , which is semantically aligned with the key frame and has richer texture details.", + "bbox": [ + 89, + 522, + 483, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditioning Augmentation", + "text_level": 1, + "bbox": [ + 89, + 590, + 341, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we perform a series of conditioning augmentation strategies on both image reference and LQ video latent in training to simulate various scenarios, thereby boosting model robustness. The augmentation includes noise augmentation, temporal augmentation and masking condition.", + "bbox": [ + 89, + 613, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noise Augmentation. To enable controllable video refinement, a noise augmentation strategy is introduced on both video and image conditions. We implement the noise augmentation on the image reference latent $y$ and the video latent $Z = \\{z^i\\}_{i=1}^I$ , following the forward diffusion as Eq.(1) with $n$ time steps, thereby producing the noisy conditional latents $\\tilde{y}_n$ and $Z_n = \\{z_n^i\\}_{i=1}^I$ . The noise augmentation set the noise level $n$ different at each training step, corrupting the low-level visual information. Hence, the noise level can be adjusted in inference to balance the visual fidelity and quality, thus obtaining the most realistic videos.", + "bbox": [ + 89, + 689, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal Augmentation. To further improve the video temporal coherence, the temporal augmentation is applied to incorporate motion clues in the original LQ video into the", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b20d5d13e15b01f53bf1629c23a572bcdefa03763b84244101e9c93632403c3a.jpg", + "image_caption": [ + "Figure 3. Illustration of Local-Aware Cross-Attention (LCA). Given the augmented video frame latent $z_{n}^{i}$ and image reference latent $y_{n}^{i}$ , LCA first extracts the feature patches using two convolutional layers, respectively. Then, the feature patches of image reference latent are reshuffled within each neighboring $d$ patches. By treating video feature patches as query and the image reference feature patches as key/value, the attention map is computed and further added by an extra local attention bias $B$ , to estimate the output texture-aligned feature embedding $e$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 84, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HQ video diffusion process. We first estimate the optical flows using RAFT model [36] on the input LQ video $Z$ . Then, the attained flows are utilized to warp the noisy image reference latent $\\tilde{y}_n$ . As such, it yields a set of corresponding warped image references latents $Y_{n} = \\{\\tilde{y}_{n}^{i}\\}_{i = 1}^{I}$ , reinforcing the motion alignment between the warped image references and the input LQ video.", + "bbox": [ + 511, + 371, + 905, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking Condition. Moreover, we leverage the masking operation on the warped image reference latents to facilitate the model training. The masking procedure is conducted on each latent $\\tilde{y}_n^i$ of warped image references using random mask ratio $r$ , generating the masked noise image reference latent $y_{n}^{i}$ . During training, the masking condition will set a portion of the original input pixels to zeros. In inference, we maintain the whole input $\\tilde{y}_n^i$ to ensure that the entire image reference data is available for the optimal result. As such, the masking stimulates the model to refer to more potential texture feature patches in the image reference for visual enhancement.", + "bbox": [ + 511, + 477, + 905, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By the sequential conditioning augmentation, we obtain the augmented video latent $Z_{n} = \\{z_{n}^{i}\\}_{i = 1}^{I}$ and the sequence of augmented image reference latents $Y_{n} = \\{y_{n}^{i}\\}$ from $Z$ and $y$ , respectively. These two latents will be further exploited for visual guidance learning in video controller.", + "bbox": [ + 511, + 657, + 905, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Local Texture Alignment", + "text_level": 1, + "bbox": [ + 511, + 742, + 741, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To encourage the video diffuser to generate more details for video enhancement, we leverage a video controller and further design a local-aware cross-attention (LCA) module to learn texture-aligned feature embedding. The local-aware cross attention concentrates on matching texture features of similar spatial regions between image reference and input LQ video frame, facilitating more efficient and precise feature selection. The texture-aligned feature embedding produced by LCA is then utilized for visual guidance learning.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17090", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a639f7a61ee54f4c74a685ef392f87fc1e45e4845be57b2074dc9a26fdf54cbb.jpg", + "table_caption": [ + "Table 1. Performance comparisons in terms of MUSIQ, DOVER and CLIP-IQA on the YouHQ40, VideoLQ and AIGC-Vid datasets." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricsBicubicDATSR[4]LaVie-SR[39]Upscale-A-Video[55]VEnhancer[14]GenVE
YouHQ40MUSIQ ↑48.2149.2264.1763.6968.9969.93
DOVER ↑40.7630.0479.6474.0677.0883.50
CLIP-IQA ↑0.3920.4180.5810.5090.5940.596
VideoLQMUSIQ ↑59.5359.1364.5060.9066.1968.23
DOVER ↑50.9432.5666.9960.3871.4774.30
CLIP-IQA ↑0.3520.3960.4590.3800.4540.464
AIGC-VidMUSIQ ↑60.3357.4863.7160.0064.3064.56
DOVER ↑55.7357.2676.8474.1876.4677.02
CLIP-IQA ↑0.6010.5870.6560.6220.6340.658
", + "bbox": [ + 93, + 95, + 901, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the learnt visual guidance steers video denoising in video diffuser for video enhancement.", + "bbox": [ + 89, + 266, + 482, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Texture-Aligning Video Controller. The right part of Figure 2 depicts the pipeline of our video controller. The video controller leverages the LCA module for texture alignment, and the cascaded encoder blocks inherited from the 3D-UNet for feature enhancement. Specifically, the video controller takes three inputs: the augmented video latent $Z_{n}$ , the sequence of augmented image reference latents $Y_{n}$ , and the denoised video latent $Z_{t}$ that is initialized from the noised $Z_{T_1}$ with time step $T_{1}$ . The LCA module, as illustrated in Figure 3, first extracts the image feature patches from the image reference latent $y_{n}^{i}$ and video frame latent $z_{n}^{i}$ using convolutional layers, and then flatten them into patch sequences. Next, LCA executes local patch reshuffling on the patch sequences to boost model training. In detail, LCA randomly reshuffles the patch order within each neighboring $d$ patches. The image reference feature patches serve as the keys $K$ and values $V$ , and the video frame feature patches are the queries $Q$ for attention computation. Particularly, we add a local attention bias matrix $B$ on the attention weights before the Softmax operation. Thus, the attention mechanism can be formulated as:", + "bbox": [ + 91, + 297, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB = \\left( \\begin{array}{c c c} b _ {1 1} & b _ {1 2} & \\dots \\\\ b _ {2 1} & b _ {2 2} & \\dots \\\\ \\vdots & \\vdots & \\ddots \\end{array} \\right), \\quad b _ {i j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} | i - j | < d \\\\ 0, & \\text {i f} | i - j | \\geq d \\end{array} \\right., \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 625, + 482, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Q} = \\operatorname {N o r m} (Q), \\quad \\tilde {K} = \\operatorname {N o r m} (K),\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 683, + 346, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\operatorname {S o f t m a x} (\\frac {\\tilde {Q} \\tilde {K} ^ {T} + B}{\\sqrt {d _ {k}}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 704, + 282, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d$ equals to the local reshuffling range and $d_{k}$ is the channel dimension. As such, the local patch alignment mainly focuses the reference texture feature learning located within a radius of $d$ , thus facilitating precise transferring of high-quality texture patterns. After that, we feed the output attention features into another convolutional layer and add it to the query features as the texture-aligned feature embedding $e$ , which is formulated as follows:", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne = \\operatorname {C o n v} (A V) + Q. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The texture-aligned feature embedding is added to the output video features of the following encoder block, and further propagated to the subsequent cascaded blocks for multi-scale visual guidance learning.", + "bbox": [ + 511, + 266, + 903, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we collect all learnt visual guidance features output from the video controller at different scales, to regulate the multi-step backward diffusion in 3D-UNet until obtaining the final output high-quality video.", + "bbox": [ + 511, + 325, + 903, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 400, + 645, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 425, + 720, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate the effectiveness of GenVE on three datasets, i.e., the public YouHQ [55] and VideoLQ [7], as well as the self-built AIGC-Vid. The YouHQ dataset comprises 38K high-quality videos for training and 40 high-quality videos for testing (namely YouHQ40), with approximately 30 frames per video. We empirically apply the multi-tier degradation model [38] to reduce the visual quality and obtain the high-low quality video pairs for training and validation. The VideoLQ dataset [7] is a real-world collection containing 50 low-quality videos for evaluation. There are 100 frames for each video. Additionally, we build the AIGC-Vid dataset, which is generated by three text-to-video advances, i.e., ModelScope [50], LaVie [39] and VideoCrafter [8]. We exploit each model to generate 93 videos based on the text prompts from VBench [20], and finally collect 279 AI-synthesized videos for evaluation.", + "bbox": [ + 511, + 446, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We implement GenVE on the PyTorch platform. The image diffuser and the video diffuser is built on the pre-trained image diffusion model SDXL [31] and video diffusion model I2VGen-XL [50], respectively. The time steps $T$ for the base video diffusion model in training is set as 1000. The denoising time steps $T_{0}$ for image reference generation is 700, and $T_{1}$ in video enhancement is 900, respectively. In training, the parameter $n$ for noise augmentation is an integer randomly sampled within the range of [1, 300]. We empirically set it to 250 in inference. The masking ratio $r$ is randomly sampled from [0, 1) in training and set to 0 in inference. We set the local reshuffling range $d$ of LCA to 7, and define the parameter $\\gamma_{t}$ in image diffuser as $\\gamma_{t} = \\frac{1}{2}\\left[\\frac{1}{2}(1 + \\cos\\left(\\frac{T - t}{T}\\pi\\right))\\right]^{3}$ . Our", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17091", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c96baac8e4c277f25b6a2f21601466d29d9e80bc4edc422dab981b04cb28a951.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 83, + 292, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/870d126a3c940aa73b3a31c2090c8d89bbc776216e2cbcd72c0afc4346cbc648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 161, + 292, + 237 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b54112942778eb6efaa101eb2226dae3f3cac3bc509f56fc227210ec21e1057b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 120, + 238, + 292, + 313 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3def86ccbc95d30516b1ab3bb62911ea8e0d65bf165b99034763e9d6038b7f9.jpg", + "image_caption": [ + "Input Video Frame", + "Figure 4. Four visual examples of video enhancement results by different approaches on the real video data from YouHQ40 and VideoLQ. The region in the red box is presented in the zoom-in view for comparison." + ], + "image_footnote": [], + "bbox": [ + 120, + 314, + 292, + 395 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e35e3a8f7844b368abe3d492d2721ef0143c95629f387dc8510948e09c9f7a5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 84, + 413, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ddaddacb3748ab57eb2d2d250bdde4f224d78e191b7a546bcf72bc4095c16999.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 161, + 413, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/762a2677c4ec4686b261d2880af3669f4d9860c8855acee62fbb70190ab31732.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 238, + 413, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6a39d0284fd16416b36660644e25b2c483575d523dc04c03ecd768c193b575a.jpg", + "image_caption": [ + "DATSR" + ], + "image_footnote": [], + "bbox": [ + 302, + 316, + 413, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6f83e9413ec207f55a715d59cb36ee3a257f617bfe5fc6056aeec3b10e14194.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 84, + 527, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5ca676453211d34c1e0366f777dd81b04c79b0016c463113d72236b89d8e8d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 161, + 527, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b6d6ed5b45c3167d5d4aa755ea3cceda27230a3d97fd02a482c20f57d36488d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 238, + 527, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da62cc89107ec116e891321beaaa2f61a75576cfa9a8690ab6ba37c0ed9554e7.jpg", + "image_caption": [ + "LaVie-SR" + ], + "image_footnote": [], + "bbox": [ + 416, + 316, + 527, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd0745d020e0cbccd6e6193c90cc506e7355cd590418f94e35a83d450117bd57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 84, + 642, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05875dcb0cb3583975ead1a20ece100761bea6207a08c2a14e7e68b4f8695c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 161, + 642, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3fc2eaa0139bcd702c38e2dd9c2c03198b8d43e140c6ab78efb138d1098cb98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 238, + 642, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/14a278d7790969eb4ee783f714aa20185f4ef529a26f265507b9fdbb93a14183.jpg", + "image_caption": [ + "Upscale-A-Video" + ], + "image_footnote": [], + "bbox": [ + 529, + 316, + 642, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/10ea8beebca93c7eb750feefc65488c16a89be6bdd2879c4bd38ce71b6104bd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 84, + 756, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ab3675295a97807d3388168035aa04da5578cf331e29ece1720e91020e37be49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 161, + 756, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e724dbc6684edbbff920540146e70534790cd794c307038eac4c8e266226967f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 238, + 756, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b697ef6c3a0f456a9095fd88df5b40b012fcbd286ae439f9a1f2e0be0501f1d9.jpg", + "image_caption": [ + "VEnhancer" + ], + "image_footnote": [], + "bbox": [ + 645, + 316, + 758, + 392 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c17ee9b0d99b369944eab2045dc1d7a4cec7cc73be4af7b3f20cc27351268d73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 84, + 872, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89c2563c9e299a33dc3b9792a8fe6869a57cf5f20f0cbf100aa03994850d9251.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 161, + 872, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b67e91160c03b8c38b389f2fc1302439e54d602153a7ae6933da0a2d3ffe6e21.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 238, + 872, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19c320db6a07206dd0a535a8ec668af9d282d7f72b63730e999a0d6755041cdb.jpg", + "image_caption": [ + "GenVE" + ], + "image_footnote": [], + "bbox": [ + 759, + 316, + 872, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GenVE is optimized using AdamW optimizer with the base learning rate $5 \\times 10^{-5}$ and batch size 8. The key frame for the image reference generation is set as the first frame of the input video by cross validation.", + "bbox": [ + 89, + 436, + 482, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous video enhancement advances [14, 55], we conduct the evaluation on video quality using three metrics, i.e., MUSIQ [22], DOVER [40], CLIP-IQA [37]. All the three metrics estimate the perceptual scores on input video frames using the pre-trained deep models for quality assessment. In detail, MUSIQ assesses the frame-wise visual quality via a pretrained image transformer. Similarly, DOVER exploits a video transformer to estimate video quality score in terms of aesthetics and technical distortion. CLIP-IQA is determined by the feature cosine similarity between the generated frames and text prompts (e.g., good quality) via CLIP model [32] for quality perception.", + "bbox": [ + 89, + 500, + 482, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 709, + 465, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our GenVE with several state-of-the-art techniques, including the reference-based SR approach DATSR [4], and diffusion-based video enhancement approaches, i.e., LaVie-SR [39], Upscale-A-Video [55], VEnhancer [14] on YouHQ40, VideoLQ and AIGC-Vid.", + "bbox": [ + 89, + 732, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. Table 1 summarizes the performances of different approaches in terms of the three metrics over three datasets. Overall, our GenVE achieves the best performances across all metrics on both real video data and AI-synthesized video data. In general, higher MUSIQ and CLIP-IQA scores indicate the better visual percep", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tual quality on the aesthetic aspect. Specifically, GenVE achieves the MUSIQ score of 69.93 and the CLIP-IQA score of 0.596 on YouHQ40, outperforming the competitor VEnhancer. Compared to VEnhancer that solely exploits the visual guidance from LQ videos, ours further involves the semantics-aligned HQ image reference as the bridge to enhance the patch-wise texture alignment in video enhancement, attaining better aesthetic quality. Similar performance trend can be also observed on AIGC-Vid with the AI-synthesized videos. Such performances also demonstrate our model capability to handle complex video degradation scenarios that are brought by the basic video generation model in text-to-video diffusion.", + "bbox": [ + 511, + 436, + 906, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. Figure 4 showcases four video enhancement examples from real data in YouHQ40 and VideoLQ. Compared to other baselines, GenVE successfully synthesizes fine-grained details (e.g., the clear house windows in the 1-st and 2-nd case) with better visual appearance. Even for the extremely degraded video (e.g., the human face in the 4-th case), GenVE can still capture the major structural contents and imagine the reasonable local parts. The results demonstrate the advantage of exploring comprehensive information of high-quality image reference to facilitate spatial details synthesis in video diffusion model. We further visualize four video enhancement cases of the AI generated videos in Figure 5 to verify the model generalization ability. As depicted in the figure, there exists some blur (the guitar in the 3rd case) in the result videos by Upscale-A-Video. The local structures of the guitar frets in 1-st case generated by VEnhancer are also unorganized.", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17092", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e07dff65030d348a5d033fa00092f872d160550788dc4d7f431836f0784ed692.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 80, + 274, + 169 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83cabc7bc58e42c047e2245f91a42fdd7d235f129d02a33e3fef9e048f2f9810.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 170, + 274, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c02700f37fa3c7f444f0ca137d9026101d5cd236e43f46c0092a3cfe987a6ac7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 260, + 274, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e39cb670fd29f5ad6ff9b61b4c69fda9c67942d116e413763bcff152a17e0b4d.jpg", + "image_caption": [ + "Input Video Frame" + ], + "image_footnote": [], + "bbox": [ + 94, + 348, + 274, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5f84acd4790a0144ff7b123460adaa6d4f10ca93eeb77d94c5867cb5023c052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 85, + 401, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f28e93856f2e262dd626ecc2a9201712a2282dbb61ad7dec0f804c62f7f32e4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 172, + 401, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/954b48783b3887052439b66fe3af1676b63299385dedee71c11223f7e3f60952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 263, + 401, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9927edb4de5cfab93d4921aeb76c94956f28fa8886ca652e22ed6039b3bd4b8d.jpg", + "image_caption": [ + "DATSR" + ], + "image_footnote": [], + "bbox": [ + 279, + 353, + 401, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a996b60f03f6e0818eab6f484b5ca817547b8277d0dd42dfab78cd42dc78b131.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 85, + 527, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e00a16eec4e5f11b7ff534e37b46163c34d9b104deb0dfc5276a76665140bf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 174, + 527, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62feeaadb48c740a8fa021d1d4b5f46f799bfe7baf5aa0bff648db1aeba74b15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 405, + 263, + 527, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b991057b76c741a5b68fa5408a90f1117a24a259c2001e368d12908b261d6266.jpg", + "image_caption": [ + "LaVie-SR" + ], + "image_footnote": [], + "bbox": [ + 405, + 353, + 527, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a18122c8d1e1811e314b2048307fde175fd7e0dbe388d6572dd40963cca77826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 85, + 651, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efd295710595a338fa204d8e60212da84c826726f1968aca43eec85454b5599f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 174, + 651, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8226e7903abcf50b9659fc00b173eff7d4512da80cc81a24012766075c9df02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 263, + 651, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/941fac3490a8f1f47d513e4ddf3561913f9c7102d4bdb936c982984b9d05e9ce.jpg", + "image_caption": [ + "Upscale-A-Video" + ], + "image_footnote": [], + "bbox": [ + 527, + 353, + 651, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/994d530f6379eb7426c24a3e11d4c8ff2da739c6b140ebaa8895196b9ffbcab0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 85, + 777, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5f848cec41cec5579cfb130d81e1b9fff96d037e77f76291b19564460b647d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 174, + 776, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b718b62722de41569cc9d0eb77ccf08a3d9757729690dbd94c18825757bdde3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 263, + 776, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35a6f7995aff3853131987fade9ab028bac39c8d803f6492444db83fc6d5a931.jpg", + "image_caption": [ + "VEnhancer" + ], + "image_footnote": [], + "bbox": [ + 651, + 353, + 776, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab36d63970e614e57d39a278528bdc339870bd34cd2a0dcc1e30ff8298948f6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 85, + 901, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beec68e6615d8fed09c2099aa7ba3ef9d244d428cf23b7f086f045e7b608016a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 174, + 901, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05b46b287d2a9ffb6b31a0800948cd872352004d55b8af8bad14f51d467bca12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 263, + 901, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6ac3e2fbeef7506a8b8979cfc2d18cf400f2822d850afcf8e1d8ba4da71a51a.jpg", + "image_caption": [ + "GenVE" + ], + "image_footnote": [], + "bbox": [ + 777, + 353, + 901, + 433 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebd9ca4e73c249d3d53ab5971e33dd342cc2d470c3d8e92d1fab73985324589d.jpg", + "image_caption": [ + "Figure 6. Comparison of temporal profile that tracks variation of the pixels in a spatial row (highlighted by the yellow line). The width of the temporal profile equals to the video width, and the height is the number of frames of generated HQ video." + ], + "image_footnote": [], + "bbox": [ + 99, + 488, + 272, + 614 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ba37f5d4c5eab2fe23cc00f95050ddbdac176cc2fa1d33cf12b01ac8d03a57b.jpg", + "image_caption": [ + "Figure 5. Four visual examples of video enhancement results by different approaches on AI-synthesized videos from AIGC-Vid. The region in the red box is presented in the zoom-in view for comparison." + ], + "image_footnote": [], + "bbox": [ + 274, + 486, + 893, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, the result videos of our GenVE contain more fancy visual details and natural local textures, confirming the merit of texture alignment learning for high-fidelity local region restoration. Moreover, as shown in the 4-th case, our GenVE nicely distinguishes the background tiles from the toy bear, and accordingly synthesizes the smooth and fuzzy textures on their surfaces, respectively. That basically indicates the local enhancement of each video patch can be accurately calibrated by visual cues from the corresponding region in the image reference, manifesting the merit of local-aware cross-attention in GenVE.", + "bbox": [ + 88, + 654, + 482, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Besides the validation of visual appearance, we also verify the temporal consistency of generated video via the temporal profile. Figure 6 depicts the temporal variation of the pixels in a specific spatial row (i.e., the yellow line) across different approaches. As highlighted by the red box, GenVE", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exhibits a more seamless and smoother transition across the temporal dimension, while the others display blurriness or discontinuities. The result demonstrates the effectiveness of the motion-based conditioning augmentation exploited in GenVE to achieve better temporal coherence.", + "bbox": [ + 511, + 654, + 906, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Model Analysis", + "text_level": 1, + "bbox": [ + 511, + 741, + 668, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on LCA and Conditioning Augmentation. We first investigate how the LCA module and conditioning augmentation strategy influence the overall performances of video enhancement. Table 2 lists the performance comparisons among variants with different combination ways of the two components. We start from the basic video diffusion model A with a controller [49], which only extracts the visual cues from low-quality videos for video enhancement. The model B upgrades the basic model A through learning", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17093", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/daf012feb1b0389f1b72077afca3b5d4057d4770f848ce49d986bd64276dbc1c.jpg", + "table_caption": [ + "Table 2. Performance comparisons among different variants of GenVE on the YouHQ40 dataset." + ], + "table_footnote": [], + "table_body": "
ModelFusionAugmentationMUSIQ↑DOVER↑
CALCANoiseMaskTemporal
A68.3477.29
B✓69.1078.75
C✓69.4579.98
C1✓✓69.5380.16
C2✓✓✓69.8882.15
C3✓✓✓✓69.9383.50
", + "bbox": [ + 91, + 109, + 477, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "feature interaction between image reference and LQ video by vanilla cross-attention (CA) to emphasize detail generation, and leads 1.46 improvement of DOVER. By replacing vanilla CA with our deliberately designed local-aware cross-attention (LCA) module, the model C boosts DOVER from 78.75 to 79.98. Compared to general cross-attention, our LCA module imposes the neighboring textures for more accurate feature selection via involving attention bias and local patch reshuffling, thereby facilitating the synthesis of fine-grained objects. The runs of model $\\mathbf{C}_1$ , $\\mathbf{C}_2$ and $\\mathbf{C}_3$ further validate different types of conditioning augmentation in GenVE. Notably, there is a large gain of 1.99 in terms of DOVER from model $\\mathbf{C}_1$ to $\\mathbf{C}_2$ , verifying the efficacy of masking conditioning to benefit feature selection and improve model capacity in training. Finally, the model $\\mathbf{C}_3$ , i.e., our GenVE, by equipping all the three augmentations for robust model optimization, shows the best performances.", + "bbox": [ + 88, + 263, + 482, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on Noise Level. Then, we study the influence of the noise level on the image reference and input LQ video in the inference stage. We show two video cases generated by GenVE with different noise levels in Figure 7. As shown in the figure, the appearance of output HQ video is closer to that of the input LQ video when the noise level is smaller. The results are expected since that the adding noise in forward diffusion is small and the backward denoising cannot change a lot. Conversely, when exploiting high noise level in video noisings, more fancy local details (e.g., the tooth of human in the 2-nd case) are presented while some visual contents can be changed. Therefore, we achieve the tradeoff control between the video quality and fidelity by additionally involving noise augmentation in diffusion model training. In evaluation, we empirically set the noise level as 250 to seek a good balance between synthesized novel details and original visual appearance.", + "bbox": [ + 89, + 521, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on Image Reference. The image reference acts as a bridge between the input LQ video and generated HQ video in our GenVE. A valid question is how different image references influence the final results of video enhancement. Figure 8 shows two visual cases of video enhancement by using two distinct image references in GenVE. Although there are substantial visual differences between the two conditioning image references, the", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c381f913790bc4ee7f47d970a603d235ff1dbf14b47f30bdaee5a5c6c09ee165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 79, + 612, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f3576bbdfd8bc395e429a2457897fbb70f19920e6628c970dacc0259eb6fa2a.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 517, + 152, + 611, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d55e4ab0db398b4c7a08da4421e13a41dd701b2b6b141b976887bd262c85de4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 79, + 709, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad068027973554dc28e039e405106f7e1ce03455d930353607f5600e3ef738f2.jpg", + "image_caption": [ + "Noise level n=100" + ], + "image_footnote": [], + "bbox": [ + 616, + 152, + 707, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5c481be2ad410f55398ef5d17e7452770ff0402d15d206d2ea0ac00c718317d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 79, + 805, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab21a0e837258f829e6b38706a6ef79d04ab9747eac4ba7599cfe66a3260c1e5.jpg", + "image_caption": [ + "Noise level $n = 200$" + ], + "image_footnote": [], + "bbox": [ + 712, + 152, + 805, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70dcb123aefe06956aed7d5dab3559c4c9c07b198cc0c25f52e7d3974b623c55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 79, + 901, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/09174802adf8eb672b744eae204185a37889bf77b194f33dc6f711c8b83ffa27.jpg", + "image_caption": [ + "Noise level $n = 300$" + ], + "image_footnote": [], + "bbox": [ + 810, + 152, + 901, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ff236f5a9a176685bfc0668fb391526f76f364887b7b54f9506578af19a934a.jpg", + "image_caption": [ + "Figure 7. Two visual examples of video enhancement results by using different noise levels in GenVE." + ], + "image_footnote": [], + "bbox": [ + 516, + 263, + 591, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a051011b9fa37183ec748d16f6fb9f1f5f7462c6798ffe9beb58a6d4d09cc40.jpg", + "image_caption": [ + "Input", + "Figure 8. Two visual examples of video enhancement results guided by two different image references in GenVE." + ], + "image_footnote": [], + "bbox": [ + 517, + 321, + 591, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d7ce51d43d7d90cfd59231c2aac6827c4ebd1c06c98853e91f62db660f697f1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 263, + 750, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a0be5c3b7156d2ea267b2623ed8d14eced0af4782a39a12df90faa945e7bae81.jpg", + "image_caption": [ + "Reference-1", + "Result-1" + ], + "image_footnote": [], + "bbox": [ + 596, + 321, + 748, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91f5cb132ad848bb494bd60a405e7308e186c2a6c1bdefbcc85a36b2a0c239e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 263, + 828, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b600df0706e2a9ad4a1ddf0a7a6b2ba4c7db9c6b33950f6e6308d6d09523561.jpg", + "image_caption": [ + "Reference-2" + ], + "image_footnote": [], + "bbox": [ + 753, + 321, + 826, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6042350be53c431e2411ab33632c5dad2d2c802fe2486ee1b55507a5e60be8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 263, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d6cf06ef4a5520532b7b2b25cb9874cc2c293d8a94c96384d8d19f59070d2940.jpg", + "image_caption": [ + "Result-2" + ], + "image_footnote": [], + "bbox": [ + 830, + 321, + 903, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "final video enhancement results are perceptually similar. We speculate that the results are contributed by the augmentation strategy of masking condition which encourages GenVE to search multiple visual cues in image reference holistically, enhancing the generation of local details.", + "bbox": [ + 511, + 422, + 905, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 512, + 638, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented GenVE that explores the image reference to boost diffusion-based video enhancement. In particular, we study the problem from the view point of leveraging image reference as the bridge for semantic and texture alignment between input low-quality video and generated high-quality video. To consolidate our idea, GenVE first exploits an image diffusion model to upscale a key frame of input video for semantics-aligned high-quality image reference generation. A video controller then transfers the patch-level texture details in high-quality image reference to the correspondence region of low-quality video for fine-grained details enhancement. In the training of GenVE, a series of conditioning augmentation strategies is conducted to further strengthen the model robustness. Experiments conducted on two public real-world video datasets, i.e., YouHQ40 and VideoLQ, as well as a self-built AI-synthesized video dataset AIGC-Vid, validate the effectiveness of the proposed GenVE for video enhancement.", + "bbox": [ + 511, + 537, + 906, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the Beijing Municipal Science and Technology Project No. Z241100001324002, and Beijing Nova Program No. 20240484681, and the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC.", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17094", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 3", + "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime Video Super-Resolution with Spatio-Temporal Networks and Motion Compensation. In CVPR, 2017. 2", + "[3] Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer. arXiv:2505.22705, 2025. 2", + "[4] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based Image Super-Resolution with Deformable Attention Transformer. In ECCV, 2022. 5, 6", + "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond. In CVPR, 2021. 2", + "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving Video SuperResolution with Enhanced Propagation and Alignment. In CVPR, 2022. 2", + "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating Tradeoffs in Real-World Video Super-Resolution. In CVPR, 2022. 5", + "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 5", + "[9] Jingyuan Chen, Fuchen Long, Jie An, Zhaofan Qiu, Ting Yao, Jiebo Luo, and Tao Mei. Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion. In AAAI, 2025. 2", + "[10] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wenggang Zhou, Jiebo Luo, and Tao Mei. Learning 3D Shape Latent for Point Cloud Completion. IEEE Trans. on Multimedia, 2024.", + "[11] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution. In CVPR, 2024.", + "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models. In ICCV, 2021.", + "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2", + "[14] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. VEnhancer: Generative Space-Time Enhancement for Video Generation. arXiv:2407.07667, 2024. 1, 2, 5, 6", + "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS Workshop, 2021. 3", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 3", + "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv:2210.02303, 2022. 1, 3", + "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. In NeurIPS, 2022. 3", + "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive Benchmark Suite for Video Generative Models. In CVPR, 2024. 5", + "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video Super-Resolution with Recurrent Structure-Detail Network. In ECCV, 2020. 2", + "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-Scale Image Quality Transformer. In ICCV, 2021. 6", + "[23] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-Correspondence Aggregation Network for Video Super-Resolution. In ECCV, 2020. 2", + "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A Video Restoration Transformer. IEEE TIP, 2022. 2", + "[25] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent Video Restoration Transformer with Guided Deformable Attention. In NeurIPS, 2022. 2", + "[26] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022. 2", + "[27] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2", + "[28] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Bi-calibration Networks for Weakly-Supervised Video Representation Learning. IJCV, 2023. 2", + "[29] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. VideoStudio: Generating Consistent-Content and Multi-Scene Videos. In ECCV, 2024. 2", + "[30] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards Photorealistic Image Generation and Editing with Text-guided Diffusion Models. In ICML, 2022. 2", + "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 5" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17095", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 2, 6", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 2", + "[34] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking Alignment in Video Super-Resolution Transformers. In NeurIPS, 2022. 2", + "[35] Gyumin Shim, Jinsun Park, and In So Kweon. Robust Reference-based Super-Resolution with Similarity-Aware Deformable Convolution. In CVPR, 2020. 2", + "[36] Zachary Teed and Jia Deng. RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In ECCV, 2020. 4", + "[37] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring CLIP for Assessing the Look and Feel of Images. In AAAI, 2023. 6", + "[38] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind SuperResolution with Pure Synthetic Data. In ICCV, 2021. 5", + "[39] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models. IJCV, 2024. 1, 2, 5, 6", + "[40] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing wen Hou, Annan Wang, Qiong Yan Wenxiu Sun, and Weisi Lin. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. In ICCV, 2023. 6", + "[41] Yanchun Xie, Jimin Xiao, Mingjie Sun, Chao Yao, and Kaizhu Huang. Feature Representation Matters: End-to-End Learning for Reference-based Image Super-resolution. In ECCV, 2020. 2", + "[42] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal Modulation Network for Controllable Space-Time Video Super-Resolution. In CVPR, 2021. 2", + "[43] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T. Freeman. Video Enhancement with Task-Oriented Flow. IJCV, 2019. 2", + "[44] Xu Yan, Weibing Zhao, Kun Yuan, Ruimao Zhang, Zhen Li, and Shuguang Cui. Towards Content-Independent Multi-Reference Super-Resolution: Adaptive Pattern Matching and Feature Aggregation. In ECCV, 2020. 2", + "[45] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning Texture Transformer Network for Image Super-Resolution. In CVPR, 2020. 2, 3", + "[46] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations. In ICCV, 2019. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient Video Super-Resolution. In ICCV, 2021. 2", + "[48] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Landmark Image Super-resolution by Retrieving Web Images. IEEE TIP, 2013. 2", + "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 7", + "[50] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv:2311.04145, 2023. 1, 2, 3, 5", + "[51] Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. Image Super-Resolution by Neural Texture Transfer. In CVPR, 2019. 2, 3", + "[52] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models. In CVPR, 2024. 2", + "[53] Zhongwei Zhang, Fuchen Long, Zhaofan Qiu, Yingwei Pan, Wu Liu, Ting Yao, and Tao Mei. MotionPro: A Precise Motion Controller for Image-to-Video Generation. In CVPR, 2025. 2", + "[54] Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, and Lu Fang. CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping. In ECCV, 2018. 2", + "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video SuperResolution. In CVPR, 2024. 1, 2, 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 556 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17096", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_model.json b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..17d615864d4680fc0ff41e40bc47095bd8d6f892 --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_model.json @@ -0,0 +1,2860 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.915, + 0.155 + ], + "angle": 0, + "content": "Aligning Global Semantics and Local Textures in Generative Video Enhancement*" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.182, + 0.845, + 0.199 + ], + "angle": 0, + "content": "Zhikai Chen\\(^{1}\\), Fuchen Long\\(^{2,\\dagger}\\), Zhaofan Qiu\\(^{2}\\), Ting Yao\\(^{2}\\), Wengang Zhou\\(^{1}\\), Jiebo Luo\\(^{3}\\), and Tao Mei\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.2, + 0.928, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\)MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China \\(^{2}\\)HiDream.ai Inc. \\(^{3}\\)University of Rochester, Rochester, NY USA" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.237, + 0.81, + 0.252 + ], + "angle": 0, + "content": "czk654@mail.ustc.edu.cn,{longfuchen, qiuzhaofan, tiyao}@hidream.ai" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.256, + 0.762, + 0.269 + ], + "angle": 0, + "content": "zhwg@ustc.edu.cn, jluo@cs.ROchester.edu, tmei@hidream.ai" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.744 + ], + "angle": 0, + "content": "Recent advances in video generation have demonstrated the utility of powerful diffusion models. One important direction among them is to enhance the visual quality of the AI-synthesized videos for artistic creation. Nevertheless, solely relying on the knowledge embedded in the pretrained video diffusion models might limit the generalization ability of local details (e.g., texture). In this paper, we address this issue by exploring the visual cues from a high-quality (HQ) image reference to facilitate visual details generation in video enhancement. We present GenVE, a new recipe of generative video enhancement framework that pursues the semantic and texture alignment between HQ image reference and denoised video in diffusion. Technically, GenVE first leverages an image diffusion model to magnify a key frame of the input video to attain a semantics-aligned HQ image reference. Then, a video controller is integrated into 3D-UNet to capture patch-level texture of the image reference to enhance fine-grained details generation at the corresponding region of low-quality (LQ) video. Moreover, a series of conditioning augmentation strategies are implemented for effective model training and algorithm robustness. Extensive experiments conducted on the public YouHQ40 and VideoLQ, as well as self-built AIGC-Vid dataset, quantitatively and qualitatively demonstrate the efficacy of our GenVE over the state-of-the-art video enhancement approaches. Source code is available at https://github.com/HiDream-ai/GenVE." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.764, + 0.222, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.866 + ], + "angle": 0, + "content": "With the rapid development of video diffusion models [18, 39], there is a growing demand to elevate the visual quality of the synthesized video contents. Video enhancement, as an emerging compute vision topic distinct from traditional video super-resolution, aims to refine the low" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.305, + 0.907, + 0.548 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.551, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Figure 1. An illustration of visual alignment in our generative video enhancement (GenVE), including the global semantic alignment between LQ video and image reference, and the local texture alignment between image reference and HQ video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.78 + ], + "angle": 0, + "content": "quality (LQ) video to achieve better appearance and motion quality, adding more innovative details and endowing smoother temporal dynamics to the input video. Several studies [14, 39, 50, 55] have tackled this issue by treating it as a conditioning video diffusion paradigm. The prior knowledge of the input LQ video is exploited to regulate the video denoising procedure to produce the high-quality (HQ) counterpart. Despite having impressive results in the realm of video enhancement, such recipe still faces the challenge of generating fine-grained details (e.g., texture) due to the lack of some content prior knowledge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Instead of solely relying on the input LQ video as the condition, we consider involving an additional knowledge connection, i.e., high-quality image reference, to bridge the gap of fine-grained details between LQ and HQ videos. One intuition behind this is that the visual quality of image is usually higher than that of video with the same resolution, which is caused by degradation brought by temporal dynamics (e.g., motion blur). Technically, we formulate our" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.361, + 0.888 + ], + "angle": 0, + "content": "*This work was performed at HiDream.ai." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.253, + 0.9 + ], + "angle": 0, + "content": "† Co-corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.361, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17087" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "reference-guided video enhancement in diffusion paradigm from two perspectives: global semantic alignment and local texture alignment. As depicted in the upper part of Figure 1, global semantic alignment is achieved by transferring the information of holistic layout (e.g., the relative position of cat and cloth) of a key frame to the HQ image reference via diffusion-based image up-scaling. Next, we develop a video controller in 3D-UNet to calibrate the video patch denoising with the guidance of local texture details (e.g., the cat eye and nose) in the obtained image reference. The learning of patch-level correspondences between denoised videos and image reference are executed by a cross-attention mechanism, boosting the fine-grained contextual interaction. Moreover, we investigate a set of augmentation approaches to enhance the robust learning of our reference-guided video enhancement model." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.333, + 0.483, + 0.635 + ], + "angle": 0, + "content": "To materialize our idea, we present GenVE to carry out semantic and texture alignment for generative video enhancement. Technically, GenVE first utilizes an image diffusion model to create a high-quality (HQ) image reference based on a key frame of input low-quality (LQ) video through forward-backward diffusion process. In each step of backward diffusion, the image reference denoising is regulated by a weighted summation of LQ key frame features to achieve global semantic alignment. Then, both of the image reference and LQ video are fed into a video controller with the newly-devised local-aware cross-attention (LCA) to attain local texture alignment between image reference and denoised video in diffusion. In LCA, the feature patch of LQ video acts as the query while that of image reference serves as the key and value for computation. A local attention bias is further exploited to impose the effects of neighboring patches around each patch in LQ video. Furthermore, GenVE involves a series of data augmentation strategies on input conditions (i.e., LQ video and image reference) to strengthen the model robustness during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.482, + 0.772 + ], + "angle": 0, + "content": "The main contribution of this paper is the proposed GenVE to explore semantic and texture alignment between image reference and denoised video in diffusion. The solution also leads to the elegant views of how to maintain the semantic correspondence in high-quality image reference synthesis, and how to leverage the texture information in the reference to benefit local detail generation in video denoising. Extensive experiments on three benchmarks verify the superiority of GenVE for video quality improvement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Traditional video super-resolution. Traditional video super-resolution (VSR) aims to up-scale video frames while keeping the fidelity, which could be categorized into two groups, i.e., sliding window-based methods and recurrent methods. Sliding window-based techniques [2, 23, 24, 42, 43, 46] rely on CNNs [28] or Transformers [26, 27] which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "aggregate the information of a sequence of low-resolution (LR) frames to predict center high-resolution (HR) frame. One representative is VRT [24] that plugs the temporal mutual attention block into transformer backbone to facilitate motion estimation, feature alignment and fusion. In contrast to aggregate information from adjacent frames, recurrent approaches [5, 6, 21, 25, 34, 47] utilize a hidden state to sequentially propagate information from all previous frames to the current frame, benefiting the frame restoration. Despite the great performance on magnifying original visual contents in the input LR video, such regression paradigm still faces the generalization problem to produce novel fine-grained details and improve the video aesthetic quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.578 + ], + "angle": 0, + "content": "Reference-based image super-resolution. There has been a series of investigation [35, 41, 44, 45, 48, 51, 54] to leverage the reference for image super-resolution (ISR). One typical way [48, 54] is to refine low-resolution (LR) image with the texture information in reference image via global visual alignment. For instance, Landmark [48] aligns the reference image with the LR image through optimizing a global registration problem and exploits the texture for visual enhancement. To fully mine the correspondence between reference image and input LR image, Zheng et al. [54] introduce cross-scale warping in CNNs to fuse features in different scales. Another direction for reference-based ISR is conducted in patch level. Recent advance TTSR [45] adopts a texture transformer to search active reference patches via cross-attention for image detail refinement. In this paper, we leverage the reference-based video enhancement recipe, and further delve into the context alignment between LQ and HQ videos from both semantic and texture perspectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.905, + 0.883 + ], + "angle": 0, + "content": "Diffusion-based video enhancement. Significant progress of video synthesis achieved by diffusion models [3, 9-13, 15, 29, 30, 32, 52, 53] encourage the development on video enhancement. Distinct from traditional video super-resolution that focuses on fidelity preserving, diffusion-based video enhancement approaches [14, 39, 50, 55] can imagine fancy details for appearance improvement. Most of them leverage the knowledge prior embedded in the pre-trained image diffusion models and formulate the problem as conditional video diffusion. For example, Zhou et al. [55] freeze the pre-trained spatial layers of the image diffusion model [33] and train the temporal latent propagation module for video enhancement. In comparison, LaVie [39] jointly trains a spatial-temporal video diffusion network and attains better temporal consistency. Later in VEnhancer [14], ControlNet [49] is further exploited to learn videow-aware conditional features to enhance the fine-grained object synthesis in video diffusion. Nevertheless, solely depending on the knowledge in pre-trained diffusion models could limit the generalization for local details (e.g., texture)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In summary, our work focuses on video enhancement" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17088" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.079, + 0.62, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.083, + 0.905, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.268, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 2. An overview of our GenVE. The input LQ video is first up-sampled to the target resolution via bicubic interpolation. The up-sampled video is compressed as the video latent \\( Z \\) and then fed into the image diffuser to synthesize an image reference latent \\( y \\). Next, the latents \\( y \\) and \\( Z \\) are converted to the augmented image reference latent \\( Y_{n} \\) and video latent \\( Z_{n} \\) as the conditional signals. Concurrently, the up-sampled video latent \\( Z \\) is noised by the Gaussian vector \\( \\epsilon \\) to obtain the noisy video latent \\( Z_{T_1} \\). Hence, the video controller exploits the two augmented latents \\( Y_{n} \\) and \\( Z_{n} \\), and the noisy video latent \\( Z_{T_1} \\) to learn a visual guidance for diffusion calibration. Finally, the HQ video latent \\( Z_{0} \\) is synthesized from \\( Z_{T_1} \\) via video diffuser based on the learnt visual guidance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.36, + 0.484, + 0.435 + ], + "angle": 0, + "content": "using image reference in diffusion models. The proposal of GenVE contributes by investigating not only how to obtain semantics-aligned image reference through diffusion forward-backward calibration, but also how to benefit patch-level texture details synthesis in video denoising." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.2, + 0.468 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.747 + ], + "angle": 0, + "content": "In this section, we present our newly-minted GenVE, pursuing global semantics alignment and local texture alignment between LQ and HQ videos for video enhancement. Figure 2 depicts an overview of this pipeline. Given a LQ video, we up-sample it through bicubic interpolation, and then encode it to the latent representation via VAE. Then, the image diffuser is exploited to refine a key frame from the up-sampled video to a semantics-aligned high-quality image reference. After that, the latents of image reference and up-sampled video are transformed by a series of conditioning augmentation strategies. The video controller performs local-aware cross-attention (LCA) between the feature patches of the augmented video frame latent and image reference latent for local texture alignment. These features are passed through cascaded 3D encoder blocks to learn the visual guidance. Finally, the video diffuser exploits the learnt visual guidance to calibrate the denoising procedure for video enhancement." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.424, + 0.773 + ], + "angle": 0, + "content": "3.1. Preliminaries: Video Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Video diffusion models [1, 18, 19, 50] typically employ a 2D-VAE as the video encoder and a 3D-UNet as the latent denoiser. Given a video with \\(I\\) frames, the VAE first encodes it into the latent code \\(Z = \\{z^i\\}_{i=1}^I\\) frame-by-frame. Each feature point in the latent denotes a local patch in the original frame. To conduct video refinement using diffusion models, we follow the natural solution and implement forward and backward diffusion procedures on the input video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.42 + ], + "angle": 0, + "content": "In the forward process, the Gaussian noise is gradually added to the video latent code \\( Z \\) with time step \\( T \\), formulated by a Markov chain. The noised latent \\( Z_{t} \\) at each forward diffusion step \\( t \\) is thus formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.43, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\nZ _ {t} = \\alpha_ {t} Z + \\sigma_ {t} \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.502 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) is the Gaussian noise and \\(\\alpha_{t},\\sigma_{t}\\) specify the noise schedule of DDPM [17], leading the proportion of noise to gradually increase over the steps \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.578 + ], + "angle": 0, + "content": "In the backward process, the data reconstruction for refinement eliminates the added noise in the forward diffusion process, which can start from a noise sample \\( Z_{t} \\) at arbitrary time step \\( t \\). Each backward diffusion step is a denoising process formulated by:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.588, + 0.905, + 0.604 + ], + "angle": 0, + "content": "\\[\nZ _ {t - 1} = \\left(Z _ {t} - \\tilde {\\sigma} _ {t} \\epsilon_ {t}\\right) / \\tilde {\\alpha} _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.705 + ], + "angle": 0, + "content": "where \\(\\tilde{\\sigma}_t\\), \\(\\tilde{\\alpha}_t\\) are the constant hyper-parameters related to time step \\(t\\) based on the noise scheduler [16, 17], and \\(\\epsilon_t = \\epsilon_\\theta(Z_t)\\) is the estimated noise vector using the denoising network \\(\\epsilon_\\theta\\). The HQ video latent \\(Z_0 = \\{z_0^i\\}_{i=1}^I\\) is thus iteratively denoised by the backward diffusion from the noised LQ video latent \\(Z_t\\) to improve the video quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.766, + 0.729 + ], + "angle": 0, + "content": "3.2. Global Semantic Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Reference-based super-resolution [45, 51] usually exploits a semantic-related HQ image reference to guide LQ enhancement. Nevertheless, the layouts are not required to be matched between each other, leading to potential spatial misalignment of the objects. Such misalignment could limit the enhancement quality of local details. To alleviate this, we propose to generate a semantics-aligned HQ image reference for reference-based video enhancement. An image diffuser first leverages the original noisy image latent to refine denoising procedure for HQ image reference generation. The generated image reference will be further em" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17089" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "ployed to transfer the patch-level texture information into LQ video frames, guiding the final video enhancement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.304 + ], + "angle": 0, + "content": "Semantics-Aligning Image Diffuser. Given the latent \\(z\\) of a key frame from the input video latent \\(Z = \\{z^i\\}_{i=1}^I\\), we first add the Gaussian noise with time step \\(T_0\\) on \\(z\\). Then, an image diffusion model (i.e., image diffuser in Figure 2) performs backward diffusion on the noisy frame latent \\(z_{T_0}\\) for image refinement. At each denoising step \\(t\\), two intermediate latents are prepared in advance. The first is the denoised frame latent \\(z_t\\) computed through the backward diffusion, and the other is a noisy frame latent \\(\\tilde{z}_t\\) with time steps \\(t\\). We refine the denoised frame latent \\(z_t\\) in each denoising step by leveraging the component from \\(\\tilde{z}_t\\) through weighted summation as:" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.314, + 0.483, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {t} = \\alpha_ {t} z + \\sigma_ {t} \\epsilon , \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.334, + 0.366, + 0.347 + ], + "angle": 0, + "content": "\\[\nz _ {t} = (1 - \\gamma_ {t}) z _ {t} + \\gamma_ {t} \\tilde {z} _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.522 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and \\(\\alpha_{t},\\sigma_{t}\\) are the constant hyperparameters as mentioned in Eq.(1). Note that the parameter \\(\\gamma_{t}\\) is defined by a cosine function on \\(t\\), decreasing from 0.5 to 0 monotonically with \\(t\\) from \\(T\\) to 0. In this way, the image diffuser introduces a substantial amount of visual content from the original frame at the early denoising phase, exactly where the latent features have rich high-level visual semantics. This denoising mechanism encourages the denoised HQ image reference to align with visual semantics of the original LQ frame, while maintaining low-level texture details as more as possible." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Through iterative latent denoising of image refinement, the image diffuser successfully derives the image reference latent \\( y \\), which is semantically aligned with the key frame and has richer texture details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.342, + 0.608 + ], + "angle": 0, + "content": "3.3. Conditioning Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Next, we perform a series of conditioning augmentation strategies on both image reference and LQ video latent in training to simulate various scenarios, thereby boosting model robustness. The augmentation includes noise augmentation, temporal augmentation and masking condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Noise Augmentation. To enable controllable video refinement, a noise augmentation strategy is introduced on both video and image conditions. We implement the noise augmentation on the image reference latent \\( y \\) and the video latent \\( Z = \\{z^i\\}_{i=1}^I \\), following the forward diffusion as Eq.(1) with \\( n \\) time steps, thereby producing the noisy conditional latents \\( \\tilde{y}_n \\) and \\( Z_n = \\{z_n^i\\}_{i=1}^I \\). The noise augmentation set the noise level \\( n \\) different at each training step, corrupting the low-level visual information. Hence, the noise level can be adjusted in inference to balance the visual fidelity and quality, thus obtaining the most realistic videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Temporal Augmentation. To further improve the video temporal coherence, the temporal augmentation is applied to incorporate motion clues in the original LQ video into the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.085, + 0.905, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.242, + 0.907, + 0.367 + ], + "angle": 0, + "content": "Figure 3. Illustration of Local-Aware Cross-Attention (LCA). Given the augmented video frame latent \\( z_{n}^{i} \\) and image reference latent \\( y_{n}^{i} \\), LCA first extracts the feature patches using two convolutional layers, respectively. Then, the feature patches of image reference latent are reshuffled within each neighboring \\( d \\) patches. By treating video feature patches as query and the image reference feature patches as key/value, the attention map is computed and further added by an extra local attention bias \\( B \\), to estimate the output texture-aligned feature embedding \\( e \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.476 + ], + "angle": 0, + "content": "HQ video diffusion process. We first estimate the optical flows using RAFT model [36] on the input LQ video \\( Z \\). Then, the attained flows are utilized to warp the noisy image reference latent \\( \\tilde{y}_n \\). As such, it yields a set of corresponding warped image references latents \\( Y_{n} = \\{\\tilde{y}_{n}^{i}\\}_{i = 1}^{I} \\), reinforcing the motion alignment between the warped image references and the input LQ video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Masking Condition. Moreover, we leverage the masking operation on the warped image reference latents to facilitate the model training. The masking procedure is conducted on each latent \\(\\tilde{y}_n^i\\) of warped image references using random mask ratio \\(r\\), generating the masked noise image reference latent \\(y_{n}^{i}\\). During training, the masking condition will set a portion of the original input pixels to zeros. In inference, we maintain the whole input \\(\\tilde{y}_n^i\\) to ensure that the entire image reference data is available for the optimal result. As such, the masking stimulates the model to refer to more potential texture feature patches in the image reference for visual enhancement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.735 + ], + "angle": 0, + "content": "By the sequential conditioning augmentation, we obtain the augmented video latent \\( Z_{n} = \\{z_{n}^{i}\\}_{i = 1}^{I} \\) and the sequence of augmented image reference latents \\( Y_{n} = \\{y_{n}^{i}\\} \\) from \\( Z \\) and \\( y \\), respectively. These two latents will be further exploited for visual guidance learning in video controller." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.743, + 0.759 + ], + "angle": 0, + "content": "3.4. Local Texture Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To encourage the video diffuser to generate more details for video enhancement, we leverage a video controller and further design a local-aware cross-attention (LCA) module to learn texture-aligned feature embedding. The local-aware cross attention concentrates on matching texture features of similar spatial regions between image reference and input LQ video frame, facilitating more efficient and precise feature selection. The texture-aligned feature embedding produced by LCA is then utilized for visual guidance learning." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17090" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.105, + 0.079, + 0.89, + 0.093 + ], + "angle": 0, + "content": "Table 1. Performance comparisons in terms of MUSIQ, DOVER and CLIP-IQA on the YouHQ40, VideoLQ and AIGC-Vid datasets." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.097, + 0.902, + 0.259 + ], + "angle": 0, + "content": "
DatasetsMetricsBicubicDATSR[4]LaVie-SR[39]Upscale-A-Video[55]VEnhancer[14]GenVE
YouHQ40MUSIQ ↑48.2149.2264.1763.6968.9969.93
DOVER ↑40.7630.0479.6474.0677.0883.50
CLIP-IQA ↑0.3920.4180.5810.5090.5940.596
VideoLQMUSIQ ↑59.5359.1364.5060.9066.1968.23
DOVER ↑50.9432.5666.9960.3871.4774.30
CLIP-IQA ↑0.3520.3960.4590.3800.4540.464
AIGC-VidMUSIQ ↑60.3357.4863.7160.0064.3064.56
DOVER ↑55.7357.2676.8474.1876.4677.02
CLIP-IQA ↑0.6010.5870.6560.6220.6340.658
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.295 + ], + "angle": 0, + "content": "Finally, the learnt visual guidance steers video denoising in video diffuser for video enhancement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.298, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Texture-Aligning Video Controller. The right part of Figure 2 depicts the pipeline of our video controller. The video controller leverages the LCA module for texture alignment, and the cascaded encoder blocks inherited from the 3D-UNet for feature enhancement. Specifically, the video controller takes three inputs: the augmented video latent \\( Z_{n} \\), the sequence of augmented image reference latents \\( Y_{n} \\), and the denoised video latent \\( Z_{t} \\) that is initialized from the noised \\( Z_{T_1} \\) with time step \\( T_{1} \\). The LCA module, as illustrated in Figure 3, first extracts the image feature patches from the image reference latent \\( y_{n}^{i} \\) and video frame latent \\( z_{n}^{i} \\) using convolutional layers, and then flatten them into patch sequences. Next, LCA executes local patch reshuffling on the patch sequences to boost model training. In detail, LCA randomly reshuffles the patch order within each neighboring \\( d \\) patches. The image reference feature patches serve as the keys \\( K \\) and values \\( V \\), and the video frame feature patches are the queries \\( Q \\) for attention computation. Particularly, we add a local attention bias matrix \\( B \\) on the attention weights before the Softmax operation. Thus, the attention mechanism can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.626, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\nB = \\left( \\begin{array}{c c c} b _ {1 1} & b _ {1 2} & \\dots \\\\ b _ {2 1} & b _ {2 2} & \\dots \\\\ \\vdots & \\vdots & \\ddots \\end{array} \\right), \\quad b _ {i j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} | i - j | < d \\\\ 0, & \\text {i f} | i - j | \\geq d \\end{array} \\right., \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.684, + 0.348, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\tilde {Q} = \\operatorname {N o r m} (Q), \\quad \\tilde {K} = \\operatorname {N o r m} (K),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.705, + 0.283, + 0.74 + ], + "angle": 0, + "content": "\\[\nA = \\operatorname {S o f t m a x} (\\frac {\\tilde {Q} \\tilde {K} ^ {T} + B}{\\sqrt {d _ {k}}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.871 + ], + "angle": 0, + "content": "where \\(d\\) equals to the local reshuffling range and \\(d_{k}\\) is the channel dimension. As such, the local patch alignment mainly focuses the reference texture feature learning located within a radius of \\(d\\), thus facilitating precise transferring of high-quality texture patterns. After that, we feed the output attention features into another convolutional layer and add it to the query features as the texture-aligned feature embedding \\(e\\), which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\ne = \\operatorname {C o n v} (A V) + Q. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.905, + 0.327 + ], + "angle": 0, + "content": "The texture-aligned feature embedding is added to the output video features of the following encoder block, and further propagated to the subsequent cascaded blocks for multi-scale visual guidance learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.388 + ], + "angle": 0, + "content": "Finally, we collect all learnt visual guidance features output from the video controller at different scales, to regulate the multi-step backward diffusion in 3D-UNet until obtaining the final output high-quality video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.646, + 0.418 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.426, + 0.722, + 0.442 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Datasets. We evaluate the effectiveness of GenVE on three datasets, i.e., the public YouHQ [55] and VideoLQ [7], as well as the self-built AIGC-Vid. The YouHQ dataset comprises 38K high-quality videos for training and 40 high-quality videos for testing (namely YouHQ40), with approximately 30 frames per video. We empirically apply the multi-tier degradation model [38] to reduce the visual quality and obtain the high-low quality video pairs for training and validation. The VideoLQ dataset [7] is a real-world collection containing 50 low-quality videos for evaluation. There are 100 frames for each video. Additionally, we build the AIGC-Vid dataset, which is generated by three text-to-video advances, i.e., ModelScope [50], LaVie [39] and VideoCrafter [8]. We exploit each model to generate 93 videos based on the text prompts from VBench [20], and finally collect 279 AI-synthesized videos for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We implement GenVE on the PyTorch platform. The image diffuser and the video diffuser is built on the pre-trained image diffusion model SDXL [31] and video diffusion model I2VGen-XL [50], respectively. The time steps \\( T \\) for the base video diffusion model in training is set as 1000. The denoising time steps \\( T_{0} \\) for image reference generation is 700, and \\( T_{1} \\) in video enhancement is 900, respectively. In training, the parameter \\( n \\) for noise augmentation is an integer randomly sampled within the range of [1, 300]. We empirically set it to 250 in inference. The masking ratio \\( r \\) is randomly sampled from [0, 1) in training and set to 0 in inference. We set the local reshuffling range \\( d \\) of LCA to 7, and define the parameter \\( \\gamma_{t} \\) in image diffuser as \\( \\gamma_{t} = \\frac{1}{2}\\left[\\frac{1}{2}(1 + \\cos\\left(\\frac{T - t}{T}\\pi\\right))\\right]^{3} \\). Our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17091" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.084, + 0.293, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.162, + 0.293, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.239, + 0.293, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.315, + 0.293, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.396, + 0.254, + 0.405 + ], + "angle": 0, + "content": "Input Video Frame" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.085, + 0.415, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.162, + 0.415, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.239, + 0.415, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.318, + 0.415, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.395, + 0.376, + 0.404 + ], + "angle": 0, + "content": "DATSR" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.085, + 0.529, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.162, + 0.529, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.239, + 0.529, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.318, + 0.529, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.395, + 0.492, + 0.404 + ], + "angle": 0, + "content": "LaVie-SR" + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.085, + 0.643, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.162, + 0.643, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.239, + 0.643, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.318, + 0.643, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.395, + 0.629, + 0.405 + ], + "angle": 0, + "content": "Upscale-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.085, + 0.758, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.162, + 0.758, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.239, + 0.758, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.318, + 0.759, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.395, + 0.73, + 0.404 + ], + "angle": 0, + "content": "VEnhancer" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.085, + 0.873, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.162, + 0.873, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.239, + 0.873, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.318, + 0.873, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.395, + 0.834, + 0.405 + ], + "angle": 0, + "content": "GenVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.407, + 0.904, + 0.434 + ], + "angle": 0, + "content": "Figure 4. Four visual examples of video enhancement results by different approaches on the real video data from YouHQ40 and VideoLQ. The region in the red box is presented in the zoom-in view for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.497 + ], + "angle": 0, + "content": "GenVE is optimized using AdamW optimizer with the base learning rate \\(5 \\times 10^{-5}\\) and batch size 8. The key frame for the image reference generation is set as the first frame of the input video by cross validation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous video enhancement advances [14, 55], we conduct the evaluation on video quality using three metrics, i.e., MUSIQ [22], DOVER [40], CLIP-IQA [37]. All the three metrics estimate the perceptual scores on input video frames using the pre-trained deep models for quality assessment. In detail, MUSIQ assesses the frame-wise visual quality via a pretrained image transformer. Similarly, DOVER exploits a video transformer to estimate video quality score in terms of aesthetics and technical distortion. CLIP-IQA is determined by the feature cosine similarity between the generated frames and text prompts (e.g., good quality) via CLIP model [32] for quality perception." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.71, + 0.466, + 0.726 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We compare our GenVE with several state-of-the-art techniques, including the reference-based SR approach DATSR [4], and diffusion-based video enhancement approaches, i.e., LaVie-SR [39], Upscale-A-Video [55], VEnhancer [14] on YouHQ40, VideoLQ and AIGC-Vid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. Table 1 summarizes the performances of different approaches in terms of the three metrics over three datasets. Overall, our GenVE achieves the best performances across all metrics on both real video data and AI-synthesized video data. In general, higher MUSIQ and CLIP-IQA scores indicate the better visual percep" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.633 + ], + "angle": 0, + "content": "tual quality on the aesthetic aspect. Specifically, GenVE achieves the MUSIQ score of 69.93 and the CLIP-IQA score of 0.596 on YouHQ40, outperforming the competitor VEnhancer. Compared to VEnhancer that solely exploits the visual guidance from LQ videos, ours further involves the semantics-aligned HQ image reference as the bridge to enhance the patch-wise texture alignment in video enhancement, attaining better aesthetic quality. Similar performance trend can be also observed on AIGC-Vid with the AI-synthesized videos. Such performances also demonstrate our model capability to handle complex video degradation scenarios that are brought by the basic video generation model in text-to-video diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative Comparison. Figure 4 showcases four video enhancement examples from real data in YouHQ40 and VideoLQ. Compared to other baselines, GenVE successfully synthesizes fine-grained details (e.g., the clear house windows in the 1-st and 2-nd case) with better visual appearance. Even for the extremely degraded video (e.g., the human face in the 4-th case), GenVE can still capture the major structural contents and imagine the reasonable local parts. The results demonstrate the advantage of exploring comprehensive information of high-quality image reference to facilitate spatial details synthesis in video diffusion model. We further visualize four video enhancement cases of the AI generated videos in Figure 5 to verify the model generalization ability. As depicted in the figure, there exists some blur (the guitar in the 3rd case) in the result videos by Upscale-A-Video. The local structures of the guitar frets in 1-st case generated by VEnhancer are also unorganized." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17092" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.082, + 0.276, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.171, + 0.276, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.261, + 0.276, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.349, + 0.276, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.439, + 0.235, + 0.449 + ], + "angle": 0, + "content": "Input Video Frame" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.086, + 0.403, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.174, + 0.403, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.264, + 0.403, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.354, + 0.403, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.439, + 0.36, + 0.449 + ], + "angle": 0, + "content": "DATSR" + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.086, + 0.528, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.175, + 0.528, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.265, + 0.528, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.406, + 0.354, + 0.528, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.439, + 0.49, + 0.449 + ], + "angle": 0, + "content": "LaVie-SR" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.086, + 0.653, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.175, + 0.653, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.265, + 0.653, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.354, + 0.653, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.439, + 0.633, + 0.45 + ], + "angle": 0, + "content": "Upscale-A-Video" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.086, + 0.778, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.175, + 0.777, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.265, + 0.777, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.354, + 0.777, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.439, + 0.745, + 0.449 + ], + "angle": 0, + "content": "VEnhancer" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.086, + 0.902, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.175, + 0.902, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.265, + 0.902, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.354, + 0.902, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.44, + 0.854, + 0.449 + ], + "angle": 0, + "content": "GenVE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.452, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Figure 5. Four visual examples of video enhancement results by different approaches on AI-synthesized videos from AIGC-Vid. The region in the red box is presented in the zoom-in view for comparison." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.489, + 0.273, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.487, + 0.895, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.618, + 0.907, + 0.647 + ], + "angle": 0, + "content": "Figure 6. Comparison of temporal profile that tracks variation of the pixels in a spatial row (highlighted by the yellow line). The width of the temporal profile equals to the video width, and the height is the number of frames of generated HQ video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.483, + 0.82 + ], + "angle": 0, + "content": "In contrast, the result videos of our GenVE contain more fancy visual details and natural local textures, confirming the merit of texture alignment learning for high-fidelity local region restoration. Moreover, as shown in the 4-th case, our GenVE nicely distinguishes the background tiles from the toy bear, and accordingly synthesizes the smooth and fuzzy textures on their surfaces, respectively. That basically indicates the local enhancement of each video patch can be accurately calibrated by visual cues from the corresponding region in the image reference, manifesting the merit of local-aware cross-attention in GenVE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Besides the validation of visual appearance, we also verify the temporal consistency of generated video via the temporal profile. Figure 6 depicts the temporal variation of the pixels in a specific spatial row (i.e., the yellow line) across different approaches. As highlighted by the red box, GenVE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.731 + ], + "angle": 0, + "content": "exhibits a more seamless and smoother transition across the temporal dimension, while the others display blurriness or discontinuities. The result demonstrates the effectiveness of the motion-based conditioning augmentation exploited in GenVE to achieve better temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.669, + 0.759 + ], + "angle": 0, + "content": "4.3. Model Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Analysis on LCA and Conditioning Augmentation. We first investigate how the LCA module and conditioning augmentation strategy influence the overall performances of video enhancement. Table 2 lists the performance comparisons among variants with different combination ways of the two components. We start from the basic video diffusion model A with a controller [49], which only extracts the visual cues from low-quality videos for video enhancement. The model B upgrades the basic model A through learning" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17093" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.079, + 0.486, + 0.107 + ], + "angle": 0, + "content": "Table 2. Performance comparisons among different variants of GenVE on the YouHQ40 dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.11, + 0.478, + 0.255 + ], + "angle": 0, + "content": "
ModelFusionAugmentationMUSIQ↑DOVER↑
CALCANoiseMaskTemporal
A68.3477.29
B✓69.1078.75
C✓69.4579.98
C1✓✓69.5380.16
C2✓✓✓69.8882.15
C3✓✓✓✓69.9383.50
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.264, + 0.483, + 0.521 + ], + "angle": 0, + "content": "feature interaction between image reference and LQ video by vanilla cross-attention (CA) to emphasize detail generation, and leads 1.46 improvement of DOVER. By replacing vanilla CA with our deliberately designed local-aware cross-attention (LCA) module, the model C boosts DOVER from 78.75 to 79.98. Compared to general cross-attention, our LCA module imposes the neighboring textures for more accurate feature selection via involving attention bias and local patch reshuffling, thereby facilitating the synthesis of fine-grained objects. The runs of model \\(\\mathbf{C}_1\\), \\(\\mathbf{C}_2\\) and \\(\\mathbf{C}_3\\) further validate different types of conditioning augmentation in GenVE. Notably, there is a large gain of 1.99 in terms of DOVER from model \\(\\mathbf{C}_1\\) to \\(\\mathbf{C}_2\\), verifying the efficacy of masking conditioning to benefit feature selection and improve model capacity in training. Finally, the model \\(\\mathbf{C}_3\\), i.e., our GenVE, by equipping all the three augmentations for robust model optimization, shows the best performances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Analysis on Noise Level. Then, we study the influence of the noise level on the image reference and input LQ video in the inference stage. We show two video cases generated by GenVE with different noise levels in Figure 7. As shown in the figure, the appearance of output HQ video is closer to that of the input LQ video when the noise level is smaller. The results are expected since that the adding noise in forward diffusion is small and the backward denoising cannot change a lot. Conversely, when exploiting high noise level in video noisings, more fancy local details (e.g., the tooth of human in the 2-nd case) are presented while some visual contents can be changed. Therefore, we achieve the tradeoff control between the video quality and fidelity by additionally involving noise augmentation in diffusion model training. In evaluation, we empirically set the noise level as 250 to seek a good balance between synthesized novel details and original visual appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Analysis on Image Reference. The image reference acts as a bridge between the input LQ video and generated HQ video in our GenVE. A valid question is how different image references influence the final results of video enhancement. Figure 8 shows two visual cases of video enhancement by using two distinct image references in GenVE. Although there are substantial visual differences between the two conditioning image references, the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.08, + 0.614, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.153, + 0.612, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.225, + 0.58, + 0.234 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.08, + 0.71, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.153, + 0.709, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.225, + 0.705, + 0.234 + ], + "angle": 0, + "content": "Noise level n=100" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.08, + 0.806, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.153, + 0.806, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.225, + 0.801, + 0.234 + ], + "angle": 0, + "content": "Noise level \\( n = 200 \\)" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.08, + 0.903, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.153, + 0.903, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.224, + 0.898, + 0.233 + ], + "angle": 0, + "content": "Noise level \\( n = 300 \\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.236, + 0.905, + 0.262 + ], + "angle": 0, + "content": "Figure 7. Two visual examples of video enhancement results by using different noise levels in GenVE." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.264, + 0.592, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.323, + 0.592, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.38, + 0.568, + 0.39 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.264, + 0.75, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.323, + 0.749, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.38, + 0.665, + 0.389 + ], + "angle": 0, + "content": "Reference-1" + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.38, + 0.731, + 0.389 + ], + "angle": 0, + "content": "Result-1" + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.264, + 0.829, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.323, + 0.828, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.38, + 0.82, + 0.389 + ], + "angle": 0, + "content": "Reference-2" + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.264, + 0.904, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.323, + 0.904, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.38, + 0.886, + 0.389 + ], + "angle": 0, + "content": "Result-2" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.39, + 0.905, + 0.417 + ], + "angle": 0, + "content": "Figure 8. Two visual examples of video enhancement results guided by two different image references in GenVE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.5 + ], + "angle": 0, + "content": "final video enhancement results are perceptually similar. We speculate that the results are contributed by the augmentation strategy of masking condition which encourages GenVE to search multiple visual cues in image reference holistically, enhancing the generation of local details." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.64, + 0.529 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.81 + ], + "angle": 0, + "content": "We have presented GenVE that explores the image reference to boost diffusion-based video enhancement. In particular, we study the problem from the view point of leveraging image reference as the bridge for semantic and texture alignment between input low-quality video and generated high-quality video. To consolidate our idea, GenVE first exploits an image diffusion model to upscale a key frame of input video for semantics-aligned high-quality image reference generation. A video controller then transfers the patch-level texture details in high-quality image reference to the correspondence region of low-quality video for fine-grained details enhancement. In the training of GenVE, a series of conditioning augmentation strategies is conducted to further strengthen the model robustness. Experiments conducted on two public real-world video datasets, i.e., YouHQ40 and VideoLQ, as well as a self-built AI-synthesized video dataset AIGC-Vid, validate the effectiveness of the proposed GenVE for video enhancement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the Beijing Municipal Science and Technology Project No. Z241100001324002, and Beijing Nova Program No. 20240484681, and the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17094" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime Video Super-Resolution with Spatio-Temporal Networks and Motion Compensation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer. arXiv:2505.22705, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based Image Super-Resolution with Deformable Attention Transformer. In ECCV, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving Video SuperResolution with Enhanced Propagation and Alignment. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating Tradeoffs in Real-World Video Super-Resolution. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[9] Jingyuan Chen, Fuchen Long, Jie An, Zhaofan Qiu, Ting Yao, Jiebo Luo, and Tao Mei. Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[10] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wenggang Zhou, Jiebo Luo, and Tao Mei. Learning 3D Shape Latent for Point Cloud Completion. IEEE Trans. on Multimedia, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[11] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[14] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. VEnhancer: Generative Space-Time Enhancement for Video Generation. arXiv:2407.07667, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS Workshop, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.15 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.153, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv:2210.02303, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.265 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.268, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive Benchmark Suite for Video Generative Models. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.353, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video Super-Resolution with Recurrent Structure-Detail Network. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-Scale Image Quality Transformer. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[23] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-Correspondence Aggregation Network for Video Super-Resolution. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.485, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A Video Restoration Transformer. IEEE TIP, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.528, + 0.905, + 0.596 + ], + "angle": 0, + "content": "[25] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent Video Restoration Transformer with Guided Deformable Attention. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[26] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[27] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[28] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Bi-calibration Networks for Weakly-Supervised Video Representation Learning. IJCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[29] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. VideoStudio: Generating Consistent-Content and Multi-Scene Videos. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards Photorealistic Image Generation and Editing with Text-guided Diffusion Models. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17095" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[34] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking Alignment in Video Super-Resolution Transformers. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[35] Gyumin Shim, Jinsun Park, and In So Kweon. Robust Reference-based Super-Resolution with Similarity-Aware Deformable Convolution. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[36] Zachary Teed and Jia Deng. RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.336, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[37] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring CLIP for Assessing the Look and Feel of Images. In AAAI, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[38] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind SuperResolution with Pure Synthetic Data. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[39] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models. IJCV, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[40] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing wen Hou, Annan Wang, Qiong Yan Wenxiu Sun, and Weisi Lin. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[41] Yanchun Xie, Jimin Xiao, Mingjie Sun, Chao Yao, and Kaizhu Huang. Feature Representation Matters: End-to-End Learning for Reference-based Image Super-resolution. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[42] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal Modulation Network for Controllable Space-Time Video Super-Resolution. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[43] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T. Freeman. Video Enhancement with Task-Oriented Flow. IJCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[44] Xu Yan, Weibing Zhao, Kun Yuan, Ruimao Zhang, Zhen Li, and Shuguang Cui. Towards Content-Independent Multi-Reference Super-Resolution: Adaptive Pattern Matching and Feature Aggregation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[45] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning Texture Transformer Network for Image Super-Resolution. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[46] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations. In ICCV, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[47] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient Video Super-Resolution. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[48] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Landmark Image Super-resolution by Retrieving Web Images. IEEE TIP, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[50] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv:2311.04145, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[51] Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. Image Super-Resolution by Neural Texture Transfer. In CVPR, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[52] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[53] Zhongwei Zhang, Fuchen Long, Zhaofan Qiu, Yingwei Pan, Wu Liu, Ting Yao, and Tao Mei. MotionPro: A Precise Motion Controller for Image-to-Video Generation. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[54] Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, and Lu Fang. CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video SuperResolution. In CVPR, 2024. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17096" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_origin.pdf b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0cce57c96d8ea9db96784b1c8a748269f356bac --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/48366b30-d2f0-4312-be7c-cf9d18325fdd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef9940e31320d7913bea2a5ff35d347614b19faff1260fbcf864dd1d7d7dd5a2 +size 13206755 diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/full.md b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b7d8ccf10527b3fa11c2e1c12ec1f017e5d7e38a --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/full.md @@ -0,0 +1,413 @@ +# Aligning Global Semantics and Local Textures in Generative Video Enhancement* + +Zhikai Chen $^{1}$ , Fuchen Long $^{2,\dagger}$ , Zhaofan Qiu $^{2}$ , Ting Yao $^{2}$ , Wengang Zhou $^{1}$ , Jiebo Luo $^{3}$ , and Tao Mei $^{2}$ + +$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China $^{2}$ HiDream.ai Inc. $^{3}$ University of Rochester, Rochester, NY USA + +czk654@mail.ustc.edu.cn,{longfuchen, qiuzhaofan, tiyao}@hidream.ai + +zhwg@ustc.edu.cn, jluo@cs.ROchester.edu, tmei@hidream.ai + +# Abstract + +Recent advances in video generation have demonstrated the utility of powerful diffusion models. One important direction among them is to enhance the visual quality of the AI-synthesized videos for artistic creation. Nevertheless, solely relying on the knowledge embedded in the pretrained video diffusion models might limit the generalization ability of local details (e.g., texture). In this paper, we address this issue by exploring the visual cues from a high-quality (HQ) image reference to facilitate visual details generation in video enhancement. We present GenVE, a new recipe of generative video enhancement framework that pursues the semantic and texture alignment between HQ image reference and denoised video in diffusion. Technically, GenVE first leverages an image diffusion model to magnify a key frame of the input video to attain a semantics-aligned HQ image reference. Then, a video controller is integrated into 3D-UNet to capture patch-level texture of the image reference to enhance fine-grained details generation at the corresponding region of low-quality (LQ) video. Moreover, a series of conditioning augmentation strategies are implemented for effective model training and algorithm robustness. Extensive experiments conducted on the public YouHQ40 and VideoLQ, as well as self-built AIGC-Vid dataset, quantitatively and qualitatively demonstrate the efficacy of our GenVE over the state-of-the-art video enhancement approaches. Source code is available at https://github.com/HiDream-ai/GenVE. + +# 1. Introduction + +With the rapid development of video diffusion models [18, 39], there is a growing demand to elevate the visual quality of the synthesized video contents. Video enhancement, as an emerging compute vision topic distinct from traditional video super-resolution, aims to refine the low + +![](images/d20047f8fd5ed59d7b05e491665c1a6bc4afdd5a168baa24ea06a5574c265e54.jpg) +Figure 1. An illustration of visual alignment in our generative video enhancement (GenVE), including the global semantic alignment between LQ video and image reference, and the local texture alignment between image reference and HQ video. + +quality (LQ) video to achieve better appearance and motion quality, adding more innovative details and endowing smoother temporal dynamics to the input video. Several studies [14, 39, 50, 55] have tackled this issue by treating it as a conditioning video diffusion paradigm. The prior knowledge of the input LQ video is exploited to regulate the video denoising procedure to produce the high-quality (HQ) counterpart. Despite having impressive results in the realm of video enhancement, such recipe still faces the challenge of generating fine-grained details (e.g., texture) due to the lack of some content prior knowledge. + +Instead of solely relying on the input LQ video as the condition, we consider involving an additional knowledge connection, i.e., high-quality image reference, to bridge the gap of fine-grained details between LQ and HQ videos. One intuition behind this is that the visual quality of image is usually higher than that of video with the same resolution, which is caused by degradation brought by temporal dynamics (e.g., motion blur). Technically, we formulate our + +reference-guided video enhancement in diffusion paradigm from two perspectives: global semantic alignment and local texture alignment. As depicted in the upper part of Figure 1, global semantic alignment is achieved by transferring the information of holistic layout (e.g., the relative position of cat and cloth) of a key frame to the HQ image reference via diffusion-based image up-scaling. Next, we develop a video controller in 3D-UNet to calibrate the video patch denoising with the guidance of local texture details (e.g., the cat eye and nose) in the obtained image reference. The learning of patch-level correspondences between denoised videos and image reference are executed by a cross-attention mechanism, boosting the fine-grained contextual interaction. Moreover, we investigate a set of augmentation approaches to enhance the robust learning of our reference-guided video enhancement model. + +To materialize our idea, we present GenVE to carry out semantic and texture alignment for generative video enhancement. Technically, GenVE first utilizes an image diffusion model to create a high-quality (HQ) image reference based on a key frame of input low-quality (LQ) video through forward-backward diffusion process. In each step of backward diffusion, the image reference denoising is regulated by a weighted summation of LQ key frame features to achieve global semantic alignment. Then, both of the image reference and LQ video are fed into a video controller with the newly-devised local-aware cross-attention (LCA) to attain local texture alignment between image reference and denoised video in diffusion. In LCA, the feature patch of LQ video acts as the query while that of image reference serves as the key and value for computation. A local attention bias is further exploited to impose the effects of neighboring patches around each patch in LQ video. Furthermore, GenVE involves a series of data augmentation strategies on input conditions (i.e., LQ video and image reference) to strengthen the model robustness during training. + +The main contribution of this paper is the proposed GenVE to explore semantic and texture alignment between image reference and denoised video in diffusion. The solution also leads to the elegant views of how to maintain the semantic correspondence in high-quality image reference synthesis, and how to leverage the texture information in the reference to benefit local detail generation in video denoising. Extensive experiments on three benchmarks verify the superiority of GenVE for video quality improvement. + +# 2. Related Work + +Traditional video super-resolution. Traditional video super-resolution (VSR) aims to up-scale video frames while keeping the fidelity, which could be categorized into two groups, i.e., sliding window-based methods and recurrent methods. Sliding window-based techniques [2, 23, 24, 42, 43, 46] rely on CNNs [28] or Transformers [26, 27] which + +aggregate the information of a sequence of low-resolution (LR) frames to predict center high-resolution (HR) frame. One representative is VRT [24] that plugs the temporal mutual attention block into transformer backbone to facilitate motion estimation, feature alignment and fusion. In contrast to aggregate information from adjacent frames, recurrent approaches [5, 6, 21, 25, 34, 47] utilize a hidden state to sequentially propagate information from all previous frames to the current frame, benefiting the frame restoration. Despite the great performance on magnifying original visual contents in the input LR video, such regression paradigm still faces the generalization problem to produce novel fine-grained details and improve the video aesthetic quality. + +Reference-based image super-resolution. There has been a series of investigation [35, 41, 44, 45, 48, 51, 54] to leverage the reference for image super-resolution (ISR). One typical way [48, 54] is to refine low-resolution (LR) image with the texture information in reference image via global visual alignment. For instance, Landmark [48] aligns the reference image with the LR image through optimizing a global registration problem and exploits the texture for visual enhancement. To fully mine the correspondence between reference image and input LR image, Zheng et al. [54] introduce cross-scale warping in CNNs to fuse features in different scales. Another direction for reference-based ISR is conducted in patch level. Recent advance TTSR [45] adopts a texture transformer to search active reference patches via cross-attention for image detail refinement. In this paper, we leverage the reference-based video enhancement recipe, and further delve into the context alignment between LQ and HQ videos from both semantic and texture perspectives. + +Diffusion-based video enhancement. Significant progress of video synthesis achieved by diffusion models [3, 9-13, 15, 29, 30, 32, 52, 53] encourage the development on video enhancement. Distinct from traditional video super-resolution that focuses on fidelity preserving, diffusion-based video enhancement approaches [14, 39, 50, 55] can imagine fancy details for appearance improvement. Most of them leverage the knowledge prior embedded in the pre-trained image diffusion models and formulate the problem as conditional video diffusion. For example, Zhou et al. [55] freeze the pre-trained spatial layers of the image diffusion model [33] and train the temporal latent propagation module for video enhancement. In comparison, LaVie [39] jointly trains a spatial-temporal video diffusion network and attains better temporal consistency. Later in VEnhancer [14], ControlNet [49] is further exploited to learn videow-aware conditional features to enhance the fine-grained object synthesis in video diffusion. Nevertheless, solely depending on the knowledge in pre-trained diffusion models could limit the generalization for local details (e.g., texture). + +In summary, our work focuses on video enhancement + +![](images/da8cbe43d01ae59b7ae4f6a5f63444e613c20494c9edc93d947b71f0267c63dd.jpg) +Figure 2. An overview of our GenVE. The input LQ video is first up-sampled to the target resolution via bicubic interpolation. The up-sampled video is compressed as the video latent $Z$ and then fed into the image diffuser to synthesize an image reference latent $y$ . Next, the latents $y$ and $Z$ are converted to the augmented image reference latent $Y_{n}$ and video latent $Z_{n}$ as the conditional signals. Concurrently, the up-sampled video latent $Z$ is noised by the Gaussian vector $\epsilon$ to obtain the noisy video latent $Z_{T_1}$ . Hence, the video controller exploits the two augmented latents $Y_{n}$ and $Z_{n}$ , and the noisy video latent $Z_{T_1}$ to learn a visual guidance for diffusion calibration. Finally, the HQ video latent $Z_{0}$ is synthesized from $Z_{T_1}$ via video diffuser based on the learnt visual guidance. + +![](images/5d88722cfd29ec5b0a7c2b5f7821c9d036bc964add3a687589fc7bd10663d95e.jpg) + +using image reference in diffusion models. The proposal of GenVE contributes by investigating not only how to obtain semantics-aligned image reference through diffusion forward-backward calibration, but also how to benefit patch-level texture details synthesis in video denoising. + +# 3. Approach + +In this section, we present our newly-minted GenVE, pursuing global semantics alignment and local texture alignment between LQ and HQ videos for video enhancement. Figure 2 depicts an overview of this pipeline. Given a LQ video, we up-sample it through bicubic interpolation, and then encode it to the latent representation via VAE. Then, the image diffuser is exploited to refine a key frame from the up-sampled video to a semantics-aligned high-quality image reference. After that, the latents of image reference and up-sampled video are transformed by a series of conditioning augmentation strategies. The video controller performs local-aware cross-attention (LCA) between the feature patches of the augmented video frame latent and image reference latent for local texture alignment. These features are passed through cascaded 3D encoder blocks to learn the visual guidance. Finally, the video diffuser exploits the learnt visual guidance to calibrate the denoising procedure for video enhancement. + +# 3.1. Preliminaries: Video Diffusion Models + +Video diffusion models [1, 18, 19, 50] typically employ a 2D-VAE as the video encoder and a 3D-UNet as the latent denoiser. Given a video with $I$ frames, the VAE first encodes it into the latent code $Z = \{z^i\}_{i=1}^I$ frame-by-frame. Each feature point in the latent denotes a local patch in the original frame. To conduct video refinement using diffusion models, we follow the natural solution and implement forward and backward diffusion procedures on the input video. + +In the forward process, the Gaussian noise is gradually added to the video latent code $Z$ with time step $T$ , formulated by a Markov chain. The noised latent $Z_{t}$ at each forward diffusion step $t$ is thus formulated as: + +$$ +Z _ {t} = \alpha_ {t} Z + \sigma_ {t} \epsilon , \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ is the Gaussian noise and $\alpha_{t},\sigma_{t}$ specify the noise schedule of DDPM [17], leading the proportion of noise to gradually increase over the steps $t$ . + +In the backward process, the data reconstruction for refinement eliminates the added noise in the forward diffusion process, which can start from a noise sample $Z_{t}$ at arbitrary time step $t$ . Each backward diffusion step is a denoising process formulated by: + +$$ +Z _ {t - 1} = \left(Z _ {t} - \tilde {\sigma} _ {t} \epsilon_ {t}\right) / \tilde {\alpha} _ {t}, \tag {2} +$$ + +where $\tilde{\sigma}_t$ , $\tilde{\alpha}_t$ are the constant hyper-parameters related to time step $t$ based on the noise scheduler [16, 17], and $\epsilon_t = \epsilon_\theta(Z_t)$ is the estimated noise vector using the denoising network $\epsilon_\theta$ . The HQ video latent $Z_0 = \{z_0^i\}_{i=1}^I$ is thus iteratively denoised by the backward diffusion from the noised LQ video latent $Z_t$ to improve the video quality. + +# 3.2. Global Semantic Alignment + +Reference-based super-resolution [45, 51] usually exploits a semantic-related HQ image reference to guide LQ enhancement. Nevertheless, the layouts are not required to be matched between each other, leading to potential spatial misalignment of the objects. Such misalignment could limit the enhancement quality of local details. To alleviate this, we propose to generate a semantics-aligned HQ image reference for reference-based video enhancement. An image diffuser first leverages the original noisy image latent to refine denoising procedure for HQ image reference generation. The generated image reference will be further em + +ployed to transfer the patch-level texture information into LQ video frames, guiding the final video enhancement. + +Semantics-Aligning Image Diffuser. Given the latent $z$ of a key frame from the input video latent $Z = \{z^i\}_{i=1}^I$ , we first add the Gaussian noise with time step $T_0$ on $z$ . Then, an image diffusion model (i.e., image diffuser in Figure 2) performs backward diffusion on the noisy frame latent $z_{T_0}$ for image refinement. At each denoising step $t$ , two intermediate latents are prepared in advance. The first is the denoised frame latent $z_t$ computed through the backward diffusion, and the other is a noisy frame latent $\tilde{z}_t$ with time steps $t$ . We refine the denoised frame latent $z_t$ in each denoising step by leveraging the component from $\tilde{z}_t$ through weighted summation as: + +$$ +\tilde {z} _ {t} = \alpha_ {t} z + \sigma_ {t} \epsilon , \tag {3} +$$ + +$$ +z _ {t} = (1 - \gamma_ {t}) z _ {t} + \gamma_ {t} \tilde {z} _ {t}, +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ and $\alpha_{t},\sigma_{t}$ are the constant hyperparameters as mentioned in Eq.(1). Note that the parameter $\gamma_{t}$ is defined by a cosine function on $t$ , decreasing from 0.5 to 0 monotonically with $t$ from $T$ to 0. In this way, the image diffuser introduces a substantial amount of visual content from the original frame at the early denoising phase, exactly where the latent features have rich high-level visual semantics. This denoising mechanism encourages the denoised HQ image reference to align with visual semantics of the original LQ frame, while maintaining low-level texture details as more as possible. + +Through iterative latent denoising of image refinement, the image diffuser successfully derives the image reference latent $y$ , which is semantically aligned with the key frame and has richer texture details. + +# 3.3. Conditioning Augmentation + +Next, we perform a series of conditioning augmentation strategies on both image reference and LQ video latent in training to simulate various scenarios, thereby boosting model robustness. The augmentation includes noise augmentation, temporal augmentation and masking condition. + +Noise Augmentation. To enable controllable video refinement, a noise augmentation strategy is introduced on both video and image conditions. We implement the noise augmentation on the image reference latent $y$ and the video latent $Z = \{z^i\}_{i=1}^I$ , following the forward diffusion as Eq.(1) with $n$ time steps, thereby producing the noisy conditional latents $\tilde{y}_n$ and $Z_n = \{z_n^i\}_{i=1}^I$ . The noise augmentation set the noise level $n$ different at each training step, corrupting the low-level visual information. Hence, the noise level can be adjusted in inference to balance the visual fidelity and quality, thus obtaining the most realistic videos. + +Temporal Augmentation. To further improve the video temporal coherence, the temporal augmentation is applied to incorporate motion clues in the original LQ video into the + +![](images/b20d5d13e15b01f53bf1629c23a572bcdefa03763b84244101e9c93632403c3a.jpg) +Figure 3. Illustration of Local-Aware Cross-Attention (LCA). Given the augmented video frame latent $z_{n}^{i}$ and image reference latent $y_{n}^{i}$ , LCA first extracts the feature patches using two convolutional layers, respectively. Then, the feature patches of image reference latent are reshuffled within each neighboring $d$ patches. By treating video feature patches as query and the image reference feature patches as key/value, the attention map is computed and further added by an extra local attention bias $B$ , to estimate the output texture-aligned feature embedding $e$ . + +HQ video diffusion process. We first estimate the optical flows using RAFT model [36] on the input LQ video $Z$ . Then, the attained flows are utilized to warp the noisy image reference latent $\tilde{y}_n$ . As such, it yields a set of corresponding warped image references latents $Y_{n} = \{\tilde{y}_{n}^{i}\}_{i = 1}^{I}$ , reinforcing the motion alignment between the warped image references and the input LQ video. + +Masking Condition. Moreover, we leverage the masking operation on the warped image reference latents to facilitate the model training. The masking procedure is conducted on each latent $\tilde{y}_n^i$ of warped image references using random mask ratio $r$ , generating the masked noise image reference latent $y_{n}^{i}$ . During training, the masking condition will set a portion of the original input pixels to zeros. In inference, we maintain the whole input $\tilde{y}_n^i$ to ensure that the entire image reference data is available for the optimal result. As such, the masking stimulates the model to refer to more potential texture feature patches in the image reference for visual enhancement. + +By the sequential conditioning augmentation, we obtain the augmented video latent $Z_{n} = \{z_{n}^{i}\}_{i = 1}^{I}$ and the sequence of augmented image reference latents $Y_{n} = \{y_{n}^{i}\}$ from $Z$ and $y$ , respectively. These two latents will be further exploited for visual guidance learning in video controller. + +# 3.4. Local Texture Alignment + +To encourage the video diffuser to generate more details for video enhancement, we leverage a video controller and further design a local-aware cross-attention (LCA) module to learn texture-aligned feature embedding. The local-aware cross attention concentrates on matching texture features of similar spatial regions between image reference and input LQ video frame, facilitating more efficient and precise feature selection. The texture-aligned feature embedding produced by LCA is then utilized for visual guidance learning. + +Table 1. Performance comparisons in terms of MUSIQ, DOVER and CLIP-IQA on the YouHQ40, VideoLQ and AIGC-Vid datasets. + +
DatasetsMetricsBicubicDATSR[4]LaVie-SR[39]Upscale-A-Video[55]VEnhancer[14]GenVE
YouHQ40MUSIQ ↑48.2149.2264.1763.6968.9969.93
DOVER ↑40.7630.0479.6474.0677.0883.50
CLIP-IQA ↑0.3920.4180.5810.5090.5940.596
VideoLQMUSIQ ↑59.5359.1364.5060.9066.1968.23
DOVER ↑50.9432.5666.9960.3871.4774.30
CLIP-IQA ↑0.3520.3960.4590.3800.4540.464
AIGC-VidMUSIQ ↑60.3357.4863.7160.0064.3064.56
DOVER ↑55.7357.2676.8474.1876.4677.02
CLIP-IQA ↑0.6010.5870.6560.6220.6340.658
+ +Finally, the learnt visual guidance steers video denoising in video diffuser for video enhancement. + +Texture-Aligning Video Controller. The right part of Figure 2 depicts the pipeline of our video controller. The video controller leverages the LCA module for texture alignment, and the cascaded encoder blocks inherited from the 3D-UNet for feature enhancement. Specifically, the video controller takes three inputs: the augmented video latent $Z_{n}$ , the sequence of augmented image reference latents $Y_{n}$ , and the denoised video latent $Z_{t}$ that is initialized from the noised $Z_{T_1}$ with time step $T_{1}$ . The LCA module, as illustrated in Figure 3, first extracts the image feature patches from the image reference latent $y_{n}^{i}$ and video frame latent $z_{n}^{i}$ using convolutional layers, and then flatten them into patch sequences. Next, LCA executes local patch reshuffling on the patch sequences to boost model training. In detail, LCA randomly reshuffles the patch order within each neighboring $d$ patches. The image reference feature patches serve as the keys $K$ and values $V$ , and the video frame feature patches are the queries $Q$ for attention computation. Particularly, we add a local attention bias matrix $B$ on the attention weights before the Softmax operation. Thus, the attention mechanism can be formulated as: + +$$ +B = \left( \begin{array}{c c c} b _ {1 1} & b _ {1 2} & \dots \\ b _ {2 1} & b _ {2 2} & \dots \\ \vdots & \vdots & \ddots \end{array} \right), \quad b _ {i j} = \left\{ \begin{array}{l l} 1, & \text {i f} | i - j | < d \\ 0, & \text {i f} | i - j | \geq d \end{array} \right., \tag {4} +$$ + +$$ +\tilde {Q} = \operatorname {N o r m} (Q), \quad \tilde {K} = \operatorname {N o r m} (K), +$$ + +$$ +A = \operatorname {S o f t m a x} (\frac {\tilde {Q} \tilde {K} ^ {T} + B}{\sqrt {d _ {k}}}), +$$ + +where $d$ equals to the local reshuffling range and $d_{k}$ is the channel dimension. As such, the local patch alignment mainly focuses the reference texture feature learning located within a radius of $d$ , thus facilitating precise transferring of high-quality texture patterns. After that, we feed the output attention features into another convolutional layer and add it to the query features as the texture-aligned feature embedding $e$ , which is formulated as follows: + +$$ +e = \operatorname {C o n v} (A V) + Q. \tag {5} +$$ + +The texture-aligned feature embedding is added to the output video features of the following encoder block, and further propagated to the subsequent cascaded blocks for multi-scale visual guidance learning. + +Finally, we collect all learnt visual guidance features output from the video controller at different scales, to regulate the multi-step backward diffusion in 3D-UNet until obtaining the final output high-quality video. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We evaluate the effectiveness of GenVE on three datasets, i.e., the public YouHQ [55] and VideoLQ [7], as well as the self-built AIGC-Vid. The YouHQ dataset comprises 38K high-quality videos for training and 40 high-quality videos for testing (namely YouHQ40), with approximately 30 frames per video. We empirically apply the multi-tier degradation model [38] to reduce the visual quality and obtain the high-low quality video pairs for training and validation. The VideoLQ dataset [7] is a real-world collection containing 50 low-quality videos for evaluation. There are 100 frames for each video. Additionally, we build the AIGC-Vid dataset, which is generated by three text-to-video advances, i.e., ModelScope [50], LaVie [39] and VideoCrafter [8]. We exploit each model to generate 93 videos based on the text prompts from VBench [20], and finally collect 279 AI-synthesized videos for evaluation. + +Implementation Details. We implement GenVE on the PyTorch platform. The image diffuser and the video diffuser is built on the pre-trained image diffusion model SDXL [31] and video diffusion model I2VGen-XL [50], respectively. The time steps $T$ for the base video diffusion model in training is set as 1000. The denoising time steps $T_{0}$ for image reference generation is 700, and $T_{1}$ in video enhancement is 900, respectively. In training, the parameter $n$ for noise augmentation is an integer randomly sampled within the range of [1, 300]. We empirically set it to 250 in inference. The masking ratio $r$ is randomly sampled from [0, 1) in training and set to 0 in inference. We set the local reshuffling range $d$ of LCA to 7, and define the parameter $\gamma_{t}$ in image diffuser as $\gamma_{t} = \frac{1}{2}\left[\frac{1}{2}(1 + \cos\left(\frac{T - t}{T}\pi\right))\right]^{3}$ . Our + +![](images/c96baac8e4c277f25b6a2f21601466d29d9e80bc4edc422dab981b04cb28a951.jpg) + +![](images/870d126a3c940aa73b3a31c2090c8d89bbc776216e2cbcd72c0afc4346cbc648.jpg) + +![](images/b54112942778eb6efaa101eb2226dae3f3cac3bc509f56fc227210ec21e1057b.jpg) + +![](images/f3def86ccbc95d30516b1ab3bb62911ea8e0d65bf165b99034763e9d6038b7f9.jpg) +Input Video Frame +Figure 4. Four visual examples of video enhancement results by different approaches on the real video data from YouHQ40 and VideoLQ. The region in the red box is presented in the zoom-in view for comparison. + +![](images/e35e3a8f7844b368abe3d492d2721ef0143c95629f387dc8510948e09c9f7a5f.jpg) + +![](images/ddaddacb3748ab57eb2d2d250bdde4f224d78e191b7a546bcf72bc4095c16999.jpg) + +![](images/762a2677c4ec4686b261d2880af3669f4d9860c8855acee62fbb70190ab31732.jpg) + +![](images/c6a39d0284fd16416b36660644e25b2c483575d523dc04c03ecd768c193b575a.jpg) +DATSR + +![](images/a6f83e9413ec207f55a715d59cb36ee3a257f617bfe5fc6056aeec3b10e14194.jpg) + +![](images/c5ca676453211d34c1e0366f777dd81b04c79b0016c463113d72236b89d8e8d6.jpg) + +![](images/b6d6ed5b45c3167d5d4aa755ea3cceda27230a3d97fd02a482c20f57d36488d7.jpg) + +![](images/da62cc89107ec116e891321beaaa2f61a75576cfa9a8690ab6ba37c0ed9554e7.jpg) +LaVie-SR + +![](images/fd0745d020e0cbccd6e6193c90cc506e7355cd590418f94e35a83d450117bd57.jpg) + +![](images/05875dcb0cb3583975ead1a20ece100761bea6207a08c2a14e7e68b4f8695c7b.jpg) + +![](images/e3fc2eaa0139bcd702c38e2dd9c2c03198b8d43e140c6ab78efb138d1098cb98.jpg) + +![](images/14a278d7790969eb4ee783f714aa20185f4ef529a26f265507b9fdbb93a14183.jpg) +Upscale-A-Video + +![](images/10ea8beebca93c7eb750feefc65488c16a89be6bdd2879c4bd38ce71b6104bd0.jpg) + +![](images/ab3675295a97807d3388168035aa04da5578cf331e29ece1720e91020e37be49.jpg) + +![](images/e724dbc6684edbbff920540146e70534790cd794c307038eac4c8e266226967f.jpg) + +![](images/b697ef6c3a0f456a9095fd88df5b40b012fcbd286ae439f9a1f2e0be0501f1d9.jpg) +VEnhancer + +![](images/c17ee9b0d99b369944eab2045dc1d7a4cec7cc73be4af7b3f20cc27351268d73.jpg) + +![](images/89c2563c9e299a33dc3b9792a8fe6869a57cf5f20f0cbf100aa03994850d9251.jpg) + +![](images/b67e91160c03b8c38b389f2fc1302439e54d602153a7ae6933da0a2d3ffe6e21.jpg) + +![](images/19c320db6a07206dd0a535a8ec668af9d282d7f72b63730e999a0d6755041cdb.jpg) +GenVE + +GenVE is optimized using AdamW optimizer with the base learning rate $5 \times 10^{-5}$ and batch size 8. The key frame for the image reference generation is set as the first frame of the input video by cross validation. + +Evaluation Metrics. Following previous video enhancement advances [14, 55], we conduct the evaluation on video quality using three metrics, i.e., MUSIQ [22], DOVER [40], CLIP-IQA [37]. All the three metrics estimate the perceptual scores on input video frames using the pre-trained deep models for quality assessment. In detail, MUSIQ assesses the frame-wise visual quality via a pretrained image transformer. Similarly, DOVER exploits a video transformer to estimate video quality score in terms of aesthetics and technical distortion. CLIP-IQA is determined by the feature cosine similarity between the generated frames and text prompts (e.g., good quality) via CLIP model [32] for quality perception. + +# 4.2. Comparisons with State-of-the-Art Methods + +We compare our GenVE with several state-of-the-art techniques, including the reference-based SR approach DATSR [4], and diffusion-based video enhancement approaches, i.e., LaVie-SR [39], Upscale-A-Video [55], VEnhancer [14] on YouHQ40, VideoLQ and AIGC-Vid. + +Quantitative Comparison. Table 1 summarizes the performances of different approaches in terms of the three metrics over three datasets. Overall, our GenVE achieves the best performances across all metrics on both real video data and AI-synthesized video data. In general, higher MUSIQ and CLIP-IQA scores indicate the better visual percep + +tual quality on the aesthetic aspect. Specifically, GenVE achieves the MUSIQ score of 69.93 and the CLIP-IQA score of 0.596 on YouHQ40, outperforming the competitor VEnhancer. Compared to VEnhancer that solely exploits the visual guidance from LQ videos, ours further involves the semantics-aligned HQ image reference as the bridge to enhance the patch-wise texture alignment in video enhancement, attaining better aesthetic quality. Similar performance trend can be also observed on AIGC-Vid with the AI-synthesized videos. Such performances also demonstrate our model capability to handle complex video degradation scenarios that are brought by the basic video generation model in text-to-video diffusion. + +Qualitative Comparison. Figure 4 showcases four video enhancement examples from real data in YouHQ40 and VideoLQ. Compared to other baselines, GenVE successfully synthesizes fine-grained details (e.g., the clear house windows in the 1-st and 2-nd case) with better visual appearance. Even for the extremely degraded video (e.g., the human face in the 4-th case), GenVE can still capture the major structural contents and imagine the reasonable local parts. The results demonstrate the advantage of exploring comprehensive information of high-quality image reference to facilitate spatial details synthesis in video diffusion model. We further visualize four video enhancement cases of the AI generated videos in Figure 5 to verify the model generalization ability. As depicted in the figure, there exists some blur (the guitar in the 3rd case) in the result videos by Upscale-A-Video. The local structures of the guitar frets in 1-st case generated by VEnhancer are also unorganized. + +![](images/e07dff65030d348a5d033fa00092f872d160550788dc4d7f431836f0784ed692.jpg) + +![](images/83cabc7bc58e42c047e2245f91a42fdd7d235f129d02a33e3fef9e048f2f9810.jpg) + +![](images/c02700f37fa3c7f444f0ca137d9026101d5cd236e43f46c0092a3cfe987a6ac7.jpg) + +![](images/e39cb670fd29f5ad6ff9b61b4c69fda9c67942d116e413763bcff152a17e0b4d.jpg) +Input Video Frame + +![](images/a5f84acd4790a0144ff7b123460adaa6d4f10ca93eeb77d94c5867cb5023c052.jpg) + +![](images/f28e93856f2e262dd626ecc2a9201712a2282dbb61ad7dec0f804c62f7f32e4f.jpg) + +![](images/954b48783b3887052439b66fe3af1676b63299385dedee71c11223f7e3f60952.jpg) + +![](images/9927edb4de5cfab93d4921aeb76c94956f28fa8886ca652e22ed6039b3bd4b8d.jpg) +DATSR + +![](images/a996b60f03f6e0818eab6f484b5ca817547b8277d0dd42dfab78cd42dc78b131.jpg) + +![](images/3e00a16eec4e5f11b7ff534e37b46163c34d9b104deb0dfc5276a76665140bf0.jpg) + +![](images/62feeaadb48c740a8fa021d1d4b5f46f799bfe7baf5aa0bff648db1aeba74b15.jpg) + +![](images/b991057b76c741a5b68fa5408a90f1117a24a259c2001e368d12908b261d6266.jpg) +LaVie-SR + +![](images/a18122c8d1e1811e314b2048307fde175fd7e0dbe388d6572dd40963cca77826.jpg) + +![](images/efd295710595a338fa204d8e60212da84c826726f1968aca43eec85454b5599f.jpg) + +![](images/d8226e7903abcf50b9659fc00b173eff7d4512da80cc81a24012766075c9df02.jpg) + +![](images/941fac3490a8f1f47d513e4ddf3561913f9c7102d4bdb936c982984b9d05e9ce.jpg) +Upscale-A-Video + +![](images/994d530f6379eb7426c24a3e11d4c8ff2da739c6b140ebaa8895196b9ffbcab0.jpg) + +![](images/d5f848cec41cec5579cfb130d81e1b9fff96d037e77f76291b19564460b647d9.jpg) + +![](images/3b718b62722de41569cc9d0eb77ccf08a3d9757729690dbd94c18825757bdde3.jpg) + +![](images/35a6f7995aff3853131987fade9ab028bac39c8d803f6492444db83fc6d5a931.jpg) +VEnhancer + +![](images/ab36d63970e614e57d39a278528bdc339870bd34cd2a0dcc1e30ff8298948f6f.jpg) + +![](images/beec68e6615d8fed09c2099aa7ba3ef9d244d428cf23b7f086f045e7b608016a.jpg) + +![](images/05b46b287d2a9ffb6b31a0800948cd872352004d55b8af8bad14f51d467bca12.jpg) + +![](images/d6ac3e2fbeef7506a8b8979cfc2d18cf400f2822d850afcf8e1d8ba4da71a51a.jpg) +GenVE + +![](images/ebd9ca4e73c249d3d53ab5971e33dd342cc2d470c3d8e92d1fab73985324589d.jpg) +Figure 6. Comparison of temporal profile that tracks variation of the pixels in a spatial row (highlighted by the yellow line). The width of the temporal profile equals to the video width, and the height is the number of frames of generated HQ video. + +![](images/5ba37f5d4c5eab2fe23cc00f95050ddbdac176cc2fa1d33cf12b01ac8d03a57b.jpg) +Figure 5. Four visual examples of video enhancement results by different approaches on AI-synthesized videos from AIGC-Vid. The region in the red box is presented in the zoom-in view for comparison. + +In contrast, the result videos of our GenVE contain more fancy visual details and natural local textures, confirming the merit of texture alignment learning for high-fidelity local region restoration. Moreover, as shown in the 4-th case, our GenVE nicely distinguishes the background tiles from the toy bear, and accordingly synthesizes the smooth and fuzzy textures on their surfaces, respectively. That basically indicates the local enhancement of each video patch can be accurately calibrated by visual cues from the corresponding region in the image reference, manifesting the merit of local-aware cross-attention in GenVE. + +Besides the validation of visual appearance, we also verify the temporal consistency of generated video via the temporal profile. Figure 6 depicts the temporal variation of the pixels in a specific spatial row (i.e., the yellow line) across different approaches. As highlighted by the red box, GenVE + +exhibits a more seamless and smoother transition across the temporal dimension, while the others display blurriness or discontinuities. The result demonstrates the effectiveness of the motion-based conditioning augmentation exploited in GenVE to achieve better temporal coherence. + +# 4.3. Model Analysis + +Analysis on LCA and Conditioning Augmentation. We first investigate how the LCA module and conditioning augmentation strategy influence the overall performances of video enhancement. Table 2 lists the performance comparisons among variants with different combination ways of the two components. We start from the basic video diffusion model A with a controller [49], which only extracts the visual cues from low-quality videos for video enhancement. The model B upgrades the basic model A through learning + +Table 2. Performance comparisons among different variants of GenVE on the YouHQ40 dataset. + +
ModelFusionAugmentationMUSIQ↑DOVER↑
CALCANoiseMaskTemporal
A68.3477.29
B✓69.1078.75
C✓69.4579.98
C1✓✓69.5380.16
C2✓✓✓69.8882.15
C3✓✓✓✓69.9383.50
+ +feature interaction between image reference and LQ video by vanilla cross-attention (CA) to emphasize detail generation, and leads 1.46 improvement of DOVER. By replacing vanilla CA with our deliberately designed local-aware cross-attention (LCA) module, the model C boosts DOVER from 78.75 to 79.98. Compared to general cross-attention, our LCA module imposes the neighboring textures for more accurate feature selection via involving attention bias and local patch reshuffling, thereby facilitating the synthesis of fine-grained objects. The runs of model $\mathbf{C}_1$ , $\mathbf{C}_2$ and $\mathbf{C}_3$ further validate different types of conditioning augmentation in GenVE. Notably, there is a large gain of 1.99 in terms of DOVER from model $\mathbf{C}_1$ to $\mathbf{C}_2$ , verifying the efficacy of masking conditioning to benefit feature selection and improve model capacity in training. Finally, the model $\mathbf{C}_3$ , i.e., our GenVE, by equipping all the three augmentations for robust model optimization, shows the best performances. + +Analysis on Noise Level. Then, we study the influence of the noise level on the image reference and input LQ video in the inference stage. We show two video cases generated by GenVE with different noise levels in Figure 7. As shown in the figure, the appearance of output HQ video is closer to that of the input LQ video when the noise level is smaller. The results are expected since that the adding noise in forward diffusion is small and the backward denoising cannot change a lot. Conversely, when exploiting high noise level in video noisings, more fancy local details (e.g., the tooth of human in the 2-nd case) are presented while some visual contents can be changed. Therefore, we achieve the tradeoff control between the video quality and fidelity by additionally involving noise augmentation in diffusion model training. In evaluation, we empirically set the noise level as 250 to seek a good balance between synthesized novel details and original visual appearance. + +Analysis on Image Reference. The image reference acts as a bridge between the input LQ video and generated HQ video in our GenVE. A valid question is how different image references influence the final results of video enhancement. Figure 8 shows two visual cases of video enhancement by using two distinct image references in GenVE. Although there are substantial visual differences between the two conditioning image references, the + +![](images/c381f913790bc4ee7f47d970a603d235ff1dbf14b47f30bdaee5a5c6c09ee165.jpg) + +![](images/7f3576bbdfd8bc395e429a2457897fbb70f19920e6628c970dacc0259eb6fa2a.jpg) +Input + +![](images/8d55e4ab0db398b4c7a08da4421e13a41dd701b2b6b141b976887bd262c85de4.jpg) + +![](images/ad068027973554dc28e039e405106f7e1ce03455d930353607f5600e3ef738f2.jpg) +Noise level n=100 + +![](images/c5c481be2ad410f55398ef5d17e7452770ff0402d15d206d2ea0ac00c718317d.jpg) + +![](images/ab21a0e837258f829e6b38706a6ef79d04ab9747eac4ba7599cfe66a3260c1e5.jpg) +Noise level $n = 200$ + +![](images/70dcb123aefe06956aed7d5dab3559c4c9c07b198cc0c25f52e7d3974b623c55.jpg) + +![](images/09174802adf8eb672b744eae204185a37889bf77b194f33dc6f711c8b83ffa27.jpg) +Noise level $n = 300$ + +![](images/6ff236f5a9a176685bfc0668fb391526f76f364887b7b54f9506578af19a934a.jpg) +Figure 7. Two visual examples of video enhancement results by using different noise levels in GenVE. + +![](images/3a051011b9fa37183ec748d16f6fb9f1f5f7462c6798ffe9beb58a6d4d09cc40.jpg) +Input +Figure 8. Two visual examples of video enhancement results guided by two different image references in GenVE. + +![](images/d7ce51d43d7d90cfd59231c2aac6827c4ebd1c06c98853e91f62db660f697f1a.jpg) + +![](images/a0be5c3b7156d2ea267b2623ed8d14eced0af4782a39a12df90faa945e7bae81.jpg) +Reference-1 +Result-1 + +![](images/91f5cb132ad848bb494bd60a405e7308e186c2a6c1bdefbcc85a36b2a0c239e8.jpg) + +![](images/4b600df0706e2a9ad4a1ddf0a7a6b2ba4c7db9c6b33950f6e6308d6d09523561.jpg) +Reference-2 + +![](images/e6042350be53c431e2411ab33632c5dad2d2c802fe2486ee1b55507a5e60be8c.jpg) + +![](images/d6cf06ef4a5520532b7b2b25cb9874cc2c293d8a94c96384d8d19f59070d2940.jpg) +Result-2 + +final video enhancement results are perceptually similar. We speculate that the results are contributed by the augmentation strategy of masking condition which encourages GenVE to search multiple visual cues in image reference holistically, enhancing the generation of local details. + +# 5. Conclusions + +We have presented GenVE that explores the image reference to boost diffusion-based video enhancement. In particular, we study the problem from the view point of leveraging image reference as the bridge for semantic and texture alignment between input low-quality video and generated high-quality video. To consolidate our idea, GenVE first exploits an image diffusion model to upscale a key frame of input video for semantics-aligned high-quality image reference generation. A video controller then transfers the patch-level texture details in high-quality image reference to the correspondence region of low-quality video for fine-grained details enhancement. In the training of GenVE, a series of conditioning augmentation strategies is conducted to further strengthen the model robustness. Experiments conducted on two public real-world video datasets, i.e., YouHQ40 and VideoLQ, as well as a self-built AI-synthesized video dataset AIGC-Vid, validate the effectiveness of the proposed GenVE for video enhancement. + +Acknowledgements. This work was supported in part by the Beijing Municipal Science and Technology Project No. Z241100001324002, and Beijing Nova Program No. 20240484681, and the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC. + +# References + +[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 3 +[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime Video Super-Resolution with Spatio-Temporal Networks and Motion Compensation. In CVPR, 2017. 2 +[3] Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer. arXiv:2505.22705, 2025. 2 +[4] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based Image Super-Resolution with Deformable Attention Transformer. In ECCV, 2022. 5, 6 +[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond. In CVPR, 2021. 2 +[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving Video SuperResolution with Enhanced Propagation and Alignment. In CVPR, 2022. 2 +[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating Tradeoffs in Real-World Video Super-Resolution. In CVPR, 2022. 5 +[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 5 +[9] Jingyuan Chen, Fuchen Long, Jie An, Zhaofan Qiu, Ting Yao, Jiebo Luo, and Tao Mei. Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion. In AAAI, 2025. 2 +[10] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wenggang Zhou, Jiebo Luo, and Tao Mei. Learning 3D Shape Latent for Point Cloud Completion. IEEE Trans. on Multimedia, 2024. +[11] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution. In CVPR, 2024. +[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models. In ICCV, 2021. +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2 +[14] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. VEnhancer: Generative Space-Time Enhancement for Video Generation. arXiv:2407.07667, 2024. 1, 2, 5, 6 +[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 2 + +[16] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS Workshop, 2021. 3 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 3 +[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv:2210.02303, 2022. 1, 3 +[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. In NeurIPS, 2022. 3 +[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive Benchmark Suite for Video Generative Models. In CVPR, 2024. 5 +[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video Super-Resolution with Recurrent Structure-Detail Network. In ECCV, 2020. 2 +[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-Scale Image Quality Transformer. In ICCV, 2021. 6 +[23] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-Correspondence Aggregation Network for Video Super-Resolution. In ECCV, 2020. 2 +[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A Video Restoration Transformer. IEEE TIP, 2022. 2 +[25] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent Video Restoration Transformer with Guided Deformable Attention. In NeurIPS, 2022. 2 +[26] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022. 2 +[27] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2 +[28] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Bi-calibration Networks for Weakly-Supervised Video Representation Learning. IJCV, 2023. 2 +[29] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. VideoStudio: Generating Consistent-Content and Multi-Scene Videos. In ECCV, 2024. 2 +[30] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards Photorealistic Image Generation and Editing with Text-guided Diffusion Models. In ICML, 2022. 2 +[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 5 + +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 2, 6 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 2 +[34] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking Alignment in Video Super-Resolution Transformers. In NeurIPS, 2022. 2 +[35] Gyumin Shim, Jinsun Park, and In So Kweon. Robust Reference-based Super-Resolution with Similarity-Aware Deformable Convolution. In CVPR, 2020. 2 +[36] Zachary Teed and Jia Deng. RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In ECCV, 2020. 4 +[37] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring CLIP for Assessing the Look and Feel of Images. In AAAI, 2023. 6 +[38] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind SuperResolution with Pure Synthetic Data. In ICCV, 2021. 5 +[39] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models. IJCV, 2024. 1, 2, 5, 6 +[40] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing wen Hou, Annan Wang, Qiong Yan Wenxiu Sun, and Weisi Lin. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. In ICCV, 2023. 6 +[41] Yanchun Xie, Jimin Xiao, Mingjie Sun, Chao Yao, and Kaizhu Huang. Feature Representation Matters: End-to-End Learning for Reference-based Image Super-resolution. In ECCV, 2020. 2 +[42] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal Modulation Network for Controllable Space-Time Video Super-Resolution. In CVPR, 2021. 2 +[43] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T. Freeman. Video Enhancement with Task-Oriented Flow. IJCV, 2019. 2 +[44] Xu Yan, Weibing Zhao, Kun Yuan, Ruimao Zhang, Zhen Li, and Shuguang Cui. Towards Content-Independent Multi-Reference Super-Resolution: Adaptive Pattern Matching and Feature Aggregation. In ECCV, 2020. 2 +[45] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning Texture Transformer Network for Image Super-Resolution. In CVPR, 2020. 2, 3 +[46] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations. In ICCV, 2019. 2 + +[47] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient Video Super-Resolution. In ICCV, 2021. 2 +[48] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Landmark Image Super-resolution by Retrieving Web Images. IEEE TIP, 2013. 2 +[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 7 +[50] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv:2311.04145, 2023. 1, 2, 3, 5 +[51] Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. Image Super-Resolution by Neural Texture Transfer. In CVPR, 2019. 2, 3 +[52] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models. In CVPR, 2024. 2 +[53] Zhongwei Zhang, Fuchen Long, Zhaofan Qiu, Yingwei Pan, Wu Liu, Ting Yao, and Tao Mei. MotionPro: A Precise Motion Controller for Image-to-Video Generation. In CVPR, 2025. 2 +[54] Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, and Lu Fang. CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping. In ECCV, 2018. 2 +[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video SuperResolution. In CVPR, 2024. 1, 2, 5, 6 \ No newline at end of file diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/images.zip b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7c5bfd14f5f074cd3012cfa755ce9ef680afc5f8 --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56da09ded87a7fd1c6d4e7e649b01b474f94e8f59b4ea5b3de98e467f4083a1d +size 827334 diff --git a/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/layout.json b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1ee7e06d117938957cbbcc1684c685cd94db6b1f --- /dev/null +++ b/2025/Aligning Global Semantics and Local Textures in Generative Video Enhancement/layout.json @@ -0,0 +1,11655 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 559, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 559, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 559, + 122 + ], + "type": "text", + "content": "Aligning Global Semantics and Local Textures in Generative Video Enhancement*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": "Zhikai Chen" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", Fuchen Long" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", Zhaofan Qiu" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", Ting Yao" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", Wengang Zhou" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", Jiebo Luo" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "text", + "content": ", and Tao Mei" + }, + { + "bbox": [ + 111, + 144, + 517, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "text", + "content": "MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China " + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "text", + "content": "HiDream.ai Inc. " + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 158, + 567, + 184 + ], + "type": "text", + "content": "University of Rochester, Rochester, NY USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 138, + 187, + 495, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 187, + 495, + 199 + ], + "spans": [ + { + "bbox": [ + 138, + 187, + 495, + 199 + ], + "type": "text", + "content": "czk654@mail.ustc.edu.cn,{longfuchen, qiuzhaofan, tiyao}@hidream.ai" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 166, + 202, + 466, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 202, + 466, + 213 + ], + "spans": [ + { + "bbox": [ + 166, + 202, + 466, + 213 + ], + "type": "text", + "content": "zhwg@ustc.edu.cn, jluo@cs.ROchester.edu, tmei@hidream.ai" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 266, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 589 + ], + "type": "text", + "content": "Recent advances in video generation have demonstrated the utility of powerful diffusion models. One important direction among them is to enhance the visual quality of the AI-synthesized videos for artistic creation. Nevertheless, solely relying on the knowledge embedded in the pretrained video diffusion models might limit the generalization ability of local details (e.g., texture). In this paper, we address this issue by exploring the visual cues from a high-quality (HQ) image reference to facilitate visual details generation in video enhancement. We present GenVE, a new recipe of generative video enhancement framework that pursues the semantic and texture alignment between HQ image reference and denoised video in diffusion. Technically, GenVE first leverages an image diffusion model to magnify a key frame of the input video to attain a semantics-aligned HQ image reference. Then, a video controller is integrated into 3D-UNet to capture patch-level texture of the image reference to enhance fine-grained details generation at the corresponding region of low-quality (LQ) video. Moreover, a series of conditioning augmentation strategies are implemented for effective model training and algorithm robustness. Extensive experiments conducted on the public YouHQ40 and VideoLQ, as well as self-built AIGC-Vid dataset, quantitatively and qualitatively demonstrate the efficacy of our GenVE over the state-of-the-art video enhancement approaches. Source code is available at https://github.com/HiDream-ai/GenVE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 135, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 625, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 685 + ], + "type": "text", + "content": "With the rapid development of video diffusion models [18, 39], there is a growing demand to elevate the visual quality of the synthesized video contents. Video enhancement, as an emerging compute vision topic distinct from traditional video super-resolution, aims to refine the low" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 241, + 555, + 434 + ], + "blocks": [ + { + "bbox": [ + 320, + 241, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 320, + 241, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 320, + 241, + 555, + 434 + ], + "type": "image", + "image_path": "d20047f8fd5ed59d7b05e491665c1a6bc4afdd5a168baa24ea06a5574c265e54.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 436, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 479 + ], + "type": "text", + "content": "Figure 1. An illustration of visual alignment in our generative video enhancement (GenVE), including the global semantic alignment between LQ video and image reference, and the local texture alignment between image reference and HQ video." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 617 + ], + "type": "text", + "content": "quality (LQ) video to achieve better appearance and motion quality, adding more innovative details and endowing smoother temporal dynamics to the input video. Several studies [14, 39, 50, 55] have tackled this issue by treating it as a conditioning video diffusion paradigm. The prior knowledge of the input LQ video is exploited to regulate the video denoising procedure to produce the high-quality (HQ) counterpart. Despite having impressive results in the realm of video enhancement, such recipe still faces the challenge of generating fine-grained details (e.g., texture) due to the lack of some content prior knowledge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "Instead of solely relying on the input LQ video as the condition, we consider involving an additional knowledge connection, i.e., high-quality image reference, to bridge the gap of fine-grained details between LQ and HQ videos. One intuition behind this is that the visual quality of image is usually higher than that of video with the same resolution, which is caused by degradation brought by temporal dynamics (e.g., motion blur). Technically, we formulate our" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 220, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 220, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 220, + 703 + ], + "type": "text", + "content": "*This work was performed at HiDream.ai." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 154, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 154, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 154, + 712 + ], + "type": "text", + "content": "† Co-corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17087" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "reference-guided video enhancement in diffusion paradigm from two perspectives: global semantic alignment and local texture alignment. As depicted in the upper part of Figure 1, global semantic alignment is achieved by transferring the information of holistic layout (e.g., the relative position of cat and cloth) of a key frame to the HQ image reference via diffusion-based image up-scaling. Next, we develop a video controller in 3D-UNet to calibrate the video patch denoising with the guidance of local texture details (e.g., the cat eye and nose) in the obtained image reference. The learning of patch-level correspondences between denoised videos and image reference are executed by a cross-attention mechanism, boosting the fine-grained contextual interaction. Moreover, we investigate a set of augmentation approaches to enhance the robust learning of our reference-guided video enhancement model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 263, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 295, + 502 + ], + "type": "text", + "content": "To materialize our idea, we present GenVE to carry out semantic and texture alignment for generative video enhancement. Technically, GenVE first utilizes an image diffusion model to create a high-quality (HQ) image reference based on a key frame of input low-quality (LQ) video through forward-backward diffusion process. In each step of backward diffusion, the image reference denoising is regulated by a weighted summation of LQ key frame features to achieve global semantic alignment. Then, both of the image reference and LQ video are fed into a video controller with the newly-devised local-aware cross-attention (LCA) to attain local texture alignment between image reference and denoised video in diffusion. In LCA, the feature patch of LQ video acts as the query while that of image reference serves as the key and value for computation. A local attention bias is further exploited to impose the effects of neighboring patches around each patch in LQ video. Furthermore, GenVE involves a series of data augmentation strategies on input conditions (i.e., LQ video and image reference) to strengthen the model robustness during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 503, + 294, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 294, + 611 + ], + "type": "text", + "content": "The main contribution of this paper is the proposed GenVE to explore semantic and texture alignment between image reference and denoised video in diffusion. The solution also leads to the elegant views of how to maintain the semantic correspondence in high-quality image reference synthesis, and how to leverage the texture information in the reference to benefit local detail generation in video denoising. Extensive experiments on three benchmarks verify the superiority of GenVE for video quality improvement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 714 + ], + "type": "text", + "content": "Traditional video super-resolution. Traditional video super-resolution (VSR) aims to up-scale video frames while keeping the fidelity, which could be categorized into two groups, i.e., sliding window-based methods and recurrent methods. Sliding window-based techniques [2, 23, 24, 42, 43, 46] rely on CNNs [28] or Transformers [26, 27] which" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "aggregate the information of a sequence of low-resolution (LR) frames to predict center high-resolution (HR) frame. One representative is VRT [24] that plugs the temporal mutual attention block into transformer backbone to facilitate motion estimation, feature alignment and fusion. In contrast to aggregate information from adjacent frames, recurrent approaches [5, 6, 21, 25, 34, 47] utilize a hidden state to sequentially propagate information from all previous frames to the current frame, benefiting the frame restoration. Despite the great performance on magnifying original visual contents in the input LR video, such regression paradigm still faces the generalization problem to produce novel fine-grained details and improve the video aesthetic quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 229, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 457 + ], + "type": "text", + "content": "Reference-based image super-resolution. There has been a series of investigation [35, 41, 44, 45, 48, 51, 54] to leverage the reference for image super-resolution (ISR). One typical way [48, 54] is to refine low-resolution (LR) image with the texture information in reference image via global visual alignment. For instance, Landmark [48] aligns the reference image with the LR image through optimizing a global registration problem and exploits the texture for visual enhancement. To fully mine the correspondence between reference image and input LR image, Zheng et al. [54] introduce cross-scale warping in CNNs to fuse features in different scales. Another direction for reference-based ISR is conducted in patch level. Recent advance TTSR [45] adopts a texture transformer to search active reference patches via cross-attention for image detail refinement. In this paper, we leverage the reference-based video enhancement recipe, and further delve into the context alignment between LQ and HQ videos from both semantic and texture perspectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 460, + 553, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 699 + ], + "type": "text", + "content": "Diffusion-based video enhancement. Significant progress of video synthesis achieved by diffusion models [3, 9-13, 15, 29, 30, 32, 52, 53] encourage the development on video enhancement. Distinct from traditional video super-resolution that focuses on fidelity preserving, diffusion-based video enhancement approaches [14, 39, 50, 55] can imagine fancy details for appearance improvement. Most of them leverage the knowledge prior embedded in the pre-trained image diffusion models and formulate the problem as conditional video diffusion. For example, Zhou et al. [55] freeze the pre-trained spatial layers of the image diffusion model [33] and train the temporal latent propagation module for video enhancement. In comparison, LaVie [39] jointly trains a spatial-temporal video diffusion network and attains better temporal consistency. Later in VEnhancer [14], ControlNet [49] is further exploited to learn videow-aware conditional features to enhance the fine-grained object synthesis in video diffusion. Nevertheless, solely depending on the knowledge in pre-trained diffusion models could limit the generalization for local details (e.g., texture)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 713 + ], + "type": "text", + "content": "In summary, our work focuses on video enhancement" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17088" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 62, + 379, + 209 + ], + "blocks": [ + { + "bbox": [ + 57, + 62, + 379, + 209 + ], + "lines": [ + { + "bbox": [ + 57, + 62, + 379, + 209 + ], + "spans": [ + { + "bbox": [ + 57, + 62, + 379, + 209 + ], + "type": "image", + "image_path": "da8cbe43d01ae59b7ae4f6a5f63444e613c20494c9edc93d947b71f0267c63dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": "Figure 2. An overview of our GenVE. The input LQ video is first up-sampled to the target resolution via bicubic interpolation. The up-sampled video is compressed as the video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " and then fed into the image diffuser to synthesize an image reference latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": ". Next, the latents " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " are converted to the augmented image reference latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Y_{n}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " and video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{n}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " as the conditional signals. Concurrently, the up-sampled video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " is noised by the Gaussian vector " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " to obtain the noisy video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{T_1}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": ". Hence, the video controller exploits the two augmented latents " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Y_{n}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{n}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": ", and the noisy video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{T_1}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " to learn a visual guidance for diffusion calibration. Finally, the HQ video latent " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{0}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " is synthesized from " + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "inline_equation", + "content": "Z_{T_1}" + }, + { + "bbox": [ + 54, + 212, + 555, + 278 + ], + "type": "text", + "content": " via video diffuser based on the learnt visual guidance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 65, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 381, + 65, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 381, + 65, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 381, + 65, + 553, + 209 + ], + "type": "image", + "image_path": "5d88722cfd29ec5b0a7c2b5f7821c9d036bc964add3a687589fc7bd10663d95e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 285, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 296, + 344 + ], + "type": "text", + "content": "using image reference in diffusion models. The proposal of GenVE contributes by investigating not only how to obtain semantics-aligned image reference through diffusion forward-backward calibration, but also how to benefit patch-level texture details synthesis in video denoising." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 122, + 370 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 122, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 122, + 370 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 376, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 591 + ], + "type": "text", + "content": "In this section, we present our newly-minted GenVE, pursuing global semantics alignment and local texture alignment between LQ and HQ videos for video enhancement. Figure 2 depicts an overview of this pipeline. Given a LQ video, we up-sample it through bicubic interpolation, and then encode it to the latent representation via VAE. Then, the image diffuser is exploited to refine a key frame from the up-sampled video to a semantics-aligned high-quality image reference. After that, the latents of image reference and up-sampled video are transformed by a series of conditioning augmentation strategies. The video controller performs local-aware cross-attention (LCA) between the feature patches of the augmented video frame latent and image reference latent for local texture alignment. These features are passed through cascaded 3D encoder blocks to learn the visual guidance. Finally, the video diffuser exploits the learnt visual guidance to calibrate the denoising procedure for video enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 259, + 612 + ], + "type": "text", + "content": "3.1. Preliminaries: Video Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": "Video diffusion models [1, 18, 19, 50] typically employ a 2D-VAE as the video encoder and a 3D-UNet as the latent denoiser. Given a video with " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " frames, the VAE first encodes it into the latent code " + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z = \\{z^i\\}_{i=1}^I" + }, + { + "bbox": [ + 54, + 617, + 295, + 713 + ], + "type": "text", + "content": " frame-by-frame. Each feature point in the latent denotes a local patch in the original frame. To conduct video refinement using diffusion models, we follow the natural solution and implement forward and backward diffusion procedures on the input video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "content": "In the forward process, the Gaussian noise is gradually added to the video latent code " + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "content": " with time step " + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "content": ", formulated by a Markov chain. The noised latent " + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "content": " at each forward diffusion step " + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 285, + 555, + 332 + ], + "type": "text", + "content": " is thus formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 398, + 340, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 340, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 398, + 340, + 553, + 354 + ], + "type": "interline_equation", + "content": "Z _ {t} = \\alpha_ {t} Z + \\sigma_ {t} \\epsilon , \\tag {1}", + "image_path": "721aa4f043da5019b00c70da0fb549da21c85027705fdce07b55225dc8d9e8ae.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "content": " is the Gaussian noise and " + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\sigma_{t}" + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "content": " specify the noise schedule of DDPM [17], leading the proportion of noise to gradually increase over the steps " + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 361, + 555, + 397 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "content": "In the backward process, the data reconstruction for refinement eliminates the added noise in the forward diffusion process, which can start from a noise sample " + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "content": " at arbitrary time step " + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "content": ". Each backward diffusion step is a denoising process formulated by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 465, + 553, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 465, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 383, + 465, + 553, + 478 + ], + "type": "interline_equation", + "content": "Z _ {t - 1} = \\left(Z _ {t} - \\tilde {\\sigma} _ {t} \\epsilon_ {t}\\right) / \\tilde {\\alpha} _ {t}, \\tag {2}", + "image_path": "0652572af3376ddfe27d1b30ddfeac42ce580cf5c02f787b8b97369fc320a986.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{\\sigma}_t" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\tilde{\\alpha}_t" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " are the constant hyper-parameters related to time step " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " based on the noise scheduler [16, 17], and " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\epsilon_t = \\epsilon_\\theta(Z_t)" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " is the estimated noise vector using the denoising network " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\epsilon_\\theta" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": ". The HQ video latent " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "Z_0 = \\{z_0^i\\}_{i=1}^I" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " is thus iteratively denoised by the backward diffusion from the noised LQ video latent " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "Z_t" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " to improve the video quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 468, + 577 + ], + "type": "text", + "content": "3.2. Global Semantic Alignment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Reference-based super-resolution [45, 51] usually exploits a semantic-related HQ image reference to guide LQ enhancement. Nevertheless, the layouts are not required to be matched between each other, leading to potential spatial misalignment of the objects. Such misalignment could limit the enhancement quality of local details. To alleviate this, we propose to generate a semantics-aligned HQ image reference for reference-based video enhancement. An image diffuser first leverages the original noisy image latent to refine denoising procedure for HQ image reference generation. The generated image reference will be further em" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17089" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "ployed to transfer the patch-level texture information into LQ video frames, guiding the final video enhancement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": "Semantics-Aligning Image Diffuser. Given the latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " of a key frame from the input video latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "Z = \\{z^i\\}_{i=1}^I" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ", we first add the Gaussian noise with time step " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "T_0" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ". Then, an image diffusion model (i.e., image diffuser in Figure 2) performs backward diffusion on the noisy frame latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "z_{T_0}" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " for image refinement. At each denoising step " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ", two intermediate latents are prepared in advance. The first is the denoised frame latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " computed through the backward diffusion, and the other is a noisy frame latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " with time steps " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": ". We refine the denoised frame latent " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " in each denoising step by leveraging the component from " + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{z}_t" + }, + { + "bbox": [ + 55, + 96, + 296, + 240 + ], + "type": "text", + "content": " through weighted summation as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 248, + 295, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 248, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 126, + 248, + 295, + 266 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {t} = \\alpha_ {t} z + \\sigma_ {t} \\epsilon , \\tag {3}", + "image_path": "8cfc17357d502e2d4c1fbb07255987f53c3af54d2a356b65fa27649349156d9c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 264, + 223, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 264, + 223, + 274 + ], + "spans": [ + { + "bbox": [ + 127, + 264, + 223, + 274 + ], + "type": "interline_equation", + "content": "z _ {t} = (1 - \\gamma_ {t}) z _ {t} + \\gamma_ {t} \\tilde {z} _ {t},", + "image_path": "cf609a0fccf1bdb3e190f8e2dd3c45ce58de29b163b083e8333e3d21891885db.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\sigma_{t}" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": " are the constant hyperparameters as mentioned in Eq.(1). Note that the parameter " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": " is defined by a cosine function on " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": ", decreasing from 0.5 to 0 monotonically with " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 282, + 296, + 413 + ], + "type": "text", + "content": " to 0. In this way, the image diffuser introduces a substantial amount of visual content from the original frame at the early denoising phase, exactly where the latent features have rich high-level visual semantics. This denoising mechanism encourages the denoised HQ image reference to align with visual semantics of the original LQ frame, while maintaining low-level texture details as more as possible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "text", + "content": "Through iterative latent denoising of image refinement, the image diffuser successfully derives the image reference latent " + }, + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 414, + 296, + 462 + ], + "type": "text", + "content": ", which is semantically aligned with the key frame and has richer texture details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 209, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 209, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 209, + 481 + ], + "type": "text", + "content": "3.3. Conditioning Augmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": "Next, we perform a series of conditioning augmentation strategies on both image reference and LQ video latent in training to simulate various scenarios, thereby boosting model robustness. The augmentation includes noise augmentation, temporal augmentation and masking condition." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": "Noise Augmentation. To enable controllable video refinement, a noise augmentation strategy is introduced on both video and image conditions. We implement the noise augmentation on the image reference latent " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": " and the video latent " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "Z = \\{z^i\\}_{i=1}^I" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": ", following the forward diffusion as Eq.(1) with " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": " time steps, thereby producing the noisy conditional latents " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{y}_n" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "Z_n = \\{z_n^i\\}_{i=1}^I" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": ". The noise augmentation set the noise level " + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 546, + 295, + 677 + ], + "type": "text", + "content": " different at each training step, corrupting the low-level visual information. Hence, the noise level can be adjusted in inference to balance the visual fidelity and quality, thus obtaining the most realistic videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Temporal Augmentation. To further improve the video temporal coherence, the temporal augmentation is applied to incorporate motion clues in the original LQ video into the" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 67, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 318, + 67, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 318, + 67, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 318, + 67, + 553, + 190 + ], + "type": "image", + "image_path": "b20d5d13e15b01f53bf1629c23a572bcdefa03763b84244101e9c93632403c3a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": "Figure 3. Illustration of Local-Aware Cross-Attention (LCA). Given the augmented video frame latent " + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "inline_equation", + "content": "z_{n}^{i}" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": " and image reference latent " + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "inline_equation", + "content": "y_{n}^{i}" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": ", LCA first extracts the feature patches using two convolutional layers, respectively. Then, the feature patches of image reference latent are reshuffled within each neighboring " + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": " patches. By treating video feature patches as query and the image reference feature patches as key/value, the attention map is computed and further added by an extra local attention bias " + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": ", to estimate the output texture-aligned feature embedding " + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 191, + 555, + 290 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": "HQ video diffusion process. We first estimate the optical flows using RAFT model [36] on the input LQ video " + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": ". Then, the attained flows are utilized to warp the noisy image reference latent " + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\tilde{y}_n" + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": ". As such, it yields a set of corresponding warped image references latents " + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "inline_equation", + "content": "Y_{n} = \\{\\tilde{y}_{n}^{i}\\}_{i = 1}^{I}" + }, + { + "bbox": [ + 313, + 294, + 554, + 376 + ], + "type": "text", + "content": ", reinforcing the motion alignment between the warped image references and the input LQ video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": "Masking Condition. Moreover, we leverage the masking operation on the warped image reference latents to facilitate the model training. The masking procedure is conducted on each latent " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{y}_n^i" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": " of warped image references using random mask ratio " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": ", generating the masked noise image reference latent " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "y_{n}^{i}" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": ". During training, the masking condition will set a portion of the original input pixels to zeros. In inference, we maintain the whole input " + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\tilde{y}_n^i" + }, + { + "bbox": [ + 313, + 378, + 554, + 520 + ], + "type": "text", + "content": " to ensure that the entire image reference data is available for the optimal result. As such, the masking stimulates the model to refer to more potential texture feature patches in the image reference for visual enhancement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "content": "By the sequential conditioning augmentation, we obtain the augmented video latent " + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "inline_equation", + "content": "Z_{n} = \\{z_{n}^{i}\\}_{i = 1}^{I}" + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "content": " and the sequence of augmented image reference latents " + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "inline_equation", + "content": "Y_{n} = \\{y_{n}^{i}\\}" + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 521, + 554, + 582 + ], + "type": "text", + "content": ", respectively. These two latents will be further exploited for visual guidance learning in video controller." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 454, + 601 + ], + "type": "text", + "content": "3.4. Local Texture Alignment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "To encourage the video diffuser to generate more details for video enhancement, we leverage a video controller and further design a local-aware cross-attention (LCA) module to learn texture-aligned feature embedding. The local-aware cross attention concentrates on matching texture features of similar spatial regions between image reference and input LQ video frame, facilitating more efficient and precise feature selection. The texture-aligned feature embedding produced by LCA is then utilized for visual guidance learning." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17090" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 76, + 552, + 205 + ], + "blocks": [ + { + "bbox": [ + 64, + 62, + 544, + 73 + ], + "lines": [ + { + "bbox": [ + 64, + 62, + 544, + 73 + ], + "spans": [ + { + "bbox": [ + 64, + 62, + 544, + 73 + ], + "type": "text", + "content": "Table 1. Performance comparisons in terms of MUSIQ, DOVER and CLIP-IQA on the YouHQ40, VideoLQ and AIGC-Vid datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 76, + 552, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 552, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 552, + 205 + ], + "type": "table", + "html": "
DatasetsMetricsBicubicDATSR[4]LaVie-SR[39]Upscale-A-Video[55]VEnhancer[14]GenVE
YouHQ40MUSIQ ↑48.2149.2264.1763.6968.9969.93
DOVER ↑40.7630.0479.6474.0677.0883.50
CLIP-IQA ↑0.3920.4180.5810.5090.5940.596
VideoLQMUSIQ ↑59.5359.1364.5060.9066.1968.23
DOVER ↑50.9432.5666.9960.3871.4774.30
CLIP-IQA ↑0.3520.3960.4590.3800.4540.464
AIGC-VidMUSIQ ↑60.3357.4863.7160.0064.3064.56
DOVER ↑55.7357.2676.8474.1876.4677.02
CLIP-IQA ↑0.6010.5870.6560.6220.6340.658
", + "image_path": "a639f7a61ee54f4c74a685ef392f87fc1e45e4845be57b2074dc9a26fdf54cbb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 233 + ], + "type": "text", + "content": "Finally, the learnt visual guidance steers video denoising in video diffuser for video enhancement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": "Texture-Aligning Video Controller. The right part of Figure 2 depicts the pipeline of our video controller. The video controller leverages the LCA module for texture alignment, and the cascaded encoder blocks inherited from the 3D-UNet for feature enhancement. Specifically, the video controller takes three inputs: the augmented video latent " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "Z_{n}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": ", the sequence of augmented image reference latents " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "Y_{n}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": ", and the denoised video latent " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "Z_{t}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " that is initialized from the noised " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "Z_{T_1}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " with time step " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": ". The LCA module, as illustrated in Figure 3, first extracts the image feature patches from the image reference latent " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "y_{n}^{i}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " and video frame latent " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "z_{n}^{i}" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " using convolutional layers, and then flatten them into patch sequences. Next, LCA executes local patch reshuffling on the patch sequences to boost model training. In detail, LCA randomly reshuffles the patch order within each neighboring " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " patches. The image reference feature patches serve as the keys " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": ", and the video frame feature patches are the queries " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " for attention computation. Particularly, we add a local attention bias matrix " + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 236, + 296, + 486 + ], + "type": "text", + "content": " on the attention weights before the Softmax operation. Thus, the attention mechanism can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 495, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 495, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 61, + 495, + 295, + 544 + ], + "type": "interline_equation", + "content": "B = \\left( \\begin{array}{c c c} b _ {1 1} & b _ {1 2} & \\dots \\\\ b _ {2 1} & b _ {2 2} & \\dots \\\\ \\vdots & \\vdots & \\ddots \\end{array} \\right), \\quad b _ {i j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} | i - j | < d \\\\ 0, & \\text {i f} | i - j | \\geq d \\end{array} \\right., \\tag {4}", + "image_path": "c979550e13e70a21adb285cc8c5c98a0a49bbc797b6852421ca9d94e22c6f15e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 541, + 212, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 541, + 212, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 541, + 212, + 555 + ], + "type": "interline_equation", + "content": "\\tilde {Q} = \\operatorname {N o r m} (Q), \\quad \\tilde {K} = \\operatorname {N o r m} (K),", + "image_path": "2cfaab780aee9516fe5fe1f884852b69217e9bdeebe2b3314276d10ce7746819.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 558, + 173, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 173, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 173, + 586 + ], + "type": "interline_equation", + "content": "A = \\operatorname {S o f t m a x} (\\frac {\\tilde {Q} \\tilde {K} ^ {T} + B}{\\sqrt {d _ {k}}}),", + "image_path": "495f2fa9a6075aae21b16b8b7c3abe8df11cc2c85b9eb67c53f8f9db590ee4a8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " equals to the local reshuffling range and " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " is the channel dimension. As such, the local patch alignment mainly focuses the reference texture feature learning located within a radius of " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ", thus facilitating precise transferring of high-quality texture patterns. After that, we feed the output attention features into another convolutional layer and add it to the query features as the texture-aligned feature embedding " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ", which is formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 129, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 129, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "e = \\operatorname {C o n v} (A V) + Q. \\tag {5}", + "image_path": "e0c03bf394401a03df92a86c70b69fdb6e7ae802d6ed333f1b21b71347fadb12.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 258 + ], + "type": "text", + "content": "The texture-aligned feature embedding is added to the output video features of the following encoder block, and further propagated to the subsequent cascaded blocks for multi-scale visual guidance learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 307 + ], + "type": "text", + "content": "Finally, we collect all learnt visual guidance features output from the video controller at different scales, to regulate the multi-step backward diffusion in 3D-UNet until obtaining the final output high-quality video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 395, + 331 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 441, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 441, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 441, + 350 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 545 + ], + "type": "text", + "content": "Datasets. We evaluate the effectiveness of GenVE on three datasets, i.e., the public YouHQ [55] and VideoLQ [7], as well as the self-built AIGC-Vid. The YouHQ dataset comprises 38K high-quality videos for training and 40 high-quality videos for testing (namely YouHQ40), with approximately 30 frames per video. We empirically apply the multi-tier degradation model [38] to reduce the visual quality and obtain the high-low quality video pairs for training and validation. The VideoLQ dataset [7] is a real-world collection containing 50 low-quality videos for evaluation. There are 100 frames for each video. Additionally, we build the AIGC-Vid dataset, which is generated by three text-to-video advances, i.e., ModelScope [50], LaVie [39] and VideoCrafter [8]. We exploit each model to generate 93 videos based on the text prompts from VBench [20], and finally collect 279 AI-synthesized videos for evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. We implement GenVE on the PyTorch platform. The image diffuser and the video diffuser is built on the pre-trained image diffusion model SDXL [31] and video diffusion model I2VGen-XL [50], respectively. The time steps " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " for the base video diffusion model in training is set as 1000. The denoising time steps " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "T_{0}" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " for image reference generation is 700, and " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " in video enhancement is 900, respectively. In training, the parameter " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " for noise augmentation is an integer randomly sampled within the range of [1, 300]. We empirically set it to 250 in inference. The masking ratio " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " is randomly sampled from [0, 1) in training and set to 0 in inference. We set the local reshuffling range " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " of LCA to 7, and define the parameter " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": " in image diffuser as " + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_{t} = \\frac{1}{2}\\left[\\frac{1}{2}(1 + \\cos\\left(\\frac{T - t}{T}\\pi\\right))\\right]^{3}" + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": ". Our" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17091" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 66, + 179, + 127 + ], + "blocks": [ + { + "bbox": [ + 74, + 66, + 179, + 127 + ], + "lines": [ + { + "bbox": [ + 74, + 66, + 179, + 127 + ], + "spans": [ + { + "bbox": [ + 74, + 66, + 179, + 127 + ], + "type": "image", + "image_path": "c96baac8e4c277f25b6a2f21601466d29d9e80bc4edc422dab981b04cb28a951.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 74, + 128, + 179, + 188 + ], + "blocks": [ + { + "bbox": [ + 74, + 128, + 179, + 188 + ], + "lines": [ + { + "bbox": [ + 74, + 128, + 179, + 188 + ], + "spans": [ + { + "bbox": [ + 74, + 128, + 179, + 188 + ], + "type": "image", + "image_path": "870d126a3c940aa73b3a31c2090c8d89bbc776216e2cbcd72c0afc4346cbc648.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 74, + 189, + 179, + 248 + ], + "blocks": [ + { + "bbox": [ + 74, + 189, + 179, + 248 + ], + "lines": [ + { + "bbox": [ + 74, + 189, + 179, + 248 + ], + "spans": [ + { + "bbox": [ + 74, + 189, + 179, + 248 + ], + "type": "image", + "image_path": "b54112942778eb6efaa101eb2226dae3f3cac3bc509f56fc227210ec21e1057b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 74, + 249, + 179, + 313 + ], + "blocks": [ + { + "bbox": [ + 74, + 249, + 179, + 313 + ], + "lines": [ + { + "bbox": [ + 74, + 249, + 179, + 313 + ], + "spans": [ + { + "bbox": [ + 74, + 249, + 179, + 313 + ], + "type": "image", + "image_path": "f3def86ccbc95d30516b1ab3bb62911ea8e0d65bf165b99034763e9d6038b7f9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 313, + 155, + 320 + ], + "lines": [ + { + "bbox": [ + 97, + 313, + 155, + 320 + ], + "spans": [ + { + "bbox": [ + 97, + 313, + 155, + 320 + ], + "type": "text", + "content": "Input Video Frame" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 322, + 553, + 343 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 553, + 343 + ], + "type": "text", + "content": "Figure 4. Four visual examples of video enhancement results by different approaches on the real video data from YouHQ40 and VideoLQ. The region in the red box is presented in the zoom-in view for comparison." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 185, + 67, + 253, + 126 + ], + "blocks": [ + { + "bbox": [ + 185, + 67, + 253, + 126 + ], + "lines": [ + { + "bbox": [ + 185, + 67, + 253, + 126 + ], + "spans": [ + { + "bbox": [ + 185, + 67, + 253, + 126 + ], + "type": "image", + "image_path": "e35e3a8f7844b368abe3d492d2721ef0143c95629f387dc8510948e09c9f7a5f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 185, + 128, + 253, + 186 + ], + "blocks": [ + { + "bbox": [ + 185, + 128, + 253, + 186 + ], + "lines": [ + { + "bbox": [ + 185, + 128, + 253, + 186 + ], + "spans": [ + { + "bbox": [ + 185, + 128, + 253, + 186 + ], + "type": "image", + "image_path": "ddaddacb3748ab57eb2d2d250bdde4f224d78e191b7a546bcf72bc4095c16999.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 185, + 189, + 253, + 247 + ], + "blocks": [ + { + "bbox": [ + 185, + 189, + 253, + 247 + ], + "lines": [ + { + "bbox": [ + 185, + 189, + 253, + 247 + ], + "spans": [ + { + "bbox": [ + 185, + 189, + 253, + 247 + ], + "type": "image", + "image_path": "762a2677c4ec4686b261d2880af3669f4d9860c8855acee62fbb70190ab31732.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 185, + 251, + 253, + 311 + ], + "blocks": [ + { + "bbox": [ + 185, + 251, + 253, + 311 + ], + "lines": [ + { + "bbox": [ + 185, + 251, + 253, + 311 + ], + "spans": [ + { + "bbox": [ + 185, + 251, + 253, + 311 + ], + "type": "image", + "image_path": "c6a39d0284fd16416b36660644e25b2c483575d523dc04c03ecd768c193b575a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 312, + 230, + 319 + ], + "lines": [ + { + "bbox": [ + 209, + 312, + 230, + 319 + ], + "spans": [ + { + "bbox": [ + 209, + 312, + 230, + 319 + ], + "type": "text", + "content": "DATSR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 255, + 67, + 323, + 126 + ], + "blocks": [ + { + "bbox": [ + 255, + 67, + 323, + 126 + ], + "lines": [ + { + "bbox": [ + 255, + 67, + 323, + 126 + ], + "spans": [ + { + "bbox": [ + 255, + 67, + 323, + 126 + ], + "type": "image", + "image_path": "a6f83e9413ec207f55a715d59cb36ee3a257f617bfe5fc6056aeec3b10e14194.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 255, + 128, + 323, + 186 + ], + "blocks": [ + { + "bbox": [ + 255, + 128, + 323, + 186 + ], + "lines": [ + { + "bbox": [ + 255, + 128, + 323, + 186 + ], + "spans": [ + { + "bbox": [ + 255, + 128, + 323, + 186 + ], + "type": "image", + "image_path": "c5ca676453211d34c1e0366f777dd81b04c79b0016c463113d72236b89d8e8d6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 255, + 189, + 323, + 247 + ], + "blocks": [ + { + "bbox": [ + 255, + 189, + 323, + 247 + ], + "lines": [ + { + "bbox": [ + 255, + 189, + 323, + 247 + ], + "spans": [ + { + "bbox": [ + 255, + 189, + 323, + 247 + ], + "type": "image", + "image_path": "b6d6ed5b45c3167d5d4aa755ea3cceda27230a3d97fd02a482c20f57d36488d7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 255, + 251, + 323, + 311 + ], + "blocks": [ + { + "bbox": [ + 255, + 251, + 323, + 311 + ], + "lines": [ + { + "bbox": [ + 255, + 251, + 323, + 311 + ], + "spans": [ + { + "bbox": [ + 255, + 251, + 323, + 311 + ], + "type": "image", + "image_path": "da62cc89107ec116e891321beaaa2f61a75576cfa9a8690ab6ba37c0ed9554e7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 312, + 301, + 319 + ], + "lines": [ + { + "bbox": [ + 274, + 312, + 301, + 319 + ], + "spans": [ + { + "bbox": [ + 274, + 312, + 301, + 319 + ], + "type": "text", + "content": "LaVie-SR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 324, + 67, + 393, + 126 + ], + "blocks": [ + { + "bbox": [ + 324, + 67, + 393, + 126 + ], + "lines": [ + { + "bbox": [ + 324, + 67, + 393, + 126 + ], + "spans": [ + { + "bbox": [ + 324, + 67, + 393, + 126 + ], + "type": "image", + "image_path": "fd0745d020e0cbccd6e6193c90cc506e7355cd590418f94e35a83d450117bd57.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 324, + 128, + 393, + 186 + ], + "blocks": [ + { + "bbox": [ + 324, + 128, + 393, + 186 + ], + "lines": [ + { + "bbox": [ + 324, + 128, + 393, + 186 + ], + "spans": [ + { + "bbox": [ + 324, + 128, + 393, + 186 + ], + "type": "image", + "image_path": "05875dcb0cb3583975ead1a20ece100761bea6207a08c2a14e7e68b4f8695c7b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 324, + 189, + 393, + 247 + ], + "blocks": [ + { + "bbox": [ + 324, + 189, + 393, + 247 + ], + "lines": [ + { + "bbox": [ + 324, + 189, + 393, + 247 + ], + "spans": [ + { + "bbox": [ + 324, + 189, + 393, + 247 + ], + "type": "image", + "image_path": "e3fc2eaa0139bcd702c38e2dd9c2c03198b8d43e140c6ab78efb138d1098cb98.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 324, + 251, + 393, + 311 + ], + "blocks": [ + { + "bbox": [ + 324, + 251, + 393, + 311 + ], + "lines": [ + { + "bbox": [ + 324, + 251, + 393, + 311 + ], + "spans": [ + { + "bbox": [ + 324, + 251, + 393, + 311 + ], + "type": "image", + "image_path": "14a278d7790969eb4ee783f714aa20185f4ef529a26f265507b9fdbb93a14183.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 312, + 384, + 320 + ], + "lines": [ + { + "bbox": [ + 334, + 312, + 384, + 320 + ], + "spans": [ + { + "bbox": [ + 334, + 312, + 384, + 320 + ], + "type": "text", + "content": "Upscale-A-Video" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 67, + 463, + 126 + ], + "blocks": [ + { + "bbox": [ + 395, + 67, + 463, + 126 + ], + "lines": [ + { + "bbox": [ + 395, + 67, + 463, + 126 + ], + "spans": [ + { + "bbox": [ + 395, + 67, + 463, + 126 + ], + "type": "image", + "image_path": "10ea8beebca93c7eb750feefc65488c16a89be6bdd2879c4bd38ce71b6104bd0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 395, + 128, + 463, + 186 + ], + "blocks": [ + { + "bbox": [ + 395, + 128, + 463, + 186 + ], + "lines": [ + { + "bbox": [ + 395, + 128, + 463, + 186 + ], + "spans": [ + { + "bbox": [ + 395, + 128, + 463, + 186 + ], + "type": "image", + "image_path": "ab3675295a97807d3388168035aa04da5578cf331e29ece1720e91020e37be49.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 395, + 189, + 463, + 247 + ], + "blocks": [ + { + "bbox": [ + 395, + 189, + 463, + 247 + ], + "lines": [ + { + "bbox": [ + 395, + 189, + 463, + 247 + ], + "spans": [ + { + "bbox": [ + 395, + 189, + 463, + 247 + ], + "type": "image", + "image_path": "e724dbc6684edbbff920540146e70534790cd794c307038eac4c8e266226967f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 395, + 251, + 464, + 311 + ], + "blocks": [ + { + "bbox": [ + 395, + 251, + 464, + 311 + ], + "lines": [ + { + "bbox": [ + 395, + 251, + 464, + 311 + ], + "spans": [ + { + "bbox": [ + 395, + 251, + 464, + 311 + ], + "type": "image", + "image_path": "b697ef6c3a0f456a9095fd88df5b40b012fcbd286ae439f9a1f2e0be0501f1d9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 312, + 446, + 319 + ], + "lines": [ + { + "bbox": [ + 412, + 312, + 446, + 319 + ], + "spans": [ + { + "bbox": [ + 412, + 312, + 446, + 319 + ], + "type": "text", + "content": "VEnhancer" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 465, + 67, + 534, + 126 + ], + "blocks": [ + { + "bbox": [ + 465, + 67, + 534, + 126 + ], + "lines": [ + { + "bbox": [ + 465, + 67, + 534, + 126 + ], + "spans": [ + { + "bbox": [ + 465, + 67, + 534, + 126 + ], + "type": "image", + "image_path": "c17ee9b0d99b369944eab2045dc1d7a4cec7cc73be4af7b3f20cc27351268d73.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 465, + 128, + 534, + 186 + ], + "blocks": [ + { + "bbox": [ + 465, + 128, + 534, + 186 + ], + "lines": [ + { + "bbox": [ + 465, + 128, + 534, + 186 + ], + "spans": [ + { + "bbox": [ + 465, + 128, + 534, + 186 + ], + "type": "image", + "image_path": "89c2563c9e299a33dc3b9792a8fe6869a57cf5f20f0cbf100aa03994850d9251.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 465, + 189, + 534, + 247 + ], + "blocks": [ + { + "bbox": [ + 465, + 189, + 534, + 247 + ], + "lines": [ + { + "bbox": [ + 465, + 189, + 534, + 247 + ], + "spans": [ + { + "bbox": [ + 465, + 189, + 534, + 247 + ], + "type": "image", + "image_path": "b67e91160c03b8c38b389f2fc1302439e54d602153a7ae6933da0a2d3ffe6e21.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 465, + 251, + 534, + 311 + ], + "blocks": [ + { + "bbox": [ + 465, + 251, + 534, + 311 + ], + "lines": [ + { + "bbox": [ + 465, + 251, + 534, + 311 + ], + "spans": [ + { + "bbox": [ + 465, + 251, + 534, + 311 + ], + "type": "image", + "image_path": "19c320db6a07206dd0a535a8ec668af9d282d7f72b63730e999a0d6755041cdb.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 312, + 510, + 320 + ], + "lines": [ + { + "bbox": [ + 488, + 312, + 510, + 320 + ], + "spans": [ + { + "bbox": [ + 488, + 312, + 510, + 320 + ], + "type": "text", + "content": "GenVE" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 346, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 393 + ], + "type": "text", + "content": "GenVE is optimized using AdamW optimizer with the base learning rate " + }, + { + "bbox": [ + 55, + 346, + 295, + 393 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 346, + 295, + 393 + ], + "type": "text", + "content": " and batch size 8. The key frame for the image reference generation is set as the first frame of the input video by cross validation." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 55, + 396, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 551 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous video enhancement advances [14, 55], we conduct the evaluation on video quality using three metrics, i.e., MUSIQ [22], DOVER [40], CLIP-IQA [37]. All the three metrics estimate the perceptual scores on input video frames using the pre-trained deep models for quality assessment. In detail, MUSIQ assesses the frame-wise visual quality via a pretrained image transformer. Similarly, DOVER exploits a video transformer to estimate video quality score in terms of aesthetics and technical distortion. CLIP-IQA is determined by the feature cosine similarity between the generated frames and text prompts (e.g., good quality) via CLIP model [32] for quality perception." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 562, + 285, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 285, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 285, + 574 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 640 + ], + "type": "text", + "content": "We compare our GenVE with several state-of-the-art techniques, including the reference-based SR approach DATSR [4], and diffusion-based video enhancement approaches, i.e., LaVie-SR [39], Upscale-A-Video [55], VEnhancer [14] on YouHQ40, VideoLQ and AIGC-Vid." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. Table 1 summarizes the performances of different approaches in terms of the three metrics over three datasets. Overall, our GenVE achieves the best performances across all metrics on both real video data and AI-synthesized video data. In general, higher MUSIQ and CLIP-IQA scores indicate the better visual percep" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 346, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 501 + ], + "type": "text", + "content": "tual quality on the aesthetic aspect. Specifically, GenVE achieves the MUSIQ score of 69.93 and the CLIP-IQA score of 0.596 on YouHQ40, outperforming the competitor VEnhancer. Compared to VEnhancer that solely exploits the visual guidance from LQ videos, ours further involves the semantics-aligned HQ image reference as the bridge to enhance the patch-wise texture alignment in video enhancement, attaining better aesthetic quality. Similar performance trend can be also observed on AIGC-Vid with the AI-synthesized videos. Such performances also demonstrate our model capability to handle complex video degradation scenarios that are brought by the basic video generation model in text-to-video diffusion." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Qualitative Comparison. Figure 4 showcases four video enhancement examples from real data in YouHQ40 and VideoLQ. Compared to other baselines, GenVE successfully synthesizes fine-grained details (e.g., the clear house windows in the 1-st and 2-nd case) with better visual appearance. Even for the extremely degraded video (e.g., the human face in the 4-th case), GenVE can still capture the major structural contents and imagine the reasonable local parts. The results demonstrate the advantage of exploring comprehensive information of high-quality image reference to facilitate spatial details synthesis in video diffusion model. We further visualize four video enhancement cases of the AI generated videos in Figure 5 to verify the model generalization ability. As depicted in the figure, there exists some blur (the guitar in the 3rd case) in the result videos by Upscale-A-Video. The local structures of the guitar frets in 1-st case generated by VEnhancer are also unorganized." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17092" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 64, + 168, + 134 + ], + "blocks": [ + { + "bbox": [ + 58, + 64, + 168, + 134 + ], + "lines": [ + { + "bbox": [ + 58, + 64, + 168, + 134 + ], + "spans": [ + { + "bbox": [ + 58, + 64, + 168, + 134 + ], + "type": "image", + "image_path": "e07dff65030d348a5d033fa00092f872d160550788dc4d7f431836f0784ed692.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 135, + 168, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 135, + 168, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 135, + 168, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 135, + 168, + 205 + ], + "type": "image", + "image_path": "83cabc7bc58e42c047e2245f91a42fdd7d235f129d02a33e3fef9e048f2f9810.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 206, + 168, + 275 + ], + "blocks": [ + { + "bbox": [ + 58, + 206, + 168, + 275 + ], + "lines": [ + { + "bbox": [ + 58, + 206, + 168, + 275 + ], + "spans": [ + { + "bbox": [ + 58, + 206, + 168, + 275 + ], + "type": "image", + "image_path": "c02700f37fa3c7f444f0ca137d9026101d5cd236e43f46c0092a3cfe987a6ac7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 276, + 168, + 346 + ], + "blocks": [ + { + "bbox": [ + 58, + 276, + 168, + 346 + ], + "lines": [ + { + "bbox": [ + 58, + 276, + 168, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 276, + 168, + 346 + ], + "type": "image", + "image_path": "e39cb670fd29f5ad6ff9b61b4c69fda9c67942d116e413763bcff152a17e0b4d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 347, + 143, + 355 + ], + "lines": [ + { + "bbox": [ + 83, + 347, + 143, + 355 + ], + "spans": [ + { + "bbox": [ + 83, + 347, + 143, + 355 + ], + "type": "text", + "content": "Input Video Frame" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 171, + 68, + 246, + 131 + ], + "blocks": [ + { + "bbox": [ + 171, + 68, + 246, + 131 + ], + "lines": [ + { + "bbox": [ + 171, + 68, + 246, + 131 + ], + "spans": [ + { + "bbox": [ + 171, + 68, + 246, + 131 + ], + "type": "image", + "image_path": "a5f84acd4790a0144ff7b123460adaa6d4f10ca93eeb77d94c5867cb5023c052.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 171, + 137, + 246, + 201 + ], + "blocks": [ + { + "bbox": [ + 171, + 137, + 246, + 201 + ], + "lines": [ + { + "bbox": [ + 171, + 137, + 246, + 201 + ], + "spans": [ + { + "bbox": [ + 171, + 137, + 246, + 201 + ], + "type": "image", + "image_path": "f28e93856f2e262dd626ecc2a9201712a2282dbb61ad7dec0f804c62f7f32e4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 171, + 209, + 246, + 272 + ], + "blocks": [ + { + "bbox": [ + 171, + 209, + 246, + 272 + ], + "lines": [ + { + "bbox": [ + 171, + 209, + 246, + 272 + ], + "spans": [ + { + "bbox": [ + 171, + 209, + 246, + 272 + ], + "type": "image", + "image_path": "954b48783b3887052439b66fe3af1676b63299385dedee71c11223f7e3f60952.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 171, + 280, + 246, + 343 + ], + "blocks": [ + { + "bbox": [ + 171, + 280, + 246, + 343 + ], + "lines": [ + { + "bbox": [ + 171, + 280, + 246, + 343 + ], + "spans": [ + { + "bbox": [ + 171, + 280, + 246, + 343 + ], + "type": "image", + "image_path": "9927edb4de5cfab93d4921aeb76c94956f28fa8886ca652e22ed6039b3bd4b8d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 347, + 220, + 355 + ], + "lines": [ + { + "bbox": [ + 198, + 347, + 220, + 355 + ], + "spans": [ + { + "bbox": [ + 198, + 347, + 220, + 355 + ], + "type": "text", + "content": "DATSR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 248, + 68, + 323, + 131 + ], + "blocks": [ + { + "bbox": [ + 248, + 68, + 323, + 131 + ], + "lines": [ + { + "bbox": [ + 248, + 68, + 323, + 131 + ], + "spans": [ + { + "bbox": [ + 248, + 68, + 323, + 131 + ], + "type": "image", + "image_path": "a996b60f03f6e0818eab6f484b5ca817547b8277d0dd42dfab78cd42dc78b131.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 248, + 138, + 323, + 201 + ], + "blocks": [ + { + "bbox": [ + 248, + 138, + 323, + 201 + ], + "lines": [ + { + "bbox": [ + 248, + 138, + 323, + 201 + ], + "spans": [ + { + "bbox": [ + 248, + 138, + 323, + 201 + ], + "type": "image", + "image_path": "3e00a16eec4e5f11b7ff534e37b46163c34d9b104deb0dfc5276a76665140bf0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 248, + 209, + 323, + 272 + ], + "blocks": [ + { + "bbox": [ + 248, + 209, + 323, + 272 + ], + "lines": [ + { + "bbox": [ + 248, + 209, + 323, + 272 + ], + "spans": [ + { + "bbox": [ + 248, + 209, + 323, + 272 + ], + "type": "image", + "image_path": "62feeaadb48c740a8fa021d1d4b5f46f799bfe7baf5aa0bff648db1aeba74b15.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 248, + 280, + 323, + 343 + ], + "blocks": [ + { + "bbox": [ + 248, + 280, + 323, + 343 + ], + "lines": [ + { + "bbox": [ + 248, + 280, + 323, + 343 + ], + "spans": [ + { + "bbox": [ + 248, + 280, + 323, + 343 + ], + "type": "image", + "image_path": "b991057b76c741a5b68fa5408a90f1117a24a259c2001e368d12908b261d6266.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 347, + 299, + 355 + ], + "lines": [ + { + "bbox": [ + 271, + 347, + 299, + 355 + ], + "spans": [ + { + "bbox": [ + 271, + 347, + 299, + 355 + ], + "type": "text", + "content": "LaVie-SR" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 323, + 68, + 399, + 131 + ], + "blocks": [ + { + "bbox": [ + 323, + 68, + 399, + 131 + ], + "lines": [ + { + "bbox": [ + 323, + 68, + 399, + 131 + ], + "spans": [ + { + "bbox": [ + 323, + 68, + 399, + 131 + ], + "type": "image", + "image_path": "a18122c8d1e1811e314b2048307fde175fd7e0dbe388d6572dd40963cca77826.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 323, + 138, + 399, + 201 + ], + "blocks": [ + { + "bbox": [ + 323, + 138, + 399, + 201 + ], + "lines": [ + { + "bbox": [ + 323, + 138, + 399, + 201 + ], + "spans": [ + { + "bbox": [ + 323, + 138, + 399, + 201 + ], + "type": "image", + "image_path": "efd295710595a338fa204d8e60212da84c826726f1968aca43eec85454b5599f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 323, + 209, + 399, + 272 + ], + "blocks": [ + { + "bbox": [ + 323, + 209, + 399, + 272 + ], + "lines": [ + { + "bbox": [ + 323, + 209, + 399, + 272 + ], + "spans": [ + { + "bbox": [ + 323, + 209, + 399, + 272 + ], + "type": "image", + "image_path": "d8226e7903abcf50b9659fc00b173eff7d4512da80cc81a24012766075c9df02.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 323, + 280, + 399, + 343 + ], + "blocks": [ + { + "bbox": [ + 323, + 280, + 399, + 343 + ], + "lines": [ + { + "bbox": [ + 323, + 280, + 399, + 343 + ], + "spans": [ + { + "bbox": [ + 323, + 280, + 399, + 343 + ], + "type": "image", + "image_path": "941fac3490a8f1f47d513e4ddf3561913f9c7102d4bdb936c982984b9d05e9ce.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 347, + 387, + 356 + ], + "lines": [ + { + "bbox": [ + 333, + 347, + 387, + 356 + ], + "spans": [ + { + "bbox": [ + 333, + 347, + 387, + 356 + ], + "type": "text", + "content": "Upscale-A-Video" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 399, + 68, + 476, + 131 + ], + "blocks": [ + { + "bbox": [ + 399, + 68, + 476, + 131 + ], + "lines": [ + { + "bbox": [ + 399, + 68, + 476, + 131 + ], + "spans": [ + { + "bbox": [ + 399, + 68, + 476, + 131 + ], + "type": "image", + "image_path": "994d530f6379eb7426c24a3e11d4c8ff2da739c6b140ebaa8895196b9ffbcab0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 399, + 138, + 475, + 201 + ], + "blocks": [ + { + "bbox": [ + 399, + 138, + 475, + 201 + ], + "lines": [ + { + "bbox": [ + 399, + 138, + 475, + 201 + ], + "spans": [ + { + "bbox": [ + 399, + 138, + 475, + 201 + ], + "type": "image", + "image_path": "d5f848cec41cec5579cfb130d81e1b9fff96d037e77f76291b19564460b647d9.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 399, + 209, + 475, + 272 + ], + "blocks": [ + { + "bbox": [ + 399, + 209, + 475, + 272 + ], + "lines": [ + { + "bbox": [ + 399, + 209, + 475, + 272 + ], + "spans": [ + { + "bbox": [ + 399, + 209, + 475, + 272 + ], + "type": "image", + "image_path": "3b718b62722de41569cc9d0eb77ccf08a3d9757729690dbd94c18825757bdde3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 399, + 280, + 475, + 343 + ], + "blocks": [ + { + "bbox": [ + 399, + 280, + 475, + 343 + ], + "lines": [ + { + "bbox": [ + 399, + 280, + 475, + 343 + ], + "spans": [ + { + "bbox": [ + 399, + 280, + 475, + 343 + ], + "type": "image", + "image_path": "35a6f7995aff3853131987fade9ab028bac39c8d803f6492444db83fc6d5a931.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 347, + 455, + 355 + ], + "lines": [ + { + "bbox": [ + 420, + 347, + 455, + 355 + ], + "spans": [ + { + "bbox": [ + 420, + 347, + 455, + 355 + ], + "type": "text", + "content": "VEnhancer" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 476, + 68, + 552, + 131 + ], + "blocks": [ + { + "bbox": [ + 476, + 68, + 552, + 131 + ], + "lines": [ + { + "bbox": [ + 476, + 68, + 552, + 131 + ], + "spans": [ + { + "bbox": [ + 476, + 68, + 552, + 131 + ], + "type": "image", + "image_path": "ab36d63970e614e57d39a278528bdc339870bd34cd2a0dcc1e30ff8298948f6f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 476, + 138, + 552, + 201 + ], + "blocks": [ + { + "bbox": [ + 476, + 138, + 552, + 201 + ], + "lines": [ + { + "bbox": [ + 476, + 138, + 552, + 201 + ], + "spans": [ + { + "bbox": [ + 476, + 138, + 552, + 201 + ], + "type": "image", + "image_path": "beec68e6615d8fed09c2099aa7ba3ef9d244d428cf23b7f086f045e7b608016a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 476, + 209, + 552, + 272 + ], + "blocks": [ + { + "bbox": [ + 476, + 209, + 552, + 272 + ], + "lines": [ + { + "bbox": [ + 476, + 209, + 552, + 272 + ], + "spans": [ + { + "bbox": [ + 476, + 209, + 552, + 272 + ], + "type": "image", + "image_path": "05b46b287d2a9ffb6b31a0800948cd872352004d55b8af8bad14f51d467bca12.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 476, + 280, + 552, + 343 + ], + "blocks": [ + { + "bbox": [ + 476, + 280, + 552, + 343 + ], + "lines": [ + { + "bbox": [ + 476, + 280, + 552, + 343 + ], + "spans": [ + { + "bbox": [ + 476, + 280, + 552, + 343 + ], + "type": "image", + "image_path": "d6ac3e2fbeef7506a8b8979cfc2d18cf400f2822d850afcf8e1d8ba4da71a51a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 348, + 522, + 355 + ], + "lines": [ + { + "bbox": [ + 501, + 348, + 522, + 355 + ], + "spans": [ + { + "bbox": [ + 501, + 348, + 522, + 355 + ], + "type": "text", + "content": "GenVE" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 61, + 387, + 167, + 487 + ], + "blocks": [ + { + "bbox": [ + 61, + 387, + 167, + 487 + ], + "lines": [ + { + "bbox": [ + 61, + 387, + 167, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 387, + 167, + 487 + ], + "type": "image", + "image_path": "ebd9ca4e73c249d3d53ab5971e33dd342cc2d470c3d8e92d1fab73985324589d.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 489, + 555, + 512 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 555, + 512 + ], + "type": "text", + "content": "Figure 6. Comparison of temporal profile that tracks variation of the pixels in a spatial row (highlighted by the yellow line). The width of the temporal profile equals to the video width, and the height is the number of frames of generated HQ video." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 168, + 385, + 547, + 487 + ], + "blocks": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "type": "text", + "content": "Figure 5. Four visual examples of video enhancement results by different approaches on AI-synthesized videos from AIGC-Vid. The region in the red box is presented in the zoom-in view for comparison." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 385, + 547, + 487 + ], + "lines": [ + { + "bbox": [ + 168, + 385, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 168, + 385, + 547, + 487 + ], + "type": "image", + "image_path": "5ba37f5d4c5eab2fe23cc00f95050ddbdac176cc2fa1d33cf12b01ac8d03a57b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 54, + 518, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 295, + 649 + ], + "type": "text", + "content": "In contrast, the result videos of our GenVE contain more fancy visual details and natural local textures, confirming the merit of texture alignment learning for high-fidelity local region restoration. Moreover, as shown in the 4-th case, our GenVE nicely distinguishes the background tiles from the toy bear, and accordingly synthesizes the smooth and fuzzy textures on their surfaces, respectively. That basically indicates the local enhancement of each video patch can be accurately calibrated by visual cues from the corresponding region in the image reference, manifesting the merit of local-aware cross-attention in GenVE." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Besides the validation of visual appearance, we also verify the temporal consistency of generated video via the temporal profile. Figure 6 depicts the temporal variation of the pixels in a specific spatial row (i.e., the yellow line) across different approaches. As highlighted by the red box, GenVE" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 518, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 578 + ], + "type": "text", + "content": "exhibits a more seamless and smoother transition across the temporal dimension, while the others display blurriness or discontinuities. The result demonstrates the effectiveness of the motion-based conditioning augmentation exploited in GenVE to achieve better temporal coherence." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 587, + 409, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 409, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 409, + 601 + ], + "type": "text", + "content": "4.3. Model Analysis" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "Analysis on LCA and Conditioning Augmentation. We first investigate how the LCA module and conditioning augmentation strategy influence the overall performances of video enhancement. Table 2 lists the performance comparisons among variants with different combination ways of the two components. We start from the basic video diffusion model A with a controller [49], which only extracts the visual cues from low-quality videos for video enhancement. The model B upgrades the basic model A through learning" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17093" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 87, + 292, + 201 + ], + "blocks": [ + { + "bbox": [ + 55, + 62, + 297, + 84 + ], + "lines": [ + { + "bbox": [ + 55, + 62, + 297, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 62, + 297, + 84 + ], + "type": "text", + "content": "Table 2. Performance comparisons among different variants of GenVE on the YouHQ40 dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 87, + 292, + 201 + ], + "lines": [ + { + "bbox": [ + 56, + 87, + 292, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 87, + 292, + 201 + ], + "type": "table", + "html": "
ModelFusionAugmentationMUSIQ↑DOVER↑
CALCANoiseMaskTemporal
A68.3477.29
B✓69.1078.75
C✓69.4579.98
C1✓✓69.5380.16
C2✓✓✓69.8882.15
C3✓✓✓✓69.9383.50
", + "image_path": "daf012feb1b0389f1b72077afca3b5d4057d4770f848ce49d986bd64276dbc1c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": "feature interaction between image reference and LQ video by vanilla cross-attention (CA) to emphasize detail generation, and leads 1.46 improvement of DOVER. By replacing vanilla CA with our deliberately designed local-aware cross-attention (LCA) module, the model C boosts DOVER from 78.75 to 79.98. Compared to general cross-attention, our LCA module imposes the neighboring textures for more accurate feature selection via involving attention bias and local patch reshuffling, thereby facilitating the synthesis of fine-grained objects. The runs of model " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_3" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": " further validate different types of conditioning augmentation in GenVE. Notably, there is a large gain of 1.99 in terms of DOVER from model " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": ", verifying the efficacy of masking conditioning to benefit feature selection and improve model capacity in training. Finally, the model " + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_3" + }, + { + "bbox": [ + 54, + 209, + 295, + 412 + ], + "type": "text", + "content": ", i.e., our GenVE, by equipping all the three augmentations for robust model optimization, shows the best performances." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 617 + ], + "type": "text", + "content": "Analysis on Noise Level. Then, we study the influence of the noise level on the image reference and input LQ video in the inference stage. We show two video cases generated by GenVE with different noise levels in Figure 7. As shown in the figure, the appearance of output HQ video is closer to that of the input LQ video when the noise level is smaller. The results are expected since that the adding noise in forward diffusion is small and the backward denoising cannot change a lot. Conversely, when exploiting high noise level in video noisings, more fancy local details (e.g., the tooth of human in the 2-nd case) are presented while some visual contents can be changed. Therefore, we achieve the tradeoff control between the video quality and fidelity by additionally involving noise augmentation in diffusion model training. In evaluation, we empirically set the noise level as 250 to seek a good balance between synthesized novel details and original visual appearance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Analysis on Image Reference. The image reference acts as a bridge between the input LQ video and generated HQ video in our GenVE. A valid question is how different image references influence the final results of video enhancement. Figure 8 shows two visual cases of video enhancement by using two distinct image references in GenVE. Although there are substantial visual differences between the two conditioning image references, the" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 63, + 375, + 121 + ], + "blocks": [ + { + "bbox": [ + 317, + 63, + 375, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 63, + 375, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 63, + 375, + 121 + ], + "type": "image", + "image_path": "c381f913790bc4ee7f47d970a603d235ff1dbf14b47f30bdaee5a5c6c09ee165.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 121, + 374, + 177 + ], + "blocks": [ + { + "bbox": [ + 317, + 121, + 374, + 177 + ], + "lines": [ + { + "bbox": [ + 317, + 121, + 374, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 374, + 177 + ], + "type": "image", + "image_path": "7f3576bbdfd8bc395e429a2457897fbb70f19920e6628c970dacc0259eb6fa2a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 178, + 354, + 185 + ], + "lines": [ + { + "bbox": [ + 337, + 178, + 354, + 185 + ], + "spans": [ + { + "bbox": [ + 337, + 178, + 354, + 185 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 376, + 63, + 434, + 121 + ], + "blocks": [ + { + "bbox": [ + 376, + 63, + 434, + 121 + ], + "lines": [ + { + "bbox": [ + 376, + 63, + 434, + 121 + ], + "spans": [ + { + "bbox": [ + 376, + 63, + 434, + 121 + ], + "type": "image", + "image_path": "8d55e4ab0db398b4c7a08da4421e13a41dd701b2b6b141b976887bd262c85de4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 377, + 121, + 433, + 177 + ], + "blocks": [ + { + "bbox": [ + 377, + 121, + 433, + 177 + ], + "lines": [ + { + "bbox": [ + 377, + 121, + 433, + 177 + ], + "spans": [ + { + "bbox": [ + 377, + 121, + 433, + 177 + ], + "type": "image", + "image_path": "ad068027973554dc28e039e405106f7e1ce03455d930353607f5600e3ef738f2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 178, + 431, + 185 + ], + "lines": [ + { + "bbox": [ + 379, + 178, + 431, + 185 + ], + "spans": [ + { + "bbox": [ + 379, + 178, + 431, + 185 + ], + "type": "text", + "content": "Noise level n=100" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 435, + 63, + 493, + 121 + ], + "blocks": [ + { + "bbox": [ + 435, + 63, + 493, + 121 + ], + "lines": [ + { + "bbox": [ + 435, + 63, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 435, + 63, + 493, + 121 + ], + "type": "image", + "image_path": "c5c481be2ad410f55398ef5d17e7452770ff0402d15d206d2ea0ac00c718317d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 436, + 121, + 493, + 177 + ], + "blocks": [ + { + "bbox": [ + 436, + 121, + 493, + 177 + ], + "lines": [ + { + "bbox": [ + 436, + 121, + 493, + 177 + ], + "spans": [ + { + "bbox": [ + 436, + 121, + 493, + 177 + ], + "type": "image", + "image_path": "ab21a0e837258f829e6b38706a6ef79d04ab9747eac4ba7599cfe66a3260c1e5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 178, + 490, + 185 + ], + "lines": [ + { + "bbox": [ + 440, + 178, + 490, + 185 + ], + "spans": [ + { + "bbox": [ + 440, + 178, + 490, + 185 + ], + "type": "text", + "content": "Noise level " + }, + { + "bbox": [ + 440, + 178, + 490, + 185 + ], + "type": "inline_equation", + "content": "n = 200" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 495, + 63, + 552, + 121 + ], + "blocks": [ + { + "bbox": [ + 495, + 63, + 552, + 121 + ], + "lines": [ + { + "bbox": [ + 495, + 63, + 552, + 121 + ], + "spans": [ + { + "bbox": [ + 495, + 63, + 552, + 121 + ], + "type": "image", + "image_path": "70dcb123aefe06956aed7d5dab3559c4c9c07b198cc0c25f52e7d3974b623c55.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 496, + 121, + 552, + 177 + ], + "blocks": [ + { + "bbox": [ + 496, + 121, + 552, + 177 + ], + "lines": [ + { + "bbox": [ + 496, + 121, + 552, + 177 + ], + "spans": [ + { + "bbox": [ + 496, + 121, + 552, + 177 + ], + "type": "image", + "image_path": "09174802adf8eb672b744eae204185a37889bf77b194f33dc6f711c8b83ffa27.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 177, + 549, + 184 + ], + "lines": [ + { + "bbox": [ + 499, + 177, + 549, + 184 + ], + "spans": [ + { + "bbox": [ + 499, + 177, + 549, + 184 + ], + "type": "text", + "content": "Noise level " + }, + { + "bbox": [ + 499, + 177, + 549, + 184 + ], + "type": "inline_equation", + "content": "n = 300" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 362, + 255 + ], + "blocks": [ + { + "bbox": [ + 314, + 186, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 314, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 314, + 186, + 553, + 207 + ], + "type": "text", + "content": "Figure 7. Two visual examples of video enhancement results by using different noise levels in GenVE." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 209, + 362, + 255 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 362, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 362, + 255 + ], + "type": "image", + "image_path": "6ff236f5a9a176685bfc0668fb391526f76f364887b7b54f9506578af19a934a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 317, + 255, + 362, + 300 + ], + "blocks": [ + { + "bbox": [ + 317, + 255, + 362, + 300 + ], + "lines": [ + { + "bbox": [ + 317, + 255, + 362, + 300 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 362, + 300 + ], + "type": "image", + "image_path": "3a051011b9fa37183ec748d16f6fb9f1f5f7462c6798ffe9beb58a6d4d09cc40.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 300, + 347, + 308 + ], + "lines": [ + { + "bbox": [ + 331, + 300, + 347, + 308 + ], + "spans": [ + { + "bbox": [ + 331, + 300, + 347, + 308 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 308, + 553, + 330 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 330 + ], + "type": "text", + "content": "Figure 8. Two visual examples of video enhancement results guided by two different image references in GenVE." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 365, + 209, + 459, + 255 + ], + "blocks": [ + { + "bbox": [ + 365, + 209, + 459, + 255 + ], + "lines": [ + { + "bbox": [ + 365, + 209, + 459, + 255 + ], + "spans": [ + { + "bbox": [ + 365, + 209, + 459, + 255 + ], + "type": "image", + "image_path": "d7ce51d43d7d90cfd59231c2aac6827c4ebd1c06c98853e91f62db660f697f1a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 365, + 255, + 458, + 300 + ], + "blocks": [ + { + "bbox": [ + 365, + 255, + 458, + 300 + ], + "lines": [ + { + "bbox": [ + 365, + 255, + 458, + 300 + ], + "spans": [ + { + "bbox": [ + 365, + 255, + 458, + 300 + ], + "type": "image", + "image_path": "a0be5c3b7156d2ea267b2623ed8d14eced0af4782a39a12df90faa945e7bae81.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 300, + 406, + 308 + ], + "lines": [ + { + "bbox": [ + 372, + 300, + 406, + 308 + ], + "spans": [ + { + "bbox": [ + 372, + 300, + 406, + 308 + ], + "type": "text", + "content": "Reference-1" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 411, + 300, + 447, + 308 + ], + "lines": [ + { + "bbox": [ + 411, + 300, + 447, + 308 + ], + "spans": [ + { + "bbox": [ + 411, + 300, + 447, + 308 + ], + "type": "text", + "content": "Result-1" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 461, + 209, + 507, + 255 + ], + "blocks": [ + { + "bbox": [ + 461, + 209, + 507, + 255 + ], + "lines": [ + { + "bbox": [ + 461, + 209, + 507, + 255 + ], + "spans": [ + { + "bbox": [ + 461, + 209, + 507, + 255 + ], + "type": "image", + "image_path": "91f5cb132ad848bb494bd60a405e7308e186c2a6c1bdefbcc85a36b2a0c239e8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 461, + 255, + 506, + 300 + ], + "blocks": [ + { + "bbox": [ + 461, + 255, + 506, + 300 + ], + "lines": [ + { + "bbox": [ + 461, + 255, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 461, + 255, + 506, + 300 + ], + "type": "image", + "image_path": "4b600df0706e2a9ad4a1ddf0a7a6b2ba4c7db9c6b33950f6e6308d6d09523561.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 300, + 501, + 308 + ], + "lines": [ + { + "bbox": [ + 467, + 300, + 501, + 308 + ], + "spans": [ + { + "bbox": [ + 467, + 300, + 501, + 308 + ], + "type": "text", + "content": "Reference-2" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 507, + 209, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 507, + 209, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 507, + 209, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 507, + 209, + 553, + 255 + ], + "type": "image", + "image_path": "e6042350be53c431e2411ab33632c5dad2d2c802fe2486ee1b55507a5e60be8c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 508, + 255, + 553, + 300 + ], + "blocks": [ + { + "bbox": [ + 508, + 255, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 508, + 255, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 508, + 255, + 553, + 300 + ], + "type": "image", + "image_path": "d6cf06ef4a5520532b7b2b25cb9874cc2c293d8a94c96384d8d19f59070d2940.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 300, + 542, + 308 + ], + "lines": [ + { + "bbox": [ + 518, + 300, + 542, + 308 + ], + "spans": [ + { + "bbox": [ + 518, + 300, + 542, + 308 + ], + "type": "text", + "content": "Result-2" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 335, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 396 + ], + "type": "text", + "content": "final video enhancement results are perceptually similar. We speculate that the results are contributed by the augmentation strategy of masking condition which encourages GenVE to search multiple visual cues in image reference holistically, enhancing the generation of local details." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 406, + 391, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 391, + 418 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 391, + 418 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 641 + ], + "type": "text", + "content": "We have presented GenVE that explores the image reference to boost diffusion-based video enhancement. In particular, we study the problem from the view point of leveraging image reference as the bridge for semantic and texture alignment between input low-quality video and generated high-quality video. To consolidate our idea, GenVE first exploits an image diffusion model to upscale a key frame of input video for semantics-aligned high-quality image reference generation. A video controller then transfers the patch-level texture details in high-quality image reference to the correspondence region of low-quality video for fine-grained details enhancement. In the training of GenVE, a series of conditioning augmentation strategies is conducted to further strengthen the model robustness. Experiments conducted on two public real-world video datasets, i.e., YouHQ40 and VideoLQ, as well as a self-built AI-synthesized video dataset AIGC-Vid, validate the effectiveness of the proposed GenVE for video enhancement." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the Beijing Municipal Science and Technology Project No. Z241100001324002, and Beijing Nova Program No. 20240484681, and the GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC, and the Supercomputing Center of the USTC." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17094" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your Latents: High-Resolution Video Synthesis with Latent Diffusion Models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Realtime Video Super-Resolution with Spatio-Temporal Networks and Motion Compensation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Qi Cai, Jingwen Chen, Yang Chen, Yehao Li, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Yiheng Zhang, Fengbin Gao, Peihan Xu, et al. HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer. arXiv:2505.22705, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 278 + ], + "type": "text", + "content": "[4] Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. Reference-based Image Super-Resolution with Deformable Attention Transformer. In ECCV, 2022. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. BasicVSR: The Search for Essential Components in Video Super-Resolution and Beyond. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 367 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving Video SuperResolution with Enhanced Propagation and Alignment. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "text", + "content": "[7] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Investigating Tradeoffs in Real-World Video Super-Resolution. In CVPR, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 445 + ], + "type": "text", + "content": "[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "text", + "content": "[9] Jingyuan Chen, Fuchen Long, Jie An, Zhaofan Qiu, Ting Yao, Jiebo Luo, and Tao Mei. Ouroboros-Diffusion: Exploring Consistent Content Generation in Tuning-free Long Video Diffusion. In AAAI, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 534 + ], + "type": "text", + "content": "[10] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wenggang Zhou, Jiebo Luo, and Tao Mei. Learning 3D Shape Latent for Point Cloud Completion. IEEE Trans. on Multimedia, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 578 + ], + "type": "text", + "content": "[11] Zhikai Chen, Fuchen Long, Zhaofan Qiu, Ting Yao, Wengang Zhou, Jiebo Luo, and Tao Mei. Learning Spatial Adaptation and Temporal Coherence in Diffusion Models for Video Super-Resolution. In CVPR, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[12] Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models. In ICCV, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 635 + ], + "type": "text", + "content": "[13] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 679 + ], + "type": "text", + "content": "[14] Jingwen He, Tianfan Xue, Dongyang Liu, Xinqi Lin, Peng Gao, Dahua Lin, Yu Qiao, Wanli Ouyang, and Ziwei Liu. VEnhancer: Generative Space-Time Enhancement for Video Generation. arXiv:2407.07667, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[15] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In ICLR, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance. In NeurIPS Workshop, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 118 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 121, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 553, + 175 + ], + "type": "text", + "content": "[18] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen Video: High Definition Video Generation with Diffusion Models. arXiv:2210.02303, 2022. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 209 + ], + "type": "text", + "content": "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video Diffusion Models. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 212, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 276 + ], + "type": "text", + "content": "[20] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive Benchmark Suite for Video Generative Models. In CVPR, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 553, + 312 + ], + "type": "text", + "content": "[21] Takashi Isobe, Xu Jia, Shuhang Gu, Songjiang Li, Shengjin Wang, and Qi Tian. Video Super-Resolution with Recurrent Structure-Detail Network. In ECCV, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 553, + 346 + ], + "type": "text", + "content": "[22] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. MUSIQ: Multi-Scale Image Quality Transformer. In ICCV, 2021. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 553, + 380 + ], + "type": "text", + "content": "[23] Wenbo Li, Xin Tao, Taian Guo, Lu Qi, Jiangbo Lu, and Jiaya Jia. MuCAN: Multi-Correspondence Aggregation Network for Video Super-Resolution. In ECCV, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 415 + ], + "type": "text", + "content": "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. VRT: A Video Restoration Transformer. IEEE TIP, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 418, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 418, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 418, + 553, + 472 + ], + "type": "text", + "content": "[25] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc Van Gool. Recurrent Video Restoration Transformer with Guided Deformable Attention. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 475, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 506 + ], + "type": "text", + "content": "[26] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Jiebo Luo, and Tao Mei. Stand-Alone Inter-Frame Attention in Video Models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 540 + ], + "type": "text", + "content": "[27] Fuchen Long, Zhaofan Qiu, Yingwei Pan, Ting Yao, Chong-Wah Ngo, and Tao Mei. Dynamic Temporal Filtering in Video Models. In ECCV, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "text", + "content": "[28] Fuchen Long, Ting Yao, Zhaofan Qiu, Xinmei Tian, Jiebo Luo, and Tao Mei. Bi-calibration Networks for Weakly-Supervised Video Representation Learning. IJCV, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 610 + ], + "type": "text", + "content": "[29] Fuchen Long, Zhaofan Qiu, Ting Yao, and Tao Mei. VideoStudio: Generating Consistent-Content and Multi-Scene Videos. In ECCV, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "text", + "content": "[30] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards Photorealistic Image Generation and Editing with Text-guided Diffusion Models. In ICML, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[31] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 5" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17095" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "text", + "content": "[34] Shuwei Shi, Jinjin Gu, Liangbin Xie, Xintao Wang, Yujiu Yang, and Chao Dong. Rethinking Alignment in Video Super-Resolution Transformers. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 239 + ], + "type": "text", + "content": "[35] Gyumin Shim, Jinsun Park, and In So Kweon. Robust Reference-based Super-Resolution with Similarity-Aware Deformable Convolution. In CVPR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 263 + ], + "type": "text", + "content": "[36] Zachary Teed and Jia Deng. RAFT: Recurrent All-pairs Field Transforms for Optical Flow. In ECCV, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 297 + ], + "type": "text", + "content": "[37] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring CLIP for Assessing the Look and Feel of Images. In AAAI, 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 331 + ], + "type": "text", + "content": "[38] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-ESRGAN: Training Real-World Blind SuperResolution with Pure Synthetic Data. In ICCV, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 333, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 409 + ], + "type": "text", + "content": "[39] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, Yuwei Guo, Tianxing Wu, Chenyang Si, Yuming Jiang, Cunjian Chen, Chen Change Loy, Bo Dai, Dahua Lin, Yu Qiao, and Ziwei Liu. LaVie: High-Quality Video Generation with Cascaded Latent Diffusion Models. IJCV, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[40] Haoning Wu, Erli Zhang, Liang Liao, Chaofeng Chen, Jing wen Hou, Annan Wang, Qiong Yan Wenxiu Sun, and Weisi Lin. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives. In ICCV, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[41] Yanchun Xie, Jimin Xiao, Mingjie Sun, Chao Yao, and Kaizhu Huang. Feature Representation Matters: End-to-End Learning for Reference-based Image Super-resolution. In ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "type": "text", + "content": "[42] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal Modulation Network for Controllable Space-Time Video Super-Resolution. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 588 + ], + "type": "text", + "content": "[43] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T. Freeman. Video Enhancement with Task-Oriented Flow. IJCV, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[44] Xu Yan, Weibing Zhao, Kun Yuan, Ruimao Zhang, Zhen Li, and Shuguang Cui. Towards Content-Independent Multi-Reference Super-Resolution: Adaptive Pattern Matching and Feature Aggregation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 667 + ], + "type": "text", + "content": "[45] Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo. Learning Texture Transformer Network for Image Super-Resolution. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[46] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, and Jiayi Ma. Progressive Fusion Video Super-Resolution Network via Exploiting Non-Local Spatio-Temporal Correlations. In ICCV, 2019. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[47] Peng Yi, Zhongyuan Wang, Kui Jiang, Junjun Jiang, Tao Lu, Xin Tian, and Jiayi Ma. Omniscient Video Super-Resolution. In ICCV, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[48] Huanjing Yue, Xiaoyan Sun, Jingyu Yang, and Feng Wu. Landmark Image Super-resolution by Retrieving Web Images. IEEE TIP, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[49] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[50] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qin, Xiang Wang, Deli Zhao, and Jingren Zhou. I2VGen-XL: High-Quality Image-to-Video Synthesis via Cascaded Diffusion Models. arXiv:2311.04145, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 262 + ], + "type": "text", + "content": "[51] Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. Image Super-Resolution by Neural Texture Transfer. In CVPR, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 307 + ], + "type": "text", + "content": "[52] Zhongwei Zhang, Fuchen Long, Yingwei Pan, Zhaofan Qiu, Ting Yao, Yang Cao, and Tao Mei. TRIP: Temporal Residual Learning with Image Noise Prior for Image-to-Video Diffusion Models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[53] Zhongwei Zhang, Fuchen Long, Zhaofan Qiu, Yingwei Pan, Wu Liu, Ting Yao, and Tao Mei. MotionPro: A Precise Motion Controller for Image-to-Video Generation. In CVPR, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[54] Haitian Zheng, Mengqi Ji, Haoqian Wang, Yebin Liu, and Lu Fang. CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping. In ECCV, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[55] Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy. Upscale-A-Video: Temporal-Consistent Diffusion Model for Real-World Video SuperResolution. In CVPR, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17096" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_content_list.json b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..70e46d66b8e122387e33b5b468d5e684f66263af --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_content_list.json @@ -0,0 +1,1494 @@ +[ + { + "type": "text", + "text": "Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching", + "text_level": 1, + "bbox": [ + 169, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yang Liu $^{1,2}$ , Wentao Feng $^{1,2}$ , Zhuoyao Liu $^{1,2}$ , Shudong Huang $^{1,2,*}$ , Jiancheng Lv $^{1,2}$", + "bbox": [ + 155, + 202, + 815, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1College of Computer Science, Sichuan University, China", + "bbox": [ + 243, + 223, + 728, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Engineering Research Center of Machine Learning and Industry Intelligence, China", + "bbox": [ + 127, + 239, + 841, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://d2s-vse.github.io", + "bbox": [ + 372, + 258, + 598, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enabling Visual Semantic Models to effectively handle multi-view description matching has been a longstanding challenge. Existing methods typically learn a set of embeddings to find the optimal match for each view's text and compute similarity. However, the visual and text embeddings learned through these approaches have limited information capacity and are prone to interference from locally similar negative samples. To address this issue, we argue that the information capacity of embeddings is crucial and propose Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE), which enhances the information capacity of sparse text by leveraging dense text distillation. Specifically, D2S-VSE is a two-stage framework. In the pretraining stage, we align images with dense text to enhance the information capacity of visual semantic embeddings. In the fine-tuning stage, we optimize two tasks simultaneously, distilling dense text embeddings to sparse text embeddings while aligning images and sparse texts, enhancing the information capacity of sparse text embeddings. Our proposed D2S-VSE model is extensively evaluated on the large-scale MS-COCO and Flickr30K datasets, demonstrating its superiority over recent state-of-the-art methods.", + "bbox": [ + 89, + 340, + 483, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the domain of vision and language, the alignment of vision and language stands as a foundational task, with the core objective of establishing a robust and precise bridge between these two distinct yet interconnected modalities. The traditional approach to cross-modal retrieval [12, 24, 31, 32] follows a standard procedure: encoding images and texts into a shared embedding space and measuring their similarity. Typically, models are optimized using a triplet loss, which ensures that matched image-text pairs exhibit higher similarity than unmatched pairs [9, 16].", + "bbox": [ + 89, + 726, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c081069e2598ab9f860806e5705792e1a064f6b608bca0533f41982fc40f83b1.jpg", + "image_caption": [ + "Figure 1. (Top) In previous methods, when addressing problems with varying information density, a set of visual embeddings is learned to match the most similar text embeddings. However, this approach results in learned embeddings with limited information capacity. (Mid) Since dense text has a higher information density than sparse text, training with aligned dense text results in image embeddings with greater information capacity. (Bottom) Sparse text embedding can be distilled through dense text embedding to enhance its information capacity, enabling the matching of text descriptions from different perspectives with a single visual embedding. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 516, + 306, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most traditional Visual Semantic Embedding (VSE) models overlook the inherent differences between image and text modalities [3, 18, 35]. Specifically, visual signals represent natural, objective records of real-world phenomena, whereas textual signals are inherently seman-", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21679", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tic and subjective, shaped by human interpretation. Consequently, the same image may be described in various ways depending on individual perspectives. Existing studies refer to this issue as the \"ambiguity issue\" [4, 36], \"sparse annotations\" [15] and \"multi-view descriptions\" [20, 33]. But, at its core, the problem stems from the inherent differences between image and text modalities, particularly the disparity in \"information density\" [26].", + "bbox": [ + 89, + 90, + 480, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [4, 20, 26, 36] propose learning a sets of embeddings from images/texts, where each embedding captures specific semantics. For each image-text pair, the most relevant embedding is selected for similarity calculation. While this approach reduces computational overhead and partially addresses information density challenges, we identify a critical limitation: As shown in the figure 1, the text in the dataset is usually short and concise. The image and text embeddings learned in this way can only capture a limited subset of features, resulting in limited information capacity being learned, and are therefore easily disturbed by negative samples with similar local semantics. Given this, we argue that the information capacity of visual semantic embeddings is crucial for the image-text matching task. The key question, then, is: how can we learn visual semantic embeddings with higher information capacity?", + "bbox": [ + 89, + 215, + 482, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel approach, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE), which enhances visual semantic embeddings by aligning both semantics and information capacity. Our key hypothesis posits that text embeddings derived from dense image captions inherently possess greater information capacity compared than those from sparse textual annotations in traditional datasets such as Flickr30K and COCO. Specifically, we introduce a two-stage training framework. In the pre-train stage, we generate dense descriptions for images using LLaVa and pre-train image-dense description pairs to align images with dense textual descriptions in the embedding space. This alignment enables the model to extract more comprehensive features, resulting in embeddings with higher information capacity. In the fine-tuning stage, we introduce a dense-to-sparse feature distillation method. During this phase, while maintaining alignment between images and sparse text annotations from the dataset, we additionally distill knowledge from dense text embeddings into sparse text embeddings to enhance their information capacity. We formulate this distillation process as a masked signal reconstruction task for sparse text embeddings. To achieve this, we implement a Transformer-based decoder module that follows the text encoder. This decoder processes short text features combined with learnable mask tokens, aiming to predict the latent context of the sparse text and thus enhance its information capacity. And in the inference phase, we only require the image and the sparse text from the dataset, without the need for dense text, thus in", + "bbox": [ + 89, + 463, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "curring no additional computational overhead.", + "bbox": [ + 511, + 90, + 818, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 531, + 107, + 870, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Importance of Information Capacity: We identify the critical role of enhancing the information capacity of visual semantic embeddings. Based on this insight, we propose a novel framework to improve the VSE model by increasing the information capacity of visual semantic embeddings.", + "- Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE): We introduce D2S-VSE, a two-stage training approach that leverages dense text to enhance the information capacity of image embeddings in the first pre-training stage. In the second stage, a distillation branch is incorporated to improve the information capacity of sparse text embedding by distilling dense text embedding from sparse text embedding, ensuring both semantic alignment and balanced information capacity between image and text embeddings.", + "- State-of-the-Art Performance: We evaluate D2S-VSE with existing fine-grained methods across various model backbones. D2S-VSE outperforms the latest state-of-the-art methods on image-text retrieval benchmarks, including Flickr30K and MS-COCO." + ], + "bbox": [ + 511, + 123, + 903, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 457, + 653, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-Text Matching", + "text_level": 1, + "bbox": [ + 511, + 483, + 714, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing Image-text retrieval approaches can be broadly categorized into two groups based on the granularity of the matching similarity: Global-level matching and Local-level matching methods. Global-level matching methods project images and texts into a shared embedding space, computing similarity using improved loss functions [5, 9] to bring semantically aligned pairs closer together. To improve the quality of the embedding space, recent related works proposed several more sophisticated networks, such as graph convolutional network [22, 37], Generalized Pooling Operator [3], and other model architectures [14, 18, 19]. In order to facilitate the closer alignment of semantically corresponding image-text pairs. Different from global matching methods, local matching methods[2, 7, 25, 34, 38, 39, 42, 43] focus more on the alignment between regions. Most of them capture the finer-grained relationship between images and texts by adopting cross-modal networks.", + "bbox": [ + 511, + 506, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Information Density Issues. Recent studies have proposed various methods to address the issue of differing information density between images and text. PCME [5] proposed a probabilistic cross-modal embedding to properly represent the one-to-many relationships in joint embedding. DivE [15] introduced a slot attention-based set prediction module, where the elements of the embedding set compete it-", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21680", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3528f257f5d9b0d3a36000fdf470d305c13e2f8bf70af8ff136da6d2153e2510.jpg", + "image_caption": [ + "Figure 2. Overview of our D2S-VSE framework. (a) Dense Text Generation: D2S-VSE utilizes the LLaVa [23] to generate dense text descriptions for images in the dataset that contain all the details of the image. (b) Pre-train With Dense Descriptions: D2S-VSE first uses contrastive learning to pre-train the model on image-dense-text pairs to enhance the information capacity of visual semantic embeddings. (c) Finetune With the Sparse Descriptions: D2S-VSE performs distillation from dense to sparse while contrasting the alignment, allowing sparse text embeddings to learn more details of dense text embeddings, thereby improving the information capacity of sparse text embeddings and better aligning them with image embeddings." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eratively to aggregate the input data. Under the framework of PCME, PCME++ [4] introduced a new probabilistic distance with a closed-form solution. AVSE [26] proposed a matching module for compute the dynamic similarity between asymmetric multi-view image and text embeddings.", + "bbox": [ + 89, + 588, + 482, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These methods learn a set of embeddedings to match ambiguous samples, but reduce the information capacity of each embedding.", + "bbox": [ + 89, + 666, + 483, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision language model with dense text.", + "text_level": 1, + "bbox": [ + 89, + 724, + 423, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of vision-language research, few studies have leveraged long-text descriptions to enhance the multimodal representation of images. DreamLip [44] utilizes a multimodal large language model to generate dense text descriptions and randomly extracts sentence fragments for pre-training, but it does not fully exploit long-text pretraining. Long-CLIP [41] fine-tunes CLIP-pre-trained on a short-text-image dataset—on long-text-image pairs to unlock long-text capabilities, yet it overlooks short-text effectiveness. LoTLIP [40] directly pre-trains on long-text", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image data and introduces corner tokens to aggregate diverse textual information, preserving the ability to handle short text.", + "bbox": [ + 511, + 588, + 906, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from the above methods that focus on pretraining, our method utilizes a very small amount of long text for pre-training, aiming to use the long text to enhance the information capacity of visual semantic embeddings.", + "bbox": [ + 511, + 633, + 906, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dense-to-Sparse Feature Distilled Visual Semantic Embedding", + "text_level": 1, + "bbox": [ + 511, + 705, + 906, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to enhance Visual Semantic Embedding (VSE) models to address the challenge of inconsistent information density between images and text. Existing VSE models typically align images with the sparse textual descriptions provided in datasets. This approach biases the image encoder toward textual perspectives, limiting its ability to fully comprehend visual content and making it difficult to handle ambiguous retrieval cases. To overcome this issue, we propose a two-stage training procedure. An overview of our D2S-VSE framework is illustrated in Figure 2.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21681", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Stage I. We leverage the pre-trained multimodal large model LLaVa to generate dense textual descriptions for images. The VSE model is then pre-trained using these image-dense text pairs. By training on data with comparable information density, the model learns embeddings with greater information capacity, enabling a more comprehensive understanding of visual content (Sec. 3.1).", + "- **Stage II.** We fine-tune the pre-trained VSE model on general datasets containing image-sparse text pairs. Simultaneously, we distill knowledge from the dense text features into the sparse text representations, enhancing their information capacity. While aligning image-text embeddings, we also align the information capacity between visual and textual modalities. This enables the VSE model to effectively handle varying information densities in images and texts (Sec. 3.2)." + ], + "bbox": [ + 88, + 90, + 480, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Pre-training with Dense Descriptions", + "text_level": 1, + "bbox": [ + 89, + 345, + 410, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the issue of inconsistent information density between images and text, we introduce a pre-training stage where the Visual Semantic Embedding (VSE) model learns from image-dense text pairs. This stage enhances the model's ability to extract rich semantic representations from images by ensuring that textual descriptions contain a level of detail comparable to the visual information.", + "bbox": [ + 89, + 369, + 483, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dense Text Generation. To generate dense textual descriptions for images, we leverage the pre-trained multimodal large model LLaVa [23], which integrates vision and language understanding to generate detailed descriptions based on image inputs. Unlike traditional dataset texts, which are often brief and sparse, LLaVa produces detailed descriptions that include object attributes, relationships, actions, and contextual information. Given a raw image $I$ , we feed it into the LLaVa with the prompt \"Describe this image in detail\". LLaVa then produces a dense textual description that captures fine-grained details of the image. The hyperparameters used for LLaVa are detailed in Sec 4.2. This process ensures that the textual input provides a comprehensive representation of the image, reducing the information density gap between modalities.", + "bbox": [ + 89, + 483, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Dense-Caption Pre-train. Once the dense captions are generated, we pre-train the VSE model using these image-dense text pairs. The objective of this stage is to enable the model to learn embeddings that effectively capture fine-grained visual semantics and their corresponding textual descriptions.", + "bbox": [ + 89, + 717, + 480, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ a contrastive learning framework where the image encoder $\\mathcal{F}_i$ and the dense text encoder $\\mathcal{F}_t^d$ map images and dense textual descriptions into a shared embedding space. Formally, given a dense textual description $T^d$ and a raw image as $I$ , our goal is to map these inputs into compact vectors $\\mathbf{t}^d \\in \\mathbb{R}^d$ and $\\mathbf{i} \\in \\mathbb{R}^d$ such that their semantic", + "bbox": [ + 89, + 810, + 480, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f64ed5e7bfcfd77a3f28f6584a0d702661448cb09ea1092c7f7d4ad395b1efb.jpg", + "image_caption": [ + "Figure 3. Architecture for the sparse text branch. Similar to the decoder in MAE [13], we introduce a set of learnable tokens in the sparse text branch to predict the implicit context of sparse text through the decoder. The learnable tokens help in reconstructing the missing semantics, allowing the model to better capture the underlying context and relationships within the sparse text." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 901, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relationships can be quantified through a similarity metric $S(\\cdot, \\cdot)$ . Specifically, the similarity function is formulated as:", + "bbox": [ + 511, + 494, + 906, + 539 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (I, T) = \\frac {\\mathbf {i} ^ {T} \\mathbf {t} ^ {d}}{| | \\mathbf {i} | | \\cdot | | \\mathbf {t} ^ {d} | |} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 551, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize the model, we use triplet loss function with the hardest negative samples to minimize the similarity between relevant pairs and to maximize the distance between irrelevant ones. The optimized loss function is mathematically formulated as:", + "bbox": [ + 511, + 593, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {p r e t r a i n}} = \\sum_ {(I, T ^ {d}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {d}) + S (I, \\hat {T} ^ {d}) \\right] _ {+} \\tag {2} \\\\ + \\left[ \\alpha - S (I, T ^ {d}) + S (\\hat {I}, T ^ {d}) \\right] _ {+} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 679, + 903, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ serves as a margin parameter and $[x]_{+} \\equiv Max(x,0)$ . In dataset $\\mathcal{D}$ , the similarity in a positive pair $S(I,T^{d})$ should be higher than that in the hardest negative pairs $S(\\hat{I},T^{d})$ and $S(I,\\hat{T}^{d})$ by a margin $\\alpha$ .", + "bbox": [ + 511, + 758, + 906, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Finetune With Sparse Description", + "text_level": 1, + "bbox": [ + 511, + 832, + 808, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage I pre-training with dense descriptions enhances the information capacity of embeddings of the VSE model, but there exists a gap between the training and testing data, as", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21682", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/55c49958cffaf328941b4ff65658bed08c2d3ddd3fbbe7dbc98488e2d3f031da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFGFlickr30k 1K TestMS-COCO 5-fold 1K TestMS-COCO 5K Test
Text RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSum
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
ViT-Base-224 + BERT-base
VSE++ [9]X71.892.896.559.484.790.9496.175.094.698.062.789.494.9514.662.387.693.443.973.683.3413.4
SCAN [17]✓69.590.995.656.483.190.0485.676.095.498.164.590.895.8520.653.981.890.042.972.382.5423.5
SGR [7]✓69.790.895.259.184.189.9488.777.295.098.065.190.795.8521.854.982.890.542.872.282.5425.8
CHAN [29]✓69.291.895.058.484.990.6489.977.195.198.165.091.096.0522.256.383.290.143.072.682.8428.0
LAPS [11]✓74.093.497.462.587.392.7507.378.795.598.366.291.396.2526.357.584.090.844.574.083.6434.4
AVSE [26]X76.094.697.562.788.493.1512.379.895.698.367.091.596.3528.558.884.391.045.174.383.9437.4
D2S-VSEX82.896.198.368.591.394.9531.980.197.099.268.192.596.7533.760.185.592.546.375.985.2445.6
ViT-Base-384 + BERT-base
VSE++ [9]X77.195.797.565.890.294.3520.577.095.798.464.691.196.2523.054.982.890.442.472.482.8425.8
SCAN [17]✓75.494.496.963.688.693.5512.576.195.598.565.191.696.3523.153.381.890.042.672.682.9423.1
SGR [7]✓76.994.998.164.288.493.3515.875.895.798.665.692.096.5524.253.381.089.642.973.183.7423.6
CHAN [29]✓75.494.597.663.288.693.1512.478.195.898.666.182.196.6527.355.683.891.243.473.683.5431.1
LAPS [11]✓79.096.098.167.390.594.5525.478.696.398.968.092.496.8531.057.484.992.546.475.885.2442.2
AVSE [26]X80.396.498.767.991.294.7529.281.197.199.068.392.797.0535.261.286.893.246.275.985.0448.3
D2S-VSEX84.197.599.170.391.695.3537.980.897.299.169.092.996.8535.860.686.593.246.876.485.7449.1
", + "bbox": [ + 98, + 88, + 898, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of experimental results using ViT-Base [8] and BERT [6] the MS-COCO and Flickr30K datasets. \"FG\" indicates whether fine-grained cross-modal alignment is applied. The best performance is highlighted in bold.", + "bbox": [ + 89, + 426, + 906, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the model is trained on dense texts and tested on sparse texts, leading to suboptimal performance. However, simply fine-tuning directly on the image-sparse text pair data will not make the sparse text embedding contain richer semantics, that is, more information capacity. To address this problem, we introduce a novel distillation mechanism while aligning, which treats sparse text as masked dense text and designs a decoder. Through distillation, the decoder can decode sparse text embeddings into dense text embeddings, thereby improving the information capacity of sparse text embeddings. This mechanism extracts sparse text embeddings from dense text embeddings while aligning images and text, thereby enhancing the information capacity of sparse text features.", + "bbox": [ + 88, + 481, + 480, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Aligning Information Capacity by Dense-to-Sparse Feature Distillation", + "text_level": 1, + "bbox": [ + 89, + 700, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure that the sparse text embedding retains as much information as the dense text embedding, we introduce a feature distillation mechanism from dense to sparse. Inspired by masked signal modeling [6, 13], we treat sparse text as a masked version of dense text. This allows us to decode and reconstruct the sparse text embedding through feature distillation, enabling the reconstructed embedding to approximate the information capacity of the dense text embedding.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer-based Decoder. To achieve this reconstruction, we design a Transformer-based decoder, as illustrated", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in Figure 3. The decoder plays a pivotal role in D2S-VSE, responsible for reconstructing the dense text embedding from the sparse text embedding by leveraging feature distillation. It learns to restore the missing contextual information in sparse text, ensuring that the final sparse text embedding retains as much semantic richness as the dense text embedding. Given a sparse text $T^s$ , we first feed it into the sparse text encoder $\\mathcal{F}_t^s$ to obtain the word embedding $\\mathbf{w}^s$ and the sparse text embedding $\\mathbf{t}^s$ . Next, learnable mask tokens with position $\\mathbf{m}$ are concatenated with $\\mathbf{w}^s$ and input into the decoder for processing. The decoder outputs the reconstructed tokens $\\bar{\\mathbf{m}}$ , which are then averaged to obtain $\\bar{\\mathbf{t}}$ . Finally, we sum $\\mathbf{t}^s$ and $\\bar{\\mathbf{t}}$ to generate the final sparse text embedding $\\hat{\\mathbf{t}}^s$ . This process can be formulated as:", + "bbox": [ + 511, + 481, + 906, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {t}} ^ {s} = \\operatorname {D e c o d e r} \\left(\\left[ \\mathbf {w} ^ {s}, \\mathbf {m} \\right]\\right) + \\mathbf {t} ^ {s} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 704, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This reconstruction process enables the mask tokens to capture the contextual information missing from the sparse text, aligning it more closely with the dense text embedding. By combining the two, we obtain a more comprehensive and enriched text feature embedding.", + "bbox": [ + 511, + 734, + 905, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dense-to-Sparse Distillation. Through the dense text pretraining in stage 1, the representation of the pre-trained VSE model can capture richer semantics, and the pre-trained dense text encoder can be used as a teacher to distill the information capacity of dense text embeddings into sparse text embeddings. This Dense-to-Sparse distillation mini-", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21683", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mizes a negative cosine similarity between then dense text embedding $\\mathbf{t}^d$ and the final sparse text embedding $\\hat{\\mathbf{t}}^s$ :", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\mathbf {t} ^ {d}, \\hat {\\mathbf {t}} ^ {s}\\right) = 1 - \\frac {\\mathbf {t} ^ {d} \\cdot \\hat {\\mathbf {t}} ^ {s}}{\\left| \\left| \\mathbf {t} ^ {d} \\right| \\right| \\cdot \\left| \\left| \\hat {\\mathbf {t}} ^ {s} \\right| \\right|} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 132, + 482, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.2. Aligning Visual Semantic Using Sparse text", + "text_level": 1, + "bbox": [ + 89, + 178, + 439, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At the same time, we also fine-tune the VSE model on image-sparse text pair data to adapt the model to common sparse text queries in cross-modal retrieval. Given image $I$ and sparse text $T^s$ , we extract image embedding $\\mathbf{i}$ and sparse text embedding $\\mathbf{t}$ and optimize the VSE model as follow:", + "bbox": [ + 89, + 196, + 483, + 287 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {(I, T ^ {s}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {s}) + S (I, \\hat {T} ^ {s}) \\right] _ {+} \\tag {5} \\\\ + \\left[ \\alpha - S (I, T ^ {s}) + S (\\hat {I}, T ^ {s}) \\right] _ {+} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 295, + 483, + 362 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the final loss function is defined as follows to perform joint optimization of the two objectives.", + "bbox": [ + 89, + 372, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 415, + 482, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 444, + 223, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Settings", + "text_level": 1, + "bbox": [ + 89, + 469, + 284, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following previous works [9], we conduct our experiments on two widely used benchmark datasets: MS-COCO [21] and Flickr30K [30]. MS-COCO consists of 123,287 images, with each image annotated with five text captions. Following the dataset split in [9], we use 113,287 images for training, 5,000 for validation, and 5,000 for testing. Flickr30K contains 31,783 images, each paired with five textual descriptions. We adopt the standard split from [9], using 29,000 images for training, 1,000 for validation, and 1,000 for testing.", + "bbox": [ + 89, + 491, + 483, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For image-text matching, we use the commonly adopted Recall@K (where $\\mathrm{K} = 1,5,10$ ), denoted as R@1, R@5, and R@10. These metrics represent the percentage of ground-truth instances retrieved within the top 1, 5, or 10 results, respectively.", + "bbox": [ + 89, + 642, + 483, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 727, + 305, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For dense text generation, we use LLaVa [23] to generate dense textual descriptions. The hyperparameters for LLaVa are set as follows: Top-P = 0.9, temperature = 0.2, and max-new-tokens = 500. We generate a dense text description for each image in the MS-COCO and Flickr30K datasets, with each dense text sentence corresponding to $\\frac{1}{5}$ of the sparse text descriptions.", + "bbox": [ + 89, + 750, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the pre-training stage, we use the AdamW optimizer [28] to train the VSE model for 30 epochs, with a minibatch size of 128. The learning rate is initially set to 0.0005", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2ba2819b5c59a3b3c4cfc40e484d785478ce96423b0757c7eb08da51658148a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Pre-trained ResNet-152 on IN + GRU
VSE∞ [3]76.594.297.756.483.489.9498.1
MV-VSE [20]79.094.997.759.184.690.6505.9
NAAF* [42]81.996.198.361.085.390.6513.2
HREM* [10]81.496.598.560.985.691.3514.2
CHAN [29]79.794.597.360.285.390.7507.7
AVSE [26]82.396.298.562.888.092.8520.6
D2S-VSE82.596.699.062.688.193.1521.9
Swin-Base-224 + BERT-base
VSE++ [9]82.596.598.970.091.495.1534.4
SCAN [17]79.095.998.267.790.694.9526.3
SGR [7]80.497.098.766.990.294.5527.6
CHAN [29])81.497.098.668.590.694.5530.6
LAPS[11]82.497.499.570.091.795.4536.3
AVSE [26]83.997.499.470.092.495.6538.7
D2S-VSE87.298.499.973.093.596.7548.7
Swin-Base-384 + BERT-base
VSE++ [9]83.397.599.271.193.296.2540.6
SCAN [17]81.996.998.970.092.795.8536.1
SGR [7]80.796.899.069.991.795.3533.4
CHAN [29]81.296.798.870.392.295.9535.0
LAPS [11]85.197.799.274.093.096.3545.3
AVSE [26]87.198.399.273.693.596.5548.2
D2S-VSE87.899.099.775.794.196.9553.2
", + "bbox": [ + 517, + 89, + 903, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of experimental results using other common backbones on the Flickr30K dataset. The best performance is highlighted in **.** indicates ensemble methods.", + "bbox": [ + 511, + 500, + 906, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for the first few epochs and is reduced by a factor of 10 during the 9th, 15th, 20th, and 25th epochs. In the fine-tuning stage, we freeze the dense text encoder and fine-tune both the image encoder and sparse text encoder. We train them using the same training strategy as in the pre-training stage and evaluate the best model on the validation set.", + "bbox": [ + 511, + 566, + 906, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For model details, we use traditional region features and employ the recently popular Vision Transformer (ViT) [8] and Swin Transformer [27] as the backbone. We set the dimension of the shared embedding space to 512. For feature aggregation, we use GPO [3] for both the image encoder and the text encoder. The decoder consists of 4 layers of Transformer, with 4 attention heads per layer. We set 100 learnable tokens to predict the context of sparse text embeddings. We provide a more detailed implementation setup in the supplementary material.", + "bbox": [ + 511, + 657, + 908, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with Recent State-of-the-arts", + "text_level": 1, + "bbox": [ + 511, + 818, + 877, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our D2S-VSE with the most recent state-of-the-art approaches on two widely used datasets, i.e., Flirkr30K and MS-COCO. For the sake of fairness, our D2S-VSE employs a diverse range of backbones to facili", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21684", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b60994adbd0779743cbcb8b7edb52e9193de6f5c14ce216b1ed81dc0ecf39168.jpg", + "image_caption": [ + "(a) The impact of Learnable Tokens" + ], + "image_footnote": [], + "bbox": [ + 99, + 94, + 356, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/73ffa7a323a7e7ae6115f36ee47f3b0d9eba35a73f870d602ccbbb05bb24a59c.jpg", + "image_caption": [ + "(b) The impact of Decoder Depth", + "Figure 4. Ablation study of the architecture of decoder on Flickr30K. If not specified, the default is: the number of the learnable tokens is 100, the decoder has 4 layers and 4 head." + ], + "image_footnote": [], + "bbox": [ + 370, + 94, + 625, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/288db231dbc846eae6a3093b2cf3ce429dde15b05ae06a89d005f626f78f8544.jpg", + "image_caption": [ + "(c) The impact of Decoder Head" + ], + "image_footnote": [], + "bbox": [ + 643, + 95, + 897, + 200 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tate comparisons with various previous methods.", + "bbox": [ + 89, + 290, + 413, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 1, we present the results of using ViT-base [8] + BERT-base [6] as the backbone on the Flickr30K and MS-COCO datasets. Our D2S-VSE surpasses all state-of-the-art methods, demonstrating remarkable performance. In this setting, we compare our method against the two leading approaches, LAPS [11] and AVSE [26], and achieve superior results across nearly all metrics. For instance, on the Flickr30K dataset, our approach outperforms AVSE in text retrieval, achieving a $6.8\\%$ improvement in $\\mathsf{R}@\\mathsf{1}$ with the ViT-base-224 backbone and a $3.8\\%$ improvement with the ViT-base-384 backbone. Similarly, for image retrieval, our method achieves gains of $5.8\\%$ and $2.4\\%$ when using the ViT-base-224 and ViT-base-384 backbones, respectively. A consistent performance boost is also observed across all metrics on the MS-COCO dataset.", + "bbox": [ + 88, + 305, + 483, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we conducted experiments using other commonly used backbones, as shown in Table 2. Specifically, we evaluated our approach on the Flickr30K dataset with Faster R-CNN [1] + GRU and Swin-Base [27] + BERT-Base as backbones. The results further confirm that our method consistently outperforms existing approaches across different backbone architectures.", + "bbox": [ + 89, + 532, + 483, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 650, + 243, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of each component of our D2S-VSE, we conduct extensive ablation studies on the Flickr30K dataset. All experiments are performed using ViT-base-224 and BERT-base as the backbone to ensure a consistent and fair comparison.", + "bbox": [ + 89, + 672, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The impact of the two-stage training framework. Our D2S-VSE introduces a two-stage training framework to enhance the information capacity of image embedding and text embedding, and achieves very good results on commonly used datasets. In order to explore the role of each component in the framework, we conducted ablation experiments as shown in Table 3.", + "bbox": [ + 89, + 750, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is the model pre-trained with dense text inherently superior, leading to better performance? To verify this, we directly tested the pre-trained model on the Flickr30K test", + "bbox": [ + 89, + 856, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/faaabf235b01f9568d33a4cd6eac32c3b34d258abf5a10c511ac27913a33d09a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PretrainFinetuneText RetrievalImage RetrievalrSum
AlignDistillR@1R@5R@10R@1R@5R@10
✓56.284.791.641.771.379.7425.2
✓76.393.997.361.287.693.2509.5
✓✓79.296.598.366.890.594.4525.7
✓✓76.094.597.362.988.393.1512.1
✓✓✓82.896.198.368.591.394.9531.9
", + "bbox": [ + 524, + 287, + 897, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation experiment on the impact of our framework on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue.", + "bbox": [ + 511, + 404, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "set. We found that due to the gap between dense and sparse text, the pre-trained model struggles to adapt to sparse text, resulting in rSum of only 425.2 on the Flickr30K test set.", + "bbox": [ + 511, + 474, + 903, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The pre-training stage enhances the information capacity of image embeddings! We observed that the R@1 of text retrieval improved by $2.9\\%$ after pre-training compared to the model without pre-training. This suggests that dense text pre-training enhances the information capacity of image embeddings, enabling the model to better align multi-view descriptions.", + "bbox": [ + 511, + 521, + 905, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dense-to-Sparse feature distillation enhances the information capacity of sparse text embeddings! Applying dense-to-sparse distillation while aligning images during the fine-tuning stage further boosts the model's retrieval performance. On the Flickr30K dataset, text retrieval and image retrieval reached $82.8\\%$ and $68.5\\%$ , respectively. This improvement stems from distillation enhances the information capacity of text embeddings, leading to better semantic consistency between image and text embeddings.", + "bbox": [ + 511, + 627, + 905, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different type of distillation functions. We used L1 distance, L2 distance, and negative cosine similarity as the loss functions for Dense-to-Sparse distillation, respectively. As shown in Table 4, the effect of negative cosine similarity distillation is much better than the other two.", + "bbox": [ + 511, + 763, + 903, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoder design. The decoder is the most important part of Dense-to-Sparse distillation. Its existence allows learnable tokens to predict possible contexts based on the word embeddings of sparse text, thereby increasing the informa", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21685", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc488757593a235bb1dd6d7232488b96f4d365e05cb246425706934912cc9ca9.jpg", + "image_caption": [ + "(a) Prefix Learnable Tokens" + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 352, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24a8c85839378f2a7f0d6fe44c3020aa2eaaa0d2e5ca316f063dadfed9f34e1a.jpg", + "image_caption": [ + "(a) Surround Learnable Tokens", + "Figure 5. Different type of the position strategy of learnable tokens. Tokens at different positions can predict the semantics of different positions in dense text." + ], + "image_footnote": [], + "bbox": [ + 408, + 88, + 586, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48b9a4dcc44b8a830bf6286b971b80357f3b8d652d5e579ff6e5636c68344f86.jpg", + "image_caption": [ + "(c) Postfix Learnable Tokens" + ], + "image_footnote": [], + "bbox": [ + 645, + 89, + 823, + 146 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34e9e49bd2ed3035ff452df12a67e81ca5b3887cc98d8fb595274e26aab6fed3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
\\( \\ell_1 \\) Distance81.995.398.268.090.894.7528.9
\\( \\ell_2 \\) Distance82.195.698.467.991.295.1530.2
Negative Cosine Similarity82.896.198.368.591.394.9531.9
", + "bbox": [ + 99, + 224, + 475, + 305 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7e9a21f87fb3d5b4621f8fcb80f323805bd9e989210217149651b5ae1c32f278.jpg", + "table_caption": [ + "Table 4. Ablation experiment on the impact of different distillation functions on Flick30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + ], + "table_footnote": [], + "table_body": "
TypeText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Prefix81.795.398.268.191.294.8529.3
Postfix82.095.698.367.891.194.7529.5
Surround82.896.198.368.591.394.9531.9
", + "bbox": [ + 517, + 224, + 903, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation experiment on the impact of the position of the learnable tokens on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue.", + "bbox": [ + 511, + 314, + 906, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion capacity of sparse text features. As shown in Figure 4, we conduct ablation studies of Learnable tokens, decoder depth, and decoder head.", + "bbox": [ + 88, + 383, + 482, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The number of tokens determines the reconstruction capability of the decoder. As shown in Figure 4a, we observed that as the number of learnable tokens in the decoder increases, the model's performance in both image retrieval and text retrieval gradually improves, stabilizing when the number of tokens reaches approximately 100. At this point, the R@1 for text retrieval and image retrieval reaches $82.8\\%$ and $68.5\\%$ , respectively. However, when the number of tokens increases to 120, the model shows no significant advantage. Therefore, considering both performance and efficiency, we set the number of tokens to 100.", + "bbox": [ + 88, + 431, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Surrounding learnable tokens can better predict the context of sparse text. In addition, we conducted an ablation experiment on the insertion position of learnable tokens. Since our motivation stems from treating sparse text as masked dense text, with learnable tokens used to reconstruct the masked semantics, their placement is crucial. As shown in Figure 5, we designed Prefix Learnable Tokens, Surround Learnable Tokens, and Postfix Learnable Tokens to explore the impact of token positioning on model performance. Table 5 presents the ablation results on the impact of different learnable token positions. We observe that the rSum index of Surround Learnable Tokens is $2.6\\%$ and $2.4\\%$ higher than that of Prefix Learnable Tokens and Postfix Learnable Tokens, respectively. This outcome is intuitive, as the exact location of sparse text semantics within dense text is unknown. The Surround insertion method allows the model to predict semantics from both the preceding and following contexts simultaneously, leading to improved performance.", + "bbox": [ + 91, + 598, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A deep decoder can improve retrieval performance. As shown in Figure 4b, as the decoder depth increases, retrieval", + "bbox": [ + 89, + 871, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performance first improves and then declines, with the best results achieved when the number of layers is set to 4. Although our approach is inspired by Masked Signal Modeling [13], our findings regarding the decoder differ. In Masked Signal Modeling, the objective is to enhance the encoder through reconstruction, where an excessively deep decoder may hinder the encoder's capability. However, in our D2S-VSE, retrieval performance relies on the predictive ability of the decoder, meaning that a deeper decoder generally leads to better results.", + "bbox": [ + 511, + 386, + 906, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our D2S-VSE can work with very few decoder heads. As shown in Figure 4c, the number of attention heads has minimal impact on performance, with only 4 heads required to achieve optimal results. Increasing the number of heads does not lead to significant improvements, as all performance differences remain within $0.2\\%$ .", + "bbox": [ + 511, + 537, + 908, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 647, + 638, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we highlight the importance of embedding information capacity in addressing the multi-view description problem in image-text matching tasks. To tackle this challenge, we propose a novel Visual Semantic Embedding (VSE) framework, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE). The core idea of D2S-VSE is to enhance the information capacity of visual semantic embeddings by aligning dense text with images during pre-training. In the fine-tuning stage, we further refine sparse text embeddings through Dense-to-Sparse feature distillation, effectively improving their information capacity and leading to better image-text matching performance. Comprehensive experiments on two widely-used benchmarks validate the effectiveness of the proposed method, leading to state-of-the-art performance.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21686", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was partially supported by the National Major Scientific Instruments and Equipments Development Project of National Natural Science Foundation of China under Grant 62427820, the National Science Foundation of China under Grant 62376175 and 22494712, the 111 Project under Grant B21044, the Science Fund for Creative Research Groups of Sichuan Province Natural Science Foundation under Grant 2024NSFTD0035, and the National Science Foundation of Sichuan Province under Grant 2025ZNSFSC0480.", + "bbox": [ + 89, + 90, + 485, + 231 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 257, + 187, + 273 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 7", + "[2] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In CVPR, pages 12655-12663, 2020. 2", + "[3] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In CVPR, pages 15789-15798, 2021. 1, 2, 6", + "[4] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 2, 3", + "[5] Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio De Rezende, Yannis Kalantidis, and Diane Larlus. Probabilistic embeddings for cross-modal retrieval. In CVPR, pages 8415-8424, 2021. 2", + "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 7", + "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In AAAI, pages 1218-1226, 2021. 2, 5, 6", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5,6,7", + "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. In BMVC, 2018. 1, 2, 5, 6", + "[10] Zheren Fu, Zhendong Mao, Yan Song, and Yongdong Zhang. Learning semantic relationship among instances for image-text matching. In CVPR, pages 15159-15168, 2023. 6", + "[11] Zheren Fu, Lei Zhang, Hou Xia, and Zhendong Mao. Linguistic-aware patch slimming framework for fine-grained cross-modal alignment. In CVPR, pages 26307-26316, 2024. 5, 6, 7", + "[12] Jieuxiang Gu, Jianfei Cai, Shafiq R Joty, Li Niu, and Gang Wang. Look, imagine and match: Improving textual-visual" + ], + "bbox": [ + 93, + 282, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cross-modal retrieval with generative models. In CVPR, pages 7181-7189, 2018. 1", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 4, 5, 8", + "[14] Yan Huang, Qi Wu, Chunfeng Song, and Liang Wang. Learning semantic concepts and order for image and sentence matching. In CVPR, pages 6163-6171, 2018. 2", + "[15] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. In CVPR, pages 23422-23431, 2023. 2", + "[16] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 1", + "[17] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In ECCV, pages 201-216, 2018. 5, 6", + "[18] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In ICCV, pages 4654-4662, 2019. 1, 2", + "[19] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Image-text embedding learning via visual and textual semantic reasoning. TPAMI, 45(1):641-656, 2022. 2", + "[20] Zheng Li, Caili Guo, Zerun Feng, Jenq-Neng Hwang, and Xijun Xue. Multi-view visual semantic embedding. In IJ-CAI, 2022. 2, 6", + "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6", + "[22] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In CVPR, pages 10921-10930, 2020. 2", + "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 3, 4, 6", + "[24] Yang Liu, Hong Liu, Huaqiu Wang, and Mengyuan Liu. Regularizing visual semantic embedding with contrastive learning for image-text matching. IEEE SPL, 2022. 1", + "[25] Yang Liu, Hong Liu, Huaqiu Wang, Fanyang Meng, and Mengyuan Liu. Bcan: Bidirectional correct attention network for cross-modal retrieval. TNNLS, 2023. 2", + "[26] Yang Liu, Mengyuan Liu, shudong Huang, and Jiancheng Lv. Asymmetric visual semantic embedding framework for efficient vision-language alignment. In AAAI, 2025. 2, 3, 5, 6, 7", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6", + "[29] Zhengxin Pan, Fangyu Wu, and Bailing Zhang. Fine-grained image-text matching by cross-modal hard aligning network. In CVPR, pages 19275-19284, 2023. 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21687", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 6", + "[31] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In ACMMM, pages 4948-4956, 2022. 1", + "[32] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. arXiv preprint arXiv:2310.17468, 2023. 1", + "[33] Leigang Qu, Meng Liu, Da Cao, Liqiang Nie, and Qi Tian. Context-aware multi-view summarization network for image-text matching. In ACMMM, pages 1047-1055, 2020. 2", + "[34] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In SIGIR, pages 1104-1113, 2021. 2", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 1", + "[36] Yale Song and Mohammad Soleymani. Polysemous visual-semantic embedding for cross-modal retrieval. In CVPR, pages 1979–1988, 2019. 2", + "[37] Haoran Wang, Ying Zhang, Zhong Ji, Yanwei Pang, and Lin Ma. Consensus-aware visual-semantic embedding for image-text matching. In ECCV, pages 18–34, 2020. 2", + "[38] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In ICCV, pages 5764-5773, 2019. 2", + "[39] Jiwei Wei, Xing Xu, Yang Yang, Yanli Ji, Zheng Wang, and Heng Tao Shen. Universal weighting metric learning for cross-modal matching. In CVPR, pages 13005-13014, 2020. 2", + "[40] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pretraining for long text understanding. Neurips, 37:64996-65019, 2024. 3", + "[41] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, pages 310–325. Springer, 2024. 3", + "[42] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In CVPR, pages 15661-15670, 2022. 2, 6", + "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In CVPR, pages 3536-3545, 2020. 2", + "[44] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamlip: Language-image pre-training with long captions. In ECCV, pages 73-90. Springer, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21688", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_model.json b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d935f8ad2fd28c12911e36c587988ff7ee40868 --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_model.json @@ -0,0 +1,2057 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.171, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.203, + 0.816, + 0.223 + ], + "angle": 0, + "content": "Yang Liu\\(^{1,2}\\), Wentao Feng\\(^{1,2}\\), Zhuoyao Liu\\(^{1,2}\\), Shudong Huang\\(^{1,2,*}\\), Jiancheng Lv\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.224, + 0.729, + 0.239 + ], + "angle": 0, + "content": "1College of Computer Science, Sichuan University, China" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.241, + 0.843, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Engineering Research Center of Machine Learning and Industry Intelligence, China" + }, + { + "type": "text", + "bbox": [ + 0.374, + 0.26, + 0.599, + 0.274 + ], + "angle": 0, + "content": "https://d2s-vse.github.io" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.675 + ], + "angle": 0, + "content": "Enabling Visual Semantic Models to effectively handle multi-view description matching has been a longstanding challenge. Existing methods typically learn a set of embeddings to find the optimal match for each view's text and compute similarity. However, the visual and text embeddings learned through these approaches have limited information capacity and are prone to interference from locally similar negative samples. To address this issue, we argue that the information capacity of embeddings is crucial and propose Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE), which enhances the information capacity of sparse text by leveraging dense text distillation. Specifically, D2S-VSE is a two-stage framework. In the pretraining stage, we align images with dense text to enhance the information capacity of visual semantic embeddings. In the fine-tuning stage, we optimize two tasks simultaneously, distilling dense text embeddings to sparse text embeddings while aligning images and sparse texts, enhancing the information capacity of sparse text embeddings. Our proposed D2S-VSE model is extensively evaluated on the large-scale MS-COCO and Flickr30K datasets, demonstrating its superiority over recent state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.702, + 0.222, + 0.717 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.879 + ], + "angle": 0, + "content": "In the domain of vision and language, the alignment of vision and language stands as a foundational task, with the core objective of establishing a robust and precise bridge between these two distinct yet interconnected modalities. The traditional approach to cross-modal retrieval [12, 24, 31, 32] follows a standard procedure: encoding images and texts into a shared embedding space and measuring their similarity. Typically, models are optimized using a triplet loss, which ensures that matched image-text pairs exhibit higher similarity than unmatched pairs [9, 16]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.308, + 0.907, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.792 + ], + "angle": 0, + "content": "Figure 1. (Top) In previous methods, when addressing problems with varying information density, a set of visual embeddings is learned to match the most similar text embeddings. However, this approach results in learned embeddings with limited information capacity. (Mid) Since dense text has a higher information density than sparse text, training with aligned dense text results in image embeddings with greater information capacity. (Bottom) Sparse text embedding can be distilled through dense text embedding to enhance its information capacity, enabling the matching of text descriptions from different perspectives with a single visual embedding. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "However, most traditional Visual Semantic Embedding (VSE) models overlook the inherent differences between image and text modalities [3, 18, 35]. Specifically, visual signals represent natural, objective records of real-world phenomena, whereas textual signals are inherently seman-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.238, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "21679" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.213 + ], + "angle": 0, + "content": "tic and subjective, shaped by human interpretation. Consequently, the same image may be described in various ways depending on individual perspectives. Existing studies refer to this issue as the \"ambiguity issue\" [4, 36], \"sparse annotations\" [15] and \"multi-view descriptions\" [20, 33]. But, at its core, the problem stems from the inherent differences between image and text modalities, particularly the disparity in \"information density\" [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.483, + 0.46 + ], + "angle": 0, + "content": "Recent studies [4, 20, 26, 36] propose learning a sets of embeddings from images/texts, where each embedding captures specific semantics. For each image-text pair, the most relevant embedding is selected for similarity calculation. While this approach reduces computational overhead and partially addresses information density challenges, we identify a critical limitation: As shown in the figure 1, the text in the dataset is usually short and concise. The image and text embeddings learned in this way can only capture a limited subset of features, resulting in limited information capacity being learned, and are therefore easily disturbed by negative samples with similar local semantics. Given this, we argue that the information capacity of visual semantic embeddings is crucial for the image-text matching task. The key question, then, is: how can we learn visual semantic embeddings with higher information capacity?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel approach, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE), which enhances visual semantic embeddings by aligning both semantics and information capacity. Our key hypothesis posits that text embeddings derived from dense image captions inherently possess greater information capacity compared than those from sparse textual annotations in traditional datasets such as Flickr30K and COCO. Specifically, we introduce a two-stage training framework. In the pre-train stage, we generate dense descriptions for images using LLaVa and pre-train image-dense description pairs to align images with dense textual descriptions in the embedding space. This alignment enables the model to extract more comprehensive features, resulting in embeddings with higher information capacity. In the fine-tuning stage, we introduce a dense-to-sparse feature distillation method. During this phase, while maintaining alignment between images and sparse text annotations from the dataset, we additionally distill knowledge from dense text embeddings into sparse text embeddings to enhance their information capacity. We formulate this distillation process as a masked signal reconstruction task for sparse text embeddings. To achieve this, we implement a Transformer-based decoder module that follows the text encoder. This decoder processes short text features combined with learnable mask tokens, aiming to predict the latent context of the sparse text and thus enhance its information capacity. And in the inference phase, we only require the image and the sparse text from the dataset, without the need for dense text, thus in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.82, + 0.106 + ], + "angle": 0, + "content": "curring no additional computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.108, + 0.872, + 0.121 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.214 + ], + "angle": 0, + "content": "- Importance of Information Capacity: We identify the critical role of enhancing the information capacity of visual semantic embeddings. Based on this insight, we propose a novel framework to improve the VSE model by increasing the information capacity of visual semantic embeddings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.365 + ], + "angle": 0, + "content": "- Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE): We introduce D2S-VSE, a two-stage training approach that leverages dense text to enhance the information capacity of image embeddings in the first pre-training stage. In the second stage, a distillation branch is incorporated to improve the information capacity of sparse text embedding by distilling dense text embedding from sparse text embedding, ensuring both semantic alignment and balanced information capacity between image and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.44 + ], + "angle": 0, + "content": "- State-of-the-Art Performance: We evaluate D2S-VSE with existing fine-grained methods across various model backbones. D2S-VSE outperforms the latest state-of-the-art methods on image-text retrieval benchmarks, including Flickr30K and MS-COCO." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.124, + 0.905, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.458, + 0.655, + 0.473 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.715, + 0.501 + ], + "angle": 0, + "content": "2.1. Image-Text Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Existing Image-text retrieval approaches can be broadly categorized into two groups based on the granularity of the matching similarity: Global-level matching and Local-level matching methods. Global-level matching methods project images and texts into a shared embedding space, computing similarity using improved loss functions [5, 9] to bring semantically aligned pairs closer together. To improve the quality of the embedding space, recent related works proposed several more sophisticated networks, such as graph convolutional network [22, 37], Generalized Pooling Operator [3], and other model architectures [14, 18, 19]. In order to facilitate the closer alignment of semantically corresponding image-text pairs. Different from global matching methods, local matching methods[2, 7, 25, 34, 38, 39, 42, 43] focus more on the alignment between regions. Most of them capture the finer-grained relationship between images and texts by adopting cross-modal networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Information Density Issues. Recent studies have proposed various methods to address the issue of differing information density between images and text. PCME [5] proposed a probabilistic cross-modal embedding to properly represent the one-to-many relationships in joint embedding. DivE [15] introduced a slot attention-based set prediction module, where the elements of the embedding set compete it-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21680" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.904, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.479, + 0.907, + 0.568 + ], + "angle": 0, + "content": "Figure 2. Overview of our D2S-VSE framework. (a) Dense Text Generation: D2S-VSE utilizes the LLaVa [23] to generate dense text descriptions for images in the dataset that contain all the details of the image. (b) Pre-train With Dense Descriptions: D2S-VSE first uses contrastive learning to pre-train the model on image-dense-text pairs to enhance the information capacity of visual semantic embeddings. (c) Finetune With the Sparse Descriptions: D2S-VSE performs distillation from dense to sparse while contrasting the alignment, allowing sparse text embeddings to learn more details of dense text embeddings, thereby improving the information capacity of sparse text embeddings and better aligning them with image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.666 + ], + "angle": 0, + "content": "eratively to aggregate the input data. Under the framework of PCME, PCME++ [4] introduced a new probabilistic distance with a closed-form solution. AVSE [26] proposed a matching module for compute the dynamic similarity between asymmetric multi-view image and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.711 + ], + "angle": 0, + "content": "These methods learn a set of embeddedings to match ambiguous samples, but reduce the information capacity of each embedding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.424, + 0.743 + ], + "angle": 0, + "content": "2.2. Vision language model with dense text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In the field of vision-language research, few studies have leveraged long-text descriptions to enhance the multimodal representation of images. DreamLip [44] utilizes a multimodal large language model to generate dense text descriptions and randomly extracts sentence fragments for pre-training, but it does not fully exploit long-text pretraining. Long-CLIP [41] fine-tunes CLIP-pre-trained on a short-text-image dataset—on long-text-image pairs to unlock long-text capabilities, yet it overlooks short-text effectiveness. LoTLIP [40] directly pre-trains on long-text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.907, + 0.633 + ], + "angle": 0, + "content": "image data and introduces corner tokens to aggregate diverse textual information, preserving the ability to handle short text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.695 + ], + "angle": 0, + "content": "Different from the above methods that focus on pretraining, our method utilizes a very small amount of long text for pre-training, aiming to use the long text to enhance the information capacity of visual semantic embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.907, + 0.743 + ], + "angle": 0, + "content": "3. Dense-to-Sparse Feature Distilled Visual Semantic Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our goal is to enhance Visual Semantic Embedding (VSE) models to address the challenge of inconsistent information density between images and text. Existing VSE models typically align images with the sparse textual descriptions provided in datasets. This approach biases the image encoder toward textual perspectives, limiting its ability to fully comprehend visual content and making it difficult to handle ambiguous retrieval cases. To overcome this issue, we propose a two-stage training procedure. An overview of our D2S-VSE framework is illustrated in Figure 2." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "- Stage I. We leverage the pre-trained multimodal large model LLaVa to generate dense textual descriptions for images. The VSE model is then pre-trained using these image-dense text pairs. By training on data with comparable information density, the model learns embeddings with greater information capacity, enabling a more comprehensive understanding of visual content (Sec. 3.1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.198, + 0.482, + 0.333 + ], + "angle": 0, + "content": "- **Stage II.** We fine-tune the pre-trained VSE model on general datasets containing image-sparse text pairs. Simultaneously, we distill knowledge from the dense text features into the sparse text representations, enhancing their information capacity. While aligning image-text embeddings, we also align the information capacity between visual and textual modalities. This enables the VSE model to effectively handle varying information densities in images and texts (Sec. 3.2)." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.347, + 0.411, + 0.364 + ], + "angle": 0, + "content": "3.1. Pre-training with Dense Descriptions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.484, + 0.477 + ], + "angle": 0, + "content": "To address the issue of inconsistent information density between images and text, we introduce a pre-training stage where the Visual Semantic Embedding (VSE) model learns from image-dense text pairs. This stage enhances the model's ability to extract rich semantic representations from images by ensuring that textual descriptions contain a level of detail comparable to the visual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.711 + ], + "angle": 0, + "content": "Dense Text Generation. To generate dense textual descriptions for images, we leverage the pre-trained multimodal large model LLaVa [23], which integrates vision and language understanding to generate detailed descriptions based on image inputs. Unlike traditional dataset texts, which are often brief and sparse, LLaVa produces detailed descriptions that include object attributes, relationships, actions, and contextual information. Given a raw image \\(I\\), we feed it into the LLaVa with the prompt \"Describe this image in detail\". LLaVa then produces a dense textual description that captures fine-grained details of the image. The hyperparameters used for LLaVa are detailed in Sec 4.2. This process ensures that the textual input provides a comprehensive representation of the image, reducing the information density gap between modalities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.482, + 0.809 + ], + "angle": 0, + "content": "Image-Dense-Caption Pre-train. Once the dense captions are generated, we pre-train the VSE model using these image-dense text pairs. The objective of this stage is to enable the model to learn embeddings that effectively capture fine-grained visual semantics and their corresponding textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We employ a contrastive learning framework where the image encoder \\(\\mathcal{F}_i\\) and the dense text encoder \\(\\mathcal{F}_t^d\\) map images and dense textual descriptions into a shared embedding space. Formally, given a dense textual description \\(T^d\\) and a raw image as \\(I\\), our goal is to map these inputs into compact vectors \\(\\mathbf{t}^d \\in \\mathbb{R}^d\\) and \\(\\mathbf{i} \\in \\mathbb{R}^d\\) such that their semantic" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.903, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.384, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Figure 3. Architecture for the sparse text branch. Similar to the decoder in MAE [13], we introduce a set of learnable tokens in the sparse text branch to predict the implicit context of sparse text through the decoder. The learnable tokens help in reconstructing the missing semantics, allowing the model to better capture the underlying context and relationships within the sparse text." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.907, + 0.54 + ], + "angle": 0, + "content": "relationships can be quantified through a similarity metric \\( S(\\cdot, \\cdot) \\). Specifically, the similarity function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.552, + 0.906, + 0.587 + ], + "angle": 0, + "content": "\\[\nS (I, T) = \\frac {\\mathbf {i} ^ {T} \\mathbf {t} ^ {d}}{| | \\mathbf {i} | | \\cdot | | \\mathbf {t} ^ {d} | |} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.906, + 0.669 + ], + "angle": 0, + "content": "To optimize the model, we use triplet loss function with the hardest negative samples to minimize the similarity between relevant pairs and to maximize the distance between irrelevant ones. The optimized loss function is mathematically formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.68, + 0.905, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {p r e t r a i n}} = \\sum_ {(I, T ^ {d}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {d}) + S (I, \\hat {T} ^ {d}) \\right] _ {+} \\tag {2} \\\\ + \\left[ \\alpha - S (I, T ^ {d}) + S (\\hat {I}, T ^ {d}) \\right] _ {+} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.759, + 0.907, + 0.824 + ], + "angle": 0, + "content": "where \\(\\alpha\\) serves as a margin parameter and \\([x]_{+} \\equiv Max(x,0)\\). In dataset \\(\\mathcal{D}\\), the similarity in a positive pair \\(S(I,T^{d})\\) should be higher than that in the hardest negative pairs \\(S(\\hat{I},T^{d})\\) and \\(S(I,\\hat{T}^{d})\\) by a margin \\(\\alpha\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.81, + 0.85 + ], + "angle": 0, + "content": "3.2. Finetune With Sparse Description" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Stage I pre-training with dense descriptions enhances the information capacity of embeddings of the VSE model, but there exists a gap between the training and testing data, as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21682" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.9, + 0.416 + ], + "angle": 0, + "content": "
MethodFGFlickr30k 1K TestMS-COCO 5-fold 1K TestMS-COCO 5K Test
Text RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSum
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
ViT-Base-224 + BERT-base
VSE++ [9]X71.892.896.559.484.790.9496.175.094.698.062.789.494.9514.662.387.693.443.973.683.3413.4
SCAN [17]✓69.590.995.656.483.190.0485.676.095.498.164.590.895.8520.653.981.890.042.972.382.5423.5
SGR [7]✓69.790.895.259.184.189.9488.777.295.098.065.190.795.8521.854.982.890.542.872.282.5425.8
CHAN [29]✓69.291.895.058.484.990.6489.977.195.198.165.091.096.0522.256.383.290.143.072.682.8428.0
LAPS [11]✓74.093.497.462.587.392.7507.378.795.598.366.291.396.2526.357.584.090.844.574.083.6434.4
AVSE [26]X76.094.697.562.788.493.1512.379.895.698.367.091.596.3528.558.884.391.045.174.383.9437.4
D2S-VSEX82.896.198.368.591.394.9531.980.197.099.268.192.596.7533.760.185.592.546.375.985.2445.6
ViT-Base-384 + BERT-base
VSE++ [9]X77.195.797.565.890.294.3520.577.095.798.464.691.196.2523.054.982.890.442.472.482.8425.8
SCAN [17]✓75.494.496.963.688.693.5512.576.195.598.565.191.696.3523.153.381.890.042.672.682.9423.1
SGR [7]✓76.994.998.164.288.493.3515.875.895.798.665.692.096.5524.253.381.089.642.973.183.7423.6
CHAN [29]✓75.494.597.663.288.693.1512.478.195.898.666.182.196.6527.355.683.891.243.473.683.5431.1
LAPS [11]✓79.096.098.167.390.594.5525.478.696.398.968.092.496.8531.057.484.992.546.475.885.2442.2
AVSE [26]X80.396.498.767.991.294.7529.281.197.199.068.392.797.0535.261.286.893.246.275.985.0448.3
D2S-VSEX84.197.599.170.391.695.3537.980.897.299.169.092.996.8535.860.686.593.246.876.485.7449.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.427, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Table 1. Comparison of experimental results using ViT-Base [8] and BERT [6] the MS-COCO and Flickr30K datasets. \"FG\" indicates whether fine-grained cross-modal alignment is applied. The best performance is highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.482, + 0.482, + 0.693 + ], + "angle": 0, + "content": "the model is trained on dense texts and tested on sparse texts, leading to suboptimal performance. However, simply fine-tuning directly on the image-sparse text pair data will not make the sparse text embedding contain richer semantics, that is, more information capacity. To address this problem, we introduce a novel distillation mechanism while aligning, which treats sparse text as masked dense text and designs a decoder. Through distillation, the decoder can decode sparse text embeddings into dense text embeddings, thereby improving the information capacity of sparse text embeddings. This mechanism extracts sparse text embeddings from dense text embeddings while aligning images and text, thereby enhancing the information capacity of sparse text features." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.731 + ], + "angle": 0, + "content": "3.2.1. Aligning Information Capacity by Dense-to-Sparse Feature Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "To ensure that the sparse text embedding retains as much information as the dense text embedding, we introduce a feature distillation mechanism from dense to sparse. Inspired by masked signal modeling [6, 13], we treat sparse text as a masked version of dense text. This allows us to decode and reconstruct the sparse text embedding through feature distillation, enabling the reconstructed embedding to approximate the information capacity of the dense text embedding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Transformer-based Decoder. To achieve this reconstruction, we design a Transformer-based decoder, as illustrated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.693 + ], + "angle": 0, + "content": "in Figure 3. The decoder plays a pivotal role in D2S-VSE, responsible for reconstructing the dense text embedding from the sparse text embedding by leveraging feature distillation. It learns to restore the missing contextual information in sparse text, ensuring that the final sparse text embedding retains as much semantic richness as the dense text embedding. Given a sparse text \\( T^s \\), we first feed it into the sparse text encoder \\( \\mathcal{F}_t^s \\) to obtain the word embedding \\( \\mathbf{w}^s \\) and the sparse text embedding \\( \\mathbf{t}^s \\). Next, learnable mask tokens with position \\( \\mathbf{m} \\) are concatenated with \\( \\mathbf{w}^s \\) and input into the decoder for processing. The decoder outputs the reconstructed tokens \\( \\bar{\\mathbf{m}} \\), which are then averaged to obtain \\( \\bar{\\mathbf{t}} \\). Finally, we sum \\( \\mathbf{t}^s \\) and \\( \\bar{\\mathbf{t}} \\) to generate the final sparse text embedding \\( \\hat{\\mathbf{t}}^s \\). This process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.705, + 0.907, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {t}} ^ {s} = \\operatorname {D e c o d e r} \\left(\\left[ \\mathbf {w} ^ {s}, \\mathbf {m} \\right]\\right) + \\mathbf {t} ^ {s} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.809 + ], + "angle": 0, + "content": "This reconstruction process enables the mask tokens to capture the contextual information missing from the sparse text, aligning it more closely with the dense text embedding. By combining the two, we obtain a more comprehensive and enriched text feature embedding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dense-to-Sparse Distillation. Through the dense text pretraining in stage 1, the representation of the pre-trained VSE model can capture richer semantics, and the pre-trained dense text encoder can be used as a teacher to distill the information capacity of dense text embeddings into sparse text embeddings. This Dense-to-Sparse distillation mini-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21683" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "mizes a negative cosine similarity between then dense text embedding \\(\\mathbf{t}^d\\) and the final sparse text embedding \\(\\hat{\\mathbf{t}}^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.133, + 0.483, + 0.171 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\mathbf {t} ^ {d}, \\hat {\\mathbf {t}} ^ {s}\\right) = 1 - \\frac {\\mathbf {t} ^ {d} \\cdot \\hat {\\mathbf {t}} ^ {s}}{\\left| \\left| \\mathbf {t} ^ {d} \\right| \\right| \\cdot \\left| \\left| \\hat {\\mathbf {t}} ^ {s} \\right| \\right|} \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.179, + 0.441, + 0.195 + ], + "angle": 0, + "content": "3.2.2. Aligning Visual Semantic Using Sparse text" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.485, + 0.288 + ], + "angle": 0, + "content": "At the same time, we also fine-tune the VSE model on image-sparse text pair data to adapt the model to common sparse text queries in cross-modal retrieval. Given image \\(I\\) and sparse text \\(T^s\\), we extract image embedding \\(\\mathbf{i}\\) and sparse text embedding \\(\\mathbf{t}\\) and optimize the VSE model as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.296, + 0.484, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {(I, T ^ {s}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {s}) + S (I, \\hat {T} ^ {s}) \\right] _ {+} \\tag {5} \\\\ + \\left[ \\alpha - S (I, T ^ {s}) + S (\\hat {I}, T ^ {s}) \\right] _ {+} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.404 + ], + "angle": 0, + "content": "In summary, the final loss function is defined as follows to perform joint optimization of the two objectives." + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.416, + 0.483, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.224, + 0.461 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.285, + 0.485 + ], + "angle": 0, + "content": "4.1. Dataset and Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Datasets. Following previous works [9], we conduct our experiments on two widely used benchmark datasets: MS-COCO [21] and Flickr30K [30]. MS-COCO consists of 123,287 images, with each image annotated with five text captions. Following the dataset split in [9], we use 113,287 images for training, 5,000 for validation, and 5,000 for testing. Flickr30K contains 31,783 images, each paired with five textual descriptions. We adopt the standard split from [9], using 29,000 images for training, 1,000 for validation, and 1,000 for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Evaluation Metrics. For image-text matching, we use the commonly adopted Recall@K (where \\(\\mathrm{K} = 1,5,10\\)), denoted as R@1, R@5, and R@10. These metrics represent the percentage of ground-truth instances retrieved within the top 1, 5, or 10 results, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.306, + 0.744 + ], + "angle": 0, + "content": "4.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "For dense text generation, we use LLaVa [23] to generate dense textual descriptions. The hyperparameters for LLaVa are set as follows: Top-P = 0.9, temperature = 0.2, and max-new-tokens = 500. We generate a dense text description for each image in the MS-COCO and Flickr30K datasets, with each dense text sentence corresponding to \\(\\frac{1}{5}\\) of the sparse text descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the pre-training stage, we use the AdamW optimizer [28] to train the VSE model for 30 epochs, with a minibatch size of 128. The learning rate is initially set to 0.0005" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.49 + ], + "angle": 0, + "content": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Pre-trained ResNet-152 on IN + GRU
VSE∞ [3]76.594.297.756.483.489.9498.1
MV-VSE [20]79.094.997.759.184.690.6505.9
NAAF* [42]81.996.198.361.085.390.6513.2
HREM* [10]81.496.598.560.985.691.3514.2
CHAN [29]79.794.597.360.285.390.7507.7
AVSE [26]82.396.298.562.888.092.8520.6
D2S-VSE82.596.699.062.688.193.1521.9
Swin-Base-224 + BERT-base
VSE++ [9]82.596.598.970.091.495.1534.4
SCAN [17]79.095.998.267.790.694.9526.3
SGR [7]80.497.098.766.990.294.5527.6
CHAN [29])81.497.098.668.590.694.5530.6
LAPS[11]82.497.499.570.091.795.4536.3
AVSE [26]83.997.499.470.092.495.6538.7
D2S-VSE87.298.499.973.093.596.7548.7
Swin-Base-384 + BERT-base
VSE++ [9]83.397.599.271.193.296.2540.6
SCAN [17]81.996.998.970.092.795.8536.1
SGR [7]80.796.899.069.991.795.3533.4
CHAN [29]81.296.798.870.392.295.9535.0
LAPS [11]85.197.799.274.093.096.3545.3
AVSE [26]87.198.399.273.693.596.5548.2
D2S-VSE87.899.099.775.794.196.9553.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.5, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Table 2. Comparison of experimental results using other common backbones on the Flickr30K dataset. The best performance is highlighted in **.** indicates ensemble methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.657 + ], + "angle": 0, + "content": "for the first few epochs and is reduced by a factor of 10 during the 9th, 15th, 20th, and 25th epochs. In the fine-tuning stage, we freeze the dense text encoder and fine-tune both the image encoder and sparse text encoder. We train them using the same training strategy as in the pre-training stage and evaluate the best model on the validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.81 + ], + "angle": 0, + "content": "For model details, we use traditional region features and employ the recently popular Vision Transformer (ViT) [8] and Swin Transformer [27] as the backbone. We set the dimension of the shared embedding space to 512. For feature aggregation, we use GPO [3] for both the image encoder and the text encoder. The decoder consists of 4 layers of Transformer, with 4 attention heads per layer. We set 100 learnable tokens to predict the context of sparse text embeddings. We provide a more detailed implementation setup in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.879, + 0.835 + ], + "angle": 0, + "content": "4.3. Comparisons with Recent State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We compare our D2S-VSE with the most recent state-of-the-art approaches on two widely used datasets, i.e., Flirkr30K and MS-COCO. For the sake of fairness, our D2S-VSE employs a diverse range of backbones to facili" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21684" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.095, + 0.357, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.211, + 0.323, + 0.223 + ], + "angle": 0, + "content": "(a) The impact of Learnable Tokens" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.095, + 0.627, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.211, + 0.588, + 0.224 + ], + "angle": 0, + "content": "(b) The impact of Decoder Depth" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.096, + 0.898, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.211, + 0.856, + 0.223 + ], + "angle": 0, + "content": "(c) The impact of Decoder Head" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.908, + 0.264 + ], + "angle": 0, + "content": "Figure 4. Ablation study of the architecture of decoder on Flickr30K. If not specified, the default is: the number of the learnable tokens is 100, the decoder has 4 layers and 4 head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.414, + 0.305 + ], + "angle": 0, + "content": "tate comparisons with various previous methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.306, + 0.484, + 0.531 + ], + "angle": 0, + "content": "In Table 1, we present the results of using ViT-base [8] + BERT-base [6] as the backbone on the Flickr30K and MS-COCO datasets. Our D2S-VSE surpasses all state-of-the-art methods, demonstrating remarkable performance. In this setting, we compare our method against the two leading approaches, LAPS [11] and AVSE [26], and achieve superior results across nearly all metrics. For instance, on the Flickr30K dataset, our approach outperforms AVSE in text retrieval, achieving a \\(6.8\\%\\) improvement in \\(\\mathsf{R}@\\mathsf{1}\\) with the ViT-base-224 backbone and a \\(3.8\\%\\) improvement with the ViT-base-384 backbone. Similarly, for image retrieval, our method achieves gains of \\(5.8\\%\\) and \\(2.4\\%\\) when using the ViT-base-224 and ViT-base-384 backbones, respectively. A consistent performance boost is also observed across all metrics on the MS-COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Additionally, we conducted experiments using other commonly used backbones, as shown in Table 2. Specifically, we evaluated our approach on the Flickr30K dataset with Faster R-CNN [1] + GRU and Swin-Base [27] + BERT-Base as backbones. The results further confirm that our method consistently outperforms existing approaches across different backbone architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.245, + 0.667 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.748 + ], + "angle": 0, + "content": "To verify the effectiveness of each component of our D2S-VSE, we conduct extensive ablation studies on the Flickr30K dataset. All experiments are performed using ViT-base-224 and BERT-base as the backbone to ensure a consistent and fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.854 + ], + "angle": 0, + "content": "The impact of the two-stage training framework. Our D2S-VSE introduces a two-stage training framework to enhance the information capacity of image embedding and text embedding, and achieves very good results on commonly used datasets. In order to explore the role of each component in the framework, we conducted ablation experiments as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Is the model pre-trained with dense text inherently superior, leading to better performance? To verify this, we directly tested the pre-trained model on the Flickr30K test" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.288, + 0.898, + 0.394 + ], + "angle": 0, + "content": "
PretrainFinetuneText RetrievalImage RetrievalrSum
AlignDistillR@1R@5R@10R@1R@5R@10
✓56.284.791.641.771.379.7425.2
✓76.393.997.361.287.693.2509.5
✓✓79.296.598.366.890.594.4525.7
✓✓76.094.597.362.988.393.1512.1
✓✓✓82.896.198.368.591.394.9531.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 3. Ablation experiment on the impact of our framework on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.521 + ], + "angle": 0, + "content": "set. We found that due to the gap between dense and sparse text, the pre-trained model struggles to adapt to sparse text, resulting in rSum of only 425.2 on the Flickr30K test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.627 + ], + "angle": 0, + "content": "The pre-training stage enhances the information capacity of image embeddings! We observed that the R@1 of text retrieval improved by \\(2.9\\%\\) after pre-training compared to the model without pre-training. This suggests that dense text pre-training enhances the information capacity of image embeddings, enabling the model to better align multi-view descriptions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Dense-to-Sparse feature distillation enhances the information capacity of sparse text embeddings! Applying dense-to-sparse distillation while aligning images during the fine-tuning stage further boosts the model's retrieval performance. On the Flickr30K dataset, text retrieval and image retrieval reached \\(82.8\\%\\) and \\(68.5\\%\\), respectively. This improvement stems from distillation enhances the information capacity of text embeddings, leading to better semantic consistency between image and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.839 + ], + "angle": 0, + "content": "Different type of distillation functions. We used L1 distance, L2 distance, and negative cosine similarity as the loss functions for Dense-to-Sparse distillation, respectively. As shown in Table 4, the effect of negative cosine similarity distillation is much better than the other two." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Decoder design. The decoder is the most important part of Dense-to-Sparse distillation. Its existence allows learnable tokens to predict possible contexts based on the word embeddings of sparse text, thereby increasing the informa" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21685" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.089, + 0.353, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.15, + 0.337, + 0.162 + ], + "angle": 0, + "content": "(a) Prefix Learnable Tokens" + }, + { + "type": "image", + "bbox": [ + 0.409, + 0.089, + 0.588, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.15, + 0.582, + 0.162 + ], + "angle": 0, + "content": "(a) Surround Learnable Tokens" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.09, + 0.824, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.15, + 0.812, + 0.162 + ], + "angle": 0, + "content": "(c) Postfix Learnable Tokens" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.175, + 0.908, + 0.204 + ], + "angle": 0, + "content": "Figure 5. Different type of the position strategy of learnable tokens. Tokens at different positions can predict the semantics of different positions in dense text." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.226, + 0.477, + 0.306 + ], + "angle": 0, + "content": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
\\( \\ell_1 \\) Distance81.995.398.268.090.894.7528.9
\\( \\ell_2 \\) Distance82.195.698.467.991.295.1530.2
Negative Cosine Similarity82.896.198.368.591.394.9531.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Table 4. Ablation experiment on the impact of different distillation functions on Flick30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.226, + 0.905, + 0.306 + ], + "angle": 0, + "content": "
TypeText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Prefix81.795.398.268.191.294.8529.3
Postfix82.095.698.367.891.194.7529.5
Surround82.896.198.368.591.394.9531.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.315, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Table 5. Ablation experiment on the impact of the position of the learnable tokens on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.483, + 0.431 + ], + "angle": 0, + "content": "tion capacity of sparse text features. As shown in Figure 4, we conduct ablation studies of Learnable tokens, decoder depth, and decoder head." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.432, + 0.483, + 0.598 + ], + "angle": 0, + "content": "The number of tokens determines the reconstruction capability of the decoder. As shown in Figure 4a, we observed that as the number of learnable tokens in the decoder increases, the model's performance in both image retrieval and text retrieval gradually improves, stabilizing when the number of tokens reaches approximately 100. At this point, the R@1 for text retrieval and image retrieval reaches \\(82.8\\%\\) and \\(68.5\\%\\), respectively. However, when the number of tokens increases to 120, the model shows no significant advantage. Therefore, considering both performance and efficiency, we set the number of tokens to 100." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.599, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Surrounding learnable tokens can better predict the context of sparse text. In addition, we conducted an ablation experiment on the insertion position of learnable tokens. Since our motivation stems from treating sparse text as masked dense text, with learnable tokens used to reconstruct the masked semantics, their placement is crucial. As shown in Figure 5, we designed Prefix Learnable Tokens, Surround Learnable Tokens, and Postfix Learnable Tokens to explore the impact of token positioning on model performance. Table 5 presents the ablation results on the impact of different learnable token positions. We observe that the rSum index of Surround Learnable Tokens is \\(2.6\\%\\) and \\(2.4\\%\\) higher than that of Prefix Learnable Tokens and Postfix Learnable Tokens, respectively. This outcome is intuitive, as the exact location of sparse text semantics within dense text is unknown. The Surround insertion method allows the model to predict semantics from both the preceding and following contexts simultaneously, leading to improved performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A deep decoder can improve retrieval performance. As shown in Figure 4b, as the decoder depth increases, retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.538 + ], + "angle": 0, + "content": "performance first improves and then declines, with the best results achieved when the number of layers is set to 4. Although our approach is inspired by Masked Signal Modeling [13], our findings regarding the decoder differ. In Masked Signal Modeling, the objective is to enhance the encoder through reconstruction, where an excessively deep decoder may hinder the encoder's capability. However, in our D2S-VSE, retrieval performance relies on the predictive ability of the decoder, meaning that a deeper decoder generally leads to better results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.63 + ], + "angle": 0, + "content": "Our D2S-VSE can work with very few decoder heads. As shown in Figure 4c, the number of attention heads has minimal impact on performance, with only 4 heads required to achieve optimal results. Increasing the number of heads does not lead to significant improvements, as all performance differences remain within \\(0.2\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.64, + 0.663 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we highlight the importance of embedding information capacity in addressing the multi-view description problem in image-text matching tasks. To tackle this challenge, we propose a novel Visual Semantic Embedding (VSE) framework, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE). The core idea of D2S-VSE is to enhance the information capacity of visual semantic embeddings by aligning dense text with images during pre-training. In the fine-tuning stage, we further refine sparse text embeddings through Dense-to-Sparse feature distillation, effectively improving their information capacity and leading to better image-text matching performance. Comprehensive experiments on two widely-used benchmarks validate the effectiveness of the proposed method, leading to state-of-the-art performance." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.232 + ], + "angle": 0, + "content": "Acknowledgements. This work was partially supported by the National Major Scientific Instruments and Equipments Development Project of National Natural Science Foundation of China under Grant 62427820, the National Science Foundation of China under Grant 62376175 and 22494712, the 111 Project under Grant B21044, the Science Fund for Creative Research Groups of Sichuan Province Natural Science Foundation under Grant 2024NSFTD0035, and the National Science Foundation of Sichuan Province under Grant 2025ZNSFSC0480." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.258, + 0.188, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.283, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[2] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In CVPR, pages 12655-12663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[3] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In CVPR, pages 15789-15798, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[4] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[5] Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio De Rezende, Yannis Kalantidis, and Diane Larlus. Probabilistic embeddings for cross-modal retrieval. In CVPR, pages 8415-8424, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In AAAI, pages 1218-1226, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. In BMVC, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[10] Zheren Fu, Zhendong Mao, Yan Song, and Yongdong Zhang. Learning semantic relationship among instances for image-text matching. In CVPR, pages 15159-15168, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[11] Zheren Fu, Lei Zhang, Hou Xia, and Zhendong Mao. Linguistic-aware patch slimming framework for fine-grained cross-modal alignment. In CVPR, pages 26307-26316, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Jieuxiang Gu, Jianfei Cai, Shafiq R Joty, Li Niu, and Gang Wang. Look, imagine and match: Improving textual-visual" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.283, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "cross-modal retrieval with generative models. In CVPR, pages 7181-7189, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[14] Yan Huang, Qi Wu, Chunfeng Song, and Liang Wang. Learning semantic concepts and order for image and sentence matching. In CVPR, pages 6163-6171, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. In CVPR, pages 23422-23431, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[16] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[17] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In ECCV, pages 201-216, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[18] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In ICCV, pages 4654-4662, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[19] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Image-text embedding learning via visual and textual semantic reasoning. TPAMI, 45(1):641-656, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[20] Zheng Li, Caili Guo, Zerun Feng, Jenq-Neng Hwang, and Xijun Xue. Multi-view visual semantic embedding. In IJ-CAI, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[22] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In CVPR, pages 10921-10930, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[24] Yang Liu, Hong Liu, Huaqiu Wang, and Mengyuan Liu. Regularizing visual semantic embedding with contrastive learning for image-text matching. IEEE SPL, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[25] Yang Liu, Hong Liu, Huaqiu Wang, Fanyang Meng, and Mengyuan Liu. Bcan: Bidirectional correct attention network for cross-modal retrieval. TNNLS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[26] Yang Liu, Mengyuan Liu, shudong Huang, and Jiancheng Lv. Asymmetric visual semantic embedding framework for efficient vision-language alignment. In AAAI, 2025. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Zhengxin Pan, Fangyu Wu, and Bailing Zhang. Fine-grained image-text matching by cross-modal hard aligning network. In CVPR, pages 19275-19284, 2023. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21687" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In ACMMM, pages 4948-4956, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[32] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. arXiv preprint arXiv:2310.17468, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[33] Leigang Qu, Meng Liu, Da Cao, Liqiang Nie, and Qi Tian. Context-aware multi-view summarization network for image-text matching. In ACMMM, pages 1047-1055, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[34] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In SIGIR, pages 1104-1113, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[36] Yale Song and Mohammad Soleymani. Polysemous visual-semantic embedding for cross-modal retrieval. In CVPR, pages 1979–1988, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[37] Haoran Wang, Ying Zhang, Zhong Ji, Yanwei Pang, and Lin Ma. Consensus-aware visual-semantic embedding for image-text matching. In ECCV, pages 18–34, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[38] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In ICCV, pages 5764-5773, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[39] Jiwei Wei, Xing Xu, Yang Yang, Yanli Ji, Zheng Wang, and Heng Tao Shen. Universal weighting metric learning for cross-modal matching. In CVPR, pages 13005-13014, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[40] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pretraining for long text understanding. Neurips, 37:64996-65019, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[41] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, pages 310–325. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[42] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In CVPR, pages 15661-15670, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In CVPR, pages 3536-3545, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[44] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamlip: Language-image pre-training with long captions. In ECCV, pages 73-90. Springer, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21688" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_origin.pdf b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7227a372c49d077ffbc8103ed15c13012b902ec4 --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/eee67cc2-5832-464d-b273-17cfe308efbf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a9b2a2c6f864701938bd4758635de874e8b4fc140f11f068bbfcb9f7ed3ebc7 +size 499255 diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/full.md b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9945dbe7450c1884c72197aad795712e1b6bf19c --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/full.md @@ -0,0 +1,289 @@ +# Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching + +Yang Liu $^{1,2}$ , Wentao Feng $^{1,2}$ , Zhuoyao Liu $^{1,2}$ , Shudong Huang $^{1,2,*}$ , Jiancheng Lv $^{1,2}$ + +1College of Computer Science, Sichuan University, China + +$^{2}$ Engineering Research Center of Machine Learning and Industry Intelligence, China + +https://d2s-vse.github.io + +# Abstract + +Enabling Visual Semantic Models to effectively handle multi-view description matching has been a longstanding challenge. Existing methods typically learn a set of embeddings to find the optimal match for each view's text and compute similarity. However, the visual and text embeddings learned through these approaches have limited information capacity and are prone to interference from locally similar negative samples. To address this issue, we argue that the information capacity of embeddings is crucial and propose Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE), which enhances the information capacity of sparse text by leveraging dense text distillation. Specifically, D2S-VSE is a two-stage framework. In the pretraining stage, we align images with dense text to enhance the information capacity of visual semantic embeddings. In the fine-tuning stage, we optimize two tasks simultaneously, distilling dense text embeddings to sparse text embeddings while aligning images and sparse texts, enhancing the information capacity of sparse text embeddings. Our proposed D2S-VSE model is extensively evaluated on the large-scale MS-COCO and Flickr30K datasets, demonstrating its superiority over recent state-of-the-art methods. + +# 1. Introduction + +In the domain of vision and language, the alignment of vision and language stands as a foundational task, with the core objective of establishing a robust and precise bridge between these two distinct yet interconnected modalities. The traditional approach to cross-modal retrieval [12, 24, 31, 32] follows a standard procedure: encoding images and texts into a shared embedding space and measuring their similarity. Typically, models are optimized using a triplet loss, which ensures that matched image-text pairs exhibit higher similarity than unmatched pairs [9, 16]. + +![](images/c081069e2598ab9f860806e5705792e1a064f6b608bca0533f41982fc40f83b1.jpg) +Figure 1. (Top) In previous methods, when addressing problems with varying information density, a set of visual embeddings is learned to match the most similar text embeddings. However, this approach results in learned embeddings with limited information capacity. (Mid) Since dense text has a higher information density than sparse text, training with aligned dense text results in image embeddings with greater information capacity. (Bottom) Sparse text embedding can be distilled through dense text embedding to enhance its information capacity, enabling the matching of text descriptions from different perspectives with a single visual embedding. (Best viewed in color) + +However, most traditional Visual Semantic Embedding (VSE) models overlook the inherent differences between image and text modalities [3, 18, 35]. Specifically, visual signals represent natural, objective records of real-world phenomena, whereas textual signals are inherently seman- + +tic and subjective, shaped by human interpretation. Consequently, the same image may be described in various ways depending on individual perspectives. Existing studies refer to this issue as the "ambiguity issue" [4, 36], "sparse annotations" [15] and "multi-view descriptions" [20, 33]. But, at its core, the problem stems from the inherent differences between image and text modalities, particularly the disparity in "information density" [26]. + +Recent studies [4, 20, 26, 36] propose learning a sets of embeddings from images/texts, where each embedding captures specific semantics. For each image-text pair, the most relevant embedding is selected for similarity calculation. While this approach reduces computational overhead and partially addresses information density challenges, we identify a critical limitation: As shown in the figure 1, the text in the dataset is usually short and concise. The image and text embeddings learned in this way can only capture a limited subset of features, resulting in limited information capacity being learned, and are therefore easily disturbed by negative samples with similar local semantics. Given this, we argue that the information capacity of visual semantic embeddings is crucial for the image-text matching task. The key question, then, is: how can we learn visual semantic embeddings with higher information capacity? + +In this paper, we introduce a novel approach, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE), which enhances visual semantic embeddings by aligning both semantics and information capacity. Our key hypothesis posits that text embeddings derived from dense image captions inherently possess greater information capacity compared than those from sparse textual annotations in traditional datasets such as Flickr30K and COCO. Specifically, we introduce a two-stage training framework. In the pre-train stage, we generate dense descriptions for images using LLaVa and pre-train image-dense description pairs to align images with dense textual descriptions in the embedding space. This alignment enables the model to extract more comprehensive features, resulting in embeddings with higher information capacity. In the fine-tuning stage, we introduce a dense-to-sparse feature distillation method. During this phase, while maintaining alignment between images and sparse text annotations from the dataset, we additionally distill knowledge from dense text embeddings into sparse text embeddings to enhance their information capacity. We formulate this distillation process as a masked signal reconstruction task for sparse text embeddings. To achieve this, we implement a Transformer-based decoder module that follows the text encoder. This decoder processes short text features combined with learnable mask tokens, aiming to predict the latent context of the sparse text and thus enhance its information capacity. And in the inference phase, we only require the image and the sparse text from the dataset, without the need for dense text, thus in + +curring no additional computational overhead. + +Our main contributions are summarized as follows: + +- Importance of Information Capacity: We identify the critical role of enhancing the information capacity of visual semantic embeddings. Based on this insight, we propose a novel framework to improve the VSE model by increasing the information capacity of visual semantic embeddings. +- Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE): We introduce D2S-VSE, a two-stage training approach that leverages dense text to enhance the information capacity of image embeddings in the first pre-training stage. In the second stage, a distillation branch is incorporated to improve the information capacity of sparse text embedding by distilling dense text embedding from sparse text embedding, ensuring both semantic alignment and balanced information capacity between image and text embeddings. +- State-of-the-Art Performance: We evaluate D2S-VSE with existing fine-grained methods across various model backbones. D2S-VSE outperforms the latest state-of-the-art methods on image-text retrieval benchmarks, including Flickr30K and MS-COCO. + +# 2. Related Work + +# 2.1. Image-Text Matching + +Existing Image-text retrieval approaches can be broadly categorized into two groups based on the granularity of the matching similarity: Global-level matching and Local-level matching methods. Global-level matching methods project images and texts into a shared embedding space, computing similarity using improved loss functions [5, 9] to bring semantically aligned pairs closer together. To improve the quality of the embedding space, recent related works proposed several more sophisticated networks, such as graph convolutional network [22, 37], Generalized Pooling Operator [3], and other model architectures [14, 18, 19]. In order to facilitate the closer alignment of semantically corresponding image-text pairs. Different from global matching methods, local matching methods[2, 7, 25, 34, 38, 39, 42, 43] focus more on the alignment between regions. Most of them capture the finer-grained relationship between images and texts by adopting cross-modal networks. + +Information Density Issues. Recent studies have proposed various methods to address the issue of differing information density between images and text. PCME [5] proposed a probabilistic cross-modal embedding to properly represent the one-to-many relationships in joint embedding. DivE [15] introduced a slot attention-based set prediction module, where the elements of the embedding set compete it- + +![](images/3528f257f5d9b0d3a36000fdf470d305c13e2f8bf70af8ff136da6d2153e2510.jpg) +Figure 2. Overview of our D2S-VSE framework. (a) Dense Text Generation: D2S-VSE utilizes the LLaVa [23] to generate dense text descriptions for images in the dataset that contain all the details of the image. (b) Pre-train With Dense Descriptions: D2S-VSE first uses contrastive learning to pre-train the model on image-dense-text pairs to enhance the information capacity of visual semantic embeddings. (c) Finetune With the Sparse Descriptions: D2S-VSE performs distillation from dense to sparse while contrasting the alignment, allowing sparse text embeddings to learn more details of dense text embeddings, thereby improving the information capacity of sparse text embeddings and better aligning them with image embeddings. + +eratively to aggregate the input data. Under the framework of PCME, PCME++ [4] introduced a new probabilistic distance with a closed-form solution. AVSE [26] proposed a matching module for compute the dynamic similarity between asymmetric multi-view image and text embeddings. + +These methods learn a set of embeddedings to match ambiguous samples, but reduce the information capacity of each embedding. + +# 2.2. Vision language model with dense text. + +In the field of vision-language research, few studies have leveraged long-text descriptions to enhance the multimodal representation of images. DreamLip [44] utilizes a multimodal large language model to generate dense text descriptions and randomly extracts sentence fragments for pre-training, but it does not fully exploit long-text pretraining. Long-CLIP [41] fine-tunes CLIP-pre-trained on a short-text-image dataset—on long-text-image pairs to unlock long-text capabilities, yet it overlooks short-text effectiveness. LoTLIP [40] directly pre-trains on long-text + +image data and introduces corner tokens to aggregate diverse textual information, preserving the ability to handle short text. + +Different from the above methods that focus on pretraining, our method utilizes a very small amount of long text for pre-training, aiming to use the long text to enhance the information capacity of visual semantic embeddings. + +# 3. Dense-to-Sparse Feature Distilled Visual Semantic Embedding + +Our goal is to enhance Visual Semantic Embedding (VSE) models to address the challenge of inconsistent information density between images and text. Existing VSE models typically align images with the sparse textual descriptions provided in datasets. This approach biases the image encoder toward textual perspectives, limiting its ability to fully comprehend visual content and making it difficult to handle ambiguous retrieval cases. To overcome this issue, we propose a two-stage training procedure. An overview of our D2S-VSE framework is illustrated in Figure 2. + +- Stage I. We leverage the pre-trained multimodal large model LLaVa to generate dense textual descriptions for images. The VSE model is then pre-trained using these image-dense text pairs. By training on data with comparable information density, the model learns embeddings with greater information capacity, enabling a more comprehensive understanding of visual content (Sec. 3.1). +- **Stage II.** We fine-tune the pre-trained VSE model on general datasets containing image-sparse text pairs. Simultaneously, we distill knowledge from the dense text features into the sparse text representations, enhancing their information capacity. While aligning image-text embeddings, we also align the information capacity between visual and textual modalities. This enables the VSE model to effectively handle varying information densities in images and texts (Sec. 3.2). + +# 3.1. Pre-training with Dense Descriptions + +To address the issue of inconsistent information density between images and text, we introduce a pre-training stage where the Visual Semantic Embedding (VSE) model learns from image-dense text pairs. This stage enhances the model's ability to extract rich semantic representations from images by ensuring that textual descriptions contain a level of detail comparable to the visual information. + +Dense Text Generation. To generate dense textual descriptions for images, we leverage the pre-trained multimodal large model LLaVa [23], which integrates vision and language understanding to generate detailed descriptions based on image inputs. Unlike traditional dataset texts, which are often brief and sparse, LLaVa produces detailed descriptions that include object attributes, relationships, actions, and contextual information. Given a raw image $I$ , we feed it into the LLaVa with the prompt "Describe this image in detail". LLaVa then produces a dense textual description that captures fine-grained details of the image. The hyperparameters used for LLaVa are detailed in Sec 4.2. This process ensures that the textual input provides a comprehensive representation of the image, reducing the information density gap between modalities. + +Image-Dense-Caption Pre-train. Once the dense captions are generated, we pre-train the VSE model using these image-dense text pairs. The objective of this stage is to enable the model to learn embeddings that effectively capture fine-grained visual semantics and their corresponding textual descriptions. + +We employ a contrastive learning framework where the image encoder $\mathcal{F}_i$ and the dense text encoder $\mathcal{F}_t^d$ map images and dense textual descriptions into a shared embedding space. Formally, given a dense textual description $T^d$ and a raw image as $I$ , our goal is to map these inputs into compact vectors $\mathbf{t}^d \in \mathbb{R}^d$ and $\mathbf{i} \in \mathbb{R}^d$ such that their semantic + +![](images/0f64ed5e7bfcfd77a3f28f6584a0d702661448cb09ea1092c7f7d4ad395b1efb.jpg) +Figure 3. Architecture for the sparse text branch. Similar to the decoder in MAE [13], we introduce a set of learnable tokens in the sparse text branch to predict the implicit context of sparse text through the decoder. The learnable tokens help in reconstructing the missing semantics, allowing the model to better capture the underlying context and relationships within the sparse text. + +relationships can be quantified through a similarity metric $S(\cdot, \cdot)$ . Specifically, the similarity function is formulated as: + +$$ +S (I, T) = \frac {\mathbf {i} ^ {T} \mathbf {t} ^ {d}}{| | \mathbf {i} | | \cdot | | \mathbf {t} ^ {d} | |} \tag {1} +$$ + +To optimize the model, we use triplet loss function with the hardest negative samples to minimize the similarity between relevant pairs and to maximize the distance between irrelevant ones. The optimized loss function is mathematically formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {p r e t r a i n}} = \sum_ {(I, T ^ {d}) \in \mathcal {D}} \left[ \alpha - S (I, T ^ {d}) + S (I, \hat {T} ^ {d}) \right] _ {+} \tag {2} \\ + \left[ \alpha - S (I, T ^ {d}) + S (\hat {I}, T ^ {d}) \right] _ {+} \\ \end{array} +$$ + +where $\alpha$ serves as a margin parameter and $[x]_{+} \equiv Max(x,0)$ . In dataset $\mathcal{D}$ , the similarity in a positive pair $S(I,T^{d})$ should be higher than that in the hardest negative pairs $S(\hat{I},T^{d})$ and $S(I,\hat{T}^{d})$ by a margin $\alpha$ . + +# 3.2. Finetune With Sparse Description + +Stage I pre-training with dense descriptions enhances the information capacity of embeddings of the VSE model, but there exists a gap between the training and testing data, as + +
MethodFGFlickr30k 1K TestMS-COCO 5-fold 1K TestMS-COCO 5K Test
Text RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSum
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
ViT-Base-224 + BERT-base
VSE++ [9]X71.892.896.559.484.790.9496.175.094.698.062.789.494.9514.662.387.693.443.973.683.3413.4
SCAN [17]✓69.590.995.656.483.190.0485.676.095.498.164.590.895.8520.653.981.890.042.972.382.5423.5
SGR [7]✓69.790.895.259.184.189.9488.777.295.098.065.190.795.8521.854.982.890.542.872.282.5425.8
CHAN [29]✓69.291.895.058.484.990.6489.977.195.198.165.091.096.0522.256.383.290.143.072.682.8428.0
LAPS [11]✓74.093.497.462.587.392.7507.378.795.598.366.291.396.2526.357.584.090.844.574.083.6434.4
AVSE [26]X76.094.697.562.788.493.1512.379.895.698.367.091.596.3528.558.884.391.045.174.383.9437.4
D2S-VSEX82.896.198.368.591.394.9531.980.197.099.268.192.596.7533.760.185.592.546.375.985.2445.6
ViT-Base-384 + BERT-base
VSE++ [9]X77.195.797.565.890.294.3520.577.095.798.464.691.196.2523.054.982.890.442.472.482.8425.8
SCAN [17]✓75.494.496.963.688.693.5512.576.195.598.565.191.696.3523.153.381.890.042.672.682.9423.1
SGR [7]✓76.994.998.164.288.493.3515.875.895.798.665.692.096.5524.253.381.089.642.973.183.7423.6
CHAN [29]✓75.494.597.663.288.693.1512.478.195.898.666.182.196.6527.355.683.891.243.473.683.5431.1
LAPS [11]✓79.096.098.167.390.594.5525.478.696.398.968.092.496.8531.057.484.992.546.475.885.2442.2
AVSE [26]X80.396.498.767.991.294.7529.281.197.199.068.392.797.0535.261.286.893.246.275.985.0448.3
D2S-VSEX84.197.599.170.391.695.3537.980.897.299.169.092.996.8535.860.686.593.246.876.485.7449.1
+ +Table 1. Comparison of experimental results using ViT-Base [8] and BERT [6] the MS-COCO and Flickr30K datasets. "FG" indicates whether fine-grained cross-modal alignment is applied. The best performance is highlighted in bold. + +the model is trained on dense texts and tested on sparse texts, leading to suboptimal performance. However, simply fine-tuning directly on the image-sparse text pair data will not make the sparse text embedding contain richer semantics, that is, more information capacity. To address this problem, we introduce a novel distillation mechanism while aligning, which treats sparse text as masked dense text and designs a decoder. Through distillation, the decoder can decode sparse text embeddings into dense text embeddings, thereby improving the information capacity of sparse text embeddings. This mechanism extracts sparse text embeddings from dense text embeddings while aligning images and text, thereby enhancing the information capacity of sparse text features. + +# 3.2.1. Aligning Information Capacity by Dense-to-Sparse Feature Distillation + +To ensure that the sparse text embedding retains as much information as the dense text embedding, we introduce a feature distillation mechanism from dense to sparse. Inspired by masked signal modeling [6, 13], we treat sparse text as a masked version of dense text. This allows us to decode and reconstruct the sparse text embedding through feature distillation, enabling the reconstructed embedding to approximate the information capacity of the dense text embedding. + +Transformer-based Decoder. To achieve this reconstruction, we design a Transformer-based decoder, as illustrated + +in Figure 3. The decoder plays a pivotal role in D2S-VSE, responsible for reconstructing the dense text embedding from the sparse text embedding by leveraging feature distillation. It learns to restore the missing contextual information in sparse text, ensuring that the final sparse text embedding retains as much semantic richness as the dense text embedding. Given a sparse text $T^s$ , we first feed it into the sparse text encoder $\mathcal{F}_t^s$ to obtain the word embedding $\mathbf{w}^s$ and the sparse text embedding $\mathbf{t}^s$ . Next, learnable mask tokens with position $\mathbf{m}$ are concatenated with $\mathbf{w}^s$ and input into the decoder for processing. The decoder outputs the reconstructed tokens $\bar{\mathbf{m}}$ , which are then averaged to obtain $\bar{\mathbf{t}}$ . Finally, we sum $\mathbf{t}^s$ and $\bar{\mathbf{t}}$ to generate the final sparse text embedding $\hat{\mathbf{t}}^s$ . This process can be formulated as: + +$$ +\hat {\mathbf {t}} ^ {s} = \operatorname {D e c o d e r} \left(\left[ \mathbf {w} ^ {s}, \mathbf {m} \right]\right) + \mathbf {t} ^ {s} \tag {3} +$$ + +This reconstruction process enables the mask tokens to capture the contextual information missing from the sparse text, aligning it more closely with the dense text embedding. By combining the two, we obtain a more comprehensive and enriched text feature embedding. + +Dense-to-Sparse Distillation. Through the dense text pretraining in stage 1, the representation of the pre-trained VSE model can capture richer semantics, and the pre-trained dense text encoder can be used as a teacher to distill the information capacity of dense text embeddings into sparse text embeddings. This Dense-to-Sparse distillation mini- + +mizes a negative cosine similarity between then dense text embedding $\mathbf{t}^d$ and the final sparse text embedding $\hat{\mathbf{t}}^s$ : + +$$ +\mathcal {L} _ {\text {d i s t i l l}} \left(\mathbf {t} ^ {d}, \hat {\mathbf {t}} ^ {s}\right) = 1 - \frac {\mathbf {t} ^ {d} \cdot \hat {\mathbf {t}} ^ {s}}{\left| \left| \mathbf {t} ^ {d} \right| \right| \cdot \left| \left| \hat {\mathbf {t}} ^ {s} \right| \right|} \tag {4} +$$ + +# 3.2.2. Aligning Visual Semantic Using Sparse text + +At the same time, we also fine-tune the VSE model on image-sparse text pair data to adapt the model to common sparse text queries in cross-modal retrieval. Given image $I$ and sparse text $T^s$ , we extract image embedding $\mathbf{i}$ and sparse text embedding $\mathbf{t}$ and optimize the VSE model as follow: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a l i g n}} = \sum_ {(I, T ^ {s}) \in \mathcal {D}} \left[ \alpha - S (I, T ^ {s}) + S (I, \hat {T} ^ {s}) \right] _ {+} \tag {5} \\ + \left[ \alpha - S (I, T ^ {s}) + S (\hat {I}, T ^ {s}) \right] _ {+} \\ \end{array} +$$ + +In summary, the final loss function is defined as follows to perform joint optimization of the two objectives. + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a l i g n}} + \mathcal {L} _ {\text {d i s t i l l}} \tag {6} +$$ + +# 4. Experiments + +# 4.1. Dataset and Settings + +Datasets. Following previous works [9], we conduct our experiments on two widely used benchmark datasets: MS-COCO [21] and Flickr30K [30]. MS-COCO consists of 123,287 images, with each image annotated with five text captions. Following the dataset split in [9], we use 113,287 images for training, 5,000 for validation, and 5,000 for testing. Flickr30K contains 31,783 images, each paired with five textual descriptions. We adopt the standard split from [9], using 29,000 images for training, 1,000 for validation, and 1,000 for testing. + +Evaluation Metrics. For image-text matching, we use the commonly adopted Recall@K (where $\mathrm{K} = 1,5,10$ ), denoted as R@1, R@5, and R@10. These metrics represent the percentage of ground-truth instances retrieved within the top 1, 5, or 10 results, respectively. + +# 4.2. Implementation details + +For dense text generation, we use LLaVa [23] to generate dense textual descriptions. The hyperparameters for LLaVa are set as follows: Top-P = 0.9, temperature = 0.2, and max-new-tokens = 500. We generate a dense text description for each image in the MS-COCO and Flickr30K datasets, with each dense text sentence corresponding to $\frac{1}{5}$ of the sparse text descriptions. + +In the pre-training stage, we use the AdamW optimizer [28] to train the VSE model for 30 epochs, with a minibatch size of 128. The learning rate is initially set to 0.0005 + +
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Pre-trained ResNet-152 on IN + GRU
VSE∞ [3]76.594.297.756.483.489.9498.1
MV-VSE [20]79.094.997.759.184.690.6505.9
NAAF* [42]81.996.198.361.085.390.6513.2
HREM* [10]81.496.598.560.985.691.3514.2
CHAN [29]79.794.597.360.285.390.7507.7
AVSE [26]82.396.298.562.888.092.8520.6
D2S-VSE82.596.699.062.688.193.1521.9
Swin-Base-224 + BERT-base
VSE++ [9]82.596.598.970.091.495.1534.4
SCAN [17]79.095.998.267.790.694.9526.3
SGR [7]80.497.098.766.990.294.5527.6
CHAN [29])81.497.098.668.590.694.5530.6
LAPS[11]82.497.499.570.091.795.4536.3
AVSE [26]83.997.499.470.092.495.6538.7
D2S-VSE87.298.499.973.093.596.7548.7
Swin-Base-384 + BERT-base
VSE++ [9]83.397.599.271.193.296.2540.6
SCAN [17]81.996.998.970.092.795.8536.1
SGR [7]80.796.899.069.991.795.3533.4
CHAN [29]81.296.798.870.392.295.9535.0
LAPS [11]85.197.799.274.093.096.3545.3
AVSE [26]87.198.399.273.693.596.5548.2
D2S-VSE87.899.099.775.794.196.9553.2
+ +Table 2. Comparison of experimental results using other common backbones on the Flickr30K dataset. The best performance is highlighted in **.** indicates ensemble methods. + +for the first few epochs and is reduced by a factor of 10 during the 9th, 15th, 20th, and 25th epochs. In the fine-tuning stage, we freeze the dense text encoder and fine-tune both the image encoder and sparse text encoder. We train them using the same training strategy as in the pre-training stage and evaluate the best model on the validation set. + +For model details, we use traditional region features and employ the recently popular Vision Transformer (ViT) [8] and Swin Transformer [27] as the backbone. We set the dimension of the shared embedding space to 512. For feature aggregation, we use GPO [3] for both the image encoder and the text encoder. The decoder consists of 4 layers of Transformer, with 4 attention heads per layer. We set 100 learnable tokens to predict the context of sparse text embeddings. We provide a more detailed implementation setup in the supplementary material. + +# 4.3. Comparisons with Recent State-of-the-arts + +We compare our D2S-VSE with the most recent state-of-the-art approaches on two widely used datasets, i.e., Flirkr30K and MS-COCO. For the sake of fairness, our D2S-VSE employs a diverse range of backbones to facili + +![](images/b60994adbd0779743cbcb8b7edb52e9193de6f5c14ce216b1ed81dc0ecf39168.jpg) +(a) The impact of Learnable Tokens + +![](images/73ffa7a323a7e7ae6115f36ee47f3b0d9eba35a73f870d602ccbbb05bb24a59c.jpg) +(b) The impact of Decoder Depth +Figure 4. Ablation study of the architecture of decoder on Flickr30K. If not specified, the default is: the number of the learnable tokens is 100, the decoder has 4 layers and 4 head. + +![](images/288db231dbc846eae6a3093b2cf3ce429dde15b05ae06a89d005f626f78f8544.jpg) +(c) The impact of Decoder Head + +tate comparisons with various previous methods. + +In Table 1, we present the results of using ViT-base [8] + BERT-base [6] as the backbone on the Flickr30K and MS-COCO datasets. Our D2S-VSE surpasses all state-of-the-art methods, demonstrating remarkable performance. In this setting, we compare our method against the two leading approaches, LAPS [11] and AVSE [26], and achieve superior results across nearly all metrics. For instance, on the Flickr30K dataset, our approach outperforms AVSE in text retrieval, achieving a $6.8\%$ improvement in $\mathsf{R}@\mathsf{1}$ with the ViT-base-224 backbone and a $3.8\%$ improvement with the ViT-base-384 backbone. Similarly, for image retrieval, our method achieves gains of $5.8\%$ and $2.4\%$ when using the ViT-base-224 and ViT-base-384 backbones, respectively. A consistent performance boost is also observed across all metrics on the MS-COCO dataset. + +Additionally, we conducted experiments using other commonly used backbones, as shown in Table 2. Specifically, we evaluated our approach on the Flickr30K dataset with Faster R-CNN [1] + GRU and Swin-Base [27] + BERT-Base as backbones. The results further confirm that our method consistently outperforms existing approaches across different backbone architectures. + +# 4.4. Ablation Study + +To verify the effectiveness of each component of our D2S-VSE, we conduct extensive ablation studies on the Flickr30K dataset. All experiments are performed using ViT-base-224 and BERT-base as the backbone to ensure a consistent and fair comparison. + +The impact of the two-stage training framework. Our D2S-VSE introduces a two-stage training framework to enhance the information capacity of image embedding and text embedding, and achieves very good results on commonly used datasets. In order to explore the role of each component in the framework, we conducted ablation experiments as shown in Table 3. + +Is the model pre-trained with dense text inherently superior, leading to better performance? To verify this, we directly tested the pre-trained model on the Flickr30K test + +
PretrainFinetuneText RetrievalImage RetrievalrSum
AlignDistillR@1R@5R@10R@1R@5R@10
✓56.284.791.641.771.379.7425.2
✓76.393.997.361.287.693.2509.5
✓✓79.296.598.366.890.594.4525.7
✓✓76.094.597.362.988.393.1512.1
✓✓✓82.896.198.368.591.394.9531.9
+ +Table 3. Ablation experiment on the impact of our framework on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue. + +set. We found that due to the gap between dense and sparse text, the pre-trained model struggles to adapt to sparse text, resulting in rSum of only 425.2 on the Flickr30K test set. + +The pre-training stage enhances the information capacity of image embeddings! We observed that the R@1 of text retrieval improved by $2.9\%$ after pre-training compared to the model without pre-training. This suggests that dense text pre-training enhances the information capacity of image embeddings, enabling the model to better align multi-view descriptions. + +Dense-to-Sparse feature distillation enhances the information capacity of sparse text embeddings! Applying dense-to-sparse distillation while aligning images during the fine-tuning stage further boosts the model's retrieval performance. On the Flickr30K dataset, text retrieval and image retrieval reached $82.8\%$ and $68.5\%$ , respectively. This improvement stems from distillation enhances the information capacity of text embeddings, leading to better semantic consistency between image and text embeddings. + +Different type of distillation functions. We used L1 distance, L2 distance, and negative cosine similarity as the loss functions for Dense-to-Sparse distillation, respectively. As shown in Table 4, the effect of negative cosine similarity distillation is much better than the other two. + +Decoder design. The decoder is the most important part of Dense-to-Sparse distillation. Its existence allows learnable tokens to predict possible contexts based on the word embeddings of sparse text, thereby increasing the informa + +![](images/dc488757593a235bb1dd6d7232488b96f4d365e05cb246425706934912cc9ca9.jpg) +(a) Prefix Learnable Tokens + +![](images/24a8c85839378f2a7f0d6fe44c3020aa2eaaa0d2e5ca316f063dadfed9f34e1a.jpg) +(a) Surround Learnable Tokens +Figure 5. Different type of the position strategy of learnable tokens. Tokens at different positions can predict the semantics of different positions in dense text. + +![](images/48b9a4dcc44b8a830bf6286b971b80357f3b8d652d5e579ff6e5636c68344f86.jpg) +(c) Postfix Learnable Tokens + +
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
\( \ell_1 \) Distance81.995.398.268.090.894.7528.9
\( \ell_2 \) Distance82.195.698.467.991.295.1530.2
Negative Cosine Similarity82.896.198.368.591.394.9531.9
+ +Table 4. Ablation experiment on the impact of different distillation functions on Flick30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue. + +
TypeText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Prefix81.795.398.268.191.294.8529.3
Postfix82.095.698.367.891.194.7529.5
Surround82.896.198.368.591.394.9531.9
+ +Table 5. Ablation experiment on the impact of the position of the learnable tokens on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue. + +tion capacity of sparse text features. As shown in Figure 4, we conduct ablation studies of Learnable tokens, decoder depth, and decoder head. + +The number of tokens determines the reconstruction capability of the decoder. As shown in Figure 4a, we observed that as the number of learnable tokens in the decoder increases, the model's performance in both image retrieval and text retrieval gradually improves, stabilizing when the number of tokens reaches approximately 100. At this point, the R@1 for text retrieval and image retrieval reaches $82.8\%$ and $68.5\%$ , respectively. However, when the number of tokens increases to 120, the model shows no significant advantage. Therefore, considering both performance and efficiency, we set the number of tokens to 100. + +Surrounding learnable tokens can better predict the context of sparse text. In addition, we conducted an ablation experiment on the insertion position of learnable tokens. Since our motivation stems from treating sparse text as masked dense text, with learnable tokens used to reconstruct the masked semantics, their placement is crucial. As shown in Figure 5, we designed Prefix Learnable Tokens, Surround Learnable Tokens, and Postfix Learnable Tokens to explore the impact of token positioning on model performance. Table 5 presents the ablation results on the impact of different learnable token positions. We observe that the rSum index of Surround Learnable Tokens is $2.6\%$ and $2.4\%$ higher than that of Prefix Learnable Tokens and Postfix Learnable Tokens, respectively. This outcome is intuitive, as the exact location of sparse text semantics within dense text is unknown. The Surround insertion method allows the model to predict semantics from both the preceding and following contexts simultaneously, leading to improved performance. + +A deep decoder can improve retrieval performance. As shown in Figure 4b, as the decoder depth increases, retrieval + +performance first improves and then declines, with the best results achieved when the number of layers is set to 4. Although our approach is inspired by Masked Signal Modeling [13], our findings regarding the decoder differ. In Masked Signal Modeling, the objective is to enhance the encoder through reconstruction, where an excessively deep decoder may hinder the encoder's capability. However, in our D2S-VSE, retrieval performance relies on the predictive ability of the decoder, meaning that a deeper decoder generally leads to better results. + +Our D2S-VSE can work with very few decoder heads. As shown in Figure 4c, the number of attention heads has minimal impact on performance, with only 4 heads required to achieve optimal results. Increasing the number of heads does not lead to significant improvements, as all performance differences remain within $0.2\%$ . + +# 5. Conclusions + +In this paper, we highlight the importance of embedding information capacity in addressing the multi-view description problem in image-text matching tasks. To tackle this challenge, we propose a novel Visual Semantic Embedding (VSE) framework, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE). The core idea of D2S-VSE is to enhance the information capacity of visual semantic embeddings by aligning dense text with images during pre-training. In the fine-tuning stage, we further refine sparse text embeddings through Dense-to-Sparse feature distillation, effectively improving their information capacity and leading to better image-text matching performance. Comprehensive experiments on two widely-used benchmarks validate the effectiveness of the proposed method, leading to state-of-the-art performance. + +Acknowledgements. This work was partially supported by the National Major Scientific Instruments and Equipments Development Project of National Natural Science Foundation of China under Grant 62427820, the National Science Foundation of China under Grant 62376175 and 22494712, the 111 Project under Grant B21044, the Science Fund for Creative Research Groups of Sichuan Province Natural Science Foundation under Grant 2024NSFTD0035, and the National Science Foundation of Sichuan Province under Grant 2025ZNSFSC0480. + +# References + +[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 7 +[2] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In CVPR, pages 12655-12663, 2020. 2 +[3] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In CVPR, pages 15789-15798, 2021. 1, 2, 6 +[4] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 2, 3 +[5] Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio De Rezende, Yannis Kalantidis, and Diane Larlus. Probabilistic embeddings for cross-modal retrieval. In CVPR, pages 8415-8424, 2021. 2 +[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 7 +[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In AAAI, pages 1218-1226, 2021. 2, 5, 6 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5,6,7 +[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. In BMVC, 2018. 1, 2, 5, 6 +[10] Zheren Fu, Zhendong Mao, Yan Song, and Yongdong Zhang. Learning semantic relationship among instances for image-text matching. In CVPR, pages 15159-15168, 2023. 6 +[11] Zheren Fu, Lei Zhang, Hou Xia, and Zhendong Mao. Linguistic-aware patch slimming framework for fine-grained cross-modal alignment. In CVPR, pages 26307-26316, 2024. 5, 6, 7 +[12] Jieuxiang Gu, Jianfei Cai, Shafiq R Joty, Li Niu, and Gang Wang. Look, imagine and match: Improving textual-visual + +cross-modal retrieval with generative models. In CVPR, pages 7181-7189, 2018. 1 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 4, 5, 8 +[14] Yan Huang, Qi Wu, Chunfeng Song, and Liang Wang. Learning semantic concepts and order for image and sentence matching. In CVPR, pages 6163-6171, 2018. 2 +[15] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. In CVPR, pages 23422-23431, 2023. 2 +[16] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 1 +[17] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In ECCV, pages 201-216, 2018. 5, 6 +[18] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In ICCV, pages 4654-4662, 2019. 1, 2 +[19] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Image-text embedding learning via visual and textual semantic reasoning. TPAMI, 45(1):641-656, 2022. 2 +[20] Zheng Li, Caili Guo, Zerun Feng, Jenq-Neng Hwang, and Xijun Xue. Multi-view visual semantic embedding. In IJ-CAI, 2022. 2, 6 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6 +[22] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In CVPR, pages 10921-10930, 2020. 2 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 3, 4, 6 +[24] Yang Liu, Hong Liu, Huaqiu Wang, and Mengyuan Liu. Regularizing visual semantic embedding with contrastive learning for image-text matching. IEEE SPL, 2022. 1 +[25] Yang Liu, Hong Liu, Huaqiu Wang, Fanyang Meng, and Mengyuan Liu. Bcan: Bidirectional correct attention network for cross-modal retrieval. TNNLS, 2023. 2 +[26] Yang Liu, Mengyuan Liu, shudong Huang, and Jiancheng Lv. Asymmetric visual semantic embedding framework for efficient vision-language alignment. In AAAI, 2025. 2, 3, 5, 6, 7 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6 +[29] Zhengxin Pan, Fangyu Wu, and Bailing Zhang. Fine-grained image-text matching by cross-modal hard aligning network. In CVPR, pages 19275-19284, 2023. 5, 6 + +[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 6 +[31] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In ACMMM, pages 4948-4956, 2022. 1 +[32] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. arXiv preprint arXiv:2310.17468, 2023. 1 +[33] Leigang Qu, Meng Liu, Da Cao, Liqiang Nie, and Qi Tian. Context-aware multi-view summarization network for image-text matching. In ACMMM, pages 1047-1055, 2020. 2 +[34] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In SIGIR, pages 1104-1113, 2021. 2 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 1 +[36] Yale Song and Mohammad Soleymani. Polysemous visual-semantic embedding for cross-modal retrieval. In CVPR, pages 1979–1988, 2019. 2 +[37] Haoran Wang, Ying Zhang, Zhong Ji, Yanwei Pang, and Lin Ma. Consensus-aware visual-semantic embedding for image-text matching. In ECCV, pages 18–34, 2020. 2 +[38] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In ICCV, pages 5764-5773, 2019. 2 +[39] Jiwei Wei, Xing Xu, Yang Yang, Yanli Ji, Zheng Wang, and Heng Tao Shen. Universal weighting metric learning for cross-modal matching. In CVPR, pages 13005-13014, 2020. 2 +[40] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pretraining for long text understanding. Neurips, 37:64996-65019, 2024. 3 +[41] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, pages 310–325. Springer, 2024. 3 +[42] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In CVPR, pages 15661-15670, 2022. 2, 6 +[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In CVPR, pages 3536-3545, 2020. 2 +[44] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamlip: Language-image pre-training with long captions. In ECCV, pages 73-90. Springer, 2024. 3 \ No newline at end of file diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/images.zip b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6dca4843455cb03839393a5de6167179f715c057 --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6ea39d4220efe05633cf20345d06dd431cfa551e6fe99bbadf5adcfc9e5b148 +size 751540 diff --git a/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/layout.json b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c319567635054697e9305d40193fe0c1386e08ff --- /dev/null +++ b/2025/Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching/layout.json @@ -0,0 +1,7345 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 104, + 103, + 507, + 140 + ], + "type": "text", + "content": "Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "spans": [ + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "content": "Yang Liu" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "content": ", Wentao Feng" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "content": ", Zhuoyao Liu" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "content": ", Shudong Huang" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "text", + "content": ", Jiancheng Lv" + }, + { + "bbox": [ + 95, + 160, + 499, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 149, + 177, + 446, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 177, + 446, + 189 + ], + "spans": [ + { + "bbox": [ + 149, + 177, + 446, + 189 + ], + "type": "text", + "content": "1College of Computer Science, Sichuan University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 190, + 515, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 190, + 515, + 203 + ], + "spans": [ + { + "bbox": [ + 78, + 190, + 515, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 190, + 515, + 203 + ], + "type": "text", + "content": "Engineering Research Center of Machine Learning and Industry Intelligence, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 228, + 205, + 366, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 205, + 366, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 205, + 366, + 217 + ], + "type": "text", + "content": "https://d2s-vse.github.io" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 270, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 534 + ], + "type": "text", + "content": "Enabling Visual Semantic Models to effectively handle multi-view description matching has been a longstanding challenge. Existing methods typically learn a set of embeddings to find the optimal match for each view's text and compute similarity. However, the visual and text embeddings learned through these approaches have limited information capacity and are prone to interference from locally similar negative samples. To address this issue, we argue that the information capacity of embeddings is crucial and propose Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE), which enhances the information capacity of sparse text by leveraging dense text distillation. Specifically, D2S-VSE is a two-stage framework. In the pretraining stage, we align images with dense text to enhance the information capacity of visual semantic embeddings. In the fine-tuning stage, we optimize two tasks simultaneously, distilling dense text embeddings to sparse text embeddings while aligning images and sparse texts, enhancing the information capacity of sparse text embeddings. Our proposed D2S-VSE model is extensively evaluated on the large-scale MS-COCO and Flickr30K datasets, demonstrating its superiority over recent state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 696 + ], + "type": "text", + "content": "In the domain of vision and language, the alignment of vision and language stands as a foundational task, with the core objective of establishing a robust and precise bridge between these two distinct yet interconnected modalities. The traditional approach to cross-modal retrieval [12, 24, 31, 32] follows a standard procedure: encoding images and texts into a shared embedding space and measuring their similarity. Typically, models are optimized using a triplet loss, which ensures that matched image-text pairs exhibit higher similarity than unmatched pairs [9, 16]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 243, + 555, + 498 + ], + "blocks": [ + { + "bbox": [ + 316, + 243, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 498 + ], + "type": "image", + "image_path": "c081069e2598ab9f860806e5705792e1a064f6b608bca0533f41982fc40f83b1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 506, + 555, + 627 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 627 + ], + "type": "text", + "content": "Figure 1. (Top) In previous methods, when addressing problems with varying information density, a set of visual embeddings is learned to match the most similar text embeddings. However, this approach results in learned embeddings with limited information capacity. (Mid) Since dense text has a higher information density than sparse text, training with aligned dense text results in image embeddings with greater information capacity. (Bottom) Sparse text embedding can be distilled through dense text embedding to enhance its information capacity, enabling the matching of text descriptions from different perspectives with a single visual embedding. (Best viewed in color)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "However, most traditional Visual Semantic Embedding (VSE) models overlook the inherent differences between image and text modalities [3, 18, 35]. Specifically, visual signals represent natural, objective records of real-world phenomena, whereas textual signals are inherently seman-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding Author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21679" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 168 + ], + "type": "text", + "content": "tic and subjective, shaped by human interpretation. Consequently, the same image may be described in various ways depending on individual perspectives. Existing studies refer to this issue as the \"ambiguity issue\" [4, 36], \"sparse annotations\" [15] and \"multi-view descriptions\" [20, 33]. But, at its core, the problem stems from the inherent differences between image and text modalities, particularly the disparity in \"information density\" [26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 364 + ], + "type": "text", + "content": "Recent studies [4, 20, 26, 36] propose learning a sets of embeddings from images/texts, where each embedding captures specific semantics. For each image-text pair, the most relevant embedding is selected for similarity calculation. While this approach reduces computational overhead and partially addresses information density challenges, we identify a critical limitation: As shown in the figure 1, the text in the dataset is usually short and concise. The image and text embeddings learned in this way can only capture a limited subset of features, resulting in limited information capacity being learned, and are therefore easily disturbed by negative samples with similar local semantics. Given this, we argue that the information capacity of visual semantic embeddings is crucial for the image-text matching task. The key question, then, is: how can we learn visual semantic embeddings with higher information capacity?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel approach, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE), which enhances visual semantic embeddings by aligning both semantics and information capacity. Our key hypothesis posits that text embeddings derived from dense image captions inherently possess greater information capacity compared than those from sparse textual annotations in traditional datasets such as Flickr30K and COCO. Specifically, we introduce a two-stage training framework. In the pre-train stage, we generate dense descriptions for images using LLaVa and pre-train image-dense description pairs to align images with dense textual descriptions in the embedding space. This alignment enables the model to extract more comprehensive features, resulting in embeddings with higher information capacity. In the fine-tuning stage, we introduce a dense-to-sparse feature distillation method. During this phase, while maintaining alignment between images and sparse text annotations from the dataset, we additionally distill knowledge from dense text embeddings into sparse text embeddings to enhance their information capacity. We formulate this distillation process as a masked signal reconstruction task for sparse text embeddings. To achieve this, we implement a Transformer-based decoder module that follows the text encoder. This decoder processes short text features combined with learnable mask tokens, aiming to predict the latent context of the sparse text and thus enhance its information capacity. And in the inference phase, we only require the image and the sparse text from the dataset, without the need for dense text, thus in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 501, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 501, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 501, + 83 + ], + "type": "text", + "content": "curring no additional computational overhead." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 85, + 533, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 85, + 533, + 95 + ], + "spans": [ + { + "bbox": [ + 325, + 85, + 533, + 95 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 98, + 553, + 348 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 98, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 553, + 169 + ], + "type": "text", + "content": "- Importance of Information Capacity: We identify the critical role of enhancing the information capacity of visual semantic embeddings. Based on this insight, we propose a novel framework to improve the VSE model by increasing the information capacity of visual semantic embeddings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 170, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 289 + ], + "type": "text", + "content": "- Dense-to-Sparse Feature Distilled Visual Semantic Embedding (D2S-VSE): We introduce D2S-VSE, a two-stage training approach that leverages dense text to enhance the information capacity of image embeddings in the first pre-training stage. In the second stage, a distillation branch is incorporated to improve the information capacity of sparse text embedding by distilling dense text embedding from sparse text embedding, ensuring both semantic alignment and balanced information capacity between image and text embeddings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 348 + ], + "type": "text", + "content": "- State-of-the-Art Performance: We evaluate D2S-VSE with existing fine-grained methods across various model backbones. D2S-VSE outperforms the latest state-of-the-art methods on image-text retrieval benchmarks, including Flickr30K and MS-COCO." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 362, + 400, + 374 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 383, + 437, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 437, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 437, + 396 + ], + "type": "text", + "content": "2.1. Image-Text Matching" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 617 + ], + "type": "text", + "content": "Existing Image-text retrieval approaches can be broadly categorized into two groups based on the granularity of the matching similarity: Global-level matching and Local-level matching methods. Global-level matching methods project images and texts into a shared embedding space, computing similarity using improved loss functions [5, 9] to bring semantically aligned pairs closer together. To improve the quality of the embedding space, recent related works proposed several more sophisticated networks, such as graph convolutional network [22, 37], Generalized Pooling Operator [3], and other model architectures [14, 18, 19]. In order to facilitate the closer alignment of semantically corresponding image-text pairs. Different from global matching methods, local matching methods[2, 7, 25, 34, 38, 39, 42, 43] focus more on the alignment between regions. Most of them capture the finer-grained relationship between images and texts by adopting cross-modal networks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Information Density Issues. Recent studies have proposed various methods to address the issue of differing information density between images and text. PCME [5] proposed a probabilistic cross-modal embedding to properly represent the one-to-many relationships in joint embedding. DivE [15] introduced a slot attention-based set prediction module, where the elements of the embedding set compete it-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21680" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 358 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 358 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 358 + ], + "type": "image", + "image_path": "3528f257f5d9b0d3a36000fdf470d305c13e2f8bf70af8ff136da6d2153e2510.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 379, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 379, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 555, + 449 + ], + "type": "text", + "content": "Figure 2. Overview of our D2S-VSE framework. (a) Dense Text Generation: D2S-VSE utilizes the LLaVa [23] to generate dense text descriptions for images in the dataset that contain all the details of the image. (b) Pre-train With Dense Descriptions: D2S-VSE first uses contrastive learning to pre-train the model on image-dense-text pairs to enhance the information capacity of visual semantic embeddings. (c) Finetune With the Sparse Descriptions: D2S-VSE performs distillation from dense to sparse while contrasting the alignment, allowing sparse text embeddings to learn more details of dense text embeddings, thereby improving the information capacity of sparse text embeddings and better aligning them with image embeddings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 466, + 295, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 527 + ], + "type": "text", + "content": "eratively to aggregate the input data. Under the framework of PCME, PCME++ [4] introduced a new probabilistic distance with a closed-form solution. AVSE [26] proposed a matching module for compute the dynamic similarity between asymmetric multi-view image and text embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 563 + ], + "type": "text", + "content": "These methods learn a set of embeddedings to match ambiguous samples, but reduce the information capacity of each embedding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 574, + 259, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 259, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 259, + 588 + ], + "type": "text", + "content": "2.2. Vision language model with dense text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "In the field of vision-language research, few studies have leveraged long-text descriptions to enhance the multimodal representation of images. DreamLip [44] utilizes a multimodal large language model to generate dense text descriptions and randomly extracts sentence fragments for pre-training, but it does not fully exploit long-text pretraining. Long-CLIP [41] fine-tunes CLIP-pre-trained on a short-text-image dataset—on long-text-image pairs to unlock long-text capabilities, yet it overlooks short-text effectiveness. LoTLIP [40] directly pre-trains on long-text" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 466, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 555, + 501 + ], + "type": "text", + "content": "image data and introduces corner tokens to aggregate diverse textual information, preserving the ability to handle short text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": "Different from the above methods that focus on pretraining, our method utilizes a very small amount of long text for pre-training, aiming to use the long text to enhance the information capacity of visual semantic embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 559, + 555, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 555, + 588 + ], + "type": "text", + "content": "3. Dense-to-Sparse Feature Distilled Visual Semantic Embedding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Our goal is to enhance Visual Semantic Embedding (VSE) models to address the challenge of inconsistent information density between images and text. Existing VSE models typically align images with the sparse textual descriptions provided in datasets. This approach biases the image encoder toward textual perspectives, limiting its ability to fully comprehend visual content and making it difficult to handle ambiguous retrieval cases. To overcome this issue, we propose a two-stage training procedure. An overview of our D2S-VSE framework is illustrated in Figure 2." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21681" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 263 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 156 + ], + "type": "text", + "content": "- Stage I. We leverage the pre-trained multimodal large model LLaVa to generate dense textual descriptions for images. The VSE model is then pre-trained using these image-dense text pairs. By training on data with comparable information density, the model learns embeddings with greater information capacity, enabling a more comprehensive understanding of visual content (Sec. 3.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 156, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 156, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 156, + 294, + 263 + ], + "type": "text", + "content": "- **Stage II.** We fine-tune the pre-trained VSE model on general datasets containing image-sparse text pairs. Simultaneously, we distill knowledge from the dense text features into the sparse text representations, enhancing their information capacity. While aligning image-text embeddings, we also align the information capacity between visual and textual modalities. This enables the VSE model to effectively handle varying information densities in images and texts (Sec. 3.2)." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 274, + 251, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 251, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 251, + 288 + ], + "type": "text", + "content": "3.1. Pre-training with Dense Descriptions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 377 + ], + "type": "text", + "content": "To address the issue of inconsistent information density between images and text, we introduce a pre-training stage where the Visual Semantic Embedding (VSE) model learns from image-dense text pairs. This stage enhances the model's ability to extract rich semantic representations from images by ensuring that textual descriptions contain a level of detail comparable to the visual information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 563 + ], + "type": "text", + "content": "Dense Text Generation. To generate dense textual descriptions for images, we leverage the pre-trained multimodal large model LLaVa [23], which integrates vision and language understanding to generate detailed descriptions based on image inputs. Unlike traditional dataset texts, which are often brief and sparse, LLaVa produces detailed descriptions that include object attributes, relationships, actions, and contextual information. Given a raw image " + }, + { + "bbox": [ + 55, + 383, + 296, + 563 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 383, + 296, + 563 + ], + "type": "text", + "content": ", we feed it into the LLaVa with the prompt \"Describe this image in detail\". LLaVa then produces a dense textual description that captures fine-grained details of the image. The hyperparameters used for LLaVa are detailed in Sec 4.2. This process ensures that the textual input provides a comprehensive representation of the image, reducing the information density gap between modalities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 568, + 294, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 294, + 640 + ], + "type": "text", + "content": "Image-Dense-Caption Pre-train. Once the dense captions are generated, we pre-train the VSE model using these image-dense text pairs. The objective of this stage is to enable the model to learn embeddings that effectively capture fine-grained visual semantics and their corresponding textual descriptions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "We employ a contrastive learning framework where the image encoder " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": " and the dense text encoder " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^d" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": " map images and dense textual descriptions into a shared embedding space. Formally, given a dense textual description " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "T^d" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": " and a raw image as " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": ", our goal is to map these inputs into compact vectors " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^d \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{i} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": " such that their semantic" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 552, + 293 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 293 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 293 + ], + "type": "image", + "image_path": "0f64ed5e7bfcfd77a3f28f6584a0d702661448cb09ea1092c7f7d4ad395b1efb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 304, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 370 + ], + "type": "text", + "content": "Figure 3. Architecture for the sparse text branch. Similar to the decoder in MAE [13], we introduce a set of learnable tokens in the sparse text branch to predict the implicit context of sparse text through the decoder. The learnable tokens help in reconstructing the missing semantics, allowing the model to better capture the underlying context and relationships within the sparse text." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "type": "text", + "content": "relationships can be quantified through a similarity metric " + }, + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "type": "inline_equation", + "content": "S(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 392, + 555, + 427 + ], + "type": "text", + "content": ". Specifically, the similarity function is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 388, + 437, + 554, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 437, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 388, + 437, + 554, + 464 + ], + "type": "interline_equation", + "content": "S (I, T) = \\frac {\\mathbf {i} ^ {T} \\mathbf {t} ^ {d}}{| | \\mathbf {i} | | \\cdot | | \\mathbf {t} ^ {d} | |} \\tag {1}", + "image_path": "6ff15ec3f11981add40bb76e4d9a721102f284fadaa6e69fc021a2710f321ce5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 554, + 529 + ], + "type": "text", + "content": "To optimize the model, we use triplet loss function with the hardest negative samples to minimize the similarity between relevant pairs and to maximize the distance between irrelevant ones. The optimized loss function is mathematically formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 538, + 553, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 538, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 323, + 538, + 553, + 591 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {p r e t r a i n}} = \\sum_ {(I, T ^ {d}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {d}) + S (I, \\hat {T} ^ {d}) \\right] _ {+} \\tag {2} \\\\ + \\left[ \\alpha - S (I, T ^ {d}) + S (\\hat {I}, T ^ {d}) \\right] _ {+} \\\\ \\end{array}", + "image_path": "4be5e71b37b98640f033429601622347d91eb96cacea821ea15ee5547cbfc08c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": " serves as a margin parameter and " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "[x]_{+} \\equiv Max(x,0)" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": ". In dataset " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": ", the similarity in a positive pair " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "S(I,T^{d})" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": " should be higher than that in the hardest negative pairs " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "S(\\hat{I},T^{d})" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "S(I,\\hat{T}^{d})" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": " by a margin " + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 601, + 555, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 659, + 495, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 495, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 495, + 673 + ], + "type": "text", + "content": "3.2. Finetune With Sparse Description" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Stage I pre-training with dense descriptions enhances the information capacity of embeddings of the VSE model, but there exists a gap between the training and testing data, as" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21682" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 550, + 329 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 550, + 329 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 550, + 329 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 550, + 329 + ], + "type": "table", + "html": "
MethodFGFlickr30k 1K TestMS-COCO 5-fold 1K TestMS-COCO 5K Test
Text RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSumText RetrievalImage RetrievalrSum
R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10R1R5R10
ViT-Base-224 + BERT-base
VSE++ [9]X71.892.896.559.484.790.9496.175.094.698.062.789.494.9514.662.387.693.443.973.683.3413.4
SCAN [17]✓69.590.995.656.483.190.0485.676.095.498.164.590.895.8520.653.981.890.042.972.382.5423.5
SGR [7]✓69.790.895.259.184.189.9488.777.295.098.065.190.795.8521.854.982.890.542.872.282.5425.8
CHAN [29]✓69.291.895.058.484.990.6489.977.195.198.165.091.096.0522.256.383.290.143.072.682.8428.0
LAPS [11]✓74.093.497.462.587.392.7507.378.795.598.366.291.396.2526.357.584.090.844.574.083.6434.4
AVSE [26]X76.094.697.562.788.493.1512.379.895.698.367.091.596.3528.558.884.391.045.174.383.9437.4
D2S-VSEX82.896.198.368.591.394.9531.980.197.099.268.192.596.7533.760.185.592.546.375.985.2445.6
ViT-Base-384 + BERT-base
VSE++ [9]X77.195.797.565.890.294.3520.577.095.798.464.691.196.2523.054.982.890.442.472.482.8425.8
SCAN [17]✓75.494.496.963.688.693.5512.576.195.598.565.191.696.3523.153.381.890.042.672.682.9423.1
SGR [7]✓76.994.998.164.288.493.3515.875.895.798.665.692.096.5524.253.381.089.642.973.183.7423.6
CHAN [29]✓75.494.597.663.288.693.1512.478.195.898.666.182.196.6527.355.683.891.243.473.683.5431.1
LAPS [11]✓79.096.098.167.390.594.5525.478.696.398.968.092.496.8531.057.484.992.546.475.885.2442.2
AVSE [26]X80.396.498.767.991.294.7529.281.197.199.068.392.797.0535.261.286.893.246.275.985.0448.3
D2S-VSEX84.197.599.170.391.695.3537.980.897.299.169.092.996.8535.860.686.593.246.876.485.7449.1
", + "image_path": "55c49958cffaf328941b4ff65658bed08c2d3ddd3fbbe7dbc98488e2d3f031da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Table 1. Comparison of experimental results using ViT-Base [8] and BERT [6] the MS-COCO and Flickr30K datasets. \"FG\" indicates whether fine-grained cross-modal alignment is applied. The best performance is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 381, + 294, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 381, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 294, + 548 + ], + "type": "text", + "content": "the model is trained on dense texts and tested on sparse texts, leading to suboptimal performance. However, simply fine-tuning directly on the image-sparse text pair data will not make the sparse text embedding contain richer semantics, that is, more information capacity. To address this problem, we introduce a novel distillation mechanism while aligning, which treats sparse text as masked dense text and designs a decoder. Through distillation, the decoder can decode sparse text embeddings into dense text embeddings, thereby improving the information capacity of sparse text embeddings. This mechanism extracts sparse text embeddings from dense text embeddings while aligning images and text, thereby enhancing the information capacity of sparse text features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 578 + ], + "type": "text", + "content": "3.2.1. Aligning Information Capacity by Dense-to-Sparse Feature Distillation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "To ensure that the sparse text embedding retains as much information as the dense text embedding, we introduce a feature distillation mechanism from dense to sparse. Inspired by masked signal modeling [6, 13], we treat sparse text as a masked version of dense text. This allows us to decode and reconstruct the sparse text embedding through feature distillation, enabling the reconstructed embedding to approximate the information capacity of the dense text embedding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Transformer-based Decoder. To achieve this reconstruction, we design a Transformer-based decoder, as illustrated" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": "in Figure 3. The decoder plays a pivotal role in D2S-VSE, responsible for reconstructing the dense text embedding from the sparse text embedding by leveraging feature distillation. It learns to restore the missing contextual information in sparse text, ensuring that the final sparse text embedding retains as much semantic richness as the dense text embedding. Given a sparse text " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "T^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": ", we first feed it into the sparse text encoder " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " to obtain the word embedding " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " and the sparse text embedding " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": ". Next, learnable mask tokens with position " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " are concatenated with " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{w}^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " and input into the decoder for processing. The decoder outputs the reconstructed tokens " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": ", which are then averaged to obtain " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{t}}" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": ". Finally, we sum " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{t}}" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": " to generate the final sparse text embedding " + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}^s" + }, + { + "bbox": [ + 313, + 381, + 555, + 548 + ], + "type": "text", + "content": ". This process can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 372, + 558, + 555, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 558, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 372, + 558, + 555, + 572 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {t}} ^ {s} = \\operatorname {D e c o d e r} \\left(\\left[ \\mathbf {w} ^ {s}, \\mathbf {m} \\right]\\right) + \\mathbf {t} ^ {s} \\tag {3}", + "image_path": "0bb7ce7902fec6470e1326c0ef2d548b664e608795598f3901a489ec57a93772.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 640 + ], + "type": "text", + "content": "This reconstruction process enables the mask tokens to capture the contextual information missing from the sparse text, aligning it more closely with the dense text embedding. By combining the two, we obtain a more comprehensive and enriched text feature embedding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "Dense-to-Sparse Distillation. Through the dense text pretraining in stage 1, the representation of the pre-trained VSE model can capture richer semantics, and the pre-trained dense text encoder can be used as a teacher to distill the information capacity of dense text embeddings into sparse text embeddings. This Dense-to-Sparse distillation mini-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21683" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "mizes a negative cosine similarity between then dense text embedding " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^d" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " and the final sparse text embedding " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}^s" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 105, + 295, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 105, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 104, + 105, + 295, + 135 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t i l l}} \\left(\\mathbf {t} ^ {d}, \\hat {\\mathbf {t}} ^ {s}\\right) = 1 - \\frac {\\mathbf {t} ^ {d} \\cdot \\hat {\\mathbf {t}} ^ {s}}{\\left| \\left| \\mathbf {t} ^ {d} \\right| \\right| \\cdot \\left| \\left| \\hat {\\mathbf {t}} ^ {s} \\right| \\right|} \\tag {4}", + "image_path": "7ced1a7d0ba42acb0aebada92324a25665992fed227b78a905386f5ba7933a25.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 269, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 269, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 269, + 154 + ], + "type": "text", + "content": "3.2.2. Aligning Visual Semantic Using Sparse text" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": "At the same time, we also fine-tune the VSE model on image-sparse text pair data to adapt the model to common sparse text queries in cross-modal retrieval. Given image " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": " and sparse text " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "T^s" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": ", we extract image embedding " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{i}" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": " and sparse text embedding " + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 156, + 296, + 228 + ], + "type": "text", + "content": " and optimize the VSE model as follow:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 234, + 296, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 234, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 71, + 234, + 296, + 287 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {a l i g n}} = \\sum_ {(I, T ^ {s}) \\in \\mathcal {D}} \\left[ \\alpha - S (I, T ^ {s}) + S (I, \\hat {T} ^ {s}) \\right] _ {+} \\tag {5} \\\\ + \\left[ \\alpha - S (I, T ^ {s}) + S (\\hat {I}, T ^ {s}) \\right] _ {+} \\\\ \\end{array}", + "image_path": "36e4616cc19963d295559d56072e74f0bf211d6224d4d71ee785fc04409cb533.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 319 + ], + "type": "text", + "content": "In summary, the final loss function is defined as follows to perform joint optimization of the two objectives." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 329, + 295, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 329, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 130, + 329, + 295, + 342 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {a l i g n}} + \\mathcal {L} _ {\\text {d i s t i l l}} \\tag {6}", + "image_path": "f963c77d0e51d50e4e5ef21302a0c7d69ded8e1b61daa9af72da40a407f3a841.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 352, + 137, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 137, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 137, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 372, + 174, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 174, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 174, + 384 + ], + "type": "text", + "content": "4.1. Dataset and Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "type": "text", + "content": "Datasets. Following previous works [9], we conduct our experiments on two widely used benchmark datasets: MS-COCO [21] and Flickr30K [30]. MS-COCO consists of 123,287 images, with each image annotated with five text captions. Following the dataset split in [9], we use 113,287 images for training, 5,000 for validation, and 5,000 for testing. Flickr30K contains 31,783 images, each paired with five textual descriptions. We adopt the standard split from [9], using 29,000 images for training, 1,000 for validation, and 1,000 for testing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "text", + "content": "Evaluation Metrics. For image-text matching, we use the commonly adopted Recall@K (where " + }, + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 1,5,10" + }, + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "text", + "content": "), denoted as R@1, R@5, and R@10. These metrics represent the percentage of ground-truth instances retrieved within the top 1, 5, or 10 results, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 576, + 187, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 187, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 187, + 589 + ], + "type": "text", + "content": "4.2. Implementation details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": "For dense text generation, we use LLaVa [23] to generate dense textual descriptions. The hyperparameters for LLaVa are set as follows: Top-P = 0.9, temperature = 0.2, and max-new-tokens = 500. We generate a dense text description for each image in the MS-COCO and Flickr30K datasets, with each dense text sentence corresponding to " + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\frac{1}{5}" + }, + { + "bbox": [ + 55, + 594, + 295, + 677 + ], + "type": "text", + "content": " of the sparse text descriptions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In the pre-training stage, we use the AdamW optimizer [28] to train the VSE model for 30 epochs, with a minibatch size of 128. The learning rate is initially set to 0.0005" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 388 + ], + "type": "table", + "html": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Pre-trained ResNet-152 on IN + GRU
VSE∞ [3]76.594.297.756.483.489.9498.1
MV-VSE [20]79.094.997.759.184.690.6505.9
NAAF* [42]81.996.198.361.085.390.6513.2
HREM* [10]81.496.598.560.985.691.3514.2
CHAN [29]79.794.597.360.285.390.7507.7
AVSE [26]82.396.298.562.888.092.8520.6
D2S-VSE82.596.699.062.688.193.1521.9
Swin-Base-224 + BERT-base
VSE++ [9]82.596.598.970.091.495.1534.4
SCAN [17]79.095.998.267.790.694.9526.3
SGR [7]80.497.098.766.990.294.5527.6
CHAN [29])81.497.098.668.590.694.5530.6
LAPS[11]82.497.499.570.091.795.4536.3
AVSE [26]83.997.499.470.092.495.6538.7
D2S-VSE87.298.499.973.093.596.7548.7
Swin-Base-384 + BERT-base
VSE++ [9]83.397.599.271.193.296.2540.6
SCAN [17]81.996.998.970.092.795.8536.1
SGR [7]80.796.899.069.991.795.3533.4
CHAN [29]81.296.798.870.392.295.9535.0
LAPS [11]85.197.799.274.093.096.3545.3
AVSE [26]87.198.399.273.693.596.5548.2
D2S-VSE87.899.099.775.794.196.9553.2
", + "image_path": "c2ba2819b5c59a3b3c4cfc40e484d785478ce96423b0757c7eb08da51658148a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "type": "text", + "content": "Table 2. Comparison of experimental results using other common backbones on the Flickr30K dataset. The best performance is highlighted in **.** indicates ensemble methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 520 + ], + "type": "text", + "content": "for the first few epochs and is reduced by a factor of 10 during the 9th, 15th, 20th, and 25th epochs. In the fine-tuning stage, we freeze the dense text encoder and fine-tune both the image encoder and sparse text encoder. We train them using the same training strategy as in the pre-training stage and evaluate the best model on the validation set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 641 + ], + "type": "text", + "content": "For model details, we use traditional region features and employ the recently popular Vision Transformer (ViT) [8] and Swin Transformer [27] as the backbone. We set the dimension of the shared embedding space to 512. For feature aggregation, we use GPO [3] for both the image encoder and the text encoder. The decoder consists of 4 layers of Transformer, with 4 attention heads per layer. We set 100 learnable tokens to predict the context of sparse text embeddings. We provide a more detailed implementation setup in the supplementary material." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 648, + 537, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 537, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 537, + 661 + ], + "type": "text", + "content": "4.3. Comparisons with Recent State-of-the-arts" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We compare our D2S-VSE with the most recent state-of-the-art approaches on two widely used datasets, i.e., Flirkr30K and MS-COCO. For the sake of fairness, our D2S-VSE employs a diverse range of backbones to facili" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21684" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 75, + 218, + 159 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 218, + 159 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 218, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 218, + 159 + ], + "type": "image", + "image_path": "b60994adbd0779743cbcb8b7edb52e9193de6f5c14ce216b1ed81dc0ecf39168.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 167, + 197, + 176 + ], + "lines": [ + { + "bbox": [ + 81, + 167, + 197, + 176 + ], + "spans": [ + { + "bbox": [ + 81, + 167, + 197, + 176 + ], + "type": "text", + "content": "(a) The impact of Learnable Tokens" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 75, + 383, + 159 + ], + "blocks": [ + { + "bbox": [ + 227, + 75, + 383, + 159 + ], + "lines": [ + { + "bbox": [ + 227, + 75, + 383, + 159 + ], + "spans": [ + { + "bbox": [ + 227, + 75, + 383, + 159 + ], + "type": "image", + "image_path": "73ffa7a323a7e7ae6115f36ee47f3b0d9eba35a73f870d602ccbbb05bb24a59c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 167, + 359, + 177 + ], + "lines": [ + { + "bbox": [ + 251, + 167, + 359, + 177 + ], + "spans": [ + { + "bbox": [ + 251, + 167, + 359, + 177 + ], + "type": "text", + "content": "(b) The impact of Decoder Depth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": "Figure 4. Ablation study of the architecture of decoder on Flickr30K. If not specified, the default is: the number of the learnable tokens is 100, the decoder has 4 layers and 4 head." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 76, + 549, + 159 + ], + "blocks": [ + { + "bbox": [ + 394, + 76, + 549, + 159 + ], + "lines": [ + { + "bbox": [ + 394, + 76, + 549, + 159 + ], + "spans": [ + { + "bbox": [ + 394, + 76, + 549, + 159 + ], + "type": "image", + "image_path": "288db231dbc846eae6a3093b2cf3ce429dde15b05ae06a89d005f626f78f8544.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 167, + 523, + 176 + ], + "lines": [ + { + "bbox": [ + 419, + 167, + 523, + 176 + ], + "spans": [ + { + "bbox": [ + 419, + 167, + 523, + 176 + ], + "type": "text", + "content": "(c) The impact of Decoder Head" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 230, + 253, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 253, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 253, + 241 + ], + "type": "text", + "content": "tate comparisons with various previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": "In Table 1, we present the results of using ViT-base [8] + BERT-base [6] as the backbone on the Flickr30K and MS-COCO datasets. Our D2S-VSE surpasses all state-of-the-art methods, demonstrating remarkable performance. In this setting, we compare our method against the two leading approaches, LAPS [11] and AVSE [26], and achieve superior results across nearly all metrics. For instance, on the Flickr30K dataset, our approach outperforms AVSE in text retrieval, achieving a " + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "inline_equation", + "content": "6.8\\%" + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": " improvement in " + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathsf{R}@\\mathsf{1}" + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": " with the ViT-base-224 backbone and a " + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": " improvement with the ViT-base-384 backbone. Similarly, for image retrieval, our method achieves gains of " + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 54, + 242, + 296, + 420 + ], + "type": "text", + "content": " when using the ViT-base-224 and ViT-base-384 backbones, respectively. A consistent performance boost is also observed across all metrics on the MS-COCO dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "content": "Additionally, we conducted experiments using other commonly used backbones, as shown in Table 2. Specifically, we evaluated our approach on the Flickr30K dataset with Faster R-CNN [1] + GRU and Swin-Base [27] + BERT-Base as backbones. The results further confirm that our method consistently outperforms existing approaches across different backbone architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 149, + 528 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "type": "text", + "content": "To verify the effectiveness of each component of our D2S-VSE, we conduct extensive ablation studies on the Flickr30K dataset. All experiments are performed using ViT-base-224 and BERT-base as the backbone to ensure a consistent and fair comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 676 + ], + "type": "text", + "content": "The impact of the two-stage training framework. Our D2S-VSE introduces a two-stage training framework to enhance the information capacity of image embedding and text embedding, and achieves very good results on commonly used datasets. In order to explore the role of each component in the framework, we conducted ablation experiments as shown in Table 3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "content": "Is the model pre-trained with dense text inherently superior, leading to better performance? To verify this, we directly tested the pre-trained model on the Flickr30K test" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 228, + 549, + 312 + ], + "blocks": [ + { + "bbox": [ + 321, + 228, + 549, + 312 + ], + "lines": [ + { + "bbox": [ + 321, + 228, + 549, + 312 + ], + "spans": [ + { + "bbox": [ + 321, + 228, + 549, + 312 + ], + "type": "table", + "html": "
PretrainFinetuneText RetrievalImage RetrievalrSum
AlignDistillR@1R@5R@10R@1R@5R@10
✓56.284.791.641.771.379.7425.2
✓76.393.997.361.287.693.2509.5
✓✓79.296.598.366.890.594.4525.7
✓✓76.094.597.362.988.393.1512.1
✓✓✓82.896.198.368.591.394.9531.9
", + "image_path": "faaabf235b01f9568d33a4cd6eac32c3b34d258abf5a10c511ac27913a33d09a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "text", + "content": "Table 3. Ablation experiment on the impact of our framework on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 376, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 412 + ], + "type": "text", + "content": "set. We found that due to the gap between dense and sparse text, the pre-trained model struggles to adapt to sparse text, resulting in rSum of only 425.2 on the Flickr30K test set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 496 + ], + "type": "text", + "content": "The pre-training stage enhances the information capacity of image embeddings! We observed that the R@1 of text retrieval improved by " + }, + { + "bbox": [ + 313, + 413, + 554, + 496 + ], + "type": "inline_equation", + "content": "2.9\\%" + }, + { + "bbox": [ + 313, + 413, + 554, + 496 + ], + "type": "text", + "content": " after pre-training compared to the model without pre-training. This suggests that dense text pre-training enhances the information capacity of image embeddings, enabling the model to better align multi-view descriptions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "content": "Dense-to-Sparse feature distillation enhances the information capacity of sparse text embeddings! Applying dense-to-sparse distillation while aligning images during the fine-tuning stage further boosts the model's retrieval performance. On the Flickr30K dataset, text retrieval and image retrieval reached " + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "inline_equation", + "content": "82.8\\%" + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 313, + 497, + 554, + 605 + ], + "type": "text", + "content": ", respectively. This improvement stems from distillation enhances the information capacity of text embeddings, leading to better semantic consistency between image and text embeddings." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 664 + ], + "type": "text", + "content": "Different type of distillation functions. We used L1 distance, L2 distance, and negative cosine similarity as the loss functions for Dense-to-Sparse distillation, respectively. As shown in Table 4, the effect of negative cosine similarity distillation is much better than the other two." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Decoder design. The decoder is the most important part of Dense-to-Sparse distillation. Its existence allows learnable tokens to predict possible contexts based on the word embeddings of sparse text, thereby increasing the informa" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21685" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 216, + 116 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 216, + 116 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 216, + 116 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 216, + 116 + ], + "type": "image", + "image_path": "dc488757593a235bb1dd6d7232488b96f4d365e05cb246425706934912cc9ca9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 118, + 206, + 128 + ], + "lines": [ + { + "bbox": [ + 115, + 118, + 206, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 118, + 206, + 128 + ], + "type": "text", + "content": "(a) Prefix Learnable Tokens" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 70, + 359, + 116 + ], + "blocks": [ + { + "bbox": [ + 250, + 70, + 359, + 116 + ], + "lines": [ + { + "bbox": [ + 250, + 70, + 359, + 116 + ], + "spans": [ + { + "bbox": [ + 250, + 70, + 359, + 116 + ], + "type": "image", + "image_path": "24a8c85839378f2a7f0d6fe44c3020aa2eaaa0d2e5ca316f063dadfed9f34e1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 118, + 356, + 128 + ], + "lines": [ + { + "bbox": [ + 255, + 118, + 356, + 128 + ], + "spans": [ + { + "bbox": [ + 255, + 118, + 356, + 128 + ], + "type": "text", + "content": "(a) Surround Learnable Tokens" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 138, + 555, + 161 + ], + "lines": [ + { + "bbox": [ + 55, + 138, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 555, + 161 + ], + "type": "text", + "content": "Figure 5. Different type of the position strategy of learnable tokens. Tokens at different positions can predict the semantics of different positions in dense text." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 395, + 71, + 504, + 116 + ], + "blocks": [ + { + "bbox": [ + 395, + 71, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 395, + 71, + 504, + 116 + ], + "spans": [ + { + "bbox": [ + 395, + 71, + 504, + 116 + ], + "type": "image", + "image_path": "48b9a4dcc44b8a830bf6286b971b80357f3b8d652d5e579ff6e5636c68344f86.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 118, + 496, + 128 + ], + "lines": [ + { + "bbox": [ + 402, + 118, + 496, + 128 + ], + "spans": [ + { + "bbox": [ + 402, + 118, + 496, + 128 + ], + "type": "text", + "content": "(c) Postfix Learnable Tokens" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 178, + 291, + 242 + ], + "blocks": [ + { + "bbox": [ + 61, + 178, + 291, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 291, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 291, + 242 + ], + "type": "table", + "html": "
MethodText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
\\( \\ell_1 \\) Distance81.995.398.268.090.894.7528.9
\\( \\ell_2 \\) Distance82.195.698.467.991.295.1530.2
Negative Cosine Similarity82.896.198.368.591.394.9531.9
", + "image_path": "34e9e49bd2ed3035ff452df12a67e81ca5b3887cc98d8fb595274e26aab6fed3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 178, + 553, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 249, + 296, + 283 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 283 + ], + "type": "text", + "content": "Table 4. Ablation experiment on the impact of different distillation functions on Flick30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 178, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 317, + 178, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 553, + 242 + ], + "type": "table", + "html": "
TypeText RetrievalImage RetrievalrSum
R@1R@5R@10R@1R@5R@10
Prefix81.795.398.268.191.294.8529.3
Postfix82.095.698.367.891.194.7529.5
Surround82.896.198.368.591.394.9531.9
", + "image_path": "7e9a21f87fb3d5b4621f8fcb80f323805bd9e989210217149651b5ae1c32f278.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 249, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 283 + ], + "type": "text", + "content": "Table 5. Ablation experiment on the impact of the position of the learnable tokens on Flickr30K using ViT-base-224 + BERT-base as backbones. The default setting is highlighted in Light Blue." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 304, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 341 + ], + "type": "text", + "content": "tion capacity of sparse text features. As shown in Figure 4, we conduct ablation studies of Learnable tokens, decoder depth, and decoder head." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "text", + "content": "The number of tokens determines the reconstruction capability of the decoder. As shown in Figure 4a, we observed that as the number of learnable tokens in the decoder increases, the model's performance in both image retrieval and text retrieval gradually improves, stabilizing when the number of tokens reaches approximately 100. At this point, the R@1 for text retrieval and image retrieval reaches " + }, + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "inline_equation", + "content": "82.8\\%" + }, + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "inline_equation", + "content": "68.5\\%" + }, + { + "bbox": [ + 54, + 342, + 295, + 473 + ], + "type": "text", + "content": ", respectively. However, when the number of tokens increases to 120, the model shows no significant advantage. Therefore, considering both performance and efficiency, we set the number of tokens to 100." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "text", + "content": "Surrounding learnable tokens can better predict the context of sparse text. In addition, we conducted an ablation experiment on the insertion position of learnable tokens. Since our motivation stems from treating sparse text as masked dense text, with learnable tokens used to reconstruct the masked semantics, their placement is crucial. As shown in Figure 5, we designed Prefix Learnable Tokens, Surround Learnable Tokens, and Postfix Learnable Tokens to explore the impact of token positioning on model performance. Table 5 presents the ablation results on the impact of different learnable token positions. We observe that the rSum index of Surround Learnable Tokens is " + }, + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 56, + 474, + 295, + 689 + ], + "type": "text", + "content": " higher than that of Prefix Learnable Tokens and Postfix Learnable Tokens, respectively. This outcome is intuitive, as the exact location of sparse text semantics within dense text is unknown. The Surround insertion method allows the model to predict semantics from both the preceding and following contexts simultaneously, leading to improved performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": "A deep decoder can improve retrieval performance. As shown in Figure 4b, as the decoder depth increases, retrieval" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 426 + ], + "type": "text", + "content": "performance first improves and then declines, with the best results achieved when the number of layers is set to 4. Although our approach is inspired by Masked Signal Modeling [13], our findings regarding the decoder differ. In Masked Signal Modeling, the objective is to enhance the encoder through reconstruction, where an excessively deep decoder may hinder the encoder's capability. However, in our D2S-VSE, retrieval performance relies on the predictive ability of the decoder, meaning that a deeper decoder generally leads to better results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 498 + ], + "type": "text", + "content": "Our D2S-VSE can work with very few decoder heads. As shown in Figure 4c, the number of attention heads has minimal impact on performance, with only 4 heads required to achieve optimal results. Increasing the number of heads does not lead to significant improvements, as all performance differences remain within " + }, + { + "bbox": [ + 313, + 426, + 556, + 498 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 391, + 525 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we highlight the importance of embedding information capacity in addressing the multi-view description problem in image-text matching tasks. To tackle this challenge, we propose a novel Visual Semantic Embedding (VSE) framework, Dense-to-Sparse Feature Distillation for Visual Semantic Embedding (D2S-VSE). The core idea of D2S-VSE is to enhance the information capacity of visual semantic embeddings by aligning dense text with images during pre-training. In the fine-tuning stage, we further refine sparse text embeddings through Dense-to-Sparse feature distillation, effectively improving their information capacity and leading to better image-text matching performance. Comprehensive experiments on two widely-used benchmarks validate the effectiveness of the proposed method, leading to state-of-the-art performance." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21686" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 183 + ], + "type": "text", + "content": "Acknowledgements. This work was partially supported by the National Major Scientific Instruments and Equipments Development Project of National Natural Science Foundation of China under Grant 62427820, the National Science Foundation of China under Grant 62376175 and 22494712, the 111 Project under Grant B21044, the Science Fund for Creative Research Groups of Sichuan Province Natural Science Foundation under Grant 2024NSFTD0035, and the National Science Foundation of Sichuan Province under Grant 2025ZNSFSC0480." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 224, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 224, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 294, + 277 + ], + "type": "text", + "content": "[1] Peter Anderson, Xiaodong He, Chris Buehler, Damien Teney, Mark Johnson, Stephen Gould, and Lei Zhang. Bottom-up and top-down attention for image captioning and visual question answering. In CVPR, pages 6077-6086, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 294, + 323 + ], + "type": "text", + "content": "[2] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In CVPR, pages 12655-12663, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 324, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 367 + ], + "type": "text", + "content": "[3] Jiacheng Chen, Hexiang Hu, Hao Wu, Yuning Jiang, and Changhu Wang. Learning the best pooling strategy for visual semantic embedding. In CVPR, pages 15789-15798, 2021. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 390 + ], + "type": "text", + "content": "[4] Sanghyuk Chun. Improved probabilistic image-text representations. arXiv preprint arXiv:2305.18171, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "text", + "content": "[5] Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio De Rezende, Yannis Kalantidis, and Diane Larlus. Probabilistic embeddings for cross-modal retrieval. In CVPR, pages 8415-8424, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[6] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 513 + ], + "type": "text", + "content": "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In AAAI, pages 1218-1226, 2021. 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 579 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.5,6,7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 581, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 294, + 613 + ], + "type": "text", + "content": "[9] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. In BMVC, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[10] Zheren Fu, Zhendong Mao, Yan Song, and Yongdong Zhang. Learning semantic relationship among instances for image-text matching. In CVPR, pages 15159-15168, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[11] Zheren Fu, Lei Zhang, Hou Xia, and Zhendong Mao. Linguistic-aware patch slimming framework for fine-grained cross-modal alignment. In CVPR, pages 26307-26316, 2024. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[12] Jieuxiang Gu, Jianfei Cai, Shafiq R Joty, Li Niu, and Gang Wang. Look, imagine and match: Improving textual-visual" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "cross-modal retrieval with generative models. In CVPR, pages 7181-7189, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 4, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[14] Yan Huang, Qi Wu, Chunfeng Song, and Liang Wang. Learning semantic concepts and order for image and sentence matching. In CVPR, pages 6163-6171, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "text", + "content": "[15] Dongwon Kim, Namyup Kim, and Suha Kwak. Improving cross-modal retrieval with set of diverse embeddings. In CVPR, pages 23422-23431, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[16] Ryan Kiros, Ruslan Salakhutdinov, and Richard S Zemel. Unifying visual-semantic embeddings with multimodal neural language models. arXiv preprint arXiv:1411.2539, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 275 + ], + "type": "text", + "content": "[17] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In ECCV, pages 201-216, 2018. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 308 + ], + "type": "text", + "content": "[18] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In ICCV, pages 4654-4662, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 342 + ], + "type": "text", + "content": "[19] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Image-text embedding learning via visual and textual semantic reasoning. TPAMI, 45(1):641-656, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 376 + ], + "type": "text", + "content": "[20] Zheng Li, Caili Guo, Zerun Feng, Jenq-Neng Hwang, and Xijun Xue. Multi-view visual semantic embedding. In IJ-CAI, 2022. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 464 + ], + "type": "text", + "content": "[22] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In CVPR, pages 10921-10930, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 467, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 498 + ], + "type": "text", + "content": "[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 533 + ], + "type": "text", + "content": "[24] Yang Liu, Hong Liu, Huaqiu Wang, and Mengyuan Liu. Regularizing visual semantic embedding with contrastive learning for image-text matching. IEEE SPL, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 567 + ], + "type": "text", + "content": "[25] Yang Liu, Hong Liu, Huaqiu Wang, Fanyang Meng, and Mengyuan Liu. Bcan: Bidirectional correct attention network for cross-modal retrieval. TNNLS, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 611 + ], + "type": "text", + "content": "[26] Yang Liu, Mengyuan Liu, shudong Huang, and Jiancheng Lv. Asymmetric visual semantic embedding framework for efficient vision-language alignment. In AAAI, 2025. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Zhengxin Pan, Fangyu Wu, and Bailing Zhang. Fine-grained image-text matching by cross-modal hard aligning network. In CVPR, pages 19275-19284, 2023. 5, 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21687" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[30] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, pages 2641–2649, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[31] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In ACMMM, pages 4948-4956, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[32] Yang Qin, Yuan Sun, Dezhong Peng, Joey Tianyi Zhou, Xi Peng, and Peng Hu. Cross-modal active complementary learning with self-refining correspondence. arXiv preprint arXiv:2310.17468, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[33] Leigang Qu, Meng Liu, Da Cao, Liqiang Nie, and Qi Tian. Context-aware multi-view summarization network for image-text matching. In ACMMM, pages 1047-1055, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[34] Leigang Qu, Meng Liu, Jianlong Wu, Zan Gao, and Liqiang Nie. Dynamic modality interaction modeling for image-text retrieval. In SIGIR, pages 1104-1113, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv preprint arXiv:2103.00020, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[36] Yale Song and Mohammad Soleymani. Polysemous visual-semantic embedding for cross-modal retrieval. In CVPR, pages 1979–1988, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[37] Haoran Wang, Ying Zhang, Zhong Ji, Yanwei Pang, and Lin Ma. Consensus-aware visual-semantic embedding for image-text matching. In ECCV, pages 18–34, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "text", + "content": "[38] Zihao Wang, Xihui Liu, Hongsheng Li, Lu Sheng, Junjie Yan, Xiaogang Wang, and Jing Shao. Camp: Cross-modal adaptive message passing for text-image retrieval. In ICCV, pages 5764-5773, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 507 + ], + "type": "text", + "content": "[39] Jiwei Wei, Xing Xu, Yang Yang, Yanli Ji, Zheng Wang, and Heng Tao Shen. Universal weighting metric learning for cross-modal matching. In CVPR, pages 13005-13014, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 564 + ], + "type": "text", + "content": "[40] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pretraining for long text understanding. Neurips, 37:64996-65019, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 599 + ], + "type": "text", + "content": "[41] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In ECCV, pages 310–325. Springer, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 633 + ], + "type": "text", + "content": "[42] Kun Zhang, Zhendong Mao, Quan Wang, and Yongdong Zhang. Negative-aware attention framework for image-text matching. In CVPR, pages 15661-15670, 2022. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "text", + "content": "[43] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In CVPR, pages 3536-3545, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[44] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamlip: Language-image pre-training with long captions. In ECCV, pages 73-90. Springer, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21688" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_content_list.json b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79c551087daa5f80ed4c20eefd3822a3b068ad1b --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_content_list.json @@ -0,0 +1,1565 @@ +[ + { + "type": "text", + "text": "Aligning Moments in Time using Video Queries", + "text_level": 1, + "bbox": [ + 258, + 130, + 738, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yogesh Kumar1* Uday Agarwal1* Manish Gupta2 Anand Mishra1 \n1Indian Institute of Technology Jodhpur 2Microsoft", + "bbox": [ + 209, + 179, + 784, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{kumar.204, agarwaluday, mishra}@iitj.ac.in, gmanish@microsoft.com", + "bbox": [ + 205, + 218, + 787, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment Transformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of $13.1\\%$ in $R@1$ and $8.1\\%$ in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a $14.7\\%$ gain in $R@1$ and a $14.4\\%$ gain in mIoU on an absolute scale over strong baselines. We make the dataset and code public at: https://github.com/vl2g/MATR.", + "bbox": [ + 89, + 300, + 483, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 752, + 220, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video moment retrieval is the task of temporally localizing the start and end times of a moment in a target video described by a given query. Although text-based video moment retrieval has been extensively explored [25, 28, 32, 33, 57], it often poses challenges for users attempting to", + "bbox": [ + 89, + 777, + 482, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text Query \"A man positions himself beneath the ball, leaping into the air, bending his knees and arching his back. As he flips backward, he connects with the ball using his foot.\"", + "bbox": [ + 514, + 268, + 905, + 301 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/98a07981df1a760724ca4b2e0df71e3e6902dfca5e11499f05c3b39d944adcca.jpg", + "image_caption": [ + "Figure 1. (Top): Activity where text falls short to explain complex action or events, such as bicycle kick, necessitating the need for a more intuitive query modality for localizing semantically matching moments. (Bottom): Shows the Vid2Vid moment retrieval setting, which is the goal of this work." + ], + "image_footnote": [], + "bbox": [ + 516, + 303, + 903, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "describe specific moments verbally. For example, consider a spectacular bicycle kick in soccer (Fig. 1 top). Although an expert can search by naming or describing it in detail, a beginner in soccer may struggle to describe this impressive move accurately. They might say something like \"The player kicked the ball while in the air\", which lacks the nuance needed to convey the athleticism and artistry involved. Further, such a description may lead to poor results for someone searching for similar moments. In contrast, if a soccer trainee shows a short video clip of a bicycle kick, the input is rich and crisp, making it easier for a retriever to locate similar moments in a target soccer match video. This approach aligns with how users naturally would prefer to search content within the same (video) modality, making it a more effective method for retrieving specific moments. Fig. 1 (bottom) shows a video bicycle kick query in soccer that is localized in a target video.", + "bbox": [ + 511, + 590, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, in this work, we study video-to-video moment retrieval (Vid2VidMR) where the aim is to temporally localize a moment in the target video with a high se", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 862, + 220, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1A moment is a continuous segment of frames within a target video that best represents the actions, events, or interactions described by the query.", + "bbox": [ + 91, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20215", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "matic match between the moment and the query video. Vid2VidMR, a task formally introduced by Feng et al. [14], has many potential application areas, such as sports video analytics, educational content creation and e-learning, and surveillance systems. This task is challenging and requires semantic frame-level alignment and modeling complex dependencies between query and target videos. The need for a semantic understanding of video content and variety in video length, context, and action speed calls for adaptive models capable of generalizing across diverse scenarios. Addressing these challenges demands advanced temporal video modeling techniques.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose MATR (Moment Alignment TRansformer) - a method that uses explicit 'dual-stage sequence alignment' to capture the required correlation and temporal details essential for accurate moment localization. By conditioning target video representations on query features, MATR produces query-aligned representations that encode the necessary correlations and temporal dependencies between the two videos. We use differentiable dynamic time warping loss [9] for aligning the query and target videos, and represent the target video by conditioning it on the query video to focus on correlated temporal features. These representations guide a classification head to discriminate relevant moments from the irrelevant background and a boundary prediction head to mark the start and end of the identified moment in the target video. Further, to enhance our model's generalization using unlabeled videos, we introduce a self-supervised pretraining strategy which involves training MATR to localize randomly sampled clips within the same video, enabling it to learn the moment localization skill in a self-supervised manner.", + "bbox": [ + 91, + 275, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate MATR on public ActivityNet-VRL [14] benchmark and on our newly introduced dataset on sports domain, viz. SportsMoments, covering two of the most popular sports, namely soccer and cricket. Our approach achieves significant performance gains, with an improvement of $13.1\\%$ and $8.1\\%$ in R@1 and mIoU, respectively, on ActivityNet-VRL outperforming the state-of-the-art methods. Furthermore, MATR outperforms the implemented strong baselines on our proposed dataset with $14.7\\%$ gains in R@1 and $14.4\\%$ gains in mIoU, all on an absolute scale.", + "bbox": [ + 89, + 595, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows: (i) We introduce MATR, which uses explicit dual-stage sequence alignment within a transformer framework between target and query video to capture temporal correlations and dependencies for accurate moment localization. (ii) We propose a self-supervised pre-training objective that enhances model initialization by understanding rich video structure without requiring any labeled data. (iii) We conduct extensive experiments and ablations to study the efficacy of our", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "framework against competitive baselines and state-of-the-art methods. Our findings offer valuable insights into our design choices, and our approach advances the state-of-the-art on Vid2VidMR.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 167, + 653, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Moment Retrieval (VMR): VMR has recently gained significant interest in the research community [1, 15, 18, 24, 25, 28, 29, 32-34, 53, 55, 57]. Unlike video action understanding tasks such as action classification [6, 13, 27, 44, 47, 54, 56, 60] or temporal action localization [43, 59, 64], VMR focuses on identifying segments that semantically align with a broader range of queries, which may describe complex and context-specific moments. Based on the query modality, existing VMR methods can be broadly grouped into (i) textual query-based approaches like Moment-DETR [25], QD-DETR [33], UniVTG [28], and (ii) video query-based approaches like GDP [7], FFI+SRM [19], SRL [51]. Our work falls under video query-based VMR (Vid2VidMR). However, in addition to developing a new approach tailored for video query, we also adapt several text-query-based methods to make them suitable for Vid2VidMR and compare our method against both text and video query-based methods.", + "bbox": [ + 511, + 191, + 903, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "**Alignment in Videos:** Alignment has been a thrust area in the video understanding community. It has been applied to a wide range of video understanding tasks, including video retrieval [52], procedural steps alignment [11], action recognition [10], anomaly detection [12], movie understanding [2] and video synchronization [35, 40]. Researchers have also explored sequence alignment for the text-VMR task, e.g., Mithun et al. [31] used sequence alignment between CNN representation of frames and GRU representation of text query to perform text-VMR. Jung et al. [20] used alignment to enhance semantic understanding between query and target video at the abstract level for the text-VMR task. However, their alignment is not at the sequence level.", + "bbox": [ + 511, + 465, + 903, + 676 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to the existing literature, MATR goes beyond traditional sequence alignment by introducing a dual-stage alignment mechanism, leveraging transformer-based feature fusion, and maintaining flexibility with multiple alignment strategies. These innovations enable it to achieve more accurate and context-aware video moment retrieval compared to existing approaches.", + "bbox": [ + 511, + 678, + 903, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The MATR Model", + "text_level": 1, + "bbox": [ + 513, + 799, + 691, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our objective is to temporally localize a moment in a target video $V_{t}$ using a query video $V_{q}$ . We represent the target and query videos as sequences of $M$ and $N$ uniformly sampled frames, respectively. We refer to this problem as Vid2VidMR. In this work, we present Moment Alignment", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20216", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c87769f42ef8917b0b561b87a735c4eb7a7050ab40d41f078aa87032b7b285da.jpg", + "image_caption": [ + "Figure 2. Moment Alignment Transformer (MATR) Architecture. We represent the target video as a query-aligned representation capturing the required correlation and dependencies with query video at the abstract level (output of encoder) and finer level (output of decoder). These features are used to extract foreground target frames, as well as localize the relevant moment in terms of start and end times via two prediction heads. The alignment is computed using representations both before and after the encoder (shown using blue dotted and red dashed lines, respectively). [Best viewed in color]." + ], + "image_footnote": [], + "bbox": [ + 98, + 87, + 903, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TRansformer (MATR) - a transformer-based architecture that leverages explicit alignment for precise moment localization. MATR transforms the target video into a query-aligned representation via a transformer framework to exploit finer-level correlation and dependencies for precise moment localization.", + "bbox": [ + 88, + 478, + 486, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation behind MATR Architecture", + "text_level": 1, + "bbox": [ + 89, + 577, + 403, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video moments vary significantly in duration, motion patterns, and visual appearance, making their retrieval inherently challenging. Additionally, preserving the correct temporal order of events is crucial for accurate localization, as moments often involve complex interactions that unfold over time. A robust Vid2VidMR model must, therefore, capture both high-level semantic relationships and fine-grained frame dependencies to ensure precise alignment between the query and target video.", + "bbox": [ + 88, + 598, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these challenges, MATR incorporates an explicit dual-stage sequence alignment strategy within a transformer-based encoder-decoder framework. This strategy enables MATR to learn a query-aligned representation of the target video by combining the abstract representation from the encoder with the refined fine-grained features from the decoder. In doing so, it facilitates precise moment localization by leveraging both global semantic alignment and fine-grained temporal dependencies between the query and target videos. Figure 2 illustrates the overall architecture of MATR, which we discuss in detail in the following.", + "bbox": [ + 88, + 734, + 486, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture Details", + "text_level": 1, + "bbox": [ + 511, + 478, + 692, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Representation. We uniformly sample frames every 2 seconds to obtain $M$ and $N$ frames from the target and query videos, respectively. We encode these frames using a frozen CLIP [39] (ViT-B/32) encoder along with a linear projection module. CLIP is applied separately to both videos to obtain target video embeddings $\\mathrm{E}_t \\in \\mathbb{R}^{M \\times d}$ and query video embeddings $\\mathrm{E}_q \\in \\mathbb{R}^{N \\times d}$ . The linear projection module is a two-layer perceptron, each with layer normalization and dropout. It maps 512-dimensional CLIP embeddings to $d$ -dimensional outputs. The resulting projected features $(\\mathrm{E}_t$ and $\\mathrm{E}_q)$ are concatenated along the sequence length dimension to form the input to the transformer encoder, denoted as $\\mathrm{E}_c = [\\mathrm{E}_t; \\mathrm{E}_q] \\in \\mathbb{R}^{(M + N) \\times d}$ .", + "bbox": [ + 511, + 502, + 906, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. The combined target and query video representation $(E_{c})$ is processed through a series of $k$ standard Transformer [48] encoder layers, each comprising a multi-head self-attention mechanism and a feed-forward network. Further, following the prior works [3, 5, 36], fixed positional encodings are added to the input of each attention layer to preserve temporal order. The encoder generates a fused representation $\\left[\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}};\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\right]$ of the target video conditioned on the query video. Here $\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}}\\in \\mathbb{R}^{M\\times d}$ and $\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\in \\mathbb{R}^{N\\times d}$ represent the target and query parts respectively.", + "bbox": [ + 511, + 700, + 908, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dual-stage Sequence Alignment. The transformer encoder computes effective features by performing a joint understanding (or fusion) of target and query video frames.", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20217", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The final goal of Vid2VidMR is to localize a moment in the target video that semantically matches the content of the query video. In other words, we would prefer the features of the moment in the target to align strongly with the features of the query video. Therefore, we perform dual-stage sequence alignment, i.e., before and after the encoder. For alignment, we choose soft-DTW [9]² and perform the alignment as follows.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-fusion alignment: Before encoder fusion, given the target and query video feature sequences, $\\mathrm{E_t} = [e_1^t,\\dots,e_M^t ]\\in \\mathbb{R}^{M\\times d}$ and $\\mathrm{E_q} = [e_1^q,\\dots,e_N^q ]\\in \\mathbb{R}^{N\\times d}$ , soft-DTW outputs a binary alignment matrix $\\mathrm{A}^{\\mathrm{pre}}$ and an alignment cost matrix $\\mathrm{C}^{\\mathrm{pre}}$ . $\\mathrm{A}^{\\mathrm{pre}}\\in \\{0,1\\}^{M\\times N}$ is an alignment matrix such that:", + "bbox": [ + 89, + 210, + 483, + 301 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {i, j} ^ {\\mathrm {p r e}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\boldsymbol {e} _ {i} ^ {t} \\text {i s m a t c h e d t o} \\boldsymbol {e} _ {j} ^ {q}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 308, + 406, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "further, soft-DTW finds the optimal alignment by minimizing the pre-fusion alignment loss $\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{pre}}$ , defined as:", + "bbox": [ + 89, + 354, + 482, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p r e}} = \\text {s o f t - D T W} _ {\\gamma} (\\mathrm {A} _ {i, j} ^ {\\text {p r e}}, \\mathrm {C} _ {i, j} ^ {\\text {p r e}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 397, + 398, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "such that aligned frames in the target are contiguous. Here $\\gamma$ is a smoothing factor for soft-min operator. The value of $\\gamma$ in Soft-DTW is selected empirically to balance smoothness and alignment fidelity. Note that $C^{\\mathrm{pre}}$ is the alignment cost matrix with elements defined by cosine similarity:", + "bbox": [ + 89, + 419, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C} _ {i, j} = 1 - \\frac {\\langle \\pmb {e} _ {i} ^ {t} , \\pmb {e} _ {j} ^ {q} \\rangle}{\\| \\pmb {e} _ {i} ^ {t} \\| \\| \\pmb {e} _ {j} ^ {q} \\|}.\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 501, + 364, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-fusion alignment enhances the semantic representation of target and query video. These enhanced representations are further processed by the encoder. The encoder outputs the fused representation of the target and query.", + "bbox": [ + 89, + 545, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-fusion alignment: Given the post-fusion target and query video feature sequences, $\\mathrm{E_t^g}\\in \\mathbb{R}^{M\\times d}$ and $\\mathrm{E_q^g}\\in$ $\\mathbb{R}^{N\\times d}$ , soft-DTW outputs a binary alignment matrix $\\mathrm{A}^{\\mathrm{post}}$ and an alignment cost matrix $\\mathrm{C}^{\\mathrm{post}}$ , soft-DTW achieves optimal alignment between $\\mathrm{E_t^g}$ and $\\mathrm{E_q^g}$ by minimizing the postfusion alignment loss $\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{post}}$ , defined as:", + "bbox": [ + 89, + 606, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p o s t}} = \\text {s o f t - D T W} _ {\\gamma} \\left(\\mathrm {A} _ {i, j} ^ {\\text {p o s t}}, \\mathrm {C} _ {i, j} ^ {\\text {p o s t}}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 715, + 401, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Post-fusion alignment refines target video features by leveraging fused query-target representations, ensuring fine-grained semantic matching for precise moment localization.", + "bbox": [ + 89, + 738, + 482, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder. The decoder in MATR further refines the target video representation by processing aligned query-target features, enabling fine-grained temporal matching and enhancing frame-level precision for accurate moment localization.", + "bbox": [ + 89, + 782, + 482, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, the post-fusion alignment matrix $\\mathbf{A}^{\\mathrm{post}}$ defines a contiguous sub-sequence $\\mathrm{E_t^g}[s : e]$ which aligns best with the query video. These aligned target features, capturing query-aligned fine-grained information, are passed as input to the decoder for further refinement. The decoder consists of $k$ Transformer layers, each containing a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward network. The input to the decoder comprises fixed size $l$ learnable query vectors denoted by $\\mathrm{Q} \\in [q_1, \\ldots, q_l] \\in \\mathbb{R}^{l \\times d}$ . These queries guide the extraction of refined features relevant to moment localization. As these queries are processed through each decoder layer, they are refined to capture intricate temporal dependencies, with positional encodings applied at each attention layer. The cross-attention layers enable interaction between $\\mathrm{E_t^g}[s : e]$ from the encoder (which serve as keys and values) and representations of moment queries (which serve as queries). The final decoder output, $\\mathrm{E_t^l} \\in \\mathbb{R}^{l \\times d}$ , provides a refined fine-grained set of target video features conditioned on the query video.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the encoder and decoder representations are combined as $\\mathrm{E_f} = [\\mathrm{E_t^g};\\mathrm{E_t^l}]$ , making the final target video representation query-aligned. The final query-aligned representation, combining the abstract and fine-grained semantics, serves as input for the prediction heads.", + "bbox": [ + 511, + 393, + 903, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction Heads. The prediction heads consist of three $(c = 3)$ sequential convolutional layers each having $d$ , $1 \\times 3$ kernels followed by a ReLU activation. Finally, a sigmoid activation function is applied to produce the foreground predictions $\\hat{f}_i$ . The model is trained using the following binary cross-entropy loss $\\mathcal{L}_{fg}$ to distinguish foreground (relevant moment) from background predictions:", + "bbox": [ + 511, + 468, + 905, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f g} = - \\left(f _ {i} \\log \\hat {f} _ {i} + (1 - f _ {i}) \\log (1 - \\hat {f} _ {i})\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 582, + 859, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{i}$ is the ground truth label, with $f_{i} = 1$ and $f_{i} = 0$ indicating foreground and background, respectively.", + "bbox": [ + 511, + 614, + 903, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to foreground classification, the boundary prediction head is designed to estimate the start and end boundaries for the moment at every position. This head shares the same initial structure as the classification head but differs in the final output, producing two channels for left and right boundary offsets relative to the current position. Specifically, given $\\mathrm{E_f}$ , the boundary prediction head outputs predicted offsets $\\hat{d}_i = \\{\\hat{d}_i^L, \\hat{d}_i^R\\}$ for each position $i \\in [0, M-1]$ , where $\\hat{d}_i^L$ and $\\hat{d}_i^R$ represent the left and right boundary offsets, respectively. The boundary prediction head is trained with a combination of smooth $L1$ and generalized intersection over union (IoU) loss [42], specifically applied to foreground positions $(f_i = 1)$ , defined as:", + "bbox": [ + 511, + 646, + 905, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {1} _ {f _ {i} = 1} \\left[ \\lambda_ {\\mathrm {L} 1} \\mathcal {L} _ {\\mathrm {L} 1} \\left(\\hat {d} _ {i}, d _ {i}\\right) + \\lambda_ {\\mathrm {I o U}} \\mathcal {L} _ {\\mathrm {I o U}} \\left(\\hat {d} _ {i}, d _ {i}\\right) \\right].\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 851, + 885, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $d_{i}$ denotes the ground truth offset. The parameters", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2Soft-DTW allows non-linear alignments between sequences of different lengths, is differentiable, robust to variations in speed, and can handle noise and outliers, though other alignment algorithms, such as TCC [12] and DropDTW [11] can also be used in our model.", + "bbox": [ + 89, + 849, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20218", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6c3e3ca473497383049fe5b203719752208f104a65d7982c83ffc5a1702f27bc.jpg", + "image_caption": [ + "Figure 3. Proposed self-supervised pre-training strategy: Given a target video, a query clip is randomly sampled from it and processed by the model. The model then predicts the boundaries of the selected clip within the target video, highlighting the corresponding frames with a green border. (Best viewed in color)" + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 483, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\lambda_{\\mathrm{L1}}$ and $\\lambda_{\\mathrm{IoU}}$ weigh the contributions of the smooth L1 and IoU losses, ensuring that boundary localization focuses on the predictions identified as foreground.", + "bbox": [ + 89, + 287, + 483, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Convolutional layers operate along the temporal axis over the concatenated encoder query-conditioned and decoder-aligned target features. This preserves temporal continuity, enabling fine-grained predictions without boundary artifacts.", + "bbox": [ + 89, + 334, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss. The overall loss function $\\mathcal{L}$ is computed as a combination of multiple objectives, accounting for alignment costs (pre- and post-fusion alignment), foreground/background classification, and boundary localization. For $S$ training samples, this multi-objective overall loss is defined as:", + "bbox": [ + 89, + 409, + 483, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{S} \\sum_ {i = 1} ^ {S} \\left(\\lambda_ {f g} \\mathcal {L} _ {f g} + \\lambda_ {\\mathrm {s e g}} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 500, + 460, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\lambda$ s control the importance of each component.", + "bbox": [ + 91, + 546, + 467, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. During inference, given a target video $V_{t}$ and a query video $V_{q}$ , we pass both through the model to generate foreground probabilities $\\{\\hat{f}_i\\}_{i=0}^{M-1}$ and boundary predictions $\\{\\hat{d}_i\\}_{i=0}^{M-1}$ using the two prediction heads. To handle the densely generated boundaries, while predicting $\\hat{d}_i$ , we apply 1-dimensional non-maximal suppression with a threshold of 0.7, filtering out highly overlapping boundaries and producing the final set of predictions. We choose $\\hat{d}_i$ corresponding to position $i$ with highest $\\hat{f}_i$ .", + "bbox": [ + 89, + 561, + 483, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Self-supervised Pre-training Strategy", + "text_level": 1, + "bbox": [ + 89, + 712, + 385, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The video representation learning community has leveraged self-supervised pre-training techniques designed for training effective encoders [22, 23, 37, 38, 45, 49, 50], which can be finetuned for enhanced performance on downstream tasks. In this work, we introduce a self-supervised pretraining objective designed to improve temporal localization capabilities without relying on labeled data. Specifically, given a target video $V_{t}$ , a query clip $V_{q}$ is randomly sampled from $V_{t}$ as shown in Fig. 3. The model is then trained to localize this query clip in $V_{t}$ . This pre-training objective closely aligns with the task of Vid2VidMR. To en", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "hance generalization capabilities and encourage the model to become robust to variations in both appearance and timing, we apply one of the following random augmentations to each of the query clips: reversing frames, adding gaussian noise, slowing down or speeding up the action, thereby doubling the number of pre-training samples. These augmentations introduce temporal and spatial variations, enabling the model to learn diverse representations.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall pre-training loss, $\\mathcal{L}_{pt}$ , combines the foreground-background classification loss, the boundary prediction loss, pre-fusion and post-fusion alignment costs. The total loss is averaged over $P$ pre-training samples as:", + "bbox": [ + 511, + 212, + 903, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p t} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\left(\\lambda_ {f g} ^ {p} \\mathcal {L} _ {\\mathrm {f g}} + \\lambda_ {s e g} ^ {p} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 281, + 895, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\lambda$ s control the importance of each component.", + "bbox": [ + 516, + 330, + 888, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Datasets", + "text_level": 1, + "bbox": [ + 511, + 361, + 609, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the following two datasets to compare our proposed approach with existing Vid2VidMR methods and other strong baselines:", + "bbox": [ + 511, + 388, + 903, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(i) ActivityNet-VRL Dataset [14]: ActivityNet-VRL is the popular benchmark dataset shared by Feng et al. [14], based on the ActivityNet video understanding benchmark [17] consisting of 200 action classes. The dataset is split into disjoint 160 classes for training and 20 classes each for the validation and test splits. Furthermore, the training set comprises of $\\sim 463\\mathrm{K}$ query-target video pairs, while the validation set and the test set have 829 and 978 query-target video pairs, respectively.", + "bbox": [ + 511, + 434, + 905, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(ii) Our proposed SportsMoments Dataset: One of the most promising applications for Vid2VidMR lies in the area of sports analytics. Although there exist large-scale sports datasets such as Sports-1M [21], they are primarily tailored for broader sports video classification tasks and do not contain fine-grained sports actions and moments, such as \"Cover Drive,\" \"Ducking a Bouncer,\" \"Goal Kick,\" or \"Penalty.\" Secondly, although ActivityNet-VRL covers a wide range of action categories (including sports actions), it only encompasses broader categories such as \"Playing Polo\" and does not focus on capturing higher-granularity events within a sport. Towards filling this gap, we introduce the SportsMoments dataset, which consists of $\\sim 770\\mathrm{K}$ query-target pairs annotated from 176.6 hours of complete match footage of two of the most popular sports, viz. soccer and cricket. We obtain a total of 80 full-length cricket and soccer full-length match videos from YouTube. We then curate a list of 29 action classes comprising 13 soccer and 16 cricket actions, respectively. Given this list, we employed two annotators, each with strong knowledge of these sports,", + "bbox": [ + 511, + 571, + 906, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3https://github.com/vl2g/MATR/tree/main/sportsmoments", + "bbox": [ + 529, + 886, + 839, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20219", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to mark the start and end timestamps for the specified actions in the videos. We split SportsMoments into training, validation and test sets. The training split consists of approximately 750K pairs spanning 16 classes, while the validation and test splits each contain 10K pairs, covering four and nine classes, respectively. Action classes are disjoint across the train, validation, and test sets, ensuring no overlap. Additionally, each split includes cricket and soccer actions for a well-rounded distribution.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we leverage the unlabeled videos from the Kinetics700 [6] dataset for pre-training.", + "bbox": [ + 89, + 227, + 480, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 273, + 323, + 290 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Baselines", + "text_level": 1, + "bbox": [ + 89, + 300, + 197, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experiment with an extensive set of strong baselines grouped into four categories as follows.", + "bbox": [ + 89, + 323, + 480, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(i) Fully-supervised Vid2VidMR methods: We compare with Vid2VidMR methods like CGBM [14], GDP [7], SRL [51], FFI+SRM [19], SST [4] and Video-level match [14]. Further, Huo et al. [19] adapt text-VMR methods like VSLNet [61], MABAN [46] and 2D-TAN [63] by replacing their text encoders with a C3D [47] feature extractor to extract query video features. All of these models have been trained on ActivityNet-VRL.", + "bbox": [ + 89, + 354, + 482, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(ii) Vision-Language Models (VLMs): VLMs like VideoLLaMa [62], Video-LLaVa [26] and TimeChat [41] have shown promising results for multimodal text-video tasks like Video VQA, captioning, etc. Unfortunately, they have not been pre-trained with video-video aligned data. First we represent the query video using a state-of-the-art captioner, i.e., mPLUG-OWL [58]. Next, the VLMs are zero-shot prompted to generate start and end moment timestamps given query caption and target video tokens.", + "bbox": [ + 89, + 476, + 482, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iii) Text-VMR methods: Moment-DETR [25], QD-DETR [33], CG-DETR [32], and UniVTG [28] have been originally proposed for text-VMR. We compare with five variants of these methods. The zero-shot variant (a) uses caption from mPLUG-OWL [58] to represent query video, and leverages the pretrained checkpoints. Variants (b) and (c) both use caption from mPLUG-OWL [58] to represent query video, but train a randomly initialized checkpoint (variant b) or finetune the pre-trained checkpoint (variant c), respectively. Variants (d) and (e) are equivalent to (b) and (c) where their CLIP text encoder is replaced by CLIP ViT/B-32 vision encoder, and therefore take query video directly as input along with the target video.", + "bbox": [ + 89, + 612, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(iv) Image-VMR methods: Our work focuses on video moment retrieval using 'video' queries. As a single image query may not effectively capture the temporal aspects of a video query, this comparison may not be appropriate. However, we still design baselines by representing the video by its key-frame (more precisely, the middle frame) using", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text-VMR methods, namely Moment-DETR, QD-DETR and UniVTG.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 132, + 730, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose hidden dimension of 1024 with $k = 4$ layers in both encoders and apply a dropout rate of 0.1 and 0.5 within the transformer and linear projection layers, respectively. Model weights are initialized using Xavier initialization [16]. To optimize the model parameters, we utilize AdamW optimizer [30] with an initial learning rate of 1e-4 and a weight decay of 1e-4. Training was done on two NVIDIA A6000 GPUs. We trained our model for 200 epochs with a batch size of 1200 while using ActivityNetVRL. For SportsMoments, we used 40 epochs with a batch size of 40. We set number of learnable queries $l = 10$ and all $\\lambda s$ in both pre-training as well as finetuning losses to 1. We make our implementation and checkpoints available at: https://github.com/vl2g/MATR.", + "bbox": [ + 511, + 155, + 906, + 367 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results and Discussion", + "text_level": 1, + "bbox": [ + 511, + 378, + 723, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results: Table 1 shows our main results where we compare MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and five variants of each of the Text-VMR methods. As mentioned before these variants correspond to zero-shot, finetuned or trained setup. As usual, trained setup implies random initialization while finetuned setup implies initialization using a pre-trained checkpoint (different for each architecture). Query video can be represented using text (T) captions obtained using mPLUGOWL [58] or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions and image query is not needed. Following previous work, we report the standard mean Intersection over Union (mIoU) and Recall@1 IoU=0.5 metrics.", + "bbox": [ + 511, + 400, + 906, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with fully-supervised Vid2VidMR methods: For a fair comparison, we directly use the reported results from original papers for fully-supervised methods on ActivityNet-VRL. Consequently, we do not report their results on SportsMoments. As shown in Table 1, MATR outperforms the best method, FFI+SRM, with an $8.1\\%$ gain in mIoU and achieves a $13.1\\%$ improvement in R@1 over GDP, which had the highest recall. These results underscore MATR's superior ability to accurately localize moments using video queries.", + "bbox": [ + 511, + 613, + 903, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with VLMs: Table 1 shows that our method MATR outperforms all four VLM-based baselines by a significant margin. In comparison, the best-performing baseline, TimeChat [41], achieves $26.4\\mathrm{mIoU}$ and $23.8\\mathrm{R}@\\mathbb{1}$ on ActivityNet-VRL, and $22.6\\mathrm{mIoU}$ and $21.3\\mathrm{R}@\\mathbb{1}$ on SportsMoments. Other methods, such as VideoLLaVA [26], Video-LLaMA [62], and Video-LLaMA2 [8], perform significantly worse. These results show that task-specific models for Vid2VidMR are significantly better.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20220", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7def681002fbc0e36db09c57ffd2fa3f932290fd358e4698645770a247fbe44f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Comparing MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and variants of Text-VMR methods. Trained implies random initialization. Finetuned implies initialization using a pre-trained (respective to each architecture) checkpoint. Query video can be represented using text (T) captions or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions is not needed. For more details of baselines and their variants, please refer to Section 5.1. Results for the overall best, best among implemented baselines and previously reported SOTA are highlighted in bold, underline and box, respectively." + ], + "table_body": "
MethodsVariantsActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Fully-supervised MethodsRandom [14]-7.316.2--
Video Match [14]-12.424.3--
SST [4]-17.133.2--
CGBM [14]-25.743.5--
GDP [7]-27.844.0--
SRL [51]-40.629.3--
2D-TAN [63]-45.339.6--
MABAN [46]-42.837.5--
VSLNet [61]-27.243.8--
FFI+SRM [19]-48.740.6--
VLMsVideo-LLaVA [26]-15.114.713.811.9
Video-LLaMA [62]-14.713.912.511.2
Video-LLaMA2 [8]-17.615.214.713.4
TimeChat [41]-26.423.822.621.3
I-VMRMoment-DETR [25]-35.632.525.218.5
QD-DETR [33]-37.238.127.720.7
UniVTG [28]-38.842.034.637.2
Text-VMR MethodsMoment-DETR [25](a) ZS+T28.222.83.70.9
(b) Trained+T37.135.831.429.2
(c) Finetuned+T40.038.935.834.0
(d) Trained+V35.934.728.724.2
(e) Finetuned+V40.039.930.425.1
QD-DETR [33](a) ZS+T25.422.26.32.6
(b) Trained+T38.439.029.827.2
(c) Finetuned+T41.645.236.033.5
(d) Trained+V39.741.027.225.1
(e) Finetuned+V42.542.735.430.6
UniVTG [28](a) ZS+T32.426.711.46.0
(b) Trained+T43.845.641.536.5
(c) Finetuned+T45.846.444.839.2
(d) Trained+V48.449.843.239.5
(e) Finetuned+V49.150.743.641.8
CG-DETR [32](a) ZS+T25.921.93.31.6
(b) Trained+T39.438.435.131.3
(c) Finetuned+T41.043.235.432.5
(d) Trained+V40.041.736.635.1
(e) Finetuned+V40.141.737.234.8
OursMATRZero-shot32.630.131.830.7
Trained+V53.254.856.252.7
Finetuned+V56.857.159.256.5
", + "bbox": [ + 89, + 87, + 495, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Text-VMR Methods: Table 1 shows that amongst all the five variants, the finetuned variants are typically better than the trained variants, i.e., variant (c) is better than (b), and (e) is better than (d). On both ActivityNet-VRL and SportsMoments, UniVTG achieves the best results among variants which use video queries. However, our MATR model outperforms UniVTG by $7.7\\%$ in mIoU and $6.4\\%$ in R@1 on ActivityNet-VRL. On SportsMoments, MATR shows a further improvement, sur", + "bbox": [ + 89, + 760, + 483, + 897 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee17677be81b50d3fab021483658d573663c1ccabb3be42a5f21c087aa03e790.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Pre-FusionPost-FusionActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
XX49.750.249.146.6
X✓52.352.953.148.2
✓X50.451.252.446.7
✓✓53.254.856.252.7
", + "bbox": [ + 540, + 88, + 880, + 164 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a8559d65e692b90aa43160292a74a9f1d4e9cbdee41a952c6454805a3c94a7b1.jpg", + "table_caption": [ + "Table 2. Advantage of explicit dual-stage alignment of MATR. For this study, experiments were performed without pre-training." + ], + "table_footnote": [], + "table_body": "
Pre-trainingAugmentationActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
X-53.254.856.252.7
✓X54.155.656.953.6
✓✓56.857.159.256.5
", + "bbox": [ + 532, + 209, + 888, + 281 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Effect of pre-training and augmentation in MATR.", + "bbox": [ + 532, + 290, + 885, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "passing UniVTG by $15.6\\%$ in mIoU and $14.7\\%$ in R@1. These gains highlight the superior performance of MATR compared to the text-VMR methods when used with video query input. When text captions are used to represent the query, we observe similar trends as in the video query input case. Among the baselines, UniVTG performs the best. However, MATR outperforms it by $11\\%$ in mIoU and $10.7\\%$ in R@1 on ActivityNet-VRL. On SportsMoments, MATR achieves $14.4\\%$ higher mIoU and $17.3\\%$ higher R@1 than UniVTG. These results demonstrate the superior performance of MATR compared to the text-VMR methods when used with text query input for Vid2VidMR.", + "bbox": [ + 511, + 314, + 906, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Image-VMR: Table 1 shows that the best-performing approach, UniVTG [28], achieves a mIoU of 38.8 and 42.0 R@1 on ActivityNet-VRL and 34.6 mIoU and R@1 37.2 on SportsMoments. Although this is comparable to previous fully supervised methods, it still lags significantly behind our MATR. This further goes to show the importance of temporal dynamics captured by video queries as compared to static image queries.", + "bbox": [ + 511, + 500, + 908, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Advantage of Dual Sequence Alignment: Table 2 shows the impact of different alignment strategies in MATR on both datasets. This study is done without any pre-training. We evaluate the effect of incorporating alignment at the pre-fusion and post-fusion stages. Our results indicate that the combined use of both pre-fusion and post-fusion alignment achieves the best performance, suggesting that alignment is important to ensure both before and after the encoder. Models with alignment either before (pre) or after (post) show moderate gains over neither. Removing post-fusion alignment hurts more than removing pre-fusion alignment, suggesting that post-fusion alignment is more important. Omitting alignment entirely results in the lowest scores. This highlights the importance of dual alignment for effectively capturing relevant temporal information for video query-based moment localization.", + "bbox": [ + 511, + 625, + 908, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Pre-training: We analyze the impact of pretraining and data augmentation on model performance in", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20221", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f34237d23fd96a08df298a0fd06983b207b140d2909476fa41ddd4e6c5e1cfa.jpg", + "image_caption": [ + "Figure 4. Visualization of Vid2VidMR on a sample from ActivityNet-VRL for calf roping action. Our proposed MATR model shows improved generalization capabilities over the best-performing baseline methods (variant (e), i.e., finetuned+V). Start and end times for the ground truth and predicted moments are shown in the brackets. Refer to supplementary material for the videos used in this figure." + ], + "image_footnote": [], + "bbox": [ + 125, + 90, + 359, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a461c2ccd27c007ebe387383b266007d7fdb85eb2701c8b2c6968741cbd4da5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 870, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0031de5ad333789d20a845489a9d2209e1647a27e799776293004eba2794a27.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Pre-fusion alignment (Apre)34.132.631.629.8
Post-fusion alignment (Apost)38.335.236.433.5
Prediction-heads56.857.159.256.5
", + "bbox": [ + 112, + 309, + 460, + 383 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b320b5ebc8b61ee559fdc7318bf7e56ee447bc5917d4862b8cab22154bc7160d.jpg", + "table_caption": [ + "Table 4. Advantage of predicting using prediction heads on decoder. We observe that directly predicting from fusion matrices is inferior as it lacks fine-grained refinement." + ], + "table_footnote": [], + "table_body": "
AlignmentActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
None50.451.852.148.6
TCC52.452.753.952.8
Drop-DTW54.254.755.654.7
soft-DTW56.857.159.256.5
", + "bbox": [ + 153, + 443, + 421, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation on alignment methods.", + "bbox": [ + 161, + 535, + 410, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. On ActivityNet-VRL, without pre-training, the model achieves $53.2\\mathrm{mIoU}$ and $54.8\\mathrm{R}@\\mathrm{l}$ . Adding pretraining without augmentation yields $54.1\\mathrm{mIoU}$ and $55.6\\mathrm{R}@\\mathrm{l}$ . The best results are obtained by using pre-training with augmentation, leading to a significant boost to $56.8\\mathrm{mIoU}$ and $57.1\\mathrm{R}@\\mathrm{l}$ . We got similar observations for the SportsMoments dataset.", + "bbox": [ + 89, + 559, + 483, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Predictions from Heads vs Alignment Matrices: The results in Table 4 show that prediction from heads outperforms the prediction from both pre-fusion alignment $(\\mathrm{A}^{\\mathrm{pre}})$ and post-fusion alignment $(\\mathrm{A}^{\\mathrm{post}})$ on both datasets. Notably, prediction heads improve mIoU and R@1 by a large margin, demonstrating their effectiveness in accurately retrieving temporal moments. Post-fusion alignment provides better results than pre-fusion alignment. However, the gains are modest compared to the substantial boost from prediction heads. This justifies the need for the decoder.", + "bbox": [ + 89, + 669, + 482, + 820 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Alignment Losses: MATR can incorporate different alignment methods. Table 5 compares the performance of different alignment methods on both ActivityNet-VRL and SportsMoments datasets. Without any alignment (\"None\"), the model achieves R@1 scores of $51.8\\%$", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on ActivityNet-VRL and $48.6\\%$ on SportsMoments. TCC alignment yields an improvement of $0.9\\%$ and $4.2\\%$ on ActivityNet-VRL and SportsMoments, respectively. Drop-DTW is better than TCC. soft-DTW alignment consistently demonstrates superior performance across both datasets, achieving the highest gains of $5.3\\%$ on ActivityNet-VRL and $7.9\\%$ on SportsMoments.", + "bbox": [ + 511, + 311, + 906, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results: We present moment retrieval results on a sample from ActivityNet-VRL in Fig. 4, where we compare our proposed approach to variant (e) (refer Sec. 5.1 for details) of the four best-performing baselines. The video query is a 5-second clip depicting calf roping4. The target video, with a duration of 24 seconds, showcases this event occurring between 2.1 and 18.1 seconds. Our proposed MATR model exhibits the highest overlap with the ground truth for the given query, accurately identifying the correct temporal boundaries. Among the baseline methods, CG-DETR and UniVTG achieve strong overlap, however, are still outperformed by our approach.", + "bbox": [ + 511, + 419, + 908, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced MATR, a robust approach for Vid2VidMR that combines abstract representation from encoder with fine-grained features from decoder conditioned on aligned query features. It captures semantic and temporal cues for precise moment localization. Our self-supervised pretraining enhances initialization and boosts performance. Extensive experiments on ActivityNet-VRL and our new SportsMoments dataset show that MATR outperforms strong baselines. Future directions of this work include exploring multimodal queries and developing scalable architectures to enhance broader applicability.", + "bbox": [ + 511, + 640, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment: This work is supported by the Microsoft Academic Partnership Grant (MAPG) 2023. Yogesh Kumar is supported by a UGC fellowship, Govt. of India.", + "bbox": [ + 511, + 806, + 908, + 853 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4A rodeo event where a rider on horseback attempts to catch and tie a calf within a timed competition.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20222", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] U. Agarwal, Y. Kumar, A. Shahid, P. Gatti, M. Gupta, and A. Mishra. CHAPVIDMR: Chapter-based video moment retrieval using natural language queries. In ICVGIP, 2024. 2", + "[2] D. M. Argaw, J.-Y. Lee, M. Woodson, I.-S. Kweon, and F. C. Heilbron. Long-range multimodal pretraining for movie understanding. ICCV, 2023. 2", + "[3] I. Bello, B. Zoph, Q. Le, A. Vaswani, and J. Schlens. Attention augmented convolutional networks. In ICCV, 2019. 3", + "[4] S. Buch, V. Escorcia, C. Shen, B. Ghanem, and J. C. Niebles. SST: single-stream temporal action proposals. In CVPR, 2017. 6, 7", + "[5] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3", + "[6] J. Carreira and A. Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 2, 6", + "[7] L. Chen, C. Lu, S. Tang, J. Xiao, D. Zhang, C. Tan, and X. Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2, 6, 7", + "[8] Z. Cheng, S. Leng, H. Zhang, Y. Xin, X. Li, G. Chen, Y. Zhu, W. Zhang, Z. Luo, D. Zhao, and L. Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6, 7", + "[9] M. Cuturi and M. Blondel. Soft-DTW: a differentiable loss function for time-series. In ICML, 2017. 2, 4", + "[10] P. Dogan, B. Li, L. Sigal, and M. Gross. A neural multi-sequence alignment technique (neumatch). In CVPR, 2018. 2", + "[11] N. Dvornik, I. Hadji, K. G. Derpanis, A. Garg, and A. D. Jepson. Drop-DTW: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 2, 4", + "[12] D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. Temporal cycle-consistency learning. In CVPR, 2019. 2, 4", + "[13] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slow-fast networks for video recognition. In ICCV, 2019. 2", + "[14] Y. Feng, L. Ma, W. Liu, T. Zhang, and J. Luo. Video re-localization. In ECCV, 2018. 2, 5, 6, 7", + "[15] J. Gao, C. Sun, Z. Yang, and R. Nevatia. TALL: temporal activity localization via language query. In ICCV, 2017. 2", + "[16] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 6" + ], + "bbox": [ + 93, + 114, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 5", + "[18] L. A. Hendricks, O. Wang, E. Shechtman, J. Sivic, T. Darrell, and B. C. Russell. Localizing moments in video with natural language. In ICCV, 2017. 2", + "[19] S. Huo, Y. Zhou, R. Wang, W. Xiang, and S. Kung. Semantic relevance learning for video-query based video moment retrieval. IEEE Trans. Multim., 25:9290-9301, 2023. 2, 6, 7", + "[20] M. Jung, Y. Jang, S. Choi, J. Kim, J.-H. Kim, and B.-T. Zhang. Background-aware moment detection for video moment retrieval. In WACV, 2025. 2", + "[21] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 5", + "[22] Y. Kumar, S. Mallick, A. Mishra, S. Rasipuram, A. Maitra, and R. Ramnani. Qdetrv: Query-guided detr for one-shot object localization in videos. In AAAI, 2024. 5", + "[23] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In ICCV, 2017. 5", + "[24] J. Lei, L. Yu, T. L. Berg, and M. Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 2", + "[25] J. Lei, T. L. Berg, and M. Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 1, 2, 6, 7", + "[26] B. Lin, B. Zhu, Y. Ye, M. Ning, P. Jin, and L. Yuan. Video-llava: Learning united visual representation by alignment before projection. In EMNLP, 2024. 6, 7", + "[27] J. Lin, C. Gan, and S. Han. TSM: temporal shift module for efficient video understanding. In ICCV, 2019. 2", + "[28] K. Q. Lin, P. Zhang, J. Chen, S. Pramanick, D. Gao, A. J. Wang, R. Yan, and M. Z. Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, 2023. 1, 2, 6, 7", + "[29] Y. Liu, S. Li, Y. Wu, C. W. Chen, Y. Shan, and X. Qie. UMT: unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2", + "[30] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 6", + "[31] N. C. Mithun, S. Paul, and A. K. Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2", + "[32] W. Moon, S. Hyun, S. B. Lee, and J. Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, abs/2311.08835, 2023. 1, 2, 6, 7" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] W. Moon, S. Hyun, S. Park, D. Park, and J. Heo. Query - dependent video representation for moment retrieval and highlight detection. In CVPR, 2023. 1, 2, 6, 7", + "[34] J. Mun, M. Cho, , and B. Han. Local-Global Video-Text Interactions for Temporal Grounding. In CVPR, 2020. 2", + "[35] F. Padua, R. Carceroni, G. Santos, and K. Kutulakos. Linear sequence-to-sequence alignment. IEEE Trans. Pattern Anal. Mach. Intell., 32:304-320, 2008. 2", + "[36] N. Parmar, A. Vaswani, J. Uszkoreit, L. Kaiser, N. Shazeer, A. Ku, and D. Tran. Image transformer. In ICML, 2018. 3", + "[37] A. Piergiovanni, A. Angelova, and M. S. Ryoo. Evolving losses for unsupervised video representation learning. In CVPR, 2020. 5", + "[38] R. Qian, T. Meng, B. Gong, M.-H. Yang, H. Wang, S. Belongie, and Y. Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 5", + "[39] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[40] Rao, Gritai, and Shah. View-invariant alignment and matching of video sequences. In ICCV, 2003. 2", + "[41] S. Ren, L. Yao, S. Li, X. Sun, and L. Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, 2024. 6, 7", + "[42] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4", + "[43] Z. Shou, D. Wang, and S.-F. Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In CVPR, 2016. 2", + "[44] K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos. In NuerIPS, 2014. 2", + "[45] N. Srivastava, E. Mansimov, and R. Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 5", + "[46] X. Sun, H. Wang, and B. He. Maban: Multi-agent boundary-aware network for natural language moment retrieval. IEEE Trans. on Image Processing, 30:5589-5599, 2021. 6, 7", + "[47] D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 2, 6", + "[48] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + ], + "bbox": [ + 89, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] G. Wang, Y. Zhou, C. Luo, W. Xie, W. Zeng, and Z. Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, 2021. 5", + "[50] J. Wang, Y. Gao, K. Li, J. Hu, X. Jiang, X. Guo, R. Ji, and X. Sun. Enhancing unsupervised video representation learning by decoupling the scene and the motion. In AAAI, 2021. 5", + "[51] R. Wang and Y. Zhou. A feature pair fusion and hierarchical learning framework for video re-localization. In ICIP, 2020. 2, 6, 7", + "[52] X. Wang, L. Zhu, and Y. Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, 2021. 2", + "[53] Y. Xiao, Z. Luo, Y. Liu, Y. Ma, H. Bian, Y. Ji, Y. Yang, and X. Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, 2024. 2", + "[54] S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018. 2", + "[55] S. Yan, X. Xiong, A. Nagrani, A. Arnab, Z. Wang, W. Ge, D. Ross, and C. Schmid. Unloc: A unified framework for video localization tasks. In ICCV, 2023. 2", + "[56] C. Yang, Y. Xu, J. Shi, B. Dai, and B. Zhou. Temporal pyramid network for action recognition. In CVPR, 2020. 2", + "[57] J. Yang, P. Wei, H. Li, and Z. Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, 2024. 1, 2", + "[58] Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou. mplug-owi2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6", + "[59] R. Zeng, W. Huang, M. Tan, Y. Rong, P. Zhao, J. Huang, and C. Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2", + "[60] C. Zhang, A. Gupta, and A. Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2", + "[61] H. Zhang, A. Sun, W. Jing, L. Zhen, J. T. Zhou, and R. S. M. Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 6, 7", + "[62] H. Zhang, X. Li, and L. Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 6, 7", + "[63] S. Zhang, H. Peng, J. Fu, and J. Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 6, 7" + ], + "bbox": [ + 514, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20224", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[64] Z. Zhu, W. Tang, L. Wang, N. Zheng, and G. Hua. Enriching local and global contexts for temporal action localization. In ICCV, 2021. 2", + "bbox": [ + 91, + 90, + 480, + 135 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20225", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_model.json b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc9d2b77591d8cf43265660d29f972991c0e8ea0 --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.259, + 0.131, + 0.74, + 0.154 + ], + "angle": 0, + "content": "Aligning Moments in Time using Video Queries" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.18, + 0.785, + 0.218 + ], + "angle": 0, + "content": "Yogesh Kumar1* Uday Agarwal1* Manish Gupta2 Anand Mishra1 \n1Indian Institute of Technology Jodhpur 2Microsoft" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.219, + 0.788, + 0.235 + ], + "angle": 0, + "content": "{kumar.204, agarwaluday, mishra}@iitj.ac.in, gmanish@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.725 + ], + "angle": 0, + "content": "Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment Transformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of \\(13.1\\%\\) in \\(R@1\\) and \\(8.1\\%\\) in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a \\(14.7\\%\\) gain in \\(R@1\\) and a \\(14.4\\%\\) gain in mIoU on an absolute scale over strong baselines. We make the dataset and code public at: https://github.com/vl2g/MATR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.753, + 0.222, + 0.768 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Video moment retrieval is the task of temporally localizing the start and end times of a moment in a target video described by a given query. Although text-based video moment retrieval has been extensively explored [25, 28, 32, 33, 57], it often poses challenges for users attempting to" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.269, + 0.906, + 0.303 + ], + "angle": 0, + "content": "Text Query \"A man positions himself beneath the ball, leaping into the air, bending his knees and arching his back. As he flips backward, he connects with the ball using his foot.\"" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.304, + 0.904, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 1. (Top): Activity where text falls short to explain complex action or events, such as bicycle kick, necessitating the need for a more intuitive query modality for localizing semantically matching moments. (Bottom): Shows the Vid2Vid moment retrieval setting, which is the goal of this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.851 + ], + "angle": 0, + "content": "describe specific moments verbally. For example, consider a spectacular bicycle kick in soccer (Fig. 1 top). Although an expert can search by naming or describing it in detail, a beginner in soccer may struggle to describe this impressive move accurately. They might say something like \"The player kicked the ball while in the air\", which lacks the nuance needed to convey the athleticism and artistry involved. Further, such a description may lead to poor results for someone searching for similar moments. In contrast, if a soccer trainee shows a short video clip of a bicycle kick, the input is rich and crisp, making it easier for a retriever to locate similar moments in a target soccer match video. This approach aligns with how users naturally would prefer to search content within the same (video) modality, making it a more effective method for retrieving specific moments. Fig. 1 (bottom) shows a video bicycle kick query in soccer that is localized in a target video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Therefore, in this work, we study video-to-video moment retrieval (Vid2VidMR) where the aim is to temporally localize a moment in the target video with a high se" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.863, + 0.221, + 0.876 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1A moment is a continuous segment of frames within a target video that best represents the actions, events, or interactions described by the query." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20215" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "matic match between the moment and the query video. Vid2VidMR, a task formally introduced by Feng et al. [14], has many potential application areas, such as sports video analytics, educational content creation and e-learning, and surveillance systems. This task is challenging and requires semantic frame-level alignment and modeling complex dependencies between query and target videos. The need for a semantic understanding of video content and variety in video length, context, and action speed calls for adaptive models capable of generalizing across diverse scenarios. Addressing these challenges demands advanced temporal video modeling techniques." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.593 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose MATR (Moment Alignment TRansformer) - a method that uses explicit 'dual-stage sequence alignment' to capture the required correlation and temporal details essential for accurate moment localization. By conditioning target video representations on query features, MATR produces query-aligned representations that encode the necessary correlations and temporal dependencies between the two videos. We use differentiable dynamic time warping loss [9] for aligning the query and target videos, and represent the target video by conditioning it on the query video to focus on correlated temporal features. These representations guide a classification head to discriminate relevant moments from the irrelevant background and a boundary prediction head to mark the start and end of the identified moment in the target video. Further, to enhance our model's generalization using unlabeled videos, we introduce a self-supervised pretraining strategy which involves training MATR to localize randomly sampled clips within the same video, enabling it to learn the moment localization skill in a self-supervised manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.483, + 0.761 + ], + "angle": 0, + "content": "We evaluate MATR on public ActivityNet-VRL [14] benchmark and on our newly introduced dataset on sports domain, viz. SportsMoments, covering two of the most popular sports, namely soccer and cricket. Our approach achieves significant performance gains, with an improvement of \\(13.1\\%\\) and \\(8.1\\%\\) in R@1 and mIoU, respectively, on ActivityNet-VRL outperforming the state-of-the-art methods. Furthermore, MATR outperforms the implemented strong baselines on our proposed dataset with \\(14.7\\%\\) gains in R@1 and \\(14.4\\%\\) gains in mIoU, all on an absolute scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our contributions are as follows: (i) We introduce MATR, which uses explicit dual-stage sequence alignment within a transformer framework between target and query video to capture temporal correlations and dependencies for accurate moment localization. (ii) We propose a self-supervised pre-training objective that enhances model initialization by understanding rich video structure without requiring any labeled data. (iii) We conduct extensive experiments and ablations to study the efficacy of our" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "framework against competitive baselines and state-of-the-art methods. Our findings offer valuable insights into our design choices, and our approach advances the state-of-the-art on Vid2VidMR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.168, + 0.655, + 0.183 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.193, + 0.905, + 0.466 + ], + "angle": 0, + "content": "Video Moment Retrieval (VMR): VMR has recently gained significant interest in the research community [1, 15, 18, 24, 25, 28, 29, 32-34, 53, 55, 57]. Unlike video action understanding tasks such as action classification [6, 13, 27, 44, 47, 54, 56, 60] or temporal action localization [43, 59, 64], VMR focuses on identifying segments that semantically align with a broader range of queries, which may describe complex and context-specific moments. Based on the query modality, existing VMR methods can be broadly grouped into (i) textual query-based approaches like Moment-DETR [25], QD-DETR [33], UniVTG [28], and (ii) video query-based approaches like GDP [7], FFI+SRM [19], SRL [51]. Our work falls under video query-based VMR (Vid2VidMR). However, in addition to developing a new approach tailored for video query, we also adapt several text-query-based methods to make them suitable for Vid2VidMR and compare our method against both text and video query-based methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.905, + 0.678 + ], + "angle": 0, + "content": "**Alignment in Videos:** Alignment has been a thrust area in the video understanding community. It has been applied to a wide range of video understanding tasks, including video retrieval [52], procedural steps alignment [11], action recognition [10], anomaly detection [12], movie understanding [2] and video synchronization [35, 40]. Researchers have also explored sequence alignment for the text-VMR task, e.g., Mithun et al. [31] used sequence alignment between CNN representation of frames and GRU representation of text query to perform text-VMR. Jung et al. [20] used alignment to enhance semantic understanding between query and target video at the abstract level for the text-VMR task. However, their alignment is not at the sequence level." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.679, + 0.905, + 0.785 + ], + "angle": 0, + "content": "Compared to the existing literature, MATR goes beyond traditional sequence alignment by introducing a dual-stage alignment mechanism, leveraging transformer-based feature fusion, and maintaining flexibility with multiple alignment strategies. These innovations enable it to achieve more accurate and context-aware video moment retrieval compared to existing approaches." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.692, + 0.815 + ], + "angle": 0, + "content": "3. The MATR Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our objective is to temporally localize a moment in a target video \\( V_{t} \\) using a query video \\( V_{q} \\). We represent the target and query videos as sequences of \\( M \\) and \\( N \\) uniformly sampled frames, respectively. We refer to this problem as Vid2VidMR. In this work, we present Moment Alignment" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20216" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.088, + 0.905, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.399, + 0.908, + 0.471 + ], + "angle": 0, + "content": "Figure 2. Moment Alignment Transformer (MATR) Architecture. We represent the target video as a query-aligned representation capturing the required correlation and dependencies with query video at the abstract level (output of encoder) and finer level (output of decoder). These features are used to extract foreground target frames, as well as localize the relevant moment in terms of start and end times via two prediction heads. The alignment is computed using representations both before and after the encoder (shown using blue dotted and red dashed lines, respectively). [Best viewed in color]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.479, + 0.487, + 0.571 + ], + "angle": 0, + "content": "TRansformer (MATR) - a transformer-based architecture that leverages explicit alignment for precise moment localization. MATR transforms the target video into a query-aligned representation via a transformer framework to exploit finer-level correlation and dependencies for precise moment localization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.405, + 0.593 + ], + "angle": 0, + "content": "3.1. Motivation behind MATR Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Video moments vary significantly in duration, motion patterns, and visual appearance, making their retrieval inherently challenging. Additionally, preserving the correct temporal order of events is crucial for accurate localization, as moments often involve complex interactions that unfold over time. A robust Vid2VidMR model must, therefore, capture both high-level semantic relationships and fine-grained frame dependencies to ensure precise alignment between the query and target video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.487, + 0.903 + ], + "angle": 0, + "content": "To address these challenges, MATR incorporates an explicit dual-stage sequence alignment strategy within a transformer-based encoder-decoder framework. This strategy enables MATR to learn a query-aligned representation of the target video by combining the abstract representation from the encoder with the refined fine-grained features from the decoder. In doing so, it facilitates precise moment localization by leveraging both global semantic alignment and fine-grained temporal dependencies between the query and target videos. Figure 2 illustrates the overall architecture of MATR, which we discuss in detail in the following." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.693, + 0.493 + ], + "angle": 0, + "content": "3.2. Architecture Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.907, + 0.699 + ], + "angle": 0, + "content": "Input Representation. We uniformly sample frames every 2 seconds to obtain \\(M\\) and \\(N\\) frames from the target and query videos, respectively. We encode these frames using a frozen CLIP [39] (ViT-B/32) encoder along with a linear projection module. CLIP is applied separately to both videos to obtain target video embeddings \\(\\mathrm{E}_t \\in \\mathbb{R}^{M \\times d}\\) and query video embeddings \\(\\mathrm{E}_q \\in \\mathbb{R}^{N \\times d}\\). The linear projection module is a two-layer perceptron, each with layer normalization and dropout. It maps 512-dimensional CLIP embeddings to \\(d\\)-dimensional outputs. The resulting projected features \\((\\mathrm{E}_t\\) and \\(\\mathrm{E}_q)\\) are concatenated along the sequence length dimension to form the input to the transformer encoder, denoted as \\(\\mathrm{E}_c = [\\mathrm{E}_t; \\mathrm{E}_q] \\in \\mathbb{R}^{(M + N) \\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Encoder. The combined target and query video representation \\((E_{c})\\) is processed through a series of \\(k\\) standard Transformer [48] encoder layers, each comprising a multi-head self-attention mechanism and a feed-forward network. Further, following the prior works [3, 5, 36], fixed positional encodings are added to the input of each attention layer to preserve temporal order. The encoder generates a fused representation \\(\\left[\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}};\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\right]\\) of the target video conditioned on the query video. Here \\(\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}}\\in \\mathbb{R}^{M\\times d}\\) and \\(\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\in \\mathbb{R}^{N\\times d}\\) represent the target and query parts respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dual-stage Sequence Alignment. The transformer encoder computes effective features by performing a joint understanding (or fusion) of target and query video frames." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20217" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.212 + ], + "angle": 0, + "content": "The final goal of Vid2VidMR is to localize a moment in the target video that semantically matches the content of the query video. In other words, we would prefer the features of the moment in the target to align strongly with the features of the query video. Therefore, we perform dual-stage sequence alignment, i.e., before and after the encoder. For alignment, we choose soft-DTW [9]² and perform the alignment as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.302 + ], + "angle": 0, + "content": "Pre-fusion alignment: Before encoder fusion, given the target and query video feature sequences, \\(\\mathrm{E_t} = [e_1^t,\\dots,e_M^t ]\\in \\mathbb{R}^{M\\times d}\\) and \\(\\mathrm{E_q} = [e_1^q,\\dots,e_N^q ]\\in \\mathbb{R}^{N\\times d}\\), soft-DTW outputs a binary alignment matrix \\(\\mathrm{A}^{\\mathrm{pre}}\\) and an alignment cost matrix \\(\\mathrm{C}^{\\mathrm{pre}}\\). \\(\\mathrm{A}^{\\mathrm{pre}}\\in \\{0,1\\}^{M\\times N}\\) is an alignment matrix such that:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.309, + 0.407, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {i, j} ^ {\\mathrm {p r e}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\boldsymbol {e} _ {i} ^ {t} \\text {i s m a t c h e d t o} \\boldsymbol {e} _ {j} ^ {q}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.389 + ], + "angle": 0, + "content": "further, soft-DTW finds the optimal alignment by minimizing the pre-fusion alignment loss \\(\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{pre}}\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.398, + 0.399, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p r e}} = \\text {s o f t - D T W} _ {\\gamma} (\\mathrm {A} _ {i, j} ^ {\\text {p r e}}, \\mathrm {C} _ {i, j} ^ {\\text {p r e}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.484, + 0.496 + ], + "angle": 0, + "content": "such that aligned frames in the target are contiguous. Here \\(\\gamma\\) is a smoothing factor for soft-min operator. The value of \\(\\gamma\\) in Soft-DTW is selected empirically to balance smoothness and alignment fidelity. Note that \\(C^{\\mathrm{pre}}\\) is the alignment cost matrix with elements defined by cosine similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.502, + 0.366, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C} _ {i, j} = 1 - \\frac {\\langle \\pmb {e} _ {i} ^ {t} , \\pmb {e} _ {j} ^ {q} \\rangle}{\\| \\pmb {e} _ {i} ^ {t} \\| \\| \\pmb {e} _ {j} ^ {q} \\|}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Pre-fusion alignment enhances the semantic representation of target and query video. These enhanced representations are further processed by the encoder. The encoder outputs the fused representation of the target and query." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.706 + ], + "angle": 0, + "content": "Post-fusion alignment: Given the post-fusion target and query video feature sequences, \\(\\mathrm{E_t^g}\\in \\mathbb{R}^{M\\times d}\\) and \\(\\mathrm{E_q^g}\\in\\) \\(\\mathbb{R}^{N\\times d}\\), soft-DTW outputs a binary alignment matrix \\(\\mathrm{A}^{\\mathrm{post}}\\) and an alignment cost matrix \\(\\mathrm{C}^{\\mathrm{post}}\\), soft-DTW achieves optimal alignment between \\(\\mathrm{E_t^g}\\) and \\(\\mathrm{E_q^g}\\) by minimizing the postfusion alignment loss \\(\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{post}}\\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.716, + 0.402, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p o s t}} = \\text {s o f t - D T W} _ {\\gamma} \\left(\\mathrm {A} _ {i, j} ^ {\\text {p o s t}}, \\mathrm {C} _ {i, j} ^ {\\text {p o s t}}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Post-fusion alignment refines target video features by leveraging fused query-target representations, ensuring fine-grained semantic matching for precise moment localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.483, + 0.845 + ], + "angle": 0, + "content": "Decoder. The decoder in MATR further refines the target video representation by processing aligned query-target features, enabling fine-grained temporal matching and enhancing frame-level precision for accurate moment localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.394 + ], + "angle": 0, + "content": "To this end, the post-fusion alignment matrix \\(\\mathbf{A}^{\\mathrm{post}}\\) defines a contiguous sub-sequence \\(\\mathrm{E_t^g}[s : e]\\) which aligns best with the query video. These aligned target features, capturing query-aligned fine-grained information, are passed as input to the decoder for further refinement. The decoder consists of \\(k\\) Transformer layers, each containing a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward network. The input to the decoder comprises fixed size \\(l\\) learnable query vectors denoted by \\(\\mathrm{Q} \\in [q_1, \\ldots, q_l] \\in \\mathbb{R}^{l \\times d}\\). These queries guide the extraction of refined features relevant to moment localization. As these queries are processed through each decoder layer, they are refined to capture intricate temporal dependencies, with positional encodings applied at each attention layer. The cross-attention layers enable interaction between \\(\\mathrm{E_t^g}[s : e]\\) from the encoder (which serve as keys and values) and representations of moment queries (which serve as queries). The final decoder output, \\(\\mathrm{E_t^l} \\in \\mathbb{R}^{l \\times d}\\), provides a refined fine-grained set of target video features conditioned on the query video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.469 + ], + "angle": 0, + "content": "Finally, the encoder and decoder representations are combined as \\(\\mathrm{E_f} = [\\mathrm{E_t^g};\\mathrm{E_t^l}]\\), making the final target video representation query-aligned. The final query-aligned representation, combining the abstract and fine-grained semantics, serves as input for the prediction heads." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.469, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Prediction Heads. The prediction heads consist of three \\((c = 3)\\) sequential convolutional layers each having \\(d\\), \\(1 \\times 3\\) kernels followed by a ReLU activation. Finally, a sigmoid activation function is applied to produce the foreground predictions \\(\\hat{f}_i\\). The model is trained using the following binary cross-entropy loss \\(\\mathcal{L}_{fg}\\) to distinguish foreground (relevant moment) from background predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.583, + 0.86, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f g} = - \\left(f _ {i} \\log \\hat {f} _ {i} + (1 - f _ {i}) \\log (1 - \\hat {f} _ {i})\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.905, + 0.647 + ], + "angle": 0, + "content": "where \\( f_{i} \\) is the ground truth label, with \\( f_{i} = 1 \\) and \\( f_{i} = 0 \\) indicating foreground and background, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.647, + 0.906, + 0.845 + ], + "angle": 0, + "content": "In addition to foreground classification, the boundary prediction head is designed to estimate the start and end boundaries for the moment at every position. This head shares the same initial structure as the classification head but differs in the final output, producing two channels for left and right boundary offsets relative to the current position. Specifically, given \\( \\mathrm{E_f} \\), the boundary prediction head outputs predicted offsets \\( \\hat{d}_i = \\{\\hat{d}_i^L, \\hat{d}_i^R\\} \\) for each position \\( i \\in [0, M-1] \\), where \\( \\hat{d}_i^L \\) and \\( \\hat{d}_i^R \\) represent the left and right boundary offsets, respectively. The boundary prediction head is trained with a combination of smooth \\( L1 \\) and generalized intersection over union (IoU) loss [42], specifically applied to foreground positions \\( (f_i = 1) \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.852, + 0.887, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {1} _ {f _ {i} = 1} \\left[ \\lambda_ {\\mathrm {L} 1} \\mathcal {L} _ {\\mathrm {L} 1} \\left(\\hat {d} _ {i}, d _ {i}\\right) + \\lambda_ {\\mathrm {I o U}} \\mathcal {L} _ {\\mathrm {I o U}} \\left(\\hat {d} _ {i}, d _ {i}\\right) \\right].\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Here, \\( d_{i} \\) denotes the ground truth offset. The parameters" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.85, + 0.484, + 0.901 + ], + "angle": 0, + "content": "2Soft-DTW allows non-linear alignments between sequences of different lengths, is differentiable, robust to variations in speed, and can handle noise and outliers, though other alignment algorithms, such as TCC [12] and DropDTW [11] can also be used in our model." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.485, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.209, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Proposed self-supervised pre-training strategy: Given a target video, a query clip is randomly sampled from it and processed by the model. The model then predicts the boundaries of the selected clip within the target video, highlighting the corresponding frames with a green border. (Best viewed in color)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.334 + ], + "angle": 0, + "content": "\\(\\lambda_{\\mathrm{L1}}\\) and \\(\\lambda_{\\mathrm{IoU}}\\) weigh the contributions of the smooth L1 and IoU losses, ensuring that boundary localization focuses on the predictions identified as foreground." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.409 + ], + "angle": 0, + "content": "Convolutional layers operate along the temporal axis over the concatenated encoder query-conditioned and decoder-aligned target features. This preserves temporal continuity, enabling fine-grained predictions without boundary artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.493 + ], + "angle": 0, + "content": "Overall Loss. The overall loss function \\(\\mathcal{L}\\) is computed as a combination of multiple objectives, accounting for alignment costs (pre- and post-fusion alignment), foreground/background classification, and boundary localization. For \\(S\\) training samples, this multi-objective overall loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.501, + 0.462, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{S} \\sum_ {i = 1} ^ {S} \\left(\\lambda_ {f g} \\mathcal {L} _ {f g} + \\lambda_ {\\mathrm {s e g}} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.547, + 0.468, + 0.562 + ], + "angle": 0, + "content": "where the \\(\\lambda\\)s control the importance of each component." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Inference. During inference, given a target video \\( V_{t} \\) and a query video \\( V_{q} \\), we pass both through the model to generate foreground probabilities \\( \\{\\hat{f}_i\\}_{i=0}^{M-1} \\) and boundary predictions \\( \\{\\hat{d}_i\\}_{i=0}^{M-1} \\) using the two prediction heads. To handle the densely generated boundaries, while predicting \\( \\hat{d}_i \\), we apply 1-dimensional non-maximal suppression with a threshold of 0.7, filtering out highly overlapping boundaries and producing the final set of predictions. We choose \\( \\hat{d}_i \\) corresponding to position \\( i \\) with highest \\( \\hat{f}_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.387, + 0.73 + ], + "angle": 0, + "content": "3.3. Self-supervised Pre-training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The video representation learning community has leveraged self-supervised pre-training techniques designed for training effective encoders [22, 23, 37, 38, 45, 49, 50], which can be finetuned for enhanced performance on downstream tasks. In this work, we introduce a self-supervised pretraining objective designed to improve temporal localization capabilities without relying on labeled data. Specifically, given a target video \\( V_{t} \\), a query clip \\( V_{q} \\) is randomly sampled from \\( V_{t} \\) as shown in Fig. 3. The model is then trained to localize this query clip in \\( V_{t} \\). This pre-training objective closely aligns with the task of Vid2VidMR. To en" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "hance generalization capabilities and encourage the model to become robust to variations in both appearance and timing, we apply one of the following random augmentations to each of the query clips: reversing frames, adding gaussian noise, slowing down or speeding up the action, thereby doubling the number of pre-training samples. These augmentations introduce temporal and spatial variations, enabling the model to learn diverse representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.27 + ], + "angle": 0, + "content": "The overall pre-training loss, \\(\\mathcal{L}_{pt}\\), combines the foreground-background classification loss, the boundary prediction loss, pre-fusion and post-fusion alignment costs. The total loss is averaged over \\(P\\) pre-training samples as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.282, + 0.897, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p t} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\left(\\lambda_ {f g} ^ {p} \\mathcal {L} _ {\\mathrm {f g}} + \\lambda_ {s e g} ^ {p} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.331, + 0.89, + 0.347 + ], + "angle": 0, + "content": "where the \\(\\lambda\\)s control the importance of each component." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.611, + 0.377 + ], + "angle": 0, + "content": "4. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.905, + 0.434 + ], + "angle": 0, + "content": "We use the following two datasets to compare our proposed approach with existing Vid2VidMR methods and other strong baselines:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.906, + 0.571 + ], + "angle": 0, + "content": "(i) ActivityNet-VRL Dataset [14]: ActivityNet-VRL is the popular benchmark dataset shared by Feng et al. [14], based on the ActivityNet video understanding benchmark [17] consisting of 200 action classes. The dataset is split into disjoint 160 classes for training and 20 classes each for the validation and test splits. Furthermore, the training set comprises of \\(\\sim 463\\mathrm{K}\\) query-target video pairs, while the validation set and the test set have 829 and 978 query-target video pairs, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.875 + ], + "angle": 0, + "content": "(ii) Our proposed SportsMoments Dataset: One of the most promising applications for Vid2VidMR lies in the area of sports analytics. Although there exist large-scale sports datasets such as Sports-1M [21], they are primarily tailored for broader sports video classification tasks and do not contain fine-grained sports actions and moments, such as \"Cover Drive,\" \"Ducking a Bouncer,\" \"Goal Kick,\" or \"Penalty.\" Secondly, although ActivityNet-VRL covers a wide range of action categories (including sports actions), it only encompasses broader categories such as \"Playing Polo\" and does not focus on capturing higher-granularity events within a sport. Towards filling this gap, we introduce the SportsMoments dataset, which consists of \\(\\sim 770\\mathrm{K}\\) query-target pairs annotated from 176.6 hours of complete match footage of two of the most popular sports, viz. soccer and cricket. We obtain a total of 80 full-length cricket and soccer full-length match videos from YouTube. We then curate a list of 29 action classes comprising 13 soccer and 16 cricket actions, respectively. Given this list, we employed two annotators, each with strong knowledge of these sports," + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.841, + 0.901 + ], + "angle": 0, + "content": "3https://github.com/vl2g/MATR/tree/main/sportsmoments" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20219" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "to mark the start and end timestamps for the specified actions in the videos. We split SportsMoments into training, validation and test sets. The training split consists of approximately 750K pairs spanning 16 classes, while the validation and test splits each contain 10K pairs, covering four and nine classes, respectively. Action classes are disjoint across the train, validation, and test sets, ensuring no overlap. Additionally, each split includes cricket and soccer actions for a well-rounded distribution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.228, + 0.482, + 0.259 + ], + "angle": 0, + "content": "Additionally, we leverage the unlabeled videos from the Kinetics700 [6] dataset for pre-training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.275, + 0.325, + 0.291 + ], + "angle": 0, + "content": "5. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.199, + 0.315 + ], + "angle": 0, + "content": "5.1. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.482, + 0.354 + ], + "angle": 0, + "content": "We experiment with an extensive set of strong baselines grouped into four categories as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.476 + ], + "angle": 0, + "content": "(i) Fully-supervised Vid2VidMR methods: We compare with Vid2VidMR methods like CGBM [14], GDP [7], SRL [51], FFI+SRM [19], SST [4] and Video-level match [14]. Further, Huo et al. [19] adapt text-VMR methods like VSLNet [61], MABAN [46] and 2D-TAN [63] by replacing their text encoders with a C3D [47] feature extractor to extract query video features. All of these models have been trained on ActivityNet-VRL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.612 + ], + "angle": 0, + "content": "(ii) Vision-Language Models (VLMs): VLMs like VideoLLaMa [62], Video-LLaVa [26] and TimeChat [41] have shown promising results for multimodal text-video tasks like Video VQA, captioning, etc. Unfortunately, they have not been pre-trained with video-video aligned data. First we represent the query video using a state-of-the-art captioner, i.e., mPLUG-OWL [58]. Next, the VLMs are zero-shot prompted to generate start and end moment timestamps given query caption and target video tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.809 + ], + "angle": 0, + "content": "(iii) Text-VMR methods: Moment-DETR [25], QD-DETR [33], CG-DETR [32], and UniVTG [28] have been originally proposed for text-VMR. We compare with five variants of these methods. The zero-shot variant (a) uses caption from mPLUG-OWL [58] to represent query video, and leverages the pretrained checkpoints. Variants (b) and (c) both use caption from mPLUG-OWL [58] to represent query video, but train a randomly initialized checkpoint (variant b) or finetune the pre-trained checkpoint (variant c), respectively. Variants (d) and (e) are equivalent to (b) and (c) where their CLIP text encoder is replaced by CLIP ViT/B-32 vision encoder, and therefore take query video directly as input along with the target video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "(iv) Image-VMR methods: Our work focuses on video moment retrieval using 'video' queries. As a single image query may not effectively capture the temporal aspects of a video query, this comparison may not be appropriate. However, we still design baselines by representing the video by its key-frame (more precisely, the middle frame) using" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Text-VMR methods, namely Moment-DETR, QD-DETR and UniVTG." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.731, + 0.149 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.907, + 0.368 + ], + "angle": 0, + "content": "We choose hidden dimension of 1024 with \\( k = 4 \\) layers in both encoders and apply a dropout rate of 0.1 and 0.5 within the transformer and linear projection layers, respectively. Model weights are initialized using Xavier initialization [16]. To optimize the model parameters, we utilize AdamW optimizer [30] with an initial learning rate of 1e-4 and a weight decay of 1e-4. Training was done on two NVIDIA A6000 GPUs. We trained our model for 200 epochs with a batch size of 1200 while using ActivityNetVRL. For SportsMoments, we used 40 epochs with a batch size of 40. We set number of learnable queries \\( l = 10 \\) and all \\( \\lambda s \\) in both pre-training as well as finetuning losses to 1. We make our implementation and checkpoints available at: https://github.com/vl2g/MATR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.379, + 0.724, + 0.393 + ], + "angle": 0, + "content": "5.3. Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Main Results: Table 1 shows our main results where we compare MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and five variants of each of the Text-VMR methods. As mentioned before these variants correspond to zero-shot, finetuned or trained setup. As usual, trained setup implies random initialization while finetuned setup implies initialization using a pre-trained checkpoint (different for each architecture). Query video can be represented using text (T) captions obtained using mPLUGOWL [58] or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions and image query is not needed. Following previous work, we report the standard mean Intersection over Union (mIoU) and Recall@1 IoU=0.5 metrics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.764 + ], + "angle": 0, + "content": "Comparison with fully-supervised Vid2VidMR methods: For a fair comparison, we directly use the reported results from original papers for fully-supervised methods on ActivityNet-VRL. Consequently, we do not report their results on SportsMoments. As shown in Table 1, MATR outperforms the best method, FFI+SRM, with an \\(8.1\\%\\) gain in mIoU and achieves a \\(13.1\\%\\) improvement in R@1 over GDP, which had the highest recall. These results underscore MATR's superior ability to accurately localize moments using video queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Comparison with VLMs: Table 1 shows that our method MATR outperforms all four VLM-based baselines by a significant margin. In comparison, the best-performing baseline, TimeChat [41], achieves \\(26.4\\mathrm{mIoU}\\) and \\(23.8\\mathrm{R}@\\mathbb{1}\\) on ActivityNet-VRL, and \\(22.6\\mathrm{mIoU}\\) and \\(21.3\\mathrm{R}@\\mathbb{1}\\) on SportsMoments. Other methods, such as VideoLLaVA [26], Video-LLaMA [62], and Video-LLaMA2 [8], perform significantly worse. These results show that task-specific models for Vid2VidMR are significantly better." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20220" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.088, + 0.496, + 0.589 + ], + "angle": 0, + "content": "
MethodsVariantsActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Fully-supervised MethodsRandom [14]-7.316.2--
Video Match [14]-12.424.3--
SST [4]-17.133.2--
CGBM [14]-25.743.5--
GDP [7]-27.844.0--
SRL [51]-40.629.3--
2D-TAN [63]-45.339.6--
MABAN [46]-42.837.5--
VSLNet [61]-27.243.8--
FFI+SRM [19]-48.740.6--
VLMsVideo-LLaVA [26]-15.114.713.811.9
Video-LLaMA [62]-14.713.912.511.2
Video-LLaMA2 [8]-17.615.214.713.4
TimeChat [41]-26.423.822.621.3
I-VMRMoment-DETR [25]-35.632.525.218.5
QD-DETR [33]-37.238.127.720.7
UniVTG [28]-38.842.034.637.2
Text-VMR MethodsMoment-DETR [25](a) ZS+T28.222.83.70.9
(b) Trained+T37.135.831.429.2
(c) Finetuned+T40.038.935.834.0
(d) Trained+V35.934.728.724.2
(e) Finetuned+V40.039.930.425.1
QD-DETR [33](a) ZS+T25.422.26.32.6
(b) Trained+T38.439.029.827.2
(c) Finetuned+T41.645.236.033.5
(d) Trained+V39.741.027.225.1
(e) Finetuned+V42.542.735.430.6
UniVTG [28](a) ZS+T32.426.711.46.0
(b) Trained+T43.845.641.536.5
(c) Finetuned+T45.846.444.839.2
(d) Trained+V48.449.843.239.5
(e) Finetuned+V49.150.743.641.8
CG-DETR [32](a) ZS+T25.921.93.31.6
(b) Trained+T39.438.435.131.3
(c) Finetuned+T41.043.235.432.5
(d) Trained+V40.041.736.635.1
(e) Finetuned+V40.141.737.234.8
OursMATRZero-shot32.630.131.830.7
Trained+V53.254.856.252.7
Finetuned+V56.857.159.256.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.597, + 0.486, + 0.751 + ], + "angle": 0, + "content": "Table 1. Comparing MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and variants of Text-VMR methods. Trained implies random initialization. Finetuned implies initialization using a pre-trained (respective to each architecture) checkpoint. Query video can be represented using text (T) captions or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions is not needed. For more details of baselines and their variants, please refer to Section 5.1. Results for the overall best, best among implemented baselines and previously reported SOTA are highlighted in bold, underline and box, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.485, + 0.898 + ], + "angle": 0, + "content": "Comparison with Text-VMR Methods: Table 1 shows that amongst all the five variants, the finetuned variants are typically better than the trained variants, i.e., variant (c) is better than (b), and (e) is better than (d). On both ActivityNet-VRL and SportsMoments, UniVTG achieves the best results among variants which use video queries. However, our MATR model outperforms UniVTG by \\(7.7\\%\\) in mIoU and \\(6.4\\%\\) in R@1 on ActivityNet-VRL. On SportsMoments, MATR shows a further improvement, sur" + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.089, + 0.882, + 0.165 + ], + "angle": 0, + "content": "
Pre-FusionPost-FusionActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
XX49.750.249.146.6
X✓52.352.953.148.2
✓X50.451.252.446.7
✓✓53.254.856.252.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.174, + 0.905, + 0.203 + ], + "angle": 0, + "content": "Table 2. Advantage of explicit dual-stage alignment of MATR. For this study, experiments were performed without pre-training." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.21, + 0.889, + 0.282 + ], + "angle": 0, + "content": "
Pre-trainingAugmentationActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
X-53.254.856.252.7
✓X54.155.656.953.6
✓✓56.857.159.256.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.291, + 0.887, + 0.306 + ], + "angle": 0, + "content": "Table 3. Effect of pre-training and augmentation in MATR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.907, + 0.497 + ], + "angle": 0, + "content": "passing UniVTG by \\(15.6\\%\\) in mIoU and \\(14.7\\%\\) in R@1. These gains highlight the superior performance of MATR compared to the text-VMR methods when used with video query input. When text captions are used to represent the query, we observe similar trends as in the video query input case. Among the baselines, UniVTG performs the best. However, MATR outperforms it by \\(11\\%\\) in mIoU and \\(10.7\\%\\) in R@1 on ActivityNet-VRL. On SportsMoments, MATR achieves \\(14.4\\%\\) higher mIoU and \\(17.3\\%\\) higher R@1 than UniVTG. These results demonstrate the superior performance of MATR compared to the text-VMR methods when used with text query input for Vid2VidMR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Comparison with Image-VMR: Table 1 shows that the best-performing approach, UniVTG [28], achieves a mIoU of 38.8 and 42.0 R@1 on ActivityNet-VRL and 34.6 mIoU and R@1 37.2 on SportsMoments. Although this is comparable to previous fully supervised methods, it still lags significantly behind our MATR. This further goes to show the importance of temporal dynamics captured by video queries as compared to static image queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.867 + ], + "angle": 0, + "content": "Advantage of Dual Sequence Alignment: Table 2 shows the impact of different alignment strategies in MATR on both datasets. This study is done without any pre-training. We evaluate the effect of incorporating alignment at the pre-fusion and post-fusion stages. Our results indicate that the combined use of both pre-fusion and post-fusion alignment achieves the best performance, suggesting that alignment is important to ensure both before and after the encoder. Models with alignment either before (pre) or after (post) show moderate gains over neither. Removing post-fusion alignment hurts more than removing pre-fusion alignment, suggesting that post-fusion alignment is more important. Omitting alignment entirely results in the lowest scores. This highlights the importance of dual alignment for effectively capturing relevant temporal information for video query-based moment localization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of Pre-training: We analyze the impact of pretraining and data augmentation on model performance in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.092, + 0.361, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.092, + 0.872, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.261, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Visualization of Vid2VidMR on a sample from ActivityNet-VRL for calf roping action. Our proposed MATR model shows improved generalization capabilities over the best-performing baseline methods (variant (e), i.e., finetuned+V). Start and end times for the ground truth and predicted moments are shown in the brackets. Refer to supplementary material for the videos used in this figure." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.31, + 0.462, + 0.384 + ], + "angle": 0, + "content": "
ActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Pre-fusion alignment (Apre)34.132.631.629.8
Post-fusion alignment (Apost)38.335.236.433.5
Prediction-heads56.857.159.256.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.437 + ], + "angle": 0, + "content": "Table 4. Advantage of predicting using prediction heads on decoder. We observe that directly predicting from fusion matrices is inferior as it lacks fine-grained refinement." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.444, + 0.422, + 0.526 + ], + "angle": 0, + "content": "
AlignmentActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
None50.451.852.148.6
TCC52.452.753.952.8
Drop-DTW54.254.755.654.7
soft-DTW56.857.159.256.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.163, + 0.536, + 0.411, + 0.55 + ], + "angle": 0, + "content": "Table 5. Ablation on alignment methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.666 + ], + "angle": 0, + "content": "Table 3. On ActivityNet-VRL, without pre-training, the model achieves \\(53.2\\mathrm{mIoU}\\) and \\(54.8\\mathrm{R}@\\mathrm{l}\\). Adding pretraining without augmentation yields \\(54.1\\mathrm{mIoU}\\) and \\(55.6\\mathrm{R}@\\mathrm{l}\\). The best results are obtained by using pre-training with augmentation, leading to a significant boost to \\(56.8\\mathrm{mIoU}\\) and \\(57.1\\mathrm{R}@\\mathrm{l}\\). We got similar observations for the SportsMoments dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Predictions from Heads vs Alignment Matrices: The results in Table 4 show that prediction from heads outperforms the prediction from both pre-fusion alignment \\((\\mathrm{A}^{\\mathrm{pre}})\\) and post-fusion alignment \\((\\mathrm{A}^{\\mathrm{post}})\\) on both datasets. Notably, prediction heads improve mIoU and R@1 by a large margin, demonstrating their effectiveness in accurately retrieving temporal moments. Post-fusion alignment provides better results than pre-fusion alignment. However, the gains are modest compared to the substantial boost from prediction heads. This justifies the need for the decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation on Alignment Losses: MATR can incorporate different alignment methods. Table 5 compares the performance of different alignment methods on both ActivityNet-VRL and SportsMoments datasets. Without any alignment (\"None\"), the model achieves R@1 scores of \\(51.8\\%\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.42 + ], + "angle": 0, + "content": "on ActivityNet-VRL and \\(48.6\\%\\) on SportsMoments. TCC alignment yields an improvement of \\(0.9\\%\\) and \\(4.2\\%\\) on ActivityNet-VRL and SportsMoments, respectively. Drop-DTW is better than TCC. soft-DTW alignment consistently demonstrates superior performance across both datasets, achieving the highest gains of \\(5.3\\%\\) on ActivityNet-VRL and \\(7.9\\%\\) on SportsMoments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.909, + 0.603 + ], + "angle": 0, + "content": "Qualitative Results: We present moment retrieval results on a sample from ActivityNet-VRL in Fig. 4, where we compare our proposed approach to variant (e) (refer Sec. 5.1 for details) of the four best-performing baselines. The video query is a 5-second clip depicting calf roping4. The target video, with a duration of 24 seconds, showcases this event occurring between 2.1 and 18.1 seconds. Our proposed MATR model exhibits the highest overlap with the ground truth for the given query, accurately identifying the correct temporal boundaries. Among the baseline methods, CG-DETR and UniVTG achieve strong overlap, however, are still outperformed by our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.634, + 0.631 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.807 + ], + "angle": 0, + "content": "We introduced MATR, a robust approach for Vid2VidMR that combines abstract representation from encoder with fine-grained features from decoder conditioned on aligned query features. It captures semantic and temporal cues for precise moment localization. Our self-supervised pretraining enhances initialization and boosts performance. Extensive experiments on ActivityNet-VRL and our new SportsMoments dataset show that MATR outperforms strong baselines. Future directions of this work include exploring multimodal queries and developing scalable architectures to enhance broader applicability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Acknowledgment: This work is supported by the Microsoft Academic Partnership Grant (MAPG) 2023. Yogesh Kumar is supported by a UGC fellowship, Govt. of India." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "4A rodeo event where a rider on horseback attempts to catch and tie a calf within a timed competition." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20222" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.116, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[1] U. Agarwal, Y. Kumar, A. Shahid, P. Gatti, M. Gupta, and A. Mishra. CHAPVIDMR: Chapter-based video moment retrieval using natural language queries. In ICVGIP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.178, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[2] D. M. Argaw, J.-Y. Lee, M. Woodson, I.-S. Kweon, and F. C. Heilbron. Long-range multimodal pretraining for movie understanding. ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.224, + 0.482, + 0.268 + ], + "angle": 0, + "content": "[3] I. Bello, B. Zoph, Q. Le, A. Vaswani, and J. Schlens. Attention augmented convolutional networks. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.271, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[4] S. Buch, V. Escorcia, C. Shen, B. Ghanem, and J. C. Niebles. SST: single-stream temporal action proposals. In CVPR, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.317, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[5] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.363, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[6] J. Carreira and A. Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.409, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[7] L. Chen, C. Lu, S. Tang, J. Xiao, D. Zhang, C. Tan, and X. Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[8] Z. Cheng, S. Leng, H. Zhang, Y. Xin, X. Li, G. Chen, Y. Zhu, W. Zhang, Z. Luo, D. Zhao, and L. Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.532, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[9] M. Cuturi and M. Blondel. Soft-DTW: a differentiable loss function for time-series. In ICML, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.564, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[10] P. Dogan, B. Li, L. Sigal, and M. Gross. A neural multi-sequence alignment technique (neumatch). In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.61, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[11] N. Dvornik, I. Hadji, K. G. Derpanis, A. Garg, and A. D. Jepson. Drop-DTW: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[12] D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. Temporal cycle-consistency learning. In CVPR, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[13] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slow-fast networks for video recognition. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.764, + 0.482, + 0.792 + ], + "angle": 0, + "content": "[14] Y. Feng, L. Ma, W. Liu, T. Zhang, and J. Luo. Video re-localization. In ECCV, 2018. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.795, + 0.482, + 0.839 + ], + "angle": 0, + "content": "[15] J. Gao, C. Sun, Z. Yang, and R. Nevatia. TALL: temporal activity localization via language query. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.841, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[16] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "[17] F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.139, + 0.905, + 0.182 + ], + "angle": 0, + "content": "[18] L. A. Hendricks, O. Wang, E. Shechtman, J. Sivic, T. Darrell, and B. C. Russell. Localizing moments in video with natural language. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.185, + 0.905, + 0.242 + ], + "angle": 0, + "content": "[19] S. Huo, Y. Zhou, R. Wang, W. Xiang, and S. Kung. Semantic relevance learning for video-query based video moment retrieval. IEEE Trans. Multim., 25:9290-9301, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.246, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[20] M. Jung, Y. Jang, S. Choi, J. Kim, J.-H. Kim, and B.-T. Zhang. Background-aware moment detection for video moment retrieval. In WACV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[21] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.352, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[22] Y. Kumar, S. Mallick, A. Mishra, S. Rasipuram, A. Maitra, and R. Ramnani. Qdetrv: Query-guided detr for one-shot object localization in videos. In AAAI, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.413, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[23] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.459, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[24] J. Lei, L. Yu, T. L. Berg, and M. Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.549 + ], + "angle": 0, + "content": "[25] J. Lei, T. L. Berg, and M. Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.551, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[26] B. Lin, B. Zhu, Y. Ye, M. Ning, P. Jin, and L. Yuan. Video-llava: Learning united visual representation by alignment before projection. In EMNLP, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.905, + 0.639 + ], + "angle": 0, + "content": "[27] J. Lin, C. Gan, and S. Han. TSM: temporal shift module for efficient video understanding. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.642, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[28] K. Q. Lin, P. Zhang, J. Chen, S. Pramanick, D. Gao, A. J. Wang, R. Yan, and M. Z. Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[29] Y. Liu, S. Li, Y. Wu, C. W. Chen, Y. Shan, and X. Qie. UMT: unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.765, + 0.905, + 0.794 + ], + "angle": 0, + "content": "[30] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.796, + 0.905, + 0.84 + ], + "angle": 0, + "content": "[31] N. C. Mithun, S. Paul, and A. K. Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.841, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[32] W. Moon, S. Hyun, S. B. Lee, and J. Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, abs/2311.08835, 2023. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20223" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.091, + 0.486, + 0.152 + ], + "angle": 0, + "content": "[33] W. Moon, S. Hyun, S. Park, D. Park, and J. Heo. Query - dependent video representation for moment retrieval and highlight detection. In CVPR, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.153, + 0.485, + 0.198 + ], + "angle": 0, + "content": "[34] J. Mun, M. Cho, , and B. Han. Local-Global Video-Text Interactions for Temporal Grounding. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.2, + 0.483, + 0.246 + ], + "angle": 0, + "content": "[35] F. Padua, R. Carceroni, G. Santos, and K. Kutulakos. Linear sequence-to-sequence alignment. IEEE Trans. Pattern Anal. Mach. Intell., 32:304-320, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.248, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[36] N. Parmar, A. Vaswani, J. Uszkoreit, L. Kaiser, N. Shazeer, A. Ku, and D. Tran. Image transformer. In ICML, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.294, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[37] A. Piergiovanni, A. Angelova, and M. S. Ryoo. Evolving losses for unsupervised video representation learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.342, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[38] R. Qian, T. Meng, B. Gong, M.-H. Yang, H. Wang, S. Belongie, and Y. Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.388, + 0.483, + 0.463 + ], + "angle": 0, + "content": "[39] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.465, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[40] Rao, Gritai, and Shah. View-invariant alignment and matching of video sequences. In ICCV, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.498, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[41] S. Ren, L. Yao, S. Li, X. Sun, and L. Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.545, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[42] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.607, + 0.483, + 0.651 + ], + "angle": 0, + "content": "[43] Z. Shou, D. Wang, and S.-F. Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.653, + 0.483, + 0.697 + ], + "angle": 0, + "content": "[44] K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos. In NuerIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.7, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[45] N. Srivastava, E. Mansimov, and R. Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.747, + 0.483, + 0.807 + ], + "angle": 0, + "content": "[46] X. Sun, H. Wang, and B. He. Maban: Multi-agent boundary-aware network for natural language moment retrieval. IEEE Trans. on Image Processing, 30:5589-5599, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.809, + 0.483, + 0.854 + ], + "angle": 0, + "content": "[47] D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.091, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.091, + 0.906, + 0.137 + ], + "angle": 0, + "content": "[49] G. Wang, Y. Zhou, C. Luo, W. Xie, W. Zeng, and Z. Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.139, + 0.906, + 0.198 + ], + "angle": 0, + "content": "[50] J. Wang, Y. Gao, K. Li, J. Hu, X. Jiang, X. Guo, R. Ji, and X. Sun. Enhancing unsupervised video representation learning by decoupling the scene and the motion. In AAAI, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.201, + 0.906, + 0.246 + ], + "angle": 0, + "content": "[51] R. Wang and Y. Zhou. A feature pair fusion and hierarchical learning framework for video re-localization. In ICIP, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.248, + 0.906, + 0.292 + ], + "angle": 0, + "content": "[52] X. Wang, L. Zhu, and Y. Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.295, + 0.906, + 0.354 + ], + "angle": 0, + "content": "[53] Y. Xiao, Z. Luo, Y. Liu, Y. Ma, H. Bian, Y. Ji, Y. Yang, and X. Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.357, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[54] S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.419, + 0.906, + 0.478 + ], + "angle": 0, + "content": "[55] S. Yan, X. Xiong, A. Nagrani, A. Arnab, Z. Wang, W. Ge, D. Ross, and C. Schmid. Unloc: A unified framework for video localization tasks. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.481, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[56] C. Yang, Y. Xu, J. Shi, B. Dai, and B. Zhou. Temporal pyramid network for action recognition. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.528, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[57] J. Yang, P. Wei, H. Li, and Z. Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.591, + 0.906, + 0.651 + ], + "angle": 0, + "content": "[58] Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou. mplug-owi2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.653, + 0.906, + 0.697 + ], + "angle": 0, + "content": "[59] R. Zeng, W. Huang, M. Tan, Y. Rong, P. Zhao, J. Huang, and C. Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.7, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[60] C. Zhang, A. Gupta, and A. Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.747, + 0.906, + 0.807 + ], + "angle": 0, + "content": "[61] H. Zhang, A. Sun, W. Jing, L. Zhen, J. T. Zhou, and R. S. M. Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.809, + 0.906, + 0.854 + ], + "angle": 0, + "content": "[62] H. Zhang, X. Li, and L. Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[63] S. Zhang, H. Peng, J. Fu, and J. Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.091, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20224" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.136 + ], + "angle": 0, + "content": "[64] Z. Zhu, W. Tang, L. Wang, N. Zheng, and G. Hua. Enriching local and global contexts for temporal action localization. In ICCV, 2021. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "20225" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_origin.pdf b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f33b5c7c0f293a950ad8b5bd83138fc1c185a565 --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/f1f18e54-b74a-4e85-8e01-d0d39e528c08_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dcf5b57dbda6c9b34b76f617832e9ae27e095f7f039e42d28ec402abf04a498 +size 2958851 diff --git a/2025/Aligning Moments in Time using Video Queries/full.md b/2025/Aligning Moments in Time using Video Queries/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cfffaa75459f202dd789241a11ca3a3e616806d2 --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/full.md @@ -0,0 +1,309 @@ +# Aligning Moments in Time using Video Queries + +Yogesh Kumar1* Uday Agarwal1* Manish Gupta2 Anand Mishra1 +1Indian Institute of Technology Jodhpur 2Microsoft + +{kumar.204, agarwaluday, mishra}@iitj.ac.in, gmanish@microsoft.com + +# Abstract + +Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment Transformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of $13.1\%$ in $R@1$ and $8.1\%$ in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a $14.7\%$ gain in $R@1$ and a $14.4\%$ gain in mIoU on an absolute scale over strong baselines. We make the dataset and code public at: https://github.com/vl2g/MATR. + +# 1. Introduction + +Video moment retrieval is the task of temporally localizing the start and end times of a moment in a target video described by a given query. Although text-based video moment retrieval has been extensively explored [25, 28, 32, 33, 57], it often poses challenges for users attempting to + +Text Query "A man positions himself beneath the ball, leaping into the air, bending his knees and arching his back. As he flips backward, he connects with the ball using his foot." + +![](images/98a07981df1a760724ca4b2e0df71e3e6902dfca5e11499f05c3b39d944adcca.jpg) +Figure 1. (Top): Activity where text falls short to explain complex action or events, such as bicycle kick, necessitating the need for a more intuitive query modality for localizing semantically matching moments. (Bottom): Shows the Vid2Vid moment retrieval setting, which is the goal of this work. + +describe specific moments verbally. For example, consider a spectacular bicycle kick in soccer (Fig. 1 top). Although an expert can search by naming or describing it in detail, a beginner in soccer may struggle to describe this impressive move accurately. They might say something like "The player kicked the ball while in the air", which lacks the nuance needed to convey the athleticism and artistry involved. Further, such a description may lead to poor results for someone searching for similar moments. In contrast, if a soccer trainee shows a short video clip of a bicycle kick, the input is rich and crisp, making it easier for a retriever to locate similar moments in a target soccer match video. This approach aligns with how users naturally would prefer to search content within the same (video) modality, making it a more effective method for retrieving specific moments. Fig. 1 (bottom) shows a video bicycle kick query in soccer that is localized in a target video. + +Therefore, in this work, we study video-to-video moment retrieval (Vid2VidMR) where the aim is to temporally localize a moment in the target video with a high se + +matic match between the moment and the query video. Vid2VidMR, a task formally introduced by Feng et al. [14], has many potential application areas, such as sports video analytics, educational content creation and e-learning, and surveillance systems. This task is challenging and requires semantic frame-level alignment and modeling complex dependencies between query and target videos. The need for a semantic understanding of video content and variety in video length, context, and action speed calls for adaptive models capable of generalizing across diverse scenarios. Addressing these challenges demands advanced temporal video modeling techniques. + +To address the aforementioned challenges, we propose MATR (Moment Alignment TRansformer) - a method that uses explicit 'dual-stage sequence alignment' to capture the required correlation and temporal details essential for accurate moment localization. By conditioning target video representations on query features, MATR produces query-aligned representations that encode the necessary correlations and temporal dependencies between the two videos. We use differentiable dynamic time warping loss [9] for aligning the query and target videos, and represent the target video by conditioning it on the query video to focus on correlated temporal features. These representations guide a classification head to discriminate relevant moments from the irrelevant background and a boundary prediction head to mark the start and end of the identified moment in the target video. Further, to enhance our model's generalization using unlabeled videos, we introduce a self-supervised pretraining strategy which involves training MATR to localize randomly sampled clips within the same video, enabling it to learn the moment localization skill in a self-supervised manner. + +We evaluate MATR on public ActivityNet-VRL [14] benchmark and on our newly introduced dataset on sports domain, viz. SportsMoments, covering two of the most popular sports, namely soccer and cricket. Our approach achieves significant performance gains, with an improvement of $13.1\%$ and $8.1\%$ in R@1 and mIoU, respectively, on ActivityNet-VRL outperforming the state-of-the-art methods. Furthermore, MATR outperforms the implemented strong baselines on our proposed dataset with $14.7\%$ gains in R@1 and $14.4\%$ gains in mIoU, all on an absolute scale. + +Our contributions are as follows: (i) We introduce MATR, which uses explicit dual-stage sequence alignment within a transformer framework between target and query video to capture temporal correlations and dependencies for accurate moment localization. (ii) We propose a self-supervised pre-training objective that enhances model initialization by understanding rich video structure without requiring any labeled data. (iii) We conduct extensive experiments and ablations to study the efficacy of our + +framework against competitive baselines and state-of-the-art methods. Our findings offer valuable insights into our design choices, and our approach advances the state-of-the-art on Vid2VidMR. + +# 2. Related Work + +Video Moment Retrieval (VMR): VMR has recently gained significant interest in the research community [1, 15, 18, 24, 25, 28, 29, 32-34, 53, 55, 57]. Unlike video action understanding tasks such as action classification [6, 13, 27, 44, 47, 54, 56, 60] or temporal action localization [43, 59, 64], VMR focuses on identifying segments that semantically align with a broader range of queries, which may describe complex and context-specific moments. Based on the query modality, existing VMR methods can be broadly grouped into (i) textual query-based approaches like Moment-DETR [25], QD-DETR [33], UniVTG [28], and (ii) video query-based approaches like GDP [7], FFI+SRM [19], SRL [51]. Our work falls under video query-based VMR (Vid2VidMR). However, in addition to developing a new approach tailored for video query, we also adapt several text-query-based methods to make them suitable for Vid2VidMR and compare our method against both text and video query-based methods. + +**Alignment in Videos:** Alignment has been a thrust area in the video understanding community. It has been applied to a wide range of video understanding tasks, including video retrieval [52], procedural steps alignment [11], action recognition [10], anomaly detection [12], movie understanding [2] and video synchronization [35, 40]. Researchers have also explored sequence alignment for the text-VMR task, e.g., Mithun et al. [31] used sequence alignment between CNN representation of frames and GRU representation of text query to perform text-VMR. Jung et al. [20] used alignment to enhance semantic understanding between query and target video at the abstract level for the text-VMR task. However, their alignment is not at the sequence level. + +Compared to the existing literature, MATR goes beyond traditional sequence alignment by introducing a dual-stage alignment mechanism, leveraging transformer-based feature fusion, and maintaining flexibility with multiple alignment strategies. These innovations enable it to achieve more accurate and context-aware video moment retrieval compared to existing approaches. + +# 3. The MATR Model + +Our objective is to temporally localize a moment in a target video $V_{t}$ using a query video $V_{q}$ . We represent the target and query videos as sequences of $M$ and $N$ uniformly sampled frames, respectively. We refer to this problem as Vid2VidMR. In this work, we present Moment Alignment + +![](images/c87769f42ef8917b0b561b87a735c4eb7a7050ab40d41f078aa87032b7b285da.jpg) +Figure 2. Moment Alignment Transformer (MATR) Architecture. We represent the target video as a query-aligned representation capturing the required correlation and dependencies with query video at the abstract level (output of encoder) and finer level (output of decoder). These features are used to extract foreground target frames, as well as localize the relevant moment in terms of start and end times via two prediction heads. The alignment is computed using representations both before and after the encoder (shown using blue dotted and red dashed lines, respectively). [Best viewed in color]. + +TRansformer (MATR) - a transformer-based architecture that leverages explicit alignment for precise moment localization. MATR transforms the target video into a query-aligned representation via a transformer framework to exploit finer-level correlation and dependencies for precise moment localization. + +# 3.1. Motivation behind MATR Architecture + +Video moments vary significantly in duration, motion patterns, and visual appearance, making their retrieval inherently challenging. Additionally, preserving the correct temporal order of events is crucial for accurate localization, as moments often involve complex interactions that unfold over time. A robust Vid2VidMR model must, therefore, capture both high-level semantic relationships and fine-grained frame dependencies to ensure precise alignment between the query and target video. + +To address these challenges, MATR incorporates an explicit dual-stage sequence alignment strategy within a transformer-based encoder-decoder framework. This strategy enables MATR to learn a query-aligned representation of the target video by combining the abstract representation from the encoder with the refined fine-grained features from the decoder. In doing so, it facilitates precise moment localization by leveraging both global semantic alignment and fine-grained temporal dependencies between the query and target videos. Figure 2 illustrates the overall architecture of MATR, which we discuss in detail in the following. + +# 3.2. Architecture Details + +Input Representation. We uniformly sample frames every 2 seconds to obtain $M$ and $N$ frames from the target and query videos, respectively. We encode these frames using a frozen CLIP [39] (ViT-B/32) encoder along with a linear projection module. CLIP is applied separately to both videos to obtain target video embeddings $\mathrm{E}_t \in \mathbb{R}^{M \times d}$ and query video embeddings $\mathrm{E}_q \in \mathbb{R}^{N \times d}$ . The linear projection module is a two-layer perceptron, each with layer normalization and dropout. It maps 512-dimensional CLIP embeddings to $d$ -dimensional outputs. The resulting projected features $(\mathrm{E}_t$ and $\mathrm{E}_q)$ are concatenated along the sequence length dimension to form the input to the transformer encoder, denoted as $\mathrm{E}_c = [\mathrm{E}_t; \mathrm{E}_q] \in \mathbb{R}^{(M + N) \times d}$ . + +Encoder. The combined target and query video representation $(E_{c})$ is processed through a series of $k$ standard Transformer [48] encoder layers, each comprising a multi-head self-attention mechanism and a feed-forward network. Further, following the prior works [3, 5, 36], fixed positional encodings are added to the input of each attention layer to preserve temporal order. The encoder generates a fused representation $\left[\mathrm{E}_{\mathrm{t}}^{\mathrm{g}};\mathrm{E}_{\mathrm{q}}^{\mathrm{g}}\right]$ of the target video conditioned on the query video. Here $\mathrm{E}_{\mathrm{t}}^{\mathrm{g}}\in \mathbb{R}^{M\times d}$ and $\mathrm{E}_{\mathrm{q}}^{\mathrm{g}}\in \mathbb{R}^{N\times d}$ represent the target and query parts respectively. + +Dual-stage Sequence Alignment. The transformer encoder computes effective features by performing a joint understanding (or fusion) of target and query video frames. + +The final goal of Vid2VidMR is to localize a moment in the target video that semantically matches the content of the query video. In other words, we would prefer the features of the moment in the target to align strongly with the features of the query video. Therefore, we perform dual-stage sequence alignment, i.e., before and after the encoder. For alignment, we choose soft-DTW [9]² and perform the alignment as follows. + +Pre-fusion alignment: Before encoder fusion, given the target and query video feature sequences, $\mathrm{E_t} = [e_1^t,\dots,e_M^t ]\in \mathbb{R}^{M\times d}$ and $\mathrm{E_q} = [e_1^q,\dots,e_N^q ]\in \mathbb{R}^{N\times d}$ , soft-DTW outputs a binary alignment matrix $\mathrm{A}^{\mathrm{pre}}$ and an alignment cost matrix $\mathrm{C}^{\mathrm{pre}}$ . $\mathrm{A}^{\mathrm{pre}}\in \{0,1\}^{M\times N}$ is an alignment matrix such that: + +$$ +\mathrm {A} _ {i, j} ^ {\mathrm {p r e}} = \left\{ \begin{array}{l l} 1 & \text {i f} \boldsymbol {e} _ {i} ^ {t} \text {i s m a t c h e d t o} \boldsymbol {e} _ {j} ^ {q}, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. +$$ + +further, soft-DTW finds the optimal alignment by minimizing the pre-fusion alignment loss $\mathcal{L}_{\mathrm{align}}^{\mathrm{pre}}$ , defined as: + +$$ +\mathcal {L} _ {\text {a l i g n}} ^ {\text {p r e}} = \text {s o f t - D T W} _ {\gamma} (\mathrm {A} _ {i, j} ^ {\text {p r e}}, \mathrm {C} _ {i, j} ^ {\text {p r e}}), +$$ + +such that aligned frames in the target are contiguous. Here $\gamma$ is a smoothing factor for soft-min operator. The value of $\gamma$ in Soft-DTW is selected empirically to balance smoothness and alignment fidelity. Note that $C^{\mathrm{pre}}$ is the alignment cost matrix with elements defined by cosine similarity: + +$$ +\mathrm {C} _ {i, j} = 1 - \frac {\langle \pmb {e} _ {i} ^ {t} , \pmb {e} _ {j} ^ {q} \rangle}{\| \pmb {e} _ {i} ^ {t} \| \| \pmb {e} _ {j} ^ {q} \|}. +$$ + +Pre-fusion alignment enhances the semantic representation of target and query video. These enhanced representations are further processed by the encoder. The encoder outputs the fused representation of the target and query. + +Post-fusion alignment: Given the post-fusion target and query video feature sequences, $\mathrm{E_t^g}\in \mathbb{R}^{M\times d}$ and $\mathrm{E_q^g}\in$ $\mathbb{R}^{N\times d}$ , soft-DTW outputs a binary alignment matrix $\mathrm{A}^{\mathrm{post}}$ and an alignment cost matrix $\mathrm{C}^{\mathrm{post}}$ , soft-DTW achieves optimal alignment between $\mathrm{E_t^g}$ and $\mathrm{E_q^g}$ by minimizing the postfusion alignment loss $\mathcal{L}_{\mathrm{align}}^{\mathrm{post}}$ , defined as: + +$$ +\mathcal {L} _ {\text {a l i g n}} ^ {\text {p o s t}} = \text {s o f t - D T W} _ {\gamma} \left(\mathrm {A} _ {i, j} ^ {\text {p o s t}}, \mathrm {C} _ {i, j} ^ {\text {p o s t}}\right). +$$ + +Post-fusion alignment refines target video features by leveraging fused query-target representations, ensuring fine-grained semantic matching for precise moment localization. + +Decoder. The decoder in MATR further refines the target video representation by processing aligned query-target features, enabling fine-grained temporal matching and enhancing frame-level precision for accurate moment localization. + +To this end, the post-fusion alignment matrix $\mathbf{A}^{\mathrm{post}}$ defines a contiguous sub-sequence $\mathrm{E_t^g}[s : e]$ which aligns best with the query video. These aligned target features, capturing query-aligned fine-grained information, are passed as input to the decoder for further refinement. The decoder consists of $k$ Transformer layers, each containing a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward network. The input to the decoder comprises fixed size $l$ learnable query vectors denoted by $\mathrm{Q} \in [q_1, \ldots, q_l] \in \mathbb{R}^{l \times d}$ . These queries guide the extraction of refined features relevant to moment localization. As these queries are processed through each decoder layer, they are refined to capture intricate temporal dependencies, with positional encodings applied at each attention layer. The cross-attention layers enable interaction between $\mathrm{E_t^g}[s : e]$ from the encoder (which serve as keys and values) and representations of moment queries (which serve as queries). The final decoder output, $\mathrm{E_t^l} \in \mathbb{R}^{l \times d}$ , provides a refined fine-grained set of target video features conditioned on the query video. + +Finally, the encoder and decoder representations are combined as $\mathrm{E_f} = [\mathrm{E_t^g};\mathrm{E_t^l}]$ , making the final target video representation query-aligned. The final query-aligned representation, combining the abstract and fine-grained semantics, serves as input for the prediction heads. + +Prediction Heads. The prediction heads consist of three $(c = 3)$ sequential convolutional layers each having $d$ , $1 \times 3$ kernels followed by a ReLU activation. Finally, a sigmoid activation function is applied to produce the foreground predictions $\hat{f}_i$ . The model is trained using the following binary cross-entropy loss $\mathcal{L}_{fg}$ to distinguish foreground (relevant moment) from background predictions: + +$$ +\mathcal {L} _ {f g} = - \left(f _ {i} \log \hat {f} _ {i} + (1 - f _ {i}) \log (1 - \hat {f} _ {i})\right), +$$ + +where $f_{i}$ is the ground truth label, with $f_{i} = 1$ and $f_{i} = 0$ indicating foreground and background, respectively. + +In addition to foreground classification, the boundary prediction head is designed to estimate the start and end boundaries for the moment at every position. This head shares the same initial structure as the classification head but differs in the final output, producing two channels for left and right boundary offsets relative to the current position. Specifically, given $\mathrm{E_f}$ , the boundary prediction head outputs predicted offsets $\hat{d}_i = \{\hat{d}_i^L, \hat{d}_i^R\}$ for each position $i \in [0, M-1]$ , where $\hat{d}_i^L$ and $\hat{d}_i^R$ represent the left and right boundary offsets, respectively. The boundary prediction head is trained with a combination of smooth $L1$ and generalized intersection over union (IoU) loss [42], specifically applied to foreground positions $(f_i = 1)$ , defined as: + +$$ +\mathcal {L} _ {\mathrm {s e g}} = \mathbb {1} _ {f _ {i} = 1} \left[ \lambda_ {\mathrm {L} 1} \mathcal {L} _ {\mathrm {L} 1} \left(\hat {d} _ {i}, d _ {i}\right) + \lambda_ {\mathrm {I o U}} \mathcal {L} _ {\mathrm {I o U}} \left(\hat {d} _ {i}, d _ {i}\right) \right]. +$$ + +Here, $d_{i}$ denotes the ground truth offset. The parameters + +![](images/6c3e3ca473497383049fe5b203719752208f104a65d7982c83ffc5a1702f27bc.jpg) +Figure 3. Proposed self-supervised pre-training strategy: Given a target video, a query clip is randomly sampled from it and processed by the model. The model then predicts the boundaries of the selected clip within the target video, highlighting the corresponding frames with a green border. (Best viewed in color) + +$\lambda_{\mathrm{L1}}$ and $\lambda_{\mathrm{IoU}}$ weigh the contributions of the smooth L1 and IoU losses, ensuring that boundary localization focuses on the predictions identified as foreground. + +Convolutional layers operate along the temporal axis over the concatenated encoder query-conditioned and decoder-aligned target features. This preserves temporal continuity, enabling fine-grained predictions without boundary artifacts. + +Overall Loss. The overall loss function $\mathcal{L}$ is computed as a combination of multiple objectives, accounting for alignment costs (pre- and post-fusion alignment), foreground/background classification, and boundary localization. For $S$ training samples, this multi-objective overall loss is defined as: + +$$ +\mathcal {L} = \frac {1}{S} \sum_ {i = 1} ^ {S} \left(\lambda_ {f g} \mathcal {L} _ {f g} + \lambda_ {\mathrm {s e g}} \mathcal {L} _ {\mathrm {s e g}} + \lambda_ {\mathrm {a l i g n}} ^ {\mathrm {p r e}} \mathcal {L} _ {\mathrm {a l i g n}} ^ {\mathrm {p r e}} + \lambda_ {\mathrm {a l i g n}} ^ {\mathrm {p o s t}} \mathcal {L} _ {\mathrm {a l i g n}} ^ {\mathrm {p o s t}}\right), +$$ + +where the $\lambda$ s control the importance of each component. + +Inference. During inference, given a target video $V_{t}$ and a query video $V_{q}$ , we pass both through the model to generate foreground probabilities $\{\hat{f}_i\}_{i=0}^{M-1}$ and boundary predictions $\{\hat{d}_i\}_{i=0}^{M-1}$ using the two prediction heads. To handle the densely generated boundaries, while predicting $\hat{d}_i$ , we apply 1-dimensional non-maximal suppression with a threshold of 0.7, filtering out highly overlapping boundaries and producing the final set of predictions. We choose $\hat{d}_i$ corresponding to position $i$ with highest $\hat{f}_i$ . + +# 3.3. Self-supervised Pre-training Strategy + +The video representation learning community has leveraged self-supervised pre-training techniques designed for training effective encoders [22, 23, 37, 38, 45, 49, 50], which can be finetuned for enhanced performance on downstream tasks. In this work, we introduce a self-supervised pretraining objective designed to improve temporal localization capabilities without relying on labeled data. Specifically, given a target video $V_{t}$ , a query clip $V_{q}$ is randomly sampled from $V_{t}$ as shown in Fig. 3. The model is then trained to localize this query clip in $V_{t}$ . This pre-training objective closely aligns with the task of Vid2VidMR. To en + +hance generalization capabilities and encourage the model to become robust to variations in both appearance and timing, we apply one of the following random augmentations to each of the query clips: reversing frames, adding gaussian noise, slowing down or speeding up the action, thereby doubling the number of pre-training samples. These augmentations introduce temporal and spatial variations, enabling the model to learn diverse representations. + +The overall pre-training loss, $\mathcal{L}_{pt}$ , combines the foreground-background classification loss, the boundary prediction loss, pre-fusion and post-fusion alignment costs. The total loss is averaged over $P$ pre-training samples as: + +$$ +\mathcal {L} _ {p t} = \frac {1}{P} \sum_ {i = 1} ^ {P} \left(\lambda_ {f g} ^ {p} \mathcal {L} _ {\mathrm {f g}} + \lambda_ {s e g} ^ {p} \mathcal {L} _ {\mathrm {s e g}} + \lambda_ {\mathrm {a l i g n}} ^ {\mathrm {p r e}} \mathcal {L} _ {\mathrm {a l i g n}} ^ {\mathrm {p r e}} + \lambda_ {\mathrm {a l i g n}} ^ {\mathrm {p o s t}} \mathcal {L} _ {\mathrm {a l i g n}} ^ {\mathrm {p o s t}}\right), +$$ + +where the $\lambda$ s control the importance of each component. + +# 4. Datasets + +We use the following two datasets to compare our proposed approach with existing Vid2VidMR methods and other strong baselines: + +(i) ActivityNet-VRL Dataset [14]: ActivityNet-VRL is the popular benchmark dataset shared by Feng et al. [14], based on the ActivityNet video understanding benchmark [17] consisting of 200 action classes. The dataset is split into disjoint 160 classes for training and 20 classes each for the validation and test splits. Furthermore, the training set comprises of $\sim 463\mathrm{K}$ query-target video pairs, while the validation set and the test set have 829 and 978 query-target video pairs, respectively. + +(ii) Our proposed SportsMoments Dataset: One of the most promising applications for Vid2VidMR lies in the area of sports analytics. Although there exist large-scale sports datasets such as Sports-1M [21], they are primarily tailored for broader sports video classification tasks and do not contain fine-grained sports actions and moments, such as "Cover Drive," "Ducking a Bouncer," "Goal Kick," or "Penalty." Secondly, although ActivityNet-VRL covers a wide range of action categories (including sports actions), it only encompasses broader categories such as "Playing Polo" and does not focus on capturing higher-granularity events within a sport. Towards filling this gap, we introduce the SportsMoments dataset, which consists of $\sim 770\mathrm{K}$ query-target pairs annotated from 176.6 hours of complete match footage of two of the most popular sports, viz. soccer and cricket. We obtain a total of 80 full-length cricket and soccer full-length match videos from YouTube. We then curate a list of 29 action classes comprising 13 soccer and 16 cricket actions, respectively. Given this list, we employed two annotators, each with strong knowledge of these sports, + +to mark the start and end timestamps for the specified actions in the videos. We split SportsMoments into training, validation and test sets. The training split consists of approximately 750K pairs spanning 16 classes, while the validation and test splits each contain 10K pairs, covering four and nine classes, respectively. Action classes are disjoint across the train, validation, and test sets, ensuring no overlap. Additionally, each split includes cricket and soccer actions for a well-rounded distribution. + +Additionally, we leverage the unlabeled videos from the Kinetics700 [6] dataset for pre-training. + +# 5. Experiments and Results + +# 5.1. Baselines + +We experiment with an extensive set of strong baselines grouped into four categories as follows. + +(i) Fully-supervised Vid2VidMR methods: We compare with Vid2VidMR methods like CGBM [14], GDP [7], SRL [51], FFI+SRM [19], SST [4] and Video-level match [14]. Further, Huo et al. [19] adapt text-VMR methods like VSLNet [61], MABAN [46] and 2D-TAN [63] by replacing their text encoders with a C3D [47] feature extractor to extract query video features. All of these models have been trained on ActivityNet-VRL. + +(ii) Vision-Language Models (VLMs): VLMs like VideoLLaMa [62], Video-LLaVa [26] and TimeChat [41] have shown promising results for multimodal text-video tasks like Video VQA, captioning, etc. Unfortunately, they have not been pre-trained with video-video aligned data. First we represent the query video using a state-of-the-art captioner, i.e., mPLUG-OWL [58]. Next, the VLMs are zero-shot prompted to generate start and end moment timestamps given query caption and target video tokens. + +(iii) Text-VMR methods: Moment-DETR [25], QD-DETR [33], CG-DETR [32], and UniVTG [28] have been originally proposed for text-VMR. We compare with five variants of these methods. The zero-shot variant (a) uses caption from mPLUG-OWL [58] to represent query video, and leverages the pretrained checkpoints. Variants (b) and (c) both use caption from mPLUG-OWL [58] to represent query video, but train a randomly initialized checkpoint (variant b) or finetune the pre-trained checkpoint (variant c), respectively. Variants (d) and (e) are equivalent to (b) and (c) where their CLIP text encoder is replaced by CLIP ViT/B-32 vision encoder, and therefore take query video directly as input along with the target video. + +(iv) Image-VMR methods: Our work focuses on video moment retrieval using 'video' queries. As a single image query may not effectively capture the temporal aspects of a video query, this comparison may not be appropriate. However, we still design baselines by representing the video by its key-frame (more precisely, the middle frame) using + +Text-VMR methods, namely Moment-DETR, QD-DETR and UniVTG. + +# 5.2. Implementation Details + +We choose hidden dimension of 1024 with $k = 4$ layers in both encoders and apply a dropout rate of 0.1 and 0.5 within the transformer and linear projection layers, respectively. Model weights are initialized using Xavier initialization [16]. To optimize the model parameters, we utilize AdamW optimizer [30] with an initial learning rate of 1e-4 and a weight decay of 1e-4. Training was done on two NVIDIA A6000 GPUs. We trained our model for 200 epochs with a batch size of 1200 while using ActivityNetVRL. For SportsMoments, we used 40 epochs with a batch size of 40. We set number of learnable queries $l = 10$ and all $\lambda s$ in both pre-training as well as finetuning losses to 1. We make our implementation and checkpoints available at: https://github.com/vl2g/MATR. + +# 5.3. Results and Discussion + +Main Results: Table 1 shows our main results where we compare MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and five variants of each of the Text-VMR methods. As mentioned before these variants correspond to zero-shot, finetuned or trained setup. As usual, trained setup implies random initialization while finetuned setup implies initialization using a pre-trained checkpoint (different for each architecture). Query video can be represented using text (T) captions obtained using mPLUGOWL [58] or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions and image query is not needed. Following previous work, we report the standard mean Intersection over Union (mIoU) and Recall@1 IoU=0.5 metrics. + +Comparison with fully-supervised Vid2VidMR methods: For a fair comparison, we directly use the reported results from original papers for fully-supervised methods on ActivityNet-VRL. Consequently, we do not report their results on SportsMoments. As shown in Table 1, MATR outperforms the best method, FFI+SRM, with an $8.1\%$ gain in mIoU and achieves a $13.1\%$ improvement in R@1 over GDP, which had the highest recall. These results underscore MATR's superior ability to accurately localize moments using video queries. + +Comparison with VLMs: Table 1 shows that our method MATR outperforms all four VLM-based baselines by a significant margin. In comparison, the best-performing baseline, TimeChat [41], achieves $26.4\mathrm{mIoU}$ and $23.8\mathrm{R}@\mathbb{1}$ on ActivityNet-VRL, and $22.6\mathrm{mIoU}$ and $21.3\mathrm{R}@\mathbb{1}$ on SportsMoments. Other methods, such as VideoLLaVA [26], Video-LLaMA [62], and Video-LLaMA2 [8], perform significantly worse. These results show that task-specific models for Vid2VidMR are significantly better. + +
MethodsVariantsActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Fully-supervised MethodsRandom [14]-7.316.2--
Video Match [14]-12.424.3--
SST [4]-17.133.2--
CGBM [14]-25.743.5--
GDP [7]-27.844.0--
SRL [51]-40.629.3--
2D-TAN [63]-45.339.6--
MABAN [46]-42.837.5--
VSLNet [61]-27.243.8--
FFI+SRM [19]-48.740.6--
VLMsVideo-LLaVA [26]-15.114.713.811.9
Video-LLaMA [62]-14.713.912.511.2
Video-LLaMA2 [8]-17.615.214.713.4
TimeChat [41]-26.423.822.621.3
I-VMRMoment-DETR [25]-35.632.525.218.5
QD-DETR [33]-37.238.127.720.7
UniVTG [28]-38.842.034.637.2
Text-VMR MethodsMoment-DETR [25](a) ZS+T28.222.83.70.9
(b) Trained+T37.135.831.429.2
(c) Finetuned+T40.038.935.834.0
(d) Trained+V35.934.728.724.2
(e) Finetuned+V40.039.930.425.1
QD-DETR [33](a) ZS+T25.422.26.32.6
(b) Trained+T38.439.029.827.2
(c) Finetuned+T41.645.236.033.5
(d) Trained+V39.741.027.225.1
(e) Finetuned+V42.542.735.430.6
UniVTG [28](a) ZS+T32.426.711.46.0
(b) Trained+T43.845.641.536.5
(c) Finetuned+T45.846.444.839.2
(d) Trained+V48.449.843.239.5
(e) Finetuned+V49.150.743.641.8
CG-DETR [32](a) ZS+T25.921.93.31.6
(b) Trained+T39.438.435.131.3
(c) Finetuned+T41.043.235.432.5
(d) Trained+V40.041.736.635.1
(e) Finetuned+V40.141.737.234.8
OursMATRZero-shot32.630.131.830.7
Trained+V53.254.856.252.7
Finetuned+V56.857.159.256.5
+ +Table 1. Comparing MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and variants of Text-VMR methods. Trained implies random initialization. Finetuned implies initialization using a pre-trained (respective to each architecture) checkpoint. Query video can be represented using text (T) captions or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions is not needed. For more details of baselines and their variants, please refer to Section 5.1. Results for the overall best, best among implemented baselines and previously reported SOTA are highlighted in bold, underline and box, respectively. + +Comparison with Text-VMR Methods: Table 1 shows that amongst all the five variants, the finetuned variants are typically better than the trained variants, i.e., variant (c) is better than (b), and (e) is better than (d). On both ActivityNet-VRL and SportsMoments, UniVTG achieves the best results among variants which use video queries. However, our MATR model outperforms UniVTG by $7.7\%$ in mIoU and $6.4\%$ in R@1 on ActivityNet-VRL. On SportsMoments, MATR shows a further improvement, sur + +
Pre-FusionPost-FusionActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
XX49.750.249.146.6
X✓52.352.953.148.2
✓X50.451.252.446.7
✓✓53.254.856.252.7
+ +Table 2. Advantage of explicit dual-stage alignment of MATR. For this study, experiments were performed without pre-training. + +
Pre-trainingAugmentationActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
X-53.254.856.252.7
✓X54.155.656.953.6
✓✓56.857.159.256.5
+ +Table 3. Effect of pre-training and augmentation in MATR. + +passing UniVTG by $15.6\%$ in mIoU and $14.7\%$ in R@1. These gains highlight the superior performance of MATR compared to the text-VMR methods when used with video query input. When text captions are used to represent the query, we observe similar trends as in the video query input case. Among the baselines, UniVTG performs the best. However, MATR outperforms it by $11\%$ in mIoU and $10.7\%$ in R@1 on ActivityNet-VRL. On SportsMoments, MATR achieves $14.4\%$ higher mIoU and $17.3\%$ higher R@1 than UniVTG. These results demonstrate the superior performance of MATR compared to the text-VMR methods when used with text query input for Vid2VidMR. + +Comparison with Image-VMR: Table 1 shows that the best-performing approach, UniVTG [28], achieves a mIoU of 38.8 and 42.0 R@1 on ActivityNet-VRL and 34.6 mIoU and R@1 37.2 on SportsMoments. Although this is comparable to previous fully supervised methods, it still lags significantly behind our MATR. This further goes to show the importance of temporal dynamics captured by video queries as compared to static image queries. + +Advantage of Dual Sequence Alignment: Table 2 shows the impact of different alignment strategies in MATR on both datasets. This study is done without any pre-training. We evaluate the effect of incorporating alignment at the pre-fusion and post-fusion stages. Our results indicate that the combined use of both pre-fusion and post-fusion alignment achieves the best performance, suggesting that alignment is important to ensure both before and after the encoder. Models with alignment either before (pre) or after (post) show moderate gains over neither. Removing post-fusion alignment hurts more than removing pre-fusion alignment, suggesting that post-fusion alignment is more important. Omitting alignment entirely results in the lowest scores. This highlights the importance of dual alignment for effectively capturing relevant temporal information for video query-based moment localization. + +Effect of Pre-training: We analyze the impact of pretraining and data augmentation on model performance in + +![](images/2f34237d23fd96a08df298a0fd06983b207b140d2909476fa41ddd4e6c5e1cfa.jpg) +Figure 4. Visualization of Vid2VidMR on a sample from ActivityNet-VRL for calf roping action. Our proposed MATR model shows improved generalization capabilities over the best-performing baseline methods (variant (e), i.e., finetuned+V). Start and end times for the ground truth and predicted moments are shown in the brackets. Refer to supplementary material for the videos used in this figure. + +![](images/4a461c2ccd27c007ebe387383b266007d7fdb85eb2701c8b2c6968741cbd4da5.jpg) + +
ActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Pre-fusion alignment (Apre)34.132.631.629.8
Post-fusion alignment (Apost)38.335.236.433.5
Prediction-heads56.857.159.256.5
+ +Table 4. Advantage of predicting using prediction heads on decoder. We observe that directly predicting from fusion matrices is inferior as it lacks fine-grained refinement. + +
AlignmentActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
None50.451.852.148.6
TCC52.452.753.952.8
Drop-DTW54.254.755.654.7
soft-DTW56.857.159.256.5
+ +Table 5. Ablation on alignment methods. + +Table 3. On ActivityNet-VRL, without pre-training, the model achieves $53.2\mathrm{mIoU}$ and $54.8\mathrm{R}@\mathrm{l}$ . Adding pretraining without augmentation yields $54.1\mathrm{mIoU}$ and $55.6\mathrm{R}@\mathrm{l}$ . The best results are obtained by using pre-training with augmentation, leading to a significant boost to $56.8\mathrm{mIoU}$ and $57.1\mathrm{R}@\mathrm{l}$ . We got similar observations for the SportsMoments dataset. + +Predictions from Heads vs Alignment Matrices: The results in Table 4 show that prediction from heads outperforms the prediction from both pre-fusion alignment $(\mathrm{A}^{\mathrm{pre}})$ and post-fusion alignment $(\mathrm{A}^{\mathrm{post}})$ on both datasets. Notably, prediction heads improve mIoU and R@1 by a large margin, demonstrating their effectiveness in accurately retrieving temporal moments. Post-fusion alignment provides better results than pre-fusion alignment. However, the gains are modest compared to the substantial boost from prediction heads. This justifies the need for the decoder. + +Ablation on Alignment Losses: MATR can incorporate different alignment methods. Table 5 compares the performance of different alignment methods on both ActivityNet-VRL and SportsMoments datasets. Without any alignment ("None"), the model achieves R@1 scores of $51.8\%$ + +on ActivityNet-VRL and $48.6\%$ on SportsMoments. TCC alignment yields an improvement of $0.9\%$ and $4.2\%$ on ActivityNet-VRL and SportsMoments, respectively. Drop-DTW is better than TCC. soft-DTW alignment consistently demonstrates superior performance across both datasets, achieving the highest gains of $5.3\%$ on ActivityNet-VRL and $7.9\%$ on SportsMoments. + +Qualitative Results: We present moment retrieval results on a sample from ActivityNet-VRL in Fig. 4, where we compare our proposed approach to variant (e) (refer Sec. 5.1 for details) of the four best-performing baselines. The video query is a 5-second clip depicting calf roping4. The target video, with a duration of 24 seconds, showcases this event occurring between 2.1 and 18.1 seconds. Our proposed MATR model exhibits the highest overlap with the ground truth for the given query, accurately identifying the correct temporal boundaries. Among the baseline methods, CG-DETR and UniVTG achieve strong overlap, however, are still outperformed by our approach. + +# 6. Conclusion + +We introduced MATR, a robust approach for Vid2VidMR that combines abstract representation from encoder with fine-grained features from decoder conditioned on aligned query features. It captures semantic and temporal cues for precise moment localization. Our self-supervised pretraining enhances initialization and boosts performance. Extensive experiments on ActivityNet-VRL and our new SportsMoments dataset show that MATR outperforms strong baselines. Future directions of this work include exploring multimodal queries and developing scalable architectures to enhance broader applicability. + +Acknowledgment: This work is supported by the Microsoft Academic Partnership Grant (MAPG) 2023. Yogesh Kumar is supported by a UGC fellowship, Govt. of India. + +# References + +[1] U. Agarwal, Y. Kumar, A. Shahid, P. Gatti, M. Gupta, and A. Mishra. CHAPVIDMR: Chapter-based video moment retrieval using natural language queries. In ICVGIP, 2024. 2 +[2] D. M. Argaw, J.-Y. Lee, M. Woodson, I.-S. Kweon, and F. C. Heilbron. Long-range multimodal pretraining for movie understanding. ICCV, 2023. 2 +[3] I. Bello, B. Zoph, Q. Le, A. Vaswani, and J. Schlens. Attention augmented convolutional networks. In ICCV, 2019. 3 +[4] S. Buch, V. Escorcia, C. Shen, B. Ghanem, and J. C. Niebles. SST: single-stream temporal action proposals. In CVPR, 2017. 6, 7 +[5] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3 +[6] J. Carreira and A. Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 2, 6 +[7] L. Chen, C. Lu, S. Tang, J. Xiao, D. Zhang, C. Tan, and X. Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2, 6, 7 +[8] Z. Cheng, S. Leng, H. Zhang, Y. Xin, X. Li, G. Chen, Y. Zhu, W. Zhang, Z. Luo, D. Zhao, and L. Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6, 7 +[9] M. Cuturi and M. Blondel. Soft-DTW: a differentiable loss function for time-series. In ICML, 2017. 2, 4 +[10] P. Dogan, B. Li, L. Sigal, and M. Gross. A neural multi-sequence alignment technique (neumatch). In CVPR, 2018. 2 +[11] N. Dvornik, I. Hadji, K. G. Derpanis, A. Garg, and A. D. Jepson. Drop-DTW: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 2, 4 +[12] D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. Temporal cycle-consistency learning. In CVPR, 2019. 2, 4 +[13] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slow-fast networks for video recognition. In ICCV, 2019. 2 +[14] Y. Feng, L. Ma, W. Liu, T. Zhang, and J. Luo. Video re-localization. In ECCV, 2018. 2, 5, 6, 7 +[15] J. Gao, C. Sun, Z. Yang, and R. Nevatia. TALL: temporal activity localization via language query. In ICCV, 2017. 2 +[16] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 6 + +[17] F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 5 +[18] L. A. Hendricks, O. Wang, E. Shechtman, J. Sivic, T. Darrell, and B. C. Russell. Localizing moments in video with natural language. In ICCV, 2017. 2 +[19] S. Huo, Y. Zhou, R. Wang, W. Xiang, and S. Kung. Semantic relevance learning for video-query based video moment retrieval. IEEE Trans. Multim., 25:9290-9301, 2023. 2, 6, 7 +[20] M. Jung, Y. Jang, S. Choi, J. Kim, J.-H. Kim, and B.-T. Zhang. Background-aware moment detection for video moment retrieval. In WACV, 2025. 2 +[21] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 5 +[22] Y. Kumar, S. Mallick, A. Mishra, S. Rasipuram, A. Maitra, and R. Ramnani. Qdetrv: Query-guided detr for one-shot object localization in videos. In AAAI, 2024. 5 +[23] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In ICCV, 2017. 5 +[24] J. Lei, L. Yu, T. L. Berg, and M. Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 2 +[25] J. Lei, T. L. Berg, and M. Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 1, 2, 6, 7 +[26] B. Lin, B. Zhu, Y. Ye, M. Ning, P. Jin, and L. Yuan. Video-llava: Learning united visual representation by alignment before projection. In EMNLP, 2024. 6, 7 +[27] J. Lin, C. Gan, and S. Han. TSM: temporal shift module for efficient video understanding. In ICCV, 2019. 2 +[28] K. Q. Lin, P. Zhang, J. Chen, S. Pramanick, D. Gao, A. J. Wang, R. Yan, and M. Z. Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, 2023. 1, 2, 6, 7 +[29] Y. Liu, S. Li, Y. Wu, C. W. Chen, Y. Shan, and X. Qie. UMT: unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2 +[30] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 6 +[31] N. C. Mithun, S. Paul, and A. K. Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2 +[32] W. Moon, S. Hyun, S. B. Lee, and J. Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, abs/2311.08835, 2023. 1, 2, 6, 7 + +[33] W. Moon, S. Hyun, S. Park, D. Park, and J. Heo. Query - dependent video representation for moment retrieval and highlight detection. In CVPR, 2023. 1, 2, 6, 7 +[34] J. Mun, M. Cho, , and B. Han. Local-Global Video-Text Interactions for Temporal Grounding. In CVPR, 2020. 2 +[35] F. Padua, R. Carceroni, G. Santos, and K. Kutulakos. Linear sequence-to-sequence alignment. IEEE Trans. Pattern Anal. Mach. Intell., 32:304-320, 2008. 2 +[36] N. Parmar, A. Vaswani, J. Uszkoreit, L. Kaiser, N. Shazeer, A. Ku, and D. Tran. Image transformer. In ICML, 2018. 3 +[37] A. Piergiovanni, A. Angelova, and M. S. Ryoo. Evolving losses for unsupervised video representation learning. In CVPR, 2020. 5 +[38] R. Qian, T. Meng, B. Gong, M.-H. Yang, H. Wang, S. Belongie, and Y. Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 5 +[39] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[40] Rao, Gritai, and Shah. View-invariant alignment and matching of video sequences. In ICCV, 2003. 2 +[41] S. Ren, L. Yao, S. Li, X. Sun, and L. Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, 2024. 6, 7 +[42] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4 +[43] Z. Shou, D. Wang, and S.-F. Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In CVPR, 2016. 2 +[44] K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos. In NuerIPS, 2014. 2 +[45] N. Srivastava, E. Mansimov, and R. Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 5 +[46] X. Sun, H. Wang, and B. He. Maban: Multi-agent boundary-aware network for natural language moment retrieval. IEEE Trans. on Image Processing, 30:5589-5599, 2021. 6, 7 +[47] D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 2, 6 +[48] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In NeurIPS, 2017. 3 + +[49] G. Wang, Y. Zhou, C. Luo, W. Xie, W. Zeng, and Z. Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, 2021. 5 +[50] J. Wang, Y. Gao, K. Li, J. Hu, X. Jiang, X. Guo, R. Ji, and X. Sun. Enhancing unsupervised video representation learning by decoupling the scene and the motion. In AAAI, 2021. 5 +[51] R. Wang and Y. Zhou. A feature pair fusion and hierarchical learning framework for video re-localization. In ICIP, 2020. 2, 6, 7 +[52] X. Wang, L. Zhu, and Y. Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, 2021. 2 +[53] Y. Xiao, Z. Luo, Y. Liu, Y. Ma, H. Bian, Y. Ji, Y. Yang, and X. Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, 2024. 2 +[54] S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018. 2 +[55] S. Yan, X. Xiong, A. Nagrani, A. Arnab, Z. Wang, W. Ge, D. Ross, and C. Schmid. Unloc: A unified framework for video localization tasks. In ICCV, 2023. 2 +[56] C. Yang, Y. Xu, J. Shi, B. Dai, and B. Zhou. Temporal pyramid network for action recognition. In CVPR, 2020. 2 +[57] J. Yang, P. Wei, H. Li, and Z. Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, 2024. 1, 2 +[58] Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou. mplug-owi2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6 +[59] R. Zeng, W. Huang, M. Tan, Y. Rong, P. Zhao, J. Huang, and C. Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2 +[60] C. Zhang, A. Gupta, and A. Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2 +[61] H. Zhang, A. Sun, W. Jing, L. Zhen, J. T. Zhou, and R. S. M. Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 6, 7 +[62] H. Zhang, X. Li, and L. Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 6, 7 +[63] S. Zhang, H. Peng, J. Fu, and J. Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 6, 7 + +[64] Z. Zhu, W. Tang, L. Wang, N. Zheng, and G. Hua. Enriching local and global contexts for temporal action localization. In ICCV, 2021. 2 \ No newline at end of file diff --git a/2025/Aligning Moments in Time using Video Queries/images.zip b/2025/Aligning Moments in Time using Video Queries/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..75a24f30d2cc77fc8c38d5022b9661ca4e08ad59 --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5db5e8359a4885c28f66079992e0424c3f2e23fc30a70587d77d04fef324c0fd +size 451500 diff --git a/2025/Aligning Moments in Time using Video Queries/layout.json b/2025/Aligning Moments in Time using Video Queries/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee56ba070fe246ddc485aa106ac37f614125ca41 --- /dev/null +++ b/2025/Aligning Moments in Time using Video Queries/layout.json @@ -0,0 +1,9174 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 158, + 103, + 452, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 103, + 452, + 121 + ], + "spans": [ + { + "bbox": [ + 158, + 103, + 452, + 121 + ], + "type": "text", + "content": "Aligning Moments in Time using Video Queries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 142, + 480, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 142, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 128, + 142, + 480, + 172 + ], + "type": "text", + "content": "Yogesh Kumar1* Uday Agarwal1* Manish Gupta2 Anand Mishra1 \n1Indian Institute of Technology Jodhpur 2Microsoft" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 173, + 482, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 173, + 482, + 186 + ], + "spans": [ + { + "bbox": [ + 126, + 173, + 482, + 186 + ], + "type": "text", + "content": "{kumar.204, agarwaluday, mishra}@iitj.ac.in, gmanish@microsoft.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": "Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment Transformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "14.7\\%" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " gain in " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 55, + 238, + 296, + 574 + ], + "type": "text", + "content": " gain in mIoU on an absolute scale over strong baselines. We make the dataset and code public at: https://github.com/vl2g/MATR." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 596, + 135, + 608 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 676 + ], + "type": "text", + "content": "Video moment retrieval is the task of temporally localizing the start and end times of a moment in a target video described by a given query. Although text-based video moment retrieval has been extensively explored [25, 28, 32, 33, 57], it often poses challenges for users attempting to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 213, + 554, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 213, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 213, + 554, + 239 + ], + "type": "text", + "content": "Text Query \"A man positions himself beneath the ball, leaping into the air, bending his knees and arching his back. As he flips backward, he connects with the ball using his foot.\"" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 240, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 316, + 240, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 398 + ], + "type": "image", + "image_path": "98a07981df1a760724ca4b2e0df71e3e6902dfca5e11499f05c3b39d944adcca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. (Top): Activity where text falls short to explain complex action or events, such as bicycle kick, necessitating the need for a more intuitive query modality for localizing semantically matching moments. (Bottom): Shows the Vid2Vid moment retrieval setting, which is the goal of this work." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 673 + ], + "type": "text", + "content": "describe specific moments verbally. For example, consider a spectacular bicycle kick in soccer (Fig. 1 top). Although an expert can search by naming or describing it in detail, a beginner in soccer may struggle to describe this impressive move accurately. They might say something like \"The player kicked the ball while in the air\", which lacks the nuance needed to convey the athleticism and artistry involved. Further, such a description may lead to poor results for someone searching for similar moments. In contrast, if a soccer trainee shows a short video clip of a bicycle kick, the input is rich and crisp, making it easier for a retriever to locate similar moments in a target soccer match video. This approach aligns with how users naturally would prefer to search content within the same (video) modality, making it a more effective method for retrieving specific moments. Fig. 1 (bottom) shows a video bicycle kick query in soccer that is localized in a target video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Therefore, in this work, we study video-to-video moment retrieval (Vid2VidMR) where the aim is to temporally localize a moment in the target video with a high se" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 683, + 135, + 693 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "type": "text", + "content": "1A moment is a continuous segment of frames within a target video that best represents the actions, events, or interactions described by the query." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20215" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "matic match between the moment and the query video. Vid2VidMR, a task formally introduced by Feng et al. [14], has many potential application areas, such as sports video analytics, educational content creation and e-learning, and surveillance systems. This task is challenging and requires semantic frame-level alignment and modeling complex dependencies between query and target videos. The need for a semantic understanding of video content and variety in video length, context, and action speed calls for adaptive models capable of generalizing across diverse scenarios. Addressing these challenges demands advanced temporal video modeling techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 469 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose MATR (Moment Alignment TRansformer) - a method that uses explicit 'dual-stage sequence alignment' to capture the required correlation and temporal details essential for accurate moment localization. By conditioning target video representations on query features, MATR produces query-aligned representations that encode the necessary correlations and temporal dependencies between the two videos. We use differentiable dynamic time warping loss [9] for aligning the query and target videos, and represent the target video by conditioning it on the query video to focus on correlated temporal features. These representations guide a classification head to discriminate relevant moments from the irrelevant background and a boundary prediction head to mark the start and end of the identified moment in the target video. Further, to enhance our model's generalization using unlabeled videos, we introduce a self-supervised pretraining strategy which involves training MATR to localize randomly sampled clips within the same video, enabling it to learn the moment localization skill in a self-supervised manner." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "content": "We evaluate MATR on public ActivityNet-VRL [14] benchmark and on our newly introduced dataset on sports domain, viz. SportsMoments, covering two of the most popular sports, namely soccer and cricket. Our approach achieves significant performance gains, with an improvement of " + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "content": " in R@1 and mIoU, respectively, on ActivityNet-VRL outperforming the state-of-the-art methods. Furthermore, MATR outperforms the implemented strong baselines on our proposed dataset with " + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "inline_equation", + "content": "14.7\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "content": " gains in R@1 and " + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 55, + 472, + 295, + 602 + ], + "type": "text", + "content": " gains in mIoU, all on an absolute scale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Our contributions are as follows: (i) We introduce MATR, which uses explicit dual-stage sequence alignment within a transformer framework between target and query video to capture temporal correlations and dependencies for accurate moment localization. (ii) We propose a self-supervised pre-training objective that enhances model initialization by understanding rich video structure without requiring any labeled data. (iii) We conduct extensive experiments and ablations to study the efficacy of our" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "framework against competitive baselines and state-of-the-art methods. Our findings offer valuable insights into our design choices, and our approach advances the state-of-the-art on Vid2VidMR." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 133, + 400, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 400, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 400, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 152, + 553, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 369 + ], + "type": "text", + "content": "Video Moment Retrieval (VMR): VMR has recently gained significant interest in the research community [1, 15, 18, 24, 25, 28, 29, 32-34, 53, 55, 57]. Unlike video action understanding tasks such as action classification [6, 13, 27, 44, 47, 54, 56, 60] or temporal action localization [43, 59, 64], VMR focuses on identifying segments that semantically align with a broader range of queries, which may describe complex and context-specific moments. Based on the query modality, existing VMR methods can be broadly grouped into (i) textual query-based approaches like Moment-DETR [25], QD-DETR [33], UniVTG [28], and (ii) video query-based approaches like GDP [7], FFI+SRM [19], SRL [51]. Our work falls under video query-based VMR (Vid2VidMR). However, in addition to developing a new approach tailored for video query, we also adapt several text-query-based methods to make them suitable for Vid2VidMR and compare our method against both text and video query-based methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 369, + 553, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 553, + 536 + ], + "type": "text", + "content": "**Alignment in Videos:** Alignment has been a thrust area in the video understanding community. It has been applied to a wide range of video understanding tasks, including video retrieval [52], procedural steps alignment [11], action recognition [10], anomaly detection [12], movie understanding [2] and video synchronization [35, 40]. Researchers have also explored sequence alignment for the text-VMR task, e.g., Mithun et al. [31] used sequence alignment between CNN representation of frames and GRU representation of text query to perform text-VMR. Jung et al. [20] used alignment to enhance semantic understanding between query and target video at the abstract level for the text-VMR task. However, their alignment is not at the sequence level." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 537, + 553, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 553, + 621 + ], + "type": "text", + "content": "Compared to the existing literature, MATR goes beyond traditional sequence alignment by introducing a dual-stage alignment mechanism, leveraging transformer-based feature fusion, and maintaining flexibility with multiple alignment strategies. These innovations enable it to achieve more accurate and context-aware video moment retrieval compared to existing approaches." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 633, + 423, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 423, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 423, + 645 + ], + "type": "text", + "content": "3. The MATR Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Our objective is to temporally localize a moment in a target video " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " using a query video " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "V_{q}" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ". We represent the target and query videos as sequences of " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " uniformly sampled frames, respectively. We refer to this problem as Vid2VidMR. In this work, we present Moment Alignment" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20216" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 69, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 60, + 69, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 60, + 69, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 60, + 69, + 553, + 308 + ], + "type": "image", + "image_path": "c87769f42ef8917b0b561b87a735c4eb7a7050ab40d41f078aa87032b7b285da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 373 + ], + "type": "text", + "content": "Figure 2. Moment Alignment Transformer (MATR) Architecture. We represent the target video as a query-aligned representation capturing the required correlation and dependencies with query video at the abstract level (output of encoder) and finer level (output of decoder). These features are used to extract foreground target frames, as well as localize the relevant moment in terms of start and end times via two prediction heads. The alignment is computed using representations both before and after the encoder (shown using blue dotted and red dashed lines, respectively). [Best viewed in color]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 379, + 298, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 379, + 298, + 452 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 298, + 452 + ], + "type": "text", + "content": "TRansformer (MATR) - a transformer-based architecture that leverages explicit alignment for precise moment localization. MATR transforms the target video into a query-aligned representation via a transformer framework to exploit finer-level correlation and dependencies for precise moment localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 457, + 247, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 247, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 247, + 469 + ], + "type": "text", + "content": "3.1. Motivation behind MATR Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": "Video moments vary significantly in duration, motion patterns, and visual appearance, making their retrieval inherently challenging. Additionally, preserving the correct temporal order of events is crucial for accurate localization, as moments often involve complex interactions that unfold over time. A robust Vid2VidMR model must, therefore, capture both high-level semantic relationships and fine-grained frame dependencies to ensure precise alignment between the query and target video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 298, + 715 + ], + "type": "text", + "content": "To address these challenges, MATR incorporates an explicit dual-stage sequence alignment strategy within a transformer-based encoder-decoder framework. This strategy enables MATR to learn a query-aligned representation of the target video by combining the abstract representation from the encoder with the refined fine-grained features from the decoder. In doing so, it facilitates precise moment localization by leveraging both global semantic alignment and fine-grained temporal dependencies between the query and target videos. Figure 2 illustrates the overall architecture of MATR, which we discuss in detail in the following." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 379, + 424, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 424, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 424, + 390 + ], + "type": "text", + "content": "3.2. Architecture Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": "Input Representation. We uniformly sample frames every 2 seconds to obtain " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": " frames from the target and query videos, respectively. We encode these frames using a frozen CLIP [39] (ViT-B/32) encoder along with a linear projection module. CLIP is applied separately to both videos to obtain target video embeddings " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_t \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": " and query video embeddings " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_q \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": ". The linear projection module is a two-layer perceptron, each with layer normalization and dropout. It maps 512-dimensional CLIP embeddings to " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": "-dimensional outputs. The resulting projected features " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "(\\mathrm{E}_t" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_q)" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": " are concatenated along the sequence length dimension to form the input to the transformer encoder, denoted as " + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_c = [\\mathrm{E}_t; \\mathrm{E}_q] \\in \\mathbb{R}^{(M + N) \\times d}" + }, + { + "bbox": [ + 313, + 398, + 555, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": "Encoder. The combined target and query video representation " + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "inline_equation", + "content": "(E_{c})" + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": " is processed through a series of " + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": " standard Transformer [48] encoder layers, each comprising a multi-head self-attention mechanism and a feed-forward network. Further, following the prior works [3, 5, 36], fixed positional encodings are added to the input of each attention layer to preserve temporal order. The encoder generates a fused representation " + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\left[\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}};\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\right]" + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": " of the target video conditioned on the query video. Here " + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\mathrm{t}}^{\\mathrm{g}}\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{\\mathrm{q}}^{\\mathrm{g}}\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 313, + 555, + 556, + 676 + ], + "type": "text", + "content": " represent the target and query parts respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Dual-stage Sequence Alignment. The transformer encoder computes effective features by performing a joint understanding (or fusion) of target and query video frames." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20217" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "The final goal of Vid2VidMR is to localize a moment in the target video that semantically matches the content of the query video. In other words, we would prefer the features of the moment in the target to align strongly with the features of the query video. Therefore, we perform dual-stage sequence alignment, i.e., before and after the encoder. For alignment, we choose soft-DTW [9]² and perform the alignment as follows." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": "Pre-fusion alignment: Before encoder fusion, given the target and query video feature sequences, " + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t} = [e_1^t,\\dots,e_M^t ]\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{E_q} = [e_1^q,\\dots,e_N^q ]\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": ", soft-DTW outputs a binary alignment matrix " + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": " and an alignment cost matrix " + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^{\\mathrm{pre}}\\in \\{0,1\\}^{M\\times N}" + }, + { + "bbox": [ + 55, + 167, + 296, + 239 + ], + "type": "text", + "content": " is an alignment matrix such that:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 244, + 249, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 244, + 249, + 275 + ], + "spans": [ + { + "bbox": [ + 102, + 244, + 249, + 275 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {i, j} ^ {\\mathrm {p r e}} = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\boldsymbol {e} _ {i} ^ {t} \\text {i s m a t c h e d t o} \\boldsymbol {e} _ {j} ^ {q}, \\\\ 0 & \\text {o t h e r w i s e .} \\end{array} \\right.", + "image_path": "da021b075e5f93b68236193a863e28b5979f008cc1162e19f0da23ee4c3425ce.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 295, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 308 + ], + "type": "text", + "content": "further, soft-DTW finds the optimal alignment by minimizing the pre-fusion alignment loss " + }, + { + "bbox": [ + 55, + 281, + 295, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 281, + 295, + 308 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 315, + 244, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 315, + 244, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 315, + 244, + 331 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p r e}} = \\text {s o f t - D T W} _ {\\gamma} (\\mathrm {A} _ {i, j} ^ {\\text {p r e}}, \\mathrm {C} _ {i, j} ^ {\\text {p r e}}),", + "image_path": "723d1f020e34171e7b2d66d7f19f6a2ce9af8fa929257e031da7e2458467eecf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": "such that aligned frames in the target are contiguous. Here " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " is a smoothing factor for soft-min operator. The value of " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " in Soft-DTW is selected empirically to balance smoothness and alignment fidelity. Note that " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{pre}}" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " is the alignment cost matrix with elements defined by cosine similarity:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 397, + 223, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 397, + 223, + 428 + ], + "spans": [ + { + "bbox": [ + 127, + 397, + 223, + 428 + ], + "type": "interline_equation", + "content": "\\mathrm {C} _ {i, j} = 1 - \\frac {\\langle \\pmb {e} _ {i} ^ {t} , \\pmb {e} _ {j} ^ {q} \\rangle}{\\| \\pmb {e} _ {i} ^ {t} \\| \\| \\pmb {e} _ {j} ^ {q} \\|}.", + "image_path": "2543b22165be5efb8c76220043dbc41e6aa41ffa33ab4122f53aa103cb31fc7d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "type": "text", + "content": "Pre-fusion alignment enhances the semantic representation of target and query video. These enhanced representations are further processed by the encoder. The encoder outputs the fused representation of the target and query." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": "Post-fusion alignment: Given the post-fusion target and query video feature sequences, " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t^g}\\in \\mathbb{R}^{M\\times d}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{E_q^g}\\in" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": ", soft-DTW outputs a binary alignment matrix " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^{\\mathrm{post}}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": " and an alignment cost matrix " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^{\\mathrm{post}}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": ", soft-DTW achieves optimal alignment between " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t^g}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathrm{E_q^g}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": " by minimizing the postfusion alignment loss " + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}^{\\mathrm{post}}" + }, + { + "bbox": [ + 55, + 480, + 296, + 559 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 567, + 246, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 567, + 246, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 246, + 582 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} ^ {\\text {p o s t}} = \\text {s o f t - D T W} _ {\\gamma} \\left(\\mathrm {A} _ {i, j} ^ {\\text {p o s t}}, \\mathrm {C} _ {i, j} ^ {\\text {p o s t}}\\right).", + "image_path": "4cc65f86c660b0ba08cf18a6affe18e57defb78214444b91c9750dd87d71077c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 585, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 295, + 620 + ], + "type": "text", + "content": "Post-fusion alignment refines target video features by leveraging fused query-target representations, ensuring fine-grained semantic matching for precise moment localization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 620, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 295, + 669 + ], + "type": "text", + "content": "Decoder. The decoder in MATR further refines the target video representation by processing aligned query-target features, enabling fine-grained temporal matching and enhancing frame-level precision for accurate moment localization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "To this end, the post-fusion alignment matrix " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\mathrm{post}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " defines a contiguous sub-sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t^g}[s : e]" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " which aligns best with the query video. These aligned target features, capturing query-aligned fine-grained information, are passed as input to the decoder for further refinement. The decoder consists of " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " Transformer layers, each containing a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward network. The input to the decoder comprises fixed size " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " learnable query vectors denoted by " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{Q} \\in [q_1, \\ldots, q_l] \\in \\mathbb{R}^{l \\times d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ". These queries guide the extraction of refined features relevant to moment localization. As these queries are processed through each decoder layer, they are refined to capture intricate temporal dependencies, with positional encodings applied at each attention layer. The cross-attention layers enable interaction between " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t^g}[s : e]" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " from the encoder (which serve as keys and values) and representations of moment queries (which serve as queries). The final decoder output, " + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "\\mathrm{E_t^l} \\in \\mathbb{R}^{l \\times d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": ", provides a refined fine-grained set of target video features conditioned on the query video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "text", + "content": "Finally, the encoder and decoder representations are combined as " + }, + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "inline_equation", + "content": "\\mathrm{E_f} = [\\mathrm{E_t^g};\\mathrm{E_t^l}]" + }, + { + "bbox": [ + 313, + 312, + 553, + 371 + ], + "type": "text", + "content": ", making the final target video representation query-aligned. The final query-aligned representation, combining the abstract and fine-grained semantics, serves as input for the prediction heads." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": "Prediction Heads. The prediction heads consist of three " + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "inline_equation", + "content": "(c = 3)" + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": " sequential convolutional layers each having " + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "inline_equation", + "content": "1 \\times 3" + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": " kernels followed by a ReLU activation. Finally, a sigmoid activation function is applied to produce the foreground predictions " + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "inline_equation", + "content": "\\hat{f}_i" + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": ". The model is trained using the following binary cross-entropy loss " + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fg}" + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": " to distinguish foreground (relevant moment) from background predictions:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 461, + 526, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 461, + 526, + 482 + ], + "spans": [ + { + "bbox": [ + 341, + 461, + 526, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f g} = - \\left(f _ {i} \\log \\hat {f} _ {i} + (1 - f _ {i}) \\log (1 - \\hat {f} _ {i})\\right),", + "image_path": "7599b2bb8e7bb77bae79d598d04c189d2cf006d72ceee27958aa4852f7dc100e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "content": " is the ground truth label, with " + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "inline_equation", + "content": "f_{i} = 1" + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "inline_equation", + "content": "f_{i} = 0" + }, + { + "bbox": [ + 313, + 487, + 553, + 512 + ], + "type": "text", + "content": " indicating foreground and background, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": "In addition to foreground classification, the boundary prediction head is designed to estimate the start and end boundaries for the moment at every position. This head shares the same initial structure as the classification head but differs in the final output, producing two channels for left and right boundary offsets relative to the current position. Specifically, given " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{E_f}" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": ", the boundary prediction head outputs predicted offsets " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\hat{d}_i = \\{\\hat{d}_i^L, \\hat{d}_i^R\\}" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": " for each position " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "i \\in [0, M-1]" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\hat{d}_i^L" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "\\hat{d}_i^R" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": " represent the left and right boundary offsets, respectively. The boundary prediction head is trained with a combination of smooth " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": " and generalized intersection over union (IoU) loss [42], specifically applied to foreground positions " + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "inline_equation", + "content": "(f_i = 1)" + }, + { + "bbox": [ + 313, + 512, + 554, + 669 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 674, + 542, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 674, + 542, + 696 + ], + "spans": [ + { + "bbox": [ + 324, + 674, + 542, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s e g}} = \\mathbb {1} _ {f _ {i} = 1} \\left[ \\lambda_ {\\mathrm {L} 1} \\mathcal {L} _ {\\mathrm {L} 1} \\left(\\hat {d} _ {i}, d _ {i}\\right) + \\lambda_ {\\mathrm {I o U}} \\mathcal {L} _ {\\mathrm {I o U}} \\left(\\hat {d} _ {i}, d _ {i}\\right) \\right].", + "image_path": "30be9cc19d843c3eb6d79a540d5840688e2442e1f8cae85c1cef039de6989fd3.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": " denotes the ground truth offset. The parameters" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 296, + 713 + ], + "type": "text", + "content": "2Soft-DTW allows non-linear alignments between sequences of different lengths, is differentiable, robust to variations in speed, and can handle noise and outliers, though other alignment algorithms, such as TCC [12] and DropDTW [11] can also be used in our model." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20218" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 296, + 157 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 296, + 157 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 296, + 157 + ], + "type": "image", + "image_path": "6c3e3ca473497383049fe5b203719752208f104a65d7982c83ffc5a1702f27bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 165, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 165, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 295, + 222 + ], + "type": "text", + "content": "Figure 3. Proposed self-supervised pre-training strategy: Given a target video, a query clip is randomly sampled from it and processed by the model. The model then predicts the boundaries of the selected clip within the target video, highlighting the corresponding frames with a green border. (Best viewed in color)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{L1}}" + }, + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{IoU}}" + }, + { + "bbox": [ + 55, + 228, + 296, + 264 + ], + "type": "text", + "content": " weigh the contributions of the smooth L1 and IoU losses, ensuring that boundary localization focuses on the predictions identified as foreground." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 323 + ], + "type": "text", + "content": "Convolutional layers operate along the temporal axis over the concatenated encoder query-conditioned and decoder-aligned target features. This preserves temporal continuity, enabling fine-grained predictions without boundary artifacts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "text", + "content": "Overall Loss. The overall loss function " + }, + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "text", + "content": " is computed as a combination of multiple objectives, accounting for alignment costs (pre- and post-fusion alignment), foreground/background classification, and boundary localization. For " + }, + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 324, + 296, + 390 + ], + "type": "text", + "content": " training samples, this multi-objective overall loss is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 396, + 282, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 282, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 282, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{S} \\sum_ {i = 1} ^ {S} \\left(\\lambda_ {f g} \\mathcal {L} _ {f g} + \\lambda_ {\\mathrm {s e g}} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),", + "image_path": "6e41d831a7864fc941c629b6ccbb0dd00896a418169278353702d6873e402f3f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 433, + 286, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 286, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 286, + 445 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 56, + 433, + 286, + 445 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 433, + 286, + 445 + ], + "type": "text", + "content": "s control the importance of each component." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": "Inference. During inference, given a target video " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": " and a query video " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "V_{q}" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": ", we pass both through the model to generate foreground probabilities " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_i\\}_{i=0}^{M-1}" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": " and boundary predictions " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\{\\hat{d}_i\\}_{i=0}^{M-1}" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": " using the two prediction heads. To handle the densely generated boundaries, while predicting " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": ", we apply 1-dimensional non-maximal suppression with a threshold of 0.7, filtering out highly overlapping boundaries and producing the final set of predictions. We choose " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\hat{d}_i" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": " corresponding to position " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": " with highest " + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\hat{f}_i" + }, + { + "bbox": [ + 55, + 445, + 296, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 564, + 236, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 236, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 236, + 578 + ], + "type": "text", + "content": "3.3. Self-supervised Pre-training Strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "The video representation learning community has leveraged self-supervised pre-training techniques designed for training effective encoders [22, 23, 37, 38, 45, 49, 50], which can be finetuned for enhanced performance on downstream tasks. In this work, we introduce a self-supervised pretraining objective designed to improve temporal localization capabilities without relying on labeled data. Specifically, given a target video " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ", a query clip " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_{q}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is randomly sampled from " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " as shown in Fig. 3. The model is then trained to localize this query clip in " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_{t}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". This pre-training objective closely aligns with the task of Vid2VidMR. To en" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "hance generalization capabilities and encourage the model to become robust to variations in both appearance and timing, we apply one of the following random augmentations to each of the query clips: reversing frames, adding gaussian noise, slowing down or speeding up the action, thereby doubling the number of pre-training samples. These augmentations introduce temporal and spatial variations, enabling the model to learn diverse representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "text", + "content": "The overall pre-training loss, " + }, + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pt}" + }, + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "text", + "content": ", combines the foreground-background classification loss, the boundary prediction loss, pre-fusion and post-fusion alignment costs. The total loss is averaged over " + }, + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 168, + 553, + 213 + ], + "type": "text", + "content": " pre-training samples as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 223, + 548, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 548, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 548, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p t} = \\frac {1}{P} \\sum_ {i = 1} ^ {P} \\left(\\lambda_ {f g} ^ {p} \\mathcal {L} _ {\\mathrm {f g}} + \\lambda_ {s e g} ^ {p} \\mathcal {L} _ {\\mathrm {s e g}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p r e}} + \\lambda_ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}} \\mathcal {L} _ {\\mathrm {a l i g n}} ^ {\\mathrm {p o s t}}\\right),", + "image_path": "385a6c258c60351bc8ae8d8ef3e2382d7bc7a5db8a8b82d68782c2c1947c69dc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 262, + 544, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 544, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 544, + 274 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 316, + 262, + 544, + 274 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 316, + 262, + 544, + 274 + ], + "type": "text", + "content": "s control the importance of each component." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 286, + 373, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 373, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 373, + 298 + ], + "type": "text", + "content": "4. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 308, + 553, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 343 + ], + "type": "text", + "content": "We use the following two datasets to compare our proposed approach with existing Vid2VidMR methods and other strong baselines:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 344, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 554, + 452 + ], + "type": "text", + "content": "(i) ActivityNet-VRL Dataset [14]: ActivityNet-VRL is the popular benchmark dataset shared by Feng et al. [14], based on the ActivityNet video understanding benchmark [17] consisting of 200 action classes. The dataset is split into disjoint 160 classes for training and 20 classes each for the validation and test splits. Furthermore, the training set comprises of " + }, + { + "bbox": [ + 313, + 344, + 554, + 452 + ], + "type": "inline_equation", + "content": "\\sim 463\\mathrm{K}" + }, + { + "bbox": [ + 313, + 344, + 554, + 452 + ], + "type": "text", + "content": " query-target video pairs, while the validation set and the test set have 829 and 978 query-target video pairs, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 453, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 693 + ], + "type": "text", + "content": "(ii) Our proposed SportsMoments Dataset: One of the most promising applications for Vid2VidMR lies in the area of sports analytics. Although there exist large-scale sports datasets such as Sports-1M [21], they are primarily tailored for broader sports video classification tasks and do not contain fine-grained sports actions and moments, such as \"Cover Drive,\" \"Ducking a Bouncer,\" \"Goal Kick,\" or \"Penalty.\" Secondly, although ActivityNet-VRL covers a wide range of action categories (including sports actions), it only encompasses broader categories such as \"Playing Polo\" and does not focus on capturing higher-granularity events within a sport. Towards filling this gap, we introduce the SportsMoments dataset, which consists of " + }, + { + "bbox": [ + 313, + 453, + 555, + 693 + ], + "type": "inline_equation", + "content": "\\sim 770\\mathrm{K}" + }, + { + "bbox": [ + 313, + 453, + 555, + 693 + ], + "type": "text", + "content": " query-target pairs annotated from 176.6 hours of complete match footage of two of the most popular sports, viz. soccer and cricket. We obtain a total of 80 full-length cricket and soccer full-length match videos from YouTube. We then curate a list of 29 action classes comprising 13 soccer and 16 cricket actions, respectively. Given this list, we employed two annotators, each with strong knowledge of these sports," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 514, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 514, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 514, + 713 + ], + "type": "text", + "content": "3https://github.com/vl2g/MATR/tree/main/sportsmoments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20219" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "to mark the start and end timestamps for the specified actions in the videos. We split SportsMoments into training, validation and test sets. The training split consists of approximately 750K pairs spanning 16 classes, while the validation and test splits each contain 10K pairs, covering four and nine classes, respectively. Action classes are disjoint across the train, validation, and test sets, ensuring no overlap. Additionally, each split includes cricket and soccer actions for a well-rounded distribution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 205 + ], + "type": "text", + "content": "Additionally, we leverage the unlabeled videos from the Kinetics700 [6] dataset for pre-training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 198, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 198, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 198, + 230 + ], + "type": "text", + "content": "5. Experiments and Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 121, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 121, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 121, + 249 + ], + "type": "text", + "content": "5.1. Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 256, + 294, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 294, + 280 + ], + "type": "text", + "content": "We experiment with an extensive set of strong baselines grouped into four categories as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 281, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 376 + ], + "type": "text", + "content": "(i) Fully-supervised Vid2VidMR methods: We compare with Vid2VidMR methods like CGBM [14], GDP [7], SRL [51], FFI+SRM [19], SST [4] and Video-level match [14]. Further, Huo et al. [19] adapt text-VMR methods like VSLNet [61], MABAN [46] and 2D-TAN [63] by replacing their text encoders with a C3D [47] feature extractor to extract query video features. All of these models have been trained on ActivityNet-VRL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 377, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 484 + ], + "type": "text", + "content": "(ii) Vision-Language Models (VLMs): VLMs like VideoLLaMa [62], Video-LLaVa [26] and TimeChat [41] have shown promising results for multimodal text-video tasks like Video VQA, captioning, etc. Unfortunately, they have not been pre-trained with video-video aligned data. First we represent the query video using a state-of-the-art captioner, i.e., mPLUG-OWL [58]. Next, the VLMs are zero-shot prompted to generate start and end moment timestamps given query caption and target video tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 485, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 640 + ], + "type": "text", + "content": "(iii) Text-VMR methods: Moment-DETR [25], QD-DETR [33], CG-DETR [32], and UniVTG [28] have been originally proposed for text-VMR. We compare with five variants of these methods. The zero-shot variant (a) uses caption from mPLUG-OWL [58] to represent query video, and leverages the pretrained checkpoints. Variants (b) and (c) both use caption from mPLUG-OWL [58] to represent query video, but train a randomly initialized checkpoint (variant b) or finetune the pre-trained checkpoint (variant c), respectively. Variants (d) and (e) are equivalent to (b) and (c) where their CLIP text encoder is replaced by CLIP ViT/B-32 vision encoder, and therefore take query video directly as input along with the target video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "(iv) Image-VMR methods: Our work focuses on video moment retrieval using 'video' queries. As a single image query may not effectively capture the temporal aspects of a video query, this comparison may not be appropriate. However, we still design baselines by representing the video by its key-frame (more precisely, the middle frame) using" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "Text-VMR methods, namely Moment-DETR, QD-DETR and UniVTG." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "text", + "content": "We choose hidden dimension of 1024 with " + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "text", + "content": " layers in both encoders and apply a dropout rate of 0.1 and 0.5 within the transformer and linear projection layers, respectively. Model weights are initialized using Xavier initialization [16]. To optimize the model parameters, we utilize AdamW optimizer [30] with an initial learning rate of 1e-4 and a weight decay of 1e-4. Training was done on two NVIDIA A6000 GPUs. We trained our model for 200 epochs with a batch size of 1200 while using ActivityNetVRL. For SportsMoments, we used 40 epochs with a batch size of 40. We set number of learnable queries " + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "inline_equation", + "content": "l = 10" + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "text", + "content": " and all " + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\lambda s" + }, + { + "bbox": [ + 313, + 123, + 555, + 291 + ], + "type": "text", + "content": " in both pre-training as well as finetuning losses to 1. We make our implementation and checkpoints available at: https://github.com/vl2g/MATR." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 443, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 443, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 443, + 311 + ], + "type": "text", + "content": "5.3. Results and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 484 + ], + "type": "text", + "content": "Main Results: Table 1 shows our main results where we compare MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and five variants of each of the Text-VMR methods. As mentioned before these variants correspond to zero-shot, finetuned or trained setup. As usual, trained setup implies random initialization while finetuned setup implies initialization using a pre-trained checkpoint (different for each architecture). Query video can be represented using text (T) captions obtained using mPLUGOWL [58] or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions and image query is not needed. Following previous work, we report the standard mean Intersection over Union (mIoU) and Recall@1 IoU=0.5 metrics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "content": "Comparison with fully-supervised Vid2VidMR methods: For a fair comparison, we directly use the reported results from original papers for fully-supervised methods on ActivityNet-VRL. Consequently, we do not report their results on SportsMoments. As shown in Table 1, MATR outperforms the best method, FFI+SRM, with an " + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "content": " gain in mIoU and achieves a " + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 605 + ], + "type": "text", + "content": " improvement in R@1 over GDP, which had the highest recall. These results underscore MATR's superior ability to accurately localize moments using video queries." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "Comparison with VLMs: Table 1 shows that our method MATR outperforms all four VLM-based baselines by a significant margin. In comparison, the best-performing baseline, TimeChat [41], achieves " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "26.4\\mathrm{mIoU}" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "23.8\\mathrm{R}@\\mathbb{1}" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " on ActivityNet-VRL, and " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "22.6\\mathrm{mIoU}" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "inline_equation", + "content": "21.3\\mathrm{R}@\\mathbb{1}" + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": " on SportsMoments. Other methods, such as VideoLLaVA [26], Video-LLaMA [62], and Video-LLaMA2 [8], perform significantly worse. These results show that task-specific models for Vid2VidMR are significantly better." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20220" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 69, + 303, + 466 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 303, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 303, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 303, + 466 + ], + "type": "table", + "html": "
MethodsVariantsActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Fully-supervised MethodsRandom [14]-7.316.2--
Video Match [14]-12.424.3--
SST [4]-17.133.2--
CGBM [14]-25.743.5--
GDP [7]-27.844.0--
SRL [51]-40.629.3--
2D-TAN [63]-45.339.6--
MABAN [46]-42.837.5--
VSLNet [61]-27.243.8--
FFI+SRM [19]-48.740.6--
VLMsVideo-LLaVA [26]-15.114.713.811.9
Video-LLaMA [62]-14.713.912.511.2
Video-LLaMA2 [8]-17.615.214.713.4
TimeChat [41]-26.423.822.621.3
I-VMRMoment-DETR [25]-35.632.525.218.5
QD-DETR [33]-37.238.127.720.7
UniVTG [28]-38.842.034.637.2
Text-VMR MethodsMoment-DETR [25](a) ZS+T28.222.83.70.9
(b) Trained+T37.135.831.429.2
(c) Finetuned+T40.038.935.834.0
(d) Trained+V35.934.728.724.2
(e) Finetuned+V40.039.930.425.1
QD-DETR [33](a) ZS+T25.422.26.32.6
(b) Trained+T38.439.029.827.2
(c) Finetuned+T41.645.236.033.5
(d) Trained+V39.741.027.225.1
(e) Finetuned+V42.542.735.430.6
UniVTG [28](a) ZS+T32.426.711.46.0
(b) Trained+T43.845.641.536.5
(c) Finetuned+T45.846.444.839.2
(d) Trained+V48.449.843.239.5
(e) Finetuned+V49.150.743.641.8
CG-DETR [32](a) ZS+T25.921.93.31.6
(b) Trained+T39.438.435.131.3
(c) Finetuned+T41.043.235.432.5
(d) Trained+V40.041.736.635.1
(e) Finetuned+V40.141.737.234.8
OursMATRZero-shot32.630.131.830.7
Trained+V53.254.856.252.7
Finetuned+V56.857.159.256.5
", + "image_path": "7def681002fbc0e36db09c57ffd2fa3f932290fd358e4698645770a247fbe44f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 472, + 297, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 472, + 297, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 297, + 594 + ], + "type": "text", + "content": "Table 1. Comparing MATR with fully-supervised Vid2VidMR methods, VLMs, image-VMR methods, and variants of Text-VMR methods. Trained implies random initialization. Finetuned implies initialization using a pre-trained (respective to each architecture) checkpoint. Query video can be represented using text (T) captions or as video (V) itself. Since MATR is inherently designed to take video query as input, experimenting with text (T) captions is not needed. For more details of baselines and their variants, please refer to Section 5.1. Results for the overall best, best among implemented baselines and previously reported SOTA are highlighted in bold, underline and box, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "text", + "content": "Comparison with Text-VMR Methods: Table 1 shows that amongst all the five variants, the finetuned variants are typically better than the trained variants, i.e., variant (c) is better than (b), and (e) is better than (d). On both ActivityNet-VRL and SportsMoments, UniVTG achieves the best results among variants which use video queries. However, our MATR model outperforms UniVTG by " + }, + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 55, + 602, + 296, + 711 + ], + "type": "text", + "content": " in R@1 on ActivityNet-VRL. On SportsMoments, MATR shows a further improvement, sur" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 331, + 70, + 539, + 130 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 539, + 130 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 539, + 130 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 539, + 130 + ], + "type": "table", + "html": "
Pre-FusionPost-FusionActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
XX49.750.249.146.6
X✓52.352.953.148.2
✓X50.451.252.446.7
✓✓53.254.856.252.7
", + "image_path": "ee17677be81b50d3fab021483658d573663c1ccabb3be42a5f21c087aa03e790.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 326, + 166, + 544, + 223 + ], + "blocks": [ + { + "bbox": [ + 314, + 137, + 553, + 160 + ], + "lines": [ + { + "bbox": [ + 314, + 137, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 314, + 137, + 553, + 160 + ], + "type": "text", + "content": "Table 2. Advantage of explicit dual-stage alignment of MATR. For this study, experiments were performed without pre-training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 166, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 326, + 166, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 326, + 166, + 544, + 223 + ], + "type": "table", + "html": "
Pre-trainingAugmentationActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
X-53.254.856.252.7
✓X54.155.656.953.6
✓✓56.857.159.256.5
", + "image_path": "a8559d65e692b90aa43160292a74a9f1d4e9cbdee41a952c6454805a3c94a7b1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 230, + 542, + 242 + ], + "lines": [ + { + "bbox": [ + 326, + 230, + 542, + 242 + ], + "spans": [ + { + "bbox": [ + 326, + 230, + 542, + 242 + ], + "type": "text", + "content": "Table 3. Effect of pre-training and augmentation in MATR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": "passing UniVTG by " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "15.6\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "14.7\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " in R@1. These gains highlight the superior performance of MATR compared to the text-VMR methods when used with video query input. When text captions are used to represent the query, we observe similar trends as in the video query input case. Among the baselines, UniVTG performs the best. However, MATR outperforms it by " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " in mIoU and " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " in R@1 on ActivityNet-VRL. On SportsMoments, MATR achieves " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " higher mIoU and " + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "inline_equation", + "content": "17.3\\%" + }, + { + "bbox": [ + 313, + 249, + 555, + 393 + ], + "type": "text", + "content": " higher R@1 than UniVTG. These results demonstrate the superior performance of MATR compared to the text-VMR methods when used with text query input for Vid2VidMR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 492 + ], + "type": "text", + "content": "Comparison with Image-VMR: Table 1 shows that the best-performing approach, UniVTG [28], achieves a mIoU of 38.8 and 42.0 R@1 on ActivityNet-VRL and 34.6 mIoU and R@1 37.2 on SportsMoments. Although this is comparable to previous fully supervised methods, it still lags significantly behind our MATR. This further goes to show the importance of temporal dynamics captured by video queries as compared to static image queries." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 686 + ], + "type": "text", + "content": "Advantage of Dual Sequence Alignment: Table 2 shows the impact of different alignment strategies in MATR on both datasets. This study is done without any pre-training. We evaluate the effect of incorporating alignment at the pre-fusion and post-fusion stages. Our results indicate that the combined use of both pre-fusion and post-fusion alignment achieves the best performance, suggesting that alignment is important to ensure both before and after the encoder. Models with alignment either before (pre) or after (post) show moderate gains over neither. Removing post-fusion alignment hurts more than removing pre-fusion alignment, suggesting that post-fusion alignment is more important. Omitting alignment entirely results in the lowest scores. This highlights the importance of dual alignment for effectively capturing relevant temporal information for video query-based moment localization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Effect of Pre-training: We analyze the impact of pretraining and data augmentation on model performance in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20221" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 72, + 220, + 193 + ], + "blocks": [ + { + "bbox": [ + 77, + 72, + 220, + 193 + ], + "lines": [ + { + "bbox": [ + 77, + 72, + 220, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 220, + 193 + ], + "type": "image", + "image_path": "2f34237d23fd96a08df298a0fd06983b207b140d2909476fa41ddd4e6c5e1cfa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 206, + 555, + 240 + ], + "type": "text", + "content": "Figure 4. Visualization of Vid2VidMR on a sample from ActivityNet-VRL for calf roping action. Our proposed MATR model shows improved generalization capabilities over the best-performing baseline methods (variant (e), i.e., finetuned+V). Start and end times for the ground truth and predicted moments are shown in the brackets. Refer to supplementary material for the videos used in this figure." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 533, + 194 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 533, + 194 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 533, + 194 + ], + "type": "image", + "image_path": "4a461c2ccd27c007ebe387383b266007d7fdb85eb2701c8b2c6968741cbd4da5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 245, + 282, + 304 + ], + "blocks": [ + { + "bbox": [ + 69, + 245, + 282, + 304 + ], + "lines": [ + { + "bbox": [ + 69, + 245, + 282, + 304 + ], + "spans": [ + { + "bbox": [ + 69, + 245, + 282, + 304 + ], + "type": "table", + "html": "
ActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
Pre-fusion alignment (Apre)34.132.631.629.8
Post-fusion alignment (Apost)38.335.236.433.5
Prediction-heads56.857.159.256.5
", + "image_path": "f0031de5ad333789d20a845489a9d2209e1647a27e799776293004eba2794a27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 94, + 351, + 258, + 416 + ], + "blocks": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 346 + ], + "type": "text", + "content": "Table 4. Advantage of predicting using prediction heads on decoder. We observe that directly predicting from fusion matrices is inferior as it lacks fine-grained refinement." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 351, + 258, + 416 + ], + "lines": [ + { + "bbox": [ + 94, + 351, + 258, + 416 + ], + "spans": [ + { + "bbox": [ + 94, + 351, + 258, + 416 + ], + "type": "table", + "html": "
AlignmentActivityNet-VRLSportsMoments
mIoUR@1mIoUR@1
None50.451.852.148.6
TCC52.452.753.952.8
Drop-DTW54.254.755.654.7
soft-DTW56.857.159.256.5
", + "image_path": "b320b5ebc8b61ee559fdc7318bf7e56ee447bc5917d4862b8cab22154bc7160d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 424, + 251, + 435 + ], + "lines": [ + { + "bbox": [ + 99, + 424, + 251, + 435 + ], + "spans": [ + { + "bbox": [ + 99, + 424, + 251, + 435 + ], + "type": "text", + "content": "Table 5. Ablation on alignment methods." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": "Table 3. On ActivityNet-VRL, without pre-training, the model achieves " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "53.2\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "54.8\\mathrm{R}@\\mathrm{l}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": ". Adding pretraining without augmentation yields " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "54.1\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "55.6\\mathrm{R}@\\mathrm{l}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": ". The best results are obtained by using pre-training with augmentation, leading to a significant boost to " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "56.8\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "inline_equation", + "content": "57.1\\mathrm{R}@\\mathrm{l}" + }, + { + "bbox": [ + 55, + 443, + 296, + 527 + ], + "type": "text", + "content": ". We got similar observations for the SportsMoments dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "text", + "content": "Predictions from Heads vs Alignment Matrices: The results in Table 4 show that prediction from heads outperforms the prediction from both pre-fusion alignment " + }, + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^{\\mathrm{pre}})" + }, + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "text", + "content": " and post-fusion alignment " + }, + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^{\\mathrm{post}})" + }, + { + "bbox": [ + 55, + 530, + 295, + 650 + ], + "type": "text", + "content": " on both datasets. Notably, prediction heads improve mIoU and R@1 by a large margin, demonstrating their effectiveness in accurately retrieving temporal moments. Post-fusion alignment provides better results than pre-fusion alignment. However, the gains are modest compared to the substantial boost from prediction heads. This justifies the need for the decoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Ablation on Alignment Losses: MATR can incorporate different alignment methods. Table 5 compares the performance of different alignment methods on both ActivityNet-VRL and SportsMoments datasets. Without any alignment (\"None\"), the model achieves R@1 scores of " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "51.8\\%" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": "on ActivityNet-VRL and " + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "inline_equation", + "content": "48.6\\%" + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": " on SportsMoments. TCC alignment yields an improvement of " + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": " on ActivityNet-VRL and SportsMoments, respectively. Drop-DTW is better than TCC. soft-DTW alignment consistently demonstrates superior performance across both datasets, achieving the highest gains of " + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": " on ActivityNet-VRL and " + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 313, + 247, + 555, + 332 + ], + "type": "text", + "content": " on SportsMoments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 556, + 477 + ], + "type": "text", + "content": "Qualitative Results: We present moment retrieval results on a sample from ActivityNet-VRL in Fig. 4, where we compare our proposed approach to variant (e) (refer Sec. 5.1 for details) of the four best-performing baselines. The video query is a 5-second clip depicting calf roping4. The target video, with a duration of 24 seconds, showcases this event occurring between 2.1 and 18.1 seconds. Our proposed MATR model exhibits the highest overlap with the ground truth for the given query, accurately identifying the correct temporal boundaries. Among the baseline methods, CG-DETR and UniVTG achieve strong overlap, however, are still outperformed by our approach." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 499 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 639 + ], + "type": "text", + "content": "We introduced MATR, a robust approach for Vid2VidMR that combines abstract representation from encoder with fine-grained features from decoder conditioned on aligned query features. It captures semantic and temporal cues for precise moment localization. Our self-supervised pretraining enhances initialization and boosts performance. Extensive experiments on ActivityNet-VRL and our new SportsMoments dataset show that MATR outperforms strong baselines. Future directions of this work include exploring multimodal queries and developing scalable architectures to enhance broader applicability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 639, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 676 + ], + "type": "text", + "content": "Acknowledgment: This work is supported by the Microsoft Academic Partnership Grant (MAPG) 2023. Yogesh Kumar is supported by a UGC fellowship, Govt. of India." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "4A rodeo event where a rider on horseback attempts to catch and tie a calf within a timed competition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20222" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 62, + 91, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 91, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 62, + 91, + 294, + 138 + ], + "type": "text", + "content": "[1] U. Agarwal, Y. Kumar, A. Shahid, P. Gatti, M. Gupta, and A. Mishra. CHAPVIDMR: Chapter-based video moment retrieval using natural language queries. In ICVGIP, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 140, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 140, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 140, + 294, + 175 + ], + "type": "text", + "content": "[2] D. M. Argaw, J.-Y. Lee, M. Woodson, I.-S. Kweon, and F. C. Heilbron. Long-range multimodal pretraining for movie understanding. ICCV, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 177, + 294, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 177, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 63, + 177, + 294, + 212 + ], + "type": "text", + "content": "[3] I. Bello, B. Zoph, Q. Le, A. Vaswani, and J. Schlens. Attention augmented convolutional networks. In ICCV, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 214, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 214, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 63, + 214, + 294, + 248 + ], + "type": "text", + "content": "[4] S. Buch, V. Escorcia, C. Shen, B. Ghanem, and J. C. Niebles. SST: single-stream temporal action proposals. In CVPR, 2017. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 251, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 251, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 63, + 251, + 294, + 285 + ], + "type": "text", + "content": "[5] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 287, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 287, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 63, + 287, + 294, + 322 + ], + "type": "text", + "content": "[6] J. Carreira and A. Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 323, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 323, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 63, + 323, + 294, + 358 + ], + "type": "text", + "content": "[7] L. Chen, C. Lu, S. Tang, J. Xiao, D. Zhang, C. Tan, and X. Li. Rethinking the bottom-up framework for query-based video localization. In AAAI, 2020. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 419 + ], + "type": "text", + "content": "[8] Z. Cheng, S. Leng, H. Zhang, Y. Xin, X. Li, G. Chen, Y. Zhu, W. Zhang, Z. Luo, D. Zhao, and L. Bing. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 421, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 421, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 63, + 421, + 294, + 444 + ], + "type": "text", + "content": "[9] M. Cuturi and M. Blondel. Soft-DTW: a differentiable loss function for time-series. In ICML, 2017. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 446, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 446, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 294, + 480 + ], + "type": "text", + "content": "[10] P. Dogan, B. Li, L. Sigal, and M. Gross. A neural multi-sequence alignment technique (neumatch). In CVPR, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 483, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 483, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 483, + 294, + 529 + ], + "type": "text", + "content": "[11] N. Dvornik, I. Hadji, K. G. Derpanis, A. Garg, and A. D. Jepson. Drop-DTW: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 531, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 294, + 566 + ], + "type": "text", + "content": "[12] D. Dwibedi, Y. Aytar, J. Tompson, P. Sermanet, and A. Zisserman. Temporal cycle-consistency learning. In CVPR, 2019. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 568, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 602 + ], + "type": "text", + "content": "[13] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slow-fast networks for video recognition. In ICCV, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 605, + 294, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 605, + 294, + 627 + ], + "spans": [ + { + "bbox": [ + 57, + 605, + 294, + 627 + ], + "type": "text", + "content": "[14] Y. Feng, L. Ma, W. Liu, T. Zhang, and J. Luo. Video re-localization. In ECCV, 2018. 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 629, + 294, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 629, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 57, + 629, + 294, + 664 + ], + "type": "text", + "content": "[15] J. Gao, C. Sun, Z. Yang, and R. Nevatia. TALL: temporal activity localization via language query. In ICCV, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 666, + 294, + 712 + ], + "type": "text", + "content": "[16] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In International Conference on Artificial Intelligence and Statistics, 2010. 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 108 + ], + "type": "text", + "content": "[17] F. C. Heilbron, V. Escorcia, B. Ghanem, and J. C. Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In CVPR, 2015. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 110, + 553, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 110, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 110, + 553, + 144 + ], + "type": "text", + "content": "[18] L. A. Hendricks, O. Wang, E. Shechtman, J. Sivic, T. Darrell, and B. C. Russell. Localizing moments in video with natural language. In ICCV, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 146, + 553, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 146, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 146, + 553, + 191 + ], + "type": "text", + "content": "[19] S. Huo, Y. Zhou, R. Wang, W. Xiang, and S. Kung. Semantic relevance learning for video-query based video moment retrieval. IEEE Trans. Multim., 25:9290-9301, 2023. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 194, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 194, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 194, + 553, + 228 + ], + "type": "text", + "content": "[20] M. Jung, Y. Jang, S. Choi, J. Kim, J.-H. Kim, and B.-T. Zhang. Background-aware moment detection for video moment retrieval. In WACV, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 230, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 276 + ], + "type": "text", + "content": "[21] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 278, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 278, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 317, + 278, + 553, + 325 + ], + "type": "text", + "content": "[22] Y. Kumar, S. Mallick, A. Mishra, S. Rasipuram, A. Maitra, and R. Ramnani. Qdetrv: Query-guided detr for one-shot object localization in videos. In AAAI, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 327, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 361 + ], + "type": "text", + "content": "[23] H.-Y. Lee, J.-B. Huang, M. Singh, and M.-H. Yang. Unsupervised representation learning by sorting sequences. In ICCV, 2017. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 363, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 363, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 363, + 553, + 397 + ], + "type": "text", + "content": "[24] J. Lei, L. Yu, T. L. Berg, and M. Bansal. TVR: A large-scale dataset for video-subtitle moment retrieval. In ECCV, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 399, + 553, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 434 + ], + "type": "text", + "content": "[25] J. Lei, T. L. Berg, and M. Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 436, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 436, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 317, + 436, + 553, + 471 + ], + "type": "text", + "content": "[26] B. Lin, B. Zhu, Y. Ye, M. Ning, P. Jin, and L. Yuan. Video-llava: Learning united visual representation by alignment before projection. In EMNLP, 2024. 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 472, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 506 + ], + "type": "text", + "content": "[27] J. Lin, C. Gan, and S. Han. TSM: temporal shift module for efficient video understanding. In ICCV, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 508, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 508, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 508, + 553, + 555 + ], + "type": "text", + "content": "[28] K. Q. Lin, P. Zhang, J. Chen, S. Pramanick, D. Gao, A. J. Wang, R. Yan, and M. Z. Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 557, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 603 + ], + "type": "text", + "content": "[29] Y. Liu, S. Li, Y. Wu, C. W. Chen, Y. Shan, and X. Qie. UMT: unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 605, + 553, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 605, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 317, + 605, + 553, + 628 + ], + "type": "text", + "content": "[30] I. Loshchilov and F. Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 630, + 553, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 630, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 317, + 630, + 553, + 665 + ], + "type": "text", + "content": "[31] N. C. Mithun, S. Paul, and A. K. Roy-Chowdhury. Weakly supervised video moment retrieval from text queries. In CVPR, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 666, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 666, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 666, + 553, + 712 + ], + "type": "text", + "content": "[32] W. Moon, S. Hyun, S. B. Lee, and J. Heo. Correlation-guided query-dependency calibration in video representation learning for temporal grounding. CoRR, abs/2311.08835, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20223" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 120 + ], + "type": "text", + "content": "[33] W. Moon, S. Hyun, S. Park, D. Park, and J. Heo. Query - dependent video representation for moment retrieval and highlight detection. In CVPR, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 156 + ], + "type": "text", + "content": "[34] J. Mun, M. Cho, , and B. Han. Local-Global Video-Text Interactions for Temporal Grounding. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 158, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 295, + 194 + ], + "type": "text", + "content": "[35] F. Padua, R. Carceroni, G. Santos, and K. Kutulakos. Linear sequence-to-sequence alignment. IEEE Trans. Pattern Anal. Mach. Intell., 32:304-320, 2008. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 196, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 295, + 231 + ], + "type": "text", + "content": "[36] N. Parmar, A. Vaswani, J. Uszkoreit, L. Kaiser, N. Shazeer, A. Ku, and D. Tran. Image transformer. In ICML, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 269 + ], + "type": "text", + "content": "[37] A. Piergiovanni, A. Angelova, and M. S. Ryoo. Evolving losses for unsupervised video representation learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 270, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 295, + 306 + ], + "type": "text", + "content": "[38] R. Qian, T. Meng, B. Gong, M.-H. Yang, H. Wang, S. Belongie, and Y. Cui. Spatiotemporal contrastive video representation learning. In CVPR, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 307, + 295, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 295, + 366 + ], + "type": "text", + "content": "[39] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 368, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 368, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 368, + 295, + 392 + ], + "type": "text", + "content": "[40] Rao, Gritai, and Shah. View-invariant alignment and matching of video sequences. In ICCV, 2003. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 430 + ], + "type": "text", + "content": "[41] S. Ren, L. Yao, S. Li, X. Sun, and L. Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 431, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 295, + 479 + ], + "type": "text", + "content": "[42] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 480, + 295, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 480, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 480, + 295, + 515 + ], + "type": "text", + "content": "[43] Z. Shou, D. Wang, and S.-F. Chang. Temporal action localization in untrimmed videos via multi-stage cnns. In CVPR, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 517, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 295, + 552 + ], + "type": "text", + "content": "[44] K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos. In NuerIPS, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 295, + 589 + ], + "type": "text", + "content": "[45] N. Srivastava, E. Mansimov, and R. Salakhudinov. Unsupervised learning of video representations using lstms. In ICML, 2015. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 295, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 639 + ], + "type": "text", + "content": "[46] X. Sun, H. Wang, and B. He. Maban: Multi-agent boundary-aware network for natural language moment retrieval. IEEE Trans. on Image Processing, 30:5589-5599, 2021. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 640, + 295, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 295, + 676 + ], + "type": "text", + "content": "[47] D. Tran, L. D. Bourdev, R. Fergus, L. Torresani, and M. Paluri. Learning spatiotemporal features with 3d convolutional networks. In ICCV, 2015. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 713 + ], + "type": "text", + "content": "[48] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 108 + ], + "type": "text", + "content": "[49] G. Wang, Y. Zhou, C. Luo, W. Xie, W. Zeng, and Z. Xiong. Unsupervised visual representation learning by tracking patches in video. In CVPR, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 110, + 554, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 110, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 110, + 554, + 156 + ], + "type": "text", + "content": "[50] J. Wang, Y. Gao, K. Li, J. Hu, X. Jiang, X. Guo, R. Ji, and X. Sun. Enhancing unsupervised video representation learning by decoupling the scene and the motion. In AAAI, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 159, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 159, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 159, + 554, + 194 + ], + "type": "text", + "content": "[51] R. Wang and Y. Zhou. A feature pair fusion and hierarchical learning framework for video re-localization. In ICIP, 2020. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 196, + 554, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 196, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 554, + 231 + ], + "type": "text", + "content": "[52] X. Wang, L. Zhu, and Y. Yang. T2vlad: global-local sequence alignment for text-video retrieval. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 233, + 554, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 233, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 315, + 233, + 554, + 280 + ], + "type": "text", + "content": "[53] Y. Xiao, Z. Luo, Y. Liu, Y. Ma, H. Bian, Y. Ji, Y. Yang, and X. Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 282, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 282, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 315, + 282, + 554, + 329 + ], + "type": "text", + "content": "[54] S. Xie, C. Sun, J. Huang, Z. Tu, and K. Murphy. Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification. In ECCV, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 331, + 554, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 331, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 315, + 331, + 554, + 378 + ], + "type": "text", + "content": "[55] S. Yan, X. Xiong, A. Nagrani, A. Arnab, Z. Wang, W. Ge, D. Ross, and C. Schmid. Unloc: A unified framework for video localization tasks. In ICCV, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 380, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 380, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 380, + 554, + 415 + ], + "type": "text", + "content": "[56] C. Yang, Y. Xu, J. Shi, B. Dai, and B. Zhou. Temporal pyramid network for action recognition. In CVPR, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 418, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 554, + 464 + ], + "type": "text", + "content": "[57] J. Yang, P. Wei, H. Li, and Z. Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 468, + 554, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 554, + 515 + ], + "type": "text", + "content": "[58] Q. Ye, H. Xu, J. Ye, M. Yan, A. Hu, H. Liu, Q. Qian, J. Zhang, F. Huang, and J. Zhou. mplug-owi2: Revolutionizing multi-modal large language model with modality collaboration. In CVPR, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 517, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 517, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 315, + 517, + 554, + 552 + ], + "type": "text", + "content": "[59] R. Zeng, W. Huang, M. Tan, Y. Rong, P. Zhao, J. Huang, and C. Gan. Graph convolutional networks for temporal action localization. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 554, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 554, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 315, + 554, + 554, + 589 + ], + "type": "text", + "content": "[60] C. Zhang, A. Gupta, and A. Zisserman. Temporal query networks for fine-grained video understanding. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 591, + 554, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 591, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 315, + 591, + 554, + 639 + ], + "type": "text", + "content": "[61] H. Zhang, A. Sun, W. Jing, L. Zhen, J. T. Zhou, and R. S. M. Goh. Natural language video localization: A revisit in span-based question answering framework. IEEE Trans. Pattern Anal. Mach. Intell., 2022. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 640, + 554, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 640, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 315, + 640, + 554, + 676 + ], + "type": "text", + "content": "[62] H. Zhang, X. Li, and L. Bing. Video-LLaMA: An instruction-tuned audio-visual language model for video understanding. In EMNLP, 2023. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 677, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 677, + 554, + 713 + ], + "type": "text", + "content": "[63] S. Zhang, H. Peng, J. Fu, and J. Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, 2020. 6, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20224" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 107 + ], + "type": "text", + "content": "[64] Z. Zhu, W. Tang, L. Wang, N. Zheng, and G. Hua. Enriching local and global contexts for temporal action localization. In ICCV, 2021. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20225" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_content_list.json b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d58fa4b15abed0098faf855c844eb6e6e11f1e73 --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_content_list.json @@ -0,0 +1,1754 @@ +[ + { + "type": "text", + "text": "Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning", + "text_level": 1, + "bbox": [ + 106, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junming Liu $^{1,2}$ , Siyuan Meng $^{2,3}$ , Yanting Gao $^{1}$ , Song Mao $^{2}$ , Pinlong Cai $^{2}$ ,", + "bbox": [ + 204, + 202, + 792, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guohang Yan $^{2}$ , Yirong Chen $^{2,4}$ , Zilin Bian $^{5}$ , Ding Wang $^{2*}$ , Botian Shi $^{2}$", + "bbox": [ + 222, + 220, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University $^{2}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 243, + 239, + 754, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ East China Normal University $^{4}$ Stanford University $^{5}$ New York University", + "bbox": [ + 187, + 256, + 808, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liu_junming6917@tongji.edu.cn wangding@pjlab.org.cn", + "bbox": [ + 264, + 277, + 720, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal reasoning in Large Language Models (LLMs) struggles with incomplete knowledge and hallucination artifacts, challenges that textual Knowledge Graphs (KGs) only partially mitigate due to their modality isolation. While Multimodal Knowledge Graphs (MMKGs) promise enhanced cross-modal understanding, their practical construction is impeded by semantic narrowness of manual text annotations and inherent noise in visual-semantic entity linkages. In this paper, we propose Vision-align-to-Language integrated Knowledge Graph (VaLiK), a novel approach for constructing MMKGs that enhances LLMs reasoning through cross-modal information supplementation. Specifically, we cascade pre-trained Vision-Language Models (VLMs) to align image features with text, transforming them into descriptions that encapsulate image-specific information. Furthermore, we developed a cross-modal similarity verification mechanism to quantify semantic consistency, effectively filtering out noise introduced during feature alignment. Even without manually annotated image captions, the refined descriptions alone suffice to construct the MMKG. Compared to conventional MMKGs construction paradigms, our approach achieves substantial storage efficiency gains while maintaining direct entity-to-image linkage capability. Experimental results on multimodal reasoning tasks demonstrate that LLMs augmented with VaLiK outperform previous state-of-the-art models. Our code is published at https://github.com/Wings-Of-Disaster/VaLiK.", + "bbox": [ + 88, + 357, + 485, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Large Language Models (LLMs) [2, 9, 25, 65] have demonstrated their superiority and versatility across various Natural Language Reasoning (NLR) tasks [8, 43, 53, 58]. To enhance LLMs into the", + "bbox": [ + 89, + 818, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dfacec5607ec5e36626acfb0e0ac18a11bebb24a4130fb85486e59c4e58c6e72.jpg", + "image_caption": [ + "Figure 1. (a) Training entity extraction models relies on extensive fine-grained annotations, increasing labeling costs. More examples are provided in Appendix B. (b) Capturing implicit semantic associations demands abstract comprehension or logical inference." + ], + "image_footnote": [], + "bbox": [ + 519, + 328, + 901, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "realm of multimodal reasoning, researchers [64, 71, 74, 79] have endeavored to equip these models with multimodal capabilities, as evidenced by advancements in Multimodal Large Language Models (MLLMs) such as BLIP-2 [40], GPT-4o [32], Janus-Pro [13], among others. Despite their notable progress, these models often experience hallucinations [5, 34], primarily arising from knowledge deficiencies due to incomplete or obsolete information.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "981", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fine-tuning LLMs demands prohibitive computational costs [31]. While text-based Knowledge Graphs (KGs) have partially addressed this limitation by efficient real-time updates [6, 62, 72], they are still restricted by modal isolation, which hinders cross-modal reasoning, as detailed in Appendix A. To bridge this semantic fragmentation, Multimodal Knowledge Graphs (MMKGs) have been developed as unified representational frameworks [10, 33, 38, 45].", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, constructing robust MMKGs faces two primary obstacles [15, 89]. First, the lack of large-scale fine-grained entity-image corpora makes it infeasible to train high-quality entity extractors, significantly constraining scalability, as illustrated in Figure 1a. Second, conventional visual relation detectors primarily identify superficial spatial interactions instead of semantic relations consistent with KGs, while frequently hallucinating implausible connections that corrupt graph integrity, as shown in Figure 1b.", + "bbox": [ + 89, + 215, + 480, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose VaLiK, short for Vision-align-to-Language integrated Knowledge Graph, a novel framework designed to empower LLMs with advanced multimodal reasoning. Unlike traditional methods that rely on text annotations for training extraction models and the knowledge construction process [54], VaLiK adopts a annotation-free approach to MMKGs construction. Specifically, we first employ several pretrained Vision-Language models (VLMs), designed based on Chain-of-Experts (CoE) principles [73], to convert visual inputs into image-specific textual descriptions through cross-modal feature alignment. This procedure eliminates the need for manually annotated image captions in both the knowledge extraction and construction phases while preserving visual details typically missing in generic text descriptions. Moreover, in contrast to existing relation detection methods that require predefined label taxonomies [16, 60, 81, 84], VaLiK excels at extracting profound semantic relationships that are both KG-compatible and capture novel associations beyond training supervision. While VLMs enable cross-modal reasoning and interpretation, they introduce spurious relational noise through hallucinated inter-modal attributions, as depicted in Figure 2. We address this limitation through cross-modal similarity recalibration, strategically filtering inconsistent information while preserving valid semantic correspondences. Finally, the purified descriptions are systematically organized into MMKGs via LLM-driven symbolic structuring [27], bridging visual and textual domains with factual consistency.", + "bbox": [ + 89, + 354, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To thoroughly evaluate the VaLiK method, we conduct a comprehensive assessment across two critical multimodal benchmarks: multimodal classification (tested on the CrisisMMD dataset [3]) and multimodal question answering (evaluated via the ScienceQA benchmark [47]). The experiments span diverse LLM architectures and MMKG construction techniques to ensure the framework's robustness.", + "bbox": [ + 89, + 795, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/70a5e7b33cf2f7b408d7e9bbab0216de573f5ba53a84b04153ef68002435eb2d.jpg", + "image_caption": [ + "Figure 2. Feature-aligned descriptions from VLMs introduce redundant and inaccurate relationship patterns." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 901, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The experimental results demonstrate that the MMKGs constructed by VaLiK achieve superior multimodal reasoning performance in LLMs while requiring substantially less storage than conventional approaches. More importantly, the proposed approach retains direct entity-to-image linkage capabilities even with the compressed graph structure.", + "bbox": [ + 511, + 316, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, VaLiK is the first framework that enables end-to-end, annotation-free, zero-shot, and storage-efficient multimodal knowledge construction with high adaptability and scalability. Our key contributions include:", + "bbox": [ + 511, + 409, + 903, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, VaLik is the first end-to-end framework to build Annotation-Free MMKGs to improve LLMs' multimodal reasoning capabilities, effectively eliminating the need for manually annotated textual material and enabling a completely autonomous multimodal knowledge generation process.", + "- We offer an innovative zero-shot method for constructing MMKG that captures deep semantic connections beyond traditional predetermined labels with an effective verification system that guarantees the accuracy of these relationships. The knowledge distillation paradigm greatly decreases storage while maintaining semantic integrity.", + "- We develop a highly modular and extensible architecture that allows VaLiK to effortlessly incorporate new models and workflows for specialized domain tasks, facilitating rapid adaptation to diverse application scenarios without incurring expensive system changes." + ], + "bbox": [ + 513, + 470, + 903, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 744, + 653, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multimodal Knowledge Graphs", + "text_level": 1, + "bbox": [ + 511, + 771, + 792, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The principal advantage of MMKGs resides in their multimodal integration beyond conventional KGs. By linking entities with corresponding visual or textual data, MMKGs introduce valuable visual and textual information to the knowledge base, substantially advancing multimodal reasoning capabilities. This combination addresses core challenges in tasks that inherently demand multimodal synergy", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "982", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "like autonomous driving [26, 28], image-text retrieval [23, 86] and robotic manipulation [51, 57]. However, constructing trustworthy MMKGs with minimal manual effort remains a critical challenge. Recent studies have proposed innovative strategies to enhance MMKG reliability and utility. For instance, Chen et al. [12] proposed MSPT, a framework addressing continual MMKG construction through gradient modulation for balanced multimodal learning and attention distillation to mitigate catastrophic forgetting. Song et al. [60] developed Scene-MMKG, integrating knowledge engineering with large language models to improve robotic manipulation by resolving data sparsity and knowledge uncertainty. Wang et al. [69] introduced TIVA-KG, the first quad-modal knowledge graph spanning text, image, video, and audio with triplet grounding, empirically validating its effectiveness in downstream tasks. While these advances enhance multimodal reasoning capabilities, their efficacy remains rooted in resource-intensive paradigms, requiring extensively annotated datasets for knowledge acquisition.", + "bbox": [ + 89, + 90, + 483, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Knowledge-Augmented Multimodal Learning", + "text_level": 1, + "bbox": [ + 89, + 388, + 478, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal learning has seen significant progress in aligning and integrating information across different data modalities [7, 44, 75]. The incorporation of structured knowledge through MMKGs further enhances these approaches, improving the reasoning capabilities and generalization across a variety of domains, such as visual question answering [50, 59, 67], recommendation systems [17, 61, 70], and classification [30, 55, 83]. Methods like GraphAdapter's dual-KG adaptation [41] and contrastive multi-relational encoding with KGs [22] inject external knowledge into models, refining their performance and improving their capability to handle complex tasks. Additionally, Lee et al. [38] proposed MR-MKG, a novel framework that constructs task-specific MMKGs to enhance multimodal reasoning in LLMs. These knowledge-augmented paradigms demonstrate superior cross-modal semantic grounding compared to unimodal approaches [14, 35]. However, their reliance on preconstructed MMKGs often leads to domain discrepancies, where generic knowledge schemas misalign with task-specific reasoning patterns, ultimately limiting contextual precision in target applications.", + "bbox": [ + 89, + 412, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 741, + 410, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The limitations of text-only LLMs in meeting increasingly complex demands have spurred extensive research [78, 82, 85] into developing LLMs capable of effectively processing and reasoning over multimodal inputs. Current research predominantly employs adapter or projection layers to connect the embedding spaces of various modality-specific encoders with the textual embedding space of LLMs [38]. For instance, foundational models like CLIP [56] and BLIP [39] pioneered cross-modal alignment by jointly training vision", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and text encoders to map images and text into a shared embedding space. Building on this, LLaVA [42] and Flamingo [4] advanced the field by integrating visual encoders with LLMs, enabling more nuanced multimodal understanding and generation. More recently, Gemini [63], Qwen2-VL [68] and GPT-4o [32] have further pushed the boundaries by scaling up multimodal pretraining and introducing sophisticated mechanisms for cross-modal interaction. However, multimodal LLMs remain prone to hallucinations. While they enhance cross-modal alignment, they neither acquire new knowledge nor avoid introducing noise through integration. To address these limitations, VaLiK \"uses the master's tools to refine the master's craft,\" first constructing MMKGs via MLLMs and then leveraging them to enhance MLLMs' reasoning capabilities.", + "bbox": [ + 511, + 90, + 906, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 332, + 604, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the technical details of VaLiK. VaLiK introduces a novel expansion-reduction paradigm for visual knowledge extraction. The architecture initially organizes several VLMs with distinct knowledge domains, designed based on CoE principles [73], to produce comprehensive textual descriptions encompassing hierarchical visual details. A cross-modal similarity verification mechanism then iteratively filters out noisy tokens through cross-modal alignment while preserving semantically salient elements. This optimization-style approach eliminates external textual dependencies while enabling effective MMKG construction. VaLiK's framework is shown in Figure 3.", + "bbox": [ + 511, + 357, + 908, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. CoE-based Visual to Language Modeling", + "text_level": 1, + "bbox": [ + 511, + 549, + 864, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent entity detection techniques [19, 80, 90] have been widely adopted for entity and relation extraction in MMKG construction. However, these methods are inherently limited by predefined categorical boundaries, lacking the capacity to recognize visual concepts outside their training vocabulary. In contrast, VLMs pretrained on web-scale corpora [11, 40, 88] exhibit broader recognition capabilities through exposure to diverse visual concepts.", + "bbox": [ + 511, + 571, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We therefore leverage pretrained VLMs to extract comprehensive visual information. This process removes the necessity for detailed fine-grained data typically required to train specialized recognition models. The generalized vision to language conversion pipeline can be formalized as:", + "bbox": [ + 511, + 691, + 906, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = \\mathcal {D} _ {\\text {t e x t}} \\left(\\mathcal {A} \\left(\\mathcal {E} _ {\\text {v i s}} (I)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 777, + 906, + 804 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ denotes for the input image, $\\mathcal{E}_{\\mathrm{vis}}$ denotes the visual encoder extracting visual features, $\\mathcal{A}$ carries out cross-modal feature alignment and interaction, and $\\mathcal{D}_{\\mathrm{text}}$ generates textual tokens through autoregressive decoding. The resulting visual description $S = \\{w_{1},\\dots,w_{n}\\}$ emerges from this multi-stage processing.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "983", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f146c2291d18be7bbdd2100fa3e32b5e7a21e72aea45eeb431f07813670ed63c.jpg", + "image_caption": [ + "Figure 3. The pipeline of VaLiK: First, large-scale visual descriptions are generated using CoE-based VLMs. Then, a similarity verification mechanism is used to prune irrelevant information. Finally, MMKGs are constructed using LLMs based on LightRAG. The constructed MMKGs can assist LLMs in multimodal reasoning, alleviating the hallucination issues caused by incomplete knowledge." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 300 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, quantitative analysis uncovers considerable discrepancies between machine-generated and human-annotated descriptions [87]. As an illustration, while utilizing BLIP-2 [40] to generate sample captions, we noted that the model outputs are markedly concise and devoid of visual specifics, as detailed in Appendix C. To bridge this gap, we implement CoE enhanced generation through cascade VLMs processing. At iteration step $t$ , each expert $E_{i}$ receives both the original visual signals $I$ and the contextual output from the preceding expert $E_{i - 1}$ :", + "bbox": [ + 88, + 363, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} ^ {(t)} = E _ {i} \\left(I, \\mathcal {S} _ {i - 1} ^ {(t - 1)}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 526, + 482, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{i - 1}^{(t - 1)}$ denotes the description from expert $E_{i - 1}$ at step $t - 1$ , with $S_0^{(t)}\\coloneqq \\emptyset$ for initialization.", + "bbox": [ + 89, + 565, + 482, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, each expert $E_{i}$ implements a unified visual-language processing task:", + "bbox": [ + 89, + 601, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. Visual Feature Extraction:", + "text_level": 1, + "bbox": [ + 89, + 635, + 302, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {i} = \\operatorname {E n c} _ {\\text {v i s}} ^ {i} (I) \\in \\mathbb {R} ^ {d _ {v} \\times N _ {p}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 661, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathsf{Enc}_{\\mathrm{vis}}^i$ denotes established visual encoder [20, 29, 46] producing $N_{p}$ patch embeddings with dimension $d_v$ .", + "bbox": [ + 109, + 691, + 482, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Cross-Modal Interaction and Generation:", + "text_level": 1, + "bbox": [ + 89, + 724, + 406, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLMs integrate pretrained learnable query embeddings $\\mathbf{Q}_i\\in \\mathbb{R}^{d_q\\times L_q}$ to interact with visual features $\\mathbf{V}_i\\in$ $\\mathbb{R}^{d_v\\times N_p}$ via cross-attention [66]:", + "bbox": [ + 109, + 739, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {H} _ {i} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {Q} _ {i}, \\mathbf {V} _ {i}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {W} _ {q} ^ {i} \\left(\\mathbf {V} _ {i} \\mathbf {W} _ {k} ^ {i}\\right) ^ {\\top}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V} _ {i} \\mathbf {W} _ {v} ^ {i}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 796, + 482, + 858 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_q^i\\in \\mathbb{R}^{d_q\\times d_k}$ , $\\mathbf{W}_k^i$ , $\\mathbf{W}_v^i\\in \\mathbb{R}^{d_v\\times d_k}$ , and $L_{q}$ denotes the predefined query length. Cross-attention serves", + "bbox": [ + 109, + 868, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as a prevalent approach, while other interaction strategies coexist [4]. The adopted VLMs in our implementation primarily rely on this approach for modality fusion.", + "bbox": [ + 531, + 363, + 903, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Text Generation:", + "text_level": 1, + "bbox": [ + 513, + 411, + 658, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The text encoder $\\mathsf{Enc}_{\\mathrm{text}}^{i}$ first processes the preceding expert's output $S_{i - 1}^{(t - 1)}$ into latent features:", + "bbox": [ + 531, + 424, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {i} = \\operatorname {E n c} _ {\\text {t e x t}} ^ {i} \\left(S _ {i - 1} ^ {(t - 1)}\\right) \\in \\mathbb {R} ^ {d _ {t} \\times L}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 472, + 903, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the text decoder $\\mathrm{Dec}_{\\mathrm{text}}^{i}$ synthesizes the final output $S_{i}^{(t)}$ by jointly conditioning on $\\mathbf{P}_i$ and $\\mathbf{H}_i$ :", + "bbox": [ + 532, + 506, + 903, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} _ {i} ^ {(t)} = \\operatorname {D e c} _ {\\text {t e x t}} ^ {i} \\left(\\mathbf {P} _ {i}, \\mathbf {H} _ {i}\\right) = \\left\\{w _ {1} ^ {(t, i)}, \\dots , w _ {m} ^ {(t, i)} \\right\\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 564, + 903, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ultimately, the final textual description $S_N^{(C)}$ is obtained after $C$ iteration steps through $N$ cascaded experts.", + "bbox": [ + 511, + 597, + 903, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Cross-Modal Similarity Verification", + "text_level": 1, + "bbox": [ + 511, + 640, + 823, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address noise in VLM-generated captions, we design a sliding window mechanism with semantic consistency verification. This method ensures that only relevant and semantically consistent segments are retained in the final description. Let $W_{k}$ denote the $k$ -th window containing $m$ consecutive tokens $\\{w_{km + 1},\\dots ,w_{(k + 1)m}\\}$ . For each window, we compute its cross-modal similarity score:", + "bbox": [ + 511, + 662, + 906, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {k} = \\frac {\\operatorname {E n c} _ {\\text {v i s}} (I) \\cdot \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right)}{\\| \\operatorname {E n c} _ {\\text {v i s}} (I) \\| \\| \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right) \\|}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 779, + 903, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathsf{Enc}_{vis/text}(\\cdot)$ adopts a lightweight CLIP [58] encoder-decoder with frozen parameters for efficient processing. The similarity score $\\alpha_{k}$ lies within the range [0, 1], with higher values indicating a stronger alignment between the visual and textual information.", + "bbox": [ + 511, + 824, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "984", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After calculating the cross-modal similarity for each window, we employ an empirical threshold $\\tau$ to filter out low-similarity windows. This threshold helps to identify and discard noisy or irrelevant sections of the generated caption that do not align well with the visual content, thereby reducing the impact of inaccurate or misleading descriptions. Formally, for each window $W_{k}$ , if $\\alpha_{k} < \\tau$ , the window is discarded as noise. This process effectively prunes windows with low similarity scores, ensuring that only semantically meaningful segments remain. The final denoised description $\\hat{S}$ is obtained by concatenating all windows $W_{k}$ for which $\\alpha_{k} \\geq \\tau$ :", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {S} = \\bigcup_ {\\alpha_ {k} \\geq \\tau} W _ {k}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 279, + 483, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our window size $m$ is flexibly determined and generally adapts dynamically to natural sentence segmentation.", + "bbox": [ + 89, + 320, + 483, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. MMKG Construction for Enhanced Reasoning", + "text_level": 1, + "bbox": [ + 89, + 358, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LLMs have become increasingly popular for identifying entities, relationships, and attributes within a corpus, which are then organized into a KG. The strength of LLM-based KG generation lies in its capacity to leverage the vast amount of knowledge encoded within these models, allowing them to detect complex and nuanced patterns across diverse data sources. This approach eliminates the need for manual annotation, enabling a highly scalable and domain-adaptive process suitable for a wide range of applications.", + "bbox": [ + 89, + 380, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by refining the generated textual description $\\hat{S}$ (VLM-based information), which is then optionally concatenated with any available external textual knowledge $T$ to form the input for KG generation. This combined input is used to generate MMKGs with the help of a LLM [21, 27], leveraging its capacity for multi-hop reasoning and dynamic knowledge integration.", + "bbox": [ + 89, + 517, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\operatorname {L L M} (\\hat {S} \\oplus T), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 632, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes optional concatenation based on the availability of $T$ . The resulting graph $\\mathcal{G}$ captures both visual and textual relationships inferred by the LLM.", + "bbox": [ + 89, + 660, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define $\\mathcal{G}$ as a set of triplets:", + "bbox": [ + 109, + 705, + 318, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\{(h, r, t) \\mid h, t \\in \\mathcal {E}, r \\in \\mathcal {R} \\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 729, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ and $\\mathcal{R}$ denote the sets of entities and relations. Entities include objects or concepts from the image or external text, while relations describe connections such as \"is a type of,\" \"part of,\" or \"has property.\" Each triplet $(h,r,t)$ links a head entity $h$ and a tail entity $t$ via relation $r$ .", + "bbox": [ + 89, + 753, + 483, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Reasoning Enhancement. To support multimodal reasoning, we retrieve relevant triplets from $\\mathcal{G}$ through structural patterns during LLMs inference:", + "bbox": [ + 89, + 830, + 483, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {q} = \\operatorname {R e t r i e v e} (q, \\mathcal {G}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Retrieve}(\\cdot)$ denotes a retrieval strategy that identifies subgraphs relevant to the query for reasoning. Detailed retrieval strategies are described in Appendix D.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The augmented prompt integrates multimodal evidence:", + "bbox": [ + 532, + 137, + 903, + 152 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {a u g}} = q \\left\\|\\left(\\bigcup_ {(h, r, t) \\in \\mathcal {G} _ {q}} [ h ] \\rightarrow r \\rightarrow [ t ]\\right). \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 165, + 903, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we incorporate the storage locations of images in the database during MMKGs construction, enabling the MMKGs to link to visual data. VaLiK enables text-only LLMs to perform multimodal reasoning through $\\mathcal{G}$ 's visual associations, while VLMs refresh knowledge representations by jointly injecting both visual and textual information, significantly mitigating hallucination risks.", + "bbox": [ + 511, + 205, + 906, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 327, + 637, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setups", + "text_level": 1, + "bbox": [ + 511, + 352, + 601, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets. We evaluate VaLiK on two multimodal reasoning benchmarks with distinct characteristics:", + "bbox": [ + 511, + 375, + 903, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- CrisisMMD [3]. This real-world disaster response dataset includes around 35,000 noisy social media postings with paired images and text, each annotated for seven catastrophe categories and four severity levels. Its realistic user-generated content with natural noise and implicit modality correlations provides a rigorous testbed for zero-shot adaptation, with good performance indicating practical relevance in real-world crisis scenarios.", + "bbox": [ + 513, + 406, + 903, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- ScienceQA [47]. This dataset contains 21,208 multimodal science questions combining textual and visual contexts, with $48.7\\%$ of instances containing images. Questions span physics, chemistry, and biology domains, requiring cross-modal reasoning between textual concepts and visual diagrams. Additionally, ScienceQA offers image captions to aid text-only LLMs in reasoning, allowing a comparison of unimodal approaches.", + "bbox": [ + 513, + 527, + 905, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Formulation. For CrisisMMD, we define three multimodal classification tasks1: (1) binary information relevance filtering, (2) fine-grained humanitarian category recognition, and (3) a consolidated taxonomy with merged categories to reduce label complexity. We omit the unimodal damage assessment to focus on multimodal aspects. For ScienceQA, we follow the original evaluation using multiple metrics: question types, contextual modalities, and educational stages. Performance is assessed through accuracy percentage across these categories.", + "bbox": [ + 511, + 648, + 906, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We conduct a comprehensive evaluation of text-only LLMs, multimodal VLMs, and KGs that enhance LLMs in multimodal reasoning.", + "bbox": [ + 511, + 799, + 903, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- For CrisisMMD, we compare text-only LLMs using few-shot prompting (LLaMA-2 [65], GPT-4 [2],", + "bbox": [ + 513, + 845, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "This setting references the repository GitHub and Abavisani et al. [1]", + "bbox": [ + 531, + 886, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "985", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ba85b4a98ffd5aca5fb455bf5872311135cd1ed5f88bf09e4b9f0d4269f53a72.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskText-only LLMsKG-Enhanced LLMs
LLaMA-2GPT-4DeepSeek-R1Qwen2.5LightRAGVaLiK
7B13B70B-7B8B32B70B7B32B72BText-onlyImage-onlyText-Image
Task 162.3263.8063.1566.8367.2363.3163.6165.5365.0467.2867.9567.4969.5268.90
Task 218.3221.8228.8747.2526.5325.4924.7721.0544.5246.9450.5145.1149.5450.02
Task 2 Merged21.4533.1536.8949.4425.8523.5621.5525.5745.3347.0750.2945.9449.0750.69
", + "bbox": [ + 94, + 88, + 903, + 181 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a322dfcaa43dd7ca6cae69d5260b7f829eb8a055cd6a1848cf1eada31544feb7.jpg", + "table_caption": [ + "Table 1. The performance evaluation of text-only LLMs using few-shot prompting without any fine-tuning on the training set. As these models handle text only, test data is formatted as unimodal text for compatibility. In our implementations, both LightRAG and VaLiK adopt Qwen2.5-7B as the base reasoning model. Bold indicates the highest value, and underline indicates the second highest." + ], + "table_footnote": [], + "table_body": "
TaskMultimodal VLMsKG-Enhanced LLMs
CLIPLLaVABLIP-2GPT-4oQwen2-VLVaLiK
ViT-L/147B13B34BFlan-T5-XLOPT-2B-I7B-I72B-I*#+~
Task 143.3654.0060.5856.4461.2938.6268.2047.5662.4565.8060.7868.4461.1168.89
Task 217.8828.0120.1425.1540.8614.2647.587.6032.6847.2125.8048.8827.2349.78
Task 2-M20.7930.6123.4425.0740.7214.2749.557.4234.2048.2827.3149.2729.0949.31
", + "bbox": [ + 94, + 246, + 919, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. The performance of multimodal VLMs and KG-enhanced LLMs. The -I suffix denotes instruction-tuned variants. Symbol markers denote KG types and models: the asterisk (*) represents image-only KG with LLaVA-34B, hash (#) indicates image-only KG using Qwen2-VL-72B-I, plus (+) denotes text-image KG with LLaVA-34B, and tilde (\\~) shows text-image KG using Qwen2-VL-72B-I.", + "bbox": [ + 89, + 359, + 906, + 405 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DeepSeek-R1 [25], Qwen-2.5 [76]) and multimodal VLMs (CLIP [56], LLaVA [42], GPT-4o [32], Qwen2-VL [68], BLIP-2 [40]).", + "bbox": [ + 102, + 428, + 480, + 472 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For ScienceQA, we compare models for general domains in zero/few-shot settings, including text-only LLMs (GPT Model [47], CoT [47], DDCoT [85]), multimodal VLMs (LG-VQA [24], LaVIN [49], BLIP-2, CCOT [52], GraphVis [18]) and Tool-LLM Chameleon [48]. These models are not specifically fine-tuned for scientific tasks, ensuring a fair evaluation of generalization capabilities.", + "- We further compare the multimodal reasoning performance of LLMs assisted by KGs, evaluating text-based KGs built with LightRAG [27], and pre-constructed MMKGs such as Visual Genome [37] and Mmkg [45]." + ], + "bbox": [ + 89, + 474, + 482, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation. For MMKG construction, we design a chain of VLMs including BLIP-2, LLaVA, and Qwen2-VL, with the CLIP-ViT-L/14 for pruning. Stronger or additional VLMs could be employed to enhance performance if more computational resources are available. We use the entire training set as the knowledge base and construct MMKGs from the extracted descriptions based on the LightRAG framework. In comparative experiments, the LightRAG method we evaluate utilizes only textual data, while VaLiK employs two configurations: (1) fully image-generated text descriptions (Image-only), and (2) original text combined with image-generated text (Text-Image). Dynamic window partitioning based on sentence length ensures syntactically coherent pruning results. Similarity thresholds are set to $\\tau = 0.25$ for CrisisMMD and $\\tau = 0.20$ for ScienceQA based on empirical evaluations to balance precision and recall. See Appendix E for selection details. We construct the", + "bbox": [ + 88, + 643, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "graph using DeepSeek-R1-70B and implement LightRAG's hybrid retrieval approach with Qwen2.5-7B. For graph construction and multimodal reasoning, we utilize $1 \\times$ NVIDIA A100-80GB GPUs. Task-specific prompts are designed to assist LLMs in multimodal reasoning evaluation.", + "bbox": [ + 511, + 428, + 906, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 513, + 653, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodal Classification Tasks. We conduct multimodal classification experiments on the CrisisMMD dataset, evaluating both text-only LLMs and multimodal VLMs. Detailed comparative results are provided in Tables 1 and 2. For text-only LLMs, we adopt Qwen2.5-7B as the foundational reasoning model. Remarkably, the VaLiK-enhanced version achieves state-of-the-art (SOTA) performance matching that of the native Qwen2.5-72B model. The image-only KG constructed through VaLiK demonstrates an average accuracy improvement of $4.41\\%$ across tasks, with the text-image variant attaining a $4.90\\%$ enhancement. These improvements significantly surpass the $1.22\\%$ gain obtained by LightRAG using textual KG. We further validate VaLiK's cross-scale applicability through evaluations on Qwen2.5-32B and 72B architectures, observing consistent $2.0\\% - 2.5\\%$ improvements. While not as significant as the 7B model's benefits, this shows that models that have substantial prior knowledge benefit less from external knowledge augmentation.", + "bbox": [ + 511, + 537, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike text-only LLMs that depend on MMKGs for visual understanding, VLMs primarily benefit from KGs integration through outdated knowledge refreshment. Due to the inherent availability of visual features during inference, VaLiK's performance gains for VLMs remain con-", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "986", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02345aa4ff865484870aefe25faa21813394a722f4f544f950b7f5f349a4dd9c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Human [47]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
GPT-4 [42]-84.0673.4587.3681.8770.7590.7384.6979.1082.69
CoT (GPT-3) [47]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
CoT (UnifiedQA) [47]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
CoT (GPT-4) [48]1T+85.4872.4490.2782.6571.4992.8986.6679.0483.99
DDCoT [85]175B80.1576.7282.8278.8972.5385.0282.8675.2180.15
Chameleon (ChatGPT) [48]175B+81.6270.6484.0079.7770.8086.6281.8676.5379.93
LG-VQA (BLIP-2) [24]---------86.32
LaVIN-13B [77]---------77.54
BLIP-2 [77]---------74.17
CCOT7B--------76.84
GraphVis [18]7B--------73.18
Qwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
Qwen2.5-72B72B79.6467.1084.9077.5665.0087.9380.2574.8578.37
Qwen2.5-7B (Mmkg) [45]7B73.9866.3778.1871.6564.3079.6576.5168.0373.47
Qwen2.5-7B (Visual Genome) [37]7B76.7867.0478.0974.0566.1979.7278.0869.6875.08
Qwen2.5-7B (VaLiK Text-only)7B84.5474.2486.9182.7472.5390.0384.5180.2882.98
Qwen2.5-7B (VaLiK Image-only)7B79.1471.5479.2777.1669.7283.1480.6573.9678.88
Qwen2.5-7B (VaLiK Text-Image)7B84.1575.1487.6482.9973.1889.6984.4080.9583.16
Qwen2.5-72B (VaLiK Text-Image)72B85.6175.9390.2784.4074.1792.3385.7982.9884.77
", + "bbox": [ + 94, + 88, + 903, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance comparison (\\%) on ScienceQA benchmark. #T-Params denotes trainable parameters. Categories: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG-Cap (image caption), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12). Method groups: (1) Human performance baseline, (2) Zero/Few-shot text-only LLMs, (3) Zero/Few-shot Multimodal VLMs, (4) LLMs enhanced with knowledge graphs for multimodal reasoning.", + "bbox": [ + 89, + 436, + 906, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strained compared to text-only counterparts. We separately applied VaLiK enhancement to Qwen2-VL-72B-Instruct and LLaVA-34B, obtaining distinct improvements: LLaVA-34B achieves accuracy gains of $2.41\\%$ (image-only KG) and $3.59\\%$ (text-image KG), while Qwen2-VL-72B-Instruct shows $1.77\\%$ and $2.23\\%$ improvements respectively under identical configurations. These experimental findings collectively demonstrate that VaLiK effectively extracts valuable signals from the training corpus and enables dynamic knowledge injection into VLMs during inference, thereby substantially alleviating hallucination phenomena. The differential improvements between Qwen2-VL-72B-Instruct and LLaVA-34B further validate the framework's adaptability across model architectures.", + "bbox": [ + 88, + 520, + 480, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we analyze the results of LLMs without KG enhancement in the tables, which generally follow the scaling law [36]. However, DeepSeek-R1 shows anomalous behavior. Through testing, we find that its reasoning process may introduce complex information that interferes with its judgment. Furthermore, empirical results show that most baseline models achieve suboptimal performance without fine-tuning. In contrast, VaLiK's automated MMKG construction framework requires no task-specific adaptation yet delivers consistent improvements.", + "bbox": [ + 89, + 732, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal Question Answering Tasks. We evalu", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ated multimodal QA performance on the ScienceQA benchmark with Qwen2.5-7B and Qwen2.5-72B as base architectures, augmented by four knowledge sources: Mmkg, Visual Genome, text-only LightRAG and VaLiK. Compared to existing zero-shot/few-shot LLMs that not specifically optimized for scientific QA, our VaLiK-enhanced Qwen2.5-72B achieved SOTA performance on $62.5\\%$ of subtasks, demonstrating particular strengths in multimodal reasoning scenarios requiring cross-modal alignment with an average accuracy gain of $6.4\\%$ over baseline models.", + "bbox": [ + 511, + 520, + 903, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our study identifies a fundamental imbalance between textual and visual knowledge representations in ScienceQA. Text-only KGs (14k entities, 18k relations) exhibit $8 \\times$ denser structured knowledge than image-only counterparts (3k concepts, 1k relations), explaining visual modality underperformance. Despite this gap, vision-KG-augmented Qwen2.5-7B still attains $4.16\\%$ accuracy gains over its non-enhanced version. Notably, our MMKG requires only 489MB storage for complete storage, while the scene graph component2 of Visual Genome alone occupies 739MB. This lightweight construction enables effective reasoning using only textual KG descriptions without raw images in resource-constrained scenarios.", + "bbox": [ + 511, + 672, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2Visual Genome", + "bbox": [ + 531, + 886, + 624, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "987", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4646fcd105f2d0a9a95e374a10d683ca7fe9398a4c070703b010b8b6c67db1fd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethod#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Image-OnlyQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B80.06 (↑3.86)70.30 (↑2.47)80.55 (↑3.28)78.05 (↑3.56)68.43 (↑2.64)83.76 (↑4.74)81.17 (↑3.45)72.71 (↑3.36)78.14 (↑3.42)
+ SV7B79.14 (↓0.92)71.54 (↑1.24)79.27 (↓1.28)77.16 (↓0.89)69.72 (↑1.29)83.14 (↓0.62)80.65 (↓0.52)73.96 (↑1.25)78.88 (↑0.74)
Text-ImageQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B81.88 (↑5.68)73.00 (↑5.17)84.00 (↑6.73)80.55 (↑6.06)70.05 (↑4.26)87.11 (↑8.09)82.01 (↑4.29)77.98 (↑8.63)80.57 (↑5.85)
+ SV7B84.15 (↑2.27)75.14 (↑2.14)87.64 (↑3.64)82.99 (↑2.44)73.18 (↑3.13)89.69 (↑2.58)84.40 (↑2.39)80.95 (↑2.97)83.16 (↑2.59)
", + "bbox": [ + 94, + 88, + 906, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e95ed00b81ec004a5771de39ec640774b155153c405c18bddcaaaf4337a28b87.jpg", + "table_caption": [ + "Table 4. Ablation study on ScienceQA benchmark (CVs: CoE-based Vision-Language Models; SV: Similarly Verification). Performance metrics include: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG (image context), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12)." + ], + "table_footnote": [], + "table_body": "
TypeMethodTask 1 (%)Task 2 (%)Task 2-Merged (%)
Image-OnlyQwen2.5-7B65.0444.5245.33
+ CVs68.11 (↑3.07)47.00 (↑2.48)46.95 (↑1.62)
+ SV69.52 (↑1.41)49.54 (↑2.54)49.07 (↑2.12)
Text-ImageQwen2.5-7B65.0444.5245.33
+ CVs68.43 (↑3.39)48.61 (↑4.09)48.97 (↑3.64)
+ SV68.90 (↑0.47)50.02 (↑1.41)50.69 (↑1.72)
", + "bbox": [ + 94, + 263, + 485, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on CrisisMMD with Qwen2.5-7B.", + "bbox": [ + 111, + 375, + 460, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 404, + 243, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation studies on CrisisMMD and ScienceQA demonstrate the specific roles of VaLiK's components. As shown in Table 4 and Table 5, the CVs (CoE-based VLM) module improves accuracy across all settings, with average gains of $+3.05\\%$ on CrisisMMD and $+4.63\\%$ on ScienceQA tasks, validating visual descriptions enhance reasoning. However, the SV (Similarly Verification) module exhibits dual effects: it significantly improves CrisisMMD metrics by pruning redundant textual descriptions, yet slightly degrades ScienceQA's image-only natural science reasoning. We hypothesize this discrepancy arises from dataset characteristics: CrisisMMD's generated captions contain substantially more redundant content, whereas ScienceQA's simpler visual scenes yield shorter descriptions. Pruning these shorter descriptions risks over-removal of critical semantics. Furthermore, different types of KGs influence the effectiveness of the components: CVs achieve greater gains in CrisisMMD's text-image fusion as original text provides complementary context, while SV shows reduced effectiveness, likely due to occasional over-pruning of cross-modal linkages. Nevertheless, both modules collectively enhance performance across configurations, demonstrating their synergistic yet context-sensitive nature.", + "bbox": [ + 89, + 426, + 483, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 787, + 258, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of VLM Quantity and Types. We evaluate the impact of varying quantities and types of VLMs on the CVs module. Our experiments reveal that Qwen2-VL generates the most visual descriptions, followed by LLaVA, while BLIP-2 produces the fewest. However, BLIP-2 demonstrates superior capability in extracting critical information", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c11325abdab9142cb8a1d96545e124e536e4c45e57e562428116da00e71ee8c.jpg", + "image_caption": [ + "Figure 4. Impact analysis of VLM quantity on CrisisMMD." + ], + "image_footnote": [], + "bbox": [ + 516, + 263, + 710, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9559482c345cc743f96b22e87bcf418e0d4c3d23fddec5c9335bc5e04f78ee59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 263, + 903, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and identifying key entity relationships within images. We therefore adopt BLIP-2 as the primary model, with LLaVA or Qwen2-VL serving as secondary/tertiary components. Adding more VLMs yields diminishing returns, due to limited entities in current images, though we hypothesize their benefits would increase for complex visual scenes with richer semantic content. This phenomenon is empirically validated by our quantitative results in Figure 4.", + "bbox": [ + 511, + 412, + 906, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Costs. Due to space limitations, we provide an overview of VaLiK's computational costs in Appendix F. Our method is significantly more cost-effective than manual annotation or LLM fine-tuning.", + "bbox": [ + 511, + 536, + 906, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 616, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multimodal reasoning in LLMs is constrained by incomplete knowledge and hallucination artifacts, limitations that persist because textual KGs cannot bridge visual-textual semantics due to their modality isolation. To bridge this gap, we propose VaLiK, a framework for constructing MMKGs through vision-language alignment, eliminating dependency on manual annotations while resolving visual-textual semantic inconsistencies. By integrating a cascade of pretrained VLMs and cross-modal verification, VaLiK converts images into structured knowledge while filtering noise. The resulting graphs enhance LLMs' reasoning with minimal storage overhead. Experiments on multimodal reasoning benchmarks show SOTA performance. VaLiK's modular design supports adaptability across domains, offering a scalable solution for autonomous knowledge synthesis. This work advances multimodal AI systems by enabling efficient integration of visual and textual data.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "988", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The research was supported by Shanghai Artificial Intelligence Laboratory, the National Key R&D Program of China (Grant No. 2022ZD0160201) and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ1100102).", + "bbox": [ + 89, + 114, + 483, + 190 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 219 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahdi Abavisani, Liwei Wu, Shengli Hu, Joel Tetreault, and Alejandro Jaimes. Multimodal categorization of crisis events in social media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 5", + "[3] Firoj Alam, Ferda Ofli, and Muhammad Imran. Crisismmd: Multimodal twitter datasets from natural disasters. Proceedings of the International AAAI Conference on Web and Social Media, 12(1), 2018. 2, 5", + "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736. Curran Associates, Inc., 2022. 3, 4", + "[5] Razvan Azamfirei, Sapna R Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 1", + "[6] Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2023. 2", + "[7] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):423-443, 2019. 3", + "[8] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17754-17762, 2024. 1", + "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems (NeurIPS), 33:22243-22255, 2020. 1", + "[10] Xiang Chen, Ningyu Zhang, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, and Huajun Chen." + ], + "bbox": [ + 93, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hybrid transformer with multi-level fusion for multimodal knowledge graph completion. In Proceedings of the International Conference on Research and Development in Information Retrieval (SIGIR), pages 904-915, 2022. 2", + "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 3", + "[12] Xiang Chen, Jingtian Zhang, Xiaohan Wang, Ningyu Zhang, Tongtong Wu, Yuxiang Wang, Yongheng Wang, and Huajun Chen. Continual multimodal knowledge graph construction. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 3", + "[13] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 1", + "[14] Yong Chen, Xinkai Ge, Shengli Yang, Linmei Hu, Jie Li, and Jinwen Zhang. A survey on multimodal knowledge graphs: Construction, completion and applications. Mathematics, 11 (8), 2023. 3", + "[15] Zhuo Chen, Yichi Zhang, Yin Fang, Yuxia Geng, Lingbing Guo, Xiang Chen, Qian Li, Wen Zhang, Jiaoyan Chen, Yushan Zhu, et al. Knowledge graphs meet multimodal learning: A comprehensive survey. arXiv preprint arXiv:2402.05391, 2024. 2", + "[16] Shiyao Cui, Jiangxia Cao, Xin Cong, Jiawei Sheng, Quanggang Li, Tingwen Liu, and Jinqiao Shi. Enhancing multimodal entity and relation extraction with variational information bottleneck. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:1274-1285, 2024. 2", + "[17] Xiaohui Cui, Xiaolong Qu, Dongmei Li, Yu Yang, Yuxun Li, and Xiaoping Zhang. Mkgcn: Multi-modal knowledge graph convolutional network for music recommender systems. *Electronics*, 12(12), 2023. 3", + "[18] Yihe Deng, Chenchen Ye, Zijie Huang, Mingyu Derek Ma, Yiwen Kou, and Wei Wang. Graphvis: Boosting llms with visual knowledge graph integration. In Advances in Neural Information Processing Systems, pages 67511-67534. Curran Associates, Inc., 2024. 6, 7", + "[19] Tausif Diwan, G. Anirudh, and Jitendra V. Tembhurne. Object detection using yolo: challenges, architectural successors, datasets and applications. Multimedia Tools Appl., 82 (6):9243-9275, 2022. 3", + "[20] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[21] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024. 5", + "[22] Quan Fang, Xiaowei Zhang, Jun Hu, Xian Wu, and Changsheng Xu. Contrastive multi-modal knowledge graph representation learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8983-8996, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "989", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Duoduo Feng, Xiangteng He, and Yuxin Peng. Mkvse: Multimodal knowledge enhanced visual-semantic embedding for image-text retrieval. ACM Trans. Multimedia Comput. Commun. Appl., 19(5), 2023. 3", + "[24] Deepanway Ghosal, Navonil Majumder, Roy Ka-Wei Lee, Rada Mihalcea, and Soujanya Poria. Language guided visual question answering: Elevate your multimodal language model using knowledge-enriched prompts. arXiv preprint arXiv:2310.20159, 2023. 6, 7", + "[25] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 6", + "[26] Yunfei Guo, Fei Yin, Xiao-hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21604-21613, 2023. 3", + "[27] ZIRUI GUO, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. LightRAG: Simple and fast retrieval-augmented generation, 2024. 2, 5, 6", + "[28] Lavdim Halilaj, Juergen Luettin, Sebastian Monka, Cory Henson, and Stefan Schmid. Knowledge graph-based integration of autonomous driving datasets. International Journal of Semantic Computing, 17(02):249-271, 2023. 3", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4", + "[30] Yang Hu, Guihua Wen, Adriane Chapman, Pei Yang, Mingnan Luo, Yingxue Xu, Dan Dai, and Wendy Hall. Graph-based visual-semantic entanglement network for zero-shot image recognition. IEEE Transactions on Multimedia, 24: 2473-2487, 2022. 3", + "[31] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. LLM-adapters: An adapter family for parameter-efficient fine-tuning of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5254-5276, Singapore, 2023. Association for Computational Linguistics. 2", + "[32] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 3, 6", + "[33] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2", + "[34] Adam Tauman Kalai and Santosh S. Vempala. Calibrated language models must hallucinate. In Proceedings of the 56th Annual ACM Symposium on Theory of Computing, page 160-171, New York, NY, USA, 2024. Association for Computing Machinery. 1", + "[35] Amar Viswanathan Kannan, Dmitriy Fradkin, Ioannis Akrotirianakis, Tugba Kulahcioglu, Arquimedes Canedo," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aditi Roy, Shih-Yuan Yu, Malawade Arnav, and Mohammad Abdullah Al Faruque. Multimodal knowledge graph for deep learning papers and code. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 3417-3420, New York, NY, USA, 2020. Association for Computing Machinery. 3", + "[36] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7", + "[37] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7", + "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10767-10782, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3", + "[39] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3", + "[40] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning (ICML), pages 19730–19742, 2023. 1, 3, 4, 6", + "[41] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466. Curran Associates, Inc., 2023. 3", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3, 6, 7", + "[43] Junming Liu, Yanting Gao, Siyuan Meng, Yifei Sun, Aoqi Wu, Yufei Jin, Yirong Chen, Ding Wang, and Guosun Zeng. Mosaic: Data-free knowledge distillation via mixture-of-experts for heterogeneous distributed environments. arXiv preprint arXiv:2505.19699, 2025. 1", + "[44] Junming Liu, Guosun Zeng, Ding Wang, Yanting Gao, and Yufei Jin. Fedrecon: Missing modality reconstruction in distributed heterogeneous environments. arXiv preprint arXiv:2504.09941, 2025.3", + "[45] Ye Liu, Hui Li, Alberto Garcia-Duran, Mathias Niepert, Daniel Onoro-Rubio, and David S Rosenblum. Mmkg: multi-modal knowledge graphs. In The Semantic Web: 16th International Conference, ESWC 2019, Portoroz, Slovenia, June 2-6, 2019, Proceedings 16, pages 459-474. Springer, 2019. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "990", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4", + "[47] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, pages 2507–2521. Curran Associates, Inc., 2022. 2, 5, 6, 7", + "[48] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In Advances in Neural Information Processing Systems, pages 43447-43478. Curran Associates, Inc., 2023. 6, 7", + "[49] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. In Advances in Neural Information Processing Systems, pages 29615-29627. Curran Associates, Inc., 2023. 6", + "[50] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[51] Runqing Miao, Qingxuan Jia, Fuchun Sun, Gang Chen, Haiming Huang, and Shengyi Miao. Semantic representation of robot manipulation with knowledge graph. Entropy, 25(4), 2023. 3", + "[52] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14420-14431, 2024. 6", + "[53] Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 1", + "[54] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 2", + "[55] Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Knowledge-aware multi-modal adaptive graph convolutional networks for fake news detection. ACM Trans. Multimedia Comput. Commun. Appl., 17(3), 2021. 3", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Brian Reily, Christopher Reardon, and Hao Zhang. Representing multi-robot structure through multimodal graph embedding for the selection of robot teams. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5576–5582, 2020. 3", + "[58] Joshua Robinson, Christopher Michael Ryting, and David Wingate. Leveraging large language models for multiple choice question answering. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4", + "[59] Hrituraj Singh, Anshul Nasery, Denil Mehta, Aishwarya Agarwal, Jatin Lamba, and Balaji Vasan Srinivasan. MIMOQA: Multimodal input multimodal output question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5317-5332, Online, 2021. Association for Computational Linguistics. 3", + "[60] Yaoxian Song, Penglei Sun, Haoyu Liu, Zhixu Li, Wei Song, Yanghua Xiao, and Xiaofang Zhou. Scene-driven multimodal knowledge graph construction for embodied ai. IEEE Transactions on Knowledge and Data Engineering, 36(11): 6962-6976, 2024. 2, 3", + "[61] Rui Sun, Xuezhi Cao, Yan Zhao, Junchen Wan, Kun Zhou, Fuzheng Zhang, Zhongyuan Wang, and Kai Zheng. Multimodal knowledge graphs for recommender systems. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 1405-1414, New York, NY, USA, 2020. Association for Computing Machinery. 3", + "[62] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2", + "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3", + "[64] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, ADITHYA JAIRAM IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1", + "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5", + "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4", + "[67] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton van den Hengel. Fvqa: Fact-based visual question an" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "991", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "swering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2413-2427, 2018. 3", + "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6", + "[69] Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In Proceedings of the 31st ACM International Conference on Multimedia, page 2391–2399, New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[70] Yuequn Wang, Liyan Dong, Hao Zhang, Xintao Ma, Yongli Li, and Minghui Sun. An enhanced multi-modal recommendation based on alternate training with knowledge graph representation. IEEE Access, 8:213012-213026, 2020. 3", + "[71] Tao Wu, Mengze Li, Jingyuan Chen, Wei Ji, Wang Lin, Jinyang Gao, Kun Kuang, Zhou Zhao, and Fei Wu. Semantic alignment for multimodal large language models. In Proceedings of the 32nd ACM International Conference on Multimedia, page 3489-3498, New York, NY, USA, 2024. Association for Computing Machinery. 1", + "[72] Yike Wu, Nan Hu, Guilin Qi, Sheng Bi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering. arXiv preprint arXiv:2309.11206, 2023. 2", + "[73] Ziyang Xiao, Dongxiang Zhang, Yangjun Wu, Lilin Xu, Yuan Jessica Wang, Xiongwei Han, Xiaojin Fu, Tao Zhong, Jia Zeng, Mingli Song, and Gang Chen. Chain-of-experts: When LLMs meet complex operations research problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[74] Dexuan Xu, Yanyuan Chen, Jieyi Wang, Yue Huang, Hanpin Wang, Zhi Jin, Hongxing Wang, Weihua Yue, Jing He, Hang Li, and Yu Huang. MLeVLM: Improve multi-level progressive capabilities based on multimodal large language model for medical visual question answering. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4977-4997, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1", + "[75] Peng Xu, Xiatian Zhu, and David A. Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 3", + "[76] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6", + "[77] Xiaocui Yang, Wenfang Wu, Shi Feng, Ming Wang, Daling Wang, Yang Li, Qi Sun, Yifei Zhang, Xiaoming Fu, and Soujanya Poria. Mm-bigbench: Evaluating multimodal models on multimodal content comprehension tasks. arXiv preprint arXiv:2310.09036, 2023. 7", + "[78] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 3", + "[79] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1", + "[80] Jingtong Yue, Zhiwei Lin, Xin Lin, Xiaoyu Zhou, Xiangtai Li, Lu Qi, Yongtao Wang, and Ming-Hsuan Yang. RobuR-CDet: Enhancing robustness of radar-camera fusion in bird's eye view for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[81] Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Huajun Chen, and Wen Zhang. Mygo: Discrete modality information as fine-grained tokens for multi-modal knowledge graph completion. CoRR, abs/2404.09468, 2024. 2", + "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 3", + "[83] Jiabao Zhao, Xin Lin, Jie Zhou, Jing Yang, Liang He, and Zhaohui Yang. Knowledge-based fine-grained classification for few-shot learning. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 3", + "[84] Changmeng Zheng, Junhao Feng, Ze Fu, Yi Cai, Qing Li, and Tao Wang. Multimodal relation extraction with efficient graph alignment. In Proceedings of the 29th ACM International Conference on Multimedia, page 5298-5306, New York, NY, USA, 2021. Association for Computing Machinery. 2", + "[85] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. In Advances in Neural Information Processing Systems, pages 5168-5191. Curran Associates, Inc., 2023. 3, 6, 7", + "[86] Juncheng Zheng, Meiyu Liang, Yang Yu, Yawen Li, and Zhe Xue. Knowledge graph enhanced multimodal transformer for image-text retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 70-82, 2024. 3", + "[87] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 4", + "[88] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[89] Xiangru Zhu, Zhixu Li, Xiaodan Wang, Xueyao Jiang, Penglei Sun, Xuwu Wang, Yanghua Xiao, and Nicholas Jing Yuan. Multi-modal knowledge graph construction and application: A survey. IEEE Transactions on Knowledge and Data Engineering, 36(2):715-735, 2024. 2", + "[90] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 896 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "992", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_model.json b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..79d124d674bc5e26046e5be075fcee8c8f04bf3e --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_model.json @@ -0,0 +1,2743 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.203, + 0.793, + 0.222 + ], + "angle": 0, + "content": "Junming Liu\\(^{1,2}\\), Siyuan Meng\\(^{2,3}\\), Yanting Gao\\(^{1}\\), Song Mao\\(^{2}\\), Pinlong Cai\\(^{2}\\)," + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.222, + 0.776, + 0.24 + ], + "angle": 0, + "content": "Guohang Yan\\(^{2}\\), Yirong Chen\\(^{2,4}\\), Zilin Bian\\(^{5}\\), Ding Wang\\(^{2*}\\), Botian Shi\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.24, + 0.756, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Tongji University \\(^{2}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.257, + 0.81, + 0.275 + ], + "angle": 0, + "content": "\\(^{3}\\)East China Normal University \\(^{4}\\)Stanford University \\(^{5}\\)New York University" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.278, + 0.722, + 0.293 + ], + "angle": 0, + "content": "liu_junming6917@tongji.edu.cn wangding@pjlab.org.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.486, + 0.766 + ], + "angle": 0, + "content": "Multimodal reasoning in Large Language Models (LLMs) struggles with incomplete knowledge and hallucination artifacts, challenges that textual Knowledge Graphs (KGs) only partially mitigate due to their modality isolation. While Multimodal Knowledge Graphs (MMKGs) promise enhanced cross-modal understanding, their practical construction is impeded by semantic narrowness of manual text annotations and inherent noise in visual-semantic entity linkages. In this paper, we propose Vision-align-to-Language integrated Knowledge Graph (VaLiK), a novel approach for constructing MMKGs that enhances LLMs reasoning through cross-modal information supplementation. Specifically, we cascade pre-trained Vision-Language Models (VLMs) to align image features with text, transforming them into descriptions that encapsulate image-specific information. Furthermore, we developed a cross-modal similarity verification mechanism to quantify semantic consistency, effectively filtering out noise introduced during feature alignment. Even without manually annotated image captions, the refined descriptions alone suffice to construct the MMKG. Compared to conventional MMKGs construction paradigms, our approach achieves substantial storage efficiency gains while maintaining direct entity-to-image linkage capability. Experimental results on multimodal reasoning tasks demonstrate that LLMs augmented with VaLiK outperform previous state-of-the-art models. Our code is published at https://github.com/Wings-Of-Disaster/VaLiK." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.794, + 0.222, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Recent advancements in Large Language Models (LLMs) [2, 9, 25, 65] have demonstrated their superiority and versatility across various Natural Language Reasoning (NLR) tasks [8, 43, 53, 58]. To enhance LLMs into the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.329, + 0.902, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.699, + 0.908, + 0.756 + ], + "angle": 0, + "content": "Figure 1. (a) Training entity extraction models relies on extensive fine-grained annotations, increasing labeling costs. More examples are provided in Appendix B. (b) Capturing implicit semantic associations demands abstract comprehension or logical inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "realm of multimodal reasoning, researchers [64, 71, 74, 79] have endeavored to equip these models with multimodal capabilities, as evidenced by advancements in Multimodal Large Language Models (MLLMs) such as BLIP-2 [40], GPT-4o [32], Janus-Pro [13], among others. Despite their notable progress, these models often experience hallucinations [5, 34], primarily arising from knowledge deficiencies due to incomplete or obsolete information." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "981" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "Fine-tuning LLMs demands prohibitive computational costs [31]. While text-based Knowledge Graphs (KGs) have partially addressed this limitation by efficient real-time updates [6, 62, 72], they are still restricted by modal isolation, which hinders cross-modal reasoning, as detailed in Appendix A. To bridge this semantic fragmentation, Multimodal Knowledge Graphs (MMKGs) have been developed as unified representational frameworks [10, 33, 38, 45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.482, + 0.352 + ], + "angle": 0, + "content": "However, constructing robust MMKGs faces two primary obstacles [15, 89]. First, the lack of large-scale fine-grained entity-image corpora makes it infeasible to train high-quality entity extractors, significantly constraining scalability, as illustrated in Figure 1a. Second, conventional visual relation detectors primarily identify superficial spatial interactions instead of semantic relations consistent with KGs, while frequently hallucinating implausible connections that corrupt graph integrity, as shown in Figure 1b." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.793 + ], + "angle": 0, + "content": "In this paper, we propose VaLiK, short for Vision-align-to-Language integrated Knowledge Graph, a novel framework designed to empower LLMs with advanced multimodal reasoning. Unlike traditional methods that rely on text annotations for training extraction models and the knowledge construction process [54], VaLiK adopts a annotation-free approach to MMKGs construction. Specifically, we first employ several pretrained Vision-Language models (VLMs), designed based on Chain-of-Experts (CoE) principles [73], to convert visual inputs into image-specific textual descriptions through cross-modal feature alignment. This procedure eliminates the need for manually annotated image captions in both the knowledge extraction and construction phases while preserving visual details typically missing in generic text descriptions. Moreover, in contrast to existing relation detection methods that require predefined label taxonomies [16, 60, 81, 84], VaLiK excels at extracting profound semantic relationships that are both KG-compatible and capture novel associations beyond training supervision. While VLMs enable cross-modal reasoning and interpretation, they introduce spurious relational noise through hallucinated inter-modal attributions, as depicted in Figure 2. We address this limitation through cross-modal similarity recalibration, strategically filtering inconsistent information while preserving valid semantic correspondences. Finally, the purified descriptions are systematically organized into MMKGs via LLM-driven symbolic structuring [27], bridging visual and textual domains with factual consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To thoroughly evaluate the VaLiK method, we conduct a comprehensive assessment across two critical multimodal benchmarks: multimodal classification (tested on the CrisisMMD dataset [3]) and multimodal question answering (evaluated via the ScienceQA benchmark [47]). The experiments span diverse LLM architectures and MMKG construction techniques to ensure the framework's robustness." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.903, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Figure 2. Feature-aligned descriptions from VLMs introduce redundant and inaccurate relationship patterns." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.409 + ], + "angle": 0, + "content": "The experimental results demonstrate that the MMKGs constructed by VaLiK achieve superior multimodal reasoning performance in LLMs while requiring substantially less storage than conventional approaches. More importantly, the proposed approach retains direct entity-to-image linkage capabilities even with the compressed graph structure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.905, + 0.47 + ], + "angle": 0, + "content": "In summary, VaLiK is the first framework that enables end-to-end, annotation-free, zero-shot, and storage-efficient multimodal knowledge construction with high adaptability and scalability. Our key contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.472, + 0.904, + 0.561 + ], + "angle": 0, + "content": "- To the best of our knowledge, VaLik is the first end-to-end framework to build Annotation-Free MMKGs to improve LLMs' multimodal reasoning capabilities, effectively eliminating the need for manually annotated textual material and enabling a completely autonomous multimodal knowledge generation process." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.563, + 0.905, + 0.652 + ], + "angle": 0, + "content": "- We offer an innovative zero-shot method for constructing MMKG that captures deep semantic connections beyond traditional predetermined labels with an effective verification system that guarantees the accuracy of these relationships. The knowledge distillation paradigm greatly decreases storage while maintaining semantic integrity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.653, + 0.905, + 0.728 + ], + "angle": 0, + "content": "- We develop a highly modular and extensible architecture that allows VaLiK to effortlessly incorporate new models and workflows for specialized domain tasks, facilitating rapid adaptation to diverse application scenarios without incurring expensive system changes." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.472, + 0.905, + 0.728 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.746, + 0.655, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.794, + 0.789 + ], + "angle": 0, + "content": "2.1. Multimodal Knowledge Graphs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The principal advantage of MMKGs resides in their multimodal integration beyond conventional KGs. By linking entities with corresponding visual or textual data, MMKGs introduce valuable visual and textual information to the knowledge base, substantially advancing multimodal reasoning capabilities. This combination addresses core challenges in tasks that inherently demand multimodal synergy" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.381 + ], + "angle": 0, + "content": "like autonomous driving [26, 28], image-text retrieval [23, 86] and robotic manipulation [51, 57]. However, constructing trustworthy MMKGs with minimal manual effort remains a critical challenge. Recent studies have proposed innovative strategies to enhance MMKG reliability and utility. For instance, Chen et al. [12] proposed MSPT, a framework addressing continual MMKG construction through gradient modulation for balanced multimodal learning and attention distillation to mitigate catastrophic forgetting. Song et al. [60] developed Scene-MMKG, integrating knowledge engineering with large language models to improve robotic manipulation by resolving data sparsity and knowledge uncertainty. Wang et al. [69] introduced TIVA-KG, the first quad-modal knowledge graph spanning text, image, video, and audio with triplet grounding, empirically validating its effectiveness in downstream tasks. While these advances enhance multimodal reasoning capabilities, their efficacy remains rooted in resource-intensive paradigms, requiring extensively annotated datasets for knowledge acquisition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.479, + 0.407 + ], + "angle": 0, + "content": "2.2. Knowledge-Augmented Multimodal Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Multimodal learning has seen significant progress in aligning and integrating information across different data modalities [7, 44, 75]. The incorporation of structured knowledge through MMKGs further enhances these approaches, improving the reasoning capabilities and generalization across a variety of domains, such as visual question answering [50, 59, 67], recommendation systems [17, 61, 70], and classification [30, 55, 83]. Methods like GraphAdapter's dual-KG adaptation [41] and contrastive multi-relational encoding with KGs [22] inject external knowledge into models, refining their performance and improving their capability to handle complex tasks. Additionally, Lee et al. [38] proposed MR-MKG, a novel framework that constructs task-specific MMKGs to enhance multimodal reasoning in LLMs. These knowledge-augmented paradigms demonstrate superior cross-modal semantic grounding compared to unimodal approaches [14, 35]. However, their reliance on preconstructed MMKGs often leads to domain discrepancies, where generic knowledge schemas misalign with task-specific reasoning patterns, ultimately limiting contextual precision in target applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.411, + 0.759 + ], + "angle": 0, + "content": "2.3. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The limitations of text-only LLMs in meeting increasingly complex demands have spurred extensive research [78, 82, 85] into developing LLMs capable of effectively processing and reasoning over multimodal inputs. Current research predominantly employs adapter or projection layers to connect the embedding spaces of various modality-specific encoders with the textual embedding space of LLMs [38]. For instance, foundational models like CLIP [56] and BLIP [39] pioneered cross-modal alignment by jointly training vision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.32 + ], + "angle": 0, + "content": "and text encoders to map images and text into a shared embedding space. Building on this, LLaVA [42] and Flamingo [4] advanced the field by integrating visual encoders with LLMs, enabling more nuanced multimodal understanding and generation. More recently, Gemini [63], Qwen2-VL [68] and GPT-4o [32] have further pushed the boundaries by scaling up multimodal pretraining and introducing sophisticated mechanisms for cross-modal interaction. However, multimodal LLMs remain prone to hallucinations. While they enhance cross-modal alignment, they neither acquire new knowledge nor avoid introducing noise through integration. To address these limitations, VaLiK \"uses the master's tools to refine the master's craft,\" first constructing MMKGs via MLLMs and then leveraging them to enhance MLLMs' reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.333, + 0.605, + 0.348 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.909, + 0.542 + ], + "angle": 0, + "content": "In this section, we present the technical details of VaLiK. VaLiK introduces a novel expansion-reduction paradigm for visual knowledge extraction. The architecture initially organizes several VLMs with distinct knowledge domains, designed based on CoE principles [73], to produce comprehensive textual descriptions encompassing hierarchical visual details. A cross-modal similarity verification mechanism then iteratively filters out noisy tokens through cross-modal alignment while preserving semantically salient elements. This optimization-style approach eliminates external textual dependencies while enabling effective MMKG construction. VaLiK's framework is shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.55, + 0.866, + 0.567 + ], + "angle": 0, + "content": "3.1. CoE-based Visual to Language Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.692 + ], + "angle": 0, + "content": "Recent entity detection techniques [19, 80, 90] have been widely adopted for entity and relation extraction in MMKG construction. However, these methods are inherently limited by predefined categorical boundaries, lacking the capacity to recognize visual concepts outside their training vocabulary. In contrast, VLMs pretrained on web-scale corpora [11, 40, 88] exhibit broader recognition capabilities through exposure to diverse visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.907, + 0.769 + ], + "angle": 0, + "content": "We therefore leverage pretrained VLMs to extract comprehensive visual information. This process removes the necessity for detailed fine-grained data typically required to train specialized recognition models. The generalized vision to language conversion pipeline can be formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.779, + 0.907, + 0.805 + ], + "angle": 0, + "content": "\\[\nS = \\mathcal {D} _ {\\text {t e x t}} \\left(\\mathcal {A} \\left(\\mathcal {E} _ {\\text {v i s}} (I)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(I\\) denotes for the input image, \\(\\mathcal{E}_{\\mathrm{vis}}\\) denotes the visual encoder extracting visual features, \\(\\mathcal{A}\\) carries out cross-modal feature alignment and interaction, and \\(\\mathcal{D}_{\\mathrm{text}}\\) generates textual tokens through autoregressive decoding. The resulting visual description \\(S = \\{w_{1},\\dots,w_{n}\\}\\) emerges from this multi-stage processing." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.309, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 3. The pipeline of VaLiK: First, large-scale visual descriptions are generated using CoE-based VLMs. Then, a similarity verification mechanism is used to prune irrelevant information. Finally, MMKGs are constructed using LLMs based on LightRAG. The constructed MMKGs can assist LLMs in multimodal reasoning, alleviating the hallucination issues caused by incomplete knowledge." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.484, + 0.515 + ], + "angle": 0, + "content": "However, quantitative analysis uncovers considerable discrepancies between machine-generated and human-annotated descriptions [87]. As an illustration, while utilizing BLIP-2 [40] to generate sample captions, we noted that the model outputs are markedly concise and devoid of visual specifics, as detailed in Appendix C. To bridge this gap, we implement CoE enhanced generation through cascade VLMs processing. At iteration step \\( t \\), each expert \\( E_{i} \\) receives both the original visual signals \\( I \\) and the contextual output from the preceding expert \\( E_{i - 1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.527, + 0.483, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} ^ {(t)} = E _ {i} \\left(I, \\mathcal {S} _ {i - 1} ^ {(t - 1)}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.601 + ], + "angle": 0, + "content": "where \\(S_{i - 1}^{(t - 1)}\\) denotes the description from expert \\(E_{i - 1}\\) at step \\(t - 1\\), with \\(S_0^{(t)}\\coloneqq \\emptyset\\) for initialization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.634 + ], + "angle": 0, + "content": "Specifically, each expert \\( E_{i} \\) implements a unified visual-language processing task:" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.303, + 0.65 + ], + "angle": 0, + "content": "1. Visual Feature Extraction:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.662, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {i} = \\operatorname {E n c} _ {\\text {v i s}} ^ {i} (I) \\in \\mathbb {R} ^ {d _ {v} \\times N _ {p}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.693, + 0.483, + 0.724 + ], + "angle": 0, + "content": "where \\(\\mathsf{Enc}_{\\mathrm{vis}}^i\\) denotes established visual encoder [20, 29, 46] producing \\(N_{p}\\) patch embeddings with dimension \\(d_v\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.408, + 0.739 + ], + "angle": 0, + "content": "2. Cross-Modal Interaction and Generation:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.74, + 0.483, + 0.787 + ], + "angle": 0, + "content": "VLMs integrate pretrained learnable query embeddings \\(\\mathbf{Q}_i\\in \\mathbb{R}^{d_q\\times L_q}\\) to interact with visual features \\(\\mathbf{V}_i\\in\\) \\(\\mathbb{R}^{d_v\\times N_p}\\) via cross-attention [66]:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.797, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {H} _ {i} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {Q} _ {i}, \\mathbf {V} _ {i}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {W} _ {q} ^ {i} \\left(\\mathbf {V} _ {i} \\mathbf {W} _ {k} ^ {i}\\right) ^ {\\top}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V} _ {i} \\mathbf {W} _ {v} ^ {i}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.869, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_q^i\\in \\mathbb{R}^{d_q\\times d_k}\\), \\(\\mathbf{W}_k^i\\), \\(\\mathbf{W}_v^i\\in \\mathbb{R}^{d_v\\times d_k}\\), and \\(L_{q}\\) denotes the predefined query length. Cross-attention serves" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.364, + 0.905, + 0.41 + ], + "angle": 0, + "content": "as a prevalent approach, while other interaction strategies coexist [4]. The adopted VLMs in our implementation primarily rely on this approach for modality fusion." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.412, + 0.659, + 0.425 + ], + "angle": 0, + "content": "3. Text Generation:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.425, + 0.905, + 0.461 + ], + "angle": 0, + "content": "The text encoder \\(\\mathsf{Enc}_{\\mathrm{text}}^{i}\\) first processes the preceding expert's output \\(S_{i - 1}^{(t - 1)}\\) into latent features:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.473, + 0.905, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {i} = \\operatorname {E n c} _ {\\text {t e x t}} ^ {i} \\left(S _ {i - 1} ^ {(t - 1)}\\right) \\in \\mathbb {R} ^ {d _ {t} \\times L}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.507, + 0.905, + 0.541 + ], + "angle": 0, + "content": "Subsequently, the text decoder \\(\\mathrm{Dec}_{\\mathrm{text}}^{i}\\) synthesizes the final output \\(S_{i}^{(t)}\\) by jointly conditioning on \\(\\mathbf{P}_i\\) and \\(\\mathbf{H}_i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.565, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} _ {i} ^ {(t)} = \\operatorname {D e c} _ {\\text {t e x t}} ^ {i} \\left(\\mathbf {P} _ {i}, \\mathbf {H} _ {i}\\right) = \\left\\{w _ {1} ^ {(t, i)}, \\dots , w _ {m} ^ {(t, i)} \\right\\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.631 + ], + "angle": 0, + "content": "Ultimately, the final textual description \\( S_N^{(C)} \\) is obtained after \\( C \\) iteration steps through \\( N \\) cascaded experts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.825, + 0.657 + ], + "angle": 0, + "content": "3.2. Cross-Modal Similarity Verification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.77 + ], + "angle": 0, + "content": "To address noise in VLM-generated captions, we design a sliding window mechanism with semantic consistency verification. This method ensures that only relevant and semantically consistent segments are retained in the final description. Let \\( W_{k} \\) denote the \\( k \\)-th window containing \\( m \\) consecutive tokens \\( \\{w_{km + 1},\\dots ,w_{(k + 1)m}\\} \\). For each window, we compute its cross-modal similarity score:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.78, + 0.905, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {k} = \\frac {\\operatorname {E n c} _ {\\text {v i s}} (I) \\cdot \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right)}{\\| \\operatorname {E n c} _ {\\text {v i s}} (I) \\| \\| \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right) \\|}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.9 + ], + "angle": 0, + "content": "where \\(\\mathsf{Enc}_{vis/text}(\\cdot)\\) adopts a lightweight CLIP [58] encoder-decoder with frozen parameters for efficient processing. The similarity score \\(\\alpha_{k}\\) lies within the range [0, 1], with higher values indicating a stronger alignment between the visual and textual information." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "After calculating the cross-modal similarity for each window, we employ an empirical threshold \\(\\tau\\) to filter out low-similarity windows. This threshold helps to identify and discard noisy or irrelevant sections of the generated caption that do not align well with the visual content, thereby reducing the impact of inaccurate or misleading descriptions. Formally, for each window \\(W_{k}\\), if \\(\\alpha_{k} < \\tau\\), the window is discarded as noise. This process effectively prunes windows with low similarity scores, ensuring that only semantically meaningful segments remain. The final denoised description \\(\\hat{S}\\) is obtained by concatenating all windows \\(W_{k}\\) for which \\(\\alpha_{k} \\geq \\tau\\):" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.28, + 0.484, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\hat {S} = \\bigcup_ {\\alpha_ {k} \\geq \\tau} W _ {k}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.352 + ], + "angle": 0, + "content": "Our window size \\( m \\) is flexibly determined and generally adapts dynamically to natural sentence segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.359, + 0.484, + 0.375 + ], + "angle": 0, + "content": "3.3. MMKG Construction for Enhanced Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.517 + ], + "angle": 0, + "content": "LLMs have become increasingly popular for identifying entities, relationships, and attributes within a corpus, which are then organized into a KG. The strength of LLM-based KG generation lies in its capacity to leverage the vast amount of knowledge encoded within these models, allowing them to detect complex and nuanced patterns across diverse data sources. This approach eliminates the need for manual annotation, enabling a highly scalable and domain-adaptive process suitable for a wide range of applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.623 + ], + "angle": 0, + "content": "We begin by refining the generated textual description \\(\\hat{S}\\) (VLM-based information), which is then optionally concatenated with any available external textual knowledge \\(T\\) to form the input for KG generation. This combined input is used to generate MMKGs with the help of a LLM [21, 27], leveraging its capacity for multi-hop reasoning and dynamic knowledge integration." + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.633, + 0.483, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\operatorname {L L M} (\\hat {S} \\oplus T), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.707 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes optional concatenation based on the availability of \\(T\\). The resulting graph \\(\\mathcal{G}\\) captures both visual and textual relationships inferred by the LLM." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.707, + 0.319, + 0.722 + ], + "angle": 0, + "content": "We define \\(\\mathcal{G}\\) as a set of triplets:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.731, + 0.483, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\{(h, r, t) \\mid h, t \\in \\mathcal {E}, r \\in \\mathcal {R} \\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.831 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) and \\(\\mathcal{R}\\) denote the sets of entities and relations. Entities include objects or concepts from the image or external text, while relations describe connections such as \"is a type of,\" \"part of,\" or \"has property.\" Each triplet \\((h,r,t)\\) links a head entity \\(h\\) and a tail entity \\(t\\) via relation \\(r\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Multimodal Reasoning Enhancement. To support multimodal reasoning, we retrieve relevant triplets from \\(\\mathcal{G}\\) through structural patterns during LLMs inference:" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {q} = \\operatorname {R e t r i e v e} (q, \\mathcal {G}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "where \\( \\text{Retrieve}(\\cdot) \\) denotes a retrieval strategy that identifies subgraphs relevant to the query for reasoning. Detailed retrieval strategies are described in Appendix D." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.138, + 0.905, + 0.153 + ], + "angle": 0, + "content": "The augmented prompt integrates multimodal evidence:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.166, + 0.905, + 0.201 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {a u g}} = q \\left\\|\\left(\\bigcup_ {(h, r, t) \\in \\mathcal {G} _ {q}} [ h ] \\rightarrow r \\rightarrow [ t ]\\right). \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Note that we incorporate the storage locations of images in the database during MMKGs construction, enabling the MMKGs to link to visual data. VaLiK enables text-only LLMs to perform multimodal reasoning through \\(\\mathcal{G}\\)'s visual associations, while VLMs refresh knowledge representations by jointly injecting both visual and textual information, significantly mitigating hallucination risks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.328, + 0.638, + 0.344 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.353, + 0.602, + 0.37 + ], + "angle": 0, + "content": "4.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.376, + 0.905, + 0.406 + ], + "angle": 0, + "content": "Evaluation Datasets. We evaluate VaLiK on two multimodal reasoning benchmarks with distinct characteristics:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.407, + 0.905, + 0.527 + ], + "angle": 0, + "content": "- CrisisMMD [3]. This real-world disaster response dataset includes around 35,000 noisy social media postings with paired images and text, each annotated for seven catastrophe categories and four severity levels. Its realistic user-generated content with natural noise and implicit modality correlations provides a rigorous testbed for zero-shot adaptation, with good performance indicating practical relevance in real-world crisis scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.528, + 0.906, + 0.648 + ], + "angle": 0, + "content": "- ScienceQA [47]. This dataset contains 21,208 multimodal science questions combining textual and visual contexts, with \\(48.7\\%\\) of instances containing images. Questions span physics, chemistry, and biology domains, requiring cross-modal reasoning between textual concepts and visual diagrams. Additionally, ScienceQA offers image captions to aid text-only LLMs in reasoning, allowing a comparison of unimodal approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.649, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Task Formulation. For CrisisMMD, we define three multimodal classification tasks1: (1) binary information relevance filtering, (2) fine-grained humanitarian category recognition, and (3) a consolidated taxonomy with merged categories to reduce label complexity. We omit the unimodal damage assessment to focus on multimodal aspects. For ScienceQA, we follow the original evaluation using multiple metrics: question types, contextual modalities, and educational stages. Performance is assessed through accuracy percentage across these categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.8, + 0.905, + 0.845 + ], + "angle": 0, + "content": "Baselines. We conduct a comprehensive evaluation of text-only LLMs, multimodal VLMs, and KGs that enhance LLMs in multimodal reasoning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.846, + 0.905, + 0.877 + ], + "angle": 0, + "content": "- For CrisisMMD, we compare text-only LLMs using few-shot prompting (LLaMA-2 [65], GPT-4 [2]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This setting references the repository GitHub and Abavisani et al. [1]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "985" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.183 + ], + "angle": 0, + "content": "
TaskText-only LLMsKG-Enhanced LLMs
LLaMA-2GPT-4DeepSeek-R1Qwen2.5LightRAGVaLiK
7B13B70B-7B8B32B70B7B32B72BText-onlyImage-onlyText-Image
Task 162.3263.8063.1566.8367.2363.3163.6165.5365.0467.2867.9567.4969.5268.90
Task 218.3221.8228.8747.2526.5325.4924.7721.0544.5246.9450.5145.1149.5450.02
Task 2 Merged21.4533.1536.8949.4425.8523.5621.5525.5745.3347.0750.2945.9449.0750.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.192, + 0.908, + 0.237 + ], + "angle": 0, + "content": "Table 1. The performance evaluation of text-only LLMs using few-shot prompting without any fine-tuning on the training set. As these models handle text only, test data is formatted as unimodal text for compatibility. In our implementations, both LightRAG and VaLiK adopt Qwen2.5-7B as the base reasoning model. Bold indicates the highest value, and underline indicates the second highest." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.247, + 0.92, + 0.351 + ], + "angle": 0, + "content": "
TaskMultimodal VLMsKG-Enhanced LLMs
CLIPLLaVABLIP-2GPT-4oQwen2-VLVaLiK
ViT-L/147B13B34BFlan-T5-XLOPT-2B-I7B-I72B-I*#+~
Task 143.3654.0060.5856.4461.2938.6268.2047.5662.4565.8060.7868.4461.1168.89
Task 217.8828.0120.1425.1540.8614.2647.587.6032.6847.2125.8048.8827.2349.78
Task 2-M20.7930.6123.4425.0740.7214.2749.557.4234.2048.2827.3149.2729.0949.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.361, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Table 2. The performance of multimodal VLMs and KG-enhanced LLMs. The -I suffix denotes instruction-tuned variants. Symbol markers denote KG types and models: the asterisk (*) represents image-only KG with LLaVA-34B, hash (#) indicates image-only KG using Qwen2-VL-72B-I, plus (+) denotes text-image KG with LLaVA-34B, and tilde (\\~) shows text-image KG using Qwen2-VL-72B-I." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.429, + 0.482, + 0.473 + ], + "angle": 0, + "content": "DeepSeek-R1 [25], Qwen-2.5 [76]) and multimodal VLMs (CLIP [56], LLaVA [42], GPT-4o [32], Qwen2-VL [68], BLIP-2 [40])." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.475, + 0.483, + 0.58 + ], + "angle": 0, + "content": "- For ScienceQA, we compare models for general domains in zero/few-shot settings, including text-only LLMs (GPT Model [47], CoT [47], DDCoT [85]), multimodal VLMs (LG-VQA [24], LaVIN [49], BLIP-2, CCOT [52], GraphVis [18]) and Tool-LLM Chameleon [48]. These models are not specifically fine-tuned for scientific tasks, ensuring a fair evaluation of generalization capabilities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.581, + 0.483, + 0.642 + ], + "angle": 0, + "content": "- We further compare the multimodal reasoning performance of LLMs assisted by KGs, evaluating text-based KGs built with LightRAG [27], and pre-constructed MMKGs such as Visual Genome [37] and Mmkg [45]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.475, + 0.483, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation. For MMKG construction, we design a chain of VLMs including BLIP-2, LLaVA, and Qwen2-VL, with the CLIP-ViT-L/14 for pruning. Stronger or additional VLMs could be employed to enhance performance if more computational resources are available. We use the entire training set as the knowledge base and construct MMKGs from the extracted descriptions based on the LightRAG framework. In comparative experiments, the LightRAG method we evaluate utilizes only textual data, while VaLiK employs two configurations: (1) fully image-generated text descriptions (Image-only), and (2) original text combined with image-generated text (Text-Image). Dynamic window partitioning based on sentence length ensures syntactically coherent pruning results. Similarity thresholds are set to \\(\\tau = 0.25\\) for CrisisMMD and \\(\\tau = 0.20\\) for ScienceQA based on empirical evaluations to balance precision and recall. See Appendix E for selection details. We construct the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.506 + ], + "angle": 0, + "content": "graph using DeepSeek-R1-70B and implement LightRAG's hybrid retrieval approach with Qwen2.5-7B. For graph construction and multimodal reasoning, we utilize \\(1 \\times\\) NVIDIA A100-80GB GPUs. Task-specific prompts are designed to assist LLMs in multimodal reasoning evaluation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.655, + 0.531 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Multimodal Classification Tasks. We conduct multimodal classification experiments on the CrisisMMD dataset, evaluating both text-only LLMs and multimodal VLMs. Detailed comparative results are provided in Tables 1 and 2. For text-only LLMs, we adopt Qwen2.5-7B as the foundational reasoning model. Remarkably, the VaLiK-enhanced version achieves state-of-the-art (SOTA) performance matching that of the native Qwen2.5-72B model. The image-only KG constructed through VaLiK demonstrates an average accuracy improvement of \\(4.41\\%\\) across tasks, with the text-image variant attaining a \\(4.90\\%\\) enhancement. These improvements significantly surpass the \\(1.22\\%\\) gain obtained by LightRAG using textual KG. We further validate VaLiK's cross-scale applicability through evaluations on Qwen2.5-32B and 72B architectures, observing consistent \\(2.0\\% - 2.5\\%\\) improvements. While not as significant as the 7B model's benefits, this shows that models that have substantial prior knowledge benefit less from external knowledge augmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Unlike text-only LLMs that depend on MMKGs for visual understanding, VLMs primarily benefit from KGs integration through outdated knowledge refreshment. Due to the inherent availability of visual features during inference, VaLiK's performance gains for VLMs remain con-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "986" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.428 + ], + "angle": 0, + "content": "
Method#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Human [47]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
GPT-4 [42]-84.0673.4587.3681.8770.7590.7384.6979.1082.69
CoT (GPT-3) [47]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
CoT (UnifiedQA) [47]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
CoT (GPT-4) [48]1T+85.4872.4490.2782.6571.4992.8986.6679.0483.99
DDCoT [85]175B80.1576.7282.8278.8972.5385.0282.8675.2180.15
Chameleon (ChatGPT) [48]175B+81.6270.6484.0079.7770.8086.6281.8676.5379.93
LG-VQA (BLIP-2) [24]---------86.32
LaVIN-13B [77]---------77.54
BLIP-2 [77]---------74.17
CCOT7B--------76.84
GraphVis [18]7B--------73.18
Qwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
Qwen2.5-72B72B79.6467.1084.9077.5665.0087.9380.2574.8578.37
Qwen2.5-7B (Mmkg) [45]7B73.9866.3778.1871.6564.3079.6576.5168.0373.47
Qwen2.5-7B (Visual Genome) [37]7B76.7867.0478.0974.0566.1979.7278.0869.6875.08
Qwen2.5-7B (VaLiK Text-only)7B84.5474.2486.9182.7472.5390.0384.5180.2882.98
Qwen2.5-7B (VaLiK Image-only)7B79.1471.5479.2777.1669.7283.1480.6573.9678.88
Qwen2.5-7B (VaLiK Text-Image)7B84.1575.1487.6482.9973.1889.6984.4080.9583.16
Qwen2.5-72B (VaLiK Text-Image)72B85.6175.9390.2784.4074.1792.3385.7982.9884.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.438, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Table 3. Performance comparison (\\%) on ScienceQA benchmark. #T-Params denotes trainable parameters. Categories: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG-Cap (image caption), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12). Method groups: (1) Human performance baseline, (2) Zero/Few-shot text-only LLMs, (3) Zero/Few-shot Multimodal VLMs, (4) LLMs enhanced with knowledge graphs for multimodal reasoning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.521, + 0.482, + 0.732 + ], + "angle": 0, + "content": "strained compared to text-only counterparts. We separately applied VaLiK enhancement to Qwen2-VL-72B-Instruct and LLaVA-34B, obtaining distinct improvements: LLaVA-34B achieves accuracy gains of \\(2.41\\%\\) (image-only KG) and \\(3.59\\%\\) (text-image KG), while Qwen2-VL-72B-Instruct shows \\(1.77\\%\\) and \\(2.23\\%\\) improvements respectively under identical configurations. These experimental findings collectively demonstrate that VaLiK effectively extracts valuable signals from the training corpus and enables dynamic knowledge injection into VLMs during inference, thereby substantially alleviating hallucination phenomena. The differential improvements between Qwen2-VL-72B-Instruct and LLaVA-34B further validate the framework's adaptability across model architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Additionally, we analyze the results of LLMs without KG enhancement in the tables, which generally follow the scaling law [36]. However, DeepSeek-R1 shows anomalous behavior. Through testing, we find that its reasoning process may introduce complex information that interferes with its judgment. Furthermore, empirical results show that most baseline models achieve suboptimal performance without fine-tuning. In contrast, VaLiK's automated MMKG construction framework requires no task-specific adaptation yet delivers consistent improvements." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Multimodal Question Answering Tasks. We evalu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.905, + 0.672 + ], + "angle": 0, + "content": "ated multimodal QA performance on the ScienceQA benchmark with Qwen2.5-7B and Qwen2.5-72B as base architectures, augmented by four knowledge sources: Mmkg, Visual Genome, text-only LightRAG and VaLiK. Compared to existing zero-shot/few-shot LLMs that not specifically optimized for scientific QA, our VaLiK-enhanced Qwen2.5-72B achieved SOTA performance on \\(62.5\\%\\) of subtasks, demonstrating particular strengths in multimodal reasoning scenarios requiring cross-modal alignment with an average accuracy gain of \\(6.4\\%\\) over baseline models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Our study identifies a fundamental imbalance between textual and visual knowledge representations in ScienceQA. Text-only KGs (14k entities, 18k relations) exhibit \\(8 \\times\\) denser structured knowledge than image-only counterparts (3k concepts, 1k relations), explaining visual modality underperformance. Despite this gap, vision-KG-augmented Qwen2.5-7B still attains \\(4.16\\%\\) accuracy gains over its non-enhanced version. Notably, our MMKG requires only 489MB storage for complete storage, while the scene graph component2 of Visual Genome alone occupies 739MB. This lightweight construction enables effective reasoning using only textual KG descriptions without raw images in resource-constrained scenarios." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.625, + 0.9 + ], + "angle": 0, + "content": "2Visual Genome" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "987" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.908, + 0.192 + ], + "angle": 0, + "content": "
TypeMethod#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Image-OnlyQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B80.06 (↑3.86)70.30 (↑2.47)80.55 (↑3.28)78.05 (↑3.56)68.43 (↑2.64)83.76 (↑4.74)81.17 (↑3.45)72.71 (↑3.36)78.14 (↑3.42)
+ SV7B79.14 (↓0.92)71.54 (↑1.24)79.27 (↓1.28)77.16 (↓0.89)69.72 (↑1.29)83.14 (↓0.62)80.65 (↓0.52)73.96 (↑1.25)78.88 (↑0.74)
Text-ImageQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B81.88 (↑5.68)73.00 (↑5.17)84.00 (↑6.73)80.55 (↑6.06)70.05 (↑4.26)87.11 (↑8.09)82.01 (↑4.29)77.98 (↑8.63)80.57 (↑5.85)
+ SV7B84.15 (↑2.27)75.14 (↑2.14)87.64 (↑3.64)82.99 (↑2.44)73.18 (↑3.13)89.69 (↑2.58)84.40 (↑2.39)80.95 (↑2.97)83.16 (↑2.59)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.21, + 0.908, + 0.255 + ], + "angle": 0, + "content": "Table 4. Ablation study on ScienceQA benchmark (CVs: CoE-based Vision-Language Models; SV: Similarly Verification). Performance metrics include: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG (image context), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.264, + 0.486, + 0.357 + ], + "angle": 0, + "content": "
TypeMethodTask 1 (%)Task 2 (%)Task 2-Merged (%)
Image-OnlyQwen2.5-7B65.0444.5245.33
+ CVs68.11 (↑3.07)47.00 (↑2.48)46.95 (↑1.62)
+ SV69.52 (↑1.41)49.54 (↑2.54)49.07 (↑2.12)
Text-ImageQwen2.5-7B65.0444.5245.33
+ CVs68.43 (↑3.39)48.61 (↑4.09)48.97 (↑3.64)
+ SV68.90 (↑0.47)50.02 (↑1.41)50.69 (↑1.72)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.375, + 0.461, + 0.39 + ], + "angle": 0, + "content": "Table 5. Ablation study on CrisisMMD with Qwen2.5-7B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.405, + 0.245, + 0.422 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.485, + 0.777 + ], + "angle": 0, + "content": "Our ablation studies on CrisisMMD and ScienceQA demonstrate the specific roles of VaLiK's components. As shown in Table 4 and Table 5, the CVs (CoE-based VLM) module improves accuracy across all settings, with average gains of \\(+3.05\\%\\) on CrisisMMD and \\(+4.63\\%\\) on ScienceQA tasks, validating visual descriptions enhance reasoning. However, the SV (Similarly Verification) module exhibits dual effects: it significantly improves CrisisMMD metrics by pruning redundant textual descriptions, yet slightly degrades ScienceQA's image-only natural science reasoning. We hypothesize this discrepancy arises from dataset characteristics: CrisisMMD's generated captions contain substantially more redundant content, whereas ScienceQA's simpler visual scenes yield shorter descriptions. Pruning these shorter descriptions risks over-removal of critical semantics. Furthermore, different types of KGs influence the effectiveness of the components: CVs achieve greater gains in CrisisMMD's text-image fusion as original text provides complementary context, while SV shows reduced effectiveness, likely due to occasional over-pruning of cross-modal linkages. Nevertheless, both modules collectively enhance performance across configurations, demonstrating their synergistic yet context-sensitive nature." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.259, + 0.804 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Impact of VLM Quantity and Types. We evaluate the impact of varying quantities and types of VLMs on the CVs module. Our experiments reveal that Qwen2-VL generates the most visual descriptions, followed by LLaVA, while BLIP-2 produces the fewest. However, BLIP-2 demonstrates superior capability in extracting critical information" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.264, + 0.712, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.264, + 0.905, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.387, + 0.887, + 0.403 + ], + "angle": 0, + "content": "Figure 4. Impact analysis of VLM quantity on CrisisMMD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.534 + ], + "angle": 0, + "content": "and identifying key entity relationships within images. We therefore adopt BLIP-2 as the primary model, with LLaVA or Qwen2-VL serving as secondary/tertiary components. Adding more VLMs yields diminishing returns, due to limited entities in current images, though we hypothesize their benefits would increase for complex visual scenes with richer semantic content. This phenomenon is empirically validated by our quantitative results in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Computational Costs. Due to space limitations, we provide an overview of VaLiK's computational costs in Appendix F. Our method is significantly more cost-effective than manual annotation or LLM fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.634, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Multimodal reasoning in LLMs is constrained by incomplete knowledge and hallucination artifacts, limitations that persist because textual KGs cannot bridge visual-textual semantics due to their modality isolation. To bridge this gap, we propose VaLiK, a framework for constructing MMKGs through vision-language alignment, eliminating dependency on manual annotations while resolving visual-textual semantic inconsistencies. By integrating a cascade of pretrained VLMs and cross-modal verification, VaLiK converts images into structured knowledge while filtering noise. The resulting graphs enhance LLMs' reasoning with minimal storage overhead. Experiments on multimodal reasoning benchmarks show SOTA performance. VaLiK's modular design supports adaptability across domains, offering a scalable solution for autonomous knowledge synthesis. This work advances multimodal AI systems by enabling efficient integration of visual and textual data." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.191 + ], + "angle": 0, + "content": "The research was supported by Shanghai Artificial Intelligence Laboratory, the National Key R&D Program of China (Grant No. 2022ZD0160201) and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ1100102)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.22 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[1] Mahdi Abavisani, Liwei Wu, Shengli Hu, Joel Tetreault, and Alejandro Jaimes. Multimodal categorization of crisis events in social media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.299, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.369, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[3] Firoj Alam, Ferda Ofli, and Muhammad Imran. Crisismmd: Multimodal twitter datasets from natural disasters. Proceedings of the International AAAI Conference on Web and Social Media, 12(1), 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.426, + 0.485, + 0.577 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736. Curran Associates, Inc., 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[5] Razvan Azamfirei, Sapna R Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[6] Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[7] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):423-443, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[8] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17754-17762, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems (NeurIPS), 33:22243-22255, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[10] Xiang Chen, Ningyu Zhang, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, and Huajun Chen." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Hybrid transformer with multi-level fusion for multimodal knowledge graph completion. In Proceedings of the International Conference on Research and Development in Information Retrieval (SIGIR), pages 904-915, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[12] Xiang Chen, Jingtian Zhang, Xiaohan Wang, Ningyu Zhang, Tongtong Wu, Yuxiang Wang, Yongheng Wang, and Huajun Chen. Continual multimodal knowledge graph construction. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[13] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.412 + ], + "angle": 0, + "content": "[14] Yong Chen, Xinkai Ge, Shengli Yang, Linmei Hu, Jie Li, and Jinwen Zhang. A survey on multimodal knowledge graphs: Construction, completion and applications. Mathematics, 11 (8), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[15] Zhuo Chen, Yichi Zhang, Yin Fang, Yuxia Geng, Lingbing Guo, Xiang Chen, Qian Li, Wen Zhang, Jiaoyan Chen, Yushan Zhu, et al. Knowledge graphs meet multimodal learning: A comprehensive survey. arXiv preprint arXiv:2402.05391, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[16] Shiyao Cui, Jiangxia Cao, Xin Cong, Jiawei Sheng, Quanggang Li, Tingwen Liu, and Jinqiao Shi. Enhancing multimodal entity and relation extraction with variational information bottleneck. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:1274-1285, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[17] Xiaohui Cui, Xiaolong Qu, Dongmei Li, Yu Yang, Yuxun Li, and Xiaoping Zhang. Mkgcn: Multi-modal knowledge graph convolutional network for music recommender systems. *Electronics*, 12(12), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[18] Yihe Deng, Chenchen Ye, Zijie Huang, Mingyu Derek Ma, Yiwen Kou, and Wei Wang. Graphvis: Boosting llms with visual knowledge graph integration. In Advances in Neural Information Processing Systems, pages 67511-67534. Curran Associates, Inc., 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[19] Tausif Diwan, G. Anirudh, and Jitendra V. Tembhurne. Object detection using yolo: challenges, architectural successors, datasets and applications. Multimedia Tools Appl., 82 (6):9243-9275, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[21] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Quan Fang, Xiaowei Zhang, Jun Hu, Xian Wu, and Changsheng Xu. Contrastive multi-modal knowledge graph representation learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8983-8996, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[23] Duoduo Feng, Xiangteng He, and Yuxin Peng. Mkvse: Multimodal knowledge enhanced visual-semantic embedding for image-text retrieval. ACM Trans. Multimedia Comput. Commun. Appl., 19(5), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[24] Deepanway Ghosal, Navonil Majumder, Roy Ka-Wei Lee, Rada Mihalcea, and Soujanya Poria. Language guided visual question answering: Elevate your multimodal language model using knowledge-enriched prompts. arXiv preprint arXiv:2310.20159, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[25] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[26] Yunfei Guo, Fei Yin, Xiao-hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21604-21613, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[27] ZIRUI GUO, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. LightRAG: Simple and fast retrieval-augmented generation, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[28] Lavdim Halilaj, Juergen Luettin, Sebastian Monka, Cory Henson, and Stefan Schmid. Knowledge graph-based integration of autonomous driving datasets. International Journal of Semantic Computing, 17(02):249-271, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[30] Yang Hu, Guihua Wen, Adriane Chapman, Pei Yang, Mingnan Luo, Yingxue Xu, Dan Dai, and Wendy Hall. Graph-based visual-semantic entanglement network for zero-shot image recognition. IEEE Transactions on Multimedia, 24: 2473-2487, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[31] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. LLM-adapters: An adapter family for parameter-efficient fine-tuning of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5254-5276, Singapore, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[32] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[33] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[34] Adam Tauman Kalai and Santosh S. Vempala. Calibrated language models must hallucinate. In Proceedings of the 56th Annual ACM Symposium on Theory of Computing, page 160-171, New York, NY, USA, 2024. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Amar Viswanathan Kannan, Dmitriy Fradkin, Ioannis Akrotirianakis, Tugba Kulahcioglu, Arquimedes Canedo," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "Aditi Roy, Shih-Yuan Yu, Malawade Arnav, and Mohammad Abdullah Al Faruque. Multimodal knowledge graph for deep learning papers and code. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 3417-3420, New York, NY, USA, 2020. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[36] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[37] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10767-10782, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[39] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[40] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning (ICML), pages 19730–19742, 2023. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[41] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466. Curran Associates, Inc., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[43] Junming Liu, Yanting Gao, Siyuan Meng, Yifei Sun, Aoqi Wu, Yufei Jin, Yirong Chen, Ding Wang, and Guosun Zeng. Mosaic: Data-free knowledge distillation via mixture-of-experts for heterogeneous distributed environments. arXiv preprint arXiv:2505.19699, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[44] Junming Liu, Guosun Zeng, Ding Wang, Yanting Gao, and Yufei Jin. Fedrecon: Missing modality reconstruction in distributed heterogeneous environments. arXiv preprint arXiv:2504.09941, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[45] Ye Liu, Hui Li, Alberto Garcia-Duran, Mathias Niepert, Daniel Onoro-Rubio, and David S Rosenblum. Mmkg: multi-modal knowledge graphs. In The Semantic Web: 16th International Conference, ESWC 2019, Portoroz, Slovenia, June 2-6, 2019, Proceedings 16, pages 459-474. Springer, 2019. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[47] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, pages 2507–2521. Curran Associates, Inc., 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[48] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In Advances in Neural Information Processing Systems, pages 43447-43478. Curran Associates, Inc., 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[49] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. In Advances in Neural Information Processing Systems, pages 29615-29627. Curran Associates, Inc., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[50] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[51] Runqing Miao, Qingxuan Jia, Fuchun Sun, Gang Chen, Haiming Huang, and Shengyi Miao. Semantic representation of robot manipulation with knowledge graph. Entropy, 25(4), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[52] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14420-14431, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[53] Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[54] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[55] Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Knowledge-aware multi-modal adaptive graph convolutional networks for fake news detection. ACM Trans. Multimedia Comput. Commun. Appl., 17(3), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[57] Brian Reily, Christopher Reardon, and Hao Zhang. Representing multi-robot structure through multimodal graph embedding for the selection of robot teams. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5576–5582, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[58] Joshua Robinson, Christopher Michael Ryting, and David Wingate. Leveraging large language models for multiple choice question answering. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[59] Hrituraj Singh, Anshul Nasery, Denil Mehta, Aishwarya Agarwal, Jatin Lamba, and Balaji Vasan Srinivasan. MIMOQA: Multimodal input multimodal output question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5317-5332, Online, 2021. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[60] Yaoxian Song, Penglei Sun, Haoyu Liu, Zhixu Li, Wei Song, Yanghua Xiao, and Xiaofang Zhou. Scene-driven multimodal knowledge graph construction for embodied ai. IEEE Transactions on Knowledge and Data Engineering, 36(11): 6962-6976, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[61] Rui Sun, Xuezhi Cao, Yan Zhao, Junchen Wan, Kun Zhou, Fuzheng Zhang, Zhongyuan Wang, and Kai Zheng. Multimodal knowledge graphs for recommender systems. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 1405-1414, New York, NY, USA, 2020. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[62] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[64] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, ADITHYA JAIRAM IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[67] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton van den Hengel. Fvqa: Fact-based visual question an" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "991" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "swering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2413-2427, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[69] Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In Proceedings of the 31st ACM International Conference on Multimedia, page 2391–2399, New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[70] Yuequn Wang, Liyan Dong, Hao Zhang, Xintao Ma, Yongli Li, and Minghui Sun. An enhanced multi-modal recommendation based on alternate training with knowledge graph representation. IEEE Access, 8:213012-213026, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.418 + ], + "angle": 0, + "content": "[71] Tao Wu, Mengze Li, Jingyuan Chen, Wei Ji, Wang Lin, Jinyang Gao, Kun Kuang, Zhou Zhao, and Fei Wu. Semantic alignment for multimodal large language models. In Proceedings of the 32nd ACM International Conference on Multimedia, page 3489-3498, New York, NY, USA, 2024. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.475 + ], + "angle": 0, + "content": "[72] Yike Wu, Nan Hu, Guilin Qi, Sheng Bi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering. arXiv preprint arXiv:2309.11206, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[73] Ziyang Xiao, Dongxiang Zhang, Yangjun Wu, Lilin Xu, Yuan Jessica Wang, Xiongwei Han, Xiaojin Fu, Tao Zhong, Jia Zeng, Mingli Song, and Gang Chen. Chain-of-experts: When LLMs meet complex operations research problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[74] Dexuan Xu, Yanyuan Chen, Jieyi Wang, Yue Huang, Hanpin Wang, Zhi Jin, Hongxing Wang, Weihua Yue, Jing He, Hang Li, and Yu Huang. MLeVLM: Improve multi-level progressive capabilities based on multimodal large language model for medical visual question answering. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4977-4997, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[75] Peng Xu, Xiatian Zhu, and David A. Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[76] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[77] Xiaocui Yang, Wenfang Wu, Shi Feng, Ming Wang, Daling Wang, Yang Li, Qi Sun, Yifei Zhang, Xiaoming Fu, and Soujanya Poria. Mm-bigbench: Evaluating multimodal models on multimodal content comprehension tasks. arXiv preprint arXiv:2310.09036, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[78] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[79] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[80] Jingtong Yue, Zhiwei Lin, Xin Lin, Xiaoyu Zhou, Xiangtai Li, Lu Qi, Yongtao Wang, and Ming-Hsuan Yang. RobuR-CDet: Enhancing robustness of radar-camera fusion in bird's eye view for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[81] Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Huajun Chen, and Wen Zhang. Mygo: Discrete modality information as fine-grained tokens for multi-modal knowledge graph completion. CoRR, abs/2404.09468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[83] Jiabao Zhao, Xin Lin, Jie Zhou, Jing Yang, Liang He, and Zhaohui Yang. Knowledge-based fine-grained classification for few-shot learning. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[84] Changmeng Zheng, Junhao Feng, Ze Fu, Yi Cai, Qing Li, and Tao Wang. Multimodal relation extraction with efficient graph alignment. In Proceedings of the 29th ACM International Conference on Multimedia, page 5298-5306, New York, NY, USA, 2021. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[85] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. In Advances in Neural Information Processing Systems, pages 5168-5191. Curran Associates, Inc., 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[86] Juncheng Zheng, Meiyu Liang, Yang Yu, Yawen Li, and Zhe Xue. Knowledge graph enhanced multimodal transformer for image-text retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 70-82, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[87] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.782 + ], + "angle": 0, + "content": "[88] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.905, + 0.854 + ], + "angle": 0, + "content": "[89] Xiangru Zhu, Zhixu Li, Xiaodan Wang, Xueyao Jiang, Penglei Sun, Xuwu Wang, Yanghua Xiao, and Nicholas Jing Yuan. Multi-modal knowledge graph construction and application: A survey. IEEE Transactions on Knowledge and Data Engineering, 36(2):715-735, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.855, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[90] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "992" + } + ] +] \ No newline at end of file diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_origin.pdf b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b7d22a10629fc751a552514748938c327db67071 --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/7ddf6f08-8bf9-46ae-a11c-83b7c2afd49c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:205002f882e74d9a89f419c4b5aad5317c0db25b5de3d34dfa8d71f8b1eecdfc +size 6450824 diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/full.md b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..aaa16aab14bab318173a2c70353793ef9b17c82b --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/full.md @@ -0,0 +1,370 @@ +# Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning + +Junming Liu $^{1,2}$ , Siyuan Meng $^{2,3}$ , Yanting Gao $^{1}$ , Song Mao $^{2}$ , Pinlong Cai $^{2}$ , + +Guohang Yan $^{2}$ , Yirong Chen $^{2,4}$ , Zilin Bian $^{5}$ , Ding Wang $^{2*}$ , Botian Shi $^{2}$ + +$^{1}$ Tongji University $^{2}$ Shanghai Artificial Intelligence Laboratory + +$^{3}$ East China Normal University $^{4}$ Stanford University $^{5}$ New York University + +liu_junming6917@tongji.edu.cn wangding@pjlab.org.cn + +# Abstract + +Multimodal reasoning in Large Language Models (LLMs) struggles with incomplete knowledge and hallucination artifacts, challenges that textual Knowledge Graphs (KGs) only partially mitigate due to their modality isolation. While Multimodal Knowledge Graphs (MMKGs) promise enhanced cross-modal understanding, their practical construction is impeded by semantic narrowness of manual text annotations and inherent noise in visual-semantic entity linkages. In this paper, we propose Vision-align-to-Language integrated Knowledge Graph (VaLiK), a novel approach for constructing MMKGs that enhances LLMs reasoning through cross-modal information supplementation. Specifically, we cascade pre-trained Vision-Language Models (VLMs) to align image features with text, transforming them into descriptions that encapsulate image-specific information. Furthermore, we developed a cross-modal similarity verification mechanism to quantify semantic consistency, effectively filtering out noise introduced during feature alignment. Even without manually annotated image captions, the refined descriptions alone suffice to construct the MMKG. Compared to conventional MMKGs construction paradigms, our approach achieves substantial storage efficiency gains while maintaining direct entity-to-image linkage capability. Experimental results on multimodal reasoning tasks demonstrate that LLMs augmented with VaLiK outperform previous state-of-the-art models. Our code is published at https://github.com/Wings-Of-Disaster/VaLiK. + +# 1. Introduction + +Recent advancements in Large Language Models (LLMs) [2, 9, 25, 65] have demonstrated their superiority and versatility across various Natural Language Reasoning (NLR) tasks [8, 43, 53, 58]. To enhance LLMs into the + +![](images/dfacec5607ec5e36626acfb0e0ac18a11bebb24a4130fb85486e59c4e58c6e72.jpg) +Figure 1. (a) Training entity extraction models relies on extensive fine-grained annotations, increasing labeling costs. More examples are provided in Appendix B. (b) Capturing implicit semantic associations demands abstract comprehension or logical inference. + +realm of multimodal reasoning, researchers [64, 71, 74, 79] have endeavored to equip these models with multimodal capabilities, as evidenced by advancements in Multimodal Large Language Models (MLLMs) such as BLIP-2 [40], GPT-4o [32], Janus-Pro [13], among others. Despite their notable progress, these models often experience hallucinations [5, 34], primarily arising from knowledge deficiencies due to incomplete or obsolete information. + +Fine-tuning LLMs demands prohibitive computational costs [31]. While text-based Knowledge Graphs (KGs) have partially addressed this limitation by efficient real-time updates [6, 62, 72], they are still restricted by modal isolation, which hinders cross-modal reasoning, as detailed in Appendix A. To bridge this semantic fragmentation, Multimodal Knowledge Graphs (MMKGs) have been developed as unified representational frameworks [10, 33, 38, 45]. + +However, constructing robust MMKGs faces two primary obstacles [15, 89]. First, the lack of large-scale fine-grained entity-image corpora makes it infeasible to train high-quality entity extractors, significantly constraining scalability, as illustrated in Figure 1a. Second, conventional visual relation detectors primarily identify superficial spatial interactions instead of semantic relations consistent with KGs, while frequently hallucinating implausible connections that corrupt graph integrity, as shown in Figure 1b. + +In this paper, we propose VaLiK, short for Vision-align-to-Language integrated Knowledge Graph, a novel framework designed to empower LLMs with advanced multimodal reasoning. Unlike traditional methods that rely on text annotations for training extraction models and the knowledge construction process [54], VaLiK adopts a annotation-free approach to MMKGs construction. Specifically, we first employ several pretrained Vision-Language models (VLMs), designed based on Chain-of-Experts (CoE) principles [73], to convert visual inputs into image-specific textual descriptions through cross-modal feature alignment. This procedure eliminates the need for manually annotated image captions in both the knowledge extraction and construction phases while preserving visual details typically missing in generic text descriptions. Moreover, in contrast to existing relation detection methods that require predefined label taxonomies [16, 60, 81, 84], VaLiK excels at extracting profound semantic relationships that are both KG-compatible and capture novel associations beyond training supervision. While VLMs enable cross-modal reasoning and interpretation, they introduce spurious relational noise through hallucinated inter-modal attributions, as depicted in Figure 2. We address this limitation through cross-modal similarity recalibration, strategically filtering inconsistent information while preserving valid semantic correspondences. Finally, the purified descriptions are systematically organized into MMKGs via LLM-driven symbolic structuring [27], bridging visual and textual domains with factual consistency. + +To thoroughly evaluate the VaLiK method, we conduct a comprehensive assessment across two critical multimodal benchmarks: multimodal classification (tested on the CrisisMMD dataset [3]) and multimodal question answering (evaluated via the ScienceQA benchmark [47]). The experiments span diverse LLM architectures and MMKG construction techniques to ensure the framework's robustness. + +![](images/70a5e7b33cf2f7b408d7e9bbab0216de573f5ba53a84b04153ef68002435eb2d.jpg) +Figure 2. Feature-aligned descriptions from VLMs introduce redundant and inaccurate relationship patterns. + +The experimental results demonstrate that the MMKGs constructed by VaLiK achieve superior multimodal reasoning performance in LLMs while requiring substantially less storage than conventional approaches. More importantly, the proposed approach retains direct entity-to-image linkage capabilities even with the compressed graph structure. + +In summary, VaLiK is the first framework that enables end-to-end, annotation-free, zero-shot, and storage-efficient multimodal knowledge construction with high adaptability and scalability. Our key contributions include: + +- To the best of our knowledge, VaLik is the first end-to-end framework to build Annotation-Free MMKGs to improve LLMs' multimodal reasoning capabilities, effectively eliminating the need for manually annotated textual material and enabling a completely autonomous multimodal knowledge generation process. +- We offer an innovative zero-shot method for constructing MMKG that captures deep semantic connections beyond traditional predetermined labels with an effective verification system that guarantees the accuracy of these relationships. The knowledge distillation paradigm greatly decreases storage while maintaining semantic integrity. +- We develop a highly modular and extensible architecture that allows VaLiK to effortlessly incorporate new models and workflows for specialized domain tasks, facilitating rapid adaptation to diverse application scenarios without incurring expensive system changes. + +# 2. Related Work + +# 2.1. Multimodal Knowledge Graphs + +The principal advantage of MMKGs resides in their multimodal integration beyond conventional KGs. By linking entities with corresponding visual or textual data, MMKGs introduce valuable visual and textual information to the knowledge base, substantially advancing multimodal reasoning capabilities. This combination addresses core challenges in tasks that inherently demand multimodal synergy + +like autonomous driving [26, 28], image-text retrieval [23, 86] and robotic manipulation [51, 57]. However, constructing trustworthy MMKGs with minimal manual effort remains a critical challenge. Recent studies have proposed innovative strategies to enhance MMKG reliability and utility. For instance, Chen et al. [12] proposed MSPT, a framework addressing continual MMKG construction through gradient modulation for balanced multimodal learning and attention distillation to mitigate catastrophic forgetting. Song et al. [60] developed Scene-MMKG, integrating knowledge engineering with large language models to improve robotic manipulation by resolving data sparsity and knowledge uncertainty. Wang et al. [69] introduced TIVA-KG, the first quad-modal knowledge graph spanning text, image, video, and audio with triplet grounding, empirically validating its effectiveness in downstream tasks. While these advances enhance multimodal reasoning capabilities, their efficacy remains rooted in resource-intensive paradigms, requiring extensively annotated datasets for knowledge acquisition. + +# 2.2. Knowledge-Augmented Multimodal Learning + +Multimodal learning has seen significant progress in aligning and integrating information across different data modalities [7, 44, 75]. The incorporation of structured knowledge through MMKGs further enhances these approaches, improving the reasoning capabilities and generalization across a variety of domains, such as visual question answering [50, 59, 67], recommendation systems [17, 61, 70], and classification [30, 55, 83]. Methods like GraphAdapter's dual-KG adaptation [41] and contrastive multi-relational encoding with KGs [22] inject external knowledge into models, refining their performance and improving their capability to handle complex tasks. Additionally, Lee et al. [38] proposed MR-MKG, a novel framework that constructs task-specific MMKGs to enhance multimodal reasoning in LLMs. These knowledge-augmented paradigms demonstrate superior cross-modal semantic grounding compared to unimodal approaches [14, 35]. However, their reliance on preconstructed MMKGs often leads to domain discrepancies, where generic knowledge schemas misalign with task-specific reasoning patterns, ultimately limiting contextual precision in target applications. + +# 2.3. Multimodal Large Language Models + +The limitations of text-only LLMs in meeting increasingly complex demands have spurred extensive research [78, 82, 85] into developing LLMs capable of effectively processing and reasoning over multimodal inputs. Current research predominantly employs adapter or projection layers to connect the embedding spaces of various modality-specific encoders with the textual embedding space of LLMs [38]. For instance, foundational models like CLIP [56] and BLIP [39] pioneered cross-modal alignment by jointly training vision + +and text encoders to map images and text into a shared embedding space. Building on this, LLaVA [42] and Flamingo [4] advanced the field by integrating visual encoders with LLMs, enabling more nuanced multimodal understanding and generation. More recently, Gemini [63], Qwen2-VL [68] and GPT-4o [32] have further pushed the boundaries by scaling up multimodal pretraining and introducing sophisticated mechanisms for cross-modal interaction. However, multimodal LLMs remain prone to hallucinations. While they enhance cross-modal alignment, they neither acquire new knowledge nor avoid introducing noise through integration. To address these limitations, VaLiK "uses the master's tools to refine the master's craft," first constructing MMKGs via MLLMs and then leveraging them to enhance MLLMs' reasoning capabilities. + +# 3. Method + +In this section, we present the technical details of VaLiK. VaLiK introduces a novel expansion-reduction paradigm for visual knowledge extraction. The architecture initially organizes several VLMs with distinct knowledge domains, designed based on CoE principles [73], to produce comprehensive textual descriptions encompassing hierarchical visual details. A cross-modal similarity verification mechanism then iteratively filters out noisy tokens through cross-modal alignment while preserving semantically salient elements. This optimization-style approach eliminates external textual dependencies while enabling effective MMKG construction. VaLiK's framework is shown in Figure 3. + +# 3.1. CoE-based Visual to Language Modeling + +Recent entity detection techniques [19, 80, 90] have been widely adopted for entity and relation extraction in MMKG construction. However, these methods are inherently limited by predefined categorical boundaries, lacking the capacity to recognize visual concepts outside their training vocabulary. In contrast, VLMs pretrained on web-scale corpora [11, 40, 88] exhibit broader recognition capabilities through exposure to diverse visual concepts. + +We therefore leverage pretrained VLMs to extract comprehensive visual information. This process removes the necessity for detailed fine-grained data typically required to train specialized recognition models. The generalized vision to language conversion pipeline can be formalized as: + +$$ +S = \mathcal {D} _ {\text {t e x t}} \left(\mathcal {A} \left(\mathcal {E} _ {\text {v i s}} (I)\right)\right), \tag {1} +$$ + +where $I$ denotes for the input image, $\mathcal{E}_{\mathrm{vis}}$ denotes the visual encoder extracting visual features, $\mathcal{A}$ carries out cross-modal feature alignment and interaction, and $\mathcal{D}_{\mathrm{text}}$ generates textual tokens through autoregressive decoding. The resulting visual description $S = \{w_{1},\dots,w_{n}\}$ emerges from this multi-stage processing. + +![](images/f146c2291d18be7bbdd2100fa3e32b5e7a21e72aea45eeb431f07813670ed63c.jpg) +Figure 3. The pipeline of VaLiK: First, large-scale visual descriptions are generated using CoE-based VLMs. Then, a similarity verification mechanism is used to prune irrelevant information. Finally, MMKGs are constructed using LLMs based on LightRAG. The constructed MMKGs can assist LLMs in multimodal reasoning, alleviating the hallucination issues caused by incomplete knowledge. + +However, quantitative analysis uncovers considerable discrepancies between machine-generated and human-annotated descriptions [87]. As an illustration, while utilizing BLIP-2 [40] to generate sample captions, we noted that the model outputs are markedly concise and devoid of visual specifics, as detailed in Appendix C. To bridge this gap, we implement CoE enhanced generation through cascade VLMs processing. At iteration step $t$ , each expert $E_{i}$ receives both the original visual signals $I$ and the contextual output from the preceding expert $E_{i - 1}$ : + +$$ +\mathcal {S} _ {i} ^ {(t)} = E _ {i} \left(I, \mathcal {S} _ {i - 1} ^ {(t - 1)}\right), \tag {2} +$$ + +where $S_{i - 1}^{(t - 1)}$ denotes the description from expert $E_{i - 1}$ at step $t - 1$ , with $S_0^{(t)}\coloneqq \emptyset$ for initialization. + +Specifically, each expert $E_{i}$ implements a unified visual-language processing task: + +# 1. Visual Feature Extraction: + +$$ +\mathbf {V} _ {i} = \operatorname {E n c} _ {\text {v i s}} ^ {i} (I) \in \mathbb {R} ^ {d _ {v} \times N _ {p}}, \tag {3} +$$ + +where $\mathsf{Enc}_{\mathrm{vis}}^i$ denotes established visual encoder [20, 29, 46] producing $N_{p}$ patch embeddings with dimension $d_v$ . + +# 2. Cross-Modal Interaction and Generation: + +VLMs integrate pretrained learnable query embeddings $\mathbf{Q}_i\in \mathbb{R}^{d_q\times L_q}$ to interact with visual features $\mathbf{V}_i\in$ $\mathbb{R}^{d_v\times N_p}$ via cross-attention [66]: + +$$ +\begin{array}{l} \mathbf {H} _ {i} = \operatorname {C r o s s A t t n} \left(\mathbf {Q} _ {i}, \mathbf {V} _ {i}\right) \\ = \operatorname {s o f t m a x} \left(\frac {\mathbf {Q} _ {i} \mathbf {W} _ {q} ^ {i} \left(\mathbf {V} _ {i} \mathbf {W} _ {k} ^ {i}\right) ^ {\top}}{\sqrt {d _ {k}}}\right) \mathbf {V} _ {i} \mathbf {W} _ {v} ^ {i}, \tag {4} \\ \end{array} +$$ + +where $\mathbf{W}_q^i\in \mathbb{R}^{d_q\times d_k}$ , $\mathbf{W}_k^i$ , $\mathbf{W}_v^i\in \mathbb{R}^{d_v\times d_k}$ , and $L_{q}$ denotes the predefined query length. Cross-attention serves + +as a prevalent approach, while other interaction strategies coexist [4]. The adopted VLMs in our implementation primarily rely on this approach for modality fusion. + +# 3. Text Generation: + +The text encoder $\mathsf{Enc}_{\mathrm{text}}^{i}$ first processes the preceding expert's output $S_{i - 1}^{(t - 1)}$ into latent features: + +$$ +\mathbf {P} _ {i} = \operatorname {E n c} _ {\text {t e x t}} ^ {i} \left(S _ {i - 1} ^ {(t - 1)}\right) \in \mathbb {R} ^ {d _ {t} \times L}. \tag {5} +$$ + +Subsequently, the text decoder $\mathrm{Dec}_{\mathrm{text}}^{i}$ synthesizes the final output $S_{i}^{(t)}$ by jointly conditioning on $\mathbf{P}_i$ and $\mathbf{H}_i$ : + +$$ +\mathcal {S} _ {i} ^ {(t)} = \operatorname {D e c} _ {\text {t e x t}} ^ {i} \left(\mathbf {P} _ {i}, \mathbf {H} _ {i}\right) = \left\{w _ {1} ^ {(t, i)}, \dots , w _ {m} ^ {(t, i)} \right\}. \tag {6} +$$ + +Ultimately, the final textual description $S_N^{(C)}$ is obtained after $C$ iteration steps through $N$ cascaded experts. + +# 3.2. Cross-Modal Similarity Verification + +To address noise in VLM-generated captions, we design a sliding window mechanism with semantic consistency verification. This method ensures that only relevant and semantically consistent segments are retained in the final description. Let $W_{k}$ denote the $k$ -th window containing $m$ consecutive tokens $\{w_{km + 1},\dots ,w_{(k + 1)m}\}$ . For each window, we compute its cross-modal similarity score: + +$$ +\alpha_ {k} = \frac {\operatorname {E n c} _ {\text {v i s}} (I) \cdot \operatorname {E n c} _ {\text {t e x t}} \left(W _ {k}\right)}{\| \operatorname {E n c} _ {\text {v i s}} (I) \| \| \operatorname {E n c} _ {\text {t e x t}} \left(W _ {k}\right) \|}, \tag {7} +$$ + +where $\mathsf{Enc}_{vis/text}(\cdot)$ adopts a lightweight CLIP [58] encoder-decoder with frozen parameters for efficient processing. The similarity score $\alpha_{k}$ lies within the range [0, 1], with higher values indicating a stronger alignment between the visual and textual information. + +After calculating the cross-modal similarity for each window, we employ an empirical threshold $\tau$ to filter out low-similarity windows. This threshold helps to identify and discard noisy or irrelevant sections of the generated caption that do not align well with the visual content, thereby reducing the impact of inaccurate or misleading descriptions. Formally, for each window $W_{k}$ , if $\alpha_{k} < \tau$ , the window is discarded as noise. This process effectively prunes windows with low similarity scores, ensuring that only semantically meaningful segments remain. The final denoised description $\hat{S}$ is obtained by concatenating all windows $W_{k}$ for which $\alpha_{k} \geq \tau$ : + +$$ +\hat {S} = \bigcup_ {\alpha_ {k} \geq \tau} W _ {k}. \tag {8} +$$ + +Our window size $m$ is flexibly determined and generally adapts dynamically to natural sentence segmentation. + +# 3.3. MMKG Construction for Enhanced Reasoning + +LLMs have become increasingly popular for identifying entities, relationships, and attributes within a corpus, which are then organized into a KG. The strength of LLM-based KG generation lies in its capacity to leverage the vast amount of knowledge encoded within these models, allowing them to detect complex and nuanced patterns across diverse data sources. This approach eliminates the need for manual annotation, enabling a highly scalable and domain-adaptive process suitable for a wide range of applications. + +We begin by refining the generated textual description $\hat{S}$ (VLM-based information), which is then optionally concatenated with any available external textual knowledge $T$ to form the input for KG generation. This combined input is used to generate MMKGs with the help of a LLM [21, 27], leveraging its capacity for multi-hop reasoning and dynamic knowledge integration. + +$$ +\mathcal {G} = \operatorname {L L M} (\hat {S} \oplus T), \tag {9} +$$ + +where $\oplus$ denotes optional concatenation based on the availability of $T$ . The resulting graph $\mathcal{G}$ captures both visual and textual relationships inferred by the LLM. + +We define $\mathcal{G}$ as a set of triplets: + +$$ +\mathcal {G} = \{(h, r, t) \mid h, t \in \mathcal {E}, r \in \mathcal {R} \}, \tag {10} +$$ + +where $\mathcal{E}$ and $\mathcal{R}$ denote the sets of entities and relations. Entities include objects or concepts from the image or external text, while relations describe connections such as "is a type of," "part of," or "has property." Each triplet $(h,r,t)$ links a head entity $h$ and a tail entity $t$ via relation $r$ . + +Multimodal Reasoning Enhancement. To support multimodal reasoning, we retrieve relevant triplets from $\mathcal{G}$ through structural patterns during LLMs inference: + +$$ +\mathcal {G} _ {q} = \operatorname {R e t r i e v e} (q, \mathcal {G}), \tag {11} +$$ + +where $\text{Retrieve}(\cdot)$ denotes a retrieval strategy that identifies subgraphs relevant to the query for reasoning. Detailed retrieval strategies are described in Appendix D. + +The augmented prompt integrates multimodal evidence: + +$$ +p _ {\mathrm {a u g}} = q \left\|\left(\bigcup_ {(h, r, t) \in \mathcal {G} _ {q}} [ h ] \rightarrow r \rightarrow [ t ]\right). \right. \tag {12} +$$ + +Note that we incorporate the storage locations of images in the database during MMKGs construction, enabling the MMKGs to link to visual data. VaLiK enables text-only LLMs to perform multimodal reasoning through $\mathcal{G}$ 's visual associations, while VLMs refresh knowledge representations by jointly injecting both visual and textual information, significantly mitigating hallucination risks. + +# 4. Experiment + +# 4.1. Setups + +Evaluation Datasets. We evaluate VaLiK on two multimodal reasoning benchmarks with distinct characteristics: + +- CrisisMMD [3]. This real-world disaster response dataset includes around 35,000 noisy social media postings with paired images and text, each annotated for seven catastrophe categories and four severity levels. Its realistic user-generated content with natural noise and implicit modality correlations provides a rigorous testbed for zero-shot adaptation, with good performance indicating practical relevance in real-world crisis scenarios. + +- ScienceQA [47]. This dataset contains 21,208 multimodal science questions combining textual and visual contexts, with $48.7\%$ of instances containing images. Questions span physics, chemistry, and biology domains, requiring cross-modal reasoning between textual concepts and visual diagrams. Additionally, ScienceQA offers image captions to aid text-only LLMs in reasoning, allowing a comparison of unimodal approaches. + +Task Formulation. For CrisisMMD, we define three multimodal classification tasks1: (1) binary information relevance filtering, (2) fine-grained humanitarian category recognition, and (3) a consolidated taxonomy with merged categories to reduce label complexity. We omit the unimodal damage assessment to focus on multimodal aspects. For ScienceQA, we follow the original evaluation using multiple metrics: question types, contextual modalities, and educational stages. Performance is assessed through accuracy percentage across these categories. + +Baselines. We conduct a comprehensive evaluation of text-only LLMs, multimodal VLMs, and KGs that enhance LLMs in multimodal reasoning. + +- For CrisisMMD, we compare text-only LLMs using few-shot prompting (LLaMA-2 [65], GPT-4 [2], + +
TaskText-only LLMsKG-Enhanced LLMs
LLaMA-2GPT-4DeepSeek-R1Qwen2.5LightRAGVaLiK
7B13B70B-7B8B32B70B7B32B72BText-onlyImage-onlyText-Image
Task 162.3263.8063.1566.8367.2363.3163.6165.5365.0467.2867.9567.4969.5268.90
Task 218.3221.8228.8747.2526.5325.4924.7721.0544.5246.9450.5145.1149.5450.02
Task 2 Merged21.4533.1536.8949.4425.8523.5621.5525.5745.3347.0750.2945.9449.0750.69
+ +Table 1. The performance evaluation of text-only LLMs using few-shot prompting without any fine-tuning on the training set. As these models handle text only, test data is formatted as unimodal text for compatibility. In our implementations, both LightRAG and VaLiK adopt Qwen2.5-7B as the base reasoning model. Bold indicates the highest value, and underline indicates the second highest. + +
TaskMultimodal VLMsKG-Enhanced LLMs
CLIPLLaVABLIP-2GPT-4oQwen2-VLVaLiK
ViT-L/147B13B34BFlan-T5-XLOPT-2B-I7B-I72B-I*#+~
Task 143.3654.0060.5856.4461.2938.6268.2047.5662.4565.8060.7868.4461.1168.89
Task 217.8828.0120.1425.1540.8614.2647.587.6032.6847.2125.8048.8827.2349.78
Task 2-M20.7930.6123.4425.0740.7214.2749.557.4234.2048.2827.3149.2729.0949.31
+ +Table 2. The performance of multimodal VLMs and KG-enhanced LLMs. The -I suffix denotes instruction-tuned variants. Symbol markers denote KG types and models: the asterisk (*) represents image-only KG with LLaVA-34B, hash (#) indicates image-only KG using Qwen2-VL-72B-I, plus (+) denotes text-image KG with LLaVA-34B, and tilde (\~) shows text-image KG using Qwen2-VL-72B-I. + +DeepSeek-R1 [25], Qwen-2.5 [76]) and multimodal VLMs (CLIP [56], LLaVA [42], GPT-4o [32], Qwen2-VL [68], BLIP-2 [40]). + +- For ScienceQA, we compare models for general domains in zero/few-shot settings, including text-only LLMs (GPT Model [47], CoT [47], DDCoT [85]), multimodal VLMs (LG-VQA [24], LaVIN [49], BLIP-2, CCOT [52], GraphVis [18]) and Tool-LLM Chameleon [48]. These models are not specifically fine-tuned for scientific tasks, ensuring a fair evaluation of generalization capabilities. +- We further compare the multimodal reasoning performance of LLMs assisted by KGs, evaluating text-based KGs built with LightRAG [27], and pre-constructed MMKGs such as Visual Genome [37] and Mmkg [45]. + +Implementation. For MMKG construction, we design a chain of VLMs including BLIP-2, LLaVA, and Qwen2-VL, with the CLIP-ViT-L/14 for pruning. Stronger or additional VLMs could be employed to enhance performance if more computational resources are available. We use the entire training set as the knowledge base and construct MMKGs from the extracted descriptions based on the LightRAG framework. In comparative experiments, the LightRAG method we evaluate utilizes only textual data, while VaLiK employs two configurations: (1) fully image-generated text descriptions (Image-only), and (2) original text combined with image-generated text (Text-Image). Dynamic window partitioning based on sentence length ensures syntactically coherent pruning results. Similarity thresholds are set to $\tau = 0.25$ for CrisisMMD and $\tau = 0.20$ for ScienceQA based on empirical evaluations to balance precision and recall. See Appendix E for selection details. We construct the + +graph using DeepSeek-R1-70B and implement LightRAG's hybrid retrieval approach with Qwen2.5-7B. For graph construction and multimodal reasoning, we utilize $1 \times$ NVIDIA A100-80GB GPUs. Task-specific prompts are designed to assist LLMs in multimodal reasoning evaluation. + +# 4.2. Main Results + +Multimodal Classification Tasks. We conduct multimodal classification experiments on the CrisisMMD dataset, evaluating both text-only LLMs and multimodal VLMs. Detailed comparative results are provided in Tables 1 and 2. For text-only LLMs, we adopt Qwen2.5-7B as the foundational reasoning model. Remarkably, the VaLiK-enhanced version achieves state-of-the-art (SOTA) performance matching that of the native Qwen2.5-72B model. The image-only KG constructed through VaLiK demonstrates an average accuracy improvement of $4.41\%$ across tasks, with the text-image variant attaining a $4.90\%$ enhancement. These improvements significantly surpass the $1.22\%$ gain obtained by LightRAG using textual KG. We further validate VaLiK's cross-scale applicability through evaluations on Qwen2.5-32B and 72B architectures, observing consistent $2.0\% - 2.5\%$ improvements. While not as significant as the 7B model's benefits, this shows that models that have substantial prior knowledge benefit less from external knowledge augmentation. + +Unlike text-only LLMs that depend on MMKGs for visual understanding, VLMs primarily benefit from KGs integration through outdated knowledge refreshment. Due to the inherent availability of visual features during inference, VaLiK's performance gains for VLMs remain con- + +
Method#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Human [47]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
GPT-4 [42]-84.0673.4587.3681.8770.7590.7384.6979.1082.69
CoT (GPT-3) [47]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
CoT (UnifiedQA) [47]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
CoT (GPT-4) [48]1T+85.4872.4490.2782.6571.4992.8986.6679.0483.99
DDCoT [85]175B80.1576.7282.8278.8972.5385.0282.8675.2180.15
Chameleon (ChatGPT) [48]175B+81.6270.6484.0079.7770.8086.6281.8676.5379.93
LG-VQA (BLIP-2) [24]---------86.32
LaVIN-13B [77]---------77.54
BLIP-2 [77]---------74.17
CCOT7B--------76.84
GraphVis [18]7B--------73.18
Qwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
Qwen2.5-72B72B79.6467.1084.9077.5665.0087.9380.2574.8578.37
Qwen2.5-7B (Mmkg) [45]7B73.9866.3778.1871.6564.3079.6576.5168.0373.47
Qwen2.5-7B (Visual Genome) [37]7B76.7867.0478.0974.0566.1979.7278.0869.6875.08
Qwen2.5-7B (VaLiK Text-only)7B84.5474.2486.9182.7472.5390.0384.5180.2882.98
Qwen2.5-7B (VaLiK Image-only)7B79.1471.5479.2777.1669.7283.1480.6573.9678.88
Qwen2.5-7B (VaLiK Text-Image)7B84.1575.1487.6482.9973.1889.6984.4080.9583.16
Qwen2.5-72B (VaLiK Text-Image)72B85.6175.9390.2784.4074.1792.3385.7982.9884.77
+ +Table 3. Performance comparison (\%) on ScienceQA benchmark. #T-Params denotes trainable parameters. Categories: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG-Cap (image caption), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12). Method groups: (1) Human performance baseline, (2) Zero/Few-shot text-only LLMs, (3) Zero/Few-shot Multimodal VLMs, (4) LLMs enhanced with knowledge graphs for multimodal reasoning. + +strained compared to text-only counterparts. We separately applied VaLiK enhancement to Qwen2-VL-72B-Instruct and LLaVA-34B, obtaining distinct improvements: LLaVA-34B achieves accuracy gains of $2.41\%$ (image-only KG) and $3.59\%$ (text-image KG), while Qwen2-VL-72B-Instruct shows $1.77\%$ and $2.23\%$ improvements respectively under identical configurations. These experimental findings collectively demonstrate that VaLiK effectively extracts valuable signals from the training corpus and enables dynamic knowledge injection into VLMs during inference, thereby substantially alleviating hallucination phenomena. The differential improvements between Qwen2-VL-72B-Instruct and LLaVA-34B further validate the framework's adaptability across model architectures. + +Additionally, we analyze the results of LLMs without KG enhancement in the tables, which generally follow the scaling law [36]. However, DeepSeek-R1 shows anomalous behavior. Through testing, we find that its reasoning process may introduce complex information that interferes with its judgment. Furthermore, empirical results show that most baseline models achieve suboptimal performance without fine-tuning. In contrast, VaLiK's automated MMKG construction framework requires no task-specific adaptation yet delivers consistent improvements. + +Multimodal Question Answering Tasks. We evalu + +ated multimodal QA performance on the ScienceQA benchmark with Qwen2.5-7B and Qwen2.5-72B as base architectures, augmented by four knowledge sources: Mmkg, Visual Genome, text-only LightRAG and VaLiK. Compared to existing zero-shot/few-shot LLMs that not specifically optimized for scientific QA, our VaLiK-enhanced Qwen2.5-72B achieved SOTA performance on $62.5\%$ of subtasks, demonstrating particular strengths in multimodal reasoning scenarios requiring cross-modal alignment with an average accuracy gain of $6.4\%$ over baseline models. + +Our study identifies a fundamental imbalance between textual and visual knowledge representations in ScienceQA. Text-only KGs (14k entities, 18k relations) exhibit $8 \times$ denser structured knowledge than image-only counterparts (3k concepts, 1k relations), explaining visual modality underperformance. Despite this gap, vision-KG-augmented Qwen2.5-7B still attains $4.16\%$ accuracy gains over its non-enhanced version. Notably, our MMKG requires only 489MB storage for complete storage, while the scene graph component2 of Visual Genome alone occupies 739MB. This lightweight construction enables effective reasoning using only textual KG descriptions without raw images in resource-constrained scenarios. + +
TypeMethod#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Image-OnlyQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B80.06 (↑3.86)70.30 (↑2.47)80.55 (↑3.28)78.05 (↑3.56)68.43 (↑2.64)83.76 (↑4.74)81.17 (↑3.45)72.71 (↑3.36)78.14 (↑3.42)
+ SV7B79.14 (↓0.92)71.54 (↑1.24)79.27 (↓1.28)77.16 (↓0.89)69.72 (↑1.29)83.14 (↓0.62)80.65 (↓0.52)73.96 (↑1.25)78.88 (↑0.74)
Text-ImageQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B81.88 (↑5.68)73.00 (↑5.17)84.00 (↑6.73)80.55 (↑6.06)70.05 (↑4.26)87.11 (↑8.09)82.01 (↑4.29)77.98 (↑8.63)80.57 (↑5.85)
+ SV7B84.15 (↑2.27)75.14 (↑2.14)87.64 (↑3.64)82.99 (↑2.44)73.18 (↑3.13)89.69 (↑2.58)84.40 (↑2.39)80.95 (↑2.97)83.16 (↑2.59)
+ +Table 4. Ablation study on ScienceQA benchmark (CVs: CoE-based Vision-Language Models; SV: Similarly Verification). Performance metrics include: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG (image context), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12). + +
TypeMethodTask 1 (%)Task 2 (%)Task 2-Merged (%)
Image-OnlyQwen2.5-7B65.0444.5245.33
+ CVs68.11 (↑3.07)47.00 (↑2.48)46.95 (↑1.62)
+ SV69.52 (↑1.41)49.54 (↑2.54)49.07 (↑2.12)
Text-ImageQwen2.5-7B65.0444.5245.33
+ CVs68.43 (↑3.39)48.61 (↑4.09)48.97 (↑3.64)
+ SV68.90 (↑0.47)50.02 (↑1.41)50.69 (↑1.72)
+ +Table 5. Ablation study on CrisisMMD with Qwen2.5-7B. + +# 4.3. Ablation Study + +Our ablation studies on CrisisMMD and ScienceQA demonstrate the specific roles of VaLiK's components. As shown in Table 4 and Table 5, the CVs (CoE-based VLM) module improves accuracy across all settings, with average gains of $+3.05\%$ on CrisisMMD and $+4.63\%$ on ScienceQA tasks, validating visual descriptions enhance reasoning. However, the SV (Similarly Verification) module exhibits dual effects: it significantly improves CrisisMMD metrics by pruning redundant textual descriptions, yet slightly degrades ScienceQA's image-only natural science reasoning. We hypothesize this discrepancy arises from dataset characteristics: CrisisMMD's generated captions contain substantially more redundant content, whereas ScienceQA's simpler visual scenes yield shorter descriptions. Pruning these shorter descriptions risks over-removal of critical semantics. Furthermore, different types of KGs influence the effectiveness of the components: CVs achieve greater gains in CrisisMMD's text-image fusion as original text provides complementary context, while SV shows reduced effectiveness, likely due to occasional over-pruning of cross-modal linkages. Nevertheless, both modules collectively enhance performance across configurations, demonstrating their synergistic yet context-sensitive nature. + +# 4.4. Further Analysis + +Impact of VLM Quantity and Types. We evaluate the impact of varying quantities and types of VLMs on the CVs module. Our experiments reveal that Qwen2-VL generates the most visual descriptions, followed by LLaVA, while BLIP-2 produces the fewest. However, BLIP-2 demonstrates superior capability in extracting critical information + +![](images/0c11325abdab9142cb8a1d96545e124e536e4c45e57e562428116da00e71ee8c.jpg) +Figure 4. Impact analysis of VLM quantity on CrisisMMD. + +![](images/9559482c345cc743f96b22e87bcf418e0d4c3d23fddec5c9335bc5e04f78ee59.jpg) + +and identifying key entity relationships within images. We therefore adopt BLIP-2 as the primary model, with LLaVA or Qwen2-VL serving as secondary/tertiary components. Adding more VLMs yields diminishing returns, due to limited entities in current images, though we hypothesize their benefits would increase for complex visual scenes with richer semantic content. This phenomenon is empirically validated by our quantitative results in Figure 4. + +Computational Costs. Due to space limitations, we provide an overview of VaLiK's computational costs in Appendix F. Our method is significantly more cost-effective than manual annotation or LLM fine-tuning. + +# 5. Conclusion + +Multimodal reasoning in LLMs is constrained by incomplete knowledge and hallucination artifacts, limitations that persist because textual KGs cannot bridge visual-textual semantics due to their modality isolation. To bridge this gap, we propose VaLiK, a framework for constructing MMKGs through vision-language alignment, eliminating dependency on manual annotations while resolving visual-textual semantic inconsistencies. By integrating a cascade of pretrained VLMs and cross-modal verification, VaLiK converts images into structured knowledge while filtering noise. The resulting graphs enhance LLMs' reasoning with minimal storage overhead. Experiments on multimodal reasoning benchmarks show SOTA performance. VaLiK's modular design supports adaptability across domains, offering a scalable solution for autonomous knowledge synthesis. This work advances multimodal AI systems by enabling efficient integration of visual and textual data. + +# 6. Acknowledgments + +The research was supported by Shanghai Artificial Intelligence Laboratory, the National Key R&D Program of China (Grant No. 2022ZD0160201) and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ1100102). + +# References + +[1] Mahdi Abavisani, Liwei Wu, Shengli Hu, Joel Tetreault, and Alejandro Jaimes. Multimodal categorization of crisis events in social media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 5 +[3] Firoj Alam, Ferda Ofli, and Muhammad Imran. Crisismmd: Multimodal twitter datasets from natural disasters. Proceedings of the International AAAI Conference on Web and Social Media, 12(1), 2018. 2, 5 +[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736. Curran Associates, Inc., 2022. 3, 4 +[5] Razvan Azamfirei, Sapna R Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 1 +[6] Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2023. 2 +[7] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):423-443, 2019. 3 +[8] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17754-17762, 2024. 1 +[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems (NeurIPS), 33:22243-22255, 2020. 1 +[10] Xiang Chen, Ningyu Zhang, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, and Huajun Chen. + +Hybrid transformer with multi-level fusion for multimodal knowledge graph completion. In Proceedings of the International Conference on Research and Development in Information Retrieval (SIGIR), pages 904-915, 2022. 2 +[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 3 +[12] Xiang Chen, Jingtian Zhang, Xiaohan Wang, Ningyu Zhang, Tongtong Wu, Yuxiang Wang, Yongheng Wang, and Huajun Chen. Continual multimodal knowledge graph construction. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 3 +[13] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 1 +[14] Yong Chen, Xinkai Ge, Shengli Yang, Linmei Hu, Jie Li, and Jinwen Zhang. A survey on multimodal knowledge graphs: Construction, completion and applications. Mathematics, 11 (8), 2023. 3 +[15] Zhuo Chen, Yichi Zhang, Yin Fang, Yuxia Geng, Lingbing Guo, Xiang Chen, Qian Li, Wen Zhang, Jiaoyan Chen, Yushan Zhu, et al. Knowledge graphs meet multimodal learning: A comprehensive survey. arXiv preprint arXiv:2402.05391, 2024. 2 +[16] Shiyao Cui, Jiangxia Cao, Xin Cong, Jiawei Sheng, Quanggang Li, Tingwen Liu, and Jinqiao Shi. Enhancing multimodal entity and relation extraction with variational information bottleneck. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:1274-1285, 2024. 2 +[17] Xiaohui Cui, Xiaolong Qu, Dongmei Li, Yu Yang, Yuxun Li, and Xiaoping Zhang. Mkgcn: Multi-modal knowledge graph convolutional network for music recommender systems. *Electronics*, 12(12), 2023. 3 +[18] Yihe Deng, Chenchen Ye, Zijie Huang, Mingyu Derek Ma, Yiwen Kou, and Wei Wang. Graphvis: Boosting llms with visual knowledge graph integration. In Advances in Neural Information Processing Systems, pages 67511-67534. Curran Associates, Inc., 2024. 6, 7 +[19] Tausif Diwan, G. Anirudh, and Jitendra V. Tembhurne. Object detection using yolo: challenges, architectural successors, datasets and applications. Multimedia Tools Appl., 82 (6):9243-9275, 2022. 3 +[20] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[21] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024. 5 +[22] Quan Fang, Xiaowei Zhang, Jun Hu, Xian Wu, and Changsheng Xu. Contrastive multi-modal knowledge graph representation learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8983-8996, 2023. 3 + +[23] Duoduo Feng, Xiangteng He, and Yuxin Peng. Mkvse: Multimodal knowledge enhanced visual-semantic embedding for image-text retrieval. ACM Trans. Multimedia Comput. Commun. Appl., 19(5), 2023. 3 +[24] Deepanway Ghosal, Navonil Majumder, Roy Ka-Wei Lee, Rada Mihalcea, and Soujanya Poria. Language guided visual question answering: Elevate your multimodal language model using knowledge-enriched prompts. arXiv preprint arXiv:2310.20159, 2023. 6, 7 +[25] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 6 +[26] Yunfei Guo, Fei Yin, Xiao-hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21604-21613, 2023. 3 +[27] ZIRUI GUO, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. LightRAG: Simple and fast retrieval-augmented generation, 2024. 2, 5, 6 +[28] Lavdim Halilaj, Juergen Luettin, Sebastian Monka, Cory Henson, and Stefan Schmid. Knowledge graph-based integration of autonomous driving datasets. International Journal of Semantic Computing, 17(02):249-271, 2023. 3 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4 +[30] Yang Hu, Guihua Wen, Adriane Chapman, Pei Yang, Mingnan Luo, Yingxue Xu, Dan Dai, and Wendy Hall. Graph-based visual-semantic entanglement network for zero-shot image recognition. IEEE Transactions on Multimedia, 24: 2473-2487, 2022. 3 +[31] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. LLM-adapters: An adapter family for parameter-efficient fine-tuning of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5254-5276, Singapore, 2023. Association for Computational Linguistics. 2 +[32] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 3, 6 +[33] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2 +[34] Adam Tauman Kalai and Santosh S. Vempala. Calibrated language models must hallucinate. In Proceedings of the 56th Annual ACM Symposium on Theory of Computing, page 160-171, New York, NY, USA, 2024. Association for Computing Machinery. 1 +[35] Amar Viswanathan Kannan, Dmitriy Fradkin, Ioannis Akrotirianakis, Tugba Kulahcioglu, Arquimedes Canedo, + +Aditi Roy, Shih-Yuan Yu, Malawade Arnav, and Mohammad Abdullah Al Faruque. Multimodal knowledge graph for deep learning papers and code. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 3417-3420, New York, NY, USA, 2020. Association for Computing Machinery. 3 +[36] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7 +[37] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7 +[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10767-10782, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3 +[39] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3 +[40] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning (ICML), pages 19730–19742, 2023. 1, 3, 4, 6 +[41] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466. Curran Associates, Inc., 2023. 3 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3, 6, 7 +[43] Junming Liu, Yanting Gao, Siyuan Meng, Yifei Sun, Aoqi Wu, Yufei Jin, Yirong Chen, Ding Wang, and Guosun Zeng. Mosaic: Data-free knowledge distillation via mixture-of-experts for heterogeneous distributed environments. arXiv preprint arXiv:2505.19699, 2025. 1 +[44] Junming Liu, Guosun Zeng, Ding Wang, Yanting Gao, and Yufei Jin. Fedrecon: Missing modality reconstruction in distributed heterogeneous environments. arXiv preprint arXiv:2504.09941, 2025.3 +[45] Ye Liu, Hui Li, Alberto Garcia-Duran, Mathias Niepert, Daniel Onoro-Rubio, and David S Rosenblum. Mmkg: multi-modal knowledge graphs. In The Semantic Web: 16th International Conference, ESWC 2019, Portoroz, Slovenia, June 2-6, 2019, Proceedings 16, pages 459-474. Springer, 2019. 2, 6, 7 + +[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4 +[47] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, pages 2507–2521. Curran Associates, Inc., 2022. 2, 5, 6, 7 +[48] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In Advances in Neural Information Processing Systems, pages 43447-43478. Curran Associates, Inc., 2023. 6, 7 +[49] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. In Advances in Neural Information Processing Systems, pages 29615-29627. Curran Associates, Inc., 2023. 6 +[50] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[51] Runqing Miao, Qingxuan Jia, Fuchun Sun, Gang Chen, Haiming Huang, and Shengyi Miao. Semantic representation of robot manipulation with knowledge graph. Entropy, 25(4), 2023. 3 +[52] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14420-14431, 2024. 6 +[53] Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 1 +[54] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 2 +[55] Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Knowledge-aware multi-modal adaptive graph convolutional networks for fake news detection. ACM Trans. Multimedia Comput. Commun. Appl., 17(3), 2021. 3 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6 + +[57] Brian Reily, Christopher Reardon, and Hao Zhang. Representing multi-robot structure through multimodal graph embedding for the selection of robot teams. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5576–5582, 2020. 3 +[58] Joshua Robinson, Christopher Michael Ryting, and David Wingate. Leveraging large language models for multiple choice question answering. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4 +[59] Hrituraj Singh, Anshul Nasery, Denil Mehta, Aishwarya Agarwal, Jatin Lamba, and Balaji Vasan Srinivasan. MIMOQA: Multimodal input multimodal output question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5317-5332, Online, 2021. Association for Computational Linguistics. 3 +[60] Yaoxian Song, Penglei Sun, Haoyu Liu, Zhixu Li, Wei Song, Yanghua Xiao, and Xiaofang Zhou. Scene-driven multimodal knowledge graph construction for embodied ai. IEEE Transactions on Knowledge and Data Engineering, 36(11): 6962-6976, 2024. 2, 3 +[61] Rui Sun, Xuezhi Cao, Yan Zhao, Junchen Wan, Kun Zhou, Fuzheng Zhang, Zhongyuan Wang, and Kai Zheng. Multimodal knowledge graphs for recommender systems. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 1405-1414, New York, NY, USA, 2020. Association for Computing Machinery. 3 +[62] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2 +[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3 +[64] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, ADITHYA JAIRAM IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1 +[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5 +[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4 +[67] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton van den Hengel. Fvqa: Fact-based visual question an + +swering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2413-2427, 2018. 3 +[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6 +[69] Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In Proceedings of the 31st ACM International Conference on Multimedia, page 2391–2399, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[70] Yuequn Wang, Liyan Dong, Hao Zhang, Xintao Ma, Yongli Li, and Minghui Sun. An enhanced multi-modal recommendation based on alternate training with knowledge graph representation. IEEE Access, 8:213012-213026, 2020. 3 +[71] Tao Wu, Mengze Li, Jingyuan Chen, Wei Ji, Wang Lin, Jinyang Gao, Kun Kuang, Zhou Zhao, and Fei Wu. Semantic alignment for multimodal large language models. In Proceedings of the 32nd ACM International Conference on Multimedia, page 3489-3498, New York, NY, USA, 2024. Association for Computing Machinery. 1 +[72] Yike Wu, Nan Hu, Guilin Qi, Sheng Bi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering. arXiv preprint arXiv:2309.11206, 2023. 2 +[73] Ziyang Xiao, Dongxiang Zhang, Yangjun Wu, Lilin Xu, Yuan Jessica Wang, Xiongwei Han, Xiaojin Fu, Tao Zhong, Jia Zeng, Mingli Song, and Gang Chen. Chain-of-experts: When LLMs meet complex operations research problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[74] Dexuan Xu, Yanyuan Chen, Jieyi Wang, Yue Huang, Hanpin Wang, Zhi Jin, Hongxing Wang, Weihua Yue, Jing He, Hang Li, and Yu Huang. MLeVLM: Improve multi-level progressive capabilities based on multimodal large language model for medical visual question answering. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4977-4997, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1 +[75] Peng Xu, Xiatian Zhu, and David A. Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 3 +[76] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6 +[77] Xiaocui Yang, Wenfang Wu, Shi Feng, Ming Wang, Daling Wang, Yang Li, Qi Sun, Yifei Zhang, Xiaoming Fu, and Soujanya Poria. Mm-bigbench: Evaluating multimodal models on multimodal content comprehension tasks. arXiv preprint arXiv:2310.09036, 2023. 7 +[78] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers + +large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 3 +[79] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1 +[80] Jingtong Yue, Zhiwei Lin, Xin Lin, Xiaoyu Zhou, Xiangtai Li, Lu Qi, Yongtao Wang, and Ming-Hsuan Yang. RobuR-CDet: Enhancing robustness of radar-camera fusion in bird's eye view for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[81] Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Huajun Chen, and Wen Zhang. Mygo: Discrete modality information as fine-grained tokens for multi-modal knowledge graph completion. CoRR, abs/2404.09468, 2024. 2 +[82] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 3 +[83] Jiabao Zhao, Xin Lin, Jie Zhou, Jing Yang, Liang He, and Zhaohui Yang. Knowledge-based fine-grained classification for few-shot learning. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 3 +[84] Changmeng Zheng, Junhao Feng, Ze Fu, Yi Cai, Qing Li, and Tao Wang. Multimodal relation extraction with efficient graph alignment. In Proceedings of the 29th ACM International Conference on Multimedia, page 5298-5306, New York, NY, USA, 2021. Association for Computing Machinery. 2 +[85] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. In Advances in Neural Information Processing Systems, pages 5168-5191. Curran Associates, Inc., 2023. 3, 6, 7 +[86] Juncheng Zheng, Meiyu Liang, Yang Yu, Yawen Li, and Zhe Xue. Knowledge graph enhanced multimodal transformer for image-text retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 70-82, 2024. 3 +[87] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 4 +[88] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 3 +[89] Xiangru Zhu, Zhixu Li, Xiaodan Wang, Xueyao Jiang, Penglei Sun, Xuwu Wang, Yanghua Xiao, and Nicholas Jing Yuan. Multi-modal knowledge graph construction and application: A survey. IEEE Transactions on Knowledge and Data Engineering, 36(2):715-735, 2024. 2 +[90] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 3 \ No newline at end of file diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/images.zip b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..571ac7b02ca504921f9b8b7453df764babb28788 --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc93d16bdb23385c6663d17bd561c541f217868d191fcefcdcc66c9fb48041f0 +size 655941 diff --git a/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/layout.json b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..588eb6a5148b27e7a84ce8f51eec4e9f7fa9d0fe --- /dev/null +++ b/2025/Aligning Vision to Language_ Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning/layout.json @@ -0,0 +1,10076 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "text", + "content": "Aligning Vision to Language: Annotation-Free Multimodal Knowledge Graph Construction for Enhanced LLMs Reasoning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "spans": [ + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": "Junming Liu" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": ", Siyuan Meng" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": ", Yanting Gao" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": ", Song Mao" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": ", Pinlong Cai" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 125, + 160, + 485, + 175 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "content": "Guohang Yan" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "content": ", Yirong Chen" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "content": ", Zilin Bian" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "content": ", Ding Wang" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "text", + "content": ", Botian Shi" + }, + { + "bbox": [ + 136, + 175, + 474, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "spans": [ + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 190, + 462, + 204 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "text", + "content": "East China Normal University " + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 115, + 203, + 495, + 217 + ], + "type": "text", + "content": "New York University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 162, + 220, + 441, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 220, + 441, + 232 + ], + "spans": [ + { + "bbox": [ + 162, + 220, + 441, + 232 + ], + "type": "text", + "content": "liu_junming6917@tongji.edu.cn wangding@pjlab.org.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 283, + 297, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 297, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 297, + 606 + ], + "type": "text", + "content": "Multimodal reasoning in Large Language Models (LLMs) struggles with incomplete knowledge and hallucination artifacts, challenges that textual Knowledge Graphs (KGs) only partially mitigate due to their modality isolation. While Multimodal Knowledge Graphs (MMKGs) promise enhanced cross-modal understanding, their practical construction is impeded by semantic narrowness of manual text annotations and inherent noise in visual-semantic entity linkages. In this paper, we propose Vision-align-to-Language integrated Knowledge Graph (VaLiK), a novel approach for constructing MMKGs that enhances LLMs reasoning through cross-modal information supplementation. Specifically, we cascade pre-trained Vision-Language Models (VLMs) to align image features with text, transforming them into descriptions that encapsulate image-specific information. Furthermore, we developed a cross-modal similarity verification mechanism to quantify semantic consistency, effectively filtering out noise introduced during feature alignment. Even without manually annotated image captions, the refined descriptions alone suffice to construct the MMKG. Compared to conventional MMKGs construction paradigms, our approach achieves substantial storage efficiency gains while maintaining direct entity-to-image linkage capability. Experimental results on multimodal reasoning tasks demonstrate that LLMs augmented with VaLiK outperform previous state-of-the-art models. Our code is published at https://github.com/Wings-Of-Disaster/VaLiK." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 696 + ], + "type": "text", + "content": "Recent advancements in Large Language Models (LLMs) [2, 9, 25, 65] have demonstrated their superiority and versatility across various Natural Language Reasoning (NLR) tasks [8, 43, 53, 58]. To enhance LLMs into the" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 260, + 552, + 543 + ], + "blocks": [ + { + "bbox": [ + 318, + 260, + 552, + 543 + ], + "lines": [ + { + "bbox": [ + 318, + 260, + 552, + 543 + ], + "spans": [ + { + "bbox": [ + 318, + 260, + 552, + 543 + ], + "type": "image", + "image_path": "dfacec5607ec5e36626acfb0e0ac18a11bebb24a4130fb85486e59c4e58c6e72.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 553, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 598 + ], + "type": "text", + "content": "Figure 1. (a) Training entity extraction models relies on extensive fine-grained annotations, increasing labeling costs. More examples are provided in Appendix B. (b) Capturing implicit semantic associations demands abstract comprehension or logical inference." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "realm of multimodal reasoning, researchers [64, 71, 74, 79] have endeavored to equip these models with multimodal capabilities, as evidenced by advancements in Multimodal Large Language Models (MLLMs) such as BLIP-2 [40], GPT-4o [32], Janus-Pro [13], among others. Despite their notable progress, these models often experience hallucinations [5, 34], primarily arising from knowledge deficiencies due to incomplete or obsolete information." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "981" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "Fine-tuning LLMs demands prohibitive computational costs [31]. While text-based Knowledge Graphs (KGs) have partially addressed this limitation by efficient real-time updates [6, 62, 72], they are still restricted by modal isolation, which hinders cross-modal reasoning, as detailed in Appendix A. To bridge this semantic fragmentation, Multimodal Knowledge Graphs (MMKGs) have been developed as unified representational frameworks [10, 33, 38, 45]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 294, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 294, + 278 + ], + "type": "text", + "content": "However, constructing robust MMKGs faces two primary obstacles [15, 89]. First, the lack of large-scale fine-grained entity-image corpora makes it infeasible to train high-quality entity extractors, significantly constraining scalability, as illustrated in Figure 1a. Second, conventional visual relation detectors primarily identify superficial spatial interactions instead of semantic relations consistent with KGs, while frequently hallucinating implausible connections that corrupt graph integrity, as shown in Figure 1b." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 628 + ], + "type": "text", + "content": "In this paper, we propose VaLiK, short for Vision-align-to-Language integrated Knowledge Graph, a novel framework designed to empower LLMs with advanced multimodal reasoning. Unlike traditional methods that rely on text annotations for training extraction models and the knowledge construction process [54], VaLiK adopts a annotation-free approach to MMKGs construction. Specifically, we first employ several pretrained Vision-Language models (VLMs), designed based on Chain-of-Experts (CoE) principles [73], to convert visual inputs into image-specific textual descriptions through cross-modal feature alignment. This procedure eliminates the need for manually annotated image captions in both the knowledge extraction and construction phases while preserving visual details typically missing in generic text descriptions. Moreover, in contrast to existing relation detection methods that require predefined label taxonomies [16, 60, 81, 84], VaLiK excels at extracting profound semantic relationships that are both KG-compatible and capture novel associations beyond training supervision. While VLMs enable cross-modal reasoning and interpretation, they introduce spurious relational noise through hallucinated inter-modal attributions, as depicted in Figure 2. We address this limitation through cross-modal similarity recalibration, strategically filtering inconsistent information while preserving valid semantic correspondences. Finally, the purified descriptions are systematically organized into MMKGs via LLM-driven symbolic structuring [27], bridging visual and textual domains with factual consistency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 713 + ], + "type": "text", + "content": "To thoroughly evaluate the VaLiK method, we conduct a comprehensive assessment across two critical multimodal benchmarks: multimodal classification (tested on the CrisisMMD dataset [3]) and multimodal question answering (evaluated via the ScienceQA benchmark [47]). The experiments span diverse LLM architectures and MMKG construction techniques to ensure the framework's robustness." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 552, + 208 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 552, + 208 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 552, + 208 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 552, + 208 + ], + "type": "image", + "image_path": "70a5e7b33cf2f7b408d7e9bbab0216de573f5ba53a84b04153ef68002435eb2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "Figure 2. Feature-aligned descriptions from VLMs introduce redundant and inaccurate relationship patterns." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 323 + ], + "type": "text", + "content": "The experimental results demonstrate that the MMKGs constructed by VaLiK achieve superior multimodal reasoning performance in LLMs while requiring substantially less storage than conventional approaches. More importantly, the proposed approach retains direct entity-to-image linkage capabilities even with the compressed graph structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 372 + ], + "type": "text", + "content": "In summary, VaLiK is the first framework that enables end-to-end, annotation-free, zero-shot, and storage-efficient multimodal knowledge construction with high adaptability and scalability. Our key contributions include:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 373, + 553, + 576 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 373, + 553, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 553, + 444 + ], + "type": "text", + "content": "- To the best of our knowledge, VaLik is the first end-to-end framework to build Annotation-Free MMKGs to improve LLMs' multimodal reasoning capabilities, effectively eliminating the need for manually annotated textual material and enabling a completely autonomous multimodal knowledge generation process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 445, + 553, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 553, + 516 + ], + "type": "text", + "content": "- We offer an innovative zero-shot method for constructing MMKG that captures deep semantic connections beyond traditional predetermined labels with an effective verification system that guarantees the accuracy of these relationships. The knowledge distillation paradigm greatly decreases storage while maintaining semantic integrity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "type": "text", + "content": "- We develop a highly modular and extensible architecture that allows VaLiK to effortlessly incorporate new models and workflows for specialized domain tasks, facilitating rapid adaptation to diverse application scenarios without incurring expensive system changes." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 590, + 400, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 590, + 400, + 602 + ], + "spans": [ + { + "bbox": [ + 314, + 590, + 400, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "type": "text", + "content": "2.1. Multimodal Knowledge Graphs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "The principal advantage of MMKGs resides in their multimodal integration beyond conventional KGs. By linking entities with corresponding visual or textual data, MMKGs introduce valuable visual and textual information to the knowledge base, substantially advancing multimodal reasoning capabilities. This combination addresses core challenges in tasks that inherently demand multimodal synergy" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "982" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 301 + ], + "type": "text", + "content": "like autonomous driving [26, 28], image-text retrieval [23, 86] and robotic manipulation [51, 57]. However, constructing trustworthy MMKGs with minimal manual effort remains a critical challenge. Recent studies have proposed innovative strategies to enhance MMKG reliability and utility. For instance, Chen et al. [12] proposed MSPT, a framework addressing continual MMKG construction through gradient modulation for balanced multimodal learning and attention distillation to mitigate catastrophic forgetting. Song et al. [60] developed Scene-MMKG, integrating knowledge engineering with large language models to improve robotic manipulation by resolving data sparsity and knowledge uncertainty. Wang et al. [69] introduced TIVA-KG, the first quad-modal knowledge graph spanning text, image, video, and audio with triplet grounding, empirically validating its effectiveness in downstream tasks. While these advances enhance multimodal reasoning capabilities, their efficacy remains rooted in resource-intensive paradigms, requiring extensively annotated datasets for knowledge acquisition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 308, + 293, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 293, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 293, + 322 + ], + "type": "text", + "content": "2.2. Knowledge-Augmented Multimodal Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 579 + ], + "type": "text", + "content": "Multimodal learning has seen significant progress in aligning and integrating information across different data modalities [7, 44, 75]. The incorporation of structured knowledge through MMKGs further enhances these approaches, improving the reasoning capabilities and generalization across a variety of domains, such as visual question answering [50, 59, 67], recommendation systems [17, 61, 70], and classification [30, 55, 83]. Methods like GraphAdapter's dual-KG adaptation [41] and contrastive multi-relational encoding with KGs [22] inject external knowledge into models, refining their performance and improving their capability to handle complex tasks. Additionally, Lee et al. [38] proposed MR-MKG, a novel framework that constructs task-specific MMKGs to enhance multimodal reasoning in LLMs. These knowledge-augmented paradigms demonstrate superior cross-modal semantic grounding compared to unimodal approaches [14, 35]. However, their reliance on preconstructed MMKGs often leads to domain discrepancies, where generic knowledge schemas misalign with task-specific reasoning patterns, ultimately limiting contextual precision in target applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 587, + 251, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 251, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 251, + 601 + ], + "type": "text", + "content": "2.3. Multimodal Large Language Models" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "The limitations of text-only LLMs in meeting increasingly complex demands have spurred extensive research [78, 82, 85] into developing LLMs capable of effectively processing and reasoning over multimodal inputs. Current research predominantly employs adapter or projection layers to connect the embedding spaces of various modality-specific encoders with the textual embedding space of LLMs [38]. For instance, foundational models like CLIP [56] and BLIP [39] pioneered cross-modal alignment by jointly training vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 253 + ], + "type": "text", + "content": "and text encoders to map images and text into a shared embedding space. Building on this, LLaVA [42] and Flamingo [4] advanced the field by integrating visual encoders with LLMs, enabling more nuanced multimodal understanding and generation. More recently, Gemini [63], Qwen2-VL [68] and GPT-4o [32] have further pushed the boundaries by scaling up multimodal pretraining and introducing sophisticated mechanisms for cross-modal interaction. However, multimodal LLMs remain prone to hallucinations. While they enhance cross-modal alignment, they neither acquire new knowledge nor avoid introducing noise through integration. To address these limitations, VaLiK \"uses the master's tools to refine the master's craft,\" first constructing MMKGs via MLLMs and then leveraging them to enhance MLLMs' reasoning capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 263, + 370, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 370, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 370, + 275 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 283, + 556, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 556, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 556, + 429 + ], + "type": "text", + "content": "In this section, we present the technical details of VaLiK. VaLiK introduces a novel expansion-reduction paradigm for visual knowledge extraction. The architecture initially organizes several VLMs with distinct knowledge domains, designed based on CoE principles [73], to produce comprehensive textual descriptions encompassing hierarchical visual details. A cross-modal similarity verification mechanism then iteratively filters out noisy tokens through cross-modal alignment while preserving semantically salient elements. This optimization-style approach eliminates external textual dependencies while enabling effective MMKG construction. VaLiK's framework is shown in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 435, + 529, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 529, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 529, + 449 + ], + "type": "text", + "content": "3.1. CoE-based Visual to Language Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "type": "text", + "content": "Recent entity detection techniques [19, 80, 90] have been widely adopted for entity and relation extraction in MMKG construction. However, these methods are inherently limited by predefined categorical boundaries, lacking the capacity to recognize visual concepts outside their training vocabulary. In contrast, VLMs pretrained on web-scale corpora [11, 40, 88] exhibit broader recognition capabilities through exposure to diverse visual concepts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 548, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 609 + ], + "type": "text", + "content": "We therefore leverage pretrained VLMs to extract comprehensive visual information. This process removes the necessity for detailed fine-grained data typically required to train specialized recognition models. The generalized vision to language conversion pipeline can be formalized as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 616, + 555, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 616, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 383, + 616, + 555, + 637 + ], + "type": "interline_equation", + "content": "S = \\mathcal {D} _ {\\text {t e x t}} \\left(\\mathcal {A} \\left(\\mathcal {E} _ {\\text {v i s}} (I)\\right)\\right), \\tag {1}", + "image_path": "d19d23eb15739c40b38c5104cccd6bdc2fd7f614911b465cec21e76bb69d0c41.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " denotes for the input image, " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " denotes the visual encoder extracting visual features, " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " carries out cross-modal feature alignment and interaction, and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " generates textual tokens through autoregressive decoding. The resulting visual description " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "S = \\{w_{1},\\dots,w_{n}\\}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " emerges from this multi-stage processing." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "983" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 238 + ], + "type": "image", + "image_path": "f146c2291d18be7bbdd2100fa3e32b5e7a21e72aea45eeb431f07813670ed63c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 555, + 281 + ], + "type": "text", + "content": "Figure 3. The pipeline of VaLiK: First, large-scale visual descriptions are generated using CoE-based VLMs. Then, a similarity verification mechanism is used to prune irrelevant information. Finally, MMKGs are constructed using LLMs based on LightRAG. The constructed MMKGs can assist LLMs in multimodal reasoning, alleviating the hallucination issues caused by incomplete knowledge." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": "However, quantitative analysis uncovers considerable discrepancies between machine-generated and human-annotated descriptions [87]. As an illustration, while utilizing BLIP-2 [40] to generate sample captions, we noted that the model outputs are markedly concise and devoid of visual specifics, as detailed in Appendix C. To bridge this gap, we implement CoE enhanced generation through cascade VLMs processing. At iteration step " + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": ", each expert " + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": " receives both the original visual signals " + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": " and the contextual output from the preceding expert " + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "inline_equation", + "content": "E_{i - 1}" + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 417, + 295, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 417, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 126, + 417, + 295, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} ^ {(t)} = E _ {i} \\left(I, \\mathcal {S} _ {i - 1} ^ {(t - 1)}\\right), \\tag {2}", + "image_path": "cc8cd53566f173b3deea26ddcb747f6b18c32dc8419234535699d34f08dea029.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "inline_equation", + "content": "S_{i - 1}^{(t - 1)}" + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "content": " denotes the description from expert " + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "inline_equation", + "content": "E_{i - 1}" + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "content": " at step " + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "inline_equation", + "content": "S_0^{(t)}\\coloneqq \\emptyset" + }, + { + "bbox": [ + 55, + 448, + 295, + 475 + ], + "type": "text", + "content": " for initialization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 476, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 502 + ], + "type": "text", + "content": "Specifically, each expert " + }, + { + "bbox": [ + 55, + 476, + 295, + 502 + ], + "type": "inline_equation", + "content": "E_{i}" + }, + { + "bbox": [ + 55, + 476, + 295, + 502 + ], + "type": "text", + "content": " implements a unified visual-language processing task:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 503, + 185, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 185, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 185, + 514 + ], + "type": "text", + "content": "1. Visual Feature Extraction:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 524, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 524, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 125, + 524, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {i} = \\operatorname {E n c} _ {\\text {v i s}} ^ {i} (I) \\in \\mathbb {R} ^ {d _ {v} \\times N _ {p}}, \\tag {3}", + "image_path": "66f47e4a8f1a81816e06304d1e933d13f899c402e5cec3ab370a3c5f59e2bce4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "inline_equation", + "content": "\\mathsf{Enc}_{\\mathrm{vis}}^i" + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "text", + "content": " denotes established visual encoder [20, 29, 46] producing " + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "text", + "content": " patch embeddings with dimension " + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "inline_equation", + "content": "d_v" + }, + { + "bbox": [ + 67, + 548, + 295, + 573 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 574, + 249, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 249, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 249, + 585 + ], + "type": "text", + "content": "2. Cross-Modal Interaction and Generation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "text", + "content": "VLMs integrate pretrained learnable query embeddings " + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i\\in \\mathbb{R}^{d_q\\times L_q}" + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "text", + "content": " to interact with visual features " + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i\\in" + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_v\\times N_p}" + }, + { + "bbox": [ + 67, + 586, + 295, + 623 + ], + "type": "text", + "content": " via cross-attention [66]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 631, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 631, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 83, + 631, + 295, + 680 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {H} _ {i} = \\operatorname {C r o s s A t t n} \\left(\\mathbf {Q} _ {i}, \\mathbf {V} _ {i}\\right) \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {Q} _ {i} \\mathbf {W} _ {q} ^ {i} \\left(\\mathbf {V} _ {i} \\mathbf {W} _ {k} ^ {i}\\right) ^ {\\top}}{\\sqrt {d _ {k}}}\\right) \\mathbf {V} _ {i} \\mathbf {W} _ {v} ^ {i}, \\tag {4} \\\\ \\end{array}", + "image_path": "93525ba17ac6f3639e6bfd29926c37b5e4442e1b9129d66450f3e54ae51c0a2f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q^i\\in \\mathbb{R}^{d_q\\times d_k}" + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k^i" + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v^i\\in \\mathbb{R}^{d_v\\times d_k}" + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{q}" + }, + { + "bbox": [ + 67, + 688, + 296, + 714 + ], + "type": "text", + "content": " denotes the predefined query length. Cross-attention serves" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 288, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 288, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 325, + 288, + 553, + 324 + ], + "type": "text", + "content": "as a prevalent approach, while other interaction strategies coexist [4]. The adopted VLMs in our implementation primarily rely on this approach for modality fusion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 326, + 403, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 326, + 403, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 326, + 403, + 336 + ], + "type": "text", + "content": "3. Text Generation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "text", + "content": "The text encoder " + }, + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "inline_equation", + "content": "\\mathsf{Enc}_{\\mathrm{text}}^{i}" + }, + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "text", + "content": " first processes the preceding expert's output " + }, + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "inline_equation", + "content": "S_{i - 1}^{(t - 1)}" + }, + { + "bbox": [ + 325, + 336, + 553, + 365 + ], + "type": "text", + "content": " into latent features:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 374, + 553, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 374, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 375, + 374, + 553, + 390 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {i} = \\operatorname {E n c} _ {\\text {t e x t}} ^ {i} \\left(S _ {i - 1} ^ {(t - 1)}\\right) \\in \\mathbb {R} ^ {d _ {t} \\times L}. \\tag {5}", + "image_path": "676733440db0b9882a960cd592abcb680fe61e0aba9dfa6d1b1e48425ea9fc62.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "content": "Subsequently, the text decoder " + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}_{\\mathrm{text}}^{i}" + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "content": " synthesizes the final output " + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "inline_equation", + "content": "S_{i}^{(t)}" + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "content": " by jointly conditioning on " + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i" + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_i" + }, + { + "bbox": [ + 326, + 401, + 553, + 428 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 339, + 447, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 447, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 339, + 447, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {S} _ {i} ^ {(t)} = \\operatorname {D e c} _ {\\text {t e x t}} ^ {i} \\left(\\mathbf {P} _ {i}, \\mathbf {H} _ {i}\\right) = \\left\\{w _ {1} ^ {(t, i)}, \\dots , w _ {m} ^ {(t, i)} \\right\\}. \\tag {6}", + "image_path": "0e27894f15d68c665c5e86d526ed483f49fafbbb9c71ce9cf51e418ed0c00788.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "text", + "content": "Ultimately, the final textual description " + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "inline_equation", + "content": "S_N^{(C)}" + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "text", + "content": " is obtained after " + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "text", + "content": " iteration steps through " + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 473, + 553, + 499 + ], + "type": "text", + "content": " cascaded experts." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 507, + 504, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 504, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 504, + 520 + ], + "type": "text", + "content": "3.2. Cross-Modal Similarity Verification" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "content": "To address noise in VLM-generated captions, we design a sliding window mechanism with semantic consistency verification. This method ensures that only relevant and semantically consistent segments are retained in the final description. Let " + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "content": "-th window containing " + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "content": " consecutive tokens " + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "inline_equation", + "content": "\\{w_{km + 1},\\dots ,w_{(k + 1)m}\\}" + }, + { + "bbox": [ + 313, + 525, + 555, + 609 + ], + "type": "text", + "content": ". For each window, we compute its cross-modal similarity score:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 364, + 617, + 553, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 617, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 364, + 617, + 553, + 645 + ], + "type": "interline_equation", + "content": "\\alpha_ {k} = \\frac {\\operatorname {E n c} _ {\\text {v i s}} (I) \\cdot \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right)}{\\| \\operatorname {E n c} _ {\\text {v i s}} (I) \\| \\| \\operatorname {E n c} _ {\\text {t e x t}} \\left(W _ {k}\\right) \\|}, \\tag {7}", + "image_path": "94158e93df6d2350058c36b406aa84f3338033517b5caa73e3a0e922316b20de.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathsf{Enc}_{vis/text}(\\cdot)" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " adopts a lightweight CLIP [58] encoder-decoder with frozen parameters for efficient processing. The similarity score " + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\alpha_{k}" + }, + { + "bbox": [ + 313, + 653, + 555, + 712 + ], + "type": "text", + "content": " lies within the range [0, 1], with higher values indicating a stronger alignment between the visual and textual information." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "984" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "After calculating the cross-modal similarity for each window, we employ an empirical threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " to filter out low-similarity windows. This threshold helps to identify and discard noisy or irrelevant sections of the generated caption that do not align well with the visual content, thereby reducing the impact of inaccurate or misleading descriptions. Formally, for each window " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\alpha_{k} < \\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", the window is discarded as noise. This process effectively prunes windows with low similarity scores, ensuring that only semantically meaningful segments remain. The final denoised description " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " is obtained by concatenating all windows " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " for which " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\alpha_{k} \\geq \\tau" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 143, + 221, + 296, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 221, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 143, + 221, + 296, + 248 + ], + "type": "interline_equation", + "content": "\\hat {S} = \\bigcup_ {\\alpha_ {k} \\geq \\tau} W _ {k}. \\tag {8}", + "image_path": "356f1eb00fa696f51141c8117686049d8d8bde11ba3ee95f9062bf6b32c13820.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 278 + ], + "type": "text", + "content": "Our window size " + }, + { + "bbox": [ + 55, + 254, + 296, + 278 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 254, + 296, + 278 + ], + "type": "text", + "content": " is flexibly determined and generally adapts dynamically to natural sentence segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 296, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 297 + ], + "type": "text", + "content": "3.3. MMKG Construction for Enhanced Reasoning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 409 + ], + "type": "text", + "content": "LLMs have become increasingly popular for identifying entities, relationships, and attributes within a corpus, which are then organized into a KG. The strength of LLM-based KG generation lies in its capacity to leverage the vast amount of knowledge encoded within these models, allowing them to detect complex and nuanced patterns across diverse data sources. This approach eliminates the need for manual annotation, enabling a highly scalable and domain-adaptive process suitable for a wide range of applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "content": "We begin by refining the generated textual description " + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "content": " (VLM-based information), which is then optionally concatenated with any available external textual knowledge " + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "content": " to form the input for KG generation. This combined input is used to generate MMKGs with the help of a LLM [21, 27], leveraging its capacity for multi-hop reasoning and dynamic knowledge integration." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 501, + 295, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 501, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 130, + 501, + 295, + 521 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\operatorname {L L M} (\\hat {S} \\oplus T), \\tag {9}", + "image_path": "38601c1d1b595202adbcef02e10f63813194bf24755c7651d3f938be61bf4af9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "text", + "content": " denotes optional concatenation based on the availability of " + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "text", + "content": ". The resulting graph " + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 523, + 295, + 559 + ], + "type": "text", + "content": " captures both visual and textual relationships inferred by the LLM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 559, + 195, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 559, + 195, + 571 + ], + "spans": [ + { + "bbox": [ + 67, + 559, + 195, + 571 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 67, + 559, + 195, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 67, + 559, + 195, + 571 + ], + "type": "text", + "content": " as a set of triplets:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 578, + 295, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 578, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 107, + 578, + 295, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\{(h, r, t) \\mid h, t \\in \\mathcal {E}, r \\in \\mathcal {R} \\}, \\tag {10}", + "image_path": "bf8e0b4ca5cd8a1c4e61d1431e1dce516e6362a5ad98e68058456a9ae74242d6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": " denote the sets of entities and relations. Entities include objects or concepts from the image or external text, while relations describe connections such as \"is a type of,\" \"part of,\" or \"has property.\" Each triplet " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "(h,r,t)" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": " links a head entity " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": " and a tail entity " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": " via relation " + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 597, + 296, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "text", + "content": "Multimodal Reasoning Enhancement. To support multimodal reasoning, we retrieve relevant triplets from " + }, + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 658, + 296, + 694 + ], + "type": "text", + "content": " through structural patterns during LLMs inference:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 132, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {q} = \\operatorname {R e t r i e v e} (q, \\mathcal {G}), \\tag {11}", + "image_path": "6ca4a1adbe4b2a148fc54ecb81ae4dd5abb2ae824d3f93fd36b3b5b91b294cae.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\text{Retrieve}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " denotes a retrieval strategy that identifies subgraphs relevant to the query for reasoning. Detailed retrieval strategies are described in Appendix D." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 109, + 553, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 553, + 121 + ], + "type": "text", + "content": "The augmented prompt integrates multimodal evidence:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 131, + 553, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 131, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 361, + 131, + 553, + 159 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {a u g}} = q \\left\\|\\left(\\bigcup_ {(h, r, t) \\in \\mathcal {G} _ {q}} [ h ] \\rightarrow r \\rightarrow [ t ]\\right). \\right. \\tag {12}", + "image_path": "a919a72d07971e9512e6adecd6c864445f2db4c4fa908665181cc2f524238b2f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 163, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 248 + ], + "type": "text", + "content": "Note that we incorporate the storage locations of images in the database during MMKGs construction, enabling the MMKGs to link to visual data. VaLiK enables text-only LLMs to perform multimodal reasoning through " + }, + { + "bbox": [ + 313, + 163, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 163, + 555, + 248 + ], + "type": "text", + "content": "'s visual associations, while VLMs refresh knowledge representations by jointly injecting both visual and textual information, significantly mitigating hallucination risks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 259, + 390, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 390, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 390, + 272 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 279, + 368, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 368, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 368, + 293 + ], + "type": "text", + "content": "4.1. Setups" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 553, + 321 + ], + "type": "text", + "content": "Evaluation Datasets. We evaluate VaLiK on two multimodal reasoning benchmarks with distinct characteristics:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 322, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 553, + 417 + ], + "type": "text", + "content": "- CrisisMMD [3]. This real-world disaster response dataset includes around 35,000 noisy social media postings with paired images and text, each annotated for seven catastrophe categories and four severity levels. Its realistic user-generated content with natural noise and implicit modality correlations provides a rigorous testbed for zero-shot adaptation, with good performance indicating practical relevance in real-world crisis scenarios." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 418, + 554, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 554, + 513 + ], + "type": "text", + "content": "- ScienceQA [47]. This dataset contains 21,208 multimodal science questions combining textual and visual contexts, with " + }, + { + "bbox": [ + 314, + 418, + 554, + 513 + ], + "type": "inline_equation", + "content": "48.7\\%" + }, + { + "bbox": [ + 314, + 418, + 554, + 513 + ], + "type": "text", + "content": " of instances containing images. Questions span physics, chemistry, and biology domains, requiring cross-modal reasoning between textual concepts and visual diagrams. Additionally, ScienceQA offers image captions to aid text-only LLMs in reasoning, allowing a comparison of unimodal approaches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "content": "Task Formulation. For CrisisMMD, we define three multimodal classification tasks1: (1) binary information relevance filtering, (2) fine-grained humanitarian category recognition, and (3) a consolidated taxonomy with merged categories to reduce label complexity. We omit the unimodal damage assessment to focus on multimodal aspects. For ScienceQA, we follow the original evaluation using multiple metrics: question types, contextual modalities, and educational stages. Performance is assessed through accuracy percentage across these categories." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 633, + 553, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 553, + 669 + ], + "type": "text", + "content": "Baselines. We conduct a comprehensive evaluation of text-only LLMs, multimodal VLMs, and KGs that enhance LLMs in multimodal reasoning." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 670, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 670, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 314, + 670, + 553, + 694 + ], + "type": "text", + "content": "- For CrisisMMD, we compare text-only LLMs using few-shot prompting (LLaMA-2 [65], GPT-4 [2]," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 553, + 713 + ], + "type": "text", + "content": "This setting references the repository GitHub and Abavisani et al. [1]" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "985" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 144 + ], + "type": "table", + "html": "
TaskText-only LLMsKG-Enhanced LLMs
LLaMA-2GPT-4DeepSeek-R1Qwen2.5LightRAGVaLiK
7B13B70B-7B8B32B70B7B32B72BText-onlyImage-onlyText-Image
Task 162.3263.8063.1566.8367.2363.3163.6165.5365.0467.2867.9567.4969.5268.90
Task 218.3221.8228.8747.2526.5325.4924.7721.0544.5246.9450.5145.1149.5450.02
Task 2 Merged21.4533.1536.8949.4425.8523.5621.5525.5745.3347.0750.2945.9449.0750.69
", + "image_path": "ba85b4a98ffd5aca5fb455bf5872311135cd1ed5f88bf09e4b9f0d4269f53a72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 195, + 563, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 555, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 555, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 555, + 187 + ], + "type": "text", + "content": "Table 1. The performance evaluation of text-only LLMs using few-shot prompting without any fine-tuning on the training set. As these models handle text only, test data is formatted as unimodal text for compatibility. In our implementations, both LightRAG and VaLiK adopt Qwen2.5-7B as the base reasoning model. Bold indicates the highest value, and underline indicates the second highest." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 195, + 563, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 195, + 563, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 195, + 563, + 277 + ], + "type": "table", + "html": "
TaskMultimodal VLMsKG-Enhanced LLMs
CLIPLLaVABLIP-2GPT-4oQwen2-VLVaLiK
ViT-L/147B13B34BFlan-T5-XLOPT-2B-I7B-I72B-I*#+~
Task 143.3654.0060.5856.4461.2938.6268.2047.5662.4565.8060.7868.4461.1168.89
Task 217.8828.0120.1425.1540.8614.2647.587.6032.6847.2125.8048.8827.2349.78
Task 2-M20.7930.6123.4425.0740.7214.2749.557.4234.2048.2827.3149.2729.0949.31
", + "image_path": "a322dfcaa43dd7ca6cae69d5260b7f829eb8a055cd6a1848cf1eada31544feb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 285, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 555, + 321 + ], + "type": "text", + "content": "Table 2. The performance of multimodal VLMs and KG-enhanced LLMs. The -I suffix denotes instruction-tuned variants. Symbol markers denote KG types and models: the asterisk (*) represents image-only KG with LLaVA-34B, hash (#) indicates image-only KG using Qwen2-VL-72B-I, plus (+) denotes text-image KG with LLaVA-34B, and tilde (\\~) shows text-image KG using Qwen2-VL-72B-I." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 63, + 339, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 339, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 63, + 339, + 294, + 374 + ], + "type": "text", + "content": "DeepSeek-R1 [25], Qwen-2.5 [76]) and multimodal VLMs (CLIP [56], LLaVA [42], GPT-4o [32], Qwen2-VL [68], BLIP-2 [40])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 295, + 508 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 376, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 459 + ], + "type": "text", + "content": "- For ScienceQA, we compare models for general domains in zero/few-shot settings, including text-only LLMs (GPT Model [47], CoT [47], DDCoT [85]), multimodal VLMs (LG-VQA [24], LaVIN [49], BLIP-2, CCOT [52], GraphVis [18]) and Tool-LLM Chameleon [48]. These models are not specifically fine-tuned for scientific tasks, ensuring a fair evaluation of generalization capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 508 + ], + "type": "text", + "content": "- We further compare the multimodal reasoning performance of LLMs assisted by KGs, evaluating text-based KGs built with LightRAG [27], and pre-constructed MMKGs such as Visual Genome [37] and Mmkg [45]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": "Implementation. For MMKG construction, we design a chain of VLMs including BLIP-2, LLaVA, and Qwen2-VL, with the CLIP-ViT-L/14 for pruning. Stronger or additional VLMs could be employed to enhance performance if more computational resources are available. We use the entire training set as the knowledge base and construct MMKGs from the extracted descriptions based on the LightRAG framework. In comparative experiments, the LightRAG method we evaluate utilizes only textual data, while VaLiK employs two configurations: (1) fully image-generated text descriptions (Image-only), and (2) original text combined with image-generated text (Text-Image). Dynamic window partitioning based on sentence length ensures syntactically coherent pruning results. Similarity thresholds are set to " + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.25" + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": " for CrisisMMD and " + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\tau = 0.20" + }, + { + "bbox": [ + 54, + 510, + 295, + 713 + ], + "type": "text", + "content": " for ScienceQA based on empirical evaluations to balance precision and recall. See Appendix E for selection details. We construct the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 339, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 400 + ], + "type": "text", + "content": "graph using DeepSeek-R1-70B and implement LightRAG's hybrid retrieval approach with Qwen2.5-7B. For graph construction and multimodal reasoning, we utilize " + }, + { + "bbox": [ + 313, + 339, + 555, + 400 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 313, + 339, + 555, + 400 + ], + "type": "text", + "content": " NVIDIA A100-80GB GPUs. Task-specific prompts are designed to assist LLMs in multimodal reasoning evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 400, + 420 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "content": "Multimodal Classification Tasks. We conduct multimodal classification experiments on the CrisisMMD dataset, evaluating both text-only LLMs and multimodal VLMs. Detailed comparative results are provided in Tables 1 and 2. For text-only LLMs, we adopt Qwen2.5-7B as the foundational reasoning model. Remarkably, the VaLiK-enhanced version achieves state-of-the-art (SOTA) performance matching that of the native Qwen2.5-72B model. The image-only KG constructed through VaLiK demonstrates an average accuracy improvement of " + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "inline_equation", + "content": "4.41\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "content": " across tasks, with the text-image variant attaining a " + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "inline_equation", + "content": "4.90\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "content": " enhancement. These improvements significantly surpass the " + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "content": " gain obtained by LightRAG using textual KG. We further validate VaLiK's cross-scale applicability through evaluations on Qwen2.5-32B and 72B architectures, observing consistent " + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "inline_equation", + "content": "2.0\\% - 2.5\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 653 + ], + "type": "text", + "content": " improvements. While not as significant as the 7B model's benefits, this shows that models that have substantial prior knowledge benefit less from external knowledge augmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Unlike text-only LLMs that depend on MMKGs for visual understanding, VLMs primarily benefit from KGs integration through outdated knowledge refreshment. Due to the inherent availability of visual features during inference, VaLiK's performance gains for VLMs remain con-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "986" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 338 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 338 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 338 + ], + "type": "table", + "html": "
Method#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Human [47]-90.2384.9787.4889.6087.5088.1091.5982.4288.40
GPT-4 [42]-84.0673.4587.3681.8770.7590.7384.6979.1082.69
CoT (GPT-3) [47]173B75.4470.8778.0974.6867.4379.9378.2369.6875.17
CoT (UnifiedQA) [47]223M71.0076.0478.9166.4266.5381.8177.0668.8274.11
CoT (GPT-4) [48]1T+85.4872.4490.2782.6571.4992.8986.6679.0483.99
DDCoT [85]175B80.1576.7282.8278.8972.5385.0282.8675.2180.15
Chameleon (ChatGPT) [48]175B+81.6270.6484.0079.7770.8086.6281.8676.5379.93
LG-VQA (BLIP-2) [24]---------86.32
LaVIN-13B [77]---------77.54
BLIP-2 [77]---------74.17
CCOT7B--------76.84
GraphVis [18]7B--------73.18
Qwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
Qwen2.5-72B72B79.6467.1084.9077.5665.0087.9380.2574.8578.37
Qwen2.5-7B (Mmkg) [45]7B73.9866.3778.1871.6564.3079.6576.5168.0373.47
Qwen2.5-7B (Visual Genome) [37]7B76.7867.0478.0974.0566.1979.7278.0869.6875.08
Qwen2.5-7B (VaLiK Text-only)7B84.5474.2486.9182.7472.5390.0384.5180.2882.98
Qwen2.5-7B (VaLiK Image-only)7B79.1471.5479.2777.1669.7283.1480.6573.9678.88
Qwen2.5-7B (VaLiK Text-Image)7B84.1575.1487.6482.9973.1889.6984.4080.9583.16
Qwen2.5-72B (VaLiK Text-Image)72B85.6175.9390.2784.4074.1792.3385.7982.9884.77
", + "image_path": "02345aa4ff865484870aefe25faa21813394a722f4f544f950b7f5f349a4dd9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 392 + ], + "type": "text", + "content": "Table 3. Performance comparison (\\%) on ScienceQA benchmark. #T-Params denotes trainable parameters. Categories: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG-Cap (image caption), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12). Method groups: (1) Human performance baseline, (2) Zero/Few-shot text-only LLMs, (3) Zero/Few-shot Multimodal VLMs, (4) LLMs enhanced with knowledge graphs for multimodal reasoning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "content": "strained compared to text-only counterparts. We separately applied VaLiK enhancement to Qwen2-VL-72B-Instruct and LLaVA-34B, obtaining distinct improvements: LLaVA-34B achieves accuracy gains of " + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "inline_equation", + "content": "2.41\\%" + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "content": " (image-only KG) and " + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "inline_equation", + "content": "3.59\\%" + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "content": " (text-image KG), while Qwen2-VL-72B-Instruct shows " + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "inline_equation", + "content": "1.77\\%" + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "inline_equation", + "content": "2.23\\%" + }, + { + "bbox": [ + 54, + 412, + 294, + 579 + ], + "type": "text", + "content": " improvements respectively under identical configurations. These experimental findings collectively demonstrate that VaLiK effectively extracts valuable signals from the training corpus and enables dynamic knowledge injection into VLMs during inference, thereby substantially alleviating hallucination phenomena. The differential improvements between Qwen2-VL-72B-Instruct and LLaVA-34B further validate the framework's adaptability across model architectures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 700 + ], + "type": "text", + "content": "Additionally, we analyze the results of LLMs without KG enhancement in the tables, which generally follow the scaling law [36]. However, DeepSeek-R1 shows anomalous behavior. Through testing, we find that its reasoning process may introduce complex information that interferes with its judgment. Furthermore, empirical results show that most baseline models achieve suboptimal performance without fine-tuning. In contrast, VaLiK's automated MMKG construction framework requires no task-specific adaptation yet delivers consistent improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "Multimodal Question Answering Tasks. We evalu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "content": "ated multimodal QA performance on the ScienceQA benchmark with Qwen2.5-7B and Qwen2.5-72B as base architectures, augmented by four knowledge sources: Mmkg, Visual Genome, text-only LightRAG and VaLiK. Compared to existing zero-shot/few-shot LLMs that not specifically optimized for scientific QA, our VaLiK-enhanced Qwen2.5-72B achieved SOTA performance on " + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "inline_equation", + "content": "62.5\\%" + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "content": " of subtasks, demonstrating particular strengths in multimodal reasoning scenarios requiring cross-modal alignment with an average accuracy gain of " + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "content": " over baseline models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "text", + "content": "Our study identifies a fundamental imbalance between textual and visual knowledge representations in ScienceQA. Text-only KGs (14k entities, 18k relations) exhibit " + }, + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "inline_equation", + "content": "8 \\times" + }, + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "text", + "content": " denser structured knowledge than image-only counterparts (3k concepts, 1k relations), explaining visual modality underperformance. Despite this gap, vision-KG-augmented Qwen2.5-7B still attains " + }, + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "inline_equation", + "content": "4.16\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 689 + ], + "type": "text", + "content": " accuracy gains over its non-enhanced version. Notably, our MMKG requires only 489MB storage for complete storage, while the scene graph component2 of Visual Genome alone occupies 739MB. This lightweight construction enables effective reasoning using only textual KG descriptions without raw images in resource-constrained scenarios." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 382, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 382, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 382, + 712 + ], + "type": "text", + "content": "2Visual Genome" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "987" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 152 + ], + "type": "table", + "html": "
TypeMethod#T-ParamSubjectContext ModalityGradeAverage
NATSOCLANTXTIMGNOG1-6G7-12
Image-OnlyQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B80.06 (↑3.86)70.30 (↑2.47)80.55 (↑3.28)78.05 (↑3.56)68.43 (↑2.64)83.76 (↑4.74)81.17 (↑3.45)72.71 (↑3.36)78.14 (↑3.42)
+ SV7B79.14 (↓0.92)71.54 (↑1.24)79.27 (↓1.28)77.16 (↓0.89)69.72 (↑1.29)83.14 (↓0.62)80.65 (↓0.52)73.96 (↑1.25)78.88 (↑0.74)
Text-ImageQwen2.5-7B7B76.2067.8377.2774.4965.7979.0277.7269.3574.72
+ CVs7B81.88 (↑5.68)73.00 (↑5.17)84.00 (↑6.73)80.55 (↑6.06)70.05 (↑4.26)87.11 (↑8.09)82.01 (↑4.29)77.98 (↑8.63)80.57 (↑5.85)
+ SV7B84.15 (↑2.27)75.14 (↑2.14)87.64 (↑3.64)82.99 (↑2.44)73.18 (↑3.13)89.69 (↑2.58)84.40 (↑2.39)80.95 (↑2.97)83.16 (↑2.59)
", + "image_path": "4646fcd105f2d0a9a95e374a10d683ca7fe9398a4c070703b010b8b6c67db1fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 209, + 297, + 282 + ], + "blocks": [ + { + "bbox": [ + 55, + 166, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 166, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 555, + 201 + ], + "type": "text", + "content": "Table 4. Ablation study on ScienceQA benchmark (CVs: CoE-based Vision-Language Models; SV: Similarly Verification). Performance metrics include: NAT (natural science), SOC (social science), LAN (language), TXT (text context), IMG (image context), NO (no context), G1-6 (grades 1-6), G7-12 (grades 7-12)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 209, + 297, + 282 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 297, + 282 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 297, + 282 + ], + "type": "table", + "html": "
TypeMethodTask 1 (%)Task 2 (%)Task 2-Merged (%)
Image-OnlyQwen2.5-7B65.0444.5245.33
+ CVs68.11 (↑3.07)47.00 (↑2.48)46.95 (↑1.62)
+ SV69.52 (↑1.41)49.54 (↑2.54)49.07 (↑2.12)
Text-ImageQwen2.5-7B65.0444.5245.33
+ CVs68.43 (↑3.39)48.61 (↑4.09)48.97 (↑3.64)
+ SV68.90 (↑0.47)50.02 (↑1.41)50.69 (↑1.72)
", + "image_path": "e95ed00b81ec004a5771de39ec640774b155153c405c18bddcaaaf4337a28b87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 297, + 282, + 308 + ], + "lines": [ + { + "bbox": [ + 68, + 297, + 282, + 308 + ], + "spans": [ + { + "bbox": [ + 68, + 297, + 282, + 308 + ], + "type": "text", + "content": "Table 5. Ablation study on CrisisMMD with Qwen2.5-7B." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 320, + 149, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 149, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 149, + 334 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "text", + "content": "Our ablation studies on CrisisMMD and ScienceQA demonstrate the specific roles of VaLiK's components. As shown in Table 4 and Table 5, the CVs (CoE-based VLM) module improves accuracy across all settings, with average gains of " + }, + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "inline_equation", + "content": "+3.05\\%" + }, + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "text", + "content": " on CrisisMMD and " + }, + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "inline_equation", + "content": "+4.63\\%" + }, + { + "bbox": [ + 55, + 338, + 296, + 615 + ], + "type": "text", + "content": " on ScienceQA tasks, validating visual descriptions enhance reasoning. However, the SV (Similarly Verification) module exhibits dual effects: it significantly improves CrisisMMD metrics by pruning redundant textual descriptions, yet slightly degrades ScienceQA's image-only natural science reasoning. We hypothesize this discrepancy arises from dataset characteristics: CrisisMMD's generated captions contain substantially more redundant content, whereas ScienceQA's simpler visual scenes yield shorter descriptions. Pruning these shorter descriptions risks over-removal of critical semantics. Furthermore, different types of KGs influence the effectiveness of the components: CVs achieve greater gains in CrisisMMD's text-image fusion as original text provides complementary context, while SV shows reduced effectiveness, likely due to occasional over-pruning of cross-modal linkages. Nevertheless, both modules collectively enhance performance across configurations, demonstrating their synergistic yet context-sensitive nature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 158, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 158, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 158, + 636 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Impact of VLM Quantity and Types. We evaluate the impact of varying quantities and types of VLMs on the CVs module. Our experiments reveal that Qwen2-VL generates the most visual descriptions, followed by LLaVA, while BLIP-2 produces the fewest. However, BLIP-2 demonstrates superior capability in extracting critical information" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 435, + 297 + ], + "blocks": [ + { + "bbox": [ + 316, + 209, + 435, + 297 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 435, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 435, + 297 + ], + "type": "image", + "image_path": "0c11325abdab9142cb8a1d96545e124e536e4c45e57e562428116da00e71ee8c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 306, + 542, + 319 + ], + "lines": [ + { + "bbox": [ + 325, + 306, + 542, + 319 + ], + "spans": [ + { + "bbox": [ + 325, + 306, + 542, + 319 + ], + "type": "text", + "content": "Figure 4. Impact analysis of VLM quantity on CrisisMMD." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 209, + 553, + 297 + ], + "blocks": [ + { + "bbox": [ + 435, + 209, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 435, + 209, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 435, + 209, + 553, + 297 + ], + "type": "image", + "image_path": "9559482c345cc743f96b22e87bcf418e0d4c3d23fddec5c9335bc5e04f78ee59.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 422 + ], + "type": "text", + "content": "and identifying key entity relationships within images. We therefore adopt BLIP-2 as the primary model, with LLaVA or Qwen2-VL serving as secondary/tertiary components. Adding more VLMs yields diminishing returns, due to limited entities in current images, though we hypothesize their benefits would increase for complex visual scenes with richer semantic content. This phenomenon is empirically validated by our quantitative results in Figure 4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 425, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 473 + ], + "type": "text", + "content": "Computational Costs. Due to space limitations, we provide an overview of VaLiK's computational costs in Appendix F. Our method is significantly more cost-effective than manual annotation or LLM fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 388, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Multimodal reasoning in LLMs is constrained by incomplete knowledge and hallucination artifacts, limitations that persist because textual KGs cannot bridge visual-textual semantics due to their modality isolation. To bridge this gap, we propose VaLiK, a framework for constructing MMKGs through vision-language alignment, eliminating dependency on manual annotations while resolving visual-textual semantic inconsistencies. By integrating a cascade of pretrained VLMs and cross-modal verification, VaLiK converts images into structured knowledge while filtering noise. The resulting graphs enhance LLMs' reasoning with minimal storage overhead. Experiments on multimodal reasoning benchmarks show SOTA performance. VaLiK's modular design supports adaptability across domains, offering a scalable solution for autonomous knowledge synthesis. This work advances multimodal AI systems by enabling efficient integration of visual and textual data." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "988" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 151 + ], + "type": "text", + "content": "The research was supported by Shanghai Artificial Intelligence Laboratory, the National Key R&D Program of China (Grant No. 2022ZD0160201) and the Science and Technology Commission of Shanghai Municipality (Grant No. 22DZ1100102)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 235 + ], + "type": "text", + "content": "[1] Mahdi Abavisani, Liwei Wu, Shengli Hu, Joel Tetreault, and Alejandro Jaimes. Multimodal categorization of crisis events in social media. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 296, + 290 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 295, + 335 + ], + "type": "text", + "content": "[3] Firoj Alam, Ferda Ofli, and Muhammad Imran. Crisismmd: Multimodal twitter datasets from natural disasters. Proceedings of the International AAAI Conference on Web and Social Media, 12(1), 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 337, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 337, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 296, + 456 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikol aj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716-23736. Curran Associates, Inc., 2022. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 458, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 490 + ], + "type": "text", + "content": "[5] Razvan Azamfirei, Sapna R Kudchadkar, and James Fackler. Large language models and the perils of their hallucinations. Critical Care, 27(1):120, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 545 + ], + "type": "text", + "content": "[6] Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-augmented language model prompting for zero-shot knowledge graph question answering. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 590 + ], + "type": "text", + "content": "[7] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2):423-443, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 635 + ], + "type": "text", + "content": "[8] Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. Proceedings of the AAAI Conference on Artificial Intelligence, 38(16):17754-17762, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 689 + ], + "type": "text", + "content": "[9] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners. Advances in neural information processing systems (NeurIPS), 33:22243-22255, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[10] Xiang Chen, Ningyu Zhang, Lei Li, Shumin Deng, Chuanqi Tan, Changliang Xu, Fei Huang, Luo Si, and Huajun Chen." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "Hybrid transformer with multi-level fusion for multimodal knowledge graph completion. In Proceedings of the International Conference on Research and Development in Information Retrieval (SIGIR), pages 904-915, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 172 + ], + "type": "text", + "content": "[11] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 228 + ], + "type": "text", + "content": "[12] Xiang Chen, Jingtian Zhang, Xiaohan Wang, Ningyu Zhang, Tongtong Wu, Yuxiang Wang, Yongheng Wang, and Huajun Chen. Continual multimodal knowledge graph construction. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[13] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "text", + "content": "[14] Yong Chen, Xinkai Ge, Shengli Yang, Linmei Hu, Jie Li, and Jinwen Zhang. A survey on multimodal knowledge graphs: Construction, completion and applications. Mathematics, 11 (8), 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[15] Zhuo Chen, Yichi Zhang, Yin Fang, Yuxia Geng, Lingbing Guo, Xiang Chen, Qian Li, Wen Zhang, Jiaoyan Chen, Yushan Zhu, et al. Knowledge graphs meet multimodal learning: A comprehensive survey. arXiv preprint arXiv:2402.05391, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[16] Shiyao Cui, Jiangxia Cao, Xin Cong, Jiawei Sheng, Quanggang Li, Tingwen Liu, and Jinqiao Shi. Enhancing multimodal entity and relation extraction with variational information bottleneck. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 32:1274-1285, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[17] Xiaohui Cui, Xiaolong Qu, Dongmei Li, Yu Yang, Yuxun Li, and Xiaoping Zhang. Mkgcn: Multi-modal knowledge graph convolutional network for music recommender systems. *Electronics*, 12(12), 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 536 + ], + "type": "text", + "content": "[18] Yihe Deng, Chenchen Ye, Zijie Huang, Mingyu Derek Ma, Yiwen Kou, and Wei Wang. Graphvis: Boosting llms with visual knowledge graph integration. In Advances in Neural Information Processing Systems, pages 67511-67534. Curran Associates, Inc., 2024. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 581 + ], + "type": "text", + "content": "[19] Tausif Diwan, G. Anirudh, and Jitendra V. Tembhurne. Object detection using yolo: challenges, architectural successors, datasets and applications. Multimedia Tools Appl., 82 (6):9243-9275, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "text", + "content": "[21] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[22] Quan Fang, Xiaowei Zhang, Jun Hu, Xian Wu, and Changsheng Xu. Contrastive multi-modal knowledge graph representation learning. IEEE Transactions on Knowledge and Data Engineering, 35(9):8983-8996, 2023. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "989" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[23] Duoduo Feng, Xiangteng He, and Yuxin Peng. Mkvse: Multimodal knowledge enhanced visual-semantic embedding for image-text retrieval. ACM Trans. Multimedia Comput. Commun. Appl., 19(5), 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[24] Deepanway Ghosal, Navonil Majumder, Roy Ka-Wei Lee, Rada Mihalcea, and Soujanya Poria. Language guided visual question answering: Elevate your multimodal language model using knowledge-enriched prompts. arXiv preprint arXiv:2310.20159, 2023. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[25] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[26] Yunfei Guo, Fei Yin, Xiao-hui Li, Xudong Yan, Tao Xue, Shuqi Mei, and Cheng-Lin Liu. Visual traffic knowledge graph generation from scene images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21604-21613, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "text", + "content": "[27] ZIRUI GUO, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. LightRAG: Simple and fast retrieval-augmented generation, 2024. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[28] Lavdim Halilaj, Juergen Luettin, Sebastian Monka, Cory Henson, and Stefan Schmid. Knowledge graph-based integration of autonomous driving datasets. International Journal of Semantic Computing, 17(02):249-271, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 403 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "text", + "content": "[30] Yang Hu, Guihua Wen, Adriane Chapman, Pei Yang, Mingnan Luo, Yingxue Xu, Dan Dai, and Wendy Hall. Graph-based visual-semantic entanglement network for zero-shot image recognition. IEEE Transactions on Multimedia, 24: 2473-2487, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 536 + ], + "type": "text", + "content": "[31] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Lee. LLM-adapters: An adapter family for parameter-efficient fine-tuning of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 5254-5276, Singapore, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 581 + ], + "type": "text", + "content": "[32] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 1, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 635 + ], + "type": "text", + "content": "[33] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "text", + "content": "[34] Adam Tauman Kalai and Santosh S. Vempala. Calibrated language models must hallucinate. In Proceedings of the 56th Annual ACM Symposium on Theory of Computing, page 160-171, New York, NY, USA, 2024. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[35] Amar Viswanathan Kannan, Dmitriy Fradkin, Ioannis Akrotirianakis, Tugba Kulahcioglu, Arquimedes Canedo," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 139 + ], + "type": "text", + "content": "Aditi Roy, Shih-Yuan Yu, Malawade Arnav, and Mohammad Abdullah Al Faruque. Multimodal knowledge graph for deep learning papers and code. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 3417-3420, New York, NY, USA, 2020. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 194 + ], + "type": "text", + "content": "[36] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 262 + ], + "type": "text", + "content": "[37] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 330 + ], + "type": "text", + "content": "[38] Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. Multimodal reasoning with multimodal knowledge graph. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10767-10782, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "text", + "content": "[39] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[40] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Proceedings of the International Conference on Machine Learning (ICML), pages 19730–19742, 2023. 1, 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[41] Xin Li, Dongze Lian, Zhihe Lu, Jiawang Bai, Zhibo Chen, and Xinchao Wang. Graphadapter: Tuning vision-language models with dual knowledge graph. In Advances in Neural Information Processing Systems, pages 13448-13466. Curran Associates, Inc., 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[43] Junming Liu, Yanting Gao, Siyuan Meng, Yifei Sun, Aoqi Wu, Yufei Jin, Yirong Chen, Ding Wang, and Guosun Zeng. Mosaic: Data-free knowledge distillation via mixture-of-experts for heterogeneous distributed environments. arXiv preprint arXiv:2505.19699, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 645 + ], + "type": "text", + "content": "[44] Junming Liu, Guosun Zeng, Ding Wang, Yanting Gao, and Yufei Jin. Fedrecon: Missing modality reconstruction in distributed heterogeneous environments. arXiv preprint arXiv:2504.09941, 2025.3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[45] Ye Liu, Hui Li, Alberto Garcia-Duran, Mathias Niepert, Daniel Onoro-Rubio, and David S Rosenblum. Mmkg: multi-modal knowledge graphs. In The Semantic Web: 16th International Conference, ESWC 2019, Portoroz, Slovenia, June 2-6, 2019, Proceedings 16, pages 459-474. Springer, 2019. 2, 6, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "990" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[46] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "type": "text", + "content": "[47] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In Advances in Neural Information Processing Systems, pages 2507–2521. Curran Associates, Inc., 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "text", + "content": "[48] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In Advances in Neural Information Processing Systems, pages 43447-43478. Curran Associates, Inc., 2023. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[49] Gen Luo, Yiyi Zhou, Tianhe Ren, Shengxin Chen, Xiaoshuai Sun, and Rongrong Ji. Cheap and quick: Efficient vision-language instruction tuning for large language models. In Advances in Neural Information Processing Systems, pages 29615-29627. Curran Associates, Inc., 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[50] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 420 + ], + "type": "text", + "content": "[51] Runqing Miao, Qingxuan Jia, Fuchun Sun, Gang Chen, Haiming Huang, and Shengyi Miao. Semantic representation of robot manipulation with knowledge graph. Entropy, 25(4), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 477 + ], + "type": "text", + "content": "[52] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14420-14431, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[53] Aske Plaat, Annie Wong, Suzan Verberne, Joost Broekens, Niki van Stein, and Thomas Back. Reasoning with large language models, a survey. arXiv preprint arXiv:2407.11511, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 590 + ], + "type": "text", + "content": "[54] Bryan A. Plummer, Liwei Wang, Chris M. Cervantes, Juan C. Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 635 + ], + "type": "text", + "content": "[55] Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. Knowledge-aware multi-modal adaptive graph convolutional networks for fake news detection. ACM Trans. Multimedia Comput. Commun. Appl., 17(3), 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 713 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3, 6" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[57] Brian Reily, Christopher Reardon, and Hao Zhang. Representing multi-robot structure through multimodal graph embedding for the selection of robot teams. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 5576–5582, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[58] Joshua Robinson, Christopher Michael Ryting, and David Wingate. Leveraging large language models for multiple choice question answering. In Proceedings of the International Conference on Learning Representations (ICLR), 2023. 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 184, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 260 + ], + "type": "text", + "content": "[59] Hrituraj Singh, Anshul Nasery, Denil Mehta, Aishwarya Agarwal, Jatin Lamba, and Balaji Vasan Srinivasan. MIMOQA: Multimodal input multimodal output question answering. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5317-5332, Online, 2021. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[60] Yaoxian Song, Penglei Sun, Haoyu Liu, Zhixu Li, Wei Song, Yanghua Xiao, and Xiaofang Zhou. Scene-driven multimodal knowledge graph construction for embodied ai. IEEE Transactions on Knowledge and Data Engineering, 36(11): 6962-6976, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 392 + ], + "type": "text", + "content": "[61] Rui Sun, Xuezhi Cao, Yan Zhao, Junchen Wan, Kun Zhou, Fuzheng Zhang, Zhongyuan Wang, and Kai Zheng. Multimodal knowledge graphs for recommender systems. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, page 1405-1414, New York, NY, USA, 2020. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 448 + ], + "type": "text", + "content": "[62] Yu Sun, Shuohuan Wang, Shikun Feng, Siyu Ding, Chao Pang, Junyuan Shang, Jiaxiang Liu, Xuyi Chen, Yanbin Zhao, Yuxiang Lu, et al. Ernie 3.0: Large-scale knowledge enhanced pre-training for language understanding and generation. arXiv preprint arXiv:2107.02137, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[63] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 580 + ], + "type": "text", + "content": "[64] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, ADITHYA JAIRAM IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[66] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[67] Peng Wang, Qi Wu, Chunhua Shen, Anthony Dick, and Anton van den Hengel. Fvqa: Fact-based visual question an" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "991" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "swering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2413-2427, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 151 + ], + "type": "text", + "content": "[68] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 218 + ], + "type": "text", + "content": "[69] Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In Proceedings of the 31st ACM International Conference on Multimedia, page 2391–2399, New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 262 + ], + "type": "text", + "content": "[70] Yuequn Wang, Liyan Dong, Hao Zhang, Xintao Ma, Yongli Li, and Minghui Sun. An enhanced multi-modal recommendation based on alternate training with knowledge graph representation. IEEE Access, 8:213012-213026, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 331 + ], + "type": "text", + "content": "[71] Tao Wu, Mengze Li, Jingyuan Chen, Wei Ji, Wang Lin, Jinyang Gao, Kun Kuang, Zhou Zhao, and Fei Wu. Semantic alignment for multimodal large language models. In Proceedings of the 32nd ACM International Conference on Multimedia, page 3489-3498, New York, NY, USA, 2024. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 376 + ], + "type": "text", + "content": "[72] Yike Wu, Nan Hu, Guilin Qi, Sheng Bi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewrite-answer: A kg-to-text enhanced llms framework for knowledge graph question answering. arXiv preprint arXiv:2309.11206, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 442 + ], + "type": "text", + "content": "[73] Ziyang Xiao, Dongxiang Zhang, Yangjun Wu, Lilin Xu, Yuan Jessica Wang, Xiongwei Han, Xiaojin Fu, Tao Zhong, Jia Zeng, Mingli Song, and Gang Chen. Chain-of-experts: When LLMs meet complex operations research problems. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 532 + ], + "type": "text", + "content": "[74] Dexuan Xu, Yanyuan Chen, Jieyi Wang, Yue Huang, Hanpin Wang, Zhi Jin, Hongxing Wang, Weihua Yue, Jing He, Hang Li, and Yu Huang. MLeVLM: Improve multi-level progressive capabilities based on multimodal large language model for medical visual question answering. In Findings of the Association for Computational Linguistics: ACL 2024, pages 4977-4997, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 533, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 295, + 577 + ], + "type": "text", + "content": "[75] Peng Xu, Xiatian Zhu, and David A. Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 578, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 295, + 622 + ], + "type": "text", + "content": "[76] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 678 + ], + "type": "text", + "content": "[77] Xiaocui Yang, Wenfang Wu, Shi Feng, Ming Wang, Daling Wang, Yang Li, Qi Sun, Yifei Zhang, Xiaoming Fu, and Soujanya Poria. Mm-bigbench: Evaluating multimodal models on multimodal content comprehension tasks. arXiv preprint arXiv:2310.09036, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[78] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 710 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[79] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 184 + ], + "type": "text", + "content": "[80] Jingtong Yue, Zhiwei Lin, Xin Lin, Xiaoyu Zhou, Xiangtai Li, Lu Qi, Yongtao Wang, and Ming-Hsuan Yang. RobuR-CDet: Enhancing robustness of radar-camera fusion in bird's eye view for 3d object detection. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[81] Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Huajun Chen, and Wen Zhang. Mygo: Discrete modality information as fine-grained tokens for multi-modal knowledge graph completion. CoRR, abs/2404.09468, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "text", + "content": "[82] Zhuosheng Zhang, Aston Zhang, Mu Li, hai zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. Transactions on Machine Learning Research, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[83] Jiabao Zhao, Xin Lin, Jie Zhou, Jing Yang, Liang He, and Zhaohui Yang. Knowledge-based fine-grained classification for few-shot learning. In 2020 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 397 + ], + "type": "text", + "content": "[84] Changmeng Zheng, Junhao Feng, Ze Fu, Yi Cai, Qing Li, and Tao Wang. Multimodal relation extraction with efficient graph alignment. In Proceedings of the 29th ACM International Conference on Multimedia, page 5298-5306, New York, NY, USA, 2021. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[85] Ge Zheng, Bin Yang, Jiajin Tang, Hong-Yu Zhou, and Sibei Yang. Ddcot: Duty-distinct chain-of-thought prompting for multimodal reasoning in language models. In Advances in Neural Information Processing Systems, pages 5168-5191. Curran Associates, Inc., 2023. 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "text", + "content": "[86] Juncheng Zheng, Meiyu Liang, Yang Yu, Yawen Li, and Zhe Xue. Knowledge graph enhanced multimodal transformer for image-text retrieval. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 70-82, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 563 + ], + "type": "text", + "content": "[87] Deyao Zhu, Jun Chen, Kilichbek Haydarov, Xiaogian Shen, Wenxuan Zhang, and Mohamed Elhoseiny. Chatgpt asks, blip-2 answers: Automatic questioning towards enriched visual descriptions. arXiv preprint arXiv:2303.06594, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 565, + 553, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 619 + ], + "type": "text", + "content": "[88] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 621, + 553, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 553, + 676 + ], + "type": "text", + "content": "[89] Xiangru Zhu, Zhixu Li, Xiaodan Wang, Xueyao Jiang, Penglei Sun, Xuwu Wang, Yanghua Xiao, and Nicholas Jing Yuan. Multi-modal knowledge graph construction and application: A survey. IEEE Transactions on Knowledge and Data Engineering, 36(2):715-735, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 710 + ], + "type": "text", + "content": "[90] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "992" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_content_list.json b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f213e4e81aba36403d15f69fbf661bac41fc190a --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_content_list.json @@ -0,0 +1,1745 @@ +[ + { + "type": "text", + "text": "All Parts Matter: A Unified Mask-Free Virtual Try-On Framework", + "text_level": 1, + "bbox": [ + 158, + 130, + 841, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenghu Du $^{1,3}$ Shengwu Xiong $^{2*}$ Yi Rong $^{1,3*}$ $^{1}$ School of Computer Science and Artificial Intelligence, Wuhan University of Technology \n $^{2}$ Interdisciplinary Artificial Intelligence Research Institute, Wuhan College \n $^{3}$ Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 142, + 179, + 854, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{duch, xiongsw, wrong}@whut.edu.cn", + "bbox": [ + 344, + 253, + 643, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current virtual try-on methods primarily enhance performance through network optimization, employing strategies such as coarse-to-fine structures and ReferenceNet to inject clothing information. However, these methods are often constrained by the limited quantity and diversity of training samples, which ultimately restricts their potential for further improvement. To address this challenge, we propose a unified, mask-free virtual try-on framework. Our approach leverages the inherent strengths of latent diffusion models to enhance the capability of each pipeline component to accurately model the target distribution, thereby achieving superior performance. Specifically, we introduce a text-driven pseudo-input preparation strategy that significantly increases the diversity of clothing regions within the generated person pseudo-samples. This encourages the generator to focus on variations in these areas and improves the model's generalization capability. Within the generator itself, we develop a gated manipulation mechanism to prevent weight forgetting and reduce training costs. Furthermore, we incorporate a texture-aware injection module to explicitly integrate human-perceptible clothing texture information into the generation process. During inference, we propose a refining conditional inference strategy that mitigates the randomness introduced by Gaussian noise. This effectively preserves identity information and fine clothing details in the final results. Extensive experiments demonstrate that our method outperforms existing state-of-the-art virtual try-on approaches.", + "bbox": [ + 88, + 335, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 789, + 220, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image-based virtual try-on (VTON) aims to realistically transfer a target garment onto the corresponding region of a reference person [3, 8, 14, 22, 39]. Owing to its strong application prospects and commercial potential, VTON has", + "bbox": [ + 89, + 814, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attracted growing research interest. Leveraging the powerful generative capabilities of diffusion models [20, 41], recent VTON methods have significantly narrowed the gap to real-world fitting.", + "bbox": [ + 511, + 305, + 906, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous VTON methods are broadly categorized into mask-based and mask-free paradigms. Mask-based approaches [3, 14, 22, 37] typically operate in a self-supervised fashion: a clothing-agnostic mask (Figure 2) is first applied to remove the garment region and adjacent body parts from the reference image, and the generator is then trained to inpaint the masked areas, thereby learning to perform the virtual try-on task. Although effective and widely adopted, these methods rely heavily on obtaining accurate masks for each reference person; even minor mask errors can introduce conspicuous artifacts [13, 21]. To mitigate mask-induced artifacts, mask-free methods grounded in pseudo-sample [8, 13, 18] have been proposed. A mask-based VTON model is first pre-trained to synthesize try-on images for arbitrary garments on the reference person, yielding paired pseudo-samples. These pseudo-samples are then fed as inputs, while the original reference image serves as the reconstruction target (ground truth), to train a main generator. Although this pipeline obviates the need for precise masks, they are limited to a single, fixed transformation for each pseudo-sample/reference-person pair. Moreover, because both garments and reference images are typically drawn from the same dataset, the generator overfits to these specific items and generalizes poorly to unseen clothing.", + "bbox": [ + 511, + 368, + 908, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are also some other methods attempt to boost performance through architectural improvements—such as coarse-to-fine networks and ReferenceNet-based clothing-feature injection [22, 37, 40]. Yet these refinements remain constrained by the input materials (human parsing, etc), causing the generator to converge to sub-optimal solutions. Finally, the randomness of Gaussian-noise sampling during inference further degrades output quality [36]—a problem often overlooked by prior work.", + "bbox": [ + 511, + 732, + 908, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose a unified mask-free virtual try-on framework that leverages the intrinsic", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19525", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strengths of latent diffusion models to enable every pipeline component to better approximate the target data distribution and thus improve overall performance. Specifically, we introduce a text-driven pseudo-input preparation strategy to augment sample diversity by enriching the clothing regions in each pseudo-sample (see Figure 2). First, a pretrained Stable Diffusion inpainting model is conditioned on text prompts generated by a large language model (e.g., DeepSeek or GPT-series models) to produce multiple try-on results for the same identity. We then perform sample selection to filter out implausible images and retain the most realistic candidate as the final pseudo-sample. This process maximizes the variability of clothing regions and mitigates overfitting of the generator during training.", + "bbox": [ + 89, + 90, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During training phase, we leverage the off-the-shelf inpainting try-on model as the generator, thereby reducing training costs. To convert this model into a mask-free generator, we introduce mask-free generation learning process. Concretely, a gated manipulation mechanism fuses clothing and person latent representations in latent space, preventing catastrophic forgetting when clothing conditioning is added and avoiding structural changes that inflate training cost. Complementarily, texture-aware injection module explicitly injects fine-grained texture cues into the cross-attention layers, sharpening the generator's perception of garment details.", + "bbox": [ + 89, + 301, + 480, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "During inference phase, to reduce the randomness of Gaussian noise, we propose a refining conditional inference in our optimization process. By optimizing the initial Gaussian noise to make it as close as possible to the distribution of real person samples, helping the generator improve the quality of the synthesized results.", + "bbox": [ + 89, + 484, + 480, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this paper makes the following main contributions:", + "bbox": [ + 89, + 575, + 480, + 602 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified mask-free virtual try-on framework, which improves the generator's performance from multiple aspects of both the training and inference processes.", + "- We present a text-driven pseudo-input preparation approach to improve the generator's generalization ability.", + "- We propose a gated manipulation to prevent weight forgetting and cut training costs.", + "- We introduce a texture-aware injection to explicitly add human-perceptible clothing texture information.", + "- We propose a refining conditional inference approach to reduce Gaussian noise randomness." + ], + "bbox": [ + 91, + 604, + 480, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 230, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mask-Free Virtual Try-On. This task requires performing virtual try-on at inference time using only a clothing image and a person image—no masks, human parsing, or other auxiliary inputs. Issenhuth et al. [21] introduced the first mask-free method, WUTON, which attains fully supervised training by distilling knowledge from a mask-based teacher", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that generates paired training samples. Because the synthetic images are treated as ground truth for the student, the teacher's defects directly bound the student's performance. Ge et al. [13] relaxed this constraint with PFAFN, feeding pseudo-samples as inputs instead of supervision, thereby allowing the student to learn a more accurate mapping. He et al. [18] extended this paradigm in Flow-Style, adding explicit garment-deformation optimization to improve fit and realism. More recently, Du et al. [8] proposed USC-PFN, a two-stage architecture that employs cycle-consistency and a weight-shared generator to refine the garment region and try-on result iteratively. In contrast to these prior works, we present a unified mask-free framework that optimizes the pseudo-sample and enhances both training and inference through multi-stage optimization, yielding higher fidelity and better generalization.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable Diffusion Models. Recent advances in diffusion models demonstrate that generation can be guided by modalities beyond text. Specifically, the integration of external models to inject external features into diffusion models has produced remarkable results. For example, ControlNet [47] introduced an external model that leverages images, skeletons, and edge maps as conditional inputs to generate corresponding outputs. IP-adaptor [45] proposed a method for generating images similar to a given reference image. T2I-adapter [31] also introduced an external, more lightweight model that leverages conditional inputs to generate corresponding images. These approaches have effectively manipulated image generation by injecting feature values via external models. Instead of adding extra network branches, following the optimization idea of latent features in latent space of [9], we embed the try-on conditions directly into the latent space, yielding a high-performance virtual try-on model with minimal overhead.", + "bbox": [ + 511, + 352, + 903, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary: Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 638, + 854, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Latent Diffusion Models (LDMs) [33] are a class of generative models that learn to synthesize data $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ by denoising latent representations $\\mathbf{z} \\in \\mathbb{R}^{h \\times w \\times c}$ in a compressed space, where typically $h \\ll H$ and $w \\ll W$ . The compression is achieved via a pre-trained autoencoder with KL-regularization (e.g. VQGAN), consisting of an encoder $\\mathcal{E}_{VAE}(\\cdot)$ and decoder $\\mathcal{D}_{VAE}(\\cdot)$ , such that", + "bbox": [ + 511, + 662, + 903, + 770 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {0} = \\mathcal {E} _ {V A E} (\\mathbf {x}), \\quad \\hat {\\mathbf {x}} = \\mathcal {D} _ {V A E} (\\mathbf {z} _ {0}), \\quad \\text {w i t h} \\quad \\hat {\\mathbf {x}} \\approx \\mathbf {x}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 782, + 903, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The latent-space diffusion process is modeled as a Markov chain that progressively adds Gaussian noise to $\\mathbf{z}_0$ (the encoded data) over $T$ steps, producing a sequence of noisy latents $\\{\\mathbf{z}_t\\}_{t=0}^T$ :", + "bbox": [ + 511, + 810, + 903, + 871 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 883, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19526", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb3df232a54430247be2ac1461589a8fc6ca3b98bc550ab2988cd9d38da90621.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 416, + 223 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/664bdaa56cd71151f8a51cab7f18780c1be152d062ae9027c480ef96a6052a09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 88, + 903, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fd003c139c0430488fc11d5746c5aeebb55e665dea8c9170bab02868069fc7fd.jpg", + "image_caption": [ + "Sect. 4.1: Text-Driven Pseudo-Input Preparation" + ], + "image_footnote": [], + "bbox": [ + 94, + 223, + 418, + 291 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/57c3c7b1cf6fe27721a00d9327d80656355970e8076e3e5194e1601fb384c5ec.jpg", + "image_caption": [ + "Sect. 4.3: Optimization", + "Figure 1. Overview of the proposed virtual try-on framework. It consists of three sub-pipelines, namely Text-Driven Pseudo-Input Preparation, Mask-Free Generation Learning, and Optimization. In text-driven pseudo-input preparation (Sect. 4.1), we first use a pre-trained Stable Diffusion inpainting model to generate $N$ different pseudo-samples $\\{I_i\\}_{i=0}^N$ based on prompts generated by DeepSeek and various noises. Then, a sample selection method is proposed to pick the best one $I_{best}$ from the $N$ pseudo-samples for the next step. In mask-free generation learning (Sect. 4.2), we design a mask-free training approach that initially uses gated manipulation to fuse clothing and person representations to prevent catastrophic forgetting, and then injects extracted clothing texture features through cross-attention modules. In optimization (Sect. 4.3), we first train a noise optimization MLP to obtain the optimized Gaussian noise that preserves as much identity and fine-grained clothing cues as possible, and then use it for the inference phase." + ], + "image_footnote": [], + "bbox": [ + 421, + 224, + 903, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\{\\beta_t\\in (0,1)\\}_{t = 1}^T$ is a fixed variance schedule. The forward process can be marginalized in closed form as", + "bbox": [ + 89, + 452, + 482, + 483 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 494, + 482, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\alpha_{t} = 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}$ . LDMs optimize a variational lower bound (VLB) on the log-likelihood of the data, simplified to the noise-prediction objective (DDPM-style):", + "bbox": [ + 89, + 523, + 483, + 584 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {z} _ {0}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 594, + 480, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ is the ground-truth noise and $\\epsilon_{\\theta}(\\cdot)$ is a U-Net parameterized by $\\theta$ .", + "bbox": [ + 89, + 628, + 483, + 659 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 672, + 225, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. This task aims to transfer the target clothing from an arbitrary clothing image $C_{un} \\in \\mathbb{R}^{H \\times W \\times 3}$ onto a reference person image $P \\in \\mathbb{R}^{H \\times W \\times 3}$ , producing a try-on result $I \\in \\mathbb{R}^{H \\times W \\times 3}$ . To train the mask-free model, $I$ (generated from the off-the-shelf mask-based try-on method) is first used as pseudo-input, we construct triplets $\\{C, I, P\\}$ for fully supervised learning of the generator $G$ with parameters $\\theta$ , i.e.,", + "bbox": [ + 89, + 696, + 482, + 818 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {\\boldsymbol {C}, \\boldsymbol {I}, \\boldsymbol {P}} \\ell_ {\\text {d i s t}} (G (\\boldsymbol {C}, \\boldsymbol {I}), \\boldsymbol {P}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 827, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\ell_{dist}$ represents a suitable loss function (e.g., L1, MSE) between the generated image $G(C,I)$ and the", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ground-truth image $\\pmb{P}$ . The clothing image $C \\in \\mathbb{R}^{H \\times W \\times 3}$ and the clothes worn by $\\pmb{P}$ are identical and also sourced from the dataset. Sections that follow detail how we obtain a suitable pseudo-input $\\pmb{I}$ for efficient training of the generator $G$ . An overview of our pipeline—comprising Text-Driven Pseudo-Input Preparation (Sect. 4.1), Mask-Free Generation Learning (Sect. 4.2), and Refining Conditional Inference (Sect. 4.3)—is provided in Figure 1.", + "bbox": [ + 511, + 452, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Text-Driven Pseudo-Input Preparation", + "text_level": 1, + "bbox": [ + 511, + 584, + 846, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous mask-free methods [13, 18] rely on off-the-shelf try-on networks to generate the pseudo-input. Although training converges, the rigid one-to-one pairing of $C_{un} / I$ and $P$ drives the generator $G$ toward sub-optimal solutions and limits generalization. Therefore, diversifying the clothing regions of $I$ (see Figure 2) is crucial for improving model performance in mask-free try-ons. By increasing clothing variability, the generator $G$ can focus more effectively on clothing features and details.", + "bbox": [ + 511, + 606, + 906, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Automatic Sample Generation. To increase the visual diversity of $\\pmb{I}$ and mitigate the generalization drop caused by the overlap between $C_{un}$ (for generating $\\pmb{I}$ ) and $\\pmb{C}$ , we leverage a Stable Diffusion inpainting model [37], $\\mathcal{F}_{\\mathrm{SD}}$ , which possesses strong generative priors, together with the large language model DeepSeek2. The resulting pipeline is", + "bbox": [ + 511, + 762, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://civitai.com/models/120149/controlnet-for-densepose", + "bbox": [ + 511, + 862, + 903, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://chat_deepseek.com", + "bbox": [ + 532, + 887, + 736, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19527", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "summarized in Equation (6).", + "bbox": [ + 89, + 90, + 282, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {I} _ {i} \\right\\} _ {i = 0} ^ {N} = \\left\\{\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {i}\\right) \\right\\} _ {i = 0} ^ {N}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 114, + 482, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P^{\\star}$ denotes the clothing-agnostic person representation, consisting of the clothing-agnostic person image, the clothing-agnostic mask, and the DensePose map [15]. $\\xi$ is a random Gaussian noise sample, $\\xi \\sim \\mathcal{N}(0,\\mathbf{I})$ . $O$ is a text prompt constructed from the template \"A photo of a person wearing {upper garment description}\", with the description generated by DeepSeek to cover a wide range of clothing styles. For each prompt, we synthesize $N = 5$ images and select the best try-on result $I_{best}$ . The resulting triplets $\\{C,I_{best},P\\}$ are used for training.", + "bbox": [ + 89, + 143, + 483, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sample Selection. To this end, we adopt the Human Preference Score $\\nu 2$ (HPS v2) model [38] to rank the generated images $\\{I_i\\}_{i=1}^N$ according to their alignment with the corresponding prompt. HPS v2 employs an image encoder $E_{IMG}(\\cdot)$ and a text encoder $E_{TEXT}(\\cdot)$ to map images and prompts into a joint embedding space, where their cosine similarity serves as the preference score, as formulated in Equation (7).", + "bbox": [ + 89, + 314, + 483, + 434 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} ^ {\\eta} (\\boldsymbol {I} _ {i}, O) = E _ {I M G} (\\boldsymbol {I} _ {i}) \\cdot E _ {T E X T} (O) \\times 1 0 0. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 445, + 482, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We select the image achieving the highest score as the final sample. However, some $I_{i}$ while semantically consistent with the prompt (see column 5 of Figure 2)—exhibit implausible appearances in the inpainting regions, such as distorted limbs. To address this issue, before using HPS v2, we introduce an appearance rationality discriminator $\\mathcal{D}_r$ that outputs a rationality score $s^r(I_i) = \\mathcal{D}_r(I_i)$ for each candidate image. This discriminator is trained by minimizing the objective given in Equation (8).", + "bbox": [ + 89, + 473, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\mathcal {D} _ {r}} - \\mathbb {E} _ {\\boldsymbol {P}, O} \\left[ \\log \\left(\\mathcal {D} _ {r} \\left(\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {\\text {b e s t}}\\right)\\right)\\right) \\right. \\tag {8} \\\\ \\left. + \\log \\left(1 - \\mathcal {D} _ {r} \\big (\\mathcal {F} _ {S D} (\\boldsymbol {P} ^ {\\star}, O, \\xi_ {w o r s t})\\right)\\right)\\left. \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 618, + 482, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\xi_{\\mathrm{best}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ denotes the Gaussian noise that yields the best result under prompt $O$ , and let $\\xi_{\\mathrm{worst}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ yield the worst. Training samples were manually screened to identify the best and worst ones. We then sort the $N$ scores $\\{s^r (I_i)\\}_{i = 1}^N$ and denote the two highest as $s_{(1)}^r$ and $s_{(2)}^r$ . The acceptance probability for choosing between them by using HPS v2 is given by $\\mathbb{I}[s^{\\eta}(s_{(1)}^{r}),s^{\\eta}(s_{(2)}^{r})]$ where $m = 0$ as the filtering threshold and $\\mathbb{I}[\\cdot ]$ is the indicator function [38, 49]. The sample satisfying this condition is retained and used to build the target triplet $\\{C,I_{best},P\\}$ .", + "bbox": [ + 89, + 684, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Mask-Free Generation Learning", + "text_level": 1, + "bbox": [ + 89, + 848, + 377, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this stage, we leverage the triplet $\\{C, I_{best}, P\\}$ to generate the final virtual try-on result $P_{rec}$ , such that $P_{rec} \\approx P$ .", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ea0dfecd90daf9807ddf00161efca625430b47b38d6b516c52c10ab6285eaf49.jpg", + "image_caption": [ + "Figure 2. Visualization of pseudo-samples. Under the same prompt, the pseudo-samples are generated from five different Gaussian noises. When a pseudo-sample is semantically aligned with the prompt, the semantic number is labeled on the image. The leftmost column shows the clothing-agnostic person and the clothing-agnostic mask." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gated Manipulation. It remains challenging to organically fuse $C$ and $I_{best}$ so that the generator $G$ can both preserve rich human details and synthesize realistic clothing, while suppressing artifacts. Existing approaches can be roughly divided into two groups: (a) Treat $C$ and $I_{best}$ as separate conditioning signals to $G$ [22, 40]. (b) Crucely concatenate them in the pixel space before feeding the result to $G$ [14]. Approach (a) demands an additional Clothing-ReferenceNet to inject the new condition via attention modules; otherwise, directly adding a new branch may trigger catastrophic forgetting of the pre-trained weights, significantly increasing training cost. Approach (b) suffers from detail loss and blurring caused by the naive pixel-level combination. To achieve high-quality try-on without extra heavy modules or training overhead, we propose gated manipulation. Concretely, we first encode $C$ and $I_{best}$ into latent features, then blend them through a learnable gate $\\kappa \\in [0,1]$ in the feature space. The fused feature is", + "bbox": [ + 511, + 306, + 906, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {B} = \\kappa \\odot \\mathbf {x} _ {C} + (1 - \\kappa) \\odot \\mathbf {x} _ {I}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 590, + 903, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa = \\varPhi_{gate}(\\mathbf{x}_I,\\mathbf{x}_C)$ $\\odot$ denotes Hadamard product. $\\mathbf{x}_I = \\mathcal{E}_{VAE}(\\mathbf{I}_{best})$ $\\mathbf{x}_C = \\mathcal{E}_{VAE}(\\hat{C})$ $\\hat{C}$ is the warped $C$ $\\varPhi_{gate}(\\cdot)$ is a gate generator pre-trained by", + "bbox": [ + 511, + 619, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {g a t e} ^ {*} = \\arg \\min _ {\\Phi_ {g a t e}} \\ell_ {d i s t} (\\mathbf {x} _ {P}, \\mathbf {x} _ {B}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 679, + 903, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_P = \\mathcal{E}_{VAE}(P)$ . Since $\\kappa$ participates in encoding rich semantic and contextual cues, the blended feature $\\mathbf{x}_B$ obtained via Equation (9) effectively suppresses redundant clothing patterns from $I_{best}$ and background clutter from $C$ . At the same time, it preserves the holistic understanding of the entire $I_{best}$ (e.g., pose and identity), thereby mitigating catastrophic forgetting that typically occurs when the conditions for the new modality are fed to the generator.", + "bbox": [ + 511, + 714, + 905, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture-aware Injection. To enhance the generator's sensitivity to clothing texture, we adopt a bank of Gabor filters [11]—oriented at $0^{\\circ}$ , $45^{\\circ}$ , $90^{\\circ}$ , and $135^{\\circ}$ —to convolve", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19528", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the clothing $C$ . Each filter produces an orientation-specific response; the maximum across these four responses is taken at every spatial location, yielding a texture map $C_g$ that highlights dominant fabric patterns while suppressing noise. A domain-specific encoder $\\Psi(\\cdot)$ [32] then projects $C_g$ into an intermediate latent representation $\\Psi(C_g)$ , which is supplied to the cross-attention layer via learned key and value projections $K^{\\star} = W_k^{(i)}\\Psi(C_g), V^{\\star} = W_v^{(i)}\\Psi(C_g)$ . So, the objective of this stage is defined as Equation (11):", + "bbox": [ + 89, + 90, + 483, + 229 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m s e} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {I} _ {\\text {b e s t}}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 239, + 482, + 257 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\left\\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 258, + 449, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_P + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}.\\| \\cdot \\| _2^2$ denote the L2 norm. Considering that $\\hat{C}$ is usually coarsely warped, in the early stages of training, we considering directing $G$ 's focus towards clothing regions by preventing learning changes in other areas caused by $I_{best}$ 's potential defects. To this end, we conduct source-to-source self-reconstruction for training, as formulated in Equation (12):", + "bbox": [ + 89, + 287, + 483, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 405, + 482, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B} ^ {\\prime}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 422, + 447, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_B^\\prime = \\kappa \\odot \\mathbf{x}_C + (1 - \\kappa)\\odot \\mathbf{x}_P$", + "bbox": [ + 89, + 453, + 352, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Optimization", + "text_level": 1, + "bbox": [ + 89, + 477, + 230, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Full Objective. The full objective functions to optimize $\\epsilon_{\\theta}$ as Equation (13):", + "bbox": [ + 89, + 500, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathcal {L} _ {m s e} + \\lambda \\cdot \\mathcal {L} _ {r e c}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 539, + 482, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyper-parameter controlling relative importance and is set to 0.1 based on functionality.", + "bbox": [ + 89, + 569, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refining Conditional Inference. During inference, traditional methods [3, 37, 40] sample noise $\\mathbf{x}_T$ from a standard Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , unrelated to the target try-on. This introduces significant initial randomness and fails to preserve identity information and clothing details in the results [36]. Inspired by [36, 49], we design a refining conditional inference to enhance the model's performance.", + "bbox": [ + 89, + 617, + 483, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the latent representation $\\mathbf{x}_B$ (Eq. (9) during training) and sample a random Gaussian noise $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , we add $\\epsilon$ to $\\mathbf{x}_B$ at time step $T$ , as represented in Equation (14):", + "bbox": [ + 89, + 723, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {B} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 777, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To refine $\\hat{\\mathbf{x}}_T$ , we first decompose it using Singular Value Decomposition (SVD), which can be formulated in:", + "bbox": [ + 89, + 804, + 483, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU \\times \\Sigma \\times V ^ {\\mathrm {T}} = S V D (\\hat {\\mathbf {x}} _ {T}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 842, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Sigma$ is a diagonal matrix containing singular values, $U$ and $V^{\\mathrm{T}}$ represents the left and right singular vectors,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0149006772c10981675f98ce38fe391c27f1a82df41b30ffb633f8374bdb013a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 94, + 898, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81762b678792ecfa4d50d1d6d02728d53508e2c3d8e2c7212fcfd6d469f95faa.jpg", + "image_caption": [ + "Diffusion-noisy $\\hat{x}_T$" + ], + "image_footnote": [], + "bbox": [ + 524, + 162, + 898, + 205 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf3f4e3fc7cd087df2458647f526f8a51d8f26a9cab07d782b1867c3fcaa613f.jpg", + "image_caption": [ + "SVD denoised", + "Figure 3. Simulation comparison of denoising performance via SVD. From top to bottom: (row 1) clean signal $\\mathbf{x}_B$ , (row 2) diffusion-noisy signal $\\hat{\\mathbf{x}}_T$ , and (row 3) SVD-denoised reconstruction. SVD effectively suppresses the diffusion noise while preserving the dominant structural features of the original signal." + ], + "image_footnote": [], + "bbox": [ + 534, + 219, + 897, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively. As shown in Figure 3, SVD can distinguish principal components from perturbations, captures key components of $\\hat{\\mathbf{x}}_T$ via singular values. To enhance semantic information, we utilize a multi-layer perceptron (MLP) with multi-head self-attention layers [34] to predict a new singular value $\\widehat{\\Sigma}$ , as represented in Equation (16):", + "bbox": [ + 511, + 376, + 905, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {T} ^ {\\prime} = \\varphi (U, \\Sigma , V ^ {\\mathrm {T}}), \\quad \\widehat {\\Sigma} = f (\\phi (\\hat {\\mathbf {x}} _ {T} ^ {\\prime}), \\mathbf {x} _ {B}) + \\Sigma , \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 476, + 905, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi (\\cdot)$ denotes the sum of all MLP-encoded inputs, $\\phi (\\cdot)$ represents the multi-head self-attention layer, and $f(\\cdot)$ is another MLP. Thus, target representation $\\mathbf{x}_T$ is reformed by using $U$ , $\\widehat{\\Sigma}$ , $V^{\\mathrm{T}}$ to align with real representation $\\tilde{\\mathbf{x}}_{GT}$ , as formulated in Equation (17):", + "bbox": [ + 511, + 506, + 905, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {T} = U \\times \\widehat {\\boldsymbol {\\Sigma}} \\times V ^ {\\mathrm {T}}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 590, + 903, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, $\\tilde{\\pmb{x}}_{GT}$ can be obtained through Equation (18):", + "bbox": [ + 532, + 619, + 897, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {G T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {P} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 646, + 903, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining $\\tilde{\\mathbf{x}}_{GT}$ and $\\mathbf{x}_T$ , we optimize $\\{\\varphi ,\\phi ,f\\}$ by minimizing Equation (19):", + "bbox": [ + 511, + 674, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\varphi , \\phi , f} \\mathbb {E} _ {\\boldsymbol {C}, \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {P}} \\left[ \\| \\mathbf {x} _ {T} - \\tilde {\\boldsymbol {x}} _ {G T} \\| _ {2} ^ {2} \\right]. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 714, + 903, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 1, during inference, given a person $\\mathbf{x}_P$ and a random clothing $\\mathcal{E}_{VAE}(C_{un}) = \\mathbf{x}_{C_{un}}$ , $\\mathbf{x}_B$ is computed as: $\\mathbf{x}_B = \\kappa \\odot \\mathbf{x}_{C_{un}} + (1 - \\kappa) \\odot \\mathbf{x}_P$ .", + "bbox": [ + 511, + 750, + 906, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 808, + 645, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 511, + 833, + 594, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct experiments on the VITON-HD [4], VITON [16], and DressCode [29] datasets. VITON dataset comprises 16,253 image pairs with a resolution of", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19529", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$256 \\times 192$ , $14,221 / 2,032$ training/testing pairs. VITON-HD dataset is composed of 11,647/2,032 training/testing pairs with a high-resolution of $512 \\times 384$ . DressCode dataset is another high-resolution dataset with a resolution of $512 \\times 384$ , which comprises 15,363 image pairs, 12,863/2,500 training/testing pairs.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Setting. Our method is implemented with PyTorch on a single NVIDIA A100 GPU. Follow the configuration of [6, 14], Stable Diffusion $\\mathrm{V}1.5^{3}$ is used as our base model. We train for 20 epochs using the AdamW optimizer [27] with a base learning rate of $1e^{-5}$ and a batch size of 2, and use the warm-up adjustment strategy.", + "bbox": [ + 89, + 200, + 483, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 301, + 382, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We adopt Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [48], Fréchet Inception Distance (FID) [19], and Kernel Inception Distance (KID) [2], for quantitative evaluation. SSIM assesses pixel-level structural fidelity between generated and real images. LPIPS measures perceptual distance via deep features trained on human judgments; lower scores indicate higher perceptual similarity. FID computes the Wasserstein-2 distance between Inception-V3 feature distributions of real and generated images; lower is better. KID estimates distributional similarity through a maximum-mean-discrepancy kernel on the same Inception features; again, lower values denote higher fidelity.", + "bbox": [ + 89, + 324, + 483, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Methods. To conduct objective quantitative experiments, we employed the following methods:", + "bbox": [ + 89, + 540, + 482, + 570 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- GAN-based Methods: CP-VTON [35], Cloth-flow [17], CP-VTON+ [28], VTNFP [46], ACGPN [42], DCTON [12], PF-AFN [13], ZFlow [7], OVNet [24], LM-VTON [26], DAFlow [1], Style-Flow [18], RT-VTON [43], DressCode [29], VITON-HD [4], HR-VITON [23], POVNet [25], GP-VTON [39], and USC-PFN [8].", + "- Diffusion-based Methods: LaDI-VTON [30], PbE [41], DCI-VTON [14], StableVITON [22], OOTDiffusion [40], Anydoor [3], TPD [44], IDM-VTON [5], StableGarment [37], and LDE-VTON [10]." + ], + "bbox": [ + 89, + 570, + 482, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. Figure 4 visualizes our results on the VITON-HD dataset [4]. GAN-based approaches—GP-VTON [39] and USC-PFN [8], for example—lack the human-centric priors inherent to diffusion models. Consequently, their outputs are overly smooth, contain visible artifacts, and struggle to disambiguate crossed arms. In contrast, our method synthesizes more realistic garment wrinkles and correctly reconstructs cross-arm poses. Diffusion model-based methods, such as DCI-", + "bbox": [ + 89, + 739, + 482, + 878 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/31cd9159376fc1e90ee8b076a89799c7ecc3c5c8632734335020d4751bcba67f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsVenueMask\\( SSIM_p \\)↑\\( FID_{up} \\)↓
VITONCVPR'18✓0.7455.71
CP-VTON [35]ECCV'18✓0.7224.45
VTNFP [46]ICCV'19✓0.80n/a
Cloth-flow [17]CVPR'19✓0.8414.43
CP-VTON+ [28]CVPRW'20✓0.7521.04
ACGPN [42]CVPR'20✓0.8416.64
LM-VTON [26]AAAI'21✓0.8517.18
DCTON [12]CVPR'21✓0.8314.82
ZFlow [7]ICCV'21✓0.8815.17
OVNet [24]CVPR'21✓0.8515.78
PF-AFN [13]CVPR'21✘0.8910.21
RT-VTON [43]CVPR'22✓n/a11.66
DAFlow [1]ECCV'22✓0.8812.05
DressCode [29]CVPR'22✓0.8913.71
POVNet [25]TPAMI'23✓0.8913.37
PbE [41]CVPR'23✓0.8312.56
USC-PFN [8]NeurIPS'23✘0.9110.47
TPD [44]CVPR'24✓0.899.58
LDE-VTON [10]AAAI'25✘0.919.86
OursThis Work✘0.919.52
", + "bbox": [ + 519, + 88, + 903, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on the VITON dataset. \"Mask\" denotes whether the mask is used during inference. Bold denotes the best result. Some data are from published literature.", + "bbox": [ + 511, + 382, + 906, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VTON [14], StableVITON [22], StableGarment [37], LDEVTON [10], and OOTDiffusion [40], limited by the diversity of training samples, learn fixed sample transformations. Consequently, they often fail to generate natural and realistic arms when dealing with unusual poses and clothing (e.g., the $2^{nd}$ row). Additionally, since the randomness of Gaussian noise during inference is not addressed, the results often lose identity information (e.g., the belt in the $1^{st}$ row) and clothing details (e.g., garment shape and texture in the $3^{rd}$ row). In contrast, our approach addresses these issues by (1) augmenting training sample diversity to improve robustness, (2) enhancing the model architecture for higher synthesis quality, and (3) optimizing the handling of Gaussian noise to preserve identity and garment details.", + "bbox": [ + 511, + 452, + 908, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. Table 1 and Table 2 present quantitative comparisons between our method and the baselines on the VITON [16], VITON-HD [4], and DressCode [29] datasets. Our approach attains the highest scores on VITON dataset, demonstrating its robustness to low-resolution inputs. On VITON-HD and DressCode datasets, it ranks either first or second. The occasional second place arises because most baselines are trained on a fixed set of clothing images, leading to overfitting between specific garments and reference identities. In contrast, we enlarge appearance diversity by synthesizing additional pseudosamples and introducing out-of-distribution information; this strategy slightly penalizes performance on test cases that closely match the original data distribution.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://huggingface.co/stable-diffusion-v1-5", + "bbox": [ + 107, + 886, + 467, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19530", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/268735c1a8ca0f803de09a0358ce7cd5da65f7dfe71966478f3b8537e696523c.jpg", + "image_caption": [ + "Figure 4. Qualitative results with different baseline methods on the VITON-HD dataset. The left area shows a visual comparison with GAN-based baseline methods, while the right area shows a comparison with diffusion model-based baseline methods. It can be seen that our method achieves the best visual performance. Red dashed box represents defects." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 903, + 342 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0dda832b872b5e34fd3e5557d639e367fe77e81541f6cc4d4f4ce6a491c09b7c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train / Test MethodsVenueVITON-HDDressCode UpperMask
\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)
VITON-HD [4]CVPR'210.8620.11712.1173.23n/an/an/an/a✓
HR-VITON [23]ECCV'220.8780.10511.2652.730.9360.06513.822.71✓
GP-VTON [39]CVPR'230.8840.0819.7011.260.7690.27020.118.17✓
LaDI-VTON [30]MM'230.8640.0969.4801.990.9150.06314.263.33✓
PbE [41]CVPR'230.8020.14311.9393.850.8970.07815.334.64✓
DCI-VTON [14]MM'230.8800.0809.4101.100.9370.04211.921.89✓
StableVITON [22]CVPR'240.8640.0849.4651.400.9110.05011.270.72✓
StableGarment [37]arXiv'240.8030.10417.1158.85n/an/an/an/a✓
IDM-VTON [5]ECCV'240.8500.0609.8421.120.8800.0569.544.32✓
AnyDoor [3]CVPR'240.8210.09910.8502.460.8990.11914.833.05✓
LDE-VTON [10]AAAI'250.8980.0819.6401.21n/an/an/an/a✓
OursThis Work0.9010.0799.3841.120.9390.04810.951.78✘
", + "bbox": [ + 96, + 407, + 901, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparisons on the VITON-HD and DressCode datasets. For LPIPS, FID, and KID, the lower the better. For SSIM, the higher the better. Underline represents second best. Some data are from published literature.", + "bbox": [ + 89, + 638, + 906, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 691, + 243, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Text-Driven Pseudo-Input Preparation.", + "text_level": 1, + "bbox": [ + 89, + 714, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 (A) and Table 3 (A) highlight the necessity of text-driven pseudo-input preparation (TDPIP). By increasing the appearance diversity of pseudo-samples, it enhances the generator's generalization ability for abnormal samples. Without this component, the arm regions of abnormal samples are poorly generated, and the clothing is not properly fitted. In contrast, with this preparation, the arms are well generated, and the clothing is effectively transferred.", + "bbox": [ + 89, + 729, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Gated Manipulation. Figure 6 demonstrates the necessity of introducing gated manipulation", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(GM) to prevent catastrophic forgetting. It can be seen that when we remove GM from the baseline method, the generator with altered input structure requires nearly 75 epochs to converge. In contrast, when GM is introduced, the model achieves good convergence after only 10 epochs of training.", + "bbox": [ + 511, + 693, + 906, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Texture-aware Injection. Figure 5 (B) and Table 3 (B) demonstrate the necessity of introducing texture-aware injection (TAI) method for capturing clothing texture details. It can be seen that when TAI is introduced into the baseline generator, the text on the target clothing image is faithfully reconstructed. In contrast, when it is removed, some text details are lost by the generator.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19531", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f9ef0e66ab130b2a0fccf4a6dc30004bcb0d0d4d0a0a9d3ae0af489fdc09ee2.jpg", + "image_caption": [ + "Figure 5. Visual ablation study of different components in our approach. (A) Text-Driven Pseudo-Input Preparation (TDPIP). (B) Texture-aware Injection (TAI). (C) Refining Conditional Inference (RCI). The blue dashed box represents the area of focus for comparison." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 906, + 241 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8a73c714aa450c95dd65cf8d8fa12a4b73456f8c445bd1e4ed8af9c402ac7636.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ConfigurationSSIMp↑FIDup↓
Baseline Framework0.8312.56
+ (A) Text-Driven Pseudo-Input0.879.86
+ (B) Texture-aware Injection0.899.61
+ (C) Refining Conditional Inference0.919.52
", + "bbox": [ + 91, + 303, + 486, + 397 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study. Performance of various configurations on VITON dataset. Baseline is a basic teacher-student architecture with PbE [41]. Bold denotes the best result.", + "bbox": [ + 89, + 406, + 483, + 449 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Refining Conditional Inference. Finally, Figure 5 (C) and Table 3 (C) demonstrate the necessity of the proposed refining conditional inference (RCI) for information preservation and reducing generation randomness. When this initial noise generation mode is introduced into the framework, the clothing details in the try-on results are well reconstructed, and there is no loss of identity information. However, when this initial noise generation mode is removed, some clothing details are lost to varying degrees.", + "bbox": [ + 88, + 476, + 482, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Limitations", + "text_level": 1, + "bbox": [ + 89, + 625, + 217, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our proposed method can generate high-quality try-on results within a mask-free framework, the target noise optimized by the proposed refining conditional inference approach is estimated from only a limited set of samples. This scarcity can introduce instability and, at the beginning of inference, may perturb the initial latent distribution. In future work, we will investigate more robust and data-driven adversarial augmentation strategies for target-noise generation to eliminate this limitation.", + "bbox": [ + 89, + 647, + 482, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 799, + 209, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a unified, mask-free virtual try-on framework that boosts generation quality by increasing training sample diversity and improving both training and inference processes. First, we introduce a text-driven pseudo-input preparation strategy to enlarge data", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f43ec90fa996478dbaeaa1b2c2ef42e31198cc77722edd2783e4c2d1852067e.jpg", + "image_caption": [ + "Figure 6. Ablation study of the Gated Manipulation. It can be seen that with the use of the gated manipulation, the generator quickly converges to a better state in the early stages of training." + ], + "image_footnote": [], + "bbox": [ + 516, + 305, + 714, + 439 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6cfc4e440a7ae9cc46db2ac5e8daa2573bf4561a0a840179845bad30a591812.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 306, + 900, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "diversity and enhance model's generalization. Second, a gated-manipulation mechanism prevents catastrophic forgetting of pre-trained model weights while cutting training costs. Third, texture-aware injection explicitly supplements human-perceptible garment details. Finally, a refining conditional-inference scheme suppresses Gaussian-noise randomness to preserve identity and fine-grained clothing cues during inference. Extensive experiments on multiple benchmarks show that our proposed method consistently outperforms previous virtual try-on work.", + "bbox": [ + 511, + 521, + 906, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was in part supported by the National Key Research and Development Program of China (Grant No. 2022ZD0160604), the National Natural Science Foundation of China (NSFC, Grant No. 62176194), the Key Research and Development Program of Hubei Province (Grant No. 2023BAB083), the Project of Sanya Yazhou Bay Science and Technology City (Grant Nos. SCKJ-JYRC-2022-76, SKJC-2022-PTDX-031), the Project of Sanya Science and Education Innovation Park of Wuhan University of Technology (Grant No. 2021KF0031), and the Huawei Kunpeng-Ascend Innovation Incentive Programme. This work was also supported in part by both computing resources from the Wuhan Supercomputing Center and the Wuhan Artificial Intelligence Computing Center.", + "bbox": [ + 511, + 691, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19532", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 409-425. Springer, 2022. 6", + "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6", + "[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 1, 5, 6, 7", + "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 5, 6, 7", + "[5] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild. In European Conference on Computer Vision, pages 206-235. Springer, 2024. 6, 7", + "[6] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, Dongmei Jiang, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 6", + "[7] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5433-5442, 2021. 6", + "[8] Chenghu Du, Shuqing Liu, Shengwu Xiong, et al. Greatness in simplicity: Unified self-cycle consistency for parser-free virtual try-on. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6", + "[9] Chenghu Du, Junyin Wang, Kai Liu, and Shengwu Xiong. Glv: Geometric correlation distillation for latent diffusion-enhanced parser-free virtual try-on. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2", + "[10] Chenghu Du, Junyin Wang, Feng Yu, and Shengwu Xiong. Latent diffusion-enhanced virtual try-on via optimized pseudo-label generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2780-2788, 2025. 6, 7", + "[11] D. Gabor. Theory of communication. part 1: The analysis of information. Journal of the Institution of Electrical Engineers - Part III: Radio and Communication Engineering, 93: 429-441, 1946. 4", + "[12] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 6", + "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1, 2, 3, 6", + "[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 1, 4, 6, 7", + "[15] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4", + "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 5, 6", + "[17] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 6", + "[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1, 2, 3, 6", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 1", + "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 1, 2", + "[22] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 1, 4, 6, 7", + "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 204-219. Springer, 2022. 6, 7", + "[24] Kedan Li, Min Jin Chong, Jeffrey Zhang, and Jingen Liu. Toward accurate and realistic outfits visualization with attention to details. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15546-15555, 2021. 6", + "[25] Kedan Li, Jeffrey Zhang, and David Forsyth. Povnet: Image-based virtual try-on through accurate warping and residual." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19533", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12222-12235, 2023. 6", + "[26] Guoqiang Liu, Dan Song, Ruofeng Tong, and Min Tang. Toward realistic virtual try-on through landmark guided shape matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2118-2126, 2021. 6", + "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[28] Matiur Rahman Minar, Thai Thanh Tuan, Heejune Ahn, Paul Rosin, and Yu-Kun Lai. Cp-vton+: Clothing shape and texture preserving image-based virtual try-on. In CVPR Workshops, pages 10-14, 2020. 6", + "[29] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 5, 6", + "[30] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 6, 7", + "[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[35] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 6", + "[36] Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, and Hongming Shan. Fldm-vton: Faithful latent diffusion model for virtual try-on. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 1362–1370. International Joint Conferences on Artificial Intelligence Organization, 2024. 1, 5", + "[37] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783, 2024. 1, 3, 5, 6, 7" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 4", + "[39] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23550-23559, 2023. 1, 6, 7", + "[40] Yuhao Xu, Tao Gu, Weifeng Chen, and Arlene Chen. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8996-9004, 2025. 1, 4, 5, 6", + "[41] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 1, 6, 7, 8", + "[42] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 6", + "[43] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3460-3469, 2022. 6", + "[44] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 6", + "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2", + "[46] Ruiyun Yu, Xiaqi Wang, and Xiaohui Xie. Vtnfp: An image-based virtual try-on network with body and clothing feature preservation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10511-10520, 2019. 6", + "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[49]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.4,5" + ], + "bbox": [ + 516, + 92, + 906, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19534", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_model.json b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6ebd9b0c7e58747ed12fafd867e3cc7e2ccfc316 --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.842, + 0.154 + ], + "angle": 0, + "content": "All Parts Matter: A Unified Mask-Free Virtual Try-On Framework" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.18, + 0.856, + 0.252 + ], + "angle": 0, + "content": "Chenghu Du\\(^{1,3}\\) Shengwu Xiong\\(^{2*}\\) Yi Rong\\(^{1,3*}\\) \n\\(^{1}\\)School of Computer Science and Artificial Intelligence, Wuhan University of Technology \n\\(^{2}\\)Interdisciplinary Artificial Intelligence Research Institute, Wuhan College \n\\(^{3}\\)Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.255, + 0.645, + 0.269 + ], + "angle": 0, + "content": "{duch, xiongsw, wrong}@whut.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.76 + ], + "angle": 0, + "content": "Current virtual try-on methods primarily enhance performance through network optimization, employing strategies such as coarse-to-fine structures and ReferenceNet to inject clothing information. However, these methods are often constrained by the limited quantity and diversity of training samples, which ultimately restricts their potential for further improvement. To address this challenge, we propose a unified, mask-free virtual try-on framework. Our approach leverages the inherent strengths of latent diffusion models to enhance the capability of each pipeline component to accurately model the target distribution, thereby achieving superior performance. Specifically, we introduce a text-driven pseudo-input preparation strategy that significantly increases the diversity of clothing regions within the generated person pseudo-samples. This encourages the generator to focus on variations in these areas and improves the model's generalization capability. Within the generator itself, we develop a gated manipulation mechanism to prevent weight forgetting and reduce training costs. Furthermore, we incorporate a texture-aware injection module to explicitly integrate human-perceptible clothing texture information into the generation process. During inference, we propose a refining conditional inference strategy that mitigates the randomness introduced by Gaussian noise. This effectively preserves identity information and fine clothing details in the final results. Extensive experiments demonstrate that our method outperforms existing state-of-the-art virtual try-on approaches." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.79, + 0.222, + 0.805 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Image-based virtual try-on (VTON) aims to realistically transfer a target garment onto the corresponding region of a reference person [3, 8, 14, 22, 39]. Owing to its strong application prospects and commercial potential, VTON has" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.366 + ], + "angle": 0, + "content": "attracted growing research interest. Leveraging the powerful generative capabilities of diffusion models [20, 41], recent VTON methods have significantly narrowed the gap to real-world fitting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Previous VTON methods are broadly categorized into mask-based and mask-free paradigms. Mask-based approaches [3, 14, 22, 37] typically operate in a self-supervised fashion: a clothing-agnostic mask (Figure 2) is first applied to remove the garment region and adjacent body parts from the reference image, and the generator is then trained to inpaint the masked areas, thereby learning to perform the virtual try-on task. Although effective and widely adopted, these methods rely heavily on obtaining accurate masks for each reference person; even minor mask errors can introduce conspicuous artifacts [13, 21]. To mitigate mask-induced artifacts, mask-free methods grounded in pseudo-sample [8, 13, 18] have been proposed. A mask-based VTON model is first pre-trained to synthesize try-on images for arbitrary garments on the reference person, yielding paired pseudo-samples. These pseudo-samples are then fed as inputs, while the original reference image serves as the reconstruction target (ground truth), to train a main generator. Although this pipeline obviates the need for precise masks, they are limited to a single, fixed transformation for each pseudo-sample/reference-person pair. Moreover, because both garments and reference images are typically drawn from the same dataset, the generator overfits to these specific items and generalizes poorly to unseen clothing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.869 + ], + "angle": 0, + "content": "There are also some other methods attempt to boost performance through architectural improvements—such as coarse-to-fine networks and ReferenceNet-based clothing-feature injection [22, 37, 40]. Yet these refinements remain constrained by the input materials (human parsing, etc), causing the generator to converge to sub-optimal solutions. Finally, the randomness of Gaussian-noise sampling during inference further degrades output quality [36]—a problem often overlooked by prior work." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we propose a unified mask-free virtual try-on framework that leverages the intrinsic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19525" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.302 + ], + "angle": 0, + "content": "strengths of latent diffusion models to enable every pipeline component to better approximate the target data distribution and thus improve overall performance. Specifically, we introduce a text-driven pseudo-input preparation strategy to augment sample diversity by enriching the clothing regions in each pseudo-sample (see Figure 2). First, a pretrained Stable Diffusion inpainting model is conditioned on text prompts generated by a large language model (e.g., DeepSeek or GPT-series models) to produce multiple try-on results for the same identity. We then perform sample selection to filter out implausible images and retain the most realistic candidate as the final pseudo-sample. This process maximizes the variability of clothing regions and mitigates overfitting of the generator during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.483 + ], + "angle": 0, + "content": "During training phase, we leverage the off-the-shelf inpainting try-on model as the generator, thereby reducing training costs. To convert this model into a mask-free generator, we introduce mask-free generation learning process. Concretely, a gated manipulation mechanism fuses clothing and person latent representations in latent space, preventing catastrophic forgetting when clothing conditioning is added and avoiding structural changes that inflate training cost. Complementarily, texture-aware injection module explicitly injects fine-grained texture cues into the cross-attention layers, sharpening the generator's perception of garment details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.482, + 0.576 + ], + "angle": 0, + "content": "During inference phase, to reduce the randomness of Gaussian noise, we propose a refining conditional inference in our optimization process. By optimizing the initial Gaussian noise to make it as close as possible to the distribution of real person samples, helping the generator improve the quality of the synthesized results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.482, + 0.603 + ], + "angle": 0, + "content": "In summary, this paper makes the following main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.605, + 0.481, + 0.65 + ], + "angle": 0, + "content": "- We propose a unified mask-free virtual try-on framework, which improves the generator's performance from multiple aspects of both the training and inference processes." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.651, + 0.481, + 0.68 + ], + "angle": 0, + "content": "- We present a text-driven pseudo-input preparation approach to improve the generator's generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.681, + 0.481, + 0.71 + ], + "angle": 0, + "content": "- We propose a gated manipulation to prevent weight forgetting and cut training costs." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.711, + 0.481, + 0.74 + ], + "angle": 0, + "content": "- We introduce a texture-aware injection to explicitly add human-perceptible clothing texture information." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.741, + 0.481, + 0.77 + ], + "angle": 0, + "content": "- We propose a refining conditional inference approach to reduce Gaussian noise randomness." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.605, + 0.481, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.232, + 0.8 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Mask-Free Virtual Try-On. This task requires performing virtual try-on at inference time using only a clothing image and a person image—no masks, human parsing, or other auxiliary inputs. Issenhuth et al. [21] introduced the first mask-free method, WUTON, which attains fully supervised training by distilling knowledge from a mask-based teacher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "that generates paired training samples. Because the synthetic images are treated as ground truth for the student, the teacher's defects directly bound the student's performance. Ge et al. [13] relaxed this constraint with PFAFN, feeding pseudo-samples as inputs instead of supervision, thereby allowing the student to learn a more accurate mapping. He et al. [18] extended this paradigm in Flow-Style, adding explicit garment-deformation optimization to improve fit and realism. More recently, Du et al. [8] proposed USC-PFN, a two-stage architecture that employs cycle-consistency and a weight-shared generator to refine the garment region and try-on result iteratively. In contrast to these prior works, we present a unified mask-free framework that optimizes the pseudo-sample and enhances both training and inference through multi-stage optimization, yielding higher fidelity and better generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.624 + ], + "angle": 0, + "content": "Controllable Diffusion Models. Recent advances in diffusion models demonstrate that generation can be guided by modalities beyond text. Specifically, the integration of external models to inject external features into diffusion models has produced remarkable results. For example, ControlNet [47] introduced an external model that leverages images, skeletons, and edge maps as conditional inputs to generate corresponding outputs. IP-adaptor [45] proposed a method for generating images similar to a given reference image. T2I-adapter [31] also introduced an external, more lightweight model that leverages conditional inputs to generate corresponding images. These approaches have effectively manipulated image generation by injecting feature values via external models. Instead of adding extra network branches, following the optimization idea of latent features in latent space of [9], we embed the try-on conditions directly into the latent space, yielding a high-performance virtual try-on model with minimal overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.639, + 0.855, + 0.656 + ], + "angle": 0, + "content": "3. Preliminary: Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.905, + 0.771 + ], + "angle": 0, + "content": "Latent Diffusion Models (LDMs) [33] are a class of generative models that learn to synthesize data \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\) by denoising latent representations \\(\\mathbf{z} \\in \\mathbb{R}^{h \\times w \\times c}\\) in a compressed space, where typically \\(h \\ll H\\) and \\(w \\ll W\\). The compression is achieved via a pre-trained autoencoder with KL-regularization (e.g. VQGAN), consisting of an encoder \\(\\mathcal{E}_{VAE}(\\cdot)\\) and decoder \\(\\mathcal{D}_{VAE}(\\cdot)\\), such that" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.783, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {0} = \\mathcal {E} _ {V A E} (\\mathbf {x}), \\quad \\hat {\\mathbf {x}} = \\mathcal {D} _ {V A E} (\\mathbf {z} _ {0}), \\quad \\text {w i t h} \\quad \\hat {\\mathbf {x}} \\approx \\mathbf {x}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.872 + ], + "angle": 0, + "content": "The latent-space diffusion process is modeled as a Markov chain that progressively adds Gaussian noise to \\(\\mathbf{z}_0\\) (the encoded data) over \\(T\\) steps, producing a sequence of noisy latents \\(\\{\\mathbf{z}_t\\}_{t=0}^T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.884, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.418, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.224, + 0.419, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.226, + 0.904, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.293, + 0.41, + 0.305 + ], + "angle": 0, + "content": "Sect. 4.1: Text-Driven Pseudo-Input Preparation" + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.293, + 0.736, + 0.304 + ], + "angle": 0, + "content": "Sect. 4.3: Optimization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.907, + 0.428 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed virtual try-on framework. It consists of three sub-pipelines, namely Text-Driven Pseudo-Input Preparation, Mask-Free Generation Learning, and Optimization. In text-driven pseudo-input preparation (Sect. 4.1), we first use a pre-trained Stable Diffusion inpainting model to generate \\(N\\) different pseudo-samples \\(\\{I_i\\}_{i=0}^N\\) based on prompts generated by DeepSeek and various noises. Then, a sample selection method is proposed to pick the best one \\(I_{best}\\) from the \\(N\\) pseudo-samples for the next step. In mask-free generation learning (Sect. 4.2), we design a mask-free training approach that initially uses gated manipulation to fuse clothing and person representations to prevent catastrophic forgetting, and then injects extracted clothing texture features through cross-attention modules. In optimization (Sect. 4.3), we first train a noise optimization MLP to obtain the optimized Gaussian noise that preserves as much identity and fine-grained clothing cues as possible, and then use it for the inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.484 + ], + "angle": 0, + "content": "where \\(\\{\\beta_t\\in (0,1)\\}_{t = 1}^T\\) is a fixed variance schedule. The forward process can be marginalized in closed form as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.495, + 0.483, + 0.513 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.484, + 0.585 + ], + "angle": 0, + "content": "with \\(\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}\\). LDMs optimize a variational lower bound (VLB) on the log-likelihood of the data, simplified to the noise-prediction objective (DDPM-style):" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.595, + 0.482, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {z} _ {0}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.66 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) is the ground-truth noise and \\(\\epsilon_{\\theta}(\\cdot)\\) is a U-Net parameterized by \\(\\theta\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.226, + 0.691 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Overview. This task aims to transfer the target clothing from an arbitrary clothing image \\(C_{un} \\in \\mathbb{R}^{H \\times W \\times 3}\\) onto a reference person image \\(P \\in \\mathbb{R}^{H \\times W \\times 3}\\), producing a try-on result \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\). To train the mask-free model, \\(I\\) (generated from the off-the-shelf mask-based try-on method) is first used as pseudo-input, we construct triplets \\(\\{C, I, P\\}\\) for fully supervised learning of the generator \\(G\\) with parameters \\(\\theta\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.828, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {\\boldsymbol {C}, \\boldsymbol {I}, \\boldsymbol {P}} \\ell_ {\\text {d i s t}} (G (\\boldsymbol {C}, \\boldsymbol {I}), \\boldsymbol {P}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\ell_{dist}\\) represents a suitable loss function (e.g., L1, MSE) between the generated image \\(G(C,I)\\) and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.576 + ], + "angle": 0, + "content": "ground-truth image \\( \\pmb{P} \\). The clothing image \\( C \\in \\mathbb{R}^{H \\times W \\times 3} \\) and the clothes worn by \\( \\pmb{P} \\) are identical and also sourced from the dataset. Sections that follow detail how we obtain a suitable pseudo-input \\( \\pmb{I} \\) for efficient training of the generator \\( G \\). An overview of our pipeline—comprising Text-Driven Pseudo-Input Preparation (Sect. 4.1), Mask-Free Generation Learning (Sect. 4.2), and Refining Conditional Inference (Sect. 4.3)—is provided in Figure 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.847, + 0.601 + ], + "angle": 0, + "content": "4.1. Text-Driven Pseudo-Input Preparation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Previous mask-free methods [13, 18] rely on off-the-shelf try-on networks to generate the pseudo-input. Although training converges, the rigid one-to-one pairing of \\( C_{un} / I \\) and \\( P \\) drives the generator \\( G \\) toward sub-optimal solutions and limits generalization. Therefore, diversifying the clothing regions of \\( I \\) (see Figure 2) is crucial for improving model performance in mask-free try-ons. By increasing clothing variability, the generator \\( G \\) can focus more effectively on clothing features and details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Automatic Sample Generation. To increase the visual diversity of \\( \\pmb{I} \\) and mitigate the generalization drop caused by the overlap between \\( C_{un} \\) (for generating \\( \\pmb{I} \\)) and \\( \\pmb{C} \\), we leverage a Stable Diffusion inpainting model [37], \\( \\mathcal{F}_{\\mathrm{SD}} \\), which possesses strong generative priors, together with the large language model DeepSeek2. The resulting pipeline is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.888 + ], + "angle": 0, + "content": "1https://civitai.com/models/120149/controlnet-for-densepose" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.888, + 0.738, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://chat_deepseek.com" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19527" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.283, + 0.106 + ], + "angle": 0, + "content": "summarized in Equation (6)." + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.116, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {I} _ {i} \\right\\} _ {i = 0} ^ {N} = \\left\\{\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {i}\\right) \\right\\} _ {i = 0} ^ {N}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.145, + 0.485, + 0.298 + ], + "angle": 0, + "content": "where \\( P^{\\star} \\) denotes the clothing-agnostic person representation, consisting of the clothing-agnostic person image, the clothing-agnostic mask, and the DensePose map [15]. \\( \\xi \\) is a random Gaussian noise sample, \\( \\xi \\sim \\mathcal{N}(0,\\mathbf{I}) \\). \\( O \\) is a text prompt constructed from the template \"A photo of a person wearing {upper garment description}\", with the description generated by DeepSeek to cover a wide range of clothing styles. For each prompt, we synthesize \\( N = 5 \\) images and select the best try-on result \\( I_{best} \\). The resulting triplets \\( \\{C,I_{best},P\\} \\) are used for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.435 + ], + "angle": 0, + "content": "Sample Selection. To this end, we adopt the Human Preference Score \\(\\nu 2\\) (HPS v2) model [38] to rank the generated images \\(\\{I_i\\}_{i=1}^N\\) according to their alignment with the corresponding prompt. HPS v2 employs an image encoder \\(E_{IMG}(\\cdot)\\) and a text encoder \\(E_{TEXT}(\\cdot)\\) to map images and prompts into a joint embedding space, where their cosine similarity serves as the preference score, as formulated in Equation (7)." + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.446, + 0.483, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} ^ {\\eta} (\\boldsymbol {I} _ {i}, O) = E _ {I M G} (\\boldsymbol {I} _ {i}) \\cdot E _ {T E X T} (O) \\times 1 0 0. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.61 + ], + "angle": 0, + "content": "We select the image achieving the highest score as the final sample. However, some \\(I_{i}\\) while semantically consistent with the prompt (see column 5 of Figure 2)—exhibit implausible appearances in the inpainting regions, such as distorted limbs. To address this issue, before using HPS v2, we introduce an appearance rationality discriminator \\(\\mathcal{D}_r\\) that outputs a rationality score \\(s^r(I_i) = \\mathcal{D}_r(I_i)\\) for each candidate image. This discriminator is trained by minimizing the objective given in Equation (8)." + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.619, + 0.483, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\mathcal {D} _ {r}} - \\mathbb {E} _ {\\boldsymbol {P}, O} \\left[ \\log \\left(\\mathcal {D} _ {r} \\left(\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {\\text {b e s t}}\\right)\\right)\\right) \\right. \\tag {8} \\\\ \\left. + \\log \\left(1 - \\mathcal {D} _ {r} \\big (\\mathcal {F} _ {S D} (\\boldsymbol {P} ^ {\\star}, O, \\xi_ {w o r s t})\\right)\\right)\\left. \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.841 + ], + "angle": 0, + "content": "where \\(\\xi_{\\mathrm{best}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) denotes the Gaussian noise that yields the best result under prompt \\(O\\), and let \\(\\xi_{\\mathrm{worst}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) yield the worst. Training samples were manually screened to identify the best and worst ones. We then sort the \\(N\\) scores \\(\\{s^r (I_i)\\}_{i = 1}^N\\) and denote the two highest as \\(s_{(1)}^r\\) and \\(s_{(2)}^r\\). The acceptance probability for choosing between them by using HPS v2 is given by \\(\\mathbb{I}[s^{\\eta}(s_{(1)}^{r}),s^{\\eta}(s_{(2)}^{r})]\\) where \\(m = 0\\) as the filtering threshold and \\(\\mathbb{I}[\\cdot ]\\) is the indicator function [38, 49]. The sample satisfying this condition is retained and used to build the target triplet \\(\\{C,I_{best},P\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.379, + 0.866 + ], + "angle": 0, + "content": "4.2. Mask-Free Generation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this stage, we leverage the triplet \\(\\{C, I_{best}, P\\}\\) to generate the final virtual try-on result \\(P_{rec}\\), such that \\(P_{rec} \\approx P\\)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.28 + ], + "angle": 0, + "content": "Figure 2. Visualization of pseudo-samples. Under the same prompt, the pseudo-samples are generated from five different Gaussian noises. When a pseudo-sample is semantically aligned with the prompt, the semantic number is labeled on the image. The leftmost column shows the clothing-agnostic person and the clothing-agnostic mask." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Gated Manipulation. It remains challenging to organically fuse \\( C \\) and \\( I_{best} \\) so that the generator \\( G \\) can both preserve rich human details and synthesize realistic clothing, while suppressing artifacts. Existing approaches can be roughly divided into two groups: (a) Treat \\( C \\) and \\( I_{best} \\) as separate conditioning signals to \\( G \\) [22, 40]. (b) Crucely concatenate them in the pixel space before feeding the result to \\( G \\) [14]. Approach (a) demands an additional Clothing-ReferenceNet to inject the new condition via attention modules; otherwise, directly adding a new branch may trigger catastrophic forgetting of the pre-trained weights, significantly increasing training cost. Approach (b) suffers from detail loss and blurring caused by the naive pixel-level combination. To achieve high-quality try-on without extra heavy modules or training overhead, we propose gated manipulation. Concretely, we first encode \\( C \\) and \\( I_{best} \\) into latent features, then blend them through a learnable gate \\( \\kappa \\in [0,1] \\) in the feature space. The fused feature is" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.592, + 0.905, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {B} = \\kappa \\odot \\mathbf {x} _ {C} + (1 - \\kappa) \\odot \\mathbf {x} _ {I}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.62, + 0.906, + 0.669 + ], + "angle": 0, + "content": "where \\(\\kappa = \\varPhi_{gate}(\\mathbf{x}_I,\\mathbf{x}_C)\\) \\(\\odot\\) denotes Hadamard product. \\(\\mathbf{x}_I = \\mathcal{E}_{VAE}(\\mathbf{I}_{best})\\) \\(\\mathbf{x}_C = \\mathcal{E}_{VAE}(\\hat{C})\\) \\(\\hat{C}\\) is the warped \\(C\\) \\(\\varPhi_{gate}(\\cdot)\\) is a gate generator pre-trained by" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.68, + 0.905, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {g a t e} ^ {*} = \\arg \\min _ {\\Phi_ {g a t e}} \\ell_ {d i s t} (\\mathbf {x} _ {P}, \\mathbf {x} _ {B}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.906, + 0.837 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_P = \\mathcal{E}_{VAE}(P)\\). Since \\(\\kappa\\) participates in encoding rich semantic and contextual cues, the blended feature \\(\\mathbf{x}_B\\) obtained via Equation (9) effectively suppresses redundant clothing patterns from \\(I_{best}\\) and background clutter from \\(C\\). At the same time, it preserves the holistic understanding of the entire \\(I_{best}\\) (e.g., pose and identity), thereby mitigating catastrophic forgetting that typically occurs when the conditions for the new modality are fed to the generator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Texture-aware Injection. To enhance the generator's sensitivity to clothing texture, we adopt a bank of Gabor filters [11]—oriented at \\(0^{\\circ}\\), \\(45^{\\circ}\\), \\(90^{\\circ}\\), and \\(135^{\\circ}\\)—to convolve" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19528" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.231 + ], + "angle": 0, + "content": "the clothing \\(C\\). Each filter produces an orientation-specific response; the maximum across these four responses is taken at every spatial location, yielding a texture map \\(C_g\\) that highlights dominant fabric patterns while suppressing noise. A domain-specific encoder \\(\\Psi(\\cdot)\\) [32] then projects \\(C_g\\) into an intermediate latent representation \\(\\Psi(C_g)\\), which is supplied to the cross-attention layer via learned key and value projections \\(K^{\\star} = W_k^{(i)}\\Psi(C_g), V^{\\star} = W_v^{(i)}\\Psi(C_g)\\). So, the objective of this stage is defined as Equation (11):" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.24, + 0.483, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m s e} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {I} _ {\\text {b e s t}}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.234, + 0.259, + 0.45, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\left[ \\left\\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.397 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_P + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}.\\| \\cdot \\| _2^2\\) denote the L2 norm. Considering that \\(\\hat{C}\\) is usually coarsely warped, in the early stages of training, we considering directing \\(G\\) 's focus towards clothing regions by preventing learning changes in other areas caused by \\(I_{best}\\)'s potential defects. To this end, we conduct source-to-source self-reconstruction for training, as formulated in Equation (12):" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.406, + 0.483, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.424, + 0.449, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B} ^ {\\prime}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.454, + 0.354, + 0.471 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_B^\\prime = \\kappa \\odot \\mathbf{x}_C + (1 - \\kappa)\\odot \\mathbf{x}_P\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.231, + 0.494 + ], + "angle": 0, + "content": "4.3. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.484, + 0.53 + ], + "angle": 0, + "content": "Full Objective. The full objective functions to optimize \\(\\epsilon_{\\theta}\\) as Equation (13):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.54, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathcal {L} _ {m s e} + \\lambda \\cdot \\mathcal {L} _ {r e c}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.601 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyper-parameter controlling relative importance and is set to 0.1 based on functionality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.723 + ], + "angle": 0, + "content": "Refining Conditional Inference. During inference, traditional methods [3, 37, 40] sample noise \\(\\mathbf{x}_T\\) from a standard Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), unrelated to the target try-on. This introduces significant initial randomness and fails to preserve identity information and clothing details in the results [36]. Inspired by [36, 49], we design a refining conditional inference to enhance the model's performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Given the latent representation \\(\\mathbf{x}_B\\) (Eq. (9) during training) and sample a random Gaussian noise \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), we add \\(\\epsilon\\) to \\(\\mathbf{x}_B\\) at time step \\(T\\), as represented in Equation (14):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.779, + 0.483, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {B} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.805, + 0.484, + 0.836 + ], + "angle": 0, + "content": "To refine \\(\\hat{\\mathbf{x}}_T\\), we first decompose it using Singular Value Decomposition (SVD), which can be formulated in:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.843, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nU \\times \\Sigma \\times V ^ {\\mathrm {T}} = S V D (\\hat {\\mathbf {x}} _ {T}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\Sigma\\) is a diagonal matrix containing singular values, \\(U\\) and \\(V^{\\mathrm{T}}\\) represents the left and right singular vectors," + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.095, + 0.9, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.673, + 0.152, + 0.774, + 0.163 + ], + "angle": 0, + "content": "Diffusion-noisy \\(\\hat{x}_T\\)" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.163, + 0.899, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.211, + 0.763, + 0.22 + ], + "angle": 0, + "content": "SVD denoised" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.22, + 0.898, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Simulation comparison of denoising performance via SVD. From top to bottom: (row 1) clean signal \\(\\mathbf{x}_B\\), (row 2) diffusion-noisy signal \\(\\hat{\\mathbf{x}}_T\\), and (row 3) SVD-denoised reconstruction. SVD effectively suppresses the diffusion noise while preserving the dominant structural features of the original signal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.906, + 0.468 + ], + "angle": 0, + "content": "respectively. As shown in Figure 3, SVD can distinguish principal components from perturbations, captures key components of \\(\\hat{\\mathbf{x}}_T\\) via singular values. To enhance semantic information, we utilize a multi-layer perceptron (MLP) with multi-head self-attention layers [34] to predict a new singular value \\(\\widehat{\\Sigma}\\), as represented in Equation (16):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.477, + 0.906, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {T} ^ {\\prime} = \\varphi (U, \\Sigma , V ^ {\\mathrm {T}}), \\quad \\widehat {\\Sigma} = f (\\phi (\\hat {\\mathbf {x}} _ {T} ^ {\\prime}), \\mathbf {x} _ {B}) + \\Sigma , \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.906, + 0.582 + ], + "angle": 0, + "content": "where \\(\\varphi (\\cdot)\\) denotes the sum of all MLP-encoded inputs, \\(\\phi (\\cdot)\\) represents the multi-head self-attention layer, and \\(f(\\cdot)\\) is another MLP. Thus, target representation \\(\\mathbf{x}_T\\) is reformed by using \\(U\\), \\(\\widehat{\\Sigma}\\), \\(V^{\\mathrm{T}}\\) to align with real representation \\(\\tilde{\\mathbf{x}}_{GT}\\), as formulated in Equation (17):" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.591, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {T} = U \\times \\widehat {\\boldsymbol {\\Sigma}} \\times V ^ {\\mathrm {T}}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.621, + 0.898, + 0.637 + ], + "angle": 0, + "content": "Similarly, \\(\\tilde{\\pmb{x}}_{GT}\\) can be obtained through Equation (18):" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.647, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {G T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {P} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.906, + 0.707 + ], + "angle": 0, + "content": "After obtaining \\(\\tilde{\\mathbf{x}}_{GT}\\) and \\(\\mathbf{x}_T\\), we optimize \\(\\{\\varphi ,\\phi ,f\\}\\) by minimizing Equation (19):" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.715, + 0.905, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\varphi , \\phi , f} \\mathbb {E} _ {\\boldsymbol {C}, \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {P}} \\left[ \\| \\mathbf {x} _ {T} - \\tilde {\\boldsymbol {x}} _ {G T} \\| _ {2} ^ {2} \\right]. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.797 + ], + "angle": 0, + "content": "As shown in Figure 1, during inference, given a person \\(\\mathbf{x}_P\\) and a random clothing \\(\\mathcal{E}_{VAE}(C_{un}) = \\mathbf{x}_{C_{un}}\\), \\(\\mathbf{x}_B\\) is computed as: \\(\\mathbf{x}_B = \\kappa \\odot \\mathbf{x}_{C_{un}} + (1 - \\kappa) \\odot \\mathbf{x}_P\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.809, + 0.646, + 0.827 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.596, + 0.85 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dataset. We conduct experiments on the VITON-HD [4], VITON [16], and DressCode [29] datasets. VITON dataset comprises 16,253 image pairs with a resolution of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19529" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.184 + ], + "angle": 0, + "content": "\\(256 \\times 192\\), \\(14,221 / 2,032\\) training/testing pairs. VITON-HD dataset is composed of 11,647/2,032 training/testing pairs with a high-resolution of \\(512 \\times 384\\). DressCode dataset is another high-resolution dataset with a resolution of \\(512 \\times 384\\), which comprises 15,363 image pairs, 12,863/2,500 training/testing pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.295 + ], + "angle": 0, + "content": "Implementation Setting. Our method is implemented with PyTorch on a single NVIDIA A100 GPU. Follow the configuration of [6, 14], Stable Diffusion \\(\\mathrm{V}1.5^{3}\\) is used as our base model. We train for 20 epochs using the AdamW optimizer [27] with a base learning rate of \\(1e^{-5}\\) and a batch size of 2, and use the warm-up adjustment strategy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.303, + 0.383, + 0.318 + ], + "angle": 0, + "content": "5.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.522 + ], + "angle": 0, + "content": "Evaluation Metrics. We adopt Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [48], Fréchet Inception Distance (FID) [19], and Kernel Inception Distance (KID) [2], for quantitative evaluation. SSIM assesses pixel-level structural fidelity between generated and real images. LPIPS measures perceptual distance via deep features trained on human judgments; lower scores indicate higher perceptual similarity. FID computes the Wasserstein-2 distance between Inception-V3 feature distributions of real and generated images; lower is better. KID estimates distributional similarity through a maximum-mean-discrepancy kernel on the same Inception features; again, lower values denote higher fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Baseline Methods. To conduct objective quantitative experiments, we employed the following methods:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.661 + ], + "angle": 0, + "content": "- GAN-based Methods: CP-VTON [35], Cloth-flow [17], CP-VTON+ [28], VTNFP [46], ACGPN [42], DCTON [12], PF-AFN [13], ZFlow [7], OVNet [24], LM-VTON [26], DAFlow [1], Style-Flow [18], RT-VTON [43], DressCode [29], VITON-HD [4], HR-VITON [23], POVNet [25], GP-VTON [39], and USC-PFN [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.722 + ], + "angle": 0, + "content": "- Diffusion-based Methods: LaDI-VTON [30], PbE [41], DCI-VTON [14], StableVITON [22], OOTDiffusion [40], Anydoor [3], TPD [44], IDM-VTON [5], StableGarment [37], and LDE-VTON [10]." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Quantitative Comparison. Figure 4 visualizes our results on the VITON-HD dataset [4]. GAN-based approaches—GP-VTON [39] and USC-PFN [8], for example—lack the human-centric priors inherent to diffusion models. Consequently, their outputs are overly smooth, contain visible artifacts, and struggle to disambiguate crossed arms. In contrast, our method synthesizes more realistic garment wrinkles and correctly reconstructs cross-arm poses. Diffusion model-based methods, such as DCI-" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.089, + 0.905, + 0.373 + ], + "angle": 0, + "content": "
MethodsVenueMask\\( SSIM_p \\)↑\\( FID_{up} \\)↓
VITONCVPR'18✓0.7455.71
CP-VTON [35]ECCV'18✓0.7224.45
VTNFP [46]ICCV'19✓0.80n/a
Cloth-flow [17]CVPR'19✓0.8414.43
CP-VTON+ [28]CVPRW'20✓0.7521.04
ACGPN [42]CVPR'20✓0.8416.64
LM-VTON [26]AAAI'21✓0.8517.18
DCTON [12]CVPR'21✓0.8314.82
ZFlow [7]ICCV'21✓0.8815.17
OVNet [24]CVPR'21✓0.8515.78
PF-AFN [13]CVPR'21✘0.8910.21
RT-VTON [43]CVPR'22✓n/a11.66
DAFlow [1]ECCV'22✓0.8812.05
DressCode [29]CVPR'22✓0.8913.71
POVNet [25]TPAMI'23✓0.8913.37
PbE [41]CVPR'23✓0.8312.56
USC-PFN [8]NeurIPS'23✘0.9110.47
TPD [44]CVPR'24✓0.899.58
LDE-VTON [10]AAAI'25✘0.919.86
OursThis Work✘0.919.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.383, + 0.907, + 0.426 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on the VITON dataset. \"Mask\" denotes whether the mask is used during inference. Bold denotes the best result. Some data are from published literature." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.668 + ], + "angle": 0, + "content": "VTON [14], StableVITON [22], StableGarment [37], LDEVTON [10], and OOTDiffusion [40], limited by the diversity of training samples, learn fixed sample transformations. Consequently, they often fail to generate natural and realistic arms when dealing with unusual poses and clothing (e.g., the \\(2^{nd}\\) row). Additionally, since the randomness of Gaussian noise during inference is not addressed, the results often lose identity information (e.g., the belt in the \\(1^{st}\\) row) and clothing details (e.g., garment shape and texture in the \\(3^{rd}\\) row). In contrast, our approach addresses these issues by (1) augmenting training sample diversity to improve robustness, (2) enhancing the model architecture for higher synthesis quality, and (3) optimizing the handling of Gaussian noise to preserve identity and garment details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison. Table 1 and Table 2 present quantitative comparisons between our method and the baselines on the VITON [16], VITON-HD [4], and DressCode [29] datasets. Our approach attains the highest scores on VITON dataset, demonstrating its robustness to low-resolution inputs. On VITON-HD and DressCode datasets, it ranks either first or second. The occasional second place arises because most baselines are trained on a fixed set of clothing images, leading to overfitting between specific garments and reference identities. In contrast, we enlarge appearance diversity by synthesizing additional pseudosamples and introducing out-of-distribution information; this strategy slightly penalizes performance on test cases that closely match the original data distribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.468, + 0.902 + ], + "angle": 0, + "content": "3https://huggingface.co/stable-diffusion-v1-5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19530" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 4. Qualitative results with different baseline methods on the VITON-HD dataset. The left area shows a visual comparison with GAN-based baseline methods, while the right area shows a comparison with diffusion model-based baseline methods. It can be seen that our method achieves the best visual performance. Red dashed box represents defects." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.409, + 0.903, + 0.628 + ], + "angle": 0, + "content": "
Train / Test MethodsVenueVITON-HDDressCode UpperMask
\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)
VITON-HD [4]CVPR'210.8620.11712.1173.23n/an/an/an/a✓
HR-VITON [23]ECCV'220.8780.10511.2652.730.9360.06513.822.71✓
GP-VTON [39]CVPR'230.8840.0819.7011.260.7690.27020.118.17✓
LaDI-VTON [30]MM'230.8640.0969.4801.990.9150.06314.263.33✓
PbE [41]CVPR'230.8020.14311.9393.850.8970.07815.334.64✓
DCI-VTON [14]MM'230.8800.0809.4101.100.9370.04211.921.89✓
StableVITON [22]CVPR'240.8640.0849.4651.400.9110.05011.270.72✓
StableGarment [37]arXiv'240.8030.10417.1158.85n/an/an/an/a✓
IDM-VTON [5]ECCV'240.8500.0609.8421.120.8800.0569.544.32✓
AnyDoor [3]CVPR'240.8210.09910.8502.460.8990.11914.833.05✓
LDE-VTON [10]AAAI'250.8980.0819.6401.21n/an/an/an/a✓
OursThis Work0.9010.0799.3841.120.9390.04810.951.78✘
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.639, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons on the VITON-HD and DressCode datasets. For LPIPS, FID, and KID, the lower the better. For SSIM, the higher the better. Underline represents second best. Some data are from published literature." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.245, + 0.709 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Effectiveness of Text-Driven Pseudo-Input Preparation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Figure 5 (A) and Table 3 (A) highlight the necessity of text-driven pseudo-input preparation (TDPIP). By increasing the appearance diversity of pseudo-samples, it enhances the generator's generalization ability for abnormal samples. Without this component, the arm regions of abnormal samples are poorly generated, and the clothing is not properly fitted. In contrast, with this preparation, the arms are well generated, and the clothing is effectively transferred." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Gated Manipulation. Figure 6 demonstrates the necessity of introducing gated manipulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.907, + 0.771 + ], + "angle": 0, + "content": "(GM) to prevent catastrophic forgetting. It can be seen that when we remove GM from the baseline method, the generator with altered input structure requires nearly 75 epochs to converge. In contrast, when GM is introduced, the model achieves good convergence after only 10 epochs of training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Texture-aware Injection. Figure 5 (B) and Table 3 (B) demonstrate the necessity of introducing texture-aware injection (TAI) method for capturing clothing texture details. It can be seen that when TAI is introduced into the baseline generator, the text on the target clothing image is faithfully reconstructed. In contrast, when it is removed, some text details are lost by the generator." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19531" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.907, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Visual ablation study of different components in our approach. (A) Text-Driven Pseudo-Input Preparation (TDPIP). (B) Texture-aware Injection (TAI). (C) Refining Conditional Inference (RCI). The blue dashed box represents the area of focus for comparison." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.304, + 0.487, + 0.398 + ], + "angle": 0, + "content": "
ConfigurationSSIMp↑FIDup↓
Baseline Framework0.8312.56
+ (A) Text-Driven Pseudo-Input0.879.86
+ (B) Texture-aware Injection0.899.61
+ (C) Refining Conditional Inference0.919.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.45 + ], + "angle": 0, + "content": "Table 3. Ablation study. Performance of various configurations on VITON dataset. Baseline is a basic teacher-student architecture with PbE [41]. Bold denotes the best result." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Effectiveness of Refining Conditional Inference. Finally, Figure 5 (C) and Table 3 (C) demonstrate the necessity of the proposed refining conditional inference (RCI) for information preservation and reducing generation randomness. When this initial noise generation mode is introduced into the framework, the clothing details in the try-on results are well reconstructed, and there is no loss of identity information. However, when this initial noise generation mode is removed, some clothing details are lost to varying degrees." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.218, + 0.64 + ], + "angle": 0, + "content": "5.4. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.784 + ], + "angle": 0, + "content": "Although our proposed method can generate high-quality try-on results within a mask-free framework, the target noise optimized by the proposed refining conditional inference approach is estimated from only a limited set of samples. This scarcity can introduce instability and, at the beginning of inference, may perturb the initial latent distribution. In future work, we will investigate more robust and data-driven adversarial augmentation strategies for target-noise generation to eliminate this limitation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.21, + 0.815 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present a unified, mask-free virtual try-on framework that boosts generation quality by increasing training sample diversity and improving both training and inference processes. First, we introduce a text-driven pseudo-input preparation strategy to enlarge data" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.306, + 0.715, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.307, + 0.901, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.453, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 6. Ablation study of the Gated Manipulation. It can be seen that with the use of the gated manipulation, the generator quickly converges to a better state in the early stages of training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.675 + ], + "angle": 0, + "content": "diversity and enhance model's generalization. Second, a gated-manipulation mechanism prevents catastrophic forgetting of pre-trained model weights while cutting training costs. Third, texture-aware injection explicitly supplements human-perceptible garment details. Finally, a refining conditional-inference scheme suppresses Gaussian-noise randomness to preserve identity and fine-grained clothing cues during inference. Extensive experiments on multiple benchmarks show that our proposed method consistently outperforms previous virtual try-on work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.909, + 0.889 + ], + "angle": 0, + "content": "Acknowledgments. This work was in part supported by the National Key Research and Development Program of China (Grant No. 2022ZD0160604), the National Natural Science Foundation of China (NSFC, Grant No. 62176194), the Key Research and Development Program of Hubei Province (Grant No. 2023BAB083), the Project of Sanya Yazhou Bay Science and Technology City (Grant Nos. SCKJ-JYRC-2022-76, SKJC-2022-PTDX-031), the Project of Sanya Science and Education Innovation Park of Wuhan University of Technology (Grant No. 2021KF0031), and the Huawei Kunpeng-Ascend Innovation Incentive Programme. This work was also supported in part by both computing resources from the Wuhan Supercomputing Center and the Wuhan Artificial Intelligence Computing Center." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19532" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 409-425. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[5] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild. In European Conference on Computer Vision, pages 206-235. Springer, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, Dongmei Jiang, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.563 + ], + "angle": 0, + "content": "[7] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5433-5442, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[8] Chenghu Du, Shuqing Liu, Shengwu Xiong, et al. Greatness in simplicity: Unified self-cycle consistency for parser-free virtual try-on. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[9] Chenghu Du, Junyin Wang, Kai Liu, and Shengwu Xiong. Glv: Geometric correlation distillation for latent diffusion-enhanced parser-free virtual try-on. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[10] Chenghu Du, Junyin Wang, Feng Yu, and Shengwu Xiong. Latent diffusion-enhanced virtual try-on via optimized pseudo-label generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2780-2788, 2025. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] D. Gabor. Theory of communication. part 1: The analysis of information. Journal of the Institution of Electrical Engineers - Part III: Radio and Communication Engineering, 93: 429-441, 1946. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[15] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[17] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[22] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 204-219. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Kedan Li, Min Jin Chong, Jeffrey Zhang, and Jingen Liu. Toward accurate and realistic outfits visualization with attention to details. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15546-15555, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Kedan Li, Jeffrey Zhang, and David Forsyth. Povnet: Image-based virtual try-on through accurate warping and residual." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19533" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12222-12235, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[26] Guoqiang Liu, Dan Song, Ruofeng Tong, and Min Tang. Toward realistic virtual try-on through landmark guided shape matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2118-2126, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.178, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[28] Matiur Rahman Minar, Thai Thanh Tuan, Heejune Ahn, Paul Rosin, and Yu-Kun Lai. Cp-vton+: Clothing shape and texture preserving image-based virtual try-on. In CVPR Workshops, pages 10-14, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[29] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[30] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.47, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.622 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[35] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[36] Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, and Hongming Shan. Fldm-vton: Faithful latent diffusion model for virtual try-on. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 1362–1370. International Joint Conferences on Artificial Intelligence Organization, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783, 2024. 1, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.16 + ], + "angle": 0, + "content": "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.246 + ], + "angle": 0, + "content": "[39] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23550-23559, 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[40] Yuhao Xu, Tao Gu, Weifeng Chen, and Arlene Chen. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8996-9004, 2025. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[41] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[42] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.907, + 0.514 + ], + "angle": 0, + "content": "[43] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3460-3469, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.907, + 0.584 + ], + "angle": 0, + "content": "[44] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[46] Ruiyun Yu, Xiaqi Wang, and Xiaohui Xie. Vtnfp: An image-based virtual try-on network with body and clothing feature preservation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10511-10520, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.713, + 0.905, + 0.768 + ], + "angle": 0, + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.77, + 0.905, + 0.838 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.84, + 0.905, + 0.895 + ], + "angle": 0, + "content": "[49]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.4,5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19534" + } + ] +] \ No newline at end of file diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_origin.pdf b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f2a3343a6d02ee9e299d79240d8923ead8c8ea7d --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/4207f645-0fb0-4a1e-b60a-46c12825cd89_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2a5352d62785b6db653f640e554b4a85fd44dd58d5f606008bc6ab00ee324c4 +size 3986255 diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/full.md b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/full.md new file mode 100644 index 0000000000000000000000000000000000000000..778e53e0fc3e3f26abc16f3648ad486a944aff5a --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/full.md @@ -0,0 +1,361 @@ +# All Parts Matter: A Unified Mask-Free Virtual Try-On Framework + +Chenghu Du $^{1,3}$ Shengwu Xiong $^{2*}$ Yi Rong $^{1,3*}$ $^{1}$ School of Computer Science and Artificial Intelligence, Wuhan University of Technology + $^{2}$ Interdisciplinary Artificial Intelligence Research Institute, Wuhan College + $^{3}$ Shanghai Artificial Intelligence Laboratory + +{duch, xiongsw, wrong}@whut.edu.cn + +# Abstract + +Current virtual try-on methods primarily enhance performance through network optimization, employing strategies such as coarse-to-fine structures and ReferenceNet to inject clothing information. However, these methods are often constrained by the limited quantity and diversity of training samples, which ultimately restricts their potential for further improvement. To address this challenge, we propose a unified, mask-free virtual try-on framework. Our approach leverages the inherent strengths of latent diffusion models to enhance the capability of each pipeline component to accurately model the target distribution, thereby achieving superior performance. Specifically, we introduce a text-driven pseudo-input preparation strategy that significantly increases the diversity of clothing regions within the generated person pseudo-samples. This encourages the generator to focus on variations in these areas and improves the model's generalization capability. Within the generator itself, we develop a gated manipulation mechanism to prevent weight forgetting and reduce training costs. Furthermore, we incorporate a texture-aware injection module to explicitly integrate human-perceptible clothing texture information into the generation process. During inference, we propose a refining conditional inference strategy that mitigates the randomness introduced by Gaussian noise. This effectively preserves identity information and fine clothing details in the final results. Extensive experiments demonstrate that our method outperforms existing state-of-the-art virtual try-on approaches. + +# 1. Introduction + +Image-based virtual try-on (VTON) aims to realistically transfer a target garment onto the corresponding region of a reference person [3, 8, 14, 22, 39]. Owing to its strong application prospects and commercial potential, VTON has + +attracted growing research interest. Leveraging the powerful generative capabilities of diffusion models [20, 41], recent VTON methods have significantly narrowed the gap to real-world fitting. + +Previous VTON methods are broadly categorized into mask-based and mask-free paradigms. Mask-based approaches [3, 14, 22, 37] typically operate in a self-supervised fashion: a clothing-agnostic mask (Figure 2) is first applied to remove the garment region and adjacent body parts from the reference image, and the generator is then trained to inpaint the masked areas, thereby learning to perform the virtual try-on task. Although effective and widely adopted, these methods rely heavily on obtaining accurate masks for each reference person; even minor mask errors can introduce conspicuous artifacts [13, 21]. To mitigate mask-induced artifacts, mask-free methods grounded in pseudo-sample [8, 13, 18] have been proposed. A mask-based VTON model is first pre-trained to synthesize try-on images for arbitrary garments on the reference person, yielding paired pseudo-samples. These pseudo-samples are then fed as inputs, while the original reference image serves as the reconstruction target (ground truth), to train a main generator. Although this pipeline obviates the need for precise masks, they are limited to a single, fixed transformation for each pseudo-sample/reference-person pair. Moreover, because both garments and reference images are typically drawn from the same dataset, the generator overfits to these specific items and generalizes poorly to unseen clothing. + +There are also some other methods attempt to boost performance through architectural improvements—such as coarse-to-fine networks and ReferenceNet-based clothing-feature injection [22, 37, 40]. Yet these refinements remain constrained by the input materials (human parsing, etc), causing the generator to converge to sub-optimal solutions. Finally, the randomness of Gaussian-noise sampling during inference further degrades output quality [36]—a problem often overlooked by prior work. + +To address these challenges, we propose a unified mask-free virtual try-on framework that leverages the intrinsic + +strengths of latent diffusion models to enable every pipeline component to better approximate the target data distribution and thus improve overall performance. Specifically, we introduce a text-driven pseudo-input preparation strategy to augment sample diversity by enriching the clothing regions in each pseudo-sample (see Figure 2). First, a pretrained Stable Diffusion inpainting model is conditioned on text prompts generated by a large language model (e.g., DeepSeek or GPT-series models) to produce multiple try-on results for the same identity. We then perform sample selection to filter out implausible images and retain the most realistic candidate as the final pseudo-sample. This process maximizes the variability of clothing regions and mitigates overfitting of the generator during training. + +During training phase, we leverage the off-the-shelf inpainting try-on model as the generator, thereby reducing training costs. To convert this model into a mask-free generator, we introduce mask-free generation learning process. Concretely, a gated manipulation mechanism fuses clothing and person latent representations in latent space, preventing catastrophic forgetting when clothing conditioning is added and avoiding structural changes that inflate training cost. Complementarily, texture-aware injection module explicitly injects fine-grained texture cues into the cross-attention layers, sharpening the generator's perception of garment details. + +During inference phase, to reduce the randomness of Gaussian noise, we propose a refining conditional inference in our optimization process. By optimizing the initial Gaussian noise to make it as close as possible to the distribution of real person samples, helping the generator improve the quality of the synthesized results. + +In summary, this paper makes the following main contributions: + +- We propose a unified mask-free virtual try-on framework, which improves the generator's performance from multiple aspects of both the training and inference processes. +- We present a text-driven pseudo-input preparation approach to improve the generator's generalization ability. +- We propose a gated manipulation to prevent weight forgetting and cut training costs. +- We introduce a texture-aware injection to explicitly add human-perceptible clothing texture information. +- We propose a refining conditional inference approach to reduce Gaussian noise randomness. + +# 2. Related Work + +Mask-Free Virtual Try-On. This task requires performing virtual try-on at inference time using only a clothing image and a person image—no masks, human parsing, or other auxiliary inputs. Issenhuth et al. [21] introduced the first mask-free method, WUTON, which attains fully supervised training by distilling knowledge from a mask-based teacher + +that generates paired training samples. Because the synthetic images are treated as ground truth for the student, the teacher's defects directly bound the student's performance. Ge et al. [13] relaxed this constraint with PFAFN, feeding pseudo-samples as inputs instead of supervision, thereby allowing the student to learn a more accurate mapping. He et al. [18] extended this paradigm in Flow-Style, adding explicit garment-deformation optimization to improve fit and realism. More recently, Du et al. [8] proposed USC-PFN, a two-stage architecture that employs cycle-consistency and a weight-shared generator to refine the garment region and try-on result iteratively. In contrast to these prior works, we present a unified mask-free framework that optimizes the pseudo-sample and enhances both training and inference through multi-stage optimization, yielding higher fidelity and better generalization. + +Controllable Diffusion Models. Recent advances in diffusion models demonstrate that generation can be guided by modalities beyond text. Specifically, the integration of external models to inject external features into diffusion models has produced remarkable results. For example, ControlNet [47] introduced an external model that leverages images, skeletons, and edge maps as conditional inputs to generate corresponding outputs. IP-adaptor [45] proposed a method for generating images similar to a given reference image. T2I-adapter [31] also introduced an external, more lightweight model that leverages conditional inputs to generate corresponding images. These approaches have effectively manipulated image generation by injecting feature values via external models. Instead of adding extra network branches, following the optimization idea of latent features in latent space of [9], we embed the try-on conditions directly into the latent space, yielding a high-performance virtual try-on model with minimal overhead. + +# 3. Preliminary: Latent Diffusion Models + +Latent Diffusion Models (LDMs) [33] are a class of generative models that learn to synthesize data $\mathbf{x} \in \mathbb{R}^{H \times W \times 3}$ by denoising latent representations $\mathbf{z} \in \mathbb{R}^{h \times w \times c}$ in a compressed space, where typically $h \ll H$ and $w \ll W$ . The compression is achieved via a pre-trained autoencoder with KL-regularization (e.g. VQGAN), consisting of an encoder $\mathcal{E}_{VAE}(\cdot)$ and decoder $\mathcal{D}_{VAE}(\cdot)$ , such that + +$$ +\mathbf {z} _ {0} = \mathcal {E} _ {V A E} (\mathbf {x}), \quad \hat {\mathbf {x}} = \mathcal {D} _ {V A E} (\mathbf {z} _ {0}), \quad \text {w i t h} \quad \hat {\mathbf {x}} \approx \mathbf {x}. \tag {1} +$$ + +The latent-space diffusion process is modeled as a Markov chain that progressively adds Gaussian noise to $\mathbf{z}_0$ (the encoded data) over $T$ steps, producing a sequence of noisy latents $\{\mathbf{z}_t\}_{t=0}^T$ : + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {2} +$$ + +![](images/fb3df232a54430247be2ac1461589a8fc6ca3b98bc550ab2988cd9d38da90621.jpg) + +![](images/664bdaa56cd71151f8a51cab7f18780c1be152d062ae9027c480ef96a6052a09.jpg) + +![](images/fd003c139c0430488fc11d5746c5aeebb55e665dea8c9170bab02868069fc7fd.jpg) +Sect. 4.1: Text-Driven Pseudo-Input Preparation + +![](images/57c3c7b1cf6fe27721a00d9327d80656355970e8076e3e5194e1601fb384c5ec.jpg) +Sect. 4.3: Optimization +Figure 1. Overview of the proposed virtual try-on framework. It consists of three sub-pipelines, namely Text-Driven Pseudo-Input Preparation, Mask-Free Generation Learning, and Optimization. In text-driven pseudo-input preparation (Sect. 4.1), we first use a pre-trained Stable Diffusion inpainting model to generate $N$ different pseudo-samples $\{I_i\}_{i=0}^N$ based on prompts generated by DeepSeek and various noises. Then, a sample selection method is proposed to pick the best one $I_{best}$ from the $N$ pseudo-samples for the next step. In mask-free generation learning (Sect. 4.2), we design a mask-free training approach that initially uses gated manipulation to fuse clothing and person representations to prevent catastrophic forgetting, and then injects extracted clothing texture features through cross-attention modules. In optimization (Sect. 4.3), we first train a noise optimization MLP to obtain the optimized Gaussian noise that preserves as much identity and fine-grained clothing cues as possible, and then use it for the inference phase. + +where $\{\beta_t\in (0,1)\}_{t = 1}^T$ is a fixed variance schedule. The forward process can be marginalized in closed form as + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {0}\right) = \mathcal {N} \left(\mathbf {z} _ {t}; \sqrt {\bar {\alpha} _ {t}} \mathbf {z} _ {0}, (1 - \bar {\alpha} _ {t}) \mathbf {I}\right), \tag {3} +$$ + +with $\alpha_{t} = 1 - \beta_{t}$ and $\bar{\alpha}_{t} = \prod_{s=1}^{t} \alpha_{s}$ . LDMs optimize a variational lower bound (VLB) on the log-likelihood of the data, simplified to the noise-prediction objective (DDPM-style): + +$$ +\min _ {\theta} \mathbb {E} _ {\mathbf {z} _ {0}, \epsilon \sim \mathcal {N} (0, \mathbf {I}), t} \left[ \| \epsilon_ {\theta} (\mathbf {z} _ {t}, t) - \epsilon \| _ {2} ^ {2} \right]. \tag {4} +$$ + +where $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ is the ground-truth noise and $\epsilon_{\theta}(\cdot)$ is a U-Net parameterized by $\theta$ . + +# 4. Methodology + +Overview. This task aims to transfer the target clothing from an arbitrary clothing image $C_{un} \in \mathbb{R}^{H \times W \times 3}$ onto a reference person image $P \in \mathbb{R}^{H \times W \times 3}$ , producing a try-on result $I \in \mathbb{R}^{H \times W \times 3}$ . To train the mask-free model, $I$ (generated from the off-the-shelf mask-based try-on method) is first used as pseudo-input, we construct triplets $\{C, I, P\}$ for fully supervised learning of the generator $G$ with parameters $\theta$ , i.e., + +$$ +\theta^ {*} = \arg \min _ {\theta} \sum_ {\boldsymbol {C}, \boldsymbol {I}, \boldsymbol {P}} \ell_ {\text {d i s t}} (G (\boldsymbol {C}, \boldsymbol {I}), \boldsymbol {P}), \tag {5} +$$ + +where $\ell_{dist}$ represents a suitable loss function (e.g., L1, MSE) between the generated image $G(C,I)$ and the + +ground-truth image $\pmb{P}$ . The clothing image $C \in \mathbb{R}^{H \times W \times 3}$ and the clothes worn by $\pmb{P}$ are identical and also sourced from the dataset. Sections that follow detail how we obtain a suitable pseudo-input $\pmb{I}$ for efficient training of the generator $G$ . An overview of our pipeline—comprising Text-Driven Pseudo-Input Preparation (Sect. 4.1), Mask-Free Generation Learning (Sect. 4.2), and Refining Conditional Inference (Sect. 4.3)—is provided in Figure 1. + +# 4.1. Text-Driven Pseudo-Input Preparation + +Previous mask-free methods [13, 18] rely on off-the-shelf try-on networks to generate the pseudo-input. Although training converges, the rigid one-to-one pairing of $C_{un} / I$ and $P$ drives the generator $G$ toward sub-optimal solutions and limits generalization. Therefore, diversifying the clothing regions of $I$ (see Figure 2) is crucial for improving model performance in mask-free try-ons. By increasing clothing variability, the generator $G$ can focus more effectively on clothing features and details. + +Automatic Sample Generation. To increase the visual diversity of $\pmb{I}$ and mitigate the generalization drop caused by the overlap between $C_{un}$ (for generating $\pmb{I}$ ) and $\pmb{C}$ , we leverage a Stable Diffusion inpainting model [37], $\mathcal{F}_{\mathrm{SD}}$ , which possesses strong generative priors, together with the large language model DeepSeek2. The resulting pipeline is + +summarized in Equation (6). + +$$ +\left\{\boldsymbol {I} _ {i} \right\} _ {i = 0} ^ {N} = \left\{\mathcal {F} _ {S D} \left(\boldsymbol {P} ^ {\star}, O, \xi_ {i}\right) \right\} _ {i = 0} ^ {N}, \tag {6} +$$ + +where $P^{\star}$ denotes the clothing-agnostic person representation, consisting of the clothing-agnostic person image, the clothing-agnostic mask, and the DensePose map [15]. $\xi$ is a random Gaussian noise sample, $\xi \sim \mathcal{N}(0,\mathbf{I})$ . $O$ is a text prompt constructed from the template "A photo of a person wearing {upper garment description}", with the description generated by DeepSeek to cover a wide range of clothing styles. For each prompt, we synthesize $N = 5$ images and select the best try-on result $I_{best}$ . The resulting triplets $\{C,I_{best},P\}$ are used for training. + +Sample Selection. To this end, we adopt the Human Preference Score $\nu 2$ (HPS v2) model [38] to rank the generated images $\{I_i\}_{i=1}^N$ according to their alignment with the corresponding prompt. HPS v2 employs an image encoder $E_{IMG}(\cdot)$ and a text encoder $E_{TEXT}(\cdot)$ to map images and prompts into a joint embedding space, where their cosine similarity serves as the preference score, as formulated in Equation (7). + +$$ +\boldsymbol {s} ^ {\eta} (\boldsymbol {I} _ {i}, O) = E _ {I M G} (\boldsymbol {I} _ {i}) \cdot E _ {T E X T} (O) \times 1 0 0. \tag {7} +$$ + +We select the image achieving the highest score as the final sample. However, some $I_{i}$ while semantically consistent with the prompt (see column 5 of Figure 2)—exhibit implausible appearances in the inpainting regions, such as distorted limbs. To address this issue, before using HPS v2, we introduce an appearance rationality discriminator $\mathcal{D}_r$ that outputs a rationality score $s^r(I_i) = \mathcal{D}_r(I_i)$ for each candidate image. This discriminator is trained by minimizing the objective given in Equation (8). + +$$ +\begin{array}{l} \min _ {\mathcal {D} _ {r}} - \mathbb {E} _ {\boldsymbol {P}, O} \left[ \log \left(\mathcal {D} _ {r} \left(\mathcal {F} _ {S D} \left(\boldsymbol {P} ^ {\star}, O, \xi_ {\text {b e s t}}\right)\right)\right) \right. \tag {8} \\ \left. + \log \left(1 - \mathcal {D} _ {r} \big (\mathcal {F} _ {S D} (\boldsymbol {P} ^ {\star}, O, \xi_ {w o r s t})\right)\right)\left. \right], \\ \end{array} +$$ + +where $\xi_{\mathrm{best}}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ denotes the Gaussian noise that yields the best result under prompt $O$ , and let $\xi_{\mathrm{worst}}\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ yield the worst. Training samples were manually screened to identify the best and worst ones. We then sort the $N$ scores $\{s^r (I_i)\}_{i = 1}^N$ and denote the two highest as $s_{(1)}^r$ and $s_{(2)}^r$ . The acceptance probability for choosing between them by using HPS v2 is given by $\mathbb{I}[s^{\eta}(s_{(1)}^{r}),s^{\eta}(s_{(2)}^{r})]$ where $m = 0$ as the filtering threshold and $\mathbb{I}[\cdot ]$ is the indicator function [38, 49]. The sample satisfying this condition is retained and used to build the target triplet $\{C,I_{best},P\}$ . + +# 4.2. Mask-Free Generation Learning + +In this stage, we leverage the triplet $\{C, I_{best}, P\}$ to generate the final virtual try-on result $P_{rec}$ , such that $P_{rec} \approx P$ . + +![](images/ea0dfecd90daf9807ddf00161efca625430b47b38d6b516c52c10ab6285eaf49.jpg) +Figure 2. Visualization of pseudo-samples. Under the same prompt, the pseudo-samples are generated from five different Gaussian noises. When a pseudo-sample is semantically aligned with the prompt, the semantic number is labeled on the image. The leftmost column shows the clothing-agnostic person and the clothing-agnostic mask. + +Gated Manipulation. It remains challenging to organically fuse $C$ and $I_{best}$ so that the generator $G$ can both preserve rich human details and synthesize realistic clothing, while suppressing artifacts. Existing approaches can be roughly divided into two groups: (a) Treat $C$ and $I_{best}$ as separate conditioning signals to $G$ [22, 40]. (b) Crucely concatenate them in the pixel space before feeding the result to $G$ [14]. Approach (a) demands an additional Clothing-ReferenceNet to inject the new condition via attention modules; otherwise, directly adding a new branch may trigger catastrophic forgetting of the pre-trained weights, significantly increasing training cost. Approach (b) suffers from detail loss and blurring caused by the naive pixel-level combination. To achieve high-quality try-on without extra heavy modules or training overhead, we propose gated manipulation. Concretely, we first encode $C$ and $I_{best}$ into latent features, then blend them through a learnable gate $\kappa \in [0,1]$ in the feature space. The fused feature is + +$$ +\mathbf {x} _ {B} = \kappa \odot \mathbf {x} _ {C} + (1 - \kappa) \odot \mathbf {x} _ {I}, \tag {9} +$$ + +where $\kappa = \varPhi_{gate}(\mathbf{x}_I,\mathbf{x}_C)$ $\odot$ denotes Hadamard product. $\mathbf{x}_I = \mathcal{E}_{VAE}(\mathbf{I}_{best})$ $\mathbf{x}_C = \mathcal{E}_{VAE}(\hat{C})$ $\hat{C}$ is the warped $C$ $\varPhi_{gate}(\cdot)$ is a gate generator pre-trained by + +$$ +\Phi_ {g a t e} ^ {*} = \arg \min _ {\Phi_ {g a t e}} \ell_ {d i s t} (\mathbf {x} _ {P}, \mathbf {x} _ {B}), \tag {10} +$$ + +where $\mathbf{x}_P = \mathcal{E}_{VAE}(P)$ . Since $\kappa$ participates in encoding rich semantic and contextual cues, the blended feature $\mathbf{x}_B$ obtained via Equation (9) effectively suppresses redundant clothing patterns from $I_{best}$ and background clutter from $C$ . At the same time, it preserves the holistic understanding of the entire $I_{best}$ (e.g., pose and identity), thereby mitigating catastrophic forgetting that typically occurs when the conditions for the new modality are fed to the generator. + +Texture-aware Injection. To enhance the generator's sensitivity to clothing texture, we adopt a bank of Gabor filters [11]—oriented at $0^{\circ}$ , $45^{\circ}$ , $90^{\circ}$ , and $135^{\circ}$ —to convolve + +the clothing $C$ . Each filter produces an orientation-specific response; the maximum across these four responses is taken at every spatial location, yielding a texture map $C_g$ that highlights dominant fabric patterns while suppressing noise. A domain-specific encoder $\Psi(\cdot)$ [32] then projects $C_g$ into an intermediate latent representation $\Psi(C_g)$ , which is supplied to the cross-attention layer via learned key and value projections $K^{\star} = W_k^{(i)}\Psi(C_g), V^{\star} = W_v^{(i)}\Psi(C_g)$ . So, the objective of this stage is defined as Equation (11): + +$$ +\mathcal {L} _ {m s e} = \mathbb {E} _ {\boldsymbol {C}, \boldsymbol {I} _ {\text {b e s t}}, \boldsymbol {P}, \epsilon \sim \mathcal {N} (0, I), t \sim U [ 0, T)} \tag {11} +$$ + +$$ +\left[ \left\| \boldsymbol {\epsilon} - \epsilon_ {\theta} \big (\mathbf {z} _ {t}, \mathbf {x} _ {B}, \boldsymbol {\Psi} (\boldsymbol {C} _ {g}), t \big) \right\| _ {2} ^ {2} \right], +$$ + +where $\mathbf{z}_t = \sqrt{\bar{\alpha}_t}\mathbf{x}_P + \sqrt{1 - \bar{\alpha}_t\epsilon}.\| \cdot \| _2^2$ denote the L2 norm. Considering that $\hat{C}$ is usually coarsely warped, in the early stages of training, we considering directing $G$ 's focus towards clothing regions by preventing learning changes in other areas caused by $I_{best}$ 's potential defects. To this end, we conduct source-to-source self-reconstruction for training, as formulated in Equation (12): + +$$ +\mathcal {L} _ {r e c} = \mathbb {E} _ {\boldsymbol {C}, \boldsymbol {P}, \epsilon \sim \mathcal {N} (0, I), t \sim U [ 0, T)} \tag {12} +$$ + +$$ +\left[ \left\| \epsilon - \epsilon_ {\theta} \big (\mathbf {z} _ {t}, \mathbf {x} _ {B} ^ {\prime}, \boldsymbol {\Psi} (\boldsymbol {C} _ {g}), t \big) \right\| _ {2} ^ {2} \right], +$$ + +where $\mathbf{x}_B^\prime = \kappa \odot \mathbf{x}_C + (1 - \kappa)\odot \mathbf{x}_P$ + +# 4.3. Optimization + +Full Objective. The full objective functions to optimize $\epsilon_{\theta}$ as Equation (13): + +$$ +\min _ {\theta} \mathcal {L} _ {m s e} + \lambda \cdot \mathcal {L} _ {r e c}, \tag {13} +$$ + +where $\lambda$ is a hyper-parameter controlling relative importance and is set to 0.1 based on functionality. + +Refining Conditional Inference. During inference, traditional methods [3, 37, 40] sample noise $\mathbf{x}_T$ from a standard Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ , unrelated to the target try-on. This introduces significant initial randomness and fails to preserve identity information and clothing details in the results [36]. Inspired by [36, 49], we design a refining conditional inference to enhance the model's performance. + +Given the latent representation $\mathbf{x}_B$ (Eq. (9) during training) and sample a random Gaussian noise $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , we add $\epsilon$ to $\mathbf{x}_B$ at time step $T$ , as represented in Equation (14): + +$$ +\hat {\mathbf {x}} _ {T} = \sqrt {\bar {\alpha} _ {T}} \mathbf {x} _ {B} + \sqrt {1 - \bar {\alpha} _ {T}} \epsilon . \tag {14} +$$ + +To refine $\hat{\mathbf{x}}_T$ , we first decompose it using Singular Value Decomposition (SVD), which can be formulated in: + +$$ +U \times \Sigma \times V ^ {\mathrm {T}} = S V D (\hat {\mathbf {x}} _ {T}), \tag {15} +$$ + +where $\Sigma$ is a diagonal matrix containing singular values, $U$ and $V^{\mathrm{T}}$ represents the left and right singular vectors, + +![](images/0149006772c10981675f98ce38fe391c27f1a82df41b30ffb633f8374bdb013a.jpg) + +![](images/81762b678792ecfa4d50d1d6d02728d53508e2c3d8e2c7212fcfd6d469f95faa.jpg) +Diffusion-noisy $\hat{x}_T$ + +![](images/cf3f4e3fc7cd087df2458647f526f8a51d8f26a9cab07d782b1867c3fcaa613f.jpg) +SVD denoised +Figure 3. Simulation comparison of denoising performance via SVD. From top to bottom: (row 1) clean signal $\mathbf{x}_B$ , (row 2) diffusion-noisy signal $\hat{\mathbf{x}}_T$ , and (row 3) SVD-denoised reconstruction. SVD effectively suppresses the diffusion noise while preserving the dominant structural features of the original signal. + +respectively. As shown in Figure 3, SVD can distinguish principal components from perturbations, captures key components of $\hat{\mathbf{x}}_T$ via singular values. To enhance semantic information, we utilize a multi-layer perceptron (MLP) with multi-head self-attention layers [34] to predict a new singular value $\widehat{\Sigma}$ , as represented in Equation (16): + +$$ +\hat {\mathbf {x}} _ {T} ^ {\prime} = \varphi (U, \Sigma , V ^ {\mathrm {T}}), \quad \widehat {\Sigma} = f (\phi (\hat {\mathbf {x}} _ {T} ^ {\prime}), \mathbf {x} _ {B}) + \Sigma , \tag {16} +$$ + +where $\varphi (\cdot)$ denotes the sum of all MLP-encoded inputs, $\phi (\cdot)$ represents the multi-head self-attention layer, and $f(\cdot)$ is another MLP. Thus, target representation $\mathbf{x}_T$ is reformed by using $U$ , $\widehat{\Sigma}$ , $V^{\mathrm{T}}$ to align with real representation $\tilde{\mathbf{x}}_{GT}$ , as formulated in Equation (17): + +$$ +\mathbf {x} _ {T} = U \times \widehat {\boldsymbol {\Sigma}} \times V ^ {\mathrm {T}}. \tag {17} +$$ + +Similarly, $\tilde{\pmb{x}}_{GT}$ can be obtained through Equation (18): + +$$ +\tilde {\mathbf {x}} _ {G T} = \sqrt {\bar {\alpha} _ {T}} \mathbf {x} _ {P} + \sqrt {1 - \bar {\alpha} _ {T}} \epsilon . \tag {18} +$$ + +After obtaining $\tilde{\mathbf{x}}_{GT}$ and $\mathbf{x}_T$ , we optimize $\{\varphi ,\phi ,f\}$ by minimizing Equation (19): + +$$ +\min _ {\varphi , \phi , f} \mathbb {E} _ {\boldsymbol {C}, \epsilon \sim \mathcal {N} (0, I), \boldsymbol {P}} \left[ \| \mathbf {x} _ {T} - \tilde {\boldsymbol {x}} _ {G T} \| _ {2} ^ {2} \right]. \tag {19} +$$ + +As shown in Figure 1, during inference, given a person $\mathbf{x}_P$ and a random clothing $\mathcal{E}_{VAE}(C_{un}) = \mathbf{x}_{C_{un}}$ , $\mathbf{x}_B$ is computed as: $\mathbf{x}_B = \kappa \odot \mathbf{x}_{C_{un}} + (1 - \kappa) \odot \mathbf{x}_P$ . + +# 5. Experiments + +# 5.1. Setup + +Dataset. We conduct experiments on the VITON-HD [4], VITON [16], and DressCode [29] datasets. VITON dataset comprises 16,253 image pairs with a resolution of + +$256 \times 192$ , $14,221 / 2,032$ training/testing pairs. VITON-HD dataset is composed of 11,647/2,032 training/testing pairs with a high-resolution of $512 \times 384$ . DressCode dataset is another high-resolution dataset with a resolution of $512 \times 384$ , which comprises 15,363 image pairs, 12,863/2,500 training/testing pairs. + +Implementation Setting. Our method is implemented with PyTorch on a single NVIDIA A100 GPU. Follow the configuration of [6, 14], Stable Diffusion $\mathrm{V}1.5^{3}$ is used as our base model. We train for 20 epochs using the AdamW optimizer [27] with a base learning rate of $1e^{-5}$ and a batch size of 2, and use the warm-up adjustment strategy. + +# 5.2. Comparison with SOTA Methods + +Evaluation Metrics. We adopt Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [48], Fréchet Inception Distance (FID) [19], and Kernel Inception Distance (KID) [2], for quantitative evaluation. SSIM assesses pixel-level structural fidelity between generated and real images. LPIPS measures perceptual distance via deep features trained on human judgments; lower scores indicate higher perceptual similarity. FID computes the Wasserstein-2 distance between Inception-V3 feature distributions of real and generated images; lower is better. KID estimates distributional similarity through a maximum-mean-discrepancy kernel on the same Inception features; again, lower values denote higher fidelity. + +Baseline Methods. To conduct objective quantitative experiments, we employed the following methods: + +- GAN-based Methods: CP-VTON [35], Cloth-flow [17], CP-VTON+ [28], VTNFP [46], ACGPN [42], DCTON [12], PF-AFN [13], ZFlow [7], OVNet [24], LM-VTON [26], DAFlow [1], Style-Flow [18], RT-VTON [43], DressCode [29], VITON-HD [4], HR-VITON [23], POVNet [25], GP-VTON [39], and USC-PFN [8]. +- Diffusion-based Methods: LaDI-VTON [30], PbE [41], DCI-VTON [14], StableVITON [22], OOTDiffusion [40], Anydoor [3], TPD [44], IDM-VTON [5], StableGarment [37], and LDE-VTON [10]. + +Quantitative Comparison. Figure 4 visualizes our results on the VITON-HD dataset [4]. GAN-based approaches—GP-VTON [39] and USC-PFN [8], for example—lack the human-centric priors inherent to diffusion models. Consequently, their outputs are overly smooth, contain visible artifacts, and struggle to disambiguate crossed arms. In contrast, our method synthesizes more realistic garment wrinkles and correctly reconstructs cross-arm poses. Diffusion model-based methods, such as DCI- + +
MethodsVenueMask\( SSIM_p \)↑\( FID_{up} \)↓
VITONCVPR'18✓0.7455.71
CP-VTON [35]ECCV'18✓0.7224.45
VTNFP [46]ICCV'19✓0.80n/a
Cloth-flow [17]CVPR'19✓0.8414.43
CP-VTON+ [28]CVPRW'20✓0.7521.04
ACGPN [42]CVPR'20✓0.8416.64
LM-VTON [26]AAAI'21✓0.8517.18
DCTON [12]CVPR'21✓0.8314.82
ZFlow [7]ICCV'21✓0.8815.17
OVNet [24]CVPR'21✓0.8515.78
PF-AFN [13]CVPR'21✘0.8910.21
RT-VTON [43]CVPR'22✓n/a11.66
DAFlow [1]ECCV'22✓0.8812.05
DressCode [29]CVPR'22✓0.8913.71
POVNet [25]TPAMI'23✓0.8913.37
PbE [41]CVPR'23✓0.8312.56
USC-PFN [8]NeurIPS'23✘0.9110.47
TPD [44]CVPR'24✓0.899.58
LDE-VTON [10]AAAI'25✘0.919.86
OursThis Work✘0.919.52
+ +Table 1. Quantitative comparisons on the VITON dataset. "Mask" denotes whether the mask is used during inference. Bold denotes the best result. Some data are from published literature. + +VTON [14], StableVITON [22], StableGarment [37], LDEVTON [10], and OOTDiffusion [40], limited by the diversity of training samples, learn fixed sample transformations. Consequently, they often fail to generate natural and realistic arms when dealing with unusual poses and clothing (e.g., the $2^{nd}$ row). Additionally, since the randomness of Gaussian noise during inference is not addressed, the results often lose identity information (e.g., the belt in the $1^{st}$ row) and clothing details (e.g., garment shape and texture in the $3^{rd}$ row). In contrast, our approach addresses these issues by (1) augmenting training sample diversity to improve robustness, (2) enhancing the model architecture for higher synthesis quality, and (3) optimizing the handling of Gaussian noise to preserve identity and garment details. + +Qualitative Comparison. Table 1 and Table 2 present quantitative comparisons between our method and the baselines on the VITON [16], VITON-HD [4], and DressCode [29] datasets. Our approach attains the highest scores on VITON dataset, demonstrating its robustness to low-resolution inputs. On VITON-HD and DressCode datasets, it ranks either first or second. The occasional second place arises because most baselines are trained on a fixed set of clothing images, leading to overfitting between specific garments and reference identities. In contrast, we enlarge appearance diversity by synthesizing additional pseudosamples and introducing out-of-distribution information; this strategy slightly penalizes performance on test cases that closely match the original data distribution. + +![](images/268735c1a8ca0f803de09a0358ce7cd5da65f7dfe71966478f3b8537e696523c.jpg) +Figure 4. Qualitative results with different baseline methods on the VITON-HD dataset. The left area shows a visual comparison with GAN-based baseline methods, while the right area shows a comparison with diffusion model-based baseline methods. It can be seen that our method achieves the best visual performance. Red dashed box represents defects. + +
Train / Test MethodsVenueVITON-HDDressCode UpperMask
\(SSIM_p\uparrow\)\(LPIPS_p\downarrow\)\(FID_{up}\downarrow\)\(KID_{up}\downarrow\)\(SSIM_p\uparrow\)\(LPIPS_p\downarrow\)\(FID_{up}\downarrow\)\(KID_{up}\downarrow\)
VITON-HD [4]CVPR'210.8620.11712.1173.23n/an/an/an/a✓
HR-VITON [23]ECCV'220.8780.10511.2652.730.9360.06513.822.71✓
GP-VTON [39]CVPR'230.8840.0819.7011.260.7690.27020.118.17✓
LaDI-VTON [30]MM'230.8640.0969.4801.990.9150.06314.263.33✓
PbE [41]CVPR'230.8020.14311.9393.850.8970.07815.334.64✓
DCI-VTON [14]MM'230.8800.0809.4101.100.9370.04211.921.89✓
StableVITON [22]CVPR'240.8640.0849.4651.400.9110.05011.270.72✓
StableGarment [37]arXiv'240.8030.10417.1158.85n/an/an/an/a✓
IDM-VTON [5]ECCV'240.8500.0609.8421.120.8800.0569.544.32✓
AnyDoor [3]CVPR'240.8210.09910.8502.460.8990.11914.833.05✓
LDE-VTON [10]AAAI'250.8980.0819.6401.21n/an/an/an/a✓
OursThis Work0.9010.0799.3841.120.9390.04810.951.78✘
+ +Table 2. Quantitative comparisons on the VITON-HD and DressCode datasets. For LPIPS, FID, and KID, the lower the better. For SSIM, the higher the better. Underline represents second best. Some data are from published literature. + +# 5.3. Ablation Study + +# Effectiveness of Text-Driven Pseudo-Input Preparation. + +Figure 5 (A) and Table 3 (A) highlight the necessity of text-driven pseudo-input preparation (TDPIP). By increasing the appearance diversity of pseudo-samples, it enhances the generator's generalization ability for abnormal samples. Without this component, the arm regions of abnormal samples are poorly generated, and the clothing is not properly fitted. In contrast, with this preparation, the arms are well generated, and the clothing is effectively transferred. + +Effectiveness of Gated Manipulation. Figure 6 demonstrates the necessity of introducing gated manipulation + +(GM) to prevent catastrophic forgetting. It can be seen that when we remove GM from the baseline method, the generator with altered input structure requires nearly 75 epochs to converge. In contrast, when GM is introduced, the model achieves good convergence after only 10 epochs of training. + +Effectiveness of Texture-aware Injection. Figure 5 (B) and Table 3 (B) demonstrate the necessity of introducing texture-aware injection (TAI) method for capturing clothing texture details. It can be seen that when TAI is introduced into the baseline generator, the text on the target clothing image is faithfully reconstructed. In contrast, when it is removed, some text details are lost by the generator. + +![](images/1f9ef0e66ab130b2a0fccf4a6dc30004bcb0d0d4d0a0a9d3ae0af489fdc09ee2.jpg) +Figure 5. Visual ablation study of different components in our approach. (A) Text-Driven Pseudo-Input Preparation (TDPIP). (B) Texture-aware Injection (TAI). (C) Refining Conditional Inference (RCI). The blue dashed box represents the area of focus for comparison. + +
ConfigurationSSIMp↑FIDup↓
Baseline Framework0.8312.56
+ (A) Text-Driven Pseudo-Input0.879.86
+ (B) Texture-aware Injection0.899.61
+ (C) Refining Conditional Inference0.919.52
+ +Table 3. Ablation study. Performance of various configurations on VITON dataset. Baseline is a basic teacher-student architecture with PbE [41]. Bold denotes the best result. + +Effectiveness of Refining Conditional Inference. Finally, Figure 5 (C) and Table 3 (C) demonstrate the necessity of the proposed refining conditional inference (RCI) for information preservation and reducing generation randomness. When this initial noise generation mode is introduced into the framework, the clothing details in the try-on results are well reconstructed, and there is no loss of identity information. However, when this initial noise generation mode is removed, some clothing details are lost to varying degrees. + +# 5.4. Limitations + +Although our proposed method can generate high-quality try-on results within a mask-free framework, the target noise optimized by the proposed refining conditional inference approach is estimated from only a limited set of samples. This scarcity can introduce instability and, at the beginning of inference, may perturb the initial latent distribution. In future work, we will investigate more robust and data-driven adversarial augmentation strategies for target-noise generation to eliminate this limitation. + +# 6. Conclusion + +In this paper, we present a unified, mask-free virtual try-on framework that boosts generation quality by increasing training sample diversity and improving both training and inference processes. First, we introduce a text-driven pseudo-input preparation strategy to enlarge data + +![](images/3f43ec90fa996478dbaeaa1b2c2ef42e31198cc77722edd2783e4c2d1852067e.jpg) +Figure 6. Ablation study of the Gated Manipulation. It can be seen that with the use of the gated manipulation, the generator quickly converges to a better state in the early stages of training. + +![](images/e6cfc4e440a7ae9cc46db2ac5e8daa2573bf4561a0a840179845bad30a591812.jpg) + +diversity and enhance model's generalization. Second, a gated-manipulation mechanism prevents catastrophic forgetting of pre-trained model weights while cutting training costs. Third, texture-aware injection explicitly supplements human-perceptible garment details. Finally, a refining conditional-inference scheme suppresses Gaussian-noise randomness to preserve identity and fine-grained clothing cues during inference. Extensive experiments on multiple benchmarks show that our proposed method consistently outperforms previous virtual try-on work. + +Acknowledgments. This work was in part supported by the National Key Research and Development Program of China (Grant No. 2022ZD0160604), the National Natural Science Foundation of China (NSFC, Grant No. 62176194), the Key Research and Development Program of Hubei Province (Grant No. 2023BAB083), the Project of Sanya Yazhou Bay Science and Technology City (Grant Nos. SCKJ-JYRC-2022-76, SKJC-2022-PTDX-031), the Project of Sanya Science and Education Innovation Park of Wuhan University of Technology (Grant No. 2021KF0031), and the Huawei Kunpeng-Ascend Innovation Incentive Programme. This work was also supported in part by both computing resources from the Wuhan Supercomputing Center and the Wuhan Artificial Intelligence Computing Center. + +# References + +[1] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 409-425. Springer, 2022. 6 +[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6 +[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 1, 5, 6, 7 +[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 5, 6, 7 +[5] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild. In European Conference on Computer Vision, pages 206-235. Springer, 2024. 6, 7 +[6] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, Dongmei Jiang, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 6 +[7] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5433-5442, 2021. 6 +[8] Chenghu Du, Shuqing Liu, Shengwu Xiong, et al. Greatness in simplicity: Unified self-cycle consistency for parser-free virtual try-on. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6 +[9] Chenghu Du, Junyin Wang, Kai Liu, and Shengwu Xiong. Glv: Geometric correlation distillation for latent diffusion-enhanced parser-free virtual try-on. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2 +[10] Chenghu Du, Junyin Wang, Feng Yu, and Shengwu Xiong. Latent diffusion-enhanced virtual try-on via optimized pseudo-label generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2780-2788, 2025. 6, 7 +[11] D. Gabor. Theory of communication. part 1: The analysis of information. Journal of the Institution of Electrical Engineers - Part III: Radio and Communication Engineering, 93: 429-441, 1946. 4 +[12] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 6 +[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling + +appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1, 2, 3, 6 +[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 1, 4, 6, 7 +[15] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4 +[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 5, 6 +[17] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 6 +[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1, 2, 3, 6 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 1 +[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 1, 2 +[22] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 1, 4, 6, 7 +[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 204-219. Springer, 2022. 6, 7 +[24] Kedan Li, Min Jin Chong, Jeffrey Zhang, and Jingen Liu. Toward accurate and realistic outfits visualization with attention to details. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15546-15555, 2021. 6 +[25] Kedan Li, Jeffrey Zhang, and David Forsyth. Povnet: Image-based virtual try-on through accurate warping and residual. + +IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12222-12235, 2023. 6 +[26] Guoqiang Liu, Dan Song, Ruofeng Tong, and Min Tang. Toward realistic virtual try-on through landmark guided shape matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2118-2126, 2021. 6 +[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[28] Matiur Rahman Minar, Thai Thanh Tuan, Heejune Ahn, Paul Rosin, and Yu-Kun Lai. Cp-vton+: Clothing shape and texture preserving image-based virtual try-on. In CVPR Workshops, pages 10-14, 2020. 6 +[29] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 5, 6 +[30] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 6, 7 +[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[35] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 6 +[36] Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, and Hongming Shan. Fldm-vton: Faithful latent diffusion model for virtual try-on. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 1362–1370. International Joint Conferences on Artificial Intelligence Organization, 2024. 1, 5 +[37] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783, 2024. 1, 3, 5, 6, 7 + +[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 4 +[39] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23550-23559, 2023. 1, 6, 7 +[40] Yuhao Xu, Tao Gu, Weifeng Chen, and Arlene Chen. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8996-9004, 2025. 1, 4, 5, 6 +[41] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 1, 6, 7, 8 +[42] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 6 +[43] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3460-3469, 2022. 6 +[44] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 6 +[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2 +[46] Ruiyun Yu, Xiaqi Wang, and Xiaohui Xie. Vtnfp: An image-based virtual try-on network with body and clothing feature preservation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10511-10520, 2019. 6 +[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[49]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.4,5 \ No newline at end of file diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/images.zip b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0fcfa5f1e772be11812549852ed579804e1a21ef --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:635cd95a686df88ab86b11d52fdc45a154fe9e8f5f3e558188e5a3bc7f877a3d +size 732515 diff --git a/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/layout.json b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b160b1f0970c637069d6bb9fde0cf4b1c57ad001 --- /dev/null +++ b/2025/All Parts Matter_ A Unified Mask-Free Virtual Try-On Framework/layout.json @@ -0,0 +1,10292 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 515, + 121 + ], + "type": "text", + "content": "All Parts Matter: A Unified Mask-Free Virtual Try-On Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "spans": [ + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": "Chenghu Du" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": " Shengwu Xiong" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": " Yi Rong" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": "School of Computer Science and Artificial Intelligence, Wuhan University of Technology \n" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": "Interdisciplinary Artificial Intelligence Research Institute, Wuhan College \n" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 87, + 142, + 523, + 199 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 211, + 201, + 394, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 201, + 394, + 213 + ], + "spans": [ + { + "bbox": [ + 211, + 201, + 394, + 213 + ], + "type": "text", + "content": "{duch, xiongsw, wrong}@whut.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 601 + ], + "type": "text", + "content": "Current virtual try-on methods primarily enhance performance through network optimization, employing strategies such as coarse-to-fine structures and ReferenceNet to inject clothing information. However, these methods are often constrained by the limited quantity and diversity of training samples, which ultimately restricts their potential for further improvement. To address this challenge, we propose a unified, mask-free virtual try-on framework. Our approach leverages the inherent strengths of latent diffusion models to enhance the capability of each pipeline component to accurately model the target distribution, thereby achieving superior performance. Specifically, we introduce a text-driven pseudo-input preparation strategy that significantly increases the diversity of clothing regions within the generated person pseudo-samples. This encourages the generator to focus on variations in these areas and improves the model's generalization capability. Within the generator itself, we develop a gated manipulation mechanism to prevent weight forgetting and reduce training costs. Furthermore, we incorporate a texture-aware injection module to explicitly integrate human-perceptible clothing texture information into the generation process. During inference, we propose a refining conditional inference strategy that mitigates the randomness introduced by Gaussian noise. This effectively preserves identity information and fine clothing details in the final results. Extensive experiments demonstrate that our method outperforms existing state-of-the-art virtual try-on approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 625, + 135, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 135, + 637 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 135, + 637 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 694 + ], + "type": "text", + "content": "Image-based virtual try-on (VTON) aims to realistically transfer a target garment onto the corresponding region of a reference person [3, 8, 14, 22, 39]. Owing to its strong application prospects and commercial potential, VTON has" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 242, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 289 + ], + "type": "text", + "content": "attracted growing research interest. Leveraging the powerful generative capabilities of diffusion models [20, 41], recent VTON methods have significantly narrowed the gap to real-world fitting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 292, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 556, + 579 + ], + "type": "text", + "content": "Previous VTON methods are broadly categorized into mask-based and mask-free paradigms. Mask-based approaches [3, 14, 22, 37] typically operate in a self-supervised fashion: a clothing-agnostic mask (Figure 2) is first applied to remove the garment region and adjacent body parts from the reference image, and the generator is then trained to inpaint the masked areas, thereby learning to perform the virtual try-on task. Although effective and widely adopted, these methods rely heavily on obtaining accurate masks for each reference person; even minor mask errors can introduce conspicuous artifacts [13, 21]. To mitigate mask-induced artifacts, mask-free methods grounded in pseudo-sample [8, 13, 18] have been proposed. A mask-based VTON model is first pre-trained to synthesize try-on images for arbitrary garments on the reference person, yielding paired pseudo-samples. These pseudo-samples are then fed as inputs, while the original reference image serves as the reconstruction target (ground truth), to train a main generator. Although this pipeline obviates the need for precise masks, they are limited to a single, fixed transformation for each pseudo-sample/reference-person pair. Moreover, because both garments and reference images are typically drawn from the same dataset, the generator overfits to these specific items and generalizes poorly to unseen clothing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 580, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 688 + ], + "type": "text", + "content": "There are also some other methods attempt to boost performance through architectural improvements—such as coarse-to-fine networks and ReferenceNet-based clothing-feature injection [22, 37, 40]. Yet these refinements remain constrained by the input materials (human parsing, etc), causing the generator to converge to sub-optimal solutions. Finally, the randomness of Gaussian-noise sampling during inference further degrades output quality [36]—a problem often overlooked by prior work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "To address these challenges, we propose a unified mask-free virtual try-on framework that leverages the intrinsic" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19525" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 239 + ], + "type": "text", + "content": "strengths of latent diffusion models to enable every pipeline component to better approximate the target data distribution and thus improve overall performance. Specifically, we introduce a text-driven pseudo-input preparation strategy to augment sample diversity by enriching the clothing regions in each pseudo-sample (see Figure 2). First, a pretrained Stable Diffusion inpainting model is conditioned on text prompts generated by a large language model (e.g., DeepSeek or GPT-series models) to produce multiple try-on results for the same identity. We then perform sample selection to filter out implausible images and retain the most realistic candidate as the final pseudo-sample. This process maximizes the variability of clothing regions and mitigates overfitting of the generator during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 294, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 382 + ], + "type": "text", + "content": "During training phase, we leverage the off-the-shelf inpainting try-on model as the generator, thereby reducing training costs. To convert this model into a mask-free generator, we introduce mask-free generation learning process. Concretely, a gated manipulation mechanism fuses clothing and person latent representations in latent space, preventing catastrophic forgetting when clothing conditioning is added and avoiding structural changes that inflate training cost. Complementarily, texture-aware injection module explicitly injects fine-grained texture cues into the cross-attention layers, sharpening the generator's perception of garment details." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 294, + 456 + ], + "type": "text", + "content": "During inference phase, to reduce the randomness of Gaussian noise, we propose a refining conditional inference in our optimization process. By optimizing the initial Gaussian noise to make it as close as possible to the distribution of real person samples, helping the generator improve the quality of the synthesized results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 294, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 294, + 477 + ], + "type": "text", + "content": "In summary, this paper makes the following main contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 479, + 294, + 609 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 479, + 294, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 514 + ], + "type": "text", + "content": "- We propose a unified mask-free virtual try-on framework, which improves the generator's performance from multiple aspects of both the training and inference processes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 515, + 294, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 538 + ], + "type": "text", + "content": "- We present a text-driven pseudo-input preparation approach to improve the generator's generalization ability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 539, + 294, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 539, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 294, + 562 + ], + "type": "text", + "content": "- We propose a gated manipulation to prevent weight forgetting and cut training costs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 563, + 294, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 294, + 586 + ], + "type": "text", + "content": "- We introduce a texture-aware injection to explicitly add human-perceptible clothing texture information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 586, + 294, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 294, + 609 + ], + "type": "text", + "content": "- We propose a refining conditional inference approach to reduce Gaussian noise randomness." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 621, + 141, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 141, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 141, + 633 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "Mask-Free Virtual Try-On. This task requires performing virtual try-on at inference time using only a clothing image and a person image—no masks, human parsing, or other auxiliary inputs. Issenhuth et al. [21] introduced the first mask-free method, WUTON, which attains fully supervised training by distilling knowledge from a mask-based teacher" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "that generates paired training samples. Because the synthetic images are treated as ground truth for the student, the teacher's defects directly bound the student's performance. Ge et al. [13] relaxed this constraint with PFAFN, feeding pseudo-samples as inputs instead of supervision, thereby allowing the student to learn a more accurate mapping. He et al. [18] extended this paradigm in Flow-Style, adding explicit garment-deformation optimization to improve fit and realism. More recently, Du et al. [8] proposed USC-PFN, a two-stage architecture that employs cycle-consistency and a weight-shared generator to refine the garment region and try-on result iteratively. In contrast to these prior works, we present a unified mask-free framework that optimizes the pseudo-sample and enhances both training and inference through multi-stage optimization, yielding higher fidelity and better generalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 279, + 553, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 494 + ], + "type": "text", + "content": "Controllable Diffusion Models. Recent advances in diffusion models demonstrate that generation can be guided by modalities beyond text. Specifically, the integration of external models to inject external features into diffusion models has produced remarkable results. For example, ControlNet [47] introduced an external model that leverages images, skeletons, and edge maps as conditional inputs to generate corresponding outputs. IP-adaptor [45] proposed a method for generating images similar to a given reference image. T2I-adapter [31] also introduced an external, more lightweight model that leverages conditional inputs to generate corresponding images. These approaches have effectively manipulated image generation by injecting feature values via external models. Instead of adding extra network branches, following the optimization idea of latent features in latent space of [9], we embed the try-on conditions directly into the latent space, yielding a high-performance virtual try-on model with minimal overhead." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 506, + 523, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 523, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 523, + 519 + ], + "type": "text", + "content": "3. Preliminary: Latent Diffusion Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": "Latent Diffusion Models (LDMs) [33] are a class of generative models that learn to synthesize data " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": " by denoising latent representations " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": " in a compressed space, where typically " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "h \\ll H" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "w \\ll W" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": ". The compression is achieved via a pre-trained autoencoder with KL-regularization (e.g. VQGAN), consisting of an encoder " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{VAE}(\\cdot)" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{VAE}(\\cdot)" + }, + { + "bbox": [ + 313, + 525, + 553, + 610 + ], + "type": "text", + "content": ", such that" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 620, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 620, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 328, + 620, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {0} = \\mathcal {E} _ {V A E} (\\mathbf {x}), \\quad \\hat {\\mathbf {x}} = \\mathcal {D} _ {V A E} (\\mathbf {z} _ {0}), \\quad \\text {w i t h} \\quad \\hat {\\mathbf {x}} \\approx \\mathbf {x}. \\tag {1}", + "image_path": "8185f02c43013e1c64f122ec3875b84bdcf25f6db7dfe68eeaf1d1b045026b9a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "text", + "content": "The latent-space diffusion process is modeled as a Markov chain that progressively adds Gaussian noise to " + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "text", + "content": " (the encoded data) over " + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "text", + "content": " steps, producing a sequence of noisy latents " + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 313, + 642, + 553, + 690 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 347, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {2}", + "image_path": "40b62da9ad465317d07cde9561745091d0c2e9bc02195e2fb5ee3387d8e679f3.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19526" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 255, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 255, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 255, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 255, + 177 + ], + "type": "image", + "image_path": "fb3df232a54430247be2ac1461589a8fc6ca3b98bc550ab2988cd9d38da90621.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 258, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 258, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 258, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "664bdaa56cd71151f8a51cab7f18780c1be152d062ae9027c480ef96a6052a09.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 177, + 256, + 231 + ], + "blocks": [ + { + "bbox": [ + 58, + 177, + 256, + 231 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 256, + 231 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 256, + 231 + ], + "type": "image", + "image_path": "fd003c139c0430488fc11d5746c5aeebb55e665dea8c9170bab02868069fc7fd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 232, + 250, + 241 + ], + "lines": [ + { + "bbox": [ + 63, + 232, + 250, + 241 + ], + "spans": [ + { + "bbox": [ + 63, + 232, + 250, + 241 + ], + "type": "text", + "content": "Sect. 4.1: Text-Driven Pseudo-Input Preparation" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 178, + 553, + 231 + ], + "blocks": [ + { + "bbox": [ + 258, + 178, + 553, + 231 + ], + "lines": [ + { + "bbox": [ + 258, + 178, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 258, + 178, + 553, + 231 + ], + "type": "image", + "image_path": "57c3c7b1cf6fe27721a00d9327d80656355970e8076e3e5194e1601fb384c5ec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 232, + 450, + 240 + ], + "lines": [ + { + "bbox": [ + 360, + 232, + 450, + 240 + ], + "spans": [ + { + "bbox": [ + 360, + 232, + 450, + 240 + ], + "type": "text", + "content": "Sect. 4.3: Optimization" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed virtual try-on framework. It consists of three sub-pipelines, namely Text-Driven Pseudo-Input Preparation, Mask-Free Generation Learning, and Optimization. In text-driven pseudo-input preparation (Sect. 4.1), we first use a pre-trained Stable Diffusion inpainting model to generate " + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "text", + "content": " different pseudo-samples " + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=0}^N" + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "text", + "content": " based on prompts generated by DeepSeek and various noises. Then, a sample selection method is proposed to pick the best one " + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 251, + 555, + 338 + ], + "type": "text", + "content": " pseudo-samples for the next step. In mask-free generation learning (Sect. 4.2), we design a mask-free training approach that initially uses gated manipulation to fuse clothing and person representations to prevent catastrophic forgetting, and then injects extracted clothing texture features through cross-attention modules. In optimization (Sect. 4.3), we first train a noise optimization MLP to obtain the optimized Gaussian noise that preserves as much identity and fine-grained clothing cues as possible, and then use it for the inference phase." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "inline_equation", + "content": "\\{\\beta_t\\in (0,1)\\}_{t = 1}^T" + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": " is a fixed variance schedule. The forward process can be marginalized in closed form as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 392, + 295, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 392, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 94, + 392, + 295, + 406 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {0}\\right) = \\mathcal {N} \\left(\\mathbf {z} _ {t}; \\sqrt {\\bar {\\alpha} _ {t}} \\mathbf {z} _ {0}, (1 - \\bar {\\alpha} _ {t}) \\mathbf {I}\\right), \\tag {3}", + "image_path": "53beac2e61cf28718814c2585ffefee8decbb88f44f6ad665a8a5117142bdc4f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 55, + 415, + 296, + 463 + ], + "type": "text", + "content": ". LDMs optimize a variational lower bound (VLB) on the log-likelihood of the data, simplified to the noise-prediction objective (DDPM-style):" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 95, + 471, + 294, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 471, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 95, + 471, + 294, + 490 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {\\mathbf {z} _ {0}, \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}), t} \\left[ \\| \\epsilon_ {\\theta} (\\mathbf {z} _ {t}, t) - \\epsilon \\| _ {2} ^ {2} \\right]. \\tag {4}", + "image_path": "e206cef3398c80bd65544ecbc8f9dfebdd550b0287859990ab7c39d57a59357e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "text", + "content": " is the ground-truth noise and " + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "text", + "content": " is a U-Net parameterized by " + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 498, + 296, + 522 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 138, + 547 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": "Overview. This task aims to transfer the target clothing from an arbitrary clothing image " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "C_{un} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": " onto a reference person image " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": ", producing a try-on result " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": ". To train the mask-free model, " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": " (generated from the off-the-shelf mask-based try-on method) is first used as pseudo-input, we construct triplets " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "\\{C, I, P\\}" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": " for fully supervised learning of the generator " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 552, + 295, + 648 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 91, + 655, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 655, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 91, + 655, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {\\boldsymbol {C}, \\boldsymbol {I}, \\boldsymbol {P}} \\ell_ {\\text {d i s t}} (G (\\boldsymbol {C}, \\boldsymbol {I}), \\boldsymbol {P}), \\tag {5}", + "image_path": "3f7b8ee4515f721b087b44e8a346fbdf5707b962c0bcc457b5a254f431b0198c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{dist}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represents a suitable loss function (e.g., L1, MSE) between the generated image " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "G(C,I)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and the" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": "ground-truth image " + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": ". The clothing image " + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": " and the clothes worn by " + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{P}" + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": " are identical and also sourced from the dataset. Sections that follow detail how we obtain a suitable pseudo-input " + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": " for efficient training of the generator " + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 358, + 555, + 456 + ], + "type": "text", + "content": ". An overview of our pipeline—comprising Text-Driven Pseudo-Input Preparation (Sect. 4.1), Mask-Free Generation Learning (Sect. 4.2), and Refining Conditional Inference (Sect. 4.3)—is provided in Figure 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 463, + 518, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 518, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 518, + 475 + ], + "type": "text", + "content": "4.1. Text-Driven Pseudo-Input Preparation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": "Previous mask-free methods [13, 18] rely on off-the-shelf try-on networks to generate the pseudo-input. Although training converges, the rigid one-to-one pairing of " + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "inline_equation", + "content": "C_{un} / I" + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": " drives the generator " + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": " toward sub-optimal solutions and limits generalization. Therefore, diversifying the clothing regions of " + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": " (see Figure 2) is crucial for improving model performance in mask-free try-ons. By increasing clothing variability, the generator " + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 480, + 555, + 589 + ], + "type": "text", + "content": " can focus more effectively on clothing features and details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": "Automatic Sample Generation. To increase the visual diversity of " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": " and mitigate the generalization drop caused by the overlap between " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "C_{un}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": " (for generating " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{I}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\pmb{C}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": ", we leverage a Stable Diffusion inpainting model [37], " + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{SD}}" + }, + { + "bbox": [ + 313, + 604, + 555, + 676 + ], + "type": "text", + "content": ", which possesses strong generative priors, together with the large language model DeepSeek2. The resulting pipeline is" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 703 + ], + "type": "text", + "content": "1https://civitai.com/models/120149/controlnet-for-densepose" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 703, + 451, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 451, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 451, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 703, + 451, + 712 + ], + "type": "text", + "content": "https://chat_deepseek.com" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19527" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "type": "text", + "content": "summarized in Equation (6)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 91, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 91, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 91, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {I} _ {i} \\right\\} _ {i = 0} ^ {N} = \\left\\{\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {i}\\right) \\right\\} _ {i = 0} ^ {N}, \\tag {6}", + "image_path": "9e1311a4e803b53e9e2df7f6c77a3cdf8d9171e42ce253353930a671a3c29cb9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "P^{\\star}" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": " denotes the clothing-agnostic person representation, consisting of the clothing-agnostic person image, the clothing-agnostic mask, and the DensePose map [15]. " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": " is a random Gaussian noise sample, " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\xi \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": " is a text prompt constructed from the template \"A photo of a person wearing {upper garment description}\", with the description generated by DeepSeek to cover a wide range of clothing styles. For each prompt, we synthesize " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "N = 5" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": " images and select the best try-on result " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": ". The resulting triplets " + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "inline_equation", + "content": "\\{C,I_{best},P\\}" + }, + { + "bbox": [ + 55, + 114, + 296, + 236 + ], + "type": "text", + "content": " are used for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "content": "Sample Selection. To this end, we adopt the Human Preference Score " + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\nu 2" + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "content": " (HPS v2) model [38] to rank the generated images " + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "content": " according to their alignment with the corresponding prompt. HPS v2 employs an image encoder " + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "inline_equation", + "content": "E_{IMG}(\\cdot)" + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "inline_equation", + "content": "E_{TEXT}(\\cdot)" + }, + { + "bbox": [ + 55, + 249, + 296, + 344 + ], + "type": "text", + "content": " to map images and prompts into a joint embedding space, where their cosine similarity serves as the preference score, as formulated in Equation (7)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 353, + 295, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 353, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 84, + 353, + 295, + 368 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} ^ {\\eta} (\\boldsymbol {I} _ {i}, O) = E _ {I M G} (\\boldsymbol {I} _ {i}) \\cdot E _ {T E X T} (O) \\times 1 0 0. \\tag {7}", + "image_path": "d7866a4067ceacb778038b41ff74646ba6d4e7f6de113e95af3792552dee0d00.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "text", + "content": "We select the image achieving the highest score as the final sample. However, some " + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "text", + "content": " while semantically consistent with the prompt (see column 5 of Figure 2)—exhibit implausible appearances in the inpainting regions, such as distorted limbs. To address this issue, before using HPS v2, we introduce an appearance rationality discriminator " + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_r" + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "text", + "content": " that outputs a rationality score " + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "inline_equation", + "content": "s^r(I_i) = \\mathcal{D}_r(I_i)" + }, + { + "bbox": [ + 55, + 375, + 296, + 483 + ], + "type": "text", + "content": " for each candidate image. This discriminator is trained by minimizing the objective given in Equation (8)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 490, + 295, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 490, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 65, + 490, + 295, + 535 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\mathcal {D} _ {r}} - \\mathbb {E} _ {\\boldsymbol {P}, O} \\left[ \\log \\left(\\mathcal {D} _ {r} \\left(\\mathcal {F} _ {S D} \\left(\\boldsymbol {P} ^ {\\star}, O, \\xi_ {\\text {b e s t}}\\right)\\right)\\right) \\right. \\tag {8} \\\\ \\left. + \\log \\left(1 - \\mathcal {D} _ {r} \\big (\\mathcal {F} _ {S D} (\\boldsymbol {P} ^ {\\star}, O, \\xi_ {w o r s t})\\right)\\right)\\left. \\right], \\\\ \\end{array}", + "image_path": "41a51afd90a6a0b97be734fa24dbb2beedc761c08e4e9564e33bca3456d223a5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{best}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " denotes the Gaussian noise that yields the best result under prompt " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\xi_{\\mathrm{worst}}\\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " yield the worst. Training samples were manually screened to identify the best and worst ones. We then sort the " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " scores " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\{s^r (I_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " and denote the two highest as " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "s_{(1)}^r" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "s_{(2)}^r" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": ". The acceptance probability for choosing between them by using HPS v2 is given by " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[s^{\\eta}(s_{(1)}^{r}),s^{\\eta}(s_{(2)}^{r})]" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "m = 0" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " as the filtering threshold and " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbb{I}[\\cdot ]" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": " is the indicator function [38, 49]. The sample satisfying this condition is retained and used to build the target triplet " + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\{C,I_{best},P\\}" + }, + { + "bbox": [ + 55, + 542, + 296, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 231, + 685 + ], + "type": "text", + "content": "4.2. Mask-Free Generation Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "In this stage, we leverage the triplet " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{C, I_{best}, P\\}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " to generate the final virtual try-on result " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{rec}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "P_{rec} \\approx P" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 148 + ], + "type": "image", + "image_path": "ea0dfecd90daf9807ddf00161efca625430b47b38d6b516c52c10ab6285eaf49.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 555, + 221 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 221 + ], + "type": "text", + "content": "Figure 2. Visualization of pseudo-samples. Under the same prompt, the pseudo-samples are generated from five different Gaussian noises. When a pseudo-sample is semantically aligned with the prompt, the semantic number is labeled on the image. The leftmost column shows the clothing-agnostic person and the clothing-agnostic mask." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": "Gated Manipulation. It remains challenging to organically fuse " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " so that the generator " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " can both preserve rich human details and synthesize realistic clothing, while suppressing artifacts. Existing approaches can be roughly divided into two groups: (a) Treat " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " as separate conditioning signals to " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " [22, 40]. (b) Crucely concatenate them in the pixel space before feeding the result to " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " [14]. Approach (a) demands an additional Clothing-ReferenceNet to inject the new condition via attention modules; otherwise, directly adding a new branch may trigger catastrophic forgetting of the pre-trained weights, significantly increasing training cost. Approach (b) suffers from detail loss and blurring caused by the naive pixel-level combination. To achieve high-quality try-on without extra heavy modules or training overhead, we propose gated manipulation. Concretely, we first encode " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " into latent features, then blend them through a learnable gate " + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "inline_equation", + "content": "\\kappa \\in [0,1]" + }, + { + "bbox": [ + 313, + 243, + 555, + 459 + ], + "type": "text", + "content": " in the feature space. The fused feature is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 368, + 468, + 553, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 468, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 368, + 468, + 553, + 482 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {B} = \\kappa \\odot \\mathbf {x} _ {C} + (1 - \\kappa) \\odot \\mathbf {x} _ {I}, \\tag {9}", + "image_path": "56369663b2e5061a057913e50924c45575ec27de12c3045ad57e64f67aeed198.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\kappa = \\varPhi_{gate}(\\mathbf{x}_I,\\mathbf{x}_C)" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "text", + "content": " denotes Hadamard product. " + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_I = \\mathcal{E}_{VAE}(\\mathbf{I}_{best})" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_C = \\mathcal{E}_{VAE}(\\hat{C})" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "text", + "content": " is the warped " + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\varPhi_{gate}(\\cdot)" + }, + { + "bbox": [ + 313, + 491, + 554, + 529 + ], + "type": "text", + "content": " is a gate generator pre-trained by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 538, + 553, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 538, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 364, + 538, + 553, + 557 + ], + "type": "interline_equation", + "content": "\\Phi_ {g a t e} ^ {*} = \\arg \\min _ {\\Phi_ {g a t e}} \\ell_ {d i s t} (\\mathbf {x} _ {P}, \\mathbf {x} _ {B}), \\tag {10}", + "image_path": "37c5587176e3d03b0cb3d85e4dfce69da13ec987afb78f324cf6417b63f0a4af.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_P = \\mathcal{E}_{VAE}(P)" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " participates in encoding rich semantic and contextual cues, the blended feature " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " obtained via Equation (9) effectively suppresses redundant clothing patterns from " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " and background clutter from " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": ". At the same time, it preserves the holistic understanding of the entire " + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 313, + 566, + 554, + 662 + ], + "type": "text", + "content": " (e.g., pose and identity), thereby mitigating catastrophic forgetting that typically occurs when the conditions for the new modality are fed to the generator." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Texture-aware Injection. To enhance the generator's sensitivity to clothing texture, we adopt a bank of Gabor filters [11]—oriented at " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "135^{\\circ}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "—to convolve" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19528" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": "the clothing " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ". Each filter produces an orientation-specific response; the maximum across these four responses is taken at every spatial location, yielding a texture map " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " that highlights dominant fabric patterns while suppressing noise. A domain-specific encoder " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\Psi(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " [32] then projects " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "C_g" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " into an intermediate latent representation " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\Psi(C_g)" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ", which is supplied to the cross-attention layer via learned key and value projections " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "K^{\\star} = W_k^{(i)}\\Psi(C_g), V^{\\star} = W_v^{(i)}\\Psi(C_g)" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ". So, the objective of this stage is defined as Equation (11):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 190, + 295, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 190, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 75, + 190, + 295, + 204 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m s e} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {I} _ {\\text {b e s t}}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {11}", + "image_path": "dc9dff663f1116d6b7cc4575da62e3ce76cb85d1711bbd11ed44a191b8cefe58.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 143, + 205, + 275, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 205, + 275, + 223 + ], + "spans": [ + { + "bbox": [ + 143, + 205, + 275, + 223 + ], + "type": "interline_equation", + "content": "\\left[ \\left\\| \\boldsymbol {\\epsilon} - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],", + "image_path": "f8e7702d5cd091b091b087aec32e70f8382e7250ba5a1949c87a0938894e11de.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t = \\sqrt{\\bar{\\alpha}_t}\\mathbf{x}_P + \\sqrt{1 - \\bar{\\alpha}_t\\epsilon}.\\| \\cdot \\| _2^2" + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "content": " denote the L2 norm. Considering that " + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "content": " is usually coarsely warped, in the early stages of training, we considering directing " + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "content": " 's focus towards clothing regions by preventing learning changes in other areas caused by " + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "inline_equation", + "content": "I_{best}" + }, + { + "bbox": [ + 55, + 228, + 296, + 314 + ], + "type": "text", + "content": "'s potential defects. To this end, we conduct source-to-source self-reconstruction for training, as formulated in Equation (12):" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 321, + 295, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 321, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 77, + 321, + 295, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\mathbb {E} _ {\\boldsymbol {C}, \\boldsymbol {P}, \\epsilon \\sim \\mathcal {N} (0, I), t \\sim U [ 0, T)} \\tag {12}", + "image_path": "04486b2b9d309e5ab01caab72dd1d5cf12cdabc0714b7c06560d41260e94713b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 140, + 335, + 274, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 335, + 274, + 354 + ], + "spans": [ + { + "bbox": [ + 140, + 335, + 274, + 354 + ], + "type": "interline_equation", + "content": "\\left[ \\left\\| \\epsilon - \\epsilon_ {\\theta} \\big (\\mathbf {z} _ {t}, \\mathbf {x} _ {B} ^ {\\prime}, \\boldsymbol {\\Psi} (\\boldsymbol {C} _ {g}), t \\big) \\right\\| _ {2} ^ {2} \\right],", + "image_path": "4b79292c5a5b8df7e4e3663ab6841061baf87d72a7f2b39d897b16a0afb2d896.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 359, + 216, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 216, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 216, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 216, + 373 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B^\\prime = \\kappa \\odot \\mathbf{x}_C + (1 - \\kappa)\\odot \\mathbf{x}_P" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 378, + 141, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 141, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 141, + 391 + ], + "type": "text", + "content": "4.3. Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 396, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 419 + ], + "type": "text", + "content": "Full Objective. The full objective functions to optimize " + }, + { + "bbox": [ + 55, + 396, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 396, + 296, + 419 + ], + "type": "text", + "content": " as Equation (13):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 427, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 427, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 129, + 427, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathcal {L} _ {m s e} + \\lambda \\cdot \\mathcal {L} _ {r e c}, \\tag {13}", + "image_path": "6139d73103d6306862ce5340083f0a9a1f7a95a70529abb5b47c949a2892d7ac.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 451, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 451, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 451, + 296, + 475 + ], + "type": "text", + "content": " is a hyper-parameter controlling relative importance and is set to 0.1 based on functionality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "text", + "content": "Refining Conditional Inference. During inference, traditional methods [3, 37, 40] sample noise " + }, + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "text", + "content": " from a standard Gaussian distribution " + }, + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 489, + 296, + 572 + ], + "type": "text", + "content": ", unrelated to the target try-on. This introduces significant initial randomness and fails to preserve identity information and clothing details in the results [36]. Inspired by [36, 49], we design a refining conditional inference to enhance the model's performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": "Given the latent representation " + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": " (Eq. (9) during training) and sample a random Gaussian noise " + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": ", we add " + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 573, + 296, + 609 + ], + "type": "text", + "content": ", as represented in Equation (14):" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 616, + 295, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 616, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 113, + 616, + 295, + 630 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {B} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {14}", + "image_path": "c430b7df4e454bd5985dd0d9c25adfa3fa0f9aba9e0133300585f2c1373a0b3a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 637, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 296, + 662 + ], + "type": "text", + "content": "To refine " + }, + { + "bbox": [ + 55, + 637, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_T" + }, + { + "bbox": [ + 55, + 637, + 296, + 662 + ], + "type": "text", + "content": ", we first decompose it using Singular Value Decomposition (SVD), which can be formulated in:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 667, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 667, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 116, + 667, + 295, + 682 + ], + "type": "interline_equation", + "content": "U \\times \\Sigma \\times V ^ {\\mathrm {T}} = S V D (\\hat {\\mathbf {x}} _ {T}), \\tag {15}", + "image_path": "a037199618d500827e1edd3722f892659f0bbacdb3b4fcd9c7a6363d98a0afa2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a diagonal matrix containing singular values, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " represents the left and right singular vectors," + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 322, + 75, + 550, + 118 + ], + "blocks": [ + { + "bbox": [ + 322, + 75, + 550, + 118 + ], + "lines": [ + { + "bbox": [ + 322, + 75, + 550, + 118 + ], + "spans": [ + { + "bbox": [ + 322, + 75, + 550, + 118 + ], + "type": "image", + "image_path": "0149006772c10981675f98ce38fe391c27f1a82df41b30ffb633f8374bdb013a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 321, + 129, + 550, + 163 + ], + "blocks": [ + { + "bbox": [ + 411, + 120, + 473, + 129 + ], + "lines": [ + { + "bbox": [ + 411, + 120, + 473, + 129 + ], + "spans": [ + { + "bbox": [ + 411, + 120, + 473, + 129 + ], + "type": "text", + "content": "Diffusion-noisy " + }, + { + "bbox": [ + 411, + 120, + 473, + 129 + ], + "type": "inline_equation", + "content": "\\hat{x}_T" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 129, + 550, + 163 + ], + "lines": [ + { + "bbox": [ + 321, + 129, + 550, + 163 + ], + "spans": [ + { + "bbox": [ + 321, + 129, + 550, + 163 + ], + "type": "image", + "image_path": "81762b678792ecfa4d50d1d6d02728d53508e2c3d8e2c7212fcfd6d469f95faa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 327, + 174, + 549, + 209 + ], + "blocks": [ + { + "bbox": [ + 419, + 167, + 466, + 174 + ], + "lines": [ + { + "bbox": [ + 419, + 167, + 466, + 174 + ], + "spans": [ + { + "bbox": [ + 419, + 167, + 466, + 174 + ], + "type": "text", + "content": "SVD denoised" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 327, + 174, + 549, + 209 + ], + "lines": [ + { + "bbox": [ + 327, + 174, + 549, + 209 + ], + "spans": [ + { + "bbox": [ + 327, + 174, + 549, + 209 + ], + "type": "image", + "image_path": "cf3f4e3fc7cd087df2458647f526f8a51d8f26a9cab07d782b1867c3fcaa613f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "type": "text", + "content": "Figure 3. Simulation comparison of denoising performance via SVD. From top to bottom: (row 1) clean signal " + }, + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "type": "text", + "content": ", (row 2) diffusion-noisy signal " + }, + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_T" + }, + { + "bbox": [ + 313, + 222, + 555, + 277 + ], + "type": "text", + "content": ", and (row 3) SVD-denoised reconstruction. SVD effectively suppresses the diffusion noise while preserving the dominant structural features of the original signal." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "text", + "content": "respectively. As shown in Figure 3, SVD can distinguish principal components from perturbations, captures key components of " + }, + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_T" + }, + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "text", + "content": " via singular values. To enhance semantic information, we utilize a multi-layer perceptron (MLP) with multi-head self-attention layers [34] to predict a new singular value " + }, + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "inline_equation", + "content": "\\widehat{\\Sigma}" + }, + { + "bbox": [ + 313, + 298, + 554, + 370 + ], + "type": "text", + "content": ", as represented in Equation (16):" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 324, + 377, + 554, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 377, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 324, + 377, + 554, + 393 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {T} ^ {\\prime} = \\varphi (U, \\Sigma , V ^ {\\mathrm {T}}), \\quad \\widehat {\\Sigma} = f (\\phi (\\hat {\\mathbf {x}} _ {T} ^ {\\prime}), \\mathbf {x} _ {B}) + \\Sigma , \\tag {16}", + "image_path": "6126e394c49f46ad4be8422dff3e7b4bdf4d694a10c27c71db37d210a659bf36.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\varphi (\\cdot)" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": " denotes the sum of all MLP-encoded inputs, " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": " represents the multi-head self-attention layer, and " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": " is another MLP. Thus, target representation " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": " is reformed by using " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\widehat{\\Sigma}" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "V^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": " to align with real representation " + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{GT}" + }, + { + "bbox": [ + 313, + 401, + 554, + 460 + ], + "type": "text", + "content": ", as formulated in Equation (17):" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 391, + 468, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 468, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 391, + 468, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {T} = U \\times \\widehat {\\boldsymbol {\\Sigma}} \\times V ^ {\\mathrm {T}}. \\tag {17}", + "image_path": "a3cbd481cfc71bdc6eadb2c1955bbb1bb961f76673d1ed1ec2162bb5ab08dc66.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 326, + 491, + 549, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 491, + 549, + 504 + ], + "spans": [ + { + "bbox": [ + 326, + 491, + 549, + 504 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 326, + 491, + 549, + 504 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_{GT}" + }, + { + "bbox": [ + 326, + 491, + 549, + 504 + ], + "type": "text", + "content": " can be obtained through Equation (18):" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 369, + 512, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 512, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 369, + 512, + 553, + 526 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {G T} = \\sqrt {\\bar {\\alpha} _ {T}} \\mathbf {x} _ {P} + \\sqrt {1 - \\bar {\\alpha} _ {T}} \\epsilon . \\tag {18}", + "image_path": "c91cd3823d88a3b1bd4102e9a00cdb6b2e068a81f23e0fdd235e15f095283593.jpg" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_{GT}" + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "text", + "content": ", we optimize " + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\{\\varphi ,\\phi ,f\\}" + }, + { + "bbox": [ + 313, + 534, + 554, + 559 + ], + "type": "text", + "content": " by minimizing Equation (19):" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 354, + 566, + 553, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 566, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 354, + 566, + 553, + 586 + ], + "type": "interline_equation", + "content": "\\min _ {\\varphi , \\phi , f} \\mathbb {E} _ {\\boldsymbol {C}, \\epsilon \\sim \\mathcal {N} (0, I), \\boldsymbol {P}} \\left[ \\| \\mathbf {x} _ {T} - \\tilde {\\boldsymbol {x}} _ {G T} \\| _ {2} ^ {2} \\right]. \\tag {19}", + "image_path": "a87fbd30cf0b9ae5189a0e4a3afaaca9cd1309aa8bd588aa45b1b50f60c88991.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "content": "As shown in Figure 1, during inference, given a person " + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_P" + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "content": " and a random clothing " + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{VAE}(C_{un}) = \\mathbf{x}_{C_{un}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B" + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "content": " is computed as: " + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_B = \\kappa \\odot \\mathbf{x}_{C_{un}} + (1 - \\kappa) \\odot \\mathbf{x}_P" + }, + { + "bbox": [ + 313, + 594, + 555, + 631 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 640, + 395, + 654 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 660, + 364, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 364, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 364, + 673 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Dataset. We conduct experiments on the VITON-HD [4], VITON [16], and DressCode [29] datasets. VITON dataset comprises 16,253 image pairs with a resolution of" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19529" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "256 \\times 192" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "14,221 / 2,032" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": " training/testing pairs. VITON-HD dataset is composed of 11,647/2,032 training/testing pairs with a high-resolution of " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "512 \\times 384" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ". DressCode dataset is another high-resolution dataset with a resolution of " + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "inline_equation", + "content": "512 \\times 384" + }, + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": ", which comprises 15,363 image pairs, 12,863/2,500 training/testing pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": "Implementation Setting. Our method is implemented with PyTorch on a single NVIDIA A100 GPU. Follow the configuration of [6, 14], Stable Diffusion " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\mathrm{V}1.5^{3}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " is used as our base model. We train for 20 epochs using the AdamW optimizer [27] with a base learning rate of " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " and a batch size of 2, and use the warm-up adjustment strategy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 234, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 234, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 234, + 251 + ], + "type": "text", + "content": "5.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 413 + ], + "type": "text", + "content": "Evaluation Metrics. We adopt Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS) [48], Fréchet Inception Distance (FID) [19], and Kernel Inception Distance (KID) [2], for quantitative evaluation. SSIM assesses pixel-level structural fidelity between generated and real images. LPIPS measures perceptual distance via deep features trained on human judgments; lower scores indicate higher perceptual similarity. FID computes the Wasserstein-2 distance between Inception-V3 feature distributions of real and generated images; lower is better. KID estimates distributional similarity through a maximum-mean-discrepancy kernel on the same Inception features; again, lower values denote higher fidelity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "type": "text", + "content": "Baseline Methods. To conduct objective quantitative experiments, we employed the following methods:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 452, + 295, + 571 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 452, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 523 + ], + "type": "text", + "content": "- GAN-based Methods: CP-VTON [35], Cloth-flow [17], CP-VTON+ [28], VTNFP [46], ACGPN [42], DCTON [12], PF-AFN [13], ZFlow [7], OVNet [24], LM-VTON [26], DAFlow [1], Style-Flow [18], RT-VTON [43], DressCode [29], VITON-HD [4], HR-VITON [23], POVNet [25], GP-VTON [39], and USC-PFN [8]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 523, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 571 + ], + "type": "text", + "content": "- Diffusion-based Methods: LaDI-VTON [30], PbE [41], DCI-VTON [14], StableVITON [22], OOTDiffusion [40], Anydoor [3], TPD [44], IDM-VTON [5], StableGarment [37], and LDE-VTON [10]." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 696 + ], + "type": "text", + "content": "Quantitative Comparison. Figure 4 visualizes our results on the VITON-HD dataset [4]. GAN-based approaches—GP-VTON [39] and USC-PFN [8], for example—lack the human-centric priors inherent to diffusion models. Consequently, their outputs are overly smooth, contain visible artifacts, and struggle to disambiguate crossed arms. In contrast, our method synthesizes more realistic garment wrinkles and correctly reconstructs cross-arm poses. Diffusion model-based methods, such as DCI-" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 553, + 295 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 295 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 295 + ], + "type": "table", + "html": "
MethodsVenueMask\\( SSIM_p \\)↑\\( FID_{up} \\)↓
VITONCVPR'18✓0.7455.71
CP-VTON [35]ECCV'18✓0.7224.45
VTNFP [46]ICCV'19✓0.80n/a
Cloth-flow [17]CVPR'19✓0.8414.43
CP-VTON+ [28]CVPRW'20✓0.7521.04
ACGPN [42]CVPR'20✓0.8416.64
LM-VTON [26]AAAI'21✓0.8517.18
DCTON [12]CVPR'21✓0.8314.82
ZFlow [7]ICCV'21✓0.8815.17
OVNet [24]CVPR'21✓0.8515.78
PF-AFN [13]CVPR'21✘0.8910.21
RT-VTON [43]CVPR'22✓n/a11.66
DAFlow [1]ECCV'22✓0.8812.05
DressCode [29]CVPR'22✓0.8913.71
POVNet [25]TPAMI'23✓0.8913.37
PbE [41]CVPR'23✓0.8312.56
USC-PFN [8]NeurIPS'23✘0.9110.47
TPD [44]CVPR'24✓0.899.58
LDE-VTON [10]AAAI'25✘0.919.86
OursThis Work✘0.919.52
", + "image_path": "31cd9159376fc1e90ee8b076a89799c7ecc3c5c8632734335020d4751bcba67f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 303, + 555, + 337 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 337 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on the VITON dataset. \"Mask\" denotes whether the mask is used during inference. Bold denotes the best result. Some data are from published literature." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "text", + "content": "VTON [14], StableVITON [22], StableGarment [37], LDEVTON [10], and OOTDiffusion [40], limited by the diversity of training samples, learn fixed sample transformations. Consequently, they often fail to generate natural and realistic arms when dealing with unusual poses and clothing (e.g., the " + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "text", + "content": " row). Additionally, since the randomness of Gaussian noise during inference is not addressed, the results often lose identity information (e.g., the belt in the " + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "text", + "content": " row) and clothing details (e.g., garment shape and texture in the " + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 313, + 358, + 556, + 529 + ], + "type": "text", + "content": " row). In contrast, our approach addresses these issues by (1) augmenting training sample diversity to improve robustness, (2) enhancing the model architecture for higher synthesis quality, and (3) optimizing the handling of Gaussian noise to preserve identity and garment details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Comparison. Table 1 and Table 2 present quantitative comparisons between our method and the baselines on the VITON [16], VITON-HD [4], and DressCode [29] datasets. Our approach attains the highest scores on VITON dataset, demonstrating its robustness to low-resolution inputs. On VITON-HD and DressCode datasets, it ranks either first or second. The occasional second place arises because most baselines are trained on a fixed set of clothing images, leading to overfitting between specific garments and reference identities. In contrast, we enlarge appearance diversity by synthesizing additional pseudosamples and introducing out-of-distribution information; this strategy slightly penalizes performance on test cases that closely match the original data distribution." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 286, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 286, + 714 + ], + "type": "text", + "content": "3https://huggingface.co/stable-diffusion-v1-5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19530" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 553, + 271 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 271 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 271 + ], + "type": "image", + "image_path": "268735c1a8ca0f803de09a0358ce7cd5da65f7dfe71966478f3b8537e696523c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 314 + ], + "type": "text", + "content": "Figure 4. Qualitative results with different baseline methods on the VITON-HD dataset. The left area shows a visual comparison with GAN-based baseline methods, while the right area shows a comparison with diffusion model-based baseline methods. It can be seen that our method achieves the best visual performance. Red dashed box represents defects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 323, + 552, + 497 + ], + "blocks": [ + { + "bbox": [ + 59, + 323, + 552, + 497 + ], + "lines": [ + { + "bbox": [ + 59, + 323, + 552, + 497 + ], + "spans": [ + { + "bbox": [ + 59, + 323, + 552, + 497 + ], + "type": "table", + "html": "
Train / Test MethodsVenueVITON-HDDressCode UpperMask
\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)\\(SSIM_p\\uparrow\\)\\(LPIPS_p\\downarrow\\)\\(FID_{up}\\downarrow\\)\\(KID_{up}\\downarrow\\)
VITON-HD [4]CVPR'210.8620.11712.1173.23n/an/an/an/a✓
HR-VITON [23]ECCV'220.8780.10511.2652.730.9360.06513.822.71✓
GP-VTON [39]CVPR'230.8840.0819.7011.260.7690.27020.118.17✓
LaDI-VTON [30]MM'230.8640.0969.4801.990.9150.06314.263.33✓
PbE [41]CVPR'230.8020.14311.9393.850.8970.07815.334.64✓
DCI-VTON [14]MM'230.8800.0809.4101.100.9370.04211.921.89✓
StableVITON [22]CVPR'240.8640.0849.4651.400.9110.05011.270.72✓
StableGarment [37]arXiv'240.8030.10417.1158.85n/an/an/an/a✓
IDM-VTON [5]ECCV'240.8500.0609.8421.120.8800.0569.544.32✓
AnyDoor [3]CVPR'240.8210.09910.8502.460.8990.11914.833.05✓
LDE-VTON [10]AAAI'250.8980.0819.6401.21n/an/an/an/a✓
OursThis Work0.9010.0799.3841.120.9390.04810.951.78✘
", + "image_path": "0dda832b872b5e34fd3e5557d639e367fe77e81541f6cc4d4f4ce6a491c09b7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 555, + 529 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons on the VITON-HD and DressCode datasets. For LPIPS, FID, and KID, the lower the better. For SSIM, the higher the better. Underline represents second best. Some data are from published literature." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 548, + 149, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 149, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 149, + 561 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 295, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 578 + ], + "type": "text", + "content": "Effectiveness of Text-Driven Pseudo-Input Preparation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": "Figure 5 (A) and Table 3 (A) highlight the necessity of text-driven pseudo-input preparation (TDPIP). By increasing the appearance diversity of pseudo-samples, it enhances the generator's generalization ability for abnormal samples. Without this component, the arm regions of abnormal samples are poorly generated, and the clothing is not properly fitted. In contrast, with this preparation, the arms are well generated, and the clothing is effectively transferred." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Effectiveness of Gated Manipulation. Figure 6 demonstrates the necessity of introducing gated manipulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 549, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 555, + 610 + ], + "type": "text", + "content": "(GM) to prevent catastrophic forgetting. It can be seen that when we remove GM from the baseline method, the generator with altered input structure requires nearly 75 epochs to converge. In contrast, when GM is introduced, the model achieves good convergence after only 10 epochs of training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness of Texture-aware Injection. Figure 5 (B) and Table 3 (B) demonstrate the necessity of introducing texture-aware injection (TAI) method for capturing clothing texture details. It can be seen that when TAI is introduced into the baseline generator, the text on the target clothing image is faithfully reconstructed. In contrast, when it is removed, some text details are lost by the generator." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19531" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 555, + 191 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 555, + 191 + ], + "type": "image", + "image_path": "1f9ef0e66ab130b2a0fccf4a6dc30004bcb0d0d4d0a0a9d3ae0af489fdc09ee2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "type": "text", + "content": "Figure 5. Visual ablation study of different components in our approach. (A) Text-Driven Pseudo-Input Preparation (TDPIP). (B) Texture-aware Injection (TAI). (C) Refining Conditional Inference (RCI). The blue dashed box represents the area of focus for comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 240, + 298, + 315 + ], + "blocks": [ + { + "bbox": [ + 56, + 240, + 298, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 240, + 298, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 298, + 315 + ], + "type": "table", + "html": "
ConfigurationSSIMp↑FIDup↓
Baseline Framework0.8312.56
+ (A) Text-Driven Pseudo-Input0.879.86
+ (B) Texture-aware Injection0.899.61
+ (C) Refining Conditional Inference0.919.52
", + "image_path": "8a73c714aa450c95dd65cf8d8fa12a4b73456f8c445bd1e4ed8af9c402ac7636.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 356 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 356 + ], + "type": "text", + "content": "Table 3. Ablation study. Performance of various configurations on VITON dataset. Baseline is a basic teacher-student architecture with PbE [41]. Bold denotes the best result." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 377, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 486 + ], + "type": "text", + "content": "Effectiveness of Refining Conditional Inference. Finally, Figure 5 (C) and Table 3 (C) demonstrate the necessity of the proposed refining conditional inference (RCI) for information preservation and reducing generation randomness. When this initial noise generation mode is introduced into the framework, the clothing details in the try-on results are well reconstructed, and there is no loss of identity information. However, when this initial noise generation mode is removed, some clothing details are lost to varying degrees." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 133, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 133, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 133, + 506 + ], + "type": "text", + "content": "5.4. Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 513, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 620 + ], + "type": "text", + "content": "Although our proposed method can generate high-quality try-on results within a mask-free framework, the target noise optimized by the proposed refining conditional inference approach is estimated from only a limited set of samples. This scarcity can introduce instability and, at the beginning of inference, may perturb the initial latent distribution. In future work, we will investigate more robust and data-driven adversarial augmentation strategies for target-noise generation to eliminate this limitation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 633, + 128, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 128, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 128, + 645 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we present a unified, mask-free virtual try-on framework that boosts generation quality by increasing training sample diversity and improving both training and inference processes. First, we introduce a text-driven pseudo-input preparation strategy to enlarge data" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 242, + 437, + 348 + ], + "blocks": [ + { + "bbox": [ + 316, + 242, + 437, + 348 + ], + "lines": [ + { + "bbox": [ + 316, + 242, + 437, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 437, + 348 + ], + "type": "image", + "image_path": "3f43ec90fa996478dbaeaa1b2c2ef42e31198cc77722edd2783e4c2d1852067e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 392 + ], + "type": "text", + "content": "Figure 6. Ablation study of the Gated Manipulation. It can be seen that with the use of the gated manipulation, the generator quickly converges to a better state in the early stages of training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 438, + 243, + 551, + 348 + ], + "blocks": [ + { + "bbox": [ + 438, + 243, + 551, + 348 + ], + "lines": [ + { + "bbox": [ + 438, + 243, + 551, + 348 + ], + "spans": [ + { + "bbox": [ + 438, + 243, + 551, + 348 + ], + "type": "image", + "image_path": "e6cfc4e440a7ae9cc46db2ac5e8daa2573bf4561a0a840179845bad30a591812.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 413, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 534 + ], + "type": "text", + "content": "diversity and enhance model's generalization. Second, a gated-manipulation mechanism prevents catastrophic forgetting of pre-trained model weights while cutting training costs. Third, texture-aware injection explicitly supplements human-perceptible garment details. Finally, a refining conditional-inference scheme suppresses Gaussian-noise randomness to preserve identity and fine-grained clothing cues during inference. Extensive experiments on multiple benchmarks show that our proposed method consistently outperforms previous virtual try-on work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 548, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 556, + 704 + ], + "type": "text", + "content": "Acknowledgments. This work was in part supported by the National Key Research and Development Program of China (Grant No. 2022ZD0160604), the National Natural Science Foundation of China (NSFC, Grant No. 62176194), the Key Research and Development Program of Hubei Province (Grant No. 2023BAB083), the Project of Sanya Yazhou Bay Science and Technology City (Grant Nos. SCKJ-JYRC-2022-76, SKJC-2022-PTDX-031), the Project of Sanya Science and Education Innovation Park of Wuhan University of Technology (Grant No. 2021KF0031), and the Huawei Kunpeng-Ascend Innovation Incentive Programme. This work was also supported in part by both computing resources from the Wuhan Supercomputing Center and the Wuhan Artificial Intelligence Computing Center." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19532" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Shuai Bai, Huiling Zhou, Zhikang Li, Chang Zhou, and Hongxia Yang. Single stage virtual try-on via deformable attention flows. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XV, pages 409-425. Springer, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 178 + ], + "type": "text", + "content": "[2] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 234 + ], + "type": "text", + "content": "[3] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6593-6602, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 290 + ], + "type": "text", + "content": "[4] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14131-14140, 2021. 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 335 + ], + "type": "text", + "content": "[5] Yisol Choi, Sangkyung Kwak, Kyungmin Lee, Hyungwon Choi, and Jinwoo Shin. Improving diffusion models for authentic virtual try-on in the wild. In European Conference on Computer Vision, pages 206-235. Springer, 2024. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "text", + "content": "[6] Zheng Chong, Xiao Dong, Haoxiang Li, Shiyue Zhang, Wenqing Zhang, Xujie Zhang, Hanqing Zhao, Dongmei Jiang, and Xiaodan Liang. Catvton: Concatenation is all you need for virtual try-on with diffusion models. arXiv preprint arXiv:2407.15886, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 445 + ], + "type": "text", + "content": "[7] Ayush Chopra, Rishabh Jain, Mayur Hemani, and Balaji Krishnamurthy. Zflow: Gated appearance flow-based virtual try-on with 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5433-5442, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 490 + ], + "type": "text", + "content": "[8] Chenghu Du, Shuqing Liu, Shengwu Xiong, et al. Greatness in simplicity: Unified self-cycle consistency for parser-free virtual try-on. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 535 + ], + "type": "text", + "content": "[9] Chenghu Du, Junyin Wang, Kai Liu, and Shengwu Xiong. Glv: Geometric correlation distillation for latent diffusion-enhanced parser-free virtual try-on. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Chenghu Du, Junyin Wang, Feng Yu, and Shengwu Xiong. Latent diffusion-enhanced virtual try-on via optimized pseudo-label generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2780-2788, 2025. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] D. Gabor. Theory of communication. part 1: The analysis of information. Journal of the Institution of Electrical Engineers - Part III: Radio and Communication Engineering, 93: 429-441, 1946. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Chongjian Ge, Yibing Song, Yuying Ge, Han Yang, Wei Liu, and Ping Luo. Disentangled cycle consistency for highly-realistic virtual try-on. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16928-16937, 2021. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 690, + 295, + 714 + ], + "type": "text", + "content": "[13] Yuying Ge, Yibing Song, Ruimao Zhang, Chongjian Ge, Wei Liu, and Ping Luo. Parser-free virtual try-on via distilling" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "appearance flows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8485-8493, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[14] Junhong Gou, Siyu Sun, Jianfu Zhang, Jianlou Si, Chen Qian, and Liqing Zhang. Taming the power of diffusion models for high-quality virtual try-on with appearance flow. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7599-7607, 2023. 1, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[15] Ríza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densespose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297-7306, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[16] Xintong Han, Zuxuan Wu, Zhe Wu, Ruichi Yu, and Larry S. Davis. Viton: An image-based virtual try-on network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7543-7552, 2018. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "text", + "content": "[17] Xintong Han, Xiaojun Hu, Weilin Huang, and Matthew R Scott. Clothflow: A flow-based model for clothed person generation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10471-10480, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 354 + ], + "type": "text", + "content": "[18] Sen He, Yi-Zhe Song, and Tao Xiang. Style-based global appearance flow for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3470-3479, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[20] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion model-based image editing: A survey. arXiv preprint arXiv:2402.17525, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 511 + ], + "type": "text", + "content": "[21] Thibaut Issenhuth, Jérémie Mary, and Clément Calauzenes. Do not mask what you do not need to mask: a parser-free virtual try-on. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 619-635. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 567 + ], + "type": "text", + "content": "[22] Jeongho Kim, Guojung Gu, Minho Park, Sunghyun Park, and Jaegul Choo. Stableviton: Learning semantic correspondence with latent diffusion model for virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8176-8185, 2024. 1, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 634 + ], + "type": "text", + "content": "[23] Sangyun Lee, Gyojung Gu, Sunghyun Park, Seunghwan Choi, and Jaegul Choo. High-resolution virtual try-on with misalignment and occlusion-handled conditions. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVII, pages 204-219. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[24] Kedan Li, Min Jin Chong, Jeffrey Zhang, and Jingen Liu. Toward accurate and realistic outfits visualization with attention to details. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15546-15555, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Kedan Li, Jeffrey Zhang, and David Forsyth. Povnet: Image-based virtual try-on through accurate warping and residual." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19533" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12222-12235, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 139 + ], + "type": "text", + "content": "[26] Guoqiang Liu, Dan Song, Ruofeng Tong, and Min Tang. Toward realistic virtual try-on through landmark guided shape matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2118-2126, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 294, + 161 + ], + "type": "text", + "content": "[27] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 205 + ], + "type": "text", + "content": "[28] Matiur Rahman Minar, Thai Thanh Tuan, Heejune Ahn, Paul Rosin, and Yu-Kun Lai. Cp-vton+: Clothing shape and texture preserving image-based virtual try-on. In CVPR Workshops, pages 10-14, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 206, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 295, + 261 + ], + "type": "text", + "content": "[29] Davide Morelli, Matteo Fincato, Marcella Cornia, Federico Landi, Fabio Cesari, and Rita Cucchiara. Dress code: High-resolution multi-category virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2231-2235, 2022. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 316 + ], + "type": "text", + "content": "[30] Davide Morelli, Alberto Baldrati, Giuseppe Cartella, Marcelia Cornia, Marco Bertini, and Rita Cucchiara. Ladi-vton: Latent diffusion textual-inversion enhanced virtual try-on. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8580-8589, 2023. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 371 + ], + "type": "text", + "content": "[31] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI conference on artificial intelligence, pages 4296–4304, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 372, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 372, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 372, + 295, + 437 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 492 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 295, + 536 + ], + "type": "text", + "content": "[34] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "text", + "content": "[35] Bochao Wang, Huabin Zheng, Xiaodan Liang, Yimin Chen, Liang Lin, and Meng Yang. Toward characteristic-preserving image-based virtual try-on network. In Proceedings of the European conference on computer vision (ECCV), pages 589-604, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 593, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 295, + 668 + ], + "type": "text", + "content": "[36] Chenhui Wang, Tao Chen, Zhihao Chen, Zhizhong Huang, Taoran Jiang, Qi Wang, and Hongming Shan. Fldm-vton: Faithful latent diffusion model for virtual try-on. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, IJCAI-24, pages 1362–1370. International Joint Conferences on Artificial Intelligence Organization, 2024. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[37] Rui Wang, Hailong Guo, Jiaming Liu, Huaxia Li, Haibo Zhao, Xu Tang, Yao Hu, Hao Tang, and Peipei Li. Stablegarment: Garment-centric generation via stable diffusion. arXiv preprint arXiv:2403.10783, 2024. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 708 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 126 + ], + "type": "text", + "content": "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "text", + "content": "[39] Zhenyu Xie, Zaiyu Huang, Xin Dong, Fuwei Zhao, Haoye Dong, Xijin Zhang, Feida Zhu, and Xiaodan Liang. Gp-vton: Towards general purpose virtual try-on via collaborative local-flow global-parsing learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23550-23559, 2023. 1, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 239 + ], + "type": "text", + "content": "[40] Yuhao Xu, Tao Gu, Weifeng Chen, and Arlene Chen. Ootdiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8996-9004, 2025. 1, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 305 + ], + "type": "text", + "content": "[41] Binxin Yang, Shuyang Gu, Bo Zhang, Ting Zhang, Xuejin Chen, Xiaoyan Sun, Dong Chen, and Fang Wen. Paint by example: Exemplar-based image editing with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18381-18391, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[42] Han Yang, Ruimao Zhang, Xiaobao Guo, Wei Liu, Wangmeng Zuo, and Ping Luo. Towards photo-realistic virtual try-on by adaptively generating-preserving image content. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7850-7859, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 555, + 407 + ], + "type": "text", + "content": "[43] Han Yang, Xinrui Yu, and Ziwei Liu. Full-range virtual try-on with recurrent tri-level transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3460-3469, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "type": "text", + "content": "[44] Xu Yang, Changxing Ding, Zhibin Hong, Junhao Huang, Jin Tao, and Xiangmin Xu. Texture-preserving diffusion models for high-fidelity virtual try-on. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7017-7026, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 506 + ], + "type": "text", + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 563 + ], + "type": "text", + "content": "[46] Ruiyun Yu, Xiaqi Wang, and Xiaohui Xie. Vtnfp: An image-based virtual try-on network with body and clothing feature preservation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10511-10520, 2019. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 564, + 553, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 564, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 316, + 564, + 553, + 608 + ], + "type": "text", + "content": "[47] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 609, + 553, + 663 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 663 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 665, + 553, + 708 + ], + "type": "text", + "content": "[49]Zikai Zhou,Shitong Shao,Lichen Bai,Zhiqiang Xu,Bo Han,and Zeke Xie. Golden noise for diffusion models: A learning framework.arXiv preprint arXiv:2411.09502, 2024.4,5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "19534" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_content_list.json b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35cc0b2d0025cc8c21e4e6f8b98db0e3e17fb743 --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_content_list.json @@ -0,0 +1,1532 @@ +[ + { + "type": "text", + "text": "All in One: Visual-Description-Guided Unified Point Cloud Segmentation", + "text_level": 1, + "bbox": [ + 127, + 130, + 869, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zongyan Han $^{1}$ , Mohamed El Amine Boudjoghra $^{2}$ , Jiahua Dong $^{1}$ , Jinhong Wang $^{1}$ , Rao Muhammad Anwer $^{1}$", + "bbox": [ + 238, + 179, + 756, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE \n $^{2}$ Technical University of Munich, Germany", + "bbox": [ + 197, + 215, + 800, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zongyan.han, jiahua.dong, jinhong.wang, rao.anwer}@mbzuai.ac.ae", + "bbox": [ + 215, + 253, + 781, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohamed.boudjoghra@tum.de", + "bbox": [ + 382, + 272, + 607, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unified segmentation of 3D point clouds is crucial for scene understanding, but is hindered by its sparse structure, limited annotations, and the challenge of distinguishing fine-grained object classes in complex environments. Existing methods often struggle to capture rich semantic and contextual information due to limited supervision and a lack of diverse multimodal cues, leading to suboptimal differentiation of classes and instances. To address these challenges, we propose VDG-Uni3DSeg, a novel framework that integrates pre-trained vision-language models (e.g., CLIP) and large language models (LLMs) to enhance 3D segmentation. By leveraging LLM-generated textual descriptions and reference images from the internet, our method incorporates rich multimodal cues, facilitating fine-grained class and instance separation. We further design a Semantic-Visual Contrastive Loss to align point features with multimodal queries and a Spatial Enhanced Module to model scene-wide relationships efficiently. Operating within a closed-set paradigm that utilizes multimodal knowledge generated offline, VDG-Uni3DSeg achieves state-of-the-art results in semantic, instance, and panoptic segmentation, offering a scalable and practical solution for 3D understanding. Our code is available at https://github.com/Hanzy1996/VDG-Uni3DSeg.", + "bbox": [ + 88, + 356, + 483, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 752, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unified 3D point cloud segmentation, comprising semantic, instance, and panoptic segmentation, has become a fundamental component of tasks such as autonomous driving, robotics, and AR/VR applications. However, progress in this field remains challenging because of the unique characteristics of point cloud data and the scarcity of annotated datasets. Unlike 2D images, which benefit from rich contextual cues such as texture and lighting, point clouds are", + "bbox": [ + 89, + 779, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/591cb123da4da34894a16bcebe0da4c5e102ddb346a0548ad9f11e9b3fa52b05.jpg", + "image_caption": [ + "(a) Existing Multi-Modal 3D Segmentation Methods" + ], + "image_footnote": [], + "bbox": [ + 517, + 319, + 903, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2e4ee34d5352b081f5518638c460f1cdd3da1f7388b21c77d18d710c8991184.jpg", + "image_caption": [ + "Figure 1. Comparison between existing multi-modal 3D segmentation methods and our proposed approach. (a) Existing methods depend on paired images and captioning models to generate descriptions, requiring real-time 2D-3D alignment, which increases computational overhead. (b) Our approach leverages pregenerated class descriptions from LLMs and reference images retrieved from the internet, establishing offline semantic references that enhance efficiency and robustness." + ], + "image_footnote": [], + "bbox": [ + 517, + 435, + 901, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "inherently sparse and unstructured, making manual annotation labor-intensive and time-consuming. These limitations have slowed progress in 3D segmentation, especially compared to the strides achieved in 2D segmentation, where large-scale datasets have significantly enhanced model performance. Although methods like OneFormer3D [27] have achieved notable success in unified 3D segmentation, they often struggle to capture fine-grained semantic details, thus limiting their effectiveness in complex 3D scenes.", + "bbox": [ + 511, + 714, + 906, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these limitations, we propose a novel multimodal framework that leverages external knowledge from pretrained large language models (LLMs) [47] and vision-", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "language models (CLIP [38]) to integrate diverse cross-modal knowledge into the 3D segmentation pipeline. These pretrained models aim to provide semantically rich textual and visual embeddings, enriching the segmentation process with fine-grained contextual information. By incorporating multimodal knowledge, our approach improves the ability of the 3D segmentation model to capture fine-grained attributes crucial to distinguishing classes and instances in complex scenes, while also minimizing reliance on extensive human annotation to provide a more efficient and scalable solution for understanding 3D point clouds.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our method first prompts an LLM to generate detailed semantic descriptions of 3D classes, capturing key attributes such as color, texture, and shape. These descriptions offer a broad understanding of the characteristics of the class, forming a strong semantic foundation. To further enhance visual context, we incorporate class-specific reference images sourced from the internet, which introduce diverse real-world variations to improve the robustness of object recognition. To effectively integrate these multi-modal cues, we encode both textual descriptions and reference images into query representations via CLIP [38]. These queries are fused with point-wise features within a unified mask decoder, allowing the model to produce semantic, instance, and panoptic segmentation masks. Additionally, we introduce a Semantic-Visual Contrastive (SVC) Loss, which aligns point features with their corresponding textual and visual embeddings, thereby enhancing class distinction and segmentation accuracy. To further capture spatial relationships and refine object differentiation, we incorporate a Spatial Enhancement Module. This module utilizes sparse attention to efficiently model scene-wide spatial dependencies, ensuring structural coherence in segmentation while maintaining computational efficiency.", + "bbox": [ + 91, + 260, + 483, + 607 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, unlike open-vocabulary methods [14, 54] that require real-time fusion of 3D scenes with paired images and text captions, our approach follows a closed-set paradigm with predefined class knowledge. We establish offline semantic references through offline-generated class descriptions and reference images (Fig. 1). During inference, our method operates without paired images, captioning models, or additional vision-language modules, enhancing efficiency and deployment practicality. By integrating offline multimodal knowledge with spatially aware feature learning, our approach achieves superior class separation and instance delineation compared to conventional closed-set 3D segmentation baselines. Both quantitative and qualitative comparisons demonstrate consistent improvements in our model performance in 3D instance, semantic, and panoptic segmentation tasks.", + "bbox": [ + 89, + 609, + 482, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 853, + 434, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce VDG-Uni3DSeg, a novel framework that integrates pretrained vision-language models and LLMs", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to enhance unified 3D segmentation.", + "bbox": [ + 527, + 90, + 771, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our method synergizes LLM-generated descriptions and unpaired reference images, enriching class representations with both generalized attributes and fine-grained visual details, thus improving class and instance differentiation.", + "- We develop an efficient mask decoder that integrates textual and visual embeddings as queries, enhancing segmentation accuracy across 3D instance, semantic, and panoptic tasks." + ], + "bbox": [ + 513, + 106, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 257, + 663, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Task specific Unimodal 3D Segmentation", + "text_level": 1, + "bbox": [ + 511, + 282, + 861, + 299 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous techniques for 3D segmentation adopt point-based or voxel-based approaches to process point clouds. Point-based methods use hand-crafted aggregation mechanisms [32, 36, 37, 46] or transformer blocks [52, 60] for direct point processing, while voxel-based methods convert point clouds into sparse voxel representations, which are then processed using dense [19] or sparse [11] 3D convolutional neural networks. Instance segmentation typically involves first performing semantic segmentation followed by per-point feature aggregation. Recent transformer-based models [42] integrate sparse CNN backbones for feature extraction with transformer decoders.", + "bbox": [ + 511, + 306, + 906, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The 3D instance segmentation task aims at predicting masks for individual objects in a 3D scene, along with a class label belonging to the set of known classes. Some methods use a bottom-up approach based on grouping learning embeddings in the latent space to facilitate clustering of object points [4, 16, 17, 23, 28, 30, 50, 59]. On the other hand, proposal-based methods adopt a top-down strategy, initially detecting 3D bounding boxes and then segmenting the object region within each box [15, 18, 33, 56, 58]. Inspired by advances in 2D work [6, 8], transformer-based architectures [48] have recently been introduced for 3D instance segmentation [1, 20, 26, 41, 43]. Mask3D [41] introduces a hybrid architecture that uses a 3D CNN backbone for the extraction of per-point characteristics and a transformer-based instance mask decoder to refine a set of queries. Building on Mask3D, [1] shows that explicit spatial and semantic supervision at the level of the 3D backbone further improves instance segmentation.", + "bbox": [ + 511, + 488, + 908, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For 3D panoptic segmentation, existing methods [34, 51, 57] perform panoptic segmentation on RGB images, lift 2D panoptic masks into 3D space, and aggregate them to generate final 3D panoptic masks.", + "bbox": [ + 511, + 760, + 906, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unified Unimodal segmentation", + "text_level": 1, + "bbox": [ + 513, + 832, + 792, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, several approaches for 2D unified segmentation have been proposed, including MaskFormer [5] and Mask2Former [7]. These methods use a transformer-based", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24836", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "architecture to predict masks for three different segmentation tasks simultaneously. OneFormer [21] builds on the same transformer design, but introduces task-conditioned queries, which help the model focus on a specific segmentation task during prediction.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the success of OneFormer [21] in 2D, a recent method OneFormer3D [27] introduces a novel multitask unified framework that simultaneously handles 3D semantic, instance, and panoptic segmentation within a single model. Unlike previous approaches, OneFormer3D unifies these tasks through semantic and instance queries in parallel within a transformer decoder, addressing the performance instability by incorporating a new query selection mechanism and an efficient matching strategy, eliminating the need for the Hungarian algorithm, which makes the training process much faster. This approach is trained once on a panoptic dataset, allowing it to outperform existing methods that are specifically tuned for each segmentation task. Although previous work has shown promising results, it still lacks essential multimodal information. In our method, we demonstrate how incorporating this information can significantly improve performance.", + "bbox": [ + 91, + 167, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 435, + 225, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce our proposed method, Visual-Description-Guided Unified 3D Segmentation (VDG-Uni3DSeg). First, we describe the 3D backbone in Sec. 3.1, which extracts point-wise features from the input point cloud. Next, we detail the multi-modal reference queries in Sec. 3.2, which integrate textual and visual information to enhance segmentation performance. We then present the mask decoder in Sec. 3.3, designed to generate task-specific masks for semantic, instance, and panoptic segmentation. Finally, we describe the loss functions used to train the proposed model in Sec. 3.4.", + "bbox": [ + 89, + 460, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Backbone", + "text_level": 1, + "bbox": [ + 89, + 633, + 230, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input to our model is a 3D scene represented as a point cloud $\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}$ containing $N$ points. Each point consists of three spatial coordinates and three color values. To expedite the training process and improve computational efficiency, we adopt the voxelization approach from [10], which reduces the point cloud to $M$ superpoints with $M < N$ , allowing the model to focus on meaningful spatial clusters. Subsequently, these superpoints are processed by a sparse 3D U-Net to generate point-wise features, $\\mathbf{X} \\in \\mathbb{R}^{M \\times d}$ , where $d$ denotes the point embedding dimensionality.", + "bbox": [ + 89, + 656, + 482, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Enhancement Module To capture the spatial relationships between point-wise features in a 3D scene, we adopt a sparse attention mechanism [9] that integrates contextual information from a randomly sampled subset of the", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "entire point cloud. Specifically, the point-wise features $\\mathbf{X}$ are first transformed into query, key, and value embeddings:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 137, + 903, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_q, \\mathbf{W}_k,$ and $\\mathbf{W}_v \\in \\mathbb{R}^{d \\times d}$ are learnable weight matrices. For each point feature $\\mathbf{x}_i$ , attention is calculated using a randomly sampled subset $S_i$ . The spatially enhanced feature is then computed as:", + "bbox": [ + 511, + 160, + 903, + 220 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {x}} _ {i} = \\sum_ {\\mathbf {v} _ {j} \\in \\mathcal {S} _ {i}} \\alpha_ {i, j} \\cdot \\mathbf {v} _ {j}, \\quad \\alpha_ {i, j} = \\frac {\\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}{\\sum_ {\\mathbf {k} _ {j} \\in \\mathcal {S} _ {i}} \\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 232, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $S_{i}$ denotes a randomly sampled subset of points from the point cloud, and $\\mathbf{q}_i,\\mathbf{k}_j,\\mathbf{v}_j\\in \\mathbb{R}^{1\\times d}$ . Using the subset $S_{i}$ , the module captures the spatial structure of the overall scene while significantly reducing computational costs. This makes it highly efficient and well-suited for large-scale 3D point cloud processing.", + "bbox": [ + 511, + 287, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Modal Reference Queries", + "text_level": 1, + "bbox": [ + 511, + 388, + 794, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach uses a unified 3D mask decoder to perform semantic, instance, and panoptic segmentation simultaneously. Specifically, we build on the Transformer-based mask decoder framework proposed in [21]. For each segmentation task, we incorporate task-specific queries that interact with the point-wise features to produce the corresponding masks. For instance segmentation, we use instance queries as defined in [21], denoted as $\\mathbf{Q}_{ins} \\in \\mathbb{R}^{K \\times d}$ . For semantic segmentation, we introduce multimodal semantic queries to generate masks as described below.", + "bbox": [ + 511, + 411, + 903, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Description Reference Queries To capture fine-grained details for accurate class distinction, we first introduce the LLM-based Description Generator (LLM-DesGen), a pipeline designed to generate descriptive visual details specific to the class. LLM-DesGen is guided by pre-defined description types and example visual descriptions for given classes. In this in-context learning manner, it can be generalized to unseen classes by learning the descriptive patterns. For each candidate class, LLM-DesGen generates $K$ descriptions for the given class, as follows:", + "bbox": [ + 511, + 563, + 906, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} ^ {c} = \\left[ t _ {1} ^ {c}, \\dots , t _ {K} ^ {c} \\right] = \\text {L L M - D e s G e n} (c), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 726, + 903, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{T}^c$ represents the descriptions of the $c$ -th class, and $t_k^c$ denotes the $k$ -th description of the $c$ -th class. Finally, we can get the set of descriptions for all classes: $\\mathbf{T} = [\\mathbf{T}^1, \\dots, \\mathbf{T}^C]$ . Next, we utilize a text encoder (i.e., CLIP [38]) to embed these descriptions into description embeddings:", + "bbox": [ + 511, + 755, + 903, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {t} = \\operatorname {C L I P - T} (\\mathbf {T}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 845, + 903, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_t = [\\mathbf{Q}_t^1,\\dots ,\\mathbf{Q}_t^C ]$ , where $\\mathbf{Q}_t^c\\in \\mathbb{R}^{K\\times d}$ denotes the $K$ queries for class $c$", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24837", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7600f259afdfada17427f6723840a75d86165e527ec3c53b33dc483a29bddb5f.jpg", + "image_caption": [ + "Figure 2. Illustration of the VDG-Uni3DSeg framework. First, we utilize a 3D encoder to extract point-wise features from the input point cloud. Then, the Spatial Enhancement Module refines these features by capturing local spatial relationships. Next, class-specific description queries are generated using LLMs, while image queries are constructed from reference images retrieved from the internet. Both description and image queries are embedded using CLIP. These queries, along with the enhanced point features, are processed through a multi-modal fusion block to generate task-specific masks. The description-based and image-based predictions are then fused to achieve robust semantic segmentation. Finally, instance and semantic segmentation masks are integrated to produce panoptic segmentation results." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 898, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Reference Queries To complement the class descriptions, we collect a set of reference images for each class by collecting images from the internet, denoted $\\mathbf{O} = [\\mathbf{O}^{1},\\dots,\\mathbf{O}^{C}]$ , where $\\mathbf{O}^c = [o_1^c,\\dots,o_L^c]$ consists of $L$ images representing the $c$ -th class. We embed these images using the CLIP image encoder [38] to generate visual embeddings as follows:", + "bbox": [ + 89, + 465, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {o} = \\operatorname {C L I P - I} (\\mathbf {O}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 583, + 482, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_o\\in \\mathbb{R}^{C\\times L\\times d}$ . Unlike multiview images in standard training datasets, these internet-sourced images provide diverse representations with broader variations, reducing reliance on extensive annotations.", + "bbox": [ + 89, + 609, + 483, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Uni3DSeg, description queries provide high-level semantic information, capturing the general appearance and characteristics of each class, while visual queries offer fine-grained visual details.", + "bbox": [ + 89, + 670, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Modal Unified Mask Decoder", + "text_level": 1, + "bbox": [ + 89, + 739, + 401, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the approach in [27], we first utilize $B$ fusion layers to align and blend the information from the point cloud, image, and description domains, enhancing the model's capacity for accurate multimodal mask prediction. Concretely, we apply cross-attention to integrate 3D information into our task-specific queries. Then, we use a self-attention mechanism and a feed-forward network to combine the cross-modal information within these queries. To preserve the original semantic and visual information, we incorporate a residual structure into the description and im", + "bbox": [ + 89, + 760, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age reference queries, as shown in Fig. 2. We then feed these refined queries and the spatially enhanced point features $\\widetilde{\\mathbf{X}}$ into the main mask decoder $D$ to generate the corresponding segmentation masks for each query:", + "bbox": [ + 511, + 465, + 906, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {i n s} = D \\left(\\mathbf {Q} _ {i n s}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {t} = D \\left(\\mathbf {Q} _ {t}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {o} = D \\left(\\mathbf {Q} _ {o}, \\tilde {\\mathbf {X}}\\right), (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 536, + 906, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{ins} \\in \\mathbb{R}^{N \\times S}$ is the predicted instance segmentation mask, $\\mathbf{M}_t \\in \\mathbb{R}^{N \\times C \\times K}$ and $\\mathbf{M}_o \\in \\mathbb{R}^{N \\times C \\times L}$ are the semantic segmentation masks predicted by the description and image queries, respectively.", + "bbox": [ + 511, + 566, + 905, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Modal Segmentation Ensemble To achieve a robust semantic segmentation result, we adopt an ensemble strategy that combines the semantic segmentation outputs of $\\mathbf{M}_d$ and $\\mathbf{M}_o$ . Concretely, for each point feature $\\mathbf{x}_m$ , the description-based semantic prediction $\\mathbf{M}_d^m \\in \\mathbb{R}^{C \\times K}$ represents the similarity scores for all descriptions for each class. To determine the best-matching description for each point, we apply max pooling on $\\mathbf{M}_d^m$ to obtain the refined description-based prediction $\\widehat{\\mathbf{M}}_d^m \\in \\mathbb{R}^C$ :", + "bbox": [ + 511, + 628, + 906, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathbf {M}} _ {t} ^ {m} = \\max _ {k} \\left(\\mathbf {M} _ {t} ^ {m}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 782, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we obtain the refined image prediction $\\widehat{\\mathbf{M}}_o^m$ . By combining description-based and image-based predictions, the ensemble approach produces a robust semantic segmentation result that leverages both high-level semantic and fine-grained visual details. The final semantic segmentation", + "bbox": [ + 511, + 823, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24838", + "bbox": [ + 478, + 989, + 517, + 1000 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "prediction is computed as:", + "bbox": [ + 89, + 90, + 267, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {s e m} ^ {m} = \\widehat {\\mathbf {M}} _ {t} ^ {m} + \\widehat {\\mathbf {M}} _ {o} ^ {m}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 114, + 482, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{sem}^{m}\\in \\mathbb{R}^{C}$ is the semantic prediction of the $m$ -th point feature. The final result of semantic segmentation is denoted as $\\mathbf{M}_{sem}\\in \\mathbb{R}^{M\\times C}$ .", + "bbox": [ + 89, + 143, + 482, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Losses", + "text_level": 1, + "bbox": [ + 89, + 198, + 248, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [21], we use a combination of cross-entropy, binary cross-entropy, and Dice losses to optimize the instance mask predictions. For semantic segmentation, in addition to cross-entropy loss, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with both description embeddings $\\mathbf{Q}_d$ and image embeddings $\\mathbf{Q}_o$ . In contrast to traditional contrastive losses that only align features with a single-modality embedding, the SVC loss promotes multimodal alignment, encouraging spatially enhanced point features $\\widetilde{\\mathbf{X}}$ to be close to both description embeddings $\\mathbf{Q}_d$ and image embeddings $\\mathbf{Q}_o$ of the correct class while pushing them farther from the embeddings of other classes. The SVC loss is defined as follows:", + "bbox": [ + 89, + 218, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S V C}} = \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {t}) + \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {o}),\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 407, + 313, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q}) = - \\mathbb {E} _ {(\\tilde {\\mathbf {x}} _ {i}, \\mathbf {q} ^ {+})} \\left[ \\log \\frac {\\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} ^ {+}\\right) / \\tau\\right)}{\\sum_ {n} \\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} _ {n}\\right) / \\tau\\right)} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 425, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_{+}$ denotes the positive description or image embedding corresponding to the same class as the point feature $\\tilde{\\mathbf{x}}_i$ , $\\mathrm{sim}(\\cdot)$ denotes the inner product similarity, and $\\tau$ is the temperature parameter, empirically set to 1.0. This loss aligns point features with class-specific embeddings, enhancing both semantic consistency and class discriminability. It also facilitates better integration of information from reference descriptions and images, contributing to a more robust segmentation.", + "bbox": [ + 89, + 465, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function for our method is defined as:", + "bbox": [ + 109, + 601, + 468, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a l l} = \\underbrace {\\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} + \\lambda_ {1} \\mathcal {L} _ {c e} ^ {i n s}} _ {\\mathcal {L} _ {i n s}} + \\underbrace {\\lambda_ {2} \\mathcal {L} _ {c e} ^ {s e m} + \\lambda_ {3} \\mathcal {L} _ {S V C}} _ {\\mathcal {L} _ {s e m}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 625, + 480, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1},\\lambda_{2}$ , and $\\lambda_{3}$ are hyperparameters that balance different loss functions.", + "bbox": [ + 89, + 675, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 718, + 223, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 742, + 290, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets We evaluate our method on three widely used datasets: S3DIS [2], ScanNet [13], and ScanNet200 [39]. The S3DIS dataset [2] comprises 272 scenes spread over six areas, with annotations for 13 semantic classes. Among these, five furniture categories are defined as thing classes, while the remaining eight are stuff classes for panoptic evaluation. Following the standard evaluation protocol, we evaluate our method on Area-5 and conduct six-fold cross-validation. ScanNet dataset [13] consists of 1,201 scans", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for training, 312 scans for validation, and 100 scans for testing. The 3D instance segmentation task includes 18 object classes, while the semantic and panoptic segmentation tasks incorporate two additional background classes. ScanNet200 dataset [39] extends ScanNet [13] with finer-grained annotations, containing 198 object classes for instance segmentation and two additional background classes. It follows the same data split as ScanNet [13].", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics For 3D semantic segmentation, we evaluate performance based on the mean Intersection over Union (mIoU). For 3D instance segmentation, we compute the mean Average Precision (mAP) across IoU thresholds ranging from $50\\%$ to $95\\%$ , with increments of $5\\%$ . Furthermore, mAP50 and mAP25 represent mAP values calculated at IoU thresholds of $50\\%$ and $25\\%$ , respectively. For S3DIS, we additionally report mean precision (mPrec) and mean recall (mRec) according to the standard evaluation protocol. 3D panoptic segmentation is evaluated using the Panoptic Quality (PQ) score [24]. We also provide separate evaluations for the categories of things and stuff, denoted $\\mathrm{PQ}_{\\mathrm{th}}$ and $\\mathrm{PQ}_{\\mathrm{st}}$ , respectively.", + "bbox": [ + 511, + 213, + 906, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We implement our method using the MMDetection3D framework [12], following OneFormer3D [27]. The training setup utilizes AdamW optimizer with an initial learning rate of 0.0001, a weight decay of 0.05, a batch size of 4, and a polynomial learning rate scheduler with a power of 0.9. We apply standard data augmentations, including horizontal flipping, random rotations along the z-axis, elastic distortion, and random scaling. For each class, we generate 10 descriptions using LLaMA 3.1 [47] and collect reference images from Bing. We then extract the description and image embeddings using CLIP [38]. These embeddings are further projected into the 3D feature space as queries via a multi-layer perceptron.", + "bbox": [ + 511, + 411, + 908, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 619, + 813, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method against state-of-the-art (SOTA) approaches on three datasets. Table 1 presents comparisons on the S3DIS [2] dataset. Across both the Area-5 and 6-fold cross-validation benchmarks, our method demonstrates strong competitive performance compared to SOTA methods. Notably, on the Area-5 benchmark, our method achieves significant improvements in instance segmentation, outperforming previous methods by a large margin. Specifically, VDG-Uni3DSeg achieves a 2.1-point gain in $\\mathrm{MAP}_{50}$ and a 1.4-point improvement in mAP over the best-performing baseline. For panoptic segmentation, our method surpasses SOTA methods with a 4.1-point gain in PQ and a remarkable 9.6-point improvement in $\\mathrm{PQ}_{\\mathrm{th}}$ . Additionally, for semantic segmentation, VDG-Uni3DSeg achieves a 1.7-point improvement in mIoU over our baseline, OneFormer3D [27]. On the 6-fold cross-validation benchmark, VDG-Uni3DSeg achieves similarly notable im", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24839", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f606d13043b0779fa46c0a3c615abb884ccbf705640aeb1865f89fc72c9170b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInstanceSemantic mIoUPQPanoptic PQthPQst
mAP50mAPmPrec50mRec50
Area-5 validation
PointGroup[22]57.861.962.1
DyCo3D[17]64.364.2
SSTNet[31]59.342.765.564.2
DKNet[53]70.865.3
HAIS[3]71.165.0
TD3D[25]65.148.674.464.8
SoftGroup[49]66.151.673.666.6
PBNet[61]66.453.574.965.4
SPFormer[44]66.872.867.1
Mask3D[42]71.957.874.363.7
SegGCN[29]63.6
MinkUNet[11]65.4
PAConv[55]66.6
KPConv[46]67.1
PointTransformer[60]70.4
PointNeXt-XL[37]70.5
PointTransformerV2[52]71.6
OneFormer3D [27]72.058.779.773.069.862.258.465.5
VDG-Uni3DSeg (Ours)74.160.181.073.971.566.368.065.3
6-fold cross-validation
PointGroup[22]64.069.669.2
HAIS[3]73.269.4
SSTNet[31]67.854.173.573.4
DKNet[53]75.371.1
TD3D[25]68.256.276.374.0
SoftGroup[49]68.954.475.369.8
SPFormer[44]69.274.071.1
PBNet[61]70.659.580.172.9
Mask3D[42]74.361.876.566.2
PointNet++[36]56.7
MinkUNet[11]69.1
KPConv[46]70.6
PointTransformer[60]73.5
PointNeXt-XL[37]74.9
OneFormer3D [27]72.860.482.372.772.868.561.574.5
VDG-Uni3DSeg (Ours)75.863.785.574.873.269.168.972.7
", + "bbox": [ + 215, + 87, + 782, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of existing segmentation methods on S3DIS. Our proposed VDG-Uni3DSeg shows significant performance improvement across instance, semantic, and panoptic segmentation tasks.", + "bbox": [ + 89, + 592, + 906, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provements across all metrics, demonstrating its robustness and effectiveness in different settings.", + "bbox": [ + 89, + 647, + 482, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents comparisons on the ScanNet [13] dataset, including results for both the validation split and the hidden test split. In the validation split, our method consistently improves performance across all segmentation tasks, including instance, semantic, and panoptic segmentation. Specifically, for instance segmentation, VDG-Uni3DSeg achieves a 1.3-point improvement in $\\mathrm{mAP}_{50}$ , demonstrating its enhanced ability to detect and distinguish individual objects. For semantic segmentation, our method achieves a 0.4-point improvement in mIoU, highlighting its improved capability to classify points accurately across the scene. Furthermore, for panoptic segmentation, VDG-Uni3DSeg also achieves improvement in PQ, demonstrating its effectiveness in jointly optimizing instance and semantic pre", + "bbox": [ + 88, + 688, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dictions. Additionally, VDG-Uni3DSeg also achieves improvements in the ScanNet hidden test set, further highlighting its strong generalization ability. Table 3 presents the results for ScanNet200 [39], which introduces fine-grained classes. VDG-Uni3DSeg achieves state-of-the-art performance in all instances, semantic, and panoptic segmentation, demonstrating its robustness in complex segmentation tasks. Although our model achieves slightly lower semantic mIoU, it remains competitive and maintains balanced performance across all tasks.", + "bbox": [ + 511, + 647, + 906, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These consistent improvements across different datasets highlight the effectiveness of using both visual and description-based references to guide the segmentation process. By integrating multi-modal information, our method enhances the understanding of objects in 3D scenes, leading to significant performance gains.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24840", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47383b30fc197f55e80ccd05d3fcfb6fa057d1cc10b6c9a29073f0450accbb77.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPresented atmAP25Instance mAP50mAPSemantic mIoUPQPanoptic PQthPQst
Validation split
NeuralBF[45]WACV'2371.155.536.0
DyCo3D[17]CVPR'2172.957.635.4
SSTNet[31]ICCV'2174.064.349.4
HAIS[3]ICCV'2175.664.443.5
DKNet[53]ICCV'2276.966.750.8
SoftGroup[49]CVPR'2278.967.645.8
PBNet[61]ICCV'2378.970.554.3
TD3D[25]WACV'2481.971.247.3
ISBNet[35]CVPR'2382.573.154.5
SPFormer[44]AAAI'2382.973.956.3
Mask3D[42]ICRA'2383.573.755.2
PointTransformer[60]ICCV'2170.6
PointNeXt-XL[37]NeurIPS'2271.5
PointMetaBase-XXL[32]CVPR'2372.8
PointTransformerV2[52]NeurIPS'2275.4
SceneGraphFusion[51]CVPR'2131.530.243.4
PanopticFusion[34]IROS'1933.530.858.4
TUPPer-Map[57]IROS'2150.247.871.5
OneFormer3D [27]CVPR'2486.177.258.575.871.370.586.1
VDG-Uni3DSeg (Ours)86.578.559.376.271.570.086.6
Hidden test split at 03 Mar: 2025
NeuralBF[45]WACV'2371.855.535.3
DyCo3D[17]CVPR'2176.164.139.5
PointGroup[22]CVPR'2077.863.640.7
SSTNet[31]ICCV'2178.969.850.6
HAIS[3]ICCV'2180.369.945.7
DKNet[53]ICCV'2281.571.853.2
ISBNet[35]CVPR'2383.575.755.9
SPFormer[44]AAAI'2385.177.054.9
SoftGroup[49]CVPR'2286.576.150.4
Mask3D[42]ICRA'2387.078.056.6
TD3D[25]WACV'2487.575.148.9
OneFormer3D [27]CVPR'2489.680.156.6
VDG-Uni3DSeg (Ours)88.080.457.6
", + "bbox": [ + 197, + 87, + 802, + 556 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ac62d8189ec140fdf814eef035b7eb386abf6fd3dcfea42df7a76781968228f.jpg", + "table_caption": [ + "Table 2. Comparison of existing segmentation methods on ScanNet. Our proposed VDG-Uni3DSeg achieves superior performance across all metrics for instance, semantic, and panoptic segmentation tasks." + ], + "table_footnote": [], + "table_body": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
PointGroup[22]24.5
PointGroup + LGround[40]26.1
TD3D[25]40.434.823.1
Mask3D[42]42.337.027.4
MinkUNet[11]25.0
MinkUNet + LGround[40]28.9
OneFormer3D [27]44.240.029.029.031.030.578.6
VDG-Uni3DSeg (Ours)45.140.029.529.731.330.977.9
", + "bbox": [ + 236, + 608, + 761, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of existing segmentation methods on the ScanNet200 validation split. Our VDG-Uni3DSeg shows competitive performance across instance, semantic, and panoptic segmentation tasks.", + "bbox": [ + 88, + 762, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 806, + 243, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 presents an ablation study evaluating the contribution of each component in our VDG-Uni3DSeg framework on S3DIS. First, removing the Spatial Enhancement Module slightly degrades the performance, particularly in instance and panoptic segmentation metrics (e.g., a 2-point", + "bbox": [ + 89, + 825, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "decrease in mAP and a 1.7-point drop in PQ). This underscores the importance of spatial information in effectively distinguishing instances within the scene. Next, we analyze the impact of reference queries. Removing either the Description Query or the Image Query leads to a significant performance drop, with a more pronounced decline when", + "bbox": [ + 511, + 806, + 908, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/18f8266bf58e657e0d60c744f0556929d3073d1dbef361b1f7d0c9b28b43633b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
baseline79.572.058.769.862.258.465.5
VDG-Uni3DSeg80.674.160.171.566.368.065.3
W/O Spatial Enhancement80.073.358.170.664.663.865.7
W/O Description Query78.472.459.670.264.868.962.2
W/O Image Query79.772.760.070.465.164.365.6
W/O SVC Loss78.072.558.369.062.261.163.8
", + "bbox": [ + 236, + 88, + 761, + 213 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study evaluating each component of our VDG-Uni3DSeg. 'W/O' indicates that the corresponding part of our VDG-Uni3DSeg is removed.", + "bbox": [ + 89, + 223, + 903, + 251 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89e9fe500a265b758ebec91f6bd1016a01ef5e438c96814b1d6921d723c6786c.jpg", + "image_caption": [ + "(a) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 102, + 261, + 256, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/40b2246015b8dea520969fa7548281312f6d2bb32297429a195bcd06aa2e3dd3.jpg", + "image_caption": [ + "(b) OneFormer3D", + "Figure 3. Qualitative results of semantic segmentation on ScanNet. Compared to OneFormer3D, the Description and Image Prediction modules improve segmentation accuracy, particularly for tables. By integrating both description and image queries, VDG-Uni3DSeg achieves more balanced and consistent segmentation, enhancing the semantic segmentation for small and challenging objects such as doors and chairs." + ], + "image_footnote": [], + "bbox": [ + 259, + 261, + 416, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/61ec7e75c39743691d59428c83b4c599b6f020d35d2fa5ea4909dd7794a1c51d.jpg", + "image_caption": [ + "(c) Description Prediction" + ], + "image_footnote": [], + "bbox": [ + 418, + 261, + 573, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64a45f970f34d1bdf44bcc39c4a1be3df4b1f861169b99177054f42e1e9fb524.jpg", + "image_caption": [ + "(d) Image Prediction" + ], + "image_footnote": [], + "bbox": [ + 576, + 261, + 730, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fefbcbbcf228e20e0e41c347524ef003e0947c2c1a491e8303dfef32b023f28.jpg", + "image_caption": [ + "(e)VDG-Uni3DSeg" + ], + "image_footnote": [], + "bbox": [ + 733, + 261, + 888, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the Description Query is removed. This demonstrates the benefit of detailed descriptions generated by LLMs in our method. Furthermore, incorporating reference images helps bridge the gap between textual descriptions and the 3D point cloud. By integrating description and image queries, our method enables each point to associate with the most relevant class representation, leading to overall improvements across all segmentation tasks. Finally, removing the SVC Loss causes the most significant drop in instance and panoptic segmentation metrics, including a 2.6-point decline in $\\mathrm{mAP}_{25}$ and a 6.9-point drop in $\\mathrm{PQ_{th}}$ . This demonstrates that SVC Loss encourage the model to learn more discriminative segmentation features by leveraging multimodal queries as anchors for feature alignment. In general, each component of our method contributes to improving feature alignment and overall segmentation performance.", + "bbox": [ + 88, + 435, + 485, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 694, + 272, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the semantic segmentation results of different methods. Compared to OneFormer3D [27], both the Description Prediction and Image Prediction modules improve segmentation accuracy by leveraging reference queries, with notable improvements in identifying tables. However, these individual modules struggle with segmenting complex objects such as chairs and doors. When combining description and image predictions, as in VDG-Uni3DSeg, the segmentation results exhibit more balanced and comprehensive improvements across all object categories. This integration results in more precise and consistent segmentation, demonstrating that while individual queries contribute to perfor", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance gains, their combined use in VDG-Uni3DSeg yields the most accurate and robust results. Other results are provided in the supplementary material.", + "bbox": [ + 511, + 435, + 905, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 496, + 633, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present VDG-Uni3DSeg, a unified 3D segmentation framework that integrates LLM-generated descriptions and internet-sourced images as reference cues to enhance segmentation performance. By leveraging multimodal information, our method improves class and instance differentiation without requiring human annotations or paired multimodal data. To further refine segmentation, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with multimodal queries to enhance feature distinctiveness. Additionally, our Spatial Enhancement Module effectively captures long-range spatial relationships, ensuring precise segmentation boundaries while balancing computational cost. Extensive experiments on multiple datasets demonstrate that VDG-Uni3DSeg achieves state-of-the-art performance across semantic, instance, and panoptic segmentation tasks. These results highlight the effectiveness of multimodal guidance in 3D segmentation and pave the way for future research in leveraging external knowledge to further improve 3D scene understanding.", + "bbox": [ + 509, + 522, + 906, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 828, + 684, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to express our sincere gratitude to all the reviewers for their valuable and constructive suggestions, which have significantly improved the quality of this work.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24842", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Salwa Al Khatib, Mohamed El Amine Boudjoghra, Jean Lahoud, and Fahad Shahbaz Khan. 3d instance segmentation via enhanced spatial and semantic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 541-550, 2023. 2", + "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 5", + "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 6, 7", + "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 2", + "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2", + "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2", + "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2", + "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2", + "[9] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 3", + "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3", + "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 6, 7", + "[12] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6", + "[14] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2", + "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2", + "[16] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 2", + "[17] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 2, 6, 7", + "[18] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2", + "[19] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2", + "[20] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[21] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 3, 5", + "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 6, 7", + "[23] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2", + "[24] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5", + "[25] Maksim Kolodiazhnyi, Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Top-down beats bottom-up in 3d" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24843", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "instance segmentation. arXiv preprint arXiv:2302.02871, 2023.6,7", + "[26] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[27] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 1, 3, 4, 5, 6, 7, 8", + "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 2", + "[29] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11611-11620, 2020. 6", + "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 2", + "[31] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6, 7", + "[32] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17682-17691, 2023. 2, 7", + "[33] Shih-Hung Liu, Shang-Yi Yu, Shao-Chi Wu, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning gaussian instance segmentation in point clouds. arXiv preprint arXiv:2007.09860, 2020. 2", + "[34] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4205-4212. IEEE, 2019. 2, 7", + "[35] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 7", + "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6", + "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 2, 6, 7", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5", + "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 5, 6", + "[40] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 7", + "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In International Conference on Robotics and Automation (ICRA), 2023. 2", + "[42] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6, 7", + "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2", + "[44] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 6, 7", + "[45] Weiwei Sun, Daniel Rebain, Renjie Liao, Vladimir Tankovich, Soroosh Yazdani, Kwang Moo Yi, and Andrea Tagliasacchi. Neuralbf: Neural bilateral filtering for top-down instance segmentation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 551-560, 2023. 7", + "[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 6", + "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 5", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[49] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24844", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6, 7", + "[50] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 2", + "[51] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2, 7", + "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 6, 7", + "[53] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 6, 7", + "[54] Zihao Xiao, Longlong Jing, Shangxuan Wu, Alex Zihao Zhu, Jingwei Ji, Chiyu Max Jiang, Wei-Chih Hung, Thomas Funkhouser, Weicheng Kuo, Anelia Angelova, et al. 3d open-vocabulary panoptic segmentation with 2d-3d vision-language distillation. In European Conference on Computer Vision, pages 21-38. Springer, 2025. 2", + "[55] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 6", + "[56] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 2", + "[57] Zhiliu Yang and Chen Liu. Tupper-map: Temporal and unified panoptic perception for 3d metric-semantic mapping. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1094–1101. IEEE, 2021. 2, 7", + "[58] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 2", + "[59] Biao Zhang and Peter Wonka. Point cloud instance segmentation using probabilistic embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8883-8892, 2021. 2", + "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7", + "[61] Weiguang Zhao, Yuyao Yan, Chaolong Yang, Jianan Ye, Xi Yang, and Kaizhu Huang. Divide and conquer: 3d point" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "cloud instance segmentation with point-wise binarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 562-571, 2023. 6, 7", + "bbox": [ + 545, + 90, + 906, + 135 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_model.json b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7e4adf54506960cb21bf688a673a8da988a62eab --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.131, + 0.87, + 0.153 + ], + "angle": 0, + "content": "All in One: Visual-Description-Guided Unified Point Cloud Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.757, + 0.217 + ], + "angle": 0, + "content": "Zongyan Han\\(^{1}\\), Mohamed El Amine Boudjoghra\\(^{2}\\), Jiahua Dong\\(^{1}\\), Jinhong Wang\\(^{1}\\), Rao Muhammad Anwer\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.802, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE \n\\(^{2}\\)Technical University of Munich, Germany" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.255, + 0.782, + 0.269 + ], + "angle": 0, + "content": "{zongyan.han, jiahua.dong, jinhong.wang, rao.anwer}@mbzuai.ac.ae" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.273, + 0.608, + 0.286 + ], + "angle": 0, + "content": "Mohamed.boudjoghra@tum.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.485, + 0.721 + ], + "angle": 0, + "content": "Unified segmentation of 3D point clouds is crucial for scene understanding, but is hindered by its sparse structure, limited annotations, and the challenge of distinguishing fine-grained object classes in complex environments. Existing methods often struggle to capture rich semantic and contextual information due to limited supervision and a lack of diverse multimodal cues, leading to suboptimal differentiation of classes and instances. To address these challenges, we propose VDG-Uni3DSeg, a novel framework that integrates pre-trained vision-language models (e.g., CLIP) and large language models (LLMs) to enhance 3D segmentation. By leveraging LLM-generated textual descriptions and reference images from the internet, our method incorporates rich multimodal cues, facilitating fine-grained class and instance separation. We further design a Semantic-Visual Contrastive Loss to align point features with multimodal queries and a Spatial Enhanced Module to model scene-wide relationships efficiently. Operating within a closed-set paradigm that utilizes multimodal knowledge generated offline, VDG-Uni3DSeg achieves state-of-the-art results in semantic, instance, and panoptic segmentation, offering a scalable and practical solution for 3D understanding. Our code is available at https://github.com/Hanzy1996/VDG-Uni3DSeg." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.222, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Unified 3D point cloud segmentation, comprising semantic, instance, and panoptic segmentation, has become a fundamental component of tasks such as autonomous driving, robotics, and AR/VR applications. However, progress in this field remains challenging because of the unique characteristics of point cloud data and the scarcity of annotated datasets. Unlike 2D images, which benefit from rich contextual cues such as texture and lighting, point clouds are" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.419, + 0.822, + 0.43 + ], + "angle": 0, + "content": "(a) Existing Multi-Modal 3D Segmentation Methods" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.436, + 0.903, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.698 + ], + "angle": 0, + "content": "Figure 1. Comparison between existing multi-modal 3D segmentation methods and our proposed approach. (a) Existing methods depend on paired images and captioning models to generate descriptions, requiring real-time 2D-3D alignment, which increases computational overhead. (b) Our approach leverages pregenerated class descriptions from LLMs and reference images retrieved from the internet, establishing offline semantic references that enhance efficiency and robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.852 + ], + "angle": 0, + "content": "inherently sparse and unstructured, making manual annotation labor-intensive and time-consuming. These limitations have slowed progress in 3D segmentation, especially compared to the strides achieved in 2D segmentation, where large-scale datasets have significantly enhanced model performance. Although methods like OneFormer3D [27] have achieved notable success in unified 3D segmentation, they often struggle to capture fine-grained semantic details, thus limiting their effectiveness in complex 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To address these limitations, we propose a novel multimodal framework that leverages external knowledge from pretrained large language models (LLMs) [47] and vision-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24835" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "language models (CLIP [38]) to integrate diverse cross-modal knowledge into the 3D segmentation pipeline. These pretrained models aim to provide semantically rich textual and visual embeddings, enriching the segmentation process with fine-grained contextual information. By incorporating multimodal knowledge, our approach improves the ability of the 3D segmentation model to capture fine-grained attributes crucial to distinguishing classes and instances in complex scenes, while also minimizing reliance on extensive human annotation to provide a more efficient and scalable solution for understanding 3D point clouds." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.261, + 0.485, + 0.608 + ], + "angle": 0, + "content": "Specifically, our method first prompts an LLM to generate detailed semantic descriptions of 3D classes, capturing key attributes such as color, texture, and shape. These descriptions offer a broad understanding of the characteristics of the class, forming a strong semantic foundation. To further enhance visual context, we incorporate class-specific reference images sourced from the internet, which introduce diverse real-world variations to improve the robustness of object recognition. To effectively integrate these multi-modal cues, we encode both textual descriptions and reference images into query representations via CLIP [38]. These queries are fused with point-wise features within a unified mask decoder, allowing the model to produce semantic, instance, and panoptic segmentation masks. Additionally, we introduce a Semantic-Visual Contrastive (SVC) Loss, which aligns point features with their corresponding textual and visual embeddings, thereby enhancing class distinction and segmentation accuracy. To further capture spatial relationships and refine object differentiation, we incorporate a Spatial Enhancement Module. This module utilizes sparse attention to efficiently model scene-wide spatial dependencies, ensuring structural coherence in segmentation while maintaining computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.851 + ], + "angle": 0, + "content": "In particular, unlike open-vocabulary methods [14, 54] that require real-time fusion of 3D scenes with paired images and text captions, our approach follows a closed-set paradigm with predefined class knowledge. We establish offline semantic references through offline-generated class descriptions and reference images (Fig. 1). During inference, our method operates without paired images, captioning models, or additional vision-language modules, enhancing efficiency and deployment practicality. By integrating offline multimodal knowledge with spatially aware feature learning, our approach achieves superior class separation and instance delineation compared to conventional closed-set 3D segmentation baselines. Both quantitative and qualitative comparisons demonstrate consistent improvements in our model performance in 3D instance, semantic, and panoptic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.854, + 0.435, + 0.868 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We introduce VDG-Uni3DSeg, a novel framework that integrates pretrained vision-language models and LLMs" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.772, + 0.107 + ], + "angle": 0, + "content": "to enhance unified 3D segmentation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.181 + ], + "angle": 0, + "content": "- Our method synergizes LLM-generated descriptions and unpaired reference images, enriching class representations with both generalized attributes and fine-grained visual details, thus improving class and instance differentiation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- We develop an efficient mask decoder that integrates textual and visual embeddings as queries, enhancing segmentation accuracy across 3D instance, semantic, and panoptic tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.258, + 0.664, + 0.274 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.862, + 0.3 + ], + "angle": 0, + "content": "2.1. Task specific Unimodal 3D Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.907, + 0.486 + ], + "angle": 0, + "content": "Previous techniques for 3D segmentation adopt point-based or voxel-based approaches to process point clouds. Point-based methods use hand-crafted aggregation mechanisms [32, 36, 37, 46] or transformer blocks [52, 60] for direct point processing, while voxel-based methods convert point clouds into sparse voxel representations, which are then processed using dense [19] or sparse [11] 3D convolutional neural networks. Instance segmentation typically involves first performing semantic segmentation followed by per-point feature aggregation. Recent transformer-based models [42] integrate sparse CNN backbones for feature extraction with transformer decoders." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.76 + ], + "angle": 0, + "content": "The 3D instance segmentation task aims at predicting masks for individual objects in a 3D scene, along with a class label belonging to the set of known classes. Some methods use a bottom-up approach based on grouping learning embeddings in the latent space to facilitate clustering of object points [4, 16, 17, 23, 28, 30, 50, 59]. On the other hand, proposal-based methods adopt a top-down strategy, initially detecting 3D bounding boxes and then segmenting the object region within each box [15, 18, 33, 56, 58]. Inspired by advances in 2D work [6, 8], transformer-based architectures [48] have recently been introduced for 3D instance segmentation [1, 20, 26, 41, 43]. Mask3D [41] introduces a hybrid architecture that uses a 3D CNN backbone for the extraction of per-point characteristics and a transformer-based instance mask decoder to refine a set of queries. Building on Mask3D, [1] shows that explicit spatial and semantic supervision at the level of the 3D backbone further improves instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.908, + 0.822 + ], + "angle": 0, + "content": "For 3D panoptic segmentation, existing methods [34, 51, 57] perform panoptic segmentation on RGB images, lift 2D panoptic masks into 3D space, and aggregate them to generate final 3D panoptic masks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.833, + 0.794, + 0.849 + ], + "angle": 0, + "content": "2.2. Unified Unimodal segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In recent years, several approaches for 2D unified segmentation have been proposed, including MaskFormer [5] and Mask2Former [7]. These methods use a transformer-based" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "architecture to predict masks for three different segmentation tasks simultaneously. OneFormer [21] builds on the same transformer design, but introduces task-conditioned queries, which help the model focus on a specific segmentation task during prediction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.168, + 0.483, + 0.424 + ], + "angle": 0, + "content": "Inspired by the success of OneFormer [21] in 2D, a recent method OneFormer3D [27] introduces a novel multitask unified framework that simultaneously handles 3D semantic, instance, and panoptic segmentation within a single model. Unlike previous approaches, OneFormer3D unifies these tasks through semantic and instance queries in parallel within a transformer decoder, addressing the performance instability by incorporating a new query selection mechanism and an efficient matching strategy, eliminating the need for the Hungarian algorithm, which makes the training process much faster. This approach is trained once on a panoptic dataset, allowing it to outperform existing methods that are specifically tuned for each segmentation task. Although previous work has shown promising results, it still lacks essential multimodal information. In our method, we demonstrate how incorporating this information can significantly improve performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.226, + 0.453 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.627 + ], + "angle": 0, + "content": "In this section, we introduce our proposed method, Visual-Description-Guided Unified 3D Segmentation (VDG-Uni3DSeg). First, we describe the 3D backbone in Sec. 3.1, which extracts point-wise features from the input point cloud. Next, we detail the multi-modal reference queries in Sec. 3.2, which integrate textual and visual information to enhance segmentation performance. We then present the mask decoder in Sec. 3.3, designed to generate task-specific masks for semantic, instance, and panoptic segmentation. Finally, we describe the loss functions used to train the proposed model in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.231, + 0.649 + ], + "angle": 0, + "content": "3.1. 3D Backbone" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.822 + ], + "angle": 0, + "content": "The input to our model is a 3D scene represented as a point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}\\) containing \\(N\\) points. Each point consists of three spatial coordinates and three color values. To expedite the training process and improve computational efficiency, we adopt the voxelization approach from [10], which reduces the point cloud to \\(M\\) superpoints with \\(M < N\\), allowing the model to focus on meaningful spatial clusters. Subsequently, these superpoints are processed by a sparse 3D U-Net to generate point-wise features, \\(\\mathbf{X} \\in \\mathbb{R}^{M \\times d}\\), where \\(d\\) denotes the point embedding dimensionality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Spatial Enhancement Module To capture the spatial relationships between point-wise features in a 3D scene, we adopt a sparse attention mechanism [9] that integrates contextual information from a randomly sampled subset of the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "entire point cloud. Specifically, the point-wise features \\(\\mathbf{X}\\) are first transformed into query, key, and value embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.138, + 0.905, + 0.154 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.222 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_q, \\mathbf{W}_k,\\) and \\(\\mathbf{W}_v \\in \\mathbb{R}^{d \\times d}\\) are learnable weight matrices. For each point feature \\(\\mathbf{x}_i\\), attention is calculated using a randomly sampled subset \\(S_i\\). The spatially enhanced feature is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.233, + 0.905, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {x}} _ {i} = \\sum_ {\\mathbf {v} _ {j} \\in \\mathcal {S} _ {i}} \\alpha_ {i, j} \\cdot \\mathbf {v} _ {j}, \\quad \\alpha_ {i, j} = \\frac {\\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}{\\sum_ {\\mathbf {k} _ {j} \\in \\mathcal {S} _ {i}} \\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.38 + ], + "angle": 0, + "content": "where \\(S_{i}\\) denotes a randomly sampled subset of points from the point cloud, and \\(\\mathbf{q}_i,\\mathbf{k}_j,\\mathbf{v}_j\\in \\mathbb{R}^{1\\times d}\\). Using the subset \\(S_{i}\\), the module captures the spatial structure of the overall scene while significantly reducing computational costs. This makes it highly efficient and well-suited for large-scale 3D point cloud processing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.795, + 0.405 + ], + "angle": 0, + "content": "3.2. Multi-Modal Reference Queries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.905, + 0.563 + ], + "angle": 0, + "content": "Our approach uses a unified 3D mask decoder to perform semantic, instance, and panoptic segmentation simultaneously. Specifically, we build on the Transformer-based mask decoder framework proposed in [21]. For each segmentation task, we incorporate task-specific queries that interact with the point-wise features to produce the corresponding masks. For instance segmentation, we use instance queries as defined in [21], denoted as \\(\\mathbf{Q}_{ins} \\in \\mathbb{R}^{K \\times d}\\). For semantic segmentation, we introduce multimodal semantic queries to generate masks as described below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Description Reference Queries To capture fine-grained details for accurate class distinction, we first introduce the LLM-based Description Generator (LLM-DesGen), a pipeline designed to generate descriptive visual details specific to the class. LLM-DesGen is guided by pre-defined description types and example visual descriptions for given classes. In this in-context learning manner, it can be generalized to unseen classes by learning the descriptive patterns. For each candidate class, LLM-DesGen generates \\(K\\) descriptions for the given class, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.727, + 0.905, + 0.744 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} ^ {c} = \\left[ t _ {1} ^ {c}, \\dots , t _ {K} ^ {c} \\right] = \\text {L L M - D e s G e n} (c), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.905, + 0.846 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}^c\\) represents the descriptions of the \\(c\\)-th class, and \\(t_k^c\\) denotes the \\(k\\)-th description of the \\(c\\)-th class. Finally, we can get the set of descriptions for all classes: \\(\\mathbf{T} = [\\mathbf{T}^1, \\dots, \\mathbf{T}^C]\\). Next, we utilize a text encoder (i.e., CLIP [38]) to embed these descriptions into description embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.847, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {t} = \\operatorname {C L I P - T} (\\mathbf {T}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_t = [\\mathbf{Q}_t^1,\\dots ,\\mathbf{Q}_t^C ]\\) , where \\(\\mathbf{Q}_t^c\\in \\mathbb{R}^{K\\times d}\\) denotes the \\(K\\) queries for class \\(c\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.9, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.356, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Illustration of the VDG-Uni3DSeg framework. First, we utilize a 3D encoder to extract point-wise features from the input point cloud. Then, the Spatial Enhancement Module refines these features by capturing local spatial relationships. Next, class-specific description queries are generated using LLMs, while image queries are constructed from reference images retrieved from the internet. Both description and image queries are embedded using CLIP. These queries, along with the enhanced point features, are processed through a multi-modal fusion block to generate task-specific masks. The description-based and image-based predictions are then fused to achieve robust semantic segmentation. Finally, instance and semantic segmentation masks are integrated to produce panoptic segmentation results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.572 + ], + "angle": 0, + "content": "Image Reference Queries To complement the class descriptions, we collect a set of reference images for each class by collecting images from the internet, denoted \\(\\mathbf{O} = [\\mathbf{O}^{1},\\dots,\\mathbf{O}^{C}]\\), where \\(\\mathbf{O}^c = [o_1^c,\\dots,o_L^c]\\) consists of \\(L\\) images representing the \\(c\\)-th class. We embed these images using the CLIP image encoder [38] to generate visual embeddings as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.584, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {o} = \\operatorname {C L I P - I} (\\mathbf {O}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.671 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_o\\in \\mathbb{R}^{C\\times L\\times d}\\). Unlike multiview images in standard training datasets, these internet-sourced images provide diverse representations with broader variations, reducing reliance on extensive annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.732 + ], + "angle": 0, + "content": "In Uni3DSeg, description queries provide high-level semantic information, capturing the general appearance and characteristics of each class, while visual queries offer fine-grained visual details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.403, + 0.755 + ], + "angle": 0, + "content": "3.3. Multi-Modal Unified Mask Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following the approach in [27], we first utilize \\(B\\) fusion layers to align and blend the information from the point cloud, image, and description domains, enhancing the model's capacity for accurate multimodal mask prediction. Concretely, we apply cross-attention to integrate 3D information into our task-specific queries. Then, we use a self-attention mechanism and a feed-forward network to combine the cross-modal information within these queries. To preserve the original semantic and visual information, we incorporate a residual structure into the description and im" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.527 + ], + "angle": 0, + "content": "age reference queries, as shown in Fig. 2. We then feed these refined queries and the spatially enhanced point features \\(\\widetilde{\\mathbf{X}}\\) into the main mask decoder \\(D\\) to generate the corresponding segmentation masks for each query:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.537, + 0.907, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {i n s} = D \\left(\\mathbf {Q} _ {i n s}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {t} = D \\left(\\mathbf {Q} _ {t}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {o} = D \\left(\\mathbf {Q} _ {o}, \\tilde {\\mathbf {X}}\\right), (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.629 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{ins} \\in \\mathbb{R}^{N \\times S}\\) is the predicted instance segmentation mask, \\(\\mathbf{M}_t \\in \\mathbb{R}^{N \\times C \\times K}\\) and \\(\\mathbf{M}_o \\in \\mathbb{R}^{N \\times C \\times L}\\) are the semantic segmentation masks predicted by the description and image queries, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.907, + 0.772 + ], + "angle": 0, + "content": "Multi-Modal Segmentation Ensemble To achieve a robust semantic segmentation result, we adopt an ensemble strategy that combines the semantic segmentation outputs of \\(\\mathbf{M}_d\\) and \\(\\mathbf{M}_o\\). Concretely, for each point feature \\(\\mathbf{x}_m\\), the description-based semantic prediction \\(\\mathbf{M}_d^m \\in \\mathbb{R}^{C \\times K}\\) represents the similarity scores for all descriptions for each class. To determine the best-matching description for each point, we apply max pooling on \\(\\mathbf{M}_d^m\\) to obtain the refined description-based prediction \\(\\widehat{\\mathbf{M}}_d^m \\in \\mathbb{R}^C\\):" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.784, + 0.907, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathbf {M}} _ {t} ^ {m} = \\max _ {k} \\left(\\mathbf {M} _ {t} ^ {m}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Similarly, we obtain the refined image prediction \\(\\widehat{\\mathbf{M}}_o^m\\). By combining description-based and image-based predictions, the ensemble approach produces a robust semantic segmentation result that leverages both high-level semantic and fine-grained visual details. The final semantic segmentation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.99, + 0.518, + 1.0 + ], + "angle": 0, + "content": "24838" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.269, + 0.107 + ], + "angle": 0, + "content": "prediction is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.116, + 0.483, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {s e m} ^ {m} = \\widehat {\\mathbf {M}} _ {t} ^ {m} + \\widehat {\\mathbf {M}} _ {o} ^ {m}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.144, + 0.483, + 0.19 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{sem}^{m}\\in \\mathbb{R}^{C}\\) is the semantic prediction of the \\(m\\)-th point feature. The final result of semantic segmentation is denoted as \\(\\mathbf{M}_{sem}\\in \\mathbb{R}^{M\\times C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.199, + 0.249, + 0.215 + ], + "angle": 0, + "content": "3.4. Training Losses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.484, + 0.403 + ], + "angle": 0, + "content": "Following [21], we use a combination of cross-entropy, binary cross-entropy, and Dice losses to optimize the instance mask predictions. For semantic segmentation, in addition to cross-entropy loss, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with both description embeddings \\(\\mathbf{Q}_d\\) and image embeddings \\(\\mathbf{Q}_o\\). In contrast to traditional contrastive losses that only align features with a single-modality embedding, the SVC loss promotes multimodal alignment, encouraging spatially enhanced point features \\(\\widetilde{\\mathbf{X}}\\) to be close to both description embeddings \\(\\mathbf{Q}_d\\) and image embeddings \\(\\mathbf{Q}_o\\) of the correct class while pushing them farther from the embeddings of other classes. The SVC loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.408, + 0.314, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S V C}} = \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {t}) + \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {o}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.426, + 0.483, + 0.461 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q}) = - \\mathbb {E} _ {(\\tilde {\\mathbf {x}} _ {i}, \\mathbf {q} ^ {+})} \\left[ \\log \\frac {\\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} ^ {+}\\right) / \\tau\\right)}{\\sum_ {n} \\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} _ {n}\\right) / \\tau\\right)} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.601 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_{+}\\) denotes the positive description or image embedding corresponding to the same class as the point feature \\(\\tilde{\\mathbf{x}}_i\\), \\(\\mathrm{sim}(\\cdot)\\) denotes the inner product similarity, and \\(\\tau\\) is the temperature parameter, empirically set to 1.0. This loss aligns point features with class-specific embeddings, enhancing both semantic consistency and class discriminability. It also facilitates better integration of information from reference descriptions and images, contributing to a more robust segmentation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.602, + 0.469, + 0.617 + ], + "angle": 0, + "content": "The overall loss function for our method is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.626, + 0.482, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a l l} = \\underbrace {\\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} + \\lambda_ {1} \\mathcal {L} _ {c e} ^ {i n s}} _ {\\mathcal {L} _ {i n s}} + \\underbrace {\\lambda_ {2} \\mathcal {L} _ {c e} ^ {s e m} + \\lambda_ {3} \\mathcal {L} _ {S V C}} _ {\\mathcal {L} _ {s e m}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.483, + 0.705 + ], + "angle": 0, + "content": "where \\(\\lambda_{1},\\lambda_{2}\\) , and \\(\\lambda_{3}\\) are hyperparameters that balance different loss functions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.224, + 0.735 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.292, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets We evaluate our method on three widely used datasets: S3DIS [2], ScanNet [13], and ScanNet200 [39]. The S3DIS dataset [2] comprises 272 scenes spread over six areas, with annotations for 13 semantic classes. Among these, five furniture categories are defined as thing classes, while the remaining eight are stuff classes for panoptic evaluation. Following the standard evaluation protocol, we evaluate our method on Area-5 and conduct six-fold cross-validation. ScanNet dataset [13] consists of 1,201 scans" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "for training, 312 scans for validation, and 100 scans for testing. The 3D instance segmentation task includes 18 object classes, while the semantic and panoptic segmentation tasks incorporate two additional background classes. ScanNet200 dataset [39] extends ScanNet [13] with finer-grained annotations, containing 198 object classes for instance segmentation and two additional background classes. It follows the same data split as ScanNet [13]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Evaluation Metrics For 3D semantic segmentation, we evaluate performance based on the mean Intersection over Union (mIoU). For 3D instance segmentation, we compute the mean Average Precision (mAP) across IoU thresholds ranging from \\(50\\%\\) to \\(95\\%\\), with increments of \\(5\\%\\). Furthermore, mAP50 and mAP25 represent mAP values calculated at IoU thresholds of \\(50\\%\\) and \\(25\\%\\), respectively. For S3DIS, we additionally report mean precision (mPrec) and mean recall (mRec) according to the standard evaluation protocol. 3D panoptic segmentation is evaluated using the Panoptic Quality (PQ) score [24]. We also provide separate evaluations for the categories of things and stuff, denoted \\(\\mathrm{PQ}_{\\mathrm{th}}\\) and \\(\\mathrm{PQ}_{\\mathrm{st}}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Implementation Details We implement our method using the MMDetection3D framework [12], following OneFormer3D [27]. The training setup utilizes AdamW optimizer with an initial learning rate of 0.0001, a weight decay of 0.05, a batch size of 4, and a polynomial learning rate scheduler with a power of 0.9. We apply standard data augmentations, including horizontal flipping, random rotations along the z-axis, elastic distortion, and random scaling. For each class, we generate 10 descriptions using LLaMA 3.1 [47] and collect reference images from Bing. We then extract the description and image embeddings using CLIP [38]. These embeddings are further projected into the 3D feature space as queries via a multi-layer perceptron." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.814, + 0.637 + ], + "angle": 0, + "content": "4.2. Comparisons with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare our method against state-of-the-art (SOTA) approaches on three datasets. Table 1 presents comparisons on the S3DIS [2] dataset. Across both the Area-5 and 6-fold cross-validation benchmarks, our method demonstrates strong competitive performance compared to SOTA methods. Notably, on the Area-5 benchmark, our method achieves significant improvements in instance segmentation, outperforming previous methods by a large margin. Specifically, VDG-Uni3DSeg achieves a 2.1-point gain in \\(\\mathrm{MAP}_{50}\\) and a 1.4-point improvement in mAP over the best-performing baseline. For panoptic segmentation, our method surpasses SOTA methods with a 4.1-point gain in PQ and a remarkable 9.6-point improvement in \\(\\mathrm{PQ}_{\\mathrm{th}}\\). Additionally, for semantic segmentation, VDG-Uni3DSeg achieves a 1.7-point improvement in mIoU over our baseline, OneFormer3D [27]. On the 6-fold cross-validation benchmark, VDG-Uni3DSeg achieves similarly notable im" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24839" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.217, + 0.088, + 0.784, + 0.584 + ], + "angle": 0, + "content": "
MethodInstanceSemantic mIoUPQPanoptic PQthPQst
mAP50mAPmPrec50mRec50
Area-5 validation
PointGroup[22]57.861.962.1
DyCo3D[17]64.364.2
SSTNet[31]59.342.765.564.2
DKNet[53]70.865.3
HAIS[3]71.165.0
TD3D[25]65.148.674.464.8
SoftGroup[49]66.151.673.666.6
PBNet[61]66.453.574.965.4
SPFormer[44]66.872.867.1
Mask3D[42]71.957.874.363.7
SegGCN[29]63.6
MinkUNet[11]65.4
PAConv[55]66.6
KPConv[46]67.1
PointTransformer[60]70.4
PointNeXt-XL[37]70.5
PointTransformerV2[52]71.6
OneFormer3D [27]72.058.779.773.069.862.258.465.5
VDG-Uni3DSeg (Ours)74.160.181.073.971.566.368.065.3
6-fold cross-validation
PointGroup[22]64.069.669.2
HAIS[3]73.269.4
SSTNet[31]67.854.173.573.4
DKNet[53]75.371.1
TD3D[25]68.256.276.374.0
SoftGroup[49]68.954.475.369.8
SPFormer[44]69.274.071.1
PBNet[61]70.659.580.172.9
Mask3D[42]74.361.876.566.2
PointNet++[36]56.7
MinkUNet[11]69.1
KPConv[46]70.6
PointTransformer[60]73.5
PointNeXt-XL[37]74.9
OneFormer3D [27]72.860.482.372.772.868.561.574.5
VDG-Uni3DSeg (Ours)75.863.785.574.873.269.168.972.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.593, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Table 1. Comparison of existing segmentation methods on S3DIS. Our proposed VDG-Uni3DSeg shows significant performance improvement across instance, semantic, and panoptic segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.483, + 0.679 + ], + "angle": 0, + "content": "provements across all metrics, demonstrating its robustness and effectiveness in different settings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 2 presents comparisons on the ScanNet [13] dataset, including results for both the validation split and the hidden test split. In the validation split, our method consistently improves performance across all segmentation tasks, including instance, semantic, and panoptic segmentation. Specifically, for instance segmentation, VDG-Uni3DSeg achieves a 1.3-point improvement in \\(\\mathrm{mAP}_{50}\\), demonstrating its enhanced ability to detect and distinguish individual objects. For semantic segmentation, our method achieves a 0.4-point improvement in mIoU, highlighting its improved capability to classify points accurately across the scene. Furthermore, for panoptic segmentation, VDG-Uni3DSeg also achieves improvement in PQ, demonstrating its effectiveness in jointly optimizing instance and semantic pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.799 + ], + "angle": 0, + "content": "dictions. Additionally, VDG-Uni3DSeg also achieves improvements in the ScanNet hidden test set, further highlighting its strong generalization ability. Table 3 presents the results for ScanNet200 [39], which introduces fine-grained classes. VDG-Uni3DSeg achieves state-of-the-art performance in all instances, semantic, and panoptic segmentation, demonstrating its robustness in complex segmentation tasks. Although our model achieves slightly lower semantic mIoU, it remains competitive and maintains balanced performance across all tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "These consistent improvements across different datasets highlight the effectiveness of using both visual and description-based references to guide the segmentation process. By integrating multi-modal information, our method enhances the understanding of objects in 3D scenes, leading to significant performance gains." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24840" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.198, + 0.088, + 0.803, + 0.557 + ], + "angle": 0, + "content": "
MethodPresented atmAP25Instance mAP50mAPSemantic mIoUPQPanoptic PQthPQst
Validation split
NeuralBF[45]WACV'2371.155.536.0
DyCo3D[17]CVPR'2172.957.635.4
SSTNet[31]ICCV'2174.064.349.4
HAIS[3]ICCV'2175.664.443.5
DKNet[53]ICCV'2276.966.750.8
SoftGroup[49]CVPR'2278.967.645.8
PBNet[61]ICCV'2378.970.554.3
TD3D[25]WACV'2481.971.247.3
ISBNet[35]CVPR'2382.573.154.5
SPFormer[44]AAAI'2382.973.956.3
Mask3D[42]ICRA'2383.573.755.2
PointTransformer[60]ICCV'2170.6
PointNeXt-XL[37]NeurIPS'2271.5
PointMetaBase-XXL[32]CVPR'2372.8
PointTransformerV2[52]NeurIPS'2275.4
SceneGraphFusion[51]CVPR'2131.530.243.4
PanopticFusion[34]IROS'1933.530.858.4
TUPPer-Map[57]IROS'2150.247.871.5
OneFormer3D [27]CVPR'2486.177.258.575.871.370.586.1
VDG-Uni3DSeg (Ours)86.578.559.376.271.570.086.6
Hidden test split at 03 Mar: 2025
NeuralBF[45]WACV'2371.855.535.3
DyCo3D[17]CVPR'2176.164.139.5
PointGroup[22]CVPR'2077.863.640.7
SSTNet[31]ICCV'2178.969.850.6
HAIS[3]ICCV'2180.369.945.7
DKNet[53]ICCV'2281.571.853.2
ISBNet[35]CVPR'2383.575.755.9
SPFormer[44]AAAI'2385.177.054.9
SoftGroup[49]CVPR'2286.576.150.4
Mask3D[42]ICRA'2387.078.056.6
TD3D[25]WACV'2487.575.148.9
OneFormer3D [27]CVPR'2489.680.156.6
VDG-Uni3DSeg (Ours)88.080.457.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.568, + 0.907, + 0.597 + ], + "angle": 0, + "content": "Table 2. Comparison of existing segmentation methods on ScanNet. Our proposed VDG-Uni3DSeg achieves superior performance across all metrics for instance, semantic, and panoptic segmentation tasks." + }, + { + "type": "table", + "bbox": [ + 0.237, + 0.609, + 0.763, + 0.753 + ], + "angle": 0, + "content": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
PointGroup[22]24.5
PointGroup + LGround[40]26.1
TD3D[25]40.434.823.1
Mask3D[42]42.337.027.4
MinkUNet[11]25.0
MinkUNet + LGround[40]28.9
OneFormer3D [27]44.240.029.029.031.030.578.6
VDG-Uni3DSeg (Ours)45.140.029.529.731.330.977.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.763, + 0.908, + 0.793 + ], + "angle": 0, + "content": "Table 3. Comparison of existing segmentation methods on the ScanNet200 validation split. Our VDG-Uni3DSeg shows competitive performance across instance, semantic, and panoptic segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.245, + 0.823 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Table 4 presents an ablation study evaluating the contribution of each component in our VDG-Uni3DSeg framework on S3DIS. First, removing the Spatial Enhancement Module slightly degrades the performance, particularly in instance and panoptic segmentation metrics (e.g., a 2-point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.808, + 0.909, + 0.9 + ], + "angle": 0, + "content": "decrease in mAP and a 1.7-point drop in PQ). This underscores the importance of spatial information in effectively distinguishing instances within the scene. Next, we analyze the impact of reference queries. Removing either the Description Query or the Image Query leads to a significant performance drop, with a more pronounced decline when" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24841" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.237, + 0.089, + 0.763, + 0.214 + ], + "angle": 0, + "content": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
baseline79.572.058.769.862.258.465.5
VDG-Uni3DSeg80.674.160.171.566.368.065.3
W/O Spatial Enhancement80.073.358.170.664.663.865.7
W/O Description Query78.472.459.670.264.868.962.2
W/O Image Query79.772.760.070.465.164.365.6
W/O SVC Loss78.072.558.369.062.261.163.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.224, + 0.905, + 0.252 + ], + "angle": 0, + "content": "Table 4. Ablation study evaluating each component of our VDG-Uni3DSeg. 'W/O' indicates that the corresponding part of our VDG-Uni3DSeg is removed." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.262, + 0.258, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.344, + 0.227, + 0.356 + ], + "angle": 0, + "content": "(a) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.262, + 0.417, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.344, + 0.388, + 0.356 + ], + "angle": 0, + "content": "(b) OneFormer3D" + }, + { + "type": "image", + "bbox": [ + 0.419, + 0.262, + 0.575, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.344, + 0.566, + 0.357 + ], + "angle": 0, + "content": "(c) Description Prediction" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.262, + 0.732, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.344, + 0.71, + 0.356 + ], + "angle": 0, + "content": "(d) Image Prediction" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.262, + 0.89, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.759, + 0.344, + 0.866, + 0.357 + ], + "angle": 0, + "content": "(e)VDG-Uni3DSeg" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.362, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of semantic segmentation on ScanNet. Compared to OneFormer3D, the Description and Image Prediction modules improve segmentation accuracy, particularly for tables. By integrating both description and image queries, VDG-Uni3DSeg achieves more balanced and consistent segmentation, enhancing the semantic segmentation for small and challenging objects such as doors and chairs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.436, + 0.486, + 0.679 + ], + "angle": 0, + "content": "the Description Query is removed. This demonstrates the benefit of detailed descriptions generated by LLMs in our method. Furthermore, incorporating reference images helps bridge the gap between textual descriptions and the 3D point cloud. By integrating description and image queries, our method enables each point to associate with the most relevant class representation, leading to overall improvements across all segmentation tasks. Finally, removing the SVC Loss causes the most significant drop in instance and panoptic segmentation metrics, including a 2.6-point decline in \\(\\mathrm{mAP}_{25}\\) and a 6.9-point drop in \\(\\mathrm{PQ_{th}}\\). This demonstrates that SVC Loss encourage the model to learn more discriminative segmentation features by leveraging multimodal queries as anchors for feature alignment. In general, each component of our method contributes to improving feature alignment and overall segmentation performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.274, + 0.71 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 illustrates the semantic segmentation results of different methods. Compared to OneFormer3D [27], both the Description Prediction and Image Prediction modules improve segmentation accuracy by leveraging reference queries, with notable improvements in identifying tables. However, these individual modules struggle with segmenting complex objects such as chairs and doors. When combining description and image predictions, as in VDG-Uni3DSeg, the segmentation results exhibit more balanced and comprehensive improvements across all object categories. This integration results in more precise and consistent segmentation, demonstrating that while individual queries contribute to perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.482 + ], + "angle": 0, + "content": "mance gains, their combined use in VDG-Uni3DSeg yields the most accurate and robust results. Other results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.634, + 0.513 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.523, + 0.907, + 0.812 + ], + "angle": 0, + "content": "We present VDG-Uni3DSeg, a unified 3D segmentation framework that integrates LLM-generated descriptions and internet-sourced images as reference cues to enhance segmentation performance. By leveraging multimodal information, our method improves class and instance differentiation without requiring human annotations or paired multimodal data. To further refine segmentation, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with multimodal queries to enhance feature distinctiveness. Additionally, our Spatial Enhancement Module effectively captures long-range spatial relationships, ensuring precise segmentation boundaries while balancing computational cost. Extensive experiments on multiple datasets demonstrate that VDG-Uni3DSeg achieves state-of-the-art performance across semantic, instance, and panoptic segmentation tasks. These results highlight the effectiveness of multimodal guidance in 3D segmentation and pave the way for future research in leveraging external knowledge to further improve 3D scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.686, + 0.846 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We would like to express our sincere gratitude to all the reviewers for their valuable and constructive suggestions, which have significantly improved the quality of this work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24842" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.185 + ], + "angle": 0, + "content": "[1] Salwa Al Khatib, Mohamed El Amine Boudjoghra, Jean Lahoud, and Fahad Shahbaz Khan. 3d instance segmentation via enhanced spatial and semantic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 541-550, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[9] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.304, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[16] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[17] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[18] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[19] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.906, + 0.608 + ], + "angle": 0, + "content": "[20] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.907, + 0.678 + ], + "angle": 0, + "content": "[21] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[23] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.906, + 0.873 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Maksim Kolodiazhnyi, Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Top-down beats bottom-up in 3d" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24843" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "instance segmentation. arXiv preprint arXiv:2302.02871, 2023.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[26] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[27] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[29] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11611-11620, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[31] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[32] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17682-17691, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[33] Shih-Hung Liu, Shang-Yi Yu, Shao-Chi Wu, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning gaussian instance segmentation in point clouds. arXiv preprint arXiv:2007.09860, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[34] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4205-4212. IEEE, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[35] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[40] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In International Conference on Robotics and Automation (ICRA), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[42] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[44] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[45] Weiwei Sun, Daniel Rebain, Renjie Liao, Vladimir Tankovich, Soroosh Yazdani, Kwang Moo Yi, and Andrea Tagliasacchi. Neuralbf: Neural bilateral filtering for top-down instance segmentation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 551-560, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[50] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[51] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[53] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[54] Zihao Xiao, Longlong Jing, Shangxuan Wu, Alex Zihao Zhu, Jingwei Ji, Chiyu Max Jiang, Wei-Chih Hung, Thomas Funkhouser, Weicheng Kuo, Anelia Angelova, et al. 3d open-vocabulary panoptic segmentation with 2d-3d vision-language distillation. In European Conference on Computer Vision, pages 21-38. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[55] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[56] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[57] Zhiliu Yang and Chen Liu. Tupper-map: Temporal and unified panoptic perception for 3d metric-semantic mapping. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1094–1101. IEEE, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[58] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[59] Biao Zhang and Peter Wonka. Point cloud instance segmentation using probabilistic embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8883-8892, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.903 + ], + "angle": 0, + "content": "[61] Weiguang Zhao, Yuyao Yan, Chaolong Yang, Jianan Ye, Xi Yang, and Kaizhu Huang. Divide and conquer: 3d point" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.907, + 0.136 + ], + "angle": 0, + "content": "cloud instance segmentation with point-wise binarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 562-571, 2023. 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24845" + } + ] +] \ No newline at end of file diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_origin.pdf b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9978186bda7f2133ea0b19be36dfe6fd911e96b8 --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/9f45fe2a-eeec-496b-b486-eee2ccb9cf69_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f47866508510e16b99b7c63a54eeaedce19d5ec09e629f9386fb6e3ab6f8a1c +size 11041485 diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/full.md b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..962d540f4353786e5cd98858aaf8e3c49225b4fe --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/full.md @@ -0,0 +1,317 @@ +# All in One: Visual-Description-Guided Unified Point Cloud Segmentation + +Zongyan Han $^{1}$ , Mohamed El Amine Boudjoghra $^{2}$ , Jiahua Dong $^{1}$ , Jinhong Wang $^{1}$ , Rao Muhammad Anwer $^{1}$ + +$^{1}$ Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE + $^{2}$ Technical University of Munich, Germany + +{zongyan.han, jiahua.dong, jinhong.wang, rao.anwer}@mbzuai.ac.ae + +Mohamed.boudjoghra@tum.de + +# Abstract + +Unified segmentation of 3D point clouds is crucial for scene understanding, but is hindered by its sparse structure, limited annotations, and the challenge of distinguishing fine-grained object classes in complex environments. Existing methods often struggle to capture rich semantic and contextual information due to limited supervision and a lack of diverse multimodal cues, leading to suboptimal differentiation of classes and instances. To address these challenges, we propose VDG-Uni3DSeg, a novel framework that integrates pre-trained vision-language models (e.g., CLIP) and large language models (LLMs) to enhance 3D segmentation. By leveraging LLM-generated textual descriptions and reference images from the internet, our method incorporates rich multimodal cues, facilitating fine-grained class and instance separation. We further design a Semantic-Visual Contrastive Loss to align point features with multimodal queries and a Spatial Enhanced Module to model scene-wide relationships efficiently. Operating within a closed-set paradigm that utilizes multimodal knowledge generated offline, VDG-Uni3DSeg achieves state-of-the-art results in semantic, instance, and panoptic segmentation, offering a scalable and practical solution for 3D understanding. Our code is available at https://github.com/Hanzy1996/VDG-Uni3DSeg. + +# 1. Introduction + +Unified 3D point cloud segmentation, comprising semantic, instance, and panoptic segmentation, has become a fundamental component of tasks such as autonomous driving, robotics, and AR/VR applications. However, progress in this field remains challenging because of the unique characteristics of point cloud data and the scarcity of annotated datasets. Unlike 2D images, which benefit from rich contextual cues such as texture and lighting, point clouds are + +![](images/591cb123da4da34894a16bcebe0da4c5e102ddb346a0548ad9f11e9b3fa52b05.jpg) +(a) Existing Multi-Modal 3D Segmentation Methods + +![](images/d2e4ee34d5352b081f5518638c460f1cdd3da1f7388b21c77d18d710c8991184.jpg) +Figure 1. Comparison between existing multi-modal 3D segmentation methods and our proposed approach. (a) Existing methods depend on paired images and captioning models to generate descriptions, requiring real-time 2D-3D alignment, which increases computational overhead. (b) Our approach leverages pregenerated class descriptions from LLMs and reference images retrieved from the internet, establishing offline semantic references that enhance efficiency and robustness. + +inherently sparse and unstructured, making manual annotation labor-intensive and time-consuming. These limitations have slowed progress in 3D segmentation, especially compared to the strides achieved in 2D segmentation, where large-scale datasets have significantly enhanced model performance. Although methods like OneFormer3D [27] have achieved notable success in unified 3D segmentation, they often struggle to capture fine-grained semantic details, thus limiting their effectiveness in complex 3D scenes. + +To address these limitations, we propose a novel multimodal framework that leverages external knowledge from pretrained large language models (LLMs) [47] and vision- + +language models (CLIP [38]) to integrate diverse cross-modal knowledge into the 3D segmentation pipeline. These pretrained models aim to provide semantically rich textual and visual embeddings, enriching the segmentation process with fine-grained contextual information. By incorporating multimodal knowledge, our approach improves the ability of the 3D segmentation model to capture fine-grained attributes crucial to distinguishing classes and instances in complex scenes, while also minimizing reliance on extensive human annotation to provide a more efficient and scalable solution for understanding 3D point clouds. + +Specifically, our method first prompts an LLM to generate detailed semantic descriptions of 3D classes, capturing key attributes such as color, texture, and shape. These descriptions offer a broad understanding of the characteristics of the class, forming a strong semantic foundation. To further enhance visual context, we incorporate class-specific reference images sourced from the internet, which introduce diverse real-world variations to improve the robustness of object recognition. To effectively integrate these multi-modal cues, we encode both textual descriptions and reference images into query representations via CLIP [38]. These queries are fused with point-wise features within a unified mask decoder, allowing the model to produce semantic, instance, and panoptic segmentation masks. Additionally, we introduce a Semantic-Visual Contrastive (SVC) Loss, which aligns point features with their corresponding textual and visual embeddings, thereby enhancing class distinction and segmentation accuracy. To further capture spatial relationships and refine object differentiation, we incorporate a Spatial Enhancement Module. This module utilizes sparse attention to efficiently model scene-wide spatial dependencies, ensuring structural coherence in segmentation while maintaining computational efficiency. + +In particular, unlike open-vocabulary methods [14, 54] that require real-time fusion of 3D scenes with paired images and text captions, our approach follows a closed-set paradigm with predefined class knowledge. We establish offline semantic references through offline-generated class descriptions and reference images (Fig. 1). During inference, our method operates without paired images, captioning models, or additional vision-language modules, enhancing efficiency and deployment practicality. By integrating offline multimodal knowledge with spatially aware feature learning, our approach achieves superior class separation and instance delineation compared to conventional closed-set 3D segmentation baselines. Both quantitative and qualitative comparisons demonstrate consistent improvements in our model performance in 3D instance, semantic, and panoptic segmentation tasks. + +Our contributions can be summarized as follows: + +- We introduce VDG-Uni3DSeg, a novel framework that integrates pretrained vision-language models and LLMs + +to enhance unified 3D segmentation. + +- Our method synergizes LLM-generated descriptions and unpaired reference images, enriching class representations with both generalized attributes and fine-grained visual details, thus improving class and instance differentiation. +- We develop an efficient mask decoder that integrates textual and visual embeddings as queries, enhancing segmentation accuracy across 3D instance, semantic, and panoptic tasks. + +# 2. Related Works + +# 2.1. Task specific Unimodal 3D Segmentation + +Previous techniques for 3D segmentation adopt point-based or voxel-based approaches to process point clouds. Point-based methods use hand-crafted aggregation mechanisms [32, 36, 37, 46] or transformer blocks [52, 60] for direct point processing, while voxel-based methods convert point clouds into sparse voxel representations, which are then processed using dense [19] or sparse [11] 3D convolutional neural networks. Instance segmentation typically involves first performing semantic segmentation followed by per-point feature aggregation. Recent transformer-based models [42] integrate sparse CNN backbones for feature extraction with transformer decoders. + +The 3D instance segmentation task aims at predicting masks for individual objects in a 3D scene, along with a class label belonging to the set of known classes. Some methods use a bottom-up approach based on grouping learning embeddings in the latent space to facilitate clustering of object points [4, 16, 17, 23, 28, 30, 50, 59]. On the other hand, proposal-based methods adopt a top-down strategy, initially detecting 3D bounding boxes and then segmenting the object region within each box [15, 18, 33, 56, 58]. Inspired by advances in 2D work [6, 8], transformer-based architectures [48] have recently been introduced for 3D instance segmentation [1, 20, 26, 41, 43]. Mask3D [41] introduces a hybrid architecture that uses a 3D CNN backbone for the extraction of per-point characteristics and a transformer-based instance mask decoder to refine a set of queries. Building on Mask3D, [1] shows that explicit spatial and semantic supervision at the level of the 3D backbone further improves instance segmentation. + +For 3D panoptic segmentation, existing methods [34, 51, 57] perform panoptic segmentation on RGB images, lift 2D panoptic masks into 3D space, and aggregate them to generate final 3D panoptic masks. + +# 2.2. Unified Unimodal segmentation + +In recent years, several approaches for 2D unified segmentation have been proposed, including MaskFormer [5] and Mask2Former [7]. These methods use a transformer-based + +architecture to predict masks for three different segmentation tasks simultaneously. OneFormer [21] builds on the same transformer design, but introduces task-conditioned queries, which help the model focus on a specific segmentation task during prediction. + +Inspired by the success of OneFormer [21] in 2D, a recent method OneFormer3D [27] introduces a novel multitask unified framework that simultaneously handles 3D semantic, instance, and panoptic segmentation within a single model. Unlike previous approaches, OneFormer3D unifies these tasks through semantic and instance queries in parallel within a transformer decoder, addressing the performance instability by incorporating a new query selection mechanism and an efficient matching strategy, eliminating the need for the Hungarian algorithm, which makes the training process much faster. This approach is trained once on a panoptic dataset, allowing it to outperform existing methods that are specifically tuned for each segmentation task. Although previous work has shown promising results, it still lacks essential multimodal information. In our method, we demonstrate how incorporating this information can significantly improve performance. + +# 3. Methodology + +In this section, we introduce our proposed method, Visual-Description-Guided Unified 3D Segmentation (VDG-Uni3DSeg). First, we describe the 3D backbone in Sec. 3.1, which extracts point-wise features from the input point cloud. Next, we detail the multi-modal reference queries in Sec. 3.2, which integrate textual and visual information to enhance segmentation performance. We then present the mask decoder in Sec. 3.3, designed to generate task-specific masks for semantic, instance, and panoptic segmentation. Finally, we describe the loss functions used to train the proposed model in Sec. 3.4. + +# 3.1. 3D Backbone + +The input to our model is a 3D scene represented as a point cloud $\mathbf{P} \in \mathbb{R}^{N \times 6}$ containing $N$ points. Each point consists of three spatial coordinates and three color values. To expedite the training process and improve computational efficiency, we adopt the voxelization approach from [10], which reduces the point cloud to $M$ superpoints with $M < N$ , allowing the model to focus on meaningful spatial clusters. Subsequently, these superpoints are processed by a sparse 3D U-Net to generate point-wise features, $\mathbf{X} \in \mathbb{R}^{M \times d}$ , where $d$ denotes the point embedding dimensionality. + +Spatial Enhancement Module To capture the spatial relationships between point-wise features in a 3D scene, we adopt a sparse attention mechanism [9] that integrates contextual information from a randomly sampled subset of the + +entire point cloud. Specifically, the point-wise features $\mathbf{X}$ are first transformed into query, key, and value embeddings: + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} _ {q}, \mathbf {K} = \mathbf {X} \mathbf {W} _ {k}, \mathbf {V} = \mathbf {X} \mathbf {W} _ {v}, \tag {1} +$$ + +where $\mathbf{W}_q, \mathbf{W}_k,$ and $\mathbf{W}_v \in \mathbb{R}^{d \times d}$ are learnable weight matrices. For each point feature $\mathbf{x}_i$ , attention is calculated using a randomly sampled subset $S_i$ . The spatially enhanced feature is then computed as: + +$$ +\widetilde {\mathbf {x}} _ {i} = \sum_ {\mathbf {v} _ {j} \in \mathcal {S} _ {i}} \alpha_ {i, j} \cdot \mathbf {v} _ {j}, \quad \alpha_ {i, j} = \frac {\exp \left(\mathbf {q} _ {i} \cdot \mathbf {k} _ {j} ^ {\top} / \sqrt {d}\right)}{\sum_ {\mathbf {k} _ {j} \in \mathcal {S} _ {i}} \exp \left(\mathbf {q} _ {i} \cdot \mathbf {k} _ {j} ^ {\top} / \sqrt {d}\right)}, \tag {2} +$$ + +where $S_{i}$ denotes a randomly sampled subset of points from the point cloud, and $\mathbf{q}_i,\mathbf{k}_j,\mathbf{v}_j\in \mathbb{R}^{1\times d}$ . Using the subset $S_{i}$ , the module captures the spatial structure of the overall scene while significantly reducing computational costs. This makes it highly efficient and well-suited for large-scale 3D point cloud processing. + +# 3.2. Multi-Modal Reference Queries + +Our approach uses a unified 3D mask decoder to perform semantic, instance, and panoptic segmentation simultaneously. Specifically, we build on the Transformer-based mask decoder framework proposed in [21]. For each segmentation task, we incorporate task-specific queries that interact with the point-wise features to produce the corresponding masks. For instance segmentation, we use instance queries as defined in [21], denoted as $\mathbf{Q}_{ins} \in \mathbb{R}^{K \times d}$ . For semantic segmentation, we introduce multimodal semantic queries to generate masks as described below. + +Description Reference Queries To capture fine-grained details for accurate class distinction, we first introduce the LLM-based Description Generator (LLM-DesGen), a pipeline designed to generate descriptive visual details specific to the class. LLM-DesGen is guided by pre-defined description types and example visual descriptions for given classes. In this in-context learning manner, it can be generalized to unseen classes by learning the descriptive patterns. For each candidate class, LLM-DesGen generates $K$ descriptions for the given class, as follows: + +$$ +\mathbf {T} ^ {c} = \left[ t _ {1} ^ {c}, \dots , t _ {K} ^ {c} \right] = \text {L L M - D e s G e n} (c), \tag {3} +$$ + +where $\mathbf{T}^c$ represents the descriptions of the $c$ -th class, and $t_k^c$ denotes the $k$ -th description of the $c$ -th class. Finally, we can get the set of descriptions for all classes: $\mathbf{T} = [\mathbf{T}^1, \dots, \mathbf{T}^C]$ . Next, we utilize a text encoder (i.e., CLIP [38]) to embed these descriptions into description embeddings: + +$$ +\mathbf {Q} _ {t} = \operatorname {C L I P - T} (\mathbf {T}), \tag {4} +$$ + +where $\mathbf{Q}_t = [\mathbf{Q}_t^1,\dots ,\mathbf{Q}_t^C ]$ , where $\mathbf{Q}_t^c\in \mathbb{R}^{K\times d}$ denotes the $K$ queries for class $c$ + +![](images/7600f259afdfada17427f6723840a75d86165e527ec3c53b33dc483a29bddb5f.jpg) +Figure 2. Illustration of the VDG-Uni3DSeg framework. First, we utilize a 3D encoder to extract point-wise features from the input point cloud. Then, the Spatial Enhancement Module refines these features by capturing local spatial relationships. Next, class-specific description queries are generated using LLMs, while image queries are constructed from reference images retrieved from the internet. Both description and image queries are embedded using CLIP. These queries, along with the enhanced point features, are processed through a multi-modal fusion block to generate task-specific masks. The description-based and image-based predictions are then fused to achieve robust semantic segmentation. Finally, instance and semantic segmentation masks are integrated to produce panoptic segmentation results. + +Image Reference Queries To complement the class descriptions, we collect a set of reference images for each class by collecting images from the internet, denoted $\mathbf{O} = [\mathbf{O}^{1},\dots,\mathbf{O}^{C}]$ , where $\mathbf{O}^c = [o_1^c,\dots,o_L^c]$ consists of $L$ images representing the $c$ -th class. We embed these images using the CLIP image encoder [38] to generate visual embeddings as follows: + +$$ +\mathbf {Q} _ {o} = \operatorname {C L I P - I} (\mathbf {O}), \tag {5} +$$ + +where $\mathbf{Q}_o\in \mathbb{R}^{C\times L\times d}$ . Unlike multiview images in standard training datasets, these internet-sourced images provide diverse representations with broader variations, reducing reliance on extensive annotations. + +In Uni3DSeg, description queries provide high-level semantic information, capturing the general appearance and characteristics of each class, while visual queries offer fine-grained visual details. + +# 3.3. Multi-Modal Unified Mask Decoder + +Following the approach in [27], we first utilize $B$ fusion layers to align and blend the information from the point cloud, image, and description domains, enhancing the model's capacity for accurate multimodal mask prediction. Concretely, we apply cross-attention to integrate 3D information into our task-specific queries. Then, we use a self-attention mechanism and a feed-forward network to combine the cross-modal information within these queries. To preserve the original semantic and visual information, we incorporate a residual structure into the description and im + +age reference queries, as shown in Fig. 2. We then feed these refined queries and the spatially enhanced point features $\widetilde{\mathbf{X}}$ into the main mask decoder $D$ to generate the corresponding segmentation masks for each query: + +$$ +\mathbf {M} _ {i n s} = D \left(\mathbf {Q} _ {i n s}, \tilde {\mathbf {X}}\right), \mathbf {M} _ {t} = D \left(\mathbf {Q} _ {t}, \tilde {\mathbf {X}}\right), \mathbf {M} _ {o} = D \left(\mathbf {Q} _ {o}, \tilde {\mathbf {X}}\right), (6) +$$ + +where $\mathbf{M}_{ins} \in \mathbb{R}^{N \times S}$ is the predicted instance segmentation mask, $\mathbf{M}_t \in \mathbb{R}^{N \times C \times K}$ and $\mathbf{M}_o \in \mathbb{R}^{N \times C \times L}$ are the semantic segmentation masks predicted by the description and image queries, respectively. + +Multi-Modal Segmentation Ensemble To achieve a robust semantic segmentation result, we adopt an ensemble strategy that combines the semantic segmentation outputs of $\mathbf{M}_d$ and $\mathbf{M}_o$ . Concretely, for each point feature $\mathbf{x}_m$ , the description-based semantic prediction $\mathbf{M}_d^m \in \mathbb{R}^{C \times K}$ represents the similarity scores for all descriptions for each class. To determine the best-matching description for each point, we apply max pooling on $\mathbf{M}_d^m$ to obtain the refined description-based prediction $\widehat{\mathbf{M}}_d^m \in \mathbb{R}^C$ : + +$$ +\widehat {\mathbf {M}} _ {t} ^ {m} = \max _ {k} \left(\mathbf {M} _ {t} ^ {m}\right). \tag {7} +$$ + +Similarly, we obtain the refined image prediction $\widehat{\mathbf{M}}_o^m$ . By combining description-based and image-based predictions, the ensemble approach produces a robust semantic segmentation result that leverages both high-level semantic and fine-grained visual details. The final semantic segmentation + +prediction is computed as: + +$$ +\mathbf {M} _ {s e m} ^ {m} = \widehat {\mathbf {M}} _ {t} ^ {m} + \widehat {\mathbf {M}} _ {o} ^ {m}, \tag {8} +$$ + +where $\mathbf{M}_{sem}^{m}\in \mathbb{R}^{C}$ is the semantic prediction of the $m$ -th point feature. The final result of semantic segmentation is denoted as $\mathbf{M}_{sem}\in \mathbb{R}^{M\times C}$ . + +# 3.4. Training Losses + +Following [21], we use a combination of cross-entropy, binary cross-entropy, and Dice losses to optimize the instance mask predictions. For semantic segmentation, in addition to cross-entropy loss, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with both description embeddings $\mathbf{Q}_d$ and image embeddings $\mathbf{Q}_o$ . In contrast to traditional contrastive losses that only align features with a single-modality embedding, the SVC loss promotes multimodal alignment, encouraging spatially enhanced point features $\widetilde{\mathbf{X}}$ to be close to both description embeddings $\mathbf{Q}_d$ and image embeddings $\mathbf{Q}_o$ of the correct class while pushing them farther from the embeddings of other classes. The SVC loss is defined as follows: + +$$ +\mathcal {L} _ {\mathrm {S V C}} = \mathcal {L} _ {\mathrm {c}} (\widetilde {\mathbf {X}}, \mathbf {Q} _ {t}) + \mathcal {L} _ {\mathrm {c}} (\widetilde {\mathbf {X}}, \mathbf {Q} _ {o}), +$$ + +$$ +\mathcal {L} _ {\mathrm {c}} (\widetilde {\mathbf {X}}, \mathbf {Q}) = - \mathbb {E} _ {(\tilde {\mathbf {x}} _ {i}, \mathbf {q} ^ {+})} \left[ \log \frac {\exp \left(\sin \left(\tilde {\mathbf {x}} _ {i} , \mathbf {q} ^ {+}\right) / \tau\right)}{\sum_ {n} \exp \left(\sin \left(\tilde {\mathbf {x}} _ {i} , \mathbf {q} _ {n}\right) / \tau\right)} \right], \tag {9} +$$ + +where $\mathbf{q}_{+}$ denotes the positive description or image embedding corresponding to the same class as the point feature $\tilde{\mathbf{x}}_i$ , $\mathrm{sim}(\cdot)$ denotes the inner product similarity, and $\tau$ is the temperature parameter, empirically set to 1.0. This loss aligns point features with class-specific embeddings, enhancing both semantic consistency and class discriminability. It also facilitates better integration of information from reference descriptions and images, contributing to a more robust segmentation. + +The overall loss function for our method is defined as: + +$$ +\mathcal {L} _ {a l l} = \underbrace {\mathcal {L} _ {b c e} + \mathcal {L} _ {d i c e} + \lambda_ {1} \mathcal {L} _ {c e} ^ {i n s}} _ {\mathcal {L} _ {i n s}} + \underbrace {\lambda_ {2} \mathcal {L} _ {c e} ^ {s e m} + \lambda_ {3} \mathcal {L} _ {S V C}} _ {\mathcal {L} _ {s e m}}, \tag {10} +$$ + +where $\lambda_{1},\lambda_{2}$ , and $\lambda_{3}$ are hyperparameters that balance different loss functions. + +# 4. Experiments + +# 4.1. Experimental Setting + +Datasets We evaluate our method on three widely used datasets: S3DIS [2], ScanNet [13], and ScanNet200 [39]. The S3DIS dataset [2] comprises 272 scenes spread over six areas, with annotations for 13 semantic classes. Among these, five furniture categories are defined as thing classes, while the remaining eight are stuff classes for panoptic evaluation. Following the standard evaluation protocol, we evaluate our method on Area-5 and conduct six-fold cross-validation. ScanNet dataset [13] consists of 1,201 scans + +for training, 312 scans for validation, and 100 scans for testing. The 3D instance segmentation task includes 18 object classes, while the semantic and panoptic segmentation tasks incorporate two additional background classes. ScanNet200 dataset [39] extends ScanNet [13] with finer-grained annotations, containing 198 object classes for instance segmentation and two additional background classes. It follows the same data split as ScanNet [13]. + +Evaluation Metrics For 3D semantic segmentation, we evaluate performance based on the mean Intersection over Union (mIoU). For 3D instance segmentation, we compute the mean Average Precision (mAP) across IoU thresholds ranging from $50\%$ to $95\%$ , with increments of $5\%$ . Furthermore, mAP50 and mAP25 represent mAP values calculated at IoU thresholds of $50\%$ and $25\%$ , respectively. For S3DIS, we additionally report mean precision (mPrec) and mean recall (mRec) according to the standard evaluation protocol. 3D panoptic segmentation is evaluated using the Panoptic Quality (PQ) score [24]. We also provide separate evaluations for the categories of things and stuff, denoted $\mathrm{PQ}_{\mathrm{th}}$ and $\mathrm{PQ}_{\mathrm{st}}$ , respectively. + +Implementation Details We implement our method using the MMDetection3D framework [12], following OneFormer3D [27]. The training setup utilizes AdamW optimizer with an initial learning rate of 0.0001, a weight decay of 0.05, a batch size of 4, and a polynomial learning rate scheduler with a power of 0.9. We apply standard data augmentations, including horizontal flipping, random rotations along the z-axis, elastic distortion, and random scaling. For each class, we generate 10 descriptions using LLaMA 3.1 [47] and collect reference images from Bing. We then extract the description and image embeddings using CLIP [38]. These embeddings are further projected into the 3D feature space as queries via a multi-layer perceptron. + +# 4.2. Comparisons with SOTA Methods + +We compare our method against state-of-the-art (SOTA) approaches on three datasets. Table 1 presents comparisons on the S3DIS [2] dataset. Across both the Area-5 and 6-fold cross-validation benchmarks, our method demonstrates strong competitive performance compared to SOTA methods. Notably, on the Area-5 benchmark, our method achieves significant improvements in instance segmentation, outperforming previous methods by a large margin. Specifically, VDG-Uni3DSeg achieves a 2.1-point gain in $\mathrm{MAP}_{50}$ and a 1.4-point improvement in mAP over the best-performing baseline. For panoptic segmentation, our method surpasses SOTA methods with a 4.1-point gain in PQ and a remarkable 9.6-point improvement in $\mathrm{PQ}_{\mathrm{th}}$ . Additionally, for semantic segmentation, VDG-Uni3DSeg achieves a 1.7-point improvement in mIoU over our baseline, OneFormer3D [27]. On the 6-fold cross-validation benchmark, VDG-Uni3DSeg achieves similarly notable im + +
MethodInstanceSemantic mIoUPQPanoptic PQthPQst
mAP50mAPmPrec50mRec50
Area-5 validation
PointGroup[22]57.861.962.1
DyCo3D[17]64.364.2
SSTNet[31]59.342.765.564.2
DKNet[53]70.865.3
HAIS[3]71.165.0
TD3D[25]65.148.674.464.8
SoftGroup[49]66.151.673.666.6
PBNet[61]66.453.574.965.4
SPFormer[44]66.872.867.1
Mask3D[42]71.957.874.363.7
SegGCN[29]63.6
MinkUNet[11]65.4
PAConv[55]66.6
KPConv[46]67.1
PointTransformer[60]70.4
PointNeXt-XL[37]70.5
PointTransformerV2[52]71.6
OneFormer3D [27]72.058.779.773.069.862.258.465.5
VDG-Uni3DSeg (Ours)74.160.181.073.971.566.368.065.3
6-fold cross-validation
PointGroup[22]64.069.669.2
HAIS[3]73.269.4
SSTNet[31]67.854.173.573.4
DKNet[53]75.371.1
TD3D[25]68.256.276.374.0
SoftGroup[49]68.954.475.369.8
SPFormer[44]69.274.071.1
PBNet[61]70.659.580.172.9
Mask3D[42]74.361.876.566.2
PointNet++[36]56.7
MinkUNet[11]69.1
KPConv[46]70.6
PointTransformer[60]73.5
PointNeXt-XL[37]74.9
OneFormer3D [27]72.860.482.372.772.868.561.574.5
VDG-Uni3DSeg (Ours)75.863.785.574.873.269.168.972.7
+ +Table 1. Comparison of existing segmentation methods on S3DIS. Our proposed VDG-Uni3DSeg shows significant performance improvement across instance, semantic, and panoptic segmentation tasks. + +provements across all metrics, demonstrating its robustness and effectiveness in different settings. + +Table 2 presents comparisons on the ScanNet [13] dataset, including results for both the validation split and the hidden test split. In the validation split, our method consistently improves performance across all segmentation tasks, including instance, semantic, and panoptic segmentation. Specifically, for instance segmentation, VDG-Uni3DSeg achieves a 1.3-point improvement in $\mathrm{mAP}_{50}$ , demonstrating its enhanced ability to detect and distinguish individual objects. For semantic segmentation, our method achieves a 0.4-point improvement in mIoU, highlighting its improved capability to classify points accurately across the scene. Furthermore, for panoptic segmentation, VDG-Uni3DSeg also achieves improvement in PQ, demonstrating its effectiveness in jointly optimizing instance and semantic pre + +dictions. Additionally, VDG-Uni3DSeg also achieves improvements in the ScanNet hidden test set, further highlighting its strong generalization ability. Table 3 presents the results for ScanNet200 [39], which introduces fine-grained classes. VDG-Uni3DSeg achieves state-of-the-art performance in all instances, semantic, and panoptic segmentation, demonstrating its robustness in complex segmentation tasks. Although our model achieves slightly lower semantic mIoU, it remains competitive and maintains balanced performance across all tasks. + +These consistent improvements across different datasets highlight the effectiveness of using both visual and description-based references to guide the segmentation process. By integrating multi-modal information, our method enhances the understanding of objects in 3D scenes, leading to significant performance gains. + +
MethodPresented atmAP25Instance mAP50mAPSemantic mIoUPQPanoptic PQthPQst
Validation split
NeuralBF[45]WACV'2371.155.536.0
DyCo3D[17]CVPR'2172.957.635.4
SSTNet[31]ICCV'2174.064.349.4
HAIS[3]ICCV'2175.664.443.5
DKNet[53]ICCV'2276.966.750.8
SoftGroup[49]CVPR'2278.967.645.8
PBNet[61]ICCV'2378.970.554.3
TD3D[25]WACV'2481.971.247.3
ISBNet[35]CVPR'2382.573.154.5
SPFormer[44]AAAI'2382.973.956.3
Mask3D[42]ICRA'2383.573.755.2
PointTransformer[60]ICCV'2170.6
PointNeXt-XL[37]NeurIPS'2271.5
PointMetaBase-XXL[32]CVPR'2372.8
PointTransformerV2[52]NeurIPS'2275.4
SceneGraphFusion[51]CVPR'2131.530.243.4
PanopticFusion[34]IROS'1933.530.858.4
TUPPer-Map[57]IROS'2150.247.871.5
OneFormer3D [27]CVPR'2486.177.258.575.871.370.586.1
VDG-Uni3DSeg (Ours)86.578.559.376.271.570.086.6
Hidden test split at 03 Mar: 2025
NeuralBF[45]WACV'2371.855.535.3
DyCo3D[17]CVPR'2176.164.139.5
PointGroup[22]CVPR'2077.863.640.7
SSTNet[31]ICCV'2178.969.850.6
HAIS[3]ICCV'2180.369.945.7
DKNet[53]ICCV'2281.571.853.2
ISBNet[35]CVPR'2383.575.755.9
SPFormer[44]AAAI'2385.177.054.9
SoftGroup[49]CVPR'2286.576.150.4
Mask3D[42]ICRA'2387.078.056.6
TD3D[25]WACV'2487.575.148.9
OneFormer3D [27]CVPR'2489.680.156.6
VDG-Uni3DSeg (Ours)88.080.457.6
+ +Table 2. Comparison of existing segmentation methods on ScanNet. Our proposed VDG-Uni3DSeg achieves superior performance across all metrics for instance, semantic, and panoptic segmentation tasks. + +
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
PointGroup[22]24.5
PointGroup + LGround[40]26.1
TD3D[25]40.434.823.1
Mask3D[42]42.337.027.4
MinkUNet[11]25.0
MinkUNet + LGround[40]28.9
OneFormer3D [27]44.240.029.029.031.030.578.6
VDG-Uni3DSeg (Ours)45.140.029.529.731.330.977.9
+ +Table 3. Comparison of existing segmentation methods on the ScanNet200 validation split. Our VDG-Uni3DSeg shows competitive performance across instance, semantic, and panoptic segmentation tasks. + +# 4.3. Ablation Study + +Table 4 presents an ablation study evaluating the contribution of each component in our VDG-Uni3DSeg framework on S3DIS. First, removing the Spatial Enhancement Module slightly degrades the performance, particularly in instance and panoptic segmentation metrics (e.g., a 2-point + +decrease in mAP and a 1.7-point drop in PQ). This underscores the importance of spatial information in effectively distinguishing instances within the scene. Next, we analyze the impact of reference queries. Removing either the Description Query or the Image Query leads to a significant performance drop, with a more pronounced decline when + +
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
baseline79.572.058.769.862.258.465.5
VDG-Uni3DSeg80.674.160.171.566.368.065.3
W/O Spatial Enhancement80.073.358.170.664.663.865.7
W/O Description Query78.472.459.670.264.868.962.2
W/O Image Query79.772.760.070.465.164.365.6
W/O SVC Loss78.072.558.369.062.261.163.8
+ +Table 4. Ablation study evaluating each component of our VDG-Uni3DSeg. 'W/O' indicates that the corresponding part of our VDG-Uni3DSeg is removed. + +![](images/89e9fe500a265b758ebec91f6bd1016a01ef5e438c96814b1d6921d723c6786c.jpg) +(a) Ground Truth + +![](images/40b2246015b8dea520969fa7548281312f6d2bb32297429a195bcd06aa2e3dd3.jpg) +(b) OneFormer3D +Figure 3. Qualitative results of semantic segmentation on ScanNet. Compared to OneFormer3D, the Description and Image Prediction modules improve segmentation accuracy, particularly for tables. By integrating both description and image queries, VDG-Uni3DSeg achieves more balanced and consistent segmentation, enhancing the semantic segmentation for small and challenging objects such as doors and chairs. + +![](images/61ec7e75c39743691d59428c83b4c599b6f020d35d2fa5ea4909dd7794a1c51d.jpg) +(c) Description Prediction + +![](images/64a45f970f34d1bdf44bcc39c4a1be3df4b1f861169b99177054f42e1e9fb524.jpg) +(d) Image Prediction + +![](images/5fefbcbbcf228e20e0e41c347524ef003e0947c2c1a491e8303dfef32b023f28.jpg) +(e)VDG-Uni3DSeg + +the Description Query is removed. This demonstrates the benefit of detailed descriptions generated by LLMs in our method. Furthermore, incorporating reference images helps bridge the gap between textual descriptions and the 3D point cloud. By integrating description and image queries, our method enables each point to associate with the most relevant class representation, leading to overall improvements across all segmentation tasks. Finally, removing the SVC Loss causes the most significant drop in instance and panoptic segmentation metrics, including a 2.6-point decline in $\mathrm{mAP}_{25}$ and a 6.9-point drop in $\mathrm{PQ_{th}}$ . This demonstrates that SVC Loss encourage the model to learn more discriminative segmentation features by leveraging multimodal queries as anchors for feature alignment. In general, each component of our method contributes to improving feature alignment and overall segmentation performance. + +# 4.4. Qualitative Results + +Fig. 3 illustrates the semantic segmentation results of different methods. Compared to OneFormer3D [27], both the Description Prediction and Image Prediction modules improve segmentation accuracy by leveraging reference queries, with notable improvements in identifying tables. However, these individual modules struggle with segmenting complex objects such as chairs and doors. When combining description and image predictions, as in VDG-Uni3DSeg, the segmentation results exhibit more balanced and comprehensive improvements across all object categories. This integration results in more precise and consistent segmentation, demonstrating that while individual queries contribute to perfor + +mance gains, their combined use in VDG-Uni3DSeg yields the most accurate and robust results. Other results are provided in the supplementary material. + +# 5. Conclusion + +We present VDG-Uni3DSeg, a unified 3D segmentation framework that integrates LLM-generated descriptions and internet-sourced images as reference cues to enhance segmentation performance. By leveraging multimodal information, our method improves class and instance differentiation without requiring human annotations or paired multimodal data. To further refine segmentation, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with multimodal queries to enhance feature distinctiveness. Additionally, our Spatial Enhancement Module effectively captures long-range spatial relationships, ensuring precise segmentation boundaries while balancing computational cost. Extensive experiments on multiple datasets demonstrate that VDG-Uni3DSeg achieves state-of-the-art performance across semantic, instance, and panoptic segmentation tasks. These results highlight the effectiveness of multimodal guidance in 3D segmentation and pave the way for future research in leveraging external knowledge to further improve 3D scene understanding. + +# 6. Acknowledgment + +We would like to express our sincere gratitude to all the reviewers for their valuable and constructive suggestions, which have significantly improved the quality of this work. + +# References + +[1] Salwa Al Khatib, Mohamed El Amine Boudjoghra, Jean Lahoud, and Fahad Shahbaz Khan. 3d instance segmentation via enhanced spatial and semantic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 541-550, 2023. 2 +[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 5 +[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 6, 7 +[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 2 +[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2 +[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2 +[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2 +[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2 +[9] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 3 +[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3 +[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 6, 7 +[12] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5 + +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6 +[14] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2 +[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2 +[16] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 2 +[17] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 2, 6, 7 +[18] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2 +[19] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2 +[20] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[21] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 3, 5 +[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 6, 7 +[23] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2 +[24] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5 +[25] Maksim Kolodiazhnyi, Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Top-down beats bottom-up in 3d + +instance segmentation. arXiv preprint arXiv:2302.02871, 2023.6,7 +[26] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[27] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 1, 3, 4, 5, 6, 7, 8 +[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 2 +[29] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11611-11620, 2020. 6 +[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 2 +[31] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6, 7 +[32] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17682-17691, 2023. 2, 7 +[33] Shih-Hung Liu, Shang-Yi Yu, Shao-Chi Wu, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning gaussian instance segmentation in point clouds. arXiv preprint arXiv:2007.09860, 2020. 2 +[34] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4205-4212. IEEE, 2019. 2, 7 +[35] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 7 +[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6 +[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved + +training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 2, 6, 7 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5 +[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 5, 6 +[40] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 7 +[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In International Conference on Robotics and Automation (ICRA), 2023. 2 +[42] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6, 7 +[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2 +[44] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 6, 7 +[45] Weiwei Sun, Daniel Rebain, Renjie Liao, Vladimir Tankovich, Soroosh Yazdani, Kwang Moo Yi, and Andrea Tagliasacchi. Neuralbf: Neural bilateral filtering for top-down instance segmentation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 551-560, 2023. 7 +[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 6 +[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 5 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[49] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on + +point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6, 7 +[50] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 2 +[51] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2, 7 +[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 6, 7 +[53] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 6, 7 +[54] Zihao Xiao, Longlong Jing, Shangxuan Wu, Alex Zihao Zhu, Jingwei Ji, Chiyu Max Jiang, Wei-Chih Hung, Thomas Funkhouser, Weicheng Kuo, Anelia Angelova, et al. 3d open-vocabulary panoptic segmentation with 2d-3d vision-language distillation. In European Conference on Computer Vision, pages 21-38. Springer, 2025. 2 +[55] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 6 +[56] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 2 +[57] Zhiliu Yang and Chen Liu. Tupper-map: Temporal and unified panoptic perception for 3d metric-semantic mapping. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1094–1101. IEEE, 2021. 2, 7 +[58] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 2 +[59] Biao Zhang and Peter Wonka. Point cloud instance segmentation using probabilistic embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8883-8892, 2021. 2 +[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7 +[61] Weiguang Zhao, Yuyao Yan, Chaolong Yang, Jianan Ye, Xi Yang, and Kaizhu Huang. Divide and conquer: 3d point + +cloud instance segmentation with point-wise binarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 562-571, 2023. 6, 7 \ No newline at end of file diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/images.zip b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e8151bdc5e2aaa5e183e2bda09abfb846981567 --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0abdc9fa58950d94878a4e01869249935f4c43f9b8deb9ea6043d2d7b037f4a7 +size 569225 diff --git a/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/layout.json b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5a95158b4aeb490369f8d7a268e073223229e3c --- /dev/null +++ b/2025/All in One_ Visual-Description-Guided Unified Point Cloud Segmentation/layout.json @@ -0,0 +1,8315 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "spans": [ + { + "bbox": [ + 78, + 103, + 532, + 121 + ], + "type": "text", + "content": "All in One: Visual-Description-Guided Unified Point Cloud Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "content": "Zongyan Han" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Mohamed El Amine Boudjoghra" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Jiahua Dong" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Jinhong Wang" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "text", + "content": ", Rao Muhammad Anwer" + }, + { + "bbox": [ + 146, + 142, + 463, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "type": "text", + "content": "Mohamed Bin Zayed University of Artificial Intelligence, Abu Dhabi, UAE \n" + }, + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 171, + 490, + 200 + ], + "type": "text", + "content": "Technical University of Munich, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 132, + 201, + 478, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 201, + 478, + 213 + ], + "spans": [ + { + "bbox": [ + 132, + 201, + 478, + 213 + ], + "type": "text", + "content": "{zongyan.han, jiahua.dong, jinhong.wang, rao.anwer}@mbzuai.ac.ae" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 234, + 216, + 372, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 216, + 372, + 226 + ], + "spans": [ + { + "bbox": [ + 234, + 216, + 372, + 226 + ], + "type": "text", + "content": "Mohamed.boudjoghra@tum.de" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 282, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 296, + 571 + ], + "type": "text", + "content": "Unified segmentation of 3D point clouds is crucial for scene understanding, but is hindered by its sparse structure, limited annotations, and the challenge of distinguishing fine-grained object classes in complex environments. Existing methods often struggle to capture rich semantic and contextual information due to limited supervision and a lack of diverse multimodal cues, leading to suboptimal differentiation of classes and instances. To address these challenges, we propose VDG-Uni3DSeg, a novel framework that integrates pre-trained vision-language models (e.g., CLIP) and large language models (LLMs) to enhance 3D segmentation. By leveraging LLM-generated textual descriptions and reference images from the internet, our method incorporates rich multimodal cues, facilitating fine-grained class and instance separation. We further design a Semantic-Visual Contrastive Loss to align point features with multimodal queries and a Spatial Enhanced Module to model scene-wide relationships efficiently. Operating within a closed-set paradigm that utilizes multimodal knowledge generated offline, VDG-Uni3DSeg achieves state-of-the-art results in semantic, instance, and panoptic segmentation, offering a scalable and practical solution for 3D understanding. Our code is available at https://github.com/Hanzy1996/VDG-Uni3DSeg." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 596, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 714 + ], + "type": "text", + "content": "Unified 3D point cloud segmentation, comprising semantic, instance, and panoptic segmentation, has become a fundamental component of tasks such as autonomous driving, robotics, and AR/VR applications. However, progress in this field remains challenging because of the unique characteristics of point cloud data and the scarcity of annotated datasets. Unlike 2D images, which benefit from rich contextual cues such as texture and lighting, point clouds are" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 253, + 553, + 331 + ], + "blocks": [ + { + "bbox": [ + 317, + 253, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 331 + ], + "type": "image", + "image_path": "591cb123da4da34894a16bcebe0da4c5e102ddb346a0548ad9f11e9b3fa52b05.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 331, + 503, + 340 + ], + "lines": [ + { + "bbox": [ + 364, + 331, + 503, + 340 + ], + "spans": [ + { + "bbox": [ + 364, + 331, + 503, + 340 + ], + "type": "text", + "content": "(a) Existing Multi-Modal 3D Segmentation Methods" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 345, + 552, + 450 + ], + "blocks": [ + { + "bbox": [ + 317, + 345, + 552, + 450 + ], + "lines": [ + { + "bbox": [ + 317, + 345, + 552, + 450 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 552, + 450 + ], + "type": "image", + "image_path": "d2e4ee34d5352b081f5518638c460f1cdd3da1f7388b21c77d18d710c8991184.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 463, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 552 + ], + "type": "text", + "content": "Figure 1. Comparison between existing multi-modal 3D segmentation methods and our proposed approach. (a) Existing methods depend on paired images and captioning models to generate descriptions, requiring real-time 2D-3D alignment, which increases computational overhead. (b) Our approach leverages pregenerated class descriptions from LLMs and reference images retrieved from the internet, establishing offline semantic references that enhance efficiency and robustness." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 566, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 674 + ], + "type": "text", + "content": "inherently sparse and unstructured, making manual annotation labor-intensive and time-consuming. These limitations have slowed progress in 3D segmentation, especially compared to the strides achieved in 2D segmentation, where large-scale datasets have significantly enhanced model performance. Although methods like OneFormer3D [27] have achieved notable success in unified 3D segmentation, they often struggle to capture fine-grained semantic details, thus limiting their effectiveness in complex 3D scenes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To address these limitations, we propose a novel multimodal framework that leverages external knowledge from pretrained large language models (LLMs) [47] and vision-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24835" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "language models (CLIP [38]) to integrate diverse cross-modal knowledge into the 3D segmentation pipeline. These pretrained models aim to provide semantically rich textual and visual embeddings, enriching the segmentation process with fine-grained contextual information. By incorporating multimodal knowledge, our approach improves the ability of the 3D segmentation model to capture fine-grained attributes crucial to distinguishing classes and instances in complex scenes, while also minimizing reliance on extensive human annotation to provide a more efficient and scalable solution for understanding 3D point clouds." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 206, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 481 + ], + "type": "text", + "content": "Specifically, our method first prompts an LLM to generate detailed semantic descriptions of 3D classes, capturing key attributes such as color, texture, and shape. These descriptions offer a broad understanding of the characteristics of the class, forming a strong semantic foundation. To further enhance visual context, we incorporate class-specific reference images sourced from the internet, which introduce diverse real-world variations to improve the robustness of object recognition. To effectively integrate these multi-modal cues, we encode both textual descriptions and reference images into query representations via CLIP [38]. These queries are fused with point-wise features within a unified mask decoder, allowing the model to produce semantic, instance, and panoptic segmentation masks. Additionally, we introduce a Semantic-Visual Contrastive (SVC) Loss, which aligns point features with their corresponding textual and visual embeddings, thereby enhancing class distinction and segmentation accuracy. To further capture spatial relationships and refine object differentiation, we incorporate a Spatial Enhancement Module. This module utilizes sparse attention to efficiently model scene-wide spatial dependencies, ensuring structural coherence in segmentation while maintaining computational efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 483, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 673 + ], + "type": "text", + "content": "In particular, unlike open-vocabulary methods [14, 54] that require real-time fusion of 3D scenes with paired images and text captions, our approach follows a closed-set paradigm with predefined class knowledge. We establish offline semantic references through offline-generated class descriptions and reference images (Fig. 1). During inference, our method operates without paired images, captioning models, or additional vision-language modules, enhancing efficiency and deployment practicality. By integrating offline multimodal knowledge with spatially aware feature learning, our approach achieves superior class separation and instance delineation compared to conventional closed-set 3D segmentation baselines. Both quantitative and qualitative comparisons demonstrate consistent improvements in our model performance in 3D instance, semantic, and panoptic segmentation tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 676, + 266, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 676, + 266, + 687 + ], + "spans": [ + { + "bbox": [ + 67, + 676, + 266, + 687 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We introduce VDG-Uni3DSeg, a novel framework that integrates pretrained vision-language models and LLMs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 323, + 72, + 472, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 472, + 84 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 472, + 84 + ], + "type": "text", + "content": "to enhance unified 3D segmentation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 84, + 553, + 192 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 143 + ], + "type": "text", + "content": "- Our method synergizes LLM-generated descriptions and unpaired reference images, enriching class representations with both generalized attributes and fine-grained visual details, thus improving class and instance differentiation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 192 + ], + "type": "text", + "content": "- We develop an efficient mask decoder that integrates textual and visual embeddings as queries, enhancing segmentation accuracy across 3D instance, semantic, and panoptic tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 204, + 406, + 217 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 527, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 527, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 527, + 237 + ], + "type": "text", + "content": "2.1. Task specific Unimodal 3D Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 384 + ], + "type": "text", + "content": "Previous techniques for 3D segmentation adopt point-based or voxel-based approaches to process point clouds. Point-based methods use hand-crafted aggregation mechanisms [32, 36, 37, 46] or transformer blocks [52, 60] for direct point processing, while voxel-based methods convert point clouds into sparse voxel representations, which are then processed using dense [19] or sparse [11] 3D convolutional neural networks. Instance segmentation typically involves first performing semantic segmentation followed by per-point feature aggregation. Recent transformer-based models [42] integrate sparse CNN backbones for feature extraction with transformer decoders." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 601 + ], + "type": "text", + "content": "The 3D instance segmentation task aims at predicting masks for individual objects in a 3D scene, along with a class label belonging to the set of known classes. Some methods use a bottom-up approach based on grouping learning embeddings in the latent space to facilitate clustering of object points [4, 16, 17, 23, 28, 30, 50, 59]. On the other hand, proposal-based methods adopt a top-down strategy, initially detecting 3D bounding boxes and then segmenting the object region within each box [15, 18, 33, 56, 58]. Inspired by advances in 2D work [6, 8], transformer-based architectures [48] have recently been introduced for 3D instance segmentation [1, 20, 26, 41, 43]. Mask3D [41] introduces a hybrid architecture that uses a 3D CNN backbone for the extraction of per-point characteristics and a transformer-based instance mask decoder to refine a set of queries. Building on Mask3D, [1] shows that explicit spatial and semantic supervision at the level of the 3D backbone further improves instance segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 651 + ], + "type": "text", + "content": "For 3D panoptic segmentation, existing methods [34, 51, 57] perform panoptic segmentation on RGB images, lift 2D panoptic masks into 3D space, and aggregate them to generate final 3D panoptic masks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 659, + 485, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 485, + 672 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 485, + 672 + ], + "type": "text", + "content": "2.2. Unified Unimodal segmentation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In recent years, several approaches for 2D unified segmentation have been proposed, including MaskFormer [5] and Mask2Former [7]. These methods use a transformer-based" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24836" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "architecture to predict masks for three different segmentation tasks simultaneously. OneFormer [21] builds on the same transformer design, but introduces task-conditioned queries, which help the model focus on a specific segmentation task during prediction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 295, + 335 + ], + "type": "text", + "content": "Inspired by the success of OneFormer [21] in 2D, a recent method OneFormer3D [27] introduces a novel multitask unified framework that simultaneously handles 3D semantic, instance, and panoptic segmentation within a single model. Unlike previous approaches, OneFormer3D unifies these tasks through semantic and instance queries in parallel within a transformer decoder, addressing the performance instability by incorporating a new query selection mechanism and an efficient matching strategy, eliminating the need for the Hungarian algorithm, which makes the training process much faster. This approach is trained once on a panoptic dataset, allowing it to outperform existing methods that are specifically tuned for each segmentation task. Although previous work has shown promising results, it still lacks essential multimodal information. In our method, we demonstrate how incorporating this information can significantly improve performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 138, + 358 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 496 + ], + "type": "text", + "content": "In this section, we introduce our proposed method, Visual-Description-Guided Unified 3D Segmentation (VDG-Uni3DSeg). First, we describe the 3D backbone in Sec. 3.1, which extracts point-wise features from the input point cloud. Next, we detail the multi-modal reference queries in Sec. 3.2, which integrate textual and visual information to enhance segmentation performance. We then present the mask decoder in Sec. 3.3, designed to generate task-specific masks for semantic, instance, and panoptic segmentation. Finally, we describe the loss functions used to train the proposed model in Sec. 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 141, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 141, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 141, + 514 + ], + "type": "text", + "content": "3.1. 3D Backbone" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": "The input to our model is a 3D scene represented as a point cloud " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N \\times 6}" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": " points. Each point consists of three spatial coordinates and three color values. To expedite the training process and improve computational efficiency, we adopt the voxelization approach from [10], which reduces the point cloud to " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": " superpoints with " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "M < N" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": ", allowing the model to focus on meaningful spatial clusters. Subsequently, these superpoints are processed by a sparse 3D U-Net to generate point-wise features, " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "text", + "content": " denotes the point embedding dimensionality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Spatial Enhancement Module To capture the spatial relationships between point-wise features in a 3D scene, we adopt a sparse attention mechanism [9] that integrates contextual information from a randomly sampled subset of the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "entire point cloud. Specifically, the point-wise features " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " are first transformed into query, key, and value embeddings:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 353, + 109, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 109, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 353, + 109, + 553, + 121 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\tag {1}", + "image_path": "7bafb800c8559cf548a7ce125fdfb140e1988b39351d3c2442faf6b9ddc922af.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q, \\mathbf{W}_k," + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "content": " are learnable weight matrices. For each point feature " + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "content": ", attention is calculated using a randomly sampled subset " + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 313, + 127, + 553, + 175 + ], + "type": "text", + "content": ". The spatially enhanced feature is then computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 184, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 184, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {x}} _ {i} = \\sum_ {\\mathbf {v} _ {j} \\in \\mathcal {S} _ {i}} \\alpha_ {i, j} \\cdot \\mathbf {v} _ {j}, \\quad \\alpha_ {i, j} = \\frac {\\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}{\\sum_ {\\mathbf {k} _ {j} \\in \\mathcal {S} _ {i}} \\exp \\left(\\mathbf {q} _ {i} \\cdot \\mathbf {k} _ {j} ^ {\\top} / \\sqrt {d}\\right)}, \\tag {2}", + "image_path": "256c2e94c761505ae6fb2ae186e0e9cb106592192181764cb4ed52671e5e80f0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "text", + "content": " denotes a randomly sampled subset of points from the point cloud, and " + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i,\\mathbf{k}_j,\\mathbf{v}_j\\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "text", + "content": ". Using the subset " + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 228, + 553, + 300 + ], + "type": "text", + "content": ", the module captures the spatial structure of the overall scene while significantly reducing computational costs. This makes it highly efficient and well-suited for large-scale 3D point cloud processing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 308, + 486, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 486, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 486, + 320 + ], + "type": "text", + "content": "3.2. Multi-Modal Reference Queries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 326, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 445 + ], + "type": "text", + "content": "Our approach uses a unified 3D mask decoder to perform semantic, instance, and panoptic segmentation simultaneously. Specifically, we build on the Transformer-based mask decoder framework proposed in [21]. For each segmentation task, we incorporate task-specific queries that interact with the point-wise features to produce the corresponding masks. For instance segmentation, we use instance queries as defined in [21], denoted as " + }, + { + "bbox": [ + 313, + 326, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{ins} \\in \\mathbb{R}^{K \\times d}" + }, + { + "bbox": [ + 313, + 326, + 553, + 445 + ], + "type": "text", + "content": ". For semantic segmentation, we introduce multimodal semantic queries to generate masks as described below." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "text", + "content": "Description Reference Queries To capture fine-grained details for accurate class distinction, we first introduce the LLM-based Description Generator (LLM-DesGen), a pipeline designed to generate descriptive visual details specific to the class. LLM-DesGen is guided by pre-defined description types and example visual descriptions for given classes. In this in-context learning manner, it can be generalized to unseen classes by learning the descriptive patterns. For each candidate class, LLM-DesGen generates " + }, + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 446, + 555, + 566 + ], + "type": "text", + "content": " descriptions for the given class, as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 575, + 553, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 575, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 353, + 575, + 553, + 589 + ], + "type": "interline_equation", + "content": "\\mathbf {T} ^ {c} = \\left[ t _ {1} ^ {c}, \\dots , t _ {K} ^ {c} \\right] = \\text {L L M - D e s G e n} (c), \\tag {3}", + "image_path": "e0b7b6b9c4cc48f28dd24f297833ac0a488fe043110b661c3f2433b5e03883d5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{T}^c" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": " represents the descriptions of the " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": "-th class, and " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "t_k^c" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": "-th description of the " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": "-th class. Finally, we can get the set of descriptions for all classes: " + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{T} = [\\mathbf{T}^1, \\dots, \\mathbf{T}^C]" + }, + { + "bbox": [ + 313, + 598, + 553, + 670 + ], + "type": "text", + "content": ". Next, we utilize a text encoder (i.e., CLIP [38]) to embed these descriptions into description embeddings:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 394, + 670, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 394, + 670, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {t} = \\operatorname {C L I P - T} (\\mathbf {T}), \\tag {4}", + "image_path": "7b780f9564793fe1e1b260f2c6d03e8b55c5588f7b15283fff7023291ad914ce.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t = [\\mathbf{Q}_t^1,\\dots ,\\mathbf{Q}_t^C ]" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_t^c\\in \\mathbb{R}^{K\\times d}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " queries for class " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "c" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24837" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 550, + 272 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 550, + 272 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 550, + 272 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 550, + 272 + ], + "type": "image", + "image_path": "7600f259afdfada17427f6723840a75d86165e527ec3c53b33dc483a29bddb5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Illustration of the VDG-Uni3DSeg framework. First, we utilize a 3D encoder to extract point-wise features from the input point cloud. Then, the Spatial Enhancement Module refines these features by capturing local spatial relationships. Next, class-specific description queries are generated using LLMs, while image queries are constructed from reference images retrieved from the internet. Both description and image queries are embedded using CLIP. These queries, along with the enhanced point features, are processed through a multi-modal fusion block to generate task-specific masks. The description-based and image-based predictions are then fused to achieve robust semantic segmentation. Finally, instance and semantic segmentation masks are integrated to produce panoptic segmentation results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "Image Reference Queries To complement the class descriptions, we collect a set of reference images for each class by collecting images from the internet, denoted " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{O} = [\\mathbf{O}^{1},\\dots,\\mathbf{O}^{C}]" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{O}^c = [o_1^c,\\dots,o_L^c]" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " images representing the " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "-th class. We embed these images using the CLIP image encoder [38] to generate visual embeddings as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 462, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 462, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 137, + 462, + 295, + 475 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {o} = \\operatorname {C L I P - I} (\\mathbf {O}), \\tag {5}", + "image_path": "73f5d34b1ab1c325b18d4ffee027aa31819a07a3d30680bf8f7baeb0fe886e3f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_o\\in \\mathbb{R}^{C\\times L\\times d}" + }, + { + "bbox": [ + 55, + 483, + 296, + 531 + ], + "type": "text", + "content": ". Unlike multiview images in standard training datasets, these internet-sourced images provide diverse representations with broader variations, reducing reliance on extensive annotations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 579 + ], + "type": "text", + "content": "In Uni3DSeg, description queries provide high-level semantic information, capturing the general appearance and characteristics of each class, while visual queries offer fine-grained visual details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 246, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 246, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 246, + 597 + ], + "type": "text", + "content": "3.3. Multi-Modal Unified Mask Decoder" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 602, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 713 + ], + "type": "text", + "content": "Following the approach in [27], we first utilize " + }, + { + "bbox": [ + 55, + 602, + 296, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 602, + 296, + 713 + ], + "type": "text", + "content": " fusion layers to align and blend the information from the point cloud, image, and description domains, enhancing the model's capacity for accurate multimodal mask prediction. Concretely, we apply cross-attention to integrate 3D information into our task-specific queries. Then, we use a self-attention mechanism and a feed-forward network to combine the cross-modal information within these queries. To preserve the original semantic and visual information, we incorporate a residual structure into the description and im" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "text", + "content": "age reference queries, as shown in Fig. 2. We then feed these refined queries and the spatially enhanced point features " + }, + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "text", + "content": " into the main mask decoder " + }, + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 369, + 555, + 417 + ], + "type": "text", + "content": " to generate the corresponding segmentation masks for each query:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 425, + 555, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 425, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 425, + 555, + 441 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {i n s} = D \\left(\\mathbf {Q} _ {i n s}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {t} = D \\left(\\mathbf {Q} _ {t}, \\tilde {\\mathbf {X}}\\right), \\mathbf {M} _ {o} = D \\left(\\mathbf {Q} _ {o}, \\tilde {\\mathbf {X}}\\right), (6)", + "image_path": "35924356805ba6eb35660320cd74a50583b712572ba3ba545e4704d2b9ae6136.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{ins} \\in \\mathbb{R}^{N \\times S}" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " is the predicted instance segmentation mask, " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t \\in \\mathbb{R}^{N \\times C \\times K}" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_o \\in \\mathbb{R}^{N \\times C \\times L}" + }, + { + "bbox": [ + 313, + 449, + 554, + 498 + ], + "type": "text", + "content": " are the semantic segmentation masks predicted by the description and image queries, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": "Multi-Modal Segmentation Ensemble To achieve a robust semantic segmentation result, we adopt an ensemble strategy that combines the semantic segmentation outputs of " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_o" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": ". Concretely, for each point feature " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_m" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": ", the description-based semantic prediction " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d^m \\in \\mathbb{R}^{C \\times K}" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": " represents the similarity scores for all descriptions for each class. To determine the best-matching description for each point, we apply max pooling on " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_d^m" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": " to obtain the refined description-based prediction " + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{M}}_d^m \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 498, + 555, + 611 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 620, + 555, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 620, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 392, + 620, + 555, + 640 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathbf {M}} _ {t} ^ {m} = \\max _ {k} \\left(\\mathbf {M} _ {t} ^ {m}\\right). \\tag {7}", + "image_path": "d32fdc0f5965a6af9b0f7778f80acbf615f51399324f9ca7a27f232a420acc84.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 652, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 555, + 715 + ], + "type": "text", + "content": "Similarly, we obtain the refined image prediction " + }, + { + "bbox": [ + 313, + 652, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{M}}_o^m" + }, + { + "bbox": [ + 313, + 652, + 555, + 715 + ], + "type": "text", + "content": ". By combining description-based and image-based predictions, the ensemble approach produces a robust semantic segmentation result that leverages both high-level semantic and fine-grained visual details. The final semantic segmentation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 784, + 317, + 792 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 784, + 317, + 792 + ], + "spans": [ + { + "bbox": [ + 293, + 784, + 317, + 792 + ], + "type": "text", + "content": "24838" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 164, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 164, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 164, + 84 + ], + "type": "text", + "content": "prediction is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 91, + 295, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 91, + 295, + 107 + ], + "spans": [ + { + "bbox": [ + 127, + 91, + 295, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {s e m} ^ {m} = \\widehat {\\mathbf {M}} _ {t} ^ {m} + \\widehat {\\mathbf {M}} _ {o} ^ {m}, \\tag {8}", + "image_path": "c194bc61fa8abe98e0d8f7c568da8183c4fde191b2f49f6846da78904eb2d453.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sem}^{m}\\in \\mathbb{R}^{C}" + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "text", + "content": " is the semantic prediction of the " + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "text", + "content": "-th point feature. The final result of semantic segmentation is denoted as " + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{sem}\\in \\mathbb{R}^{M\\times C}" + }, + { + "bbox": [ + 55, + 114, + 295, + 150 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 157, + 152, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 152, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 152, + 170 + ], + "type": "text", + "content": "3.4. Training Losses" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": "Following [21], we use a combination of cross-entropy, binary cross-entropy, and Dice losses to optimize the instance mask predictions. For semantic segmentation, in addition to cross-entropy loss, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with both description embeddings " + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_d" + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": " and image embeddings " + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_o" + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": ". In contrast to traditional contrastive losses that only align features with a single-modality embedding, the SVC loss promotes multimodal alignment, encouraging spatially enhanced point features " + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": " to be close to both description embeddings " + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_d" + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": " and image embeddings " + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_o" + }, + { + "bbox": [ + 55, + 173, + 296, + 319 + ], + "type": "text", + "content": " of the correct class while pushing them farther from the embeddings of other classes. The SVC loss is defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 192, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 192, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 192, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S V C}} = \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {t}) + \\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q} _ {o}),", + "image_path": "b8890d50c563c39ba99d58a5e86f647f2f11a3db0f4a3790dbe33ac180d506a8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} (\\widetilde {\\mathbf {X}}, \\mathbf {Q}) = - \\mathbb {E} _ {(\\tilde {\\mathbf {x}} _ {i}, \\mathbf {q} ^ {+})} \\left[ \\log \\frac {\\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} ^ {+}\\right) / \\tau\\right)}{\\sum_ {n} \\exp \\left(\\sin \\left(\\tilde {\\mathbf {x}} _ {i} , \\mathbf {q} _ {n}\\right) / \\tau\\right)} \\right], \\tag {9}", + "image_path": "aa82fb3e77db0bd6d30c1d9c0586ae150dc941d44911db5ba232f9e492041082.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{+}" + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "content": " denotes the positive description or image embedding corresponding to the same class as the point feature " + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "content": " denotes the inner product similarity, and " + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 369, + 296, + 475 + ], + "type": "text", + "content": " is the temperature parameter, empirically set to 1.0. This loss aligns point features with class-specific embeddings, enhancing both semantic consistency and class discriminability. It also facilitates better integration of information from reference descriptions and images, contributing to a more robust segmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 476, + 287, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 476, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 67, + 476, + 287, + 488 + ], + "type": "text", + "content": "The overall loss function for our method is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 495, + 294, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 495, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 65, + 495, + 294, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a l l} = \\underbrace {\\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} + \\lambda_ {1} \\mathcal {L} _ {c e} ^ {i n s}} _ {\\mathcal {L} _ {i n s}} + \\underbrace {\\lambda_ {2} \\mathcal {L} _ {c e} ^ {s e m} + \\lambda_ {3} \\mathcal {L} _ {S V C}} _ {\\mathcal {L} _ {s e m}}, \\tag {10}", + "image_path": "4d0315691346f431fcddf889ea1e0fbf17d2dcfa400fc36fae975bdc472b8534.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2}" + }, + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 55, + 535, + 295, + 558 + ], + "type": "text", + "content": " are hyperparameters that balance different loss functions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 588, + 178, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 178, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 178, + 601 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Datasets We evaluate our method on three widely used datasets: S3DIS [2], ScanNet [13], and ScanNet200 [39]. The S3DIS dataset [2] comprises 272 scenes spread over six areas, with annotations for 13 semantic classes. Among these, five furniture categories are defined as thing classes, while the remaining eight are stuff classes for panoptic evaluation. Following the standard evaluation protocol, we evaluate our method on Area-5 and conduct six-fold cross-validation. ScanNet dataset [13] consists of 1,201 scans" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "for training, 312 scans for validation, and 100 scans for testing. The 3D instance segmentation task includes 18 object classes, while the semantic and panoptic segmentation tasks incorporate two additional background classes. ScanNet200 dataset [39] extends ScanNet [13] with finer-grained annotations, containing 198 object classes for instance segmentation and two additional background classes. It follows the same data split as ScanNet [13]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": "Evaluation Metrics For 3D semantic segmentation, we evaluate performance based on the mean Intersection over Union (mIoU). For 3D instance segmentation, we compute the mean Average Precision (mAP) across IoU thresholds ranging from " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ", with increments of " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ". Furthermore, mAP50 and mAP25 represent mAP values calculated at IoU thresholds of " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ", respectively. For S3DIS, we additionally report mean precision (mPrec) and mean recall (mRec) according to the standard evaluation protocol. 3D panoptic segmentation is evaluated using the Panoptic Quality (PQ) score [24]. We also provide separate evaluations for the categories of things and stuff, denoted " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}_{\\mathrm{st}}" + }, + { + "bbox": [ + 313, + 169, + 555, + 324 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 326, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 482 + ], + "type": "text", + "content": "Implementation Details We implement our method using the MMDetection3D framework [12], following OneFormer3D [27]. The training setup utilizes AdamW optimizer with an initial learning rate of 0.0001, a weight decay of 0.05, a batch size of 4, and a polynomial learning rate scheduler with a power of 0.9. We apply standard data augmentations, including horizontal flipping, random rotations along the z-axis, elastic distortion, and random scaling. For each class, we generate 10 descriptions using LLaMA 3.1 [47] and collect reference images from Bing. We then extract the description and image embeddings using CLIP [38]. These embeddings are further projected into the 3D feature space as queries via a multi-layer perceptron." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 491, + 498, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 498, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 498, + 504 + ], + "type": "text", + "content": "4.2. Comparisons with SOTA Methods" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "We compare our method against state-of-the-art (SOTA) approaches on three datasets. Table 1 presents comparisons on the S3DIS [2] dataset. Across both the Area-5 and 6-fold cross-validation benchmarks, our method demonstrates strong competitive performance compared to SOTA methods. Notably, on the Area-5 benchmark, our method achieves significant improvements in instance segmentation, outperforming previous methods by a large margin. Specifically, VDG-Uni3DSeg achieves a 2.1-point gain in " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MAP}_{50}" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " and a 1.4-point improvement in mAP over the best-performing baseline. For panoptic segmentation, our method surpasses SOTA methods with a 4.1-point gain in PQ and a remarkable 9.6-point improvement in " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ}_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ". Additionally, for semantic segmentation, VDG-Uni3DSeg achieves a 1.7-point improvement in mIoU over our baseline, OneFormer3D [27]. On the 6-fold cross-validation benchmark, VDG-Uni3DSeg achieves similarly notable im" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24839" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 132, + 69, + 479, + 462 + ], + "blocks": [ + { + "bbox": [ + 132, + 69, + 479, + 462 + ], + "lines": [ + { + "bbox": [ + 132, + 69, + 479, + 462 + ], + "spans": [ + { + "bbox": [ + 132, + 69, + 479, + 462 + ], + "type": "table", + "html": "
MethodInstanceSemantic mIoUPQPanoptic PQthPQst
mAP50mAPmPrec50mRec50
Area-5 validation
PointGroup[22]57.861.962.1
DyCo3D[17]64.364.2
SSTNet[31]59.342.765.564.2
DKNet[53]70.865.3
HAIS[3]71.165.0
TD3D[25]65.148.674.464.8
SoftGroup[49]66.151.673.666.6
PBNet[61]66.453.574.965.4
SPFormer[44]66.872.867.1
Mask3D[42]71.957.874.363.7
SegGCN[29]63.6
MinkUNet[11]65.4
PAConv[55]66.6
KPConv[46]67.1
PointTransformer[60]70.4
PointNeXt-XL[37]70.5
PointTransformerV2[52]71.6
OneFormer3D [27]72.058.779.773.069.862.258.465.5
VDG-Uni3DSeg (Ours)74.160.181.073.971.566.368.065.3
6-fold cross-validation
PointGroup[22]64.069.669.2
HAIS[3]73.269.4
SSTNet[31]67.854.173.573.4
DKNet[53]75.371.1
TD3D[25]68.256.276.374.0
SoftGroup[49]68.954.475.369.8
SPFormer[44]69.274.071.1
PBNet[61]70.659.580.172.9
Mask3D[42]74.361.876.566.2
PointNet++[36]56.7
MinkUNet[11]69.1
KPConv[46]70.6
PointTransformer[60]73.5
PointNeXt-XL[37]74.9
OneFormer3D [27]72.860.482.372.772.868.561.574.5
VDG-Uni3DSeg (Ours)75.863.785.574.873.269.168.972.7
", + "image_path": "f606d13043b0779fa46c0a3c615abb884ccbf705640aeb1865f89fc72c9170b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 469, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 469, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 555, + 492 + ], + "type": "text", + "content": "Table 1. Comparison of existing segmentation methods on S3DIS. Our proposed VDG-Uni3DSeg shows significant performance improvement across instance, semantic, and panoptic segmentation tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 295, + 537 + ], + "type": "text", + "content": "provements across all metrics, demonstrating its robustness and effectiveness in different settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "content": "Table 2 presents comparisons on the ScanNet [13] dataset, including results for both the validation split and the hidden test split. In the validation split, our method consistently improves performance across all segmentation tasks, including instance, semantic, and panoptic segmentation. Specifically, for instance segmentation, VDG-Uni3DSeg achieves a 1.3-point improvement in " + }, + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 54, + 545, + 295, + 713 + ], + "type": "text", + "content": ", demonstrating its enhanced ability to detect and distinguish individual objects. For semantic segmentation, our method achieves a 0.4-point improvement in mIoU, highlighting its improved capability to classify points accurately across the scene. Furthermore, for panoptic segmentation, VDG-Uni3DSeg also achieves improvement in PQ, demonstrating its effectiveness in jointly optimizing instance and semantic pre" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 513, + 555, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 632 + ], + "type": "text", + "content": "dictions. Additionally, VDG-Uni3DSeg also achieves improvements in the ScanNet hidden test set, further highlighting its strong generalization ability. Table 3 presents the results for ScanNet200 [39], which introduces fine-grained classes. VDG-Uni3DSeg achieves state-of-the-art performance in all instances, semantic, and panoptic segmentation, demonstrating its robustness in complex segmentation tasks. Although our model achieves slightly lower semantic mIoU, it remains competitive and maintains balanced performance across all tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "These consistent improvements across different datasets highlight the effectiveness of using both visual and description-based references to guide the segmentation process. By integrating multi-modal information, our method enhances the understanding of objects in 3D scenes, leading to significant performance gains." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24840" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 121, + 69, + 491, + 441 + ], + "blocks": [ + { + "bbox": [ + 121, + 69, + 491, + 441 + ], + "lines": [ + { + "bbox": [ + 121, + 69, + 491, + 441 + ], + "spans": [ + { + "bbox": [ + 121, + 69, + 491, + 441 + ], + "type": "table", + "html": "
MethodPresented atmAP25Instance mAP50mAPSemantic mIoUPQPanoptic PQthPQst
Validation split
NeuralBF[45]WACV'2371.155.536.0
DyCo3D[17]CVPR'2172.957.635.4
SSTNet[31]ICCV'2174.064.349.4
HAIS[3]ICCV'2175.664.443.5
DKNet[53]ICCV'2276.966.750.8
SoftGroup[49]CVPR'2278.967.645.8
PBNet[61]ICCV'2378.970.554.3
TD3D[25]WACV'2481.971.247.3
ISBNet[35]CVPR'2382.573.154.5
SPFormer[44]AAAI'2382.973.956.3
Mask3D[42]ICRA'2383.573.755.2
PointTransformer[60]ICCV'2170.6
PointNeXt-XL[37]NeurIPS'2271.5
PointMetaBase-XXL[32]CVPR'2372.8
PointTransformerV2[52]NeurIPS'2275.4
SceneGraphFusion[51]CVPR'2131.530.243.4
PanopticFusion[34]IROS'1933.530.858.4
TUPPer-Map[57]IROS'2150.247.871.5
OneFormer3D [27]CVPR'2486.177.258.575.871.370.586.1
VDG-Uni3DSeg (Ours)86.578.559.376.271.570.086.6
Hidden test split at 03 Mar: 2025
NeuralBF[45]WACV'2371.855.535.3
DyCo3D[17]CVPR'2176.164.139.5
PointGroup[22]CVPR'2077.863.640.7
SSTNet[31]ICCV'2178.969.850.6
HAIS[3]ICCV'2180.369.945.7
DKNet[53]ICCV'2281.571.853.2
ISBNet[35]CVPR'2383.575.755.9
SPFormer[44]AAAI'2385.177.054.9
SoftGroup[49]CVPR'2286.576.150.4
Mask3D[42]ICRA'2387.078.056.6
TD3D[25]WACV'2487.575.148.9
OneFormer3D [27]CVPR'2489.680.156.6
VDG-Uni3DSeg (Ours)88.080.457.6
", + "image_path": "47383b30fc197f55e80ccd05d3fcfb6fa057d1cc10b6c9a29073f0450accbb77.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 145, + 482, + 466, + 596 + ], + "blocks": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 472 + ], + "type": "text", + "content": "Table 2. Comparison of existing segmentation methods on ScanNet. Our proposed VDG-Uni3DSeg achieves superior performance across all metrics for instance, semantic, and panoptic segmentation tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 145, + 482, + 466, + 596 + ], + "lines": [ + { + "bbox": [ + 145, + 482, + 466, + 596 + ], + "spans": [ + { + "bbox": [ + 145, + 482, + 466, + 596 + ], + "type": "table", + "html": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
PointGroup[22]24.5
PointGroup + LGround[40]26.1
TD3D[25]40.434.823.1
Mask3D[42]42.337.027.4
MinkUNet[11]25.0
MinkUNet + LGround[40]28.9
OneFormer3D [27]44.240.029.029.031.030.578.6
VDG-Uni3DSeg (Ours)45.140.029.529.731.330.977.9
", + "image_path": "6ac62d8189ec140fdf814eef035b7eb386abf6fd3dcfea42df7a76781968228f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 604, + 555, + 628 + ], + "lines": [ + { + "bbox": [ + 54, + 604, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 604, + 555, + 628 + ], + "type": "text", + "content": "Table 3. Comparison of existing segmentation methods on the ScanNet200 validation split. Our VDG-Uni3DSeg shows competitive performance across instance, semantic, and panoptic segmentation tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 639, + 149, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 149, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 149, + 651 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 297, + 715 + ], + "type": "text", + "content": "Table 4 presents an ablation study evaluating the contribution of each component in our VDG-Uni3DSeg framework on S3DIS. First, removing the Spatial Enhancement Module slightly degrades the performance, particularly in instance and panoptic segmentation metrics (e.g., a 2-point" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 639, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 556, + 712 + ], + "type": "text", + "content": "decrease in mAP and a 1.7-point drop in PQ). This underscores the importance of spatial information in effectively distinguishing instances within the scene. Next, we analyze the impact of reference queries. Removing either the Description Query or the Image Query leads to a significant performance drop, with a more pronounced decline when" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24841" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 145, + 70, + 466, + 169 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 466, + 169 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 466, + 169 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 466, + 169 + ], + "type": "table", + "html": "
MethodInstanceSemanticPanoptic
mAP25mAP50mAPmIoUPQPQthPQst
baseline79.572.058.769.862.258.465.5
VDG-Uni3DSeg80.674.160.171.566.368.065.3
W/O Spatial Enhancement80.073.358.170.664.663.865.7
W/O Description Query78.472.459.670.264.868.962.2
W/O Image Query79.772.760.070.465.164.365.6
W/O SVC Loss78.072.558.369.062.261.163.8
", + "image_path": "18f8266bf58e657e0d60c744f0556929d3073d1dbef361b1f7d0c9b28b43633b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 553, + 199 + ], + "type": "text", + "content": "Table 4. Ablation study evaluating each component of our VDG-Uni3DSeg. 'W/O' indicates that the corresponding part of our VDG-Uni3DSeg is removed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 207, + 157, + 271 + ], + "blocks": [ + { + "bbox": [ + 63, + 207, + 157, + 271 + ], + "lines": [ + { + "bbox": [ + 63, + 207, + 157, + 271 + ], + "spans": [ + { + "bbox": [ + 63, + 207, + 157, + 271 + ], + "type": "image", + "image_path": "89e9fe500a265b758ebec91f6bd1016a01ef5e438c96814b1d6921d723c6786c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 272, + 138, + 281 + ], + "lines": [ + { + "bbox": [ + 83, + 272, + 138, + 281 + ], + "spans": [ + { + "bbox": [ + 83, + 272, + 138, + 281 + ], + "type": "text", + "content": "(a) Ground Truth" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 207, + 255, + 271 + ], + "blocks": [ + { + "bbox": [ + 159, + 207, + 255, + 271 + ], + "lines": [ + { + "bbox": [ + 159, + 207, + 255, + 271 + ], + "spans": [ + { + "bbox": [ + 159, + 207, + 255, + 271 + ], + "type": "image", + "image_path": "40b2246015b8dea520969fa7548281312f6d2bb32297429a195bcd06aa2e3dd3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 272, + 237, + 281 + ], + "lines": [ + { + "bbox": [ + 178, + 272, + 237, + 281 + ], + "spans": [ + { + "bbox": [ + 178, + 272, + 237, + 281 + ], + "type": "text", + "content": "(b) OneFormer3D" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 286, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 286, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 555, + 329 + ], + "type": "text", + "content": "Figure 3. Qualitative results of semantic segmentation on ScanNet. Compared to OneFormer3D, the Description and Image Prediction modules improve segmentation accuracy, particularly for tables. By integrating both description and image queries, VDG-Uni3DSeg achieves more balanced and consistent segmentation, enhancing the semantic segmentation for small and challenging objects such as doors and chairs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 256, + 207, + 351, + 271 + ], + "blocks": [ + { + "bbox": [ + 256, + 207, + 351, + 271 + ], + "lines": [ + { + "bbox": [ + 256, + 207, + 351, + 271 + ], + "spans": [ + { + "bbox": [ + 256, + 207, + 351, + 271 + ], + "type": "image", + "image_path": "61ec7e75c39743691d59428c83b4c599b6f020d35d2fa5ea4909dd7794a1c51d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 272, + 346, + 282 + ], + "lines": [ + { + "bbox": [ + 261, + 272, + 346, + 282 + ], + "spans": [ + { + "bbox": [ + 261, + 272, + 346, + 282 + ], + "type": "text", + "content": "(c) Description Prediction" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 353, + 207, + 447, + 271 + ], + "blocks": [ + { + "bbox": [ + 353, + 207, + 447, + 271 + ], + "lines": [ + { + "bbox": [ + 353, + 207, + 447, + 271 + ], + "spans": [ + { + "bbox": [ + 353, + 207, + 447, + 271 + ], + "type": "image", + "image_path": "64a45f970f34d1bdf44bcc39c4a1be3df4b1f861169b99177054f42e1e9fb524.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 366, + 272, + 434, + 281 + ], + "lines": [ + { + "bbox": [ + 366, + 272, + 434, + 281 + ], + "spans": [ + { + "bbox": [ + 366, + 272, + 434, + 281 + ], + "type": "text", + "content": "(d) Image Prediction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 449, + 207, + 544, + 271 + ], + "blocks": [ + { + "bbox": [ + 449, + 207, + 544, + 271 + ], + "lines": [ + { + "bbox": [ + 449, + 207, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 449, + 207, + 544, + 271 + ], + "type": "image", + "image_path": "5fefbcbbcf228e20e0e41c347524ef003e0947c2c1a491e8303dfef32b023f28.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 464, + 272, + 529, + 282 + ], + "lines": [ + { + "bbox": [ + 464, + 272, + 529, + 282 + ], + "spans": [ + { + "bbox": [ + 464, + 272, + 529, + 282 + ], + "type": "text", + "content": "(e)VDG-Uni3DSeg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "text", + "content": "the Description Query is removed. This demonstrates the benefit of detailed descriptions generated by LLMs in our method. Furthermore, incorporating reference images helps bridge the gap between textual descriptions and the 3D point cloud. By integrating description and image queries, our method enables each point to associate with the most relevant class representation, leading to overall improvements across all segmentation tasks. Finally, removing the SVC Loss causes the most significant drop in instance and panoptic segmentation metrics, including a 2.6-point decline in " + }, + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{25}" + }, + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "text", + "content": " and a 6.9-point drop in " + }, + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "inline_equation", + "content": "\\mathrm{PQ_{th}}" + }, + { + "bbox": [ + 54, + 345, + 297, + 537 + ], + "type": "text", + "content": ". This demonstrates that SVC Loss encourage the model to learn more discriminative segmentation features by leveraging multimodal queries as anchors for feature alignment. In general, each component of our method contributes to improving feature alignment and overall segmentation performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 550, + 167, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 167, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 167, + 562 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Fig. 3 illustrates the semantic segmentation results of different methods. Compared to OneFormer3D [27], both the Description Prediction and Image Prediction modules improve segmentation accuracy by leveraging reference queries, with notable improvements in identifying tables. However, these individual modules struggle with segmenting complex objects such as chairs and doors. When combining description and image predictions, as in VDG-Uni3DSeg, the segmentation results exhibit more balanced and comprehensive improvements across all object categories. This integration results in more precise and consistent segmentation, demonstrating that while individual queries contribute to perfor" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 345, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 381 + ], + "type": "text", + "content": "mance gains, their combined use in VDG-Uni3DSeg yields the most accurate and robust results. Other results are provided in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 388, + 406 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 414, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 414, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 312, + 414, + 555, + 643 + ], + "type": "text", + "content": "We present VDG-Uni3DSeg, a unified 3D segmentation framework that integrates LLM-generated descriptions and internet-sourced images as reference cues to enhance segmentation performance. By leveraging multimodal information, our method improves class and instance differentiation without requiring human annotations or paired multimodal data. To further refine segmentation, we introduce the Semantic-Visual Contrastive (SVC) Loss, which aligns point features with multimodal queries to enhance feature distinctiveness. Additionally, our Spatial Enhancement Module effectively captures long-range spatial relationships, ensuring precise segmentation boundaries while balancing computational cost. Extensive experiments on multiple datasets demonstrate that VDG-Uni3DSeg achieves state-of-the-art performance across semantic, instance, and panoptic segmentation tasks. These results highlight the effectiveness of multimodal guidance in 3D segmentation and pave the way for future research in leveraging external knowledge to further improve 3D scene understanding." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 656, + 419, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 419, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 419, + 670 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "We would like to express our sincere gratitude to all the reviewers for their valuable and constructive suggestions, which have significantly improved the quality of this work." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24842" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 146 + ], + "type": "text", + "content": "[1] Salwa Al Khatib, Mohamed El Amine Boudjoghra, Jean Lahoud, and Fahad Shahbaz Khan. 3d instance segmentation via enhanced spatial and semantic supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 541-550, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "text", + "content": "[4] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3d instance segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15467-15476, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 360 + ], + "type": "text", + "content": "[5] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 294, + 407 + ], + "type": "text", + "content": "[6] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "text", + "content": "[7] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "text", + "content": "[8] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1290–1299, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 294, + 555 + ], + "type": "text", + "content": "[9] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 610 + ], + "type": "text", + "content": "[10] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[11] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] MMDetection3D Contributors. MMDetection3D: OpenMMLab next-generation platform for general 3D object detection. https://github.com/open-mmlab/mmdetection3d, 2020.5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "text", + "content": "[14] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 185, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 239 + ], + "type": "text", + "content": "[15] Francis Engelmann, Martin Bokeloh, Alireza Fathi, Bastian Leibe, and Matthias Nießner. 3d-mpa: Multi-proposal aggregation for 3d semantic instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9031-9040, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 240, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 554, + 282 + ], + "type": "text", + "content": "[16] Lei Han, Tian Zheng, Lan Xu, and Lu Fang. Occuseg: Occupancy-aware 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2940-2949, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 337 + ], + "type": "text", + "content": "[17] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3d: Robust instance segmentation of 3d point clouds through dynamic convolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 354-363, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[18] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "text", + "content": "[19] Ji Hou, Angela Dai, and Matthias Nießner. 3d-sis: 3d semantic instance segmentation of rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4421-4430, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 427, + 554, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 554, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 554, + 481 + ], + "type": "text", + "content": "[20] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 483, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 555, + 536 + ], + "type": "text", + "content": "[21] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2989-2998, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 592 + ], + "type": "text", + "content": "[22] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 646 + ], + "type": "text", + "content": "[23] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition, pages 4867-4876, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 554, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 554, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 554, + 691 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dollar. Panoptic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9404-9413, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 693, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 693, + 554, + 713 + ], + "type": "text", + "content": "[25] Maksim Kolodiazhnyi, Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Top-down beats bottom-up in 3d" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24843" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "instance segmentation. arXiv preprint arXiv:2302.02871, 2023.6,7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 151 + ], + "type": "text", + "content": "[26] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[27] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[28] Jean Lahoud, Bernard Ghanem, Marc Pollefeys, and Martin R Oswald. 3d instance segmentation via multi-task metric learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9256-9266, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 307 + ], + "type": "text", + "content": "[29] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11611-11620, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[30] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[31] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3d scenes using semantic superpoint tree networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2783-2792, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 296, + 476 + ], + "type": "text", + "content": "[32] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17682-17691, 2023. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "text", + "content": "[33] Shih-Hung Liu, Shang-Yi Yu, Shao-Chi Wu, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning gaussian instance segmentation in point clouds. arXiv preprint arXiv:2007.09860, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[34] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4205-4212. IEEE, 2019. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[35] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13550-13559, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[36] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3, 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[39] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[40] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In European Conference on Computer Vision, pages 125-141. Springer, 2022. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "text", + "content": "[41] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask Transformer for 3D Semantic Instance Segmentation. In International Conference on Robotics and Automation (ICRA), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[42] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[43] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[44] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3d scene instance segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2393-2401, 2023. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 521 + ], + "type": "text", + "content": "[45] Weiwei Sun, Daniel Rebain, Renjie Liao, Vladimir Tankovich, Soroosh Yazdani, Kwang Moo Yi, and Andrea Tagliasacchi. Neuralbf: Neural bilateral filtering for top-down instance segmentation on point clouds. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 551-560, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 587 + ], + "type": "text", + "content": "[46] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 644 + ], + "type": "text", + "content": "[47] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 689 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24844" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2708-2717, 2022. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 162 + ], + "type": "text", + "content": "[50] Weiyue Wang, Ronald Yu, Qiangui Huang, and Ulrich Neumann. Sgpn: Similarity group proposal network for 3d point cloud instance segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2569-2578, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 218 + ], + "type": "text", + "content": "[51] Shun-Cheng Wu, Johanna Wald, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scenegraphfusion: Incremental 3d scene graph prediction from rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7515-7525, 2021. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 262 + ], + "type": "text", + "content": "[52] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[53] Yizheng Wu, Min Shi, Shuaiyuan Du, Hao Lu, Zhiguo Cao, and Weicai Zhong. 3d instances as 1d kernels. In European Conference on Computer Vision, pages 235-252. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 376 + ], + "type": "text", + "content": "[54] Zihao Xiao, Longlong Jing, Shangxuan Wu, Alex Zihao Zhu, Jingwei Ji, Chiyu Max Jiang, Wei-Chih Hung, Thomas Funkhouser, Weicheng Kuo, Anelia Angelova, et al. 3d open-vocabulary panoptic segmentation with 2d-3d vision-language distillation. In European Conference on Computer Vision, pages 21-38. Springer, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 432 + ], + "type": "text", + "content": "[55] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3173-3182, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "text", + "content": "[56] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3d instance segmentation on point clouds. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 296, + 543 + ], + "type": "text", + "content": "[57] Zhiliu Yang and Chen Liu. Tupper-map: Temporal and unified panoptic perception for 3d metric-semantic mapping. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1094–1101. IEEE, 2021. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 600 + ], + "type": "text", + "content": "[58] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3947-3956, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 644 + ], + "type": "text", + "content": "[59] Biao Zhang and Peter Wonka. Point cloud instance segmentation using probabilistic embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8883-8892, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 296, + 690 + ], + "type": "text", + "content": "[60] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 715 + ], + "type": "text", + "content": "[61] Weiguang Zhao, Yuyao Yan, Chaolong Yang, Jianan Ye, Xi Yang, and Kaizhu Huang. Divide and conquer: 3d point" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 555, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 555, + 107 + ], + "type": "text", + "content": "cloud instance segmentation with point-wise binarization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 562-571, 2023. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24845" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_content_list.json b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..233bc4e59d15c1130e92795ff9a87296a0743b9e --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_content_list.json @@ -0,0 +1,1803 @@ +[ + { + "type": "text", + "text": "AllGCD: Leveraging All Unlabeled Data for Generalized Category Discovery", + "text_level": 1, + "bbox": [ + 109, + 130, + 885, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinzi Cao $^{1,2}$ , Ke Chen $^{1,2}$ , Feidiao Yang $^{2}$ , Xiawu Zheng $^{3}$ , Yonghong Tian $^{2,4}$ , Yutong Lu $^{1*}$", + "bbox": [ + 153, + 179, + 851, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Sun Yat-sen University, 2 Peng Cheng Laboratory, 3 Xiamen University, 4 Peking University", + "bbox": [ + 132, + 200, + 866, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{caoxz, chenk68}@mail2.sysu.edu.cn yangfd@pcl.ac.cn", + "bbox": [ + 264, + 224, + 730, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhengxiawu@xmu.edu.cn yhtian@pku.edu.cn luyutong@mail.sysu.edu.cn", + "bbox": [ + 191, + 244, + 795, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generalized Category Discovery (GCD) aims to identify both known and novel categories in unlabeled data by leveraging knowledge from labeled datasets. Current methods employ supervised contrastive learning on labeled data to capture known category structures but neglect unlabeled data, limiting their effectiveness in classifying novel classes, especially in fine-grained open-set detection where subtle class differences are crucial. To address this issue, we propose a novel learning approach, AllGCD, which seamlessly integrates all unlabeled data into contrastive learning to enhance the discrimination of novel classes. Specifically, we introduce two key techniques: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first refines intra-class compactness within known categories by integrating potential known samples into labeled data. This process refines the decision boundaries of known categories, reducing ambiguity when distinguishing novel categories. Building on this, Inter-CU further strengthens inter-class separation between known and novel categories by applying global contrastive learning to the class distribution in the unlabeled data. By jointly leveraging Intra-CL and Inter-CU, AllGCD effectively improves both intra-class compactness and inter-class separation, effectively enhancing the discriminability between known and novel classes. Experiments demonstrate that AllGCD significantly improves novel classes accuracy, e.g., achieving increases of $7.4\\%$ on CUB and $7.5\\%$ on Stanford Cars.", + "bbox": [ + 89, + 325, + 483, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 776, + 220, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved outstanding performance in computer vision tasks [4, 7, 15, 31, 32, 36, 40], especially in image classification [14, 16, 20, 33, 34, 36]. However, most conventional methods are developed in a closed setting where all training and test data belong to pre-defined", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fa28c778791a9f66da1ff2941555d0ed0b452a00512dd08b112e10a483e9ba7.jpg", + "image_caption": [ + "(a) 'New' acc w/o CL in SimGCD" + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 712, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f56d58f8e2732f1cc23fca788ac7fc9a6c657ad3308aee748903890a3907b090.jpg", + "image_caption": [ + "(b) Attentions w/o unlabeled data", + "Figure 1. Comparison of 'New' accuracy and attention w/o unlabeled data in supervised contrastive learning (S-CL). (a) reveals that S-CL solely with labeled data fails to classify novel samples (orange bar vs. blue bar) in the baseline SimGCD. In contrast, our method with unlabeled data (green bar) significantly improves accuracy. (b) shows that our AllGCD highlights more class-specific object regions by incorporating unlabeled data." + ], + "image_footnote": [], + "bbox": [ + 718, + 292, + 903, + 414 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "known) classes. It weakens the capability of these methods to handle novel categories, limiting the scope of their applications [8, 11, 26, 46]. To overcome this issue, the Generalized Category Discovery (GCD) paradigm [1, 3, 28, 41, 44, 45, 49, 50] was introduced for open-world learning. It identifies both known and novel categories from unlabeled data by leveraging knowledge learned from labeled data with known classes.", + "bbox": [ + 511, + 566, + 906, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key strategy for discovering novel classes is to transfer knowledge from labeled data to promote the learning of novel classes. Based on this idea, previous GCD methods [3, 41, 42, 44, 45] utilize contrastive learning [17, 39] to learn the representations of classification within labeled data. Initially, Vaze et al. [41] adopt a two-stage pipeline: contrastive learning for representation, followed by $k$ -means for classifying unlabeled data. However, clustering becomes computationally expensive as data scales. To reduce this, Wen et al. [45] proposed SimGCD, a parametric approach that replaces clustering with a classifier and learns class prototypes via joint contrastive learning. Building on this, Wang et al. [44] enhanced SimGCD with data prompts for better performance, while Cao et al. [3]", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3293", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "addressed SimGCD's limitation in retaining known-class knowledge during transfer by introducing a regularization that improves both known and novel class performance.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite advancements in parametric methods, recent studies reveal that supervised contrastive learning (S-CL), a key component of contrastive learning, has a limited effect in discovering novel classes. This is evident from the SimGCD baseline in Fig. 1a, where 'New' accuracy shows little improvement between that without S-CL and S-CL with only labeled data (orange vs blue bars) across both fine-grained and general datasets. In particular, accuracy even slightly declines when S-CL is limited to labeled data, as observed in Stanford Cars and CIFAR100. This indicates that current S-CL methods, constrained to labeled data, are inadequate for discovering novel classes, as the labeled data encompasses only known classes and omit the novel ones. Therefore, the key to improving the discovery of novel categories is to incorporate unlabeled data into S-CL, as it contains both known and novel samples.", + "bbox": [ + 89, + 138, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this observation, we propose AllGCD, a straightforward approach that integrates all unlabeled data into contrastive learning to facilitate the novel category discovery. Specifically, AllGCD comprises two main components: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). To establish a stronger foundation for Inter-CU, Intra-CL first enhances intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data. Building on this refined knowledge, Inter-CU further improves inter-class separation between known and novel categories through global contrastive learning on their class distribution in unlabeled data. Finally, by combining Intra-CL and Inter-CU, we significantly improve novel category representations, as shown by the green bar in Fig. 1a.", + "bbox": [ + 89, + 382, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our approach on six datasets, including three fine-grained datasets (CUB [43], Stanford Cars [18], FGVC-Aircraft [25]), three generic image recognition datasets (CIFAR100 [19] and ImageNet-100 [6]), and the challenging Herbarium-19 dataset [37]. As shown in Fig. 1a, our method outperforms strong baselines in recognizing novel ('New') classes, with a $7.4\\%$ improvement over SimGCD on CUB. Additionally, Fig. 1b highlights that AllGCD focuses on more class-specific regions than the baseline. In summary, our main contributions are as follows:", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose AllGCD to overcome the limitations of contrastive learning in parametric methods by utilizing all unlabeled data for novel class discovery.", + "- We introduce Intra-class Contrast in Labeled Data (IntraCL) to enhance intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data.", + "- Building on this refined knowledge, we design Inter-class" + ], + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrast in Unlabeled Data (Inter-CU) to improve interclass separation between known and novel categories by leveraging the global class distribution in unlabeled data.", + "bbox": [ + 526, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive results show that our method significantly improves the performance on novel classes, achieving a $7.4\\%$ increase on CUB compared to the baseline.", + "bbox": [ + 514, + 136, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 195, + 663, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel Category Discovery (NCD) [12] assumes unlabeled data contains only novel classes and was first formalized as deep transfer clustering, where novel classes are clustered using knowledge from known classes. Various methods address this challenge, including Han et al. [12, 13] who use self-supervised learning and ranking statistics for knowledge transfer, Zhong et al. [52] who propose a mixup strategy [48] to combine old and novel classes, Gu et al. [10] who introduce a knowledge distillation framework for regularizing novel class learning, Zang et al. [47] who transfer knowledge from a source domain to a label-scarce target domain without label constraints, and Liu et al. [23] who apply NCD in ultra-fine-grained visual categorization.", + "bbox": [ + 511, + 220, + 906, + 417 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalized Category Discovery (GCD) [41] extends NCD by assuming that unlabeled data also contains known class images. Vaze et al. [41] formalize this with contrastive learning in representation space during training and clustering at inference. To reduce clustering cost, Wen et al. [45] introduce SimGCD, using a parametric classifier to learn class representation centers, serving as a strong baseline. Cao et al. [3] note SimGCD's issue of forgetting known classes while learning novel ones, proposing entropy regularization for potential known samples. Wang et al. [44] further extend SimGCD by optimizing model and data parameters for better alignment with the GCD task. Despite their performance, these methods rely on limited labeled images, constraining novel class discovery.", + "bbox": [ + 511, + 417, + 905, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning (CL) spans unsupervised [39] and supervised [17] variants for learning from unlabeled and labeled data, respectively. Unsupervised CL treats two augmented views of the same image as positives and all others as negatives, while supervised CL (S-CL) considers samples from the same class as positives. Vaze et al. [41] first combine both in GCD to enhance image representations. Choi et al. [5] employ mean-shifted embeddings for CL, and Pu et al. [28] propose DCCL, which explores conceptual-level CL to model relations between known and novel categories. However, these methods primarily apply S-CL within non-parametric GCD frameworks. In contrast, recent parametric approaches [3, 44, 45] demonstrate that S-CL can effectively train classifiers by learning class prototypes. Nonetheless, these works still rely heavily on limited labeled data, underutilizing the rich supervision signal from abundant unlabeled samples, which ultimately constrains the full potential of contrastive learning.", + "bbox": [ + 511, + 628, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3294", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries and Analysis", + "text_level": 1, + "bbox": [ + 89, + 89, + 341, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. GCD setting and training", + "text_level": 1, + "bbox": [ + 89, + 114, + 323, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Unlike traditional closed-set classification, where training and test classes coincide $(\\mathcal{V}_l)$ , GCD aims to identify novel classes $\\mathcal{Y}_u$ in the test data. Given labeled dataset $\\mathcal{D}^l = \\{(\\pmb{x}_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_l$ and unlabeled dataset $\\mathcal{D}^u = \\{(x_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_u$ , GCD seeks to classify both known and novel samples in $\\mathcal{D}^u$ using knowledge from $\\mathcal{D}^l$ . The total class number $K = |\\mathcal{Y}_l\\cup \\mathcal{Y}_u|$ is assumed known, following prior work [3, 9, 13, 44, 45].", + "bbox": [ + 89, + 136, + 483, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. SimGCD [45] consists of a Vision Transformer (ViT) encoder [7], denoted as $\\Phi$ , along with a classification head $f$ to generate pseudo-labels for self-distillation [2, 4], and a multi-layer perception (MLP) projection head $g$ for representation contrastive learning.", + "bbox": [ + 89, + 257, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Representation Learning. To learn representations for distinguishing samples, Vaze et al. [41] apply unsupervised contrastive learning across all data and supervised contrastive learning (S-CL) on labeled data. Given two views, $\\boldsymbol{x}_i$ and $\\boldsymbol{x}_i'$ , of an identical image as a positive pair in a minibatch $B$ , the unsupervised contrastive loss encourages their representations to be close in an unsupervised manner:", + "bbox": [ + 89, + 333, + 483, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e p}} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {i} ^ {\\prime} / \\tau_ {u}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {u}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 449, + 482, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\boldsymbol{z}_i = g \\circ \\Phi(\\boldsymbol{x}_i)$ and $\\tau_u$ is a temperature. Meanwhile, the supervised contrastive loss minimizes the distance between same-class samples in feature space, defined as:", + "bbox": [ + 89, + 498, + 483, + 544 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e p}} ^ {s} = \\frac {1}{| B ^ {l} |} \\sum_ {i \\in B ^ {l}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {q} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 555, + 482, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B^l$ is the labeled subset of $B$ , and $\\mathcal{N}_i$ is the set of images with the same label as $x_i$ . The overall contrastive learning loss for representation is:", + "bbox": [ + 89, + 607, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e p}} = (1 - \\lambda) \\mathcal {L} _ {\\text {r e p}} ^ {u} + \\lambda \\mathcal {L} _ {\\text {r e p}} ^ {s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 665, + 482, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the balance between unsupervised and supervised contrastive learning.", + "bbox": [ + 89, + 694, + 482, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parametric Clustering. Instead of the time-consuming $k$ -means, Wen et al. [45] propose an efficient parametric classifier $f$ using a set of prototypes $\\mathcal{C} = \\{c_1, \\dots, c_K\\}$ , where $K$ is the total classes number. For an image view $x_i$ , the prediction is made by calculating the cosine similarity between its feature, $\\Phi(x_i)$ , and each prototype $c_k$ :", + "bbox": [ + 89, + 724, + 483, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {i} ^ {(k)} = \\frac {\\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k}) / \\tau_ {s})}{\\sum_ {k ^ {\\prime}} \\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k ^ {\\prime}}) / \\tau_ {s})}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 825, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, the soft pseudo-labels $\\pmb{q}_i'$ for another view $\\pmb{x}_i'$ are obtained using a smaller $\\tau_t$ , and the cross-entropy loss is", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "applied for prototype learning using these pseudo-labels $q_{i}^{\\prime}$ :", + "bbox": [ + 513, + 90, + 903, + 107 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} = - \\frac {1}{| B |} \\sum_ {i \\in B} \\sum_ {k} \\boldsymbol {q} ^ {\\prime (k)} \\log \\boldsymbol {p} ^ {(k)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 117, + 903, + 154 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, SimGCD employs a class mean entropy regularization [2], $H(\\overline{p}) = -\\sum_{k} \\overline{p}^{(k)} \\log \\overline{p}^{(k)}$ , where $\\overline{p}$ is the batch-wise class-average prediction, forcing the model to focus on novel classes. The total classifier is:", + "bbox": [ + 511, + 165, + 906, + 224 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c l s}} = \\lambda \\mathcal {L} _ {\\mathrm {c l s}} ^ {s} + (1 - \\lambda) \\left(\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} - \\varepsilon H\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 238, + 903, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{cls}}^s$ is the cross-entropy loss for labeled data.", + "bbox": [ + 511, + 267, + 859, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limited contrastive learning", + "text_level": 1, + "bbox": [ + 513, + 292, + 769, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis. Although contrastive learning (CL) enhances representation discriminability of known classes in", + "bbox": [ + 511, + 314, + 905, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SimGCD, it applies only to labeled data, as shown in the upper-left corner of Fig. 2. In the upper-left corner, same-colored entries along the vertical and horizontal indices correspond to samples from the same class, while different colors denote distinct classes. Self-contrast cases (blue-filled) on the diagonal are masked out. Obviously, abundant unlabeled data, shown with a gray background, are excluded from the", + "bbox": [ + 511, + 345, + 723, + 555 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65ce49b823f47ba9ad2363b27c39efa9647633ddf83df8bcd55202838d8cb719.jpg", + "image_caption": [ + "Figure 2. A class mask: Samples from the same class are shown in the same color in a batch $B$ ." + ], + "image_footnote": [], + "bbox": [ + 733, + 359, + 906, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CL due to the absence of class labels. As a result, the model lacks direct contrastive supervision on novel samples, making it harder to separate them in feature space.", + "bbox": [ + 511, + 556, + 905, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 614, + 604, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome the limited contrastive learning issue discussed in subsection 3.2, we propose a straightforward approach called AllGCD to integrate supervised contrastive learning across all unlabeled data directly. Specifically, we detail our proposed Intra-class Contrast in Labeled Data (Intra-CL) in subsection 4.1 and Inter-class Contrast in Unlabeled Data (Inter-CU) in subsection 4.2.", + "bbox": [ + 511, + 641, + 906, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Intra-class Contrast in Labeled Data", + "text_level": 1, + "bbox": [ + 511, + 757, + 831, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intra-CL first strengthens intra-class compactness within known classes for better separation of known and novel categories in Inter-CU. To achieve this, we use baseline SimGCD [45] to identify potential known samples in unlabeled data and incorporate them into the labeled set. However, as shown by the orange lines in Fig. 4a and Fig. 4b, SimGCD's known-class accuracy declines sharply during training, compromising the quality of potential known", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3295", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/27fa5cf4ef6b67909731898e0504ed3b87a3d8293e4ce126ac0952b7f948f2d5.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed AllGCD. It is composed of Intra-CL and Inter-CU. First, Intra-CL (Sec. 4.1) enhances knowledge of known classes by incrementally collecting reliable pseudo-labels $\\mathcal{V}^*$ within unlabeled data. Once this knowledge is strengthened, Inter-CU (Sec. 4.2) further promotes the separation of known and novel classes through contrastive learning between them." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "samples. To overcome this, we propose a voting strategy that utilizes the model's prior knowledge to identify potential known samples early. Specifically, Intra-CL comprises three steps: (1) Selecting potential known samples within unlabeled data as in [3]; (2) Storing these known samples into a dictionary $\\mathcal{Q}$ ; (3) Voting pseudo-labels $\\mathcal{Y}^*$ from $\\mathcal{Q}$ based on their most frequent label across training iterations.", + "bbox": [ + 88, + 473, + 480, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step (1): Selecting potential known samples within unlabeled data.", + "bbox": [ + 89, + 579, + 482, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this step, we identify potential known samples by selecting instances where both views of the same image exhibit high confidence $(\\delta)$ and are classified as known categories $(\\mathcal{V}_l)$ . The specific calculation process is as follows.", + "bbox": [ + 89, + 609, + 482, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $B^u$ denote the unlabeled subset of $B$ , where each image in this batch has a unique index $\\mathcal{U}_B = \\{u_1, u_2, \\ldots, u_b\\}$ . Given the predictions $p_i$ and $q_i'$ from two views of an image in $B^u$ , as defined in Eq. (4). We first identify high-confidence samples by selecting those with predictions $p_i$ and $q_i'$ both exceed a confidence threshold $\\delta$ : $\\mathbf{M}_{\\delta} = \\mathbb{1}(\\max(p_i) > \\delta \\land \\max(q_i') > \\delta)$ .", + "bbox": [ + 89, + 670, + 482, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we classify a sample as known only if the predicted label from the two views match and both belong to known class set $\\mathcal{Y}_l$ : $\\mathbf{M}_{\\mathrm{known}} = \\mathbb{1}((\\hat{y}_i = \\hat{y}_i') \\wedge (\\hat{y}_i \\in \\mathcal{Y}_l))$ .", + "bbox": [ + 89, + 775, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the selected known samples within the unlabeled subset $B^{u}$ are determined as:", + "bbox": [ + 89, + 821, + 482, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} = \\left(\\mathbf {M} _ {\\delta} \\odot \\mathbf {M} _ {\\text {k n o w n}}\\right) \\odot \\hat {\\mathcal {Y}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 857, + 482, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\delta}$ and $\\mathbf{M}_{\\mathrm{known}}$ are binary (0-1) vectors of length b,", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\odot$ is element-wise multiplication, and $\\hat{\\mathcal{Y}} = \\{\\hat{y}_i \\mid i \\in [1,b]\\}$ represents the predicted labels across the batch of unlabeled samples. Thus, $\\hat{\\mathcal{Y}}_{\\mathrm{select}}$ identifies the known samples $(\\mathbf{M}_{\\mathrm{known}} = 1)$ that also have high confidence $(\\mathbf{M}_{\\delta} = 1)$ .", + "bbox": [ + 511, + 472, + 903, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step (2): Storing above predicted known samples.", + "bbox": [ + 531, + 534, + 882, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To leverage the model's early ability to recognize known classes, we initialize a dictionary $\\mathcal{Q}$ to store the pseudolabels $\\hat{\\mathcal{Y}}_{\\mathrm{select}}$ of each image based on its unique index $\\mathcal{U}_B$ .", + "bbox": [ + 511, + 549, + 903, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, for each unlabeled subset $B^u$ , we update the dictionary $\\mathcal{Q}:\\mathcal{U}_B\\to \\hat{\\mathcal{V}}_{\\mathrm{select}}$ as:", + "bbox": [ + 511, + 594, + 903, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} \\left(u _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {Q} \\left(u _ {i}\\right) \\cup \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\in \\operatorname {d o m} (\\mathcal {Q}), \\\\ \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\notin \\operatorname {d o m} (\\mathcal {Q}). \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 635, + 903, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step (3): Voting on pseudo-labels for unlabeled data based on frequency.", + "bbox": [ + 511, + 686, + 903, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain a reliable pseudo-label for each unlabeled image, we apply a voting mechanism over the stored pseudolabels $\\mathcal{Q}$ across training batches. Specifically, for a batch of unlabeled samples, we select the most frequently assigned pseudo-label from its historical records in $\\mathcal{Q}(u_i)$ :", + "bbox": [ + 511, + 715, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} ^ {*} = \\underset {y} {\\arg \\max } \\left(\\operatorname {f r e q} (\\hat {y} \\in \\mathcal {Q} (u _ {i})), \\forall u _ {i} \\in \\mathcal {U} _ {B}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 803, + 903, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{Y}^* = \\{y_1^*, y_2^*, \\ldots, y_b^*\\}$ denotes the pseudo-labels for unlabeled samples in $B^u$ . The updated mini-batch labels, incorporating both original labeled data $\\mathcal{Y}_l$ and pseudolabels $\\mathcal{Y}^*$ , are given by $\\mathcal{Y}_{\\mathrm{intra}} = \\mathcal{Y}_l \\cup \\mathcal{Y}^*$ . Consequently,", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3296", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a7ad01421434797eab5d06eab787b1bee2bde3f215b9297fe8f642b17b9118f1.jpg", + "image_caption": [ + "(a) CUB" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 285, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7637b0bf5a8c93340ed67f55718bf7797a8125cfe8b45ab42085f6b4d443fc3f.jpg", + "image_caption": [ + "(b) Stanford Cars", + "Figure 4. Known-class accuracy on unlabeled data during training. SimGCD, using S-CL without unlabeled data, shows a sharp drop on CUB and Stanford Cars. In contrast, Intra-class leverages potential known samples in unlabeled data, improving the knowledge of known classes." + ], + "image_footnote": [], + "bbox": [ + 289, + 88, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the supervised contrastive loss in Eq. (2) is updated with pseudo known samples from unlabeled data as:", + "bbox": [ + 89, + 316, + 483, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t r a}} = \\frac {1}{| B ^ {\\text {i n t r a}} |} \\sum_ {i \\in B ^ {\\text {i n t r a}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 359, + 488, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B^{\\mathrm{intra}}$ is the newly labeled subset of mini-batch $B$ with labels $\\mathcal{V}_{\\mathrm{intra}}$ , and $\\mathcal{N}_i$ indexes same-label images as $x_i$ in the batch. After applying Intra-CL, the accuracy of known classes in unlabeled data remains more stable (Fig. 4), indicating more compact representations within known classes.", + "bbox": [ + 89, + 412, + 483, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Inter-class Contrast in Unlabeled Data", + "text_level": 1, + "bbox": [ + 89, + 513, + 421, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While Inter-CL enhances discrimination among known classes, it lacks contrastive interactions with novel classes, limiting novel category accuracy (Fig. 1a). To address this, we propose Inter-CU, which enforces global contrastive learning on the unlabeled class distribution. Specifically, once the model has stabilized on known and novel categories from epoch $e$ , we use the Hungarian algorithm [21] on the predictions of all unlabeled data to optimize the class distribution of known and novel categories. To ensure a high-quality distribution, we retain only high-confidence predictions $(P > \\gamma)$ for global contrastive learning.", + "bbox": [ + 89, + 536, + 483, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formally, given the unlabeled dataset $\\mathcal{D}^u = \\{(x_i,y_i)\\} \\in \\mathcal{X}\\times \\mathcal{Y}_u$ we obtain classifier predictions $P = [p_1,p_2,\\dots ,p_N]$ via Eq. (4), where $N$ is the total number of unlabeled samples. The minimum-cost assignment function $\\pi$ is defined as:", + "bbox": [ + 89, + 702, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {*} = \\arg \\min _ {\\pi} \\sum_ {i = 1} ^ {N} - P _ {i, \\pi (i)}, \\quad \\pi (i) \\in \\{1, 2, \\dots , K \\}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 787, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pi^{*}$ assigns each sample in $\\mathcal{D}^u$ to one of $K$ categories, minimizing the sum of costs $-P_{i,\\pi (i)}$ . The Hungarian algorithm efficiently solves this assignment problem by finding the optimal mapping with minimal total cost. Therefore, the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 One Training Procedure for AllGCD" + ], + "code_body": "Input: Encoder $\\Phi$ , classifier head $f$ , projector $g$ , labeled data $\\mathcal{D}^l$ , unlabeled data $\\mathcal{D}^u$ , hyperparameters $\\delta, e$ . \nOutput: $\\Phi, f$ , and $g$ ; \n1: Initialize dictionary $\\mathcal{Q} \\gets \\emptyset$ ; \n2: for epoch = 1 to total epochs do \n3: if epoch ≥ e then \n4: Get assignment $\\mathcal{Y}_{\\text{inter}}$ for $D^u$ using Eq. (12); \n5: end if \n6: for $(x, x') \\in \\mathcal{D}^l \\cup \\mathcal{D}^u$ , $y \\in \\mathcal{Y}_l$ , $y_{\\text{inter}} \\in \\mathcal{Y}_{\\text{inter}}$ do \n7: $z \\gets g \\circ \\Phi(x)$ , $z' \\gets g \\circ \\Phi(x')$ ; \n8: $p \\gets f \\circ \\Phi(x)$ , $q' \\gets f \\circ \\Phi(x')$ ; \n9: Compute $\\mathcal{L}_{\\text{cls}}^s(p, y) + \\mathcal{L}_{\\text{cls}}^s(q', y)$ ; ▷ For $\\mathcal{D}^l$ \n10: Compute $\\mathcal{L}_{\\text{dis}}^u(p, q')$ using Eq. (5); \n11: Compute Max class Entropy $H(p)$ ; \n12: Compute $\\mathcal{L}_{\\text{cls}} \\gets \\lambda \\mathcal{L}_{\\text{cls}}^s + (1 - \\lambda) (\\mathcal{L}_{\\text{dis}}^u - \\varepsilon H)$ ; \n13: Step 1: Generate $\\hat{y} \\gets (p, q', \\delta)$ ; ▷ For $\\mathcal{D}^u$ \n14: Step 2: Store $\\hat{y}$ into $\\mathcal{Q}$ using Eq. (8); \n15: Step 3: Vote for $y^*$ from $\\mathcal{Q}$ using Eq. (9); \n16: Combine with pseudo-labels $y_{\\text{intra}} \\gets y \\cup y^*$ ; \n17: Compute $\\mathcal{L}_{\\text{rep}}^u(z, z')$ using Eq. (1); \n18: Compute $\\mathcal{L}_{\\text{intra}}(z, z', y_{\\text{intra}})$ using Eq. (10); \n19: Compute $\\mathcal{L}_{\\text{inter}}(z, z', y_{\\text{inter}})$ using Eq. (12); \n20: $\\mathcal{L}_{\\text{all}} = (1 - \\lambda) \\mathcal{L}_{\\text{rep}}^u + \\lambda (\\mathcal{L}_{\\text{intra}} + \\mathcal{L}_{\\text{inter}}) + \\mathcal{L}_{\\text{cls}}$ ; \n21: Update $\\Phi, f$ , and $g$ by SGD [29]; \n22: end for \n23: end for", + "bbox": [ + 516, + 109, + 906, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "final assignments for unlabeled data are:", + "bbox": [ + 513, + 532, + 782, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Y} _ {\\text {i n t e r}} = \\left\\{\\hat {y} _ {i} = \\pi^ {*} (i) \\mid i = 1, 2, \\dots , N \\right\\} \\odot \\mathbf {M} _ {\\text {i n t e r}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 561, + 906, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{M}_{\\mathrm{inter}}$ is a binary mask that applies a confidence threshold $\\gamma$ to filters out low-confidence predictions: $\\mathbf{M}_{\\mathrm{inter}} = \\{i \\mid \\max(\\pmb{p}_i) > \\gamma\\}$ . Only high-confidence samples contribute to the global contrastive loss:", + "bbox": [ + 511, + 592, + 906, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t e r}} = \\frac {1}{| B ^ {\\text {i n t e r}} |} \\sum_ {i \\in B ^ {\\text {i n t e r}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 666, + 919, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B^{\\mathrm{inter}}$ corresponds to the refined labeled subset with pseudo-labels $\\mathcal{V}_{\\mathrm{inter}}$ . This approach enables global contrastive learning on both known and novel samples, enhancing category discovery, particularly for novel classes.", + "bbox": [ + 511, + 719, + 906, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall. Utilizing all unlabeled data in supervised contrastive learning significantly boosts class discovery, especially for novel classes. This overall training loss is updated as:", + "bbox": [ + 511, + 781, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l l}} = (1 - \\lambda) \\mathcal {L} _ {\\text {i n t r a}} + \\lambda \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {c l s}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 844, + 903, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ controls the trade-off between Intra-CL and Inter-CU. Algorithm 1 details one training step of AllGCD.", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3297", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0160e97e1c46783e2345ebc2ea312fdf32f680d51c44808f96814f8525a0dd66.jpg", + "table_caption": [ + "Table 1. Comparison with State-of-the-Art GCD Methods, Evaluated with and without Ground-Truth $K$ for Clustering. Bold denotes the best results." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
(a) Clustering with the ground-truth number of classes K given
k-means[24]52.052.250.872.775.571.334.338.932.112.810.613.816.014.416.813.012.213.4
RS+[13]58.277.619.337.161.624.833.351.624.228.361.812.126.936.422.227.955.812.8
UNO+[9]69.580.647.270.395.057.935.149.028.135.570.518.640.356.432.228.353.714.7
ORCA[22]69.077.452.073.592.663.935.345.630.223.550.110.722.031.817.120.930.915.5
GCD [41]73.076.266.574.189.866.351.356.648.739.057.629.945.041.146.935.451.027.0
DCCL [28]75.376.870.280.590.576.263.560.864.9---43.155.736.2---
PromptCAL [50]81.284.275.383.192.778.362.964.462.150.270.140.652.252.252.337.052.028.9
SimGCD [45]80.181.277.883.093.177.960.365.657.753.871.945.054.259.151.844.058.036.4
InfoSieve [30]78.382.270.580.593.873.869.477.965.255.774.846.456.363.752.541.055.433.2
SPTNet [44]81.384.375.685.493.281.465.868.865.159.079.249.359.361.858.143.458.735.2
LegoGCD [3]81.881.482.586.394.582.163.871.959.857.375.748.455.061.551.745.157.438.5
CMS [5]82.385.775.584.795.679.268.276.564.056.063.452.356.976.147.636.454.926.4
AllGCD (ours)82.382.382.286.594.782.368.475.165.160.577.052.557.462.454.945.258.338.0
(b) Clustering without the ground-truth number of classes K given
GCD [41]70.877.657.077.991.171.351.156.448.439.158.629.7---37.251.729.4
GPC [51]75.484.660.175.393.466.752.055.547.538.258.927.443.340.744.836.551.727.9
SimGCD [45]80.181.277.881.791.276.861.566.459.149.165.141.3------
CMS [5]79.683.272.381.395.674.264.468.262.451.768.943.455.260.652.437.456.527.1
AllGCD (ours)81.583.378.484.494.380.266.772.164.058.875.750.656.764.253.045.758.738.5
", + "bbox": [ + 94, + 127, + 906, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 435, + 223, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 462, + 279, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our approach on three fine-grained datasets: CUB [43], Stanford Cars [18], and FGVC-Aircraft [25], as well as two generic image recognition datasets, CIFAR100 [19] and ImageNet-100 [6]. Additionally, we test on a more challenging long-tailed fine-grained Herbarium 19 [37]. Following [41], we randomly sample $50\\%$ of seen categories as labeled data $\\mathcal{D}^l$ , with the remaining images from seen and novel categories forming the unlabeled set $\\mathcal{D}^u$ . Details are in the supplementary material.", + "bbox": [ + 89, + 484, + 483, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. We use clustering accuracy (ACC) to evaluate model performance, following [41, 45]. During evaluation, we compare the ground-truth labels $y_{i}$ with the predicted labels $\\hat{y}_i$ , and calculate ACC as $ACC = \\frac{1}{|\\mathcal{D}^u|}\\sum_{i=1}^{|\\mathcal{D}^u|}\\mathbb{1}(y_i = p(\\hat{y}_i))$ , where $p$ is determined using the Hungarian optimal assignment algorithm [21].", + "bbox": [ + 89, + 623, + 483, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. AllGCD is built upon SimGCD [45], using a ViT-B/16 backbone [7] pretrained with DINO [4]. We use the class token to train a classifier and fine-tune only the last attention block. We use a batch size of 128 and an initial learning rate of 0.1, decayed over 200 epochs using a cosine schedule. Following [41], we set $\\lambda = 0.35$ , $\\tau_{u} = 0.07$ , and $\\tau_{c} = 1.0$ for representation learning, and use $\\tau_{s} = 0.1$ , with $\\tau_{t}$ decayed from 0.07 to 0.04 over 30 epochs for classifier training. Pseudo-label thresholds $\\delta$ and $\\gamma$ range from 0.6 to 0.9. Based on the convergence time of 'New' classes in SimGCD (Fig. 7a), the Inter-CU start epoch $e$ is selected", + "bbox": [ + 88, + 719, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "from $\\{0,20,40,60,80\\}$ . All experiments are conducted using PyTorch on NVIDIA Tesla V100 GPUs.", + "bbox": [ + 511, + 436, + 906, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparison With the State of the Arts", + "text_level": 1, + "bbox": [ + 511, + 481, + 846, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of AllGCD, we evaluate it under both known and unknown total class numbers $K$ : (1) In Table 1(a), we assume $K$ is known, consistent with prior works [3, 9, 13, 44, 45]; (2) In Table 1(b), we report results where $K$ is unknown, estimating it using clustering techniques as in [5, 41, 51].", + "bbox": [ + 511, + 503, + 906, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation with ground-truth class number $K$ . Table 1(a) compares state-of-the-art methods assuming a known class number $K$ . Our method significantly enhances 'New' category accuracy compared to the baseline SimGCD, with notable improvements of $7.4\\%$ , $7.5\\%$ , and $4.4\\%$ on CUB, Stanford Cars, and ImageNet-100, respectively. Furthermore, AllGCD remains competitive with SPTNet, surpassing it by $2.6\\%$ and $1.5\\%$ on the 'All' classes in CUB and Stanford Cars, respectively. Importantly, our method attains comparable performance to SPTNet using only 200 training epochs, compared to SPTNet's 1000 epochs, showing superior computational efficiency. Though AllGCD trails CMS [5] by $2.7\\%$ in 'Old' accuracy on CIFAR100, it outperforms CMS by $6.7\\%$ in 'New' accuracy, highlighting its strength in novel category discovery.", + "bbox": [ + 511, + 597, + 908, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation with estimated number of clusters. Table 1 (b) presents a comparison conducted without access to the ground-truth class number $K$ , instead using an estimated $K$ derived from an off-the-shelf method [41], as shown in Table 2. Table 2 shows our approach provides a more accu", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3298", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4b3355ea115230c676dcf5a3324abd862d7c332c409c33988b8ce29d0f84ba30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓
GT100-100-200-196-100-683-
GCD [41]1000109923115.523017.3--52023.8
DCCL [28]146461292917291920.02----
SimGCD [45]1000109923115.523017.3--52023.8
GPC [51]1000103321262010.03----
CMS [5]955116161701515620.490106228.9
AllGCD (ours)95510882136.52107.110996366.9
", + "bbox": [ + 94, + 102, + 632, + 200 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5b23e4e9c3d533c2abd4f0032b5260a0af76f97a98d9a3dbc8ec634a8b82e50.jpg", + "image_caption": [ + "(c) SimGCD [45]" + ], + "image_footnote": [], + "bbox": [ + 635, + 88, + 764, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c4744ce16b86f4a8182e54cd6fef3ff65ca8a488a6c9258548ec8f4d7a00d19.jpg", + "image_caption": [ + "(d) AllGCD" + ], + "image_footnote": [], + "bbox": [ + 772, + 88, + 897, + 184 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c902b2c84bbdf34a1915d2414bafe091fedade6ca1d147a92a6bbfc492a9f360.jpg", + "table_caption": [ + "Table 2. Estimated class number and error rate for $K$ using off-the-self method [41]." + ], + "table_footnote": [], + "table_body": "
δCUBCIFAR100
AllOldNewAllOldNew
0.7068.274.664.980.881.380.5
0.7568.475.165.181.481.881.3
0.8068.275.764.582.283.181.8
0.8567.274.763.482.382.382.2
0.9066.273.362.581.681.981.5
", + "bbox": [ + 94, + 253, + 388, + 356 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Effect of confidence $\\delta$ in Intra-CL.", + "bbox": [ + 109, + 367, + 367, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/075ea49b201b9b00582a424f7d97d9b0d3c73b0393dae54aef3b3d3ec219cee0.jpg", + "image_caption": [ + "(a) CUB with $\\delta = 0.75$", + "Figure 6. Logit distributions of unlabeled data." + ], + "image_footnote": [], + "bbox": [ + 395, + 256, + 643, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f3ebc4713b108c3dcc7eabc7183de9ca176223cb488b99bf5eff325d04aae7b.jpg", + "image_caption": [ + "Figure 5. T-SNE visualization of 10 random novel classes from CIFAR100.", + "(b) CIFAR100 with $\\delta = 0.85$" + ], + "image_footnote": [], + "bbox": [ + 648, + 256, + 897, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rate estimation of $K$ , with fewer errors in determining cluster centers than SimGCD and CMS, except for CIFAR100 in SimGCD. Under this improved estimation, our approach consistently outperforms state-of-the-art methods across all datasets. Notably, performance gains are particularly significant on the Stanford Cars, ImageNet-100, and Herbarium-19 benchmarks, where AllGCD achieves 7.2%, 6.0%, and 9.1% higher accuracy in the 'New' category compared to the second-best results. These findings highlight the effectiveness of our method in utilizing all available unlabeled data to learn more discriminative representations.", + "bbox": [ + 88, + 407, + 483, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Visualizations and Attention Distributions. To demonstrate the effectiveness of our approach in representation learning, we visualize 10 randomly selected novel classes from CIFAR100. As shown in Fig. 5, AllGCD produces more compact and well-separated clusters, whereas SimGCD exhibits considerable cluster overlap. Additionally, we visualize attention on both known and novel samples from unlabeled data in Table 4. Ours focuses on class-specific regions (e.g., the bird's wing) and suppressing background noise (e.g., the window in the second row), highlighting that leveraging unlabeled data enhances attention to relevant regions and improves class discriminability.", + "bbox": [ + 89, + 575, + 483, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 770, + 243, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we present ablation studies to assess the effectiveness of Intra-CL and Inter-CU in our AllGCD model. The evaluation includes fine-grained datasets such as CUB and Stanford Cars, alongside generic image recognition datasets CIFAR100 and ImageNet-100.", + "bbox": [ + 89, + 792, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different confidence $\\delta$ . Table 3 shows the effect of varying the confidence threshold $\\delta$ in Intra-CL on", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54ca2209fb4ecf5c0b695874eb21c5f039e1dd879cf2836559d3e3e377227e71.jpg", + "table_caption": [ + "Table 4. Average attention across 12 heads in the final layer of the image encoder on CUB and FGVC-Aircraft. The known sample (top row) and novel sample (bottom row) in AllGCD show improved focus on object regions with reduced background noise." + ], + "table_footnote": [], + "table_body": "
ImageDINOSimGCDAllGCD
", + "bbox": [ + 516, + 473, + 906, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CUB and CIFAR100. Overall, the accuracy for both 'Old' and 'New' classes initially improves with increasing $\\delta$ but slightly drops beyond the optimal point. The fine-grained CUB dataset performs best at $\\delta = 0.75$ , while the generic dataset CIFAR100 peaks at $\\delta = 0.85$ . This difference arises from the varying proportions of unlabeled samples preserved in fine-grained and generic datasets. Higher thresholds, such as 0.85, help retain sufficient known samples in CIFAR100 (see Fig. 5b), but may exclude too many in CUB (see Fig. 5a), resulting in suboptimal performance for the fine-grained dataset. Therefore, we set $\\delta = 0.75$ for fine-grained datasets and $\\delta = 0.85$ for generic datasets.", + "bbox": [ + 511, + 657, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of different confidence $\\gamma$ . Similar to $\\delta$ in Intra-CL for selecting high-confidence known samples, $\\gamma$ filters high-quality distributions for inter-class contrastive learning in Inter-CU. Table 5 shows its effect on Stanford Cars and", + "bbox": [ + 511, + 839, + 908, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3299", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8bb4d79cd5fe3adea853372932fc75ec9ab24333ef4354415df4f9d1a9ccdb58.jpg", + "table_caption": [ + "Table 5. Ablation study on different thresholds $\\gamma$ in Inter-CU." + ], + "table_footnote": [], + "table_body": "
γStanford CarsImageNet-100
AllOldNewAllOldNew
0.6060.076.852.285.693.881.5
0.6560.577.052.586.094.181.7
0.7059.876.752.186.594.782.3
0.7558.975.151.286.394.982.1
0.8057.673.750.186.194.481.9
", + "bbox": [ + 112, + 114, + 460, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e619490587e27e7bc1f7a1da542599a6480115addc413cf15fd64272cdf93efb.jpg", + "table_caption": [ + "Table 6. Effectiveness of each component of our method. Intra-CL and Inter-CU are described in Sec. 4.1 and 4.2, respectively. Row (1) shows the baseline SimGCD results." + ], + "table_footnote": [], + "table_body": "
IndexSimGCDIntra-CLInter-CUCUBCIFAR100
AllOldNewAllOldNew
(1)✓60.365.657.780.181.277.8
(2)✓✓65.273.161.381.281.879.6
(3)✓✓67.172.863.881.681.581.7
(4)✓✓✓68.475.165.182.382.382.2
", + "bbox": [ + 93, + 303, + 482, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ImageNet-100. Specifically, we fix $\\delta = 0.75$ for Stanford Cars and $\\delta = 0.85$ for ImageNet-100 based on the $\\delta$ ablation. Intuitively, Stanford Cars performs best with $\\gamma = 0.65$ , while ImageNet-100 performs best with $\\gamma = 0.7$ . Similar to Intra-CL, this difference is mainly due to the varying number of selected samples at each threshold: $6.1\\mathrm{k}$ vs. $95.3\\mathrm{k}$ (see Data in appendix). ImageNet-100 retains more high-confidence samples at a stricter threshold (e.g., 0.75), while fine-grained datasets yield fewer, affecting Inter-CU's effectiveness. Therefore, we set $\\gamma = 0.65$ or 0.6 for fine-grained datasets and $\\gamma = 0.70$ or 0.75 for generic ones.", + "bbox": [ + 88, + 401, + 482, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of each proposed component. Table 6 presents the ablation study on our proposed Intra-CL and Inter-CU. Compared to SimGCD (Row (1)), Intra-CL (Row (2)) improves 'Old' class accuracy by $7.5\\%$ on CUB and $0.6\\%$ on CIFAR100, demonstrating its effectiveness in refining known sample selection for supervised contrastive learning. This also leads to a $3.6\\%$ and $1.8\\%$ improvement in 'New' accuracy on CUB and CIFAR100, respectively. Inter-CU (Row (3)) further enhances 'New' accuracy by $6.1\\%$ on CUB and $3.9\\%$ on CIFAR100 by fully utilizing unlabeled data for novel class discovery. Combining both (Row (4)) achieves the best overall performance, confirming their complementary benefits in improving recognition of both known and novel classes.", + "bbox": [ + 89, + 568, + 482, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of starting epoch $e$ in Inter-CU. Fig. 7b reports 'New' accuracy for different Inter-CU start epochs $e$ . Fig. 7a shows that SimGCD struggles to separate known and novel classes early (e.g., during the first 20 epochs). This difficulty leads to an unreliable class distribution, which negatively impacts Inter-CU since it depends on well-separated inter-class relationships. To address this, we evaluate $e \\in \\{0, 20, 40, 60, 80\\}$ , where $e = 0$ denotes All-", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb54a748e68428cd78d7492e6d2ba1f0bb6c3c0ec76b49a76c81d5ada3ff9c63.jpg", + "image_caption": [ + "(a) 'New' ACC. in SimGCD [45]" + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 710, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d399f2d06f5bc41ae1d043ea53d33b019ee402d29042153c9cd3cbd7fe5a6220.jpg", + "image_caption": [ + "(b) 'New' ACC. @ Epochs", + "Figure 7. 'New' class accuracy on unlabeled data. (a) In SimGCD, 'New' ACC. remains low in the early training stages (e.g. below $40\\%$ for FGVC-Aircraft and Stanford Cars at epoch $\\leq 20$ .) (b) Inter-CU improves accuracy more effectively in the middle and late stages ( $e.g. e \\geq 40$ )." + ], + "image_footnote": [], + "bbox": [ + 714, + 88, + 898, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GCD without Inter-CU (only Intra-CL). Results show that Stanford Cars and FGVC-Aircraft peak at later epochs (e.g., $e = 80$ ), while CUB and CIFAR100 peak at $e = 40$ . This aligns with CUB's higher early-stage known class recognition compared to Stanford Cars and FGVC-Aircraft (60% vs. ≈43%), enabling earlier effective separation of known and novel samples. CIFAR100 is less sensitive to $e$ , reaching approximately 70% accuracy by epoch 40, yielding stable class distributions for Inter-CU.", + "bbox": [ + 511, + 311, + 906, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 464, + 633, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper identifies that previous supervised contrastive learning (S-CL) in parametric GCD algorithms limits novel category discovery due to its restricted use of labeled data. To address this, we propose AllGCD, a straightforward approach that leverages all unlabeled data containing novel classes into S-CL. Specifically, we introduce two strategies named Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first enhances the intra-class compactness within known classes by incorporating potential known samples from unlabeled data into the original S-CL. Building on this, Inter-CU further separates known and novel classes by applying S-CL to a global class distribution. By combining these strategies, AllGCD maximizes the potential of S-CL, achieving substantial improvements in the discovery of novel categories. Extensive experiments show that AllGCD significantly outperforms baselines and achieves competitive results compared to other GCD methods.", + "bbox": [ + 511, + 489, + 906, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 777, + 671, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Guangdong Science and Technology Program (Grant No. 2024B0101040005), the National Natural Science Foundation of China (Grant No. 62332002 and 61825101), the Shenzhen High-Level Talent Team Program, and the Shenzhen Science and Technology Innovation Commission (Grant No. KQTD20240729102051063).", + "bbox": [ + 511, + 801, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3300", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 122, + 187, + 137 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12527-12535, 2023. 1", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Proceedings of the European conference on Computer Vision (ECCV), pages 456-473, 2022. 3", + "[3] Xinzi Cao, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16880-16889, 2024. 1, 2, 3, 4, 6", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of IEEE/CVF International Conference on Computer Vision (CVPR), pages 9630-9640, 2021. 1, 3, 6", + "[5] Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23094-23104, 2024. 2, 6, 7, 1", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 6", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 1, 3, 6", + "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14064-14073, 2022. 1", + "[9] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264-9272, 2021. 3, 6", + "[10] Peiyan Gu, Chuyu Zhang, Ruijie Xu, and Xuming He. Classrelation knowledge distillation for novel class discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16428-16437, 2023. 2" + ], + "bbox": [ + 94, + 147, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 1", + "[12] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 8400-8408, 2019. 2", + "[13] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6767-6781, 2022. 2, 3, 6", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 1", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2980-2988, 2017. 1", + "[16] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 1", + "[17] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2", + "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCV), pages 554–561, 2013. 2, 6, 1", + "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 1", + "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), pages 1106-1114, 2012. 1", + "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 6", + "[22] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJ-CAI, pages 4002–4010. ijcai.org, 2023. 6", + "[23] Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17679-17688, 2024. 2", + "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3301", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297, 1967. 6", + "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 6, 1", + "[26] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. In Proceedings of Conference on European Conference on Computer Vision (ECCV), pages 728-755, 2022. 1", + "[27] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 2", + "[28] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptual contrastive learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7579-7588, 2023. 1, 2, 6, 7", + "[29] Ning Qian. On the momentum term in gradient descent learning algorithms. Neural Networks, 12(1):145-151, 1999. 5", + "[30] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6", + "[31] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 91-99, 2015. 1", + "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 1", + "[33] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 1", + "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1", + "[35] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023. 1", + "[36] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 1" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 2, 6, 1", + "[38] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Proceedings of the European conference on Computer Vision (ECCV), pages 776-794, 2020. 1", + "[39] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 1, 2", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1", + "[41] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482-7491, 2022. 1, 2, 3, 6, 7", + "[42] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2", + "[43] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 1", + "[44] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In Proceedings of International Conference on Learning Representations (ICLR), 2024. 1, 2, 3, 6", + "[45] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16590-16600, 2023. 1, 2, 3, 6, 7, 8", + "[46] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 46(7):5092-5113, 2024. 1", + "[47] Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z. Li. Boosting novel category discovery over domains with soft contrastive learning and all in one classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11824-11833, 2023. 2", + "[48] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proceedings of International Conference on Learning Representations (ICLR), 2018. 2", + "[49] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. In Proceedings of International Conference on Machine Learning, (ICML), pages 41498-41509, 2023. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3302", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3479-3488, 2023. 1, 6", + "[51] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16623-16633, 2023. 6, 7, 1", + "[52] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9462-9470, 2021. 2" + ], + "bbox": [ + 91, + 92, + 486, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3303", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_model.json b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d3b7b3f423bc8e91bff6e1477119ede4a9d1d294 --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_model.json @@ -0,0 +1,2510 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.887, + 0.155 + ], + "angle": 0, + "content": "AllGCD: Leveraging All Unlabeled Data for Generalized Category Discovery" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.18, + 0.852, + 0.2 + ], + "angle": 0, + "content": "Xinzi Cao\\(^{1,2}\\), Ke Chen\\(^{1,2}\\), Feidiao Yang\\(^{2}\\), Xiawu Zheng\\(^{3}\\), Yonghong Tian\\(^{2,4}\\), Yutong Lu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.202, + 0.867, + 0.221 + ], + "angle": 0, + "content": "1 Sun Yat-sen University, 2 Peng Cheng Laboratory, 3 Xiamen University, 4 Peking University" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.226, + 0.732, + 0.242 + ], + "angle": 0, + "content": "{caoxz, chenk68}@mail2.sysu.edu.cn yangfd@pcl.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.245, + 0.797, + 0.259 + ], + "angle": 0, + "content": "zhengxiawu@xmu.edu.cn yhtian@pku.edu.cn luyutong@mail.sysu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.485, + 0.749 + ], + "angle": 0, + "content": "Generalized Category Discovery (GCD) aims to identify both known and novel categories in unlabeled data by leveraging knowledge from labeled datasets. Current methods employ supervised contrastive learning on labeled data to capture known category structures but neglect unlabeled data, limiting their effectiveness in classifying novel classes, especially in fine-grained open-set detection where subtle class differences are crucial. To address this issue, we propose a novel learning approach, AllGCD, which seamlessly integrates all unlabeled data into contrastive learning to enhance the discrimination of novel classes. Specifically, we introduce two key techniques: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first refines intra-class compactness within known categories by integrating potential known samples into labeled data. This process refines the decision boundaries of known categories, reducing ambiguity when distinguishing novel categories. Building on this, Inter-CU further strengthens inter-class separation between known and novel categories by applying global contrastive learning to the class distribution in the unlabeled data. By jointly leveraging Intra-CL and Inter-CU, AllGCD effectively improves both intra-class compactness and inter-class separation, effectively enhancing the discriminability between known and novel classes. Experiments demonstrate that AllGCD significantly improves novel classes accuracy, e.g., achieving increases of \\(7.4\\%\\) on CUB and \\(7.5\\%\\) on Stanford Cars." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.222, + 0.793 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.803, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Deep learning has achieved outstanding performance in computer vision tasks [4, 7, 15, 31, 32, 36, 40], especially in image classification [14, 16, 20, 33, 34, 36]. However, most conventional methods are developed in a closed setting where all training and test data belong to pre-defined" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.713, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.416, + 0.706, + 0.428 + ], + "angle": 0, + "content": "(a) 'New' acc w/o CL in SimGCD" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.294, + 0.905, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.416, + 0.896, + 0.428 + ], + "angle": 0, + "content": "(b) Attentions w/o unlabeled data" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.44, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Comparison of 'New' accuracy and attention w/o unlabeled data in supervised contrastive learning (S-CL). (a) reveals that S-CL solely with labeled data fails to classify novel samples (orange bar vs. blue bar) in the baseline SimGCD. In contrast, our method with unlabeled data (green bar) significantly improves accuracy. (b) shows that our AllGCD highlights more class-specific object regions by incorporating unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.687 + ], + "angle": 0, + "content": "known) classes. It weakens the capability of these methods to handle novel categories, limiting the scope of their applications [8, 11, 26, 46]. To overcome this issue, the Generalized Category Discovery (GCD) paradigm [1, 3, 28, 41, 44, 45, 49, 50] was introduced for open-world learning. It identifies both known and novel categories from unlabeled data by leveraging knowledge learned from labeled data with known classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A key strategy for discovering novel classes is to transfer knowledge from labeled data to promote the learning of novel classes. Based on this idea, previous GCD methods [3, 41, 42, 44, 45] utilize contrastive learning [17, 39] to learn the representations of classification within labeled data. Initially, Vaze et al. [41] adopt a two-stage pipeline: contrastive learning for representation, followed by \\( k \\)-means for classifying unlabeled data. However, clustering becomes computationally expensive as data scales. To reduce this, Wen et al. [45] proposed SimGCD, a parametric approach that replaces clustering with a classifier and learns class prototypes via joint contrastive learning. Building on this, Wang et al. [44] enhanced SimGCD with data prompts for better performance, while Cao et al. [3]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3293" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "addressed SimGCD's limitation in retaining known-class knowledge during transfer by introducing a regularization that improves both known and novel class performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Despite advancements in parametric methods, recent studies reveal that supervised contrastive learning (S-CL), a key component of contrastive learning, has a limited effect in discovering novel classes. This is evident from the SimGCD baseline in Fig. 1a, where 'New' accuracy shows little improvement between that without S-CL and S-CL with only labeled data (orange vs blue bars) across both fine-grained and general datasets. In particular, accuracy even slightly declines when S-CL is limited to labeled data, as observed in Stanford Cars and CIFAR100. This indicates that current S-CL methods, constrained to labeled data, are inadequate for discovering novel classes, as the labeled data encompasses only known classes and omit the novel ones. Therefore, the key to improving the discovery of novel categories is to incorporate unlabeled data into S-CL, as it contains both known and novel samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Motivated by this observation, we propose AllGCD, a straightforward approach that integrates all unlabeled data into contrastive learning to facilitate the novel category discovery. Specifically, AllGCD comprises two main components: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). To establish a stronger foundation for Inter-CU, Intra-CL first enhances intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data. Building on this refined knowledge, Inter-CU further improves inter-class separation between known and novel categories through global contrastive learning on their class distribution in unlabeled data. Finally, by combining Intra-CL and Inter-CU, we significantly improve novel category representations, as shown by the green bar in Fig. 1a." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We validate our approach on six datasets, including three fine-grained datasets (CUB [43], Stanford Cars [18], FGVC-Aircraft [25]), three generic image recognition datasets (CIFAR100 [19] and ImageNet-100 [6]), and the challenging Herbarium-19 dataset [37]. As shown in Fig. 1a, our method outperforms strong baselines in recognizing novel ('New') classes, with a \\(7.4\\%\\) improvement over SimGCD on CUB. Additionally, Fig. 1b highlights that AllGCD focuses on more class-specific regions than the baseline. In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- We propose AllGCD to overcome the limitations of contrastive learning in parametric methods by utilizing all unlabeled data for novel class discovery." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- We introduce Intra-class Contrast in Labeled Data (IntraCL) to enhance intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Building on this refined knowledge, we design Inter-class" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "Contrast in Unlabeled Data (Inter-CU) to improve interclass separation between known and novel categories by leveraging the global class distribution in unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- Extensive results show that our method significantly improves the performance on novel classes, achieving a \\(7.4\\%\\) increase on CUB compared to the baseline." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.196, + 0.664, + 0.212 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Novel Category Discovery (NCD) [12] assumes unlabeled data contains only novel classes and was first formalized as deep transfer clustering, where novel classes are clustered using knowledge from known classes. Various methods address this challenge, including Han et al. [12, 13] who use self-supervised learning and ranking statistics for knowledge transfer, Zhong et al. [52] who propose a mixup strategy [48] to combine old and novel classes, Gu et al. [10] who introduce a knowledge distillation framework for regularizing novel class learning, Zang et al. [47] who transfer knowledge from a source domain to a label-scarce target domain without label constraints, and Liu et al. [23] who apply NCD in ultra-fine-grained visual categorization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Generalized Category Discovery (GCD) [41] extends NCD by assuming that unlabeled data also contains known class images. Vaze et al. [41] formalize this with contrastive learning in representation space during training and clustering at inference. To reduce clustering cost, Wen et al. [45] introduce SimGCD, using a parametric classifier to learn class representation centers, serving as a strong baseline. Cao et al. [3] note SimGCD's issue of forgetting known classes while learning novel ones, proposing entropy regularization for potential known samples. Wang et al. [44] further extend SimGCD by optimizing model and data parameters for better alignment with the GCD task. Despite their performance, these methods rely on limited labeled images, constraining novel class discovery." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Contrastive learning (CL) spans unsupervised [39] and supervised [17] variants for learning from unlabeled and labeled data, respectively. Unsupervised CL treats two augmented views of the same image as positives and all others as negatives, while supervised CL (S-CL) considers samples from the same class as positives. Vaze et al. [41] first combine both in GCD to enhance image representations. Choi et al. [5] employ mean-shifted embeddings for CL, and Pu et al. [28] propose DCCL, which explores conceptual-level CL to model relations between known and novel categories. However, these methods primarily apply S-CL within non-parametric GCD frameworks. In contrast, recent parametric approaches [3, 44, 45] demonstrate that S-CL can effectively train classifiers by learning class prototypes. Nonetheless, these works still rely heavily on limited labeled data, underutilizing the rich supervision signal from abundant unlabeled samples, which ultimately constrains the full potential of contrastive learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3294" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.342, + 0.108 + ], + "angle": 0, + "content": "3. Preliminaries and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.325, + 0.132 + ], + "angle": 0, + "content": "3.1. GCD setting and training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.258 + ], + "angle": 0, + "content": "Problem Formulation. Unlike traditional closed-set classification, where training and test classes coincide \\((\\mathcal{V}_l)\\), GCD aims to identify novel classes \\(\\mathcal{Y}_u\\) in the test data. Given labeled dataset \\(\\mathcal{D}^l = \\{(\\pmb{x}_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_l\\) and unlabeled dataset \\(\\mathcal{D}^u = \\{(x_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_u\\), GCD seeks to classify both known and novel samples in \\(\\mathcal{D}^u\\) using knowledge from \\(\\mathcal{D}^l\\). The total class number \\(K = |\\mathcal{Y}_l\\cup \\mathcal{Y}_u|\\) is assumed known, following prior work [3, 9, 13, 44, 45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.333 + ], + "angle": 0, + "content": "Architecture. SimGCD [45] consists of a Vision Transformer (ViT) encoder [7], denoted as \\(\\Phi\\), along with a classification head \\(f\\) to generate pseudo-labels for self-distillation [2, 4], and a multi-layer perception (MLP) projection head \\(g\\) for representation contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Representation Learning. To learn representations for distinguishing samples, Vaze et al. [41] apply unsupervised contrastive learning across all data and supervised contrastive learning (S-CL) on labeled data. Given two views, \\( \\boldsymbol{x}_i \\) and \\( \\boldsymbol{x}_i' \\), of an identical image as a positive pair in a minibatch \\( B \\), the unsupervised contrastive loss encourages their representations to be close in an unsupervised manner:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.45, + 0.483, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e p}} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {i} ^ {\\prime} / \\tau_ {u}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {u}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.545 + ], + "angle": 0, + "content": "where \\( \\boldsymbol{z}_i = g \\circ \\Phi(\\boldsymbol{x}_i) \\) and \\( \\tau_u \\) is a temperature. Meanwhile, the supervised contrastive loss minimizes the distance between same-class samples in feature space, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.556, + 0.483, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e p}} ^ {s} = \\frac {1}{| B ^ {l} |} \\sum_ {i \\in B ^ {l}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {q} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.655 + ], + "angle": 0, + "content": "where \\( B^l \\) is the labeled subset of \\( B \\), and \\( \\mathcal{N}_i \\) is the set of images with the same label as \\( x_i \\). The overall contrastive learning loss for representation is:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.666, + 0.483, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e p}} = (1 - \\lambda) \\mathcal {L} _ {\\text {r e p}} ^ {u} + \\lambda \\mathcal {L} _ {\\text {r e p}} ^ {s}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.724 + ], + "angle": 0, + "content": "where \\(\\lambda\\) controls the balance between unsupervised and supervised contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.725, + 0.484, + 0.816 + ], + "angle": 0, + "content": "Parametric Clustering. Instead of the time-consuming \\( k \\)-means, Wen et al. [45] propose an efficient parametric classifier \\( f \\) using a set of prototypes \\( \\mathcal{C} = \\{c_1, \\dots, c_K\\} \\), where \\( K \\) is the total classes number. For an image view \\( x_i \\), the prediction is made by calculating the cosine similarity between its feature, \\( \\Phi(x_i) \\), and each prototype \\( c_k \\):" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.826, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {i} ^ {(k)} = \\frac {\\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k}) / \\tau_ {s})}{\\sum_ {k ^ {\\prime}} \\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k ^ {\\prime}}) / \\tau_ {s})}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Similarly, the soft pseudo-labels \\( \\pmb{q}_i' \\) for another view \\( \\pmb{x}_i' \\) are obtained using a smaller \\( \\tau_t \\), and the cross-entropy loss is" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.108 + ], + "angle": 0, + "content": "applied for prototype learning using these pseudo-labels \\(q_{i}^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.118, + 0.905, + 0.155 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} = - \\frac {1}{| B |} \\sum_ {i \\in B} \\sum_ {k} \\boldsymbol {q} ^ {\\prime (k)} \\log \\boldsymbol {p} ^ {(k)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.166, + 0.907, + 0.226 + ], + "angle": 0, + "content": "Additionally, SimGCD employs a class mean entropy regularization [2], \\( H(\\overline{p}) = -\\sum_{k} \\overline{p}^{(k)} \\log \\overline{p}^{(k)} \\), where \\( \\overline{p} \\) is the batch-wise class-average prediction, forcing the model to focus on novel classes. The total classifier is:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.239, + 0.905, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c l s}} = \\lambda \\mathcal {L} _ {\\mathrm {c l s}} ^ {s} + (1 - \\lambda) \\left(\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} - \\varepsilon H\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.86, + 0.284 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{cls}}^s\\) is the cross-entropy loss for labeled data." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.293, + 0.771, + 0.31 + ], + "angle": 0, + "content": "3.2. Limited contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.906, + 0.346 + ], + "angle": 0, + "content": "Analysis. Although contrastive learning (CL) enhances representation discriminability of known classes in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.346, + 0.724, + 0.556 + ], + "angle": 0, + "content": "SimGCD, it applies only to labeled data, as shown in the upper-left corner of Fig. 2. In the upper-left corner, same-colored entries along the vertical and horizontal indices correspond to samples from the same class, while different colors denote distinct classes. Self-contrast cases (blue-filled) on the diagonal are masked out. Obviously, abundant unlabeled data, shown with a gray background, are excluded from the" + }, + { + "type": "image", + "bbox": [ + 0.734, + 0.361, + 0.907, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.493, + 0.906, + 0.548 + ], + "angle": 0, + "content": "Figure 2. A class mask: Samples from the same class are shown in the same color in a batch \\(B\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.906, + 0.603 + ], + "angle": 0, + "content": "CL due to the absence of class labels. As a result, the model lacks direct contrastive supervision on novel samples, making it harder to separate them in feature space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.616, + 0.605, + 0.632 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.748 + ], + "angle": 0, + "content": "To overcome the limited contrastive learning issue discussed in subsection 3.2, we propose a straightforward approach called AllGCD to integrate supervised contrastive learning across all unlabeled data directly. Specifically, we detail our proposed Intra-class Contrast in Labeled Data (Intra-CL) in subsection 4.1 and Inter-class Contrast in Unlabeled Data (Inter-CU) in subsection 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.833, + 0.773 + ], + "angle": 0, + "content": "4.1. Intra-class Contrast in Labeled Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Intra-CL first strengthens intra-class compactness within known classes for better separation of known and novel categories in Inter-CU. To achieve this, we use baseline SimGCD [45] to identify potential known samples in unlabeled data and incorporate them into the labeled set. However, as shown by the orange lines in Fig. 4a and Fig. 4b, SimGCD's known-class accuracy declines sharply during training, compromising the quality of potential known" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "3295" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.907, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.405, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed AllGCD. It is composed of Intra-CL and Inter-CU. First, Intra-CL (Sec. 4.1) enhances knowledge of known classes by incrementally collecting reliable pseudo-labels \\(\\mathcal{V}^*\\) within unlabeled data. Once this knowledge is strengthened, Inter-CU (Sec. 4.2) further promotes the separation of known and novel classes through contrastive learning between them." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.482, + 0.58 + ], + "angle": 0, + "content": "samples. To overcome this, we propose a voting strategy that utilizes the model's prior knowledge to identify potential known samples early. Specifically, Intra-CL comprises three steps: (1) Selecting potential known samples within unlabeled data as in [3]; (2) Storing these known samples into a dictionary \\(\\mathcal{Q}\\); (3) Voting pseudo-labels \\(\\mathcal{Y}^*\\) from \\(\\mathcal{Q}\\) based on their most frequent label across training iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.608 + ], + "angle": 0, + "content": "Step (1): Selecting potential known samples within unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.67 + ], + "angle": 0, + "content": "In this step, we identify potential known samples by selecting instances where both views of the same image exhibit high confidence \\((\\delta)\\) and are classified as known categories \\((\\mathcal{V}_l)\\). The specific calculation process is as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Let \\(B^u\\) denote the unlabeled subset of \\(B\\), where each image in this batch has a unique index \\(\\mathcal{U}_B = \\{u_1, u_2, \\ldots, u_b\\}\\). Given the predictions \\(p_i\\) and \\(q_i'\\) from two views of an image in \\(B^u\\), as defined in Eq. (4). We first identify high-confidence samples by selecting those with predictions \\(p_i\\) and \\(q_i'\\) both exceed a confidence threshold \\(\\delta\\): \\(\\mathbf{M}_{\\delta} = \\mathbb{1}(\\max(p_i) > \\delta \\land \\max(q_i') > \\delta)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Next, we classify a sample as known only if the predicted label from the two views match and both belong to known class set \\(\\mathcal{Y}_l\\): \\(\\mathbf{M}_{\\mathrm{known}} = \\mathbb{1}((\\hat{y}_i = \\hat{y}_i') \\wedge (\\hat{y}_i \\in \\mathcal{Y}_l))\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Finally, the selected known samples within the unlabeled subset \\( B^{u} \\) are determined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.858, + 0.483, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} = \\left(\\mathbf {M} _ {\\delta} \\odot \\mathbf {M} _ {\\text {k n o w n}}\\right) \\odot \\hat {\\mathcal {Y}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\delta}\\) and \\(\\mathbf{M}_{\\mathrm{known}}\\) are binary (0-1) vectors of length b," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.905, + 0.534 + ], + "angle": 0, + "content": "\\(\\odot\\) is element-wise multiplication, and \\(\\hat{\\mathcal{Y}} = \\{\\hat{y}_i \\mid i \\in [1,b]\\}\\) represents the predicted labels across the batch of unlabeled samples. Thus, \\(\\hat{\\mathcal{Y}}_{\\mathrm{select}}\\) identifies the known samples \\((\\mathbf{M}_{\\mathrm{known}} = 1)\\) that also have high confidence \\((\\mathbf{M}_{\\delta} = 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.535, + 0.883, + 0.55 + ], + "angle": 0, + "content": "Step (2): Storing above predicted known samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.905, + 0.595 + ], + "angle": 0, + "content": "To leverage the model's early ability to recognize known classes, we initialize a dictionary \\(\\mathcal{Q}\\) to store the pseudolabels \\(\\hat{\\mathcal{Y}}_{\\mathrm{select}}\\) of each image based on its unique index \\(\\mathcal{U}_B\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.626 + ], + "angle": 0, + "content": "Concretely, for each unlabeled subset \\(B^u\\), we update the dictionary \\(\\mathcal{Q}:\\mathcal{U}_B\\to \\hat{\\mathcal{V}}_{\\mathrm{select}}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.636, + 0.905, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} \\left(u _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {Q} \\left(u _ {i}\\right) \\cup \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\in \\operatorname {d o m} (\\mathcal {Q}), \\\\ \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\notin \\operatorname {d o m} (\\mathcal {Q}). \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.687, + 0.905, + 0.717 + ], + "angle": 0, + "content": "Step (3): Voting on pseudo-labels for unlabeled data based on frequency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.793 + ], + "angle": 0, + "content": "To obtain a reliable pseudo-label for each unlabeled image, we apply a voting mechanism over the stored pseudolabels \\(\\mathcal{Q}\\) across training batches. Specifically, for a batch of unlabeled samples, we select the most frequently assigned pseudo-label from its historical records in \\(\\mathcal{Q}(u_i)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.804, + 0.905, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} ^ {*} = \\underset {y} {\\arg \\max } \\left(\\operatorname {f r e q} (\\hat {y} \\in \\mathcal {Q} (u _ {i})), \\forall u _ {i} \\in \\mathcal {U} _ {B}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{Y}^* = \\{y_1^*, y_2^*, \\ldots, y_b^*\\}\\) denotes the pseudo-labels for unlabeled samples in \\(B^u\\). The updated mini-batch labels, incorporating both original labeled data \\(\\mathcal{Y}_l\\) and pseudolabels \\(\\mathcal{Y}^*\\), are given by \\(\\mathcal{Y}_{\\mathrm{intra}} = \\mathcal{Y}_l \\cup \\mathcal{Y}^*\\). Consequently," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.287, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.215, + 0.215, + 0.226 + ], + "angle": 0, + "content": "(a) CUB" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.089, + 0.481, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.215, + 0.431, + 0.226 + ], + "angle": 0, + "content": "(b) Stanford Cars" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Known-class accuracy on unlabeled data during training. SimGCD, using S-CL without unlabeled data, shows a sharp drop on CUB and Stanford Cars. In contrast, Intra-class leverages potential known samples in unlabeled data, improving the knowledge of known classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.348 + ], + "angle": 0, + "content": "the supervised contrastive loss in Eq. (2) is updated with pseudo known samples from unlabeled data as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.36, + 0.49, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t r a}} = \\frac {1}{| B ^ {\\text {i n t r a}} |} \\sum_ {i \\in B ^ {\\text {i n t r a}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.489 + ], + "angle": 0, + "content": "where \\( B^{\\mathrm{intra}} \\) is the newly labeled subset of mini-batch \\( B \\) with labels \\( \\mathcal{V}_{\\mathrm{intra}} \\), and \\( \\mathcal{N}_i \\) indexes same-label images as \\( x_i \\) in the batch. After applying Intra-CL, the accuracy of known classes in unlabeled data remains more stable (Fig. 4), indicating more compact representations within known classes." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.514, + 0.423, + 0.529 + ], + "angle": 0, + "content": "4.2. Inter-class Contrast in Unlabeled Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.702 + ], + "angle": 0, + "content": "While Inter-CL enhances discrimination among known classes, it lacks contrastive interactions with novel classes, limiting novel category accuracy (Fig. 1a). To address this, we propose Inter-CU, which enforces global contrastive learning on the unlabeled class distribution. Specifically, once the model has stabilized on known and novel categories from epoch \\( e \\), we use the Hungarian algorithm [21] on the predictions of all unlabeled data to optimize the class distribution of known and novel categories. To ensure a high-quality distribution, we retain only high-confidence predictions \\( (P > \\gamma) \\) for global contrastive learning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.703, + 0.484, + 0.779 + ], + "angle": 0, + "content": "Formally, given the unlabeled dataset \\(\\mathcal{D}^u = \\{(x_i,y_i)\\} \\in \\mathcal{X}\\times \\mathcal{Y}_u\\) we obtain classifier predictions \\(P = [p_1,p_2,\\dots ,p_N]\\) via Eq. (4), where \\(N\\) is the total number of unlabeled samples. The minimum-cost assignment function \\(\\pi\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.788, + 0.484, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\pi^ {*} = \\arg \\min _ {\\pi} \\sum_ {i = 1} ^ {N} - P _ {i, \\pi (i)}, \\quad \\pi (i) \\in \\{1, 2, \\dots , K \\}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pi^{*}\\) assigns each sample in \\(\\mathcal{D}^u\\) to one of \\(K\\) categories, minimizing the sum of costs \\(-P_{i,\\pi (i)}\\). The Hungarian algorithm efficiently solves this assignment problem by finding the optimal mapping with minimal total cost. Therefore, the" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.848, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 One Training Procedure for AllGCD" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.11, + 0.907, + 0.506 + ], + "angle": 0, + "content": "Input: Encoder \\(\\Phi\\), classifier head \\(f\\), projector \\(g\\), labeled data \\(\\mathcal{D}^l\\), unlabeled data \\(\\mathcal{D}^u\\), hyperparameters \\(\\delta, e\\). \nOutput: \\(\\Phi, f\\), and \\(g\\); \n1: Initialize dictionary \\(\\mathcal{Q} \\gets \\emptyset\\); \n2: for epoch = 1 to total epochs do \n3: if epoch ≥ e then \n4: Get assignment \\(\\mathcal{Y}_{\\text{inter}}\\) for \\(D^u\\) using Eq. (12); \n5: end if \n6: for \\((x, x') \\in \\mathcal{D}^l \\cup \\mathcal{D}^u\\), \\(y \\in \\mathcal{Y}_l\\), \\(y_{\\text{inter}} \\in \\mathcal{Y}_{\\text{inter}}\\) do \n7: \\(z \\gets g \\circ \\Phi(x)\\), \\(z' \\gets g \\circ \\Phi(x')\\); \n8: \\(p \\gets f \\circ \\Phi(x)\\), \\(q' \\gets f \\circ \\Phi(x')\\); \n9: Compute \\(\\mathcal{L}_{\\text{cls}}^s(p, y) + \\mathcal{L}_{\\text{cls}}^s(q', y)\\); ▷ For \\(\\mathcal{D}^l\\) \n10: Compute \\(\\mathcal{L}_{\\text{dis}}^u(p, q')\\) using Eq. (5); \n11: Compute Max class Entropy \\(H(p)\\); \n12: Compute \\(\\mathcal{L}_{\\text{cls}} \\gets \\lambda \\mathcal{L}_{\\text{cls}}^s + (1 - \\lambda) (\\mathcal{L}_{\\text{dis}}^u - \\varepsilon H)\\); \n13: Step 1: Generate \\(\\hat{y} \\gets (p, q', \\delta)\\); ▷ For \\(\\mathcal{D}^u\\) \n14: Step 2: Store \\(\\hat{y}\\) into \\(\\mathcal{Q}\\) using Eq. (8); \n15: Step 3: Vote for \\(y^*\\) from \\(\\mathcal{Q}\\) using Eq. (9); \n16: Combine with pseudo-labels \\(y_{\\text{intra}} \\gets y \\cup y^*\\); \n17: Compute \\(\\mathcal{L}_{\\text{rep}}^u(z, z')\\) using Eq. (1); \n18: Compute \\(\\mathcal{L}_{\\text{intra}}(z, z', y_{\\text{intra}})\\) using Eq. (10); \n19: Compute \\(\\mathcal{L}_{\\text{inter}}(z, z', y_{\\text{inter}})\\) using Eq. (12); \n20: \\(\\mathcal{L}_{\\text{all}} = (1 - \\lambda) \\mathcal{L}_{\\text{rep}}^u + \\lambda (\\mathcal{L}_{\\text{intra}} + \\mathcal{L}_{\\text{inter}}) + \\mathcal{L}_{\\text{cls}}\\); \n21: Update \\(\\Phi, f\\), and \\(g\\) by SGD [29]; \n22: end for \n23: end for" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.534, + 0.783, + 0.549 + ], + "angle": 0, + "content": "final assignments for unlabeled data are:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.563, + 0.907, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Y} _ {\\text {i n t e r}} = \\left\\{\\hat {y} _ {i} = \\pi^ {*} (i) \\mid i = 1, 2, \\dots , N \\right\\} \\odot \\mathbf {M} _ {\\text {i n t e r}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.907, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}_{\\mathrm{inter}}\\) is a binary mask that applies a confidence threshold \\(\\gamma\\) to filters out low-confidence predictions: \\(\\mathbf{M}_{\\mathrm{inter}} = \\{i \\mid \\max(\\pmb{p}_i) > \\gamma\\}\\). Only high-confidence samples contribute to the global contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.667, + 0.92, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t e r}} = \\frac {1}{| B ^ {\\text {i n t e r}} |} \\sum_ {i \\in B ^ {\\text {i n t e r}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.781 + ], + "angle": 0, + "content": "where \\( B^{\\mathrm{inter}} \\) corresponds to the refined labeled subset with pseudo-labels \\( \\mathcal{V}_{\\mathrm{inter}} \\). This approach enables global contrastive learning on both known and novel samples, enhancing category discovery, particularly for novel classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Overall. Utilizing all unlabeled data in supervised contrastive learning significantly boosts class discovery, especially for novel classes. This overall training loss is updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.845, + 0.905, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l l}} = (1 - \\lambda) \\mathcal {L} _ {\\text {i n t r a}} + \\lambda \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {c l s}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) controls the trade-off between Intra-CL and Inter-CU. Algorithm 1 details one training step of AllGCD." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3297" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Comparison with State-of-the-Art GCD Methods, Evaluated with and without Ground-Truth \\( K \\) for Clustering. Bold denotes the best results." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.128, + 0.907, + 0.411 + ], + "angle": 0, + "content": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
(a) Clustering with the ground-truth number of classes K given
k-means[24]52.052.250.872.775.571.334.338.932.112.810.613.816.014.416.813.012.213.4
RS+[13]58.277.619.337.161.624.833.351.624.228.361.812.126.936.422.227.955.812.8
UNO+[9]69.580.647.270.395.057.935.149.028.135.570.518.640.356.432.228.353.714.7
ORCA[22]69.077.452.073.592.663.935.345.630.223.550.110.722.031.817.120.930.915.5
GCD [41]73.076.266.574.189.866.351.356.648.739.057.629.945.041.146.935.451.027.0
DCCL [28]75.376.870.280.590.576.263.560.864.9---43.155.736.2---
PromptCAL [50]81.284.275.383.192.778.362.964.462.150.270.140.652.252.252.337.052.028.9
SimGCD [45]80.181.277.883.093.177.960.365.657.753.871.945.054.259.151.844.058.036.4
InfoSieve [30]78.382.270.580.593.873.869.477.965.255.774.846.456.363.752.541.055.433.2
SPTNet [44]81.384.375.685.493.281.465.868.865.159.079.249.359.361.858.143.458.735.2
LegoGCD [3]81.881.482.586.394.582.163.871.959.857.375.748.455.061.551.745.157.438.5
CMS [5]82.385.775.584.795.679.268.276.564.056.063.452.356.976.147.636.454.926.4
AllGCD (ours)82.382.382.286.594.782.368.475.165.160.577.052.557.462.454.945.258.338.0
(b) Clustering without the ground-truth number of classes K given
GCD [41]70.877.657.077.991.171.351.156.448.439.158.629.7---37.251.729.4
GPC [51]75.484.660.175.393.466.752.055.547.538.258.927.443.340.744.836.551.727.9
SimGCD [45]80.181.277.881.791.276.861.566.459.149.165.141.3------
CMS [5]79.683.272.381.395.674.264.468.262.451.768.943.455.260.652.437.456.527.1
AllGCD (ours)81.583.378.484.494.380.266.772.164.058.875.750.656.764.253.045.758.738.5
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.224, + 0.453 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.463, + 0.281, + 0.479 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Datasets. We evaluate our approach on three fine-grained datasets: CUB [43], Stanford Cars [18], and FGVC-Aircraft [25], as well as two generic image recognition datasets, CIFAR100 [19] and ImageNet-100 [6]. Additionally, we test on a more challenging long-tailed fine-grained Herbarium 19 [37]. Following [41], we randomly sample \\(50\\%\\) of seen categories as labeled data \\(\\mathcal{D}^l\\), with the remaining images from seen and novel categories forming the unlabeled set \\(\\mathcal{D}^u\\). Details are in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Evaluation protocol. We use clustering accuracy (ACC) to evaluate model performance, following [41, 45]. During evaluation, we compare the ground-truth labels \\( y_{i} \\) with the predicted labels \\( \\hat{y}_i \\), and calculate ACC as \\( ACC = \\frac{1}{|\\mathcal{D}^u|}\\sum_{i=1}^{|\\mathcal{D}^u|}\\mathbb{1}(y_i = p(\\hat{y}_i)) \\), where \\( p \\) is determined using the Hungarian optimal assignment algorithm [21]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation details. AllGCD is built upon SimGCD [45], using a ViT-B/16 backbone [7] pretrained with DINO [4]. We use the class token to train a classifier and fine-tune only the last attention block. We use a batch size of 128 and an initial learning rate of 0.1, decayed over 200 epochs using a cosine schedule. Following [41], we set \\(\\lambda = 0.35\\), \\(\\tau_{u} = 0.07\\), and \\(\\tau_{c} = 1.0\\) for representation learning, and use \\(\\tau_{s} = 0.1\\), with \\(\\tau_{t}\\) decayed from 0.07 to 0.04 over 30 epochs for classifier training. Pseudo-label thresholds \\(\\delta\\) and \\(\\gamma\\) range from 0.6 to 0.9. Based on the convergence time of 'New' classes in SimGCD (Fig. 7a), the Inter-CU start epoch \\(e\\) is selected" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.907, + 0.468 + ], + "angle": 0, + "content": "from \\(\\{0,20,40,60,80\\}\\). All experiments are conducted using PyTorch on NVIDIA Tesla V100 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.847, + 0.497 + ], + "angle": 0, + "content": "5.2. Comparison With the State of the Arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.907, + 0.596 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of AllGCD, we evaluate it under both known and unknown total class numbers \\( K \\): (1) In Table 1(a), we assume \\( K \\) is known, consistent with prior works [3, 9, 13, 44, 45]; (2) In Table 1(b), we report results where \\( K \\) is unknown, estimating it using clustering techniques as in [5, 41, 51]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Evaluation with ground-truth class number \\( K \\). Table 1(a) compares state-of-the-art methods assuming a known class number \\( K \\). Our method significantly enhances 'New' category accuracy compared to the baseline SimGCD, with notable improvements of \\( 7.4\\% \\), \\( 7.5\\% \\), and \\( 4.4\\% \\) on CUB, Stanford Cars, and ImageNet-100, respectively. Furthermore, AllGCD remains competitive with SPTNet, surpassing it by \\( 2.6\\% \\) and \\( 1.5\\% \\) on the 'All' classes in CUB and Stanford Cars, respectively. Importantly, our method attains comparable performance to SPTNet using only 200 training epochs, compared to SPTNet's 1000 epochs, showing superior computational efficiency. Though AllGCD trails CMS [5] by \\( 2.7\\% \\) in 'Old' accuracy on CIFAR100, it outperforms CMS by \\( 6.7\\% \\) in 'New' accuracy, highlighting its strength in novel category discovery." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation with estimated number of clusters. Table 1 (b) presents a comparison conducted without access to the ground-truth class number \\( K \\), instead using an estimated \\( K \\) derived from an off-the-shelf method [41], as shown in Table 2. Table 2 shows our approach provides a more accu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3298" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.103, + 0.633, + 0.201 + ], + "angle": 0, + "content": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓
GT100-100-200-196-100-683-
GCD [41]1000109923115.523017.3--52023.8
DCCL [28]146461292917291920.02----
SimGCD [45]1000109923115.523017.3--52023.8
GPC [51]1000103321262010.03----
CMS [5]955116161701515620.490106228.9
AllGCD (ours)95510882136.52107.110996366.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.211, + 0.611, + 0.226 + ], + "angle": 0, + "content": "Table 2. Estimated class number and error rate for \\( K \\) using off-the-self method [41]." + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.089, + 0.765, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.187, + 0.746, + 0.2 + ], + "angle": 0, + "content": "(c) SimGCD [45]" + }, + { + "type": "image", + "bbox": [ + 0.774, + 0.089, + 0.898, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.187, + 0.869, + 0.199 + ], + "angle": 0, + "content": "(d) AllGCD" + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.212, + 0.905, + 0.24 + ], + "angle": 0, + "content": "Figure 5. T-SNE visualization of 10 random novel classes from CIFAR100." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.254, + 0.389, + 0.357 + ], + "angle": 0, + "content": "
δCUBCIFAR100
AllOldNewAllOldNew
0.7068.274.664.980.881.380.5
0.7568.475.165.181.481.881.3
0.8068.275.764.582.283.181.8
0.8567.274.763.482.382.382.2
0.9066.273.362.581.681.981.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.368, + 0.369, + 0.381 + ], + "angle": 0, + "content": "Table 3. Effect of confidence \\( \\delta \\) in Intra-CL." + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.257, + 0.644, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.457, + 0.353, + 0.583, + 0.364 + ], + "angle": 0, + "content": "(a) CUB with \\(\\delta = 0.75\\)" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.257, + 0.898, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.353, + 0.852, + 0.364 + ], + "angle": 0, + "content": "(b) CIFAR100 with \\(\\delta = 0.85\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.368, + 0.788, + 0.381 + ], + "angle": 0, + "content": "Figure 6. Logit distributions of unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.409, + 0.484, + 0.575 + ], + "angle": 0, + "content": "rate estimation of \\( K \\), with fewer errors in determining cluster centers than SimGCD and CMS, except for CIFAR100 in SimGCD. Under this improved estimation, our approach consistently outperforms state-of-the-art methods across all datasets. Notably, performance gains are particularly significant on the Stanford Cars, ImageNet-100, and Herbarium-19 benchmarks, where AllGCD achieves 7.2%, 6.0%, and 9.1% higher accuracy in the 'New' category compared to the second-best results. These findings highlight the effectiveness of our method in utilizing all available unlabeled data to learn more discriminative representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Feature Visualizations and Attention Distributions. To demonstrate the effectiveness of our approach in representation learning, we visualize 10 randomly selected novel classes from CIFAR100. As shown in Fig. 5, AllGCD produces more compact and well-separated clusters, whereas SimGCD exhibits considerable cluster overlap. Additionally, we visualize attention on both known and novel samples from unlabeled data in Table 4. Ours focuses on class-specific regions (e.g., the bird's wing) and suppressing background noise (e.g., the window in the second row), highlighting that leveraging unlabeled data enhances attention to relevant regions and improves class discriminability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.244, + 0.787 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.869 + ], + "angle": 0, + "content": "In this section, we present ablation studies to assess the effectiveness of Intra-CL and Inter-CU in our AllGCD model. The evaluation includes fine-grained datasets such as CUB and Stanford Cars, alongside generic image recognition datasets CIFAR100 and ImageNet-100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effect of different confidence \\(\\delta\\). Table 3 shows the effect of varying the confidence threshold \\(\\delta\\) in Intra-CL on" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.407, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Table 4. Average attention across 12 heads in the final layer of the image encoder on CUB and FGVC-Aircraft. The known sample (top row) and novel sample (bottom row) in AllGCD show improved focus on object regions with reduced background noise." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.474, + 0.907, + 0.628 + ], + "angle": 0, + "content": "
ImageDINOSimGCDAllGCD
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.839 + ], + "angle": 0, + "content": "CUB and CIFAR100. Overall, the accuracy for both 'Old' and 'New' classes initially improves with increasing \\(\\delta\\) but slightly drops beyond the optimal point. The fine-grained CUB dataset performs best at \\(\\delta = 0.75\\), while the generic dataset CIFAR100 peaks at \\(\\delta = 0.85\\). This difference arises from the varying proportions of unlabeled samples preserved in fine-grained and generic datasets. Higher thresholds, such as 0.85, help retain sufficient known samples in CIFAR100 (see Fig. 5b), but may exclude too many in CUB (see Fig. 5a), resulting in suboptimal performance for the fine-grained dataset. Therefore, we set \\(\\delta = 0.75\\) for fine-grained datasets and \\(\\delta = 0.85\\) for generic datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Effect of different confidence \\(\\gamma\\). Similar to \\(\\delta\\) in Intra-CL for selecting high-confidence known samples, \\(\\gamma\\) filters high-quality distributions for inter-class contrastive learning in Inter-CU. Table 5 shows its effect on Stanford Cars and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3299" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.089, + 0.471, + 0.104 + ], + "angle": 0, + "content": "Table 5. Ablation study on different thresholds \\( \\gamma \\) in Inter-CU." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.115, + 0.462, + 0.238 + ], + "angle": 0, + "content": "
γStanford CarsImageNet-100
AllOldNewAllOldNew
0.6060.076.852.285.693.881.5
0.6560.577.052.586.094.181.7
0.7059.876.752.186.594.782.3
0.7558.975.151.286.394.982.1
0.8057.673.750.186.194.481.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.293 + ], + "angle": 0, + "content": "Table 6. Effectiveness of each component of our method. Intra-CL and Inter-CU are described in Sec. 4.1 and 4.2, respectively. Row (1) shows the baseline SimGCD results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.304, + 0.483, + 0.376 + ], + "angle": 0, + "content": "
IndexSimGCDIntra-CLInter-CUCUBCIFAR100
AllOldNewAllOldNew
(1)✓60.365.657.780.181.277.8
(2)✓✓65.273.161.381.281.879.6
(3)✓✓67.172.863.881.681.581.7
(4)✓✓✓68.475.165.182.382.382.2
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.483, + 0.569 + ], + "angle": 0, + "content": "ImageNet-100. Specifically, we fix \\(\\delta = 0.75\\) for Stanford Cars and \\(\\delta = 0.85\\) for ImageNet-100 based on the \\(\\delta\\) ablation. Intuitively, Stanford Cars performs best with \\(\\gamma = 0.65\\), while ImageNet-100 performs best with \\(\\gamma = 0.7\\). Similar to Intra-CL, this difference is mainly due to the varying number of selected samples at each threshold: \\(6.1\\mathrm{k}\\) vs. \\(95.3\\mathrm{k}\\) (see Data in appendix). ImageNet-100 retains more high-confidence samples at a stricter threshold (e.g., 0.75), while fine-grained datasets yield fewer, affecting Inter-CU's effectiveness. Therefore, we set \\(\\gamma = 0.65\\) or 0.6 for fine-grained datasets and \\(\\gamma = 0.70\\) or 0.75 for generic ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Effect of each proposed component. Table 6 presents the ablation study on our proposed Intra-CL and Inter-CU. Compared to SimGCD (Row (1)), Intra-CL (Row (2)) improves 'Old' class accuracy by \\(7.5\\%\\) on CUB and \\(0.6\\%\\) on CIFAR100, demonstrating its effectiveness in refining known sample selection for supervised contrastive learning. This also leads to a \\(3.6\\%\\) and \\(1.8\\%\\) improvement in 'New' accuracy on CUB and CIFAR100, respectively. Inter-CU (Row (3)) further enhances 'New' accuracy by \\(6.1\\%\\) on CUB and \\(3.9\\%\\) on CIFAR100 by fully utilizing unlabeled data for novel class discovery. Combining both (Row (4)) achieves the best overall performance, confirming their complementary benefits in improving recognition of both known and novel classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effect of starting epoch \\( e \\) in Inter-CU. Fig. 7b reports 'New' accuracy for different Inter-CU start epochs \\( e \\). Fig. 7a shows that SimGCD struggles to separate known and novel classes early (e.g., during the first 20 epochs). This difficulty leads to an unreliable class distribution, which negatively impacts Inter-CU since it depends on well-separated inter-class relationships. To address this, we evaluate \\( e \\in \\{0, 20, 40, 60, 80\\} \\), where \\( e = 0 \\) denotes All-" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.712, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.19, + 0.704, + 0.203 + ], + "angle": 0, + "content": "(a) 'New' ACC. in SimGCD [45]" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.089, + 0.9, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.19, + 0.88, + 0.203 + ], + "angle": 0, + "content": "(b) 'New' ACC. @ Epochs" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.215, + 0.907, + 0.286 + ], + "angle": 0, + "content": "Figure 7. 'New' class accuracy on unlabeled data. (a) In SimGCD, 'New' ACC. remains low in the early training stages (e.g. below \\(40\\%\\) for FGVC-Aircraft and Stanford Cars at epoch \\(\\leq 20\\).) (b) Inter-CU improves accuracy more effectively in the middle and late stages (\\(e.g. e \\geq 40\\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.907, + 0.449 + ], + "angle": 0, + "content": "GCD without Inter-CU (only Intra-CL). Results show that Stanford Cars and FGVC-Aircraft peak at later epochs (e.g., \\( e = 80 \\)), while CUB and CIFAR100 peak at \\( e = 40 \\). This aligns with CUB's higher early-stage known class recognition compared to Stanford Cars and FGVC-Aircraft (60% vs. ≈43%), enabling earlier effective separation of known and novel samples. CIFAR100 is less sensitive to \\( e \\), reaching approximately 70% accuracy by epoch 40, yielding stable class distributions for Inter-CU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.634, + 0.481 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.763 + ], + "angle": 0, + "content": "This paper identifies that previous supervised contrastive learning (S-CL) in parametric GCD algorithms limits novel category discovery due to its restricted use of labeled data. To address this, we propose AllGCD, a straightforward approach that leverages all unlabeled data containing novel classes into S-CL. Specifically, we introduce two strategies named Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first enhances the intra-class compactness within known classes by incorporating potential known samples from unlabeled data into the original S-CL. Building on this, Inter-CU further separates known and novel classes by applying S-CL to a global class distribution. By combining these strategies, AllGCD maximizes the potential of S-CL, achieving substantial improvements in the discovery of novel categories. Extensive experiments show that AllGCD significantly outperforms baselines and achieves competitive results compared to other GCD methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.673, + 0.795 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the Guangdong Science and Technology Program (Grant No. 2024B0101040005), the National Natural Science Foundation of China (Grant No. 62332002 and 61825101), the Shenzhen High-Level Talent Team Program, and the Shenzhen Science and Technology Innovation Commission (Grant No. KQTD20240729102051063)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3300" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.123, + 0.188, + 0.138 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.486, + 0.216 + ], + "angle": 0, + "content": "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12527-12535, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.219, + 0.485, + 0.302 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Proceedings of the European conference on Computer Vision (ECCV), pages 456-473, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[3] Xinzi Cao, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16880-16889, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.376, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of IEEE/CVF International Conference on Computer Vision (CVPR), pages 9630-9640, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.447, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[5] Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23094-23104, 2024. 2, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14064-14073, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.76, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264-9272, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.831, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[10] Peiyan Gu, Chuyu Zhang, Ruijie Xu, and Xuming He. Classrelation knowledge distillation for novel class discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16428-16437, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.148, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[12] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 8400-8408, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[13] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6767-6781, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2980-2988, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[16] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[17] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCV), pages 554–561, 2013. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), pages 1106-1114, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[22] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJ-CAI, pages 4002–4010. ijcai.org, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17679-17688, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3301" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297, 1967. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[26] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. In Proceedings of Conference on European Conference on Computer Vision (ECCV), pages 728-755, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[27] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.467 + ], + "angle": 0, + "content": "[28] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptual contrastive learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7579-7588, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[29] Ning Qian. On the momentum term in gradient descent learning algorithms. Neural Networks, 12(1):145-151, 1999. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[30] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[31] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 91-99, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[33] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[35] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[36] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[37] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[38] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Proceedings of the European conference on Computer Vision (ECCV), pages 776-794, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.236 + ], + "angle": 0, + "content": "[39] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.239, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.907, + 0.368 + ], + "angle": 0, + "content": "[41] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482-7491, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.37, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[42] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[43] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[44] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In Proceedings of International Conference on Learning Representations (ICLR), 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[45] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16590-16600, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[46] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 46(7):5092-5113, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[47] Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z. Li. Boosting novel category discovery over domains with soft contrastive learning and all in one classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11824-11833, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[48] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proceedings of International Conference on Learning Representations (ICLR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[49] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. In Proceedings of International Conference on Machine Learning, (ICML), pages 41498-41509, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3302" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.176 + ], + "angle": 0, + "content": "[50] Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3479-3488, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[51] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16623-16633, 2023. 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[52] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9462-9470, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3303" + } + ] +] \ No newline at end of file diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_origin.pdf b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca0389c8e7a28bd8c4dd4f620137fc47556db8ee --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/87a08d4a-611f-4092-84c0-36e8cdf6fd1f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a1413ed5bc92b615cd1706641d88bf5b2f6900767512ab008618bc96487eb8b +size 5856280 diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/full.md b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c78eec00483086d4abfaac7e13a6cb5eefdcc10b --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/full.md @@ -0,0 +1,395 @@ +# AllGCD: Leveraging All Unlabeled Data for Generalized Category Discovery + +Xinzi Cao $^{1,2}$ , Ke Chen $^{1,2}$ , Feidiao Yang $^{2}$ , Xiawu Zheng $^{3}$ , Yonghong Tian $^{2,4}$ , Yutong Lu $^{1*}$ + +1 Sun Yat-sen University, 2 Peng Cheng Laboratory, 3 Xiamen University, 4 Peking University + +{caoxz, chenk68}@mail2.sysu.edu.cn yangfd@pcl.ac.cn + +zhengxiawu@xmu.edu.cn yhtian@pku.edu.cn luyutong@mail.sysu.edu.cn + +# Abstract + +Generalized Category Discovery (GCD) aims to identify both known and novel categories in unlabeled data by leveraging knowledge from labeled datasets. Current methods employ supervised contrastive learning on labeled data to capture known category structures but neglect unlabeled data, limiting their effectiveness in classifying novel classes, especially in fine-grained open-set detection where subtle class differences are crucial. To address this issue, we propose a novel learning approach, AllGCD, which seamlessly integrates all unlabeled data into contrastive learning to enhance the discrimination of novel classes. Specifically, we introduce two key techniques: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first refines intra-class compactness within known categories by integrating potential known samples into labeled data. This process refines the decision boundaries of known categories, reducing ambiguity when distinguishing novel categories. Building on this, Inter-CU further strengthens inter-class separation between known and novel categories by applying global contrastive learning to the class distribution in the unlabeled data. By jointly leveraging Intra-CL and Inter-CU, AllGCD effectively improves both intra-class compactness and inter-class separation, effectively enhancing the discriminability between known and novel classes. Experiments demonstrate that AllGCD significantly improves novel classes accuracy, e.g., achieving increases of $7.4\%$ on CUB and $7.5\%$ on Stanford Cars. + +# 1. Introduction + +Deep learning has achieved outstanding performance in computer vision tasks [4, 7, 15, 31, 32, 36, 40], especially in image classification [14, 16, 20, 33, 34, 36]. However, most conventional methods are developed in a closed setting where all training and test data belong to pre-defined + +![](images/7fa28c778791a9f66da1ff2941555d0ed0b452a00512dd08b112e10a483e9ba7.jpg) +(a) 'New' acc w/o CL in SimGCD + +![](images/f56d58f8e2732f1cc23fca788ac7fc9a6c657ad3308aee748903890a3907b090.jpg) +(b) Attentions w/o unlabeled data +Figure 1. Comparison of 'New' accuracy and attention w/o unlabeled data in supervised contrastive learning (S-CL). (a) reveals that S-CL solely with labeled data fails to classify novel samples (orange bar vs. blue bar) in the baseline SimGCD. In contrast, our method with unlabeled data (green bar) significantly improves accuracy. (b) shows that our AllGCD highlights more class-specific object regions by incorporating unlabeled data. + +known) classes. It weakens the capability of these methods to handle novel categories, limiting the scope of their applications [8, 11, 26, 46]. To overcome this issue, the Generalized Category Discovery (GCD) paradigm [1, 3, 28, 41, 44, 45, 49, 50] was introduced for open-world learning. It identifies both known and novel categories from unlabeled data by leveraging knowledge learned from labeled data with known classes. + +A key strategy for discovering novel classes is to transfer knowledge from labeled data to promote the learning of novel classes. Based on this idea, previous GCD methods [3, 41, 42, 44, 45] utilize contrastive learning [17, 39] to learn the representations of classification within labeled data. Initially, Vaze et al. [41] adopt a two-stage pipeline: contrastive learning for representation, followed by $k$ -means for classifying unlabeled data. However, clustering becomes computationally expensive as data scales. To reduce this, Wen et al. [45] proposed SimGCD, a parametric approach that replaces clustering with a classifier and learns class prototypes via joint contrastive learning. Building on this, Wang et al. [44] enhanced SimGCD with data prompts for better performance, while Cao et al. [3] + +addressed SimGCD's limitation in retaining known-class knowledge during transfer by introducing a regularization that improves both known and novel class performance. + +Despite advancements in parametric methods, recent studies reveal that supervised contrastive learning (S-CL), a key component of contrastive learning, has a limited effect in discovering novel classes. This is evident from the SimGCD baseline in Fig. 1a, where 'New' accuracy shows little improvement between that without S-CL and S-CL with only labeled data (orange vs blue bars) across both fine-grained and general datasets. In particular, accuracy even slightly declines when S-CL is limited to labeled data, as observed in Stanford Cars and CIFAR100. This indicates that current S-CL methods, constrained to labeled data, are inadequate for discovering novel classes, as the labeled data encompasses only known classes and omit the novel ones. Therefore, the key to improving the discovery of novel categories is to incorporate unlabeled data into S-CL, as it contains both known and novel samples. + +Motivated by this observation, we propose AllGCD, a straightforward approach that integrates all unlabeled data into contrastive learning to facilitate the novel category discovery. Specifically, AllGCD comprises two main components: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). To establish a stronger foundation for Inter-CU, Intra-CL first enhances intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data. Building on this refined knowledge, Inter-CU further improves inter-class separation between known and novel categories through global contrastive learning on their class distribution in unlabeled data. Finally, by combining Intra-CL and Inter-CU, we significantly improve novel category representations, as shown by the green bar in Fig. 1a. + +We validate our approach on six datasets, including three fine-grained datasets (CUB [43], Stanford Cars [18], FGVC-Aircraft [25]), three generic image recognition datasets (CIFAR100 [19] and ImageNet-100 [6]), and the challenging Herbarium-19 dataset [37]. As shown in Fig. 1a, our method outperforms strong baselines in recognizing novel ('New') classes, with a $7.4\%$ improvement over SimGCD on CUB. Additionally, Fig. 1b highlights that AllGCD focuses on more class-specific regions than the baseline. In summary, our main contributions are as follows: + +- We propose AllGCD to overcome the limitations of contrastive learning in parametric methods by utilizing all unlabeled data for novel class discovery. +- We introduce Intra-class Contrast in Labeled Data (IntraCL) to enhance intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data. +- Building on this refined knowledge, we design Inter-class + +Contrast in Unlabeled Data (Inter-CU) to improve interclass separation between known and novel categories by leveraging the global class distribution in unlabeled data. + +- Extensive results show that our method significantly improves the performance on novel classes, achieving a $7.4\%$ increase on CUB compared to the baseline. + +# 2. Related Works + +Novel Category Discovery (NCD) [12] assumes unlabeled data contains only novel classes and was first formalized as deep transfer clustering, where novel classes are clustered using knowledge from known classes. Various methods address this challenge, including Han et al. [12, 13] who use self-supervised learning and ranking statistics for knowledge transfer, Zhong et al. [52] who propose a mixup strategy [48] to combine old and novel classes, Gu et al. [10] who introduce a knowledge distillation framework for regularizing novel class learning, Zang et al. [47] who transfer knowledge from a source domain to a label-scarce target domain without label constraints, and Liu et al. [23] who apply NCD in ultra-fine-grained visual categorization. + +Generalized Category Discovery (GCD) [41] extends NCD by assuming that unlabeled data also contains known class images. Vaze et al. [41] formalize this with contrastive learning in representation space during training and clustering at inference. To reduce clustering cost, Wen et al. [45] introduce SimGCD, using a parametric classifier to learn class representation centers, serving as a strong baseline. Cao et al. [3] note SimGCD's issue of forgetting known classes while learning novel ones, proposing entropy regularization for potential known samples. Wang et al. [44] further extend SimGCD by optimizing model and data parameters for better alignment with the GCD task. Despite their performance, these methods rely on limited labeled images, constraining novel class discovery. + +Contrastive learning (CL) spans unsupervised [39] and supervised [17] variants for learning from unlabeled and labeled data, respectively. Unsupervised CL treats two augmented views of the same image as positives and all others as negatives, while supervised CL (S-CL) considers samples from the same class as positives. Vaze et al. [41] first combine both in GCD to enhance image representations. Choi et al. [5] employ mean-shifted embeddings for CL, and Pu et al. [28] propose DCCL, which explores conceptual-level CL to model relations between known and novel categories. However, these methods primarily apply S-CL within non-parametric GCD frameworks. In contrast, recent parametric approaches [3, 44, 45] demonstrate that S-CL can effectively train classifiers by learning class prototypes. Nonetheless, these works still rely heavily on limited labeled data, underutilizing the rich supervision signal from abundant unlabeled samples, which ultimately constrains the full potential of contrastive learning. + +# 3. Preliminaries and Analysis + +# 3.1. GCD setting and training + +Problem Formulation. Unlike traditional closed-set classification, where training and test classes coincide $(\mathcal{V}_l)$ , GCD aims to identify novel classes $\mathcal{Y}_u$ in the test data. Given labeled dataset $\mathcal{D}^l = \{(\pmb{x}_i,y_i)\} \subset \mathcal{X}\times \mathcal{Y}_l$ and unlabeled dataset $\mathcal{D}^u = \{(x_i,y_i)\} \subset \mathcal{X}\times \mathcal{Y}_u$ , GCD seeks to classify both known and novel samples in $\mathcal{D}^u$ using knowledge from $\mathcal{D}^l$ . The total class number $K = |\mathcal{Y}_l\cup \mathcal{Y}_u|$ is assumed known, following prior work [3, 9, 13, 44, 45]. + +Architecture. SimGCD [45] consists of a Vision Transformer (ViT) encoder [7], denoted as $\Phi$ , along with a classification head $f$ to generate pseudo-labels for self-distillation [2, 4], and a multi-layer perception (MLP) projection head $g$ for representation contrastive learning. + +Representation Learning. To learn representations for distinguishing samples, Vaze et al. [41] apply unsupervised contrastive learning across all data and supervised contrastive learning (S-CL) on labeled data. Given two views, $\boldsymbol{x}_i$ and $\boldsymbol{x}_i'$ , of an identical image as a positive pair in a minibatch $B$ , the unsupervised contrastive loss encourages their representations to be close in an unsupervised manner: + +$$ +\mathcal {L} _ {\mathrm {r e p}} ^ {u} = \frac {1}{| B |} \sum_ {i \in B} - \log \frac {\exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {i} ^ {\prime} / \tau_ {u}\right)}{\sum_ {n \in B , n \neq i} \exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {n} ^ {\prime} / \tau_ {u}\right)}, \tag {1} +$$ + +where $\boldsymbol{z}_i = g \circ \Phi(\boldsymbol{x}_i)$ and $\tau_u$ is a temperature. Meanwhile, the supervised contrastive loss minimizes the distance between same-class samples in feature space, defined as: + +$$ +\mathcal {L} _ {\text {r e p}} ^ {s} = \frac {1}{| B ^ {l} |} \sum_ {i \in B ^ {l}} \frac {1}{| \mathcal {N} _ {i} |} \sum_ {q \in N _ {i}} - \log \frac {\exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {q} ^ {\prime} / \tau_ {c}\right)}{\sum_ {n \in B , n \neq i} \exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {n} ^ {\prime} / \tau_ {c}\right)}, \tag {2} +$$ + +where $B^l$ is the labeled subset of $B$ , and $\mathcal{N}_i$ is the set of images with the same label as $x_i$ . The overall contrastive learning loss for representation is: + +$$ +\mathcal {L} _ {\text {r e p}} = (1 - \lambda) \mathcal {L} _ {\text {r e p}} ^ {u} + \lambda \mathcal {L} _ {\text {r e p}} ^ {s}, \tag {3} +$$ + +where $\lambda$ controls the balance between unsupervised and supervised contrastive learning. + +Parametric Clustering. Instead of the time-consuming $k$ -means, Wen et al. [45] propose an efficient parametric classifier $f$ using a set of prototypes $\mathcal{C} = \{c_1, \dots, c_K\}$ , where $K$ is the total classes number. For an image view $x_i$ , the prediction is made by calculating the cosine similarity between its feature, $\Phi(x_i)$ , and each prototype $c_k$ : + +$$ +\boldsymbol {p} _ {i} ^ {(k)} = \frac {\exp (\cos (\Phi (\boldsymbol {x} _ {i}) , \boldsymbol {c} _ {k}) / \tau_ {s})}{\sum_ {k ^ {\prime}} \exp (\cos (\Phi (\boldsymbol {x} _ {i}) , \boldsymbol {c} _ {k ^ {\prime}}) / \tau_ {s})}. \tag {4} +$$ + +Similarly, the soft pseudo-labels $\pmb{q}_i'$ for another view $\pmb{x}_i'$ are obtained using a smaller $\tau_t$ , and the cross-entropy loss is + +applied for prototype learning using these pseudo-labels $q_{i}^{\prime}$ : + +$$ +\mathcal {L} _ {\mathrm {d i s}} ^ {u} = - \frac {1}{| B |} \sum_ {i \in B} \sum_ {k} \boldsymbol {q} ^ {\prime (k)} \log \boldsymbol {p} ^ {(k)}. \tag {5} +$$ + +Additionally, SimGCD employs a class mean entropy regularization [2], $H(\overline{p}) = -\sum_{k} \overline{p}^{(k)} \log \overline{p}^{(k)}$ , where $\overline{p}$ is the batch-wise class-average prediction, forcing the model to focus on novel classes. The total classifier is: + +$$ +\mathcal {L} _ {\mathrm {c l s}} = \lambda \mathcal {L} _ {\mathrm {c l s}} ^ {s} + (1 - \lambda) \left(\mathcal {L} _ {\mathrm {d i s}} ^ {u} - \varepsilon H\right), \tag {6} +$$ + +where $\mathcal{L}_{\mathrm{cls}}^s$ is the cross-entropy loss for labeled data. + +# 3.2. Limited contrastive learning + +Analysis. Although contrastive learning (CL) enhances representation discriminability of known classes in + +SimGCD, it applies only to labeled data, as shown in the upper-left corner of Fig. 2. In the upper-left corner, same-colored entries along the vertical and horizontal indices correspond to samples from the same class, while different colors denote distinct classes. Self-contrast cases (blue-filled) on the diagonal are masked out. Obviously, abundant unlabeled data, shown with a gray background, are excluded from the + +![](images/65ce49b823f47ba9ad2363b27c39efa9647633ddf83df8bcd55202838d8cb719.jpg) +Figure 2. A class mask: Samples from the same class are shown in the same color in a batch $B$ . + +CL due to the absence of class labels. As a result, the model lacks direct contrastive supervision on novel samples, making it harder to separate them in feature space. + +# 4. Method + +To overcome the limited contrastive learning issue discussed in subsection 3.2, we propose a straightforward approach called AllGCD to integrate supervised contrastive learning across all unlabeled data directly. Specifically, we detail our proposed Intra-class Contrast in Labeled Data (Intra-CL) in subsection 4.1 and Inter-class Contrast in Unlabeled Data (Inter-CU) in subsection 4.2. + +# 4.1. Intra-class Contrast in Labeled Data + +Intra-CL first strengthens intra-class compactness within known classes for better separation of known and novel categories in Inter-CU. To achieve this, we use baseline SimGCD [45] to identify potential known samples in unlabeled data and incorporate them into the labeled set. However, as shown by the orange lines in Fig. 4a and Fig. 4b, SimGCD's known-class accuracy declines sharply during training, compromising the quality of potential known + +![](images/27fa5cf4ef6b67909731898e0504ed3b87a3d8293e4ce126ac0952b7f948f2d5.jpg) +Figure 3. Illustration of the proposed AllGCD. It is composed of Intra-CL and Inter-CU. First, Intra-CL (Sec. 4.1) enhances knowledge of known classes by incrementally collecting reliable pseudo-labels $\mathcal{V}^*$ within unlabeled data. Once this knowledge is strengthened, Inter-CU (Sec. 4.2) further promotes the separation of known and novel classes through contrastive learning between them. + +samples. To overcome this, we propose a voting strategy that utilizes the model's prior knowledge to identify potential known samples early. Specifically, Intra-CL comprises three steps: (1) Selecting potential known samples within unlabeled data as in [3]; (2) Storing these known samples into a dictionary $\mathcal{Q}$ ; (3) Voting pseudo-labels $\mathcal{Y}^*$ from $\mathcal{Q}$ based on their most frequent label across training iterations. + +Step (1): Selecting potential known samples within unlabeled data. + +In this step, we identify potential known samples by selecting instances where both views of the same image exhibit high confidence $(\delta)$ and are classified as known categories $(\mathcal{V}_l)$ . The specific calculation process is as follows. + +Let $B^u$ denote the unlabeled subset of $B$ , where each image in this batch has a unique index $\mathcal{U}_B = \{u_1, u_2, \ldots, u_b\}$ . Given the predictions $p_i$ and $q_i'$ from two views of an image in $B^u$ , as defined in Eq. (4). We first identify high-confidence samples by selecting those with predictions $p_i$ and $q_i'$ both exceed a confidence threshold $\delta$ : $\mathbf{M}_{\delta} = \mathbb{1}(\max(p_i) > \delta \land \max(q_i') > \delta)$ . + +Next, we classify a sample as known only if the predicted label from the two views match and both belong to known class set $\mathcal{Y}_l$ : $\mathbf{M}_{\mathrm{known}} = \mathbb{1}((\hat{y}_i = \hat{y}_i') \wedge (\hat{y}_i \in \mathcal{Y}_l))$ . + +Finally, the selected known samples within the unlabeled subset $B^{u}$ are determined as: + +$$ +\hat {\mathcal {Y}} _ {\text {s e l e c t}} = \left(\mathbf {M} _ {\delta} \odot \mathbf {M} _ {\text {k n o w n}}\right) \odot \hat {\mathcal {Y}}, \tag {7} +$$ + +where $\mathbf{M}_{\delta}$ and $\mathbf{M}_{\mathrm{known}}$ are binary (0-1) vectors of length b, + +$\odot$ is element-wise multiplication, and $\hat{\mathcal{Y}} = \{\hat{y}_i \mid i \in [1,b]\}$ represents the predicted labels across the batch of unlabeled samples. Thus, $\hat{\mathcal{Y}}_{\mathrm{select}}$ identifies the known samples $(\mathbf{M}_{\mathrm{known}} = 1)$ that also have high confidence $(\mathbf{M}_{\delta} = 1)$ . + +Step (2): Storing above predicted known samples. + +To leverage the model's early ability to recognize known classes, we initialize a dictionary $\mathcal{Q}$ to store the pseudolabels $\hat{\mathcal{Y}}_{\mathrm{select}}$ of each image based on its unique index $\mathcal{U}_B$ . + +Concretely, for each unlabeled subset $B^u$ , we update the dictionary $\mathcal{Q}:\mathcal{U}_B\to \hat{\mathcal{V}}_{\mathrm{select}}$ as: + +$$ +\mathcal {Q} \left(u _ {i}\right) = \left\{ \begin{array}{l l} \mathcal {Q} \left(u _ {i}\right) \cup \left\{\hat {y} _ {i} \in \hat {\mathcal {Y}} _ {\text {s e l e c t}} \right\} & \text {i f} u _ {i} \in \operatorname {d o m} (\mathcal {Q}), \\ \left\{\hat {y} _ {i} \in \hat {\mathcal {Y}} _ {\text {s e l e c t}} \right\} & \text {i f} u _ {i} \notin \operatorname {d o m} (\mathcal {Q}). \end{array} \right. \tag {8} +$$ + +Step (3): Voting on pseudo-labels for unlabeled data based on frequency. + +To obtain a reliable pseudo-label for each unlabeled image, we apply a voting mechanism over the stored pseudolabels $\mathcal{Q}$ across training batches. Specifically, for a batch of unlabeled samples, we select the most frequently assigned pseudo-label from its historical records in $\mathcal{Q}(u_i)$ : + +$$ +\mathcal {Y} ^ {*} = \underset {y} {\arg \max } \left(\operatorname {f r e q} (\hat {y} \in \mathcal {Q} (u _ {i})), \forall u _ {i} \in \mathcal {U} _ {B}\right), \tag {9} +$$ + +where $\mathcal{Y}^* = \{y_1^*, y_2^*, \ldots, y_b^*\}$ denotes the pseudo-labels for unlabeled samples in $B^u$ . The updated mini-batch labels, incorporating both original labeled data $\mathcal{Y}_l$ and pseudolabels $\mathcal{Y}^*$ , are given by $\mathcal{Y}_{\mathrm{intra}} = \mathcal{Y}_l \cup \mathcal{Y}^*$ . Consequently, + +![](images/a7ad01421434797eab5d06eab787b1bee2bde3f215b9297fe8f642b17b9118f1.jpg) +(a) CUB + +![](images/7637b0bf5a8c93340ed67f55718bf7797a8125cfe8b45ab42085f6b4d443fc3f.jpg) +(b) Stanford Cars +Figure 4. Known-class accuracy on unlabeled data during training. SimGCD, using S-CL without unlabeled data, shows a sharp drop on CUB and Stanford Cars. In contrast, Intra-class leverages potential known samples in unlabeled data, improving the knowledge of known classes. + +the supervised contrastive loss in Eq. (2) is updated with pseudo known samples from unlabeled data as: + +$$ +\mathcal {L} _ {\text {i n t r a}} = \frac {1}{| B ^ {\text {i n t r a}} |} \sum_ {i \in B ^ {\text {i n t r a}}} \frac {1}{| \mathcal {N} _ {i} |} \sum_ {j \in \mathcal {N} _ {i}} - \log \frac {\exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {j} ^ {\prime} / \tau_ {c}\right)}{\sum_ {n} \mathbb {1} _ {[ n \neq i ]} \exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {n} ^ {\prime} / \tau_ {c}\right)}, \tag {10} +$$ + +where $B^{\mathrm{intra}}$ is the newly labeled subset of mini-batch $B$ with labels $\mathcal{V}_{\mathrm{intra}}$ , and $\mathcal{N}_i$ indexes same-label images as $x_i$ in the batch. After applying Intra-CL, the accuracy of known classes in unlabeled data remains more stable (Fig. 4), indicating more compact representations within known classes. + +# 4.2. Inter-class Contrast in Unlabeled Data + +While Inter-CL enhances discrimination among known classes, it lacks contrastive interactions with novel classes, limiting novel category accuracy (Fig. 1a). To address this, we propose Inter-CU, which enforces global contrastive learning on the unlabeled class distribution. Specifically, once the model has stabilized on known and novel categories from epoch $e$ , we use the Hungarian algorithm [21] on the predictions of all unlabeled data to optimize the class distribution of known and novel categories. To ensure a high-quality distribution, we retain only high-confidence predictions $(P > \gamma)$ for global contrastive learning. + +Formally, given the unlabeled dataset $\mathcal{D}^u = \{(x_i,y_i)\} \in \mathcal{X}\times \mathcal{Y}_u$ we obtain classifier predictions $P = [p_1,p_2,\dots ,p_N]$ via Eq. (4), where $N$ is the total number of unlabeled samples. The minimum-cost assignment function $\pi$ is defined as: + +$$ +\pi^ {*} = \arg \min _ {\pi} \sum_ {i = 1} ^ {N} - P _ {i, \pi (i)}, \quad \pi (i) \in \{1, 2, \dots , K \}, \tag {11} +$$ + +where $\pi^{*}$ assigns each sample in $\mathcal{D}^u$ to one of $K$ categories, minimizing the sum of costs $-P_{i,\pi (i)}$ . The Hungarian algorithm efficiently solves this assignment problem by finding the optimal mapping with minimal total cost. Therefore, the + +Algorithm 1 One Training Procedure for AllGCD +Input: Encoder $\Phi$ , classifier head $f$ , projector $g$ , labeled data $\mathcal{D}^l$ , unlabeled data $\mathcal{D}^u$ , hyperparameters $\delta, e$ . +Output: $\Phi, f$ , and $g$ ; +1: Initialize dictionary $\mathcal{Q} \gets \emptyset$ ; +2: for epoch = 1 to total epochs do +3: if epoch ≥ e then +4: Get assignment $\mathcal{Y}_{\text{inter}}$ for $D^u$ using Eq. (12); +5: end if +6: for $(x, x') \in \mathcal{D}^l \cup \mathcal{D}^u$ , $y \in \mathcal{Y}_l$ , $y_{\text{inter}} \in \mathcal{Y}_{\text{inter}}$ do +7: $z \gets g \circ \Phi(x)$ , $z' \gets g \circ \Phi(x')$ ; +8: $p \gets f \circ \Phi(x)$ , $q' \gets f \circ \Phi(x')$ ; +9: Compute $\mathcal{L}_{\text{cls}}^s(p, y) + \mathcal{L}_{\text{cls}}^s(q', y)$ ; ▷ For $\mathcal{D}^l$ +10: Compute $\mathcal{L}_{\text{dis}}^u(p, q')$ using Eq. (5); +11: Compute Max class Entropy $H(p)$ ; +12: Compute $\mathcal{L}_{\text{cls}} \gets \lambda \mathcal{L}_{\text{cls}}^s + (1 - \lambda) (\mathcal{L}_{\text{dis}}^u - \varepsilon H)$ ; +13: Step 1: Generate $\hat{y} \gets (p, q', \delta)$ ; ▷ For $\mathcal{D}^u$ +14: Step 2: Store $\hat{y}$ into $\mathcal{Q}$ using Eq. (8); +15: Step 3: Vote for $y^*$ from $\mathcal{Q}$ using Eq. (9); +16: Combine with pseudo-labels $y_{\text{intra}} \gets y \cup y^*$ ; +17: Compute $\mathcal{L}_{\text{rep}}^u(z, z')$ using Eq. (1); +18: Compute $\mathcal{L}_{\text{intra}}(z, z', y_{\text{intra}})$ using Eq. (10); +19: Compute $\mathcal{L}_{\text{inter}}(z, z', y_{\text{inter}})$ using Eq. (12); +20: $\mathcal{L}_{\text{all}} = (1 - \lambda) \mathcal{L}_{\text{rep}}^u + \lambda (\mathcal{L}_{\text{intra}} + \mathcal{L}_{\text{inter}}) + \mathcal{L}_{\text{cls}}$ ; +21: Update $\Phi, f$ , and $g$ by SGD [29]; +22: end for +23: end for + +final assignments for unlabeled data are: + +$$ +\mathcal {Y} _ {\text {i n t e r}} = \left\{\hat {y} _ {i} = \pi^ {*} (i) \mid i = 1, 2, \dots , N \right\} \odot \mathbf {M} _ {\text {i n t e r}}, \tag {12} +$$ + +where $\mathbf{M}_{\mathrm{inter}}$ is a binary mask that applies a confidence threshold $\gamma$ to filters out low-confidence predictions: $\mathbf{M}_{\mathrm{inter}} = \{i \mid \max(\pmb{p}_i) > \gamma\}$ . Only high-confidence samples contribute to the global contrastive loss: + +$$ +\mathcal {L} _ {\text {i n t e r}} = \frac {1}{| B ^ {\text {i n t e r}} |} \sum_ {i \in B ^ {\text {i n t e r}}} \frac {1}{| \mathcal {N} _ {i} |} \sum_ {j \in \mathcal {N} _ {i}} - \log \frac {\exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {j} ^ {\prime} / \tau_ {\mathrm {c}}\right)}{\sum_ {n} \mathbb {1} _ {[ n \neq i ]} \exp \left(\boldsymbol {z} _ {i} \cdot \boldsymbol {z} _ {n} ^ {\prime} / \tau_ {\mathrm {c}}\right)}, \tag {13} +$$ + +where $B^{\mathrm{inter}}$ corresponds to the refined labeled subset with pseudo-labels $\mathcal{V}_{\mathrm{inter}}$ . This approach enables global contrastive learning on both known and novel samples, enhancing category discovery, particularly for novel classes. + +Overall. Utilizing all unlabeled data in supervised contrastive learning significantly boosts class discovery, especially for novel classes. This overall training loss is updated as: + +$$ +\mathcal {L} _ {\text {a l l}} = (1 - \lambda) \mathcal {L} _ {\text {i n t r a}} + \lambda \mathcal {L} _ {\text {i n t e r}} + \mathcal {L} _ {\text {c l s}}, \tag {14} +$$ + +where $\lambda$ controls the trade-off between Intra-CL and Inter-CU. Algorithm 1 details one training step of AllGCD. + +Table 1. Comparison with State-of-the-Art GCD Methods, Evaluated with and without Ground-Truth $K$ for Clustering. Bold denotes the best results. + +
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
(a) Clustering with the ground-truth number of classes K given
k-means[24]52.052.250.872.775.571.334.338.932.112.810.613.816.014.416.813.012.213.4
RS+[13]58.277.619.337.161.624.833.351.624.228.361.812.126.936.422.227.955.812.8
UNO+[9]69.580.647.270.395.057.935.149.028.135.570.518.640.356.432.228.353.714.7
ORCA[22]69.077.452.073.592.663.935.345.630.223.550.110.722.031.817.120.930.915.5
GCD [41]73.076.266.574.189.866.351.356.648.739.057.629.945.041.146.935.451.027.0
DCCL [28]75.376.870.280.590.576.263.560.864.9---43.155.736.2---
PromptCAL [50]81.284.275.383.192.778.362.964.462.150.270.140.652.252.252.337.052.028.9
SimGCD [45]80.181.277.883.093.177.960.365.657.753.871.945.054.259.151.844.058.036.4
InfoSieve [30]78.382.270.580.593.873.869.477.965.255.774.846.456.363.752.541.055.433.2
SPTNet [44]81.384.375.685.493.281.465.868.865.159.079.249.359.361.858.143.458.735.2
LegoGCD [3]81.881.482.586.394.582.163.871.959.857.375.748.455.061.551.745.157.438.5
CMS [5]82.385.775.584.795.679.268.276.564.056.063.452.356.976.147.636.454.926.4
AllGCD (ours)82.382.382.286.594.782.368.475.165.160.577.052.557.462.454.945.258.338.0
(b) Clustering without the ground-truth number of classes K given
GCD [41]70.877.657.077.991.171.351.156.448.439.158.629.7---37.251.729.4
GPC [51]75.484.660.175.393.466.752.055.547.538.258.927.443.340.744.836.551.727.9
SimGCD [45]80.181.277.881.791.276.861.566.459.149.165.141.3------
CMS [5]79.683.272.381.395.674.264.468.262.451.768.943.455.260.652.437.456.527.1
AllGCD (ours)81.583.378.484.494.380.266.772.164.058.875.750.656.764.253.045.758.738.5
+ +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. We evaluate our approach on three fine-grained datasets: CUB [43], Stanford Cars [18], and FGVC-Aircraft [25], as well as two generic image recognition datasets, CIFAR100 [19] and ImageNet-100 [6]. Additionally, we test on a more challenging long-tailed fine-grained Herbarium 19 [37]. Following [41], we randomly sample $50\%$ of seen categories as labeled data $\mathcal{D}^l$ , with the remaining images from seen and novel categories forming the unlabeled set $\mathcal{D}^u$ . Details are in the supplementary material. + +Evaluation protocol. We use clustering accuracy (ACC) to evaluate model performance, following [41, 45]. During evaluation, we compare the ground-truth labels $y_{i}$ with the predicted labels $\hat{y}_i$ , and calculate ACC as $ACC = \frac{1}{|\mathcal{D}^u|}\sum_{i=1}^{|\mathcal{D}^u|}\mathbb{1}(y_i = p(\hat{y}_i))$ , where $p$ is determined using the Hungarian optimal assignment algorithm [21]. + +Implementation details. AllGCD is built upon SimGCD [45], using a ViT-B/16 backbone [7] pretrained with DINO [4]. We use the class token to train a classifier and fine-tune only the last attention block. We use a batch size of 128 and an initial learning rate of 0.1, decayed over 200 epochs using a cosine schedule. Following [41], we set $\lambda = 0.35$ , $\tau_{u} = 0.07$ , and $\tau_{c} = 1.0$ for representation learning, and use $\tau_{s} = 0.1$ , with $\tau_{t}$ decayed from 0.07 to 0.04 over 30 epochs for classifier training. Pseudo-label thresholds $\delta$ and $\gamma$ range from 0.6 to 0.9. Based on the convergence time of 'New' classes in SimGCD (Fig. 7a), the Inter-CU start epoch $e$ is selected + +from $\{0,20,40,60,80\}$ . All experiments are conducted using PyTorch on NVIDIA Tesla V100 GPUs. + +# 5.2. Comparison With the State of the Arts + +To demonstrate the effectiveness of AllGCD, we evaluate it under both known and unknown total class numbers $K$ : (1) In Table 1(a), we assume $K$ is known, consistent with prior works [3, 9, 13, 44, 45]; (2) In Table 1(b), we report results where $K$ is unknown, estimating it using clustering techniques as in [5, 41, 51]. + +Evaluation with ground-truth class number $K$ . Table 1(a) compares state-of-the-art methods assuming a known class number $K$ . Our method significantly enhances 'New' category accuracy compared to the baseline SimGCD, with notable improvements of $7.4\%$ , $7.5\%$ , and $4.4\%$ on CUB, Stanford Cars, and ImageNet-100, respectively. Furthermore, AllGCD remains competitive with SPTNet, surpassing it by $2.6\%$ and $1.5\%$ on the 'All' classes in CUB and Stanford Cars, respectively. Importantly, our method attains comparable performance to SPTNet using only 200 training epochs, compared to SPTNet's 1000 epochs, showing superior computational efficiency. Though AllGCD trails CMS [5] by $2.7\%$ in 'Old' accuracy on CIFAR100, it outperforms CMS by $6.7\%$ in 'New' accuracy, highlighting its strength in novel category discovery. + +Evaluation with estimated number of clusters. Table 1 (b) presents a comparison conducted without access to the ground-truth class number $K$ , instead using an estimated $K$ derived from an off-the-shelf method [41], as shown in Table 2. Table 2 shows our approach provides a more accu + +
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓
GT100-100-200-196-100-683-
GCD [41]1000109923115.523017.3--52023.8
DCCL [28]146461292917291920.02----
SimGCD [45]1000109923115.523017.3--52023.8
GPC [51]1000103321262010.03----
CMS [5]955116161701515620.490106228.9
AllGCD (ours)95510882136.52107.110996366.9
+ +![](images/c5b23e4e9c3d533c2abd4f0032b5260a0af76f97a98d9a3dbc8ec634a8b82e50.jpg) +(c) SimGCD [45] + +![](images/1c4744ce16b86f4a8182e54cd6fef3ff65ca8a488a6c9258548ec8f4d7a00d19.jpg) +(d) AllGCD + +Table 2. Estimated class number and error rate for $K$ using off-the-self method [41]. + +
δCUBCIFAR100
AllOldNewAllOldNew
0.7068.274.664.980.881.380.5
0.7568.475.165.181.481.881.3
0.8068.275.764.582.283.181.8
0.8567.274.763.482.382.382.2
0.9066.273.362.581.681.981.5
+ +Table 3. Effect of confidence $\delta$ in Intra-CL. + +![](images/075ea49b201b9b00582a424f7d97d9b0d3c73b0393dae54aef3b3d3ec219cee0.jpg) +(a) CUB with $\delta = 0.75$ +Figure 6. Logit distributions of unlabeled data. + +![](images/4f3ebc4713b108c3dcc7eabc7183de9ca176223cb488b99bf5eff325d04aae7b.jpg) +Figure 5. T-SNE visualization of 10 random novel classes from CIFAR100. +(b) CIFAR100 with $\delta = 0.85$ + +rate estimation of $K$ , with fewer errors in determining cluster centers than SimGCD and CMS, except for CIFAR100 in SimGCD. Under this improved estimation, our approach consistently outperforms state-of-the-art methods across all datasets. Notably, performance gains are particularly significant on the Stanford Cars, ImageNet-100, and Herbarium-19 benchmarks, where AllGCD achieves 7.2%, 6.0%, and 9.1% higher accuracy in the 'New' category compared to the second-best results. These findings highlight the effectiveness of our method in utilizing all available unlabeled data to learn more discriminative representations. + +Feature Visualizations and Attention Distributions. To demonstrate the effectiveness of our approach in representation learning, we visualize 10 randomly selected novel classes from CIFAR100. As shown in Fig. 5, AllGCD produces more compact and well-separated clusters, whereas SimGCD exhibits considerable cluster overlap. Additionally, we visualize attention on both known and novel samples from unlabeled data in Table 4. Ours focuses on class-specific regions (e.g., the bird's wing) and suppressing background noise (e.g., the window in the second row), highlighting that leveraging unlabeled data enhances attention to relevant regions and improves class discriminability. + +# 5.3. Ablation Study + +In this section, we present ablation studies to assess the effectiveness of Intra-CL and Inter-CU in our AllGCD model. The evaluation includes fine-grained datasets such as CUB and Stanford Cars, alongside generic image recognition datasets CIFAR100 and ImageNet-100. + +Effect of different confidence $\delta$ . Table 3 shows the effect of varying the confidence threshold $\delta$ in Intra-CL on + +Table 4. Average attention across 12 heads in the final layer of the image encoder on CUB and FGVC-Aircraft. The known sample (top row) and novel sample (bottom row) in AllGCD show improved focus on object regions with reduced background noise. + +
ImageDINOSimGCDAllGCD
+ +CUB and CIFAR100. Overall, the accuracy for both 'Old' and 'New' classes initially improves with increasing $\delta$ but slightly drops beyond the optimal point. The fine-grained CUB dataset performs best at $\delta = 0.75$ , while the generic dataset CIFAR100 peaks at $\delta = 0.85$ . This difference arises from the varying proportions of unlabeled samples preserved in fine-grained and generic datasets. Higher thresholds, such as 0.85, help retain sufficient known samples in CIFAR100 (see Fig. 5b), but may exclude too many in CUB (see Fig. 5a), resulting in suboptimal performance for the fine-grained dataset. Therefore, we set $\delta = 0.75$ for fine-grained datasets and $\delta = 0.85$ for generic datasets. + +Effect of different confidence $\gamma$ . Similar to $\delta$ in Intra-CL for selecting high-confidence known samples, $\gamma$ filters high-quality distributions for inter-class contrastive learning in Inter-CU. Table 5 shows its effect on Stanford Cars and + +Table 5. Ablation study on different thresholds $\gamma$ in Inter-CU. + +
γStanford CarsImageNet-100
AllOldNewAllOldNew
0.6060.076.852.285.693.881.5
0.6560.577.052.586.094.181.7
0.7059.876.752.186.594.782.3
0.7558.975.151.286.394.982.1
0.8057.673.750.186.194.481.9
+ +Table 6. Effectiveness of each component of our method. Intra-CL and Inter-CU are described in Sec. 4.1 and 4.2, respectively. Row (1) shows the baseline SimGCD results. + +
IndexSimGCDIntra-CLInter-CUCUBCIFAR100
AllOldNewAllOldNew
(1)✓60.365.657.780.181.277.8
(2)✓✓65.273.161.381.281.879.6
(3)✓✓67.172.863.881.681.581.7
(4)✓✓✓68.475.165.182.382.382.2
+ +ImageNet-100. Specifically, we fix $\delta = 0.75$ for Stanford Cars and $\delta = 0.85$ for ImageNet-100 based on the $\delta$ ablation. Intuitively, Stanford Cars performs best with $\gamma = 0.65$ , while ImageNet-100 performs best with $\gamma = 0.7$ . Similar to Intra-CL, this difference is mainly due to the varying number of selected samples at each threshold: $6.1\mathrm{k}$ vs. $95.3\mathrm{k}$ (see Data in appendix). ImageNet-100 retains more high-confidence samples at a stricter threshold (e.g., 0.75), while fine-grained datasets yield fewer, affecting Inter-CU's effectiveness. Therefore, we set $\gamma = 0.65$ or 0.6 for fine-grained datasets and $\gamma = 0.70$ or 0.75 for generic ones. + +Effect of each proposed component. Table 6 presents the ablation study on our proposed Intra-CL and Inter-CU. Compared to SimGCD (Row (1)), Intra-CL (Row (2)) improves 'Old' class accuracy by $7.5\%$ on CUB and $0.6\%$ on CIFAR100, demonstrating its effectiveness in refining known sample selection for supervised contrastive learning. This also leads to a $3.6\%$ and $1.8\%$ improvement in 'New' accuracy on CUB and CIFAR100, respectively. Inter-CU (Row (3)) further enhances 'New' accuracy by $6.1\%$ on CUB and $3.9\%$ on CIFAR100 by fully utilizing unlabeled data for novel class discovery. Combining both (Row (4)) achieves the best overall performance, confirming their complementary benefits in improving recognition of both known and novel classes. + +Effect of starting epoch $e$ in Inter-CU. Fig. 7b reports 'New' accuracy for different Inter-CU start epochs $e$ . Fig. 7a shows that SimGCD struggles to separate known and novel classes early (e.g., during the first 20 epochs). This difficulty leads to an unreliable class distribution, which negatively impacts Inter-CU since it depends on well-separated inter-class relationships. To address this, we evaluate $e \in \{0, 20, 40, 60, 80\}$ , where $e = 0$ denotes All- + +![](images/eb54a748e68428cd78d7492e6d2ba1f0bb6c3c0ec76b49a76c81d5ada3ff9c63.jpg) +(a) 'New' ACC. in SimGCD [45] + +![](images/d399f2d06f5bc41ae1d043ea53d33b019ee402d29042153c9cd3cbd7fe5a6220.jpg) +(b) 'New' ACC. @ Epochs +Figure 7. 'New' class accuracy on unlabeled data. (a) In SimGCD, 'New' ACC. remains low in the early training stages (e.g. below $40\%$ for FGVC-Aircraft and Stanford Cars at epoch $\leq 20$ .) (b) Inter-CU improves accuracy more effectively in the middle and late stages ( $e.g. e \geq 40$ ). + +GCD without Inter-CU (only Intra-CL). Results show that Stanford Cars and FGVC-Aircraft peak at later epochs (e.g., $e = 80$ ), while CUB and CIFAR100 peak at $e = 40$ . This aligns with CUB's higher early-stage known class recognition compared to Stanford Cars and FGVC-Aircraft (60% vs. ≈43%), enabling earlier effective separation of known and novel samples. CIFAR100 is less sensitive to $e$ , reaching approximately 70% accuracy by epoch 40, yielding stable class distributions for Inter-CU. + +# 6. Conclusion + +This paper identifies that previous supervised contrastive learning (S-CL) in parametric GCD algorithms limits novel category discovery due to its restricted use of labeled data. To address this, we propose AllGCD, a straightforward approach that leverages all unlabeled data containing novel classes into S-CL. Specifically, we introduce two strategies named Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first enhances the intra-class compactness within known classes by incorporating potential known samples from unlabeled data into the original S-CL. Building on this, Inter-CU further separates known and novel classes by applying S-CL to a global class distribution. By combining these strategies, AllGCD maximizes the potential of S-CL, achieving substantial improvements in the discovery of novel categories. Extensive experiments show that AllGCD significantly outperforms baselines and achieves competitive results compared to other GCD methods. + +# Acknowledgments + +This work was supported by the Guangdong Science and Technology Program (Grant No. 2024B0101040005), the National Natural Science Foundation of China (Grant No. 62332002 and 61825101), the Shenzhen High-Level Talent Team Program, and the Shenzhen Science and Technology Innovation Commission (Grant No. KQTD20240729102051063). + +# References + +[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12527-12535, 2023. 1 +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Proceedings of the European conference on Computer Vision (ECCV), pages 456-473, 2022. 3 +[3] Xinzi Cao, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16880-16889, 2024. 1, 2, 3, 4, 6 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of IEEE/CVF International Conference on Computer Vision (CVPR), pages 9630-9640, 2021. 1, 3, 6 +[5] Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23094-23104, 2024. 2, 6, 7, 1 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 6 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 1, 3, 6 +[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14064-14073, 2022. 1 +[9] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264-9272, 2021. 3, 6 +[10] Peiyan Gu, Chuyu Zhang, Ruijie Xu, and Xuming He. Classrelation knowledge distillation for novel class discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16428-16437, 2023. 2 + +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 1 +[12] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 8400-8408, 2019. 2 +[13] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6767-6781, 2022. 2, 3, 6 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 1 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2980-2988, 2017. 1 +[16] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 1 +[17] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2 +[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCV), pages 554–561, 2013. 2, 6, 1 +[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 1 +[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), pages 1106-1114, 2012. 1 +[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 6 +[22] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJ-CAI, pages 4002–4010. ijcai.org, 2023. 6 +[23] Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17679-17688, 2024. 2 +[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of + +the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297, 1967. 6 +[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 6, 1 +[26] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. In Proceedings of Conference on European Conference on Computer Vision (ECCV), pages 728-755, 2022. 1 +[27] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 2 +[28] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptual contrastive learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7579-7588, 2023. 1, 2, 6, 7 +[29] Ning Qian. On the momentum term in gradient descent learning algorithms. Neural Networks, 12(1):145-151, 1999. 5 +[30] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6 +[31] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 91-99, 2015. 1 +[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 1 +[33] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 1 +[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1 +[35] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023. 1 +[36] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 1 + +[37] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 2, 6, 1 +[38] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Proceedings of the European conference on Computer Vision (ECCV), pages 776-794, 2020. 1 +[39] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 1, 2 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1 +[41] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482-7491, 2022. 1, 2, 3, 6, 7 +[42] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2 +[43] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 1 +[44] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In Proceedings of International Conference on Learning Representations (ICLR), 2024. 1, 2, 3, 6 +[45] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16590-16600, 2023. 1, 2, 3, 6, 7, 8 +[46] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 46(7):5092-5113, 2024. 1 +[47] Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z. Li. Boosting novel category discovery over domains with soft contrastive learning and all in one classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11824-11833, 2023. 2 +[48] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proceedings of International Conference on Learning Representations (ICLR), 2018. 2 +[49] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. In Proceedings of International Conference on Machine Learning, (ICML), pages 41498-41509, 2023. 1 + +[50] Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3479-3488, 2023. 1, 6 +[51] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16623-16633, 2023. 6, 7, 1 +[52] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9462-9470, 2021. 2 \ No newline at end of file diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/images.zip b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4500c157ffe51942d5a14a8bf5c8992bb68c2be1 --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0c3bb2455e81c2b25ba2912ed946f854cde6f970c0b519c360a8175bd834db7 +size 658870 diff --git a/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/layout.json b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8e4ca9f03e7e3a7767f2f0f613da54d93ef28ca5 --- /dev/null +++ b/2025/AllGCD_ Leveraging All Unlabeled Data for Generalized Category Discovery/layout.json @@ -0,0 +1,11940 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 542, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 542, + 122 + ], + "type": "text", + "content": "AllGCD: Leveraging All Unlabeled Data for Generalized Category Discovery" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": "Xinzi Cao" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": ", Ke Chen" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": ", Feidiao Yang" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": ", Xiawu Zheng" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": ", Yonghong Tian" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "text", + "content": ", Yutong Lu" + }, + { + "bbox": [ + 94, + 142, + 521, + 158 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 159, + 530, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 159, + 530, + 175 + ], + "spans": [ + { + "bbox": [ + 81, + 159, + 530, + 175 + ], + "type": "text", + "content": "1 Sun Yat-sen University, 2 Peng Cheng Laboratory, 3 Xiamen University, 4 Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 162, + 178, + 447, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 178, + 447, + 191 + ], + "spans": [ + { + "bbox": [ + 162, + 178, + 447, + 191 + ], + "type": "text", + "content": "{caoxz, chenk68}@mail2.sysu.edu.cn yangfd@pcl.ac.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 194, + 487, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 194, + 487, + 205 + ], + "spans": [ + { + "bbox": [ + 117, + 194, + 487, + 205 + ], + "type": "text", + "content": "zhengxiawu@xmu.edu.cn yhtian@pku.edu.cn luyutong@mail.sysu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "text", + "content": "Generalized Category Discovery (GCD) aims to identify both known and novel categories in unlabeled data by leveraging knowledge from labeled datasets. Current methods employ supervised contrastive learning on labeled data to capture known category structures but neglect unlabeled data, limiting their effectiveness in classifying novel classes, especially in fine-grained open-set detection where subtle class differences are crucial. To address this issue, we propose a novel learning approach, AllGCD, which seamlessly integrates all unlabeled data into contrastive learning to enhance the discrimination of novel classes. Specifically, we introduce two key techniques: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first refines intra-class compactness within known categories by integrating potential known samples into labeled data. This process refines the decision boundaries of known categories, reducing ambiguity when distinguishing novel categories. Building on this, Inter-CU further strengthens inter-class separation between known and novel categories by applying global contrastive learning to the class distribution in the unlabeled data. By jointly leveraging Intra-CL and Inter-CU, AllGCD effectively improves both intra-class compactness and inter-class separation, effectively enhancing the discriminability between known and novel classes. Experiments demonstrate that AllGCD significantly improves novel classes accuracy, e.g., achieving increases of " + }, + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "text", + "content": " on CUB and " + }, + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 55, + 258, + 296, + 593 + ], + "type": "text", + "content": " on Stanford Cars." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 135, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Deep learning has achieved outstanding performance in computer vision tasks [4, 7, 15, 31, 32, 36, 40], especially in image classification [14, 16, 20, 33, 34, 36]. However, most conventional methods are developed in a closed setting where all training and test data belong to pre-defined" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 436, + 327 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 436, + 327 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 436, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 436, + 327 + ], + "type": "image", + "image_path": "7fa28c778791a9f66da1ff2941555d0ed0b452a00512dd08b112e10a483e9ba7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 329, + 432, + 338 + ], + "lines": [ + { + "bbox": [ + 319, + 329, + 432, + 338 + ], + "spans": [ + { + "bbox": [ + 319, + 329, + 432, + 338 + ], + "type": "text", + "content": "(a) 'New' acc w/o CL in SimGCD" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 232, + 553, + 328 + ], + "blocks": [ + { + "bbox": [ + 440, + 232, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 440, + 232, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 440, + 232, + 553, + 328 + ], + "type": "image", + "image_path": "f56d58f8e2732f1cc23fca788ac7fc9a6c657ad3308aee748903890a3907b090.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 329, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 440, + 329, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 440, + 329, + 548, + 338 + ], + "type": "text", + "content": "(b) Attentions w/o unlabeled data" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Comparison of 'New' accuracy and attention w/o unlabeled data in supervised contrastive learning (S-CL). (a) reveals that S-CL solely with labeled data fails to classify novel samples (orange bar vs. blue bar) in the baseline SimGCD. In contrast, our method with unlabeled data (green bar) significantly improves accuracy. (b) shows that our AllGCD highlights more class-specific object regions by incorporating unlabeled data." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 544 + ], + "type": "text", + "content": "known) classes. It weakens the capability of these methods to handle novel categories, limiting the scope of their applications [8, 11, 26, 46]. To overcome this issue, the Generalized Category Discovery (GCD) paradigm [1, 3, 28, 41, 44, 45, 49, 50] was introduced for open-world learning. It identifies both known and novel categories from unlabeled data by leveraging knowledge learned from labeled data with known classes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "A key strategy for discovering novel classes is to transfer knowledge from labeled data to promote the learning of novel classes. Based on this idea, previous GCD methods [3, 41, 42, 44, 45] utilize contrastive learning [17, 39] to learn the representations of classification within labeled data. Initially, Vaze et al. [41] adopt a two-stage pipeline: contrastive learning for representation, followed by " + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "-means for classifying unlabeled data. However, clustering becomes computationally expensive as data scales. To reduce this, Wen et al. [45] proposed SimGCD, a parametric approach that replaces clustering with a classifier and learns class prototypes via joint contrastive learning. Building on this, Wang et al. [44] enhanced SimGCD with data prompts for better performance, while Cao et al. [3]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3293" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "addressed SimGCD's limitation in retaining known-class knowledge during transfer by introducing a regularization that improves both known and novel class performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 301 + ], + "type": "text", + "content": "Despite advancements in parametric methods, recent studies reveal that supervised contrastive learning (S-CL), a key component of contrastive learning, has a limited effect in discovering novel classes. This is evident from the SimGCD baseline in Fig. 1a, where 'New' accuracy shows little improvement between that without S-CL and S-CL with only labeled data (orange vs blue bars) across both fine-grained and general datasets. In particular, accuracy even slightly declines when S-CL is limited to labeled data, as observed in Stanford Cars and CIFAR100. This indicates that current S-CL methods, constrained to labeled data, are inadequate for discovering novel classes, as the labeled data encompasses only known classes and omit the novel ones. Therefore, the key to improving the discovery of novel categories is to incorporate unlabeled data into S-CL, as it contains both known and novel samples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 303, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 495 + ], + "type": "text", + "content": "Motivated by this observation, we propose AllGCD, a straightforward approach that integrates all unlabeled data into contrastive learning to facilitate the novel category discovery. Specifically, AllGCD comprises two main components: Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). To establish a stronger foundation for Inter-CU, Intra-CL first enhances intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data. Building on this refined knowledge, Inter-CU further improves inter-class separation between known and novel categories through global contrastive learning on their class distribution in unlabeled data. Finally, by combining Intra-CL and Inter-CU, we significantly improve novel category representations, as shown by the green bar in Fig. 1a." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "We validate our approach on six datasets, including three fine-grained datasets (CUB [43], Stanford Cars [18], FGVC-Aircraft [25]), three generic image recognition datasets (CIFAR100 [19] and ImageNet-100 [6]), and the challenging Herbarium-19 dataset [37]. As shown in Fig. 1a, our method outperforms strong baselines in recognizing novel ('New') classes, with a " + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": " improvement over SimGCD on CUB. Additionally, Fig. 1b highlights that AllGCD focuses on more class-specific regions than the baseline. In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 653 + ], + "type": "text", + "content": "- We propose AllGCD to overcome the limitations of contrastive learning in parametric methods by utilizing all unlabeled data for novel class discovery." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": "- We introduce Intra-class Contrast in Labeled Data (IntraCL) to enhance intra-class compactness within known categories by incorporating potential known samples from unlabeled data into the labeled data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "- Building on this refined knowledge, we design Inter-class" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 108 + ], + "type": "text", + "content": "Contrast in Unlabeled Data (Inter-CU) to improve interclass separation between known and novel categories by leveraging the global class distribution in unlabeled data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 553, + 144 + ], + "type": "text", + "content": "- Extensive results show that our method significantly improves the performance on novel classes, achieving a " + }, + { + "bbox": [ + 315, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 315, + 108, + 553, + 144 + ], + "type": "text", + "content": " increase on CUB compared to the baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 155, + 406, + 167 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 175, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 331 + ], + "type": "text", + "content": "Novel Category Discovery (NCD) [12] assumes unlabeled data contains only novel classes and was first formalized as deep transfer clustering, where novel classes are clustered using knowledge from known classes. Various methods address this challenge, including Han et al. [12, 13] who use self-supervised learning and ranking statistics for knowledge transfer, Zhong et al. [52] who propose a mixup strategy [48] to combine old and novel classes, Gu et al. [10] who introduce a knowledge distillation framework for regularizing novel class learning, Zang et al. [47] who transfer knowledge from a source domain to a label-scarce target domain without label constraints, and Liu et al. [23] who apply NCD in ultra-fine-grained visual categorization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 498 + ], + "type": "text", + "content": "Generalized Category Discovery (GCD) [41] extends NCD by assuming that unlabeled data also contains known class images. Vaze et al. [41] formalize this with contrastive learning in representation space during training and clustering at inference. To reduce clustering cost, Wen et al. [45] introduce SimGCD, using a parametric classifier to learn class representation centers, serving as a strong baseline. Cao et al. [3] note SimGCD's issue of forgetting known classes while learning novel ones, proposing entropy regularization for potential known samples. Wang et al. [44] further extend SimGCD by optimizing model and data parameters for better alignment with the GCD task. Despite their performance, these methods rely on limited labeled images, constraining novel class discovery." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": "Contrastive learning (CL) spans unsupervised [39] and supervised [17] variants for learning from unlabeled and labeled data, respectively. Unsupervised CL treats two augmented views of the same image as positives and all others as negatives, while supervised CL (S-CL) considers samples from the same class as positives. Vaze et al. [41] first combine both in GCD to enhance image representations. Choi et al. [5] employ mean-shifted embeddings for CL, and Pu et al. [28] propose DCCL, which explores conceptual-level CL to model relations between known and novel categories. However, these methods primarily apply S-CL within non-parametric GCD frameworks. In contrast, recent parametric approaches [3, 44, 45] demonstrate that S-CL can effectively train classifiers by learning class prototypes. Nonetheless, these works still rely heavily on limited labeled data, underutilizing the rich supervision signal from abundant unlabeled samples, which ultimately constrains the full potential of contrastive learning." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3294" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 209, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 209, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 209, + 85 + ], + "type": "text", + "content": "3. Preliminaries and Analysis" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 198, + 104 + ], + "type": "text", + "content": "3.1. GCD setting and training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": "Problem Formulation. Unlike traditional closed-set classification, where training and test classes coincide " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_l)" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ", GCD aims to identify novel classes " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_u" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " in the test data. Given labeled dataset " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l = \\{(\\pmb{x}_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_l" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " and unlabeled dataset " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{(x_i,y_i)\\} \\subset \\mathcal{X}\\times \\mathcal{Y}_u" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ", GCD seeks to classify both known and novel samples in " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " using knowledge from " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": ". The total class number " + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "inline_equation", + "content": "K = |\\mathcal{Y}_l\\cup \\mathcal{Y}_u|" + }, + { + "bbox": [ + 55, + 108, + 296, + 204 + ], + "type": "text", + "content": " is assumed known, following prior work [3, 9, 13, 44, 45]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "text", + "content": "Architecture. SimGCD [45] consists of a Vision Transformer (ViT) encoder [7], denoted as " + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "text", + "content": ", along with a classification head " + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "text", + "content": " to generate pseudo-labels for self-distillation [2, 4], and a multi-layer perception (MLP) projection head " + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 204, + 296, + 263 + ], + "type": "text", + "content": " for representation contrastive learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": "Representation Learning. To learn representations for distinguishing samples, Vaze et al. [41] apply unsupervised contrastive learning across all data and supervised contrastive learning (S-CL) on labeled data. Given two views, " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_i'" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ", of an identical image as a positive pair in a minibatch " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ", the unsupervised contrastive loss encourages their representations to be close in an unsupervised manner:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 356, + 295, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 356, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 64, + 356, + 295, + 385 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e p}} ^ {u} = \\frac {1}{| B |} \\sum_ {i \\in B} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {i} ^ {\\prime} / \\tau_ {u}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {u}\\right)}, \\tag {1}", + "image_path": "a77343f9e23ab0ecd128b0b855315daec66d3984b640ba0c5ddeac89bdc17286.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_i = g \\circ \\Phi(\\boldsymbol{x}_i)" + }, + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\tau_u" + }, + { + "bbox": [ + 55, + 395, + 296, + 431 + ], + "type": "text", + "content": " is a temperature. Meanwhile, the supervised contrastive loss minimizes the distance between same-class samples in feature space, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 440, + 295, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 440, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 58, + 440, + 295, + 481 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e p}} ^ {s} = \\frac {1}{| B ^ {l} |} \\sum_ {i \\in B ^ {l}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {q \\in N _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {q} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n \\in B , n \\neq i} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {2}", + "image_path": "346de545b1ac319825fa4eb4011a1a417121a6e35fdc3c5ff6e6021061f97543.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "B^l" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": " is the labeled subset of " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": " is the set of images with the same label as " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": ". The overall contrastive learning loss for representation is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 527, + 295, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 527, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 116, + 527, + 295, + 542 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e p}} = (1 - \\lambda) \\mathcal {L} _ {\\text {r e p}} ^ {u} + \\lambda \\mathcal {L} _ {\\text {r e p}} ^ {s}, \\tag {3}", + "image_path": "091281e34b79e189f2d9d2951cfc9bb44b64c33d66e6d55418e86396105d5bfd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 550, + 295, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 550, + 295, + 573 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 550, + 295, + 573 + ], + "type": "text", + "content": " controls the balance between unsupervised and supervised contrastive learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": "Parametric Clustering. Instead of the time-consuming " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": "-means, Wen et al. [45] propose an efficient parametric classifier " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": " using a set of prototypes " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_1, \\dots, c_K\\}" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": " is the total classes number. For an image view " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": ", the prediction is made by calculating the cosine similarity between its feature, " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\Phi(x_i)" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": ", and each prototype " + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "inline_equation", + "content": "c_k" + }, + { + "bbox": [ + 55, + 574, + 296, + 646 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 91, + 654, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 654, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 91, + 654, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {i} ^ {(k)} = \\frac {\\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k}) / \\tau_ {s})}{\\sum_ {k ^ {\\prime}} \\exp (\\cos (\\Phi (\\boldsymbol {x} _ {i}) , \\boldsymbol {c} _ {k ^ {\\prime}}) / \\tau_ {s})}. \\tag {4}", + "image_path": "18635ed259a7ba743c0f9ad87bb1e514cdc59cd8dd3e7f46f005fe62ed2d44a7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Similarly, the soft pseudo-labels " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{q}_i'" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " for another view " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i'" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " are obtained using a smaller " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_t" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", and the cross-entropy loss is" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 72, + 553, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 85 + ], + "type": "text", + "content": "applied for prototype learning using these pseudo-labels " + }, + { + "bbox": [ + 314, + 72, + 553, + 85 + ], + "type": "inline_equation", + "content": "q_{i}^{\\prime}" + }, + { + "bbox": [ + 314, + 72, + 553, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 93, + 553, + 122 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 93, + 553, + 122 + ], + "spans": [ + { + "bbox": [ + 361, + 93, + 553, + 122 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} = - \\frac {1}{| B |} \\sum_ {i \\in B} \\sum_ {k} \\boldsymbol {q} ^ {\\prime (k)} \\log \\boldsymbol {p} ^ {(k)}. \\tag {5}", + "image_path": "eddb49fe9c75ca59ac239cbadb61392ee392b22c35f91009026563bc1271d13b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "text", + "content": "Additionally, SimGCD employs a class mean entropy regularization [2], " + }, + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "inline_equation", + "content": "H(\\overline{p}) = -\\sum_{k} \\overline{p}^{(k)} \\log \\overline{p}^{(k)}" + }, + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\overline{p}" + }, + { + "bbox": [ + 313, + 131, + 555, + 178 + ], + "type": "text", + "content": " is the batch-wise class-average prediction, forcing the model to focus on novel classes. The total classifier is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 189, + 553, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 189, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 356, + 189, + 553, + 202 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c l s}} = \\lambda \\mathcal {L} _ {\\mathrm {c l s}} ^ {s} + (1 - \\lambda) \\left(\\mathcal {L} _ {\\mathrm {d i s}} ^ {u} - \\varepsilon H\\right), \\tag {6}", + "image_path": "9ea49f9cc239e0274c9fe80ac11a68e1f9fedec79b5bf1b94a6f364a01da4b6e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 212, + 526, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 526, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 526, + 224 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 212, + 526, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}^s" + }, + { + "bbox": [ + 313, + 212, + 526, + 224 + ], + "type": "text", + "content": " is the cross-entropy loss for labeled data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 232, + 471, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 232, + 471, + 245 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 471, + 245 + ], + "type": "text", + "content": "3.2. Limited contrastive learning" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 249, + 554, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 554, + 274 + ], + "type": "text", + "content": "Analysis. Although contrastive learning (CL) enhances representation discriminability of known classes in" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 274, + 443, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 443, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 443, + 440 + ], + "type": "text", + "content": "SimGCD, it applies only to labeled data, as shown in the upper-left corner of Fig. 2. In the upper-left corner, same-colored entries along the vertical and horizontal indices correspond to samples from the same class, while different colors denote distinct classes. Self-contrast cases (blue-filled) on the diagonal are masked out. Obviously, abundant unlabeled data, shown with a gray background, are excluded from the" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 449, + 285, + 555, + 388 + ], + "blocks": [ + { + "bbox": [ + 449, + 285, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 449, + 285, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 449, + 285, + 555, + 388 + ], + "type": "image", + "image_path": "65ce49b823f47ba9ad2363b27c39efa9647633ddf83df8bcd55202838d8cb719.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 390, + 554, + 434 + ], + "lines": [ + { + "bbox": [ + 461, + 390, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 461, + 390, + 554, + 434 + ], + "type": "text", + "content": "Figure 2. A class mask: Samples from the same class are shown in the same color in a batch " + }, + { + "bbox": [ + 461, + 390, + 554, + 434 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 461, + 390, + 554, + 434 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 441, + 554, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 477 + ], + "type": "text", + "content": "CL due to the absence of class labels. As a result, the model lacks direct contrastive supervision on novel samples, making it harder to separate them in feature space." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 487, + 370, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 370, + 500 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 370, + 500 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 592 + ], + "type": "text", + "content": "To overcome the limited contrastive learning issue discussed in subsection 3.2, we propose a straightforward approach called AllGCD to integrate supervised contrastive learning across all unlabeled data directly. Specifically, we detail our proposed Intra-class Contrast in Labeled Data (Intra-CL) in subsection 4.1 and Inter-class Contrast in Unlabeled Data (Inter-CU) in subsection 4.2." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 600, + 509, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 509, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 509, + 612 + ], + "type": "text", + "content": "4.1. Intra-class Contrast in Labeled Data" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Intra-CL first strengthens intra-class compactness within known classes for better separation of known and novel categories in Inter-CU. To achieve this, we use baseline SimGCD [45] to identify potential known samples in unlabeled data and incorporate them into the labeled set. However, as shown by the orange lines in Fig. 4a and Fig. 4b, SimGCD's known-class accuracy declines sharply during training, compromising the quality of potential known" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3295" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 311 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 311 + ], + "type": "image", + "image_path": "27fa5cf4ef6b67909731898e0504ed3b87a3d8293e4ce126ac0952b7f948f2d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed AllGCD. It is composed of Intra-CL and Inter-CU. First, Intra-CL (Sec. 4.1) enhances knowledge of known classes by incrementally collecting reliable pseudo-labels " + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^*" + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "text", + "content": " within unlabeled data. Once this knowledge is strengthened, Inter-CU (Sec. 4.2) further promotes the separation of known and novel classes through contrastive learning between them." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "text", + "content": "samples. To overcome this, we propose a voting strategy that utilizes the model's prior knowledge to identify potential known samples early. Specifically, Intra-CL comprises three steps: (1) Selecting potential known samples within unlabeled data as in [3]; (2) Storing these known samples into a dictionary " + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "text", + "content": "; (3) Voting pseudo-labels " + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^*" + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 54, + 375, + 294, + 459 + ], + "type": "text", + "content": " based on their most frequent label across training iterations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 481 + ], + "type": "text", + "content": "Step (1): Selecting potential known samples within unlabeled data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "text", + "content": "In this step, we identify potential known samples by selecting instances where both views of the same image exhibit high confidence " + }, + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "text", + "content": " and are classified as known categories " + }, + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "inline_equation", + "content": "(\\mathcal{V}_l)" + }, + { + "bbox": [ + 55, + 483, + 295, + 530 + ], + "type": "text", + "content": ". The specific calculation process is as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "B^u" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": " denote the unlabeled subset of " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": ", where each image in this batch has a unique index " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_B = \\{u_1, u_2, \\ldots, u_b\\}" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": ". Given the predictions " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "q_i'" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": " from two views of an image in " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "B^u" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": ", as defined in Eq. (4). We first identify high-confidence samples by selecting those with predictions " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "q_i'" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": " both exceed a confidence threshold " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\delta} = \\mathbb{1}(\\max(p_i) > \\delta \\land \\max(q_i') > \\delta)" + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": "Next, we classify a sample as known only if the predicted label from the two views match and both belong to known class set " + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_l" + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{known}} = \\mathbb{1}((\\hat{y}_i = \\hat{y}_i') \\wedge (\\hat{y}_i \\in \\mathcal{Y}_l))" + }, + { + "bbox": [ + 55, + 614, + 295, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 673 + ], + "type": "text", + "content": "Finally, the selected known samples within the unlabeled subset " + }, + { + "bbox": [ + 55, + 651, + 295, + 673 + ], + "type": "inline_equation", + "content": "B^{u}" + }, + { + "bbox": [ + 55, + 651, + 295, + 673 + ], + "type": "text", + "content": " are determined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 679, + 295, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 679, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 111, + 679, + 295, + 694 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} = \\left(\\mathbf {M} _ {\\delta} \\odot \\mathbf {M} _ {\\text {k n o w n}}\\right) \\odot \\hat {\\mathcal {Y}}, \\tag {7}", + "image_path": "aea04ea9b0b85d7a1b5531ebcb83233de5599a60f0b50f66d6c48767e11521d1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\delta}" + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{known}}" + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": " are binary (0-1) vectors of length b," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "content": " is element-wise multiplication, and " + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}} = \\{\\hat{y}_i \\mid i \\in [1,b]\\}" + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "content": " represents the predicted labels across the batch of unlabeled samples. Thus, " + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}_{\\mathrm{select}}" + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "content": " identifies the known samples " + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}_{\\mathrm{known}} = 1)" + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "content": " that also have high confidence " + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "inline_equation", + "content": "(\\mathbf{M}_{\\delta} = 1)" + }, + { + "bbox": [ + 313, + 374, + 553, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 423, + 540, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 423, + 540, + 435 + ], + "spans": [ + { + "bbox": [ + 325, + 423, + 540, + 435 + ], + "type": "text", + "content": "Step (2): Storing above predicted known samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": "To leverage the model's early ability to recognize known classes, we initialize a dictionary " + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": " to store the pseudolabels " + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}_{\\mathrm{select}}" + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": " of each image based on its unique index " + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_B" + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "content": "Concretely, for each unlabeled subset " + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "inline_equation", + "content": "B^u" + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "content": ", we update the dictionary " + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}:\\mathcal{U}_B\\to \\hat{\\mathcal{V}}_{\\mathrm{select}}" + }, + { + "bbox": [ + 313, + 471, + 553, + 495 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 503, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 503, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 319, + 503, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} \\left(u _ {i}\\right) = \\left\\{ \\begin{array}{l l} \\mathcal {Q} \\left(u _ {i}\\right) \\cup \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\in \\operatorname {d o m} (\\mathcal {Q}), \\\\ \\left\\{\\hat {y} _ {i} \\in \\hat {\\mathcal {Y}} _ {\\text {s e l e c t}} \\right\\} & \\text {i f} u _ {i} \\notin \\operatorname {d o m} (\\mathcal {Q}). \\end{array} \\right. \\tag {8}", + "image_path": "b59eb2aa5c90aa77bd93ee2672914ce3b439ec5fd98155b1547e3dd5b5e1f05f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 567 + ], + "type": "text", + "content": "Step (3): Voting on pseudo-labels for unlabeled data based on frequency." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": "To obtain a reliable pseudo-label for each unlabeled image, we apply a voting mechanism over the stored pseudolabels " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": " across training batches. Specifically, for a batch of unlabeled samples, we select the most frequently assigned pseudo-label from its historical records in " + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(u_i)" + }, + { + "bbox": [ + 313, + 567, + 554, + 628 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 636, + 553, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 636, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 339, + 636, + 553, + 657 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} ^ {*} = \\underset {y} {\\arg \\max } \\left(\\operatorname {f r e q} (\\hat {y} \\in \\mathcal {Q} (u _ {i})), \\forall u _ {i} \\in \\mathcal {U} _ {B}\\right), \\tag {9}", + "image_path": "7792f0bafd63f6e51814d3cb40f742312c398f9af96eb4e4fe7047571eab701e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^* = \\{y_1^*, y_2^*, \\ldots, y_b^*\\}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " denotes the pseudo-labels for unlabeled samples in " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "B^u" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". The updated mini-batch labels, incorporating both original labeled data " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_l" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " and pseudolabels " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}^*" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", are given by " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\mathrm{intra}} = \\mathcal{Y}_l \\cup \\mathcal{Y}^*" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". Consequently," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3296" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 175, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 175, + 167 + ], + "type": "image", + "image_path": "a7ad01421434797eab5d06eab787b1bee2bde3f215b9297fe8f642b17b9118f1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 170, + 131, + 178 + ], + "lines": [ + { + "bbox": [ + 102, + 170, + 131, + 178 + ], + "spans": [ + { + "bbox": [ + 102, + 170, + 131, + 178 + ], + "type": "text", + "content": "(a) CUB" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 70, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 177, + 70, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 177, + 70, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 177, + 70, + 294, + 167 + ], + "type": "image", + "image_path": "7637b0bf5a8c93340ed67f55718bf7797a8125cfe8b45ab42085f6b4d443fc3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 170, + 263, + 178 + ], + "lines": [ + { + "bbox": [ + 206, + 170, + 263, + 178 + ], + "spans": [ + { + "bbox": [ + 206, + 170, + 263, + 178 + ], + "type": "text", + "content": "(b) Stanford Cars" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 189, + 296, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 243 + ], + "type": "text", + "content": "Figure 4. Known-class accuracy on unlabeled data during training. SimGCD, using S-CL without unlabeled data, shows a sharp drop on CUB and Stanford Cars. In contrast, Intra-class leverages potential known samples in unlabeled data, improving the knowledge of known classes." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 275 + ], + "type": "text", + "content": "the supervised contrastive loss in Eq. (2) is updated with pseudo known samples from unlabeled data as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 285, + 299, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 299, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 299, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t r a}} = \\frac {1}{| B ^ {\\text {i n t r a}} |} \\sum_ {i \\in B ^ {\\text {i n t r a}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {c}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {c}\\right)}, \\tag {10}", + "image_path": "ad67d5d8dee293e94e4c07f44fe20c13a2a236415c1b88039c45a90d9e8d6bb3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "B^{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " is the newly labeled subset of mini-batch " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " with labels " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " indexes same-label images as " + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 327, + 296, + 387 + ], + "type": "text", + "content": " in the batch. After applying Intra-CL, the accuracy of known classes in unlabeled data remains more stable (Fig. 4), indicating more compact representations within known classes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 407, + 258, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 258, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 258, + 418 + ], + "type": "text", + "content": "4.2. Inter-class Contrast in Unlabeled Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "text", + "content": "While Inter-CL enhances discrimination among known classes, it lacks contrastive interactions with novel classes, limiting novel category accuracy (Fig. 1a). To address this, we propose Inter-CU, which enforces global contrastive learning on the unlabeled class distribution. Specifically, once the model has stabilized on known and novel categories from epoch " + }, + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "text", + "content": ", we use the Hungarian algorithm [21] on the predictions of all unlabeled data to optimize the class distribution of known and novel categories. To ensure a high-quality distribution, we retain only high-confidence predictions " + }, + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "inline_equation", + "content": "(P > \\gamma)" + }, + { + "bbox": [ + 55, + 425, + 296, + 555 + ], + "type": "text", + "content": " for global contrastive learning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": "Formally, given the unlabeled dataset " + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u = \\{(x_i,y_i)\\} \\in \\mathcal{X}\\times \\mathcal{Y}_u" + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": " we obtain classifier predictions " + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "inline_equation", + "content": "P = [p_1,p_2,\\dots ,p_N]" + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": " via Eq. (4), where " + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": " is the total number of unlabeled samples. The minimum-cost assignment function " + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 556, + 296, + 616 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 624, + 296, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 624, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 63, + 624, + 296, + 657 + ], + "type": "interline_equation", + "content": "\\pi^ {*} = \\arg \\min _ {\\pi} \\sum_ {i = 1} ^ {N} - P _ {i, \\pi (i)}, \\quad \\pi (i) \\in \\{1, 2, \\dots , K \\}, \\tag {11}", + "image_path": "07ea1caff4b3e5811b769c7cdb7af6971b9ee962dfe3f9ad819369dc7a8c5659.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pi^{*}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " assigns each sample in " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " to one of " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " categories, minimizing the sum of costs " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "-P_{i,\\pi (i)}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". The Hungarian algorithm efficiently solves this assignment problem by finding the optimal mapping with minimal total cost. Therefore, the" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 555, + 400 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 518, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 518, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 518, + 84 + ], + "type": "text", + "content": "Algorithm 1 One Training Procedure for AllGCD" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "Input: Encoder " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", classifier head " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", projector " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", labeled data " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", unlabeled data " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", hyperparameters " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\delta, e" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ". \nOutput: " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\Phi, f" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n1: Initialize dictionary " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} \\gets \\emptyset" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n2: for epoch = 1 to total epochs do \n3: if epoch ≥ e then \n4: Get assignment " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{\\text{inter}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "D^u" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (12); \n5: end if \n6: for " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "(x, x') \\in \\mathcal{D}^l \\cup \\mathcal{D}^u" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}_l" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "y_{\\text{inter}} \\in \\mathcal{Y}_{\\text{inter}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " do \n7: " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "z \\gets g \\circ \\Phi(x)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "z' \\gets g \\circ \\Phi(x')" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n8: " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "p \\gets f \\circ \\Phi(x)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "q' \\gets f \\circ \\Phi(x')" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n9: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cls}}^s(p, y) + \\mathcal{L}_{\\text{cls}}^s(q', y)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; ▷ For " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " \n10: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{dis}}^u(p, q')" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (5); \n11: Compute Max class Entropy " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "H(p)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n12: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{cls}} \\gets \\lambda \\mathcal{L}_{\\text{cls}}^s + (1 - \\lambda) (\\mathcal{L}_{\\text{dis}}^u - \\varepsilon H)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n13: Step 1: Generate " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\hat{y} \\gets (p, q', \\delta)" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; ▷ For " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " \n14: Step 2: Store " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (8); \n15: Step 3: Vote for " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "y^*" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (9); \n16: Combine with pseudo-labels " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "y_{\\text{intra}} \\gets y \\cup y^*" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n17: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{rep}}^u(z, z')" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (1); \n18: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{intra}}(z, z', y_{\\text{intra}})" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (10); \n19: Compute " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{inter}}(z, z', y_{\\text{inter}})" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " using Eq. (12); \n20: " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{all}} = (1 - \\lambda) \\mathcal{L}_{\\text{rep}}^u + \\lambda (\\mathcal{L}_{\\text{intra}} + \\mathcal{L}_{\\text{inter}}) + \\mathcal{L}_{\\text{cls}}" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": "; \n21: Update " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "\\Phi, f" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 316, + 87, + 555, + 400 + ], + "type": "text", + "content": " by SGD [29]; \n22: end for \n23: end for" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "bbox": [ + 314, + 422, + 479, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 479, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 479, + 434 + ], + "type": "text", + "content": "final assignments for unlabeled data are:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 445, + 555, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 445, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 325, + 445, + 555, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {Y} _ {\\text {i n t e r}} = \\left\\{\\hat {y} _ {i} = \\pi^ {*} (i) \\mid i = 1, 2, \\dots , N \\right\\} \\odot \\mathbf {M} _ {\\text {i n t e r}}, \\tag {12}", + "image_path": "85787d4b0b3f30e8de6362baafe04839009848405ff2fe6923b1c3cf5e039cd7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "content": " is a binary mask that applies a confidence threshold " + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "content": " to filters out low-confidence predictions: " + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{inter}} = \\{i \\mid \\max(\\pmb{p}_i) > \\gamma\\}" + }, + { + "bbox": [ + 313, + 469, + 555, + 518 + ], + "type": "text", + "content": ". Only high-confidence samples contribute to the global contrastive loss:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 528, + 563, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 563, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 563, + 570 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t e r}} = \\frac {1}{| B ^ {\\text {i n t e r}} |} \\sum_ {i \\in B ^ {\\text {i n t e r}}} \\frac {1}{| \\mathcal {N} _ {i} |} \\sum_ {j \\in \\mathcal {N} _ {i}} - \\log \\frac {\\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {j} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}{\\sum_ {n} \\mathbb {1} _ {[ n \\neq i ]} \\exp \\left(\\boldsymbol {z} _ {i} \\cdot \\boldsymbol {z} _ {n} ^ {\\prime} / \\tau_ {\\mathrm {c}}\\right)}, \\tag {13}", + "image_path": "1882eb4032098b39454157ffe051664ecb6b1803368be152d68682c58805d40d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "inline_equation", + "content": "B^{\\mathrm{inter}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "content": " corresponds to the refined labeled subset with pseudo-labels " + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\mathrm{inter}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 618 + ], + "type": "text", + "content": ". This approach enables global contrastive learning on both known and novel samples, enhancing category discovery, particularly for novel classes." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 619, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 665 + ], + "type": "text", + "content": "Overall. Utilizing all unlabeled data in supervised contrastive learning significantly boosts class discovery, especially for novel classes. This overall training loss is updated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 359, + 669, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 669, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 359, + 669, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l l}} = (1 - \\lambda) \\mathcal {L} _ {\\text {i n t r a}} + \\lambda \\mathcal {L} _ {\\text {i n t e r}} + \\mathcal {L} _ {\\text {c l s}}, \\tag {14}", + "image_path": "19867779e7540f6c76bf50bc6953f8bd3d2713dab41923d89d556d63e315e1c0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " controls the trade-off between Intra-CL and Inter-CU. Algorithm 1 details one training step of AllGCD." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3297" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 101, + 555, + 325 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Comparison with State-of-the-Art GCD Methods, Evaluated with and without Ground-Truth " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " for Clustering. Bold denotes the best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 101, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 58, + 101, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 58, + 101, + 555, + 325 + ], + "type": "table", + "html": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
AllOldNewAllOldNewAllOldNewAllOldNewAllOldNewAllOldNew
(a) Clustering with the ground-truth number of classes K given
k-means[24]52.052.250.872.775.571.334.338.932.112.810.613.816.014.416.813.012.213.4
RS+[13]58.277.619.337.161.624.833.351.624.228.361.812.126.936.422.227.955.812.8
UNO+[9]69.580.647.270.395.057.935.149.028.135.570.518.640.356.432.228.353.714.7
ORCA[22]69.077.452.073.592.663.935.345.630.223.550.110.722.031.817.120.930.915.5
GCD [41]73.076.266.574.189.866.351.356.648.739.057.629.945.041.146.935.451.027.0
DCCL [28]75.376.870.280.590.576.263.560.864.9---43.155.736.2---
PromptCAL [50]81.284.275.383.192.778.362.964.462.150.270.140.652.252.252.337.052.028.9
SimGCD [45]80.181.277.883.093.177.960.365.657.753.871.945.054.259.151.844.058.036.4
InfoSieve [30]78.382.270.580.593.873.869.477.965.255.774.846.456.363.752.541.055.433.2
SPTNet [44]81.384.375.685.493.281.465.868.865.159.079.249.359.361.858.143.458.735.2
LegoGCD [3]81.881.482.586.394.582.163.871.959.857.375.748.455.061.551.745.157.438.5
CMS [5]82.385.775.584.795.679.268.276.564.056.063.452.356.976.147.636.454.926.4
AllGCD (ours)82.382.382.286.594.782.368.475.165.160.577.052.557.462.454.945.258.338.0
(b) Clustering without the ground-truth number of classes K given
GCD [41]70.877.657.077.991.171.351.156.448.439.158.629.7---37.251.729.4
GPC [51]75.484.660.175.393.466.752.055.547.538.258.927.443.340.744.836.551.727.9
SimGCD [45]80.181.277.881.791.276.861.566.459.149.165.141.3------
CMS [5]79.683.272.381.395.674.264.468.262.451.768.943.455.260.652.437.456.527.1
AllGCD (ours)81.583.378.484.494.380.266.772.164.058.875.750.656.764.253.045.758.738.5
", + "image_path": "0160e97e1c46783e2345ebc2ea312fdf32f680d51c44808f96814f8525a0dd66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 137, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 137, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 137, + 358 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 366, + 171, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 171, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 171, + 379 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "text", + "content": "Datasets. We evaluate our approach on three fine-grained datasets: CUB [43], Stanford Cars [18], and FGVC-Aircraft [25], as well as two generic image recognition datasets, CIFAR100 [19] and ImageNet-100 [6]. Additionally, we test on a more challenging long-tailed fine-grained Herbarium 19 [37]. Following [41], we randomly sample " + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "text", + "content": " of seen categories as labeled data " + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^l" + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "text", + "content": ", with the remaining images from seen and novel categories forming the unlabeled set " + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^u" + }, + { + "bbox": [ + 55, + 384, + 296, + 492 + ], + "type": "text", + "content": ". Details are in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "content": "Evaluation protocol. We use clustering accuracy (ACC) to evaluate model performance, following [41, 45]. During evaluation, we compare the ground-truth labels " + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "content": " with the predicted labels " + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "content": ", and calculate ACC as " + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "inline_equation", + "content": "ACC = \\frac{1}{|\\mathcal{D}^u|}\\sum_{i=1}^{|\\mathcal{D}^u|}\\mathbb{1}(y_i = p(\\hat{y}_i))" + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 494, + 296, + 568 + ], + "type": "text", + "content": " is determined using the Hungarian optimal assignment algorithm [21]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": "Implementation details. AllGCD is built upon SimGCD [45], using a ViT-B/16 backbone [7] pretrained with DINO [4]. We use the class token to train a classifier and fine-tune only the last attention block. We use a batch size of 128 and an initial learning rate of 0.1, decayed over 200 epochs using a cosine schedule. Following [41], we set " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 0.35" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{u} = 0.07" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{c} = 1.0" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": " for representation learning, and use " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{s} = 0.1" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tau_{t}" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": " decayed from 0.07 to 0.04 over 30 epochs for classifier training. Pseudo-label thresholds " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": " range from 0.6 to 0.9. Based on the convergence time of 'New' classes in SimGCD (Fig. 7a), the Inter-CU start epoch " + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 54, + 570, + 296, + 714 + ], + "type": "text", + "content": " is selected" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "content": "from " + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "inline_equation", + "content": "\\{0,20,40,60,80\\}" + }, + { + "bbox": [ + 313, + 346, + 555, + 370 + ], + "type": "text", + "content": ". All experiments are conducted using PyTorch on NVIDIA Tesla V100 GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 381, + 518, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 518, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 518, + 393 + ], + "type": "text", + "content": "5.2. Comparison With the State of the Arts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "text", + "content": "To demonstrate the effectiveness of AllGCD, we evaluate it under both known and unknown total class numbers " + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "text", + "content": ": (1) In Table 1(a), we assume " + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "text", + "content": " is known, consistent with prior works [3, 9, 13, 44, 45]; (2) In Table 1(b), we report results where " + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 399, + 555, + 472 + ], + "type": "text", + "content": " is unknown, estimating it using clustering techniques as in [5, 41, 51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": "Evaluation with ground-truth class number " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": ". Table 1(a) compares state-of-the-art methods assuming a known class number " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": ". Our method significantly enhances 'New' category accuracy compared to the baseline SimGCD, with notable improvements of " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": " on CUB, Stanford Cars, and ImageNet-100, respectively. Furthermore, AllGCD remains competitive with SPTNet, surpassing it by " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": " on the 'All' classes in CUB and Stanford Cars, respectively. Importantly, our method attains comparable performance to SPTNet using only 200 training epochs, compared to SPTNet's 1000 epochs, showing superior computational efficiency. Though AllGCD trails CMS [5] by " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": " in 'Old' accuracy on CIFAR100, it outperforms CMS by " + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 653 + ], + "type": "text", + "content": " in 'New' accuracy, highlighting its strength in novel category discovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "Evaluation with estimated number of clusters. Table 1 (b) presents a comparison conducted without access to the ground-truth class number " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ", instead using an estimated " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " derived from an off-the-shelf method [41], as shown in Table 2. Table 2 shows our approach provides a more accu" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3298" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 81, + 387, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 81, + 387, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 81, + 387, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 81, + 387, + 159 + ], + "type": "table", + "html": "
MethodCIFAR100ImageNet-100CUBStanford CarsFGVC-AircraftHerbarium 19
KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓KErr(%) ↓
GT100-100-200-196-100-683-
GCD [41]1000109923115.523017.3--52023.8
DCCL [28]146461292917291920.02----
SimGCD [45]1000109923115.523017.3--52023.8
GPC [51]1000103321262010.03----
CMS [5]955116161701515620.490106228.9
AllGCD (ours)95510882136.52107.110996366.9
", + "image_path": "4b3355ea115230c676dcf5a3324abd862d7c332c409c33988b8ce29d0f84ba30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 468, + 145 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 468, + 145 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 468, + 145 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 468, + 145 + ], + "type": "image", + "image_path": "c5b23e4e9c3d533c2abd4f0032b5260a0af76f97a98d9a3dbc8ec634a8b82e50.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 148, + 456, + 158 + ], + "lines": [ + { + "bbox": [ + 399, + 148, + 456, + 158 + ], + "spans": [ + { + "bbox": [ + 399, + 148, + 456, + 158 + ], + "type": "text", + "content": "(c) SimGCD [45]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 473, + 70, + 549, + 146 + ], + "blocks": [ + { + "bbox": [ + 473, + 70, + 549, + 146 + ], + "lines": [ + { + "bbox": [ + 473, + 70, + 549, + 146 + ], + "spans": [ + { + "bbox": [ + 473, + 70, + 549, + 146 + ], + "type": "image", + "image_path": "1c4744ce16b86f4a8182e54cd6fef3ff65ca8a488a6c9258548ec8f4d7a00d19.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 148, + 531, + 157 + ], + "lines": [ + { + "bbox": [ + 492, + 148, + 531, + 157 + ], + "spans": [ + { + "bbox": [ + 492, + 148, + 531, + 157 + ], + "type": "text", + "content": "(d) AllGCD" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 201, + 238, + 282 + ], + "blocks": [ + { + "bbox": [ + 67, + 167, + 373, + 178 + ], + "lines": [ + { + "bbox": [ + 67, + 167, + 373, + 178 + ], + "spans": [ + { + "bbox": [ + 67, + 167, + 373, + 178 + ], + "type": "text", + "content": "Table 2. Estimated class number and error rate for " + }, + { + "bbox": [ + 67, + 167, + 373, + 178 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 67, + 167, + 373, + 178 + ], + "type": "text", + "content": " using off-the-self method [41]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 201, + 238, + 282 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 238, + 282 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 238, + 282 + ], + "type": "table", + "html": "
δCUBCIFAR100
AllOldNewAllOldNew
0.7068.274.664.980.881.380.5
0.7568.475.165.181.481.881.3
0.8068.275.764.582.283.181.8
0.8567.274.763.482.382.382.2
0.9066.273.362.581.681.981.5
", + "image_path": "c902b2c84bbdf34a1915d2414bafe091fedade6ca1d147a92a6bbfc492a9f360.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 291, + 225, + 301 + ], + "lines": [ + { + "bbox": [ + 67, + 291, + 225, + 301 + ], + "spans": [ + { + "bbox": [ + 67, + 291, + 225, + 301 + ], + "type": "text", + "content": "Table 3. Effect of confidence " + }, + { + "bbox": [ + 67, + 291, + 225, + 301 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 67, + 291, + 225, + 301 + ], + "type": "text", + "content": " in Intra-CL." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 242, + 203, + 394, + 277 + ], + "blocks": [ + { + "bbox": [ + 242, + 203, + 394, + 277 + ], + "lines": [ + { + "bbox": [ + 242, + 203, + 394, + 277 + ], + "spans": [ + { + "bbox": [ + 242, + 203, + 394, + 277 + ], + "type": "image", + "image_path": "075ea49b201b9b00582a424f7d97d9b0d3c73b0393dae54aef3b3d3ec219cee0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 279, + 279, + 356, + 288 + ], + "lines": [ + { + "bbox": [ + 279, + 279, + 356, + 288 + ], + "spans": [ + { + "bbox": [ + 279, + 279, + 356, + 288 + ], + "type": "text", + "content": "(a) CUB with " + }, + { + "bbox": [ + 279, + 279, + 356, + 288 + ], + "type": "inline_equation", + "content": "\\delta = 0.75" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 291, + 482, + 301 + ], + "lines": [ + { + "bbox": [ + 311, + 291, + 482, + 301 + ], + "spans": [ + { + "bbox": [ + 311, + 291, + 482, + 301 + ], + "type": "text", + "content": "Figure 6. Logit distributions of unlabeled data." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 397, + 203, + 549, + 277 + ], + "blocks": [ + { + "bbox": [ + 384, + 167, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 384, + 167, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 384, + 167, + 553, + 190 + ], + "type": "text", + "content": "Figure 5. T-SNE visualization of 10 random novel classes from CIFAR100." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 203, + 549, + 277 + ], + "lines": [ + { + "bbox": [ + 397, + 203, + 549, + 277 + ], + "spans": [ + { + "bbox": [ + 397, + 203, + 549, + 277 + ], + "type": "image", + "image_path": "4f3ebc4713b108c3dcc7eabc7183de9ca176223cb488b99bf5eff325d04aae7b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 279, + 521, + 288 + ], + "lines": [ + { + "bbox": [ + 425, + 279, + 521, + 288 + ], + "spans": [ + { + "bbox": [ + 425, + 279, + 521, + 288 + ], + "type": "text", + "content": "(b) CIFAR100 with " + }, + { + "bbox": [ + 425, + 279, + 521, + 288 + ], + "type": "inline_equation", + "content": "\\delta = 0.85" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 323, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 296, + 455 + ], + "type": "text", + "content": "rate estimation of " + }, + { + "bbox": [ + 54, + 323, + 296, + 455 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 323, + 296, + 455 + ], + "type": "text", + "content": ", with fewer errors in determining cluster centers than SimGCD and CMS, except for CIFAR100 in SimGCD. Under this improved estimation, our approach consistently outperforms state-of-the-art methods across all datasets. Notably, performance gains are particularly significant on the Stanford Cars, ImageNet-100, and Herbarium-19 benchmarks, where AllGCD achieves 7.2%, 6.0%, and 9.1% higher accuracy in the 'New' category compared to the second-best results. These findings highlight the effectiveness of our method in utilizing all available unlabeled data to learn more discriminative representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 601 + ], + "type": "text", + "content": "Feature Visualizations and Attention Distributions. To demonstrate the effectiveness of our approach in representation learning, we visualize 10 randomly selected novel classes from CIFAR100. As shown in Fig. 5, AllGCD produces more compact and well-separated clusters, whereas SimGCD exhibits considerable cluster overlap. Additionally, we visualize attention on both known and novel samples from unlabeled data in Table 4. Ours focuses on class-specific regions (e.g., the bird's wing) and suppressing background noise (e.g., the window in the second row), highlighting that leveraging unlabeled data enhances attention to relevant regions and improves class discriminability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 610, + 149, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 149, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 149, + 623 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 628, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 688 + ], + "type": "text", + "content": "In this section, we present ablation studies to assess the effectiveness of Intra-CL and Inter-CU in our AllGCD model. The evaluation includes fine-grained datasets such as CUB and Stanford Cars, alongside generic image recognition datasets CIFAR100 and ImageNet-100." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Effect of different confidence " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". Table 3 shows the effect of varying the confidence threshold " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " in Intra-CL on" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 316, + 375, + 555, + 497 + ], + "blocks": [ + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "type": "text", + "content": "Table 4. Average attention across 12 heads in the final layer of the image encoder on CUB and FGVC-Aircraft. The known sample (top row) and novel sample (bottom row) in AllGCD show improved focus on object regions with reduced background noise." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 375, + 555, + 497 + ], + "lines": [ + { + "bbox": [ + 316, + 375, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 555, + 497 + ], + "type": "table", + "html": "
ImageDINOSimGCDAllGCD
", + "image_path": "54ca2209fb4ecf5c0b695874eb21c5f039e1dd879cf2836559d3e3e377227e71.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": "CUB and CIFAR100. Overall, the accuracy for both 'Old' and 'New' classes initially improves with increasing " + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": " but slightly drops beyond the optimal point. The fine-grained CUB dataset performs best at " + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\delta = 0.75" + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": ", while the generic dataset CIFAR100 peaks at " + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\delta = 0.85" + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": ". This difference arises from the varying proportions of unlabeled samples preserved in fine-grained and generic datasets. Higher thresholds, such as 0.85, help retain sufficient known samples in CIFAR100 (see Fig. 5b), but may exclude too many in CUB (see Fig. 5a), resulting in suboptimal performance for the fine-grained dataset. Therefore, we set " + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\delta = 0.75" + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": " for fine-grained datasets and " + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\delta = 0.85" + }, + { + "bbox": [ + 313, + 521, + 555, + 664 + ], + "type": "text", + "content": " for generic datasets." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": "Effect of different confidence " + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": ". Similar to " + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": " in Intra-CL for selecting high-confidence known samples, " + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": " filters high-quality distributions for inter-class contrastive learning in Inter-CU. Table 5 shows its effect on Stanford Cars and" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3299" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 91, + 282, + 188 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 288, + 82 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 288, + 82 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 288, + 82 + ], + "type": "text", + "content": "Table 5. Ablation study on different thresholds " + }, + { + "bbox": [ + 63, + 70, + 288, + 82 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 63, + 70, + 288, + 82 + ], + "type": "text", + "content": " in Inter-CU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 91, + 282, + 188 + ], + "lines": [ + { + "bbox": [ + 69, + 91, + 282, + 188 + ], + "spans": [ + { + "bbox": [ + 69, + 91, + 282, + 188 + ], + "type": "table", + "html": "
γStanford CarsImageNet-100
AllOldNewAllOldNew
0.6060.076.852.285.693.881.5
0.6560.577.052.586.094.181.7
0.7059.876.752.186.594.782.3
0.7558.975.151.286.394.982.1
0.8057.673.750.186.194.481.9
", + "image_path": "8bb4d79cd5fe3adea853372932fc75ec9ab24333ef4354415df4f9d1a9ccdb58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 240, + 295, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 232 + ], + "type": "text", + "content": "Table 6. Effectiveness of each component of our method. Intra-CL and Inter-CU are described in Sec. 4.1 and 4.2, respectively. Row (1) shows the baseline SimGCD results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 240, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 240, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 295, + 297 + ], + "type": "table", + "html": "
IndexSimGCDIntra-CLInter-CUCUBCIFAR100
AllOldNewAllOldNew
(1)✓60.365.657.780.181.277.8
(2)✓✓65.273.161.381.281.879.6
(3)✓✓67.172.863.881.681.581.7
(4)✓✓✓68.475.165.182.382.382.2
", + "image_path": "e619490587e27e7bc1f7a1da542599a6480115addc413cf15fd64272cdf93efb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": "ImageNet-100. Specifically, we fix " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\delta = 0.75" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " for Stanford Cars and " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\delta = 0.85" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " for ImageNet-100 based on the " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " ablation. Intuitively, Stanford Cars performs best with " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\gamma = 0.65" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": ", while ImageNet-100 performs best with " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\gamma = 0.7" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": ". Similar to Intra-CL, this difference is mainly due to the varying number of selected samples at each threshold: " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "6.1\\mathrm{k}" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "95.3\\mathrm{k}" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " (see Data in appendix). ImageNet-100 retains more high-confidence samples at a stricter threshold (e.g., 0.75), while fine-grained datasets yield fewer, affecting Inter-CU's effectiveness. Therefore, we set " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\gamma = 0.65" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " or 0.6 for fine-grained datasets and " + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "inline_equation", + "content": "\\gamma = 0.70" + }, + { + "bbox": [ + 54, + 318, + 295, + 450 + ], + "type": "text", + "content": " or 0.75 for generic ones." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "Effect of each proposed component. Table 6 presents the ablation study on our proposed Intra-CL and Inter-CU. Compared to SimGCD (Row (1)), Intra-CL (Row (2)) improves 'Old' class accuracy by " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " on CUB and " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " on CIFAR100, demonstrating its effectiveness in refining known sample selection for supervised contrastive learning. This also leads to a " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " improvement in 'New' accuracy on CUB and CIFAR100, respectively. Inter-CU (Row (3)) further enhances 'New' accuracy by " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "6.1\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " on CUB and " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " on CIFAR100 by fully utilizing unlabeled data for novel class discovery. Combining both (Row (4)) achieves the best overall performance, confirming their complementary benefits in improving recognition of both known and novel classes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Effect of starting epoch " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " in Inter-CU. Fig. 7b reports 'New' accuracy for different Inter-CU start epochs " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ". Fig. 7a shows that SimGCD struggles to separate known and novel classes early (e.g., during the first 20 epochs). This difficulty leads to an unreliable class distribution, which negatively impacts Inter-CU since it depends on well-separated inter-class relationships. To address this, we evaluate " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "e \\in \\{0, 20, 40, 60, 80\\}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "e = 0" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " denotes All-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 435, + 148 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 435, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 435, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 435, + 148 + ], + "type": "image", + "image_path": "eb54a748e68428cd78d7492e6d2ba1f0bb6c3c0ec76b49a76c81d5ada3ff9c63.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 150, + 430, + 160 + ], + "lines": [ + { + "bbox": [ + 321, + 150, + 430, + 160 + ], + "spans": [ + { + "bbox": [ + 321, + 150, + 430, + 160 + ], + "type": "text", + "content": "(a) 'New' ACC. in SimGCD [45]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 437, + 70, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 437, + 70, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 437, + 70, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 437, + 70, + 550, + 148 + ], + "type": "image", + "image_path": "d399f2d06f5bc41ae1d043ea53d33b019ee402d29042153c9cd3cbd7fe5a6220.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 150, + 538, + 160 + ], + "lines": [ + { + "bbox": [ + 450, + 150, + 538, + 160 + ], + "spans": [ + { + "bbox": [ + 450, + 150, + 538, + 160 + ], + "type": "text", + "content": "(b) 'New' ACC. @ Epochs" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "text", + "content": "Figure 7. 'New' class accuracy on unlabeled data. (a) In SimGCD, 'New' ACC. remains low in the early training stages (e.g. below " + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "text", + "content": " for FGVC-Aircraft and Stanford Cars at epoch " + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "inline_equation", + "content": "\\leq 20" + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "text", + "content": ".) (b) Inter-CU improves accuracy more effectively in the middle and late stages (" + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "inline_equation", + "content": "e.g. e \\geq 40" + }, + { + "bbox": [ + 313, + 170, + 555, + 226 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "text", + "content": "GCD without Inter-CU (only Intra-CL). Results show that Stanford Cars and FGVC-Aircraft peak at later epochs (e.g., " + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "inline_equation", + "content": "e = 80" + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "text", + "content": "), while CUB and CIFAR100 peak at " + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "inline_equation", + "content": "e = 40" + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "text", + "content": ". This aligns with CUB's higher early-stage known class recognition compared to Stanford Cars and FGVC-Aircraft (60% vs. ≈43%), enabling earlier effective separation of known and novel samples. CIFAR100 is less sensitive to " + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 247, + 555, + 355 + ], + "type": "text", + "content": ", reaching approximately 70% accuracy by epoch 40, yielding stable class distributions for Inter-CU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 368, + 388, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 388, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 388, + 380 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "content": "This paper identifies that previous supervised contrastive learning (S-CL) in parametric GCD algorithms limits novel category discovery due to its restricted use of labeled data. To address this, we propose AllGCD, a straightforward approach that leverages all unlabeled data containing novel classes into S-CL. Specifically, we introduce two strategies named Intra-class Contrast in Labeled Data (Intra-CL) and Inter-class Contrast in Unlabeled Data (Inter-CU). Intra-CL first enhances the intra-class compactness within known classes by incorporating potential known samples from unlabeled data into the original S-CL. Building on this, Inter-CU further separates known and novel classes by applying S-CL to a global class distribution. By combining these strategies, AllGCD maximizes the potential of S-CL, achieving substantial improvements in the discovery of novel categories. Extensive experiments show that AllGCD significantly outperforms baselines and achieves competitive results compared to other GCD methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 411, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 411, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 411, + 629 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by the Guangdong Science and Technology Program (Grant No. 2024B0101040005), the National Natural Science Foundation of China (Grant No. 62332002 and 61825101), the Shenzhen High-Level Talent Team Program, and the Shenzhen Science and Technology Innovation Commission (Grant No. KQTD20240729102051063)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3300" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 115, + 109 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 117, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 117, + 297, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 297, + 171 + ], + "type": "text", + "content": "[1] Wenbin An, Feng Tian, Qinghua Zheng, Wei Ding, QianYing Wang, and Ping Chen. Generalized category discovery with decoupled prototypical network. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 12527-12535, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 173, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 173, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 296, + 239 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Mike Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In Proceedings of the European conference on Computer Vision (ECCV), pages 456-473, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 296 + ], + "type": "text", + "content": "[3] Xinzi Cao, Yutong Lu, and Yonghong Tian. Solving the catastrophic forgetting problem in generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16880-16889, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 297, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 353 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of IEEE/CVF International Conference on Computer Vision (CVPR), pages 9630-9640, 2021. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "text", + "content": "[5] Sua Choi, Dahyun Kang, and Minsu Cho. Contrastive mean-shift learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23094-23104, 2024. 2, 6, 7, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 464 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 467, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 544 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In Proceedings of International Conference on Learning Representations (ICLR), 2021. 1, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14064-14073, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 296, + 656 + ], + "type": "text", + "content": "[9] Enrico Fini, Enver Sangineto, Stephane Lathuilière, Zhun Zhong, Moin Nabi, and Elisa Ricci. A unified objective for novel class discovery. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264-9272, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 658, + 295, + 712 + ], + "type": "text", + "content": "[10] Peiyan Gu, Chuyu Zhang, Ruijie Xu, and Xuming He. Classrelation knowledge distillation for novel class discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16428-16437, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Proceedings of International Conference on Learning Representations (ICLR), 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[12] Kai Han, Andrea Vedaldi, and Andrew Zisserman. Learning to discover novel visual categories via deep transfer clustering. In Proceedings of IEEE/CVF International Conference on Computer Vision (ICCV), pages 8400-8408, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 554, + 215 + ], + "type": "text", + "content": "[13] Kai Han, Sylvestre-Alvise Rebuffi, Sébastien Ehrhardt, Andrea Vedaldi, and Andrew Zisserman. Autonovel: Automatically discovering and learning novel visual categories. IEEE Trans. Pattern Anal. Mach. Intell., 44(10):6767-6781, 2022. 2, 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 260 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross B. Girshick. Mask R-CNN. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2980-2988, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 360 + ], + "type": "text", + "content": "[16] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261-2269, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "type": "text", + "content": "[17] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 415, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 469 + ], + "type": "text", + "content": "[18] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of IEEE/CVF International Conference on Computer Vision Workshops (ICCV), pages 554–561, 2013. 2, 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 471, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 554, + 492 + ], + "type": "text", + "content": "[19] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 2, 6, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 547 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), pages 1106-1114, 2012. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 580 + ], + "type": "text", + "content": "[21] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 554, + 635 + ], + "type": "text", + "content": "[22] Jiaming Liu, Yangqiming Wang, Tongze Zhang, Yulu Fan, Qinli Yang, and Junming Shao. Open-world semi-supervised novel class discovery. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJ-CAI, pages 4002–4010. ijcai.org, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 690 + ], + "type": "text", + "content": "[23] Yu Liu, Yaqi Cai, Qi Jia, Binglin Qiu, Weimin Wang, and Nan Pu. Novel class discovery for ultra-fine-grained visual categorization. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17679-17688, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[24] James MacQueen et al. Some methods for classification and analysis of multivariate observations. In Proceedings of" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3301" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "the fifth Berkeley symposium on mathematical statistics and probability, pages 281-297, 1967. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[25] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 2, 6, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 215 + ], + "type": "text", + "content": "[26] Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. In Proceedings of Conference on European Conference on Computer Vision (ECCV), pages 728-755, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 315 + ], + "type": "text", + "content": "[27] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 316, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 369 + ], + "type": "text", + "content": "[28] Nan Pu, Zhun Zhong, and Nicu Sebe. Dynamic conceptual contrastive learning for generalized category discovery. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7579-7588, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 371, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 403 + ], + "type": "text", + "content": "[29] Ning Qian. On the momentum term in gradient descent learning algorithms. Neural Networks, 12(1):145-151, 1999. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[30] Sarah Rastegar, Hazel Doughty, and Cees Snoek. Learn to categorize or categorize to learn? self-coding for generalized category discovery. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "text", + "content": "[31] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 91-99, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "text", + "content": "[32] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 591 + ], + "type": "text", + "content": "[33] Mark Sandler, Andrew G. Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 296, + 624 + ], + "type": "text", + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 647 + ], + "type": "text", + "content": "[35] Yiyou Sun and Yixuan Li. Opencon: Open-world contrastive learning. Trans. Mach. Learn. Res., 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[36] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1-9, 2015. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Kiat Chuan Tan, Yulong Liu, Barbara Ambrose, Melissa Tulig, and Serge Belongie. The herbarium challenge 2019 dataset. arXiv preprint arXiv:1906.05372, 2019. 2, 6, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "text", + "content": "[38] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. In Proceedings of the European conference on Computer Vision (ECCV), pages 776-794, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 186 + ], + "type": "text", + "content": "[39] Aäron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. CoRR, abs/1807.03748, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 189, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 189, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 189, + 553, + 243 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of Advances in Neural Information Processing Systems (NeurIPS), pages 5998-6008, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 246, + 555, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 555, + 291 + ], + "type": "text", + "content": "[41] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Generalized category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7482-7491, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "text", + "content": "[42] Sagar Vaze, Andrea Vedaldi, and Andrew Zisserman. No representation rules them all in category discovery. In Proceedings of Conference on Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[43] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 2, 6, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 373, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 428 + ], + "type": "text", + "content": "[44] Hongjun Wang, Sagar Vaze, and Kai Han. Sptnet: An efficient alternative framework for generalized category discovery with spatial prompt tuning. In Proceedings of International Conference on Learning Representations (ICLR), 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "text", + "content": "[45] Xin Wen, Bingchen Zhao, and Xiaojuan Qi. Parametric classification for generalized category discovery: A baseline study. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16590-16600, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 542 + ], + "type": "text", + "content": "[46] Jianzong Wu, Xiangtai Li, Shilin Xu, Haobo Yuan, Henghui Ding, Yibo Yang, Xia Li, Jiangning Zhang, Yunhai Tong, Xudong Jiang, Bernard Ghanem, and Dacheng Tao. Towards open vocabulary learning: A survey. IEEE Trans. Pattern Anal. Mach. Intell. (TPAMI), 46(7):5092-5113, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 609 + ], + "type": "text", + "content": "[47] Zelin Zang, Lei Shang, Senqiao Yang, Fei Wang, Baigui Sun, Xuansong Xie, and Stan Z. Li. Boosting novel category discovery over domains with soft contrastive learning and all in one classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11824-11833, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 656 + ], + "type": "text", + "content": "[48] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In Proceedings of International Conference on Learning Representations (ICLR), 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[49] Jie Zhang, Xiaosong Ma, Song Guo, and Wenchao Xu. Towards unbiased training in federated open-world semi-supervised learning. In Proceedings of International Conference on Machine Learning, (ICML), pages 41498-41509, 2023. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3302" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 251 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 139 + ], + "type": "text", + "content": "[50] Sheng Zhang, Salman H. Khan, Zhiqiang Shen, Muzammal Naseer, Guangyi Chen, and Fahad Shahbaz Khan. Prompt- cal: Contrastive affinity learning via auxiliary prompts for generalized novel category discovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3479-3488, 2023. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 195 + ], + "type": "text", + "content": "[51] Bingchen Zhao, Xin Wen, and Kai Han. Learning semi-supervised gaussian mixture models for generalized category discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16623-16633, 2023. 6, 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 251 + ], + "type": "text", + "content": "[52] Zhun Zhong, Linchao Zhu, Zhiming Luo, Shaozi Li, Yi Yang, and Nicu Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9462-9470, 2021. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3303" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_content_list.json b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..25712745e9b57131af837722bae53d02e4ef350a --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_content_list.json @@ -0,0 +1,1466 @@ +[ + { + "type": "text", + "text": "AllTracker: Efficient Dense Point Tracking at High Resolution", + "text_level": 1, + "bbox": [ + 183, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adam W. Harley $^{1}$ Yang You $^{1}$ Xinglong Sun $^{1}$ Yang Zheng $^{1}$ Nikhil Raghuraman $^{1}$ Yunqi Gu $^{1}$ Sheldon Liang $^{2}$ Wen-Hsuan Chu $^{2}$ Achal Dave $^{3}$ Suya You $^{4}$ Rares Ambrus $^{3}$ Katerina Fragkiadaki $^{2}$ Leonidas Guibas $^{1}$", + "bbox": [ + 155, + 179, + 841, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stanford University $^{2}$ Carnegie Mellon University $^{3}$ Toyota Research Institute $^{4}$ Army Research Laboratory", + "bbox": [ + 161, + 234, + 831, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce AllTracker: a model that estimates long-range point tracks by way of estimating the flow field between a query frame and every other frame of a video. Unlike existing point tracking methods, our approach delivers high-resolution and dense (all-pixel) correspondence fields, which can be visualized as flow maps. Unlike existing optical flow methods, our approach corresponds one frame to hundreds of subsequent frames, rather than just the next frame. We develop a new architecture for this task, blending techniques from existing work in optical flow and point tracking: the model performs iterative inference on low-resolution grids of correspondence estimates, propagating information spatially via 2D convolution layers, and propagating information temporally via pixel-aligned attention layers. The model is fast and parameter-efficient (16 million parameters), and delivers state-of-the-art point tracking accuracy at high resolution (i.e., tracking $768 \\times 1024$ pixels, on a 40G GPU). A benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets, and we find that doing so is crucial for top performance. We provide an extensive ablation study on our architecture details and training recipe, making it clear which details matter most. Our code and model weights are available: https://alltracker.github.io", + "bbox": [ + 88, + 316, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating the long-range trajectories of arbitrary points in the world, using sequences of 2D images as input, has been a concrete and competitive challenge in computer vision since at least 2006 [40]. The utility of optical flow (i.e., the instantaneous velocity of pixels [16]) toward this goal has long been obvious, yet it has remained challenging to upgrade flows into long-range tracks.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instantaneous flows can be \"chained\" into multi-frame tracks, by interpolating in one flow field at the endpoints of the previous flow field, but imperfect flows will accumulate drift, and such chains must also be carefully stopped at", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8645867ce1ed03f51ab9fbee66039f8841c0d6dc9e2de09081f41068717e510.jpg", + "image_caption": [ + "Figure 1. AllTracker estimates high-resolution optical flow between a \"query frame\" and every other frame of a video, using a sliding-window strategy. Point samples from these outputs can be interpreted as long-term point trajectories." + ], + "image_footnote": [], + "bbox": [ + 521, + 284, + 888, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "occlusions [40]. An attractive shortcut here is to directly compute the flow between a reference frame and each other frame, and thus track drift-free and across occlusions, but estimating \"long-range flow\" becomes increasingly difficult as the time interval widens between the reference frame and the target, due to increasing variation in perspective, illumination, and scene geometry.", + "bbox": [ + 511, + 609, + 906, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In light of these challenges, recent work has side-stepped optical flow and created a line of bespoke point trackers focused on learning multi-frame temporal priors for reducing drift and tracking through occlusions [12, 19, 24, 25, 59]. These works explicitly highlight that the limited temporal context in flow-based methods is a severe weakness, and have made substantial progress by addressing this. However, the proposed solutions add temporal awareness at the cost of spatial awareness, and only deliver tracks for sparse sets of points. Recent attempts at \"dense\" point tracking [27, 37, 50] are not as accurate as the latest sparse trackers, and struggle with high-resolution input.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5253", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we demonstrate that learnable multi-frame temporal priors can be built together with high-resolution spatial awareness, by casting point tracking as a multi-frame long-range optical flow problem, as illustrated in Figure 1.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our design combines ideas from point tracking literature and optical flow literature, arriving at a composite architecture with broader capability. Following the trend in both areas, the heart of the model is a recurrent module that iteratively improves motion estimates [19, 47]. This module relies on information from spatial cross-correlations, giving it a strong inductive bias for feature-matching, which speeds up training. From point tracking literature, we borrow a per-pixel temporal module, to learn a motion prior and track through occlusions [19]. From optical flow literature, we borrow the idea of performing the majority of processing on a low-resolution grid, to allow fast spatial message-passing via 2D convolutions, and recover spatial precision at the end of the architecture with an upsampling layer [47, 54]. In relation to other point trackers, the key novelty of our approach is that we frame the problem as long-range dense flow, which makes sparse methods (of similar speed and accuracy) redundant. In relation to other optical flow models, the key novelty of our approach is that we solve a window of flow problems simultaneously, instead of frame-by-frame; the information shared within and across windows unlocks the capability to resolve flows across wide time intervals.", + "bbox": [ + 91, + 154, + 485, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A secondary benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets. We therefore build a mix of publicly-available datasets to support our model, spanning different time-ranges, annotation densities, and resolutions, and train with uniform sampling across this mix. We find that this strategy, paired with a long training schedule, is crucial for top performance.", + "bbox": [ + 89, + 487, + 483, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our method blends techniques and datasets from optical flow literature and point tracking literature, and the resulting model is a state-of-the-art point tracker that operates in high resolution at full density. Inspired by CoTracker [25] and highlighting our capability for all-pixel tracking, we name our approach AllTracker. We will release our model, our data, and our code.", + "bbox": [ + 89, + 594, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 719, + 232, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optical flow The introduction of the concept of optic flow can be attributed to Gibson [16] (see Niehorster's review [38]). In computer vision, optical flow methods generally take two consecutive frames of video as input, and estimate the motion that relates the pixels of the first frame to the pixels of the second frame [21, 32]. The dominant technique, from classic approaches to current learning-based approaches, is to iterate on a solution: first, candidate correspondences are initialized (e.g., with zero-motion or constant velocity), then the \"costs\" of these correspondences are computed by measuring how", + "bbox": [ + 89, + 750, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "well the appearance features match, and this appearance cost is combined with smoothness terms or learned priors, and then better correspondences are sought in the neighborhood of the current solution [4, 22, 42, 47, 58]. Our work uses the same basic technique.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The current state of the art in optical flow estimation is SEA-RAFT [54]. SEA-RAFT first directly estimates a low-resolution flow field [14, 22], then computes appearance features with a ResNet-34 [20], and computes pixel-to-pixel correlations with these features [58], then iteratively estimates refinements to the flow via 2D convolutions that take flows and correlations as input, and finally upsamples the flow to full resolution via a pixel-shuffle layer [41]. Our work takes inspiration from SEA-RAFT (and its predecessor RAFT [47]) in the way that we iteratively refine low-resolution flows and upsample to full resolution, but we operate on multiple frames at a time (i.e., 16 frames instead of 2), and we add temporal attention layers to help information propagate across the time axis.", + "bbox": [ + 511, + 167, + 908, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Flow-based point tracking Optical flow is often used as a building block toward multi-frame tracking. The standard technique in this space is to \"chain\" flow vectors end-to-end to form longer tracks, and guard against drift by monitoring for occlusions [40, 45]. Recent techniques skip past occlusions by using multistep flows [7-9, 36], or attempt to track through occlusions using a learned temporal prior [5] (inspired by non-flow point trackers). These methods deliver dense multi-frame trajectories similar to our method, but first require an expensive pre-processing stage in which optical flows are estimated by a different model, while our method handles the full problem efficiently with a single model.", + "bbox": [ + 511, + 382, + 908, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We highlight two concurrent works which propose a flow-based method somewhat similar to our own: DTF [50] and DELTA [37]. Both iteratively estimate flows to relate a reference frame to other frames, but unlike our work, these papers propose special-purpose transformer-based methods in which global spatial message-passing is approximated by cross-attending to sparse \"anchor\" or \"centroid\" tokens. Our method simply uses 2D convolutions for this step, more like SEA-RAFT [54]. We note also that DTF and DELTA both struggle with high-resolution inputs (running out of memory on our 40G GPUs), despite DELTA using low-resolution inference and high-resolution upsampling, comparable to SEA-RAFT and our approach. In performance, DTF is not competitive with recent point trackers (while ours exceeds them); DELTA's main model requires depth input, but we compare against its 2D variant and outperform it.", + "bbox": [ + 511, + 565, + 908, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point trackers without flow Many classic methods track points directly, without relying on optical flow as a submodule [32, 49]. Harley et al. [19] recently introduced a new method in this space, with some components similar to flow models [47], but tracking points independently, and adding a multi-frame inference window which enabled the model", + "bbox": [ + 511, + 809, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5254", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to track through occlusions. That work, along with a new \"Tracking Any Point\" benchmark that appeared soon afterward [11], spurred much follow-up effort, involving multipoint context [24, 25], better initialization schemes [12], wider correlation context [2, 6], new transformer-based designs [29, 30], and post-hoc densification methods [27].", + "bbox": [ + 88, + 90, + 482, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The current state of the art in point tracking is CoTracker3 [24]. CoTracker3 first computes feature maps for all frames, then initializes tracks for a sparse set of query points, then retrieves local correlations based on these tracks, and interleaves modules which pass information within tracks (temporally) and across tracks (spatially). The main novelty is in the spatial propagation: each point attends to a set of latent \"virtual points\" which learn a compressed representation of the video motions. While CoTracker3 delivers very accurate tracks, it requires users to pre-select a sparse set of points to track, and results vary based on how the queries are distributed [25]. CoTracker3 is more efficient than its predecessors [19, 25], but is still limited to tracking a few thousand points at a time. Our approach is simpler and more memory-efficient: we track on a low-resolution grid, where we achieve message-passing with simple 2D convolutions, and we recover spatial precision at the end of the architecture with a fast upsampling layer [41]. We note that this efficient upsampling technique was known well before the recent resurgence of point trackers [41], but perhaps its relevance was not fully appreciated.", + "bbox": [ + 93, + 188, + 485, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training data and self-supervision Current methods in optical flow and point tracking are data-driven, making data and supervision choices crucial. In optical flow, state-of-the-art methods rely on a combination of mostly synthetic datasets, including FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. In point tracking, the main training datasets are Kubric [17], FlyingThings++ [19], and PointOdyssey [59]. There is an art to creating a curriculum from diverse datasets to optimize downstream performance on particular benchmarks [54], but in our work we simply concatenate the datasets and shuffle the samples. We note that all previous point tracking methods did not use flow data, perhaps because most models do not support 2-frame inference or dense output; we therefore also provide experiments using Kubric alone.", + "bbox": [ + 89, + 515, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the fact that the training data for point trackers is all synthetic, several groups have explored self-supervision methods based on boot-strapping pre-trained models using pseudo-labels computed on real videos [13, 24, 44]. However, these schemes offer only minor gains over the supervised weights where bootstrapping begins. In this work, we avoid any pseudo-labelling and simply add more synthetic data, capitalizing on our model's ability to accept supervision from optical flow data.", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0cbe47e5a494155d34f69be48ce6a97f09f26d1b43b72db8d812b3ca206fc0e5.jpg", + "image_caption": [ + "Figure 2. AllTracker architecture. First, we compute feature maps for all frames, and copy the zeroth (query) feature map to every timestep, and compute multi-scale cost volumes. Then, we iterate a recurrent module, which references the query feature map and cost volume pyramid at each timestep, and estimates a low-resolution correspondence field, using interleaved 2D convolutions and pixel-aligned temporal attentions. The output of the RNN is upsampled into high-resolution optical flow maps, which relate all pixels of the zeroth frame to every other frame." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. AllTracker", + "text_level": 1, + "bbox": [ + 513, + 470, + 630, + 488 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model takes a video as input, along with a \"query\" index, specifying which frame's pixels to track, and it outputs full-timespan tracking for all of the pixels in that frame.", + "bbox": [ + 511, + 500, + 906, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concretely, we are given as input a video of shape $T, H, W, 3$ , where $T$ is the number of frames, and $H, W$ indicate the spatial resolution of each frame. We are also given an index $t \\in T$ , indicating which frame has the pixels we need to track. Our final output is a tensor shaped $T, H, W, 4$ : the first two channels are optical flow maps indicating the offset that takes each pixel in the query frame to its correspondence in every other frame, and the next two channels estimate visibility and confidence.", + "bbox": [ + 511, + 550, + 908, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A brief summary of the method is as follows. We operate in sliding window fashion across $T$ , processing subsequences of length $S$ , advancing at a stride of $S / 2$ . We begin by quickly computing low-resolution feature maps for the subsequence, yielding a feature volume with dimensions $S, H / 8, W / 8, D$ , where $D$ is the channel dimension of the features. We then initialize a low-resolution tensor of outputs, shaped $S, H / 8, W / 8, 4$ , and iteratively revise this tensor using cross correlations and features across space and time as reference. Finally, we upsample the outputs, producing full resolution optical flow, visibility, and confidence, shaped $S, H, W, 4$ . We then advance our window by $S / 2$ , using previous estimates as initialization, and repeat inference. Figure 2 shows the method in a diagram.", + "bbox": [ + 511, + 689, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5255", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Encoding", + "text_level": 1, + "bbox": [ + 89, + 90, + 200, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our first stage encodes the input video frames into low-resolution feature maps. We achieve this with a ConvNeXt-Tiny [31] model (see Sec. 3.6 for implementation details). This encoder compresses a subsequence input of shape $S, H, W, 3$ to the shape $S, H/8, W/8, D$ , where $D$ is the embedding dimension (in our case $D = 256$ ), $S$ is the subsequence window length (in our case $S = 16$ ). When the full video length $T$ is not too large (e.g., at training time), we compute all of the feature maps in parallel.", + "bbox": [ + 89, + 112, + 485, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then isolate the feature map of the query frame and tile it to the length of the subsequence, so that we have a copy for each timestep, as illustrated in Figure 2.", + "bbox": [ + 89, + 248, + 483, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Computing appearance similarity", + "text_level": 1, + "bbox": [ + 89, + 303, + 385, + 319 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the feature maps, we build a multi-scale 4D correlation volume, capturing appearance-based tracking cues. We implement this in two steps: first we convert each feature map into a feature pyramid, by average-pooling at multiple strides (in our case 5 strides: $\\{1,2,4,8,16\\}$ ), and then we cross-correlate the query feature map with these pyramids. That is, for each feature vector in the query feature map, we compute its dot product across each timestep's pyramid.", + "bbox": [ + 89, + 324, + 483, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The output of this step can be understood as a large collection of heatmaps: one heatmap per tracked pixel, per scale, per timestep. The model will index into these heatmaps to retrieve information on where correspondences lie. When a target's correspondence is visible at a certain timestep, we expect the respective heatmap to have a strong peak at the location of the correspondence.", + "bbox": [ + 89, + 446, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Track initialization", + "text_level": 1, + "bbox": [ + 89, + 560, + 274, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next initialize estimates for tracking coordinates, visibility, and confidence, by creating a tensor with shape $S$ , $H/8$ , $W/8$ , 4, where the coordinate channels are initialized with a 2D meshgrid, and visibility and confidence are initialized with zeros. If we are on a window beyond the first, we populate this tensor with the estimates that overlap, and copy forward the values from the last estimated timestep.", + "bbox": [ + 89, + 582, + 485, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Iterative refinement", + "text_level": 1, + "bbox": [ + 89, + 696, + 279, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The most important stage of our model is the iterative refinement stage, where we update all of the estimates. We describe the main points of the module here, but also provide extensive details (and a diagram) in the supplementary.", + "bbox": [ + 89, + 719, + 483, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each timestep, for each pixel in the query frame, we have: a feature vector $\\mathbf{f}$ (length $D$ ), visibility and confidence estimates $\\mathbf{v}$ , $\\mathbf{c}$ (together length 2), a position estimate $\\mathbf{p}$ (length 2), and a correlation pyramid $\\{\\mathbf{C}_1, \\mathbf{C}_2, \\ldots\\}$ shaped $\\{H/8 \\times W/8, H/16 \\times W/16, \\ldots\\}$ . We convert this data into a \"local\" representation suitable for convolutions, as follows. For the position data, we replace the absolute position estimates $\\mathbf{p}$ with motion estimates, by subtracting the source", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "positions: $\\mathbf{m} = \\mathbf{p} - \\mathbf{p}_0$ . For each level of the correlation pyramid, we extract a small patch centered at $\\mathbf{p}$ , and flatten these patches into a single vector $\\mathbf{q}$ of length $L \\cdot (2R + 1)^2$ , where $R$ is a radius parameter for the patches (e.g., 4) and $L$ is the number of levels in the pyramid (e.g., 5). All together, we have $\\mathbf{f}, \\mathbf{v}, \\mathbf{c}, \\mathbf{m}, \\mathbf{q}$ (per pixel, per timestep), which in our setup makes 665 channels.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inside the recurrent module, we process the input data using interleaved spatial and temporal blocks. Each spatial block is a 2D ConvNeXt block; each temporal block is a pixel-aligned transformer block. By pixel-aligned, we mean that attention only happens along the temporal axis (i.e., with cost quadratic in $S = 16$ ), and this is done for every pixel in parallel. We note that an important convenience of our design is that all of the tensors in this stage are aligned with the \"query\" frame; therefore pixel-aligned attention is attention between corresponding pixels. After the interleaved spatial and temporal blocks propagate tracking-related information across our window, we decode explicit revisions for visibility, confidence, and motion.", + "bbox": [ + 511, + 196, + 908, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply visibility, confidence, and motion revisions via simple summation with the previous values $\\mathbf{x}_{\\mathrm{new}} = \\mathbf{x}_{\\mathrm{old}} + \\delta \\mathbf{x}$ for $\\mathbf{x} \\in \\{\\mathbf{v}, \\mathbf{c}, \\mathbf{m}\\}$ , where $\\delta \\mathbf{x}$ denotes an explicit revision produced by the model.", + "bbox": [ + 511, + 393, + 906, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We additionally decode weights for a pixel-shuffle upsampling step [41, 47]. We apply this upsampling to our visibility, confidence, and motion maps, bringing them from $1/8$ resolution to full resolution.", + "bbox": [ + 511, + 454, + 908, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We iterate our recurrent \"refinement\" stage 4 times, sharing weights. At training time, we use the outputs from every refinement step for supervision, and at test time we only use the final iteration's output.", + "bbox": [ + 511, + 513, + 908, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Model training", + "text_level": 1, + "bbox": [ + 513, + 580, + 666, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the datasets which we use for training, we have supervision in the form of either optical flow or sparse point tracks, which we simply treat as trajectories of varying lengths. We use an $L_{1}$ loss between estimated trajectories and ground truth, with a higher weight on loss for visible points:", + "bbox": [ + 511, + 603, + 908, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t r a c k}} = \\alpha \\sum_ {k} ^ {K} \\gamma^ {K - k} \\left(\\mathbb {1} _ {\\text {o c c}} / 5 + \\mathbb {1} _ {\\text {v i s}}\\right) | | P _ {k} - \\hat {P} | | _ {1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 686, + 906, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{k}$ is a set of estimated trajectories at refinement step $k$ , $\\hat{P}$ is the corresponding ground truth, $\\gamma$ (set to 0.8) makes later refinement steps weigh more in the loss, and $\\alpha$ is a balancing hyperparameter (set to 0.05).", + "bbox": [ + 511, + 732, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We supervise our visibility and confidence maps with a binary cross entropy loss. We ask the visibility estimates to match ground truth binary labels: $L_{\\mathrm{vis}} = \\sum_k^K \\mathrm{BCE}(V, \\hat{V})$ . We ask the confidence estimates to reflect whether or not the corresponding position estimates are within 12 pixels of ground truth [12, 24]: $L_{\\mathrm{conf}} = \\sum_k^K \\mathrm{BCE}(C, \\mathbb{1}[\\|X_k - \\hat{X}\\|_2 < 12])$ . We apply these losses at every refinement step.", + "bbox": [ + 511, + 792, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5256", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To supervise the model with sparse annotations, we use the coordinates of the ground truth to sample our corresponding estimates, and apply the loss at these sparse locations. We use bilinear sampling for trajectories, and nearest-neighbor sampling for visibility and confidence.", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 176, + 305, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide extensive implementation details in the supplementary material, but describe the main details here.", + "bbox": [ + 89, + 196, + 485, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our CNN backbone is based on a pre-trained ConvNeXt-Tiny [31]. We use the first three \"blocks\" of this architecture, totaling 12.72 million parameters. We convert the third block from stride 2 to stride 1, by applying bicubic interpolation to the $2 \\times 2$ stride-2 kernel to create a $3 \\times 3$ stride-1 kernel, and re-scale these weights according to the change in area (i.e., scaling by $4/9$ ).", + "bbox": [ + 88, + 227, + 485, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our full model is 16.48 million parameters. We train it using 8 A100-40G GPUs. We train in two stages: first on Kubric [17] for 200,000 steps at a learning rate of $5e - 4$ , and then on a mix of point tracking and optical flow datasets for 400,000 iterations at a learning rate of $1e - 5$ . We use the standard augmentations from prior work in point tracking [19] and optical flow [47], which consist of random shifting and scaling, color jitter, and square occlusions.", + "bbox": [ + 88, + 333, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our optical flow datasets include FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. Our point tracking datasets include Kubric [17], DynamicReplica [23], and PointOdyssey [59]. We also follow Harley et al. [19] and compute point tracks from optical flows where possible, making FlyingThings++, Monkaa++, Driving++, and Spring++.", + "bbox": [ + 88, + 454, + 485, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For trajectory estimates, we apply the model's additive revisions directly in pixel coordinates. We also supervise in pixel coordinates, which is a reason to scale the loss by the factor $\\alpha = 0.05$ . For visibility and confidence estimates, we apply the model's revisions on logits (rather than on probabilities), but we apply a sigmoid to these values before they are input to the refinement block, to stabilize their range.", + "bbox": [ + 88, + 575, + 485, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 694, + 223, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section summarizes the key results, which concern the accuracy of our tracker compared to prior state of the art, on a diverse array of test benchmarks. We provide additional details for our experiments in the appendices.", + "bbox": [ + 89, + 719, + 483, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics and benchmarks Our main metric is $\\delta_{\\mathrm{avg}}$ : an accuracy metric with max value 100, capturing how closely the estimated trajectory positions follow the ground truth positions. This is defined as the average of multiple $\\delta_{k}$ metrics, where $\\delta_{k}$ equals 100 when the estimate is within $k$ pixels of ground truth, measuring at $256 \\times 256$ resolution:", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {k} = 1 0 0 \\cdot \\mathbb {1} [ | | \\mathbf {p} - \\hat {\\mathbf {p}} | | _ {2} < k ]. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 883, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Averaging $\\delta_{k}$ over $k\\in \\{1,2,4,8,16\\}$ yields $\\delta_{\\mathrm{avg}}$ . We prioritize this over other considered metrics (e.g., L2 [19], PCK [19], AJ [11]) because we find it is both interpretable and robust to outliers. In certain benchmarks we also compute occlusion classification accuracy, and Average Jaccard (AJ) which mixes tracking accuracy with occlusion accuracy.", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate on a total of nine publicly-available benchmarks with point annotations, covering a wide set of domains that include animals (BADJA [3], Horse10 [33]), YouTube videos (TAPVid-DAVIS [11], TAPVid-Kinetics [11]), surveillance camera recordings (CroHD [46]), egocentric recordings (EgoPoints [10]), and robotics data (RGB-Stacking [28], RoboTAP [52]). We trim the video lengths to a maximum of 600 frames, and on the larger datasets (CroHD, DriveTrack, EgoPoints, Horse10, Kinetics, RoboTAP) we only track points from the first available query frame, which we find gives similar results to using all possible queries and full video lengths. We argue that using a wide array of benchmarks helps flatten the effects of label noise, and gives a better sense of the overall performance in practice. We encourage future work to follow our example, but we note that our full evaluation is time-consuming (with some baselines requiring multiple days to finish their pass).", + "bbox": [ + 511, + 181, + 908, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines We benchmark a variety of recent point trackers and optical flow models in our tests, including the optical flow models RAFT [47] and SEA-RAFT [54], the long-term flow method AccFlow [55], and point trackers PIPs++ [59], LocoTrack [6], DELTA [37], BootsTAPIR [13]. We make close comparisons against DELTA [37] and CoTracker3 [24], which are the most recent state-of-the-art point trackers. DELTA is relevant because it is a concurrent work focused on dense tracking. CoTracker3 is relevant because it outperforms all past work. It has two variants: one trained exclusively on Kubric, and another version finetuned on pseudolabels produced by an ensemble of point trackers on a curated set of 15,000 real videos. This finetuned version of CoTracker3 outperforms all past work. We note that the performance of CoTracker3 varies depending on how the query points are grouped [25]. The paper suggests to run each query in a separate forward pass, while also adding \"support\" points around the query and a sparse grid covering the image, but in our multi-benchmark evaluation this would be prohibitively expensive (e.g., weeks). We simply give CoTracker3 all queries at once, supplemented by a sparse grid of points around the image. Comparing with values reported by the authors (reproduced in the supplementary), we find that our setup over-estimates CoTracker3's accuracy, but makes our evaluation tractable.", + "bbox": [ + 511, + 441, + 908, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AllTracker variants In addition to our main model, which is 16.48 million parameters and trained on a mix of datasets,", + "bbox": [ + 511, + 821, + 906, + 853 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1The CoTracker3 repo also includes an \"offline\" variant, but it is not consistently better than the standard version, and in any case its memory consumption is outside our compute budget.", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5257", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/58631c245c45923755bfa7568d9532fcaa53f8c99e62a5628aa06a21f2ce6ea9.jpg", + "table_caption": [ + "Table 1. Comparison against recent point trackers and optical flow models, across nine datasets. We evaluate $\\delta_{\\mathrm{avg}}$ (higher is better), using an input resolution of $384\\times 512$ . The benchmarks are BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], EgoPoints [10], Horse10 [33], TAPVid-Kinetics [11], RGB-Stacking [28], and RoboTAP [52]." + ], + "table_footnote": [], + "table_body": "
MethodParams.TrainingBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
RAFT [47]5.26Flow mix23.729.348.544.841.027.864.382.872.248.3
SEA-RAFT [54]19.66Flow mix23.921.948.749.444.033.164.385.767.648.7
AccFlow [55]11.76Flow mix10.322.223.526.44.012.138.863.257.928.7
PIPs++ [59]17.57PointOdyssey34.127.562.551.338.521.464.270.473.449.3
LocoTrack [6]11.52Kubric41.443.168.066.558.448.970.080.376.961.5
BootsTAPIR [13]54.70Kubric+15M42.734.967.966.956.848.870.681.078.260.9
DELTA [37]59.17Kubric44.642.975.367.840.341.866.583.074.859.7
CoTracker2 [25]45.43Kubric40.031.770.967.843.233.965.873.47355.5
CoTracker3-Kub [24]25.39Kubric47.548.977.469.858.047.570.683.477.264.5
CoTracker3 [24]25.39Kubric+15k48.344.577.169.860.447.171.884.281.665.0
AllTracker-Tiny-Kub6.29Kubric45.439.673.765.155.945.270.686.179.362.3
AllTracker-Tiny6.29Kubric+mix47.539.874.363.958.345.571.588.180.763.3
AllTracker-Kub16.48Kubric46.442.375.266.160.349.071.390.182.264.8
AllTracker16.48Kubric+mix51.544.076.365.862.549.072.390.083.466.1
", + "bbox": [ + 107, + 137, + 885, + 359 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/65532211527a6b1dc486f6d00a1b04bb74b94b9c8c5b6e649ce0429dfaa81dff.jpg", + "table_caption": [ + "Table 2. High-resolution comparison between our model and CoTracker3 [24], which is the previous state of the art. We evaluate $\\delta_{\\mathrm{avg}}$ (higher is better) on nine point tracking benchmarks, at resolutions $448 \\times 768$ and $768 \\times 1024$ . Parameter counts are in millions." + ], + "table_footnote": [], + "table_body": "
ModelParams.ResolutionBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
CoTracker3-Kub [24]25.39448 × 76849.759.978.870.257.450.670.481.176.666.1
CoTracker3 [24]25.39448 × 76850.757.979.570.861.351.172.382.881.367.5
AllTracker-Tiny-Kub6.29448 × 76847.748.076.367.857.348.371.284.879.064.5
AllTracker-Tiny6.29448 × 76849.748.877.067.860.749.272.388.180.066.0
AllTracker-Kub16.48448 × 76849.851.677.668.361.352.671.990.482.067.3
AllTracker16.48448 × 76852.551.278.868.364.252.573.090.683.468.3
CoTracker3-Kub [24]25.39768 × 102447.862.077.072.051.946.467.076.274.563.9
CoTracker3 [24]25.39768 × 102449.864.379.672.458.448.571.177.980.266.9
AllTracker-Tiny-Kub6.29768 × 102448.456.978.571.255.149.271.180.878.165.5
AllTracker-Tiny6.29768 × 102451.657.079.171.159.250.772.487.479.267.5
AllTracker-Kub16.48768 × 102451.760.279.871.259.154.471.789.281.768.8
AllTracker16.48768 × 102453.653.480.672.364.354.673.190.683.269.5
", + "bbox": [ + 112, + 407, + 880, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we evaluate a \"tiny\" version, which is trained similarly to the main model but uses a cheaper CNN backbone and totals only 6.29 million parameters (see details in supplemental). For both versions, we additionally report the results with models trained exclusively on Kubric.", + "bbox": [ + 89, + 636, + 482, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main results", + "text_level": 1, + "bbox": [ + 89, + 724, + 223, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method's ability to track arbitrary points in diverse videos of various lengths, and compare against the state of the art in Table 1. For this evaluation we resize all input videos to $384 \\times 512$ . We note that DELTA (the only other dense point tracker) often runs out of memory when testing on a 40G GPU, so we use a 96G GPU to evaluate it.", + "bbox": [ + 89, + 748, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in the table, even our weakest variant, AllTracker-Tiny-Kub, is competitive with state of the art on many datasets, despite being only 6.29M parameters and training on a single synthetic dataset. Our full AllTracker", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "model outperforms all other models on average, with the closest competitor being the variant of CoTracker3 which relied on an expensive bootstrapping scheme (noted by “+15k” in the table). This CoTracker3 model wins on three datasets (CroHD, Davis, DriveTrack) and our model wins on the rest and wins on average (66.1 vs. 65.0). Our method has a substantial gain over past work on the RGB-Stacking benchmark, which has many query points inside textureless regions; our model's good performance here suggests that it is able to incorporate spatial context from a wider area than the previous models. We note that CoTracker3's bootstrapping technique could be combined with our contributions, but it is simpler to gather additional synthetic data as we have done.", + "bbox": [ + 511, + 636, + 906, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing models trained on Kubric to models trained with a broader data distribution, we see that the additional training data does not reliably improve performance in the surveillance dataset CroHD [46] or the driving dataset Driv-", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5258", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34c670fcf7adb627938c65d0cfba3dd20e63ed38658d79043f95913863e5bd64.jpg", + "image_caption": [ + "Figure 3. AllTracker (top right corner) delivers accurate multi-frame tracks at the throughput of an optical flow model." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 459, + 258 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9d16c6fb4521441705bfe48097e863f07c9ffa4057391aa63d4793a6476197ad.jpg", + "table_caption": [ + "Table 3. Average Jaccard evaluation on TAP-Vid datasets." + ], + "table_footnote": [], + "table_body": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker362.953.569.566.363.1
AllTracker63.359.181.171.968.9
", + "bbox": [ + 120, + 329, + 449, + 387 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2269b1f1e434a36af212032cb3ba2cce70056897781d59f46672d634cd22a433.jpg", + "table_caption": [ + "Table 4. Occlusion accuracy evaluation on TAP-Vid datasets." + ], + "table_footnote": [], + "table_body": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker390.185.691.689.889.3
AllTracker90.190.392.892.891.5
", + "bbox": [ + 122, + 417, + 449, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eTrack [1]. This suggests room for improvement, in terms of perhaps data balancing or model capacity.", + "bbox": [ + 89, + 494, + 482, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-resolution performance Focusing on the comparison between our model and CoTracker3, we evaluate these models at higher resolutions in Table 2. Here we see that our model's performance reliably increases at higher resolutions, while CoTracker3 plateaus after $448 \\times 768$ , leading to the surprising result that AllTracker-Tiny outperforms CoTracker3 at $768 \\times 1024$ . Our model is also more memory efficient: with AllTracker we can process videos at $768 \\times 1024$ (and produce 786,432 tracks at once) on a 40G A100 GPU, while with CoTracker3, despite only tracking sparse points, we encountered out-of-memory errors until eventually performing the tests on a 96G H100 GPU. Our model's relative memory efficiency is partly from using a spatial stride of 8 in the encoder, while CoTracker3 uses stride 4.", + "bbox": [ + 88, + 529, + 485, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Speed We compare the throughput of our model to other point trackers and optical flow models in Figure 3: our model runs at approximately the speed of optical flow methods, while achieving accuracies higher than the point trackers.", + "bbox": [ + 89, + 744, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Realtime inference Our model's sliding-window strategy can be run in \"streaming\" fashion real-time, at a penalty to accuracy: at $512 \\times 512$ our model runs at 57.9 FPS and achieves $62.6\\delta_{\\mathrm{avg}}$ (vs. 66.1 normally); BootsTAPIR [13] has also published a realtime variant, which runs at 21.4 FPS and reaches $62.2\\delta_{\\mathrm{avg}}$ .", + "bbox": [ + 89, + 809, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. A transformer-based temporal module performs better than mixer-based or convolution-based variants.", + "bbox": [ + 511, + 89, + 694, + 143 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Representing motion as long-range flow works better than instantaneous velocity, and positional embeddings do not help.", + "bbox": [ + 702, + 89, + 901, + 145 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a0be16cbca0277bdfac0ea7b95c2b8b91f7378623a71aadd8978bb163e3251c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Time componentAcc.
Transformer [51]56.8
MLP-Mixer [48]56.1
ConvNext-1D [31]54.6
Conv-1D55.8
", + "bbox": [ + 524, + 151, + 681, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/426281266c3cc324e2fddf2693eca4502c0ac0be096eafc16ff306d211e4dc53.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Motion representationAcc.
0 → t56.8
emb(0 → t)54.8
t-1 → t56.4
emb(t-1 → t)55.3
", + "bbox": [ + 714, + 151, + 890, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ec01aef321cb2890d73e5bcea765f8ff981a3c7591e515993fb1859feeb7f2ac.jpg", + "table_caption": [ + "Table 7. Hyperparameter search: It is best to use a ConvNeXt [31] backbone, 3 refinement blocks, and radius-4 correlations at 5 scales." + ], + "table_footnote": [], + "table_body": "
BackboneRefine blocksCorr. radiusCorr. scalesAcc.
ConvNeXt34556.8
ConvNeXt w/o pret.34554.9
BasicEncoder34556.1
ConvNeXt33556.5
ConvNeXt35555.5
ConvNeXt34455.5
ConvNeXt346Err.
ConvNeXt24555.7
ConvNeXt44556.1
", + "bbox": [ + 524, + 277, + 892, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additional metrics Evaluating AJ and occlusion accuracy on TAPVid datasets, we find that AllTracker obtains substantially better AJ than CoTracker3: 68.9 vs. 63.1 on average (see Table 3, and slightly better occlusion accuracy: 91.5 vs. 89.3 (see Table 4). We perform this evaluation only on the TAP-Vid datasets because we find that other datasets' visibility labels are not as reliable.", + "bbox": [ + 511, + 484, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on flow benchmarks While traditional optical flow is not our focus, it is interesting to inspect the model's accuracy in this task. We submitted AllTracker to the official SINTEL test benchmark, yielding end-point error scores of 1.673 on \"clean\" and 3.244 on \"final\"; this is not as good as SEA-RAFT [54] (1.309 / 2.601) but comparable to GMFlow [57] (1.736 / 2.902). We note that it is common in optical flow literature to produce a different model for each benchmark (finetuned with a particular data mix and resolution), but for this test we simply use our main model as-is. Qualitatively, AllTracker's flow maps appear to be coarser than the ones from SEA-RAFT [54], suggesting that the model is underfitting on this task. We provide more optical flow results in the supplementary.", + "bbox": [ + 511, + 595, + 908, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results We visualize the long-range flow maps for our model and selected baselines in Figure 4. Note that computing the dense flow map for the sparse methods takes multiple minutes, while our method (and the flow methods) produce it in under a second. We find that these flow visualizations reveal differences in the spatial coherence of the", + "bbox": [ + 511, + 809, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5259", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/feff13cbeb6eaf1cbb87d5b03c3bdc7ff18e6cef5c9879d8b6ffeb892d8f6f47.jpg", + "image_caption": [ + "Figure 4. AllTracker produces accurate displacement fields across dozens of frames. Prior optical flow methods struggle to make correspondences across wide time gaps, while our model uses temporal priors to resolve the ambiguity; prior point trackers take multiple minutes to produce output at this density, and show splotchy pattern errors, while our method produces coherent output in less than a second." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tracks: sparse point trackers from the past few years (PIPs++, LocoTrack, CoTracker3) show progressive improvement but include splotchy pattern errors in the motion fields. The optical flow methods (AccFlow, RAFT, SEA-RAFT) show better spatial coherence, but the estimates are unreliable when the displacements are too large. Our method produces motion fields that are both coherent and accurate.", + "bbox": [ + 89, + 273, + 485, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 89, + 388, + 202, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We verify our design choices in an exhaustive series of ablation studies. In these experiments, we train each model for 100,000 iterations with 2 GPUs, using Kubric, with 4 inference iterations per sample, at a learning rate of 4e-4, with videos of size $24 \\times 384 \\times 512$ and $56 \\times 256 \\times 256$ . We test at resolution $384 \\times 512$ on a validation dataset that we create from BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], Horse10 [33], and RoboTAP [52], and report the mean $\\delta_{\\mathrm{avg}}$ across this set. We find that this truncated training and evaluation setup is crucial for enabling a thorough exploration of the model space. We note that the absolute values of these ablation experiments should not be compared to the main experiments, but can be compared to each other.", + "bbox": [ + 89, + 411, + 483, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal module Prior work has explored different network architectures for learning temporal priors: PIPs [19] used an MLP-Mixer [48], $\\mathrm{PIPs}++$ [59] and TAPIR [12] used 1D convolutions (which might reasonably be upgraded to 1D ConvNeXt layers [31]), and CoTracker [25] used a transformer [51]. In Table 5 we compare these in our setup and find that a transformer works best. We additionally note that the transformer option is most amenable to changing the window size, which is important for jointly training on optical flow (e.g., convolution kernels of size 3 cannot apply to an input of length 2).", + "bbox": [ + 89, + 609, + 483, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion representation Most point trackers use sinusoidal position embeddings of motion, and methods vary on whether displacements should be relative to the query frame or to an adjacent frame. Table 6 shows that displacement relative to the query frame (i.e., frame 0) works best. This choice has the additional benefit of merging the problem with long-range optical flow. Table 6 also shows that sinusoidal embeddings do not help.", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Backbone Most point trackers use a \"BasicEncoder\" backbone which originated from RAFT [47]. We show in Table 7 that a pre-trained ConvNeXt backbone performs better, while a ConvNeXt backbone trained from scratch is worse.", + "bbox": [ + 511, + 273, + 906, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters Our main model uses 3 refinement blocks, and radius-4 correlations computed at 5 scales. We demonstrate in Table 7 that the neighboring alternatives are worse. We note that using 6 scales produces a runtime error, as $256 \\times 256$ input is already $1 \\times 1$ after 5 scales.", + "bbox": [ + 511, + 339, + 906, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Huber vs. L1 CoTracker3 [24] and TAPIR [12] recommend a Huber loss instead of the L1 used in PIPs [19]; in our setup this is not helpful: 54.9 (Huber) vs. 56.8 (L1).", + "bbox": [ + 511, + 420, + 906, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does frame ordering matter? We train a model with shuffled frames, to disentangle the benefits of joint multi-frame inference from the benefits of temporal continuity: we find that the impact varies from dataset to dataset, but shuffling is worse on average: 56.3 vs. 56.8.", + "bbox": [ + 511, + 470, + 908, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 566, + 769, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AllTracker is an approach to point tracking that treats the task as multi-frame optical flow. Our model delivers state-of-the-art performance on point tracking benchmarks, and produces dense output at high resolution, which past point trackers have struggled to do. AllTracker makes sparse point trackers (of similar speed and accuracy) redundant, but interestingly it does not make optical flow methods redundant: it does not outperform the state-of-the-art optical flow methods on optical flow estimation. The model appears to be underfitting on short-range motion estimation, suggesting that better models might be obtained with greater compute. A related limitation to address is the temporal window size: with bigger GPUs it may be possible to simply train and test with wider windows [24], and resolve longer occlusions. A broader area for future work is to add awareness about physical and common-sense constraints on motion, perhaps using 3D [37, 56] or more expressive model designs [18, 60].", + "bbox": [ + 511, + 593, + 908, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the TRI University 2.0 program, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship.", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5260", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arjun Balasingam, Joseph Chandler, Chenning Li, Zhoutong Zhang, and Hari Balakrishnan. Drivetrack: A benchmark for long-range point tracking in real-world videos. In CVPR, 2024. 6, 7, 8, 12, 13", + "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands context features. NeurIPS, 2024. 3", + "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and SMAL: Recovering the shape and motion of animals from video. In ACCV, 2018. 5, 6, 8, 12, 13", + "[4] Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching in variational motion estimation. TPAMI, 33(3):500-513, 2010. 2", + "[5] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In CVPR, 2024. 2", + "[6] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seunglyong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In ECCV, 2025. 3, 5, 6", + "[7] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Dense long-term motion estimation via statistical multi-step flow. In VISAPP, 2014. 2", + "[8] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Multi-reference combinatorial strategy towards longer long-term dense motion estimation. Computer Vision and Image Understanding, 150:66-80, 2016.", + "[9] Tomas Crivelli, Pierre-Henri Conze, Philippe Robert, and Patrick Pérez. From optical flow to dense long term correspondences. In International Conference on Image Processing, 2012. 2", + "[10] Ahmad Darkhalil, Rhodri Guerrier, Adam W Harley, and Dima Damen. Egopoints: Advancing point tracking for egocentric videos. arXiv:2412.04592, 2024. 5, 6, 12", + "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. In NeurIPS Datasets and Benchmarks, 2022. 3, 5, 6, 8, 12, 13", + "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In ICCV, 2023. 1, 3, 4, 8", + "[13] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In ACCV, 2024. 3, 5, 6, 7", + "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ig, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, 2015. 2, 3, 5", + "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 2013. 3, 5" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] James J Gibson. The perception of the visual world. 1950. 1, 2", + "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 3, 5", + "[18] Adam Harley, Yang You, Yang Zheng, Xinglong Sun, Nikhil Raghuraman, Sheldon Liang, Wen-Hsuan Chu, Suya You, Achal Dave, Pavel Tokmakov, et al. TAG: Tracking at any granularity. 2024. 8", + "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 3, 5, 8", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2", + "[21] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 2", + "[22] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 2", + "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 5", + "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv:2410.11831, 2024. 1, 3, 4, 5, 6, 8, 12", + "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 1, 2, 3, 5, 6, 8", + "[26] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The HCI benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In CVPR Workshops, 2016. 3, 5", + "[27] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In CVPR, 2024. 1, 3, 13", + "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In CoRL, 2021. 5, 6, 12", + "[29] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Feng Li, Tianhe Ren, Bohan Li, and Lei Zhang. TAPTRv2: Attention-based position update improves tracking any point. In NeurIPS, 2024. 3", + "[30] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. TAPTR: Tracking any point with transformers as detection. In ECCV, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5261", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5, 7, 8", + "[32] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 2", + "[33] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 5, 6, 8, 12, 13", + "[34] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 3, 5", + "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 3, 5", + "[36] Michal Neoral, Jonás Šerych, and Jérí Matas. Mft: Long-term tracking of every pixel. In WACV, 2024. 2", + "[37] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 1, 2, 5, 6, 8", + "[38] Diederick C Niehorster. Optic flow: a history. i-Perception, 12(6):20416695211055766, 2021. 2", + "[39] Stephan R Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In ICCV, 2017. 3, 5", + "[40] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In CVPR, 2006. 1, 2", + "[41] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3, 4", + "[42] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In CVPR, 2018. 2, 13", + "[43] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In CVPR, 2021. 3, 5", + "[44] Xinglong Sun, Adam W Harley, and Leonidas J Guibas. Refining pre-trained motion models. In ICRA, 2024. 3", + "[45] Narayanan Sundaram, Thomas Brox, and Kurt Keutzer. Dense point trajectories by GPU-accelerated large displacement optical flow. In ECCV, 2010. 2", + "[46] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In CVPR, 2021. 5, 6, 8, 12, 13", + "[47] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 4, 5, 6, 8, 13", + "[48] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-mixer: An all-mlp architecture for vision. arXiv:2105.01601, 2021. 7, 8" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Carlo Tomasi and Takeo Kanade. Detection and tracking of point. IJCV, 9:137-154, 1991. 2", + "[50] Marc Tournadre, Catherine Soladié, Nicolas Stoiber, and Pierre-Yves Richard. Dense trajectory fields: Consistent and efficient spatio-temporal pixel tracking. In ACCV, 2024. 1, 2", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 7, 8", + "[52] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. RoboTAP: Tracking arbitrary points for few-shot visual imitation. In ICRA, 2024. 5, 6, 8, 12, 13", + "[53] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual slam. In IROS, 2020. 3, 5", + "[54] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, 2024. 2, 3, 5, 6, 7, 11, 13", + "[55] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: Backward accumulation for long-range optical flow. In ICCV, 2023. 5, 6, 13", + "[56] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In CVPR, 2024. 8", + "[57] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 7, 13", + "[58] Jia Xu, René Ranftl, and Vladlen Koltun. Accurate optical flow via direct cost volume processing. In CVPR, 2017. 2", + "[59] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 1, 3, 5, 6, 8", + "[60] Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi SM Sajjadi, Sarath Chandar, and Ross Goroshin. TAPNext: Tracking any point (TAP) as next token prediction. arXiv:2504.05579, 2025.8" + ], + "bbox": [ + 514, + 92, + 906, + 672 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5262", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_model.json b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a82d7cf884d332d279a5137ff5b89b99e2ec0bf --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.184, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "AllTracker: Efficient Dense Point Tracking at High Resolution" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.842, + 0.234 + ], + "angle": 0, + "content": "Adam W. Harley\\(^{1}\\) Yang You\\(^{1}\\) Xinglong Sun\\(^{1}\\) Yang Zheng\\(^{1}\\) Nikhil Raghuraman\\(^{1}\\) Yunqi Gu\\(^{1}\\) Sheldon Liang\\(^{2}\\) Wen-Hsuan Chu\\(^{2}\\) Achal Dave\\(^{3}\\) Suya You\\(^{4}\\) Rares Ambrus\\(^{3}\\) Katerina Fragkiadaki\\(^{2}\\) Leonidas Guibas\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.162, + 0.236, + 0.833, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Stanford University \\(^{2}\\)Carnegie Mellon University \\(^{3}\\)Toyota Research Institute \\(^{4}\\)Army Research Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.681 + ], + "angle": 0, + "content": "We introduce AllTracker: a model that estimates long-range point tracks by way of estimating the flow field between a query frame and every other frame of a video. Unlike existing point tracking methods, our approach delivers high-resolution and dense (all-pixel) correspondence fields, which can be visualized as flow maps. Unlike existing optical flow methods, our approach corresponds one frame to hundreds of subsequent frames, rather than just the next frame. We develop a new architecture for this task, blending techniques from existing work in optical flow and point tracking: the model performs iterative inference on low-resolution grids of correspondence estimates, propagating information spatially via 2D convolution layers, and propagating information temporally via pixel-aligned attention layers. The model is fast and parameter-efficient (16 million parameters), and delivers state-of-the-art point tracking accuracy at high resolution (i.e., tracking \\(768 \\times 1024\\) pixels, on a 40G GPU). A benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets, and we find that doing so is crucial for top performance. We provide an extensive ablation study on our architecture details and training recipe, making it clear which details matter most. Our code and model weights are available: https://alltracker.github.io" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Estimating the long-range trajectories of arbitrary points in the world, using sequences of 2D images as input, has been a concrete and competitive challenge in computer vision since at least 2006 [40]. The utility of optical flow (i.e., the instantaneous velocity of pixels [16]) toward this goal has long been obvious, yet it has remained challenging to upgrade flows into long-range tracks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Instantaneous flows can be \"chained\" into multi-frame tracks, by interpolating in one flow field at the endpoints of the previous flow field, but imperfect flows will accumulate drift, and such chains must also be carefully stopped at" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.285, + 0.889, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.52, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 1. AllTracker estimates high-resolution optical flow between a \"query frame\" and every other frame of a video, using a sliding-window strategy. Point samples from these outputs can be interpreted as long-term point trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.717 + ], + "angle": 0, + "content": "occlusions [40]. An attractive shortcut here is to directly compute the flow between a reference frame and each other frame, and thus track drift-free and across occlusions, but estimating \"long-range flow\" becomes increasingly difficult as the time interval widens between the reference frame and the target, due to increasing variation in perspective, illumination, and scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In light of these challenges, recent work has side-stepped optical flow and created a line of bespoke point trackers focused on learning multi-frame temporal priors for reducing drift and tracking through occlusions [12, 19, 24, 25, 59]. These works explicitly highlight that the limited temporal context in flow-based methods is a severe weakness, and have made substantial progress by addressing this. However, the proposed solutions add temporal awareness at the cost of spatial awareness, and only deliver tracks for sparse sets of points. Recent attempts at \"dense\" point tracking [27, 37, 50] are not as accurate as the latest sparse trackers, and struggle with high-resolution input." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5253" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "In this paper, we demonstrate that learnable multi-frame temporal priors can be built together with high-resolution spatial awareness, by casting point tracking as a multi-frame long-range optical flow problem, as illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.486, + 0.486 + ], + "angle": 0, + "content": "Our design combines ideas from point tracking literature and optical flow literature, arriving at a composite architecture with broader capability. Following the trend in both areas, the heart of the model is a recurrent module that iteratively improves motion estimates [19, 47]. This module relies on information from spatial cross-correlations, giving it a strong inductive bias for feature-matching, which speeds up training. From point tracking literature, we borrow a per-pixel temporal module, to learn a motion prior and track through occlusions [19]. From optical flow literature, we borrow the idea of performing the majority of processing on a low-resolution grid, to allow fast spatial message-passing via 2D convolutions, and recover spatial precision at the end of the architecture with an upsampling layer [47, 54]. In relation to other point trackers, the key novelty of our approach is that we frame the problem as long-range dense flow, which makes sparse methods (of similar speed and accuracy) redundant. In relation to other optical flow models, the key novelty of our approach is that we solve a window of flow problems simultaneously, instead of frame-by-frame; the information shared within and across windows unlocks the capability to resolve flows across wide time intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.485, + 0.594 + ], + "angle": 0, + "content": "A secondary benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets. We therefore build a mix of publicly-available datasets to support our model, spanning different time-ranges, annotation densities, and resolutions, and train with uniform sampling across this mix. We find that this strategy, paired with a long training schedule, is crucial for top performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.485, + 0.701 + ], + "angle": 0, + "content": "In summary, our method blends techniques and datasets from optical flow literature and point tracking literature, and the resulting model is a state-of-the-art point tracker that operates in high resolution at full density. Inspired by CoTracker [25] and highlighting our capability for all-pixel tracking, we name our approach AllTracker. We will release our model, our data, and our code." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.72, + 0.233, + 0.735 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Optical flow The introduction of the concept of optic flow can be attributed to Gibson [16] (see Niehorster's review [38]). In computer vision, optical flow methods generally take two consecutive frames of video as input, and estimate the motion that relates the pixels of the first frame to the pixels of the second frame [21, 32]. The dominant technique, from classic approaches to current learning-based approaches, is to iterate on a solution: first, candidate correspondences are initialized (e.g., with zero-motion or constant velocity), then the \"costs\" of these correspondences are computed by measuring how" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "well the appearance features match, and this appearance cost is combined with smoothness terms or learned priors, and then better correspondences are sought in the neighborhood of the current solution [4, 22, 42, 47, 58]. Our work uses the same basic technique." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.909, + 0.379 + ], + "angle": 0, + "content": "The current state of the art in optical flow estimation is SEA-RAFT [54]. SEA-RAFT first directly estimates a low-resolution flow field [14, 22], then computes appearance features with a ResNet-34 [20], and computes pixel-to-pixel correlations with these features [58], then iteratively estimates refinements to the flow via 2D convolutions that take flows and correlations as input, and finally upsamples the flow to full resolution via a pixel-shuffle layer [41]. Our work takes inspiration from SEA-RAFT (and its predecessor RAFT [47]) in the way that we iteratively refine low-resolution flows and upsample to full resolution, but we operate on multiple frames at a time (i.e., 16 frames instead of 2), and we add temporal attention layers to help information propagate across the time axis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.909, + 0.565 + ], + "angle": 0, + "content": "Flow-based point tracking Optical flow is often used as a building block toward multi-frame tracking. The standard technique in this space is to \"chain\" flow vectors end-to-end to form longer tracks, and guard against drift by monitoring for occlusions [40, 45]. Recent techniques skip past occlusions by using multistep flows [7-9, 36], or attempt to track through occlusions using a learned temporal prior [5] (inspired by non-flow point trackers). These methods deliver dense multi-frame trajectories similar to our method, but first require an expensive pre-processing stage in which optical flows are estimated by a different model, while our method handles the full problem efficiently with a single model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.807 + ], + "angle": 0, + "content": "We highlight two concurrent works which propose a flow-based method somewhat similar to our own: DTF [50] and DELTA [37]. Both iteratively estimate flows to relate a reference frame to other frames, but unlike our work, these papers propose special-purpose transformer-based methods in which global spatial message-passing is approximated by cross-attending to sparse \"anchor\" or \"centroid\" tokens. Our method simply uses 2D convolutions for this step, more like SEA-RAFT [54]. We note also that DTF and DELTA both struggle with high-resolution inputs (running out of memory on our 40G GPUs), despite DELTA using low-resolution inference and high-resolution upsampling, comparable to SEA-RAFT and our approach. In performance, DTF is not competitive with recent point trackers (while ours exceeds them); DELTA's main model requires depth input, but we compare against its 2D variant and outperform it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Point trackers without flow Many classic methods track points directly, without relying on optical flow as a submodule [32, 49]. Harley et al. [19] recently introduced a new method in this space, with some components similar to flow models [47], but tracking points independently, and adding a multi-frame inference window which enabled the model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5254" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.184 + ], + "angle": 0, + "content": "to track through occlusions. That work, along with a new \"Tracking Any Point\" benchmark that appeared soon afterward [11], spurred much follow-up effort, involving multipoint context [24, 25], better initialization schemes [12], wider correlation context [2, 6], new transformer-based designs [29, 30], and post-hoc densification methods [27]." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.189, + 0.486, + 0.507 + ], + "angle": 0, + "content": "The current state of the art in point tracking is CoTracker3 [24]. CoTracker3 first computes feature maps for all frames, then initializes tracks for a sparse set of query points, then retrieves local correlations based on these tracks, and interleaves modules which pass information within tracks (temporally) and across tracks (spatially). The main novelty is in the spatial propagation: each point attends to a set of latent \"virtual points\" which learn a compressed representation of the video motions. While CoTracker3 delivers very accurate tracks, it requires users to pre-select a sparse set of points to track, and results vary based on how the queries are distributed [25]. CoTracker3 is more efficient than its predecessors [19, 25], but is still limited to tracking a few thousand points at a time. Our approach is simpler and more memory-efficient: we track on a low-resolution grid, where we achieve message-passing with simple 2D convolutions, and we recover spatial precision at the end of the architecture with a fast upsampling layer [41]. We note that this efficient upsampling technique was known well before the recent resurgence of point trackers [41], but perhaps its relevance was not fully appreciated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.76 + ], + "angle": 0, + "content": "Training data and self-supervision Current methods in optical flow and point tracking are data-driven, making data and supervision choices crucial. In optical flow, state-of-the-art methods rely on a combination of mostly synthetic datasets, including FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. In point tracking, the main training datasets are Kubric [17], FlyingThings++ [19], and PointOdyssey [59]. There is an art to creating a curriculum from diverse datasets to optimize downstream performance on particular benchmarks [54], but in our work we simply concatenate the datasets and shuffle the samples. We note that all previous point tracking methods did not use flow data, perhaps because most models do not support 2-frame inference or dense output; we therefore also provide experiments using Kubric alone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Motivated by the fact that the training data for point trackers is all synthetic, several groups have explored self-supervision methods based on boot-strapping pre-trained models using pseudo-labels computed on real videos [13, 24, 44]. However, these schemes offer only minor gains over the supervised weights where bootstrapping begins. In this work, we avoid any pseudo-labelling and simply add more synthetic data, capitalizing on our model's ability to accept supervision from optical flow data." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.315, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 2. AllTracker architecture. First, we compute feature maps for all frames, and copy the zeroth (query) feature map to every timestep, and compute multi-scale cost volumes. Then, we iterate a recurrent module, which references the query feature map and cost volume pyramid at each timestep, and estimates a low-resolution correspondence field, using interleaved 2D convolutions and pixel-aligned temporal attentions. The output of the RNN is upsampled into high-resolution optical flow maps, which relate all pixels of the zeroth frame to every other frame." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.472, + 0.631, + 0.489 + ], + "angle": 0, + "content": "3. AllTracker" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Our model takes a video as input, along with a \"query\" index, specifying which frame's pixels to track, and it outputs full-timespan tracking for all of the pixels in that frame." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.687 + ], + "angle": 0, + "content": "Concretely, we are given as input a video of shape \\( T, H, W, 3 \\), where \\( T \\) is the number of frames, and \\( H, W \\) indicate the spatial resolution of each frame. We are also given an index \\( t \\in T \\), indicating which frame has the pixels we need to track. Our final output is a tensor shaped \\( T, H, W, 4 \\): the first two channels are optical flow maps indicating the offset that takes each pixel in the query frame to its correspondence in every other frame, and the next two channels estimate visibility and confidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.911, + 0.902 + ], + "angle": 0, + "content": "A brief summary of the method is as follows. We operate in sliding window fashion across \\( T \\), processing subsequences of length \\( S \\), advancing at a stride of \\( S / 2 \\). We begin by quickly computing low-resolution feature maps for the subsequence, yielding a feature volume with dimensions \\( S, H / 8, W / 8, D \\), where \\( D \\) is the channel dimension of the features. We then initialize a low-resolution tensor of outputs, shaped \\( S, H / 8, W / 8, 4 \\), and iteratively revise this tensor using cross correlations and features across space and time as reference. Finally, we upsample the outputs, producing full resolution optical flow, visibility, and confidence, shaped \\( S, H, W, 4 \\). We then advance our window by \\( S / 2 \\), using previous estimates as initialization, and repeat inference. Figure 2 shows the method in a diagram." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5255" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.202, + 0.108 + ], + "angle": 0, + "content": "3.1. Encoding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.248 + ], + "angle": 0, + "content": "Our first stage encodes the input video frames into low-resolution feature maps. We achieve this with a ConvNeXt-Tiny [31] model (see Sec. 3.6 for implementation details). This encoder compresses a subsequence input of shape \\( S, H, W, 3 \\) to the shape \\( S, H/8, W/8, D \\), where \\( D \\) is the embedding dimension (in our case \\( D = 256 \\)), \\( S \\) is the subsequence window length (in our case \\( S = 16 \\)). When the full video length \\( T \\) is not too large (e.g., at training time), we compute all of the feature maps in parallel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.485, + 0.295 + ], + "angle": 0, + "content": "We then isolate the feature map of the query frame and tile it to the length of the subsequence, so that we have a copy for each timestep, as illustrated in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.386, + 0.32 + ], + "angle": 0, + "content": "3.2. Computing appearance similarity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.485, + 0.446 + ], + "angle": 0, + "content": "Using the feature maps, we build a multi-scale 4D correlation volume, capturing appearance-based tracking cues. We implement this in two steps: first we convert each feature map into a feature pyramid, by average-pooling at multiple strides (in our case 5 strides: \\(\\{1,2,4,8,16\\}\\)), and then we cross-correlate the query feature map with these pyramids. That is, for each feature vector in the query feature map, we compute its dot product across each timestep's pyramid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.485, + 0.552 + ], + "angle": 0, + "content": "The output of this step can be understood as a large collection of heatmaps: one heatmap per tracked pixel, per scale, per timestep. The model will index into these heatmaps to retrieve information on where correspondences lie. When a target's correspondence is visible at a certain timestep, we expect the respective heatmap to have a strong peak at the location of the correspondence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.275, + 0.575 + ], + "angle": 0, + "content": "3.3. Track initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.486, + 0.689 + ], + "angle": 0, + "content": "We next initialize estimates for tracking coordinates, visibility, and confidence, by creating a tensor with shape \\( S \\), \\( H/8 \\), \\( W/8 \\), 4, where the coordinate channels are initialized with a 2D meshgrid, and visibility and confidence are initialized with zeros. If we are on a window beyond the first, we populate this tensor with the estimates that overlap, and copy forward the values from the last estimated timestep." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.281, + 0.713 + ], + "angle": 0, + "content": "3.4. Iterative refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.78 + ], + "angle": 0, + "content": "The most important stage of our model is the iterative refinement stage, where we update all of the estimates. We describe the main points of the module here, but also provide extensive details (and a diagram) in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "For each timestep, for each pixel in the query frame, we have: a feature vector \\(\\mathbf{f}\\) (length \\(D\\)), visibility and confidence estimates \\(\\mathbf{v}\\), \\(\\mathbf{c}\\) (together length 2), a position estimate \\(\\mathbf{p}\\) (length 2), and a correlation pyramid \\(\\{\\mathbf{C}_1, \\mathbf{C}_2, \\ldots\\}\\) shaped \\(\\{H/8 \\times W/8, H/16 \\times W/16, \\ldots\\}\\). We convert this data into a \"local\" representation suitable for convolutions, as follows. For the position data, we replace the absolute position estimates \\(\\mathbf{p}\\) with motion estimates, by subtracting the source" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.197 + ], + "angle": 0, + "content": "positions: \\(\\mathbf{m} = \\mathbf{p} - \\mathbf{p}_0\\). For each level of the correlation pyramid, we extract a small patch centered at \\(\\mathbf{p}\\), and flatten these patches into a single vector \\(\\mathbf{q}\\) of length \\(L \\cdot (2R + 1)^2\\), where \\(R\\) is a radius parameter for the patches (e.g., 4) and \\(L\\) is the number of levels in the pyramid (e.g., 5). All together, we have \\(\\mathbf{f}, \\mathbf{v}, \\mathbf{c}, \\mathbf{m}, \\mathbf{q}\\) (per pixel, per timestep), which in our setup makes 665 channels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.909, + 0.393 + ], + "angle": 0, + "content": "Inside the recurrent module, we process the input data using interleaved spatial and temporal blocks. Each spatial block is a 2D ConvNeXt block; each temporal block is a pixel-aligned transformer block. By pixel-aligned, we mean that attention only happens along the temporal axis (i.e., with cost quadratic in \\( S = 16 \\)), and this is done for every pixel in parallel. We note that an important convenience of our design is that all of the tensors in this stage are aligned with the \"query\" frame; therefore pixel-aligned attention is attention between corresponding pixels. After the interleaved spatial and temporal blocks propagate tracking-related information across our window, we decode explicit revisions for visibility, confidence, and motion." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.907, + 0.454 + ], + "angle": 0, + "content": "We apply visibility, confidence, and motion revisions via simple summation with the previous values \\(\\mathbf{x}_{\\mathrm{new}} = \\mathbf{x}_{\\mathrm{old}} + \\delta \\mathbf{x}\\) for \\(\\mathbf{x} \\in \\{\\mathbf{v}, \\mathbf{c}, \\mathbf{m}\\}\\), where \\(\\delta \\mathbf{x}\\) denotes an explicit revision produced by the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.909, + 0.514 + ], + "angle": 0, + "content": "We additionally decode weights for a pixel-shuffle upsampling step [41, 47]. We apply this upsampling to our visibility, confidence, and motion maps, bringing them from \\(1/8\\) resolution to full resolution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.909, + 0.575 + ], + "angle": 0, + "content": "We iterate our recurrent \"refinement\" stage 4 times, sharing weights. At training time, we use the outputs from every refinement step for supervision, and at test time we only use the final iteration's output." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.582, + 0.667, + 0.598 + ], + "angle": 0, + "content": "3.5. Model training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.909, + 0.68 + ], + "angle": 0, + "content": "In the datasets which we use for training, we have supervision in the form of either optical flow or sparse point tracks, which we simply treat as trajectories of varying lengths. We use an \\( L_{1} \\) loss between estimated trajectories and ground truth, with a higher weight on loss for visible points:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.687, + 0.907, + 0.728 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t r a c k}} = \\alpha \\sum_ {k} ^ {K} \\gamma^ {K - k} \\left(\\mathbb {1} _ {\\text {o c c}} / 5 + \\mathbb {1} _ {\\text {v i s}}\\right) | | P _ {k} - \\hat {P} | | _ {1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.733, + 0.907, + 0.793 + ], + "angle": 0, + "content": "where \\(P_{k}\\) is a set of estimated trajectories at refinement step \\(k\\), \\(\\hat{P}\\) is the corresponding ground truth, \\(\\gamma\\) (set to 0.8) makes later refinement steps weigh more in the loss, and \\(\\alpha\\) is a balancing hyperparameter (set to 0.05)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We supervise our visibility and confidence maps with a binary cross entropy loss. We ask the visibility estimates to match ground truth binary labels: \\( L_{\\mathrm{vis}} = \\sum_k^K \\mathrm{BCE}(V, \\hat{V}) \\). We ask the confidence estimates to reflect whether or not the corresponding position estimates are within 12 pixels of ground truth [12, 24]: \\( L_{\\mathrm{conf}} = \\sum_k^K \\mathrm{BCE}(C, \\mathbb{1}[\\|X_k - \\hat{X}\\|_2 < 12]) \\). We apply these losses at every refinement step." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5256" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "To supervise the model with sparse annotations, we use the coordinates of the ground truth to sample our corresponding estimates, and apply the loss at these sparse locations. We use bilinear sampling for trajectories, and nearest-neighbor sampling for visibility and confidence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.177, + 0.307, + 0.192 + ], + "angle": 0, + "content": "3.6. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.486, + 0.227 + ], + "angle": 0, + "content": "We provide extensive implementation details in the supplementary material, but describe the main details here." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.228, + 0.486, + 0.333 + ], + "angle": 0, + "content": "Our CNN backbone is based on a pre-trained ConvNeXt-Tiny [31]. We use the first three \"blocks\" of this architecture, totaling 12.72 million parameters. We convert the third block from stride 2 to stride 1, by applying bicubic interpolation to the \\(2 \\times 2\\) stride-2 kernel to create a \\(3 \\times 3\\) stride-1 kernel, and re-scale these weights according to the change in area (i.e., scaling by \\(4/9\\))." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.485, + 0.456 + ], + "angle": 0, + "content": "Our full model is 16.48 million parameters. We train it using 8 A100-40G GPUs. We train in two stages: first on Kubric [17] for 200,000 steps at a learning rate of \\(5e - 4\\), and then on a mix of point tracking and optical flow datasets for 400,000 iterations at a learning rate of \\(1e - 5\\). We use the standard augmentations from prior work in point tracking [19] and optical flow [47], which consist of random shifting and scaling, color jitter, and square occlusions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.455, + 0.486, + 0.576 + ], + "angle": 0, + "content": "Our optical flow datasets include FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. Our point tracking datasets include Kubric [17], DynamicReplica [23], and PointOdyssey [59]. We also follow Harley et al. [19] and compute point tracks from optical flows where possible, making FlyingThings++, Monkaa++, Driving++, and Spring++." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.576, + 0.486, + 0.683 + ], + "angle": 0, + "content": "For trajectory estimates, we apply the model's additive revisions directly in pixel coordinates. We also supervise in pixel coordinates, which is a reason to scale the loss by the factor \\(\\alpha = 0.05\\). For visibility and confidence estimates, we apply the model's revisions on logits (rather than on probabilities), but we apply a sigmoid to these values before they are input to the refinement block, to stabilize their range." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.224, + 0.712 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.781 + ], + "angle": 0, + "content": "This section summarizes the key results, which concern the accuracy of our tracker compared to prior state of the art, on a diverse array of test benchmarks. We provide additional details for our experiments in the appendices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Metrics and benchmarks Our main metric is \\(\\delta_{\\mathrm{avg}}\\): an accuracy metric with max value 100, capturing how closely the estimated trajectory positions follow the ground truth positions. This is defined as the average of multiple \\(\\delta_{k}\\) metrics, where \\(\\delta_{k}\\) equals 100 when the estimate is within \\(k\\) pixels of ground truth, measuring at \\(256 \\times 256\\) resolution:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\delta_ {k} = 1 0 0 \\cdot \\mathbb {1} [ | | \\mathbf {p} - \\hat {\\mathbf {p}} | | _ {2} < k ]. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.182 + ], + "angle": 0, + "content": "Averaging \\(\\delta_{k}\\) over \\(k\\in \\{1,2,4,8,16\\}\\) yields \\(\\delta_{\\mathrm{avg}}\\). We prioritize this over other considered metrics (e.g., L2 [19], PCK [19], AJ [11]) because we find it is both interpretable and robust to outliers. In certain benchmarks we also compute occlusion classification accuracy, and Average Jaccard (AJ) which mixes tracking accuracy with occlusion accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.909, + 0.44 + ], + "angle": 0, + "content": "We evaluate on a total of nine publicly-available benchmarks with point annotations, covering a wide set of domains that include animals (BADJA [3], Horse10 [33]), YouTube videos (TAPVid-DAVIS [11], TAPVid-Kinetics [11]), surveillance camera recordings (CroHD [46]), egocentric recordings (EgoPoints [10]), and robotics data (RGB-Stacking [28], RoboTAP [52]). We trim the video lengths to a maximum of 600 frames, and on the larger datasets (CroHD, DriveTrack, EgoPoints, Horse10, Kinetics, RoboTAP) we only track points from the first available query frame, which we find gives similar results to using all possible queries and full video lengths. We argue that using a wide array of benchmarks helps flatten the effects of label noise, and gives a better sense of the overall performance in practice. We encourage future work to follow our example, but we note that our full evaluation is time-consuming (with some baselines requiring multiple days to finish their pass)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.819 + ], + "angle": 0, + "content": "Baselines We benchmark a variety of recent point trackers and optical flow models in our tests, including the optical flow models RAFT [47] and SEA-RAFT [54], the long-term flow method AccFlow [55], and point trackers PIPs++ [59], LocoTrack [6], DELTA [37], BootsTAPIR [13]. We make close comparisons against DELTA [37] and CoTracker3 [24], which are the most recent state-of-the-art point trackers. DELTA is relevant because it is a concurrent work focused on dense tracking. CoTracker3 is relevant because it outperforms all past work. It has two variants: one trained exclusively on Kubric, and another version finetuned on pseudolabels produced by an ensemble of point trackers on a curated set of 15,000 real videos. This finetuned version of CoTracker3 outperforms all past work. We note that the performance of CoTracker3 varies depending on how the query points are grouped [25]. The paper suggests to run each query in a separate forward pass, while also adding \"support\" points around the query and a sparse grid covering the image, but in our multi-benchmark evaluation this would be prohibitively expensive (e.g., weeks). We simply give CoTracker3 all queries at once, supplemented by a sparse grid of points around the image. Comparing with values reported by the authors (reproduced in the supplementary), we find that our setup over-estimates CoTracker3's accuracy, but makes our evaluation tractable." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.823, + 0.908, + 0.854 + ], + "angle": 0, + "content": "AllTracker variants In addition to our main model, which is 16.48 million parameters and trained on a mix of datasets," + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "1The CoTracker3 repo also includes an \"offline\" variant, but it is not consistently better than the standard version, and in any case its memory consumption is outside our compute budget." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5257" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison against recent point trackers and optical flow models, across nine datasets. We evaluate \\(\\delta_{\\mathrm{avg}}\\) (higher is better), using an input resolution of \\(384\\times 512\\). The benchmarks are BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], EgoPoints [10], Horse10 [33], TAPVid-Kinetics [11], RGB-Stacking [28], and RoboTAP [52]." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.138, + 0.887, + 0.361 + ], + "angle": 0, + "content": "
MethodParams.TrainingBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
RAFT [47]5.26Flow mix23.729.348.544.841.027.864.382.872.248.3
SEA-RAFT [54]19.66Flow mix23.921.948.749.444.033.164.385.767.648.7
AccFlow [55]11.76Flow mix10.322.223.526.44.012.138.863.257.928.7
PIPs++ [59]17.57PointOdyssey34.127.562.551.338.521.464.270.473.449.3
LocoTrack [6]11.52Kubric41.443.168.066.558.448.970.080.376.961.5
BootsTAPIR [13]54.70Kubric+15M42.734.967.966.956.848.870.681.078.260.9
DELTA [37]59.17Kubric44.642.975.367.840.341.866.583.074.859.7
CoTracker2 [25]45.43Kubric40.031.770.967.843.233.965.873.47355.5
CoTracker3-Kub [24]25.39Kubric47.548.977.469.858.047.570.683.477.264.5
CoTracker3 [24]25.39Kubric+15k48.344.577.169.860.447.171.884.281.665.0
AllTracker-Tiny-Kub6.29Kubric45.439.673.765.155.945.270.686.179.362.3
AllTracker-Tiny6.29Kubric+mix47.539.874.363.958.345.571.588.180.763.3
AllTracker-Kub16.48Kubric46.442.375.266.160.349.071.390.182.264.8
AllTracker16.48Kubric+mix51.544.076.365.862.549.072.390.083.466.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Table 2. High-resolution comparison between our model and CoTracker3 [24], which is the previous state of the art. We evaluate \\( \\delta_{\\mathrm{avg}} \\) (higher is better) on nine point tracking benchmarks, at resolutions \\( 448 \\times 768 \\) and \\( 768 \\times 1024 \\). Parameter counts are in millions." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.409, + 0.882, + 0.611 + ], + "angle": 0, + "content": "
ModelParams.ResolutionBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
CoTracker3-Kub [24]25.39448 × 76849.759.978.870.257.450.670.481.176.666.1
CoTracker3 [24]25.39448 × 76850.757.979.570.861.351.172.382.881.367.5
AllTracker-Tiny-Kub6.29448 × 76847.748.076.367.857.348.371.284.879.064.5
AllTracker-Tiny6.29448 × 76849.748.877.067.860.749.272.388.180.066.0
AllTracker-Kub16.48448 × 76849.851.677.668.361.352.671.990.482.067.3
AllTracker16.48448 × 76852.551.278.868.364.252.573.090.683.468.3
CoTracker3-Kub [24]25.39768 × 102447.862.077.072.051.946.467.076.274.563.9
CoTracker3 [24]25.39768 × 102449.864.379.672.458.448.571.177.980.266.9
AllTracker-Tiny-Kub6.29768 × 102448.456.978.571.255.149.271.180.878.165.5
AllTracker-Tiny6.29768 × 102451.657.079.171.159.250.772.487.479.267.5
AllTracker-Kub16.48768 × 102451.760.279.871.259.154.471.789.281.768.8
AllTracker16.48768 × 102453.653.480.672.364.354.673.190.683.269.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.713 + ], + "angle": 0, + "content": "we evaluate a \"tiny\" version, which is trained similarly to the main model but uses a cheaper CNN backbone and totals only 6.29 million parameters (see details in supplemental). For both versions, we additionally report the results with models trained exclusively on Kubric." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.225, + 0.74 + ], + "angle": 0, + "content": "4.1. Main results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We evaluate our method's ability to track arbitrary points in diverse videos of various lengths, and compare against the state of the art in Table 1. For this evaluation we resize all input videos to \\(384 \\times 512\\). We note that DELTA (the only other dense point tracker) often runs out of memory when testing on a 40G GPU, so we use a 96G GPU to evaluate it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As shown in the table, even our weakest variant, AllTracker-Tiny-Kub, is competitive with state of the art on many datasets, despite being only 6.29M parameters and training on a single synthetic dataset. Our full AllTracker" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.908, + 0.834 + ], + "angle": 0, + "content": "model outperforms all other models on average, with the closest competitor being the variant of CoTracker3 which relied on an expensive bootstrapping scheme (noted by “+15k” in the table). This CoTracker3 model wins on three datasets (CroHD, Davis, DriveTrack) and our model wins on the rest and wins on average (66.1 vs. 65.0). Our method has a substantial gain over past work on the RGB-Stacking benchmark, which has many query points inside textureless regions; our model's good performance here suggests that it is able to incorporate spatial context from a wider area than the previous models. We note that CoTracker3's bootstrapping technique could be combined with our contributions, but it is simpler to gather additional synthetic data as we have done." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparing models trained on Kubric to models trained with a broader data distribution, we see that the additional training data does not reliably improve performance in the surveillance dataset CroHD [46] or the driving dataset Driv-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5258" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.46, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.265, + 0.486, + 0.294 + ], + "angle": 0, + "content": "Figure 3. AllTracker (top right corner) delivers accurate multi-frame tracks at the throughput of an optical flow model." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.31, + 0.456, + 0.324 + ], + "angle": 0, + "content": "Table 3. Average Jaccard evaluation on TAP-Vid datasets." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.33, + 0.45, + 0.388 + ], + "angle": 0, + "content": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker362.953.569.566.363.1
AllTracker63.359.181.171.968.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.397, + 0.466, + 0.411 + ], + "angle": 0, + "content": "Table 4. Occlusion accuracy evaluation on TAP-Vid datasets." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.418, + 0.45, + 0.476 + ], + "angle": 0, + "content": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker390.185.691.689.889.3
AllTracker90.190.392.892.891.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.527 + ], + "angle": 0, + "content": "eTrack [1]. This suggests room for improvement, in terms of perhaps data balancing or model capacity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.486, + 0.743 + ], + "angle": 0, + "content": "High-resolution performance Focusing on the comparison between our model and CoTracker3, we evaluate these models at higher resolutions in Table 2. Here we see that our model's performance reliably increases at higher resolutions, while CoTracker3 plateaus after \\(448 \\times 768\\), leading to the surprising result that AllTracker-Tiny outperforms CoTracker3 at \\(768 \\times 1024\\). Our model is also more memory efficient: with AllTracker we can process videos at \\(768 \\times 1024\\) (and produce 786,432 tracks at once) on a 40G A100 GPU, while with CoTracker3, despite only tracking sparse points, we encountered out-of-memory errors until eventually performing the tests on a 96G H100 GPU. Our model's relative memory efficiency is partly from using a spatial stride of 8 in the encoder, while CoTracker3 uses stride 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Speed We compare the throughput of our model to other point trackers and optical flow models in Figure 3: our model runs at approximately the speed of optical flow methods, while achieving accuracies higher than the point trackers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Realtime inference Our model's sliding-window strategy can be run in \"streaming\" fashion real-time, at a penalty to accuracy: at \\(512 \\times 512\\) our model runs at 57.9 FPS and achieves \\(62.6\\delta_{\\mathrm{avg}}\\) (vs. 66.1 normally); BootsTAPIR [13] has also published a realtime variant, which runs at 21.4 FPS and reaches \\(62.2\\delta_{\\mathrm{avg}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.09, + 0.696, + 0.145 + ], + "angle": 0, + "content": "Table 5. A transformer-based temporal module performs better than mixer-based or convolution-based variants." + }, + { + "type": "text", + "bbox": [ + 0.703, + 0.09, + 0.903, + 0.146 + ], + "angle": 0, + "content": "Table 6. Representing motion as long-range flow works better than instantaneous velocity, and positional embeddings do not help." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.152, + 0.683, + 0.237 + ], + "angle": 0, + "content": "
Time componentAcc.
Transformer [51]56.8
MLP-Mixer [48]56.1
ConvNext-1D [31]54.6
Conv-1D55.8
" + }, + { + "type": "table", + "bbox": [ + 0.715, + 0.152, + 0.892, + 0.237 + ], + "angle": 0, + "content": "
Motion representationAcc.
0 → t56.8
emb(0 → t)54.8
t-1 → t56.4
emb(t-1 → t)55.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Table 7. Hyperparameter search: It is best to use a ConvNeXt [31] backbone, 3 refinement blocks, and radius-4 correlations at 5 scales." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.279, + 0.893, + 0.466 + ], + "angle": 0, + "content": "
BackboneRefine blocksCorr. radiusCorr. scalesAcc.
ConvNeXt34556.8
ConvNeXt w/o pret.34554.9
BasicEncoder34556.1
ConvNeXt33556.5
ConvNeXt35555.5
ConvNeXt34455.5
ConvNeXt346Err.
ConvNeXt24555.7
ConvNeXt44556.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Additional metrics Evaluating AJ and occlusion accuracy on TAPVid datasets, we find that AllTracker obtains substantially better AJ than CoTracker3: 68.9 vs. 63.1 on average (see Table 3, and slightly better occlusion accuracy: 91.5 vs. 89.3 (see Table 4). We perform this evaluation only on the TAP-Vid datasets because we find that other datasets' visibility labels are not as reliable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.807 + ], + "angle": 0, + "content": "Performance on flow benchmarks While traditional optical flow is not our focus, it is interesting to inspect the model's accuracy in this task. We submitted AllTracker to the official SINTEL test benchmark, yielding end-point error scores of 1.673 on \"clean\" and 3.244 on \"final\"; this is not as good as SEA-RAFT [54] (1.309 / 2.601) but comparable to GMFlow [57] (1.736 / 2.902). We note that it is common in optical flow literature to produce a different model for each benchmark (finetuned with a particular data mix and resolution), but for this test we simply use our main model as-is. Qualitatively, AllTracker's flow maps appear to be coarser than the ones from SEA-RAFT [54], suggesting that the model is underfitting on this task. We provide more optical flow results in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Qualitative results We visualize the long-range flow maps for our model and selected baselines in Figure 4. Note that computing the dense flow map for the sparse methods takes multiple minutes, while our method (and the flow methods) produce it in under a second. We find that these flow visualizations reveal differences in the spatial coherence of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5259" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.219, + 0.908, + 0.262 + ], + "angle": 0, + "content": "Figure 4. AllTracker produces accurate displacement fields across dozens of frames. Prior optical flow methods struggle to make correspondences across wide time gaps, while our model uses temporal priors to resolve the ambiguity; prior point trackers take multiple minutes to produce output at this density, and show splotchy pattern errors, while our method produces coherent output in less than a second." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.486, + 0.381 + ], + "angle": 0, + "content": "tracks: sparse point trackers from the past few years (PIPs++, LocoTrack, CoTracker3) show progressive improvement but include splotchy pattern errors in the motion fields. The optical flow methods (AccFlow, RAFT, SEA-RAFT) show better spatial coherence, but the estimates are unreliable when the displacements are too large. Our method produces motion fields that are both coherent and accurate." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.203, + 0.403 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.485, + 0.608 + ], + "angle": 0, + "content": "We verify our design choices in an exhaustive series of ablation studies. In these experiments, we train each model for 100,000 iterations with 2 GPUs, using Kubric, with 4 inference iterations per sample, at a learning rate of 4e-4, with videos of size \\(24 \\times 384 \\times 512\\) and \\(56 \\times 256 \\times 256\\). We test at resolution \\(384 \\times 512\\) on a validation dataset that we create from BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], Horse10 [33], and RoboTAP [52], and report the mean \\(\\delta_{\\mathrm{avg}}\\) across this set. We find that this truncated training and evaluation setup is crucial for enabling a thorough exploration of the model space. We note that the absolute values of these ablation experiments should not be compared to the main experiments, but can be compared to each other." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Temporal module Prior work has explored different network architectures for learning temporal priors: PIPs [19] used an MLP-Mixer [48], \\(\\mathrm{PIPs}++\\) [59] and TAPIR [12] used 1D convolutions (which might reasonably be upgraded to 1D ConvNeXt layers [31]), and CoTracker [25] used a transformer [51]. In Table 5 we compare these in our setup and find that a transformer works best. We additionally note that the transformer option is most amenable to changing the window size, which is important for jointly training on optical flow (e.g., convolution kernels of size 3 cannot apply to an input of length 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Motion representation Most point trackers use sinusoidal position embeddings of motion, and methods vary on whether displacements should be relative to the query frame or to an adjacent frame. Table 6 shows that displacement relative to the query frame (i.e., frame 0) works best. This choice has the additional benefit of merging the problem with long-range optical flow. Table 6 also shows that sinusoidal embeddings do not help." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.335 + ], + "angle": 0, + "content": "Backbone Most point trackers use a \"BasicEncoder\" backbone which originated from RAFT [47]. We show in Table 7 that a pre-trained ConvNeXt backbone performs better, while a ConvNeXt backbone trained from scratch is worse." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Hyperparameters Our main model uses 3 refinement blocks, and radius-4 correlations computed at 5 scales. We demonstrate in Table 7 that the neighboring alternatives are worse. We note that using 6 scales produces a runtime error, as \\(256 \\times 256\\) input is already \\(1 \\times 1\\) after 5 scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.908, + 0.467 + ], + "angle": 0, + "content": "Huber vs. L1 CoTracker3 [24] and TAPIR [12] recommend a Huber loss instead of the L1 used in PIPs [19]; in our setup this is not helpful: 54.9 (Huber) vs. 56.8 (L1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.547 + ], + "angle": 0, + "content": "Does frame ordering matter? We train a model with shuffled frames, to disentangle the benefits of joint multi-frame inference from the benefits of temporal continuity: we find that the impact varies from dataset to dataset, but shuffling is worse on average: 56.3 vs. 56.8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.567, + 0.771, + 0.582 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.909, + 0.851 + ], + "angle": 0, + "content": "AllTracker is an approach to point tracking that treats the task as multi-frame optical flow. Our model delivers state-of-the-art performance on point tracking benchmarks, and produces dense output at high resolution, which past point trackers have struggled to do. AllTracker makes sparse point trackers (of similar speed and accuracy) redundant, but interestingly it does not make optical flow methods redundant: it does not outperform the state-of-the-art optical flow methods on optical flow estimation. The model appears to be underfitting on short-range motion estimation, suggesting that better models might be obtained with greater compute. A related limitation to address is the temporal window size: with bigger GPUs it may be possible to simply train and test with wider windows [24], and resolve longer occlusions. A broader area for future work is to add awareness about physical and common-sense constraints on motion, perhaps using 3D [37, 56] or more expressive model designs [18, 60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the TRI University 2.0 program, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5260" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.17 + ], + "angle": 0, + "content": "[1] Arjun Balasingam, Joseph Chandler, Chenning Li, Zhoutong Zhang, and Hari Balakrishnan. Drivetrack: A benchmark for long-range point tracking in real-world videos. In CVPR, 2024. 6, 7, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.485, + 0.226 + ], + "angle": 0, + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands context features. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and SMAL: Recovering the shape and motion of animals from video. In ACCV, 2018. 5, 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[4] Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching in variational motion estimation. TPAMI, 33(3):500-513, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[5] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.484, + 0.414 + ], + "angle": 0, + "content": "[6] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seunglyong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In ECCV, 2025. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[7] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Dense long-term motion estimation via statistical multi-step flow. In VISAPP, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[8] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Multi-reference combinatorial strategy towards longer long-term dense motion estimation. Computer Vision and Image Understanding, 150:66-80, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[9] Tomas Crivelli, Pierre-Henri Conze, Philippe Robert, and Patrick Pérez. From optical flow to dense long term correspondences. In International Conference on Image Processing, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[10] Ahmad Darkhalil, Rhodri Guerrier, Adam W Harley, and Dima Damen. Egopoints: Advancing point tracking for egocentric videos. arXiv:2412.04592, 2024. 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. In NeurIPS Datasets and Benchmarks, 2022. 3, 5, 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In ICCV, 2023. 1, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[13] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In ACCV, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ig, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, 2015. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 2013. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.119 + ], + "angle": 0, + "content": "[16] James J Gibson. The perception of the visual world. 1950. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.908, + 0.178 + ], + "angle": 0, + "content": "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[18] Adam Harley, Yang You, Yang Zheng, Xinglong Sun, Nikhil Raghuraman, Sheldon Liang, Wen-Hsuan Chu, Suya You, Achal Dave, Pavel Tokmakov, et al. TAG: Tracking at any granularity. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.907, + 0.28 + ], + "angle": 0, + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.908, + 0.322 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.353 + ], + "angle": 0, + "content": "[21] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.356, + 0.907, + 0.41 + ], + "angle": 0, + "content": "[22] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.468 + ], + "angle": 0, + "content": "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.908, + 0.54 + ], + "angle": 0, + "content": "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv:2410.11831, 2024. 1, 3, 4, 5, 6, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.908, + 0.598 + ], + "angle": 0, + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.908, + 0.684 + ], + "angle": 0, + "content": "[26] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The HCI benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In CVPR Workshops, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.908, + 0.727 + ], + "angle": 0, + "content": "[27] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In CVPR, 2024. 1, 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.907, + 0.799 + ], + "angle": 0, + "content": "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In CoRL, 2021. 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.856 + ], + "angle": 0, + "content": "[29] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Feng Li, Tianhe Ren, Bohan Li, and Lei Zhang. TAPTRv2: Attention-based position update improves tracking any point. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[30] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. TAPTR: Tracking any point with transformers as detection. In ECCV, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5261" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[31] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[32] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[33] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 5, 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[34] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[36] Michal Neoral, Jonás Šerych, and Jérí Matas. Mft: Long-term tracking of every pixel. In WACV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.432 + ], + "angle": 0, + "content": "[37] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[38] Diederick C Niehorster. Optic flow: a history. i-Perception, 12(6):20416695211055766, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[39] Stephan R Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In ICCV, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.484, + 0.518 + ], + "angle": 0, + "content": "[40] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In CVPR, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[41] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[42] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In CVPR, 2018. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[43] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In CVPR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[44] Xinglong Sun, Adam W Harley, and Leonidas J Guibas. Refining pre-trained motion models. In ICRA, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[45] Narayanan Sundaram, Thomas Brox, and Kurt Keutzer. Dense point trajectories by GPU-accelerated large displacement optical flow. In ECCV, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[46] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In CVPR, 2021. 5, 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[47] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 4, 5, 6, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[48] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-mixer: An all-mlp architecture for vision. arXiv:2105.01601, 2021. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "[49] Carlo Tomasi and Takeo Kanade. Detection and tracking of point. IJCV, 9:137-154, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[50] Marc Tournadre, Catherine Soladié, Nicolas Stoiber, and Pierre-Yves Richard. Dense trajectory fields: Consistent and efficient spatio-temporal pixel tracking. In ACCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.22 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[52] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. RoboTAP: Tracking arbitrary points for few-shot visual imitation. In ICRA, 2024. 5, 6, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[53] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual slam. In IROS, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[54] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, 2024. 2, 3, 5, 6, 7, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[55] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: Backward accumulation for long-range optical flow. In ICCV, 2023. 5, 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.476 + ], + "angle": 0, + "content": "[56] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.907, + 0.518 + ], + "angle": 0, + "content": "[57] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[58] Jia Xu, René Ranftl, and Vladlen Koltun. Accurate optical flow via direct cost volume processing. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[59] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 1, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[60] Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi SM Sajjadi, Sarath Chandar, and Ross Goroshin. TAPNext: Tracking any point (TAP) as next token prediction. arXiv:2504.05579, 2025.8" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5262" + } + ] +] \ No newline at end of file diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_origin.pdf b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..930b2df23ef1ce0435b37724a4fb031b8109a337 --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/da1d8f6b-2320-4654-bad6-ea377ecc5c87_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cab04ce8eed9cbb2a4117fc6b47a15888a99ca31bd6bd3dcf3b73e0242de73d8 +size 6096889 diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/full.md b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b8617ebe8f30350aec69384d34b0e0cff83d1d3 --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/full.md @@ -0,0 +1,289 @@ +# AllTracker: Efficient Dense Point Tracking at High Resolution + +Adam W. Harley $^{1}$ Yang You $^{1}$ Xinglong Sun $^{1}$ Yang Zheng $^{1}$ Nikhil Raghuraman $^{1}$ Yunqi Gu $^{1}$ Sheldon Liang $^{2}$ Wen-Hsuan Chu $^{2}$ Achal Dave $^{3}$ Suya You $^{4}$ Rares Ambrus $^{3}$ Katerina Fragkiadaki $^{2}$ Leonidas Guibas $^{1}$ + +$^{1}$ Stanford University $^{2}$ Carnegie Mellon University $^{3}$ Toyota Research Institute $^{4}$ Army Research Laboratory + +# Abstract + +We introduce AllTracker: a model that estimates long-range point tracks by way of estimating the flow field between a query frame and every other frame of a video. Unlike existing point tracking methods, our approach delivers high-resolution and dense (all-pixel) correspondence fields, which can be visualized as flow maps. Unlike existing optical flow methods, our approach corresponds one frame to hundreds of subsequent frames, rather than just the next frame. We develop a new architecture for this task, blending techniques from existing work in optical flow and point tracking: the model performs iterative inference on low-resolution grids of correspondence estimates, propagating information spatially via 2D convolution layers, and propagating information temporally via pixel-aligned attention layers. The model is fast and parameter-efficient (16 million parameters), and delivers state-of-the-art point tracking accuracy at high resolution (i.e., tracking $768 \times 1024$ pixels, on a 40G GPU). A benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets, and we find that doing so is crucial for top performance. We provide an extensive ablation study on our architecture details and training recipe, making it clear which details matter most. Our code and model weights are available: https://alltracker.github.io + +# 1. Introduction + +Estimating the long-range trajectories of arbitrary points in the world, using sequences of 2D images as input, has been a concrete and competitive challenge in computer vision since at least 2006 [40]. The utility of optical flow (i.e., the instantaneous velocity of pixels [16]) toward this goal has long been obvious, yet it has remained challenging to upgrade flows into long-range tracks. + +Instantaneous flows can be "chained" into multi-frame tracks, by interpolating in one flow field at the endpoints of the previous flow field, but imperfect flows will accumulate drift, and such chains must also be carefully stopped at + +![](images/e8645867ce1ed03f51ab9fbee66039f8841c0d6dc9e2de09081f41068717e510.jpg) +Figure 1. AllTracker estimates high-resolution optical flow between a "query frame" and every other frame of a video, using a sliding-window strategy. Point samples from these outputs can be interpreted as long-term point trajectories. + +occlusions [40]. An attractive shortcut here is to directly compute the flow between a reference frame and each other frame, and thus track drift-free and across occlusions, but estimating "long-range flow" becomes increasingly difficult as the time interval widens between the reference frame and the target, due to increasing variation in perspective, illumination, and scene geometry. + +In light of these challenges, recent work has side-stepped optical flow and created a line of bespoke point trackers focused on learning multi-frame temporal priors for reducing drift and tracking through occlusions [12, 19, 24, 25, 59]. These works explicitly highlight that the limited temporal context in flow-based methods is a severe weakness, and have made substantial progress by addressing this. However, the proposed solutions add temporal awareness at the cost of spatial awareness, and only deliver tracks for sparse sets of points. Recent attempts at "dense" point tracking [27, 37, 50] are not as accurate as the latest sparse trackers, and struggle with high-resolution input. + +In this paper, we demonstrate that learnable multi-frame temporal priors can be built together with high-resolution spatial awareness, by casting point tracking as a multi-frame long-range optical flow problem, as illustrated in Figure 1. + +Our design combines ideas from point tracking literature and optical flow literature, arriving at a composite architecture with broader capability. Following the trend in both areas, the heart of the model is a recurrent module that iteratively improves motion estimates [19, 47]. This module relies on information from spatial cross-correlations, giving it a strong inductive bias for feature-matching, which speeds up training. From point tracking literature, we borrow a per-pixel temporal module, to learn a motion prior and track through occlusions [19]. From optical flow literature, we borrow the idea of performing the majority of processing on a low-resolution grid, to allow fast spatial message-passing via 2D convolutions, and recover spatial precision at the end of the architecture with an upsampling layer [47, 54]. In relation to other point trackers, the key novelty of our approach is that we frame the problem as long-range dense flow, which makes sparse methods (of similar speed and accuracy) redundant. In relation to other optical flow models, the key novelty of our approach is that we solve a window of flow problems simultaneously, instead of frame-by-frame; the information shared within and across windows unlocks the capability to resolve flows across wide time intervals. + +A secondary benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets. We therefore build a mix of publicly-available datasets to support our model, spanning different time-ranges, annotation densities, and resolutions, and train with uniform sampling across this mix. We find that this strategy, paired with a long training schedule, is crucial for top performance. + +In summary, our method blends techniques and datasets from optical flow literature and point tracking literature, and the resulting model is a state-of-the-art point tracker that operates in high resolution at full density. Inspired by CoTracker [25] and highlighting our capability for all-pixel tracking, we name our approach AllTracker. We will release our model, our data, and our code. + +# 2. Related Work + +Optical flow The introduction of the concept of optic flow can be attributed to Gibson [16] (see Niehorster's review [38]). In computer vision, optical flow methods generally take two consecutive frames of video as input, and estimate the motion that relates the pixels of the first frame to the pixels of the second frame [21, 32]. The dominant technique, from classic approaches to current learning-based approaches, is to iterate on a solution: first, candidate correspondences are initialized (e.g., with zero-motion or constant velocity), then the "costs" of these correspondences are computed by measuring how + +well the appearance features match, and this appearance cost is combined with smoothness terms or learned priors, and then better correspondences are sought in the neighborhood of the current solution [4, 22, 42, 47, 58]. Our work uses the same basic technique. + +The current state of the art in optical flow estimation is SEA-RAFT [54]. SEA-RAFT first directly estimates a low-resolution flow field [14, 22], then computes appearance features with a ResNet-34 [20], and computes pixel-to-pixel correlations with these features [58], then iteratively estimates refinements to the flow via 2D convolutions that take flows and correlations as input, and finally upsamples the flow to full resolution via a pixel-shuffle layer [41]. Our work takes inspiration from SEA-RAFT (and its predecessor RAFT [47]) in the way that we iteratively refine low-resolution flows and upsample to full resolution, but we operate on multiple frames at a time (i.e., 16 frames instead of 2), and we add temporal attention layers to help information propagate across the time axis. + +Flow-based point tracking Optical flow is often used as a building block toward multi-frame tracking. The standard technique in this space is to "chain" flow vectors end-to-end to form longer tracks, and guard against drift by monitoring for occlusions [40, 45]. Recent techniques skip past occlusions by using multistep flows [7-9, 36], or attempt to track through occlusions using a learned temporal prior [5] (inspired by non-flow point trackers). These methods deliver dense multi-frame trajectories similar to our method, but first require an expensive pre-processing stage in which optical flows are estimated by a different model, while our method handles the full problem efficiently with a single model. + +We highlight two concurrent works which propose a flow-based method somewhat similar to our own: DTF [50] and DELTA [37]. Both iteratively estimate flows to relate a reference frame to other frames, but unlike our work, these papers propose special-purpose transformer-based methods in which global spatial message-passing is approximated by cross-attending to sparse "anchor" or "centroid" tokens. Our method simply uses 2D convolutions for this step, more like SEA-RAFT [54]. We note also that DTF and DELTA both struggle with high-resolution inputs (running out of memory on our 40G GPUs), despite DELTA using low-resolution inference and high-resolution upsampling, comparable to SEA-RAFT and our approach. In performance, DTF is not competitive with recent point trackers (while ours exceeds them); DELTA's main model requires depth input, but we compare against its 2D variant and outperform it. + +Point trackers without flow Many classic methods track points directly, without relying on optical flow as a submodule [32, 49]. Harley et al. [19] recently introduced a new method in this space, with some components similar to flow models [47], but tracking points independently, and adding a multi-frame inference window which enabled the model + +to track through occlusions. That work, along with a new "Tracking Any Point" benchmark that appeared soon afterward [11], spurred much follow-up effort, involving multipoint context [24, 25], better initialization schemes [12], wider correlation context [2, 6], new transformer-based designs [29, 30], and post-hoc densification methods [27]. + +The current state of the art in point tracking is CoTracker3 [24]. CoTracker3 first computes feature maps for all frames, then initializes tracks for a sparse set of query points, then retrieves local correlations based on these tracks, and interleaves modules which pass information within tracks (temporally) and across tracks (spatially). The main novelty is in the spatial propagation: each point attends to a set of latent "virtual points" which learn a compressed representation of the video motions. While CoTracker3 delivers very accurate tracks, it requires users to pre-select a sparse set of points to track, and results vary based on how the queries are distributed [25]. CoTracker3 is more efficient than its predecessors [19, 25], but is still limited to tracking a few thousand points at a time. Our approach is simpler and more memory-efficient: we track on a low-resolution grid, where we achieve message-passing with simple 2D convolutions, and we recover spatial precision at the end of the architecture with a fast upsampling layer [41]. We note that this efficient upsampling technique was known well before the recent resurgence of point trackers [41], but perhaps its relevance was not fully appreciated. + +Training data and self-supervision Current methods in optical flow and point tracking are data-driven, making data and supervision choices crucial. In optical flow, state-of-the-art methods rely on a combination of mostly synthetic datasets, including FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. In point tracking, the main training datasets are Kubric [17], FlyingThings++ [19], and PointOdyssey [59]. There is an art to creating a curriculum from diverse datasets to optimize downstream performance on particular benchmarks [54], but in our work we simply concatenate the datasets and shuffle the samples. We note that all previous point tracking methods did not use flow data, perhaps because most models do not support 2-frame inference or dense output; we therefore also provide experiments using Kubric alone. + +Motivated by the fact that the training data for point trackers is all synthetic, several groups have explored self-supervision methods based on boot-strapping pre-trained models using pseudo-labels computed on real videos [13, 24, 44]. However, these schemes offer only minor gains over the supervised weights where bootstrapping begins. In this work, we avoid any pseudo-labelling and simply add more synthetic data, capitalizing on our model's ability to accept supervision from optical flow data. + +![](images/0cbe47e5a494155d34f69be48ce6a97f09f26d1b43b72db8d812b3ca206fc0e5.jpg) +Figure 2. AllTracker architecture. First, we compute feature maps for all frames, and copy the zeroth (query) feature map to every timestep, and compute multi-scale cost volumes. Then, we iterate a recurrent module, which references the query feature map and cost volume pyramid at each timestep, and estimates a low-resolution correspondence field, using interleaved 2D convolutions and pixel-aligned temporal attentions. The output of the RNN is upsampled into high-resolution optical flow maps, which relate all pixels of the zeroth frame to every other frame. + +# 3. AllTracker + +Our model takes a video as input, along with a "query" index, specifying which frame's pixels to track, and it outputs full-timespan tracking for all of the pixels in that frame. + +Concretely, we are given as input a video of shape $T, H, W, 3$ , where $T$ is the number of frames, and $H, W$ indicate the spatial resolution of each frame. We are also given an index $t \in T$ , indicating which frame has the pixels we need to track. Our final output is a tensor shaped $T, H, W, 4$ : the first two channels are optical flow maps indicating the offset that takes each pixel in the query frame to its correspondence in every other frame, and the next two channels estimate visibility and confidence. + +A brief summary of the method is as follows. We operate in sliding window fashion across $T$ , processing subsequences of length $S$ , advancing at a stride of $S / 2$ . We begin by quickly computing low-resolution feature maps for the subsequence, yielding a feature volume with dimensions $S, H / 8, W / 8, D$ , where $D$ is the channel dimension of the features. We then initialize a low-resolution tensor of outputs, shaped $S, H / 8, W / 8, 4$ , and iteratively revise this tensor using cross correlations and features across space and time as reference. Finally, we upsample the outputs, producing full resolution optical flow, visibility, and confidence, shaped $S, H, W, 4$ . We then advance our window by $S / 2$ , using previous estimates as initialization, and repeat inference. Figure 2 shows the method in a diagram. + +# 3.1. Encoding + +Our first stage encodes the input video frames into low-resolution feature maps. We achieve this with a ConvNeXt-Tiny [31] model (see Sec. 3.6 for implementation details). This encoder compresses a subsequence input of shape $S, H, W, 3$ to the shape $S, H/8, W/8, D$ , where $D$ is the embedding dimension (in our case $D = 256$ ), $S$ is the subsequence window length (in our case $S = 16$ ). When the full video length $T$ is not too large (e.g., at training time), we compute all of the feature maps in parallel. + +We then isolate the feature map of the query frame and tile it to the length of the subsequence, so that we have a copy for each timestep, as illustrated in Figure 2. + +# 3.2. Computing appearance similarity + +Using the feature maps, we build a multi-scale 4D correlation volume, capturing appearance-based tracking cues. We implement this in two steps: first we convert each feature map into a feature pyramid, by average-pooling at multiple strides (in our case 5 strides: $\{1,2,4,8,16\}$ ), and then we cross-correlate the query feature map with these pyramids. That is, for each feature vector in the query feature map, we compute its dot product across each timestep's pyramid. + +The output of this step can be understood as a large collection of heatmaps: one heatmap per tracked pixel, per scale, per timestep. The model will index into these heatmaps to retrieve information on where correspondences lie. When a target's correspondence is visible at a certain timestep, we expect the respective heatmap to have a strong peak at the location of the correspondence. + +# 3.3. Track initialization + +We next initialize estimates for tracking coordinates, visibility, and confidence, by creating a tensor with shape $S$ , $H/8$ , $W/8$ , 4, where the coordinate channels are initialized with a 2D meshgrid, and visibility and confidence are initialized with zeros. If we are on a window beyond the first, we populate this tensor with the estimates that overlap, and copy forward the values from the last estimated timestep. + +# 3.4. Iterative refinement + +The most important stage of our model is the iterative refinement stage, where we update all of the estimates. We describe the main points of the module here, but also provide extensive details (and a diagram) in the supplementary. + +For each timestep, for each pixel in the query frame, we have: a feature vector $\mathbf{f}$ (length $D$ ), visibility and confidence estimates $\mathbf{v}$ , $\mathbf{c}$ (together length 2), a position estimate $\mathbf{p}$ (length 2), and a correlation pyramid $\{\mathbf{C}_1, \mathbf{C}_2, \ldots\}$ shaped $\{H/8 \times W/8, H/16 \times W/16, \ldots\}$ . We convert this data into a "local" representation suitable for convolutions, as follows. For the position data, we replace the absolute position estimates $\mathbf{p}$ with motion estimates, by subtracting the source + +positions: $\mathbf{m} = \mathbf{p} - \mathbf{p}_0$ . For each level of the correlation pyramid, we extract a small patch centered at $\mathbf{p}$ , and flatten these patches into a single vector $\mathbf{q}$ of length $L \cdot (2R + 1)^2$ , where $R$ is a radius parameter for the patches (e.g., 4) and $L$ is the number of levels in the pyramid (e.g., 5). All together, we have $\mathbf{f}, \mathbf{v}, \mathbf{c}, \mathbf{m}, \mathbf{q}$ (per pixel, per timestep), which in our setup makes 665 channels. + +Inside the recurrent module, we process the input data using interleaved spatial and temporal blocks. Each spatial block is a 2D ConvNeXt block; each temporal block is a pixel-aligned transformer block. By pixel-aligned, we mean that attention only happens along the temporal axis (i.e., with cost quadratic in $S = 16$ ), and this is done for every pixel in parallel. We note that an important convenience of our design is that all of the tensors in this stage are aligned with the "query" frame; therefore pixel-aligned attention is attention between corresponding pixels. After the interleaved spatial and temporal blocks propagate tracking-related information across our window, we decode explicit revisions for visibility, confidence, and motion. + +We apply visibility, confidence, and motion revisions via simple summation with the previous values $\mathbf{x}_{\mathrm{new}} = \mathbf{x}_{\mathrm{old}} + \delta \mathbf{x}$ for $\mathbf{x} \in \{\mathbf{v}, \mathbf{c}, \mathbf{m}\}$ , where $\delta \mathbf{x}$ denotes an explicit revision produced by the model. + +We additionally decode weights for a pixel-shuffle upsampling step [41, 47]. We apply this upsampling to our visibility, confidence, and motion maps, bringing them from $1/8$ resolution to full resolution. + +We iterate our recurrent "refinement" stage 4 times, sharing weights. At training time, we use the outputs from every refinement step for supervision, and at test time we only use the final iteration's output. + +# 3.5. Model training + +In the datasets which we use for training, we have supervision in the form of either optical flow or sparse point tracks, which we simply treat as trajectories of varying lengths. We use an $L_{1}$ loss between estimated trajectories and ground truth, with a higher weight on loss for visible points: + +$$ +L _ {\text {t r a c k}} = \alpha \sum_ {k} ^ {K} \gamma^ {K - k} \left(\mathbb {1} _ {\text {o c c}} / 5 + \mathbb {1} _ {\text {v i s}}\right) | | P _ {k} - \hat {P} | | _ {1}, \tag {1} +$$ + +where $P_{k}$ is a set of estimated trajectories at refinement step $k$ , $\hat{P}$ is the corresponding ground truth, $\gamma$ (set to 0.8) makes later refinement steps weigh more in the loss, and $\alpha$ is a balancing hyperparameter (set to 0.05). + +We supervise our visibility and confidence maps with a binary cross entropy loss. We ask the visibility estimates to match ground truth binary labels: $L_{\mathrm{vis}} = \sum_k^K \mathrm{BCE}(V, \hat{V})$ . We ask the confidence estimates to reflect whether or not the corresponding position estimates are within 12 pixels of ground truth [12, 24]: $L_{\mathrm{conf}} = \sum_k^K \mathrm{BCE}(C, \mathbb{1}[\|X_k - \hat{X}\|_2 < 12])$ . We apply these losses at every refinement step. + +To supervise the model with sparse annotations, we use the coordinates of the ground truth to sample our corresponding estimates, and apply the loss at these sparse locations. We use bilinear sampling for trajectories, and nearest-neighbor sampling for visibility and confidence. + +# 3.6. Implementation details + +We provide extensive implementation details in the supplementary material, but describe the main details here. + +Our CNN backbone is based on a pre-trained ConvNeXt-Tiny [31]. We use the first three "blocks" of this architecture, totaling 12.72 million parameters. We convert the third block from stride 2 to stride 1, by applying bicubic interpolation to the $2 \times 2$ stride-2 kernel to create a $3 \times 3$ stride-1 kernel, and re-scale these weights according to the change in area (i.e., scaling by $4/9$ ). + +Our full model is 16.48 million parameters. We train it using 8 A100-40G GPUs. We train in two stages: first on Kubric [17] for 200,000 steps at a learning rate of $5e - 4$ , and then on a mix of point tracking and optical flow datasets for 400,000 iterations at a learning rate of $1e - 5$ . We use the standard augmentations from prior work in point tracking [19] and optical flow [47], which consist of random shifting and scaling, color jitter, and square occlusions. + +Our optical flow datasets include FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. Our point tracking datasets include Kubric [17], DynamicReplica [23], and PointOdyssey [59]. We also follow Harley et al. [19] and compute point tracks from optical flows where possible, making FlyingThings++, Monkaa++, Driving++, and Spring++. + +For trajectory estimates, we apply the model's additive revisions directly in pixel coordinates. We also supervise in pixel coordinates, which is a reason to scale the loss by the factor $\alpha = 0.05$ . For visibility and confidence estimates, we apply the model's revisions on logits (rather than on probabilities), but we apply a sigmoid to these values before they are input to the refinement block, to stabilize their range. + +# 4. Experiments + +This section summarizes the key results, which concern the accuracy of our tracker compared to prior state of the art, on a diverse array of test benchmarks. We provide additional details for our experiments in the appendices. + +Metrics and benchmarks Our main metric is $\delta_{\mathrm{avg}}$ : an accuracy metric with max value 100, capturing how closely the estimated trajectory positions follow the ground truth positions. This is defined as the average of multiple $\delta_{k}$ metrics, where $\delta_{k}$ equals 100 when the estimate is within $k$ pixels of ground truth, measuring at $256 \times 256$ resolution: + +$$ +\delta_ {k} = 1 0 0 \cdot \mathbb {1} [ | | \mathbf {p} - \hat {\mathbf {p}} | | _ {2} < k ]. \tag {2} +$$ + +Averaging $\delta_{k}$ over $k\in \{1,2,4,8,16\}$ yields $\delta_{\mathrm{avg}}$ . We prioritize this over other considered metrics (e.g., L2 [19], PCK [19], AJ [11]) because we find it is both interpretable and robust to outliers. In certain benchmarks we also compute occlusion classification accuracy, and Average Jaccard (AJ) which mixes tracking accuracy with occlusion accuracy. + +We evaluate on a total of nine publicly-available benchmarks with point annotations, covering a wide set of domains that include animals (BADJA [3], Horse10 [33]), YouTube videos (TAPVid-DAVIS [11], TAPVid-Kinetics [11]), surveillance camera recordings (CroHD [46]), egocentric recordings (EgoPoints [10]), and robotics data (RGB-Stacking [28], RoboTAP [52]). We trim the video lengths to a maximum of 600 frames, and on the larger datasets (CroHD, DriveTrack, EgoPoints, Horse10, Kinetics, RoboTAP) we only track points from the first available query frame, which we find gives similar results to using all possible queries and full video lengths. We argue that using a wide array of benchmarks helps flatten the effects of label noise, and gives a better sense of the overall performance in practice. We encourage future work to follow our example, but we note that our full evaluation is time-consuming (with some baselines requiring multiple days to finish their pass). + +Baselines We benchmark a variety of recent point trackers and optical flow models in our tests, including the optical flow models RAFT [47] and SEA-RAFT [54], the long-term flow method AccFlow [55], and point trackers PIPs++ [59], LocoTrack [6], DELTA [37], BootsTAPIR [13]. We make close comparisons against DELTA [37] and CoTracker3 [24], which are the most recent state-of-the-art point trackers. DELTA is relevant because it is a concurrent work focused on dense tracking. CoTracker3 is relevant because it outperforms all past work. It has two variants: one trained exclusively on Kubric, and another version finetuned on pseudolabels produced by an ensemble of point trackers on a curated set of 15,000 real videos. This finetuned version of CoTracker3 outperforms all past work. We note that the performance of CoTracker3 varies depending on how the query points are grouped [25]. The paper suggests to run each query in a separate forward pass, while also adding "support" points around the query and a sparse grid covering the image, but in our multi-benchmark evaluation this would be prohibitively expensive (e.g., weeks). We simply give CoTracker3 all queries at once, supplemented by a sparse grid of points around the image. Comparing with values reported by the authors (reproduced in the supplementary), we find that our setup over-estimates CoTracker3's accuracy, but makes our evaluation tractable. + +AllTracker variants In addition to our main model, which is 16.48 million parameters and trained on a mix of datasets, + +Table 1. Comparison against recent point trackers and optical flow models, across nine datasets. We evaluate $\delta_{\mathrm{avg}}$ (higher is better), using an input resolution of $384\times 512$ . The benchmarks are BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], EgoPoints [10], Horse10 [33], TAPVid-Kinetics [11], RGB-Stacking [28], and RoboTAP [52]. + +
MethodParams.TrainingBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
RAFT [47]5.26Flow mix23.729.348.544.841.027.864.382.872.248.3
SEA-RAFT [54]19.66Flow mix23.921.948.749.444.033.164.385.767.648.7
AccFlow [55]11.76Flow mix10.322.223.526.44.012.138.863.257.928.7
PIPs++ [59]17.57PointOdyssey34.127.562.551.338.521.464.270.473.449.3
LocoTrack [6]11.52Kubric41.443.168.066.558.448.970.080.376.961.5
BootsTAPIR [13]54.70Kubric+15M42.734.967.966.956.848.870.681.078.260.9
DELTA [37]59.17Kubric44.642.975.367.840.341.866.583.074.859.7
CoTracker2 [25]45.43Kubric40.031.770.967.843.233.965.873.47355.5
CoTracker3-Kub [24]25.39Kubric47.548.977.469.858.047.570.683.477.264.5
CoTracker3 [24]25.39Kubric+15k48.344.577.169.860.447.171.884.281.665.0
AllTracker-Tiny-Kub6.29Kubric45.439.673.765.155.945.270.686.179.362.3
AllTracker-Tiny6.29Kubric+mix47.539.874.363.958.345.571.588.180.763.3
AllTracker-Kub16.48Kubric46.442.375.266.160.349.071.390.182.264.8
AllTracker16.48Kubric+mix51.544.076.365.862.549.072.390.083.466.1
+ +Table 2. High-resolution comparison between our model and CoTracker3 [24], which is the previous state of the art. We evaluate $\delta_{\mathrm{avg}}$ (higher is better) on nine point tracking benchmarks, at resolutions $448 \times 768$ and $768 \times 1024$ . Parameter counts are in millions. + +
ModelParams.ResolutionBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
CoTracker3-Kub [24]25.39448 × 76849.759.978.870.257.450.670.481.176.666.1
CoTracker3 [24]25.39448 × 76850.757.979.570.861.351.172.382.881.367.5
AllTracker-Tiny-Kub6.29448 × 76847.748.076.367.857.348.371.284.879.064.5
AllTracker-Tiny6.29448 × 76849.748.877.067.860.749.272.388.180.066.0
AllTracker-Kub16.48448 × 76849.851.677.668.361.352.671.990.482.067.3
AllTracker16.48448 × 76852.551.278.868.364.252.573.090.683.468.3
CoTracker3-Kub [24]25.39768 × 102447.862.077.072.051.946.467.076.274.563.9
CoTracker3 [24]25.39768 × 102449.864.379.672.458.448.571.177.980.266.9
AllTracker-Tiny-Kub6.29768 × 102448.456.978.571.255.149.271.180.878.165.5
AllTracker-Tiny6.29768 × 102451.657.079.171.159.250.772.487.479.267.5
AllTracker-Kub16.48768 × 102451.760.279.871.259.154.471.789.281.768.8
AllTracker16.48768 × 102453.653.480.672.364.354.673.190.683.269.5
+ +we evaluate a "tiny" version, which is trained similarly to the main model but uses a cheaper CNN backbone and totals only 6.29 million parameters (see details in supplemental). For both versions, we additionally report the results with models trained exclusively on Kubric. + +# 4.1. Main results + +We evaluate our method's ability to track arbitrary points in diverse videos of various lengths, and compare against the state of the art in Table 1. For this evaluation we resize all input videos to $384 \times 512$ . We note that DELTA (the only other dense point tracker) often runs out of memory when testing on a 40G GPU, so we use a 96G GPU to evaluate it. + +As shown in the table, even our weakest variant, AllTracker-Tiny-Kub, is competitive with state of the art on many datasets, despite being only 6.29M parameters and training on a single synthetic dataset. Our full AllTracker + +model outperforms all other models on average, with the closest competitor being the variant of CoTracker3 which relied on an expensive bootstrapping scheme (noted by “+15k” in the table). This CoTracker3 model wins on three datasets (CroHD, Davis, DriveTrack) and our model wins on the rest and wins on average (66.1 vs. 65.0). Our method has a substantial gain over past work on the RGB-Stacking benchmark, which has many query points inside textureless regions; our model's good performance here suggests that it is able to incorporate spatial context from a wider area than the previous models. We note that CoTracker3's bootstrapping technique could be combined with our contributions, but it is simpler to gather additional synthetic data as we have done. + +Comparing models trained on Kubric to models trained with a broader data distribution, we see that the additional training data does not reliably improve performance in the surveillance dataset CroHD [46] or the driving dataset Driv- + +![](images/34c670fcf7adb627938c65d0cfba3dd20e63ed38658d79043f95913863e5bd64.jpg) +Figure 3. AllTracker (top right corner) delivers accurate multi-frame tracks at the throughput of an optical flow model. + +Table 3. Average Jaccard evaluation on TAP-Vid datasets. + +
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker362.953.569.566.363.1
AllTracker63.359.181.171.968.9
+ +Table 4. Occlusion accuracy evaluation on TAP-Vid datasets. + +
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker390.185.691.689.889.3
AllTracker90.190.392.892.891.5
+ +eTrack [1]. This suggests room for improvement, in terms of perhaps data balancing or model capacity. + +High-resolution performance Focusing on the comparison between our model and CoTracker3, we evaluate these models at higher resolutions in Table 2. Here we see that our model's performance reliably increases at higher resolutions, while CoTracker3 plateaus after $448 \times 768$ , leading to the surprising result that AllTracker-Tiny outperforms CoTracker3 at $768 \times 1024$ . Our model is also more memory efficient: with AllTracker we can process videos at $768 \times 1024$ (and produce 786,432 tracks at once) on a 40G A100 GPU, while with CoTracker3, despite only tracking sparse points, we encountered out-of-memory errors until eventually performing the tests on a 96G H100 GPU. Our model's relative memory efficiency is partly from using a spatial stride of 8 in the encoder, while CoTracker3 uses stride 4. + +Speed We compare the throughput of our model to other point trackers and optical flow models in Figure 3: our model runs at approximately the speed of optical flow methods, while achieving accuracies higher than the point trackers. + +Realtime inference Our model's sliding-window strategy can be run in "streaming" fashion real-time, at a penalty to accuracy: at $512 \times 512$ our model runs at 57.9 FPS and achieves $62.6\delta_{\mathrm{avg}}$ (vs. 66.1 normally); BootsTAPIR [13] has also published a realtime variant, which runs at 21.4 FPS and reaches $62.2\delta_{\mathrm{avg}}$ . + +Table 5. A transformer-based temporal module performs better than mixer-based or convolution-based variants. + +Table 6. Representing motion as long-range flow works better than instantaneous velocity, and positional embeddings do not help. + +
Time componentAcc.
Transformer [51]56.8
MLP-Mixer [48]56.1
ConvNext-1D [31]54.6
Conv-1D55.8
+ +
Motion representationAcc.
0 → t56.8
emb(0 → t)54.8
t-1 → t56.4
emb(t-1 → t)55.3
+ +Table 7. Hyperparameter search: It is best to use a ConvNeXt [31] backbone, 3 refinement blocks, and radius-4 correlations at 5 scales. + +
BackboneRefine blocksCorr. radiusCorr. scalesAcc.
ConvNeXt34556.8
ConvNeXt w/o pret.34554.9
BasicEncoder34556.1
ConvNeXt33556.5
ConvNeXt35555.5
ConvNeXt34455.5
ConvNeXt346Err.
ConvNeXt24555.7
ConvNeXt44556.1
+ +Additional metrics Evaluating AJ and occlusion accuracy on TAPVid datasets, we find that AllTracker obtains substantially better AJ than CoTracker3: 68.9 vs. 63.1 on average (see Table 3, and slightly better occlusion accuracy: 91.5 vs. 89.3 (see Table 4). We perform this evaluation only on the TAP-Vid datasets because we find that other datasets' visibility labels are not as reliable. + +Performance on flow benchmarks While traditional optical flow is not our focus, it is interesting to inspect the model's accuracy in this task. We submitted AllTracker to the official SINTEL test benchmark, yielding end-point error scores of 1.673 on "clean" and 3.244 on "final"; this is not as good as SEA-RAFT [54] (1.309 / 2.601) but comparable to GMFlow [57] (1.736 / 2.902). We note that it is common in optical flow literature to produce a different model for each benchmark (finetuned with a particular data mix and resolution), but for this test we simply use our main model as-is. Qualitatively, AllTracker's flow maps appear to be coarser than the ones from SEA-RAFT [54], suggesting that the model is underfitting on this task. We provide more optical flow results in the supplementary. + +Qualitative results We visualize the long-range flow maps for our model and selected baselines in Figure 4. Note that computing the dense flow map for the sparse methods takes multiple minutes, while our method (and the flow methods) produce it in under a second. We find that these flow visualizations reveal differences in the spatial coherence of the + +![](images/feff13cbeb6eaf1cbb87d5b03c3bdc7ff18e6cef5c9879d8b6ffeb892d8f6f47.jpg) +Figure 4. AllTracker produces accurate displacement fields across dozens of frames. Prior optical flow methods struggle to make correspondences across wide time gaps, while our model uses temporal priors to resolve the ambiguity; prior point trackers take multiple minutes to produce output at this density, and show splotchy pattern errors, while our method produces coherent output in less than a second. + +tracks: sparse point trackers from the past few years (PIPs++, LocoTrack, CoTracker3) show progressive improvement but include splotchy pattern errors in the motion fields. The optical flow methods (AccFlow, RAFT, SEA-RAFT) show better spatial coherence, but the estimates are unreliable when the displacements are too large. Our method produces motion fields that are both coherent and accurate. + +# 4.2. Ablations + +We verify our design choices in an exhaustive series of ablation studies. In these experiments, we train each model for 100,000 iterations with 2 GPUs, using Kubric, with 4 inference iterations per sample, at a learning rate of 4e-4, with videos of size $24 \times 384 \times 512$ and $56 \times 256 \times 256$ . We test at resolution $384 \times 512$ on a validation dataset that we create from BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], Horse10 [33], and RoboTAP [52], and report the mean $\delta_{\mathrm{avg}}$ across this set. We find that this truncated training and evaluation setup is crucial for enabling a thorough exploration of the model space. We note that the absolute values of these ablation experiments should not be compared to the main experiments, but can be compared to each other. + +Temporal module Prior work has explored different network architectures for learning temporal priors: PIPs [19] used an MLP-Mixer [48], $\mathrm{PIPs}++$ [59] and TAPIR [12] used 1D convolutions (which might reasonably be upgraded to 1D ConvNeXt layers [31]), and CoTracker [25] used a transformer [51]. In Table 5 we compare these in our setup and find that a transformer works best. We additionally note that the transformer option is most amenable to changing the window size, which is important for jointly training on optical flow (e.g., convolution kernels of size 3 cannot apply to an input of length 2). + +Motion representation Most point trackers use sinusoidal position embeddings of motion, and methods vary on whether displacements should be relative to the query frame or to an adjacent frame. Table 6 shows that displacement relative to the query frame (i.e., frame 0) works best. This choice has the additional benefit of merging the problem with long-range optical flow. Table 6 also shows that sinusoidal embeddings do not help. + +Backbone Most point trackers use a "BasicEncoder" backbone which originated from RAFT [47]. We show in Table 7 that a pre-trained ConvNeXt backbone performs better, while a ConvNeXt backbone trained from scratch is worse. + +Hyperparameters Our main model uses 3 refinement blocks, and radius-4 correlations computed at 5 scales. We demonstrate in Table 7 that the neighboring alternatives are worse. We note that using 6 scales produces a runtime error, as $256 \times 256$ input is already $1 \times 1$ after 5 scales. + +Huber vs. L1 CoTracker3 [24] and TAPIR [12] recommend a Huber loss instead of the L1 used in PIPs [19]; in our setup this is not helpful: 54.9 (Huber) vs. 56.8 (L1). + +Does frame ordering matter? We train a model with shuffled frames, to disentangle the benefits of joint multi-frame inference from the benefits of temporal continuity: we find that the impact varies from dataset to dataset, but shuffling is worse on average: 56.3 vs. 56.8. + +# 5. Conclusion and Limitations + +AllTracker is an approach to point tracking that treats the task as multi-frame optical flow. Our model delivers state-of-the-art performance on point tracking benchmarks, and produces dense output at high resolution, which past point trackers have struggled to do. AllTracker makes sparse point trackers (of similar speed and accuracy) redundant, but interestingly it does not make optical flow methods redundant: it does not outperform the state-of-the-art optical flow methods on optical flow estimation. The model appears to be underfitting on short-range motion estimation, suggesting that better models might be obtained with greater compute. A related limitation to address is the temporal window size: with bigger GPUs it may be possible to simply train and test with wider windows [24], and resolve longer occlusions. A broader area for future work is to add awareness about physical and common-sense constraints on motion, perhaps using 3D [37, 56] or more expressive model designs [18, 60]. + +Acknowledgments This work was supported by the TRI University 2.0 program, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship. + +# References + +[1] Arjun Balasingam, Joseph Chandler, Chenning Li, Zhoutong Zhang, and Hari Balakrishnan. Drivetrack: A benchmark for long-range point tracking in real-world videos. In CVPR, 2024. 6, 7, 8, 12, 13 +[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands context features. NeurIPS, 2024. 3 +[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and SMAL: Recovering the shape and motion of animals from video. In ACCV, 2018. 5, 6, 8, 12, 13 +[4] Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching in variational motion estimation. TPAMI, 33(3):500-513, 2010. 2 +[5] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In CVPR, 2024. 2 +[6] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seunglyong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In ECCV, 2025. 3, 5, 6 +[7] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Dense long-term motion estimation via statistical multi-step flow. In VISAPP, 2014. 2 +[8] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Multi-reference combinatorial strategy towards longer long-term dense motion estimation. Computer Vision and Image Understanding, 150:66-80, 2016. +[9] Tomas Crivelli, Pierre-Henri Conze, Philippe Robert, and Patrick Pérez. From optical flow to dense long term correspondences. In International Conference on Image Processing, 2012. 2 +[10] Ahmad Darkhalil, Rhodri Guerrier, Adam W Harley, and Dima Damen. Egopoints: Advancing point tracking for egocentric videos. arXiv:2412.04592, 2024. 5, 6, 12 +[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. In NeurIPS Datasets and Benchmarks, 2022. 3, 5, 6, 8, 12, 13 +[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In ICCV, 2023. 1, 3, 4, 8 +[13] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In ACCV, 2024. 3, 5, 6, 7 +[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ig, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, 2015. 2, 3, 5 +[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 2013. 3, 5 + +[16] James J Gibson. The perception of the visual world. 1950. 1, 2 +[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 3, 5 +[18] Adam Harley, Yang You, Yang Zheng, Xinglong Sun, Nikhil Raghuraman, Sheldon Liang, Wen-Hsuan Chu, Suya You, Achal Dave, Pavel Tokmakov, et al. TAG: Tracking at any granularity. 2024. 8 +[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 3, 5, 8 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2 +[21] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 2 +[22] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 2 +[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 5 +[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv:2410.11831, 2024. 1, 3, 4, 5, 6, 8, 12 +[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 1, 2, 3, 5, 6, 8 +[26] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The HCI benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In CVPR Workshops, 2016. 3, 5 +[27] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In CVPR, 2024. 1, 3, 13 +[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In CoRL, 2021. 5, 6, 12 +[29] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Feng Li, Tianhe Ren, Bohan Li, and Lei Zhang. TAPTRv2: Attention-based position update improves tracking any point. In NeurIPS, 2024. 3 +[30] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. TAPTR: Tracking any point with transformers as detection. In ECCV, 2024. 3 + +[31] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5, 7, 8 +[32] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 2 +[33] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 5, 6, 8, 12, 13 +[34] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 3, 5 +[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 3, 5 +[36] Michal Neoral, Jonás Šerych, and Jérí Matas. Mft: Long-term tracking of every pixel. In WACV, 2024. 2 +[37] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 1, 2, 5, 6, 8 +[38] Diederick C Niehorster. Optic flow: a history. i-Perception, 12(6):20416695211055766, 2021. 2 +[39] Stephan R Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In ICCV, 2017. 3, 5 +[40] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In CVPR, 2006. 1, 2 +[41] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3, 4 +[42] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In CVPR, 2018. 2, 13 +[43] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In CVPR, 2021. 3, 5 +[44] Xinglong Sun, Adam W Harley, and Leonidas J Guibas. Refining pre-trained motion models. In ICRA, 2024. 3 +[45] Narayanan Sundaram, Thomas Brox, and Kurt Keutzer. Dense point trajectories by GPU-accelerated large displacement optical flow. In ECCV, 2010. 2 +[46] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In CVPR, 2021. 5, 6, 8, 12, 13 +[47] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 4, 5, 6, 8, 13 +[48] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-mixer: An all-mlp architecture for vision. arXiv:2105.01601, 2021. 7, 8 + +[49] Carlo Tomasi and Takeo Kanade. Detection and tracking of point. IJCV, 9:137-154, 1991. 2 +[50] Marc Tournadre, Catherine Soladié, Nicolas Stoiber, and Pierre-Yves Richard. Dense trajectory fields: Consistent and efficient spatio-temporal pixel tracking. In ACCV, 2024. 1, 2 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 7, 8 +[52] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. RoboTAP: Tracking arbitrary points for few-shot visual imitation. In ICRA, 2024. 5, 6, 8, 12, 13 +[53] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual slam. In IROS, 2020. 3, 5 +[54] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, 2024. 2, 3, 5, 6, 7, 11, 13 +[55] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: Backward accumulation for long-range optical flow. In ICCV, 2023. 5, 6, 13 +[56] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In CVPR, 2024. 8 +[57] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 7, 13 +[58] Jia Xu, René Ranftl, and Vladlen Koltun. Accurate optical flow via direct cost volume processing. In CVPR, 2017. 2 +[59] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 1, 3, 5, 6, 8 +[60] Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi SM Sajjadi, Sarath Chandar, and Ross Goroshin. TAPNext: Tracking any point (TAP) as next token prediction. arXiv:2504.05579, 2025.8 \ No newline at end of file diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/images.zip b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c5da3173e42aaca3d87acb018b8334b57d049e03 --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c70356f6ffb07f8d6c82874bb12b6ae4af8522977cb66e8ee7b0b9c816b0130 +size 490763 diff --git a/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/layout.json b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ee97c022270404139aca0e1e2726e6d31e33d333 --- /dev/null +++ b/2025/AllTracker_ Efficient Dense Point Tracking at High Resolution/layout.json @@ -0,0 +1,8491 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "type": "text", + "content": "AllTracker: Efficient Dense Point Tracking at High Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": "Adam W. Harley" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Yang You" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Xinglong Sun" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Yang Zheng" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Nikhil Raghuraman" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Yunqi Gu" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Sheldon Liang" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Wen-Hsuan Chu" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Achal Dave" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Suya You" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Rares Ambrus" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Katerina Fragkiadaki" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "text", + "content": " Leonidas Guibas" + }, + { + "bbox": [ + 95, + 142, + 515, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "spans": [ + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "text", + "content": "Toyota Research Institute " + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 99, + 186, + 509, + 200 + ], + "type": "text", + "content": "Army Research Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "text", + "content": "We introduce AllTracker: a model that estimates long-range point tracks by way of estimating the flow field between a query frame and every other frame of a video. Unlike existing point tracking methods, our approach delivers high-resolution and dense (all-pixel) correspondence fields, which can be visualized as flow maps. Unlike existing optical flow methods, our approach corresponds one frame to hundreds of subsequent frames, rather than just the next frame. We develop a new architecture for this task, blending techniques from existing work in optical flow and point tracking: the model performs iterative inference on low-resolution grids of correspondence estimates, propagating information spatially via 2D convolution layers, and propagating information temporally via pixel-aligned attention layers. The model is fast and parameter-efficient (16 million parameters), and delivers state-of-the-art point tracking accuracy at high resolution (i.e., tracking " + }, + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "inline_equation", + "content": "768 \\times 1024" + }, + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "text", + "content": " pixels, on a 40G GPU). A benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets, and we find that doing so is crucial for top performance. We provide an extensive ablation study on our architecture details and training recipe, making it clear which details matter most. Our code and model weights are available: https://alltracker.github.io" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "Estimating the long-range trajectories of arbitrary points in the world, using sequences of 2D images as input, has been a concrete and competitive challenge in computer vision since at least 2006 [40]. The utility of optical flow (i.e., the instantaneous velocity of pixels [16]) toward this goal has long been obvious, yet it has remained challenging to upgrade flows into long-range tracks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Instantaneous flows can be \"chained\" into multi-frame tracks, by interpolating in one flow field at the endpoints of the previous flow field, but imperfect flows will accumulate drift, and such chains must also be carefully stopped at" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 225, + 544, + 404 + ], + "blocks": [ + { + "bbox": [ + 319, + 225, + 544, + 404 + ], + "lines": [ + { + "bbox": [ + 319, + 225, + 544, + 404 + ], + "spans": [ + { + "bbox": [ + 319, + 225, + 544, + 404 + ], + "type": "image", + "image_path": "e8645867ce1ed03f51ab9fbee66039f8841c0d6dc9e2de09081f41068717e510.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 456 + ], + "type": "text", + "content": "Figure 1. AllTracker estimates high-resolution optical flow between a \"query frame\" and every other frame of a video, using a sliding-window strategy. Point samples from these outputs can be interpreted as long-term point trajectories." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 567 + ], + "type": "text", + "content": "occlusions [40]. An attractive shortcut here is to directly compute the flow between a reference frame and each other frame, and thus track drift-free and across occlusions, but estimating \"long-range flow\" becomes increasingly difficult as the time interval widens between the reference frame and the target, due to increasing variation in perspective, illumination, and scene geometry." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "In light of these challenges, recent work has side-stepped optical flow and created a line of bespoke point trackers focused on learning multi-frame temporal priors for reducing drift and tracking through occlusions [12, 19, 24, 25, 59]. These works explicitly highlight that the limited temporal context in flow-based methods is a severe weakness, and have made substantial progress by addressing this. However, the proposed solutions add temporal awareness at the cost of spatial awareness, and only deliver tracks for sparse sets of points. Recent attempts at \"dense\" point tracking [27, 37, 50] are not as accurate as the latest sparse trackers, and struggle with high-resolution input." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5253" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "In this paper, we demonstrate that learnable multi-frame temporal priors can be built together with high-resolution spatial awareness, by casting point tracking as a multi-frame long-range optical flow problem, as illustrated in Figure 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 297, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 297, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 297, + 384 + ], + "type": "text", + "content": "Our design combines ideas from point tracking literature and optical flow literature, arriving at a composite architecture with broader capability. Following the trend in both areas, the heart of the model is a recurrent module that iteratively improves motion estimates [19, 47]. This module relies on information from spatial cross-correlations, giving it a strong inductive bias for feature-matching, which speeds up training. From point tracking literature, we borrow a per-pixel temporal module, to learn a motion prior and track through occlusions [19]. From optical flow literature, we borrow the idea of performing the majority of processing on a low-resolution grid, to allow fast spatial message-passing via 2D convolutions, and recover spatial precision at the end of the architecture with an upsampling layer [47, 54]. In relation to other point trackers, the key novelty of our approach is that we frame the problem as long-range dense flow, which makes sparse methods (of similar speed and accuracy) redundant. In relation to other optical flow models, the key novelty of our approach is that we solve a window of flow problems simultaneously, instead of frame-by-frame; the information shared within and across windows unlocks the capability to resolve flows across wide time intervals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 386, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 470 + ], + "type": "text", + "content": "A secondary benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets. We therefore build a mix of publicly-available datasets to support our model, spanning different time-ranges, annotation densities, and resolutions, and train with uniform sampling across this mix. We find that this strategy, paired with a long training schedule, is crucial for top performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 555 + ], + "type": "text", + "content": "In summary, our method blends techniques and datasets from optical flow literature and point tracking literature, and the resulting model is a state-of-the-art point tracker that operates in high resolution at full density. Inspired by CoTracker [25] and highlighting our capability for all-pixel tracking, we name our approach AllTracker. We will release our model, our data, and our code." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 142, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 142, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 142, + 582 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 713 + ], + "type": "text", + "content": "Optical flow The introduction of the concept of optic flow can be attributed to Gibson [16] (see Niehorster's review [38]). In computer vision, optical flow methods generally take two consecutive frames of video as input, and estimate the motion that relates the pixels of the first frame to the pixels of the second frame [21, 32]. The dominant technique, from classic approaches to current learning-based approaches, is to iterate on a solution: first, candidate correspondences are initialized (e.g., with zero-motion or constant velocity), then the \"costs\" of these correspondences are computed by measuring how" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "well the appearance features match, and this appearance cost is combined with smoothness terms or learned priors, and then better correspondences are sought in the neighborhood of the current solution [4, 22, 42, 47, 58]. Our work uses the same basic technique." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 556, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 300 + ], + "type": "text", + "content": "The current state of the art in optical flow estimation is SEA-RAFT [54]. SEA-RAFT first directly estimates a low-resolution flow field [14, 22], then computes appearance features with a ResNet-34 [20], and computes pixel-to-pixel correlations with these features [58], then iteratively estimates refinements to the flow via 2D convolutions that take flows and correlations as input, and finally upsamples the flow to full resolution via a pixel-shuffle layer [41]. Our work takes inspiration from SEA-RAFT (and its predecessor RAFT [47]) in the way that we iteratively refine low-resolution flows and upsample to full resolution, but we operate on multiple frames at a time (i.e., 16 frames instead of 2), and we add temporal attention layers to help information propagate across the time axis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 303, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 556, + 447 + ], + "type": "text", + "content": "Flow-based point tracking Optical flow is often used as a building block toward multi-frame tracking. The standard technique in this space is to \"chain\" flow vectors end-to-end to form longer tracks, and guard against drift by monitoring for occlusions [40, 45]. Recent techniques skip past occlusions by using multistep flows [7-9, 36], or attempt to track through occlusions using a learned temporal prior [5] (inspired by non-flow point trackers). These methods deliver dense multi-frame trajectories similar to our method, but first require an expensive pre-processing stage in which optical flows are estimated by a different model, while our method handles the full problem efficiently with a single model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 448, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 639 + ], + "type": "text", + "content": "We highlight two concurrent works which propose a flow-based method somewhat similar to our own: DTF [50] and DELTA [37]. Both iteratively estimate flows to relate a reference frame to other frames, but unlike our work, these papers propose special-purpose transformer-based methods in which global spatial message-passing is approximated by cross-attending to sparse \"anchor\" or \"centroid\" tokens. Our method simply uses 2D convolutions for this step, more like SEA-RAFT [54]. We note also that DTF and DELTA both struggle with high-resolution inputs (running out of memory on our 40G GPUs), despite DELTA using low-resolution inference and high-resolution upsampling, comparable to SEA-RAFT and our approach. In performance, DTF is not competitive with recent point trackers (while ours exceeds them); DELTA's main model requires depth input, but we compare against its 2D variant and outperform it." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 713 + ], + "type": "text", + "content": "Point trackers without flow Many classic methods track points directly, without relying on optical flow as a submodule [32, 49]. Harley et al. [19] recently introduced a new method in this space, with some components similar to flow models [47], but tracking points independently, and adding a multi-frame inference window which enabled the model" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5254" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 145 + ], + "type": "text", + "content": "to track through occlusions. That work, along with a new \"Tracking Any Point\" benchmark that appeared soon afterward [11], spurred much follow-up effort, involving multipoint context [24, 25], better initialization schemes [12], wider correlation context [2, 6], new transformer-based designs [29, 30], and post-hoc densification methods [27]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 149, + 297, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 149, + 297, + 401 + ], + "spans": [ + { + "bbox": [ + 57, + 149, + 297, + 401 + ], + "type": "text", + "content": "The current state of the art in point tracking is CoTracker3 [24]. CoTracker3 first computes feature maps for all frames, then initializes tracks for a sparse set of query points, then retrieves local correlations based on these tracks, and interleaves modules which pass information within tracks (temporally) and across tracks (spatially). The main novelty is in the spatial propagation: each point attends to a set of latent \"virtual points\" which learn a compressed representation of the video motions. While CoTracker3 delivers very accurate tracks, it requires users to pre-select a sparse set of points to track, and results vary based on how the queries are distributed [25]. CoTracker3 is more efficient than its predecessors [19, 25], but is still limited to tracking a few thousand points at a time. Our approach is simpler and more memory-efficient: we track on a low-resolution grid, where we achieve message-passing with simple 2D convolutions, and we recover spatial precision at the end of the architecture with a fast upsampling layer [41]. We note that this efficient upsampling technique was known well before the recent resurgence of point trackers [41], but perhaps its relevance was not fully appreciated." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 408, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 601 + ], + "type": "text", + "content": "Training data and self-supervision Current methods in optical flow and point tracking are data-driven, making data and supervision choices crucial. In optical flow, state-of-the-art methods rely on a combination of mostly synthetic datasets, including FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. In point tracking, the main training datasets are Kubric [17], FlyingThings++ [19], and PointOdyssey [59]. There is an art to creating a curriculum from diverse datasets to optimize downstream performance on particular benchmarks [54], but in our work we simply concatenate the datasets and shuffle the samples. We note that all previous point tracking methods did not use flow data, perhaps because most models do not support 2-frame inference or dense output; we therefore also provide experiments using Kubric alone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Motivated by the fact that the training data for point trackers is all synthetic, several groups have explored self-supervision methods based on boot-strapping pre-trained models using pseudo-labels computed on real videos [13, 24, 44]. However, these schemes offer only minor gains over the supervised weights where bootstrapping begins. In this work, we avoid any pseudo-labelling and simply add more synthetic data, capitalizing on our model's ability to accept supervision from optical flow data." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 242 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 242 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 242 + ], + "type": "image", + "image_path": "0cbe47e5a494155d34f69be48ce6a97f09f26d1b43b72db8d812b3ca206fc0e5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 249, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. AllTracker architecture. First, we compute feature maps for all frames, and copy the zeroth (query) feature map to every timestep, and compute multi-scale cost volumes. Then, we iterate a recurrent module, which references the query feature map and cost volume pyramid at each timestep, and estimates a low-resolution correspondence field, using interleaved 2D convolutions and pixel-aligned temporal attentions. The output of the RNN is upsampled into high-resolution optical flow maps, which relate all pixels of the zeroth frame to every other frame." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 373, + 386, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 373, + 386, + 387 + ], + "spans": [ + { + "bbox": [ + 314, + 373, + 386, + 387 + ], + "type": "text", + "content": "3. AllTracker" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 433 + ], + "type": "text", + "content": "Our model takes a video as input, along with a \"query\" index, specifying which frame's pixels to track, and it outputs full-timespan tracking for all of the pixels in that frame." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": "Concretely, we are given as input a video of shape " + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "inline_equation", + "content": "T, H, W, 3" + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": " is the number of frames, and " + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": " indicate the spatial resolution of each frame. We are also given an index " + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "inline_equation", + "content": "t \\in T" + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": ", indicating which frame has the pixels we need to track. Our final output is a tensor shaped " + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "inline_equation", + "content": "T, H, W, 4" + }, + { + "bbox": [ + 313, + 436, + 556, + 544 + ], + "type": "text", + "content": ": the first two channels are optical flow maps indicating the offset that takes each pixel in the query frame to its correspondence in every other frame, and the next two channels estimate visibility and confidence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": "A brief summary of the method is as follows. We operate in sliding window fashion across " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ", processing subsequences of length " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ", advancing at a stride of " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S / 2" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ". We begin by quickly computing low-resolution feature maps for the subsequence, yielding a feature volume with dimensions " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S, H / 8, W / 8, D" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": " is the channel dimension of the features. We then initialize a low-resolution tensor of outputs, shaped " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S, H / 8, W / 8, 4" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ", and iteratively revise this tensor using cross correlations and features across space and time as reference. Finally, we upsample the outputs, producing full resolution optical flow, visibility, and confidence, shaped " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S, H, W, 4" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ". We then advance our window by " + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "inline_equation", + "content": "S / 2" + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": ", using previous estimates as initialization, and repeat inference. Figure 2 shows the method in a diagram." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5255" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 123, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 123, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 123, + 85 + ], + "type": "text", + "content": "3.1. Encoding" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": "Our first stage encodes the input video frames into low-resolution feature maps. We achieve this with a ConvNeXt-Tiny [31] model (see Sec. 3.6 for implementation details). This encoder compresses a subsequence input of shape " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "S, H, W, 3" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": " to the shape " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "S, H/8, W/8, D" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": " is the embedding dimension (in our case " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "D = 256" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": " is the subsequence window length (in our case " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "S = 16" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": "). When the full video length " + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 89, + 297, + 196 + ], + "type": "text", + "content": " is not too large (e.g., at training time), we compute all of the feature maps in parallel." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 197, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 233 + ], + "type": "text", + "content": "We then isolate the feature map of the query frame and tile it to the length of the subsequence, so that we have a copy for each timestep, as illustrated in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 236, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 236, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 236, + 253 + ], + "type": "text", + "content": "3.2. Computing appearance similarity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 353 + ], + "type": "text", + "content": "Using the feature maps, we build a multi-scale 4D correlation volume, capturing appearance-based tracking cues. We implement this in two steps: first we convert each feature map into a feature pyramid, by average-pooling at multiple strides (in our case 5 strides: " + }, + { + "bbox": [ + 55, + 257, + 296, + 353 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8,16\\}" + }, + { + "bbox": [ + 55, + 257, + 296, + 353 + ], + "type": "text", + "content": "), and then we cross-correlate the query feature map with these pyramids. That is, for each feature vector in the query feature map, we compute its dot product across each timestep's pyramid." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": "The output of this step can be understood as a large collection of heatmaps: one heatmap per tracked pixel, per scale, per timestep. The model will index into these heatmaps to retrieve information on where correspondences lie. When a target's correspondence is visible at a certain timestep, we expect the respective heatmap to have a strong peak at the location of the correspondence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 168, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 168, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 168, + 455 + ], + "type": "text", + "content": "3.3. Track initialization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "text", + "content": "We next initialize estimates for tracking coordinates, visibility, and confidence, by creating a tensor with shape " + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "inline_equation", + "content": "H/8" + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "inline_equation", + "content": "W/8" + }, + { + "bbox": [ + 55, + 461, + 297, + 545 + ], + "type": "text", + "content": ", 4, where the coordinate channels are initialized with a 2D meshgrid, and visibility and confidence are initialized with zeros. If we are on a window beyond the first, we populate this tensor with the estimates that overlap, and copy forward the values from the last estimated timestep." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 171, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 171, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 171, + 564 + ], + "type": "text", + "content": "3.4. Iterative refinement" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 617 + ], + "type": "text", + "content": "The most important stage of our model is the iterative refinement stage, where we update all of the estimates. We describe the main points of the module here, but also provide extensive details (and a diagram) in the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "For each timestep, for each pixel in the query frame, we have: a feature vector " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " (length " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "), visibility and confidence estimates " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " (together length 2), a position estimate " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " (length 2), and a correlation pyramid " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{C}_1, \\mathbf{C}_2, \\ldots\\}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " shaped " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{H/8 \\times W/8, H/16 \\times W/16, \\ldots\\}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": ". We convert this data into a \"local\" representation suitable for convolutions, as follows. For the position data, we replace the absolute position estimates " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " with motion estimates, by subtracting the source" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "positions: " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{m} = \\mathbf{p} - \\mathbf{p}_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ". For each level of the correlation pyramid, we extract a small patch centered at " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", and flatten these patches into a single vector " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "L \\cdot (2R + 1)^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " is a radius parameter for the patches (e.g., 4) and " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " is the number of levels in the pyramid (e.g., 5). All together, we have " + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{f}, \\mathbf{v}, \\mathbf{c}, \\mathbf{m}, \\mathbf{q}" + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": " (per pixel, per timestep), which in our setup makes 665 channels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 556, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 556, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 556, + 311 + ], + "type": "text", + "content": "Inside the recurrent module, we process the input data using interleaved spatial and temporal blocks. Each spatial block is a 2D ConvNeXt block; each temporal block is a pixel-aligned transformer block. By pixel-aligned, we mean that attention only happens along the temporal axis (i.e., with cost quadratic in " + }, + { + "bbox": [ + 313, + 156, + 556, + 311 + ], + "type": "inline_equation", + "content": "S = 16" + }, + { + "bbox": [ + 313, + 156, + 556, + 311 + ], + "type": "text", + "content": "), and this is done for every pixel in parallel. We note that an important convenience of our design is that all of the tensors in this stage are aligned with the \"query\" frame; therefore pixel-aligned attention is attention between corresponding pixels. After the interleaved spatial and temporal blocks propagate tracking-related information across our window, we decode explicit revisions for visibility, confidence, and motion." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": "We apply visibility, confidence, and motion revisions via simple summation with the previous values " + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{new}} = \\mathbf{x}_{\\mathrm{old}} + \\delta \\mathbf{x}" + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\{\\mathbf{v}, \\mathbf{c}, \\mathbf{m}\\}" + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "inline_equation", + "content": "\\delta \\mathbf{x}" + }, + { + "bbox": [ + 313, + 312, + 555, + 359 + ], + "type": "text", + "content": " denotes an explicit revision produced by the model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 360, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 556, + 407 + ], + "type": "text", + "content": "We additionally decode weights for a pixel-shuffle upsampling step [41, 47]. We apply this upsampling to our visibility, confidence, and motion maps, bringing them from " + }, + { + "bbox": [ + 313, + 360, + 556, + 407 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 313, + 360, + 556, + 407 + ], + "type": "text", + "content": " resolution to full resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 407, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 455 + ], + "type": "text", + "content": "We iterate our recurrent \"refinement\" stage 4 times, sharing weights. At training time, we use the outputs from every refinement step for supervision, and at test time we only use the final iteration's output." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 460, + 408, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 460, + 408, + 473 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 408, + 473 + ], + "type": "text", + "content": "3.5. Model training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 478, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 556, + 538 + ], + "type": "text", + "content": "In the datasets which we use for training, we have supervision in the form of either optical flow or sparse point tracks, which we simply treat as trajectories of varying lengths. We use an " + }, + { + "bbox": [ + 313, + 478, + 556, + 538 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 478, + 556, + 538 + ], + "type": "text", + "content": " loss between estimated trajectories and ground truth, with a higher weight on loss for visible points:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 336, + 544, + 555, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 544, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 336, + 544, + 555, + 576 + ], + "type": "interline_equation", + "content": "L _ {\\text {t r a c k}} = \\alpha \\sum_ {k} ^ {K} \\gamma^ {K - k} \\left(\\mathbb {1} _ {\\text {o c c}} / 5 + \\mathbb {1} _ {\\text {v i s}}\\right) | | P _ {k} - \\hat {P} | | _ {1}, \\tag {1}", + "image_path": "a150918ace85e04dfb072500ee55dd288910368cc6802ef41ac51f5205a214d1.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "inline_equation", + "content": "P_{k}" + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": " is a set of estimated trajectories at refinement step " + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": " is the corresponding ground truth, " + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": " (set to 0.8) makes later refinement steps weigh more in the loss, and " + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": " is a balancing hyperparameter (set to 0.05)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": "We supervise our visibility and confidence maps with a binary cross entropy loss. We ask the visibility estimates to match ground truth binary labels: " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{vis}} = \\sum_k^K \\mathrm{BCE}(V, \\hat{V})" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": ". We ask the confidence estimates to reflect whether or not the corresponding position estimates are within 12 pixels of ground truth [12, 24]: " + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{conf}} = \\sum_k^K \\mathrm{BCE}(C, \\mathbb{1}[\\|X_k - \\hat{X}\\|_2 < 12])" + }, + { + "bbox": [ + 313, + 628, + 556, + 714 + ], + "type": "text", + "content": ". We apply these losses at every refinement step." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5256" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "To supervise the model with sparse annotations, we use the coordinates of the ground truth to sample our corresponding estimates, and apply the loss at these sparse locations. We use bilinear sampling for trajectories, and nearest-neighbor sampling for visibility and confidence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 187, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 187, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 187, + 152 + ], + "type": "text", + "content": "3.6. Implementation details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 297, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 297, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 297, + 179 + ], + "type": "text", + "content": "We provide extensive implementation details in the supplementary material, but describe the main details here." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "text", + "content": "Our CNN backbone is based on a pre-trained ConvNeXt-Tiny [31]. We use the first three \"blocks\" of this architecture, totaling 12.72 million parameters. We convert the third block from stride 2 to stride 1, by applying bicubic interpolation to the " + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "text", + "content": " stride-2 kernel to create a " + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "text", + "content": " stride-1 kernel, and re-scale these weights according to the change in area (i.e., scaling by " + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "inline_equation", + "content": "4/9" + }, + { + "bbox": [ + 54, + 180, + 297, + 263 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "text", + "content": "Our full model is 16.48 million parameters. We train it using 8 A100-40G GPUs. We train in two stages: first on Kubric [17] for 200,000 steps at a learning rate of " + }, + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "inline_equation", + "content": "5e - 4" + }, + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "text", + "content": ", and then on a mix of point tracking and optical flow datasets for 400,000 iterations at a learning rate of " + }, + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "inline_equation", + "content": "1e - 5" + }, + { + "bbox": [ + 54, + 264, + 296, + 361 + ], + "type": "text", + "content": ". We use the standard augmentations from prior work in point tracking [19] and optical flow [47], which consist of random shifting and scaling, color jitter, and square occlusions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 360, + 297, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 360, + 297, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 360, + 297, + 456 + ], + "type": "text", + "content": "Our optical flow datasets include FlyingChairs [14], FlyingThings3D [34], Monkaa [34], Driving [34], AutoFlow [43], Spring [35], VIPER [39], HD1K [26], KITTI [15], and TartanAir [53]. Our point tracking datasets include Kubric [17], DynamicReplica [23], and PointOdyssey [59]. We also follow Harley et al. [19] and compute point tracks from optical flows where possible, making FlyingThings++, Monkaa++, Driving++, and Spring++." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 456, + 297, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 456, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 54, + 456, + 297, + 540 + ], + "type": "text", + "content": "For trajectory estimates, we apply the model's additive revisions directly in pixel coordinates. We also supervise in pixel coordinates, which is a reason to scale the loss by the factor " + }, + { + "bbox": [ + 54, + 456, + 297, + 540 + ], + "type": "inline_equation", + "content": "\\alpha = 0.05" + }, + { + "bbox": [ + 54, + 456, + 297, + 540 + ], + "type": "text", + "content": ". For visibility and confidence estimates, we apply the model's revisions on logits (rather than on probabilities), but we apply a sigmoid to these values before they are input to the refinement block, to stabilize their range." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 137, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 618 + ], + "type": "text", + "content": "This section summarizes the key results, which concern the accuracy of our tracker compared to prior state of the art, on a diverse array of test benchmarks. We provide additional details for our experiments in the appendices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "Metrics and benchmarks Our main metric is " + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": ": an accuracy metric with max value 100, capturing how closely the estimated trajectory positions follow the ground truth positions. This is defined as the average of multiple " + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": " metrics, where " + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": " equals 100 when the estimate is within " + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": " pixels of ground truth, measuring at " + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": " resolution:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 114, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\delta_ {k} = 1 0 0 \\cdot \\mathbb {1} [ | | \\mathbf {p} - \\hat {\\mathbf {p}} | | _ {2} < k ]. \\tag {2}", + "image_path": "e1040b19f068f48951d7d7b023eaf18432474dca2fd3b485976d809b2f3be9b6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "Averaging " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "k\\in \\{1,2,4,8,16\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " yields " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ". We prioritize this over other considered metrics (e.g., L2 [19], PCK [19], AJ [11]) because we find it is both interpretable and robust to outliers. In certain benchmarks we also compute occlusion classification accuracy, and Average Jaccard (AJ) which mixes tracking accuracy with occlusion accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 144, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 556, + 348 + ], + "type": "text", + "content": "We evaluate on a total of nine publicly-available benchmarks with point annotations, covering a wide set of domains that include animals (BADJA [3], Horse10 [33]), YouTube videos (TAPVid-DAVIS [11], TAPVid-Kinetics [11]), surveillance camera recordings (CroHD [46]), egocentric recordings (EgoPoints [10]), and robotics data (RGB-Stacking [28], RoboTAP [52]). We trim the video lengths to a maximum of 600 frames, and on the larger datasets (CroHD, DriveTrack, EgoPoints, Horse10, Kinetics, RoboTAP) we only track points from the first available query frame, which we find gives similar results to using all possible queries and full video lengths. We argue that using a wide array of benchmarks helps flatten the effects of label noise, and gives a better sense of the overall performance in practice. We encourage future work to follow our example, but we note that our full evaluation is time-consuming (with some baselines requiring multiple days to finish their pass)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 350, + 556, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 648 + ], + "type": "text", + "content": "Baselines We benchmark a variety of recent point trackers and optical flow models in our tests, including the optical flow models RAFT [47] and SEA-RAFT [54], the long-term flow method AccFlow [55], and point trackers PIPs++ [59], LocoTrack [6], DELTA [37], BootsTAPIR [13]. We make close comparisons against DELTA [37] and CoTracker3 [24], which are the most recent state-of-the-art point trackers. DELTA is relevant because it is a concurrent work focused on dense tracking. CoTracker3 is relevant because it outperforms all past work. It has two variants: one trained exclusively on Kubric, and another version finetuned on pseudolabels produced by an ensemble of point trackers on a curated set of 15,000 real videos. This finetuned version of CoTracker3 outperforms all past work. We note that the performance of CoTracker3 varies depending on how the query points are grouped [25]. The paper suggests to run each query in a separate forward pass, while also adding \"support\" points around the query and a sparse grid covering the image, but in our multi-benchmark evaluation this would be prohibitively expensive (e.g., weeks). We simply give CoTracker3 all queries at once, supplemented by a sparse grid of points around the image. Comparing with values reported by the authors (reproduced in the supplementary), we find that our setup over-estimates CoTracker3's accuracy, but makes our evaluation tractable." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 651, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 555, + 676 + ], + "type": "text", + "content": "AllTracker variants In addition to our main model, which is 16.48 million parameters and trained on a mix of datasets," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "1The CoTracker3 repo also includes an \"offline\" variant, but it is not consistently better than the standard version, and in any case its memory consumption is outside our compute budget." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5257" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 109, + 542, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Comparison against recent point trackers and optical flow models, across nine datasets. We evaluate " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " (higher is better), using an input resolution of " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "384\\times 512" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ". The benchmarks are BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], EgoPoints [10], Horse10 [33], TAPVid-Kinetics [11], RGB-Stacking [28], and RoboTAP [52]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 109, + 542, + 285 + ], + "lines": [ + { + "bbox": [ + 66, + 109, + 542, + 285 + ], + "spans": [ + { + "bbox": [ + 66, + 109, + 542, + 285 + ], + "type": "table", + "html": "
MethodParams.TrainingBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
RAFT [47]5.26Flow mix23.729.348.544.841.027.864.382.872.248.3
SEA-RAFT [54]19.66Flow mix23.921.948.749.444.033.164.385.767.648.7
AccFlow [55]11.76Flow mix10.322.223.526.44.012.138.863.257.928.7
PIPs++ [59]17.57PointOdyssey34.127.562.551.338.521.464.270.473.449.3
LocoTrack [6]11.52Kubric41.443.168.066.558.448.970.080.376.961.5
BootsTAPIR [13]54.70Kubric+15M42.734.967.966.956.848.870.681.078.260.9
DELTA [37]59.17Kubric44.642.975.367.840.341.866.583.074.859.7
CoTracker2 [25]45.43Kubric40.031.770.967.843.233.965.873.47355.5
CoTracker3-Kub [24]25.39Kubric47.548.977.469.858.047.570.683.477.264.5
CoTracker3 [24]25.39Kubric+15k48.344.577.169.860.447.171.884.281.665.0
AllTracker-Tiny-Kub6.29Kubric45.439.673.765.155.945.270.686.179.362.3
AllTracker-Tiny6.29Kubric+mix47.539.874.363.958.345.571.588.180.763.3
AllTracker-Kub16.48Kubric46.442.375.266.160.349.071.390.182.264.8
AllTracker16.48Kubric+mix51.544.076.365.862.549.072.390.083.466.1
", + "image_path": "58631c245c45923755bfa7568d9532fcaa53f8c99e62a5628aa06a21f2ce6ea9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 323, + 539, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": "Table 2. High-resolution comparison between our model and CoTracker3 [24], which is the previous state of the art. We evaluate " + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": " (higher is better) on nine point tracking benchmarks, at resolutions " + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "inline_equation", + "content": "448 \\times 768" + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "inline_equation", + "content": "768 \\times 1024" + }, + { + "bbox": [ + 55, + 297, + 555, + 319 + ], + "type": "text", + "content": ". Parameter counts are in millions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 323, + 539, + 483 + ], + "lines": [ + { + "bbox": [ + 69, + 323, + 539, + 483 + ], + "spans": [ + { + "bbox": [ + 69, + 323, + 539, + 483 + ], + "type": "table", + "html": "
ModelParams.ResolutionBad.Cro.Dav.Dri.Ego.Hor.Kin.Rgb.Rob.Avg.
CoTracker3-Kub [24]25.39448 × 76849.759.978.870.257.450.670.481.176.666.1
CoTracker3 [24]25.39448 × 76850.757.979.570.861.351.172.382.881.367.5
AllTracker-Tiny-Kub6.29448 × 76847.748.076.367.857.348.371.284.879.064.5
AllTracker-Tiny6.29448 × 76849.748.877.067.860.749.272.388.180.066.0
AllTracker-Kub16.48448 × 76849.851.677.668.361.352.671.990.482.067.3
AllTracker16.48448 × 76852.551.278.868.364.252.573.090.683.468.3
CoTracker3-Kub [24]25.39768 × 102447.862.077.072.051.946.467.076.274.563.9
CoTracker3 [24]25.39768 × 102449.864.379.672.458.448.571.177.980.266.9
AllTracker-Tiny-Kub6.29768 × 102448.456.978.571.255.149.271.180.878.165.5
AllTracker-Tiny6.29768 × 102451.657.079.171.159.250.772.487.479.267.5
AllTracker-Kub16.48768 × 102451.760.279.871.259.154.471.789.281.768.8
AllTracker16.48768 × 102453.653.480.672.364.354.673.190.683.269.5
", + "image_path": "65532211527a6b1dc486f6d00a1b04bb74b94b9c8c5b6e649ce0429dfaa81dff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 564 + ], + "type": "text", + "content": "we evaluate a \"tiny\" version, which is trained similarly to the main model but uses a cheaper CNN backbone and totals only 6.29 million parameters (see details in supplemental). For both versions, we additionally report the results with models trained exclusively on Kubric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 137, + 586 + ], + "type": "text", + "content": "4.1. Main results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 593, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 664 + ], + "type": "text", + "content": "We evaluate our method's ability to track arbitrary points in diverse videos of various lengths, and compare against the state of the art in Table 1. For this evaluation we resize all input videos to " + }, + { + "bbox": [ + 55, + 593, + 295, + 664 + ], + "type": "inline_equation", + "content": "384 \\times 512" + }, + { + "bbox": [ + 55, + 593, + 295, + 664 + ], + "type": "text", + "content": ". We note that DELTA (the only other dense point tracker) often runs out of memory when testing on a 40G GPU, so we use a 96G GPU to evaluate it." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "As shown in the table, even our weakest variant, AllTracker-Tiny-Kub, is competitive with state of the art on many datasets, despite being only 6.29M parameters and training on a single synthetic dataset. Our full AllTracker" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 660 + ], + "type": "text", + "content": "model outperforms all other models on average, with the closest competitor being the variant of CoTracker3 which relied on an expensive bootstrapping scheme (noted by “+15k” in the table). This CoTracker3 model wins on three datasets (CroHD, Davis, DriveTrack) and our model wins on the rest and wins on average (66.1 vs. 65.0). Our method has a substantial gain over past work on the RGB-Stacking benchmark, which has many query points inside textureless regions; our model's good performance here suggests that it is able to incorporate spatial context from a wider area than the previous models. We note that CoTracker3's bootstrapping technique could be combined with our contributions, but it is simpler to gather additional synthetic data as we have done." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Comparing models trained on Kubric to models trained with a broader data distribution, we see that the additional training data does not reliably improve performance in the surveillance dataset CroHD [46] or the driving dataset Driv-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5258" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 281, + 205 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 281, + 205 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 281, + 205 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 281, + 205 + ], + "type": "image", + "image_path": "34c670fcf7adb627938c65d0cfba3dd20e63ed38658d79043f95913863e5bd64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 297, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 297, + 232 + ], + "type": "text", + "content": "Figure 3. AllTracker (top right corner) delivers accurate multi-frame tracks at the throughput of an optical flow model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 261, + 275, + 307 + ], + "blocks": [ + { + "bbox": [ + 69, + 245, + 279, + 256 + ], + "lines": [ + { + "bbox": [ + 69, + 245, + 279, + 256 + ], + "spans": [ + { + "bbox": [ + 69, + 245, + 279, + 256 + ], + "type": "text", + "content": "Table 3. Average Jaccard evaluation on TAP-Vid datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 261, + 275, + 307 + ], + "lines": [ + { + "bbox": [ + 74, + 261, + 275, + 307 + ], + "spans": [ + { + "bbox": [ + 74, + 261, + 275, + 307 + ], + "type": "table", + "html": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker362.953.569.566.363.1
AllTracker63.359.181.171.968.9
", + "image_path": "9d16c6fb4521441705bfe48097e863f07c9ffa4057391aa63d4793a6476197ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 75, + 331, + 275, + 376 + ], + "blocks": [ + { + "bbox": [ + 63, + 314, + 285, + 325 + ], + "lines": [ + { + "bbox": [ + 63, + 314, + 285, + 325 + ], + "spans": [ + { + "bbox": [ + 63, + 314, + 285, + 325 + ], + "type": "text", + "content": "Table 4. Occlusion accuracy evaluation on TAP-Vid datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 331, + 275, + 376 + ], + "lines": [ + { + "bbox": [ + 75, + 331, + 275, + 376 + ], + "spans": [ + { + "bbox": [ + 75, + 331, + 275, + 376 + ], + "type": "table", + "html": "
ModelDav.Kin.Rgb.Rob.Avg.
CoTracker390.185.691.689.889.3
AllTracker90.190.392.892.891.5
", + "image_path": "2269b1f1e434a36af212032cb3ba2cce70056897781d59f46672d634cd22a433.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 295, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 417 + ], + "type": "text", + "content": "eTrack [1]. This suggests room for improvement, in terms of perhaps data balancing or model capacity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "text", + "content": "High-resolution performance Focusing on the comparison between our model and CoTracker3, we evaluate these models at higher resolutions in Table 2. Here we see that our model's performance reliably increases at higher resolutions, while CoTracker3 plateaus after " + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "inline_equation", + "content": "448 \\times 768" + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "text", + "content": ", leading to the surprising result that AllTracker-Tiny outperforms CoTracker3 at " + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "inline_equation", + "content": "768 \\times 1024" + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "text", + "content": ". Our model is also more memory efficient: with AllTracker we can process videos at " + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "inline_equation", + "content": "768 \\times 1024" + }, + { + "bbox": [ + 54, + 419, + 297, + 588 + ], + "type": "text", + "content": " (and produce 786,432 tracks at once) on a 40G A100 GPU, while with CoTracker3, despite only tracking sparse points, we encountered out-of-memory errors until eventually performing the tests on a 96G H100 GPU. Our model's relative memory efficiency is partly from using a spatial stride of 8 in the encoder, while CoTracker3 uses stride 4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 639 + ], + "type": "text", + "content": "Speed We compare the throughput of our model to other point trackers and optical flow models in Figure 3: our model runs at approximately the speed of optical flow methods, while achieving accuracies higher than the point trackers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "Realtime inference Our model's sliding-window strategy can be run in \"streaming\" fashion real-time, at a penalty to accuracy: at " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": " our model runs at 57.9 FPS and achieves " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "62.6\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": " (vs. 66.1 normally); BootsTAPIR [13] has also published a realtime variant, which runs at 21.4 FPS and reaches " + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "inline_equation", + "content": "62.2\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 641, + 297, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 71, + 425, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 425, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 425, + 114 + ], + "type": "text", + "content": "Table 5. A transformer-based temporal module performs better than mixer-based or convolution-based variants." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 430, + 71, + 552, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 71, + 552, + 115 + ], + "spans": [ + { + "bbox": [ + 430, + 71, + 552, + 115 + ], + "type": "text", + "content": "Table 6. Representing motion as long-range flow works better than instantaneous velocity, and positional embeddings do not help." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 321, + 120, + 417, + 187 + ], + "blocks": [ + { + "bbox": [ + 321, + 120, + 417, + 187 + ], + "lines": [ + { + "bbox": [ + 321, + 120, + 417, + 187 + ], + "spans": [ + { + "bbox": [ + 321, + 120, + 417, + 187 + ], + "type": "table", + "html": "
Time componentAcc.
Transformer [51]56.8
MLP-Mixer [48]56.1
ConvNext-1D [31]54.6
Conv-1D55.8
", + "image_path": "4a0be16cbca0277bdfac0ea7b95c2b8b91f7378623a71aadd8978bb163e3251c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 437, + 120, + 545, + 187 + ], + "blocks": [ + { + "bbox": [ + 437, + 120, + 545, + 187 + ], + "lines": [ + { + "bbox": [ + 437, + 120, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 437, + 120, + 545, + 187 + ], + "type": "table", + "html": "
Motion representationAcc.
0 → t56.8
emb(0 → t)54.8
t-1 → t56.4
emb(t-1 → t)55.3
", + "image_path": "426281266c3cc324e2fddf2693eca4502c0ac0be096eafc16ff306d211e4dc53.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 220, + 546, + 369 + ], + "blocks": [ + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 216 + ], + "type": "text", + "content": "Table 7. Hyperparameter search: It is best to use a ConvNeXt [31] backbone, 3 refinement blocks, and radius-4 correlations at 5 scales." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 220, + 546, + 369 + ], + "lines": [ + { + "bbox": [ + 321, + 220, + 546, + 369 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 546, + 369 + ], + "type": "table", + "html": "
BackboneRefine blocksCorr. radiusCorr. scalesAcc.
ConvNeXt34556.8
ConvNeXt w/o pret.34554.9
BasicEncoder34556.1
ConvNeXt33556.5
ConvNeXt35555.5
ConvNeXt34455.5
ConvNeXt346Err.
ConvNeXt24555.7
ConvNeXt44556.1
", + "image_path": "ec01aef321cb2890d73e5bcea765f8ff981a3c7591e515993fb1859feeb7f2ac.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 468 + ], + "type": "text", + "content": "Additional metrics Evaluating AJ and occlusion accuracy on TAPVid datasets, we find that AllTracker obtains substantially better AJ than CoTracker3: 68.9 vs. 63.1 on average (see Table 3, and slightly better occlusion accuracy: 91.5 vs. 89.3 (see Table 4). We perform this evaluation only on the TAP-Vid datasets because we find that other datasets' visibility labels are not as reliable." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 639 + ], + "type": "text", + "content": "Performance on flow benchmarks While traditional optical flow is not our focus, it is interesting to inspect the model's accuracy in this task. We submitted AllTracker to the official SINTEL test benchmark, yielding end-point error scores of 1.673 on \"clean\" and 3.244 on \"final\"; this is not as good as SEA-RAFT [54] (1.309 / 2.601) but comparable to GMFlow [57] (1.736 / 2.902). We note that it is common in optical flow literature to produce a different model for each benchmark (finetuned with a particular data mix and resolution), but for this test we simply use our main model as-is. Qualitatively, AllTracker's flow maps appear to be coarser than the ones from SEA-RAFT [54], suggesting that the model is underfitting on this task. We provide more optical flow results in the supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "content": "Qualitative results We visualize the long-range flow maps for our model and selected baselines in Figure 4. Note that computing the dense flow map for the sparse methods takes multiple minutes, while our method (and the flow methods) produce it in under a second. We find that these flow visualizations reveal differences in the spatial coherence of the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5259" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 166 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 166 + ], + "type": "image", + "image_path": "feff13cbeb6eaf1cbb87d5b03c3bdc7ff18e6cef5c9879d8b6ffeb892d8f6f47.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 173, + 555, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 173, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 555, + 207 + ], + "type": "text", + "content": "Figure 4. AllTracker produces accurate displacement fields across dozens of frames. Prior optical flow methods struggle to make correspondences across wide time gaps, while our model uses temporal priors to resolve the ambiguity; prior point trackers take multiple minutes to produce output at this density, and show splotchy pattern errors, while our method produces coherent output in less than a second." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 297, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 297, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 297, + 301 + ], + "type": "text", + "content": "tracks: sparse point trackers from the past few years (PIPs++, LocoTrack, CoTracker3) show progressive improvement but include splotchy pattern errors in the motion fields. The optical flow methods (AccFlow, RAFT, SEA-RAFT) show better spatial coherence, but the estimates are unreliable when the displacements are too large. Our method produces motion fields that are both coherent and accurate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 124, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 124, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 124, + 319 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "content": "We verify our design choices in an exhaustive series of ablation studies. In these experiments, we train each model for 100,000 iterations with 2 GPUs, using Kubric, with 4 inference iterations per sample, at a learning rate of 4e-4, with videos of size " + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "inline_equation", + "content": "24 \\times 384 \\times 512" + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "inline_equation", + "content": "56 \\times 256 \\times 256" + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "content": ". We test at resolution " + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "inline_equation", + "content": "384 \\times 512" + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "content": " on a validation dataset that we create from BADJA [3], CroHD [46], TAPVid-DAVIS [11], DriveTrack [1], Horse10 [33], and RoboTAP [52], and report the mean " + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathrm{avg}}" + }, + { + "bbox": [ + 55, + 326, + 296, + 481 + ], + "type": "text", + "content": " across this set. We find that this truncated training and evaluation setup is crucial for enabling a thorough exploration of the model space. We note that the absolute values of these ablation experiments should not be compared to the main experiments, but can be compared to each other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": "Temporal module Prior work has explored different network architectures for learning temporal priors: PIPs [19] used an MLP-Mixer [48], " + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{PIPs}++" + }, + { + "bbox": [ + 55, + 483, + 296, + 615 + ], + "type": "text", + "content": " [59] and TAPIR [12] used 1D convolutions (which might reasonably be upgraded to 1D ConvNeXt layers [31]), and CoTracker [25] used a transformer [51]. In Table 5 we compare these in our setup and find that a transformer works best. We additionally note that the transformer option is most amenable to changing the window size, which is important for jointly training on optical flow (e.g., convolution kernels of size 3 cannot apply to an input of length 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "Motion representation Most point trackers use sinusoidal position embeddings of motion, and methods vary on whether displacements should be relative to the query frame or to an adjacent frame. Table 6 shows that displacement relative to the query frame (i.e., frame 0) works best. This choice has the additional benefit of merging the problem with long-range optical flow. Table 6 also shows that sinusoidal embeddings do not help." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 265 + ], + "type": "text", + "content": "Backbone Most point trackers use a \"BasicEncoder\" backbone which originated from RAFT [47]. We show in Table 7 that a pre-trained ConvNeXt backbone performs better, while a ConvNeXt backbone trained from scratch is worse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": "Hyperparameters Our main model uses 3 refinement blocks, and radius-4 correlations computed at 5 scales. We demonstrate in Table 7 that the neighboring alternatives are worse. We note that using 6 scales produces a runtime error, as " + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": " input is already " + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 269, + 555, + 329 + ], + "type": "text", + "content": " after 5 scales." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 369 + ], + "type": "text", + "content": "Huber vs. L1 CoTracker3 [24] and TAPIR [12] recommend a Huber loss instead of the L1 used in PIPs [19]; in our setup this is not helpful: 54.9 (Huber) vs. 56.8 (L1)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 556, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 433 + ], + "type": "text", + "content": "Does frame ordering matter? We train a model with shuffled frames, to disentangle the benefits of joint multi-frame inference from the benefits of temporal continuity: we find that the impact varies from dataset to dataset, but shuffling is worse on average: 56.3 vs. 56.8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 471, + 460 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 470, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 556, + 673 + ], + "type": "text", + "content": "AllTracker is an approach to point tracking that treats the task as multi-frame optical flow. Our model delivers state-of-the-art performance on point tracking benchmarks, and produces dense output at high resolution, which past point trackers have struggled to do. AllTracker makes sparse point trackers (of similar speed and accuracy) redundant, but interestingly it does not make optical flow methods redundant: it does not outperform the state-of-the-art optical flow methods on optical flow estimation. The model appears to be underfitting on short-range motion estimation, suggesting that better models might be obtained with greater compute. A related limitation to address is the temporal window size: with bigger GPUs it may be possible to simply train and test with wider windows [24], and resolve longer occlusions. A broader area for future work is to add awareness about physical and common-sense constraints on motion, perhaps using 3D [37, 56] or more expressive model designs [18, 60]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the TRI University 2.0 program, ARL grant W911NF-21-2-0104, and a Vannevar Bush Faculty Fellowship." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5260" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "text", + "content": "[1] Arjun Balasingam, Joseph Chandler, Chenning Li, Zhoutong Zhang, and Hari Balakrishnan. Drivetrack: A benchmark for long-range point tracking in real-world videos. In CVPR, 2024. 6, 7, 8, 12, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "text", + "content": "[2] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands context features. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 224 + ], + "type": "text", + "content": "[3] Benjamin Biggs, Thomas Roddick, Andrew Fitzgibbon, and Roberto Cipolla. Creatures great and SMAL: Recovering the shape and motion of animals from video. In ACCV, 2018. 5, 6, 8, 12, 13" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 296, + 258 + ], + "type": "text", + "content": "[4] Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching in variational motion estimation. TPAMI, 33(3):500-513, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[5] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 296, + 327 + ], + "type": "text", + "content": "[6] Seokju Cho, Jiahui Huang, Jisu Nam, Honggyu An, Seunglyong Kim, and Joon-Young Lee. Local all-pair correspondence for point tracking. In ECCV, 2025. 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 294, + 362 + ], + "type": "text", + "content": "[7] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Dense long-term motion estimation via statistical multi-step flow. In VISAPP, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "text", + "content": "[8] Pierre-Henri Conze, Philippe Robert, Tomas Crivelli, and Luce Morin. Multi-reference combinatorial strategy towards longer long-term dense motion estimation. Computer Vision and Image Understanding, 150:66-80, 2016." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 410, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 452 + ], + "type": "text", + "content": "[9] Tomas Crivelli, Pierre-Henri Conze, Philippe Robert, and Patrick Pérez. From optical flow to dense long term correspondences. In International Conference on Image Processing, 2012. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 454, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 296, + 486 + ], + "type": "text", + "content": "[10] Ahmad Darkhalil, Rhodri Guerrier, Adam W Harley, and Dima Damen. Egopoints: Advancing point tracking for egocentric videos. arXiv:2412.04592, 2024. 5, 6, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 296, + 542 + ], + "type": "text", + "content": "[11] Carl Doersch, Ankush Gupta, Larisa Markeeeva, Adrià Recasens, Lucas Smaira, Yusuf Aytar, João Carreira, Andrew Zisserman, and Yi Yang. TAP-Vid: A benchmark for tracking any point in a video. In NeurIPS Datasets and Benchmarks, 2022. 3, 5, 6, 8, 12, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "text", + "content": "[12] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. TAPIR: Tracking any point with per-frame initialization and temporal refinement. In ICCV, 2023. 1, 3, 4, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 634 + ], + "type": "text", + "content": "[13] Carl Doersch, Pauline Luc, Yi Yang, Dilara Gokay, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ignacio Rocco, Ross Goroshin, João Carreira, et al. Bootstrap: Bootstrapped training for tracking-any-point. In ACCV, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 678 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Philipp Fischer, Eddy Ig, Philip Hausser, Caner Hazirbas, Vladimir Golkov, Patrick Van Der Smagt, Daniel Cremers, and Thomas Brox. Flownet: Learning optical flow with convolutional networks. In ICCV, 2015. 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "text", + "content": "[15] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. International Journal of Robotics Research, 2013. 3, 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 94 + ], + "type": "text", + "content": "[16] James J Gibson. The perception of the visual world. 1950. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "type": "text", + "content": "[17] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 143, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 555, + 186 + ], + "type": "text", + "content": "[18] Adam Harley, Yang You, Yang Zheng, Xinglong Sun, Nikhil Raghuraman, Sheldon Liang, Wen-Hsuan Chu, Suya You, Achal Dave, Pavel Tokmakov, et al. TAG: Tracking at any granularity. 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 189, + 555, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 555, + 221 + ], + "type": "text", + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In ECCV, 2022. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 555, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 555, + 255 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 258, + 553, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 279 + ], + "type": "text", + "content": "[21] Berthold KP Horn and Brian G Schunck. Determining optical flow. Artificial intelligence, 17(1-3):185-203, 1981. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 281, + 555, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 281, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 281, + 555, + 324 + ], + "type": "text", + "content": "[22] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 370 + ], + "type": "text", + "content": "[23] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. CVPR, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 373, + 555, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 555, + 427 + ], + "type": "text", + "content": "[24] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker3: Simpler and better point tracking by pseudolabelling real videos. arXiv:2410.11831, 2024. 1, 3, 4, 5, 6, 8, 12" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 555, + 473 + ], + "type": "text", + "content": "[25] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker: It is better to track together. In ECCV, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 475, + 555, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 541 + ], + "type": "text", + "content": "[26] Daniel Kondermann, Rahul Nair, Katrin Honauer, Karsten Krispin, Jonas Andrulis, Alexander Brock, Burkhard Gussefeld, Mohsen Rahimimoghaddam, Sabine Hofmann, Claus Brenner, et al. The HCI benchmark suite: Stereo and flow ground truth with uncertainties for urban autonomous driving. In CVPR Workshops, 2016. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 555, + 575 + ], + "type": "text", + "content": "[27] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: connecting the dots. In CVPR, 2024. 1, 3, 13" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 555, + 632 + ], + "type": "text", + "content": "[28] Alex X Lee, Coline Manon Devin, Yuxiang Zhou, Thomas Lampe, Konstantinos Bousmalis, Jost Tobias Springenberg, Arunkumar Byravan, Abbas Abdelmaleki, Nimrod Gileadi, David Khosid, et al. Beyond pick-and-place: Tackling robotic stacking of diverse shapes. In CoRL, 2021. 5, 6, 12" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 555, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 677 + ], + "type": "text", + "content": "[29] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Feng Li, Tianhe Ren, Bohan Li, and Lei Zhang. TAPTRv2: Attention-based position update improves tracking any point. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[30] Hongyang Li, Hao Zhang, Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, and Lei Zhang. TAPTR: Tracking any point with transformers as detection. In ECCV, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5261" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[31] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, 2022. 4, 5, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "text", + "content": "[32] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 184 + ], + "type": "text", + "content": "[33] Alexander Mathis, Thomas Biasi, Steffen Schneider, Mert Yuksekgonul, Byron Rogers, Matthias Bethge, and Mackenzie W Mathis. Pretraining boosts out-of-domain robustness for pose estimation. In WACV, 2021. 5, 6, 8, 12, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[34] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[35] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In CVPR, 2023. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 297 + ], + "type": "text", + "content": "[36] Michal Neoral, Jonás Šerych, and Jérí Matas. Mft: Long-term tracking of every pixel. In WACV, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 342 + ], + "type": "text", + "content": "[37] Tuan Duc Ngo, Peiye Zhuang, Chuang Gan, Evangelos Kalogerakis, Sergey Tulyakov, Hsin-Ying Lee, and Chaoyang Wang. DELTA: Dense efficient long-range 3d tracking for any video. In ICLR, 2025. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 364 + ], + "type": "text", + "content": "[38] Diederick C Niehorster. Optic flow: a history. i-Perception, 12(6):20416695211055766, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "type": "text", + "content": "[39] Stephan R Richter, Zeeshan Hayden, and Vladlen Koltun. Playing for benchmarks. In ICCV, 2017. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 296, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 410 + ], + "type": "text", + "content": "[40] P. Sand and S. Teller. Particle video: Long-range motion estimation using point trajectories. In CVPR, 2006. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 411, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 465 + ], + "type": "text", + "content": "[41] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 296, + 499 + ], + "type": "text", + "content": "[42] Deqing Sun, Xiaodong Yang, Ming-Yu Liu, and Jan Kautz. PWC-Net: CNNs for optical flow using pyramid, warping, and cost volume. In CVPR, 2018. 2, 13" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 544 + ], + "type": "text", + "content": "[43] Deqing Sun, Daniel Vlasic, Charles Herrmann, Varun Jampani, Michael Krainin, Huiwen Chang, Ramin Zabih, William T Freeman, and Ce Liu. Autoflow: Learning a better training set for optical flow. In CVPR, 2021. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 545, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 567 + ], + "type": "text", + "content": "[44] Xinglong Sun, Adam W Harley, and Leonidas J Guibas. Refining pre-trained motion models. In ICRA, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "type": "text", + "content": "[45] Narayanan Sundaram, Thomas Brox, and Kurt Keutzer. Dense point trajectories by GPU-accelerated large displacement optical flow. In ECCV, 2010. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 602, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 634 + ], + "type": "text", + "content": "[46] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In CVPR, 2021. 5, 6, 8, 12, 13" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 657 + ], + "type": "text", + "content": "[47] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 2, 4, 5, 6, 8, 13" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[48] Ilya O. Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, and Alexey Dosovitskiy. MLP-mixer: An all-mlp architecture for vision. arXiv:2105.01601, 2021. 7, 8" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 533 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 555, + 95 + ], + "type": "text", + "content": "[49] Carlo Tomasi and Takeo Kanade. Detection and tracking of point. IJCV, 9:137-154, 1991. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[50] Marc Tournadre, Catherine Soladié, Nicolas Stoiber, and Pierre-Yves Richard. Dense trajectory fields: Consistent and efficient spatio-temporal pixel tracking. In ACCV, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 174 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[52] Mel Vecerik, Carl Doersch, Yi Yang, Todor Davchev, Yusuf Aytar, Guangyao Zhou, Raia Hadsell, Lourdes Agapito, and Jon Scholz. RoboTAP: Tracking arbitrary points for few-shot visual imitation. In ICRA, 2024. 5, 6, 8, 12, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "text", + "content": "[53] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual slam. In IROS, 2020. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 264, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 297 + ], + "type": "text", + "content": "[54] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In ECCV, 2024. 2, 3, 5, 6, 7, 11, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 342 + ], + "type": "text", + "content": "[55] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: Backward accumulation for long-range optical flow. In ICCV, 2023. 5, 6, 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 343, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 376 + ], + "type": "text", + "content": "[56] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In CVPR, 2024. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 377, + 555, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 410 + ], + "type": "text", + "content": "[57] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, and Dacheng Tao. Gmflow: Learning optical flow via global matching. In CVPR, 2022. 7, 13" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 411, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 555, + 432 + ], + "type": "text", + "content": "[58] Jia Xu, René Ranftl, and Vladlen Koltun. Accurate optical flow via direct cost volume processing. In CVPR, 2017. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "text", + "content": "[59] Yang Zheng, Adam W. Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J. Guibas. Pointodyssey: A large-scale synthetic dataset for long-term point tracking. In ICCV, 2023. 1, 3, 5, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 479, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 555, + 533 + ], + "type": "text", + "content": "[60] Artem Zholus, Carl Doersch, Yi Yang, Skanda Koppula, Viorica Patraucean, Xu Owen He, Ignacio Rocco, Mehdi SM Sajjadi, Sarath Chandar, and Ross Goroshin. TAPNext: Tracking any point (TAP) as next token prediction. arXiv:2504.05579, 2025.8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5262" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_content_list.json b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6814157a10356df98a044ed01bb2628db56b3ff3 --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_content_list.json @@ -0,0 +1,1693 @@ +[ + { + "type": "text", + "text": "Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation", + "text_level": 1, + "bbox": [ + 138, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuchang Ye1 Usman Naseem2 Mingyuan Meng1 Jinman Kim1 \n1The University of Sydney, Sydney, Australia \n2Macquarie University, Sydney, Australia", + "bbox": [ + 217, + 202, + 777, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shuchang.ye@sydney.edu.au", + "bbox": [ + 251, + 260, + 493, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "usman.naseem@mq.edu.au", + "bbox": [ + 526, + 260, + 730, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mmen2292@uni.sydney.edu.au", + "bbox": [ + 254, + 277, + 488, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jinman.kim@sydney.edu.au", + "bbox": [ + 522, + 277, + 736, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Medical language-guided segmentation, integrating textual clinical reports as auxiliary guidance to enhance image segmentation, has demonstrated significant improvements over unimodal approaches. However, its inherent reliance on paired image-text input, which we refer to as \"textual reliance\", presents two fundamental limitations: 1) many medical segmentation datasets lack paired reports, leaving a substantial portion of image-only data underutilized for training; and 2) inference is limited to retrospective analysis of cases with paired reports, limiting its applicability in most clinical scenarios where segmentation typically precedes reporting. To address these limitations, we propose ProLearn, the first Prototype-driven Learning framework for language-guided segmentation that fundamentally alleviates textual reliance. At its core, we introduce a novel Prototype-driven Semantic Approximation (PSA) module to enable approximation of semantic guidance from textual input. PSA initializes a discrete and compact prototype space by distilling segmentation-relevant semantics from textual reports. Once initialized, it supports a query-and-respond mechanism which approximates semantic guidance for images without textual input, thereby alleviating textual reliance. Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms state-of-the-art language-guided methods when limited text is available. The code is available at https://github.com/ShuchangYe-bib/ProLearn.", + "bbox": [ + 88, + 358, + 483, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmentation is an essential tool in medical image analysis, supporting clinical workflows by enabling precise delineation of anatomical structures, identification of pathological regions, and facilitating targeted interventions [24]. Its applications span critical tasks such as disease diag", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21b6ad183903f790dc44cbaaf35c6381287e5acddb4d911f27d7f48b0511f5d5.jpg", + "image_caption": [ + "Figure 1. Comparison of vision-language paradigms for medical image segmentation. a) VLP: Pretrained on large-scale image-text pairs and then fine-tuned on image-only data from the target dataset. b) Language-guided Segmentation: Requires strict image-text pairs during both training and inference. c) ProLearn: initializes with a limited amount of paired image-text data to construct query-response spaces. After initialization, it enables learning with limited textual input and performs inference without text." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 901, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nosis, treatment planning, and surgery support [1]. Deep learning has revolutionized segmentation, making it more accurate, reliable, and widely applicable in clinical practice. Unimodal (image-only) segmentation methods [43], such as U-Net [32] and its extensions [33], including U-Net++ [49], Attention U-Net [28], and Trans U-Net [8], have been widely adopted in medical imaging. In recent years, multimodal segmentation methods that leverage textual clinical reports as complementary information have gained wide attention for their ability to transcend the performance limits of unimodal segmentation [6, 13, 44].", + "bbox": [ + 509, + 641, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The exploration of multimodal segmentation [6, 13, 44] began with vision-language pretraining (VLP) [31], as shown in Figure 1a, where models pre-trained on paired textual and visual data demonstrated improved visual understanding and performance when finetuned on target image-only segmentation datasets [23, 37, 39]. How", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22316", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4427badb1869cef6a057c0149df79cb47e5a02d5d4ecd7fef0cf77fab078b12c.jpg", + "image_caption": [ + "Figure 2. Flow diagram comparison between SGSeg [42] and our proposed ProLearn. a) SGSeg: image $\\rightarrow$ LLM $\\rightarrow$ report $\\rightarrow$ BERT $\\rightarrow$ embedding. b) ProLearn: image $\\rightarrow$ PSA $\\rightarrow$ embedding." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 310, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7e559e8b4d082ce9cfdfc13e14b5458e020992dcdb9f18f8d66c1863b5e6479.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ever, these pretraining-based approaches fail to fully exploit the disease-specific information embedded in the target dataset's reports, as the general knowledge learned during pretraining often lacks the domain-specific details essential for precise disease lesion segmentation. To address this limitation, language-guided segmentation has been proposed as a promising approach [15, 21, 46, 48], which takes textual clinical report inputs as auxiliary semantic guidance for image segmentation (see Figure 1b). It has achieved remarkable success by leveraging textual abnormality description as explicit semantic guidance to segmentation, thus outperforming pretraining-based approaches and setting new benchmarks in medical image segmentation.", + "bbox": [ + 88, + 277, + 482, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, language-guided segmentation methods have inherent reliance on paired image-text data due to the requirement for textual reports as auxiliary inputs. This reliance is referred to as textual reliance in this study and has incurred notable limitations during the training and inference stages. In the training stage, its reliance on paired image-text data leads to the underutilization of image-only data, which constitutes the majority of available datasets [2]. In the inference stage, reliance on textual descriptions during inference confines its use to retrospective analysis, misaligning with most clinical workflows where segmentation is required preemptively for tasks such as preoperative planning [12], diagnostic decision making [4, 25, 26], and real-time procedural guidance [14, 19, 38].", + "bbox": [ + 89, + 476, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our previous study, Self-Guided Segmentation (SGSeg) [42], made a preliminary attempt to eliminate textural reliance during inference in language-guided segmentation by leveraging large language models (LLMs) to generate synthetic reports to compensate for the missing textual input, as illustrated in Figure 2a. However, the integration of LLMs substantially increases the model size and inference time, making the approach unsuitable for deployment on edge devices [7] or for real-time applications such as image-guided surgery [20]. Moreover, textual reliance during training remains an open issue.", + "bbox": [ + 89, + 686, + 482, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that the critical guidance in language-guided segmentation is not the entire clinical report, which is often verbose and cluttered with irrelevant information, but", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rather the specific segmentation-relevant semantic features embedded within it. Our investigation further indicates that the semantic space of medical reports is inherently constrained, consisting of a finite set of distinct representations relevant to segmentation tasks, as clinical reports typically adhere to standardized medical terminologies, which result in a relatively closed vocabulary.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these insights, we propose ProLearn (see Figure 1c), a lightweight and efficient Prototype Learning framework that fundamentally alleviates textual reliance during both training and inference. ProLearn introduces a Prototype-Driven Semantic Approximation (PSA) module, which enables the model to approximate semantic guidance without the need for textual input. Specifically, PSA constructs a discrete and compact prototype space by distilling segmentation-relevant concepts from clinical reports. It then provides a query-and-respond mechanism to support interaction between segmentation models and the prototype space, where unseen semantics are approximated by weighted aggregation of the existing prototypes based on similarity. Therefore, PSA enables segmentation models to query by image feature and receive responded semantics feature as guidance for feature maps refinement, as illustrated in Figure 2b.", + "bbox": [ + 511, + 198, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are as follows:", + "bbox": [ + 511, + 455, + 852, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, our proposed ProLearn is the first work to alleviate textual reliance in medical language-guided segmentation in both training and inference.", + "- We introduce a novel PSA module that supports learning with both paired image-text data and image-only data while enabling inference without textual input by querying a learned prototype space to provide semantic guidance for segmentation.", + "- Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms language-guided segmentation methods in limited text availability settings (see Section 5.1) and surpasses state-of-the-art unimodal segmentation models (see Section 5.2). Compared to SGSeg, ProLearn achieves a $1000 \\times$ reduction in parameters and $100 \\times$ faster inference speed (see Section 5.3)." + ], + "bbox": [ + 511, + 472, + 903, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 744, + 653, + 761 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Language-guided Segmentation", + "text_level": 1, + "bbox": [ + 511, + 771, + 792, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-guided segmentation aims to address the gap that the target dataset's reports are not fully exploited for learning in conventional unimodal [5, 8, 28, 32, 49] or VLP methods [23, 31, 37, 39]. LViT [21] was the first work in language-guided segmentation, which takes both images and textual reports as input to train a multimodal-input segmentation model. They annotated existing pub-", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22317", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fde88a958ef02a100a8a3c17c1d4a7495d147678cba75054ddc56d658a77a1d.jpg", + "image_caption": [ + "Figure 3. The overview of the ProLearn framework for alleviating textual reliance in medical language-guided segmentation. a) Language-guided U-Net: The U-Net encoder extracts image features, which are used to query the PSA module. The decoder is then guided by PSA's response to decode the segmentation mask. b) PSA Initialization: Clinical reports are processed and abstracted into a discrete prototype space, representing segmentation-relevant semantic and spatial information. c) PSA Querying and Responding: PSA receives image feature queries, selects relevant prototype candidates, and responds with an aggregated representation to approximate the pathological region's segmentation-relevant semantic representation." + ], + "image_footnote": [], + "bbox": [ + 160, + 92, + 848, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "licly available segmentation datasets, QaTa-COV19 [10] and MosMedData+ [27], with corresponding clinical reports. To fuse the text features with feature maps in the U-Net, LViT adopted an early fusion approach, which introduced a Pixel-Level Attention Module (PLAM) to involve textual features as semantic guidance. LViT showed consistent performance over image-only and VLP segmentation methods, which demonstrated the importance and potential of textual semantic guidance from clinical reports in target datasets.", + "bbox": [ + 88, + 540, + 480, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Subsequent methods proposed more flexible and robust feature fusion modules for language-guided segmentation. GuideSeg [48] moved away from early fusion while adopting a late-fusion strategy that fused textual and visual features at the decoding stage, where the text features were better preserved and more effectively influenced the segmentation process. MA adaptor [46] addressed the unidirectional flow of textual semantics seen in previous frameworks. It introduced a bidirectional adaptor connecting multiple layers of unimodal encoders, facilitating mutual information exchange between text and image representations at various scales. LGA [15] adopted a parameter-efficient fin-tuning strategy that preserved the original parameters of", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "large segmentation foundation models [24]. These works further prove that integrating target datasets' textual reports as guidance can significantly improve the performance of segmentation models.", + "bbox": [ + 511, + 540, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These methods suffered from textual reliance. The reliance during training restricted the utilization of large portions of medical datasets that lacked paired textual reports. The reliance during inference limits their practical utility in most clinical scenarios using segmentation without reports. SGSeg [42] attempted to release the textual reliance at inference by training a LLM [35] to generate needed clinical reports from images. Nevertheless, the inclusion of LLMs increased model size and inference time, making it unsuitable for edge devices and real-time applications. The textual reliance during training still remained unsolved. Our proposed ProLearn fundamentally alleviates textual reliance in training and inference, and its prototype design PSA significantly reduces the number of parameters and inference time compared to LLM-based SGSeg.", + "bbox": [ + 511, + 603, + 908, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prototype Learning", + "text_level": 1, + "bbox": [ + 511, + 845, + 702, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prototype learning draws on the principle that images and text can be effectively captured in discrete prototype repre", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22318", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cab19660a2132f4f0970e8b50b497a84e4517b12d08c682f03b73f49d5391cb4.jpg", + "image_caption": [ + "Figure 4. Attention-guided surrogate label extraction." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 480, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentations rather than relying on freshly built embeddings for every instance [36]. Each prototype functions as a canonical reference, reflecting the typical features of its corresponding class. Such a strategy is widely adopted in classification [40], where an unseen query is associated with the class whose prototype lies closest in a learned space. By prioritizing the reuse of well-established representations, prototype-based methods frequently attain top-tier performance under data-scarce settings [34]. Beyond unimodal applications, prototype learning has demonstrated its potential in multimodal tasks, effectively aligning features across modalities [9], such as images and text. Inspired by the principles of expressing unseen queries via finite existing prototypes, the proposed ProLearn extracts semantic information from textual reports as prototypes, enabling representing semantic guidance from existing image-text pairs for image-only input during training and inference, with improved parameter, computation, and data efficiency.", + "bbox": [ + 89, + 267, + 483, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 551, + 225, + 569 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 shows the overall architecture of ProLearn, a framework that leverages an efficient prototype learning module called Prototype-Driven Semantic Approximation (PSA) alongside a Language-guided U-Net backbone for segmentation. PSA is designed to (i) reduce the need for textual annotations during training by using available text only once to initialize a discrete prototype space and (ii) remove the need for textual input entirely at inference.", + "bbox": [ + 89, + 577, + 483, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. PSA Initialization", + "text_level": 1, + "bbox": [ + 89, + 707, + 264, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The PSA initialization is a one-time process that constructs a queryable prototype space before training. Given that a training set includes $K$ paired image-text samples and other image-only samples, where the $i$ -th sample is denoted as $\\langle I_i, T_i \\rangle$ , with $I_i$ as the image and $T_i$ as its associated report. Each paired sample is processed by a trained domain-specific vision-language encoder (BioMedCLIP [45]), $f_{\\mathrm{enc}}^I$ and $f_{\\mathrm{enc}}^T$ to extract their semantic features $e_i^I$ and $e_i^T$ :", + "bbox": [ + 89, + 729, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {I} = f _ {\\mathrm {e n c}} ^ {I} (I _ {i}), \\quad e _ {i} ^ {T} = f _ {\\mathrm {e n c}} ^ {T} (T _ {i}), \\quad e _ {i} ^ {I}, e _ {i} ^ {T} \\in \\mathbb {R} ^ {D} \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 857, + 483, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surrogate Label Extraction: To extract finite", + "bbox": [ + 109, + 885, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "segmentation-relevant semantics in the clinical reports, we cluster the image-text pairs into $N$ semantic surrogate labels, as shown in Figure 4.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As many clinical reports are verbose, we first isolate the tokens most relevant to segmentation. To achieve this, we utilize the cross-attention module of a separately trained Language-guided U-Net, which has the same architecture as the language-guided U-Net in Figure 3, to assess token relevance in all $K$ available image-text pairs $\\{\\langle I_i,T_i\\rangle \\mid i = 1,\\dots ,K\\}$ . Specifically, for each text input $T_{i}$ paired with an image $I_{i}$ , the cross-attention weight $\\alpha_{j}$ computed for each token $t_j$ in $T_{i}$ indicates the relevance of each token $t_j$ in guiding the segmentation of $I_{i}$ . Tokens whose attention scores exceed a threshold $\\tau$ are retained, resulting in a shorter segmentation-relevant sentence for each $T_{i}$ , which we denote $T_{i}^{\\mathrm{selected}}$ :", + "bbox": [ + 511, + 136, + 906, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {\\text {s e l e c t e d}} = \\left\\{t _ {j} \\mid \\alpha_ {j} > \\tau , t _ {j} \\in T _ {i} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 339, + 903, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then feed $T_{i}^{\\mathrm{selected}}$ into the text encoder $f_{\\mathrm{enc}}^{T}$ , obtaining a semantic feature $e_{i}^{\\mathrm{sem}}$ :", + "bbox": [ + 511, + 364, + 903, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {\\text {s e m}} = f _ {\\text {e n c}} ^ {T} \\left(T _ {i} ^ {\\text {s e l e c t e d}}\\right), \\quad \\forall i \\in \\{1, \\dots , K \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 402, + 903, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then group similar textual semantics together using the hierarchical density-based clustering algorithm, HDBSCAN [3]. The clustering results in $N$ surrogate labels $\\{l_1,l_2,\\dots ,l_N\\}$ . Each surrogate label $l_{i}$ encapsulates a distinct segmentation-relevant textual semantics, with its corresponding cluster $\\mathcal{C}_i$ representing the grouped features:", + "bbox": [ + 511, + 428, + 905, + 518 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i} = \\left\\{\\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\mid \\operatorname {H D B S C A N} \\left(e _ {i} ^ {\\text {s e m}}\\right) = l _ {i} \\right\\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 525, + 903, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prototype Space Construction: While the above surrogate labels capture sparse textual semantics, medical images often convey richer and more fine-grained information. To make the prototype space compact and more representative for image queries, each text-based surrogate label cluster $\\mathcal{C}_i$ is further subdivided into sub-clusters $\\mathcal{C}_{ij}$ via K-means clustering algorithm [22]:", + "bbox": [ + 511, + 550, + 905, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i j} = \\left\\{\\left\\langle e _ {k} ^ {I}, e _ {k} ^ {T} \\right\\rangle \\mid \\text {K - M e a n} \\left(e _ {k} ^ {I}\\right) = j, \\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\in \\mathcal {C} _ {i} \\right\\}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 662, + 903, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each sub-cluster $\\mathcal{C}_{ij}$ , instead of taking the centroid itself, we locate the sample $c_{ij}$ closest to the cluster's centroid and treat it as a fine representation of disease lesion. This approach reduces the influence of outliers that can significantly distort the centroid. Given $c_{ij} = \\langle e_k^I, e_k^T \\rangle$ , we assign its containing image feature $e_k^I$ as visual query prototype $q_{ij}$ and text feaître $e_k^T$ as textual respond prototype $r_{ij}$ . By repeating this across all sub-clusters, we construct a discrete and compact initial prototype space $S$ , consisting of a query space $S^Q$ and a response space $S^R$ , both of dimension $N \\times M \\times D$ :", + "bbox": [ + 511, + 688, + 905, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left(\\mathcal {S} ^ {Q}, \\mathcal {S} ^ {R}\\right) = \\bigcup_ {i = 1} ^ {N} \\bigcup_ {j = 1} ^ {M} \\langle q _ {i j}, r _ {i j} \\rangle . \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 859, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22319", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each visual query prototype $q_{ij}$ in $\\mathcal{S}^Q$ is directly linked to its corresponding textual response prototype $r_{ij}$ in $\\mathcal{S}^R$ :", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i j} \\longrightarrow r _ {i j}, \\quad \\forall i \\in \\{1, \\dots , N \\}, \\quad j \\in \\{1, \\dots , M \\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 132, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This prototype space is dynamically learned during the training process of ProLearn.", + "bbox": [ + 89, + 159, + 482, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. PSA Querying and Responding", + "text_level": 1, + "bbox": [ + 89, + 196, + 367, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During both training and inference, the Language-guided U-Net $f_{\\mathrm{seg}}$ queries the query space $S^Q$ using an image feature $q^*$ , encoded by the image encoder $f_{\\mathrm{enc}}^I$ from an image input $I^*$ , enabling the use of image-only samples. Through the query-and-response mechanism of the PSA module, it responds with an approximate textual semantic feature $r^*$ which guides $f_{\\mathrm{seg}}$ without textual input.", + "bbox": [ + 89, + 219, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PSA Querying: The PSA querying process involves searching the most relevant visual queries from the query prototype space $S^Q$ . Given the encoded image feature $q*$ , the PSA module computes the cosine similarity scores $s_{ij}$ between $q^*$ and each query prototype $q_{ij}$ :", + "bbox": [ + 89, + 325, + 483, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} j = s \\left(q ^ {*}, q _ {i j}\\right) = \\frac {q ^ {*} \\cdot q _ {i j}}{\\| q ^ {*} \\| \\| q _ {i j} \\|} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 411, + 482, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After ranking these similarity scores, the PSA module selects the top- $k$ query prototypes $Q^{*}$ that best match $q^{*}$ :", + "bbox": [ + 89, + 452, + 482, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {*} = \\left\\{q _ {i j} \\mid \\forall i, j \\in \\arg \\operatorname {t o p} _ {k} \\left(s _ {i j}\\right) \\right\\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 493, + 482, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PSA Responding: The PSA responding process finds the corresponding response prototypes $R^{*}$ that are linked to the selected query prototypes $Q^{*}$ :", + "bbox": [ + 89, + 520, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {*} = \\left\\{r _ {i j} \\mid q _ {i j} \\in Q ^ {*}, q _ {i j} \\longrightarrow r _ {i j} \\right\\}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 575, + 482, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These response prototypes $R^{*}$ are referred to as candidates in Figure 3b. The candidates are then aggregated using a similarity-weighted sum, where the weights are computed through the softmax function over the similarity scores:", + "bbox": [ + 89, + 603, + 483, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr ^ {*} = \\sum_ {r _ {i} \\in R ^ {*}} w _ {i} r _ {i}, \\quad w _ {i} = \\frac {\\exp \\left(s \\left(q ^ {*} , q _ {i}\\right)\\right)}{\\sum_ {q _ {j} \\in Q ^ {*}} \\exp \\left(s \\left(q ^ {*} , q _ {j}\\right)\\right)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 672, + 482, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At this point, the PSA responding process is complete. We then feed $r^*$ into the decoding process of Language-guided U-Net, providing an approximated semantic guidance.", + "bbox": [ + 89, + 719, + 483, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 777, + 282, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 803, + 192, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our approach, we used the two well-benchmarked datasets: QaTa-COV19 [10], MosMedData+ [27] and Kvasir-SEG [18], which are commonly adopted for performance comparison in language-guided segmentation [15, 21, 42, 46, 48].", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QaTa-COV19 is a large-scale dataset consisting of 9,258 chest X-ray images with manually annotated COVID-19 [41] lesion masks. To facilitate language-guided segmentation, LViT [21] extends this dataset with textual descriptions detailing bilateral lung infections, the number of affected regions, and their spatial localization within the lungs. We adopt the official dataset split: 5,716 images for training, 1,429 for validation, and 2,113 for testing.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MosMedData+ comprises 2,729 CT slices depicting pulmonary infections. Similar to QaTa-COV19, LViT augments this dataset with text-based annotations to support language-guided segmentation tasks. We adopt the official dataset split: 2,183 slices for training, 273 for validation, and 273 for testing.", + "bbox": [ + 511, + 210, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Kvasir-SEG is a publicly available dataset comprising 1,000 colonoscopy images of gastrointestinal polyps with corresponding pixel-level segmentation masks. We follow a standard 8:1:1 split for training, validation, and testing.", + "bbox": [ + 511, + 301, + 903, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 371, + 697, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To quantitatively evaluate segmentation performance, we used the metrics which are used in the previous similar studies [15, 21, 42, 46, 48]: Dice coefficient [11] and the mean Intersection over Union (mIoU [17]), two widely used metrics for measuring spatial overlap between predicted and ground truth segmentation masks. The formulations are defined in Equations 12 and 13.", + "bbox": [ + 511, + 393, + 903, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {D i c e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\sum_ {c = 1} ^ {C} \\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\sum_ {c = 1} ^ {C} \\left(\\left| P _ {i} ^ {(c)} \\right| + \\left| G _ {i} ^ {(c)} \\right|\\right)} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 508, + 903, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {m I o U} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {\\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\left| P _ {i} ^ {(c)} \\cup G _ {i} ^ {(c)} \\right|} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 561, + 903, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ represents the number of images in the dataset, $C$ denotes the number of semantic categories, and $P_{i}^{(c)}$ and $G_{i}^{(c)}$ correspond to the predicted and ground truth segmentation masks for class $c$ in image $i$ , respectively.", + "bbox": [ + 511, + 608, + 903, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Experimental Design", + "text_level": 1, + "bbox": [ + 511, + 681, + 710, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate ProLearn, we compare it against both multimodal and unimodal segmentation methods, analyzing its performance under various realistic clinical scenarios.", + "bbox": [ + 511, + 704, + 903, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Limited availability of text: To illustrate the limitations of language-guided segmentation methods that require strict image-report pairing and highlight the importance of training with both image-text-paired and image-only data, we simulate a real-world clinical scenario where detailed radiology reports are often unavailable for a significant portion of the dataset. Due to the lack of large-scale benchmarks containing both paired and unpaired examples, we construct a counterfactual setting by progressively reducing the proportion of image-text pairs in the training data to $50\\%$ ,", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22320", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fac14836dd2854a198e3ff5e7fcb9178b6ff8bc7173b09c08cabe96a7e9e433d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetModelDicemIoU
50%25%10%5%1%50%25%10%5%1%
QaTa-COV19 [10]LViT [21]0.84160.82020.80040.76380.70060.73200.70120.67470.62480.5490
GuideSeg [48]0.86330.85240.84020.82400.73330.75950.74280.72440.70070.5789
SGSeg [42]0.86410.85740.84230.80570.73070.76070.75040.72760.67460.5757
ProLearn0.86670.85980.85830.85730.85660.77210.76900.75730.75580.7553
MosMedData+ [27]LViT [21]0.71890.66080.55010.50690.16770.56960.50420.41080.35380.1015
GuideSeg [48]0.75080.73930.68980.63750.42350.60890.58640.52650.46780.2686
SGSeg [42]0.74550.74390.69500.64650.34520.59430.59220.53250.47760.2086
ProLearn0.75390.75120.74240.73790.72180.61260.61090.60870.60690.6032
Kvasir-SEG [18]LViT [21]0.76690.64240.57190.54820.42720.62280.47690.40250.37820.2729
GuideSeg [48]0.88480.83900.77540.74970.56150.79390.72280.63600.60430.4008
SGSeg [42]0.87690.83040.80250.75260.54060.78080.70990.67020.60340.3705
ProLearn0.89830.89460.88980.88230.87180.81620.81010.80200.79050.7729
", + "bbox": [ + 143, + 88, + 857, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison of language-guided segmentation models in simulated scenarios with limited text supervision on the QaTa-COV19, MosMedData+ and Kvasir-SEG dataset. The best results are highlighted in bold.", + "bbox": [ + 89, + 316, + 906, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "25%, 10%, 5%, and 1%, while discarding the remaining unpaired images to simulate their inaccessibility. We compare ProLearn against state-of-the-art language-guided segmentation models: LViT [21], GuideSeg [48], and SGSeg [42].", + "bbox": [ + 89, + 369, + 482, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation in real-world (image-only) setting: In the majority of clinical scenarios, such as real-time procedural guidance and decision support, segmentation is used without text. To align with real-world settings, we evaluate ProLearn in a strictly \"image-only\" setting, where no text input is provided during either training or inference on the target datasets. Under these conditions, we compare ProLearn to established unimodal segmentation methods and vision-language pretraining models (CLIP [31] and GLoRIA [16]) adapted for image-only use.", + "bbox": [ + 89, + 446, + 483, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Prototype vs. LLM: To demonstrate our PSA's advantages against LLMs in real-world deployment, we focus on inference speed, model size, and time complexity. We specifically compare ProLearn with SGSeg [42], a recent method that leverages LLMs (e.g., GPT-2 [30], Llama3 [29]) to generate synthetic textual reports at inference in order to compensate for missing textual input.", + "bbox": [ + 89, + 613, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Analysis: We provide visual comparisons against state-of-the-art language-guided segmentation models under progressively lower text availability settings. Specifically, We visualize both final segmentation outputs and corresponding saliency maps.", + "bbox": [ + 89, + 734, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameter Sensitivity Analysis: We analyze how model performance varies as we adjust two key hyperparameters: the number of prototypes $M$ , which governs the compactness of the prototype space, and the number of candidate responses $k$ .", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e4168c5036418f1d68efd61ac74629bf9beb361b781d233e054f5bc95167a3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 369, + 717, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/142f38d89ba4e96bcfd0433b749a37434384396318f686e39a3865dd6f6b66a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 369, + 895, + 473 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4463e7ad9077d5b242d49f4e2c41be8d7e5e39eac577ad4855e16a4be65537c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 479, + 717, + 583 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ccd2da5a06773d45c4c2f4a4d747f88395603bad1c8f67bcb6f7e7d005f971a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 722, + 479, + 895, + 583 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/11fb212e2d1fae579770553c467dc5e46561234825e22220c06a51a54e568552.jpg", + "image_caption": [ + "Figure 5. Performance degradation as the ratio of text availability decreases. The plots show Dice and nIoU metrics for a) the QaTa-COV19, b) the MosMedData+ and c) the Kvasir-SEG dataset." + ], + "image_footnote": [], + "bbox": [ + 532, + 592, + 717, + 691 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee722b6c26e182e31005dd66cdd682043d76095ef09649b7eb306d4e19209f32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 593, + 893, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results and Discussion", + "text_level": 1, + "bbox": [ + 511, + 775, + 728, + 791 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison with State-Of-The-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 801, + 892, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents our comparative results under clinical scenarios where paired textual reports are limited (50%, 25%, 10%, 5%, and 1%). ProLearn achieved higher Dice and mIoU scores than existing language-guided methods in all settings, and this performance gap increased as the frac", + "bbox": [ + 511, + 824, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22321", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/75e90768487d1333aaca4b05f246fa7ba8b4a469a7689bb2bba969b8fb262d93.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelQaTa-COV19MosMedData+Kvasir-SEG
DicemIoUDicemIoUDicemIoU
U-Net [32]0.8190.6920.6380.5050.1950.182
U-Net++ [49]0.8230.7060.7140.5820.2800.180
Attention U-Net [28]0.8220.7010.6640.5280.3640.226
Trans U-Net [8]0.8060.6870.7020.5750.0480.100
Swin U-Net [5]0.8360.7240.6690.5310.3980.246
CLIP [31]0.7980.7070.7200.596--
GLoRIA [16]0.7990.7070.7220.602--
MedSAM [24]0.7300.6190.5090.371--
BiomedParse [47]0.7810.6820.6710.5530.8280.721
ProLearn (1%)0.8570.7550.7220.6030.8720.773
ProLearn (5%)0.8570.7560.7380.6070.8820.790
ProLearn (10%)0.8580.7570.7420.6090.8890.802
", + "bbox": [ + 91, + 89, + 491, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion of available text decreased. For example, in the $1\\%$ text in MosMedData+, ProLearn retained a Dice score of 0.7218, compared to SGSeg 0.3452 and LViT 0.1677. This trend underscores the effectiveness of ProLearn in learning from image-text and image-only data, allowing it to outperform approaches that require textual input for both training and inference. Although SGSeg also requires no text at inference, its performance remained below ProLearn due to its limited capacity to exploit learned semantics when text is scarce. Further discussions on the comparison between SGSeg and ProLearn can be found in Section 5.3.", + "bbox": [ + 88, + 398, + 482, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance degradation analysis, visualized in Figure 5, further illustrates the importance of reducing textual reliance in training. ProLearn exhibited the most minor degradation, preserving robust performance even under severely limited text supervision.", + "bbox": [ + 89, + 571, + 483, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison in real-world (image-only) setting", + "text_level": 1, + "bbox": [ + 89, + 676, + 482, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluated the effectiveness of ProLearn's use of a target dataset's auxiliary text under near \"image-only\" conditions, as most clinical workflows require segmentation models operating without textual input. As shown in Table 2, ProLearn outperformed both unimodal approaches (U-Net, Attention U-Net, Swin U-Net, etc.) and well-recognized VLP methods (CLIP, GLoRIA), even when only $10\\%$ , $5\\%$ , or $1\\%$ of the paired reports were available. This demonstrates the importance of textual semantics within the target dataset. Unlike unimodal and VLP-based methods, which fail to leverage clinical reports during fine-tuning on target datasets, ProLearn incorporates available image-report pairs in the learning process.", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e82611590829cf3788a2adb3164bcee9482a05e8dd74f215436894e129039fd.jpg", + "table_caption": [ + "Table 2. Performance comparison of methods in image-only settings where paired reports are excluded entirely from the target datasets' training and inference. The best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
ProLearnSGSeg
MethodQuery & respond from prototype matrixGenerate text tokens one by one (autoregressive) & text embedding
No. of inferenceSingle inferenceMultiple inference
TimeO(1)O(n) depends on the no. of tokens
", + "bbox": [ + 522, + 89, + 903, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0edecf992e88319dde3b78c2ac5ce6556da282a83368b4355f482e12ff2e72dd.jpg", + "image_caption": [ + "Figure 6. Comparison between the proposed prototype-driven approach, ProLearn, and the LLM-based approach, SGSeg. Top: The theoretical analysis of time complexity. Bottom: The model size and inference time comparison." + ], + "image_footnote": [], + "bbox": [ + 526, + 160, + 696, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff891e1d07e66ea52635536f76d2e4d8292a6b87723e20c87d942c70d94de6fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 160, + 895, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Prototype vs. LLM", + "text_level": 1, + "bbox": [ + 513, + 359, + 697, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prototype learning enables constant-time inference of $\\mathcal{O}(1)$ because it uses a pre-defined, finite prototype space that is independent of both image and text size. At the same time, LLM applies autoregressive image-to-text generation, which has a linear time complexity $\\mathcal{O}(n)$ , where $n$ represents the number of tokens in the generated text (see Figure 6, top). As shown in Figure 6, bottom, prototype learning achieves an inference time of $4\\mathrm{ms}$ , making it $100\\times$ faster than GPT-2 (136ms) and $300\\times$ faster than Llama3 (1.2s). Moreover, the prototype model has only 1M parameters, which is $1000\\times$ smaller than large language models such as GPT-2 (1.5B parameters) and Llama3 (7B parameters). These properties make prototype learning highly efficient for real-time applications and suitable for deployment on resource-constrained edge devices.", + "bbox": [ + 511, + 388, + 906, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Visualization", + "text_level": 1, + "bbox": [ + 513, + 645, + 651, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7 presents a qualitative comparison of ProLearn with state-of-the-art approaches on the QaTa-COV19, MosMedData+ and Kvasir-SEG datasets. As textual availability decreased, existing language-guided segmentation models, which rely heavily on image-text pairs, experienced a pronounced drop in performance. In contrast, ProLearn effectively learned from both image-text and image-only data, demonstrating stable results even under limited textual guidance. The same findings can be observed in saliency maps that pinpoint the regions on which each model was focused. Under decreased text guidance, conventional models produced unstable or diffuse attention patterns, compromising their segmentation accuracy. However, ProLearn, aided by semantic approximation, retained more coherent attention focused on areas of the lesion.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22322", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8ff8901badf26303f47fc3eff26e9167cb96d6d40de5a69b721d20f8c52f16d.jpg", + "image_caption": [ + "Figure 7. Visualization comparison of language-guided segmentation methods under different text availabilities. The upper row shows segmentation outputs on CXR_S09687-E18404-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_4_85506_1(case19_16 from the MosMedData+ dataset (middle) and ju5xkwzxf0z0818gk4xabdm from the Kvasir-SEG dataset (right). Blue regions indicate groundtruth pixels missed by the model, while red regions indicate pixels mistakenly predicted as lesion. The lower row presents saliency map interpretability on CXR_S09346-E23164-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_40_86625_0(case18_53 from the MosMedData+ dataset (middle) and cju2tzypl4wss0799ow05oxb9 from the Kvasir-SEG dataset (right)." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/458e4c6ec4b7712acd49fb8eebfd8346d83baf62b5833ab23e2aae487e0ab5b3.jpg", + "image_caption": [ + "Figure 8. Hyperparameter sensitivity analysis with varying candidate and prototype configurations. a) Effect of varying $k$ , number of responding prototype vectors on Dice score; b) Effect of varying $M$ , number of prototypes per surrogate label on Dice score." + ], + "image_footnote": [], + "bbox": [ + 107, + 459, + 285, + 554 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6bb5ae8bebaa1c389b5f689e2b7b0ea4bb9df91d68972ce8e794c7df7517e492.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 458, + 473, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Hyperparameter Sensitivity", + "text_level": 1, + "bbox": [ + 89, + 648, + 341, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the effect of the number of candidates per response $k$ and the number of prototypes per surrogate label $M$ , we presented the performance comparison by varying their values for segmentation, as shown in Figure 8. We adjusted one hyperparameter at a time while keeping the other fixed. The experimental results indicated that our framework remained stable across a broad range of hyperparameter values.", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, $k$ governs the trade-off between information diversity and noise. A small $k$ limits the model's ability to capture subtle segmentation patterns due to insufficient variability. As $k$ increases, richer patterns are incorporated, improving performance. However, when $k$ is too large, irrelevant patterns introduce excessive noise, leading to degraded segmentation accuracy.", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For $M$ , increasing the number of prototypes enhances performance by capturing diverse segmentation characteristics. However, when $M$ exceeds a certain threshold, prototype redundancy arises, leading to overlapping or irrelevant representations that blur the segmentation boundaries. Thus, balancing $M$ is crucial; too few prototypes hinder representational capacity, while an excessive count dilutes the model's focus on meaningful patterns.", + "bbox": [ + 511, + 459, + 906, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We investigated the issue of textual reliance in medical language-guided segmentation, which limits both the applicability of segmentation in clinical workflows and the exploration of image-only training data. To address this issue, we have presented ProLearn, a prototype-driven framework that fundamentally alleviates textual reliance in medical language-guided segmentation. Our experiments demonstrated that ProLearn effectively learns from both image-only and image-text data, making efficient use of the target dataset's textual descriptions. ProLearn showed stable performance under different text availability and outperformed state-of-the-art image-only approaches under minimal text supervision.", + "bbox": [ + 511, + 626, + 906, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works: In this work, we focused on demonstrating the core idea of ProLearn in single-label 2D segmentation. Future directions involve exploring multi-label segmentation, broader imaging modalities, and other language-guided vision tasks.", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22323", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Australian Research Council (ARC) under Grant IM240100224.", + "bbox": [ + 89, + 114, + 480, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 187, + 174 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10076-10095, 2024. 1", + "[2] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21438-21451, 2023. 2", + "[3] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 4", + "[4] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE transactions on medical imaging, 33(2):577-590, 2013. 2", + "[5] H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 2, 7", + "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1", + "[7] Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):1655-1674, 2019. 2", + "[8] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 7", + "[9] Pujin Cheng, Li Lin, Junyan Lyu, Yijin Huang, Wenhan Luo, and Xiaoying Tang. Prior: Prototype representation joint learning from medical images and reports. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21361-21371, 2023. 4", + "[10] Aysen Degerli, Serkan Kiranyaz, Muhammad E. H. Chowdhury, and Moncef Gabbouj. Osegnet: Operational segmentation network for Covid-19 detection using chest x-ray images. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2306–2310, 2022. 3, 5, 6", + "[11] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 5", + "[12] Vincenzo Ferrari, Marina Carbone, Carla Cappelli, Luigi Boni, Franca Melfi, Mauro Ferrari, Franco Mosca, and Andrea Pietrabissa. Value of multidetector computed tomogra" + ], + "bbox": [ + 93, + 183, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "phy image segmentation for preoperative planning in general surgery. Surgical endoscopy, 26:616-626, 2012. 2", + "[13] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7606-7623, 2022. 1", + "[14] Matthew Stephen Holden, Tamas Ungi, Derek Sargent, Robert C McGraw, Elvis CS Chen, Sugantha Ganapathy, Terry M Peters, and Gabor Fichtinger. Feasibility of real-time workflow segmentation for tracked needle interventions. IEEE Transactions on Biomedical Engineering, 61(6): 1720-1728, 2014. 2", + "[15] Jihong Hu, Yinhao Li, Hao Sun, Yu Song, Chujie Zhang, Lanfen Lin, and Yen-Wei Chen. Lga: A language guide adapter for advancing the sam model's capabilities in medical image segmentation. page 610–620, Berlin, Heidelberg, 2024. Springer-Verlag. 2, 3, 5", + "[16] Shih-Cheng Huang, Liyue Shen, Matthew P. Lungren, and Serena Yeung. Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3922-3931, 2021. 6, 7", + "[17] Paul Jaccard. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547-579, 1901. 5", + "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In International conference on multimedia modeling, pages 451-462. Springer, 2019. 5, 6", + "[19] Tina Kapur, Jan Egger, Jagadeesan Jayender, Matthew Toews, and William M Wells. Registration and segmentation for image-guided therapy. Intraoperative Imaging and Image-Guided Therapy, pages 79-91, 2014. 2", + "[20] Stijn Keereweer, Jeroen DF Kerrebijn, Pieter BAA Van Driel, Bangwen Xie, Eric L Kaijzel, Thomas JA Snoeks, Ivo Que, Merlijn Hutteman, Joost R Van Der Vorst, J Sven D Mieog, et al. Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199-207, 2011. 2", + "[21] Zihan Li, Yunxiang Li, Qingde Li, Puyang Wang, Dazhou Guo, Le Lu, Dakai Jin, You Zhang, and Qingqi Hong. Lvit: language meets vision transformer in medical image segmentation. IEEE transactions on medical imaging, 2023. 2, 5, 6", + "[22] S. Lloyd. Least squares quantization in pcm. IEEE Transactions on Information Theory, 28(2):129-137, 1982. 4", + "[23] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 1, 2", + "[24] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3, 7", + "[25] Dhiraj Maji, Prarthana Sigedar, and Munendra Singh. Attention res-unet with guided decoder for semantic segmentation" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22324", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of brain tumors. Biomedical Signal Processing and Control, 71:103077, 2022. 2", + "[26] Epimack Michael, He Ma, Hong Li, Frank Kulwa, and Jing Li. Breast cancer segmentation methods: current status and future potentials. *BioMed research international*, 2021(1): 9962109, 2021. 2", + "[27] Sergey Morozov, Anna Andreychenko, Nikolay Pavlov, Anton Vladzymyrskyy, Natalya Ledikhova, Victor Gombolevskiy, Ivan Blokhin, Pavel Gelezhe, Anna Gonchar, Valeria Chernina, et al. Mosmeddata: Chest ct scans with Covid-19 related findings. 2020. 3, 5, 6", + "[28] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, and et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 7", + "[29] Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards building multilingual language model for medicine. Nature Communications, 15(1):8384, 2024. 6", + "[30] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7", + "[32] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1, 2, 7", + "[33] N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031-82057, 2021. 1", + "[34] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4", + "[35] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 3", + "[36] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4", + "[37] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 1, 2", + "[38] Sudanthi Wijewickrema, Yun Zhou, James Bailey, Gregor Kennedy, and Stephen O'Leary. Provision of automated step-by-step procedural guidance in virtual reality surgery" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "simulation. In Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology, pages 69-72, 2016. 2", + "[39] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503-17512, 2023. 1, 2", + "[40] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 4", + "[41] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 5", + "[42] Shuchang Ye, Mingyuan Meng, Mingjian Li, Dagan Feng, and Jinman Kim. Enabling text-free inference in language-guided segmentation of chest x-rays via self-guidance. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2024. 2, 3, 5, 6", + "[43] Xiao-Xia Yin, Le Sun, Yuhan Fu, Ruiliang Lu, and Yanchun Zhang. [retracted] u-net-based medical image segmentation. Journal of healthcare engineering, 2022(1):4189781, 2022. 1", + "[44] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[45] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Angela Crabtree, Brian Piening, Carlo Bifulco, Matthew P. Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon. A multimodal biomedical foundation model trained from fifteen million image-text pairs. NEJM AI, 2(1):A1oa2400640, 2025. 4", + "[46] Xu Zhang, Bo Ni, Yang Yang, and Lefei Zhang. Madapter: A better interaction between image and language for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2024. 2, 3, 5", + "[47] Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Hin Lee, Ho Sid Kiblawi, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, Brian Piening, Carlo Bifulco, Mu Wei, Hoifung Poon, and Sheng Wang. A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities. Nature Methods, 22(1):166-176, 2025. 7", + "[48] Yi Zhong, Mengqiu Xu, Kongming Liang, Kaixin Chen, and Ming Wu. Ariadne's thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In International Conference on Medical Image Computing and" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22325", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer-Assisted Intervention, pages 724-733. Springer, 2023. 2, 3, 5, 6", + "[49] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. 1, 2, 7" + ], + "bbox": [ + 91, + 90, + 482, + 232 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22326", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_model.json b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4469b37e131ad9f514599913e3fd568479331335 --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.13, + 0.86, + 0.177 + ], + "angle": 0, + "content": "Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.778, + 0.258 + ], + "angle": 0, + "content": "Shuchang Ye1 Usman Naseem2 Mingyuan Meng1 Jinman Kim1 \n1The University of Sydney, Sydney, Australia \n2Macquarie University, Sydney, Australia" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.261, + 0.495, + 0.275 + ], + "angle": 0, + "content": "shuchang.ye@sydney.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.261, + 0.731, + 0.274 + ], + "angle": 0, + "content": "usman.naseem@mq.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.255, + 0.279, + 0.489, + 0.292 + ], + "angle": 0, + "content": "mmen2292@uni.sydney.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.279, + 0.737, + 0.292 + ], + "angle": 0, + "content": "jinman.kim@sydney.edu.au" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.485, + 0.768 + ], + "angle": 0, + "content": "Medical language-guided segmentation, integrating textual clinical reports as auxiliary guidance to enhance image segmentation, has demonstrated significant improvements over unimodal approaches. However, its inherent reliance on paired image-text input, which we refer to as \"textual reliance\", presents two fundamental limitations: 1) many medical segmentation datasets lack paired reports, leaving a substantial portion of image-only data underutilized for training; and 2) inference is limited to retrospective analysis of cases with paired reports, limiting its applicability in most clinical scenarios where segmentation typically precedes reporting. To address these limitations, we propose ProLearn, the first Prototype-driven Learning framework for language-guided segmentation that fundamentally alleviates textual reliance. At its core, we introduce a novel Prototype-driven Semantic Approximation (PSA) module to enable approximation of semantic guidance from textual input. PSA initializes a discrete and compact prototype space by distilling segmentation-relevant semantics from textual reports. Once initialized, it supports a query-and-respond mechanism which approximates semantic guidance for images without textual input, thereby alleviating textual reliance. Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms state-of-the-art language-guided methods when limited text is available. The code is available at https://github.com/ShuchangYe-bib/ProLearn." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.799, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Segmentation is an essential tool in medical image analysis, supporting clinical workflows by enabling precise delineation of anatomical structures, identification of pathological regions, and facilitating targeted interventions [24]. Its applications span critical tasks such as disease diag" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.326, + 0.903, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Figure 1. Comparison of vision-language paradigms for medical image segmentation. a) VLP: Pretrained on large-scale image-text pairs and then fine-tuned on image-only data from the target dataset. b) Language-guided Segmentation: Requires strict image-text pairs during both training and inference. c) ProLearn: initializes with a limited amount of paired image-text data to construct query-response spaces. After initialization, it enables learning with limited textual input and performs inference without text." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.642, + 0.907, + 0.809 + ], + "angle": 0, + "content": "nosis, treatment planning, and surgery support [1]. Deep learning has revolutionized segmentation, making it more accurate, reliable, and widely applicable in clinical practice. Unimodal (image-only) segmentation methods [43], such as U-Net [32] and its extensions [33], including U-Net++ [49], Attention U-Net [28], and Trans U-Net [8], have been widely adopted in medical imaging. In recent years, multimodal segmentation methods that leverage textual clinical reports as complementary information have gained wide attention for their ability to transcend the performance limits of unimodal segmentation [6, 13, 44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The exploration of multimodal segmentation [6, 13, 44] began with vision-language pretraining (VLP) [31], as shown in Figure 1a, where models pre-trained on paired textual and visual data demonstrated improved visual understanding and performance when finetuned on target image-only segmentation datasets [23, 37, 39]. How" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22316" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.312, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.091, + 0.482, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.207, + 0.485, + 0.25 + ], + "angle": 0, + "content": "Figure 2. Flow diagram comparison between SGSeg [42] and our proposed ProLearn. a) SGSeg: image \\(\\rightarrow\\) LLM \\(\\rightarrow\\) report \\(\\rightarrow\\) BERT \\(\\rightarrow\\) embedding. b) ProLearn: image \\(\\rightarrow\\) PSA \\(\\rightarrow\\) embedding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.278, + 0.483, + 0.475 + ], + "angle": 0, + "content": "ever, these pretraining-based approaches fail to fully exploit the disease-specific information embedded in the target dataset's reports, as the general knowledge learned during pretraining often lacks the domain-specific details essential for precise disease lesion segmentation. To address this limitation, language-guided segmentation has been proposed as a promising approach [15, 21, 46, 48], which takes textual clinical report inputs as auxiliary semantic guidance for image segmentation (see Figure 1b). It has achieved remarkable success by leveraging textual abnormality description as explicit semantic guidance to segmentation, thus outperforming pretraining-based approaches and setting new benchmarks in medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.687 + ], + "angle": 0, + "content": "However, language-guided segmentation methods have inherent reliance on paired image-text data due to the requirement for textual reports as auxiliary inputs. This reliance is referred to as textual reliance in this study and has incurred notable limitations during the training and inference stages. In the training stage, its reliance on paired image-text data leads to the underutilization of image-only data, which constitutes the majority of available datasets [2]. In the inference stage, reliance on textual descriptions during inference confines its use to retrospective analysis, misaligning with most clinical workflows where segmentation is required preemptively for tasks such as preoperative planning [12], diagnostic decision making [4, 25, 26], and real-time procedural guidance [14, 19, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Our previous study, Self-Guided Segmentation (SGSeg) [42], made a preliminary attempt to eliminate textural reliance during inference in language-guided segmentation by leveraging large language models (LLMs) to generate synthetic reports to compensate for the missing textual input, as illustrated in Figure 2a. However, the integration of LLMs substantially increases the model size and inference time, making the approach unsuitable for deployment on edge devices [7] or for real-time applications such as image-guided surgery [20]. Moreover, textual reliance during training remains an open issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We argue that the critical guidance in language-guided segmentation is not the entire clinical report, which is often verbose and cluttered with irrelevant information, but" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "rather the specific segmentation-relevant semantic features embedded within it. Our investigation further indicates that the semantic space of medical reports is inherently constrained, consisting of a finite set of distinct representations relevant to segmentation tasks, as clinical reports typically adhere to standardized medical terminologies, which result in a relatively closed vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.456 + ], + "angle": 0, + "content": "Building on these insights, we propose ProLearn (see Figure 1c), a lightweight and efficient Prototype Learning framework that fundamentally alleviates textual reliance during both training and inference. ProLearn introduces a Prototype-Driven Semantic Approximation (PSA) module, which enables the model to approximate semantic guidance without the need for textual input. Specifically, PSA constructs a discrete and compact prototype space by distilling segmentation-relevant concepts from clinical reports. It then provides a query-and-respond mechanism to support interaction between segmentation models and the prototype space, where unseen semantics are approximated by weighted aggregation of the existing prototypes based on similarity. Therefore, PSA enables segmentation models to query by image feature and receive responded semantics feature as guidance for feature maps refinement, as illustrated in Figure 2b." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.854, + 0.47 + ], + "angle": 0, + "content": "The main contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.905, + 0.532 + ], + "angle": 0, + "content": "- To the best of our knowledge, our proposed ProLearn is the first work to alleviate textual reliance in medical language-guided segmentation in both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.905, + 0.608 + ], + "angle": 0, + "content": "- We introduce a novel PSA module that supports learning with both paired image-text data and image-only data while enabling inference without textual input by querying a learned prototype space to provide semantic guidance for segmentation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.608, + 0.905, + 0.73 + ], + "angle": 0, + "content": "- Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms language-guided segmentation methods in limited text availability settings (see Section 5.1) and surpasses state-of-the-art unimodal segmentation models (see Section 5.2). Compared to SGSeg, ProLearn achieves a \\(1000 \\times\\) reduction in parameters and \\(100 \\times\\) faster inference speed (see Section 5.3)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.473, + 0.905, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.655, + 0.762 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.794, + 0.789 + ], + "angle": 0, + "content": "2.1. Language-guided Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Language-guided segmentation aims to address the gap that the target dataset's reports are not fully exploited for learning in conventional unimodal [5, 8, 28, 32, 49] or VLP methods [23, 31, 37, 39]. LViT [21] was the first work in language-guided segmentation, which takes both images and textual reports as input to train a multimodal-input segmentation model. They annotated existing pub-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22317" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.093, + 0.849, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.431, + 0.908, + 0.515 + ], + "angle": 0, + "content": "Figure 3. The overview of the ProLearn framework for alleviating textual reliance in medical language-guided segmentation. a) Language-guided U-Net: The U-Net encoder extracts image features, which are used to query the PSA module. The decoder is then guided by PSA's response to decode the segmentation mask. b) PSA Initialization: Clinical reports are processed and abstracted into a discrete prototype space, representing segmentation-relevant semantic and spatial information. c) PSA Querying and Responding: PSA receives image feature queries, selects relevant prototype candidates, and responds with an aggregated representation to approximate the pathological region's segmentation-relevant semantic representation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.541, + 0.482, + 0.692 + ], + "angle": 0, + "content": "licly available segmentation datasets, QaTa-COV19 [10] and MosMedData+ [27], with corresponding clinical reports. To fuse the text features with feature maps in the U-Net, LViT adopted an early fusion approach, which introduced a Pixel-Level Attention Module (PLAM) to involve textual features as semantic guidance. LViT showed consistent performance over image-only and VLP segmentation methods, which demonstrated the importance and potential of textual semantic guidance from clinical reports in target datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Subsequent methods proposed more flexible and robust feature fusion modules for language-guided segmentation. GuideSeg [48] moved away from early fusion while adopting a late-fusion strategy that fused textual and visual features at the decoding stage, where the text features were better preserved and more effectively influenced the segmentation process. MA adaptor [46] addressed the unidirectional flow of textual semantics seen in previous frameworks. It introduced a bidirectional adaptor connecting multiple layers of unimodal encoders, facilitating mutual information exchange between text and image representations at various scales. LGA [15] adopted a parameter-efficient fin-tuning strategy that preserved the original parameters of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.601 + ], + "angle": 0, + "content": "large segmentation foundation models [24]. These works further prove that integrating target datasets' textual reports as guidance can significantly improve the performance of segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.909, + 0.832 + ], + "angle": 0, + "content": "These methods suffered from textual reliance. The reliance during training restricted the utilization of large portions of medical datasets that lacked paired textual reports. The reliance during inference limits their practical utility in most clinical scenarios using segmentation without reports. SGSeg [42] attempted to release the textual reliance at inference by training a LLM [35] to generate needed clinical reports from images. Nevertheless, the inclusion of LLMs increased model size and inference time, making it unsuitable for edge devices and real-time applications. The textual reliance during training still remained unsolved. Our proposed ProLearn fundamentally alleviates textual reliance in training and inference, and its prototype design PSA significantly reduces the number of parameters and inference time compared to LLM-based SGSeg." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.704, + 0.864 + ], + "angle": 0, + "content": "2.2. Prototype Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Prototype learning draws on the principle that images and text can be effectively captured in discrete prototype repre" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22318" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.482, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.229, + 0.446, + 0.243 + ], + "angle": 0, + "content": "Figure 4. Attention-guided surrogate label extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.485, + 0.542 + ], + "angle": 0, + "content": "sentations rather than relying on freshly built embeddings for every instance [36]. Each prototype functions as a canonical reference, reflecting the typical features of its corresponding class. Such a strategy is widely adopted in classification [40], where an unseen query is associated with the class whose prototype lies closest in a learned space. By prioritizing the reuse of well-established representations, prototype-based methods frequently attain top-tier performance under data-scarce settings [34]. Beyond unimodal applications, prototype learning has demonstrated its potential in multimodal tasks, effectively aligning features across modalities [9], such as images and text. Inspired by the principles of expressing unseen queries via finite existing prototypes, the proposed ProLearn extracts semantic information from textual reports as prototypes, enabling representing semantic guidance from existing image-text pairs for image-only input during training and inference, with improved parameter, computation, and data efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.553, + 0.226, + 0.57 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Figure 3 shows the overall architecture of ProLearn, a framework that leverages an efficient prototype learning module called Prototype-Driven Semantic Approximation (PSA) alongside a Language-guided U-Net backbone for segmentation. PSA is designed to (i) reduce the need for textual annotations during training by using available text only once to initialize a discrete prototype space and (ii) remove the need for textual input entirely at inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.265, + 0.723 + ], + "angle": 0, + "content": "3.1. PSA Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.852 + ], + "angle": 0, + "content": "The PSA initialization is a one-time process that constructs a queryable prototype space before training. Given that a training set includes \\(K\\) paired image-text samples and other image-only samples, where the \\(i\\)-th sample is denoted as \\(\\langle I_i, T_i \\rangle\\), with \\(I_i\\) as the image and \\(T_i\\) as its associated report. Each paired sample is processed by a trained domain-specific vision-language encoder (BioMedCLIP [45]), \\(f_{\\mathrm{enc}}^I\\) and \\(f_{\\mathrm{enc}}^T\\) to extract their semantic features \\(e_i^I\\) and \\(e_i^T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.858, + 0.484, + 0.878 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {I} = f _ {\\mathrm {e n c}} ^ {I} (I _ {i}), \\quad e _ {i} ^ {T} = f _ {\\mathrm {e n c}} ^ {T} (T _ {i}), \\quad e _ {i} ^ {I}, e _ {i} ^ {T} \\in \\mathbb {R} ^ {D} \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Surrogate Label Extraction: To extract finite" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "segmentation-relevant semantics in the clinical reports, we cluster the image-text pairs into \\(N\\) semantic surrogate labels, as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.907, + 0.333 + ], + "angle": 0, + "content": "As many clinical reports are verbose, we first isolate the tokens most relevant to segmentation. To achieve this, we utilize the cross-attention module of a separately trained Language-guided U-Net, which has the same architecture as the language-guided U-Net in Figure 3, to assess token relevance in all \\( K \\) available image-text pairs \\( \\{\\langle I_i,T_i\\rangle \\mid i = 1,\\dots ,K\\} \\). Specifically, for each text input \\( T_{i} \\) paired with an image \\( I_{i} \\), the cross-attention weight \\( \\alpha_{j} \\) computed for each token \\( t_j \\) in \\( T_{i} \\) indicates the relevance of each token \\( t_j \\) in guiding the segmentation of \\( I_{i} \\). Tokens whose attention scores exceed a threshold \\( \\tau \\) are retained, resulting in a shorter segmentation-relevant sentence for each \\( T_{i} \\), which we denote \\( T_{i}^{\\mathrm{selected}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.34, + 0.905, + 0.358 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {\\text {s e l e c t e d}} = \\left\\{t _ {j} \\mid \\alpha_ {j} > \\tau , t _ {j} \\in T _ {i} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.905, + 0.396 + ], + "angle": 0, + "content": "We then feed \\( T_{i}^{\\mathrm{selected}} \\) into the text encoder \\( f_{\\mathrm{enc}}^{T} \\), obtaining a semantic feature \\( e_{i}^{\\mathrm{sem}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.403, + 0.905, + 0.422 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {\\text {s e m}} = f _ {\\text {e n c}} ^ {T} \\left(T _ {i} ^ {\\text {s e l e c t e d}}\\right), \\quad \\forall i \\in \\{1, \\dots , K \\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.906, + 0.519 + ], + "angle": 0, + "content": "We then group similar textual semantics together using the hierarchical density-based clustering algorithm, HDBSCAN [3]. The clustering results in \\(N\\) surrogate labels \\(\\{l_1,l_2,\\dots ,l_N\\}\\). Each surrogate label \\(l_{i}\\) encapsulates a distinct segmentation-relevant textual semantics, with its corresponding cluster \\(\\mathcal{C}_i\\) representing the grouped features:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.526, + 0.905, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i} = \\left\\{\\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\mid \\operatorname {H D B S C A N} \\left(e _ {i} ^ {\\text {s e m}}\\right) = l _ {i} \\right\\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Prototype Space Construction: While the above surrogate labels capture sparse textual semantics, medical images often convey richer and more fine-grained information. To make the prototype space compact and more representative for image queries, each text-based surrogate label cluster \\(\\mathcal{C}_i\\) is further subdivided into sub-clusters \\(\\mathcal{C}_{ij}\\) via K-means clustering algorithm [22]:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.663, + 0.905, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i j} = \\left\\{\\left\\langle e _ {k} ^ {I}, e _ {k} ^ {T} \\right\\rangle \\mid \\text {K - M e a n} \\left(e _ {k} ^ {I}\\right) = j, \\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\in \\mathcal {C} _ {i} \\right\\}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.855 + ], + "angle": 0, + "content": "For each sub-cluster \\( \\mathcal{C}_{ij} \\), instead of taking the centroid itself, we locate the sample \\( c_{ij} \\) closest to the cluster's centroid and treat it as a fine representation of disease lesion. This approach reduces the influence of outliers that can significantly distort the centroid. Given \\( c_{ij} = \\langle e_k^I, e_k^T \\rangle \\), we assign its containing image feature \\( e_k^I \\) as visual query prototype \\( q_{ij} \\) and text feaître \\( e_k^T \\) as textual respond prototype \\( r_{ij} \\). By repeating this across all sub-clusters, we construct a discrete and compact initial prototype space \\( S \\), consisting of a query space \\( S^Q \\) and a response space \\( S^R \\), both of dimension \\( N \\times M \\times D \\):" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.861, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left(\\mathcal {S} ^ {Q}, \\mathcal {S} ^ {R}\\right) = \\bigcup_ {i = 1} ^ {N} \\bigcup_ {j = 1} ^ {M} \\langle q _ {i j}, r _ {i j} \\rangle . \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22319" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "Each visual query prototype \\(q_{ij}\\) in \\(\\mathcal{S}^Q\\) is directly linked to its corresponding textual response prototype \\(r_{ij}\\) in \\(\\mathcal{S}^R\\):" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.133, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nq _ {i j} \\longrightarrow r _ {i j}, \\quad \\forall i \\in \\{1, \\dots , N \\}, \\quad j \\in \\{1, \\dots , M \\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.19 + ], + "angle": 0, + "content": "This prototype space is dynamically learned during the training process of ProLearn." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.368, + 0.215 + ], + "angle": 0, + "content": "3.2. PSA Querying and Responding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.326 + ], + "angle": 0, + "content": "During both training and inference, the Language-guided U-Net \\( f_{\\mathrm{seg}} \\) queries the query space \\( S^Q \\) using an image feature \\( q^* \\), encoded by the image encoder \\( f_{\\mathrm{enc}}^I \\) from an image input \\( I^* \\), enabling the use of image-only samples. Through the query-and-response mechanism of the PSA module, it responds with an approximate textual semantic feature \\( r^* \\) which guides \\( f_{\\mathrm{seg}} \\) without textual input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.403 + ], + "angle": 0, + "content": "PSA Querying: The PSA querying process involves searching the most relevant visual queries from the query prototype space \\( S^Q \\). Given the encoded image feature \\( q* \\), the PSA module computes the cosine similarity scores \\( s_{ij} \\) between \\( q^* \\) and each query prototype \\( q_{ij} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.412, + 0.483, + 0.446 + ], + "angle": 0, + "content": "\\[\ns _ {i} j = s \\left(q ^ {*}, q _ {i j}\\right) = \\frac {q ^ {*} \\cdot q _ {i j}}{\\| q ^ {*} \\| \\| q _ {i j} \\|} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.483, + 0.484 + ], + "angle": 0, + "content": "After ranking these similarity scores, the PSA module selects the top-\\(k\\) query prototypes \\(Q^{*}\\) that best match \\(q^{*}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.494, + 0.483, + 0.512 + ], + "angle": 0, + "content": "\\[\nQ ^ {*} = \\left\\{q _ {i j} \\mid \\forall i, j \\in \\arg \\operatorname {t o p} _ {k} \\left(s _ {i j}\\right) \\right\\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.567 + ], + "angle": 0, + "content": "PSA Responding: The PSA responding process finds the corresponding response prototypes \\( R^{*} \\) that are linked to the selected query prototypes \\( Q^{*} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.577, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\nR ^ {*} = \\left\\{r _ {i j} \\mid q _ {i j} \\in Q ^ {*}, q _ {i j} \\longrightarrow r _ {i j} \\right\\}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.664 + ], + "angle": 0, + "content": "These response prototypes \\( R^{*} \\) are referred to as candidates in Figure 3b. The candidates are then aggregated using a similarity-weighted sum, where the weights are computed through the softmax function over the similarity scores:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.673, + 0.483, + 0.711 + ], + "angle": 0, + "content": "\\[\nr ^ {*} = \\sum_ {r _ {i} \\in R ^ {*}} w _ {i} r _ {i}, \\quad w _ {i} = \\frac {\\exp \\left(s \\left(q ^ {*} , q _ {i}\\right)\\right)}{\\sum_ {q _ {j} \\in Q ^ {*}} \\exp \\left(s \\left(q ^ {*} , q _ {j}\\right)\\right)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.766 + ], + "angle": 0, + "content": "At this point, the PSA responding process is complete. We then feed \\( r^* \\) into the decoding process of Language-guided U-Net, providing an approximated semantic guidance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.779, + 0.284, + 0.797 + ], + "angle": 0, + "content": "4. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.194, + 0.818 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To evaluate our approach, we used the two well-benchmarked datasets: QaTa-COV19 [10], MosMedData+ [27] and Kvasir-SEG [18], which are commonly adopted for performance comparison in language-guided segmentation [15, 21, 42, 46, 48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "QaTa-COV19 is a large-scale dataset consisting of 9,258 chest X-ray images with manually annotated COVID-19 [41] lesion masks. To facilitate language-guided segmentation, LViT [21] extends this dataset with textual descriptions detailing bilateral lung infections, the number of affected regions, and their spatial localization within the lungs. We adopt the official dataset split: 5,716 images for training, 1,429 for validation, and 2,113 for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.905, + 0.303 + ], + "angle": 0, + "content": "MosMedData+ comprises 2,729 CT slices depicting pulmonary infections. Similar to QaTa-COV19, LViT augments this dataset with text-based annotations to support language-guided segmentation tasks. We adopt the official dataset split: 2,183 slices for training, 273 for validation, and 273 for testing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Kvasir-SEG is a publicly available dataset comprising 1,000 colonoscopy images of gastrointestinal polyps with corresponding pixel-level segmentation masks. We follow a standard 8:1:1 split for training, validation, and testing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.698, + 0.387 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.5 + ], + "angle": 0, + "content": "To quantitatively evaluate segmentation performance, we used the metrics which are used in the previous similar studies [15, 21, 42, 46, 48]: Dice coefficient [11] and the mean Intersection over Union (mIoU [17]), two widely used metrics for measuring spatial overlap between predicted and ground truth segmentation masks. The formulations are defined in Equations 12 and 13." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.51, + 0.905, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\operatorname {D i c e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\sum_ {c = 1} ^ {C} \\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\sum_ {c = 1} ^ {C} \\left(\\left| P _ {i} ^ {(c)} \\right| + \\left| G _ {i} ^ {(c)} \\right|\\right)} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.563, + 0.905, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\mathrm {m I o U} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {\\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\left| P _ {i} ^ {(c)} \\cup G _ {i} ^ {(c)} \\right|} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.905, + 0.675 + ], + "angle": 0, + "content": "where \\(N\\) represents the number of images in the dataset, \\(C\\) denotes the number of semantic categories, and \\(P_{i}^{(c)}\\) and \\(G_{i}^{(c)}\\) correspond to the predicted and ground truth segmentation masks for class \\(c\\) in image \\(i\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.712, + 0.699 + ], + "angle": 0, + "content": "4.3. Experimental Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.75 + ], + "angle": 0, + "content": "To evaluate ProLearn, we compare it against both multimodal and unimodal segmentation methods, analyzing its performance under various realistic clinical scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Limited availability of text: To illustrate the limitations of language-guided segmentation methods that require strict image-report pairing and highlight the importance of training with both image-text-paired and image-only data, we simulate a real-world clinical scenario where detailed radiology reports are often unavailable for a significant portion of the dataset. Due to the lack of large-scale benchmarks containing both paired and unpaired examples, we construct a counterfactual setting by progressively reducing the proportion of image-text pairs in the training data to \\(50\\%\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22320" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.089, + 0.858, + 0.295 + ], + "angle": 0, + "content": "
DatasetModelDicemIoU
50%25%10%5%1%50%25%10%5%1%
QaTa-COV19 [10]LViT [21]0.84160.82020.80040.76380.70060.73200.70120.67470.62480.5490
GuideSeg [48]0.86330.85240.84020.82400.73330.75950.74280.72440.70070.5789
SGSeg [42]0.86410.85740.84230.80570.73070.76070.75040.72760.67460.5757
ProLearn0.86670.85980.85830.85730.85660.77210.76900.75730.75580.7553
MosMedData+ [27]LViT [21]0.71890.66080.55010.50690.16770.56960.50420.41080.35380.1015
GuideSeg [48]0.75080.73930.68980.63750.42350.60890.58640.52650.46780.2686
SGSeg [42]0.74550.74390.69500.64650.34520.59430.59220.53250.47760.2086
ProLearn0.75390.75120.74240.73790.72180.61260.61090.60870.60690.6032
Kvasir-SEG [18]LViT [21]0.76690.64240.57190.54820.42720.62280.47690.40250.37820.2729
GuideSeg [48]0.88480.83900.77540.74970.56150.79390.72280.63600.60430.4008
SGSeg [42]0.87690.83040.80250.75260.54060.78080.70990.67020.60340.3705
ProLearn0.89830.89460.88980.88230.87180.81620.81010.80200.79050.7729
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.907, + 0.347 + ], + "angle": 0, + "content": "Table 1. Performance comparison of language-guided segmentation models in simulated scenarios with limited text supervision on the QaTa-COV19, MosMedData+ and Kvasir-SEG dataset. The best results are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.434 + ], + "angle": 0, + "content": "25%, 10%, 5%, and 1%, while discarding the remaining unpaired images to simulate their inaccessibility. We compare ProLearn against state-of-the-art language-guided segmentation models: LViT [21], GuideSeg [48], and SGSeg [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.6 + ], + "angle": 0, + "content": "Segmentation in real-world (image-only) setting: In the majority of clinical scenarios, such as real-time procedural guidance and decision support, segmentation is used without text. To align with real-world settings, we evaluate ProLearn in a strictly \"image-only\" setting, where no text input is provided during either training or inference on the target datasets. Under these conditions, we compare ProLearn to established unimodal segmentation methods and vision-language pretraining models (CLIP [31] and GLoRIA [16]) adapted for image-only use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.721 + ], + "angle": 0, + "content": "Prototype vs. LLM: To demonstrate our PSA's advantages against LLMs in real-world deployment, we focus on inference speed, model size, and time complexity. We specifically compare ProLearn with SGSeg [42], a recent method that leverages LLMs (e.g., GPT-2 [30], Llama3 [29]) to generate synthetic textual reports at inference in order to compensate for missing textual input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.812 + ], + "angle": 0, + "content": "Qualitative Analysis: We provide visual comparisons against state-of-the-art language-guided segmentation models under progressively lower text availability settings. Specifically, We visualize both final segmentation outputs and corresponding saliency maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Hyperparameter Sensitivity Analysis: We analyze how model performance varies as we adjust two key hyperparameters: the number of prototypes \\( M \\), which governs the compactness of the prototype space, and the number of candidate responses \\( k \\)." + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.371, + 0.718, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.371, + 0.897, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.481, + 0.718, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.481, + 0.897, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.593, + 0.718, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.594, + 0.894, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Figure 5. Performance degradation as the ratio of text availability decreases. The plots show Dice and nIoU metrics for a) the QaTa-COV19, b) the MosMedData+ and c) the Kvasir-SEG dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.73, + 0.792 + ], + "angle": 0, + "content": "5. Results and Discussion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.893, + 0.818 + ], + "angle": 0, + "content": "5.1. Comparison with State-Of-The-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Table 1 presents our comparative results under clinical scenarios where paired textual reports are limited (50%, 25%, 10%, 5%, and 1%). ProLearn achieved higher Dice and mIoU scores than existing language-guided methods in all settings, and this performance gap increased as the frac" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22321" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.493, + 0.284 + ], + "angle": 0, + "content": "
ModelQaTa-COV19MosMedData+Kvasir-SEG
DicemIoUDicemIoUDicemIoU
U-Net [32]0.8190.6920.6380.5050.1950.182
U-Net++ [49]0.8230.7060.7140.5820.2800.180
Attention U-Net [28]0.8220.7010.6640.5280.3640.226
Trans U-Net [8]0.8060.6870.7020.5750.0480.100
Swin U-Net [5]0.8360.7240.6690.5310.3980.246
CLIP [31]0.7980.7070.7200.596--
GLoRIA [16]0.7990.7070.7220.602--
MedSAM [24]0.7300.6190.5090.371--
BiomedParse [47]0.7810.6820.6710.5530.8280.721
ProLearn (1%)0.8570.7550.7220.6030.8720.773
ProLearn (5%)0.8570.7560.7380.6070.8820.790
ProLearn (10%)0.8580.7570.7420.6090.8890.802
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.485, + 0.361 + ], + "angle": 0, + "content": "Table 2. Performance comparison of methods in image-only settings where paired reports are excluded entirely from the target datasets' training and inference. The best results are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.567 + ], + "angle": 0, + "content": "tion of available text decreased. For example, in the \\(1\\%\\) text in MosMedData+, ProLearn retained a Dice score of 0.7218, compared to SGSeg 0.3452 and LViT 0.1677. This trend underscores the effectiveness of ProLearn in learning from image-text and image-only data, allowing it to outperform approaches that require textual input for both training and inference. Although SGSeg also requires no text at inference, its performance remained below ProLearn due to its limited capacity to exploit learned semantics when text is scarce. Further discussions on the comparison between SGSeg and ProLearn can be found in Section 5.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.484, + 0.649 + ], + "angle": 0, + "content": "The performance degradation analysis, visualized in Figure 5, further illustrates the importance of reducing textual reliance in training. ProLearn exhibited the most minor degradation, preserving robust performance even under severely limited text supervision." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.694 + ], + "angle": 0, + "content": "5.2. Comparison in real-world (image-only) setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further evaluated the effectiveness of ProLearn's use of a target dataset's auxiliary text under near \"image-only\" conditions, as most clinical workflows require segmentation models operating without textual input. As shown in Table 2, ProLearn outperformed both unimodal approaches (U-Net, Attention U-Net, Swin U-Net, etc.) and well-recognized VLP methods (CLIP, GLoRIA), even when only \\(10\\%\\), \\(5\\%\\), or \\(1\\%\\) of the paired reports were available. This demonstrates the importance of textual semantics within the target dataset. Unlike unimodal and VLP-based methods, which fail to leverage clinical reports during fine-tuning on target datasets, ProLearn incorporates available image-report pairs in the learning process." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.09, + 0.905, + 0.158 + ], + "angle": 0, + "content": "
ProLearnSGSeg
MethodQuery & respond from prototype matrixGenerate text tokens one by one (autoregressive) & text embedding
No. of inferenceSingle inferenceMultiple inference
TimeO(1)O(n) depends on the no. of tokens
" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.161, + 0.697, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.161, + 0.897, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.266, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 6. Comparison between the proposed prototype-driven approach, ProLearn, and the LLM-based approach, SGSeg. Top: The theoretical analysis of time complexity. Bottom: The model size and inference time comparison." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.361, + 0.699, + 0.377 + ], + "angle": 0, + "content": "5.3. Prototype vs. LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.907, + 0.618 + ], + "angle": 0, + "content": "Prototype learning enables constant-time inference of \\(\\mathcal{O}(1)\\) because it uses a pre-defined, finite prototype space that is independent of both image and text size. At the same time, LLM applies autoregressive image-to-text generation, which has a linear time complexity \\(\\mathcal{O}(n)\\), where \\(n\\) represents the number of tokens in the generated text (see Figure 6, top). As shown in Figure 6, bottom, prototype learning achieves an inference time of \\(4\\mathrm{ms}\\), making it \\(100\\times\\) faster than GPT-2 (136ms) and \\(300\\times\\) faster than Llama3 (1.2s). Moreover, the prototype model has only 1M parameters, which is \\(1000\\times\\) smaller than large language models such as GPT-2 (1.5B parameters) and Llama3 (7B parameters). These properties make prototype learning highly efficient for real-time applications and suitable for deployment on resource-constrained edge devices." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.652, + 0.661 + ], + "angle": 0, + "content": "5.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Figure 7 presents a qualitative comparison of ProLearn with state-of-the-art approaches on the QaTa-COV19, MosMedData+ and Kvasir-SEG datasets. As textual availability decreased, existing language-guided segmentation models, which rely heavily on image-text pairs, experienced a pronounced drop in performance. In contrast, ProLearn effectively learned from both image-text and image-only data, demonstrating stable results even under limited textual guidance. The same findings can be observed in saliency maps that pinpoint the regions on which each model was focused. Under decreased text guidance, conventional models produced unstable or diffuse attention patterns, compromising their segmentation accuracy. However, ProLearn, aided by semantic approximation, retained more coherent attention focused on areas of the lesion." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22322" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.907, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 7. Visualization comparison of language-guided segmentation methods under different text availabilities. The upper row shows segmentation outputs on CXR_S09687-E18404-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_4_85506_1(case19_16 from the MosMedData+ dataset (middle) and ju5xkwzxf0z0818gk4xabdm from the Kvasir-SEG dataset (right). Blue regions indicate groundtruth pixels missed by the model, while red regions indicate pixels mistakenly predicted as lesion. The lower row presents saliency map interpretability on CXR_S09346-E23164-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_40_86625_0(case18_53 from the MosMedData+ dataset (middle) and cju2tzypl4wss0799ow05oxb9 from the Kvasir-SEG dataset (right)." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.46, + 0.287, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.459, + 0.474, + 0.554 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.621 + ], + "angle": 0, + "content": "Figure 8. Hyperparameter sensitivity analysis with varying candidate and prototype configurations. a) Effect of varying \\( k \\), number of responding prototype vectors on Dice score; b) Effect of varying \\( M \\), number of prototypes per surrogate label on Dice score." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.342, + 0.667 + ], + "angle": 0, + "content": "5.5. Hyperparameter Sensitivity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.793 + ], + "angle": 0, + "content": "To investigate the effect of the number of candidates per response \\( k \\) and the number of prototypes per surrogate label \\( M \\), we presented the performance comparison by varying their values for segmentation, as shown in Figure 8. We adjusted one hyperparameter at a time while keeping the other fixed. The experimental results indicated that our framework remained stable across a broad range of hyperparameter values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Specifically, \\( k \\) governs the trade-off between information diversity and noise. A small \\( k \\) limits the model's ability to capture subtle segmentation patterns due to insufficient variability. As \\( k \\) increases, richer patterns are incorporated, improving performance. However, when \\( k \\) is too large, irrelevant patterns introduce excessive noise, leading to degraded segmentation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.583 + ], + "angle": 0, + "content": "For \\( M \\), increasing the number of prototypes enhances performance by capturing diverse segmentation characteristics. However, when \\( M \\) exceeds a certain threshold, prototype redundancy arises, leading to overlapping or irrelevant representations that blur the segmentation boundaries. Thus, balancing \\( M \\) is crucial; too few prototypes hinder representational capacity, while an excessive count dilutes the model's focus on meaningful patterns." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.6, + 0.634, + 0.616 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.824 + ], + "angle": 0, + "content": "We investigated the issue of textual reliance in medical language-guided segmentation, which limits both the applicability of segmentation in clinical workflows and the exploration of image-only training data. To address this issue, we have presented ProLearn, a prototype-driven framework that fundamentally alleviates textual reliance in medical language-guided segmentation. Our experiments demonstrated that ProLearn effectively learns from both image-only and image-text data, making efficient use of the target dataset's textual descriptions. ProLearn showed stable performance under different text availability and outperformed state-of-the-art image-only approaches under minimal text supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Limitations and Future Works: In this work, we focused on demonstrating the core idea of ProLearn in single-label 2D segmentation. Future directions involve exploring multi-label segmentation, broader imaging modalities, and other language-guided vision tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22323" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.482, + 0.145 + ], + "angle": 0, + "content": "This work was supported in part by the Australian Research Council (ARC) under Grant IM240100224." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.188, + 0.175 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.184, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10076-10095, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.269, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[2] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21438-21451, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[3] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[4] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE transactions on medical imaging, 33(2):577-590, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[5] H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.536, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.632 + ], + "angle": 0, + "content": "[7] Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):1655-1674, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.483, + 0.689 + ], + "angle": 0, + "content": "[8] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[9] Pujin Cheng, Li Lin, Junyan Lyu, Yijin Huang, Wenhan Luo, and Xiaoying Tang. Prior: Prototype representation joint learning from medical images and reports. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21361-21371, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[10] Aysen Degerli, Serkan Kiranyaz, Muhammad E. H. Chowdhury, and Moncef Gabbouj. Osegnet: Operational segmentation network for Covid-19 detection using chest x-ray images. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2306–2310, 2022. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[11] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Vincenzo Ferrari, Marina Carbone, Carla Cappelli, Luigi Boni, Franca Melfi, Mauro Ferrari, Franco Mosca, and Andrea Pietrabissa. Value of multidetector computed tomogra" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.184, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "phy image segmentation for preoperative planning in general surgery. Surgical endoscopy, 26:616-626, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[13] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7606-7623, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[14] Matthew Stephen Holden, Tamas Ungi, Derek Sargent, Robert C McGraw, Elvis CS Chen, Sugantha Ganapathy, Terry M Peters, and Gabor Fichtinger. Feasibility of real-time workflow segmentation for tracked needle interventions. IEEE Transactions on Biomedical Engineering, 61(6): 1720-1728, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[15] Jihong Hu, Yinhao Li, Hao Sun, Yu Song, Chujie Zhang, Lanfen Lin, and Yen-Wei Chen. Lga: A language guide adapter for advancing the sam model's capabilities in medical image segmentation. page 610–620, Berlin, Heidelberg, 2024. Springer-Verlag. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[16] Shih-Cheng Huang, Liyue Shen, Matthew P. Lungren, and Serena Yeung. Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3922-3931, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Paul Jaccard. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547-579, 1901. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In International conference on multimedia modeling, pages 451-462. Springer, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[19] Tina Kapur, Jan Egger, Jagadeesan Jayender, Matthew Toews, and William M Wells. Registration and segmentation for image-guided therapy. Intraoperative Imaging and Image-Guided Therapy, pages 79-91, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.675 + ], + "angle": 0, + "content": "[20] Stijn Keereweer, Jeroen DF Kerrebijn, Pieter BAA Van Driel, Bangwen Xie, Eric L Kaijzel, Thomas JA Snoeks, Ivo Que, Merlijn Hutteman, Joost R Van Der Vorst, J Sven D Mieog, et al. Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199-207, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[21] Zihan Li, Yunxiang Li, Qingde Li, Puyang Wang, Dazhou Guo, Le Lu, Dakai Jin, You Zhang, and Qingqi Hong. Lvit: language meets vision transformer in medical image segmentation. IEEE transactions on medical imaging, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[22] S. Lloyd. Least squares quantization in pcm. IEEE Transactions on Information Theory, 28(2):129-137, 1982. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.908, + 0.831 + ], + "angle": 0, + "content": "[23] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Dhiraj Maji, Prarthana Sigedar, and Munendra Singh. Attention res-unet with guided decoder for semantic segmentation" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22324" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "of brain tumors. Biomedical Signal Processing and Control, 71:103077, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[26] Epimack Michael, He Ma, Hong Li, Frank Kulwa, and Jing Li. Breast cancer segmentation methods: current status and future potentials. *BioMed research international*, 2021(1): 9962109, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Sergey Morozov, Anna Andreychenko, Nikolay Pavlov, Anton Vladzymyrskyy, Natalya Ledikhova, Victor Gombolevskiy, Ivan Blokhin, Pavel Gelezhe, Anna Gonchar, Valeria Chernina, et al. Mosmeddata: Chest ct scans with Covid-19 related findings. 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, and et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards building multilingual language model for medicine. Nature Communications, 15(1):8384, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[30] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[32] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[33] N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031-82057, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[35] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[36] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[37] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Sudanthi Wijewickrema, Yun Zhou, James Bailey, Gregor Kennedy, and Stephen O'Leary. Provision of automated step-by-step procedural guidance in virtual reality surgery" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "simulation. In Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology, pages 69-72, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[39] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503-17512, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[40] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[41] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[42] Shuchang Ye, Mingyuan Meng, Mingjian Li, Dagan Feng, and Jinman Kim. Enabling text-free inference in language-guided segmentation of chest x-rays via self-guidance. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[43] Xiao-Xia Yin, Le Sun, Yuhan Fu, Ruiliang Lu, and Yanchun Zhang. [retracted] u-net-based medical image segmentation. Journal of healthcare engineering, 2022(1):4189781, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[44] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[45] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Angela Crabtree, Brian Piening, Carlo Bifulco, Matthew P. Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon. A multimodal biomedical foundation model trained from fifteen million image-text pairs. NEJM AI, 2(1):A1oa2400640, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[46] Xu Zhang, Bo Ni, Yang Yang, and Lefei Zhang. Madapter: A better interaction between image and language for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[47] Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Hin Lee, Ho Sid Kiblawi, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, Brian Piening, Carlo Bifulco, Mu Wei, Hoifung Poon, and Sheng Wang. A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities. Nature Methods, 22(1):166-176, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[48] Yi Zhong, Mengqiu Xu, Kongming Liang, Kaixin Chen, and Ming Wu. Ariadne's thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In International Conference on Medical Image Computing and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22325" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Computer-Assisted Intervention, pages 724-733. Springer, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[49] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22326" + } + ] +] \ No newline at end of file diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_origin.pdf b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d256daa4bdd6e1e2be0ba8bbbdc87d2cbd59a94 --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/bb0e3137-699a-4301-8697-612d0c48d0ca_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41ec49458ad2fa04931ff3ff180d8d232a3b39defa52855e02079779a1777220 +size 2655812 diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/full.md b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c4a7990fd155cccc8488f005c7bc2c0998a935a9 --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/full.md @@ -0,0 +1,339 @@ +# Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation + +Shuchang Ye1 Usman Naseem2 Mingyuan Meng1 Jinman Kim1 +1The University of Sydney, Sydney, Australia +2Macquarie University, Sydney, Australia + +shuchang.ye@sydney.edu.au + +usman.naseem@mq.edu.au + +mmen2292@uni.sydney.edu.au + +jinman.kim@sydney.edu.au + +# Abstract + +Medical language-guided segmentation, integrating textual clinical reports as auxiliary guidance to enhance image segmentation, has demonstrated significant improvements over unimodal approaches. However, its inherent reliance on paired image-text input, which we refer to as "textual reliance", presents two fundamental limitations: 1) many medical segmentation datasets lack paired reports, leaving a substantial portion of image-only data underutilized for training; and 2) inference is limited to retrospective analysis of cases with paired reports, limiting its applicability in most clinical scenarios where segmentation typically precedes reporting. To address these limitations, we propose ProLearn, the first Prototype-driven Learning framework for language-guided segmentation that fundamentally alleviates textual reliance. At its core, we introduce a novel Prototype-driven Semantic Approximation (PSA) module to enable approximation of semantic guidance from textual input. PSA initializes a discrete and compact prototype space by distilling segmentation-relevant semantics from textual reports. Once initialized, it supports a query-and-respond mechanism which approximates semantic guidance for images without textual input, thereby alleviating textual reliance. Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms state-of-the-art language-guided methods when limited text is available. The code is available at https://github.com/ShuchangYe-bib/ProLearn. + +# 1. Introduction + +Segmentation is an essential tool in medical image analysis, supporting clinical workflows by enabling precise delineation of anatomical structures, identification of pathological regions, and facilitating targeted interventions [24]. Its applications span critical tasks such as disease diag + +![](images/21b6ad183903f790dc44cbaaf35c6381287e5acddb4d911f27d7f48b0511f5d5.jpg) +Figure 1. Comparison of vision-language paradigms for medical image segmentation. a) VLP: Pretrained on large-scale image-text pairs and then fine-tuned on image-only data from the target dataset. b) Language-guided Segmentation: Requires strict image-text pairs during both training and inference. c) ProLearn: initializes with a limited amount of paired image-text data to construct query-response spaces. After initialization, it enables learning with limited textual input and performs inference without text. + +nosis, treatment planning, and surgery support [1]. Deep learning has revolutionized segmentation, making it more accurate, reliable, and widely applicable in clinical practice. Unimodal (image-only) segmentation methods [43], such as U-Net [32] and its extensions [33], including U-Net++ [49], Attention U-Net [28], and Trans U-Net [8], have been widely adopted in medical imaging. In recent years, multimodal segmentation methods that leverage textual clinical reports as complementary information have gained wide attention for their ability to transcend the performance limits of unimodal segmentation [6, 13, 44]. + +The exploration of multimodal segmentation [6, 13, 44] began with vision-language pretraining (VLP) [31], as shown in Figure 1a, where models pre-trained on paired textual and visual data demonstrated improved visual understanding and performance when finetuned on target image-only segmentation datasets [23, 37, 39]. How + +![](images/4427badb1869cef6a057c0149df79cb47e5a02d5d4ecd7fef0cf77fab078b12c.jpg) +Figure 2. Flow diagram comparison between SGSeg [42] and our proposed ProLearn. a) SGSeg: image $\rightarrow$ LLM $\rightarrow$ report $\rightarrow$ BERT $\rightarrow$ embedding. b) ProLearn: image $\rightarrow$ PSA $\rightarrow$ embedding. + +![](images/a7e559e8b4d082ce9cfdfc13e14b5458e020992dcdb9f18f8d66c1863b5e6479.jpg) + +ever, these pretraining-based approaches fail to fully exploit the disease-specific information embedded in the target dataset's reports, as the general knowledge learned during pretraining often lacks the domain-specific details essential for precise disease lesion segmentation. To address this limitation, language-guided segmentation has been proposed as a promising approach [15, 21, 46, 48], which takes textual clinical report inputs as auxiliary semantic guidance for image segmentation (see Figure 1b). It has achieved remarkable success by leveraging textual abnormality description as explicit semantic guidance to segmentation, thus outperforming pretraining-based approaches and setting new benchmarks in medical image segmentation. + +However, language-guided segmentation methods have inherent reliance on paired image-text data due to the requirement for textual reports as auxiliary inputs. This reliance is referred to as textual reliance in this study and has incurred notable limitations during the training and inference stages. In the training stage, its reliance on paired image-text data leads to the underutilization of image-only data, which constitutes the majority of available datasets [2]. In the inference stage, reliance on textual descriptions during inference confines its use to retrospective analysis, misaligning with most clinical workflows where segmentation is required preemptively for tasks such as preoperative planning [12], diagnostic decision making [4, 25, 26], and real-time procedural guidance [14, 19, 38]. + +Our previous study, Self-Guided Segmentation (SGSeg) [42], made a preliminary attempt to eliminate textural reliance during inference in language-guided segmentation by leveraging large language models (LLMs) to generate synthetic reports to compensate for the missing textual input, as illustrated in Figure 2a. However, the integration of LLMs substantially increases the model size and inference time, making the approach unsuitable for deployment on edge devices [7] or for real-time applications such as image-guided surgery [20]. Moreover, textual reliance during training remains an open issue. + +We argue that the critical guidance in language-guided segmentation is not the entire clinical report, which is often verbose and cluttered with irrelevant information, but + +rather the specific segmentation-relevant semantic features embedded within it. Our investigation further indicates that the semantic space of medical reports is inherently constrained, consisting of a finite set of distinct representations relevant to segmentation tasks, as clinical reports typically adhere to standardized medical terminologies, which result in a relatively closed vocabulary. + +Building on these insights, we propose ProLearn (see Figure 1c), a lightweight and efficient Prototype Learning framework that fundamentally alleviates textual reliance during both training and inference. ProLearn introduces a Prototype-Driven Semantic Approximation (PSA) module, which enables the model to approximate semantic guidance without the need for textual input. Specifically, PSA constructs a discrete and compact prototype space by distilling segmentation-relevant concepts from clinical reports. It then provides a query-and-respond mechanism to support interaction between segmentation models and the prototype space, where unseen semantics are approximated by weighted aggregation of the existing prototypes based on similarity. Therefore, PSA enables segmentation models to query by image feature and receive responded semantics feature as guidance for feature maps refinement, as illustrated in Figure 2b. + +The main contributions of this work are as follows: + +- To the best of our knowledge, our proposed ProLearn is the first work to alleviate textual reliance in medical language-guided segmentation in both training and inference. +- We introduce a novel PSA module that supports learning with both paired image-text data and image-only data while enabling inference without textual input by querying a learned prototype space to provide semantic guidance for segmentation. +- Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms language-guided segmentation methods in limited text availability settings (see Section 5.1) and surpasses state-of-the-art unimodal segmentation models (see Section 5.2). Compared to SGSeg, ProLearn achieves a $1000 \times$ reduction in parameters and $100 \times$ faster inference speed (see Section 5.3). + +# 2. Related Work + +# 2.1. Language-guided Segmentation + +Language-guided segmentation aims to address the gap that the target dataset's reports are not fully exploited for learning in conventional unimodal [5, 8, 28, 32, 49] or VLP methods [23, 31, 37, 39]. LViT [21] was the first work in language-guided segmentation, which takes both images and textual reports as input to train a multimodal-input segmentation model. They annotated existing pub- + +![](images/2fde88a958ef02a100a8a3c17c1d4a7495d147678cba75054ddc56d658a77a1d.jpg) +Figure 3. The overview of the ProLearn framework for alleviating textual reliance in medical language-guided segmentation. a) Language-guided U-Net: The U-Net encoder extracts image features, which are used to query the PSA module. The decoder is then guided by PSA's response to decode the segmentation mask. b) PSA Initialization: Clinical reports are processed and abstracted into a discrete prototype space, representing segmentation-relevant semantic and spatial information. c) PSA Querying and Responding: PSA receives image feature queries, selects relevant prototype candidates, and responds with an aggregated representation to approximate the pathological region's segmentation-relevant semantic representation. + +licly available segmentation datasets, QaTa-COV19 [10] and MosMedData+ [27], with corresponding clinical reports. To fuse the text features with feature maps in the U-Net, LViT adopted an early fusion approach, which introduced a Pixel-Level Attention Module (PLAM) to involve textual features as semantic guidance. LViT showed consistent performance over image-only and VLP segmentation methods, which demonstrated the importance and potential of textual semantic guidance from clinical reports in target datasets. + +Subsequent methods proposed more flexible and robust feature fusion modules for language-guided segmentation. GuideSeg [48] moved away from early fusion while adopting a late-fusion strategy that fused textual and visual features at the decoding stage, where the text features were better preserved and more effectively influenced the segmentation process. MA adaptor [46] addressed the unidirectional flow of textual semantics seen in previous frameworks. It introduced a bidirectional adaptor connecting multiple layers of unimodal encoders, facilitating mutual information exchange between text and image representations at various scales. LGA [15] adopted a parameter-efficient fin-tuning strategy that preserved the original parameters of + +large segmentation foundation models [24]. These works further prove that integrating target datasets' textual reports as guidance can significantly improve the performance of segmentation models. + +These methods suffered from textual reliance. The reliance during training restricted the utilization of large portions of medical datasets that lacked paired textual reports. The reliance during inference limits their practical utility in most clinical scenarios using segmentation without reports. SGSeg [42] attempted to release the textual reliance at inference by training a LLM [35] to generate needed clinical reports from images. Nevertheless, the inclusion of LLMs increased model size and inference time, making it unsuitable for edge devices and real-time applications. The textual reliance during training still remained unsolved. Our proposed ProLearn fundamentally alleviates textual reliance in training and inference, and its prototype design PSA significantly reduces the number of parameters and inference time compared to LLM-based SGSeg. + +# 2.2. Prototype Learning + +Prototype learning draws on the principle that images and text can be effectively captured in discrete prototype repre + +![](images/cab19660a2132f4f0970e8b50b497a84e4517b12d08c682f03b73f49d5391cb4.jpg) +Figure 4. Attention-guided surrogate label extraction. + +sentations rather than relying on freshly built embeddings for every instance [36]. Each prototype functions as a canonical reference, reflecting the typical features of its corresponding class. Such a strategy is widely adopted in classification [40], where an unseen query is associated with the class whose prototype lies closest in a learned space. By prioritizing the reuse of well-established representations, prototype-based methods frequently attain top-tier performance under data-scarce settings [34]. Beyond unimodal applications, prototype learning has demonstrated its potential in multimodal tasks, effectively aligning features across modalities [9], such as images and text. Inspired by the principles of expressing unseen queries via finite existing prototypes, the proposed ProLearn extracts semantic information from textual reports as prototypes, enabling representing semantic guidance from existing image-text pairs for image-only input during training and inference, with improved parameter, computation, and data efficiency. + +# 3. Methodology + +Figure 3 shows the overall architecture of ProLearn, a framework that leverages an efficient prototype learning module called Prototype-Driven Semantic Approximation (PSA) alongside a Language-guided U-Net backbone for segmentation. PSA is designed to (i) reduce the need for textual annotations during training by using available text only once to initialize a discrete prototype space and (ii) remove the need for textual input entirely at inference. + +# 3.1. PSA Initialization + +The PSA initialization is a one-time process that constructs a queryable prototype space before training. Given that a training set includes $K$ paired image-text samples and other image-only samples, where the $i$ -th sample is denoted as $\langle I_i, T_i \rangle$ , with $I_i$ as the image and $T_i$ as its associated report. Each paired sample is processed by a trained domain-specific vision-language encoder (BioMedCLIP [45]), $f_{\mathrm{enc}}^I$ and $f_{\mathrm{enc}}^T$ to extract their semantic features $e_i^I$ and $e_i^T$ : + +$$ +e _ {i} ^ {I} = f _ {\mathrm {e n c}} ^ {I} (I _ {i}), \quad e _ {i} ^ {T} = f _ {\mathrm {e n c}} ^ {T} (T _ {i}), \quad e _ {i} ^ {I}, e _ {i} ^ {T} \in \mathbb {R} ^ {D} \quad (1) +$$ + +Surrogate Label Extraction: To extract finite + +segmentation-relevant semantics in the clinical reports, we cluster the image-text pairs into $N$ semantic surrogate labels, as shown in Figure 4. + +As many clinical reports are verbose, we first isolate the tokens most relevant to segmentation. To achieve this, we utilize the cross-attention module of a separately trained Language-guided U-Net, which has the same architecture as the language-guided U-Net in Figure 3, to assess token relevance in all $K$ available image-text pairs $\{\langle I_i,T_i\rangle \mid i = 1,\dots ,K\}$ . Specifically, for each text input $T_{i}$ paired with an image $I_{i}$ , the cross-attention weight $\alpha_{j}$ computed for each token $t_j$ in $T_{i}$ indicates the relevance of each token $t_j$ in guiding the segmentation of $I_{i}$ . Tokens whose attention scores exceed a threshold $\tau$ are retained, resulting in a shorter segmentation-relevant sentence for each $T_{i}$ , which we denote $T_{i}^{\mathrm{selected}}$ : + +$$ +T _ {i} ^ {\text {s e l e c t e d}} = \left\{t _ {j} \mid \alpha_ {j} > \tau , t _ {j} \in T _ {i} \right\}. \tag {2} +$$ + +We then feed $T_{i}^{\mathrm{selected}}$ into the text encoder $f_{\mathrm{enc}}^{T}$ , obtaining a semantic feature $e_{i}^{\mathrm{sem}}$ : + +$$ +e _ {i} ^ {\text {s e m}} = f _ {\text {e n c}} ^ {T} \left(T _ {i} ^ {\text {s e l e c t e d}}\right), \quad \forall i \in \{1, \dots , K \} \tag {3} +$$ + +We then group similar textual semantics together using the hierarchical density-based clustering algorithm, HDBSCAN [3]. The clustering results in $N$ surrogate labels $\{l_1,l_2,\dots ,l_N\}$ . Each surrogate label $l_{i}$ encapsulates a distinct segmentation-relevant textual semantics, with its corresponding cluster $\mathcal{C}_i$ representing the grouped features: + +$$ +\mathcal {C} _ {i} = \left\{\left\langle e _ {j} ^ {I}, e _ {j} ^ {T} \right\rangle \mid \operatorname {H D B S C A N} \left(e _ {i} ^ {\text {s e m}}\right) = l _ {i} \right\}. \tag {4} +$$ + +Prototype Space Construction: While the above surrogate labels capture sparse textual semantics, medical images often convey richer and more fine-grained information. To make the prototype space compact and more representative for image queries, each text-based surrogate label cluster $\mathcal{C}_i$ is further subdivided into sub-clusters $\mathcal{C}_{ij}$ via K-means clustering algorithm [22]: + +$$ +\mathcal {C} _ {i j} = \left\{\left\langle e _ {k} ^ {I}, e _ {k} ^ {T} \right\rangle \mid \text {K - M e a n} \left(e _ {k} ^ {I}\right) = j, \left\langle e _ {j} ^ {I}, e _ {j} ^ {T} \right\rangle \in \mathcal {C} _ {i} \right\}. \tag {5} +$$ + +For each sub-cluster $\mathcal{C}_{ij}$ , instead of taking the centroid itself, we locate the sample $c_{ij}$ closest to the cluster's centroid and treat it as a fine representation of disease lesion. This approach reduces the influence of outliers that can significantly distort the centroid. Given $c_{ij} = \langle e_k^I, e_k^T \rangle$ , we assign its containing image feature $e_k^I$ as visual query prototype $q_{ij}$ and text feaître $e_k^T$ as textual respond prototype $r_{ij}$ . By repeating this across all sub-clusters, we construct a discrete and compact initial prototype space $S$ , consisting of a query space $S^Q$ and a response space $S^R$ , both of dimension $N \times M \times D$ : + +$$ +\mathcal {S} = \left(\mathcal {S} ^ {Q}, \mathcal {S} ^ {R}\right) = \bigcup_ {i = 1} ^ {N} \bigcup_ {j = 1} ^ {M} \langle q _ {i j}, r _ {i j} \rangle . \tag {6} +$$ + +Each visual query prototype $q_{ij}$ in $\mathcal{S}^Q$ is directly linked to its corresponding textual response prototype $r_{ij}$ in $\mathcal{S}^R$ : + +$$ +q _ {i j} \longrightarrow r _ {i j}, \quad \forall i \in \{1, \dots , N \}, \quad j \in \{1, \dots , M \}. \tag {7} +$$ + +This prototype space is dynamically learned during the training process of ProLearn. + +# 3.2. PSA Querying and Responding + +During both training and inference, the Language-guided U-Net $f_{\mathrm{seg}}$ queries the query space $S^Q$ using an image feature $q^*$ , encoded by the image encoder $f_{\mathrm{enc}}^I$ from an image input $I^*$ , enabling the use of image-only samples. Through the query-and-response mechanism of the PSA module, it responds with an approximate textual semantic feature $r^*$ which guides $f_{\mathrm{seg}}$ without textual input. + +PSA Querying: The PSA querying process involves searching the most relevant visual queries from the query prototype space $S^Q$ . Given the encoded image feature $q*$ , the PSA module computes the cosine similarity scores $s_{ij}$ between $q^*$ and each query prototype $q_{ij}$ : + +$$ +s _ {i} j = s \left(q ^ {*}, q _ {i j}\right) = \frac {q ^ {*} \cdot q _ {i j}}{\| q ^ {*} \| \| q _ {i j} \|} \tag {8} +$$ + +After ranking these similarity scores, the PSA module selects the top- $k$ query prototypes $Q^{*}$ that best match $q^{*}$ : + +$$ +Q ^ {*} = \left\{q _ {i j} \mid \forall i, j \in \arg \operatorname {t o p} _ {k} \left(s _ {i j}\right) \right\}. \tag {9} +$$ + +PSA Responding: The PSA responding process finds the corresponding response prototypes $R^{*}$ that are linked to the selected query prototypes $Q^{*}$ : + +$$ +R ^ {*} = \left\{r _ {i j} \mid q _ {i j} \in Q ^ {*}, q _ {i j} \longrightarrow r _ {i j} \right\}. \tag {10} +$$ + +These response prototypes $R^{*}$ are referred to as candidates in Figure 3b. The candidates are then aggregated using a similarity-weighted sum, where the weights are computed through the softmax function over the similarity scores: + +$$ +r ^ {*} = \sum_ {r _ {i} \in R ^ {*}} w _ {i} r _ {i}, \quad w _ {i} = \frac {\exp \left(s \left(q ^ {*} , q _ {i}\right)\right)}{\sum_ {q _ {j} \in Q ^ {*}} \exp \left(s \left(q ^ {*} , q _ {j}\right)\right)}. \tag {11} +$$ + +At this point, the PSA responding process is complete. We then feed $r^*$ into the decoding process of Language-guided U-Net, providing an approximated semantic guidance. + +# 4. Experimental Setup + +# 4.1. Datasets + +To evaluate our approach, we used the two well-benchmarked datasets: QaTa-COV19 [10], MosMedData+ [27] and Kvasir-SEG [18], which are commonly adopted for performance comparison in language-guided segmentation [15, 21, 42, 46, 48]. + +QaTa-COV19 is a large-scale dataset consisting of 9,258 chest X-ray images with manually annotated COVID-19 [41] lesion masks. To facilitate language-guided segmentation, LViT [21] extends this dataset with textual descriptions detailing bilateral lung infections, the number of affected regions, and their spatial localization within the lungs. We adopt the official dataset split: 5,716 images for training, 1,429 for validation, and 2,113 for testing. + +MosMedData+ comprises 2,729 CT slices depicting pulmonary infections. Similar to QaTa-COV19, LViT augments this dataset with text-based annotations to support language-guided segmentation tasks. We adopt the official dataset split: 2,183 slices for training, 273 for validation, and 273 for testing. + +Kvasir-SEG is a publicly available dataset comprising 1,000 colonoscopy images of gastrointestinal polyps with corresponding pixel-level segmentation masks. We follow a standard 8:1:1 split for training, validation, and testing. + +# 4.2. Evaluation Metrics + +To quantitatively evaluate segmentation performance, we used the metrics which are used in the previous similar studies [15, 21, 42, 46, 48]: Dice coefficient [11] and the mean Intersection over Union (mIoU [17]), two widely used metrics for measuring spatial overlap between predicted and ground truth segmentation masks. The formulations are defined in Equations 12 and 13. + +$$ +\operatorname {D i c e} = \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {2 \sum_ {c = 1} ^ {C} \left| P _ {i} ^ {(c)} \cap G _ {i} ^ {(c)} \right|}{\sum_ {c = 1} ^ {C} \left(\left| P _ {i} ^ {(c)} \right| + \left| G _ {i} ^ {(c)} \right|\right)} \tag {12} +$$ + +$$ +\mathrm {m I o U} = \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {1}{C} \sum_ {c = 1} ^ {C} \frac {\left| P _ {i} ^ {(c)} \cap G _ {i} ^ {(c)} \right|}{\left| P _ {i} ^ {(c)} \cup G _ {i} ^ {(c)} \right|} \tag {13} +$$ + +where $N$ represents the number of images in the dataset, $C$ denotes the number of semantic categories, and $P_{i}^{(c)}$ and $G_{i}^{(c)}$ correspond to the predicted and ground truth segmentation masks for class $c$ in image $i$ , respectively. + +# 4.3. Experimental Design + +To evaluate ProLearn, we compare it against both multimodal and unimodal segmentation methods, analyzing its performance under various realistic clinical scenarios. + +Limited availability of text: To illustrate the limitations of language-guided segmentation methods that require strict image-report pairing and highlight the importance of training with both image-text-paired and image-only data, we simulate a real-world clinical scenario where detailed radiology reports are often unavailable for a significant portion of the dataset. Due to the lack of large-scale benchmarks containing both paired and unpaired examples, we construct a counterfactual setting by progressively reducing the proportion of image-text pairs in the training data to $50\%$ , + +
DatasetModelDicemIoU
50%25%10%5%1%50%25%10%5%1%
QaTa-COV19 [10]LViT [21]0.84160.82020.80040.76380.70060.73200.70120.67470.62480.5490
GuideSeg [48]0.86330.85240.84020.82400.73330.75950.74280.72440.70070.5789
SGSeg [42]0.86410.85740.84230.80570.73070.76070.75040.72760.67460.5757
ProLearn0.86670.85980.85830.85730.85660.77210.76900.75730.75580.7553
MosMedData+ [27]LViT [21]0.71890.66080.55010.50690.16770.56960.50420.41080.35380.1015
GuideSeg [48]0.75080.73930.68980.63750.42350.60890.58640.52650.46780.2686
SGSeg [42]0.74550.74390.69500.64650.34520.59430.59220.53250.47760.2086
ProLearn0.75390.75120.74240.73790.72180.61260.61090.60870.60690.6032
Kvasir-SEG [18]LViT [21]0.76690.64240.57190.54820.42720.62280.47690.40250.37820.2729
GuideSeg [48]0.88480.83900.77540.74970.56150.79390.72280.63600.60430.4008
SGSeg [42]0.87690.83040.80250.75260.54060.78080.70990.67020.60340.3705
ProLearn0.89830.89460.88980.88230.87180.81620.81010.80200.79050.7729
+ +Table 1. Performance comparison of language-guided segmentation models in simulated scenarios with limited text supervision on the QaTa-COV19, MosMedData+ and Kvasir-SEG dataset. The best results are highlighted in bold. + +25%, 10%, 5%, and 1%, while discarding the remaining unpaired images to simulate their inaccessibility. We compare ProLearn against state-of-the-art language-guided segmentation models: LViT [21], GuideSeg [48], and SGSeg [42]. + +Segmentation in real-world (image-only) setting: In the majority of clinical scenarios, such as real-time procedural guidance and decision support, segmentation is used without text. To align with real-world settings, we evaluate ProLearn in a strictly "image-only" setting, where no text input is provided during either training or inference on the target datasets. Under these conditions, we compare ProLearn to established unimodal segmentation methods and vision-language pretraining models (CLIP [31] and GLoRIA [16]) adapted for image-only use. + +Prototype vs. LLM: To demonstrate our PSA's advantages against LLMs in real-world deployment, we focus on inference speed, model size, and time complexity. We specifically compare ProLearn with SGSeg [42], a recent method that leverages LLMs (e.g., GPT-2 [30], Llama3 [29]) to generate synthetic textual reports at inference in order to compensate for missing textual input. + +Qualitative Analysis: We provide visual comparisons against state-of-the-art language-guided segmentation models under progressively lower text availability settings. Specifically, We visualize both final segmentation outputs and corresponding saliency maps. + +Hyperparameter Sensitivity Analysis: We analyze how model performance varies as we adjust two key hyperparameters: the number of prototypes $M$ , which governs the compactness of the prototype space, and the number of candidate responses $k$ . + +![](images/5e4168c5036418f1d68efd61ac74629bf9beb361b781d233e054f5bc95167a3a.jpg) + +![](images/142f38d89ba4e96bcfd0433b749a37434384396318f686e39a3865dd6f6b66a9.jpg) + +![](images/4463e7ad9077d5b242d49f4e2c41be8d7e5e39eac577ad4855e16a4be65537c8.jpg) + +![](images/9ccd2da5a06773d45c4c2f4a4d747f88395603bad1c8f67bcb6f7e7d005f971a.jpg) + +![](images/11fb212e2d1fae579770553c467dc5e46561234825e22220c06a51a54e568552.jpg) +Figure 5. Performance degradation as the ratio of text availability decreases. The plots show Dice and nIoU metrics for a) the QaTa-COV19, b) the MosMedData+ and c) the Kvasir-SEG dataset. + +![](images/ee722b6c26e182e31005dd66cdd682043d76095ef09649b7eb306d4e19209f32.jpg) + +# 5. Results and Discussion + +# 5.1. Comparison with State-Of-The-Art Methods + +Table 1 presents our comparative results under clinical scenarios where paired textual reports are limited (50%, 25%, 10%, 5%, and 1%). ProLearn achieved higher Dice and mIoU scores than existing language-guided methods in all settings, and this performance gap increased as the frac + +
ModelQaTa-COV19MosMedData+Kvasir-SEG
DicemIoUDicemIoUDicemIoU
U-Net [32]0.8190.6920.6380.5050.1950.182
U-Net++ [49]0.8230.7060.7140.5820.2800.180
Attention U-Net [28]0.8220.7010.6640.5280.3640.226
Trans U-Net [8]0.8060.6870.7020.5750.0480.100
Swin U-Net [5]0.8360.7240.6690.5310.3980.246
CLIP [31]0.7980.7070.7200.596--
GLoRIA [16]0.7990.7070.7220.602--
MedSAM [24]0.7300.6190.5090.371--
BiomedParse [47]0.7810.6820.6710.5530.8280.721
ProLearn (1%)0.8570.7550.7220.6030.8720.773
ProLearn (5%)0.8570.7560.7380.6070.8820.790
ProLearn (10%)0.8580.7570.7420.6090.8890.802
+ +tion of available text decreased. For example, in the $1\%$ text in MosMedData+, ProLearn retained a Dice score of 0.7218, compared to SGSeg 0.3452 and LViT 0.1677. This trend underscores the effectiveness of ProLearn in learning from image-text and image-only data, allowing it to outperform approaches that require textual input for both training and inference. Although SGSeg also requires no text at inference, its performance remained below ProLearn due to its limited capacity to exploit learned semantics when text is scarce. Further discussions on the comparison between SGSeg and ProLearn can be found in Section 5.3. + +The performance degradation analysis, visualized in Figure 5, further illustrates the importance of reducing textual reliance in training. ProLearn exhibited the most minor degradation, preserving robust performance even under severely limited text supervision. + +# 5.2. Comparison in real-world (image-only) setting + +We further evaluated the effectiveness of ProLearn's use of a target dataset's auxiliary text under near "image-only" conditions, as most clinical workflows require segmentation models operating without textual input. As shown in Table 2, ProLearn outperformed both unimodal approaches (U-Net, Attention U-Net, Swin U-Net, etc.) and well-recognized VLP methods (CLIP, GLoRIA), even when only $10\%$ , $5\%$ , or $1\%$ of the paired reports were available. This demonstrates the importance of textual semantics within the target dataset. Unlike unimodal and VLP-based methods, which fail to leverage clinical reports during fine-tuning on target datasets, ProLearn incorporates available image-report pairs in the learning process. + +Table 2. Performance comparison of methods in image-only settings where paired reports are excluded entirely from the target datasets' training and inference. The best results are highlighted in bold. + +
ProLearnSGSeg
MethodQuery & respond from prototype matrixGenerate text tokens one by one (autoregressive) & text embedding
No. of inferenceSingle inferenceMultiple inference
TimeO(1)O(n) depends on the no. of tokens
+ +![](images/0edecf992e88319dde3b78c2ac5ce6556da282a83368b4355f482e12ff2e72dd.jpg) +Figure 6. Comparison between the proposed prototype-driven approach, ProLearn, and the LLM-based approach, SGSeg. Top: The theoretical analysis of time complexity. Bottom: The model size and inference time comparison. + +![](images/ff891e1d07e66ea52635536f76d2e4d8292a6b87723e20c87d942c70d94de6fc.jpg) + +# 5.3. Prototype vs. LLM + +Prototype learning enables constant-time inference of $\mathcal{O}(1)$ because it uses a pre-defined, finite prototype space that is independent of both image and text size. At the same time, LLM applies autoregressive image-to-text generation, which has a linear time complexity $\mathcal{O}(n)$ , where $n$ represents the number of tokens in the generated text (see Figure 6, top). As shown in Figure 6, bottom, prototype learning achieves an inference time of $4\mathrm{ms}$ , making it $100\times$ faster than GPT-2 (136ms) and $300\times$ faster than Llama3 (1.2s). Moreover, the prototype model has only 1M parameters, which is $1000\times$ smaller than large language models such as GPT-2 (1.5B parameters) and Llama3 (7B parameters). These properties make prototype learning highly efficient for real-time applications and suitable for deployment on resource-constrained edge devices. + +# 5.4. Visualization + +Figure 7 presents a qualitative comparison of ProLearn with state-of-the-art approaches on the QaTa-COV19, MosMedData+ and Kvasir-SEG datasets. As textual availability decreased, existing language-guided segmentation models, which rely heavily on image-text pairs, experienced a pronounced drop in performance. In contrast, ProLearn effectively learned from both image-text and image-only data, demonstrating stable results even under limited textual guidance. The same findings can be observed in saliency maps that pinpoint the regions on which each model was focused. Under decreased text guidance, conventional models produced unstable or diffuse attention patterns, compromising their segmentation accuracy. However, ProLearn, aided by semantic approximation, retained more coherent attention focused on areas of the lesion. + +![](images/c8ff8901badf26303f47fc3eff26e9167cb96d6d40de5a69b721d20f8c52f16d.jpg) +Figure 7. Visualization comparison of language-guided segmentation methods under different text availabilities. The upper row shows segmentation outputs on CXR_S09687-E18404-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_4_85506_1(case19_16 from the MosMedData+ dataset (middle) and ju5xkwzxf0z0818gk4xabdm from the Kvasir-SEG dataset (right). Blue regions indicate groundtruth pixels missed by the model, while red regions indicate pixels mistakenly predicted as lesion. The lower row presents saliency map interpretability on CXR_S09346-E23164-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_40_86625_0(case18_53 from the MosMedData+ dataset (middle) and cju2tzypl4wss0799ow05oxb9 from the Kvasir-SEG dataset (right). + +![](images/458e4c6ec4b7712acd49fb8eebfd8346d83baf62b5833ab23e2aae487e0ab5b3.jpg) +Figure 8. Hyperparameter sensitivity analysis with varying candidate and prototype configurations. a) Effect of varying $k$ , number of responding prototype vectors on Dice score; b) Effect of varying $M$ , number of prototypes per surrogate label on Dice score. + +![](images/6bb5ae8bebaa1c389b5f689e2b7b0ea4bb9df91d68972ce8e794c7df7517e492.jpg) + +# 5.5. Hyperparameter Sensitivity + +To investigate the effect of the number of candidates per response $k$ and the number of prototypes per surrogate label $M$ , we presented the performance comparison by varying their values for segmentation, as shown in Figure 8. We adjusted one hyperparameter at a time while keeping the other fixed. The experimental results indicated that our framework remained stable across a broad range of hyperparameter values. + +Specifically, $k$ governs the trade-off between information diversity and noise. A small $k$ limits the model's ability to capture subtle segmentation patterns due to insufficient variability. As $k$ increases, richer patterns are incorporated, improving performance. However, when $k$ is too large, irrelevant patterns introduce excessive noise, leading to degraded segmentation accuracy. + +For $M$ , increasing the number of prototypes enhances performance by capturing diverse segmentation characteristics. However, when $M$ exceeds a certain threshold, prototype redundancy arises, leading to overlapping or irrelevant representations that blur the segmentation boundaries. Thus, balancing $M$ is crucial; too few prototypes hinder representational capacity, while an excessive count dilutes the model's focus on meaningful patterns. + +# 6. Conclusion + +We investigated the issue of textual reliance in medical language-guided segmentation, which limits both the applicability of segmentation in clinical workflows and the exploration of image-only training data. To address this issue, we have presented ProLearn, a prototype-driven framework that fundamentally alleviates textual reliance in medical language-guided segmentation. Our experiments demonstrated that ProLearn effectively learns from both image-only and image-text data, making efficient use of the target dataset's textual descriptions. ProLearn showed stable performance under different text availability and outperformed state-of-the-art image-only approaches under minimal text supervision. + +Limitations and Future Works: In this work, we focused on demonstrating the core idea of ProLearn in single-label 2D segmentation. Future directions involve exploring multi-label segmentation, broader imaging modalities, and other language-guided vision tasks. + +# 7. Acknowledgement + +This work was supported in part by the Australian Research Council (ARC) under Grant IM240100224. + +# References + +[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10076-10095, 2024. 1 +[2] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21438-21451, 2023. 2 +[3] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 4 +[4] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE transactions on medical imaging, 33(2):577-590, 2013. 2 +[5] H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 2, 7 +[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1 +[7] Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):1655-1674, 2019. 2 +[8] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 7 +[9] Pujin Cheng, Li Lin, Junyan Lyu, Yijin Huang, Wenhan Luo, and Xiaoying Tang. Prior: Prototype representation joint learning from medical images and reports. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21361-21371, 2023. 4 +[10] Aysen Degerli, Serkan Kiranyaz, Muhammad E. H. Chowdhury, and Moncef Gabbouj. Osegnet: Operational segmentation network for Covid-19 detection using chest x-ray images. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2306–2310, 2022. 3, 5, 6 +[11] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 5 +[12] Vincenzo Ferrari, Marina Carbone, Carla Cappelli, Luigi Boni, Franca Melfi, Mauro Ferrari, Franco Mosca, and Andrea Pietrabissa. Value of multidetector computed tomogra + +phy image segmentation for preoperative planning in general surgery. Surgical endoscopy, 26:616-626, 2012. 2 +[13] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7606-7623, 2022. 1 +[14] Matthew Stephen Holden, Tamas Ungi, Derek Sargent, Robert C McGraw, Elvis CS Chen, Sugantha Ganapathy, Terry M Peters, and Gabor Fichtinger. Feasibility of real-time workflow segmentation for tracked needle interventions. IEEE Transactions on Biomedical Engineering, 61(6): 1720-1728, 2014. 2 +[15] Jihong Hu, Yinhao Li, Hao Sun, Yu Song, Chujie Zhang, Lanfen Lin, and Yen-Wei Chen. Lga: A language guide adapter for advancing the sam model's capabilities in medical image segmentation. page 610–620, Berlin, Heidelberg, 2024. Springer-Verlag. 2, 3, 5 +[16] Shih-Cheng Huang, Liyue Shen, Matthew P. Lungren, and Serena Yeung. Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3922-3931, 2021. 6, 7 +[17] Paul Jaccard. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547-579, 1901. 5 +[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In International conference on multimedia modeling, pages 451-462. Springer, 2019. 5, 6 +[19] Tina Kapur, Jan Egger, Jagadeesan Jayender, Matthew Toews, and William M Wells. Registration and segmentation for image-guided therapy. Intraoperative Imaging and Image-Guided Therapy, pages 79-91, 2014. 2 +[20] Stijn Keereweer, Jeroen DF Kerrebijn, Pieter BAA Van Driel, Bangwen Xie, Eric L Kaijzel, Thomas JA Snoeks, Ivo Que, Merlijn Hutteman, Joost R Van Der Vorst, J Sven D Mieog, et al. Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199-207, 2011. 2 +[21] Zihan Li, Yunxiang Li, Qingde Li, Puyang Wang, Dazhou Guo, Le Lu, Dakai Jin, You Zhang, and Qingqi Hong. Lvit: language meets vision transformer in medical image segmentation. IEEE transactions on medical imaging, 2023. 2, 5, 6 +[22] S. Lloyd. Least squares quantization in pcm. IEEE Transactions on Information Theory, 28(2):129-137, 1982. 4 +[23] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 1, 2 +[24] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3, 7 +[25] Dhiraj Maji, Prarthana Sigedar, and Munendra Singh. Attention res-unet with guided decoder for semantic segmentation + +of brain tumors. Biomedical Signal Processing and Control, 71:103077, 2022. 2 +[26] Epimack Michael, He Ma, Hong Li, Frank Kulwa, and Jing Li. Breast cancer segmentation methods: current status and future potentials. *BioMed research international*, 2021(1): 9962109, 2021. 2 +[27] Sergey Morozov, Anna Andreychenko, Nikolay Pavlov, Anton Vladzymyrskyy, Natalya Ledikhova, Victor Gombolevskiy, Ivan Blokhin, Pavel Gelezhe, Anna Gonchar, Valeria Chernina, et al. Mosmeddata: Chest ct scans with Covid-19 related findings. 2020. 3, 5, 6 +[28] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, and et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 7 +[29] Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards building multilingual language model for medicine. Nature Communications, 15(1):8384, 2024. 6 +[30] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7 +[32] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1, 2, 7 +[33] N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031-82057, 2021. 1 +[34] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4 +[35] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 3 +[36] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4 +[37] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 1, 2 +[38] Sudanthi Wijewickrema, Yun Zhou, James Bailey, Gregor Kennedy, and Stephen O'Leary. Provision of automated step-by-step procedural guidance in virtual reality surgery + +simulation. In Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology, pages 69-72, 2016. 2 +[39] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503-17512, 2023. 1, 2 +[40] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 4 +[41] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 5 +[42] Shuchang Ye, Mingyuan Meng, Mingjian Li, Dagan Feng, and Jinman Kim. Enabling text-free inference in language-guided segmentation of chest x-rays via self-guidance. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2024. 2, 3, 5, 6 +[43] Xiao-Xia Yin, Le Sun, Yuhan Fu, Ruiliang Lu, and Yanchun Zhang. [retracted] u-net-based medical image segmentation. Journal of healthcare engineering, 2022(1):4189781, 2022. 1 +[44] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[45] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Angela Crabtree, Brian Piening, Carlo Bifulco, Matthew P. Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon. A multimodal biomedical foundation model trained from fifteen million image-text pairs. NEJM AI, 2(1):A1oa2400640, 2025. 4 +[46] Xu Zhang, Bo Ni, Yang Yang, and Lefei Zhang. Madapter: A better interaction between image and language for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2024. 2, 3, 5 +[47] Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Hin Lee, Ho Sid Kiblawi, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, Brian Piening, Carlo Bifulco, Mu Wei, Hoifung Poon, and Sheng Wang. A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities. Nature Methods, 22(1):166-176, 2025. 7 +[48] Yi Zhong, Mengqiu Xu, Kongming Liang, Kaixin Chen, and Ming Wu. Ariadne's thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In International Conference on Medical Image Computing and + +Computer-Assisted Intervention, pages 724-733. Springer, 2023. 2, 3, 5, 6 +[49] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. 1, 2, 7 \ No newline at end of file diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/images.zip b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..433aaa1ec9b30edb6e9116621d6c478ca29e30ac --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2175479e5aa0140a7dc47d5909d51df5dafd885ff68b407a577ff15a3225576e +size 658733 diff --git a/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/layout.json b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..285ab6d40a0f623c9dccfc449177e4f033bca921 --- /dev/null +++ b/2025/Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation/layout.json @@ -0,0 +1,9069 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 140 + ], + "type": "text", + "content": "Alleviating Textual Reliance in Medical Language-guided Segmentation via Prototype-driven Semantic Approximation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 476, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 476, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 476, + 204 + ], + "type": "text", + "content": "Shuchang Ye1 Usman Naseem2 Mingyuan Meng1 Jinman Kim1 \n1The University of Sydney, Sydney, Australia \n2Macquarie University, Sydney, Australia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 206, + 302, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 206, + 302, + 217 + ], + "spans": [ + { + "bbox": [ + 154, + 206, + 302, + 217 + ], + "type": "text", + "content": "shuchang.ye@sydney.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 206, + 447, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 206, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 322, + 206, + 447, + 217 + ], + "type": "text", + "content": "usman.naseem@mq.edu.au" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 156, + 220, + 299, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 220, + 299, + 231 + ], + "spans": [ + { + "bbox": [ + 156, + 220, + 299, + 231 + ], + "type": "text", + "content": "mmen2292@uni.sydney.edu.au" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 320, + 220, + 451, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 451, + 231 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 451, + 231 + ], + "type": "text", + "content": "jinman.kim@sydney.edu.au" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 284, + 296, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 296, + 608 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 296, + 608 + ], + "type": "text", + "content": "Medical language-guided segmentation, integrating textual clinical reports as auxiliary guidance to enhance image segmentation, has demonstrated significant improvements over unimodal approaches. However, its inherent reliance on paired image-text input, which we refer to as \"textual reliance\", presents two fundamental limitations: 1) many medical segmentation datasets lack paired reports, leaving a substantial portion of image-only data underutilized for training; and 2) inference is limited to retrospective analysis of cases with paired reports, limiting its applicability in most clinical scenarios where segmentation typically precedes reporting. To address these limitations, we propose ProLearn, the first Prototype-driven Learning framework for language-guided segmentation that fundamentally alleviates textual reliance. At its core, we introduce a novel Prototype-driven Semantic Approximation (PSA) module to enable approximation of semantic guidance from textual input. PSA initializes a discrete and compact prototype space by distilling segmentation-relevant semantics from textual reports. Once initialized, it supports a query-and-respond mechanism which approximates semantic guidance for images without textual input, thereby alleviating textual reliance. Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms state-of-the-art language-guided methods when limited text is available. The code is available at https://github.com/ShuchangYe-bib/ProLearn." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Segmentation is an essential tool in medical image analysis, supporting clinical workflows by enabling precise delineation of anatomical structures, identification of pathological regions, and facilitating targeted interventions [24]. Its applications span critical tasks such as disease diag" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 552, + 389 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 552, + 389 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 552, + 389 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 552, + 389 + ], + "type": "image", + "image_path": "21b6ad183903f790dc44cbaaf35c6381287e5acddb4d911f27d7f48b0511f5d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 484 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 484 + ], + "type": "text", + "content": "Figure 1. Comparison of vision-language paradigms for medical image segmentation. a) VLP: Pretrained on large-scale image-text pairs and then fine-tuned on image-only data from the target dataset. b) Language-guided Segmentation: Requires strict image-text pairs during both training and inference. c) ProLearn: initializes with a limited amount of paired image-text data to construct query-response spaces. After initialization, it enables learning with limited textual input and performs inference without text." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 508, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 508, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 508, + 555, + 640 + ], + "type": "text", + "content": "nosis, treatment planning, and surgery support [1]. Deep learning has revolutionized segmentation, making it more accurate, reliable, and widely applicable in clinical practice. Unimodal (image-only) segmentation methods [43], such as U-Net [32] and its extensions [33], including U-Net++ [49], Attention U-Net [28], and Trans U-Net [8], have been widely adopted in medical imaging. In recent years, multimodal segmentation methods that leverage textual clinical reports as complementary information have gained wide attention for their ability to transcend the performance limits of unimodal segmentation [6, 13, 44]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "The exploration of multimodal segmentation [6, 13, 44] began with vision-language pretraining (VLP) [31], as shown in Figure 1a, where models pre-trained on paired textual and visual data demonstrated improved visual understanding and performance when finetuned on target image-only segmentation datasets [23, 37, 39]. How" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22316" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 190, + 155 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 190, + 155 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 190, + 155 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 190, + 155 + ], + "type": "image", + "image_path": "4427badb1869cef6a057c0149df79cb47e5a02d5d4ecd7fef0cf77fab078b12c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": "Figure 2. Flow diagram comparison between SGSeg [42] and our proposed ProLearn. a) SGSeg: image " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " LLM " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " report " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " BERT " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " embedding. b) ProLearn: image " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " PSA " + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 163, + 296, + 198 + ], + "type": "text", + "content": " embedding." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 294, + 156 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 294, + 156 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 294, + 156 + ], + "type": "image", + "image_path": "a7e559e8b4d082ce9cfdfc13e14b5458e020992dcdb9f18f8d66c1863b5e6479.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 220, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 295, + 376 + ], + "type": "text", + "content": "ever, these pretraining-based approaches fail to fully exploit the disease-specific information embedded in the target dataset's reports, as the general knowledge learned during pretraining often lacks the domain-specific details essential for precise disease lesion segmentation. To address this limitation, language-guided segmentation has been proposed as a promising approach [15, 21, 46, 48], which takes textual clinical report inputs as auxiliary semantic guidance for image segmentation (see Figure 1b). It has achieved remarkable success by leveraging textual abnormality description as explicit semantic guidance to segmentation, thus outperforming pretraining-based approaches and setting new benchmarks in medical image segmentation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 544 + ], + "type": "text", + "content": "However, language-guided segmentation methods have inherent reliance on paired image-text data due to the requirement for textual reports as auxiliary inputs. This reliance is referred to as textual reliance in this study and has incurred notable limitations during the training and inference stages. In the training stage, its reliance on paired image-text data leads to the underutilization of image-only data, which constitutes the majority of available datasets [2]. In the inference stage, reliance on textual descriptions during inference confines its use to retrospective analysis, misaligning with most clinical workflows where segmentation is required preemptively for tasks such as preoperative planning [12], diagnostic decision making [4, 25, 26], and real-time procedural guidance [14, 19, 38]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 677 + ], + "type": "text", + "content": "Our previous study, Self-Guided Segmentation (SGSeg) [42], made a preliminary attempt to eliminate textural reliance during inference in language-guided segmentation by leveraging large language models (LLMs) to generate synthetic reports to compensate for the missing textual input, as illustrated in Figure 2a. However, the integration of LLMs substantially increases the model size and inference time, making the approach unsuitable for deployment on edge devices [7] or for real-time applications such as image-guided surgery [20]. Moreover, textual reliance during training remains an open issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We argue that the critical guidance in language-guided segmentation is not the entire clinical report, which is often verbose and cluttered with irrelevant information, but" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "rather the specific segmentation-relevant semantic features embedded within it. Our investigation further indicates that the semantic space of medical reports is inherently constrained, consisting of a finite set of distinct representations relevant to segmentation tasks, as clinical reports typically adhere to standardized medical terminologies, which result in a relatively closed vocabulary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 157, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 361 + ], + "type": "text", + "content": "Building on these insights, we propose ProLearn (see Figure 1c), a lightweight and efficient Prototype Learning framework that fundamentally alleviates textual reliance during both training and inference. ProLearn introduces a Prototype-Driven Semantic Approximation (PSA) module, which enables the model to approximate semantic guidance without the need for textual input. Specifically, PSA constructs a discrete and compact prototype space by distilling segmentation-relevant concepts from clinical reports. It then provides a query-and-respond mechanism to support interaction between segmentation models and the prototype space, where unseen semantics are approximated by weighted aggregation of the existing prototypes based on similarity. Therefore, PSA enables segmentation models to query by image feature and receive responded semantics feature as guidance for feature maps refinement, as illustrated in Figure 2b." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 522, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 522, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 522, + 372 + ], + "type": "text", + "content": "The main contributions of this work are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 553, + 578 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 374, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 553, + 421 + ], + "type": "text", + "content": "- To the best of our knowledge, our proposed ProLearn is the first work to alleviate textual reliance in medical language-guided segmentation in both training and inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 481 + ], + "type": "text", + "content": "- We introduce a novel PSA module that supports learning with both paired image-text data and image-only data while enabling inference without textual input by querying a learned prototype space to provide semantic guidance for segmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "text", + "content": "- Extensive experiments on QaTa-COV19, MosMedData+ and Kvasir-SEG demonstrate that ProLearn outperforms language-guided segmentation methods in limited text availability settings (see Section 5.1) and surpasses state-of-the-art unimodal segmentation models (see Section 5.2). Compared to SGSeg, ProLearn achieves a " + }, + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "inline_equation", + "content": "1000 \\times" + }, + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "text", + "content": " reduction in parameters and " + }, + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 313, + 481, + 553, + 578 + ], + "type": "text", + "content": " faster inference speed (see Section 5.3)." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 400, + 603 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 485, + 624 + ], + "type": "text", + "content": "2.1. Language-guided Segmentation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Language-guided segmentation aims to address the gap that the target dataset's reports are not fully exploited for learning in conventional unimodal [5, 8, 28, 32, 49] or VLP methods [23, 31, 37, 39]. LViT [21] was the first work in language-guided segmentation, which takes both images and textual reports as input to train a multimodal-input segmentation model. They annotated existing pub-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22317" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 73, + 519, + 319 + ], + "blocks": [ + { + "bbox": [ + 98, + 73, + 519, + 319 + ], + "lines": [ + { + "bbox": [ + 98, + 73, + 519, + 319 + ], + "spans": [ + { + "bbox": [ + 98, + 73, + 519, + 319 + ], + "type": "image", + "image_path": "2fde88a958ef02a100a8a3c17c1d4a7495d147678cba75054ddc56d658a77a1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 341, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 54, + 341, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 555, + 407 + ], + "type": "text", + "content": "Figure 3. The overview of the ProLearn framework for alleviating textual reliance in medical language-guided segmentation. a) Language-guided U-Net: The U-Net encoder extracts image features, which are used to query the PSA module. The decoder is then guided by PSA's response to decode the segmentation mask. b) PSA Initialization: Clinical reports are processed and abstracted into a discrete prototype space, representing segmentation-relevant semantic and spatial information. c) PSA Querying and Responding: PSA receives image feature queries, selects relevant prototype candidates, and responds with an aggregated representation to approximate the pathological region's segmentation-relevant semantic representation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 428, + 294, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 428, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 428, + 294, + 548 + ], + "type": "text", + "content": "licly available segmentation datasets, QaTa-COV19 [10] and MosMedData+ [27], with corresponding clinical reports. To fuse the text features with feature maps in the U-Net, LViT adopted an early fusion approach, which introduced a Pixel-Level Attention Module (PLAM) to involve textual features as semantic guidance. LViT showed consistent performance over image-only and VLP segmentation methods, which demonstrated the importance and potential of textual semantic guidance from clinical reports in target datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Subsequent methods proposed more flexible and robust feature fusion modules for language-guided segmentation. GuideSeg [48] moved away from early fusion while adopting a late-fusion strategy that fused textual and visual features at the decoding stage, where the text features were better preserved and more effectively influenced the segmentation process. MA adaptor [46] addressed the unidirectional flow of textual semantics seen in previous frameworks. It introduced a bidirectional adaptor connecting multiple layers of unimodal encoders, facilitating mutual information exchange between text and image representations at various scales. LGA [15] adopted a parameter-efficient fin-tuning strategy that preserved the original parameters of" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 428, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 475 + ], + "type": "text", + "content": "large segmentation foundation models [24]. These works further prove that integrating target datasets' textual reports as guidance can significantly improve the performance of segmentation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 478, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 556, + 658 + ], + "type": "text", + "content": "These methods suffered from textual reliance. The reliance during training restricted the utilization of large portions of medical datasets that lacked paired textual reports. The reliance during inference limits their practical utility in most clinical scenarios using segmentation without reports. SGSeg [42] attempted to release the textual reliance at inference by training a LLM [35] to generate needed clinical reports from images. Nevertheless, the inclusion of LLMs increased model size and inference time, making it unsuitable for edge devices and real-time applications. The textual reliance during training still remained unsolved. Our proposed ProLearn fundamentally alleviates textual reliance in training and inference, and its prototype design PSA significantly reduces the number of parameters and inference time compared to LLM-based SGSeg." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 670, + 430, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 430, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 430, + 684 + ], + "type": "text", + "content": "2.2. Prototype Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Prototype learning draws on the principle that images and text can be effectively captured in discrete prototype repre" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22318" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 294, + 174 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 294, + 174 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 294, + 174 + ], + "type": "image", + "image_path": "cab19660a2132f4f0970e8b50b497a84e4517b12d08c682f03b73f49d5391cb4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 181, + 272, + 192 + ], + "lines": [ + { + "bbox": [ + 77, + 181, + 272, + 192 + ], + "spans": [ + { + "bbox": [ + 77, + 181, + 272, + 192 + ], + "type": "text", + "content": "Figure 4. Attention-guided surrogate label extraction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 429 + ], + "type": "text", + "content": "sentations rather than relying on freshly built embeddings for every instance [36]. Each prototype functions as a canonical reference, reflecting the typical features of its corresponding class. Such a strategy is widely adopted in classification [40], where an unseen query is associated with the class whose prototype lies closest in a learned space. By prioritizing the reuse of well-established representations, prototype-based methods frequently attain top-tier performance under data-scarce settings [34]. Beyond unimodal applications, prototype learning has demonstrated its potential in multimodal tasks, effectively aligning features across modalities [9], such as images and text. Inspired by the principles of expressing unseen queries via finite existing prototypes, the proposed ProLearn extracts semantic information from textual reports as prototypes, enabling representing semantic guidance from existing image-text pairs for image-only input during training and inference, with improved parameter, computation, and data efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 138, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 138, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 138, + 451 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 554 + ], + "type": "text", + "content": "Figure 3 shows the overall architecture of ProLearn, a framework that leverages an efficient prototype learning module called Prototype-Driven Semantic Approximation (PSA) alongside a Language-guided U-Net backbone for segmentation. PSA is designed to (i) reduce the need for textual annotations during training by using available text only once to initialize a discrete prototype space and (ii) remove the need for textual input entirely at inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 560, + 162, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 162, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 162, + 572 + ], + "type": "text", + "content": "3.1. PSA Initialization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": "The PSA initialization is a one-time process that constructs a queryable prototype space before training. Given that a training set includes " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " paired image-text samples and other image-only samples, where the " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": "-th sample is denoted as " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\langle I_i, T_i \\rangle" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " as the image and " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " as its associated report. Each paired sample is processed by a trained domain-specific vision-language encoder (BioMedCLIP [45]), " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}^I" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}^T" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " to extract their semantic features " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "e_i^I" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "inline_equation", + "content": "e_i^T" + }, + { + "bbox": [ + 55, + 578, + 296, + 674 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 679, + 296, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 679, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 78, + 679, + 296, + 695 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {I} = f _ {\\mathrm {e n c}} ^ {I} (I _ {i}), \\quad e _ {i} ^ {T} = f _ {\\mathrm {e n c}} ^ {T} (T _ {i}), \\quad e _ {i} ^ {I}, e _ {i} ^ {T} \\in \\mathbb {R} ^ {D} \\quad (1)", + "image_path": "0c1f1b9c1f532a28b30454bed27019ffbdd3e3e01992ffbc8124be242b23b4e3.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 714 + ], + "type": "text", + "content": "Surrogate Label Extraction: To extract finite" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "segmentation-relevant semantics in the clinical reports, we cluster the image-text pairs into " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " semantic surrogate labels, as shown in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": "As many clinical reports are verbose, we first isolate the tokens most relevant to segmentation. To achieve this, we utilize the cross-attention module of a separately trained Language-guided U-Net, which has the same architecture as the language-guided U-Net in Figure 3, to assess token relevance in all " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " available image-text pairs " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\{\\langle I_i,T_i\\rangle \\mid i = 1,\\dots ,K\\}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": ". Specifically, for each text input " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " paired with an image " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": ", the cross-attention weight " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " computed for each token " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " indicates the relevance of each token " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " in guiding the segmentation of " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": ". Tokens whose attention scores exceed a threshold " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": " are retained, resulting in a shorter segmentation-relevant sentence for each " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": ", which we denote " + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "inline_equation", + "content": "T_{i}^{\\mathrm{selected}}" + }, + { + "bbox": [ + 313, + 108, + 555, + 263 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 269, + 553, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 269, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 365, + 269, + 553, + 283 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {\\text {s e l e c t e d}} = \\left\\{t _ {j} \\mid \\alpha_ {j} > \\tau , t _ {j} \\in T _ {i} \\right\\}. \\tag {2}", + "image_path": "f854356aa55171ca741a5b936f1c4b487ca1ec1605b592eafe34c9dac2a74ed5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "content": "We then feed " + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "inline_equation", + "content": "T_{i}^{\\mathrm{selected}}" + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "content": " into the text encoder " + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}^{T}" + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "content": ", obtaining a semantic feature " + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "inline_equation", + "content": "e_{i}^{\\mathrm{sem}}" + }, + { + "bbox": [ + 313, + 289, + 553, + 313 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 319, + 553, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 319, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 347, + 319, + 553, + 334 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {\\text {s e m}} = f _ {\\text {e n c}} ^ {T} \\left(T _ {i} ^ {\\text {s e l e c t e d}}\\right), \\quad \\forall i \\in \\{1, \\dots , K \\} \\tag {3}", + "image_path": "bf9913bedb464981e585bbba5b2c5ce03897c7533ca110284db73dd062fc65ea.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": "We then group similar textual semantics together using the hierarchical density-based clustering algorithm, HDBSCAN [3]. The clustering results in " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " surrogate labels " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\{l_1,l_2,\\dots ,l_N\\}" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": ". Each surrogate label " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " encapsulates a distinct segmentation-relevant textual semantics, with its corresponding cluster " + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 339, + 554, + 411 + ], + "type": "text", + "content": " representing the grouped features:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 416, + 553, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 416, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 347, + 416, + 553, + 431 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i} = \\left\\{\\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\mid \\operatorname {H D B S C A N} \\left(e _ {i} ^ {\\text {s e m}}\\right) = l _ {i} \\right\\}. \\tag {4}", + "image_path": "ac7e0229612d9ad93b3ebe03e76318f8f27d3aa1af03ec648a1147b2c513dc1c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "text", + "content": "Prototype Space Construction: While the above surrogate labels capture sparse textual semantics, medical images often convey richer and more fine-grained information. To make the prototype space compact and more representative for image queries, each text-based surrogate label cluster " + }, + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "text", + "content": " is further subdivided into sub-clusters " + }, + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{ij}" + }, + { + "bbox": [ + 313, + 436, + 554, + 520 + ], + "type": "text", + "content": " via K-means clustering algorithm [22]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 322, + 525, + 553, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 525, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 322, + 525, + 553, + 540 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i j} = \\left\\{\\left\\langle e _ {k} ^ {I}, e _ {k} ^ {T} \\right\\rangle \\mid \\text {K - M e a n} \\left(e _ {k} ^ {I}\\right) = j, \\left\\langle e _ {j} ^ {I}, e _ {j} ^ {T} \\right\\rangle \\in \\mathcal {C} _ {i} \\right\\}. \\tag {5}", + "image_path": "dc6533a812bb462746bba203ea3681a156a6e46179a43af1e6aeaa7dcd67efe0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": "For each sub-cluster " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{ij}" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ", instead of taking the centroid itself, we locate the sample " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "c_{ij}" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": " closest to the cluster's centroid and treat it as a fine representation of disease lesion. This approach reduces the influence of outliers that can significantly distort the centroid. Given " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "c_{ij} = \\langle e_k^I, e_k^T \\rangle" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ", we assign its containing image feature " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "e_k^I" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": " as visual query prototype " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": " and text feaître " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "e_k^T" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": " as textual respond prototype " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ". By repeating this across all sub-clusters, we construct a discrete and compact initial prototype space " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ", consisting of a query space " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "S^Q" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": " and a response space " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "S^R" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ", both of dimension " + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "inline_equation", + "content": "N \\times M \\times D" + }, + { + "bbox": [ + 313, + 545, + 554, + 677 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 681, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 681, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 361, + 681, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left(\\mathcal {S} ^ {Q}, \\mathcal {S} ^ {R}\\right) = \\bigcup_ {i = 1} ^ {N} \\bigcup_ {j = 1} ^ {M} \\langle q _ {i j}, r _ {i j} \\rangle . \\tag {6}", + "image_path": "a241cee94d3e63f1e73902a6d2c92a8d9856b66216b9152c54007f80d0597672.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22319" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "Each visual query prototype " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^Q" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is directly linked to its corresponding textual response prototype " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "r_{ij}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{S}^R" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 105, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 105, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 63, + 105, + 295, + 119 + ], + "type": "interline_equation", + "content": "q _ {i j} \\longrightarrow r _ {i j}, \\quad \\forall i \\in \\{1, \\dots , N \\}, \\quad j \\in \\{1, \\dots , M \\}. \\tag {7}", + "image_path": "be28ebbf1a17dd63097464bcba9e7e260b173aaaae428ff36b2ac9db423d56d9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 295, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 150 + ], + "type": "text", + "content": "This prototype space is dynamically learned during the training process of ProLearn." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 156, + 225, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 225, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 225, + 170 + ], + "type": "text", + "content": "3.2. PSA Querying and Responding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": "During both training and inference, the Language-guided U-Net " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{seg}}" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": " queries the query space " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "S^Q" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": " using an image feature " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "q^*" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": ", encoded by the image encoder " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{enc}}^I" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": " from an image input " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "I^*" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": ", enabling the use of image-only samples. Through the query-and-response mechanism of the PSA module, it responds with an approximate textual semantic feature " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "r^*" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": " which guides " + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{seg}}" + }, + { + "bbox": [ + 55, + 174, + 296, + 258 + ], + "type": "text", + "content": " without textual input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": "PSA Querying: The PSA querying process involves searching the most relevant visual queries from the query prototype space " + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "inline_equation", + "content": "S^Q" + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": ". Given the encoded image feature " + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "inline_equation", + "content": "q*" + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": ", the PSA module computes the cosine similarity scores " + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "inline_equation", + "content": "s_{ij}" + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "inline_equation", + "content": "q^*" + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": " and each query prototype " + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "inline_equation", + "content": "q_{ij}" + }, + { + "bbox": [ + 55, + 258, + 296, + 319 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 326, + 295, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 326, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 113, + 326, + 295, + 353 + ], + "type": "interline_equation", + "content": "s _ {i} j = s \\left(q ^ {*}, q _ {i j}\\right) = \\frac {q ^ {*} \\cdot q _ {i j}}{\\| q ^ {*} \\| \\| q _ {i j} \\|} \\tag {8}", + "image_path": "aa5d039eb92c71c5cd8c04f892ce8d5106604aeaeacf9953f5c2e09d57da77de.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": "After ranking these similarity scores, the PSA module selects the top-" + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": " query prototypes " + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "inline_equation", + "content": "Q^{*}" + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": " that best match " + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "inline_equation", + "content": "q^{*}" + }, + { + "bbox": [ + 55, + 358, + 295, + 383 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 102, + 391, + 295, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 391, + 295, + 405 + ], + "spans": [ + { + "bbox": [ + 102, + 391, + 295, + 405 + ], + "type": "interline_equation", + "content": "Q ^ {*} = \\left\\{q _ {i j} \\mid \\forall i, j \\in \\arg \\operatorname {t o p} _ {k} \\left(s _ {i j}\\right) \\right\\}. \\tag {9}", + "image_path": "c58caad5a6df58a7e422ca009e41ced423b12009ce90c8c9e1008aba96278a1f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "content": "PSA Responding: The PSA responding process finds the corresponding response prototypes " + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "content": " that are linked to the selected query prototypes " + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "inline_equation", + "content": "Q^{*}" + }, + { + "bbox": [ + 55, + 412, + 296, + 449 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 456, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 456, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 102, + 456, + 295, + 471 + ], + "type": "interline_equation", + "content": "R ^ {*} = \\left\\{r _ {i j} \\mid q _ {i j} \\in Q ^ {*}, q _ {i j} \\longrightarrow r _ {i j} \\right\\}. \\tag {10}", + "image_path": "165bcbc65b5e495251e053c0d57cc2a3b34f9a65cb7e9a47a1b93b40d0022fbd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 478, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 525 + ], + "type": "text", + "content": "These response prototypes " + }, + { + "bbox": [ + 55, + 478, + 296, + 525 + ], + "type": "inline_equation", + "content": "R^{*}" + }, + { + "bbox": [ + 55, + 478, + 296, + 525 + ], + "type": "text", + "content": " are referred to as candidates in Figure 3b. The candidates are then aggregated using a similarity-weighted sum, where the weights are computed through the softmax function over the similarity scores:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 64, + 533, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 533, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 64, + 533, + 295, + 563 + ], + "type": "interline_equation", + "content": "r ^ {*} = \\sum_ {r _ {i} \\in R ^ {*}} w _ {i} r _ {i}, \\quad w _ {i} = \\frac {\\exp \\left(s \\left(q ^ {*} , q _ {i}\\right)\\right)}{\\sum_ {q _ {j} \\in Q ^ {*}} \\exp \\left(s \\left(q ^ {*} , q _ {j}\\right)\\right)}. \\tag {11}", + "image_path": "dea81b5cbef7126f71ec88e73a6a89d8f746aac13b6859285c238c2a3fa6b1f5.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 570, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 606 + ], + "type": "text", + "content": "At this point, the PSA responding process is complete. We then feed " + }, + { + "bbox": [ + 55, + 570, + 296, + 606 + ], + "type": "inline_equation", + "content": "r^*" + }, + { + "bbox": [ + 55, + 570, + 296, + 606 + ], + "type": "text", + "content": " into the decoding process of Language-guided U-Net, providing an approximated semantic guidance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 616, + 173, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 173, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 173, + 631 + ], + "type": "text", + "content": "4. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 636, + 118, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 118, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 118, + 647 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To evaluate our approach, we used the two well-benchmarked datasets: QaTa-COV19 [10], MosMedData+ [27] and Kvasir-SEG [18], which are commonly adopted for performance comparison in language-guided segmentation [15, 21, 42, 46, 48]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "QaTa-COV19 is a large-scale dataset consisting of 9,258 chest X-ray images with manually annotated COVID-19 [41] lesion masks. To facilitate language-guided segmentation, LViT [21] extends this dataset with textual descriptions detailing bilateral lung infections, the number of affected regions, and their spatial localization within the lungs. We adopt the official dataset split: 5,716 images for training, 1,429 for validation, and 2,113 for testing." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 167, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 239 + ], + "type": "text", + "content": "MosMedData+ comprises 2,729 CT slices depicting pulmonary infections. Similar to QaTa-COV19, LViT augments this dataset with text-based annotations to support language-guided segmentation tasks. We adopt the official dataset split: 2,183 slices for training, 273 for validation, and 273 for testing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 239, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 288 + ], + "type": "text", + "content": "Kvasir-SEG is a publicly available dataset comprising 1,000 colonoscopy images of gastrointestinal polyps with corresponding pixel-level segmentation masks. We follow a standard 8:1:1 split for training, validation, and testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 294, + 427, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 427, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 427, + 306 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 396 + ], + "type": "text", + "content": "To quantitatively evaluate segmentation performance, we used the metrics which are used in the previous similar studies [15, 21, 42, 46, 48]: Dice coefficient [11] and the mean Intersection over Union (mIoU [17]), two widely used metrics for measuring spatial overlap between predicted and ground truth segmentation masks. The formulations are defined in Equations 12 and 13." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 353, + 403, + 553, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 403, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 353, + 403, + 553, + 437 + ], + "type": "interline_equation", + "content": "\\operatorname {D i c e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {2 \\sum_ {c = 1} ^ {C} \\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\sum_ {c = 1} ^ {C} \\left(\\left| P _ {i} ^ {(c)} \\right| + \\left| G _ {i} ^ {(c)} \\right|\\right)} \\tag {12}", + "image_path": "14122476d11bb592dec14e205a4d284890304420fa5ee7175826d2e4ac4068bf.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 358, + 445, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 445, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 358, + 445, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\mathrm {m I o U} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\frac {\\left| P _ {i} ^ {(c)} \\cap G _ {i} ^ {(c)} \\right|}{\\left| P _ {i} ^ {(c)} \\cup G _ {i} ^ {(c)} \\right|} \\tag {13}", + "image_path": "04a3a4f82277bef3df22dd8bf5fcdec25618dc869413381bdf28e6e5d9edc32b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": " represents the number of images in the dataset, " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": " denotes the number of semantic categories, and " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "P_{i}^{(c)}" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "G_{i}^{(c)}" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": " correspond to the predicted and ground truth segmentation masks for class " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": " in image " + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 482, + 553, + 534 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 435, + 553 + ], + "type": "text", + "content": "4.3. Experimental Design" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "content": "To evaluate ProLearn, we compare it against both multimodal and unimodal segmentation methods, analyzing its performance under various realistic clinical scenarios." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Limited availability of text: To illustrate the limitations of language-guided segmentation methods that require strict image-report pairing and highlight the importance of training with both image-text-paired and image-only data, we simulate a real-world clinical scenario where detailed radiology reports are often unavailable for a significant portion of the dataset. Due to the lack of large-scale benchmarks containing both paired and unpaired examples, we construct a counterfactual setting by progressively reducing the proportion of image-text pairs in the training data to " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22320" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 525, + 233 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 525, + 233 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 525, + 233 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 525, + 233 + ], + "type": "table", + "html": "
DatasetModelDicemIoU
50%25%10%5%1%50%25%10%5%1%
QaTa-COV19 [10]LViT [21]0.84160.82020.80040.76380.70060.73200.70120.67470.62480.5490
GuideSeg [48]0.86330.85240.84020.82400.73330.75950.74280.72440.70070.5789
SGSeg [42]0.86410.85740.84230.80570.73070.76070.75040.72760.67460.5757
ProLearn0.86670.85980.85830.85730.85660.77210.76900.75730.75580.7553
MosMedData+ [27]LViT [21]0.71890.66080.55010.50690.16770.56960.50420.41080.35380.1015
GuideSeg [48]0.75080.73930.68980.63750.42350.60890.58640.52650.46780.2686
SGSeg [42]0.74550.74390.69500.64650.34520.59430.59220.53250.47760.2086
ProLearn0.75390.75120.74240.73790.72180.61260.61090.60870.60690.6032
Kvasir-SEG [18]LViT [21]0.76690.64240.57190.54820.42720.62280.47690.40250.37820.2729
GuideSeg [48]0.88480.83900.77540.74970.56150.79390.72280.63600.60430.4008
SGSeg [42]0.87690.83040.80250.75260.54060.78080.70990.67020.60340.3705
ProLearn0.89830.89460.88980.88230.87180.81620.81010.80200.79050.7729
", + "image_path": "fac14836dd2854a198e3ff5e7fcb9178b6ff8bc7173b09c08cabe96a7e9e433d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "type": "text", + "content": "Table 1. Performance comparison of language-guided segmentation models in simulated scenarios with limited text supervision on the QaTa-COV19, MosMedData+ and Kvasir-SEG dataset. The best results are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 293, + 295, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 343 + ], + "type": "text", + "content": "25%, 10%, 5%, and 1%, while discarding the remaining unpaired images to simulate their inaccessibility. We compare ProLearn against state-of-the-art language-guided segmentation models: LViT [21], GuideSeg [48], and SGSeg [42]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 475 + ], + "type": "text", + "content": "Segmentation in real-world (image-only) setting: In the majority of clinical scenarios, such as real-time procedural guidance and decision support, segmentation is used without text. To align with real-world settings, we evaluate ProLearn in a strictly \"image-only\" setting, where no text input is provided during either training or inference on the target datasets. Under these conditions, we compare ProLearn to established unimodal segmentation methods and vision-language pretraining models (CLIP [31] and GLoRIA [16]) adapted for image-only use." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 571 + ], + "type": "text", + "content": "Prototype vs. LLM: To demonstrate our PSA's advantages against LLMs in real-world deployment, we focus on inference speed, model size, and time complexity. We specifically compare ProLearn with SGSeg [42], a recent method that leverages LLMs (e.g., GPT-2 [30], Llama3 [29]) to generate synthetic textual reports at inference in order to compensate for missing textual input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "content": "Qualitative Analysis: We provide visual comparisons against state-of-the-art language-guided segmentation models under progressively lower text availability settings. Specifically, We visualize both final segmentation outputs and corresponding saliency maps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Hyperparameter Sensitivity Analysis: We analyze how model performance varies as we adjust two key hyperparameters: the number of prototypes " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", which governs the compactness of the prototype space, and the number of candidate responses " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 293, + 439, + 376 + ], + "blocks": [ + { + "bbox": [ + 326, + 293, + 439, + 376 + ], + "lines": [ + { + "bbox": [ + 326, + 293, + 439, + 376 + ], + "spans": [ + { + "bbox": [ + 326, + 293, + 439, + 376 + ], + "type": "image", + "image_path": "5e4168c5036418f1d68efd61ac74629bf9beb361b781d233e054f5bc95167a3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 442, + 293, + 548, + 375 + ], + "blocks": [ + { + "bbox": [ + 442, + 293, + 548, + 375 + ], + "lines": [ + { + "bbox": [ + 442, + 293, + 548, + 375 + ], + "spans": [ + { + "bbox": [ + 442, + 293, + 548, + 375 + ], + "type": "image", + "image_path": "142f38d89ba4e96bcfd0433b749a37434384396318f686e39a3865dd6f6b66a9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 380, + 439, + 462 + ], + "blocks": [ + { + "bbox": [ + 326, + 380, + 439, + 462 + ], + "lines": [ + { + "bbox": [ + 326, + 380, + 439, + 462 + ], + "spans": [ + { + "bbox": [ + 326, + 380, + 439, + 462 + ], + "type": "image", + "image_path": "4463e7ad9077d5b242d49f4e2c41be8d7e5e39eac577ad4855e16a4be65537c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 442, + 380, + 548, + 462 + ], + "blocks": [ + { + "bbox": [ + 442, + 380, + 548, + 462 + ], + "lines": [ + { + "bbox": [ + 442, + 380, + 548, + 462 + ], + "spans": [ + { + "bbox": [ + 442, + 380, + 548, + 462 + ], + "type": "image", + "image_path": "9ccd2da5a06773d45c4c2f4a4d747f88395603bad1c8f67bcb6f7e7d005f971a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 469, + 439, + 548 + ], + "blocks": [ + { + "bbox": [ + 326, + 469, + 439, + 548 + ], + "lines": [ + { + "bbox": [ + 326, + 469, + 439, + 548 + ], + "spans": [ + { + "bbox": [ + 326, + 469, + 439, + 548 + ], + "type": "image", + "image_path": "11fb212e2d1fae579770553c467dc5e46561234825e22220c06a51a54e568552.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 558, + 555, + 592 + ], + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 592 + ], + "type": "text", + "content": "Figure 5. Performance degradation as the ratio of text availability decreases. The plots show Dice and nIoU metrics for a) the QaTa-COV19, b) the MosMedData+ and c) the Kvasir-SEG dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 444, + 470, + 547, + 548 + ], + "blocks": [ + { + "bbox": [ + 444, + 470, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 444, + 470, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 444, + 470, + 547, + 548 + ], + "type": "image", + "image_path": "ee722b6c26e182e31005dd66cdd682043d76095ef09649b7eb306d4e19209f32.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 614, + 446, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 446, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 446, + 627 + ], + "type": "text", + "content": "5. Results and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 546, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 546, + 647 + ], + "type": "text", + "content": "5.1. Comparison with State-Of-The-Art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": "Table 1 presents our comparative results under clinical scenarios where paired textual reports are limited (50%, 25%, 10%, 5%, and 1%). ProLearn achieved higher Dice and mIoU scores than existing language-guided methods in all settings, and this performance gap increased as the frac" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22321" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 301, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 301, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 301, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 301, + 224 + ], + "type": "table", + "html": "
ModelQaTa-COV19MosMedData+Kvasir-SEG
DicemIoUDicemIoUDicemIoU
U-Net [32]0.8190.6920.6380.5050.1950.182
U-Net++ [49]0.8230.7060.7140.5820.2800.180
Attention U-Net [28]0.8220.7010.6640.5280.3640.226
Trans U-Net [8]0.8060.6870.7020.5750.0480.100
Swin U-Net [5]0.8360.7240.6690.5310.3980.246
CLIP [31]0.7980.7070.7200.596--
GLoRIA [16]0.7990.7070.7220.602--
MedSAM [24]0.7300.6190.5090.371--
BiomedParse [47]0.7810.6820.6710.5530.8280.721
ProLearn (1%)0.8570.7550.7220.6030.8720.773
ProLearn (5%)0.8570.7560.7380.6070.8820.790
ProLearn (10%)0.8580.7570.7420.6090.8890.802
", + "image_path": "75e90768487d1333aaca4b05f246fa7ba8b4a469a7689bb2bba969b8fb262d93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 316, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 449 + ], + "type": "text", + "content": "tion of available text decreased. For example, in the " + }, + { + "bbox": [ + 54, + 316, + 295, + 449 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 54, + 316, + 295, + 449 + ], + "type": "text", + "content": " text in MosMedData+, ProLearn retained a Dice score of 0.7218, compared to SGSeg 0.3452 and LViT 0.1677. This trend underscores the effectiveness of ProLearn in learning from image-text and image-only data, allowing it to outperform approaches that require textual input for both training and inference. Although SGSeg also requires no text at inference, its performance remained below ProLearn due to its limited capacity to exploit learned semantics when text is scarce. Further discussions on the comparison between SGSeg and ProLearn can be found in Section 5.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 296, + 514 + ], + "type": "text", + "content": "The performance degradation analysis, visualized in Figure 5, further illustrates the importance of reducing textual reliance in training. ProLearn exhibited the most minor degradation, preserving robust performance even under severely limited text supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 536, + 295, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 549 + ], + "type": "text", + "content": "5.2. Comparison in real-world (image-only) setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "We further evaluated the effectiveness of ProLearn's use of a target dataset's auxiliary text under near \"image-only\" conditions, as most clinical workflows require segmentation models operating without textual input. As shown in Table 2, ProLearn outperformed both unimodal approaches (U-Net, Attention U-Net, Swin U-Net, etc.) and well-recognized VLP methods (CLIP, GLoRIA), even when only " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " of the paired reports were available. This demonstrates the importance of textual semantics within the target dataset. Unlike unimodal and VLP-based methods, which fail to leverage clinical reports during fine-tuning on target datasets, ProLearn incorporates available image-report pairs in the learning process." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 71, + 553, + 125 + ], + "blocks": [ + { + "bbox": [ + 55, + 241, + 296, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 285 + ], + "type": "text", + "content": "Table 2. Performance comparison of methods in image-only settings where paired reports are excluded entirely from the target datasets' training and inference. The best results are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 71, + 553, + 125 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 553, + 125 + ], + "type": "table", + "html": "
ProLearnSGSeg
MethodQuery & respond from prototype matrixGenerate text tokens one by one (autoregressive) & text embedding
No. of inferenceSingle inferenceMultiple inference
TimeO(1)O(n) depends on the no. of tokens
", + "image_path": "0e82611590829cf3788a2adb3164bcee9482a05e8dd74f215436894e129039fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 322, + 127, + 426, + 198 + ], + "blocks": [ + { + "bbox": [ + 322, + 127, + 426, + 198 + ], + "lines": [ + { + "bbox": [ + 322, + 127, + 426, + 198 + ], + "spans": [ + { + "bbox": [ + 322, + 127, + 426, + 198 + ], + "type": "image", + "image_path": "0edecf992e88319dde3b78c2ac5ce6556da282a83368b4355f482e12ff2e72dd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 210, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 255 + ], + "type": "text", + "content": "Figure 6. Comparison between the proposed prototype-driven approach, ProLearn, and the LLM-based approach, SGSeg. Top: The theoretical analysis of time complexity. Bottom: The model size and inference time comparison." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 436, + 127, + 548, + 198 + ], + "blocks": [ + { + "bbox": [ + 436, + 127, + 548, + 198 + ], + "lines": [ + { + "bbox": [ + 436, + 127, + 548, + 198 + ], + "spans": [ + { + "bbox": [ + 436, + 127, + 548, + 198 + ], + "type": "image", + "image_path": "ff891e1d07e66ea52635536f76d2e4d8292a6b87723e20c87d942c70d94de6fc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 285, + 427, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 427, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 427, + 298 + ], + "type": "text", + "content": "5.3. Prototype vs. LLM" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": "Prototype learning enables constant-time inference of " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(1)" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": " because it uses a pre-defined, finite prototype space that is independent of both image and text size. At the same time, LLM applies autoregressive image-to-text generation, which has a linear time complexity " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(n)" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": " represents the number of tokens in the generated text (see Figure 6, top). As shown in Figure 6, bottom, prototype learning achieves an inference time of " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "4\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": ", making it " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": " faster than GPT-2 (136ms) and " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "300\\times" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": " faster than Llama3 (1.2s). Moreover, the prototype model has only 1M parameters, which is " + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "inline_equation", + "content": "1000\\times" + }, + { + "bbox": [ + 313, + 308, + 555, + 489 + ], + "type": "text", + "content": " smaller than large language models such as GPT-2 (1.5B parameters) and Llama3 (7B parameters). These properties make prototype learning highly efficient for real-time applications and suitable for deployment on resource-constrained edge devices." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 511, + 399, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 399, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 399, + 523 + ], + "type": "text", + "content": "5.4. Visualization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "Figure 7 presents a qualitative comparison of ProLearn with state-of-the-art approaches on the QaTa-COV19, MosMedData+ and Kvasir-SEG datasets. As textual availability decreased, existing language-guided segmentation models, which rely heavily on image-text pairs, experienced a pronounced drop in performance. In contrast, ProLearn effectively learned from both image-text and image-only data, demonstrating stable results even under limited textual guidance. The same findings can be observed in saliency maps that pinpoint the regions on which each model was focused. Under decreased text guidance, conventional models produced unstable or diffuse attention patterns, compromising their segmentation accuracy. However, ProLearn, aided by semantic approximation, retained more coherent attention focused on areas of the lesion." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22322" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 261 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 261 + ], + "type": "image", + "image_path": "c8ff8901badf26303f47fc3eff26e9167cb96d6d40de5a69b721d20f8c52f16d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 555, + 344 + ], + "type": "text", + "content": "Figure 7. Visualization comparison of language-guided segmentation methods under different text availabilities. The upper row shows segmentation outputs on CXR_S09687-E18404-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_4_85506_1(case19_16 from the MosMedData+ dataset (middle) and ju5xkwzxf0z0818gk4xabdm from the Kvasir-SEG dataset (right). Blue regions indicate groundtruth pixels missed by the model, while red regions indicate pixels mistakenly predicted as lesion. The lower row presents saliency map interpretability on CXR_S09346-E23164-R1 from the QaTa-COV19 dataset (left), Jun_radiopaedia_40_86625_0(case18_53 from the MosMedData+ dataset (middle) and cju2tzypl4wss0799ow05oxb9 from the Kvasir-SEG dataset (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 364, + 175, + 439 + ], + "blocks": [ + { + "bbox": [ + 66, + 364, + 175, + 439 + ], + "lines": [ + { + "bbox": [ + 66, + 364, + 175, + 439 + ], + "spans": [ + { + "bbox": [ + 66, + 364, + 175, + 439 + ], + "type": "image", + "image_path": "458e4c6ec4b7712acd49fb8eebfd8346d83baf62b5833ab23e2aae487e0ab5b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "type": "text", + "content": "Figure 8. Hyperparameter sensitivity analysis with varying candidate and prototype configurations. a) Effect of varying " + }, + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "type": "text", + "content": ", number of responding prototype vectors on Dice score; b) Effect of varying " + }, + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 446, + 295, + 491 + ], + "type": "text", + "content": ", number of prototypes per surrogate label on Dice score." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 363, + 290, + 438 + ], + "blocks": [ + { + "bbox": [ + 181, + 363, + 290, + 438 + ], + "lines": [ + { + "bbox": [ + 181, + 363, + 290, + 438 + ], + "spans": [ + { + "bbox": [ + 181, + 363, + 290, + 438 + ], + "type": "image", + "image_path": "6bb5ae8bebaa1c389b5f689e2b7b0ea4bb9df91d68972ce8e794c7df7517e492.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 209, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 209, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 209, + 528 + ], + "type": "text", + "content": "5.5. Hyperparameter Sensitivity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "To investigate the effect of the number of candidates per response " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": " and the number of prototypes per surrogate label " + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": ", we presented the performance comparison by varying their values for segmentation, as shown in Figure 8. We adjusted one hyperparameter at a time while keeping the other fixed. The experimental results indicated that our framework remained stable across a broad range of hyperparameter values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " governs the trade-off between information diversity and noise. A small " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " limits the model's ability to capture subtle segmentation patterns due to insufficient variability. As " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " increases, richer patterns are incorporated, improving performance. However, when " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " is too large, irrelevant patterns introduce excessive noise, leading to degraded segmentation accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "content": ", increasing the number of prototypes enhances performance by capturing diverse segmentation characteristics. However, when " + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "content": " exceeds a certain threshold, prototype redundancy arises, leading to overlapping or irrelevant representations that blur the segmentation boundaries. Thus, balancing " + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 364, + 555, + 461 + ], + "type": "text", + "content": " is crucial; too few prototypes hinder representational capacity, while an excessive count dilutes the model's focus on meaningful patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 388, + 487 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "type": "text", + "content": "We investigated the issue of textual reliance in medical language-guided segmentation, which limits both the applicability of segmentation in clinical workflows and the exploration of image-only training data. To address this issue, we have presented ProLearn, a prototype-driven framework that fundamentally alleviates textual reliance in medical language-guided segmentation. Our experiments demonstrated that ProLearn effectively learns from both image-only and image-text data, making efficient use of the target dataset's textual descriptions. ProLearn showed stable performance under different text availability and outperformed state-of-the-art image-only approaches under minimal text supervision." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Limitations and Future Works: In this work, we focused on demonstrating the core idea of ProLearn in single-label 2D segmentation. Future directions involve exploring multi-label segmentation, broader imaging modalities, and other language-guided vision tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22323" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 114 + ], + "type": "text", + "content": "This work was supported in part by the Australian Research Council (ARC) under Grant IM240100224." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 115, + 138 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 145, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 145, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 145, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 145, + 296, + 212 + ], + "type": "text", + "content": "[1] Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland, Yiwei Jia, Atlas Haddadi Avval, Afshin Bozorgpour, Sanaz Karimijafarbigloo, Joseph Paul Cohen, Ehsan Adeli, and Dorit Merhof. Medical image segmentation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):10076-10095, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 267 + ], + "type": "text", + "content": "[2] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 21438-21451, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 295, + 312 + ], + "type": "text", + "content": "[3] Ricardo JGB Campello, Davoud Moulavi, and Jörg Sander. Density-based clustering based on hierarchical density estimates. In Pacific-Asia conference on knowledge discovery and data mining, pages 160–172. Springer, 2013. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 313, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 295, + 378 + ], + "type": "text", + "content": "[4] Sema Candemir, Stefan Jaeger, Kannappan Palaniappan, Jonathan P Musco, Rahul K Singh, Zhiyun Xue, Alexandros Karargyris, Sameer Antani, George Thoma, and Clement J McDonald. Lung segmentation in chest radiographs using anatomical atlases with nonrigid registration. IEEE transactions on medical imaging, 33(2):577-590, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 423 + ], + "type": "text", + "content": "[5] H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In European conference on computer vision, pages 205-218. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 424, + 295, + 467 + ], + "type": "text", + "content": "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 295, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 500 + ], + "type": "text", + "content": "[7] Jiasi Chen and Xukan Ran. Deep learning with edge computing: A review. Proceedings of the IEEE, 107(8):1655-1674, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 545 + ], + "type": "text", + "content": "[8] J. Chen, Y. Lu, Q. Yu, X. Luo, E. Adeli, Y. Wang, L. Lu, A. L. Yuille, and Y. Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 1, 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 547, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 601 + ], + "type": "text", + "content": "[9] Pujin Cheng, Li Lin, Junyan Lyu, Yijin Huang, Wenhan Luo, and Xiaoying Tang. Prior: Prototype representation joint learning from medical images and reports. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21361-21371, 2023. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 656 + ], + "type": "text", + "content": "[10] Aysen Degerli, Serkan Kiranyaz, Muhammad E. H. Chowdhury, and Moncef Gabbouj. Osegnet: Operational segmentation network for Covid-19 detection using chest x-ray images. In 2022 IEEE International Conference on Image Processing (ICIP), pages 2306–2310, 2022. 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 679 + ], + "type": "text", + "content": "[11] Lee R Dice. Measures of the amount of ecologic association between species. Ecology, 26(3):297-302, 1945. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Vincenzo Ferrari, Marina Carbone, Carla Cappelli, Luigi Boni, Franca Melfi, Mauro Ferrari, Franco Mosca, and Andrea Pietrabissa. Value of multidetector computed tomogra" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "phy image segmentation for preoperative planning in general surgery. Surgical endoscopy, 26:616-626, 2012. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "type": "text", + "content": "[13] Jing Gu, Eliana Stefani, Qi Wu, Jesse Thomason, and Xin Wang. Vision-and-language navigation: A survey of tasks, methods, and future directions. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7606-7623, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 226 + ], + "type": "text", + "content": "[14] Matthew Stephen Holden, Tamas Ungi, Derek Sargent, Robert C McGraw, Elvis CS Chen, Sugantha Ganapathy, Terry M Peters, and Gabor Fichtinger. Feasibility of real-time workflow segmentation for tracked needle interventions. IEEE Transactions on Biomedical Engineering, 61(6): 1720-1728, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 282 + ], + "type": "text", + "content": "[15] Jihong Hu, Yinhao Li, Hao Sun, Yu Song, Chujie Zhang, Lanfen Lin, and Yen-Wei Chen. Lga: A language guide adapter for advancing the sam model's capabilities in medical image segmentation. page 610–620, Berlin, Heidelberg, 2024. Springer-Verlag. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 338 + ], + "type": "text", + "content": "[16] Shih-Cheng Huang, Liyue Shen, Matthew P. Lungren, and Serena Yeung. Gloria: A multimodal global-local representation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3922-3931, 2021. 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[17] Paul Jaccard. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bull Soc Vaudoise Sci Nat, 37:547-579, 1901. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 426 + ], + "type": "text", + "content": "[18] Debesh Jha, Pia H Smedsrud, Michael A Riegler, Pål Halvorsen, Thomas De Lange, Dag Johansen, and Håvard D Johansen. Kvasir-seg: A segmented polyp dataset. In International conference on multimedia modeling, pages 451-462. Springer, 2019. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 553, + 470 + ], + "type": "text", + "content": "[19] Tina Kapur, Jan Egger, Jagadeesan Jayender, Matthew Toews, and William M Wells. Registration and segmentation for image-guided therapy. Intraoperative Imaging and Image-Guided Therapy, pages 79-91, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 534 + ], + "type": "text", + "content": "[20] Stijn Keereweer, Jeroen DF Kerrebijn, Pieter BAA Van Driel, Bangwen Xie, Eric L Kaijzel, Thomas JA Snoeks, Ivo Que, Merlijn Hutteman, Joost R Van Der Vorst, J Sven D Mieog, et al. Optical image-guided surgery—where do we stand? Molecular Imaging and Biology, 13:199-207, 2011. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 590 + ], + "type": "text", + "content": "[21] Zihan Li, Yunxiang Li, Qingde Li, Puyang Wang, Dazhou Guo, Le Lu, Dakai Jin, You Zhang, and Qingqi Hong. Lvit: language meets vision transformer in medical image segmentation. IEEE transactions on medical imaging, 2023. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "text", + "content": "[22] S. Lloyd. Least squares quantization in pcm. IEEE Transactions on Information Theory, 28(2):129-137, 1982. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 615, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 658 + ], + "type": "text", + "content": "[23] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "text", + "content": "[24] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Dhiraj Maji, Prarthana Sigedar, and Munendra Singh. Attention res-unet with guided decoder for semantic segmentation" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22324" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "of brain tumors. Biomedical Signal Processing and Control, 71:103077, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[26] Epimack Michael, He Ma, Hong Li, Frank Kulwa, and Jing Li. Breast cancer segmentation methods: current status and future potentials. *BioMed research international*, 2021(1): 9962109, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[27] Sergey Morozov, Anna Andreychenko, Nikolay Pavlov, Anton Vladzymyrskyy, Natalya Ledikhova, Victor Gombolevskiy, Ivan Blokhin, Pavel Gelezhe, Anna Gonchar, Valeria Chernina, et al. Mosmeddata: Chest ct scans with Covid-19 related findings. 2020. 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[28] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, and et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Pengcheng Qiu, Chaoyi Wu, Xiaoman Zhang, Weixiong Lin, Haicheng Wang, Ya Zhang, Yanfeng Wang, and Weidi Xie. Towards building multilingual language model for medicine. Nature Communications, 15(1):8384, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[30] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 397 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 465 + ], + "type": "text", + "content": "[32] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1, 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[33] N. Siddique, S. Paheding, C. P. Elkin, and V. Devabhaktuni. U-net and its variants for medical image segmentation: A review of theory and applications. IEEE Access, 9:82031-82057, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[34] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[35] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez, Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature medicine, 29(8):1930-1940, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "text", + "content": "[36] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[37] Zhaoqing Wang, Yu Lu, Qiang Li, Xunqiang Tao, Yandong Guo, Mingming Gong, and Tongliang Liu. Cris: Clipdriven referring image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11686-11695, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "text", + "content": "[38] Sudanthi Wijewickrema, Yun Zhou, James Bailey, Gregor Kennedy, and Stephen O'Leary. Provision of automated step-by-step procedural guidance in virtual reality surgery" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "simulation. In Proceedings of the 22nd ACM Conference on Virtual Reality Software and Technology, pages 69-72, 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[39] Zunnan Xu, Zhihong Chen, Yong Zhang, Yibing Song, Xiang Wan, and Guanbin Li. Bridging vision and language encoders: Parameter-efficient tuning for referring image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17503-17512, 2023. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[40] Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3474-3482, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 275 + ], + "type": "text", + "content": "[41] Li Yang, Shasha Liu, Jinyan Liu, Zhixin Zhang, Xiaochun Wan, Bo Huang, Youhai Chen, and Yi Zhang. Covid-19: immunopathogenesis and immunotherapeutics. Signal transduction and targeted therapy, 5(1):128, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "text", + "content": "[42] Shuchang Ye, Mingyuan Meng, Mingjian Li, Dagan Feng, and Jinman Kim. Enabling text-free inference in language-guided segmentation of chest x-rays via self-guidance. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 242-252. Springer, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "text", + "content": "[43] Xiao-Xia Yin, Le Sun, Yuhan Fu, Ruiliang Lu, and Yanchun Zhang. [retracted] u-net-based medical image segmentation. Journal of healthcare engineering, 2022(1):4189781, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 389, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 432 + ], + "type": "text", + "content": "[44] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 533 + ], + "type": "text", + "content": "[45] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, Cliff Wong, Andrea Tupini, Yu Wang, Matt Mazzola, Swadheen Shukla, Lars Liden, Jianfeng Gao, Angela Crabtree, Brian Piening, Carlo Bifulco, Matthew P. Lungren, Tristan Naumann, Sheng Wang, and Hoifung Poon. A multimodal biomedical foundation model trained from fifteen million image-text pairs. NEJM AI, 2(1):A1oa2400640, 2025. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "text", + "content": "[46] Xu Zhang, Bo Ni, Yang Yang, and Lefei Zhang. Madapter: A better interaction between image and language for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 425-434. Springer, 2024. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 668 + ], + "type": "text", + "content": "[47] Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Hin Lee, Ho Sid Kiblawi, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, Brian Piening, Carlo Bifulco, Mu Wei, Hoifung Poon, and Sheng Wang. A foundation model for joint segmentation, detection and recognition of biomedical objects across nine modalities. Nature Methods, 22(1):166-176, 2025. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[48] Yi Zhong, Mengqiu Xu, Kongming Liang, Kaixin Chen, and Ming Wu. Ariadne's thread: Using text prompts to improve segmentation of infected areas from chest x-ray images. In International Conference on Medical Image Computing and" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22325" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 184 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Computer-Assisted Intervention, pages 724-733. Springer, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 184 + ], + "type": "text", + "content": "[49] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support: 4th International Workshop, DLMIA 2018, and 8th International Workshop, ML-CDS 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, Proceedings 4, pages 3-11. Springer, 2018. 1, 2, 7" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22326" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_content_list.json b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b87831c9e60a79e72f277c1a487d364b733d9866 --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_content_list.json @@ -0,0 +1,1674 @@ +[ + { + "type": "text", + "text": "Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization", + "text_level": 1, + "bbox": [ + 109, + 128, + 885, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weiying Xie $^{1}$ Zihan Meng $^{1*}$ Jitao Ma $^{1}$ Wenjin Guo $^{1}$ Haowei Li $^{1}$", + "bbox": [ + 218, + 202, + 777, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haonan Qin $^{1}$ Leyuan Fang $^{2}$ Yunsong Li $^{1}$", + "bbox": [ + 321, + 220, + 671, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xidian University $^{2}$ Hunan University", + "bbox": [ + 341, + 239, + 656, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Quantization-aware Training (QAT) enables deep models to adapt to precision loss by simulating quantization. However, existing methods often converge to sub-optimal solutions due to inadequate exploration of quantization solution space. To address this, we propose a novel QAT method, Allowing Oscillation Quantization (AOQ), which expands the reachable solution space through weight oscillation. Notably, unlike previous methods that suppress oscillation throughout training, AOQ actively encourages it in the earlier stages to explore diverse quantization configurations, and suppresses it later to ensure convergence. In addition, by decoupling quantization thresholds and levels, AOQ promotes meaningful oscillation and improves the stability of learnable quantization parameters. Extensive experiments across various models, including ResNet, MobileNet, DeiT and Swin Transformer, demonstrate the effectiveness of our method. Specifically, with 2-bit quantization, AOQ achieves a $0.4\\% \\sim 2.2\\%$ accuracy improvement on ImageNet compared to state-of-the-art methods. Our implementation is available at https://github.com/muzenc/AOQ.", + "bbox": [ + 89, + 321, + 483, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 654, + 220, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Model quantization [31, 46, 49] has attracted growing attention for enabling efficient model inference while substantially reducing hardware resource consumption. By reducing the bit-width of model parameters, quantization effectively reduces memory footprint and computational complexity, thereby enhancing inference speed and energy efficiency. Thus, quantization has been widely applied in tasks such as image classification [27, 32, 39], object detection [3, 42, 48] and natural language processing [1, 10, 11].", + "bbox": [ + 89, + 680, + 482, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its advantages, quantization shrinks the dynamic range of parameters, often resulting in information loss and degraded performance. Quantization-aware training (QAT) significantly mitigates the impact of quantization on perfor", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30f7d84627e4b80f2d787dee8a1fbdeafc8a22528e7492819284c059066e6636.jpg", + "image_caption": [ + "- Discrete Solution", + "Solution Space", + "$\\bigcirc$ Contour", + "Expanded Solution Space", + "2D quantization problem:", + "subject to $q(\\cdot)$ is quantization function", + "Figure 1. Illustration of solving a 2D quantization problem using previous methods vs. proposed method. The two plots are contour maps of the function $f(w_{1},w_{2})$ , where darker regions indicate lower function value. The yellow and orange curves show the solution trajectories of the previous method and our method, respectively. In contrast, AOQ explores a broader solution space and finds a better solution." + ], + "image_footnote": [], + "bbox": [ + 522, + 295, + 901, + 438 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mance compared to post-training quantization [16, 24, 35], but a noticeable performance gap remains under extremely low-bit quantization scenarios, such as 2-bit quantization.", + "bbox": [ + 511, + 657, + 903, + 703 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mainstream QAT techniques [12, 32, 46] simulate quantization effects by inserting fake quantization nodes during training, allowing the model to adapt to low-precision constraints. However, these nodes are non-differentiable, leading prior works [13, 23, 45] to employ biased gradient approximations. Recent studies [8, 30, 36] show that such biased gradients can result in weight oscillation, hindering the network's convergence to a better local minimum. Nevertheless, an overlooked insight is that QAT relies on weight oscillation to explore more quantized solutions. In existing methods [12, 13, 18, 32], the frequency of weight oscillation is relatively low, and directly suppressing oscillations may limit the model's ability to escape suboptimal quanti", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24615", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zation bins. Specifically, weights seldom cross quantization thresholds, instead making insignificant shifts within quantization bins.", + "bbox": [ + 89, + 90, + 480, + 133 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We investigate QAT from the perspective of solving nonlinear mixed-integer programming problems, drawing an analogy to the branch-and-bound method with depth-first search (DFS). While DFS systematically explores all possible solutions through branching and backtracking, QAT lacks an exhaustive exploration mechanism and instead performs local updates guided by gradients. Our analysis reveals that in existing methods, most weights infrequently change their quantization bin assignments, thereby restricting the search to a small subset of possible quantization solutions and resulting in a limited solution space. As shown in Figure 1, this limited search space often leads to premature convergence, leaving potentially optimal configurations unexamined. According to [37], low bit-width quantization introduces sharp curvature in the loss landscape, indirectly demonstrating that gradient-based optimization is more likely to get trapped in poor local minima.", + "bbox": [ + 89, + 136, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problems mentioned above, we propose a novel QAT approach, named Allowing Oscillation Quantization (AOQ). Unlike previous methods that suppress weight oscillation throughout the entire optimization process, AOQ strategically leverages weight oscillation as a positive exploration mechanism. We encourage weight oscillation to expand the solution space of the weights and only suppress oscillation in the later stages of training. Moreover, AOQ addresses a key limitation of traditional QAT methods, where the coupling of quantization thresholds and levels often causes inaccurate updates to the quantization parameters and hinders model convergence [30]. To overcome this, AOQ decouples thresholds and levels, allowing thresholds to be manually adjusted to encourage oscillation while setting quantization levels as learnable parameters directly optimized with the task loss. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 393, + 482, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We empirically show that fake quantization-based QAT methods explore a limited subset of quantized solutions and are prone to suboptimal convergence. To the best of our knowledge, this is the first systematic demonstration of this phenomenon.", + "- We propose AOQ to expand the quantization solution space of the weights by leveraging weight oscillation. Moreover, we improve the stability of learnable quantization parameters by decoupling the thresholds and levels.", + "- Extensive experiments in CNN and Vision Transformer architectures demonstrate the effectiveness of AOQ in extremely low-bit quantization." + ], + "bbox": [ + 89, + 651, + 482, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 844, + 232, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "QAT was initially proposed for binary quantization [7], and has since become the most popular quantization method in", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "precision-sensitive scenarios. Most QAT-based methods insert fake quantization nodes into the model. These nodes simulate the effects of quantization during forward propagation and pass the approximated gradients during backward propagation [18, 19, 26, 27, 43, 46]. Some other methods [8, 38] add independent pseudo quantization noise during training, enabling effective QAT without relying on gradient approximations. This work focuses on the former.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The former is non-differentiable at the fake quantization nodes, preventing optimization and convergence of the quantized model. To address this problem, earlier works approximate gradients using straight-through estimator (STE) [22, 45] and its variants [13, 23, 44], while other approaches compute gradients using constructed neural networks or manual derivation [4, 25]. Among them, STE is widely used in quantization methods due to its robustness, broad applicability, and straightforward implementation. However, it inevitably brings the problem of gradient mismatch: gradients of weights are computed based on their quantized values, instead of their floating-point values. Previous research demonstrated that following the STE approach can cause weight oscillation during training due to biased gradients [8]. It has been suggested that weight oscillation prevents the network from converging to the best local minimum, and that suppressing oscillation earlier is beneficial for the optimization process [36]. The oscillation dampening and iterative weight freezing methods it proposed overcome weight oscillation. In addition, OFQ [30] observed that weight oscillation can introduce noise to the gradient of the learnable scaling factor (quantization level) and cause it to oscillate. The latter, in turn, changes the quantization threshold and aggravates the weight oscillation. OFQ deems weight oscillation harmful and suppresses it; in contrast, AOQ dispels this long-standing misconception, theoretically and experimentally proving its value.", + "bbox": [ + 511, + 215, + 903, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, several works have delved into searching for optimal scaling factor to determine quantization thresholds and levels. For instance, trainable clipping parameters have been introduced to determine the appropriate quantization scale [5, 49]. Esser et al. [12] provided a simple yet effective way to approximate the gradient of learnable scaling factor, which is sensitive to quantized state transitions. It has been widely adopted as the de facto approach for QAT. Another line of research decouples quantization thresholds and levels, proposing increasing the quantization representation ability by replacing uniform quantization with nonuniform quantization [32]. The approach has demonstrated excellent performance in quantization of activation values.", + "bbox": [ + 511, + 626, + 903, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While we acknowledge that weight oscillation impacts final convergence, we argue that it is not entirely detrimental to optimization. In layer-by-layer quantization, each weight represents a dimension within the quantization solution space, and weight oscillation plays a crucial role in", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rapidly exploring optimal quantization solutions within limited training iterations. Moreover, our research shows that in existing methods, the oscillation frequency of weights at quantization thresholds is quite low, motivating us to increase this frequency. To this end, we delve into the working mechanism of QAT, and our proposed AOQ method combines the benefits of enhancing weight oscillation to expand the solution space and suppressing weight oscillation to ensure stable convergence.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 238, + 225, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 262, + 230, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One of the most effective ways to quantize a neural network is by inserting fake quantization nodes to simulate the quantization effect during training. In weight quantization, the goal is to identify the quantization bin each weight falls into and assign it the corresponding quantization level. At present, the most successful methods are based on LSQ [12], and its quantization function is:", + "bbox": [ + 89, + 284, + 483, + 390 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} = q (w; s, Q _ {n}, Q _ {p}) = s \\cdot c l i p \\left(\\left\\lfloor \\frac {w}{s} \\right\\rceil , Q _ {n}, Q _ {p}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 393, + 482, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $clip(\\cdot, \\alpha, \\beta)$ is a truncated function with lower bound $\\alpha$ and upper bound $\\beta$ , $\\lfloor \\cdot \\rceil$ is a rounding operator, $s$ is a scaling factor, $Q_{n}$ and $Q_{p}$ are the quantization ranges.", + "bbox": [ + 89, + 426, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the quantization function is non-differentiable, the gradient of $w$ cannot be directly computed during backpropagation. Thus, gradient approximation is used to keep the training going, such as straight through estimator (STE) [2]:", + "bbox": [ + 89, + 472, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial w} \\approx \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\mathbf {1} _ {Q _ {n} \\leq w / s \\leq Q _ {p}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 535, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{1}$ is an indicator function. When $Q_{n} \\leq w / s \\leq Q_{p}$ , it is 1, otherwise it is 0. There is no gradient for $w$ outside the quantization range. In LSQ, the gradient of the scaling factor $s$ is computed as:", + "bbox": [ + 89, + 569, + 483, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial L}{\\partial s} = \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial s},\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 633, + 254, + 665 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\hat {w}}{\\partial s} = \\left\\{ \\begin{array}{c l} - \\frac {w}{s} + \\left\\lfloor \\frac {w}{s} \\right\\rceil & \\text {i f} Q _ {n} < \\frac {w}{s} < Q _ {p} \\\\ Q _ {n} & \\text {i f} \\frac {w}{s} \\leq Q _ {n} \\\\ Q _ {p} & \\text {i f} \\frac {w}{s} \\geq Q _ {p} \\end{array} . \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 666, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limited Solution Space of Weights", + "text_level": 1, + "bbox": [ + 89, + 718, + 393, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural network quantization can be formulated as a mixed-integer nonlinear programming (MINLP) problem. It involves converting floating-point weights and activations into a finite set of discrete values while minimizing performance degradation. In this formulation, quantized weights and activations act as integer variables, while the nonlinearity stems from forward propagation and activation functions. Continuous variables, such as quantization parameters, are also involved. The optimization objective is:", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {w}} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\mathcal {L} \\left(f _ {\\text {q u a n t}} (x; q (\\mathbf {w})), y\\right) \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 882, + 482, + 904 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ is the loss function, $\\mathbf{w}$ is the set of all weights, $(x,y)\\sim D$ represents samples drawn from the dataset $D$ , $f_{\\mathrm{quant}}(\\cdot)$ represents the model's output with activation quantization, and $q(\\cdot)$ is the quantization function. $q(\\mathbf{w})$ consists of integer variables $\\mathbf{Q}$ and continuous parameters $s,z$ :", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\forall \\mathbf {Q} \\in [ 0, 2 ^ {b} - 1 ], \\mathbf {Q} \\in \\mathbb {Z}, b \\in \\mathbb {N} _ {*}, q (\\mathbf {w}) = s \\cdot (\\mathbf {Q} - z), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 176, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}$ is the set of integers mapped from $\\mathbf{w}$ , $s$ is the scaling factor, $z$ is the zero point, and $b$ is the bit-width.", + "bbox": [ + 511, + 205, + 903, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "QAT-based methods serve as heuristics for solving this MINLP problem. These methods select an initial solution and iteratively refine it. Drawing inspiration from research on mixed-integer programming [6, 47], we observe that QAT shares similarities with depth-first search (DFS) in branch-and-bound algorithms. Both rely on branching and backtracking: QAT chooses quantization strategies per weight (branching), and uses gradient updates (implicit backtracking) to adjust solutions. However, unlike exact DFS, which systematically explores the entire solution space, QAT heuristics prioritize fast convergence, often leading to limited exploration and suboptimal results. We compare QAT and DFS from two key perspectives:", + "bbox": [ + 511, + 236, + 905, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Branching of variables In DFS, the algorithm explicitly branches on relaxed variables to construct the search tree. Likewise, in QAT, weights oscillating near quantization thresholds effectively induce branching behavior, determining whether a weight falls into one bin or another.", + "bbox": [ + 511, + 431, + 905, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Backtracking mechanism DFS performs explicit backtracking: when a branch leads to no improvement, it reverts and explores alternatives, enabling a global search. In contrast, QAT relies on local gradients to implicitly adjust weights. This often causes the model to converge prematurely and miss better configurations.", + "bbox": [ + 511, + 508, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, both DFS and QAT can be regarded as rule-based search strategies. However, under low-bit constraints, QAT's gradient-based updates are more prone to becoming trapped in local optima due to the steep curvature of the loss surface [37]. Motivated by this limitation, we aim to enhance QAT by expanding its effective solution space, similar in spirit to the global search behavior exhibited by exact DFS, in order to achieve improved performance.", + "bbox": [ + 511, + 598, + 905, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assess the ability of existing methods to explore the weight solution space by measuring how often weights cross quantization thresholds. As shown in Table 1, we use LSQ [12] to quantize ResNet-18 on the ImageNet [9]. \"SGD\" denotes an SGD optimizer with an initial learning rate of $1\\mathrm{e} - 2$ and cosine learning rate decay [34], while \"Adam\" denotes an Adam optimizer with a learning rate of $1\\mathrm{e} - 3$ , both following the training strategies in [12]. We find that the frequency of weights crossing quantization thresholds is very low, all below $0.5\\%$ , with most untruncated weights concentrated in the $[0,0.01\\%)$ range. This indicates that previous gradient-based optimization methods have a", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24617", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b2032bd9f14ac44bfa9db6a342421f95bcc094f4e0faaa95608cd76bf6e85bbc.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 101, + 98, + 352, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1debe54953a29a64dfce8ac3bca93787021eb555642c9246736f47024319937.jpg", + "image_caption": [ + "(b)", + "Figure 2. Histograms of pre-trained weights mapped to different quantization values: (a) lossless solution; (b) LSQ; and (c) our AOQ. The red, green, blue, and yellow bars represent histograms of pre-trained weights mapped to the quantized values -2, -1, 0, and 1, respectively." + ], + "image_footnote": [], + "bbox": [ + 370, + 98, + 625, + 267 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2500b8fb35bf5f9d3a718943f21fa2750211f5d62f31b9404771397013e2e587.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 651, + 98, + 901, + 267 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/edcc9962ea0f8d79bf1169b6d79d2c8a4f8c85148312e64ed8c9ee86fe7688fb.jpg", + "table_caption": [ + "Table 1. Measuring the solution space of weights on 2-bit quantized ResNet-18. The first column is the frequency of weights crossing the quantization thresholds, calculated as the number of crossings divided by the total iterations. \"Truncated\" refers to weights whose frequency of crossing quantization thresholds falls between 0 and $0.01\\%$ due to being clipped by the truncation function. \"Not Truncated\" refers to not being clipped by the truncation function. $\\%$ of weights\" refers to the percentage of weights within a specific frequency range." + ], + "table_footnote": [], + "table_body": "
Frequency Range% of weights
SGDAdam
[0, 0.01%) - Truncated14.5%11.8%
[0, 0.01%) - Not Truncated47.2%12.6%
[0.01%, 0.05%)5.1%19.1%
[0.05%, 0.1%)8.4%24.5%
[0.1%, 0.5%)24.8%32.0%
", + "bbox": [ + 122, + 489, + 447, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "weak ability to explore the solution space, potentially hindering the model from finding an optimal solution.", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further illustrate how limited exploration affects quantization quality, we construct a toy example. We first generate an integer set $\\bar{w} \\in \\mathbb{R}^{10000}$ containing only values -2, -1, 0, and 1. We then derive matrix $\\mathbf{A} \\in \\mathbb{R}^{2000 \\times 10000}$ and vector $\\mathbf{b} \\in \\mathbb{R}^{2000}$ satisfying $\\mathbf{A}\\bar{w} = \\mathbf{b}$ , establishing $\\bar{w}$ as our known lossless integer solution. Clearly, $\\mathbf{A}$ is not a full-rank matrix, which indicates that the linear equation system has multiple solutions, similar to the multi-solution nature of neural networks. Solving $\\mathbf{A}w = \\mathbf{b}$ with the constraint $w \\in \\{-2s + z, -s + z, z, s + z\\}^{10000}$ , where $s, z \\in \\mathbb{R}$ , simulates the quantization process. Conceptually, this setup mimics training a fully connected network with 2-bit quantization, 10,000 input units, one output unit, and 2,000 samples, using mean squared error loss and Adam optimizer with a learning rate of $1 \\times 10^{-3}$ .", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first train the model with floating-point weights until convergence, obtaining pre-trained weights, similar to the standard QAT procedure. These weights are then mapped to quantized values. In the lossless solution, the scaling factor $s$ is 1 and $z$ is 0. Figure 2a shows the distribution of pre-trained weights mapped to each quantized value under the lossless solution. Each quantization level corresponds to a wide range of pre-trained weights, indicating that allowing distant weights to map to the same quantized value can preserve model quality. In contrast, Figure 2b shows the distribution after freezing $s$ to 1 and $z$ to 0 and using LSQ [12] for quantization. In this setting, pre-trained weights are mapped only to the nearest quantized values, restricting the solution space. Such methods, by mapping only to nearby values, may result in more mapping errors and does not fully leverage the broader solution space to minimize quantization errors.", + "bbox": [ + 511, + 354, + 906, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Allowing Oscillation Quantization", + "text_level": 1, + "bbox": [ + 511, + 622, + 812, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In previous methods, many weights seldom oscillate near quantization thresholds, limiting their ability to switch quantization bins and often resulting in suboptimal solutions. We speculate that in QAT, especially in low-bit settings, weight transitions between different quantization levels make gradient direction unstable. This instability causes weights to fluctuate within a narrow range during training, preventing them from crossing thresholds. While some studies [30, 36] consider weight oscillation harmful to the whole optimization process, we believe that oscillation actually promotes solution diversity, increasing the likelihood of discovering better optima. Specifically, oscillation acts as an adaptive perturbation mechanism that enables the model to continuously test new quantization configurations and explore more quantized solutions within a given number of iterations, helping avoid premature convergence. This phenomenon resembles \"random perturbation\" strategies used", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24618", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f14cd86f4a1f22fc20c3ef8920b81b9f19d6f5b71c3f207fb42164909c344abe.jpg", + "image_caption": [ + "Figure 3. Oscillation patterns of two weights in layer1.0.conv1 of ResNet-18 during early training. We apply AOQ under a 2-bit quantization setting and monitor quantization assignments at the end of each epoch." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 478, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in integer programming [21, 40], where moderate instability prevents the model from becoming locked into overly deterministic paths. Based on this intuition, we propose to actively promote oscillation during early training phases to encourage exploration, then gradually reduce it in later stages to ensure convergence stability.", + "bbox": [ + 88, + 321, + 482, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To increase the number of weights that cross thresholds without disrupting QAT's gradient-based backtracking mechanism, we examine how quantization parameters affect oscillation. Inspired by [30], we observe that modifying quantization thresholds can induce more weights to oscillate around them, as shown in Figure 3. Specifically, when the intervals between quantization thresholds and the differences among quantization levels decrease, weights tend to transition from intra-bin oscillation to threshold-adjacent oscillation. This behavior is driven by two factors. First, the shifting thresholds force certain weights to cross them, introducing noise into the network. The facilitating effect of smaller quantization threshold intervals on weight oscillation will be demonstrated in Section 3.4. Second, the reduced differences between quantization levels lower the penalty for mapping errors when floating-point weights are mapped to quantized values, making weights more likely to oscillate near the thresholds, as smaller differences between quantization levels tend to smooth the loss surface.", + "bbox": [ + 91, + 412, + 483, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To leverage the weight oscillation mentioned above, we decouple the quantization thresholds and levels and introduce a three-stage training strategy. Specifically, we manually adjust the quantization thresholds to gradually reduce their intervals, while setting the quantization levels as learnable parameters to optimize during training. Additionally, we introduce differentiability for clipped weights in fake quantization to address the issue observed in Section 3.2, where the absence of gradients in clipped weights (as shown in Eq. (2)) hinders model fine-tuning. In the final stage, oscillation is suppressed to stabilize convergence. The complete strategy is outlined below:", + "bbox": [ + 89, + 700, + 482, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Stage 1: Inducing oscillation. For weight quantization,", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we define the intervals between thresholds as $s_{th}$ and between quantization levels as $s_{le}$ , both initialized to $\\frac{\\sigma_w}{2^{b - 2}}$ where $\\sigma_w$ is the standard deviation of weights in a layer, and $b$ is the bit-width. In this stage, we manually reduce $s_{th}$ and $s_{le}$ to promote oscillation at thresholds.", + "bbox": [ + 524, + 90, + 903, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Stage 2: Learning quantization levels. We fix $s_{th}$ and set $s_{le}$ as a learnable parameter for optimization. In Stage 1, $s_{le}$ is manually reduced, which leads to higher loss due to poor representation of weights by the quantization levels. As training progresses and the task loss decreases, $s_{le}$ is learned to increase automatically. This widens the intervals between quantization levels, makes the loss landscape steeper near thresholds, and reduces oscillation by discouraging weights from frequently crossing thresholds. The quantizer is defined as:", + "bbox": [ + 513, + 167, + 906, + 318 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {w} = \\left\\{ \\begin{array}{l l} \\operatorname {l e v e l} [ 0 ] & \\text {i f} w < \\operatorname {t h r e s h} [ 0 ] \\\\ \\operatorname {l e v e l} [ i ] & \\text {i f t h r e s h} [ i - 1 ] \\leq w < \\operatorname {t h r e s h} [ i ] \\\\ \\operatorname {l e v e l} \\left[ 2 ^ {b} - 1 \\right] & \\text {i f} w \\geq \\operatorname {t h r e s h} \\left[ 2 ^ {b} - 2 \\right] \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 329, + 903, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i\\in \\{1,2,\\ldots ,2^b -2\\}$ and $\\mathrm{thresh}[i]$ and level[i] represent the $(\\mathrm{i + 1})$ -th threshold and level, respectively. The optimized quantized weights are denoted as:", + "bbox": [ + 526, + 397, + 903, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {w} = \\hat {w} + w - w. \\det a c h (\\cdot). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 455, + 903, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This formulation, based on the STE, enables distinct gradient flows: $\\hat{w}$ determines the gradient for $s_{le}$ , while $w$ receives gradients directly from the task loss, even when clipped. This decoupling avoids the zero-gradient issue and enables more effective optimization of both weights and quantizer parameters.", + "bbox": [ + 524, + 483, + 905, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Stage 3: Dampening oscillation. In the third stage, we employ the Oscillation dampening method outlined in [36], which adds a regularization term to the loss:", + "bbox": [ + 513, + 574, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda \\mathcal {L} _ {\\text {d a m p e n}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 632, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter. The dampening loss is:", + "bbox": [ + 526, + 660, + 877, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d a m p e n}} = \\left\\| \\hat {w} - \\operatorname {c l i p} \\left(w, Q _ {n}, Q _ {p}\\right) \\right\\| _ {F} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 685, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which penalizes weights lingering near quantization thresholds, encouraging them to converge to the nearest quantization centers.", + "bbox": [ + 524, + 715, + 903, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Details regarding the hyperparameters and transition schedule for each training stage are provided in Section 4.1.", + "bbox": [ + 511, + 763, + 903, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 511, + 801, + 707, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we theoretically analyze the effectiveness of our approach and compare it against prior differentiable quantization parameter methods.", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first show the advantage of using smaller quantization threshold intervals. Let $W$ be the set of pre-trained", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24619", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "weights, and $Q$ be a set composed of quantized values. Weight quantization is a mapping from the $W$ to the $Q$ . For a given weight $w \\in W$ , we define its probability of being quantized to $Q_{i} \\in Q$ as the transition probability $P(Q_{i} \\mid w)$ , where $i \\in \\{0, 1, \\dots, 2^{b} - 1\\}$ and $b$ is the bit-width. By Bayes' theorem:", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{P (w)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 191, + 482, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P(Q_{i})$ is the proportion of the number of weights mapped to $Q_{i}$ to the total number of weights, $P(w)$ is the probability distribution of weights, and $P(w|Q_i)$ is the conditional distribution of weights mapped to $Q_{i}$ . The total probability formula applied to Eq. (10) is:", + "bbox": [ + 89, + 237, + 483, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{\\sum_ {i} P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 323, + 482, + 357 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 2a, $P(w \\mid Q_i)$ approximately follows a Gaussian distribution, so the total probability formula can be transformed into a Gaussian mixture model:", + "bbox": [ + 89, + 368, + 483, + 412 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP (w) = \\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i}, \\sigma_ {i}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 422, + 482, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\pi_i, \\mu_i, \\sigma_i$ denote the weight, mean and variance of the Gaussian component, respectively. This transformation does not effect the theoretical conclusion, just for analytical convenience. Eq. (11) is transformed into:", + "bbox": [ + 89, + 465, + 483, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(Q _ {i} \\mid w\\right) = \\frac {\\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}{\\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 537, + 482, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The weight $w$ is assigned to the quantization value $Q_{j}$ that maximizes this probability:", + "bbox": [ + 89, + 580, + 483, + 612 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ni f \\forall i, \\pi_ {j} \\mathcal {N} (w \\mid \\mu_ {j}, \\sigma_ {j}) \\geq \\pi_ {i} \\mathcal {N} (w \\mid \\mu_ {i}, \\sigma_ {i}) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 623, + 482, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "then $w\\to Q_j$", + "bbox": [ + 129, + 643, + 230, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, when the quantization threshold is set at the horizontal coordinate value of the intersection between two adjacent Gaussian components, it ensures optimal probabilistic mapping of weights. In Figure 2a, we observe that in the lossless solution for this example, the variances of each Gaussian component are approximately equal, denoted as $\\sigma$ . Thus, the horizontal coordinate value at the intersection of adjacent Gaussian components is:", + "bbox": [ + 89, + 670, + 483, + 791 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {t h r e s h} [ i ] = \\frac {\\mu_ {i + 1} + \\mu_ {i}}{2} - \\frac {\\log \\left(\\frac {\\pi_ {i + 1}}{\\pi_ {i}}\\right) \\cdot \\sigma^ {2}}{\\mu_ {i + 1} - \\mu_ {i}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 801, + 482, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This analysis suggests that, when Gaussian component means are close, using smaller threshold intervals improves mapping accuracy. It explains why our method encourages", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tighter thresholds in early training to support more effective weight assignment in QAT.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we show that AOQ provides a more direct and accurate link between quantization levels and task loss than previous approaches. In prior methods [5, 12, 49], learnable quantization parameters jointly determine both thresholds and levels. This coupling introduces approximation errors in the gradient computation of quantization levels. For instance, in LSQ [12], the gradients of the scaling factors are shown in Eq. (3). The derivative of weights with respect to the scaling factor aids in optimizing quantization thresholds but makes the gradient of the quantization levels inaccurate. In contrast, AOQ decouples thresholds from levels. It directly optimizes quantization levels based on Eq. (6) and Eq. (7), allowing a more faithful gradient flow that better aligns with task loss.", + "bbox": [ + 511, + 121, + 905, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, AOQ effectively addresses the impact of weight oscillation on final convergence, as well as on batch normalization and layer normalization. After exploring a broader solution space, we use Oscillation Dampening in the final stage to gradually guide weights out of oscillatory regions. As training continues, the exponential moving average used in batch normalization and layer normalization becomes increasingly stable, reinforcing consistent model behavior and improving convergence quality.", + "bbox": [ + 511, + 332, + 905, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 481, + 645, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the proposed Allowing Oscillation Quantization (AOQ) on the ILSVRC12 ImageNet dataset [9], using a range of models including ResNet [15], MobileNet-V2 [17], DeiT [41], and Swin Transformer [33]. Our experiments focus on low-bit weight quantization. For activation quantization, we adopt LSQ [12] and N2UQ [32].", + "bbox": [ + 511, + 507, + 905, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 606, + 702, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ResNet and MobileNet We follow the training schemes in [5, 12, 32], initializing quantized networks from full-precision pre-trained models. Training is performed using the Adam optimizer [20] with a linear learning rate decay, starting from $1 \\times 10^{-3}$ . Following LSQ [12], we apply knowledge distillation and train for 150 epochs. All convolutional and fully connected layers are quantized, except for the first and last layers.", + "bbox": [ + 511, + 628, + 905, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DeiT and Swin Transformer We use the official implementations and adopt training settings from [28-30], including knowledge distillation and a training duration of 300 epochs. As in prior work, the patch embedding and classification layers are quantized to 8 bits [28-30].", + "bbox": [ + 511, + 750, + 905, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parameter Setup The batch size is set to 512, and the weight decay is set to 0. The three training stages occupy $1/5$ , $2/5$ , and $2/5$ of the total training epochs, respectively. We reduce $s_{\\mathrm{th}}$ and $s_{\\mathrm{le}}$ to $30\\% - 35\\%$ of their initial values in the first stage. As shown in Fig. 4, a broad grid search", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24620", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42e30db4e0e4a1713efc32754cb24c6d3cb49dc78904fb680b9a3bb0dd68681c.jpg", + "table_caption": [ + "Table 2. Accuracy comparison to the state-of-the-art quantization methods with ResNet structure on ImageNet dataset. Note that W/A denote the bit-widths of weights and activations, respectively. FP denotes the top-1 accuracy of the full-precision models." + ], + "table_footnote": [], + "table_body": "
NetworkMethodBit-width (W/A)Accuracy (%) Top-1Bit-width (W/A)Accuracy(%) Top-1Bit-width (W/A)Accuracy(%) Top-1
ResNet-18 (FP: 71.8)LQ-Nets [46]2/264.93/368.24/469.3
DSQ [13]2/265.23/368.74/469.6
QIL [19]2/265.73/369.24/470.1
APoT [27]2/267.33/369.94/470.7
LSQ [12]2/267.63/370.24/471.1
LCQ [43]2/268.93/370.64/471.5
N2UQ [32]2/269.43/371.94/472.9
LSQ+AOQ2/269.73/371.44/472.0
N2UQ+AOQ2/270.33/372.14/472.7
ResNet-34 (FP:74.9)LQ-Nets [46]2/269.83/371.94/4-
DSQ [13]2/270.03/372.54/472.8
QIL [19]2/270.63/373.14/473.7
APoT [27]2/270.93/373.44/473.8
LSQ [12]2/271.63/373.44/474.1
LCQ [43]2/272.73/374.04/474.3
N2UQ [32]2/273.33/375.24/476.0
LSQ+AOQ2/273.53/374.94/475.5
N2UQ+AOQ2/273.93/375.54/476.0
ResNet-50 (FP: 77.0)LQ-Nets [46]2/271.53/374.24/475.1
APoT [27]2/273.43/375.84/476.6
LSQ [12]2/273.73/375.84/476.7
LCQ [43]2/275.13/376.34/476.6
N2UQ [32]2/275.83/377.54/478.0
LSQ+AOQ2/276.33/377.54/477.9
N2UQ+AOQ2/276.43/377.84/478.1
", + "bbox": [ + 140, + 127, + 856, + 510 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83a4b8f167de8458948e7fdcaed756a05136101bab787c9e1e07d387073d366e.jpg", + "image_caption": [ + "(a) ResNet-18" + ], + "image_footnote": [], + "bbox": [ + 94, + 541, + 282, + 647 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4fd1ea6ad70f8a2c172724c4ffdeb0ed1df4dfdac76f34074340a01ad5201a9.jpg", + "image_caption": [ + "(b) DeiT-T", + "Figure 4. Accuracy heatmaps on ImageNet under different narrowing schedules. $\\alpha$ denotes the final-to-initial ratio of $s_{le}$ and $s_{th}$ in stage one, and $\\beta$ is the proportion of total epochs allocated to this stage. AOQ maintains stable performance across a wide range of $(\\alpha, \\beta)$ settings on both (a) 2-bit ResNet-18 and (b) 2-bit DeiT-T, indicating schedule robustness across architectures." + ], + "image_footnote": [], + "bbox": [ + 282, + 541, + 480, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "confirms that a common schedule performs robustly across different architectures in practice. For the final stage, the hyper-parameter $\\lambda$ is chosen such that $\\lambda \\mathcal{L}_{\\mathrm{dampen}}$ is two to three orders of magnitude lower than $\\mathcal{L}_{\\mathrm{task}}$ .", + "bbox": [ + 89, + 779, + 483, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Experimental Results on Toy Example", + "text_level": 1, + "bbox": [ + 89, + 848, + 421, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first revisit the toy example from Section 3.2 to evaluate the effect of oscillation on quantization quality. To", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78e094382a8c8fb6520420029da6247389ac8e95ae55767ad42c1e963d380ba0.jpg", + "image_caption": [ + "Figure 5. Loss curves and minimum loss obtained from solving the toy example using various methods." + ], + "image_footnote": [], + "bbox": [ + 535, + 546, + 712, + 650 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5b216ad1fa8d5e32ef0a8286d72767051cc4043c71ff926a8a6db47c0352450f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodLoss
LSQ51.2
LSQ+OD(final)16.7
LSQ+OD(enti)26.2
AOQ5.6
", + "bbox": [ + 718, + 553, + 893, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrate that weight oscillation can expand the solution space and improve quantization outcomes, we apply the Oscillation Dampening (OD) [36] to suppress oscillation either during the final training stage or throughout the entire training process. We also evaluate the performance of AOQ using the toy example. As shown in Figure 5, allowing early-stage oscillation improves QAT performance, and AOQ converges faster to a better solution. When using fixed quantization levels $\\{-2, -1, 0, 1\\}$ in the third stage, Figure 2c shows that AOQ enables pre-trained weights to be mapped to more diverse and distant quantized values, expanding the solution space compared to Figure 2b.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24621", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f1235511c25efbed4c1c06f165bf1ea4d72d262f58a3a8e2ac3bc35defe3772.jpg", + "table_caption": [ + "Table 3. Top-1 accuracy (\\%) comparison for methods suppressing weight oscillation throughout the process in MobileNet-V2." + ], + "table_footnote": [], + "table_body": "
NetworkMethodW3A3W4A4
MobileNet-V2 (FP: 72.1)LSQ [12]65.369.5
LSQ+BR [14]67.470.4
LSQ+Dampen [36]67.870.5
N2UQ [32]69.072.1
LSQ+AOQ69.171.8
N2UQ+AOQ69.972.5
", + "bbox": [ + 104, + 128, + 468, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f70394efd0c25523d87cef901ae77f84170a43045d659c3ff276c82358987193.jpg", + "table_caption": [ + "Table 4. Top-1 accuracy (\\%) comparison with the state-of-the-art methods in the ViT architecture." + ], + "table_footnote": [], + "table_body": "
NetworkMethodW2A2W3A3
DeiT-T (FP: 72.1)LSQ [12]54.568.1
QViT* [28]50.467.1
OFQ [30]64.372.7
LSQ+AOQ66.572.3
DeiT-S (FP: 79.9)LSQ [12]68.077.8
QViT* [28]68.778.5
OFQ [30]75.779.6
LSQ+AOQ76.779.8
Swin-T (FP: 81.2)LSQ [12]70.479.0
QViT* [28]73.980.1
OFQ [30]78.581.1
LSQ+AOQ78.981.2
", + "bbox": [ + 101, + 303, + 472, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Accuracy Evaluation", + "text_level": 1, + "bbox": [ + 89, + 559, + 290, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2 reports the accuracy of AOQ on ResNet. When paired with different activation quantization techniques, AOQ consistently improves performance. For instance, using AOQ with LSQ yields $0.9\\% - 2.6\\%$ accuracy gains. When combined with N2UQ, AOQ outperforms many traditional methods. Its advantage becomes more pronounced in low-bit networks with a large number of weights, where the expanded solution space is more effectively utilized. In the 2-bit ResNet-50 case, AOQ reduces the accuracy gap with full-precision to just $0.6\\%$ .", + "bbox": [ + 88, + 580, + 482, + 733 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the performance of the AOQ method on MobileNet-V2. Compared to other methods, AOQ has a natural advantage in quantizing MobileNet. Due to significant variations in weight distributions across channels in the depth-wise separable convolution layers, traditional quantization methods often map all weights in certain channels to a single quantized value, resulting in the loss of important information. AOQ mitigates this by adaptively shrinking threshold intervals during training, promoting threshold-crossing behavior and improving quantization granularity. This effect is illustrated in Figure 3. Overall, AOQ outper", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/98f6b5a351f9f2c0dc7a2d7101355c22621277de1ac14fbac46e37a25c8bcc51.jpg", + "table_caption": [ + "Table 5. Effects of different components of our strategy on the final performance." + ], + "table_footnote": [], + "table_body": "
MethodAcc (%)
Baseline [12]73.7
- Weight quantization76.6
+ Expanding solution space of weights75.2
+ Delayed weight oscillation suppression75.6
+ Both76.3
Corresponding real-valued network77.0
", + "bbox": [ + 539, + 128, + 879, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "forms other QAT approaches, including those that suppress oscillation throughout [14, 36].", + "bbox": [ + 511, + 272, + 903, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate AOQ on Vision Transformers (ViT), we conducted 2-bit and 3-bit quantization experiments on DeiT and Swin Transformer. We observe that QViT [28] does not fully quantize activations in the fully connected and attention layers, limiting hardware acceleration. For fair comparison, we modify their implementation accordingly and report the results as QViT* in Table 4. The results show that AOQ significantly outperforms LSQ and QViT by suppressing weight oscillation, boosting 2-bit DeiT-T accuracy by $12.0\\%$ and $16.1\\%$ , respectively. By expanding the solution space, AOQ also surpasses OFQ, the prior state-of-the-art, under 2-bit quantization.", + "bbox": [ + 511, + 303, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 492, + 666, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate the impact of each improvement on the final performance, we conduct an ablation study on the 2-bit quantized ResNet-50 network on ImageNet. As shown in Table 5, the proposed expanded solution space and delayed weight oscillation suppression improve the accuracy by $1.5\\%$ and $1.9\\%$ , respectively over the baseline [12]. By using both techniques, we reduced the accuracy gap with the corresponding real-valued ResNet-50 to only $0.7\\%$ , and the accuracy is only reduced by $0.3\\%$ compared to the baseline method without weight quantization.", + "bbox": [ + 511, + 513, + 908, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we reveal the weakness of existing QAT methods in exploring the quantization solution space of weights within low bit-width constraints, which adversely impacts the performance of quantized networks. We highlight the importance of weight oscillation during early training, as it helps the optimizer escape local minima and explore a broader set of quantized configurations. Based on this insight, we propose a simple yet effective method that encourages weight oscillation in the early stage to expand the solution space, followed by oscillation suppression to stabilize convergence. Experiments demonstrate that our algorithm quantizes various models to low-bits and achieves state-of-the-art accuracy on ImageNet.", + "bbox": [ + 511, + 703, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24622", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under Grant 62322117, 62425109, 62371365, U24B20136, and U22B2014.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Haoli Bai, Wei Zhang, Lu Hou, Lifeng Shang, Jing Jin, Xin Jiang, Qun Liu, Michael R. Lyu, and Irwin King. Binarybert: Pushing the limit of bert quantization. In Annual Meeting of the Association for Computational Linguistics, 2020. 1", + "[2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3", + "[3] Hanlin Chen, Li'an Zhuo, Baochang Zhang, Xiawu Zheng, Jianzhuang Liu, Rongrong Ji, David S. Doermann, and Guodong Guo. Binarized neural architecture search for efficient object recognition. International Journal of Computer Vision, 129:501-516, 2020. 1", + "[4] Shangyu Chen, Wenya Wang, and Sinno Jialin Pan. Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization. In Neural Information Processing Systems, 2019. 2", + "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2, 6", + "[6] Michele Conforti, Gérard Cornuéjols, Giacomo Zambelli, Michele Conforti, Gérard Cornuéjols, and Giacomo Zambelli. Integer programming models. Springer, 2014. 3", + "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Neural Information Processing Systems, 2015. 2", + "[8] Alexandre Défossez, Yossi Adi, and Gabriel Synnaeve. Differentiable model compression via pseudo quantization noise. Transactions on Machine Learning Research, 2021. 1, 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 3, 6", + "[10] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. In Neural Information Processing Systems, pages 30318-30332, 2022. 1", + "[11] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Neural Information Processing Systems, 2024. 1", + "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 1, 2, 3, 4, 6, 7, 8", + "[13] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tian-Hao Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differen" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF International Conference on Computer Vision, pages 4851-4860, 2019. 1, 2, 7", + "[14] Tiantian Han, Dong Li, Ji Liu, Lu Tian, and Yi Shan. Improving low-precision network quantization via bin regularization. In IEEE/CVF International Conference on Computer Vision, pages 5241–5250, 2021. 8", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6", + "[16] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In Neural Information Processing Systems, 2024. 1", + "[17] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6", + "[18] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew G. Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2704-2713, 2017. 1, 2", + "[19] San Hun Jung, Changyong Son, Seohyung Lee, JinWoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4345-4354, 2018. 2, 7", + "[20] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[21] Scott Kirkpatrick, Charles D. Gelatt, and Mario P. Vecchi. Optimization by simulated annealing. Science, 220:671-680, 1983. 5", + "[22] Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-50. Springer, 2002. 2", + "[23] Junghyup Lee, Dohyung Kim, and Bumsub Ham. Network quantization with element-wise gradient scaling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6444-6453, 2021. 1, 2", + "[24] Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dongsoo Lee. Flexround: Learnable rounding based on elementwise division for post-training quantization. In International Conference on Machine Learning, 2023. 1", + "[25] Cong Leng, Hao Li, Shenghuo Zhu, and Rong Jin. Extremely low bit neural network: Squeeze the last bit out with admm. In AAAI Conference on Artificial Intelligence, 2017. 2", + "[26] Xinlin Li, Bang Liu, Ruizhi Yang, Vanessa Courville, Chao Xing, and V. Nia. Denseshift : Towards accurate and efficient low-bit power-of-two quantization. In IEEE/CVF International Conference on Computer Vision, pages 16964-16974, 2022. 2", + "[27] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24623", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "neural networks. In International Conference on Learning Representations, 2019. 1, 2, 7", + "[28] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Neural Information Processing Systems, pages 34451-34463, 2022. 6, 8", + "[29] Zhexin Li, Tong Yang, Peisong Wang, and Jian Cheng. Qvit: Fully differentiable quantization for vision transformer. arXiv preprint arXiv:2201.07703, 2022.", + "[30] Shi Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning, 2023. 1, 2, 4, 5, 6, 8", + "[31] Zechun Liu, Baoyuan Wu, Wenhan Luo, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision, pages 722-737, 2018. 1", + "[32] Zechun Liu, Kwang-Ting Cheng, Dong Huang, Eric P. Xing, and Zhiqiang Shen. Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4932-4942, 2021. 1, 2, 6, 7, 8", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 6", + "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 3", + "[35] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 1", + "[36] Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort. Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning, 2022. 1, 2, 4, 5, 7, 8", + "[37] Yury Nahshan, Brian Chmiel, Chaim Baskin, Evgenii Zheltonozhskii, Ron Banner, Alexander M. Bronstein, and Avi Mendelson. Loss aware post-training quantization. Machine Learning, 110:3245 - 3262, 2019. 2, 3", + "[38] Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eunhyeok Park. Nipq: Noise proxy-based integrated pseudo-quantization. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3852–3861, 2022. 2", + "[39] Zhenhong Sun, Ce Ge, Junyan Wang, Ming Lin, Hesen Chen, Hao Li, and Xiuyu Sun. Entropy-driven mixed-precision quantization for deep network design. In Neural Information Processing Systems, 2022. 1", + "[40] Charles E. Taylor. Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence. The Quarterly Review of Biology, 69:88-89, 1994. 5" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 6", + "[42] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2046-2055, 2020. 1", + "[43] Kohei Yamamoto. Learnable companding quantization for accurate low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5027-5036, 2021. 2, 7", + "[44] Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xian-sheng Hua. Quantization networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7316, 2019. 2", + "[45] Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, and Jack Xin. Understanding straight-through estimator in training activation quantized neural nets. In International Conference on Learning Representations, 2019. 1, 2", + "[46] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In European Conference on Computer Vision, 2018. 1, 2, 7", + "[47] Jiayi Zhang, Chang Liu, Junchi Yan, Xijun Li, Hui-Ling Zhen, and Min jie Yuan. A survey for solving mixed integer programming via machine learning. Neurocomputing, 519:205-217, 2022. 3", + "[48] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yuan Du, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. In IEEE/CVF International Conference on Computer Visio, pages 3802-3812, 2023. 1", + "[49] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1, 2, 6" + ], + "bbox": [ + 514, + 92, + 903, + 656 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24624", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_model.json b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1240f658bba27333244ca105cc2aeabcea5fc5dc --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.887, + 0.177 + ], + "angle": 0, + "content": "Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.203, + 0.778, + 0.222 + ], + "angle": 0, + "content": "Weiying Xie\\(^{1}\\) Zihan Meng\\(^{1*}\\) Jitao Ma\\(^{1}\\) Wenjin Guo\\(^{1}\\) Haowei Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.323, + 0.222, + 0.673, + 0.24 + ], + "angle": 0, + "content": "Haonan Qin\\(^{1}\\) Leyuan Fang\\(^{2}\\) Yunsong Li\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.24, + 0.657, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Xidian University \\(^{2}\\)Hunan University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.485, + 0.626 + ], + "angle": 0, + "content": "Quantization-aware Training (QAT) enables deep models to adapt to precision loss by simulating quantization. However, existing methods often converge to sub-optimal solutions due to inadequate exploration of quantization solution space. To address this, we propose a novel QAT method, Allowing Oscillation Quantization (AOQ), which expands the reachable solution space through weight oscillation. Notably, unlike previous methods that suppress oscillation throughout training, AOQ actively encourages it in the earlier stages to explore diverse quantization configurations, and suppresses it later to ensure convergence. In addition, by decoupling quantization thresholds and levels, AOQ promotes meaningful oscillation and improves the stability of learnable quantization parameters. Extensive experiments across various models, including ResNet, MobileNet, DeiT and Swin Transformer, demonstrate the effectiveness of our method. Specifically, with 2-bit quantization, AOQ achieves a \\(0.4\\% \\sim 2.2\\%\\) accuracy improvement on ImageNet compared to state-of-the-art methods. Our implementation is available at https://github.com/muzenc/AOQ." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.222, + 0.672 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Model quantization [31, 46, 49] has attracted growing attention for enabling efficient model inference while substantially reducing hardware resource consumption. By reducing the bit-width of model parameters, quantization effectively reduces memory footprint and computational complexity, thereby enhancing inference speed and energy efficiency. Thus, quantization has been widely applied in tasks such as image classification [27, 32, 39], object detection [3, 42, 48] and natural language processing [1, 10, 11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Despite its advantages, quantization shrinks the dynamic range of parameters, often resulting in information loss and degraded performance. Quantization-aware training (QAT) significantly mitigates the impact of quantization on perfor" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.296, + 0.903, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.448, + 0.665, + 0.46 + ], + "angle": 0, + "content": "- Discrete Solution" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.447, + 0.824, + 0.46 + ], + "angle": 0, + "content": "Solution Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.466, + 0.622, + 0.478 + ], + "angle": 0, + "content": "\\(\\bigcirc\\) Contour" + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.465, + 0.874, + 0.479 + ], + "angle": 0, + "content": "Expanded Solution Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.486, + 0.78, + 0.498 + ], + "angle": 0, + "content": "2D quantization problem:" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.502, + 0.875, + 0.52 + ], + "angle": 0, + "content": "subject to \\(q(\\cdot)\\) is quantization function" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Figure 1. Illustration of solving a 2D quantization problem using previous methods vs. proposed method. The two plots are contour maps of the function \\( f(w_{1},w_{2}) \\), where darker regions indicate lower function value. The yellow and orange curves show the solution trajectories of the previous method and our method, respectively. In contrast, AOQ explores a broader solution space and finds a better solution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.704 + ], + "angle": 0, + "content": "mance compared to post-training quantization [16, 24, 35], but a noticeable performance gap remains under extremely low-bit quantization scenarios, such as 2-bit quantization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Mainstream QAT techniques [12, 32, 46] simulate quantization effects by inserting fake quantization nodes during training, allowing the model to adapt to low-precision constraints. However, these nodes are non-differentiable, leading prior works [13, 23, 45] to employ biased gradient approximations. Recent studies [8, 30, 36] show that such biased gradients can result in weight oscillation, hindering the network's convergence to a better local minimum. Nevertheless, an overlooked insight is that QAT relies on weight oscillation to explore more quantized solutions. In existing methods [12, 13, 18, 32], the frequency of weight oscillation is relatively low, and directly suppressing oscillations may limit the model's ability to escape suboptimal quanti" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24615" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "zation bins. Specifically, weights seldom cross quantization thresholds, instead making insignificant shifts within quantization bins." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.394 + ], + "angle": 0, + "content": "We investigate QAT from the perspective of solving nonlinear mixed-integer programming problems, drawing an analogy to the branch-and-bound method with depth-first search (DFS). While DFS systematically explores all possible solutions through branching and backtracking, QAT lacks an exhaustive exploration mechanism and instead performs local updates guided by gradients. Our analysis reveals that in existing methods, most weights infrequently change their quantization bin assignments, thereby restricting the search to a small subset of possible quantization solutions and resulting in a limited solution space. As shown in Figure 1, this limited search space often leads to premature convergence, leaving potentially optimal configurations unexamined. According to [37], low bit-width quantization introduces sharp curvature in the loss landscape, indirectly demonstrating that gradient-based optimization is more likely to get trapped in poor local minima." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.649 + ], + "angle": 0, + "content": "To address the problems mentioned above, we propose a novel QAT approach, named Allowing Oscillation Quantization (AOQ). Unlike previous methods that suppress weight oscillation throughout the entire optimization process, AOQ strategically leverages weight oscillation as a positive exploration mechanism. We encourage weight oscillation to expand the solution space of the weights and only suppress oscillation in the later stages of training. Moreover, AOQ addresses a key limitation of traditional QAT methods, where the coupling of quantization thresholds and levels often causes inaccurate updates to the quantization parameters and hinders model convergence [30]. To overcome this, AOQ decouples thresholds and levels, allowing thresholds to be manually adjusted to encourage oscillation while setting quantization levels as learnable parameters directly optimized with the task loss. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.652, + 0.483, + 0.726 + ], + "angle": 0, + "content": "- We empirically show that fake quantization-based QAT methods explore a limited subset of quantized solutions and are prone to suboptimal convergence. To the best of our knowledge, this is the first systematic demonstration of this phenomenon." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.728, + 0.483, + 0.787 + ], + "angle": 0, + "content": "- We propose AOQ to expand the quantization solution space of the weights by leveraging weight oscillation. Moreover, we improve the stability of learnable quantization parameters by decoupling the thresholds and levels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.788, + 0.483, + 0.833 + ], + "angle": 0, + "content": "- Extensive experiments in CNN and Vision Transformer architectures demonstrate the effectiveness of AOQ in extremely low-bit quantization." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.652, + 0.483, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.233, + 0.86 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "QAT was initially proposed for binary quantization [7], and has since become the most popular quantization method in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "precision-sensitive scenarios. Most QAT-based methods insert fake quantization nodes into the model. These nodes simulate the effects of quantization during forward propagation and pass the approximated gradients during backward propagation [18, 19, 26, 27, 43, 46]. Some other methods [8, 38] add independent pseudo quantization noise during training, enabling effective QAT without relying on gradient approximations. This work focuses on the former." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.905, + 0.624 + ], + "angle": 0, + "content": "The former is non-differentiable at the fake quantization nodes, preventing optimization and convergence of the quantized model. To address this problem, earlier works approximate gradients using straight-through estimator (STE) [22, 45] and its variants [13, 23, 44], while other approaches compute gradients using constructed neural networks or manual derivation [4, 25]. Among them, STE is widely used in quantization methods due to its robustness, broad applicability, and straightforward implementation. However, it inevitably brings the problem of gradient mismatch: gradients of weights are computed based on their quantized values, instead of their floating-point values. Previous research demonstrated that following the STE approach can cause weight oscillation during training due to biased gradients [8]. It has been suggested that weight oscillation prevents the network from converging to the best local minimum, and that suppressing oscillation earlier is beneficial for the optimization process [36]. The oscillation dampening and iterative weight freezing methods it proposed overcome weight oscillation. In addition, OFQ [30] observed that weight oscillation can introduce noise to the gradient of the learnable scaling factor (quantization level) and cause it to oscillate. The latter, in turn, changes the quantization threshold and aggravates the weight oscillation. OFQ deems weight oscillation harmful and suppresses it; in contrast, AOQ dispels this long-standing misconception, theoretically and experimentally proving its value." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.822 + ], + "angle": 0, + "content": "Moreover, several works have delved into searching for optimal scaling factor to determine quantization thresholds and levels. For instance, trainable clipping parameters have been introduced to determine the appropriate quantization scale [5, 49]. Esser et al. [12] provided a simple yet effective way to approximate the gradient of learnable scaling factor, which is sensitive to quantized state transitions. It has been widely adopted as the de facto approach for QAT. Another line of research decouples quantization thresholds and levels, proposing increasing the quantization representation ability by replacing uniform quantization with nonuniform quantization [32]. The approach has demonstrated excellent performance in quantization of activation values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "While we acknowledge that weight oscillation impacts final convergence, we argue that it is not entirely detrimental to optimization. In layer-by-layer quantization, each weight represents a dimension within the quantization solution space, and weight oscillation plays a crucial role in" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "rapidly exploring optimal quantization solutions within limited training iterations. Moreover, our research shows that in existing methods, the oscillation frequency of weights at quantization thresholds is quite low, motivating us to increase this frequency. To this end, we delve into the working mechanism of QAT, and our proposed AOQ method combines the benefits of enhancing weight oscillation to expand the solution space and suppressing weight oscillation to ensure stable convergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.239, + 0.226, + 0.256 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.263, + 0.232, + 0.278 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.285, + 0.484, + 0.391 + ], + "angle": 0, + "content": "One of the most effective ways to quantize a neural network is by inserting fake quantization nodes to simulate the quantization effect during training. In weight quantization, the goal is to identify the quantization bin each weight falls into and assign it the corresponding quantization level. At present, the most successful methods are based on LSQ [12], and its quantization function is:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.395, + 0.483, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\hat {w} = q (w; s, Q _ {n}, Q _ {p}) = s \\cdot c l i p \\left(\\left\\lfloor \\frac {w}{s} \\right\\rceil , Q _ {n}, Q _ {p}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.472 + ], + "angle": 0, + "content": "where \\(clip(\\cdot, \\alpha, \\beta)\\) is a truncated function with lower bound \\(\\alpha\\) and upper bound \\(\\beta\\), \\(\\lfloor \\cdot \\rceil\\) is a rounding operator, \\(s\\) is a scaling factor, \\(Q_{n}\\) and \\(Q_{p}\\) are the quantization ranges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.533 + ], + "angle": 0, + "content": "Since the quantization function is non-differentiable, the gradient of \\( w \\) cannot be directly computed during backpropagation. Thus, gradient approximation is used to keep the training going, such as straight through estimator (STE) [2]:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.536, + 0.483, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial w} \\approx \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\mathbf {1} _ {Q _ {n} \\leq w / s \\leq Q _ {p}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}\\) is an indicator function. When \\(Q_{n} \\leq w / s \\leq Q_{p}\\), it is 1, otherwise it is 0. There is no gradient for \\(w\\) outside the quantization range. In LSQ, the gradient of the scaling factor \\(s\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.635, + 0.256, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial L}{\\partial s} = \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial s},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.667, + 0.483, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\hat {w}}{\\partial s} = \\left\\{ \\begin{array}{c l} - \\frac {w}{s} + \\left\\lfloor \\frac {w}{s} \\right\\rceil & \\text {i f} Q _ {n} < \\frac {w}{s} < Q _ {p} \\\\ Q _ {n} & \\text {i f} \\frac {w}{s} \\leq Q _ {n} \\\\ Q _ {p} & \\text {i f} \\frac {w}{s} \\geq Q _ {p} \\end{array} . \\right. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.394, + 0.736 + ], + "angle": 0, + "content": "3.2. Limited Solution Space of Weights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Neural network quantization can be formulated as a mixed-integer nonlinear programming (MINLP) problem. It involves converting floating-point weights and activations into a finite set of discrete values while minimizing performance degradation. In this formulation, quantized weights and activations act as integer variables, while the nonlinearity stems from forward propagation and activation functions. Continuous variables, such as quantization parameters, are also involved. The optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.883, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {w}} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\mathcal {L} \\left(f _ {\\text {q u a n t}} (x; q (\\mathbf {w})), y\\right) \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) is the loss function, \\(\\mathbf{w}\\) is the set of all weights, \\((x,y)\\sim D\\) represents samples drawn from the dataset \\(D\\), \\(f_{\\mathrm{quant}}(\\cdot)\\) represents the model's output with activation quantization, and \\(q(\\cdot)\\) is the quantization function. \\(q(\\mathbf{w})\\) consists of integer variables \\(\\mathbf{Q}\\) and continuous parameters \\(s,z\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.177, + 0.907, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\forall \\mathbf {Q} \\in [ 0, 2 ^ {b} - 1 ], \\mathbf {Q} \\in \\mathbb {Z}, b \\in \\mathbb {N} _ {*}, q (\\mathbf {w}) = s \\cdot (\\mathbf {Q} - z), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.904, + 0.236 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}\\) is the set of integers mapped from \\(\\mathbf{w}\\), \\(s\\) is the scaling factor, \\(z\\) is the zero point, and \\(b\\) is the bit-width." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.906, + 0.433 + ], + "angle": 0, + "content": "QAT-based methods serve as heuristics for solving this MINLP problem. These methods select an initial solution and iteratively refine it. Drawing inspiration from research on mixed-integer programming [6, 47], we observe that QAT shares similarities with depth-first search (DFS) in branch-and-bound algorithms. Both rely on branching and backtracking: QAT chooses quantization strategies per weight (branching), and uses gradient updates (implicit backtracking) to adjust solutions. However, unlike exact DFS, which systematically explores the entire solution space, QAT heuristics prioritize fast convergence, often leading to limited exploration and suboptimal results. We compare QAT and DFS from two key perspectives:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.906, + 0.509 + ], + "angle": 0, + "content": "- Branching of variables In DFS, the algorithm explicitly branches on relaxed variables to construct the search tree. Likewise, in QAT, weights oscillating near quantization thresholds effectively induce branching behavior, determining whether a weight falls into one bin or another." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.906, + 0.599 + ], + "angle": 0, + "content": "- Backtracking mechanism DFS performs explicit backtracking: when a branch leads to no improvement, it reverts and explores alternatives, enabling a global search. In contrast, QAT relies on local gradients to implicitly adjust weights. This often causes the model to converge prematurely and miss better configurations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.599, + 0.906, + 0.719 + ], + "angle": 0, + "content": "In summary, both DFS and QAT can be regarded as rule-based search strategies. However, under low-bit constraints, QAT's gradient-based updates are more prone to becoming trapped in local optima due to the steep curvature of the loss surface [37]. Motivated by this limitation, we aim to enhance QAT by expanding its effective solution space, similar in spirit to the global search behavior exhibited by exact DFS, in order to achieve improved performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We assess the ability of existing methods to explore the weight solution space by measuring how often weights cross quantization thresholds. As shown in Table 1, we use LSQ [12] to quantize ResNet-18 on the ImageNet [9]. \"SGD\" denotes an SGD optimizer with an initial learning rate of \\(1\\mathrm{e} - 2\\) and cosine learning rate decay [34], while \"Adam\" denotes an Adam optimizer with a learning rate of \\(1\\mathrm{e} - 3\\), both following the training strategies in [12]. We find that the frequency of weights crossing quantization thresholds is very low, all below \\(0.5\\%\\), with most untruncated weights concentrated in the \\([0,0.01\\%)\\) range. This indicates that previous gradient-based optimization methods have a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.099, + 0.353, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.216, + 0.278, + 0.234, + 0.289 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.099, + 0.627, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.277, + 0.51, + 0.289 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.099, + 0.902, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.277, + 0.784, + 0.289 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.908, + 0.332 + ], + "angle": 0, + "content": "Figure 2. Histograms of pre-trained weights mapped to different quantization values: (a) lossless solution; (b) LSQ; and (c) our AOQ. The red, green, blue, and yellow bars represent histograms of pre-trained weights mapped to the quantized values -2, -1, 0, and 1, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.354, + 0.484, + 0.48 + ], + "angle": 0, + "content": "Table 1. Measuring the solution space of weights on 2-bit quantized ResNet-18. The first column is the frequency of weights crossing the quantization thresholds, calculated as the number of crossings divided by the total iterations. \"Truncated\" refers to weights whose frequency of crossing quantization thresholds falls between 0 and \\(0.01\\%\\) due to being clipped by the truncation function. \"Not Truncated\" refers to not being clipped by the truncation function. \\(\\%\\) of weights\" refers to the percentage of weights within a specific frequency range." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.49, + 0.449, + 0.62 + ], + "angle": 0, + "content": "
Frequency Range% of weights
SGDAdam
[0, 0.01%) - Truncated14.5%11.8%
[0, 0.01%) - Not Truncated47.2%12.6%
[0.01%, 0.05%)5.1%19.1%
[0.05%, 0.1%)8.4%24.5%
[0.1%, 0.5%)24.8%32.0%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.674 + ], + "angle": 0, + "content": "weak ability to explore the solution space, potentially hindering the model from finding an optimal solution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To further illustrate how limited exploration affects quantization quality, we construct a toy example. We first generate an integer set \\(\\bar{w} \\in \\mathbb{R}^{10000}\\) containing only values -2, -1, 0, and 1. We then derive matrix \\(\\mathbf{A} \\in \\mathbb{R}^{2000 \\times 10000}\\) and vector \\(\\mathbf{b} \\in \\mathbb{R}^{2000}\\) satisfying \\(\\mathbf{A}\\bar{w} = \\mathbf{b}\\), establishing \\(\\bar{w}\\) as our known lossless integer solution. Clearly, \\(\\mathbf{A}\\) is not a full-rank matrix, which indicates that the linear equation system has multiple solutions, similar to the multi-solution nature of neural networks. Solving \\(\\mathbf{A}w = \\mathbf{b}\\) with the constraint \\(w \\in \\{-2s + z, -s + z, z, s + z\\}^{10000}\\), where \\(s, z \\in \\mathbb{R}\\), simulates the quantization process. Conceptually, this setup mimics training a fully connected network with 2-bit quantization, 10,000 input units, one output unit, and 2,000 samples, using mean squared error loss and Adam optimizer with a learning rate of \\(1 \\times 10^{-3}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.613 + ], + "angle": 0, + "content": "We first train the model with floating-point weights until convergence, obtaining pre-trained weights, similar to the standard QAT procedure. These weights are then mapped to quantized values. In the lossless solution, the scaling factor \\( s \\) is 1 and \\( z \\) is 0. Figure 2a shows the distribution of pre-trained weights mapped to each quantized value under the lossless solution. Each quantization level corresponds to a wide range of pre-trained weights, indicating that allowing distant weights to map to the same quantized value can preserve model quality. In contrast, Figure 2b shows the distribution after freezing \\( s \\) to 1 and \\( z \\) to 0 and using LSQ [12] for quantization. In this setting, pre-trained weights are mapped only to the nearest quantized values, restricting the solution space. Such methods, by mapping only to nearby values, may result in more mapping errors and does not fully leverage the broader solution space to minimize quantization errors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.813, + 0.639 + ], + "angle": 0, + "content": "3.3. Allowing Oscillation Quantization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In previous methods, many weights seldom oscillate near quantization thresholds, limiting their ability to switch quantization bins and often resulting in suboptimal solutions. We speculate that in QAT, especially in low-bit settings, weight transitions between different quantization levels make gradient direction unstable. This instability causes weights to fluctuate within a narrow range during training, preventing them from crossing thresholds. While some studies [30, 36] consider weight oscillation harmful to the whole optimization process, we believe that oscillation actually promotes solution diversity, increasing the likelihood of discovering better optima. Specifically, oscillation acts as an adaptive perturbation mechanism that enables the model to continuously test new quantization configurations and explore more quantized solutions within a given number of iterations, helping avoid premature convergence. This phenomenon resembles \"random perturbation\" strategies used" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.48, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.295 + ], + "angle": 0, + "content": "Figure 3. Oscillation patterns of two weights in layer1.0.conv1 of ResNet-18 during early training. We apply AOQ under a 2-bit quantization setting and monitor quantization assignments at the end of each epoch." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.483, + 0.413 + ], + "angle": 0, + "content": "in integer programming [21, 40], where moderate instability prevents the model from becoming locked into overly deterministic paths. Based on this intuition, we propose to actively promote oscillation during early training phases to encourage exploration, then gradually reduce it in later stages to ensure convergence stability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.699 + ], + "angle": 0, + "content": "To increase the number of weights that cross thresholds without disrupting QAT's gradient-based backtracking mechanism, we examine how quantization parameters affect oscillation. Inspired by [30], we observe that modifying quantization thresholds can induce more weights to oscillate around them, as shown in Figure 3. Specifically, when the intervals between quantization thresholds and the differences among quantization levels decrease, weights tend to transition from intra-bin oscillation to threshold-adjacent oscillation. This behavior is driven by two factors. First, the shifting thresholds force certain weights to cross them, introducing noise into the network. The facilitating effect of smaller quantization threshold intervals on weight oscillation will be demonstrated in Section 3.4. Second, the reduced differences between quantization levels lower the penalty for mapping errors when floating-point weights are mapped to quantized values, making weights more likely to oscillate near the thresholds, as smaller differences between quantization levels tend to smooth the loss surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.882 + ], + "angle": 0, + "content": "To leverage the weight oscillation mentioned above, we decouple the quantization thresholds and levels and introduce a three-stage training strategy. Specifically, we manually adjust the quantization thresholds to gradually reduce their intervals, while setting the quantization levels as learnable parameters to optimize during training. Additionally, we introduce differentiability for clipped weights in fake quantization to address the issue observed in Section 3.2, where the absence of gradients in clipped weights (as shown in Eq. (2)) hinders model fine-tuning. In the final stage, oscillation is suppressed to stabilize convergence. The complete strategy is outlined below:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Stage 1: Inducing oscillation. For weight quantization," + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "we define the intervals between thresholds as \\( s_{th} \\) and between quantization levels as \\( s_{le} \\), both initialized to \\( \\frac{\\sigma_w}{2^{b - 2}} \\) where \\( \\sigma_w \\) is the standard deviation of weights in a layer, and \\( b \\) is the bit-width. In this stage, we manually reduce \\( s_{th} \\) and \\( s_{le} \\) to promote oscillation at thresholds." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.907, + 0.319 + ], + "angle": 0, + "content": "- Stage 2: Learning quantization levels. We fix \\( s_{th} \\) and set \\( s_{le} \\) as a learnable parameter for optimization. In Stage 1, \\( s_{le} \\) is manually reduced, which leads to higher loss due to poor representation of weights by the quantization levels. As training progresses and the task loss decreases, \\( s_{le} \\) is learned to increase automatically. This widens the intervals between quantization levels, makes the loss landscape steeper near thresholds, and reduces oscillation by discouraging weights from frequently crossing thresholds. The quantizer is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\hat {w} = \\left\\{ \\begin{array}{l l} \\operatorname {l e v e l} [ 0 ] & \\text {i f} w < \\operatorname {t h r e s h} [ 0 ] \\\\ \\operatorname {l e v e l} [ i ] & \\text {i f t h r e s h} [ i - 1 ] \\leq w < \\operatorname {t h r e s h} [ i ] \\\\ \\operatorname {l e v e l} \\left[ 2 ^ {b} - 1 \\right] & \\text {i f} w \\geq \\operatorname {t h r e s h} \\left[ 2 ^ {b} - 2 \\right] \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.398, + 0.905, + 0.445 + ], + "angle": 0, + "content": "where \\(i\\in \\{1,2,\\ldots ,2^b -2\\}\\) and \\(\\mathrm{thresh}[i]\\) and level[i] represent the \\((\\mathrm{i + 1})\\) -th threshold and level, respectively. The optimized quantized weights are denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.457, + 0.905, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\tilde {w} = \\hat {w} + w - w. \\det a c h (\\cdot). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.484, + 0.906, + 0.574 + ], + "angle": 0, + "content": "This formulation, based on the STE, enables distinct gradient flows: \\(\\hat{w}\\) determines the gradient for \\(s_{le}\\), while \\(w\\) receives gradients directly from the task loss, even when clipped. This decoupling avoids the zero-gradient issue and enables more effective optimization of both weights and quantizer parameters." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.575, + 0.906, + 0.62 + ], + "angle": 0, + "content": "- Stage 3: Dampening oscillation. In the third stage, we employ the Oscillation dampening method outlined in [36], which adds a regularization term to the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.633, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda \\mathcal {L} _ {\\text {d a m p e n}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.661, + 0.878, + 0.675 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter. The dampening loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.686, + 0.905, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d a m p e n}} = \\left\\| \\hat {w} - \\operatorname {c l i p} \\left(w, Q _ {n}, Q _ {p}\\right) \\right\\| _ {F} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.716, + 0.905, + 0.761 + ], + "angle": 0, + "content": "which penalizes weights lingering near quantization thresholds, encouraging them to converge to the nearest quantization centers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.764, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Details regarding the hyperparameters and transition schedule for each training stage are provided in Section 4.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.708, + 0.82 + ], + "angle": 0, + "content": "3.4. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.87 + ], + "angle": 0, + "content": "In this section, we theoretically analyze the effectiveness of our approach and compare it against prior differentiable quantization parameter methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We first show the advantage of using smaller quantization threshold intervals. Let \\( W \\) be the set of pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "weights, and \\( Q \\) be a set composed of quantized values. Weight quantization is a mapping from the \\( W \\) to the \\( Q \\). For a given weight \\( w \\in W \\), we define its probability of being quantized to \\( Q_{i} \\in Q \\) as the transition probability \\( P(Q_{i} \\mid w) \\), where \\( i \\in \\{0, 1, \\dots, 2^{b} - 1\\} \\) and \\( b \\) is the bit-width. By Bayes' theorem:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.193, + 0.483, + 0.227 + ], + "angle": 0, + "content": "\\[\nP \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{P (w)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.484, + 0.314 + ], + "angle": 0, + "content": "where \\( P(Q_{i}) \\) is the proportion of the number of weights mapped to \\( Q_{i} \\) to the total number of weights, \\( P(w) \\) is the probability distribution of weights, and \\( P(w|Q_i) \\) is the conditional distribution of weights mapped to \\( Q_{i} \\). The total probability formula applied to Eq. (10) is:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.324, + 0.483, + 0.358 + ], + "angle": 0, + "content": "\\[\nP \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{\\sum_ {i} P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.414 + ], + "angle": 0, + "content": "As shown in Figure 2a, \\( P(w \\mid Q_i) \\) approximately follows a Gaussian distribution, so the total probability formula can be transformed into a Gaussian mixture model:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.424, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\nP (w) = \\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i}, \\sigma_ {i}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.528 + ], + "angle": 0, + "content": "where \\(\\pi_i, \\mu_i, \\sigma_i\\) denote the weight, mean and variance of the Gaussian component, respectively. This transformation does not effect the theoretical conclusion, just for analytical convenience. Eq. (11) is transformed into:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.538, + 0.483, + 0.572 + ], + "angle": 0, + "content": "\\[\nP \\left(Q _ {i} \\mid w\\right) = \\frac {\\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}{\\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.613 + ], + "angle": 0, + "content": "The weight \\( w \\) is assigned to the quantization value \\( Q_{j} \\) that maximizes this probability:" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.624, + 0.483, + 0.647 + ], + "angle": 0, + "content": "\\[\ni f \\forall i, \\pi_ {j} \\mathcal {N} (w \\mid \\mu_ {j}, \\sigma_ {j}) \\geq \\pi_ {i} \\mathcal {N} (w \\mid \\mu_ {i}, \\sigma_ {i}) \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.644, + 0.232, + 0.66 + ], + "angle": 0, + "content": "then \\(w\\to Q_j\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.792 + ], + "angle": 0, + "content": "Therefore, when the quantization threshold is set at the horizontal coordinate value of the intersection between two adjacent Gaussian components, it ensures optimal probabilistic mapping of weights. In Figure 2a, we observe that in the lossless solution for this example, the variances of each Gaussian component are approximately equal, denoted as \\(\\sigma\\). Thus, the horizontal coordinate value at the intersection of adjacent Gaussian components is:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.803, + 0.483, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\operatorname {t h r e s h} [ i ] = \\frac {\\mu_ {i + 1} + \\mu_ {i}}{2} - \\frac {\\log \\left(\\frac {\\pi_ {i + 1}}{\\pi_ {i}}\\right) \\cdot \\sigma^ {2}}{\\mu_ {i + 1} - \\mu_ {i}}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This analysis suggests that, when Gaussian component means are close, using smaller threshold intervals improves mapping accuracy. It explains why our method encourages" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "tighter thresholds in early training to support more effective weight assignment in QAT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.332 + ], + "angle": 0, + "content": "Next, we show that AOQ provides a more direct and accurate link between quantization levels and task loss than previous approaches. In prior methods [5, 12, 49], learnable quantization parameters jointly determine both thresholds and levels. This coupling introduces approximation errors in the gradient computation of quantization levels. For instance, in LSQ [12], the gradients of the scaling factors are shown in Eq. (3). The derivative of weights with respect to the scaling factor aids in optimizing quantization thresholds but makes the gradient of the quantization levels inaccurate. In contrast, AOQ decouples thresholds from levels. It directly optimizes quantization levels based on Eq. (6) and Eq. (7), allowing a more faithful gradient flow that better aligns with task loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.47 + ], + "angle": 0, + "content": "Finally, AOQ effectively addresses the impact of weight oscillation on final convergence, as well as on batch normalization and layer normalization. After exploring a broader solution space, we use Oscillation Dampening in the final stage to gradually guide weights out of oscillatory regions. As training continues, the exponential moving average used in batch normalization and layer normalization becomes increasingly stable, reinforcing consistent model behavior and improving convergence quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.646, + 0.5 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.6 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the proposed Allowing Oscillation Quantization (AOQ) on the ILSVRC12 ImageNet dataset [9], using a range of models including ResNet [15], MobileNet-V2 [17], DeiT [41], and Swin Transformer [33]. Our experiments focus on low-bit weight quantization. For activation quantization, we adopt LSQ [12] and N2UQ [32]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.704, + 0.623 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.75 + ], + "angle": 0, + "content": "ResNet and MobileNet We follow the training schemes in [5, 12, 32], initializing quantized networks from full-precision pre-trained models. Training is performed using the Adam optimizer [20] with a linear learning rate decay, starting from \\(1 \\times 10^{-3}\\). Following LSQ [12], we apply knowledge distillation and train for 150 epochs. All convolutional and fully connected layers are quantized, except for the first and last layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.826 + ], + "angle": 0, + "content": "DeiT and Swin Transformer We use the official implementations and adopt training settings from [28-30], including knowledge distillation and a training duration of 300 epochs. As in prior work, the patch embedding and classification layers are quantized to 8 bits [28-30]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Parameter Setup The batch size is set to 512, and the weight decay is set to 0. The three training stages occupy \\(1/5\\), \\(2/5\\), and \\(2/5\\) of the total training epochs, respectively. We reduce \\(s_{\\mathrm{th}}\\) and \\(s_{\\mathrm{le}}\\) to \\(30\\% - 35\\%\\) of their initial values in the first stage. As shown in Fig. 4, a broad grid search" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24620" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Accuracy comparison to the state-of-the-art quantization methods with ResNet structure on ImageNet dataset. Note that W/A denote the bit-widths of weights and activations, respectively. FP denotes the top-1 accuracy of the full-precision models." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.128, + 0.857, + 0.511 + ], + "angle": 0, + "content": "
NetworkMethodBit-width (W/A)Accuracy (%) Top-1Bit-width (W/A)Accuracy(%) Top-1Bit-width (W/A)Accuracy(%) Top-1
ResNet-18 (FP: 71.8)LQ-Nets [46]2/264.93/368.24/469.3
DSQ [13]2/265.23/368.74/469.6
QIL [19]2/265.73/369.24/470.1
APoT [27]2/267.33/369.94/470.7
LSQ [12]2/267.63/370.24/471.1
LCQ [43]2/268.93/370.64/471.5
N2UQ [32]2/269.43/371.94/472.9
LSQ+AOQ2/269.73/371.44/472.0
N2UQ+AOQ2/270.33/372.14/472.7
ResNet-34 (FP:74.9)LQ-Nets [46]2/269.83/371.94/4-
DSQ [13]2/270.03/372.54/472.8
QIL [19]2/270.63/373.14/473.7
APoT [27]2/270.93/373.44/473.8
LSQ [12]2/271.63/373.44/474.1
LCQ [43]2/272.73/374.04/474.3
N2UQ [32]2/273.33/375.24/476.0
LSQ+AOQ2/273.53/374.94/475.5
N2UQ+AOQ2/273.93/375.54/476.0
ResNet-50 (FP: 77.0)LQ-Nets [46]2/271.53/374.24/475.1
APoT [27]2/273.43/375.84/476.6
LSQ [12]2/273.73/375.84/476.7
LCQ [43]2/275.13/376.34/476.6
N2UQ [32]2/275.83/377.54/478.0
LSQ+AOQ2/276.33/377.54/477.9
N2UQ+AOQ2/276.43/377.84/478.1
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.542, + 0.283, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.653, + 0.228, + 0.664 + ], + "angle": 0, + "content": "(a) ResNet-18" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.542, + 0.482, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.653, + 0.416, + 0.664 + ], + "angle": 0, + "content": "(b) DeiT-T" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.677, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Figure 4. Accuracy heatmaps on ImageNet under different narrowing schedules. \\(\\alpha\\) denotes the final-to-initial ratio of \\(s_{le}\\) and \\(s_{th}\\) in stage one, and \\(\\beta\\) is the proportion of total epochs allocated to this stage. AOQ maintains stable performance across a wide range of \\((\\alpha, \\beta)\\) settings on both (a) 2-bit ResNet-18 and (b) 2-bit DeiT-T, indicating schedule robustness across architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.842 + ], + "angle": 0, + "content": "confirms that a common schedule performs robustly across different architectures in practice. For the final stage, the hyper-parameter \\(\\lambda\\) is chosen such that \\(\\lambda \\mathcal{L}_{\\mathrm{dampen}}\\) is two to three orders of magnitude lower than \\(\\mathcal{L}_{\\mathrm{task}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.849, + 0.423, + 0.866 + ], + "angle": 0, + "content": "4.2. Experimental Results on Toy Example" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We first revisit the toy example from Section 3.2 to evaluate the effect of oscillation on quantization quality. To" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.547, + 0.713, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.72, + 0.554, + 0.895, + 0.641 + ], + "angle": 0, + "content": "
MethodLoss
LSQ51.2
LSQ+OD(final)16.7
LSQ+OD(enti)26.2
AOQ5.6
" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.663, + 0.907, + 0.691 + ], + "angle": 0, + "content": "Figure 5. Loss curves and minimum loss obtained from solving the toy example using various methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "demonstrate that weight oscillation can expand the solution space and improve quantization outcomes, we apply the Oscillation Dampening (OD) [36] to suppress oscillation either during the final training stage or throughout the entire training process. We also evaluate the performance of AOQ using the toy example. As shown in Figure 5, allowing early-stage oscillation improves QAT performance, and AOQ converges faster to a better solution. When using fixed quantization levels \\(\\{-2, -1, 0, 1\\}\\) in the third stage, Figure 2c shows that AOQ enables pre-trained weights to be mapped to more diverse and distant quantized values, expanding the solution space compared to Figure 2b." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24621" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.12 + ], + "angle": 0, + "content": "Table 3. Top-1 accuracy (\\%) comparison for methods suppressing weight oscillation throughout the process in MobileNet-V2." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.129, + 0.47, + 0.25 + ], + "angle": 0, + "content": "
NetworkMethodW3A3W4A4
MobileNet-V2 (FP: 72.1)LSQ [12]65.369.5
LSQ+BR [14]67.470.4
LSQ+Dampen [36]67.870.5
N2UQ [32]69.072.1
LSQ+AOQ69.171.8
N2UQ+AOQ69.972.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.485, + 0.294 + ], + "angle": 0, + "content": "Table 4. Top-1 accuracy (\\%) comparison with the state-of-the-art methods in the ViT architecture." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.304, + 0.473, + 0.533 + ], + "angle": 0, + "content": "
NetworkMethodW2A2W3A3
DeiT-T (FP: 72.1)LSQ [12]54.568.1
QViT* [28]50.467.1
OFQ [30]64.372.7
LSQ+AOQ66.572.3
DeiT-S (FP: 79.9)LSQ [12]68.077.8
QViT* [28]68.778.5
OFQ [30]75.779.6
LSQ+AOQ76.779.8
Swin-T (FP: 81.2)LSQ [12]70.479.0
QViT* [28]73.980.1
OFQ [30]78.581.1
LSQ+AOQ78.981.2
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.56, + 0.291, + 0.575 + ], + "angle": 0, + "content": "4.3. Accuracy Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Table 2 reports the accuracy of AOQ on ResNet. When paired with different activation quantization techniques, AOQ consistently improves performance. For instance, using AOQ with LSQ yields \\(0.9\\% - 2.6\\%\\) accuracy gains. When combined with N2UQ, AOQ outperforms many traditional methods. Its advantage becomes more pronounced in low-bit networks with a large number of weights, where the expanded solution space is more effectively utilized. In the 2-bit ResNet-50 case, AOQ reduces the accuracy gap with full-precision to just \\(0.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Table 3 shows the performance of the AOQ method on MobileNet-V2. Compared to other methods, AOQ has a natural advantage in quantizing MobileNet. Due to significant variations in weight distributions across channels in the depth-wise separable convolution layers, traditional quantization methods often map all weights in certain channels to a single quantized value, resulting in the loss of important information. AOQ mitigates this by adaptively shrinking threshold intervals during training, promoting threshold-crossing behavior and improving quantization granularity. This effect is illustrated in Figure 3. Overall, AOQ outper" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Table 5. Effects of different components of our strategy on the final performance." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.129, + 0.88, + 0.25 + ], + "angle": 0, + "content": "
MethodAcc (%)
Baseline [12]73.7
- Weight quantization76.6
+ Expanding solution space of weights75.2
+ Delayed weight oscillation suppression75.6
+ Both76.3
Corresponding real-valued network77.0
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.304 + ], + "angle": 0, + "content": "forms other QAT approaches, including those that suppress oscillation throughout [14, 36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.485 + ], + "angle": 0, + "content": "To validate AOQ on Vision Transformers (ViT), we conducted 2-bit and 3-bit quantization experiments on DeiT and Swin Transformer. We observe that QViT [28] does not fully quantize activations in the fully connected and attention layers, limiting hardware acceleration. For fair comparison, we modify their implementation accordingly and report the results as QViT* in Table 4. The results show that AOQ significantly outperforms LSQ and QViT by suppressing weight oscillation, boosting 2-bit DeiT-T accuracy by \\(12.0\\%\\) and \\(16.1\\%\\), respectively. By expanding the solution space, AOQ also surpasses OFQ, the prior state-of-the-art, under 2-bit quantization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.493, + 0.667, + 0.51 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.909, + 0.668 + ], + "angle": 0, + "content": "To investigate the impact of each improvement on the final performance, we conduct an ablation study on the 2-bit quantized ResNet-50 network on ImageNet. As shown in Table 5, the proposed expanded solution space and delayed weight oscillation suppression improve the accuracy by \\(1.5\\%\\) and \\(1.9\\%\\), respectively over the baseline [12]. By using both techniques, we reduced the accuracy gap with the corresponding real-valued ResNet-50 to only \\(0.7\\%\\), and the accuracy is only reduced by \\(0.3\\%\\) compared to the baseline method without weight quantization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we reveal the weakness of existing QAT methods in exploring the quantization solution space of weights within low bit-width constraints, which adversely impacts the performance of quantized networks. We highlight the importance of weight oscillation during early training, as it helps the optimizer escape local minima and explore a broader set of quantized configurations. Based on this insight, we propose a simple yet effective method that encourages weight oscillation in the early stage to expand the solution space, followed by oscillation suppression to stabilize convergence. Experiments demonstrate that our algorithm quantizes various models to low-bits and achieves state-of-the-art accuracy on ImageNet." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24622" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under Grant 62322117, 62425109, 62371365, U24B20136, and U22B2014." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Haoli Bai, Wei Zhang, Lu Hou, Lifeng Shang, Jing Jin, Xin Jiang, Qun Liu, Michael R. Lyu, and Irwin King. Binarybert: Pushing the limit of bert quantization. In Annual Meeting of the Association for Computational Linguistics, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[3] Hanlin Chen, Li'an Zhuo, Baochang Zhang, Xiawu Zheng, Jianzhuang Liu, Rongrong Ji, David S. Doermann, and Guodong Guo. Binarized neural architecture search for efficient object recognition. International Journal of Computer Vision, 129:501-516, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[4] Shangyu Chen, Wenya Wang, and Sinno Jialin Pan. Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization. In Neural Information Processing Systems, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[6] Michele Conforti, Gérard Cornuéjols, Giacomo Zambelli, Michele Conforti, Gérard Cornuéjols, and Giacomo Zambelli. Integer programming models. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Neural Information Processing Systems, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.66 + ], + "angle": 0, + "content": "[8] Alexandre Défossez, Yossi Adi, and Gabriel Synnaeve. Differentiable model compression via pseudo quantization noise. Transactions on Machine Learning Research, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. In Neural Information Processing Systems, pages 30318-30332, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[11] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tian-Hao Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differen" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.135 + ], + "angle": 0, + "content": "tiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF International Conference on Computer Vision, pages 4851-4860, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[14] Tiantian Han, Dong Li, Ji Liu, Lu Tian, and Yi Shan. Improving low-precision network quantization via bin regularization. In IEEE/CVF International Conference on Computer Vision, pages 5241–5250, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[16] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In Neural Information Processing Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.371 + ], + "angle": 0, + "content": "[17] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[18] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew G. Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2704-2713, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[19] San Hun Jung, Changyong Son, Seohyung Lee, JinWoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4345-4354, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.565 + ], + "angle": 0, + "content": "[20] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[21] Scott Kirkpatrick, Charles D. Gelatt, and Mario P. Vecchi. Optimization by simulated annealing. Science, 220:671-680, 1983. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[22] Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-50. Springer, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[23] Junghyup Lee, Dohyung Kim, and Bumsub Ham. Network quantization with element-wise gradient scaling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6444-6453, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[24] Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dongsoo Lee. Flexround: Learnable rounding based on elementwise division for post-training quantization. In International Conference on Machine Learning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[25] Cong Leng, Hao Li, Shenghuo Zhu, and Rong Jin. Extremely low bit neural network: Squeeze the last bit out with admm. In AAAI Conference on Artificial Intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[26] Xinlin Li, Bang Liu, Ruizhi Yang, Vanessa Courville, Chao Xing, and V. Nia. Denseshift : Towards accurate and efficient low-bit power-of-two quantization. In IEEE/CVF International Conference on Computer Vision, pages 16964-16974, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24623" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "neural networks. In International Conference on Learning Representations, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[28] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Neural Information Processing Systems, pages 34451-34463, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[29] Zhexin Li, Tong Yang, Peisong Wang, and Jian Cheng. Qvit: Fully differentiable quantization for vision transformer. arXiv preprint arXiv:2201.07703, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[30] Shi Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning, 2023. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[31] Zechun Liu, Baoyuan Wu, Wenhan Luo, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision, pages 722-737, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[32] Zechun Liu, Kwang-Ting Cheng, Dong Huang, Eric P. Xing, and Zhiqiang Shen. Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4932-4942, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[35] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[36] Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort. Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning, 2022. 1, 2, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.719 + ], + "angle": 0, + "content": "[37] Yury Nahshan, Brian Chmiel, Chaim Baskin, Evgenii Zheltonozhskii, Ron Banner, Alexander M. Bronstein, and Avi Mendelson. Loss aware post-training quantization. Machine Learning, 110:3245 - 3262, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[38] Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eunhyeok Park. Nipq: Noise proxy-based integrated pseudo-quantization. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3852–3861, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[39] Zhenhong Sun, Ce Ge, Junyan Wang, Ming Lin, Hesen Chen, Hao Li, and Xiuyu Sun. Entropy-driven mixed-precision quantization for deep network design. In Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Charles E. Taylor. Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence. The Quarterly Review of Biology, 69:88-89, 1994. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[42] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2046-2055, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[43] Kohei Yamamoto. Learnable companding quantization for accurate low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5027-5036, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[44] Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xian-sheng Hua. Quantization networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7316, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[45] Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, and Jack Xin. Understanding straight-through estimator in training activation quantized neural nets. In International Conference on Learning Representations, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[46] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In European Conference on Computer Vision, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[47] Jiayi Zhang, Chang Liu, Junchi Yan, Xijun Li, Hui-Ling Zhen, and Min jie Yuan. A survey for solving mixed integer programming via machine learning. Neurocomputing, 519:205-217, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.599 + ], + "angle": 0, + "content": "[48] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yuan Du, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. In IEEE/CVF International Conference on Computer Visio, pages 3802-3812, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[49] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24624" + } + ] +] \ No newline at end of file diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_origin.pdf b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b560eff39cc895551f809790058349eac3af192b --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/5d6dba58-3126-413c-a7af-b25ffe7d92f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e07e6e1ce999d7f9704238e312af74cb493ca65c2c6b17d888f9c7943b0d517 +size 845000 diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/full.md b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..91de3d7935ce99d5ae3aac4f517cd062fe6987a5 --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/full.md @@ -0,0 +1,354 @@ +# Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization + +Weiying Xie $^{1}$ Zihan Meng $^{1*}$ Jitao Ma $^{1}$ Wenjin Guo $^{1}$ Haowei Li $^{1}$ + +Haonan Qin $^{1}$ Leyuan Fang $^{2}$ Yunsong Li $^{1}$ + +$^{1}$ Xidian University $^{2}$ Hunan University + +# Abstract + +Quantization-aware Training (QAT) enables deep models to adapt to precision loss by simulating quantization. However, existing methods often converge to sub-optimal solutions due to inadequate exploration of quantization solution space. To address this, we propose a novel QAT method, Allowing Oscillation Quantization (AOQ), which expands the reachable solution space through weight oscillation. Notably, unlike previous methods that suppress oscillation throughout training, AOQ actively encourages it in the earlier stages to explore diverse quantization configurations, and suppresses it later to ensure convergence. In addition, by decoupling quantization thresholds and levels, AOQ promotes meaningful oscillation and improves the stability of learnable quantization parameters. Extensive experiments across various models, including ResNet, MobileNet, DeiT and Swin Transformer, demonstrate the effectiveness of our method. Specifically, with 2-bit quantization, AOQ achieves a $0.4\% \sim 2.2\%$ accuracy improvement on ImageNet compared to state-of-the-art methods. Our implementation is available at https://github.com/muzenc/AOQ. + +# 1. Introduction + +Model quantization [31, 46, 49] has attracted growing attention for enabling efficient model inference while substantially reducing hardware resource consumption. By reducing the bit-width of model parameters, quantization effectively reduces memory footprint and computational complexity, thereby enhancing inference speed and energy efficiency. Thus, quantization has been widely applied in tasks such as image classification [27, 32, 39], object detection [3, 42, 48] and natural language processing [1, 10, 11]. + +Despite its advantages, quantization shrinks the dynamic range of parameters, often resulting in information loss and degraded performance. Quantization-aware training (QAT) significantly mitigates the impact of quantization on perfor + +![](images/30f7d84627e4b80f2d787dee8a1fbdeafc8a22528e7492819284c059066e6636.jpg) +- Discrete Solution +Solution Space +$\bigcirc$ Contour +Expanded Solution Space +2D quantization problem: +subject to $q(\cdot)$ is quantization function +Figure 1. Illustration of solving a 2D quantization problem using previous methods vs. proposed method. The two plots are contour maps of the function $f(w_{1},w_{2})$ , where darker regions indicate lower function value. The yellow and orange curves show the solution trajectories of the previous method and our method, respectively. In contrast, AOQ explores a broader solution space and finds a better solution. + +mance compared to post-training quantization [16, 24, 35], but a noticeable performance gap remains under extremely low-bit quantization scenarios, such as 2-bit quantization. + +Mainstream QAT techniques [12, 32, 46] simulate quantization effects by inserting fake quantization nodes during training, allowing the model to adapt to low-precision constraints. However, these nodes are non-differentiable, leading prior works [13, 23, 45] to employ biased gradient approximations. Recent studies [8, 30, 36] show that such biased gradients can result in weight oscillation, hindering the network's convergence to a better local minimum. Nevertheless, an overlooked insight is that QAT relies on weight oscillation to explore more quantized solutions. In existing methods [12, 13, 18, 32], the frequency of weight oscillation is relatively low, and directly suppressing oscillations may limit the model's ability to escape suboptimal quanti + +zation bins. Specifically, weights seldom cross quantization thresholds, instead making insignificant shifts within quantization bins. + +We investigate QAT from the perspective of solving nonlinear mixed-integer programming problems, drawing an analogy to the branch-and-bound method with depth-first search (DFS). While DFS systematically explores all possible solutions through branching and backtracking, QAT lacks an exhaustive exploration mechanism and instead performs local updates guided by gradients. Our analysis reveals that in existing methods, most weights infrequently change their quantization bin assignments, thereby restricting the search to a small subset of possible quantization solutions and resulting in a limited solution space. As shown in Figure 1, this limited search space often leads to premature convergence, leaving potentially optimal configurations unexamined. According to [37], low bit-width quantization introduces sharp curvature in the loss landscape, indirectly demonstrating that gradient-based optimization is more likely to get trapped in poor local minima. + +To address the problems mentioned above, we propose a novel QAT approach, named Allowing Oscillation Quantization (AOQ). Unlike previous methods that suppress weight oscillation throughout the entire optimization process, AOQ strategically leverages weight oscillation as a positive exploration mechanism. We encourage weight oscillation to expand the solution space of the weights and only suppress oscillation in the later stages of training. Moreover, AOQ addresses a key limitation of traditional QAT methods, where the coupling of quantization thresholds and levels often causes inaccurate updates to the quantization parameters and hinders model convergence [30]. To overcome this, AOQ decouples thresholds and levels, allowing thresholds to be manually adjusted to encourage oscillation while setting quantization levels as learnable parameters directly optimized with the task loss. Our contributions can be summarized as follows: + +- We empirically show that fake quantization-based QAT methods explore a limited subset of quantized solutions and are prone to suboptimal convergence. To the best of our knowledge, this is the first systematic demonstration of this phenomenon. +- We propose AOQ to expand the quantization solution space of the weights by leveraging weight oscillation. Moreover, we improve the stability of learnable quantization parameters by decoupling the thresholds and levels. +- Extensive experiments in CNN and Vision Transformer architectures demonstrate the effectiveness of AOQ in extremely low-bit quantization. + +# 2. Related Work + +QAT was initially proposed for binary quantization [7], and has since become the most popular quantization method in + +precision-sensitive scenarios. Most QAT-based methods insert fake quantization nodes into the model. These nodes simulate the effects of quantization during forward propagation and pass the approximated gradients during backward propagation [18, 19, 26, 27, 43, 46]. Some other methods [8, 38] add independent pseudo quantization noise during training, enabling effective QAT without relying on gradient approximations. This work focuses on the former. + +The former is non-differentiable at the fake quantization nodes, preventing optimization and convergence of the quantized model. To address this problem, earlier works approximate gradients using straight-through estimator (STE) [22, 45] and its variants [13, 23, 44], while other approaches compute gradients using constructed neural networks or manual derivation [4, 25]. Among them, STE is widely used in quantization methods due to its robustness, broad applicability, and straightforward implementation. However, it inevitably brings the problem of gradient mismatch: gradients of weights are computed based on their quantized values, instead of their floating-point values. Previous research demonstrated that following the STE approach can cause weight oscillation during training due to biased gradients [8]. It has been suggested that weight oscillation prevents the network from converging to the best local minimum, and that suppressing oscillation earlier is beneficial for the optimization process [36]. The oscillation dampening and iterative weight freezing methods it proposed overcome weight oscillation. In addition, OFQ [30] observed that weight oscillation can introduce noise to the gradient of the learnable scaling factor (quantization level) and cause it to oscillate. The latter, in turn, changes the quantization threshold and aggravates the weight oscillation. OFQ deems weight oscillation harmful and suppresses it; in contrast, AOQ dispels this long-standing misconception, theoretically and experimentally proving its value. + +Moreover, several works have delved into searching for optimal scaling factor to determine quantization thresholds and levels. For instance, trainable clipping parameters have been introduced to determine the appropriate quantization scale [5, 49]. Esser et al. [12] provided a simple yet effective way to approximate the gradient of learnable scaling factor, which is sensitive to quantized state transitions. It has been widely adopted as the de facto approach for QAT. Another line of research decouples quantization thresholds and levels, proposing increasing the quantization representation ability by replacing uniform quantization with nonuniform quantization [32]. The approach has demonstrated excellent performance in quantization of activation values. + +While we acknowledge that weight oscillation impacts final convergence, we argue that it is not entirely detrimental to optimization. In layer-by-layer quantization, each weight represents a dimension within the quantization solution space, and weight oscillation plays a crucial role in + +rapidly exploring optimal quantization solutions within limited training iterations. Moreover, our research shows that in existing methods, the oscillation frequency of weights at quantization thresholds is quite low, motivating us to increase this frequency. To this end, we delve into the working mechanism of QAT, and our proposed AOQ method combines the benefits of enhancing weight oscillation to expand the solution space and suppressing weight oscillation to ensure stable convergence. + +# 3. Methodology + +# 3.1. Preliminaries + +One of the most effective ways to quantize a neural network is by inserting fake quantization nodes to simulate the quantization effect during training. In weight quantization, the goal is to identify the quantization bin each weight falls into and assign it the corresponding quantization level. At present, the most successful methods are based on LSQ [12], and its quantization function is: + +$$ +\hat {w} = q (w; s, Q _ {n}, Q _ {p}) = s \cdot c l i p \left(\left\lfloor \frac {w}{s} \right\rceil , Q _ {n}, Q _ {p}\right), \tag {1} +$$ + +where $clip(\cdot, \alpha, \beta)$ is a truncated function with lower bound $\alpha$ and upper bound $\beta$ , $\lfloor \cdot \rceil$ is a rounding operator, $s$ is a scaling factor, $Q_{n}$ and $Q_{p}$ are the quantization ranges. + +Since the quantization function is non-differentiable, the gradient of $w$ cannot be directly computed during backpropagation. Thus, gradient approximation is used to keep the training going, such as straight through estimator (STE) [2]: + +$$ +\frac {\partial L}{\partial w} \approx \frac {\partial L}{\partial \hat {w}} \cdot \mathbf {1} _ {Q _ {n} \leq w / s \leq Q _ {p}}, \tag {2} +$$ + +where $\mathbf{1}$ is an indicator function. When $Q_{n} \leq w / s \leq Q_{p}$ , it is 1, otherwise it is 0. There is no gradient for $w$ outside the quantization range. In LSQ, the gradient of the scaling factor $s$ is computed as: + +$$ +\frac {\partial L}{\partial s} = \frac {\partial L}{\partial \hat {w}} \cdot \frac {\partial \hat {w}}{\partial s}, +$$ + +$$ +\frac {\partial \hat {w}}{\partial s} = \left\{ \begin{array}{c l} - \frac {w}{s} + \left\lfloor \frac {w}{s} \right\rceil & \text {i f} Q _ {n} < \frac {w}{s} < Q _ {p} \\ Q _ {n} & \text {i f} \frac {w}{s} \leq Q _ {n} \\ Q _ {p} & \text {i f} \frac {w}{s} \geq Q _ {p} \end{array} . \right. \tag {3} +$$ + +# 3.2. Limited Solution Space of Weights + +Neural network quantization can be formulated as a mixed-integer nonlinear programming (MINLP) problem. It involves converting floating-point weights and activations into a finite set of discrete values while minimizing performance degradation. In this formulation, quantized weights and activations act as integer variables, while the nonlinearity stems from forward propagation and activation functions. Continuous variables, such as quantization parameters, are also involved. The optimization objective is: + +$$ +\min _ {\mathbf {w}} \mathbb {E} _ {(x, y) \sim D} \left[ \mathcal {L} \left(f _ {\text {q u a n t}} (x; q (\mathbf {w})), y\right) \right], \tag {4} +$$ + +where $\mathcal{L}(\cdot)$ is the loss function, $\mathbf{w}$ is the set of all weights, $(x,y)\sim D$ represents samples drawn from the dataset $D$ , $f_{\mathrm{quant}}(\cdot)$ represents the model's output with activation quantization, and $q(\cdot)$ is the quantization function. $q(\mathbf{w})$ consists of integer variables $\mathbf{Q}$ and continuous parameters $s,z$ : + +$$ +\forall \mathbf {Q} \in [ 0, 2 ^ {b} - 1 ], \mathbf {Q} \in \mathbb {Z}, b \in \mathbb {N} _ {*}, q (\mathbf {w}) = s \cdot (\mathbf {Q} - z), \tag {5} +$$ + +where $\mathbf{Q}$ is the set of integers mapped from $\mathbf{w}$ , $s$ is the scaling factor, $z$ is the zero point, and $b$ is the bit-width. + +QAT-based methods serve as heuristics for solving this MINLP problem. These methods select an initial solution and iteratively refine it. Drawing inspiration from research on mixed-integer programming [6, 47], we observe that QAT shares similarities with depth-first search (DFS) in branch-and-bound algorithms. Both rely on branching and backtracking: QAT chooses quantization strategies per weight (branching), and uses gradient updates (implicit backtracking) to adjust solutions. However, unlike exact DFS, which systematically explores the entire solution space, QAT heuristics prioritize fast convergence, often leading to limited exploration and suboptimal results. We compare QAT and DFS from two key perspectives: + +- Branching of variables In DFS, the algorithm explicitly branches on relaxed variables to construct the search tree. Likewise, in QAT, weights oscillating near quantization thresholds effectively induce branching behavior, determining whether a weight falls into one bin or another. + +- Backtracking mechanism DFS performs explicit backtracking: when a branch leads to no improvement, it reverts and explores alternatives, enabling a global search. In contrast, QAT relies on local gradients to implicitly adjust weights. This often causes the model to converge prematurely and miss better configurations. + +In summary, both DFS and QAT can be regarded as rule-based search strategies. However, under low-bit constraints, QAT's gradient-based updates are more prone to becoming trapped in local optima due to the steep curvature of the loss surface [37]. Motivated by this limitation, we aim to enhance QAT by expanding its effective solution space, similar in spirit to the global search behavior exhibited by exact DFS, in order to achieve improved performance. + +We assess the ability of existing methods to explore the weight solution space by measuring how often weights cross quantization thresholds. As shown in Table 1, we use LSQ [12] to quantize ResNet-18 on the ImageNet [9]. "SGD" denotes an SGD optimizer with an initial learning rate of $1\mathrm{e} - 2$ and cosine learning rate decay [34], while "Adam" denotes an Adam optimizer with a learning rate of $1\mathrm{e} - 3$ , both following the training strategies in [12]. We find that the frequency of weights crossing quantization thresholds is very low, all below $0.5\%$ , with most untruncated weights concentrated in the $[0,0.01\%)$ range. This indicates that previous gradient-based optimization methods have a + +![](images/b2032bd9f14ac44bfa9db6a342421f95bcc094f4e0faaa95608cd76bf6e85bbc.jpg) +(a) + +![](images/b1debe54953a29a64dfce8ac3bca93787021eb555642c9246736f47024319937.jpg) +(b) +Figure 2. Histograms of pre-trained weights mapped to different quantization values: (a) lossless solution; (b) LSQ; and (c) our AOQ. The red, green, blue, and yellow bars represent histograms of pre-trained weights mapped to the quantized values -2, -1, 0, and 1, respectively. + +![](images/2500b8fb35bf5f9d3a718943f21fa2750211f5d62f31b9404771397013e2e587.jpg) +(c) + +Table 1. Measuring the solution space of weights on 2-bit quantized ResNet-18. The first column is the frequency of weights crossing the quantization thresholds, calculated as the number of crossings divided by the total iterations. "Truncated" refers to weights whose frequency of crossing quantization thresholds falls between 0 and $0.01\%$ due to being clipped by the truncation function. "Not Truncated" refers to not being clipped by the truncation function. $\%$ of weights" refers to the percentage of weights within a specific frequency range. + +
Frequency Range% of weights
SGDAdam
[0, 0.01%) - Truncated14.5%11.8%
[0, 0.01%) - Not Truncated47.2%12.6%
[0.01%, 0.05%)5.1%19.1%
[0.05%, 0.1%)8.4%24.5%
[0.1%, 0.5%)24.8%32.0%
+ +weak ability to explore the solution space, potentially hindering the model from finding an optimal solution. + +To further illustrate how limited exploration affects quantization quality, we construct a toy example. We first generate an integer set $\bar{w} \in \mathbb{R}^{10000}$ containing only values -2, -1, 0, and 1. We then derive matrix $\mathbf{A} \in \mathbb{R}^{2000 \times 10000}$ and vector $\mathbf{b} \in \mathbb{R}^{2000}$ satisfying $\mathbf{A}\bar{w} = \mathbf{b}$ , establishing $\bar{w}$ as our known lossless integer solution. Clearly, $\mathbf{A}$ is not a full-rank matrix, which indicates that the linear equation system has multiple solutions, similar to the multi-solution nature of neural networks. Solving $\mathbf{A}w = \mathbf{b}$ with the constraint $w \in \{-2s + z, -s + z, z, s + z\}^{10000}$ , where $s, z \in \mathbb{R}$ , simulates the quantization process. Conceptually, this setup mimics training a fully connected network with 2-bit quantization, 10,000 input units, one output unit, and 2,000 samples, using mean squared error loss and Adam optimizer with a learning rate of $1 \times 10^{-3}$ . + +We first train the model with floating-point weights until convergence, obtaining pre-trained weights, similar to the standard QAT procedure. These weights are then mapped to quantized values. In the lossless solution, the scaling factor $s$ is 1 and $z$ is 0. Figure 2a shows the distribution of pre-trained weights mapped to each quantized value under the lossless solution. Each quantization level corresponds to a wide range of pre-trained weights, indicating that allowing distant weights to map to the same quantized value can preserve model quality. In contrast, Figure 2b shows the distribution after freezing $s$ to 1 and $z$ to 0 and using LSQ [12] for quantization. In this setting, pre-trained weights are mapped only to the nearest quantized values, restricting the solution space. Such methods, by mapping only to nearby values, may result in more mapping errors and does not fully leverage the broader solution space to minimize quantization errors. + +# 3.3. Allowing Oscillation Quantization + +In previous methods, many weights seldom oscillate near quantization thresholds, limiting their ability to switch quantization bins and often resulting in suboptimal solutions. We speculate that in QAT, especially in low-bit settings, weight transitions between different quantization levels make gradient direction unstable. This instability causes weights to fluctuate within a narrow range during training, preventing them from crossing thresholds. While some studies [30, 36] consider weight oscillation harmful to the whole optimization process, we believe that oscillation actually promotes solution diversity, increasing the likelihood of discovering better optima. Specifically, oscillation acts as an adaptive perturbation mechanism that enables the model to continuously test new quantization configurations and explore more quantized solutions within a given number of iterations, helping avoid premature convergence. This phenomenon resembles "random perturbation" strategies used + +![](images/f14cd86f4a1f22fc20c3ef8920b81b9f19d6f5b71c3f207fb42164909c344abe.jpg) +Figure 3. Oscillation patterns of two weights in layer1.0.conv1 of ResNet-18 during early training. We apply AOQ under a 2-bit quantization setting and monitor quantization assignments at the end of each epoch. + +in integer programming [21, 40], where moderate instability prevents the model from becoming locked into overly deterministic paths. Based on this intuition, we propose to actively promote oscillation during early training phases to encourage exploration, then gradually reduce it in later stages to ensure convergence stability. + +To increase the number of weights that cross thresholds without disrupting QAT's gradient-based backtracking mechanism, we examine how quantization parameters affect oscillation. Inspired by [30], we observe that modifying quantization thresholds can induce more weights to oscillate around them, as shown in Figure 3. Specifically, when the intervals between quantization thresholds and the differences among quantization levels decrease, weights tend to transition from intra-bin oscillation to threshold-adjacent oscillation. This behavior is driven by two factors. First, the shifting thresholds force certain weights to cross them, introducing noise into the network. The facilitating effect of smaller quantization threshold intervals on weight oscillation will be demonstrated in Section 3.4. Second, the reduced differences between quantization levels lower the penalty for mapping errors when floating-point weights are mapped to quantized values, making weights more likely to oscillate near the thresholds, as smaller differences between quantization levels tend to smooth the loss surface. + +To leverage the weight oscillation mentioned above, we decouple the quantization thresholds and levels and introduce a three-stage training strategy. Specifically, we manually adjust the quantization thresholds to gradually reduce their intervals, while setting the quantization levels as learnable parameters to optimize during training. Additionally, we introduce differentiability for clipped weights in fake quantization to address the issue observed in Section 3.2, where the absence of gradients in clipped weights (as shown in Eq. (2)) hinders model fine-tuning. In the final stage, oscillation is suppressed to stabilize convergence. The complete strategy is outlined below: + +- Stage 1: Inducing oscillation. For weight quantization, + +we define the intervals between thresholds as $s_{th}$ and between quantization levels as $s_{le}$ , both initialized to $\frac{\sigma_w}{2^{b - 2}}$ where $\sigma_w$ is the standard deviation of weights in a layer, and $b$ is the bit-width. In this stage, we manually reduce $s_{th}$ and $s_{le}$ to promote oscillation at thresholds. + +- Stage 2: Learning quantization levels. We fix $s_{th}$ and set $s_{le}$ as a learnable parameter for optimization. In Stage 1, $s_{le}$ is manually reduced, which leads to higher loss due to poor representation of weights by the quantization levels. As training progresses and the task loss decreases, $s_{le}$ is learned to increase automatically. This widens the intervals between quantization levels, makes the loss landscape steeper near thresholds, and reduces oscillation by discouraging weights from frequently crossing thresholds. The quantizer is defined as: + +$$ +\hat {w} = \left\{ \begin{array}{l l} \operatorname {l e v e l} [ 0 ] & \text {i f} w < \operatorname {t h r e s h} [ 0 ] \\ \operatorname {l e v e l} [ i ] & \text {i f t h r e s h} [ i - 1 ] \leq w < \operatorname {t h r e s h} [ i ] \\ \operatorname {l e v e l} \left[ 2 ^ {b} - 1 \right] & \text {i f} w \geq \operatorname {t h r e s h} \left[ 2 ^ {b} - 2 \right] \end{array} \right. \tag {6} +$$ + +where $i\in \{1,2,\ldots ,2^b -2\}$ and $\mathrm{thresh}[i]$ and level[i] represent the $(\mathrm{i + 1})$ -th threshold and level, respectively. The optimized quantized weights are denoted as: + +$$ +\tilde {w} = \hat {w} + w - w. \det a c h (\cdot). \tag {7} +$$ + +This formulation, based on the STE, enables distinct gradient flows: $\hat{w}$ determines the gradient for $s_{le}$ , while $w$ receives gradients directly from the task loss, even when clipped. This decoupling avoids the zero-gradient issue and enables more effective optimization of both weights and quantizer parameters. + +- Stage 3: Dampening oscillation. In the third stage, we employ the Oscillation dampening method outlined in [36], which adds a regularization term to the loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t a s k}} + \lambda \mathcal {L} _ {\text {d a m p e n}}, \tag {8} +$$ + +where $\lambda$ is a hyperparameter. The dampening loss is: + +$$ +\mathcal {L} _ {\text {d a m p e n}} = \left\| \hat {w} - \operatorname {c l i p} \left(w, Q _ {n}, Q _ {p}\right) \right\| _ {F} ^ {2}, \tag {9} +$$ + +which penalizes weights lingering near quantization thresholds, encouraging them to converge to the nearest quantization centers. + +Details regarding the hyperparameters and transition schedule for each training stage are provided in Section 4.1. + +# 3.4. Theoretical Analysis + +In this section, we theoretically analyze the effectiveness of our approach and compare it against prior differentiable quantization parameter methods. + +We first show the advantage of using smaller quantization threshold intervals. Let $W$ be the set of pre-trained + +weights, and $Q$ be a set composed of quantized values. Weight quantization is a mapping from the $W$ to the $Q$ . For a given weight $w \in W$ , we define its probability of being quantized to $Q_{i} \in Q$ as the transition probability $P(Q_{i} \mid w)$ , where $i \in \{0, 1, \dots, 2^{b} - 1\}$ and $b$ is the bit-width. By Bayes' theorem: + +$$ +P \left(Q _ {i} \mid w\right) = \frac {P \left(Q _ {i}\right) \cdot P \left(w \mid Q _ {i}\right)}{P (w)}, \tag {10} +$$ + +where $P(Q_{i})$ is the proportion of the number of weights mapped to $Q_{i}$ to the total number of weights, $P(w)$ is the probability distribution of weights, and $P(w|Q_i)$ is the conditional distribution of weights mapped to $Q_{i}$ . The total probability formula applied to Eq. (10) is: + +$$ +P \left(Q _ {i} \mid w\right) = \frac {P \left(Q _ {i}\right) \cdot P \left(w \mid Q _ {i}\right)}{\sum_ {i} P \left(Q _ {i}\right) \cdot P \left(w \mid Q _ {i}\right)}. \tag {11} +$$ + +As shown in Figure 2a, $P(w \mid Q_i)$ approximately follows a Gaussian distribution, so the total probability formula can be transformed into a Gaussian mixture model: + +$$ +P (w) = \sum_ {i} \pi_ {i} \mathcal {N} \left(w \mid \mu_ {i}, \sigma_ {i}\right), \tag {12} +$$ + +where $\pi_i, \mu_i, \sigma_i$ denote the weight, mean and variance of the Gaussian component, respectively. This transformation does not effect the theoretical conclusion, just for analytical convenience. Eq. (11) is transformed into: + +$$ +P \left(Q _ {i} \mid w\right) = \frac {\pi_ {i} \mathcal {N} \left(w \mid \mu_ {i} , \sigma_ {i}\right)}{\sum_ {i} \pi_ {i} \mathcal {N} \left(w \mid \mu_ {i} , \sigma_ {i}\right)}. \tag {13} +$$ + +The weight $w$ is assigned to the quantization value $Q_{j}$ that maximizes this probability: + +$$ +i f \forall i, \pi_ {j} \mathcal {N} (w \mid \mu_ {j}, \sigma_ {j}) \geq \pi_ {i} \mathcal {N} (w \mid \mu_ {i}, \sigma_ {i}) \tag {14} +$$ + +then $w\to Q_j$ + +Therefore, when the quantization threshold is set at the horizontal coordinate value of the intersection between two adjacent Gaussian components, it ensures optimal probabilistic mapping of weights. In Figure 2a, we observe that in the lossless solution for this example, the variances of each Gaussian component are approximately equal, denoted as $\sigma$ . Thus, the horizontal coordinate value at the intersection of adjacent Gaussian components is: + +$$ +\operatorname {t h r e s h} [ i ] = \frac {\mu_ {i + 1} + \mu_ {i}}{2} - \frac {\log \left(\frac {\pi_ {i + 1}}{\pi_ {i}}\right) \cdot \sigma^ {2}}{\mu_ {i + 1} - \mu_ {i}}. \tag {15} +$$ + +This analysis suggests that, when Gaussian component means are close, using smaller threshold intervals improves mapping accuracy. It explains why our method encourages + +tighter thresholds in early training to support more effective weight assignment in QAT. + +Next, we show that AOQ provides a more direct and accurate link between quantization levels and task loss than previous approaches. In prior methods [5, 12, 49], learnable quantization parameters jointly determine both thresholds and levels. This coupling introduces approximation errors in the gradient computation of quantization levels. For instance, in LSQ [12], the gradients of the scaling factors are shown in Eq. (3). The derivative of weights with respect to the scaling factor aids in optimizing quantization thresholds but makes the gradient of the quantization levels inaccurate. In contrast, AOQ decouples thresholds from levels. It directly optimizes quantization levels based on Eq. (6) and Eq. (7), allowing a more faithful gradient flow that better aligns with task loss. + +Finally, AOQ effectively addresses the impact of weight oscillation on final convergence, as well as on batch normalization and layer normalization. After exploring a broader solution space, we use Oscillation Dampening in the final stage to gradually guide weights out of oscillatory regions. As training continues, the exponential moving average used in batch normalization and layer normalization becomes increasingly stable, reinforcing consistent model behavior and improving convergence quality. + +# 4. Experiments + +We evaluate the effectiveness of the proposed Allowing Oscillation Quantization (AOQ) on the ILSVRC12 ImageNet dataset [9], using a range of models including ResNet [15], MobileNet-V2 [17], DeiT [41], and Swin Transformer [33]. Our experiments focus on low-bit weight quantization. For activation quantization, we adopt LSQ [12] and N2UQ [32]. + +# 4.1. Experimental Setup + +ResNet and MobileNet We follow the training schemes in [5, 12, 32], initializing quantized networks from full-precision pre-trained models. Training is performed using the Adam optimizer [20] with a linear learning rate decay, starting from $1 \times 10^{-3}$ . Following LSQ [12], we apply knowledge distillation and train for 150 epochs. All convolutional and fully connected layers are quantized, except for the first and last layers. + +DeiT and Swin Transformer We use the official implementations and adopt training settings from [28-30], including knowledge distillation and a training duration of 300 epochs. As in prior work, the patch embedding and classification layers are quantized to 8 bits [28-30]. + +Parameter Setup The batch size is set to 512, and the weight decay is set to 0. The three training stages occupy $1/5$ , $2/5$ , and $2/5$ of the total training epochs, respectively. We reduce $s_{\mathrm{th}}$ and $s_{\mathrm{le}}$ to $30\% - 35\%$ of their initial values in the first stage. As shown in Fig. 4, a broad grid search + +Table 2. Accuracy comparison to the state-of-the-art quantization methods with ResNet structure on ImageNet dataset. Note that W/A denote the bit-widths of weights and activations, respectively. FP denotes the top-1 accuracy of the full-precision models. + +
NetworkMethodBit-width (W/A)Accuracy (%) Top-1Bit-width (W/A)Accuracy(%) Top-1Bit-width (W/A)Accuracy(%) Top-1
ResNet-18 (FP: 71.8)LQ-Nets [46]2/264.93/368.24/469.3
DSQ [13]2/265.23/368.74/469.6
QIL [19]2/265.73/369.24/470.1
APoT [27]2/267.33/369.94/470.7
LSQ [12]2/267.63/370.24/471.1
LCQ [43]2/268.93/370.64/471.5
N2UQ [32]2/269.43/371.94/472.9
LSQ+AOQ2/269.73/371.44/472.0
N2UQ+AOQ2/270.33/372.14/472.7
ResNet-34 (FP:74.9)LQ-Nets [46]2/269.83/371.94/4-
DSQ [13]2/270.03/372.54/472.8
QIL [19]2/270.63/373.14/473.7
APoT [27]2/270.93/373.44/473.8
LSQ [12]2/271.63/373.44/474.1
LCQ [43]2/272.73/374.04/474.3
N2UQ [32]2/273.33/375.24/476.0
LSQ+AOQ2/273.53/374.94/475.5
N2UQ+AOQ2/273.93/375.54/476.0
ResNet-50 (FP: 77.0)LQ-Nets [46]2/271.53/374.24/475.1
APoT [27]2/273.43/375.84/476.6
LSQ [12]2/273.73/375.84/476.7
LCQ [43]2/275.13/376.34/476.6
N2UQ [32]2/275.83/377.54/478.0
LSQ+AOQ2/276.33/377.54/477.9
N2UQ+AOQ2/276.43/377.84/478.1
+ +![](images/83a4b8f167de8458948e7fdcaed756a05136101bab787c9e1e07d387073d366e.jpg) +(a) ResNet-18 + +![](images/b4fd1ea6ad70f8a2c172724c4ffdeb0ed1df4dfdac76f34074340a01ad5201a9.jpg) +(b) DeiT-T +Figure 4. Accuracy heatmaps on ImageNet under different narrowing schedules. $\alpha$ denotes the final-to-initial ratio of $s_{le}$ and $s_{th}$ in stage one, and $\beta$ is the proportion of total epochs allocated to this stage. AOQ maintains stable performance across a wide range of $(\alpha, \beta)$ settings on both (a) 2-bit ResNet-18 and (b) 2-bit DeiT-T, indicating schedule robustness across architectures. + +confirms that a common schedule performs robustly across different architectures in practice. For the final stage, the hyper-parameter $\lambda$ is chosen such that $\lambda \mathcal{L}_{\mathrm{dampen}}$ is two to three orders of magnitude lower than $\mathcal{L}_{\mathrm{task}}$ . + +# 4.2. Experimental Results on Toy Example + +We first revisit the toy example from Section 3.2 to evaluate the effect of oscillation on quantization quality. To + +![](images/78e094382a8c8fb6520420029da6247389ac8e95ae55767ad42c1e963d380ba0.jpg) +Figure 5. Loss curves and minimum loss obtained from solving the toy example using various methods. + +
MethodLoss
LSQ51.2
LSQ+OD(final)16.7
LSQ+OD(enti)26.2
AOQ5.6
+ +demonstrate that weight oscillation can expand the solution space and improve quantization outcomes, we apply the Oscillation Dampening (OD) [36] to suppress oscillation either during the final training stage or throughout the entire training process. We also evaluate the performance of AOQ using the toy example. As shown in Figure 5, allowing early-stage oscillation improves QAT performance, and AOQ converges faster to a better solution. When using fixed quantization levels $\{-2, -1, 0, 1\}$ in the third stage, Figure 2c shows that AOQ enables pre-trained weights to be mapped to more diverse and distant quantized values, expanding the solution space compared to Figure 2b. + +Table 3. Top-1 accuracy (\%) comparison for methods suppressing weight oscillation throughout the process in MobileNet-V2. + +
NetworkMethodW3A3W4A4
MobileNet-V2 (FP: 72.1)LSQ [12]65.369.5
LSQ+BR [14]67.470.4
LSQ+Dampen [36]67.870.5
N2UQ [32]69.072.1
LSQ+AOQ69.171.8
N2UQ+AOQ69.972.5
+ +Table 4. Top-1 accuracy (\%) comparison with the state-of-the-art methods in the ViT architecture. + +
NetworkMethodW2A2W3A3
DeiT-T (FP: 72.1)LSQ [12]54.568.1
QViT* [28]50.467.1
OFQ [30]64.372.7
LSQ+AOQ66.572.3
DeiT-S (FP: 79.9)LSQ [12]68.077.8
QViT* [28]68.778.5
OFQ [30]75.779.6
LSQ+AOQ76.779.8
Swin-T (FP: 81.2)LSQ [12]70.479.0
QViT* [28]73.980.1
OFQ [30]78.581.1
LSQ+AOQ78.981.2
+ +# 4.3. Accuracy Evaluation + +Table 2 reports the accuracy of AOQ on ResNet. When paired with different activation quantization techniques, AOQ consistently improves performance. For instance, using AOQ with LSQ yields $0.9\% - 2.6\%$ accuracy gains. When combined with N2UQ, AOQ outperforms many traditional methods. Its advantage becomes more pronounced in low-bit networks with a large number of weights, where the expanded solution space is more effectively utilized. In the 2-bit ResNet-50 case, AOQ reduces the accuracy gap with full-precision to just $0.6\%$ . + +Table 3 shows the performance of the AOQ method on MobileNet-V2. Compared to other methods, AOQ has a natural advantage in quantizing MobileNet. Due to significant variations in weight distributions across channels in the depth-wise separable convolution layers, traditional quantization methods often map all weights in certain channels to a single quantized value, resulting in the loss of important information. AOQ mitigates this by adaptively shrinking threshold intervals during training, promoting threshold-crossing behavior and improving quantization granularity. This effect is illustrated in Figure 3. Overall, AOQ outper + +Table 5. Effects of different components of our strategy on the final performance. + +
MethodAcc (%)
Baseline [12]73.7
- Weight quantization76.6
+ Expanding solution space of weights75.2
+ Delayed weight oscillation suppression75.6
+ Both76.3
Corresponding real-valued network77.0
+ +forms other QAT approaches, including those that suppress oscillation throughout [14, 36]. + +To validate AOQ on Vision Transformers (ViT), we conducted 2-bit and 3-bit quantization experiments on DeiT and Swin Transformer. We observe that QViT [28] does not fully quantize activations in the fully connected and attention layers, limiting hardware acceleration. For fair comparison, we modify their implementation accordingly and report the results as QViT* in Table 4. The results show that AOQ significantly outperforms LSQ and QViT by suppressing weight oscillation, boosting 2-bit DeiT-T accuracy by $12.0\%$ and $16.1\%$ , respectively. By expanding the solution space, AOQ also surpasses OFQ, the prior state-of-the-art, under 2-bit quantization. + +# 4.4. Ablation Study + +To investigate the impact of each improvement on the final performance, we conduct an ablation study on the 2-bit quantized ResNet-50 network on ImageNet. As shown in Table 5, the proposed expanded solution space and delayed weight oscillation suppression improve the accuracy by $1.5\%$ and $1.9\%$ , respectively over the baseline [12]. By using both techniques, we reduced the accuracy gap with the corresponding real-valued ResNet-50 to only $0.7\%$ , and the accuracy is only reduced by $0.3\%$ compared to the baseline method without weight quantization. + +# 5. Conclusion + +In this work, we reveal the weakness of existing QAT methods in exploring the quantization solution space of weights within low bit-width constraints, which adversely impacts the performance of quantized networks. We highlight the importance of weight oscillation during early training, as it helps the optimizer escape local minima and explore a broader set of quantized configurations. Based on this insight, we propose a simple yet effective method that encourages weight oscillation in the early stage to expand the solution space, followed by oscillation suppression to stabilize convergence. Experiments demonstrate that our algorithm quantizes various models to low-bits and achieves state-of-the-art accuracy on ImageNet. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China under Grant 62322117, 62425109, 62371365, U24B20136, and U22B2014. + +# References + +[1] Haoli Bai, Wei Zhang, Lu Hou, Lifeng Shang, Jing Jin, Xin Jiang, Qun Liu, Michael R. Lyu, and Irwin King. Binarybert: Pushing the limit of bert quantization. In Annual Meeting of the Association for Computational Linguistics, 2020. 1 +[2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3 +[3] Hanlin Chen, Li'an Zhuo, Baochang Zhang, Xiawu Zheng, Jianzhuang Liu, Rongrong Ji, David S. Doermann, and Guodong Guo. Binarized neural architecture search for efficient object recognition. International Journal of Computer Vision, 129:501-516, 2020. 1 +[4] Shangyu Chen, Wenya Wang, and Sinno Jialin Pan. Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization. In Neural Information Processing Systems, 2019. 2 +[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2, 6 +[6] Michele Conforti, Gérard Cornuéjols, Giacomo Zambelli, Michele Conforti, Gérard Cornuéjols, and Giacomo Zambelli. Integer programming models. Springer, 2014. 3 +[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Neural Information Processing Systems, 2015. 2 +[8] Alexandre Défossez, Yossi Adi, and Gabriel Synnaeve. Differentiable model compression via pseudo quantization noise. Transactions on Machine Learning Research, 2021. 1, 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 3, 6 +[10] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. In Neural Information Processing Systems, pages 30318-30332, 2022. 1 +[11] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Neural Information Processing Systems, 2024. 1 +[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 1, 2, 3, 4, 6, 7, 8 +[13] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tian-Hao Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differen + +tiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF International Conference on Computer Vision, pages 4851-4860, 2019. 1, 2, 7 +[14] Tiantian Han, Dong Li, Ji Liu, Lu Tian, and Yi Shan. Improving low-precision network quantization via bin regularization. In IEEE/CVF International Conference on Computer Vision, pages 5241–5250, 2021. 8 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[16] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In Neural Information Processing Systems, 2024. 1 +[17] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6 +[18] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew G. Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2704-2713, 2017. 1, 2 +[19] San Hun Jung, Changyong Son, Seohyung Lee, JinWoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4345-4354, 2018. 2, 7 +[20] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[21] Scott Kirkpatrick, Charles D. Gelatt, and Mario P. Vecchi. Optimization by simulated annealing. Science, 220:671-680, 1983. 5 +[22] Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-50. Springer, 2002. 2 +[23] Junghyup Lee, Dohyung Kim, and Bumsub Ham. Network quantization with element-wise gradient scaling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6444-6453, 2021. 1, 2 +[24] Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dongsoo Lee. Flexround: Learnable rounding based on elementwise division for post-training quantization. In International Conference on Machine Learning, 2023. 1 +[25] Cong Leng, Hao Li, Shenghuo Zhu, and Rong Jin. Extremely low bit neural network: Squeeze the last bit out with admm. In AAAI Conference on Artificial Intelligence, 2017. 2 +[26] Xinlin Li, Bang Liu, Ruizhi Yang, Vanessa Courville, Chao Xing, and V. Nia. Denseshift : Towards accurate and efficient low-bit power-of-two quantization. In IEEE/CVF International Conference on Computer Vision, pages 16964-16974, 2022. 2 +[27] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for + +neural networks. In International Conference on Learning Representations, 2019. 1, 2, 7 +[28] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Neural Information Processing Systems, pages 34451-34463, 2022. 6, 8 +[29] Zhexin Li, Tong Yang, Peisong Wang, and Jian Cheng. Qvit: Fully differentiable quantization for vision transformer. arXiv preprint arXiv:2201.07703, 2022. +[30] Shi Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning, 2023. 1, 2, 4, 5, 6, 8 +[31] Zechun Liu, Baoyuan Wu, Wenhan Luo, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision, pages 722-737, 2018. 1 +[32] Zechun Liu, Kwang-Ting Cheng, Dong Huang, Eric P. Xing, and Zhiqiang Shen. Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4932-4942, 2021. 1, 2, 6, 7, 8 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 6 +[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 3 +[35] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 1 +[36] Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort. Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning, 2022. 1, 2, 4, 5, 7, 8 +[37] Yury Nahshan, Brian Chmiel, Chaim Baskin, Evgenii Zheltonozhskii, Ron Banner, Alexander M. Bronstein, and Avi Mendelson. Loss aware post-training quantization. Machine Learning, 110:3245 - 3262, 2019. 2, 3 +[38] Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eunhyeok Park. Nipq: Noise proxy-based integrated pseudo-quantization. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3852–3861, 2022. 2 +[39] Zhenhong Sun, Ce Ge, Junyan Wang, Ming Lin, Hesen Chen, Hao Li, and Xiuyu Sun. Entropy-driven mixed-precision quantization for deep network design. In Neural Information Processing Systems, 2022. 1 +[40] Charles E. Taylor. Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence. The Quarterly Review of Biology, 69:88-89, 1994. 5 + +[41] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 6 +[42] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2046-2055, 2020. 1 +[43] Kohei Yamamoto. Learnable companding quantization for accurate low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5027-5036, 2021. 2, 7 +[44] Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xian-sheng Hua. Quantization networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7316, 2019. 2 +[45] Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, and Jack Xin. Understanding straight-through estimator in training activation quantized neural nets. In International Conference on Learning Representations, 2019. 1, 2 +[46] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In European Conference on Computer Vision, 2018. 1, 2, 7 +[47] Jiayi Zhang, Chang Liu, Junchi Yan, Xijun Li, Hui-Ling Zhen, and Min jie Yuan. A survey for solving mixed integer programming via machine learning. Neurocomputing, 519:205-217, 2022. 3 +[48] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yuan Du, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. In IEEE/CVF International Conference on Computer Visio, pages 3802-3812, 2023. 1 +[49] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1, 2, 6 \ No newline at end of file diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/images.zip b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a104457fa61e7e7671b1597a299376b1fb6c1ab --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:648d10071c48fb0299874e0b3aaa4544c31068b77deabfeaf0d9b079de313342 +size 492076 diff --git a/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/layout.json b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7b5c3d26deebb52f7b7d50ac795c951d26a3f1d6 --- /dev/null +++ b/2025/Allowing Oscillation Quantization_ Overcoming Solution Space Limitation in Low Bit-Width Quantization/layout.json @@ -0,0 +1,9597 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 542, + 140 + ], + "type": "text", + "content": "Allowing Oscillation Quantization: Overcoming Solution Space Limitation in Low Bit-Width Quantization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "content": "Weiying Xie" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "content": " Zihan Meng" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "content": " Jitao Ma" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "content": " Wenjin Guo" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "text", + "content": " Haowei Li" + }, + { + "bbox": [ + 134, + 160, + 476, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "spans": [ + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "text", + "content": "Haonan Qin" + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "text", + "content": " Leyuan Fang" + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "text", + "content": " Yunsong Li" + }, + { + "bbox": [ + 197, + 175, + 411, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "spans": [ + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "type": "text", + "content": "Xidian University " + }, + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 190, + 402, + 203 + ], + "type": "text", + "content": "Hunan University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "text", + "content": "Quantization-aware Training (QAT) enables deep models to adapt to precision loss by simulating quantization. However, existing methods often converge to sub-optimal solutions due to inadequate exploration of quantization solution space. To address this, we propose a novel QAT method, Allowing Oscillation Quantization (AOQ), which expands the reachable solution space through weight oscillation. Notably, unlike previous methods that suppress oscillation throughout training, AOQ actively encourages it in the earlier stages to explore diverse quantization configurations, and suppresses it later to ensure convergence. In addition, by decoupling quantization thresholds and levels, AOQ promotes meaningful oscillation and improves the stability of learnable quantization parameters. Extensive experiments across various models, including ResNet, MobileNet, DeiT and Swin Transformer, demonstrate the effectiveness of our method. Specifically, with 2-bit quantization, AOQ achieves a " + }, + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "inline_equation", + "content": "0.4\\% \\sim 2.2\\%" + }, + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "text", + "content": " accuracy improvement on ImageNet compared to state-of-the-art methods. Our implementation is available at https://github.com/muzenc/AOQ." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 135, + 532 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 539, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 647 + ], + "type": "text", + "content": "Model quantization [31, 46, 49] has attracted growing attention for enabling efficient model inference while substantially reducing hardware resource consumption. By reducing the bit-width of model parameters, quantization effectively reduces memory footprint and computational complexity, thereby enhancing inference speed and energy efficiency. Thus, quantization has been widely applied in tasks such as image classification [27, 32, 39], object detection [3, 42, 48] and natural language processing [1, 10, 11]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "Despite its advantages, quantization shrinks the dynamic range of parameters, often resulting in information loss and degraded performance. Quantization-aware training (QAT) significantly mitigates the impact of quantization on perfor" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 234, + 552, + 347 + ], + "blocks": [ + { + "bbox": [ + 320, + 234, + 552, + 347 + ], + "lines": [ + { + "bbox": [ + 320, + 234, + 552, + 347 + ], + "spans": [ + { + "bbox": [ + 320, + 234, + 552, + 347 + ], + "type": "image", + "image_path": "30f7d84627e4b80f2d787dee8a1fbdeafc8a22528e7492819284c059066e6636.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 354, + 406, + 364 + ], + "lines": [ + { + "bbox": [ + 342, + 354, + 406, + 364 + ], + "spans": [ + { + "bbox": [ + 342, + 354, + 406, + 364 + ], + "type": "text", + "content": "- Discrete Solution" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 354, + 504, + 364 + ], + "lines": [ + { + "bbox": [ + 440, + 354, + 504, + 364 + ], + "spans": [ + { + "bbox": [ + 440, + 354, + 504, + 364 + ], + "type": "text", + "content": "Solution Space" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 369, + 380, + 378 + ], + "lines": [ + { + "bbox": [ + 340, + 369, + 380, + 378 + ], + "spans": [ + { + "bbox": [ + 340, + 369, + 380, + 378 + ], + "type": "inline_equation", + "content": "\\bigcirc" + }, + { + "bbox": [ + 340, + 369, + 380, + 378 + ], + "type": "text", + "content": " Contour" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 368, + 534, + 379 + ], + "lines": [ + { + "bbox": [ + 440, + 368, + 534, + 379 + ], + "spans": [ + { + "bbox": [ + 440, + 368, + 534, + 379 + ], + "type": "text", + "content": "Expanded Solution Space" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 384, + 477, + 394 + ], + "lines": [ + { + "bbox": [ + 394, + 384, + 477, + 394 + ], + "spans": [ + { + "bbox": [ + 394, + 384, + 477, + 394 + ], + "type": "text", + "content": "2D quantization problem:" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 345, + 397, + 535, + 411 + ], + "lines": [ + { + "bbox": [ + 345, + 397, + 535, + 411 + ], + "spans": [ + { + "bbox": [ + 345, + 397, + 535, + 411 + ], + "type": "text", + "content": "subject to " + }, + { + "bbox": [ + 345, + 397, + 535, + 411 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 345, + 397, + 535, + 411 + ], + "type": "text", + "content": " is quantization function" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 423, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. Illustration of solving a 2D quantization problem using previous methods vs. proposed method. The two plots are contour maps of the function " + }, + { + "bbox": [ + 313, + 423, + 555, + 499 + ], + "type": "inline_equation", + "content": "f(w_{1},w_{2})" + }, + { + "bbox": [ + 313, + 423, + 555, + 499 + ], + "type": "text", + "content": ", where darker regions indicate lower function value. The yellow and orange curves show the solution trajectories of the previous method and our method, respectively. In contrast, AOQ explores a broader solution space and finds a better solution." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 553, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 557 + ], + "type": "text", + "content": "mance compared to post-training quantization [16, 24, 35], but a noticeable performance gap remains under extremely low-bit quantization scenarios, such as 2-bit quantization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "Mainstream QAT techniques [12, 32, 46] simulate quantization effects by inserting fake quantization nodes during training, allowing the model to adapt to low-precision constraints. However, these nodes are non-differentiable, leading prior works [13, 23, 45] to employ biased gradient approximations. Recent studies [8, 30, 36] show that such biased gradients can result in weight oscillation, hindering the network's convergence to a better local minimum. Nevertheless, an overlooked insight is that QAT relies on weight oscillation to explore more quantized solutions. In existing methods [12, 13, 18, 32], the frequency of weight oscillation is relatively low, and directly suppressing oscillations may limit the model's ability to escape suboptimal quanti" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24615" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "zation bins. Specifically, weights seldom cross quantization thresholds, instead making insignificant shifts within quantization bins." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 312 + ], + "type": "text", + "content": "We investigate QAT from the perspective of solving nonlinear mixed-integer programming problems, drawing an analogy to the branch-and-bound method with depth-first search (DFS). While DFS systematically explores all possible solutions through branching and backtracking, QAT lacks an exhaustive exploration mechanism and instead performs local updates guided by gradients. Our analysis reveals that in existing methods, most weights infrequently change their quantization bin assignments, thereby restricting the search to a small subset of possible quantization solutions and resulting in a limited solution space. As shown in Figure 1, this limited search space often leads to premature convergence, leaving potentially optimal configurations unexamined. According to [37], low bit-width quantization introduces sharp curvature in the loss landscape, indirectly demonstrating that gradient-based optimization is more likely to get trapped in poor local minima." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 312, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 514 + ], + "type": "text", + "content": "To address the problems mentioned above, we propose a novel QAT approach, named Allowing Oscillation Quantization (AOQ). Unlike previous methods that suppress weight oscillation throughout the entire optimization process, AOQ strategically leverages weight oscillation as a positive exploration mechanism. We encourage weight oscillation to expand the solution space of the weights and only suppress oscillation in the later stages of training. Moreover, AOQ addresses a key limitation of traditional QAT methods, where the coupling of quantization thresholds and levels often causes inaccurate updates to the quantization parameters and hinders model convergence [30]. To overcome this, AOQ decouples thresholds and levels, allowing thresholds to be manually adjusted to encourage oscillation while setting quantization levels as learnable parameters directly optimized with the task loss. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 516, + 295, + 659 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 516, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 574 + ], + "type": "text", + "content": "- We empirically show that fake quantization-based QAT methods explore a limited subset of quantized solutions and are prone to suboptimal convergence. To the best of our knowledge, this is the first systematic demonstration of this phenomenon." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 623 + ], + "type": "text", + "content": "- We propose AOQ to expand the quantization solution space of the weights by leveraging weight oscillation. Moreover, we improve the stability of learnable quantization parameters by decoupling the thresholds and levels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 659 + ], + "type": "text", + "content": "- Extensive experiments in CNN and Vision Transformer architectures demonstrate the effectiveness of AOQ in extremely low-bit quantization." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "QAT was initially proposed for binary quantization [7], and has since become the most popular quantization method in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "precision-sensitive scenarios. Most QAT-based methods insert fake quantization nodes into the model. These nodes simulate the effects of quantization during forward propagation and pass the approximated gradients during backward propagation [18, 19, 26, 27, 43, 46]. Some other methods [8, 38] add independent pseudo quantization noise during training, enabling effective QAT without relying on gradient approximations. This work focuses on the former." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 171, + 553, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 494 + ], + "type": "text", + "content": "The former is non-differentiable at the fake quantization nodes, preventing optimization and convergence of the quantized model. To address this problem, earlier works approximate gradients using straight-through estimator (STE) [22, 45] and its variants [13, 23, 44], while other approaches compute gradients using constructed neural networks or manual derivation [4, 25]. Among them, STE is widely used in quantization methods due to its robustness, broad applicability, and straightforward implementation. However, it inevitably brings the problem of gradient mismatch: gradients of weights are computed based on their quantized values, instead of their floating-point values. Previous research demonstrated that following the STE approach can cause weight oscillation during training due to biased gradients [8]. It has been suggested that weight oscillation prevents the network from converging to the best local minimum, and that suppressing oscillation earlier is beneficial for the optimization process [36]. The oscillation dampening and iterative weight freezing methods it proposed overcome weight oscillation. In addition, OFQ [30] observed that weight oscillation can introduce noise to the gradient of the learnable scaling factor (quantization level) and cause it to oscillate. The latter, in turn, changes the quantization threshold and aggravates the weight oscillation. OFQ deems weight oscillation harmful and suppresses it; in contrast, AOQ dispels this long-standing misconception, theoretically and experimentally proving its value." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 496, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 651 + ], + "type": "text", + "content": "Moreover, several works have delved into searching for optimal scaling factor to determine quantization thresholds and levels. For instance, trainable clipping parameters have been introduced to determine the appropriate quantization scale [5, 49]. Esser et al. [12] provided a simple yet effective way to approximate the gradient of learnable scaling factor, which is sensitive to quantized state transitions. It has been widely adopted as the de facto approach for QAT. Another line of research decouples quantization thresholds and levels, proposing increasing the quantization representation ability by replacing uniform quantization with nonuniform quantization [32]. The approach has demonstrated excellent performance in quantization of activation values." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "While we acknowledge that weight oscillation impacts final convergence, we argue that it is not entirely detrimental to optimization. In layer-by-layer quantization, each weight represents a dimension within the quantization solution space, and weight oscillation plays a crucial role in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24616" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "rapidly exploring optimal quantization solutions within limited training iterations. Moreover, our research shows that in existing methods, the oscillation frequency of weights at quantization thresholds is quite low, motivating us to increase this frequency. To this end, we delve into the working mechanism of QAT, and our proposed AOQ method combines the benefits of enhancing weight oscillation to expand the solution space and suppressing weight oscillation to ensure stable convergence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 189, + 138, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 138, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 138, + 202 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 208, + 141, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 141, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 141, + 220 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 225, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 296, + 309 + ], + "type": "text", + "content": "One of the most effective ways to quantize a neural network is by inserting fake quantization nodes to simulate the quantization effect during training. In weight quantization, the goal is to identify the quantization bin each weight falls into and assign it the corresponding quantization level. At present, the most successful methods are based on LSQ [12], and its quantization function is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 312, + 295, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 312, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 312, + 295, + 335 + ], + "type": "interline_equation", + "content": "\\hat {w} = q (w; s, Q _ {n}, Q _ {p}) = s \\cdot c l i p \\left(\\left\\lfloor \\frac {w}{s} \\right\\rceil , Q _ {n}, Q _ {p}\\right), \\tag {1}", + "image_path": "251f47a75e2144a61c727e2e73f7a49d862dfb435cc990f74790a339f6b86cea.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "clip(\\cdot, \\alpha, \\beta)" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " is a truncated function with lower bound " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " and upper bound " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rceil" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " is a rounding operator, " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " is a scaling factor, " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "Q_{n}" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "inline_equation", + "content": "Q_{p}" + }, + { + "bbox": [ + 55, + 338, + 296, + 373 + ], + "type": "text", + "content": " are the quantization ranges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 374, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 422 + ], + "type": "text", + "content": "Since the quantization function is non-differentiable, the gradient of " + }, + { + "bbox": [ + 55, + 374, + 296, + 422 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 374, + 296, + 422 + ], + "type": "text", + "content": " cannot be directly computed during backpropagation. Thus, gradient approximation is used to keep the training going, such as straight through estimator (STE) [2]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 424, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 424, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 120, + 424, + 295, + 449 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial w} \\approx \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\mathbf {1} _ {Q _ {n} \\leq w / s \\leq Q _ {p}}, \\tag {2}", + "image_path": "bf4705108ffd95b4a4dabf27ee4590aa725dbffd46abaa1bd5024463b8ce8992.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " is an indicator function. When " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "Q_{n} \\leq w / s \\leq Q_{p}" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": ", it is 1, otherwise it is 0. There is no gradient for " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " outside the quantization range. In LSQ, the gradient of the scaling factor " + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 451, + 296, + 499 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 502, + 156, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 502, + 156, + 527 + ], + "spans": [ + { + "bbox": [ + 84, + 502, + 156, + 527 + ], + "type": "interline_equation", + "content": "\\frac {\\partial L}{\\partial s} = \\frac {\\partial L}{\\partial \\hat {w}} \\cdot \\frac {\\partial \\hat {w}}{\\partial s},", + "image_path": "cb09d2c9a588390d28ce1886eff22d08c460e1a62ddd54b4727d9aad4b98e4b0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 84, + 528, + 295, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 528, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 84, + 528, + 295, + 567 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\hat {w}}{\\partial s} = \\left\\{ \\begin{array}{c l} - \\frac {w}{s} + \\left\\lfloor \\frac {w}{s} \\right\\rceil & \\text {i f} Q _ {n} < \\frac {w}{s} < Q _ {p} \\\\ Q _ {n} & \\text {i f} \\frac {w}{s} \\leq Q _ {n} \\\\ Q _ {p} & \\text {i f} \\frac {w}{s} \\geq Q _ {p} \\end{array} . \\right. \\tag {3}", + "image_path": "4075e837b19c3bfa60d968a35ac4c8627439a67ec975c1ed6f7ae81b81f4d517.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 241, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 241, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 241, + 582 + ], + "type": "text", + "content": "3.2. Limited Solution Space of Weights" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Neural network quantization can be formulated as a mixed-integer nonlinear programming (MINLP) problem. It involves converting floating-point weights and activations into a finite set of discrete values while minimizing performance degradation. In this formulation, quantized weights and activations act as integer variables, while the nonlinearity stems from forward propagation and activation functions. Continuous variables, such as quantization parameters, are also involved. The optimization objective is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 699, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 699, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 96, + 699, + 295, + 716 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {w}} \\mathbb {E} _ {(x, y) \\sim D} \\left[ \\mathcal {L} \\left(f _ {\\text {q u a n t}} (x; q (\\mathbf {w})), y\\right) \\right], \\tag {4}", + "image_path": "75268f225fbea38b8e5f46c7464b8be7ea704863dd568f2b466bc3a02eddeec4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " is the loss function, " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " is the set of all weights, " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "(x,y)\\sim D" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " represents samples drawn from the dataset " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{quant}}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " represents the model's output with activation quantization, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "q(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " is the quantization function. " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "q(\\mathbf{w})" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " consists of integer variables " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and continuous parameters " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "s,z" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 318, + 140, + 555, + 156 + ], + "type": "interline_equation", + "content": "\\forall \\mathbf {Q} \\in [ 0, 2 ^ {b} - 1 ], \\mathbf {Q} \\in \\mathbb {Z}, b \\in \\mathbb {N} _ {*}, q (\\mathbf {w}) = s \\cdot (\\mathbf {Q} - z), \\tag {5}", + "image_path": "87fdc2f6170b4f6a5f97b8a0f8cf2235efb61bfbdcf22e8830c06b0635b4a134.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": " is the set of integers mapped from " + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": " is the scaling factor, " + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": " is the zero point, and " + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 163, + 553, + 186 + ], + "type": "text", + "content": " is the bit-width." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 187, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 554, + 342 + ], + "type": "text", + "content": "QAT-based methods serve as heuristics for solving this MINLP problem. These methods select an initial solution and iteratively refine it. Drawing inspiration from research on mixed-integer programming [6, 47], we observe that QAT shares similarities with depth-first search (DFS) in branch-and-bound algorithms. Both rely on branching and backtracking: QAT chooses quantization strategies per weight (branching), and uses gradient updates (implicit backtracking) to adjust solutions. However, unlike exact DFS, which systematically explores the entire solution space, QAT heuristics prioritize fast convergence, often leading to limited exploration and suboptimal results. We compare QAT and DFS from two key perspectives:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 403 + ], + "type": "text", + "content": "- Branching of variables In DFS, the algorithm explicitly branches on relaxed variables to construct the search tree. Likewise, in QAT, weights oscillating near quantization thresholds effectively induce branching behavior, determining whether a weight falls into one bin or another." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 403, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 474 + ], + "type": "text", + "content": "- Backtracking mechanism DFS performs explicit backtracking: when a branch leads to no improvement, it reverts and explores alternatives, enabling a global search. In contrast, QAT relies on local gradients to implicitly adjust weights. This often causes the model to converge prematurely and miss better configurations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 569 + ], + "type": "text", + "content": "In summary, both DFS and QAT can be regarded as rule-based search strategies. However, under low-bit constraints, QAT's gradient-based updates are more prone to becoming trapped in local optima due to the steep curvature of the loss surface [37]. Motivated by this limitation, we aim to enhance QAT by expanding its effective solution space, similar in spirit to the global search behavior exhibited by exact DFS, in order to achieve improved performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We assess the ability of existing methods to explore the weight solution space by measuring how often weights cross quantization thresholds. As shown in Table 1, we use LSQ [12] to quantize ResNet-18 on the ImageNet [9]. \"SGD\" denotes an SGD optimizer with an initial learning rate of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{e} - 2" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " and cosine learning rate decay [34], while \"Adam\" denotes an Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{e} - 3" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", both following the training strategies in [12]. We find that the frequency of weights crossing quantization thresholds is very low, all below " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", with most untruncated weights concentrated in the " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "[0,0.01\\%)" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " range. This indicates that previous gradient-based optimization methods have a" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24617" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 78, + 216, + 212 + ], + "blocks": [ + { + "bbox": [ + 62, + 78, + 216, + 212 + ], + "lines": [ + { + "bbox": [ + 62, + 78, + 216, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 78, + 216, + 212 + ], + "type": "image", + "image_path": "b2032bd9f14ac44bfa9db6a342421f95bcc094f4e0faaa95608cd76bf6e85bbc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 220, + 143, + 228 + ], + "lines": [ + { + "bbox": [ + 132, + 220, + 143, + 228 + ], + "spans": [ + { + "bbox": [ + 132, + 220, + 143, + 228 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 78, + 383, + 212 + ], + "blocks": [ + { + "bbox": [ + 227, + 78, + 383, + 212 + ], + "lines": [ + { + "bbox": [ + 227, + 78, + 383, + 212 + ], + "spans": [ + { + "bbox": [ + 227, + 78, + 383, + 212 + ], + "type": "image", + "image_path": "b1debe54953a29a64dfce8ac3bca93787021eb555642c9246736f47024319937.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 219, + 312, + 228 + ], + "lines": [ + { + "bbox": [ + 299, + 219, + 312, + 228 + ], + "spans": [ + { + "bbox": [ + 299, + 219, + 312, + 228 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 555, + 262 + ], + "type": "text", + "content": "Figure 2. Histograms of pre-trained weights mapped to different quantization values: (a) lossless solution; (b) LSQ; and (c) our AOQ. The red, green, blue, and yellow bars represent histograms of pre-trained weights mapped to the quantized values -2, -1, 0, and 1, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 399, + 78, + 552, + 212 + ], + "blocks": [ + { + "bbox": [ + 399, + 78, + 552, + 212 + ], + "lines": [ + { + "bbox": [ + 399, + 78, + 552, + 212 + ], + "spans": [ + { + "bbox": [ + 399, + 78, + 552, + 212 + ], + "type": "image", + "image_path": "2500b8fb35bf5f9d3a718943f21fa2750211f5d62f31b9404771397013e2e587.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 219, + 479, + 228 + ], + "lines": [ + { + "bbox": [ + 468, + 219, + 479, + 228 + ], + "spans": [ + { + "bbox": [ + 468, + 219, + 479, + 228 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 75, + 388, + 274, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "text", + "content": "Table 1. Measuring the solution space of weights on 2-bit quantized ResNet-18. The first column is the frequency of weights crossing the quantization thresholds, calculated as the number of crossings divided by the total iterations. \"Truncated\" refers to weights whose frequency of crossing quantization thresholds falls between 0 and " + }, + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "inline_equation", + "content": "0.01\\%" + }, + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "text", + "content": " due to being clipped by the truncation function. \"Not Truncated\" refers to not being clipped by the truncation function. " + }, + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 55, + 280, + 296, + 380 + ], + "type": "text", + "content": " of weights\" refers to the percentage of weights within a specific frequency range." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 388, + 274, + 491 + ], + "lines": [ + { + "bbox": [ + 75, + 388, + 274, + 491 + ], + "spans": [ + { + "bbox": [ + 75, + 388, + 274, + 491 + ], + "type": "table", + "html": "
Frequency Range% of weights
SGDAdam
[0, 0.01%) - Truncated14.5%11.8%
[0, 0.01%) - Not Truncated47.2%12.6%
[0.01%, 0.05%)5.1%19.1%
[0.05%, 0.1%)8.4%24.5%
[0.1%, 0.5%)24.8%32.0%
", + "image_path": "edcc9962ea0f8d79bf1169b6d79d2c8a4f8c85148312e64ed8c9ee86fe7688fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "weak ability to explore the solution space, potentially hindering the model from finding an optimal solution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "To further illustrate how limited exploration affects quantization quality, we construct a toy example. We first generate an integer set " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\bar{w} \\in \\mathbb{R}^{10000}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " containing only values -2, -1, 0, and 1. We then derive matrix " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{2000 \\times 10000}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " and vector " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^{2000}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " satisfying " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\bar{w} = \\mathbf{b}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", establishing " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " as our known lossless integer solution. Clearly, " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " is not a full-rank matrix, which indicates that the linear equation system has multiple solutions, similar to the multi-solution nature of neural networks. Solving " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{A}w = \\mathbf{b}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": " with the constraint " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "w \\in \\{-2s + z, -s + z, z, s + z\\}^{10000}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "s, z \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": ", simulates the quantization process. Conceptually, this setup mimics training a fully connected network with 2-bit quantization, 10,000 input units, one output unit, and 2,000 samples, using mean squared error loss and Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "content": "We first train the model with floating-point weights until convergence, obtaining pre-trained weights, similar to the standard QAT procedure. These weights are then mapped to quantized values. In the lossless solution, the scaling factor " + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "content": " is 1 and " + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "content": " is 0. Figure 2a shows the distribution of pre-trained weights mapped to each quantized value under the lossless solution. Each quantization level corresponds to a wide range of pre-trained weights, indicating that allowing distant weights to map to the same quantized value can preserve model quality. In contrast, Figure 2b shows the distribution after freezing " + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "content": " to 1 and " + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 281, + 555, + 485 + ], + "type": "text", + "content": " to 0 and using LSQ [12] for quantization. In this setting, pre-trained weights are mapped only to the nearest quantized values, restricting the solution space. Such methods, by mapping only to nearby values, may result in more mapping errors and does not fully leverage the broader solution space to minimize quantization errors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 493, + 497, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 497, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 497, + 506 + ], + "type": "text", + "content": "3.3. Allowing Oscillation Quantization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "In previous methods, many weights seldom oscillate near quantization thresholds, limiting their ability to switch quantization bins and often resulting in suboptimal solutions. We speculate that in QAT, especially in low-bit settings, weight transitions between different quantization levels make gradient direction unstable. This instability causes weights to fluctuate within a narrow range during training, preventing them from crossing thresholds. While some studies [30, 36] consider weight oscillation harmful to the whole optimization process, we believe that oscillation actually promotes solution diversity, increasing the likelihood of discovering better optima. Specifically, oscillation acts as an adaptive perturbation mechanism that enables the model to continuously test new quantization configurations and explore more quantized solutions within a given number of iterations, helping avoid premature convergence. This phenomenon resembles \"random perturbation\" strategies used" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24618" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 293, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 178 + ], + "type": "image", + "image_path": "f14cd86f4a1f22fc20c3ef8920b81b9f19d6f5b71c3f207fb42164909c344abe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 296, + 233 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 233 + ], + "type": "text", + "content": "Figure 3. Oscillation patterns of two weights in layer1.0.conv1 of ResNet-18 during early training. We apply AOQ under a 2-bit quantization setting and monitor quantization assignments at the end of each epoch." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 255, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 295, + 327 + ], + "type": "text", + "content": "in integer programming [21, 40], where moderate instability prevents the model from becoming locked into overly deterministic paths. Based on this intuition, we propose to actively promote oscillation during early training phases to encourage exploration, then gradually reduce it in later stages to ensure convergence stability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 327, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 553 + ], + "type": "text", + "content": "To increase the number of weights that cross thresholds without disrupting QAT's gradient-based backtracking mechanism, we examine how quantization parameters affect oscillation. Inspired by [30], we observe that modifying quantization thresholds can induce more weights to oscillate around them, as shown in Figure 3. Specifically, when the intervals between quantization thresholds and the differences among quantization levels decrease, weights tend to transition from intra-bin oscillation to threshold-adjacent oscillation. This behavior is driven by two factors. First, the shifting thresholds force certain weights to cross them, introducing noise into the network. The facilitating effect of smaller quantization threshold intervals on weight oscillation will be demonstrated in Section 3.4. Second, the reduced differences between quantization levels lower the penalty for mapping errors when floating-point weights are mapped to quantized values, making weights more likely to oscillate near the thresholds, as smaller differences between quantization levels tend to smooth the loss surface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 698 + ], + "type": "text", + "content": "To leverage the weight oscillation mentioned above, we decouple the quantization thresholds and levels and introduce a three-stage training strategy. Specifically, we manually adjust the quantization thresholds to gradually reduce their intervals, while setting the quantization levels as learnable parameters to optimize during training. Additionally, we introduce differentiability for clipped weights in fake quantization to address the issue observed in Section 3.2, where the absence of gradients in clipped weights (as shown in Eq. (2)) hinders model fine-tuning. In the final stage, oscillation is suppressed to stabilize convergence. The complete strategy is outlined below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- Stage 1: Inducing oscillation. For weight quantization," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": "we define the intervals between thresholds as " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "s_{th}" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " and between quantization levels as " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": ", both initialized to " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\frac{\\sigma_w}{2^{b - 2}}" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\sigma_w" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " is the standard deviation of weights in a layer, and " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " is the bit-width. In this stage, we manually reduce " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "s_{th}" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 321, + 72, + 553, + 132 + ], + "type": "text", + "content": " to promote oscillation at thresholds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "content": "- Stage 2: Learning quantization levels. We fix " + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "inline_equation", + "content": "s_{th}" + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "content": " as a learnable parameter for optimization. In Stage 1, " + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "content": " is manually reduced, which leads to higher loss due to poor representation of weights by the quantization levels. As training progresses and the task loss decreases, " + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 314, + 133, + 555, + 252 + ], + "type": "text", + "content": " is learned to increase automatically. This widens the intervals between quantization levels, makes the loss landscape steeper near thresholds, and reduces oscillation by discouraging weights from frequently crossing thresholds. The quantizer is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 261, + 553, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 322, + 261, + 553, + 315 + ], + "type": "interline_equation", + "content": "\\hat {w} = \\left\\{ \\begin{array}{l l} \\operatorname {l e v e l} [ 0 ] & \\text {i f} w < \\operatorname {t h r e s h} [ 0 ] \\\\ \\operatorname {l e v e l} [ i ] & \\text {i f t h r e s h} [ i - 1 ] \\leq w < \\operatorname {t h r e s h} [ i ] \\\\ \\operatorname {l e v e l} \\left[ 2 ^ {b} - 1 \\right] & \\text {i f} w \\geq \\operatorname {t h r e s h} \\left[ 2 ^ {b} - 2 \\right] \\end{array} \\right. \\tag {6}", + "image_path": "f7d85b04e31569e90fe9b7a44c073f40413ee930143f4f056ffc54681cdc542e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,\\ldots ,2^b -2\\}" + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{thresh}[i]" + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "text", + "content": " and level[i] represent the " + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "inline_equation", + "content": "(\\mathrm{i + 1})" + }, + { + "bbox": [ + 322, + 315, + 553, + 352 + ], + "type": "text", + "content": " -th threshold and level, respectively. The optimized quantized weights are denoted as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 380, + 361, + 553, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 361, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 380, + 361, + 553, + 374 + ], + "type": "interline_equation", + "content": "\\tilde {w} = \\hat {w} + w - w. \\det a c h (\\cdot). \\tag {7}", + "image_path": "01cfe5af3b893cd69f6fd9828d3f1d708ef253b55f4571c5e7b4cdcf564b5eb8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "text", + "content": "This formulation, based on the STE, enables distinct gradient flows: " + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "text", + "content": " determines the gradient for " + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 321, + 383, + 554, + 454 + ], + "type": "text", + "content": " receives gradients directly from the task loss, even when clipped. This decoupling avoids the zero-gradient issue and enables more effective optimization of both weights and quantizer parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 455, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 455, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 455, + 554, + 491 + ], + "type": "text", + "content": "- Stage 3: Dampening oscillation. In the third stage, we employ the Oscillation dampening method outlined in [36], which adds a regularization term to the loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 392, + 501, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 501, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 392, + 501, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda \\mathcal {L} _ {\\text {d a m p e n}}, \\tag {8}", + "image_path": "44731e1f10f7afa26d9e827e6c569fbea2e2ff2b8b87e140877b07374096ff8d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 523, + 537, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 523, + 537, + 534 + ], + "spans": [ + { + "bbox": [ + 322, + 523, + 537, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 523, + 537, + 534 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 322, + 523, + 537, + 534 + ], + "type": "text", + "content": " is a hyperparameter. The dampening loss is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 362, + 543, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 543, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 362, + 543, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d a m p e n}} = \\left\\| \\hat {w} - \\operatorname {c l i p} \\left(w, Q _ {n}, Q _ {p}\\right) \\right\\| _ {F} ^ {2}, \\tag {9}", + "image_path": "17080133e4614b41419033429f446cd0c93abadb3c5ad4f3ce547cd9ce2b72d2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 567, + 553, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 567, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 321, + 567, + 553, + 602 + ], + "type": "text", + "content": "which penalizes weights lingering near quantization thresholds, encouraging them to converge to the nearest quantization centers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 628 + ], + "type": "text", + "content": "Details regarding the hyperparameters and transition schedule for each training stage are provided in Section 4.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 635, + 433, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 433, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 433, + 649 + ], + "type": "text", + "content": "3.4. Theoretical Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "In this section, we theoretically analyze the effectiveness of our approach and compare it against prior differentiable quantization parameter methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We first show the advantage of using smaller quantization threshold intervals. Let " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " be the set of pre-trained" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24619" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "weights, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " be a set composed of quantized values. Weight quantization is a mapping from the " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ". For a given weight " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "w \\in W" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", we define its probability of being quantized to " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "Q_{i} \\in Q" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " as the transition probability " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "P(Q_{i} \\mid w)" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "i \\in \\{0, 1, \\dots, 2^{b} - 1\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " is the bit-width. By Bayes' theorem:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 152, + 295, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 152, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 102, + 152, + 295, + 179 + ], + "type": "interline_equation", + "content": "P \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{P (w)}, \\tag {10}", + "image_path": "98b53017829a1522da64dc8a63b10cd20a07f55b6783921b24a3e7b5beae0f47.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "P(Q_{i})" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " is the proportion of the number of weights mapped to " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " to the total number of weights, " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "P(w)" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " is the probability distribution of weights, and " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "P(w|Q_i)" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": " is the conditional distribution of weights mapped to " + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 188, + 296, + 248 + ], + "type": "text", + "content": ". The total probability formula applied to Eq. (10) is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 256, + 295, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 256, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 94, + 256, + 295, + 283 + ], + "type": "interline_equation", + "content": "P \\left(Q _ {i} \\mid w\\right) = \\frac {P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}{\\sum_ {i} P \\left(Q _ {i}\\right) \\cdot P \\left(w \\mid Q _ {i}\\right)}. \\tag {11}", + "image_path": "8adcd4638d6291fdee4ba3ae120245a9599eb98574ced2bffa4e966528e1fc05.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 292, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 327 + ], + "type": "text", + "content": "As shown in Figure 2a, " + }, + { + "bbox": [ + 55, + 292, + 296, + 327 + ], + "type": "inline_equation", + "content": "P(w \\mid Q_i)" + }, + { + "bbox": [ + 55, + 292, + 296, + 327 + ], + "type": "text", + "content": " approximately follows a Gaussian distribution, so the total probability formula can be transformed into a Gaussian mixture model:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 335, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 335, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 112, + 335, + 295, + 360 + ], + "type": "interline_equation", + "content": "P (w) = \\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i}, \\sigma_ {i}\\right), \\tag {12}", + "image_path": "f1911abed27ac4ebcecc3471d2074b3069baf8c2316d2b3dd7a5989c3ab540b4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\pi_i, \\mu_i, \\sigma_i" + }, + { + "bbox": [ + 55, + 369, + 296, + 418 + ], + "type": "text", + "content": " denote the weight, mean and variance of the Gaussian component, respectively. This transformation does not effect the theoretical conclusion, just for analytical convenience. Eq. (11) is transformed into:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 426, + 295, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 426, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 101, + 426, + 295, + 453 + ], + "type": "interline_equation", + "content": "P \\left(Q _ {i} \\mid w\\right) = \\frac {\\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}{\\sum_ {i} \\pi_ {i} \\mathcal {N} \\left(w \\mid \\mu_ {i} , \\sigma_ {i}\\right)}. \\tag {13}", + "image_path": "5328fd42421a615a0cc7fcd9edaccc57037d305fb27341cbfe788bc6a34dd77a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "content": "The weight " + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "content": " is assigned to the quantization value " + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "inline_equation", + "content": "Q_{j}" + }, + { + "bbox": [ + 55, + 460, + 296, + 485 + ], + "type": "text", + "content": " that maximizes this probability:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 494, + 295, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 494, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 78, + 494, + 295, + 512 + ], + "type": "interline_equation", + "content": "i f \\forall i, \\pi_ {j} \\mathcal {N} (w \\mid \\mu_ {j}, \\sigma_ {j}) \\geq \\pi_ {i} \\mathcal {N} (w \\mid \\mu_ {i}, \\sigma_ {i}) \\tag {14}", + "image_path": "af590599238f37ee70da78608a94551e8ddeae7a885438bda6717dac1751bc71.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 510, + 141, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 510, + 141, + 522 + ], + "spans": [ + { + "bbox": [ + 79, + 510, + 141, + 522 + ], + "type": "text", + "content": "then " + }, + { + "bbox": [ + 79, + 510, + 141, + 522 + ], + "type": "inline_equation", + "content": "w\\to Q_j" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 531, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 627 + ], + "type": "text", + "content": "Therefore, when the quantization threshold is set at the horizontal coordinate value of the intersection between two adjacent Gaussian components, it ensures optimal probabilistic mapping of weights. In Figure 2a, we observe that in the lossless solution for this example, the variances of each Gaussian component are approximately equal, denoted as " + }, + { + "bbox": [ + 55, + 531, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 531, + 296, + 627 + ], + "type": "text", + "content": ". Thus, the horizontal coordinate value at the intersection of adjacent Gaussian components is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 79, + 635, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 635, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 79, + 635, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\operatorname {t h r e s h} [ i ] = \\frac {\\mu_ {i + 1} + \\mu_ {i}}{2} - \\frac {\\log \\left(\\frac {\\pi_ {i + 1}}{\\pi_ {i}}\\right) \\cdot \\sigma^ {2}}{\\mu_ {i + 1} - \\mu_ {i}}. \\tag {15}", + "image_path": "a9d466a53b1af3289b17f2a258d33cfda9fb9139763643d12b1c9a6c5b03fa60.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "This analysis suggests that, when Gaussian component means are close, using smaller threshold intervals improves mapping accuracy. It explains why our method encourages" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "tighter thresholds in early training to support more effective weight assignment in QAT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 96, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 262 + ], + "type": "text", + "content": "Next, we show that AOQ provides a more direct and accurate link between quantization levels and task loss than previous approaches. In prior methods [5, 12, 49], learnable quantization parameters jointly determine both thresholds and levels. This coupling introduces approximation errors in the gradient computation of quantization levels. For instance, in LSQ [12], the gradients of the scaling factors are shown in Eq. (3). The derivative of weights with respect to the scaling factor aids in optimizing quantization thresholds but makes the gradient of the quantization levels inaccurate. In contrast, AOQ decouples thresholds from levels. It directly optimizes quantization levels based on Eq. (6) and Eq. (7), allowing a more faithful gradient flow that better aligns with task loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 263, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 372 + ], + "type": "text", + "content": "Finally, AOQ effectively addresses the impact of weight oscillation on final convergence, as well as on batch normalization and layer normalization. After exploring a broader solution space, we use Oscillation Dampening in the final stage to gradually guide weights out of oscillatory regions. As training continues, the exponential moving average used in batch normalization and layer normalization becomes increasingly stable, reinforcing consistent model behavior and improving convergence quality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 381, + 395, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 395, + 396 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 395, + 396 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 402, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 475 + ], + "type": "text", + "content": "We evaluate the effectiveness of the proposed Allowing Oscillation Quantization (AOQ) on the ILSVRC12 ImageNet dataset [9], using a range of models including ResNet [15], MobileNet-V2 [17], DeiT [41], and Swin Transformer [33]. Our experiments focus on low-bit weight quantization. For activation quantization, we adopt LSQ [12] and N2UQ [32]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 480, + 430, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 430, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 430, + 493 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "text", + "content": "ResNet and MobileNet We follow the training schemes in [5, 12, 32], initializing quantized networks from full-precision pre-trained models. Training is performed using the Adam optimizer [20] with a linear learning rate decay, starting from " + }, + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 498, + 554, + 594 + ], + "type": "text", + "content": ". Following LSQ [12], we apply knowledge distillation and train for 150 epochs. All convolutional and fully connected layers are quantized, except for the first and last layers." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 654 + ], + "type": "text", + "content": "DeiT and Swin Transformer We use the official implementations and adopt training settings from [28-30], including knowledge distillation and a training duration of 300 epochs. As in prior work, the patch embedding and classification layers are quantized to 8 bits [28-30]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Parameter Setup The batch size is set to 512, and the weight decay is set to 0. The three training stages occupy " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "1/5" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "2/5" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "2/5" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " of the total training epochs, respectively. We reduce " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{le}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "30\\% - 35\\%" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " of their initial values in the first stage. As shown in Fig. 4, a broad grid search" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24620" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 101, + 524, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Accuracy comparison to the state-of-the-art quantization methods with ResNet structure on ImageNet dataset. Note that W/A denote the bit-widths of weights and activations, respectively. FP denotes the top-1 accuracy of the full-precision models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 101, + 524, + 404 + ], + "lines": [ + { + "bbox": [ + 86, + 101, + 524, + 404 + ], + "spans": [ + { + "bbox": [ + 86, + 101, + 524, + 404 + ], + "type": "table", + "html": "
NetworkMethodBit-width (W/A)Accuracy (%) Top-1Bit-width (W/A)Accuracy(%) Top-1Bit-width (W/A)Accuracy(%) Top-1
ResNet-18 (FP: 71.8)LQ-Nets [46]2/264.93/368.24/469.3
DSQ [13]2/265.23/368.74/469.6
QIL [19]2/265.73/369.24/470.1
APoT [27]2/267.33/369.94/470.7
LSQ [12]2/267.63/370.24/471.1
LCQ [43]2/268.93/370.64/471.5
N2UQ [32]2/269.43/371.94/472.9
LSQ+AOQ2/269.73/371.44/472.0
N2UQ+AOQ2/270.33/372.14/472.7
ResNet-34 (FP:74.9)LQ-Nets [46]2/269.83/371.94/4-
DSQ [13]2/270.03/372.54/472.8
QIL [19]2/270.63/373.14/473.7
APoT [27]2/270.93/373.44/473.8
LSQ [12]2/271.63/373.44/474.1
LCQ [43]2/272.73/374.04/474.3
N2UQ [32]2/273.33/375.24/476.0
LSQ+AOQ2/273.53/374.94/475.5
N2UQ+AOQ2/273.93/375.54/476.0
ResNet-50 (FP: 77.0)LQ-Nets [46]2/271.53/374.24/475.1
APoT [27]2/273.43/375.84/476.6
LSQ [12]2/273.73/375.84/476.7
LCQ [43]2/275.13/376.34/476.6
N2UQ [32]2/275.83/377.54/478.0
LSQ+AOQ2/276.33/377.54/477.9
N2UQ+AOQ2/276.43/377.84/478.1
", + "image_path": "42e30db4e0e4a1713efc32754cb24c6d3cb49dc78904fb680b9a3bb0dd68681c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 429, + 173, + 513 + ], + "blocks": [ + { + "bbox": [ + 58, + 429, + 173, + 513 + ], + "lines": [ + { + "bbox": [ + 58, + 429, + 173, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 429, + 173, + 513 + ], + "type": "image", + "image_path": "83a4b8f167de8458948e7fdcaed756a05136101bab787c9e1e07d387073d366e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 517, + 139, + 525 + ], + "lines": [ + { + "bbox": [ + 93, + 517, + 139, + 525 + ], + "spans": [ + { + "bbox": [ + 93, + 517, + 139, + 525 + ], + "type": "text", + "content": "(a) ResNet-18" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 429, + 294, + 514 + ], + "blocks": [ + { + "bbox": [ + 173, + 429, + 294, + 514 + ], + "lines": [ + { + "bbox": [ + 173, + 429, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 173, + 429, + 294, + 514 + ], + "type": "image", + "image_path": "b4fd1ea6ad70f8a2c172724c4ffdeb0ed1df4dfdac76f34074340a01ad5201a9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 517, + 254, + 525 + ], + "lines": [ + { + "bbox": [ + 218, + 517, + 254, + 525 + ], + "spans": [ + { + "bbox": [ + 218, + 517, + 254, + 525 + ], + "type": "text", + "content": "(b) DeiT-T" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": "Figure 4. Accuracy heatmaps on ImageNet under different narrowing schedules. " + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": " denotes the final-to-initial ratio of " + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "inline_equation", + "content": "s_{le}" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "inline_equation", + "content": "s_{th}" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": " in stage one, and " + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": " is the proportion of total epochs allocated to this stage. AOQ maintains stable performance across a wide range of " + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "inline_equation", + "content": "(\\alpha, \\beta)" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": " settings on both (a) 2-bit ResNet-18 and (b) 2-bit DeiT-T, indicating schedule robustness across architectures." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": "confirms that a common schedule performs robustly across different architectures in practice. For the final stage, the hyper-parameter " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " is chosen such that " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\lambda \\mathcal{L}_{\\mathrm{dampen}}" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": " is two to three orders of magnitude lower than " + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 55, + 617, + 296, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 258, + 685 + ], + "type": "text", + "content": "4.2. Experimental Results on Toy Example" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We first revisit the toy example from Section 3.2 to evaluate the effect of oscillation on quantization quality. To" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 433, + 436, + 515 + ], + "blocks": [ + { + "bbox": [ + 328, + 433, + 436, + 515 + ], + "lines": [ + { + "bbox": [ + 328, + 433, + 436, + 515 + ], + "spans": [ + { + "bbox": [ + 328, + 433, + 436, + 515 + ], + "type": "image", + "image_path": "78e094382a8c8fb6520420029da6247389ac8e95ae55767ad42c1e963d380ba0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 525, + 555, + 547 + ], + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 547 + ], + "type": "text", + "content": "Figure 5. Loss curves and minimum loss obtained from solving the toy example using various methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 440, + 438, + 547, + 507 + ], + "blocks": [ + { + "bbox": [ + 440, + 438, + 547, + 507 + ], + "lines": [ + { + "bbox": [ + 440, + 438, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 440, + 438, + 547, + 507 + ], + "type": "table", + "html": "
MethodLoss
LSQ51.2
LSQ+OD(final)16.7
LSQ+OD(enti)26.2
AOQ5.6
", + "image_path": "5b216ad1fa8d5e32ef0a8286d72767051cc4043c71ff926a8a6db47c0352450f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "demonstrate that weight oscillation can expand the solution space and improve quantization outcomes, we apply the Oscillation Dampening (OD) [36] to suppress oscillation either during the final training stage or throughout the entire training process. We also evaluate the performance of AOQ using the toy example. As shown in Figure 5, allowing early-stage oscillation improves QAT performance, and AOQ converges faster to a better solution. When using fixed quantization levels " + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{-2, -1, 0, 1\\}" + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": " in the third stage, Figure 2c shows that AOQ enables pre-trained weights to be mapped to more diverse and distant quantized values, expanding the solution space compared to Figure 2b." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24621" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 64, + 102, + 287, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "type": "text", + "content": "Table 3. Top-1 accuracy (\\%) comparison for methods suppressing weight oscillation throughout the process in MobileNet-V2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 102, + 287, + 198 + ], + "lines": [ + { + "bbox": [ + 64, + 102, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 287, + 198 + ], + "type": "table", + "html": "
NetworkMethodW3A3W4A4
MobileNet-V2 (FP: 72.1)LSQ [12]65.369.5
LSQ+BR [14]67.470.4
LSQ+Dampen [36]67.870.5
N2UQ [32]69.072.1
LSQ+AOQ69.171.8
N2UQ+AOQ69.972.5
", + "image_path": "8f1235511c25efbed4c1c06f165bf1ea4d72d262f58a3a8e2ac3bc35defe3772.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 62, + 240, + 289, + 422 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 232 + ], + "type": "text", + "content": "Table 4. Top-1 accuracy (\\%) comparison with the state-of-the-art methods in the ViT architecture." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 240, + 289, + 422 + ], + "lines": [ + { + "bbox": [ + 62, + 240, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 289, + 422 + ], + "type": "table", + "html": "
NetworkMethodW2A2W3A3
DeiT-T (FP: 72.1)LSQ [12]54.568.1
QViT* [28]50.467.1
OFQ [30]64.372.7
LSQ+AOQ66.572.3
DeiT-S (FP: 79.9)LSQ [12]68.077.8
QViT* [28]68.778.5
OFQ [30]75.779.6
LSQ+AOQ76.779.8
Swin-T (FP: 81.2)LSQ [12]70.479.0
QViT* [28]73.980.1
OFQ [30]78.581.1
LSQ+AOQ78.981.2
", + "image_path": "f70394efd0c25523d87cef901ae77f84170a43045d659c3ff276c82358987193.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 178, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 178, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 178, + 455 + ], + "type": "text", + "content": "4.3. Accuracy Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "Table 2 reports the accuracy of AOQ on ResNet. When paired with different activation quantization techniques, AOQ consistently improves performance. For instance, using AOQ with LSQ yields " + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "0.9\\% - 2.6\\%" + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": " accuracy gains. When combined with N2UQ, AOQ outperforms many traditional methods. Its advantage becomes more pronounced in low-bit networks with a large number of weights, where the expanded solution space is more effectively utilized. In the 2-bit ResNet-50 case, AOQ reduces the accuracy gap with full-precision to just " + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 54, + 460, + 295, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "Table 3 shows the performance of the AOQ method on MobileNet-V2. Compared to other methods, AOQ has a natural advantage in quantizing MobileNet. Due to significant variations in weight distributions across channels in the depth-wise separable convolution layers, traditional quantization methods often map all weights in certain channels to a single quantized value, resulting in the loss of important information. AOQ mitigates this by adaptively shrinking threshold intervals during training, promoting threshold-crossing behavior and improving quantization granularity. This effect is illustrated in Figure 3. Overall, AOQ outper" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 330, + 102, + 538, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 5. Effects of different components of our strategy on the final performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 102, + 538, + 198 + ], + "lines": [ + { + "bbox": [ + 330, + 102, + 538, + 198 + ], + "spans": [ + { + "bbox": [ + 330, + 102, + 538, + 198 + ], + "type": "table", + "html": "
MethodAcc (%)
Baseline [12]73.7
- Weight quantization76.6
+ Expanding solution space of weights75.2
+ Delayed weight oscillation suppression75.6
+ Both76.3
Corresponding real-valued network77.0
", + "image_path": "98f6b5a351f9f2c0dc7a2d7101355c22621277de1ac14fbac46e37a25c8bcc51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 240 + ], + "type": "text", + "content": "forms other QAT approaches, including those that suppress oscillation throughout [14, 36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "text", + "content": "To validate AOQ on Vision Transformers (ViT), we conducted 2-bit and 3-bit quantization experiments on DeiT and Swin Transformer. We observe that QViT [28] does not fully quantize activations in the fully connected and attention layers, limiting hardware acceleration. For fair comparison, we modify their implementation accordingly and report the results as QViT* in Table 4. The results show that AOQ significantly outperforms LSQ and QViT by suppressing weight oscillation, boosting 2-bit DeiT-T accuracy by " + }, + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "inline_equation", + "content": "12.0\\%" + }, + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "inline_equation", + "content": "16.1\\%" + }, + { + "bbox": [ + 313, + 240, + 555, + 384 + ], + "type": "text", + "content": ", respectively. By expanding the solution space, AOQ also surpasses OFQ, the prior state-of-the-art, under 2-bit quantization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 390, + 408, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 408, + 403 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "content": "To investigate the impact of each improvement on the final performance, we conduct an ablation study on the 2-bit quantized ResNet-50 network on ImageNet. As shown in Table 5, the proposed expanded solution space and delayed weight oscillation suppression improve the accuracy by " + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "content": ", respectively over the baseline [12]. By using both techniques, we reduced the accuracy gap with the corresponding real-valued ResNet-50 to only " + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "content": ", and the accuracy is only reduced by " + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 407, + 556, + 529 + ], + "type": "text", + "content": " compared to the baseline method without weight quantization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 715 + ], + "type": "text", + "content": "In this work, we reveal the weakness of existing QAT methods in exploring the quantization solution space of weights within low bit-width constraints, which adversely impacts the performance of quantized networks. We highlight the importance of weight oscillation during early training, as it helps the optimizer escape local minima and explore a broader set of quantized configurations. Based on this insight, we propose a simple yet effective method that encourages weight oscillation in the early stage to expand the solution space, followed by oscillation suppression to stabilize convergence. Experiments demonstrate that our algorithm quantizes various models to low-bits and achieves state-of-the-art accuracy on ImageNet." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24622" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under Grant 62322117, 62425109, 62371365, U24B20136, and U22B2014." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Haoli Bai, Wei Zhang, Lu Hou, Lifeng Shang, Jing Jin, Xin Jiang, Qun Liu, Michael R. Lyu, and Irwin King. Binarybert: Pushing the limit of bert quantization. In Annual Meeting of the Association for Computational Linguistics, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "text", + "content": "[2] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 296, + 300 + ], + "type": "text", + "content": "[3] Hanlin Chen, Li'an Zhuo, Baochang Zhang, Xiawu Zheng, Jianzhuang Liu, Rongrong Ji, David S. Doermann, and Guodong Guo. Binarized neural architecture search for efficient object recognition. International Journal of Computer Vision, 129:501-516, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 301, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 296, + 345 + ], + "type": "text", + "content": "[4] Shangyu Chen, Wenya Wang, and Sinno Jialin Pan. Metaquant: Learning to quantize by learning to penetrate non-differentiable quantization. In Neural Information Processing Systems, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 400 + ], + "type": "text", + "content": "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 435 + ], + "type": "text", + "content": "[6] Michele Conforti, Gérard Cornuéjols, Giacomo Zambelli, Michele Conforti, Gérard Cornuéjols, and Giacomo Zambelli. Integer programming models. Springer, 2014. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "text", + "content": "[7] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. In Neural Information Processing Systems, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 480, + 296, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 522 + ], + "type": "text", + "content": "[8] Alexandre Défossez, Yossi Adi, and Gabriel Synnaeve. Differentiable model compression via pseudo quantization noise. Transactions on Machine Learning Research, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 524, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 296, + 568 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 248-255. IEEE, 2009. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 296, + 612 + ], + "type": "text", + "content": "[10] Tim Dettmers, Mike Lewis, Younes Belkada, and Luke Zettlemoyer. Gpt3. int8(): 8-bit matrix multiplication for transformers at scale. In Neural Information Processing Systems, pages 30318-30332, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 646 + ], + "type": "text", + "content": "[11] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Ruihao Gong, Xianglong Liu, Shenghu Jiang, Tian-Hao Li, Peng Hu, Jiazhen Lin, Fengwei Yu, and Junjie Yan. Differen" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "tiable soft quantization: Bridging full-precision and low-bit neural networks. In IEEE/CVF International Conference on Computer Vision, pages 4851-4860, 2019. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[14] Tiantian Han, Dong Li, Ji Liu, Lu Tian, and Yi Shan. Improving low-precision network quantization via bin regularization. In IEEE/CVF International Conference on Computer Vision, pages 5241–5250, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 194 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 237 + ], + "type": "text", + "content": "[16] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Ptqd: Accurate post-training quantization for diffusion models. In Neural Information Processing Systems, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "text", + "content": "[17] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 294, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 360 + ], + "type": "text", + "content": "[18] Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew G. Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2704-2713, 2017. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 426 + ], + "type": "text", + "content": "[19] San Hun Jung, Changyong Son, Seohyung Lee, JinWoo Son, Jae-Joon Han, Youngjun Kwak, Sung Ju Hwang, and Changkyu Choi. Learning to quantize deep networks by optimizing quantization intervals with task loss. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4345-4354, 2018. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 554, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 447 + ], + "type": "text", + "content": "[20] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "text", + "content": "[21] Scott Kirkpatrick, Charles D. Gelatt, and Mario P. Vecchi. Optimization by simulated annealing. Science, 220:671-680, 1983. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 514 + ], + "type": "text", + "content": "[22] Yann LeCun, Léon Bottou, Genevieve B Orr, and Klaus-Robert Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9-50. Springer, 2002. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 554, + 558 + ], + "type": "text", + "content": "[23] Junghyup Lee, Dohyung Kim, and Bumsub Ham. Network quantization with element-wise gradient scaling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6444-6453, 2021. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "type": "text", + "content": "[24] Jung Hyun Lee, Jeonghoon Kim, Se Jung Kwon, and Dongsoo Lee. Flexround: Learnable rounding based on elementwise division for post-training quantization. In International Conference on Machine Learning, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[25] Cong Leng, Hao Li, Shenghuo Zhu, and Rong Jin. Extremely low bit neural network: Squeeze the last bit out with admm. In AAAI Conference on Artificial Intelligence, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "type": "text", + "content": "[26] Xinlin Li, Bang Liu, Ruizhi Yang, Vanessa Courville, Chao Xing, and V. Nia. Denseshift : Towards accurate and efficient low-bit power-of-two quantization. In IEEE/CVF International Conference on Computer Vision, pages 16964-16974, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[27] Yuhang Li, Xin Dong, and Wei Wang. Additive powers-of-two quantization: An efficient non-uniform discretization for" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24623" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "neural networks. In International Conference on Learning Representations, 2019. 1, 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[28] Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-vit: Accurate and fully quantized low-bit vision transformer. In Neural Information Processing Systems, pages 34451-34463, 2022. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[29] Zhexin Li, Tong Yang, Peisong Wang, and Jian Cheng. Qvit: Fully differentiable quantization for vision transformer. arXiv preprint arXiv:2201.07703, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[30] Shi Liu, Zechun Liu, and Kwang-Ting Cheng. Oscillation-free quantization for low-bit vision transformers. In International Conference on Machine Learning, 2023. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 272 + ], + "type": "text", + "content": "[31] Zechun Liu, Baoyuan Wu, Wenhan Luo, Xin Yang, Wei Liu, and Kwang-Ting Cheng. Bi-real net: Enhancing the performance of 1-bit cnns with improved representational capability and advanced training algorithm. In European Conference on Computer Vision, pages 722-737, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 337 + ], + "type": "text", + "content": "[32] Zechun Liu, Kwang-Ting Cheng, Dong Huang, Eric P. Xing, and Zhiqiang Shen. Nonuniform-to-uniform quantization: Towards accurate quantization via generalized straight-through estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4932-4942, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 392 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 9992-10002, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 426 + ], + "type": "text", + "content": "[34] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 480 + ], + "type": "text", + "content": "[35] Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197-7206. PMLR, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 295, + 525 + ], + "type": "text", + "content": "[36] Markus Nagel, Marios Fournarakis, Yelysei Bondarenko, and Tijmen Blankevoort. Overcoming oscillations in quantization-aware training. In International Conference on Machine Learning, 2022. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 569 + ], + "type": "text", + "content": "[37] Yury Nahshan, Brian Chmiel, Chaim Baskin, Evgenii Zheltonozhskii, Ron Banner, Alexander M. Bronstein, and Avi Mendelson. Loss aware post-training quantization. Machine Learning, 110:3245 - 3262, 2019. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[38] Juncheol Shin, Junhyuk So, Sein Park, Seungyeop Kang, Sungjoo Yoo, and Eunhyeok Park. Nipq: Noise proxy-based integrated pseudo-quantization. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3852–3861, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 669 + ], + "type": "text", + "content": "[39] Zhenhong Sun, Ce Ge, Junyan Wang, Ming Lin, Hesen Chen, Hao Li, and Xiuyu Sun. Entropy-driven mixed-precision quantization for deep network design. In Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[40] Charles E. Taylor. Adaptation in natural and artificial systems: An introductory analysis with applications to biology, control, and artificial intelligence. The Quarterly Review of Biology, 69:88-89, 1994. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 520 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 126 + ], + "type": "text", + "content": "[41] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[42] Ziwei Wang, Ziyi Wu, Jiwen Lu, and Jie Zhou. Bidet: An efficient binarized object detector. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2046-2055, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "text", + "content": "[43] Kohei Yamamoto. Learnable companding quantization for accurate low-bit neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5027-5036, 2021. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[44] Jiwei Yang, Xu Shen, Jun Xing, Xinmei Tian, Houqiang Li, Bing Deng, Jianqiang Huang, and Xian-sheng Hua. Quantization networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7308-7316, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 317 + ], + "type": "text", + "content": "[45] Penghang Yin, Jiancheng Lyu, Shuai Zhang, Stanley Osher, Yingyong Qi, and Jack Xin. Understanding straight-through estimator in training activation quantized neural nets. In International Conference on Learning Representations, 2019. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[46] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In European Conference on Computer Vision, 2018. 1, 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 407 + ], + "type": "text", + "content": "[47] Jiayi Zhang, Chang Liu, Junchi Yan, Xijun Li, Hui-Ling Zhen, and Min jie Yuan. A survey for solving mixed integer programming via machine learning. Neurocomputing, 519:205-217, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 408, + 553, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 474 + ], + "type": "text", + "content": "[48] Yifan Zhang, Zhen Dong, Huanrui Yang, Ming Lu, Cheng-Ching Tseng, Yuan Du, Kurt Keutzer, Li Du, and Shanghang Zhang. Qd-bev: Quantization-aware view-guided distillation for multi-view 3d object detection. In IEEE/CVF International Conference on Computer Visio, pages 3802-3812, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 520 + ], + "type": "text", + "content": "[49] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1, 2, 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24624" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_content_list.json b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6dd0064a7d782fa25c34ec976dcc2eb55f98e18d --- /dev/null +++ b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "Always Skip Attention", + "text_level": 1, + "bbox": [ + 382, + 130, + 612, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yiping Ji $^{1,2}$ , Hemanth Saratchandran $^{1}$ , Peyman Moghadam $^{2,3}$ , Simon Lucey $^{1}$ $^{1}$ Adelaide University, $^{2}$ CSIRO, $^{3}$ Queensland University of Technology", + "bbox": [ + 194, + 179, + 799, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yiping.ji, hemanth.saratchandran, simon.lucy}@adelaide.edu.au, {yiping.ji, peyman.moghadam}@csiro.au", + "bbox": [ + 218, + 219, + 774, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d557aff84ed6bb07db54af823bbea325dbabb942af77bd77072f197d35abd6e9.jpg", + "image_caption": [ + "(a) Architecture" + ], + "image_footnote": [], + "bbox": [ + 94, + 263, + 205, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/79fcdb7e27381b66ffa8d51cf6358191763ed54e29d4133d72f97258dcdf90d2.jpg", + "image_caption": [ + "(b) FFN and SAB output embedding condition number using different configurations" + ], + "image_footnote": [], + "bbox": [ + 207, + 272, + 555, + 482 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d73e83f7d50976ffadb9c8aeb611c2f79b43596064c42de14eba1028f5e5c199.jpg", + "image_caption": [ + "(c) Performance after removing skip connections", + "Figure 1. Removing skip connections from the Self-Attention Block (SAB) versus the Feedforward Network (FFN) in Vision Transformer (ViT) models results in different performance drop. (a) Model architecture of ViT consisting of a SAB, and a FFN in each layer. $①$ represents the output of the self-attention transformation, while $③$ represents the output of the MLP transformation. $②$ and $④$ represent identity mappings, referred to as skip connections, in the SAB and FFN, respectively. (b) The condition number of the different output embeddings for each hidden layer. A ViT-Tiny model is trained in a typical manner with skip connections on both SAB and FFN blocks (i.e., $① + ② + ③ + ④$ ). It shows that the SAB output embedding without skip connection is poorly conditioned $(①)$ , with a condition number significantly higher than that of the other three configurations involving ViT blocks. (c) Classification results on three different datasets using the ViT-Tiny model under three different architectures at train time. The result shows the effects of removing skip connections at train time on the SAB and FFN blocks (either $②$ or $④$ ) in all layers of ViT-Tiny. We observe that removing $②$ results in a catastrophic performance drop, whereas removing $④$ (as predicted) has a modest impact on performance. This difference becomes more pronounced as the scale of the dataset increases." + ], + "image_footnote": [], + "bbox": [ + 558, + 277, + 903, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 700, + 326, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (e.g., CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill", + "bbox": [ + 88, + 734, + 485, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying (TG), a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods.", + "bbox": [ + 511, + 702, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 811, + 645, + 827 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have demonstrated impressive performance on various computer vision and robotic tasks, such as object detection, semantic segmentation, video understanding, visual-language learning, and many oth-", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23115", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ers [8, 12, 17, 19, 33]. Much of this success is attributed to the self-attention mechanism, commonly referred to as Self-Attention Block (SAB), which allows ViTs to selectively attend to relevant parts of the input sequence when generating each element of the output sequence. Another essential component is the MLP, or Feedforward Networks (FFN), which facilitates intra-token communication across channel dimensions (see Fig. 1). Finally, identity mappings, commonly referred to as skip connections, are incorporated into both the SAB and FFN to enhance model performance.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that the Jacobian of the SAB is disproportionately ill-conditioned compared to other components, notably the FFN block. A poorly conditioned ${}^{1}$ Jacobian is fundamentally detrimental to gradient descent training [1, 13, 20, 24, 25], impeding both convergence and stability. A natural strategy to address this issue is to explicitly model the Jacobian of the entire network. However, this approach is impractical due to the sheer scale of modern transformer architectures. To proceed, we introduce the following simplifying assumptions:", + "bbox": [ + 89, + 242, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assumptions", + "text_level": 1, + "bbox": [ + 98, + 409, + 192, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The condition of the network Jacobian is bound by the most ill-conditioned sub-block Jacobian.", + "2. A proxy for the condition of the sub-block Jacobian is the condition of the output embedding of that block." + ], + "bbox": [ + 98, + 425, + 465, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although we do not provide formal proofs for these assumptions, we offer indirect validation through extensive empirical analysis. First, we show that better conditioning of a sub-block's output embeddings consistently correlates with improved empirical performance across multiple ViT benchmarks. Second, we demonstrate that this improvement in embedding condition is accompanied by a measurable enhancement in the sub-block's Jacobian condition", + "bbox": [ + 89, + 500, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skip in Transformers. Skip connections have become an indispensable component in modern transformer architectures [29] and have been empirically proven to be a de facto component. In Fig. 1 (b), we analyze a pre-trained ViT-Tiny model by measuring the condition numbers of the SAB and FFN output embeddings, both with and without skip connections. Our observations reveal that the SAB output embedding becomes highly ill-conditioned in the absence of skip connections, with a condition number approaching $e^6$ . In contrast, the other three configurations have relatively low condition numbers, around $e^3$ . In other words, the SAB output embedding without skip connections has significantly worse condition than the other configurations. To further investigate the impact of skip connections, we train the ViT-Tiny models from scratch under three config-", + "bbox": [ + 89, + 625, + 482, + 852 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c911ac0c9883163a64b9628d1eba9dd28af0452d086b6e5092eec04d10a79de.jpg", + "image_caption": [ + "Figure 2. Training loss for three different configurations of ViT-Tiny models trained on the Tiny-ImageNet dataset." + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "urations: the standard SAB and FFN with skip connections, SAB without skip connections, and FFN without skip connections.", + "bbox": [ + 511, + 342, + 903, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Fig. 1 (c), we observe a modest drop in classification accuracy $(2\\%)$ a ViT-Tiny model trained on the CIFAR-10 dataset when skip connections are removed from the FFN while retained in the SAB. However, removing skip connections from the SAB, while keeping them in the FFN, leads to a near-catastrophic performance drop of approximately $22\\%$ . Moreover, as dataset size increases, models without SAB skip connections degrade substantially in performance; on larger-scale datasets such as Tiny-[ImageNet, these models fail to train altogether. To the best of our knowledge, this empirical finding has not been reported in prior literature. Additionally, the training loss curves for the three configurations are shown in Fig. 2. The results demonstrate that the SAB without skip connections, which is significantly ill-conditioned, exhibits much slower convergence and diverges after 30 epochs. Interestingly, this critical dependence on skip connections is unique to transformer architectures.", + "bbox": [ + 511, + 388, + 905, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Skip in ConvMixer. To further highlight the poor conditioning of the SAB, we conducted a similar experiment on a modern CNN ConvMixer [28]. ConvMixer has a similar architecture and competitive performance to ViT with the exception that self-attention is replaced by a convolution block to spatially mix tokens. Unlike ViT, training ConvMixer-Tiny with and without skip connections, we observed that the performance remains effectively unchanged, with variations of less than $\\pm 0.2\\%$ . Further details are shown in Fig. 6 in Appendix. This contrast further suggests that the SAB introduces an extreme conditioning issue, making skip connections indispensable for stable training in ViTs.", + "bbox": [ + 511, + 664, + 905, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows.", + "bbox": [ + 532, + 854, + 870, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We present a proposition that characterizes why SAB output embedding without skip connection is funda", + "bbox": [ + 514, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1The condition number of a full-rank matrix is defined by the ratio of the largest singular value to the smallest singular value. A lower condition number implies better condition.", + "bbox": [ + 89, + 862, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23116", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mentally ill-conditioned, which challenges training convergence and stability (Sec. 4.1).", + "bbox": [ + 114, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. A theoretical analysis on the role of skip connection within the SAB is undertaken. We demonstrate that it significantly improves the condition of the block's output embedding, enhancing stability and performance (Sec. 4.2).", + "3. Finally, we propose a novel approach—Token Greying (TG)—to better pre-condition the input tokens. This step, when used in conjunction with conventional skip connections—improves ViT in both supervised and self-supervised settings (Sec. 5)." + ], + "bbox": [ + 89, + 121, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our central contribution in this paper is not TG itself, but the insight that the SAB within modern ViTs is intrinsically ill-conditioned. Our hope is that this insight opens up brand new lines of inquiry for the vision community for more effective and efficient ViT design.", + "bbox": [ + 89, + 272, + 480, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 359, + 227, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Attention. Originally proposed by Vaswani et al. [29], the self-attention mechanism has become central to transformer architectures, enabling the capture of long-range dependencies by dynamically weighting interactions between tokens because it captures long-range dependencies by dynamically weighting interactions between tokens. Building on this foundation, transformer-based large language models, such as Llama [6] and Deepseek [15], have significantly advanced language understanding. In computer vision, researchers introduced Vision Transformers (ViTs) [5] for image classification, achieving superior results compared to traditional convolutional neural networks. Moreover, researchers have trained Diffusion Transformers [21], replacing the commonly used U-Net backbone with a transformer that operates on latent patches and inherits the excellent scaling properties of the transformer model class. Similarly, self-attention mechanisms have been successfully applied in speech recognition and physics-informed neural networks, further demonstrating their versatility across diverse applications [3, 32].", + "bbox": [ + 89, + 383, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skip connections. He et al. [11] introduced skip connections in ResNet to make deep networks easier to optimize and to enhance accuracy by enabling significantly increased network depth. Since their introduction, skip connections have been extensively studied and applied across various Convolutional Neural Network (CNN) architectures. Building on these ideas, Veit et al. [30] offer a novel interpretation of residual networks, demonstrating that they can be viewed as a collection of multiple paths of varying lengths, which contributes to the network's flexibility and resilience. Furthermore, Hayou et al. [9] addresses gradient stability in deep ResNets, noting that while skip connections alleviate the issue of vanishing gradients, they may also intro-", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duce gradient explosion at extreme depths. To counter this, they propose scaling the skip connections according to the layer index, thereby stabilizing gradients and ensuring network expressivity even in the limit of infinite depth. Recently, in Table 3 of [16], the authors reported that VGG—a model without skip connections—achieves comparable results to ResNet—a model with skip connections—on modern datasets. To the best of our knowledge, only one paper challenges the skip connection in transformer architectures. In [10], the authors attempted to train deep transformer networks without using skip connections, achieving performance comparable to standard models. However, this approach requires five times more iterations, and the reasons why skip connections are crucial for self-attention-based transformers remain unresolved.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 332, + 638, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we define the self-attention blocks used in the Vision Transformer (ViT) architecture and establish the mathematical notation for key quantities referenced in subsequent sections. For a comprehensive overview of transformers, we refer readers to [5].", + "bbox": [ + 511, + 358, + 903, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A Vision Transformer architecture consists of $L$ stacked self-attention blocks (SABs) and feedforward networks (FFNs). An identity mapping, commonly referred to as a skip connection, is applied to connect the inputs and outputs of the transformation in both SAB and FFN.", + "bbox": [ + 511, + 434, + 903, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Attention Blocks in ViT. Formally, given an input sequence $\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}$ , with $n$ tokens of dimension $d$ , a SAB is defined as:", + "bbox": [ + 511, + 529, + 903, + 574 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {o u t}} := \\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 588, + 903, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the self-attention output embedding is", + "bbox": [ + 511, + 617, + 810, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) = \\eta \\left(\\mathbf {Q K} ^ {\\mathbf {T}}\\right) \\mathbf {V}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 647, + 903, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{Q} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{Q}},\\mathbf{K} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{K}}$ and $\\mathbf{V} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{V}}$ with learnable parameters $\\mathbf{W}^{\\mathrm{Q}},\\mathbf{W}^{\\mathrm{K}}$ , and $\\mathbf{W}^{\\mathrm{V}}\\in \\mathbb{R}^{d\\times d}$ . $\\eta (\\cdot)$ is an activation function and is set by default as softmax function in [5]. Additionally, some studies [22, 23, 26] have explored the possibilities of using alternative activation functions. In particular, good performance is achieved by using linear attention with an appropriate scale. It allows us to characterize why the SAB output embedding without a skip connection is ill-conditioned.", + "bbox": [ + 511, + 672, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For general vision transformers, multiple heads $\\mathrm{SA}(\\mathbf{X}_{\\mathrm{in}})_i$ , where $1 \\leq i \\leq h$ , are commonly used. Learnable matrices are divided into head numbers $h$ , such that $\\mathbf{W}_{\\mathrm{Q}}, \\mathbf{W}_{\\mathrm{K}}$ , and $\\mathbf{W}_{\\mathrm{V}} \\in \\mathbb{R}^{h \\times d \\times d_h}$ , where $d_h = \\frac{d}{h}$ . Then all outputs of each attention head are concatenated together before the skip connection.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23117", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feed Forward Networks in ViT. Formally, given an input sequence $\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}$ , with $n$ tokens of dimension $d$ , an FFN is defined as:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {\\text {o u t}} := \\mathbf {W} _ {\\text {d o w n}} \\left(g \\left(\\mathbf {W} _ {\\text {u p}} \\left(\\mathbf {X} _ {\\text {i n}}\\right)\\right)\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 148, + 482, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathrm{up}}\\in \\mathbb{R}^{d\\times 4d}$ is up projection, $\\mathbf{W}_{\\mathrm{down}}\\in \\mathbb{R}^{4d\\times d}$ is down projection, and $g(\\cdot)$ is an activation function.", + "bbox": [ + 89, + 176, + 483, + 209 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conditioning of matrices. Formally, for a rectangular full rank matrix $\\mathbf{A} \\in \\mathbb{R}^{n \\times d}$ , the condition number of $\\mathbf{A}$ is defined below:", + "bbox": [ + 89, + 224, + 483, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (\\mathbf {A}) = \\| \\mathbf {A} \\| _ {2} \\| \\mathbf {A} ^ {\\dagger} \\| _ {2} = \\frac {\\sigma_ {\\max } (\\mathbf {A})}{\\sigma_ {\\min } (\\mathbf {A})}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 281, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ is the matrix operator norm induced by the Euclidean norm. We use the Euclidean norm in this paper. $\\mathbf{A}^{\\dagger}$ is the pseudo inverse of the matrix $\\mathbf{A}$ . $\\sigma_{\\mathrm{max}}$ and $\\sigma_{\\mathrm{min}}$ are the maximal and minimal singular values of $\\mathbf{A}$ respectively. We will use Eq. (4) as a metric for analysis in future sections.", + "bbox": [ + 89, + 320, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Theoretical analysis", + "text_level": 1, + "bbox": [ + 89, + 410, + 282, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we provide a theoretical analysis showing that the output embedding of the SAB without skip connections is highly ill-conditioned. This matters as the condition of the output embedding is a proxy for the condition of the sub-block's Jacobian. We then demonstrate that an essential function of the skip connection is to improve this conditioning.", + "bbox": [ + 89, + 435, + 483, + 541 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Self-attention is ill-conditioned", + "text_level": 1, + "bbox": [ + 89, + 550, + 362, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We attempt to validate our claim for the simpler case of linear attention.", + "bbox": [ + 89, + 573, + 483, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 4.1. Assume $\\mathbf{X} \\in \\mathbb{R}^{n \\times d}$ , $\\mathbf{W}_{\\mathbf{Q}}$ , and $\\mathbf{W}_{\\mathbf{K}}$ and $\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}$ have entries that are independently drawn from a distribution with zero mean. The condition number of the SAB output embedding (without skip-connection) can be expressed as:", + "bbox": [ + 86, + 613, + 483, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa \\left(\\mathbf {X} \\mathbf {W} _ {Q} \\mathbf {W} _ {K} ^ {T} \\mathbf {X} ^ {T} \\mathbf {X} \\mathbf {W} _ {V}\\right) \\leq C \\cdot \\left(\\frac {\\sigma_ {\\text {m a x}}}{\\sigma_ {\\text {m i n}}}\\right) ^ {3}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 699, + 482, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{\\mathrm{max}}$ and $\\sigma_{\\mathrm{min}}$ are the maximal and minimal singular values of $\\mathbf{X}$ . $C$ is a fixed constant that is statistically close to unity.", + "bbox": [ + 89, + 763, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 4.1 assumes the use of linear attention (i.e., no softmax) as advocated for in [26]. This was done to offer a simplified theoretical exposition on how the SAB without skip connection can affect the condition of the output embedding. The proposition shows the condition number of the output embedding without skip connection is bounded", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5ba6082307532d2f6dab6594eef31d6e2a6cab166bb6691c441b1da5a00ebc4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 103, + 903, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f22015f645c58bf02500a14df6568b8fd01412dadc5a8448f859c10195db50fd.jpg", + "image_caption": [ + "Figure 3. Condition numbers across all layers of trained ViT-Tiny on Tiny-ImageNet datasets. We use $10\\%$ validation sets to plot the above figures. Upper: fixed scale linear self-attention. Lower: softmax self-attention" + ], + "image_footnote": [], + "bbox": [ + 517, + 273, + 901, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by the cube of the condition number of the input matrix $\\mathbf{X}$ . Unless the condition of $\\mathbf{X}$ is unity, the output of the SAB without skip connection will therefore result in a poorly conditioned embedding. Applying this same process across multiple layers within a transformer results in a highly ill-conditioned outcome.", + "bbox": [ + 511, + 525, + 905, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 illustrates our theoretical analysis by showing that the linear self-attention operation increases the condition number of embeddings from $e^3$ to $e^6$ , and we can empirically extend this analysis to softmax self-attention. The output embedding of the softmax SAB without skip connections is also significantly ill-conditioned.", + "bbox": [ + 511, + 616, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, for condition numbers of FFN output embeddings without skip connection using linear activation $g$ we have:", + "bbox": [ + 511, + 705, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (\\mathbf {M L P} (\\mathbf {X})) = \\kappa (\\mathbf {W} _ {\\mathrm {d o w n}} \\mathbf {W} _ {\\mathrm {u p}} \\mathbf {X}) \\leq C _ {\\mathrm {d o w n}} \\cdot C _ {\\mathrm {u p}} \\cdot \\frac {\\sigma_ {\\operatorname* {m a x}}}{\\sigma_ {\\operatorname* {m i n}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 758, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sigma_{\\mathrm{max}}$ and $\\sigma_{\\mathrm{min}}$ are maximal singular values of $\\mathbf{X}$ computed using SVD. Since $\\mathbf{W}_{\\mathrm{down}}$ and $\\mathbf{W}_{\\mathrm{up}}$ do not depend on data $\\mathbf{X}$ , we denote $\\kappa (\\mathbf{W}_{\\mathrm{down}}) = C_{\\mathrm{down}}$ and $\\kappa (\\mathbf{W}_{\\mathrm{up}}) = C_{\\mathrm{up}}$ .", + "bbox": [ + 511, + 795, + 905, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to the self-attention mechanism, the MLP transformation does not worsen conditioning as severely, since it has a much lower bound on the condition number. In Fig. 1 (b), the output embeddings of the FFN without", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23118", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "skip connection is observed to have a significantly better condition than the output embedding of the SAB without skip connection. Therefore, in (c), the performance drop after removing the FFN's skip connections is less severe than after removing the SAB's skip connections.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Skip connection improves condition", + "text_level": 1, + "bbox": [ + 89, + 176, + 401, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this subsection, we theoretically demonstrate that the skip connection improves the condition of the self-attention output embeddings.", + "bbox": [ + 89, + 199, + 483, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous works [4, 10, 18] have demonstrated that transformers, without skip connections, converge very slowly or may even become un-trainable. In [10], the authors show that without skip connections, output embeddings of the SAB quickly converge to rank 1. They hypothesize that this is due to the inability of the transformer to propagate signals through the self-attention mechanism. This rank-1 output embedding implies an infinite condition number. The authors introduce an inductive bias into the self-attention mechanism to enable what they refer to as \"better signal propagation\". While their approach matches the performance of standard (vanilla) models, it incurs a substantial computational cost—running approximately five times slower. We hypothesize that the success of this method could be directly attributed to its ability to improve the condition of the self-attention output embeddings.", + "bbox": [ + 89, + 244, + 483, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This observation further motivates our view that the primary role of skip connections may be to regularize the condition of the self-attention output embeddings", + "bbox": [ + 89, + 486, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4.2. Let $\\mathbf{X} \\in \\mathbb{R}^{n \\times d}$ and $\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}$ have entries that are independently drawn from a distribution with zero mean. We define $\\mathbf{M} = \\mathbf{W}_{\\mathbf{Q}} \\mathbf{W}_{\\mathbf{K}}^{\\mathrm{T}} \\mathbf{X}^{\\mathrm{T}} \\mathbf{X} \\mathbf{W}_{\\mathbf{V}}$ and assume $\\mathbf{M}$ is the positive semi-definite matrix. We have the following bounds on the condition numbers $\\kappa$ .", + "bbox": [ + 88, + 541, + 483, + 632 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (\\mathbf {X M} + \\mathbf {X}) \\ll \\kappa (\\mathbf {X M}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 643, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 4.2 shows that during training, the condition number of the SAB output embedding has a much lower bound than that of the SAB output embedding without skip connections. Furthermore, with other conditions being equal, better condition of SAB can contribute to faster convergence and more stable gradient updates, reducing the risk of gradient explosion or vanishing. This ensures the signal can propagate well during the training process in the Vision Transformers.", + "bbox": [ + 89, + 672, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In other words, the self-attention mechanism enables tokens to communicate by focusing on different parts of the input, thereby capturing correlations between elements within the same sequence. However, one critical drawback is that it is ill-conditioned and challenging to train. In contrast, although skip connections do not facilitate communi", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 SVD Token Graying" + ], + "code_body": "1: $\\triangleright$ Input: training data, amplification coefficient $\\epsilon \\in (0,1]$ \n2: $\\triangleright$ Training: \n3: for each minibatch $\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k$ do \n4: Convert image $\\mathbf{x}_i$ into token $\\mathbf{X}$ \n5: $\\mathbf{U\\Sigma V^T} \\leftarrow \\mathbf{X}$ $\\triangleright$ Compute SVD \n6: $\\tilde{\\Sigma} \\leftarrow \\frac{\\Sigma}{\\max(\\Sigma)}$ $\\triangleright$ Normalize elements $\\in (0,1]$ \n7: $\\tilde{\\Sigma} \\leftarrow \\tilde{\\Sigma}^\\epsilon$ $\\triangleright$ Amplify \n8: $\\tilde{\\mathbf{X}} \\leftarrow \\mathbf{U}\\tilde{\\Sigma}\\mathbf{V^T}$ $\\triangleright$ SVD reconstruct \n9: Patch Embedding: $\\mathbf{Z} \\leftarrow$ PatchEmbed(\\tilde{\\mathbf{X}}) \n10: Forward Pass: $\\hat{y} \\leftarrow$ Model(\\mathbf{Z}) \n11: Compute Loss: $\\mathcal{L} \\leftarrow$ Loss(\\hat{y},y_i) \n12: Backward Pass: Compute gradients $\\nabla_{\\theta}\\mathcal{L}$ \n13: Parameter Update: $\\theta \\leftarrow \\theta - \\eta\\nabla_{\\theta}\\mathcal{L}$ \n14: end for", + "bbox": [ + 519, + 109, + 906, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cation between tokens, they help regularize the condition of output embeddings and aid in the optimization process.", + "bbox": [ + 511, + 371, + 906, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, empirical results demonstrate that in both the linear and softmax self-attention cases, adding skip connections significantly improves the condition of output embeddings.", + "bbox": [ + 511, + 401, + 906, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Methodology", + "text_level": 1, + "bbox": [ + 511, + 477, + 648, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce Token Graying (TG), a simple yet effective method that employs Singular Value Decomposition and Discrete Cosine Transform techniques to reconstruct better-conditioned input tokens.", + "bbox": [ + 511, + 503, + 905, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Singular Value Decomposition (SVD)", + "text_level": 1, + "bbox": [ + 511, + 574, + 834, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Singular Value Decomposition of a matrix $\\mathbf{X} \\in \\mathbb{R}^{n \\times d}$ is the factorization of $\\mathbf{X}$ into the product of three matrices $\\mathbf{X} = \\mathbf{U} \\boldsymbol{\\Sigma} \\mathbf{V}^{\\mathrm{T}}$ where the columns of $\\mathbf{U} \\in \\mathbb{R}^{n \\times n}$ and $\\mathbf{V} \\in \\mathbb{R}^{d \\times d}$ are orthonormal and $\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{n \\times d}$ is a rectangular diagonal matrix whose diagonal elements are non-negative in descending order and represent the singular values such that $\\boldsymbol{\\Sigma}_{ii} = \\sigma_i$ . As observed in Eq. (4), we can reduce the condition number by increasing the minimal singular value while keeping the maximal singular value unchanged. Hence, our goal is to reconstruct $\\tilde{\\mathbf{X}} = \\mathbf{U} \\boldsymbol{\\tilde{\\Sigma}} \\mathbf{V}^{\\mathrm{T}}$ , where $\\boldsymbol{\\tilde{\\Sigma}}$ is formed by amplifying non-maximal elements of $\\boldsymbol{\\Sigma}$ while keeping the maximal element unchanged. Our SVD token graying pipeline is presented in Algorithm 1.", + "bbox": [ + 511, + 597, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, directly applying SVD to matrices is computationally expensive, with a cost of $O(nd \\cdot \\min(n, d))$ leading to longer training time. Tab. 1 demonstrates that training a ViT is significantly slower when using SVD reconstruction. Therefore, in the next subsection, we introduce a more efficient Token Graying (TG) method using the discrete cosine transform.", + "bbox": [ + 511, + 795, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23119", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7de2d55ac5052375a5e29a8acbc3808312a3754ceb27ddde41eb46c3af249f42.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsViT-BViT-B + SVDTGViT-B + DCTTG
Time (days)0.7234.5520.732
", + "bbox": [ + 98, + 89, + 480, + 131 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. ViT-Base training time w/o and w/ token graying methods on ImageNet-1K dataset.", + "bbox": [ + 89, + 141, + 482, + 170 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Discrete Cosine Transform (DCT)", + "text_level": 1, + "bbox": [ + 89, + 194, + 387, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A Discrete Cosine Transform (DCT) expresses a finite sequence of data points as a sum of cosine functions oscillating at different frequencies. It is real-valued and has an inverse, which is often simply called the Inverse DCT (IDCT).", + "bbox": [ + 89, + 215, + 483, + 276 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There are several variants of the DCT and in this work we choose DCT-II which is one of the most commonly used forms. Given a real-valued sequence data $\\mathbf{x} \\in \\mathbb{R}^{N \\times 1}$ , the 1D DCT sequence $\\hat{\\mathbf{x}}$ is defined as:", + "bbox": [ + 89, + 277, + 483, + 338 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mathbf {x}} _ {k} = \\alpha_ {k} \\sum_ {i = 0} ^ {N - 1} x _ {i} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right], \\mathrm {f o r} k = 0, 1, \\dots , N - 1 \\\\ \\text {w h e r e} \\alpha_ {k} = \\left\\{ \\begin{array}{l} \\sqrt {\\frac {1}{N}}, \\text {i f} k = 0 \\\\ \\sqrt {\\frac {2}{N}}, \\text {i f} k \\neq 0 \\end{array} \\right. \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 359, + 483, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simplicity, 1D DCT is a linear transformation, Eq. (8) can be written with the formula $DCT(\\mathbf{x}) = \\hat{\\mathbf{x}} = D\\mathbf{x}$ , where $D_{i,k}$ is the 1D DCT basis and orthogonal.", + "bbox": [ + 89, + 465, + 483, + 512 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i, k} = \\alpha_ {k} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 532, + 482, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the reconstruction process, since the DCT expresses a signal as a sum of cosine functions with different frequencies, the IDCT combines these frequency components to recover the original signal. Hence, $\\mathrm{IDCT}(\\hat{x}) = \\tilde{\\mathbf{x}} = D^T\\hat{\\mathbf{x}}$ . In our case, we have a token matrix $\\mathbf{X}\\in \\mathbb{R}^{n\\times d}$ and we generalize the 1D DCT to a 2D DCT using the formula $\\hat{\\mathbf{X}} = D\\mathbf{X}D^{T}$ . In natural images, the dominant singular vector, associated with the largest singular value, typically captures the low frequency content of the image. Intuitively, this observation suggests that DCT reconstruction can serve as an approximation for SVD reconstruction. Our DCT token graying pipeline is presented in Algorithm 2.", + "bbox": [ + 89, + 575, + 482, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DCT avoids the expensive computational cost of SVD and has a complexity of $O(nd \\cdot \\log(nd))$ . Tab. 1 shows that training ViT using the DCT algorithm is significantly faster than SVD algorithm.", + "bbox": [ + 89, + 756, + 483, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 223, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have emerged as powerful models in the field of computer vision, demonstrating remarkable performance across a variety of tasks. This section is", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 2 DCT Token Graying" + ], + "code_body": "1: $\\triangleright$ Input: training data, amplification coefficient $\\epsilon \\in (0,1]$ \n2: Training: \n3: for each minibatch $\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k \\sim$ training data do \n4: Convert image $\\mathbf{x}_i$ into token $\\mathbf{X}$ \n5: $D\\hat{\\mathbf{X}}D^T \\leftarrow \\mathbf{X}$ $\\triangleright$ Compute DCT \n6: $\\hat{\\mathbf{X}}_{\\mathrm{norm}} \\leftarrow |\\hat{\\mathbf{X}}| / \\max(|\\hat{\\mathbf{X}}|)$ $\\triangleright$ Normalize elements $\\in (0,1]$ \n7: $\\hat{\\mathbf{X}} \\leftarrow \\hat{\\mathbf{X}}_{\\mathrm{norm}}^\\epsilon \\cdot \\mathrm{sign}(\\hat{\\mathbf{X}}) \\cdot \\max(|\\hat{\\mathbf{X}}|)$ $\\triangleright$ Amplify \n8: $\\hat{\\mathbf{X}} \\leftarrow \\hat{D}^T\\hat{\\mathbf{X}}\\hat{D}$ $\\triangleright$ Inverse DCT \n9: Patch Embedding: $\\mathbf{Z} \\leftarrow$ PatchEmbed( $\\hat{\\mathbf{X}}$ ) \n10: Forward Pass: $\\hat{y} \\leftarrow$ Model( $\\mathbf{Z}$ ) \n11: Compute Loss: $\\mathcal{L} \\leftarrow$ Loss( $\\hat{y}, y_i$ ) \n12: Backward Pass: Compute gradients $\\nabla_\\theta \\mathcal{L}$ \n13: Parameter Update: $\\theta \\leftarrow \\theta - \\eta \\nabla_\\theta \\mathcal{L}$ \n14: end for", + "bbox": [ + 519, + 109, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dedicated to validating and analyzing our proposed algorithm for both supervised and self-supervised learning in ViTs. For all experiments, unless specified otherwise, we use $\\epsilon = 0.95$ .", + "bbox": [ + 511, + 382, + 906, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Supervised learning ViT", + "text_level": 1, + "bbox": [ + 511, + 454, + 736, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we validate our methods in supervised learning using different types and scales of ViTs. Cross-entropy loss objective is used. We use PyTorch and Timm libraries.", + "bbox": [ + 511, + 476, + 905, + 536 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ViT [5]: is the pioneering work that interprets an image as a sequence of patches and processes it by a standard Transformer encoder as used in NLP. This simple, yet scalable, strategy works surprisingly well when coupled with pretraining on large datasets. We train the ViT-Tiny, which has 12 layers and 3 heads, with a head dimension of 64 and a token dimension of 192 on the Tiny-[ImageNet] dataset [14]. Additionally, we train ViT-Base, which consists of 12 layers and 12 heads, each with a head dimension of 64 and a token dimension of 768 on the ImageNet-1K dataset [2].", + "- Swin [17] is a hierarchical vision transformer designed to focus on local attention first and expand globally. It improves computational efficiency and locality modeling compared to standard ViTs. We train on the Swin small model with a patch size of 4 and a window size of 7 on the ImageNet-1K dataset.", + "- CaiT [27] is a variant Vision Transformer that aims at increasing the stability of the optimization when ViTs go deeper. We train CaiT small, which has 24 layers, on the ImageNet-1K dataset.", + "- PVT [31] an enhanced version of the Pyramid Vision Transformer that utilizes a hierarchical architecture to ex" + ], + "bbox": [ + 511, + 544, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23120", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f8adc1d0cc14bd7e6c1a9508b8df64fddf56cc26690069983122e2f95c93460.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc (%)Top-5 Acc (%)
1/k(QK)43.062.7
1/k(QK) + SVD44.368.8
1/k(QK) + DCT44.268.8
softmax(QK)43.067.3
softmax(QK) + SVD44.769.7
softmax(QK) + DCT44.869.7
", + "bbox": [ + 91, + 101, + 496, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4b33c550490601f560c0283bd6730f2a12a66efc6e550890edcc501525d15348.jpg", + "table_caption": [ + "Table 2. ViT-Tiny results on Tiny-ImageNet datasets. In the upper part, we show results for linear self-attention, where $k = 16$ is chosen based on [26] w/ and w/o our method. On the lower part, we show regular softmax self-attention results w/ and w/o our method." + ], + "table_footnote": [], + "table_body": "
MethodεTop-1 Acc (%)κinκout
ViT-Base-81.06.726.74
ViT-Base + SVDTG0.981.26.646.66
0.881.26.476.66
0.781.46.156.17
0.681.45.735.71
0.581.05.295.25
ViT-Base + DCTTG0.9781.16.476.49
0.9581.36.426.43
0.9381.26.336.35
0.981.16.256.25
0.8581.06.016.02
", + "bbox": [ + 101, + 327, + 478, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. ViT-Base results using SVD and DCT token graying methods with different amplification coefficient $\\epsilon$ . $\\kappa_{\\mathrm{in}}$ and $\\kappa_{\\mathrm{out}}$ are the average token condition numbers in log scale before and after each SAB.", + "bbox": [ + 89, + 526, + 483, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tract multi-scale features for improved performance in computer vision tasks. It incorporates efficient attention mechanisms and refined positional encoding to boost accuracy and reduce computational complexity.", + "bbox": [ + 89, + 604, + 483, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. In Tab. 2, we first demonstrate the results for the linear case and the regular softmax case using our methods on a small model (ViT-Tiny) and the small-scale TinyImageNet dataset. Given that Theorem Proposition 4.2 and Proposition Proposition 4.1 are mathematically explained using linear self-attention, we train the ViT-Tiny model with linear self-attention, which aligns well with our analysis in Fig. 3. Furthermore, as shown in Fig. 4, both SVDTG and DCTTG lead to better conditioning of SAB output embeddings throughout training epochs. Next, we train the model on a larger dataset using the default softmax self-attention, further validating that our approach generalizes to softmax self-attention. Moreover, SVDTG is the most direct method to improve the conditioning of a matrix, and as demonstrated in Tab. 3, it leads to notable performance gains.", + "bbox": [ + 89, + 672, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa422d58e87cec12de2196802ebbc98278cc0889416f479d1a5470ae053eabf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 246 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b70abd157dd403e423c8e45644f041348a8a1498b175ef3318e083bd2746f44.jpg", + "image_caption": [ + "Figure 4. We compute the condition number of the SAB block's output embeddings during ViT-Tiny training using SVDTG and DCTTG with different $\\epsilon$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 250, + 903, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To address the high computational cost—approximately six times slower than regular ViT training—we approximate SVDTG using DCTTG, which achieves better token embeddings conditioning and yields comparable results too. As shown in Fig. 5, our DCTTG method produces condition numbers across all layers that indicate better token conditioning in almost every layer. Finally, we evaluate our methods on different variants of ViTs trained on ImageNet-1K datasets, where our DCTTG method demonstrates superior performance compared to vanilla models in terms of Top-1 and Top-5 classification accuracy, while also achieving better conditioning of SAB output embeddings across all layers in the ViT-B model.", + "bbox": [ + 511, + 491, + 906, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Self-supervised learning ViT", + "text_level": 1, + "bbox": [ + 511, + 700, + 769, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we validate our methods on self-supervised learning models as well as on fine-tuning pretrained models [7, 12]. Self-supervised learning is a powerful technique in which models learn representations directly from unlabeled data. Masked Image Modeling (MIM) [12] learns representations by reconstructing images that have been corrupted through masking.", + "bbox": [ + 511, + 724, + 905, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-training. In the pre-training stage, we pre-train our model on the ImageNet-1k training set without using ground-truth labels in a self-supervised manner. We use a batch size of 2048, a base learning rate of $1.5 \\times 10^{-4}$ , and a", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23121", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/975f543b1ab1a1f5abe96ba9cfd299d027913e3f0eac67cadfd52b7ed7c3afab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc (%)Top-5 Acc (%)
ViT-S80.295.1
ViT-S + DCTTG80.495.2
ViT-B81.095.3
ViT-B + DCTTG81.395.4
Swin-S81.395.6
Swin-S + DCTTG81.695.6
CaiT-S82.696.1
CaiT-S + DCTTG82.796.3
PVT V2 b382.996.0
PVT V2 b3 + DCTTG83.096.1
", + "bbox": [ + 91, + 88, + 498, + 296 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e7e7342aa6c476d69917c3e7a8aa0edcca9ac887a87de8ae65a5728057b630f.jpg", + "image_caption": [ + "Figure 5. Condition numbers of all layers in the trained ViT-Base model with and without the DCTTG method." + ], + "image_footnote": [], + "bbox": [ + 94, + 348, + 478, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fcec65867870f30dffe3773915013670363c66ddc098f4f0a4a8d26f2b736eef.jpg", + "table_caption": [ + "Table 4. Top-1 and Top-5 classification accuracy on ImageNet-1K dataset using DCT token graying on different ViTs configurations." + ], + "table_footnote": [], + "table_body": "
MethodTop-1 Acc (%)Top-5 Acc (%)
MAE83.096.4
MAE + DCTTG83.296.6
", + "bbox": [ + 104, + 594, + 470, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. MAE pretrained ViT-Base results finetuned on ImageNet dataset-1k. We evaluate our MSVR method on self-attention using the power iteration method.", + "bbox": [ + 89, + 667, + 482, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "weight decay of 0.05, and we train for 400 epochs. Our loss function computes the mean squared error (MSE) between the reconstructed and original images in pixel space.", + "bbox": [ + 89, + 737, + 482, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning and reuslts. In the finetuning stage, we fine-tune the pre-trained models on ImageNet-1k for classification tasks using cross-entropy loss. We use a batch size of 512, a base learning rate of $5 \\times 10^{-4}$ , a weight decay of 0.05, and train for 100 epochs. As shown in Tab. 5, our DCTTG method outperforms the vanilla MAE model by $0.2\\%$ in both top-1 and top-5 classification accuracy.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitation", + "text_level": 1, + "bbox": [ + 513, + 89, + 627, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work establishes a theoretical foundation demonstrating why skip connections act as essential conditioning regularizers within self-attention blocks. Additionally, we introduce a method designed to further enhance the output embeddings of self-attention blocks. However, certain limitations remain. For instance, while our regularizer is effective, training in low-precision settings may pose challenges due to the DCT operation, which involves many multiplications and summations that are sensitive to quantization errors. Additionally, for some variants of ViTs, the performance improvement might be marginal.", + "bbox": [ + 511, + 114, + 906, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 296, + 633, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper provides a theoretical demonstration that the output embeddings of linear self-attention blocks, when lacking skip connections, are inherently ill-conditioned, a finding that empirically extends to conventional softmax self-attention blocks as well. Furthermore, we show that skip connections serve as a powerful regularizer by significantly improving the conditioning of self-attention blocks' output embeddings. In addition, we use the condition of the self-attention output embeddings as a proxy for that of its Jacobian, empirically demonstrating that input tokens with improved conditioning lead to a better-conditioned Jacobian. Building on this insight, we introduce SVDTG and DCTTG, simple yet effective methods designed to improve the condition of input tokens. We then validate their effectiveness on both supervised and unsupervised learning tasks. Ultimately, we hope these insights will guide the vision community in designing more effective and efficient ViT architectures for downstream applications.", + "bbox": [ + 511, + 321, + 906, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the Australian Institute for Machine Learning (AIML) and the CSIRO's Data61 Embodied AI Cluster.", + "bbox": [ + 511, + 601, + 906, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 514, + 661, + 609, + 678 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Paul Albert, Frederic Z Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Randlora: Full-rank parameter-efficient fine-tuning of large models. In International Conference on Learning Representations (ICLR), 2025. 2", + "[2] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 6", + "[3] Linhao Dong, Shuang Xu, and Bo Xu. Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5884-5888, 2018. 3", + "[4] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas." + ], + "bbox": [ + 521, + 686, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23122", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 5", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 3, 6", + "[6] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3", + "[7] Maryam Haghighat, Peyman Moghadam, Shaheer Mohamed, and Piotr Koniusz. Pre-training with Random Orthogonal Projection Image Modeling. In International Conference on Learning Representations (ICLR), 2024. 7", + "[8] Stephen Hausler and Peyman Moghadam. Pair-VPR: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers. IEEE Robotics and Automation Letters, 2025. 2", + "[9] Soufiane Hayou, Eugenio Clerico, Bobby He, George Deligiannidis, Arnaud Doucet, and Judith Rousseau. Stable resnet. In International Conference on Artificial Intelligence and Statistics, pages 1324-1332. PMLR, 2021. 3", + "[10] Bobby He, James Martens, Guodong Zhang, Aleksandar Botev, Andrew Brock, et al. Deep transformers without shortcuts: Modifying self-attention for faithful signal propagation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7", + "[13] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. In The International Conference on Learning Representations, 2025. 2", + "[14] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[15] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 3", + "[16] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6", + "[18] Lorenzo Noci, Chuning Li, Mufan Li, Bobby He, Thomas Hofmann, Chris J Maddison, and Dan Roy. The shaped" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transformer: Attention models in the infinite depth-andwidth limit. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[19] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2", + "[20] Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks, 2013. 2", + "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3", + "[22] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, et al. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 3", + "[23] Jason Ramapuram, Federico Danieli, Eeshan Gunesh Dhekane, Floris Weers, Dan Busbridge, et al. Theory, analysis, and best practices for sigmoid self-attention. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[24] Hemanth Saratchandran and Simon Lucey. Enhancing transformers through conditioned embedded tokens. arXiv preprint arXiv:2505.12789, 2025. 2", + "[25] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2024. 2", + "[26] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking softmax: Self-attention with polynomial activations, 2024. 3, 4, 7", + "[27] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 32-42, 2021. 6", + "[28] Asher Trockman and J. Zico Kolter. Patches are all you need?, 2022. 2, 10", + "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3", + "[30] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in neural information processing systems, 29, 2016. 3", + "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 6", + "[32] Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks. In The International Conference on Learning Representations (ICLR), 2024. 3", + "[33] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23123", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_model.json b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c22a63bf3d48f03b702643fde452a30965c115f --- /dev/null +++ b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_model.json @@ -0,0 +1,2088 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.0, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.383, + 0.131, + 0.614, + 0.154 + ], + "angle": 0, + "content": "Always Skip Attention" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.18, + 0.8, + 0.217 + ], + "angle": 0, + "content": "Yiping Ji\\(^{1,2}\\), Hemanth Saratchandran\\(^{1}\\), Peyman Moghadam\\(^{2,3}\\), Simon Lucey\\(^{1}\\) \n\\(^{1}\\)Adelaide University, \\(^{2}\\)CSIRO, \\(^{3}\\)Queensland University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.22, + 0.775, + 0.252 + ], + "angle": 0, + "content": "{yiping.ji, hemanth.saratchandran, simon.lucy}@adelaide.edu.au, {yiping.ji, peyman.moghadam}@csiro.au" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.264, + 0.207, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.487, + 0.201, + 0.501 + ], + "angle": 0, + "content": "(a) Architecture" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.273, + 0.557, + 0.483 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.488, + 0.541, + 0.517 + ], + "angle": 0, + "content": "(b) FFN and SAB output embedding condition number using different configurations" + }, + { + "type": "image", + "bbox": [ + 0.559, + 0.278, + 0.904, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.487, + 0.882, + 0.502 + ], + "angle": 0, + "content": "(c) Performance after removing skip connections" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.528, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Figure 1. Removing skip connections from the Self-Attention Block (SAB) versus the Feedforward Network (FFN) in Vision Transformer (ViT) models results in different performance drop. (a) Model architecture of ViT consisting of a SAB, and a FFN in each layer. \\(①\\) represents the output of the self-attention transformation, while \\(③\\) represents the output of the MLP transformation. \\(②\\) and \\(④\\) represent identity mappings, referred to as skip connections, in the SAB and FFN, respectively. (b) The condition number of the different output embeddings for each hidden layer. A ViT-Tiny model is trained in a typical manner with skip connections on both SAB and FFN blocks (i.e., \\(① + ② + ③ + ④\\)). It shows that the SAB output embedding without skip connection is poorly conditioned \\((①)\\), with a condition number significantly higher than that of the other three configurations involving ViT blocks. (c) Classification results on three different datasets using the ViT-Tiny model under three different architectures at train time. The result shows the effects of removing skip connections at train time on the SAB and FFN blocks (either \\(②\\) or \\(④\\)) in all layers of ViT-Tiny. We observe that removing \\(②\\) results in a catastrophic performance drop, whereas removing \\(④\\) (as predicted) has a modest impact on performance. This difference becomes more pronounced as the scale of the dataset increases." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.701, + 0.327, + 0.716 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (e.g., CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.794 + ], + "angle": 0, + "content": "conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying (TG), a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.812, + 0.646, + 0.828 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have demonstrated impressive performance on various computer vision and robotic tasks, such as object detection, semantic segmentation, video understanding, visual-language learning, and many oth-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23115" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "ers [8, 12, 17, 19, 33]. Much of this success is attributed to the self-attention mechanism, commonly referred to as Self-Attention Block (SAB), which allows ViTs to selectively attend to relevant parts of the input sequence when generating each element of the output sequence. Another essential component is the MLP, or Feedforward Networks (FFN), which facilitates intra-token communication across channel dimensions (see Fig. 1). Finally, identity mappings, commonly referred to as skip connections, are incorporated into both the SAB and FFN to enhance model performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.395 + ], + "angle": 0, + "content": "We argue that the Jacobian of the SAB is disproportionately ill-conditioned compared to other components, notably the FFN block. A poorly conditioned \\( {}^{1} \\) Jacobian is fundamentally detrimental to gradient descent training [1, 13, 20, 24, 25], impeding both convergence and stability. A natural strategy to address this issue is to explicitly model the Jacobian of the entire network. However, this approach is impractical due to the sheer scale of modern transformer architectures. To proceed, we introduce the following simplifying assumptions:" + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.41, + 0.194, + 0.424 + ], + "angle": 0, + "content": "Assumptions" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.426, + 0.466, + 0.453 + ], + "angle": 0, + "content": "1. The condition of the network Jacobian is bound by the most ill-conditioned sub-block Jacobian." + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.455, + 0.466, + 0.484 + ], + "angle": 0, + "content": "2. A proxy for the condition of the sub-block Jacobian is the condition of the output embedding of that block." + }, + { + "type": "list", + "bbox": [ + 0.099, + 0.426, + 0.466, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.621 + ], + "angle": 0, + "content": "Although we do not provide formal proofs for these assumptions, we offer indirect validation through extensive empirical analysis. First, we show that better conditioning of a sub-block's output embeddings consistently correlates with improved empirical performance across multiple ViT benchmarks. Second, we demonstrate that this improvement in embedding condition is accompanied by a measurable enhancement in the sub-block's Jacobian condition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Skip in Transformers. Skip connections have become an indispensable component in modern transformer architectures [29] and have been empirically proven to be a de facto component. In Fig. 1 (b), we analyze a pre-trained ViT-Tiny model by measuring the condition numbers of the SAB and FFN output embeddings, both with and without skip connections. Our observations reveal that the SAB output embedding becomes highly ill-conditioned in the absence of skip connections, with a condition number approaching \\( e^6 \\). In contrast, the other three configurations have relatively low condition numbers, around \\( e^3 \\). In other words, the SAB output embedding without skip connections has significantly worse condition than the other configurations. To further investigate the impact of skip connections, we train the ViT-Tiny models from scratch under three config-" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.091, + 0.895, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.287, + 0.905, + 0.315 + ], + "angle": 0, + "content": "Figure 2. Training loss for three different configurations of ViT-Tiny models trained on the Tiny-ImageNet dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.905, + 0.387 + ], + "angle": 0, + "content": "urations: the standard SAB and FFN with skip connections, SAB without skip connections, and FFN without skip connections." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.906, + 0.66 + ], + "angle": 0, + "content": "In Fig. 1 (c), we observe a modest drop in classification accuracy \\((2\\%)\\) a ViT-Tiny model trained on the CIFAR-10 dataset when skip connections are removed from the FFN while retained in the SAB. However, removing skip connections from the SAB, while keeping them in the FFN, leads to a near-catastrophic performance drop of approximately \\(22\\%\\). Moreover, as dataset size increases, models without SAB skip connections degrade substantially in performance; on larger-scale datasets such as Tiny-[ImageNet, these models fail to train altogether. To the best of our knowledge, this empirical finding has not been reported in prior literature. Additionally, the training loss curves for the three configurations are shown in Fig. 2. The results demonstrate that the SAB without skip connections, which is significantly ill-conditioned, exhibits much slower convergence and diverges after 30 epochs. Interestingly, this critical dependence on skip connections is unique to transformer architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.846 + ], + "angle": 0, + "content": "Skip in ConvMixer. To further highlight the poor conditioning of the SAB, we conducted a similar experiment on a modern CNN ConvMixer [28]. ConvMixer has a similar architecture and competitive performance to ViT with the exception that self-attention is replaced by a convolution block to spatially mix tokens. Unlike ViT, training ConvMixer-Tiny with and without skip connections, we observed that the performance remains effectively unchanged, with variations of less than \\(\\pm 0.2\\%\\). Further details are shown in Fig. 6 in Appendix. This contrast further suggests that the SAB introduces an extreme conditioning issue, making skip connections indispensable for stable training in ViTs." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.855, + 0.871, + 0.869 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1. We present a proposition that characterizes why SAB output embedding without skip connection is funda" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.9 + ], + "angle": 0, + "content": "1The condition number of a full-rank matrix is defined by the ratio of the largest singular value to the smallest singular value. A lower condition number implies better condition." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.116, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "mentally ill-conditioned, which challenges training convergence and stability (Sec. 4.1)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.122, + 0.482, + 0.196 + ], + "angle": 0, + "content": "2. A theoretical analysis on the role of skip connection within the SAB is undertaken. We demonstrate that it significantly improves the condition of the block's output embedding, enhancing stability and performance (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.197, + 0.482, + 0.272 + ], + "angle": 0, + "content": "3. Finally, we propose a novel approach—Token Greying (TG)—to better pre-condition the input tokens. This step, when used in conjunction with conventional skip connections—improves ViT in both supervised and self-supervised settings (Sec. 5)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.122, + 0.482, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.349 + ], + "angle": 0, + "content": "Our central contribution in this paper is not TG itself, but the insight that the SAB within modern ViTs is intrinsically ill-conditioned. Our hope is that this insight opens up brand new lines of inquiry for the vision community for more effective and efficient ViT design." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.228, + 0.375 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.485, + 0.688 + ], + "angle": 0, + "content": "Self-Attention. Originally proposed by Vaswani et al. [29], the self-attention mechanism has become central to transformer architectures, enabling the capture of long-range dependencies by dynamically weighting interactions between tokens because it captures long-range dependencies by dynamically weighting interactions between tokens. Building on this foundation, transformer-based large language models, such as Llama [6] and Deepseek [15], have significantly advanced language understanding. In computer vision, researchers introduced Vision Transformers (ViTs) [5] for image classification, achieving superior results compared to traditional convolutional neural networks. Moreover, researchers have trained Diffusion Transformers [21], replacing the commonly used U-Net backbone with a transformer that operates on latent patches and inherits the excellent scaling properties of the transformer model class. Similarly, self-attention mechanisms have been successfully applied in speech recognition and physics-informed neural networks, further demonstrating their versatility across diverse applications [3, 32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Skip connections. He et al. [11] introduced skip connections in ResNet to make deep networks easier to optimize and to enhance accuracy by enabling significantly increased network depth. Since their introduction, skip connections have been extensively studied and applied across various Convolutional Neural Network (CNN) architectures. Building on these ideas, Veit et al. [30] offer a novel interpretation of residual networks, demonstrating that they can be viewed as a collection of multiple paths of varying lengths, which contributes to the network's flexibility and resilience. Furthermore, Hayou et al. [9] addresses gradient stability in deep ResNets, noting that while skip connections alleviate the issue of vanishing gradients, they may also intro-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "duce gradient explosion at extreme depths. To counter this, they propose scaling the skip connections according to the layer index, thereby stabilizing gradients and ensuring network expressivity even in the limit of infinite depth. Recently, in Table 3 of [16], the authors reported that VGG—a model without skip connections—achieves comparable results to ResNet—a model with skip connections—on modern datasets. To the best of our knowledge, only one paper challenges the skip connection in transformer architectures. In [10], the authors attempted to train deep transformer networks without using skip connections, achieving performance comparable to standard models. However, this approach requires five times more iterations, and the reasons why skip connections are crucial for self-attention-based transformers remain unresolved." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.333, + 0.64, + 0.351 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.905, + 0.434 + ], + "angle": 0, + "content": "In this section, we define the self-attention blocks used in the Vision Transformer (ViT) architecture and establish the mathematical notation for key quantities referenced in subsequent sections. For a comprehensive overview of transformers, we refer readers to [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.905, + 0.51 + ], + "angle": 0, + "content": "A Vision Transformer architecture consists of \\(L\\) stacked self-attention blocks (SABs) and feedforward networks (FFNs). An identity mapping, commonly referred to as a skip connection, is applied to connect the inputs and outputs of the transformation in both SAB and FFN." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Self-Attention Blocks in ViT. Formally, given an input sequence \\(\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}\\), with \\(n\\) tokens of dimension \\(d\\), a SAB is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.589, + 0.905, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {o u t}} := \\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.618, + 0.812, + 0.634 + ], + "angle": 0, + "content": "where the self-attention output embedding is" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.648, + 0.905, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) = \\eta \\left(\\mathbf {Q K} ^ {\\mathbf {T}}\\right) \\mathbf {V}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.809 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{Q}},\\mathbf{K} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{K}}\\) and \\(\\mathbf{V} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{V}}\\) with learnable parameters \\(\\mathbf{W}^{\\mathrm{Q}},\\mathbf{W}^{\\mathrm{K}}\\), and \\(\\mathbf{W}^{\\mathrm{V}}\\in \\mathbb{R}^{d\\times d}\\). \\(\\eta (\\cdot)\\) is an activation function and is set by default as softmax function in [5]. Additionally, some studies [22, 23, 26] have explored the possibilities of using alternative activation functions. In particular, good performance is achieved by using linear attention with an appropriate scale. It allows us to characterize why the SAB output embedding without a skip connection is ill-conditioned." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "For general vision transformers, multiple heads \\(\\mathrm{SA}(\\mathbf{X}_{\\mathrm{in}})_i\\), where \\(1 \\leq i \\leq h\\), are commonly used. Learnable matrices are divided into head numbers \\(h\\), such that \\(\\mathbf{W}_{\\mathrm{Q}}, \\mathbf{W}_{\\mathrm{K}}\\), and \\(\\mathbf{W}_{\\mathrm{V}} \\in \\mathbb{R}^{h \\times d \\times d_h}\\), where \\(d_h = \\frac{d}{h}\\). Then all outputs of each attention head are concatenated together before the skip connection." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23117" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "Feed Forward Networks in ViT. Formally, given an input sequence \\(\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}\\), with \\(n\\) tokens of dimension \\(d\\), an FFN is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.159, + 0.149, + 0.483, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {\\text {o u t}} := \\mathbf {W} _ {\\text {d o w n}} \\left(g \\left(\\mathbf {W} _ {\\text {u p}} \\left(\\mathbf {X} _ {\\text {i n}}\\right)\\right)\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.21 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathrm{up}}\\in \\mathbb{R}^{d\\times 4d}\\) is up projection, \\(\\mathbf{W}_{\\mathrm{down}}\\in \\mathbb{R}^{4d\\times d}\\) is down projection, and \\(g(\\cdot)\\) is an activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.484, + 0.271 + ], + "angle": 0, + "content": "Conditioning of matrices. Formally, for a rectangular full rank matrix \\(\\mathbf{A} \\in \\mathbb{R}^{n \\times d}\\), the condition number of \\(\\mathbf{A}\\) is defined below:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.282, + 0.484, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\kappa (\\mathbf {A}) = \\| \\mathbf {A} \\| _ {2} \\| \\mathbf {A} ^ {\\dagger} \\| _ {2} = \\frac {\\sigma_ {\\max } (\\mathbf {A})}{\\sigma_ {\\min } (\\mathbf {A})}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.398 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) is the matrix operator norm induced by the Euclidean norm. We use the Euclidean norm in this paper. \\(\\mathbf{A}^{\\dagger}\\) is the pseudo inverse of the matrix \\(\\mathbf{A}\\). \\(\\sigma_{\\mathrm{max}}\\) and \\(\\sigma_{\\mathrm{min}}\\) are the maximal and minimal singular values of \\(\\mathbf{A}\\) respectively. We will use Eq. (4) as a metric for analysis in future sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.411, + 0.284, + 0.429 + ], + "angle": 0, + "content": "4. Theoretical analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.542 + ], + "angle": 0, + "content": "In this section, we provide a theoretical analysis showing that the output embedding of the SAB without skip connections is highly ill-conditioned. This matters as the condition of the output embedding is a proxy for the condition of the sub-block's Jacobian. We then demonstrate that an essential function of the skip connection is to improve this conditioning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.364, + 0.567 + ], + "angle": 0, + "content": "4.1. Self-attention is ill-conditioned" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.603 + ], + "angle": 0, + "content": "We attempt to validate our claim for the simpler case of linear attention." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.614, + 0.485, + 0.69 + ], + "angle": 0, + "content": "Proposition 4.1. Assume \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d}\\), \\(\\mathbf{W}_{\\mathbf{Q}}\\), and \\(\\mathbf{W}_{\\mathbf{K}}\\) and \\(\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}\\) have entries that are independently drawn from a distribution with zero mean. The condition number of the SAB output embedding (without skip-connection) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.7, + 0.483, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\kappa \\left(\\mathbf {X} \\mathbf {W} _ {Q} \\mathbf {W} _ {K} ^ {T} \\mathbf {X} ^ {T} \\mathbf {X} \\mathbf {W} _ {V}\\right) \\leq C \\cdot \\left(\\frac {\\sigma_ {\\text {m a x}}}{\\sigma_ {\\text {m i n}}}\\right) ^ {3}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.81 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\mathrm{max}}\\) and \\(\\sigma_{\\mathrm{min}}\\) are the maximal and minimal singular values of \\(\\mathbf{X}\\). \\(C\\) is a fixed constant that is statistically close to unity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Proposition 4.1 assumes the use of linear attention (i.e., no softmax) as advocated for in [26]. This was done to offer a simplified theoretical exposition on how the SAB without skip connection can affect the condition of the output embedding. The proposition shows the condition number of the output embedding without skip connection is bounded" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.104, + 0.905, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.274, + 0.903, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.445, + 0.907, + 0.5 + ], + "angle": 0, + "content": "Figure 3. Condition numbers across all layers of trained ViT-Tiny on Tiny-ImageNet datasets. We use \\(10\\%\\) validation sets to plot the above figures. Upper: fixed scale linear self-attention. Lower: softmax self-attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.906, + 0.615 + ], + "angle": 0, + "content": "by the cube of the condition number of the input matrix \\(\\mathbf{X}\\). Unless the condition of \\(\\mathbf{X}\\) is unity, the output of the SAB without skip connection will therefore result in a poorly conditioned embedding. Applying this same process across multiple layers within a transformer results in a highly ill-conditioned outcome." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.906, + 0.706 + ], + "angle": 0, + "content": "Fig. 3 illustrates our theoretical analysis by showing that the linear self-attention operation increases the condition number of embeddings from \\( e^3 \\) to \\( e^6 \\), and we can empirically extend this analysis to softmax self-attention. The output embedding of the softmax SAB without skip connections is also significantly ill-conditioned." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.751 + ], + "angle": 0, + "content": "Further, for condition numbers of FFN output embeddings without skip connection using linear activation \\(g\\) we have:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.759, + 0.907, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\kappa (\\mathbf {M L P} (\\mathbf {X})) = \\kappa (\\mathbf {W} _ {\\mathrm {d o w n}} \\mathbf {W} _ {\\mathrm {u p}} \\mathbf {X}) \\leq C _ {\\mathrm {d o w n}} \\cdot C _ {\\mathrm {u p}} \\cdot \\frac {\\sigma_ {\\operatorname* {m a x}}}{\\sigma_ {\\operatorname* {m i n}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.841 + ], + "angle": 0, + "content": "where \\(\\sigma_{\\mathrm{max}}\\) and \\(\\sigma_{\\mathrm{min}}\\) are maximal singular values of \\(\\mathbf{X}\\) computed using SVD. Since \\(\\mathbf{W}_{\\mathrm{down}}\\) and \\(\\mathbf{W}_{\\mathrm{up}}\\) do not depend on data \\(\\mathbf{X}\\), we denote \\(\\kappa (\\mathbf{W}_{\\mathrm{down}}) = C_{\\mathrm{down}}\\) and \\(\\kappa (\\mathbf{W}_{\\mathrm{up}}) = C_{\\mathrm{up}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Compared to the self-attention mechanism, the MLP transformation does not worsen conditioning as severely, since it has a much lower bound on the condition number. In Fig. 1 (b), the output embeddings of the FFN without" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23118" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "skip connection is observed to have a significantly better condition than the output embedding of the SAB without skip connection. Therefore, in (c), the performance drop after removing the FFN's skip connections is less severe than after removing the SAB's skip connections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.177, + 0.402, + 0.193 + ], + "angle": 0, + "content": "4.2. Skip connection improves condition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.484, + 0.245 + ], + "angle": 0, + "content": "In this subsection, we theoretically demonstrate that the skip connection improves the condition of the self-attention output embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Previous works [4, 10, 18] have demonstrated that transformers, without skip connections, converge very slowly or may even become un-trainable. In [10], the authors show that without skip connections, output embeddings of the SAB quickly converge to rank 1. They hypothesize that this is due to the inability of the transformer to propagate signals through the self-attention mechanism. This rank-1 output embedding implies an infinite condition number. The authors introduce an inductive bias into the self-attention mechanism to enable what they refer to as \"better signal propagation\". While their approach matches the performance of standard (vanilla) models, it incurs a substantial computational cost—running approximately five times slower. We hypothesize that the success of this method could be directly attributed to its ability to improve the condition of the self-attention output embeddings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.533 + ], + "angle": 0, + "content": "This observation further motivates our view that the primary role of skip connections may be to regularize the condition of the self-attention output embeddings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.542, + 0.485, + 0.633 + ], + "angle": 0, + "content": "Proposition 4.2. Let \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d}\\) and \\(\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}\\) have entries that are independently drawn from a distribution with zero mean. We define \\(\\mathbf{M} = \\mathbf{W}_{\\mathbf{Q}} \\mathbf{W}_{\\mathbf{K}}^{\\mathrm{T}} \\mathbf{X}^{\\mathrm{T}} \\mathbf{X} \\mathbf{W}_{\\mathbf{V}}\\) and assume \\(\\mathbf{M}\\) is the positive semi-definite matrix. We have the following bounds on the condition numbers \\(\\kappa\\)." + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.645, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\kappa (\\mathbf {X M} + \\mathbf {X}) \\ll \\kappa (\\mathbf {X M}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Proposition 4.2 shows that during training, the condition number of the SAB output embedding has a much lower bound than that of the SAB output embedding without skip connections. Furthermore, with other conditions being equal, better condition of SAB can contribute to faster convergence and more stable gradient updates, reducing the risk of gradient explosion or vanishing. This ensures the signal can propagate well during the training process in the Vision Transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In other words, the self-attention mechanism enables tokens to communicate by focusing on different parts of the input, thereby capturing correlations between elements within the same sequence. However, one critical drawback is that it is ill-conditioned and challenging to train. In contrast, although skip connections do not facilitate communi" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.091, + 0.743, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 SVD Token Graying" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.11, + 0.907, + 0.345 + ], + "angle": 0, + "content": "1: \\(\\triangleright\\) Input: training data, amplification coefficient \\(\\epsilon \\in (0,1]\\) \n2: \\(\\triangleright\\) Training: \n3: for each minibatch \\(\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k\\) do \n4: Convert image \\(\\mathbf{x}_i\\) into token \\(\\mathbf{X}\\) \n5: \\(\\mathbf{U\\Sigma V^T} \\leftarrow \\mathbf{X}\\) \\(\\triangleright\\) Compute SVD \n6: \\(\\tilde{\\Sigma} \\leftarrow \\frac{\\Sigma}{\\max(\\Sigma)}\\) \\(\\triangleright\\)Normalize elements \\(\\in (0,1]\\) \n7: \\(\\tilde{\\Sigma} \\leftarrow \\tilde{\\Sigma}^\\epsilon\\) \\(\\triangleright\\)Amplify \n8: \\(\\tilde{\\mathbf{X}} \\leftarrow \\mathbf{U}\\tilde{\\Sigma}\\mathbf{V^T}\\) \\(\\triangleright\\)SVD reconstruct \n9: Patch Embedding: \\(\\mathbf{Z} \\leftarrow\\) PatchEmbed(\\tilde{\\mathbf{X}}) \n10: Forward Pass: \\(\\hat{y} \\leftarrow\\) Model(\\mathbf{Z}) \n11: Compute Loss: \\(\\mathcal{L} \\leftarrow\\) Loss(\\hat{y},y_i) \n12: Backward Pass: Compute gradients \\(\\nabla_{\\theta}\\mathcal{L}\\) \n13: Parameter Update: \\(\\theta \\leftarrow \\theta - \\eta\\nabla_{\\theta}\\mathcal{L}\\) \n14: end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.402 + ], + "angle": 0, + "content": "cation between tokens, they help regularize the condition of output embeddings and aid in the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.463 + ], + "angle": 0, + "content": "As shown in Fig. 3, empirical results demonstrate that in both the linear and softmax self-attention cases, adding skip connections significantly improves the condition of output embeddings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.478, + 0.649, + 0.496 + ], + "angle": 0, + "content": "5. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.565 + ], + "angle": 0, + "content": "In this section, we introduce Token Graying (TG), a simple yet effective method that employs Singular Value Decomposition and Discrete Cosine Transform techniques to reconstruct better-conditioned input tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.836, + 0.592 + ], + "angle": 0, + "content": "5.1. Singular Value Decomposition (SVD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Singular Value Decomposition of a matrix \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d}\\) is the factorization of \\(\\mathbf{X}\\) into the product of three matrices \\(\\mathbf{X} = \\mathbf{U} \\boldsymbol{\\Sigma} \\mathbf{V}^{\\mathrm{T}}\\) where the columns of \\(\\mathbf{U} \\in \\mathbb{R}^{n \\times n}\\) and \\(\\mathbf{V} \\in \\mathbb{R}^{d \\times d}\\) are orthonormal and \\(\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{n \\times d}\\) is a rectangular diagonal matrix whose diagonal elements are non-negative in descending order and represent the singular values such that \\(\\boldsymbol{\\Sigma}_{ii} = \\sigma_i\\). As observed in Eq. (4), we can reduce the condition number by increasing the minimal singular value while keeping the maximal singular value unchanged. Hence, our goal is to reconstruct \\(\\tilde{\\mathbf{X}} = \\mathbf{U} \\boldsymbol{\\tilde{\\Sigma}} \\mathbf{V}^{\\mathrm{T}}\\), where \\(\\boldsymbol{\\tilde{\\Sigma}}\\) is formed by amplifying non-maximal elements of \\(\\boldsymbol{\\Sigma}\\) while keeping the maximal element unchanged. Our SVD token graying pipeline is presented in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.9 + ], + "angle": 0, + "content": "However, directly applying SVD to matrices is computationally expensive, with a cost of \\( O(nd \\cdot \\min(n, d)) \\) leading to longer training time. Tab. 1 demonstrates that training a ViT is significantly slower when using SVD reconstruction. Therefore, in the next subsection, we introduce a more efficient Token Graying (TG) method using the discrete cosine transform." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23119" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.09, + 0.481, + 0.132 + ], + "angle": 0, + "content": "
MethodsViT-BViT-B + SVDTGViT-B + DCTTG
Time (days)0.7234.5520.732
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.142, + 0.483, + 0.171 + ], + "angle": 0, + "content": "Table 1. ViT-Base training time w/o and w/ token graying methods on ImageNet-1K dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.195, + 0.388, + 0.211 + ], + "angle": 0, + "content": "5.2. Discrete Cosine Transform (DCT)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.277 + ], + "angle": 0, + "content": "A Discrete Cosine Transform (DCT) expresses a finite sequence of data points as a sum of cosine functions oscillating at different frequencies. It is real-valued and has an inverse, which is often simply called the Inverse DCT (IDCT)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.278, + 0.484, + 0.339 + ], + "angle": 0, + "content": "There are several variants of the DCT and in this work we choose DCT-II which is one of the most commonly used forms. Given a real-valued sequence data \\(\\mathbf{x} \\in \\mathbb{R}^{N \\times 1}\\), the 1D DCT sequence \\(\\hat{\\mathbf{x}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.361, + 0.484, + 0.465 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mathbf {x}} _ {k} = \\alpha_ {k} \\sum_ {i = 0} ^ {N - 1} x _ {i} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right], \\mathrm {f o r} k = 0, 1, \\dots , N - 1 \\\\ \\text {w h e r e} \\alpha_ {k} = \\left\\{ \\begin{array}{l} \\sqrt {\\frac {1}{N}}, \\text {i f} k = 0 \\\\ \\sqrt {\\frac {2}{N}}, \\text {i f} k \\neq 0 \\end{array} \\right. \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.513 + ], + "angle": 0, + "content": "Simplicity, 1D DCT is a linear transformation, Eq. (8) can be written with the formula \\(DCT(\\mathbf{x}) = \\hat{\\mathbf{x}} = D\\mathbf{x}\\), where \\(D_{i,k}\\) is the 1D DCT basis and orthogonal." + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.533, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nD _ {i, k} = \\alpha_ {k} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.756 + ], + "angle": 0, + "content": "For the reconstruction process, since the DCT expresses a signal as a sum of cosine functions with different frequencies, the IDCT combines these frequency components to recover the original signal. Hence, \\(\\mathrm{IDCT}(\\hat{x}) = \\tilde{\\mathbf{x}} = D^T\\hat{\\mathbf{x}}\\). In our case, we have a token matrix \\(\\mathbf{X}\\in \\mathbb{R}^{n\\times d}\\) and we generalize the 1D DCT to a 2D DCT using the formula \\(\\hat{\\mathbf{X}} = D\\mathbf{X}D^{T}\\). In natural images, the dominant singular vector, associated with the largest singular value, typically captures the low frequency content of the image. Intuitively, this observation suggests that DCT reconstruction can serve as an approximation for SVD reconstruction. Our DCT token graying pipeline is presented in Algorithm 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.817 + ], + "angle": 0, + "content": "DCT avoids the expensive computational cost of SVD and has a complexity of \\( O(nd \\cdot \\log(nd)) \\). Tab. 1 shows that training ViT using the DCT algorithm is significantly faster than SVD algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.224, + 0.848 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have emerged as powerful models in the field of computer vision, demonstrating remarkable performance across a variety of tasks. This section is" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.743, + 0.107 + ], + "angle": 0, + "content": "Algorithm 2 DCT Token Graying" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.111, + 0.907, + 0.356 + ], + "angle": 0, + "content": "1: \\(\\triangleright\\) Input: training data, amplification coefficient \\(\\epsilon \\in (0,1]\\) \n2: Training: \n3: for each minibatch \\(\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k \\sim\\) training data do \n4: Convert image \\(\\mathbf{x}_i\\) into token \\(\\mathbf{X}\\) \n5: \\(D\\hat{\\mathbf{X}}D^T \\leftarrow \\mathbf{X}\\) \\(\\triangleright\\) Compute DCT \n6: \\(\\hat{\\mathbf{X}}_{\\mathrm{norm}} \\leftarrow |\\hat{\\mathbf{X}}| / \\max(|\\hat{\\mathbf{X}}|)\\) \\(\\triangleright\\)Normalize elements \\(\\in (0,1]\\) \n7: \\(\\hat{\\mathbf{X}} \\leftarrow \\hat{\\mathbf{X}}_{\\mathrm{norm}}^\\epsilon \\cdot \\mathrm{sign}(\\hat{\\mathbf{X}}) \\cdot \\max(|\\hat{\\mathbf{X}}|)\\) \\(\\triangleright\\)Amplify \n8: \\(\\hat{\\mathbf{X}} \\leftarrow \\hat{D}^T\\hat{\\mathbf{X}}\\hat{D}\\) \\(\\triangleright\\)Inverse DCT \n9: Patch Embedding: \\(\\mathbf{Z} \\leftarrow\\) PatchEmbed(\\(\\hat{\\mathbf{X}}\\)) \n10: Forward Pass: \\(\\hat{y} \\leftarrow\\) Model(\\(\\mathbf{Z}\\)) \n11: Compute Loss: \\(\\mathcal{L} \\leftarrow\\) Loss(\\(\\hat{y}, y_i\\)) \n12: Backward Pass: Compute gradients \\(\\nabla_\\theta \\mathcal{L}\\) \n13: Parameter Update: \\(\\theta \\leftarrow \\theta - \\eta \\nabla_\\theta \\mathcal{L}\\) \n14: end for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.443 + ], + "angle": 0, + "content": "dedicated to validating and analyzing our proposed algorithm for both supervised and self-supervised learning in ViTs. For all experiments, unless specified otherwise, we use \\(\\epsilon = 0.95\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.738, + 0.471 + ], + "angle": 0, + "content": "6.1. Supervised learning ViT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.537 + ], + "angle": 0, + "content": "In this subsection, we validate our methods in supervised learning using different types and scales of ViTs. Cross-entropy loss objective is used. We use PyTorch and Timm libraries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.697 + ], + "angle": 0, + "content": "- ViT [5]: is the pioneering work that interprets an image as a sequence of patches and processes it by a standard Transformer encoder as used in NLP. This simple, yet scalable, strategy works surprisingly well when coupled with pretraining on large datasets. We train the ViT-Tiny, which has 12 layers and 3 heads, with a head dimension of 64 and a token dimension of 192 on the Tiny-[ImageNet] dataset [14]. Additionally, we train ViT-Base, which consists of 12 layers and 12 heads, each with a head dimension of 64 and a token dimension of 768 on the ImageNet-1K dataset [2]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.906, + 0.795 + ], + "angle": 0, + "content": "- Swin [17] is a hierarchical vision transformer designed to focus on local attention first and expand globally. It improves computational efficiency and locality modeling compared to standard ViTs. We train on the Swin small model with a patch size of 4 and a window size of 7 on the ImageNet-1K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.906, + 0.863 + ], + "angle": 0, + "content": "- CaiT [27] is a variant Vision Transformer that aims at increasing the stability of the optimization when ViTs go deeper. We train CaiT small, which has 24 layers, on the ImageNet-1K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "- PVT [31] an enhanced version of the Pyramid Vision Transformer that utilizes a hierarchical architecture to ex" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23120" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.102, + 0.498, + 0.231 + ], + "angle": 0, + "content": "
MethodTop-1 Acc (%)Top-5 Acc (%)
1/k(QK)43.062.7
1/k(QK) + SVD44.368.8
1/k(QK) + DCT44.268.8
softmax(QK)43.067.3
softmax(QK) + SVD44.769.7
softmax(QK) + DCT44.869.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.24, + 0.485, + 0.31 + ], + "angle": 0, + "content": "Table 2. ViT-Tiny results on Tiny-ImageNet datasets. In the upper part, we show results for linear self-attention, where \\( k = 16 \\) is chosen based on [26] w/ and w/o our method. On the lower part, we show regular softmax self-attention results w/ and w/o our method." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.328, + 0.48, + 0.516 + ], + "angle": 0, + "content": "
MethodεTop-1 Acc (%)κinκout
ViT-Base-81.06.726.74
ViT-Base + SVDTG0.981.26.646.66
0.881.26.476.66
0.781.46.156.17
0.681.45.735.71
0.581.05.295.25
ViT-Base + DCTTG0.9781.16.476.49
0.9581.36.426.43
0.9381.26.336.35
0.981.16.256.25
0.8581.06.016.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.584 + ], + "angle": 0, + "content": "Table 3. ViT-Base results using SVD and DCT token graying methods with different amplification coefficient \\(\\epsilon\\). \\(\\kappa_{\\mathrm{in}}\\) and \\(\\kappa_{\\mathrm{out}}\\) are the average token condition numbers in log scale before and after each SAB." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.667 + ], + "angle": 0, + "content": "tract multi-scale features for improved performance in computer vision tasks. It incorporates efficient attention mechanisms and refined positional encoding to boost accuracy and reduce computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Results. In Tab. 2, we first demonstrate the results for the linear case and the regular softmax case using our methods on a small model (ViT-Tiny) and the small-scale TinyImageNet dataset. Given that Theorem Proposition 4.2 and Proposition Proposition 4.1 are mathematically explained using linear self-attention, we train the ViT-Tiny model with linear self-attention, which aligns well with our analysis in Fig. 3. Furthermore, as shown in Fig. 4, both SVDTG and DCTTG lead to better conditioning of SAB output embeddings throughout training epochs. Next, we train the model on a larger dataset using the default softmax self-attention, further validating that our approach generalizes to softmax self-attention. Moreover, SVDTG is the most direct method to improve the conditioning of a matrix, and as demonstrated in Tab. 3, it leads to notable performance gains." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.42, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Figure 4. We compute the condition number of the SAB block's output embeddings during ViT-Tiny training using SVDTG and DCTTG with different \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.689 + ], + "angle": 0, + "content": "To address the high computational cost—approximately six times slower than regular ViT training—we approximate SVDTG using DCTTG, which achieves better token embeddings conditioning and yields comparable results too. As shown in Fig. 5, our DCTTG method produces condition numbers across all layers that indicate better token conditioning in almost every layer. Finally, we evaluate our methods on different variants of ViTs trained on ImageNet-1K datasets, where our DCTTG method demonstrates superior performance compared to vanilla models in terms of Top-1 and Top-5 classification accuracy, while also achieving better conditioning of SAB output embeddings across all layers in the ViT-B model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.702, + 0.771, + 0.719 + ], + "angle": 0, + "content": "6.2. Self-supervised learning ViT" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.906, + 0.833 + ], + "angle": 0, + "content": "In this subsection, we validate our methods on self-supervised learning models as well as on fine-tuning pretrained models [7, 12]. Self-supervised learning is a powerful technique in which models learn representations directly from unlabeled data. Masked Image Modeling (MIM) [12] learns representations by reconstructing images that have been corrupted through masking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Pre-training. In the pre-training stage, we pre-train our model on the ImageNet-1k training set without using ground-truth labels in a self-supervised manner. We use a batch size of 2048, a base learning rate of \\(1.5 \\times 10^{-4}\\), and a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "23121" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.499, + 0.297 + ], + "angle": 0, + "content": "
MethodTop-1 Acc (%)Top-5 Acc (%)
ViT-S80.295.1
ViT-S + DCTTG80.495.2
ViT-B81.095.3
ViT-B + DCTTG81.395.4
Swin-S81.395.6
Swin-S + DCTTG81.695.6
CaiT-S82.696.1
CaiT-S + DCTTG82.796.3
PVT V2 b382.996.0
PVT V2 b3 + DCTTG83.096.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.337 + ], + "angle": 0, + "content": "Table 4. Top-1 and Top-5 classification accuracy on ImageNet-1K dataset using DCT token graying on different ViTs configurations." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.349, + 0.48, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.483, + 0.575 + ], + "angle": 0, + "content": "Figure 5. Condition numbers of all layers in the trained ViT-Base model with and without the DCTTG method." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.595, + 0.472, + 0.659 + ], + "angle": 0, + "content": "
MethodTop-1 Acc (%)Top-5 Acc (%)
MAE83.096.4
MAE + DCTTG83.296.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.711 + ], + "angle": 0, + "content": "Table 5. MAE pretrained ViT-Base results finetuned on ImageNet dataset-1k. We evaluate our MSVR method on self-attention using the power iteration method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.786 + ], + "angle": 0, + "content": "weight decay of 0.05, and we train for 400 epochs. Our loss function computes the mean squared error (MSE) between the reconstructed and original images in pixel space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Finetuning and reuslts. In the finetuning stage, we fine-tune the pre-trained models on ImageNet-1k for classification tasks using cross-entropy loss. We use a batch size of 512, a base learning rate of \\(5 \\times 10^{-4}\\), a weight decay of 0.05, and train for 100 epochs. As shown in Tab. 5, our DCTTG method outperforms the vanilla MAE model by \\(0.2\\%\\) in both top-1 and top-5 classification accuracy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.628, + 0.106 + ], + "angle": 0, + "content": "7. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Our work establishes a theoretical foundation demonstrating why skip connections act as essential conditioning regularizers within self-attention blocks. Additionally, we introduce a method designed to further enhance the output embeddings of self-attention blocks. However, certain limitations remain. For instance, while our regularizer is effective, training in low-precision settings may pose challenges due to the DCT operation, which involves many multiplications and summations that are sensitive to quantization errors. Additionally, for some variants of ViTs, the performance improvement might be marginal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.634, + 0.314 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.596 + ], + "angle": 0, + "content": "This paper provides a theoretical demonstration that the output embeddings of linear self-attention blocks, when lacking skip connections, are inherently ill-conditioned, a finding that empirically extends to conventional softmax self-attention blocks as well. Furthermore, we show that skip connections serve as a powerful regularizer by significantly improving the conditioning of self-attention blocks' output embeddings. In addition, we use the condition of the self-attention output embeddings as a proxy for that of its Jacobian, empirically demonstrating that input tokens with improved conditioning lead to a better-conditioned Jacobian. Building on this insight, we introduce SVDTG and DCTTG, simple yet effective methods designed to improve the condition of input tokens. We then validate their effectiveness on both supervised and unsupervised learning tasks. Ultimately, we hope these insights will guide the vision community in designing more effective and efficient ViT architectures for downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.602, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the Australian Institute for Machine Learning (AIML) and the CSIRO's Data61 Embodied AI Cluster." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.662, + 0.611, + 0.679 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.687, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[1] Paul Albert, Frederic Z Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Randlora: Full-rank parameter-efficient fine-tuning of large models. In International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.758, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[2] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.816, + 0.905, + 0.883 + ], + "angle": 0, + "content": "[3] Linhao Dong, Shuang Xu, and Bo Xu. Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5884-5888, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[4] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.687, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23122" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.136, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.209, + 0.483, + 0.25 + ], + "angle": 0, + "content": "[6] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[7] Maryam Haghighat, Peyman Moghadam, Shaheer Mohamed, and Piotr Koniusz. Pre-training with Random Orthogonal Projection Image Modeling. In International Conference on Learning Representations (ICLR), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.483, + 0.364 + ], + "angle": 0, + "content": "[8] Stephen Hausler and Peyman Moghadam. Pair-VPR: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers. IEEE Robotics and Automation Letters, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.367, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[9] Soufiane Hayou, Eugenio Clerico, Bobby He, George Deligiannidis, Arnaud Doucet, and Judith Rousseau. Stable resnet. In International Conference on Artificial Intelligence and Statistics, pages 1324-1332. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.425, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[10] Bobby He, James Martens, Guodong Zhang, Aleksandar Botev, Andrew Brock, et al. Deep transformers without shortcuts: Modifying self-attention for faithful signal propagation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.554, + 0.482, + 0.623 + ], + "angle": 0, + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.626, + 0.482, + 0.681 + ], + "angle": 0, + "content": "[13] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. In The International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.684, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[14] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.714, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[15] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[16] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[18] Lorenzo Noci, Chuning Li, Mufan Li, Bobby He, Thomas Hofmann, Chris J Maddison, and Dan Roy. The shaped" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "transformer: Attention models in the infinite depth-andwidth limit. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[19] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[20] Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[22] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, et al. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[23] Jason Ramapuram, Federico Danieli, Eeshan Gunesh Dhekane, Floris Weers, Dan Busbridge, et al. Theory, analysis, and best practices for sigmoid self-attention. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[24] Hemanth Saratchandran and Simon Lucey. Enhancing transformers through conditioned embedded tokens. arXiv preprint arXiv:2505.12789, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[25] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[26] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking softmax: Self-attention with polynomial activations, 2024. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[27] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 32-42, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[28] Asher Trockman and J. Zico Kolter. Patches are all you need?, 2022. 2, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[30] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in neural information processing systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[32] Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks. In The International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23123" + } + ] +] \ No newline at end of file diff --git a/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_origin.pdf b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..56d89521e171ac21580cc9e46cdca835bc19b71d --- /dev/null +++ b/2025/Always Skip Attention/569226ba-e747-4584-b317-a54b6e4497d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c133420f433dd92cef40ad44b238ec06dc34e17d04b8fa89e6f52a8c02a907e0 +size 677191 diff --git a/2025/Always Skip Attention/full.md b/2025/Always Skip Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..378e048feb0f90c59c6a5f8b378a570ba2d8dc1c --- /dev/null +++ b/2025/Always Skip Attention/full.md @@ -0,0 +1,336 @@ +# Always Skip Attention + +Yiping Ji $^{1,2}$ , Hemanth Saratchandran $^{1}$ , Peyman Moghadam $^{2,3}$ , Simon Lucey $^{1}$ $^{1}$ Adelaide University, $^{2}$ CSIRO, $^{3}$ Queensland University of Technology + +{yiping.ji, hemanth.saratchandran, simon.lucy}@adelaide.edu.au, {yiping.ji, peyman.moghadam}@csiro.au + +![](images/d557aff84ed6bb07db54af823bbea325dbabb942af77bd77072f197d35abd6e9.jpg) +(a) Architecture + +![](images/79fcdb7e27381b66ffa8d51cf6358191763ed54e29d4133d72f97258dcdf90d2.jpg) +(b) FFN and SAB output embedding condition number using different configurations + +![](images/d73e83f7d50976ffadb9c8aeb611c2f79b43596064c42de14eba1028f5e5c199.jpg) +(c) Performance after removing skip connections +Figure 1. Removing skip connections from the Self-Attention Block (SAB) versus the Feedforward Network (FFN) in Vision Transformer (ViT) models results in different performance drop. (a) Model architecture of ViT consisting of a SAB, and a FFN in each layer. $①$ represents the output of the self-attention transformation, while $③$ represents the output of the MLP transformation. $②$ and $④$ represent identity mappings, referred to as skip connections, in the SAB and FFN, respectively. (b) The condition number of the different output embeddings for each hidden layer. A ViT-Tiny model is trained in a typical manner with skip connections on both SAB and FFN blocks (i.e., $① + ② + ③ + ④$ ). It shows that the SAB output embedding without skip connection is poorly conditioned $(①)$ , with a condition number significantly higher than that of the other three configurations involving ViT blocks. (c) Classification results on three different datasets using the ViT-Tiny model under three different architectures at train time. The result shows the effects of removing skip connections at train time on the SAB and FFN blocks (either $②$ or $④$ ) in all layers of ViT-Tiny. We observe that removing $②$ results in a catastrophic performance drop, whereas removing $④$ (as predicted) has a modest impact on performance. This difference becomes more pronounced as the scale of the dataset increases. + +# Abstract + +We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (e.g., CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill + +conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying (TG), a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods. + +# 1. Introduction + +Vision Transformers (ViTs) have demonstrated impressive performance on various computer vision and robotic tasks, such as object detection, semantic segmentation, video understanding, visual-language learning, and many oth- + +ers [8, 12, 17, 19, 33]. Much of this success is attributed to the self-attention mechanism, commonly referred to as Self-Attention Block (SAB), which allows ViTs to selectively attend to relevant parts of the input sequence when generating each element of the output sequence. Another essential component is the MLP, or Feedforward Networks (FFN), which facilitates intra-token communication across channel dimensions (see Fig. 1). Finally, identity mappings, commonly referred to as skip connections, are incorporated into both the SAB and FFN to enhance model performance. + +We argue that the Jacobian of the SAB is disproportionately ill-conditioned compared to other components, notably the FFN block. A poorly conditioned ${}^{1}$ Jacobian is fundamentally detrimental to gradient descent training [1, 13, 20, 24, 25], impeding both convergence and stability. A natural strategy to address this issue is to explicitly model the Jacobian of the entire network. However, this approach is impractical due to the sheer scale of modern transformer architectures. To proceed, we introduce the following simplifying assumptions: + +# Assumptions + +1. The condition of the network Jacobian is bound by the most ill-conditioned sub-block Jacobian. +2. A proxy for the condition of the sub-block Jacobian is the condition of the output embedding of that block. + +Although we do not provide formal proofs for these assumptions, we offer indirect validation through extensive empirical analysis. First, we show that better conditioning of a sub-block's output embeddings consistently correlates with improved empirical performance across multiple ViT benchmarks. Second, we demonstrate that this improvement in embedding condition is accompanied by a measurable enhancement in the sub-block's Jacobian condition + +Skip in Transformers. Skip connections have become an indispensable component in modern transformer architectures [29] and have been empirically proven to be a de facto component. In Fig. 1 (b), we analyze a pre-trained ViT-Tiny model by measuring the condition numbers of the SAB and FFN output embeddings, both with and without skip connections. Our observations reveal that the SAB output embedding becomes highly ill-conditioned in the absence of skip connections, with a condition number approaching $e^6$ . In contrast, the other three configurations have relatively low condition numbers, around $e^3$ . In other words, the SAB output embedding without skip connections has significantly worse condition than the other configurations. To further investigate the impact of skip connections, we train the ViT-Tiny models from scratch under three config- + +![](images/1c911ac0c9883163a64b9628d1eba9dd28af0452d086b6e5092eec04d10a79de.jpg) +Figure 2. Training loss for three different configurations of ViT-Tiny models trained on the Tiny-ImageNet dataset. + +urations: the standard SAB and FFN with skip connections, SAB without skip connections, and FFN without skip connections. + +In Fig. 1 (c), we observe a modest drop in classification accuracy $(2\%)$ a ViT-Tiny model trained on the CIFAR-10 dataset when skip connections are removed from the FFN while retained in the SAB. However, removing skip connections from the SAB, while keeping them in the FFN, leads to a near-catastrophic performance drop of approximately $22\%$ . Moreover, as dataset size increases, models without SAB skip connections degrade substantially in performance; on larger-scale datasets such as Tiny-[ImageNet, these models fail to train altogether. To the best of our knowledge, this empirical finding has not been reported in prior literature. Additionally, the training loss curves for the three configurations are shown in Fig. 2. The results demonstrate that the SAB without skip connections, which is significantly ill-conditioned, exhibits much slower convergence and diverges after 30 epochs. Interestingly, this critical dependence on skip connections is unique to transformer architectures. + +Skip in ConvMixer. To further highlight the poor conditioning of the SAB, we conducted a similar experiment on a modern CNN ConvMixer [28]. ConvMixer has a similar architecture and competitive performance to ViT with the exception that self-attention is replaced by a convolution block to spatially mix tokens. Unlike ViT, training ConvMixer-Tiny with and without skip connections, we observed that the performance remains effectively unchanged, with variations of less than $\pm 0.2\%$ . Further details are shown in Fig. 6 in Appendix. This contrast further suggests that the SAB introduces an extreme conditioning issue, making skip connections indispensable for stable training in ViTs. + +Our main contributions are summarized as follows. + +1. We present a proposition that characterizes why SAB output embedding without skip connection is funda + +mentally ill-conditioned, which challenges training convergence and stability (Sec. 4.1). + +2. A theoretical analysis on the role of skip connection within the SAB is undertaken. We demonstrate that it significantly improves the condition of the block's output embedding, enhancing stability and performance (Sec. 4.2). +3. Finally, we propose a novel approach—Token Greying (TG)—to better pre-condition the input tokens. This step, when used in conjunction with conventional skip connections—improves ViT in both supervised and self-supervised settings (Sec. 5). + +Our central contribution in this paper is not TG itself, but the insight that the SAB within modern ViTs is intrinsically ill-conditioned. Our hope is that this insight opens up brand new lines of inquiry for the vision community for more effective and efficient ViT design. + +# 2. Related work + +Self-Attention. Originally proposed by Vaswani et al. [29], the self-attention mechanism has become central to transformer architectures, enabling the capture of long-range dependencies by dynamically weighting interactions between tokens because it captures long-range dependencies by dynamically weighting interactions between tokens. Building on this foundation, transformer-based large language models, such as Llama [6] and Deepseek [15], have significantly advanced language understanding. In computer vision, researchers introduced Vision Transformers (ViTs) [5] for image classification, achieving superior results compared to traditional convolutional neural networks. Moreover, researchers have trained Diffusion Transformers [21], replacing the commonly used U-Net backbone with a transformer that operates on latent patches and inherits the excellent scaling properties of the transformer model class. Similarly, self-attention mechanisms have been successfully applied in speech recognition and physics-informed neural networks, further demonstrating their versatility across diverse applications [3, 32]. + +Skip connections. He et al. [11] introduced skip connections in ResNet to make deep networks easier to optimize and to enhance accuracy by enabling significantly increased network depth. Since their introduction, skip connections have been extensively studied and applied across various Convolutional Neural Network (CNN) architectures. Building on these ideas, Veit et al. [30] offer a novel interpretation of residual networks, demonstrating that they can be viewed as a collection of multiple paths of varying lengths, which contributes to the network's flexibility and resilience. Furthermore, Hayou et al. [9] addresses gradient stability in deep ResNets, noting that while skip connections alleviate the issue of vanishing gradients, they may also intro- + +duce gradient explosion at extreme depths. To counter this, they propose scaling the skip connections according to the layer index, thereby stabilizing gradients and ensuring network expressivity even in the limit of infinite depth. Recently, in Table 3 of [16], the authors reported that VGG—a model without skip connections—achieves comparable results to ResNet—a model with skip connections—on modern datasets. To the best of our knowledge, only one paper challenges the skip connection in transformer architectures. In [10], the authors attempted to train deep transformer networks without using skip connections, achieving performance comparable to standard models. However, this approach requires five times more iterations, and the reasons why skip connections are crucial for self-attention-based transformers remain unresolved. + +# 3. Preliminary + +In this section, we define the self-attention blocks used in the Vision Transformer (ViT) architecture and establish the mathematical notation for key quantities referenced in subsequent sections. For a comprehensive overview of transformers, we refer readers to [5]. + +A Vision Transformer architecture consists of $L$ stacked self-attention blocks (SABs) and feedforward networks (FFNs). An identity mapping, commonly referred to as a skip connection, is applied to connect the inputs and outputs of the transformation in both SAB and FFN. + +Self-Attention Blocks in ViT. Formally, given an input sequence $\mathbf{X}_{\mathrm{in}} \in \mathbb{R}^{n \times d}$ , with $n$ tokens of dimension $d$ , a SAB is defined as: + +$$ +\mathbf {X} _ {\text {o u t}} := \operatorname {S A} \left(\mathbf {X} _ {\text {i n}}\right) + \mathbf {X} _ {\text {i n}}, \tag {1} +$$ + +where the self-attention output embedding is + +$$ +\operatorname {S A} \left(\mathbf {X} _ {\text {i n}}\right) = \eta \left(\mathbf {Q K} ^ {\mathbf {T}}\right) \mathbf {V}, \tag {2} +$$ + +where $\mathbf{Q} = \mathbf{X}_{\mathrm{in}}\mathbf{W}_{\mathrm{Q}},\mathbf{K} = \mathbf{X}_{\mathrm{in}}\mathbf{W}_{\mathrm{K}}$ and $\mathbf{V} = \mathbf{X}_{\mathrm{in}}\mathbf{W}_{\mathrm{V}}$ with learnable parameters $\mathbf{W}^{\mathrm{Q}},\mathbf{W}^{\mathrm{K}}$ , and $\mathbf{W}^{\mathrm{V}}\in \mathbb{R}^{d\times d}$ . $\eta (\cdot)$ is an activation function and is set by default as softmax function in [5]. Additionally, some studies [22, 23, 26] have explored the possibilities of using alternative activation functions. In particular, good performance is achieved by using linear attention with an appropriate scale. It allows us to characterize why the SAB output embedding without a skip connection is ill-conditioned. + +For general vision transformers, multiple heads $\mathrm{SA}(\mathbf{X}_{\mathrm{in}})_i$ , where $1 \leq i \leq h$ , are commonly used. Learnable matrices are divided into head numbers $h$ , such that $\mathbf{W}_{\mathrm{Q}}, \mathbf{W}_{\mathrm{K}}$ , and $\mathbf{W}_{\mathrm{V}} \in \mathbb{R}^{h \times d \times d_h}$ , where $d_h = \frac{d}{h}$ . Then all outputs of each attention head are concatenated together before the skip connection. + +Feed Forward Networks in ViT. Formally, given an input sequence $\mathbf{X}_{\mathrm{in}} \in \mathbb{R}^{n \times d}$ , with $n$ tokens of dimension $d$ , an FFN is defined as: + +$$ +\mathbf {X} _ {\text {o u t}} := \mathbf {W} _ {\text {d o w n}} \left(g \left(\mathbf {W} _ {\text {u p}} \left(\mathbf {X} _ {\text {i n}}\right)\right)\right) + \mathbf {X} _ {\text {i n}}, \tag {3} +$$ + +where $\mathbf{W}_{\mathrm{up}}\in \mathbb{R}^{d\times 4d}$ is up projection, $\mathbf{W}_{\mathrm{down}}\in \mathbb{R}^{4d\times d}$ is down projection, and $g(\cdot)$ is an activation function. + +Conditioning of matrices. Formally, for a rectangular full rank matrix $\mathbf{A} \in \mathbb{R}^{n \times d}$ , the condition number of $\mathbf{A}$ is defined below: + +$$ +\kappa (\mathbf {A}) = \| \mathbf {A} \| _ {2} \| \mathbf {A} ^ {\dagger} \| _ {2} = \frac {\sigma_ {\max } (\mathbf {A})}{\sigma_ {\min } (\mathbf {A})}, \tag {4} +$$ + +where $\| \cdot \| _2$ is the matrix operator norm induced by the Euclidean norm. We use the Euclidean norm in this paper. $\mathbf{A}^{\dagger}$ is the pseudo inverse of the matrix $\mathbf{A}$ . $\sigma_{\mathrm{max}}$ and $\sigma_{\mathrm{min}}$ are the maximal and minimal singular values of $\mathbf{A}$ respectively. We will use Eq. (4) as a metric for analysis in future sections. + +# 4. Theoretical analysis + +In this section, we provide a theoretical analysis showing that the output embedding of the SAB without skip connections is highly ill-conditioned. This matters as the condition of the output embedding is a proxy for the condition of the sub-block's Jacobian. We then demonstrate that an essential function of the skip connection is to improve this conditioning. + +# 4.1. Self-attention is ill-conditioned + +We attempt to validate our claim for the simpler case of linear attention. + +Proposition 4.1. Assume $\mathbf{X} \in \mathbb{R}^{n \times d}$ , $\mathbf{W}_{\mathbf{Q}}$ , and $\mathbf{W}_{\mathbf{K}}$ and $\mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{d \times d}$ have entries that are independently drawn from a distribution with zero mean. The condition number of the SAB output embedding (without skip-connection) can be expressed as: + +$$ +\kappa \left(\mathbf {X} \mathbf {W} _ {Q} \mathbf {W} _ {K} ^ {T} \mathbf {X} ^ {T} \mathbf {X} \mathbf {W} _ {V}\right) \leq C \cdot \left(\frac {\sigma_ {\text {m a x}}}{\sigma_ {\text {m i n}}}\right) ^ {3}, \tag {5} +$$ + +where $\sigma_{\mathrm{max}}$ and $\sigma_{\mathrm{min}}$ are the maximal and minimal singular values of $\mathbf{X}$ . $C$ is a fixed constant that is statistically close to unity. + +Proposition 4.1 assumes the use of linear attention (i.e., no softmax) as advocated for in [26]. This was done to offer a simplified theoretical exposition on how the SAB without skip connection can affect the condition of the output embedding. The proposition shows the condition number of the output embedding without skip connection is bounded + +![](images/5ba6082307532d2f6dab6594eef31d6e2a6cab166bb6691c441b1da5a00ebc4e.jpg) + +![](images/f22015f645c58bf02500a14df6568b8fd01412dadc5a8448f859c10195db50fd.jpg) +Figure 3. Condition numbers across all layers of trained ViT-Tiny on Tiny-ImageNet datasets. We use $10\%$ validation sets to plot the above figures. Upper: fixed scale linear self-attention. Lower: softmax self-attention + +by the cube of the condition number of the input matrix $\mathbf{X}$ . Unless the condition of $\mathbf{X}$ is unity, the output of the SAB without skip connection will therefore result in a poorly conditioned embedding. Applying this same process across multiple layers within a transformer results in a highly ill-conditioned outcome. + +Fig. 3 illustrates our theoretical analysis by showing that the linear self-attention operation increases the condition number of embeddings from $e^3$ to $e^6$ , and we can empirically extend this analysis to softmax self-attention. The output embedding of the softmax SAB without skip connections is also significantly ill-conditioned. + +Further, for condition numbers of FFN output embeddings without skip connection using linear activation $g$ we have: + +$$ +\kappa (\mathbf {M L P} (\mathbf {X})) = \kappa (\mathbf {W} _ {\mathrm {d o w n}} \mathbf {W} _ {\mathrm {u p}} \mathbf {X}) \leq C _ {\mathrm {d o w n}} \cdot C _ {\mathrm {u p}} \cdot \frac {\sigma_ {\operatorname* {m a x}}}{\sigma_ {\operatorname* {m i n}}}, \tag {6} +$$ + +where $\sigma_{\mathrm{max}}$ and $\sigma_{\mathrm{min}}$ are maximal singular values of $\mathbf{X}$ computed using SVD. Since $\mathbf{W}_{\mathrm{down}}$ and $\mathbf{W}_{\mathrm{up}}$ do not depend on data $\mathbf{X}$ , we denote $\kappa (\mathbf{W}_{\mathrm{down}}) = C_{\mathrm{down}}$ and $\kappa (\mathbf{W}_{\mathrm{up}}) = C_{\mathrm{up}}$ . + +Compared to the self-attention mechanism, the MLP transformation does not worsen conditioning as severely, since it has a much lower bound on the condition number. In Fig. 1 (b), the output embeddings of the FFN without + +skip connection is observed to have a significantly better condition than the output embedding of the SAB without skip connection. Therefore, in (c), the performance drop after removing the FFN's skip connections is less severe than after removing the SAB's skip connections. + +# 4.2. Skip connection improves condition + +In this subsection, we theoretically demonstrate that the skip connection improves the condition of the self-attention output embeddings. + +Previous works [4, 10, 18] have demonstrated that transformers, without skip connections, converge very slowly or may even become un-trainable. In [10], the authors show that without skip connections, output embeddings of the SAB quickly converge to rank 1. They hypothesize that this is due to the inability of the transformer to propagate signals through the self-attention mechanism. This rank-1 output embedding implies an infinite condition number. The authors introduce an inductive bias into the self-attention mechanism to enable what they refer to as "better signal propagation". While their approach matches the performance of standard (vanilla) models, it incurs a substantial computational cost—running approximately five times slower. We hypothesize that the success of this method could be directly attributed to its ability to improve the condition of the self-attention output embeddings. + +This observation further motivates our view that the primary role of skip connections may be to regularize the condition of the self-attention output embeddings + +Proposition 4.2. Let $\mathbf{X} \in \mathbb{R}^{n \times d}$ and $\mathbf{W}_{\mathbf{Q}}, \mathbf{W}_{\mathbf{K}}, \mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{d \times d}$ have entries that are independently drawn from a distribution with zero mean. We define $\mathbf{M} = \mathbf{W}_{\mathbf{Q}} \mathbf{W}_{\mathbf{K}}^{\mathrm{T}} \mathbf{X}^{\mathrm{T}} \mathbf{X} \mathbf{W}_{\mathbf{V}}$ and assume $\mathbf{M}$ is the positive semi-definite matrix. We have the following bounds on the condition numbers $\kappa$ . + +$$ +\kappa (\mathbf {X M} + \mathbf {X}) \ll \kappa (\mathbf {X M}). \tag {7} +$$ + +Proposition 4.2 shows that during training, the condition number of the SAB output embedding has a much lower bound than that of the SAB output embedding without skip connections. Furthermore, with other conditions being equal, better condition of SAB can contribute to faster convergence and more stable gradient updates, reducing the risk of gradient explosion or vanishing. This ensures the signal can propagate well during the training process in the Vision Transformers. + +In other words, the self-attention mechanism enables tokens to communicate by focusing on different parts of the input, thereby capturing correlations between elements within the same sequence. However, one critical drawback is that it is ill-conditioned and challenging to train. In contrast, although skip connections do not facilitate communi + +Algorithm 1 SVD Token Graying +1: $\triangleright$ Input: training data, amplification coefficient $\epsilon \in (0,1]$ +2: $\triangleright$ Training: +3: for each minibatch $\{(\mathbf{x}_i,y_i)\}_{i=1}^k$ do +4: Convert image $\mathbf{x}_i$ into token $\mathbf{X}$ +5: $\mathbf{U\Sigma V^T} \leftarrow \mathbf{X}$ $\triangleright$ Compute SVD +6: $\tilde{\Sigma} \leftarrow \frac{\Sigma}{\max(\Sigma)}$ $\triangleright$ Normalize elements $\in (0,1]$ +7: $\tilde{\Sigma} \leftarrow \tilde{\Sigma}^\epsilon$ $\triangleright$ Amplify +8: $\tilde{\mathbf{X}} \leftarrow \mathbf{U}\tilde{\Sigma}\mathbf{V^T}$ $\triangleright$ SVD reconstruct +9: Patch Embedding: $\mathbf{Z} \leftarrow$ PatchEmbed(\tilde{\mathbf{X}}) +10: Forward Pass: $\hat{y} \leftarrow$ Model(\mathbf{Z}) +11: Compute Loss: $\mathcal{L} \leftarrow$ Loss(\hat{y},y_i) +12: Backward Pass: Compute gradients $\nabla_{\theta}\mathcal{L}$ +13: Parameter Update: $\theta \leftarrow \theta - \eta\nabla_{\theta}\mathcal{L}$ +14: end for + +cation between tokens, they help regularize the condition of output embeddings and aid in the optimization process. + +As shown in Fig. 3, empirical results demonstrate that in both the linear and softmax self-attention cases, adding skip connections significantly improves the condition of output embeddings. + +# 5. Methodology + +In this section, we introduce Token Graying (TG), a simple yet effective method that employs Singular Value Decomposition and Discrete Cosine Transform techniques to reconstruct better-conditioned input tokens. + +# 5.1. Singular Value Decomposition (SVD) + +Singular Value Decomposition of a matrix $\mathbf{X} \in \mathbb{R}^{n \times d}$ is the factorization of $\mathbf{X}$ into the product of three matrices $\mathbf{X} = \mathbf{U} \boldsymbol{\Sigma} \mathbf{V}^{\mathrm{T}}$ where the columns of $\mathbf{U} \in \mathbb{R}^{n \times n}$ and $\mathbf{V} \in \mathbb{R}^{d \times d}$ are orthonormal and $\boldsymbol{\Sigma} \in \mathbb{R}^{n \times d}$ is a rectangular diagonal matrix whose diagonal elements are non-negative in descending order and represent the singular values such that $\boldsymbol{\Sigma}_{ii} = \sigma_i$ . As observed in Eq. (4), we can reduce the condition number by increasing the minimal singular value while keeping the maximal singular value unchanged. Hence, our goal is to reconstruct $\tilde{\mathbf{X}} = \mathbf{U} \boldsymbol{\tilde{\Sigma}} \mathbf{V}^{\mathrm{T}}$ , where $\boldsymbol{\tilde{\Sigma}}$ is formed by amplifying non-maximal elements of $\boldsymbol{\Sigma}$ while keeping the maximal element unchanged. Our SVD token graying pipeline is presented in Algorithm 1. + +However, directly applying SVD to matrices is computationally expensive, with a cost of $O(nd \cdot \min(n, d))$ leading to longer training time. Tab. 1 demonstrates that training a ViT is significantly slower when using SVD reconstruction. Therefore, in the next subsection, we introduce a more efficient Token Graying (TG) method using the discrete cosine transform. + +
MethodsViT-BViT-B + SVDTGViT-B + DCTTG
Time (days)0.7234.5520.732
+ +Table 1. ViT-Base training time w/o and w/ token graying methods on ImageNet-1K dataset. + +# 5.2. Discrete Cosine Transform (DCT) + +A Discrete Cosine Transform (DCT) expresses a finite sequence of data points as a sum of cosine functions oscillating at different frequencies. It is real-valued and has an inverse, which is often simply called the Inverse DCT (IDCT). + +There are several variants of the DCT and in this work we choose DCT-II which is one of the most commonly used forms. Given a real-valued sequence data $\mathbf{x} \in \mathbb{R}^{N \times 1}$ , the 1D DCT sequence $\hat{\mathbf{x}}$ is defined as: + +$$ +\begin{array}{l} \hat {\mathbf {x}} _ {k} = \alpha_ {k} \sum_ {i = 0} ^ {N - 1} x _ {i} \cos \left[ \frac {\pi (2 i + 1) k}{2 N} \right], \mathrm {f o r} k = 0, 1, \dots , N - 1 \\ \text {w h e r e} \alpha_ {k} = \left\{ \begin{array}{l} \sqrt {\frac {1}{N}}, \text {i f} k = 0 \\ \sqrt {\frac {2}{N}}, \text {i f} k \neq 0 \end{array} \right. \tag {8} \\ \end{array} +$$ + +Simplicity, 1D DCT is a linear transformation, Eq. (8) can be written with the formula $DCT(\mathbf{x}) = \hat{\mathbf{x}} = D\mathbf{x}$ , where $D_{i,k}$ is the 1D DCT basis and orthogonal. + +$$ +D _ {i, k} = \alpha_ {k} \cos \left[ \frac {\pi (2 i + 1) k}{2 N} \right]. \tag {9} +$$ + +For the reconstruction process, since the DCT expresses a signal as a sum of cosine functions with different frequencies, the IDCT combines these frequency components to recover the original signal. Hence, $\mathrm{IDCT}(\hat{x}) = \tilde{\mathbf{x}} = D^T\hat{\mathbf{x}}$ . In our case, we have a token matrix $\mathbf{X}\in \mathbb{R}^{n\times d}$ and we generalize the 1D DCT to a 2D DCT using the formula $\hat{\mathbf{X}} = D\mathbf{X}D^{T}$ . In natural images, the dominant singular vector, associated with the largest singular value, typically captures the low frequency content of the image. Intuitively, this observation suggests that DCT reconstruction can serve as an approximation for SVD reconstruction. Our DCT token graying pipeline is presented in Algorithm 2. + +DCT avoids the expensive computational cost of SVD and has a complexity of $O(nd \cdot \log(nd))$ . Tab. 1 shows that training ViT using the DCT algorithm is significantly faster than SVD algorithm. + +# 6. Experiments + +Vision Transformers (ViTs) have emerged as powerful models in the field of computer vision, demonstrating remarkable performance across a variety of tasks. This section is + +Algorithm 2 DCT Token Graying +1: $\triangleright$ Input: training data, amplification coefficient $\epsilon \in (0,1]$ +2: Training: +3: for each minibatch $\{(\mathbf{x}_i,y_i)\}_{i=1}^k \sim$ training data do +4: Convert image $\mathbf{x}_i$ into token $\mathbf{X}$ +5: $D\hat{\mathbf{X}}D^T \leftarrow \mathbf{X}$ $\triangleright$ Compute DCT +6: $\hat{\mathbf{X}}_{\mathrm{norm}} \leftarrow |\hat{\mathbf{X}}| / \max(|\hat{\mathbf{X}}|)$ $\triangleright$ Normalize elements $\in (0,1]$ +7: $\hat{\mathbf{X}} \leftarrow \hat{\mathbf{X}}_{\mathrm{norm}}^\epsilon \cdot \mathrm{sign}(\hat{\mathbf{X}}) \cdot \max(|\hat{\mathbf{X}}|)$ $\triangleright$ Amplify +8: $\hat{\mathbf{X}} \leftarrow \hat{D}^T\hat{\mathbf{X}}\hat{D}$ $\triangleright$ Inverse DCT +9: Patch Embedding: $\mathbf{Z} \leftarrow$ PatchEmbed( $\hat{\mathbf{X}}$ ) +10: Forward Pass: $\hat{y} \leftarrow$ Model( $\mathbf{Z}$ ) +11: Compute Loss: $\mathcal{L} \leftarrow$ Loss( $\hat{y}, y_i$ ) +12: Backward Pass: Compute gradients $\nabla_\theta \mathcal{L}$ +13: Parameter Update: $\theta \leftarrow \theta - \eta \nabla_\theta \mathcal{L}$ +14: end for + +dedicated to validating and analyzing our proposed algorithm for both supervised and self-supervised learning in ViTs. For all experiments, unless specified otherwise, we use $\epsilon = 0.95$ . + +# 6.1. Supervised learning ViT + +In this subsection, we validate our methods in supervised learning using different types and scales of ViTs. Cross-entropy loss objective is used. We use PyTorch and Timm libraries. + +- ViT [5]: is the pioneering work that interprets an image as a sequence of patches and processes it by a standard Transformer encoder as used in NLP. This simple, yet scalable, strategy works surprisingly well when coupled with pretraining on large datasets. We train the ViT-Tiny, which has 12 layers and 3 heads, with a head dimension of 64 and a token dimension of 192 on the Tiny-[ImageNet] dataset [14]. Additionally, we train ViT-Base, which consists of 12 layers and 12 heads, each with a head dimension of 64 and a token dimension of 768 on the ImageNet-1K dataset [2]. +- Swin [17] is a hierarchical vision transformer designed to focus on local attention first and expand globally. It improves computational efficiency and locality modeling compared to standard ViTs. We train on the Swin small model with a patch size of 4 and a window size of 7 on the ImageNet-1K dataset. +- CaiT [27] is a variant Vision Transformer that aims at increasing the stability of the optimization when ViTs go deeper. We train CaiT small, which has 24 layers, on the ImageNet-1K dataset. +- PVT [31] an enhanced version of the Pyramid Vision Transformer that utilizes a hierarchical architecture to ex + +
MethodTop-1 Acc (%)Top-5 Acc (%)
1/k(QK)43.062.7
1/k(QK) + SVD44.368.8
1/k(QK) + DCT44.268.8
softmax(QK)43.067.3
softmax(QK) + SVD44.769.7
softmax(QK) + DCT44.869.7
+ +Table 2. ViT-Tiny results on Tiny-ImageNet datasets. In the upper part, we show results for linear self-attention, where $k = 16$ is chosen based on [26] w/ and w/o our method. On the lower part, we show regular softmax self-attention results w/ and w/o our method. + +
MethodεTop-1 Acc (%)κinκout
ViT-Base-81.06.726.74
ViT-Base + SVDTG0.981.26.646.66
0.881.26.476.66
0.781.46.156.17
0.681.45.735.71
0.581.05.295.25
ViT-Base + DCTTG0.9781.16.476.49
0.9581.36.426.43
0.9381.26.336.35
0.981.16.256.25
0.8581.06.016.02
+ +Table 3. ViT-Base results using SVD and DCT token graying methods with different amplification coefficient $\epsilon$ . $\kappa_{\mathrm{in}}$ and $\kappa_{\mathrm{out}}$ are the average token condition numbers in log scale before and after each SAB. + +tract multi-scale features for improved performance in computer vision tasks. It incorporates efficient attention mechanisms and refined positional encoding to boost accuracy and reduce computational complexity. + +Results. In Tab. 2, we first demonstrate the results for the linear case and the regular softmax case using our methods on a small model (ViT-Tiny) and the small-scale TinyImageNet dataset. Given that Theorem Proposition 4.2 and Proposition Proposition 4.1 are mathematically explained using linear self-attention, we train the ViT-Tiny model with linear self-attention, which aligns well with our analysis in Fig. 3. Furthermore, as shown in Fig. 4, both SVDTG and DCTTG lead to better conditioning of SAB output embeddings throughout training epochs. Next, we train the model on a larger dataset using the default softmax self-attention, further validating that our approach generalizes to softmax self-attention. Moreover, SVDTG is the most direct method to improve the conditioning of a matrix, and as demonstrated in Tab. 3, it leads to notable performance gains. + +![](images/aa422d58e87cec12de2196802ebbc98278cc0889416f479d1a5470ae053eabf0.jpg) + +![](images/0b70abd157dd403e423c8e45644f041348a8a1498b175ef3318e083bd2746f44.jpg) +Figure 4. We compute the condition number of the SAB block's output embeddings during ViT-Tiny training using SVDTG and DCTTG with different $\epsilon$ . + +To address the high computational cost—approximately six times slower than regular ViT training—we approximate SVDTG using DCTTG, which achieves better token embeddings conditioning and yields comparable results too. As shown in Fig. 5, our DCTTG method produces condition numbers across all layers that indicate better token conditioning in almost every layer. Finally, we evaluate our methods on different variants of ViTs trained on ImageNet-1K datasets, where our DCTTG method demonstrates superior performance compared to vanilla models in terms of Top-1 and Top-5 classification accuracy, while also achieving better conditioning of SAB output embeddings across all layers in the ViT-B model. + +# 6.2. Self-supervised learning ViT + +In this subsection, we validate our methods on self-supervised learning models as well as on fine-tuning pretrained models [7, 12]. Self-supervised learning is a powerful technique in which models learn representations directly from unlabeled data. Masked Image Modeling (MIM) [12] learns representations by reconstructing images that have been corrupted through masking. + +Pre-training. In the pre-training stage, we pre-train our model on the ImageNet-1k training set without using ground-truth labels in a self-supervised manner. We use a batch size of 2048, a base learning rate of $1.5 \times 10^{-4}$ , and a + +
MethodTop-1 Acc (%)Top-5 Acc (%)
ViT-S80.295.1
ViT-S + DCTTG80.495.2
ViT-B81.095.3
ViT-B + DCTTG81.395.4
Swin-S81.395.6
Swin-S + DCTTG81.695.6
CaiT-S82.696.1
CaiT-S + DCTTG82.796.3
PVT V2 b382.996.0
PVT V2 b3 + DCTTG83.096.1
+ +![](images/5e7e7342aa6c476d69917c3e7a8aa0edcca9ac887a87de8ae65a5728057b630f.jpg) +Figure 5. Condition numbers of all layers in the trained ViT-Base model with and without the DCTTG method. + +Table 4. Top-1 and Top-5 classification accuracy on ImageNet-1K dataset using DCT token graying on different ViTs configurations. + +
MethodTop-1 Acc (%)Top-5 Acc (%)
MAE83.096.4
MAE + DCTTG83.296.6
+ +Table 5. MAE pretrained ViT-Base results finetuned on ImageNet dataset-1k. We evaluate our MSVR method on self-attention using the power iteration method. + +weight decay of 0.05, and we train for 400 epochs. Our loss function computes the mean squared error (MSE) between the reconstructed and original images in pixel space. + +Finetuning and reuslts. In the finetuning stage, we fine-tune the pre-trained models on ImageNet-1k for classification tasks using cross-entropy loss. We use a batch size of 512, a base learning rate of $5 \times 10^{-4}$ , a weight decay of 0.05, and train for 100 epochs. As shown in Tab. 5, our DCTTG method outperforms the vanilla MAE model by $0.2\%$ in both top-1 and top-5 classification accuracy. + +# 7. Limitation + +Our work establishes a theoretical foundation demonstrating why skip connections act as essential conditioning regularizers within self-attention blocks. Additionally, we introduce a method designed to further enhance the output embeddings of self-attention blocks. However, certain limitations remain. For instance, while our regularizer is effective, training in low-precision settings may pose challenges due to the DCT operation, which involves many multiplications and summations that are sensitive to quantization errors. Additionally, for some variants of ViTs, the performance improvement might be marginal. + +# 8. Conclusion + +This paper provides a theoretical demonstration that the output embeddings of linear self-attention blocks, when lacking skip connections, are inherently ill-conditioned, a finding that empirically extends to conventional softmax self-attention blocks as well. Furthermore, we show that skip connections serve as a powerful regularizer by significantly improving the conditioning of self-attention blocks' output embeddings. In addition, we use the condition of the self-attention output embeddings as a proxy for that of its Jacobian, empirically demonstrating that input tokens with improved conditioning lead to a better-conditioned Jacobian. Building on this insight, we introduce SVDTG and DCTTG, simple yet effective methods designed to improve the condition of input tokens. We then validate their effectiveness on both supervised and unsupervised learning tasks. Ultimately, we hope these insights will guide the vision community in designing more effective and efficient ViT architectures for downstream applications. + +Acknowledgements. This work was supported by the Australian Institute for Machine Learning (AIML) and the CSIRO's Data61 Embodied AI Cluster. + +# References + +[1] Paul Albert, Frederic Z Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Randlora: Full-rank parameter-efficient fine-tuning of large models. In International Conference on Learning Representations (ICLR), 2025. 2 +[2] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 6 +[3] Linhao Dong, Shuang Xu, and Bo Xu. Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5884-5888, 2018. 3 +[4] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas. + +Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 5 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 3, 6 +[6] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3 +[7] Maryam Haghighat, Peyman Moghadam, Shaheer Mohamed, and Piotr Koniusz. Pre-training with Random Orthogonal Projection Image Modeling. In International Conference on Learning Representations (ICLR), 2024. 7 +[8] Stephen Hausler and Peyman Moghadam. Pair-VPR: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers. IEEE Robotics and Automation Letters, 2025. 2 +[9] Soufiane Hayou, Eugenio Clerico, Bobby He, George Deligiannidis, Arnaud Doucet, and Judith Rousseau. Stable resnet. In International Conference on Artificial Intelligence and Statistics, pages 1324-1332. PMLR, 2021. 3 +[10] Bobby He, James Martens, Guodong Zhang, Aleksandar Botev, Andrew Brock, et al. Deep transformers without shortcuts: Modifying self-attention for faithful signal propagation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7 +[13] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. In The International Conference on Learning Representations, 2025. 2 +[14] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[15] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 3 +[16] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? In The Thirteenth International Conference on Learning Representations, 2025. 3 +[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6 +[18] Lorenzo Noci, Chuning Li, Mufan Li, Bobby He, Thomas Hofmann, Chris J Maddison, and Dan Roy. The shaped + +transformer: Attention models in the infinite depth-andwidth limit. Advances in Neural Information Processing Systems, 36, 2024. 5 +[19] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2 +[20] Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks, 2013. 2 +[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3 +[22] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, et al. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 3 +[23] Jason Ramapuram, Federico Danieli, Eeshan Gunesh Dhekane, Floris Weers, Dan Busbridge, et al. Theory, analysis, and best practices for sigmoid self-attention. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[24] Hemanth Saratchandran and Simon Lucey. Enhancing transformers through conditioned embedded tokens. arXiv preprint arXiv:2505.12789, 2025. 2 +[25] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2024. 2 +[26] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking softmax: Self-attention with polynomial activations, 2024. 3, 4, 7 +[27] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 32-42, 2021. 6 +[28] Asher Trockman and J. Zico Kolter. Patches are all you need?, 2022. 2, 10 +[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3 +[30] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in neural information processing systems, 29, 2016. 3 +[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 6 +[32] Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks. In The International Conference on Learning Representations (ICLR), 2024. 3 +[33] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 2 \ No newline at end of file diff --git a/2025/Always Skip Attention/images.zip b/2025/Always Skip Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c368e362280449b75b15869445d2667fba8fba31 --- /dev/null +++ b/2025/Always Skip Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3f3379f0880a9568d7303aaaf09afd786c19927f4ac7d11b70499336bc33277 +size 393510 diff --git a/2025/Always Skip Attention/layout.json b/2025/Always Skip Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..137007e2023fb5f2fe3b86027653dea9a0ccbe62 --- /dev/null +++ b/2025/Always Skip Attention/layout.json @@ -0,0 +1,9293 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 234, + 103, + 375, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 103, + 375, + 121 + ], + "spans": [ + { + "bbox": [ + 234, + 103, + 375, + 121 + ], + "type": "text", + "content": "Always Skip Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": "Yiping Ji" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": ", Hemanth Saratchandran" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": ", Peyman Moghadam" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": ", Simon Lucey" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": "Adelaide University, " + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": "CSIRO, " + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 119, + 142, + 489, + 171 + ], + "type": "text", + "content": "Queensland University of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 134, + 174, + 474, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 174, + 474, + 199 + ], + "spans": [ + { + "bbox": [ + 134, + 174, + 474, + 199 + ], + "type": "text", + "content": "{yiping.ji, hemanth.saratchandran, simon.lucy}@adelaide.edu.au, {yiping.ji, peyman.moghadam}@csiro.au" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 209, + 126, + 380 + ], + "blocks": [ + { + "bbox": [ + 58, + 209, + 126, + 380 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 126, + 380 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 126, + 380 + ], + "type": "image", + "image_path": "d557aff84ed6bb07db54af823bbea325dbabb942af77bd77072f197d35abd6e9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 385, + 123, + 396 + ], + "lines": [ + { + "bbox": [ + 62, + 385, + 123, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 123, + 396 + ], + "type": "text", + "content": "(a) Architecture" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 127, + 216, + 340, + 382 + ], + "blocks": [ + { + "bbox": [ + 127, + 216, + 340, + 382 + ], + "lines": [ + { + "bbox": [ + 127, + 216, + 340, + 382 + ], + "spans": [ + { + "bbox": [ + 127, + 216, + 340, + 382 + ], + "type": "image", + "image_path": "79fcdb7e27381b66ffa8d51cf6358191763ed54e29d4133d72f97258dcdf90d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 386, + 331, + 409 + ], + "lines": [ + { + "bbox": [ + 156, + 386, + 331, + 409 + ], + "spans": [ + { + "bbox": [ + 156, + 386, + 331, + 409 + ], + "type": "text", + "content": "(b) FFN and SAB output embedding condition number using different configurations" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 342, + 220, + 553, + 381 + ], + "blocks": [ + { + "bbox": [ + 342, + 220, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 342, + 220, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 342, + 220, + 553, + 381 + ], + "type": "image", + "image_path": "d73e83f7d50976ffadb9c8aeb611c2f79b43596064c42de14eba1028f5e5c199.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 385, + 539, + 397 + ], + "lines": [ + { + "bbox": [ + 356, + 385, + 539, + 397 + ], + "spans": [ + { + "bbox": [ + 356, + 385, + 539, + 397 + ], + "type": "text", + "content": "(c) Performance after removing skip connections" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "lines": [ + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": "Figure 1. Removing skip connections from the Self-Attention Block (SAB) versus the Feedforward Network (FFN) in Vision Transformer (ViT) models results in different performance drop. (a) Model architecture of ViT consisting of a SAB, and a FFN in each layer. " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " represents the output of the self-attention transformation, while " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " represents the output of the MLP transformation. " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " represent identity mappings, referred to as skip connections, in the SAB and FFN, respectively. (b) The condition number of the different output embeddings for each hidden layer. A ViT-Tiny model is trained in a typical manner with skip connections on both SAB and FFN blocks (i.e., " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "① + ② + ③ + ④" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": "). It shows that the SAB output embedding without skip connection is poorly conditioned " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "(①)" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": ", with a condition number significantly higher than that of the other three configurations involving ViT blocks. (c) Classification results on three different datasets using the ViT-Tiny model under three different architectures at train time. The result shows the effects of removing skip connections at train time on the SAB and FFN blocks (either " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": ") in all layers of ViT-Tiny. We observe that removing " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " results in a catastrophic performance drop, whereas removing " + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 54, + 418, + 555, + 538 + ], + "type": "text", + "content": " (as predicted) has a modest impact on performance. This difference becomes more pronounced as the scale of the dataset increases." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "spans": [ + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "content": "We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (e.g., CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 628 + ], + "type": "text", + "content": "conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying (TG), a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 643, + 395, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 643, + 395, + 655 + ], + "spans": [ + { + "bbox": [ + 314, + 643, + 395, + 655 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have demonstrated impressive performance on various computer vision and robotic tasks, such as object detection, semantic segmentation, video understanding, visual-language learning, and many oth-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23115" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "ers [8, 12, 17, 19, 33]. Much of this success is attributed to the self-attention mechanism, commonly referred to as Self-Attention Block (SAB), which allows ViTs to selectively attend to relevant parts of the input sequence when generating each element of the output sequence. Another essential component is the MLP, or Feedforward Networks (FFN), which facilitates intra-token communication across channel dimensions (see Fig. 1). Finally, identity mappings, commonly referred to as skip connections, are incorporated into both the SAB and FFN to enhance model performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 312 + ], + "type": "text", + "content": "We argue that the Jacobian of the SAB is disproportionately ill-conditioned compared to other components, notably the FFN block. A poorly conditioned " + }, + { + "bbox": [ + 55, + 192, + 295, + 312 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 55, + 192, + 295, + 312 + ], + "type": "text", + "content": " Jacobian is fundamentally detrimental to gradient descent training [1, 13, 20, 24, 25], impeding both convergence and stability. A natural strategy to address this issue is to explicitly model the Jacobian of the entire network. However, this approach is impractical due to the sheer scale of modern transformer architectures. To proceed, we introduce the following simplifying assumptions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 60, + 324, + 118, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 324, + 118, + 335 + ], + "spans": [ + { + "bbox": [ + 60, + 324, + 118, + 335 + ], + "type": "text", + "content": "Assumptions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 337, + 285, + 383 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 60, + 337, + 285, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 337, + 285, + 358 + ], + "spans": [ + { + "bbox": [ + 60, + 337, + 285, + 358 + ], + "type": "text", + "content": "1. The condition of the network Jacobian is bound by the most ill-conditioned sub-block Jacobian." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 360, + 285, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 360, + 285, + 383 + ], + "spans": [ + { + "bbox": [ + 60, + 360, + 285, + 383 + ], + "type": "text", + "content": "2. A proxy for the condition of the sub-block Jacobian is the condition of the output embedding of that block." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 396, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 491 + ], + "type": "text", + "content": "Although we do not provide formal proofs for these assumptions, we offer indirect validation through extensive empirical analysis. First, we show that better conditioning of a sub-block's output embeddings consistently correlates with improved empirical performance across multiple ViT benchmarks. Second, we demonstrate that this improvement in embedding condition is accompanied by a measurable enhancement in the sub-block's Jacobian condition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "text", + "content": "Skip in Transformers. Skip connections have become an indispensable component in modern transformer architectures [29] and have been empirically proven to be a de facto component. In Fig. 1 (b), we analyze a pre-trained ViT-Tiny model by measuring the condition numbers of the SAB and FFN output embeddings, both with and without skip connections. Our observations reveal that the SAB output embedding becomes highly ill-conditioned in the absence of skip connections, with a condition number approaching " + }, + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "inline_equation", + "content": "e^6" + }, + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "text", + "content": ". In contrast, the other three configurations have relatively low condition numbers, around " + }, + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "inline_equation", + "content": "e^3" + }, + { + "bbox": [ + 55, + 495, + 295, + 675 + ], + "type": "text", + "content": ". In other words, the SAB output embedding without skip connections has significantly worse condition than the other configurations. To further investigate the impact of skip connections, we train the ViT-Tiny models from scratch under three config-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 547, + 216 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 547, + 216 + ], + "type": "image", + "image_path": "1c911ac0c9883163a64b9628d1eba9dd28af0452d086b6e5092eec04d10a79de.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 553, + 249 + ], + "type": "text", + "content": "Figure 2. Training loss for three different configurations of ViT-Tiny models trained on the Tiny-ImageNet dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 271, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 306 + ], + "type": "text", + "content": "urations: the standard SAB and FFN with skip connections, SAB without skip connections, and FFN without skip connections." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "text", + "content": "In Fig. 1 (c), we observe a modest drop in classification accuracy " + }, + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "inline_equation", + "content": "(2\\%)" + }, + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "text", + "content": " a ViT-Tiny model trained on the CIFAR-10 dataset when skip connections are removed from the FFN while retained in the SAB. However, removing skip connections from the SAB, while keeping them in the FFN, leads to a near-catastrophic performance drop of approximately " + }, + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 313, + 308, + 554, + 522 + ], + "type": "text", + "content": ". Moreover, as dataset size increases, models without SAB skip connections degrade substantially in performance; on larger-scale datasets such as Tiny-[ImageNet, these models fail to train altogether. To the best of our knowledge, this empirical finding has not been reported in prior literature. Additionally, the training loss curves for the three configurations are shown in Fig. 2. The results demonstrate that the SAB without skip connections, which is significantly ill-conditioned, exhibits much slower convergence and diverges after 30 epochs. Interestingly, this critical dependence on skip connections is unique to transformer architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 526, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 670 + ], + "type": "text", + "content": "Skip in ConvMixer. To further highlight the poor conditioning of the SAB, we conducted a similar experiment on a modern CNN ConvMixer [28]. ConvMixer has a similar architecture and competitive performance to ViT with the exception that self-attention is replaced by a convolution block to spatially mix tokens. Unlike ViT, training ConvMixer-Tiny with and without skip connections, we observed that the performance remains effectively unchanged, with variations of less than " + }, + { + "bbox": [ + 313, + 526, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\pm 0.2\\%" + }, + { + "bbox": [ + 313, + 526, + 554, + 670 + ], + "type": "text", + "content": ". Further details are shown in Fig. 6 in Appendix. This contrast further suggests that the SAB introduces an extreme conditioning issue, making skip connections indispensable for stable training in ViTs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 677, + 533, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 677, + 533, + 688 + ], + "spans": [ + { + "bbox": [ + 326, + 677, + 533, + 688 + ], + "type": "text", + "content": "Our main contributions are summarized as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 689, + 553, + 713 + ], + "type": "text", + "content": "1. We present a proposition that characterizes why SAB output embedding without skip connection is funda" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 712 + ], + "type": "text", + "content": "1The condition number of a full-rank matrix is defined by the ratio of the largest singular value to the smallest singular value. A lower condition number implies better condition." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23116" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 294, + 95 + ], + "type": "text", + "content": "mentally ill-conditioned, which challenges training convergence and stability (Sec. 4.1)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 294, + 215 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 96, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 294, + 155 + ], + "type": "text", + "content": "2. A theoretical analysis on the role of skip connection within the SAB is undertaken. We demonstrate that it significantly improves the condition of the block's output embedding, enhancing stability and performance (Sec. 4.2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 215 + ], + "type": "text", + "content": "3. Finally, we propose a novel approach—Token Greying (TG)—to better pre-condition the input tokens. This step, when used in conjunction with conventional skip connections—improves ViT in both supervised and self-supervised settings (Sec. 5)." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 216, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 276 + ], + "type": "text", + "content": "Our central contribution in this paper is not TG itself, but the insight that the SAB within modern ViTs is intrinsically ill-conditioned. Our hope is that this insight opens up brand new lines of inquiry for the vision community for more effective and efficient ViT design." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 139, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 139, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 139, + 297 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 304, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 544 + ], + "type": "text", + "content": "Self-Attention. Originally proposed by Vaswani et al. [29], the self-attention mechanism has become central to transformer architectures, enabling the capture of long-range dependencies by dynamically weighting interactions between tokens because it captures long-range dependencies by dynamically weighting interactions between tokens. Building on this foundation, transformer-based large language models, such as Llama [6] and Deepseek [15], have significantly advanced language understanding. In computer vision, researchers introduced Vision Transformers (ViTs) [5] for image classification, achieving superior results compared to traditional convolutional neural networks. Moreover, researchers have trained Diffusion Transformers [21], replacing the commonly used U-Net backbone with a transformer that operates on latent patches and inherits the excellent scaling properties of the transformer model class. Similarly, self-attention mechanisms have been successfully applied in speech recognition and physics-informed neural networks, further demonstrating their versatility across diverse applications [3, 32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Skip connections. He et al. [11] introduced skip connections in ResNet to make deep networks easier to optimize and to enhance accuracy by enabling significantly increased network depth. Since their introduction, skip connections have been extensively studied and applied across various Convolutional Neural Network (CNN) architectures. Building on these ideas, Veit et al. [30] offer a novel interpretation of residual networks, demonstrating that they can be viewed as a collection of multiple paths of varying lengths, which contributes to the network's flexibility and resilience. Furthermore, Hayou et al. [9] addresses gradient stability in deep ResNets, noting that while skip connections alleviate the issue of vanishing gradients, they may also intro-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "duce gradient explosion at extreme depths. To counter this, they propose scaling the skip connections according to the layer index, thereby stabilizing gradients and ensuring network expressivity even in the limit of infinite depth. Recently, in Table 3 of [16], the authors reported that VGG—a model without skip connections—achieves comparable results to ResNet—a model with skip connections—on modern datasets. To the best of our knowledge, only one paper challenges the skip connection in transformer architectures. In [10], the authors attempted to train deep transformer networks without using skip connections, achieving performance comparable to standard models. However, this approach requires five times more iterations, and the reasons why skip connections are crucial for self-attention-based transformers remain unresolved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 263, + 391, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 391, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 391, + 277 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 284, + 553, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 553, + 343 + ], + "type": "text", + "content": "In this section, we define the self-attention blocks used in the Vision Transformer (ViT) architecture and establish the mathematical notation for key quantities referenced in subsequent sections. For a comprehensive overview of transformers, we refer readers to [5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 344, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 553, + 403 + ], + "type": "text", + "content": "A Vision Transformer architecture consists of " + }, + { + "bbox": [ + 313, + 344, + 553, + 403 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 344, + 553, + 403 + ], + "type": "text", + "content": " stacked self-attention blocks (SABs) and feedforward networks (FFNs). An identity mapping, commonly referred to as a skip connection, is applied to connect the inputs and outputs of the transformation in both SAB and FFN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": "Self-Attention Blocks in ViT. Formally, given an input sequence " + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": " tokens of dimension " + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": ", a SAB is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 466, + 553, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 466, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 383, + 466, + 553, + 479 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {o u t}} := \\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {1}", + "image_path": "2388011fa4c0b5722d180e04e3b273d9f282da6c186d38eb6386a1bd2543d3c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 489, + 496, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 496, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 496, + 502 + ], + "type": "text", + "content": "where the self-attention output embedding is" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 513, + 553, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 513, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 383, + 513, + 553, + 528 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} \\left(\\mathbf {X} _ {\\text {i n}}\\right) = \\eta \\left(\\mathbf {Q K} ^ {\\mathbf {T}}\\right) \\mathbf {V}, \\tag {2}", + "image_path": "c57829efb03219da08096fe17a2c8b4ef040444f51ee64acffbbef137a385e06.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{Q} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{Q}},\\mathbf{K} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{K}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\mathbf{X}_{\\mathrm{in}}\\mathbf{W}_{\\mathrm{V}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{Q}},\\mathbf{W}^{\\mathrm{K}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^{\\mathrm{V}}\\in \\mathbb{R}^{d\\times d}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\eta (\\cdot)" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " is an activation function and is set by default as softmax function in [5]. Additionally, some studies [22, 23, 26] have explored the possibilities of using alternative activation functions. In particular, good performance is achieved by using linear attention with an appropriate scale. It allows us to characterize why the SAB output embedding without a skip connection is ill-conditioned." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "For general vision transformers, multiple heads " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SA}(\\mathbf{X}_{\\mathrm{in}})_i" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "1 \\leq i \\leq h" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", are commonly used. Learnable matrices are divided into head numbers " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{Q}}, \\mathbf{W}_{\\mathrm{K}}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{V}} \\in \\mathbb{R}^{h \\times d \\times d_h}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "d_h = \\frac{d}{h}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". Then all outputs of each attention head are concatenated together before the skip connection." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23117" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "Feed Forward Networks in ViT. Formally, given an input sequence " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{\\mathrm{in}} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " tokens of dimension " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": ", an FFN is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 118, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 118, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 97, + 118, + 295, + 133 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {\\text {o u t}} := \\mathbf {W} _ {\\text {d o w n}} \\left(g \\left(\\mathbf {W} _ {\\text {u p}} \\left(\\mathbf {X} _ {\\text {i n}}\\right)\\right)\\right) + \\mathbf {X} _ {\\text {i n}}, \\tag {3}", + "image_path": "c7ff2503729013d8d260de7e5346e79d58e1c3954d17f3dae0ae87da24dc7587.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up}}\\in \\mathbb{R}^{d\\times 4d}" + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "text", + "content": " is up projection, " + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}}\\in \\mathbb{R}^{4d\\times d}" + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "text", + "content": " is down projection, and " + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 140, + 296, + 166 + ], + "type": "text", + "content": " is an activation function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "text", + "content": "Conditioning of matrices. Formally, for a rectangular full rank matrix " + }, + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "text", + "content": ", the condition number of " + }, + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 178, + 296, + 214 + ], + "type": "text", + "content": " is defined below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 223, + 296, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 223, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 296, + 251 + ], + "type": "interline_equation", + "content": "\\kappa (\\mathbf {A}) = \\| \\mathbf {A} \\| _ {2} \\| \\mathbf {A} ^ {\\dagger} \\| _ {2} = \\frac {\\sigma_ {\\max } (\\mathbf {A})}{\\sigma_ {\\min } (\\mathbf {A})}, \\tag {4}", + "image_path": "c9af16fcf3f365e872e862484c017f39cfb8bf7b7e14834a86e73dba45d07993.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": " is the matrix operator norm induced by the Euclidean norm. We use the Euclidean norm in this paper. " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^{\\dagger}" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": " is the pseudo inverse of the matrix " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": " are the maximal and minimal singular values of " + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 254, + 296, + 315 + ], + "type": "text", + "content": " respectively. We will use Eq. (4) as a metric for analysis in future sections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 173, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 173, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 173, + 339 + ], + "type": "text", + "content": "4. Theoretical analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 345, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 429 + ], + "type": "text", + "content": "In this section, we provide a theoretical analysis showing that the output embedding of the SAB without skip connections is highly ill-conditioned. This matters as the condition of the output embedding is a proxy for the condition of the sub-block's Jacobian. We then demonstrate that an essential function of the skip connection is to improve this conditioning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 436, + 222, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 222, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 222, + 449 + ], + "type": "text", + "content": "4.1. Self-attention is ill-conditioned" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 454, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 477 + ], + "type": "text", + "content": "We attempt to validate our claim for the simpler case of linear attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "content": "Proposition 4.1. Assume " + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}" + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 53, + 486, + 296, + 546 + ], + "type": "text", + "content": " have entries that are independently drawn from a distribution with zero mean. The condition number of the SAB output embedding (without skip-connection) can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 554, + 295, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 554, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 89, + 554, + 295, + 585 + ], + "type": "interline_equation", + "content": "\\kappa \\left(\\mathbf {X} \\mathbf {W} _ {Q} \\mathbf {W} _ {K} ^ {T} \\mathbf {X} ^ {T} \\mathbf {X} \\mathbf {W} _ {V}\\right) \\leq C \\cdot \\left(\\frac {\\sigma_ {\\text {m a x}}}{\\sigma_ {\\text {m i n}}}\\right) ^ {3}, \\tag {5}", + "image_path": "0444696b7723db71935fc8d755d8f25508089066ca4b29c95e9f6d2b293166a7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": " are the maximal and minimal singular values of " + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 605, + 295, + 641 + ], + "type": "text", + "content": " is a fixed constant that is statistically close to unity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Proposition 4.1 assumes the use of linear attention (i.e., no softmax) as advocated for in [26]. This was done to offer a simplified theoretical exposition on how the SAB without skip connection can affect the condition of the output embedding. The proposition shows the condition number of the output embedding without skip connection is bounded" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 82, + 553, + 207 + ], + "blocks": [ + { + "bbox": [ + 317, + 82, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 317, + 82, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 82, + 553, + 207 + ], + "type": "image", + "image_path": "5ba6082307532d2f6dab6594eef31d6e2a6cab166bb6691c441b1da5a00ebc4e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 217, + 552, + 342 + ], + "blocks": [ + { + "bbox": [ + 317, + 217, + 552, + 342 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 552, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 552, + 342 + ], + "type": "image", + "image_path": "f22015f645c58bf02500a14df6568b8fd01412dadc5a8448f859c10195db50fd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 396 + ], + "type": "text", + "content": "Figure 3. Condition numbers across all layers of trained ViT-Tiny on Tiny-ImageNet datasets. We use " + }, + { + "bbox": [ + 313, + 352, + 555, + 396 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 352, + 555, + 396 + ], + "type": "text", + "content": " validation sets to plot the above figures. Upper: fixed scale linear self-attention. Lower: softmax self-attention" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "text", + "content": "by the cube of the condition number of the input matrix " + }, + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "text", + "content": ". Unless the condition of " + }, + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 416, + 554, + 487 + ], + "type": "text", + "content": " is unity, the output of the SAB without skip connection will therefore result in a poorly conditioned embedding. Applying this same process across multiple layers within a transformer results in a highly ill-conditioned outcome." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "text", + "content": "Fig. 3 illustrates our theoretical analysis by showing that the linear self-attention operation increases the condition number of embeddings from " + }, + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "inline_equation", + "content": "e^3" + }, + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "inline_equation", + "content": "e^6" + }, + { + "bbox": [ + 313, + 488, + 554, + 559 + ], + "type": "text", + "content": ", and we can empirically extend this analysis to softmax self-attention. The output embedding of the softmax SAB without skip connections is also significantly ill-conditioned." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 559, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 594 + ], + "type": "text", + "content": "Further, for condition numbers of FFN output embeddings without skip connection using linear activation " + }, + { + "bbox": [ + 313, + 559, + 554, + 594 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 559, + 554, + 594 + ], + "type": "text", + "content": " we have:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 601, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 601, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 318, + 601, + 555, + 624 + ], + "type": "interline_equation", + "content": "\\kappa (\\mathbf {M L P} (\\mathbf {X})) = \\kappa (\\mathbf {W} _ {\\mathrm {d o w n}} \\mathbf {W} _ {\\mathrm {u p}} \\mathbf {X}) \\leq C _ {\\mathrm {d o w n}} \\cdot C _ {\\mathrm {u p}} \\cdot \\frac {\\sigma_ {\\operatorname* {m a x}}}{\\sigma_ {\\operatorname* {m i n}}}, \\tag {6}", + "image_path": "e8a8774d39cf5f0f9fd08328040149aca4fec6b7420c0e5cba0ea96030bbecd9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " are maximal singular values of " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " computed using SVD. Since " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{down}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{up}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " do not depend on data " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": ", we denote " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\kappa (\\mathbf{W}_{\\mathrm{down}}) = C_{\\mathrm{down}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "inline_equation", + "content": "\\kappa (\\mathbf{W}_{\\mathrm{up}}) = C_{\\mathrm{up}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Compared to the self-attention mechanism, the MLP transformation does not worsen conditioning as severely, since it has a much lower bound on the condition number. In Fig. 1 (b), the output embeddings of the FFN without" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23118" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "skip connection is observed to have a significantly better condition than the output embedding of the SAB without skip connection. Therefore, in (c), the performance drop after removing the FFN's skip connections is less severe than after removing the SAB's skip connections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 246, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 246, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 246, + 152 + ], + "type": "text", + "content": "4.2. Skip connection improves condition" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 158, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 296, + 194 + ], + "type": "text", + "content": "In this subsection, we theoretically demonstrate that the skip connection improves the condition of the self-attention output embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 194, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 385 + ], + "type": "text", + "content": "Previous works [4, 10, 18] have demonstrated that transformers, without skip connections, converge very slowly or may even become un-trainable. In [10], the authors show that without skip connections, output embeddings of the SAB quickly converge to rank 1. They hypothesize that this is due to the inability of the transformer to propagate signals through the self-attention mechanism. This rank-1 output embedding implies an infinite condition number. The authors introduce an inductive bias into the self-attention mechanism to enable what they refer to as \"better signal propagation\". While their approach matches the performance of standard (vanilla) models, it incurs a substantial computational cost—running approximately five times slower. We hypothesize that the success of this method could be directly attributed to its ability to improve the condition of the self-attention output embeddings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 422 + ], + "type": "text", + "content": "This observation further motivates our view that the primary role of skip connections may be to regularize the condition of the self-attention output embeddings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": "Proposition 4.2. Let " + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": " have entries that are independently drawn from a distribution with zero mean. We define " + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{M} = \\mathbf{W}_{\\mathbf{Q}} \\mathbf{W}_{\\mathbf{K}}^{\\mathrm{T}} \\mathbf{X}^{\\mathrm{T}} \\mathbf{X} \\mathbf{W}_{\\mathbf{V}}" + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": " and assume " + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": " is the positive semi-definite matrix. We have the following bounds on the condition numbers " + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 54, + 429, + 296, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 510, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 510, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 121, + 510, + 295, + 525 + ], + "type": "interline_equation", + "content": "\\kappa (\\mathbf {X M} + \\mathbf {X}) \\ll \\kappa (\\mathbf {X M}). \\tag {7}", + "image_path": "9a4edf47086f5b54ff092253ff25df3e9b67d87de964c36f63b28bd842088755.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "type": "text", + "content": "Proposition 4.2 shows that during training, the condition number of the SAB output embedding has a much lower bound than that of the SAB output embedding without skip connections. Furthermore, with other conditions being equal, better condition of SAB can contribute to faster convergence and more stable gradient updates, reducing the risk of gradient explosion or vanishing. This ensures the signal can propagate well during the training process in the Vision Transformers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "In other words, the self-attention mechanism enables tokens to communicate by focusing on different parts of the input, thereby capturing correlations between elements within the same sequence. However, one critical drawback is that it is ill-conditioned and challenging to train. In contrast, although skip connections do not facilitate communi" + } + ] + } + ], + "index": 8 + }, + { + "type": "code", + "bbox": [ + 318, + 87, + 555, + 273 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "type": "text", + "content": "Algorithm 1 SVD Token Graying" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " Input: training data, amplification coefficient " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0,1]" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " Training: \n3: for each minibatch " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " do \n4: Convert image " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " into token " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{U\\Sigma V^T} \\leftarrow \\mathbf{X}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " Compute SVD \n6: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\Sigma} \\leftarrow \\frac{\\Sigma}{\\max(\\Sigma)}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": "Normalize elements " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\in (0,1]" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\Sigma} \\leftarrow \\tilde{\\Sigma}^\\epsilon" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": "Amplify \n8: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}} \\leftarrow \\mathbf{U}\\tilde{\\Sigma}\\mathbf{V^T}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": "SVD reconstruct \n9: Patch Embedding: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " PatchEmbed(\\tilde{\\mathbf{X}}) \n10: Forward Pass: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\hat{y} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " Model(\\mathbf{Z}) \n11: Compute Loss: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " Loss(\\hat{y},y_i) \n12: Backward Pass: Compute gradients " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\nabla_{\\theta}\\mathcal{L}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " \n13: Parameter Update: " + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "inline_equation", + "content": "\\theta \\leftarrow \\theta - \\eta\\nabla_{\\theta}\\mathcal{L}" + }, + { + "bbox": [ + 318, + 87, + 555, + 273 + ], + "type": "text", + "content": " \n14: end for" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 294, + 555, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 318 + ], + "type": "text", + "content": "cation between tokens, they help regularize the condition of output embeddings and aid in the optimization process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "content": "As shown in Fig. 3, empirical results demonstrate that in both the linear and softmax self-attention cases, adding skip connections significantly improves the condition of output embeddings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 397, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 397, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 397, + 392 + ], + "type": "text", + "content": "5. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 399, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 447 + ], + "type": "text", + "content": "In this section, we introduce Token Graying (TG), a simple yet effective method that employs Singular Value Decomposition and Discrete Cosine Transform techniques to reconstruct better-conditioned input tokens." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 455, + 511, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 511, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 511, + 468 + ], + "type": "text", + "content": "5.1. Singular Value Decomposition (SVD)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": "Singular Value Decomposition of a matrix " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " is the factorization of " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " into the product of three matrices " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\mathbf{U} \\boldsymbol{\\Sigma} \\mathbf{V}^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " where the columns of " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " are orthonormal and " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " is a rectangular diagonal matrix whose diagonal elements are non-negative in descending order and represent the singular values such that " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}_{ii} = \\sigma_i" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": ". As observed in Eq. (4), we can reduce the condition number by increasing the minimal singular value while keeping the maximal singular value unchanged. Hence, our goal is to reconstruct " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}} = \\mathbf{U} \\boldsymbol{\\tilde{\\Sigma}} \\mathbf{V}^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\tilde{\\Sigma}}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " is formed by amplifying non-maximal elements of " + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}" + }, + { + "bbox": [ + 313, + 473, + 555, + 629 + ], + "type": "text", + "content": " while keeping the maximal element unchanged. Our SVD token graying pipeline is presented in Algorithm 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": "However, directly applying SVD to matrices is computationally expensive, with a cost of " + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "inline_equation", + "content": "O(nd \\cdot \\min(n, d))" + }, + { + "bbox": [ + 313, + 630, + 554, + 712 + ], + "type": "text", + "content": " leading to longer training time. Tab. 1 demonstrates that training a ViT is significantly slower when using SVD reconstruction. Therefore, in the next subsection, we introduce a more efficient Token Graying (TG) method using the discrete cosine transform." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23119" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 71, + 294, + 104 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 294, + 104 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 294, + 104 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 294, + 104 + ], + "type": "table", + "html": "
MethodsViT-BViT-B + SVDTGViT-B + DCTTG
Time (days)0.7234.5520.732
", + "image_path": "7de2d55ac5052375a5e29a8acbc3808312a3754ceb27ddde41eb46c3af249f42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 295, + 135 + ], + "type": "text", + "content": "Table 1. ViT-Base training time w/o and w/ token graying methods on ImageNet-1K dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 154, + 237, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 237, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 237, + 167 + ], + "type": "text", + "content": "5.2. Discrete Cosine Transform (DCT)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 219 + ], + "type": "text", + "content": "A Discrete Cosine Transform (DCT) expresses a finite sequence of data points as a sum of cosine functions oscillating at different frequencies. It is real-valued and has an inverse, which is often simply called the Inverse DCT (IDCT)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "text", + "content": "There are several variants of the DCT and in this work we choose DCT-II which is one of the most commonly used forms. Given a real-valued sequence data " + }, + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "text", + "content": ", the 1D DCT sequence " + }, + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 220, + 296, + 268 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 285, + 296, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 368 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mathbf {x}} _ {k} = \\alpha_ {k} \\sum_ {i = 0} ^ {N - 1} x _ {i} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right], \\mathrm {f o r} k = 0, 1, \\dots , N - 1 \\\\ \\text {w h e r e} \\alpha_ {k} = \\left\\{ \\begin{array}{l} \\sqrt {\\frac {1}{N}}, \\text {i f} k = 0 \\\\ \\sqrt {\\frac {2}{N}}, \\text {i f} k \\neq 0 \\end{array} \\right. \\tag {8} \\\\ \\end{array}", + "image_path": "0d2466a41fef55a66854535b03c9210753ffcee086bf61e03418e4af475d79b9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "text", + "content": "Simplicity, 1D DCT is a linear transformation, Eq. (8) can be written with the formula " + }, + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "inline_equation", + "content": "DCT(\\mathbf{x}) = \\hat{\\mathbf{x}} = D\\mathbf{x}" + }, + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "inline_equation", + "content": "D_{i,k}" + }, + { + "bbox": [ + 55, + 369, + 296, + 406 + ], + "type": "text", + "content": " is the 1D DCT basis and orthogonal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 422, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 422, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 113, + 422, + 295, + 449 + ], + "type": "interline_equation", + "content": "D _ {i, k} = \\alpha_ {k} \\cos \\left[ \\frac {\\pi (2 i + 1) k}{2 N} \\right]. \\tag {9}", + "image_path": "51cfe3402c133dd842f01c7e958b1f8c4a8978f148d3e570430da4fe39016486.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": "For the reconstruction process, since the DCT expresses a signal as a sum of cosine functions with different frequencies, the IDCT combines these frequency components to recover the original signal. Hence, " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathrm{IDCT}(\\hat{x}) = \\tilde{\\mathbf{x}} = D^T\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": ". In our case, we have a token matrix " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": " and we generalize the 1D DCT to a 2D DCT using the formula " + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} = D\\mathbf{X}D^{T}" + }, + { + "bbox": [ + 55, + 456, + 295, + 598 + ], + "type": "text", + "content": ". In natural images, the dominant singular vector, associated with the largest singular value, typically captures the low frequency content of the image. Intuitively, this observation suggests that DCT reconstruction can serve as an approximation for SVD reconstruction. Our DCT token graying pipeline is presented in Algorithm 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "content": "DCT avoids the expensive computational cost of SVD and has a complexity of " + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "inline_equation", + "content": "O(nd \\cdot \\log(nd))" + }, + { + "bbox": [ + 55, + 599, + 296, + 647 + ], + "type": "text", + "content": ". Tab. 1 shows that training ViT using the DCT algorithm is significantly faster than SVD algorithm." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 137, + 671 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have emerged as powerful models in the field of computer vision, demonstrating remarkable performance across a variety of tasks. This section is" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 318, + 87, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 454, + 84 + ], + "type": "text", + "content": "Algorithm 2 DCT Token Graying" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " Input: training data, amplification coefficient " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0,1]" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " \n2: Training: \n3: for each minibatch " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\{(\\mathbf{x}_i,y_i)\\}_{i=1}^k \\sim" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " training data do \n4: Convert image " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " into token " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "D\\hat{\\mathbf{X}}D^T \\leftarrow \\mathbf{X}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " Compute DCT \n6: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}_{\\mathrm{norm}} \\leftarrow |\\hat{\\mathbf{X}}| / \\max(|\\hat{\\mathbf{X}}|)" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": "Normalize elements " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\in (0,1]" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\leftarrow \\hat{\\mathbf{X}}_{\\mathrm{norm}}^\\epsilon \\cdot \\mathrm{sign}(\\hat{\\mathbf{X}}) \\cdot \\max(|\\hat{\\mathbf{X}}|)" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": "Amplify \n8: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}} \\leftarrow \\hat{D}^T\\hat{\\mathbf{X}}\\hat{D}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": "Inverse DCT \n9: Patch Embedding: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " PatchEmbed(" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": ") \n10: Forward Pass: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{y} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " Model(" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": ") \n11: Compute Loss: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\leftarrow" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " Loss(" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{y}, y_i" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": ") \n12: Backward Pass: Compute gradients " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\nabla_\\theta \\mathcal{L}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " \n13: Parameter Update: " + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\theta \\leftarrow \\theta - \\eta \\nabla_\\theta \\mathcal{L}" + }, + { + "bbox": [ + 318, + 87, + 555, + 281 + ], + "type": "text", + "content": " \n14: end for" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 303, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 350 + ], + "type": "text", + "content": "dedicated to validating and analyzing our proposed algorithm for both supervised and self-supervised learning in ViTs. For all experiments, unless specified otherwise, we use " + }, + { + "bbox": [ + 313, + 303, + 555, + 350 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.95" + }, + { + "bbox": [ + 313, + 303, + 555, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 360, + 451, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 451, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 451, + 373 + ], + "type": "text", + "content": "6.1. Supervised learning ViT" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 425 + ], + "type": "text", + "content": "In this subsection, we validate our methods in supervised learning using different types and scales of ViTs. Cross-entropy loss objective is used. We use PyTorch and Timm libraries." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 552 + ], + "type": "text", + "content": "- ViT [5]: is the pioneering work that interprets an image as a sequence of patches and processes it by a standard Transformer encoder as used in NLP. This simple, yet scalable, strategy works surprisingly well when coupled with pretraining on large datasets. We train the ViT-Tiny, which has 12 layers and 3 heads, with a head dimension of 64 and a token dimension of 192 on the Tiny-[ImageNet] dataset [14]. Additionally, we train ViT-Base, which consists of 12 layers and 12 heads, each with a head dimension of 64 and a token dimension of 768 on the ImageNet-1K dataset [2]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 554, + 629 + ], + "type": "text", + "content": "- Swin [17] is a hierarchical vision transformer designed to focus on local attention first and expand globally. It improves computational efficiency and locality modeling compared to standard ViTs. We train on the Swin small model with a patch size of 4 and a window size of 7 on the ImageNet-1K dataset." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 635, + 554, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 683 + ], + "type": "text", + "content": "- CaiT [27] is a variant Vision Transformer that aims at increasing the stability of the optimization when ViTs go deeper. We train CaiT small, which has 24 layers, on the ImageNet-1K dataset." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "- PVT [31] an enhanced version of the Pyramid Vision Transformer that utilizes a hierarchical architecture to ex" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23120" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 80, + 304, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 80, + 304, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 80, + 304, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 80, + 304, + 182 + ], + "type": "table", + "html": "
MethodTop-1 Acc (%)Top-5 Acc (%)
1/k(QK)43.062.7
1/k(QK) + SVD44.368.8
1/k(QK) + DCT44.268.8
softmax(QK)43.067.3
softmax(QK) + SVD44.769.7
softmax(QK) + DCT44.869.7
", + "image_path": "3f8adc1d0cc14bd7e6c1a9508b8df64fddf56cc26690069983122e2f95c93460.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 259, + 293, + 408 + ], + "blocks": [ + { + "bbox": [ + 55, + 190, + 296, + 245 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 245 + ], + "type": "text", + "content": "Table 2. ViT-Tiny results on Tiny-ImageNet datasets. In the upper part, we show results for linear self-attention, where " + }, + { + "bbox": [ + 55, + 190, + 296, + 245 + ], + "type": "inline_equation", + "content": "k = 16" + }, + { + "bbox": [ + 55, + 190, + 296, + 245 + ], + "type": "text", + "content": " is chosen based on [26] w/ and w/o our method. On the lower part, we show regular softmax self-attention results w/ and w/o our method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 259, + 293, + 408 + ], + "lines": [ + { + "bbox": [ + 62, + 259, + 293, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 293, + 408 + ], + "type": "table", + "html": "
MethodεTop-1 Acc (%)κinκout
ViT-Base-81.06.726.74
ViT-Base + SVDTG0.981.26.646.66
0.881.26.476.66
0.781.46.156.17
0.681.45.735.71
0.581.05.295.25
ViT-Base + DCTTG0.9781.16.476.49
0.9581.36.426.43
0.9381.26.336.35
0.981.16.256.25
0.8581.06.016.02
", + "image_path": "4b33c550490601f560c0283bd6730f2a12a66efc6e550890edcc501525d15348.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "text", + "content": "Table 3. ViT-Base results using SVD and DCT token graying methods with different amplification coefficient " + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\kappa_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\kappa_{\\mathrm{out}}" + }, + { + "bbox": [ + 55, + 417, + 296, + 462 + ], + "type": "text", + "content": " are the average token condition numbers in log scale before and after each SAB." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": "tract multi-scale features for improved performance in computer vision tasks. It incorporates efficient attention mechanisms and refined positional encoding to boost accuracy and reduce computational complexity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "content": "Results. In Tab. 2, we first demonstrate the results for the linear case and the regular softmax case using our methods on a small model (ViT-Tiny) and the small-scale TinyImageNet dataset. Given that Theorem Proposition 4.2 and Proposition Proposition 4.1 are mathematically explained using linear self-attention, we train the ViT-Tiny model with linear self-attention, which aligns well with our analysis in Fig. 3. Furthermore, as shown in Fig. 4, both SVDTG and DCTTG lead to better conditioning of SAB output embeddings throughout training epochs. Next, we train the model on a larger dataset using the default softmax self-attention, further validating that our approach generalizes to softmax self-attention. Moreover, SVDTG is the most direct method to improve the conditioning of a matrix, and as demonstrated in Tab. 3, it leads to notable performance gains." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 195 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 195 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 195 + ], + "type": "image", + "image_path": "aa422d58e87cec12de2196802ebbc98278cc0889416f479d1a5470ae053eabf0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 198, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 317, + 198, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 322 + ], + "type": "image", + "image_path": "0b70abd157dd403e423c8e45644f041348a8a1498b175ef3318e083bd2746f44.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 332, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 365 + ], + "type": "text", + "content": "Figure 4. We compute the condition number of the SAB block's output embeddings during ViT-Tiny training using SVDTG and DCTTG with different " + }, + { + "bbox": [ + 313, + 332, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 332, + 555, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 545 + ], + "type": "text", + "content": "To address the high computational cost—approximately six times slower than regular ViT training—we approximate SVDTG using DCTTG, which achieves better token embeddings conditioning and yields comparable results too. As shown in Fig. 5, our DCTTG method produces condition numbers across all layers that indicate better token conditioning in almost every layer. Finally, we evaluate our methods on different variants of ViTs trained on ImageNet-1K datasets, where our DCTTG method demonstrates superior performance compared to vanilla models in terms of Top-1 and Top-5 classification accuracy, while also achieving better conditioning of SAB output embeddings across all layers in the ViT-B model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 471, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 471, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 471, + 569 + ], + "type": "text", + "content": "6.2. Self-supervised learning ViT" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 554, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 554, + 659 + ], + "type": "text", + "content": "In this subsection, we validate our methods on self-supervised learning models as well as on fine-tuning pretrained models [7, 12]. Self-supervised learning is a powerful technique in which models learn representations directly from unlabeled data. Masked Image Modeling (MIM) [12] learns representations by reconstructing images that have been corrupted through masking." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Pre-training. In the pre-training stage, we pre-train our model on the ImageNet-1k training set without using ground-truth labels in a self-supervised manner. We use a batch size of 2048, a base learning rate of " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "1.5 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", and a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "23121" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 305, + 235 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 305, + 235 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 305, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 305, + 235 + ], + "type": "table", + "html": "
MethodTop-1 Acc (%)Top-5 Acc (%)
ViT-S80.295.1
ViT-S + DCTTG80.495.2
ViT-B81.095.3
ViT-B + DCTTG81.395.4
Swin-S81.395.6
Swin-S + DCTTG81.695.6
CaiT-S82.696.1
CaiT-S + DCTTG82.796.3
PVT V2 b382.996.0
PVT V2 b3 + DCTTG83.096.1
", + "image_path": "975f543b1ab1a1f5abe96ba9cfd299d027913e3f0eac67cadfd52b7ed7c3afab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 276, + 293, + 422 + ], + "blocks": [ + { + "bbox": [ + 58, + 276, + 293, + 422 + ], + "lines": [ + { + "bbox": [ + 58, + 276, + 293, + 422 + ], + "spans": [ + { + "bbox": [ + 58, + 276, + 293, + 422 + ], + "type": "image", + "image_path": "5e7e7342aa6c476d69917c3e7a8aa0edcca9ac887a87de8ae65a5728057b630f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 431, + 295, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 295, + 455 + ], + "type": "text", + "content": "Figure 5. Condition numbers of all layers in the trained ViT-Base model with and without the DCTTG method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 64, + 471, + 288, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 266 + ], + "type": "text", + "content": "Table 4. Top-1 and Top-5 classification accuracy on ImageNet-1K dataset using DCT token graying on different ViTs configurations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 471, + 288, + 521 + ], + "lines": [ + { + "bbox": [ + 64, + 471, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 64, + 471, + 288, + 521 + ], + "type": "table", + "html": "
MethodTop-1 Acc (%)Top-5 Acc (%)
MAE83.096.4
MAE + DCTTG83.296.6
", + "image_path": "fcec65867870f30dffe3773915013670363c66ddc098f4f0a4a8d26f2b736eef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 563 + ], + "type": "text", + "content": "Table 5. MAE pretrained ViT-Base results finetuned on ImageNet dataset-1k. We evaluate our MSVR method on self-attention using the power iteration method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 584, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 622 + ], + "type": "text", + "content": "weight decay of 0.05, and we train for 400 epochs. Our loss function computes the mean squared error (MSE) between the reconstructed and original images in pixel space." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Finetuning and reuslts. In the finetuning stage, we fine-tune the pre-trained models on ImageNet-1k for classification tasks using cross-entropy loss. We use a batch size of 512, a base learning rate of " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", a weight decay of 0.05, and train for 100 epochs. As shown in Tab. 5, our DCTTG method outperforms the vanilla MAE model by " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " in both top-1 and top-5 classification accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 71, + 384, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 384, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 384, + 83 + ], + "type": "text", + "content": "7. Limitation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 224 + ], + "type": "text", + "content": "Our work establishes a theoretical foundation demonstrating why skip connections act as essential conditioning regularizers within self-attention blocks. Additionally, we introduce a method designed to further enhance the output embeddings of self-attention blocks. However, certain limitations remain. For instance, while our regularizer is effective, training in low-precision settings may pose challenges due to the DCT operation, which involves many multiplications and summations that are sensitive to quantization errors. Additionally, for some variants of ViTs, the performance improvement might be marginal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 235, + 388, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 388, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 388, + 248 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 472 + ], + "type": "text", + "content": "This paper provides a theoretical demonstration that the output embeddings of linear self-attention blocks, when lacking skip connections, are inherently ill-conditioned, a finding that empirically extends to conventional softmax self-attention blocks as well. Furthermore, we show that skip connections serve as a powerful regularizer by significantly improving the conditioning of self-attention blocks' output embeddings. In addition, we use the condition of the self-attention output embeddings as a proxy for that of its Jacobian, empirically demonstrating that input tokens with improved conditioning lead to a better-conditioned Jacobian. Building on this insight, we introduce SVDTG and DCTTG, simple yet effective methods designed to improve the condition of input tokens. We then validate their effectiveness on both supervised and unsupervised learning tasks. Ultimately, we hope these insights will guide the vision community in designing more effective and efficient ViT architectures for downstream applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 476, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 513 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the Australian Institute for Machine Learning (AIML) and the CSIRO's Data61 Embodied AI Cluster." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 524, + 373, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 524, + 373, + 537 + ], + "spans": [ + { + "bbox": [ + 315, + 524, + 373, + 537 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 544, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 319, + 544, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 544, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 319, + 544, + 555, + 599 + ], + "type": "text", + "content": "[1] Paul Albert, Frederic Z Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Randlora: Full-rank parameter-efficient fine-tuning of large models. In International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 600, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 600, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 319, + 600, + 554, + 644 + ], + "type": "text", + "content": "[2] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE conference on computer vision and pattern recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 646, + 553, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 646, + 553, + 699 + ], + "spans": [ + { + "bbox": [ + 319, + 646, + 553, + 699 + ], + "type": "text", + "content": "[3] Linhao Dong, Shuang Xu, and Bo Xu. Speech-transformer: a no-recurrence sequence-to-sequence model for speech recognition. In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 5884-5888, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 319, + 702, + 553, + 714 + ], + "type": "text", + "content": "[4] Yihe Dong, Jean-Baptiste Cordonnier, and Andreas Loukas." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23122" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 106 + ], + "type": "text", + "content": "Attention is not all you need: Pure attention loses rank doubly exponentially with depth. In International Conference on Machine Learning, pages 2793-2803. PMLR, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 107, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 107, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 62, + 107, + 295, + 163 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 165, + 295, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 165, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 295, + 198 + ], + "type": "text", + "content": "[6] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "type": "text", + "content": "[7] Maryam Haghighat, Peyman Moghadam, Shaheer Mohamed, and Piotr Koniusz. Pre-training with Random Orthogonal Projection Image Modeling. In International Conference on Learning Representations (ICLR), 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 245, + 295, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 295, + 288 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 295, + 288 + ], + "type": "text", + "content": "[8] Stephen Hausler and Peyman Moghadam. Pair-VPR: Place-aware pre-training and contrastive pair classification for visual place recognition with vision transformers. IEEE Robotics and Automation Letters, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 290, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 290, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 290, + 294, + 334 + ], + "type": "text", + "content": "[9] Soufiane Hayou, Eugenio Clerico, Bobby He, George Deligiannidis, Arnaud Doucet, and Judith Rousseau. Stable resnet. In International Conference on Artificial Intelligence and Statistics, pages 1324-1332. PMLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 336, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 336, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 57, + 336, + 295, + 392 + ], + "type": "text", + "content": "[10] Bobby He, James Martens, Guodong Zhang, Aleksandar Botev, Andrew Brock, et al. Deep transformers without shortcuts: Modifying self-attention for faithful signal propagation. In The Eleventh International Conference on Learning Representations, 2023. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 393, + 294, + 437 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 438, + 294, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 294, + 493 + ], + "type": "text", + "content": "[12] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 495, + 294, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 495, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 57, + 495, + 294, + 539 + ], + "type": "text", + "content": "[13] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. In The International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 541, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 541, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 57, + 541, + 294, + 563 + ], + "type": "text", + "content": "[14] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 565, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 565, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 565, + 294, + 597 + ], + "type": "text", + "content": "[15] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 632 + ], + "type": "text", + "content": "[16] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[17] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[18] Lorenzo Noci, Chuning Li, Mufan Li, Bobby He, Thomas Hofmann, Chris J Maddison, and Dan Roy. The shaped" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "transformer: Attention models in the infinite depth-andwidth limit. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 148 + ], + "type": "text", + "content": "[19] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 150, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 172 + ], + "type": "text", + "content": "[20] Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks, 2013. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[21] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[22] Zhen Qin, Weixuan Sun, Hui Deng, Dongxu Li, Yunshen Wei, Baohong Lv, et al. cosformer: Rethinking softmax in attention. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[23] Jason Ramapuram, Federico Danieli, Eeshan Gunesh Dhekane, Floris Weers, Dan Busbridge, et al. Theory, analysis, and best practices for sigmoid self-attention. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[24] Hemanth Saratchandran and Simon Lucey. Enhancing transformers through conditioned embedded tokens. arXiv preprint arXiv:2505.12789, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[25] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "text", + "content": "[26] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking softmax: Self-attention with polynomial activations, 2024. 3, 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[27] Hugo Touvron, Matthieu Cord, Alexandre Sablayrolles, Gabriel Synnaeve, and Hervé Jégou. Going deeper with image transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 32-42, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 472, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 491 + ], + "type": "text", + "content": "[28] Asher Trockman and J. Zico Kolter. Patches are all you need?, 2022. 2, 10" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 494, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 514 + ], + "type": "text", + "content": "[29] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 516, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 553, + 557 + ], + "type": "text", + "content": "[30] Andreas Veit, Michael J Wilber, and Serge Belongie. Residual networks behave like ensembles of relatively shallow networks. Advances in neural information processing systems, 29, 2016. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[32] Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks. In The International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[33] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23123" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_content_list.json b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40c0107689b9eb49b7573545303ce17730e3275a --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_content_list.json @@ -0,0 +1,1903 @@ +[ + { + "type": "text", + "text": "Amodal Depth Anything: Amodal Depth Estimation in the Wild", + "text_level": 1, + "bbox": [ + 176, + 130, + 821, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyu Li $^{1}$ , Mykola Lavreniuk $^{2}$ , Jian Shi $^{1}$ , Shariq Farooq Bhat $^{1}$ , Peter Wonka $^{1}$ $^{1}$ KAUST, $^{2}$ Space Research Institute NASU-SSAU", + "bbox": [ + 184, + 179, + 810, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/802c0d29d73ea09201971555c48ab6dab09b47ad141358a0e80a5820f6bb4cb7.jpg", + "image_caption": [ + "Figure 1. Amodal Depth Estimation in the Wild. For each image, we present the amodal depth estimation with the target object outlined in black. Our model demonstrates strong generalization across diverse scenes, including both indoor and outdoor environments for accurate depth estimation for occluded parts of objects. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 93, + 250, + 903, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 585, + 328, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present", + "bbox": [ + 88, + 628, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions (Fig. 1). Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a $50.7\\%$ improvement in RMSE over the previous SoTA on the ADIW dataset.", + "bbox": [ + 511, + 585, + 908, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 782, + 645, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular depth estimation is a foundational task in computer vision and generative modeling [5, 10, 27] as it provides depth perception from a single image without stereo cues. However, while recent methods, e.g. [2, 4, 10, 23, 46], focus solely on estimating depth for visible pixels, humans can intuitively perceive the complete 3D", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9673", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dcacca2ba2921de9904d9143e0c51cd53f5b980dc24232d8a2689b550c5099d7.jpg", + "image_caption": [ + "Figure 2. Amodal Depth Estimation Pipeline. Given an input image, users can generate the amodal mask for the depth estimator in two ways: (1) Model Heuristics: click the target object, apply SAM [21] to generate modal mask, then use amodal segmentation methods to estimate amodal mask, (2) Human Heuristics: manually draw the amodal mask. Our model estimates amodal depth based on original observation image $I_{o}$ , the observed depth map $D_{o}$ , and the amodal mask $M_{a}$ . Note that we can get visible and invisible masks from both pipelines. Our model only focuses on the amodal depth (i.e., invisible part), whereas the visible part depth can be directly derived from the observed depth map $D_{o}$ ." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "geometry of objects, even when only parts of them are visible. Amodal depth estimation is the task of predicting depth values for the occluded (invisible) parts of objects [17, 38]. It enables accurate 3D reconstruction and novel view synthesis from a single view, significantly reducing dependence on costly multi-view systems or LiDAR [1]. This capability is crucial for advancing AR/VR, robotics, autonomous driving and digital twin technologies [1]. In this paper, we aim to address this under-explored question: Can models effectively perceive the geometry of the invisible parts of objects in a scene?", + "bbox": [ + 89, + 415, + 483, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike inpainting methods [11, 39, 50], which reconstruct missing image regions, amodal depth estimation is a novel task that extends traditional amodal segmentation [28, 48, 49] into the depth domain, by predicting the depth of occluded object parts. In this task, given an input image with a target amodal mask, the objective is to infer depth values for the occluded object regions. While amodal segmentation benefits from human annotations to collect training samples [17, 38], there is currently no device capable of collecting ground-truth depth data for occluded parts of objects at scale in real-world scenes.", + "bbox": [ + 89, + 582, + 482, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, prior approaches to amodal depth estimation have relied on synthetic datasets [17, 38]. However, generating these datasets is both time-consuming and hard to scale, often requiring manually placing occluders one by one [17, 38]. Additionally, synthetic data lacks the complexity and diversity of real-world scenes, resulting in a domain gap that limits the generalization of models trained on such data. Furthermore, these previous approaches only consider metric depth, where the goal is to estimate real-world distances for occluded regions. Re", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "liaence on metric depth, which naturally struggles to generalize to unseen scenarios with limited data [7, 29], exacerbates the models' poor zero-shot performance on real-world images [6, 34, 46].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth. Recent advancements in depth estimation models, such as Depth Anything [46], have enabled the generation of high-quality relative depth maps from natural images. Focusing on relative depth allows us to train models using real-world data, leveraging the depth relationships within scenes to achieve better generalization without relying on precise metric measurements.", + "bbox": [ + 511, + 152, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a new large-scale real-world dataset, Amodal Depth In the Wild (ADIW), to facilitate the training of models for amodal depth estimation. Our data generation pipeline leverages segmentation datasets to create high-quality datasets for amodal relative depth estimation with scalable annotations. We adopt a compositing approach to create training pairs, where objects are sampled and composited into scenes. Corresponding depth maps are estimated using the large depth foundation model [47]. Given the challenges posed by occlusion in relative depth estimation, we employ a scale-and-shift alignment technique to refine and blend the depth values, ensuring consistency in the final ground-truth depth map used for training.", + "bbox": [ + 511, + 289, + 906, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present two complementary frameworks for amodal relative depth estimation, targeting both deterministic and generative model classes. For deterministic models, we propose Amodal-DAV2, an extension of the Depth Anything V2 (DAV2) model [47]. This framework incorporates additional guidance layers to reason about depth in occluded regions. For generative models, we present Amodal-DepthFM, an adaptation of the DepthFM model [14], which uses a conditional flow matching approach to estimate depth in occluded regions. In both frameworks, we leverage the power of large pre-trained models with minimal modifications to effectively guide them toward predicting depth for occluded parts of objects.", + "bbox": [ + 511, + 487, + 908, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results indicate that our method achieves SoTA amodal depth estimation on the ADIW dataset and demonstrates strong zero-shot performance on real-world images (Fig. 1). Comprehensive ablation studies further validate the effectiveness of our design choices, showing that object-level supervision and guidance signals significantly enhance prediction accuracy. Moreover, we demonstrate the flexibility of our models in generating diverse plausible structures for occluded regions, illustrating their adaptability to different occlusion scenarios. In summary, our contributions are:", + "bbox": [ + 511, + 684, + 908, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A novel task formulation of amodal depth estimation focusing on relative depth, enabling improved generalization capabilities compared to previous metric-based", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9674", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "amodal depth estimation.", + "bbox": [ + 102, + 90, + 272, + 104 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ADIW, a large-scale real-world dataset, generated from real-world images using a scalable pipeline with segmentation datasets and compositing techniques, coupled with a scale-and-shift alignment strategy.", + "- Two novel frameworks for amodal depth estimation, Amodal-DAV2 and Amodal-DepthFM, leveraging largescale pre-trained models with minimal modifications to achieve high-quality amodal depth predictions." + ], + "bbox": [ + 89, + 106, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 243, + 230, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 89, + 268, + 349, + 285 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular depth estimation, which predicts depth from a single image, has made substantial progress in recent years [2, 4, 10, 19, 24, 46]. Early methods focused on domain-specific metric depth estimation, assuming that training and test images share similar characteristics [3, 10, 12, 23, 25]. However, this in-domain focus poses challenges for generalization to unseen domains. To address this, recent research has shifted toward cross-domain relative depth estimation, where models infer the relative depth relationships between pixels [4, 6, 34, 46, 47]. For instance, MiDaS [34] aggregates multiple datasets and trains the model in a relative depth setting, achieving superior zero-shot performance. ZoeDepth [4] demonstrates that a strong relative depth model can effectively generalize to metric depth estimation through fine-tuning. Depth Anything [46] adopts a semi-supervised strategy that combines large-scale labeled and unlabeled data to further boost the model performance. Generative approaches have also emerged, with methods like Marigold [19, 22, 53] repurposing denoising UNet models [37] alongside fixed VAE encoders [42] for depth estimation [19], and DepthFM integrates flow matching principles into the depth estimation pipeline [14]. Despite advancements in monocular depth estimation, a major challenge remains: depth estimation in occluded regions. Existing methods focus only on visible areas, leaving unseen parts unaddressed. Our work addresses this gap by enabling direct depth estimation for invisible object regions.", + "bbox": [ + 89, + 291, + 483, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Amodal Perception", + "text_level": 1, + "bbox": [ + 89, + 710, + 274, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior work in amodal perception has primarily focused on tasks such as inpainting [9, 28, 49], segmentation [20, 26, 31, 54], and detection [15, 18]. Recently, amodal depth estimation has emerged as a new challenge in amodal perception [17, 38], which aims to predict the depth of occluded regions of objects. For instance, [38] proposed the Amodal-SynthDrive dataset for synthetic driving scenes and adopted VIP-DeepLab [32] with multi-headed outputs to estimate the depth at different occlusion levels. Similarly, [17] introduced the indoor-focused Amodal-3D-FRONT dataset and an iterative approach leveraging amodal masks to estimate", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "depth in occluded areas. These works rely heavily on synthetic datasets due to the difficulty of capturing ground-truth depth behind occlusions in real-world settings. This reliance on synthetic data limits model generalization to real-world scenes, and their focus on metric depth increases the difficulty of accurate depth prediction in the zero-shot setting. In contrast, our approach adopts a relative depth-based solution using large-scale pre-trained models and realistic synthetically generated datasets, enabling strong generalization to diverse real-world images.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Depth Inpainting", + "text_level": 1, + "bbox": [ + 511, + 253, + 683, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth inpainting, or depth completion, traditionally focuses on filling in missing depth values for visible regions based on sparse depth inputs [41, 43-45, 52]. Some methods aim to inpaint depth around occlusions to create novel 3D views [11, 16, 39], this differs from amodal depth estimation, which specifically targets object-level occlusions. For example, 3D Photography [39] uses shared edge guidance to combine RGB and depth inpainting, while SLIDE [16] introduces a soft-layering strategy to preserve visual details in novel views. Recently, diffusion models have shown strong image generation capabilities [37]. Invisible Stitch [11] uses a Stable Diffusion model variant [37] to fill in missing image regions, generating novel views [30], followed by a depth painting network to fill holes in the depth map with guidance from the novel-view image. Unlike these methods, which implicitly or explicitly rely on color priors and focus on visual quality, our approach directly estimates occluded region depth, prioritizing depth accuracy independently of color information or visual coherence. This distinction shifts the focus from achieving realistic view synthesis to predicting occluded depths with high geometric accuracy, an approach that enhances scene understanding for real-world applications.", + "bbox": [ + 511, + 276, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 638, + 604, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition", + "text_level": 1, + "bbox": [ + 511, + 665, + 666, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Amodal depth estimation extends the concept of amodal segmentation by predicting depth information for occluded areas. In traditional amodal segmentation [28, 48, 49], given an input image and a segmentation mask of the visible portion of an object, the goal is to predict the complete object mask, including the occluded (invisible) part. Similarly, amodal depth estimation aims to estimate the depth values for the occluded regions, given an input observation image $I_{o}$ , a corresponding observation depth map $D_{o}$ of the input image, and a target amodal segmentation mask $M_{a}$ .", + "bbox": [ + 511, + 688, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from earlier works that formulate amodal depth estimation as a metric depth estimation task [17, 38], we propose a novel formulation where the goal is to predict the relative depth of the occluded parts of the object for", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9675", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62511e5e64361c171db9de542dd0681acf898dc4d3dbfcc758080d516d4821af.jpg", + "image_caption": [ + "Figure 3. Constructing Training Data. We use the method from [28] to convert an initial segmentation dataset into a whole-object dataset. Next, we sample and composite images to create training pairs. Due to occluders, the relative depth predictions differ between the composite and background images, so we apply scale-and-shift alignment for consistent depth blending." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 480, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the input image in the wild. Recent models, such as Depth Anything [46, 47] generate high-quality relative depth maps from natural images, allowing us to train amodal depth models using real-world data as shown in Sec. 3.2.", + "bbox": [ + 89, + 363, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Collection", + "text_level": 1, + "bbox": [ + 89, + 436, + 267, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present Amodal Depth In the Wild (ADIW), a large-scale dataset specifically designed for predicting occluded object relative depths in real-world scenes. Creating a natural image dataset at scale for amodal depth estimation is challenging due to the lack of annotations for hidden regions. No existing device can capture ground-truth depth data for occluded parts at scale in real-world settings. Previous efforts have generally relied on synthetic datasets [17, 38], which, although valuable, fall short in capturing the complexity and diversity of real-world scenes. Additionally, these approaches require manually placing occluders, limiting scalability. ADIW overcomes these by generating training data from real-world images.", + "bbox": [ + 88, + 460, + 482, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We create paired data by overlaying objects over natural images following [28], as illustrated in Fig. 3. Our method leverages segmentation datasets, which are more scalable than traditional depth estimation datasets [21, 35, 46]. Specifically, we employ the Segment Anything model [21] to automatically generate segmentation masks from the SA-1B dataset, forming the initial segmentation dataset. Next, we apply the heuristic algorithm from [28] to filter out incomplete objects, creating a whole-object dataset.", + "bbox": [ + 89, + 657, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, each image $I_{b}$ (the background image) contains at least one complete object. We then sample an occluder object $I_{f}$ (the foreground image) and superimpose it onto $I_{b}$ , forming the assembled observation image $I_{o}$ . Both $I_{o}$ and $I_{b}$ are then processed through the Depth Anything V2 model [47] (ViT-G) to obtain relative depth maps $D_{o}$ and $D_{b}$ , respectively. Importantly, the depth val", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ues for the background object in $D_{o}$ and $D_{b}$ differ due to the presence of the foreground object, which changes perceived depth. Both depth maps are scaled to the range [0, 1].", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure consistent training labels, we apply the scale-and-shift alignment algorithm [34] to align the background object's depth values across the two depth maps. The scale factor $s$ and shift factor $t$ are computed as:", + "bbox": [ + 511, + 136, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {N} \\left(s d _ {i} ^ {b} + t - d _ {i} ^ {o}\\right) ^ {2}, i \\in M _ {\\text {v i s}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 207, + 903, + 248 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_i^o$ and $d_i^b$ are the depth values of pixels in the visible part of the background object in $D_o$ and $D_b$ , respectively, and $N$ denotes the total number of valid pixels in the visible mask $M_{vis}$ of the background object. The aligned depth map $D_{aligned}$ is then calculated as:", + "bbox": [ + 511, + 260, + 903, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\text {a l i g n e d}} = s D _ {b} + t, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 348, + 903, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "serving as the ground-truth map for model training. This procedure generates a dataset of 564K images with amodal depth labels.", + "bbox": [ + 511, + 375, + 903, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown by [46], relative depth estimators [6, 34] generalize better than metric depth estimators [2, 23, 25]. Following this insight, we produce relative depth maps for amodal depth estimation in natural scenes. Moreover, our data generation approach is also adaptable for metric amodal depth estimation by omitting the scale-and-shift alignment and utilizing metric depth models [7, 29].", + "bbox": [ + 511, + 420, + 905, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Amodal Depth Estimator", + "text_level": 1, + "bbox": [ + 511, + 535, + 743, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we aim to leverage large pre-trained depth models by fine-tuning them specifically for amodal depth estimation. Our goal is to make minimal modifications to the original network architectures, preserving the capabilities of the pre-trained weights while introducing the necessary adjustments to guide the model for amodal depth prediction. By integrating additional guidance into the networks, we enable them to predict depth values for occluded parts of objects.", + "bbox": [ + 511, + 556, + 905, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth estimation models fall into two types: deterministic and generative models [13]. In this section, we introduce a dedicated strategy for adapting each model class to the amodal depth estimation task, as illustrated in Fig. 4 and Fig. 5. For the deterministic model, we adopt the Depth Anything V2 (DAV2) [47], a highly representative and top-performing pre-trained model. For the generative model, we select DepthFM [14], known for its superior ability to capture depth details and fast inference speed.", + "bbox": [ + 511, + 693, + 905, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1.Amodal-DAV2", + "text_level": 1, + "bbox": [ + 511, + 835, + 661, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Model Structure. The Amodal-DAV2 framework introduces subtle structural modifications to the Depth Anything V2 (DAV2) model's image encoder to retain its pre-trained", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9676", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1903eaed9dfb9a773c0be23d931b6240bc082cba83b75120ac537b8c27c5b333.jpg", + "image_caption": [ + "Figure 4. Amodal-DAV2 Framework Structure. Amodal-DAV2 modifies the DAV2 image encoder to take additional guidance channels along with RGB." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "knowledge while enhancing its capacity in amodal depth estimation. The original DAV2 model employs a Vision Transformer (ViT) architecture [8], which begins with an RGB convolution applied to the input image. As shown in Fig. 4, we add a Guidance Convolution (Conv) [40] layer in parallel with the RGB Conv layer, enabling the encoder to accept additional guidance channels for the observation depth map $D_{o}$ and amodal mask $M_{a}$ . To ensure compatibility with the pre-trained model, the kernel weights of the Guidance Conv layer are initialized to zero, allowing the model to initially ignore the additional guidance information. Additionally, we incorporate layer normalization into the input features of the DPT head [33] to stabilize training and improve convergence.", + "bbox": [ + 88, + 257, + 482, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training method. We fine-tune the entire Amodal-DAV2 framework during training, utilizing the standard scale-invariant loss $\\mathcal{L}_{si}$ [10] as our objective function:", + "bbox": [ + 89, + 469, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s i} = \\alpha \\sqrt {\\frac {1}{N} \\sum_ {i \\in M _ {a}} g _ {i} ^ {2} - \\frac {\\lambda}{N ^ {2}} (\\sum_ {i \\in M _ {a}} g _ {i}) ^ {2}},\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 523, + 423, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $g_{i} = \\log \\tilde{d}_{i} - \\log d_{i}$ , with $\\tilde{d}_i$ and $d_{i}$ representing the predicted depth and the ground truth depth, respectively. $N$ denotes the number of valid pixels on the amodal mask $M_{a}$ . Although our primary focus is on the accuracy of depth predictions for occluded (invisible) parts of the objects, we supervise the model using the entire object's depth. This holistic approach helps the model better understand the overall scene structure, leading to improved performance.", + "bbox": [ + 89, + 575, + 483, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Amodal-DepthFM", + "text_level": 1, + "bbox": [ + 89, + 704, + 266, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Framework and Training. We adapt DepthFM [14] for amodal depth estimation using a similar approach. Given pairs of an image $I$ , its observation depth map $D_{o}$ , and corresponding amodal mask $M_{o}$ , we fine-tune a conditional flow matching model to achieve amodal depth estimation. The objective is defined as:", + "bbox": [ + 89, + 722, + 483, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {t, z, p (x _ {0})} | | v _ {\\theta} (t, \\phi_ {t} (x _ {0})) - (x _ {1} - x _ {0}) | |,\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 824, + 434, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{1}$ represents the encoded depth samples in the latent space, and the starting point $x_{0}$ corresponds to an encoded representation of the input image. The latent flow", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0a2b84a36ae8f7e333c124ea9a87255d38fc9f50ec911b06cc8ab8798af07603.jpg", + "image_caption": [ + "Figure 5. Amodal-DepthFM Framework Structure. Amodal-DepthFM modifies the DepthFM denoising UNet encoder to take additional guidance channels along with RGB latent code." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is conditioned on the guidance provided by $z$ , which includes the input image latent code, observation depth map $D_{o}$ , and amodal mask $M_{a}$ . Here, $t$ denotes the timestamp, and $\\phi_t(x|z) = tx_1 + (1 - t)x_0$ .", + "bbox": [ + 511, + 334, + 905, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we apply noise augmentation, introducing Gaussian noise $\\mathcal{N}(x,\\sigma_{min})$ to the data samples. Consequently, the conditional probability path is modeled as $p_t(x|z) = \\mathcal{N}(x|tx_1 + (1 - t)x_0,\\sigma_{\\mathrm{min}}^2\\mathbf{I})$ . For additional details, we refer readers to [14].", + "bbox": [ + 511, + 397, + 905, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the U-Net architecture used for the flow network $v_{\\theta}$ , we modify the first input Conv layer to accommodate the additional guidance channels. Specifically, we extend the channel dimensions to accept extra guidance information $D_{o}$ and $M_{a}$ alongside the image latent code. To effectively leverage the pre-trained model, we initialize the first Conv layer with a combination of pre-trained weights for the primary eight channels (corresponding to the depth and image latent codes) and zero-initialized weights for the two additional channels dedicated to guidance inputs $D_{o}$ and $M_{a}$ .", + "bbox": [ + 511, + 474, + 905, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scale-and-Shift Alignment Inference. While the Amodal-DAV2 model directly regresses the amodal depth map, the Amodal-DepthFM model operates in the latent space, predicting the amodal depth latent code, which is then decoded into a depth map using a latent decoder. However, learning consistent amodal depth can be challenging for A modal-DepthFM. To enhance the predicted depth map, we employ the scale-and-shift alignment [34, 36] during inference.", + "bbox": [ + 511, + 627, + 905, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rather than relying solely on the model's output, we enhance the prediction by blending it with the observed depth map using a scale-and-shift alignment over the shared visible regions, as described in Eq. 1 and Eq. 2. This technique leverages the information from shared regions between the observation depth map and the amodal prediction. By calculating optimal scale and shift factors, we align the depth values of the occluded regions with the visible areas, thereby enhancing the coherence and consistency of the final depth map.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9677", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/27b6df8bd02d5b590a43cb04a5af0f9553720936bcf9722cb8901da98be82ea8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 88, + 230, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0fa99f835a456786b8a28dc0b5ab4f9c3823547461050df2a97a1e995ef5a379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 189, + 230, + 287 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/442229eb04f48ea0acaa2904722fb5551f6a4d596ed9f4382b04eee6e0f222f2.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 101, + 287, + 230, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3286d24224a3cb55896fbbc6dd93a7c1f420f3234acfdb7c1aa8f5ec2c6f69d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 88, + 362, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7bf66e680d4469415237bb7bf23718a6034cdd5487ecca3948857bd771ca0a2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 188, + 362, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08bcbf48136ef3f613985fcb4ae985123bea05428dec724ed5e6cb38f05f980e.jpg", + "image_caption": [ + "IS [11]" + ], + "image_footnote": [], + "bbox": [ + 233, + 286, + 362, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/95c903c0c91e6b1e719c4a82f2038ae328e9fae73ef7ab639060d8ab11b591e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 496, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12dfe083ac19ec224365c40ec6e5c1b79e3add275615f59a885508bd36be4ce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 188, + 496, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86385e5711cae45d605950561ae65d8d24ae4f787d813f0f832c0747b37a9d7d.jpg", + "image_caption": [ + "Jo et al. [17]" + ], + "image_footnote": [], + "bbox": [ + 367, + 286, + 496, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85ad5f0bc96ad19a4ef1d7c231c749146fbc2758369a62d5777d64ea2a2c79ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 630, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a97049af47e9c64e4a47591c6b950256ee388b319cd469fb04116be8da9b9dc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 188, + 629, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/248f706b255cf266f610749425b798051519d230381c64efc424eec691c3e5ec.jpg", + "image_caption": [ + "pix2gestalt [28]", + "Figure 6. Qualitative Comparisons on the Validation Set of ADiW. Since Invisible Stitch (IS) [11] and Jo et al. [17] use metric depth estimation models [4], the depth maps are shifted compared with other methods that use relative depth estimators. Only officially released models are taken into account in this qualitative comparison." + ], + "image_footnote": [], + "bbox": [ + 501, + 286, + 629, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eafc825da420f810ebacccecb3955628ed9a238d47621cd5eb3099a2b9ef93b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 88, + 761, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77209a82aa9c81a1ca5bc4c265e39a62062f17eae9ef77b64b86616161a7f8be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 188, + 761, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c9d5d71c10befaf9e0c6d4cb2121149e8fca189227964809a1b32eab84d842d.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 633, + 286, + 761, + 386 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c2c51cae0a7eb39fa31601768fa7fe3d274e26bfd1a9ad481c0544ce59ce94b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 88, + 895, + 188 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f6a85dda5fcaf639f232ff13b6bc5095a5a7d59318dbd9b049485d1bd6cc17b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 188, + 895, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc733a17c8ee8adba2925a9bf83eb715f5949934ab604866357147871914840b.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 764, + 286, + 895, + 386 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/078a6dbf0e9c12f1b9c62a162e9752f1f6839250e735a55f5af9bfa2ede86bdc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEasyMid.HardOverall
RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑
pix2gestalt‡[28]5.06790.3534.81890.2715.64184.0645.11488.717
Invisible Stitch‡[11]7.58483.3177.43581.3597.39380.0367.48181.781
Sekkat et al.†[38]11.19551.05211.40451.26711.15049.22211.26449.222
Sekkat et al.†‡[38]5.24588.7055.25488.7225.04487.4235.19488.367
Jo et al.†[17]10.26056.11810.22755.6379.98254.68710.17255.545
Jo et al.‡[17]7.55181.9336.63382.7336.48882.7936.93982.451
Jo et al.†‡[17]4.62489.8484.77789.9154.72887.2564.71289.177
Amodal-DAV2-S3.57493.6503.74193.6663.73492.1253.68293.251
Amodal-DAV2-B3.48293.8713.62094.0023.64092.6223.57893.590
Amodal-DAV2-L3.32494.4113.46094.5663.47693.3093.41894.179
Amodal-DepthFM5.35488.4275.39889.2415.49886.9705.41088.353
Amodal-DepthFM‡4.62292.4944.50092.9824.88391.0484.64592.295
", + "bbox": [ + 96, + 472, + 901, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparisons on ADIW Dataset. We compare our methods against other possible solutions. Results from Invisible Stitch [11] and Jo et al. [17] are not directly comparable, as they use metric depth estimation models [4]. $\\dagger$ : Models retrained on our dataset for fair comparison. $\\ddagger$ : Scale-and-shift alignment applied for consistent prediction handling (see 3.3). Note that Amodal-DAV2 does not rely on this alignment approach. Best results are in bold.", + "bbox": [ + 89, + 702, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 784, + 223, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Metrics", + "text_level": 1, + "bbox": [ + 89, + 813, + 187, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the standard evaluation protocol proposed in previous monocular metric depth estimation works [2, 10] to evaluate the effectiveness of our framework. We use the root mean squared error (RMSE), the $\\log_{10}$ error, and the", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "accuracy under the threshold $\\delta$ . Metrics consider the invisible parts of objects, with the RMSE scaled by a factor of 100 for better illustration. We also categorize objects based on their visible ratios into three difficulty levels: easy (0.75, 1], medium (0.5, 0.75], and hard (0, 0.5]. Metrics are calculated separately for each difficulty level to provide a comprehensive evaluation.", + "bbox": [ + 511, + 785, + 906, + 891 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9678", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/093e3d9537f688da919e9f816d2691290937495bacc57b0341bd498694ae186f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRMSE↓log10↓δ(%)↑
①w/o D0, Ma7.5498.60770.607
②w/o Ma4.3694.32091.193
③w/ align3.8783.65992.037
④Ours (Full)3.6823.53893.251
⑤Lsilog for inv. only3.8453.75192.608
⑥Lssi + align4.0153.94891.852
", + "bbox": [ + 101, + 88, + 475, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 273, + 307, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We split our dataset into a training set of approximately 559K samples and a validation set of 4K samples. The Amodal-DAV2 model is trained with a batch size of 32, a learning rate of $1\\mathrm{e}^{-5}$ , and 50K iterations, using the scale-invariant log (silog) loss with $\\lambda = 0.85$ , following previous works [2, 23, 25]. The Amodal-DepthFM model is trained with a batch size of 128, a learning rate of $3\\mathrm{e}^{-5}$ , and 15K iterations. Both models are initialized with depth-pretrained parameters before fine-tuning for a modal depth estimation. We use the Adam optimizer with exponential learning rate decay. To stabilize training, we apply a max gradient norm clip of 0.01. Data augmentation is minimal, with only horizontal flipping. All experiments are conducted on 4 NVIDIA A100 GPUs. By default, we evaluate model performance using the final checkpoint after training. We use our Amodal-DAV2-L for all visualization results as default.", + "bbox": [ + 88, + 295, + 482, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 546, + 228, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Amodal Depth Methods. We compare our methods with existing amodal depth estimation approaches. As these methods were initially trained on synthetic datasets and relied on metric depth formulations, we retrained them on our dataset and adapted them to relative depth estimation. Results in Tab. 1 and Fig. 6 demonstrate a significant performance gap between the previous SoTA amodal depth model[17] and our methods, even after output alignment with our strategy. Notably, Amodal-DepthFM outperforms prior approaches even without the alignment, while our best model, Amodal-DAV2-L, improves performance by $27.4\\%$ in terms of RMSE, setting a new SoTA. Previous methods, designed for amodal metric depth estimation, struggle to generalize across varied natural image depth ranges. In contrast, our approach leverages pretrained DAV2 [47] and DepthFM [14] models, which benefit from extensive prior knowledge of geometry and color, resulting in more accurate amodal depth predictions.", + "bbox": [ + 91, + 568, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Other Solutions. We also evaluate two alternative solutions: Invisible Stitch [11] and pix2gestalt [28]. (1) For Invisible Stitch, we use SDXL [30] with a ground-truth amodal mask and image cap", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6e16d07fb62dce682f238dd1372f3c44f9637939628bed54f5a4fd8ad9c49111.jpg", + "table_caption": [ + "Table 2. Ablation Study for Amodal-DAV2. We investigate the impact of different guidance signals, supervision strategies, and inference techniques." + ], + "table_footnote": [], + "table_body": "
MethodRMSE↓log10↓δ(%)↑
①w/o align & Do, Ma10.21112.56356.271
②w/o align & Ma5.2836.23183.799
③w/o align5.4105.23588.353
④Ours (Full)4.6453.55392.295
⑤Lrmse for inv. only5.6363.91191.739
⑥Lrmse for scene4.6083.80991.149
", + "bbox": [ + 522, + 89, + 897, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation Study for Amodal-DepthFM. We explore the effectiveness of various guidance signals, supervision strategies, and the alignment inference for Amodal-DepthFM. Unlike Amodal-DAV2, the alignment strategy significantly improves the performance of Amodal-DepthFM.", + "bbox": [ + 511, + 202, + 905, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion to guide the inpainting. The inpainted image serves as guidance for the stitcher model [11] to perform depth inpainting, and the predicted depth is aligned with the observation depth map. (2) For pix2gestalt [28], we conduct amodal inpainting followed by depth estimation using DAV2 (ViT-G), aligning the predicted depth with the observation depth. Additional implementation details are in the supplementary materials.", + "bbox": [ + 511, + 301, + 903, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Fig. 6 and Fig. 7, the accuracy of amodal depth estimation relies heavily on the quality of RGB inpainting in both methods. Even with the ground-truth amodal mask, SD-XL struggles to accurately inpaint target areas, and the stretcher model fails to predict accurate depth for occluded parts, leaving ghosting artifacts in the predicted depth map. Similarly, pix2gestalt suffers cascade errors due to inaccurate inpainting with uncontrollable amodal shapes. Moreover, since DAV2 is trained on complete natural images, applying it directly to inpainted outputs from pix2gestalt (with single foreground objects on white backgrounds) leads to performance degradation. For example, the window boundary in the first case in Fig. 6 is missing from the predicted depth map. In contrast, our amodal depth models directly regress the depth of invisible parts without relying on RGB information, using only the amodal mask as guidance. These results demonstrate that our approach provides strong geometric priors, which could also serve as a useful condition for inpainting methods [51].", + "bbox": [ + 511, + 424, + 905, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies and Discussion", + "text_level": 1, + "bbox": [ + 511, + 724, + 795, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Guidance and Supervision Strategies. We conduct ablation experiments on Amodal-DAV2-S and Amodal-DepthFM to assess each framework component's contribution, with results in Tab. 2 and Tab. 3, respectively. For both models, guidance from the observation depth map $D_{o}$ and amodal mask $M_{a}$ is crucial for optimal performance (1, 2). While the primary goal and evaluation protocol focus is estimating depth for occluded parts of objects, both frameworks benefit from object-level supervision that includes visible regions (4, 5). Interestingly, the scale-and-shift align", + "bbox": [ + 511, + 750, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9679", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fe3af7970e7b7418b6f075b596e070705a2f1d952bb392f4284f12de286de9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 209, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2bd26d3e13cbed7dff88f3938f31074d92338017808152b9bba10252b2446d12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 172, + 207, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b7b19e74f08faaf30d2d5f2b486a7eb7fb559a2328847ba426471508099ffd9b.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 101, + 256, + 209, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0e5a45974ccadbf09823dd72d819a981d9332a6f2bd352cb176e64e5bd7641f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 90, + 323, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0c4c9f6185e8491b4193ead3acc74fdae37351f7c9f9099af64638b921471d86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 215, + 172, + 323, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/05b2221ce96e24779eac7e310790e92b738afc49b920ce75c660e85412a16b7f.jpg", + "image_caption": [ + "SD-XL [30]" + ], + "image_footnote": [], + "bbox": [ + 215, + 256, + 323, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f15514eafe16a9afcaf31c919c76ed90c2c2b25a29e4bed393d07ecf1f955e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 90, + 437, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ca773620b14d2ca46110baa10c722a002f6775bfccd9a7f0f56afb7f56de8750.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 172, + 437, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51e09fbf810bcd1eff2b47293744fa1ae41843f6de731dc9c6ffdeb495e1746e.jpg", + "image_caption": [ + "IS [11]" + ], + "image_footnote": [], + "bbox": [ + 331, + 256, + 437, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97f7e436e5a43da0e90912a6f88b6e8a4271b33f4334165c1878db4937b5f39c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 90, + 552, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a9c3a8bc3e96d445e2a4fc0ce2e6b600ba83239076a401adb26074cf05f6ca6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 446, + 172, + 550, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83f256ccdd6dfb5f8d7652cac1f6564fb6e4ec98571752e01b1b6a7f3a6ad0ef.jpg", + "image_caption": [ + "pix2gestalt [28]" + ], + "image_footnote": [], + "bbox": [ + 446, + 256, + 550, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/683a0f9b3a41e48f33d2409f4584a48b4c6ac5d745775aef971b23726578eedd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 90, + 666, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b9e32f1a5c685e44ae78cd92909188cb4c7b01a0033a19e4c91aa13a734fb70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 172, + 666, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d54c8718535dcdff9abd4bbbf84d85ba21e90713865277b2145f7495e69e0832.jpg", + "image_caption": [ + "pix2gestalt [28]" + ], + "image_footnote": [], + "bbox": [ + 560, + 256, + 666, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d26a5b16a6f7fba7c337730e538bfa509cb15d1f92d64d3c8d9ffa73915ebfc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 90, + 779, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2278828038f0232a7459cce193c55ec759924a39b53bb78bca47b6f0fc1cbac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 673, + 172, + 779, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d852c171bd387eecba0ffa12672ec666abfbec47babea4495bbcd79b093026b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 673, + 256, + 779, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82808fb121eaf921de0fb957cb8a11f571f9fe51f7302519100cc3d08a872cec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 90, + 895, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3eba01883d4649c04f9433c9ca77c06ee275f667dc2e1252822da2a354eebb93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 172, + 895, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4e2a0e71bf0d5c416a0bb1ca7f86c8dff73e53778232b38a12aea94cb6c9cd6f.jpg", + "image_caption": [ + "$D_{o}$" + ], + "image_footnote": [], + "bbox": [ + 789, + 256, + 895, + 337 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36fc49f35d9c6fb1880fce4f30b0951b185a8b42385785201dfeb6b2e11651ad.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparisons on Images in the Wild. Invisible Stitch (IS) [11] uses SD-XL [30] with a ground-truth visible mask and image caption to inpaint occluded areas, while pix2gestalt [28] completes the occluded areas via amodal inpainting. Both methods suffer from inaccurate shape completion and cascade depth errors. The third row highlights the fragile shape completion of previous work: both SD-XL and pix2gestalt fail to remove the foreground cabinets correctly. In the Human Heuristics mode (Fig. 2), our method infers without any RGB inpainting or completion prior and achieves reasonable depth estimations for occluded areas." + ], + "image_footnote": [], + "bbox": [ + 93, + 464, + 191, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56a121a7613b4099c1e03fe992fe356ae89ab9c38410e0f1ea75345fe3c95f66.jpg", + "image_caption": [ + "Figure 8. Results with Different Mask Guidance. Guided by various amodal masks, our model successfully predicts the corresponding amodal depth maps for the target objects in the image, showcasing flexibility in estimating occluded regions." + ], + "image_footnote": [], + "bbox": [ + 191, + 464, + 287, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04b428f399dc37639a109ec5826d6aabe374656eec0b061cbd11a28da6e1bee6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 464, + 385, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbe1f1eb3ffe267a4955a01c453509d7008fbe38aceb101ced48b1f346e78ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 464, + 480, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment during inference improves Amodal-DepthFM performance but decreases that of Amodal-DAV2-S, indicating that inconsistent depth estimation may hinder Amodal-DepthFM's overall performance (③).", + "bbox": [ + 89, + 625, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Varying Predictions with Different Guidance. Fig. 8 demonstrates the flexibility of our amodal depth estimation model in generating diverse, plausible predictions based on different amodal masks as guidance. This capability is crucial when multiple valid interpretations exist for occluded object parts. For example, the right-hand chair occluded by a wall could vary in width or even be connected to another occluded chair on the left. Our Amodal-DAV2-L adapts the amodal depth output according to the provided mask.", + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Edge Region v.s., Non-Edge Region. In this experiment, we extract edge areas from the ground-truth depth map and calculate metrics for both edge and non-edge regions in the invisible area. The overall RMSE from Amodal-DAV2-S is 3.682, with the RMSE for the edge region significantly", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "higher at 7.378, compared to only 3.515 for non-edge regions. This discrepancy highlights the challenges in estimating amodal depth with fine-grained details. This finding aligns with human perception, where it is generally more difficult to accurately recover fine-grained geometry than the coarser structures for occluded areas.", + "bbox": [ + 511, + 465, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 571, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a novel approach to amodal depth estimation, focusing on predicting the depth of invisible parts of objects in natural scenes. Our work introduces Amodal Depth In the Wild (ADIW), a large-scale dataset that leverages segmentation datasets and a compositing pipeline, enabling high-quality, real-world amodal depth annotations. We proposed two complementary frameworks for amodal depth estimation: Amodal-DAV2 and Amodal-DepthFM. Amodal-DAV2 leverages the deterministic capabilities of Depth Anything V2 to achieve state-of-the-art performance in relative depth estimation, while A modal-DepthFM, built on a generative flow matching paradigm, and excels in providing finer details and sharper boundaries in occluded regions. Our experiments highlight that both models benefit from object-level supervision and the importance of guidance signals for improving amodal depth prediction accuracy. This work not only sets a new benchmark for amodal depth estimation but also opens the door for future research in occluded geometry understanding and improving applications such as inpainting and scene reconstruction.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9680", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This publication is supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940.", + "bbox": [ + 89, + 114, + 483, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiayang Ao, Qiuhong Ke, and Krista A. Ehinger. Image amodal completion: A survey. In Computer Vision and Image Understanding. Elsevier, 2023. 2", + "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1, 3, 4, 6, 7", + "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3", + "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 3, 6", + "[5] Amlaan Bhoi. Monocular depth estimation: A survey. arXiv preprint arXiv:1901.09402, 2019. 1", + "[6] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3.1 - a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 2, 3, 4", + "[7] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 4", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[9] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, pages 6144-6153, 2018. 3", + "[10] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 1, 3, 5, 6", + "[11] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 2, 3, 6, 7, 8", + "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 3", + "[13] Yongtao Ge, Guangkai Xu, Zhiyue Zhao, Libo Sun, Zheng Huang, Yanlong Sun, Hao Chen, and Chunhua Shen. Geobench: Benchmarking and analyzing monocular geometry estimation models. arXiv preprint arXiv:2406.12671, 2024. 4" + ], + "bbox": [ + 93, + 213, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 2, 3, 4, 5, 7", + "[15] Cheng-Yen Hsieh, Tarasha Khurana, Achal Dave, and Deva Ramanan. Tracking any object amodally. arXiv preprint arXiv:2312.12433, 2023. 3", + "[16] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12518-12527, 2021. 3", + "[17] Seong-Uk Jo, Du Yeol Lee, and Chae Eun Rhee. Occlusion-aware amodal depth estimation for enhancing 3d reconstruction from a single image. IEEE Access, 2024. 2, 3, 4, 6, 7", + "[18] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Amodal completion and size constancy in natural scenes. In ICCV, pages 127-135, 2015. 3", + "[19] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, pages 9492-9502, 2024. 3", + "[20] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, pages 4019–4028, 2021. 3", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4", + "[22] Mykola Lavreniuk, Shariq Farooq Bhat, Matthias Muller, and Peter Wonka. Evp: Enhanced visual perception using inverse multi-attentive feature refinement and regularized image-text alignment. In European Conference on Computer Vision Workshops (ECCVW), 2024. 3", + "[23] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 1, 3, 4, 7", + "[24] Zhenyu Li, Shariq Farooq Bhat, and Peter Wonka. Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284, 2023. 3", + "[25] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, pages 1-18, 2023. 3, 4, 7", + "[26] Ruoshi Liu and Carl Vondrick. Humans as light bulbs: 3d human reconstruction from thermal reflection. In CVPR, pages 12531-12542, 2023. 3", + "[27] Alican Mertan, Damien Jade Duff, and Gozde Unal. Single image depth estimation: An overview. Digital Signal Processing, 123:103441, 2022. 1", + "[28] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt:" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9681", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amodal segmentation by synthesizing wholes. In CVPR, pages 3931-3940. IEEE Computer Society, 2024. 2, 3, 4, 6, 7, 8", + "[29] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116, 2024. 2, 4", + "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8", + "[31] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, pages 3014-3023, 2019. 3", + "[32] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, pages 3997-4008, 2021. 3", + "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 5", + "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2022. 2, 3, 4, 5", + "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4", + "[36] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 5", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3", + "[38] Ahmed Rida Sekkat, Rohit Mohan, Oliver Sawade, Elmar Matthes, and Abhinav Valada. Amodalsynthdrive: A synthetic amodal perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 2024. 2, 3, 4, 6", + "[39] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In CVPR, pages 8028-8038, 2020. 2, 3", + "[40] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 5", + "[41] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 3", + "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3", + "[43] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In CVPR, pages 9422-9432, 2023. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024.", + "[45] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 3", + "[46] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891, 2024. 1, 2, 3, 4", + "[47] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 2, 3, 4, 7", + "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, pages 28003-28013, 2024. 2, 3", + "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2, 3", + "[50] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2", + "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 7", + "[52] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 3", + "[53] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5729-5739, 2023. 3", + "[54] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In CVPR, pages 1464-1472, 2017. 3" + ], + "bbox": [ + 516, + 92, + 903, + 643 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_model.json b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_model.json new file mode 100644 index 0000000000000000000000000000000000000000..351029b859dfba84a80912d643dc04dca220b583 --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.823, + 0.154 + ], + "angle": 0, + "content": "Amodal Depth Anything: Amodal Depth Estimation in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.18, + 0.812, + 0.218 + ], + "angle": 0, + "content": "Zhenyu Li\\(^{1}\\), Mykola Lavreniuk\\(^{2}\\), Jian Shi\\(^{1}\\), Shariq Farooq Bhat\\(^{1}\\), Peter Wonka\\(^{1}\\) \n\\(^{1}\\)KAUST, \\(^{2}\\)Space Research Institute NASU-SSAU" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.251, + 0.904, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.53, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Figure 1. Amodal Depth Estimation in the Wild. For each image, we present the amodal depth estimation with the target object outlined in black. Our model demonstrates strong generalization across diverse scenes, including both indoor and outdoor environments for accurate depth estimation for occluded parts of objects. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.586, + 0.33, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.753 + ], + "angle": 0, + "content": "two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions (Fig. 1). Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a \\(50.7\\%\\) improvement in RMSE over the previous SoTA on the ADIW dataset." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.646, + 0.8 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Monocular depth estimation is a foundational task in computer vision and generative modeling [5, 10, 27] as it provides depth perception from a single image without stereo cues. However, while recent methods, e.g. [2, 4, 10, 23, 46], focus solely on estimating depth for visible pixels, humans can intuitively perceive the complete 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.485, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Amodal Depth Estimation Pipeline. Given an input image, users can generate the amodal mask for the depth estimator in two ways: (1) Model Heuristics: click the target object, apply SAM [21] to generate modal mask, then use amodal segmentation methods to estimate amodal mask, (2) Human Heuristics: manually draw the amodal mask. Our model estimates amodal depth based on original observation image \\( I_{o} \\), the observed depth map \\( D_{o} \\), and the amodal mask \\( M_{a} \\). Note that we can get visible and invisible masks from both pipelines. Our model only focuses on the amodal depth (i.e., invisible part), whereas the visible part depth can be directly derived from the observed depth map \\( D_{o} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.583 + ], + "angle": 0, + "content": "geometry of objects, even when only parts of them are visible. Amodal depth estimation is the task of predicting depth values for the occluded (invisible) parts of objects [17, 38]. It enables accurate 3D reconstruction and novel view synthesis from a single view, significantly reducing dependence on costly multi-view systems or LiDAR [1]. This capability is crucial for advancing AR/VR, robotics, autonomous driving and digital twin technologies [1]. In this paper, we aim to address this under-explored question: Can models effectively perceive the geometry of the invisible parts of objects in a scene?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Unlike inpainting methods [11, 39, 50], which reconstruct missing image regions, amodal depth estimation is a novel task that extends traditional amodal segmentation [28, 48, 49] into the depth domain, by predicting the depth of occluded object parts. In this task, given an input image with a target amodal mask, the objective is to infer depth values for the occluded object regions. While amodal segmentation benefits from human annotations to collect training samples [17, 38], there is currently no device capable of collecting ground-truth depth data for occluded parts of objects at scale in real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address this challenge, prior approaches to amodal depth estimation have relied on synthetic datasets [17, 38]. However, generating these datasets is both time-consuming and hard to scale, often requiring manually placing occluders one by one [17, 38]. Additionally, synthetic data lacks the complexity and diversity of real-world scenes, resulting in a domain gap that limits the generalization of models trained on such data. Furthermore, these previous approaches only consider metric depth, where the goal is to estimate real-world distances for occluded regions. Re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "liaence on metric depth, which naturally struggles to generalize to unseen scenarios with limited data [7, 29], exacerbates the models' poor zero-shot performance on real-world images [6, 34, 46]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.289 + ], + "angle": 0, + "content": "To overcome these limitations, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth. Recent advancements in depth estimation models, such as Depth Anything [46], have enabled the generation of high-quality relative depth maps from natural images. Focusing on relative depth allows us to train models using real-world data, leveraging the depth relationships within scenes to achieve better generalization without relying on precise metric measurements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.908, + 0.487 + ], + "angle": 0, + "content": "We introduce a new large-scale real-world dataset, Amodal Depth In the Wild (ADIW), to facilitate the training of models for amodal depth estimation. Our data generation pipeline leverages segmentation datasets to create high-quality datasets for amodal relative depth estimation with scalable annotations. We adopt a compositing approach to create training pairs, where objects are sampled and composited into scenes. Corresponding depth maps are estimated using the large depth foundation model [47]. Given the challenges posed by occlusion in relative depth estimation, we employ a scale-and-shift alignment technique to refine and blend the depth values, ensuring consistency in the final ground-truth depth map used for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.684 + ], + "angle": 0, + "content": "In this work, we present two complementary frameworks for amodal relative depth estimation, targeting both deterministic and generative model classes. For deterministic models, we propose Amodal-DAV2, an extension of the Depth Anything V2 (DAV2) model [47]. This framework incorporates additional guidance layers to reason about depth in occluded regions. For generative models, we present Amodal-DepthFM, an adaptation of the DepthFM model [14], which uses a conditional flow matching approach to estimate depth in occluded regions. In both frameworks, we leverage the power of large pre-trained models with minimal modifications to effectively guide them toward predicting depth for occluded parts of objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.851 + ], + "angle": 0, + "content": "Experimental results indicate that our method achieves SoTA amodal depth estimation on the ADIW dataset and demonstrates strong zero-shot performance on real-world images (Fig. 1). Comprehensive ablation studies further validate the effectiveness of our design choices, showing that object-level supervision and guidance signals significantly enhance prediction accuracy. Moreover, we demonstrate the flexibility of our models in generating diverse plausible structures for occluded regions, illustrating their adaptability to different occlusion scenarios. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "- A novel task formulation of amodal depth estimation focusing on relative depth, enabling improved generalization capabilities compared to previous metric-based" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.274, + 0.106 + ], + "angle": 0, + "content": "amodal depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.107, + 0.482, + 0.167 + ], + "angle": 0, + "content": "- ADIW, a large-scale real-world dataset, generated from real-world images using a scalable pipeline with segmentation datasets and compositing techniques, coupled with a scale-and-shift alignment strategy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.482, + 0.227 + ], + "angle": 0, + "content": "- Two novel frameworks for amodal depth estimation, Amodal-DAV2 and Amodal-DepthFM, leveraging largescale pre-trained models with minimal modifications to achieve high-quality amodal depth predictions." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.107, + 0.482, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.244, + 0.232, + 0.259 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.269, + 0.351, + 0.286 + ], + "angle": 0, + "content": "2.1. Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.485, + 0.701 + ], + "angle": 0, + "content": "Monocular depth estimation, which predicts depth from a single image, has made substantial progress in recent years [2, 4, 10, 19, 24, 46]. Early methods focused on domain-specific metric depth estimation, assuming that training and test images share similar characteristics [3, 10, 12, 23, 25]. However, this in-domain focus poses challenges for generalization to unseen domains. To address this, recent research has shifted toward cross-domain relative depth estimation, where models infer the relative depth relationships between pixels [4, 6, 34, 46, 47]. For instance, MiDaS [34] aggregates multiple datasets and trains the model in a relative depth setting, achieving superior zero-shot performance. ZoeDepth [4] demonstrates that a strong relative depth model can effectively generalize to metric depth estimation through fine-tuning. Depth Anything [46] adopts a semi-supervised strategy that combines large-scale labeled and unlabeled data to further boost the model performance. Generative approaches have also emerged, with methods like Marigold [19, 22, 53] repurposing denoising UNet models [37] alongside fixed VAE encoders [42] for depth estimation [19], and DepthFM integrates flow matching principles into the depth estimation pipeline [14]. Despite advancements in monocular depth estimation, a major challenge remains: depth estimation in occluded regions. Existing methods focus only on visible areas, leaving unseen parts unaddressed. Our work addresses this gap by enabling direct depth estimation for invisible object regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.275, + 0.729 + ], + "angle": 0, + "content": "2.2. Amodal Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prior work in amodal perception has primarily focused on tasks such as inpainting [9, 28, 49], segmentation [20, 26, 31, 54], and detection [15, 18]. Recently, amodal depth estimation has emerged as a new challenge in amodal perception [17, 38], which aims to predict the depth of occluded regions of objects. For instance, [38] proposed the Amodal-SynthDrive dataset for synthetic driving scenes and adopted VIP-DeepLab [32] with multi-headed outputs to estimate the depth at different occlusion levels. Similarly, [17] introduced the indoor-focused Amodal-3D-FRONT dataset and an iterative approach leveraging amodal masks to estimate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "depth in occluded areas. These works rely heavily on synthetic datasets due to the difficulty of capturing ground-truth depth behind occlusions in real-world settings. This reliance on synthetic data limits model generalization to real-world scenes, and their focus on metric depth increases the difficulty of accurate depth prediction in the zero-shot setting. In contrast, our approach adopts a relative depth-based solution using large-scale pre-trained models and realistic synthetically generated datasets, enabling strong generalization to diverse real-world images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.254, + 0.684, + 0.271 + ], + "angle": 0, + "content": "2.3. Depth Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.908, + 0.626 + ], + "angle": 0, + "content": "Depth inpainting, or depth completion, traditionally focuses on filling in missing depth values for visible regions based on sparse depth inputs [41, 43-45, 52]. Some methods aim to inpaint depth around occlusions to create novel 3D views [11, 16, 39], this differs from amodal depth estimation, which specifically targets object-level occlusions. For example, 3D Photography [39] uses shared edge guidance to combine RGB and depth inpainting, while SLIDE [16] introduces a soft-layering strategy to preserve visual details in novel views. Recently, diffusion models have shown strong image generation capabilities [37]. Invisible Stitch [11] uses a Stable Diffusion model variant [37] to fill in missing image regions, generating novel views [30], followed by a depth painting network to fill holes in the depth map with guidance from the novel-view image. Unlike these methods, which implicitly or explicitly rely on color priors and focus on visual quality, our approach directly estimates occluded region depth, prioritizing depth accuracy independently of color information or visual coherence. This distinction shifts the focus from achieving realistic view synthesis to predicting occluded depths with high geometric accuracy, an approach that enhances scene understanding for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.64, + 0.605, + 0.656 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.667, + 0.681 + ], + "angle": 0, + "content": "3.1. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Amodal depth estimation extends the concept of amodal segmentation by predicting depth information for occluded areas. In traditional amodal segmentation [28, 48, 49], given an input image and a segmentation mask of the visible portion of an object, the goal is to predict the complete object mask, including the occluded (invisible) part. Similarly, amodal depth estimation aims to estimate the depth values for the occluded regions, given an input observation image \\( I_{o} \\), a corresponding observation depth map \\( D_{o} \\) of the input image, and a target amodal segmentation mask \\( M_{a} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Different from earlier works that formulate amodal depth estimation as a metric depth estimation task [17, 38], we propose a novel formulation where the goal is to predict the relative depth of the occluded parts of the object for" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9675" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.482, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.485, + 0.336 + ], + "angle": 0, + "content": "Figure 3. Constructing Training Data. We use the method from [28] to convert an initial segmentation dataset into a whole-object dataset. Next, we sample and composite images to create training pairs. Due to occluders, the relative depth predictions differ between the composite and background images, so we apply scale-and-shift alignment for consistent depth blending." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.484, + 0.424 + ], + "angle": 0, + "content": "the input image in the wild. Recent models, such as Depth Anything [46, 47] generate high-quality relative depth maps from natural images, allowing us to train amodal depth models using real-world data as shown in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.269, + 0.452 + ], + "angle": 0, + "content": "3.2. Dataset Collection" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.656 + ], + "angle": 0, + "content": "We present Amodal Depth In the Wild (ADIW), a large-scale dataset specifically designed for predicting occluded object relative depths in real-world scenes. Creating a natural image dataset at scale for amodal depth estimation is challenging due to the lack of annotations for hidden regions. No existing device can capture ground-truth depth data for occluded parts at scale in real-world settings. Previous efforts have generally relied on synthetic datasets [17, 38], which, although valuable, fall short in capturing the complexity and diversity of real-world scenes. Additionally, these approaches require manually placing occluders, limiting scalability. ADIW overcomes these by generating training data from real-world images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.794 + ], + "angle": 0, + "content": "We create paired data by overlaying objects over natural images following [28], as illustrated in Fig. 3. Our method leverages segmentation datasets, which are more scalable than traditional depth estimation datasets [21, 35, 46]. Specifically, we employ the Segment Anything model [21] to automatically generate segmentation masks from the SA-1B dataset, forming the initial segmentation dataset. Next, we apply the heuristic algorithm from [28] to filter out incomplete objects, creating a whole-object dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our approach, each image \\(I_{b}\\) (the background image) contains at least one complete object. We then sample an occluder object \\(I_{f}\\) (the foreground image) and superimpose it onto \\(I_{b}\\), forming the assembled observation image \\(I_{o}\\). Both \\(I_{o}\\) and \\(I_{b}\\) are then processed through the Depth Anything V2 model [47] (ViT-G) to obtain relative depth maps \\(D_{o}\\) and \\(D_{b}\\), respectively. Importantly, the depth val" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ues for the background object in \\(D_{o}\\) and \\(D_{b}\\) differ due to the presence of the foreground object, which changes perceived depth. Both depth maps are scaled to the range [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.198 + ], + "angle": 0, + "content": "To ensure consistent training labels, we apply the scale-and-shift alignment algorithm [34] to align the background object's depth values across the two depth maps. The scale factor \\( s \\) and shift factor \\( t \\) are computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.208, + 0.905, + 0.249 + ], + "angle": 0, + "content": "\\[\n(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {N} \\left(s d _ {i} ^ {b} + t - d _ {i} ^ {o}\\right) ^ {2}, i \\in M _ {\\text {v i s}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.905, + 0.337 + ], + "angle": 0, + "content": "where \\(d_i^o\\) and \\(d_i^b\\) are the depth values of pixels in the visible part of the background object in \\(D_o\\) and \\(D_b\\), respectively, and \\(N\\) denotes the total number of valid pixels in the visible mask \\(M_{vis}\\) of the background object. The aligned depth map \\(D_{aligned}\\) is then calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.639, + 0.349, + 0.905, + 0.365 + ], + "angle": 0, + "content": "\\[\nD _ {\\text {a l i g n e d}} = s D _ {b} + t, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.905, + 0.42 + ], + "angle": 0, + "content": "serving as the ground-truth map for model training. This procedure generates a dataset of 564K images with amodal depth labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.906, + 0.527 + ], + "angle": 0, + "content": "As shown by [46], relative depth estimators [6, 34] generalize better than metric depth estimators [2, 23, 25]. Following this insight, we produce relative depth maps for amodal depth estimation in natural scenes. Moreover, our data generation approach is also adaptable for metric amodal depth estimation by omitting the scale-and-shift alignment and utilizing metric depth models [7, 29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.744, + 0.551 + ], + "angle": 0, + "content": "3.3. Amodal Depth Estimator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.906, + 0.693 + ], + "angle": 0, + "content": "In this work, we aim to leverage large pre-trained depth models by fine-tuning them specifically for amodal depth estimation. Our goal is to make minimal modifications to the original network architectures, preserving the capabilities of the pre-trained weights while introducing the necessary adjustments to guide the model for amodal depth prediction. By integrating additional guidance into the networks, we enable them to predict depth values for occluded parts of objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.906, + 0.83 + ], + "angle": 0, + "content": "Depth estimation models fall into two types: deterministic and generative models [13]. In this section, we introduce a dedicated strategy for adapting each model class to the amodal depth estimation task, as illustrated in Fig. 4 and Fig. 5. For the deterministic model, we adopt the Depth Anything V2 (DAV2) [47], a highly representative and top-performing pre-trained model. For the generative model, we select DepthFM [14], known for its superior ability to capture depth details and fast inference speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.663, + 0.85 + ], + "angle": 0, + "content": "3.3.1.Amodal-DAV2" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Model Structure. The Amodal-DAV2 framework introduces subtle structural modifications to the Depth Anything V2 (DAV2) model's image encoder to retain its pre-trained" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9676" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.091, + 0.482, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.191, + 0.485, + 0.234 + ], + "angle": 0, + "content": "Figure 4. Amodal-DAV2 Framework Structure. Amodal-DAV2 modifies the DAV2 image encoder to take additional guidance channels along with RGB." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.258, + 0.483, + 0.469 + ], + "angle": 0, + "content": "knowledge while enhancing its capacity in amodal depth estimation. The original DAV2 model employs a Vision Transformer (ViT) architecture [8], which begins with an RGB convolution applied to the input image. As shown in Fig. 4, we add a Guidance Convolution (Conv) [40] layer in parallel with the RGB Conv layer, enabling the encoder to accept additional guidance channels for the observation depth map \\(D_{o}\\) and amodal mask \\(M_{a}\\). To ensure compatibility with the pre-trained model, the kernel weights of the Guidance Conv layer are initialized to zero, allowing the model to initially ignore the additional guidance information. Additionally, we incorporate layer normalization into the input features of the DPT head [33] to stabilize training and improve convergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Training method. We fine-tune the entire Amodal-DAV2 framework during training, utilizing the standard scale-invariant loss \\(\\mathcal{L}_{si}\\) [10] as our objective function:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.525, + 0.424, + 0.566 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s i} = \\alpha \\sqrt {\\frac {1}{N} \\sum_ {i \\in M _ {a}} g _ {i} ^ {2} - \\frac {\\lambda}{N ^ {2}} (\\sum_ {i \\in M _ {a}} g _ {i}) ^ {2}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.7 + ], + "angle": 0, + "content": "where \\( g_{i} = \\log \\tilde{d}_{i} - \\log d_{i} \\), with \\( \\tilde{d}_i \\) and \\( d_{i} \\) representing the predicted depth and the ground truth depth, respectively. \\( N \\) denotes the number of valid pixels on the amodal mask \\( M_{a} \\). Although our primary focus is on the accuracy of depth predictions for occluded (invisible) parts of the objects, we supervise the model using the entire object's depth. This holistic approach helps the model better understand the overall scene structure, leading to improved performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.705, + 0.267, + 0.72 + ], + "angle": 0, + "content": "3.3.2. Amodal-DepthFM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.814 + ], + "angle": 0, + "content": "Framework and Training. We adapt DepthFM [14] for amodal depth estimation using a similar approach. Given pairs of an image \\(I\\), its observation depth map \\(D_{o}\\), and corresponding amodal mask \\(M_{o}\\), we fine-tune a conditional flow matching model to achieve amodal depth estimation. The objective is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.825, + 0.435, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {t, z, p (x _ {0})} | | v _ {\\theta} (t, \\phi_ {t} (x _ {0})) - (x _ {1} - x _ {0}) | |,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( x_{1} \\) represents the encoded depth samples in the latent space, and the starting point \\( x_{0} \\) corresponds to an encoded representation of the input image. The latent flow" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.263, + 0.908, + 0.307 + ], + "angle": 0, + "content": "Figure 5. Amodal-DepthFM Framework Structure. Amodal-DepthFM modifies the DepthFM denoising UNet encoder to take additional guidance channels along with RGB latent code." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.397 + ], + "angle": 0, + "content": "is conditioned on the guidance provided by \\(z\\), which includes the input image latent code, observation depth map \\(D_{o}\\), and amodal mask \\(M_{a}\\). Here, \\(t\\) denotes the timestamp, and \\(\\phi_t(x|z) = tx_1 + (1 - t)x_0\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.906, + 0.473 + ], + "angle": 0, + "content": "During training, we apply noise augmentation, introducing Gaussian noise \\(\\mathcal{N}(x,\\sigma_{min})\\) to the data samples. Consequently, the conditional probability path is modeled as \\(p_t(x|z) = \\mathcal{N}(x|tx_1 + (1 - t)x_0,\\sigma_{\\mathrm{min}}^2\\mathbf{I})\\). For additional details, we refer readers to [14]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.627 + ], + "angle": 0, + "content": "In the U-Net architecture used for the flow network \\( v_{\\theta} \\), we modify the first input Conv layer to accommodate the additional guidance channels. Specifically, we extend the channel dimensions to accept extra guidance information \\( D_{o} \\) and \\( M_{a} \\) alongside the image latent code. To effectively leverage the pre-trained model, we initialize the first Conv layer with a combination of pre-trained weights for the primary eight channels (corresponding to the depth and image latent codes) and zero-initialized weights for the two additional channels dedicated to guidance inputs \\( D_{o} \\) and \\( M_{a} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Scale-and-Shift Alignment Inference. While the Amodal-DAV2 model directly regresses the amodal depth map, the Amodal-DepthFM model operates in the latent space, predicting the amodal depth latent code, which is then decoded into a depth map using a latent decoder. However, learning consistent amodal depth can be challenging for A modal-DepthFM. To enhance the predicted depth map, we employ the scale-and-shift alignment [34, 36] during inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Rather than relying solely on the model's output, we enhance the prediction by blending it with the observed depth map using a scale-and-shift alignment over the shared visible regions, as described in Eq. 1 and Eq. 2. This technique leverages the information from shared regions between the observation depth map and the amodal prediction. By calculating optimal scale and shift factors, we align the depth values of the occluded regions with the visible areas, thereby enhancing the coherence and consistency of the final depth map." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.231, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.19, + 0.231, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.288, + 0.231, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.391, + 0.184, + 0.404 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.089, + 0.364, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.189, + 0.364, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.287, + 0.364, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.391, + 0.322, + 0.404 + ], + "angle": 0, + "content": "IS [11]" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.497, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.189, + 0.497, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.287, + 0.497, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.391, + 0.471, + 0.404 + ], + "angle": 0, + "content": "Jo et al. [17]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.631, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.189, + 0.63, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.287, + 0.63, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.391, + 0.613, + 0.405 + ], + "angle": 0, + "content": "pix2gestalt [28]" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.089, + 0.763, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.189, + 0.763, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.287, + 0.762, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.391, + 0.714, + 0.403 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.089, + 0.897, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.189, + 0.897, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.287, + 0.897, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.82, + 0.391, + 0.843, + 0.403 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.417, + 0.908, + 0.46 + ], + "angle": 0, + "content": "Figure 6. Qualitative Comparisons on the Validation Set of ADiW. Since Invisible Stitch (IS) [11] and Jo et al. [17] use metric depth estimation models [4], the depth maps are shifted compared with other methods that use relative depth estimators. Only officially released models are taken into account in this qualitative comparison." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.473, + 0.903, + 0.693 + ], + "angle": 0, + "content": "
MethodEasyMid.HardOverall
RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑
pix2gestalt‡[28]5.06790.3534.81890.2715.64184.0645.11488.717
Invisible Stitch‡[11]7.58483.3177.43581.3597.39380.0367.48181.781
Sekkat et al.†[38]11.19551.05211.40451.26711.15049.22211.26449.222
Sekkat et al.†‡[38]5.24588.7055.25488.7225.04487.4235.19488.367
Jo et al.†[17]10.26056.11810.22755.6379.98254.68710.17255.545
Jo et al.‡[17]7.55181.9336.63382.7336.48882.7936.93982.451
Jo et al.†‡[17]4.62489.8484.77789.9154.72887.2564.71289.177
Amodal-DAV2-S3.57493.6503.74193.6663.73492.1253.68293.251
Amodal-DAV2-B3.48293.8713.62094.0023.64092.6223.57893.590
Amodal-DAV2-L3.32494.4113.46094.5663.47693.3093.41894.179
Amodal-DepthFM5.35488.4275.39889.2415.49886.9705.41088.353
Amodal-DepthFM‡4.62292.4944.50092.9824.88391.0484.64592.295
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.703, + 0.907, + 0.76 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons on ADIW Dataset. We compare our methods against other possible solutions. Results from Invisible Stitch [11] and Jo et al. [17] are not directly comparable, as they use metric depth estimation models [4]. \\(\\dagger\\): Models retrained on our dataset for fair comparison. \\(\\ddagger\\): Scale-and-shift alignment applied for consistent prediction handling (see 3.3). Note that Amodal-DAV2 does not rely on this alignment approach. Best results are in bold." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.224, + 0.803 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.188, + 0.829 + ], + "angle": 0, + "content": "4.1. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "We follow the standard evaluation protocol proposed in previous monocular metric depth estimation works [2, 10] to evaluate the effectiveness of our framework. We use the root mean squared error (RMSE), the \\(\\log_{10}\\) error, and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.907, + 0.892 + ], + "angle": 0, + "content": "accuracy under the threshold \\(\\delta\\). Metrics consider the invisible parts of objects, with the RMSE scaled by a factor of 100 for better illustration. We also categorize objects based on their visible ratios into three difficulty levels: easy (0.75, 1], medium (0.5, 0.75], and hard (0, 0.5]. Metrics are calculated separately for each difficulty level to provide a comprehensive evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9678" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.477, + 0.196 + ], + "angle": 0, + "content": "
MethodRMSE↓log10↓δ(%)↑
①w/o D0, Ma7.5498.60770.607
②w/o Ma4.3694.32091.193
③w/ align3.8783.65992.037
④Ours (Full)3.6823.53893.251
⑤Lsilog for inv. only3.8453.75192.608
⑥Lssi + align4.0153.94891.852
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Table 2. Ablation Study for Amodal-DAV2. We investigate the impact of different guidance signals, supervision strategies, and inference techniques." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.274, + 0.308, + 0.29 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.483, + 0.538 + ], + "angle": 0, + "content": "We split our dataset into a training set of approximately 559K samples and a validation set of 4K samples. The Amodal-DAV2 model is trained with a batch size of 32, a learning rate of \\(1\\mathrm{e}^{-5}\\), and 50K iterations, using the scale-invariant log (silog) loss with \\(\\lambda = 0.85\\), following previous works [2, 23, 25]. The Amodal-DepthFM model is trained with a batch size of 128, a learning rate of \\(3\\mathrm{e}^{-5}\\), and 15K iterations. Both models are initialized with depth-pretrained parameters before fine-tuning for a modal depth estimation. We use the Adam optimizer with exponential learning rate decay. To stabilize training, we apply a max gradient norm clip of 0.01. Data augmentation is minimal, with only horizontal flipping. All experiments are conducted on 4 NVIDIA A100 GPUs. By default, we evaluate model performance using the final checkpoint after training. We use our Amodal-DAV2-L for all visualization results as default." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.547, + 0.229, + 0.561 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.569, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Comparison with Amodal Depth Methods. We compare our methods with existing amodal depth estimation approaches. As these methods were initially trained on synthetic datasets and relied on metric depth formulations, we retrained them on our dataset and adapted them to relative depth estimation. Results in Tab. 1 and Fig. 6 demonstrate a significant performance gap between the previous SoTA amodal depth model[17] and our methods, even after output alignment with our strategy. Notably, Amodal-DepthFM outperforms prior approaches even without the alignment, while our best model, Amodal-DAV2-L, improves performance by \\(27.4\\%\\) in terms of RMSE, setting a new SoTA. Previous methods, designed for amodal metric depth estimation, struggle to generalize across varied natural image depth ranges. In contrast, our approach leverages pretrained DAV2 [47] and DepthFM [14] models, which benefit from extensive prior knowledge of geometry and color, resulting in more accurate amodal depth predictions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparison with Other Solutions. We also evaluate two alternative solutions: Invisible Stitch [11] and pix2gestalt [28]. (1) For Invisible Stitch, we use SDXL [30] with a ground-truth amodal mask and image cap" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.09, + 0.898, + 0.193 + ], + "angle": 0, + "content": "
MethodRMSE↓log10↓δ(%)↑
①w/o align & Do, Ma10.21112.56356.271
②w/o align & Ma5.2836.23183.799
③w/o align5.4105.23588.353
④Ours (Full)4.6453.55392.295
⑤Lrmse for inv. only5.6363.91191.739
⑥Lrmse for scene4.6083.80991.149
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.203, + 0.906, + 0.273 + ], + "angle": 0, + "content": "Table 3. Ablation Study for Amodal-DepthFM. We explore the effectiveness of various guidance signals, supervision strategies, and the alignment inference for Amodal-DepthFM. Unlike Amodal-DAV2, the alignment strategy significantly improves the performance of Amodal-DepthFM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.905, + 0.424 + ], + "angle": 0, + "content": "tion to guide the inpainting. The inpainted image serves as guidance for the stitcher model [11] to perform depth inpainting, and the predicted depth is aligned with the observation depth map. (2) For pix2gestalt [28], we conduct amodal inpainting followed by depth estimation using DAV2 (ViT-G), aligning the predicted depth with the observation depth. Additional implementation details are in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.713 + ], + "angle": 0, + "content": "As illustrated in Fig. 6 and Fig. 7, the accuracy of amodal depth estimation relies heavily on the quality of RGB inpainting in both methods. Even with the ground-truth amodal mask, SD-XL struggles to accurately inpaint target areas, and the stretcher model fails to predict accurate depth for occluded parts, leaving ghosting artifacts in the predicted depth map. Similarly, pix2gestalt suffers cascade errors due to inaccurate inpainting with uncontrollable amodal shapes. Moreover, since DAV2 is trained on complete natural images, applying it directly to inpainted outputs from pix2gestalt (with single foreground objects on white backgrounds) leads to performance degradation. For example, the window boundary in the first case in Fig. 6 is missing from the predicted depth map. In contrast, our amodal depth models directly regress the depth of invisible parts without relying on RGB information, using only the amodal mask as guidance. These results demonstrate that our approach provides strong geometric priors, which could also serve as a useful condition for inpainting methods [51]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.796, + 0.74 + ], + "angle": 0, + "content": "4.4. Ablation Studies and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Guidance and Supervision Strategies. We conduct ablation experiments on Amodal-DAV2-S and Amodal-DepthFM to assess each framework component's contribution, with results in Tab. 2 and Tab. 3, respectively. For both models, guidance from the observation depth map \\(D_{o}\\) and amodal mask \\(M_{a}\\) is crucial for optimal performance (1, 2). While the primary goal and evaluation protocol focus is estimating depth for occluded parts of objects, both frameworks benefit from object-level supervision that includes visible regions (4, 5). Interestingly, the scale-and-shift align" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9679" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.21, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.174, + 0.209, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.257, + 0.21, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.345, + 0.174, + 0.359 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.092, + 0.324, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.174, + 0.324, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.257, + 0.325, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.344, + 0.307, + 0.358 + ], + "angle": 0, + "content": "SD-XL [30]" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.092, + 0.439, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.174, + 0.439, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.257, + 0.439, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.344, + 0.407, + 0.358 + ], + "angle": 0, + "content": "IS [11]" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.092, + 0.553, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.174, + 0.552, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.257, + 0.552, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.451, + 0.344, + 0.547, + 0.359 + ], + "angle": 0, + "content": "pix2gestalt [28]" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.092, + 0.667, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.174, + 0.667, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.257, + 0.667, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.344, + 0.662, + 0.359 + ], + "angle": 0, + "content": "pix2gestalt [28]" + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.092, + 0.781, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.174, + 0.781, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.674, + 0.257, + 0.781, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.345, + 0.744, + 0.358 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.092, + 0.896, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.174, + 0.896, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.257, + 0.896, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.345, + 0.853, + 0.358 + ], + "angle": 0, + "content": "\\(D_{o}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparisons on Images in the Wild. Invisible Stitch (IS) [11] uses SD-XL [30] with a ground-truth visible mask and image caption to inpaint occluded areas, while pix2gestalt [28] completes the occluded areas via amodal inpainting. Both methods suffer from inaccurate shape completion and cascade depth errors. The third row highlights the fragile shape completion of previous work: both SD-XL and pix2gestalt fail to remove the foreground cabinets correctly. In the Human Heuristics mode (Fig. 2), our method infers without any RGB inpainting or completion prior and achieves reasonable depth estimations for occluded areas." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.465, + 0.192, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.465, + 0.289, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.465, + 0.387, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.465, + 0.482, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Figure 8. Results with Different Mask Guidance. Guided by various amodal masks, our model successfully predicts the corresponding amodal depth maps for the target objects in the image, showcasing flexibility in estimating occluded regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.687 + ], + "angle": 0, + "content": "ment during inference improves Amodal-DepthFM performance but decreases that of Amodal-DAV2-S, indicating that inconsistent depth estimation may hinder Amodal-DepthFM's overall performance (③)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Varying Predictions with Different Guidance. Fig. 8 demonstrates the flexibility of our amodal depth estimation model in generating diverse, plausible predictions based on different amodal masks as guidance. This capability is crucial when multiple valid interpretations exist for occluded object parts. For example, the right-hand chair occluded by a wall could vary in width or even be connected to another occluded chair on the left. Our Amodal-DAV2-L adapts the amodal depth output according to the provided mask." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Edge Region v.s., Non-Edge Region. In this experiment, we extract edge areas from the ground-truth depth map and calculate metrics for both edge and non-edge regions in the invisible area. The overall RMSE from Amodal-DAV2-S is 3.682, with the RMSE for the edge region significantly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.557 + ], + "angle": 0, + "content": "higher at 7.378, compared to only 3.515 for non-edge regions. This discrepancy highlights the challenges in estimating amodal depth with fine-grained details. This finding aligns with human perception, where it is generally more difficult to accurately recover fine-grained geometry than the coarser structures for occluded areas." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.573, + 0.634, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We presented a novel approach to amodal depth estimation, focusing on predicting the depth of invisible parts of objects in natural scenes. Our work introduces Amodal Depth In the Wild (ADIW), a large-scale dataset that leverages segmentation datasets and a compositing pipeline, enabling high-quality, real-world amodal depth annotations. We proposed two complementary frameworks for amodal depth estimation: Amodal-DAV2 and Amodal-DepthFM. Amodal-DAV2 leverages the deterministic capabilities of Depth Anything V2 to achieve state-of-the-art performance in relative depth estimation, while A modal-DepthFM, built on a generative flow matching paradigm, and excels in providing finer details and sharper boundaries in occluded regions. Our experiments highlight that both models benefit from object-level supervision and the importance of guidance signals for improving amodal depth prediction accuracy. This work not only sets a new benchmark for amodal depth estimation but also opens the door for future research in occluded geometry understanding and improving applications such as inpainting and scene reconstruction." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9680" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.175 + ], + "angle": 0, + "content": "This publication is supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Jiayang Ao, Qiuhong Ke, and Krista A. Ehinger. Image amodal completion: A survey. In Computer Vision and Image Understanding. Elsevier, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[5] Amlaan Bhoi. Monocular depth estimation: A survey. arXiv preprint arXiv:1901.09402, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[6] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3.1 - a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[7] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[9] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, pages 6144-6153, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[10] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[11] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[13] Yongtao Ge, Guangkai Xu, Zhiyue Zhao, Libo Sun, Zheng Huang, Yanlong Sun, Hao Chen, and Chunhua Shen. Geobench: Benchmarking and analyzing monocular geometry estimation models. arXiv preprint arXiv:2406.12671, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[14] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[15] Cheng-Yen Hsieh, Tarasha Khurana, Achal Dave, and Deva Ramanan. Tracking any object amodally. arXiv preprint arXiv:2312.12433, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[16] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12518-12527, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[17] Seong-Uk Jo, Du Yeol Lee, and Chae Eun Rhee. Occlusion-aware amodal depth estimation for enhancing 3d reconstruction from a single image. IEEE Access, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[18] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Amodal completion and size constancy in natural scenes. In ICCV, pages 127-135, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, pages 9492-9502, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[20] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, pages 4019–4028, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[22] Mykola Lavreniuk, Shariq Farooq Bhat, Matthias Muller, and Peter Wonka. Evp: Enhanced visual perception using inverse multi-attentive feature refinement and regularized image-text alignment. In European Conference on Computer Vision Workshops (ECCVW), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[23] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[24] Zhenyu Li, Shariq Farooq Bhat, and Peter Wonka. Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, pages 1-18, 2023. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[26] Ruoshi Liu and Carl Vondrick. Humans as light bulbs: 3d human reconstruction from thermal reflection. In CVPR, pages 12531-12542, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[27] Alican Mertan, Damien Jade Duff, and Gozde Unal. Single image depth estimation: An overview. Digital Signal Processing, 123:103441, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[28] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "Amodal segmentation by synthesizing wholes. In CVPR, pages 3931-3940. IEEE Computer Society, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[29] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[31] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, pages 3014-3023, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[32] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, pages 3997-4008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[36] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[38] Ahmed Rida Sekkat, Rohit Mohan, Oliver Sawade, Elmar Matthes, and Abhinav Valada. Amodalsynthdrive: A synthetic amodal perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[39] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In CVPR, pages 8028-8038, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[40] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[41] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In CVPR, pages 9422-9432, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[44] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[45] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[46] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891, 2024. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[47] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, pages 28003-28013, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[50] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[52] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[53] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5729-5739, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[54] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In CVPR, pages 1464-1472, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9682" + } + ] +] \ No newline at end of file diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_origin.pdf b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8c5428130b5ce21ca7beb81f14e9844151df12da --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/dc30b08f-8ee9-4bca-9c02-871cd4377982_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06689a00c6fa17fb15ca874c98f1cd8767fe3f4fd608d92ce455c0fd94063cf6 +size 3065971 diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/full.md b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ea7ad59d2facf8df7ec27de244bfba4d901f347 --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/full.md @@ -0,0 +1,360 @@ +# Amodal Depth Anything: Amodal Depth Estimation in the Wild + +Zhenyu Li $^{1}$ , Mykola Lavreniuk $^{2}$ , Jian Shi $^{1}$ , Shariq Farooq Bhat $^{1}$ , Peter Wonka $^{1}$ $^{1}$ KAUST, $^{2}$ Space Research Institute NASU-SSAU + +![](images/802c0d29d73ea09201971555c48ab6dab09b47ad141358a0e80a5820f6bb4cb7.jpg) +Figure 1. Amodal Depth Estimation in the Wild. For each image, we present the amodal depth estimation with the target object outlined in black. Our model demonstrates strong generalization across diverse scenes, including both indoor and outdoor environments for accurate depth estimation for occluded parts of objects. Best viewed in color. + +# Abstract + +Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present + +two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions (Fig. 1). Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a $50.7\%$ improvement in RMSE over the previous SoTA on the ADIW dataset. + +# 1. Introduction + +Monocular depth estimation is a foundational task in computer vision and generative modeling [5, 10, 27] as it provides depth perception from a single image without stereo cues. However, while recent methods, e.g. [2, 4, 10, 23, 46], focus solely on estimating depth for visible pixels, humans can intuitively perceive the complete 3D + +![](images/dcacca2ba2921de9904d9143e0c51cd53f5b980dc24232d8a2689b550c5099d7.jpg) +Figure 2. Amodal Depth Estimation Pipeline. Given an input image, users can generate the amodal mask for the depth estimator in two ways: (1) Model Heuristics: click the target object, apply SAM [21] to generate modal mask, then use amodal segmentation methods to estimate amodal mask, (2) Human Heuristics: manually draw the amodal mask. Our model estimates amodal depth based on original observation image $I_{o}$ , the observed depth map $D_{o}$ , and the amodal mask $M_{a}$ . Note that we can get visible and invisible masks from both pipelines. Our model only focuses on the amodal depth (i.e., invisible part), whereas the visible part depth can be directly derived from the observed depth map $D_{o}$ . + +geometry of objects, even when only parts of them are visible. Amodal depth estimation is the task of predicting depth values for the occluded (invisible) parts of objects [17, 38]. It enables accurate 3D reconstruction and novel view synthesis from a single view, significantly reducing dependence on costly multi-view systems or LiDAR [1]. This capability is crucial for advancing AR/VR, robotics, autonomous driving and digital twin technologies [1]. In this paper, we aim to address this under-explored question: Can models effectively perceive the geometry of the invisible parts of objects in a scene? + +Unlike inpainting methods [11, 39, 50], which reconstruct missing image regions, amodal depth estimation is a novel task that extends traditional amodal segmentation [28, 48, 49] into the depth domain, by predicting the depth of occluded object parts. In this task, given an input image with a target amodal mask, the objective is to infer depth values for the occluded object regions. While amodal segmentation benefits from human annotations to collect training samples [17, 38], there is currently no device capable of collecting ground-truth depth data for occluded parts of objects at scale in real-world scenes. + +To address this challenge, prior approaches to amodal depth estimation have relied on synthetic datasets [17, 38]. However, generating these datasets is both time-consuming and hard to scale, often requiring manually placing occluders one by one [17, 38]. Additionally, synthetic data lacks the complexity and diversity of real-world scenes, resulting in a domain gap that limits the generalization of models trained on such data. Furthermore, these previous approaches only consider metric depth, where the goal is to estimate real-world distances for occluded regions. Re + +liaence on metric depth, which naturally struggles to generalize to unseen scenarios with limited data [7, 29], exacerbates the models' poor zero-shot performance on real-world images [6, 34, 46]. + +To overcome these limitations, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth. Recent advancements in depth estimation models, such as Depth Anything [46], have enabled the generation of high-quality relative depth maps from natural images. Focusing on relative depth allows us to train models using real-world data, leveraging the depth relationships within scenes to achieve better generalization without relying on precise metric measurements. + +We introduce a new large-scale real-world dataset, Amodal Depth In the Wild (ADIW), to facilitate the training of models for amodal depth estimation. Our data generation pipeline leverages segmentation datasets to create high-quality datasets for amodal relative depth estimation with scalable annotations. We adopt a compositing approach to create training pairs, where objects are sampled and composited into scenes. Corresponding depth maps are estimated using the large depth foundation model [47]. Given the challenges posed by occlusion in relative depth estimation, we employ a scale-and-shift alignment technique to refine and blend the depth values, ensuring consistency in the final ground-truth depth map used for training. + +In this work, we present two complementary frameworks for amodal relative depth estimation, targeting both deterministic and generative model classes. For deterministic models, we propose Amodal-DAV2, an extension of the Depth Anything V2 (DAV2) model [47]. This framework incorporates additional guidance layers to reason about depth in occluded regions. For generative models, we present Amodal-DepthFM, an adaptation of the DepthFM model [14], which uses a conditional flow matching approach to estimate depth in occluded regions. In both frameworks, we leverage the power of large pre-trained models with minimal modifications to effectively guide them toward predicting depth for occluded parts of objects. + +Experimental results indicate that our method achieves SoTA amodal depth estimation on the ADIW dataset and demonstrates strong zero-shot performance on real-world images (Fig. 1). Comprehensive ablation studies further validate the effectiveness of our design choices, showing that object-level supervision and guidance signals significantly enhance prediction accuracy. Moreover, we demonstrate the flexibility of our models in generating diverse plausible structures for occluded regions, illustrating their adaptability to different occlusion scenarios. In summary, our contributions are: + +- A novel task formulation of amodal depth estimation focusing on relative depth, enabling improved generalization capabilities compared to previous metric-based + +amodal depth estimation. + +- ADIW, a large-scale real-world dataset, generated from real-world images using a scalable pipeline with segmentation datasets and compositing techniques, coupled with a scale-and-shift alignment strategy. +- Two novel frameworks for amodal depth estimation, Amodal-DAV2 and Amodal-DepthFM, leveraging largescale pre-trained models with minimal modifications to achieve high-quality amodal depth predictions. + +# 2. Related Work + +# 2.1. Monocular Depth Estimation + +Monocular depth estimation, which predicts depth from a single image, has made substantial progress in recent years [2, 4, 10, 19, 24, 46]. Early methods focused on domain-specific metric depth estimation, assuming that training and test images share similar characteristics [3, 10, 12, 23, 25]. However, this in-domain focus poses challenges for generalization to unseen domains. To address this, recent research has shifted toward cross-domain relative depth estimation, where models infer the relative depth relationships between pixels [4, 6, 34, 46, 47]. For instance, MiDaS [34] aggregates multiple datasets and trains the model in a relative depth setting, achieving superior zero-shot performance. ZoeDepth [4] demonstrates that a strong relative depth model can effectively generalize to metric depth estimation through fine-tuning. Depth Anything [46] adopts a semi-supervised strategy that combines large-scale labeled and unlabeled data to further boost the model performance. Generative approaches have also emerged, with methods like Marigold [19, 22, 53] repurposing denoising UNet models [37] alongside fixed VAE encoders [42] for depth estimation [19], and DepthFM integrates flow matching principles into the depth estimation pipeline [14]. Despite advancements in monocular depth estimation, a major challenge remains: depth estimation in occluded regions. Existing methods focus only on visible areas, leaving unseen parts unaddressed. Our work addresses this gap by enabling direct depth estimation for invisible object regions. + +# 2.2. Amodal Perception + +Prior work in amodal perception has primarily focused on tasks such as inpainting [9, 28, 49], segmentation [20, 26, 31, 54], and detection [15, 18]. Recently, amodal depth estimation has emerged as a new challenge in amodal perception [17, 38], which aims to predict the depth of occluded regions of objects. For instance, [38] proposed the Amodal-SynthDrive dataset for synthetic driving scenes and adopted VIP-DeepLab [32] with multi-headed outputs to estimate the depth at different occlusion levels. Similarly, [17] introduced the indoor-focused Amodal-3D-FRONT dataset and an iterative approach leveraging amodal masks to estimate + +depth in occluded areas. These works rely heavily on synthetic datasets due to the difficulty of capturing ground-truth depth behind occlusions in real-world settings. This reliance on synthetic data limits model generalization to real-world scenes, and their focus on metric depth increases the difficulty of accurate depth prediction in the zero-shot setting. In contrast, our approach adopts a relative depth-based solution using large-scale pre-trained models and realistic synthetically generated datasets, enabling strong generalization to diverse real-world images. + +# 2.3. Depth Inpainting + +Depth inpainting, or depth completion, traditionally focuses on filling in missing depth values for visible regions based on sparse depth inputs [41, 43-45, 52]. Some methods aim to inpaint depth around occlusions to create novel 3D views [11, 16, 39], this differs from amodal depth estimation, which specifically targets object-level occlusions. For example, 3D Photography [39] uses shared edge guidance to combine RGB and depth inpainting, while SLIDE [16] introduces a soft-layering strategy to preserve visual details in novel views. Recently, diffusion models have shown strong image generation capabilities [37]. Invisible Stitch [11] uses a Stable Diffusion model variant [37] to fill in missing image regions, generating novel views [30], followed by a depth painting network to fill holes in the depth map with guidance from the novel-view image. Unlike these methods, which implicitly or explicitly rely on color priors and focus on visual quality, our approach directly estimates occluded region depth, prioritizing depth accuracy independently of color information or visual coherence. This distinction shifts the focus from achieving realistic view synthesis to predicting occluded depths with high geometric accuracy, an approach that enhances scene understanding for real-world applications. + +# 3. Method + +# 3.1. Task Definition + +Amodal depth estimation extends the concept of amodal segmentation by predicting depth information for occluded areas. In traditional amodal segmentation [28, 48, 49], given an input image and a segmentation mask of the visible portion of an object, the goal is to predict the complete object mask, including the occluded (invisible) part. Similarly, amodal depth estimation aims to estimate the depth values for the occluded regions, given an input observation image $I_{o}$ , a corresponding observation depth map $D_{o}$ of the input image, and a target amodal segmentation mask $M_{a}$ . + +Different from earlier works that formulate amodal depth estimation as a metric depth estimation task [17, 38], we propose a novel formulation where the goal is to predict the relative depth of the occluded parts of the object for + +![](images/62511e5e64361c171db9de542dd0681acf898dc4d3dbfcc758080d516d4821af.jpg) +Figure 3. Constructing Training Data. We use the method from [28] to convert an initial segmentation dataset into a whole-object dataset. Next, we sample and composite images to create training pairs. Due to occluders, the relative depth predictions differ between the composite and background images, so we apply scale-and-shift alignment for consistent depth blending. + +the input image in the wild. Recent models, such as Depth Anything [46, 47] generate high-quality relative depth maps from natural images, allowing us to train amodal depth models using real-world data as shown in Sec. 3.2. + +# 3.2. Dataset Collection + +We present Amodal Depth In the Wild (ADIW), a large-scale dataset specifically designed for predicting occluded object relative depths in real-world scenes. Creating a natural image dataset at scale for amodal depth estimation is challenging due to the lack of annotations for hidden regions. No existing device can capture ground-truth depth data for occluded parts at scale in real-world settings. Previous efforts have generally relied on synthetic datasets [17, 38], which, although valuable, fall short in capturing the complexity and diversity of real-world scenes. Additionally, these approaches require manually placing occluders, limiting scalability. ADIW overcomes these by generating training data from real-world images. + +We create paired data by overlaying objects over natural images following [28], as illustrated in Fig. 3. Our method leverages segmentation datasets, which are more scalable than traditional depth estimation datasets [21, 35, 46]. Specifically, we employ the Segment Anything model [21] to automatically generate segmentation masks from the SA-1B dataset, forming the initial segmentation dataset. Next, we apply the heuristic algorithm from [28] to filter out incomplete objects, creating a whole-object dataset. + +In our approach, each image $I_{b}$ (the background image) contains at least one complete object. We then sample an occluder object $I_{f}$ (the foreground image) and superimpose it onto $I_{b}$ , forming the assembled observation image $I_{o}$ . Both $I_{o}$ and $I_{b}$ are then processed through the Depth Anything V2 model [47] (ViT-G) to obtain relative depth maps $D_{o}$ and $D_{b}$ , respectively. Importantly, the depth val + +ues for the background object in $D_{o}$ and $D_{b}$ differ due to the presence of the foreground object, which changes perceived depth. Both depth maps are scaled to the range [0, 1]. + +To ensure consistent training labels, we apply the scale-and-shift alignment algorithm [34] to align the background object's depth values across the two depth maps. The scale factor $s$ and shift factor $t$ are computed as: + +$$ +(s, t) = \operatorname {a r g m i n} _ {s, t} \sum_ {i = 1} ^ {N} \left(s d _ {i} ^ {b} + t - d _ {i} ^ {o}\right) ^ {2}, i \in M _ {\text {v i s}}, \tag {1} +$$ + +where $d_i^o$ and $d_i^b$ are the depth values of pixels in the visible part of the background object in $D_o$ and $D_b$ , respectively, and $N$ denotes the total number of valid pixels in the visible mask $M_{vis}$ of the background object. The aligned depth map $D_{aligned}$ is then calculated as: + +$$ +D _ {\text {a l i g n e d}} = s D _ {b} + t, \tag {2} +$$ + +serving as the ground-truth map for model training. This procedure generates a dataset of 564K images with amodal depth labels. + +As shown by [46], relative depth estimators [6, 34] generalize better than metric depth estimators [2, 23, 25]. Following this insight, we produce relative depth maps for amodal depth estimation in natural scenes. Moreover, our data generation approach is also adaptable for metric amodal depth estimation by omitting the scale-and-shift alignment and utilizing metric depth models [7, 29]. + +# 3.3. Amodal Depth Estimator + +In this work, we aim to leverage large pre-trained depth models by fine-tuning them specifically for amodal depth estimation. Our goal is to make minimal modifications to the original network architectures, preserving the capabilities of the pre-trained weights while introducing the necessary adjustments to guide the model for amodal depth prediction. By integrating additional guidance into the networks, we enable them to predict depth values for occluded parts of objects. + +Depth estimation models fall into two types: deterministic and generative models [13]. In this section, we introduce a dedicated strategy for adapting each model class to the amodal depth estimation task, as illustrated in Fig. 4 and Fig. 5. For the deterministic model, we adopt the Depth Anything V2 (DAV2) [47], a highly representative and top-performing pre-trained model. For the generative model, we select DepthFM [14], known for its superior ability to capture depth details and fast inference speed. + +# 3.3.1.Amodal-DAV2 + +Model Structure. The Amodal-DAV2 framework introduces subtle structural modifications to the Depth Anything V2 (DAV2) model's image encoder to retain its pre-trained + +![](images/1903eaed9dfb9a773c0be23d931b6240bc082cba83b75120ac537b8c27c5b333.jpg) +Figure 4. Amodal-DAV2 Framework Structure. Amodal-DAV2 modifies the DAV2 image encoder to take additional guidance channels along with RGB. + +knowledge while enhancing its capacity in amodal depth estimation. The original DAV2 model employs a Vision Transformer (ViT) architecture [8], which begins with an RGB convolution applied to the input image. As shown in Fig. 4, we add a Guidance Convolution (Conv) [40] layer in parallel with the RGB Conv layer, enabling the encoder to accept additional guidance channels for the observation depth map $D_{o}$ and amodal mask $M_{a}$ . To ensure compatibility with the pre-trained model, the kernel weights of the Guidance Conv layer are initialized to zero, allowing the model to initially ignore the additional guidance information. Additionally, we incorporate layer normalization into the input features of the DPT head [33] to stabilize training and improve convergence. + +Training method. We fine-tune the entire Amodal-DAV2 framework during training, utilizing the standard scale-invariant loss $\mathcal{L}_{si}$ [10] as our objective function: + +$$ +\mathcal {L} _ {s i} = \alpha \sqrt {\frac {1}{N} \sum_ {i \in M _ {a}} g _ {i} ^ {2} - \frac {\lambda}{N ^ {2}} (\sum_ {i \in M _ {a}} g _ {i}) ^ {2}}, +$$ + +where $g_{i} = \log \tilde{d}_{i} - \log d_{i}$ , with $\tilde{d}_i$ and $d_{i}$ representing the predicted depth and the ground truth depth, respectively. $N$ denotes the number of valid pixels on the amodal mask $M_{a}$ . Although our primary focus is on the accuracy of depth predictions for occluded (invisible) parts of the objects, we supervise the model using the entire object's depth. This holistic approach helps the model better understand the overall scene structure, leading to improved performance. + +# 3.3.2. Amodal-DepthFM + +Framework and Training. We adapt DepthFM [14] for amodal depth estimation using a similar approach. Given pairs of an image $I$ , its observation depth map $D_{o}$ , and corresponding amodal mask $M_{o}$ , we fine-tune a conditional flow matching model to achieve amodal depth estimation. The objective is defined as: + +$$ +\min _ {\theta} \mathbb {E} _ {t, z, p (x _ {0})} | | v _ {\theta} (t, \phi_ {t} (x _ {0})) - (x _ {1} - x _ {0}) | |, +$$ + +where $x_{1}$ represents the encoded depth samples in the latent space, and the starting point $x_{0}$ corresponds to an encoded representation of the input image. The latent flow + +![](images/0a2b84a36ae8f7e333c124ea9a87255d38fc9f50ec911b06cc8ab8798af07603.jpg) +Figure 5. Amodal-DepthFM Framework Structure. Amodal-DepthFM modifies the DepthFM denoising UNet encoder to take additional guidance channels along with RGB latent code. + +is conditioned on the guidance provided by $z$ , which includes the input image latent code, observation depth map $D_{o}$ , and amodal mask $M_{a}$ . Here, $t$ denotes the timestamp, and $\phi_t(x|z) = tx_1 + (1 - t)x_0$ . + +During training, we apply noise augmentation, introducing Gaussian noise $\mathcal{N}(x,\sigma_{min})$ to the data samples. Consequently, the conditional probability path is modeled as $p_t(x|z) = \mathcal{N}(x|tx_1 + (1 - t)x_0,\sigma_{\mathrm{min}}^2\mathbf{I})$ . For additional details, we refer readers to [14]. + +In the U-Net architecture used for the flow network $v_{\theta}$ , we modify the first input Conv layer to accommodate the additional guidance channels. Specifically, we extend the channel dimensions to accept extra guidance information $D_{o}$ and $M_{a}$ alongside the image latent code. To effectively leverage the pre-trained model, we initialize the first Conv layer with a combination of pre-trained weights for the primary eight channels (corresponding to the depth and image latent codes) and zero-initialized weights for the two additional channels dedicated to guidance inputs $D_{o}$ and $M_{a}$ . + +Scale-and-Shift Alignment Inference. While the Amodal-DAV2 model directly regresses the amodal depth map, the Amodal-DepthFM model operates in the latent space, predicting the amodal depth latent code, which is then decoded into a depth map using a latent decoder. However, learning consistent amodal depth can be challenging for A modal-DepthFM. To enhance the predicted depth map, we employ the scale-and-shift alignment [34, 36] during inference. + +Rather than relying solely on the model's output, we enhance the prediction by blending it with the observed depth map using a scale-and-shift alignment over the shared visible regions, as described in Eq. 1 and Eq. 2. This technique leverages the information from shared regions between the observation depth map and the amodal prediction. By calculating optimal scale and shift factors, we align the depth values of the occluded regions with the visible areas, thereby enhancing the coherence and consistency of the final depth map. + +![](images/27b6df8bd02d5b590a43cb04a5af0f9553720936bcf9722cb8901da98be82ea8.jpg) + +![](images/0fa99f835a456786b8a28dc0b5ab4f9c3823547461050df2a97a1e995ef5a379.jpg) + +![](images/442229eb04f48ea0acaa2904722fb5551f6a4d596ed9f4382b04eee6e0f222f2.jpg) +Input + +![](images/3286d24224a3cb55896fbbc6dd93a7c1f420f3234acfdb7c1aa8f5ec2c6f69d1.jpg) + +![](images/7bf66e680d4469415237bb7bf23718a6034cdd5487ecca3948857bd771ca0a2f.jpg) + +![](images/08bcbf48136ef3f613985fcb4ae985123bea05428dec724ed5e6cb38f05f980e.jpg) +IS [11] + +![](images/95c903c0c91e6b1e719c4a82f2038ae328e9fae73ef7ab639060d8ab11b591e2.jpg) + +![](images/12dfe083ac19ec224365c40ec6e5c1b79e3add275615f59a885508bd36be4ce9.jpg) + +![](images/86385e5711cae45d605950561ae65d8d24ae4f787d813f0f832c0747b37a9d7d.jpg) +Jo et al. [17] + +![](images/85ad5f0bc96ad19a4ef1d7c231c749146fbc2758369a62d5777d64ea2a2c79ce.jpg) + +![](images/a97049af47e9c64e4a47591c6b950256ee388b319cd469fb04116be8da9b9dc1.jpg) + +![](images/248f706b255cf266f610749425b798051519d230381c64efc424eec691c3e5ec.jpg) +pix2gestalt [28] +Figure 6. Qualitative Comparisons on the Validation Set of ADiW. Since Invisible Stitch (IS) [11] and Jo et al. [17] use metric depth estimation models [4], the depth maps are shifted compared with other methods that use relative depth estimators. Only officially released models are taken into account in this qualitative comparison. + +![](images/eafc825da420f810ebacccecb3955628ed9a238d47621cd5eb3099a2b9ef93b4.jpg) + +![](images/77209a82aa9c81a1ca5bc4c265e39a62062f17eae9ef77b64b86616161a7f8be.jpg) + +![](images/5c9d5d71c10befaf9e0c6d4cb2121149e8fca189227964809a1b32eab84d842d.jpg) +Ours + +![](images/8c2c51cae0a7eb39fa31601768fa7fe3d274e26bfd1a9ad481c0544ce59ce94b.jpg) + +![](images/8f6a85dda5fcaf639f232ff13b6bc5095a5a7d59318dbd9b049485d1bd6cc17b.jpg) + +![](images/fc733a17c8ee8adba2925a9bf83eb715f5949934ab604866357147871914840b.jpg) +GT + +
MethodEasyMid.HardOverall
RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑
pix2gestalt‡[28]5.06790.3534.81890.2715.64184.0645.11488.717
Invisible Stitch‡[11]7.58483.3177.43581.3597.39380.0367.48181.781
Sekkat et al.†[38]11.19551.05211.40451.26711.15049.22211.26449.222
Sekkat et al.†‡[38]5.24588.7055.25488.7225.04487.4235.19488.367
Jo et al.†[17]10.26056.11810.22755.6379.98254.68710.17255.545
Jo et al.‡[17]7.55181.9336.63382.7336.48882.7936.93982.451
Jo et al.†‡[17]4.62489.8484.77789.9154.72887.2564.71289.177
Amodal-DAV2-S3.57493.6503.74193.6663.73492.1253.68293.251
Amodal-DAV2-B3.48293.8713.62094.0023.64092.6223.57893.590
Amodal-DAV2-L3.32494.4113.46094.5663.47693.3093.41894.179
Amodal-DepthFM5.35488.4275.39889.2415.49886.9705.41088.353
Amodal-DepthFM‡4.62292.4944.50092.9824.88391.0484.64592.295
+ +Table 1. Quantitative Comparisons on ADIW Dataset. We compare our methods against other possible solutions. Results from Invisible Stitch [11] and Jo et al. [17] are not directly comparable, as they use metric depth estimation models [4]. $\dagger$ : Models retrained on our dataset for fair comparison. $\ddagger$ : Scale-and-shift alignment applied for consistent prediction handling (see 3.3). Note that Amodal-DAV2 does not rely on this alignment approach. Best results are in bold. + +# 4. Experiments + +# 4.1. Metrics + +We follow the standard evaluation protocol proposed in previous monocular metric depth estimation works [2, 10] to evaluate the effectiveness of our framework. We use the root mean squared error (RMSE), the $\log_{10}$ error, and the + +accuracy under the threshold $\delta$ . Metrics consider the invisible parts of objects, with the RMSE scaled by a factor of 100 for better illustration. We also categorize objects based on their visible ratios into three difficulty levels: easy (0.75, 1], medium (0.5, 0.75], and hard (0, 0.5]. Metrics are calculated separately for each difficulty level to provide a comprehensive evaluation. + +
MethodRMSE↓log10↓δ(%)↑
①w/o D0, Ma7.5498.60770.607
②w/o Ma4.3694.32091.193
③w/ align3.8783.65992.037
④Ours (Full)3.6823.53893.251
⑤Lsilog for inv. only3.8453.75192.608
⑥Lssi + align4.0153.94891.852
+ +# 4.2. Implementation Details + +We split our dataset into a training set of approximately 559K samples and a validation set of 4K samples. The Amodal-DAV2 model is trained with a batch size of 32, a learning rate of $1\mathrm{e}^{-5}$ , and 50K iterations, using the scale-invariant log (silog) loss with $\lambda = 0.85$ , following previous works [2, 23, 25]. The Amodal-DepthFM model is trained with a batch size of 128, a learning rate of $3\mathrm{e}^{-5}$ , and 15K iterations. Both models are initialized with depth-pretrained parameters before fine-tuning for a modal depth estimation. We use the Adam optimizer with exponential learning rate decay. To stabilize training, we apply a max gradient norm clip of 0.01. Data augmentation is minimal, with only horizontal flipping. All experiments are conducted on 4 NVIDIA A100 GPUs. By default, we evaluate model performance using the final checkpoint after training. We use our Amodal-DAV2-L for all visualization results as default. + +# 4.3. Main Results + +Comparison with Amodal Depth Methods. We compare our methods with existing amodal depth estimation approaches. As these methods were initially trained on synthetic datasets and relied on metric depth formulations, we retrained them on our dataset and adapted them to relative depth estimation. Results in Tab. 1 and Fig. 6 demonstrate a significant performance gap between the previous SoTA amodal depth model[17] and our methods, even after output alignment with our strategy. Notably, Amodal-DepthFM outperforms prior approaches even without the alignment, while our best model, Amodal-DAV2-L, improves performance by $27.4\%$ in terms of RMSE, setting a new SoTA. Previous methods, designed for amodal metric depth estimation, struggle to generalize across varied natural image depth ranges. In contrast, our approach leverages pretrained DAV2 [47] and DepthFM [14] models, which benefit from extensive prior knowledge of geometry and color, resulting in more accurate amodal depth predictions. + +Comparison with Other Solutions. We also evaluate two alternative solutions: Invisible Stitch [11] and pix2gestalt [28]. (1) For Invisible Stitch, we use SDXL [30] with a ground-truth amodal mask and image cap + +Table 2. Ablation Study for Amodal-DAV2. We investigate the impact of different guidance signals, supervision strategies, and inference techniques. + +
MethodRMSE↓log10↓δ(%)↑
①w/o align & Do, Ma10.21112.56356.271
②w/o align & Ma5.2836.23183.799
③w/o align5.4105.23588.353
④Ours (Full)4.6453.55392.295
⑤Lrmse for inv. only5.6363.91191.739
⑥Lrmse for scene4.6083.80991.149
+ +Table 3. Ablation Study for Amodal-DepthFM. We explore the effectiveness of various guidance signals, supervision strategies, and the alignment inference for Amodal-DepthFM. Unlike Amodal-DAV2, the alignment strategy significantly improves the performance of Amodal-DepthFM. + +tion to guide the inpainting. The inpainted image serves as guidance for the stitcher model [11] to perform depth inpainting, and the predicted depth is aligned with the observation depth map. (2) For pix2gestalt [28], we conduct amodal inpainting followed by depth estimation using DAV2 (ViT-G), aligning the predicted depth with the observation depth. Additional implementation details are in the supplementary materials. + +As illustrated in Fig. 6 and Fig. 7, the accuracy of amodal depth estimation relies heavily on the quality of RGB inpainting in both methods. Even with the ground-truth amodal mask, SD-XL struggles to accurately inpaint target areas, and the stretcher model fails to predict accurate depth for occluded parts, leaving ghosting artifacts in the predicted depth map. Similarly, pix2gestalt suffers cascade errors due to inaccurate inpainting with uncontrollable amodal shapes. Moreover, since DAV2 is trained on complete natural images, applying it directly to inpainted outputs from pix2gestalt (with single foreground objects on white backgrounds) leads to performance degradation. For example, the window boundary in the first case in Fig. 6 is missing from the predicted depth map. In contrast, our amodal depth models directly regress the depth of invisible parts without relying on RGB information, using only the amodal mask as guidance. These results demonstrate that our approach provides strong geometric priors, which could also serve as a useful condition for inpainting methods [51]. + +# 4.4. Ablation Studies and Discussion + +Guidance and Supervision Strategies. We conduct ablation experiments on Amodal-DAV2-S and Amodal-DepthFM to assess each framework component's contribution, with results in Tab. 2 and Tab. 3, respectively. For both models, guidance from the observation depth map $D_{o}$ and amodal mask $M_{a}$ is crucial for optimal performance (1, 2). While the primary goal and evaluation protocol focus is estimating depth for occluded parts of objects, both frameworks benefit from object-level supervision that includes visible regions (4, 5). Interestingly, the scale-and-shift align + +![](images/7fe3af7970e7b7418b6f075b596e070705a2f1d952bb392f4284f12de286de9c.jpg) + +![](images/2bd26d3e13cbed7dff88f3938f31074d92338017808152b9bba10252b2446d12.jpg) + +![](images/b7b19e74f08faaf30d2d5f2b486a7eb7fb559a2328847ba426471508099ffd9b.jpg) +Input + +![](images/0e5a45974ccadbf09823dd72d819a981d9332a6f2bd352cb176e64e5bd7641f9.jpg) + +![](images/0c4c9f6185e8491b4193ead3acc74fdae37351f7c9f9099af64638b921471d86.jpg) + +![](images/05b2221ce96e24779eac7e310790e92b738afc49b920ce75c660e85412a16b7f.jpg) +SD-XL [30] + +![](images/1f15514eafe16a9afcaf31c919c76ed90c2c2b25a29e4bed393d07ecf1f955e5.jpg) + +![](images/ca773620b14d2ca46110baa10c722a002f6775bfccd9a7f0f56afb7f56de8750.jpg) + +![](images/51e09fbf810bcd1eff2b47293744fa1ae41843f6de731dc9c6ffdeb495e1746e.jpg) +IS [11] + +![](images/97f7e436e5a43da0e90912a6f88b6e8a4271b33f4334165c1878db4937b5f39c.jpg) + +![](images/5a9c3a8bc3e96d445e2a4fc0ce2e6b600ba83239076a401adb26074cf05f6ca6.jpg) + +![](images/83f256ccdd6dfb5f8d7652cac1f6564fb6e4ec98571752e01b1b6a7f3a6ad0ef.jpg) +pix2gestalt [28] + +![](images/683a0f9b3a41e48f33d2409f4584a48b4c6ac5d745775aef971b23726578eedd.jpg) + +![](images/7b9e32f1a5c685e44ae78cd92909188cb4c7b01a0033a19e4c91aa13a734fb70.jpg) + +![](images/d54c8718535dcdff9abd4bbbf84d85ba21e90713865277b2145f7495e69e0832.jpg) +pix2gestalt [28] + +![](images/d26a5b16a6f7fba7c337730e538bfa509cb15d1f92d64d3c8d9ffa73915ebfc5.jpg) + +![](images/2278828038f0232a7459cce193c55ec759924a39b53bb78bca47b6f0fc1cbac4.jpg) + +![](images/9d852c171bd387eecba0ffa12672ec666abfbec47babea4495bbcd79b093026b.jpg) +Ours + +![](images/82808fb121eaf921de0fb957cb8a11f571f9fe51f7302519100cc3d08a872cec.jpg) + +![](images/3eba01883d4649c04f9433c9ca77c06ee275f667dc2e1252822da2a354eebb93.jpg) + +![](images/4e2a0e71bf0d5c416a0bb1ca7f86c8dff73e53778232b38a12aea94cb6c9cd6f.jpg) +$D_{o}$ + +![](images/36fc49f35d9c6fb1880fce4f30b0951b185a8b42385785201dfeb6b2e11651ad.jpg) +Figure 7. Qualitative Comparisons on Images in the Wild. Invisible Stitch (IS) [11] uses SD-XL [30] with a ground-truth visible mask and image caption to inpaint occluded areas, while pix2gestalt [28] completes the occluded areas via amodal inpainting. Both methods suffer from inaccurate shape completion and cascade depth errors. The third row highlights the fragile shape completion of previous work: both SD-XL and pix2gestalt fail to remove the foreground cabinets correctly. In the Human Heuristics mode (Fig. 2), our method infers without any RGB inpainting or completion prior and achieves reasonable depth estimations for occluded areas. + +![](images/56a121a7613b4099c1e03fe992fe356ae89ab9c38410e0f1ea75345fe3c95f66.jpg) +Figure 8. Results with Different Mask Guidance. Guided by various amodal masks, our model successfully predicts the corresponding amodal depth maps for the target objects in the image, showcasing flexibility in estimating occluded regions. + +![](images/04b428f399dc37639a109ec5826d6aabe374656eec0b061cbd11a28da6e1bee6.jpg) + +![](images/dbe1f1eb3ffe267a4955a01c453509d7008fbe38aceb101ced48b1f346e78ba0.jpg) + +ment during inference improves Amodal-DepthFM performance but decreases that of Amodal-DAV2-S, indicating that inconsistent depth estimation may hinder Amodal-DepthFM's overall performance (③). + +Varying Predictions with Different Guidance. Fig. 8 demonstrates the flexibility of our amodal depth estimation model in generating diverse, plausible predictions based on different amodal masks as guidance. This capability is crucial when multiple valid interpretations exist for occluded object parts. For example, the right-hand chair occluded by a wall could vary in width or even be connected to another occluded chair on the left. Our Amodal-DAV2-L adapts the amodal depth output according to the provided mask. + +Edge Region v.s., Non-Edge Region. In this experiment, we extract edge areas from the ground-truth depth map and calculate metrics for both edge and non-edge regions in the invisible area. The overall RMSE from Amodal-DAV2-S is 3.682, with the RMSE for the edge region significantly + +higher at 7.378, compared to only 3.515 for non-edge regions. This discrepancy highlights the challenges in estimating amodal depth with fine-grained details. This finding aligns with human perception, where it is generally more difficult to accurately recover fine-grained geometry than the coarser structures for occluded areas. + +# 5. Conclusion + +We presented a novel approach to amodal depth estimation, focusing on predicting the depth of invisible parts of objects in natural scenes. Our work introduces Amodal Depth In the Wild (ADIW), a large-scale dataset that leverages segmentation datasets and a compositing pipeline, enabling high-quality, real-world amodal depth annotations. We proposed two complementary frameworks for amodal depth estimation: Amodal-DAV2 and Amodal-DepthFM. Amodal-DAV2 leverages the deterministic capabilities of Depth Anything V2 to achieve state-of-the-art performance in relative depth estimation, while A modal-DepthFM, built on a generative flow matching paradigm, and excels in providing finer details and sharper boundaries in occluded regions. Our experiments highlight that both models benefit from object-level supervision and the importance of guidance signals for improving amodal depth prediction accuracy. This work not only sets a new benchmark for amodal depth estimation but also opens the door for future research in occluded geometry understanding and improving applications such as inpainting and scene reconstruction. + +# Acknowledgements + +This publication is supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. + +# References + +[1] Jiayang Ao, Qiuhong Ke, and Krista A. Ehinger. Image amodal completion: A survey. In Computer Vision and Image Understanding. Elsevier, 2023. 2 +[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1, 3, 4, 6, 7 +[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3 +[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 3, 6 +[5] Amlaan Bhoi. Monocular depth estimation: A survey. arXiv preprint arXiv:1901.09402, 2019. 1 +[6] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3.1 - a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 2, 3, 4 +[7] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 4 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[9] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, pages 6144-6153, 2018. 3 +[10] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 1, 3, 5, 6 +[11] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 2, 3, 6, 7, 8 +[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 3 +[13] Yongtao Ge, Guangkai Xu, Zhiyue Zhao, Libo Sun, Zheng Huang, Yanlong Sun, Hao Chen, and Chunhua Shen. Geobench: Benchmarking and analyzing monocular geometry estimation models. arXiv preprint arXiv:2406.12671, 2024. 4 + +[14] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 2, 3, 4, 5, 7 +[15] Cheng-Yen Hsieh, Tarasha Khurana, Achal Dave, and Deva Ramanan. Tracking any object amodally. arXiv preprint arXiv:2312.12433, 2023. 3 +[16] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12518-12527, 2021. 3 +[17] Seong-Uk Jo, Du Yeol Lee, and Chae Eun Rhee. Occlusion-aware amodal depth estimation for enhancing 3d reconstruction from a single image. IEEE Access, 2024. 2, 3, 4, 6, 7 +[18] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Amodal completion and size constancy in natural scenes. In ICCV, pages 127-135, 2015. 3 +[19] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, pages 9492-9502, 2024. 3 +[20] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, pages 4019–4028, 2021. 3 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4 +[22] Mykola Lavreniuk, Shariq Farooq Bhat, Matthias Muller, and Peter Wonka. Evp: Enhanced visual perception using inverse multi-attentive feature refinement and regularized image-text alignment. In European Conference on Computer Vision Workshops (ECCVW), 2024. 3 +[23] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 1, 3, 4, 7 +[24] Zhenyu Li, Shariq Farooq Bhat, and Peter Wonka. Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284, 2023. 3 +[25] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, pages 1-18, 2023. 3, 4, 7 +[26] Ruoshi Liu and Carl Vondrick. Humans as light bulbs: 3d human reconstruction from thermal reflection. In CVPR, pages 12531-12542, 2023. 3 +[27] Alican Mertan, Damien Jade Duff, and Gozde Unal. Single image depth estimation: An overview. Digital Signal Processing, 123:103441, 2022. 1 +[28] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: + +Amodal segmentation by synthesizing wholes. In CVPR, pages 3931-3940. IEEE Computer Society, 2024. 2, 3, 4, 6, 7, 8 +[29] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116, 2024. 2, 4 +[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8 +[31] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, pages 3014-3023, 2019. 3 +[32] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, pages 3997-4008, 2021. 3 +[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 5 +[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2022. 2, 3, 4, 5 +[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4 +[36] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 5 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3 +[38] Ahmed Rida Sekkat, Rohit Mohan, Oliver Sawade, Elmar Matthes, and Abhinav Valada. Amodalsynthdrive: A synthetic amodal perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 2024. 2, 3, 4, 6 +[39] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In CVPR, pages 8028-8038, 2020. 2, 3 +[40] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 5 +[41] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 3 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3 +[43] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In CVPR, pages 9422-9432, 2023. 3 + +[44] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024. +[45] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 3 +[46] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891, 2024. 1, 2, 3, 4 +[47] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 2, 3, 4, 7 +[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, pages 28003-28013, 2024. 2, 3 +[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2, 3 +[50] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2 +[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 7 +[52] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 3 +[53] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5729-5739, 2023. 3 +[54] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In CVPR, pages 1464-1472, 2017. 3 \ No newline at end of file diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/images.zip b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08c3e9e3f7ddb2941df5ffdbd52e0e8afca319e5 --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4661fe2b624eea0446a411c4f73c1b46507a82031dbcb2a0019bd48aeaf4307 +size 658773 diff --git a/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/layout.json b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7e83614db7ab78fd56da86b3a78db696c706fa9e --- /dev/null +++ b/2025/Amodal Depth Anything_ Amodal Depth Estimation in the Wild/layout.json @@ -0,0 +1,9600 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 503, + 121 + ], + "type": "text", + "content": "Amodal Depth Anything: Amodal Depth Estimation in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": "Zhenyu Li" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": ", Mykola Lavreniuk" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": ", Jian Shi" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": ", Shariq Farooq Bhat" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": ", Peter Wonka" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": "KAUST, " + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 496, + 172 + ], + "type": "text", + "content": "Space Research Institute NASU-SSAU" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 198, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 57, + 198, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 57, + 198, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 553, + 411 + ], + "type": "image", + "image_path": "802c0d29d73ea09201971555c48ab6dab09b47ad141358a0e80a5820f6bb4cb7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 419, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 54, + 419, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 555, + 455 + ], + "type": "text", + "content": "Figure 1. Amodal Depth Estimation in the Wild. For each image, we present the amodal depth estimation with the target object outlined in black. Our model demonstrates strong generalization across diverse scenes, including both indoor and outdoor environments for accurate depth estimation for occluded parts of objects. Best viewed in color." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 464, + 201, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 464, + 201, + 477 + ], + "spans": [ + { + "bbox": [ + 151, + 464, + 201, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 714 + ], + "type": "text", + "content": "Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 464, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 596 + ], + "type": "text", + "content": "two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions (Fig. 1). Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a " + }, + { + "bbox": [ + 313, + 464, + 556, + 596 + ], + "type": "inline_equation", + "content": "50.7\\%" + }, + { + "bbox": [ + 313, + 464, + 556, + 596 + ], + "type": "text", + "content": " improvement in RMSE over the previous SoTA on the ADIW dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 620, + 395, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 395, + 633 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 395, + 633 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Monocular depth estimation is a foundational task in computer vision and generative modeling [5, 10, 27] as it provides depth perception from a single image without stereo cues. However, while recent methods, e.g. [2, 4, 10, 23, 46], focus solely on estimating depth for visible pixels, humans can intuitively perceive the complete 3D" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9673" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "type": "image", + "image_path": "dcacca2ba2921de9904d9143e0c51cd53f5b980dc24232d8a2689b550c5099d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "text", + "content": "Figure 2. Amodal Depth Estimation Pipeline. Given an input image, users can generate the amodal mask for the depth estimator in two ways: (1) Model Heuristics: click the target object, apply SAM [21] to generate modal mask, then use amodal segmentation methods to estimate amodal mask, (2) Human Heuristics: manually draw the amodal mask. Our model estimates amodal depth based on original observation image " + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "text", + "content": ", the observed depth map " + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "text", + "content": ", and the amodal mask " + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "text", + "content": ". Note that we can get visible and invisible masks from both pipelines. Our model only focuses on the amodal depth (i.e., invisible part), whereas the visible part depth can be directly derived from the observed depth map " + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 55, + 185, + 296, + 307 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 329, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 461 + ], + "type": "text", + "content": "geometry of objects, even when only parts of them are visible. Amodal depth estimation is the task of predicting depth values for the occluded (invisible) parts of objects [17, 38]. It enables accurate 3D reconstruction and novel view synthesis from a single view, significantly reducing dependence on costly multi-view systems or LiDAR [1]. This capability is crucial for advancing AR/VR, robotics, autonomous driving and digital twin technologies [1]. In this paper, we aim to address this under-explored question: Can models effectively perceive the geometry of the invisible parts of objects in a scene?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 593 + ], + "type": "text", + "content": "Unlike inpainting methods [11, 39, 50], which reconstruct missing image regions, amodal depth estimation is a novel task that extends traditional amodal segmentation [28, 48, 49] into the depth domain, by predicting the depth of occluded object parts. In this task, given an input image with a target amodal mask, the objective is to infer depth values for the occluded object regions. While amodal segmentation benefits from human annotations to collect training samples [17, 38], there is currently no device capable of collecting ground-truth depth data for occluded parts of objects at scale in real-world scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To address this challenge, prior approaches to amodal depth estimation have relied on synthetic datasets [17, 38]. However, generating these datasets is both time-consuming and hard to scale, often requiring manually placing occluders one by one [17, 38]. Additionally, synthetic data lacks the complexity and diversity of real-world scenes, resulting in a domain gap that limits the generalization of models trained on such data. Furthermore, these previous approaches only consider metric depth, where the goal is to estimate real-world distances for occluded regions. Re" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "liaence on metric depth, which naturally struggles to generalize to unseen scenarios with limited data [7, 29], exacerbates the models' poor zero-shot performance on real-world images [6, 34, 46]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 228 + ], + "type": "text", + "content": "To overcome these limitations, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth. Recent advancements in depth estimation models, such as Depth Anything [46], have enabled the generation of high-quality relative depth maps from natural images. Focusing on relative depth allows us to train models using real-world data, leveraging the depth relationships within scenes to achieve better generalization without relying on precise metric measurements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 385 + ], + "type": "text", + "content": "We introduce a new large-scale real-world dataset, Amodal Depth In the Wild (ADIW), to facilitate the training of models for amodal depth estimation. Our data generation pipeline leverages segmentation datasets to create high-quality datasets for amodal relative depth estimation with scalable annotations. We adopt a compositing approach to create training pairs, where objects are sampled and composited into scenes. Corresponding depth maps are estimated using the large depth foundation model [47]. Given the challenges posed by occlusion in relative depth estimation, we employ a scale-and-shift alignment technique to refine and blend the depth values, ensuring consistency in the final ground-truth depth map used for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 556, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 541 + ], + "type": "text", + "content": "In this work, we present two complementary frameworks for amodal relative depth estimation, targeting both deterministic and generative model classes. For deterministic models, we propose Amodal-DAV2, an extension of the Depth Anything V2 (DAV2) model [47]. This framework incorporates additional guidance layers to reason about depth in occluded regions. For generative models, we present Amodal-DepthFM, an adaptation of the DepthFM model [14], which uses a conditional flow matching approach to estimate depth in occluded regions. In both frameworks, we leverage the power of large pre-trained models with minimal modifications to effectively guide them toward predicting depth for occluded parts of objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 542, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 673 + ], + "type": "text", + "content": "Experimental results indicate that our method achieves SoTA amodal depth estimation on the ADIW dataset and demonstrates strong zero-shot performance on real-world images (Fig. 1). Comprehensive ablation studies further validate the effectiveness of our design choices, showing that object-level supervision and guidance signals significantly enhance prediction accuracy. Moreover, we demonstrate the flexibility of our models in generating diverse plausible structures for occluded regions, illustrating their adaptability to different occlusion scenarios. In summary, our contributions are:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "- A novel task formulation of amodal depth estimation focusing on relative depth, enabling improved generalization capabilities compared to previous metric-based" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9674" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 167, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 167, + 83 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 167, + 83 + ], + "type": "text", + "content": "amodal depth estimation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 179 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 84, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 132 + ], + "type": "text", + "content": "- ADIW, a large-scale real-world dataset, generated from real-world images using a scalable pipeline with segmentation datasets and compositing techniques, coupled with a scale-and-shift alignment strategy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 294, + 179 + ], + "type": "text", + "content": "- Two novel frameworks for amodal depth estimation, Amodal-DAV2 and Amodal-DepthFM, leveraging largescale pre-trained models with minimal modifications to achieve high-quality amodal depth predictions." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 141, + 205 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 213, + 214, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 214, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 214, + 226 + ], + "type": "text", + "content": "2.1. Monocular Depth Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 231, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 555 + ], + "type": "text", + "content": "Monocular depth estimation, which predicts depth from a single image, has made substantial progress in recent years [2, 4, 10, 19, 24, 46]. Early methods focused on domain-specific metric depth estimation, assuming that training and test images share similar characteristics [3, 10, 12, 23, 25]. However, this in-domain focus poses challenges for generalization to unseen domains. To address this, recent research has shifted toward cross-domain relative depth estimation, where models infer the relative depth relationships between pixels [4, 6, 34, 46, 47]. For instance, MiDaS [34] aggregates multiple datasets and trains the model in a relative depth setting, achieving superior zero-shot performance. ZoeDepth [4] demonstrates that a strong relative depth model can effectively generalize to metric depth estimation through fine-tuning. Depth Anything [46] adopts a semi-supervised strategy that combines large-scale labeled and unlabeled data to further boost the model performance. Generative approaches have also emerged, with methods like Marigold [19, 22, 53] repurposing denoising UNet models [37] alongside fixed VAE encoders [42] for depth estimation [19], and DepthFM integrates flow matching principles into the depth estimation pipeline [14]. Despite advancements in monocular depth estimation, a major challenge remains: depth estimation in occluded regions. Existing methods focus only on visible areas, leaving unseen parts unaddressed. Our work addresses this gap by enabling direct depth estimation for invisible object regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 168, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 168, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 168, + 577 + ], + "type": "text", + "content": "2.2. Amodal Perception" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Prior work in amodal perception has primarily focused on tasks such as inpainting [9, 28, 49], segmentation [20, 26, 31, 54], and detection [15, 18]. Recently, amodal depth estimation has emerged as a new challenge in amodal perception [17, 38], which aims to predict the depth of occluded regions of objects. For instance, [38] proposed the Amodal-SynthDrive dataset for synthetic driving scenes and adopted VIP-DeepLab [32] with multi-headed outputs to estimate the depth at different occlusion levels. Similarly, [17] introduced the indoor-focused Amodal-3D-FRONT dataset and an iterative approach leveraging amodal masks to estimate" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "depth in occluded areas. These works rely heavily on synthetic datasets due to the difficulty of capturing ground-truth depth behind occlusions in real-world settings. This reliance on synthetic data limits model generalization to real-world scenes, and their focus on metric depth increases the difficulty of accurate depth prediction in the zero-shot setting. In contrast, our approach adopts a relative depth-based solution using large-scale pre-trained models and realistic synthetically generated datasets, enabling strong generalization to diverse real-world images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 201, + 418, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 418, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 418, + 214 + ], + "type": "text", + "content": "2.3. Depth Inpainting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 219, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 495 + ], + "type": "text", + "content": "Depth inpainting, or depth completion, traditionally focuses on filling in missing depth values for visible regions based on sparse depth inputs [41, 43-45, 52]. Some methods aim to inpaint depth around occlusions to create novel 3D views [11, 16, 39], this differs from amodal depth estimation, which specifically targets object-level occlusions. For example, 3D Photography [39] uses shared edge guidance to combine RGB and depth inpainting, while SLIDE [16] introduces a soft-layering strategy to preserve visual details in novel views. Recently, diffusion models have shown strong image generation capabilities [37]. Invisible Stitch [11] uses a Stable Diffusion model variant [37] to fill in missing image regions, generating novel views [30], followed by a depth painting network to fill holes in the depth map with guidance from the novel-view image. Unlike these methods, which implicitly or explicitly rely on color priors and focus on visual quality, our approach directly estimates occluded region depth, prioritizing depth accuracy independently of color information or visual coherence. This distinction shifts the focus from achieving realistic view synthesis to predicting occluded depths with high geometric accuracy, an approach that enhances scene understanding for real-world applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 370, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 370, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 370, + 519 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 527, + 408, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 408, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 408, + 539 + ], + "type": "text", + "content": "3.1. Task Definition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": "Amodal depth estimation extends the concept of amodal segmentation by predicting depth information for occluded areas. In traditional amodal segmentation [28, 48, 49], given an input image and a segmentation mask of the visible portion of an object, the goal is to predict the complete object mask, including the occluded (invisible) part. Similarly, amodal depth estimation aims to estimate the depth values for the occluded regions, given an input observation image " + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": ", a corresponding observation depth map " + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": " of the input image, and a target amodal segmentation mask " + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 545, + 553, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Different from earlier works that formulate amodal depth estimation as a metric depth estimation task [17, 38], we propose a novel formulation where the goal is to predict the relative depth of the occluded parts of the object for" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9675" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 294, + 191 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 294, + 191 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 294, + 191 + ], + "type": "image", + "image_path": "62511e5e64361c171db9de542dd0681acf898dc4d3dbfcc758080d516d4821af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 296, + 266 + ], + "type": "text", + "content": "Figure 3. Constructing Training Data. We use the method from [28] to convert an initial segmentation dataset into a whole-object dataset. Next, we sample and composite images to create training pairs. Due to occluders, the relative depth predictions differ between the composite and background images, so we apply scale-and-shift alignment for consistent depth blending." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 288, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 335 + ], + "type": "text", + "content": "the input image in the wild. Recent models, such as Depth Anything [46, 47] generate high-quality relative depth maps from natural images, allowing us to train amodal depth models using real-world data as shown in Sec. 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 164, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 164, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 164, + 357 + ], + "type": "text", + "content": "3.2. Dataset Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 519 + ], + "type": "text", + "content": "We present Amodal Depth In the Wild (ADIW), a large-scale dataset specifically designed for predicting occluded object relative depths in real-world scenes. Creating a natural image dataset at scale for amodal depth estimation is challenging due to the lack of annotations for hidden regions. No existing device can capture ground-truth depth data for occluded parts at scale in real-world settings. Previous efforts have generally relied on synthetic datasets [17, 38], which, although valuable, fall short in capturing the complexity and diversity of real-world scenes. Additionally, these approaches require manually placing occluders, limiting scalability. ADIW overcomes these by generating training data from real-world images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "content": "We create paired data by overlaying objects over natural images following [28], as illustrated in Fig. 3. Our method leverages segmentation datasets, which are more scalable than traditional depth estimation datasets [21, 35, 46]. Specifically, we employ the Segment Anything model [21] to automatically generate segmentation masks from the SA-1B dataset, forming the initial segmentation dataset. Next, we apply the heuristic algorithm from [28] to filter out incomplete objects, creating a whole-object dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "In our approach, each image " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " (the background image) contains at least one complete object. We then sample an occluder object " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " (the foreground image) and superimpose it onto " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", forming the assembled observation image " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " are then processed through the Depth Anything V2 model [47] (ViT-G) to obtain relative depth maps " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_{b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", respectively. Importantly, the depth val" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ues for the background object in " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "D_{b}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " differ due to the presence of the foreground object, which changes perceived depth. Both depth maps are scaled to the range [0, 1]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "content": "To ensure consistent training labels, we apply the scale-and-shift alignment algorithm [34] to align the background object's depth values across the two depth maps. The scale factor " + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "content": " and shift factor " + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "content": " are computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 164, + 553, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 164, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 326, + 164, + 553, + 197 + ], + "type": "interline_equation", + "content": "(s, t) = \\operatorname {a r g m i n} _ {s, t} \\sum_ {i = 1} ^ {N} \\left(s d _ {i} ^ {b} + t - d _ {i} ^ {o}\\right) ^ {2}, i \\in M _ {\\text {v i s}}, \\tag {1}", + "image_path": "860f06f261f1fbef0209a2caa810be05e23b10db8df71e924f6db6a3e4b28aad.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "d_i^o" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "d_i^b" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " are the depth values of pixels in the visible part of the background object in " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "D_o" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "D_b" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " denotes the total number of valid pixels in the visible mask " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "M_{vis}" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " of the background object. The aligned depth map " + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "inline_equation", + "content": "D_{aligned}" + }, + { + "bbox": [ + 313, + 206, + 553, + 266 + ], + "type": "text", + "content": " is then calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 391, + 276, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 276, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 391, + 276, + 553, + 289 + ], + "type": "interline_equation", + "content": "D _ {\\text {a l i g n e d}} = s D _ {b} + t, \\tag {2}", + "image_path": "476042a597b97209a441ecbd1ab6cc2921f8b7ac24c0f2844fb765c279484ce1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 553, + 332 + ], + "type": "text", + "content": "serving as the ground-truth map for model training. This procedure generates a dataset of 564K images with amodal depth labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 554, + 417 + ], + "type": "text", + "content": "As shown by [46], relative depth estimators [6, 34] generalize better than metric depth estimators [2, 23, 25]. Following this insight, we produce relative depth maps for amodal depth estimation in natural scenes. Moreover, our data generation approach is also adaptable for metric amodal depth estimation by omitting the scale-and-shift alignment and utilizing metric depth models [7, 29]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 424, + 455, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 455, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 455, + 436 + ], + "type": "text", + "content": "3.3. Amodal Depth Estimator" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 554, + 548 + ], + "type": "text", + "content": "In this work, we aim to leverage large pre-trained depth models by fine-tuning them specifically for amodal depth estimation. Our goal is to make minimal modifications to the original network architectures, preserving the capabilities of the pre-trained weights while introducing the necessary adjustments to guide the model for amodal depth prediction. By integrating additional guidance into the networks, we enable them to predict depth values for occluded parts of objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 549, + 554, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 554, + 657 + ], + "type": "text", + "content": "Depth estimation models fall into two types: deterministic and generative models [13]. In this section, we introduce a dedicated strategy for adapting each model class to the amodal depth estimation task, as illustrated in Fig. 4 and Fig. 5. For the deterministic model, we adopt the Depth Anything V2 (DAV2) [47], a highly representative and top-performing pre-trained model. For the generative model, we select DepthFM [14], known for its superior ability to capture depth details and fast inference speed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 662, + 405, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 405, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 405, + 673 + ], + "type": "text", + "content": "3.3.1.Amodal-DAV2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Model Structure. The Amodal-DAV2 framework introduces subtle structural modifications to the Depth Anything V2 (DAV2) model's image encoder to retain its pre-trained" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9676" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 294, + 144 + ], + "type": "image", + "image_path": "1903eaed9dfb9a773c0be23d931b6240bc082cba83b75120ac537b8c27c5b333.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 151, + 296, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 185 + ], + "type": "text", + "content": "Figure 4. Amodal-DAV2 Framework Structure. Amodal-DAV2 modifies the DAV2 image encoder to take additional guidance channels along with RGB." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "text", + "content": "knowledge while enhancing its capacity in amodal depth estimation. The original DAV2 model employs a Vision Transformer (ViT) architecture [8], which begins with an RGB convolution applied to the input image. As shown in Fig. 4, we add a Guidance Convolution (Conv) [40] layer in parallel with the RGB Conv layer, enabling the encoder to accept additional guidance channels for the observation depth map " + }, + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "text", + "content": " and amodal mask " + }, + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 54, + 204, + 295, + 371 + ], + "type": "text", + "content": ". To ensure compatibility with the pre-trained model, the kernel weights of the Guidance Conv layer are initialized to zero, allowing the model to initially ignore the additional guidance information. Additionally, we incorporate layer normalization into the input features of the DPT head [33] to stabilize training and improve convergence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 372, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 407 + ], + "type": "text", + "content": "Training method. We fine-tune the entire Amodal-DAV2 framework during training, utilizing the standard scale-invariant loss " + }, + { + "bbox": [ + 55, + 372, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{si}" + }, + { + "bbox": [ + 55, + 372, + 296, + 407 + ], + "type": "text", + "content": " [10] as our objective function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 415, + 259, + 448 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 415, + 259, + 448 + ], + "spans": [ + { + "bbox": [ + 93, + 415, + 259, + 448 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s i} = \\alpha \\sqrt {\\frac {1}{N} \\sum_ {i \\in M _ {a}} g _ {i} ^ {2} - \\frac {\\lambda}{N ^ {2}} (\\sum_ {i \\in M _ {a}} g _ {i}) ^ {2}},", + "image_path": "3c8b56753cb2f94cbd960887171534c654bf2dd0b3e6175e44b69040e36d83d9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "inline_equation", + "content": "g_{i} = \\log \\tilde{d}_{i} - \\log d_{i}" + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\tilde{d}_i" + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": " representing the predicted depth and the ground truth depth, respectively. " + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": " denotes the number of valid pixels on the amodal mask " + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 55, + 456, + 296, + 554 + ], + "type": "text", + "content": ". Although our primary focus is on the accuracy of depth predictions for occluded (invisible) parts of the objects, we supervise the model using the entire object's depth. This holistic approach helps the model better understand the overall scene structure, leading to improved performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 163, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 163, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 163, + 570 + ], + "type": "text", + "content": "3.3.2. Amodal-DepthFM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": "Framework and Training. We adapt DepthFM [14] for amodal depth estimation using a similar approach. Given pairs of an image " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": ", its observation depth map " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": ", and corresponding amodal mask " + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "inline_equation", + "content": "M_{o}" + }, + { + "bbox": [ + 55, + 572, + 296, + 644 + ], + "type": "text", + "content": ", we fine-tune a conditional flow matching model to achieve amodal depth estimation. The objective is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 653, + 266, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 653, + 266, + 671 + ], + "spans": [ + { + "bbox": [ + 85, + 653, + 266, + 671 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {t, z, p (x _ {0})} | | v _ {\\theta} (t, \\phi_ {t} (x _ {0})) - (x _ {1} - x _ {0}) | |,", + "image_path": "53e26249525f3a8daa0c580d9e06d1fa29539ddd8517e2dd3383dbb2a8f643ea.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " represents the encoded depth samples in the latent space, and the starting point " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " corresponds to an encoded representation of the input image. The latent flow" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 200 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 200 + ], + "type": "image", + "image_path": "0a2b84a36ae8f7e333c124ea9a87255d38fc9f50ec911b06cc8ab8798af07603.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 208, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 243 + ], + "type": "text", + "content": "Figure 5. Amodal-DepthFM Framework Structure. Amodal-DepthFM modifies the DepthFM denoising UNet encoder to take additional guidance channels along with RGB latent code." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": "is conditioned on the guidance provided by " + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": ", which includes the input image latent code, observation depth map " + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": ", and amodal mask " + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": " denotes the timestamp, and " + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "inline_equation", + "content": "\\phi_t(x|z) = tx_1 + (1 - t)x_0" + }, + { + "bbox": [ + 313, + 265, + 554, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "text", + "content": "During training, we apply noise augmentation, introducing Gaussian noise " + }, + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(x,\\sigma_{min})" + }, + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "text", + "content": " to the data samples. Consequently, the conditional probability path is modeled as " + }, + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "inline_equation", + "content": "p_t(x|z) = \\mathcal{N}(x|tx_1 + (1 - t)x_0,\\sigma_{\\mathrm{min}}^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 315, + 554, + 374 + ], + "type": "text", + "content": ". For additional details, we refer readers to [14]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": "In the U-Net architecture used for the flow network " + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": ", we modify the first input Conv layer to accommodate the additional guidance channels. Specifically, we extend the channel dimensions to accept extra guidance information " + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": " alongside the image latent code. To effectively leverage the pre-trained model, we initialize the first Conv layer with a combination of pre-trained weights for the primary eight channels (corresponding to the depth and image latent codes) and zero-initialized weights for the two additional channels dedicated to guidance inputs " + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 376, + 554, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 593 + ], + "type": "text", + "content": "Scale-and-Shift Alignment Inference. While the Amodal-DAV2 model directly regresses the amodal depth map, the Amodal-DepthFM model operates in the latent space, predicting the amodal depth latent code, which is then decoded into a depth map using a latent decoder. However, learning consistent amodal depth can be challenging for A modal-DepthFM. To enhance the predicted depth map, we employ the scale-and-shift alignment [34, 36] during inference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Rather than relying solely on the model's output, we enhance the prediction by blending it with the observed depth map using a scale-and-shift alignment over the shared visible regions, as described in Eq. 1 and Eq. 2. This technique leverages the information from shared regions between the observation depth map and the amodal prediction. By calculating optimal scale and shift factors, we align the depth values of the occluded regions with the visible areas, thereby enhancing the coherence and consistency of the final depth map." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9677" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 141, + 150 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 141, + 150 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 141, + 150 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 141, + 150 + ], + "type": "image", + "image_path": "27b6df8bd02d5b590a43cb04a5af0f9553720936bcf9722cb8901da98be82ea8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 150, + 141, + 228 + ], + "blocks": [ + { + "bbox": [ + 62, + 150, + 141, + 228 + ], + "lines": [ + { + "bbox": [ + 62, + 150, + 141, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 150, + 141, + 228 + ], + "type": "image", + "image_path": "0fa99f835a456786b8a28dc0b5ab4f9c3823547461050df2a97a1e995ef5a379.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 228, + 141, + 306 + ], + "blocks": [ + { + "bbox": [ + 62, + 228, + 141, + 306 + ], + "lines": [ + { + "bbox": [ + 62, + 228, + 141, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 141, + 306 + ], + "type": "image", + "image_path": "442229eb04f48ea0acaa2904722fb5551f6a4d596ed9f4382b04eee6e0f222f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 309, + 112, + 319 + ], + "lines": [ + { + "bbox": [ + 91, + 309, + 112, + 319 + ], + "spans": [ + { + "bbox": [ + 91, + 309, + 112, + 319 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 70, + 222, + 149 + ], + "blocks": [ + { + "bbox": [ + 143, + 70, + 222, + 149 + ], + "lines": [ + { + "bbox": [ + 143, + 70, + 222, + 149 + ], + "spans": [ + { + "bbox": [ + 143, + 70, + 222, + 149 + ], + "type": "image", + "image_path": "3286d24224a3cb55896fbbc6dd93a7c1f420f3234acfdb7c1aa8f5ec2c6f69d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 143, + 149, + 222, + 227 + ], + "blocks": [ + { + "bbox": [ + 143, + 149, + 222, + 227 + ], + "lines": [ + { + "bbox": [ + 143, + 149, + 222, + 227 + ], + "spans": [ + { + "bbox": [ + 143, + 149, + 222, + 227 + ], + "type": "image", + "image_path": "7bf66e680d4469415237bb7bf23718a6034cdd5487ecca3948857bd771ca0a2f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 143, + 227, + 222, + 306 + ], + "blocks": [ + { + "bbox": [ + 143, + 227, + 222, + 306 + ], + "lines": [ + { + "bbox": [ + 143, + 227, + 222, + 306 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 222, + 306 + ], + "type": "image", + "image_path": "08bcbf48136ef3f613985fcb4ae985123bea05428dec724ed5e6cb38f05f980e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 309, + 197, + 319 + ], + "lines": [ + { + "bbox": [ + 170, + 309, + 197, + 319 + ], + "spans": [ + { + "bbox": [ + 170, + 309, + 197, + 319 + ], + "type": "text", + "content": "IS [11]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 304, + 149 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 304, + 149 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 304, + 149 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 304, + 149 + ], + "type": "image", + "image_path": "95c903c0c91e6b1e719c4a82f2038ae328e9fae73ef7ab639060d8ab11b591e2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 225, + 149, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 225, + 149, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 225, + 149, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 225, + 149, + 304, + 227 + ], + "type": "image", + "image_path": "12dfe083ac19ec224365c40ec6e5c1b79e3add275615f59a885508bd36be4ce9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 227, + 304, + 306 + ], + "blocks": [ + { + "bbox": [ + 225, + 227, + 304, + 306 + ], + "lines": [ + { + "bbox": [ + 225, + 227, + 304, + 306 + ], + "spans": [ + { + "bbox": [ + 225, + 227, + 304, + 306 + ], + "type": "image", + "image_path": "86385e5711cae45d605950561ae65d8d24ae4f787d813f0f832c0747b37a9d7d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 240, + 309, + 288, + 319 + ], + "lines": [ + { + "bbox": [ + 240, + 309, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 240, + 309, + 288, + 319 + ], + "type": "text", + "content": "Jo et al. [17]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 386, + 149 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 386, + 149 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 386, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 386, + 149 + ], + "type": "image", + "image_path": "85ad5f0bc96ad19a4ef1d7c231c749146fbc2758369a62d5777d64ea2a2c79ce.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 149, + 385, + 227 + ], + "blocks": [ + { + "bbox": [ + 307, + 149, + 385, + 227 + ], + "lines": [ + { + "bbox": [ + 307, + 149, + 385, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 385, + 227 + ], + "type": "image", + "image_path": "a97049af47e9c64e4a47591c6b950256ee388b319cd469fb04116be8da9b9dc1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 227, + 385, + 306 + ], + "blocks": [ + { + "bbox": [ + 307, + 227, + 385, + 306 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 385, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 385, + 306 + ], + "type": "image", + "image_path": "248f706b255cf266f610749425b798051519d230381c64efc424eec691c3e5ec.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 309, + 375, + 320 + ], + "lines": [ + { + "bbox": [ + 316, + 309, + 375, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 375, + 320 + ], + "type": "text", + "content": "pix2gestalt [28]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 555, + 364 + ], + "type": "text", + "content": "Figure 6. Qualitative Comparisons on the Validation Set of ADiW. Since Invisible Stitch (IS) [11] and Jo et al. [17] use metric depth estimation models [4], the depth maps are shifted compared with other methods that use relative depth estimators. Only officially released models are taken into account in this qualitative comparison." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 466, + 149 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 466, + 149 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 466, + 149 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 466, + 149 + ], + "type": "image", + "image_path": "eafc825da420f810ebacccecb3955628ed9a238d47621cd5eb3099a2b9ef93b4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 388, + 149, + 466, + 227 + ], + "blocks": [ + { + "bbox": [ + 388, + 149, + 466, + 227 + ], + "lines": [ + { + "bbox": [ + 388, + 149, + 466, + 227 + ], + "spans": [ + { + "bbox": [ + 388, + 149, + 466, + 227 + ], + "type": "image", + "image_path": "77209a82aa9c81a1ca5bc4c265e39a62062f17eae9ef77b64b86616161a7f8be.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 388, + 227, + 466, + 306 + ], + "blocks": [ + { + "bbox": [ + 388, + 227, + 466, + 306 + ], + "lines": [ + { + "bbox": [ + 388, + 227, + 466, + 306 + ], + "spans": [ + { + "bbox": [ + 388, + 227, + 466, + 306 + ], + "type": "image", + "image_path": "5c9d5d71c10befaf9e0c6d4cb2121149e8fca189227964809a1b32eab84d842d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 309, + 436, + 319 + ], + "lines": [ + { + "bbox": [ + 417, + 309, + 436, + 319 + ], + "spans": [ + { + "bbox": [ + 417, + 309, + 436, + 319 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 70, + 548, + 149 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 548, + 149 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 548, + 149 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 548, + 149 + ], + "type": "image", + "image_path": "8c2c51cae0a7eb39fa31601768fa7fe3d274e26bfd1a9ad481c0544ce59ce94b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 468, + 149, + 548, + 227 + ], + "blocks": [ + { + "bbox": [ + 468, + 149, + 548, + 227 + ], + "lines": [ + { + "bbox": [ + 468, + 149, + 548, + 227 + ], + "spans": [ + { + "bbox": [ + 468, + 149, + 548, + 227 + ], + "type": "image", + "image_path": "8f6a85dda5fcaf639f232ff13b6bc5095a5a7d59318dbd9b049485d1bd6cc17b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 468, + 227, + 548, + 306 + ], + "blocks": [ + { + "bbox": [ + 468, + 227, + 548, + 306 + ], + "lines": [ + { + "bbox": [ + 468, + 227, + 548, + 306 + ], + "spans": [ + { + "bbox": [ + 468, + 227, + 548, + 306 + ], + "type": "image", + "image_path": "fc733a17c8ee8adba2925a9bf83eb715f5949934ab604866357147871914840b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 309, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 501, + 309, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 501, + 309, + 515, + 319 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 59, + 374, + 552, + 548 + ], + "blocks": [ + { + "bbox": [ + 59, + 374, + 552, + 548 + ], + "lines": [ + { + "bbox": [ + 59, + 374, + 552, + 548 + ], + "spans": [ + { + "bbox": [ + 59, + 374, + 552, + 548 + ], + "type": "table", + "html": "
MethodEasyMid.HardOverall
RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑RMSE↓δ(%)↑
pix2gestalt‡[28]5.06790.3534.81890.2715.64184.0645.11488.717
Invisible Stitch‡[11]7.58483.3177.43581.3597.39380.0367.48181.781
Sekkat et al.†[38]11.19551.05211.40451.26711.15049.22211.26449.222
Sekkat et al.†‡[38]5.24588.7055.25488.7225.04487.4235.19488.367
Jo et al.†[17]10.26056.11810.22755.6379.98254.68710.17255.545
Jo et al.‡[17]7.55181.9336.63382.7336.48882.7936.93982.451
Jo et al.†‡[17]4.62489.8484.77789.9154.72887.2564.71289.177
Amodal-DAV2-S3.57493.6503.74193.6663.73492.1253.68293.251
Amodal-DAV2-B3.48293.8713.62094.0023.64092.6223.57893.590
Amodal-DAV2-L3.32494.4113.46094.5663.47693.3093.41894.179
Amodal-DepthFM5.35488.4275.39889.2415.49886.9705.41088.353
Amodal-DepthFM‡4.62292.4944.50092.9824.88391.0484.64592.295
", + "image_path": "078a6dbf0e9c12f1b9c62a162e9752f1f6839250e735a55f5af9bfa2ede86bdc.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons on ADIW Dataset. We compare our methods against other possible solutions. Results from Invisible Stitch [11] and Jo et al. [17] are not directly comparable, as they use metric depth estimation models [4]. " + }, + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "type": "text", + "content": ": Models retrained on our dataset for fair comparison. " + }, + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 556, + 555, + 601 + ], + "type": "text", + "content": ": Scale-and-shift alignment applied for consistent prediction handling (see 3.3). Note that Amodal-DAV2 does not rely on this alignment approach. Best results are in bold." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 137, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 644, + 115, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 115, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 115, + 656 + ], + "type": "text", + "content": "4.1. Metrics" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We follow the standard evaluation protocol proposed in previous monocular metric depth estimation works [2, 10] to evaluate the effectiveness of our framework. We use the root mean squared error (RMSE), the " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\log_{10}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " error, and the" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 622, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 555, + 706 + ], + "type": "text", + "content": "accuracy under the threshold " + }, + { + "bbox": [ + 313, + 622, + 555, + 706 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 622, + 555, + 706 + ], + "type": "text", + "content": ". Metrics consider the invisible parts of objects, with the RMSE scaled by a factor of 100 for better illustration. We also categorize objects based on their visible ratios into three difficulty levels: easy (0.75, 1], medium (0.5, 0.75], and hard (0, 0.5]. Metrics are calculated separately for each difficulty level to provide a comprehensive evaluation." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9678" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 291, + 155 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 291, + 155 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 291, + 155 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 291, + 155 + ], + "type": "table", + "html": "
MethodRMSE↓log10↓δ(%)↑
①w/o D0, Ma7.5498.60770.607
②w/o Ma4.3694.32091.193
③w/ align3.8783.65992.037
④Ours (Full)3.6823.53893.251
⑤Lsilog for inv. only3.8453.75192.608
⑥Lssi + align4.0153.94891.852
", + "image_path": "093e3d9537f688da919e9f816d2691290937495bacc57b0341bd498694ae186f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 188, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 188, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 188, + 229 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "text", + "content": "We split our dataset into a training set of approximately 559K samples and a validation set of 4K samples. The Amodal-DAV2 model is trained with a batch size of 32, a learning rate of " + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "inline_equation", + "content": "1\\mathrm{e}^{-5}" + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "text", + "content": ", and 50K iterations, using the scale-invariant log (silog) loss with " + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "inline_equation", + "content": "\\lambda = 0.85" + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "text", + "content": ", following previous works [2, 23, 25]. The Amodal-DepthFM model is trained with a batch size of 128, a learning rate of " + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "inline_equation", + "content": "3\\mathrm{e}^{-5}" + }, + { + "bbox": [ + 54, + 234, + 295, + 426 + ], + "type": "text", + "content": ", and 15K iterations. Both models are initialized with depth-pretrained parameters before fine-tuning for a modal depth estimation. We use the Adam optimizer with exponential learning rate decay. To stabilize training, we apply a max gradient norm clip of 0.01. Data augmentation is minimal, with only horizontal flipping. All experiments are conducted on 4 NVIDIA A100 GPUs. By default, we evaluate model performance using the final checkpoint after training. We use our Amodal-DAV2-L for all visualization results as default." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 140, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 140, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 140, + 444 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 450, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 450, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 56, + 450, + 295, + 665 + ], + "type": "text", + "content": "Comparison with Amodal Depth Methods. We compare our methods with existing amodal depth estimation approaches. As these methods were initially trained on synthetic datasets and relied on metric depth formulations, we retrained them on our dataset and adapted them to relative depth estimation. Results in Tab. 1 and Fig. 6 demonstrate a significant performance gap between the previous SoTA amodal depth model[17] and our methods, even after output alignment with our strategy. Notably, Amodal-DepthFM outperforms prior approaches even without the alignment, while our best model, Amodal-DAV2-L, improves performance by " + }, + { + "bbox": [ + 56, + 450, + 295, + 665 + ], + "type": "inline_equation", + "content": "27.4\\%" + }, + { + "bbox": [ + 56, + 450, + 295, + 665 + ], + "type": "text", + "content": " in terms of RMSE, setting a new SoTA. Previous methods, designed for amodal metric depth estimation, struggle to generalize across varied natural image depth ranges. In contrast, our approach leverages pretrained DAV2 [47] and DepthFM [14] models, which benefit from extensive prior knowledge of geometry and color, resulting in more accurate amodal depth predictions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Comparison with Other Solutions. We also evaluate two alternative solutions: Invisible Stitch [11] and pix2gestalt [28]. (1) For Invisible Stitch, we use SDXL [30] with a ground-truth amodal mask and image cap" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 71, + 549, + 152 + ], + "blocks": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "type": "text", + "content": "Table 2. Ablation Study for Amodal-DAV2. We investigate the impact of different guidance signals, supervision strategies, and inference techniques." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 71, + 549, + 152 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 549, + 152 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 549, + 152 + ], + "type": "table", + "html": "
MethodRMSE↓log10↓δ(%)↑
①w/o align & Do, Ma10.21112.56356.271
②w/o align & Ma5.2836.23183.799
③w/o align5.4105.23588.353
④Ours (Full)4.6453.55392.295
⑤Lrmse for inv. only5.6363.91191.739
⑥Lrmse for scene4.6083.80991.149
", + "image_path": "6e16d07fb62dce682f238dd1372f3c44f9637939628bed54f5a4fd8ad9c49111.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 160, + 554, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 554, + 216 + ], + "type": "text", + "content": "Table 3. Ablation Study for Amodal-DepthFM. We explore the effectiveness of various guidance signals, supervision strategies, and the alignment inference for Amodal-DepthFM. Unlike Amodal-DAV2, the alignment strategy significantly improves the performance of Amodal-DepthFM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 239, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 335 + ], + "type": "text", + "content": "tion to guide the inpainting. The inpainted image serves as guidance for the stitcher model [11] to perform depth inpainting, and the predicted depth is aligned with the observation depth map. (2) For pix2gestalt [28], we conduct amodal inpainting followed by depth estimation using DAV2 (ViT-G), aligning the predicted depth with the observation depth. Additional implementation details are in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 336, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 564 + ], + "type": "text", + "content": "As illustrated in Fig. 6 and Fig. 7, the accuracy of amodal depth estimation relies heavily on the quality of RGB inpainting in both methods. Even with the ground-truth amodal mask, SD-XL struggles to accurately inpaint target areas, and the stretcher model fails to predict accurate depth for occluded parts, leaving ghosting artifacts in the predicted depth map. Similarly, pix2gestalt suffers cascade errors due to inaccurate inpainting with uncontrollable amodal shapes. Moreover, since DAV2 is trained on complete natural images, applying it directly to inpainted outputs from pix2gestalt (with single foreground objects on white backgrounds) leads to performance degradation. For example, the window boundary in the first case in Fig. 6 is missing from the predicted depth map. In contrast, our amodal depth models directly regress the depth of invisible parts without relying on RGB information, using only the amodal mask as guidance. These results demonstrate that our approach provides strong geometric priors, which could also serve as a useful condition for inpainting methods [51]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 487, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 487, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 487, + 586 + ], + "type": "text", + "content": "4.4. Ablation Studies and Discussion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": "Guidance and Supervision Strategies. We conduct ablation experiments on Amodal-DAV2-S and Amodal-DepthFM to assess each framework component's contribution, with results in Tab. 2 and Tab. 3, respectively. For both models, guidance from the observation depth map " + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "inline_equation", + "content": "D_{o}" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": " and amodal mask " + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": " is crucial for optimal performance (1, 2). While the primary goal and evaluation protocol focus is estimating depth for occluded parts of objects, both frameworks benefit from object-level supervision that includes visible regions (4, 5). Interestingly, the scale-and-shift align" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9679" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 128, + 137 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 128, + 137 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 128, + 137 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 128, + 137 + ], + "type": "image", + "image_path": "7fe3af7970e7b7418b6f075b596e070705a2f1d952bb392f4284f12de286de9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 137, + 127, + 203 + ], + "blocks": [ + { + "bbox": [ + 62, + 137, + 127, + 203 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 127, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 127, + 203 + ], + "type": "image", + "image_path": "2bd26d3e13cbed7dff88f3938f31074d92338017808152b9bba10252b2446d12.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 203, + 128, + 267 + ], + "blocks": [ + { + "bbox": [ + 62, + 203, + 128, + 267 + ], + "lines": [ + { + "bbox": [ + 62, + 203, + 128, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 128, + 267 + ], + "type": "image", + "image_path": "b7b19e74f08faaf30d2d5f2b486a7eb7fb559a2328847ba426471508099ffd9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 273, + 106, + 284 + ], + "lines": [ + { + "bbox": [ + 84, + 273, + 106, + 284 + ], + "spans": [ + { + "bbox": [ + 84, + 273, + 106, + 284 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 132, + 72, + 198, + 137 + ], + "blocks": [ + { + "bbox": [ + 132, + 72, + 198, + 137 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 198, + 137 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 198, + 137 + ], + "type": "image", + "image_path": "0e5a45974ccadbf09823dd72d819a981d9332a6f2bd352cb176e64e5bd7641f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 132, + 137, + 198, + 202 + ], + "blocks": [ + { + "bbox": [ + 132, + 137, + 198, + 202 + ], + "lines": [ + { + "bbox": [ + 132, + 137, + 198, + 202 + ], + "spans": [ + { + "bbox": [ + 132, + 137, + 198, + 202 + ], + "type": "image", + "image_path": "0c4c9f6185e8491b4193ead3acc74fdae37351f7c9f9099af64638b921471d86.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 132, + 203, + 198, + 267 + ], + "blocks": [ + { + "bbox": [ + 132, + 203, + 198, + 267 + ], + "lines": [ + { + "bbox": [ + 132, + 203, + 198, + 267 + ], + "spans": [ + { + "bbox": [ + 132, + 203, + 198, + 267 + ], + "type": "image", + "image_path": "05b2221ce96e24779eac7e310790e92b738afc49b920ce75c660e85412a16b7f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 272, + 187, + 283 + ], + "lines": [ + { + "bbox": [ + 143, + 272, + 187, + 283 + ], + "spans": [ + { + "bbox": [ + 143, + 272, + 187, + 283 + ], + "type": "text", + "content": "SD-XL [30]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 203, + 72, + 268, + 137 + ], + "blocks": [ + { + "bbox": [ + 203, + 72, + 268, + 137 + ], + "lines": [ + { + "bbox": [ + 203, + 72, + 268, + 137 + ], + "spans": [ + { + "bbox": [ + 203, + 72, + 268, + 137 + ], + "type": "image", + "image_path": "1f15514eafe16a9afcaf31c919c76ed90c2c2b25a29e4bed393d07ecf1f955e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 203, + 137, + 268, + 202 + ], + "blocks": [ + { + "bbox": [ + 203, + 137, + 268, + 202 + ], + "lines": [ + { + "bbox": [ + 203, + 137, + 268, + 202 + ], + "spans": [ + { + "bbox": [ + 203, + 137, + 268, + 202 + ], + "type": "image", + "image_path": "ca773620b14d2ca46110baa10c722a002f6775bfccd9a7f0f56afb7f56de8750.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 203, + 268, + 267 + ], + "blocks": [ + { + "bbox": [ + 203, + 203, + 268, + 267 + ], + "lines": [ + { + "bbox": [ + 203, + 203, + 268, + 267 + ], + "spans": [ + { + "bbox": [ + 203, + 203, + 268, + 267 + ], + "type": "image", + "image_path": "51e09fbf810bcd1eff2b47293744fa1ae41843f6de731dc9c6ffdeb495e1746e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 272, + 249, + 283 + ], + "lines": [ + { + "bbox": [ + 221, + 272, + 249, + 283 + ], + "spans": [ + { + "bbox": [ + 221, + 272, + 249, + 283 + ], + "type": "text", + "content": "IS [11]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 273, + 72, + 338, + 137 + ], + "blocks": [ + { + "bbox": [ + 273, + 72, + 338, + 137 + ], + "lines": [ + { + "bbox": [ + 273, + 72, + 338, + 137 + ], + "spans": [ + { + "bbox": [ + 273, + 72, + 338, + 137 + ], + "type": "image", + "image_path": "97f7e436e5a43da0e90912a6f88b6e8a4271b33f4334165c1878db4937b5f39c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 273, + 137, + 337, + 202 + ], + "blocks": [ + { + "bbox": [ + 273, + 137, + 337, + 202 + ], + "lines": [ + { + "bbox": [ + 273, + 137, + 337, + 202 + ], + "spans": [ + { + "bbox": [ + 273, + 137, + 337, + 202 + ], + "type": "image", + "image_path": "5a9c3a8bc3e96d445e2a4fc0ce2e6b600ba83239076a401adb26074cf05f6ca6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 273, + 203, + 337, + 267 + ], + "blocks": [ + { + "bbox": [ + 273, + 203, + 337, + 267 + ], + "lines": [ + { + "bbox": [ + 273, + 203, + 337, + 267 + ], + "spans": [ + { + "bbox": [ + 273, + 203, + 337, + 267 + ], + "type": "image", + "image_path": "83f256ccdd6dfb5f8d7652cac1f6564fb6e4ec98571752e01b1b6a7f3a6ad0ef.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 276, + 272, + 334, + 284 + ], + "lines": [ + { + "bbox": [ + 276, + 272, + 334, + 284 + ], + "spans": [ + { + "bbox": [ + 276, + 272, + 334, + 284 + ], + "type": "text", + "content": "pix2gestalt [28]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 343, + 72, + 408, + 137 + ], + "blocks": [ + { + "bbox": [ + 343, + 72, + 408, + 137 + ], + "lines": [ + { + "bbox": [ + 343, + 72, + 408, + 137 + ], + "spans": [ + { + "bbox": [ + 343, + 72, + 408, + 137 + ], + "type": "image", + "image_path": "683a0f9b3a41e48f33d2409f4584a48b4c6ac5d745775aef971b23726578eedd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 343, + 137, + 408, + 202 + ], + "blocks": [ + { + "bbox": [ + 343, + 137, + 408, + 202 + ], + "lines": [ + { + "bbox": [ + 343, + 137, + 408, + 202 + ], + "spans": [ + { + "bbox": [ + 343, + 137, + 408, + 202 + ], + "type": "image", + "image_path": "7b9e32f1a5c685e44ae78cd92909188cb4c7b01a0033a19e4c91aa13a734fb70.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 343, + 203, + 408, + 267 + ], + "blocks": [ + { + "bbox": [ + 343, + 203, + 408, + 267 + ], + "lines": [ + { + "bbox": [ + 343, + 203, + 408, + 267 + ], + "spans": [ + { + "bbox": [ + 343, + 203, + 408, + 267 + ], + "type": "image", + "image_path": "d54c8718535dcdff9abd4bbbf84d85ba21e90713865277b2145f7495e69e0832.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 272, + 405, + 284 + ], + "lines": [ + { + "bbox": [ + 346, + 272, + 405, + 284 + ], + "spans": [ + { + "bbox": [ + 346, + 272, + 405, + 284 + ], + "type": "text", + "content": "pix2gestalt [28]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 412, + 72, + 477, + 137 + ], + "blocks": [ + { + "bbox": [ + 412, + 72, + 477, + 137 + ], + "lines": [ + { + "bbox": [ + 412, + 72, + 477, + 137 + ], + "spans": [ + { + "bbox": [ + 412, + 72, + 477, + 137 + ], + "type": "image", + "image_path": "d26a5b16a6f7fba7c337730e538bfa509cb15d1f92d64d3c8d9ffa73915ebfc5.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 412, + 137, + 477, + 202 + ], + "blocks": [ + { + "bbox": [ + 412, + 137, + 477, + 202 + ], + "lines": [ + { + "bbox": [ + 412, + 137, + 477, + 202 + ], + "spans": [ + { + "bbox": [ + 412, + 137, + 477, + 202 + ], + "type": "image", + "image_path": "2278828038f0232a7459cce193c55ec759924a39b53bb78bca47b6f0fc1cbac4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 412, + 203, + 477, + 267 + ], + "blocks": [ + { + "bbox": [ + 412, + 203, + 477, + 267 + ], + "lines": [ + { + "bbox": [ + 412, + 203, + 477, + 267 + ], + "spans": [ + { + "bbox": [ + 412, + 203, + 477, + 267 + ], + "type": "image", + "image_path": "9d852c171bd387eecba0ffa12672ec666abfbec47babea4495bbcd79b093026b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 273, + 455, + 283 + ], + "lines": [ + { + "bbox": [ + 435, + 273, + 455, + 283 + ], + "spans": [ + { + "bbox": [ + 435, + 273, + 455, + 283 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 483, + 72, + 548, + 137 + ], + "blocks": [ + { + "bbox": [ + 483, + 72, + 548, + 137 + ], + "lines": [ + { + "bbox": [ + 483, + 72, + 548, + 137 + ], + "spans": [ + { + "bbox": [ + 483, + 72, + 548, + 137 + ], + "type": "image", + "image_path": "82808fb121eaf921de0fb957cb8a11f571f9fe51f7302519100cc3d08a872cec.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 483, + 137, + 548, + 202 + ], + "blocks": [ + { + "bbox": [ + 483, + 137, + 548, + 202 + ], + "lines": [ + { + "bbox": [ + 483, + 137, + 548, + 202 + ], + "spans": [ + { + "bbox": [ + 483, + 137, + 548, + 202 + ], + "type": "image", + "image_path": "3eba01883d4649c04f9433c9ca77c06ee275f667dc2e1252822da2a354eebb93.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 483, + 203, + 548, + 267 + ], + "blocks": [ + { + "bbox": [ + 483, + 203, + 548, + 267 + ], + "lines": [ + { + "bbox": [ + 483, + 203, + 548, + 267 + ], + "spans": [ + { + "bbox": [ + 483, + 203, + 548, + 267 + ], + "type": "image", + "image_path": "4e2a0e71bf0d5c416a0bb1ca7f86c8dff73e53778232b38a12aea94cb6c9cd6f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 273, + 522, + 283 + ], + "lines": [ + { + "bbox": [ + 509, + 273, + 522, + 283 + ], + "spans": [ + { + "bbox": [ + 509, + 273, + 522, + 283 + ], + "type": "inline_equation", + "content": "D_{o}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 57, + 368, + 117, + 418 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 350 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparisons on Images in the Wild. Invisible Stitch (IS) [11] uses SD-XL [30] with a ground-truth visible mask and image caption to inpaint occluded areas, while pix2gestalt [28] completes the occluded areas via amodal inpainting. Both methods suffer from inaccurate shape completion and cascade depth errors. The third row highlights the fragile shape completion of previous work: both SD-XL and pix2gestalt fail to remove the foreground cabinets correctly. In the Human Heuristics mode (Fig. 2), our method infers without any RGB inpainting or completion prior and achieves reasonable depth estimations for occluded areas." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 368, + 117, + 418 + ], + "lines": [ + { + "bbox": [ + 57, + 368, + 117, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 368, + 117, + 418 + ], + "type": "image", + "image_path": "36fc49f35d9c6fb1880fce4f30b0951b185a8b42385785201dfeb6b2e11651ad.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 117, + 368, + 176, + 418 + ], + "blocks": [ + { + "bbox": [ + 117, + 368, + 176, + 418 + ], + "lines": [ + { + "bbox": [ + 117, + 368, + 176, + 418 + ], + "spans": [ + { + "bbox": [ + 117, + 368, + 176, + 418 + ], + "type": "image", + "image_path": "56a121a7613b4099c1e03fe992fe356ae89ab9c38410e0f1ea75345fe3c95f66.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 428, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 473 + ], + "type": "text", + "content": "Figure 8. Results with Different Mask Guidance. Guided by various amodal masks, our model successfully predicts the corresponding amodal depth maps for the target objects in the image, showcasing flexibility in estimating occluded regions." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 177, + 368, + 236, + 418 + ], + "blocks": [ + { + "bbox": [ + 177, + 368, + 236, + 418 + ], + "lines": [ + { + "bbox": [ + 177, + 368, + 236, + 418 + ], + "spans": [ + { + "bbox": [ + 177, + 368, + 236, + 418 + ], + "type": "image", + "image_path": "04b428f399dc37639a109ec5826d6aabe374656eec0b061cbd11a28da6e1bee6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 236, + 368, + 294, + 418 + ], + "blocks": [ + { + "bbox": [ + 236, + 368, + 294, + 418 + ], + "lines": [ + { + "bbox": [ + 236, + 368, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 236, + 368, + 294, + 418 + ], + "type": "image", + "image_path": "dbe1f1eb3ffe267a4955a01c453509d7008fbe38aceb101ced48b1f346e78ba0.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 495, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 544 + ], + "type": "text", + "content": "ment during inference improves Amodal-DepthFM performance but decreases that of Amodal-DAV2-S, indicating that inconsistent depth estimation may hinder Amodal-DepthFM's overall performance (③)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "content": "Varying Predictions with Different Guidance. Fig. 8 demonstrates the flexibility of our amodal depth estimation model in generating diverse, plausible predictions based on different amodal masks as guidance. This capability is crucial when multiple valid interpretations exist for occluded object parts. For example, the right-hand chair occluded by a wall could vary in width or even be connected to another occluded chair on the left. Our Amodal-DAV2-L adapts the amodal depth output according to the provided mask." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Edge Region v.s., Non-Edge Region. In this experiment, we extract edge areas from the ground-truth depth map and calculate metrics for both edge and non-edge regions in the invisible area. The overall RMSE from Amodal-DAV2-S is 3.682, with the RMSE for the edge region significantly" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 441 + ], + "type": "text", + "content": "higher at 7.378, compared to only 3.515 for non-edge regions. This discrepancy highlights the challenges in estimating amodal depth with fine-grained details. This finding aligns with human perception, where it is generally more difficult to accurately recover fine-grained geometry than the coarser structures for occluded areas." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 388, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "We presented a novel approach to amodal depth estimation, focusing on predicting the depth of invisible parts of objects in natural scenes. Our work introduces Amodal Depth In the Wild (ADIW), a large-scale dataset that leverages segmentation datasets and a compositing pipeline, enabling high-quality, real-world amodal depth annotations. We proposed two complementary frameworks for amodal depth estimation: Amodal-DAV2 and Amodal-DepthFM. Amodal-DAV2 leverages the deterministic capabilities of Depth Anything V2 to achieve state-of-the-art performance in relative depth estimation, while A modal-DepthFM, built on a generative flow matching paradigm, and excels in providing finer details and sharper boundaries in occluded regions. Our experiments highlight that both models benefit from object-level supervision and the importance of guidance signals for improving amodal depth prediction accuracy. This work not only sets a new benchmark for amodal depth estimation but also opens the door for future research in occluded geometry understanding and improving applications such as inpainting and scene reconstruction." + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9680" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 138 + ], + "type": "text", + "content": "This publication is supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 201 + ], + "type": "text", + "content": "[1] Jiayang Ao, Qiuhong Ke, and Krista A. Ehinger. Image amodal completion: A survey. In Computer Vision and Image Understanding. Elsevier, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 236 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In CVPR, pages 4009-4018, 2021. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Ibraheem Alhashim, and Peter Wonka. Localbins: Improving depth estimation by learning local distributions. In European Conference on Computer Vision, pages 480-496. Springer, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 281, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 295, + 324 + ], + "type": "text", + "content": "[4] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 346 + ], + "type": "text", + "content": "[5] Amlaan Bhoi. Monocular depth estimation: A survey. arXiv preprint arXiv:1901.09402, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 348, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 294, + 380 + ], + "type": "text", + "content": "[6] Reiner Birkl, Diana Wofk, and Matthias Muller. Midas v3.1 - a model zoo for robust monocular relative depth estimation. arXiv preprint arXiv:2307.14460, 2023. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 381, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 295, + 434 + ], + "type": "text", + "content": "[7] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 501 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 502, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 534 + ], + "type": "text", + "content": "[9] Kiana Ehsani, Roozbeh Mottaghi, and Ali Farhadi. Segan: Segmenting and generating the invisible. In CVPR, pages 6144-6153, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 568 + ], + "type": "text", + "content": "[10] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. NeurIPS, 27, 2014. 1, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 613 + ], + "type": "text", + "content": "[11] Paul Engstler, Andrea Vedaldi, Iro Laina, and Christian Rupprecht. Invisible stitch: Generating smooth 3d scenes with depth inpainting. arXiv preprint arXiv:2404.19758, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 656 + ], + "type": "text", + "content": "[12] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[13] Yongtao Ge, Guangkai Xu, Zhiyue Zhao, Libo Sun, Zheng Huang, Yanlong Sun, Hao Chen, and Chunhua Shen. Geobench: Benchmarking and analyzing monocular geometry estimation models. arXiv preprint arXiv:2406.12671, 2024. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[14] Ming Gui, Johannes S Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Björn Ommer. Depthfm: Fast monocular depth estimation with flow matching. arXiv preprint arXiv:2403.13788, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 161 + ], + "type": "text", + "content": "[15] Cheng-Yen Hsieh, Tarasha Khurana, Achal Dave, and Deva Ramanan. Tracking any object amodally. arXiv preprint arXiv:2312.12433, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 227 + ], + "type": "text", + "content": "[16] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In ICCV, pages 12518-12527, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 272 + ], + "type": "text", + "content": "[17] Seong-Uk Jo, Du Yeol Lee, and Chae Eun Rhee. Occlusion-aware amodal depth estimation for enhancing 3d reconstruction from a single image. IEEE Access, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 308 + ], + "type": "text", + "content": "[18] Abhishek Kar, Shubham Tulsiani, Joao Carreira, and Jitendra Malik. Amodal completion and size constancy in natural scenes. In ICCV, pages 127-135, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[19] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, pages 9492-9502, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[20] Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion-aware instance segmentation with overlapping bilayers. In CVPR, pages 4019–4028, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 487 + ], + "type": "text", + "content": "[22] Mykola Lavreniuk, Shariq Farooq Bhat, Matthias Muller, and Peter Wonka. Evp: Enhanced visual perception using inverse multi-attentive feature refinement and regularized image-text alignment. In European Conference on Computer Vision Workshops (ECCVW), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "text", + "content": "[23] Zhenyu Li, Xuyang Wang, Xianming Liu, and Junjun Jiang. Binsformer: Revisiting adaptive bins for monocular depth estimation. arXiv preprint arXiv:2204.00987, 2022. 1, 3, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 577 + ], + "type": "text", + "content": "[24] Zhenyu Li, Shariq Farooq Bhat, and Peter Wonka. Patchfusion: An end-to-end tile-based framework for high-resolution monocular metric depth estimation. arXiv preprint arXiv:2312.02284, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 622 + ], + "type": "text", + "content": "[25] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, pages 1-18, 2023. 3, 4, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 655 + ], + "type": "text", + "content": "[26] Ruoshi Liu and Carl Vondrick. Humans as light bulbs: 3d human reconstruction from thermal reflection. In CVPR, pages 12531-12542, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[27] Alican Mertan, Damien Jade Duff, and Gozde Unal. Single image depth estimation: An overview. Digital Signal Processing, 123:103441, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[28] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9681" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "Amodal segmentation by synthesizing wholes. In CVPR, pages 3931-3940. IEEE Computer Society, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 149 + ], + "type": "text", + "content": "[29] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, pages 10106-10116, 2024. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[30] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "text", + "content": "[31] Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In CVPR, pages 3014-3023, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[32] Siyuan Qiao, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. Vip-deeplab: Learning visual perception with depth-aware video panoptic segmentation. In CVPR, pages 3997-4008, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 315 + ], + "type": "text", + "content": "[33] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, pages 12179-12188, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE TPAMI, 44(3), 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 414 + ], + "type": "text", + "content": "[35] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "text", + "content": "[36] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In CVPR, pages 3762-3772, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "text", + "content": "[38] Ahmed Rida Sekkat, Rohit Mohan, Oliver Sawade, Elmar Matthes, and Abhinav Valada. Amodalsynthdrive: A synthetic amodal perception dataset for autonomous driving. IEEE Robotics and Automation Letters, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "text", + "content": "[39] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In CVPR, pages 8028-8038, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 624 + ], + "type": "text", + "content": "[40] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 658 + ], + "type": "text", + "content": "[41] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In CVPR, pages 9763-9772, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "text", + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 30, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[43] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In CVPR, pages 9422-9432, 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[44] Yufei Wang, Ge Zhang, Shaoqian Wang, Bo Li, Qi Liu, Le Hui, and Yuchao Dai. Improving depth completion via depth feature upsampling. In CVPR, pages 21104-21113, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[45] Zhiqiang Yan, Yuankai Lin, Kun Wang, Yupeng Zheng, Yufei Wang, Zhenyu Zhang, Jun Li, and Jian Yang. Tri-perspective view decomposition for geometry-aware depth completion. In CVPR, pages 4874-4884, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[46] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. arXiv preprint arXiv:2401.10891, 2024. 1, 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 238 + ], + "type": "text", + "content": "[47] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[48] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. A modal ground truth and completion in the wild. In CVPR, pages 28003-28013, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[49] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[50] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In CVPR, pages 3784-3792, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 376 + ], + "type": "text", + "content": "[51] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836-3847, 2023. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 421 + ], + "type": "text", + "content": "[52] Youmin Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and Stefano Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In CVPR, pages 18527-18536, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 476 + ], + "type": "text", + "content": "[53] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5729-5739, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[54] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In CVPR, pages 1464-1472, 2017. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9682" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_content_list.json b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56d10674e528970f912ba25ba559fe8072477ce2 --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_content_list.json @@ -0,0 +1,2641 @@ +[ + { + "type": "text", + "text": "Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images", + "text_level": 1, + "bbox": [ + 165, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianhao Wu $^{1*}$ , Chuanxia Zheng $^{2\\dagger}$ , Frank Guan $^{3}$ , Andrea Vedaldi $^{2}$ , Tat-Jen Cham $^{1}$ $^{*}$ S-Lab, $^{1}$ Nanyang Technological University, $^{2}$ VGG, University of Oxford \n $^{3}$ Singapore Institute of Technology", + "bbox": [ + 174, + 179, + 820, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://sm0kywu.github.io/Amodal3R/", + "bbox": [ + 380, + 247, + 612, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db0f3e260a940e93d176cfcacfda74c200ff73d2202082359e50a6d358d3a7e1.jpg", + "image_caption": [ + "3D Recast", + "Input Image", + "Stochastic sample 1" + ], + "image_footnote": [], + "bbox": [ + 99, + 292, + 321, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4be72a6742cbba6fb6e0aa6ec50793ba1faff5009bfb21fdf47d0948e8aeff5f.jpg", + "image_caption": [ + "Stochastic sample 2" + ], + "image_footnote": [], + "bbox": [ + 330, + 277, + 470, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7042bf3f32f04a0978ae549a353e1b5a2825646847ab201d5bcb3d0a4fb37a3e.jpg", + "image_caption": [ + "In-the-wild Reconstruction", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 478, + 292, + 616, + 397 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8b639ca0d9316b8f72ff13100dfe1cd2bf2e8bfe27f71d18e823960df0befc9.jpg", + "image_caption": [ + "Stochastic sample 1" + ], + "image_footnote": [], + "bbox": [ + 616, + 286, + 756, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4bd49bd69f8a811e2e55f5c2ce9daeebf959d908174fb4622f831ed364f0e8fa.jpg", + "image_caption": [ + "Stochastic sample 2" + ], + "image_footnote": [], + "bbox": [ + 763, + 285, + 903, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c125d3a9582a8e5d07e1a535e32ca82429b560e62da9f5848bd3c787846ec90.jpg", + "image_caption": [ + "3D Scene Decomposition", + "Input Image", + "Figure 1. Example results of Amodal3R. Given partially visible objects within images (occluded regions are shown in black, visible areas in red outlines), our Amodal3R generates diverse semantically meaningful 3D assets with reasonable geometry and plausible appearance. We sample multiple times to get diverse results from the same occluded input. Trained on synthetic datasets, it generalizes well to out-of-domain test examples from real scenes, where most objects are partially visible, and reconstructs reasonable 3D assets." + ], + "image_footnote": [], + "bbox": [ + 107, + 436, + 220, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73906cc49696dba9ce937eb09543a55bea4527421a3d01dd6b8f693af5fc1014.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 436, + 444, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22060cd08dc52324121605dc107f208dc0f6955396270030ef0833c53c03ddd9.jpg", + "image_caption": [ + "Decomposed 3D assets" + ], + "image_footnote": [], + "bbox": [ + 452, + 436, + 669, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/39b7ea3e1098b79d12fdccc7992e0db1a7fdd8ecd2c429d3b867541db7111446.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 436, + 895, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 619, + 326, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing image-to-3D models assume that objects are fully visible, ignoring occlusions that commonly occur in real-world scenarios. In this paper, we introduce Amodal3R, a conditional image-to-3D model designed to reconstruct plausible 3D geometry and appearance from partial observations. We extend a \"foundation\" 3D generator by introducing a visible mask-weighted attention mechanism and an occlusion-aware attention layer that explicitly leverage visible and occlusion priors to guide the reconstruction process. We demonstrate that, by training solely on synthetic data, Amodal3R learns to recover full 3D objects even in the presence of occlusions in real scenes. It substantially outperforms state-of-the-art methods that independently perform 2D amodal completion followed by 3D reconstruction, thereby establishing a new benchmark for occlusion-aware 3D reconstruction.", + "bbox": [ + 88, + 640, + 485, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 619, + 643, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans possess a remarkable ability to infer the complete 3D shape and appearance of objects from single views, even when those objects are partly hidden behind occluders. This ability, known as amodal completion, is key to operating in complex real-world environments, where objects are often partially occluded by surrounding clutter. However, existing image-based 3D reconstruction models [5, 6, 23, 41, 47, 48, 65, 70, 71, 75, 77, 80, 88, 89] fail to recover full 3D assets when the object is partially occluded. We thus consider the problem of reconstructing 3D objects from one or more partially occluded views, a novel task that we call amodal 3D reconstruction.", + "bbox": [ + 509, + 651, + 906, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amodal 3D reconstruction is a challenging task that requires reconstructing an object's photorealistic 3D geometry and appearance, including the occluded regions, which are inherently ambiguous. Previous approaches [7, 54] typ", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project Lead.", + "bbox": [ + 107, + 886, + 189, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9181", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ically decompose the task into a two-stage pipeline: first performing 2D amodal completion [85, 86, 91], followed by conventional 3D reconstruction [41, 75, 89]. While these two-stage pipelines are straightforward to implement, they have some drawbacks. First, 2D amodal completion methods rely predominantly on appearance-only priors, without incorporating explicit information on the 3D structure. This lack of 3D geometric reasoning limits their ability to generate physically plausible completions. Second, some 3D reconstruction methods depend on, or benefit from, multiview inputs. In this case, 2D amodal completion may lack multi-view consistency, particularly when performed independently for different views, which confuses the 3D reconstruction process.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present Amodal3R, a novel single-stage paradigm for amodal 3D reconstruction that surpasses previous state-of-the-art approaches. Amodal3R fine-tunes a pre-trained 3D generator by introducing a dedicated branch that explicitly models occlusions. Its key advantage is performing reconstruction, completion, and occlusion reasoning directly in a 3D latent space rather than using a two-stage scheme. With this, the model can reconstruct both visible and occluded regions of the object coherently and plausibly. Specifically, we introduce a mask-weighted cross-attention to highlight visible information, along with a new occlusion-aware attention layer to guide completion in occluded areas. Together, these components enable the model to focus more on the visible parts of the object without significantly perturbing the pre-trained model.", + "bbox": [ + 89, + 311, + 482, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We designed several benchmarks to evaluate the performance of Amodal3R, including object-level Google Scanned Objects [16] and Toys4K [67] augmented with occlusions, 3D scenes from Replica [68], and real-world monocular images. Without relying on additional heuristics, Amodal3R achieves significantly superior performance compared to state-of-the-art models and generalizes well to different datasets, including real ones.", + "bbox": [ + 89, + 547, + 482, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are as follows:", + "bbox": [ + 109, + 679, + 447, + 694 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a feed-forward 3D reconstruction model that directly reconstructs complete and high-quality 3D objects from one or more partially occluded views.", + "- We introduce a mask-weighted cross-attention mechanism and an occlusion-aware layer to inject occlusion awareness into an existing 3D reconstruction model, improving both the geometry and appearance of reconstructed objects when they are partially occluded.", + "- We demonstrate, via experiments on the 3D scene dataset Replica and real-world images, that our one-stage pipeline significantly outperforms existing two-stage approaches, establishing a new benchmark for amodal 3D reconstruction." + ], + "bbox": [ + 91, + 704, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background", + "text_level": 1, + "bbox": [ + 513, + 89, + 640, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Amodal Completion. Recent advances in 2D amodal completion [2, 54, 83, 85, 86, 91] have achieved significant success in reconstructing occluded regions of objects in 2D images. While these methods show promise for 3D generation pipelines [2, 54], they still have limitations. First, 2D amodal completion models have limited 3D geometric priors, as they interpret images as 2D patterns. While they excel at completing textures, they may generate physically implausible geometries when handling complex occlusions. This stems from their lack of explicit 3D reasoning and reliance on 2D appearance priors, without true volumetric understanding. Second, for models that use multi-view images as input, the results of 2D amodal completion are often inconsistent across views. Inconsistent views need to be handled by the 3D reconstructors, which causes confusion (see Sec. 4). Although there has been significant work on multiview consistent generation [7, 62, 70, 72, 81, 89, 94], multiview consistent completion is less explored. RenderDiffusion [1] contains an example, but the results are often blurry or lack details. This motivates our 3D-centric reconstruction framework.", + "bbox": [ + 511, + 114, + 903, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Shape Completion. Several methods [12, 14, 69, 95] start from a partial 3D reconstruction, then complete it in 3D. They use encoder-decoder architectures [24] or volumetric representations [10, 29] to robustly recover 3D shape, but not texture, which is necessary in many applications. They also still require recovering the partial 3D geometry from an occluded image, a challenge in itself. Furthermore, they ignore the input image when completing the object in 3D, disregarding the occlusion pattern that caused the 3D geometry to be recovered only partially. Recent work [11, 82] utilizes the partially visible object as input specifically for 3D shape completion. However, it focuses solely on geometry reconstruction, without recovering the object's texture and appearance. In contrast, our approach is end-to-end, recovering the complete 3D shape and appearance of the object from the occluded image.", + "bbox": [ + 511, + 431, + 903, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Generative Models. Early advancements in 3D generation are based on GANs [19], exploring various 3D representations such as point clouds [25, 39], voxel grids [84, 96], view sets [48, 55], NeRF [4, 5, 50, 60], SDF [18], and 3D Gaussian mixtures [78]. While GANs can capture complex 3D structures, they struggle with stability, scalability, and data diversity. The focus then shifted to diffusion models [22, 59, 66], which are more capable and versatile. They were first applied to novel view synthesis [76] in image space, before expanding to model a variety of 3D representations, including point clouds [44, 45, 49, 79, 93], voxel grids [37, 46], triplanes [64, 98], meshes [20, 43], and 3D Gaussian mixtures [9, 35, 38, 87]. Autoregressive models [8, 47, 65] have also been explored for mesh generation, focusing on improving the topology of the final 3D mesh.", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9182", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec17aea1d842cf4d5ac780498601b8d2b71d4395e183436383899c5837c417a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 92, + 897, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3a7b54826228d84368a9b61c041874faaffaf8f69e5964d42e366eec13c32630.jpg", + "image_caption": [ + "Figure 2. Overview of Amodal3R. Given an image as input and the regions of interest, Amodal3R first extracts the partially visible target object, along with the visibility and occlusion masks using an off-the-shelf 2D segmenter. It then applies DINOV2 [53] to extract features $c_{\\mathrm{dino}}$ as additional conditioning for the 3D reconstructor. To enhance occlusion reasoning, each transformer block incorporates a mask-weighted cross-attention (via $c_{\\mathrm{vis}}$ ) and occlusion-aware attention layer (via $c_{\\mathrm{occ}}$ ), ensuring the 3D reconstructor accurately perceives visible information while effectively inferring occluded parts. For conditioning details, see Sec. 3.1." + ], + "image_footnote": [], + "bbox": [ + 98, + 227, + 898, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An alternative to performing diffusion in 2D image space [62, 70, 89] is to use 3D latent spaces [28, 51, 58, 63, 74, 80]. Such methods typically consist of two stages: the first for generating geometry and the second for generating appearance, and are capable of high-quality imaged-to-3D generation. However, they assume that input objects are fully visible, which limits their application to real-world scenes, where occlusions are frequent. Here, we extend such approaches to generate high-quality 3D assets from occluded input images — a challenging task that requires inferring complete 3D objects from partial information.", + "bbox": [ + 88, + 455, + 480, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 633, + 183, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider an image $x$ containing a partially occluded view of an object of interest $o_i$ . We wish to reconstruct the complete 3D shape and appearance $y$ of the object $o_i$ . This task is conceptually similar to existing image-to-3D reconstruction, except that here the object is partially occluded instead of being fully visible in $x$ . We call this problem amodal 3D reconstruction.", + "bbox": [ + 88, + 657, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Amodal3R (Fig. 2), a new method for amodal 3D reconstruction. Unlike recent two-stage methods [7, 54] that first perform 2D amodal completion followed by 3D reconstruction, Amodal3R is an end-to-end occlusion-aware 3D reconstruction model that generates complete geometry and appearance directly within the 3D space. Formally, Amodal3R is a conditional generator $v_{\\theta}(y|o_i,M_{\\mathrm{vis}},M_{\\mathrm{occ}})$ that takes as input the image $x$ , centered on a partially visible object $o_i$ , as well as the visibility", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mask $M_{\\mathrm{vis}}$ and occlusion mask $M_{\\mathrm{occ}}$ . The visibility mask $M_{\\mathrm{vis}}$ highlights the pixels in $x$ that correspond to object $o_i$ , while $M_{\\mathrm{occ}}$ marks the pixels occupied by occluders, i.e., all other objects that potentially obscure $o_i$ . For real images, these masks can be efficiently obtained using pre-trained 2D segmentation models like Segment Anything [34]. By providing point coordinate prompts for the object of interest $o_i$ and its occluders, the segmentation model can generate the corresponding masks. In cases where multiple occluders are present, or when occluders fragment the visible area of the target object, a sequential process is employed. Specifically, point prompts for each visible/occluding region are provided to the 2D segmenter one at a time, with the model generating an individual mask for each region. These masks are then aggregated to form comprehensive visibility and occlusion masks.", + "bbox": [ + 511, + 455, + 906, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The challenges for Amodal3R are how to: (1) produce a plausible and complete 3D shape and appearance from partial observations, even in the absence of multi-view information; and (2) ensure 3D consistency in terms of geometry and photometry, maintaining seamless visual coherence between visible and generated regions.", + "bbox": [ + 511, + 698, + 908, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Mask-Conditional Generative Models", + "text_level": 1, + "bbox": [ + 511, + 801, + 841, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To fine-tune an image-to-3D model for partially occluded inputs, the key change is to condition the transformer $v_{\\theta}$ not only on the object of interest $o_i$ , but also on the visibility and occlusion masks, i.e., $M_{\\mathrm{vis}}$ and $M_{\\mathrm{occ}}$ . A naive approach is to patchify or embed the masks into tokens $(c_{\\mathrm{vis}}, c_{\\mathrm{occ}})$ ,", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9183", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87381d95c462752cdc81640281a904c781af6b72aae913533c761f2102a88737.jpg", + "image_caption": [ + "(a) TRELLIS [80]", + "Figure 3. The transformer structure of Amodal3R. Compared with the original TRELLIS [80] design, we further introduce the mask-weighted cross-attention and occlusion-aware layer. It applies to both sparse structure and SLAT diffusion models." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 243, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b19bdfaaecc7a073d5f373b9f114d67bb5ce602a4bcad024930a9af08e96a53.jpg", + "image_caption": [ + "(b) Amodal3R (Ours)" + ], + "image_footnote": [], + "bbox": [ + 250, + 88, + 480, + 304 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which can be concatenated to the image tokens $\\pmb{e}_{\\mathrm{dino}}$ and processed by cross-attention as before. However, the RGB $\\pmb{o}_i$ contains significantly more information than the binary masks $M_{\\mathrm{vis}}$ and $M_{\\mathrm{occ}}$ . The transformer, which is initially trained with RGB only, may simply ignore the mask information. This is compounded by the fact that learning to use this new information, which involves a new type of token that is incompatible with image tokens, may require aggressive fine-tuning of the transformer.", + "bbox": [ + 89, + 407, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To sidestep this problem, inspired by masked attention approaches in 2D completion [92], we introduce mask-weighted cross-attention and an occlusion-aware attention layer to better utilize the visibility mask $M_{\\mathrm{vis}}$ and the occlusion mask $M_{\\mathrm{occ}}$ without disrupting the pre-trained 3D generator. These are described next.", + "bbox": [ + 89, + 542, + 483, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask-weighted Cross-Attention. A key novel component of Amodal3R is mask-weighted cross-attention, which allows the model to focus its attention on the visible parts of the object. We implement this mechanism in each attention block in the transformers $v_{\\theta}$ of a pre-trained 3D generator. Given the latent tokens $\\ell \\in \\mathbb{R}^{L \\times C}$ input to a cross-attention layer as well as the image features $c_{\\mathrm{dino}} \\in \\mathbb{R}^{K \\times C'}$ of the partially visible object, cross-attention computes the similarity score matrix via", + "bbox": [ + 89, + 633, + 483, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} = W _ {q} \\boldsymbol {\\ell}, \\quad [ \\boldsymbol {k}, \\boldsymbol {v} ] = W _ {k v} \\boldsymbol {c} _ {\\mathrm {d i n o}}, \\quad S = \\boldsymbol {q} \\boldsymbol {k} ^ {\\top} / \\sqrt {D}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 779, + 482, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{q}$ and $W_{kv}$ are learnable projections that map the latents $\\ell$ to the query $q$ , and the conditioning image feature $c_{\\mathrm{dino}}$ to the key $k$ and the value $v$ , respectively, while $D$ is the dimension of the query and key vectors.", + "bbox": [ + 89, + 809, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We wish to bias the token similarity matrix $S \\in \\mathbb{R}^{L \\times K}$ towards the visible parts of the object. Recall that $K$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e2d2a57d59bd8efdf730ab223444dac9787b762c1c812448af756fe46115ecca.jpg", + "image_caption": [ + "Figure 4. 3D-consistent mask example. Given a 3D mesh, we render selected triangles in a distinct color from the others to generate multi-view consistent masks. It allows the evaluation of multi-view methods in handling contact occlusion. (The occluded regions are shown in red.)" + ], + "image_footnote": [], + "bbox": [ + 531, + 87, + 581, + 157 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/211eab21b339562a7b9734004b576c27eb64f521933ae646676588415d5a3267.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 89, + 673, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6e1959f584b0d93c348d0d6641a7cac5286305a31fe747f49260119ea0f248a2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 90, + 740, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c82569277c7393f8b8bce254ed494b5e4abb07f88512c3fdd1a8e86563b580f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 90, + 805, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/932a594172e1a1eec2ea793d674805e1bcd98db4e07ceba4702b4ebfc7c48e12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 830, + 90, + 875, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is the number of tokens $c_{\\mathrm{dino}}$ extracted by DINOv2 [52] from the occluded image $x \\cdot M_{\\mathrm{vis}}$ . Each token thus corresponds to a $P \\times P$ patch in the input image (where $P = 14$ ). We extract analogous flattened 1-D tokens $c_{\\mathrm{vis}} = [\\rho(M_{\\mathrm{vis},1}), \\dots, \\rho(M_{\\mathrm{vis},K})]$ from the visibility mask by calculating the fraction of $P \\times P$ pixels that are visible in the $j$ -th image patch $M_{\\mathrm{vis},j}$ . $c_{\\mathrm{vis},j} = \\rho(M_{\\mathrm{vis},j}) \\in [0,1]$ . We then use these quantities to bias the computation of the attention matrix $A = \\text{softmax}(S + \\log c_{\\mathrm{vis}}) \\in [0,1]^{L \\times M}$ via broadcasting. Hence:", + "bbox": [ + 511, + 273, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i j} = \\frac {\\boldsymbol {c} _ {\\text {v i s} , j} \\exp \\left(S _ {i j}\\right)}{\\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {\\text {v i s} , k} \\exp \\left(S _ {i k}\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 439, + 905, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this manner, the visibility flag modulates the attention matrix $A$ smoothly. If there are no visible pixels in a patch $j$ , then $A_{ij} = 0$ , so the corresponding image tokens are skipped in cross-attention. While we have illustrated how this works for a single head, in practice Amodal3R uses a multi-head transformer, to which Eq. (2) extends trivially. Please see the supplementary materials for more details.", + "bbox": [ + 511, + 488, + 905, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Occlusion-aware Attention Layer. To prevent the network from arbitrarily extending the geometry beyond the occluded regions, we further introduce an occlusion-aware attention layer. The key insight is that specifying only visible information is insufficient; it is equally important to distinguish foreground occluders from the background, as this explicitly identifies the potential regions requiring completion. Namely, when a pixel is marked as invisible in the mask $M_{\\mathrm{vis}}$ , this might be (i) because there is an occluder in front of that pixel (so the pixel could plausibly belong to the object of interest if the occlusion were removed), or (ii) because the pixel is entirely off the object. This information is encoded by the mask $M_{\\mathrm{occ}}$ .", + "bbox": [ + 511, + 597, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To help the model distinguish between visible, occluded, and background areas, we add one more cross-attention layer placed immediately after the mask-weighted cross-attention layer, and pool solely the occlusion mask $M_{\\mathrm{occ}}$ . We encode the occlusion mask $M_{\\mathrm{occ}}$ as a sequence of flattened 1-D tokens $c_{\\mathrm{occ}} = [\\rho(M_{\\mathrm{occ},1}), \\ldots, \\rho(M_{\\mathrm{occ},M})]$ , and then simply pool $c_{\\mathrm{occ}}$ using a cross-attention layer.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9184", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Simulating Occluded 3D Data", + "text_level": 1, + "bbox": [ + 89, + 90, + 356, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train and evaluate our model, we require a dataset of 3D assets imaged in scenarios with clutter and occlusions. We curated a large synthetic 3D dataset to train our model, as collecting such 3D assets in the real world is challenging.", + "bbox": [ + 89, + 112, + 480, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Random 2D Occlusions. Each training sample $(o, x, M_{\\mathrm{vis}}, M_{\\mathrm{occ}})$ consists of a 3D object $o$ (from which ground truth latents can be obtained by using the encoders of Appendix A.1), an image $x$ with partial occlusions, and corresponding visibility and occlusion masks $M_{\\mathrm{vis}}$ and $M_{\\mathrm{occ}}$ . In a real scenario, occlusions arise from other objects in the scene. In a multi-view setting, like the one discussed below, these occlusions must be consistent across views, reflecting the underlying scene geometry. However, because our model is trained for single-view reconstruction, we can simulate occlusions by randomly masking parts of the object after rendering it in 2D. This is simpler and allows us to generate fresh occlusion patterns each time a view is sampled for training.", + "bbox": [ + 89, + 172, + 482, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thus, given $o$ and an image $x$ rendered from a random viewpoint, we generate random occlusion masks as follows. Inspired by work on 2D completion [26, 40, 56, 90, 92], we randomly place lines, ellipses, and rectangles, simulating diverse masking patterns. The union of these random shapes gives us the occlusion mask $M_{\\mathrm{occ}}$ . Then, if $M_{\\mathrm{obj}}$ is the mask of the unoccluded object, the visible mask is given by $M_{\\mathrm{vis}} = M_{\\mathrm{obj}} \\odot (1 - M_{\\mathrm{occ}})$ . Examples of such patterns are given in the supplementary material.", + "bbox": [ + 89, + 385, + 482, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D-consistent Occlusions. In a real scene imaged from multiple views, occlusions are not random but geometrically consistent across different viewpoints, as they result from other objects. This is especially true for contact occlusions, where parts of an object remain occluded by other objects from all viewpoints. To evaluate the model's performance under such challenging conditions, 3D-consistent masks are required.", + "bbox": [ + 89, + 521, + 482, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To efficiently generate such contact occlusion masks, we leverage 3D meshes during rendering. Starting from a randomly chosen triangle, we apply a random-walk strategy to iteratively select neighboring triangles, forming continuous occluded regions until the predefined mask ratio is met. By rendering these masked meshes using the same camera parameters as the RGB images, we ensure multi-view consistency in the generated occlusion masks (see Fig. 4).", + "bbox": [ + 89, + 642, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Reconstruction with Multi-view Input", + "text_level": 1, + "bbox": [ + 89, + 772, + 418, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our flow-based model performs multiple denoising steps and does not require known camera poses for input views, Amodal3R can benefit from multi-view reference images $\\mathcal{X} = \\{x_i\\}_{i=1}^N$ as conditioning inputs at different steps of the denoising process. While multi-view conditioning naturally improves reconstruction performance, a potential issue is that an image used earlier in the denoising process", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "has a stronger influence on the final 3D geometry. This is because early denoising steps establish the coarse geometry, whereas later steps refine higher-frequency details [32, 33]. Therefore, we prioritize input images based on their visibility. Specifically, when experimenting with multi-view inputs, we sort the images in proportion to the object visibility $|M_{\\mathrm{vis}}|$ , ensuring that images with higher visibility are used as primary inputs.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 234, + 645, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Settings", + "text_level": 1, + "bbox": [ + 511, + 262, + 705, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Amodal3R is trained on a combination of 3D synthetic datasets: 3D-FUTURE (9,472 objects [17]), ABO (4,485 objects [13]), and HSSD (6,670 objects [31]). Once trained, we first assessed its effectiveness on Toys4K (randomly sampling 1,500 objects [67]) and Google Scanned Objects (GSO) (1,030 objects [16]), which were excluded from the training data for both our model and the baselines. During inference, a 3D-consistent mask occludes each input object, and each view is augmented with additional randomly generated occlusion areas. This ensures that the model cannot simply extract the region to be completed from the occlusion masks. To further assess out-of-domain generalization in practical applications, we also evaluated all models on the 3D scene dataset Replica [68] as well as on in-the-wild images.", + "bbox": [ + 511, + 287, + 903, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. To measure the quality of the rendered images, we use the Fréchet Inception Distance (FID) [21] and the Kernel Inception Distance (KID) [3]. To assess the quality of the reconstructed 3D geometry, we use the Coverage Score (COV), the Point cloud FID (P-FID) [49], and the Minimum Matching Distance (MMD) using the Chamfer Distance (CD). COV measures the diversity of the results, while P-FID and MMD measure the quality of the 3D reconstruction. We also use the CLIP score [57] to evaluate the consistency between each generated and ground-truth object pair.", + "bbox": [ + 511, + 517, + 905, + 683 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. Most image-to-3D generators are trained on fully visible inputs. To ensure fair comparison, we completed the partially visible 2D objects before passing them to 3D generators. In particular, we use pix2gestalt [54], a state-of-the-art 2D amodal completion network. Using this protocol, we compared Amodal3R to state-of-the-art methods such as TRELLIS [80], Real3D [27], GaussianAnything [36]ICLR'25,and LaRa [6]ECCV'24.", + "bbox": [ + 511, + 686, + 903, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Amodal3R is trained on 4 A100 GPUs (40G) for 20K steps with a batch size of 16, taking approximately one day. We select TRELLIS [80] as underlying model because it is the latest 3D generator and is publicly available. However, our method is not specific to this model.", + "bbox": [ + 511, + 809, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9185", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e6a50c551235f4381cb26706eb19bfd086f8d3620ec0937e5dc58f841e641cd2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]92.261.300.7434.6935.925.03
Real3D [27]1pix2gestalt [54]91.212.020.7523.9219.619.21
TRELLIS [80]1pix2gestalt [54]58.825.870.7626.4331.654.17
Amodal3R (Ours)1—30.640.350.817.6939.613.62
LaRa [6]4pix2gestalt [54]172.844.540.7066.3424.568.11
LaRa [6]4pix2gestalt [54]+MV97.532.630.7521.8026.218.61
TRELLIS [80]4pix2gestalt [54]65.696.920.7824.6432.334.26
TRELLIS [80]4pix2gestalt [54]+MV60.371.850.8319.6831.754.21
Amodal3R (Ours)4—26.270.220.845.0338.743.61
", + "bbox": [ + 94, + 88, + 906, + 268 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d2cb8eaebe6634f7469d32f79bd74c40b9f40b6b49a1e79a776b6a95949543f.jpg", + "table_caption": [ + "Table 1. Amodal 3D reconstruction results on GSO [16]. Here, V-num denotes the number of input views, and 2D-Comp refers to the 2D completion method. For single-view image-to-3D, we first complete occluded objects using pix2gestalt [54] before passing them to various 3D models. For multi-view image-to-3D, we adopt two variants: (1) pix2gestalt [54] is applied independently on each view; (2) pix2gestalt $[54] + \\mathrm{MV}$ : a single-view completion is generated first, followed by multi-view diffusion [61] to synthesize 4 views as inputs. Without bells and whistles, Amodal3R consistently outperforms state-of-the-art models across all evaluation metrics." + ], + "table_footnote": [], + "table_body": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]57.171.220.8021.9733.567.23
Real3D [27]1pix2gestalt [54]59.921.630.7923.3124.359.53
TRELLIS [80]1pix2gestalt [54]43.056.830.8026.0426.286.87
Amodal3R (Ours)1—23.450.420.835.0037.095.89
LaRa [6]4pix2gestalt [54]123.523.610.7545.9127.899.67
LaRa [6]4pix2gestalt [54]+MV75.334.140.8013.0024.8210.93
TRELLIS [80]4pix2gestalt [54]46.348.770.8128.7625.357.13
TRELLIS [80]4pix2gestalt [54]+MV43.007.530.8124.4126.557.05
Amodal3R (Ours)4—20.930.500.853.7839.035.75
", + "bbox": [ + 94, + 354, + 906, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Amodal 3D reconstruction results on Toys4K [67]. The experimental setting is the same to Tab. 1, except for the dataset.", + "bbox": [ + 107, + 541, + 885, + 555 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/214c14a2442947413628d675f0dc806eac9a9b619143d2876fd720ddb04267b6.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 98, + 568, + 176, + 670 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d09515060d8f065dc68a83ff275e786b139d6dc6bf88deeda9e4bd95c9b5c0b.jpg", + "image_caption": [ + "pix2gestalt [54]" + ], + "image_footnote": [], + "bbox": [ + 184, + 571, + 241, + 669 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42865b2963f051b0c2bff5ffc35995faa5cab09351613b701818ec6602ab8fd8.jpg", + "image_caption": [ + "Figure 5. Single-view amodal 3D reconstruction. The occlusion regions are shown in black and the visible regions are highlighted with red outlines. More examples are provided in supplementary material Fig. D.4" + ], + "image_footnote": [], + "bbox": [ + 263, + 571, + 330, + 669 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c2763fdf1c66b93cc21539d425558ff93e45d8f13c0c5da52e3e678eff76f7a1.jpg", + "image_caption": [ + "GaussianAnything [36]" + ], + "image_footnote": [], + "bbox": [ + 334, + 571, + 401, + 669 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbdc468da5e748eb5f62a19b180df9909d73e5150ed73d6a08080db0e2afb4ae.jpg", + "image_caption": [ + "Real3D [27]" + ], + "image_footnote": [], + "bbox": [ + 419, + 571, + 482, + 669 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cd07fa7f5d6b40ebaebb2d0c5ae2177b0dedaa1f1b4a8d9716bf88166a01b7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 571, + 553, + 667 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ed7d60ced0c102b5269e611f751e0e04ab271d3a3ece8cefb85e6d32796eafd.jpg", + "image_caption": [ + "TRELLIS [80]" + ], + "image_footnote": [], + "bbox": [ + 586, + 573, + 633, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/379e0e796dc1c2e971bd5e5916cb2f25f540ccfed2e459e2df28035bbc850c2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 573, + 710, + 662 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/837c8a94c76ee900b4119ba1434ef012529765f892c9fef8e3d7a31849774486.jpg", + "image_caption": [ + "Amodal3R (Ours)" + ], + "image_footnote": [], + "bbox": [ + 740, + 570, + 803, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 743, + 228, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. We compare Amodal3R with prior methods on amodal 3D reconstruction in Tabs. 1 and 2. Amodal3R significantly outperforms the baselines across all metrics in both single- and multi-view settings, demonstrating its effectiveness. Notably, two-stage methods using multiple views (\"4 V-num + pix2gestalt\") perform worse than their single-view counterparts (\"1 V-num + pix2gestalt\"). This shows that inconsistent 2D completion", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "degrades 3D reconstruction (Sec. 2). In contrast, Amodal3R avoids this issue by operating directly in 3D space and benefits from additional occluded views as expected.", + "bbox": [ + 511, + 744, + 905, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Qualitative results are shown in Figs. 5 and 6 and in the supplementary material Figs. D.4 and D.5. For all baselines, pix2gestalt is first applied for 2D amodal completion (second column), and the corresponding completed images are passed to each baseline image-to-3D generator. Amodal3R consistently produces high-quality 3D assets under both single- and multi-view settings, even", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9186", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/12012618131c21e5dab9e6f613af87cc1a0dcbc2136a2a7bf79f8f68b8228445.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 197, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1947c7a55137f0e34a70b742a96e9cf6f01a1ef8656f708836959fa26593cdf.jpg", + "image_caption": [ + "pix2gestalt [54]" + ], + "image_footnote": [], + "bbox": [ + 210, + 88, + 279, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45d750e591eb5be7d2bce1b5c60b0ac898451b57f911f0c101146bdfc512c33e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 88, + 387, + 128 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45897ab0ddaa06acd8c18b9318729497506e1506fcced3531a81b2f89c09532a.jpg", + "image_caption": [ + "LaRa [6]" + ], + "image_footnote": [], + "bbox": [ + 310, + 138, + 380, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f199b91890a543943aee5d14751d56d92d35f08a66e6fcfe03090295c0d965b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 95, + 473, + 125 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/169fbf26b83e089ce7fe23f2286a3902a20d903d4affd660041b5ed69ddad98c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 142, + 473, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6554d658abf7a597ba389b01081b18ecf8e82f2bbc113efa95bca9da7f768f63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 90, + 586, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff823111c9b70405d6c45af704cfa0a836289b2b3850deced980ff5166b7e264.jpg", + "image_caption": [ + "TRELLIS [80]" + ], + "image_footnote": [], + "bbox": [ + 514, + 137, + 584, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e531f073587deb24f188aa5a57000be49699f533947ee2fb2bb1b68528d8ccad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 90, + 676, + 131 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f7403491e87d1bd5e14a6ed2d3ea06cafb71daf13387f7a90236d248aeb5ba3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 141, + 679, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6790e50b47b03e37fc50e00599c442571ea9baf83c3f4ff8960de4f09a771123.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 90, + 792, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d214b6b7760f05c0c3c52637bc5ff6baca125c5df226f7fa47a0f85b78a853d8.jpg", + "image_caption": [ + "Amodal3R (Ours)", + "Figure 6. Multi-view amodal 3D reconstruction. The above results are reconstructed using 4 occluded input views. Due to limited space, we present only the best results for LaRa and TRELLIS under the \"pix2gestalt+MV\" setting. We apply 3d-consistent mask and random extended occlusion areas. More examples are provided in supplementary material Fig. D.5." + ], + "image_footnote": [], + "bbox": [ + 715, + 135, + 792, + 185 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1cb5a8c474a13b4076e5c3119e6d82d418f88a69167b4365b548debcdde86cf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 92, + 883, + 133 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7498ce14f93ae2c932cd88d4ba37575e8adf7627855d1c507d76e6f6ee88b53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 136, + 888, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55be6d9d173b001d1b42e5a16a60ea950a5815f07e0e6e1da5032576fc55f759.jpg", + "image_caption": [ + "Input", + "' Naive concatenation", + "Figure 7. Ablation examples. Naive concatenation fails to reconstruct complete shape and appearance. Mask-weighted attention alone extends geometry into background regions, while occlusion-aware attention alone cannot guarantee photorealistic appearance." + ], + "image_footnote": [], + "bbox": [ + 91, + 272, + 160, + 420 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/535de559e9e64624d482c201449869b4f691749929ca466b2ba824ad8392d7dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 166, + 280, + 225, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b03a90ec60dde9fb4c280130c6844cb500af15892bb7bb790a67fde2001adbc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 334, + 227, + 362 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2d24585d24a0d43522a7ee702bde2614e7b12ab0d70c736989d762513479e42c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 375, + 220, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2fd0b4a2bc4cc74d8d9fe4607d26c742c3bdd95e9600d43431d943dd377f8034.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 272, + 310, + 328 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad19ffd4eb033e8daa9763f2619f2d325b35208ea5f8d5c191acee1f40db0150.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 330, + 313, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d3cca3cb212f5ab952596e9352fc814ed08e754a914d31f28c9473b84f92b40.jpg", + "image_caption": [ + "w/only $M_{\\mathrm{vis}}$ attention" + ], + "image_footnote": [], + "bbox": [ + 254, + 377, + 299, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ffa1cfdc0ab7af2637d775a87d8c18e8e912644e517147ce22c31d6bbc67675.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 281, + 385, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/98cd5d400fbaed8d15d5ddd5f788da08cafdf1f1c12993f866dcc5a8feb294c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 335, + 387, + 362 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76b15a1e0fb1d39530419956091de5892730b41beac49a64ece981401612de8f.jpg", + "image_caption": [ + "w/only $M_{\\mathrm{occ}}$ attention" + ], + "image_footnote": [], + "bbox": [ + 338, + 377, + 383, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48ae97f8280fd4eb1d96592a5215d8dc71da9dbf33a861e42ca530b9cccbeec7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 277, + 467, + 323 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6835bb035d609bf6ba38348e36d53a49eac7ab22e4a7c4a8afedaa49f7d85fb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 413, + 334, + 470, + 362 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/959175325ad9a896663bb3b676db8e8cff8d45ca0065f9e5a4fb82de74ddc81c.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 421, + 376, + 462, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in challenging scenarios. In contrast, 2D completion inconsistencies accumulate as more views are added, especially when pix2gestalt is more uncertain, confusing the downstream reconstructor models. For instance, in the giraffe example in Fig. 5, the pix2gestalt completion fails to capture the overall structure of the 3D object, which in turn leads the 3D generator models to misinterpret it as a woodstick-like shape. Instead, Amodal3R reconstructs the 3D geometry and appearance well, maintaining good alignment with the partially visible inputs. These results underscore the effectiveness of Amodal3R and highlight the benefit of integrating occlusion reasoning directly in 3D space, eliminating the reliance on separate, potentially inconsistent, 2D completion stages.", + "bbox": [ + 89, + 527, + 483, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 753, + 243, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted several ablations to study the impact of the different components of Amodal3R and report the results in Tab. 3 and Fig. 7. For these experiments, we test single-view image-to-3D reconstruction on the GSO dataset.", + "bbox": [ + 89, + 777, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Naive Conditioning. We first evaluated a version of the model that still conditions the reconstruction on the visibility and occlusion masks, but without using the modules of Sec. 3.1. Instead, we directly concatenate the tokens $c_{\\mathrm{vis}}$", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ddc4efba93fa58ff72c0ca6e351a7c0231f790f55991c88bc2e393f81a58fd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAppearanceGeometry
FID ↓KID(%) ↓COV(%) ↑MMD(‰) ↓
naive conditioning31.960.4937.963.61
w/ only mask-weighted attention30.530.3836.903.69
w/ only occlusion-aware layer31.770.5740.193.51
full model (Ours)30.640.3539.613.62
", + "bbox": [ + 516, + 273, + 905, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablations on different mask conditioning designs.", + "bbox": [ + 532, + 364, + 883, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to $c_{\\mathrm{dino}}$ to form the condition for the cross-attention layer. The results (first row in Tab. 3 and second column in Fig. 7) show that, while the resulting model can still perform basic completion, the textures in the occluded regions are inconsistent with those in the visible ones, and the reconstructed geometry is inaccurate, e.g. the hole in the shoe.", + "bbox": [ + 511, + 409, + 906, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask-weighted Attention. To evaluate the effectiveness of our proposed mask-weighted attention mechanism, we integrate it into the training while omitting the occlusion-aware layer. The results demonstrate improved visual quality, especially in capturing texture details in the visible areas. However, the geometries exhibit deficiencies, as seen in the problematic shoe in Fig. 7 (first row), and the toy monster with a broken back (third row).", + "bbox": [ + 511, + 500, + 905, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Occlusion-aware Layer. The integration of the occlusion-aware layer improves the geometry both quantitatively and qualitatively. This aligns with our motivation for introducing an additional cross-attention layer, aimed at reconstructing the visible areas via the image-conditioned layer and occluded areas via the subsequent layer. However, the occlusion-aware layer alone still results in unsatisfactory appearances, which again indirectly attests to the effectiveness of the mask-weighted attention mechanism.", + "bbox": [ + 511, + 621, + 905, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Full Model. Consequently, the full model integrating both modules achieves optimal 3D generation results characterized by complete geometry and consistent textures.", + "bbox": [ + 511, + 758, + 903, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Real-World Generation / Completion", + "text_level": 1, + "bbox": [ + 511, + 816, + 834, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Amodal3R demonstrates strong generalization to out-of-domain test examples, primarily because it builds upon a large-scale pre-trained 3D generator and is fine-tuned with diverse categories. To demonstrate this advantage,", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9187", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0aaa4a1e38acdffb89a296b94d52034c83d55d827f052e369128603ee27b9ea3.jpg", + "image_caption": [ + "Occluded input" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 183, + 348 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd9906c266acf5dba296545ee4c74eb83b3bd899f98e2c37e9f02386537e8754.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 104, + 261, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6255e3e76a9b3d590968e813e66631f1bea605af0f2ac52acdc079fadd1eacf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 104, + 344, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fa5c180b999464af1fdc01181a725ffe146219a7762a97543d5a8b1b4ac4c42.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 102, + 483, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f10aee28cacabda7128574d2eae3a69cb0d67b96aa299faf18e5be799b306367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 107, + 606, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/892deecb97e98270146a0d38533fa2d5c34f4035e00f5925d4e7e03b6ff31120.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 106, + 733, + 148 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/234d82c8065bcddf1d2049287d52b9664374f9fface24fcba575e534966411d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 98, + 879, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e36b8a031e27126923f0ced3c2ab4180a4bac07bc33df0bdaf9b49b53757226c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 358, + 181, + 421 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82dea0ecb153627d533b3aac6897890423a40e2c23dd3df859e0e95839deb30d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 181, + 240, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88295939781841cee35d5000c351c08e248212fab62f8c07c3f561f1a3889ca1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 202, + 243, + 250, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc064faa16ef879a9ba1d032fee422fb0ebc947eabfea37da35ad56efcbabf12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 214, + 296, + 243, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ff6bc3ed69e3e2674d70fa394d723bcb8970c492a0659966af6727f07648192.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 181, + 277, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a4314c862580b2aa2fce1ac227cac194689dde71da5b50611a71b7a236e2b41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 181, + 323, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d9057c53bb1aed5ecf98e5fed336de69e42395bf867e96f5ebb0b28e771fcd2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 229, + 480, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8f7dcc9b77e7dd831b84e0fc47252353b7c026800f899e87b8272a5eab1a687e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 291, + 331, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2571264e7593b19c26b1ee6901744dc8e429fa68c88e75dfc330069516431d8e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 282, + 470, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b348933600ad1b2ca1ac0b04a6f11de2ddc290c3c10fe390247968962e7e83e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 178, + 545, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3f47ac1d5f26df7b563f91770e9388b5209e1d152b5e68c1a5b0ecc481442a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 560, + 176, + 602, + 204 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdab1d71ab835dfbf75ed3891d78e1e9dd2884ad01ce645725e44565f26d6905.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 176, + 661, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/779221d84eac13762e3104fc796d1250c707fe75e07cd2bd744cfa1eb962af37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 176, + 720, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3812f3278f3a6124fa65304bb1fb01e03cd8551b3e7f80e5e82ae36ada5851e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 175, + 807, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18cde7d1bc2554bec861d740a3074698dd941324d4d785f080a7db93745fc3bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 176, + 877, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f2e244f019b483441bf1b023c3042682d39627a63eadbcd551087476d6c8165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 495, + 234, + 604, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dc23b4a45ae4854a6f2a8956b0389f2020b1b7c916bbd755a40886f1655b56b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 232, + 725, + 282 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cf392de632adcf2d68ffa9409dd67b0b8a481bd3f1a1614627d0b9a9912e4579.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 295, + 733, + 327 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/70020367524cbc699a41c68ab03082351c1238ec07afadaf4072d8b3414b7190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 226, + 879, + 285 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebdeda1de9d2f58051643f58f0aafffba36868ae1a7ec8e7faf7e5c6f5f6fafe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 294, + 895, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77e1b7449a788711e712931136f03a029ebda3a502be37bc41a9a790c689ba8b.jpg", + "image_caption": [ + "Occluded input", + "Figure 8. Examples on Replica [68] and in-the-wild images. The target objects and the occluders are shown in red and green outlines. More examples are provided in supplementary material Figs. D.7 to D.9." + ], + "image_footnote": [], + "bbox": [ + 96, + 422, + 181, + 484 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ace62434057fe4fb29c912e2103f8da59747ea756046d5a2b9791ef453b560a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 372, + 289, + 419 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/185102de64684adf0ad58d98b905b6426df9d917e537d74a4542a7a1aebc8fc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 426, + 285, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31b127e645d074e2b254bf4a71b0909fa02ecb6743ee5223b33dfb650101491b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 373, + 370, + 420 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/289d9b695e7ba4ba8013fd3ee4cc17b39ff7fafa036e8844ab8e49af2b771168.jpg", + "image_caption": [ + "Novel views" + ], + "image_footnote": [], + "bbox": [ + 295, + 426, + 367, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fe967532a0c24f45888599ae0b356d60e934b6c115bd8f680688d8a313e5cb0.jpg", + "image_caption": [ + "pix2gestalt [54] GaussianAnything [36]", + "Real3D [27]" + ], + "image_footnote": [], + "bbox": [ + 388, + 359, + 537, + 428 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a87f5edfdfea7b2908f2f1152168e5d1f6034398e5b237e129cf4821cc533de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 426, + 457, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2acc76aecb295dfc1a8c335dadb975480f04c941e533914e6d2fc45b1e46d92c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 359, + 540, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c863e58eeb415d78874e23b004d7563bee499fda1fa4df113ed32f1dbb27686d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 426, + 539, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92dfd49f4bbb50ad9c2fb36968106be55a1d8652bb5a0beeaa3cecc05f9b4e9e.jpg", + "image_caption": [ + "TRELLIS [80]" + ], + "image_footnote": [], + "bbox": [ + 571, + 372, + 640, + 419 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3f4adb75a662be7f373e48ff20b88e463399b648cd18d9c0de7665505e87bd4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 372, + 718, + 421 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8446728d4dd3a897b765ea0c0f6f4388fff96bfe0fb74902dea7d8d9730c2ad6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 426, + 640, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd104cfb159e20527df1235e54d67489c2e59aa683914fe2e44f69546062d9eb.jpg", + "image_caption": [ + "Normal maps" + ], + "image_footnote": [], + "bbox": [ + 645, + 426, + 715, + 476 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/599cd74b2b6add55f3c3bd9a79332d182be0b6e97457cdd4bf02904893136add.jpg", + "image_caption": [ + "Amodal3R (Ours)" + ], + "image_footnote": [], + "bbox": [ + 730, + 359, + 879, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d38f75edd1ea05c2b78c64e0e16f7abaa3e2f6bc04b3f720a3b7a564bebf594b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 426, + 879, + 476 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0e9a7a60e563272d71d507f13b6fabfdccb2d9da73d1afa31ffe8f952ec5453d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodP-FID↓COV(%) ↑MMD(‰) ↓
GaussianAnything91.3761.2225.00
Real3D79.5555.1018.28
TRELLIS128.5259.1723.95
Amodal3R61.4659.1810.98
", + "bbox": [ + 104, + 559, + 470, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative results on the Replica dataset [68].", + "bbox": [ + 109, + 661, + 464, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we conducted scene-level evaluations. Qualitative results are shown in Fig. 8, which include COCOA [97] (first two rows), Replica [68] (third and fourth rows), and in-the-wild images (fifth and sixth rows). Here, we used Segment Anything [34] to obtain the visibility and occlusion masks. The results show that Amodal3R generates photorealistic 3D assets, whereas pix2gestalt fails to infer complete shapes from the same inputs, leading to unsatisfactory results from subsequent image-to-3D models. We also visualize the colored normal maps, which show that the results of Amodal3R are simultaneously rich in geometric and textural detail.", + "bbox": [ + 89, + 703, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative results are presented in Tab. 4. As real-world datasets with instance 3D amodal annotations for", + "bbox": [ + 89, + 869, + 483, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "both geometry and appearance are limited, here we focus on geometry evaluation by curating a set of 49 completed 3D instances from the Replica scenes. Furthermore, we used the Umeyama algorithm [73] to match the generated meshes with the ground truth meshes in terms of scale and pose. It can be observed that our Amodal3R is still on par with or better than the baselines on the scene-level dataset.", + "bbox": [ + 511, + 561, + 906, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 679, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced Amodal3R, a feed-forward approach to reconstruct the complete photorealistic 3D assets from partially visible images. The key to our success is that we construct a mask-weighted cross-attention mechanism and an occlusion-aware layer to effectively exploit visible and occluded information. Compared to the state-of-the-art methods that rely on sequential 2D completion followed by 3D generation, our Amodal3R achieves remarkably better performance by operating directly in 3D space. Furthermore, results on in-the-wild images indicate its potential for subsequent applications in 3D decomposition, marking a step towards robust 3D asset reconstruction in real-world environments with complex occlusion.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9188", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). This research is also supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 RG107/24. Dr. Chuanxia Zheng is supported by EPSRC SYN3D EP/Z001811/1.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2", + "[2] Jiayang Ao, Yanbei Jiang, Qiuhong Ke, and Krista A Ehinger. Open-world amodal appearance completion. arXiv preprint arXiv:2411.13019, 2024. 2", + "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 5, 3", + "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 5799-5809, 2021. 2", + "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2", + "[6] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2024. 1, 5, 6, 7", + "[7] Minghao Chen, Roman Shapovalov, Iro Laina, Tom Monnier, Jianyuan Wang, David Novotny, and Andrea Vedaldi. Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3", + "[8] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2", + "[9] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Neural Information Processing Systems (NeurIPS), 2024. 2", + "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2" + ], + "bbox": [ + 93, + 257, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Junhyeong Cho, Kim Youwang, Hunmin Yang, and TaeHyun Oh. Robust 3d shape reconstruction in zero-shot from a single image in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2", + "[12] Ruihang Chu, Enze Xie, Shentong Mo, Zhenguo Li, Matthias Nießner, Chi-Wing Fu, and Jiaya Jia. Diffcomplete: Diffusion-based generative 3d shape completion. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[13] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5", + "[14] Ruikai Cui, Weizhe Liu, Weixuan Sun, Senbo Wang, Taizhang Shang, Yang Li, Xibin Song, Han Yan, Zhennan Wu, Shenzhen Chen, et al. Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2", + "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 2, 5, 6, 3", + "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 5", + "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2", + "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2", + "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 3", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9189", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1", + "[24] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI conference on artificial intelligence, pages 10997-11004, 2020. 2", + "[25] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7662-7670, 2020. 2", + "[26] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (TOG), 36(4):107, 2017. 5", + "[27] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. 5, 6, 8, 2", + "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3", + "[29] Yoni Kasten, Ohad Rahamim, and Gal Chechik. Point cloud completion with pretrained text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1", + "[31] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 5", + "[32] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 5", + "[33] Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, et al. Leveraging early-stage robustness in diffusion models for efficient and high-quality image synthesis. Advances in Neural Information Processing Systems, 36:1229-1244, 2023. 5", + "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 3, 8", + "[35] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 2", + "[36] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 5, 6, 8, 2, 3", + "[37] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12642-12651, 2023. 2", + "[38] Xinhai Li, Huaibin Wang, and Kuo-Kun Tseng. Gaussian diffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221, 2023. 2", + "[39] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2", + "[40] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 5", + "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 9298–9309, 2023. 1, 2", + "[42] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1", + "[43] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 9970-9980, 2024. 2", + "[44] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2837-2845, 2021. 2", + "[45] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 2", + "[46] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4328-4338, 2023. 2", + "[47] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), pages 7220-7229. PMLR, 2020. 1, 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9190", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 7588–7597, 2019. 1, 2", + "[49] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 5, 3", + "[50] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11453-11464, 2021. 2", + "[51] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc V Gool, and Sergey Tulyakov. Autodecoding latent 3d diffusion models. Advances in Neural Information Processing Systems, 36:67021-67047, 2023. 3", + "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 1", + "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 3, 1", + "[54] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3931-3940. IEEE Computer Society, 2024. 1, 2, 3, 5, 6, 7, 8, 4", + "[55] Eunbyung Park, Jimei Yang, Ersin Yumer, Duygu Ceylan, and Alexander C Berg. Transformation-grounded image generation network for novel 3d view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3500-3509, 2017. 2", + "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544. IEEE, 2016. 5", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763. PmLR, 2021. 5, 3", + "[58] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 2", + "[60] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33:20154-20166, 2020. 2", + "[61] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero $123++$ : a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 6, 4", + "[62] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3", + "[63] Wu Shuang, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems, 37:121859–121881, 2025. 3", + "[64] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2", + "[65] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2", + "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning (ICML), pages 2256-2265. pmlr, 2015. 2", + "[67] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798-1808, 2021. 2, 5, 6, 3", + "[68] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 8", + "[69] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1955-1964, 2018. 2", + "[70] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8863-8873, 2023. 1, 2, 3", + "[71] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9191", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1", + "[72] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision, pages 175-191. Springer, 2024. 2", + "[73] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on pattern analysis and machine intelligence, 13(4): 376-380, 2002. 8", + "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3", + "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457. Springer, 2024. 1, 2", + "[76] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2", + "[77] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 1", + "[78] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In European Conference on Computer Vision (ECCV), pages 456-473. Springer, 2024. 2", + "[79] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2", + "[80] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 1, 3, 4, 5, 6, 7, 8, 2", + "[81] Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, Soren Pirk, and Arie E Kaufman. Carve3d: Improving multi-view reconstruction consistency for diffusion models with rl finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6369-6379, 2024. 2", + "[82] Yiheng Xiong and Angela Dai. Pt43d: A probabilistic transformer for generating 3d shapes from single highly-ambiguous rgb images. arXiv preprint arXiv:2405.11914, 2024. 2", + "[83] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Amodal completion via progressive mixed context diffusion. In Proceed-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9109, 2024. 2", + "[84] Bo Yang, Hongkai Wen, Sen Wang, Ronald Clark, Andrew Markham, and Niki Trigoni. 3d object reconstruction from a single depth view with adversarial learning. In Proceedings of the IEEE international conference on computer vision workshops (ICCVW), pages 679-688, 2017. 2", + "[85] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28003-28013, 2024. 2", + "[86] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3784-3792, 2020. 2", + "[87] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2", + "[88] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1", + "[89] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9720-9731, 2024. 1, 2, 3", + "[90] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1438-1447, 2019. 5", + "[91] Chuanxia Zheng, Duy-Son Dao, Guoxian Song, Tat-Jen Cham, and Jianfei Cai. Visiting the invisible: Layer-by-layer completed scene decomposition. International Journal of Computer Vision (IJCV), 129:3195-3215, 2021. 2", + "[92] Chuanxia Zheng, Tat-Jen Cham, Jianfei Cai, and Dinh Phung. Bridging global context interactions for high-fidelity image completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11512-11522, 2022. 4, 5", + "[93] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 2", + "[94] Xin-Yang Zheng, Hao Pan, Yu-Xiao Guo, Xin Tong, and Yang Liu. Mvd^2: Efficient multiview 3d reconstruction for multiview diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2", + "[95] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2" + ], + "bbox": [ + 517, + 92, + 903, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9192", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[96] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. Advances in neural information processing systems (NeurIPS), 31, 2018. 2", + "[97] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1464-1472, 2017. 8", + "[98] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 301 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "9193", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_model.json b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59080fc8f447a0b81e464c9abe9cc1616cb61fc0 --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_model.json @@ -0,0 +1,3900 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.166, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.18, + 0.821, + 0.247 + ], + "angle": 0, + "content": "Tianhao Wu\\(^{1*}\\), Chuanxia Zheng\\(^{2\\dagger}\\), Frank Guan\\(^{3}\\), Andrea Vedaldi\\(^{2}\\), Tat-Jen Cham\\(^{1}\\) \n\\(^{*}\\)S-Lab, \\(^{1}\\)Nanyang Technological University, \\(^{2}\\)VGG, University of Oxford \n\\(^{3}\\)Singapore Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.248, + 0.613, + 0.264 + ], + "angle": 0, + "content": "https://sm0kywu.github.io/Amodal3R/" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.279, + 0.173, + 0.292 + ], + "angle": 0, + "content": "3D Recast" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.293, + 0.323, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.407, + 0.18, + 0.419 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.407, + 0.312, + 0.419 + ], + "angle": 0, + "content": "Stochastic sample 1" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.278, + 0.472, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.407, + 0.462, + 0.419 + ], + "angle": 0, + "content": "Stochastic sample 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.279, + 0.671, + 0.293 + ], + "angle": 0, + "content": "In-the-wild Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.479, + 0.293, + 0.617, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.405, + 0.592, + 0.419 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.287, + 0.758, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.405, + 0.753, + 0.418 + ], + "angle": 0, + "content": "Stochastic sample 1" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.286, + 0.904, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.405, + 0.896, + 0.418 + ], + "angle": 0, + "content": "Stochastic sample 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.422, + 0.273, + 0.436 + ], + "angle": 0, + "content": "3D Scene Decomposition" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.437, + 0.221, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.525, + 0.204, + 0.539 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.437, + 0.446, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.437, + 0.671, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.437, + 0.896, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.525, + 0.636, + 0.538 + ], + "angle": 0, + "content": "Decomposed 3D assets" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.551, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 1. Example results of Amodal3R. Given partially visible objects within images (occluded regions are shown in black, visible areas in red outlines), our Amodal3R generates diverse semantically meaningful 3D assets with reasonable geometry and plausible appearance. We sample multiple times to get diverse results from the same occluded input. Trained on synthetic datasets, it generalizes well to out-of-domain test examples from real scenes, where most objects are partially visible, and reconstructs reasonable 3D assets." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.62, + 0.327, + 0.635 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.641, + 0.486, + 0.881 + ], + "angle": 0, + "content": "Most existing image-to-3D models assume that objects are fully visible, ignoring occlusions that commonly occur in real-world scenarios. In this paper, we introduce Amodal3R, a conditional image-to-3D model designed to reconstruct plausible 3D geometry and appearance from partial observations. We extend a \"foundation\" 3D generator by introducing a visible mask-weighted attention mechanism and an occlusion-aware attention layer that explicitly leverage visible and occlusion priors to guide the reconstruction process. We demonstrate that, by training solely on synthetic data, Amodal3R learns to recover full 3D objects even in the presence of occlusions in real scenes. It substantially outperforms state-of-the-art methods that independently perform 2D amodal completion followed by 3D reconstruction, thereby establishing a new benchmark for occlusion-aware 3D reconstruction." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.19, + 0.9 + ], + "angle": 0, + "content": "†Project Lead." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.62, + 0.645, + 0.635 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.652, + 0.907, + 0.833 + ], + "angle": 0, + "content": "Humans possess a remarkable ability to infer the complete 3D shape and appearance of objects from single views, even when those objects are partly hidden behind occluders. This ability, known as amodal completion, is key to operating in complex real-world environments, where objects are often partially occluded by surrounding clutter. However, existing image-based 3D reconstruction models [5, 6, 23, 41, 47, 48, 65, 70, 71, 75, 77, 80, 88, 89] fail to recover full 3D assets when the object is partially occluded. We thus consider the problem of reconstructing 3D objects from one or more partially occluded views, a novel task that we call amodal 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Amodal 3D reconstruction is a challenging task that requires reconstructing an object's photorealistic 3D geometry and appearance, including the occluded regions, which are inherently ambiguous. Previous approaches [7, 54] typ" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "ically decompose the task into a two-stage pipeline: first performing 2D amodal completion [85, 86, 91], followed by conventional 3D reconstruction [41, 75, 89]. While these two-stage pipelines are straightforward to implement, they have some drawbacks. First, 2D amodal completion methods rely predominantly on appearance-only priors, without incorporating explicit information on the 3D structure. This lack of 3D geometric reasoning limits their ability to generate physically plausible completions. Second, some 3D reconstruction methods depend on, or benefit from, multiview inputs. In this case, 2D amodal completion may lack multi-view consistency, particularly when performed independently for different views, which confuses the 3D reconstruction process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.483, + 0.541 + ], + "angle": 0, + "content": "In this paper, we present Amodal3R, a novel single-stage paradigm for amodal 3D reconstruction that surpasses previous state-of-the-art approaches. Amodal3R fine-tunes a pre-trained 3D generator by introducing a dedicated branch that explicitly models occlusions. Its key advantage is performing reconstruction, completion, and occlusion reasoning directly in a 3D latent space rather than using a two-stage scheme. With this, the model can reconstruct both visible and occluded regions of the object coherently and plausibly. Specifically, we introduce a mask-weighted cross-attention to highlight visible information, along with a new occlusion-aware attention layer to guide completion in occluded areas. Together, these components enable the model to focus more on the visible parts of the object without significantly perturbing the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.671 + ], + "angle": 0, + "content": "We designed several benchmarks to evaluate the performance of Amodal3R, including object-level Google Scanned Objects [16] and Toys4K [67] augmented with occlusions, 3D scenes from Replica [68], and real-world monocular images. Without relying on additional heuristics, Amodal3R achieves significantly superior performance compared to state-of-the-art models and generalizes well to different datasets, including real ones." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.68, + 0.449, + 0.695 + ], + "angle": 0, + "content": "In summary, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.705, + 0.483, + 0.749 + ], + "angle": 0, + "content": "- We propose a feed-forward 3D reconstruction model that directly reconstructs complete and high-quality 3D objects from one or more partially occluded views." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- We introduce a mask-weighted cross-attention mechanism and an occlusion-aware layer to inject occlusion awareness into an existing 3D reconstruction model, improving both the geometry and appearance of reconstructed objects when they are partially occluded." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- We demonstrate, via experiments on the 3D scene dataset Replica and real-world images, that our one-stage pipeline significantly outperforms existing two-stage approaches, establishing a new benchmark for amodal 3D reconstruction." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.705, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.642, + 0.108 + ], + "angle": 0, + "content": "2. Background" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.431 + ], + "angle": 0, + "content": "2D Amodal Completion. Recent advances in 2D amodal completion [2, 54, 83, 85, 86, 91] have achieved significant success in reconstructing occluded regions of objects in 2D images. While these methods show promise for 3D generation pipelines [2, 54], they still have limitations. First, 2D amodal completion models have limited 3D geometric priors, as they interpret images as 2D patterns. While they excel at completing textures, they may generate physically implausible geometries when handling complex occlusions. This stems from their lack of explicit 3D reasoning and reliance on 2D appearance priors, without true volumetric understanding. Second, for models that use multi-view images as input, the results of 2D amodal completion are often inconsistent across views. Inconsistent views need to be handled by the 3D reconstructors, which causes confusion (see Sec. 4). Although there has been significant work on multiview consistent generation [7, 62, 70, 72, 81, 89, 94], multiview consistent completion is less explored. RenderDiffusion [1] contains an example, but the results are often blurry or lack details. This motivates our 3D-centric reconstruction framework." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.674 + ], + "angle": 0, + "content": "3D Shape Completion. Several methods [12, 14, 69, 95] start from a partial 3D reconstruction, then complete it in 3D. They use encoder-decoder architectures [24] or volumetric representations [10, 29] to robustly recover 3D shape, but not texture, which is necessary in many applications. They also still require recovering the partial 3D geometry from an occluded image, a challenge in itself. Furthermore, they ignore the input image when completing the object in 3D, disregarding the occlusion pattern that caused the 3D geometry to be recovered only partially. Recent work [11, 82] utilizes the partially visible object as input specifically for 3D shape completion. However, it focuses solely on geometry reconstruction, without recovering the object's texture and appearance. In contrast, our approach is end-to-end, recovering the complete 3D shape and appearance of the object from the occluded image." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3D Generative Models. Early advancements in 3D generation are based on GANs [19], exploring various 3D representations such as point clouds [25, 39], voxel grids [84, 96], view sets [48, 55], NeRF [4, 5, 50, 60], SDF [18], and 3D Gaussian mixtures [78]. While GANs can capture complex 3D structures, they struggle with stability, scalability, and data diversity. The focus then shifted to diffusion models [22, 59, 66], which are more capable and versatile. They were first applied to novel view synthesis [76] in image space, before expanding to model a variety of 3D representations, including point clouds [44, 45, 49, 79, 93], voxel grids [37, 46], triplanes [64, 98], meshes [20, 43], and 3D Gaussian mixtures [9, 35, 38, 87]. Autoregressive models [8, 47, 65] have also been explored for mesh generation, focusing on improving the topology of the final 3D mesh." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9182" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.898, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.228, + 0.9, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Figure 2. Overview of Amodal3R. Given an image as input and the regions of interest, Amodal3R first extracts the partially visible target object, along with the visibility and occlusion masks using an off-the-shelf 2D segmenter. It then applies DINOV2 [53] to extract features \\( c_{\\mathrm{dino}} \\) as additional conditioning for the 3D reconstructor. To enhance occlusion reasoning, each transformer block incorporates a mask-weighted cross-attention (via \\( c_{\\mathrm{vis}} \\)) and occlusion-aware attention layer (via \\( c_{\\mathrm{occ}} \\)), ensuring the 3D reconstructor accurately perceives visible information while effectively inferring occluded parts. For conditioning details, see Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.456, + 0.482, + 0.624 + ], + "angle": 0, + "content": "An alternative to performing diffusion in 2D image space [62, 70, 89] is to use 3D latent spaces [28, 51, 58, 63, 74, 80]. Such methods typically consist of two stages: the first for generating geometry and the second for generating appearance, and are capable of high-quality imaged-to-3D generation. However, they assume that input objects are fully visible, which limits their application to real-world scenes, where occlusions are frequent. Here, we extend such approaches to generate high-quality 3D assets from occluded input images — a challenging task that requires inferring complete 3D objects from partial information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.184, + 0.649 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Consider an image \\( x \\) containing a partially occluded view of an object of interest \\( o_i \\). We wish to reconstruct the complete 3D shape and appearance \\( y \\) of the object \\( o_i \\). This task is conceptually similar to existing image-to-3D reconstruction, except that here the object is partially occluded instead of being fully visible in \\( x \\). We call this problem amodal 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We introduce Amodal3R (Fig. 2), a new method for amodal 3D reconstruction. Unlike recent two-stage methods [7, 54] that first perform 2D amodal completion followed by 3D reconstruction, Amodal3R is an end-to-end occlusion-aware 3D reconstruction model that generates complete geometry and appearance directly within the 3D space. Formally, Amodal3R is a conditional generator \\( v_{\\theta}(y|o_i,M_{\\mathrm{vis}},M_{\\mathrm{occ}}) \\) that takes as input the image \\( x \\), centered on a partially visible object \\( o_i \\), as well as the visibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.697 + ], + "angle": 0, + "content": "mask \\( M_{\\mathrm{vis}} \\) and occlusion mask \\( M_{\\mathrm{occ}} \\). The visibility mask \\( M_{\\mathrm{vis}} \\) highlights the pixels in \\( x \\) that correspond to object \\( o_i \\), while \\( M_{\\mathrm{occ}} \\) marks the pixels occupied by occluders, i.e., all other objects that potentially obscure \\( o_i \\). For real images, these masks can be efficiently obtained using pre-trained 2D segmentation models like Segment Anything [34]. By providing point coordinate prompts for the object of interest \\( o_i \\) and its occluders, the segmentation model can generate the corresponding masks. In cases where multiple occluders are present, or when occluders fragment the visible area of the target object, a sequential process is employed. Specifically, point prompts for each visible/occluding region are provided to the 2D segmenter one at a time, with the model generating an individual mask for each region. These masks are then aggregated to form comprehensive visibility and occlusion masks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.909, + 0.79 + ], + "angle": 0, + "content": "The challenges for Amodal3R are how to: (1) produce a plausible and complete 3D shape and appearance from partial observations, even in the absence of multi-view information; and (2) ensure 3D consistency in terms of geometry and photometry, maintaining seamless visual coherence between visible and generated regions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.842, + 0.818 + ], + "angle": 0, + "content": "3.1. Mask-Conditional Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To fine-tune an image-to-3D model for partially occluded inputs, the key change is to condition the transformer \\( v_{\\theta} \\) not only on the object of interest \\( o_i \\), but also on the visibility and occlusion masks, i.e., \\( M_{\\mathrm{vis}} \\) and \\( M_{\\mathrm{occ}} \\). A naive approach is to patchify or embed the masks into tokens \\( (c_{\\mathrm{vis}}, c_{\\mathrm{occ}}) \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9183" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.245, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.308, + 0.22, + 0.32 + ], + "angle": 0, + "content": "(a) TRELLIS [80]" + }, + { + "type": "image", + "bbox": [ + 0.25, + 0.089, + 0.481, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.308, + 0.43, + 0.32 + ], + "angle": 0, + "content": "(b) Amodal3R (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.381 + ], + "angle": 0, + "content": "Figure 3. The transformer structure of Amodal3R. Compared with the original TRELLIS [80] design, we further introduce the mask-weighted cross-attention and occlusion-aware layer. It applies to both sparse structure and SLAT diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.543 + ], + "angle": 0, + "content": "which can be concatenated to the image tokens \\( \\pmb{e}_{\\mathrm{dino}} \\) and processed by cross-attention as before. However, the RGB \\( \\pmb{o}_i \\) contains significantly more information than the binary masks \\( M_{\\mathrm{vis}} \\) and \\( M_{\\mathrm{occ}} \\). The transformer, which is initially trained with RGB only, may simply ignore the mask information. This is compounded by the fact that learning to use this new information, which involves a new type of token that is incompatible with image tokens, may require aggressive fine-tuning of the transformer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.484, + 0.634 + ], + "angle": 0, + "content": "To sidestep this problem, inspired by masked attention approaches in 2D completion [92], we introduce mask-weighted cross-attention and an occlusion-aware attention layer to better utilize the visibility mask \\( M_{\\mathrm{vis}} \\) and the occlusion mask \\( M_{\\mathrm{occ}} \\) without disrupting the pre-trained 3D generator. These are described next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.77 + ], + "angle": 0, + "content": "Mask-weighted Cross-Attention. A key novel component of Amodal3R is mask-weighted cross-attention, which allows the model to focus its attention on the visible parts of the object. We implement this mechanism in each attention block in the transformers \\( v_{\\theta} \\) of a pre-trained 3D generator. Given the latent tokens \\( \\ell \\in \\mathbb{R}^{L \\times C} \\) input to a cross-attention layer as well as the image features \\( c_{\\mathrm{dino}} \\in \\mathbb{R}^{K \\times C'} \\) of the partially visible object, cross-attention computes the similarity score matrix via" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.78, + 0.483, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} = W _ {q} \\boldsymbol {\\ell}, \\quad [ \\boldsymbol {k}, \\boldsymbol {v} ] = W _ {k v} \\boldsymbol {c} _ {\\mathrm {d i n o}}, \\quad S = \\boldsymbol {q} \\boldsymbol {k} ^ {\\top} / \\sqrt {D}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\( W_{q} \\) and \\( W_{kv} \\) are learnable projections that map the latents \\( \\ell \\) to the query \\( q \\), and the conditioning image feature \\( c_{\\mathrm{dino}} \\) to the key \\( k \\) and the value \\( v \\), respectively, while \\( D \\) is the dimension of the query and key vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We wish to bias the token similarity matrix \\( S \\in \\mathbb{R}^{L \\times K} \\) towards the visible parts of the object. Recall that \\( K \\)" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.088, + 0.583, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.09, + 0.674, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.091, + 0.741, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.091, + 0.806, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.831, + 0.092, + 0.877, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.174, + 0.907, + 0.244 + ], + "angle": 0, + "content": "Figure 4. 3D-consistent mask example. Given a 3D mesh, we render selected triangles in a distinct color from the others to generate multi-view consistent masks. It allows the evaluation of multi-view methods in handling contact occlusion. (The occluded regions are shown in red.)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.427 + ], + "angle": 0, + "content": "is the number of tokens \\( c_{\\mathrm{dino}} \\) extracted by DINOv2 [52] from the occluded image \\( x \\cdot M_{\\mathrm{vis}} \\). Each token thus corresponds to a \\( P \\times P \\) patch in the input image (where \\( P = 14 \\)). We extract analogous flattened 1-D tokens \\( c_{\\mathrm{vis}} = [\\rho(M_{\\mathrm{vis},1}), \\dots, \\rho(M_{\\mathrm{vis},K})] \\) from the visibility mask by calculating the fraction of \\( P \\times P \\) pixels that are visible in the \\( j \\)-th image patch \\( M_{\\mathrm{vis},j} \\). \\( c_{\\mathrm{vis},j} = \\rho(M_{\\mathrm{vis},j}) \\in [0,1] \\). We then use these quantities to bias the computation of the attention matrix \\( A = \\text{softmax}(S + \\log c_{\\mathrm{vis}}) \\in [0,1]^{L \\times M} \\) via broadcasting. Hence:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.44, + 0.906, + 0.477 + ], + "angle": 0, + "content": "\\[\nA _ {i j} = \\frac {\\boldsymbol {c} _ {\\text {v i s} , j} \\exp \\left(S _ {i j}\\right)}{\\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {\\text {v i s} , k} \\exp \\left(S _ {i k}\\right)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.595 + ], + "angle": 0, + "content": "In this manner, the visibility flag modulates the attention matrix \\( A \\) smoothly. If there are no visible pixels in a patch \\( j \\), then \\( A_{ij} = 0 \\), so the corresponding image tokens are skipped in cross-attention. While we have illustrated how this works for a single head, in practice Amodal3R uses a multi-head transformer, to which Eq. (2) extends trivially. Please see the supplementary materials for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.793 + ], + "angle": 0, + "content": "Occlusion-aware Attention Layer. To prevent the network from arbitrarily extending the geometry beyond the occluded regions, we further introduce an occlusion-aware attention layer. The key insight is that specifying only visible information is insufficient; it is equally important to distinguish foreground occluders from the background, as this explicitly identifies the potential regions requiring completion. Namely, when a pixel is marked as invisible in the mask \\( M_{\\mathrm{vis}} \\), this might be (i) because there is an occluder in front of that pixel (so the pixel could plausibly belong to the object of interest if the occlusion were removed), or (ii) because the pixel is entirely off the object. This information is encoded by the mask \\( M_{\\mathrm{occ}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To help the model distinguish between visible, occluded, and background areas, we add one more cross-attention layer placed immediately after the mask-weighted cross-attention layer, and pool solely the occlusion mask \\( M_{\\mathrm{occ}} \\). We encode the occlusion mask \\( M_{\\mathrm{occ}} \\) as a sequence of flattened 1-D tokens \\( c_{\\mathrm{occ}} = [\\rho(M_{\\mathrm{occ},1}), \\ldots, \\rho(M_{\\mathrm{occ},M})] \\), and then simply pool \\( c_{\\mathrm{occ}} \\) using a cross-attention layer." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9184" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.357, + 0.108 + ], + "angle": 0, + "content": "3.2. Simulating Occluded 3D Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.482, + 0.174 + ], + "angle": 0, + "content": "To train and evaluate our model, we require a dataset of 3D assets imaged in scenarios with clutter and occlusions. We curated a large synthetic 3D dataset to train our model, as collecting such 3D assets in the real world is challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.385 + ], + "angle": 0, + "content": "Random 2D Occlusions. Each training sample \\((o, x, M_{\\mathrm{vis}}, M_{\\mathrm{occ}})\\) consists of a 3D object \\(o\\) (from which ground truth latents can be obtained by using the encoders of Appendix A.1), an image \\(x\\) with partial occlusions, and corresponding visibility and occlusion masks \\(M_{\\mathrm{vis}}\\) and \\(M_{\\mathrm{occ}}\\). In a real scenario, occlusions arise from other objects in the scene. In a multi-view setting, like the one discussed below, these occlusions must be consistent across views, reflecting the underlying scene geometry. However, because our model is trained for single-view reconstruction, we can simulate occlusions by randomly masking parts of the object after rendering it in 2D. This is simpler and allows us to generate fresh occlusion patterns each time a view is sampled for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.521 + ], + "angle": 0, + "content": "Thus, given \\( o \\) and an image \\( x \\) rendered from a random viewpoint, we generate random occlusion masks as follows. Inspired by work on 2D completion [26, 40, 56, 90, 92], we randomly place lines, ellipses, and rectangles, simulating diverse masking patterns. The union of these random shapes gives us the occlusion mask \\( M_{\\mathrm{occ}} \\). Then, if \\( M_{\\mathrm{obj}} \\) is the mask of the unoccluded object, the visible mask is given by \\( M_{\\mathrm{vis}} = M_{\\mathrm{obj}} \\odot (1 - M_{\\mathrm{occ}}) \\). Examples of such patterns are given in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.643 + ], + "angle": 0, + "content": "3D-consistent Occlusions. In a real scene imaged from multiple views, occlusions are not random but geometrically consistent across different viewpoints, as they result from other objects. This is especially true for contact occlusions, where parts of an object remain occluded by other objects from all viewpoints. To evaluate the model's performance under such challenging conditions, 3D-consistent masks are required." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.764 + ], + "angle": 0, + "content": "To efficiently generate such contact occlusion masks, we leverage 3D meshes during rendering. Starting from a randomly chosen triangle, we apply a random-walk strategy to iteratively select neighboring triangles, forming continuous occluded regions until the predefined mask ratio is met. By rendering these masked meshes using the same camera parameters as the RGB images, we ensure multi-view consistency in the generated occlusion masks (see Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.419, + 0.79 + ], + "angle": 0, + "content": "3.3. Reconstruction with Multi-view Input" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Since our flow-based model performs multiple denoising steps and does not require known camera poses for input views, Amodal3R can benefit from multi-view reference images \\(\\mathcal{X} = \\{x_i\\}_{i=1}^N\\) as conditioning inputs at different steps of the denoising process. While multi-view conditioning naturally improves reconstruction performance, a potential issue is that an image used earlier in the denoising process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "has a stronger influence on the final 3D geometry. This is because early denoising steps establish the coarse geometry, whereas later steps refine higher-frequency details [32, 33]. Therefore, we prioritize input images based on their visibility. Specifically, when experimenting with multi-view inputs, we sort the images in proportion to the object visibility \\( |M_{\\mathrm{vis}}| \\), ensuring that images with higher visibility are used as primary inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.235, + 0.646, + 0.253 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.263, + 0.706, + 0.28 + ], + "angle": 0, + "content": "4.1. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.515 + ], + "angle": 0, + "content": "Datasets. Amodal3R is trained on a combination of 3D synthetic datasets: 3D-FUTURE (9,472 objects [17]), ABO (4,485 objects [13]), and HSSD (6,670 objects [31]). Once trained, we first assessed its effectiveness on Toys4K (randomly sampling 1,500 objects [67]) and Google Scanned Objects (GSO) (1,030 objects [16]), which were excluded from the training data for both our model and the baselines. During inference, a 3D-consistent mask occludes each input object, and each view is augmented with additional randomly generated occlusion areas. This ensures that the model cannot simply extract the region to be completed from the occlusion masks. To further assess out-of-domain generalization in practical applications, we also evaluated all models on the 3D scene dataset Replica [68] as well as on in-the-wild images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.906, + 0.684 + ], + "angle": 0, + "content": "Metrics. To measure the quality of the rendered images, we use the Fréchet Inception Distance (FID) [21] and the Kernel Inception Distance (KID) [3]. To assess the quality of the reconstructed 3D geometry, we use the Coverage Score (COV), the Point cloud FID (P-FID) [49], and the Minimum Matching Distance (MMD) using the Chamfer Distance (CD). COV measures the diversity of the results, while P-FID and MMD measure the quality of the 3D reconstruction. We also use the CLIP score [57] to evaluate the consistency between each generated and ground-truth object pair." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.905, + 0.808 + ], + "angle": 0, + "content": "Baselines. Most image-to-3D generators are trained on fully visible inputs. To ensure fair comparison, we completed the partially visible 2D objects before passing them to 3D generators. In particular, we use pix2gestalt [54], a state-of-the-art 2D amodal completion network. Using this protocol, we compared Amodal3R to state-of-the-art methods such as TRELLIS [80], Real3D [27], GaussianAnything [36]ICLR'25,and LaRa [6]ECCV'24." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Implementation Details. Amodal3R is trained on 4 A100 GPUs (40G) for 20K steps with a batch size of 16, taking approximately one day. We select TRELLIS [80] as underlying model because it is the latest 3D generator and is publicly available. However, our method is not specific to this model." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9185" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.269 + ], + "angle": 0, + "content": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]92.261.300.7434.6935.925.03
Real3D [27]1pix2gestalt [54]91.212.020.7523.9219.619.21
TRELLIS [80]1pix2gestalt [54]58.825.870.7626.4331.654.17
Amodal3R (Ours)1—30.640.350.817.6939.613.62
LaRa [6]4pix2gestalt [54]172.844.540.7066.3424.568.11
LaRa [6]4pix2gestalt [54]+MV97.532.630.7521.8026.218.61
TRELLIS [80]4pix2gestalt [54]65.696.920.7824.6432.334.26
TRELLIS [80]4pix2gestalt [54]+MV60.371.850.8319.6831.754.21
Amodal3R (Ours)4—26.270.220.845.0338.743.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.273, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Table 1. Amodal 3D reconstruction results on GSO [16]. Here, V-num denotes the number of input views, and 2D-Comp refers to the 2D completion method. For single-view image-to-3D, we first complete occluded objects using pix2gestalt [54] before passing them to various 3D models. For multi-view image-to-3D, we adopt two variants: (1) pix2gestalt [54] is applied independently on each view; (2) pix2gestalt \\([54] + \\mathrm{MV}\\): a single-view completion is generated first, followed by multi-view diffusion [61] to synthesize 4 views as inputs. Without bells and whistles, Amodal3R consistently outperforms state-of-the-art models across all evaluation metrics." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.356, + 0.907, + 0.538 + ], + "angle": 0, + "content": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]57.171.220.8021.9733.567.23
Real3D [27]1pix2gestalt [54]59.921.630.7923.3124.359.53
TRELLIS [80]1pix2gestalt [54]43.056.830.8026.0426.286.87
Amodal3R (Ours)1—23.450.420.835.0037.095.89
LaRa [6]4pix2gestalt [54]123.523.610.7545.9127.899.67
LaRa [6]4pix2gestalt [54]+MV75.334.140.8013.0024.8210.93
TRELLIS [80]4pix2gestalt [54]46.348.770.8128.7625.357.13
TRELLIS [80]4pix2gestalt [54]+MV43.007.530.8124.4126.557.05
Amodal3R (Ours)4—20.930.500.853.7839.035.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.542, + 0.887, + 0.556 + ], + "angle": 0, + "content": "Table 2. Amodal 3D reconstruction results on Toys4K [67]. The experimental setting is the same to Tab. 1, except for the dataset." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.569, + 0.177, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.673, + 0.16, + 0.685 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.185, + 0.572, + 0.243, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.673, + 0.263, + 0.685 + ], + "angle": 0, + "content": "pix2gestalt [54]" + }, + { + "type": "image", + "bbox": [ + 0.264, + 0.572, + 0.331, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.673, + 0.4, + 0.685 + ], + "angle": 0, + "content": "GaussianAnything [36]" + }, + { + "type": "image", + "bbox": [ + 0.335, + 0.573, + 0.402, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.572, + 0.483, + 0.67 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.672, + 0.528, + 0.684 + ], + "angle": 0, + "content": "Real3D [27]" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.573, + 0.555, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.574, + 0.635, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.574, + 0.711, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.672, + 0.693, + 0.684 + ], + "angle": 0, + "content": "TRELLIS [80]" + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.571, + 0.805, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.672, + 0.853, + 0.684 + ], + "angle": 0, + "content": "Amodal3R (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.69, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 5. Single-view amodal 3D reconstruction. The occlusion regions are shown in black and the visible regions are highlighted with red outlines. More examples are provided in supplementary material Fig. D.4" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.744, + 0.229, + 0.758 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We compare Amodal3R with prior methods on amodal 3D reconstruction in Tabs. 1 and 2. Amodal3R significantly outperforms the baselines across all metrics in both single- and multi-view settings, demonstrating its effectiveness. Notably, two-stage methods using multiple views (\"4 V-num + pix2gestalt\") perform worse than their single-view counterparts (\"1 V-num + pix2gestalt\"). This shows that inconsistent 2D completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.906, + 0.79 + ], + "angle": 0, + "content": "degrades 3D reconstruction (Sec. 2). In contrast, Amodal3R avoids this issue by operating directly in 3D space and benefits from additional occluded views as expected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results. Qualitative results are shown in Figs. 5 and 6 and in the supplementary material Figs. D.4 and D.5. For all baselines, pix2gestalt is first applied for 2D amodal completion (second column), and the corresponding completed images are passed to each baseline image-to-3D generator. Amodal3R consistently produces high-quality 3D assets under both single- and multi-view settings, even" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9186" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.198, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.191, + 0.16, + 0.202 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.089, + 0.281, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.203, + 0.191, + 0.287, + 0.203 + ], + "angle": 0, + "content": "pix2gestalt [54]" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.089, + 0.388, + 0.129 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.139, + 0.382, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.191, + 0.423, + 0.203 + ], + "angle": 0, + "content": "LaRa [6]" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.096, + 0.474, + 0.126 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.143, + 0.474, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.091, + 0.588, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.138, + 0.586, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.091, + 0.677, + 0.132 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.142, + 0.681, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.191, + 0.641, + 0.203 + ], + "angle": 0, + "content": "TRELLIS [80]" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.091, + 0.793, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.136, + 0.793, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.817, + 0.093, + 0.885, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.137, + 0.89, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.191, + 0.853, + 0.202 + ], + "angle": 0, + "content": "Amodal3R (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.907, + 0.25 + ], + "angle": 0, + "content": "Figure 6. Multi-view amodal 3D reconstruction. The above results are reconstructed using 4 occluded input views. Due to limited space, we present only the best results for LaRa and TRELLIS under the \"pix2gestalt+MV\" setting. We apply 3d-consistent mask and random extended occlusion areas. More examples are provided in supplementary material Fig. D.5." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.273, + 0.161, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.424, + 0.135, + 0.435 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.281, + 0.227, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.335, + 0.228, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.376, + 0.221, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.417, + 0.232, + 0.437 + ], + "angle": 0, + "content": "' Naive concatenation" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.273, + 0.311, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.332, + 0.314, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.378, + 0.3, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.415, + 0.309, + 0.437 + ], + "angle": 0, + "content": "w/only \\(M_{\\mathrm{vis}}\\) attention" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.282, + 0.387, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.336, + 0.388, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.378, + 0.385, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.416, + 0.394, + 0.437 + ], + "angle": 0, + "content": "w/only \\(M_{\\mathrm{occ}}\\) attention" + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.279, + 0.468, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.415, + 0.335, + 0.471, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.377, + 0.464, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.417, + 0.455, + 0.437 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.499 + ], + "angle": 0, + "content": "Figure 7. Ablation examples. Naive concatenation fails to reconstruct complete shape and appearance. Mask-weighted attention alone extends geometry into background regions, while occlusion-aware attention alone cannot guarantee photorealistic appearance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.741 + ], + "angle": 0, + "content": "in challenging scenarios. In contrast, 2D completion inconsistencies accumulate as more views are added, especially when pix2gestalt is more uncertain, confusing the downstream reconstructor models. For instance, in the giraffe example in Fig. 5, the pix2gestalt completion fails to capture the overall structure of the 3D object, which in turn leads the 3D generator models to misinterpret it as a woodstick-like shape. Instead, Amodal3R reconstructs the 3D geometry and appearance well, maintaining good alignment with the partially visible inputs. These results underscore the effectiveness of Amodal3R and highlight the benefit of integrating occlusion reasoning directly in 3D space, eliminating the reliance on separate, potentially inconsistent, 2D completion stages." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.244, + 0.77 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.839 + ], + "angle": 0, + "content": "We conducted several ablations to study the impact of the different components of Amodal3R and report the results in Tab. 3 and Fig. 7. For these experiments, we test single-view image-to-3D reconstruction on the GSO dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Naive Conditioning. We first evaluated a version of the model that still conditions the reconstruction on the visibility and occlusion masks, but without using the modules of Sec. 3.1. Instead, we directly concatenate the tokens \\( c_{\\mathrm{vis}} \\)" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.36 + ], + "angle": 0, + "content": "
MethodAppearanceGeometry
FID ↓KID(%) ↓COV(%) ↑MMD(‰) ↓
naive conditioning31.960.4937.963.61
w/ only mask-weighted attention30.530.3836.903.69
w/ only occlusion-aware layer31.770.5740.193.51
full model (Ours)30.640.3539.613.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.533, + 0.366, + 0.885, + 0.38 + ], + "angle": 0, + "content": "Table 3. Ablations on different mask conditioning designs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.499 + ], + "angle": 0, + "content": "to \\( c_{\\mathrm{dino}} \\) to form the condition for the cross-attention layer. The results (first row in Tab. 3 and second column in Fig. 7) show that, while the resulting model can still perform basic completion, the textures in the occluded regions are inconsistent with those in the visible ones, and the reconstructed geometry is inaccurate, e.g. the hole in the shoe." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Mask-weighted Attention. To evaluate the effectiveness of our proposed mask-weighted attention mechanism, we integrate it into the training while omitting the occlusion-aware layer. The results demonstrate improved visual quality, especially in capturing texture details in the visible areas. However, the geometries exhibit deficiencies, as seen in the problematic shoe in Fig. 7 (first row), and the toy monster with a broken back (third row)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.906, + 0.758 + ], + "angle": 0, + "content": "Occlusion-aware Layer. The integration of the occlusion-aware layer improves the geometry both quantitatively and qualitatively. This aligns with our motivation for introducing an additional cross-attention layer, aimed at reconstructing the visible areas via the image-conditioned layer and occluded areas via the subsequent layer. However, the occlusion-aware layer alone still results in unsatisfactory appearances, which again indirectly attests to the effectiveness of the mask-weighted attention mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.905, + 0.805 + ], + "angle": 0, + "content": "Full Model. Consequently, the full model integrating both modules achieves optimal 3D generation results characterized by complete geometry and consistent textures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.836, + 0.834 + ], + "angle": 0, + "content": "4.4. Real-World Generation / Completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Amodal3R demonstrates strong generalization to out-of-domain test examples, primarily because it builds upon a large-scale pre-trained 3D generator and is fine-tuned with diverse categories. To demonstrate this advantage," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9187" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.184, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.349, + 0.228, + 0.36 + ], + "angle": 0, + "content": "Occluded input" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.106, + 0.262, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.106, + 0.346, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.103, + 0.485, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.108, + 0.607, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.107, + 0.735, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.099, + 0.88, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.359, + 0.183, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.183, + 0.241, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.204, + 0.244, + 0.25, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.215, + 0.297, + 0.245, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.183, + 0.278, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.183, + 0.325, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.231, + 0.481, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.292, + 0.332, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.284, + 0.471, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.179, + 0.547, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.178, + 0.603, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.178, + 0.662, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.178, + 0.721, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.176, + 0.808, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.178, + 0.879, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.496, + 0.236, + 0.606, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.233, + 0.726, + 0.283 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.296, + 0.734, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.227, + 0.88, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.295, + 0.896, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.349, + 0.482, + 0.361 + ], + "angle": 0, + "content": "pix2gestalt [54] GaussianAnything [36]" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.423, + 0.183, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.488, + 0.18, + 0.5 + ], + "angle": 0, + "content": "Occluded input" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.373, + 0.29, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.428, + 0.286, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.374, + 0.371, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.427, + 0.368, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.361, + 0.539, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.428, + 0.459, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.361, + 0.541, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.427, + 0.54, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.488, + 0.41, + 0.499 + ], + "angle": 0, + "content": "Novel views" + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.348, + 0.583, + 0.36 + ], + "angle": 0, + "content": "Real3D [27]" + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.348, + 0.714, + 0.36 + ], + "angle": 0, + "content": "TRELLIS [80]" + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.348, + 0.862, + 0.36 + ], + "angle": 0, + "content": "Amodal3R (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.373, + 0.642, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.373, + 0.719, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.428, + 0.642, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.427, + 0.716, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.36, + 0.88, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.428, + 0.88, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.488, + 0.756, + 0.5 + ], + "angle": 0, + "content": "Normal maps" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.507, + 0.905, + 0.536 + ], + "angle": 0, + "content": "Figure 8. Examples on Replica [68] and in-the-wild images. The target objects and the occluders are shown in red and green outlines. More examples are provided in supplementary material Figs. D.7 to D.9." + }, + { + "type": "table", + "bbox": [ + 0.105, + 0.56, + 0.472, + 0.652 + ], + "angle": 0, + "content": "
MethodP-FID↓COV(%) ↑MMD(‰) ↓
GaussianAnything91.3761.2225.00
Real3D79.5555.1018.28
TRELLIS128.5259.1723.95
Amodal3R61.4659.1810.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.662, + 0.465, + 0.677 + ], + "angle": 0, + "content": "Table 4. Quantitative results on the Replica dataset [68]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.87 + ], + "angle": 0, + "content": "we conducted scene-level evaluations. Qualitative results are shown in Fig. 8, which include COCOA [97] (first two rows), Replica [68] (third and fourth rows), and in-the-wild images (fifth and sixth rows). Here, we used Segment Anything [34] to obtain the visibility and occlusion masks. The results show that Amodal3R generates photorealistic 3D assets, whereas pix2gestalt fails to infer complete shapes from the same inputs, leading to unsatisfactory results from subsequent image-to-3D models. We also visualize the colored normal maps, which show that the results of Amodal3R are simultaneously rich in geometric and textural detail." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Quantitative results are presented in Tab. 4. As real-world datasets with instance 3D amodal annotations for" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.669 + ], + "angle": 0, + "content": "both geometry and appearance are limited, here we focus on geometry evaluation by curating a set of 49 completed 3D instances from the Replica scenes. Furthermore, we used the Umeyama algorithm [73] to match the generated meshes with the ground truth meshes in terms of scale and pose. It can be observed that our Amodal3R is still on par with or better than the baselines on the scene-level dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have introduced Amodal3R, a feed-forward approach to reconstruct the complete photorealistic 3D assets from partially visible images. The key to our success is that we construct a mask-weighted cross-attention mechanism and an occlusion-aware layer to effectively exploit visible and occluded information. Compared to the state-of-the-art methods that rely on sequential 2D completion followed by 3D generation, our Amodal3R achieves remarkably better performance by operating directly in 3D space. Furthermore, results on in-the-wild images indicate its potential for subsequent applications in 3D decomposition, marking a step towards robust 3D asset reconstruction in real-world environments with complex occlusion." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9188" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). This research is also supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 RG107/24. Dr. Chuanxia Zheng is supported by EPSRC SYN3D EP/Z001811/1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[2] Jiayang Ao, Yanbei Jiang, Qiuhong Ke, and Krista A Ehinger. Open-world amodal appearance completion. arXiv preprint arXiv:2411.13019, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.484, + 0.425 + ], + "angle": 0, + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.426, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[6] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2024. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[7] Minghao Chen, Roman Shapovalov, Iro Laina, Tom Monnier, Jianyuan Wang, David Novotny, and Andrea Vedaldi. Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[9] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.258, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[11] Junhyeong Cho, Kim Youwang, Hunmin Yang, and TaeHyun Oh. Robust 3d shape reconstruction in zero-shot from a single image in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.162, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Ruihang Chu, Enze Xie, Shentong Mo, Zhenguo Li, Matthias Nießner, Chi-Wing Fu, and Jiaya Jia. Diffcomplete: Diffusion-based generative 3d shape completion. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.906, + 0.314 + ], + "angle": 0, + "content": "[13] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.315, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[14] Ruikai Cui, Weizhe Liu, Weixuan Sun, Senbo Wang, Taizhang Shang, Yang Li, Xibin Song, Han Yan, Zhennan Wu, Shenzhen Chen, et al. Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.469, + 0.906, + 0.551 + ], + "angle": 0, + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 2, 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.552, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.608, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9189" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[24] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI conference on artificial intelligence, pages 10997-11004, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[25] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7662-7670, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[26] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (TOG), 36(4):107, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[27] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. 5, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[29] Yoni Kasten, Ohad Rahamim, and Gal Chechik. Point cloud completion with pretrained text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[31] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[32] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[33] Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, et al. Leveraging early-stage robustness in diffusion models for efficient and high-quality image synthesis. Advances in Neural Information Processing Systems, 36:1229-1244, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 5, 6, 8, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[37] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12642-12651, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[38] Xinhai Li, Huaibin Wang, and Kuo-Kun Tseng. Gaussian diffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[39] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[40] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 9298–9309, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[42] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[43] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 9970-9980, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[44] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[45] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[46] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4328-4338, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[47] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), pages 7220-7229. PMLR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9190" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 7588–7597, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[49] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[50] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11453-11464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[51] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc V Gool, and Sergey Tulyakov. Autodecoding latent 3d diffusion models. Advances in Neural Information Processing Systems, 36:67021-67047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[54] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3931-3940. IEEE Computer Society, 2024. 1, 2, 3, 5, 6, 7, 8, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[55] Eunbyung Park, Jimei Yang, Ersin Yumer, Duygu Ceylan, and Alexander C Berg. Transformation-grounded image generation network for novel 3d view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3500-3509, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763. PmLR, 2021. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[58] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.189 + ], + "angle": 0, + "content": "[60] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33:20154-20166, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[61] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero \\(123++\\): a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[63] Wu Shuang, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems, 37:121859–121881, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[64] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.907, + 0.524 + ], + "angle": 0, + "content": "[65] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.593 + ], + "angle": 0, + "content": "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning (ICML), pages 2256-2265. pmlr, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.662 + ], + "angle": 0, + "content": "[67] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798-1808, 2021. 2, 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[68] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[69] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1955-1964, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[70] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8863-8873, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[71] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9191" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.127, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[72] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision, pages 175-191. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[73] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on pattern analysis and machine intelligence, 13(4): 376-380, 2002. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[76] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[77] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[78] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In European Conference on Computer Vision (ECCV), pages 456-473. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[79] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[80] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 1, 3, 4, 5, 6, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.817 + ], + "angle": 0, + "content": "[81] Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, Soren Pirk, and Arie E Kaufman. Carve3d: Improving multi-view reconstruction consistency for diffusion models with rl finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6369-6379, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[82] Yiheng Xiong and Angela Dai. Pt43d: A probabilistic transformer for generating 3d shapes from single highly-ambiguous rgb images. arXiv preprint arXiv:2405.11914, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[83] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Amodal completion via progressive mixed context diffusion. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9109, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[84] Bo Yang, Hongkai Wen, Sen Wang, Ronald Clark, Andrew Markham, and Niki Trigoni. 3d object reconstruction from a single depth view with adversarial learning. In Proceedings of the IEEE international conference on computer vision workshops (ICCVW), pages 679-688, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[85] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28003-28013, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[86] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3784-3792, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[87] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[88] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.47, + 0.905, + 0.536 + ], + "angle": 0, + "content": "[89] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9720-9731, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[90] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1438-1447, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[91] Chuanxia Zheng, Duy-Son Dao, Guoxian Song, Tat-Jen Cham, and Jianfei Cai. Visiting the invisible: Layer-by-layer completed scene decomposition. International Journal of Computer Vision (IJCV), 129:3195-3215, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[92] Chuanxia Zheng, Tat-Jen Cham, Jianfei Cai, and Dinh Phung. Bridging global context interactions for high-fidelity image completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11512-11522, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[93] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[94] Xin-Yang Zheng, Hao Pan, Yu-Xiao Guo, Xin Tong, and Yang Liu. Mvd^2: Efficient multiview 3d reconstruction for multiview diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[95] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9192" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[96] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. Advances in neural information processing systems (NeurIPS), 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[97] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1464-1472, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[98] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "9193" + } + ] +] \ No newline at end of file diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_origin.pdf b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2528d73fe6ba03169b5730a8289b6b7116ed9880 --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/88b04b9b-d30d-4c9a-91cb-8a21cd28defd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b59f17b9e032733e57fd84823bd4e29a61f4392c1d22c001161eae13ba404481 +size 2783705 diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/full.md b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b96612742ba2a517ba79943eee34810ebbfa77bb --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/full.md @@ -0,0 +1,516 @@ +# Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images + +Tianhao Wu $^{1*}$ , Chuanxia Zheng $^{2\dagger}$ , Frank Guan $^{3}$ , Andrea Vedaldi $^{2}$ , Tat-Jen Cham $^{1}$ $^{*}$ S-Lab, $^{1}$ Nanyang Technological University, $^{2}$ VGG, University of Oxford + $^{3}$ Singapore Institute of Technology + +https://sm0kywu.github.io/Amodal3R/ + +![](images/db0f3e260a940e93d176cfcacfda74c200ff73d2202082359e50a6d358d3a7e1.jpg) +3D Recast +Input Image +Stochastic sample 1 + +![](images/4be72a6742cbba6fb6e0aa6ec50793ba1faff5009bfb21fdf47d0948e8aeff5f.jpg) +Stochastic sample 2 + +![](images/7042bf3f32f04a0978ae549a353e1b5a2825646847ab201d5bcb3d0a4fb37a3e.jpg) +In-the-wild Reconstruction +Input Image + +![](images/e8b639ca0d9316b8f72ff13100dfe1cd2bf2e8bfe27f71d18e823960df0befc9.jpg) +Stochastic sample 1 + +![](images/4bd49bd69f8a811e2e55f5c2ce9daeebf959d908174fb4622f831ed364f0e8fa.jpg) +Stochastic sample 2 + +![](images/1c125d3a9582a8e5d07e1a535e32ca82429b560e62da9f5848bd3c787846ec90.jpg) +3D Scene Decomposition +Input Image +Figure 1. Example results of Amodal3R. Given partially visible objects within images (occluded regions are shown in black, visible areas in red outlines), our Amodal3R generates diverse semantically meaningful 3D assets with reasonable geometry and plausible appearance. We sample multiple times to get diverse results from the same occluded input. Trained on synthetic datasets, it generalizes well to out-of-domain test examples from real scenes, where most objects are partially visible, and reconstructs reasonable 3D assets. + +![](images/73906cc49696dba9ce937eb09543a55bea4527421a3d01dd6b8f693af5fc1014.jpg) + +![](images/22060cd08dc52324121605dc107f208dc0f6955396270030ef0833c53c03ddd9.jpg) +Decomposed 3D assets + +![](images/39b7ea3e1098b79d12fdccc7992e0db1a7fdd8ecd2c429d3b867541db7111446.jpg) + +# Abstract + +Most existing image-to-3D models assume that objects are fully visible, ignoring occlusions that commonly occur in real-world scenarios. In this paper, we introduce Amodal3R, a conditional image-to-3D model designed to reconstruct plausible 3D geometry and appearance from partial observations. We extend a "foundation" 3D generator by introducing a visible mask-weighted attention mechanism and an occlusion-aware attention layer that explicitly leverage visible and occlusion priors to guide the reconstruction process. We demonstrate that, by training solely on synthetic data, Amodal3R learns to recover full 3D objects even in the presence of occlusions in real scenes. It substantially outperforms state-of-the-art methods that independently perform 2D amodal completion followed by 3D reconstruction, thereby establishing a new benchmark for occlusion-aware 3D reconstruction. + +# 1. Introduction + +Humans possess a remarkable ability to infer the complete 3D shape and appearance of objects from single views, even when those objects are partly hidden behind occluders. This ability, known as amodal completion, is key to operating in complex real-world environments, where objects are often partially occluded by surrounding clutter. However, existing image-based 3D reconstruction models [5, 6, 23, 41, 47, 48, 65, 70, 71, 75, 77, 80, 88, 89] fail to recover full 3D assets when the object is partially occluded. We thus consider the problem of reconstructing 3D objects from one or more partially occluded views, a novel task that we call amodal 3D reconstruction. + +Amodal 3D reconstruction is a challenging task that requires reconstructing an object's photorealistic 3D geometry and appearance, including the occluded regions, which are inherently ambiguous. Previous approaches [7, 54] typ + +ically decompose the task into a two-stage pipeline: first performing 2D amodal completion [85, 86, 91], followed by conventional 3D reconstruction [41, 75, 89]. While these two-stage pipelines are straightforward to implement, they have some drawbacks. First, 2D amodal completion methods rely predominantly on appearance-only priors, without incorporating explicit information on the 3D structure. This lack of 3D geometric reasoning limits their ability to generate physically plausible completions. Second, some 3D reconstruction methods depend on, or benefit from, multiview inputs. In this case, 2D amodal completion may lack multi-view consistency, particularly when performed independently for different views, which confuses the 3D reconstruction process. + +In this paper, we present Amodal3R, a novel single-stage paradigm for amodal 3D reconstruction that surpasses previous state-of-the-art approaches. Amodal3R fine-tunes a pre-trained 3D generator by introducing a dedicated branch that explicitly models occlusions. Its key advantage is performing reconstruction, completion, and occlusion reasoning directly in a 3D latent space rather than using a two-stage scheme. With this, the model can reconstruct both visible and occluded regions of the object coherently and plausibly. Specifically, we introduce a mask-weighted cross-attention to highlight visible information, along with a new occlusion-aware attention layer to guide completion in occluded areas. Together, these components enable the model to focus more on the visible parts of the object without significantly perturbing the pre-trained model. + +We designed several benchmarks to evaluate the performance of Amodal3R, including object-level Google Scanned Objects [16] and Toys4K [67] augmented with occlusions, 3D scenes from Replica [68], and real-world monocular images. Without relying on additional heuristics, Amodal3R achieves significantly superior performance compared to state-of-the-art models and generalizes well to different datasets, including real ones. + +In summary, our main contributions are as follows: + +- We propose a feed-forward 3D reconstruction model that directly reconstructs complete and high-quality 3D objects from one or more partially occluded views. +- We introduce a mask-weighted cross-attention mechanism and an occlusion-aware layer to inject occlusion awareness into an existing 3D reconstruction model, improving both the geometry and appearance of reconstructed objects when they are partially occluded. +- We demonstrate, via experiments on the 3D scene dataset Replica and real-world images, that our one-stage pipeline significantly outperforms existing two-stage approaches, establishing a new benchmark for amodal 3D reconstruction. + +# 2. Background + +2D Amodal Completion. Recent advances in 2D amodal completion [2, 54, 83, 85, 86, 91] have achieved significant success in reconstructing occluded regions of objects in 2D images. While these methods show promise for 3D generation pipelines [2, 54], they still have limitations. First, 2D amodal completion models have limited 3D geometric priors, as they interpret images as 2D patterns. While they excel at completing textures, they may generate physically implausible geometries when handling complex occlusions. This stems from their lack of explicit 3D reasoning and reliance on 2D appearance priors, without true volumetric understanding. Second, for models that use multi-view images as input, the results of 2D amodal completion are often inconsistent across views. Inconsistent views need to be handled by the 3D reconstructors, which causes confusion (see Sec. 4). Although there has been significant work on multiview consistent generation [7, 62, 70, 72, 81, 89, 94], multiview consistent completion is less explored. RenderDiffusion [1] contains an example, but the results are often blurry or lack details. This motivates our 3D-centric reconstruction framework. + +3D Shape Completion. Several methods [12, 14, 69, 95] start from a partial 3D reconstruction, then complete it in 3D. They use encoder-decoder architectures [24] or volumetric representations [10, 29] to robustly recover 3D shape, but not texture, which is necessary in many applications. They also still require recovering the partial 3D geometry from an occluded image, a challenge in itself. Furthermore, they ignore the input image when completing the object in 3D, disregarding the occlusion pattern that caused the 3D geometry to be recovered only partially. Recent work [11, 82] utilizes the partially visible object as input specifically for 3D shape completion. However, it focuses solely on geometry reconstruction, without recovering the object's texture and appearance. In contrast, our approach is end-to-end, recovering the complete 3D shape and appearance of the object from the occluded image. + +3D Generative Models. Early advancements in 3D generation are based on GANs [19], exploring various 3D representations such as point clouds [25, 39], voxel grids [84, 96], view sets [48, 55], NeRF [4, 5, 50, 60], SDF [18], and 3D Gaussian mixtures [78]. While GANs can capture complex 3D structures, they struggle with stability, scalability, and data diversity. The focus then shifted to diffusion models [22, 59, 66], which are more capable and versatile. They were first applied to novel view synthesis [76] in image space, before expanding to model a variety of 3D representations, including point clouds [44, 45, 49, 79, 93], voxel grids [37, 46], triplanes [64, 98], meshes [20, 43], and 3D Gaussian mixtures [9, 35, 38, 87]. Autoregressive models [8, 47, 65] have also been explored for mesh generation, focusing on improving the topology of the final 3D mesh. + +![](images/ec17aea1d842cf4d5ac780498601b8d2b71d4395e183436383899c5837c417a0.jpg) + +![](images/3a7b54826228d84368a9b61c041874faaffaf8f69e5964d42e366eec13c32630.jpg) +Figure 2. Overview of Amodal3R. Given an image as input and the regions of interest, Amodal3R first extracts the partially visible target object, along with the visibility and occlusion masks using an off-the-shelf 2D segmenter. It then applies DINOV2 [53] to extract features $c_{\mathrm{dino}}$ as additional conditioning for the 3D reconstructor. To enhance occlusion reasoning, each transformer block incorporates a mask-weighted cross-attention (via $c_{\mathrm{vis}}$ ) and occlusion-aware attention layer (via $c_{\mathrm{occ}}$ ), ensuring the 3D reconstructor accurately perceives visible information while effectively inferring occluded parts. For conditioning details, see Sec. 3.1. + +An alternative to performing diffusion in 2D image space [62, 70, 89] is to use 3D latent spaces [28, 51, 58, 63, 74, 80]. Such methods typically consist of two stages: the first for generating geometry and the second for generating appearance, and are capable of high-quality imaged-to-3D generation. However, they assume that input objects are fully visible, which limits their application to real-world scenes, where occlusions are frequent. Here, we extend such approaches to generate high-quality 3D assets from occluded input images — a challenging task that requires inferring complete 3D objects from partial information. + +# 3. Method + +Consider an image $x$ containing a partially occluded view of an object of interest $o_i$ . We wish to reconstruct the complete 3D shape and appearance $y$ of the object $o_i$ . This task is conceptually similar to existing image-to-3D reconstruction, except that here the object is partially occluded instead of being fully visible in $x$ . We call this problem amodal 3D reconstruction. + +We introduce Amodal3R (Fig. 2), a new method for amodal 3D reconstruction. Unlike recent two-stage methods [7, 54] that first perform 2D amodal completion followed by 3D reconstruction, Amodal3R is an end-to-end occlusion-aware 3D reconstruction model that generates complete geometry and appearance directly within the 3D space. Formally, Amodal3R is a conditional generator $v_{\theta}(y|o_i,M_{\mathrm{vis}},M_{\mathrm{occ}})$ that takes as input the image $x$ , centered on a partially visible object $o_i$ , as well as the visibility + +mask $M_{\mathrm{vis}}$ and occlusion mask $M_{\mathrm{occ}}$ . The visibility mask $M_{\mathrm{vis}}$ highlights the pixels in $x$ that correspond to object $o_i$ , while $M_{\mathrm{occ}}$ marks the pixels occupied by occluders, i.e., all other objects that potentially obscure $o_i$ . For real images, these masks can be efficiently obtained using pre-trained 2D segmentation models like Segment Anything [34]. By providing point coordinate prompts for the object of interest $o_i$ and its occluders, the segmentation model can generate the corresponding masks. In cases where multiple occluders are present, or when occluders fragment the visible area of the target object, a sequential process is employed. Specifically, point prompts for each visible/occluding region are provided to the 2D segmenter one at a time, with the model generating an individual mask for each region. These masks are then aggregated to form comprehensive visibility and occlusion masks. + +The challenges for Amodal3R are how to: (1) produce a plausible and complete 3D shape and appearance from partial observations, even in the absence of multi-view information; and (2) ensure 3D consistency in terms of geometry and photometry, maintaining seamless visual coherence between visible and generated regions. + +# 3.1. Mask-Conditional Generative Models + +To fine-tune an image-to-3D model for partially occluded inputs, the key change is to condition the transformer $v_{\theta}$ not only on the object of interest $o_i$ , but also on the visibility and occlusion masks, i.e., $M_{\mathrm{vis}}$ and $M_{\mathrm{occ}}$ . A naive approach is to patchify or embed the masks into tokens $(c_{\mathrm{vis}}, c_{\mathrm{occ}})$ , + +![](images/87381d95c462752cdc81640281a904c781af6b72aae913533c761f2102a88737.jpg) +(a) TRELLIS [80] +Figure 3. The transformer structure of Amodal3R. Compared with the original TRELLIS [80] design, we further introduce the mask-weighted cross-attention and occlusion-aware layer. It applies to both sparse structure and SLAT diffusion models. + +![](images/4b19bdfaaecc7a073d5f373b9f114d67bb5ce602a4bcad024930a9af08e96a53.jpg) +(b) Amodal3R (Ours) + +which can be concatenated to the image tokens $\pmb{e}_{\mathrm{dino}}$ and processed by cross-attention as before. However, the RGB $\pmb{o}_i$ contains significantly more information than the binary masks $M_{\mathrm{vis}}$ and $M_{\mathrm{occ}}$ . The transformer, which is initially trained with RGB only, may simply ignore the mask information. This is compounded by the fact that learning to use this new information, which involves a new type of token that is incompatible with image tokens, may require aggressive fine-tuning of the transformer. + +To sidestep this problem, inspired by masked attention approaches in 2D completion [92], we introduce mask-weighted cross-attention and an occlusion-aware attention layer to better utilize the visibility mask $M_{\mathrm{vis}}$ and the occlusion mask $M_{\mathrm{occ}}$ without disrupting the pre-trained 3D generator. These are described next. + +Mask-weighted Cross-Attention. A key novel component of Amodal3R is mask-weighted cross-attention, which allows the model to focus its attention on the visible parts of the object. We implement this mechanism in each attention block in the transformers $v_{\theta}$ of a pre-trained 3D generator. Given the latent tokens $\ell \in \mathbb{R}^{L \times C}$ input to a cross-attention layer as well as the image features $c_{\mathrm{dino}} \in \mathbb{R}^{K \times C'}$ of the partially visible object, cross-attention computes the similarity score matrix via + +$$ +\boldsymbol {q} = W _ {q} \boldsymbol {\ell}, \quad [ \boldsymbol {k}, \boldsymbol {v} ] = W _ {k v} \boldsymbol {c} _ {\mathrm {d i n o}}, \quad S = \boldsymbol {q} \boldsymbol {k} ^ {\top} / \sqrt {D}, \tag {1} +$$ + +where $W_{q}$ and $W_{kv}$ are learnable projections that map the latents $\ell$ to the query $q$ , and the conditioning image feature $c_{\mathrm{dino}}$ to the key $k$ and the value $v$ , respectively, while $D$ is the dimension of the query and key vectors. + +We wish to bias the token similarity matrix $S \in \mathbb{R}^{L \times K}$ towards the visible parts of the object. Recall that $K$ + +![](images/e2d2a57d59bd8efdf730ab223444dac9787b762c1c812448af756fe46115ecca.jpg) +Figure 4. 3D-consistent mask example. Given a 3D mesh, we render selected triangles in a distinct color from the others to generate multi-view consistent masks. It allows the evaluation of multi-view methods in handling contact occlusion. (The occluded regions are shown in red.) + +![](images/211eab21b339562a7b9734004b576c27eb64f521933ae646676588415d5a3267.jpg) + +![](images/6e1959f584b0d93c348d0d6641a7cac5286305a31fe747f49260119ea0f248a2.jpg) + +![](images/3c82569277c7393f8b8bce254ed494b5e4abb07f88512c3fdd1a8e86563b580f.jpg) + +![](images/932a594172e1a1eec2ea793d674805e1bcd98db4e07ceba4702b4ebfc7c48e12.jpg) + +is the number of tokens $c_{\mathrm{dino}}$ extracted by DINOv2 [52] from the occluded image $x \cdot M_{\mathrm{vis}}$ . Each token thus corresponds to a $P \times P$ patch in the input image (where $P = 14$ ). We extract analogous flattened 1-D tokens $c_{\mathrm{vis}} = [\rho(M_{\mathrm{vis},1}), \dots, \rho(M_{\mathrm{vis},K})]$ from the visibility mask by calculating the fraction of $P \times P$ pixels that are visible in the $j$ -th image patch $M_{\mathrm{vis},j}$ . $c_{\mathrm{vis},j} = \rho(M_{\mathrm{vis},j}) \in [0,1]$ . We then use these quantities to bias the computation of the attention matrix $A = \text{softmax}(S + \log c_{\mathrm{vis}}) \in [0,1]^{L \times M}$ via broadcasting. Hence: + +$$ +A _ {i j} = \frac {\boldsymbol {c} _ {\text {v i s} , j} \exp \left(S _ {i j}\right)}{\sum_ {k = 1} ^ {K} \boldsymbol {c} _ {\text {v i s} , k} \exp \left(S _ {i k}\right)}. \tag {2} +$$ + +In this manner, the visibility flag modulates the attention matrix $A$ smoothly. If there are no visible pixels in a patch $j$ , then $A_{ij} = 0$ , so the corresponding image tokens are skipped in cross-attention. While we have illustrated how this works for a single head, in practice Amodal3R uses a multi-head transformer, to which Eq. (2) extends trivially. Please see the supplementary materials for more details. + +Occlusion-aware Attention Layer. To prevent the network from arbitrarily extending the geometry beyond the occluded regions, we further introduce an occlusion-aware attention layer. The key insight is that specifying only visible information is insufficient; it is equally important to distinguish foreground occluders from the background, as this explicitly identifies the potential regions requiring completion. Namely, when a pixel is marked as invisible in the mask $M_{\mathrm{vis}}$ , this might be (i) because there is an occluder in front of that pixel (so the pixel could plausibly belong to the object of interest if the occlusion were removed), or (ii) because the pixel is entirely off the object. This information is encoded by the mask $M_{\mathrm{occ}}$ . + +To help the model distinguish between visible, occluded, and background areas, we add one more cross-attention layer placed immediately after the mask-weighted cross-attention layer, and pool solely the occlusion mask $M_{\mathrm{occ}}$ . We encode the occlusion mask $M_{\mathrm{occ}}$ as a sequence of flattened 1-D tokens $c_{\mathrm{occ}} = [\rho(M_{\mathrm{occ},1}), \ldots, \rho(M_{\mathrm{occ},M})]$ , and then simply pool $c_{\mathrm{occ}}$ using a cross-attention layer. + +# 3.2. Simulating Occluded 3D Data + +To train and evaluate our model, we require a dataset of 3D assets imaged in scenarios with clutter and occlusions. We curated a large synthetic 3D dataset to train our model, as collecting such 3D assets in the real world is challenging. + +Random 2D Occlusions. Each training sample $(o, x, M_{\mathrm{vis}}, M_{\mathrm{occ}})$ consists of a 3D object $o$ (from which ground truth latents can be obtained by using the encoders of Appendix A.1), an image $x$ with partial occlusions, and corresponding visibility and occlusion masks $M_{\mathrm{vis}}$ and $M_{\mathrm{occ}}$ . In a real scenario, occlusions arise from other objects in the scene. In a multi-view setting, like the one discussed below, these occlusions must be consistent across views, reflecting the underlying scene geometry. However, because our model is trained for single-view reconstruction, we can simulate occlusions by randomly masking parts of the object after rendering it in 2D. This is simpler and allows us to generate fresh occlusion patterns each time a view is sampled for training. + +Thus, given $o$ and an image $x$ rendered from a random viewpoint, we generate random occlusion masks as follows. Inspired by work on 2D completion [26, 40, 56, 90, 92], we randomly place lines, ellipses, and rectangles, simulating diverse masking patterns. The union of these random shapes gives us the occlusion mask $M_{\mathrm{occ}}$ . Then, if $M_{\mathrm{obj}}$ is the mask of the unoccluded object, the visible mask is given by $M_{\mathrm{vis}} = M_{\mathrm{obj}} \odot (1 - M_{\mathrm{occ}})$ . Examples of such patterns are given in the supplementary material. + +3D-consistent Occlusions. In a real scene imaged from multiple views, occlusions are not random but geometrically consistent across different viewpoints, as they result from other objects. This is especially true for contact occlusions, where parts of an object remain occluded by other objects from all viewpoints. To evaluate the model's performance under such challenging conditions, 3D-consistent masks are required. + +To efficiently generate such contact occlusion masks, we leverage 3D meshes during rendering. Starting from a randomly chosen triangle, we apply a random-walk strategy to iteratively select neighboring triangles, forming continuous occluded regions until the predefined mask ratio is met. By rendering these masked meshes using the same camera parameters as the RGB images, we ensure multi-view consistency in the generated occlusion masks (see Fig. 4). + +# 3.3. Reconstruction with Multi-view Input + +Since our flow-based model performs multiple denoising steps and does not require known camera poses for input views, Amodal3R can benefit from multi-view reference images $\mathcal{X} = \{x_i\}_{i=1}^N$ as conditioning inputs at different steps of the denoising process. While multi-view conditioning naturally improves reconstruction performance, a potential issue is that an image used earlier in the denoising process + +has a stronger influence on the final 3D geometry. This is because early denoising steps establish the coarse geometry, whereas later steps refine higher-frequency details [32, 33]. Therefore, we prioritize input images based on their visibility. Specifically, when experimenting with multi-view inputs, we sort the images in proportion to the object visibility $|M_{\mathrm{vis}}|$ , ensuring that images with higher visibility are used as primary inputs. + +# 4. Experiments + +# 4.1. Experiment Settings + +Datasets. Amodal3R is trained on a combination of 3D synthetic datasets: 3D-FUTURE (9,472 objects [17]), ABO (4,485 objects [13]), and HSSD (6,670 objects [31]). Once trained, we first assessed its effectiveness on Toys4K (randomly sampling 1,500 objects [67]) and Google Scanned Objects (GSO) (1,030 objects [16]), which were excluded from the training data for both our model and the baselines. During inference, a 3D-consistent mask occludes each input object, and each view is augmented with additional randomly generated occlusion areas. This ensures that the model cannot simply extract the region to be completed from the occlusion masks. To further assess out-of-domain generalization in practical applications, we also evaluated all models on the 3D scene dataset Replica [68] as well as on in-the-wild images. + +Metrics. To measure the quality of the rendered images, we use the Fréchet Inception Distance (FID) [21] and the Kernel Inception Distance (KID) [3]. To assess the quality of the reconstructed 3D geometry, we use the Coverage Score (COV), the Point cloud FID (P-FID) [49], and the Minimum Matching Distance (MMD) using the Chamfer Distance (CD). COV measures the diversity of the results, while P-FID and MMD measure the quality of the 3D reconstruction. We also use the CLIP score [57] to evaluate the consistency between each generated and ground-truth object pair. + +Baselines. Most image-to-3D generators are trained on fully visible inputs. To ensure fair comparison, we completed the partially visible 2D objects before passing them to 3D generators. In particular, we use pix2gestalt [54], a state-of-the-art 2D amodal completion network. Using this protocol, we compared Amodal3R to state-of-the-art methods such as TRELLIS [80], Real3D [27], GaussianAnything [36]ICLR'25,and LaRa [6]ECCV'24. + +Implementation Details. Amodal3R is trained on 4 A100 GPUs (40G) for 20K steps with a batch size of 16, taking approximately one day. We select TRELLIS [80] as underlying model because it is the latest 3D generator and is publicly available. However, our method is not specific to this model. + +
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]92.261.300.7434.6935.925.03
Real3D [27]1pix2gestalt [54]91.212.020.7523.9219.619.21
TRELLIS [80]1pix2gestalt [54]58.825.870.7626.4331.654.17
Amodal3R (Ours)1—30.640.350.817.6939.613.62
LaRa [6]4pix2gestalt [54]172.844.540.7066.3424.568.11
LaRa [6]4pix2gestalt [54]+MV97.532.630.7521.8026.218.61
TRELLIS [80]4pix2gestalt [54]65.696.920.7824.6432.334.26
TRELLIS [80]4pix2gestalt [54]+MV60.371.850.8319.6831.754.21
Amodal3R (Ours)4—26.270.220.845.0338.743.61
+ +Table 1. Amodal 3D reconstruction results on GSO [16]. Here, V-num denotes the number of input views, and 2D-Comp refers to the 2D completion method. For single-view image-to-3D, we first complete occluded objects using pix2gestalt [54] before passing them to various 3D models. For multi-view image-to-3D, we adopt two variants: (1) pix2gestalt [54] is applied independently on each view; (2) pix2gestalt $[54] + \mathrm{MV}$ : a single-view completion is generated first, followed by multi-view diffusion [61] to synthesize 4 views as inputs. Without bells and whistles, Amodal3R consistently outperforms state-of-the-art models across all evaluation metrics. + +
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]57.171.220.8021.9733.567.23
Real3D [27]1pix2gestalt [54]59.921.630.7923.3124.359.53
TRELLIS [80]1pix2gestalt [54]43.056.830.8026.0426.286.87
Amodal3R (Ours)1—23.450.420.835.0037.095.89
LaRa [6]4pix2gestalt [54]123.523.610.7545.9127.899.67
LaRa [6]4pix2gestalt [54]+MV75.334.140.8013.0024.8210.93
TRELLIS [80]4pix2gestalt [54]46.348.770.8128.7625.357.13
TRELLIS [80]4pix2gestalt [54]+MV43.007.530.8124.4126.557.05
Amodal3R (Ours)4—20.930.500.853.7839.035.75
+ +Table 2. Amodal 3D reconstruction results on Toys4K [67]. The experimental setting is the same to Tab. 1, except for the dataset. + +![](images/214c14a2442947413628d675f0dc806eac9a9b619143d2876fd720ddb04267b6.jpg) +Input + +![](images/5d09515060d8f065dc68a83ff275e786b139d6dc6bf88deeda9e4bd95c9b5c0b.jpg) +pix2gestalt [54] + +![](images/42865b2963f051b0c2bff5ffc35995faa5cab09351613b701818ec6602ab8fd8.jpg) +Figure 5. Single-view amodal 3D reconstruction. The occlusion regions are shown in black and the visible regions are highlighted with red outlines. More examples are provided in supplementary material Fig. D.4 + +![](images/c2763fdf1c66b93cc21539d425558ff93e45d8f13c0c5da52e3e678eff76f7a1.jpg) +GaussianAnything [36] + +![](images/cbdc468da5e748eb5f62a19b180df9909d73e5150ed73d6a08080db0e2afb4ae.jpg) +Real3D [27] + +![](images/9cd07fa7f5d6b40ebaebb2d0c5ae2177b0dedaa1f1b4a8d9716bf88166a01b7d.jpg) + +![](images/0ed7d60ced0c102b5269e611f751e0e04ab271d3a3ece8cefb85e6d32796eafd.jpg) +TRELLIS [80] + +![](images/379e0e796dc1c2e971bd5e5916cb2f25f540ccfed2e459e2df28035bbc850c2f.jpg) + +![](images/837c8a94c76ee900b4119ba1434ef012529765f892c9fef8e3d7a31849774486.jpg) +Amodal3R (Ours) + +# 4.2. Main Results + +Quantitative Results. We compare Amodal3R with prior methods on amodal 3D reconstruction in Tabs. 1 and 2. Amodal3R significantly outperforms the baselines across all metrics in both single- and multi-view settings, demonstrating its effectiveness. Notably, two-stage methods using multiple views ("4 V-num + pix2gestalt") perform worse than their single-view counterparts ("1 V-num + pix2gestalt"). This shows that inconsistent 2D completion + +degrades 3D reconstruction (Sec. 2). In contrast, Amodal3R avoids this issue by operating directly in 3D space and benefits from additional occluded views as expected. + +Qualitative Results. Qualitative results are shown in Figs. 5 and 6 and in the supplementary material Figs. D.4 and D.5. For all baselines, pix2gestalt is first applied for 2D amodal completion (second column), and the corresponding completed images are passed to each baseline image-to-3D generator. Amodal3R consistently produces high-quality 3D assets under both single- and multi-view settings, even + +![](images/12012618131c21e5dab9e6f613af87cc1a0dcbc2136a2a7bf79f8f68b8228445.jpg) +Input + +![](images/a1947c7a55137f0e34a70b742a96e9cf6f01a1ef8656f708836959fa26593cdf.jpg) +pix2gestalt [54] + +![](images/45d750e591eb5be7d2bce1b5c60b0ac898451b57f911f0c101146bdfc512c33e.jpg) + +![](images/45897ab0ddaa06acd8c18b9318729497506e1506fcced3531a81b2f89c09532a.jpg) +LaRa [6] + +![](images/f199b91890a543943aee5d14751d56d92d35f08a66e6fcfe03090295c0d965b0.jpg) + +![](images/169fbf26b83e089ce7fe23f2286a3902a20d903d4affd660041b5ed69ddad98c.jpg) + +![](images/6554d658abf7a597ba389b01081b18ecf8e82f2bbc113efa95bca9da7f768f63.jpg) + +![](images/ff823111c9b70405d6c45af704cfa0a836289b2b3850deced980ff5166b7e264.jpg) +TRELLIS [80] + +![](images/e531f073587deb24f188aa5a57000be49699f533947ee2fb2bb1b68528d8ccad.jpg) + +![](images/f7403491e87d1bd5e14a6ed2d3ea06cafb71daf13387f7a90236d248aeb5ba3a.jpg) + +![](images/6790e50b47b03e37fc50e00599c442571ea9baf83c3f4ff8960de4f09a771123.jpg) + +![](images/d214b6b7760f05c0c3c52637bc5ff6baca125c5df226f7fa47a0f85b78a853d8.jpg) +Amodal3R (Ours) +Figure 6. Multi-view amodal 3D reconstruction. The above results are reconstructed using 4 occluded input views. Due to limited space, we present only the best results for LaRa and TRELLIS under the "pix2gestalt+MV" setting. We apply 3d-consistent mask and random extended occlusion areas. More examples are provided in supplementary material Fig. D.5. + +![](images/1cb5a8c474a13b4076e5c3119e6d82d418f88a69167b4365b548debcdde86cf9.jpg) + +![](images/a7498ce14f93ae2c932cd88d4ba37575e8adf7627855d1c507d76e6f6ee88b53.jpg) + +![](images/55be6d9d173b001d1b42e5a16a60ea950a5815f07e0e6e1da5032576fc55f759.jpg) +Input +' Naive concatenation +Figure 7. Ablation examples. Naive concatenation fails to reconstruct complete shape and appearance. Mask-weighted attention alone extends geometry into background regions, while occlusion-aware attention alone cannot guarantee photorealistic appearance. + +![](images/535de559e9e64624d482c201449869b4f691749929ca466b2ba824ad8392d7dd.jpg) + +![](images/b03a90ec60dde9fb4c280130c6844cb500af15892bb7bb790a67fde2001adbc0.jpg) + +![](images/2d24585d24a0d43522a7ee702bde2614e7b12ab0d70c736989d762513479e42c.jpg) + +![](images/2fd0b4a2bc4cc74d8d9fe4607d26c742c3bdd95e9600d43431d943dd377f8034.jpg) + +![](images/ad19ffd4eb033e8daa9763f2619f2d325b35208ea5f8d5c191acee1f40db0150.jpg) + +![](images/6d3cca3cb212f5ab952596e9352fc814ed08e754a914d31f28c9473b84f92b40.jpg) +w/only $M_{\mathrm{vis}}$ attention + +![](images/7ffa1cfdc0ab7af2637d775a87d8c18e8e912644e517147ce22c31d6bbc67675.jpg) + +![](images/98cd5d400fbaed8d15d5ddd5f788da08cafdf1f1c12993f866dcc5a8feb294c6.jpg) + +![](images/76b15a1e0fb1d39530419956091de5892730b41beac49a64ece981401612de8f.jpg) +w/only $M_{\mathrm{occ}}$ attention + +![](images/48ae97f8280fd4eb1d96592a5215d8dc71da9dbf33a861e42ca530b9cccbeec7.jpg) + +![](images/6835bb035d609bf6ba38348e36d53a49eac7ab22e4a7c4a8afedaa49f7d85fb6.jpg) + +![](images/959175325ad9a896663bb3b676db8e8cff8d45ca0065f9e5a4fb82de74ddc81c.jpg) +Full model + +in challenging scenarios. In contrast, 2D completion inconsistencies accumulate as more views are added, especially when pix2gestalt is more uncertain, confusing the downstream reconstructor models. For instance, in the giraffe example in Fig. 5, the pix2gestalt completion fails to capture the overall structure of the 3D object, which in turn leads the 3D generator models to misinterpret it as a woodstick-like shape. Instead, Amodal3R reconstructs the 3D geometry and appearance well, maintaining good alignment with the partially visible inputs. These results underscore the effectiveness of Amodal3R and highlight the benefit of integrating occlusion reasoning directly in 3D space, eliminating the reliance on separate, potentially inconsistent, 2D completion stages. + +# 4.3. Ablation Study + +We conducted several ablations to study the impact of the different components of Amodal3R and report the results in Tab. 3 and Fig. 7. For these experiments, we test single-view image-to-3D reconstruction on the GSO dataset. + +Naive Conditioning. We first evaluated a version of the model that still conditions the reconstruction on the visibility and occlusion masks, but without using the modules of Sec. 3.1. Instead, we directly concatenate the tokens $c_{\mathrm{vis}}$ + +
MethodAppearanceGeometry
FID ↓KID(%) ↓COV(%) ↑MMD(‰) ↓
naive conditioning31.960.4937.963.61
w/ only mask-weighted attention30.530.3836.903.69
w/ only occlusion-aware layer31.770.5740.193.51
full model (Ours)30.640.3539.613.62
+ +Table 3. Ablations on different mask conditioning designs. + +to $c_{\mathrm{dino}}$ to form the condition for the cross-attention layer. The results (first row in Tab. 3 and second column in Fig. 7) show that, while the resulting model can still perform basic completion, the textures in the occluded regions are inconsistent with those in the visible ones, and the reconstructed geometry is inaccurate, e.g. the hole in the shoe. + +Mask-weighted Attention. To evaluate the effectiveness of our proposed mask-weighted attention mechanism, we integrate it into the training while omitting the occlusion-aware layer. The results demonstrate improved visual quality, especially in capturing texture details in the visible areas. However, the geometries exhibit deficiencies, as seen in the problematic shoe in Fig. 7 (first row), and the toy monster with a broken back (third row). + +Occlusion-aware Layer. The integration of the occlusion-aware layer improves the geometry both quantitatively and qualitatively. This aligns with our motivation for introducing an additional cross-attention layer, aimed at reconstructing the visible areas via the image-conditioned layer and occluded areas via the subsequent layer. However, the occlusion-aware layer alone still results in unsatisfactory appearances, which again indirectly attests to the effectiveness of the mask-weighted attention mechanism. + +Full Model. Consequently, the full model integrating both modules achieves optimal 3D generation results characterized by complete geometry and consistent textures. + +# 4.4. Real-World Generation / Completion + +Amodal3R demonstrates strong generalization to out-of-domain test examples, primarily because it builds upon a large-scale pre-trained 3D generator and is fine-tuned with diverse categories. To demonstrate this advantage, + +![](images/0aaa4a1e38acdffb89a296b94d52034c83d55d827f052e369128603ee27b9ea3.jpg) +Occluded input + +![](images/cd9906c266acf5dba296545ee4c74eb83b3bd899f98e2c37e9f02386537e8754.jpg) + +![](images/6255e3e76a9b3d590968e813e66631f1bea605af0f2ac52acdc079fadd1eacf9.jpg) + +![](images/3fa5c180b999464af1fdc01181a725ffe146219a7762a97543d5a8b1b4ac4c42.jpg) + +![](images/f10aee28cacabda7128574d2eae3a69cb0d67b96aa299faf18e5be799b306367.jpg) + +![](images/892deecb97e98270146a0d38533fa2d5c34f4035e00f5925d4e7e03b6ff31120.jpg) + +![](images/234d82c8065bcddf1d2049287d52b9664374f9fface24fcba575e534966411d2.jpg) + +![](images/e36b8a031e27126923f0ced3c2ab4180a4bac07bc33df0bdaf9b49b53757226c.jpg) + +![](images/82dea0ecb153627d533b3aac6897890423a40e2c23dd3df859e0e95839deb30d.jpg) + +![](images/88295939781841cee35d5000c351c08e248212fab62f8c07c3f561f1a3889ca1.jpg) + +![](images/fc064faa16ef879a9ba1d032fee422fb0ebc947eabfea37da35ad56efcbabf12.jpg) + +![](images/0ff6bc3ed69e3e2674d70fa394d723bcb8970c492a0659966af6727f07648192.jpg) + +![](images/8a4314c862580b2aa2fce1ac227cac194689dde71da5b50611a71b7a236e2b41.jpg) + +![](images/d9057c53bb1aed5ecf98e5fed336de69e42395bf867e96f5ebb0b28e771fcd2b.jpg) + +![](images/8f7dcc9b77e7dd831b84e0fc47252353b7c026800f899e87b8272a5eab1a687e.jpg) + +![](images/2571264e7593b19c26b1ee6901744dc8e429fa68c88e75dfc330069516431d8e.jpg) + +![](images/1b348933600ad1b2ca1ac0b04a6f11de2ddc290c3c10fe390247968962e7e83e.jpg) + +![](images/d3f47ac1d5f26df7b563f91770e9388b5209e1d152b5e68c1a5b0ecc481442a8.jpg) + +![](images/fdab1d71ab835dfbf75ed3891d78e1e9dd2884ad01ce645725e44565f26d6905.jpg) + +![](images/779221d84eac13762e3104fc796d1250c707fe75e07cd2bd744cfa1eb962af37.jpg) + +![](images/3812f3278f3a6124fa65304bb1fb01e03cd8551b3e7f80e5e82ae36ada5851e5.jpg) + +![](images/18cde7d1bc2554bec861d740a3074698dd941324d4d785f080a7db93745fc3bf.jpg) + +![](images/0f2e244f019b483441bf1b023c3042682d39627a63eadbcd551087476d6c8165.jpg) + +![](images/dc23b4a45ae4854a6f2a8956b0389f2020b1b7c916bbd755a40886f1655b56b6.jpg) + +![](images/cf392de632adcf2d68ffa9409dd67b0b8a481bd3f1a1614627d0b9a9912e4579.jpg) + +![](images/70020367524cbc699a41c68ab03082351c1238ec07afadaf4072d8b3414b7190.jpg) + +![](images/ebdeda1de9d2f58051643f58f0aafffba36868ae1a7ec8e7faf7e5c6f5f6fafe.jpg) + +![](images/77e1b7449a788711e712931136f03a029ebda3a502be37bc41a9a790c689ba8b.jpg) +Occluded input +Figure 8. Examples on Replica [68] and in-the-wild images. The target objects and the occluders are shown in red and green outlines. More examples are provided in supplementary material Figs. D.7 to D.9. + +![](images/9ace62434057fe4fb29c912e2103f8da59747ea756046d5a2b9791ef453b560a.jpg) + +![](images/185102de64684adf0ad58d98b905b6426df9d917e537d74a4542a7a1aebc8fc1.jpg) + +![](images/31b127e645d074e2b254bf4a71b0909fa02ecb6743ee5223b33dfb650101491b.jpg) + +![](images/289d9b695e7ba4ba8013fd3ee4cc17b39ff7fafa036e8844ab8e49af2b771168.jpg) +Novel views + +![](images/1fe967532a0c24f45888599ae0b356d60e934b6c115bd8f680688d8a313e5cb0.jpg) +pix2gestalt [54] GaussianAnything [36] +Real3D [27] + +![](images/8a87f5edfdfea7b2908f2f1152168e5d1f6034398e5b237e129cf4821cc533de.jpg) + +![](images/2acc76aecb295dfc1a8c335dadb975480f04c941e533914e6d2fc45b1e46d92c.jpg) + +![](images/c863e58eeb415d78874e23b004d7563bee499fda1fa4df113ed32f1dbb27686d.jpg) + +![](images/92dfd49f4bbb50ad9c2fb36968106be55a1d8652bb5a0beeaa3cecc05f9b4e9e.jpg) +TRELLIS [80] + +![](images/3f4adb75a662be7f373e48ff20b88e463399b648cd18d9c0de7665505e87bd4b.jpg) + +![](images/8446728d4dd3a897b765ea0c0f6f4388fff96bfe0fb74902dea7d8d9730c2ad6.jpg) + +![](images/fd104cfb159e20527df1235e54d67489c2e59aa683914fe2e44f69546062d9eb.jpg) +Normal maps + +![](images/599cd74b2b6add55f3c3bd9a79332d182be0b6e97457cdd4bf02904893136add.jpg) +Amodal3R (Ours) + +![](images/d38f75edd1ea05c2b78c64e0e16f7abaa3e2f6bc04b3f720a3b7a564bebf594b.jpg) + +
MethodP-FID↓COV(%) ↑MMD(‰) ↓
GaussianAnything91.3761.2225.00
Real3D79.5555.1018.28
TRELLIS128.5259.1723.95
Amodal3R61.4659.1810.98
+ +Table 4. Quantitative results on the Replica dataset [68]. + +we conducted scene-level evaluations. Qualitative results are shown in Fig. 8, which include COCOA [97] (first two rows), Replica [68] (third and fourth rows), and in-the-wild images (fifth and sixth rows). Here, we used Segment Anything [34] to obtain the visibility and occlusion masks. The results show that Amodal3R generates photorealistic 3D assets, whereas pix2gestalt fails to infer complete shapes from the same inputs, leading to unsatisfactory results from subsequent image-to-3D models. We also visualize the colored normal maps, which show that the results of Amodal3R are simultaneously rich in geometric and textural detail. + +Quantitative results are presented in Tab. 4. As real-world datasets with instance 3D amodal annotations for + +both geometry and appearance are limited, here we focus on geometry evaluation by curating a set of 49 completed 3D instances from the Replica scenes. Furthermore, we used the Umeyama algorithm [73] to match the generated meshes with the ground truth meshes in terms of scale and pose. It can be observed that our Amodal3R is still on par with or better than the baselines on the scene-level dataset. + +# 5. Conclusion + +We have introduced Amodal3R, a feed-forward approach to reconstruct the complete photorealistic 3D assets from partially visible images. The key to our success is that we construct a mask-weighted cross-attention mechanism and an occlusion-aware layer to effectively exploit visible and occluded information. Compared to the state-of-the-art methods that rely on sequential 2D completion followed by 3D generation, our Amodal3R achieves remarkably better performance by operating directly in 3D space. Furthermore, results on in-the-wild images indicate its potential for subsequent applications in 3D decomposition, marking a step towards robust 3D asset reconstruction in real-world environments with complex occlusion. + +# Acknowledgements + +This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). This research is also supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 RG107/24. Dr. Chuanxia Zheng is supported by EPSRC SYN3D EP/Z001811/1. + +# References + +[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2 +[2] Jiayang Ao, Yanbei Jiang, Qiuhong Ke, and Krista A Ehinger. Open-world amodal appearance completion. arXiv preprint arXiv:2411.13019, 2024. 2 +[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 5, 3 +[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 5799-5809, 2021. 2 +[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2 +[6] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2024. 1, 5, 6, 7 +[7] Minghao Chen, Roman Shapovalov, Iro Laina, Tom Monnier, Jianyuan Wang, David Novotny, and Andrea Vedaldi. Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3 +[8] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2 +[9] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Neural Information Processing Systems (NeurIPS), 2024. 2 +[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2 + +[11] Junhyeong Cho, Kim Youwang, Hunmin Yang, and TaeHyun Oh. Robust 3d shape reconstruction in zero-shot from a single image in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2 +[12] Ruihang Chu, Enze Xie, Shentong Mo, Zhenguo Li, Matthias Nießner, Chi-Wing Fu, and Jiaya Jia. Diffcomplete: Diffusion-based generative 3d shape completion. Advances in Neural Information Processing Systems, 36, 2024. 2 +[13] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5 +[14] Ruikai Cui, Weizhe Liu, Weixuan Sun, Senbo Wang, Taizhang Shang, Yang Li, Xibin Song, Han Yan, Zhennan Wu, Shenzhen Chen, et al. Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2 +[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 3 +[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 2, 5, 6, 3 +[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 5 +[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2 +[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2 +[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 3 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2 + +[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1 +[24] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI conference on artificial intelligence, pages 10997-11004, 2020. 2 +[25] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7662-7670, 2020. 2 +[26] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (TOG), 36(4):107, 2017. 5 +[27] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. 5, 6, 8, 2 +[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3 +[29] Yoni Kasten, Ohad Rahamim, and Gal Chechik. Point cloud completion with pretrained text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1 +[31] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 5 +[32] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 5 +[33] Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, et al. Leveraging early-stage robustness in diffusion models for efficient and high-quality image synthesis. Advances in Neural Information Processing Systems, 36:1229-1244, 2023. 5 +[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 3, 8 +[35] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas + +Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 2 +[36] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 5, 6, 8, 2, 3 +[37] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12642-12651, 2023. 2 +[38] Xinhai Li, Huaibin Wang, and Kuo-Kun Tseng. Gaussian diffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221, 2023. 2 +[39] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2 +[40] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 5 +[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 9298–9309, 2023. 1, 2 +[42] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1 +[43] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 9970-9980, 2024. 2 +[44] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2837-2845, 2021. 2 +[45] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 2 +[46] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4328-4338, 2023. 2 +[47] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), pages 7220-7229. PMLR, 2020. 1, 2 + +[48] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 7588–7597, 2019. 1, 2 +[49] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 5, 3 +[50] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11453-11464, 2021. 2 +[51] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc V Gool, and Sergey Tulyakov. Autodecoding latent 3d diffusion models. Advances in Neural Information Processing Systems, 36:67021-67047, 2023. 3 +[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 1 +[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 3, 1 +[54] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3931-3940. IEEE Computer Society, 2024. 1, 2, 3, 5, 6, 7, 8, 4 +[55] Eunbyung Park, Jimei Yang, Ersin Yumer, Duygu Ceylan, and Alexander C Berg. Transformation-grounded image generation network for novel 3d view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3500-3509, 2017. 2 +[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544. IEEE, 2016. 5 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763. PmLR, 2021. 5, 3 +[58] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image + +synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 2 +[60] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33:20154-20166, 2020. 2 +[61] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero $123++$ : a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 6, 4 +[62] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3 +[63] Wu Shuang, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems, 37:121859–121881, 2025. 3 +[64] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2 +[65] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2 +[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning (ICML), pages 2256-2265. pmlr, 2015. 2 +[67] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798-1808, 2021. 2, 5, 6, 3 +[68] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 8 +[69] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1955-1964, 2018. 2 +[70] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8863-8873, 2023. 1, 2, 3 +[71] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1 +[72] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision, pages 175-191. Springer, 2024. 2 +[73] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on pattern analysis and machine intelligence, 13(4): 376-380, 2002. 8 +[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3 +[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457. Springer, 2024. 1, 2 +[76] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2 +[77] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 1 +[78] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In European Conference on Computer Vision (ECCV), pages 456-473. Springer, 2024. 2 +[79] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2 +[80] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 1, 3, 4, 5, 6, 7, 8, 2 +[81] Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, Soren Pirk, and Arie E Kaufman. Carve3d: Improving multi-view reconstruction consistency for diffusion models with rl finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6369-6379, 2024. 2 +[82] Yiheng Xiong and Angela Dai. Pt43d: A probabilistic transformer for generating 3d shapes from single highly-ambiguous rgb images. arXiv preprint arXiv:2405.11914, 2024. 2 +[83] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Amodal completion via progressive mixed context diffusion. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9109, 2024. 2 +[84] Bo Yang, Hongkai Wen, Sen Wang, Ronald Clark, Andrew Markham, and Niki Trigoni. 3d object reconstruction from a single depth view with adversarial learning. In Proceedings of the IEEE international conference on computer vision workshops (ICCVW), pages 679-688, 2017. 2 +[85] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28003-28013, 2024. 2 +[86] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3784-3792, 2020. 2 +[87] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2 +[88] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1 +[89] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9720-9731, 2024. 1, 2, 3 +[90] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1438-1447, 2019. 5 +[91] Chuanxia Zheng, Duy-Son Dao, Guoxian Song, Tat-Jen Cham, and Jianfei Cai. Visiting the invisible: Layer-by-layer completed scene decomposition. International Journal of Computer Vision (IJCV), 129:3195-3215, 2021. 2 +[92] Chuanxia Zheng, Tat-Jen Cham, Jianfei Cai, and Dinh Phung. Bridging global context interactions for high-fidelity image completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11512-11522, 2022. 4, 5 +[93] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 2 +[94] Xin-Yang Zheng, Hao Pan, Yu-Xiao Guo, Xin Tong, and Yang Liu. Mvd^2: Efficient multiview 3d reconstruction for multiview diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[95] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2 + +[96] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. Advances in neural information processing systems (NeurIPS), 31, 2018. 2 +[97] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1464-1472, 2017. 8 +[98] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2 \ No newline at end of file diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/images.zip b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..952d26dc94b36cfc22bb441c0010371263931fe9 --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee9ba3a7d3ac97e5f08b3c967dcf484a0fb68e990b1ba042d2bf520085d651aa +size 682694 diff --git a/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/layout.json b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a9077109ec6c14dcdc64f67f778fe410fb679ef --- /dev/null +++ b/2025/Amodal3R_ Amodal 3D Reconstruction from Occluded 2D Images/layout.json @@ -0,0 +1,14551 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 101, + 103, + 510, + 121 + ], + "type": "text", + "content": "Amodal3R: Amodal 3D Reconstruction from Occluded 2D Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "spans": [ + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": "Tianhao Wu" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": ", Chuanxia Zheng" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": ", Frank Guan" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": ", Andrea Vedaldi" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": ", Tat-Jen Cham" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{*}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": "S-Lab, " + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": "Nanyang Technological University, " + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": "VGG, University of Oxford \n" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 142, + 502, + 195 + ], + "type": "text", + "content": "Singapore Institute of Technology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 233, + 196, + 375, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 196, + 375, + 209 + ], + "spans": [ + { + "bbox": [ + 233, + 196, + 375, + 209 + ], + "type": "text", + "content": "https://sm0kywu.github.io/Amodal3R/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 232, + 197, + 322 + ], + "blocks": [ + { + "bbox": [ + 59, + 220, + 105, + 231 + ], + "lines": [ + { + "bbox": [ + 59, + 220, + 105, + 231 + ], + "spans": [ + { + "bbox": [ + 59, + 220, + 105, + 231 + ], + "type": "text", + "content": "3D Recast" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 232, + 197, + 322 + ], + "lines": [ + { + "bbox": [ + 61, + 232, + 197, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 197, + 322 + ], + "type": "image", + "image_path": "db0f3e260a940e93d176cfcacfda74c200ff73d2202082359e50a6d358d3a7e1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 322, + 110, + 331 + ], + "lines": [ + { + "bbox": [ + 64, + 322, + 110, + 331 + ], + "spans": [ + { + "bbox": [ + 64, + 322, + 110, + 331 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 118, + 322, + 190, + 331 + ], + "lines": [ + { + "bbox": [ + 118, + 322, + 190, + 331 + ], + "spans": [ + { + "bbox": [ + 118, + 322, + 190, + 331 + ], + "type": "text", + "content": "Stochastic sample 1" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 202, + 220, + 288, + 322 + ], + "blocks": [ + { + "bbox": [ + 202, + 220, + 288, + 322 + ], + "lines": [ + { + "bbox": [ + 202, + 220, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 202, + 220, + 288, + 322 + ], + "type": "image", + "image_path": "4be72a6742cbba6fb6e0aa6ec50793ba1faff5009bfb21fdf47d0948e8aeff5f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 322, + 282, + 331 + ], + "lines": [ + { + "bbox": [ + 208, + 322, + 282, + 331 + ], + "spans": [ + { + "bbox": [ + 208, + 322, + 282, + 331 + ], + "type": "text", + "content": "Stochastic sample 2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 293, + 232, + 377, + 315 + ], + "blocks": [ + { + "bbox": [ + 293, + 220, + 410, + 232 + ], + "lines": [ + { + "bbox": [ + 293, + 220, + 410, + 232 + ], + "spans": [ + { + "bbox": [ + 293, + 220, + 410, + 232 + ], + "type": "text", + "content": "In-the-wild Reconstruction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 293, + 232, + 377, + 315 + ], + "lines": [ + { + "bbox": [ + 293, + 232, + 377, + 315 + ], + "spans": [ + { + "bbox": [ + 293, + 232, + 377, + 315 + ], + "type": "image", + "image_path": "7042bf3f32f04a0978ae549a353e1b5a2825646847ab201d5bcb3d0a4fb37a3e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 320, + 362, + 331 + ], + "lines": [ + { + "bbox": [ + 315, + 320, + 362, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 362, + 331 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 377, + 227, + 463, + 319 + ], + "blocks": [ + { + "bbox": [ + 377, + 227, + 463, + 319 + ], + "lines": [ + { + "bbox": [ + 377, + 227, + 463, + 319 + ], + "spans": [ + { + "bbox": [ + 377, + 227, + 463, + 319 + ], + "type": "image", + "image_path": "e8b639ca0d9316b8f72ff13100dfe1cd2bf2e8bfe27f71d18e823960df0befc9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 320, + 460, + 331 + ], + "lines": [ + { + "bbox": [ + 388, + 320, + 460, + 331 + ], + "spans": [ + { + "bbox": [ + 388, + 320, + 460, + 331 + ], + "type": "text", + "content": "Stochastic sample 1" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 467, + 226, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 467, + 226, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 467, + 226, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 467, + 226, + 553, + 319 + ], + "type": "image", + "image_path": "4bd49bd69f8a811e2e55f5c2ce9daeebf959d908174fb4622f831ed364f0e8fa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 320, + 548, + 331 + ], + "lines": [ + { + "bbox": [ + 474, + 320, + 548, + 331 + ], + "spans": [ + { + "bbox": [ + 474, + 320, + 548, + 331 + ], + "type": "text", + "content": "Stochastic sample 2" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 66, + 346, + 135, + 415 + ], + "blocks": [ + { + "bbox": [ + 59, + 334, + 167, + 345 + ], + "lines": [ + { + "bbox": [ + 59, + 334, + 167, + 345 + ], + "spans": [ + { + "bbox": [ + 59, + 334, + 167, + 345 + ], + "type": "text", + "content": "3D Scene Decomposition" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 346, + 135, + 415 + ], + "lines": [ + { + "bbox": [ + 66, + 346, + 135, + 415 + ], + "spans": [ + { + "bbox": [ + 66, + 346, + 135, + 415 + ], + "type": "image", + "image_path": "1c125d3a9582a8e5d07e1a535e32ca82429b560e62da9f5848bd3c787846ec90.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 415, + 124, + 426 + ], + "lines": [ + { + "bbox": [ + 78, + 415, + 124, + 426 + ], + "spans": [ + { + "bbox": [ + 78, + 415, + 124, + 426 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. Example results of Amodal3R. Given partially visible objects within images (occluded regions are shown in black, visible areas in red outlines), our Amodal3R generates diverse semantically meaningful 3D assets with reasonable geometry and plausible appearance. We sample multiple times to get diverse results from the same occluded input. Trained on synthetic datasets, it generalizes well to out-of-domain test examples from real scenes, where most objects are partially visible, and reconstructs reasonable 3D assets." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 140, + 346, + 272, + 415 + ], + "blocks": [ + { + "bbox": [ + 140, + 346, + 272, + 415 + ], + "lines": [ + { + "bbox": [ + 140, + 346, + 272, + 415 + ], + "spans": [ + { + "bbox": [ + 140, + 346, + 272, + 415 + ], + "type": "image", + "image_path": "73906cc49696dba9ce937eb09543a55bea4527421a3d01dd6b8f693af5fc1014.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 277, + 346, + 410, + 415 + ], + "blocks": [ + { + "bbox": [ + 277, + 346, + 410, + 415 + ], + "lines": [ + { + "bbox": [ + 277, + 346, + 410, + 415 + ], + "spans": [ + { + "bbox": [ + 277, + 346, + 410, + 415 + ], + "type": "image", + "image_path": "22060cd08dc52324121605dc107f208dc0f6955396270030ef0833c53c03ddd9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 415, + 389, + 426 + ], + "lines": [ + { + "bbox": [ + 302, + 415, + 389, + 426 + ], + "spans": [ + { + "bbox": [ + 302, + 415, + 389, + 426 + ], + "type": "text", + "content": "Decomposed 3D assets" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 414, + 346, + 548, + 415 + ], + "blocks": [ + { + "bbox": [ + 414, + 346, + 548, + 415 + ], + "lines": [ + { + "bbox": [ + 414, + 346, + 548, + 415 + ], + "spans": [ + { + "bbox": [ + 414, + 346, + 548, + 415 + ], + "type": "image", + "image_path": "39b7ea3e1098b79d12fdccc7992e0db1a7fdd8ecd2c429d3b867541db7111446.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 151, + 491, + 200, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 491, + 200, + 502 + ], + "spans": [ + { + "bbox": [ + 151, + 491, + 200, + 502 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 54, + 507, + 297, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 507, + 297, + 697 + ], + "spans": [ + { + "bbox": [ + 54, + 507, + 297, + 697 + ], + "type": "text", + "content": "Most existing image-to-3D models assume that objects are fully visible, ignoring occlusions that commonly occur in real-world scenarios. In this paper, we introduce Amodal3R, a conditional image-to-3D model designed to reconstruct plausible 3D geometry and appearance from partial observations. We extend a \"foundation\" 3D generator by introducing a visible mask-weighted attention mechanism and an occlusion-aware attention layer that explicitly leverage visible and occlusion priors to guide the reconstruction process. We demonstrate that, by training solely on synthetic data, Amodal3R learns to recover full 3D objects even in the presence of occlusions in real scenes. It substantially outperforms state-of-the-art methods that independently perform 2D amodal completion followed by 3D reconstruction, thereby establishing a new benchmark for occlusion-aware 3D reconstruction." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 491, + 394, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 394, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 394, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 516, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 516, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 312, + 516, + 555, + 659 + ], + "type": "text", + "content": "Humans possess a remarkable ability to infer the complete 3D shape and appearance of objects from single views, even when those objects are partly hidden behind occluders. This ability, known as amodal completion, is key to operating in complex real-world environments, where objects are often partially occluded by surrounding clutter. However, existing image-based 3D reconstruction models [5, 6, 23, 41, 47, 48, 65, 70, 71, 75, 77, 80, 88, 89] fail to recover full 3D assets when the object is partially occluded. We thus consider the problem of reconstructing 3D objects from one or more partially occluded views, a novel task that we call amodal 3D reconstruction." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Amodal 3D reconstruction is a challenging task that requires reconstructing an object's photorealistic 3D geometry and appearance, including the occluded regions, which are inherently ambiguous. Previous approaches [7, 54] typ" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 116, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 116, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 116, + 712 + ], + "type": "text", + "content": "†Project Lead." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9181" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "ically decompose the task into a two-stage pipeline: first performing 2D amodal completion [85, 86, 91], followed by conventional 3D reconstruction [41, 75, 89]. While these two-stage pipelines are straightforward to implement, they have some drawbacks. First, 2D amodal completion methods rely predominantly on appearance-only priors, without incorporating explicit information on the 3D structure. This lack of 3D geometric reasoning limits their ability to generate physically plausible completions. Second, some 3D reconstruction methods depend on, or benefit from, multiview inputs. In this case, 2D amodal completion may lack multi-view consistency, particularly when performed independently for different views, which confuses the 3D reconstruction process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 295, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 428 + ], + "type": "text", + "content": "In this paper, we present Amodal3R, a novel single-stage paradigm for amodal 3D reconstruction that surpasses previous state-of-the-art approaches. Amodal3R fine-tunes a pre-trained 3D generator by introducing a dedicated branch that explicitly models occlusions. Its key advantage is performing reconstruction, completion, and occlusion reasoning directly in a 3D latent space rather than using a two-stage scheme. With this, the model can reconstruct both visible and occluded regions of the object coherently and plausibly. Specifically, we introduce a mask-weighted cross-attention to highlight visible information, along with a new occlusion-aware attention layer to guide completion in occluded areas. Together, these components enable the model to focus more on the visible parts of the object without significantly perturbing the pre-trained model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 434, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 531 + ], + "type": "text", + "content": "We designed several benchmarks to evaluate the performance of Amodal3R, including object-level Google Scanned Objects [16] and Toys4K [67] augmented with occlusions, 3D scenes from Replica [68], and real-world monocular images. Without relying on additional heuristics, Amodal3R achieves significantly superior performance compared to state-of-the-art models and generalizes well to different datasets, including real ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 538, + 274, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 538, + 274, + 550 + ], + "spans": [ + { + "bbox": [ + 67, + 538, + 274, + 550 + ], + "type": "text", + "content": "In summary, our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 558, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 558, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 593 + ], + "type": "text", + "content": "- We propose a feed-forward 3D reconstruction model that directly reconstructs complete and high-quality 3D objects from one or more partially occluded views." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 295, + 653 + ], + "type": "text", + "content": "- We introduce a mask-weighted cross-attention mechanism and an occlusion-aware layer to inject occlusion awareness into an existing 3D reconstruction model, improving both the geometry and appearance of reconstructed objects when they are partially occluded." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 654, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 712 + ], + "type": "text", + "content": "- We demonstrate, via experiments on the 3D scene dataset Replica and real-world images, that our one-stage pipeline significantly outperforms existing two-stage approaches, establishing a new benchmark for amodal 3D reconstruction." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 392, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 392, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 392, + 85 + ], + "type": "text", + "content": "2. Background" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 341 + ], + "type": "text", + "content": "2D Amodal Completion. Recent advances in 2D amodal completion [2, 54, 83, 85, 86, 91] have achieved significant success in reconstructing occluded regions of objects in 2D images. While these methods show promise for 3D generation pipelines [2, 54], they still have limitations. First, 2D amodal completion models have limited 3D geometric priors, as they interpret images as 2D patterns. While they excel at completing textures, they may generate physically implausible geometries when handling complex occlusions. This stems from their lack of explicit 3D reasoning and reliance on 2D appearance priors, without true volumetric understanding. Second, for models that use multi-view images as input, the results of 2D amodal completion are often inconsistent across views. Inconsistent views need to be handled by the 3D reconstructors, which causes confusion (see Sec. 4). Although there has been significant work on multiview consistent generation [7, 62, 70, 72, 81, 89, 94], multiview consistent completion is less explored. RenderDiffusion [1] contains an example, but the results are often blurry or lack details. This motivates our 3D-centric reconstruction framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 533 + ], + "type": "text", + "content": "3D Shape Completion. Several methods [12, 14, 69, 95] start from a partial 3D reconstruction, then complete it in 3D. They use encoder-decoder architectures [24] or volumetric representations [10, 29] to robustly recover 3D shape, but not texture, which is necessary in many applications. They also still require recovering the partial 3D geometry from an occluded image, a challenge in itself. Furthermore, they ignore the input image when completing the object in 3D, disregarding the occlusion pattern that caused the 3D geometry to be recovered only partially. Recent work [11, 82] utilizes the partially visible object as input specifically for 3D shape completion. However, it focuses solely on geometry reconstruction, without recovering the object's texture and appearance. In contrast, our approach is end-to-end, recovering the complete 3D shape and appearance of the object from the occluded image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "3D Generative Models. Early advancements in 3D generation are based on GANs [19], exploring various 3D representations such as point clouds [25, 39], voxel grids [84, 96], view sets [48, 55], NeRF [4, 5, 50, 60], SDF [18], and 3D Gaussian mixtures [78]. While GANs can capture complex 3D structures, they struggle with stability, scalability, and data diversity. The focus then shifted to diffusion models [22, 59, 66], which are more capable and versatile. They were first applied to novel view synthesis [76] in image space, before expanding to model a variety of 3D representations, including point clouds [44, 45, 49, 79, 93], voxel grids [37, 46], triplanes [64, 98], meshes [20, 43], and 3D Gaussian mixtures [9, 35, 38, 87]. Autoregressive models [8, 47, 65] have also been explored for mesh generation, focusing on improving the topology of the final 3D mesh." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9182" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 549, + 178 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 549, + 178 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 549, + 178 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 549, + 178 + ], + "type": "image", + "image_path": "ec17aea1d842cf4d5ac780498601b8d2b71d4395e183436383899c5837c417a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 180, + 550, + 277 + ], + "blocks": [ + { + "bbox": [ + 60, + 180, + 550, + 277 + ], + "lines": [ + { + "bbox": [ + 60, + 180, + 550, + 277 + ], + "spans": [ + { + "bbox": [ + 60, + 180, + 550, + 277 + ], + "type": "image", + "image_path": "3a7b54826228d84368a9b61c041874faaffaf8f69e5964d42e366eec13c32630.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "text", + "content": "Figure 2. Overview of Amodal3R. Given an image as input and the regions of interest, Amodal3R first extracts the partially visible target object, along with the visibility and occlusion masks using an off-the-shelf 2D segmenter. It then applies DINOV2 [53] to extract features " + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{dino}}" + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "text", + "content": " as additional conditioning for the 3D reconstructor. To enhance occlusion reasoning, each transformer block incorporates a mask-weighted cross-attention (via " + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{vis}}" + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "text", + "content": ") and occlusion-aware attention layer (via " + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{occ}}" + }, + { + "bbox": [ + 54, + 284, + 555, + 341 + ], + "type": "text", + "content": "), ensuring the 3D reconstructor accurately perceives visible information while effectively inferring occluded parts. For conditioning details, see Sec. 3.1." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 361, + 294, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 294, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 294, + 494 + ], + "type": "text", + "content": "An alternative to performing diffusion in 2D image space [62, 70, 89] is to use 3D latent spaces [28, 51, 58, 63, 74, 80]. Such methods typically consist of two stages: the first for generating geometry and the second for generating appearance, and are capable of high-quality imaged-to-3D generation. However, they assume that input objects are fully visible, which limits their application to real-world scenes, where occlusions are frequent. Here, we extend such approaches to generate high-quality 3D assets from occluded input images — a challenging task that requires inferring complete 3D objects from partial information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 112, + 514 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": "Consider an image " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": " containing a partially occluded view of an object of interest " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": ". We wish to reconstruct the complete 3D shape and appearance " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": " of the object " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": ". This task is conceptually similar to existing image-to-3D reconstruction, except that here the object is partially occluded instead of being fully visible in " + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 521, + 295, + 605 + ], + "type": "text", + "content": ". We call this problem amodal 3D reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "We introduce Amodal3R (Fig. 2), a new method for amodal 3D reconstruction. Unlike recent two-stage methods [7, 54] that first perform 2D amodal completion followed by 3D reconstruction, Amodal3R is an end-to-end occlusion-aware 3D reconstruction model that generates complete geometry and appearance directly within the 3D space. Formally, Amodal3R is a conditional generator " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "v_{\\theta}(y|o_i,M_{\\mathrm{vis}},M_{\\mathrm{occ}})" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " that takes as input the image " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", centered on a partially visible object " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ", as well as the visibility" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": "mask " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": " and occlusion mask " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": ". The visibility mask " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": " highlights the pixels in " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": " that correspond to object " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": " marks the pixels occupied by occluders, i.e., all other objects that potentially obscure " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": ". For real images, these masks can be efficiently obtained using pre-trained 2D segmentation models like Segment Anything [34]. By providing point coordinate prompts for the object of interest " + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 361, + 555, + 552 + ], + "type": "text", + "content": " and its occluders, the segmentation model can generate the corresponding masks. In cases where multiple occluders are present, or when occluders fragment the visible area of the target object, a sequential process is employed. Specifically, point prompts for each visible/occluding region are provided to the 2D segmenter one at a time, with the model generating an individual mask for each region. These masks are then aggregated to form comprehensive visibility and occlusion masks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 553, + 556, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 556, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 556, + 625 + ], + "type": "text", + "content": "The challenges for Amodal3R are how to: (1) produce a plausible and complete 3D shape and appearance from partial observations, even in the absence of multi-view information; and (2) ensure 3D consistency in terms of geometry and photometry, maintaining seamless visual coherence between visible and generated regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 635, + 515, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 515, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 515, + 647 + ], + "type": "text", + "content": "3.1. Mask-Conditional Generative Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "To fine-tune an image-to-3D model for partially occluded inputs, the key change is to condition the transformer " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " not only on the object of interest " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "o_i" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", but also on the visibility and occlusion masks, i.e., " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ". A naive approach is to patchify or embed the masks into tokens " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "(c_{\\mathrm{vis}}, c_{\\mathrm{occ}})" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9183" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 149, + 241 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 149, + 241 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 149, + 241 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 149, + 241 + ], + "type": "image", + "image_path": "87381d95c462752cdc81640281a904c781af6b72aae913533c761f2102a88737.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 243, + 134, + 253 + ], + "lines": [ + { + "bbox": [ + 74, + 243, + 134, + 253 + ], + "spans": [ + { + "bbox": [ + 74, + 243, + 134, + 253 + ], + "type": "text", + "content": "(a) TRELLIS [80]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 257, + 296, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 301 + ], + "type": "text", + "content": "Figure 3. The transformer structure of Amodal3R. Compared with the original TRELLIS [80] design, we further introduce the mask-weighted cross-attention and occlusion-aware layer. It applies to both sparse structure and SLAT diffusion models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 153, + 70, + 294, + 241 + ], + "blocks": [ + { + "bbox": [ + 153, + 70, + 294, + 241 + ], + "lines": [ + { + "bbox": [ + 153, + 70, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 153, + 70, + 294, + 241 + ], + "type": "image", + "image_path": "4b19bdfaaecc7a073d5f373b9f114d67bb5ce602a4bcad024930a9af08e96a53.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 243, + 263, + 253 + ], + "lines": [ + { + "bbox": [ + 191, + 243, + 263, + 253 + ], + "spans": [ + { + "bbox": [ + 191, + 243, + 263, + 253 + ], + "type": "text", + "content": "(b) Amodal3R (Ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": "which can be concatenated to the image tokens " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{e}_{\\mathrm{dino}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": " and processed by cross-attention as before. However, the RGB " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\pmb{o}_i" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": " contains significantly more information than the binary masks " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": ". The transformer, which is initially trained with RGB only, may simply ignore the mask information. This is compounded by the fact that learning to use this new information, which involves a new type of token that is incompatible with image tokens, may require aggressive fine-tuning of the transformer." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "content": "To sidestep this problem, inspired by masked attention approaches in 2D completion [92], we introduce mask-weighted cross-attention and an occlusion-aware attention layer to better utilize the visibility mask " + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "content": " and the occlusion mask " + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 502 + ], + "type": "text", + "content": " without disrupting the pre-trained 3D generator. These are described next." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "text", + "content": "Mask-weighted Cross-Attention. A key novel component of Amodal3R is mask-weighted cross-attention, which allows the model to focus its attention on the visible parts of the object. We implement this mechanism in each attention block in the transformers " + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "text", + "content": " of a pre-trained 3D generator. Given the latent tokens " + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathbb{R}^{L \\times C}" + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "text", + "content": " input to a cross-attention layer as well as the image features " + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{dino}} \\in \\mathbb{R}^{K \\times C'}" + }, + { + "bbox": [ + 55, + 502, + 296, + 609 + ], + "type": "text", + "content": " of the partially visible object, cross-attention computes the similarity score matrix via" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 617, + 295, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 617, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 67, + 617, + 295, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} = W _ {q} \\boldsymbol {\\ell}, \\quad [ \\boldsymbol {k}, \\boldsymbol {v} ] = W _ {k v} \\boldsymbol {c} _ {\\mathrm {d i n o}}, \\quad S = \\boldsymbol {q} \\boldsymbol {k} ^ {\\top} / \\sqrt {D}, \\tag {1}", + "image_path": "a5a546756878afc656bd36f61d96851f637a4c1bb06b925f172baf4c0c8ecbe2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "W_{q}" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "W_{kv}" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " are learnable projections that map the latents " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " to the query " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": ", and the conditioning image feature " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{dino}}" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " to the key " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " and the value " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": ", respectively, while " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " is the dimension of the query and key vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We wish to bias the token similarity matrix " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{L \\times K}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " towards the visible parts of the object. Recall that " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 69, + 356, + 125 + ], + "blocks": [ + { + "bbox": [ + 325, + 69, + 356, + 125 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 356, + 125 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 356, + 125 + ], + "type": "image", + "image_path": "e2d2a57d59bd8efdf730ab223444dac9787b762c1c812448af756fe46115ecca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 137, + 555, + 193 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 193 + ], + "type": "text", + "content": "Figure 4. 3D-consistent mask example. Given a 3D mesh, we render selected triangles in a distinct color from the others to generate multi-view consistent masks. It allows the evaluation of multi-view methods in handling contact occlusion. (The occluded regions are shown in red.)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 381, + 71, + 412, + 120 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 412, + 120 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 412, + 120 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 412, + 120 + ], + "type": "image", + "image_path": "211eab21b339562a7b9734004b576c27eb64f521933ae646676588415d5a3267.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 72, + 453, + 119 + ], + "blocks": [ + { + "bbox": [ + 425, + 72, + 453, + 119 + ], + "lines": [ + { + "bbox": [ + 425, + 72, + 453, + 119 + ], + "spans": [ + { + "bbox": [ + 425, + 72, + 453, + 119 + ], + "type": "image", + "image_path": "6e1959f584b0d93c348d0d6641a7cac5286305a31fe747f49260119ea0f248a2.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 466, + 72, + 493, + 121 + ], + "blocks": [ + { + "bbox": [ + 466, + 72, + 493, + 121 + ], + "lines": [ + { + "bbox": [ + 466, + 72, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 466, + 72, + 493, + 121 + ], + "type": "image", + "image_path": "3c82569277c7393f8b8bce254ed494b5e4abb07f88512c3fdd1a8e86563b580f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 508, + 72, + 536, + 117 + ], + "blocks": [ + { + "bbox": [ + 508, + 72, + 536, + 117 + ], + "lines": [ + { + "bbox": [ + 508, + 72, + 536, + 117 + ], + "spans": [ + { + "bbox": [ + 508, + 72, + 536, + 117 + ], + "type": "image", + "image_path": "932a594172e1a1eec2ea793d674805e1bcd98db4e07ceba4702b4ebfc7c48e12.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": "is the number of tokens " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{dino}}" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": " extracted by DINOv2 [52] from the occluded image " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "x \\cdot M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": ". Each token thus corresponds to a " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": " patch in the input image (where " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "P = 14" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": "). We extract analogous flattened 1-D tokens " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{vis}} = [\\rho(M_{\\mathrm{vis},1}), \\dots, \\rho(M_{\\mathrm{vis},K})]" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": " from the visibility mask by calculating the fraction of " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": " pixels that are visible in the " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": "-th image patch " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis},j}" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{vis},j} = \\rho(M_{\\mathrm{vis},j}) \\in [0,1]" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": ". We then use these quantities to bias the computation of the attention matrix " + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "inline_equation", + "content": "A = \\text{softmax}(S + \\log c_{\\mathrm{vis}}) \\in [0,1]^{L \\times M}" + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": " via broadcasting. Hence:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 373, + 348, + 554, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 348, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 373, + 348, + 554, + 377 + ], + "type": "interline_equation", + "content": "A _ {i j} = \\frac {\\boldsymbol {c} _ {\\text {v i s} , j} \\exp \\left(S _ {i j}\\right)}{\\sum_ {k = 1} ^ {K} \\boldsymbol {c} _ {\\text {v i s} , k} \\exp \\left(S _ {i k}\\right)}. \\tag {2}", + "image_path": "f1675eeb565e2d24c86aa26ee02de33586ac9086d6157ffbb1ac8923a87b7de8.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": "In this manner, the visibility flag modulates the attention matrix " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": " smoothly. If there are no visible pixels in a patch " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "inline_equation", + "content": "A_{ij} = 0" + }, + { + "bbox": [ + 313, + 387, + 554, + 471 + ], + "type": "text", + "content": ", so the corresponding image tokens are skipped in cross-attention. While we have illustrated how this works for a single head, in practice Amodal3R uses a multi-head transformer, to which Eq. (2) extends trivially. Please see the supplementary materials for more details." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "content": "Occlusion-aware Attention Layer. To prevent the network from arbitrarily extending the geometry beyond the occluded regions, we further introduce an occlusion-aware attention layer. The key insight is that specifying only visible information is insufficient; it is equally important to distinguish foreground occluders from the background, as this explicitly identifies the potential regions requiring completion. Namely, when a pixel is marked as invisible in the mask " + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "content": ", this might be (i) because there is an occluder in front of that pixel (so the pixel could plausibly belong to the object of interest if the occlusion were removed), or (ii) because the pixel is entirely off the object. This information is encoded by the mask " + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 473, + 554, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "To help the model distinguish between visible, occluded, and background areas, we add one more cross-attention layer placed immediately after the mask-weighted cross-attention layer, and pool solely the occlusion mask " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ". We encode the occlusion mask " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " as a sequence of flattened 1-D tokens " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{occ}} = [\\rho(M_{\\mathrm{occ},1}), \\ldots, \\rho(M_{\\mathrm{occ},M})]" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ", and then simply pool " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " using a cross-attention layer." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9184" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 218, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 218, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 218, + 85 + ], + "type": "text", + "content": "3.2. Simulating Occluded 3D Data" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 294, + 137 + ], + "type": "text", + "content": "To train and evaluate our model, we require a dataset of 3D assets imaged in scenarios with clutter and occlusions. We curated a large synthetic 3D dataset to train our model, as collecting such 3D assets in the real world is challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": "Random 2D Occlusions. Each training sample " + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "inline_equation", + "content": "(o, x, M_{\\mathrm{vis}}, M_{\\mathrm{occ}})" + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": " consists of a 3D object " + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": " (from which ground truth latents can be obtained by using the encoders of Appendix A.1), an image " + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": " with partial occlusions, and corresponding visibility and occlusion masks " + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 137, + 295, + 304 + ], + "type": "text", + "content": ". In a real scenario, occlusions arise from other objects in the scene. In a multi-view setting, like the one discussed below, these occlusions must be consistent across views, reflecting the underlying scene geometry. However, because our model is trained for single-view reconstruction, we can simulate occlusions by randomly masking parts of the object after rendering it in 2D. This is simpler and allows us to generate fresh occlusion patterns each time a view is sampled for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": "Thus, given " + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": " and an image " + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": " rendered from a random viewpoint, we generate random occlusion masks as follows. Inspired by work on 2D completion [26, 40, 56, 90, 92], we randomly place lines, ellipses, and rectangles, simulating diverse masking patterns. The union of these random shapes gives us the occlusion mask " + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": ". Then, if " + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{obj}}" + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": " is the mask of the unoccluded object, the visible mask is given by " + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}} = M_{\\mathrm{obj}} \\odot (1 - M_{\\mathrm{occ}})" + }, + { + "bbox": [ + 55, + 305, + 295, + 412 + ], + "type": "text", + "content": ". Examples of such patterns are given in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "content": "3D-consistent Occlusions. In a real scene imaged from multiple views, occlusions are not random but geometrically consistent across different viewpoints, as they result from other objects. This is especially true for contact occlusions, where parts of an object remain occluded by other objects from all viewpoints. To evaluate the model's performance under such challenging conditions, 3D-consistent masks are required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 605 + ], + "type": "text", + "content": "To efficiently generate such contact occlusion masks, we leverage 3D meshes during rendering. Starting from a randomly chosen triangle, we apply a random-walk strategy to iteratively select neighboring triangles, forming continuous occluded regions until the predefined mask ratio is met. By rendering these masked meshes using the same camera parameters as the RGB images, we ensure multi-view consistency in the generated occlusion masks (see Fig. 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 612, + 256, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 256, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 256, + 625 + ], + "type": "text", + "content": "3.3. Reconstruction with Multi-view Input" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Since our flow-based model performs multiple denoising steps and does not require known camera poses for input views, Amodal3R can benefit from multi-view reference images " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " as conditioning inputs at different steps of the denoising process. While multi-view conditioning naturally improves reconstruction performance, a potential issue is that an image used earlier in the denoising process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "has a stronger influence on the final 3D geometry. This is because early denoising steps establish the coarse geometry, whereas later steps refine higher-frequency details [32, 33]. Therefore, we prioritize input images based on their visibility. Specifically, when experimenting with multi-view inputs, we sort the images in proportion to the object visibility " + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "inline_equation", + "content": "|M_{\\mathrm{vis}}|" + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": ", ensuring that images with higher visibility are used as primary inputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 186, + 395, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 395, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 395, + 200 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 208, + 432, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 432, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 432, + 221 + ], + "type": "text", + "content": "4.1. Experiment Settings" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "type": "text", + "content": "Datasets. Amodal3R is trained on a combination of 3D synthetic datasets: 3D-FUTURE (9,472 objects [17]), ABO (4,485 objects [13]), and HSSD (6,670 objects [31]). Once trained, we first assessed its effectiveness on Toys4K (randomly sampling 1,500 objects [67]) and Google Scanned Objects (GSO) (1,030 objects [16]), which were excluded from the training data for both our model and the baselines. During inference, a 3D-consistent mask occludes each input object, and each view is augmented with additional randomly generated occlusion areas. This ensures that the model cannot simply extract the region to be completed from the occlusion masks. To further assess out-of-domain generalization in practical applications, we also evaluated all models on the 3D scene dataset Replica [68] as well as on in-the-wild images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 410, + 554, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 554, + 541 + ], + "type": "text", + "content": "Metrics. To measure the quality of the rendered images, we use the Fréchet Inception Distance (FID) [21] and the Kernel Inception Distance (KID) [3]. To assess the quality of the reconstructed 3D geometry, we use the Coverage Score (COV), the Point cloud FID (P-FID) [49], and the Minimum Matching Distance (MMD) using the Chamfer Distance (CD). COV measures the diversity of the results, while P-FID and MMD measure the quality of the 3D reconstruction. We also use the CLIP score [57] to evaluate the consistency between each generated and ground-truth object pair." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 639 + ], + "type": "text", + "content": "Baselines. Most image-to-3D generators are trained on fully visible inputs. To ensure fair comparison, we completed the partially visible 2D objects before passing them to 3D generators. In particular, we use pix2gestalt [54], a state-of-the-art 2D amodal completion network. Using this protocol, we compared Amodal3R to state-of-the-art methods such as TRELLIS [80], Real3D [27], GaussianAnything [36]ICLR'25,and LaRa [6]ECCV'24." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 712 + ], + "type": "text", + "content": "Implementation Details. Amodal3R is trained on 4 A100 GPUs (40G) for 20K steps with a batch size of 16, taking approximately one day. We select TRELLIS [80] as underlying model because it is the latest 3D generator and is publicly available. However, our method is not specific to this model." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9185" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 213 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 213 + ], + "type": "table", + "html": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]92.261.300.7434.6935.925.03
Real3D [27]1pix2gestalt [54]91.212.020.7523.9219.619.21
TRELLIS [80]1pix2gestalt [54]58.825.870.7626.4331.654.17
Amodal3R (Ours)1—30.640.350.817.6939.613.62
LaRa [6]4pix2gestalt [54]172.844.540.7066.3424.568.11
LaRa [6]4pix2gestalt [54]+MV97.532.630.7521.8026.218.61
TRELLIS [80]4pix2gestalt [54]65.696.920.7824.6432.334.26
TRELLIS [80]4pix2gestalt [54]+MV60.371.850.8319.6831.754.21
Amodal3R (Ours)4—26.270.220.845.0338.743.61
", + "image_path": "e6a50c551235f4381cb26706eb19bfd086f8d3620ec0937e5dc58f841e641cd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 281, + 555, + 426 + ], + "blocks": [ + { + "bbox": [ + 55, + 216, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 555, + 271 + ], + "type": "text", + "content": "Table 1. Amodal 3D reconstruction results on GSO [16]. Here, V-num denotes the number of input views, and 2D-Comp refers to the 2D completion method. For single-view image-to-3D, we first complete occluded objects using pix2gestalt [54] before passing them to various 3D models. For multi-view image-to-3D, we adopt two variants: (1) pix2gestalt [54] is applied independently on each view; (2) pix2gestalt " + }, + { + "bbox": [ + 55, + 216, + 555, + 271 + ], + "type": "inline_equation", + "content": "[54] + \\mathrm{MV}" + }, + { + "bbox": [ + 55, + 216, + 555, + 271 + ], + "type": "text", + "content": ": a single-view completion is generated first, followed by multi-view diffusion [61] to synthesize 4 views as inputs. Without bells and whistles, Amodal3R consistently outperforms state-of-the-art models across all evaluation metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 281, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 58, + 281, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 58, + 281, + 555, + 426 + ], + "type": "table", + "html": "
MethodV-num2D-CompAppearanceGeometry
FID ↓KID(%) ↓CLIP ↑P-FID ↓COV(%) ↑MMD(‰) ↓
GaussianAnything [36]1pix2gestalt [54]57.171.220.8021.9733.567.23
Real3D [27]1pix2gestalt [54]59.921.630.7923.3124.359.53
TRELLIS [80]1pix2gestalt [54]43.056.830.8026.0426.286.87
Amodal3R (Ours)1—23.450.420.835.0037.095.89
LaRa [6]4pix2gestalt [54]123.523.610.7545.9127.899.67
LaRa [6]4pix2gestalt [54]+MV75.334.140.8013.0024.8210.93
TRELLIS [80]4pix2gestalt [54]46.348.770.8128.7625.357.13
TRELLIS [80]4pix2gestalt [54]+MV43.007.530.8124.4126.557.05
Amodal3R (Ours)4—20.930.500.853.7839.035.75
", + "image_path": "1d2cb8eaebe6634f7469d32f79bd74c40b9f40b6b49a1e79a776b6a95949543f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 429, + 542, + 440 + ], + "lines": [ + { + "bbox": [ + 66, + 429, + 542, + 440 + ], + "spans": [ + { + "bbox": [ + 66, + 429, + 542, + 440 + ], + "type": "text", + "content": "Table 2. Amodal 3D reconstruction results on Toys4K [67]. The experimental setting is the same to Tab. 1, except for the dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 60, + 450, + 108, + 531 + ], + "blocks": [ + { + "bbox": [ + 60, + 450, + 108, + 531 + ], + "lines": [ + { + "bbox": [ + 60, + 450, + 108, + 531 + ], + "spans": [ + { + "bbox": [ + 60, + 450, + 108, + 531 + ], + "type": "image", + "image_path": "214c14a2442947413628d675f0dc806eac9a9b619143d2876fd720ddb04267b6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 533, + 97, + 542 + ], + "lines": [ + { + "bbox": [ + 78, + 533, + 97, + 542 + ], + "spans": [ + { + "bbox": [ + 78, + 533, + 97, + 542 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 113, + 453, + 148, + 530 + ], + "blocks": [ + { + "bbox": [ + 113, + 453, + 148, + 530 + ], + "lines": [ + { + "bbox": [ + 113, + 453, + 148, + 530 + ], + "spans": [ + { + "bbox": [ + 113, + 453, + 148, + 530 + ], + "type": "image", + "image_path": "5d09515060d8f065dc68a83ff275e786b139d6dc6bf88deeda9e4bd95c9b5c0b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 533, + 160, + 542 + ], + "lines": [ + { + "bbox": [ + 109, + 533, + 160, + 542 + ], + "spans": [ + { + "bbox": [ + 109, + 533, + 160, + 542 + ], + "type": "text", + "content": "pix2gestalt [54]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 161, + 453, + 202, + 530 + ], + "blocks": [ + { + "bbox": [ + 161, + 453, + 202, + 530 + ], + "lines": [ + { + "bbox": [ + 161, + 453, + 202, + 530 + ], + "spans": [ + { + "bbox": [ + 161, + 453, + 202, + 530 + ], + "type": "image", + "image_path": "42865b2963f051b0c2bff5ffc35995faa5cab09351613b701818ec6602ab8fd8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 546, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 546, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 555, + 568 + ], + "type": "text", + "content": "Figure 5. Single-view amodal 3D reconstruction. The occlusion regions are shown in black and the visible regions are highlighted with red outlines. More examples are provided in supplementary material Fig. D.4" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 205, + 453, + 246, + 530 + ], + "blocks": [ + { + "bbox": [ + 168, + 533, + 244, + 542 + ], + "lines": [ + { + "bbox": [ + 168, + 533, + 244, + 542 + ], + "spans": [ + { + "bbox": [ + 168, + 533, + 244, + 542 + ], + "type": "text", + "content": "GaussianAnything [36]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 453, + 246, + 530 + ], + "lines": [ + { + "bbox": [ + 205, + 453, + 246, + 530 + ], + "spans": [ + { + "bbox": [ + 205, + 453, + 246, + 530 + ], + "type": "image", + "image_path": "c2763fdf1c66b93cc21539d425558ff93e45d8f13c0c5da52e3e678eff76f7a1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 257, + 453, + 295, + 530 + ], + "blocks": [ + { + "bbox": [ + 257, + 453, + 295, + 530 + ], + "lines": [ + { + "bbox": [ + 257, + 453, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 257, + 453, + 295, + 530 + ], + "type": "image", + "image_path": "cbdc468da5e748eb5f62a19b180df9909d73e5150ed73d6a08080db0e2afb4ae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 532, + 323, + 541 + ], + "lines": [ + { + "bbox": [ + 281, + 532, + 323, + 541 + ], + "spans": [ + { + "bbox": [ + 281, + 532, + 323, + 541 + ], + "type": "text", + "content": "Real3D [27]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 301, + 453, + 339, + 529 + ], + "blocks": [ + { + "bbox": [ + 301, + 453, + 339, + 529 + ], + "lines": [ + { + "bbox": [ + 301, + 453, + 339, + 529 + ], + "spans": [ + { + "bbox": [ + 301, + 453, + 339, + 529 + ], + "type": "image", + "image_path": "9cd07fa7f5d6b40ebaebb2d0c5ae2177b0dedaa1f1b4a8d9716bf88166a01b7d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 359, + 454, + 388, + 527 + ], + "blocks": [ + { + "bbox": [ + 359, + 454, + 388, + 527 + ], + "lines": [ + { + "bbox": [ + 359, + 454, + 388, + 527 + ], + "spans": [ + { + "bbox": [ + 359, + 454, + 388, + 527 + ], + "type": "image", + "image_path": "0ed7d60ced0c102b5269e611f751e0e04ab271d3a3ece8cefb85e6d32796eafd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 532, + 424, + 541 + ], + "lines": [ + { + "bbox": [ + 375, + 532, + 424, + 541 + ], + "spans": [ + { + "bbox": [ + 375, + 532, + 424, + 541 + ], + "type": "text", + "content": "TRELLIS [80]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 454, + 435, + 525 + ], + "blocks": [ + { + "bbox": [ + 395, + 454, + 435, + 525 + ], + "lines": [ + { + "bbox": [ + 395, + 454, + 435, + 525 + ], + "spans": [ + { + "bbox": [ + 395, + 454, + 435, + 525 + ], + "type": "image", + "image_path": "379e0e796dc1c2e971bd5e5916cb2f25f540ccfed2e459e2df28035bbc850c2f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 453, + 452, + 492, + 529 + ], + "blocks": [ + { + "bbox": [ + 453, + 452, + 492, + 529 + ], + "lines": [ + { + "bbox": [ + 453, + 452, + 492, + 529 + ], + "spans": [ + { + "bbox": [ + 453, + 452, + 492, + 529 + ], + "type": "image", + "image_path": "837c8a94c76ee900b4119ba1434ef012529765f892c9fef8e3d7a31849774486.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 532, + 522, + 541 + ], + "lines": [ + { + "bbox": [ + 462, + 532, + 522, + 541 + ], + "spans": [ + { + "bbox": [ + 462, + 532, + 522, + 541 + ], + "type": "text", + "content": "Amodal3R (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 589, + 140, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 140, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 140, + 600 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Results. We compare Amodal3R with prior methods on amodal 3D reconstruction in Tabs. 1 and 2. Amodal3R significantly outperforms the baselines across all metrics in both single- and multi-view settings, demonstrating its effectiveness. Notably, two-stage methods using multiple views (\"4 V-num + pix2gestalt\") perform worse than their single-view counterparts (\"1 V-num + pix2gestalt\"). This shows that inconsistent 2D completion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 590, + 554, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 625 + ], + "type": "text", + "content": "degrades 3D reconstruction (Sec. 2). In contrast, Amodal3R avoids this issue by operating directly in 3D space and benefits from additional occluded views as expected." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Results. Qualitative results are shown in Figs. 5 and 6 and in the supplementary material Figs. D.4 and D.5. For all baselines, pix2gestalt is first applied for 2D amodal completion (second column), and the corresponding completed images are passed to each baseline image-to-3D generator. Amodal3R consistently produces high-quality 3D assets under both single- and multi-view settings, even" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9186" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 121, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 121, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 121, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 121, + 150 + ], + "type": "image", + "image_path": "12012618131c21e5dab9e6f613af87cc1a0dcbc2136a2a7bf79f8f68b8228445.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 151, + 97, + 159 + ], + "lines": [ + { + "bbox": [ + 78, + 151, + 97, + 159 + ], + "spans": [ + { + "bbox": [ + 78, + 151, + 97, + 159 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 70, + 171, + 145 + ], + "blocks": [ + { + "bbox": [ + 129, + 70, + 171, + 145 + ], + "lines": [ + { + "bbox": [ + 129, + 70, + 171, + 145 + ], + "spans": [ + { + "bbox": [ + 129, + 70, + 171, + 145 + ], + "type": "image", + "image_path": "a1947c7a55137f0e34a70b742a96e9cf6f01a1ef8656f708836959fa26593cdf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 151, + 175, + 160 + ], + "lines": [ + { + "bbox": [ + 124, + 151, + 175, + 160 + ], + "spans": [ + { + "bbox": [ + 124, + 151, + 175, + 160 + ], + "type": "text", + "content": "pix2gestalt [54]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 237, + 102 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 237, + 102 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 237, + 102 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 237, + 102 + ], + "type": "image", + "image_path": "45d750e591eb5be7d2bce1b5c60b0ac898451b57f911f0c101146bdfc512c33e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 190, + 110, + 233, + 143 + ], + "blocks": [ + { + "bbox": [ + 190, + 110, + 233, + 143 + ], + "lines": [ + { + "bbox": [ + 190, + 110, + 233, + 143 + ], + "spans": [ + { + "bbox": [ + 190, + 110, + 233, + 143 + ], + "type": "image", + "image_path": "45897ab0ddaa06acd8c18b9318729497506e1506fcced3531a81b2f89c09532a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 151, + 258, + 160 + ], + "lines": [ + { + "bbox": [ + 228, + 151, + 258, + 160 + ], + "spans": [ + { + "bbox": [ + 228, + 151, + 258, + 160 + ], + "type": "text", + "content": "LaRa [6]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 245, + 76, + 290, + 99 + ], + "blocks": [ + { + "bbox": [ + 245, + 76, + 290, + 99 + ], + "lines": [ + { + "bbox": [ + 245, + 76, + 290, + 99 + ], + "spans": [ + { + "bbox": [ + 245, + 76, + 290, + 99 + ], + "type": "image", + "image_path": "f199b91890a543943aee5d14751d56d92d35f08a66e6fcfe03090295c0d965b0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 245, + 113, + 290, + 143 + ], + "blocks": [ + { + "bbox": [ + 245, + 113, + 290, + 143 + ], + "lines": [ + { + "bbox": [ + 245, + 113, + 290, + 143 + ], + "spans": [ + { + "bbox": [ + 245, + 113, + 290, + 143 + ], + "type": "image", + "image_path": "169fbf26b83e089ce7fe23f2286a3902a20d903d4affd660041b5ed69ddad98c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 72, + 359, + 106 + ], + "blocks": [ + { + "bbox": [ + 314, + 72, + 359, + 106 + ], + "lines": [ + { + "bbox": [ + 314, + 72, + 359, + 106 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 359, + 106 + ], + "type": "image", + "image_path": "6554d658abf7a597ba389b01081b18ecf8e82f2bbc113efa95bca9da7f768f63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 109, + 358, + 143 + ], + "blocks": [ + { + "bbox": [ + 315, + 109, + 358, + 143 + ], + "lines": [ + { + "bbox": [ + 315, + 109, + 358, + 143 + ], + "spans": [ + { + "bbox": [ + 315, + 109, + 358, + 143 + ], + "type": "image", + "image_path": "ff823111c9b70405d6c45af704cfa0a836289b2b3850deced980ff5166b7e264.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 151, + 392, + 160 + ], + "lines": [ + { + "bbox": [ + 343, + 151, + 392, + 160 + ], + "spans": [ + { + "bbox": [ + 343, + 151, + 392, + 160 + ], + "type": "text", + "content": "TRELLIS [80]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 372, + 72, + 414, + 104 + ], + "blocks": [ + { + "bbox": [ + 372, + 72, + 414, + 104 + ], + "lines": [ + { + "bbox": [ + 372, + 72, + 414, + 104 + ], + "spans": [ + { + "bbox": [ + 372, + 72, + 414, + 104 + ], + "type": "image", + "image_path": "e531f073587deb24f188aa5a57000be49699f533947ee2fb2bb1b68528d8ccad.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 372, + 112, + 416, + 143 + ], + "blocks": [ + { + "bbox": [ + 372, + 112, + 416, + 143 + ], + "lines": [ + { + "bbox": [ + 372, + 112, + 416, + 143 + ], + "spans": [ + { + "bbox": [ + 372, + 112, + 416, + 143 + ], + "type": "image", + "image_path": "f7403491e87d1bd5e14a6ed2d3ea06cafb71daf13387f7a90236d248aeb5ba3a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 438, + 72, + 485, + 106 + ], + "blocks": [ + { + "bbox": [ + 438, + 72, + 485, + 106 + ], + "lines": [ + { + "bbox": [ + 438, + 72, + 485, + 106 + ], + "spans": [ + { + "bbox": [ + 438, + 72, + 485, + 106 + ], + "type": "image", + "image_path": "6790e50b47b03e37fc50e00599c442571ea9baf83c3f4ff8960de4f09a771123.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 438, + 107, + 485, + 147 + ], + "blocks": [ + { + "bbox": [ + 438, + 107, + 485, + 147 + ], + "lines": [ + { + "bbox": [ + 438, + 107, + 485, + 147 + ], + "spans": [ + { + "bbox": [ + 438, + 107, + 485, + 147 + ], + "type": "image", + "image_path": "d214b6b7760f05c0c3c52637bc5ff6baca125c5df226f7fa47a0f85b78a853d8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 151, + 522, + 159 + ], + "lines": [ + { + "bbox": [ + 462, + 151, + 522, + 159 + ], + "spans": [ + { + "bbox": [ + 462, + 151, + 522, + 159 + ], + "type": "text", + "content": "Amodal3R (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 555, + 198 + ], + "type": "text", + "content": "Figure 6. Multi-view amodal 3D reconstruction. The above results are reconstructed using 4 occluded input views. Due to limited space, we present only the best results for LaRa and TRELLIS under the \"pix2gestalt+MV\" setting. We apply 3d-consistent mask and random extended occlusion areas. More examples are provided in supplementary material Fig. D.5." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 500, + 73, + 541, + 106 + ], + "blocks": [ + { + "bbox": [ + 500, + 73, + 541, + 106 + ], + "lines": [ + { + "bbox": [ + 500, + 73, + 541, + 106 + ], + "spans": [ + { + "bbox": [ + 500, + 73, + 541, + 106 + ], + "type": "image", + "image_path": "1cb5a8c474a13b4076e5c3119e6d82d418f88a69167b4365b548debcdde86cf9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 498, + 108, + 544, + 144 + ], + "blocks": [ + { + "bbox": [ + 498, + 108, + 544, + 144 + ], + "lines": [ + { + "bbox": [ + 498, + 108, + 544, + 144 + ], + "spans": [ + { + "bbox": [ + 498, + 108, + 544, + 144 + ], + "type": "image", + "image_path": "a7498ce14f93ae2c932cd88d4ba37575e8adf7627855d1c507d76e6f6ee88b53.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 56, + 216, + 98, + 333 + ], + "blocks": [ + { + "bbox": [ + 56, + 216, + 98, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 216, + 98, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 98, + 333 + ], + "type": "image", + "image_path": "55be6d9d173b001d1b42e5a16a60ea950a5815f07e0e6e1da5032576fc55f759.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 335, + 82, + 344 + ], + "lines": [ + { + "bbox": [ + 64, + 335, + 82, + 344 + ], + "spans": [ + { + "bbox": [ + 64, + 335, + 82, + 344 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 95, + 330, + 141, + 346 + ], + "lines": [ + { + "bbox": [ + 95, + 330, + 141, + 346 + ], + "spans": [ + { + "bbox": [ + 95, + 330, + 141, + 346 + ], + "type": "text", + "content": "' Naive concatenation" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 350, + 295, + 395 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 395 + ], + "type": "text", + "content": "Figure 7. Ablation examples. Naive concatenation fails to reconstruct complete shape and appearance. Mask-weighted attention alone extends geometry into background regions, while occlusion-aware attention alone cannot guarantee photorealistic appearance." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 102, + 222, + 138, + 255 + ], + "blocks": [ + { + "bbox": [ + 102, + 222, + 138, + 255 + ], + "lines": [ + { + "bbox": [ + 102, + 222, + 138, + 255 + ], + "spans": [ + { + "bbox": [ + 102, + 222, + 138, + 255 + ], + "type": "image", + "image_path": "535de559e9e64624d482c201449869b4f691749929ca466b2ba824ad8392d7dd.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 101, + 265, + 139, + 287 + ], + "blocks": [ + { + "bbox": [ + 101, + 265, + 139, + 287 + ], + "lines": [ + { + "bbox": [ + 101, + 265, + 139, + 287 + ], + "spans": [ + { + "bbox": [ + 101, + 265, + 139, + 287 + ], + "type": "image", + "image_path": "b03a90ec60dde9fb4c280130c6844cb500af15892bb7bb790a67fde2001adbc0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 105, + 297, + 135, + 327 + ], + "blocks": [ + { + "bbox": [ + 105, + 297, + 135, + 327 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 135, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 135, + 327 + ], + "type": "image", + "image_path": "2d24585d24a0d43522a7ee702bde2614e7b12ab0d70c736989d762513479e42c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 146, + 216, + 190, + 260 + ], + "blocks": [ + { + "bbox": [ + 146, + 216, + 190, + 260 + ], + "lines": [ + { + "bbox": [ + 146, + 216, + 190, + 260 + ], + "spans": [ + { + "bbox": [ + 146, + 216, + 190, + 260 + ], + "type": "image", + "image_path": "2fd0b4a2bc4cc74d8d9fe4607d26c742c3bdd95e9600d43431d943dd377f8034.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 146, + 262, + 192, + 289 + ], + "blocks": [ + { + "bbox": [ + 146, + 262, + 192, + 289 + ], + "lines": [ + { + "bbox": [ + 146, + 262, + 192, + 289 + ], + "spans": [ + { + "bbox": [ + 146, + 262, + 192, + 289 + ], + "type": "image", + "image_path": "ad19ffd4eb033e8daa9763f2619f2d325b35208ea5f8d5c191acee1f40db0150.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 156, + 299, + 183, + 327 + ], + "blocks": [ + { + "bbox": [ + 156, + 299, + 183, + 327 + ], + "lines": [ + { + "bbox": [ + 156, + 299, + 183, + 327 + ], + "spans": [ + { + "bbox": [ + 156, + 299, + 183, + 327 + ], + "type": "image", + "image_path": "6d3cca3cb212f5ab952596e9352fc814ed08e754a914d31f28c9473b84f92b40.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 328, + 189, + 346 + ], + "lines": [ + { + "bbox": [ + 146, + 328, + 189, + 346 + ], + "spans": [ + { + "bbox": [ + 146, + 328, + 189, + 346 + ], + "type": "text", + "content": "w/only " + }, + { + "bbox": [ + 146, + 328, + 189, + 346 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{vis}}" + }, + { + "bbox": [ + 146, + 328, + 189, + 346 + ], + "type": "text", + "content": " attention" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 203, + 223, + 236, + 255 + ], + "blocks": [ + { + "bbox": [ + 203, + 223, + 236, + 255 + ], + "lines": [ + { + "bbox": [ + 203, + 223, + 236, + 255 + ], + "spans": [ + { + "bbox": [ + 203, + 223, + 236, + 255 + ], + "type": "image", + "image_path": "7ffa1cfdc0ab7af2637d775a87d8c18e8e912644e517147ce22c31d6bbc67675.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 202, + 266, + 237, + 287 + ], + "blocks": [ + { + "bbox": [ + 202, + 266, + 237, + 287 + ], + "lines": [ + { + "bbox": [ + 202, + 266, + 237, + 287 + ], + "spans": [ + { + "bbox": [ + 202, + 266, + 237, + 287 + ], + "type": "image", + "image_path": "98cd5d400fbaed8d15d5ddd5f788da08cafdf1f1c12993f866dcc5a8feb294c6.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 207, + 299, + 235, + 327 + ], + "blocks": [ + { + "bbox": [ + 207, + 299, + 235, + 327 + ], + "lines": [ + { + "bbox": [ + 207, + 299, + 235, + 327 + ], + "spans": [ + { + "bbox": [ + 207, + 299, + 235, + 327 + ], + "type": "image", + "image_path": "76b15a1e0fb1d39530419956091de5892730b41beac49a64ece981401612de8f.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 329, + 241, + 346 + ], + "lines": [ + { + "bbox": [ + 196, + 329, + 241, + 346 + ], + "spans": [ + { + "bbox": [ + 196, + 329, + 241, + 346 + ], + "type": "text", + "content": "w/only " + }, + { + "bbox": [ + 196, + 329, + 241, + 346 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{occ}}" + }, + { + "bbox": [ + 196, + 329, + 241, + 346 + ], + "type": "text", + "content": " attention" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 253, + 220, + 286, + 256 + ], + "blocks": [ + { + "bbox": [ + 253, + 220, + 286, + 256 + ], + "lines": [ + { + "bbox": [ + 253, + 220, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 253, + 220, + 286, + 256 + ], + "type": "image", + "image_path": "48ae97f8280fd4eb1d96592a5215d8dc71da9dbf33a861e42ca530b9cccbeec7.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 253, + 265, + 288, + 287 + ], + "blocks": [ + { + "bbox": [ + 253, + 265, + 288, + 287 + ], + "lines": [ + { + "bbox": [ + 253, + 265, + 288, + 287 + ], + "spans": [ + { + "bbox": [ + 253, + 265, + 288, + 287 + ], + "type": "image", + "image_path": "6835bb035d609bf6ba38348e36d53a49eac7ab22e4a7c4a8afedaa49f7d85fb6.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 258, + 298, + 283, + 327 + ], + "blocks": [ + { + "bbox": [ + 258, + 298, + 283, + 327 + ], + "lines": [ + { + "bbox": [ + 258, + 298, + 283, + 327 + ], + "spans": [ + { + "bbox": [ + 258, + 298, + 283, + 327 + ], + "type": "image", + "image_path": "959175325ad9a896663bb3b676db8e8cff8d45ca0065f9e5a4fb82de74ddc81c.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 330, + 278, + 346 + ], + "lines": [ + { + "bbox": [ + 257, + 330, + 278, + 346 + ], + "spans": [ + { + "bbox": [ + 257, + 330, + 278, + 346 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 418, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 586 + ], + "type": "text", + "content": "in challenging scenarios. In contrast, 2D completion inconsistencies accumulate as more views are added, especially when pix2gestalt is more uncertain, confusing the downstream reconstructor models. For instance, in the giraffe example in Fig. 5, the pix2gestalt completion fails to capture the overall structure of the 3D object, which in turn leads the 3D generator models to misinterpret it as a woodstick-like shape. Instead, Amodal3R reconstructs the 3D geometry and appearance well, maintaining good alignment with the partially visible inputs. These results underscore the effectiveness of Amodal3R and highlight the benefit of integrating occlusion reasoning directly in 3D space, eliminating the reliance on separate, potentially inconsistent, 2D completion stages." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 597, + 149, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 149, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 149, + 609 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 664 + ], + "type": "text", + "content": "We conducted several ablations to study the impact of the different components of Amodal3R and report the results in Tab. 3 and Fig. 7. For these experiments, we test single-view image-to-3D reconstruction on the GSO dataset." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Naive Conditioning. We first evaluated a version of the model that still conditions the reconstruction on the visibility and occlusion masks, but without using the modules of Sec. 3.1. Instead, we directly concatenate the tokens " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{vis}}" + } + ] + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 316, + 217, + 554, + 285 + ], + "blocks": [ + { + "bbox": [ + 316, + 217, + 554, + 285 + ], + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 285 + ], + "type": "table", + "html": "
MethodAppearanceGeometry
FID ↓KID(%) ↓COV(%) ↑MMD(‰) ↓
naive conditioning31.960.4937.963.61
w/ only mask-weighted attention30.530.3836.903.69
w/ only occlusion-aware layer31.770.5740.193.51
full model (Ours)30.640.3539.613.62
", + "image_path": "6ddc4efba93fa58ff72c0ca6e351a7c0231f790f55991c88bc2e393f81a58fd0.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "table_body" + } + ], + "index": 43 + }, + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "type": "text", + "content": "Table 3. Ablations on different mask conditioning designs." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{dino}}" + }, + { + "bbox": [ + 313, + 324, + 555, + 395 + ], + "type": "text", + "content": " to form the condition for the cross-attention layer. The results (first row in Tab. 3 and second column in Fig. 7) show that, while the resulting model can still perform basic completion, the textures in the occluded regions are inconsistent with those in the visible ones, and the reconstructed geometry is inaccurate, e.g. the hole in the shoe." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 313, + 396, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 491 + ], + "type": "text", + "content": "Mask-weighted Attention. To evaluate the effectiveness of our proposed mask-weighted attention mechanism, we integrate it into the training while omitting the occlusion-aware layer. The results demonstrate improved visual quality, especially in capturing texture details in the visible areas. However, the geometries exhibit deficiencies, as seen in the problematic shoe in Fig. 7 (first row), and the toy monster with a broken back (third row)." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 313, + 492, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 600 + ], + "type": "text", + "content": "Occlusion-aware Layer. The integration of the occlusion-aware layer improves the geometry both quantitatively and qualitatively. This aligns with our motivation for introducing an additional cross-attention layer, aimed at reconstructing the visible areas via the image-conditioned layer and occluded areas via the subsequent layer. However, the occlusion-aware layer alone still results in unsatisfactory appearances, which again indirectly attests to the effectiveness of the mask-weighted attention mechanism." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 601, + 553, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 637 + ], + "type": "text", + "content": "Full Model. Consequently, the full model integrating both modules achieves optimal 3D generation results characterized by complete geometry and consistent textures." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 647, + 511, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 511, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 511, + 660 + ], + "type": "text", + "content": "4.4. Real-World Generation / Completion" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Amodal3R demonstrates strong generalization to out-of-domain test examples, primarily because it builds upon a large-scale pre-trained 3D generator and is fine-tuned with diverse categories. To demonstrate this advantage," + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9187" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 112, + 276 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 112, + 276 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 112, + 276 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 112, + 276 + ], + "type": "image", + "image_path": "0aaa4a1e38acdffb89a296b94d52034c83d55d827f052e369128603ee27b9ea3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 276, + 139, + 285 + ], + "lines": [ + { + "bbox": [ + 88, + 276, + 139, + 285 + ], + "spans": [ + { + "bbox": [ + 88, + 276, + 139, + 285 + ], + "type": "text", + "content": "Occluded input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 83, + 160, + 110 + ], + "blocks": [ + { + "bbox": [ + 120, + 83, + 160, + 110 + ], + "lines": [ + { + "bbox": [ + 120, + 83, + 160, + 110 + ], + "spans": [ + { + "bbox": [ + 120, + 83, + 160, + 110 + ], + "type": "image", + "image_path": "cd9906c266acf5dba296545ee4c74eb83b3bd899f98e2c37e9f02386537e8754.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 174, + 83, + 211, + 110 + ], + "blocks": [ + { + "bbox": [ + 174, + 83, + 211, + 110 + ], + "lines": [ + { + "bbox": [ + 174, + 83, + 211, + 110 + ], + "spans": [ + { + "bbox": [ + 174, + 83, + 211, + 110 + ], + "type": "image", + "image_path": "6255e3e76a9b3d590968e813e66631f1bea605af0f2ac52acdc079fadd1eacf9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 220, + 81, + 296, + 118 + ], + "blocks": [ + { + "bbox": [ + 220, + 81, + 296, + 118 + ], + "lines": [ + { + "bbox": [ + 220, + 81, + 296, + 118 + ], + "spans": [ + { + "bbox": [ + 220, + 81, + 296, + 118 + ], + "type": "image", + "image_path": "3fa5c180b999464af1fdc01181a725ffe146219a7762a97543d5a8b1b4ac4c42.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 85, + 371, + 119 + ], + "blocks": [ + { + "bbox": [ + 302, + 85, + 371, + 119 + ], + "lines": [ + { + "bbox": [ + 302, + 85, + 371, + 119 + ], + "spans": [ + { + "bbox": [ + 302, + 85, + 371, + 119 + ], + "type": "image", + "image_path": "f10aee28cacabda7128574d2eae3a69cb0d67b96aa299faf18e5be799b306367.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 84, + 449, + 118 + ], + "blocks": [ + { + "bbox": [ + 378, + 84, + 449, + 118 + ], + "lines": [ + { + "bbox": [ + 378, + 84, + 449, + 118 + ], + "spans": [ + { + "bbox": [ + 378, + 84, + 449, + 118 + ], + "type": "image", + "image_path": "892deecb97e98270146a0d38533fa2d5c34f4035e00f5925d4e7e03b6ff31120.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 463, + 78, + 538, + 125 + ], + "blocks": [ + { + "bbox": [ + 463, + 78, + 538, + 125 + ], + "lines": [ + { + "bbox": [ + 463, + 78, + 538, + 125 + ], + "spans": [ + { + "bbox": [ + 463, + 78, + 538, + 125 + ], + "type": "image", + "image_path": "234d82c8065bcddf1d2049287d52b9664374f9fface24fcba575e534966411d2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 60, + 284, + 111, + 334 + ], + "blocks": [ + { + "bbox": [ + 60, + 284, + 111, + 334 + ], + "lines": [ + { + "bbox": [ + 60, + 284, + 111, + 334 + ], + "spans": [ + { + "bbox": [ + 60, + 284, + 111, + 334 + ], + "type": "image", + "image_path": "e36b8a031e27126923f0ced3c2ab4180a4bac07bc33df0bdaf9b49b53757226c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 121, + 144, + 147, + 156 + ], + "blocks": [ + { + "bbox": [ + 121, + 144, + 147, + 156 + ], + "lines": [ + { + "bbox": [ + 121, + 144, + 147, + 156 + ], + "spans": [ + { + "bbox": [ + 121, + 144, + 147, + 156 + ], + "type": "image", + "image_path": "82dea0ecb153627d533b3aac6897890423a40e2c23dd3df859e0e95839deb30d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 124, + 193, + 153, + 213 + ], + "blocks": [ + { + "bbox": [ + 124, + 193, + 153, + 213 + ], + "lines": [ + { + "bbox": [ + 124, + 193, + 153, + 213 + ], + "spans": [ + { + "bbox": [ + 124, + 193, + 153, + 213 + ], + "type": "image", + "image_path": "88295939781841cee35d5000c351c08e248212fab62f8c07c3f561f1a3889ca1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 131, + 235, + 149, + 266 + ], + "blocks": [ + { + "bbox": [ + 131, + 235, + 149, + 266 + ], + "lines": [ + { + "bbox": [ + 131, + 235, + 149, + 266 + ], + "spans": [ + { + "bbox": [ + 131, + 235, + 149, + 266 + ], + "type": "image", + "image_path": "fc064faa16ef879a9ba1d032fee422fb0ebc947eabfea37da35ad56efcbabf12.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 163, + 144, + 170, + 155 + ], + "blocks": [ + { + "bbox": [ + 163, + 144, + 170, + 155 + ], + "lines": [ + { + "bbox": [ + 163, + 144, + 170, + 155 + ], + "spans": [ + { + "bbox": [ + 163, + 144, + 170, + 155 + ], + "type": "image", + "image_path": "0ff6bc3ed69e3e2674d70fa394d723bcb8970c492a0659966af6727f07648192.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 179, + 144, + 198, + 156 + ], + "blocks": [ + { + "bbox": [ + 179, + 144, + 198, + 156 + ], + "lines": [ + { + "bbox": [ + 179, + 144, + 198, + 156 + ], + "spans": [ + { + "bbox": [ + 179, + 144, + 198, + 156 + ], + "type": "image", + "image_path": "8a4314c862580b2aa2fce1ac227cac194689dde71da5b50611a71b7a236e2b41.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 182, + 294, + 215 + ], + "blocks": [ + { + "bbox": [ + 208, + 182, + 294, + 215 + ], + "lines": [ + { + "bbox": [ + 208, + 182, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 208, + 182, + 294, + 215 + ], + "type": "image", + "image_path": "d9057c53bb1aed5ecf98e5fed336de69e42395bf867e96f5ebb0b28e771fcd2b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 181, + 231, + 203, + 268 + ], + "blocks": [ + { + "bbox": [ + 181, + 231, + 203, + 268 + ], + "lines": [ + { + "bbox": [ + 181, + 231, + 203, + 268 + ], + "spans": [ + { + "bbox": [ + 181, + 231, + 203, + 268 + ], + "type": "image", + "image_path": "8f7dcc9b77e7dd831b84e0fc47252353b7c026800f899e87b8272a5eab1a687e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 229, + 224, + 288, + 266 + ], + "blocks": [ + { + "bbox": [ + 229, + 224, + 288, + 266 + ], + "lines": [ + { + "bbox": [ + 229, + 224, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 229, + 224, + 288, + 266 + ], + "type": "image", + "image_path": "2571264e7593b19c26b1ee6901744dc8e429fa68c88e75dfc330069516431d8e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 304, + 141, + 334, + 160 + ], + "blocks": [ + { + "bbox": [ + 304, + 141, + 334, + 160 + ], + "lines": [ + { + "bbox": [ + 304, + 141, + 334, + 160 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 334, + 160 + ], + "type": "image", + "image_path": "1b348933600ad1b2ca1ac0b04a6f11de2ddc290c3c10fe390247968962e7e83e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 343, + 140, + 369, + 162 + ], + "blocks": [ + { + "bbox": [ + 343, + 140, + 369, + 162 + ], + "lines": [ + { + "bbox": [ + 343, + 140, + 369, + 162 + ], + "spans": [ + { + "bbox": [ + 343, + 140, + 369, + 162 + ], + "type": "image", + "image_path": "d3f47ac1d5f26df7b563f91770e9388b5209e1d152b5e68c1a5b0ecc481442a8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 386, + 140, + 405, + 158 + ], + "blocks": [ + { + "bbox": [ + 386, + 140, + 405, + 158 + ], + "lines": [ + { + "bbox": [ + 386, + 140, + 405, + 158 + ], + "spans": [ + { + "bbox": [ + 386, + 140, + 405, + 158 + ], + "type": "image", + "image_path": "fdab1d71ab835dfbf75ed3891d78e1e9dd2884ad01ce645725e44565f26d6905.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 422, + 140, + 441, + 158 + ], + "blocks": [ + { + "bbox": [ + 422, + 140, + 441, + 158 + ], + "lines": [ + { + "bbox": [ + 422, + 140, + 441, + 158 + ], + "spans": [ + { + "bbox": [ + 422, + 140, + 441, + 158 + ], + "type": "image", + "image_path": "779221d84eac13762e3104fc796d1250c707fe75e07cd2bd744cfa1eb962af37.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 465, + 139, + 494, + 159 + ], + "blocks": [ + { + "bbox": [ + 465, + 139, + 494, + 159 + ], + "lines": [ + { + "bbox": [ + 465, + 139, + 494, + 159 + ], + "spans": [ + { + "bbox": [ + 465, + 139, + 494, + 159 + ], + "type": "image", + "image_path": "3812f3278f3a6124fa65304bb1fb01e03cd8551b3e7f80e5e82ae36ada5851e5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 507, + 140, + 537, + 164 + ], + "blocks": [ + { + "bbox": [ + 507, + 140, + 537, + 164 + ], + "lines": [ + { + "bbox": [ + 507, + 140, + 537, + 164 + ], + "spans": [ + { + "bbox": [ + 507, + 140, + 537, + 164 + ], + "type": "image", + "image_path": "18cde7d1bc2554bec861d740a3074698dd941324d4d785f080a7db93745fc3bf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 303, + 186, + 370, + 218 + ], + "blocks": [ + { + "bbox": [ + 303, + 186, + 370, + 218 + ], + "lines": [ + { + "bbox": [ + 303, + 186, + 370, + 218 + ], + "spans": [ + { + "bbox": [ + 303, + 186, + 370, + 218 + ], + "type": "image", + "image_path": "0f2e244f019b483441bf1b023c3042682d39627a63eadbcd551087476d6c8165.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 383, + 184, + 444, + 224 + ], + "blocks": [ + { + "bbox": [ + 383, + 184, + 444, + 224 + ], + "lines": [ + { + "bbox": [ + 383, + 184, + 444, + 224 + ], + "spans": [ + { + "bbox": [ + 383, + 184, + 444, + 224 + ], + "type": "image", + "image_path": "dc23b4a45ae4854a6f2a8956b0389f2020b1b7c916bbd755a40886f1655b56b6.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 380, + 234, + 449, + 259 + ], + "blocks": [ + { + "bbox": [ + 380, + 234, + 449, + 259 + ], + "lines": [ + { + "bbox": [ + 380, + 234, + 449, + 259 + ], + "spans": [ + { + "bbox": [ + 380, + 234, + 449, + 259 + ], + "type": "image", + "image_path": "cf392de632adcf2d68ffa9409dd67b0b8a481bd3f1a1614627d0b9a9912e4579.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 463, + 179, + 538, + 226 + ], + "blocks": [ + { + "bbox": [ + 463, + 179, + 538, + 226 + ], + "lines": [ + { + "bbox": [ + 463, + 179, + 538, + 226 + ], + "spans": [ + { + "bbox": [ + 463, + 179, + 538, + 226 + ], + "type": "image", + "image_path": "70020367524cbc699a41c68ab03082351c1238ec07afadaf4072d8b3414b7190.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 462, + 233, + 548, + 262 + ], + "blocks": [ + { + "bbox": [ + 462, + 233, + 548, + 262 + ], + "lines": [ + { + "bbox": [ + 462, + 233, + 548, + 262 + ], + "spans": [ + { + "bbox": [ + 462, + 233, + 548, + 262 + ], + "type": "image", + "image_path": "ebdeda1de9d2f58051643f58f0aafffba36868ae1a7ec8e7faf7e5c6f5f6fafe.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 59, + 335, + 111, + 384 + ], + "blocks": [ + { + "bbox": [ + 59, + 335, + 111, + 384 + ], + "lines": [ + { + "bbox": [ + 59, + 335, + 111, + 384 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 111, + 384 + ], + "type": "image", + "image_path": "77e1b7449a788711e712931136f03a029ebda3a502be37bc41a9a790c689ba8b.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 386, + 110, + 396 + ], + "lines": [ + { + "bbox": [ + 59, + 386, + 110, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 386, + 110, + 396 + ], + "type": "text", + "content": "Occluded input" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 401, + 553, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 401, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 553, + 424 + ], + "type": "text", + "content": "Figure 8. Examples on Replica [68] and in-the-wild images. The target objects and the occluders are shown in red and green outlines. More examples are provided in supplementary material Figs. D.7 to D.9." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 133, + 295, + 177, + 332 + ], + "blocks": [ + { + "bbox": [ + 133, + 295, + 177, + 332 + ], + "lines": [ + { + "bbox": [ + 133, + 295, + 177, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 295, + 177, + 332 + ], + "type": "image", + "image_path": "9ace62434057fe4fb29c912e2103f8da59747ea756046d5a2b9791ef453b560a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 134, + 338, + 175, + 377 + ], + "blocks": [ + { + "bbox": [ + 134, + 338, + 175, + 377 + ], + "lines": [ + { + "bbox": [ + 134, + 338, + 175, + 377 + ], + "spans": [ + { + "bbox": [ + 134, + 338, + 175, + 377 + ], + "type": "image", + "image_path": "185102de64684adf0ad58d98b905b6426df9d917e537d74a4542a7a1aebc8fc1.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 184, + 296, + 227, + 333 + ], + "blocks": [ + { + "bbox": [ + 184, + 296, + 227, + 333 + ], + "lines": [ + { + "bbox": [ + 184, + 296, + 227, + 333 + ], + "spans": [ + { + "bbox": [ + 184, + 296, + 227, + 333 + ], + "type": "image", + "image_path": "31b127e645d074e2b254bf4a71b0909fa02ecb6743ee5223b33dfb650101491b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 181, + 338, + 225, + 377 + ], + "blocks": [ + { + "bbox": [ + 181, + 338, + 225, + 377 + ], + "lines": [ + { + "bbox": [ + 181, + 338, + 225, + 377 + ], + "spans": [ + { + "bbox": [ + 181, + 338, + 225, + 377 + ], + "type": "image", + "image_path": "289d9b695e7ba4ba8013fd3ee4cc17b39ff7fafa036e8844ab8e49af2b771168.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 386, + 250, + 395 + ], + "lines": [ + { + "bbox": [ + 208, + 386, + 250, + 395 + ], + "spans": [ + { + "bbox": [ + 208, + 386, + 250, + 395 + ], + "type": "text", + "content": "Novel views" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 238, + 285, + 329, + 339 + ], + "blocks": [ + { + "bbox": [ + 164, + 276, + 294, + 285 + ], + "lines": [ + { + "bbox": [ + 164, + 276, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 164, + 276, + 294, + 285 + ], + "type": "text", + "content": "pix2gestalt [54] GaussianAnything [36]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 238, + 285, + 329, + 339 + ], + "lines": [ + { + "bbox": [ + 238, + 285, + 329, + 339 + ], + "spans": [ + { + "bbox": [ + 238, + 285, + 329, + 339 + ], + "type": "image", + "image_path": "1fe967532a0c24f45888599ae0b356d60e934b6c115bd8f680688d8a313e5cb0.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 275, + 356, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 275, + 356, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 275, + 356, + 285 + ], + "type": "text", + "content": "Real3D [27]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 238, + 338, + 280, + 377 + ], + "blocks": [ + { + "bbox": [ + 238, + 338, + 280, + 377 + ], + "lines": [ + { + "bbox": [ + 238, + 338, + 280, + 377 + ], + "spans": [ + { + "bbox": [ + 238, + 338, + 280, + 377 + ], + "type": "image", + "image_path": "8a87f5edfdfea7b2908f2f1152168e5d1f6034398e5b237e129cf4821cc533de.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 286, + 285, + 331, + 338 + ], + "blocks": [ + { + "bbox": [ + 286, + 285, + 331, + 338 + ], + "lines": [ + { + "bbox": [ + 286, + 285, + 331, + 338 + ], + "spans": [ + { + "bbox": [ + 286, + 285, + 331, + 338 + ], + "type": "image", + "image_path": "2acc76aecb295dfc1a8c335dadb975480f04c941e533914e6d2fc45b1e46d92c.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 286, + 338, + 330, + 377 + ], + "blocks": [ + { + "bbox": [ + 286, + 338, + 330, + 377 + ], + "lines": [ + { + "bbox": [ + 286, + 338, + 330, + 377 + ], + "spans": [ + { + "bbox": [ + 286, + 338, + 330, + 377 + ], + "type": "image", + "image_path": "c863e58eeb415d78874e23b004d7563bee499fda1fa4df113ed32f1dbb27686d.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 350, + 295, + 392, + 332 + ], + "blocks": [ + { + "bbox": [ + 388, + 275, + 436, + 285 + ], + "lines": [ + { + "bbox": [ + 388, + 275, + 436, + 285 + ], + "spans": [ + { + "bbox": [ + 388, + 275, + 436, + 285 + ], + "type": "text", + "content": "TRELLIS [80]" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 350, + 295, + 392, + 332 + ], + "lines": [ + { + "bbox": [ + 350, + 295, + 392, + 332 + ], + "spans": [ + { + "bbox": [ + 350, + 295, + 392, + 332 + ], + "type": "image", + "image_path": "92dfd49f4bbb50ad9c2fb36968106be55a1d8652bb5a0beeaa3cecc05f9b4e9e.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 399, + 295, + 440, + 334 + ], + "blocks": [ + { + "bbox": [ + 399, + 295, + 440, + 334 + ], + "lines": [ + { + "bbox": [ + 399, + 295, + 440, + 334 + ], + "spans": [ + { + "bbox": [ + 399, + 295, + 440, + 334 + ], + "type": "image", + "image_path": "3f4adb75a662be7f373e48ff20b88e463399b648cd18d9c0de7665505e87bd4b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 351, + 338, + 392, + 377 + ], + "blocks": [ + { + "bbox": [ + 351, + 338, + 392, + 377 + ], + "lines": [ + { + "bbox": [ + 351, + 338, + 392, + 377 + ], + "spans": [ + { + "bbox": [ + 351, + 338, + 392, + 377 + ], + "type": "image", + "image_path": "8446728d4dd3a897b765ea0c0f6f4388fff96bfe0fb74902dea7d8d9730c2ad6.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 395, + 338, + 438, + 377 + ], + "blocks": [ + { + "bbox": [ + 395, + 338, + 438, + 377 + ], + "lines": [ + { + "bbox": [ + 395, + 338, + 438, + 377 + ], + "spans": [ + { + "bbox": [ + 395, + 338, + 438, + 377 + ], + "type": "image", + "image_path": "fd104cfb159e20527df1235e54d67489c2e59aa683914fe2e44f69546062d9eb.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 386, + 462, + 396 + ], + "lines": [ + { + "bbox": [ + 417, + 386, + 462, + 396 + ], + "spans": [ + { + "bbox": [ + 417, + 386, + 462, + 396 + ], + "type": "text", + "content": "Normal maps" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 447, + 285, + 538, + 342 + ], + "blocks": [ + { + "bbox": [ + 468, + 275, + 527, + 285 + ], + "lines": [ + { + "bbox": [ + 468, + 275, + 527, + 285 + ], + "spans": [ + { + "bbox": [ + 468, + 275, + 527, + 285 + ], + "type": "text", + "content": "Amodal3R (Ours)" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 285, + 538, + 342 + ], + "lines": [ + { + "bbox": [ + 447, + 285, + 538, + 342 + ], + "spans": [ + { + "bbox": [ + 447, + 285, + 538, + 342 + ], + "type": "image", + "image_path": "599cd74b2b6add55f3c3bd9a79332d182be0b6e97457cdd4bf02904893136add.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 447, + 338, + 538, + 377 + ], + "blocks": [ + { + "bbox": [ + 447, + 338, + 538, + 377 + ], + "lines": [ + { + "bbox": [ + 447, + 338, + 538, + 377 + ], + "spans": [ + { + "bbox": [ + 447, + 338, + 538, + 377 + ], + "type": "image", + "image_path": "d38f75edd1ea05c2b78c64e0e16f7abaa3e2f6bc04b3f720a3b7a564bebf594b.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "table", + "bbox": [ + 64, + 443, + 288, + 516 + ], + "blocks": [ + { + "bbox": [ + 64, + 443, + 288, + 516 + ], + "lines": [ + { + "bbox": [ + 64, + 443, + 288, + 516 + ], + "spans": [ + { + "bbox": [ + 64, + 443, + 288, + 516 + ], + "type": "table", + "html": "
MethodP-FID↓COV(%) ↑MMD(‰) ↓
GaussianAnything91.3761.2225.00
Real3D79.5555.1018.28
TRELLIS128.5259.1723.95
Amodal3R61.4659.1810.98
", + "image_path": "0e9a7a60e563272d71d507f13b6fabfdccb2d9da73d1afa31ffe8f952ec5453d.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "table_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 67, + 524, + 284, + 536 + ], + "lines": [ + { + "bbox": [ + 67, + 524, + 284, + 536 + ], + "spans": [ + { + "bbox": [ + 67, + 524, + 284, + 536 + ], + "type": "text", + "content": "Table 4. Quantitative results on the Replica dataset [68]." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 689 + ], + "type": "text", + "content": "we conducted scene-level evaluations. Qualitative results are shown in Fig. 8, which include COCOA [97] (first two rows), Replica [68] (third and fourth rows), and in-the-wild images (fifth and sixth rows). Here, we used Segment Anything [34] to obtain the visibility and occlusion masks. The results show that Amodal3R generates photorealistic 3D assets, whereas pix2gestalt fails to infer complete shapes from the same inputs, leading to unsatisfactory results from subsequent image-to-3D models. We also visualize the colored normal maps, which show that the results of Amodal3R are simultaneously rich in geometric and textural detail." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "content": "Quantitative results are presented in Tab. 4. As real-world datasets with instance 3D amodal annotations for" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 445, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 529 + ], + "type": "text", + "content": "both geometry and appearance are limited, here we focus on geometry evaluation by curating a set of 49 completed 3D instances from the Replica scenes. Furthermore, we used the Umeyama algorithm [73] to match the generated meshes with the ground truth meshes in terms of scale and pose. It can be observed that our Amodal3R is still on par with or better than the baselines on the scene-level dataset." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "We have introduced Amodal3R, a feed-forward approach to reconstruct the complete photorealistic 3D assets from partially visible images. The key to our success is that we construct a mask-weighted cross-attention mechanism and an occlusion-aware layer to effectively exploit visible and occluded information. Compared to the state-of-the-art methods that rely on sequential 2D completion followed by 3D generation, our Amodal3R achieves remarkably better performance by operating directly in 3D space. Furthermore, results on in-the-wild images indicate its potential for subsequent applications in 3D decomposition, marking a step towards robust 3D asset reconstruction in real-world environments with complex occlusion." + } + ] + } + ], + "index": 57 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9188" + } + ] + } + ], + "index": 58 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This study is supported under the RIE2020 Industry Alignment Fund-Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contributions from the industry partner(s). This research is also supported by the Ministry of Education, Singapore, under its Academic Research Fund Tier 1 RG107/24. Dr. Chuanxia Zheng is supported by EPSRC SYN3D EP/Z001811/1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 204, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 204, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 296, + 270 + ], + "type": "text", + "content": "[1] Titas Anciukevicius, Zexiang Xu, Matthew Fisher, Paul Henderson, Hakan Bilen, Niloy J Mitra, and Paul Guerrero. Renderdiffusion: Image diffusion for 3d reconstruction, inpainting and generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12608-12618, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 304 + ], + "type": "text", + "content": "[2] Jiayang Ao, Yanbei Jiang, Qiuhong Ke, and Krista A Ehinger. Open-world amodal appearance completion. arXiv preprint arXiv:2411.13019, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 296, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 296, + 336 + ], + "type": "text", + "content": "[3] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 5, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 337, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 337, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 337, + 296, + 392 + ], + "type": "text", + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 446 + ], + "type": "text", + "content": "[5] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 448, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 296, + 491 + ], + "type": "text", + "content": "[6] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2024. 1, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 492, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 296, + 557 + ], + "type": "text", + "content": "[7] Minghao Chen, Roman Shapovalov, Iro Laina, Tom Monnier, Jianyuan Wang, David Novotny, and Andrea Vedaldi. Partgen: Part-level 3d generation and reconstruction with multi-view diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "text", + "content": "[8] Yiwen Chen, Tong He, Di Huang, Weicai Ye, Sijin Chen, Jiaxiang Tang, Xin Chen, Zhongang Cai, Lei Yang, Gang Yu, et al. Meshanything: Artist-created mesh generation with autoregressive transformers. arXiv preprint arXiv:2406.10163, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 658 + ], + "type": "text", + "content": "[9] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexander G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4456-4465, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[11] Junhyeong Cho, Kim Youwang, Hunmin Yang, and TaeHyun Oh. Robust 3d shape reconstruction in zero-shot from a single image in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 128, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 128, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 554, + 172 + ], + "type": "text", + "content": "[12] Ruihang Chu, Enze Xie, Shentong Mo, Zhenguo Li, Matthias Nießner, Chi-Wing Fu, and Jiaya Jia. Diffcomplete: Diffusion-based generative 3d shape completion. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 554, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 554, + 248 + ], + "type": "text", + "content": "[13] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21126-21136, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 249, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 249, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 249, + 554, + 316 + ], + "type": "text", + "content": "[14] Ruikai Cui, Weizhe Liu, Weixuan Sun, Senbo Wang, Taizhang Shang, Yang Li, Xibin Song, Han Yan, Zhennan Wu, Shenzhen Chen, et al. Neusdfusion: A spatial-aware generative model for 3d shape completion, reconstruction, and generation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "text", + "content": "[15] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 317, + 316, + 554, + 370 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 371, + 554, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 371, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 371, + 554, + 436 + ], + "type": "text", + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. In 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560. IEEE, 2022. 2, 5, 6, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 437, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 554, + 480 + ], + "type": "text", + "content": "[17] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129:3313-3337, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 481, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 481, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 481, + 554, + 536 + ], + "type": "text", + "content": "[18] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. Advances In Neural Information Processing Systems, 35:31841-31854, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 537, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 580 + ], + "type": "text", + "content": "[19] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 581, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 554, + 613 + ], + "type": "text", + "content": "[20] Anchit Gupta, Wenhan Xiong, Yixin Nie, Ian Jones, and Barlas Oğuz. 3dgen: Triplane latent diffusion for textured mesh generation. arXiv preprint arXiv:2303.05371, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 668 + ], + "type": "text", + "content": "[21] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), pages 6840-6851, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9189" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. In International Conference on Learning Representations (ICLR), 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[24] Tao Hu, Zhizhong Han, and Matthias Zwicker. 3d shape completion with multi-view consistent inference. In Proceedings of the AAAI conference on artificial intelligence, pages 10997-11004, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[25] Zitian Huang, Yikuan Yu, Jiawen Xu, Feng Ni, and Xinyi Le. Pf-net: Point fractal network for 3d point cloud completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 7662-7670, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[26] Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. Globally and locally consistent image completion. ACM Transactions on Graphics (TOG), 36(4):107, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[27] Hanwen Jiang, Qixing Huang, and Georgios Pavlakos. Real3d: Scaling up large reconstruction models with real-world images. arXiv preprint arXiv:2406.08479, 2024. 5, 6, 8, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "text", + "content": "[28] Heewoo Jun and Alex Nichol. Shap-e: Generating conditional 3d implicit functions. arXiv preprint arXiv:2305.02463, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 387 + ], + "type": "text", + "content": "[29] Yoni Kasten, Ohad Rahamim, and Gal Chechik. Point cloud completion with pretrained text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[30] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 520 + ], + "type": "text", + "content": "[31] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": "[32] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffus ionclip: Text-guided diffusion models for robust image man i plication. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[33] Yulhwa Kim, Dongwon Jo, Hyesung Jeon, Taesu Kim, Daehyun Ahn, Hyungjun Kim, et al. Leveraging early-stage robustness in diffusion models for efficient and high-quality image synthesis. Advances in Neural Information Processing Systems, 36:1229-1244, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[34] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Yushi Lan, Feitong Tan, Di Qiu, Qiangeng Xu, Kyle Genova, Zeng Huang, Sean Fanello, Rohit Pandey, Thomas" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Funkhouser, Chen Change Loy, et al. Gaussian3diff: 3d gaussian diffusion for 3d full head synthesis and editing. arXiv preprint arXiv:2312.03763, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[36] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. arXiv preprint arXiv:2411.08033, 2024. 5, 6, 8, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[37] Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusionsdf: Text-to-shape via voxelized diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12642-12651, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[38] Xinhai Li, Huaibin Wang, and Kuo-Kun Tseng. Gaussian diffusion: 3d gaussian splatting for denoising diffusion probabilistic models with structured noise. arXiv preprint arXiv:2311.11221, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 293 + ], + "type": "text", + "content": "[39] Chen-Hsuan Lin, Chen Kong, and Simon Lucey. Learning efficient point cloud generation for dense 3d object reconstruction. In proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[40] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image inpainting for irregular holes using partial convolutions. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 404 + ], + "type": "text", + "content": "[41] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 9298–9309, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[42] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 504 + ], + "type": "text", + "content": "[43] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 9970-9980, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 505, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 555, + 548 + ], + "type": "text", + "content": "[44] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 602 + ], + "type": "text", + "content": "[45] Luke Melas-Kyriazi, Christian Rupprecht, and Andrea Vedaldi. Pc2: Projection-conditioned point cloud diffusion for single-image 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12923-12932, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "type": "text", + "content": "[46] Norman Müller, Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Bulo, Peter Kontschieder, and Matthias Nießner. Diffrf: Rendering-guided 3d radiance field diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4328-4338, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[47] Charlie Nash, Yaroslav Ganin, SM Ali Eslami, and Peter Battaglia. *Polygen: An autoregressive generative model of 3d meshes*. In International Conference on Machine Learning (ICML), pages 7220-7229. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9190" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[48] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 7588–7597, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[49] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2, 5, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[50] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11453-11464, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[51] Evangelos Ntavelis, Aliaksandr Siarohin, Kyle Olszewski, Chaoyang Wang, Luc V Gool, and Sergey Tulyakov. Autodecoding latent 3d diffusion models. Advances in Neural Information Processing Systems, 36:67021-67047, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[52] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 4, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 396 + ], + "type": "text", + "content": "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 3, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "text", + "content": "[54] Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. pix2gestalt: Amodal segmentation by synthesizing wholes. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3931-3940. IEEE Computer Society, 2024. 1, 2, 3, 5, 6, 7, 8, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[55] Eunbyung Park, Jimei Yang, Ersin Yumer, Duygu Ceylan, and Alexander C Berg. Transformation-grounded image generation network for novel 3d view synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3500-3509, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": "[56] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2536-2544. IEEE, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 644 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (ICML), pages 8748-8763. PmLR, 2021. 5, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[58] Barbara Roessle, Norman Müller, Lorenzo Porzi, Samuel Rota Bulò, Peter Kontschieder, Angela Dai, and Matthias Nießner. L3dg: Latent 3d gaussian diffusion. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 554, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 149 + ], + "type": "text", + "content": "[60] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems (NeurIPS), 33:20154-20166, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "text", + "content": "[61] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero " + }, + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "inline_equation", + "content": "123++" + }, + { + "bbox": [ + 316, + 151, + 554, + 205 + ], + "type": "text", + "content": ": a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110, 2023. 6, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 239 + ], + "type": "text", + "content": "[62] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 293 + ], + "type": "text", + "content": "[63] Wu Shuang, Youtian Lin, Yifei Zeng, Feihu Zhang, Jingxi Xu, Philip Torr, Xun Cao, and Yao Yao. Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems, 37:121859–121881, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 295, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 554, + 348 + ], + "type": "text", + "content": "[64] J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 415 + ], + "type": "text", + "content": "[65] Yawar Siddiqui, Antonio Alliegro, Alexey Artemov, Tatiana Tommasi, Daniele Sirigatti, Vladislav Rosov, Angela Dai, and Matthias Nießner. Meshgpt: Generating triangle meshes with decoder-only transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19615-19625, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 554, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 469 + ], + "type": "text", + "content": "[66] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning (ICML), pages 2256-2265. pmlr, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 472, + 554, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 524 + ], + "type": "text", + "content": "[67] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1798-1808, 2021. 2, 5, 6, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 580 + ], + "type": "text", + "content": "[68] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[69] David Stutz and Andreas Geiger. Learning 3d shape completion from laser scan data with weak supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1955-1964, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 626, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 554, + 679 + ], + "type": "text", + "content": "[70] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Viewset diffusion: (0-)image-conditioned 3d generative models from 2d data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8863-8873, 2023. 1, 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[71] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9191" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition (CVPR), pages 10208-10217, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[72] Shitao Tang, Jiacheng Chen, Dilin Wang, Chengzhou Tang, Fuyang Zhang, Yuchen Fan, Vikas Chandra, Yasutaka Furukawa, and Rakesh Ranjan. Mvdiffusion++: A dense high-resolution multi-view diffusion model for single or sparse-view 3d object reconstruction. In European Conference on Computer Vision, pages 175-191. Springer, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[73] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on pattern analysis and machine intelligence, 13(4): 376-380, 2002. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 250 + ], + "type": "text", + "content": "[74] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 316 + ], + "type": "text", + "content": "[75] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision (ECCV), pages 439-457. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 360 + ], + "type": "text", + "content": "[76] Daniel Watson, William Chan, Ricardo Martin-Brualla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "text", + "content": "[77] Daniel Watson, William Chan, Ricardo Martin Brulla, Jonathan Ho, Andrea Tagliasacchi, and Mohammad Norouzi. Novel view synthesis with diffusion models. In The Eleventh International Conference on Learning Representations (ICLR), 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "type": "text", + "content": "[78] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. In European Conference on Computer Vision (ECCV), pages 456-473. Springer, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 472, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 525 + ], + "type": "text", + "content": "[79] Zijie Wu, Yaonan Wang, Mingtao Feng, He Xie, and Ajmal Mian. Sketch and text guided diffusion model for colored point cloud generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8929-8939, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[80] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 1, 3, 4, 5, 6, 7, 8, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 647 + ], + "type": "text", + "content": "[81] Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, Soren Pirk, and Arie E Kaufman. Carve3d: Improving multi-view reconstruction consistency for diffusion models with rl finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6369-6379, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[82] Yiheng Xiong and Angela Dai. Pt43d: A probabilistic transformer for generating 3d shapes from single highly-ambiguous rgb images. arXiv preprint arXiv:2405.11914, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[83] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Amodal completion via progressive mixed context diffusion. In Proceed-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9109, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 150 + ], + "type": "text", + "content": "[84] Bo Yang, Hongkai Wen, Sen Wang, Ronald Clark, Andrew Markham, and Niki Trigoni. 3d object reconstruction from a single depth view with adversarial learning. In Proceedings of the IEEE international conference on computer vision workshops (ICCVW), pages 679-688, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 204 + ], + "type": "text", + "content": "[85] Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zisserman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28003-28013, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 259 + ], + "type": "text", + "content": "[86] Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene deocclusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 3784-3792, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 315 + ], + "type": "text", + "content": "[87] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 370 + ], + "type": "text", + "content": "[88] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 372, + 553, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 372, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 553, + 424 + ], + "type": "text", + "content": "[89] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9720-9731, 2024. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "text", + "content": "[90] Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Pluralistic image completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1438-1447, 2019. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 514 + ], + "type": "text", + "content": "[91] Chuanxia Zheng, Duy-Son Dao, Guoxian Song, Tat-Jen Cham, and Jianfei Cai. Visiting the invisible: Layer-by-layer completed scene decomposition. International Journal of Computer Vision (IJCV), 129:3195-3215, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 569 + ], + "type": "text", + "content": "[92] Chuanxia Zheng, Tat-Jen Cham, Jianfei Cai, and Dinh Phung. Bridging global context interactions for high-fidelity image completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11512-11522, 2022. 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[93] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[94] Xin-Yang Zheng, Hao Pan, Yu-Xiao Guo, Xin Tong, and Yang Liu. Mvd^2: Efficient multiview 3d reconstruction for multiview diffusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[95] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9192" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 239 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[96] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. Advances in neural information processing systems (NeurIPS), 31, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 173 + ], + "type": "text", + "content": "[97] Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Dóllár. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1464-1472, 2017. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 239 + ], + "type": "text", + "content": "[98] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9193" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_content_list.json b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8de9d5d77fd364ae04ed3d6315146b606a08f323 --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "An Efficient Hybrid Vision Transformer for TinyML Applications", + "text_level": 1, + "bbox": [ + 166, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fanhong Zeng $^{1,2}$ Huanan Li $^{1,2}$ Juntao Guan $^{1,3*}$", + "bbox": [ + 297, + 179, + 704, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rui Fan $^{1,2}$ Tong Wu $^{1,2}$ Xilong Wang $^{1,2}$ Rui Lai $^{1,2*}$", + "bbox": [ + 287, + 198, + 736, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Key Laboratory of Analog Integrated Circuits and Systems (Ministry of Education)", + "bbox": [ + 222, + 217, + 774, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Faculty of Integrated Circuit, Xidian University, Xi'an, China", + "bbox": [ + 292, + 234, + 704, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hangzhou Institute of Technology, Xidian University, Hangzhou, China", + "bbox": [ + 258, + 252, + 738, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enable the deployment of Vision Transformers on resource-constrained mobile and edge devices, the development of efficient ViT models has attracted significant attention. Researchers achieved remarkable improvements in accuracy and speed by optimizing attention mechanisms and integrating lightweight CNN modules. However, existing designs often overlook runtime overhead from memory-bound operations and the shift in feature characteristics from spatial-dominant to semantic-dominant as networks deepen. This work introduces TinyNeXt, a family of efficient hybrid ViTs for TinyML, featuring Lean Single-Head Self-Attention to minimize memory-bound operations, and a macro design tailored to feature characteristics at different stages. TinyNeXt strikes a better accuraciespeed trade-off across diverse tasks and hardware platforms, outperforming state-of-the-art models of comparable scale. For instance, our TinyNeXt-T achieves a remarkable $71.5\\%$ top-1 accuracy with only 1.0M parameters on ImageNet-1K. Furthermore, compared to recent efficient models like MobileViT-XXS and MobileViT-XS, TinyNeXt-S and TinyNeXt-M achieve $3.7\\% /0.5\\%$ higher accuracy, respectively, while running $2.1\\times /2.6\\times$ faster on Nvidia Jetson Nano. The code and trained models are available at https://github.com/yuffeenn/TinyNeXt.", + "bbox": [ + 88, + 335, + 485, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Vision Transformers (ViTs) [11] have demonstrated superior performance across diverse computer vision tasks, including classification[44], detection [2, 59], and segmentation [5, 54, 61]. As a key component of ViT, Multi-Head Self-Attention (MHSA) captures interrelations between any two spatial locations, significantly enhancing global information comprehension. Nonetheless, the quadratic complexity and substantial memory demands", + "bbox": [ + 89, + 753, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9c0a96ddb012f6c180bb16cf2288e59f7ab4148706435201f44a3af723c9f299.jpg", + "image_caption": [ + "Figure 1. Throughput vs Accuracy Comparison." + ], + "image_footnote": [], + "bbox": [ + 519, + 308, + 901, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of MHSA pose significant challenges for deploying ViTs on edge devices.", + "bbox": [ + 511, + 580, + 906, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the above challenges, researchers have introduced a range of lightweight design strategies. One approach integrates Convolutional Neural Networks (CNNs) with ViTs, such as employing lightweight convolutions in early stages [25, 26, 53] to reduce computations. Another approach focuses on optimizing self-attention mechanisms to reduce the number of visual tokens, thereby decreasing computational costs. This includes restricting attention to specific regions [10, 30, 45, 63] and dynamically pruning redundant tokens [15, 40, 57]. Leveraging CNN-Transformer hybrids and innovative attention modules, current lightweight ViT [26, 60] models have progressively surpassed lightweight CNN models[41, 43] in terms of accuracy and speed.", + "bbox": [ + 509, + 613, + 906, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, prior works have overlooked several critical aspects in the design of efficient ViTs, both at micro and macro architectural levels. At the micro level, dense matrix multiplication operations are commonly regarded as the throughput bottleneck in attention module inference. Many", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59b607e1be7661a3ce6b8d7944e7f2234a49d23a4b5b810bc57cb8be29280d03.jpg", + "image_caption": [ + "Figure 2. Runtime Profiling of Different Vision Transformers. Beyond matrix multiplication, both DeiT and Swin models involve costly memory-bound operations, such as Reshape, Copy, and LayerNorm. Notably, in the Swin-S model, the latency of these memory-bound operations exceeds $50\\%$ . PyTorch Profiler [39] was used as the latency analysis tool." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 462, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e1fd852072007a3b4556abfb7151b41d1c323c6fa9264b5146dc969f09116131.jpg", + "image_caption": [ + "Figure 3. Feature maps across stages of PVT [48], showing the transition from spatial to channel dominance." + ], + "image_footnote": [], + "bbox": [ + 109, + 359, + 464, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "attention variants aim to reduce theoretical FLOPs but inadvertently introduce significant runtime overhead from noncomputational operations. As illustrated in Fig. 2, a substantial portion of Swin-S's [30] runtime is consumed by memory-bound operations, including tensor reshaping, data copying, and layer normalization. This observation highlights the need for hardware-friendly attention mechanisms to achieve efficient inference.", + "bbox": [ + 89, + 597, + 482, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At the macro design level, driven by computational cost considerations, existing efficient ViTs [16, 34, 58] often adopt a hybrid architecture: using embedding layers or hierarchical downsampling convolutional layers to reduce spatial resolution, followed by stacking transformer blocks in the later stages. While this design achieves a better tradeoff between accuracy and latency, it may overlook the evolution of feature representations across different network depths. Our preliminary investigation reveals an intriguing phenomenon: as the network deepens, feature characteristics gradually shift from being spatial-dominant to semantic-dominant. This phenomenon is clearly illustrated", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in Fig 3, which visualizes the feature maps at different stages of the Pyramid Vision Transformer [49] architecture. Notably, when feature maps are downsampled to a $7 \\times 7$ resolution in the last stage, spatial information diminishes significantly, indicating a fundamental shift in feature representation.", + "bbox": [ + 511, + 90, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This observation aligns with previous findings in the Non-Local Networks [50], where researchers systematically investigated the effect of adding non-local blocks at different stages of ResNet [19] architectures. Their results showed that incorporating non-local blocks in the last stage (res5), where spatial dimensions are reduced to $7 \\times 7$ , yielded minimal performance improvements. These findings suggest that the common practice of placing spatial attention blocks in the last stage may not be optimal.", + "bbox": [ + 511, + 181, + 906, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these findings, we propose TinyNeXt, a family of efficient hybrid Vision Transformers for Tiny Machine Learning (TinyML) applications. In micro design, we introduce Lean Single-Head Self-Attention (LeanSHSA), which employs a single-head architecture to minimize latency from reshape and softmax operations, while merging projection matrices to form a more streamlined attention mechanism. Additionally, we refine the macro design of ViT architectures by tailoring module to feature characteristics at different stages. Experiments demonstrate that our models achieve superior parameter efficiency, latency, and accuracy compared to state-of-the-art (SOTA) lightweight models, while avoiding complex designs. Specifically, TinyNeXt-T achieves $71.5\\%$ top-1 accuracy on ImageNet-1K with only 1.0M parameters. Furthermore, TinyNeXt-S/M attains $72.7\\% / 75.3\\%$ top-1 accuracy with $1.3 / 2.3\\mathrm{M}$ parameters and $14.3 / 19.4~\\mathrm{ms}$ latency on Nvidia Jetson Nano, outperforming EMO-1M/2M [60] by $+1.2\\uparrow / + 0.2\\uparrow$ with a speedup of $1.83\\times / 1.84\\times$ . When integrated with SSDlite framework for object detection, TinyNeXt-S/M achieves $22.4 / 25.0\\mathrm{mAP}$ with $2.3 / 3.3\\mathrm{M}$ parameters, surpassing MobileViTv2 [35] by $+1.2\\uparrow / + 0.4\\uparrow$ with a speedup of $1.2\\times / 1.4\\times$ on Nvidia RTX 3090. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 316, + 908, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Lean Single Head Self-Attention, a simplified and efficient attention mechanism that minimizes memory-inefficient operations and runs faster than traditional MHSA.", + "- We refine the macro design of lightweight ViTs by introducing stage-specific feature extraction modules, replacing spatial attention with channel-wise attention in the last stage.", + "- We conduct comprehensive experiments demonstrating that TinyNeXt achieves a superior accuracy-speed trade-off across diverse vision tasks and hardware platforms." + ], + "bbox": [ + 511, + 665, + 903, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 844, + 656, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Over the past decade, CNNs have significantly influenced computer vision, becoming the basis for many tasks. Mod-", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19915", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els like AlexNet [24], VGGNet [42], ResNet [19], and SENet [23] have rapidly advanced, substantially enhancing accuracy across various vision tasks. But their substantial parameter and computational costs impede deployment on mobile and edge devices. To address these challenges, previous work has proposed numerous lightweight techniques, including efficient network design [22, 27, 41, 56], network architecture search [43, 64], low-bit quantization [12, 37, 62], knowledge distillation [20, 55], and network pruning [17, 36]. However, CNNs are spatially local and lack global interaction between the features, limiting further performance improvements.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent years have witnessed the emergence of Transformer-based models outperforming CNN-based models in the computer vision domain. Dosovitskiy et al. [11] first introduced the transformer architecture to vision tasks and ViT's excellent performance has inspired numerous subsequent improvements. Subsequent research has reduced ViT's limitations through techniques such as distillation token [44, 52], local attention mechanism [30], and hierarchical architecture [48, 49].", + "bbox": [ + 89, + 272, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, to deploy ViTs on resource-constrained devices, efficient ViT designs are rapidly evolving. At the micro design level, in addition to embedding depth-wise convolutions in attention modules, numerous works have proposed various attention variants to reduce computational complexity. Swin Transformer [30] and follow-up works [10, 45, 63] limit self-attention computation to crafted local windows, interacting window information through additional operations. MobileViT [34] reduces computational complexity by decreasing the number of visual tokens in self-attention via unfold operation. Unfortunately, as MHSA inherently involves numerous memory-inefficient operations, these attention variants inevitably introduce additional overhead, resulting in suboptimal speed. EfficientViT [29] and SHViT [58] conduct a systematic analysis of redundancy in MHSA and propose a set of memory-efficient design principles. Their models strike a good trade-off between speed and accuracy at the cost of lower parameter efficiency. At the macro design level, hybrid architectures [16, 18, 60] that employ lightweight convolutions in early stages and transformers in later stages have become increasingly popular. This design introduces locality to ViTs and avoids using MHSA in high-resolution early stages to reduce computational costs.", + "bbox": [ + 91, + 409, + 482, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 784, + 181, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the quadratic computational complexity of the self-attention mechanism, prior research has predominantly focused on designing efficient ViTs from a computational cost perspective. While these efforts have yielded promising results, current designs overlook several important factors, resulting in suboptimal model performance and inference", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "speed. This section first motivates the need for improved designs by analyzing the limitations of existing approaches from both micro and macro perspectives. Subsequently, we introduce our optimized designs, aiming for a better tradeoff between model size, accuracy, and speed.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 511, + 179, + 633, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Microscopic Perspective: Recent studies have proposed various attention variants to optimize computational efficiency. However, most designs primarily focus on theoretical FLOPs reduction without considering practical constraints. Fig. 2 presents a latency breakdown in Swin-S [30] (window-based attention), revealing substantial overhead from memory-inefficient operations. Moreover, some attention variants[1, 8] often require custom CUDA kernel implementations and specialized optimization techniques, increasing development complexity and limit deployment flexibility across heterogeneous computing platforms. The absence of hardware-aware design creates a discrepancy between theoretical efficiency metrics and actual runtime performance, particularly in resource-constrained environments. Therefore, a simplified yet effective attention module with spatial feature learning capabilities is crucial for deploying ViTs in TinyML applications.", + "bbox": [ + 511, + 202, + 903, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Macroscopic Perspective: From the perspective of macroscopic network architecture, extensive research [16, 25, 58] has employed either embedding layers with aggressive downsampling factors or hierarchical downsampling modules to reduce spatial resolution, subsequently cascading transformer blocks at the latter stages where the resolution has already shrunk enough. It is well known that MHSA possesses a robust capability for extracting global spatial information. However, through visualizing the feature maps, it has been observed that features exhibit more pronounced semantic characteristics in the last stage of the network, rather than spatial characteristics. To investigate the potential impact of this architectural mismatch on model performance, this study conducts preliminary experiments on models such as MobileViT[34] and SHViT[58], replacing the attention modules in the last stage of the network with Squeeze-and-Excitation (SE) [23] modules. As shown in Tab. 1, this strategic adjustment yields accuracy improvements on the ImageNet-1K benchmark.", + "bbox": [ + 511, + 459, + 903, + 746 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/5096dc8c8eba52bf3ab1c9dea8affcffeea18f1ed46fc98b46019502ecc0962d.jpg", + "table_caption": [ + "Table 1. Replacing the last stage module with SE, † denotes our reproduction." + ], + "table_footnote": [], + "table_body": "
ModelParams (M)Top-1 (%)Flops (M)
MobileViT-XXS[34]†1.370.0418
MobileViT-XXS[34] with SE1.370.8411
SHViT-S2[58]†11.474.9366
SHViT-S2[58] with SE11.475.3352
", + "bbox": [ + 517, + 801, + 908, + 880 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19916", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Lean Single-Head Self-Attention", + "text_level": 1, + "bbox": [ + 89, + 90, + 377, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The MHSA module is a key component of the transformer, designed to effectively capture global spatial information from input tokens. In particular, for token embeddings $\\mathbf{X} \\in \\mathbb{R}^{n \\times d}$ , the MHSA is computed as follows:", + "bbox": [ + 89, + 112, + 483, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\quad [ \\mathbf {Q} _ {1}, \\mathbf {Q} _ {2}, \\dots , \\mathbf {Q} _ {h} ] = \\mathbf {Q}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 181, + 413, + 199 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\quad \\left[ \\mathbf {K} _ {1}, \\mathbf {K} _ {2}, \\dots , \\mathbf {K} _ {h} \\right] = \\mathbf {K} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 200, + 480, + 217 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\quad [ \\mathbf {V} _ {1}, \\mathbf {V} _ {2}, \\dots , \\mathbf {V} _ {h} ] = \\mathbf {V}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 219, + 415, + 237 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {i} = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {T}}{\\sqrt {d _ {h}}} \\right] \\cdot \\mathbf {V} _ {i} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 247, + 480, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} = \\operatorname {C o n c a t} \\left[ \\mathbf {A} _ {1}, \\mathbf {A} _ {2}, \\dots , \\mathbf {A} _ {h} \\right] \\cdot \\mathbf {W} _ {o} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 292, + 480, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\mathbf{W}_q\\in \\mathbb{R}^{d\\times d_q}$ , $\\mathbf{W}_k\\in \\mathbb{R}^{d\\times d_k}$ , $\\mathbf{W}_v\\in \\mathbb{R}^{d\\times d_v}$ denote the projection matrices for queries, keys, and values, respectively. Generally, it holds that $d_{q} = d_{k} = d_{v} = d_{h}\\cdot h$ , where $d_h$ represents the dimension of each head and $h$ denotes the number of heads.", + "bbox": [ + 89, + 314, + 483, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From the above equations, it is evident that the MHSA module involves multiple linear transformations and matrix multiplications, contributing significantly to inference latency. In addition to dense matrix multiplication operations, the MHSA module also encompasses memory inefficient operations such as reshape and softmax. These operations significantly impact memory usage and access patterns during computation, thereby affecting the actual inference speed of the model.", + "bbox": [ + 89, + 391, + 483, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When $h$ equals one, MHSA degenerates into Single-Head Self-Attention (SHSA). In this simplified configuration, the computational process eliminates the need for reshape operations that would otherwise partition query, key, and value matrices into multiple sub-matrices. Furthermore, since both the memory requirements for storing attention maps and the computational complexity of softmax operations scale linearly with the number of attention heads, the reduction to single-head attention significantly decreases memory consumption and latency associated with memory-inefficient operations.", + "bbox": [ + 89, + 527, + 483, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\left(\\mathbf {X W} _ {q}\\right) \\left(\\mathbf {X W} _ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] \\cdot \\left(\\mathbf {X W} _ {v}\\right) \\mathbf {W} _ {o} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 702, + 482, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computational complexity of matrix multiplication in SHSA is $O(n^{2} \\cdot d_{h})$ . To reduce computational costs, Non-Local Networks [50] employ a bottleneck strategy by setting $d_{h}$ to $\\frac{1}{2}d$ , while SHViT [58] adopts a sampling approach by only taking a small portion of $X$ as input. Both approaches effectively reduce the head dimension, thereby decreasing the computational overhead. However, while reducing the head dimension can alleviate computational burden, it may compromise the model's ability to capture complex relationships and features. As", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f68d0d734f15a6a1d94e27208de58449aa4c8ad9b24fde96fbbd79f6d921c0fa.jpg", + "image_caption": [ + "a) Multi-Head Self-Attention" + ], + "image_footnote": [], + "bbox": [ + 537, + 90, + 648, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a6b4462abc4a569d245303ac2926e86dbb27fd4a430e5afbc7a621c0df4f3e7d.jpg", + "image_caption": [ + "b) Single-Head Self-Attention" + ], + "image_footnote": [], + "bbox": [ + 653, + 90, + 764, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0874ad63621494d8295fbefb3512ab83aab6b6b499d253c0e3b14847bab08e8c.jpg", + "image_caption": [ + "c) Lean Single-Head Self-Attention" + ], + "image_footnote": [], + "bbox": [ + 769, + 90, + 882, + 191 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8ebf034e1e0577f63cbb9420ffe4d55fdab333a1e53ca294af6cf8a8bbbc07a6.jpg", + "image_caption": [ + "Figure 4. Comparison with different attention modules.", + "MatMul", + "Reshape", + "Figure 5. Runtime breakdown comparison of different attention modules. The results are measured on an NVIDIA RTX 3090 and an AMD EPYC 7542 32-Core Processor with configurations $b = 128$ (batch size), $n = 196$ (number of tokens), $d = 192$ (token dimensionality), and $h = 3$ (number of heads for MHSA), using PyTorch Profiler [39]." + ], + "image_footnote": [], + "bbox": [ + 540, + 247, + 705, + 439 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1578acb4f00aa6e0d23e966b1af14f07758e2c70bfb97af660f59d66ee9acfd.jpg", + "image_caption": [ + "Elementwise Mul", + "Softmax" + ], + "image_footnote": [], + "bbox": [ + 709, + 248, + 877, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the dimension of each attention head decreases, the feature space it can represent becomes increasingly constrained, ultimately limiting the model's expressive capacity. Through mathematical analysis, we observe that $\\mathbf{Q}\\mathbf{K}^{\\mathrm{T}} = (\\mathbf{X}\\mathbf{W}_{q})(\\mathbf{X}\\mathbf{W}_{k})^{T} = \\mathbf{X}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T})\\mathbf{X}^{T}$ . The absence of non-linearity between $Q$ and $K$ imposes a rank constraint: $\\mathrm{rank}(\\mathbf{Q}\\mathbf{K}) \\leq \\mathrm{rank}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T}) \\leq d$ . This constraint reveals a critical trade-off. When $d_{h} < d$ , the model's expressiveness suffers from rank deficiency. Conversely, $d_{h} > d$ introduces redundant parameters without theoretical benefits.", + "bbox": [ + 511, + 582, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L e a n S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {X W} _ {1} \\mathbf {X} ^ {T}}{\\sqrt {d}} \\right] \\cdot \\mathbf {X W} _ {2} \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 771, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To optimize this trade-off, we propose Lean Single-Head Self-Attention(LeanSHSA), as illustrated in Eq. 5. This approach sets $d_h = d$ and merges projection matrices: $\\mathbf{W}_q\\mathbf{W}_k^T$ into $\\mathbf{W}_1$ and $\\mathbf{W}_v\\mathbf{W}_o$ into $\\mathbf{W}_2$ . By reducing four weight matrices to two, LeanSHSA signifi", + "bbox": [ + 511, + 824, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19917", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/564e3e9be8a38feca0db4661caf6501a4cefbc5144625824c7ec05c8aecddd52.jpg", + "image_caption": [ + "b) ConvBlock", + "c) SpatialBlock", + "d) ChannelBlock", + "Figure 6. Overview of TinyNeXt: A hierarchical architecture featuring a convolutional stem, four progressively designed stages with specific modules (ConvBlock, SpatialBlock, ChannelBlock), and a final classification head. $H$ and $W$ represent input image spatial dimensions, with $L_{i}$ indicating the block count per stage." + ], + "image_footnote": [], + "bbox": [ + 140, + 97, + 861, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cantly simplifies the attention mechanism while maintaining its theoretical expressiveness. Furthermore, the simplified matrix operations align better with modern hardware accelerators, enabling more efficient parallel processing and better utilization of computing resources. Through these simplifications, LeanSHSA effectively reduces the latency caused by memory-inefficient operations, achieving a near $2 \\times$ improvement in CPU throughput compared to MHSA, as demonstrated in Fig. 5.", + "bbox": [ + 89, + 585, + 482, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 89, + 738, + 225, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on our previous analysis and findings, we systematically optimize ViT architectures by adapting module designs to the evolving feature characteristics across network stages. Fig. 6 depicts an overview of the proposed TinyNeXt architecture. TinyNeXt follows the pyramid architecture, comprising a sequential arrangement of a stem, four stages of varying resolutions, and a classification head, with downsampling layers connecting the successive stages.", + "bbox": [ + 89, + 762, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The stem layer serves as the initial feature extraction", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "module, processing the input image and producing a feature representation. Similar to [25, 47], we stack two standard $3 \\times 3$ convolutional layers with a stride of 2 as the stem, downsampling the input image by a factor of 4.", + "bbox": [ + 511, + 585, + 906, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the early network stages (Stage 1 and Stage 2), characterized by localized feature representations, we employ the ConvBlock, which originates from the inverted residual block [41]. This choice is driven by the need to efficiently capture local spatial information, such as texture patterns and edge details, which are predominantly concentrated in the initial layers of the network. By leveraging depth-wise separable convolutions and an efficient expansion-compression structure, it effectively extracts fine-grained details and patterns from the input data while maintaining computational efficiency.", + "bbox": [ + 511, + 652, + 908, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the middle stage (Stage 3), after multiple downsampling operations, the feature scales become more moderate, and the low-frequency spatial information is retained. In this case, we have employed the SpatialBlock to effectively extract the long-range dependencies between features. The", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19918", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/13b1a108e26ea8399222ac3e2826e17f71452ed724ff08d145bc605831236df4.jpg", + "table_caption": [ + "Table 2. Classification performance on ImageNet-1K. Throughput is measured on an Nvidia RTX 3090 GPU with batch size of 512 for GPU and AMD EPYC 7542 32-Core Processor with batch size of 16 for CPU. Larger throughput means faster inference speed. FLOP count is computed by fvcore library [14]." + ], + "table_footnote": [], + "table_body": "
ModelInputParams (M)FLOPs (M)Throughput (img/s)Top-1 (%)Top-5 (%)
GPUCPU
ShuffleNetv2-1.0 [32]2242.31461208510469.488.9
PVTv2-b0 [49]2243.757225862670.5-
MobileNetv1 [22]2244.257953104570.6-
EfficientViT-M2 [43]2244.22011668814170.890.2
MobileOne-S0 [46]2242.127584107971.4-
TinyNeXt-T2241.025972578571.590.3
MobileNetv2-1.0 [41]2243.530045723872.091.0
MobileViTv2-0.5 [35]2561.446644352270.2-
MobileViT-XXS [34]2561.341033742369.0-
EdgeNeXt-XXS [33]2561.326146796471.2-
EMO-1M [60]2241.326139494071.590.4
TinyNeXt-S2241.330465557572.791.0
MobileNetv2-1.4 [41]2246.158530911974.7-
MobileNetv3-L-1.0 [21]2245.421958646175.2-
MobileOne-S1 [46]2244.882545543175.9-
EdgeViT-XXS [38]2244.155733263674.4-
MobileViT-XS [34]2562.39862028874.8-
EdgeNeXt-XS [33]2562.353829063275.0-
EMO-2M [60]2242.343927792975.192.2
TinyNeXt-M2242.347548895175.392.3
", + "bbox": [ + 120, + 142, + 874, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SpatialBlock is mainly composed of three components: a LeanSHSA layer for comprehensive global context modeling, a $3 \\times 3$ lightweight Depthwise Convolution (DWConv) layer dedicated to local feature aggregation or conditional position embedding [6, 53], and a Multilayer Perceptron (MLP) for channel feature interaction. Formally, the SpatialBlock can be represented as follows:", + "bbox": [ + 89, + 527, + 483, + 635 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {1} = \\operatorname {L e a n S H S A} \\left(\\mathrm {B N} \\left(X _ {\\text {i n}}\\right)\\right) + X _ {\\text {i n}},\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 642, + 408, + 659 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {2} = \\operatorname {D W C o n v} \\left(X _ {1}\\right) + X _ {1},\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 661, + 352, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nX _ {\\text {o u t}} = \\operatorname {M L P} (\\operatorname {B N} (X _ {2})) + X _ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 680, + 367, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the last stage (Stage 4), as feature scales further contract and spatial characteristics diminish, semantic representations become increasingly pronounced. Consequently, we deploy the ChannelBlock, which leverages a SE module to dynamically recalibrate channel-wise features. This module, structurally akin to the SpatialBlock but replacing the LeanSHSA layer with an SE module, enables adaptive feature emphasis and enhances the network's discriminative potential.", + "bbox": [ + 89, + 704, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, in designing TinyNeXt, we have avoided overly complex operations, focusing instead on selecting the most appropriate basic modules based on the evolving characteristics of feature. This approach ensures a balance between", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "computational efficiency and feature representation capability. We build three TinyNext variants with different setting of depth, width and expansion ratio, the detailed model specifications are demonstrated as Tab 3.", + "bbox": [ + 511, + 527, + 906, + 589 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/90d1b5eabca3d4f406d2a725d3532cec5415ac5a62a9a2d05523cef41370c7f2.jpg", + "table_caption": [ + "Table 3. Architecture details of TinyNeXt model variants." + ], + "table_footnote": [], + "table_body": "
ModelsTinyNeXt-TTinyNeXt-STinyNeXt-M
Depth[3, 3, 6, 2][3, 3, 8, 3][4, 4, 9, 4]
Width[32, 64, 96, 192][32, 64, 96, 192][32, 64, 128, 256]
Exp.Ratio[2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 1.5]
", + "bbox": [ + 519, + 628, + 897, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 717, + 645, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Image Classification", + "text_level": 1, + "bbox": [ + 511, + 742, + 705, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We conduct image classification experiments on ImageNet-1K [9] using PyTorch 2.3.0 [39] and Timm library 0.5.4 [51]. All models are trained from scratch at a resolution of $224 \\times 224$ for 300 epochs, utilizing 8 Nvidia RTX 3090 GPUs. We employ the AdamW optimizer [31], warm-up training with 5 epochs and a cosine annealing learning rate scheduler. The total batch size is set as 2,048 and the initial learning rate is $6 \\times 10^{-3}$ with weight decay of $2.5 \\times 10^{-2}$ . The data augmentations ap", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19919", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "plied during training included Random Resized Crop, Random Horizontal Flip, and Color Jitter, without the use of Mixup, CutMix, or Random Erasing. Moreover, we evaluated the model's throughput across desktop GPU and CPU platforms. On the GPU platform, throughput measurements were conducted using an NVIDIA RTX 3090 with a batch size of 512. For CPU throughput evaluation, we employed an AMD EPYC 7542 32-Core Processor, performing measurements with a batch size of 16 in a single-threaded.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. TinyNeXt is systematically benchmarked against SOTA lightweight models across multiple scales, with the comprehensive evaluation results detailed in Tab. 2. TinyNeXt achieves competitive or superior results in terms of accuracy, parameter efficiency, and inference throughput. TinyNeXt-T achieves an impressive $71.5\\%$ top-1 accuracy with only 1.0M parameters, demonstrating notable improvements over MobileNetv1 [22] and PVTv2-b0 [48]. Specifically, it achieves an accuracy gain of $0.9\\% \\uparrow$ and $1.0\\% \\uparrow$ , respectively, while reducing parameters by $3.2\\mathrm{M} \\downarrow$ and $2.7\\mathrm{M} \\downarrow$ , with corresponding GPU throughput enhancements of $1.4\\times$ and $2.8\\times$ . At the small scale, TinyNeXt-S, with a comparable parameters of 1.3M to MobileViT-XXS, EdgeNeXt-XXS, and EMO-1M [33, 34, 60], significantly outperforms these models in both accuracy and multi-platform throughput. It achieves a $3.7\\% \\uparrow$ improvement in top-1 accuracy over MobileViT-XXS, with GPU throughput increasing by $1.9\\times$ . Compared to the classic lightweight CNN model MobileNetv2 [41], TinyNeXt-S demonstrates remarkable performance, improving $0.7\\% \\uparrow$ top-1 accuracy and reducing $37\\% \\downarrow$ parameters. Moreover, TinyNeXt-M achieves the highest top-1 accuracy of $75.3\\%$ under the same parameter constraints when compared with MobileViT-XS, EdgeNeXt-XS, and EMO-2M [33, 34, 60], while delivering an exceptional CPU throughput that is $6.4\\times$ that of MobileViT-XS. With its excellent balanced accuracy and efficiency, TinyNeXt demonstrates potential as a viable option for resource-constrained applications.", + "bbox": [ + 91, + 227, + 486, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 89, + 665, + 272, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object Detection. We integrate TinyNeXt into the SSDLite framework [21] for object detection on the MS COCO 2017 dataset [28] at $320 \\times 320$ resolution. The implementation leverages the MMDetection library [3], employing the AdamW optimizer [31] with a batch size of 64, and inference throughput (images/s) is benchmarked on an NVIDIA RTX 3090 GPU. As shown in Table 4, TinyNeXt demonstrates superior computational efficiency with competitive accuracy. Specifically, TinyNeXt-S achieves $22.4\\mathrm{mAP}$ using only $2.3\\mathrm{M}$ parameters and 2,517 images/s throughput. It reduces parameters by $2.8 / 2.0\\mathrm{M}$ compared to MobileNetv1/v2[22, 41], and surpasses MobileViTv2-0.5/EMO-1M[35, 60] by $1.2 / 0.4\\mathrm{mAP}$ with $1.2 / 1.5 \\times$ throughput. The medium-scale TinyNeXt-M", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/60821d57e77543e7f5d5bb183446cd87da4f8e8e7b1b5ec3f17030d658b8eea8.jpg", + "table_caption": [ + "Table 4. Object detection results on MS COCO 2017 with the SSDLite framework." + ], + "table_footnote": [], + "table_body": "
BackboneParams↓Throughput↑mAP↑
MobileNetv1 [22]5.1M228222.2
MobileNetv2 [41]4.3M202922.1
MobileViT-XXS [34]1.7M152819.9
MobileViTv2-0.5[35]2.0M205821.2
EMO-1M[60]2.3M168122.0
TinyNeXt-S2.3M251722.4
MobileViTv2-0.75[35]3.6M138824.6
EMO-2M[60]3.3M119425.2
TinyNeXt-M3.3M191125.0
", + "bbox": [ + 517, + 128, + 903, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f1cb72b520c6ecaf273613631aa410f952df86d902d3b52415f769ca0df363db.jpg", + "table_caption": [ + "Table 5. Semantic segmentation results on PASCAL VOC 2012 with the DeepLabv3 framework." + ], + "table_footnote": [], + "table_body": "
BackboneParams↓Flops↓mIOU↑
MobileNetV2 [41]4.5M5.8G75.7
MobileViTv2-0.75[35]9.6M40.0G75.1
TinyNeXt-S3.5M3.5G75.5
TinyNeXt-M4.9M5.1G76.9
", + "bbox": [ + 524, + 351, + 890, + 436 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45238727ed2c0bd67ad42c8135a26e46d31135b98531bc869c82a3d02251e868.jpg", + "table_caption": [ + "Table 6. Performance Comparison on Edge Devices. The last two columns present inference latencies measured on edge computing platforms: NVIDIA Jetson NANO and ARM Cortex-A57." + ], + "table_footnote": [], + "table_body": "
ModelInputParams (M)Top1 (%)NANO (ms)Cortex-A57 (ms)
MobileViT-XXS [34]25621.369.030.1467.9
EdgeNeXt-XXS [33]25621.371.233.5242.8
EMO-1M [60]22421.371.526.2526.4
TinyNeXt-S22421.372.714.3227.7
MobileViT-XS [34]25622.374.750.8859.1
EdgeNeXt-XS [33]25622.375.050.8442.4
EMO-2M [60]22422.375.135.7759.2
TinyNeXt-M22422.375.319.4311.6
", + "bbox": [ + 517, + 510, + 903, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "attains $25.0\\mathrm{mAP}$ with 3.3M parameters, matching EMO-2M in accuracy while providing $1.6\\times$ faster inference.", + "bbox": [ + 511, + 715, + 903, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Segmentation. We integrate TinyNeXt into the DeepLabv3 framework [4] for semantic segmentation on the Pascal VOC 2012 dataset [13] at $512 \\times 512$ resolution. Our implementation utilizes the MMSegmentation library [7], trains with the AdamW optimizer [31] for 40K iterations using a total batch size of 16. As summarized in Table 5, both variants provide superior accuracy and efficiency. Specifically, TinyNeXt-M surpasses MobileNetV2 by $1.2\\%$ mIOU with 0.7G fewer FLOPs, further demonstrating its effectiveness as a vision backbone across tasks.", + "bbox": [ + 509, + 750, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19920", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Inference on Edge Devices", + "text_level": 1, + "bbox": [ + 89, + 90, + 328, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation Details. To validate our model's potential for TinyML applications, we conduct a comprehensive evaluation of TinyNeXt's inference performance against SOTA models, including MobileViT [34], EdgeNeXt [33], and EMO [60], across diverse edge computing platforms. Specifically, we evaluate their inference capabilities on edge GPU and CPU architectures, utilizing platforms such as the NVIDIA Jetson Nano and ARM Cortex-A57. For each platform, we implement a unified testing flow: an initial 10s warm-up phase followed by 60s of continuous inference to accurately measure average latency.", + "bbox": [ + 89, + 112, + 480, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results. TinyNeXt achieves a superior accuracy-latency trade-off across all evaluated edge platforms. With 1.3M parameters, TinyNeXt-S demonstrates $1.8 \\times$ faster inference on NVIDIA Jetson Nano while maintaining $1.2\\%$ higher top-1 accuracy, and reduces ARM Cortex-A57 latency by $56.7\\%$ compared to EMO-1M [60]. At 2.3M parameters, TinyNeXt-M further keeps these advantages, achieving $1.8 \\times$ speedup over EMO-2M [60] on edge GPU with $0.2\\%$ accuracy improvement, while delivering $2.8 \\times$ faster CPU execution than MobileViT-XS [34].", + "bbox": [ + 89, + 279, + 480, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 440, + 243, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies on key design elements of the proposed TinyNeXt using the ImageNet-1K dataset [9]. To highlight architectural differences and reduce training time, all models are trained for 100 epochs [16, 29]. Throughout the ablation experiments, we consistently preserve the baseline architecture's original depth and width configurations. This approach facilitates a precise evaluation of individual design choices' performance impact.", + "bbox": [ + 89, + 462, + 482, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the attention block. These ablation experiments validate LeanSHSA's efficiency through substitutions with standard MHSA and SHSA modules. As shown in Tab. 7 (#2 and #3), replacing LeanSHSA with 4-head MHSA increases parameters by $0.3\\mathrm{M}$ for merely $0.2\\%$ accuracy gain, while incurring $19\\%$ throughput penalty. This empirically demonstrates that LeanSHSA's streamlined computational design effectively enhances inference efficiency without compromising global feature extraction capabilities. Furthermore, adopting a bottlenecked SHSA approach marginally reduces accuracy by $0.2\\%$ despite equivalent parameter count, with only a $5\\%$ speed gain. This suggests the importance of maintaining head dimensionality for better feature extraction in single-head architecture.", + "bbox": [ + 89, + 598, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the last stage choice. This ablation experiment evaluates the effectiveness of TinyNeXt's macro architecture by modifying the feature extraction module in the last stage. As shown in Tab. 7 (#4, #5, #6), we replace the channel-wise attention module (SE) with various global spatial attention modules, including MHSA (with 4 heads", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4fc9ff745698fc33b23fa23df9e42287e001d2784e2c901f68ec72709bce81ec.jpg", + "table_caption": [ + "Table 7. Ablation Study of TinyNeXt-M on ImageNet-1K [9]. The last three columns present model parameters (M), top-1 accuracy (\\%), and GPU throughput (images/s)." + ], + "table_footnote": [], + "table_body": "
#AblationParamsTop-1Throughput
1TinyNeXt-M (baseline)2.374.14889
2LeanSHSA → MHSA2.674.33950
3LeanSHSA → SHSA2.373.95137
4Last Stage with MHSA2.774.14718
5Last Stage with SHSA2.773.64788
6Last Stage with LeanSHSA2.773.94743
7BN → LN2.373.84479
", + "bbox": [ + 526, + 142, + 893, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and a head dimension of 32), SHSA, and LeanSHSA. These substitutions increase the model size by 0.4M parameters compared to the baseline, but yield no accuracy gains and even slightly reduce throughput. For example, using MHSA in the last stage (#4) raises parameters from 2.3M to 2.7M and cuts throughput by nearly 200 images/s, while the top-1 accuracy remains unchanged at $74.1\\%$ . Similar trends are observed with SHSA and LeanSHSA variants. These results demonstrate that channel-wise attention is more effective in deep layers, where feature representations shift from being spatially dominant to semantically dominant, offering superior semantic discrimination.", + "bbox": [ + 511, + 299, + 906, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of the normalization layer. We investigate the impact of normalization layers by replacing Batch Normalization (BN) with Layer Normalization (LN) in SpatialBlock and ChannelBlock. LN requires dynamic mean and variance computation during inference, which poses challenges for edge device deployment. Empirical evaluation (#7) reveals that replacing BN with LN results in a $8\\%$ reduction in inference throughput and a marginal $0.3\\%$ decline in top-1 accuracy. Consequently, we retain BN as the preferred normalization strategy in our TinyNeXt architecture.", + "bbox": [ + 511, + 483, + 908, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 657, + 633, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper analyzes the limitations of current efficient ViTs at both micro and macro levels and proposes corresponding optimizations. Specifically, we introduce a simplified and efficient attention variant, LeanSHSA, and enhance model performance through stage-specific macro designs. Extensive experiments demonstrate TinyNeXt's exceptional trade-off of model size, accuracy, and speed across diverse vision tasks and hardware platforms.", + "bbox": [ + 511, + 685, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The computational complexity of Lean-SHSA remains quadratic with visual tokens, leading to high costs in high-resolution tasks and restricting its applicability in such scenarios. In future work, we aim to explore dynamic attention mechanisms while minimizing the overhead of memory-bound operations.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19921", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science and Technology Innovation 2030-Major Projects under Grant 2021ZD0114400, in part by the Young Scientists Fund of the National Natural Science Foundation of China No.62304162, in part by the China Postdoctoral Science Foundation under Grant Number 2024M762532, in part by the Postdoctoral Fellowship Program of CPSF under Grant GZC20241313, in part by Shaanxi Provincial Natural Science Foundation for Basic Research Program under Grant 2024JC-YBMS-794, and in part by the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24090.", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 309, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV). Springer, 2020. 1", + "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7", + "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1", + "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 6", + "[7] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7", + "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 3", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 8", + "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of" + ], + "bbox": [ + 94, + 335, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1, 3", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3", + "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 3", + "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7", + "[14] facebookresearch. fvcore library. https://github.com/facebookresearch/fvcore, 2019.6", + "[15] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision, pages 396-414. Springer, 2022. 1", + "[16] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 2, 3, 8", + "[17] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 3", + "[18] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189, 2023. 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3", + "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3", + "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 6, 7", + "[22] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3, 6, 7", + "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3", + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19922", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 1, 3, 5", + "[26] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16889-16900, 2023. 1", + "[27] Yishi Li, Yuhao Zhang, and Rui Lai. Tinypillarnet: Tiny pillar-based network for 3d point cloud object detection at edge. IEEE Transactions on Circuits and Systems for Video Technology, 34(3):1772-1785, 2023. 3", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7", + "[29] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3, 8", + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3", + "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 7", + "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6", + "[33] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. In European conference on computer vision, pages 3-20. Springer, 2022. 6, 7, 8", + "[34] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 3, 6, 7, 8", + "[35] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 2, 6, 7", + "[36] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 3", + "[37] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, pages 294-311. Springer, 2022. 6", + "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 2, 4, 6", + "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1", + "[41] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 5, 6, 7", + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3", + "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1, 3, 6", + "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 3", + "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 1, 3", + "[46] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7907-7917, 2023. 6", + "[47] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15909-15920, 2024. 5", + "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2, 3, 7", + "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2, 3, 6" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19923", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2, 4", + "[51] Ross Wightman et al. Pytorch image models, 2019. 6", + "[52] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European conference on computer vision, pages 68-85. Springer, 2022. 3", + "[53] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 1, 6", + "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 12077-12090, 2021. 1", + "[55] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3", + "[56] Kunran Xu, Yishi Li, Huawei Zhang, Rui Lai, and Lin Gu. Etinynet: Extremely tiny network for tinyml. In Proceedings of the AAAI conference on artificial intelligence, pages 4628-4636, 2022. 3", + "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1", + "[58] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5756-5767, 2024. 2, 3, 4", + "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jianming Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2022. 1", + "[60] Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, and Chengjie Wang. Rethinking mobile block for efficient attention-based models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1389-1400. IEEE Computer Society, 2023. 1, 2, 3, 6, 7, 8", + "[61] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6881-6890, 2021. 1", + "[62] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 1, 3", + "[64] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018. 3" + ], + "bbox": [ + 516, + 90, + 903, + 229 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19924", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_model.json b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35f5c8be95f2b2ae18906d066fb0fe0331e1ddb9 --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "An Efficient Hybrid Vision Transformer for TinyML Applications" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.18, + 0.705, + 0.199 + ], + "angle": 0, + "content": "Fanhong Zeng\\(^{1,2}\\) Huanan Li\\(^{1,2}\\) Juntao Guan\\(^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.199, + 0.737, + 0.218 + ], + "angle": 0, + "content": "Rui Fan\\(^{1,2}\\) Tong Wu\\(^{1,2}\\) Xilong Wang\\(^{1,2}\\) Rui Lai\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.218, + 0.776, + 0.235 + ], + "angle": 0, + "content": "1 Key Laboratory of Analog Integrated Circuits and Systems (Ministry of Education)" + }, + { + "type": "text", + "bbox": [ + 0.294, + 0.235, + 0.705, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Faculty of Integrated Circuit, Xidian University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.259, + 0.253, + 0.739, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\)Hangzhou Institute of Technology, Xidian University, Hangzhou, China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.701 + ], + "angle": 0, + "content": "To enable the deployment of Vision Transformers on resource-constrained mobile and edge devices, the development of efficient ViT models has attracted significant attention. Researchers achieved remarkable improvements in accuracy and speed by optimizing attention mechanisms and integrating lightweight CNN modules. However, existing designs often overlook runtime overhead from memory-bound operations and the shift in feature characteristics from spatial-dominant to semantic-dominant as networks deepen. This work introduces TinyNeXt, a family of efficient hybrid ViTs for TinyML, featuring Lean Single-Head Self-Attention to minimize memory-bound operations, and a macro design tailored to feature characteristics at different stages. TinyNeXt strikes a better accuraciespeed trade-off across diverse tasks and hardware platforms, outperforming state-of-the-art models of comparable scale. For instance, our TinyNeXt-T achieves a remarkable \\(71.5\\%\\) top-1 accuracy with only 1.0M parameters on ImageNet-1K. Furthermore, compared to recent efficient models like MobileViT-XXS and MobileViT-XS, TinyNeXt-S and TinyNeXt-M achieve \\(3.7\\% /0.5\\%\\) higher accuracy, respectively, while running \\(2.1\\times /2.6\\times\\) faster on Nvidia Jetson Nano. The code and trained models are available at https://github.com/yuffeenn/TinyNeXt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Recently, Vision Transformers (ViTs) [11] have demonstrated superior performance across diverse computer vision tasks, including classification[44], detection [2, 59], and segmentation [5, 54, 61]. As a key component of ViT, Multi-Head Self-Attention (MHSA) captures interrelations between any two spatial locations, significantly enhancing global information comprehension. Nonetheless, the quadratic complexity and substantial memory demands" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.309, + 0.903, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.54, + 0.862, + 0.555 + ], + "angle": 0, + "content": "Figure 1. Throughput vs Accuracy Comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.613 + ], + "angle": 0, + "content": "of MHSA pose significant challenges for deploying ViTs on edge devices." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.907, + 0.825 + ], + "angle": 0, + "content": "To tackle the above challenges, researchers have introduced a range of lightweight design strategies. One approach integrates Convolutional Neural Networks (CNNs) with ViTs, such as employing lightweight convolutions in early stages [25, 26, 53] to reduce computations. Another approach focuses on optimizing self-attention mechanisms to reduce the number of visual tokens, thereby decreasing computational costs. This includes restricting attention to specific regions [10, 30, 45, 63] and dynamically pruning redundant tokens [15, 40, 57]. Leveraging CNN-Transformer hybrids and innovative attention modules, current lightweight ViT [26, 60] models have progressively surpassed lightweight CNN models[41, 43] in terms of accuracy and speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, prior works have overlooked several critical aspects in the design of efficient ViTs, both at micro and macro architectural levels. At the micro level, dense matrix multiplication operations are commonly regarded as the throughput bottleneck in attention module inference. Many" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.464, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.486, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Runtime Profiling of Different Vision Transformers. Beyond matrix multiplication, both DeiT and Swin models involve costly memory-bound operations, such as Reshape, Copy, and LayerNorm. Notably, in the Swin-S model, the latency of these memory-bound operations exceeds \\(50\\%\\). PyTorch Profiler [39] was used as the latency analysis tool." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.361, + 0.465, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.571 + ], + "angle": 0, + "content": "Figure 3. Feature maps across stages of PVT [48], showing the transition from spatial to channel dominance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.718 + ], + "angle": 0, + "content": "attention variants aim to reduce theoretical FLOPs but inadvertently introduce significant runtime overhead from noncomputational operations. As illustrated in Fig. 2, a substantial portion of Swin-S's [30] runtime is consumed by memory-bound operations, including tensor reshaping, data copying, and layer normalization. This observation highlights the need for hardware-friendly attention mechanisms to achieve efficient inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "At the macro design level, driven by computational cost considerations, existing efficient ViTs [16, 34, 58] often adopt a hybrid architecture: using embedding layers or hierarchical downsampling convolutional layers to reduce spatial resolution, followed by stacking transformer blocks in the later stages. While this design achieves a better tradeoff between accuracy and latency, it may overlook the evolution of feature representations across different network depths. Our preliminary investigation reveals an intriguing phenomenon: as the network deepens, feature characteristics gradually shift from being spatial-dominant to semantic-dominant. This phenomenon is clearly illustrated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.181 + ], + "angle": 0, + "content": "in Fig 3, which visualizes the feature maps at different stages of the Pyramid Vision Transformer [49] architecture. Notably, when feature maps are downsampled to a \\(7 \\times 7\\) resolution in the last stage, spatial information diminishes significantly, indicating a fundamental shift in feature representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.318 + ], + "angle": 0, + "content": "This observation aligns with previous findings in the Non-Local Networks [50], where researchers systematically investigated the effect of adding non-local blocks at different stages of ResNet [19] architectures. Their results showed that incorporating non-local blocks in the last stage (res5), where spatial dimensions are reduced to \\(7 \\times 7\\), yielded minimal performance improvements. These findings suggest that the common practice of placing spatial attention blocks in the last stage may not be optimal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.909, + 0.664 + ], + "angle": 0, + "content": "Based on these findings, we propose TinyNeXt, a family of efficient hybrid Vision Transformers for Tiny Machine Learning (TinyML) applications. In micro design, we introduce Lean Single-Head Self-Attention (LeanSHSA), which employs a single-head architecture to minimize latency from reshape and softmax operations, while merging projection matrices to form a more streamlined attention mechanism. Additionally, we refine the macro design of ViT architectures by tailoring module to feature characteristics at different stages. Experiments demonstrate that our models achieve superior parameter efficiency, latency, and accuracy compared to state-of-the-art (SOTA) lightweight models, while avoiding complex designs. Specifically, TinyNeXt-T achieves \\(71.5\\%\\) top-1 accuracy on ImageNet-1K with only 1.0M parameters. Furthermore, TinyNeXt-S/M attains \\(72.7\\% / 75.3\\%\\) top-1 accuracy with \\(1.3 / 2.3\\mathrm{M}\\) parameters and \\(14.3 / 19.4~\\mathrm{ms}\\) latency on Nvidia Jetson Nano, outperforming EMO-1M/2M [60] by \\(+1.2\\uparrow / + 0.2\\uparrow\\) with a speedup of \\(1.83\\times / 1.84\\times\\). When integrated with SSDlite framework for object detection, TinyNeXt-S/M achieves \\(22.4 / 25.0\\mathrm{mAP}\\) with \\(2.3 / 3.3\\mathrm{M}\\) parameters, surpassing MobileViTv2 [35] by \\(+1.2\\uparrow / + 0.4\\uparrow\\) with a speedup of \\(1.2\\times / 1.4\\times\\) on Nvidia RTX 3090. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.725 + ], + "angle": 0, + "content": "- We propose Lean Single Head Self-Attention, a simplified and efficient attention mechanism that minimizes memory-inefficient operations and runs faster than traditional MHSA." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.726, + 0.905, + 0.786 + ], + "angle": 0, + "content": "- We refine the macro design of lightweight ViTs by introducing stage-specific feature extraction modules, replacing spatial attention with channel-wise attention in the last stage." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.787, + 0.905, + 0.832 + ], + "angle": 0, + "content": "- We conduct comprehensive experiments demonstrating that TinyNeXt achieves a superior accuracy-speed trade-off across diverse vision tasks and hardware platforms." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.666, + 0.905, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.657, + 0.861 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Over the past decade, CNNs have significantly influenced computer vision, becoming the basis for many tasks. Mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19915" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.272 + ], + "angle": 0, + "content": "els like AlexNet [24], VGGNet [42], ResNet [19], and SENet [23] have rapidly advanced, substantially enhancing accuracy across various vision tasks. But their substantial parameter and computational costs impede deployment on mobile and edge devices. To address these challenges, previous work has proposed numerous lightweight techniques, including efficient network design [22, 27, 41, 56], network architecture search [43, 64], low-bit quantization [12, 37, 62], knowledge distillation [20, 55], and network pruning [17, 36]. However, CNNs are spatially local and lack global interaction between the features, limiting further performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.408 + ], + "angle": 0, + "content": "Recent years have witnessed the emergence of Transformer-based models outperforming CNN-based models in the computer vision domain. Dosovitskiy et al. [11] first introduced the transformer architecture to vision tasks and ViT's excellent performance has inspired numerous subsequent improvements. Subsequent research has reduced ViT's limitations through techniques such as distillation token [44, 52], local attention mechanism [30], and hierarchical architecture [48, 49]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.41, + 0.483, + 0.771 + ], + "angle": 0, + "content": "Currently, to deploy ViTs on resource-constrained devices, efficient ViT designs are rapidly evolving. At the micro design level, in addition to embedding depth-wise convolutions in attention modules, numerous works have proposed various attention variants to reduce computational complexity. Swin Transformer [30] and follow-up works [10, 45, 63] limit self-attention computation to crafted local windows, interacting window information through additional operations. MobileViT [34] reduces computational complexity by decreasing the number of visual tokens in self-attention via unfold operation. Unfortunately, as MHSA inherently involves numerous memory-inefficient operations, these attention variants inevitably introduce additional overhead, resulting in suboptimal speed. EfficientViT [29] and SHViT [58] conduct a systematic analysis of redundancy in MHSA and propose a set of memory-efficient design principles. Their models strike a good trade-off between speed and accuracy at the cost of lower parameter efficiency. At the macro design level, hybrid architectures [16, 18, 60] that employ lightweight convolutions in early stages and transformers in later stages have become increasingly popular. This design introduces locality to ViTs and avoids using MHSA in high-resolution early stages to reduce computational costs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.182, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Given the quadratic computational complexity of the self-attention mechanism, prior research has predominantly focused on designing efficient ViTs from a computational cost perspective. While these efforts have yielded promising results, current designs overlook several important factors, resulting in suboptimal model performance and inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "speed. This section first motivates the need for improved designs by analyzing the limitations of existing approaches from both micro and macro perspectives. Subsequently, we introduce our optimized designs, aiming for a better tradeoff between model size, accuracy, and speed." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.18, + 0.635, + 0.194 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.905, + 0.459 + ], + "angle": 0, + "content": "Microscopic Perspective: Recent studies have proposed various attention variants to optimize computational efficiency. However, most designs primarily focus on theoretical FLOPs reduction without considering practical constraints. Fig. 2 presents a latency breakdown in Swin-S [30] (window-based attention), revealing substantial overhead from memory-inefficient operations. Moreover, some attention variants[1, 8] often require custom CUDA kernel implementations and specialized optimization techniques, increasing development complexity and limit deployment flexibility across heterogeneous computing platforms. The absence of hardware-aware design creates a discrepancy between theoretical efficiency metrics and actual runtime performance, particularly in resource-constrained environments. Therefore, a simplified yet effective attention module with spatial feature learning capabilities is crucial for deploying ViTs in TinyML applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.905, + 0.747 + ], + "angle": 0, + "content": "Macroscopic Perspective: From the perspective of macroscopic network architecture, extensive research [16, 25, 58] has employed either embedding layers with aggressive downsampling factors or hierarchical downsampling modules to reduce spatial resolution, subsequently cascading transformer blocks at the latter stages where the resolution has already shrunk enough. It is well known that MHSA possesses a robust capability for extracting global spatial information. However, through visualizing the feature maps, it has been observed that features exhibit more pronounced semantic characteristics in the last stage of the network, rather than spatial characteristics. To investigate the potential impact of this architectural mismatch on model performance, this study conducts preliminary experiments on models such as MobileViT[34] and SHViT[58], replacing the attention modules in the last stage of the network with Squeeze-and-Excitation (SE) [23] modules. As shown in Tab. 1, this strategic adjustment yields accuracy improvements on the ImageNet-1K benchmark." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.762, + 0.905, + 0.79 + ], + "angle": 0, + "content": "Table 1. Replacing the last stage module with SE, † denotes our reproduction." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.802, + 0.909, + 0.881 + ], + "angle": 0, + "content": "
ModelParams (M)Top-1 (%)Flops (M)
MobileViT-XXS[34]†1.370.0418
MobileViT-XXS[34] with SE1.370.8411
SHViT-S2[58]†11.474.9366
SHViT-S2[58] with SE11.475.3352
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.378, + 0.108 + ], + "angle": 0, + "content": "3.2. Lean Single-Head Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.173 + ], + "angle": 0, + "content": "The MHSA module is a key component of the transformer, designed to effectively capture global spatial information from input tokens. In particular, for token embeddings \\(\\mathbf{X} \\in \\mathbb{R}^{n \\times d}\\), the MHSA is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.183, + 0.415, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\quad [ \\mathbf {Q} _ {1}, \\mathbf {Q} _ {2}, \\dots , \\mathbf {Q} _ {h} ] = \\mathbf {Q}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.202, + 0.482, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\quad \\left[ \\mathbf {K} _ {1}, \\mathbf {K} _ {2}, \\dots , \\mathbf {K} _ {h} \\right] = \\mathbf {K} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.22, + 0.416, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\quad [ \\mathbf {V} _ {1}, \\mathbf {V} _ {2}, \\dots , \\mathbf {V} _ {h} ] = \\mathbf {V}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.248, + 0.482, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {i} = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {T}}{\\sqrt {d _ {h}}} \\right] \\cdot \\mathbf {V} _ {i} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.293, + 0.482, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} = \\operatorname {C o n c a t} \\left[ \\mathbf {A} _ {1}, \\mathbf {A} _ {2}, \\dots , \\mathbf {A} _ {h} \\right] \\cdot \\mathbf {W} _ {o} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.484, + 0.391 + ], + "angle": 0, + "content": "Here \\(\\mathbf{W}_q\\in \\mathbb{R}^{d\\times d_q}\\), \\(\\mathbf{W}_k\\in \\mathbb{R}^{d\\times d_k}\\), \\(\\mathbf{W}_v\\in \\mathbb{R}^{d\\times d_v}\\) denote the projection matrices for queries, keys, and values, respectively. Generally, it holds that \\(d_{q} = d_{k} = d_{v} = d_{h}\\cdot h\\), where \\(d_h\\) represents the dimension of each head and \\(h\\) denotes the number of heads." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.527 + ], + "angle": 0, + "content": "From the above equations, it is evident that the MHSA module involves multiple linear transformations and matrix multiplications, contributing significantly to inference latency. In addition to dense matrix multiplication operations, the MHSA module also encompasses memory inefficient operations such as reshape and softmax. These operations significantly impact memory usage and access patterns during computation, thereby affecting the actual inference speed of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.695 + ], + "angle": 0, + "content": "When \\( h \\) equals one, MHSA degenerates into Single-Head Self-Attention (SHSA). In this simplified configuration, the computational process eliminates the need for reshape operations that would otherwise partition query, key, and value matrices into multiple sub-matrices. Furthermore, since both the memory requirements for storing attention maps and the computational complexity of softmax operations scale linearly with the number of attention heads, the reduction to single-head attention significantly decreases memory consumption and latency associated with memory-inefficient operations." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathbf {S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\left(\\mathbf {X W} _ {q}\\right) \\left(\\mathbf {X W} _ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] \\cdot \\left(\\mathbf {X W} _ {v}\\right) \\mathbf {W} _ {o} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The computational complexity of matrix multiplication in SHSA is \\( O(n^{2} \\cdot d_{h}) \\). To reduce computational costs, Non-Local Networks [50] employ a bottleneck strategy by setting \\( d_{h} \\) to \\( \\frac{1}{2}d \\), while SHViT [58] adopts a sampling approach by only taking a small portion of \\( X \\) as input. Both approaches effectively reduce the head dimension, thereby decreasing the computational overhead. However, while reducing the head dimension can alleviate computational burden, it may compromise the model's ability to capture complex relationships and features. As" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.091, + 0.65, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.195, + 0.641, + 0.204 + ], + "angle": 0, + "content": "a) Multi-Head Self-Attention" + }, + { + "type": "image", + "bbox": [ + 0.654, + 0.091, + 0.765, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.195, + 0.756, + 0.204 + ], + "angle": 0, + "content": "b) Single-Head Self-Attention" + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.091, + 0.883, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.195, + 0.885, + 0.204 + ], + "angle": 0, + "content": "c) Lean Single-Head Self-Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.218, + 0.874, + 0.232 + ], + "angle": 0, + "content": "Figure 4. Comparison with different attention modules." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.248, + 0.707, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.445, + 0.613, + 0.454 + ], + "angle": 0, + "content": "MatMul" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.445, + 0.684, + 0.455 + ], + "angle": 0, + "content": "Reshape" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.445, + 0.787, + 0.455 + ], + "angle": 0, + "content": "Elementwise Mul" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.445, + 0.855, + 0.455 + ], + "angle": 0, + "content": "Softmax" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.249, + 0.879, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Figure 5. Runtime breakdown comparison of different attention modules. The results are measured on an NVIDIA RTX 3090 and an AMD EPYC 7542 32-Core Processor with configurations \\( b = 128 \\) (batch size), \\( n = 196 \\) (number of tokens), \\( d = 192 \\) (token dimensionality), and \\( h = 3 \\) (number of heads for MHSA), using PyTorch Profiler [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.75 + ], + "angle": 0, + "content": "the dimension of each attention head decreases, the feature space it can represent becomes increasingly constrained, ultimately limiting the model's expressive capacity. Through mathematical analysis, we observe that \\(\\mathbf{Q}\\mathbf{K}^{\\mathrm{T}} = (\\mathbf{X}\\mathbf{W}_{q})(\\mathbf{X}\\mathbf{W}_{k})^{T} = \\mathbf{X}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T})\\mathbf{X}^{T}\\). The absence of non-linearity between \\(Q\\) and \\(K\\) imposes a rank constraint: \\(\\mathrm{rank}(\\mathbf{Q}\\mathbf{K}) \\leq \\mathrm{rank}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T}) \\leq d\\). This constraint reveals a critical trade-off. When \\(d_{h} < d\\), the model's expressiveness suffers from rank deficiency. Conversely, \\(d_{h} > d\\) introduces redundant parameters without theoretical benefits." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.772, + 0.907, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L e a n S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {X W} _ {1} \\mathbf {X} ^ {T}}{\\sqrt {d}} \\right] \\cdot \\mathbf {X W} _ {2} \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To optimize this trade-off, we propose Lean Single-Head Self-Attention(LeanSHSA), as illustrated in Eq. 5. This approach sets \\( d_h = d \\) and merges projection matrices: \\( \\mathbf{W}_q\\mathbf{W}_k^T \\) into \\( \\mathbf{W}_1 \\) and \\( \\mathbf{W}_v\\mathbf{W}_o \\) into \\( \\mathbf{W}_2 \\). By reducing four weight matrices to two, LeanSHSA signifi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19917" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.098, + 0.862, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.483, + 0.264, + 0.498 + ], + "angle": 0, + "content": "b) ConvBlock" + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.484, + 0.445, + 0.498 + ], + "angle": 0, + "content": "c) SpatialBlock" + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.484, + 0.63, + 0.497 + ], + "angle": 0, + "content": "d) ChannelBlock" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 6. Overview of TinyNeXt: A hierarchical architecture featuring a convolutional stem, four progressively designed stages with specific modules (ConvBlock, SpatialBlock, ChannelBlock), and a final classification head. \\(H\\) and \\(W\\) represent input image spatial dimensions, with \\(L_{i}\\) indicating the block count per stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.483, + 0.723 + ], + "angle": 0, + "content": "cantly simplifies the attention mechanism while maintaining its theoretical expressiveness. Furthermore, the simplified matrix operations align better with modern hardware accelerators, enabling more efficient parallel processing and better utilization of computing resources. Through these simplifications, LeanSHSA effectively reduces the latency caused by memory-inefficient operations, achieving a near \\(2 \\times\\) improvement in CPU throughput compared to MHSA, as demonstrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.226, + 0.753 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Based on our previous analysis and findings, we systematically optimize ViT architectures by adapting module designs to the evolving feature characteristics across network stages. Fig. 6 depicts an overview of the proposed TinyNeXt architecture. TinyNeXt follows the pyramid architecture, comprising a sequential arrangement of a stem, four stages of varying resolutions, and a classification head, with downsampling layers connecting the successive stages." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The stem layer serves as the initial feature extraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.907, + 0.648 + ], + "angle": 0, + "content": "module, processing the input image and producing a feature representation. Similar to [25, 47], we stack two standard \\(3 \\times 3\\) convolutional layers with a stride of 2 as the stem, downsampling the input image by a factor of 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.82 + ], + "angle": 0, + "content": "In the early network stages (Stage 1 and Stage 2), characterized by localized feature representations, we employ the ConvBlock, which originates from the inverted residual block [41]. This choice is driven by the need to efficiently capture local spatial information, such as texture patterns and edge details, which are predominantly concentrated in the initial layers of the network. By leveraging depth-wise separable convolutions and an efficient expansion-compression structure, it effectively extracts fine-grained details and patterns from the input data while maintaining computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In the middle stage (Stage 3), after multiple downsampling operations, the feature scales become more moderate, and the low-frequency spatial information is retained. In this case, we have employed the SpatialBlock to effectively extract the long-range dependencies between features. The" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19918" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 2. Classification performance on ImageNet-1K. Throughput is measured on an Nvidia RTX 3090 GPU with batch size of 512 for GPU and AMD EPYC 7542 32-Core Processor with batch size of 16 for CPU. Larger throughput means faster inference speed. FLOP count is computed by fvcore library [14]." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.143, + 0.875, + 0.507 + ], + "angle": 0, + "content": "
ModelInputParams (M)FLOPs (M)Throughput (img/s)Top-1 (%)Top-5 (%)
GPUCPU
ShuffleNetv2-1.0 [32]2242.31461208510469.488.9
PVTv2-b0 [49]2243.757225862670.5-
MobileNetv1 [22]2244.257953104570.6-
EfficientViT-M2 [43]2244.22011668814170.890.2
MobileOne-S0 [46]2242.127584107971.4-
TinyNeXt-T2241.025972578571.590.3
MobileNetv2-1.0 [41]2243.530045723872.091.0
MobileViTv2-0.5 [35]2561.446644352270.2-
MobileViT-XXS [34]2561.341033742369.0-
EdgeNeXt-XXS [33]2561.326146796471.2-
EMO-1M [60]2241.326139494071.590.4
TinyNeXt-S2241.330465557572.791.0
MobileNetv2-1.4 [41]2246.158530911974.7-
MobileNetv3-L-1.0 [21]2245.421958646175.2-
MobileOne-S1 [46]2244.882545543175.9-
EdgeViT-XXS [38]2244.155733263674.4-
MobileViT-XS [34]2562.39862028874.8-
EdgeNeXt-XS [33]2562.353829063275.0-
EMO-2M [60]2242.343927792975.192.2
TinyNeXt-M2242.347548895175.392.3
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.529, + 0.484, + 0.636 + ], + "angle": 0, + "content": "SpatialBlock is mainly composed of three components: a LeanSHSA layer for comprehensive global context modeling, a \\(3 \\times 3\\) lightweight Depthwise Convolution (DWConv) layer dedicated to local feature aggregation or conditional position embedding [6, 53], and a Multilayer Perceptron (MLP) for channel feature interaction. Formally, the SpatialBlock can be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.643, + 0.409, + 0.66 + ], + "angle": 0, + "content": "\\[\nX _ {1} = \\operatorname {L e a n S H S A} \\left(\\mathrm {B N} \\left(X _ {\\text {i n}}\\right)\\right) + X _ {\\text {i n}},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.662, + 0.354, + 0.678 + ], + "angle": 0, + "content": "\\[\nX _ {2} = \\operatorname {D W C o n v} \\left(X _ {1}\\right) + X _ {1},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.681, + 0.369, + 0.697 + ], + "angle": 0, + "content": "\\[\nX _ {\\text {o u t}} = \\operatorname {M L P} (\\operatorname {B N} (X _ {2})) + X _ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In the last stage (Stage 4), as feature scales further contract and spatial characteristics diminish, semantic representations become increasingly pronounced. Consequently, we deploy the ChannelBlock, which leverages a SE module to dynamically recalibrate channel-wise features. This module, structurally akin to the SpatialBlock but replacing the LeanSHSA layer with an SE module, enables adaptive feature emphasis and enhances the network's discriminative potential." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Overall, in designing TinyNeXt, we have avoided overly complex operations, focusing instead on selecting the most appropriate basic modules based on the evolving characteristics of feature. This approach ensures a balance between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.59 + ], + "angle": 0, + "content": "computational efficiency and feature representation capability. We build three TinyNext variants with different setting of depth, width and expansion ratio, the detailed model specifications are demonstrated as Tab 3." + }, + { + "type": "table_caption", + "bbox": [ + 0.538, + 0.604, + 0.882, + 0.618 + ], + "angle": 0, + "content": "Table 3. Architecture details of TinyNeXt model variants." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.629, + 0.898, + 0.689 + ], + "angle": 0, + "content": "
ModelsTinyNeXt-TTinyNeXt-STinyNeXt-M
Depth[3, 3, 6, 2][3, 3, 8, 3][4, 4, 9, 4]
Width[32, 64, 96, 192][32, 64, 96, 192][32, 64, 128, 256]
Exp.Ratio[2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 1.5]
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.718, + 0.646, + 0.735 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.706, + 0.759 + ], + "angle": 0, + "content": "4.1. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We conduct image classification experiments on ImageNet-1K [9] using PyTorch 2.3.0 [39] and Timm library 0.5.4 [51]. All models are trained from scratch at a resolution of \\(224 \\times 224\\) for 300 epochs, utilizing 8 Nvidia RTX 3090 GPUs. We employ the AdamW optimizer [31], warm-up training with 5 epochs and a cosine annealing learning rate scheduler. The total batch size is set as 2,048 and the initial learning rate is \\(6 \\times 10^{-3}\\) with weight decay of \\(2.5 \\times 10^{-2}\\). The data augmentations ap" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.227 + ], + "angle": 0, + "content": "plied during training included Random Resized Crop, Random Horizontal Flip, and Color Jitter, without the use of Mixup, CutMix, or Random Erasing. Moreover, we evaluated the model's throughput across desktop GPU and CPU platforms. On the GPU platform, throughput measurements were conducted using an NVIDIA RTX 3090 with a batch size of 512. For CPU throughput evaluation, we employed an AMD EPYC 7542 32-Core Processor, performing measurements with a batch size of 16 in a single-threaded." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.487, + 0.653 + ], + "angle": 0, + "content": "Results. TinyNeXt is systematically benchmarked against SOTA lightweight models across multiple scales, with the comprehensive evaluation results detailed in Tab. 2. TinyNeXt achieves competitive or superior results in terms of accuracy, parameter efficiency, and inference throughput. TinyNeXt-T achieves an impressive \\(71.5\\%\\) top-1 accuracy with only 1.0M parameters, demonstrating notable improvements over MobileNetv1 [22] and PVTv2-b0 [48]. Specifically, it achieves an accuracy gain of \\(0.9\\% \\uparrow\\) and \\(1.0\\% \\uparrow\\), respectively, while reducing parameters by \\(3.2\\mathrm{M} \\downarrow\\) and \\(2.7\\mathrm{M} \\downarrow\\), with corresponding GPU throughput enhancements of \\(1.4\\times\\) and \\(2.8\\times\\). At the small scale, TinyNeXt-S, with a comparable parameters of 1.3M to MobileViT-XXS, EdgeNeXt-XXS, and EMO-1M [33, 34, 60], significantly outperforms these models in both accuracy and multi-platform throughput. It achieves a \\(3.7\\% \\uparrow\\) improvement in top-1 accuracy over MobileViT-XXS, with GPU throughput increasing by \\(1.9\\times\\). Compared to the classic lightweight CNN model MobileNetv2 [41], TinyNeXt-S demonstrates remarkable performance, improving \\(0.7\\% \\uparrow\\) top-1 accuracy and reducing \\(37\\% \\downarrow\\) parameters. Moreover, TinyNeXt-M achieves the highest top-1 accuracy of \\(75.3\\%\\) under the same parameter constraints when compared with MobileViT-XS, EdgeNeXt-XS, and EMO-2M [33, 34, 60], while delivering an exceptional CPU throughput that is \\(6.4\\times\\) that of MobileViT-XS. With its excellent balanced accuracy and efficiency, TinyNeXt demonstrates potential as a viable option for resource-constrained applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.273, + 0.681 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Object Detection. We integrate TinyNeXt into the SSDLite framework [21] for object detection on the MS COCO 2017 dataset [28] at \\(320 \\times 320\\) resolution. The implementation leverages the MMDetection library [3], employing the AdamW optimizer [31] with a batch size of 64, and inference throughput (images/s) is benchmarked on an NVIDIA RTX 3090 GPU. As shown in Table 4, TinyNeXt demonstrates superior computational efficiency with competitive accuracy. Specifically, TinyNeXt-S achieves \\(22.4\\mathrm{mAP}\\) using only \\(2.3\\mathrm{M}\\) parameters and 2,517 images/s throughput. It reduces parameters by \\(2.8 / 2.0\\mathrm{M}\\) compared to MobileNetv1/v2[22, 41], and surpasses MobileViTv2-0.5/EMO-1M[35, 60] by \\(1.2 / 0.4\\mathrm{mAP}\\) with \\(1.2 / 1.5 \\times\\) throughput. The medium-scale TinyNeXt-M" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 4. Object detection results on MS COCO 2017 with the SSDLite framework." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.129, + 0.905, + 0.293 + ], + "angle": 0, + "content": "
BackboneParams↓Throughput↑mAP↑
MobileNetv1 [22]5.1M228222.2
MobileNetv2 [41]4.3M202922.1
MobileViT-XXS [34]1.7M152819.9
MobileViTv2-0.5[35]2.0M205821.2
EMO-1M[60]2.3M168122.0
TinyNeXt-S2.3M251722.4
MobileViTv2-0.75[35]3.6M138824.6
EMO-2M[60]3.3M119425.2
TinyNeXt-M3.3M191125.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.312, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Table 5. Semantic segmentation results on PASCAL VOC 2012 with the DeepLabv3 framework." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.352, + 0.892, + 0.438 + ], + "angle": 0, + "content": "
BackboneParams↓Flops↓mIOU↑
MobileNetV2 [41]4.5M5.8G75.7
MobileViTv2-0.75[35]9.6M40.0G75.1
TinyNeXt-S3.5M3.5G75.5
TinyNeXt-M4.9M5.1G76.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.458, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Table 6. Performance Comparison on Edge Devices. The last two columns present inference latencies measured on edge computing platforms: NVIDIA Jetson NANO and ARM Cortex-A57." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.683 + ], + "angle": 0, + "content": "
ModelInputParams (M)Top1 (%)NANO (ms)Cortex-A57 (ms)
MobileViT-XXS [34]25621.369.030.1467.9
EdgeNeXt-XXS [33]25621.371.233.5242.8
EMO-1M [60]22421.371.526.2526.4
TinyNeXt-S22421.372.714.3227.7
MobileViT-XS [34]25622.374.750.8859.1
EdgeNeXt-XS [33]25622.375.050.8442.4
EMO-2M [60]22422.375.135.7759.2
TinyNeXt-M22422.375.319.4311.6
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.905, + 0.746 + ], + "angle": 0, + "content": "attains \\(25.0\\mathrm{mAP}\\) with 3.3M parameters, matching EMO-2M in accuracy while providing \\(1.6\\times\\) faster inference." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Semantic Segmentation. We integrate TinyNeXt into the DeepLabv3 framework [4] for semantic segmentation on the Pascal VOC 2012 dataset [13] at \\(512 \\times 512\\) resolution. Our implementation utilizes the MMSegmentation library [7], trains with the AdamW optimizer [31] for 40K iterations using a total batch size of 16. As summarized in Table 5, both variants provide superior accuracy and efficiency. Specifically, TinyNeXt-M surpasses MobileNetV2 by \\(1.2\\%\\) mIOU with 0.7G fewer FLOPs, further demonstrating its effectiveness as a vision backbone across tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19920" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.329, + 0.108 + ], + "angle": 0, + "content": "4.3. Inference on Edge Devices" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.482, + 0.279 + ], + "angle": 0, + "content": "Implementation Details. To validate our model's potential for TinyML applications, we conduct a comprehensive evaluation of TinyNeXt's inference performance against SOTA models, including MobileViT [34], EdgeNeXt [33], and EMO [60], across diverse edge computing platforms. Specifically, we evaluate their inference capabilities on edge GPU and CPU architectures, utilizing platforms such as the NVIDIA Jetson Nano and ARM Cortex-A57. For each platform, we implement a unified testing flow: an initial 10s warm-up phase followed by 60s of continuous inference to accurately measure average latency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.482, + 0.431 + ], + "angle": 0, + "content": "Results. TinyNeXt achieves a superior accuracy-latency trade-off across all evaluated edge platforms. With 1.3M parameters, TinyNeXt-S demonstrates \\(1.8 \\times\\) faster inference on NVIDIA Jetson Nano while maintaining \\(1.2\\%\\) higher top-1 accuracy, and reduces ARM Cortex-A57 latency by \\(56.7\\%\\) compared to EMO-1M [60]. At 2.3M parameters, TinyNeXt-M further keeps these advantages, achieving \\(1.8 \\times\\) speedup over EMO-2M [60] on edge GPU with \\(0.2\\%\\) accuracy improvement, while delivering \\(2.8 \\times\\) faster CPU execution than MobileViT-XS [34]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.245, + 0.457 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.483, + 0.599 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies on key design elements of the proposed TinyNeXt using the ImageNet-1K dataset [9]. To highlight architectural differences and reduce training time, all models are trained for 100 epochs [16, 29]. Throughout the ablation experiments, we consistently preserve the baseline architecture's original depth and width configurations. This approach facilitates a precise evaluation of individual design choices' performance impact." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Impact of the attention block. These ablation experiments validate LeanSHSA's efficiency through substitutions with standard MHSA and SHSA modules. As shown in Tab. 7 (#2 and #3), replacing LeanSHSA with 4-head MHSA increases parameters by \\(0.3\\mathrm{M}\\) for merely \\(0.2\\%\\) accuracy gain, while incurring \\(19\\%\\) throughput penalty. This empirically demonstrates that LeanSHSA's streamlined computational design effectively enhances inference efficiency without compromising global feature extraction capabilities. Furthermore, adopting a bottlenecked SHSA approach marginally reduces accuracy by \\(0.2\\%\\) despite equivalent parameter count, with only a \\(5\\%\\) speed gain. This suggests the importance of maintaining head dimensionality for better feature extraction in single-head architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Impact of the last stage choice. This ablation experiment evaluates the effectiveness of TinyNeXt's macro architecture by modifying the feature extraction module in the last stage. As shown in Tab. 7 (#4, #5, #6), we replace the channel-wise attention module (SE) with various global spatial attention modules, including MHSA (with 4 heads" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Table 7. Ablation Study of TinyNeXt-M on ImageNet-1K [9]. The last three columns present model parameters (M), top-1 accuracy (\\%), and GPU throughput (images/s)." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.143, + 0.894, + 0.27 + ], + "angle": 0, + "content": "
#AblationParamsTop-1Throughput
1TinyNeXt-M (baseline)2.374.14889
2LeanSHSA → MHSA2.674.33950
3LeanSHSA → SHSA2.373.95137
4Last Stage with MHSA2.774.14718
5Last Stage with SHSA2.773.64788
6Last Stage with LeanSHSA2.773.94743
7BN → LN2.373.84479
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.481 + ], + "angle": 0, + "content": "and a head dimension of 32), SHSA, and LeanSHSA. These substitutions increase the model size by 0.4M parameters compared to the baseline, but yield no accuracy gains and even slightly reduce throughput. For example, using MHSA in the last stage (#4) raises parameters from 2.3M to 2.7M and cuts throughput by nearly 200 images/s, while the top-1 accuracy remains unchanged at \\(74.1\\%\\). Similar trends are observed with SHSA and LeanSHSA variants. These results demonstrate that channel-wise attention is more effective in deep layers, where feature representations shift from being spatially dominant to semantically dominant, offering superior semantic discrimination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.909, + 0.636 + ], + "angle": 0, + "content": "Impact of the normalization layer. We investigate the impact of normalization layers by replacing Batch Normalization (BN) with Layer Normalization (LN) in SpatialBlock and ChannelBlock. LN requires dynamic mean and variance computation during inference, which poses challenges for edge device deployment. Empirical evaluation (#7) reveals that replacing BN with LN results in a \\(8\\%\\) reduction in inference throughput and a marginal \\(0.3\\%\\) decline in top-1 accuracy. Consequently, we retain BN as the preferred normalization strategy in our TinyNeXt architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.634, + 0.674 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.808 + ], + "angle": 0, + "content": "This paper analyzes the limitations of current efficient ViTs at both micro and macro levels and proposes corresponding optimizations. Specifically, we introduce a simplified and efficient attention variant, LeanSHSA, and enhance model performance through stage-specific macro designs. Extensive experiments demonstrate TinyNeXt's exceptional trade-off of model size, accuracy, and speed across diverse vision tasks and hardware platforms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations. The computational complexity of Lean-SHSA remains quadratic with visual tokens, leading to high costs in high-resolution tasks and restricting its applicability in such scenarios. In future work, we aim to explore dynamic attention mechanisms while minimizing the overhead of memory-bound operations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19921" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work was supported in part by the National Science and Technology Innovation 2030-Major Projects under Grant 2021ZD0114400, in part by the Young Scientists Fund of the National Natural Science Foundation of China No.62304162, in part by the China Postdoctoral Science Foundation under Grant Number 2024M762532, in part by the Postdoctoral Fellowship Program of CPSF under Grant GZC20241313, in part by Shaanxi Provincial Natural Science Foundation for Basic Research Program under Grant 2024JC-YBMS-794, and in part by the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24090." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.31, + 0.188, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[1] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.379, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV). Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.436, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.563, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[7] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.846, + 0.486, + 0.902 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.336, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.302 + ], + "angle": 0, + "content": "[14] facebookresearch. fvcore library. https://github.com/facebookresearch/fvcore, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[15] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision, pages 396-414. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[16] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[17] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[18] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[22] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19922" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[26] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16889-16900, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[27] Yishi Li, Yuhao Zhang, and Rui Lai. Tinypillarnet: Tiny pillar-based network for 3d point cloud object detection at edge. IEEE Transactions on Circuits and Systems for Video Technology, 34(3):1772-1785, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[29] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. In European conference on computer vision, pages 3-20. Springer, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[34] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[35] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[36] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[38] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, pages 294-311. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[41] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[46] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7907-7917, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[47] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15909-15920, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19923" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.445, + 0.163 + ], + "angle": 0, + "content": "[51] Ross Wightman et al. Pytorch image models, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[52] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European conference on computer vision, pages 68-85. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[53] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 12077-12090, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.405 + ], + "angle": 0, + "content": "[55] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[56] Kunran Xu, Yishi Li, Huawei Zhang, Rui Lai, and Lin Gu. Etinynet: Extremely tiny network for tinyml. In Proceedings of the AAAI conference on artificial intelligence, pages 4628-4636, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[58] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5756-5767, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jianming Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.663, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[60] Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, and Chengjie Wang. Rethinking mobile block for efficient attention-based models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1389-1400. IEEE Computer Society, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[61] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6881-6890, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[63] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[64] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19924" + } + ] +] \ No newline at end of file diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_origin.pdf b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4f2ae0662b7862c4c34f1905327f85d464ab6d49 --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/7d996d12-432e-4c35-99e7-c13b06e6d6ec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cee581c1a5fbc7c830e7cc36fd30d457ae593d7e1bc5e099690e379bdda9c563 +size 770059 diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/full.md b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b5b9e7eb1812409312bf16295ce8e856090b0bca --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/full.md @@ -0,0 +1,320 @@ +# An Efficient Hybrid Vision Transformer for TinyML Applications + +Fanhong Zeng $^{1,2}$ Huanan Li $^{1,2}$ Juntao Guan $^{1,3*}$ + +Rui Fan $^{1,2}$ Tong Wu $^{1,2}$ Xilong Wang $^{1,2}$ Rui Lai $^{1,2*}$ + +1 Key Laboratory of Analog Integrated Circuits and Systems (Ministry of Education) + +$^{2}$ Faculty of Integrated Circuit, Xidian University, Xi'an, China + +$^{3}$ Hangzhou Institute of Technology, Xidian University, Hangzhou, China + +# Abstract + +To enable the deployment of Vision Transformers on resource-constrained mobile and edge devices, the development of efficient ViT models has attracted significant attention. Researchers achieved remarkable improvements in accuracy and speed by optimizing attention mechanisms and integrating lightweight CNN modules. However, existing designs often overlook runtime overhead from memory-bound operations and the shift in feature characteristics from spatial-dominant to semantic-dominant as networks deepen. This work introduces TinyNeXt, a family of efficient hybrid ViTs for TinyML, featuring Lean Single-Head Self-Attention to minimize memory-bound operations, and a macro design tailored to feature characteristics at different stages. TinyNeXt strikes a better accuraciespeed trade-off across diverse tasks and hardware platforms, outperforming state-of-the-art models of comparable scale. For instance, our TinyNeXt-T achieves a remarkable $71.5\%$ top-1 accuracy with only 1.0M parameters on ImageNet-1K. Furthermore, compared to recent efficient models like MobileViT-XXS and MobileViT-XS, TinyNeXt-S and TinyNeXt-M achieve $3.7\% /0.5\%$ higher accuracy, respectively, while running $2.1\times /2.6\times$ faster on Nvidia Jetson Nano. The code and trained models are available at https://github.com/yuffeenn/TinyNeXt. + +# 1. Introduction + +Recently, Vision Transformers (ViTs) [11] have demonstrated superior performance across diverse computer vision tasks, including classification[44], detection [2, 59], and segmentation [5, 54, 61]. As a key component of ViT, Multi-Head Self-Attention (MHSA) captures interrelations between any two spatial locations, significantly enhancing global information comprehension. Nonetheless, the quadratic complexity and substantial memory demands + +![](images/9c0a96ddb012f6c180bb16cf2288e59f7ab4148706435201f44a3af723c9f299.jpg) +Figure 1. Throughput vs Accuracy Comparison. + +of MHSA pose significant challenges for deploying ViTs on edge devices. + +To tackle the above challenges, researchers have introduced a range of lightweight design strategies. One approach integrates Convolutional Neural Networks (CNNs) with ViTs, such as employing lightweight convolutions in early stages [25, 26, 53] to reduce computations. Another approach focuses on optimizing self-attention mechanisms to reduce the number of visual tokens, thereby decreasing computational costs. This includes restricting attention to specific regions [10, 30, 45, 63] and dynamically pruning redundant tokens [15, 40, 57]. Leveraging CNN-Transformer hybrids and innovative attention modules, current lightweight ViT [26, 60] models have progressively surpassed lightweight CNN models[41, 43] in terms of accuracy and speed. + +However, prior works have overlooked several critical aspects in the design of efficient ViTs, both at micro and macro architectural levels. At the micro level, dense matrix multiplication operations are commonly regarded as the throughput bottleneck in attention module inference. Many + +![](images/59b607e1be7661a3ce6b8d7944e7f2234a49d23a4b5b810bc57cb8be29280d03.jpg) +Figure 2. Runtime Profiling of Different Vision Transformers. Beyond matrix multiplication, both DeiT and Swin models involve costly memory-bound operations, such as Reshape, Copy, and LayerNorm. Notably, in the Swin-S model, the latency of these memory-bound operations exceeds $50\%$ . PyTorch Profiler [39] was used as the latency analysis tool. + +![](images/e1fd852072007a3b4556abfb7151b41d1c323c6fa9264b5146dc969f09116131.jpg) +Figure 3. Feature maps across stages of PVT [48], showing the transition from spatial to channel dominance. + +attention variants aim to reduce theoretical FLOPs but inadvertently introduce significant runtime overhead from noncomputational operations. As illustrated in Fig. 2, a substantial portion of Swin-S's [30] runtime is consumed by memory-bound operations, including tensor reshaping, data copying, and layer normalization. This observation highlights the need for hardware-friendly attention mechanisms to achieve efficient inference. + +At the macro design level, driven by computational cost considerations, existing efficient ViTs [16, 34, 58] often adopt a hybrid architecture: using embedding layers or hierarchical downsampling convolutional layers to reduce spatial resolution, followed by stacking transformer blocks in the later stages. While this design achieves a better tradeoff between accuracy and latency, it may overlook the evolution of feature representations across different network depths. Our preliminary investigation reveals an intriguing phenomenon: as the network deepens, feature characteristics gradually shift from being spatial-dominant to semantic-dominant. This phenomenon is clearly illustrated + +in Fig 3, which visualizes the feature maps at different stages of the Pyramid Vision Transformer [49] architecture. Notably, when feature maps are downsampled to a $7 \times 7$ resolution in the last stage, spatial information diminishes significantly, indicating a fundamental shift in feature representation. + +This observation aligns with previous findings in the Non-Local Networks [50], where researchers systematically investigated the effect of adding non-local blocks at different stages of ResNet [19] architectures. Their results showed that incorporating non-local blocks in the last stage (res5), where spatial dimensions are reduced to $7 \times 7$ , yielded minimal performance improvements. These findings suggest that the common practice of placing spatial attention blocks in the last stage may not be optimal. + +Based on these findings, we propose TinyNeXt, a family of efficient hybrid Vision Transformers for Tiny Machine Learning (TinyML) applications. In micro design, we introduce Lean Single-Head Self-Attention (LeanSHSA), which employs a single-head architecture to minimize latency from reshape and softmax operations, while merging projection matrices to form a more streamlined attention mechanism. Additionally, we refine the macro design of ViT architectures by tailoring module to feature characteristics at different stages. Experiments demonstrate that our models achieve superior parameter efficiency, latency, and accuracy compared to state-of-the-art (SOTA) lightweight models, while avoiding complex designs. Specifically, TinyNeXt-T achieves $71.5\%$ top-1 accuracy on ImageNet-1K with only 1.0M parameters. Furthermore, TinyNeXt-S/M attains $72.7\% / 75.3\%$ top-1 accuracy with $1.3 / 2.3\mathrm{M}$ parameters and $14.3 / 19.4~\mathrm{ms}$ latency on Nvidia Jetson Nano, outperforming EMO-1M/2M [60] by $+1.2\uparrow / + 0.2\uparrow$ with a speedup of $1.83\times / 1.84\times$ . When integrated with SSDlite framework for object detection, TinyNeXt-S/M achieves $22.4 / 25.0\mathrm{mAP}$ with $2.3 / 3.3\mathrm{M}$ parameters, surpassing MobileViTv2 [35] by $+1.2\uparrow / + 0.4\uparrow$ with a speedup of $1.2\times / 1.4\times$ on Nvidia RTX 3090. Our contributions can be summarized as follows: + +- We propose Lean Single Head Self-Attention, a simplified and efficient attention mechanism that minimizes memory-inefficient operations and runs faster than traditional MHSA. +- We refine the macro design of lightweight ViTs by introducing stage-specific feature extraction modules, replacing spatial attention with channel-wise attention in the last stage. +- We conduct comprehensive experiments demonstrating that TinyNeXt achieves a superior accuracy-speed trade-off across diverse vision tasks and hardware platforms. + +# 2. Related Work + +Over the past decade, CNNs have significantly influenced computer vision, becoming the basis for many tasks. Mod- + +els like AlexNet [24], VGGNet [42], ResNet [19], and SENet [23] have rapidly advanced, substantially enhancing accuracy across various vision tasks. But their substantial parameter and computational costs impede deployment on mobile and edge devices. To address these challenges, previous work has proposed numerous lightweight techniques, including efficient network design [22, 27, 41, 56], network architecture search [43, 64], low-bit quantization [12, 37, 62], knowledge distillation [20, 55], and network pruning [17, 36]. However, CNNs are spatially local and lack global interaction between the features, limiting further performance improvements. + +Recent years have witnessed the emergence of Transformer-based models outperforming CNN-based models in the computer vision domain. Dosovitskiy et al. [11] first introduced the transformer architecture to vision tasks and ViT's excellent performance has inspired numerous subsequent improvements. Subsequent research has reduced ViT's limitations through techniques such as distillation token [44, 52], local attention mechanism [30], and hierarchical architecture [48, 49]. + +Currently, to deploy ViTs on resource-constrained devices, efficient ViT designs are rapidly evolving. At the micro design level, in addition to embedding depth-wise convolutions in attention modules, numerous works have proposed various attention variants to reduce computational complexity. Swin Transformer [30] and follow-up works [10, 45, 63] limit self-attention computation to crafted local windows, interacting window information through additional operations. MobileViT [34] reduces computational complexity by decreasing the number of visual tokens in self-attention via unfold operation. Unfortunately, as MHSA inherently involves numerous memory-inefficient operations, these attention variants inevitably introduce additional overhead, resulting in suboptimal speed. EfficientViT [29] and SHViT [58] conduct a systematic analysis of redundancy in MHSA and propose a set of memory-efficient design principles. Their models strike a good trade-off between speed and accuracy at the cost of lower parameter efficiency. At the macro design level, hybrid architectures [16, 18, 60] that employ lightweight convolutions in early stages and transformers in later stages have become increasingly popular. This design introduces locality to ViTs and avoids using MHSA in high-resolution early stages to reduce computational costs. + +# 3. Method + +Given the quadratic computational complexity of the self-attention mechanism, prior research has predominantly focused on designing efficient ViTs from a computational cost perspective. While these efforts have yielded promising results, current designs overlook several important factors, resulting in suboptimal model performance and inference + +speed. This section first motivates the need for improved designs by analyzing the limitations of existing approaches from both micro and macro perspectives. Subsequently, we introduce our optimized designs, aiming for a better tradeoff between model size, accuracy, and speed. + +# 3.1. Motivation + +Microscopic Perspective: Recent studies have proposed various attention variants to optimize computational efficiency. However, most designs primarily focus on theoretical FLOPs reduction without considering practical constraints. Fig. 2 presents a latency breakdown in Swin-S [30] (window-based attention), revealing substantial overhead from memory-inefficient operations. Moreover, some attention variants[1, 8] often require custom CUDA kernel implementations and specialized optimization techniques, increasing development complexity and limit deployment flexibility across heterogeneous computing platforms. The absence of hardware-aware design creates a discrepancy between theoretical efficiency metrics and actual runtime performance, particularly in resource-constrained environments. Therefore, a simplified yet effective attention module with spatial feature learning capabilities is crucial for deploying ViTs in TinyML applications. + +Macroscopic Perspective: From the perspective of macroscopic network architecture, extensive research [16, 25, 58] has employed either embedding layers with aggressive downsampling factors or hierarchical downsampling modules to reduce spatial resolution, subsequently cascading transformer blocks at the latter stages where the resolution has already shrunk enough. It is well known that MHSA possesses a robust capability for extracting global spatial information. However, through visualizing the feature maps, it has been observed that features exhibit more pronounced semantic characteristics in the last stage of the network, rather than spatial characteristics. To investigate the potential impact of this architectural mismatch on model performance, this study conducts preliminary experiments on models such as MobileViT[34] and SHViT[58], replacing the attention modules in the last stage of the network with Squeeze-and-Excitation (SE) [23] modules. As shown in Tab. 1, this strategic adjustment yields accuracy improvements on the ImageNet-1K benchmark. + +Table 1. Replacing the last stage module with SE, † denotes our reproduction. + +
ModelParams (M)Top-1 (%)Flops (M)
MobileViT-XXS[34]†1.370.0418
MobileViT-XXS[34] with SE1.370.8411
SHViT-S2[58]†11.474.9366
SHViT-S2[58] with SE11.475.3352
+ +# 3.2. Lean Single-Head Self-Attention + +The MHSA module is a key component of the transformer, designed to effectively capture global spatial information from input tokens. In particular, for token embeddings $\mathbf{X} \in \mathbb{R}^{n \times d}$ , the MHSA is computed as follows: + +$$ +\mathbf {Q} = \mathbf {X} \mathbf {W} _ {q}, \quad [ \mathbf {Q} _ {1}, \mathbf {Q} _ {2}, \dots , \mathbf {Q} _ {h} ] = \mathbf {Q} +$$ + +$$ +\mathbf {K} = \mathbf {X} \mathbf {W} _ {k}, \quad \left[ \mathbf {K} _ {1}, \mathbf {K} _ {2}, \dots , \mathbf {K} _ {h} \right] = \mathbf {K} \tag {1} +$$ + +$$ +\mathbf {V} = \mathbf {X} \mathbf {W} _ {v}, \quad [ \mathbf {V} _ {1}, \mathbf {V} _ {2}, \dots , \mathbf {V} _ {h} ] = \mathbf {V} +$$ + +$$ +\mathbf {A} _ {i} = \operatorname {S o f t m a x} \left[ \frac {\mathbf {Q} _ {i} \mathbf {K} _ {i} ^ {T}}{\sqrt {d _ {h}}} \right] \cdot \mathbf {V} _ {i} \tag {2} +$$ + +$$ +\hat {\mathbf {X}} = \operatorname {C o n c a t} \left[ \mathbf {A} _ {1}, \mathbf {A} _ {2}, \dots , \mathbf {A} _ {h} \right] \cdot \mathbf {W} _ {o} \tag {3} +$$ + +Here $\mathbf{W}_q\in \mathbb{R}^{d\times d_q}$ , $\mathbf{W}_k\in \mathbb{R}^{d\times d_k}$ , $\mathbf{W}_v\in \mathbb{R}^{d\times d_v}$ denote the projection matrices for queries, keys, and values, respectively. Generally, it holds that $d_{q} = d_{k} = d_{v} = d_{h}\cdot h$ , where $d_h$ represents the dimension of each head and $h$ denotes the number of heads. + +From the above equations, it is evident that the MHSA module involves multiple linear transformations and matrix multiplications, contributing significantly to inference latency. In addition to dense matrix multiplication operations, the MHSA module also encompasses memory inefficient operations such as reshape and softmax. These operations significantly impact memory usage and access patterns during computation, thereby affecting the actual inference speed of the model. + +When $h$ equals one, MHSA degenerates into Single-Head Self-Attention (SHSA). In this simplified configuration, the computational process eliminates the need for reshape operations that would otherwise partition query, key, and value matrices into multiple sub-matrices. Furthermore, since both the memory requirements for storing attention maps and the computational complexity of softmax operations scale linearly with the number of attention heads, the reduction to single-head attention significantly decreases memory consumption and latency associated with memory-inefficient operations. + +$$ +\mathbf {S H S A} (X) = \operatorname {S o f t m a x} \left[ \frac {\left(\mathbf {X W} _ {q}\right) \left(\mathbf {X W} _ {k}\right) ^ {T}}{\sqrt {d}} \right] \cdot \left(\mathbf {X W} _ {v}\right) \mathbf {W} _ {o} \tag {4} +$$ + +The computational complexity of matrix multiplication in SHSA is $O(n^{2} \cdot d_{h})$ . To reduce computational costs, Non-Local Networks [50] employ a bottleneck strategy by setting $d_{h}$ to $\frac{1}{2}d$ , while SHViT [58] adopts a sampling approach by only taking a small portion of $X$ as input. Both approaches effectively reduce the head dimension, thereby decreasing the computational overhead. However, while reducing the head dimension can alleviate computational burden, it may compromise the model's ability to capture complex relationships and features. As + +![](images/f68d0d734f15a6a1d94e27208de58449aa4c8ad9b24fde96fbbd79f6d921c0fa.jpg) +a) Multi-Head Self-Attention + +![](images/a6b4462abc4a569d245303ac2926e86dbb27fd4a430e5afbc7a621c0df4f3e7d.jpg) +b) Single-Head Self-Attention + +![](images/0874ad63621494d8295fbefb3512ab83aab6b6b499d253c0e3b14847bab08e8c.jpg) +c) Lean Single-Head Self-Attention + +![](images/8ebf034e1e0577f63cbb9420ffe4d55fdab333a1e53ca294af6cf8a8bbbc07a6.jpg) +Figure 4. Comparison with different attention modules. +MatMul +Reshape +Figure 5. Runtime breakdown comparison of different attention modules. The results are measured on an NVIDIA RTX 3090 and an AMD EPYC 7542 32-Core Processor with configurations $b = 128$ (batch size), $n = 196$ (number of tokens), $d = 192$ (token dimensionality), and $h = 3$ (number of heads for MHSA), using PyTorch Profiler [39]. + +![](images/b1578acb4f00aa6e0d23e966b1af14f07758e2c70bfb97af660f59d66ee9acfd.jpg) +Elementwise Mul +Softmax + +the dimension of each attention head decreases, the feature space it can represent becomes increasingly constrained, ultimately limiting the model's expressive capacity. Through mathematical analysis, we observe that $\mathbf{Q}\mathbf{K}^{\mathrm{T}} = (\mathbf{X}\mathbf{W}_{q})(\mathbf{X}\mathbf{W}_{k})^{T} = \mathbf{X}(\mathbf{W}_{q}\mathbf{W}_{k}^{T})\mathbf{X}^{T}$ . The absence of non-linearity between $Q$ and $K$ imposes a rank constraint: $\mathrm{rank}(\mathbf{Q}\mathbf{K}) \leq \mathrm{rank}(\mathbf{W}_{q}\mathbf{W}_{k}^{T}) \leq d$ . This constraint reveals a critical trade-off. When $d_{h} < d$ , the model's expressiveness suffers from rank deficiency. Conversely, $d_{h} > d$ introduces redundant parameters without theoretical benefits. + +$$ +\mathbf {L e a n S H S A} (X) = \operatorname {S o f t m a x} \left[ \frac {\mathbf {X W} _ {1} \mathbf {X} ^ {T}}{\sqrt {d}} \right] \cdot \mathbf {X W} _ {2} \quad (5) +$$ + +To optimize this trade-off, we propose Lean Single-Head Self-Attention(LeanSHSA), as illustrated in Eq. 5. This approach sets $d_h = d$ and merges projection matrices: $\mathbf{W}_q\mathbf{W}_k^T$ into $\mathbf{W}_1$ and $\mathbf{W}_v\mathbf{W}_o$ into $\mathbf{W}_2$ . By reducing four weight matrices to two, LeanSHSA signifi + +![](images/564e3e9be8a38feca0db4661caf6501a4cefbc5144625824c7ec05c8aecddd52.jpg) +b) ConvBlock +c) SpatialBlock +d) ChannelBlock +Figure 6. Overview of TinyNeXt: A hierarchical architecture featuring a convolutional stem, four progressively designed stages with specific modules (ConvBlock, SpatialBlock, ChannelBlock), and a final classification head. $H$ and $W$ represent input image spatial dimensions, with $L_{i}$ indicating the block count per stage. + +cantly simplifies the attention mechanism while maintaining its theoretical expressiveness. Furthermore, the simplified matrix operations align better with modern hardware accelerators, enabling more efficient parallel processing and better utilization of computing resources. Through these simplifications, LeanSHSA effectively reduces the latency caused by memory-inefficient operations, achieving a near $2 \times$ improvement in CPU throughput compared to MHSA, as demonstrated in Fig. 5. + +# 3.3. Architecture + +Based on our previous analysis and findings, we systematically optimize ViT architectures by adapting module designs to the evolving feature characteristics across network stages. Fig. 6 depicts an overview of the proposed TinyNeXt architecture. TinyNeXt follows the pyramid architecture, comprising a sequential arrangement of a stem, four stages of varying resolutions, and a classification head, with downsampling layers connecting the successive stages. + +The stem layer serves as the initial feature extraction + +module, processing the input image and producing a feature representation. Similar to [25, 47], we stack two standard $3 \times 3$ convolutional layers with a stride of 2 as the stem, downsampling the input image by a factor of 4. + +In the early network stages (Stage 1 and Stage 2), characterized by localized feature representations, we employ the ConvBlock, which originates from the inverted residual block [41]. This choice is driven by the need to efficiently capture local spatial information, such as texture patterns and edge details, which are predominantly concentrated in the initial layers of the network. By leveraging depth-wise separable convolutions and an efficient expansion-compression structure, it effectively extracts fine-grained details and patterns from the input data while maintaining computational efficiency. + +In the middle stage (Stage 3), after multiple downsampling operations, the feature scales become more moderate, and the low-frequency spatial information is retained. In this case, we have employed the SpatialBlock to effectively extract the long-range dependencies between features. The + +Table 2. Classification performance on ImageNet-1K. Throughput is measured on an Nvidia RTX 3090 GPU with batch size of 512 for GPU and AMD EPYC 7542 32-Core Processor with batch size of 16 for CPU. Larger throughput means faster inference speed. FLOP count is computed by fvcore library [14]. + +
ModelInputParams (M)FLOPs (M)Throughput (img/s)Top-1 (%)Top-5 (%)
GPUCPU
ShuffleNetv2-1.0 [32]2242.31461208510469.488.9
PVTv2-b0 [49]2243.757225862670.5-
MobileNetv1 [22]2244.257953104570.6-
EfficientViT-M2 [43]2244.22011668814170.890.2
MobileOne-S0 [46]2242.127584107971.4-
TinyNeXt-T2241.025972578571.590.3
MobileNetv2-1.0 [41]2243.530045723872.091.0
MobileViTv2-0.5 [35]2561.446644352270.2-
MobileViT-XXS [34]2561.341033742369.0-
EdgeNeXt-XXS [33]2561.326146796471.2-
EMO-1M [60]2241.326139494071.590.4
TinyNeXt-S2241.330465557572.791.0
MobileNetv2-1.4 [41]2246.158530911974.7-
MobileNetv3-L-1.0 [21]2245.421958646175.2-
MobileOne-S1 [46]2244.882545543175.9-
EdgeViT-XXS [38]2244.155733263674.4-
MobileViT-XS [34]2562.39862028874.8-
EdgeNeXt-XS [33]2562.353829063275.0-
EMO-2M [60]2242.343927792975.192.2
TinyNeXt-M2242.347548895175.392.3
+ +SpatialBlock is mainly composed of three components: a LeanSHSA layer for comprehensive global context modeling, a $3 \times 3$ lightweight Depthwise Convolution (DWConv) layer dedicated to local feature aggregation or conditional position embedding [6, 53], and a Multilayer Perceptron (MLP) for channel feature interaction. Formally, the SpatialBlock can be represented as follows: + +$$ +X _ {1} = \operatorname {L e a n S H S A} \left(\mathrm {B N} \left(X _ {\text {i n}}\right)\right) + X _ {\text {i n}}, +$$ + +$$ +X _ {2} = \operatorname {D W C o n v} \left(X _ {1}\right) + X _ {1}, +$$ + +$$ +X _ {\text {o u t}} = \operatorname {M L P} (\operatorname {B N} (X _ {2})) + X _ {2}. +$$ + +In the last stage (Stage 4), as feature scales further contract and spatial characteristics diminish, semantic representations become increasingly pronounced. Consequently, we deploy the ChannelBlock, which leverages a SE module to dynamically recalibrate channel-wise features. This module, structurally akin to the SpatialBlock but replacing the LeanSHSA layer with an SE module, enables adaptive feature emphasis and enhances the network's discriminative potential. + +Overall, in designing TinyNeXt, we have avoided overly complex operations, focusing instead on selecting the most appropriate basic modules based on the evolving characteristics of feature. This approach ensures a balance between + +computational efficiency and feature representation capability. We build three TinyNext variants with different setting of depth, width and expansion ratio, the detailed model specifications are demonstrated as Tab 3. + +Table 3. Architecture details of TinyNeXt model variants. + +
ModelsTinyNeXt-TTinyNeXt-STinyNeXt-M
Depth[3, 3, 6, 2][3, 3, 8, 3][4, 4, 9, 4]
Width[32, 64, 96, 192][32, 64, 96, 192][32, 64, 128, 256]
Exp.Ratio[2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 1.5]
+ +# 4. Experiments + +# 4.1. Image Classification + +Implementation Details. We conduct image classification experiments on ImageNet-1K [9] using PyTorch 2.3.0 [39] and Timm library 0.5.4 [51]. All models are trained from scratch at a resolution of $224 \times 224$ for 300 epochs, utilizing 8 Nvidia RTX 3090 GPUs. We employ the AdamW optimizer [31], warm-up training with 5 epochs and a cosine annealing learning rate scheduler. The total batch size is set as 2,048 and the initial learning rate is $6 \times 10^{-3}$ with weight decay of $2.5 \times 10^{-2}$ . The data augmentations ap + +plied during training included Random Resized Crop, Random Horizontal Flip, and Color Jitter, without the use of Mixup, CutMix, or Random Erasing. Moreover, we evaluated the model's throughput across desktop GPU and CPU platforms. On the GPU platform, throughput measurements were conducted using an NVIDIA RTX 3090 with a batch size of 512. For CPU throughput evaluation, we employed an AMD EPYC 7542 32-Core Processor, performing measurements with a batch size of 16 in a single-threaded. + +Results. TinyNeXt is systematically benchmarked against SOTA lightweight models across multiple scales, with the comprehensive evaluation results detailed in Tab. 2. TinyNeXt achieves competitive or superior results in terms of accuracy, parameter efficiency, and inference throughput. TinyNeXt-T achieves an impressive $71.5\%$ top-1 accuracy with only 1.0M parameters, demonstrating notable improvements over MobileNetv1 [22] and PVTv2-b0 [48]. Specifically, it achieves an accuracy gain of $0.9\% \uparrow$ and $1.0\% \uparrow$ , respectively, while reducing parameters by $3.2\mathrm{M} \downarrow$ and $2.7\mathrm{M} \downarrow$ , with corresponding GPU throughput enhancements of $1.4\times$ and $2.8\times$ . At the small scale, TinyNeXt-S, with a comparable parameters of 1.3M to MobileViT-XXS, EdgeNeXt-XXS, and EMO-1M [33, 34, 60], significantly outperforms these models in both accuracy and multi-platform throughput. It achieves a $3.7\% \uparrow$ improvement in top-1 accuracy over MobileViT-XXS, with GPU throughput increasing by $1.9\times$ . Compared to the classic lightweight CNN model MobileNetv2 [41], TinyNeXt-S demonstrates remarkable performance, improving $0.7\% \uparrow$ top-1 accuracy and reducing $37\% \downarrow$ parameters. Moreover, TinyNeXt-M achieves the highest top-1 accuracy of $75.3\%$ under the same parameter constraints when compared with MobileViT-XS, EdgeNeXt-XS, and EMO-2M [33, 34, 60], while delivering an exceptional CPU throughput that is $6.4\times$ that of MobileViT-XS. With its excellent balanced accuracy and efficiency, TinyNeXt demonstrates potential as a viable option for resource-constrained applications. + +# 4.2. Downstream Tasks + +Object Detection. We integrate TinyNeXt into the SSDLite framework [21] for object detection on the MS COCO 2017 dataset [28] at $320 \times 320$ resolution. The implementation leverages the MMDetection library [3], employing the AdamW optimizer [31] with a batch size of 64, and inference throughput (images/s) is benchmarked on an NVIDIA RTX 3090 GPU. As shown in Table 4, TinyNeXt demonstrates superior computational efficiency with competitive accuracy. Specifically, TinyNeXt-S achieves $22.4\mathrm{mAP}$ using only $2.3\mathrm{M}$ parameters and 2,517 images/s throughput. It reduces parameters by $2.8 / 2.0\mathrm{M}$ compared to MobileNetv1/v2[22, 41], and surpasses MobileViTv2-0.5/EMO-1M[35, 60] by $1.2 / 0.4\mathrm{mAP}$ with $1.2 / 1.5 \times$ throughput. The medium-scale TinyNeXt-M + +Table 4. Object detection results on MS COCO 2017 with the SSDLite framework. + +
BackboneParams↓Throughput↑mAP↑
MobileNetv1 [22]5.1M228222.2
MobileNetv2 [41]4.3M202922.1
MobileViT-XXS [34]1.7M152819.9
MobileViTv2-0.5[35]2.0M205821.2
EMO-1M[60]2.3M168122.0
TinyNeXt-S2.3M251722.4
MobileViTv2-0.75[35]3.6M138824.6
EMO-2M[60]3.3M119425.2
TinyNeXt-M3.3M191125.0
+ +Table 5. Semantic segmentation results on PASCAL VOC 2012 with the DeepLabv3 framework. + +
BackboneParams↓Flops↓mIOU↑
MobileNetV2 [41]4.5M5.8G75.7
MobileViTv2-0.75[35]9.6M40.0G75.1
TinyNeXt-S3.5M3.5G75.5
TinyNeXt-M4.9M5.1G76.9
+ +Table 6. Performance Comparison on Edge Devices. The last two columns present inference latencies measured on edge computing platforms: NVIDIA Jetson NANO and ARM Cortex-A57. + +
ModelInputParams (M)Top1 (%)NANO (ms)Cortex-A57 (ms)
MobileViT-XXS [34]25621.369.030.1467.9
EdgeNeXt-XXS [33]25621.371.233.5242.8
EMO-1M [60]22421.371.526.2526.4
TinyNeXt-S22421.372.714.3227.7
MobileViT-XS [34]25622.374.750.8859.1
EdgeNeXt-XS [33]25622.375.050.8442.4
EMO-2M [60]22422.375.135.7759.2
TinyNeXt-M22422.375.319.4311.6
+ +attains $25.0\mathrm{mAP}$ with 3.3M parameters, matching EMO-2M in accuracy while providing $1.6\times$ faster inference. + +Semantic Segmentation. We integrate TinyNeXt into the DeepLabv3 framework [4] for semantic segmentation on the Pascal VOC 2012 dataset [13] at $512 \times 512$ resolution. Our implementation utilizes the MMSegmentation library [7], trains with the AdamW optimizer [31] for 40K iterations using a total batch size of 16. As summarized in Table 5, both variants provide superior accuracy and efficiency. Specifically, TinyNeXt-M surpasses MobileNetV2 by $1.2\%$ mIOU with 0.7G fewer FLOPs, further demonstrating its effectiveness as a vision backbone across tasks. + +# 4.3. Inference on Edge Devices + +Implementation Details. To validate our model's potential for TinyML applications, we conduct a comprehensive evaluation of TinyNeXt's inference performance against SOTA models, including MobileViT [34], EdgeNeXt [33], and EMO [60], across diverse edge computing platforms. Specifically, we evaluate their inference capabilities on edge GPU and CPU architectures, utilizing platforms such as the NVIDIA Jetson Nano and ARM Cortex-A57. For each platform, we implement a unified testing flow: an initial 10s warm-up phase followed by 60s of continuous inference to accurately measure average latency. + +Results. TinyNeXt achieves a superior accuracy-latency trade-off across all evaluated edge platforms. With 1.3M parameters, TinyNeXt-S demonstrates $1.8 \times$ faster inference on NVIDIA Jetson Nano while maintaining $1.2\%$ higher top-1 accuracy, and reduces ARM Cortex-A57 latency by $56.7\%$ compared to EMO-1M [60]. At 2.3M parameters, TinyNeXt-M further keeps these advantages, achieving $1.8 \times$ speedup over EMO-2M [60] on edge GPU with $0.2\%$ accuracy improvement, while delivering $2.8 \times$ faster CPU execution than MobileViT-XS [34]. + +# 4.4. Ablation Study + +In this section, we conduct ablation studies on key design elements of the proposed TinyNeXt using the ImageNet-1K dataset [9]. To highlight architectural differences and reduce training time, all models are trained for 100 epochs [16, 29]. Throughout the ablation experiments, we consistently preserve the baseline architecture's original depth and width configurations. This approach facilitates a precise evaluation of individual design choices' performance impact. + +Impact of the attention block. These ablation experiments validate LeanSHSA's efficiency through substitutions with standard MHSA and SHSA modules. As shown in Tab. 7 (#2 and #3), replacing LeanSHSA with 4-head MHSA increases parameters by $0.3\mathrm{M}$ for merely $0.2\%$ accuracy gain, while incurring $19\%$ throughput penalty. This empirically demonstrates that LeanSHSA's streamlined computational design effectively enhances inference efficiency without compromising global feature extraction capabilities. Furthermore, adopting a bottlenecked SHSA approach marginally reduces accuracy by $0.2\%$ despite equivalent parameter count, with only a $5\%$ speed gain. This suggests the importance of maintaining head dimensionality for better feature extraction in single-head architecture. + +Impact of the last stage choice. This ablation experiment evaluates the effectiveness of TinyNeXt's macro architecture by modifying the feature extraction module in the last stage. As shown in Tab. 7 (#4, #5, #6), we replace the channel-wise attention module (SE) with various global spatial attention modules, including MHSA (with 4 heads + +Table 7. Ablation Study of TinyNeXt-M on ImageNet-1K [9]. The last three columns present model parameters (M), top-1 accuracy (\%), and GPU throughput (images/s). + +
#AblationParamsTop-1Throughput
1TinyNeXt-M (baseline)2.374.14889
2LeanSHSA → MHSA2.674.33950
3LeanSHSA → SHSA2.373.95137
4Last Stage with MHSA2.774.14718
5Last Stage with SHSA2.773.64788
6Last Stage with LeanSHSA2.773.94743
7BN → LN2.373.84479
+ +and a head dimension of 32), SHSA, and LeanSHSA. These substitutions increase the model size by 0.4M parameters compared to the baseline, but yield no accuracy gains and even slightly reduce throughput. For example, using MHSA in the last stage (#4) raises parameters from 2.3M to 2.7M and cuts throughput by nearly 200 images/s, while the top-1 accuracy remains unchanged at $74.1\%$ . Similar trends are observed with SHSA and LeanSHSA variants. These results demonstrate that channel-wise attention is more effective in deep layers, where feature representations shift from being spatially dominant to semantically dominant, offering superior semantic discrimination. + +Impact of the normalization layer. We investigate the impact of normalization layers by replacing Batch Normalization (BN) with Layer Normalization (LN) in SpatialBlock and ChannelBlock. LN requires dynamic mean and variance computation during inference, which poses challenges for edge device deployment. Empirical evaluation (#7) reveals that replacing BN with LN results in a $8\%$ reduction in inference throughput and a marginal $0.3\%$ decline in top-1 accuracy. Consequently, we retain BN as the preferred normalization strategy in our TinyNeXt architecture. + +# 5. Conclusion + +This paper analyzes the limitations of current efficient ViTs at both micro and macro levels and proposes corresponding optimizations. Specifically, we introduce a simplified and efficient attention variant, LeanSHSA, and enhance model performance through stage-specific macro designs. Extensive experiments demonstrate TinyNeXt's exceptional trade-off of model size, accuracy, and speed across diverse vision tasks and hardware platforms. + +Limitations. The computational complexity of Lean-SHSA remains quadratic with visual tokens, leading to high costs in high-resolution tasks and restricting its applicability in such scenarios. In future work, we aim to explore dynamic attention mechanisms while minimizing the overhead of memory-bound operations. + +# Acknowledgments + +This work was supported in part by the National Science and Technology Innovation 2030-Major Projects under Grant 2021ZD0114400, in part by the Young Scientists Fund of the National Natural Science Foundation of China No.62304162, in part by the China Postdoctoral Science Foundation under Grant Number 2024M762532, in part by the Postdoctoral Fellowship Program of CPSF under Grant GZC20241313, in part by Shaanxi Provincial Natural Science Foundation for Basic Research Program under Grant 2024JC-YBMS-794, and in part by the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24090. + +# References + +[1] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV). Springer, 2020. 1 +[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7 +[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1 +[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 6 +[7] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7 +[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 3 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 8 +[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1, 3 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3 +[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 3 +[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7 +[14] facebookresearch. fvcore library. https://github.com/facebookresearch/fvcore, 2019.6 +[15] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision, pages 396-414. Springer, 2022. 1 +[16] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 2, 3, 8 +[17] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 3 +[18] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189, 2023. 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3 +[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3 +[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 6, 7 +[22] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3, 6, 7 +[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3 +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3 + +[25] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 1, 3, 5 +[26] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16889-16900, 2023. 1 +[27] Yishi Li, Yuhao Zhang, and Rui Lai. Tinypillarnet: Tiny pillar-based network for 3d point cloud object detection at edge. IEEE Transactions on Circuits and Systems for Video Technology, 34(3):1772-1785, 2023. 3 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7 +[29] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3, 8 +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3 +[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 7 +[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6 +[33] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. In European conference on computer vision, pages 3-20. Springer, 2022. 6, 7, 8 +[34] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 3, 6, 7, 8 +[35] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 2, 6, 7 +[36] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 3 +[37] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 3 + +[38] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, pages 294-311. Springer, 2022. 6 +[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 2, 4, 6 +[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1 +[41] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 5, 6, 7 +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3 +[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1, 3, 6 +[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 3 +[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 1, 3 +[46] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7907-7917, 2023. 6 +[47] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15909-15920, 2024. 5 +[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2, 3, 7 +[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2, 3, 6 + +[50] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2, 4 +[51] Ross Wightman et al. Pytorch image models, 2019. 6 +[52] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European conference on computer vision, pages 68-85. Springer, 2022. 3 +[53] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 1, 6 +[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 12077-12090, 2021. 1 +[55] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3 +[56] Kunran Xu, Yishi Li, Huawei Zhang, Rui Lai, and Lin Gu. Etinynet: Extremely tiny network for tinyml. In Proceedings of the AAAI conference on artificial intelligence, pages 4628-4636, 2022. 3 +[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1 +[58] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5756-5767, 2024. 2, 3, 4 +[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jianming Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2022. 1 +[60] Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, and Chengjie Wang. Rethinking mobile block for efficient attention-based models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1389-1400. IEEE Computer Society, 2023. 1, 2, 3, 6, 7, 8 +[61] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6881-6890, 2021. 1 +[62] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 3 + +[63] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 1, 3 +[64] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018. 3 \ No newline at end of file diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/images.zip b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..554df8e4246b1fe71168e0cb086a52a3600e6939 --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95bbdc277cf0456812953fba608b35960612056013e9631144cd91ea8f9f556f +size 625308 diff --git a/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/layout.json b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25703e3cd4b1a4959bd787c6082b47f998b9b872 --- /dev/null +++ b/2025/An Efficient Hybrid Vision Transformer for TinyML Applications/layout.json @@ -0,0 +1,9053 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 509, + 121 + ], + "type": "text", + "content": "An Efficient Hybrid Vision Transformer for TinyML Applications" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "spans": [ + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "text", + "content": "Fanhong Zeng" + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "text", + "content": " Huanan Li" + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "text", + "content": " Juntao Guan" + }, + { + "bbox": [ + 182, + 142, + 431, + 157 + ], + "type": "inline_equation", + "content": "^{1,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "spans": [ + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "text", + "content": "Rui Fan" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "text", + "content": " Tong Wu" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "text", + "content": " Xilong Wang" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "text", + "content": " Rui Lai" + }, + { + "bbox": [ + 176, + 157, + 451, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 136, + 172, + 474, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 172, + 474, + 186 + ], + "spans": [ + { + "bbox": [ + 136, + 172, + 474, + 186 + ], + "type": "text", + "content": "1 Key Laboratory of Analog Integrated Circuits and Systems (Ministry of Education)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 179, + 186, + 431, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 186, + 431, + 199 + ], + "spans": [ + { + "bbox": [ + 179, + 186, + 431, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 179, + 186, + 431, + 199 + ], + "type": "text", + "content": "Faculty of Integrated Circuit, Xidian University, Xi'an, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 158, + 200, + 452, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 158, + 200, + 452, + 213 + ], + "spans": [ + { + "bbox": [ + 158, + 200, + 452, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 158, + 200, + 452, + 213 + ], + "type": "text", + "content": "Hangzhou Institute of Technology, Xidian University, Hangzhou, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "text", + "content": "To enable the deployment of Vision Transformers on resource-constrained mobile and edge devices, the development of efficient ViT models has attracted significant attention. Researchers achieved remarkable improvements in accuracy and speed by optimizing attention mechanisms and integrating lightweight CNN modules. However, existing designs often overlook runtime overhead from memory-bound operations and the shift in feature characteristics from spatial-dominant to semantic-dominant as networks deepen. This work introduces TinyNeXt, a family of efficient hybrid ViTs for TinyML, featuring Lean Single-Head Self-Attention to minimize memory-bound operations, and a macro design tailored to feature characteristics at different stages. TinyNeXt strikes a better accuraciespeed trade-off across diverse tasks and hardware platforms, outperforming state-of-the-art models of comparable scale. For instance, our TinyNeXt-T achieves a remarkable " + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "text", + "content": " top-1 accuracy with only 1.0M parameters on ImageNet-1K. Furthermore, compared to recent efficient models like MobileViT-XXS and MobileViT-XS, TinyNeXt-S and TinyNeXt-M achieve " + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "inline_equation", + "content": "3.7\\% /0.5\\%" + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "text", + "content": " higher accuracy, respectively, while running " + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "inline_equation", + "content": "2.1\\times /2.6\\times" + }, + { + "bbox": [ + 54, + 266, + 297, + 555 + ], + "type": "text", + "content": " faster on Nvidia Jetson Nano. The code and trained models are available at https://github.com/yuffeenn/TinyNeXt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "content": "Recently, Vision Transformers (ViTs) [11] have demonstrated superior performance across diverse computer vision tasks, including classification[44], detection [2, 59], and segmentation [5, 54, 61]. As a key component of ViT, Multi-Head Self-Attention (MHSA) captures interrelations between any two spatial locations, significantly enhancing global information comprehension. Nonetheless, the quadratic complexity and substantial memory demands" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 244, + 552, + 417 + ], + "blocks": [ + { + "bbox": [ + 318, + 244, + 552, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 244, + 552, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 244, + 552, + 417 + ], + "type": "image", + "image_path": "9c0a96ddb012f6c180bb16cf2288e59f7ab4148706435201f44a3af723c9f299.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 427, + 527, + 439 + ], + "lines": [ + { + "bbox": [ + 340, + 427, + 527, + 439 + ], + "spans": [ + { + "bbox": [ + 340, + 427, + 527, + 439 + ], + "type": "text", + "content": "Figure 1. Throughput vs Accuracy Comparison." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 485 + ], + "type": "text", + "content": "of MHSA pose significant challenges for deploying ViTs on edge devices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 486, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 555, + 653 + ], + "type": "text", + "content": "To tackle the above challenges, researchers have introduced a range of lightweight design strategies. One approach integrates Convolutional Neural Networks (CNNs) with ViTs, such as employing lightweight convolutions in early stages [25, 26, 53] to reduce computations. Another approach focuses on optimizing self-attention mechanisms to reduce the number of visual tokens, thereby decreasing computational costs. This includes restricting attention to specific regions [10, 30, 45, 63] and dynamically pruning redundant tokens [15, 40, 57]. Leveraging CNN-Transformer hybrids and innovative attention modules, current lightweight ViT [26, 60] models have progressively surpassed lightweight CNN models[41, 43] in terms of accuracy and speed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "However, prior works have overlooked several critical aspects in the design of efficient ViTs, both at micro and macro architectural levels. At the micro level, dense matrix multiplication operations are commonly regarded as the throughput bottleneck in attention module inference. Many" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19914" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 283, + 198 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 283, + 198 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 283, + 198 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 283, + 198 + ], + "type": "image", + "image_path": "59b607e1be7661a3ce6b8d7944e7f2234a49d23a4b5b810bc57cb8be29280d03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 297, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 297, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 297, + 275 + ], + "type": "text", + "content": "Figure 2. Runtime Profiling of Different Vision Transformers. Beyond matrix multiplication, both DeiT and Swin models involve costly memory-bound operations, such as Reshape, Copy, and LayerNorm. Notably, in the Swin-S model, the latency of these memory-bound operations exceeds " + }, + { + "bbox": [ + 55, + 206, + 297, + 275 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 206, + 297, + 275 + ], + "type": "text", + "content": ". PyTorch Profiler [39] was used as the latency analysis tool." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 285, + 284, + 421 + ], + "blocks": [ + { + "bbox": [ + 67, + 285, + 284, + 421 + ], + "lines": [ + { + "bbox": [ + 67, + 285, + 284, + 421 + ], + "spans": [ + { + "bbox": [ + 67, + 285, + 284, + 421 + ], + "type": "image", + "image_path": "e1fd852072007a3b4556abfb7151b41d1c323c6fa9264b5146dc969f09116131.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 452 + ], + "type": "text", + "content": "Figure 3. Feature maps across stages of PVT [48], showing the transition from spatial to channel dominance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 568 + ], + "type": "text", + "content": "attention variants aim to reduce theoretical FLOPs but inadvertently introduce significant runtime overhead from noncomputational operations. As illustrated in Fig. 2, a substantial portion of Swin-S's [30] runtime is consumed by memory-bound operations, including tensor reshaping, data copying, and layer normalization. This observation highlights the need for hardware-friendly attention mechanisms to achieve efficient inference." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "At the macro design level, driven by computational cost considerations, existing efficient ViTs [16, 34, 58] often adopt a hybrid architecture: using embedding layers or hierarchical downsampling convolutional layers to reduce spatial resolution, followed by stacking transformer blocks in the later stages. While this design achieves a better tradeoff between accuracy and latency, it may overlook the evolution of feature representations across different network depths. Our preliminary investigation reveals an intriguing phenomenon: as the network deepens, feature characteristics gradually shift from being spatial-dominant to semantic-dominant. This phenomenon is clearly illustrated" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": "in Fig 3, which visualizes the feature maps at different stages of the Pyramid Vision Transformer [49] architecture. Notably, when feature maps are downsampled to a " + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 72, + 553, + 143 + ], + "type": "text", + "content": " resolution in the last stage, spatial information diminishes significantly, indicating a fundamental shift in feature representation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "content": "This observation aligns with previous findings in the Non-Local Networks [50], where researchers systematically investigated the effect of adding non-local blocks at different stages of ResNet [19] architectures. Their results showed that incorporating non-local blocks in the last stage (res5), where spatial dimensions are reduced to " + }, + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 144, + 555, + 251 + ], + "type": "text", + "content": ", yielded minimal performance improvements. These findings suggest that the common practice of placing spatial attention blocks in the last stage may not be optimal." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": "Based on these findings, we propose TinyNeXt, a family of efficient hybrid Vision Transformers for Tiny Machine Learning (TinyML) applications. In micro design, we introduce Lean Single-Head Self-Attention (LeanSHSA), which employs a single-head architecture to minimize latency from reshape and softmax operations, while merging projection matrices to form a more streamlined attention mechanism. Additionally, we refine the macro design of ViT architectures by tailoring module to feature characteristics at different stages. Experiments demonstrate that our models achieve superior parameter efficiency, latency, and accuracy compared to state-of-the-art (SOTA) lightweight models, while avoiding complex designs. Specifically, TinyNeXt-T achieves " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " top-1 accuracy on ImageNet-1K with only 1.0M parameters. Furthermore, TinyNeXt-S/M attains " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "72.7\\% / 75.3\\%" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " top-1 accuracy with " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "1.3 / 2.3\\mathrm{M}" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " parameters and " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "14.3 / 19.4~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " latency on Nvidia Jetson Nano, outperforming EMO-1M/2M [60] by " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "+1.2\\uparrow / + 0.2\\uparrow" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " with a speedup of " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "1.83\\times / 1.84\\times" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": ". When integrated with SSDlite framework for object detection, TinyNeXt-S/M achieves " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "22.4 / 25.0\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "2.3 / 3.3\\mathrm{M}" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " parameters, surpassing MobileViTv2 [35] by " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "+1.2\\uparrow / + 0.4\\uparrow" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " with a speedup of " + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "inline_equation", + "content": "1.2\\times / 1.4\\times" + }, + { + "bbox": [ + 313, + 251, + 556, + 525 + ], + "type": "text", + "content": " on Nvidia RTX 3090. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 527, + 553, + 658 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 527, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 553, + 574 + ], + "type": "text", + "content": "- We propose Lean Single Head Self-Attention, a simplified and efficient attention mechanism that minimizes memory-inefficient operations and runs faster than traditional MHSA." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 553, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 553, + 622 + ], + "type": "text", + "content": "- We refine the macro design of lightweight ViTs by introducing stage-specific feature extraction modules, replacing spatial attention with channel-wise attention in the last stage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 623, + 553, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 553, + 658 + ], + "type": "text", + "content": "- We conduct comprehensive experiments demonstrating that TinyNeXt achieves a superior accuracy-speed trade-off across diverse vision tasks and hardware platforms." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 669, + 402, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 402, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 402, + 681 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Over the past decade, CNNs have significantly influenced computer vision, becoming the basis for many tasks. Mod-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19915" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "els like AlexNet [24], VGGNet [42], ResNet [19], and SENet [23] have rapidly advanced, substantially enhancing accuracy across various vision tasks. But their substantial parameter and computational costs impede deployment on mobile and edge devices. To address these challenges, previous work has proposed numerous lightweight techniques, including efficient network design [22, 27, 41, 56], network architecture search [43, 64], low-bit quantization [12, 37, 62], knowledge distillation [20, 55], and network pruning [17, 36]. However, CNNs are spatially local and lack global interaction between the features, limiting further performance improvements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 323 + ], + "type": "text", + "content": "Recent years have witnessed the emergence of Transformer-based models outperforming CNN-based models in the computer vision domain. Dosovitskiy et al. [11] first introduced the transformer architecture to vision tasks and ViT's excellent performance has inspired numerous subsequent improvements. Subsequent research has reduced ViT's limitations through techniques such as distillation token [44, 52], local attention mechanism [30], and hierarchical architecture [48, 49]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 324, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 295, + 610 + ], + "type": "text", + "content": "Currently, to deploy ViTs on resource-constrained devices, efficient ViT designs are rapidly evolving. At the micro design level, in addition to embedding depth-wise convolutions in attention modules, numerous works have proposed various attention variants to reduce computational complexity. Swin Transformer [30] and follow-up works [10, 45, 63] limit self-attention computation to crafted local windows, interacting window information through additional operations. MobileViT [34] reduces computational complexity by decreasing the number of visual tokens in self-attention via unfold operation. Unfortunately, as MHSA inherently involves numerous memory-inefficient operations, these attention variants inevitably introduce additional overhead, resulting in suboptimal speed. EfficientViT [29] and SHViT [58] conduct a systematic analysis of redundancy in MHSA and propose a set of memory-efficient design principles. Their models strike a good trade-off between speed and accuracy at the cost of lower parameter efficiency. At the macro design level, hybrid architectures [16, 18, 60] that employ lightweight convolutions in early stages and transformers in later stages have become increasingly popular. This design introduces locality to ViTs and avoids using MHSA in high-resolution early stages to reduce computational costs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 111, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "Given the quadratic computational complexity of the self-attention mechanism, prior research has predominantly focused on designing efficient ViTs from a computational cost perspective. While these efforts have yielded promising results, current designs overlook several important factors, resulting in suboptimal model performance and inference" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "speed. This section first motivates the need for improved designs by analyzing the limitations of existing approaches from both micro and macro perspectives. Subsequently, we introduce our optimized designs, aiming for a better tradeoff between model size, accuracy, and speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 142, + 388, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 388, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 388, + 153 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 160, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 553, + 363 + ], + "type": "text", + "content": "Microscopic Perspective: Recent studies have proposed various attention variants to optimize computational efficiency. However, most designs primarily focus on theoretical FLOPs reduction without considering practical constraints. Fig. 2 presents a latency breakdown in Swin-S [30] (window-based attention), revealing substantial overhead from memory-inefficient operations. Moreover, some attention variants[1, 8] often require custom CUDA kernel implementations and specialized optimization techniques, increasing development complexity and limit deployment flexibility across heterogeneous computing platforms. The absence of hardware-aware design creates a discrepancy between theoretical efficiency metrics and actual runtime performance, particularly in resource-constrained environments. Therefore, a simplified yet effective attention module with spatial feature learning capabilities is crucial for deploying ViTs in TinyML applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 364, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 553, + 591 + ], + "type": "text", + "content": "Macroscopic Perspective: From the perspective of macroscopic network architecture, extensive research [16, 25, 58] has employed either embedding layers with aggressive downsampling factors or hierarchical downsampling modules to reduce spatial resolution, subsequently cascading transformer blocks at the latter stages where the resolution has already shrunk enough. It is well known that MHSA possesses a robust capability for extracting global spatial information. However, through visualizing the feature maps, it has been observed that features exhibit more pronounced semantic characteristics in the last stage of the network, rather than spatial characteristics. To investigate the potential impact of this architectural mismatch on model performance, this study conducts preliminary experiments on models such as MobileViT[34] and SHViT[58], replacing the attention modules in the last stage of the network with Squeeze-and-Excitation (SE) [23] modules. As shown in Tab. 1, this strategic adjustment yields accuracy improvements on the ImageNet-1K benchmark." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 635, + 556, + 697 + ], + "blocks": [ + { + "bbox": [ + 313, + 603, + 553, + 625 + ], + "lines": [ + { + "bbox": [ + 313, + 603, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 553, + 625 + ], + "type": "text", + "content": "Table 1. Replacing the last stage module with SE, † denotes our reproduction." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 635, + 556, + 697 + ], + "lines": [ + { + "bbox": [ + 317, + 635, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 556, + 697 + ], + "type": "table", + "html": "
ModelParams (M)Top-1 (%)Flops (M)
MobileViT-XXS[34]†1.370.0418
MobileViT-XXS[34] with SE1.370.8411
SHViT-S2[58]†11.474.9366
SHViT-S2[58] with SE11.475.3352
", + "image_path": "5096dc8c8eba52bf3ab1c9dea8affcffeea18f1ed46fc98b46019502ecc0962d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19916" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 231, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 231, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 231, + 85 + ], + "type": "text", + "content": "3.2. Lean Single-Head Self-Attention" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": "The MHSA module is a key component of the transformer, designed to effectively capture global spatial information from input tokens. In particular, for token embeddings " + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": ", the MHSA is computed as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 144, + 253, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 144, + 253, + 158 + ], + "spans": [ + { + "bbox": [ + 94, + 144, + 253, + 158 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {X} \\mathbf {W} _ {q}, \\quad [ \\mathbf {Q} _ {1}, \\mathbf {Q} _ {2}, \\dots , \\mathbf {Q} _ {h} ] = \\mathbf {Q}", + "image_path": "aa2cd641d5ca08c3f05af6cb8d3cf9518687234f592f98043e2e23be17afdf48.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 159, + 294, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 159, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 96, + 159, + 294, + 172 + ], + "type": "interline_equation", + "content": "\\mathbf {K} = \\mathbf {X} \\mathbf {W} _ {k}, \\quad \\left[ \\mathbf {K} _ {1}, \\mathbf {K} _ {2}, \\dots , \\mathbf {K} _ {h} \\right] = \\mathbf {K} \\tag {1}", + "image_path": "b9130963b8f351d3a40258dedfcc81a65a0ee7b528cf37f99084b3b38045677e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 174, + 254, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 174, + 254, + 188 + ], + "spans": [ + { + "bbox": [ + 96, + 174, + 254, + 188 + ], + "type": "interline_equation", + "content": "\\mathbf {V} = \\mathbf {X} \\mathbf {W} _ {v}, \\quad [ \\mathbf {V} _ {1}, \\mathbf {V} _ {2}, \\dots , \\mathbf {V} _ {h} ] = \\mathbf {V}", + "image_path": "19890d7d654f05642df7b77ccd7c0a0f271fbdbeb2e9b9b38146cc45e281f757.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 196, + 294, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 196, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 113, + 196, + 294, + 224 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {i} = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {Q} _ {i} \\mathbf {K} _ {i} ^ {T}}{\\sqrt {d _ {h}}} \\right] \\cdot \\mathbf {V} _ {i} \\tag {2}", + "image_path": "22f41b76035f4197683c6b16cbbd6516131ee0a143699c4d18a54b31c53cfecc.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 232, + 294, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 232, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 99, + 232, + 294, + 247 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} = \\operatorname {C o n c a t} \\left[ \\mathbf {A} _ {1}, \\mathbf {A} _ {2}, \\dots , \\mathbf {A} _ {h} \\right] \\cdot \\mathbf {W} _ {o} \\tag {3}", + "image_path": "3a2831af60d52f42abc69900df49334b111c65d58b0a3f8147d7419517a53c13.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q\\in \\mathbb{R}^{d\\times d_q}" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_k\\in \\mathbb{R}^{d\\times d_k}" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v\\in \\mathbb{R}^{d\\times d_v}" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": " denote the projection matrices for queries, keys, and values, respectively. Generally, it holds that " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "d_{q} = d_{k} = d_{v} = d_{h}\\cdot h" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": " represents the dimension of each head and " + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 249, + 296, + 309 + ], + "type": "text", + "content": " denotes the number of heads." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 417 + ], + "type": "text", + "content": "From the above equations, it is evident that the MHSA module involves multiple linear transformations and matrix multiplications, contributing significantly to inference latency. In addition to dense matrix multiplication operations, the MHSA module also encompasses memory inefficient operations such as reshape and softmax. These operations significantly impact memory usage and access patterns during computation, thereby affecting the actual inference speed of the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "text", + "content": " equals one, MHSA degenerates into Single-Head Self-Attention (SHSA). In this simplified configuration, the computational process eliminates the need for reshape operations that would otherwise partition query, key, and value matrices into multiple sub-matrices. Furthermore, since both the memory requirements for storing attention maps and the computational complexity of softmax operations scale linearly with the number of attention heads, the reduction to single-head attention significantly decreases memory consumption and latency associated with memory-inefficient operations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 556, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 594 + ], + "type": "interline_equation", + "content": "\\mathbf {S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\left(\\mathbf {X W} _ {q}\\right) \\left(\\mathbf {X W} _ {k}\\right) ^ {T}}{\\sqrt {d}} \\right] \\cdot \\left(\\mathbf {X W} _ {v}\\right) \\mathbf {W} _ {o} \\tag {4}", + "image_path": "3236d7f04b4f0d203867ced3409120c06040635f32e2dbe170be783feae00b16.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "The computational complexity of matrix multiplication in SHSA is " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "O(n^{2} \\cdot d_{h})" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". To reduce computational costs, Non-Local Networks [50] employ a bottleneck strategy by setting " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}d" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", while SHViT [58] adopts a sampling approach by only taking a small portion of " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " as input. Both approaches effectively reduce the head dimension, thereby decreasing the computational overhead. However, while reducing the head dimension can alleviate computational burden, it may compromise the model's ability to capture complex relationships and features. As" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 329, + 72, + 397, + 152 + ], + "blocks": [ + { + "bbox": [ + 329, + 72, + 397, + 152 + ], + "lines": [ + { + "bbox": [ + 329, + 72, + 397, + 152 + ], + "spans": [ + { + "bbox": [ + 329, + 72, + 397, + 152 + ], + "type": "image", + "image_path": "f68d0d734f15a6a1d94e27208de58449aa4c8ad9b24fde96fbbd79f6d921c0fa.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 154, + 392, + 161 + ], + "lines": [ + { + "bbox": [ + 333, + 154, + 392, + 161 + ], + "spans": [ + { + "bbox": [ + 333, + 154, + 392, + 161 + ], + "type": "text", + "content": "a) Multi-Head Self-Attention" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 400, + 72, + 468, + 152 + ], + "blocks": [ + { + "bbox": [ + 400, + 72, + 468, + 152 + ], + "lines": [ + { + "bbox": [ + 400, + 72, + 468, + 152 + ], + "spans": [ + { + "bbox": [ + 400, + 72, + 468, + 152 + ], + "type": "image", + "image_path": "a6b4462abc4a569d245303ac2926e86dbb27fd4a430e5afbc7a621c0df4f3e7d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 154, + 462, + 161 + ], + "lines": [ + { + "bbox": [ + 402, + 154, + 462, + 161 + ], + "spans": [ + { + "bbox": [ + 402, + 154, + 462, + 161 + ], + "type": "text", + "content": "b) Single-Head Self-Attention" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 540, + 152 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 540, + 152 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 540, + 152 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 540, + 152 + ], + "type": "image", + "image_path": "0874ad63621494d8295fbefb3512ab83aab6b6b499d253c0e3b14847bab08e8c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 154, + 541, + 161 + ], + "lines": [ + { + "bbox": [ + 471, + 154, + 541, + 161 + ], + "spans": [ + { + "bbox": [ + 471, + 154, + 541, + 161 + ], + "type": "text", + "content": "c) Lean Single-Head Self-Attention" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 331, + 196, + 432, + 348 + ], + "blocks": [ + { + "bbox": [ + 333, + 172, + 534, + 183 + ], + "lines": [ + { + "bbox": [ + 333, + 172, + 534, + 183 + ], + "spans": [ + { + "bbox": [ + 333, + 172, + 534, + 183 + ], + "type": "text", + "content": "Figure 4. Comparison with different attention modules." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 196, + 432, + 348 + ], + "lines": [ + { + "bbox": [ + 331, + 196, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 331, + 196, + 432, + 348 + ], + "type": "image", + "image_path": "8ebf034e1e0577f63cbb9420ffe4d55fdab333a1e53ca294af6cf8a8bbbc07a6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 352, + 375, + 359 + ], + "lines": [ + { + "bbox": [ + 342, + 352, + 375, + 359 + ], + "spans": [ + { + "bbox": [ + 342, + 352, + 375, + 359 + ], + "type": "text", + "content": "MatMul" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 352, + 418, + 360 + ], + "lines": [ + { + "bbox": [ + 381, + 352, + 418, + 360 + ], + "spans": [ + { + "bbox": [ + 381, + 352, + 418, + 360 + ], + "type": "text", + "content": "Reshape" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "text", + "content": "Figure 5. Runtime breakdown comparison of different attention modules. The results are measured on an NVIDIA RTX 3090 and an AMD EPYC 7542 32-Core Processor with configurations " + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "inline_equation", + "content": "b = 128" + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "text", + "content": " (batch size), " + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "inline_equation", + "content": "n = 196" + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "text", + "content": " (number of tokens), " + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "inline_equation", + "content": "d = 192" + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "text", + "content": " (token dimensionality), and " + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "inline_equation", + "content": "h = 3" + }, + { + "bbox": [ + 313, + 374, + 555, + 441 + ], + "type": "text", + "content": " (number of heads for MHSA), using PyTorch Profiler [39]." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 434, + 197, + 537, + 348 + ], + "blocks": [ + { + "bbox": [ + 425, + 352, + 481, + 360 + ], + "lines": [ + { + "bbox": [ + 425, + 352, + 481, + 360 + ], + "spans": [ + { + "bbox": [ + 425, + 352, + 481, + 360 + ], + "type": "text", + "content": "Elementwise Mul" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 352, + 523, + 360 + ], + "lines": [ + { + "bbox": [ + 487, + 352, + 523, + 360 + ], + "spans": [ + { + "bbox": [ + 487, + 352, + 523, + 360 + ], + "type": "text", + "content": "Softmax" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 434, + 197, + 537, + 348 + ], + "lines": [ + { + "bbox": [ + 434, + 197, + 537, + 348 + ], + "spans": [ + { + "bbox": [ + 434, + 197, + 537, + 348 + ], + "type": "image", + "image_path": "b1578acb4f00aa6e0d23e966b1af14f07758e2c70bfb97af660f59d66ee9acfd.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": "the dimension of each attention head decreases, the feature space it can represent becomes increasingly constrained, ultimately limiting the model's expressive capacity. Through mathematical analysis, we observe that " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}\\mathbf{K}^{\\mathrm{T}} = (\\mathbf{X}\\mathbf{W}_{q})(\\mathbf{X}\\mathbf{W}_{k})^{T} = \\mathbf{X}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T})\\mathbf{X}^{T}" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": ". The absence of non-linearity between " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": " imposes a rank constraint: " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{rank}(\\mathbf{Q}\\mathbf{K}) \\leq \\mathrm{rank}(\\mathbf{W}_{q}\\mathbf{W}_{k}^{T}) \\leq d" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": ". This constraint reveals a critical trade-off. When " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "d_{h} < d" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": ", the model's expressiveness suffers from rank deficiency. Conversely, " + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "inline_equation", + "content": "d_{h} > d" + }, + { + "bbox": [ + 313, + 461, + 555, + 594 + ], + "type": "text", + "content": " introduces redundant parameters without theoretical benefits." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 323, + 611, + 555, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 611, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 323, + 611, + 555, + 640 + ], + "type": "interline_equation", + "content": "\\mathbf {L e a n S H S A} (X) = \\operatorname {S o f t m a x} \\left[ \\frac {\\mathbf {X W} _ {1} \\mathbf {X} ^ {T}}{\\sqrt {d}} \\right] \\cdot \\mathbf {X W} _ {2} \\quad (5)", + "image_path": "7e9ba1f41a40f0937ee0a65ecb200b3f3752abd14149db4ab4e702b0c6b09030.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": "To optimize this trade-off, we propose Lean Single-Head Self-Attention(LeanSHSA), as illustrated in Eq. 5. This approach sets " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "d_h = d" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " and merges projection matrices: " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_q\\mathbf{W}_k^T" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_1" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_v\\mathbf{W}_o" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_2" + }, + { + "bbox": [ + 313, + 653, + 555, + 714 + ], + "type": "text", + "content": ". By reducing four weight matrices to two, LeanSHSA signifi" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "19917" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 77, + 527, + 380 + ], + "blocks": [ + { + "bbox": [ + 86, + 77, + 527, + 380 + ], + "lines": [ + { + "bbox": [ + 86, + 77, + 527, + 380 + ], + "spans": [ + { + "bbox": [ + 86, + 77, + 527, + 380 + ], + "type": "image", + "image_path": "564e3e9be8a38feca0db4661caf6501a4cefbc5144625824c7ec05c8aecddd52.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 382, + 161, + 394 + ], + "lines": [ + { + "bbox": [ + 102, + 382, + 161, + 394 + ], + "spans": [ + { + "bbox": [ + 102, + 382, + 161, + 394 + ], + "type": "text", + "content": "b) ConvBlock" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 207, + 383, + 272, + 394 + ], + "lines": [ + { + "bbox": [ + 207, + 383, + 272, + 394 + ], + "spans": [ + { + "bbox": [ + 207, + 383, + 272, + 394 + ], + "type": "text", + "content": "c) SpatialBlock" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 383, + 385, + 393 + ], + "lines": [ + { + "bbox": [ + 315, + 383, + 385, + 393 + ], + "spans": [ + { + "bbox": [ + 315, + 383, + 385, + 393 + ], + "type": "text", + "content": "d) ChannelBlock" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "text", + "content": "Figure 6. Overview of TinyNeXt: A hierarchical architecture featuring a convolutional stem, four progressively designed stages with specific modules (ConvBlock, SpatialBlock, ChannelBlock), and a final classification head. " + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "text", + "content": " represent input image spatial dimensions, with " + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "inline_equation", + "content": "L_{i}" + }, + { + "bbox": [ + 55, + 410, + 555, + 444 + ], + "type": "text", + "content": " indicating the block count per stage." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 464, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 572 + ], + "type": "text", + "content": "cantly simplifies the attention mechanism while maintaining its theoretical expressiveness. Furthermore, the simplified matrix operations align better with modern hardware accelerators, enabling more efficient parallel processing and better utilization of computing resources. Through these simplifications, LeanSHSA effectively reduces the latency caused by memory-inefficient operations, achieving a near " + }, + { + "bbox": [ + 55, + 464, + 295, + 572 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 464, + 295, + 572 + ], + "type": "text", + "content": " improvement in CPU throughput compared to MHSA, as demonstrated in Fig. 5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 585, + 138, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 138, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 138, + 596 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "type": "text", + "content": "Based on our previous analysis and findings, we systematically optimize ViT architectures by adapting module designs to the evolving feature characteristics across network stages. Fig. 6 depicts an overview of the proposed TinyNeXt architecture. TinyNeXt follows the pyramid architecture, comprising a sequential arrangement of a stem, four stages of varying resolutions, and a classification head, with downsampling layers connecting the successive stages." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The stem layer serves as the initial feature extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 464, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 513 + ], + "type": "text", + "content": "module, processing the input image and producing a feature representation. Similar to [25, 47], we stack two standard " + }, + { + "bbox": [ + 313, + 464, + 555, + 513 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 464, + 555, + 513 + ], + "type": "text", + "content": " convolutional layers with a stride of 2 as the stem, downsampling the input image by a factor of 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 517, + 556, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 649 + ], + "type": "text", + "content": "In the early network stages (Stage 1 and Stage 2), characterized by localized feature representations, we employ the ConvBlock, which originates from the inverted residual block [41]. This choice is driven by the need to efficiently capture local spatial information, such as texture patterns and edge details, which are predominantly concentrated in the initial layers of the network. By leveraging depth-wise separable convolutions and an efficient expansion-compression structure, it effectively extracts fine-grained details and patterns from the input data while maintaining computational efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "In the middle stage (Stage 3), after multiple downsampling operations, the feature scales become more moderate, and the low-frequency spatial information is retained. In this case, we have employed the SpatialBlock to effectively extract the long-range dependencies between features. The" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19918" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 113, + 535, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Classification performance on ImageNet-1K. Throughput is measured on an Nvidia RTX 3090 GPU with batch size of 512 for GPU and AMD EPYC 7542 32-Core Processor with batch size of 16 for CPU. Larger throughput means faster inference speed. FLOP count is computed by fvcore library [14]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 113, + 535, + 401 + ], + "lines": [ + { + "bbox": [ + 74, + 113, + 535, + 401 + ], + "spans": [ + { + "bbox": [ + 74, + 113, + 535, + 401 + ], + "type": "table", + "html": "
ModelInputParams (M)FLOPs (M)Throughput (img/s)Top-1 (%)Top-5 (%)
GPUCPU
ShuffleNetv2-1.0 [32]2242.31461208510469.488.9
PVTv2-b0 [49]2243.757225862670.5-
MobileNetv1 [22]2244.257953104570.6-
EfficientViT-M2 [43]2244.22011668814170.890.2
MobileOne-S0 [46]2242.127584107971.4-
TinyNeXt-T2241.025972578571.590.3
MobileNetv2-1.0 [41]2243.530045723872.091.0
MobileViTv2-0.5 [35]2561.446644352270.2-
MobileViT-XXS [34]2561.341033742369.0-
EdgeNeXt-XXS [33]2561.326146796471.2-
EMO-1M [60]2241.326139494071.590.4
TinyNeXt-S2241.330465557572.791.0
MobileNetv2-1.4 [41]2246.158530911974.7-
MobileNetv3-L-1.0 [21]2245.421958646175.2-
MobileOne-S1 [46]2244.882545543175.9-
EdgeViT-XXS [38]2244.155733263674.4-
MobileViT-XS [34]2562.39862028874.8-
EdgeNeXt-XS [33]2562.353829063275.0-
EMO-2M [60]2242.343927792975.192.2
TinyNeXt-M2242.347548895175.392.3
", + "image_path": "13b1a108e26ea8399222ac3e2826e17f71452ed724ff08d145bc605831236df4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 418, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 503 + ], + "type": "text", + "content": "SpatialBlock is mainly composed of three components: a LeanSHSA layer for comprehensive global context modeling, a " + }, + { + "bbox": [ + 55, + 418, + 296, + 503 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 418, + 296, + 503 + ], + "type": "text", + "content": " lightweight Depthwise Convolution (DWConv) layer dedicated to local feature aggregation or conditional position embedding [6, 53], and a Multilayer Perceptron (MLP) for channel feature interaction. Formally, the SpatialBlock can be represented as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 509, + 250, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 509, + 250, + 522 + ], + "spans": [ + { + "bbox": [ + 101, + 509, + 250, + 522 + ], + "type": "interline_equation", + "content": "X _ {1} = \\operatorname {L e a n S H S A} \\left(\\mathrm {B N} \\left(X _ {\\text {i n}}\\right)\\right) + X _ {\\text {i n}},", + "image_path": "4bda1130189d2d122ed431352c75ef1550058bd88b6b07b2a8bf871047ae7e12.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 524, + 216, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 524, + 216, + 536 + ], + "spans": [ + { + "bbox": [ + 102, + 524, + 216, + 536 + ], + "type": "interline_equation", + "content": "X _ {2} = \\operatorname {D W C o n v} \\left(X _ {1}\\right) + X _ {1},", + "image_path": "1e2e12f2264e65b5a99bc404dad72886413cdfb345d043872a2ed84003af119f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 539, + 225, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 539, + 225, + 552 + ], + "spans": [ + { + "bbox": [ + 102, + 539, + 225, + 552 + ], + "type": "interline_equation", + "content": "X _ {\\text {o u t}} = \\operatorname {M L P} (\\operatorname {B N} (X _ {2})) + X _ {2}.", + "image_path": "bde80398d836d21237f9f0cd64f449b77f10b87aa3499cf1e15cee1d4f3eaaf0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 665 + ], + "type": "text", + "content": "In the last stage (Stage 4), as feature scales further contract and spatial characteristics diminish, semantic representations become increasingly pronounced. Consequently, we deploy the ChannelBlock, which leverages a SE module to dynamically recalibrate channel-wise features. This module, structurally akin to the SpatialBlock but replacing the LeanSHSA layer with an SE module, enables adaptive feature emphasis and enhances the network's discriminative potential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Overall, in designing TinyNeXt, we have avoided overly complex operations, focusing instead on selecting the most appropriate basic modules based on the evolving characteristics of feature. This approach ensures a balance between" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 418, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 467 + ], + "type": "text", + "content": "computational efficiency and feature representation capability. We build three TinyNext variants with different setting of depth, width and expansion ratio, the detailed model specifications are demonstrated as Tab 3." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 498, + 549, + 545 + ], + "blocks": [ + { + "bbox": [ + 329, + 478, + 539, + 489 + ], + "lines": [ + { + "bbox": [ + 329, + 478, + 539, + 489 + ], + "spans": [ + { + "bbox": [ + 329, + 478, + 539, + 489 + ], + "type": "text", + "content": "Table 3. Architecture details of TinyNeXt model variants." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 498, + 549, + 545 + ], + "lines": [ + { + "bbox": [ + 318, + 498, + 549, + 545 + ], + "spans": [ + { + "bbox": [ + 318, + 498, + 549, + 545 + ], + "type": "table", + "html": "
ModelsTinyNeXt-TTinyNeXt-STinyNeXt-M
Depth[3, 3, 6, 2][3, 3, 8, 3][4, 4, 9, 4]
Width[32, 64, 96, 192][32, 64, 96, 192][32, 64, 128, 256]
Exp.Ratio[2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 2.0][2.0, 2.0, 2.0, 1.5]
", + "image_path": "90d1b5eabca3d4f406d2a725d3532cec5415ac5a62a9a2d05523cef41370c7f2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 568, + 395, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 395, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 395, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "type": "text", + "content": "4.1. Image Classification" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Implementation Details. We conduct image classification experiments on ImageNet-1K [9] using PyTorch 2.3.0 [39] and Timm library 0.5.4 [51]. All models are trained from scratch at a resolution of " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " for 300 epochs, utilizing 8 Nvidia RTX 3090 GPUs. We employ the AdamW optimizer [31], warm-up training with 5 epochs and a cosine annealing learning rate scheduler. The total batch size is set as 2,048 and the initial learning rate is " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "6 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " with weight decay of " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-2}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". The data augmentations ap" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19919" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "plied during training included Random Resized Crop, Random Horizontal Flip, and Color Jitter, without the use of Mixup, CutMix, or Random Erasing. Moreover, we evaluated the model's throughput across desktop GPU and CPU platforms. On the GPU platform, throughput measurements were conducted using an NVIDIA RTX 3090 with a batch size of 512. For CPU throughput evaluation, we employed an AMD EPYC 7542 32-Core Processor, performing measurements with a batch size of 16 in a single-threaded." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": "Results. TinyNeXt is systematically benchmarked against SOTA lightweight models across multiple scales, with the comprehensive evaluation results detailed in Tab. 2. TinyNeXt achieves competitive or superior results in terms of accuracy, parameter efficiency, and inference throughput. TinyNeXt-T achieves an impressive " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "71.5\\%" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " top-1 accuracy with only 1.0M parameters, demonstrating notable improvements over MobileNetv1 [22] and PVTv2-b0 [48]. Specifically, it achieves an accuracy gain of " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "0.9\\% \\uparrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "1.0\\% \\uparrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": ", respectively, while reducing parameters by " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "3.2\\mathrm{M} \\downarrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "2.7\\mathrm{M} \\downarrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": ", with corresponding GPU throughput enhancements of " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "1.4\\times" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "2.8\\times" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": ". At the small scale, TinyNeXt-S, with a comparable parameters of 1.3M to MobileViT-XXS, EdgeNeXt-XXS, and EMO-1M [33, 34, 60], significantly outperforms these models in both accuracy and multi-platform throughput. It achieves a " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "3.7\\% \\uparrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " improvement in top-1 accuracy over MobileViT-XXS, with GPU throughput increasing by " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "1.9\\times" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": ". Compared to the classic lightweight CNN model MobileNetv2 [41], TinyNeXt-S demonstrates remarkable performance, improving " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "0.7\\% \\uparrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " top-1 accuracy and reducing " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "37\\% \\downarrow" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " parameters. Moreover, TinyNeXt-M achieves the highest top-1 accuracy of " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "75.3\\%" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " under the same parameter constraints when compared with MobileViT-XS, EdgeNeXt-XS, and EMO-2M [33, 34, 60], while delivering an exceptional CPU throughput that is " + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "inline_equation", + "content": "6.4\\times" + }, + { + "bbox": [ + 56, + 180, + 298, + 517 + ], + "type": "text", + "content": " that of MobileViT-XS. With its excellent balanced accuracy and efficiency, TinyNeXt demonstrates potential as a viable option for resource-constrained applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 527, + 167, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 167, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 167, + 539 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Object Detection. We integrate TinyNeXt into the SSDLite framework [21] for object detection on the MS COCO 2017 dataset [28] at " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "320 \\times 320" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " resolution. The implementation leverages the MMDetection library [3], employing the AdamW optimizer [31] with a batch size of 64, and inference throughput (images/s) is benchmarked on an NVIDIA RTX 3090 GPU. As shown in Table 4, TinyNeXt demonstrates superior computational efficiency with competitive accuracy. Specifically, TinyNeXt-S achieves " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "22.4\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " using only " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{M}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " parameters and 2,517 images/s throughput. It reduces parameters by " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.8 / 2.0\\mathrm{M}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " compared to MobileNetv1/v2[22, 41], and surpasses MobileViTv2-0.5/EMO-1M[35, 60] by " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.2 / 0.4\\mathrm{mAP}" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.2 / 1.5 \\times" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " throughput. The medium-scale TinyNeXt-M" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 102, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 4. Object detection results on MS COCO 2017 with the SSDLite framework." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 102, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 102, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 102, + 553, + 232 + ], + "type": "table", + "html": "
BackboneParams↓Throughput↑mAP↑
MobileNetv1 [22]5.1M228222.2
MobileNetv2 [41]4.3M202922.1
MobileViT-XXS [34]1.7M152819.9
MobileViTv2-0.5[35]2.0M205821.2
EMO-1M[60]2.3M168122.0
TinyNeXt-S2.3M251722.4
MobileViTv2-0.75[35]3.6M138824.6
EMO-2M[60]3.3M119425.2
TinyNeXt-M3.3M191125.0
", + "image_path": "60821d57e77543e7f5d5bb183446cd87da4f8e8e7b1b5ec3f17030d658b8eea8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 278, + 545, + 346 + ], + "blocks": [ + { + "bbox": [ + 313, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 270 + ], + "type": "text", + "content": "Table 5. Semantic segmentation results on PASCAL VOC 2012 with the DeepLabv3 framework." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 278, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 321, + 278, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 321, + 278, + 545, + 346 + ], + "type": "table", + "html": "
BackboneParams↓Flops↓mIOU↑
MobileNetV2 [41]4.5M5.8G75.7
MobileViTv2-0.75[35]9.6M40.0G75.1
TinyNeXt-S3.5M3.5G75.5
TinyNeXt-M4.9M5.1G76.9
", + "image_path": "f1cb72b520c6ecaf273613631aa410f952df86d902d3b52415f769ca0df363db.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 404, + 553, + 540 + ], + "blocks": [ + { + "bbox": [ + 313, + 362, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 396 + ], + "type": "text", + "content": "Table 6. Performance Comparison on Edge Devices. The last two columns present inference latencies measured on edge computing platforms: NVIDIA Jetson NANO and ARM Cortex-A57." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 404, + 553, + 540 + ], + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 540 + ], + "type": "table", + "html": "
ModelInputParams (M)Top1 (%)NANO (ms)Cortex-A57 (ms)
MobileViT-XXS [34]25621.369.030.1467.9
EdgeNeXt-XXS [33]25621.371.233.5242.8
EMO-1M [60]22421.371.526.2526.4
TinyNeXt-S22421.372.714.3227.7
MobileViT-XS [34]25622.374.750.8859.1
EdgeNeXt-XS [33]25622.375.050.8442.4
EMO-2M [60]22422.375.135.7759.2
TinyNeXt-M22422.375.319.4311.6
", + "image_path": "45238727ed2c0bd67ad42c8135a26e46d31135b98531bc869c82a3d02251e868.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "text", + "content": "attains " + }, + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "inline_equation", + "content": "25.0\\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "text", + "content": " with 3.3M parameters, matching EMO-2M in accuracy while providing " + }, + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "inline_equation", + "content": "1.6\\times" + }, + { + "bbox": [ + 313, + 567, + 553, + 590 + ], + "type": "text", + "content": " faster inference." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": "Semantic Segmentation. We integrate TinyNeXt into the DeepLabv3 framework [4] for semantic segmentation on the Pascal VOC 2012 dataset [13] at " + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": " resolution. Our implementation utilizes the MMSegmentation library [7], trains with the AdamW optimizer [31] for 40K iterations using a total batch size of 16. As summarized in Table 5, both variants provide superior accuracy and efficiency. Specifically, TinyNeXt-M surpasses MobileNetV2 by " + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": " mIOU with 0.7G fewer FLOPs, further demonstrating its effectiveness as a vision backbone across tasks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19920" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "text", + "content": "4.3. Inference on Edge Devices" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 294, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 294, + 220 + ], + "type": "text", + "content": "Implementation Details. To validate our model's potential for TinyML applications, we conduct a comprehensive evaluation of TinyNeXt's inference performance against SOTA models, including MobileViT [34], EdgeNeXt [33], and EMO [60], across diverse edge computing platforms. Specifically, we evaluate their inference capabilities on edge GPU and CPU architectures, utilizing platforms such as the NVIDIA Jetson Nano and ARM Cortex-A57. For each platform, we implement a unified testing flow: an initial 10s warm-up phase followed by 60s of continuous inference to accurately measure average latency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": "Results. TinyNeXt achieves a superior accuracy-latency trade-off across all evaluated edge platforms. With 1.3M parameters, TinyNeXt-S demonstrates " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " faster inference on NVIDIA Jetson Nano while maintaining " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " higher top-1 accuracy, and reduces ARM Cortex-A57 latency by " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "56.7\\%" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " compared to EMO-1M [60]. At 2.3M parameters, TinyNeXt-M further keeps these advantages, achieving " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " speedup over EMO-2M [60] on edge GPU with " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " accuracy improvement, while delivering " + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "inline_equation", + "content": "2.8 \\times" + }, + { + "bbox": [ + 55, + 221, + 294, + 341 + ], + "type": "text", + "content": " faster CPU execution than MobileViT-XS [34]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 149, + 361 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 295, + 474 + ], + "type": "text", + "content": "In this section, we conduct ablation studies on key design elements of the proposed TinyNeXt using the ImageNet-1K dataset [9]. To highlight architectural differences and reduce training time, all models are trained for 100 epochs [16, 29]. Throughout the ablation experiments, we consistently preserve the baseline architecture's original depth and width configurations. This approach facilitates a precise evaluation of individual design choices' performance impact." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": "Impact of the attention block. These ablation experiments validate LeanSHSA's efficiency through substitutions with standard MHSA and SHSA modules. As shown in Tab. 7 (#2 and #3), replacing LeanSHSA with 4-head MHSA increases parameters by " + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "0.3\\mathrm{M}" + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": " for merely " + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": " accuracy gain, while incurring " + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": " throughput penalty. This empirically demonstrates that LeanSHSA's streamlined computational design effectively enhances inference efficiency without compromising global feature extraction capabilities. Furthermore, adopting a bottlenecked SHSA approach marginally reduces accuracy by " + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": " despite equivalent parameter count, with only a " + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 474, + 295, + 641 + ], + "type": "text", + "content": " speed gain. This suggests the importance of maintaining head dimensionality for better feature extraction in single-head architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Impact of the last stage choice. This ablation experiment evaluates the effectiveness of TinyNeXt's macro architecture by modifying the feature extraction module in the last stage. As shown in Tab. 7 (#4, #5, #6), we replace the channel-wise attention module (SE) with various global spatial attention modules, including MHSA (with 4 heads" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 322, + 113, + 547, + 213 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 105 + ], + "type": "text", + "content": "Table 7. Ablation Study of TinyNeXt-M on ImageNet-1K [9]. The last three columns present model parameters (M), top-1 accuracy (\\%), and GPU throughput (images/s)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 113, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 322, + 113, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 322, + 113, + 547, + 213 + ], + "type": "table", + "html": "
#AblationParamsTop-1Throughput
1TinyNeXt-M (baseline)2.374.14889
2LeanSHSA → MHSA2.674.33950
3LeanSHSA → SHSA2.373.95137
4Last Stage with MHSA2.774.14718
5Last Stage with SHSA2.773.64788
6Last Stage with LeanSHSA2.773.94743
7BN → LN2.373.84479
", + "image_path": "4fc9ff745698fc33b23fa23df9e42287e001d2784e2c901f68ec72709bce81ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 237, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 380 + ], + "type": "text", + "content": "and a head dimension of 32), SHSA, and LeanSHSA. These substitutions increase the model size by 0.4M parameters compared to the baseline, but yield no accuracy gains and even slightly reduce throughput. For example, using MHSA in the last stage (#4) raises parameters from 2.3M to 2.7M and cuts throughput by nearly 200 images/s, while the top-1 accuracy remains unchanged at " + }, + { + "bbox": [ + 313, + 237, + 555, + 380 + ], + "type": "inline_equation", + "content": "74.1\\%" + }, + { + "bbox": [ + 313, + 237, + 555, + 380 + ], + "type": "text", + "content": ". Similar trends are observed with SHSA and LeanSHSA variants. These results demonstrate that channel-wise attention is more effective in deep layers, where feature representations shift from being spatially dominant to semantically dominant, offering superior semantic discrimination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "text", + "content": "Impact of the normalization layer. We investigate the impact of normalization layers by replacing Batch Normalization (BN) with Layer Normalization (LN) in SpatialBlock and ChannelBlock. LN requires dynamic mean and variance computation during inference, which poses challenges for edge device deployment. Empirical evaluation (#7) reveals that replacing BN with LN results in a " + }, + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "text", + "content": " reduction in inference throughput and a marginal " + }, + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 383, + 556, + 503 + ], + "type": "text", + "content": " decline in top-1 accuracy. Consequently, we retain BN as the preferred normalization strategy in our TinyNeXt architecture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 388, + 533 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 543, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 639 + ], + "type": "text", + "content": "This paper analyzes the limitations of current efficient ViTs at both micro and macro levels and proposes corresponding optimizations. Specifically, we introduce a simplified and efficient attention variant, LeanSHSA, and enhance model performance through stage-specific macro designs. Extensive experiments demonstrate TinyNeXt's exceptional trade-off of model size, accuracy, and speed across diverse vision tasks and hardware platforms." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Limitations. The computational complexity of Lean-SHSA remains quadratic with visual tokens, leading to high costs in high-resolution tasks and restricting its applicability in such scenarios. In future work, we aim to explore dynamic attention mechanisms while minimizing the overhead of memory-bound operations." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19921" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "This work was supported in part by the National Science and Technology Innovation 2030-Major Projects under Grant 2021ZD0114400, in part by the Young Scientists Fund of the National Natural Science Foundation of China No.62304162, in part by the China Postdoctoral Science Foundation under Grant Number 2024M762532, in part by the Postdoctoral Fellowship Program of CPSF under Grant GZC20241313, in part by Shaanxi Provincial Natural Science Foundation for Basic Research Program under Grant 2024JC-YBMS-794, and in part by the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24090." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 266, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 266, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 296, + 297 + ], + "type": "text", + "content": "[1] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 300, + 296, + 343 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV). Springer, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 345, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 345, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 296, + 398 + ], + "type": "text", + "content": "[3] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 296, + 443 + ], + "type": "text", + "content": "[4] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 499 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 534 + ], + "type": "text", + "content": "[6] Xiangxiang Chu, Zhi Tian, Bo Zhang, Xinlong Wang, and Chunhua Shen. Conditional positional encodings for vision transformers. arXiv preprint arXiv:2102.10882, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 296, + 578 + ], + "type": "text", + "content": "[7] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 296, + 623 + ], + "type": "text", + "content": "[8] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 668 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 670, + 297, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 297, + 714 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 172 + ], + "type": "text", + "content": "[12] Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. arXiv preprint arXiv:1902.08153, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 216 + ], + "type": "text", + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "text", + "content": "[14] facebookresearch. fvcore library. https://github.com/facebookresearch/fvcore, 2019.6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 304 + ], + "type": "text", + "content": "[15] Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari, Sunando Sengupta, Hamid Reza Vaezi Joze, Eric Sommerlade, Hamed Pirsiavash, and Jürgen Gall. Adaptive token sampling for efficient vision transformers. In European Conference on Computer Vision, pages 396-414. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 306, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 554, + 370 + ], + "type": "text", + "content": "[16] Benjamin Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, and Matthijs Douze. Levit: a vision transformer in convnet's clothing for faster inference. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12259-12269, 2021. 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 414 + ], + "type": "text", + "content": "[17] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[18] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. arXiv preprint arXiv:2306.06189, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "text", + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 554, + 601 + ], + "type": "text", + "content": "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[22] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 669 + ], + "type": "text", + "content": "[23] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19922" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 1, 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[26] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16889-16900, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "text", + "content": "[27] Yishi Li, Yuhao Zhang, and Rui Lai. Tinypillarnet: Tiny pillar-based network for 3d point cloud object detection at edge. IEEE Transactions on Circuits and Systems for Video Technology, 34(3):1772-1785, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 297 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "text", + "content": "[29] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 432 + ], + "type": "text", + "content": "[31] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[32] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European conference on computer vision (ECCV), pages 116-131, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "text", + "content": "[33] Muhammad Maaz, Abdelrahman Shaker, Hisham Cholakkal, Salman Khan, Syed Waqas Zamir, Rao Muhammad Anwer, and Fahad Shahbaz Khan. Edgenext: efficiently amalgamated cnn-transformer architecture for mobile vision applications. In European conference on computer vision, pages 3-20. Springer, 2022. 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[34] Sachin Mehta and Mohammad Rastegari. Mobilevit: lightweight, general-purpose, and mobile-friendly vision transformer. arXiv preprint arXiv:2110.02178, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "text", + "content": "[35] Sachin Mehta and Mohammad Rastegari. Separable self-attention for mobile vision transformers. arXiv preprint arXiv:2206.02680, 2022. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[36] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[37] Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yely-sei Bondarenko, Mart Van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295, 2021. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 129 + ], + "type": "text", + "content": "[38] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martinez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In European Conference on Computer Vision, pages 294-311. Springer, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[39] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 2, 4, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 553, + 250 + ], + "type": "text", + "content": "[40] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 308 + ], + "type": "text", + "content": "[41] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4510-4520, 2018. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[43] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[44] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 489 + ], + "type": "text", + "content": "[45] Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. Maxvit: Multi-axis vision transformer. In European conference on computer vision, pages 459-479. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 491, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 544 + ], + "type": "text", + "content": "[46] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7907-7917, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[47] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15909-15920, 2024. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 667 + ], + "type": "text", + "content": "[48] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[49] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2, 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19923" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[50] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7794-7803, 2018. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 272, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 272, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 272, + 129 + ], + "type": "text", + "content": "[51] Ross Wightman et al. Pytorch image models, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 174 + ], + "type": "text", + "content": "[52] Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. In European conference on computer vision, pages 68-85. Springer, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "text", + "content": "[53] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollar, and Ross Girshick. Early convolutions help transformers see better. Advances in neural information processing systems, 34:30392-30400, 2021. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 275 + ], + "type": "text", + "content": "[54] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Advances in Neural Information Processing Systems (NeurIPS), pages 12077-12090, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 320 + ], + "type": "text", + "content": "[55] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In European conference on computer vision, pages 664-680. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 364 + ], + "type": "text", + "content": "[56] Kunran Xu, Yishi Li, Huawei Zhang, Rui Lai, and Lin Gu. Etinynet: Extremely tiny network for tinyml. In Proceedings of the AAAI conference on artificial intelligence, pages 4628-4636, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 422 + ], + "type": "text", + "content": "[57] Hongxu Yin, Arash Vahdat, Jose M Alvarez, Arun Mallya, Jan Kautz, and Pavlo Molchanov. A-vit: Adaptive tokens for efficient vision transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10809-10818, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 466 + ], + "type": "text", + "content": "[58] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5756-5767, 2024. 2, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 523 + ], + "type": "text", + "content": "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jianming Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In International Conference on Learning Representations (ICLR), 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 590 + ], + "type": "text", + "content": "[60] Jiangning Zhang, Xiangtai Li, Jian Li, Liang Liu, Zhucun Xue, Boshen Zhang, Zhengkai Jiang, Tianxin Huang, Yabiao Wang, and Chengjie Wang. Rethinking mobile block for efficient attention-based models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1389-1400. IEEE Computer Society, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 591, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 667 + ], + "type": "text", + "content": "[61] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6881-6890, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[62] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 182 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[63] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[64] Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8697-8710, 2018. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "19924" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_content_list.json b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..051bfe5a54ccf986edc4ce5ea9116f0f0b79f51c --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_content_list.json @@ -0,0 +1,1155 @@ +[ + { + "type": "text", + "text": "An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval", + "text_level": 1, + "bbox": [ + 156, + 128, + 843, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeseok Byun $^{1*}$ Seokhyeon Jeong $^{1*}$ Wonjae Kim $^{2}$ Sanghyuk Chun $^{2\\dagger}$ Taesup Moon $^{1,3\\dagger}$", + "bbox": [ + 142, + 202, + 859, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of ECE, Seoul National University $^{2}$ NAVER AI Lab $^{3}$ ASRI/INMC/IPAI/AIIS, Seoul National University", + "bbox": [ + 230, + 237, + 764, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches. The mainstream Zero-Shot (ZS) CIR methods bypass the need for expensive training CIR triplets by projecting image embeddings into the text token embedding space, forming a composed query for retrieval. However, we highlight an inherent limitation in these projection-based CIR: a task discrepancy of text encoders between the original pre-training task of the encoders (text $\\leftrightarrow$ image) and the target CIR task (image + text $\\leftrightarrow$ image), which potentially negatively impacts CIR performance. To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. Instead, we introduce Reducing Task Discrepancy of Text Encoders (RTD), an efficient text-only post-hoc framework that complements projection-based CIR methods. We devise a novel frozen-target text contrastive learning designed to enhance the capability of the text encoder for CIR. We also propose two key enhancements: (1) a hard negative-based refined batch sampling strategy and (2) a refined concatenation scheme to further mitigate training-inference discrepancy. Integrating RTD into state-of-the-art projection-based methods achieves performance comparable to, or even surpassing, resource-intensive state-of-the-art synthetic CIR triplet-based approaches, only with 23 minutes of additional training on 4 A100 GPUs. Our code is available in https://github.com/jaeseokbyun/RTD.", + "bbox": [ + 91, + 340, + 483, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 780, + 220, + 795 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Composed Image Retrieval (CIR) aims at retrieving a target image that closely resembles a reference image while reflecting the changes described in a conditioning text. Using a", + "bbox": [ + 89, + 806, + 485, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "query composed of image and text allows users to conduct more precise and flexible searches by specifying the desired modifications to the image through text. Supervised CIR methods [1, 11, 18] have been introduced to fuse information from bi-modal query, using labeled data from the target domain in the form of triplets $(I_r, T_c, I_t)$ , in which $I_r$ is a reference image, $T_c$ is a conditioning text, and $I_t$ is a target image. However, unlike the typical web-crawled image-text datasets [31], acquiring sufficient triplets for training needs expensive manual human annotations. To overcome the dependency on small-scale and human-verified triplets of the target domain, several works utilize the power of recent generative models. For example, a line of studies [14, 19, 37, 40] uses text-to-image models like IP2P [4] or large-language models (LLM) [5, 35] to synthesize large-scale CIR triplets for training, in place of the target-domain CIR triplets. While these methods achieve strong performance, they are often impractical due to the high computational and memory requirements for utilizing generative models.", + "bbox": [ + 511, + 310, + 906, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another approach for removing the dependency on the CIR triplets, which often is referred to as projection-based ZS-CIR [3, 15, 23, 30, 33, 34], employs an integrable projection module on top of the pre-trained, frozen, and shared embedding space, such as CLIP [29]. Namely, a projection module \\(\\phi\\), which maps a CLIP image embedding to the CLIP text token embedding space, can be trained by solely using images [3, 30] or texts [15]. During inference, as illustrated in Fig. 1, these methods first project the embedding of the query image to a text token embedding \\([\\\\) ]\\) using the function \\(\\phi\\). This embedding is then combined with the conditioning text \\([T_c]\\) to create the prompt \"a photo of \\([\\\\) ]\\) that \\([T_c]\\)\", which is used as a query for the text-to-image retrieval.", + "bbox": [ + 511, + 598, + 908, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The core assumption of the projection-based CIR is that the pre-trained text encoder should be robust enough to combine information from both the projected text token embedding and the conditioning text. However, we argue that this can cause a significant task discrepancy for the pre-trained text encoder between the original image-text alignment pretraining task (of CLIP) and the CIR task. For example, in", + "bbox": [ + 511, + 795, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 863, + 217, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors. (E-mail: sanghyukc@princeton.edu, tsmoon@snu.ac.kr)", + "bbox": [ + 91, + 876, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3895", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a14d8142e868fc30bd171bcafe84ad2a7d31792c340085672451184cfcd03fc2.jpg", + "image_caption": [ + "Figure 1. The task discrepancy of projection-based ZS-CIR methods between the pre-training task (image-text alignment) and the ZS-CIR task (image-text composition)." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 903, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig. 1, consider an ideal caption that accurately describes the target image. Since the CLIP text and image encoders are learned through contrastive learning, we can expect that the target image embedding (Fig. 1c) will align well with the embedding (Fig. 1b) of the ideal caption. In contrast, in the projection-based CIR, the text encoder receives a concatenated caption that combines the projected token $[\\$ ]$ and the conditioning text, not an ideal caption. However, as noted by Baldrati et al. [2], the text encoder is not typically trained to encode complex textual modifications (e.g., addition or replacement) to the reference image, which are common in conditioning texts. As a result, there is no guarantee that the textual embedding of the concatenated caption (Fig. 1a) closely aligns with the target image embedding (Fig. 1c), which will undermine the retrieval performance.", + "bbox": [ + 88, + 347, + 485, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To that end, we devise a complementary post-processing approach for existing projection-based CIR methods that reduces the task discrepancy of the text encoder efficiently. Ideally, both image and text backbones would be updated using expensive CIR triplets $(I_r, T_c, I_t)$ ; however, we achieve this using cheap text triplets and only update the text encoder. These text triplets $(T_r, T_c, T_t)$ consist of a reference caption $(T_r, e.g., “a dog catching a frisbee”)$ , a conditioning text $(T_c, e.g., “change dog to cat”)$ , and a target caption $(T_t, e.g., “a cat catching a frisbee”)$ , respectively. Given that the reference captions $T_r$ are readily available from conventional caption datasets, $T_c$ and $T_t$ can be automatically generated without any human labor [24, 38] or intensive resources [14, 19, 37, 40]. Using these triplets, we introduce a frozen-target text contrastive learning, in which the text encoder updates the embedding of the concatenated caption $T_{r + c}$ of $T_r$ and $T_c$ (e.g., “a dog catching a frisbee”+ “change the dog to a cat”) to align closely with the fixed embedding of the target caption $T_t$ from the frozen text encoder. We further enhance this text-only training approach with two key components: a batch sampling strategy that ensures hard neg-", + "bbox": [ + 91, + 584, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "atives in each mini-batch and a refined concatenation scheme for $T_{r}$ and $T_{c}$ designed to reduce the training-inference task discrepancy caused by our text-only training framework.", + "bbox": [ + 511, + 347, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experimental results demonstrate that our proposed method, dubbed as RTD (Reducing Task Discrepancy of Text Encoders), consistently and considerably improves the performance in diverse evaluation datasets (CIRR [24], CIRCO [3], FashionIQ [38], COCO object composition [30], and GeneCIS [36]), when integrated with existing projection-based CIR methods (SEARLE [3], Pic2Word [30], LinCIR [15], Context-I2W [34], and FTI4CIR [23]). Consequently, compared to resource-intensive state-of-the-art synthetic CIR triplets-based method like MagicLens [40], RTD achieves comparable results, with less than a $1\\%$ gap in CIRR R@1 and $0.1\\%$ in FashionIQ R@10 when combined with FTI4CIR (ViT-L/14). When using larger or fine-tuned backbones (ViT-G/14 or FSC-CLIP [27]) with LinCIR, RTD achieves superior results than MagicLens in those metrics.", + "bbox": [ + 511, + 420, + 908, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We note that RTD achieves the above results with significantly higher training efficiency in both data generation and training time. Specifically, our rule-based text triplet generation requires less than 10 minutes to construct 1M text triplets $(570\\times$ faster than CIR triplet generation from Compodiff [14]). With 4 NVIDIA A100 GPUs, the additional training time for RTD is just 23 minutes for training the ViT-L/14 backbone, making it approximately 10 to $100\\times$ faster than synthetic CIR triplets-based methods. Even with a significantly larger backbone (ViT-G/14), training takes only 2.5 hours, remaining negligible in comparison. Moreover, we investigate the effectiveness of RTD across various text triplet generation strategies and different backbone sizes (ViT-B/32, ViT-L/14, ViT-G/14) and types (FSC-CLIP [27], CLIP [29]), showing its reproducibility and compatibility.", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3896", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Composed Image Retrieval. Unlike supervised CIR methods [1, 2, 11, 18], projection-based CIR methods [3, 12, 15, 23, 30, 33, 34] are built upon the frozen CLIP model, where a projection module $\\phi$ is trained without CIR triplets. Each method employs a different training scheme for different training schemes for $\\phi$ (See Sec. 4.1 for details). Several approaches avoid the need for any training by employing interpolation techniques [16], while others leverage powerful yet resource-intensive models such as LLMs and captioners [17, 39]. While our approach is built upon projection-based CIR methods, its training strategy is closely related to another category of CIR methods [14, 19, 37, 40] that use synthetically generated expensive CIR triplets. However, our method stands out by utilizing text triplets and updating only the text encoder, resulting in more efficient training while maintaining strong performance.", + "bbox": [ + 89, + 114, + 485, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task discrepancy between the CLIP pre-training task and CIR. Combiner [2] updates the text encoder to minimize the gap between the target caption feature and the sum of the reference image and conditioning text features. FashionERN [8] addresses reference image dominance by introducing a separate branch to amplify the impact of the conditioning text. However, both Combiner and FashionERN require expensive CIR triplets $(I_r, T_c, I_t)$ for training, whereas our approach uses cheap, automatically generated text-only triplets $(T_r, T_c, T_t)$ . As another example, Chen and Lai [6] synthesize a triplet of an original image, its caption, and the masked image, treating them as the target image, conditioning text, and reference image, respectively. This approach, however, still has a gap between conditioning text (e.g., \"change dog to cat\") and image caption (e.g., \"a dog catching a frisbee\"); furthermore, it needs the full fine-tuning of the CLIP model, resulting in changing the visual embeddings in the retrieval database. In contrast, RTD directly uses the conditioning texts for training and does not change the target visual encoder, enabling the reuse of pre-extracted CLIP embeddings. Lastly, CIREVL [17] or LDRE [39] reduce task discrepancy by generating descriptive captions of the composed query using a large captioning model [21] and LLM [5]. While effective without any training, they rely on expensive inference steps and require carefully crafted prompts by skilled users. Whereas, RTD is fully automated, human-free, and more efficient during inference.", + "bbox": [ + 89, + 375, + 486, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Main Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 232, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Obtaining text triplets", + "text_level": 1, + "bbox": [ + 89, + 816, + 299, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the task discrepancy, we collect cheaply and automatically generated text triplets $(T_r, T_c, T_t)$ , instead of directly using the expensive CIR triplets $(I_r, T_c, I_t)$ . Given reference caption $T_r$ from conventional caption datasets,", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "conditioning text $T_{c}$ and target text $T_{t}$ can be generated using two strategies: via large language models (LLMs) [4, 14, 19, 37] or, more efficiently, through rule-based templates [14]. We investigate both strategies and demonstrate that RTD consistently improves performance across them.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the LLM-based strategy, we use the publicly available text triplets by CompoDiff [14], which are employed in our main experiments if not specified. These triplets are generated by taking a given caption $T_r$ as an input of the fine-tuned LLM, whose output predicts the corresponding conditioning text $T_c$ and the target caption $T_t$ . Previous works, such as IP2P [4], CoVR [37], and CASE [19], have also explored generating text triplets using LLMs, differing in LLM model types, input data, and fine-tuning strategies. The original purpose of these text triplet generations is to construct CIR triplets $(I_r, T_c, I_t)$ , but all these works also release the corresponding text triplets used for their CIR triplet construction. In addition to these publicly available text triplets, we also implement and evaluate an efficient in-context learning generation strategy using LLaMA3-8B [13] without additional fine-tuning. We conduct experiments with all the aforementioned text triplets in Tab. 6 and observe that RTD consistently delivers significant enhancements.", + "bbox": [ + 511, + 167, + 908, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For cheap, rule-based strategy, we can extract a keyword (e.g., \"dog\") from $T_{r}$ (\"a dog catching a frisbee\") and replace it with a randomly chosen keyword (e.g., \"cat\") [14], forming $T_{c}$ . The conditioning text is then generated automatically by using pre-defined templates (e.g., \"change [original keyword] to [altered keyword]\"). Our experiments show that this simple rule-based variant performs similarly to the LLM-based one. More explanations and examples of each strategy are provided in Suppl. A.2, and more details of generation costs are in Suppl. C.", + "bbox": [ + 511, + 440, + 910, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Frozen-target text contrastive learning", + "text_level": 1, + "bbox": [ + 511, + 604, + 846, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now, we explain our approach to update the text encoder for mitigating the task discrepancy solely with the generated text triplets $(T_r, T_c, T_t)$ . We first assume that there exists a pretrained projection module $\\phi$ obtained by the projection-based CIR methods [3, 15, 30]. Recall that for a given reference image $I_r$ and conditioning text $T_c$ , the final composed feature is generated by passing the text prompt \"a photo of $\\phi(\\psi_V(I_r))$ that $T_c$ \" to the text encoder $\\psi_T$ , where $\\psi_V$ is the visual encoder and $\\phi$ is the projection module (See Fig. 1). We aim to update the text encoder $\\psi_T$ to reduce the discrepancy between the pretext task and projection-based CIR task using the text triplets while maintaining $\\psi_V$ and $\\phi$ frozen.", + "bbox": [ + 511, + 628, + 908, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Frozen-target text contrastive loss. We apply contrastive learning using a paired caption $(T_{r + c}, T_t)$ , where $T_{r + c}$ denotes a concatenated caption of reference caption $T_r$ and conditioning caption $T_c$ . Namely, we let the representation of the concatenated caption closely approximate that of the target caption. However, solely updating the text encoder", + "bbox": [ + 511, + 810, + 910, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3897", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a5e7d43e47e88ef4a5ae72637812a8905eee82d0d7055088e9ac20f4606a392.jpg", + "image_caption": [ + "Figure 2. Overview of RTD." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 903, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while fixing the image encoder can break the alignment between image and text encoders. To prevent the issue, we extract the text embedding of $T_{t}$ using the frozen text encoder $\\psi_{T}$ , while the concatenated caption $T_{r + c}$ is extracted from the learnable text encoder $\\psi_T^{tr}$ , initialized from $\\psi_{T}$ . Here, we assume that as the target caption $T_{t}$ is a standard caption, a text embedding $\\psi_{T}(T_{t})$ , is well-aligned with the frozen image embedding space. Based on this assumption, we freeze the target textual embedding to maintain the pretrained alignment as we learn new relationships. As shown in Sec. 4.3, freezing the target embedding is crucial for fine-tuning with our objective.", + "bbox": [ + 88, + 358, + 483, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now, we introduce our frozen-target text contrastive loss $\\mathcal{L}_{FCL}$ using two text encoders: a frozen pre-trained text encoder $\\psi_T$ and a learnable text encoder $\\psi_T^{tr}$ which is initialized with $\\psi_T$ . Textual latent embeddings $\\tilde{t}_{r + c}$ and $t_t$ are extracted from $\\psi_T^{tr}$ and $\\psi_T$ , respectively. Namely, $\\tilde{t}_{r + c} = \\psi_T^{tr}(E_w^{tr}(T_{r + c}))$ and $t_t = \\psi_T(E_w(T_t))$ , where $E_w$ is a word embedding layer. We aim to tune $\\psi_T^{tr}$ to minimize the distance between the concatenated textual embedding $\\tilde{t}_{r + c}$ and the target textual embedding $t_t$ while maximizing the distance from other textual embeddings within the batch. We employ a symmetric InfoNCE loss [7, 10], as follows:", + "bbox": [ + 89, + 539, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {F C L} = \\frac {1}{B} \\sum_ {k = 1} ^ {B} - \\log \\left(\\frac {e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {k}) / \\tau)}}{\\sum_ {j = 1} ^ {B} e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {j}) / \\tau)} + \\sum_ {j \\neq k} e ^ {(c (t _ {t} ^ {k} , t _ {t} ^ {j}) / \\tau)}}\\right) \\\\ - \\log \\left(\\frac {e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {k}\\right) / \\tau\\right)}}{\\sum_ {j = 1} ^ {B} e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)} + \\sum_ {j \\neq k} e ^ {\\left(c \\left(\\tilde {t} _ {r + c} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)}}\\right) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 726, + 482, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $c(\\cdot, \\cdot)$ denotes the cosine similarity, $B$ is the batch size, and $\\tau$ is a temperature.", + "bbox": [ + 89, + 854, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refined batch sampling strategy for hard negatives. To", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "further enhance the efficacy of updating the text encoder, we devise a simple yet effective batch sampling strategy that incorporates pairs of $(T_{r + c}, T_t)$ and $(T_r, T_r)$ within the same batch. For example, as presented in Fig. 2, a pair such as $(T_{r + c}: \\text{“A dog catching a frisbee} + \\text{change dog to cat}$ , $T_t: \\text{“A cat catching a frisbee})$ is sampled along with its corresponding reference pair $(T_r: \\text{“A dog catching a frisbee}$ , $T_r: \\text{“A dog catching a frisbee})$ in the same batch. This setup ensures that the concatenated text $T_{r + c}$ and its corresponding reference text $T_r$ implicitly act as hard negatives for each other, as their semantics are much more similar $(T_{r + c}$ is derived from $T_r)$ than those of other randomly sampled texts in the batch. Explicitly distinguishing the embedding of $T_{r + c}$ from that of $T_r$ conceptually aligns well with CIR task, as it encourages the model to better capture modifications (in $T_c$ ). Moreover, we believe including $(T_r, T_r)$ pairs in the contrastive learning helps the learnable text encoder $\\psi_T^{tr}$ remain closely aligned with the pre-trained encoder $\\psi_T$ .", + "bbox": [ + 511, + 358, + 906, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refined concatenation of reference and conditioning texts. As we use \"a photo of $[\\$ ]$ that $[T_c]$ \" for inference, a naive concatenation strategy also can suffer from training-inference task discrepancy. To tackle this issue, instead of simply concatenating the $T_r$ and $T_c$ , we use the prompt \"a photo of $[\\$ ]$ that $[T_c]$ \" for updating the text encoder, where $[\\$ ]$ is obtained by the reference caption $T_r$ with the projection module $\\phi$ . Instead of obtaining a pseudo-word token with latent image embedding $v$ , we utilize a textual latent embedding from the reference caption $T_r$ , i.e., $\\phi (t_r)$ . However, Gu et al. [15] showed that naively replacing the image encoder with the text encoder for the input of $\\phi$ will suffer from the modality gap [22], a phenomenon where text and image embeddings have a gap between them. We tackle this issue by injecting random noise into the textual token representation before it is processed by $\\phi$ , following Gu et al. [15]. More analyses on variations of noise are in Suppl. B.10.", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3898", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates the overview of RTD. We use CLIP backbone and pre-trained projection module $\\phi$ produced by the existing projection-based CIR methods. The text encoder is trained using the proposed loss function (Eq. (1)) while applying the refined batch sampling and concatenation scheme. During inference, the procedure mirrors that of existing projection-based CIR methods, except we utilize the updated text encoder $\\psi_T^{tr}$ instead of the frozen one $\\psi_T$ . Note that our method only updates the text encoder while the image encoder and the projection module are frozen.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 1: Low data acquisition cost. The text triplets we use can be obtained at a significantly lower cost than CIR triplets. First, the text triplet generation process requires only an easily obtainable caption dataset for generation, whereas other approaches that generate CIR triplets [19, 37, 40] necessitate image or video datasets along with an additional collection phase for semantically similar images or videos. Second, text triplet generation can avoid resource-intensive text-to-image generation [4, 14], making it $15 \\times$ faster than the CIR triplet generation process used in CompoDiff [14]. Furthermore, if we choose the rule-based text triplet generation strategy, the process becomes $570 \\times$ faster than the full CIR triplet generation with images; 1M text triplets generation takes just 0.1 hours. Lastly, the storage for the generated text triplets takes up only 100MB, whereas storing a similar quantity of images requires significantly more space (e.g., 400GB for CC3M [32]). More details can be found in Suppl. C.", + "bbox": [ + 91, + 243, + 485, + 501 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/16d49cb5275da5096f940ef47c9ae4163ec2b0f16a6da7a7a7ee4611e6c2b230.jpg", + "table_caption": [ + "Table 1. Training and inference time comparisons. Purple denotes the additional resource required for RTD. Inference time is measured on a single A100 GPU for a single image. Note that the training hours for SEARLE, LinCIR for 8 A100 GPUs [15] and their RTD variants are trained on 4 A100 GPUs. CompoDiff is trained on 128 A100 and MagicLens is trained on 128 TPUs." + ], + "table_footnote": [], + "table_body": "
MethodCLIP BackboneTraining (h)Inference (s)
SEARLEL/144.2h0.02s
SEARLE (+ RTD)L/144.58h (+0.38h)0.02s
LinCIRL/140.5h0.02s
LinCIR (+ RTD)L/140.88h (+0.38h)0.02s
LinCIRG/140.8h0.05s
LinCIR (+ RTD)G/143.3h (+2.5h)0.05s
CompoDiffL/14231h0.12s
MagicLensL/146hunknown
CIRVLL/14-3.02s
", + "bbox": [ + 93, + 595, + 478, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark 2: Training and inference efficiency. As reported in Tab. 1, the additional training cost of RTD is small; 0.38 hours for ViT-L/14 on 4 A100 GPUs. In contrast, CompoDiff and MagicLens require 231 hours on 128 A100 GPUs and 6 hours on 128 TPUs, respectively, making the cost of RTD negligible. Even with a larger ViT-G/14 backbone, training RTD takes only 2.5 hours on 4 A100 GPUs, still faster than synthetic CIR triplet-based methods and even comparable to SEARLE in ViT-L/14 (4.2h). This training efficiency mainly stems from the advantage of text-only training. As highlighted by Gu et al. [15], the training complexity for the text", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encoder is much lower than that for the visual encoder due to the relatively short token lengths of texts ( $\\approx 12$ ) compared to images (256). The average inference time of the CLIP ViT-L/14 image encoder is $\\times 3.5$ times slower than that of the text encoder. Another source of efficiency is that RTD requires relatively few iterations (at most 2000, and significantly fewer in most cases), as the text encoders are already pre-trained and only require minor adjustments to learn the relationships between text triplets. Moreover, in Suppl. B.4, we present a more efficient implementation by selectively updating only a few layers of the text encoder. Along with efficient training, RTD retains the fast inference speed of the original projection-based methods, remaining $150\\times$ faster than training-free CIREVL.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 320, + 645, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 345, + 700, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use the AdamW optimizer [25] with a weight decay of 0.01. The learning rate is set to $10^{-5}$ , with a batch size of 512. We select the text encoder model with the best zero-shot CIRR [24] dev R@1 score for evaluating RTD. We mainly use the visual and textual encoders of the CLIP ViT-L/14 [29] as our backbone. Unless otherwise noted, we use LLM-based 2.5M text triplets provided by CompoDiff [14] for the training. We set the $\\tau$ as 0.07 in Eq. (1) and scale the standard deviation of Gaussian distribution as 0.5 for the noise injection. More results on various noise distributions can be found in the Suppl. B.10. All experiments were conducted using four NVIDIA A100 GPUs with Python 3.8 and Pytorch [28].", + "bbox": [ + 511, + 369, + 908, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation datasets and metrics. We compare projection-based CIR methods on five benchmark datasets: CIRR [24], CIRCO [3], FashionIQ [38] (FIQ), COCO object composition [30], and GeneCIS [36]. Details of each dataset are in the Suppl. A.1. For CIRR, FIQ, COCO, and GeneCIS, we have reported their recall scores at the top K retrieval results $(\\mathbb{R}@\\mathbb{K})$ . Since the CIRCO dataset includes multiple positive images for each query, we use a ranking-based metric; mean Average Precision scores at the top K results $(\\mathrm{mAP}@\\mathrm{K})$ [9, 26]. For the main results, we compare the results on the FIQ validation split, as well as the test sets of CIRR and CIRCO. For the ablation studies and analyses, the validation splits of these three datasets are utilized. GeneCIS and COCO object composition results are in the Suppl. B.1.", + "bbox": [ + 511, + 566, + 908, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We evaluate the effect of our method when combined with publicly available projection-based ZS-CIR methods: Pic2Word [30], SEARLE [3], LinCIR [15], Context-I2W [34], and FTI4CIR [23]. All these methods share a similar core concept shown in Fig. 1, but use different training schemes. Pic2Word [30] optimizes contrastive loss between the image embedding and its projected text embedding of \"a photo of [\\$\" to obtain the projection module $\\phi$ . Similarly,", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3899", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/5ca1cfcd261a71d8b274f23d493b35eb285070dd1ebe44bd2e9d0e43bbe16808.jpg", + "table_caption": [ + "Table 2. Comparison with other baselines. Note that this comparison is not entirely fair due to differences in backbone models and training data. Purple denotes the performance gain from our method, while red and blue highlight the best and second-best scores, respectively." + ], + "table_footnote": [], + "table_body": "
MethodBackboneCIRRCIRCOFashionIQ
R@1R@5R@10mAP@5mAP@10mAP@25R@10R@50
Pic2Word [30]CLIP ViT-L/1424.251.564.18.39.110.125.344.9
+RTD27.9 (+3.6)56.2 (+4.8)68.5 (+4.4)9.1 (+0.9)9.6 (+0.5)10.7 (+0.6)27.6 (+2.3)48.9 (+4.0)
SEARLE [3]CLIP ViT-L/1424.952.365.711.612.714.325.045.3
+RTD26.6 (+1.7)56.2 (+3.9)69.0 (+3.3)16.5 (+4.9)17.9 (+5.2)19.8 (+5.4)29.3 (+4.4)50.7 (+5.4)
Context-I2W [34]CLIP ViT-L/1425.655.468.6---27.949.1
+RTD29.2 (+4.6)58.4 (+3.0)70.5 (+1.9)---28.1 (+0.2)49.5 (+0.4)
FTI4CIR [23]CLIP ViT-L/1425.955.667.715.116.318.129.450.9
+RTD29.1 (+3.2)58.8 (+3.2)70.4 (+2.7)16.2 (+1.1)17.4 (+1.1)19.4 (+1.3)30.6 (+1.2)51.7 (+0.8)
LinCIR [15]CLIP ViT-L/1423.852.966.513.014.115.827.447.7
+RTD26.6 (+2.9)56.2 (+3.3)69.0 (+2.5)17.1 (+4.1)18.1 (+4.0)20.1 (+4.3)30.2 (+2.8)51.1 (+3.4)
LinCIR [15]FSC-CLIP [27] ViT-L/1433.663.174.514.215.417.028.349.5
+RTD35.9 (+2.3)67.5 (+4.4)78.1 (+3.6)18.3 (+4.1)19.7 (+4.3)21.3 (+4.3)31.2 (+2.9)52.3 (+2.8)
KEDs [33]CLIP ViT-L/1426.454.867.2---26.847.9
CIREVL [17]CLIP ViT-L/1424.552.364.918.619.020.928.648.6
LDRE [39]CLIP ViT-L/1426.555.667.523.424.026.428.550.5
MT-CIR [6]CLIP ViT-L/1425.554.667.610.411.613.035.457.4
MagicLens [40]CLIP ViT-L/1430.161.774.429.630.833.430.752.5
Compodiff [14]CLIP ViT-L/1426.755.072.612.613.415.836.048.6
CoVR [37]BLIP ViT-L/1439.368.278.9---27.744.6
CASE [19]BLIP ViT-L/1435.465.878.5-----
LinCIR [15]CLIP ViT-G/1434.964.576.120.621.924.144.565.5
+RTD36.3 (+1.4)67.5 (+3.0)78.3 (+2.2)21.1 (+0.5)22.3 (+0.4)24.5 (+0.4)46.2 (+1.7)67.3 (+1.8)
Compodiff [17]CLIP ViT-G/1434.764.375.115.317.719.439.051.7
CIREVL [17]CLIP ViT-G/1426.755.174.526.827.630.032.252.4
LDRE [39]CLIP ViT-G/1436.266.477.331.132.235.032.555.4
", + "bbox": [ + 94, + 119, + 901, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SEARLE [3] employs a two-stage approach, starting with an optimization-based textual inversion phase followed by a distillation phase for the projection module $\\phi$ . LinCIR [15] introduces a language-only self-supervised task based on keyword token replacement, where the model encourages the original text embedding and the replaced text embedding (keyword tokens are replaced with the projection of the original text embedding via $\\phi$ ) to be aligned. Context-I2W [34] refines the projection module by selecting relevant visual information [34]. Lastly, FTI4CIR [23] separately maps images into subject- and attribute-oriented pseudo-word tokens [23]. Details on how RTD is combined with them can be found in Suppl. A.3.", + "bbox": [ + 88, + 441, + 485, + 640 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train these methods with the CLIP ViT-L/14 in our main experiments. To further verify the compatibility of RTD, we also report results for the ViT-B/32 backbone with SEARLE, Pic2Word, and LinCIR. Moreover, we conduct experiments with a larger backbone (ViT-G/14) and a fine-tuned CLIP model for enhanced compositionality (FSC-CLIP [27]) for LinCIR, enabled by its fast training capability. We use the publicly available pre-trained model for SEARLE (ViT-B/32, ViT-L/14), Pic2Word (ViT-L/14), Context-I2W (ViT-L/14), and FTI4CIR (ViT-L/14). Otherwise, we reproduce the results using the official implementation. When reproducing, we adhere to the same settings in the original papers. For example, we select the final last epoch model for the Pic2Word ViT-B/32 model and choose the model based on the best zero-shot CIRR dev R@1 score for LinCIR. We additionally compare our method with a diverse set of CIR approaches, including another projection-based", + "bbox": [ + 89, + 643, + 486, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ZS-CIR method (KEDs [33]), another attempt to address task discrepancy (MT-CIR [6]), approaches leveraging synthetically generated CIR triplets (CoVR [37], CASE [19], MagicLens [40] and Compodiff [14]), and the training-free methods (CIREvL [17], LDRE [39]).", + "bbox": [ + 511, + 441, + 908, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main results", + "text_level": 1, + "bbox": [ + 511, + 529, + 648, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 shows the overview of comparison results with state-of-the-art CIR methods. First, we assess the impact of integrating RTD with existing projection-based CIR methods (SEARLE, Pic2Word, LinCIR, Context-I2W, and FTI4CIR) across CIRR, CIRCO, and FashionIQ. RTD consistently boosts performance, yielding an average improvement of over 2.78 points. This trend also holds for the GeneCIS and COCO object composition datasets, as detailed in Suppl. B.1.", + "bbox": [ + 511, + 551, + 908, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we compare the integration of RTD with those leveraging synthetically generated CIR triplets (CoVR, CASE, MagicLens, and Compodiff). We focus on this comparison to ensure fairness, as leveraging text triplets and updating the textual backbone of RTD may not be fully aligned with the standard projection-based CIR setting. We observe that RTD delivers competitive or superior performance compared to these resource-intensive synthetic CIR triplet-based methods, while being significantly more efficient; over $10 - 100 \\times$ faster, as noted in Remark-2 of Sec. 3. For example, with the same CLIP ViT-L/14 backbone, LinCIR + RTD outperforms MagicLens and CoVR in FashionIQ R@10 and R@50 while FTI4CIR + RTD or Context-I2W + RTD achieve performance on par with or exceeding that of MagicLens, MT-CIR, and Compodiff in CIRR metrics.", + "bbox": [ + 509, + 674, + 910, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3900", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When combined with larger (ViT-G/14) or fine-tuned (FSC-CLIP) backbones, RTD achieves the best or second-best result on the CIRR and FashionIQ benchmarks. We believe this flexibility of using different backbones underscores the benefit of efficient text-only training of RTD, inherited by LinCIR [15]. We reemphasize that even with a larger backbone like ViT-G/14, RTD remains much faster in training than other synthetic triplet-based methods, as shown in Tab. 1.", + "bbox": [ + 89, + 90, + 483, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, across both backbones (ViT-L/14, G/14), LinCIR + RTD mostly outperforms training-free CIREVL and LDRE, except on CIRCO metrics. As shown in Tab. 1, CIREVL is significantly slower $(150\\times)$ than projection-based methods (including RTD), while LDRE incurs even larger inference time due to its ensemble-based strategy.", + "bbox": [ + 88, + 214, + 483, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3144acdbf4109eb5f4341a17ffc3a9784c9901afdf42918cdfedbf17c9226343.jpg", + "table_caption": [ + "Table 3. Ablation study. Unlike in Tab. 2, for ablation studies and analyses, validation splits of three CIR datasets are used for evaluation. We measure the impact of FCL loss (Eq. (1)), refined batch sampling (RB), and refined concatenation scheme (RC). All models are based on LinCIR ViT-L/14. The first row denotes the vanilla LinCIR without RTD. \"Avg\" denotes the average of all reported metrics. Bold indicates the best result." + ], + "table_footnote": [], + "table_body": "
TextFCL Frozen TargetRBRCCIRR (R) @5@10CIRCO (mAP) @10@25FIQ (R) @10@50Avg
--XX54.367.812.714.527.447.737.4
\\( \\left( {{T}_{r},{T}_{r}}\\right) \\)✓XX56.069.713.415.228.248.838.5
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓XX58.271.514.416.026.947.939.2
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓X58.271.315.016.727.449.339.6
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)X✓X54.367.012.213.625.045.336.3
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓✓57.971.116.117.830.251.140.7
", + "bbox": [ + 93, + 416, + 478, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 542, + 251, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 presents the effectiveness of the proposed components: frozen-target text contrastive loss (FCL), refined batch sampling (RB), and refined concatenation scheme (RC). All evaluation results are on the validation splits. All model variants use ViT-L/14 and a projection module $\\phi$ from LinCIR, making the results in row 1 indicative of the original performance of LinCIR. We first compare the impact of the text pairs fed into FCL loss. We compare our design choice $(T_{r + c}, T_t)$ from the generated text triplets with $(T_r, T_r)$ , which is the sole option for constructing a pair given a single conventional caption $T_r$ . The results demonstrate that, on average, using generated triplets (3rd row) is more effective than using original conventional text pairs (2nd row), particularly in the CIRR and CIRCO datasets. In addition, RB (4th row) and RC (6th row) significantly enhance the overall performance, demonstrating the effectiveness of these components. Finally, we measure the impact of using the frozen target textual embedding, denoted as \"Frozen Target\" in the table. Significant performance degradation is observed when the learnable text encoder is used for extracting the embedding of the target caption $T_t$ (5th row) compared to the frozen-target case (4th row), supporting the importance of our design choice.", + "bbox": [ + 89, + 568, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/51ce49fa1ce222417919520fc9b3b2511c0b99d7a07ab348d441162a59cba52a.jpg", + "table_caption": [ + "Table 4. T2I retrieval performance of different text encoders on CIRCO val set. \"Update (pair)\" refers to the setting in the second row of Tab. 3, which uses contrastive learning with the pair $(T_r, T_r)$ . Bold indicates the best result, excluding the first row (oracle case)." + ], + "table_footnote": [], + "table_body": "
QueryText encodermAP@5mAP@10mAP@25
\\(T_t\\)Frozen18.9619.3121.05
\\(T_{r+c}\\)Frozen10.1210.7112.34
\\(T_{r+c}\\)Update (pair)10.5211.1512.72
\\(T_{r+c}\\)RTD15.1215.8017.77
", + "bbox": [ + 555, + 138, + 856, + 215 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analyses on our core motivation", + "text_level": 1, + "bbox": [ + 511, + 224, + 797, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments to validate our core motivation (reducing task discrepancy) and verify that the observed gains stem from it. Other details follow Sec. 4.3.", + "bbox": [ + 511, + 248, + 906, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Can RTD really reduce the task discrepancy of the text encoder? We first quantitatively verify whether RTD indeed reduces the task discrepancy. We first conduct a controlled experiment that measures the text-to-image (T2I) retrieval performance of the text encoder with conditional texts. We retrieve the target images $I_{t}$ with the concatenated text query $T_{r + c}$ or the ideal target caption $T_{t}$ . If our text encoder successfully handles the discrepancy due to the concatenated caption, the text encoder updated by RTD will perform better than the frozen one or \"Update (pair)\", which is updated by contrastive learning with the $(T_{r}, T_{r})$ pair (corresponding to row 2 of Tab. 3). We use the CLIP ViT-L/14 and CIRCO [3] validation set for evaluation. Since the CIRCO dataset only has CIR triplets $(I_{r}, T_{c}, I_{t})$ , we use the BLIP [20] captioner to generate $T_{r}$ and $T_{t}$ corresponding to the $I_{r}$ and $I_{t}$ , respectively. Here, the simple concatenation scheme is applied for the text query $T_{r + c}$ in all cases. Tab. 4 shows that when the text encoder is either frozen or updated with $(T_{r}, T_{r})$ , the retrieval results using the concatenated caption $T_{r + c}$ are significantly worse than those using the target caption $T_{t}$ . It supports the claim that the frozen text encoder suffers from the negative effects of task discrepancy between the pretext and CIR tasks. Moreover, it suggests that simply updating the text encoder with $(T_{r}, T_{r})$ pair fails to reduce this discrepancy. In contrast, the text encoder updated by RTD shows a significant improvement over the frozen text encoder or one updated with $(T_{r}, T_{r})$ , showing that it successfully reduces the task discrepancy.", + "bbox": [ + 511, + 294, + 908, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We additionally measure the average cosine similarity between the composed textual features with the prompt \"a photo of $\\phi (\\psi_V(I_r))$ that $T_{c}$ \" (Fig. 1a) and the target image features (Fig. 1c). The similarity is measured by the LinCIR ViT-L/14 model on the CIRCO validation split. When we use the frozen text encoder $(\\psi_T)$ , the average similarity is 0.1. By changing the text encoder to our updated text encoder $(\\psi_T^{tr})$ , the similarity becomes 0.29 (+0.19). This result shows again that RTD successfully aligns the composed query features using $\\phi$ to the frozen CLIP image features.", + "bbox": [ + 511, + 718, + 908, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Is all the gain of RTD from naive text-encoder updating? To verify that our improvements cannot be achieved solely", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3901", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "by updating the text encoder backbone without considering the task discrepancy, we additionally measure the results of previous methods (Pic2Word and LinCIR) when naively updating the text encoders. Namely, after training $\\phi$ while keeping all other networks frozen as in previous methods, we additionally update the text encoder using the original loss function, while fixing other modules including $\\phi$ . We denote this update rule as \"naive\" in the Tab. 5. Unlike RTD, we observe that just naively updating the text encoder (\"naive\") significantly degrades the performance of the baseline. The results indicate that merely updating the text backbone is not beneficial for CIR; instead, mitigating task discrepancy through RTD is necessary.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb8535e31adb1e09390a1af50f37a0bd6e3345d9dfd326ae372aabc0f9f3112d.jpg", + "table_caption": [ + "Table 5. Impact of the update scheme. Two update schemes are compared: (1) using the original objective from baseline (\"naive\") and (2) using RTD. For a fair comparison, in both schemes, $\\phi$ is updated first and $\\psi_T$ is updated top on the frozen modules. Other details are the same as Tab. 3." + ], + "table_footnote": [], + "table_body": "
CIRR (R)CIRCO (mAP)FIQ (R)Avg
@5@10@10@25@10@50
Pic2Word51.464.48.810.125.344.932.2
+naive19.227.51.31.64.411.210.9
+ RTD56.669.88.89.827.648.936.9
LinCIR54.367.812.714.527.447.736.9
+naive52.766.811.413.026.345.935.5
+ RTD57.971.116.117.830.251.140.7
", + "bbox": [ + 94, + 373, + 478, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Next, instead of using the original objective from baselines, we update the text encoder with conventional contrastive learning using the pair $(T_r, T_r)$ , which corresponds exactly to row 2 in Tab. 3 (row 3 in Tab. 4). This setup can also serve as a proxy for the standalone impact of breaking the alignment of pre-trained knowledge, as noted in [2]. As demonstrated in Sec. 4.3, the large gap (more than 2 points) between RTD (last row in Tab. 3) and this proxy (row 2 in Tab. 3) again shows that our gain cannot be attributed to this simple alternative (naïve text encoder tuning).", + "bbox": [ + 89, + 506, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74c016a39698e76fab097c94a755f2f123fedb0780cb4d21b0a88e3882babda9.jpg", + "image_caption": [ + "Figure 3. Impact of sizes of backbone. The results of RTD combined with Pic2Word [30], SEARLE [3], and LinCIR [15] across different CLIP backbones (ViT-B/32 and ViT-L/14) are shown. Here, the score is the same metric in \"Avg\" in Tab. 3 and other details are the same as Tab. 3. Full results are in the Suppl. B.2." + ], + "image_footnote": [], + "bbox": [ + 93, + 665, + 480, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Compatibility across backbone sizes", + "text_level": 1, + "bbox": [ + 511, + 90, + 826, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the Fig. 3, we observe that the integration of our approach with projection-based CIR methods significantly improves the performance across pioneering projection-based ZS-CIR methods (SEARLE, Pic2Word, and LinCIR) and all backbones (ViT-B/32 and ViT-L/14). For example, regardless of the choice of projection module $\\phi$ and backbones, the minimum performance gain for average scores is greater than 2.8 points. The results of RTD using the larger backbone (ViTG/14) can be found in Tab. 2. More detailed and qualitative results are provided in Suppl. B.3 and Suppl. D, respectively.", + "bbox": [ + 511, + 112, + 906, + 265 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5e7d165879cdac16ead15ff572085d2f4ef69d12bb52ad018af05690234ccac7.jpg", + "table_caption": [ + "Table 6. The effectiveness of different types of text triplets for RTD. \"In-context\" denotes an efficient implementation using in-context learning with LLaMA3-8B, without fine-tuning. Details and examples of each text triplet dataset are summarized in Tab. S1 and Tab. S2, respectively. Other details are the same as Tab. 3." + ], + "table_footnote": [], + "table_body": "
SourceLLMCIRR (R) @5 @10CIRCO (mAP) @10 @25FIQ (R) @10 @50Avg
LinCIR--54.367.812.714.527.447.737.4
+RTDRule-basedX56.770.315.017.030.451.940.2 (+ 2.9)
IP2P [4]✓58.771.615.918.029.650.740.7 (+ 3.3)
Compodiff [14]✓57.971.116.117.830.251.140.7 (+ 3.3)
In-context✓59.371.815.817.529.751.440.9 (+ 3.5)
CoVR [37]✓59.872.615.417.029.650.841.9 (+ 4.5)
CASE [19]✓56.369.311.112.726.647.837.7 (+ 0.3)
", + "bbox": [ + 516, + 343, + 901, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Impact of the text triplet generation strategies", + "text_level": 1, + "bbox": [ + 511, + 446, + 901, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As explained in Sec. 3.1, we evaluate RTD using both 1) publicly available LLM-based text triplets (from IP2P, Compodiff, CoVR, and CASE) along with efficient in-context learning-based text triplets, and 2) LLM-free, rule-based triplets. Tab. 6 shows that RTD consistently improves CIR performance across them (+3.3 for IP2P, +3.3 for Compodiff, +3.5 for in-context learning, +4.5 for CoVR, and +0.31 for CASE, and rule-based triplets achieve 2.9, respectively). We believe this result shows the reproducibility and consistency of RTD, with the rule-based triplets performing comparably to LLM-generated ones, indicating that efficient rule-based triplets are sufficient to achieve strong CIR performance. The marginal improvement in CASE is largely due to the poor quality of text triplets resulting from its construction pipeline that prioritizes CIR triplet quality over text triplet quality, as shown in Tab. S1. Further details can be found in Suppl. A.2, and additional analyses, including data scales related to text triplets, are provided in Suppl. B.5.", + "bbox": [ + 511, + 468, + 908, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented RTD, a novel post-processing approach that can be easily integrated into existing projection-based CIR methods, aimed at reducing task discrepancy of text encoders. Empirical evaluations demonstrate that RTD significantly boosts the performance of existing projection-based CIR methods across diverse datasets and model backbones, competing with or outperforming other resource-intensive CIR methods with much greater efficiency.", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3902", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Research Foundation of Korea (NRF) grant [No. 2021R1A2C2007884, No. RS-2025-02263628], the Institute of Information & communications Technology Planning & Evaluation (IITP) grants [RS-2021-II212068, RS-2022-II220113, RS-2022-II220959, RS-2021-II211343], and the BK21 FOUR Education and Research Program for Future ICT Pioneers (Seoul National University), funded by the Korean government (MSIT).", + "bbox": [ + 89, + 114, + 485, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 250, + 187, + 265 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 3", + "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, 2022. 2, 3, 8", + "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 5, 6, 7, 8", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 5, 8", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 3", + "[6] Junyang Chen and Hanjiang Lai. Pretrain like you inference: Masked tuning improves zero-shot composed image retrieval. arXiv preprint arXiv:2311.07622, 2023. 3, 6", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4", + "[8] Yanzhe Chen, Huasong Zhong, Xiangteng He, Yuxin Peng, Jiahuan Zhou, and Lele Cheng. Fashionern: Enhance-and-refine network for composed fashion image retrieval. In AAAI, 2024. 3", + "[9] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image-caption associations for ms-coco. In ECCV, 2022. 5", + "[10] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “this is my unicorn, fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 4", + "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 3", + "[12] Yongchao Du, Min Wang, Wengang Zhou, Shuping Hui, and Houqiang Li. Image2sentence based asymmetrical zero-shot composed image retrieval. In ICLR, 2024. 3", + "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The" + ], + "bbox": [ + 93, + 273, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3", + "[14] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. TMLR, 2023. 1, 2, 3, 5, 6, 8", + "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoonhoon Kang, and Sangwoo Yun. Language-only efficient training of zero-shot composed image retrieval. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[16] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 3", + "[17] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 3, 6", + "[18] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, 2021. 1, 3", + "[19] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6, 8", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 7", + "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3", + "[22] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4", + "[23] Haoqiang Lin, Haokun Wen, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Fine-grained textual inversion network for zero-shot composed image retrieval. In SIGIR, 2024. 1, 2, 3, 5, 6", + "[24] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 5", + "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5", + "[26] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In ECCV, 2020. 5", + "[27] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pretrained vlms for improving vision-linguistic compositionality. In EMNLP, 2024. 2, 6", + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 5", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3903", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5", + "[30] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 5, 6, 8", + "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Dataset and Benchmark, 2022. 1", + "[32] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5", + "[33] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 1, 3, 6", + "[34] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6", + "[35] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[36] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5", + "[37] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. In AAAI, 2024. 1, 2, 3, 5, 6, 8", + "[38] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2, 5", + "[39] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In SIGIR, 2024. 3, 6", + "[40] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 1, 2, 3, 5, 6" + ], + "bbox": [ + 91, + 90, + 483, + 741 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_model.json b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e4e04ab05c30aa023b2b29f84f7ab242f6d7b78 --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_model.json @@ -0,0 +1,1639 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.13, + 0.844, + 0.177 + ], + "angle": 0, + "content": "An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.861, + 0.223 + ], + "angle": 0, + "content": "Jaeseok Byun\\(^{1*}\\) Seokhyeon Jeong\\(^{1*}\\) Wonjae Kim\\(^{2}\\) Sanghyuk Chun\\(^{2\\dagger}\\) Taesup Moon\\(^{1,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.238, + 0.766, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of ECE, Seoul National University \\(^{2}\\)NAVER AI Lab \\(^{3}\\)ASRI/INMC/IPAI/AIIS, Seoul National University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.342, + 0.485, + 0.751 + ], + "angle": 0, + "content": "Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches. The mainstream Zero-Shot (ZS) CIR methods bypass the need for expensive training CIR triplets by projecting image embeddings into the text token embedding space, forming a composed query for retrieval. However, we highlight an inherent limitation in these projection-based CIR: a task discrepancy of text encoders between the original pre-training task of the encoders (text \\(\\leftrightarrow\\) image) and the target CIR task (image + text \\(\\leftrightarrow\\) image), which potentially negatively impacts CIR performance. To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. Instead, we introduce Reducing Task Discrepancy of Text Encoders (RTD), an efficient text-only post-hoc framework that complements projection-based CIR methods. We devise a novel frozen-target text contrastive learning designed to enhance the capability of the text encoder for CIR. We also propose two key enhancements: (1) a hard negative-based refined batch sampling strategy and (2) a refined concatenation scheme to further mitigate training-inference discrepancy. Integrating RTD into state-of-the-art projection-based methods achieves performance comparable to, or even surpassing, resource-intensive state-of-the-art synthetic CIR triplet-based approaches, only with 23 minutes of additional training on 4 A100 GPUs. Our code is available in https://github.com/jaeseokbyun/RTD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.781, + 0.222, + 0.796 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.486, + 0.853 + ], + "angle": 0, + "content": "Composed Image Retrieval (CIR) aims at retrieving a target image that closely resembles a reference image while reflecting the changes described in a conditioning text. Using a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.908, + 0.599 + ], + "angle": 0, + "content": "query composed of image and text allows users to conduct more precise and flexible searches by specifying the desired modifications to the image through text. Supervised CIR methods [1, 11, 18] have been introduced to fuse information from bi-modal query, using labeled data from the target domain in the form of triplets \\((I_r, T_c, I_t)\\), in which \\(I_r\\) is a reference image, \\(T_c\\) is a conditioning text, and \\(I_t\\) is a target image. However, unlike the typical web-crawled image-text datasets [31], acquiring sufficient triplets for training needs expensive manual human annotations. To overcome the dependency on small-scale and human-verified triplets of the target domain, several works utilize the power of recent generative models. For example, a line of studies [14, 19, 37, 40] uses text-to-image models like IP2P [4] or large-language models (LLM) [5, 35] to synthesize large-scale CIR triplets for training, in place of the target-domain CIR triplets. While these methods achieve strong performance, they are often impractical due to the high computational and memory requirements for utilizing generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Another approach for removing the dependency on the CIR triplets, which often is referred to as projection-based ZS-CIR [3, 15, 23, 30, 33, 34], employs an integrable projection module on top of the pre-trained, frozen, and shared embedding space, such as CLIP [29]. Namely, a projection module \\(\\phi\\), which maps a CLIP image embedding to the CLIP text token embedding space, can be trained by solely using images [3, 30] or texts [15]. During inference, as illustrated in Fig. 1, these methods first project the embedding of the query image to a text token embedding \\([\\\\) ]\\) using the function \\(\\phi\\). This embedding is then combined with the conditioning text \\([T_c]\\) to create the prompt \"a photo of \\([\\\\) ]\\) that \\([T_c]\\)\", which is used as a query for the text-to-image retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The core assumption of the projection-based CIR is that the pre-trained text encoder should be robust enough to combine information from both the projected text token embedding and the conditioning text. However, we argue that this can cause a significant task discrepancy for the pre-trained text encoder between the original image-text alignment pretraining task (of CLIP) and the CIR task. For example, in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.864, + 0.218, + 0.876 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.877, + 0.483, + 0.9 + ], + "angle": 0, + "content": "†Corresponding authors. (E-mail: sanghyukc@princeton.edu, tsmoon@snu.ac.kr)" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3895" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.904, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.311, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 1. The task discrepancy of projection-based ZS-CIR methods between the pre-training task (image-text alignment) and the ZS-CIR task (image-text composition)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.486, + 0.575 + ], + "angle": 0, + "content": "Fig. 1, consider an ideal caption that accurately describes the target image. Since the CLIP text and image encoders are learned through contrastive learning, we can expect that the target image embedding (Fig. 1c) will align well with the embedding (Fig. 1b) of the ideal caption. In contrast, in the projection-based CIR, the text encoder receives a concatenated caption that combines the projected token \\([\\\\) ]$ and the conditioning text, not an ideal caption. However, as noted by Baldrati et al. [2], the text encoder is not typically trained to encode complex textual modifications (e.g., addition or replacement) to the reference image, which are common in conditioning texts. As a result, there is no guarantee that the textual embedding of the concatenated caption (Fig. 1a) closely aligns with the target image embedding (Fig. 1c), which will undermine the retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.585, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To that end, we devise a complementary post-processing approach for existing projection-based CIR methods that reduces the task discrepancy of the text encoder efficiently. Ideally, both image and text backbones would be updated using expensive CIR triplets \\((I_r, T_c, I_t)\\); however, we achieve this using cheap text triplets and only update the text encoder. These text triplets \\((T_r, T_c, T_t)\\) consist of a reference caption \\((T_r, e.g., “a dog catching a frisbee”)\\), a conditioning text \\((T_c, e.g., “change dog to cat”)\\), and a target caption \\((T_t, e.g., “a cat catching a frisbee”)\\), respectively. Given that the reference captions \\(T_r\\) are readily available from conventional caption datasets, \\(T_c\\) and \\(T_t\\) can be automatically generated without any human labor [24, 38] or intensive resources [14, 19, 37, 40]. Using these triplets, we introduce a frozen-target text contrastive learning, in which the text encoder updates the embedding of the concatenated caption \\(T_{r + c}\\) of \\(T_r\\) and \\(T_c\\) (e.g., “a dog catching a frisbee”+ “change the dog to a cat”) to align closely with the fixed embedding of the target caption \\(T_t\\) from the frozen text encoder. We further enhance this text-only training approach with two key components: a batch sampling strategy that ensures hard neg-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.908, + 0.394 + ], + "angle": 0, + "content": "atives in each mini-batch and a refined concatenation scheme for \\( T_{r} \\) and \\( T_{c} \\) designed to reduce the training-inference task discrepancy caused by our text-only training framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.909, + 0.648 + ], + "angle": 0, + "content": "Our experimental results demonstrate that our proposed method, dubbed as RTD (Reducing Task Discrepancy of Text Encoders), consistently and considerably improves the performance in diverse evaluation datasets (CIRR [24], CIRCO [3], FashionIQ [38], COCO object composition [30], and GeneCIS [36]), when integrated with existing projection-based CIR methods (SEARLE [3], Pic2Word [30], LinCIR [15], Context-I2W [34], and FTI4CIR [23]). Consequently, compared to resource-intensive state-of-the-art synthetic CIR triplets-based method like MagicLens [40], RTD achieves comparable results, with less than a \\(1\\%\\) gap in CIRR R@1 and \\(0.1\\%\\) in FashionIQ R@10 when combined with FTI4CIR (ViT-L/14). When using larger or fine-tuned backbones (ViT-G/14 or FSC-CLIP [27]) with LinCIR, RTD achieves superior results than MagicLens in those metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We note that RTD achieves the above results with significantly higher training efficiency in both data generation and training time. Specifically, our rule-based text triplet generation requires less than 10 minutes to construct 1M text triplets \\((570\\times\\) faster than CIR triplet generation from Compodiff [14]). With 4 NVIDIA A100 GPUs, the additional training time for RTD is just 23 minutes for training the ViT-L/14 backbone, making it approximately 10 to \\(100\\times\\) faster than synthetic CIR triplets-based methods. Even with a significantly larger backbone (ViT-G/14), training takes only 2.5 hours, remaining negligible in comparison. Moreover, we investigate the effectiveness of RTD across various text triplet generation strategies and different backbone sizes (ViT-B/32, ViT-L/14, ViT-G/14) and types (FSC-CLIP [27], CLIP [29]), showing its reproducibility and compatibility." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3896" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.358 + ], + "angle": 0, + "content": "Composed Image Retrieval. Unlike supervised CIR methods [1, 2, 11, 18], projection-based CIR methods [3, 12, 15, 23, 30, 33, 34] are built upon the frozen CLIP model, where a projection module \\(\\phi\\) is trained without CIR triplets. Each method employs a different training scheme for different training schemes for \\(\\phi\\) (See Sec. 4.1 for details). Several approaches avoid the need for any training by employing interpolation techniques [16], while others leverage powerful yet resource-intensive models such as LLMs and captioners [17, 39]. While our approach is built upon projection-based CIR methods, its training strategy is closely related to another category of CIR methods [14, 19, 37, 40] that use synthetically generated expensive CIR triplets. However, our method stands out by utilizing text triplets and updating only the text encoder, resulting in more efficient training while maintaining strong performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.487, + 0.784 + ], + "angle": 0, + "content": "Task discrepancy between the CLIP pre-training task and CIR. Combiner [2] updates the text encoder to minimize the gap between the target caption feature and the sum of the reference image and conditioning text features. FashionERN [8] addresses reference image dominance by introducing a separate branch to amplify the impact of the conditioning text. However, both Combiner and FashionERN require expensive CIR triplets \\((I_r, T_c, I_t)\\) for training, whereas our approach uses cheap, automatically generated text-only triplets \\((T_r, T_c, T_t)\\). As another example, Chen and Lai [6] synthesize a triplet of an original image, its caption, and the masked image, treating them as the target image, conditioning text, and reference image, respectively. This approach, however, still has a gap between conditioning text (e.g., \"change dog to cat\") and image caption (e.g., \"a dog catching a frisbee\"); furthermore, it needs the full fine-tuning of the CLIP model, resulting in changing the visual embeddings in the retrieval database. In contrast, RTD directly uses the conditioning texts for training and does not change the target visual encoder, enabling the reuse of pre-extracted CLIP embeddings. Lastly, CIREVL [17] or LDRE [39] reduce task discrepancy by generating descriptive captions of the composed query using a large captioning model [21] and LLM [5]. While effective without any training, they rely on expensive inference steps and require carefully crafted prompts by skilled users. Whereas, RTD is fully automated, human-free, and more efficient during inference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.233, + 0.81 + ], + "angle": 0, + "content": "3. Main Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.3, + 0.835 + ], + "angle": 0, + "content": "3.1. Obtaining text triplets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To address the task discrepancy, we collect cheaply and automatically generated text triplets \\((T_r, T_c, T_t)\\), instead of directly using the expensive CIR triplets \\((I_r, T_c, I_t)\\). Given reference caption \\(T_r\\) from conventional caption datasets," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "conditioning text \\( T_{c} \\) and target text \\( T_{t} \\) can be generated using two strategies: via large language models (LLMs) [4, 14, 19, 37] or, more efficiently, through rule-based templates [14]. We investigate both strategies and demonstrate that RTD consistently improves performance across them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.909, + 0.44 + ], + "angle": 0, + "content": "For the LLM-based strategy, we use the publicly available text triplets by CompoDiff [14], which are employed in our main experiments if not specified. These triplets are generated by taking a given caption \\( T_r \\) as an input of the fine-tuned LLM, whose output predicts the corresponding conditioning text \\( T_c \\) and the target caption \\( T_t \\). Previous works, such as IP2P [4], CoVR [37], and CASE [19], have also explored generating text triplets using LLMs, differing in LLM model types, input data, and fine-tuning strategies. The original purpose of these text triplet generations is to construct CIR triplets \\( (I_r, T_c, I_t) \\), but all these works also release the corresponding text triplets used for their CIR triplet construction. In addition to these publicly available text triplets, we also implement and evaluate an efficient in-context learning generation strategy using LLaMA3-8B [13] without additional fine-tuning. We conduct experiments with all the aforementioned text triplets in Tab. 6 and observe that RTD consistently delivers significant enhancements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.911, + 0.593 + ], + "angle": 0, + "content": "For cheap, rule-based strategy, we can extract a keyword (e.g., \"dog\") from \\( T_{r} \\) (\"a dog catching a frisbee\") and replace it with a randomly chosen keyword (e.g., \"cat\") [14], forming \\( T_{c} \\). The conditioning text is then generated automatically by using pre-defined templates (e.g., \"change [original keyword] to [altered keyword]\"). Our experiments show that this simple rule-based variant performs similarly to the LLM-based one. More explanations and examples of each strategy are provided in Suppl. A.2, and more details of generation costs are in Suppl. C." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.848, + 0.621 + ], + "angle": 0, + "content": "3.2. Frozen-target text contrastive learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Now, we explain our approach to update the text encoder for mitigating the task discrepancy solely with the generated text triplets \\((T_r, T_c, T_t)\\). We first assume that there exists a pretrained projection module \\(\\phi\\) obtained by the projection-based CIR methods [3, 15, 30]. Recall that for a given reference image \\(I_r\\) and conditioning text \\(T_c\\), the final composed feature is generated by passing the text prompt \"a photo of \\(\\phi(\\psi_V(I_r))\\) that \\(T_c\\)\" to the text encoder \\(\\psi_T\\), where \\(\\psi_V\\) is the visual encoder and \\(\\phi\\) is the projection module (See Fig. 1). We aim to update the text encoder \\(\\psi_T\\) to reduce the discrepancy between the pretext task and projection-based CIR task using the text triplets while maintaining \\(\\psi_V\\) and \\(\\phi\\) frozen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Frozen-target text contrastive loss. We apply contrastive learning using a paired caption \\((T_{r + c}, T_t)\\), where \\(T_{r + c}\\) denotes a concatenated caption of reference caption \\(T_r\\) and conditioning caption \\(T_c\\). Namely, we let the representation of the concatenated caption closely approximate that of the target caption. However, solely updating the text encoder" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3897" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.905, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.331, + 0.584, + 0.344 + ], + "angle": 0, + "content": "Figure 2. Overview of RTD." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.484, + 0.54 + ], + "angle": 0, + "content": "while fixing the image encoder can break the alignment between image and text encoders. To prevent the issue, we extract the text embedding of \\( T_{t} \\) using the frozen text encoder \\( \\psi_{T} \\), while the concatenated caption \\( T_{r + c} \\) is extracted from the learnable text encoder \\( \\psi_T^{tr} \\), initialized from \\( \\psi_{T} \\). Here, we assume that as the target caption \\( T_{t} \\) is a standard caption, a text embedding \\( \\psi_{T}(T_{t}) \\), is well-aligned with the frozen image embedding space. Based on this assumption, we freeze the target textual embedding to maintain the pretrained alignment as we learn new relationships. As shown in Sec. 4.3, freezing the target embedding is crucial for fine-tuning with our objective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.707 + ], + "angle": 0, + "content": "Now, we introduce our frozen-target text contrastive loss \\(\\mathcal{L}_{FCL}\\) using two text encoders: a frozen pre-trained text encoder \\(\\psi_T\\) and a learnable text encoder \\(\\psi_T^{tr}\\) which is initialized with \\(\\psi_T\\). Textual latent embeddings \\(\\tilde{t}_{r + c}\\) and \\(t_t\\) are extracted from \\(\\psi_T^{tr}\\) and \\(\\psi_T\\), respectively. Namely, \\(\\tilde{t}_{r + c} = \\psi_T^{tr}(E_w^{tr}(T_{r + c}))\\) and \\(t_t = \\psi_T(E_w(T_t))\\), where \\(E_w\\) is a word embedding layer. We aim to tune \\(\\psi_T^{tr}\\) to minimize the distance between the concatenated textual embedding \\(\\tilde{t}_{r + c}\\) and the target textual embedding \\(t_t\\) while maximizing the distance from other textual embeddings within the batch. We employ a symmetric InfoNCE loss [7, 10], as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.727, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {F C L} = \\frac {1}{B} \\sum_ {k = 1} ^ {B} - \\log \\left(\\frac {e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {k}) / \\tau)}}{\\sum_ {j = 1} ^ {B} e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {j}) / \\tau)} + \\sum_ {j \\neq k} e ^ {(c (t _ {t} ^ {k} , t _ {t} ^ {j}) / \\tau)}}\\right) \\\\ - \\log \\left(\\frac {e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {k}\\right) / \\tau\\right)}}{\\sum_ {j = 1} ^ {B} e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)} + \\sum_ {j \\neq k} e ^ {\\left(c \\left(\\tilde {t} _ {r + c} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)}}\\right) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.885 + ], + "angle": 0, + "content": "in which \\(c(\\cdot, \\cdot)\\) denotes the cosine similarity, \\(B\\) is the batch size, and \\(\\tau\\) is a temperature." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Refined batch sampling strategy for hard negatives. To" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.631 + ], + "angle": 0, + "content": "further enhance the efficacy of updating the text encoder, we devise a simple yet effective batch sampling strategy that incorporates pairs of \\((T_{r + c}, T_t)\\) and \\((T_r, T_r)\\) within the same batch. For example, as presented in Fig. 2, a pair such as \\((T_{r + c}: \\text{“A dog catching a frisbee} + \\text{change dog to cat}\\), \\(T_t: \\text{“A cat catching a frisbee})\\) is sampled along with its corresponding reference pair \\((T_r: \\text{“A dog catching a frisbee}\\), \\(T_r: \\text{“A dog catching a frisbee})\\) in the same batch. This setup ensures that the concatenated text \\(T_{r + c}\\) and its corresponding reference text \\(T_r\\) implicitly act as hard negatives for each other, as their semantics are much more similar \\((T_{r + c}\\) is derived from \\(T_r)\\) than those of other randomly sampled texts in the batch. Explicitly distinguishing the embedding of \\(T_{r + c}\\) from that of \\(T_r\\) conceptually aligns well with CIR task, as it encourages the model to better capture modifications (in \\(T_c\\)). Moreover, we believe including \\((T_r, T_r)\\) pairs in the contrastive learning helps the learnable text encoder \\(\\psi_T^{tr}\\) remain closely aligned with the pre-trained encoder \\(\\psi_T\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Refined concatenation of reference and conditioning texts. As we use \"a photo of \\([\\\\) ]$ that \\([T_c]\\)\" for inference, a naive concatenation strategy also can suffer from training-inference task discrepancy. To tackle this issue, instead of simply concatenating the \\( T_r \\) and \\( T_c \\), we use the prompt \"a photo of \\([\\\\) ]$ that \\([T_c]\\)\" for updating the text encoder, where \\([\\\\) ]$ is obtained by the reference caption \\( T_r \\) with the projection module \\( \\phi \\). Instead of obtaining a pseudo-word token with latent image embedding \\( v \\), we utilize a textual latent embedding from the reference caption \\( T_r \\), i.e., \\( \\phi (t_r) \\). However, Gu et al. [15] showed that naively replacing the image encoder with the text encoder for the input of \\( \\phi \\) will suffer from the modality gap [22], a phenomenon where text and image embeddings have a gap between them. We tackle this issue by injecting random noise into the textual token representation before it is processed by \\( \\phi \\), following Gu et al. [15]. More analyses on variations of noise are in Suppl. B.10." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3898" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Fig. 2 illustrates the overview of RTD. We use CLIP backbone and pre-trained projection module \\(\\phi\\) produced by the existing projection-based CIR methods. The text encoder is trained using the proposed loss function (Eq. (1)) while applying the refined batch sampling and concatenation scheme. During inference, the procedure mirrors that of existing projection-based CIR methods, except we utilize the updated text encoder \\(\\psi_T^{tr}\\) instead of the frozen one \\(\\psi_T\\). Note that our method only updates the text encoder while the image encoder and the projection module are frozen." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.244, + 0.486, + 0.502 + ], + "angle": 0, + "content": "Remark 1: Low data acquisition cost. The text triplets we use can be obtained at a significantly lower cost than CIR triplets. First, the text triplet generation process requires only an easily obtainable caption dataset for generation, whereas other approaches that generate CIR triplets [19, 37, 40] necessitate image or video datasets along with an additional collection phase for semantically similar images or videos. Second, text triplet generation can avoid resource-intensive text-to-image generation [4, 14], making it \\(15 \\times\\) faster than the CIR triplet generation process used in CompoDiff [14]. Furthermore, if we choose the rule-based text triplet generation strategy, the process becomes \\(570 \\times\\) faster than the full CIR triplet generation with images; 1M text triplets generation takes just 0.1 hours. Lastly, the storage for the generated text triplets takes up only 100MB, whereas storing a similar quantity of images requires significantly more space (e.g., 400GB for CC3M [32]). More details can be found in Suppl. C." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.511, + 0.485, + 0.594 + ], + "angle": 0, + "content": "Table 1. Training and inference time comparisons. Purple denotes the additional resource required for RTD. Inference time is measured on a single A100 GPU for a single image. Note that the training hours for SEARLE, LinCIR for 8 A100 GPUs [15] and their RTD variants are trained on 4 A100 GPUs. CompoDiff is trained on 128 A100 and MagicLens is trained on 128 TPUs." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.597, + 0.479, + 0.729 + ], + "angle": 0, + "content": "
MethodCLIP BackboneTraining (h)Inference (s)
SEARLEL/144.2h0.02s
SEARLE (+ RTD)L/144.58h (+0.38h)0.02s
LinCIRL/140.5h0.02s
LinCIR (+ RTD)L/140.88h (+0.38h)0.02s
LinCIRG/140.8h0.05s
LinCIR (+ RTD)G/143.3h (+2.5h)0.05s
CompoDiffL/14231h0.12s
MagicLensL/146hunknown
CIRVLL/14-3.02s
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Remark 2: Training and inference efficiency. As reported in Tab. 1, the additional training cost of RTD is small; 0.38 hours for ViT-L/14 on 4 A100 GPUs. In contrast, CompoDiff and MagicLens require 231 hours on 128 A100 GPUs and 6 hours on 128 TPUs, respectively, making the cost of RTD negligible. Even with a larger ViT-G/14 backbone, training RTD takes only 2.5 hours on 4 A100 GPUs, still faster than synthetic CIR triplet-based methods and even comparable to SEARLE in ViT-L/14 (4.2h). This training efficiency mainly stems from the advantage of text-only training. As highlighted by Gu et al. [15], the training complexity for the text" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.304 + ], + "angle": 0, + "content": "encoder is much lower than that for the visual encoder due to the relatively short token lengths of texts (\\(\\approx 12\\)) compared to images (256). The average inference time of the CLIP ViT-L/14 image encoder is \\(\\times 3.5\\) times slower than that of the text encoder. Another source of efficiency is that RTD requires relatively few iterations (at most 2000, and significantly fewer in most cases), as the text encoders are already pre-trained and only require minor adjustments to learn the relationships between text triplets. Moreover, in Suppl. B.4, we present a more efficient implementation by selectively updating only a few layers of the text encoder. Along with efficient training, RTD retains the fast inference speed of the original projection-based methods, remaining \\(150\\times\\) faster than training-free CIREVL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.321, + 0.646, + 0.338 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.702, + 0.364 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.567 + ], + "angle": 0, + "content": "Implementation details. We use the AdamW optimizer [25] with a weight decay of 0.01. The learning rate is set to \\(10^{-5}\\), with a batch size of 512. We select the text encoder model with the best zero-shot CIRR [24] dev R@1 score for evaluating RTD. We mainly use the visual and textual encoders of the CLIP ViT-L/14 [29] as our backbone. Unless otherwise noted, we use LLM-based 2.5M text triplets provided by CompoDiff [14] for the training. We set the \\(\\tau\\) as 0.07 in Eq. (1) and scale the standard deviation of Gaussian distribution as 0.5 for the noise injection. More results on various noise distributions can be found in the Suppl. B.10. All experiments were conducted using four NVIDIA A100 GPUs with Python 3.8 and Pytorch [28]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Evaluation datasets and metrics. We compare projection-based CIR methods on five benchmark datasets: CIRR [24], CIRCO [3], FashionIQ [38] (FIQ), COCO object composition [30], and GeneCIS [36]. Details of each dataset are in the Suppl. A.1. For CIRR, FIQ, COCO, and GeneCIS, we have reported their recall scores at the top K retrieval results \\((\\mathbb{R}@\\mathbb{K})\\). Since the CIRCO dataset includes multiple positive images for each query, we use a ranking-based metric; mean Average Precision scores at the top K results \\((\\mathrm{mAP}@\\mathrm{K})\\) [9, 26]. For the main results, we compare the results on the FIQ validation split, as well as the test sets of CIRR and CIRCO. For the ablation studies and analyses, the validation splits of these three datasets are utilized. GeneCIS and COCO object composition results are in the Suppl. B.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Baselines. We evaluate the effect of our method when combined with publicly available projection-based ZS-CIR methods: Pic2Word [30], SEARLE [3], LinCIR [15], Context-I2W [34], and FTI4CIR [23]. All these methods share a similar core concept shown in Fig. 1, but use different training schemes. Pic2Word [30] optimizes contrastive loss between the image embedding and its projected text embedding of \"a photo of [\\$\" to obtain the projection module \\(\\phi\\). Similarly," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3899" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison with other baselines. Note that this comparison is not entirely fair due to differences in backbone models and training data. Purple denotes the performance gain from our method, while red and blue highlight the best and second-best scores, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.121, + 0.902, + 0.433 + ], + "angle": 0, + "content": "
MethodBackboneCIRRCIRCOFashionIQ
R@1R@5R@10mAP@5mAP@10mAP@25R@10R@50
Pic2Word [30]CLIP ViT-L/1424.251.564.18.39.110.125.344.9
+RTD27.9 (+3.6)56.2 (+4.8)68.5 (+4.4)9.1 (+0.9)9.6 (+0.5)10.7 (+0.6)27.6 (+2.3)48.9 (+4.0)
SEARLE [3]CLIP ViT-L/1424.952.365.711.612.714.325.045.3
+RTD26.6 (+1.7)56.2 (+3.9)69.0 (+3.3)16.5 (+4.9)17.9 (+5.2)19.8 (+5.4)29.3 (+4.4)50.7 (+5.4)
Context-I2W [34]CLIP ViT-L/1425.655.468.6---27.949.1
+RTD29.2 (+4.6)58.4 (+3.0)70.5 (+1.9)---28.1 (+0.2)49.5 (+0.4)
FTI4CIR [23]CLIP ViT-L/1425.955.667.715.116.318.129.450.9
+RTD29.1 (+3.2)58.8 (+3.2)70.4 (+2.7)16.2 (+1.1)17.4 (+1.1)19.4 (+1.3)30.6 (+1.2)51.7 (+0.8)
LinCIR [15]CLIP ViT-L/1423.852.966.513.014.115.827.447.7
+RTD26.6 (+2.9)56.2 (+3.3)69.0 (+2.5)17.1 (+4.1)18.1 (+4.0)20.1 (+4.3)30.2 (+2.8)51.1 (+3.4)
LinCIR [15]FSC-CLIP [27] ViT-L/1433.663.174.514.215.417.028.349.5
+RTD35.9 (+2.3)67.5 (+4.4)78.1 (+3.6)18.3 (+4.1)19.7 (+4.3)21.3 (+4.3)31.2 (+2.9)52.3 (+2.8)
KEDs [33]CLIP ViT-L/1426.454.867.2---26.847.9
CIREVL [17]CLIP ViT-L/1424.552.364.918.619.020.928.648.6
LDRE [39]CLIP ViT-L/1426.555.667.523.424.026.428.550.5
MT-CIR [6]CLIP ViT-L/1425.554.667.610.411.613.035.457.4
MagicLens [40]CLIP ViT-L/1430.161.774.429.630.833.430.752.5
Compodiff [14]CLIP ViT-L/1426.755.072.612.613.415.836.048.6
CoVR [37]BLIP ViT-L/1439.368.278.9---27.744.6
CASE [19]BLIP ViT-L/1435.465.878.5-----
LinCIR [15]CLIP ViT-G/1434.964.576.120.621.924.144.565.5
+RTD36.3 (+1.4)67.5 (+3.0)78.3 (+2.2)21.1 (+0.5)22.3 (+0.4)24.5 (+0.4)46.2 (+1.7)67.3 (+1.8)
Compodiff [17]CLIP ViT-G/1434.764.375.115.317.719.439.051.7
CIREVL [17]CLIP ViT-G/1426.755.174.526.827.630.032.252.4
LDRE [39]CLIP ViT-G/1436.266.477.331.132.235.032.555.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.486, + 0.641 + ], + "angle": 0, + "content": "SEARLE [3] employs a two-stage approach, starting with an optimization-based textual inversion phase followed by a distillation phase for the projection module \\(\\phi\\). LinCIR [15] introduces a language-only self-supervised task based on keyword token replacement, where the model encourages the original text embedding and the replaced text embedding (keyword tokens are replaced with the projection of the original text embedding via \\(\\phi\\)) to be aligned. Context-I2W [34] refines the projection module by selecting relevant visual information [34]. Lastly, FTI4CIR [23] separately maps images into subject- and attribute-oriented pseudo-word tokens [23]. Details on how RTD is combined with them can be found in Suppl. A.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.487, + 0.902 + ], + "angle": 0, + "content": "We train these methods with the CLIP ViT-L/14 in our main experiments. To further verify the compatibility of RTD, we also report results for the ViT-B/32 backbone with SEARLE, Pic2Word, and LinCIR. Moreover, we conduct experiments with a larger backbone (ViT-G/14) and a fine-tuned CLIP model for enhanced compositionality (FSC-CLIP [27]) for LinCIR, enabled by its fast training capability. We use the publicly available pre-trained model for SEARLE (ViT-B/32, ViT-L/14), Pic2Word (ViT-L/14), Context-I2W (ViT-L/14), and FTI4CIR (ViT-L/14). Otherwise, we reproduce the results using the official implementation. When reproducing, we adhere to the same settings in the original papers. For example, we select the final last epoch model for the Pic2Word ViT-B/32 model and choose the model based on the best zero-shot CIRR dev R@1 score for LinCIR. We additionally compare our method with a diverse set of CIR approaches, including another projection-based" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.909, + 0.52 + ], + "angle": 0, + "content": "ZS-CIR method (KEDs [33]), another attempt to address task discrepancy (MT-CIR [6]), approaches leveraging synthetically generated CIR triplets (CoVR [37], CASE [19], MagicLens [40] and Compodiff [14]), and the training-free methods (CIREvL [17], LDRE [39])." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.649, + 0.545 + ], + "angle": 0, + "content": "4.2. Main results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.674 + ], + "angle": 0, + "content": "Tab. 2 shows the overview of comparison results with state-of-the-art CIR methods. First, we assess the impact of integrating RTD with existing projection-based CIR methods (SEARLE, Pic2Word, LinCIR, Context-I2W, and FTI4CIR) across CIRR, CIRCO, and FashionIQ. RTD consistently boosts performance, yielding an average improvement of over 2.78 points. This trend also holds for the GeneCIS and COCO object composition datasets, as detailed in Suppl. B.1." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Second, we compare the integration of RTD with those leveraging synthetically generated CIR triplets (CoVR, CASE, MagicLens, and Compodiff). We focus on this comparison to ensure fairness, as leveraging text triplets and updating the textual backbone of RTD may not be fully aligned with the standard projection-based CIR setting. We observe that RTD delivers competitive or superior performance compared to these resource-intensive synthetic CIR triplet-based methods, while being significantly more efficient; over \\(10 - 100 \\times\\) faster, as noted in Remark-2 of Sec. 3. For example, with the same CLIP ViT-L/14 backbone, LinCIR + RTD outperforms MagicLens and CoVR in FashionIQ R@10 and R@50 while FTI4CIR + RTD or Context-I2W + RTD achieve performance on par with or exceeding that of MagicLens, MT-CIR, and Compodiff in CIRR metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3900" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.212 + ], + "angle": 0, + "content": "When combined with larger (ViT-G/14) or fine-tuned (FSC-CLIP) backbones, RTD achieves the best or second-best result on the CIRR and FashionIQ benchmarks. We believe this flexibility of using different backbones underscores the benefit of efficient text-only training of RTD, inherited by LinCIR [15]. We reemphasize that even with a larger backbone like ViT-G/14, RTD remains much faster in training than other synthetic triplet-based methods, as shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.215, + 0.485, + 0.307 + ], + "angle": 0, + "content": "Lastly, across both backbones (ViT-L/14, G/14), LinCIR + RTD mostly outperforms training-free CIREVL and LDRE, except on CIRCO metrics. As shown in Tab. 1, CIREVL is significantly slower \\((150\\times)\\) than projection-based methods (including RTD), while LDRE incurs even larger inference time due to its ensemble-based strategy." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.319, + 0.485, + 0.415 + ], + "angle": 0, + "content": "Table 3. Ablation study. Unlike in Tab. 2, for ablation studies and analyses, validation splits of three CIR datasets are used for evaluation. We measure the impact of FCL loss (Eq. (1)), refined batch sampling (RB), and refined concatenation scheme (RC). All models are based on LinCIR ViT-L/14. The first row denotes the vanilla LinCIR without RTD. \"Avg\" denotes the average of all reported metrics. Bold indicates the best result." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.417, + 0.48, + 0.515 + ], + "angle": 0, + "content": "
TextFCL Frozen TargetRBRCCIRR (R) @5@10CIRCO (mAP) @10@25FIQ (R) @10@50Avg
--XX54.367.812.714.527.447.737.4
\\( \\left( {{T}_{r},{T}_{r}}\\right) \\)✓XX56.069.713.415.228.248.838.5
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓XX58.271.514.416.026.947.939.2
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓X58.271.315.016.727.449.339.6
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)X✓X54.367.012.213.625.045.336.3
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓✓57.971.116.117.830.251.140.7
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.253, + 0.558 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Tab. 3 presents the effectiveness of the proposed components: frozen-target text contrastive loss (FCL), refined batch sampling (RB), and refined concatenation scheme (RC). All evaluation results are on the validation splits. All model variants use ViT-L/14 and a projection module \\(\\phi\\) from LinCIR, making the results in row 1 indicative of the original performance of LinCIR. We first compare the impact of the text pairs fed into FCL loss. We compare our design choice \\((T_{r + c}, T_t)\\) from the generated text triplets with \\((T_r, T_r)\\), which is the sole option for constructing a pair given a single conventional caption \\(T_r\\). The results demonstrate that, on average, using generated triplets (3rd row) is more effective than using original conventional text pairs (2nd row), particularly in the CIRR and CIRCO datasets. In addition, RB (4th row) and RC (6th row) significantly enhance the overall performance, demonstrating the effectiveness of these components. Finally, we measure the impact of using the frozen target textual embedding, denoted as \"Frozen Target\" in the table. Significant performance degradation is observed when the learnable text encoder is used for extracting the embedding of the target caption \\(T_t\\) (5th row) compared to the frozen-target case (4th row), supporting the importance of our design choice." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.083, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Table 4. T2I retrieval performance of different text encoders on CIRCO val set. \"Update (pair)\" refers to the setting in the second row of Tab. 3, which uses contrastive learning with the pair \\((T_r, T_r)\\). Bold indicates the best result, excluding the first row (oracle case)." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.14, + 0.857, + 0.217 + ], + "angle": 0, + "content": "
QueryText encodermAP@5mAP@10mAP@25
\\(T_t\\)Frozen18.9619.3121.05
\\(T_{r+c}\\)Frozen10.1210.7112.34
\\(T_{r+c}\\)Update (pair)10.5211.1512.72
\\(T_{r+c}\\)RTD15.1215.8017.77
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.798, + 0.24 + ], + "angle": 0, + "content": "4.4. Analyses on our core motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.249, + 0.907, + 0.293 + ], + "angle": 0, + "content": "We conduct experiments to validate our core motivation (reducing task discrepancy) and verify that the observed gains stem from it. Other details follow Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.909, + 0.717 + ], + "angle": 0, + "content": "Can RTD really reduce the task discrepancy of the text encoder? We first quantitatively verify whether RTD indeed reduces the task discrepancy. We first conduct a controlled experiment that measures the text-to-image (T2I) retrieval performance of the text encoder with conditional texts. We retrieve the target images \\( I_{t} \\) with the concatenated text query \\( T_{r + c} \\) or the ideal target caption \\( T_{t} \\). If our text encoder successfully handles the discrepancy due to the concatenated caption, the text encoder updated by RTD will perform better than the frozen one or \"Update (pair)\", which is updated by contrastive learning with the \\( (T_{r}, T_{r}) \\) pair (corresponding to row 2 of Tab. 3). We use the CLIP ViT-L/14 and CIRCO [3] validation set for evaluation. Since the CIRCO dataset only has CIR triplets \\( (I_{r}, T_{c}, I_{t}) \\), we use the BLIP [20] captioner to generate \\( T_{r} \\) and \\( T_{t} \\) corresponding to the \\( I_{r} \\) and \\( I_{t} \\), respectively. Here, the simple concatenation scheme is applied for the text query \\( T_{r + c} \\) in all cases. Tab. 4 shows that when the text encoder is either frozen or updated with \\( (T_{r}, T_{r}) \\), the retrieval results using the concatenated caption \\( T_{r + c} \\) are significantly worse than those using the target caption \\( T_{t} \\). It supports the claim that the frozen text encoder suffers from the negative effects of task discrepancy between the pretext and CIR tasks. Moreover, it suggests that simply updating the text encoder with \\( (T_{r}, T_{r}) \\) pair fails to reduce this discrepancy. In contrast, the text encoder updated by RTD shows a significant improvement over the frozen text encoder or one updated with \\( (T_{r}, T_{r}) \\), showing that it successfully reduces the task discrepancy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.87 + ], + "angle": 0, + "content": "We additionally measure the average cosine similarity between the composed textual features with the prompt \"a photo of \\(\\phi (\\psi_V(I_r))\\) that \\(T_{c}\\)\" (Fig. 1a) and the target image features (Fig. 1c). The similarity is measured by the LinCIR ViT-L/14 model on the CIRCO validation split. When we use the frozen text encoder \\((\\psi_T)\\), the average similarity is 0.1. By changing the text encoder to our updated text encoder \\((\\psi_T^{tr})\\), the similarity becomes 0.29 (+0.19). This result shows again that RTD successfully aligns the composed query features using \\(\\phi\\) to the frozen CLIP image features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Is all the gain of RTD from naive text-encoder updating? To verify that our improvements cannot be achieved solely" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3901" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.289 + ], + "angle": 0, + "content": "by updating the text encoder backbone without considering the task discrepancy, we additionally measure the results of previous methods (Pic2Word and LinCIR) when naively updating the text encoders. Namely, after training \\(\\phi\\) while keeping all other networks frozen as in previous methods, we additionally update the text encoder using the original loss function, while fixing other modules including \\(\\phi\\). We denote this update rule as \"naive\" in the Tab. 5. Unlike RTD, we observe that just naively updating the text encoder (\"naive\") significantly degrades the performance of the baseline. The results indicate that merely updating the text backbone is not beneficial for CIR; instead, mitigating task discrepancy through RTD is necessary." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.37 + ], + "angle": 0, + "content": "Table 5. Impact of the update scheme. Two update schemes are compared: (1) using the original objective from baseline (\"naive\") and (2) using RTD. For a fair comparison, in both schemes, \\( \\phi \\) is updated first and \\( \\psi_T \\) is updated top on the frozen modules. Other details are the same as Tab. 3." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.374, + 0.48, + 0.496 + ], + "angle": 0, + "content": "
CIRR (R)CIRCO (mAP)FIQ (R)Avg
@5@10@10@25@10@50
Pic2Word51.464.48.810.125.344.932.2
+naive19.227.51.31.64.411.210.9
+ RTD56.669.88.89.827.648.936.9
LinCIR54.367.812.714.527.447.736.9
+naive52.766.811.413.026.345.935.5
+ RTD57.971.116.117.830.251.140.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Next, instead of using the original objective from baselines, we update the text encoder with conventional contrastive learning using the pair \\((T_r, T_r)\\), which corresponds exactly to row 2 in Tab. 3 (row 3 in Tab. 4). This setup can also serve as a proxy for the standalone impact of breaking the alignment of pre-trained knowledge, as noted in [2]. As demonstrated in Sec. 4.3, the large gap (more than 2 points) between RTD (last row in Tab. 3) and this proxy (row 2 in Tab. 3) again shows that our gain cannot be attributed to this simple alternative (naïve text encoder tuning)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.666, + 0.482, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Figure 3. Impact of sizes of backbone. The results of RTD combined with Pic2Word [30], SEARLE [3], and LinCIR [15] across different CLIP backbones (ViT-B/32 and ViT-L/14) are shown. Here, the score is the same metric in \"Avg\" in Tab. 3 and other details are the same as Tab. 3. Full results are in the Suppl. B.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.828, + 0.108 + ], + "angle": 0, + "content": "4.5. Compatibility across backbone sizes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.908, + 0.266 + ], + "angle": 0, + "content": "In the Fig. 3, we observe that the integration of our approach with projection-based CIR methods significantly improves the performance across pioneering projection-based ZS-CIR methods (SEARLE, Pic2Word, and LinCIR) and all backbones (ViT-B/32 and ViT-L/14). For example, regardless of the choice of projection module \\(\\phi\\) and backbones, the minimum performance gain for average scores is greater than 2.8 points. The results of RTD using the larger backbone (ViTG/14) can be found in Tab. 2. More detailed and qualitative results are provided in Suppl. B.3 and Suppl. D, respectively." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.272, + 0.909, + 0.342 + ], + "angle": 0, + "content": "Table 6. The effectiveness of different types of text triplets for RTD. \"In-context\" denotes an efficient implementation using in-context learning with LLaMA3-8B, without fine-tuning. Details and examples of each text triplet dataset are summarized in Tab. S1 and Tab. S2, respectively. Other details are the same as Tab. 3." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.344, + 0.902, + 0.445 + ], + "angle": 0, + "content": "
SourceLLMCIRR (R) @5 @10CIRCO (mAP) @10 @25FIQ (R) @10 @50Avg
LinCIR--54.367.812.714.527.447.737.4
+RTDRule-basedX56.770.315.017.030.451.940.2 (+ 2.9)
IP2P [4]✓58.771.615.918.029.650.740.7 (+ 3.3)
Compodiff [14]✓57.971.116.117.830.251.140.7 (+ 3.3)
In-context✓59.371.815.817.529.751.440.9 (+ 3.5)
CoVR [37]✓59.872.615.417.029.650.841.9 (+ 4.5)
CASE [19]✓56.369.311.112.726.647.837.7 (+ 0.3)
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.448, + 0.902, + 0.464 + ], + "angle": 0, + "content": "4.6. Impact of the text triplet generation strategies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.742 + ], + "angle": 0, + "content": "As explained in Sec. 3.1, we evaluate RTD using both 1) publicly available LLM-based text triplets (from IP2P, Compodiff, CoVR, and CASE) along with efficient in-context learning-based text triplets, and 2) LLM-free, rule-based triplets. Tab. 6 shows that RTD consistently improves CIR performance across them (+3.3 for IP2P, +3.3 for Compodiff, +3.5 for in-context learning, +4.5 for CoVR, and +0.31 for CASE, and rule-based triplets achieve 2.9, respectively). We believe this result shows the reproducibility and consistency of RTD, with the rule-based triplets performing comparably to LLM-generated ones, indicating that efficient rule-based triplets are sufficient to achieve strong CIR performance. The marginal improvement in CASE is largely due to the poor quality of text triplets resulting from its construction pipeline that prioritizes CIR triplet quality over text triplet quality, as shown in Tab. S1. Further details can be found in Suppl. A.2, and additional analyses, including data scales related to text triplets, are provided in Suppl. B.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.634, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We presented RTD, a novel post-processing approach that can be easily integrated into existing projection-based CIR methods, aimed at reducing task discrepancy of text encoders. Empirical evaluations demonstrate that RTD significantly boosts the performance of existing projection-based CIR methods across diverse datasets and model backbones, competing with or outperforming other resource-intensive CIR methods with much greater efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3902" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.237 + ], + "angle": 0, + "content": "This work was supported in part by National Research Foundation of Korea (NRF) grant [No. 2021R1A2C2007884, No. RS-2025-02263628], the Institute of Information & communications Technology Planning & Evaluation (IITP) grants [RS-2021-II212068, RS-2022-II220113, RS-2022-II220959, RS-2021-II211343], and the BK21 FOUR Education and Research Program for Future ICT Pioneers (Seoul National University), funded by the Korean government (MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.25, + 0.188, + 0.266 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Junyang Chen and Hanjiang Lai. Pretrain like you inference: Masked tuning improves zero-shot composed image retrieval. arXiv preprint arXiv:2311.07622, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] Yanzhe Chen, Huasong Zhong, Xiangteng He, Yuxin Peng, Jiahuan Zhou, and Lele Cheng. Fashionern: Enhance-and-refine network for composed fashion image retrieval. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[9] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image-caption associations for ms-coco. In ECCV, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “this is my unicorn, fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Yongchao Du, Min Wang, Wengang Zhou, Shuping Hui, and Houqiang Li. Image2sentence based asymmetrical zero-shot composed image retrieval. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.274, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.119 + ], + "angle": 0, + "content": "llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[14] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. TMLR, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoonhoon Kang, and Sangwoo Yun. Language-only efficient training of zero-shot composed image retrieval. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.908, + 0.288 + ], + "angle": 0, + "content": "[16] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.333 + ], + "angle": 0, + "content": "[17] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[18] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[19] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.461 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[22] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.907, + 0.63 + ], + "angle": 0, + "content": "[23] Haoqiang Lin, Haokun Wen, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Fine-grained textual inversion network for zero-shot composed image retrieval. In SIGIR, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[24] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[26] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[27] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pretrained vlms for improving vision-linguistic compositionality. In EMNLP, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3903" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[30] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Dataset and Benchmark, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[32] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[33] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.481, + 0.417 + ], + "angle": 0, + "content": "[34] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[35] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.529 + ], + "angle": 0, + "content": "[36] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[37] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. In AAAI, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[38] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[39] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In SIGIR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.742 + ], + "angle": 0, + "content": "[40] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.742 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3904" + } + ] +] \ No newline at end of file diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_origin.pdf b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..638eeafd641ac4b44b6e9e557f4f7063a6e4dec6 --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/a5a58fbf-87d0-4c6b-8e78-509d0d6c7143_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a10d3a13793bed634dbc22bdb5b7be67825d4bd980391bccfea697e1e76f0d80 +size 977331 diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/full.md b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5d1e0e18b318abb599033da52bd35428bb19f6e0 --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/full.md @@ -0,0 +1,217 @@ +# An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval + +Jaeseok Byun $^{1*}$ Seokhyeon Jeong $^{1*}$ Wonjae Kim $^{2}$ Sanghyuk Chun $^{2\dagger}$ Taesup Moon $^{1,3\dagger}$ + +$^{1}$ Department of ECE, Seoul National University $^{2}$ NAVER AI Lab $^{3}$ ASRI/INMC/IPAI/AIIS, Seoul National University + +# Abstract + +Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches. The mainstream Zero-Shot (ZS) CIR methods bypass the need for expensive training CIR triplets by projecting image embeddings into the text token embedding space, forming a composed query for retrieval. However, we highlight an inherent limitation in these projection-based CIR: a task discrepancy of text encoders between the original pre-training task of the encoders (text $\leftrightarrow$ image) and the target CIR task (image + text $\leftrightarrow$ image), which potentially negatively impacts CIR performance. To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. Instead, we introduce Reducing Task Discrepancy of Text Encoders (RTD), an efficient text-only post-hoc framework that complements projection-based CIR methods. We devise a novel frozen-target text contrastive learning designed to enhance the capability of the text encoder for CIR. We also propose two key enhancements: (1) a hard negative-based refined batch sampling strategy and (2) a refined concatenation scheme to further mitigate training-inference discrepancy. Integrating RTD into state-of-the-art projection-based methods achieves performance comparable to, or even surpassing, resource-intensive state-of-the-art synthetic CIR triplet-based approaches, only with 23 minutes of additional training on 4 A100 GPUs. Our code is available in https://github.com/jaeseokbyun/RTD. + +# 1. Introduction + +Composed Image Retrieval (CIR) aims at retrieving a target image that closely resembles a reference image while reflecting the changes described in a conditioning text. Using a + +query composed of image and text allows users to conduct more precise and flexible searches by specifying the desired modifications to the image through text. Supervised CIR methods [1, 11, 18] have been introduced to fuse information from bi-modal query, using labeled data from the target domain in the form of triplets $(I_r, T_c, I_t)$ , in which $I_r$ is a reference image, $T_c$ is a conditioning text, and $I_t$ is a target image. However, unlike the typical web-crawled image-text datasets [31], acquiring sufficient triplets for training needs expensive manual human annotations. To overcome the dependency on small-scale and human-verified triplets of the target domain, several works utilize the power of recent generative models. For example, a line of studies [14, 19, 37, 40] uses text-to-image models like IP2P [4] or large-language models (LLM) [5, 35] to synthesize large-scale CIR triplets for training, in place of the target-domain CIR triplets. While these methods achieve strong performance, they are often impractical due to the high computational and memory requirements for utilizing generative models. + +Another approach for removing the dependency on the CIR triplets, which often is referred to as projection-based ZS-CIR [3, 15, 23, 30, 33, 34], employs an integrable projection module on top of the pre-trained, frozen, and shared embedding space, such as CLIP [29]. Namely, a projection module \(\phi\), which maps a CLIP image embedding to the CLIP text token embedding space, can be trained by solely using images [3, 30] or texts [15]. During inference, as illustrated in Fig. 1, these methods first project the embedding of the query image to a text token embedding \([\\) ]\) using the function \(\phi\). This embedding is then combined with the conditioning text \([T_c]\) to create the prompt "a photo of \([\\) ]\) that \([T_c]\)", which is used as a query for the text-to-image retrieval. + +The core assumption of the projection-based CIR is that the pre-trained text encoder should be robust enough to combine information from both the projected text token embedding and the conditioning text. However, we argue that this can cause a significant task discrepancy for the pre-trained text encoder between the original image-text alignment pretraining task (of CLIP) and the CIR task. For example, in + +![](images/a14d8142e868fc30bd171bcafe84ad2a7d31792c340085672451184cfcd03fc2.jpg) +Figure 1. The task discrepancy of projection-based ZS-CIR methods between the pre-training task (image-text alignment) and the ZS-CIR task (image-text composition). + +Fig. 1, consider an ideal caption that accurately describes the target image. Since the CLIP text and image encoders are learned through contrastive learning, we can expect that the target image embedding (Fig. 1c) will align well with the embedding (Fig. 1b) of the ideal caption. In contrast, in the projection-based CIR, the text encoder receives a concatenated caption that combines the projected token $[\$ ]$ and the conditioning text, not an ideal caption. However, as noted by Baldrati et al. [2], the text encoder is not typically trained to encode complex textual modifications (e.g., addition or replacement) to the reference image, which are common in conditioning texts. As a result, there is no guarantee that the textual embedding of the concatenated caption (Fig. 1a) closely aligns with the target image embedding (Fig. 1c), which will undermine the retrieval performance. + +To that end, we devise a complementary post-processing approach for existing projection-based CIR methods that reduces the task discrepancy of the text encoder efficiently. Ideally, both image and text backbones would be updated using expensive CIR triplets $(I_r, T_c, I_t)$ ; however, we achieve this using cheap text triplets and only update the text encoder. These text triplets $(T_r, T_c, T_t)$ consist of a reference caption $(T_r, e.g., “a dog catching a frisbee”)$ , a conditioning text $(T_c, e.g., “change dog to cat”)$ , and a target caption $(T_t, e.g., “a cat catching a frisbee”)$ , respectively. Given that the reference captions $T_r$ are readily available from conventional caption datasets, $T_c$ and $T_t$ can be automatically generated without any human labor [24, 38] or intensive resources [14, 19, 37, 40]. Using these triplets, we introduce a frozen-target text contrastive learning, in which the text encoder updates the embedding of the concatenated caption $T_{r + c}$ of $T_r$ and $T_c$ (e.g., “a dog catching a frisbee”+ “change the dog to a cat”) to align closely with the fixed embedding of the target caption $T_t$ from the frozen text encoder. We further enhance this text-only training approach with two key components: a batch sampling strategy that ensures hard neg- + +atives in each mini-batch and a refined concatenation scheme for $T_{r}$ and $T_{c}$ designed to reduce the training-inference task discrepancy caused by our text-only training framework. + +Our experimental results demonstrate that our proposed method, dubbed as RTD (Reducing Task Discrepancy of Text Encoders), consistently and considerably improves the performance in diverse evaluation datasets (CIRR [24], CIRCO [3], FashionIQ [38], COCO object composition [30], and GeneCIS [36]), when integrated with existing projection-based CIR methods (SEARLE [3], Pic2Word [30], LinCIR [15], Context-I2W [34], and FTI4CIR [23]). Consequently, compared to resource-intensive state-of-the-art synthetic CIR triplets-based method like MagicLens [40], RTD achieves comparable results, with less than a $1\%$ gap in CIRR R@1 and $0.1\%$ in FashionIQ R@10 when combined with FTI4CIR (ViT-L/14). When using larger or fine-tuned backbones (ViT-G/14 or FSC-CLIP [27]) with LinCIR, RTD achieves superior results than MagicLens in those metrics. + +We note that RTD achieves the above results with significantly higher training efficiency in both data generation and training time. Specifically, our rule-based text triplet generation requires less than 10 minutes to construct 1M text triplets $(570\times$ faster than CIR triplet generation from Compodiff [14]). With 4 NVIDIA A100 GPUs, the additional training time for RTD is just 23 minutes for training the ViT-L/14 backbone, making it approximately 10 to $100\times$ faster than synthetic CIR triplets-based methods. Even with a significantly larger backbone (ViT-G/14), training takes only 2.5 hours, remaining negligible in comparison. Moreover, we investigate the effectiveness of RTD across various text triplet generation strategies and different backbone sizes (ViT-B/32, ViT-L/14, ViT-G/14) and types (FSC-CLIP [27], CLIP [29]), showing its reproducibility and compatibility. + +# 2. Related Work + +Composed Image Retrieval. Unlike supervised CIR methods [1, 2, 11, 18], projection-based CIR methods [3, 12, 15, 23, 30, 33, 34] are built upon the frozen CLIP model, where a projection module $\phi$ is trained without CIR triplets. Each method employs a different training scheme for different training schemes for $\phi$ (See Sec. 4.1 for details). Several approaches avoid the need for any training by employing interpolation techniques [16], while others leverage powerful yet resource-intensive models such as LLMs and captioners [17, 39]. While our approach is built upon projection-based CIR methods, its training strategy is closely related to another category of CIR methods [14, 19, 37, 40] that use synthetically generated expensive CIR triplets. However, our method stands out by utilizing text triplets and updating only the text encoder, resulting in more efficient training while maintaining strong performance. + +Task discrepancy between the CLIP pre-training task and CIR. Combiner [2] updates the text encoder to minimize the gap between the target caption feature and the sum of the reference image and conditioning text features. FashionERN [8] addresses reference image dominance by introducing a separate branch to amplify the impact of the conditioning text. However, both Combiner and FashionERN require expensive CIR triplets $(I_r, T_c, I_t)$ for training, whereas our approach uses cheap, automatically generated text-only triplets $(T_r, T_c, T_t)$ . As another example, Chen and Lai [6] synthesize a triplet of an original image, its caption, and the masked image, treating them as the target image, conditioning text, and reference image, respectively. This approach, however, still has a gap between conditioning text (e.g., "change dog to cat") and image caption (e.g., "a dog catching a frisbee"); furthermore, it needs the full fine-tuning of the CLIP model, resulting in changing the visual embeddings in the retrieval database. In contrast, RTD directly uses the conditioning texts for training and does not change the target visual encoder, enabling the reuse of pre-extracted CLIP embeddings. Lastly, CIREVL [17] or LDRE [39] reduce task discrepancy by generating descriptive captions of the composed query using a large captioning model [21] and LLM [5]. While effective without any training, they rely on expensive inference steps and require carefully crafted prompts by skilled users. Whereas, RTD is fully automated, human-free, and more efficient during inference. + +# 3. Main Method + +# 3.1. Obtaining text triplets + +To address the task discrepancy, we collect cheaply and automatically generated text triplets $(T_r, T_c, T_t)$ , instead of directly using the expensive CIR triplets $(I_r, T_c, I_t)$ . Given reference caption $T_r$ from conventional caption datasets, + +conditioning text $T_{c}$ and target text $T_{t}$ can be generated using two strategies: via large language models (LLMs) [4, 14, 19, 37] or, more efficiently, through rule-based templates [14]. We investigate both strategies and demonstrate that RTD consistently improves performance across them. + +For the LLM-based strategy, we use the publicly available text triplets by CompoDiff [14], which are employed in our main experiments if not specified. These triplets are generated by taking a given caption $T_r$ as an input of the fine-tuned LLM, whose output predicts the corresponding conditioning text $T_c$ and the target caption $T_t$ . Previous works, such as IP2P [4], CoVR [37], and CASE [19], have also explored generating text triplets using LLMs, differing in LLM model types, input data, and fine-tuning strategies. The original purpose of these text triplet generations is to construct CIR triplets $(I_r, T_c, I_t)$ , but all these works also release the corresponding text triplets used for their CIR triplet construction. In addition to these publicly available text triplets, we also implement and evaluate an efficient in-context learning generation strategy using LLaMA3-8B [13] without additional fine-tuning. We conduct experiments with all the aforementioned text triplets in Tab. 6 and observe that RTD consistently delivers significant enhancements. + +For cheap, rule-based strategy, we can extract a keyword (e.g., "dog") from $T_{r}$ ("a dog catching a frisbee") and replace it with a randomly chosen keyword (e.g., "cat") [14], forming $T_{c}$ . The conditioning text is then generated automatically by using pre-defined templates (e.g., "change [original keyword] to [altered keyword]"). Our experiments show that this simple rule-based variant performs similarly to the LLM-based one. More explanations and examples of each strategy are provided in Suppl. A.2, and more details of generation costs are in Suppl. C. + +# 3.2. Frozen-target text contrastive learning + +Now, we explain our approach to update the text encoder for mitigating the task discrepancy solely with the generated text triplets $(T_r, T_c, T_t)$ . We first assume that there exists a pretrained projection module $\phi$ obtained by the projection-based CIR methods [3, 15, 30]. Recall that for a given reference image $I_r$ and conditioning text $T_c$ , the final composed feature is generated by passing the text prompt "a photo of $\phi(\psi_V(I_r))$ that $T_c$ " to the text encoder $\psi_T$ , where $\psi_V$ is the visual encoder and $\phi$ is the projection module (See Fig. 1). We aim to update the text encoder $\psi_T$ to reduce the discrepancy between the pretext task and projection-based CIR task using the text triplets while maintaining $\psi_V$ and $\phi$ frozen. + +Frozen-target text contrastive loss. We apply contrastive learning using a paired caption $(T_{r + c}, T_t)$ , where $T_{r + c}$ denotes a concatenated caption of reference caption $T_r$ and conditioning caption $T_c$ . Namely, we let the representation of the concatenated caption closely approximate that of the target caption. However, solely updating the text encoder + +![](images/9a5e7d43e47e88ef4a5ae72637812a8905eee82d0d7055088e9ac20f4606a392.jpg) +Figure 2. Overview of RTD. + +while fixing the image encoder can break the alignment between image and text encoders. To prevent the issue, we extract the text embedding of $T_{t}$ using the frozen text encoder $\psi_{T}$ , while the concatenated caption $T_{r + c}$ is extracted from the learnable text encoder $\psi_T^{tr}$ , initialized from $\psi_{T}$ . Here, we assume that as the target caption $T_{t}$ is a standard caption, a text embedding $\psi_{T}(T_{t})$ , is well-aligned with the frozen image embedding space. Based on this assumption, we freeze the target textual embedding to maintain the pretrained alignment as we learn new relationships. As shown in Sec. 4.3, freezing the target embedding is crucial for fine-tuning with our objective. + +Now, we introduce our frozen-target text contrastive loss $\mathcal{L}_{FCL}$ using two text encoders: a frozen pre-trained text encoder $\psi_T$ and a learnable text encoder $\psi_T^{tr}$ which is initialized with $\psi_T$ . Textual latent embeddings $\tilde{t}_{r + c}$ and $t_t$ are extracted from $\psi_T^{tr}$ and $\psi_T$ , respectively. Namely, $\tilde{t}_{r + c} = \psi_T^{tr}(E_w^{tr}(T_{r + c}))$ and $t_t = \psi_T(E_w(T_t))$ , where $E_w$ is a word embedding layer. We aim to tune $\psi_T^{tr}$ to minimize the distance between the concatenated textual embedding $\tilde{t}_{r + c}$ and the target textual embedding $t_t$ while maximizing the distance from other textual embeddings within the batch. We employ a symmetric InfoNCE loss [7, 10], as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {F C L} = \frac {1}{B} \sum_ {k = 1} ^ {B} - \log \left(\frac {e ^ {(c (\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {k}) / \tau)}}{\sum_ {j = 1} ^ {B} e ^ {(c (\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {j}) / \tau)} + \sum_ {j \neq k} e ^ {(c (t _ {t} ^ {k} , t _ {t} ^ {j}) / \tau)}}\right) \\ - \log \left(\frac {e ^ {\left(c \left(t _ {t} ^ {k} , \tilde {t} _ {r + c} ^ {k}\right) / \tau\right)}}{\sum_ {j = 1} ^ {B} e ^ {\left(c \left(t _ {t} ^ {k} , \tilde {t} _ {r + c} ^ {j}\right) / \tau\right)} + \sum_ {j \neq k} e ^ {\left(c \left(\tilde {t} _ {r + c} ^ {k} , \tilde {t} _ {r + c} ^ {j}\right) / \tau\right)}}\right) \tag {1} \\ \end{array} +$$ + +in which $c(\cdot, \cdot)$ denotes the cosine similarity, $B$ is the batch size, and $\tau$ is a temperature. + +Refined batch sampling strategy for hard negatives. To + +further enhance the efficacy of updating the text encoder, we devise a simple yet effective batch sampling strategy that incorporates pairs of $(T_{r + c}, T_t)$ and $(T_r, T_r)$ within the same batch. For example, as presented in Fig. 2, a pair such as $(T_{r + c}: \text{“A dog catching a frisbee} + \text{change dog to cat}$ , $T_t: \text{“A cat catching a frisbee})$ is sampled along with its corresponding reference pair $(T_r: \text{“A dog catching a frisbee}$ , $T_r: \text{“A dog catching a frisbee})$ in the same batch. This setup ensures that the concatenated text $T_{r + c}$ and its corresponding reference text $T_r$ implicitly act as hard negatives for each other, as their semantics are much more similar $(T_{r + c}$ is derived from $T_r)$ than those of other randomly sampled texts in the batch. Explicitly distinguishing the embedding of $T_{r + c}$ from that of $T_r$ conceptually aligns well with CIR task, as it encourages the model to better capture modifications (in $T_c$ ). Moreover, we believe including $(T_r, T_r)$ pairs in the contrastive learning helps the learnable text encoder $\psi_T^{tr}$ remain closely aligned with the pre-trained encoder $\psi_T$ . + +Refined concatenation of reference and conditioning texts. As we use "a photo of $[\$ ]$ that $[T_c]$ " for inference, a naive concatenation strategy also can suffer from training-inference task discrepancy. To tackle this issue, instead of simply concatenating the $T_r$ and $T_c$ , we use the prompt "a photo of $[\$ ]$ that $[T_c]$ " for updating the text encoder, where $[\$ ]$ is obtained by the reference caption $T_r$ with the projection module $\phi$ . Instead of obtaining a pseudo-word token with latent image embedding $v$ , we utilize a textual latent embedding from the reference caption $T_r$ , i.e., $\phi (t_r)$ . However, Gu et al. [15] showed that naively replacing the image encoder with the text encoder for the input of $\phi$ will suffer from the modality gap [22], a phenomenon where text and image embeddings have a gap between them. We tackle this issue by injecting random noise into the textual token representation before it is processed by $\phi$ , following Gu et al. [15]. More analyses on variations of noise are in Suppl. B.10. + +Fig. 2 illustrates the overview of RTD. We use CLIP backbone and pre-trained projection module $\phi$ produced by the existing projection-based CIR methods. The text encoder is trained using the proposed loss function (Eq. (1)) while applying the refined batch sampling and concatenation scheme. During inference, the procedure mirrors that of existing projection-based CIR methods, except we utilize the updated text encoder $\psi_T^{tr}$ instead of the frozen one $\psi_T$ . Note that our method only updates the text encoder while the image encoder and the projection module are frozen. + +Remark 1: Low data acquisition cost. The text triplets we use can be obtained at a significantly lower cost than CIR triplets. First, the text triplet generation process requires only an easily obtainable caption dataset for generation, whereas other approaches that generate CIR triplets [19, 37, 40] necessitate image or video datasets along with an additional collection phase for semantically similar images or videos. Second, text triplet generation can avoid resource-intensive text-to-image generation [4, 14], making it $15 \times$ faster than the CIR triplet generation process used in CompoDiff [14]. Furthermore, if we choose the rule-based text triplet generation strategy, the process becomes $570 \times$ faster than the full CIR triplet generation with images; 1M text triplets generation takes just 0.1 hours. Lastly, the storage for the generated text triplets takes up only 100MB, whereas storing a similar quantity of images requires significantly more space (e.g., 400GB for CC3M [32]). More details can be found in Suppl. C. + +Table 1. Training and inference time comparisons. Purple denotes the additional resource required for RTD. Inference time is measured on a single A100 GPU for a single image. Note that the training hours for SEARLE, LinCIR for 8 A100 GPUs [15] and their RTD variants are trained on 4 A100 GPUs. CompoDiff is trained on 128 A100 and MagicLens is trained on 128 TPUs. + +
MethodCLIP BackboneTraining (h)Inference (s)
SEARLEL/144.2h0.02s
SEARLE (+ RTD)L/144.58h (+0.38h)0.02s
LinCIRL/140.5h0.02s
LinCIR (+ RTD)L/140.88h (+0.38h)0.02s
LinCIRG/140.8h0.05s
LinCIR (+ RTD)G/143.3h (+2.5h)0.05s
CompoDiffL/14231h0.12s
MagicLensL/146hunknown
CIRVLL/14-3.02s
+ +Remark 2: Training and inference efficiency. As reported in Tab. 1, the additional training cost of RTD is small; 0.38 hours for ViT-L/14 on 4 A100 GPUs. In contrast, CompoDiff and MagicLens require 231 hours on 128 A100 GPUs and 6 hours on 128 TPUs, respectively, making the cost of RTD negligible. Even with a larger ViT-G/14 backbone, training RTD takes only 2.5 hours on 4 A100 GPUs, still faster than synthetic CIR triplet-based methods and even comparable to SEARLE in ViT-L/14 (4.2h). This training efficiency mainly stems from the advantage of text-only training. As highlighted by Gu et al. [15], the training complexity for the text + +encoder is much lower than that for the visual encoder due to the relatively short token lengths of texts ( $\approx 12$ ) compared to images (256). The average inference time of the CLIP ViT-L/14 image encoder is $\times 3.5$ times slower than that of the text encoder. Another source of efficiency is that RTD requires relatively few iterations (at most 2000, and significantly fewer in most cases), as the text encoders are already pre-trained and only require minor adjustments to learn the relationships between text triplets. Moreover, in Suppl. B.4, we present a more efficient implementation by selectively updating only a few layers of the text encoder. Along with efficient training, RTD retains the fast inference speed of the original projection-based methods, remaining $150\times$ faster than training-free CIREVL. + +# 4. Experiments + +# 4.1. Experimental setup + +Implementation details. We use the AdamW optimizer [25] with a weight decay of 0.01. The learning rate is set to $10^{-5}$ , with a batch size of 512. We select the text encoder model with the best zero-shot CIRR [24] dev R@1 score for evaluating RTD. We mainly use the visual and textual encoders of the CLIP ViT-L/14 [29] as our backbone. Unless otherwise noted, we use LLM-based 2.5M text triplets provided by CompoDiff [14] for the training. We set the $\tau$ as 0.07 in Eq. (1) and scale the standard deviation of Gaussian distribution as 0.5 for the noise injection. More results on various noise distributions can be found in the Suppl. B.10. All experiments were conducted using four NVIDIA A100 GPUs with Python 3.8 and Pytorch [28]. + +Evaluation datasets and metrics. We compare projection-based CIR methods on five benchmark datasets: CIRR [24], CIRCO [3], FashionIQ [38] (FIQ), COCO object composition [30], and GeneCIS [36]. Details of each dataset are in the Suppl. A.1. For CIRR, FIQ, COCO, and GeneCIS, we have reported their recall scores at the top K retrieval results $(\mathbb{R}@\mathbb{K})$ . Since the CIRCO dataset includes multiple positive images for each query, we use a ranking-based metric; mean Average Precision scores at the top K results $(\mathrm{mAP}@\mathrm{K})$ [9, 26]. For the main results, we compare the results on the FIQ validation split, as well as the test sets of CIRR and CIRCO. For the ablation studies and analyses, the validation splits of these three datasets are utilized. GeneCIS and COCO object composition results are in the Suppl. B.1. + +Baselines. We evaluate the effect of our method when combined with publicly available projection-based ZS-CIR methods: Pic2Word [30], SEARLE [3], LinCIR [15], Context-I2W [34], and FTI4CIR [23]. All these methods share a similar core concept shown in Fig. 1, but use different training schemes. Pic2Word [30] optimizes contrastive loss between the image embedding and its projected text embedding of "a photo of [\$" to obtain the projection module $\phi$ . Similarly, + +Table 2. Comparison with other baselines. Note that this comparison is not entirely fair due to differences in backbone models and training data. Purple denotes the performance gain from our method, while red and blue highlight the best and second-best scores, respectively. + +
MethodBackboneCIRRCIRCOFashionIQ
R@1R@5R@10mAP@5mAP@10mAP@25R@10R@50
Pic2Word [30]CLIP ViT-L/1424.251.564.18.39.110.125.344.9
+RTD27.9 (+3.6)56.2 (+4.8)68.5 (+4.4)9.1 (+0.9)9.6 (+0.5)10.7 (+0.6)27.6 (+2.3)48.9 (+4.0)
SEARLE [3]CLIP ViT-L/1424.952.365.711.612.714.325.045.3
+RTD26.6 (+1.7)56.2 (+3.9)69.0 (+3.3)16.5 (+4.9)17.9 (+5.2)19.8 (+5.4)29.3 (+4.4)50.7 (+5.4)
Context-I2W [34]CLIP ViT-L/1425.655.468.6---27.949.1
+RTD29.2 (+4.6)58.4 (+3.0)70.5 (+1.9)---28.1 (+0.2)49.5 (+0.4)
FTI4CIR [23]CLIP ViT-L/1425.955.667.715.116.318.129.450.9
+RTD29.1 (+3.2)58.8 (+3.2)70.4 (+2.7)16.2 (+1.1)17.4 (+1.1)19.4 (+1.3)30.6 (+1.2)51.7 (+0.8)
LinCIR [15]CLIP ViT-L/1423.852.966.513.014.115.827.447.7
+RTD26.6 (+2.9)56.2 (+3.3)69.0 (+2.5)17.1 (+4.1)18.1 (+4.0)20.1 (+4.3)30.2 (+2.8)51.1 (+3.4)
LinCIR [15]FSC-CLIP [27] ViT-L/1433.663.174.514.215.417.028.349.5
+RTD35.9 (+2.3)67.5 (+4.4)78.1 (+3.6)18.3 (+4.1)19.7 (+4.3)21.3 (+4.3)31.2 (+2.9)52.3 (+2.8)
KEDs [33]CLIP ViT-L/1426.454.867.2---26.847.9
CIREVL [17]CLIP ViT-L/1424.552.364.918.619.020.928.648.6
LDRE [39]CLIP ViT-L/1426.555.667.523.424.026.428.550.5
MT-CIR [6]CLIP ViT-L/1425.554.667.610.411.613.035.457.4
MagicLens [40]CLIP ViT-L/1430.161.774.429.630.833.430.752.5
Compodiff [14]CLIP ViT-L/1426.755.072.612.613.415.836.048.6
CoVR [37]BLIP ViT-L/1439.368.278.9---27.744.6
CASE [19]BLIP ViT-L/1435.465.878.5-----
LinCIR [15]CLIP ViT-G/1434.964.576.120.621.924.144.565.5
+RTD36.3 (+1.4)67.5 (+3.0)78.3 (+2.2)21.1 (+0.5)22.3 (+0.4)24.5 (+0.4)46.2 (+1.7)67.3 (+1.8)
Compodiff [17]CLIP ViT-G/1434.764.375.115.317.719.439.051.7
CIREVL [17]CLIP ViT-G/1426.755.174.526.827.630.032.252.4
LDRE [39]CLIP ViT-G/1436.266.477.331.132.235.032.555.4
+ +SEARLE [3] employs a two-stage approach, starting with an optimization-based textual inversion phase followed by a distillation phase for the projection module $\phi$ . LinCIR [15] introduces a language-only self-supervised task based on keyword token replacement, where the model encourages the original text embedding and the replaced text embedding (keyword tokens are replaced with the projection of the original text embedding via $\phi$ ) to be aligned. Context-I2W [34] refines the projection module by selecting relevant visual information [34]. Lastly, FTI4CIR [23] separately maps images into subject- and attribute-oriented pseudo-word tokens [23]. Details on how RTD is combined with them can be found in Suppl. A.3. + +We train these methods with the CLIP ViT-L/14 in our main experiments. To further verify the compatibility of RTD, we also report results for the ViT-B/32 backbone with SEARLE, Pic2Word, and LinCIR. Moreover, we conduct experiments with a larger backbone (ViT-G/14) and a fine-tuned CLIP model for enhanced compositionality (FSC-CLIP [27]) for LinCIR, enabled by its fast training capability. We use the publicly available pre-trained model for SEARLE (ViT-B/32, ViT-L/14), Pic2Word (ViT-L/14), Context-I2W (ViT-L/14), and FTI4CIR (ViT-L/14). Otherwise, we reproduce the results using the official implementation. When reproducing, we adhere to the same settings in the original papers. For example, we select the final last epoch model for the Pic2Word ViT-B/32 model and choose the model based on the best zero-shot CIRR dev R@1 score for LinCIR. We additionally compare our method with a diverse set of CIR approaches, including another projection-based + +ZS-CIR method (KEDs [33]), another attempt to address task discrepancy (MT-CIR [6]), approaches leveraging synthetically generated CIR triplets (CoVR [37], CASE [19], MagicLens [40] and Compodiff [14]), and the training-free methods (CIREvL [17], LDRE [39]). + +# 4.2. Main results + +Tab. 2 shows the overview of comparison results with state-of-the-art CIR methods. First, we assess the impact of integrating RTD with existing projection-based CIR methods (SEARLE, Pic2Word, LinCIR, Context-I2W, and FTI4CIR) across CIRR, CIRCO, and FashionIQ. RTD consistently boosts performance, yielding an average improvement of over 2.78 points. This trend also holds for the GeneCIS and COCO object composition datasets, as detailed in Suppl. B.1. + +Second, we compare the integration of RTD with those leveraging synthetically generated CIR triplets (CoVR, CASE, MagicLens, and Compodiff). We focus on this comparison to ensure fairness, as leveraging text triplets and updating the textual backbone of RTD may not be fully aligned with the standard projection-based CIR setting. We observe that RTD delivers competitive or superior performance compared to these resource-intensive synthetic CIR triplet-based methods, while being significantly more efficient; over $10 - 100 \times$ faster, as noted in Remark-2 of Sec. 3. For example, with the same CLIP ViT-L/14 backbone, LinCIR + RTD outperforms MagicLens and CoVR in FashionIQ R@10 and R@50 while FTI4CIR + RTD or Context-I2W + RTD achieve performance on par with or exceeding that of MagicLens, MT-CIR, and Compodiff in CIRR metrics. + +When combined with larger (ViT-G/14) or fine-tuned (FSC-CLIP) backbones, RTD achieves the best or second-best result on the CIRR and FashionIQ benchmarks. We believe this flexibility of using different backbones underscores the benefit of efficient text-only training of RTD, inherited by LinCIR [15]. We reemphasize that even with a larger backbone like ViT-G/14, RTD remains much faster in training than other synthetic triplet-based methods, as shown in Tab. 1. + +Lastly, across both backbones (ViT-L/14, G/14), LinCIR + RTD mostly outperforms training-free CIREVL and LDRE, except on CIRCO metrics. As shown in Tab. 1, CIREVL is significantly slower $(150\times)$ than projection-based methods (including RTD), while LDRE incurs even larger inference time due to its ensemble-based strategy. + +Table 3. Ablation study. Unlike in Tab. 2, for ablation studies and analyses, validation splits of three CIR datasets are used for evaluation. We measure the impact of FCL loss (Eq. (1)), refined batch sampling (RB), and refined concatenation scheme (RC). All models are based on LinCIR ViT-L/14. The first row denotes the vanilla LinCIR without RTD. "Avg" denotes the average of all reported metrics. Bold indicates the best result. + +
TextFCL Frozen TargetRBRCCIRR (R) @5@10CIRCO (mAP) @10@25FIQ (R) @10@50Avg
--XX54.367.812.714.527.447.737.4
\( \left( {{T}_{r},{T}_{r}}\right) \)✓XX56.069.713.415.228.248.838.5
\( \left( {{T}_{r + c},{T}_{t}}\right) \)✓XX58.271.514.416.026.947.939.2
\( \left( {{T}_{r + c},{T}_{t}}\right) \)✓✓X58.271.315.016.727.449.339.6
\( \left( {{T}_{r + c},{T}_{t}}\right) \)X✓X54.367.012.213.625.045.336.3
\( \left( {{T}_{r + c},{T}_{t}}\right) \)✓✓✓57.971.116.117.830.251.140.7
+ +# 4.3. Ablation studies + +Tab. 3 presents the effectiveness of the proposed components: frozen-target text contrastive loss (FCL), refined batch sampling (RB), and refined concatenation scheme (RC). All evaluation results are on the validation splits. All model variants use ViT-L/14 and a projection module $\phi$ from LinCIR, making the results in row 1 indicative of the original performance of LinCIR. We first compare the impact of the text pairs fed into FCL loss. We compare our design choice $(T_{r + c}, T_t)$ from the generated text triplets with $(T_r, T_r)$ , which is the sole option for constructing a pair given a single conventional caption $T_r$ . The results demonstrate that, on average, using generated triplets (3rd row) is more effective than using original conventional text pairs (2nd row), particularly in the CIRR and CIRCO datasets. In addition, RB (4th row) and RC (6th row) significantly enhance the overall performance, demonstrating the effectiveness of these components. Finally, we measure the impact of using the frozen target textual embedding, denoted as "Frozen Target" in the table. Significant performance degradation is observed when the learnable text encoder is used for extracting the embedding of the target caption $T_t$ (5th row) compared to the frozen-target case (4th row), supporting the importance of our design choice. + +Table 4. T2I retrieval performance of different text encoders on CIRCO val set. "Update (pair)" refers to the setting in the second row of Tab. 3, which uses contrastive learning with the pair $(T_r, T_r)$ . Bold indicates the best result, excluding the first row (oracle case). + +
QueryText encodermAP@5mAP@10mAP@25
\(T_t\)Frozen18.9619.3121.05
\(T_{r+c}\)Frozen10.1210.7112.34
\(T_{r+c}\)Update (pair)10.5211.1512.72
\(T_{r+c}\)RTD15.1215.8017.77
+ +# 4.4. Analyses on our core motivation + +We conduct experiments to validate our core motivation (reducing task discrepancy) and verify that the observed gains stem from it. Other details follow Sec. 4.3. + +Can RTD really reduce the task discrepancy of the text encoder? We first quantitatively verify whether RTD indeed reduces the task discrepancy. We first conduct a controlled experiment that measures the text-to-image (T2I) retrieval performance of the text encoder with conditional texts. We retrieve the target images $I_{t}$ with the concatenated text query $T_{r + c}$ or the ideal target caption $T_{t}$ . If our text encoder successfully handles the discrepancy due to the concatenated caption, the text encoder updated by RTD will perform better than the frozen one or "Update (pair)", which is updated by contrastive learning with the $(T_{r}, T_{r})$ pair (corresponding to row 2 of Tab. 3). We use the CLIP ViT-L/14 and CIRCO [3] validation set for evaluation. Since the CIRCO dataset only has CIR triplets $(I_{r}, T_{c}, I_{t})$ , we use the BLIP [20] captioner to generate $T_{r}$ and $T_{t}$ corresponding to the $I_{r}$ and $I_{t}$ , respectively. Here, the simple concatenation scheme is applied for the text query $T_{r + c}$ in all cases. Tab. 4 shows that when the text encoder is either frozen or updated with $(T_{r}, T_{r})$ , the retrieval results using the concatenated caption $T_{r + c}$ are significantly worse than those using the target caption $T_{t}$ . It supports the claim that the frozen text encoder suffers from the negative effects of task discrepancy between the pretext and CIR tasks. Moreover, it suggests that simply updating the text encoder with $(T_{r}, T_{r})$ pair fails to reduce this discrepancy. In contrast, the text encoder updated by RTD shows a significant improvement over the frozen text encoder or one updated with $(T_{r}, T_{r})$ , showing that it successfully reduces the task discrepancy. + +We additionally measure the average cosine similarity between the composed textual features with the prompt "a photo of $\phi (\psi_V(I_r))$ that $T_{c}$ " (Fig. 1a) and the target image features (Fig. 1c). The similarity is measured by the LinCIR ViT-L/14 model on the CIRCO validation split. When we use the frozen text encoder $(\psi_T)$ , the average similarity is 0.1. By changing the text encoder to our updated text encoder $(\psi_T^{tr})$ , the similarity becomes 0.29 (+0.19). This result shows again that RTD successfully aligns the composed query features using $\phi$ to the frozen CLIP image features. + +Is all the gain of RTD from naive text-encoder updating? To verify that our improvements cannot be achieved solely + +by updating the text encoder backbone without considering the task discrepancy, we additionally measure the results of previous methods (Pic2Word and LinCIR) when naively updating the text encoders. Namely, after training $\phi$ while keeping all other networks frozen as in previous methods, we additionally update the text encoder using the original loss function, while fixing other modules including $\phi$ . We denote this update rule as "naive" in the Tab. 5. Unlike RTD, we observe that just naively updating the text encoder ("naive") significantly degrades the performance of the baseline. The results indicate that merely updating the text backbone is not beneficial for CIR; instead, mitigating task discrepancy through RTD is necessary. + +Table 5. Impact of the update scheme. Two update schemes are compared: (1) using the original objective from baseline ("naive") and (2) using RTD. For a fair comparison, in both schemes, $\phi$ is updated first and $\psi_T$ is updated top on the frozen modules. Other details are the same as Tab. 3. + +
CIRR (R)CIRCO (mAP)FIQ (R)Avg
@5@10@10@25@10@50
Pic2Word51.464.48.810.125.344.932.2
+naive19.227.51.31.64.411.210.9
+ RTD56.669.88.89.827.648.936.9
LinCIR54.367.812.714.527.447.736.9
+naive52.766.811.413.026.345.935.5
+ RTD57.971.116.117.830.251.140.7
+ +Next, instead of using the original objective from baselines, we update the text encoder with conventional contrastive learning using the pair $(T_r, T_r)$ , which corresponds exactly to row 2 in Tab. 3 (row 3 in Tab. 4). This setup can also serve as a proxy for the standalone impact of breaking the alignment of pre-trained knowledge, as noted in [2]. As demonstrated in Sec. 4.3, the large gap (more than 2 points) between RTD (last row in Tab. 3) and this proxy (row 2 in Tab. 3) again shows that our gain cannot be attributed to this simple alternative (naïve text encoder tuning). + +![](images/74c016a39698e76fab097c94a755f2f123fedb0780cb4d21b0a88e3882babda9.jpg) +Figure 3. Impact of sizes of backbone. The results of RTD combined with Pic2Word [30], SEARLE [3], and LinCIR [15] across different CLIP backbones (ViT-B/32 and ViT-L/14) are shown. Here, the score is the same metric in "Avg" in Tab. 3 and other details are the same as Tab. 3. Full results are in the Suppl. B.2. + +# 4.5. Compatibility across backbone sizes + +In the Fig. 3, we observe that the integration of our approach with projection-based CIR methods significantly improves the performance across pioneering projection-based ZS-CIR methods (SEARLE, Pic2Word, and LinCIR) and all backbones (ViT-B/32 and ViT-L/14). For example, regardless of the choice of projection module $\phi$ and backbones, the minimum performance gain for average scores is greater than 2.8 points. The results of RTD using the larger backbone (ViTG/14) can be found in Tab. 2. More detailed and qualitative results are provided in Suppl. B.3 and Suppl. D, respectively. + +Table 6. The effectiveness of different types of text triplets for RTD. "In-context" denotes an efficient implementation using in-context learning with LLaMA3-8B, without fine-tuning. Details and examples of each text triplet dataset are summarized in Tab. S1 and Tab. S2, respectively. Other details are the same as Tab. 3. + +
SourceLLMCIRR (R) @5 @10CIRCO (mAP) @10 @25FIQ (R) @10 @50Avg
LinCIR--54.367.812.714.527.447.737.4
+RTDRule-basedX56.770.315.017.030.451.940.2 (+ 2.9)
IP2P [4]✓58.771.615.918.029.650.740.7 (+ 3.3)
Compodiff [14]✓57.971.116.117.830.251.140.7 (+ 3.3)
In-context✓59.371.815.817.529.751.440.9 (+ 3.5)
CoVR [37]✓59.872.615.417.029.650.841.9 (+ 4.5)
CASE [19]✓56.369.311.112.726.647.837.7 (+ 0.3)
+ +# 4.6. Impact of the text triplet generation strategies + +As explained in Sec. 3.1, we evaluate RTD using both 1) publicly available LLM-based text triplets (from IP2P, Compodiff, CoVR, and CASE) along with efficient in-context learning-based text triplets, and 2) LLM-free, rule-based triplets. Tab. 6 shows that RTD consistently improves CIR performance across them (+3.3 for IP2P, +3.3 for Compodiff, +3.5 for in-context learning, +4.5 for CoVR, and +0.31 for CASE, and rule-based triplets achieve 2.9, respectively). We believe this result shows the reproducibility and consistency of RTD, with the rule-based triplets performing comparably to LLM-generated ones, indicating that efficient rule-based triplets are sufficient to achieve strong CIR performance. The marginal improvement in CASE is largely due to the poor quality of text triplets resulting from its construction pipeline that prioritizes CIR triplet quality over text triplet quality, as shown in Tab. S1. Further details can be found in Suppl. A.2, and additional analyses, including data scales related to text triplets, are provided in Suppl. B.5. + +# 5. Conclusion + +We presented RTD, a novel post-processing approach that can be easily integrated into existing projection-based CIR methods, aimed at reducing task discrepancy of text encoders. Empirical evaluations demonstrate that RTD significantly boosts the performance of existing projection-based CIR methods across diverse datasets and model backbones, competing with or outperforming other resource-intensive CIR methods with much greater efficiency. + +# Acknowledgment + +This work was supported in part by National Research Foundation of Korea (NRF) grant [No. 2021R1A2C2007884, No. RS-2025-02263628], the Institute of Information & communications Technology Planning & Evaluation (IITP) grants [RS-2021-II212068, RS-2022-II220113, RS-2022-II220959, RS-2021-II211343], and the BK21 FOUR Education and Research Program for Future ICT Pioneers (Seoul National University), funded by the Korean government (MSIT). + +# References + +[1] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 3 +[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, 2022. 2, 3, 8 +[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 5, 6, 7, 8 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 5, 8 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 3 +[6] Junyang Chen and Hanjiang Lai. Pretrain like you inference: Masked tuning improves zero-shot composed image retrieval. arXiv preprint arXiv:2311.07622, 2023. 3, 6 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4 +[8] Yanzhe Chen, Huasong Zhong, Xiangteng He, Yuxin Peng, Jiahuan Zhou, and Lele Cheng. Fashionern: Enhance-and-refine network for composed fashion image retrieval. In AAAI, 2024. 3 +[9] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image-caption associations for ms-coco. In ECCV, 2022. 5 +[10] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “this is my unicorn, fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 4 +[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 3 +[12] Yongchao Du, Min Wang, Wengang Zhou, Shuping Hui, and Houqiang Li. Image2sentence based asymmetrical zero-shot composed image retrieval. In ICLR, 2024. 3 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The + +llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3 +[14] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. TMLR, 2023. 1, 2, 3, 5, 6, 8 +[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoonhoon Kang, and Sangwoo Yun. Language-only efficient training of zero-shot composed image retrieval. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[16] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 3 +[17] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 3, 6 +[18] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, 2021. 1, 3 +[19] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6, 8 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 7 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3 +[22] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4 +[23] Haoqiang Lin, Haokun Wen, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Fine-grained textual inversion network for zero-shot composed image retrieval. In SIGIR, 2024. 1, 2, 3, 5, 6 +[24] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 5 +[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5 +[26] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In ECCV, 2020. 5 +[27] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pretrained vlms for improving vision-linguistic compositionality. In EMNLP, 2024. 2, 6 +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 5 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5 +[30] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 5, 6, 8 +[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Dataset and Benchmark, 2022. 1 +[32] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5 +[33] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 1, 3, 6 +[34] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6 +[35] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[36] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5 +[37] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. In AAAI, 2024. 1, 2, 3, 5, 6, 8 +[38] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2, 5 +[39] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In SIGIR, 2024. 3, 6 +[40] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 1, 2, 3, 5, 6 \ No newline at end of file diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/images.zip b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e039864080541807cf2e03ad06bbcc01ff33ac23 --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38836d96e9c81ffabf8c783c78b035a6d4413921ecdd0963bca542ce86830ff3 +size 488141 diff --git a/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/layout.json b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5226066d136144a1766e5f84924cf2b9136b8a77 --- /dev/null +++ b/2025/An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval/layout.json @@ -0,0 +1,8312 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "spans": [ + { + "bbox": [ + 96, + 102, + 516, + 140 + ], + "type": "text", + "content": "An Efficient Post-hoc Framework for Reducing Task Discrepancy of Text Encoders for Composed Image Retrieval" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "content": "Jaeseok Byun" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "content": " Seokhyeon Jeong" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "content": " Wonjae Kim" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "content": " Sanghyuk Chun" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "text", + "content": " Taesup Moon" + }, + { + "bbox": [ + 87, + 160, + 526, + 176 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "text", + "content": "Department of ECE, Seoul National University " + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "text", + "content": "NAVER AI Lab " + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 141, + 188, + 468, + 217 + ], + "type": "text", + "content": "ASRI/INMC/IPAI/AIIS, Seoul National University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "text", + "content": "Composed Image Retrieval (CIR) aims to retrieve a target image based on a reference image and conditioning text, enabling controllable image searches. The mainstream Zero-Shot (ZS) CIR methods bypass the need for expensive training CIR triplets by projecting image embeddings into the text token embedding space, forming a composed query for retrieval. However, we highlight an inherent limitation in these projection-based CIR: a task discrepancy of text encoders between the original pre-training task of the encoders (text " + }, + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "text", + "content": " image) and the target CIR task (image + text " + }, + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\leftrightarrow" + }, + { + "bbox": [ + 56, + 270, + 296, + 594 + ], + "type": "text", + "content": " image), which potentially negatively impacts CIR performance. To reduce such a discrepancy, a naive solution would be to train both image and text encoders with CIR triplets in a supervised manner. Instead, we introduce Reducing Task Discrepancy of Text Encoders (RTD), an efficient text-only post-hoc framework that complements projection-based CIR methods. We devise a novel frozen-target text contrastive learning designed to enhance the capability of the text encoder for CIR. We also propose two key enhancements: (1) a hard negative-based refined batch sampling strategy and (2) a refined concatenation scheme to further mitigate training-inference discrepancy. Integrating RTD into state-of-the-art projection-based methods achieves performance comparable to, or even surpassing, resource-intensive state-of-the-art synthetic CIR triplet-based approaches, only with 23 minutes of additional training on 4 A100 GPUs. Our code is available in https://github.com/jaeseokbyun/RTD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 618, + 135, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 618, + 135, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 618, + 135, + 630 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 297, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 297, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 297, + 675 + ], + "type": "text", + "content": "Composed Image Retrieval (CIR) aims at retrieving a target image that closely resembles a reference image while reflecting the changes described in a conditioning text. Using a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": "query composed of image and text allows users to conduct more precise and flexible searches by specifying the desired modifications to the image through text. Supervised CIR methods [1, 11, 18] have been introduced to fuse information from bi-modal query, using labeled data from the target domain in the form of triplets " + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "inline_equation", + "content": "(I_r, T_c, I_t)" + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": " is a reference image, " + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": " is a conditioning text, and " + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "inline_equation", + "content": "I_t" + }, + { + "bbox": [ + 313, + 246, + 555, + 474 + ], + "type": "text", + "content": " is a target image. However, unlike the typical web-crawled image-text datasets [31], acquiring sufficient triplets for training needs expensive manual human annotations. To overcome the dependency on small-scale and human-verified triplets of the target domain, several works utilize the power of recent generative models. For example, a line of studies [14, 19, 37, 40] uses text-to-image models like IP2P [4] or large-language models (LLM) [5, 35] to synthesize large-scale CIR triplets for training, in place of the target-domain CIR triplets. While these methods achieve strong performance, they are often impractical due to the high computational and memory requirements for utilizing generative models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 629 + ], + "type": "text", + "content": "Another approach for removing the dependency on the CIR triplets, which often is referred to as projection-based ZS-CIR [3, 15, 23, 30, 33, 34], employs an integrable projection module on top of the pre-trained, frozen, and shared embedding space, such as CLIP [29]. Namely, a projection module \\(\\phi\\), which maps a CLIP image embedding to the CLIP text token embedding space, can be trained by solely using images [3, 30] or texts [15]. During inference, as illustrated in Fig. 1, these methods first project the embedding of the query image to a text token embedding \\([\\\\) ]\\) using the function \\(\\phi\\). This embedding is then combined with the conditioning text \\([T_c]\\) to create the prompt \"a photo of \\([\\\\) ]\\) that \\([T_c]\\)\", which is used as a query for the text-to-image retrieval." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 714 + ], + "type": "text", + "content": "The core assumption of the projection-based CIR is that the pre-trained text encoder should be robust enough to combine information from both the projected text token embedding and the conditioning text. However, we argue that this can cause a significant task discrepancy for the pre-trained text encoder between the original image-text alignment pretraining task (of CLIP) and the CIR task. For example, in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 684, + 133, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 684, + 133, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 684, + 133, + 693 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 694, + 295, + 712 + ], + "type": "text", + "content": "†Corresponding authors. (E-mail: sanghyukc@princeton.edu, tsmoon@snu.ac.kr)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3895" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 553, + 244 + ], + "type": "image", + "image_path": "a14d8142e868fc30bd171bcafe84ad2a7d31792c340085672451184cfcd03fc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 555, + 269 + ], + "type": "text", + "content": "Figure 1. The task discrepancy of projection-based ZS-CIR methods between the pre-training task (image-text alignment) and the ZS-CIR task (image-text composition)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 275, + 297, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 297, + 455 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 297, + 455 + ], + "type": "text", + "content": "Fig. 1, consider an ideal caption that accurately describes the target image. Since the CLIP text and image encoders are learned through contrastive learning, we can expect that the target image embedding (Fig. 1c) will align well with the embedding (Fig. 1b) of the ideal caption. In contrast, in the projection-based CIR, the text encoder receives a concatenated caption that combines the projected token " + }, + { + "bbox": [ + 54, + 275, + 297, + 455 + ], + "type": "inline_equation", + "content": "[\\" + }, + { + "bbox": [ + 54, + 275, + 297, + 455 + ], + "type": "text", + "content": " ]$ and the conditioning text, not an ideal caption. However, as noted by Baldrati et al. [2], the text encoder is not typically trained to encode complex textual modifications (e.g., addition or replacement) to the reference image, which are common in conditioning texts. As a result, there is no guarantee that the textual embedding of the concatenated caption (Fig. 1a) closely aligns with the target image embedding (Fig. 1c), which will undermine the retrieval performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": "To that end, we devise a complementary post-processing approach for existing projection-based CIR methods that reduces the task discrepancy of the text encoder efficiently. Ideally, both image and text backbones would be updated using expensive CIR triplets " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "(I_r, T_c, I_t)" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": "; however, we achieve this using cheap text triplets and only update the text encoder. These text triplets " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_r, T_c, T_t)" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " consist of a reference caption " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_r, e.g., “a dog catching a frisbee”)" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": ", a conditioning text " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_c, e.g., “change dog to cat”)" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": ", and a target caption " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_t, e.g., “a cat catching a frisbee”)" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": ", respectively. Given that the reference captions " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " are readily available from conventional caption datasets, " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " can be automatically generated without any human labor [24, 38] or intensive resources [14, 19, 37, 40]. Using these triplets, we introduce a frozen-target text contrastive learning, in which the text encoder updates the embedding of the concatenated caption " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " (e.g., “a dog catching a frisbee”+ “change the dog to a cat”) to align closely with the fixed embedding of the target caption " + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 56, + 463, + 297, + 714 + ], + "type": "text", + "content": " from the frozen text encoder. We further enhance this text-only training approach with two key components: a batch sampling strategy that ensures hard neg-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "text", + "content": "atives in each mini-batch and a refined concatenation scheme for " + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "inline_equation", + "content": "T_{r}" + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 313, + 275, + 555, + 312 + ], + "type": "text", + "content": " designed to reduce the training-inference task discrepancy caused by our text-only training framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "text", + "content": "Our experimental results demonstrate that our proposed method, dubbed as RTD (Reducing Task Discrepancy of Text Encoders), consistently and considerably improves the performance in diverse evaluation datasets (CIRR [24], CIRCO [3], FashionIQ [38], COCO object composition [30], and GeneCIS [36]), when integrated with existing projection-based CIR methods (SEARLE [3], Pic2Word [30], LinCIR [15], Context-I2W [34], and FTI4CIR [23]). Consequently, compared to resource-intensive state-of-the-art synthetic CIR triplets-based method like MagicLens [40], RTD achieves comparable results, with less than a " + }, + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "text", + "content": " gap in CIRR R@1 and " + }, + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 313, + 333, + 556, + 513 + ], + "type": "text", + "content": " in FashionIQ R@10 when combined with FTI4CIR (ViT-L/14). When using larger or fine-tuned backbones (ViT-G/14 or FSC-CLIP [27]) with LinCIR, RTD achieves superior results than MagicLens in those metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "We note that RTD achieves the above results with significantly higher training efficiency in both data generation and training time. Specifically, our rule-based text triplet generation requires less than 10 minutes to construct 1M text triplets " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "(570\\times" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " faster than CIR triplet generation from Compodiff [14]). With 4 NVIDIA A100 GPUs, the additional training time for RTD is just 23 minutes for training the ViT-L/14 backbone, making it approximately 10 to " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " faster than synthetic CIR triplets-based methods. Even with a significantly larger backbone (ViT-G/14), training takes only 2.5 hours, remaining negligible in comparison. Moreover, we investigate the effectiveness of RTD across various text triplet generation strategies and different backbone sizes (ViT-B/32, ViT-L/14, ViT-G/14) and types (FSC-CLIP [27], CLIP [29]), showing its reproducibility and compatibility." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3896" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "content": "Composed Image Retrieval. Unlike supervised CIR methods [1, 2, 11, 18], projection-based CIR methods [3, 12, 15, 23, 30, 33, 34] are built upon the frozen CLIP model, where a projection module " + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "content": " is trained without CIR triplets. Each method employs a different training scheme for different training schemes for " + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 91, + 297, + 283 + ], + "type": "text", + "content": " (See Sec. 4.1 for details). Several approaches avoid the need for any training by employing interpolation techniques [16], while others leverage powerful yet resource-intensive models such as LLMs and captioners [17, 39]. While our approach is built upon projection-based CIR methods, its training strategy is closely related to another category of CIR methods [14, 19, 37, 40] that use synthetically generated expensive CIR triplets. However, our method stands out by utilizing text triplets and updating only the text encoder, resulting in more efficient training while maintaining strong performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "text", + "content": "Task discrepancy between the CLIP pre-training task and CIR. Combiner [2] updates the text encoder to minimize the gap between the target caption feature and the sum of the reference image and conditioning text features. FashionERN [8] addresses reference image dominance by introducing a separate branch to amplify the impact of the conditioning text. However, both Combiner and FashionERN require expensive CIR triplets " + }, + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "inline_equation", + "content": "(I_r, T_c, I_t)" + }, + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "text", + "content": " for training, whereas our approach uses cheap, automatically generated text-only triplets " + }, + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "inline_equation", + "content": "(T_r, T_c, T_t)" + }, + { + "bbox": [ + 55, + 297, + 298, + 620 + ], + "type": "text", + "content": ". As another example, Chen and Lai [6] synthesize a triplet of an original image, its caption, and the masked image, treating them as the target image, conditioning text, and reference image, respectively. This approach, however, still has a gap between conditioning text (e.g., \"change dog to cat\") and image caption (e.g., \"a dog catching a frisbee\"); furthermore, it needs the full fine-tuning of the CLIP model, resulting in changing the visual embeddings in the retrieval database. In contrast, RTD directly uses the conditioning texts for training and does not change the target visual encoder, enabling the reuse of pre-extracted CLIP embeddings. Lastly, CIREVL [17] or LDRE [39] reduce task discrepancy by generating descriptive captions of the composed query using a large captioning model [21] and LLM [5]. While effective without any training, they rely on expensive inference steps and require carefully crafted prompts by skilled users. Whereas, RTD is fully automated, human-free, and more efficient during inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 628, + 142, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 142, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 142, + 641 + ], + "type": "text", + "content": "3. Main Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 647, + 183, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 183, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 183, + 661 + ], + "type": "text", + "content": "3.1. Obtaining text triplets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "To address the task discrepancy, we collect cheaply and automatically generated text triplets " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "(T_r, T_c, T_t)" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": ", instead of directly using the expensive CIR triplets " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "(I_r, T_c, I_t)" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": ". Given reference caption " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": " from conventional caption datasets," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "conditioning text " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " and target text " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " can be generated using two strategies: via large language models (LLMs) [4, 14, 19, 37] or, more efficiently, through rule-based templates [14]. We investigate both strategies and demonstrate that RTD consistently improves performance across them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "content": "For the LLM-based strategy, we use the publicly available text triplets by CompoDiff [14], which are employed in our main experiments if not specified. These triplets are generated by taking a given caption " + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "content": " as an input of the fine-tuned LLM, whose output predicts the corresponding conditioning text " + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "content": " and the target caption " + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "content": ". Previous works, such as IP2P [4], CoVR [37], and CASE [19], have also explored generating text triplets using LLMs, differing in LLM model types, input data, and fine-tuning strategies. The original purpose of these text triplet generations is to construct CIR triplets " + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "inline_equation", + "content": "(I_r, T_c, I_t)" + }, + { + "bbox": [ + 313, + 133, + 556, + 348 + ], + "type": "text", + "content": ", but all these works also release the corresponding text triplets used for their CIR triplet construction. In addition to these publicly available text triplets, we also implement and evaluate an efficient in-context learning generation strategy using LLaMA3-8B [13] without additional fine-tuning. We conduct experiments with all the aforementioned text triplets in Tab. 6 and observe that RTD consistently delivers significant enhancements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "text", + "content": "For cheap, rule-based strategy, we can extract a keyword (e.g., \"dog\") from " + }, + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "inline_equation", + "content": "T_{r}" + }, + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "text", + "content": " (\"a dog catching a frisbee\") and replace it with a randomly chosen keyword (e.g., \"cat\") [14], forming " + }, + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 313, + 349, + 557, + 469 + ], + "type": "text", + "content": ". The conditioning text is then generated automatically by using pre-defined templates (e.g., \"change [original keyword] to [altered keyword]\"). Our experiments show that this simple rule-based variant performs similarly to the LLM-based one. More explanations and examples of each strategy are provided in Suppl. A.2, and more details of generation costs are in Suppl. C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 518, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 518, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 518, + 491 + ], + "type": "text", + "content": "3.2. Frozen-target text contrastive learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": "Now, we explain our approach to update the text encoder for mitigating the task discrepancy solely with the generated text triplets " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "(T_r, T_c, T_t)" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": ". We first assume that there exists a pretrained projection module " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " obtained by the projection-based CIR methods [3, 15, 30]. Recall that for a given reference image " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " and conditioning text " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": ", the final composed feature is generated by passing the text prompt \"a photo of " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\phi(\\psi_V(I_r))" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " that " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": "\" to the text encoder " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\psi_V" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " is the visual encoder and " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " is the projection module (See Fig. 1). We aim to update the text encoder " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " to reduce the discrepancy between the pretext task and projection-based CIR task using the text triplets while maintaining " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\psi_V" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 498, + 556, + 640 + ], + "type": "text", + "content": " frozen." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "content": "Frozen-target text contrastive loss. We apply contrastive learning using a paired caption " + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "inline_equation", + "content": "(T_{r + c}, T_t)" + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "content": " denotes a concatenated caption of reference caption " + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "content": " and conditioning caption " + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 642, + 557, + 715 + ], + "type": "text", + "content": ". Namely, we let the representation of the concatenated caption closely approximate that of the target caption. However, solely updating the text encoder" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3897" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 553, + 261 + ], + "type": "image", + "image_path": "9a5e7d43e47e88ef4a5ae72637812a8905eee82d0d7055088e9ac20f4606a392.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 262, + 357, + 272 + ], + "lines": [ + { + "bbox": [ + 252, + 262, + 357, + 272 + ], + "spans": [ + { + "bbox": [ + 252, + 262, + 357, + 272 + ], + "type": "text", + "content": "Figure 2. Overview of RTD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": "while fixing the image encoder can break the alignment between image and text encoders. To prevent the issue, we extract the text embedding of " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": " using the frozen text encoder " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\psi_{T}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": ", while the concatenated caption " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": " is extracted from the learnable text encoder " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": ", initialized from " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\psi_{T}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": ". Here, we assume that as the target caption " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": " is a standard caption, a text embedding " + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\psi_{T}(T_{t})" + }, + { + "bbox": [ + 54, + 284, + 296, + 427 + ], + "type": "text", + "content": ", is well-aligned with the frozen image embedding space. Based on this assumption, we freeze the target textual embedding to maintain the pretrained alignment as we learn new relationships. As shown in Sec. 4.3, freezing the target embedding is crucial for fine-tuning with our objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": "Now, we introduce our frozen-target text contrastive loss " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FCL}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " using two text encoders: a frozen pre-trained text encoder " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " and a learnable text encoder " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " which is initialized with " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": ". Textual latent embeddings " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{t}_{r + c}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "t_t" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " are extracted from " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": ", respectively. Namely, " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{t}_{r + c} = \\psi_T^{tr}(E_w^{tr}(T_{r + c}))" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "t_t = \\psi_T(E_w(T_t))" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "E_w" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " is a word embedding layer. We aim to tune " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " to minimize the distance between the concatenated textual embedding " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\tilde{t}_{r + c}" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " and the target textual embedding " + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "inline_equation", + "content": "t_t" + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": " while maximizing the distance from other textual embeddings within the batch. We employ a symmetric InfoNCE loss [7, 10], as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 575, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 575, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 59, + 575, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {F C L} = \\frac {1}{B} \\sum_ {k = 1} ^ {B} - \\log \\left(\\frac {e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {k}) / \\tau)}}{\\sum_ {j = 1} ^ {B} e ^ {(c (\\tilde {t} _ {r + c} ^ {k} , t _ {t} ^ {j}) / \\tau)} + \\sum_ {j \\neq k} e ^ {(c (t _ {t} ^ {k} , t _ {t} ^ {j}) / \\tau)}}\\right) \\\\ - \\log \\left(\\frac {e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {k}\\right) / \\tau\\right)}}{\\sum_ {j = 1} ^ {B} e ^ {\\left(c \\left(t _ {t} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)} + \\sum_ {j \\neq k} e ^ {\\left(c \\left(\\tilde {t} _ {r + c} ^ {k} , \\tilde {t} _ {r + c} ^ {j}\\right) / \\tau\\right)}}\\right) \\tag {1} \\\\ \\end{array}", + "image_path": "c61da5565a29775044752058682e460f4c786dde33db6f42f9efd4a396c8b955.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "inline_equation", + "content": "c(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": " denotes the cosine similarity, " + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": " is the batch size, and " + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": " is a temperature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Refined batch sampling strategy for hard negatives. To" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": "further enhance the efficacy of updating the text encoder, we devise a simple yet effective batch sampling strategy that incorporates pairs of " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_{r + c}, T_t)" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_r, T_r)" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " within the same batch. For example, as presented in Fig. 2, a pair such as " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_{r + c}: \\text{“A dog catching a frisbee} + \\text{change dog to cat}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_t: \\text{“A cat catching a frisbee})" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " is sampled along with its corresponding reference pair " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_r: \\text{“A dog catching a frisbee}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_r: \\text{“A dog catching a frisbee})" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " in the same batch. This setup ensures that the concatenated text " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " and its corresponding reference text " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " implicitly act as hard negatives for each other, as their semantics are much more similar " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_{r + c}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_r)" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " than those of other randomly sampled texts in the batch. Explicitly distinguishing the embedding of " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " from that of " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " conceptually aligns well with CIR task, as it encourages the model to better capture modifications (in " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": "). Moreover, we believe including " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "(T_r, T_r)" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " pairs in the contrastive learning helps the learnable text encoder " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": " remain closely aligned with the pre-trained encoder " + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 313, + 284, + 555, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Refined concatenation of reference and conditioning texts. As we use \"a photo of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "[\\" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " ]$ that " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "[T_c]" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "\" for inference, a naive concatenation strategy also can suffer from training-inference task discrepancy. To tackle this issue, instead of simply concatenating the " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_c" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", we use the prompt \"a photo of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "[\\" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " ]$ that " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "[T_c]" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "\" for updating the text encoder, where " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "[\\" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " ]$ is obtained by the reference caption " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " with the projection module " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ". Instead of obtaining a pseudo-word token with latent image embedding " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", we utilize a textual latent embedding from the reference caption " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi (t_r)" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ". However, Gu et al. [15] showed that naively replacing the image encoder with the text encoder for the input of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " will suffer from the modality gap [22], a phenomenon where text and image embeddings have a gap between them. We tackle this issue by injecting random noise into the textual token representation before it is processed by " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", following Gu et al. [15]. More analyses on variations of noise are in Suppl. B.10." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3898" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Fig. 2 illustrates the overview of RTD. We use CLIP backbone and pre-trained projection module " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " produced by the existing projection-based CIR methods. The text encoder is trained using the proposed loss function (Eq. (1)) while applying the refined batch sampling and concatenation scheme. During inference, the procedure mirrors that of existing projection-based CIR methods, except we utilize the updated text encoder " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\psi_T^{tr}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " instead of the frozen one " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". Note that our method only updates the text encoder while the image encoder and the projection module are frozen." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "text", + "content": "Remark 1: Low data acquisition cost. The text triplets we use can be obtained at a significantly lower cost than CIR triplets. First, the text triplet generation process requires only an easily obtainable caption dataset for generation, whereas other approaches that generate CIR triplets [19, 37, 40] necessitate image or video datasets along with an additional collection phase for semantically similar images or videos. Second, text triplet generation can avoid resource-intensive text-to-image generation [4, 14], making it " + }, + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "inline_equation", + "content": "15 \\times" + }, + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "text", + "content": " faster than the CIR triplet generation process used in CompoDiff [14]. Furthermore, if we choose the rule-based text triplet generation strategy, the process becomes " + }, + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "inline_equation", + "content": "570 \\times" + }, + { + "bbox": [ + 56, + 193, + 297, + 397 + ], + "type": "text", + "content": " faster than the full CIR triplet generation with images; 1M text triplets generation takes just 0.1 hours. Lastly, the storage for the generated text triplets takes up only 100MB, whereas storing a similar quantity of images requires significantly more space (e.g., 400GB for CC3M [32]). More details can be found in Suppl. C." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 472, + 293, + 577 + ], + "blocks": [ + { + "bbox": [ + 55, + 404, + 296, + 470 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 470 + ], + "type": "text", + "content": "Table 1. Training and inference time comparisons. Purple denotes the additional resource required for RTD. Inference time is measured on a single A100 GPU for a single image. Note that the training hours for SEARLE, LinCIR for 8 A100 GPUs [15] and their RTD variants are trained on 4 A100 GPUs. CompoDiff is trained on 128 A100 and MagicLens is trained on 128 TPUs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 472, + 293, + 577 + ], + "lines": [ + { + "bbox": [ + 57, + 472, + 293, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 472, + 293, + 577 + ], + "type": "table", + "html": "
MethodCLIP BackboneTraining (h)Inference (s)
SEARLEL/144.2h0.02s
SEARLE (+ RTD)L/144.58h (+0.38h)0.02s
LinCIRL/140.5h0.02s
LinCIR (+ RTD)L/140.88h (+0.38h)0.02s
LinCIRG/140.8h0.05s
LinCIR (+ RTD)G/143.3h (+2.5h)0.05s
CompoDiffL/14231h0.12s
MagicLensL/146hunknown
CIRVLL/14-3.02s
", + "image_path": "16d49cb5275da5096f940ef47c9ae4163ec2b0f16a6da7a7a7ee4611e6c2b230.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Remark 2: Training and inference efficiency. As reported in Tab. 1, the additional training cost of RTD is small; 0.38 hours for ViT-L/14 on 4 A100 GPUs. In contrast, CompoDiff and MagicLens require 231 hours on 128 A100 GPUs and 6 hours on 128 TPUs, respectively, making the cost of RTD negligible. Even with a larger ViT-G/14 backbone, training RTD takes only 2.5 hours on 4 A100 GPUs, still faster than synthetic CIR triplet-based methods and even comparable to SEARLE in ViT-L/14 (4.2h). This training efficiency mainly stems from the advantage of text-only training. As highlighted by Gu et al. [15], the training complexity for the text" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "encoder is much lower than that for the visual encoder due to the relatively short token lengths of texts (" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\approx 12" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ") compared to images (256). The average inference time of the CLIP ViT-L/14 image encoder is " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\times 3.5" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " times slower than that of the text encoder. Another source of efficiency is that RTD requires relatively few iterations (at most 2000, and significantly fewer in most cases), as the text encoders are already pre-trained and only require minor adjustments to learn the relationships between text triplets. Moreover, in Suppl. B.4, we present a more efficient implementation by selectively updating only a few layers of the text encoder. Along with efficient training, RTD retains the fast inference speed of the original projection-based methods, remaining " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "150\\times" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " faster than training-free CIREVL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 254, + 395, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 395, + 267 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 395, + 267 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 274, + 429, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 429, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 429, + 288 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": "Implementation details. We use the AdamW optimizer [25] with a weight decay of 0.01. The learning rate is set to " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": ", with a batch size of 512. We select the text encoder model with the best zero-shot CIRR [24] dev R@1 score for evaluating RTD. We mainly use the visual and textual encoders of the CLIP ViT-L/14 [29] as our backbone. Unless otherwise noted, we use LLM-based 2.5M text triplets provided by CompoDiff [14] for the training. We set the " + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 293, + 556, + 449 + ], + "type": "text", + "content": " as 0.07 in Eq. (1) and scale the standard deviation of Gaussian distribution as 0.5 for the noise injection. More results on various noise distributions can be found in the Suppl. B.10. All experiments were conducted using four NVIDIA A100 GPUs with Python 3.8 and Pytorch [28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "content": "Evaluation datasets and metrics. We compare projection-based CIR methods on five benchmark datasets: CIRR [24], CIRCO [3], FashionIQ [38] (FIQ), COCO object composition [30], and GeneCIS [36]. Details of each dataset are in the Suppl. A.1. For CIRR, FIQ, COCO, and GeneCIS, we have reported their recall scores at the top K retrieval results " + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{K})" + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "content": ". Since the CIRCO dataset includes multiple positive images for each query, we use a ranking-based metric; mean Average Precision scores at the top K results " + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}@\\mathrm{K})" + }, + { + "bbox": [ + 313, + 449, + 556, + 616 + ], + "type": "text", + "content": " [9, 26]. For the main results, we compare the results on the FIQ validation split, as well as the test sets of CIRR and CIRCO. For the ablation studies and analyses, the validation splits of these three datasets are utilized. GeneCIS and COCO object composition results are in the Suppl. B.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Baselines. We evaluate the effect of our method when combined with publicly available projection-based ZS-CIR methods: Pic2Word [30], SEARLE [3], LinCIR [15], Context-I2W [34], and FTI4CIR [23]. All these methods share a similar core concept shown in Fig. 1, but use different training schemes. Pic2Word [30] optimizes contrastive loss between the image embedding and its projected text embedding of \"a photo of [\\$\" to obtain the projection module " + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": ". Similarly," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3899" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 552, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Comparison with other baselines. Note that this comparison is not entirely fair due to differences in backbone models and training data. Purple denotes the performance gain from our method, while red and blue highlight the best and second-best scores, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 552, + 342 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 552, + 342 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 552, + 342 + ], + "type": "table", + "html": "
MethodBackboneCIRRCIRCOFashionIQ
R@1R@5R@10mAP@5mAP@10mAP@25R@10R@50
Pic2Word [30]CLIP ViT-L/1424.251.564.18.39.110.125.344.9
+RTD27.9 (+3.6)56.2 (+4.8)68.5 (+4.4)9.1 (+0.9)9.6 (+0.5)10.7 (+0.6)27.6 (+2.3)48.9 (+4.0)
SEARLE [3]CLIP ViT-L/1424.952.365.711.612.714.325.045.3
+RTD26.6 (+1.7)56.2 (+3.9)69.0 (+3.3)16.5 (+4.9)17.9 (+5.2)19.8 (+5.4)29.3 (+4.4)50.7 (+5.4)
Context-I2W [34]CLIP ViT-L/1425.655.468.6---27.949.1
+RTD29.2 (+4.6)58.4 (+3.0)70.5 (+1.9)---28.1 (+0.2)49.5 (+0.4)
FTI4CIR [23]CLIP ViT-L/1425.955.667.715.116.318.129.450.9
+RTD29.1 (+3.2)58.8 (+3.2)70.4 (+2.7)16.2 (+1.1)17.4 (+1.1)19.4 (+1.3)30.6 (+1.2)51.7 (+0.8)
LinCIR [15]CLIP ViT-L/1423.852.966.513.014.115.827.447.7
+RTD26.6 (+2.9)56.2 (+3.3)69.0 (+2.5)17.1 (+4.1)18.1 (+4.0)20.1 (+4.3)30.2 (+2.8)51.1 (+3.4)
LinCIR [15]FSC-CLIP [27] ViT-L/1433.663.174.514.215.417.028.349.5
+RTD35.9 (+2.3)67.5 (+4.4)78.1 (+3.6)18.3 (+4.1)19.7 (+4.3)21.3 (+4.3)31.2 (+2.9)52.3 (+2.8)
KEDs [33]CLIP ViT-L/1426.454.867.2---26.847.9
CIREVL [17]CLIP ViT-L/1424.552.364.918.619.020.928.648.6
LDRE [39]CLIP ViT-L/1426.555.667.523.424.026.428.550.5
MT-CIR [6]CLIP ViT-L/1425.554.667.610.411.613.035.457.4
MagicLens [40]CLIP ViT-L/1430.161.774.429.630.833.430.752.5
Compodiff [14]CLIP ViT-L/1426.755.072.612.613.415.836.048.6
CoVR [37]BLIP ViT-L/1439.368.278.9---27.744.6
CASE [19]BLIP ViT-L/1435.465.878.5-----
LinCIR [15]CLIP ViT-G/1434.964.576.120.621.924.144.565.5
+RTD36.3 (+1.4)67.5 (+3.0)78.3 (+2.2)21.1 (+0.5)22.3 (+0.4)24.5 (+0.4)46.2 (+1.7)67.3 (+1.8)
Compodiff [17]CLIP ViT-G/1434.764.375.115.317.719.439.051.7
CIREVL [17]CLIP ViT-G/1426.755.174.526.827.630.032.252.4
LDRE [39]CLIP ViT-G/1436.266.477.331.132.235.032.555.4
", + "image_path": "5ca1cfcd261a71d8b274f23d493b35eb285070dd1ebe44bd2e9d0e43bbe16808.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "text", + "content": "SEARLE [3] employs a two-stage approach, starting with an optimization-based textual inversion phase followed by a distillation phase for the projection module " + }, + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "text", + "content": ". LinCIR [15] introduces a language-only self-supervised task based on keyword token replacement, where the model encourages the original text embedding and the replaced text embedding (keyword tokens are replaced with the projection of the original text embedding via " + }, + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 350, + 297, + 507 + ], + "type": "text", + "content": ") to be aligned. Context-I2W [34] refines the projection module by selecting relevant visual information [34]. Lastly, FTI4CIR [23] separately maps images into subject- and attribute-oriented pseudo-word tokens [23]. Details on how RTD is combined with them can be found in Suppl. A.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 298, + 714 + ], + "type": "text", + "content": "We train these methods with the CLIP ViT-L/14 in our main experiments. To further verify the compatibility of RTD, we also report results for the ViT-B/32 backbone with SEARLE, Pic2Word, and LinCIR. Moreover, we conduct experiments with a larger backbone (ViT-G/14) and a fine-tuned CLIP model for enhanced compositionality (FSC-CLIP [27]) for LinCIR, enabled by its fast training capability. We use the publicly available pre-trained model for SEARLE (ViT-B/32, ViT-L/14), Pic2Word (ViT-L/14), Context-I2W (ViT-L/14), and FTI4CIR (ViT-L/14). Otherwise, we reproduce the results using the official implementation. When reproducing, we adhere to the same settings in the original papers. For example, we select the final last epoch model for the Pic2Word ViT-B/32 model and choose the model based on the best zero-shot CIRR dev R@1 score for LinCIR. We additionally compare our method with a diverse set of CIR approaches, including another projection-based" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 350, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 411 + ], + "type": "text", + "content": "ZS-CIR method (KEDs [33]), another attempt to address task discrepancy (MT-CIR [6]), approaches leveraging synthetically generated CIR triplets (CoVR [37], CASE [19], MagicLens [40] and Compodiff [14]), and the training-free methods (CIREvL [17], LDRE [39])." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 419, + 397, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 397, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 397, + 431 + ], + "type": "text", + "content": "4.2. Main results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "type": "text", + "content": "Tab. 2 shows the overview of comparison results with state-of-the-art CIR methods. First, we assess the impact of integrating RTD with existing projection-based CIR methods (SEARLE, Pic2Word, LinCIR, Context-I2W, and FTI4CIR) across CIRR, CIRCO, and FashionIQ. RTD consistently boosts performance, yielding an average improvement of over 2.78 points. This trend also holds for the GeneCIS and COCO object composition datasets, as detailed in Suppl. B.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "text", + "content": "Second, we compare the integration of RTD with those leveraging synthetically generated CIR triplets (CoVR, CASE, MagicLens, and Compodiff). We focus on this comparison to ensure fairness, as leveraging text triplets and updating the textual backbone of RTD may not be fully aligned with the standard projection-based CIR setting. We observe that RTD delivers competitive or superior performance compared to these resource-intensive synthetic CIR triplet-based methods, while being significantly more efficient; over " + }, + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "10 - 100 \\times" + }, + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "text", + "content": " faster, as noted in Remark-2 of Sec. 3. For example, with the same CLIP ViT-L/14 backbone, LinCIR + RTD outperforms MagicLens and CoVR in FashionIQ R@10 and R@50 while FTI4CIR + RTD or Context-I2W + RTD achieve performance on par with or exceeding that of MagicLens, MT-CIR, and Compodiff in CIRR metrics." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3900" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 167 + ], + "type": "text", + "content": "When combined with larger (ViT-G/14) or fine-tuned (FSC-CLIP) backbones, RTD achieves the best or second-best result on the CIRR and FashionIQ benchmarks. We believe this flexibility of using different backbones underscores the benefit of efficient text-only training of RTD, inherited by LinCIR [15]. We reemphasize that even with a larger backbone like ViT-G/14, RTD remains much faster in training than other synthetic triplet-based methods, as shown in Tab. 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 170, + 296, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 170, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 170, + 296, + 243 + ], + "type": "text", + "content": "Lastly, across both backbones (ViT-L/14, G/14), LinCIR + RTD mostly outperforms training-free CIREVL and LDRE, except on CIRCO metrics. As shown in Tab. 1, CIREVL is significantly slower " + }, + { + "bbox": [ + 54, + 170, + 296, + 243 + ], + "type": "inline_equation", + "content": "(150\\times)" + }, + { + "bbox": [ + 54, + 170, + 296, + 243 + ], + "type": "text", + "content": " than projection-based methods (including RTD), while LDRE incurs even larger inference time due to its ensemble-based strategy." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 330, + 293, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 296, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 328 + ], + "type": "text", + "content": "Table 3. Ablation study. Unlike in Tab. 2, for ablation studies and analyses, validation splits of three CIR datasets are used for evaluation. We measure the impact of FCL loss (Eq. (1)), refined batch sampling (RB), and refined concatenation scheme (RC). All models are based on LinCIR ViT-L/14. The first row denotes the vanilla LinCIR without RTD. \"Avg\" denotes the average of all reported metrics. Bold indicates the best result." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 330, + 293, + 407 + ], + "lines": [ + { + "bbox": [ + 57, + 330, + 293, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 330, + 293, + 407 + ], + "type": "table", + "html": "
TextFCL Frozen TargetRBRCCIRR (R) @5@10CIRCO (mAP) @10@25FIQ (R) @10@50Avg
--XX54.367.812.714.527.447.737.4
\\( \\left( {{T}_{r},{T}_{r}}\\right) \\)✓XX56.069.713.415.228.248.838.5
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓XX58.271.514.416.026.947.939.2
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓X58.271.315.016.727.449.339.6
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)X✓X54.367.012.213.625.045.336.3
\\( \\left( {{T}_{r + c},{T}_{t}}\\right) \\)✓✓✓57.971.116.117.830.251.140.7
", + "image_path": "3144acdbf4109eb5f4341a17ffc3a9784c9901afdf42918cdfedbf17c9226343.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 154, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 154, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 154, + 441 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": "Tab. 3 presents the effectiveness of the proposed components: frozen-target text contrastive loss (FCL), refined batch sampling (RB), and refined concatenation scheme (RC). All evaluation results are on the validation splits. All model variants use ViT-L/14 and a projection module " + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": " from LinCIR, making the results in row 1 indicative of the original performance of LinCIR. We first compare the impact of the text pairs fed into FCL loss. We compare our design choice " + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_{r + c}, T_t)" + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": " from the generated text triplets with " + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "inline_equation", + "content": "(T_r, T_r)" + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": ", which is the sole option for constructing a pair given a single conventional caption " + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_r" + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": ". The results demonstrate that, on average, using generated triplets (3rd row) is more effective than using original conventional text pairs (2nd row), particularly in the CIRR and CIRCO datasets. In addition, RB (4th row) and RC (6th row) significantly enhance the overall performance, demonstrating the effectiveness of these components. Finally, we measure the impact of using the frozen target textual embedding, denoted as \"Frozen Target\" in the table. Significant performance degradation is observed when the learnable text encoder is used for extracting the embedding of the target caption " + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "inline_equation", + "content": "T_t" + }, + { + "bbox": [ + 55, + 450, + 297, + 714 + ], + "type": "text", + "content": " (5th row) compared to the frozen-target case (4th row), supporting the importance of our design choice." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 110, + 524, + 171 + ], + "blocks": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "lines": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "type": "text", + "content": "Table 4. T2I retrieval performance of different text encoders on CIRCO val set. \"Update (pair)\" refers to the setting in the second row of Tab. 3, which uses contrastive learning with the pair " + }, + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "type": "inline_equation", + "content": "(T_r, T_r)" + }, + { + "bbox": [ + 313, + 65, + 555, + 109 + ], + "type": "text", + "content": ". Bold indicates the best result, excluding the first row (oracle case)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 110, + 524, + 171 + ], + "lines": [ + { + "bbox": [ + 340, + 110, + 524, + 171 + ], + "spans": [ + { + "bbox": [ + 340, + 110, + 524, + 171 + ], + "type": "table", + "html": "
QueryText encodermAP@5mAP@10mAP@25
\\(T_t\\)Frozen18.9619.3121.05
\\(T_{r+c}\\)Frozen10.1210.7112.34
\\(T_{r+c}\\)Update (pair)10.5211.1512.72
\\(T_{r+c}\\)RTD15.1215.8017.77
", + "image_path": "51ce49fa1ce222417919520fc9b3b2511c0b99d7a07ab348d441162a59cba52a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 178, + 488, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 488, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 488, + 190 + ], + "type": "text", + "content": "4.4. Analyses on our core motivation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 197, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 555, + 232 + ], + "type": "text", + "content": "We conduct experiments to validate our core motivation (reducing task discrepancy) and verify that the observed gains stem from it. Other details follow Sec. 4.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": "Can RTD really reduce the task discrepancy of the text encoder? We first quantitatively verify whether RTD indeed reduces the task discrepancy. We first conduct a controlled experiment that measures the text-to-image (T2I) retrieval performance of the text encoder with conditional texts. We retrieve the target images " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " with the concatenated text query " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " or the ideal target caption " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ". If our text encoder successfully handles the discrepancy due to the concatenated caption, the text encoder updated by RTD will perform better than the frozen one or \"Update (pair)\", which is updated by contrastive learning with the " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "(T_{r}, T_{r})" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " pair (corresponding to row 2 of Tab. 3). We use the CLIP ViT-L/14 and CIRCO [3] validation set for evaluation. Since the CIRCO dataset only has CIR triplets " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "(I_{r}, T_{c}, I_{t})" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ", we use the BLIP [20] captioner to generate " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{r}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ", respectively. Here, the simple concatenation scheme is applied for the text query " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " in all cases. Tab. 4 shows that when the text encoder is either frozen or updated with " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "(T_{r}, T_{r})" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ", the retrieval results using the concatenated caption " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{r + c}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " are significantly worse than those using the target caption " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "T_{t}" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ". It supports the claim that the frozen text encoder suffers from the negative effects of task discrepancy between the pretext and CIR tasks. Moreover, it suggests that simply updating the text encoder with " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "(T_{r}, T_{r})" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": " pair fails to reduce this discrepancy. In contrast, the text encoder updated by RTD shows a significant improvement over the frozen text encoder or one updated with " + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "inline_equation", + "content": "(T_{r}, T_{r})" + }, + { + "bbox": [ + 313, + 233, + 556, + 567 + ], + "type": "text", + "content": ", showing that it successfully reduces the task discrepancy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": "We additionally measure the average cosine similarity between the composed textual features with the prompt \"a photo of " + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\phi (\\psi_V(I_r))" + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": " that " + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "inline_equation", + "content": "T_{c}" + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": "\" (Fig. 1a) and the target image features (Fig. 1c). The similarity is measured by the LinCIR ViT-L/14 model on the CIRCO validation split. When we use the frozen text encoder " + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "inline_equation", + "content": "(\\psi_T)" + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": ", the average similarity is 0.1. By changing the text encoder to our updated text encoder " + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "inline_equation", + "content": "(\\psi_T^{tr})" + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": ", the similarity becomes 0.29 (+0.19). This result shows again that RTD successfully aligns the composed query features using " + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 569, + 556, + 689 + ], + "type": "text", + "content": " to the frozen CLIP image features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Is all the gain of RTD from naive text-encoder updating? To verify that our improvements cannot be achieved solely" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3901" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "by updating the text encoder backbone without considering the task discrepancy, we additionally measure the results of previous methods (Pic2Word and LinCIR) when naively updating the text encoders. Namely, after training " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " while keeping all other networks frozen as in previous methods, we additionally update the text encoder using the original loss function, while fixing other modules including " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ". We denote this update rule as \"naive\" in the Tab. 5. Unlike RTD, we observe that just naively updating the text encoder (\"naive\") significantly degrades the performance of the baseline. The results indicate that merely updating the text backbone is not beneficial for CIR; instead, mitigating task discrepancy through RTD is necessary." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 296, + 293, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "type": "text", + "content": "Table 5. Impact of the update scheme. Two update schemes are compared: (1) using the original objective from baseline (\"naive\") and (2) using RTD. For a fair comparison, in both schemes, " + }, + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "type": "text", + "content": " is updated first and " + }, + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\psi_T" + }, + { + "bbox": [ + 55, + 238, + 296, + 293 + ], + "type": "text", + "content": " is updated top on the frozen modules. Other details are the same as Tab. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 296, + 293, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 296, + 293, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 296, + 293, + 392 + ], + "type": "table", + "html": "
CIRR (R)CIRCO (mAP)FIQ (R)Avg
@5@10@10@25@10@50
Pic2Word51.464.48.810.125.344.932.2
+naive19.227.51.31.64.411.210.9
+ RTD56.669.88.89.827.648.936.9
LinCIR54.367.812.714.527.447.736.9
+naive52.766.811.413.026.345.935.5
+ RTD57.971.116.117.830.251.140.7
", + "image_path": "bb8535e31adb1e09390a1af50f37a0bd6e3345d9dfd326ae372aabc0f9f3112d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "content": "Next, instead of using the original objective from baselines, we update the text encoder with conventional contrastive learning using the pair " + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "inline_equation", + "content": "(T_r, T_r)" + }, + { + "bbox": [ + 55, + 401, + 296, + 521 + ], + "type": "text", + "content": ", which corresponds exactly to row 2 in Tab. 3 (row 3 in Tab. 4). This setup can also serve as a proxy for the standalone impact of breaking the alignment of pre-trained knowledge, as noted in [2]. As demonstrated in Sec. 4.3, the large gap (more than 2 points) between RTD (last row in Tab. 3) and this proxy (row 2 in Tab. 3) again shows that our gain cannot be attributed to this simple alternative (naïve text encoder tuning)." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 527, + 294, + 641 + ], + "blocks": [ + { + "bbox": [ + 57, + 527, + 294, + 641 + ], + "lines": [ + { + "bbox": [ + 57, + 527, + 294, + 641 + ], + "spans": [ + { + "bbox": [ + 57, + 527, + 294, + 641 + ], + "type": "image", + "image_path": "74c016a39698e76fab097c94a755f2f123fedb0780cb4d21b0a88e3882babda9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 642, + 296, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 696 + ], + "type": "text", + "content": "Figure 3. Impact of sizes of backbone. The results of RTD combined with Pic2Word [30], SEARLE [3], and LinCIR [15] across different CLIP backbones (ViT-B/32 and ViT-L/14) are shown. Here, the score is the same metric in \"Avg\" in Tab. 3 and other details are the same as Tab. 3. Full results are in the Suppl. B.2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 506, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 506, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 506, + 85 + ], + "type": "text", + "content": "4.5. Compatibility across backbone sizes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 555, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 210 + ], + "type": "text", + "content": "In the Fig. 3, we observe that the integration of our approach with projection-based CIR methods significantly improves the performance across pioneering projection-based ZS-CIR methods (SEARLE, Pic2Word, and LinCIR) and all backbones (ViT-B/32 and ViT-L/14). For example, regardless of the choice of projection module " + }, + { + "bbox": [ + 313, + 89, + 555, + 210 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 89, + 555, + 210 + ], + "type": "text", + "content": " and backbones, the minimum performance gain for average scores is greater than 2.8 points. The results of RTD using the larger backbone (ViTG/14) can be found in Tab. 2. More detailed and qualitative results are provided in Suppl. B.3 and Suppl. D, respectively." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 272, + 552, + 352 + ], + "blocks": [ + { + "bbox": [ + 313, + 215, + 556, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 556, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 556, + 270 + ], + "type": "text", + "content": "Table 6. The effectiveness of different types of text triplets for RTD. \"In-context\" denotes an efficient implementation using in-context learning with LLaMA3-8B, without fine-tuning. Details and examples of each text triplet dataset are summarized in Tab. S1 and Tab. S2, respectively. Other details are the same as Tab. 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 272, + 552, + 352 + ], + "lines": [ + { + "bbox": [ + 316, + 272, + 552, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 552, + 352 + ], + "type": "table", + "html": "
SourceLLMCIRR (R) @5 @10CIRCO (mAP) @10 @25FIQ (R) @10 @50Avg
LinCIR--54.367.812.714.527.447.737.4
+RTDRule-basedX56.770.315.017.030.451.940.2 (+ 2.9)
IP2P [4]✓58.771.615.918.029.650.740.7 (+ 3.3)
Compodiff [14]✓57.971.116.117.830.251.140.7 (+ 3.3)
In-context✓59.371.815.817.529.751.440.9 (+ 3.5)
CoVR [37]✓59.872.615.417.029.650.841.9 (+ 4.5)
CASE [19]✓56.369.311.112.726.647.837.7 (+ 0.3)
", + "image_path": "5e7d165879cdac16ead15ff572085d2f4ef69d12bb52ad018af05690234ccac7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 552, + 367 + ], + "type": "text", + "content": "4.6. Impact of the text triplet generation strategies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 371, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 587 + ], + "type": "text", + "content": "As explained in Sec. 3.1, we evaluate RTD using both 1) publicly available LLM-based text triplets (from IP2P, Compodiff, CoVR, and CASE) along with efficient in-context learning-based text triplets, and 2) LLM-free, rule-based triplets. Tab. 6 shows that RTD consistently improves CIR performance across them (+3.3 for IP2P, +3.3 for Compodiff, +3.5 for in-context learning, +4.5 for CoVR, and +0.31 for CASE, and rule-based triplets achieve 2.9, respectively). We believe this result shows the reproducibility and consistency of RTD, with the rule-based triplets performing comparably to LLM-generated ones, indicating that efficient rule-based triplets are sufficient to achieve strong CIR performance. The marginal improvement in CASE is largely due to the poor quality of text triplets resulting from its construction pipeline that prioritizes CIR triplet quality over text triplet quality, as shown in Tab. S1. Further details can be found in Suppl. A.2, and additional analyses, including data scales related to text triplets, are provided in Suppl. B.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "We presented RTD, a novel post-processing approach that can be easily integrated into existing projection-based CIR methods, aimed at reducing task discrepancy of text encoders. Empirical evaluations demonstrate that RTD significantly boosts the performance of existing projection-based CIR methods across diverse datasets and model backbones, competing with or outperforming other resource-intensive CIR methods with much greater efficiency." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3902" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "content": "This work was supported in part by National Research Foundation of Korea (NRF) grant [No. 2021R1A2C2007884, No. RS-2025-02263628], the Institute of Information & communications Technology Planning & Evaluation (IITP) grants [RS-2021-II212068, RS-2022-II220113, RS-2022-II220959, RS-2021-II211343], and the BK21 FOUR Education and Research Program for Future ICT Pioneers (Seoul National University), funded by the Korean government (MSIT)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 217, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 217, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 251 + ], + "type": "text", + "content": "[1] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining clip-based features. In CVPR, 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 295 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Conditioned and composed image retrieval combining and partially fine-tuning clip-based features. In CVPR, 2022. 2, 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 329 + ], + "type": "text", + "content": "[3] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In ICCV, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 363 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, 2023. 1, 3, 5, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 408 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 442 + ], + "type": "text", + "content": "[6] Junyang Chen and Hanjiang Lai. Pretrain like you inference: Masked tuning improves zero-shot composed image retrieval. arXiv preprint arXiv:2311.07622, 2023. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 476 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 478, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 520 + ], + "type": "text", + "content": "[8] Yanzhe Chen, Huasong Zhong, Xiangteng He, Yuxin Peng, Jiahuan Zhou, and Lele Cheng. Fashionern: Enhance-and-refine network for composed fashion image retrieval. In AAAI, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[9] Sanghyuk Chun, Wonjae Kim, Song Park, Minsuk Chang, and Seong Joon Oh. Eccv caption: Correcting false negatives by collecting machine-and-human-verified image-caption associations for ms-coco. In ECCV, 2022. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 600 + ], + "type": "text", + "content": "[10] Niv Cohen, Rinon Gal, Eli A Meirom, Gal Chechik, and Yuval Atzmon. “this is my unicorn, fluffy”: Personalizing frozen vision-language representations. In ECCV, 2022. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "text", + "content": "[11] Ginger Delmas, Rafael S Rezende, Gabriela Csurka, and Diane Larlus. Artemis: Attention-based retrieval with text-explicit matching and implicit similarity. In ICLR, 2022. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[12] Yongchao Du, Min Wang, Wengang Zhou, Shuping Hui, and Houqiang Li. Image2sentence based asymmetrical zero-shot composed image retrieval. In ICLR, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "text", + "content": "llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[14] Geonmo Gu, Sanghyuk Chun, HeeJae Jun, Yoohoon Kang, Wonjae Kim, and Sangdoo Yun. Compodiff: Versatile composed image retrieval with latent diffusion. TMLR, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "text", + "content": "[15] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, , Yoonhoon Kang, and Sangwoo Yun. Language-only efficient training of zero-shot composed image retrieval. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 228 + ], + "type": "text", + "content": "[16] Young Kyun Jang, Dat Huynh, Ashish Shah, Wen-Kai Chen, and Ser-Nam Lim. Spherical linear interpolation and text-anchoring for zero-shot composed image retrieval. In ECCV, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 230, + 555, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 263 + ], + "type": "text", + "content": "[17] Shyamgopal Karthik, Karsten Roth, Massimiliano Mancini, and Zeynep Akata. Vision-by-language for training-free compositional image retrieval. In ICLR, 2024. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 265, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 555, + 296 + ], + "type": "text", + "content": "[18] Seungmin Lee, Dongwan Kim, and Bohyung Han. Cosmo: Content-style modulation for image retrieval with text feedback. In CVPR, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 330 + ], + "type": "text", + "content": "[19] Matan Levy, Rami Ben-Ari, Nir Darshan, and Dani Lischinski. Data roaming and quality assessment for composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 365 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 366, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 555, + 408 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 453 + ], + "type": "text", + "content": "[22] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In NeurIPS, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 498 + ], + "type": "text", + "content": "[23] Haoqiang Lin, Haokun Wen, Xuemeng Song, Meng Liu, Yu-peng Hu, and Liqiang Nie. Fine-grained textual inversion network for zero-shot composed image retrieval. In SIGIR, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "text", + "content": "[24] Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pretrained vision-and-language models. In ICCV, 2021. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 535, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 554, + 555 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 557, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 578 + ], + "type": "text", + "content": "[26] Kevin Musgrave, Serge Belongie, and Ser-Nam Lim. A metric learning reality check. In ECCV, 2020. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 623 + ], + "type": "text", + "content": "[27] Youngtaek Oh, Jae Won Cho, Dong-Jin Kim, In So Kweon, and Junmo Kim. Preserving multi-modal capabilities of pretrained vlms for improving vision-linguistic compositionality. In EMNLP, 2024. 2, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 678 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3903" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 587 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[30] Kuniaki Saito, Kihyuk Sohn, Xiang Zhang, Chun-Liang Li, Chen-Yu Lee, Kate Saenko, and Tomas Pfister. Pic2word: Mapping pictures to words for zero-shot composed image retrieval. In CVPR, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 205 + ], + "type": "text", + "content": "[31] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS Dataset and Benchmark, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 250 + ], + "type": "text", + "content": "[32] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 285 + ], + "type": "text", + "content": "[33] Yucheng Suo, Fan Ma, Linchao Zhu, and Yi Yang. Knowledge-enhanced dual-stream zero-shot composed image retrieval. In CVPR, 2024. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[34] Yuanmin Tang, Jing Yu, Keke Gai, Jiamin Zhuang, Gang Xiong, Yue Hu, and Qi Wu. Context-i2w: Mapping images to context-dependent words for accurate zero-shot composed image retrieval. In AAAI, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "type": "text", + "content": "[35] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 418 + ], + "type": "text", + "content": "[36] Sagar Vaze, Nicolas Carion, and Ishan Misra. Genecis: A benchmark for general conditional image similarity. In CVPR, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 295, + 453 + ], + "type": "text", + "content": "[37] Lucas Ventura, Antoine Yang, Cordelia Schmid, and Gül Varol. CoVR: Learning composed video retrieval from web video captions. In AAAI, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 498 + ], + "type": "text", + "content": "[38] Hui Wu, Yupeng Gao, Xiaoxiao Guo, Ziad Al-Halah, Steven Rennie, Kristen Grauman, and Rogerio Feris. Fashion iq: A new dataset towards retrieving images by natural language feedback. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 542 + ], + "type": "text", + "content": "[39] Zhenyu Yang, Dizhan Xue, Shengsheng Qian, Weiming Dong, and Changsheng Xu. Ldre: Llm-based divergent reasoning and ensemble for zero-shot composed image retrieval. In SIGIR, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 587 + ], + "type": "text", + "content": "[40] Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions. In ICML, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3904" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_content_list.json b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0bebeaeceb89ff358fee56f8f30bed6e278e8818 --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_content_list.json @@ -0,0 +1,1635 @@ +[ + { + "type": "text", + "text": "An Empirical Study of Autoregressive Pre-training from Videos", + "text_level": 1, + "bbox": [ + 178, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jathushan Rajasegaran1,2 Ilija Radosavovic2 Rahul Ravishankar2 Yossi Gandelsman2 \nChristoph Feichtenhofer1 Jitendra Malik1,2 \n1Meta FAIR 2University of California, Berkeley", + "bbox": [ + 133, + 179, + 846, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We empirically study autoregressive pre-training from unlabeled videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pretrained on a diverse dataset of unlabeled videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotic manipulation. Our results demonstrate that, despite minimal inductive biases, autoregressive video pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. https://brjathu.github.io/toto", + "bbox": [ + 88, + 300, + 485, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 571, + 220, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a paper published in 1951, Shannon, having just published the foundational papers of information theory, proposed a \"guessing game\" of next word prediction to estimate the entropy of English [51]. Nearly 70 years later, training a high-capacity transformer neural network [58] on this task, provided the generative pre-training backbone for Large Language Models [5, 13, 40, 41].", + "bbox": [ + 89, + 597, + 482, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Less well known is the fact that in 1954, Fred Attneave [2] proposed an analog of Shannon's task for images. To quote \"We may divide the picture into arbitrarily small elements which we \"transmit\" to a subject (S) in a cumulative sequence, having them guess at the color of each successive element until they are correct. This method of analysis resembles the scanning process used in television and facsimile systems and accomplishes the like purpose of transforming two spatial dimensions into a single sequence in time\". Of course this leaves open the question of what is the \"element\". Is it a pixel? A patch? A semantic entity?", + "bbox": [ + 89, + 704, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While Attneave was concerned with images, in the context of 2024, we have to note the \"Big Visual Data\" is in", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "video, not images. According to the web, YouTube alone has 156 million hours of video, and 30 thousand hours of video are being uploaded every hour. While there are concerns that most of the text available on the Internet has already been used by the leading language models, in video we are barely started on the journey of big data exploitation.", + "bbox": [ + 511, + 268, + 903, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning from this data brings a number of challenges. The vast majority of the video data is unlabeled, or at least not consistently labeled. Therefore, an effective learning approach should be able to learn from pixels alone. Indeed, human children form rich representations and models of the world purely from observing raw video [38], which further motivates the setting of learning from unlabeled videos. What is the right video architecture and the training objective? And what is the right visual \"element\" to operate over?", + "bbox": [ + 511, + 363, + 905, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we empirically study autoregressive pretraining from unlabeled videos. To perform our empirical study, we construct a family of autoregressive video models which we call Toto. We treat videos as sequences of visual tokens and train a transformer model on the next token prediction task. We use a causal transformer model with improvements from the LLaMa [55] architecture. We use dVAE [45] to tokenize frames into discrete tokens. Treating videos as sequences of tokens enables us to jointly train on both videos and images using a unified format. We construct a diverse dataset of videos and images comprising over 1 trillion visual tokens. Our models are first pretrained on this dataset and then evaluated on various downstream tasks. We extract visual representations using attention pooling from relevant layers of the model.", + "bbox": [ + 511, + 502, + 905, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We evaluate our models on a range of downstream tasks from image and video recognition, video forecasting, semi-supervised tracking, object permanence and robotic manipulation tasks in both simulation and the real world. We consider different design choices for core components of our approach. In the case of tokenizers, we consider dVAE [45], VQGAN [47] and continuous patch-normalized [25] tokens. We also consider different model architectures including LLaMA [55], GPT2 [41] and Mamba [23].", + "bbox": [ + 511, + 731, + 905, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We find that, for tokenization autoregressive models based on discrete and continuous patch-normalized [25]", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19108", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fce19da66bba7fb84e3a749d31807d434cfa1df660dfe70f57658336d89e94bc.jpg", + "image_caption": [ + "Figure 1. Overall Framework. Starting with images and video frames from a collection of datasets, we tokenize each frame/image into discrete visual tokens independently. We pre-train the transformer by predicting the next visual tokens, with a context length of 4K tokens of images or video frames. Once trained, we take the intermediate representations and evaluate them on various tasks." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 238, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a3ed2da9f518d054759fde1f086f7fc70da95e61b95280867b6b9a4c32f0cd8b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 90, + 751, + 257 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2a037c324d8b065ab735a91909140961c48e7d7a8eb5b88e7ebba5f81a14e2a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 759, + 90, + 895, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tokens perform similarly on ImageNet classification task. For efficient pre-training, starting with lower resolution and fine-tuning at higher resolution gives better performance and RoPE [54] helps with adopting to higher resolution. For measuring the representation quality in decoder-only models, due to skewed nature of the receptive field we use attention pooling over average pooling. We find that in decoder-only models, for all tasks and models sizes the middle layer gives the best performance.", + "bbox": [ + 88, + 340, + 480, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we study the scaling behaviors of autoregressive vision models, which scales with more compute but still at a slower rate compared to large language models. We will release code and models to facilitate future research.", + "bbox": [ + 89, + 479, + 482, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 556, + 227, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representation Learning for Vision: Over the years self-supervised pre-training has proven to be effective in many areas including language, vision, and robotics. [63] and SimCLR [10] showed that instance discrimination training can learn strong discriminative features. MoCo [24] and DINO [7] showed the effectiveness of strong visual representations on various tasks. Differently, BEiT [3] and MAE [25] used masked autoencoding for learning image representations. ST-MAE [19] and VideoMAE [59] extended this masked modeling approach to videos, by masking a large amount of tokens during pre-training and predict the masked tokens with a light-weight decoder.", + "bbox": [ + 89, + 582, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive Modeling of Vision: Generative autoregressive pre-training learns to directly model the data distribution. In language models, generative pre-training has become the standard for training large models. For autoregressive pre-training in vision, rCNN [46], PixelCNN [56] and PixelRNN [57] proposed generating pixels one by one using convolution and bidirectional LSTMs. With the introduction of the transformers [58], ImageTransformers [37] showed generating pixels with causal local attention per", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "forms better than previous CNN and RNN-based methods. While all of these methods focused on the generation quality of the pixels, iGPT [9] showed that generative pretraining is also a good way to learn strong visual representations for recognition tasks. [26] showed scaling behaviors of autoregressive image and video models. AIM [14] on the other hand uses patch embedding rather than any pretrained models for tokenization, however, it trains on Data Filtering Networks [17] with clip filtered data. Compared to these works, we do not use any supervision during our pre-training and utilizes image and videos jointly. VisionMamba [68] also showed how to utilize sequence models with bidirectional state-space modeling for supervised vision tasks. [62] showed autoregressive video generation for promotable video generations.", + "bbox": [ + 511, + 340, + 906, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation of Vision Representations: Most video pretraining models are evaluated on semantic tasks like ImageNet [12] and Kinetics [29]. Additionally to the standard evaluation, we evaluate our models on semi-supervised tracking task on DAVIS [39], action forecasting on Ego4D [21], object permanence on CATER [20] and on robot tasks in simulation [64] and in the real world [43].", + "bbox": [ + 511, + 566, + 908, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 511, + 679, + 624, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We treat images and videos as sequences of visual patch tokens and train a casual transformer model to predict the next patch tokens. This is akin to the next token prediction in large language models. From the vast collection of images and videos, every patch is tokenized into a discrete token, and the transformer is trained to predict the next token, using raster scan ordering. We pre-train our models on over one trillion tokens. Finally, we evaluate the learned representations of these models on various downstream tasks including image classification, action classification, action anticipation, video tracking, object permanence, and robotic manipulation tasks. We also study the scaling behaviors of our models for compute optimal training.", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19109", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/69d6a067bbfbe103aeea571f4c318348c4e740d8fecbd5f7c95ccafd1d2cc393.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParamsDimensionHeadsLayers
base120m7681212
large280m10241616
1b1.1b20481622
", + "bbox": [ + 93, + 88, + 480, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Model Architecture: We pre-train models at different scales, only on visual tokens from images and videos.", + "bbox": [ + 89, + 170, + 482, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Pre-training", + "text_level": 1, + "bbox": [ + 89, + 215, + 223, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a large collection of images and videos, we tokenize all of them into a 1D sequence using raster scan ordering. This produces a dataset of tokens, $\\{x_1^j,x_2^j,x_3^j,\\dots,x_n^j\\}$ where $j$ is the sample either from a video or an image and $n$ is the number of tokens in an image or a video. We model the density $p(x)$ as:", + "bbox": [ + 89, + 238, + 483, + 330 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np \\left(x ^ {j}\\right) = \\prod_ {i = 1} ^ {n} p \\left(x _ {i} ^ {j} \\mid x _ {i - 1} ^ {j}, x _ {i - 2} ^ {j}, \\dots , x _ {1} ^ {j}, \\theta\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 344, + 482, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\theta$ is the model parameters, which can be optimized by minimizing the negative log-likelihood loss:", + "bbox": [ + 89, + 397, + 483, + 428 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e - t r a i n}} = \\underset {x ^ {j} \\sim X} {\\mathbb {E}} - \\log p \\left(x ^ {j}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 440, + 482, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using this loss, we pre-train our models at different sizes on over 1T of tokens. The tokens come from images and videos. Figure 2 shows the training loss of 3 differently sized models with 120M, 280M and 1.1B parameters.", + "bbox": [ + 89, + 478, + 483, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture", + "text_level": 1, + "bbox": [ + 89, + 551, + 225, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model is a transformer [58] with causal attention. We apply recent advancements in language modeling such as pre-norm using RMSNorm [66], SwiGLU activation [52], and RoPE positional embeddings [54], following LLaMa [55].", + "bbox": [ + 89, + 574, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a model with $L$ layers, we define $H^l$ as the intermediate representations after layer $l$ , $0 \\leq l \\leq L$ . The intermediate representations after layer $l + 1$ , $H^{l + 1}$ , given as:", + "bbox": [ + 89, + 651, + 483, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {H} ^ {l + 1} = H ^ {l} + \\operatorname {M H S A} (\\operatorname {R M S - n o r m} (H ^ {l})) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 708, + 482, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {l + 1} = \\widehat {H} ^ {l + 1} + \\operatorname {M L P} (\\operatorname {R M S - n o r m} (\\widehat {H} ^ {l + 1})), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 729, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where MHSA is a multi-head self attention layer, MLP is a multi-layer perceptron with SwiGLU activations.", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train our models for the next token prediction task at different scales (base, large and 1b models). For more architecture details see Table 1. We train all these models with a batch size of 1M tokens. We use AdamW [33] with a maximum learning rate of $3e - 4$ , and $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.95$ . We decay the learning rate with a cosine schedule, after 2000 warm-up steps [55].", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9ef25640e4d5f02b99dcec6243b12b909acafcd81fdbef0ab3c4cb11d0f3375.jpg", + "image_caption": [ + "Figure 2. Training Loss Curves: We show the training loss curves for base, large, and 1b models trained with tokens from dVAE [45] with a vocabulary size of 8k and context length of 4k tokens (equivalent to 16 images or video frames)." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 895, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Dataset", + "text_level": 1, + "bbox": [ + 513, + 409, + 609, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train our model, we compile a large dataset from a number of different sources. Table 2 shows the total number of images and videos used for training data, the total number of tokens, as well as the number of hours of videos in each dataset. Together these datasets contain over 100,000 hours of video data and about 2.5 trillion visual tokens. During training, each mini-batch is sampled at different ratios of datasets. Each batch approximately contains $20\\%$ of ImageNet images, $10\\%$ of Ego4D videos, $10\\%$ of Kinetics videos, and $60\\%$ of HowTo100m videos. Our full training utilized about 1 trillion tokens.", + "bbox": [ + 511, + 431, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Tokenization", + "text_level": 1, + "bbox": [ + 511, + 606, + 650, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use dVAE tokenizer with a vocabulary of 8k tokens, from Dall-E [45] as our tokenizer. Using an image-based tokenizer allows training on both images and videos and testing on respective downstream tasks. While VQGAN [15] tokenizers provide sharper images, these models were trained with perceptual loss [28, 30], thus indirectly ingesting ImageNet label information via VGG-net [53]. We also explored SD-VAE based random tokenizer, please see Appendix A.4.", + "bbox": [ + 511, + 628, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "All raw pixel frames or images are tokenized into 256 discrete tokens. We take a video and resize it such that its shortest size is $R$ pixels, and then take a random crop of $R \\times R \\times T$ , and sample every 4 frames where $T$ is the number of frames. We use dVAE [45] with the vocabulary of 8k entries to tokenize every frame independently. For dVAE we set $R = 128$ , to get $16 \\times 16$ discrete tokens. Once every frame is mapped into a set of discrete tokens we have $T \\times 256$ tokens per each video. We pre-train all the mod-", + "bbox": [ + 511, + 765, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19110", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/0524008b48b6656c9bbc8c2163353a8c4e440b3210b74f7a1520f8fe75dff1a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetsInstancesTokensHours
ImageNet13.9M3.6B-
Kinetics-6000.53M41.3B1496
Ego4D52.1K103B3750
HowTo100m1.172M2560B92627
", + "bbox": [ + 122, + 88, + 450, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Pre-training Dataset: We use both image datasets (Imagenet [48]) and video datasets (Kinetics600 [8], Ego4D [21], HowTo100m [34]) with different mixing ratios during the pretraining of our models. The whole training data contains about 100,000 hours of videos.", + "bbox": [ + 89, + 205, + 480, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "els with $T = 16$ , thus all the models are per-trained for a context length of 4096 tokens.", + "bbox": [ + 89, + 286, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When training with images and videos, 16 video frames are sampled to create 4k tokens. For images, we randomly sample 16 images and create a sequence of 16 image frames to generate 4k tokens. Finally, we add start and end tokens for each sequence, for videos we use [1] as the start token, and for images we use [3] as the start token, and all sequences have an end token of [2].", + "bbox": [ + 89, + 316, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Downstream Transfer", + "text_level": 1, + "bbox": [ + 89, + 431, + 294, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The idea of large pre-trained models is that they were trained at a large compute scale, and then these models can be easily used for various downstream tasks without requiring task-specific design or lots of computing for transfer. The learned representations are general enough to transfer to various tasks. We evaluate our models on the intermediate features with linear and attention probing [31].", + "bbox": [ + 89, + 454, + 482, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For linear probing the model, we apply global average pooling [32] over the tokens from different layers to get the intermediate representation. We train a linear layer on top of this representation on the downstream task. MAE [25] or DINO [7] have a uniform structure when it comes to which token attends to which tokens, however in autoregressive sequence modeling later tokens attend to more tokens than the tokens at the beginning. Due to this skewed nature, equally weighting all the tokens affects the downstream performance. Attention pooling is an alternative to average pooling that allows to dynamically weight the tokens, ideally giving more weight to tokens that see more tokens. This requires learning $W_{k}$ and $W_{v}$ matrices and a query token $q$ . The query token cross-attends to the intermediate tokens and combines them into a single vector. While this function is not linear anymore, it has been shown to learn better representations in recent works [14].", + "bbox": [ + 89, + 561, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 220, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evaluate our pre-trained models on various downstream tasks such as ImageNet classification, Kinetics action recognition, Ego4D action anticipation, Semi-Supervised", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f37372431bdc0dcb47c1654075b4ebd2e3d26c6be27b11880d40aee3ff30867.jpg", + "image_caption": [ + "Figure 3. 1-gram Distribution of Various Tokens: This Figure shows the distribution of 1-gram tokens of various tokenizers (dVAE [45], VQGAN-1k, VQGAN-16k [15]) on Imagenet validation set. Note that, dVAE has almost full convergence of the tokens while VQGAN has less than $50\\%$ coverage of the tokens." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 645, + 156 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28d79a5c981899eb1db81caf03e3a87c9a7458d29a95aa10b1a1daab39411605.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 89, + 772, + 156 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48ac95f631af77be480ce08c80a91971f8fa86feecdb0e1c95516a85d6412fa9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 89, + 901, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tracking, and Robotic manipulation tasks. First, we discuss various design choices for pre-training and evaluation strategies for our method. All the models for studying the design choices are large models trained for 400 epochs on the ImageNet-1k dataset.", + "bbox": [ + 511, + 265, + 903, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Design Choices", + "text_level": 1, + "bbox": [ + 511, + 349, + 666, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenizer: The are various options available for tokenizing an image or a video. We could use discrete tokenizers such as dVAE, and VQGAN, or simple patch-based continuous tokenization. To study the behavior of various tokenizers we pre-train a Toto-large model on ImageNet for 400 epochs. Using linear probing at an optimal intermediate layer, we evaluate the accuracy of the models on ImageNet classification task.", + "bbox": [ + 511, + 372, + 905, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 3 shows linear probing accuracy when trained with various tokenizers. VQGAN [15] and dVAE [45] perform similarly with the same resolutions. However, VQGAN is contaminated with ImageNet label information via the perceptual loss. In addition, as shown in Figure 3, dVAE tokens have full coverage compared to VQGAN tokens on their 1-gram distributions. Please see the supplementary materials for details. Regressing normalized-patch targets from patch embeddings performs slightly worse than classifying discrete tokens as targets. Additionally, discrete tokens as targets and patch embeddings as inputs perform poorly compared to other methods at the given input-output resolutions. Overall, Table 3 shows that various ways of tokenization have little effect on ImageNet linear probing accuracy.", + "bbox": [ + 511, + 492, + 905, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to Probe: As discussed in Section 3.5 we probe the pre-trained models at the same layer with attention pooling and average pooling, followed by linear layer. Table 13 shows attention pooling performs $7.9\\%$ higher than average pooling on the ImageNet classification task. For attention pooling, we keep the embedding dimension the same as the intermediate feature dimensions.", + "bbox": [ + 511, + 704, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Resolution: When training with dVAE tokens, a 256x256 image results in 1024 tokens, this is four times more number of tokens compared to patch embeddings or VQGAN tokens. If we reduce the number of tokens to 256, then the effective image resolution becomes $128 \\times 128$ . Table 4 shows a clear drop in performance when pre-training the", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/031fa5e9cd2f6f5e76343c3bab9e0233632f08248e43f30621b67ad3eaa52de3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input-TargetTokensVocabularyTop1
VQGAN-VQGAN16x1616k61.3
VQGAN-VQGAN16x161k61.1
dVAE-dVAE32x328k61.2
dVAE-dVAE16x168k53.2
patch-patch16x16-60.6
patch-dVAE16x168k58.5
", + "bbox": [ + 101, + 88, + 470, + 227 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/991b76c54fc559bad53020cf3b58aa4d6964c3e435f742775074547eea458cfb.jpg", + "table_caption": [ + "Table 3. ImageNet Linear Probing Accuracy with Various Tokenizers: We compare discrete (dVAE, VQGAN) and patch embedding as input and target for pre-training our models. ImageNet top-1 accuracies are computed by linear probing at the 9th layer of the large model." + ], + "table_footnote": [], + "table_body": "
MethodCompute (MACs)Top1
dVAE/161.42 × 101753.2
dVAE/325.68 × 101761.2
dVAE/16→322.13 × 101763.2
dVAE/16→32†2.13 × 101764.4
", + "bbox": [ + 114, + 321, + 457, + 426 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c1f0eb78104f45d6e2ea9d6b209e9c24222d999d0c115ecda1118891a7d04ae6.jpg", + "table_caption": [ + "Table 4. Token Resolution: While the performance is lower for a low-resolution model, when finetuned for next-patch prediction at a higher resolution, its performance surpasses the full-resolution pre-trained model. $\\dagger$ Base values of the RoPE is 50,000." + ], + "table_footnote": [], + "table_body": "
MethodTokensPoolingTop1
dVAE16x16Average53.2
dVAE16x16Attention61.1
", + "bbox": [ + 169, + 513, + 401, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model at $128 \\times 128$ resolution. However, due to the use of relative positional embeddings (RoPE [54]), we can easily finetune the $128 \\times 128$ (or $16 \\times 16$ token equivalent) model for higher resolution. Surprisingly, this does better than pre-training at $256 \\times 256$ resolution and requires only one epoch of finetuning. Not only does this improve the performance, but the pre-training also becomes cheaper compared to full-resolution pre-training. Additionally, Fine-tuning with higher base values of the RoPE embeddings (50,000) leads to better accuracy.", + "bbox": [ + 89, + 702, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture: We train various language models from GPT2 [41] with absolute sine-cosine positional embeddings, and non-transformer based model Mamba [23] only", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/95f33d64982b167c5f9f9691bee470652055d0e6f9776b7e7db997b515ab67a5.jpg", + "table_caption": [ + "Table 5. Attention vs Average Pooling: When probed at the same layers, attention pooling performs much better than average pooling of intermediate tokens." + ], + "table_footnote": [], + "table_body": "
ModelParamsTop1
GPT2 [41]280 m48.5
Mamba [23]290 m40.7
LLaMA [55]280 m53.2
", + "bbox": [ + 601, + 89, + 821, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 6. Architecture: We compare sequence modeling architectures LLaMA [55], GPT2 [41], and non-transformer models, Mamba [23] on ImageNet linear probing task.", + "bbox": [ + 511, + 166, + 903, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "using dVAE tokens. We mimicked the GPT2 architecture and do architecture comparisons. We compare these models with LLaMA [55]. We evaluate linear probing performance at each layer of these models and report the best performance in Table 6.", + "bbox": [ + 511, + 233, + 903, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Probing Layer: When probing the pre-trained models, especially the decoder-only model best performance is observed at the middle layers. This behavior is first observed in iGPT [9]. Figure 4 shows the peak performance on recognition occurs at about $50\\%$ of the depth of the model. This behavior holds across all model sizes. While in MAE [25] and BEiT [3] encoder-decoder models, due to the uneven nature of the encoder and decoder, the best features are observed at the top of the encoder layers. However, on decoder-only models with uniformly distributed layers, the last layers perform worse on recognition tasks, mainly because these layers are trained to reconstruct the input. More probing results with various tokenizers, resolutions, and probing methods are shown in the supplementary material.", + "bbox": [ + 511, + 309, + 906, + 521 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eee90fbf8aa5fa427afc518ab23b49f934ba29cb5dc72f3d9cf9c2bba989ed5e.jpg", + "image_caption": [ + "Figure 4. Probing at Different Layers: We show the attention-probing performance at each layer of our three models. Peak performance is observed at around $50\\%$ depth of the models." + ], + "image_footnote": [], + "bbox": [ + 521, + 536, + 898, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Image Recognition", + "text_level": 1, + "bbox": [ + 511, + 818, + 694, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To measure the representation quality of our pre-trained models, we evaluate our models on ImageNet-1k [12] classification. We apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19112", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/25c513f6692fec35f3d9e67c7d70c134aa0ab27d1b22376e4b9d2b225181bc85.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArch#θTop1
Discriminative Approaches
SimCLR [10]†RN50x29474.2
BYOL [22]†RN50x29477.4
SwAV [6]†RN50x29473.5
DINO [7]ViT-B/88680.1
DINOv2 [36]ViT-g/14101186.4
Generative Approaches
BEiT-L [3]ViT-L/1430762.2
AIM [14]ViT-1B/14120080.6
MAE [25]ViT-H/1463280.9
iGPT-L [9]†GPT-2138665.2
iGPT-XL [9]†GPT-2680172.0
Toto-baseLLaMA12064.7
Toto-largeLLaMA28071.1
Toto-1bLLaMA110075.3
", + "bbox": [ + 135, + 88, + 439, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the highest classification accuracy. We fine-tune the pretrained models further by applying self-supervised next token prediction loss in Eq 2, together with cross-entropy loss applied for probing layers (with stop-gradients). We train the probing layers for 90 epochs, with a learning rate of $6e^{-5}$ . We also use layer decay of 0.9 to reduce the learning rate at the early layers of the model. During this stage, all the models are fine tuned with $32 \\times 32$ token resolution, on the self-supervised loss, and increase the base value of the RoPE [54] embeddings from 10,000 to 50,000 support larger resolution.", + "bbox": [ + 88, + 506, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 7 shows the ImageNet top-1 accuracy of our base, large and 1b models. First, there is a clear difference in terms of classification performance when it comes to discriminative models versus generative models. Instance discriminative models such as SimCLR [10], and DINO [7] are trained to separate samples from each other and they are designed to perform well on discriminative tasks. On the other hand, generative models are just trying to model the data distribution. While achieving comparable performance to other generative models on image recognition, among autoregressive generative models, our model achieved the highest top-1 accuracy. The scaling of data, and the use of tokens instead of pixels, allows our one billion parameter model to achieve similar performance compared to iGPT [9] 7 billion models.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4dcf443ed9e5c2c5c695f2cc16ebd56844a3fa13a2bbd02bb72096f933445906.jpg", + "table_caption": [ + "Table 7. ImageNet Results: We compare discriminative and generative models on ImageNet [12] recognition task. While achieving comparable performance among generative models, our models model achieves the highest accuracy on autoregressive modeling. $\\dagger$ models are evaluated with linear probing. $\\# \\theta$ in millions." + ], + "table_footnote": [], + "table_body": "
MethodArchTop1
Discriminative Approaches
I-JEPA [1]ViT-H/1674.5
OpenCLIP [11]ViT-G/1483.3
DINOv2 [36]ViT-g/1484.4
InternVideo [61]-73.7
Generative Approaches
Hiera [49]Hiera-H/1477.0
MVD [60]ViT-H/1479.4
VideoMAE [59]ViT-L/1479.8
Toto-baseLLaMA59.3
Toto-largeLLaMA65.3
Toto-1bLLaMA74.4
", + "bbox": [ + 570, + 88, + 851, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8. K400 Results: We compare discriminative and generative models on Kinetics-400 [29] action recognition task. While achieving comparable performance among generative models, our models are the first to show the competitive performance on K400 with autoregressive pre-training, and shows scaling with large model sizes.", + "bbox": [ + 511, + 345, + 906, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Action Recognition", + "text_level": 1, + "bbox": [ + 511, + 445, + 697, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use Kinetics-400 (K400) [29] for evaluating our models on action recognition tasks. Similar to ImageNet evaluation, we apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with the highest action classification accuracy. We also fine-tune the pre-trained models on a self-supervised next-patch prediction task while training the probing layers with a classification loss. All our video models are trained with 16 frames, thus with a context length of 4096 tokens per video. When evaluating videos, we follow the protocol in SlowFast [18]. Unlike ImageNet where we evaluate the models at $256 \\times 256$ resolution, on videos we only evaluate our models at $128 \\times 128$ resolution, to keep the number of tokens in a similar budget.", + "bbox": [ + 509, + 468, + 906, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 8 shows the Kinetics-400 top-1 accuracy of our base, large and 1b models. Similar to ImageNet results in Table 7, we see that discriminately trained models perform better than generative models. Our models achieve comparable performance among generative models, and first to show competitive performance on action recognition with autoregressive generative modeling. All the models are trained and evaluated with 16 frames with a stride of 4 frames.", + "bbox": [ + 509, + 681, + 908, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Action Forecasting", + "text_level": 1, + "bbox": [ + 511, + 830, + 694, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the Kinetics dataset captures internet-style exocentric videos, Ego4D [21] videos capture egocentric videos of daily life. A general vision model should be able to rea", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19113", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47b1cb6a3c29fa56c8d50fbeaf1fda670d84f5e570cc90e7b16b505f64ae7f25.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNounN+VN+TTCOverall
FRCNN+Rnd [21]17.551.563.210.34
FRCNN+SF [21]17.555.195.372.07
Hiera-large [49]14.056.034.532.12
StillFast [44]16.207.474.942.48
VideoMAE-large [59]15.166.725.262.55
MAE-ST-large [19]13.716.634.942.60
Toto-large15.206.755.412.70
", + "bbox": [ + 91, + 88, + 483, + 233 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "son about both exo and ego-centric videos. Task-wise, Kinetics requires the model to reason about the action using full context (e.g., the model has seen the action), while the Ego4D short-term action anticipation v1 task requires models to predict future actions from past context. We use our models as the backbone for the pyramid network used in StillFast [44] to extract tokens at 5 layers and fuse them with the pyramid network. We fully fine-tuned our model with self-supervised next-patch loss along with task-related losses, and we observed that having the self-supervised loss improves the overall performance. Table 9 shows the performance of our large model on the Ego4D short-term action anticipation task. This task requires predicting the object to be interacted with (noun) and the type of interaction (verb) as well as time to contact (ttc) from the last seen frame to an estimated time between object-hand contact. As shown in Table 9, these tasks are difficult with maximum overall mean-average precision of 2.70.", + "bbox": [ + 88, + 325, + 483, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Video Tracking", + "text_level": 1, + "bbox": [ + 89, + 606, + 246, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our pre-trained models on the label propagation task on the DAVIS dataset [39], following the evaluation protocol in [27]. In contrast to the previous tasks, such as classification, this evaluation does not require fine-tuning or probing. We first use the features from the last $n$ frames to find the nearest neighbor patch in the current frame. We then propagate the masks from the previous frames to the current frame. We compare with Dino [7] and MAE [25] in Table 10 and show qualitative results in Figure 7.", + "bbox": [ + 89, + 628, + 483, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Robotics", + "text_level": 1, + "bbox": [ + 89, + 773, + 194, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we study the effectiveness of our pre-trained representations for robotic manipulation. We consider tasks in simulation and in the real world. Due to the runtime constraints on real hardware, we use the Toto-base models. Despite being a small model, Toto-base can achieve better performance in simulation and on-par performance to state-of-the-art robot models in real-world experiments.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b70bf8ea58a31f736548e2f0dc3cf682e1a19040de0750e9f316be4b47e423c6.jpg", + "table_caption": [ + "Table 9. Ego4D Results: Our model achieves comparable mean-average precision compared to previous work. We compare our method with, FRCNN+Rnd [21], FRCNN+SF [21], Hiera [49], StillFast [44], VideoMAE [59], and MAE-ST [19]." + ], + "table_footnote": [], + "table_body": "
Method (Res/Patch)J&FJF
DINO-base (224/8)54.352.556.1
DINO-base (224/16)33.136.230.1
MAE-base (224/16)31.534.128.9
Toto-base (256/8)42.041.243.1
Toto-large (256/8)44.844.445.1
Toto-1b (256/8)46.145.846.4
Toto-large (512/8)62.459.265.6
", + "bbox": [ + 514, + 88, + 905, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 10. DAVIS Tracking: We report J, F, and J&F scores at the peak layers of each model. We achieves comparable performance as DINO and at large resolution (512), it outperforms all methods. Our model achieve best performance at the middle layer, see 9.", + "bbox": [ + 511, + 250, + 906, + 306 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/119e676c787020145b1a2588468ddb1c7e50f1bd2fb29696c4c21df68f147483.jpg", + "image_caption": [ + "(a) Franka Pick" + ], + "image_footnote": [], + "bbox": [ + 516, + 324, + 705, + 431 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/65de3d672e8e31f7b287f4b6a0d2e5a688742ea3503ef1823dedcd3b0d0d818e.jpg", + "image_caption": [ + "(b) Kuka Pick", + "Figure 5. Robot Manipulation with Reinforcement Learning: We compare MAE-base [42] with Toto-base pre-trained models in simulation following [64]. We find that the Toto outperforms the baseline in sample complexity, across two robots and two tasks." + ], + "image_footnote": [], + "bbox": [ + 709, + 323, + 897, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Simulation Experiments: Following the protocols in MVP [64], we use our visual pre-trained models to embed pixel observations. The model is frozen and we only take tokens at an intermediate layer, apply average pooling, and learn the linear layer on top to embed pixel observations. These observations are used to train DAgger policies for 2 tasks: Franka-pick 5a, Kuka-pick 5b. Figure 5 shows the mean success rate over training steps. Compared to the MVP baseline, our model was able to learn these tasks faster with better sample efficiency across robots and tasks. For fair comparisons, we use the best MAE model from MVP [42] which is trained on ImageNet [12], Ego4D [21] and 100DOH [50] datasets.", + "bbox": [ + 511, + 536, + 906, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world Experiments: Next, we evaluate our pre-trained models on hardware. We follow the setup from [42]. We extract vision features using a pre-trained vision encoder and train a controller on top of frozen representations using behavior cloning. Specifically, we consider a cube picking tasks using a 7 DoF Franka robot, shown in Figure 8. We use the demonstrations provided by [43]. In Table 11 we compare our model to a vision encoder from [42]. We report the success rate over 16 trials with variations in object position and orientation. Our model performs favorably to a vision encoder pre-trained for robotics.", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19114", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2d5c4c1901b7a1b7d077858526cf583d9764dcfa7d02ae6ec642a3a136bedf8b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model# TrajSuccess
MVP24075%
Toto-base24063%
", + "bbox": [ + 165, + 88, + 410, + 150 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2cf6667ee06f7f382c9c3d2fd3a6a1625a0d2d5d3100a16f9bed1241c7b8e441.jpg", + "table_caption": [ + "Table 11. Robotics, Real-world Experiments: We compare MVP [42] and Toto on a Franka cube-picking task in the real world. Features from both models are pre-trained, frozen, and passed into a learning module trained with behavior cloning using the same demonstrations. We see that our approach performs comparably to the state-of-the-art vision backbone for robotics, despite not being designed with the robotic application in mind." + ], + "table_footnote": [], + "table_body": "
MethodModel16 frame32 frame
V3DResNet55.269.7
TFC V3DResNet54.670.2
Toto-largeLLaMa62.872.9
", + "bbox": [ + 119, + 263, + 455, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 12. Object Permanence: In the CATER [20] object localization task, the object is hidden under or obstructed by other objects and the goal is to predict its coarse location. Our model outperforms previous methods at 16 and 32 temporal resolutions.", + "bbox": [ + 89, + 364, + 482, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Object Permanence", + "text_level": 1, + "bbox": [ + 89, + 439, + 277, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To quantitatively measure the performance of how well the model understands object permanence, we evaluate our models on the CATER localization task [20]. Here, a ball is moving in the scene, and the task is to find its location in the 6 by 6 grid. We fine-tune our Toto-large model on this task at temporal resolutions of 16 and 32 frames. In both cases, our pre-trained models outperform the models trained specifically for this task, such as V3D [67], TFC-V3D [67]. Table 12 shows the performance on the CATER snitch localization task, and Toto-large achieve $62.8\\%$ and $70.9\\%$ performance with 16 and 32 frames respectively.", + "bbox": [ + 89, + 460, + 483, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.8. Compute Optimal Scaling", + "text_level": 1, + "bbox": [ + 89, + 637, + 328, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study the scaling properties of Toto using $\\mu$ -Parameterization [65]. First we train various models, a1-a6, with linearly increasing hidden size and number of layers (Table 17). All models use the VQGAN tokenizer [15]. We then optimize the learning rate for these models, with $\\mu$ -Parameterization [65]. Figure 11 shows the optimal learning rate of $2^{-7}$ for all of the model widths. Once we find the optimal learning rate, we train a1-a6 models on our data mixture (Table 2). Figure 6 shows the loss vs. training compute of Toto models. We see a clear power law relationship between the compute and validation loss. Based on these experiments Toto shows a power law of $L(C) = 7.32 \\cdot C^{-0.0378}$ . For comparison, the GPT-3 power law relationship [4] is $L(C) = 2.57 \\cdot C^{-0.048}$ . While these are not comparable directly, the scaling coefficients indicate how much change in the loss to expect for additional com", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1f099cfe78788d43d19228d218d4bae1dcc9c06f02bc263f2ca86c9ab55ff8a.jpg", + "image_caption": [ + "Figure 6. Scaling Toto: We train multiple variants of Toto, with increasing hidden size and depth, with optimal learning rates. We plot the validation loss vs the compute spent on training in MACs. This shows a clear scaling behavior with optimal compute." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 890, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pute. This suggests that autoregressive video models, such as Toto, scale but at a slower rate than language models. In language processing, tokenization is typically lossless and highly compressed, refined over thousands of years. In contrast, video inherently contains significant redundancy across subsequent frames—if a video lacks this redundancy, it wouldn’t truly qualify as a video. This redundancy, however, is essential for effective self-supervised training. With better tokenizers that yield more compressed and semantically rich representations, we can build more scalable video encoders and use scaling laws as benchmarks. Moreover, scaling laws could help us curate more scalable data and inform the design of model architectures, rather than relying solely on single benchmark metrics. See the limitations section in the Appendix for further discussion.", + "bbox": [ + 511, + 435, + 906, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We empirically study autoregressive pre-training from videos. We curate a large video dataset and conduct a largescale evaluation across a range of diverse tasks, including image recognition, video classification, video forecasting, object tracking, object permanence, and robotic manipulation. We perform extensive ablation studies to understand different design choices and compare autoregressive video pre-training to strong baselines across different tasks. We find that, despite minimal inductive biases, our approach achieves competitive performance across all tasks. Finally, we study the scaling behavior of autoregressive video models, and show that they scale with compute, but at a slower rate than autoregressive language models.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Andrea Madotto, Po-Yao (Bernie) Huang, and Shiry Ginosar for helpful discussions. We're grateful to Ronghang Hu and Xinlei Chen for their help with TPU setup and code bases. We also thank Baifeng Shi for helping us with robots evaluations. We thank Valentin Gabeur and Neerja Thakkar for their valuable feedback on the paper. This work was supported by the FAIR-BAIR program as well as ONR MURI (N00014-21-1-2801).", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 250, + 187, + 265 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 6", + "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 1954. 1", + "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 5, 6, 12", + "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 8", + "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1", + "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6", + "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 6, 7, 12", + "[8] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 4", + "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2, 5, 6, 12, 14", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6", + "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6" + ], + "bbox": [ + 93, + 273, + 483, + 896 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 5, 6, 7", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 1", + "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 2, 4, 6", + "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868–12878, 2020. 3, 4, 8, 14", + "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 12", + "[17] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 2", + "[18] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6", + "[19] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2, 7", + "[20] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 2, 8", + "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4, 6, 7", + "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6", + "[23] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 5", + "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + ], + "bbox": [ + 516, + 92, + 905, + 895 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19116", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 5, 6, 7, 12", + "[26] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2", + "[27] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7, 13", + "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 3", + "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 12", + "[30] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In Proceedings of The 33rd International Conference on Machine Learning, pages 1558-1566, New York, New York, USA, 2016. PMLR. 3", + "[31] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 3744-3753. PMLR, 2019. 4", + "[32] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013. 4", + "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3", + "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 4", + "[35] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 12", + "[36] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6", + "[37] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International conference on machine learning, pages 4055-4064. PMLR, 2018. 2" + ], + "bbox": [ + 91, + 90, + 485, + 897 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Jean Piaget. The construction of reality in the child. Routledge, 1954. 1", + "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7", + "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 1", + "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2019. 1, 5", + "[42] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, 2022. 7, 8", + "[43] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In Conference on Robot Learning, 2023. 2, 7", + "[44] Francesco Ragusa, Giovanni Maria Farinella, and Antonino Furnari. Stillfast: An end-to-end approach for short-term object interaction anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635-3644, 2023. 7", + "[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3, 4, 14", + "[46] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 2", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 4", + "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 6, 7, 14", + "[50] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 7", + "[51] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 1951. 1" + ], + "bbox": [ + 516, + 90, + 906, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19117", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 3", + "[53] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3", + "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 3, 5, 6", + "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 14", + "[56] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2", + "[57] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2, 12", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3", + "[59] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 2, 6, 7, 14", + "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang." + ], + "bbox": [ + 89, + 92, + 483, + 570 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, 2023. 6", + "[61] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6", + "[62] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2", + "[63] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2", + "[64] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2, 7", + "[65] Greg Yang, Edward J. Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer, 2022. 8, 14", + "[66] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[67] Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning. arXiv preprint arXiv:2203.05928, 2022. 8", + "[68] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19118", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_model.json b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f9732678a8534fc10b45b31ac75487bc088d0d7a --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.179, + 0.131, + 0.82, + 0.153 + ], + "angle": 0, + "content": "An Empirical Study of Autoregressive Pre-training from Videos" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.18, + 0.847, + 0.235 + ], + "angle": 0, + "content": "Jathushan Rajasegaran1,2 Ilija Radosavovic2 Rahul Ravishankar2 Yossi Gandelsman2 \nChristoph Feichtenhofer1 Jitendra Malik1,2 \n1Meta FAIR 2University of California, Berkeley" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.486, + 0.558 + ], + "angle": 0, + "content": "We empirically study autoregressive pre-training from unlabeled videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pretrained on a diverse dataset of unlabeled videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotic manipulation. Our results demonstrate that, despite minimal inductive biases, autoregressive video pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. https://brjathu.github.io/toto" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.222, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.705 + ], + "angle": 0, + "content": "In a paper published in 1951, Shannon, having just published the foundational papers of information theory, proposed a \"guessing game\" of next word prediction to estimate the entropy of English [51]. Nearly 70 years later, training a high-capacity transformer neural network [58] on this task, provided the generative pre-training backbone for Large Language Models [5, 13, 40, 41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Less well known is the fact that in 1954, Fred Attneave [2] proposed an analog of Shannon's task for images. To quote \"We may divide the picture into arbitrarily small elements which we \"transmit\" to a subject (S) in a cumulative sequence, having them guess at the color of each successive element until they are correct. This method of analysis resembles the scanning process used in television and facsimile systems and accomplishes the like purpose of transforming two spatial dimensions into a single sequence in time\". Of course this leaves open the question of what is the \"element\". Is it a pixel? A patch? A semantic entity?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While Attneave was concerned with images, in the context of 2024, we have to note the \"Big Visual Data\" is in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.905, + 0.361 + ], + "angle": 0, + "content": "video, not images. According to the web, YouTube alone has 156 million hours of video, and 30 thousand hours of video are being uploaded every hour. While there are concerns that most of the text available on the Internet has already been used by the leading language models, in video we are barely started on the journey of big data exploitation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Learning from this data brings a number of challenges. The vast majority of the video data is unlabeled, or at least not consistently labeled. Therefore, an effective learning approach should be able to learn from pixels alone. Indeed, human children form rich representations and models of the world purely from observing raw video [38], which further motivates the setting of learning from unlabeled videos. What is the right video architecture and the training objective? And what is the right visual \"element\" to operate over?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.73 + ], + "angle": 0, + "content": "In this paper, we empirically study autoregressive pretraining from unlabeled videos. To perform our empirical study, we construct a family of autoregressive video models which we call Toto. We treat videos as sequences of visual tokens and train a transformer model on the next token prediction task. We use a causal transformer model with improvements from the LLaMa [55] architecture. We use dVAE [45] to tokenize frames into discrete tokens. Treating videos as sequences of tokens enables us to jointly train on both videos and images using a unified format. We construct a diverse dataset of videos and images comprising over 1 trillion visual tokens. Our models are first pretrained on this dataset and then evaluated on various downstream tasks. We extract visual representations using attention pooling from relevant layers of the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.868 + ], + "angle": 0, + "content": "We evaluate our models on a range of downstream tasks from image and video recognition, video forecasting, semi-supervised tracking, object permanence and robotic manipulation tasks in both simulation and the real world. We consider different design choices for core components of our approach. In the case of tokenizers, we consider dVAE [45], VQGAN [47] and continuous patch-normalized [25] tokens. We also consider different model architectures including LLaMA [55], GPT2 [41] and Mamba [23]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We find that, for tokenization autoregressive models based on discrete and continuous patch-normalized [25]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19108" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.092, + 0.239, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.091, + 0.753, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.761, + 0.092, + 0.896, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.273, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 1. Overall Framework. Starting with images and video frames from a collection of datasets, we tokenize each frame/image into discrete visual tokens independently. We pre-train the transformer by predicting the next visual tokens, with a context length of 4K tokens of images or video frames. Once trained, we take the intermediate representations and evaluate them on various tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.482, + 0.478 + ], + "angle": 0, + "content": "tokens perform similarly on ImageNet classification task. For efficient pre-training, starting with lower resolution and fine-tuning at higher resolution gives better performance and RoPE [54] helps with adopting to higher resolution. For measuring the representation quality in decoder-only models, due to skewed nature of the receptive field we use attention pooling over average pooling. We find that in decoder-only models, for all tasks and models sizes the middle layer gives the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.54 + ], + "angle": 0, + "content": "Finally, we study the scaling behaviors of autoregressive vision models, which scales with more compute but still at a slower rate compared to large language models. We will release code and models to facilitate future research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.228, + 0.572 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Representation Learning for Vision: Over the years self-supervised pre-training has proven to be effective in many areas including language, vision, and robotics. [63] and SimCLR [10] showed that instance discrimination training can learn strong discriminative features. MoCo [24] and DINO [7] showed the effectiveness of strong visual representations on various tasks. Differently, BEiT [3] and MAE [25] used masked autoencoding for learning image representations. ST-MAE [19] and VideoMAE [59] extended this masked modeling approach to videos, by masking a large amount of tokens during pre-training and predict the masked tokens with a light-weight decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Autoregressive Modeling of Vision: Generative autoregressive pre-training learns to directly model the data distribution. In language models, generative pre-training has become the standard for training large models. For autoregressive pre-training in vision, rCNN [46], PixelCNN [56] and PixelRNN [57] proposed generating pixels one by one using convolution and bidirectional LSTMs. With the introduction of the transformers [58], ImageTransformers [37] showed generating pixels with causal local attention per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.568 + ], + "angle": 0, + "content": "forms better than previous CNN and RNN-based methods. While all of these methods focused on the generation quality of the pixels, iGPT [9] showed that generative pretraining is also a good way to learn strong visual representations for recognition tasks. [26] showed scaling behaviors of autoregressive image and video models. AIM [14] on the other hand uses patch embedding rather than any pretrained models for tokenization, however, it trains on Data Filtering Networks [17] with clip filtered data. Compared to these works, we do not use any supervision during our pre-training and utilizes image and videos jointly. VisionMamba [68] also showed how to utilize sequence models with bidirectional state-space modeling for supervised vision tasks. [62] showed autoregressive video generation for promotable video generations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.676 + ], + "angle": 0, + "content": "Evaluation of Vision Representations: Most video pretraining models are evaluated on semantic tasks like ImageNet [12] and Kinetics [29]. Additionally to the standard evaluation, we evaluate our models on semi-supervised tracking task on DAVIS [39], action forecasting on Ego4D [21], object permanence on CATER [20] and on robot tasks in simulation [64] and in the real world [43]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.625, + 0.697 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We treat images and videos as sequences of visual patch tokens and train a casual transformer model to predict the next patch tokens. This is akin to the next token prediction in large language models. From the vast collection of images and videos, every patch is tokenized into a discrete token, and the transformer is trained to predict the next token, using raster scan ordering. We pre-train our models on over one trillion tokens. Finally, we evaluate the learned representations of these models on various downstream tasks including image classification, action classification, action anticipation, video tracking, object permanence, and robotic manipulation tasks. We also study the scaling behaviors of our models for compute optimal training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19109" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.162 + ], + "angle": 0, + "content": "
ModelParamsDimensionHeadsLayers
base120m7681212
large280m10241616
1b1.1b20481622
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.171, + 0.483, + 0.199 + ], + "angle": 0, + "content": "Table 1. Model Architecture: We pre-train models at different scales, only on visual tokens from images and videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.217, + 0.224, + 0.233 + ], + "angle": 0, + "content": "3.1. Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.331 + ], + "angle": 0, + "content": "Given a large collection of images and videos, we tokenize all of them into a 1D sequence using raster scan ordering. This produces a dataset of tokens, \\(\\{x_1^j,x_2^j,x_3^j,\\dots,x_n^j\\}\\) where \\(j\\) is the sample either from a video or an image and \\(n\\) is the number of tokens in an image or a video. We model the density \\(p(x)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.345, + 0.483, + 0.383 + ], + "angle": 0, + "content": "\\[\np \\left(x ^ {j}\\right) = \\prod_ {i = 1} ^ {n} p \\left(x _ {i} ^ {j} \\mid x _ {i - 1} ^ {j}, x _ {i - 2} ^ {j}, \\dots , x _ {1} ^ {j}, \\theta\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Here, \\(\\theta\\) is the model parameters, which can be optimized by minimizing the negative log-likelihood loss:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.441, + 0.483, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e - t r a i n}} = \\underset {x ^ {j} \\sim X} {\\mathbb {E}} - \\log p \\left(x ^ {j}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Using this loss, we pre-train our models at different sizes on over 1T of tokens. The tokens come from images and videos. Figure 2 shows the training loss of 3 differently sized models with 120M, 280M and 1.1B parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.226, + 0.567 + ], + "angle": 0, + "content": "3.2. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.65 + ], + "angle": 0, + "content": "Our model is a transformer [58] with causal attention. We apply recent advancements in language modeling such as pre-norm using RMSNorm [66], SwiGLU activation [52], and RoPE positional embeddings [54], following LLaMa [55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.698 + ], + "angle": 0, + "content": "For a model with \\(L\\) layers, we define \\(H^l\\) as the intermediate representations after layer \\(l\\), \\(0 \\leq l \\leq L\\). The intermediate representations after layer \\(l + 1\\), \\(H^{l + 1}\\), given as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.709, + 0.483, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\widehat {H} ^ {l + 1} = H ^ {l} + \\operatorname {M H S A} (\\operatorname {R M S - n o r m} (H ^ {l})) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.731, + 0.483, + 0.751 + ], + "angle": 0, + "content": "\\[\nH ^ {l + 1} = \\widehat {H} ^ {l + 1} + \\operatorname {M L P} (\\operatorname {R M S - n o r m} (\\widehat {H} ^ {l + 1})), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Where MHSA is a multi-head self attention layer, MLP is a multi-layer perceptron with SwiGLU activations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We train our models for the next token prediction task at different scales (base, large and 1b models). For more architecture details see Table 1. We train all these models with a batch size of 1M tokens. We use AdamW [33] with a maximum learning rate of \\(3e - 4\\), and \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.95\\). We decay the learning rate with a cosine schedule, after 2000 warm-up steps [55]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.896, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.327, + 0.906, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Training Loss Curves: We show the training loss curves for base, large, and 1b models trained with tokens from dVAE [45] with a vocabulary size of 8k and context length of 4k tokens (equivalent to 16 images or video frames)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.61, + 0.424 + ], + "angle": 0, + "content": "3.3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.598 + ], + "angle": 0, + "content": "To train our model, we compile a large dataset from a number of different sources. Table 2 shows the total number of images and videos used for training data, the total number of tokens, as well as the number of hours of videos in each dataset. Together these datasets contain over 100,000 hours of video data and about 2.5 trillion visual tokens. During training, each mini-batch is sampled at different ratios of datasets. Each batch approximately contains \\(20\\%\\) of ImageNet images, \\(10\\%\\) of Ego4D videos, \\(10\\%\\) of Kinetics videos, and \\(60\\%\\) of HowTo100m videos. Our full training utilized about 1 trillion tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.651, + 0.622 + ], + "angle": 0, + "content": "3.4. Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.765 + ], + "angle": 0, + "content": "We use dVAE tokenizer with a vocabulary of 8k tokens, from Dall-E [45] as our tokenizer. Using an image-based tokenizer allows training on both images and videos and testing on respective downstream tasks. While VQGAN [15] tokenizers provide sharper images, these models were trained with perceptual loss [28, 30], thus indirectly ingesting ImageNet label information via VGG-net [53]. We also explored SD-VAE based random tokenizer, please see Appendix A.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.901 + ], + "angle": 0, + "content": "All raw pixel frames or images are tokenized into 256 discrete tokens. We take a video and resize it such that its shortest size is \\( R \\) pixels, and then take a random crop of \\( R \\times R \\times T \\), and sample every 4 frames where \\( T \\) is the number of frames. We use dVAE [45] with the vocabulary of 8k entries to tokenize every frame independently. For dVAE we set \\( R = 128 \\), to get \\( 16 \\times 16 \\) discrete tokens. Once every frame is mapped into a set of discrete tokens we have \\( T \\times 256 \\) tokens per each video. We pre-train all the mod-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19110" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.452, + 0.184 + ], + "angle": 0, + "content": "
DatasetsInstancesTokensHours
ImageNet13.9M3.6B-
Kinetics-6000.53M41.3B1496
Ego4D52.1K103B3750
HowTo100m1.172M2560B92627
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.206, + 0.482, + 0.275 + ], + "angle": 0, + "content": "Table 2. Pre-training Dataset: We use both image datasets (Imagenet [48]) and video datasets (Kinetics600 [8], Ego4D [21], HowTo100m [34]) with different mixing ratios during the pretraining of our models. The whole training data contains about 100,000 hours of videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.317 + ], + "angle": 0, + "content": "els with \\( T = 16 \\), thus all the models are per-trained for a context length of 4096 tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.424 + ], + "angle": 0, + "content": "When training with images and videos, 16 video frames are sampled to create 4k tokens. For images, we randomly sample 16 images and create a sequence of 16 image frames to generate 4k tokens. Finally, we add start and end tokens for each sequence, for videos we use [1] as the start token, and for images we use [3] as the start token, and all sequences have an end token of [2]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.433, + 0.295, + 0.447 + ], + "angle": 0, + "content": "3.5. Downstream Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.483, + 0.56 + ], + "angle": 0, + "content": "The idea of large pre-trained models is that they were trained at a large compute scale, and then these models can be easily used for various downstream tasks without requiring task-specific design or lots of computing for transfer. The learned representations are general enough to transfer to various tasks. We evaluate our models on the intermediate features with linear and attention probing [31]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.818 + ], + "angle": 0, + "content": "For linear probing the model, we apply global average pooling [32] over the tokens from different layers to get the intermediate representation. We train a linear layer on top of this representation on the downstream task. MAE [25] or DINO [7] have a uniform structure when it comes to which token attends to which tokens, however in autoregressive sequence modeling later tokens attend to more tokens than the tokens at the beginning. Due to this skewed nature, equally weighting all the tokens affects the downstream performance. Attention pooling is an alternative to average pooling that allows to dynamically weight the tokens, ideally giving more weight to tokens that see more tokens. This requires learning \\( W_{k} \\) and \\( W_{v} \\) matrices and a query token \\( q \\). The query token cross-attends to the intermediate tokens and combines them into a single vector. While this function is not linear anymore, it has been shown to learn better representations in recent works [14]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.222, + 0.847 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate our pre-trained models on various downstream tasks such as ImageNet classification, Kinetics action recognition, Ego4D action anticipation, Semi-Supervised" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.646, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.09, + 0.773, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.09, + 0.903, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.169, + 0.905, + 0.238 + ], + "angle": 0, + "content": "Figure 3. 1-gram Distribution of Various Tokens: This Figure shows the distribution of 1-gram tokens of various tokenizers (dVAE [45], VQGAN-1k, VQGAN-16k [15]) on Imagenet validation set. Note that, dVAE has almost full convergence of the tokens while VQGAN has less than \\(50\\%\\) coverage of the tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.905, + 0.34 + ], + "angle": 0, + "content": "tracking, and Robotic manipulation tasks. First, we discuss various design choices for pre-training and evaluation strategies for our method. All the models for studying the design choices are large models trained for 400 epochs on the ImageNet-1k dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.351, + 0.667, + 0.366 + ], + "angle": 0, + "content": "4.1. Design Choices" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Tokenizer: The are various options available for tokenizing an image or a video. We could use discrete tokenizers such as dVAE, and VQGAN, or simple patch-based continuous tokenization. To study the behavior of various tokenizers we pre-train a Toto-large model on ImageNet for 400 epochs. Using linear probing at an optimal intermediate layer, we evaluate the accuracy of the models on ImageNet classification task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Table 3 shows linear probing accuracy when trained with various tokenizers. VQGAN [15] and dVAE [45] perform similarly with the same resolutions. However, VQGAN is contaminated with ImageNet label information via the perceptual loss. In addition, as shown in Figure 3, dVAE tokens have full coverage compared to VQGAN tokens on their 1-gram distributions. Please see the supplementary materials for details. Regressing normalized-patch targets from patch embeddings performs slightly worse than classifying discrete tokens as targets. Additionally, discrete tokens as targets and patch embeddings as inputs perform poorly compared to other methods at the given input-output resolutions. Overall, Table 3 shows that various ways of tokenization have little effect on ImageNet linear probing accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.809 + ], + "angle": 0, + "content": "How to Probe: As discussed in Section 3.5 we probe the pre-trained models at the same layer with attention pooling and average pooling, followed by linear layer. Table 13 shows attention pooling performs \\(7.9\\%\\) higher than average pooling on the ImageNet classification task. For attention pooling, we keep the embedding dimension the same as the intermediate feature dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Resolution: When training with dVAE tokens, a 256x256 image results in 1024 tokens, this is four times more number of tokens compared to patch embeddings or VQGAN tokens. If we reduce the number of tokens to 256, then the effective image resolution becomes \\(128 \\times 128\\). Table 4 shows a clear drop in performance when pre-training the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19111" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.472, + 0.228 + ], + "angle": 0, + "content": "
Input-TargetTokensVocabularyTop1
VQGAN-VQGAN16x1616k61.3
VQGAN-VQGAN16x161k61.1
dVAE-dVAE32x328k61.2
dVAE-dVAE16x168k53.2
patch-patch16x16-60.6
patch-dVAE16x168k58.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.245, + 0.486, + 0.317 + ], + "angle": 0, + "content": "Table 3. ImageNet Linear Probing Accuracy with Various Tokenizers: We compare discrete (dVAE, VQGAN) and patch embedding as input and target for pre-training our models. ImageNet top-1 accuracies are computed by linear probing at the 9th layer of the large model." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.323, + 0.458, + 0.428 + ], + "angle": 0, + "content": "
MethodCompute (MACs)Top1
dVAE/161.42 × 101753.2
dVAE/325.68 × 101761.2
dVAE/16→322.13 × 101763.2
dVAE/16→32†2.13 × 101764.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.443, + 0.485, + 0.501 + ], + "angle": 0, + "content": "Table 4. Token Resolution: While the performance is lower for a low-resolution model, when finetuned for next-patch prediction at a higher resolution, its performance surpasses the full-resolution pre-trained model. \\(\\dagger\\) Base values of the RoPE is 50,000." + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.514, + 0.402, + 0.61 + ], + "angle": 0, + "content": "
MethodTokensPoolingTop1
dVAE16x16Average53.2
dVAE16x16Attention61.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.677 + ], + "angle": 0, + "content": "Table 5. Attention vs Average Pooling: When probed at the same layers, attention pooling performs much better than average pooling of intermediate tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.856 + ], + "angle": 0, + "content": "model at \\(128 \\times 128\\) resolution. However, due to the use of relative positional embeddings (RoPE [54]), we can easily finetune the \\(128 \\times 128\\) (or \\(16 \\times 16\\) token equivalent) model for higher resolution. Surprisingly, this does better than pre-training at \\(256 \\times 256\\) resolution and requires only one epoch of finetuning. Not only does this improve the performance, but the pre-training also becomes cheaper compared to full-resolution pre-training. Additionally, Fine-tuning with higher base values of the RoPE embeddings (50,000) leads to better accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Architecture: We train various language models from GPT2 [41] with absolute sine-cosine positional embeddings, and non-transformer based model Mamba [23] only" + }, + { + "type": "table", + "bbox": [ + 0.602, + 0.09, + 0.822, + 0.157 + ], + "angle": 0, + "content": "
ModelParamsTop1
GPT2 [41]280 m48.5
Mamba [23]290 m40.7
LLaMA [55]280 m53.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.167, + 0.905, + 0.21 + ], + "angle": 0, + "content": "Table 6. Architecture: We compare sequence modeling architectures LLaMA [55], GPT2 [41], and non-transformer models, Mamba [23] on ImageNet linear probing task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.905, + 0.308 + ], + "angle": 0, + "content": "using dVAE tokens. We mimicked the GPT2 architecture and do architecture comparisons. We compare these models with LLaMA [55]. We evaluate linear probing performance at each layer of these models and report the best performance in Table 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Probing Layer: When probing the pre-trained models, especially the decoder-only model best performance is observed at the middle layers. This behavior is first observed in iGPT [9]. Figure 4 shows the peak performance on recognition occurs at about \\(50\\%\\) of the depth of the model. This behavior holds across all model sizes. While in MAE [25] and BEiT [3] encoder-decoder models, due to the uneven nature of the encoder and decoder, the best features are observed at the top of the encoder layers. However, on decoder-only models with uniformly distributed layers, the last layers perform worse on recognition tasks, mainly because these layers are trained to reconstruct the input. More probing results with various tokenizers, resolutions, and probing methods are shown in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.537, + 0.9, + 0.737 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.752, + 0.908, + 0.796 + ], + "angle": 0, + "content": "Figure 4. Probing at Different Layers: We show the attention-probing performance at each layer of our three models. Peak performance is observed at around \\(50\\%\\) depth of the models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.695, + 0.835 + ], + "angle": 0, + "content": "4.2. Image Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "To measure the representation quality of our pre-trained models, we evaluate our models on ImageNet-1k [12] classification. We apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19112" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.137, + 0.089, + 0.44, + 0.396 + ], + "angle": 0, + "content": "
MethodArch#θTop1
Discriminative Approaches
SimCLR [10]†RN50x29474.2
BYOL [22]†RN50x29477.4
SwAV [6]†RN50x29473.5
DINO [7]ViT-B/88680.1
DINOv2 [36]ViT-g/14101186.4
Generative Approaches
BEiT-L [3]ViT-L/1430762.2
AIM [14]ViT-1B/14120080.6
MAE [25]ViT-H/1463280.9
iGPT-L [9]†GPT-2138665.2
iGPT-XL [9]†GPT-2680172.0
Toto-baseLLaMA12064.7
Toto-largeLLaMA28071.1
Toto-1bLLaMA110075.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.407, + 0.482, + 0.477 + ], + "angle": 0, + "content": "Table 7. ImageNet Results: We compare discriminative and generative models on ImageNet [12] recognition task. While achieving comparable performance among generative models, our models model achieves the highest accuracy on autoregressive modeling. \\(\\dagger\\) models are evaluated with linear probing. \\(\\# \\theta\\) in millions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.672 + ], + "angle": 0, + "content": "the highest classification accuracy. We fine-tune the pretrained models further by applying self-supervised next token prediction loss in Eq 2, together with cross-entropy loss applied for probing layers (with stop-gradients). We train the probing layers for 90 epochs, with a learning rate of \\(6e^{-5}\\). We also use layer decay of 0.9 to reduce the learning rate at the early layers of the model. During this stage, all the models are fine tuned with \\(32 \\times 32\\) token resolution, on the self-supervised loss, and increase the base value of the RoPE [54] embeddings from 10,000 to 50,000 support larger resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Table 7 shows the ImageNet top-1 accuracy of our base, large and 1b models. First, there is a clear difference in terms of classification performance when it comes to discriminative models versus generative models. Instance discriminative models such as SimCLR [10], and DINO [7] are trained to separate samples from each other and they are designed to perform well on discriminative tasks. On the other hand, generative models are just trying to model the data distribution. While achieving comparable performance to other generative models on image recognition, among autoregressive generative models, our model achieved the highest top-1 accuracy. The scaling of data, and the use of tokens instead of pixels, allows our one billion parameter model to achieve similar performance compared to iGPT [9] 7 billion models." + }, + { + "type": "table", + "bbox": [ + 0.571, + 0.089, + 0.852, + 0.337 + ], + "angle": 0, + "content": "
MethodArchTop1
Discriminative Approaches
I-JEPA [1]ViT-H/1674.5
OpenCLIP [11]ViT-G/1483.3
DINOv2 [36]ViT-g/1484.4
InternVideo [61]-73.7
Generative Approaches
Hiera [49]Hiera-H/1477.0
MVD [60]ViT-H/1479.4
VideoMAE [59]ViT-L/1479.8
Toto-baseLLaMA59.3
Toto-largeLLaMA65.3
Toto-1bLLaMA74.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.347, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 8. K400 Results: We compare discriminative and generative models on Kinetics-400 [29] action recognition task. While achieving comparable performance among generative models, our models are the first to show the competitive performance on K400 with autoregressive pre-training, and shows scaling with large model sizes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.446, + 0.698, + 0.461 + ], + "angle": 0, + "content": "4.3. Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.469, + 0.907, + 0.68 + ], + "angle": 0, + "content": "We use Kinetics-400 (K400) [29] for evaluating our models on action recognition tasks. Similar to ImageNet evaluation, we apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with the highest action classification accuracy. We also fine-tune the pre-trained models on a self-supervised next-patch prediction task while training the probing layers with a classification loss. All our video models are trained with 16 frames, thus with a context length of 4096 tokens per video. When evaluating videos, we follow the protocol in SlowFast [18]. Unlike ImageNet where we evaluate the models at \\(256 \\times 256\\) resolution, on videos we only evaluate our models at \\(128 \\times 128\\) resolution, to keep the number of tokens in a similar budget." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.682, + 0.909, + 0.817 + ], + "angle": 0, + "content": "Table 8 shows the Kinetics-400 top-1 accuracy of our base, large and 1b models. Similar to ImageNet results in Table 7, we see that discriminately trained models perform better than generative models. Our models achieve comparable performance among generative models, and first to show competitive performance on action recognition with autoregressive generative modeling. All the models are trained and evaluated with 16 frames with a stride of 4 frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.696, + 0.848 + ], + "angle": 0, + "content": "4.4. Action Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "While the Kinetics dataset captures internet-style exocentric videos, Ego4D [21] videos capture egocentric videos of daily life. A general vision model should be able to rea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19113" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.485, + 0.234 + ], + "angle": 0, + "content": "
MethodNounN+VN+TTCOverall
FRCNN+Rnd [21]17.551.563.210.34
FRCNN+SF [21]17.555.195.372.07
Hiera-large [49]14.056.034.532.12
StillFast [44]16.207.474.942.48
VideoMAE-large [59]15.166.725.262.55
MAE-ST-large [19]13.716.634.942.60
Toto-large15.206.755.412.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.252, + 0.483, + 0.309 + ], + "angle": 0, + "content": "Table 9. Ego4D Results: Our model achieves comparable mean-average precision compared to previous work. We compare our method with, FRCNN+Rnd [21], FRCNN+SF [21], Hiera [49], StillFast [44], VideoMAE [59], and MAE-ST [19]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.485, + 0.599 + ], + "angle": 0, + "content": "son about both exo and ego-centric videos. Task-wise, Kinetics requires the model to reason about the action using full context (e.g., the model has seen the action), while the Ego4D short-term action anticipation v1 task requires models to predict future actions from past context. We use our models as the backbone for the pyramid network used in StillFast [44] to extract tokens at 5 layers and fuse them with the pyramid network. We fully fine-tuned our model with self-supervised next-patch loss along with task-related losses, and we observed that having the self-supervised loss improves the overall performance. Table 9 shows the performance of our large model on the Ego4D short-term action anticipation task. This task requires predicting the object to be interacted with (noun) and the type of interaction (verb) as well as time to contact (ttc) from the last seen frame to an estimated time between object-hand contact. As shown in Table 9, these tasks are difficult with maximum overall mean-average precision of 2.70." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.248, + 0.623 + ], + "angle": 0, + "content": "4.5. Video Tracking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.765 + ], + "angle": 0, + "content": "We evaluate our pre-trained models on the label propagation task on the DAVIS dataset [39], following the evaluation protocol in [27]. In contrast to the previous tasks, such as classification, this evaluation does not require fine-tuning or probing. We first use the features from the last \\( n \\) frames to find the nearest neighbor patch in the current frame. We then propagate the masks from the previous frames to the current frame. We compare with Dino [7] and MAE [25] in Table 10 and show qualitative results in Figure 7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.774, + 0.196, + 0.789 + ], + "angle": 0, + "content": "4.6. Robotics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we study the effectiveness of our pre-trained representations for robotic manipulation. We consider tasks in simulation and in the real world. Due to the runtime constraints on real hardware, we use the Toto-base models. Despite being a small model, Toto-base can achieve better performance in simulation and on-par performance to state-of-the-art robot models in real-world experiments." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.24 + ], + "angle": 0, + "content": "
Method (Res/Patch)J&FJF
DINO-base (224/8)54.352.556.1
DINO-base (224/16)33.136.230.1
MAE-base (224/16)31.534.128.9
Toto-base (256/8)42.041.243.1
Toto-large (256/8)44.844.445.1
Toto-1b (256/8)46.145.846.4
Toto-large (512/8)62.459.265.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 10. DAVIS Tracking: We report J, F, and J&F scores at the peak layers of each model. We achieves comparable performance as DINO and at large resolution (512), it outperforms all methods. Our model achieve best performance at the middle layer, see 9." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.325, + 0.707, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.436, + 0.655, + 0.447 + ], + "angle": 0, + "content": "(a) Franka Pick" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.324, + 0.898, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.436, + 0.844, + 0.447 + ], + "angle": 0, + "content": "(b) Kuka Pick" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.46, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Figure 5. Robot Manipulation with Reinforcement Learning: We compare MAE-base [42] with Toto-base pre-trained models in simulation following [64]. We find that the Toto outperforms the baseline in sample complexity, across two robots and two tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Simulation Experiments: Following the protocols in MVP [64], we use our visual pre-trained models to embed pixel observations. The model is frozen and we only take tokens at an intermediate layer, apply average pooling, and learn the linear layer on top to embed pixel observations. These observations are used to train DAgger policies for 2 tasks: Franka-pick 5a, Kuka-pick 5b. Figure 5 shows the mean success rate over training steps. Compared to the MVP baseline, our model was able to learn these tasks faster with better sample efficiency across robots and tasks. For fair comparisons, we use the best MAE model from MVP [42] which is trained on ImageNet [12], Ego4D [21] and 100DOH [50] datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Real-world Experiments: Next, we evaluate our pre-trained models on hardware. We follow the setup from [42]. We extract vision features using a pre-trained vision encoder and train a controller on top of frozen representations using behavior cloning. Specifically, we consider a cube picking tasks using a 7 DoF Franka robot, shown in Figure 8. We use the demonstrations provided by [43]. In Table 11 we compare our model to a vision encoder from [42]. We report the success rate over 16 trials with variations in object position and orientation. Our model performs favorably to a vision encoder pre-trained for robotics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19114" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.166, + 0.089, + 0.411, + 0.151 + ], + "angle": 0, + "content": "
Model# TrajSuccess
MVP24075%
Toto-base24063%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.162, + 0.483, + 0.26 + ], + "angle": 0, + "content": "Table 11. Robotics, Real-world Experiments: We compare MVP [42] and Toto on a Franka cube-picking task in the real world. Features from both models are pre-trained, frozen, and passed into a learning module trained with behavior cloning using the same demonstrations. We see that our approach performs comparably to the state-of-the-art vision backbone for robotics, despite not being designed with the robotic application in mind." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.265, + 0.457, + 0.356 + ], + "angle": 0, + "content": "
MethodModel16 frame32 frame
V3DResNet55.269.7
TFC V3DResNet54.670.2
Toto-largeLLaMa62.872.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.366, + 0.483, + 0.422 + ], + "angle": 0, + "content": "Table 12. Object Permanence: In the CATER [20] object localization task, the object is hidden under or obstructed by other objects and the goal is to predict its coarse location. Our model outperforms previous methods at 16 and 32 temporal resolutions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.279, + 0.456 + ], + "angle": 0, + "content": "4.7. Object Permanence" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.629 + ], + "angle": 0, + "content": "To quantitatively measure the performance of how well the model understands object permanence, we evaluate our models on the CATER localization task [20]. Here, a ball is moving in the scene, and the task is to find its location in the 6 by 6 grid. We fine-tune our Toto-large model on this task at temporal resolutions of 16 and 32 frames. In both cases, our pre-trained models outperform the models trained specifically for this task, such as V3D [67], TFC-V3D [67]. Table 12 shows the performance on the CATER snitch localization task, and Toto-large achieve \\(62.8\\%\\) and \\(70.9\\%\\) performance with 16 and 32 frames respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.329, + 0.654 + ], + "angle": 0, + "content": "4.8. Compute Optimal Scaling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We study the scaling properties of Toto using \\(\\mu\\)-Parameterization [65]. First we train various models, a1-a6, with linearly increasing hidden size and number of layers (Table 17). All models use the VQGAN tokenizer [15]. We then optimize the learning rate for these models, with \\(\\mu\\)-Parameterization [65]. Figure 11 shows the optimal learning rate of \\(2^{-7}\\) for all of the model widths. Once we find the optimal learning rate, we train a1-a6 models on our data mixture (Table 2). Figure 6 shows the loss vs. training compute of Toto models. We see a clear power law relationship between the compute and validation loss. Based on these experiments Toto shows a power law of \\(L(C) = 7.32 \\cdot C^{-0.0378}\\). For comparison, the GPT-3 power law relationship [4] is \\(L(C) = 2.57 \\cdot C^{-0.048}\\). While these are not comparable directly, the scaling coefficients indicate how much change in the loss to expect for additional com" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.092, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.359, + 0.907, + 0.416 + ], + "angle": 0, + "content": "Figure 6. Scaling Toto: We train multiple variants of Toto, with increasing hidden size and depth, with optimal learning rates. We plot the validation loss vs the compute spent on training in MACs. This shows a clear scaling behavior with optimal compute." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.907, + 0.662 + ], + "angle": 0, + "content": "pute. This suggests that autoregressive video models, such as Toto, scale but at a slower rate than language models. In language processing, tokenization is typically lossless and highly compressed, refined over thousands of years. In contrast, video inherently contains significant redundancy across subsequent frames—if a video lacks this redundancy, it wouldn’t truly qualify as a video. This redundancy, however, is essential for effective self-supervised training. With better tokenizers that yield more compressed and semantically rich representations, we can build more scalable video encoders and use scaling laws as benchmarks. Moreover, scaling laws could help us curate more scalable data and inform the design of model architectures, rather than relying solely on single benchmark metrics. See the limitations section in the Appendix for further discussion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We empirically study autoregressive pre-training from videos. We curate a large video dataset and conduct a largescale evaluation across a range of diverse tasks, including image recognition, video classification, video forecasting, object tracking, object permanence, and robotic manipulation. We perform extensive ablation studies to understand different design choices and compare autoregressive video pre-training to strong baselines across different tasks. We find that, despite minimal inductive biases, our approach achieves competitive performance across all tasks. Finally, we study the scaling behavior of autoregressive video models, and show that they scale with compute, but at a slower rate than autoregressive language models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19115" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "We thank Andrea Madotto, Po-Yao (Bernie) Huang, and Shiry Ginosar for helpful discussions. We're grateful to Ronghang Hu and Xinlei Chen for their help with TPU setup and code bases. We also thank Baifeng Shi for helping us with robots evaluations. We thank Valentin Gabeur and Neerja Thakkar for their valuable feedback on the paper. This work was supported by the FAIR-BAIR program as well as ONR MURI (N00014-21-1-2801)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.25, + 0.188, + 0.266 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.36, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 1954. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.483, + 0.43 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 5, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[8] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2, 5, 6, 12, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.814, + 0.483, + 0.897 + ], + "angle": 0, + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.274, + 0.484, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.148 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.906, + 0.261 + ], + "angle": 0, + "content": "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868–12878, 2020. 3, 4, 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[19] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[20] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.7, + 0.906, + 0.783 + ], + "angle": 0, + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.826 + ], + "angle": 0, + "content": "[23] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.828, + 0.905, + 0.896 + ], + "angle": 0, + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "19116" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 5, 6, 7, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[26] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[27] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[30] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In Proceedings of The 33rd International Conference on Machine Learning, pages 1558-1566, New York, New York, USA, 2016. PMLR. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.495, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[31] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 3744-3753. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.565, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[32] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.592, + 0.484, + 0.62 + ], + "angle": 0, + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[35] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.772, + 0.484, + 0.841 + ], + "angle": 0, + "content": "[36] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.841, + 0.484, + 0.898 + ], + "angle": 0, + "content": "[37] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International conference on machine learning, pages 4055-4064. PMLR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "[38] Jean Piaget. The construction of reality in the child. Routledge, 1954. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.177 + ], + "angle": 0, + "content": "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[42] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[43] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In Conference on Robot Learning, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[44] Francesco Ragusa, Giovanni Maria Farinella, and Antonino Furnari. Stillfast: An end-to-end approach for short-term object interaction anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635-3644, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3, 4, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[46] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.907, + 0.644 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.907, + 0.714 + ], + "angle": 0, + "content": "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.906, + 0.854 + ], + "angle": 0, + "content": "[50] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.855, + 0.906, + 0.884 + ], + "angle": 0, + "content": "[51] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 1951. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19117" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[52] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.122, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[53] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.164, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.291, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[56] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.347, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[57] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.403, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.459, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[59] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 2, 6, 7, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.542, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[61] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[62] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[63] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[64] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[65] Greg Yang, Edward J. Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer, 2022. 8, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[66] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[67] Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning. arXiv preprint arXiv:2203.05928, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[68] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19118" + } + ] +] \ No newline at end of file diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_origin.pdf b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3f3c9c4a20c9ee93f9adc7e9b834652267b3d4f --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/0fb150f7-5e61-492a-a2d2-7d4a735f9e3b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22335d430ca0d54f692d2582d945444307832786e2cdbdff06232cc71c1ed95d +size 1663011 diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/full.md b/2025/An Empirical Study of Autoregressive Pre-training from Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5bed1442d24452ad812a689e59c75adf902d1d4f --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/full.md @@ -0,0 +1,328 @@ +# An Empirical Study of Autoregressive Pre-training from Videos + +Jathushan Rajasegaran1,2 Ilija Radosavovic2 Rahul Ravishankar2 Yossi Gandelsman2 +Christoph Feichtenhofer1 Jitendra Malik1,2 +1Meta FAIR 2University of California, Berkeley + +# Abstract + +We empirically study autoregressive pre-training from unlabeled videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pretrained on a diverse dataset of unlabeled videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotic manipulation. Our results demonstrate that, despite minimal inductive biases, autoregressive video pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. https://brjathu.github.io/toto + +# 1. Introduction + +In a paper published in 1951, Shannon, having just published the foundational papers of information theory, proposed a "guessing game" of next word prediction to estimate the entropy of English [51]. Nearly 70 years later, training a high-capacity transformer neural network [58] on this task, provided the generative pre-training backbone for Large Language Models [5, 13, 40, 41]. + +Less well known is the fact that in 1954, Fred Attneave [2] proposed an analog of Shannon's task for images. To quote "We may divide the picture into arbitrarily small elements which we "transmit" to a subject (S) in a cumulative sequence, having them guess at the color of each successive element until they are correct. This method of analysis resembles the scanning process used in television and facsimile systems and accomplishes the like purpose of transforming two spatial dimensions into a single sequence in time". Of course this leaves open the question of what is the "element". Is it a pixel? A patch? A semantic entity? + +While Attneave was concerned with images, in the context of 2024, we have to note the "Big Visual Data" is in + +video, not images. According to the web, YouTube alone has 156 million hours of video, and 30 thousand hours of video are being uploaded every hour. While there are concerns that most of the text available on the Internet has already been used by the leading language models, in video we are barely started on the journey of big data exploitation. + +Learning from this data brings a number of challenges. The vast majority of the video data is unlabeled, or at least not consistently labeled. Therefore, an effective learning approach should be able to learn from pixels alone. Indeed, human children form rich representations and models of the world purely from observing raw video [38], which further motivates the setting of learning from unlabeled videos. What is the right video architecture and the training objective? And what is the right visual "element" to operate over? + +In this paper, we empirically study autoregressive pretraining from unlabeled videos. To perform our empirical study, we construct a family of autoregressive video models which we call Toto. We treat videos as sequences of visual tokens and train a transformer model on the next token prediction task. We use a causal transformer model with improvements from the LLaMa [55] architecture. We use dVAE [45] to tokenize frames into discrete tokens. Treating videos as sequences of tokens enables us to jointly train on both videos and images using a unified format. We construct a diverse dataset of videos and images comprising over 1 trillion visual tokens. Our models are first pretrained on this dataset and then evaluated on various downstream tasks. We extract visual representations using attention pooling from relevant layers of the model. + +We evaluate our models on a range of downstream tasks from image and video recognition, video forecasting, semi-supervised tracking, object permanence and robotic manipulation tasks in both simulation and the real world. We consider different design choices for core components of our approach. In the case of tokenizers, we consider dVAE [45], VQGAN [47] and continuous patch-normalized [25] tokens. We also consider different model architectures including LLaMA [55], GPT2 [41] and Mamba [23]. + +We find that, for tokenization autoregressive models based on discrete and continuous patch-normalized [25] + +![](images/fce19da66bba7fb84e3a749d31807d434cfa1df660dfe70f57658336d89e94bc.jpg) +Figure 1. Overall Framework. Starting with images and video frames from a collection of datasets, we tokenize each frame/image into discrete visual tokens independently. We pre-train the transformer by predicting the next visual tokens, with a context length of 4K tokens of images or video frames. Once trained, we take the intermediate representations and evaluate them on various tasks. + +![](images/a3ed2da9f518d054759fde1f086f7fc70da95e61b95280867b6b9a4c32f0cd8b.jpg) + +![](images/2a037c324d8b065ab735a91909140961c48e7d7a8eb5b88e7ebba5f81a14e2a0.jpg) + +tokens perform similarly on ImageNet classification task. For efficient pre-training, starting with lower resolution and fine-tuning at higher resolution gives better performance and RoPE [54] helps with adopting to higher resolution. For measuring the representation quality in decoder-only models, due to skewed nature of the receptive field we use attention pooling over average pooling. We find that in decoder-only models, for all tasks and models sizes the middle layer gives the best performance. + +Finally, we study the scaling behaviors of autoregressive vision models, which scales with more compute but still at a slower rate compared to large language models. We will release code and models to facilitate future research. + +# 2. Related work + +Representation Learning for Vision: Over the years self-supervised pre-training has proven to be effective in many areas including language, vision, and robotics. [63] and SimCLR [10] showed that instance discrimination training can learn strong discriminative features. MoCo [24] and DINO [7] showed the effectiveness of strong visual representations on various tasks. Differently, BEiT [3] and MAE [25] used masked autoencoding for learning image representations. ST-MAE [19] and VideoMAE [59] extended this masked modeling approach to videos, by masking a large amount of tokens during pre-training and predict the masked tokens with a light-weight decoder. + +Autoregressive Modeling of Vision: Generative autoregressive pre-training learns to directly model the data distribution. In language models, generative pre-training has become the standard for training large models. For autoregressive pre-training in vision, rCNN [46], PixelCNN [56] and PixelRNN [57] proposed generating pixels one by one using convolution and bidirectional LSTMs. With the introduction of the transformers [58], ImageTransformers [37] showed generating pixels with causal local attention per + +forms better than previous CNN and RNN-based methods. While all of these methods focused on the generation quality of the pixels, iGPT [9] showed that generative pretraining is also a good way to learn strong visual representations for recognition tasks. [26] showed scaling behaviors of autoregressive image and video models. AIM [14] on the other hand uses patch embedding rather than any pretrained models for tokenization, however, it trains on Data Filtering Networks [17] with clip filtered data. Compared to these works, we do not use any supervision during our pre-training and utilizes image and videos jointly. VisionMamba [68] also showed how to utilize sequence models with bidirectional state-space modeling for supervised vision tasks. [62] showed autoregressive video generation for promotable video generations. + +Evaluation of Vision Representations: Most video pretraining models are evaluated on semantic tasks like ImageNet [12] and Kinetics [29]. Additionally to the standard evaluation, we evaluate our models on semi-supervised tracking task on DAVIS [39], action forecasting on Ego4D [21], object permanence on CATER [20] and on robot tasks in simulation [64] and in the real world [43]. + +# 3. Approach + +We treat images and videos as sequences of visual patch tokens and train a casual transformer model to predict the next patch tokens. This is akin to the next token prediction in large language models. From the vast collection of images and videos, every patch is tokenized into a discrete token, and the transformer is trained to predict the next token, using raster scan ordering. We pre-train our models on over one trillion tokens. Finally, we evaluate the learned representations of these models on various downstream tasks including image classification, action classification, action anticipation, video tracking, object permanence, and robotic manipulation tasks. We also study the scaling behaviors of our models for compute optimal training. + +
ModelParamsDimensionHeadsLayers
base120m7681212
large280m10241616
1b1.1b20481622
+ +Table 1. Model Architecture: We pre-train models at different scales, only on visual tokens from images and videos. + +# 3.1. Pre-training + +Given a large collection of images and videos, we tokenize all of them into a 1D sequence using raster scan ordering. This produces a dataset of tokens, $\{x_1^j,x_2^j,x_3^j,\dots,x_n^j\}$ where $j$ is the sample either from a video or an image and $n$ is the number of tokens in an image or a video. We model the density $p(x)$ as: + +$$ +p \left(x ^ {j}\right) = \prod_ {i = 1} ^ {n} p \left(x _ {i} ^ {j} \mid x _ {i - 1} ^ {j}, x _ {i - 2} ^ {j}, \dots , x _ {1} ^ {j}, \theta\right) \tag {1} +$$ + +Here, $\theta$ is the model parameters, which can be optimized by minimizing the negative log-likelihood loss: + +$$ +\mathcal {L} _ {\text {p r e - t r a i n}} = \underset {x ^ {j} \sim X} {\mathbb {E}} - \log p \left(x ^ {j}\right). \tag {2} +$$ + +Using this loss, we pre-train our models at different sizes on over 1T of tokens. The tokens come from images and videos. Figure 2 shows the training loss of 3 differently sized models with 120M, 280M and 1.1B parameters. + +# 3.2. Architecture + +Our model is a transformer [58] with causal attention. We apply recent advancements in language modeling such as pre-norm using RMSNorm [66], SwiGLU activation [52], and RoPE positional embeddings [54], following LLaMa [55]. + +For a model with $L$ layers, we define $H^l$ as the intermediate representations after layer $l$ , $0 \leq l \leq L$ . The intermediate representations after layer $l + 1$ , $H^{l + 1}$ , given as: + +$$ +\widehat {H} ^ {l + 1} = H ^ {l} + \operatorname {M H S A} (\operatorname {R M S - n o r m} (H ^ {l})) \tag {3} +$$ + +$$ +H ^ {l + 1} = \widehat {H} ^ {l + 1} + \operatorname {M L P} (\operatorname {R M S - n o r m} (\widehat {H} ^ {l + 1})), \tag {4} +$$ + +Where MHSA is a multi-head self attention layer, MLP is a multi-layer perceptron with SwiGLU activations. + +We train our models for the next token prediction task at different scales (base, large and 1b models). For more architecture details see Table 1. We train all these models with a batch size of 1M tokens. We use AdamW [33] with a maximum learning rate of $3e - 4$ , and $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ . We decay the learning rate with a cosine schedule, after 2000 warm-up steps [55]. + +![](images/d9ef25640e4d5f02b99dcec6243b12b909acafcd81fdbef0ab3c4cb11d0f3375.jpg) +Figure 2. Training Loss Curves: We show the training loss curves for base, large, and 1b models trained with tokens from dVAE [45] with a vocabulary size of 8k and context length of 4k tokens (equivalent to 16 images or video frames). + +# 3.3. Dataset + +To train our model, we compile a large dataset from a number of different sources. Table 2 shows the total number of images and videos used for training data, the total number of tokens, as well as the number of hours of videos in each dataset. Together these datasets contain over 100,000 hours of video data and about 2.5 trillion visual tokens. During training, each mini-batch is sampled at different ratios of datasets. Each batch approximately contains $20\%$ of ImageNet images, $10\%$ of Ego4D videos, $10\%$ of Kinetics videos, and $60\%$ of HowTo100m videos. Our full training utilized about 1 trillion tokens. + +# 3.4. Tokenization + +We use dVAE tokenizer with a vocabulary of 8k tokens, from Dall-E [45] as our tokenizer. Using an image-based tokenizer allows training on both images and videos and testing on respective downstream tasks. While VQGAN [15] tokenizers provide sharper images, these models were trained with perceptual loss [28, 30], thus indirectly ingesting ImageNet label information via VGG-net [53]. We also explored SD-VAE based random tokenizer, please see Appendix A.4. + +All raw pixel frames or images are tokenized into 256 discrete tokens. We take a video and resize it such that its shortest size is $R$ pixels, and then take a random crop of $R \times R \times T$ , and sample every 4 frames where $T$ is the number of frames. We use dVAE [45] with the vocabulary of 8k entries to tokenize every frame independently. For dVAE we set $R = 128$ , to get $16 \times 16$ discrete tokens. Once every frame is mapped into a set of discrete tokens we have $T \times 256$ tokens per each video. We pre-train all the mod- + +
DatasetsInstancesTokensHours
ImageNet13.9M3.6B-
Kinetics-6000.53M41.3B1496
Ego4D52.1K103B3750
HowTo100m1.172M2560B92627
+ +Table 2. Pre-training Dataset: We use both image datasets (Imagenet [48]) and video datasets (Kinetics600 [8], Ego4D [21], HowTo100m [34]) with different mixing ratios during the pretraining of our models. The whole training data contains about 100,000 hours of videos. + +els with $T = 16$ , thus all the models are per-trained for a context length of 4096 tokens. + +When training with images and videos, 16 video frames are sampled to create 4k tokens. For images, we randomly sample 16 images and create a sequence of 16 image frames to generate 4k tokens. Finally, we add start and end tokens for each sequence, for videos we use [1] as the start token, and for images we use [3] as the start token, and all sequences have an end token of [2]. + +# 3.5. Downstream Transfer + +The idea of large pre-trained models is that they were trained at a large compute scale, and then these models can be easily used for various downstream tasks without requiring task-specific design or lots of computing for transfer. The learned representations are general enough to transfer to various tasks. We evaluate our models on the intermediate features with linear and attention probing [31]. + +For linear probing the model, we apply global average pooling [32] over the tokens from different layers to get the intermediate representation. We train a linear layer on top of this representation on the downstream task. MAE [25] or DINO [7] have a uniform structure when it comes to which token attends to which tokens, however in autoregressive sequence modeling later tokens attend to more tokens than the tokens at the beginning. Due to this skewed nature, equally weighting all the tokens affects the downstream performance. Attention pooling is an alternative to average pooling that allows to dynamically weight the tokens, ideally giving more weight to tokens that see more tokens. This requires learning $W_{k}$ and $W_{v}$ matrices and a query token $q$ . The query token cross-attends to the intermediate tokens and combines them into a single vector. While this function is not linear anymore, it has been shown to learn better representations in recent works [14]. + +# 4. Experiments + +We evaluate our pre-trained models on various downstream tasks such as ImageNet classification, Kinetics action recognition, Ego4D action anticipation, Semi-Supervised + +![](images/2f37372431bdc0dcb47c1654075b4ebd2e3d26c6be27b11880d40aee3ff30867.jpg) +Figure 3. 1-gram Distribution of Various Tokens: This Figure shows the distribution of 1-gram tokens of various tokenizers (dVAE [45], VQGAN-1k, VQGAN-16k [15]) on Imagenet validation set. Note that, dVAE has almost full convergence of the tokens while VQGAN has less than $50\%$ coverage of the tokens. + +![](images/28d79a5c981899eb1db81caf03e3a87c9a7458d29a95aa10b1a1daab39411605.jpg) + +![](images/48ac95f631af77be480ce08c80a91971f8fa86feecdb0e1c95516a85d6412fa9.jpg) + +tracking, and Robotic manipulation tasks. First, we discuss various design choices for pre-training and evaluation strategies for our method. All the models for studying the design choices are large models trained for 400 epochs on the ImageNet-1k dataset. + +# 4.1. Design Choices + +Tokenizer: The are various options available for tokenizing an image or a video. We could use discrete tokenizers such as dVAE, and VQGAN, or simple patch-based continuous tokenization. To study the behavior of various tokenizers we pre-train a Toto-large model on ImageNet for 400 epochs. Using linear probing at an optimal intermediate layer, we evaluate the accuracy of the models on ImageNet classification task. + +Table 3 shows linear probing accuracy when trained with various tokenizers. VQGAN [15] and dVAE [45] perform similarly with the same resolutions. However, VQGAN is contaminated with ImageNet label information via the perceptual loss. In addition, as shown in Figure 3, dVAE tokens have full coverage compared to VQGAN tokens on their 1-gram distributions. Please see the supplementary materials for details. Regressing normalized-patch targets from patch embeddings performs slightly worse than classifying discrete tokens as targets. Additionally, discrete tokens as targets and patch embeddings as inputs perform poorly compared to other methods at the given input-output resolutions. Overall, Table 3 shows that various ways of tokenization have little effect on ImageNet linear probing accuracy. + +How to Probe: As discussed in Section 3.5 we probe the pre-trained models at the same layer with attention pooling and average pooling, followed by linear layer. Table 13 shows attention pooling performs $7.9\%$ higher than average pooling on the ImageNet classification task. For attention pooling, we keep the embedding dimension the same as the intermediate feature dimensions. + +Resolution: When training with dVAE tokens, a 256x256 image results in 1024 tokens, this is four times more number of tokens compared to patch embeddings or VQGAN tokens. If we reduce the number of tokens to 256, then the effective image resolution becomes $128 \times 128$ . Table 4 shows a clear drop in performance when pre-training the + +
Input-TargetTokensVocabularyTop1
VQGAN-VQGAN16x1616k61.3
VQGAN-VQGAN16x161k61.1
dVAE-dVAE32x328k61.2
dVAE-dVAE16x168k53.2
patch-patch16x16-60.6
patch-dVAE16x168k58.5
+ +Table 3. ImageNet Linear Probing Accuracy with Various Tokenizers: We compare discrete (dVAE, VQGAN) and patch embedding as input and target for pre-training our models. ImageNet top-1 accuracies are computed by linear probing at the 9th layer of the large model. + +
MethodCompute (MACs)Top1
dVAE/161.42 × 101753.2
dVAE/325.68 × 101761.2
dVAE/16→322.13 × 101763.2
dVAE/16→32†2.13 × 101764.4
+ +Table 4. Token Resolution: While the performance is lower for a low-resolution model, when finetuned for next-patch prediction at a higher resolution, its performance surpasses the full-resolution pre-trained model. $\dagger$ Base values of the RoPE is 50,000. + +
MethodTokensPoolingTop1
dVAE16x16Average53.2
dVAE16x16Attention61.1
+ +model at $128 \times 128$ resolution. However, due to the use of relative positional embeddings (RoPE [54]), we can easily finetune the $128 \times 128$ (or $16 \times 16$ token equivalent) model for higher resolution. Surprisingly, this does better than pre-training at $256 \times 256$ resolution and requires only one epoch of finetuning. Not only does this improve the performance, but the pre-training also becomes cheaper compared to full-resolution pre-training. Additionally, Fine-tuning with higher base values of the RoPE embeddings (50,000) leads to better accuracy. + +Architecture: We train various language models from GPT2 [41] with absolute sine-cosine positional embeddings, and non-transformer based model Mamba [23] only + +Table 5. Attention vs Average Pooling: When probed at the same layers, attention pooling performs much better than average pooling of intermediate tokens. + +
ModelParamsTop1
GPT2 [41]280 m48.5
Mamba [23]290 m40.7
LLaMA [55]280 m53.2
+ +Table 6. Architecture: We compare sequence modeling architectures LLaMA [55], GPT2 [41], and non-transformer models, Mamba [23] on ImageNet linear probing task. + +using dVAE tokens. We mimicked the GPT2 architecture and do architecture comparisons. We compare these models with LLaMA [55]. We evaluate linear probing performance at each layer of these models and report the best performance in Table 6. + +Probing Layer: When probing the pre-trained models, especially the decoder-only model best performance is observed at the middle layers. This behavior is first observed in iGPT [9]. Figure 4 shows the peak performance on recognition occurs at about $50\%$ of the depth of the model. This behavior holds across all model sizes. While in MAE [25] and BEiT [3] encoder-decoder models, due to the uneven nature of the encoder and decoder, the best features are observed at the top of the encoder layers. However, on decoder-only models with uniformly distributed layers, the last layers perform worse on recognition tasks, mainly because these layers are trained to reconstruct the input. More probing results with various tokenizers, resolutions, and probing methods are shown in the supplementary material. + +![](images/eee90fbf8aa5fa427afc518ab23b49f934ba29cb5dc72f3d9cf9c2bba989ed5e.jpg) +Figure 4. Probing at Different Layers: We show the attention-probing performance at each layer of our three models. Peak performance is observed at around $50\%$ depth of the models. + +# 4.2. Image Recognition + +To measure the representation quality of our pre-trained models, we evaluate our models on ImageNet-1k [12] classification. We apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with + +
MethodArch#θTop1
Discriminative Approaches
SimCLR [10]†RN50x29474.2
BYOL [22]†RN50x29477.4
SwAV [6]†RN50x29473.5
DINO [7]ViT-B/88680.1
DINOv2 [36]ViT-g/14101186.4
Generative Approaches
BEiT-L [3]ViT-L/1430762.2
AIM [14]ViT-1B/14120080.6
MAE [25]ViT-H/1463280.9
iGPT-L [9]†GPT-2138665.2
iGPT-XL [9]†GPT-2680172.0
Toto-baseLLaMA12064.7
Toto-largeLLaMA28071.1
Toto-1bLLaMA110075.3
+ +the highest classification accuracy. We fine-tune the pretrained models further by applying self-supervised next token prediction loss in Eq 2, together with cross-entropy loss applied for probing layers (with stop-gradients). We train the probing layers for 90 epochs, with a learning rate of $6e^{-5}$ . We also use layer decay of 0.9 to reduce the learning rate at the early layers of the model. During this stage, all the models are fine tuned with $32 \times 32$ token resolution, on the self-supervised loss, and increase the base value of the RoPE [54] embeddings from 10,000 to 50,000 support larger resolution. + +Table 7 shows the ImageNet top-1 accuracy of our base, large and 1b models. First, there is a clear difference in terms of classification performance when it comes to discriminative models versus generative models. Instance discriminative models such as SimCLR [10], and DINO [7] are trained to separate samples from each other and they are designed to perform well on discriminative tasks. On the other hand, generative models are just trying to model the data distribution. While achieving comparable performance to other generative models on image recognition, among autoregressive generative models, our model achieved the highest top-1 accuracy. The scaling of data, and the use of tokens instead of pixels, allows our one billion parameter model to achieve similar performance compared to iGPT [9] 7 billion models. + +Table 7. ImageNet Results: We compare discriminative and generative models on ImageNet [12] recognition task. While achieving comparable performance among generative models, our models model achieves the highest accuracy on autoregressive modeling. $\dagger$ models are evaluated with linear probing. $\# \theta$ in millions. + +
MethodArchTop1
Discriminative Approaches
I-JEPA [1]ViT-H/1674.5
OpenCLIP [11]ViT-G/1483.3
DINOv2 [36]ViT-g/1484.4
InternVideo [61]-73.7
Generative Approaches
Hiera [49]Hiera-H/1477.0
MVD [60]ViT-H/1479.4
VideoMAE [59]ViT-L/1479.8
Toto-baseLLaMA59.3
Toto-largeLLaMA65.3
Toto-1bLLaMA74.4
+ +Table 8. K400 Results: We compare discriminative and generative models on Kinetics-400 [29] action recognition task. While achieving comparable performance among generative models, our models are the first to show the competitive performance on K400 with autoregressive pre-training, and shows scaling with large model sizes. + +# 4.3. Action Recognition + +We use Kinetics-400 (K400) [29] for evaluating our models on action recognition tasks. Similar to ImageNet evaluation, we apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with the highest action classification accuracy. We also fine-tune the pre-trained models on a self-supervised next-patch prediction task while training the probing layers with a classification loss. All our video models are trained with 16 frames, thus with a context length of 4096 tokens per video. When evaluating videos, we follow the protocol in SlowFast [18]. Unlike ImageNet where we evaluate the models at $256 \times 256$ resolution, on videos we only evaluate our models at $128 \times 128$ resolution, to keep the number of tokens in a similar budget. + +Table 8 shows the Kinetics-400 top-1 accuracy of our base, large and 1b models. Similar to ImageNet results in Table 7, we see that discriminately trained models perform better than generative models. Our models achieve comparable performance among generative models, and first to show competitive performance on action recognition with autoregressive generative modeling. All the models are trained and evaluated with 16 frames with a stride of 4 frames. + +# 4.4. Action Forecasting + +While the Kinetics dataset captures internet-style exocentric videos, Ego4D [21] videos capture egocentric videos of daily life. A general vision model should be able to rea + +
MethodNounN+VN+TTCOverall
FRCNN+Rnd [21]17.551.563.210.34
FRCNN+SF [21]17.555.195.372.07
Hiera-large [49]14.056.034.532.12
StillFast [44]16.207.474.942.48
VideoMAE-large [59]15.166.725.262.55
MAE-ST-large [19]13.716.634.942.60
Toto-large15.206.755.412.70
+ +son about both exo and ego-centric videos. Task-wise, Kinetics requires the model to reason about the action using full context (e.g., the model has seen the action), while the Ego4D short-term action anticipation v1 task requires models to predict future actions from past context. We use our models as the backbone for the pyramid network used in StillFast [44] to extract tokens at 5 layers and fuse them with the pyramid network. We fully fine-tuned our model with self-supervised next-patch loss along with task-related losses, and we observed that having the self-supervised loss improves the overall performance. Table 9 shows the performance of our large model on the Ego4D short-term action anticipation task. This task requires predicting the object to be interacted with (noun) and the type of interaction (verb) as well as time to contact (ttc) from the last seen frame to an estimated time between object-hand contact. As shown in Table 9, these tasks are difficult with maximum overall mean-average precision of 2.70. + +# 4.5. Video Tracking + +We evaluate our pre-trained models on the label propagation task on the DAVIS dataset [39], following the evaluation protocol in [27]. In contrast to the previous tasks, such as classification, this evaluation does not require fine-tuning or probing. We first use the features from the last $n$ frames to find the nearest neighbor patch in the current frame. We then propagate the masks from the previous frames to the current frame. We compare with Dino [7] and MAE [25] in Table 10 and show qualitative results in Figure 7. + +# 4.6. Robotics + +In this section, we study the effectiveness of our pre-trained representations for robotic manipulation. We consider tasks in simulation and in the real world. Due to the runtime constraints on real hardware, we use the Toto-base models. Despite being a small model, Toto-base can achieve better performance in simulation and on-par performance to state-of-the-art robot models in real-world experiments. + +Table 9. Ego4D Results: Our model achieves comparable mean-average precision compared to previous work. We compare our method with, FRCNN+Rnd [21], FRCNN+SF [21], Hiera [49], StillFast [44], VideoMAE [59], and MAE-ST [19]. + +
Method (Res/Patch)J&FJF
DINO-base (224/8)54.352.556.1
DINO-base (224/16)33.136.230.1
MAE-base (224/16)31.534.128.9
Toto-base (256/8)42.041.243.1
Toto-large (256/8)44.844.445.1
Toto-1b (256/8)46.145.846.4
Toto-large (512/8)62.459.265.6
+ +Table 10. DAVIS Tracking: We report J, F, and J&F scores at the peak layers of each model. We achieves comparable performance as DINO and at large resolution (512), it outperforms all methods. Our model achieve best performance at the middle layer, see 9. + +![](images/119e676c787020145b1a2588468ddb1c7e50f1bd2fb29696c4c21df68f147483.jpg) +(a) Franka Pick + +![](images/65de3d672e8e31f7b287f4b6a0d2e5a688742ea3503ef1823dedcd3b0d0d818e.jpg) +(b) Kuka Pick +Figure 5. Robot Manipulation with Reinforcement Learning: We compare MAE-base [42] with Toto-base pre-trained models in simulation following [64]. We find that the Toto outperforms the baseline in sample complexity, across two robots and two tasks. + +Simulation Experiments: Following the protocols in MVP [64], we use our visual pre-trained models to embed pixel observations. The model is frozen and we only take tokens at an intermediate layer, apply average pooling, and learn the linear layer on top to embed pixel observations. These observations are used to train DAgger policies for 2 tasks: Franka-pick 5a, Kuka-pick 5b. Figure 5 shows the mean success rate over training steps. Compared to the MVP baseline, our model was able to learn these tasks faster with better sample efficiency across robots and tasks. For fair comparisons, we use the best MAE model from MVP [42] which is trained on ImageNet [12], Ego4D [21] and 100DOH [50] datasets. + +Real-world Experiments: Next, we evaluate our pre-trained models on hardware. We follow the setup from [42]. We extract vision features using a pre-trained vision encoder and train a controller on top of frozen representations using behavior cloning. Specifically, we consider a cube picking tasks using a 7 DoF Franka robot, shown in Figure 8. We use the demonstrations provided by [43]. In Table 11 we compare our model to a vision encoder from [42]. We report the success rate over 16 trials with variations in object position and orientation. Our model performs favorably to a vision encoder pre-trained for robotics. + +
Model# TrajSuccess
MVP24075%
Toto-base24063%
+ +Table 11. Robotics, Real-world Experiments: We compare MVP [42] and Toto on a Franka cube-picking task in the real world. Features from both models are pre-trained, frozen, and passed into a learning module trained with behavior cloning using the same demonstrations. We see that our approach performs comparably to the state-of-the-art vision backbone for robotics, despite not being designed with the robotic application in mind. + +
MethodModel16 frame32 frame
V3DResNet55.269.7
TFC V3DResNet54.670.2
Toto-largeLLaMa62.872.9
+ +Table 12. Object Permanence: In the CATER [20] object localization task, the object is hidden under or obstructed by other objects and the goal is to predict its coarse location. Our model outperforms previous methods at 16 and 32 temporal resolutions. + +# 4.7. Object Permanence + +To quantitatively measure the performance of how well the model understands object permanence, we evaluate our models on the CATER localization task [20]. Here, a ball is moving in the scene, and the task is to find its location in the 6 by 6 grid. We fine-tune our Toto-large model on this task at temporal resolutions of 16 and 32 frames. In both cases, our pre-trained models outperform the models trained specifically for this task, such as V3D [67], TFC-V3D [67]. Table 12 shows the performance on the CATER snitch localization task, and Toto-large achieve $62.8\%$ and $70.9\%$ performance with 16 and 32 frames respectively. + +# 4.8. Compute Optimal Scaling + +We study the scaling properties of Toto using $\mu$ -Parameterization [65]. First we train various models, a1-a6, with linearly increasing hidden size and number of layers (Table 17). All models use the VQGAN tokenizer [15]. We then optimize the learning rate for these models, with $\mu$ -Parameterization [65]. Figure 11 shows the optimal learning rate of $2^{-7}$ for all of the model widths. Once we find the optimal learning rate, we train a1-a6 models on our data mixture (Table 2). Figure 6 shows the loss vs. training compute of Toto models. We see a clear power law relationship between the compute and validation loss. Based on these experiments Toto shows a power law of $L(C) = 7.32 \cdot C^{-0.0378}$ . For comparison, the GPT-3 power law relationship [4] is $L(C) = 2.57 \cdot C^{-0.048}$ . While these are not comparable directly, the scaling coefficients indicate how much change in the loss to expect for additional com + +![](images/c1f099cfe78788d43d19228d218d4bae1dcc9c06f02bc263f2ca86c9ab55ff8a.jpg) +Figure 6. Scaling Toto: We train multiple variants of Toto, with increasing hidden size and depth, with optimal learning rates. We plot the validation loss vs the compute spent on training in MACs. This shows a clear scaling behavior with optimal compute. + +pute. This suggests that autoregressive video models, such as Toto, scale but at a slower rate than language models. In language processing, tokenization is typically lossless and highly compressed, refined over thousands of years. In contrast, video inherently contains significant redundancy across subsequent frames—if a video lacks this redundancy, it wouldn’t truly qualify as a video. This redundancy, however, is essential for effective self-supervised training. With better tokenizers that yield more compressed and semantically rich representations, we can build more scalable video encoders and use scaling laws as benchmarks. Moreover, scaling laws could help us curate more scalable data and inform the design of model architectures, rather than relying solely on single benchmark metrics. See the limitations section in the Appendix for further discussion. + +# 5. Conclusion + +We empirically study autoregressive pre-training from videos. We curate a large video dataset and conduct a largescale evaluation across a range of diverse tasks, including image recognition, video classification, video forecasting, object tracking, object permanence, and robotic manipulation. We perform extensive ablation studies to understand different design choices and compare autoregressive video pre-training to strong baselines across different tasks. We find that, despite minimal inductive biases, our approach achieves competitive performance across all tasks. Finally, we study the scaling behavior of autoregressive video models, and show that they scale with compute, but at a slower rate than autoregressive language models. + +# 6. Acknowledgments + +We thank Andrea Madotto, Po-Yao (Bernie) Huang, and Shiry Ginosar for helpful discussions. We're grateful to Ronghang Hu and Xinlei Chen for their help with TPU setup and code bases. We also thank Baifeng Shi for helping us with robots evaluations. We thank Valentin Gabeur and Neerja Thakkar for their valuable feedback on the paper. This work was supported by the FAIR-BAIR program as well as ONR MURI (N00014-21-1-2801). + +# References + +[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 6 +[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 1954. 1 +[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 5, 6, 12 +[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 8 +[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1 +[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6 +[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 6, 7, 12 +[8] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 4 +[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2, 5, 6, 12, 14 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6 +[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6 + +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 5, 6, 7 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 1 +[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 2, 4, 6 +[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868–12878, 2020. 3, 4, 8, 14 +[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 12 +[17] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 2 +[18] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6 +[19] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2, 7 +[20] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 2, 8 +[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4, 6, 7 +[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6 +[23] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 5 +[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 + +[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 5, 6, 7, 12 +[26] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2 +[27] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7, 13 +[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 3 +[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 12 +[30] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In Proceedings of The 33rd International Conference on Machine Learning, pages 1558-1566, New York, New York, USA, 2016. PMLR. 3 +[31] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 3744-3753. PMLR, 2019. 4 +[32] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013. 4 +[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3 +[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 4 +[35] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 12 +[36] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[37] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International conference on machine learning, pages 4055-4064. PMLR, 2018. 2 + +[38] Jean Piaget. The construction of reality in the child. Routledge, 1954. 1 +[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7 +[40] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 1 +[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2019. 1, 5 +[42] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, 2022. 7, 8 +[43] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In Conference on Robot Learning, 2023. 2, 7 +[44] Francesco Ragusa, Giovanni Maria Farinella, and Antonino Furnari. Stillfast: An end-to-end approach for short-term object interaction anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635-3644, 2023. 7 +[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3, 4, 14 +[46] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 2 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 4 +[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 6, 7, 14 +[50] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 7 +[51] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 1951. 1 + +[52] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 3 +[53] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3 +[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 3, 5, 6 +[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 14 +[56] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2 +[57] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2, 12 +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3 +[59] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 2, 6, 7, 14 +[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang. + +Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, 2023. 6 +[61] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6 +[62] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2 +[63] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2 +[64] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2, 7 +[65] Greg Yang, Edward J. Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer, 2022. 8, 14 +[66] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 3 +[67] Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning. arXiv preprint arXiv:2203.05928, 2022. 8 +[68] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2 \ No newline at end of file diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/images.zip b/2025/An Empirical Study of Autoregressive Pre-training from Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b1ea6ee39e3b700be5eb719a71519bb1c7721385 --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f01189e87576d79250e9215b242f75d3e961c5a4ab259f0c0e568f613ba9a3f5 +size 508232 diff --git a/2025/An Empirical Study of Autoregressive Pre-training from Videos/layout.json b/2025/An Empirical Study of Autoregressive Pre-training from Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..494bb0b7a78cd29b09d7029cffcf06b22d992898 --- /dev/null +++ b/2025/An Empirical Study of Autoregressive Pre-training from Videos/layout.json @@ -0,0 +1,8407 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 109, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 109, + 103, + 501, + 121 + ], + "type": "text", + "content": "An Empirical Study of Autoregressive Pre-training from Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 142, + 518, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 142, + 518, + 186 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 518, + 186 + ], + "type": "text", + "content": "Jathushan Rajasegaran1,2 Ilija Radosavovic2 Rahul Ravishankar2 Yossi Gandelsman2 \nChristoph Feichtenhofer1 Jitendra Malik1,2 \n1Meta FAIR 2University of California, Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 297, + 441 + ], + "type": "text", + "content": "We empirically study autoregressive pre-training from unlabeled videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pretrained on a diverse dataset of unlabeled videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotic manipulation. Our results demonstrate that, despite minimal inductive biases, autoregressive video pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. https://brjathu.github.io/toto" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 135, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 558 + ], + "type": "text", + "content": "In a paper published in 1951, Shannon, having just published the foundational papers of information theory, proposed a \"guessing game\" of next word prediction to estimate the entropy of English [51]. Nearly 70 years later, training a high-capacity transformer neural network [58] on this task, provided the generative pre-training backbone for Large Language Models [5, 13, 40, 41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": "Less well known is the fact that in 1954, Fred Attneave [2] proposed an analog of Shannon's task for images. To quote \"We may divide the picture into arbitrarily small elements which we \"transmit\" to a subject (S) in a cumulative sequence, having them guess at the color of each successive element until they are correct. This method of analysis resembles the scanning process used in television and facsimile systems and accomplishes the like purpose of transforming two spatial dimensions into a single sequence in time\". Of course this leaves open the question of what is the \"element\". Is it a pixel? A patch? A semantic entity?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "While Attneave was concerned with images, in the context of 2024, we have to note the \"Big Visual Data\" is in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 285 + ], + "type": "text", + "content": "video, not images. According to the web, YouTube alone has 156 million hours of video, and 30 thousand hours of video are being uploaded every hour. While there are concerns that most of the text available on the Internet has already been used by the leading language models, in video we are barely started on the journey of big data exploitation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": "Learning from this data brings a number of challenges. The vast majority of the video data is unlabeled, or at least not consistently labeled. Therefore, an effective learning approach should be able to learn from pixels alone. Indeed, human children form rich representations and models of the world purely from observing raw video [38], which further motivates the setting of learning from unlabeled videos. What is the right video architecture and the training objective? And what is the right visual \"element\" to operate over?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 398, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 578 + ], + "type": "text", + "content": "In this paper, we empirically study autoregressive pretraining from unlabeled videos. To perform our empirical study, we construct a family of autoregressive video models which we call Toto. We treat videos as sequences of visual tokens and train a transformer model on the next token prediction task. We use a causal transformer model with improvements from the LLaMa [55] architecture. We use dVAE [45] to tokenize frames into discrete tokens. Treating videos as sequences of tokens enables us to jointly train on both videos and images using a unified format. We construct a diverse dataset of videos and images comprising over 1 trillion visual tokens. Our models are first pretrained on this dataset and then evaluated on various downstream tasks. We extract visual representations using attention pooling from relevant layers of the model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 579, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 687 + ], + "type": "text", + "content": "We evaluate our models on a range of downstream tasks from image and video recognition, video forecasting, semi-supervised tracking, object permanence and robotic manipulation tasks in both simulation and the real world. We consider different design choices for core components of our approach. In the case of tokenizers, we consider dVAE [45], VQGAN [47] and continuous patch-normalized [25] tokens. We also consider different model architectures including LLaMA [55], GPT2 [41] and Mamba [23]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We find that, for tokenization autoregressive models based on discrete and continuous patch-normalized [25]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19108" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 146, + 203 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 146, + 203 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 146, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 146, + 203 + ], + "type": "image", + "image_path": "fce19da66bba7fb84e3a749d31807d434cfa1df660dfe70f57658336d89e94bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 216, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 216, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 555, + 251 + ], + "type": "text", + "content": "Figure 1. Overall Framework. Starting with images and video frames from a collection of datasets, we tokenize each frame/image into discrete visual tokens independently. We pre-train the transformer by predicting the next visual tokens, with a context length of 4K tokens of images or video frames. Once trained, we take the intermediate representations and evaluate them on various tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 72, + 460, + 204 + ], + "blocks": [ + { + "bbox": [ + 151, + 72, + 460, + 204 + ], + "lines": [ + { + "bbox": [ + 151, + 72, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 460, + 204 + ], + "type": "image", + "image_path": "a3ed2da9f518d054759fde1f086f7fc70da95e61b95280867b6b9a4c32f0cd8b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 465, + 72, + 548, + 204 + ], + "blocks": [ + { + "bbox": [ + 465, + 72, + 548, + 204 + ], + "lines": [ + { + "bbox": [ + 465, + 72, + 548, + 204 + ], + "spans": [ + { + "bbox": [ + 465, + 72, + 548, + 204 + ], + "type": "image", + "image_path": "2a037c324d8b065ab735a91909140961c48e7d7a8eb5b88e7ebba5f81a14e2a0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 270, + 294, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 294, + 378 + ], + "type": "text", + "content": "tokens perform similarly on ImageNet classification task. For efficient pre-training, starting with lower resolution and fine-tuning at higher resolution gives better performance and RoPE [54] helps with adopting to higher resolution. For measuring the representation quality in decoder-only models, due to skewed nature of the receptive field we use attention pooling over average pooling. We find that in decoder-only models, for all tasks and models sizes the middle layer gives the best performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 427 + ], + "type": "text", + "content": "Finally, we study the scaling behaviors of autoregressive vision models, which scales with more compute but still at a slower rate compared to large language models. We will release code and models to facilitate future research." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 139, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 139, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 139, + 453 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 605 + ], + "type": "text", + "content": "Representation Learning for Vision: Over the years self-supervised pre-training has proven to be effective in many areas including language, vision, and robotics. [63] and SimCLR [10] showed that instance discrimination training can learn strong discriminative features. MoCo [24] and DINO [7] showed the effectiveness of strong visual representations on various tasks. Differently, BEiT [3] and MAE [25] used masked autoencoding for learning image representations. ST-MAE [19] and VideoMAE [59] extended this masked modeling approach to videos, by masking a large amount of tokens during pre-training and predict the masked tokens with a light-weight decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Autoregressive Modeling of Vision: Generative autoregressive pre-training learns to directly model the data distribution. In language models, generative pre-training has become the standard for training large models. For autoregressive pre-training in vision, rCNN [46], PixelCNN [56] and PixelRNN [57] proposed generating pixels one by one using convolution and bidirectional LSTMs. With the introduction of the transformers [58], ImageTransformers [37] showed generating pixels with causal local attention per" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 449 + ], + "type": "text", + "content": "forms better than previous CNN and RNN-based methods. While all of these methods focused on the generation quality of the pixels, iGPT [9] showed that generative pretraining is also a good way to learn strong visual representations for recognition tasks. [26] showed scaling behaviors of autoregressive image and video models. AIM [14] on the other hand uses patch embedding rather than any pretrained models for tokenization, however, it trains on Data Filtering Networks [17] with clip filtered data. Compared to these works, we do not use any supervision during our pre-training and utilizes image and videos jointly. VisionMamba [68] also showed how to utilize sequence models with bidirectional state-space modeling for supervised vision tasks. [62] showed autoregressive video generation for promotable video generations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 535 + ], + "type": "text", + "content": "Evaluation of Vision Representations: Most video pretraining models are evaluated on semantic tasks like ImageNet [12] and Kinetics [29]. Additionally to the standard evaluation, we evaluate our models on semi-supervised tracking task on DAVIS [39], action forecasting on Ego4D [21], object permanence on CATER [20] and on robot tasks in simulation [64] and in the real world [43]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 538, + 382, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 382, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 382, + 552 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "We treat images and videos as sequences of visual patch tokens and train a casual transformer model to predict the next patch tokens. This is akin to the next token prediction in large language models. From the vast collection of images and videos, every patch is tokenized into a discrete token, and the transformer is trained to predict the next token, using raster scan ordering. We pre-train our models on over one trillion tokens. Finally, we evaluate the learned representations of these models on various downstream tasks including image classification, action classification, action anticipation, video tracking, object permanence, and robotic manipulation tasks. We also study the scaling behaviors of our models for compute optimal training." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19109" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 128 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 128 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 128 + ], + "type": "table", + "html": "
ModelParamsDimensionHeadsLayers
base120m7681212
large280m10241616
1b1.1b20481622
", + "image_path": "69d6a067bbfbe103aeea571f4c318348c4e740d8fecbd5f7c95ccafd1d2cc393.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 295, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 135, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 295, + 157 + ], + "type": "text", + "content": "Table 1. Model Architecture: We pre-train models at different scales, only on visual tokens from images and videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 171, + 137, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 137, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 137, + 184 + ], + "type": "text", + "content": "3.1. Pre-training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "content": "Given a large collection of images and videos, we tokenize all of them into a 1D sequence using raster scan ordering. This produces a dataset of tokens, " + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "inline_equation", + "content": "\\{x_1^j,x_2^j,x_3^j,\\dots,x_n^j\\}" + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "content": " is the sample either from a video or an image and " + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "content": " is the number of tokens in an image or a video. We model the density " + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "inline_equation", + "content": "p(x)" + }, + { + "bbox": [ + 55, + 189, + 296, + 262 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 273, + 295, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 273, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 95, + 273, + 295, + 303 + ], + "type": "interline_equation", + "content": "p \\left(x ^ {j}\\right) = \\prod_ {i = 1} ^ {n} p \\left(x _ {i} ^ {j} \\mid x _ {i - 1} ^ {j}, x _ {i - 2} ^ {j}, \\dots , x _ {1} ^ {j}, \\theta\\right) \\tag {1}", + "image_path": "7372c2460de6582045ff340f749c843384300ceaa75303da5c651018a46c369a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 315, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 339 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 315, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 315, + 296, + 339 + ], + "type": "text", + "content": " is the model parameters, which can be optimized by minimizing the negative log-likelihood loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 349, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 349, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 295, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e - t r a i n}} = \\underset {x ^ {j} \\sim X} {\\mathbb {E}} - \\log p \\left(x ^ {j}\\right). \\tag {2}", + "image_path": "395caf1f21850a112af6225eccb716acbd33d5b9ee7c98fabb53ca074698ca8a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 379, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 428 + ], + "type": "text", + "content": "Using this loss, we pre-train our models at different sizes on over 1T of tokens. The tokens come from images and videos. Figure 2 shows the training loss of 3 differently sized models with 120M, 280M and 1.1B parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 138, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 138, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 138, + 449 + ], + "type": "text", + "content": "3.2. Architecture" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 455, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 514 + ], + "type": "text", + "content": "Our model is a transformer [58] with causal attention. We apply recent advancements in language modeling such as pre-norm using RMSNorm [66], SwiGLU activation [52], and RoPE positional embeddings [54], following LLaMa [55]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": "For a model with " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": " layers, we define " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "H^l" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": " as the intermediate representations after layer " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "0 \\leq l \\leq L" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": ". The intermediate representations after layer " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "inline_equation", + "content": "H^{l + 1}" + }, + { + "bbox": [ + 55, + 516, + 296, + 552 + ], + "type": "text", + "content": ", given as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 561, + 295, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 561, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 85, + 561, + 295, + 577 + ], + "type": "interline_equation", + "content": "\\widehat {H} ^ {l + 1} = H ^ {l} + \\operatorname {M H S A} (\\operatorname {R M S - n o r m} (H ^ {l})) \\tag {3}", + "image_path": "8cc3355371293428873cd96d8fb7d7fa1458640fba6b1d4c37a48e49a2378462.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 86, + 578, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 578, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 86, + 578, + 295, + 594 + ], + "type": "interline_equation", + "content": "H ^ {l + 1} = \\widehat {H} ^ {l + 1} + \\operatorname {M L P} (\\operatorname {R M S - n o r m} (\\widehat {H} ^ {l + 1})), \\tag {4}", + "image_path": "c2998f1aac6083462800a3d3bb4d0154315266ad041dfdd3f161e1b7dd3beed2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "Where MHSA is a multi-head self attention layer, MLP is a multi-layer perceptron with SwiGLU activations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "We train our models for the next token prediction task at different scales (base, large and 1b models). For more architecture details see Table 1. We train all these models with a batch size of 1M tokens. We use AdamW [33] with a maximum learning rate of " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "3e - 4" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". We decay the learning rate with a cosine schedule, after 2000 warm-up steps [55]." + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 548, + 249 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 548, + 249 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 548, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 548, + 249 + ], + "type": "image", + "image_path": "d9ef25640e4d5f02b99dcec6243b12b909acafcd81fdbef0ab3c4cb11d0f3375.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 554, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 304 + ], + "type": "text", + "content": "Figure 2. Training Loss Curves: We show the training loss curves for base, large, and 1b models trained with tokens from dVAE [45] with a vocabulary size of 8k and context length of 4k tokens (equivalent to 16 images or video frames)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 324, + 373, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 373, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 373, + 335 + ], + "type": "text", + "content": "3.3. Dataset" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "content": "To train our model, we compile a large dataset from a number of different sources. Table 2 shows the total number of images and videos used for training data, the total number of tokens, as well as the number of hours of videos in each dataset. Together these datasets contain over 100,000 hours of video data and about 2.5 trillion visual tokens. During training, each mini-batch is sampled at different ratios of datasets. Each batch approximately contains " + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "content": " of ImageNet images, " + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "content": " of Ego4D videos, " + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "content": " of Kinetics videos, and " + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 473 + ], + "type": "text", + "content": " of HowTo100m videos. Our full training utilized about 1 trillion tokens." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 480, + 398, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 398, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 398, + 492 + ], + "type": "text", + "content": "3.4. Tokenization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 605 + ], + "type": "text", + "content": "We use dVAE tokenizer with a vocabulary of 8k tokens, from Dall-E [45] as our tokenizer. Using an image-based tokenizer allows training on both images and videos and testing on respective downstream tasks. While VQGAN [15] tokenizers provide sharper images, these models were trained with perceptual loss [28, 30], thus indirectly ingesting ImageNet label information via VGG-net [53]. We also explored SD-VAE based random tokenizer, please see Appendix A.4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": "All raw pixel frames or images are tokenized into 256 discrete tokens. We take a video and resize it such that its shortest size is " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " pixels, and then take a random crop of " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "R \\times R \\times T" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ", and sample every 4 frames where " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " is the number of frames. We use dVAE [45] with the vocabulary of 8k entries to tokenize every frame independently. For dVAE we set " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "R = 128" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": ", to get " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " discrete tokens. Once every frame is mapped into a set of discrete tokens we have " + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "inline_equation", + "content": "T \\times 256" + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": " tokens per each video. We pre-train all the mod-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19110" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 276, + 145 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 276, + 145 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 276, + 145 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 276, + 145 + ], + "type": "table", + "html": "
DatasetsInstancesTokensHours
ImageNet13.9M3.6B-
Kinetics-6000.53M41.3B1496
Ego4D52.1K103B3750
HowTo100m1.172M2560B92627
", + "image_path": "0524008b48b6656c9bbc8c2163353a8c4e440b3210b74f7a1520f8fe75dff1a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 294, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 294, + 217 + ], + "type": "text", + "content": "Table 2. Pre-training Dataset: We use both image datasets (Imagenet [48]) and video datasets (Kinetics600 [8], Ego4D [21], HowTo100m [34]) with different mixing ratios during the pretraining of our models. The whole training data contains about 100,000 hours of videos." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "content": "els with " + }, + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "inline_equation", + "content": "T = 16" + }, + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "content": ", thus all the models are per-trained for a context length of 4096 tokens." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 335 + ], + "type": "text", + "content": "When training with images and videos, 16 video frames are sampled to create 4k tokens. For images, we randomly sample 16 images and create a sequence of 16 image frames to generate 4k tokens. Finally, we add start and end tokens for each sequence, for videos we use [1] as the start token, and for images we use [3] as the start token, and all sequences have an end token of [2]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 180, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 180, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 180, + 354 + ], + "type": "text", + "content": "3.5. Downstream Transfer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 443 + ], + "type": "text", + "content": "The idea of large pre-trained models is that they were trained at a large compute scale, and then these models can be easily used for various downstream tasks without requiring task-specific design or lots of computing for transfer. The learned representations are general enough to transfer to various tasks. We evaluate our models on the intermediate features with linear and attention probing [31]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "text", + "content": "For linear probing the model, we apply global average pooling [32] over the tokens from different layers to get the intermediate representation. We train a linear layer on top of this representation on the downstream task. MAE [25] or DINO [7] have a uniform structure when it comes to which token attends to which tokens, however in autoregressive sequence modeling later tokens attend to more tokens than the tokens at the beginning. Due to this skewed nature, equally weighting all the tokens affects the downstream performance. Attention pooling is an alternative to average pooling that allows to dynamically weight the tokens, ideally giving more weight to tokens that see more tokens. This requires learning " + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "inline_equation", + "content": "W_{k}" + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "text", + "content": " matrices and a query token " + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 445, + 295, + 647 + ], + "type": "text", + "content": ". The query token cross-attends to the intermediate tokens and combines them into a single vector. While this function is not linear anymore, it has been shown to learn better representations in recent works [14]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 135, + 670 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We evaluate our pre-trained models on various downstream tasks such as ImageNet classification, Kinetics action recognition, Ego4D action anticipation, Semi-Supervised" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 395, + 124 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 395, + 124 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 395, + 124 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 395, + 124 + ], + "type": "image", + "image_path": "2f37372431bdc0dcb47c1654075b4ebd2e3d26c6be27b11880d40aee3ff30867.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 133, + 553, + 188 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 188 + ], + "type": "text", + "content": "Figure 3. 1-gram Distribution of Various Tokens: This Figure shows the distribution of 1-gram tokens of various tokenizers (dVAE [45], VQGAN-1k, VQGAN-16k [15]) on Imagenet validation set. Note that, dVAE has almost full convergence of the tokens while VQGAN has less than " + }, + { + "bbox": [ + 313, + 133, + 553, + 188 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 133, + 553, + 188 + ], + "type": "text", + "content": " coverage of the tokens." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 71, + 473, + 124 + ], + "blocks": [ + { + "bbox": [ + 395, + 71, + 473, + 124 + ], + "lines": [ + { + "bbox": [ + 395, + 71, + 473, + 124 + ], + "spans": [ + { + "bbox": [ + 395, + 71, + 473, + 124 + ], + "type": "image", + "image_path": "28d79a5c981899eb1db81caf03e3a87c9a7458d29a95aa10b1a1daab39411605.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 474, + 71, + 552, + 124 + ], + "blocks": [ + { + "bbox": [ + 474, + 71, + 552, + 124 + ], + "lines": [ + { + "bbox": [ + 474, + 71, + 552, + 124 + ], + "spans": [ + { + "bbox": [ + 474, + 71, + 552, + 124 + ], + "type": "image", + "image_path": "48ac95f631af77be480ce08c80a91971f8fa86feecdb0e1c95516a85d6412fa9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "type": "text", + "content": "tracking, and Robotic manipulation tasks. First, we discuss various design choices for pre-training and evaluation strategies for our method. All the models for studying the design choices are large models trained for 400 epochs on the ImageNet-1k dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 277, + 408, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 408, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 408, + 289 + ], + "type": "text", + "content": "4.1. Design Choices" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 295, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 389 + ], + "type": "text", + "content": "Tokenizer: The are various options available for tokenizing an image or a video. We could use discrete tokenizers such as dVAE, and VQGAN, or simple patch-based continuous tokenization. To study the behavior of various tokenizers we pre-train a Toto-large model on ImageNet for 400 epochs. Using linear probing at an optimal intermediate layer, we evaluate the accuracy of the models on ImageNet classification task." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 390, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 557 + ], + "type": "text", + "content": "Table 3 shows linear probing accuracy when trained with various tokenizers. VQGAN [15] and dVAE [45] perform similarly with the same resolutions. However, VQGAN is contaminated with ImageNet label information via the perceptual loss. In addition, as shown in Figure 3, dVAE tokens have full coverage compared to VQGAN tokens on their 1-gram distributions. Please see the supplementary materials for details. Regressing normalized-patch targets from patch embeddings performs slightly worse than classifying discrete tokens as targets. Additionally, discrete tokens as targets and patch embeddings as inputs perform poorly compared to other methods at the given input-output resolutions. Overall, Table 3 shows that various ways of tokenization have little effect on ImageNet linear probing accuracy." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "content": "How to Probe: As discussed in Section 3.5 we probe the pre-trained models at the same layer with attention pooling and average pooling, followed by linear layer. Table 13 shows attention pooling performs " + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "content": " higher than average pooling on the ImageNet classification task. For attention pooling, we keep the embedding dimension the same as the intermediate feature dimensions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Resolution: When training with dVAE tokens, a 256x256 image results in 1024 tokens, this is four times more number of tokens compared to patch embeddings or VQGAN tokens. If we reduce the number of tokens to 256, then the effective image resolution becomes " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": ". Table 4 shows a clear drop in performance when pre-training the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19111" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 288, + 180 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 288, + 180 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 288, + 180 + ], + "type": "table", + "html": "
Input-TargetTokensVocabularyTop1
VQGAN-VQGAN16x1616k61.3
VQGAN-VQGAN16x161k61.1
dVAE-dVAE32x328k61.2
dVAE-dVAE16x168k53.2
patch-patch16x16-60.6
patch-dVAE16x168k58.5
", + "image_path": "031fa5e9cd2f6f5e76343c3bab9e0233632f08248e43f30621b67ad3eaa52de3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 255, + 280, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 194, + 297, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 297, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 297, + 251 + ], + "type": "text", + "content": "Table 3. ImageNet Linear Probing Accuracy with Various Tokenizers: We compare discrete (dVAE, VQGAN) and patch embedding as input and target for pre-training our models. ImageNet top-1 accuracies are computed by linear probing at the 9th layer of the large model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 255, + 280, + 338 + ], + "lines": [ + { + "bbox": [ + 70, + 255, + 280, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 255, + 280, + 338 + ], + "type": "table", + "html": "
MethodCompute (MACs)Top1
dVAE/161.42 × 101753.2
dVAE/325.68 × 101761.2
dVAE/16→322.13 × 101763.2
dVAE/16→32†2.13 × 101764.4
", + "image_path": "991b76c54fc559bad53020cf3b58aa4d6964c3e435f742775074547eea458cfb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 104, + 407, + 246, + 483 + ], + "blocks": [ + { + "bbox": [ + 55, + 350, + 296, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 396 + ], + "type": "text", + "content": "Table 4. Token Resolution: While the performance is lower for a low-resolution model, when finetuned for next-patch prediction at a higher resolution, its performance surpasses the full-resolution pre-trained model. " + }, + { + "bbox": [ + 55, + 350, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 350, + 296, + 396 + ], + "type": "text", + "content": " Base values of the RoPE is 50,000." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 407, + 246, + 483 + ], + "lines": [ + { + "bbox": [ + 104, + 407, + 246, + 483 + ], + "spans": [ + { + "bbox": [ + 104, + 407, + 246, + 483 + ], + "type": "table", + "html": "
MethodTokensPoolingTop1
dVAE16x16Average53.2
dVAE16x16Attention61.1
", + "image_path": "c1f0eb78104f45d6e2ea9d6b209e9c24222d999d0c115ecda1118891a7d04ae6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "content": "model at " + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "content": " resolution. However, due to the use of relative positional embeddings (RoPE [54]), we can easily finetune the " + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "content": " token equivalent) model for higher resolution. Surprisingly, this does better than pre-training at " + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 556, + 295, + 677 + ], + "type": "text", + "content": " resolution and requires only one epoch of finetuning. Not only does this improve the performance, but the pre-training also becomes cheaper compared to full-resolution pre-training. Additionally, Fine-tuning with higher base values of the RoPE embeddings (50,000) leads to better accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Architecture: We train various language models from GPT2 [41] with absolute sine-cosine positional embeddings, and non-transformer based model Mamba [23] only" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 368, + 71, + 503, + 124 + ], + "blocks": [ + { + "bbox": [ + 55, + 501, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 536 + ], + "type": "text", + "content": "Table 5. Attention vs Average Pooling: When probed at the same layers, attention pooling performs much better than average pooling of intermediate tokens." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 368, + 71, + 503, + 124 + ], + "lines": [ + { + "bbox": [ + 368, + 71, + 503, + 124 + ], + "spans": [ + { + "bbox": [ + 368, + 71, + 503, + 124 + ], + "type": "table", + "html": "
ModelParamsTop1
GPT2 [41]280 m48.5
Mamba [23]290 m40.7
LLaMA [55]280 m53.2
", + "image_path": "95f33d64982b167c5f9f9691bee470652055d0e6f9776b7e7db997b515ab67a5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 132, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 132, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 553, + 166 + ], + "type": "text", + "content": "Table 6. Architecture: We compare sequence modeling architectures LLaMA [55], GPT2 [41], and non-transformer models, Mamba [23] on ImageNet linear probing task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 185, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 243 + ], + "type": "text", + "content": "using dVAE tokens. We mimicked the GPT2 architecture and do architecture comparisons. We compare these models with LLaMA [55]. We evaluate linear probing performance at each layer of these models and report the best performance in Table 6." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 245, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 413 + ], + "type": "text", + "content": "Probing Layer: When probing the pre-trained models, especially the decoder-only model best performance is observed at the middle layers. This behavior is first observed in iGPT [9]. Figure 4 shows the peak performance on recognition occurs at about " + }, + { + "bbox": [ + 313, + 245, + 555, + 413 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 413 + ], + "type": "text", + "content": " of the depth of the model. This behavior holds across all model sizes. While in MAE [25] and BEiT [3] encoder-decoder models, due to the uneven nature of the encoder and decoder, the best features are observed at the top of the encoder layers. However, on decoder-only models with uniformly distributed layers, the last layers perform worse on recognition tasks, mainly because these layers are trained to reconstruct the input. More probing results with various tokenizers, resolutions, and probing methods are shown in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 425, + 550, + 583 + ], + "blocks": [ + { + "bbox": [ + 319, + 425, + 550, + 583 + ], + "lines": [ + { + "bbox": [ + 319, + 425, + 550, + 583 + ], + "spans": [ + { + "bbox": [ + 319, + 425, + 550, + 583 + ], + "type": "image", + "image_path": "eee90fbf8aa5fa427afc518ab23b49f934ba29cb5dc72f3d9cf9c2bba989ed5e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 595, + 555, + 630 + ], + "lines": [ + { + "bbox": [ + 313, + 595, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 555, + 630 + ], + "type": "text", + "content": "Figure 4. Probing at Different Layers: We show the attention-probing performance at each layer of our three models. Peak performance is observed at around " + }, + { + "bbox": [ + 313, + 595, + 555, + 630 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 595, + 555, + 630 + ], + "type": "text", + "content": " depth of the models." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 648, + 425, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 425, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 425, + 661 + ], + "type": "text", + "content": "4.2. Image Recognition" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "To measure the representation quality of our pre-trained models, we evaluate our models on ImageNet-1k [12] classification. We apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19112" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 269, + 313 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 269, + 313 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 269, + 313 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 269, + 313 + ], + "type": "table", + "html": "
MethodArch#θTop1
Discriminative Approaches
SimCLR [10]†RN50x29474.2
BYOL [22]†RN50x29477.4
SwAV [6]†RN50x29473.5
DINO [7]ViT-B/88680.1
DINOv2 [36]ViT-g/14101186.4
Generative Approaches
BEiT-L [3]ViT-L/1430762.2
AIM [14]ViT-1B/14120080.6
MAE [25]ViT-H/1463280.9
iGPT-L [9]†GPT-2138665.2
iGPT-XL [9]†GPT-2680172.0
Toto-baseLLaMA12064.7
Toto-largeLLaMA28071.1
Toto-1bLLaMA110075.3
", + "image_path": "25c513f6692fec35f3d9e67c7d70c134aa0ab27d1b22376e4b9d2b225181bc85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "text", + "content": "the highest classification accuracy. We fine-tune the pretrained models further by applying self-supervised next token prediction loss in Eq 2, together with cross-entropy loss applied for probing layers (with stop-gradients). We train the probing layers for 90 epochs, with a learning rate of " + }, + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "inline_equation", + "content": "6e^{-5}" + }, + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "text", + "content": ". We also use layer decay of 0.9 to reduce the learning rate at the early layers of the model. During this stage, all the models are fine tuned with " + }, + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 401, + 295, + 532 + ], + "type": "text", + "content": " token resolution, on the self-supervised loss, and increase the base value of the RoPE [54] embeddings from 10,000 to 50,000 support larger resolution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Table 7 shows the ImageNet top-1 accuracy of our base, large and 1b models. First, there is a clear difference in terms of classification performance when it comes to discriminative models versus generative models. Instance discriminative models such as SimCLR [10], and DINO [7] are trained to separate samples from each other and they are designed to perform well on discriminative tasks. On the other hand, generative models are just trying to model the data distribution. While achieving comparable performance to other generative models on image recognition, among autoregressive generative models, our model achieved the highest top-1 accuracy. The scaling of data, and the use of tokens instead of pixels, allows our one billion parameter model to achieve similar performance compared to iGPT [9] 7 billion models." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 349, + 70, + 521, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "lines": [ + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "type": "text", + "content": "Table 7. ImageNet Results: We compare discriminative and generative models on ImageNet [12] recognition task. While achieving comparable performance among generative models, our models model achieves the highest accuracy on autoregressive modeling. " + }, + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "type": "text", + "content": " models are evaluated with linear probing. " + }, + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "type": "inline_equation", + "content": "\\# \\theta" + }, + { + "bbox": [ + 55, + 322, + 294, + 377 + ], + "type": "text", + "content": " in millions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 349, + 70, + 521, + 266 + ], + "lines": [ + { + "bbox": [ + 349, + 70, + 521, + 266 + ], + "spans": [ + { + "bbox": [ + 349, + 70, + 521, + 266 + ], + "type": "table", + "html": "
MethodArchTop1
Discriminative Approaches
I-JEPA [1]ViT-H/1674.5
OpenCLIP [11]ViT-G/1483.3
DINOv2 [36]ViT-g/1484.4
InternVideo [61]-73.7
Generative Approaches
Hiera [49]Hiera-H/1477.0
MVD [60]ViT-H/1479.4
VideoMAE [59]ViT-L/1479.8
Toto-baseLLaMA59.3
Toto-largeLLaMA65.3
Toto-1bLLaMA74.4
", + "image_path": "4dcf443ed9e5c2c5c695f2cc16ebd56844a3fa13a2bbd02bb72096f933445906.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 274, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 340 + ], + "type": "text", + "content": "Table 8. K400 Results: We compare discriminative and generative models on Kinetics-400 [29] action recognition task. While achieving comparable performance among generative models, our models are the first to show the competitive performance on K400 with autoregressive pre-training, and shows scaling with large model sizes." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 353, + 427, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 427, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 427, + 365 + ], + "type": "text", + "content": "4.3. Action Recognition" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "text", + "content": "We use Kinetics-400 (K400) [29] for evaluating our models on action recognition tasks. Similar to ImageNet evaluation, we apply a probe at each layer of the model, with attention pooling, and choose the optimal layer with the highest action classification accuracy. We also fine-tune the pre-trained models on a self-supervised next-patch prediction task while training the probing layers with a classification loss. All our video models are trained with 16 frames, thus with a context length of 4096 tokens per video. When evaluating videos, we follow the protocol in SlowFast [18]. Unlike ImageNet where we evaluate the models at " + }, + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "text", + "content": " resolution, on videos we only evaluate our models at " + }, + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 312, + 371, + 555, + 538 + ], + "type": "text", + "content": " resolution, to keep the number of tokens in a similar budget." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 540, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 540, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 312, + 540, + 556, + 647 + ], + "type": "text", + "content": "Table 8 shows the Kinetics-400 top-1 accuracy of our base, large and 1b models. Similar to ImageNet results in Table 7, we see that discriminately trained models perform better than generative models. Our models achieve comparable performance among generative models, and first to show competitive performance on action recognition with autoregressive generative modeling. All the models are trained and evaluated with 16 frames with a stride of 4 frames." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 425, + 671 + ], + "type": "text", + "content": "4.4. Action Forecasting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "While the Kinetics dataset captures internet-style exocentric videos, Ego4D [21] videos capture egocentric videos of daily life. A general vision model should be able to rea" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19113" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 185 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 185 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 185 + ], + "type": "table", + "html": "
MethodNounN+VN+TTCOverall
FRCNN+Rnd [21]17.551.563.210.34
FRCNN+SF [21]17.555.195.372.07
Hiera-large [49]14.056.034.532.12
StillFast [44]16.207.474.942.48
VideoMAE-large [59]15.166.725.262.55
MAE-ST-large [19]13.716.634.942.60
Toto-large15.206.755.412.70
", + "image_path": "47b1cb6a3c29fa56c8d50fbeaf1fda670d84f5e570cc90e7b16b505f64ae7f25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 296, + 474 + ], + "type": "text", + "content": "son about both exo and ego-centric videos. Task-wise, Kinetics requires the model to reason about the action using full context (e.g., the model has seen the action), while the Ego4D short-term action anticipation v1 task requires models to predict future actions from past context. We use our models as the backbone for the pyramid network used in StillFast [44] to extract tokens at 5 layers and fuse them with the pyramid network. We fully fine-tuned our model with self-supervised next-patch loss along with task-related losses, and we observed that having the self-supervised loss improves the overall performance. Table 9 shows the performance of our large model on the Ego4D short-term action anticipation task. This task requires predicting the object to be interacted with (noun) and the type of interaction (verb) as well as time to contact (ttc) from the last seen frame to an estimated time between object-hand contact. As shown in Table 9, these tasks are difficult with maximum overall mean-average precision of 2.70." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 151, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 151, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 151, + 493 + ], + "type": "text", + "content": "4.5. Video Tracking" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 605 + ], + "type": "text", + "content": "We evaluate our pre-trained models on the label propagation task on the DAVIS dataset [39], following the evaluation protocol in [27]. In contrast to the previous tasks, such as classification, this evaluation does not require fine-tuning or probing. We first use the features from the last " + }, + { + "bbox": [ + 55, + 498, + 296, + 605 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 498, + 296, + 605 + ], + "type": "text", + "content": " frames to find the nearest neighbor patch in the current frame. We then propagate the masks from the previous frames to the current frame. We compare with Dino [7] and MAE [25] in Table 10 and show qualitative results in Figure 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 613, + 119, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 119, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 119, + 624 + ], + "type": "text", + "content": "4.6. Robotics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In this section, we study the effectiveness of our pre-trained representations for robotic manipulation. We consider tasks in simulation and in the real world. Due to the runtime constraints on real hardware, we use the Toto-base models. Despite being a small model, Toto-base can achieve better performance in simulation and on-par performance to state-of-the-art robot models in real-world experiments." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 199, + 295, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 295, + 244 + ], + "type": "text", + "content": "Table 9. Ego4D Results: Our model achieves comparable mean-average precision compared to previous work. We compare our method with, FRCNN+Rnd [21], FRCNN+SF [21], Hiera [49], StillFast [44], VideoMAE [59], and MAE-ST [19]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 190 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 190 + ], + "type": "table", + "html": "
Method (Res/Patch)J&FJF
DINO-base (224/8)54.352.556.1
DINO-base (224/16)33.136.230.1
MAE-base (224/16)31.534.128.9
Toto-base (256/8)42.041.243.1
Toto-large (256/8)44.844.445.1
Toto-1b (256/8)46.145.846.4
Toto-large (512/8)62.459.265.6
", + "image_path": "b70bf8ea58a31f736548e2f0dc3cf682e1a19040de0750e9f316be4b47e423c6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 243 + ], + "type": "text", + "content": "Table 10. DAVIS Tracking: We report J, F, and J&F scores at the peak layers of each model. We achieves comparable performance as DINO and at large resolution (512), it outperforms all methods. Our model achieve best performance at the middle layer, see 9." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 257, + 432, + 342 + ], + "blocks": [ + { + "bbox": [ + 316, + 257, + 432, + 342 + ], + "lines": [ + { + "bbox": [ + 316, + 257, + 432, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 432, + 342 + ], + "type": "image", + "image_path": "119e676c787020145b1a2588468ddb1c7e50f1bd2fb29696c4c21df68f147483.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 345, + 400, + 354 + ], + "lines": [ + { + "bbox": [ + 349, + 345, + 400, + 354 + ], + "spans": [ + { + "bbox": [ + 349, + 345, + 400, + 354 + ], + "type": "text", + "content": "(a) Franka Pick" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 256, + 549, + 342 + ], + "blocks": [ + { + "bbox": [ + 434, + 256, + 549, + 342 + ], + "lines": [ + { + "bbox": [ + 434, + 256, + 549, + 342 + ], + "spans": [ + { + "bbox": [ + 434, + 256, + 549, + 342 + ], + "type": "image", + "image_path": "65de3d672e8e31f7b287f4b6a0d2e5a688742ea3503ef1823dedcd3b0d0d818e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 345, + 516, + 354 + ], + "lines": [ + { + "bbox": [ + 469, + 345, + 516, + 354 + ], + "spans": [ + { + "bbox": [ + 469, + 345, + 516, + 354 + ], + "type": "text", + "content": "(b) Kuka Pick" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 364, + 555, + 409 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 409 + ], + "type": "text", + "content": "Figure 5. Robot Manipulation with Reinforcement Learning: We compare MAE-base [42] with Toto-base pre-trained models in simulation following [64]. We find that the Toto outperforms the baseline in sample complexity, across two robots and two tasks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 425, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 581 + ], + "type": "text", + "content": "Simulation Experiments: Following the protocols in MVP [64], we use our visual pre-trained models to embed pixel observations. The model is frozen and we only take tokens at an intermediate layer, apply average pooling, and learn the linear layer on top to embed pixel observations. These observations are used to train DAgger policies for 2 tasks: Franka-pick 5a, Kuka-pick 5b. Figure 5 shows the mean success rate over training steps. Compared to the MVP baseline, our model was able to learn these tasks faster with better sample efficiency across robots and tasks. For fair comparisons, we use the best MAE model from MVP [42] which is trained on ImageNet [12], Ego4D [21] and 100DOH [50] datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Real-world Experiments: Next, we evaluate our pre-trained models on hardware. We follow the setup from [42]. We extract vision features using a pre-trained vision encoder and train a controller on top of frozen representations using behavior cloning. Specifically, we consider a cube picking tasks using a 7 DoF Franka robot, shown in Figure 8. We use the demonstrations provided by [43]. In Table 11 we compare our model to a vision encoder from [42]. We report the success rate over 16 trials with variations in object position and orientation. Our model performs favorably to a vision encoder pre-trained for robotics." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19114" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 101, + 70, + 251, + 119 + ], + "blocks": [ + { + "bbox": [ + 101, + 70, + 251, + 119 + ], + "lines": [ + { + "bbox": [ + 101, + 70, + 251, + 119 + ], + "spans": [ + { + "bbox": [ + 101, + 70, + 251, + 119 + ], + "type": "table", + "html": "
Model# TrajSuccess
MVP24075%
Toto-base24063%
", + "image_path": "2d5c4c1901b7a1b7d077858526cf583d9764dcfa7d02ae6ec642a3a136bedf8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 209, + 279, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 128, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 205 + ], + "type": "text", + "content": "Table 11. Robotics, Real-world Experiments: We compare MVP [42] and Toto on a Franka cube-picking task in the real world. Features from both models are pre-trained, frozen, and passed into a learning module trained with behavior cloning using the same demonstrations. We see that our approach performs comparably to the state-of-the-art vision backbone for robotics, despite not being designed with the robotic application in mind." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 209, + 279, + 281 + ], + "lines": [ + { + "bbox": [ + 73, + 209, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 73, + 209, + 279, + 281 + ], + "type": "table", + "html": "
MethodModel16 frame32 frame
V3DResNet55.269.7
TFC V3DResNet54.670.2
Toto-largeLLaMa62.872.9
", + "image_path": "2cf6667ee06f7f382c9c3d2fd3a6a1625a0d2d5d3100a16f9bed1241c7b8e441.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 334 + ], + "type": "text", + "content": "Table 12. Object Permanence: In the CATER [20] object localization task, the object is hidden under or obstructed by other objects and the goal is to predict its coarse location. Our model outperforms previous methods at 16 and 32 temporal resolutions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 348, + 170, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 170, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 170, + 361 + ], + "type": "text", + "content": "4.7. Object Permanence" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "content": "To quantitatively measure the performance of how well the model understands object permanence, we evaluate our models on the CATER localization task [20]. Here, a ball is moving in the scene, and the task is to find its location in the 6 by 6 grid. We fine-tune our Toto-large model on this task at temporal resolutions of 16 and 32 frames. In both cases, our pre-trained models outperform the models trained specifically for this task, such as V3D [67], TFC-V3D [67]. Table 12 shows the performance on the CATER snitch localization task, and Toto-large achieve " + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "inline_equation", + "content": "62.8\\%" + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "inline_equation", + "content": "70.9\\%" + }, + { + "bbox": [ + 55, + 365, + 296, + 498 + ], + "type": "text", + "content": " performance with 16 and 32 frames respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 201, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 201, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 201, + 517 + ], + "type": "text", + "content": "4.8. Compute Optimal Scaling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "We study the scaling properties of Toto using " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "-Parameterization [65]. First we train various models, a1-a6, with linearly increasing hidden size and number of layers (Table 17). All models use the VQGAN tokenizer [15]. We then optimize the learning rate for these models, with " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "-Parameterization [65]. Figure 11 shows the optimal learning rate of " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2^{-7}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " for all of the model widths. Once we find the optimal learning rate, we train a1-a6 models on our data mixture (Table 2). Figure 6 shows the loss vs. training compute of Toto models. We see a clear power law relationship between the compute and validation loss. Based on these experiments Toto shows a power law of " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "L(C) = 7.32 \\cdot C^{-0.0378}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ". For comparison, the GPT-3 power law relationship [4] is " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "L(C) = 2.57 \\cdot C^{-0.048}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ". While these are not comparable directly, the scaling coefficients indicate how much change in the loss to expect for additional com" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 274 + ], + "type": "image", + "image_path": "c1f099cfe78788d43d19228d218d4bae1dcc9c06f02bc263f2ca86c9ab55ff8a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 284, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 329 + ], + "type": "text", + "content": "Figure 6. Scaling Toto: We train multiple variants of Toto, with increasing hidden size and depth, with optimal learning rates. We plot the validation loss vs the compute spent on training in MACs. This shows a clear scaling behavior with optimal compute." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 345, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 524 + ], + "type": "text", + "content": "pute. This suggests that autoregressive video models, such as Toto, scale but at a slower rate than language models. In language processing, tokenization is typically lossless and highly compressed, refined over thousands of years. In contrast, video inherently contains significant redundancy across subsequent frames—if a video lacks this redundancy, it wouldn’t truly qualify as a video. This redundancy, however, is essential for effective self-supervised training. With better tokenizers that yield more compressed and semantically rich representations, we can build more scalable video encoders and use scaling laws as benchmarks. Moreover, scaling laws could help us curate more scalable data and inform the design of model architectures, rather than relying solely on single benchmark metrics. See the limitations section in the Appendix for further discussion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 537, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "We empirically study autoregressive pre-training from videos. We curate a large video dataset and conduct a largescale evaluation across a range of diverse tasks, including image recognition, video classification, video forecasting, object tracking, object permanence, and robotic manipulation. We perform extensive ablation studies to understand different design choices and compare autoregressive video pre-training to strong baselines across different tasks. We find that, despite minimal inductive biases, our approach achieves competitive performance across all tasks. Finally, we study the scaling behavior of autoregressive video models, and show that they scale with compute, but at a slower rate than autoregressive language models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19115" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "We thank Andrea Madotto, Po-Yao (Bernie) Huang, and Shiry Ginosar for helpful discussions. We're grateful to Ronghang Hu and Xinlei Chen for their help with TPU setup and code bases. We also thank Baifeng Shi for helping us with robots evaluations. We thank Valentin Gabeur and Neerja Thakkar for their valuable feedback on the paper. This work was supported by the FAIR-BAIR program as well as ONR MURI (N00014-21-1-2801)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 115, + 210 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 217, + 296, + 710 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 217, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 283 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 285, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 295, + 306 + ], + "type": "text", + "content": "[2] Fred Attneave. Some informational aspects of visual perception. Psychological review, 1954. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 340 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 2, 5, 6, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 363 + ], + "type": "text", + "content": "[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 407 + ], + "type": "text", + "content": "[5] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 463 + ], + "type": "text", + "content": "[6] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 520 + ], + "type": "text", + "content": "[7] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 4, 6, 7, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 521, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 295, + 553 + ], + "type": "text", + "content": "[8] Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "text", + "content": "[9] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 2, 5, 6, 12, 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 600, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 643 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 644, + 295, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 644, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 57, + 644, + 295, + 710 + ], + "type": "text", + "content": "[11] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 709 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 117 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 150 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 152, + 554, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 554, + 206 + ], + "type": "text", + "content": "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 2, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[15] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12868–12878, 2020. 3, 4, 8, 14" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 253, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 554, + 317 + ], + "type": "text", + "content": "[16] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 12" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[17] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv preprint arXiv:2309.17425, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 397 + ], + "type": "text", + "content": "[18] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 441 + ], + "type": "text", + "content": "[19] Christoph Feichtenhofer, Yanghao Li, Kaiming He, et al. Masked autoencoders as spatiotemporal learners. Advances in neural information processing systems, 35:35946-35958, 2022. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 475 + ], + "type": "text", + "content": "[20] Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744, 2019. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 477, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 553, + 552 + ], + "type": "text", + "content": "[21] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 554, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 554, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 317, + 554, + 554, + 620 + ], + "type": "text", + "content": "[22] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "text", + "content": "[23] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "spans": [ + { + "bbox": [ + 316, + 655, + 553, + 709 + ], + "type": "text", + "content": "[24] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19116" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 711 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[25] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 5, 6, 7, 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[26] Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling. arXiv preprint arXiv:2010.14701, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 226 + ], + "type": "text", + "content": "[27] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 7, 13" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "text", + "content": "[28] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 335 + ], + "type": "text", + "content": "[29] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 6, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 336, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 296, + 392 + ], + "type": "text", + "content": "[30] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In Proceedings of The 33rd International Conference on Machine Learning, pages 1558-1566, New York, New York, USA, 2016. PMLR. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 392, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 296, + 446 + ], + "type": "text", + "content": "[31] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In Proceedings of the 36th International Conference on Machine Learning, pages 3744-3753. PMLR, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 447, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 296, + 468 + ], + "type": "text", + "content": "[32] Min Lin, Qiang Chen, and Shuicheng Yan. Network in network. arXiv preprint arXiv:1312.4400, 2013. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 491 + ], + "type": "text", + "content": "[33] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 555 + ], + "type": "text", + "content": "[34] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 610 + ], + "type": "text", + "content": "[35] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 611, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 296, + 666 + ], + "type": "text", + "content": "[36] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 666, + 296, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 666, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 666, + 296, + 711 + ], + "type": "text", + "content": "[37] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International conference on machine learning, pages 4055-4064. PMLR, 2018. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 700 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 95 + ], + "type": "text", + "content": "[38] Jean Piaget. The construction of reality in the child. Routledge, 1954. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[39] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 174 + ], + "type": "text", + "content": "[40] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 207 + ], + "type": "text", + "content": "[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. 2019. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "text", + "content": "[42] Ilija Radosavovic, Tete Xiao, Stephen James, Pieter Abbeel, Jitendra Malik, and Trevor Darrell. Real-world robot learning with masked visual pre-training. In Conference on Robot Learning, 2022. 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[43] Ilija Radosavovic, Baifeng Shi, Letian Fu, Ken Goldberg, Trevor Darrell, and Jitendra Malik. Robot learning with sensorimotor pre-training. In Conference on Robot Learning, 2023. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 354 + ], + "type": "text", + "content": "[44] Francesco Ragusa, Giovanni Maria Farinella, and Antonino Furnari. Stillfast: An end-to-end approach for short-term object interaction anticipation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3635-3644, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 408 + ], + "type": "text", + "content": "[45] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3, 4, 14" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[46] MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. arXiv preprint arXiv:1412.6604, 2014. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 510 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 565 + ], + "type": "text", + "content": "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 566, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 554, + 621 + ], + "type": "text", + "content": "[49] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. arXiv preprint arXiv:2306.00989, 2023. 6, 7, 14" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 622, + 554, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 554, + 676 + ], + "type": "text", + "content": "[50] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 677, + 554, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 554, + 700 + ], + "type": "text", + "content": "[51] Claude E Shannon. Prediction and entropy of printed english. Bell system technical journal, 1951. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19117" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 452 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 95 + ], + "type": "text", + "content": "[52] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 129 + ], + "type": "text", + "content": "[53] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 129, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 129, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 129, + 295, + 173 + ], + "type": "text", + "content": "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 228 + ], + "type": "text", + "content": "[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 14" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 230, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 230, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 295, + 273 + ], + "type": "text", + "content": "[56] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 57, + 274, + 295, + 317 + ], + "type": "text", + "content": "[57] Aäron Van Den Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 2, 12" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 319, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 319, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 295, + 361 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 363, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 363, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 57, + 363, + 295, + 427 + ], + "type": "text", + "content": "[59] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yi-nan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 2, 6, 7, 14" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 429, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 429, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 57, + 429, + 295, + 452 + ], + "type": "text", + "content": "[60] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Lu Yuan, and Yu-Gang Jiang." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 453 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Masked video distillation: Rethinking masked feature modeling for self-supervised video representation learning. In CVPR, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[61] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Intervideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[62] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. arXiv preprint arXiv:1906.02634, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 250 + ], + "type": "text", + "content": "[63] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3733-3742, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[64] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[65] Greg Yang, Edward J. Hu, Igor Babuschkin, Szymon Sidor, Xiaodong Liu, David Farhi, Nick Ryder, Jakub Pachocki, Weizhu Chen, and Jianfeng Gao. Tensor programs v: Tuning large neural networks via zero-shot hyperparameter transfer, 2022. 8, 14" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 373 + ], + "type": "text", + "content": "[66] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 407 + ], + "type": "text", + "content": "[67] Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning. arXiv preprint arXiv:2203.05928, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[68] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19118" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_content_list.json b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..44761b549799dd9e86c816df818a6fcbeeb5aa06 --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "An Information-Theoretic Regularizer for Lossy Neural Image Compression", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingwen Zhang $^{1}$ , Meng Wang $^{2*}$ , Xihua Sheng $^{1}$ , Peilin Chen $^{1}$ , Junru Li $^{3}$ , Li Zhang $^{3}$ , Shiqi Wang $^{1}$", + "bbox": [ + 256, + 179, + 740, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, City University of Hong Kong, Hong Kong, China $^{2}$ School of Data Science, Lingnan University, Hong Kong, China $^{3}$ Bytedance Inc., San Diego, USA", + "bbox": [ + 158, + 215, + 836, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lossy image compression networks aim to minimize the latent entropy of images while adhering to specific distortion constraints. However, optimizing the neural network can be challenging due to its nature of learning quantized latent representations. In this paper, our key finding is that minimizing the latent entropy is, to some extent, equivalent to maximizing the conditional source entropy, an insight that is deeply rooted in information-theoretic equalities. Building on this insight, we propose a novel structural regularization method for the neural image compression task by incorporating the negative conditional source entropy into the training objective, such that both the optimization efficacy and the model's generalization ability can be promoted. The proposed information-theoretic regularizer is interpretable, plug-and-play, and imposes no inference overheads. Extensive experiments demonstrate its superiority in regularizing the models and further squeezing bits from the latent representation across various compression structures and unseen domains. Code is available at Github.", + "bbox": [ + 88, + 335, + 485, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 648, + 220, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lossy image compression, or more broadly, lossy data compression, is grounded in a fundamental theorem in information theory: the rate-distortion theorem [1]. Essentially, this theorem establishes a theoretical lower bound for lossy compression and provides valuable guidance for designing practical compression algorithms [2, 3].", + "bbox": [ + 89, + 674, + 482, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A basic lossy image compression model is illustrated in Fig.1(a). Let $\\mathbf{X}$ , $\\hat{\\mathbf{X}}$ , and $\\mathbf{U}$ denote the source, the reconstruction, and the discrete index (or the latent), respectively. The encoder is a deterministic mapping $\\mathrm{Q} \\colon \\mathcal{X}^N \\to \\mathcal{U}$ where $\\mathcal{X}^N$ denotes a $N$ -dimensional image space and $\\mathcal{U} = \\{1, 2, \\dots, M\\}$ denotes the discrete index space. The decoder", + "bbox": [ + 89, + 763, + 482, + 857 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2653bdbcd142fbec9a2e1c5551274f1d4e6a38adade3615fe13c7a42b36c7e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 303, + 666, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ea905594142a6088fcd23c8c6a4c9f48b1d979f342a9aaaf4f34258c68f18706.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 301, + 893, + 395 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db1c111c8e4e030aba4d06968a12c94ccd1998e5d7520a827623981b5458bfa4.jpg", + "image_caption": [ + "(c)", + "Figure 1. (a) Direct coding model; (b) Information diagram for the direct coding, wherein $H(\\mathbf{X}) = H(\\mathbf{X}|\\hat{\\mathbf{X}}) + H(\\mathbf{U})$ and $H(\\mathbf{X})$ is fixed for any known source; (c) Transform coding model (with possible side information $\\hat{\\mathbf{Z}}$ [4])." + ], + "image_footnote": [], + "bbox": [ + 586, + 411, + 828, + 502 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is a bijection mapping $\\mathrm{Q}^{-1} \\colon \\mathcal{U} \\to \\hat{\\mathcal{X}}^N$ , where the reconstruction space $\\hat{\\mathcal{X}}^N = \\{c_1, \\dots, c_M\\}$ forms a $N$ -dimensional codebook. For index $U$ , entropy coding is applied to losslessly convert the data into the bitstream, which can then be decoded back to $U$ on the decoder side. We may refer to such a model as the direct coding model for its structural simplicity. Subsequently, following the rate-distortion theorem [1], given a bounded distortion measure $d(\\cdot)$ and a distortion constraint $D$ , the rate of transmitting $U$ is lower bounded by the informational rate-distortion function:", + "bbox": [ + 511, + 590, + 906, + 743 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\nR (D) = \\min _ {p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}: \\mathbb {E} _ {p _ {\\boldsymbol {X}} p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}} [ d (\\boldsymbol {X}, \\hat {\\boldsymbol {X}}) ] < D} \\frac {1}{N} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 750, + 906, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "i.e., the minimum mutual information obtained by traversing all the possible encoder-decoder (or quantizer) designs $p_{\\hat{X} |X}$ . Theoretically, this bound can be achieved asymptotically using vector quantization as $N \\to \\infty$ [5]. However, since the encoding complexity grows exponentially with the image dimension $N$ , a suboptimal yet more practical solution is the transform coding [6]. As depicted in", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (mengwang7@ln.edu.hk). This work was supported in part by ITF Project GHP/044/21SZ, in part by the Start-up Grant SUG-007/2425, and in part by Faculty Research Grant SDS24A6.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15573", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fig.1(c), for the transform coding model, a deterministic analysis transform $\\mathrm{T_A}$ : $\\mathcal{X}^N \\to \\mathcal{Y}^M$ and a deterministic synthesis transform $\\mathrm{T_S}$ : $\\hat{\\mathcal{Y}}^M \\to \\hat{\\mathcal{X}}^N$ are incorporated into the compression chain. The analysis transform is designed to decorrelate the source $X$ , such that a much simpler quantization scheme, e.g., uniform scalar quantization: $\\boldsymbol{u}_i = \\mathrm{round}(\\boldsymbol{y}_i)$ , can be independently applied to each dimension of the transformed coefficients $Y$ . Building on this framework and leveraging the superior nonlinear capabilities of neural networks, numerous neural image compression networks [4, 7-13] are proposed, continually pushing the frontier of lossy image compression.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typically, by parameterizing the $\\mathrm{T_A}$ and $\\mathrm{T_S}$ in Fig.1(c) as neural networks, the rate-distortion training objective is formulated as [7]:", + "bbox": [ + 89, + 272, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {R \\approx H (U)} \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} [ - \\log q _ {\\phi} (\\boldsymbol {U}) ]} + \\lambda \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} | | \\boldsymbol {X} - \\hat {\\boldsymbol {X}} | | _ {2} ^ {2}} _ {D}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 327, + 482, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $R$ is the estimation of latent entropy $H(\\pmb{U})$ ; $q_{\\phi}$ is the neural entropy model (or estimated distribution) for $\\pmb{U}$ , whose modeling accuracy can be significantly enhanced by transmitting a small amount of side information $\\hat{\\pmb{Z}}$ (as depicted in Fig.1(c)) as part of the parameters $\\phi$ [4]; the distortion $D$ is the $l_{2}$ distance between $\\pmb{X}$ and $\\hat{\\pmb{X}}$ ; $\\lambda$ is the Lagrangian multiplier, controlling the trade-off between $R$ and $D$ . Herein, due to the existence of quantization $\\mathrm{Q}(\\cdot)$ , $\\pmb{U}$ is discrete in essence. Optimizing the network is then challenging as the gradient is almost zero everywhere for $\\mathrm{Q}(\\cdot)$ . To effectively train the networks, a surrogate is introduced during training, e.g., additive uniform noise (AUN) [7]:", + "bbox": [ + 89, + 377, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {U} = \\mathrm {Q} (\\boldsymbol {Y}) \\approx \\boldsymbol {Y} + [ - 0. 5, 0. 5 ] ^ {M}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 585, + 482, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such that the backward gradient of $\\mathrm{Q}(\\cdot)$ is one. However, since such approximation inevitably results in the train-test mismatch, and thus high gradient estimation error [14], both the efficiency and efficacy of the optimization are hindered. To address this, many research efforts are devoted to designing advanced surrogates and optimization methods [14-18].", + "bbox": [ + 89, + 613, + 482, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a completely distinct view on facilitating the optimization of the compression network: regularization. Conceivably, imposing structural constraints on the network optimization can effectively regularize the gradient estimation error, promote convergence, and reduce overfitting. In the context of generative variational autoencoders [19], structural regularization techniques [20-25] have been effectively applied to improve authentic data generation. However, since neural compression networks fundamentally differ from generative autoencoders in terms of the minimum latent entropy constraint, to the best of our knowledge, no structural regularization methods tailored to this constraint have been explored. In light of this, in this", + "bbox": [ + 89, + 705, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paper, we present the first study of information-theoretic regularization for the neural image compression task. In particular, through an in-depth information-theoretic analysis of compression models, we reveal that minimizing the latent entropy $H(\\mathbf{U})$ is, to some extent, equally important as maximizing the conditional source entropy $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ . Subsequently, by enforcing the network to simultaneously minimize $-H(\\mathbf{X}|\\hat{\\mathbf{X}})$ , both the in-domain compact representation and out-of-domain generalization ability are experimentally improved.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 258, + 660, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural image compression networks. Neural image compression networks are primarily composed of transform, quantization, and latent entropy models. The transform module maps the input source into a latent representation so that the dependencies of the source domain can be more easily handled. Unlike the linear transform [26] used in traditional hybrid coding, the neural transform is inherently nonlinear, enabling more effective decorrelation of the source data. In the literature, various neural transform architectures are proposed. Representative architectures include the recurrent neural network (RNN) [27], convolutional neural network (CNN) [7], non-local attention [28], transformer [29], mixed transformer-CNN [30], state-space models [13] and invertible networks [31]. Subsequently, the continuous transformed coefficients are quantized into the discrete representation, wherein the uniform scalar quantization [7] is commonly adopted. As vector quantization can, in principle, better approach the rate-distortion bound (Eqn.(1)), low-complexity vector quantization [15, 32, 33] and dependent scalar quantization [34, 35] schemes are also studied. Finally, once the quantized latent representations are obtained, a neural entropy model is essential for performing entropy coding. Herein, since the latent cannot be perfectly decorrelated, an accurate entropy model that can better capture the dependencies and closely match the ground-truth latent distribution is crucial for the overall network performance. Prominent research directions such as spatial-channel context mining [8, 9, 12, 36], expressive latent distribution modeling [10, 37, 38] and latent architecture designs [4, 11, 39] are explored.", + "bbox": [ + 511, + 286, + 906, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimization over quantized latent representations. The surrogate-based methods [7, 14-17] have been shown to effectively address the non-differentiable quantization function during end-to-end optimization. In addition to AUN [7], various quantization surrogates such as soft assignment [15], Gumbel-softmax trick [16], and soft rounding function [17] have been studied. However, as the surrogate introduces a train-test mismatch in forward calculations, the annealing techniques [14-17] are then proposed to", + "bbox": [ + 511, + 763, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15574", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "control the \"softness\" of surrogates, gradually approximating the hard quantization function as training progresses. The straight-through estimator (STE) [40] provides an alternative approach to address the train-test mismatch by enforcing an identity mapping during the backward calculation, thereby enabling direct hard quantization in the forward pass. Nevertheless, STE ultimately introduces a significant gradient estimation bias for the analysis transform. To counter this, an additional fine-tuning stage for the entropy model and synthesis transform is suggested [18].", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While numerous network optimization methods have been developed from a quantization perspective, the regularization view has largely escaped research attention. To the best of our knowledge, the most related efforts are regularization methods for variational autoencoders (VAEs), such as regularization on latent priors [20-22], reconstruction [23, 24], and consistency [25]. However, as VAEs are designed for generative modeling, these methods are not applicable to optimizing compression networks. To bridge this knowledge gap, this paper proposes a structural regularizer for neural image compression and conducts extensive experiments to demonstrate its superiority.", + "bbox": [ + 89, + 242, + 483, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 439, + 225, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we begin by deriving the equivalent maximization form of the latent entropy minimization problem for both the direct and transform coding models. Building on this derivation, we propose a novel regularization method for optimizing the neural compression network. Note that during our derivations, we use the naive transform coding model without the side information branch (i.e., the $\\hat{Z}$ depicted in Fig 1(c)). This significantly simplifies our analysis without compromising much practicality, as it has been shown that the information flowing into the side branch is negligible compared to the latent branch [4].", + "bbox": [ + 89, + 465, + 483, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation: from minimization to maximization", + "text_level": 1, + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lemma 1. Considering a deterministic quantization process $Q(\\cdot)$ and a deterministic dequantization process $Q^{-1}(\\cdot)$ for the direct coding model illustrated in Fig. 1(a), the following equalities hold automatically:", + "bbox": [ + 89, + 681, + 482, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) & (4) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) & (5) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) & (6) \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 753, + 482, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. Eqn.(4) follows the facts that $H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0$ , meaning there is no uncertainty in $\\hat{\\mathbf{X}}$ given the observation of $\\mathbf{X}$ , and $I(\\mathbf{X};\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}) - H(\\hat{\\mathbf{X}}|\\mathbf{X})$ . Similarly, Eqn.(5) follows the fact that $H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0$ and $H(\\hat{\\mathbf{X}}) + H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\mathbf{U}) + H(\\hat{\\mathbf{X}}|\\mathbf{U})$ . Eqn.(6) is imme", + "bbox": [ + 89, + 823, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dately apparent from Eqn.(4) and Eqn.(5). Additional details are provided in the Supplementary Material (Sec. 6).", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. For the direct coding model elaborated in Lemma 1, we have", + "bbox": [ + 511, + 140, + 903, + 170 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH (\\boldsymbol {U}) = H (\\boldsymbol {X}) - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 183, + 903, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The informational diagram is thus illustrated as Fig.1(b), given the non-negativity of discrete entropy $H(\\cdot)$ .", + "bbox": [ + 511, + 215, + 903, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. This is immediately apparent from Eqn.(6), as the right-hand-side of Eqn.(7) is by definition $I(\\pmb {X};\\hat{\\pmb{X}})$", + "bbox": [ + 511, + 258, + 903, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Theorem 1, for the direct coding model, it is clear that minimizing the latent entropy $H(\\mathbf{U})$ can be equivalently achieved by maximizing the conditional source entropy $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ , for any known source $\\mathbf{X}$ . In practice, this aligns with the training setting of the neural compression network, wherein the training set is given in advance, and thus $H(\\mathbf{X})$ is fixed. However, as the prevailing neural compression networks are based on the transform coding model, in the following discussion, we extend this conclusion to the transform coding setting.", + "bbox": [ + 511, + 306, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lemma 2. Considering a transform coding model built on the direct coding model elaborated in Lemma 1 with a deterministic analysis transform $T_{A}$ and a deterministic synthesis transform $T_{S}$ (Fig. 1(c)), the following equalities hold automatically:", + "bbox": [ + 511, + 472, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) + H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 560, + 903, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. Compared to the proof of Lemma 1, for the transform coding, we still have $H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0$ and $H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0$ . However, since the synthesis transform $T_{S}$ is not guaranteed to be a bijection mapping, the $H(\\mathbf{U}|\\hat{\\mathbf{X}}) = 0$ does not always hold, leading to the correction of Eqn.(9) and thus Eqn.(10). For more details, readers may refer to the Supplementary Material (Sec. 7).", + "bbox": [ + 511, + 628, + 905, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 2. For the transform coding model elaborated in Lemma 2, we have", + "bbox": [ + 511, + 753, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH (\\boldsymbol {U}) = H (\\boldsymbol {X}) - \\left(H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}})\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 797, + 903, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof. This is immediately apparent from Eqn.(10).", + "bbox": [ + 511, + 830, + 857, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remarks. Two conclusions can be drawn for the transform coding model.", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15575", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Equivalence of minimization and maximization. From Theorem 2, for any known source $\\mathbf{X}$ , minimizing $H(\\mathbf{U})$ can be equivalently achieved by maximizing the $H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})$ . For a bijection synthesis function $T_{S}$ , of which $H(\\mathbf{U}|\\hat{\\mathbf{X}})$ is zero, this maximization objective is reduced to $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ , which is the same as the direct coding;", + "(b) Achieving $R(D)$ by minimizing $H(\\pmb{U})$ . From Eqn.(10), minimizing $H(\\pmb{U})$ is simultaneously minimizing $I(\\pmb{X};\\hat{\\pmb{X}})$ and $H(\\pmb{U}|\\hat{\\pmb{X}})$ . In the ideal case, the informational lower bound given in Eqn.(1) is achieved with $H(\\pmb{U}|\\hat{\\pmb{X}})$ being reduced to 0." + ], + "bbox": [ + 94, + 90, + 483, + 273 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6bcb1680f39bd2d58e0c52077b769a750e0c7f366f92710bb39426740177ba1a.jpg", + "image_caption": [ + "Figure 2. An illustration of the proposed regularization method, wherein an additional source entropy model $q_{\\theta}$ is introduced. The AE and AD represent arithmetic encoding and arithmetic decoding, respectively. The side information branch (i.e., the $\\hat{Z}$ branch) is considered as part of the latent entropy model, which necessitates bits transmission." + ], + "image_footnote": [], + "bbox": [ + 114, + 287, + 454, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Training Strategy", + "text_level": 1, + "bbox": [ + 91, + 561, + 297, + 575 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for each training step do", + "2: Sample a minibatch of training images.", + "3: Perform the forward pass of the compression network and source entropy model $q_{\\theta}$ .", + "4: Compute the backward gradient and update the compression network based on Eqn.(12).", + "5: Freeze the compression network.", + "6: Perform the forward pass of $q_{\\theta}$ .", + "7: Compute the backward gradient and update the source entropy model $q_{\\theta}$ based on Eqn.(13).", + "8: end for" + ], + "bbox": [ + 101, + 580, + 482, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Information-theoretic regularizer", + "text_level": 1, + "bbox": [ + 89, + 772, + 383, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In principle, optimizing neural compression networks with a quantization surrogate, e.g., AUN, introduces gradient estimation errors [14], thereby deviating the objective of minimizing the latent entropy $H(\\mathbf{U})$ under certain distortion constraints (Eqn.(2)), harming both the efficiency and efficacy of the optimization. Motivated by the equivalence of the minimization and maximization optimization discussed", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Sec. 3.1, we propose to address this issue by imposing structural maximization constraints on the optimization process. As such, the vanilla objective of minimizing $H(\\mathbf{U})$ can be further enhanced through this maximization regularization.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Theorem 2, an ideal gradient direction for model updating is one that simultaneously maximizes $H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})$ , which can be further interpreted as maximizing $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ while minimizing $H(\\mathbf{U}|\\hat{\\mathbf{X}})$ . To integrate this regularization into the training objective, two parameterized distributions $q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})$ and $q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}})$ are necessitated. However, accurately modeling an entropy model $q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}})$ alongside the compression network is quite challenging, as it involves learning another \"encoder\" function from the reconstruction $\\hat{\\mathbf{X}}$ to the latent $\\mathbf{U}$ . We experimentally find that a poorly learned $H(\\mathbf{U}|\\hat{\\mathbf{X}})$ regularizer may harm the optimization performance, and that the increased training complexity is also burdensome. To balance practical feasibility with regularization accuracy, we include only the first $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ term for regularization, which is an upper bound for the ground-truth $H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})$ objective. The entropy modeling becomes more tractable since the reconstruction $\\hat{\\mathbf{X}}$ is closely aligned with the source $\\mathbf{X}$ . The primary rationale for such simplification is that although maximizing $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ alone is less precise than $H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})$ , it still provides a meaningful constraint on the \"noisy\" gradient, yielding better gradient direction than no regularization. Another justification is that since we will retain the $H(\\mathbf{U})$ minimization term in the overall objective, this $H(\\mathbf{U}|\\hat{\\mathbf{X}})$ gap, in principle (Eqn.(11)), should decrease as training progresses, which ensures that the regularization effect remains aligned with the ideal maximization objective. The resulting training objective is thus given by:", + "bbox": [ + 511, + 167, + 906, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min R + \\lambda D + \\alpha \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} \\left[ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\right]} _ {\\approx - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}})}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 595, + 903, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R + \\lambda D$ follows the rate-distortion loss in Eqn.(2); $-H(\\mathbf{X}|\\hat{\\mathbf{X}})$ is the proposed regularizer; $q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})$ is the estimated source distribution; $\\alpha$ is regularization factor, controlling the degree of regularization. An illustration for this regularization method is given in Fig.(2). In addition to the compression network, $q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})$ is introduced, which is parameterized by a neural network and can be optimized via the maximum likelihood criteria [41]:", + "bbox": [ + 511, + 642, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\mathbb {E} _ {\\boldsymbol {X}} [ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) ]. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 770, + 903, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This results in a max-min optimization problem where minimizing $-H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})$ (Eqn.(12)) simultaneously creates a conflict. To resolve this, a GAN-style two-stage joint training strategy [42] is employed: At each training step, we first update the compression network using Eqn.(12), then freeze it and update $q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})$ using Eqn.(13). We summarize the strategy in Algorithm 1.", + "bbox": [ + 511, + 794, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15576", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/38fa7dc18bec11553ab0f5a84ff182b713062af3a4c1b0221f143fba9f9dc7a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 89, + 243, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9f24ee091835ee48c889f36d3a02d26b1f2aa3a6eaeaf5640f07f9fff9d554a9.jpg", + "image_caption": [ + "Figure 3. Conditional source entropy modeling for (a) hyperprior [4], (b) autoregressive [8] and attention [10], (c) EILC [43], and (d) MLIC++ [12]. The architecture and module designs are aligned with the latent entropy model. Additional details regarding the specific modules are provided in the Supplementary Material (Sec. 9)." + ], + "image_footnote": [], + "bbox": [ + 99, + 166, + 236, + 292 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3eaf34fbfde9d337f1fe3961e3024d6e7ef90a76df19d45a15f239c1266fd376.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 114, + 575, + 294 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b93818c47499f3a27ff75f292f5372684cbb4fae645961d619278d4728e2f5a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 108, + 895, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 373, + 223, + 390 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 398, + 277, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compression models. Three classic image compression models are used in exhaustive validation experiments: the hyperprior context model [4] (with scale and mean [8]), the joint hyperprior and spatial autoregressive context model [8], and the attention-based transform model [10]. We may refer to these models as hyperprior, autoregressive, and attention, respectively. Moreover, we conduct the experiments on two advanced compression models: ELIC [43] and MLIC++ [12]. Notably, ELIC utilizes checkerboard and channel-conditioned latent entropy contexts, while MLIC++ builds upon ELIC by introducing additional attention-based context designs and is identified as one of the state-of-the-art (SOTA) models.", + "bbox": [ + 89, + 421, + 483, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train these models from scratch using the vanilla rate-distortion loss (Eqn.(2)), and compare their performance when trained with the proposed regularization method (Algorithm 1) under an equal number of training steps. For the implementations of the hyperprior, autoregressive, and attention models, we adopt CompressAI's re-implementation [44], where the quantization surrogate is ANU. For ELIC and MLIC++, we use the open re-implementations from [45] and the official implementation, respectively, both of which employ a mixed ANU-STE quantization surrogate [9].", + "bbox": [ + 89, + 619, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditional source entropy modeling. For each compression model, we apply exactly the same entropy modeling architecture as the latent entropy modeling $q_{\\phi}(\\boldsymbol{U}|\\hat{\\boldsymbol{Z}})$ . Our experiments indicate that aligning the entropy model between the latent and source is critical for effective regularization; deviations in either direction (stronger or weaker", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "source models) degrade compression performance. For the hyperprior model [4], $q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})$ is modeled as the factorized Gaussian distribution conditioned on $\\hat{\\boldsymbol{X}}$ , i.e.,", + "bbox": [ + 511, + 375, + 906, + 420 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\sim \\prod_ {i = 1} ^ {N} \\mathbb {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 428, + 906, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\Phi \\triangleq \\left(\\mu_ {1}, \\dots , \\mu_ {N}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {N} ^ {2}\\right) = f _ {\\theta} (\\hat {\\boldsymbol {X}}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 472, + 906, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $f_{\\theta}$ is modeled by convolutional layers; $\\mathbb{N}$ denotes the Gaussian distribution; $\\Phi$ denote the mean and variance; $N$ denotes the dimension of both $X$ and $\\hat{X}$ . For both the autoregressive and attention models, the mean and variance are determined by the additional spatial context, i.e.,", + "bbox": [ + 511, + 500, + 905, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mu_ {1}, \\dots , \\mu_ {i}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {i} ^ {2}\\right) = f _ {\\theta} \\left(\\hat {\\boldsymbol {X}}, \\boldsymbol {X} _ {< i}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 584, + 906, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{ 1)$ slice, both the $\\Phi_{\\text{anchor}}^{k}$ and $\\Phi_{\\text{non-anchor}}^{k}$ are conditioned on extra channel context from previous slices $X^{Test setRegularized v.s. UnregularizedhyperpriorautoregressiveattentionELICMLIC++Kodak-1.14%-1.57%-0.84%-1.24%-1.82%Tecnick-0.83%-0.78%-0.91%-2.00%-4.01%CLIC-0.66%-0.83%-1.01%-1.45%-3.25%Natural Avg.-0.88%-1.06%-0.92%-1.56%-3.03%Pixel-style-2.20%-1.96%-0.47%-2.71%-3.08%Screen-1.15%-2.01%-0.69%-1.02%-1.72%Game-1.31%-0.83%-0.82%-2.04%-3.44%Pathology-1.11%-0.81%-2.38%-1.82%-0.43%", + "bbox": [ + 516, + 648, + 903, + 835 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1The Crater dataset is excluded due to insufficient documentation.", + "bbox": [ + 529, + 875, + 880, + 886 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2Folder: BRACS_Roi\\latest_version\\test\\0_N", + "bbox": [ + 532, + 887, + 779, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15578", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3ecf35ee4d10ed0c30c11f9ef9557f0deabd766c4b92af07e195f17f833d29b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 145, + 88, + 433, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/511e0d8346e0cbe8a3cbc7bf6b31c9e4ebf19e3179813dd78d538355741921b3.jpg", + "image_caption": [ + "Figure 5. Compression performance on Kodak. Our regularizer achieves $-1.13\\%$ , $-1.57\\%$ , $-0.84\\%$ , $-1.24\\%$ , and $-1.82\\%$ BD-Rates for the hyperprior, autoregressive, attention models, ELIC, and MLIC++, respectively." + ], + "image_footnote": [], + "bbox": [ + 143, + 266, + 433, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 542, + 184, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on natural images. In Fig.4, we present the BD-Rate comparison at four different training steps. The anchor is trained with the vanilla rate-distortion loss (Eqn.(2)) under an equal number of training steps. The regularization factors $\\alpha$ for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models are set to 0.1, 1, 1, 0.1, and 0.1, respectively. It can be observed that, for training steps beyond $0.5 \\times 10^{6}$ across all five models, our proposed regularization method can achieve better compression performance than the vanilla optimization method. In particular, at the $2 \\times 10^{6}$ step, the average BD-Rates for hyperprior, autoregressive, and attention are $-0.88\\%$ , $-1.06\\%$ and $-0.92\\%$ , respectively; at the $1 \\times 10^{6}$ step, the average BD-Rates for ELIC and MLIC++ are $-1.56\\%$ and $-3.03\\%$ , respectively. Meanwhile, the results demonstrate that the regularization effect varies across models as training progresses: For the hyperprior, autoregressive, and MLIC++ models, after a brief convergence period, the BD-Rates are gradually improved from 0 to around $-1\\%$ . For the ELIC model, the regularization is most effective during the earlier training stages, achieving over $2.5\\%$ BD-Rates saving at the $0.25 \\times 10^{6}$ step. For the attention model, the regularization", + "bbox": [ + 89, + 568, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves a relatively stable coding gain.", + "bbox": [ + 511, + 90, + 777, + 106 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 5, we present the detailed rate-distortion curves on the Kodak dataset. For the hyperprior, autoregressive, and attention models, the results correspond to the $2 \\times 10^{6}$ training step, while for the ELIC and MLIC++ models, the results are from the $1 \\times 10^{6}$ step. The results indicate that the introduction of the additional conditional source entropy regularization leads to a slight reduction in the bit rate for most of the points, although a few exceptions exhibit an increase due to training noise. From the form of regularization in Eqn.(12), i.e., $-\\alpha H(\\mathbf{X}|\\hat{\\mathbf{X}})$ , the extent of such bit rate reduction depends on both the regularization factor $\\alpha$ and the estimated conditional source entropy $H(\\mathbf{X}|\\hat{\\mathbf{X}})$ .", + "bbox": [ + 511, + 107, + 906, + 289 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizing on unseen domains. In machine learning, regularization plays an essential role in preventing overfitting and enhancing generalization. In Table 1, we directly apply the models trained on natural images to the four unseen domains and evaluate the effectiveness of our regularization method. The results indicate that, across all new domains, regularized models consistently outperform unregularized ones. Performance improvements vary by domain and model, with notable gains in some cases. For instance, in the pathology domain with the attention model, the regularization achieves a BD-Rate of $-2.38\\%$ , more than double the improvement observed in the natural domain.", + "bbox": [ + 511, + 305, + 908, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of $\\alpha$ and longer training. In Fig.(6)(a), we continue training both the unregularized and regularized hyperprior models to $4\\times 10^{6}$ steps and demonstrate the effect of the regularization factor $\\alpha$ on the performance. Herein, with extended training, the regularization consistently enhances optimization. The best performance, with $\\alpha = 1$ , is achieved at the $2.5\\times 10^{6}$ step, showing a BD-Rate of $-1.24\\%$ . As training progresses to $4\\times 10^{6}$ steps, the gain reduces to around $0.7\\%$ . For the effect of hyperparameter $\\alpha$ , it is shown that either the value is too large or too small will not benefit the optimization, which can be explained as follows. As discussed in Sec. 3.2, our $\\alpha H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})$ maximization regularization is a bound upper of the ground-truth objective $\\alpha (H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}}) - H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}}))$ . When $\\alpha$ is large, this $H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})$ gap is also amplified, which may harm the regularization accuracy and thus the optimization performance. While one might consider addressing this through additional $H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})$ estimation using Algorithm 1, such an approach proves generally ineffective. The challenge lies in accurately learning a distribution $q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})$ that maps $\\hat{\\boldsymbol{X}}$ to $\\boldsymbol{U}$ , as any discrepancy between $q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})$ and the true $p(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})$ would itself reduce accuracy. Furthermore, the incurred computational overhead is unfavorable. We provide more results in Supplementary Material (Sec. 10). When $\\alpha$ is small, the regularizer has little contribution to the overall loss, and the optimization is reduced to the vanilla objective.", + "bbox": [ + 511, + 505, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15579", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c62d64e8c6bd727c998b9c6214ec2809de4d450b60475f5af0fb4c2a5d9d25a.jpg", + "image_caption": [ + "Figure 6. (a) Effects of regularization factor $\\alpha$ . (b) Regularization performance for the hyperprior compression model with the autoregressive regularizer; (c) the autoregressive compression model with the hyperprior regularizer." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 356, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef32409542c5519175c4fd7d8c235cfb7e4e72799f45f8ca1a09841292cd1771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 88, + 620, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c2270423262d3ec580e415003fe6a60b2d5d84a4ac6a3b89eead9c14fd1a15f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 88, + 893, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the alignment of the entropy models. In our experimental setup, we emphasize the advantages of aligned entropy modeling between latent and source distributions. Herein, we extend the analysis to explore cases where the entropy models are unaligned, investigating two scenarios: a compression model equipped with (a) a stronger source entropy model and (b) a weaker source entropy model. In particular, for the source models of the hyperprior and autoregressive depicted in Fig.3(a) and Fig.3(b), respectively, we consider the autoregressive entropy model with additional spatial contexts to be a more accurate and thus \"stronger\" model. Thereafter, we train the hyperprior compression model with a autoregressive source regularizer and the autoregressive compression model with a hyperprior source regularizer, and the results are depicted in Fig.6(b) and Fig.6(c), respectively. The values of $\\alpha$ are still set to 0.1 and 1 for the hyperprior and autoregressive regularizers, respectively. The performance of the aligned entropy model is also included for reference. Herein, it can be observed that neither a stronger nor a weaker source entropy model consistently facilitates the optimization performance of the compression models. Specifically, when a stronger autoregressive regularizer is imposed on the hyperprior network, the average BD-Rate remains positive throughout training. When a weaker hyperprior regularizer is employed by the autoregressive network, the optimization is temporarily improved at the $1.5 \\times 10^{6}$ training step, with a BD-Rate of $-0.20\\%$ . In contrast, aligning the entropy models leads to significant performance improvements.", + "bbox": [ + 91, + 306, + 483, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Limitations", + "text_level": 1, + "bbox": [ + 89, + 756, + 217, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our regularization method is applied exclusively during the network training stage and therefore introduces no additional inference complexity. However, since it involves training an additional source entropy model $q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})$ , the training complexity is inevitably increased. Assuming the complexities of the source and latent entropy models are approximately equal, the complexity overhead ratio can be calculated as $\\frac{C_L}{C_L + C_T} = \\frac{1}{1 + C_T / C_L}$ , which is closely related", + "bbox": [ + 89, + 779, + 483, + 905 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the complexity ratio between the transform backbone and the latent, i.e., $C_T / C_L$ . For $\\lambda = 0.0018$ , we train $1 \\times 10^5$ steps for each model on the same machine (Nvidia RTX 4090) and compute the time ratio. As shown in Table 2, compared to the unregularized baseline, the training time of our proposed method increases by $28.8\\%$ , $18.2\\%$ , $8.1\\%$ , $39.3\\%$ , and $29.8\\%$ for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models, respectively. Notably, for the attention model, the overhead is the least significant. This can be attributed to the fact that the attention-based transform backbone is significantly more complex than the entropy model. To further reduce training complexity, one can employ lightweight convolutional layers, such as depth-wise separable convolution [56], or improve the iterative training strategy outlined in Algorithm 1.", + "bbox": [ + 511, + 306, + 906, + 532 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8598167f121ca923c57111eeb39fe4e6409e98db4cbb7a1a707e1578ebf6d4d5.jpg", + "table_caption": [ + "Table 2. Training complexity of the proposed regularizer compared to the unregularized anchor." + ], + "table_footnote": [], + "table_body": "
hyperpriorautoregressiveattentionELICMLIC++
Training time128.8%118.2%108.1%139.3%129.8%
", + "bbox": [ + 516, + 583, + 903, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper investigates a structural entropy regularizer for lossy neural image compression. Specifically, motivated by the information-theoretic analysis of compression models, we demonstrate that, in addition to minimizing latent entropy, maximizing conditional source entropy can be equally important. Subsequently, by regularizing the neural networks to simultaneously maximize the conditional source entropy during end-to-end optimization, we experimentally observe improvements in both the in-domain compact representation and the out-of-domain generalization abilities of compression models. Additionally, as the primary overhead of this regularizer comes from training an additional source entropy model, the proposed method imposes no added inference complexity, opening up a new space for regularizing neural compression networks.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15580", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] C. E. Shannon et al., \"Coding theorems for a discrete source with a fidelity criterion,\" IRE Nat. Conv. Rec, vol. 4, no. 142-163, p. 1, 1959. 1", + "[2] K. Sayood, Introduction to data compression. Morgan Kaufmann, 2017. 1", + "[3] Y. Yang, S. Mandt, L. Theis, et al., \"An introduction to neural data compression,\" Foundations and Trends® in Computer Graphics and Vision, vol. 15, no. 2, pp. 113-200, 2023. 1", + "[4] J. Balle, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, \"Variational image compression with a scale hyperprior,\" arXiv preprint arXiv:1802.01436, 2018. 1, 2, 3, 5, 6", + "[5] T. Berger, “Rate-distortion theory,” Wiley Encyclopedia of Telecommunications, 2003. 1", + "[6] V. K. Goyal, “Theoretical foundations of transform coding,” IEEE Signal Processing Magazine, vol. 18, no. 5, pp. 9–21, 2001. 1", + "[7] J. Balle, V. Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” arXiv preprint arXiv:1611.01704, 2016. 2", + "[8] D. Minnen, J. Balle, and G. D. Toderici, \"Joint autoregressive and hierarchical priors for learned image compression,\" Advances in neural information processing systems, vol. 31, 2018. 2, 5, 6, 1", + "[9] D. Minnen and S. Singh, \"Channel-wise autoregressive entropy models for learned image compression,\" in 2020 IEEE International Conference on Image Processing (ICIP), pp. 3339-3343, IEEE, 2020. 2, 5", + "[10] Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, \"Learned image compression with discretized gaussian mixture likelihoods and attention modules,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7939-7948, 2020. 2, 5, 6, 1", + "[11] D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin, \"Checkerboard context model for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14771-14780, 2021. 2", + "[12] W. Jiang and R. Wang, \"MLIC++: Linear complexity multi-reference entropy modeling for learned image compression,\" in ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 5, 6", + "[13] S. Qin, J. Wang, Y. Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y. Wang, \"MambaVC: Learned visual compression with selective state spaces,\" arXiv preprint arXiv:2405.15413, 2024. 2", + "[14] H. Zhang, L. Li, and D. Liu, \"On uniform scalar quantization for learned image compression,\" arXiv preprint arXiv:2309.17051, 2023. 2, 4", + "[15] E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Timofte, L. Benini, and L. V. Gool, \"Soft-to-hard vector quantization for end-to-end learning compressible representations,\" Advances in neural information processing systems, vol. 30, 2017. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Y. Yang, R. Bamler, and S. Mandt, \"Improving inference for neural image compression,\" Advances in Neural Information Processing Systems, vol. 33, pp. 573-584, 2020. 2", + "[17] E. Agustsson and L. Theis, \"Universally quantized neural compression,\" Advances in neural information processing systems, vol. 33, pp. 12367-12376, 2020. 2", + "[18] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Soft then hard: Rethinking the quantization in neural image compression,\" in International Conference on Machine Learning, pp. 3920-3929, PMLR, 2021. 2, 3", + "[19] D. P. Kingma, M. Welling, et al., \"An introduction to variational autoencoders,\" Foundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307-392, 2019. 2", + "[20] J. Zhao, Y. Kim, K. Zhang, A. Rush, and Y. LeCun, \"Adversarily regularized autoencoders,\" in International conference on machine learning, pp. 5902-5911, PMLR, 2018. 2, 3", + "[21] H. Xu, D. Luo, R. Henao, S. Shah, and L. Carin, \"Learning autoencoders with relational regularization,\" in International Conference on Machine Learning, pp. 10576-10586, PMLR, 2020.", + "[22] H. Wu and M. Flierl, \"Vector quantization-based regularization for autoencoders,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 6380-6387, 2020. 3", + "[23] L. Le, A. Patterson, and M. White, \"Supervised autoencoders: Improving generalization performance with unsupervised regularizers,\" Advances in neural information processing systems, vol. 31, 2018. 3", + "[24] T. Ma, J. Chen, and C. Xiao, \"Constrained generation of semantically valid graphs via regularizing variational autoencoders,\" Advances in Neural Information Processing Systems, vol. 31, 2018. 3", + "[25] S. Sinha and A. B. Dieng, \"Consistency regularization for variational auto-encoders,\" Advances in Neural Information Processing Systems, vol. 34, pp. 12943–12954, 2021. 2, 3", + "[26] H. Ochoa-Dominguez and K. R. Rao, Discrete cosine transform. CRC Press, 2019. 2", + "[27] G. Toderici, S. M. O'Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, \"Variable rate image compression with recurrent neural networks,\" arXiv preprint arXiv:1511.06085, 2015. 2", + "[28] H. Liu, T. Chen, P. Guo, Q. Shen, X. Cao, Y. Wang, and Z. Ma, “Non-local attention optimized deep image compression,” arXiv preprint arXiv:1904.09757, 2019. 2", + "[29] M. Lu, P. Guo, H. Shi, C. Cao, and Z. Ma, \"Transformer-based image compression,\" arXiv preprint arXiv:2111.06707, 2021. 2", + "[30] J. Liu, H. Sun, and J. Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14388–14397, 2023. 2", + "[31] H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, \"End-to-end optimized versatile image compression with wavelet-like transform,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1247-1263, 2020. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15581", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] X. Zhang and X. Wu, \"LVQAC: Lattice vector quantization coupled with spatially adaptive companding for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10239-10248, 2023. 2", + "[33] R. Feng, Z. Guo, W. Li, and Z. Chen, \"NVTC: Nonlinear vector transform coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6101-6110, 2023. 2", + "[34] K. Suhring, M. Schäfer, J. Pfaff, H. Schwarz, D. Marpe, and T. Wiegand, \"Trellis-coded quantization for end-to-end learned image compression,\" in 2022 IEEE International Conference on Image Processing (ICIP), pp. 3306-3310, IEEE, 2022. 2", + "[35] Z. Ge, S. Ma, W. Gao, J. Pan, and C. Jia, \"NLIC: Nonuniform quantization based learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2", + "[36] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Causal contextual prediction for learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2329-2341, 2021. 2", + "[37] X. Zhu, J. Song, L. Gao, F. Zheng, and H. T. Shen, \"Unified multivariate gaussian mixture for efficient neural image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17612-17621, 2022. 2", + "[38] H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, \"Learned image compression with gaussian-laplacian-logistic mixture model and concatenated residual modules,\" IEEE Transactions on Image Processing, vol. 32, pp. 2063-2076, 2023. 2", + "[39] Y. Hu, W. Yang, and J. Liu, \"Coarse-to-fine hyper-prior modeling for learned image compression,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 11013-11020, 2020. 2", + "[40] L. Theis, W. Shi, A. Cunningham, and F. Huszár, \"Lossy image compression with compressive autoencoders,\" arXiv preprint arXiv:1703.00395, 2017. 3", + "[41] I. Goodfellow, \"Deep learning,\" 2016. 4", + "[42] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, \"Generative adversarial nets,\" Advances in neural information processing systems, vol. 27, 2014. 4", + "[43] D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang, \"ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5718-5727, 2022. 5, 6, 2", + "[44] J. Bégaint, F. Racapé, S. Feltman, and A. Pushparaja, \"Compressai: a pytorch library and evaluation platform for end-to-end compression research,\" arXiv preprint arXiv:2011.03029, 2020. 5, 1", + "[45] W. Jiang, “Unofficial elic.” https://github.com/JiangWeibeta/ELIC, 2022.5", + "[46] J. Liu, G. Lu, Z. Hu, and D. Xu, \"A unified end-to-end framework for efficient deep image compression,\" arXiv preprint arXiv:2002.03370, 2020. 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 6", + "[48] E. Kodak, \"Kodak lossless true color image suite (photocd pcd0992),\" URL http://rOk.us/graphics/kodak, vol. 6, p. 2, 1993. 6", + "[49] \"6th Challenge on Learned Image Compression.\" https://www.compression.cc/tasks/index.html. Accessed: 2024-11-13. 6", + "[50] N. Asuni and A. Giachetti, \"Testimages: a large-scale archive for testing visual devices and basic image processing algorithms,\" in STAG, pp. 63-70, 2014. 6", + "[51] T. Zhang, H. Zhang, Y. Li, L. Li, and D. Liu, “Few-shot domain adaptation for learned image compression,” arXiv preprint arXiv:2409.11111, 2024. 6", + "[52] Y. Lv, J. Xiang, J. Zhang, W. Yang, X. Han, and W. Yang, \"Dynamic low-rank instance adaptation for universal neural image compression,\" in Proceedings of the 31st ACM International Conference on Multimedia, pp. 632-642, 2023. 6", + "[53] J. Yang, S. Shen, H. Yue, and K. Li, \"Implicit transformer network for screen content image continuous superresolution,\" Advances in Neural Information Processing Systems, vol. 34, pp. 13304-13315, 2021. 6", + "[54] X. Min, K. Ma, K. Gu, G. Zhai, Z. Wang, and W. Lin, \"Unified blind quality assessment of compressed natural, graphic, and screen content images,\" IEEE Transactions on Image Processing, vol. 26, no. 11, pp. 5462-5474, 2017. 6", + "[55] N. Brancati, A. M. Anniciello, P. Pati, D. Riccio, G. Scognamiglio, G. Jaume, G. De Pietro, M. Di Bonito, A. Fon-cubierta, G. Botti, et al., \"BRACS: A dataset for breast carcinoma subtyping in h&e histology images,\" Database, vol. 2022, p. baac093, 2022. 6", + "[56] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1251–1258, 2017. 8, 2", + "[57] G. Bjontegaard, \"Calculation of average PSNR differences between RD-curves,\" ITU SG16 Doc. VCEG-M33, 2001. 2" + ], + "bbox": [ + 514, + 90, + 903, + 616 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15582", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_model.json b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f31094b20b57e2fa46d0916cfb21e3621b04d8b4 --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.885, + 0.154 + ], + "angle": 0, + "content": "An Information-Theoretic Regularizer for Lossy Neural Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.18, + 0.741, + 0.218 + ], + "angle": 0, + "content": "Yingwen Zhang\\(^{1}\\), Meng Wang\\(^{2*}\\), Xihua Sheng\\(^{1}\\), Peilin Chen\\(^{1}\\), Junru Li\\(^{3}\\), Li Zhang\\(^{3}\\), Shiqi Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.217, + 0.838, + 0.271 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science, City University of Hong Kong, Hong Kong, China \\(^{2}\\)School of Data Science, Lingnan University, Hong Kong, China \\(^{3}\\)Bytedance Inc., San Diego, USA" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.486, + 0.624 + ], + "angle": 0, + "content": "Lossy image compression networks aim to minimize the latent entropy of images while adhering to specific distortion constraints. However, optimizing the neural network can be challenging due to its nature of learning quantized latent representations. In this paper, our key finding is that minimizing the latent entropy is, to some extent, equivalent to maximizing the conditional source entropy, an insight that is deeply rooted in information-theoretic equalities. Building on this insight, we propose a novel structural regularization method for the neural image compression task by incorporating the negative conditional source entropy into the training objective, such that both the optimization efficacy and the model's generalization ability can be promoted. The proposed information-theoretic regularizer is interpretable, plug-and-play, and imposes no inference overheads. Extensive experiments demonstrate its superiority in regularizing the models and further squeezing bits from the latent representation across various compression structures and unseen domains. Code is available at Github." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.222, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Lossy image compression, or more broadly, lossy data compression, is grounded in a fundamental theorem in information theory: the rate-distortion theorem [1]. Essentially, this theorem establishes a theoretical lower bound for lossy compression and provides valuable guidance for designing practical compression algorithms [2, 3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.858 + ], + "angle": 0, + "content": "A basic lossy image compression model is illustrated in Fig.1(a). Let \\( \\mathbf{X} \\), \\( \\hat{\\mathbf{X}} \\), and \\( \\mathbf{U} \\) denote the source, the reconstruction, and the discrete index (or the latent), respectively. The encoder is a deterministic mapping \\( \\mathrm{Q} \\colon \\mathcal{X}^N \\to \\mathcal{U} \\) where \\( \\mathcal{X}^N \\) denotes a \\( N \\)-dimensional image space and \\( \\mathcal{U} = \\{1, 2, \\dots, M\\} \\) denotes the discrete index space. The decoder" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.304, + 0.668, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.303, + 0.895, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.485, + 0.576, + 0.5 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.412, + 0.83, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.571 + ], + "angle": 0, + "content": "Figure 1. (a) Direct coding model; (b) Information diagram for the direct coding, wherein \\( H(\\mathbf{X}) = H(\\mathbf{X}|\\hat{\\mathbf{X}}) + H(\\mathbf{U}) \\) and \\( H(\\mathbf{X}) \\) is fixed for any known source; (c) Transform coding model (with possible side information \\( \\hat{\\mathbf{Z}} \\) [4])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.744 + ], + "angle": 0, + "content": "is a bijection mapping \\( \\mathrm{Q}^{-1} \\colon \\mathcal{U} \\to \\hat{\\mathcal{X}}^N \\), where the reconstruction space \\( \\hat{\\mathcal{X}}^N = \\{c_1, \\dots, c_M\\} \\) forms a \\( N \\)-dimensional codebook. For index \\( U \\), entropy coding is applied to losslessly convert the data into the bitstream, which can then be decoded back to \\( U \\) on the decoder side. We may refer to such a model as the direct coding model for its structural simplicity. Subsequently, following the rate-distortion theorem [1], given a bounded distortion measure \\( d(\\cdot) \\) and a distortion constraint \\( D \\), the rate of transmitting \\( U \\) is lower bounded by the informational rate-distortion function:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.751, + 0.907, + 0.788 + ], + "angle": 0, + "content": "\\[\nR (D) = \\min _ {p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}: \\mathbb {E} _ {p _ {\\boldsymbol {X}} p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}} [ d (\\boldsymbol {X}, \\hat {\\boldsymbol {X}}) ] < D} \\frac {1}{N} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "i.e., the minimum mutual information obtained by traversing all the possible encoder-decoder (or quantizer) designs \\( p_{\\hat{X} |X} \\). Theoretically, this bound can be achieved asymptotically using vector quantization as \\( N \\to \\infty \\) [5]. However, since the encoding complexity grows exponentially with the image dimension \\( N \\), a suboptimal yet more practical solution is the transform coding [6]. As depicted in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (mengwang7@ln.edu.hk). This work was supported in part by ITF Project GHP/044/21SZ, in part by the Start-up Grant SUG-007/2425, and in part by Faculty Research Grant SDS24A6." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15573" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "Fig.1(c), for the transform coding model, a deterministic analysis transform \\( \\mathrm{T_A} \\): \\( \\mathcal{X}^N \\to \\mathcal{Y}^M \\) and a deterministic synthesis transform \\( \\mathrm{T_S} \\): \\( \\hat{\\mathcal{Y}}^M \\to \\hat{\\mathcal{X}}^N \\) are incorporated into the compression chain. The analysis transform is designed to decorrelate the source \\( X \\), such that a much simpler quantization scheme, e.g., uniform scalar quantization: \\( \\boldsymbol{u}_i = \\mathrm{round}(\\boldsymbol{y}_i) \\), can be independently applied to each dimension of the transformed coefficients \\( Y \\). Building on this framework and leveraging the superior nonlinear capabilities of neural networks, numerous neural image compression networks [4, 7-13] are proposed, continually pushing the frontier of lossy image compression." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Typically, by parameterizing the \\(\\mathrm{T_A}\\) and \\(\\mathrm{T_S}\\) in Fig.1(c) as neural networks, the rate-distortion training objective is formulated as [7]:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.328, + 0.483, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\min _ {R \\approx H (U)} \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} [ - \\log q _ {\\phi} (\\boldsymbol {U}) ]} + \\lambda \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} | | \\boldsymbol {X} - \\hat {\\boldsymbol {X}} | | _ {2} ^ {2}} _ {D}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.56 + ], + "angle": 0, + "content": "where \\(R\\) is the estimation of latent entropy \\(H(\\pmb{U})\\); \\(q_{\\phi}\\) is the neural entropy model (or estimated distribution) for \\(\\pmb{U}\\), whose modeling accuracy can be significantly enhanced by transmitting a small amount of side information \\(\\hat{\\pmb{Z}}\\) (as depicted in Fig.1(c)) as part of the parameters \\(\\phi\\) [4]; the distortion \\(D\\) is the \\(l_{2}\\) distance between \\(\\pmb{X}\\) and \\(\\hat{\\pmb{X}}\\); \\(\\lambda\\) is the Lagrangian multiplier, controlling the trade-off between \\(R\\) and \\(D\\). Herein, due to the existence of quantization \\(\\mathrm{Q}(\\cdot)\\), \\(\\pmb{U}\\) is discrete in essence. Optimizing the network is then challenging as the gradient is almost zero everywhere for \\(\\mathrm{Q}(\\cdot)\\). To effectively train the networks, a surrogate is introduced during training, e.g., additive uniform noise (AUN) [7]:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.586, + 0.483, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {U} = \\mathrm {Q} (\\boldsymbol {Y}) \\approx \\boldsymbol {Y} + [ - 0. 5, 0. 5 ] ^ {M}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.705 + ], + "angle": 0, + "content": "such that the backward gradient of \\( \\mathrm{Q}(\\cdot) \\) is one. However, since such approximation inevitably results in the train-test mismatch, and thus high gradient estimation error [14], both the efficiency and efficacy of the optimization are hindered. To address this, many research efforts are devoted to designing advanced surrogates and optimization methods [14-18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a completely distinct view on facilitating the optimization of the compression network: regularization. Conceivably, imposing structural constraints on the network optimization can effectively regularize the gradient estimation error, promote convergence, and reduce overfitting. In the context of generative variational autoencoders [19], structural regularization techniques [20-25] have been effectively applied to improve authentic data generation. However, since neural compression networks fundamentally differ from generative autoencoders in terms of the minimum latent entropy constraint, to the best of our knowledge, no structural regularization methods tailored to this constraint have been explored. In light of this, in this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "paper, we present the first study of information-theoretic regularization for the neural image compression task. In particular, through an in-depth information-theoretic analysis of compression models, we reveal that minimizing the latent entropy \\( H(\\mathbf{U}) \\) is, to some extent, equally important as maximizing the conditional source entropy \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\). Subsequently, by enforcing the network to simultaneously minimize \\( -H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\), both the in-domain compact representation and out-of-domain generalization ability are experimentally improved." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.26, + 0.661, + 0.276 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.907, + 0.74 + ], + "angle": 0, + "content": "Neural image compression networks. Neural image compression networks are primarily composed of transform, quantization, and latent entropy models. The transform module maps the input source into a latent representation so that the dependencies of the source domain can be more easily handled. Unlike the linear transform [26] used in traditional hybrid coding, the neural transform is inherently nonlinear, enabling more effective decorrelation of the source data. In the literature, various neural transform architectures are proposed. Representative architectures include the recurrent neural network (RNN) [27], convolutional neural network (CNN) [7], non-local attention [28], transformer [29], mixed transformer-CNN [30], state-space models [13] and invertible networks [31]. Subsequently, the continuous transformed coefficients are quantized into the discrete representation, wherein the uniform scalar quantization [7] is commonly adopted. As vector quantization can, in principle, better approach the rate-distortion bound (Eqn.(1)), low-complexity vector quantization [15, 32, 33] and dependent scalar quantization [34, 35] schemes are also studied. Finally, once the quantized latent representations are obtained, a neural entropy model is essential for performing entropy coding. Herein, since the latent cannot be perfectly decorrelated, an accurate entropy model that can better capture the dependencies and closely match the ground-truth latent distribution is crucial for the overall network performance. Prominent research directions such as spatial-channel context mining [8, 9, 12, 36], expressive latent distribution modeling [10, 37, 38] and latent architecture designs [4, 11, 39] are explored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Optimization over quantized latent representations. The surrogate-based methods [7, 14-17] have been shown to effectively address the non-differentiable quantization function during end-to-end optimization. In addition to AUN [7], various quantization surrogates such as soft assignment [15], Gumbel-softmax trick [16], and soft rounding function [17] have been studied. However, as the surrogate introduces a train-test mismatch in forward calculations, the annealing techniques [14-17] are then proposed to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15574" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "control the \"softness\" of surrogates, gradually approximating the hard quantization function as training progresses. The straight-through estimator (STE) [40] provides an alternative approach to address the train-test mismatch by enforcing an identity mapping during the backward calculation, thereby enabling direct hard quantization in the forward pass. Nevertheless, STE ultimately introduces a significant gradient estimation bias for the analysis transform. To counter this, an additional fine-tuning stage for the entropy model and synthesis transform is suggested [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.425 + ], + "angle": 0, + "content": "While numerous network optimization methods have been developed from a quantization perspective, the regularization view has largely escaped research attention. To the best of our knowledge, the most related efforts are regularization methods for variational autoencoders (VAEs), such as regularization on latent priors [20-22], reconstruction [23, 24], and consistency [25]. However, as VAEs are designed for generative modeling, these methods are not applicable to optimizing compression networks. To bridge this knowledge gap, this paper proposes a structural regularizer for neural image compression and conducts extensive experiments to demonstrate its superiority." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.226, + 0.458 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.466, + 0.484, + 0.633 + ], + "angle": 0, + "content": "In this section, we begin by deriving the equivalent maximization form of the latent entropy minimization problem for both the direct and transform coding models. Building on this derivation, we propose a novel regularization method for optimizing the neural compression network. Note that during our derivations, we use the naive transform coding model without the side information branch (i.e., the \\(\\hat{Z}\\) depicted in Fig 1(c)). This significantly simplifies our analysis without compromising much practicality, as it has been shown that the information flowing into the side branch is negligible compared to the latent branch [4]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.644, + 0.483, + 0.674 + ], + "angle": 0, + "content": "3.1. Motivation: from minimization to maximization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Lemma 1. Considering a deterministic quantization process \\( Q(\\cdot) \\) and a deterministic dequantization process \\( Q^{-1}(\\cdot) \\) for the direct coding model illustrated in Fig. 1(a), the following equalities hold automatically:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.754, + 0.483, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) & (4) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) & (5) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) & (6) \\end{array} \\right.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Proof. Eqn.(4) follows the facts that \\( H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0 \\), meaning there is no uncertainty in \\( \\hat{\\mathbf{X}} \\) given the observation of \\( \\mathbf{X} \\), and \\( I(\\mathbf{X};\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}) - H(\\hat{\\mathbf{X}}|\\mathbf{X}) \\). Similarly, Eqn.(5) follows the fact that \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0 \\) and \\( H(\\hat{\\mathbf{X}}) + H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\mathbf{U}) + H(\\hat{\\mathbf{X}}|\\mathbf{U}) \\). Eqn.(6) is imme" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "dately apparent from Eqn.(4) and Eqn.(5). Additional details are provided in the Supplementary Material (Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.141, + 0.905, + 0.171 + ], + "angle": 0, + "content": "Theorem 1. For the direct coding model elaborated in Lemma 1, we have" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.184, + 0.905, + 0.203 + ], + "angle": 0, + "content": "\\[\nH (\\boldsymbol {U}) = H (\\boldsymbol {X}) - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.216, + 0.905, + 0.247 + ], + "angle": 0, + "content": "The informational diagram is thus illustrated as Fig.1(b), given the non-negativity of discrete entropy \\( H(\\cdot) \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Proof. This is immediately apparent from Eqn.(6), as the right-hand-side of Eqn.(7) is by definition \\(I(\\pmb {X};\\hat{\\pmb{X}})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.907, + 0.461 + ], + "angle": 0, + "content": "From Theorem 1, for the direct coding model, it is clear that minimizing the latent entropy \\( H(\\mathbf{U}) \\) can be equivalently achieved by maximizing the conditional source entropy \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\), for any known source \\( \\mathbf{X} \\). In practice, this aligns with the training setting of the neural compression network, wherein the training set is given in advance, and thus \\( H(\\mathbf{X}) \\) is fixed. However, as the prevailing neural compression networks are based on the transform coding model, in the following discussion, we extend this conclusion to the transform coding setting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.473, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Lemma 2. Considering a transform coding model built on the direct coding model elaborated in Lemma 1 with a deterministic analysis transform \\( T_{A} \\) and a deterministic synthesis transform \\( T_{S} \\) (Fig. 1(c)), the following equalities hold automatically:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.561, + 0.905, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) + H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Proof. Compared to the proof of Lemma 1, for the transform coding, we still have \\( H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0 \\) and \\( H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0 \\). However, since the synthesis transform \\( T_{S} \\) is not guaranteed to be a bijection mapping, the \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) = 0 \\) does not always hold, leading to the correction of Eqn.(9) and thus Eqn.(10). For more details, readers may refer to the Supplementary Material (Sec. 7)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.755, + 0.906, + 0.785 + ], + "angle": 0, + "content": "Theorem 2. For the transform coding model elaborated in Lemma 2, we have" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.798, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\nH (\\boldsymbol {U}) = H (\\boldsymbol {X}) - \\left(H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}})\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.832, + 0.859, + 0.848 + ], + "angle": 0, + "content": "Proof. This is immediately apparent from Eqn.(10)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Remarks. Two conclusions can be drawn for the transform coding model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.097, + 0.092, + 0.483, + 0.197 + ], + "angle": 0, + "content": "(a) Equivalence of minimization and maximization. From Theorem 2, for any known source \\( \\mathbf{X} \\), minimizing \\( H(\\mathbf{U}) \\) can be equivalently achieved by maximizing the \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\). For a bijection synthesis function \\( T_{S} \\), of which \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) is zero, this maximization objective is reduced to \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\), which is the same as the direct coding;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.198, + 0.484, + 0.275 + ], + "angle": 0, + "content": "(b) Achieving \\( R(D) \\) by minimizing \\( H(\\pmb{U}) \\). From Eqn.(10), minimizing \\( H(\\pmb{U}) \\) is simultaneously minimizing \\( I(\\pmb{X};\\hat{\\pmb{X}}) \\) and \\( H(\\pmb{U}|\\hat{\\pmb{X}}) \\). In the ideal case, the informational lower bound given in Eqn.(1) is achieved with \\( H(\\pmb{U}|\\hat{\\pmb{X}}) \\) being reduced to 0." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.092, + 0.484, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.289, + 0.455, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Figure 2. An illustration of the proposed regularization method, wherein an additional source entropy model \\( q_{\\theta} \\) is introduced. The AE and AD represent arithmetic encoding and arithmetic decoding, respectively. The side information branch (i.e., the \\( \\hat{Z} \\) branch) is considered as part of the latent entropy model, which necessitates bits transmission." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.562, + 0.298, + 0.577 + ], + "angle": 0, + "content": "Algorithm 1 Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.582, + 0.285, + 0.596 + ], + "angle": 0, + "content": "1: for each training step do" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.597, + 0.403, + 0.611 + ], + "angle": 0, + "content": "2: Sample a minibatch of training images." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.612, + 0.483, + 0.642 + ], + "angle": 0, + "content": "3: Perform the forward pass of the compression network and source entropy model \\( q_{\\theta} \\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.642, + 0.483, + 0.672 + ], + "angle": 0, + "content": "4: Compute the backward gradient and update the compression network based on Eqn.(12)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.672, + 0.363, + 0.686 + ], + "angle": 0, + "content": "5: Freeze the compression network." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.688, + 0.354, + 0.702 + ], + "angle": 0, + "content": "6: Perform the forward pass of \\(q_{\\theta}\\)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.732 + ], + "angle": 0, + "content": "7: Compute the backward gradient and update the source entropy model \\( q_{\\theta} \\) based on Eqn.(13)." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.733, + 0.174, + 0.745 + ], + "angle": 0, + "content": "8: end for" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.582, + 0.483, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.385, + 0.79 + ], + "angle": 0, + "content": "3.2. Information-theoretic regularizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In principle, optimizing neural compression networks with a quantization surrogate, e.g., AUN, introduces gradient estimation errors [14], thereby deviating the objective of minimizing the latent entropy \\( H(\\mathbf{U}) \\) under certain distortion constraints (Eqn.(2)), harming both the efficiency and efficacy of the optimization. Motivated by the equivalence of the minimization and maximization optimization discussed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "in Sec. 3.1, we propose to address this issue by imposing structural maximization constraints on the optimization process. As such, the vanilla objective of minimizing \\( H(\\mathbf{U}) \\) can be further enhanced through this maximization regularization." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.593 + ], + "angle": 0, + "content": "From Theorem 2, an ideal gradient direction for model updating is one that simultaneously maximizes \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\), which can be further interpreted as maximizing \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\) while minimizing \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\). To integrate this regularization into the training objective, two parameterized distributions \\( q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}}) \\) and \\( q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) are necessitated. However, accurately modeling an entropy model \\( q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) alongside the compression network is quite challenging, as it involves learning another \"encoder\" function from the reconstruction \\( \\hat{\\mathbf{X}} \\) to the latent \\( \\mathbf{U} \\). We experimentally find that a poorly learned \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) regularizer may harm the optimization performance, and that the increased training complexity is also burdensome. To balance practical feasibility with regularization accuracy, we include only the first \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\) term for regularization, which is an upper bound for the ground-truth \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) objective. The entropy modeling becomes more tractable since the reconstruction \\( \\hat{\\mathbf{X}} \\) is closely aligned with the source \\( \\mathbf{X} \\). The primary rationale for such simplification is that although maximizing \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) \\) alone is less precise than \\( H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\), it still provides a meaningful constraint on the \"noisy\" gradient, yielding better gradient direction than no regularization. Another justification is that since we will retain the \\( H(\\mathbf{U}) \\) minimization term in the overall objective, this \\( H(\\mathbf{U}|\\hat{\\mathbf{X}}) \\) gap, in principle (Eqn.(11)), should decrease as training progresses, which ensures that the regularization effect remains aligned with the ideal maximization objective. The resulting training objective is thus given by:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.597, + 0.905, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\min R + \\lambda D + \\alpha \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} \\left[ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\right]} _ {\\approx - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}})}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.907, + 0.765 + ], + "angle": 0, + "content": "where \\(R + \\lambda D\\) follows the rate-distortion loss in Eqn.(2); \\(-H(\\mathbf{X}|\\hat{\\mathbf{X}})\\) is the proposed regularizer; \\(q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})\\) is the estimated source distribution; \\(\\alpha\\) is regularization factor, controlling the degree of regularization. An illustration for this regularization method is given in Fig.(2). In addition to the compression network, \\(q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})\\) is introduced, which is parameterized by a neural network and can be optimized via the maximum likelihood criteria [41]:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.771, + 0.905, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\max \\mathbb {E} _ {\\boldsymbol {X}} [ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) ]. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.902 + ], + "angle": 0, + "content": "This results in a max-min optimization problem where minimizing \\(-H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})\\) (Eqn.(12)) simultaneously creates a conflict. To resolve this, a GAN-style two-stage joint training strategy [42] is employed: At each training step, we first update the compression network using Eqn.(12), then freeze it and update \\(q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})\\) using Eqn.(13). We summarize the strategy in Algorithm 1." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15576" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.244, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.167, + 0.237, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.116, + 0.576, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.109, + 0.896, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 3. Conditional source entropy modeling for (a) hyperprior [4], (b) autoregressive [8] and attention [10], (c) EILC [43], and (d) MLIC++ [12]. The architecture and module designs are aligned with the latent entropy model. Additional details regarding the specific modules are provided in the Supplementary Material (Sec. 9)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.224, + 0.391 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.399, + 0.279, + 0.417 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.619 + ], + "angle": 0, + "content": "Compression models. Three classic image compression models are used in exhaustive validation experiments: the hyperprior context model [4] (with scale and mean [8]), the joint hyperprior and spatial autoregressive context model [8], and the attention-based transform model [10]. We may refer to these models as hyperprior, autoregressive, and attention, respectively. Moreover, we conduct the experiments on two advanced compression models: ELIC [43] and MLIC++ [12]. Notably, ELIC utilizes checkerboard and channel-conditioned latent entropy contexts, while MLIC++ builds upon ELIC by introducing additional attention-based context designs and is identified as one of the state-of-the-art (SOTA) models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.788 + ], + "angle": 0, + "content": "We train these models from scratch using the vanilla rate-distortion loss (Eqn.(2)), and compare their performance when trained with the proposed regularization method (Algorithm 1) under an equal number of training steps. For the implementations of the hyperprior, autoregressive, and attention models, we adopt CompressAI's re-implementation [44], where the quantization surrogate is ANU. For ELIC and MLIC++, we use the open re-implementations from [45] and the official implementation, respectively, both of which employ a mixed ANU-STE quantization surrogate [9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Conditional source entropy modeling. For each compression model, we apply exactly the same entropy modeling architecture as the latent entropy modeling \\( q_{\\phi}(\\boldsymbol{U}|\\hat{\\boldsymbol{Z}}) \\). Our experiments indicate that aligning the entropy model between the latent and source is critical for effective regularization; deviations in either direction (stronger or weaker" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.375, + 0.907, + 0.421 + ], + "angle": 0, + "content": "source models) degrade compression performance. For the hyperprior model [4], \\( q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}}) \\) is modeled as the factorized Gaussian distribution conditioned on \\( \\hat{\\boldsymbol{X}} \\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.429, + 0.907, + 0.47 + ], + "angle": 0, + "content": "\\[\nq _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\sim \\prod_ {i = 1} ^ {N} \\mathbb {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.473, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\Phi \\triangleq \\left(\\mu_ {1}, \\dots , \\mu_ {N}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {N} ^ {2}\\right) = f _ {\\theta} (\\hat {\\boldsymbol {X}}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.577 + ], + "angle": 0, + "content": "and \\( f_{\\theta} \\) is modeled by convolutional layers; \\( \\mathbb{N} \\) denotes the Gaussian distribution; \\( \\Phi \\) denote the mean and variance; \\( N \\) denotes the dimension of both \\( X \\) and \\( \\hat{X} \\). For both the autoregressive and attention models, the mean and variance are determined by the additional spatial context, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.585, + 0.907, + 0.605 + ], + "angle": 0, + "content": "\\[\n\\left(\\mu_ {1}, \\dots , \\mu_ {i}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {i} ^ {2}\\right) = f _ {\\theta} \\left(\\hat {\\boldsymbol {X}}, \\boldsymbol {X} _ {< i}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.902 + ], + "angle": 0, + "content": "where \\(X_{ 1)\\) slice, both the \\(\\Phi_{\\text{anchor}}^{k}\\) and \\(\\Phi_{\\text{non-anchor}}^{k}\\) are conditioned on extra channel context from previous slices \\(X^{Test setRegularized v.s. UnregularizedhyperpriorautoregressiveattentionELICMLIC++Kodak-1.14%-1.57%-0.84%-1.24%-1.82%Tecnick-0.83%-0.78%-0.91%-2.00%-4.01%CLIC-0.66%-0.83%-1.01%-1.45%-3.25%Natural Avg.-0.88%-1.06%-0.92%-1.56%-3.03%Pixel-style-2.20%-1.96%-0.47%-2.71%-3.08%Screen-1.15%-2.01%-0.69%-1.02%-1.72%Game-1.31%-0.83%-0.82%-2.04%-3.44%Pathology-1.11%-0.81%-2.38%-1.82%-0.43%" + }, + { + "type": "page_footnote", + "bbox": [ + 0.53, + 0.875, + 0.881, + 0.887 + ], + "angle": 0, + "content": "1The Crater dataset is excluded due to insufficient documentation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.533, + 0.888, + 0.78, + 0.9 + ], + "angle": 0, + "content": "2Folder: BRACS_Roi\\latest_version\\test\\0_N" + }, + { + "type": "list", + "bbox": [ + 0.53, + 0.875, + 0.881, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15578" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.089, + 0.434, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.267, + 0.434, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.486, + 0.511 + ], + "angle": 0, + "content": "Figure 5. Compression performance on Kodak. Our regularizer achieves \\(-1.13\\%\\), \\(-1.57\\%\\), \\(-0.84\\%\\), \\(-1.24\\%\\), and \\(-1.82\\%\\) BD-Rates for the hyperprior, autoregressive, attention models, ELIC, and MLIC++, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.185, + 0.558 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Performance on natural images. In Fig.4, we present the BD-Rate comparison at four different training steps. The anchor is trained with the vanilla rate-distortion loss (Eqn.(2)) under an equal number of training steps. The regularization factors \\(\\alpha\\) for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models are set to 0.1, 1, 1, 0.1, and 0.1, respectively. It can be observed that, for training steps beyond \\(0.5 \\times 10^{6}\\) across all five models, our proposed regularization method can achieve better compression performance than the vanilla optimization method. In particular, at the \\(2 \\times 10^{6}\\) step, the average BD-Rates for hyperprior, autoregressive, and attention are \\(-0.88\\%\\), \\(-1.06\\%\\) and \\(-0.92\\%\\), respectively; at the \\(1 \\times 10^{6}\\) step, the average BD-Rates for ELIC and MLIC++ are \\(-1.56\\%\\) and \\(-3.03\\%\\), respectively. Meanwhile, the results demonstrate that the regularization effect varies across models as training progresses: For the hyperprior, autoregressive, and MLIC++ models, after a brief convergence period, the BD-Rates are gradually improved from 0 to around \\(-1\\%\\). For the ELIC model, the regularization is most effective during the earlier training stages, achieving over \\(2.5\\%\\) BD-Rates saving at the \\(0.25 \\times 10^{6}\\) step. For the attention model, the regularization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.779, + 0.107 + ], + "angle": 0, + "content": "achieves a relatively stable coding gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.29 + ], + "angle": 0, + "content": "In Fig. 5, we present the detailed rate-distortion curves on the Kodak dataset. For the hyperprior, autoregressive, and attention models, the results correspond to the \\(2 \\times 10^{6}\\) training step, while for the ELIC and MLIC++ models, the results are from the \\(1 \\times 10^{6}\\) step. The results indicate that the introduction of the additional conditional source entropy regularization leads to a slight reduction in the bit rate for most of the points, although a few exceptions exhibit an increase due to training noise. From the form of regularization in Eqn.(12), i.e., \\(-\\alpha H(\\mathbf{X}|\\hat{\\mathbf{X}})\\), the extent of such bit rate reduction depends on both the regularization factor \\(\\alpha\\) and the estimated conditional source entropy \\(H(\\mathbf{X}|\\hat{\\mathbf{X}})\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.909, + 0.489 + ], + "angle": 0, + "content": "Generalizing on unseen domains. In machine learning, regularization plays an essential role in preventing overfitting and enhancing generalization. In Table 1, we directly apply the models trained on natural images to the four unseen domains and evaluate the effectiveness of our regularization method. The results indicate that, across all new domains, regularized models consistently outperform unregularized ones. Performance improvements vary by domain and model, with notable gains in some cases. For instance, in the pathology domain with the attention model, the regularization achieves a BD-Rate of \\(-2.38\\%\\), more than double the improvement observed in the natural domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of \\(\\alpha\\) and longer training. In Fig.(6)(a), we continue training both the unregularized and regularized hyperprior models to \\(4\\times 10^{6}\\) steps and demonstrate the effect of the regularization factor \\(\\alpha\\) on the performance. Herein, with extended training, the regularization consistently enhances optimization. The best performance, with \\(\\alpha = 1\\), is achieved at the \\(2.5\\times 10^{6}\\) step, showing a BD-Rate of \\(-1.24\\%\\). As training progresses to \\(4\\times 10^{6}\\) steps, the gain reduces to around \\(0.7\\%\\). For the effect of hyperparameter \\(\\alpha\\), it is shown that either the value is too large or too small will not benefit the optimization, which can be explained as follows. As discussed in Sec. 3.2, our \\(\\alpha H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})\\) maximization regularization is a bound upper of the ground-truth objective \\(\\alpha (H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}}) - H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}}))\\). When \\(\\alpha\\) is large, this \\(H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})\\) gap is also amplified, which may harm the regularization accuracy and thus the optimization performance. While one might consider addressing this through additional \\(H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})\\) estimation using Algorithm 1, such an approach proves generally ineffective. The challenge lies in accurately learning a distribution \\(q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})\\) that maps \\(\\hat{\\boldsymbol{X}}\\) to \\(\\boldsymbol{U}\\), as any discrepancy between \\(q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})\\) and the true \\(p(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})\\) would itself reduce accuracy. Furthermore, the incurred computational overhead is unfavorable. We provide more results in Supplementary Material (Sec. 10). When \\(\\alpha\\) is small, the regularizer has little contribution to the overall loss, and the optimization is reduced to the vanilla objective." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15579" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.357, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.089, + 0.621, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.089, + 0.895, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.905, + 0.281 + ], + "angle": 0, + "content": "Figure 6. (a) Effects of regularization factor \\(\\alpha\\). (b) Regularization performance for the hyperprior compression model with the autoregressive regularizer; (c) the autoregressive compression model with the hyperprior regularizer." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.307, + 0.485, + 0.745 + ], + "angle": 0, + "content": "On the alignment of the entropy models. In our experimental setup, we emphasize the advantages of aligned entropy modeling between latent and source distributions. Herein, we extend the analysis to explore cases where the entropy models are unaligned, investigating two scenarios: a compression model equipped with (a) a stronger source entropy model and (b) a weaker source entropy model. In particular, for the source models of the hyperprior and autoregressive depicted in Fig.3(a) and Fig.3(b), respectively, we consider the autoregressive entropy model with additional spatial contexts to be a more accurate and thus \"stronger\" model. Thereafter, we train the hyperprior compression model with a autoregressive source regularizer and the autoregressive compression model with a hyperprior source regularizer, and the results are depicted in Fig.6(b) and Fig.6(c), respectively. The values of \\(\\alpha\\) are still set to 0.1 and 1 for the hyperprior and autoregressive regularizers, respectively. The performance of the aligned entropy model is also included for reference. Herein, it can be observed that neither a stronger nor a weaker source entropy model consistently facilitates the optimization performance of the compression models. Specifically, when a stronger autoregressive regularizer is imposed on the hyperprior network, the average BD-Rate remains positive throughout training. When a weaker hyperprior regularizer is employed by the autoregressive network, the optimization is temporarily improved at the \\(1.5 \\times 10^{6}\\) training step, with a BD-Rate of \\(-0.20\\%\\). In contrast, aligning the entropy models leads to significant performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.218, + 0.772 + ], + "angle": 0, + "content": "4.3. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.906 + ], + "angle": 0, + "content": "Our regularization method is applied exclusively during the network training stage and therefore introduces no additional inference complexity. However, since it involves training an additional source entropy model \\( q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}}) \\), the training complexity is inevitably increased. Assuming the complexities of the source and latent entropy models are approximately equal, the complexity overhead ratio can be calculated as \\( \\frac{C_L}{C_L + C_T} = \\frac{1}{1 + C_T / C_L} \\), which is closely related" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.908, + 0.534 + ], + "angle": 0, + "content": "to the complexity ratio between the transform backbone and the latent, i.e., \\( C_T / C_L \\). For \\( \\lambda = 0.0018 \\), we train \\( 1 \\times 10^5 \\) steps for each model on the same machine (Nvidia RTX 4090) and compute the time ratio. As shown in Table 2, compared to the unregularized baseline, the training time of our proposed method increases by \\( 28.8\\% \\), \\( 18.2\\% \\), \\( 8.1\\% \\), \\( 39.3\\% \\), and \\( 29.8\\% \\) for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models, respectively. Notably, for the attention model, the overhead is the least significant. This can be attributed to the fact that the attention-based transform backbone is significantly more complex than the entropy model. To further reduce training complexity, one can employ lightweight convolutional layers, such as depth-wise separable convolution [56], or improve the iterative training strategy outlined in Algorithm 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.544, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Table 2. Training complexity of the proposed regularizer compared to the unregularized anchor." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "
hyperpriorautoregressiveattentionELICMLIC++
Training time128.8%118.2%108.1%139.3%129.8%
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper investigates a structural entropy regularizer for lossy neural image compression. Specifically, motivated by the information-theoretic analysis of compression models, we demonstrate that, in addition to minimizing latent entropy, maximizing conditional source entropy can be equally important. Subsequently, by regularizing the neural networks to simultaneously maximize the conditional source entropy during end-to-end optimization, we experimentally observe improvements in both the in-domain compact representation and the out-of-domain generalization abilities of compression models. Additionally, as the primary overhead of this regularizer comes from training an additional source entropy model, the proposed method imposes no added inference complexity, opening up a new space for regularizing neural compression networks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15580" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.157 + ], + "angle": 0, + "content": "[1] C. E. Shannon et al., \"Coding theorems for a discrete source with a fidelity criterion,\" IRE Nat. Conv. Rec, vol. 4, no. 142-163, p. 1, 1959. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.16, + 0.483, + 0.187 + ], + "angle": 0, + "content": "[2] K. Sayood, Introduction to data compression. Morgan Kaufmann, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.19, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[3] Y. Yang, S. Mandt, L. Theis, et al., \"An introduction to neural data compression,\" Foundations and Trends® in Computer Graphics and Vision, vol. 15, no. 2, pp. 113-200, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.235, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[4] J. Balle, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, \"Variational image compression with a scale hyperprior,\" arXiv preprint arXiv:1802.01436, 2018. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.293, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[5] T. Berger, “Rate-distortion theory,” Wiley Encyclopedia of Telecommunications, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.323, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[6] V. K. Goyal, “Theoretical foundations of transform coding,” IEEE Signal Processing Magazine, vol. 18, no. 5, pp. 9–21, 2001. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.367, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[7] J. Balle, V. Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” arXiv preprint arXiv:1611.01704, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.483, + 0.466 + ], + "angle": 0, + "content": "[8] D. Minnen, J. Balle, and G. D. Toderici, \"Joint autoregressive and hierarchical priors for learned image compression,\" Advances in neural information processing systems, vol. 31, 2018. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[9] D. Minnen and S. Singh, \"Channel-wise autoregressive entropy models for learned image compression,\" in 2020 IEEE International Conference on Image Processing (ICIP), pp. 3339-3343, IEEE, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.528, + 0.483, + 0.597 + ], + "angle": 0, + "content": "[10] Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, \"Learned image compression with discretized gaussian mixture likelihoods and attention modules,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7939-7948, 2020. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.6, + 0.483, + 0.667 + ], + "angle": 0, + "content": "[11] D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin, \"Checkerboard context model for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14771-14780, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.671, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[12] W. Jiang and R. Wang, \"MLIC++: Linear complexity multi-reference entropy modeling for learned image compression,\" in ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[13] S. Qin, J. Wang, Y. Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y. Wang, \"MambaVC: Learned visual compression with selective state spaces,\" arXiv preprint arXiv:2405.15413, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[14] H. Zhang, L. Li, and D. Liu, \"On uniform scalar quantization for learned image compression,\" arXiv preprint arXiv:2309.17051, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[15] E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Timofte, L. Benini, and L. V. Gool, \"Soft-to-hard vector quantization for end-to-end learning compressible representations,\" Advances in neural information processing systems, vol. 30, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[16] Y. Yang, R. Bamler, and S. Mandt, \"Improving inference for neural image compression,\" Advances in Neural Information Processing Systems, vol. 33, pp. 573-584, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[17] E. Agustsson and L. Theis, \"Universally quantized neural compression,\" Advances in neural information processing systems, vol. 33, pp. 12367-12376, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.906, + 0.235 + ], + "angle": 0, + "content": "[18] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Soft then hard: Rethinking the quantization in neural image compression,\" in International Conference on Machine Learning, pp. 3920-3929, PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.906, + 0.279 + ], + "angle": 0, + "content": "[19] D. P. Kingma, M. Welling, et al., \"An introduction to variational autoencoders,\" Foundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307-392, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.282, + 0.906, + 0.336 + ], + "angle": 0, + "content": "[20] J. Zhao, Y. Kim, K. Zhang, A. Rush, and Y. LeCun, \"Adversarily regularized autoencoders,\" in International conference on machine learning, pp. 5902-5911, PMLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.906, + 0.393 + ], + "angle": 0, + "content": "[21] H. Xu, D. Luo, R. Henao, S. Shah, and L. Carin, \"Learning autoencoders with relational regularization,\" in International Conference on Machine Learning, pp. 10576-10586, PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.397, + 0.906, + 0.45 + ], + "angle": 0, + "content": "[22] H. Wu and M. Flierl, \"Vector quantization-based regularization for autoencoders,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 6380-6387, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.906, + 0.509 + ], + "angle": 0, + "content": "[23] L. Le, A. Patterson, and M. White, \"Supervised autoencoders: Improving generalization performance with unsupervised regularizers,\" Advances in neural information processing systems, vol. 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.906, + 0.567 + ], + "angle": 0, + "content": "[24] T. Ma, J. Chen, and C. Xiao, \"Constrained generation of semantically valid graphs via regularizing variational autoencoders,\" Advances in Neural Information Processing Systems, vol. 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.906, + 0.611 + ], + "angle": 0, + "content": "[25] S. Sinha and A. B. Dieng, \"Consistency regularization for variational auto-encoders,\" Advances in Neural Information Processing Systems, vol. 34, pp. 12943–12954, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[26] H. Ochoa-Dominguez and K. R. Rao, Discrete cosine transform. CRC Press, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.644, + 0.906, + 0.698 + ], + "angle": 0, + "content": "[27] G. Toderici, S. M. O'Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, \"Variable rate image compression with recurrent neural networks,\" arXiv preprint arXiv:1511.06085, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.701, + 0.906, + 0.742 + ], + "angle": 0, + "content": "[28] H. Liu, T. Chen, P. Guo, Q. Shen, X. Cao, Y. Wang, and Z. Ma, “Non-local attention optimized deep image compression,” arXiv preprint arXiv:1904.09757, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.906, + 0.785 + ], + "angle": 0, + "content": "[29] M. Lu, P. Guo, H. Shi, C. Cao, and Z. Ma, \"Transformer-based image compression,\" arXiv preprint arXiv:2111.06707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.788, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[30] J. Liu, H. Sun, and J. Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14388–14397, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[31] H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, \"End-to-end optimized versatile image compression with wavelet-like transform,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1247-1263, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15581" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[32] X. Zhang and X. Wu, \"LVQAC: Lattice vector quantization coupled with spatially adaptive companding for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10239-10248, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[33] R. Feng, Z. Guo, W. Li, and Z. Chen, \"NVTC: Nonlinear vector transform coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6101-6110, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[34] K. Suhring, M. Schäfer, J. Pfaff, H. Schwarz, D. Marpe, and T. Wiegand, \"Trellis-coded quantization for end-to-end learned image compression,\" in 2022 IEEE International Conference on Image Processing (ICIP), pp. 3306-3310, IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.287, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[35] Z. Ge, S. Ma, W. Gao, J. Pan, and C. Jia, \"NLIC: Nonuniform quantization based learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[36] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Causal contextual prediction for learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2329-2341, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[37] X. Zhu, J. Song, L. Gao, F. Zheng, and H. T. Shen, \"Unified multivariate gaussian mixture for efficient neural image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17612-17621, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[38] H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, \"Learned image compression with gaussian-laplacian-logistic mixture model and concatenated residual modules,\" IEEE Transactions on Image Processing, vol. 32, pp. 2063-2076, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[39] Y. Hu, W. Yang, and J. Liu, \"Coarse-to-fine hyper-prior modeling for learned image compression,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 11013-11020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[40] L. Theis, W. Shi, A. Cunningham, and F. Huszár, \"Lossy image compression with compressive autoencoders,\" arXiv preprint arXiv:1703.00395, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.365, + 0.65 + ], + "angle": 0, + "content": "[41] I. Goodfellow, \"Deep learning,\" 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[42] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, \"Generative adversarial nets,\" Advances in neural information processing systems, vol. 27, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[43] D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang, \"ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5718-5727, 2022. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[44] J. Bégaint, F. Racapé, S. Feltman, and A. Pushparaja, \"Compressai: a pytorch library and evaluation platform for end-to-end compression research,\" arXiv preprint arXiv:2011.03029, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[45] W. Jiang, “Unofficial elic.” https://github.com/JiangWeibeta/ELIC, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[46] J. Liu, G. Lu, Z. Hu, and D. Xu, \"A unified end-to-end framework for efficient deep image compression,\" arXiv preprint arXiv:2002.03370, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[47] D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[48] E. Kodak, \"Kodak lossless true color image suite (photocd pcd0992),\" URL http://rOk.us/graphics/kodak, vol. 6, p. 2, 1993. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[49] \"6th Challenge on Learned Image Compression.\" https://www.compression.cc/tasks/index.html. Accessed: 2024-11-13. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[50] N. Asuni and A. Giachetti, \"Testimages: a large-scale archive for testing visual devices and basic image processing algorithms,\" in STAG, pp. 63-70, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[51] T. Zhang, H. Zhang, Y. Li, L. Li, and D. Liu, “Few-shot domain adaptation for learned image compression,” arXiv preprint arXiv:2409.11111, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.293, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[52] Y. Lv, J. Xiang, J. Zhang, W. Yang, X. Han, and W. Yang, \"Dynamic low-rank instance adaptation for universal neural image compression,\" in Proceedings of the 31st ACM International Conference on Multimedia, pp. 632-642, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.349, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[53] J. Yang, S. Shen, H. Yue, and K. Li, \"Implicit transformer network for screen content image continuous superresolution,\" Advances in Neural Information Processing Systems, vol. 34, pp. 13304-13315, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.407, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[54] X. Min, K. Ma, K. Gu, G. Zhai, Z. Wang, and W. Lin, \"Unified blind quality assessment of compressed natural, graphic, and screen content images,\" IEEE Transactions on Image Processing, vol. 26, no. 11, pp. 5462-5474, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.463, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[55] N. Brancati, A. M. Anniciello, P. Pati, D. Riccio, G. Scognamiglio, G. Jaume, G. De Pietro, M. Di Bonito, A. Fon-cubierta, G. Botti, et al., \"BRACS: A dataset for breast carcinoma subtyping in h&e histology images,\" Database, vol. 2022, p. baac093, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.533, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[56] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1251–1258, 2017. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.589, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[57] G. Bjontegaard, \"Calculation of average PSNR differences between RD-curves,\" ITU SG16 Doc. VCEG-M33, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15582" + } + ] +] \ No newline at end of file diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_origin.pdf b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d2c80f0bc128bb10515b5918687aa702c1dd184 --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/cb73e4dd-eb4f-4fff-b4a2-c49c8fa70b5b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9da0632fa05383fccd49a3fd6461ce2a1658c1bef0f10c74ef24db3ab056c1d +size 4975462 diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/full.md b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ded41f7870fde9364f568e49a86b5ba407244d7a --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/full.md @@ -0,0 +1,307 @@ +# An Information-Theoretic Regularizer for Lossy Neural Image Compression + +Yingwen Zhang $^{1}$ , Meng Wang $^{2*}$ , Xihua Sheng $^{1}$ , Peilin Chen $^{1}$ , Junru Li $^{3}$ , Li Zhang $^{3}$ , Shiqi Wang $^{1}$ + +$^{1}$ Department of Computer Science, City University of Hong Kong, Hong Kong, China $^{2}$ School of Data Science, Lingnan University, Hong Kong, China $^{3}$ Bytedance Inc., San Diego, USA + +# Abstract + +Lossy image compression networks aim to minimize the latent entropy of images while adhering to specific distortion constraints. However, optimizing the neural network can be challenging due to its nature of learning quantized latent representations. In this paper, our key finding is that minimizing the latent entropy is, to some extent, equivalent to maximizing the conditional source entropy, an insight that is deeply rooted in information-theoretic equalities. Building on this insight, we propose a novel structural regularization method for the neural image compression task by incorporating the negative conditional source entropy into the training objective, such that both the optimization efficacy and the model's generalization ability can be promoted. The proposed information-theoretic regularizer is interpretable, plug-and-play, and imposes no inference overheads. Extensive experiments demonstrate its superiority in regularizing the models and further squeezing bits from the latent representation across various compression structures and unseen domains. Code is available at Github. + +# 1. Introduction + +Lossy image compression, or more broadly, lossy data compression, is grounded in a fundamental theorem in information theory: the rate-distortion theorem [1]. Essentially, this theorem establishes a theoretical lower bound for lossy compression and provides valuable guidance for designing practical compression algorithms [2, 3]. + +A basic lossy image compression model is illustrated in Fig.1(a). Let $\mathbf{X}$ , $\hat{\mathbf{X}}$ , and $\mathbf{U}$ denote the source, the reconstruction, and the discrete index (or the latent), respectively. The encoder is a deterministic mapping $\mathrm{Q} \colon \mathcal{X}^N \to \mathcal{U}$ where $\mathcal{X}^N$ denotes a $N$ -dimensional image space and $\mathcal{U} = \{1, 2, \dots, M\}$ denotes the discrete index space. The decoder + +![](images/a2653bdbcd142fbec9a2e1c5551274f1d4e6a38adade3615fe13c7a42b36c7e9.jpg) + +![](images/ea905594142a6088fcd23c8c6a4c9f48b1d979f342a9aaaf4f34258c68f18706.jpg) + +![](images/db1c111c8e4e030aba4d06968a12c94ccd1998e5d7520a827623981b5458bfa4.jpg) +(c) +Figure 1. (a) Direct coding model; (b) Information diagram for the direct coding, wherein $H(\mathbf{X}) = H(\mathbf{X}|\hat{\mathbf{X}}) + H(\mathbf{U})$ and $H(\mathbf{X})$ is fixed for any known source; (c) Transform coding model (with possible side information $\hat{\mathbf{Z}}$ [4]). + +is a bijection mapping $\mathrm{Q}^{-1} \colon \mathcal{U} \to \hat{\mathcal{X}}^N$ , where the reconstruction space $\hat{\mathcal{X}}^N = \{c_1, \dots, c_M\}$ forms a $N$ -dimensional codebook. For index $U$ , entropy coding is applied to losslessly convert the data into the bitstream, which can then be decoded back to $U$ on the decoder side. We may refer to such a model as the direct coding model for its structural simplicity. Subsequently, following the rate-distortion theorem [1], given a bounded distortion measure $d(\cdot)$ and a distortion constraint $D$ , the rate of transmitting $U$ is lower bounded by the informational rate-distortion function: + +$$ +R (D) = \min _ {p _ {\hat {\boldsymbol {X}} | \boldsymbol {X}}: \mathbb {E} _ {p _ {\boldsymbol {X}} p _ {\hat {\boldsymbol {X}} | \boldsymbol {X}}} [ d (\boldsymbol {X}, \hat {\boldsymbol {X}}) ] < D} \frac {1}{N} I (\boldsymbol {X}; \hat {\boldsymbol {X}}), \tag {1} +$$ + +i.e., the minimum mutual information obtained by traversing all the possible encoder-decoder (or quantizer) designs $p_{\hat{X} |X}$ . Theoretically, this bound can be achieved asymptotically using vector quantization as $N \to \infty$ [5]. However, since the encoding complexity grows exponentially with the image dimension $N$ , a suboptimal yet more practical solution is the transform coding [6]. As depicted in + +Fig.1(c), for the transform coding model, a deterministic analysis transform $\mathrm{T_A}$ : $\mathcal{X}^N \to \mathcal{Y}^M$ and a deterministic synthesis transform $\mathrm{T_S}$ : $\hat{\mathcal{Y}}^M \to \hat{\mathcal{X}}^N$ are incorporated into the compression chain. The analysis transform is designed to decorrelate the source $X$ , such that a much simpler quantization scheme, e.g., uniform scalar quantization: $\boldsymbol{u}_i = \mathrm{round}(\boldsymbol{y}_i)$ , can be independently applied to each dimension of the transformed coefficients $Y$ . Building on this framework and leveraging the superior nonlinear capabilities of neural networks, numerous neural image compression networks [4, 7-13] are proposed, continually pushing the frontier of lossy image compression. + +Typically, by parameterizing the $\mathrm{T_A}$ and $\mathrm{T_S}$ in Fig.1(c) as neural networks, the rate-distortion training objective is formulated as [7]: + +$$ +\min _ {R \approx H (U)} \underbrace {\mathbb {E} _ {\boldsymbol {X}} [ - \log q _ {\phi} (\boldsymbol {U}) ]} + \lambda \underbrace {\mathbb {E} _ {\boldsymbol {X}} | | \boldsymbol {X} - \hat {\boldsymbol {X}} | | _ {2} ^ {2}} _ {D}, \tag {2} +$$ + +where $R$ is the estimation of latent entropy $H(\pmb{U})$ ; $q_{\phi}$ is the neural entropy model (or estimated distribution) for $\pmb{U}$ , whose modeling accuracy can be significantly enhanced by transmitting a small amount of side information $\hat{\pmb{Z}}$ (as depicted in Fig.1(c)) as part of the parameters $\phi$ [4]; the distortion $D$ is the $l_{2}$ distance between $\pmb{X}$ and $\hat{\pmb{X}}$ ; $\lambda$ is the Lagrangian multiplier, controlling the trade-off between $R$ and $D$ . Herein, due to the existence of quantization $\mathrm{Q}(\cdot)$ , $\pmb{U}$ is discrete in essence. Optimizing the network is then challenging as the gradient is almost zero everywhere for $\mathrm{Q}(\cdot)$ . To effectively train the networks, a surrogate is introduced during training, e.g., additive uniform noise (AUN) [7]: + +$$ +\boldsymbol {U} = \mathrm {Q} (\boldsymbol {Y}) \approx \boldsymbol {Y} + [ - 0. 5, 0. 5 ] ^ {M}, \tag {3} +$$ + +such that the backward gradient of $\mathrm{Q}(\cdot)$ is one. However, since such approximation inevitably results in the train-test mismatch, and thus high gradient estimation error [14], both the efficiency and efficacy of the optimization are hindered. To address this, many research efforts are devoted to designing advanced surrogates and optimization methods [14-18]. + +In this paper, we present a completely distinct view on facilitating the optimization of the compression network: regularization. Conceivably, imposing structural constraints on the network optimization can effectively regularize the gradient estimation error, promote convergence, and reduce overfitting. In the context of generative variational autoencoders [19], structural regularization techniques [20-25] have been effectively applied to improve authentic data generation. However, since neural compression networks fundamentally differ from generative autoencoders in terms of the minimum latent entropy constraint, to the best of our knowledge, no structural regularization methods tailored to this constraint have been explored. In light of this, in this + +paper, we present the first study of information-theoretic regularization for the neural image compression task. In particular, through an in-depth information-theoretic analysis of compression models, we reveal that minimizing the latent entropy $H(\mathbf{U})$ is, to some extent, equally important as maximizing the conditional source entropy $H(\mathbf{X}|\hat{\mathbf{X}})$ . Subsequently, by enforcing the network to simultaneously minimize $-H(\mathbf{X}|\hat{\mathbf{X}})$ , both the in-domain compact representation and out-of-domain generalization ability are experimentally improved. + +# 2. Related works + +Neural image compression networks. Neural image compression networks are primarily composed of transform, quantization, and latent entropy models. The transform module maps the input source into a latent representation so that the dependencies of the source domain can be more easily handled. Unlike the linear transform [26] used in traditional hybrid coding, the neural transform is inherently nonlinear, enabling more effective decorrelation of the source data. In the literature, various neural transform architectures are proposed. Representative architectures include the recurrent neural network (RNN) [27], convolutional neural network (CNN) [7], non-local attention [28], transformer [29], mixed transformer-CNN [30], state-space models [13] and invertible networks [31]. Subsequently, the continuous transformed coefficients are quantized into the discrete representation, wherein the uniform scalar quantization [7] is commonly adopted. As vector quantization can, in principle, better approach the rate-distortion bound (Eqn.(1)), low-complexity vector quantization [15, 32, 33] and dependent scalar quantization [34, 35] schemes are also studied. Finally, once the quantized latent representations are obtained, a neural entropy model is essential for performing entropy coding. Herein, since the latent cannot be perfectly decorrelated, an accurate entropy model that can better capture the dependencies and closely match the ground-truth latent distribution is crucial for the overall network performance. Prominent research directions such as spatial-channel context mining [8, 9, 12, 36], expressive latent distribution modeling [10, 37, 38] and latent architecture designs [4, 11, 39] are explored. + +Optimization over quantized latent representations. The surrogate-based methods [7, 14-17] have been shown to effectively address the non-differentiable quantization function during end-to-end optimization. In addition to AUN [7], various quantization surrogates such as soft assignment [15], Gumbel-softmax trick [16], and soft rounding function [17] have been studied. However, as the surrogate introduces a train-test mismatch in forward calculations, the annealing techniques [14-17] are then proposed to + +control the "softness" of surrogates, gradually approximating the hard quantization function as training progresses. The straight-through estimator (STE) [40] provides an alternative approach to address the train-test mismatch by enforcing an identity mapping during the backward calculation, thereby enabling direct hard quantization in the forward pass. Nevertheless, STE ultimately introduces a significant gradient estimation bias for the analysis transform. To counter this, an additional fine-tuning stage for the entropy model and synthesis transform is suggested [18]. + +While numerous network optimization methods have been developed from a quantization perspective, the regularization view has largely escaped research attention. To the best of our knowledge, the most related efforts are regularization methods for variational autoencoders (VAEs), such as regularization on latent priors [20-22], reconstruction [23, 24], and consistency [25]. However, as VAEs are designed for generative modeling, these methods are not applicable to optimizing compression networks. To bridge this knowledge gap, this paper proposes a structural regularizer for neural image compression and conducts extensive experiments to demonstrate its superiority. + +# 3. Methodology + +In this section, we begin by deriving the equivalent maximization form of the latent entropy minimization problem for both the direct and transform coding models. Building on this derivation, we propose a novel regularization method for optimizing the neural compression network. Note that during our derivations, we use the naive transform coding model without the side information branch (i.e., the $\hat{Z}$ depicted in Fig 1(c)). This significantly simplifies our analysis without compromising much practicality, as it has been shown that the information flowing into the side branch is negligible compared to the latent branch [4]. + +# 3.1. Motivation: from minimization to maximization + +Lemma 1. Considering a deterministic quantization process $Q(\cdot)$ and a deterministic dequantization process $Q^{-1}(\cdot)$ for the direct coding model illustrated in Fig. 1(a), the following equalities hold automatically: + +$$ +\left\{ \begin{array}{l l} I (\boldsymbol {X}; \hat {\boldsymbol {X}}) = H (\hat {\boldsymbol {X}}) & (4) \\ H (\hat {\boldsymbol {X}}) = H (\boldsymbol {U}) & (5) \\ H (\boldsymbol {U}) = I (\boldsymbol {X}; \hat {\boldsymbol {X}}) & (6) \end{array} \right. +$$ + +Proof. Eqn.(4) follows the facts that $H(\hat{\mathbf{X}}|\mathbf{X}) = 0$ , meaning there is no uncertainty in $\hat{\mathbf{X}}$ given the observation of $\mathbf{X}$ , and $I(\mathbf{X};\hat{\mathbf{X}}) = H(\hat{\mathbf{X}}) - H(\hat{\mathbf{X}}|\mathbf{X})$ . Similarly, Eqn.(5) follows the fact that $H(\mathbf{U}|\hat{\mathbf{X}}) = H(\hat{\mathbf{X}}|\mathbf{U}) = 0$ and $H(\hat{\mathbf{X}}) + H(\mathbf{U}|\hat{\mathbf{X}}) = H(\mathbf{U}) + H(\hat{\mathbf{X}}|\mathbf{U})$ . Eqn.(6) is imme + +dately apparent from Eqn.(4) and Eqn.(5). Additional details are provided in the Supplementary Material (Sec. 6). + +Theorem 1. For the direct coding model elaborated in Lemma 1, we have + +$$ +H (\boldsymbol {U}) = H (\boldsymbol {X}) - H (\boldsymbol {X} | \hat {\boldsymbol {X}}). \tag {7} +$$ + +The informational diagram is thus illustrated as Fig.1(b), given the non-negativity of discrete entropy $H(\cdot)$ . + +Proof. This is immediately apparent from Eqn.(6), as the right-hand-side of Eqn.(7) is by definition $I(\pmb {X};\hat{\pmb{X}})$ + +From Theorem 1, for the direct coding model, it is clear that minimizing the latent entropy $H(\mathbf{U})$ can be equivalently achieved by maximizing the conditional source entropy $H(\mathbf{X}|\hat{\mathbf{X}})$ , for any known source $\mathbf{X}$ . In practice, this aligns with the training setting of the neural compression network, wherein the training set is given in advance, and thus $H(\mathbf{X})$ is fixed. However, as the prevailing neural compression networks are based on the transform coding model, in the following discussion, we extend this conclusion to the transform coding setting. + +Lemma 2. Considering a transform coding model built on the direct coding model elaborated in Lemma 1 with a deterministic analysis transform $T_{A}$ and a deterministic synthesis transform $T_{S}$ (Fig. 1(c)), the following equalities hold automatically: + +$$ +\left\{ \begin{array}{l} I (\boldsymbol {X}; \hat {\boldsymbol {X}}) = H (\hat {\boldsymbol {X}}) \\ H (\hat {\boldsymbol {X}}) = H (\boldsymbol {U}) - H (\boldsymbol {U} | \hat {\boldsymbol {X}}) \\ H (\boldsymbol {U}) = I (\boldsymbol {X}; \hat {\boldsymbol {X}}) + H (\boldsymbol {U} | \hat {\boldsymbol {X}}) \end{array} \right. \tag {8} +$$ + +Proof. Compared to the proof of Lemma 1, for the transform coding, we still have $H(\hat{\mathbf{X}}|\mathbf{X}) = 0$ and $H(\hat{\mathbf{X}}|\mathbf{U}) = 0$ . However, since the synthesis transform $T_{S}$ is not guaranteed to be a bijection mapping, the $H(\mathbf{U}|\hat{\mathbf{X}}) = 0$ does not always hold, leading to the correction of Eqn.(9) and thus Eqn.(10). For more details, readers may refer to the Supplementary Material (Sec. 7). + +Theorem 2. For the transform coding model elaborated in Lemma 2, we have + +$$ +H (\boldsymbol {U}) = H (\boldsymbol {X}) - \left(H (\boldsymbol {X} | \hat {\boldsymbol {X}}) - H (\boldsymbol {U} | \hat {\boldsymbol {X}})\right). \tag {11} +$$ + +Proof. This is immediately apparent from Eqn.(10). + +Remarks. Two conclusions can be drawn for the transform coding model. + +(a) Equivalence of minimization and maximization. From Theorem 2, for any known source $\mathbf{X}$ , minimizing $H(\mathbf{U})$ can be equivalently achieved by maximizing the $H(\mathbf{X}|\hat{\mathbf{X}}) - H(\mathbf{U}|\hat{\mathbf{X}})$ . For a bijection synthesis function $T_{S}$ , of which $H(\mathbf{U}|\hat{\mathbf{X}})$ is zero, this maximization objective is reduced to $H(\mathbf{X}|\hat{\mathbf{X}})$ , which is the same as the direct coding; +(b) Achieving $R(D)$ by minimizing $H(\pmb{U})$ . From Eqn.(10), minimizing $H(\pmb{U})$ is simultaneously minimizing $I(\pmb{X};\hat{\pmb{X}})$ and $H(\pmb{U}|\hat{\pmb{X}})$ . In the ideal case, the informational lower bound given in Eqn.(1) is achieved with $H(\pmb{U}|\hat{\pmb{X}})$ being reduced to 0. + +![](images/6bcb1680f39bd2d58e0c52077b769a750e0c7f366f92710bb39426740177ba1a.jpg) +Figure 2. An illustration of the proposed regularization method, wherein an additional source entropy model $q_{\theta}$ is introduced. The AE and AD represent arithmetic encoding and arithmetic decoding, respectively. The side information branch (i.e., the $\hat{Z}$ branch) is considered as part of the latent entropy model, which necessitates bits transmission. + +# Algorithm 1 Training Strategy + +1: for each training step do +2: Sample a minibatch of training images. +3: Perform the forward pass of the compression network and source entropy model $q_{\theta}$ . +4: Compute the backward gradient and update the compression network based on Eqn.(12). +5: Freeze the compression network. +6: Perform the forward pass of $q_{\theta}$ . +7: Compute the backward gradient and update the source entropy model $q_{\theta}$ based on Eqn.(13). +8: end for + +# 3.2. Information-theoretic regularizer + +In principle, optimizing neural compression networks with a quantization surrogate, e.g., AUN, introduces gradient estimation errors [14], thereby deviating the objective of minimizing the latent entropy $H(\mathbf{U})$ under certain distortion constraints (Eqn.(2)), harming both the efficiency and efficacy of the optimization. Motivated by the equivalence of the minimization and maximization optimization discussed + +in Sec. 3.1, we propose to address this issue by imposing structural maximization constraints on the optimization process. As such, the vanilla objective of minimizing $H(\mathbf{U})$ can be further enhanced through this maximization regularization. + +From Theorem 2, an ideal gradient direction for model updating is one that simultaneously maximizes $H(\mathbf{X}|\hat{\mathbf{X}}) - H(\mathbf{U}|\hat{\mathbf{X}})$ , which can be further interpreted as maximizing $H(\mathbf{X}|\hat{\mathbf{X}})$ while minimizing $H(\mathbf{U}|\hat{\mathbf{X}})$ . To integrate this regularization into the training objective, two parameterized distributions $q_{\theta}(\mathbf{X}|\hat{\mathbf{X}})$ and $q_{\varphi}(\mathbf{U}|\hat{\mathbf{X}})$ are necessitated. However, accurately modeling an entropy model $q_{\varphi}(\mathbf{U}|\hat{\mathbf{X}})$ alongside the compression network is quite challenging, as it involves learning another "encoder" function from the reconstruction $\hat{\mathbf{X}}$ to the latent $\mathbf{U}$ . We experimentally find that a poorly learned $H(\mathbf{U}|\hat{\mathbf{X}})$ regularizer may harm the optimization performance, and that the increased training complexity is also burdensome. To balance practical feasibility with regularization accuracy, we include only the first $H(\mathbf{X}|\hat{\mathbf{X}})$ term for regularization, which is an upper bound for the ground-truth $H(\mathbf{X}|\hat{\mathbf{X}}) - H(\mathbf{U}|\hat{\mathbf{X}})$ objective. The entropy modeling becomes more tractable since the reconstruction $\hat{\mathbf{X}}$ is closely aligned with the source $\mathbf{X}$ . The primary rationale for such simplification is that although maximizing $H(\mathbf{X}|\hat{\mathbf{X}})$ alone is less precise than $H(\mathbf{X}|\hat{\mathbf{X}}) - H(\mathbf{U}|\hat{\mathbf{X}})$ , it still provides a meaningful constraint on the "noisy" gradient, yielding better gradient direction than no regularization. Another justification is that since we will retain the $H(\mathbf{U})$ minimization term in the overall objective, this $H(\mathbf{U}|\hat{\mathbf{X}})$ gap, in principle (Eqn.(11)), should decrease as training progresses, which ensures that the regularization effect remains aligned with the ideal maximization objective. The resulting training objective is thus given by: + +$$ +\min R + \lambda D + \alpha \underbrace {\mathbb {E} _ {\boldsymbol {X}} \left[ \log q _ {\theta} (\boldsymbol {X} | \hat {\boldsymbol {X}}) \right]} _ {\approx - H (\boldsymbol {X} | \hat {\boldsymbol {X}})}, \tag {12} +$$ + +where $R + \lambda D$ follows the rate-distortion loss in Eqn.(2); $-H(\mathbf{X}|\hat{\mathbf{X}})$ is the proposed regularizer; $q_{\theta}(\mathbf{X}|\hat{\mathbf{X}})$ is the estimated source distribution; $\alpha$ is regularization factor, controlling the degree of regularization. An illustration for this regularization method is given in Fig.(2). In addition to the compression network, $q_{\theta}(\mathbf{X}|\hat{\mathbf{X}})$ is introduced, which is parameterized by a neural network and can be optimized via the maximum likelihood criteria [41]: + +$$ +\max \mathbb {E} _ {\boldsymbol {X}} [ \log q _ {\theta} (\boldsymbol {X} | \hat {\boldsymbol {X}}) ]. \tag {13} +$$ + +This results in a max-min optimization problem where minimizing $-H(\boldsymbol{X}|\hat{\boldsymbol{X}})$ (Eqn.(12)) simultaneously creates a conflict. To resolve this, a GAN-style two-stage joint training strategy [42] is employed: At each training step, we first update the compression network using Eqn.(12), then freeze it and update $q_{\theta}(\boldsymbol{X}|\hat{\boldsymbol{X}})$ using Eqn.(13). We summarize the strategy in Algorithm 1. + +![](images/38fa7dc18bec11553ab0f5a84ff182b713062af3a4c1b0221f143fba9f9dc7a4.jpg) + +![](images/9f24ee091835ee48c889f36d3a02d26b1f2aa3a6eaeaf5640f07f9fff9d554a9.jpg) +Figure 3. Conditional source entropy modeling for (a) hyperprior [4], (b) autoregressive [8] and attention [10], (c) EILC [43], and (d) MLIC++ [12]. The architecture and module designs are aligned with the latent entropy model. Additional details regarding the specific modules are provided in the Supplementary Material (Sec. 9). + +![](images/3eaf34fbfde9d337f1fe3961e3024d6e7ef90a76df19d45a15f239c1266fd376.jpg) + +![](images/b93818c47499f3a27ff75f292f5372684cbb4fae645961d619278d4728e2f5a4.jpg) + +# 4. Experiments + +# 4.1. Experimental setup + +Compression models. Three classic image compression models are used in exhaustive validation experiments: the hyperprior context model [4] (with scale and mean [8]), the joint hyperprior and spatial autoregressive context model [8], and the attention-based transform model [10]. We may refer to these models as hyperprior, autoregressive, and attention, respectively. Moreover, we conduct the experiments on two advanced compression models: ELIC [43] and MLIC++ [12]. Notably, ELIC utilizes checkerboard and channel-conditioned latent entropy contexts, while MLIC++ builds upon ELIC by introducing additional attention-based context designs and is identified as one of the state-of-the-art (SOTA) models. + +We train these models from scratch using the vanilla rate-distortion loss (Eqn.(2)), and compare their performance when trained with the proposed regularization method (Algorithm 1) under an equal number of training steps. For the implementations of the hyperprior, autoregressive, and attention models, we adopt CompressAI's re-implementation [44], where the quantization surrogate is ANU. For ELIC and MLIC++, we use the open re-implementations from [45] and the official implementation, respectively, both of which employ a mixed ANU-STE quantization surrogate [9]. + +Conditional source entropy modeling. For each compression model, we apply exactly the same entropy modeling architecture as the latent entropy modeling $q_{\phi}(\boldsymbol{U}|\hat{\boldsymbol{Z}})$ . Our experiments indicate that aligning the entropy model between the latent and source is critical for effective regularization; deviations in either direction (stronger or weaker + +source models) degrade compression performance. For the hyperprior model [4], $q_{\theta}(\boldsymbol{X}|\hat{\boldsymbol{X}})$ is modeled as the factorized Gaussian distribution conditioned on $\hat{\boldsymbol{X}}$ , i.e., + +$$ +q _ {\theta} (\boldsymbol {X} | \hat {\boldsymbol {X}}) \sim \prod_ {i = 1} ^ {N} \mathbb {N} \left(\mu_ {i}, \sigma_ {i} ^ {2}\right), \tag {14} +$$ + +$$ +\text {w h e r e} \Phi \triangleq \left(\mu_ {1}, \dots , \mu_ {N}, \sigma_ {1} ^ {2}, \dots , \sigma_ {N} ^ {2}\right) = f _ {\theta} (\hat {\boldsymbol {X}}), \tag {15} +$$ + +and $f_{\theta}$ is modeled by convolutional layers; $\mathbb{N}$ denotes the Gaussian distribution; $\Phi$ denote the mean and variance; $N$ denotes the dimension of both $X$ and $\hat{X}$ . For both the autoregressive and attention models, the mean and variance are determined by the additional spatial context, i.e., + +$$ +\left(\mu_ {1}, \dots , \mu_ {i}, \sigma_ {1} ^ {2}, \dots , \sigma_ {i} ^ {2}\right) = f _ {\theta} \left(\hat {\boldsymbol {X}}, \boldsymbol {X} _ {< i}\right), \tag {16} +$$ + +where $X_{ 1)$ slice, both the $\Phi_{\text{anchor}}^{k}$ and $\Phi_{\text{non-anchor}}^{k}$ are conditioned on extra channel context from previous slices $X^{Test setRegularized v.s. UnregularizedhyperpriorautoregressiveattentionELICMLIC++Kodak-1.14%-1.57%-0.84%-1.24%-1.82%Tecnick-0.83%-0.78%-0.91%-2.00%-4.01%CLIC-0.66%-0.83%-1.01%-1.45%-3.25%Natural Avg.-0.88%-1.06%-0.92%-1.56%-3.03%Pixel-style-2.20%-1.96%-0.47%-2.71%-3.08%Screen-1.15%-2.01%-0.69%-1.02%-1.72%Game-1.31%-0.83%-0.82%-2.04%-3.44%Pathology-1.11%-0.81%-2.38%-1.82%-0.43% + +![](images/e3ecf35ee4d10ed0c30c11f9ef9557f0deabd766c4b92af07e195f17f833d29b.jpg) + +![](images/511e0d8346e0cbe8a3cbc7bf6b31c9e4ebf19e3179813dd78d538355741921b3.jpg) +Figure 5. Compression performance on Kodak. Our regularizer achieves $-1.13\%$ , $-1.57\%$ , $-0.84\%$ , $-1.24\%$ , and $-1.82\%$ BD-Rates for the hyperprior, autoregressive, attention models, ELIC, and MLIC++, respectively. + +# 4.2. Results + +Performance on natural images. In Fig.4, we present the BD-Rate comparison at four different training steps. The anchor is trained with the vanilla rate-distortion loss (Eqn.(2)) under an equal number of training steps. The regularization factors $\alpha$ for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models are set to 0.1, 1, 1, 0.1, and 0.1, respectively. It can be observed that, for training steps beyond $0.5 \times 10^{6}$ across all five models, our proposed regularization method can achieve better compression performance than the vanilla optimization method. In particular, at the $2 \times 10^{6}$ step, the average BD-Rates for hyperprior, autoregressive, and attention are $-0.88\%$ , $-1.06\%$ and $-0.92\%$ , respectively; at the $1 \times 10^{6}$ step, the average BD-Rates for ELIC and MLIC++ are $-1.56\%$ and $-3.03\%$ , respectively. Meanwhile, the results demonstrate that the regularization effect varies across models as training progresses: For the hyperprior, autoregressive, and MLIC++ models, after a brief convergence period, the BD-Rates are gradually improved from 0 to around $-1\%$ . For the ELIC model, the regularization is most effective during the earlier training stages, achieving over $2.5\%$ BD-Rates saving at the $0.25 \times 10^{6}$ step. For the attention model, the regularization + +achieves a relatively stable coding gain. + +In Fig. 5, we present the detailed rate-distortion curves on the Kodak dataset. For the hyperprior, autoregressive, and attention models, the results correspond to the $2 \times 10^{6}$ training step, while for the ELIC and MLIC++ models, the results are from the $1 \times 10^{6}$ step. The results indicate that the introduction of the additional conditional source entropy regularization leads to a slight reduction in the bit rate for most of the points, although a few exceptions exhibit an increase due to training noise. From the form of regularization in Eqn.(12), i.e., $-\alpha H(\mathbf{X}|\hat{\mathbf{X}})$ , the extent of such bit rate reduction depends on both the regularization factor $\alpha$ and the estimated conditional source entropy $H(\mathbf{X}|\hat{\mathbf{X}})$ . + +Generalizing on unseen domains. In machine learning, regularization plays an essential role in preventing overfitting and enhancing generalization. In Table 1, we directly apply the models trained on natural images to the four unseen domains and evaluate the effectiveness of our regularization method. The results indicate that, across all new domains, regularized models consistently outperform unregularized ones. Performance improvements vary by domain and model, with notable gains in some cases. For instance, in the pathology domain with the attention model, the regularization achieves a BD-Rate of $-2.38\%$ , more than double the improvement observed in the natural domain. + +Effect of $\alpha$ and longer training. In Fig.(6)(a), we continue training both the unregularized and regularized hyperprior models to $4\times 10^{6}$ steps and demonstrate the effect of the regularization factor $\alpha$ on the performance. Herein, with extended training, the regularization consistently enhances optimization. The best performance, with $\alpha = 1$ , is achieved at the $2.5\times 10^{6}$ step, showing a BD-Rate of $-1.24\%$ . As training progresses to $4\times 10^{6}$ steps, the gain reduces to around $0.7\%$ . For the effect of hyperparameter $\alpha$ , it is shown that either the value is too large or too small will not benefit the optimization, which can be explained as follows. As discussed in Sec. 3.2, our $\alpha H(\boldsymbol{X}|\hat{\boldsymbol{X}})$ maximization regularization is a bound upper of the ground-truth objective $\alpha (H(\boldsymbol{X}|\hat{\boldsymbol{X}}) - H(\boldsymbol{U}|\hat{\boldsymbol{X}}))$ . When $\alpha$ is large, this $H(\boldsymbol{U}|\hat{\boldsymbol{X}})$ gap is also amplified, which may harm the regularization accuracy and thus the optimization performance. While one might consider addressing this through additional $H(\boldsymbol{U}|\hat{\boldsymbol{X}})$ estimation using Algorithm 1, such an approach proves generally ineffective. The challenge lies in accurately learning a distribution $q_{\varphi}(\boldsymbol{U}|\hat{\boldsymbol{X}})$ that maps $\hat{\boldsymbol{X}}$ to $\boldsymbol{U}$ , as any discrepancy between $q_{\varphi}(\boldsymbol{U}|\hat{\boldsymbol{X}})$ and the true $p(\boldsymbol{U}|\hat{\boldsymbol{X}})$ would itself reduce accuracy. Furthermore, the incurred computational overhead is unfavorable. We provide more results in Supplementary Material (Sec. 10). When $\alpha$ is small, the regularizer has little contribution to the overall loss, and the optimization is reduced to the vanilla objective. + +![](images/4c62d64e8c6bd727c998b9c6214ec2809de4d450b60475f5af0fb4c2a5d9d25a.jpg) +Figure 6. (a) Effects of regularization factor $\alpha$ . (b) Regularization performance for the hyperprior compression model with the autoregressive regularizer; (c) the autoregressive compression model with the hyperprior regularizer. + +![](images/ef32409542c5519175c4fd7d8c235cfb7e4e72799f45f8ca1a09841292cd1771.jpg) + +![](images/1c2270423262d3ec580e415003fe6a60b2d5d84a4ac6a3b89eead9c14fd1a15f.jpg) + +On the alignment of the entropy models. In our experimental setup, we emphasize the advantages of aligned entropy modeling between latent and source distributions. Herein, we extend the analysis to explore cases where the entropy models are unaligned, investigating two scenarios: a compression model equipped with (a) a stronger source entropy model and (b) a weaker source entropy model. In particular, for the source models of the hyperprior and autoregressive depicted in Fig.3(a) and Fig.3(b), respectively, we consider the autoregressive entropy model with additional spatial contexts to be a more accurate and thus "stronger" model. Thereafter, we train the hyperprior compression model with a autoregressive source regularizer and the autoregressive compression model with a hyperprior source regularizer, and the results are depicted in Fig.6(b) and Fig.6(c), respectively. The values of $\alpha$ are still set to 0.1 and 1 for the hyperprior and autoregressive regularizers, respectively. The performance of the aligned entropy model is also included for reference. Herein, it can be observed that neither a stronger nor a weaker source entropy model consistently facilitates the optimization performance of the compression models. Specifically, when a stronger autoregressive regularizer is imposed on the hyperprior network, the average BD-Rate remains positive throughout training. When a weaker hyperprior regularizer is employed by the autoregressive network, the optimization is temporarily improved at the $1.5 \times 10^{6}$ training step, with a BD-Rate of $-0.20\%$ . In contrast, aligning the entropy models leads to significant performance improvements. + +# 4.3. Limitations + +Our regularization method is applied exclusively during the network training stage and therefore introduces no additional inference complexity. However, since it involves training an additional source entropy model $q_{\theta}(\mathbf{X}|\hat{\mathbf{X}})$ , the training complexity is inevitably increased. Assuming the complexities of the source and latent entropy models are approximately equal, the complexity overhead ratio can be calculated as $\frac{C_L}{C_L + C_T} = \frac{1}{1 + C_T / C_L}$ , which is closely related + +to the complexity ratio between the transform backbone and the latent, i.e., $C_T / C_L$ . For $\lambda = 0.0018$ , we train $1 \times 10^5$ steps for each model on the same machine (Nvidia RTX 4090) and compute the time ratio. As shown in Table 2, compared to the unregularized baseline, the training time of our proposed method increases by $28.8\%$ , $18.2\%$ , $8.1\%$ , $39.3\%$ , and $29.8\%$ for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models, respectively. Notably, for the attention model, the overhead is the least significant. This can be attributed to the fact that the attention-based transform backbone is significantly more complex than the entropy model. To further reduce training complexity, one can employ lightweight convolutional layers, such as depth-wise separable convolution [56], or improve the iterative training strategy outlined in Algorithm 1. + +Table 2. Training complexity of the proposed regularizer compared to the unregularized anchor. + +
hyperpriorautoregressiveattentionELICMLIC++
Training time128.8%118.2%108.1%139.3%129.8%
+ +# 5. Conclusion + +This paper investigates a structural entropy regularizer for lossy neural image compression. Specifically, motivated by the information-theoretic analysis of compression models, we demonstrate that, in addition to minimizing latent entropy, maximizing conditional source entropy can be equally important. Subsequently, by regularizing the neural networks to simultaneously maximize the conditional source entropy during end-to-end optimization, we experimentally observe improvements in both the in-domain compact representation and the out-of-domain generalization abilities of compression models. Additionally, as the primary overhead of this regularizer comes from training an additional source entropy model, the proposed method imposes no added inference complexity, opening up a new space for regularizing neural compression networks. + +# References + +[1] C. E. Shannon et al., "Coding theorems for a discrete source with a fidelity criterion," IRE Nat. Conv. Rec, vol. 4, no. 142-163, p. 1, 1959. 1 +[2] K. Sayood, Introduction to data compression. Morgan Kaufmann, 2017. 1 +[3] Y. Yang, S. Mandt, L. Theis, et al., "An introduction to neural data compression," Foundations and Trends® in Computer Graphics and Vision, vol. 15, no. 2, pp. 113-200, 2023. 1 +[4] J. Balle, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, "Variational image compression with a scale hyperprior," arXiv preprint arXiv:1802.01436, 2018. 1, 2, 3, 5, 6 +[5] T. Berger, “Rate-distortion theory,” Wiley Encyclopedia of Telecommunications, 2003. 1 +[6] V. K. Goyal, “Theoretical foundations of transform coding,” IEEE Signal Processing Magazine, vol. 18, no. 5, pp. 9–21, 2001. 1 +[7] J. Balle, V. Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” arXiv preprint arXiv:1611.01704, 2016. 2 +[8] D. Minnen, J. Balle, and G. D. Toderici, "Joint autoregressive and hierarchical priors for learned image compression," Advances in neural information processing systems, vol. 31, 2018. 2, 5, 6, 1 +[9] D. Minnen and S. Singh, "Channel-wise autoregressive entropy models for learned image compression," in 2020 IEEE International Conference on Image Processing (ICIP), pp. 3339-3343, IEEE, 2020. 2, 5 +[10] Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, "Learned image compression with discretized gaussian mixture likelihoods and attention modules," in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7939-7948, 2020. 2, 5, 6, 1 +[11] D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin, "Checkerboard context model for efficient learned image compression," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14771-14780, 2021. 2 +[12] W. Jiang and R. Wang, "MLIC++: Linear complexity multi-reference entropy modeling for learned image compression," in ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 5, 6 +[13] S. Qin, J. Wang, Y. Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y. Wang, "MambaVC: Learned visual compression with selective state spaces," arXiv preprint arXiv:2405.15413, 2024. 2 +[14] H. Zhang, L. Li, and D. Liu, "On uniform scalar quantization for learned image compression," arXiv preprint arXiv:2309.17051, 2023. 2, 4 +[15] E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Timofte, L. Benini, and L. V. Gool, "Soft-to-hard vector quantization for end-to-end learning compressible representations," Advances in neural information processing systems, vol. 30, 2017. 2 + +[16] Y. Yang, R. Bamler, and S. Mandt, "Improving inference for neural image compression," Advances in Neural Information Processing Systems, vol. 33, pp. 573-584, 2020. 2 +[17] E. Agustsson and L. Theis, "Universally quantized neural compression," Advances in neural information processing systems, vol. 33, pp. 12367-12376, 2020. 2 +[18] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, "Soft then hard: Rethinking the quantization in neural image compression," in International Conference on Machine Learning, pp. 3920-3929, PMLR, 2021. 2, 3 +[19] D. P. Kingma, M. Welling, et al., "An introduction to variational autoencoders," Foundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307-392, 2019. 2 +[20] J. Zhao, Y. Kim, K. Zhang, A. Rush, and Y. LeCun, "Adversarily regularized autoencoders," in International conference on machine learning, pp. 5902-5911, PMLR, 2018. 2, 3 +[21] H. Xu, D. Luo, R. Henao, S. Shah, and L. Carin, "Learning autoencoders with relational regularization," in International Conference on Machine Learning, pp. 10576-10586, PMLR, 2020. +[22] H. Wu and M. Flierl, "Vector quantization-based regularization for autoencoders," in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 6380-6387, 2020. 3 +[23] L. Le, A. Patterson, and M. White, "Supervised autoencoders: Improving generalization performance with unsupervised regularizers," Advances in neural information processing systems, vol. 31, 2018. 3 +[24] T. Ma, J. Chen, and C. Xiao, "Constrained generation of semantically valid graphs via regularizing variational autoencoders," Advances in Neural Information Processing Systems, vol. 31, 2018. 3 +[25] S. Sinha and A. B. Dieng, "Consistency regularization for variational auto-encoders," Advances in Neural Information Processing Systems, vol. 34, pp. 12943–12954, 2021. 2, 3 +[26] H. Ochoa-Dominguez and K. R. Rao, Discrete cosine transform. CRC Press, 2019. 2 +[27] G. Toderici, S. M. O'Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, "Variable rate image compression with recurrent neural networks," arXiv preprint arXiv:1511.06085, 2015. 2 +[28] H. Liu, T. Chen, P. Guo, Q. Shen, X. Cao, Y. Wang, and Z. Ma, “Non-local attention optimized deep image compression,” arXiv preprint arXiv:1904.09757, 2019. 2 +[29] M. Lu, P. Guo, H. Shi, C. Cao, and Z. Ma, "Transformer-based image compression," arXiv preprint arXiv:2111.06707, 2021. 2 +[30] J. Liu, H. Sun, and J. Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14388–14397, 2023. 2 +[31] H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, "End-to-end optimized versatile image compression with wavelet-like transform," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1247-1263, 2020. 2 + +[32] X. Zhang and X. Wu, "LVQAC: Lattice vector quantization coupled with spatially adaptive companding for efficient learned image compression," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10239-10248, 2023. 2 +[33] R. Feng, Z. Guo, W. Li, and Z. Chen, "NVTC: Nonlinear vector transform coding," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6101-6110, 2023. 2 +[34] K. Suhring, M. Schäfer, J. Pfaff, H. Schwarz, D. Marpe, and T. Wiegand, "Trellis-coded quantization for end-to-end learned image compression," in 2022 IEEE International Conference on Image Processing (ICIP), pp. 3306-3310, IEEE, 2022. 2 +[35] Z. Ge, S. Ma, W. Gao, J. Pan, and C. Jia, "NLIC: Nonuniform quantization based learned image compression," IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2 +[36] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, "Causal contextual prediction for learned image compression," IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2329-2341, 2021. 2 +[37] X. Zhu, J. Song, L. Gao, F. Zheng, and H. T. Shen, "Unified multivariate gaussian mixture for efficient neural image compression," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17612-17621, 2022. 2 +[38] H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, "Learned image compression with gaussian-laplacian-logistic mixture model and concatenated residual modules," IEEE Transactions on Image Processing, vol. 32, pp. 2063-2076, 2023. 2 +[39] Y. Hu, W. Yang, and J. Liu, "Coarse-to-fine hyper-prior modeling for learned image compression," in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 11013-11020, 2020. 2 +[40] L. Theis, W. Shi, A. Cunningham, and F. Huszár, "Lossy image compression with compressive autoencoders," arXiv preprint arXiv:1703.00395, 2017. 3 +[41] I. Goodfellow, "Deep learning," 2016. 4 +[42] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, "Generative adversarial nets," Advances in neural information processing systems, vol. 27, 2014. 4 +[43] D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang, "ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5718-5727, 2022. 5, 6, 2 +[44] J. Bégaint, F. Racapé, S. Feltman, and A. Pushparaja, "Compressai: a pytorch library and evaluation platform for end-to-end compression research," arXiv preprint arXiv:2011.03029, 2020. 5, 1 +[45] W. Jiang, “Unofficial elic.” https://github.com/JiangWeibeta/ELIC, 2022.5 +[46] J. Liu, G. Lu, Z. Hu, and D. Xu, "A unified end-to-end framework for efficient deep image compression," arXiv preprint arXiv:2002.03370, 2020. 6 + +[47] D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 6 +[48] E. Kodak, "Kodak lossless true color image suite (photocd pcd0992)," URL http://rOk.us/graphics/kodak, vol. 6, p. 2, 1993. 6 +[49] "6th Challenge on Learned Image Compression." https://www.compression.cc/tasks/index.html. Accessed: 2024-11-13. 6 +[50] N. Asuni and A. Giachetti, "Testimages: a large-scale archive for testing visual devices and basic image processing algorithms," in STAG, pp. 63-70, 2014. 6 +[51] T. Zhang, H. Zhang, Y. Li, L. Li, and D. Liu, “Few-shot domain adaptation for learned image compression,” arXiv preprint arXiv:2409.11111, 2024. 6 +[52] Y. Lv, J. Xiang, J. Zhang, W. Yang, X. Han, and W. Yang, "Dynamic low-rank instance adaptation for universal neural image compression," in Proceedings of the 31st ACM International Conference on Multimedia, pp. 632-642, 2023. 6 +[53] J. Yang, S. Shen, H. Yue, and K. Li, "Implicit transformer network for screen content image continuous superresolution," Advances in Neural Information Processing Systems, vol. 34, pp. 13304-13315, 2021. 6 +[54] X. Min, K. Ma, K. Gu, G. Zhai, Z. Wang, and W. Lin, "Unified blind quality assessment of compressed natural, graphic, and screen content images," IEEE Transactions on Image Processing, vol. 26, no. 11, pp. 5462-5474, 2017. 6 +[55] N. Brancati, A. M. Anniciello, P. Pati, D. Riccio, G. Scognamiglio, G. Jaume, G. De Pietro, M. Di Bonito, A. Fon-cubierta, G. Botti, et al., "BRACS: A dataset for breast carcinoma subtyping in h&e histology images," Database, vol. 2022, p. baac093, 2022. 6 +[56] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1251–1258, 2017. 8, 2 +[57] G. Bjontegaard, "Calculation of average PSNR differences between RD-curves," ITU SG16 Doc. VCEG-M33, 2001. 2 \ No newline at end of file diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/images.zip b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e0d44f26f23a3db184c25b1fe2e890e19b79067a --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1241ed9fd8a3611fb476ca78a63fd6113f9ff7f6dc0d8bb766bf097ba2c30dc5 +size 462978 diff --git a/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/layout.json b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e7af64e1b232ee44cd1ed3b31f50d373016ef07 --- /dev/null +++ b/2025/An Information-Theoretic Regularizer for Lossy Neural Image Compression/layout.json @@ -0,0 +1,11087 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 121 + ], + "type": "text", + "content": "An Information-Theoretic Regularizer for Lossy Neural Image Compression" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": "Yingwen Zhang" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Meng Wang" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Xihua Sheng" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Peilin Chen" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Junru Li" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Li Zhang" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "text", + "content": ", Shiqi Wang" + }, + { + "bbox": [ + 157, + 142, + 453, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "spans": [ + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "text", + "content": "Department of Computer Science, City University of Hong Kong, Hong Kong, China " + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "text", + "content": "School of Data Science, Lingnan University, Hong Kong, China " + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 171, + 512, + 214 + ], + "type": "text", + "content": "Bytedance Inc., San Diego, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 494 + ], + "type": "text", + "content": "Lossy image compression networks aim to minimize the latent entropy of images while adhering to specific distortion constraints. However, optimizing the neural network can be challenging due to its nature of learning quantized latent representations. In this paper, our key finding is that minimizing the latent entropy is, to some extent, equivalent to maximizing the conditional source entropy, an insight that is deeply rooted in information-theoretic equalities. Building on this insight, we propose a novel structural regularization method for the neural image compression task by incorporating the negative conditional source entropy into the training objective, such that both the optimization efficacy and the model's generalization ability can be promoted. The proposed information-theoretic regularizer is interpretable, plug-and-play, and imposes no inference overheads. Extensive experiments demonstrate its superiority in regularizing the models and further squeezing bits from the latent representation across various compression structures and unseen domains. Code is available at Github." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 135, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 135, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 135, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 605 + ], + "type": "text", + "content": "Lossy image compression, or more broadly, lossy data compression, is grounded in a fundamental theorem in information theory: the rate-distortion theorem [1]. Essentially, this theorem establishes a theoretical lower bound for lossy compression and provides valuable guidance for designing practical compression algorithms [2, 3]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": "A basic lossy image compression model is illustrated in Fig.1(a). Let " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": " denote the source, the reconstruction, and the discrete index (or the latent), respectively. The encoder is a deterministic mapping " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\mathrm{Q} \\colon \\mathcal{X}^N \\to \\mathcal{U}" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^N" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": "-dimensional image space and " + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "inline_equation", + "content": "\\mathcal{U} = \\{1, 2, \\dots, M\\}" + }, + { + "bbox": [ + 55, + 605, + 295, + 679 + ], + "type": "text", + "content": " denotes the discrete index space. The decoder" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 240, + 408, + 314 + ], + "blocks": [ + { + "bbox": [ + 322, + 240, + 408, + 314 + ], + "lines": [ + { + "bbox": [ + 322, + 240, + 408, + 314 + ], + "spans": [ + { + "bbox": [ + 322, + 240, + 408, + 314 + ], + "type": "image", + "image_path": "a2653bdbcd142fbec9a2e1c5551274f1d4e6a38adade3615fe13c7a42b36c7e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 415, + 239, + 547, + 313 + ], + "blocks": [ + { + "bbox": [ + 415, + 239, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 415, + 239, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 415, + 239, + 547, + 313 + ], + "type": "image", + "image_path": "ea905594142a6088fcd23c8c6a4c9f48b1d979f342a9aaaf4f34258c68f18706.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 359, + 326, + 507, + 398 + ], + "blocks": [ + { + "bbox": [ + 340, + 384, + 352, + 396 + ], + "lines": [ + { + "bbox": [ + 340, + 384, + 352, + 396 + ], + "spans": [ + { + "bbox": [ + 340, + 384, + 352, + 396 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 359, + 326, + 507, + 398 + ], + "lines": [ + { + "bbox": [ + 359, + 326, + 507, + 398 + ], + "spans": [ + { + "bbox": [ + 359, + 326, + 507, + 398 + ], + "type": "image", + "image_path": "db1c111c8e4e030aba4d06968a12c94ccd1998e5d7520a827623981b5458bfa4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "text", + "content": "Figure 1. (a) Direct coding model; (b) Information diagram for the direct coding, wherein " + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}) = H(\\mathbf{X}|\\hat{\\mathbf{X}}) + H(\\mathbf{U})" + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "text", + "content": " is fixed for any known source; (c) Transform coding model (with possible side information " + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 313, + 406, + 555, + 452 + ], + "type": "text", + "content": " [4])." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": "is a bijection mapping " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}^{-1} \\colon \\mathcal{U} \\to \\hat{\\mathcal{X}}^N" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": ", where the reconstruction space " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}^N = \\{c_1, \\dots, c_M\\}" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": " forms a " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": "-dimensional codebook. For index " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": ", entropy coding is applied to losslessly convert the data into the bitstream, which can then be decoded back to " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": " on the decoder side. We may refer to such a model as the direct coding model for its structural simplicity. Subsequently, following the rate-distortion theorem [1], given a bounded distortion measure " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": " and a distortion constraint " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": ", the rate of transmitting " + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 468, + 555, + 589 + ], + "type": "text", + "content": " is lower bounded by the informational rate-distortion function:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 594, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 594, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 328, + 594, + 555, + 624 + ], + "type": "interline_equation", + "content": "R (D) = \\min _ {p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}: \\mathbb {E} _ {p _ {\\boldsymbol {X}} p _ {\\hat {\\boldsymbol {X}} | \\boldsymbol {X}}} [ d (\\boldsymbol {X}, \\hat {\\boldsymbol {X}}) ] < D} \\frac {1}{N} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}), \\tag {1}", + "image_path": "c72a222cbc64a8220ad434072424150a2cb647848ce1f8a557095fe0d8970d26.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "i.e., the minimum mutual information obtained by traversing all the possible encoder-decoder (or quantizer) designs " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "p_{\\hat{X} |X}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". Theoretically, this bound can be achieved asymptotically using vector quantization as " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N \\to \\infty" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " [5]. However, since the encoding complexity grows exponentially with the image dimension " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", a suboptimal yet more practical solution is the transform coding [6]. As depicted in" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "*Corresponding author (mengwang7@ln.edu.hk). This work was supported in part by ITF Project GHP/044/21SZ, in part by the Start-up Grant SUG-007/2425, and in part by Faculty Research Grant SDS24A6." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15573" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "Fig.1(c), for the transform coding model, a deterministic analysis transform " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\mathrm{T_A}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^N \\to \\mathcal{Y}^M" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " and a deterministic synthesis transform " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\mathrm{T_S}" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}^M \\to \\hat{\\mathcal{X}}^N" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": " are incorporated into the compression chain. The analysis transform is designed to decorrelate the source " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ", such that a much simpler quantization scheme, e.g., uniform scalar quantization: " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_i = \\mathrm{round}(\\boldsymbol{y}_i)" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ", can be independently applied to each dimension of the transformed coefficients " + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": ". Building on this framework and leveraging the superior nonlinear capabilities of neural networks, numerous neural image compression networks [4, 7-13] are proposed, continually pushing the frontier of lossy image compression." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "text", + "content": "Typically, by parameterizing the " + }, + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{T_A}" + }, + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{T_S}" + }, + { + "bbox": [ + 55, + 216, + 294, + 251 + ], + "type": "text", + "content": " in Fig.1(c) as neural networks, the rate-distortion training objective is formulated as [7]:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 87, + 259, + 295, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 259, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 87, + 259, + 295, + 291 + ], + "type": "interline_equation", + "content": "\\min _ {R \\approx H (U)} \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} [ - \\log q _ {\\phi} (\\boldsymbol {U}) ]} + \\lambda \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} | | \\boldsymbol {X} - \\hat {\\boldsymbol {X}} | | _ {2} ^ {2}} _ {D}, \\tag {2}", + "image_path": "cdfdde44acbf5ad635e8c09c265554f31c042aa7c2fac33c3eee43e20c785b78.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " is the estimation of latent entropy " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "H(\\pmb{U})" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "q_{\\phi}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " is the neural entropy model (or estimated distribution) for " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": ", whose modeling accuracy can be significantly enhanced by transmitting a small amount of side information " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{Z}}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " (as depicted in Fig.1(c)) as part of the parameters " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " [4]; the distortion " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " distance between " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{X}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{X}}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " is the Lagrangian multiplier, controlling the trade-off between " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": ". Herein, due to the existence of quantization " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}(\\cdot)" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": " is discrete in essence. Optimizing the network is then challenging as the gradient is almost zero everywhere for " + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}(\\cdot)" + }, + { + "bbox": [ + 55, + 299, + 295, + 443 + ], + "type": "text", + "content": ". To effectively train the networks, a surrogate is introduced during training, e.g., additive uniform noise (AUN) [7]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 464, + 295, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 464, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 295, + 478 + ], + "type": "interline_equation", + "content": "\\boldsymbol {U} = \\mathrm {Q} (\\boldsymbol {Y}) \\approx \\boldsymbol {Y} + [ - 0. 5, 0. 5 ] ^ {M}, \\tag {3}", + "image_path": "d9b93e3eb374dfd872a122020b0d6dbabe43f1e9e83ab50717e572cd60b42091.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": "such that the backward gradient of " + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathrm{Q}(\\cdot)" + }, + { + "bbox": [ + 55, + 486, + 295, + 558 + ], + "type": "text", + "content": " is one. However, since such approximation inevitably results in the train-test mismatch, and thus high gradient estimation error [14], both the efficiency and efficacy of the optimization are hindered. To address this, many research efforts are devoted to designing advanced surrogates and optimization methods [14-18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we present a completely distinct view on facilitating the optimization of the compression network: regularization. Conceivably, imposing structural constraints on the network optimization can effectively regularize the gradient estimation error, promote convergence, and reduce overfitting. In the context of generative variational autoencoders [19], structural regularization techniques [20-25] have been effectively applied to improve authentic data generation. However, since neural compression networks fundamentally differ from generative autoencoders in terms of the minimum latent entropy constraint, to the best of our knowledge, no structural regularization methods tailored to this constraint have been explored. In light of this, in this" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "paper, we present the first study of information-theoretic regularization for the neural image compression task. In particular, through an in-depth information-theoretic analysis of compression models, we reveal that minimizing the latent entropy " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": " is, to some extent, equally important as maximizing the conditional source entropy " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ". Subsequently, by enforcing the network to simultaneously minimize " + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "inline_equation", + "content": "-H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": ", both the in-domain compact representation and out-of-domain generalization ability are experimentally improved." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 205, + 404, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 404, + 218 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 404, + 218 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 227, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 586 + ], + "type": "text", + "content": "Neural image compression networks. Neural image compression networks are primarily composed of transform, quantization, and latent entropy models. The transform module maps the input source into a latent representation so that the dependencies of the source domain can be more easily handled. Unlike the linear transform [26] used in traditional hybrid coding, the neural transform is inherently nonlinear, enabling more effective decorrelation of the source data. In the literature, various neural transform architectures are proposed. Representative architectures include the recurrent neural network (RNN) [27], convolutional neural network (CNN) [7], non-local attention [28], transformer [29], mixed transformer-CNN [30], state-space models [13] and invertible networks [31]. Subsequently, the continuous transformed coefficients are quantized into the discrete representation, wherein the uniform scalar quantization [7] is commonly adopted. As vector quantization can, in principle, better approach the rate-distortion bound (Eqn.(1)), low-complexity vector quantization [15, 32, 33] and dependent scalar quantization [34, 35] schemes are also studied. Finally, once the quantized latent representations are obtained, a neural entropy model is essential for performing entropy coding. Herein, since the latent cannot be perfectly decorrelated, an accurate entropy model that can better capture the dependencies and closely match the ground-truth latent distribution is crucial for the overall network performance. Prominent research directions such as spatial-channel context mining [8, 9, 12, 36], expressive latent distribution modeling [10, 37, 38] and latent architecture designs [4, 11, 39] are explored." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 713 + ], + "type": "text", + "content": "Optimization over quantized latent representations. The surrogate-based methods [7, 14-17] have been shown to effectively address the non-differentiable quantization function during end-to-end optimization. In addition to AUN [7], various quantization surrogates such as soft assignment [15], Gumbel-softmax trick [16], and soft rounding function [17] have been studied. However, as the surrogate introduces a train-test mismatch in forward calculations, the annealing techniques [14-17] are then proposed to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15574" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "control the \"softness\" of surrogates, gradually approximating the hard quantization function as training progresses. The straight-through estimator (STE) [40] provides an alternative approach to address the train-test mismatch by enforcing an identity mapping during the backward calculation, thereby enabling direct hard quantization in the forward pass. Nevertheless, STE ultimately introduces a significant gradient estimation bias for the analysis transform. To counter this, an additional fine-tuning stage for the entropy model and synthesis transform is suggested [18]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 336 + ], + "type": "text", + "content": "While numerous network optimization methods have been developed from a quantization perspective, the regularization view has largely escaped research attention. To the best of our knowledge, the most related efforts are regularization methods for variational autoencoders (VAEs), such as regularization on latent priors [20-22], reconstruction [23, 24], and consistency [25]. However, as VAEs are designed for generative modeling, these methods are not applicable to optimizing compression networks. To bridge this knowledge gap, this paper proposes a structural regularizer for neural image compression and conducts extensive experiments to demonstrate its superiority." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 138, + 362 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": "In this section, we begin by deriving the equivalent maximization form of the latent entropy minimization problem for both the direct and transform coding models. Building on this derivation, we propose a novel regularization method for optimizing the neural compression network. Note that during our derivations, we use the naive transform coding model without the side information branch (i.e., the " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": " depicted in Fig 1(c)). This significantly simplifies our analysis without compromising much practicality, as it has been shown that the information flowing into the side branch is negligible compared to the latent branch [4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "3.1. Motivation: from minimization to maximization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "content": "Lemma 1. Considering a deterministic quantization process " + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "content": " and a deterministic dequantization process " + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "inline_equation", + "content": "Q^{-1}(\\cdot)" + }, + { + "bbox": [ + 55, + 540, + 295, + 588 + ], + "type": "text", + "content": " for the direct coding model illustrated in Fig. 1(a), the following equalities hold automatically:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 130, + 597, + 295, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 597, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 130, + 597, + 295, + 642 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) & (4) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) & (5) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) & (6) \\end{array} \\right.", + "image_path": "07197278c422e34546dfe0eb72c020507753383d171876d076c84661a8e49b13.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "Proof. Eqn.(4) follows the facts that " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", meaning there is no uncertainty in " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " given the observation of " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "I(\\mathbf{X};\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}) - H(\\hat{\\mathbf{X}}|\\mathbf{X})" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ". Similarly, Eqn.(5) follows the fact that " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "H(\\hat{\\mathbf{X}}) + H(\\mathbf{U}|\\hat{\\mathbf{X}}) = H(\\mathbf{U}) + H(\\hat{\\mathbf{X}}|\\mathbf{U})" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ". Eqn.(6) is imme" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "dately apparent from Eqn.(4) and Eqn.(5). Additional details are provided in the Supplementary Material (Sec. 6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 553, + 135 + ], + "type": "text", + "content": "Theorem 1. For the direct coding model elaborated in Lemma 1, we have" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 145, + 553, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 145, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 372, + 145, + 553, + 160 + ], + "type": "interline_equation", + "content": "H (\\boldsymbol {U}) = H (\\boldsymbol {X}) - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}). \\tag {7}", + "image_path": "8f6114ddae926219c082d9e7e9d3d96f5a390455311b926205515af4992c96d8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "content": "The informational diagram is thus illustrated as Fig.1(b), given the non-negativity of discrete entropy " + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 313, + 171, + 553, + 195 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 230 + ], + "type": "text", + "content": "Proof. This is immediately apparent from Eqn.(6), as the right-hand-side of Eqn.(7) is by definition " + }, + { + "bbox": [ + 313, + 205, + 553, + 230 + ], + "type": "inline_equation", + "content": "I(\\pmb {X};\\hat{\\pmb{X}})" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "content": "From Theorem 1, for the direct coding model, it is clear that minimizing the latent entropy " + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "content": " can be equivalently achieved by maximizing the conditional source entropy " + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "content": ", for any known source " + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "content": ". In practice, this aligns with the training setting of the neural compression network, wherein the training set is given in advance, and thus " + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 243, + 555, + 365 + ], + "type": "text", + "content": " is fixed. However, as the prevailing neural compression networks are based on the transform coding model, in the following discussion, we extend this conclusion to the transform coding setting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "text", + "content": "Lemma 2. Considering a transform coding model built on the direct coding model elaborated in Lemma 1 with a deterministic analysis transform " + }, + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "inline_equation", + "content": "T_{A}" + }, + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "text", + "content": " and a deterministic synthesis transform " + }, + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "inline_equation", + "content": "T_{S}" + }, + { + "bbox": [ + 313, + 374, + 555, + 434 + ], + "type": "text", + "content": " (Fig. 1(c)), the following equalities hold automatically:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 444, + 553, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 444, + 553, + 490 + ], + "spans": [ + { + "bbox": [ + 364, + 444, + 553, + 490 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) = H (\\hat {\\boldsymbol {X}}) \\\\ H (\\hat {\\boldsymbol {X}}) = H (\\boldsymbol {U}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\\\ H (\\boldsymbol {U}) = I (\\boldsymbol {X}; \\hat {\\boldsymbol {X}}) + H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}}) \\end{array} \\right. \\tag {8}", + "image_path": "938e6f60635a339dcd61190bf08c2d7b1ba080f083ffabca881b7b9fdf7c40b7.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "content": "Proof. Compared to the proof of Lemma 1, for the transform coding, we still have " + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "inline_equation", + "content": "H(\\hat{\\mathbf{X}}|\\mathbf{X}) = 0" + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "inline_equation", + "content": "H(\\hat{\\mathbf{X}}|\\mathbf{U}) = 0" + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "content": ". However, since the synthesis transform " + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "inline_equation", + "content": "T_{S}" + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "content": " is not guaranteed to be a bijection mapping, the " + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}}) = 0" + }, + { + "bbox": [ + 313, + 498, + 554, + 583 + ], + "type": "text", + "content": " does not always hold, leading to the correction of Eqn.(9) and thus Eqn.(10). For more details, readers may refer to the Supplementary Material (Sec. 7)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "content": "Theorem 2. For the transform coding model elaborated in Lemma 2, we have" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 334, + 632, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 632, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 334, + 632, + 553, + 646 + ], + "type": "interline_equation", + "content": "H (\\boldsymbol {U}) = H (\\boldsymbol {X}) - \\left(H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) - H (\\boldsymbol {U} | \\hat {\\boldsymbol {X}})\\right). \\tag {11}", + "image_path": "5a72ab8e939f413d4879a3ec4e3a208f3b5f3d2b78d2ad5b92a8149f8b91e559.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 658, + 525, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 525, + 671 + ], + "type": "text", + "content": "Proof. This is immediately apparent from Eqn.(10)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Remarks. Two conclusions can be drawn for the transform coding model." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15575" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 72, + 296, + 217 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": "(a) Equivalence of minimization and maximization. From Theorem 2, for any known source " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": ", minimizing " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": " can be equivalently achieved by maximizing the " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": ". For a bijection synthesis function " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "T_{S}" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": ", of which " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": " is zero, this maximization objective is reduced to " + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 59, + 72, + 295, + 156 + ], + "type": "text", + "content": ", which is the same as the direct coding;" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": "(b) Achieving " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "R(D)" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": " by minimizing " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "H(\\pmb{U})" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": ". From Eqn.(10), minimizing " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "H(\\pmb{U})" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": " is simultaneously minimizing " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "I(\\pmb{X};\\hat{\\pmb{X}})" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "H(\\pmb{U}|\\hat{\\pmb{X}})" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": ". In the ideal case, the informational lower bound given in Eqn.(1) is achieved with " + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "inline_equation", + "content": "H(\\pmb{U}|\\hat{\\pmb{X}})" + }, + { + "bbox": [ + 58, + 156, + 296, + 217 + ], + "type": "text", + "content": " being reduced to 0." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 70, + 228, + 278, + 344 + ], + "blocks": [ + { + "bbox": [ + 70, + 228, + 278, + 344 + ], + "lines": [ + { + "bbox": [ + 70, + 228, + 278, + 344 + ], + "spans": [ + { + "bbox": [ + 70, + 228, + 278, + 344 + ], + "type": "image", + "image_path": "6bcb1680f39bd2d58e0c52077b769a750e0c7f366f92710bb39426740177ba1a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "text", + "content": "Figure 2. An illustration of the proposed regularization method, wherein an additional source entropy model " + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "inline_equation", + "content": "q_{\\theta}" + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "text", + "content": " is introduced. The AE and AD represent arithmetic encoding and arithmetic decoding, respectively. The side information branch (i.e., the " + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\hat{Z}" + }, + { + "bbox": [ + 55, + 354, + 296, + 420 + ], + "type": "text", + "content": " branch) is considered as part of the latent entropy model, which necessitates bits transmission." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 445, + 182, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 182, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 182, + 456 + ], + "type": "text", + "content": "Algorithm 1 Training Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 460, + 295, + 590 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 460, + 174, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 174, + 472 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 174, + 472 + ], + "type": "text", + "content": "1: for each training step do" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 472, + 246, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 246, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 246, + 483 + ], + "type": "text", + "content": "2: Sample a minibatch of training images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 484, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 484, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 484, + 295, + 508 + ], + "type": "text", + "content": "3: Perform the forward pass of the compression network and source entropy model " + }, + { + "bbox": [ + 62, + 484, + 295, + 508 + ], + "type": "inline_equation", + "content": "q_{\\theta}" + }, + { + "bbox": [ + 62, + 484, + 295, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 508, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 508, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 508, + 295, + 532 + ], + "type": "text", + "content": "4: Compute the backward gradient and update the compression network based on Eqn.(12)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 532, + 222, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 222, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 222, + 543 + ], + "type": "text", + "content": "5: Freeze the compression network." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 216, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 216, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 216, + 555 + ], + "type": "text", + "content": "6: Perform the forward pass of " + }, + { + "bbox": [ + 62, + 544, + 216, + 555 + ], + "type": "inline_equation", + "content": "q_{\\theta}" + }, + { + "bbox": [ + 62, + 544, + 216, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 556, + 294, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 579 + ], + "type": "text", + "content": "7: Compute the backward gradient and update the source entropy model " + }, + { + "bbox": [ + 62, + 556, + 294, + 579 + ], + "type": "inline_equation", + "content": "q_{\\theta}" + }, + { + "bbox": [ + 62, + 556, + 294, + 579 + ], + "type": "text", + "content": " based on Eqn.(13)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 580, + 106, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 106, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 106, + 590 + ], + "type": "text", + "content": "8: end for" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 612, + 235, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 235, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 235, + 625 + ], + "type": "text", + "content": "3.2. Information-theoretic regularizer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In principle, optimizing neural compression networks with a quantization surrogate, e.g., AUN, introduces gradient estimation errors [14], thereby deviating the objective of minimizing the latent entropy " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " under certain distortion constraints (Eqn.(2)), harming both the efficiency and efficacy of the optimization. Motivated by the equivalence of the minimization and maximization optimization discussed" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "in Sec. 3.1, we propose to address this issue by imposing structural maximization constraints on the optimization process. As such, the vanilla objective of minimizing " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": " can be further enhanced through this maximization regularization." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": "From Theorem 2, an ideal gradient direction for model updating is one that simultaneously maximizes " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": ", which can be further interpreted as maximizing " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " while minimizing " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": ". To integrate this regularization into the training objective, two parameterized distributions " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " are necessitated. However, accurately modeling an entropy model " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " alongside the compression network is quite challenging, as it involves learning another \"encoder\" function from the reconstruction " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " to the latent " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": ". We experimentally find that a poorly learned " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " regularizer may harm the optimization performance, and that the increased training complexity is also burdensome. To balance practical feasibility with regularization accuracy, we include only the first " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " term for regularization, which is an upper bound for the ground-truth " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " objective. The entropy modeling becomes more tractable since the reconstruction " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " is closely aligned with the source " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": ". The primary rationale for such simplification is that although maximizing " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " alone is less precise than " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}}) - H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": ", it still provides a meaningful constraint on the \"noisy\" gradient, yielding better gradient direction than no regularization. Another justification is that since we will retain the " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " minimization term in the overall objective, this " + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "inline_equation", + "content": "H(\\mathbf{U}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 133, + 555, + 469 + ], + "type": "text", + "content": " gap, in principle (Eqn.(11)), should decrease as training progresses, which ensures that the regularization effect remains aligned with the ideal maximization objective. The resulting training objective is thus given by:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 472, + 553, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 472, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 356, + 472, + 553, + 505 + ], + "type": "interline_equation", + "content": "\\min R + \\lambda D + \\alpha \\underbrace {\\mathbb {E} _ {\\boldsymbol {X}} \\left[ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\right]} _ {\\approx - H (\\boldsymbol {X} | \\hat {\\boldsymbol {X}})}, \\tag {12}", + "image_path": "944ccc5503d4f53249d4bd5457f8d4403c06961cebc2abe744b4518ac0c55b58.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "R + \\lambda D" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": " follows the rate-distortion loss in Eqn.(2); " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "-H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": " is the proposed regularizer; " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": " is the estimated source distribution; " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": " is regularization factor, controlling the degree of regularization. An illustration for this regularization method is given in Fig.(2). In addition to the compression network, " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": " is introduced, which is parameterized by a neural network and can be optimized via the maximum likelihood criteria [41]:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 382, + 610, + 553, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 610, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 382, + 610, + 553, + 624 + ], + "type": "interline_equation", + "content": "\\max \\mathbb {E} _ {\\boldsymbol {X}} [ \\log q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) ]. \\tag {13}", + "image_path": "6c7ddbd377c73c96a88eb9245d3fffd76c98fdbe23cbbc5f2ce3b2b63219dc6d.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": "This results in a max-min optimization problem where minimizing " + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "inline_equation", + "content": "-H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": " (Eqn.(12)) simultaneously creates a conflict. To resolve this, a GAN-style two-stage joint training strategy [42] is employed: At each training step, we first update the compression network using Eqn.(12), then freeze it and update " + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 629, + 554, + 714 + ], + "type": "text", + "content": " using Eqn.(13). We summarize the strategy in Algorithm 1." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15576" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 149, + 122 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 149, + 122 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 149, + 122 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 149, + 122 + ], + "type": "image", + "image_path": "38fa7dc18bec11553ab0f5a84ff182b713062af3a4c1b0221f143fba9f9dc7a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 132, + 145, + 232 + ], + "blocks": [ + { + "bbox": [ + 61, + 132, + 145, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 132, + 145, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 132, + 145, + 232 + ], + "type": "image", + "image_path": "9f24ee091835ee48c889f36d3a02d26b1f2aa3a6eaeaf5640f07f9fff9d554a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 277 + ], + "type": "text", + "content": "Figure 3. Conditional source entropy modeling for (a) hyperprior [4], (b) autoregressive [8] and attention [10], (c) EILC [43], and (d) MLIC++ [12]. The architecture and module designs are aligned with the latent entropy model. Additional details regarding the specific modules are provided in the Supplementary Material (Sec. 9)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 91, + 352, + 233 + ], + "blocks": [ + { + "bbox": [ + 170, + 91, + 352, + 233 + ], + "lines": [ + { + "bbox": [ + 170, + 91, + 352, + 233 + ], + "spans": [ + { + "bbox": [ + 170, + 91, + 352, + 233 + ], + "type": "image", + "image_path": "3eaf34fbfde9d337f1fe3961e3024d6e7ef90a76df19d45a15f239c1266fd376.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 362, + 86, + 548, + 234 + ], + "blocks": [ + { + "bbox": [ + 362, + 86, + 548, + 234 + ], + "lines": [ + { + "bbox": [ + 362, + 86, + 548, + 234 + ], + "spans": [ + { + "bbox": [ + 362, + 86, + 548, + 234 + ], + "type": "image", + "image_path": "b93818c47499f3a27ff75f292f5372684cbb4fae645961d619278d4728e2f5a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 137, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 137, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 137, + 309 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 316, + 170, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 170, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 170, + 330 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 334, + 296, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 490 + ], + "type": "text", + "content": "Compression models. Three classic image compression models are used in exhaustive validation experiments: the hyperprior context model [4] (with scale and mean [8]), the joint hyperprior and spatial autoregressive context model [8], and the attention-based transform model [10]. We may refer to these models as hyperprior, autoregressive, and attention, respectively. Moreover, we conduct the experiments on two advanced compression models: ELIC [43] and MLIC++ [12]. Notably, ELIC utilizes checkerboard and channel-conditioned latent entropy contexts, while MLIC++ builds upon ELIC by introducing additional attention-based context designs and is identified as one of the state-of-the-art (SOTA) models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 624 + ], + "type": "text", + "content": "We train these models from scratch using the vanilla rate-distortion loss (Eqn.(2)), and compare their performance when trained with the proposed regularization method (Algorithm 1) under an equal number of training steps. For the implementations of the hyperprior, autoregressive, and attention models, we adopt CompressAI's re-implementation [44], where the quantization surrogate is ANU. For ELIC and MLIC++, we use the open re-implementations from [45] and the official implementation, respectively, both of which employ a mixed ANU-STE quantization surrogate [9]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Conditional source entropy modeling. For each compression model, we apply exactly the same entropy modeling architecture as the latent entropy modeling " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "q_{\\phi}(\\boldsymbol{U}|\\hat{\\boldsymbol{Z}})" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". Our experiments indicate that aligning the entropy model between the latent and source is critical for effective regularization; deviations in either direction (stronger or weaker" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "content": "source models) degrade compression performance. For the hyperprior model [4], " + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "content": " is modeled as the factorized Gaussian distribution conditioned on " + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{X}}" + }, + { + "bbox": [ + 313, + 297, + 555, + 333 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 339, + 555, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 339, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 377, + 339, + 555, + 372 + ], + "type": "interline_equation", + "content": "q _ {\\theta} (\\boldsymbol {X} | \\hat {\\boldsymbol {X}}) \\sim \\prod_ {i = 1} ^ {N} \\mathbb {N} \\left(\\mu_ {i}, \\sigma_ {i} ^ {2}\\right), \\tag {14}", + "image_path": "308868d8419ad388c67eb398c77781024448722144462867c73b9e911a394338.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 374, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 374, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 331, + 374, + 555, + 389 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\Phi \\triangleq \\left(\\mu_ {1}, \\dots , \\mu_ {N}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {N} ^ {2}\\right) = f _ {\\theta} (\\hat {\\boldsymbol {X}}), \\tag {15}", + "image_path": "6347e79edd407796df94cfeb0c03b94f7791126e3e58c50462dab9958af0cfbe.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " is modeled by convolutional layers; " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\mathbb{N}" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " denotes the Gaussian distribution; " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " denote the mean and variance; " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " denotes the dimension of both " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": ". For both the autoregressive and attention models, the mean and variance are determined by the additional spatial context, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 355, + 463, + 555, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 463, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 355, + 463, + 555, + 479 + ], + "type": "interline_equation", + "content": "\\left(\\mu_ {1}, \\dots , \\mu_ {i}, \\sigma_ {1} ^ {2}, \\dots , \\sigma_ {i} ^ {2}\\right) = f _ {\\theta} \\left(\\hat {\\boldsymbol {X}}, \\boldsymbol {X} _ {< i}\\right), \\tag {16}", + "image_path": "d8edb7da9734e886fdd921690164856e945033d2e81594c4b6d279a131ae9d05.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "inline_equation", + "content": "X_{ 1)" + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "text", + "content": " slice, both the " + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_{\\text{anchor}}^{k}" + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\Phi_{\\text{non-anchor}}^{k}" + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "text", + "content": " are conditioned on extra channel context from previous slices " + }, + { + "bbox": [ + 313, + 485, + 556, + 714 + ], + "type": "inline_equation", + "content": "X^{Test setRegularized v.s. UnregularizedhyperpriorautoregressiveattentionELICMLIC++Kodak-1.14%-1.57%-0.84%-1.24%-1.82%Tecnick-0.83%-0.78%-0.91%-2.00%-4.01%CLIC-0.66%-0.83%-1.01%-1.45%-3.25%Natural Avg.-0.88%-1.06%-0.92%-1.56%-3.03%Pixel-style-2.20%-1.96%-0.47%-2.71%-3.08%Screen-1.15%-2.01%-0.69%-1.02%-1.72%Game-1.31%-0.83%-0.82%-2.04%-3.44%Pathology-1.11%-0.81%-2.38%-1.82%-0.43%", + "image_path": "9a2691937fe36e9a2b5b8c2d847501948c593da84c4bb102c78e9e61eb1ebd9b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 693, + 539, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 539, + 702 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 539, + 702 + ], + "type": "text", + "content": "1The Crater dataset is excluded due to insufficient documentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 703, + 477, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 477, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 477, + 712 + ], + "type": "text", + "content": "2Folder: BRACS_Roi\\latest_version\\test\\0_N" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15578" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 70, + 265, + 208 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 265, + 208 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 265, + 208 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 265, + 208 + ], + "type": "image", + "image_path": "e3ecf35ee4d10ed0c30c11f9ef9557f0deabd766c4b92af07e195f17f833d29b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 88, + 211, + 265, + 350 + ], + "blocks": [ + { + "bbox": [ + 88, + 211, + 265, + 350 + ], + "lines": [ + { + "bbox": [ + 88, + 211, + 265, + 350 + ], + "spans": [ + { + "bbox": [ + 88, + 211, + 265, + 350 + ], + "type": "image", + "image_path": "511e0d8346e0cbe8a3cbc7bf6b31c9e4ebf19e3179813dd78d538355741921b3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": "Figure 5. Compression performance on Kodak. Our regularizer achieves " + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "inline_equation", + "content": "-1.13\\%" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "inline_equation", + "content": "-1.57\\%" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "inline_equation", + "content": "-0.84\\%" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "inline_equation", + "content": "-1.24\\%" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "inline_equation", + "content": "-1.82\\%" + }, + { + "bbox": [ + 55, + 358, + 297, + 404 + ], + "type": "text", + "content": " BD-Rates for the hyperprior, autoregressive, attention models, ELIC, and MLIC++, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 430, + 113, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 113, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 113, + 441 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": "Performance on natural images. In Fig.4, we present the BD-Rate comparison at four different training steps. The anchor is trained with the vanilla rate-distortion loss (Eqn.(2)) under an equal number of training steps. The regularization factors " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models are set to 0.1, 1, 1, 0.1, and 0.1, respectively. It can be observed that, for training steps beyond " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.5 \\times 10^{6}" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " across all five models, our proposed regularization method can achieve better compression performance than the vanilla optimization method. In particular, at the " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "2 \\times 10^{6}" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " step, the average BD-Rates for hyperprior, autoregressive, and attention are " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-0.88\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-1.06\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-0.92\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": ", respectively; at the " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "1 \\times 10^{6}" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " step, the average BD-Rates for ELIC and MLIC++ are " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-1.56\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-3.03\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": ", respectively. Meanwhile, the results demonstrate that the regularization effect varies across models as training progresses: For the hyperprior, autoregressive, and MLIC++ models, after a brief convergence period, the BD-Rates are gradually improved from 0 to around " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "-1\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": ". For the ELIC model, the regularization is most effective during the earlier training stages, achieving over " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " BD-Rates saving at the " + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "inline_equation", + "content": "0.25 \\times 10^{6}" + }, + { + "bbox": [ + 55, + 450, + 296, + 715 + ], + "type": "text", + "content": " step. For the attention model, the regularization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 476, + 84 + ], + "type": "text", + "content": "achieves a relatively stable coding gain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": "In Fig. 5, we present the detailed rate-distortion curves on the Kodak dataset. For the hyperprior, autoregressive, and attention models, the results correspond to the " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "2 \\times 10^{6}" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " training step, while for the ELIC and MLIC++ models, the results are from the " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "1 \\times 10^{6}" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " step. The results indicate that the introduction of the additional conditional source entropy regularization leads to a slight reduction in the bit rate for most of the points, although a few exceptions exhibit an increase due to training noise. From the form of regularization in Eqn.(12), i.e., " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "-\\alpha H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": ", the extent of such bit rate reduction depends on both the regularization factor " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " and the estimated conditional source entropy " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "H(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 242, + 556, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 556, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 556, + 387 + ], + "type": "text", + "content": "Generalizing on unseen domains. In machine learning, regularization plays an essential role in preventing overfitting and enhancing generalization. In Table 1, we directly apply the models trained on natural images to the four unseen domains and evaluate the effectiveness of our regularization method. The results indicate that, across all new domains, regularized models consistently outperform unregularized ones. Performance improvements vary by domain and model, with notable gains in some cases. For instance, in the pathology domain with the attention model, the regularization achieves a BD-Rate of " + }, + { + "bbox": [ + 313, + 242, + 556, + 387 + ], + "type": "inline_equation", + "content": "-2.38\\%" + }, + { + "bbox": [ + 313, + 242, + 556, + 387 + ], + "type": "text", + "content": ", more than double the improvement observed in the natural domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " and longer training. In Fig.(6)(a), we continue training both the unregularized and regularized hyperprior models to " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "4\\times 10^{6}" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " steps and demonstrate the effect of the regularization factor " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " on the performance. Herein, with extended training, the regularization consistently enhances optimization. The best performance, with " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ", is achieved at the " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "2.5\\times 10^{6}" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " step, showing a BD-Rate of " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "-1.24\\%" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ". As training progresses to " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "4\\times 10^{6}" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " steps, the gain reduces to around " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ". For the effect of hyperparameter " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ", it is shown that either the value is too large or too small will not benefit the optimization, which can be explained as follows. As discussed in Sec. 3.2, our " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " maximization regularization is a bound upper of the ground-truth objective " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha (H(\\boldsymbol{X}|\\hat{\\boldsymbol{X}}) - H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}}))" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " is large, this " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " gap is also amplified, which may harm the regularization accuracy and thus the optimization performance. While one might consider addressing this through additional " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "H(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " estimation using Algorithm 1, such an approach proves generally ineffective. The challenge lies in accurately learning a distribution " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " that maps " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{X}}" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{U}" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": ", as any discrepancy between " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " and the true " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{U}|\\hat{\\boldsymbol{X}})" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " would itself reduce accuracy. Furthermore, the incurred computational overhead is unfavorable. We provide more results in Supplementary Material (Sec. 10). When " + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 400, + 556, + 715 + ], + "type": "text", + "content": " is small, the regularizer has little contribution to the overall loss, and the optimization is reduced to the vanilla objective." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15579" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 218, + 190 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 218, + 190 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 218, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 218, + 190 + ], + "type": "image", + "image_path": "4c62d64e8c6bd727c998b9c6214ec2809de4d450b60475f5af0fb4c2a5d9d25a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 553, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 553, + 222 + ], + "type": "text", + "content": "Figure 6. (a) Effects of regularization factor " + }, + { + "bbox": [ + 55, + 199, + 553, + 222 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 199, + 553, + 222 + ], + "type": "text", + "content": ". (b) Regularization performance for the hyperprior compression model with the autoregressive regularizer; (c) the autoregressive compression model with the hyperprior regularizer." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 222, + 70, + 380, + 190 + ], + "blocks": [ + { + "bbox": [ + 222, + 70, + 380, + 190 + ], + "lines": [ + { + "bbox": [ + 222, + 70, + 380, + 190 + ], + "spans": [ + { + "bbox": [ + 222, + 70, + 380, + 190 + ], + "type": "image", + "image_path": "ef32409542c5519175c4fd7d8c235cfb7e4e72799f45f8ca1a09841292cd1771.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 388, + 70, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 388, + 70, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 388, + 70, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 388, + 70, + 547, + 190 + ], + "type": "image", + "image_path": "1c2270423262d3ec580e415003fe6a60b2d5d84a4ac6a3b89eead9c14fd1a15f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "text", + "content": "On the alignment of the entropy models. In our experimental setup, we emphasize the advantages of aligned entropy modeling between latent and source distributions. Herein, we extend the analysis to explore cases where the entropy models are unaligned, investigating two scenarios: a compression model equipped with (a) a stronger source entropy model and (b) a weaker source entropy model. In particular, for the source models of the hyperprior and autoregressive depicted in Fig.3(a) and Fig.3(b), respectively, we consider the autoregressive entropy model with additional spatial contexts to be a more accurate and thus \"stronger\" model. Thereafter, we train the hyperprior compression model with a autoregressive source regularizer and the autoregressive compression model with a hyperprior source regularizer, and the results are depicted in Fig.6(b) and Fig.6(c), respectively. The values of " + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "text", + "content": " are still set to 0.1 and 1 for the hyperprior and autoregressive regularizers, respectively. The performance of the aligned entropy model is also included for reference. Herein, it can be observed that neither a stronger nor a weaker source entropy model consistently facilitates the optimization performance of the compression models. Specifically, when a stronger autoregressive regularizer is imposed on the hyperprior network, the average BD-Rate remains positive throughout training. When a weaker hyperprior regularizer is employed by the autoregressive network, the optimization is temporarily improved at the " + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "inline_equation", + "content": "1.5 \\times 10^{6}" + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "text", + "content": " training step, with a BD-Rate of " + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "inline_equation", + "content": "-0.20\\%" + }, + { + "bbox": [ + 56, + 243, + 296, + 590 + ], + "type": "text", + "content": ". In contrast, aligning the entropy models leads to significant performance improvements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 133, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 133, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 133, + 611 + ], + "type": "text", + "content": "4.3. Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "text", + "content": "Our regularization method is applied exclusively during the network training stage and therefore introduces no additional inference complexity. However, since it involves training an additional source entropy model " + }, + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "inline_equation", + "content": "q_{\\theta}(\\mathbf{X}|\\hat{\\mathbf{X}})" + }, + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "text", + "content": ", the training complexity is inevitably increased. Assuming the complexities of the source and latent entropy models are approximately equal, the complexity overhead ratio can be calculated as " + }, + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "inline_equation", + "content": "\\frac{C_L}{C_L + C_T} = \\frac{1}{1 + C_T / C_L}" + }, + { + "bbox": [ + 55, + 617, + 296, + 717 + ], + "type": "text", + "content": ", which is closely related" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": "to the complexity ratio between the transform backbone and the latent, i.e., " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "C_T / C_L" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "\\lambda = 0.0018" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ", we train " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "1 \\times 10^5" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": " steps for each model on the same machine (Nvidia RTX 4090) and compute the time ratio. As shown in Table 2, compared to the unregularized baseline, the training time of our proposed method increases by " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "28.8\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "18.2\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "39.3\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "inline_equation", + "content": "29.8\\%" + }, + { + "bbox": [ + 313, + 243, + 555, + 422 + ], + "type": "text", + "content": " for the hyperprior, autoregressive, attention, ELIC, and MLIC++ models, respectively. Notably, for the attention model, the overhead is the least significant. This can be attributed to the fact that the attention-based transform backbone is significantly more complex than the entropy model. To further reduce training complexity, one can employ lightweight convolutional layers, such as depth-wise separable convolution [56], or improve the iterative training strategy outlined in Algorithm 1." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 462, + 553, + 495 + ], + "blocks": [ + { + "bbox": [ + 313, + 430, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 453 + ], + "type": "text", + "content": "Table 2. Training complexity of the proposed regularizer compared to the unregularized anchor." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "table", + "html": "
hyperpriorautoregressiveattentionELICMLIC++
Training time128.8%118.2%108.1%139.3%129.8%
", + "image_path": "8598167f121ca923c57111eeb39fe4e6409e98db4cbb7a1a707e1578ebf6d4d5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 388, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "This paper investigates a structural entropy regularizer for lossy neural image compression. Specifically, motivated by the information-theoretic analysis of compression models, we demonstrate that, in addition to minimizing latent entropy, maximizing conditional source entropy can be equally important. Subsequently, by regularizing the neural networks to simultaneously maximize the conditional source entropy during end-to-end optimization, we experimentally observe improvements in both the in-domain compact representation and the out-of-domain generalization abilities of compression models. Additionally, as the primary overhead of this regularizer comes from training an additional source entropy model, the proposed method imposes no added inference complexity, opening up a new space for regularizing neural compression networks." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15580" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 124 + ], + "type": "text", + "content": "[1] C. E. Shannon et al., \"Coding theorems for a discrete source with a fidelity criterion,\" IRE Nat. Conv. Rec, vol. 4, no. 142-163, p. 1, 1959. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 126, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 295, + 148 + ], + "type": "text", + "content": "[2] K. Sayood, Introduction to data compression. Morgan Kaufmann, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 150, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 150, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 62, + 150, + 295, + 184 + ], + "type": "text", + "content": "[3] Y. Yang, S. Mandt, L. Theis, et al., \"An introduction to neural data compression,\" Foundations and Trends® in Computer Graphics and Vision, vol. 15, no. 2, pp. 113-200, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "text", + "content": "[4] J. Balle, D. Minnen, S. Singh, S. J. Hwang, and N. Johnston, \"Variational image compression with a scale hyperprior,\" arXiv preprint arXiv:1802.01436, 2018. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 253 + ], + "type": "text", + "content": "[5] T. Berger, “Rate-distortion theory,” Wiley Encyclopedia of Telecommunications, 2003. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 294, + 287 + ], + "type": "text", + "content": "[6] V. K. Goyal, “Theoretical foundations of transform coding,” IEEE Signal Processing Magazine, vol. 18, no. 5, pp. 9–21, 2001. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 290, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 290, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 290, + 295, + 323 + ], + "type": "text", + "content": "[7] J. Balle, V. Laparra, and E. P. Simoncelli, “End-to-end optimized image compression,” arXiv preprint arXiv:1611.01704, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 295, + 369 + ], + "type": "text", + "content": "[8] D. Minnen, J. Balle, and G. D. Toderici, \"Joint autoregressive and hierarchical priors for learned image compression,\" Advances in neural information processing systems, vol. 31, 2018. 2, 5, 6, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 372, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 295, + 415 + ], + "type": "text", + "content": "[9] D. Minnen and S. Singh, \"Channel-wise autoregressive entropy models for learned image compression,\" in 2020 IEEE International Conference on Image Processing (ICIP), pp. 3339-3343, IEEE, 2020. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 418, + 295, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 418, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 418, + 295, + 472 + ], + "type": "text", + "content": "[10] Z. Cheng, H. Sun, M. Takeuchi, and J. Katto, \"Learned image compression with discretized gaussian mixture likelihoods and attention modules,\" in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 7939-7948, 2020. 2, 5, 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 475, + 295, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 295, + 528 + ], + "type": "text", + "content": "[11] D. He, Y. Zheng, B. Sun, Y. Wang, and H. Qin, \"Checkerboard context model for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14771-14780, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 531, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 531, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 295, + 575 + ], + "type": "text", + "content": "[12] W. Jiang and R. Wang, \"MLIC++: Linear complexity multi-reference entropy modeling for learned image compression,\" in ICML 2023 Workshop Neural Compression: From Information Theory to Applications, 2023. 2, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 620 + ], + "type": "text", + "content": "[13] S. Qin, J. Wang, Y. Zhou, B. Chen, T. Luo, B. An, T. Dai, S. Xia, and Y. Wang, \"MambaVC: Learned visual compression with selective state spaces,\" arXiv preprint arXiv:2405.15413, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[14] H. Zhang, L. Li, and D. Liu, \"On uniform scalar quantization for learned image compression,\" arXiv preprint arXiv:2309.17051, 2023. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[15] E. Agustsson, F. Mentzer, M. Tschannen, L. Cavigelli, R. Timofte, L. Benini, and L. V. Gool, \"Soft-to-hard vector quantization for end-to-end learning compressible representations,\" Advances in neural information processing systems, vol. 30, 2017. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[16] Y. Yang, R. Bamler, and S. Mandt, \"Improving inference for neural image compression,\" Advances in Neural Information Processing Systems, vol. 33, pp. 573-584, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 554, + 140 + ], + "type": "text", + "content": "[17] E. Agustsson and L. Theis, \"Universally quantized neural compression,\" Advances in neural information processing systems, vol. 33, pp. 12367-12376, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 143, + 554, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 554, + 186 + ], + "type": "text", + "content": "[18] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Soft then hard: Rethinking the quantization in neural image compression,\" in International Conference on Machine Learning, pp. 3920-3929, PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 189, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 554, + 220 + ], + "type": "text", + "content": "[19] D. P. Kingma, M. Welling, et al., \"An introduction to variational autoencoders,\" Foundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307-392, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 223, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 223, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 223, + 554, + 266 + ], + "type": "text", + "content": "[20] J. Zhao, Y. Kim, K. Zhang, A. Rush, and Y. LeCun, \"Adversarily regularized autoencoders,\" in International conference on machine learning, pp. 5902-5911, PMLR, 2018. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 269, + 554, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 554, + 311 + ], + "type": "text", + "content": "[21] H. Xu, D. Luo, R. Henao, S. Shah, and L. Carin, \"Learning autoencoders with relational regularization,\" in International Conference on Machine Learning, pp. 10576-10586, PMLR, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 314, + 554, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 314, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 314, + 554, + 356 + ], + "type": "text", + "content": "[22] H. Wu and M. Flierl, \"Vector quantization-based regularization for autoencoders,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 6380-6387, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 360, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 554, + 403 + ], + "type": "text", + "content": "[23] L. Le, A. Patterson, and M. White, \"Supervised autoencoders: Improving generalization performance with unsupervised regularizers,\" Advances in neural information processing systems, vol. 31, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 405, + 554, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 554, + 449 + ], + "type": "text", + "content": "[24] T. Ma, J. Chen, and C. Xiao, \"Constrained generation of semantically valid graphs via regularizing variational autoencoders,\" Advances in Neural Information Processing Systems, vol. 31, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 451, + 554, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 554, + 483 + ], + "type": "text", + "content": "[25] S. Sinha and A. B. Dieng, \"Consistency regularization for variational auto-encoders,\" Advances in Neural Information Processing Systems, vol. 34, pp. 12943–12954, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 486, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 554, + 507 + ], + "type": "text", + "content": "[26] H. Ochoa-Dominguez and K. R. Rao, Discrete cosine transform. CRC Press, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 510, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 554, + 552 + ], + "type": "text", + "content": "[27] G. Toderici, S. M. O'Malley, S. J. Hwang, D. Vincent, D. Minnen, S. Baluja, M. Covell, and R. Sukthankar, \"Variable rate image compression with recurrent neural networks,\" arXiv preprint arXiv:1511.06085, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 555, + 554, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 587 + ], + "type": "text", + "content": "[28] H. Liu, T. Chen, P. Guo, Q. Shen, X. Cao, Y. Wang, and Z. Ma, “Non-local attention optimized deep image compression,” arXiv preprint arXiv:1904.09757, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 590, + 554, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 621 + ], + "type": "text", + "content": "[29] M. Lu, P. Guo, H. Shi, C. Cao, and Z. Ma, \"Transformer-based image compression,\" arXiv preprint arXiv:2111.06707, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 667 + ], + "type": "text", + "content": "[30] J. Liu, H. Sun, and J. Katto, “Learned image compression with mixed transformer-cnn architectures,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 14388–14397, 2023. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[31] H. Ma, D. Liu, N. Yan, H. Li, and F. Wu, \"End-to-end optimized versatile image compression with wavelet-like transform,\" IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 3, pp. 1247-1263, 2020. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15581" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[32] X. Zhang and X. Wu, \"LVQAC: Lattice vector quantization coupled with spatially adaptive companding for efficient learned image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10239-10248, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[33] R. Feng, Z. Guo, W. Li, and Z. Chen, \"NVTC: Nonlinear vector transform coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6101-6110, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 225 + ], + "type": "text", + "content": "[34] K. Suhring, M. Schäfer, J. Pfaff, H. Schwarz, D. Marpe, and T. Wiegand, \"Trellis-coded quantization for end-to-end learned image compression,\" in 2022 IEEE International Conference on Image Processing (ICIP), pp. 3306-3310, IEEE, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 227, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 227, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 294, + 271 + ], + "type": "text", + "content": "[35] Z. Ge, S. Ma, W. Gao, J. Pan, and C. Jia, \"NLIC: Nonuniform quantization based learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[36] Z. Guo, Z. Zhang, R. Feng, and Z. Chen, \"Causal contextual prediction for learned image compression,\" IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 4, pp. 2329-2341, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 369 + ], + "type": "text", + "content": "[37] X. Zhu, J. Song, L. Gao, F. Zheng, and H. T. Shen, \"Unified multivariate gaussian mixture for efficient neural image compression,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17612-17621, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[38] H. Fu, F. Liang, J. Lin, B. Li, M. Akbari, J. Liang, G. Zhang, D. Liu, C. Tu, and J. Han, \"Learned image compression with gaussian-laplacian-logistic mixture model and concatenated residual modules,\" IEEE Transactions on Image Processing, vol. 32, pp. 2063-2076, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[39] Y. Hu, W. Yang, and J. Liu, \"Coarse-to-fine hyper-prior modeling for learned image compression,\" in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 11013-11020, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[40] L. Theis, W. Shi, A. Cunningham, and F. Huszár, \"Lossy image compression with compressive autoencoders,\" arXiv preprint arXiv:1703.00395, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 223, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 223, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 223, + 514 + ], + "type": "text", + "content": "[41] I. Goodfellow, \"Deep learning,\" 2016. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 558 + ], + "type": "text", + "content": "[42] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, \"Generative adversarial nets,\" Advances in neural information processing systems, vol. 27, 2014. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[43] D. He, Z. Yang, W. Peng, R. Ma, H. Qin, and Y. Wang, \"ELIC: Efficient learned image compression with unevenly grouped space-channel contextual adaptive coding,\" in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5718-5727, 2022. 5, 6, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[44] J. Bégaint, F. Racapé, S. Feltman, and A. Pushparaja, \"Compressai: a pytorch library and evaluation platform for end-to-end compression research,\" arXiv preprint arXiv:2011.03029, 2020. 5, 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "text", + "content": "[45] W. Jiang, “Unofficial elic.” https://github.com/JiangWeibeta/ELIC, 2022.5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[46] J. Liu, G. Lu, Z. Hu, and D. Xu, \"A unified end-to-end framework for efficient deep image compression,\" arXiv preprint arXiv:2002.03370, 2020. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 488 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 95 + ], + "type": "text", + "content": "[47] D. P. Kingma, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 96, + 553, + 128 + ], + "type": "text", + "content": "[48] E. Kodak, \"Kodak lossless true color image suite (photocd pcd0992),\" URL http://rOk.us/graphics/kodak, vol. 6, p. 2, 1993. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 162 + ], + "type": "text", + "content": "[49] \"6th Challenge on Learned Image Compression.\" https://www.compression.cc/tasks/index.html. Accessed: 2024-11-13. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 164, + 553, + 196 + ], + "type": "text", + "content": "[50] N. Asuni and A. Giachetti, \"Testimages: a large-scale archive for testing visual devices and basic image processing algorithms,\" in STAG, pp. 63-70, 2014. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 553, + 230 + ], + "type": "text", + "content": "[51] T. Zhang, H. Zhang, Y. Li, L. Li, and D. Liu, “Few-shot domain adaptation for learned image compression,” arXiv preprint arXiv:2409.11111, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 232, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 232, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 232, + 553, + 275 + ], + "type": "text", + "content": "[52] Y. Lv, J. Xiang, J. Zhang, W. Yang, X. Han, and W. Yang, \"Dynamic low-rank instance adaptation for universal neural image compression,\" in Proceedings of the 31st ACM International Conference on Multimedia, pp. 632-642, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 276, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 320 + ], + "type": "text", + "content": "[53] J. Yang, S. Shen, H. Yue, and K. Li, \"Implicit transformer network for screen content image continuous superresolution,\" Advances in Neural Information Processing Systems, vol. 34, pp. 13304-13315, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 553, + 365 + ], + "type": "text", + "content": "[54] X. Min, K. Ma, K. Gu, G. Zhai, Z. Wang, and W. Lin, \"Unified blind quality assessment of compressed natural, graphic, and screen content images,\" IEEE Transactions on Image Processing, vol. 26, no. 11, pp. 5462-5474, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 366, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 366, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 315, + 366, + 553, + 421 + ], + "type": "text", + "content": "[55] N. Brancati, A. M. Anniciello, P. Pati, D. Riccio, G. Scognamiglio, G. Jaume, G. De Pietro, M. Di Bonito, A. Fon-cubierta, G. Botti, et al., \"BRACS: A dataset for breast carcinoma subtyping in h&e histology images,\" Database, vol. 2022, p. baac093, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 553, + 464 + ], + "type": "text", + "content": "[56] F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1251–1258, 2017. 8, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 466, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 553, + 488 + ], + "type": "text", + "content": "[57] G. Bjontegaard, \"Calculation of average PSNR differences between RD-curves,\" ITU SG16 Doc. VCEG-M33, 2001. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15582" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_content_list.json b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f34417ee6c07353fb54f02fd9a22a3437dddd243 --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_content_list.json @@ -0,0 +1,1760 @@ +[ + { + "type": "text", + "text": "An Inversion-based Measure of Memorization for Diffusion Models", + "text_level": 1, + "bbox": [ + 158, + 130, + 839, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhe Ma, Qingming Li, Xuhong Zhang, Tianyu Du, Ruixiao Lin, Zonghui Wang* Shouling Ji, Wenzhi Chen* Zhejiang University, Hangzhou, China", + "bbox": [ + 243, + 180, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mz.rs, liqm, zhangxuhong, zjradty, linruixiao, zhwang, sji, chenwz}@zju.edu.cn", + "bbox": [ + 178, + 236, + 813, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The past few years have witnessed substantial advances in image generation powered by diffusion models. However, it was shown that diffusion models are susceptible to training data memorization, raising significant concerns regarding copyright infringement and privacy invasion. This study delves into a rigorous analysis of memorization in diffusion models. We introduce InvMM1, an inversion-based measure of memorization, which is based on inverting a sensitive latent noise distribution accounting for the replication of an image. For accurate estimation of the measure, we propose an adaptive algorithm that balances the normality and sensitivity of the noise distribution. Comprehensive experiments across four datasets, conducted on both unconditional and text-guided diffusion models, demonstrate that InvMM provides a reliable and complete quantification of memorization. Notably, InvMM is commensurable between samples, reveals the true extent of memorization from an adversarial standpoint and implies how memorization differs from membership. In practice, it serves as an auditing tool for developers to reliably assess the risk of memorization, thereby contributing to the enhancement of trustworthiness and privacy-preserving capabilities of diffusion models.", + "bbox": [ + 88, + 318, + 485, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 683, + 220, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion Models (DMs) [22, 50] have shown impressive capabilities in the generation of images [41, 43], videos [6], 3D point cloud [35], etc. These techniques lay the foundation for commercial systems or communities such as Stable Diffusion [43], Midjourney [1], DALL-E [3, 41] and Imagegen [44], which have attracted millions of active users. The popularity of diffusion models stems from their hierarchical denoising process, which offers high stability for training on billions of data [47] and scalability for multimodal conditional generation [41, 44].", + "bbox": [ + 89, + 709, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aade137ef013e9e5b15ef790874e3223b1f161b84c406e1f4c0e41bb4bfffbdb.jpg", + "image_caption": [ + "Figure 1. Schematic illustration of diffusion model inversion, optionally with conditions." + ], + "image_footnote": [], + "bbox": [ + 537, + 287, + 883, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, large-scale datasets used to train the prevalent DMs, e.g., the open-source image-caption dataset LAION [47], are widely acknowledged to contain content that will raise concerns regarding copyright infringement and privacy invasion. For instance, recent reports revealed that LAION could refer to photographers' works without authorization [18] and include private medical photographs [2]. With the uncurated data for training, diffusion models are likely to generate content that infringes the creators' rights or exposes private information [9, 48]. In this work, we focus on memorization in DMs, one of the most critical issues related to training data misuse.", + "bbox": [ + 511, + 503, + 906, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Memorization usually manifests itself at sampling time, when a model generates images closely resembling training images. It can turn into a problem both technically and practically. Technically, heavy memorization indicates failure of generalization to some degree because a probabilistic generative model is supposed to produce novel and diversified images rather than frequently replicate those in the training set. Practically, memorization might also raise ethical and legal concerns. First, even the model developers are authorized to train their model with protected images, the image owners will never expect their works to be exposed to arbitrary users due to memorization as this would cause indisciplinable dissemination, violating the contextual integrity [37]. Second, in past years, text-to-image models", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 107, + 875, + 240, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Maryeon/InvMM", + "bbox": [ + 109, + 887, + 369, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16959", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "have been facing lawsuits for generating derivative images that mimic the style of artists. Compared to derivative generations whose legality is still in pending [46], exact replication of copyrighted images is undisputedly intolerable. Finally, research on privacy protection [25, 38] has proposed to use synthetic data in place of real data to prevent sharing of private information. For this goal, potential memorization must be carefully circumvented.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [9, 51, 52, 58] have confirmed the existence of some heavily memorized training data in (text-guided) DMs. They either draw random samples from the model and then detect risky generations of low diversity [9], search for nearest neighbors in the training set [51, 52], or directly identify abnormal prompts [58]. However, conducting a practical analysis of memorization with these approaches still faces challenges. Transforming the memorization of an image to the presence or absence in samples from the model is not a reliable solution. While certain images do not appear in a set of samples, it does not demonstrate that they are not highly memorized [55]. For DMs with large latent space, adequate samples should be tested to ensure as complete coverage of the latent space as possible, which could be up to millions of tests [9, 58]. Furthermore, things will get much more complicated in conditional DMs, e.g., text-to-image generation. The additional conditions expand the input space of DMs, making the memorization of an image also depend on the specific condition. Existing studies exploit training captions to discover highly memorized training images or perform memorization detection [42, 61]. However, attributing the memorization risk of an image to its coupled training caption is incomplete. It still remains unknown that whether the other images not triggered by their coupled captions are also to a large extent of memorization and to be triggered by potential prompts.", + "bbox": [ + 91, + 215, + 486, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are threefold. (1) To overcome the challenges, we propose an Inversion-based Measure of Memorization (InvMM) for DMs, a continuous instance-level measure that takes an image as input, outputs a scalar quantifying its level of memorization. As shown in Fig. 1, the inversion process locates a sensitive latent noise distribution and optionally another condition distribution for the target image, such that generating with noise and condition drawn from the inverted distributions will produce samples exactly the same as the target image. InvMM is then defined as the minimum KL divergence of the sensitive noise distribution to the standard Gaussian distribution. Compared to existing methods, InvMM is more reliable and complete in that it proactively searches for all potential sensitive inputs and does not rely on limited test cases from the training set (e.g., captions). (2) To calculate the measure, we design an adaptive optimization algorithm motivated by the TCP congestion control protocol, which is demonstrated to be effective to alleviate overestimate and underestimate", + "bbox": [ + 89, + 613, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of memorization. (3) Through comprehensive experiments across four datasets, encompassing both unconditional and text-guided DMs, we present the superiority of InvMM in analyzing memorization. Compared to alternative memorization measures, detection metrics and membership inference metrics, InvMM is commensurable between samples, reflects the true memorization level from an adversarial perspective and clarifies the fundamental difference between memorization and membership. As an auditing tool, it has great potential in discovering highly memorized images, validating safety against memorization for any image, determining appropriate training settings, etc.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 290, + 651, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Memorization of generative models", + "text_level": 1, + "bbox": [ + 511, + 316, + 820, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Memorization has raised wide attention on various generative models, including GAN [17], VAE [30], language models [7] and DM [50]. There have been studies on training algorithms [4, 22, 45] and evaluation protocols [19] to improve the generalization ability of generative models that would potentially circumvent memorization. Empirical and theoretical studies also find that dataset complexity and size [15, 51], training time [54], training data duplication [10, 60] and model capacity [5, 10] are influential factors of memorization. Several works [54, 59] focus on the latent space (noise space) to analyze memorization. Nagarajan et al. [54] demonstrate that when trained on sufficient random noise vectors, a generator is guaranteed to memorize training data. Webster et al. [59] measure overfitting of GANs via the reconstruction error at the optimal noise vector. We take a similar perspective on the latent space to analyze memorization while aim to reach an instance-level measure of memorization rather than reveal overall properties of models. In addition, Van der Burg et al. [55] also propose an instance-level measure of memorization as the changed likelihood of one sample when removing it from the training set. However, the measure is not necessarily comparable between datasets, which is important to understand the true level of memorization.", + "bbox": [ + 511, + 340, + 906, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically for diffusion models, recent research [9, 51, 52] demonstrates that text-to-image diffusion models suffer from memorization. Some heavily memorized images are discovered from numerous generated samples by membership inference [9] or searching for the most similar training images with image matching models [51]. Webster [58] further provides more efficient attacks to extract training images by identifying abnormal prompts. Subsequent works [42, 61] focus on prompt detection that will trigger training image replication and memorization mitigation [12] at inference time. In response to the disclosed threats, Vyas et al. [56] propose a provable copyright protection method to prevent replication of sensitive training", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16960", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "images. Our method improves upon existing methods by being commensurable between different samples, basing on images themselves rather than prompts, and working for any samples (not necessarily training samples). A recent work by Jiang et al. [24] also considers prompt-free memorization. However, their method extends Wen et al.'s findings [61] and is restricted to text-guided diffusion models. Our method is established on the property of latent space thus applies to both unconditional and conditional diffusion models.", + "bbox": [ + 89, + 90, + 483, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Inversion of diffusion models", + "text_level": 1, + "bbox": [ + 89, + 250, + 349, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inversion techniques in diffusion models are widely studied for image editing [16, 36, 63], copyright protection [62], generated content detection [11, 57], etc. Through inversion, the object, style and concept contained in the source images can be compressed in latent noises or input token embeddings (for text-to-image models). Then the inverted latent noises or input token embeddings are utilized to generate similar images that preserve the desired content.", + "bbox": [ + 89, + 272, + 483, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage inversion techniques to analyze memorization in diffusion models. In contrast, we invert a sample to a distribution consisting of all feasible latent vectors rather than to a single vector [11, 16, 36, 57, 62]. We also adopt more efficient training objective for inversion, without the need to iteratively invert a specific sampling procedure [11, 36, 57, 62].", + "bbox": [ + 89, + 393, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 89, + 511, + 220, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Basically, DMs [22, 50] are likelihood-based generative models that learn to maximize the likelihood $p_{\\theta}(\\pmb{x}_0)$ of observed data $\\pmb{x}_0$ . Directly calculating $p_{\\theta}(\\pmb{x}_0)$ is intractable for continuous data such as image. By contrasting to a sequence of latent variables $x_{1:T}$ , the log-likelihood $\\log p_{\\theta}(\\pmb{x}_0)$ can be maximized indirectly by maximizing its variational lower bound [29, 34]:", + "bbox": [ + 89, + 537, + 483, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) = \\log \\int p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0: T}\\right) d \\boldsymbol {x} _ {1: T} (1) \\\\ \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0 : T}\\right)}{q \\left(\\boldsymbol {x} _ {1 : T} \\mid \\boldsymbol {x} _ {0}\\right)} \\right] (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 652, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assume both transitions $p_{\\pmb{\\theta}}(\\pmb{x}_{0:T})$ and $q(\\pmb{x}_{1:T}|\\pmb{x}_0)$ are Markov chains, i.e., $q(\\pmb{x}_{1:T}|\\pmb{x}_0) = \\prod_{t=1}^{T} q(\\pmb{x}_t|\\pmb{x}_{t-1})$ , $p_{\\pmb{\\theta}}(\\pmb{x}_{0:T}) = p(\\pmb{x}_T)\\prod_{t=1}^{T} p_{\\pmb{\\theta}}(\\pmb{x}_{t-1}|\\pmb{x}_t)$ , the variational lower bound in Eq. (2) expands as:", + "bbox": [ + 89, + 726, + 483, + 791 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {x} _ {1}\\right) \\right] \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} [ D _ {\\mathrm {K L}} (q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right)) ] (3) \\\\ - D _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\| p \\left(\\boldsymbol {x} _ {T}\\right)\\right) (4) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 800, + 482, + 896 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p(\\boldsymbol{x}_T)$ is generally assumed to be a simple distribution such as standard Gaussian $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transition $q(\\pmb{x}_{1:T}|\\pmb{x}_0)$ is not learned but replaced by a manually crafted non-parametric diffusion process. The transition $p_{\\theta}(\\pmb{x}_{0:T})$ is the corresponding reverse (or generation, sampling) process. The diffusion process gradually adds Gaussian noises to the input $\\pmb{x}_0$ according to a weight schedule $\\alpha_{1:T}$ :", + "bbox": [ + 511, + 121, + 905, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t - 1}, \\boldsymbol {\\epsilon} _ {t - 1} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 224, + 903, + 241 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {0}, \\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 244, + 903, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ decreases over time and finally to almost zero at the last step $T$ , such that $q(\\pmb{x}_T | \\pmb{x}_0)$ approaches to a standard Gaussian and thus Eq. (4) approximately equals to 0.", + "bbox": [ + 511, + 273, + 905, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "If the diffusion process is divided into sufficient steps, each reverse transition $p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t)$ can also be approximated by a Gaussian transformation with learnable parameters that is trained to match the corresponding posterior diffusion transition $q(\\pmb{x}_{t-1}|\\pmb{x}_t,\\pmb{x}_0)$ . Omitting further details [34], the following simplified objective [22] as an estimate of the variational lower bound is commonly used for training:", + "bbox": [ + 511, + 335, + 905, + 457 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl _ {d e} = \\mathbb {E} _ {\\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t \\sim \\mathcal {U} \\{1, T \\}} \\left[ \\| \\boldsymbol {\\epsilon} _ {0} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 468, + 903, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a neural network that takes the noisy data $\\mathbf{x}_t$ and timestep $t$ as input and predicts the added noise $\\epsilon_0$ . After training, the vanilla sampling procedure starts with a random Gaussian noise $\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ and removes the predicted noise stepwise. However, it is extremely slow to generate an image as it must invoke the network $\\epsilon_{\\theta}$ for $T$ times (typically 1000 steps). To mitigate the problem, a variety of efficient sampling algorithms are proposed, such as DDIM sampler [53] and PLMS sampler [33].", + "bbox": [ + 511, + 497, + 905, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Inversion-based measure", + "text_level": 1, + "bbox": [ + 511, + 647, + 746, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image $\\boldsymbol{x}_0$ , we aim to measure to what extent $\\boldsymbol{x}_0$ is memorized by a diffusion model (DM). We propose to invert the image into a latent noise distribution, such that noise vectors drawn from it will lead to replication of the image $\\boldsymbol{x}_0$ . Our memorization measure is established in this inversion process. In this section, we introduce a sensitivity definition of the inverted noise distribution (Sec. 4.1), elaborate the memorization measure (Sec. 4.2) and give an adaptive algorithm for its calculation (Sec. 4.3). Ultimately, we present an example to measure memorization in conditional DM, specifically text-to-image DM (Sec. 4.4).", + "bbox": [ + 511, + 672, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Sensitivity of latent distribution", + "text_level": 1, + "bbox": [ + 511, + 848, + 792, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diversity of DM generated content lies in the randomness of latent noise vectors. Starting from different noise", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16961", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vectors, the generated contents generally vary. In the context of memorization, this motivates us to search for sensitive noise vectors that will replicate the target image. We then figure out \"how many\" sensitive noise vectors there are and adopt it as a meaningful measure.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Considering the fact that the common latent noise is sampled from a standard Gaussian distribution, we correspondingly assume the sensitive noise vectors follow a distribution $q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)$ with learnable parameters $\\varphi$ . The sensitivity function $S(q_{\\varphi})$ of a noise distribution can be defined as the probability that generated image $G(\\epsilon_0)$ obtained from random noise $\\epsilon_0 \\sim q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)$ replicates the target image:", + "bbox": [ + 89, + 167, + 482, + 275 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left(q _ {\\varphi}\\right) = \\int \\mathbf {1} _ {d \\left(G \\left(\\boldsymbol {\\epsilon} _ {0}\\right), \\boldsymbol {x} _ {0}\\right) \\leq \\beta} q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) d \\boldsymbol {\\epsilon} _ {0} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 285, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_v$ is an indicator function that equals 1 if $v$ is true and 0 otherwise, $d(\\cdot ,\\cdot)$ is a distance metric between two images and $\\beta$ is an appropriate threshold.", + "bbox": [ + 89, + 328, + 482, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The choice of the distance metric depends on specific applications. In applications that accept derivative generation, a strict metric that only detects eidetic generation can be utilized. While under some circumstances even similar generation should be circumvented, a looser metric is needed. In this work, we consider exact replication that the generated image and target image are exactly the same. This consideration to some extent reduces the ambiguity of determining the similarity of two images. Following the investigation of Somepalli et al. [51], we implement the metric as the Euclidean distance between embeddings extracted by the pretrained copy detection model SSCD [39].", + "bbox": [ + 89, + 375, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Quantify memorization", + "text_level": 1, + "bbox": [ + 89, + 566, + 307, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the definition of sensitivity, we propose an Inversion-based Measure of Memorization (InvMM), as the best achievable average normality of $q_{\\varphi}(\\epsilon | \\boldsymbol{x})$ under the condition that $S(q_{\\varphi}) = 1$ :", + "bbox": [ + 89, + 590, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {\\varphi} \\frac {1}{N} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x}\\right) \\| p \\left(\\boldsymbol {\\epsilon} _ {0}\\right)\\right) \\\\ \\text {s . t .} \\mathcal {S} \\left(q _ {\\varphi}\\right) = 1 \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 662, + 480, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the dimensionality of latent noise space and $p(\\epsilon_0)$ is the prior distribution used in DM, which is often selected to be standard Gaussian. The normality of inverted noise distribution is the KL divergence to the prior. A lower value of InvMM means greater degree of memorization.", + "bbox": [ + 89, + 726, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explaining InvMM. Partitioning the normality gives us a more intuitional understanding of the measure. Notice that", + "bbox": [ + 89, + 803, + 483, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} | \\boldsymbol {x} _ {0}\\right) \\| p (\\epsilon_ {0})\\right) = \\\\ \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), p \\left(\\epsilon_ {0}\\right)\\right) - \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right)\\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 861, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8693c96cea934401c724b0022745b09ec19951ada2eeabab4e48522f41f49a67.jpg", + "image_caption": [ + "(a) Optimal" + ], + "image_footnote": [], + "bbox": [ + 566, + 92, + 653, + 180 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c8cef596b7cc85255f22c75a5f3bd0477629e5b8602e7f4b124a296366615f9.jpg", + "image_caption": [ + "(b) Overestimate", + "(c)Underestimate" + ], + "image_footnote": [], + "bbox": [ + 663, + 92, + 754, + 181 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/22f216fac8e19c8856f63f792ea342611b9259c4446177d177231e17ffe96e7d.jpg", + "image_caption": [ + "Figure 2. Illustration of our memorization measure. The gray blocks indicate the distribution of sensitive latent noise vectors." + ], + "image_footnote": [], + "bbox": [ + 772, + 101, + 851, + 175 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0))$ is the cross entropy between the inverted distribution and the prior distribution, $\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0))$ is the differential entropy of the inverted distribution. Minimizing the normality thus encourages a distribution as close to the prior distribution as possible, meanwhile with the largest entropy. For the cross entropy, following the Jensen's inequality, $-\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0)) = \\mathbb{E}_{q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)}[\\log p(\\epsilon_0)] \\leq \\log \\int q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)d\\epsilon_0$ . The term $q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)$ is proportional to the probability of a random noise $\\epsilon_0$ , with probability density $q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)$ , to be sampled from the prior distribution with probability density $p(\\epsilon_0)$ . In the sense, minimizing the cross-entropy improves the probability that sensitive noise vectors regarding an image $\\pmb{x}_0$ are sampled in the normal usage of DM. For the differential entropy, maximizing it will produce a noise distribution involving \"as many\" sensitive noise vectors as possible. In the extreme case, suppose a uniform distribution over the entire latent space having the largest entropy, it means that any noise in the space can cause replication of the target training sample.", + "bbox": [ + 511, + 273, + 906, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "InvMM requires to achieve the best trade-off between improving normality and meeting the sensitivity condition. Different images to varied extent of memorization may reach different equilibrium states. Figure 2 illustrates the optimal solution compared to suboptimal choices. Overestimate of memorization produces a distribution with good normality but low sensitivity (Fig. 2b). Conservative inversion of the latent distribution perfectly meets the sensitivity condition while underestimates memorization (Fig. 2c).", + "bbox": [ + 511, + 579, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Adaptive algorithm", + "text_level": 1, + "bbox": [ + 511, + 733, + 699, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then introduce an adaptive algorithm to solve the inversion problem.", + "bbox": [ + 511, + 758, + 903, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transforming the inversion problem. It is intractable to directly calculate the sensitivity condition. For diffusion models, it can be achieved by minimizing the denoising error $l_{de}$ (see Eq. (7)) as a proxy. As can be expected, when $l_{de}$ gradually decreases, $S(q_{\\varphi})$ will approach 1.", + "bbox": [ + 511, + 791, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, denote the average KL divergence in Eq. (9) by $l_{kl}$ , the optimization problem can be solved by La", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16962", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "grangian method:", + "bbox": [ + 89, + 90, + 210, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\varphi , \\lambda} l _ {d e} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right) + \\lambda l _ {k l} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right), \\lambda \\geq 0 \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 119, + 482, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\geq 0$ is a multiplier. We attach the multiplier to $l_{kl}$ to ensure adequate optimization of the denoising error $l_{de}$ .", + "bbox": [ + 89, + 152, + 482, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Different from the standard $\\epsilon_0$ -prediction denoising error, we adopt $x_0$ -prediction denoising error, which shows more stable inversion efficacy in practice:", + "bbox": [ + 89, + 184, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} l _ {d e} = \\mathbb {E} _ {q _ {\\boldsymbol {\\varphi}} (\\epsilon_ {0} | \\boldsymbol {x} _ {0}), t} \\left[ \\| \\boldsymbol {x} _ {0} - \\boldsymbol {x} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\\\ = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\mathbb {E} _ {q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), t} \\left[ \\| \\epsilon_ {0} - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 239, + 482, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t$ is obtained as in Eq. (6) with added noise drawn from $q_{\\varphi}$ rather than standard Gaussian.", + "bbox": [ + 89, + 301, + 482, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the noise distribution $q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)$ , we assume it to be a multivariate Gaussian $\\mathcal{N}(\\boldsymbol{\\mu},\\boldsymbol{\\sigma}^2)$ with learnable mean $\\pmb{\\mu}$ and diagonal variance $\\sigma^2$ , i.e., $\\varphi = (\\pmb{\\mu},\\pmb{\\sigma}^2)$ . Then InvMM has a closed-form expression $\\frac{1}{2N}\\sum_i(\\pmb{\\mu}_i^2 +\\pmb{\\sigma}_i^2 -\\log \\pmb{\\sigma}_i^2 -1)$ . Although the true distribution of sensitive latent noise might not be fully captured by a Gaussian, we choose this modeling for simplicity and find it effective in experiments.", + "bbox": [ + 89, + 333, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 InvMM: adaptive weight tuning algorithm." + ], + "code_body": "Input: Noise prediction network $\\epsilon_{\\theta}$ , target image $x_0$ , optimization iteration $T$ , observation cycle $C$ , weight increment $\\delta$ , threshold $\\xi$ , learning rate $\\gamma$ \n1: Initialize $\\varphi \\gets (\\mu \\gets 0, \\sigma \\gets 1)$ , $\\lambda \\gets 1, l_{de}^p \\gets +\\infty$ , success $\\gets$ FALSE \n2: for $i \\gets 1$ to $T$ do \n3: $\\mathcal{L} \\gets l_{de}(\\boldsymbol{x}_0; \\varphi) + \\lambda l_{kl}(\\boldsymbol{x}_0; \\varphi)$ \n4: $\\varphi \\gets \\varphi - \\gamma \\nabla_{\\varphi} \\mathcal{L}$ \n5: if $i > 1$ and $i \\% C = 0$ then \n6: if $l_{de} - l_{de}^p < \\xi$ then \n7: $\\lambda \\gets \\lambda / 2$ \n8: else \n9: $\\lambda \\gets \\lambda + \\delta$ \n10: $l_{de}^p \\gets l_{de}$ \n11: if $S(q_\\varphi) = 1$ then \n12: success $\\leftarrow$ TRUE \n13: break \n14: else \n15: $\\lambda \\gets \\lambda + \\delta$ \n16: if success then \n17: return $\\frac{1}{2N} \\sum_i (\\mu_i^2 + \\sigma_i^2 - \\log \\sigma_i^2 - 1)$ \n18: else \n19: return $+\\infty$", + "bbox": [ + 94, + 473, + 486, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive weight tuning. To achieve the best trade-off, the weight $\\lambda$ should be adjusted appropriately for different samples. However, it cannot be optimized via gradient descent together with $\\varphi$ . As the gradient w.r.t. $\\lambda$ is $l_{kl}$ , which", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is always nonnegative, $\\lambda$ will gradually decrease to 0 and $l_{kl}$ will be totally ignored. To solve the problem, we propose an algorithm that allows adaptive tuning of $\\lambda$ during inversion, motivated by the additive increase/multiplicative decrease scheme in congestion control of TCP protocol, as in Algorithm 1, line 6-9.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we monitor the denoising error $l_{de}$ every $C$ steps during the optimization process and divide the weight $\\lambda$ by half if no significant improvement of denoising error is observed in the cycle. Otherwise, $\\lambda$ slowly increases by a small increment $\\delta$ each step. In this way, $\\lambda$ is adaptively adjusted for different samples to balance the normality and denoising error. When the normality is too strong to further reduce the denoising error, thus to meet the sensitivity condition, it is immediately down-weighted. Otherwise, we gradually strengthen it for better normality. We also perform early stop (line 11-13) when the sensitivity condition is reached to avoid unnecessary normality degradation. The sensitivity condition is tested via the Monte Carlo method. Random latent vectors are drawn to generate samples. If the distances between all of them and the target image are lower than the threshold $\\beta$ , then the inversion is successful. There are cases that only partial samples replicate that target image in the end. We still consider such cases as insignificant memorization and flag them as unsuccessful.", + "bbox": [ + 511, + 181, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Text-to-image DM", + "text_level": 1, + "bbox": [ + 513, + 481, + 691, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For text-to-image DM, we model the prompt part as a distribution as well and optimize it simultaneously with the noise distribution. Suppose an input prompt is tokenized into maximum $M$ tokens $\\omega = [\\omega_{1},\\omega_{2},\\dots,\\omega_{M}]$ . Motivated by Guo et al. [20], we learn a $|\\mathcal{V}|$ -dimensional categorical distribution over the vocabulary $\\mathcal{V}$ for each token $\\omega_{i}$ . Let token $\\omega_{i}$ follows a categorical distribution with probabilities $\\pi_{i,j}$ , where $\\sum_{j=1}^{|\\mathcal{V}|}\\pi_{i,j} = 1$ . $\\pi_{i,j}$ can be parameterized and optimized.", + "bbox": [ + 511, + 505, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because sampling from categorical distribution is non-differentiable, we adopt the Gumbel-Softmax reparameterization [23] to obtain smoothed sample $\\tilde{\\omega}_i$ . Given the smoothed token $\\tilde{\\omega}_i$ , it can be converted to token embeddings as a linear combination of vocabulary embeddings $e(\\tilde{\\omega}_i) = \\sum_{j=1}^{|\\mathcal{V}|} \\tilde{\\omega}_{i,j} e(j)$ , where $e(j)$ is the pretrained token embedding of the $j$ -th token. A proper choice of the prompt distribution is critical for inversion on text-to-image DMs. We provide a detailed analysis in Appendix Sec. 9.", + "bbox": [ + 511, + 643, + 905, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 797, + 638, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct a series of experiments from different perspectives to demonstrate the validity of InvMM. First, we conduct controlled studies to verify the effectiveness of the adaptive algorithm (Sec. 5.2). Based on this, we show that InvMM can reasonably elucidate the influence", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16963", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4813e931ea36e647500fa196e4d782a40002627e60c75eb8762b1d7172665c3a.jpg", + "image_caption": [ + "Figure 3. The influence of weight $\\lambda$ on measuring memorization." + ], + "image_footnote": [], + "bbox": [ + 119, + 92, + 457, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of various factors on memorization (Sec. 5.3) and identify highly memorized images (Sec. 5.4). Finally, we discuss how memorization differs from membership (Sec. 5.5).", + "bbox": [ + 89, + 244, + 482, + 291 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment setup", + "text_level": 1, + "bbox": [ + 89, + 300, + 264, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We use four datasets for evaluation, i.e., CIFAR-10 [31], CelebAHQ [26], FFHQ [27] and LAION-Aesthetics V2 [47]. For subsequent experiments, we construct four randomly sampled subsets of CelebAHQ and FFHQ: CelebAHQ-250, CelebAHQ-2.5k, FFHQ-600 and FFHQ-6k. We ensure that the larger subset includes the smaller subset. The LAION dataset is extremely large. To keep computational costs manageable, we collect a subset of images. We first include 76 heavily memorized images discovered by Webster [58]. We call this set of images the confirmed set. In addition, we collected another 50 images with high duplication rates according to the statistics from Webster et al [60]. This set of suspects, named as the suspicious set, are assured to be disjoint with the confirmed set. Images in the confirmed set can be triggered by their training captions while those in the suspicious set cannot. Finally, another 50 randomly sampled images are used for comparison, referred to as the normal set.", + "bbox": [ + 89, + 323, + 483, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion Models. Following the implementations of DDPM [22] and Latent Diffusion Model (LDM) [43], we train DDPM on CIFAR-10 and LDM on the subsets of CelebAHQ and FFHQ. Pre-trained LDMs on the whole CelebAHQ and FFHQ are also used. For text-to-image generation, we utilize Stable Diffusion (SD) v1.4, v2.1 and v3.5 trained on LAION.", + "bbox": [ + 89, + 595, + 483, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inversion. The experiment setting and additional results are deferred to Appendix Sec. 8.", + "bbox": [ + 89, + 700, + 482, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Adaptive algorithm", + "text_level": 1, + "bbox": [ + 89, + 742, + 277, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The core design of Algorithm 1 is the dynamic reweighting of the two loss items. To verify its effectiveness, we compare to fixed weights $\\lambda = 1$ and $\\lambda = 0$ . $\\lambda = 1$ equals DM's original training objective (Eqs. (3) and (4)) and $\\lambda = 0$ totally ignores the normality constraint.", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A comparison of the resulting memorization score under different settings can be observed in Fig. 3. Fixing $\\lambda$ to 1 cannot successfully invert most images: $92.4\\%$ and $99.7\\%$ images of CelebAHQ and FFHQ result in infinite", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/afc7359e3bf733e790b4837efe24af4f9053d825b4c428a1fd9f1919c696c31c.jpg", + "image_caption": [ + "(a) CIFAR-10 (DDPM)" + ], + "image_footnote": [], + "bbox": [ + 552, + 87, + 709, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cbab21a67950b263c4df02e499421375d803ae2b9124a6c9ce6ed16403c1782.jpg", + "image_caption": [ + "(b) LAION (SD)" + ], + "image_footnote": [], + "bbox": [ + 710, + 88, + 867, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8509469aed809b5e7f006632522b02cd7a0bd58dd90fd21021f3bdbd9e722bc6.jpg", + "image_caption": [ + "(c) CelebAHQ (LDM)" + ], + "image_footnote": [], + "bbox": [ + 552, + 224, + 709, + 347 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c07f7eee74e85a37e97ce2de550bb087a13fcd96f61b0629fe15196d7537033b.jpg", + "image_caption": [ + "(d) FFHQ (LDM)", + "Figure 4. Random samples generated using latent noise vectors drawn from inverted distributions. The original training images are in the first column within each block." + ], + "image_footnote": [], + "bbox": [ + 710, + 224, + 867, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "memorization scores. In contrast, both $\\lambda = 0$ and adaptive reweighting successfully invert all images. However, Fixing $\\lambda$ to 0 leads to larger resulting memorization scores, which indicates underestimation of the extent of memorization.", + "bbox": [ + 511, + 431, + 905, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 showcases the generation results with the learned distributions. With the adaptive algorithm, we successfully perform inversion across various datasets and DMs that the randomly generated samples closely resemble the training images. The results form the premise of InvMM that memorization of samples is measured in the state they are made replicable.", + "bbox": [ + 511, + 491, + 905, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Influence factors", + "text_level": 1, + "bbox": [ + 511, + 606, + 679, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both theoretical analysis and empirical investigation point out that dataset complexity [15, 52], dataset size [15, 52], cardinality of latent vector set [54] and data duplication [60] are important factors for memorization. We thus verify the correctness of our measure by utilizing it to describe how memorization trends vary with changes in these factors.", + "bbox": [ + 511, + 628, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset complexity, size and training epoch. We begin by simultaneously validating dataset complexity, size and cardinality of latent vector set on CelebAHQ and FFHQ. Following Feng et al. [15], we quantify dataset complexity with Intrinsic Dimensionality (ID) [32]. The cardinality of latent vector set can be controlled by adjusting training epochs (if in each iteration a fresh random latent vector is picked). However, we argue that previous works [15, 51] obfuscate the effect of dataset size and training epochs, as models are trained for different epochs (probably more) on smaller subsets to the full dataset. Consequently, on smaller subsets, each training image would see different number of", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16964", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4d1f16b45c49edd9e1af1c325a3cefe3dc3f41c6631558aedbc85e55eeadba7.jpg", + "image_caption": [ + "Figure 5. Dataset complexity & size vs. memorization. Models are trained for 943 epochs on CelebAHQ and 446 epochs on FFHQ." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 464, + 195 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16a4ac4915f23c078a5ddf9269ee5bf89ba5c915521825c31780f30a252ad3f4.jpg", + "image_caption": [ + "Figure 6. Training epoch vs. memorization. Models are trained for multiples of the default epochs." + ], + "image_footnote": [], + "bbox": [ + 111, + 262, + 464, + 368 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be263d4e47840684b09f3eef8498f1596034b2e7ae6bbe6be9a46c09052a0fbf.jpg", + "image_caption": [ + "Figure 7. Duplication vs. memorization." + ], + "image_footnote": [], + "bbox": [ + 111, + 422, + 464, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a09685385a9420faf2a8a648f48ffcea348e5994d9418f4e1033878785a5823b.jpg", + "image_caption": [ + "Figure 8. Dedduplication vs. memorization." + ], + "image_footnote": [], + "bbox": [ + 111, + 558, + 464, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "latent vectors. Therefore, in our experiment, we decouple these two factors by training for the same epochs as on the full dataset when studying the influence of dataset size.", + "bbox": [ + 89, + 718, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be observed in Fig. 5, dataset complexity and size seem not present high correlation with memorization, as indicated by our measure. When trained for longer time (Fig. 6), each model presents a larger extent of memorization. The results suggest that dataset complexity and size are not independent determining factors, and InvMM accurately reflects the impact of these three factors.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data duplication. We then validate the effect of data duplication in DMs' training set through our measure. We", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae3932e3f976816946359a17afca31ddfed9745fcac3bf7aed08d4897ecef76c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricDDPM (CIFAR-10)SD v1.4 (LAION)
AUCTPR@1%FPRAUCTPR@1%FPR
ε0-loss0.9810.4140.9550.000
x0-loss0.1800.0000.9750.342
MLOO [55]0.4180.000--
Wen [61]--0.9990.868
Ren [42]--0.9750.355
InvMM0.9970.8990.9740.908
", + "bbox": [ + 539, + 88, + 880, + 204 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Detection performance on DDPM and SD v1.4.", + "bbox": [ + 540, + 214, + 877, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/738b0ece1612e53c175970844ff5b77c61f3897c419cdf15b176aebb71afb08d.jpg", + "image_caption": [ + "(a)", + "Figure 9. (a) Memorization score distribution on three datasets. (b) Adversarial prompt for an example from the suspicious set." + ], + "image_footnote": [], + "bbox": [ + 540, + 244, + 705, + 371 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c4d49bbcfca264b237986853ff93da2e73d281ac130b4bdfad5392c8e1aefbd.jpg", + "image_caption": [ + "Prompt:", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 725, + 244, + 883, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "<|endoftext|>eaglegotcha \nforever walaccidents concludes \ntrailer disodana crying ner banana!# conan \"\"gazing \n<|endoftext|>shark", + "bbox": [ + 732, + 318, + 874, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "train a series of LDMs on CelebAHQ-2.5k and FFHQ-6k, with their subsets CelebAHQ-250 and FFHQ-600 duplicated 5, 10, 15 and 20 times. Figure 7 depicts the memorization scores of duplicated images. As can be observed, data duplication rapidly worsens memorization in DMs and our measure is sensitive to these changes. With increasing duplication times, the larger FFHQ-6k dataset presents a stronger resistance to memorization.", + "bbox": [ + 511, + 445, + 903, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For CIFAR-10 and LAION which originally contain duplicates, we observe the effect of dedduplication. We deduplicate CIFAR-10 with SSCD embeddings and DBSCAN clustering [14]. This results in 666 identified clusters, only one image of which is retained for training. The memorization is evaluated on the retained 666 images before and after dedduplication. On LAION, as compared to SD v1.4, SD v2.1 and v3.5 deduplicate the training set. Memorization is evaluated on the confirmed and suspicious sets. Figure 8 shows the results, with infinite scores on the top. As can be observed, InvMM clearly exhibits the benefits that dedduplication brings.", + "bbox": [ + 511, + 566, + 906, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Detection", + "text_level": 1, + "bbox": [ + 511, + 757, + 624, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For InvMM to be a correct measure of memorization, it should be capable of distinguishing highly memorized images from normal ones. To construct a reliable test set, we conduct a nearest neighbor test following previous studies [9, 51, 59] to discover highly memorized images from DM's training set, i.e., highly memorized images consist of training images that can appear at sampling time. Specifically, we consider DDPM trained on CIFAR-10, LDMs", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16965", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecf7f3b0f8c5a99cfe6a7764d4f1cbc14195af6ccea105c77a61f434163181fd.jpg", + "image_caption": [ + "Figure 10. Performance comparison of membership inference on FFHQ-6k, across different training epochs and duplication times." + ], + "image_footnote": [], + "bbox": [ + 117, + 93, + 285, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10e9dcea9605bbf1ebf04c48da64361236540c32a9fd50cfa70310663eccc841.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 93, + 454, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trained on CelebAHQ and FFHQ, and SDs trained on LAION. With their official implementations, only DDPM and SD v1.4 present obvious replications, which are thus used for evaluation. For DDPM, the 99 discovered training images appearing in 1M random samples are used as positive samples and another 1k different training images are included as negative samples; for SD v1.4, in accordance with the setting of previous works, we deem images in the confirmed set as positive and those in the suspicious and normal set as negative, considering that captions in the confirmed set can cause replication of their paired images while those in the suspicious and normal set cannot.", + "bbox": [ + 88, + 261, + 482, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1 presents the detection performance assessed via AUC and TPR@1%FPR. For the unconditional DDPM on CIFAR-10, training loss metrics (the first 2 rows) show decent results and the original $\\epsilon_0$ -loss works much better than the $x_0$ -loss. The differential metric $M^{\\mathrm{LOO}}$ is less effective than the $\\epsilon_0$ -loss and InvMM. The results indicate that the compared metrics are not able to quantify memorization, as they are not fundamentally commensurable between samples. InvMM measures memorization of each sample by contrasting its sensitive noise distribution to a common standard Gaussian, thus having the same scale.", + "bbox": [ + 88, + 445, + 482, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For SD v1.4, we compare InvMM against training loss metrics and two state-of-the-art detection metrics [42, 61]. Since all compared approaches require access to image captions, we implement InvMM with fixed prompts during inversion. As shown in Tab. 1, our measure achieves performance comparable to state-of-the-art metrics. The results of compared metrics seem to demonstrate a strong ability to distinguish memorization thus far. However, we argue that the prompt-based task formulation employed in existing studies offers an overly optimistic understanding of memorization. It is because these techniques are actually to predict whether the captions will replicate their paired images, while overlook potential prompts that might cause replication. As observed in Fig. 9a, when simultaneously inverting the input prompt, our measure reveals that some images in the suspicious set are also memorized to a similar extent as those in the confirmed set (due to duplication). Applying GCG [64] attack on images with the lowest InvMM scores in the suspicious set provides additional evidence, as illus", + "bbox": [ + 91, + 613, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "trated in Fig. 9b: images previously deemed “not memorized” according to their training captions can indeed be replicated by potential adversarial prompts (see Appendix Fig. 13 for more examples). Hence, going beyond the confines of existing research, our measure describes worst-case memorization from an adversarial standpoint, making it more reliable in security-critical scenarios.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Comparison to membership", + "text_level": 1, + "bbox": [ + 511, + 205, + 764, + 222 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A close topic to memorization is membership inference [8, 13, 49], which reveals the leakage risk of membership information that whether an image is used for training. The common implicit assumption is that member samples are \"memorized\" more than hold-out samples [13], rendering them distinguishable through statistics like (variants of) training loss or differential loss [8]. Interestingly, the principle of InvMM can provide an explicit clarification of the difference between membership and memorization.", + "bbox": [ + 511, + 227, + 903, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A comparison to a recent membership inference metric SecMI [13] on FFHQ-6k, as shown in Fig. 10, presents a positive correlation between them. As factors that exacerbate memorization changes, both SecMI and InvMM show increasing performance. There is also a performance gap between SecMI and InvMM, indicating that certain samples from the hold-out set are memorized more than some member samples. This mainly results from the gap between model's metric (training loss) and human's metric (perceptual similarity). Membership inference leverages the model's metric for prediction, while memorization should resort to perceptual metric, e.g., simulated by SSCD. What level of loss leads to image replication depends both on the model and image itself. Low training loss does not necessarily lead to data replication, but would probably distinguish training data from never accessed. On the contrary, high training loss signifying non-membership may still be enough for replication. Therefore, membership does not deterministically lead to high memorization and vice versa. InvMM differs from membership metrics by focusing on the perceptual level and adaptively handles the specific loss landscapes of different samples.", + "bbox": [ + 511, + 364, + 903, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we dive into measuring memorization for diffusion models. Comprehensive experiments demonstrate the ability of the proposed measure to correctly quantify memorization. In short, an image to a higher extent of memorization is able to be inverted to a more standard and diverse latent distribution. However, our method is also limited by its simple modeling of the sensitive latent noise as a multivariate normal. During our research, we observed many clues confirming that the true distribution is complex. Future work might explore more advanced methods, e.g., a generative network, to fit the sensitive noise.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16966", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly supported by the National Key Research and Development Program of China under No. 2022YFB3102100, NSFC under No. U244120033, U24A20336, 62172243, 62402425 and 62402418, the China Postdoctoral Science Foundation under No. 2024M762829, the Zhejiang Provincial Natural Science Foundation under No. LD24F020002, the “Pioneer and Leading Goose” R&D Program of Zhejiang under No. 2025C01082, 2025C02033, 2025C02263 and 2025C02034, the Zhejiang Provincial Priority-Funded Postdoctoral Research Project under No. ZJ2024001 and the Key R&D Program of Ningbo under No. 2024Z115.", + "bbox": [ + 89, + 114, + 485, + 297 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 310, + 187, + 325 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Midjourney. https://www.midjourney.com. 1", + "[2] Artist finds private medical record photos in popular AI training data set. https://arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-photos-in-popular-ai-training-data-set/.1", + "[3] DALL·E 3. https://openai.com/dall-e-3.1", + "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2", + "[5] Sanjeev Arora and Yi Zhang. Do gans actually learn the distribution? an empirical study. arXiv preprint arXiv:1706.08224, 2017. 2", + "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2", + "[8] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 8", + "[9] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 7" + ], + "bbox": [ + 99, + 335, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[11] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10759-10769, 2024. 3", + "[12] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 2", + "[13] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? In International Conference on Machine Learning, pages 8717-8730. PMLR, 2023. 8", + "[14] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 7", + "[15] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701-6710, 2021. 2, 6", + "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[18] Andres Guadamuz. Photographer sues LAION for copyright infringement. https://www.technollama.co.uk/ photographer-sues-laion-for-copyright-infringement, 2023.1", + "[19] Ishaan Gulrajani, Colin Raffel, and Luke Metz. Towards gan benchmarks which require generalization. In International Conference on Learning Representations, 2019. 2", + "[20] Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. Gradient-based adversarial attacks against text transformers. arXiv preprint arXiv:2104.13733, 2021. 5", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 6", + "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5", + "[24] Yue Jiang, Haokun Lin, Yang Bai, Bo Peng, Zhili Liu, Yueming Lyu, Yong Yang, Xingzheng, and Jing Dong. Image-level memorization detection via inversion-based inference" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16967", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perturbation. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[25] James Jordon, Jinsung Yoon, and Mihaela Van Der Schaar. Pate-gan: Generating synthetic data with differential privacy guarantees. In International Conference on Learning Representations, 2018. 2", + "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6", + "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[28] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2", + "[29] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 3", + "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2", + "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[32] Elizaveta Levina and Peter Bickel. Maximum likelihood estimation of intrinsic dimension. Advances in neural information processing systems, 17, 2004. 6", + "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 3", + "[34] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3", + "[35] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 1", + "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3", + "[37] Helen Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119, 2004. 1", + "[38] Kai Packhäuser, Lukas Folle, Florian Thamm, and Andreas Maier. Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classification systems. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 2", + "[39] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 4", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[42] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, pages 340-356. Springer, 2024. 2, 7, 8", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6", + "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2", + "[46] Pamela Samuelson. Generative ai meets copyright. Science, 381(6654):158-161, 2023. 2", + "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1, 6", + "[48] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1", + "[49] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 8", + "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2, 3", + "[51] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery?" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16968", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 2, 4, 6, 7", + "[52] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 2, 6", + "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 2", + "[54] N Vaishnavh, C Raffel, and IJ Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2, 6", + "[55] Gerrit van den Burg and Chris Williams. On memorization in probabilistic deep generative models. Advances in Neural Information Processing Systems, 34:27916-27928, 2021. 2, 7", + "[56] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 2", + "[57] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 3", + "[58] Ryan Webster. A reproducible extraction of training images from diffusion models. CoRR, abs/2305.08694, 2023. 2, 6", + "[59] Ryan Webster, Julien Rabin, Loic Simon, and Frédéric Jurie. Detecting overfitting of deep generative networks via latent recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11273-11282, 2019. 2, 7", + "[60] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of LAION-2B. arXiv preprint arXiv:2303.12733, 2023. 2, 6", + "[61] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 7, 8", + "[62] Xiaoyu Wu, Yang Hua, Chumeng Liang, Jiaru Zhang, Hao Wang, Tao Song, and Haibing Guan. Cgi-dm: Digital copyright authentication for diffusion models via contrasting gradient inversion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10812-10821. IEEE Computer Society, 2024. 3", + "[63] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3", + "[64] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 8, 2" + ], + "bbox": [ + 91, + 90, + 482, + 869 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16969", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_model.json b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7115f118d6af4a89d5447b5752eca8106ce5ddd5 --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.16, + 0.131, + 0.84, + 0.152 + ], + "angle": 0, + "content": "An Inversion-based Measure of Memorization for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.181, + 0.751, + 0.234 + ], + "angle": 0, + "content": "Zhe Ma, Qingming Li, Xuhong Zhang, Tianyu Du, Ruixiao Lin, Zonghui Wang* Shouling Ji, Wenzhi Chen* Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.237, + 0.815, + 0.253 + ], + "angle": 0, + "content": "{mz.rs, liqm, zhangxuhong, zjradty, linruixiao, zhwang, sji, chenwz}@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.486, + 0.653 + ], + "angle": 0, + "content": "The past few years have witnessed substantial advances in image generation powered by diffusion models. However, it was shown that diffusion models are susceptible to training data memorization, raising significant concerns regarding copyright infringement and privacy invasion. This study delves into a rigorous analysis of memorization in diffusion models. We introduce InvMM1, an inversion-based measure of memorization, which is based on inverting a sensitive latent noise distribution accounting for the replication of an image. For accurate estimation of the measure, we propose an adaptive algorithm that balances the normality and sensitivity of the noise distribution. Comprehensive experiments across four datasets, conducted on both unconditional and text-guided diffusion models, demonstrate that InvMM provides a reliable and complete quantification of memorization. Notably, InvMM is commensurable between samples, reveals the true extent of memorization from an adversarial standpoint and implies how memorization differs from membership. In practice, it serves as an auditing tool for developers to reliably assess the risk of memorization, thereby contributing to the enhancement of trustworthiness and privacy-preserving capabilities of diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.222, + 0.699 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Diffusion Models (DMs) [22, 50] have shown impressive capabilities in the generation of images [41, 43], videos [6], 3D point cloud [35], etc. These techniques lay the foundation for commercial systems or communities such as Stable Diffusion [43], Midjourney [1], DALL-E [3, 41] and Imagegen [44], which have attracted millions of active users. The popularity of diffusion models stems from their hierarchical denoising process, which offers high stability for training on billions of data [47] and scalability for multimodal conditional generation [41, 44]." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.289, + 0.885, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.446, + 0.907, + 0.473 + ], + "angle": 0, + "content": "Figure 1. Schematic illustration of diffusion model inversion, optionally with conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.907, + 0.685 + ], + "angle": 0, + "content": "However, large-scale datasets used to train the prevalent DMs, e.g., the open-source image-caption dataset LAION [47], are widely acknowledged to contain content that will raise concerns regarding copyright infringement and privacy invasion. For instance, recent reports revealed that LAION could refer to photographers' works without authorization [18] and include private medical photographs [2]. With the uncurated data for training, diffusion models are likely to generate content that infringes the creators' rights or exposes private information [9, 48]. In this work, we focus on memorization in DMs, one of the most critical issues related to training data misuse." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Memorization usually manifests itself at sampling time, when a model generates images closely resembling training images. It can turn into a problem both technically and practically. Technically, heavy memorization indicates failure of generalization to some degree because a probabilistic generative model is supposed to produce novel and diversified images rather than frequently replicate those in the training set. Practically, memorization might also raise ethical and legal concerns. First, even the model developers are authorized to train their model with protected images, the image owners will never expect their works to be exposed to arbitrary users due to memorization as this would cause indisciplinable dissemination, violating the contextual integrity [37]. Second, in past years, text-to-image models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.241, + 0.888 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.37, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/Maryeon/InvMM" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.37, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16959" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "have been facing lawsuits for generating derivative images that mimic the style of artists. Compared to derivative generations whose legality is still in pending [46], exact replication of copyrighted images is undisputedly intolerable. Finally, research on privacy protection [25, 38] has proposed to use synthetic data in place of real data to prevent sharing of private information. For this goal, potential memorization must be carefully circumvented." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.217, + 0.487, + 0.61 + ], + "angle": 0, + "content": "Recent studies [9, 51, 52, 58] have confirmed the existence of some heavily memorized training data in (text-guided) DMs. They either draw random samples from the model and then detect risky generations of low diversity [9], search for nearest neighbors in the training set [51, 52], or directly identify abnormal prompts [58]. However, conducting a practical analysis of memorization with these approaches still faces challenges. Transforming the memorization of an image to the presence or absence in samples from the model is not a reliable solution. While certain images do not appear in a set of samples, it does not demonstrate that they are not highly memorized [55]. For DMs with large latent space, adequate samples should be tested to ensure as complete coverage of the latent space as possible, which could be up to millions of tests [9, 58]. Furthermore, things will get much more complicated in conditional DMs, e.g., text-to-image generation. The additional conditions expand the input space of DMs, making the memorization of an image also depend on the specific condition. Existing studies exploit training captions to discover highly memorized training images or perform memorization detection [42, 61]. However, attributing the memorization risk of an image to its coupled training caption is incomplete. It still remains unknown that whether the other images not triggered by their coupled captions are also to a large extent of memorization and to be triggered by potential prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Our contributions are threefold. (1) To overcome the challenges, we propose an Inversion-based Measure of Memorization (InvMM) for DMs, a continuous instance-level measure that takes an image as input, outputs a scalar quantifying its level of memorization. As shown in Fig. 1, the inversion process locates a sensitive latent noise distribution and optionally another condition distribution for the target image, such that generating with noise and condition drawn from the inverted distributions will produce samples exactly the same as the target image. InvMM is then defined as the minimum KL divergence of the sensitive noise distribution to the standard Gaussian distribution. Compared to existing methods, InvMM is more reliable and complete in that it proactively searches for all potential sensitive inputs and does not rely on limited test cases from the training set (e.g., captions). (2) To calculate the measure, we design an adaptive optimization algorithm motivated by the TCP congestion control protocol, which is demonstrated to be effective to alleviate overestimate and underestimate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.273 + ], + "angle": 0, + "content": "of memorization. (3) Through comprehensive experiments across four datasets, encompassing both unconditional and text-guided DMs, we present the superiority of InvMM in analyzing memorization. Compared to alternative memorization measures, detection metrics and membership inference metrics, InvMM is commensurable between samples, reflects the true memorization level from an adversarial perspective and clarifies the fundamental difference between memorization and membership. As an auditing tool, it has great potential in discovering highly memorized images, validating safety against memorization for any image, determining appropriate training settings, etc." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.291, + 0.652, + 0.307 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.821, + 0.334 + ], + "angle": 0, + "content": "2.1. Memorization of generative models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Memorization has raised wide attention on various generative models, including GAN [17], VAE [30], language models [7] and DM [50]. There have been studies on training algorithms [4, 22, 45] and evaluation protocols [19] to improve the generalization ability of generative models that would potentially circumvent memorization. Empirical and theoretical studies also find that dataset complexity and size [15, 51], training time [54], training data duplication [10, 60] and model capacity [5, 10] are influential factors of memorization. Several works [54, 59] focus on the latent space (noise space) to analyze memorization. Nagarajan et al. [54] demonstrate that when trained on sufficient random noise vectors, a generator is guaranteed to memorize training data. Webster et al. [59] measure overfitting of GANs via the reconstruction error at the optimal noise vector. We take a similar perspective on the latent space to analyze memorization while aim to reach an instance-level measure of memorization rather than reveal overall properties of models. In addition, Van der Burg et al. [55] also propose an instance-level measure of memorization as the changed likelihood of one sample when removing it from the training set. However, the measure is not necessarily comparable between datasets, which is important to understand the true level of memorization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Specifically for diffusion models, recent research [9, 51, 52] demonstrates that text-to-image diffusion models suffer from memorization. Some heavily memorized images are discovered from numerous generated samples by membership inference [9] or searching for the most similar training images with image matching models [51]. Webster [58] further provides more efficient attacks to extract training images by identifying abnormal prompts. Subsequent works [42, 61] focus on prompt detection that will trigger training image replication and memorization mitigation [12] at inference time. In response to the disclosed threats, Vyas et al. [56] propose a provable copyright protection method to prevent replication of sensitive training" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16960" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.242 + ], + "angle": 0, + "content": "images. Our method improves upon existing methods by being commensurable between different samples, basing on images themselves rather than prompts, and working for any samples (not necessarily training samples). A recent work by Jiang et al. [24] also considers prompt-free memorization. However, their method extends Wen et al.'s findings [61] and is restricted to text-guided diffusion models. Our method is established on the property of latent space thus applies to both unconditional and conditional diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.351, + 0.266 + ], + "angle": 0, + "content": "2.2. Inversion of diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.394 + ], + "angle": 0, + "content": "Inversion techniques in diffusion models are widely studied for image editing [16, 36, 63], copyright protection [62], generated content detection [11, 57], etc. Through inversion, the object, style and concept contained in the source images can be compressed in latent noises or input token embeddings (for text-to-image models). Then the inverted latent noises or input token embeddings are utilized to generate similar images that preserve the desired content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.501 + ], + "angle": 0, + "content": "We leverage inversion techniques to analyze memorization in diffusion models. In contrast, we invert a sample to a distribution consisting of all feasible latent vectors rather than to a single vector [11, 16, 36, 57, 62]. We also adopt more efficient training objective for inversion, without the need to iteratively invert a specific sampling procedure [11, 36, 57, 62]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.221, + 0.53 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Basically, DMs [22, 50] are likelihood-based generative models that learn to maximize the likelihood \\( p_{\\theta}(\\pmb{x}_0) \\) of observed data \\( \\pmb{x}_0 \\). Directly calculating \\( p_{\\theta}(\\pmb{x}_0) \\) is intractable for continuous data such as image. By contrasting to a sequence of latent variables \\( x_{1:T} \\), the log-likelihood \\( \\log p_{\\theta}(\\pmb{x}_0) \\) can be maximized indirectly by maximizing its variational lower bound [29, 34]:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.653, + 0.483, + 0.719 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) = \\log \\int p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0: T}\\right) d \\boldsymbol {x} _ {1: T} (1) \\\\ \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0 : T}\\right)}{q \\left(\\boldsymbol {x} _ {1 : T} \\mid \\boldsymbol {x} _ {0}\\right)} \\right] (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.484, + 0.792 + ], + "angle": 0, + "content": "Assume both transitions \\( p_{\\pmb{\\theta}}(\\pmb{x}_{0:T}) \\) and \\( q(\\pmb{x}_{1:T}|\\pmb{x}_0) \\) are Markov chains, i.e., \\( q(\\pmb{x}_{1:T}|\\pmb{x}_0) = \\prod_{t=1}^{T} q(\\pmb{x}_t|\\pmb{x}_{t-1}) \\), \\( p_{\\pmb{\\theta}}(\\pmb{x}_{0:T}) = p(\\pmb{x}_T)\\prod_{t=1}^{T} p_{\\pmb{\\theta}}(\\pmb{x}_{t-1}|\\pmb{x}_t) \\), the variational lower bound in Eq. (2) expands as:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.801, + 0.483, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {x} _ {1}\\right) \\right] \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} [ D _ {\\mathrm {K L}} (q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right)) ] (3) \\\\ - D _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\| p \\left(\\boldsymbol {x} _ {T}\\right)\\right) (4) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "where \\(p(\\boldsymbol{x}_T)\\) is generally assumed to be a simple distribution such as standard Gaussian \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.213 + ], + "angle": 0, + "content": "The transition \\( q(\\pmb{x}_{1:T}|\\pmb{x}_0) \\) is not learned but replaced by a manually crafted non-parametric diffusion process. The transition \\( p_{\\theta}(\\pmb{x}_{0:T}) \\) is the corresponding reverse (or generation, sampling) process. The diffusion process gradually adds Gaussian noises to the input \\( \\pmb{x}_0 \\) according to a weight schedule \\( \\alpha_{1:T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.225, + 0.905, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t - 1}, \\boldsymbol {\\epsilon} _ {t - 1} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.246, + 0.905, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {0}, \\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.906, + 0.336 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i\\) decreases over time and finally to almost zero at the last step \\(T\\), such that \\(q(\\pmb{x}_T | \\pmb{x}_0)\\) approaches to a standard Gaussian and thus Eq. (4) approximately equals to 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.458 + ], + "angle": 0, + "content": "If the diffusion process is divided into sufficient steps, each reverse transition \\( p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t) \\) can also be approximated by a Gaussian transformation with learnable parameters that is trained to match the corresponding posterior diffusion transition \\( q(\\pmb{x}_{t-1}|\\pmb{x}_t,\\pmb{x}_0) \\). Omitting further details [34], the following simplified objective [22] as an estimate of the variational lower bound is commonly used for training:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.469, + 0.905, + 0.488 + ], + "angle": 0, + "content": "\\[\nl _ {d e} = \\mathbb {E} _ {\\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t \\sim \\mathcal {U} \\{1, T \\}} \\left[ \\| \\boldsymbol {\\epsilon} _ {0} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.906, + 0.635 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a neural network that takes the noisy data \\(\\mathbf{x}_t\\) and timestep \\(t\\) as input and predicts the added noise \\(\\epsilon_0\\). After training, the vanilla sampling procedure starts with a random Gaussian noise \\(\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) and removes the predicted noise stepwise. However, it is extremely slow to generate an image as it must invoke the network \\(\\epsilon_{\\theta}\\) for \\(T\\) times (typically 1000 steps). To mitigate the problem, a variety of efficient sampling algorithms are proposed, such as DDIM sampler [53] and PLMS sampler [33]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.748, + 0.664 + ], + "angle": 0, + "content": "4. Inversion-based measure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Given an image \\( \\boldsymbol{x}_0 \\), we aim to measure to what extent \\( \\boldsymbol{x}_0 \\) is memorized by a diffusion model (DM). We propose to invert the image into a latent noise distribution, such that noise vectors drawn from it will lead to replication of the image \\( \\boldsymbol{x}_0 \\). Our memorization measure is established in this inversion process. In this section, we introduce a sensitivity definition of the inverted noise distribution (Sec. 4.1), elaborate the memorization measure (Sec. 4.2) and give an adaptive algorithm for its calculation (Sec. 4.3). Ultimately, we present an example to measure memorization in conditional DM, specifically text-to-image DM (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.794, + 0.865 + ], + "angle": 0, + "content": "4.1. Sensitivity of latent distribution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The diversity of DM generated content lies in the randomness of latent noise vectors. Starting from different noise" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16961" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "vectors, the generated contents generally vary. In the context of memorization, this motivates us to search for sensitive noise vectors that will replicate the target image. We then figure out \"how many\" sensitive noise vectors there are and adopt it as a meaningful measure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Considering the fact that the common latent noise is sampled from a standard Gaussian distribution, we correspondingly assume the sensitive noise vectors follow a distribution \\(q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)\\) with learnable parameters \\(\\varphi\\). The sensitivity function \\(S(q_{\\varphi})\\) of a noise distribution can be defined as the probability that generated image \\(G(\\epsilon_0)\\) obtained from random noise \\(\\epsilon_0 \\sim q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)\\) replicates the target image:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.286, + 0.483, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left(q _ {\\varphi}\\right) = \\int \\mathbf {1} _ {d \\left(G \\left(\\boldsymbol {\\epsilon} _ {0}\\right), \\boldsymbol {x} _ {0}\\right) \\leq \\beta} q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) d \\boldsymbol {\\epsilon} _ {0} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.375 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_v\\) is an indicator function that equals 1 if \\(v\\) is true and 0 otherwise, \\(d(\\cdot ,\\cdot)\\) is a distance metric between two images and \\(\\beta\\) is an appropriate threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.484, + 0.557 + ], + "angle": 0, + "content": "The choice of the distance metric depends on specific applications. In applications that accept derivative generation, a strict metric that only detects eidetic generation can be utilized. While under some circumstances even similar generation should be circumvented, a looser metric is needed. In this work, we consider exact replication that the generated image and target image are exactly the same. This consideration to some extent reduces the ambiguity of determining the similarity of two images. Following the investigation of Somepalli et al. [51], we implement the metric as the Euclidean distance between embeddings extracted by the pretrained copy detection model SSCD [39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.308, + 0.584 + ], + "angle": 0, + "content": "4.2. Quantify memorization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Based on the definition of sensitivity, we propose an Inversion-based Measure of Memorization (InvMM), as the best achievable average normality of \\( q_{\\varphi}(\\epsilon | \\boldsymbol{x}) \\) under the condition that \\( S(q_{\\varphi}) = 1 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.664, + 0.482, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {\\varphi} \\frac {1}{N} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x}\\right) \\| p \\left(\\boldsymbol {\\epsilon} _ {0}\\right)\\right) \\\\ \\text {s . t .} \\mathcal {S} \\left(q _ {\\varphi}\\right) = 1 \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.727, + 0.483, + 0.803 + ], + "angle": 0, + "content": "where \\(N\\) is the dimensionality of latent noise space and \\(p(\\epsilon_0)\\) is the prior distribution used in DM, which is often selected to be standard Gaussian. The normality of inverted noise distribution is the KL divergence to the prior. A lower value of InvMM means greater degree of memorization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.804, + 0.484, + 0.833 + ], + "angle": 0, + "content": "Explaining InvMM. Partitioning the normality gives us a more intuitional understanding of the measure. Notice that" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.862, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} | \\boldsymbol {x} _ {0}\\right) \\| p (\\epsilon_ {0})\\right) = \\\\ \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), p \\left(\\epsilon_ {0}\\right)\\right) - \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right)\\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.568, + 0.093, + 0.655, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.189, + 0.642, + 0.202 + ], + "angle": 0, + "content": "(a) Optimal" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.093, + 0.755, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.189, + 0.756, + 0.201 + ], + "angle": 0, + "content": "(b) Overestimate" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.102, + 0.852, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.19, + 0.859, + 0.201 + ], + "angle": 0, + "content": "(c)Underestimate" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.241 + ], + "angle": 0, + "content": "Figure 2. Illustration of our memorization measure. The gray blocks indicate the distribution of sensitive latent noise vectors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.577 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0))\\) is the cross entropy between the inverted distribution and the prior distribution, \\(\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0))\\) is the differential entropy of the inverted distribution. Minimizing the normality thus encourages a distribution as close to the prior distribution as possible, meanwhile with the largest entropy. For the cross entropy, following the Jensen's inequality, \\(-\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0)) = \\mathbb{E}_{q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)}[\\log p(\\epsilon_0)] \\leq \\log \\int q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)d\\epsilon_0\\). The term \\(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)\\) is proportional to the probability of a random noise \\(\\epsilon_0\\), with probability density \\(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)\\), to be sampled from the prior distribution with probability density \\(p(\\epsilon_0)\\). In the sense, minimizing the cross-entropy improves the probability that sensitive noise vectors regarding an image \\(\\pmb{x}_0\\) are sampled in the normal usage of DM. For the differential entropy, maximizing it will produce a noise distribution involving \"as many\" sensitive noise vectors as possible. In the extreme case, suppose a uniform distribution over the entire latent space having the largest entropy, it means that any noise in the space can cause replication of the target training sample." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.716 + ], + "angle": 0, + "content": "InvMM requires to achieve the best trade-off between improving normality and meeting the sensitivity condition. Different images to varied extent of memorization may reach different equilibrium states. Figure 2 illustrates the optimal solution compared to suboptimal choices. Overestimate of memorization produces a distribution with good normality but low sensitivity (Fig. 2b). Conservative inversion of the latent distribution perfectly meets the sensitivity condition while underestimates memorization (Fig. 2c)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.734, + 0.7, + 0.75 + ], + "angle": 0, + "content": "4.3. Adaptive algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.905, + 0.789 + ], + "angle": 0, + "content": "We then introduce an adaptive algorithm to solve the inversion problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.869 + ], + "angle": 0, + "content": "Transforming the inversion problem. It is intractable to directly calculate the sensitivity condition. For diffusion models, it can be achieved by minimizing the denoising error \\( l_{de} \\) (see Eq. (7)) as a proxy. As can be expected, when \\( l_{de} \\) gradually decreases, \\( S(q_{\\varphi}) \\) will approach 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Moreover, denote the average KL divergence in Eq. (9) by \\(l_{kl}\\), the optimization problem can be solved by La" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16962" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.212, + 0.108 + ], + "angle": 0, + "content": "grangian method:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.12, + 0.483, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\varphi , \\lambda} l _ {d e} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right) + \\lambda l _ {k l} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right), \\lambda \\geq 0 \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.185 + ], + "angle": 0, + "content": "where \\(\\lambda \\geq 0\\) is a multiplier. We attach the multiplier to \\(l_{kl}\\) to ensure adequate optimization of the denoising error \\(l_{de}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.484, + 0.232 + ], + "angle": 0, + "content": "Different from the standard \\(\\epsilon_0\\)-prediction denoising error, we adopt \\(x_0\\)-prediction denoising error, which shows more stable inversion efficacy in practice:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.241, + 0.483, + 0.293 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} l _ {d e} = \\mathbb {E} _ {q _ {\\boldsymbol {\\varphi}} (\\epsilon_ {0} | \\boldsymbol {x} _ {0}), t} \\left[ \\| \\boldsymbol {x} _ {0} - \\boldsymbol {x} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\\\ = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\mathbb {E} _ {q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), t} \\left[ \\| \\epsilon_ {0} - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.483, + 0.334 + ], + "angle": 0, + "content": "where \\( \\pmb{x}_t \\) is obtained as in Eq. (6) with added noise drawn from \\( q_{\\varphi} \\) rather than standard Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.44 + ], + "angle": 0, + "content": "For the noise distribution \\(q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)\\), we assume it to be a multivariate Gaussian \\(\\mathcal{N}(\\boldsymbol{\\mu},\\boldsymbol{\\sigma}^2)\\) with learnable mean \\(\\pmb{\\mu}\\) and diagonal variance \\(\\sigma^2\\), i.e., \\(\\varphi = (\\pmb{\\mu},\\pmb{\\sigma}^2)\\). Then InvMM has a closed-form expression \\(\\frac{1}{2N}\\sum_i(\\pmb{\\mu}_i^2 +\\pmb{\\sigma}_i^2 -\\log \\pmb{\\sigma}_i^2 -1)\\). Although the true distribution of sensitive latent noise might not be fully captured by a Gaussian, we choose this modeling for simplicity and find it effective in experiments." + }, + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.455, + 0.463, + 0.47 + ], + "angle": 0, + "content": "Algorithm 1 InvMM: adaptive weight tuning algorithm." + }, + { + "type": "algorithm", + "bbox": [ + 0.095, + 0.474, + 0.487, + 0.824 + ], + "angle": 0, + "content": "Input: Noise prediction network \\(\\epsilon_{\\theta}\\), target image \\(x_0\\), optimization iteration \\(T\\), observation cycle \\(C\\), weight increment \\(\\delta\\), threshold \\(\\xi\\), learning rate \\(\\gamma\\) \n1: Initialize \\(\\varphi \\gets (\\mu \\gets 0, \\sigma \\gets 1)\\), \\(\\lambda \\gets 1, l_{de}^p \\gets +\\infty\\), success \\(\\gets\\) FALSE \n2: for \\(i \\gets 1\\) to \\(T\\) do \n3: \\(\\mathcal{L} \\gets l_{de}(\\boldsymbol{x}_0; \\varphi) + \\lambda l_{kl}(\\boldsymbol{x}_0; \\varphi)\\) \n4: \\(\\varphi \\gets \\varphi - \\gamma \\nabla_{\\varphi} \\mathcal{L}\\) \n5: if \\(i > 1\\) and \\(i \\% C = 0\\) then \n6: if \\(l_{de} - l_{de}^p < \\xi\\) then \n7: \\(\\lambda \\gets \\lambda / 2\\) \n8: else \n9: \\(\\lambda \\gets \\lambda + \\delta\\) \n10: \\(l_{de}^p \\gets l_{de}\\) \n11: if \\(S(q_\\varphi) = 1\\) then \n12: success \\(\\leftarrow\\) TRUE \n13: break \n14: else \n15: \\(\\lambda \\gets \\lambda + \\delta\\) \n16: if success then \n17: return \\(\\frac{1}{2N} \\sum_i (\\mu_i^2 + \\sigma_i^2 - \\log \\sigma_i^2 - 1)\\) \n18: else \n19: return \\(+\\infty\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Adaptive weight tuning. To achieve the best trade-off, the weight \\(\\lambda\\) should be adjusted appropriately for different samples. However, it cannot be optimized via gradient descent together with \\(\\varphi\\). As the gradient w.r.t. \\(\\lambda\\) is \\(l_{kl}\\), which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "is always nonnegative, \\(\\lambda\\) will gradually decrease to 0 and \\(l_{kl}\\) will be totally ignored. To solve the problem, we propose an algorithm that allows adaptive tuning of \\(\\lambda\\) during inversion, motivated by the additive increase/multiplicative decrease scheme in congestion control of TCP protocol, as in Algorithm 1, line 6-9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.471 + ], + "angle": 0, + "content": "Specifically, we monitor the denoising error \\( l_{de} \\) every \\( C \\) steps during the optimization process and divide the weight \\( \\lambda \\) by half if no significant improvement of denoising error is observed in the cycle. Otherwise, \\( \\lambda \\) slowly increases by a small increment \\( \\delta \\) each step. In this way, \\( \\lambda \\) is adaptively adjusted for different samples to balance the normality and denoising error. When the normality is too strong to further reduce the denoising error, thus to meet the sensitivity condition, it is immediately down-weighted. Otherwise, we gradually strengthen it for better normality. We also perform early stop (line 11-13) when the sensitivity condition is reached to avoid unnecessary normality degradation. The sensitivity condition is tested via the Monte Carlo method. Random latent vectors are drawn to generate samples. If the distances between all of them and the target image are lower than the threshold \\( \\beta \\), then the inversion is successful. There are cases that only partial samples replicate that target image in the end. We still consider such cases as insignificant memorization and flag them as unsuccessful." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.482, + 0.692, + 0.499 + ], + "angle": 0, + "content": "4.4. Text-to-image DM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.644 + ], + "angle": 0, + "content": "For text-to-image DM, we model the prompt part as a distribution as well and optimize it simultaneously with the noise distribution. Suppose an input prompt is tokenized into maximum \\(M\\) tokens \\(\\omega = [\\omega_{1},\\omega_{2},\\dots,\\omega_{M}]\\). Motivated by Guo et al. [20], we learn a \\(|\\mathcal{V}|\\)-dimensional categorical distribution over the vocabulary \\(\\mathcal{V}\\) for each token \\(\\omega_{i}\\). Let token \\(\\omega_{i}\\) follows a categorical distribution with probabilities \\(\\pi_{i,j}\\), where \\(\\sum_{j=1}^{|\\mathcal{V}|}\\pi_{i,j} = 1\\). \\(\\pi_{i,j}\\) can be parameterized and optimized." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.906, + 0.783 + ], + "angle": 0, + "content": "Because sampling from categorical distribution is non-differentiable, we adopt the Gumbel-Softmax reparameterization [23] to obtain smoothed sample \\(\\tilde{\\omega}_i\\). Given the smoothed token \\(\\tilde{\\omega}_i\\), it can be converted to token embeddings as a linear combination of vocabulary embeddings \\(e(\\tilde{\\omega}_i) = \\sum_{j=1}^{|\\mathcal{V}|} \\tilde{\\omega}_{i,j} e(j)\\), where \\(e(j)\\) is the pretrained token embedding of the \\(j\\)-th token. A proper choice of the prompt distribution is critical for inversion on text-to-image DMs. We provide a detailed analysis in Appendix Sec. 9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.799, + 0.64, + 0.816 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct a series of experiments from different perspectives to demonstrate the validity of InvMM. First, we conduct controlled studies to verify the effectiveness of the adaptive algorithm (Sec. 5.2). Based on this, we show that InvMM can reasonably elucidate the influence" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16963" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.093, + 0.458, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.214, + 0.48, + 0.228 + ], + "angle": 0, + "content": "Figure 3. The influence of weight \\(\\lambda\\) on measuring memorization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.292 + ], + "angle": 0, + "content": "of various factors on memorization (Sec. 5.3) and identify highly memorized images (Sec. 5.4). Finally, we discuss how memorization differs from membership (Sec. 5.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.265, + 0.318 + ], + "angle": 0, + "content": "5.1. Experiment setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.594 + ], + "angle": 0, + "content": "Dataset. We use four datasets for evaluation, i.e., CIFAR-10 [31], CelebAHQ [26], FFHQ [27] and LAION-Aesthetics V2 [47]. For subsequent experiments, we construct four randomly sampled subsets of CelebAHQ and FFHQ: CelebAHQ-250, CelebAHQ-2.5k, FFHQ-600 and FFHQ-6k. We ensure that the larger subset includes the smaller subset. The LAION dataset is extremely large. To keep computational costs manageable, we collect a subset of images. We first include 76 heavily memorized images discovered by Webster [58]. We call this set of images the confirmed set. In addition, we collected another 50 images with high duplication rates according to the statistics from Webster et al [60]. This set of suspects, named as the suspicious set, are assured to be disjoint with the confirmed set. Images in the confirmed set can be triggered by their training captions while those in the suspicious set cannot. Finally, another 50 randomly sampled images are used for comparison, referred to as the normal set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Diffusion Models. Following the implementations of DDPM [22] and Latent Diffusion Model (LDM) [43], we train DDPM on CIFAR-10 and LDM on the subsets of CelebAHQ and FFHQ. Pre-trained LDMs on the whole CelebAHQ and FFHQ are also used. For text-to-image generation, we utilize Stable Diffusion (SD) v1.4, v2.1 and v3.5 trained on LAION." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.701, + 0.483, + 0.733 + ], + "angle": 0, + "content": "Inversion. The experiment setting and additional results are deferred to Appendix Sec. 8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.279, + 0.759 + ], + "angle": 0, + "content": "5.2. Adaptive algorithm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The core design of Algorithm 1 is the dynamic reweighting of the two loss items. To verify its effectiveness, we compare to fixed weights \\(\\lambda = 1\\) and \\(\\lambda = 0\\). \\(\\lambda = 1\\) equals DM's original training objective (Eqs. (3) and (4)) and \\(\\lambda = 0\\) totally ignores the normality constraint." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "A comparison of the resulting memorization score under different settings can be observed in Fig. 3. Fixing \\(\\lambda\\) to 1 cannot successfully invert most images: \\(92.4\\%\\) and \\(99.7\\%\\) images of CelebAHQ and FFHQ result in infinite" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.088, + 0.71, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.213, + 0.693, + 0.224 + ], + "angle": 0, + "content": "(a) CIFAR-10 (DDPM)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.868, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.213, + 0.835, + 0.224 + ], + "angle": 0, + "content": "(b) LAION (SD)" + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.225, + 0.71, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.349, + 0.692, + 0.361 + ], + "angle": 0, + "content": "(c) CelebAHQ (LDM)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.225, + 0.869, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.349, + 0.838, + 0.361 + ], + "angle": 0, + "content": "(d) FFHQ (LDM)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.374, + 0.906, + 0.415 + ], + "angle": 0, + "content": "Figure 4. Random samples generated using latent noise vectors drawn from inverted distributions. The original training images are in the first column within each block." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.492 + ], + "angle": 0, + "content": "memorization scores. In contrast, both \\(\\lambda = 0\\) and adaptive reweighting successfully invert all images. However, Fixing \\(\\lambda\\) to 0 leads to larger resulting memorization scores, which indicates underestimation of the extent of memorization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Figure 4 showcases the generation results with the learned distributions. With the adaptive algorithm, we successfully perform inversion across various datasets and DMs that the randomly generated samples closely resemble the training images. The results form the premise of InvMM that memorization of samples is measured in the state they are made replicable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.68, + 0.622 + ], + "angle": 0, + "content": "5.3. Influence factors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Both theoretical analysis and empirical investigation point out that dataset complexity [15, 52], dataset size [15, 52], cardinality of latent vector set [54] and data duplication [60] are important factors for memorization. We thus verify the correctness of our measure by utilizing it to describe how memorization trends vary with changes in these factors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dataset complexity, size and training epoch. We begin by simultaneously validating dataset complexity, size and cardinality of latent vector set on CelebAHQ and FFHQ. Following Feng et al. [15], we quantify dataset complexity with Intrinsic Dimensionality (ID) [32]. The cardinality of latent vector set can be controlled by adjusting training epochs (if in each iteration a fresh random latent vector is picked). However, we argue that previous works [15, 51] obfuscate the effect of dataset size and training epochs, as models are trained for different epochs (probably more) on smaller subsets to the full dataset. Consequently, on smaller subsets, each training image would see different number of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16964" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.465, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.206, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Figure 5. Dataset complexity & size vs. memorization. Models are trained for 943 epochs on CelebAHQ and 446 epochs on FFHQ." + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.263, + 0.465, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Figure 6. Training epoch vs. memorization. Models are trained for multiples of the default epochs." + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.423, + 0.465, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.541, + 0.408, + 0.555 + ], + "angle": 0, + "content": "Figure 7. Duplication vs. memorization." + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.559, + 0.465, + 0.667 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.676, + 0.414, + 0.691 + ], + "angle": 0, + "content": "Figure 8. Dedduplication vs. memorization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.764 + ], + "angle": 0, + "content": "latent vectors. Therefore, in our experiment, we decouple these two factors by training for the same epochs as on the full dataset when studying the influence of dataset size." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.87 + ], + "angle": 0, + "content": "As can be observed in Fig. 5, dataset complexity and size seem not present high correlation with memorization, as indicated by our measure. When trained for longer time (Fig. 6), each model presents a larger extent of memorization. The results suggest that dataset complexity and size are not independent determining factors, and InvMM accurately reflects the impact of these three factors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Data duplication. We then validate the effect of data duplication in DMs' training set through our measure. We" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.089, + 0.882, + 0.205 + ], + "angle": 0, + "content": "
MetricDDPM (CIFAR-10)SD v1.4 (LAION)
AUCTPR@1%FPRAUCTPR@1%FPR
ε0-loss0.9810.4140.9550.000
x0-loss0.1800.0000.9750.342
MLOO [55]0.4180.000--
Wen [61]--0.9990.868
Ren [42]--0.9750.355
InvMM0.9970.8990.9740.908
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.215, + 0.878, + 0.228 + ], + "angle": 0, + "content": "Table 1. Detection performance on DDPM and SD v1.4." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.245, + 0.707, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.377, + 0.631, + 0.389 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.246, + 0.884, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.308, + 0.775, + 0.318 + ], + "angle": 0, + "content": "Prompt:" + }, + { + "type": "text", + "bbox": [ + 0.733, + 0.319, + 0.875, + 0.369 + ], + "angle": 0, + "content": "<|endoftext|>eaglegotcha \nforever walaccidents concludes \ntrailer disodana crying ner banana!# conan \"\"gazing \n<|endoftext|>shark" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.377, + 0.814, + 0.388 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.402, + 0.905, + 0.43 + ], + "angle": 0, + "content": "Figure 9. (a) Memorization score distribution on three datasets. (b) Adversarial prompt for an example from the suspicious set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.905, + 0.566 + ], + "angle": 0, + "content": "train a series of LDMs on CelebAHQ-2.5k and FFHQ-6k, with their subsets CelebAHQ-250 and FFHQ-600 duplicated 5, 10, 15 and 20 times. Figure 7 depicts the memorization scores of duplicated images. As can be observed, data duplication rapidly worsens memorization in DMs and our measure is sensitive to these changes. With increasing duplication times, the larger FFHQ-6k dataset presents a stronger resistance to memorization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.749 + ], + "angle": 0, + "content": "For CIFAR-10 and LAION which originally contain duplicates, we observe the effect of dedduplication. We deduplicate CIFAR-10 with SSCD embeddings and DBSCAN clustering [14]. This results in 666 identified clusters, only one image of which is retained for training. The memorization is evaluated on the retained 666 images before and after dedduplication. On LAION, as compared to SD v1.4, SD v2.1 and v3.5 deduplicate the training set. Memorization is evaluated on the confirmed and suspicious sets. Figure 8 shows the results, with infinite scores on the top. As can be observed, InvMM clearly exhibits the benefits that dedduplication brings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.625, + 0.772 + ], + "angle": 0, + "content": "5.4. Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "For InvMM to be a correct measure of memorization, it should be capable of distinguishing highly memorized images from normal ones. To construct a reliable test set, we conduct a nearest neighbor test following previous studies [9, 51, 59] to discover highly memorized images from DM's training set, i.e., highly memorized images consist of training images that can appear at sampling time. Specifically, we consider DDPM trained on CIFAR-10, LDMs" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16965" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.094, + 0.287, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.094, + 0.455, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.214, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Figure 10. Performance comparison of membership inference on FFHQ-6k, across different training epochs and duplication times." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.262, + 0.483, + 0.444 + ], + "angle": 0, + "content": "trained on CelebAHQ and FFHQ, and SDs trained on LAION. With their official implementations, only DDPM and SD v1.4 present obvious replications, which are thus used for evaluation. For DDPM, the 99 discovered training images appearing in 1M random samples are used as positive samples and another 1k different training images are included as negative samples; for SD v1.4, in accordance with the setting of previous works, we deem images in the confirmed set as positive and those in the suspicious and normal set as negative, considering that captions in the confirmed set can cause replication of their paired images while those in the suspicious and normal set cannot." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Table 1 presents the detection performance assessed via AUC and TPR@1%FPR. For the unconditional DDPM on CIFAR-10, training loss metrics (the first 2 rows) show decent results and the original \\(\\epsilon_0\\)-loss works much better than the \\(x_0\\)-loss. The differential metric \\(M^{\\mathrm{LOO}}\\) is less effective than the \\(\\epsilon_0\\)-loss and InvMM. The results indicate that the compared metrics are not able to quantify memorization, as they are not fundamentally commensurable between samples. InvMM measures memorization of each sample by contrasting its sensitive noise distribution to a common standard Gaussian, thus having the same scale." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For SD v1.4, we compare InvMM against training loss metrics and two state-of-the-art detection metrics [42, 61]. Since all compared approaches require access to image captions, we implement InvMM with fixed prompts during inversion. As shown in Tab. 1, our measure achieves performance comparable to state-of-the-art metrics. The results of compared metrics seem to demonstrate a strong ability to distinguish memorization thus far. However, we argue that the prompt-based task formulation employed in existing studies offers an overly optimistic understanding of memorization. It is because these techniques are actually to predict whether the captions will replicate their paired images, while overlook potential prompts that might cause replication. As observed in Fig. 9a, when simultaneously inverting the input prompt, our measure reveals that some images in the suspicious set are also memorized to a similar extent as those in the confirmed set (due to duplication). Applying GCG [64] attack on images with the lowest InvMM scores in the suspicious set provides additional evidence, as illus" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "trated in Fig. 9b: images previously deemed “not memorized” according to their training captions can indeed be replicated by potential adversarial prompts (see Appendix Fig. 13 for more examples). Hence, going beyond the confines of existing research, our measure describes worst-case memorization from an adversarial standpoint, making it more reliable in security-critical scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.207, + 0.766, + 0.223 + ], + "angle": 0, + "content": "5.5. Comparison to membership" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.363 + ], + "angle": 0, + "content": "A close topic to memorization is membership inference [8, 13, 49], which reveals the leakage risk of membership information that whether an image is used for training. The common implicit assumption is that member samples are \"memorized\" more than hold-out samples [13], rendering them distinguishable through statistics like (variants of) training loss or differential loss [8]. Interestingly, the principle of InvMM can provide an explicit clarification of the difference between membership and memorization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.697 + ], + "angle": 0, + "content": "A comparison to a recent membership inference metric SecMI [13] on FFHQ-6k, as shown in Fig. 10, presents a positive correlation between them. As factors that exacerbate memorization changes, both SecMI and InvMM show increasing performance. There is also a performance gap between SecMI and InvMM, indicating that certain samples from the hold-out set are memorized more than some member samples. This mainly results from the gap between model's metric (training loss) and human's metric (perceptual similarity). Membership inference leverages the model's metric for prediction, while memorization should resort to perceptual metric, e.g., simulated by SSCD. What level of loss leads to image replication depends both on the model and image itself. Low training loss does not necessarily lead to data replication, but would probably distinguish training data from never accessed. On the contrary, high training loss signifying non-membership may still be enough for replication. Therefore, membership does not deterministically lead to high memorization and vice versa. InvMM differs from membership metrics by focusing on the perceptual level and adaptively handles the specific loss landscapes of different samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.634, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we dive into measuring memorization for diffusion models. Comprehensive experiments demonstrate the ability of the proposed measure to correctly quantify memorization. In short, an image to a higher extent of memorization is able to be inverted to a more standard and diverse latent distribution. However, our method is also limited by its simple modeling of the sensitive latent noise as a multivariate normal. During our research, we observed many clues confirming that the true distribution is complex. Future work might explore more advanced methods, e.g., a generative network, to fit the sensitive noise." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16966" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.298 + ], + "angle": 0, + "content": "This work was partly supported by the National Key Research and Development Program of China under No. 2022YFB3102100, NSFC under No. U244120033, U24A20336, 62172243, 62402425 and 62402418, the China Postdoctoral Science Foundation under No. 2024M762829, the Zhejiang Provincial Natural Science Foundation under No. LD24F020002, the “Pioneer and Leading Goose” R&D Program of Zhejiang under No. 2025C01082, 2025C02033, 2025C02263 and 2025C02034, the Zhejiang Provincial Priority-Funded Postdoctoral Research Project under No. ZJ2024001 and the Key R&D Program of Ningbo under No. 2024Z115." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.311, + 0.188, + 0.326 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.336, + 0.45, + 0.35 + ], + "angle": 0, + "content": "[1] Midjourney. https://www.midjourney.com. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.483, + 0.42 + ], + "angle": 0, + "content": "[2] Artist finds private medical record photos in popular AI training data set. https://arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-photos-in-popular-ai-training-data-set/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.422, + 0.455, + 0.435 + ], + "angle": 0, + "content": "[3] DALL·E 3. https://openai.com/dall-e-3.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.491 + ], + "angle": 0, + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[5] Sanjeev Arora and Yi Zhang. Do gans actually learn the distribution? an empirical study. arXiv preprint arXiv:1706.08224, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[8] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[9] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.336, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10759-10769, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.908, + 0.289 + ], + "angle": 0, + "content": "[12] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[13] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? In International Conference on Machine Learning, pages 8717-8730. PMLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.348, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[14] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.404, + 0.906, + 0.472 + ], + "angle": 0, + "content": "[15] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701-6710, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.475, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.546, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[18] Andres Guadamuz. Photographer sues LAION for copyright infringement. https://www.technollama.co.uk/ photographer-sues-laion-for-copyright-infringement, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[19] Ishaan Gulrajani, Colin Raffel, and Luke Metz. Towards gan benchmarks which require generalization. In International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.906, + 0.743 + ], + "angle": 0, + "content": "[20] Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. Gradient-based adversarial attacks against text transformers. arXiv preprint arXiv:2104.13733, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Yue Jiang, Haokun Lin, Yang Bai, Bo Peng, Zhili Liu, Yueming Lyu, Yong Yang, Xingzheng, and Jing Dong. Image-level memorization detection via inversion-based inference" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16967" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "perturbation. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[25] James Jordon, Jinsung Yoon, and Mihaela Van Der Schaar. Pate-gan: Generating synthetic data with differential privacy guarantees. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.483, + 0.235 + ], + "angle": 0, + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.351 + ], + "angle": 0, + "content": "[28] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[29] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.468, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[32] Elizaveta Levina and Peter Bickel. Maximum likelihood estimation of intrinsic dimension. Advances in neural information processing systems, 17, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[34] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[35] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[37] Helen Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[38] Kai Packhäuser, Lukas Folle, Florian Thamm, and Andreas Maier. Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classification systems. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[42] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, pages 340-356. Springer, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.904, + 0.579 + ], + "angle": 0, + "content": "[46] Pamela Samuelson. Generative ai meets copyright. Science, 381(6654):158-161, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[48] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[49] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery?" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16968" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[52] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[54] N Vaishnavh, C Raffel, and IJ Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[55] Gerrit van den Burg and Chris Williams. On memorization in probabilistic deep generative models. Advances in Neural Information Processing Systems, 34:27916-27928, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[56] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[57] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[58] Ryan Webster. A reproducible extraction of training images from diffusion models. CoRR, abs/2305.08694, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[59] Ryan Webster, Julien Rabin, Loic Simon, and Frédéric Jurie. Detecting overfitting of deep generative networks via latent recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11273-11282, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[60] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of LAION-2B. arXiv preprint arXiv:2303.12733, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[61] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[62] Xiaoyu Wu, Yang Hua, Chumeng Liang, Jiaru Zhang, Hao Wang, Tao Song, and Haibing Guan. Cgi-dm: Digital copyright authentication for diffusion models via contrasting gradient inversion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10812-10821. IEEE Computer Society, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[63] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[64] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 8, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16969" + } + ] +] \ No newline at end of file diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_origin.pdf b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2592577f294af2753ff235a63f53dd051c61e0ef --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/4b89bc09-28c7-40c0-9f29-6da101712bc2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a801e5f145684daad35b0365d5193aa00517c4c012f785aa5c059a2284298f51 +size 9828583 diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/full.md b/2025/An Inversion-based Measure of Memorization for Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0a53e1169b20716768e075e1a8feecc9e503ec17 --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/full.md @@ -0,0 +1,380 @@ +# An Inversion-based Measure of Memorization for Diffusion Models + +Zhe Ma, Qingming Li, Xuhong Zhang, Tianyu Du, Ruixiao Lin, Zonghui Wang* Shouling Ji, Wenzhi Chen* Zhejiang University, Hangzhou, China + +{mz.rs, liqm, zhangxuhong, zjradty, linruixiao, zhwang, sji, chenwz}@zju.edu.cn + +# Abstract + +The past few years have witnessed substantial advances in image generation powered by diffusion models. However, it was shown that diffusion models are susceptible to training data memorization, raising significant concerns regarding copyright infringement and privacy invasion. This study delves into a rigorous analysis of memorization in diffusion models. We introduce InvMM1, an inversion-based measure of memorization, which is based on inverting a sensitive latent noise distribution accounting for the replication of an image. For accurate estimation of the measure, we propose an adaptive algorithm that balances the normality and sensitivity of the noise distribution. Comprehensive experiments across four datasets, conducted on both unconditional and text-guided diffusion models, demonstrate that InvMM provides a reliable and complete quantification of memorization. Notably, InvMM is commensurable between samples, reveals the true extent of memorization from an adversarial standpoint and implies how memorization differs from membership. In practice, it serves as an auditing tool for developers to reliably assess the risk of memorization, thereby contributing to the enhancement of trustworthiness and privacy-preserving capabilities of diffusion models. + +# 1. Introduction + +Diffusion Models (DMs) [22, 50] have shown impressive capabilities in the generation of images [41, 43], videos [6], 3D point cloud [35], etc. These techniques lay the foundation for commercial systems or communities such as Stable Diffusion [43], Midjourney [1], DALL-E [3, 41] and Imagegen [44], which have attracted millions of active users. The popularity of diffusion models stems from their hierarchical denoising process, which offers high stability for training on billions of data [47] and scalability for multimodal conditional generation [41, 44]. + +![](images/aade137ef013e9e5b15ef790874e3223b1f161b84c406e1f4c0e41bb4bfffbdb.jpg) +Figure 1. Schematic illustration of diffusion model inversion, optionally with conditions. + +However, large-scale datasets used to train the prevalent DMs, e.g., the open-source image-caption dataset LAION [47], are widely acknowledged to contain content that will raise concerns regarding copyright infringement and privacy invasion. For instance, recent reports revealed that LAION could refer to photographers' works without authorization [18] and include private medical photographs [2]. With the uncurated data for training, diffusion models are likely to generate content that infringes the creators' rights or exposes private information [9, 48]. In this work, we focus on memorization in DMs, one of the most critical issues related to training data misuse. + +Memorization usually manifests itself at sampling time, when a model generates images closely resembling training images. It can turn into a problem both technically and practically. Technically, heavy memorization indicates failure of generalization to some degree because a probabilistic generative model is supposed to produce novel and diversified images rather than frequently replicate those in the training set. Practically, memorization might also raise ethical and legal concerns. First, even the model developers are authorized to train their model with protected images, the image owners will never expect their works to be exposed to arbitrary users due to memorization as this would cause indisciplinable dissemination, violating the contextual integrity [37]. Second, in past years, text-to-image models + +have been facing lawsuits for generating derivative images that mimic the style of artists. Compared to derivative generations whose legality is still in pending [46], exact replication of copyrighted images is undisputedly intolerable. Finally, research on privacy protection [25, 38] has proposed to use synthetic data in place of real data to prevent sharing of private information. For this goal, potential memorization must be carefully circumvented. + +Recent studies [9, 51, 52, 58] have confirmed the existence of some heavily memorized training data in (text-guided) DMs. They either draw random samples from the model and then detect risky generations of low diversity [9], search for nearest neighbors in the training set [51, 52], or directly identify abnormal prompts [58]. However, conducting a practical analysis of memorization with these approaches still faces challenges. Transforming the memorization of an image to the presence or absence in samples from the model is not a reliable solution. While certain images do not appear in a set of samples, it does not demonstrate that they are not highly memorized [55]. For DMs with large latent space, adequate samples should be tested to ensure as complete coverage of the latent space as possible, which could be up to millions of tests [9, 58]. Furthermore, things will get much more complicated in conditional DMs, e.g., text-to-image generation. The additional conditions expand the input space of DMs, making the memorization of an image also depend on the specific condition. Existing studies exploit training captions to discover highly memorized training images or perform memorization detection [42, 61]. However, attributing the memorization risk of an image to its coupled training caption is incomplete. It still remains unknown that whether the other images not triggered by their coupled captions are also to a large extent of memorization and to be triggered by potential prompts. + +Our contributions are threefold. (1) To overcome the challenges, we propose an Inversion-based Measure of Memorization (InvMM) for DMs, a continuous instance-level measure that takes an image as input, outputs a scalar quantifying its level of memorization. As shown in Fig. 1, the inversion process locates a sensitive latent noise distribution and optionally another condition distribution for the target image, such that generating with noise and condition drawn from the inverted distributions will produce samples exactly the same as the target image. InvMM is then defined as the minimum KL divergence of the sensitive noise distribution to the standard Gaussian distribution. Compared to existing methods, InvMM is more reliable and complete in that it proactively searches for all potential sensitive inputs and does not rely on limited test cases from the training set (e.g., captions). (2) To calculate the measure, we design an adaptive optimization algorithm motivated by the TCP congestion control protocol, which is demonstrated to be effective to alleviate overestimate and underestimate + +of memorization. (3) Through comprehensive experiments across four datasets, encompassing both unconditional and text-guided DMs, we present the superiority of InvMM in analyzing memorization. Compared to alternative memorization measures, detection metrics and membership inference metrics, InvMM is commensurable between samples, reflects the true memorization level from an adversarial perspective and clarifies the fundamental difference between memorization and membership. As an auditing tool, it has great potential in discovering highly memorized images, validating safety against memorization for any image, determining appropriate training settings, etc. + +# 2. Related work + +# 2.1. Memorization of generative models + +Memorization has raised wide attention on various generative models, including GAN [17], VAE [30], language models [7] and DM [50]. There have been studies on training algorithms [4, 22, 45] and evaluation protocols [19] to improve the generalization ability of generative models that would potentially circumvent memorization. Empirical and theoretical studies also find that dataset complexity and size [15, 51], training time [54], training data duplication [10, 60] and model capacity [5, 10] are influential factors of memorization. Several works [54, 59] focus on the latent space (noise space) to analyze memorization. Nagarajan et al. [54] demonstrate that when trained on sufficient random noise vectors, a generator is guaranteed to memorize training data. Webster et al. [59] measure overfitting of GANs via the reconstruction error at the optimal noise vector. We take a similar perspective on the latent space to analyze memorization while aim to reach an instance-level measure of memorization rather than reveal overall properties of models. In addition, Van der Burg et al. [55] also propose an instance-level measure of memorization as the changed likelihood of one sample when removing it from the training set. However, the measure is not necessarily comparable between datasets, which is important to understand the true level of memorization. + +Specifically for diffusion models, recent research [9, 51, 52] demonstrates that text-to-image diffusion models suffer from memorization. Some heavily memorized images are discovered from numerous generated samples by membership inference [9] or searching for the most similar training images with image matching models [51]. Webster [58] further provides more efficient attacks to extract training images by identifying abnormal prompts. Subsequent works [42, 61] focus on prompt detection that will trigger training image replication and memorization mitigation [12] at inference time. In response to the disclosed threats, Vyas et al. [56] propose a provable copyright protection method to prevent replication of sensitive training + +images. Our method improves upon existing methods by being commensurable between different samples, basing on images themselves rather than prompts, and working for any samples (not necessarily training samples). A recent work by Jiang et al. [24] also considers prompt-free memorization. However, their method extends Wen et al.'s findings [61] and is restricted to text-guided diffusion models. Our method is established on the property of latent space thus applies to both unconditional and conditional diffusion models. + +# 2.2. Inversion of diffusion models + +Inversion techniques in diffusion models are widely studied for image editing [16, 36, 63], copyright protection [62], generated content detection [11, 57], etc. Through inversion, the object, style and concept contained in the source images can be compressed in latent noises or input token embeddings (for text-to-image models). Then the inverted latent noises or input token embeddings are utilized to generate similar images that preserve the desired content. + +We leverage inversion techniques to analyze memorization in diffusion models. In contrast, we invert a sample to a distribution consisting of all feasible latent vectors rather than to a single vector [11, 16, 36, 57, 62]. We also adopt more efficient training objective for inversion, without the need to iteratively invert a specific sampling procedure [11, 36, 57, 62]. + +# 3. Background + +Basically, DMs [22, 50] are likelihood-based generative models that learn to maximize the likelihood $p_{\theta}(\pmb{x}_0)$ of observed data $\pmb{x}_0$ . Directly calculating $p_{\theta}(\pmb{x}_0)$ is intractable for continuous data such as image. By contrasting to a sequence of latent variables $x_{1:T}$ , the log-likelihood $\log p_{\theta}(\pmb{x}_0)$ can be maximized indirectly by maximizing its variational lower bound [29, 34]: + +$$ +\begin{array}{l} \log p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {0}\right) = \log \int p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {0: T}\right) d \boldsymbol {x} _ {1: T} (1) \\ \geq \mathbb {E} _ {q \left(\boldsymbol {x} _ {1: T} \mid \boldsymbol {x} _ {0}\right)} \left[ \log \frac {p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {0 : T}\right)}{q \left(\boldsymbol {x} _ {1 : T} \mid \boldsymbol {x} _ {0}\right)} \right] (2) \\ \end{array} +$$ + +Assume both transitions $p_{\pmb{\theta}}(\pmb{x}_{0:T})$ and $q(\pmb{x}_{1:T}|\pmb{x}_0)$ are Markov chains, i.e., $q(\pmb{x}_{1:T}|\pmb{x}_0) = \prod_{t=1}^{T} q(\pmb{x}_t|\pmb{x}_{t-1})$ , $p_{\pmb{\theta}}(\pmb{x}_{0:T}) = p(\pmb{x}_T)\prod_{t=1}^{T} p_{\pmb{\theta}}(\pmb{x}_{t-1}|\pmb{x}_t)$ , the variational lower bound in Eq. (2) expands as: + +$$ +\begin{array}{l} \log p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {0}\right) \geq \mathbb {E} _ {q \left(\boldsymbol {x} _ {1} \mid \boldsymbol {x} _ {0}\right)} \left[ \log p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {0} \mid \boldsymbol {x} _ {1}\right) \right] \\ - \sum_ {t = 2} ^ {T} \mathbb {E} _ {q \left(\boldsymbol {x} _ {t} \mid \boldsymbol {x} _ {0}\right)} [ D _ {\mathrm {K L}} (q \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}\right) \| p _ {\boldsymbol {\theta}} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}\right)) ] (3) \\ - D _ {\mathrm {K L}} \left(q \left(\boldsymbol {x} _ {T} \mid \boldsymbol {x} _ {0}\right) \| p \left(\boldsymbol {x} _ {T}\right)\right) (4) \\ \end{array} +$$ + +where $p(\boldsymbol{x}_T)$ is generally assumed to be a simple distribution such as standard Gaussian $\mathcal{N}(\mathbf{0},\mathbf{I})$ + +The transition $q(\pmb{x}_{1:T}|\pmb{x}_0)$ is not learned but replaced by a manually crafted non-parametric diffusion process. The transition $p_{\theta}(\pmb{x}_{0:T})$ is the corresponding reverse (or generation, sampling) process. The diffusion process gradually adds Gaussian noises to the input $\pmb{x}_0$ according to a weight schedule $\alpha_{1:T}$ : + +$$ +\boldsymbol {x} _ {t} = \sqrt {\alpha_ {t}} \boldsymbol {x} _ {t - 1} + \sqrt {1 - \alpha_ {t}} \boldsymbol {\epsilon} _ {t - 1}, \boldsymbol {\epsilon} _ {t - 1} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}) \tag {5} +$$ + +$$ +\boldsymbol {x} _ {t} = \sqrt {\bar {\alpha} _ {t}} \boldsymbol {x} _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \boldsymbol {\epsilon} _ {0}, \boldsymbol {\epsilon} _ {0} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}) \tag {6} +$$ + +where $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ decreases over time and finally to almost zero at the last step $T$ , such that $q(\pmb{x}_T | \pmb{x}_0)$ approaches to a standard Gaussian and thus Eq. (4) approximately equals to 0. + +If the diffusion process is divided into sufficient steps, each reverse transition $p_{\theta}(\pmb{x}_{t-1}|\pmb{x}_t)$ can also be approximated by a Gaussian transformation with learnable parameters that is trained to match the corresponding posterior diffusion transition $q(\pmb{x}_{t-1}|\pmb{x}_t,\pmb{x}_0)$ . Omitting further details [34], the following simplified objective [22] as an estimate of the variational lower bound is commonly used for training: + +$$ +l _ {d e} = \mathbb {E} _ {\boldsymbol {\epsilon} _ {0} \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), t \sim \mathcal {U} \{1, T \}} \left[ \| \boldsymbol {\epsilon} _ {0} - \boldsymbol {\epsilon} _ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2} \right] \tag {7} +$$ + +where $\epsilon_{\theta}$ is a neural network that takes the noisy data $\mathbf{x}_t$ and timestep $t$ as input and predicts the added noise $\epsilon_0$ . After training, the vanilla sampling procedure starts with a random Gaussian noise $\mathbf{x}_T \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ and removes the predicted noise stepwise. However, it is extremely slow to generate an image as it must invoke the network $\epsilon_{\theta}$ for $T$ times (typically 1000 steps). To mitigate the problem, a variety of efficient sampling algorithms are proposed, such as DDIM sampler [53] and PLMS sampler [33]. + +# 4. Inversion-based measure + +Given an image $\boldsymbol{x}_0$ , we aim to measure to what extent $\boldsymbol{x}_0$ is memorized by a diffusion model (DM). We propose to invert the image into a latent noise distribution, such that noise vectors drawn from it will lead to replication of the image $\boldsymbol{x}_0$ . Our memorization measure is established in this inversion process. In this section, we introduce a sensitivity definition of the inverted noise distribution (Sec. 4.1), elaborate the memorization measure (Sec. 4.2) and give an adaptive algorithm for its calculation (Sec. 4.3). Ultimately, we present an example to measure memorization in conditional DM, specifically text-to-image DM (Sec. 4.4). + +# 4.1. Sensitivity of latent distribution + +The diversity of DM generated content lies in the randomness of latent noise vectors. Starting from different noise + +vectors, the generated contents generally vary. In the context of memorization, this motivates us to search for sensitive noise vectors that will replicate the target image. We then figure out "how many" sensitive noise vectors there are and adopt it as a meaningful measure. + +Considering the fact that the common latent noise is sampled from a standard Gaussian distribution, we correspondingly assume the sensitive noise vectors follow a distribution $q_{\varphi}(\epsilon_0|\boldsymbol{x}_0)$ with learnable parameters $\varphi$ . The sensitivity function $S(q_{\varphi})$ of a noise distribution can be defined as the probability that generated image $G(\epsilon_0)$ obtained from random noise $\epsilon_0 \sim q_{\varphi}(\epsilon_0|\boldsymbol{x}_0)$ replicates the target image: + +$$ +\mathcal {S} \left(q _ {\varphi}\right) = \int \mathbf {1} _ {d \left(G \left(\boldsymbol {\epsilon} _ {0}\right), \boldsymbol {x} _ {0}\right) \leq \beta} q _ {\varphi} \left(\boldsymbol {\epsilon} _ {0} \mid \boldsymbol {x} _ {0}\right) d \boldsymbol {\epsilon} _ {0} \tag {8} +$$ + +where $\mathbf{1}_v$ is an indicator function that equals 1 if $v$ is true and 0 otherwise, $d(\cdot ,\cdot)$ is a distance metric between two images and $\beta$ is an appropriate threshold. + +The choice of the distance metric depends on specific applications. In applications that accept derivative generation, a strict metric that only detects eidetic generation can be utilized. While under some circumstances even similar generation should be circumvented, a looser metric is needed. In this work, we consider exact replication that the generated image and target image are exactly the same. This consideration to some extent reduces the ambiguity of determining the similarity of two images. Following the investigation of Somepalli et al. [51], we implement the metric as the Euclidean distance between embeddings extracted by the pretrained copy detection model SSCD [39]. + +# 4.2. Quantify memorization + +Based on the definition of sensitivity, we propose an Inversion-based Measure of Memorization (InvMM), as the best achievable average normality of $q_{\varphi}(\epsilon | \boldsymbol{x})$ under the condition that $S(q_{\varphi}) = 1$ : + +$$ +\begin{array}{l} \min _ {\varphi} \frac {1}{N} D _ {\mathrm {K L}} \left(q _ {\varphi} \left(\boldsymbol {\epsilon} _ {0} \mid \boldsymbol {x}\right) \| p \left(\boldsymbol {\epsilon} _ {0}\right)\right) \\ \text {s . t .} \mathcal {S} \left(q _ {\varphi}\right) = 1 \tag {9} \\ \end{array} +$$ + +where $N$ is the dimensionality of latent noise space and $p(\epsilon_0)$ is the prior distribution used in DM, which is often selected to be standard Gaussian. The normality of inverted noise distribution is the KL divergence to the prior. A lower value of InvMM means greater degree of memorization. + +Explaining InvMM. Partitioning the normality gives us a more intuitional understanding of the measure. Notice that + +$$ +\begin{array}{l} D _ {\mathrm {K L}} \left(q _ {\varphi} \left(\epsilon_ {0} | \boldsymbol {x} _ {0}\right) \| p (\epsilon_ {0})\right) = \\ \mathrm {H} \left(q _ {\varphi} \left(\epsilon_ {0} \mid \boldsymbol {x} _ {0}\right), p \left(\epsilon_ {0}\right)\right) - \mathrm {H} \left(q _ {\varphi} \left(\epsilon_ {0} \mid \boldsymbol {x} _ {0}\right)\right) \tag {10} \\ \end{array} +$$ + +![](images/8693c96cea934401c724b0022745b09ec19951ada2eeabab4e48522f41f49a67.jpg) +(a) Optimal + +![](images/3c8cef596b7cc85255f22c75a5f3bd0477629e5b8602e7f4b124a296366615f9.jpg) +(b) Overestimate +(c)Underestimate + +![](images/22f216fac8e19c8856f63f792ea342611b9259c4446177d177231e17ffe96e7d.jpg) +Figure 2. Illustration of our memorization measure. The gray blocks indicate the distribution of sensitive latent noise vectors. + +where $\mathrm{H}(q_{\varphi}(\epsilon_0|\pmb{x}_0), p(\epsilon_0))$ is the cross entropy between the inverted distribution and the prior distribution, $\mathrm{H}(q_{\varphi}(\epsilon_0|\pmb{x}_0))$ is the differential entropy of the inverted distribution. Minimizing the normality thus encourages a distribution as close to the prior distribution as possible, meanwhile with the largest entropy. For the cross entropy, following the Jensen's inequality, $-\mathrm{H}(q_{\varphi}(\epsilon_0|\pmb{x}_0), p(\epsilon_0)) = \mathbb{E}_{q_{\varphi}(\epsilon_0|\pmb{x}_0)}[\log p(\epsilon_0)] \leq \log \int q_{\varphi}(\epsilon_0|\pmb{x}_0)p(\epsilon_0)d\epsilon_0$ . The term $q_{\varphi}(\epsilon_0|\pmb{x}_0)p(\epsilon_0)$ is proportional to the probability of a random noise $\epsilon_0$ , with probability density $q_{\varphi}(\epsilon_0|\pmb{x}_0)$ , to be sampled from the prior distribution with probability density $p(\epsilon_0)$ . In the sense, minimizing the cross-entropy improves the probability that sensitive noise vectors regarding an image $\pmb{x}_0$ are sampled in the normal usage of DM. For the differential entropy, maximizing it will produce a noise distribution involving "as many" sensitive noise vectors as possible. In the extreme case, suppose a uniform distribution over the entire latent space having the largest entropy, it means that any noise in the space can cause replication of the target training sample. + +InvMM requires to achieve the best trade-off between improving normality and meeting the sensitivity condition. Different images to varied extent of memorization may reach different equilibrium states. Figure 2 illustrates the optimal solution compared to suboptimal choices. Overestimate of memorization produces a distribution with good normality but low sensitivity (Fig. 2b). Conservative inversion of the latent distribution perfectly meets the sensitivity condition while underestimates memorization (Fig. 2c). + +# 4.3. Adaptive algorithm + +We then introduce an adaptive algorithm to solve the inversion problem. + +Transforming the inversion problem. It is intractable to directly calculate the sensitivity condition. For diffusion models, it can be achieved by minimizing the denoising error $l_{de}$ (see Eq. (7)) as a proxy. As can be expected, when $l_{de}$ gradually decreases, $S(q_{\varphi})$ will approach 1. + +Moreover, denote the average KL divergence in Eq. (9) by $l_{kl}$ , the optimization problem can be solved by La + +grangian method: + +$$ +\min _ {\varphi , \lambda} l _ {d e} \left(\boldsymbol {x} _ {0}; \boldsymbol {\varphi}\right) + \lambda l _ {k l} \left(\boldsymbol {x} _ {0}; \boldsymbol {\varphi}\right), \lambda \geq 0 \tag {11} +$$ + +where $\lambda \geq 0$ is a multiplier. We attach the multiplier to $l_{kl}$ to ensure adequate optimization of the denoising error $l_{de}$ . + +Different from the standard $\epsilon_0$ -prediction denoising error, we adopt $x_0$ -prediction denoising error, which shows more stable inversion efficacy in practice: + +$$ +\begin{array}{l} l _ {d e} = \mathbb {E} _ {q _ {\boldsymbol {\varphi}} (\epsilon_ {0} | \boldsymbol {x} _ {0}), t} \left[ \| \boldsymbol {x} _ {0} - \boldsymbol {x} _ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2} \right] \\ = \frac {1 - \bar {\alpha} _ {t}}{\bar {\alpha} _ {t}} \mathbb {E} _ {q _ {\varphi} \left(\epsilon_ {0} \mid \boldsymbol {x} _ {0}\right), t} \left[ \| \epsilon_ {0} - \epsilon_ {\boldsymbol {\theta}} (\boldsymbol {x} _ {t}, t) \| _ {2} ^ {2} \right] \tag {12} \\ \end{array} +$$ + +where $\pmb{x}_t$ is obtained as in Eq. (6) with added noise drawn from $q_{\varphi}$ rather than standard Gaussian. + +For the noise distribution $q_{\varphi}(\epsilon_0|\boldsymbol{x}_0)$ , we assume it to be a multivariate Gaussian $\mathcal{N}(\boldsymbol{\mu},\boldsymbol{\sigma}^2)$ with learnable mean $\pmb{\mu}$ and diagonal variance $\sigma^2$ , i.e., $\varphi = (\pmb{\mu},\pmb{\sigma}^2)$ . Then InvMM has a closed-form expression $\frac{1}{2N}\sum_i(\pmb{\mu}_i^2 +\pmb{\sigma}_i^2 -\log \pmb{\sigma}_i^2 -1)$ . Although the true distribution of sensitive latent noise might not be fully captured by a Gaussian, we choose this modeling for simplicity and find it effective in experiments. + +Algorithm 1 InvMM: adaptive weight tuning algorithm. +Input: Noise prediction network $\epsilon_{\theta}$ , target image $x_0$ , optimization iteration $T$ , observation cycle $C$ , weight increment $\delta$ , threshold $\xi$ , learning rate $\gamma$ +1: Initialize $\varphi \gets (\mu \gets 0, \sigma \gets 1)$ , $\lambda \gets 1, l_{de}^p \gets +\infty$ , success $\gets$ FALSE +2: for $i \gets 1$ to $T$ do +3: $\mathcal{L} \gets l_{de}(\boldsymbol{x}_0; \varphi) + \lambda l_{kl}(\boldsymbol{x}_0; \varphi)$ +4: $\varphi \gets \varphi - \gamma \nabla_{\varphi} \mathcal{L}$ +5: if $i > 1$ and $i \% C = 0$ then +6: if $l_{de} - l_{de}^p < \xi$ then +7: $\lambda \gets \lambda / 2$ +8: else +9: $\lambda \gets \lambda + \delta$ +10: $l_{de}^p \gets l_{de}$ +11: if $S(q_\varphi) = 1$ then +12: success $\leftarrow$ TRUE +13: break +14: else +15: $\lambda \gets \lambda + \delta$ +16: if success then +17: return $\frac{1}{2N} \sum_i (\mu_i^2 + \sigma_i^2 - \log \sigma_i^2 - 1)$ +18: else +19: return $+\infty$ + +Adaptive weight tuning. To achieve the best trade-off, the weight $\lambda$ should be adjusted appropriately for different samples. However, it cannot be optimized via gradient descent together with $\varphi$ . As the gradient w.r.t. $\lambda$ is $l_{kl}$ , which + +is always nonnegative, $\lambda$ will gradually decrease to 0 and $l_{kl}$ will be totally ignored. To solve the problem, we propose an algorithm that allows adaptive tuning of $\lambda$ during inversion, motivated by the additive increase/multiplicative decrease scheme in congestion control of TCP protocol, as in Algorithm 1, line 6-9. + +Specifically, we monitor the denoising error $l_{de}$ every $C$ steps during the optimization process and divide the weight $\lambda$ by half if no significant improvement of denoising error is observed in the cycle. Otherwise, $\lambda$ slowly increases by a small increment $\delta$ each step. In this way, $\lambda$ is adaptively adjusted for different samples to balance the normality and denoising error. When the normality is too strong to further reduce the denoising error, thus to meet the sensitivity condition, it is immediately down-weighted. Otherwise, we gradually strengthen it for better normality. We also perform early stop (line 11-13) when the sensitivity condition is reached to avoid unnecessary normality degradation. The sensitivity condition is tested via the Monte Carlo method. Random latent vectors are drawn to generate samples. If the distances between all of them and the target image are lower than the threshold $\beta$ , then the inversion is successful. There are cases that only partial samples replicate that target image in the end. We still consider such cases as insignificant memorization and flag them as unsuccessful. + +# 4.4. Text-to-image DM + +For text-to-image DM, we model the prompt part as a distribution as well and optimize it simultaneously with the noise distribution. Suppose an input prompt is tokenized into maximum $M$ tokens $\omega = [\omega_{1},\omega_{2},\dots,\omega_{M}]$ . Motivated by Guo et al. [20], we learn a $|\mathcal{V}|$ -dimensional categorical distribution over the vocabulary $\mathcal{V}$ for each token $\omega_{i}$ . Let token $\omega_{i}$ follows a categorical distribution with probabilities $\pi_{i,j}$ , where $\sum_{j=1}^{|\mathcal{V}|}\pi_{i,j} = 1$ . $\pi_{i,j}$ can be parameterized and optimized. + +Because sampling from categorical distribution is non-differentiable, we adopt the Gumbel-Softmax reparameterization [23] to obtain smoothed sample $\tilde{\omega}_i$ . Given the smoothed token $\tilde{\omega}_i$ , it can be converted to token embeddings as a linear combination of vocabulary embeddings $e(\tilde{\omega}_i) = \sum_{j=1}^{|\mathcal{V}|} \tilde{\omega}_{i,j} e(j)$ , where $e(j)$ is the pretrained token embedding of the $j$ -th token. A proper choice of the prompt distribution is critical for inversion on text-to-image DMs. We provide a detailed analysis in Appendix Sec. 9. + +# 5. Experiment + +In this section, we conduct a series of experiments from different perspectives to demonstrate the validity of InvMM. First, we conduct controlled studies to verify the effectiveness of the adaptive algorithm (Sec. 5.2). Based on this, we show that InvMM can reasonably elucidate the influence + +![](images/4813e931ea36e647500fa196e4d782a40002627e60c75eb8762b1d7172665c3a.jpg) +Figure 3. The influence of weight $\lambda$ on measuring memorization. + +of various factors on memorization (Sec. 5.3) and identify highly memorized images (Sec. 5.4). Finally, we discuss how memorization differs from membership (Sec. 5.5). + +# 5.1. Experiment setup + +Dataset. We use four datasets for evaluation, i.e., CIFAR-10 [31], CelebAHQ [26], FFHQ [27] and LAION-Aesthetics V2 [47]. For subsequent experiments, we construct four randomly sampled subsets of CelebAHQ and FFHQ: CelebAHQ-250, CelebAHQ-2.5k, FFHQ-600 and FFHQ-6k. We ensure that the larger subset includes the smaller subset. The LAION dataset is extremely large. To keep computational costs manageable, we collect a subset of images. We first include 76 heavily memorized images discovered by Webster [58]. We call this set of images the confirmed set. In addition, we collected another 50 images with high duplication rates according to the statistics from Webster et al [60]. This set of suspects, named as the suspicious set, are assured to be disjoint with the confirmed set. Images in the confirmed set can be triggered by their training captions while those in the suspicious set cannot. Finally, another 50 randomly sampled images are used for comparison, referred to as the normal set. + +Diffusion Models. Following the implementations of DDPM [22] and Latent Diffusion Model (LDM) [43], we train DDPM on CIFAR-10 and LDM on the subsets of CelebAHQ and FFHQ. Pre-trained LDMs on the whole CelebAHQ and FFHQ are also used. For text-to-image generation, we utilize Stable Diffusion (SD) v1.4, v2.1 and v3.5 trained on LAION. + +Inversion. The experiment setting and additional results are deferred to Appendix Sec. 8. + +# 5.2. Adaptive algorithm + +The core design of Algorithm 1 is the dynamic reweighting of the two loss items. To verify its effectiveness, we compare to fixed weights $\lambda = 1$ and $\lambda = 0$ . $\lambda = 1$ equals DM's original training objective (Eqs. (3) and (4)) and $\lambda = 0$ totally ignores the normality constraint. + +A comparison of the resulting memorization score under different settings can be observed in Fig. 3. Fixing $\lambda$ to 1 cannot successfully invert most images: $92.4\%$ and $99.7\%$ images of CelebAHQ and FFHQ result in infinite + +![](images/afc7359e3bf733e790b4837efe24af4f9053d825b4c428a1fd9f1919c696c31c.jpg) +(a) CIFAR-10 (DDPM) + +![](images/7cbab21a67950b263c4df02e499421375d803ae2b9124a6c9ce6ed16403c1782.jpg) +(b) LAION (SD) + +![](images/8509469aed809b5e7f006632522b02cd7a0bd58dd90fd21021f3bdbd9e722bc6.jpg) +(c) CelebAHQ (LDM) + +![](images/c07f7eee74e85a37e97ce2de550bb087a13fcd96f61b0629fe15196d7537033b.jpg) +(d) FFHQ (LDM) +Figure 4. Random samples generated using latent noise vectors drawn from inverted distributions. The original training images are in the first column within each block. + +memorization scores. In contrast, both $\lambda = 0$ and adaptive reweighting successfully invert all images. However, Fixing $\lambda$ to 0 leads to larger resulting memorization scores, which indicates underestimation of the extent of memorization. + +Figure 4 showcases the generation results with the learned distributions. With the adaptive algorithm, we successfully perform inversion across various datasets and DMs that the randomly generated samples closely resemble the training images. The results form the premise of InvMM that memorization of samples is measured in the state they are made replicable. + +# 5.3. Influence factors + +Both theoretical analysis and empirical investigation point out that dataset complexity [15, 52], dataset size [15, 52], cardinality of latent vector set [54] and data duplication [60] are important factors for memorization. We thus verify the correctness of our measure by utilizing it to describe how memorization trends vary with changes in these factors. + +Dataset complexity, size and training epoch. We begin by simultaneously validating dataset complexity, size and cardinality of latent vector set on CelebAHQ and FFHQ. Following Feng et al. [15], we quantify dataset complexity with Intrinsic Dimensionality (ID) [32]. The cardinality of latent vector set can be controlled by adjusting training epochs (if in each iteration a fresh random latent vector is picked). However, we argue that previous works [15, 51] obfuscate the effect of dataset size and training epochs, as models are trained for different epochs (probably more) on smaller subsets to the full dataset. Consequently, on smaller subsets, each training image would see different number of + +![](images/f4d1f16b45c49edd9e1af1c325a3cefe3dc3f41c6631558aedbc85e55eeadba7.jpg) +Figure 5. Dataset complexity & size vs. memorization. Models are trained for 943 epochs on CelebAHQ and 446 epochs on FFHQ. + +![](images/16a4ac4915f23c078a5ddf9269ee5bf89ba5c915521825c31780f30a252ad3f4.jpg) +Figure 6. Training epoch vs. memorization. Models are trained for multiples of the default epochs. + +![](images/be263d4e47840684b09f3eef8498f1596034b2e7ae6bbe6be9a46c09052a0fbf.jpg) +Figure 7. Duplication vs. memorization. + +![](images/a09685385a9420faf2a8a648f48ffcea348e5994d9418f4e1033878785a5823b.jpg) +Figure 8. Dedduplication vs. memorization. + +latent vectors. Therefore, in our experiment, we decouple these two factors by training for the same epochs as on the full dataset when studying the influence of dataset size. + +As can be observed in Fig. 5, dataset complexity and size seem not present high correlation with memorization, as indicated by our measure. When trained for longer time (Fig. 6), each model presents a larger extent of memorization. The results suggest that dataset complexity and size are not independent determining factors, and InvMM accurately reflects the impact of these three factors. + +Data duplication. We then validate the effect of data duplication in DMs' training set through our measure. We + +
MetricDDPM (CIFAR-10)SD v1.4 (LAION)
AUCTPR@1%FPRAUCTPR@1%FPR
ε0-loss0.9810.4140.9550.000
x0-loss0.1800.0000.9750.342
MLOO [55]0.4180.000--
Wen [61]--0.9990.868
Ren [42]--0.9750.355
InvMM0.9970.8990.9740.908
+ +Table 1. Detection performance on DDPM and SD v1.4. + +![](images/738b0ece1612e53c175970844ff5b77c61f3897c419cdf15b176aebb71afb08d.jpg) +(a) +Figure 9. (a) Memorization score distribution on three datasets. (b) Adversarial prompt for an example from the suspicious set. + +![](images/5c4d49bbcfca264b237986853ff93da2e73d281ac130b4bdfad5392c8e1aefbd.jpg) +Prompt: +(b) + +<|endoftext|>eaglegotcha +forever walaccidents concludes +trailer disodana crying ner banana!# conan ""gazing +<|endoftext|>shark + +train a series of LDMs on CelebAHQ-2.5k and FFHQ-6k, with their subsets CelebAHQ-250 and FFHQ-600 duplicated 5, 10, 15 and 20 times. Figure 7 depicts the memorization scores of duplicated images. As can be observed, data duplication rapidly worsens memorization in DMs and our measure is sensitive to these changes. With increasing duplication times, the larger FFHQ-6k dataset presents a stronger resistance to memorization. + +For CIFAR-10 and LAION which originally contain duplicates, we observe the effect of dedduplication. We deduplicate CIFAR-10 with SSCD embeddings and DBSCAN clustering [14]. This results in 666 identified clusters, only one image of which is retained for training. The memorization is evaluated on the retained 666 images before and after dedduplication. On LAION, as compared to SD v1.4, SD v2.1 and v3.5 deduplicate the training set. Memorization is evaluated on the confirmed and suspicious sets. Figure 8 shows the results, with infinite scores on the top. As can be observed, InvMM clearly exhibits the benefits that dedduplication brings. + +# 5.4. Detection + +For InvMM to be a correct measure of memorization, it should be capable of distinguishing highly memorized images from normal ones. To construct a reliable test set, we conduct a nearest neighbor test following previous studies [9, 51, 59] to discover highly memorized images from DM's training set, i.e., highly memorized images consist of training images that can appear at sampling time. Specifically, we consider DDPM trained on CIFAR-10, LDMs + +![](images/ecf7f3b0f8c5a99cfe6a7764d4f1cbc14195af6ccea105c77a61f434163181fd.jpg) +Figure 10. Performance comparison of membership inference on FFHQ-6k, across different training epochs and duplication times. + +![](images/10e9dcea9605bbf1ebf04c48da64361236540c32a9fd50cfa70310663eccc841.jpg) + +trained on CelebAHQ and FFHQ, and SDs trained on LAION. With their official implementations, only DDPM and SD v1.4 present obvious replications, which are thus used for evaluation. For DDPM, the 99 discovered training images appearing in 1M random samples are used as positive samples and another 1k different training images are included as negative samples; for SD v1.4, in accordance with the setting of previous works, we deem images in the confirmed set as positive and those in the suspicious and normal set as negative, considering that captions in the confirmed set can cause replication of their paired images while those in the suspicious and normal set cannot. + +Table 1 presents the detection performance assessed via AUC and TPR@1%FPR. For the unconditional DDPM on CIFAR-10, training loss metrics (the first 2 rows) show decent results and the original $\epsilon_0$ -loss works much better than the $x_0$ -loss. The differential metric $M^{\mathrm{LOO}}$ is less effective than the $\epsilon_0$ -loss and InvMM. The results indicate that the compared metrics are not able to quantify memorization, as they are not fundamentally commensurable between samples. InvMM measures memorization of each sample by contrasting its sensitive noise distribution to a common standard Gaussian, thus having the same scale. + +For SD v1.4, we compare InvMM against training loss metrics and two state-of-the-art detection metrics [42, 61]. Since all compared approaches require access to image captions, we implement InvMM with fixed prompts during inversion. As shown in Tab. 1, our measure achieves performance comparable to state-of-the-art metrics. The results of compared metrics seem to demonstrate a strong ability to distinguish memorization thus far. However, we argue that the prompt-based task formulation employed in existing studies offers an overly optimistic understanding of memorization. It is because these techniques are actually to predict whether the captions will replicate their paired images, while overlook potential prompts that might cause replication. As observed in Fig. 9a, when simultaneously inverting the input prompt, our measure reveals that some images in the suspicious set are also memorized to a similar extent as those in the confirmed set (due to duplication). Applying GCG [64] attack on images with the lowest InvMM scores in the suspicious set provides additional evidence, as illus + +trated in Fig. 9b: images previously deemed “not memorized” according to their training captions can indeed be replicated by potential adversarial prompts (see Appendix Fig. 13 for more examples). Hence, going beyond the confines of existing research, our measure describes worst-case memorization from an adversarial standpoint, making it more reliable in security-critical scenarios. + +# 5.5. Comparison to membership + +A close topic to memorization is membership inference [8, 13, 49], which reveals the leakage risk of membership information that whether an image is used for training. The common implicit assumption is that member samples are "memorized" more than hold-out samples [13], rendering them distinguishable through statistics like (variants of) training loss or differential loss [8]. Interestingly, the principle of InvMM can provide an explicit clarification of the difference between membership and memorization. + +A comparison to a recent membership inference metric SecMI [13] on FFHQ-6k, as shown in Fig. 10, presents a positive correlation between them. As factors that exacerbate memorization changes, both SecMI and InvMM show increasing performance. There is also a performance gap between SecMI and InvMM, indicating that certain samples from the hold-out set are memorized more than some member samples. This mainly results from the gap between model's metric (training loss) and human's metric (perceptual similarity). Membership inference leverages the model's metric for prediction, while memorization should resort to perceptual metric, e.g., simulated by SSCD. What level of loss leads to image replication depends both on the model and image itself. Low training loss does not necessarily lead to data replication, but would probably distinguish training data from never accessed. On the contrary, high training loss signifying non-membership may still be enough for replication. Therefore, membership does not deterministically lead to high memorization and vice versa. InvMM differs from membership metrics by focusing on the perceptual level and adaptively handles the specific loss landscapes of different samples. + +# 6. Conclusion + +In this work, we dive into measuring memorization for diffusion models. Comprehensive experiments demonstrate the ability of the proposed measure to correctly quantify memorization. In short, an image to a higher extent of memorization is able to be inverted to a more standard and diverse latent distribution. However, our method is also limited by its simple modeling of the sensitive latent noise as a multivariate normal. During our research, we observed many clues confirming that the true distribution is complex. Future work might explore more advanced methods, e.g., a generative network, to fit the sensitive noise. + +# Acknowledgement + +This work was partly supported by the National Key Research and Development Program of China under No. 2022YFB3102100, NSFC under No. U244120033, U24A20336, 62172243, 62402425 and 62402418, the China Postdoctoral Science Foundation under No. 2024M762829, the Zhejiang Provincial Natural Science Foundation under No. LD24F020002, the “Pioneer and Leading Goose” R&D Program of Zhejiang under No. 2025C01082, 2025C02033, 2025C02263 and 2025C02034, the Zhejiang Provincial Priority-Funded Postdoctoral Research Project under No. ZJ2024001 and the Key R&D Program of Ningbo under No. 2024Z115. + +# References + +[1] Midjourney. https://www.midjourney.com. 1 +[2] Artist finds private medical record photos in popular AI training data set. https://arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-photos-in-popular-ai-training-data-set/.1 +[3] DALL·E 3. https://openai.com/dall-e-3.1 +[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2 +[5] Sanjeev Arora and Yi Zhang. Do gans actually learn the distribution? an empirical study. arXiv preprint arXiv:1706.08224, 2017. 2 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2 +[8] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 8 +[9] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 7 + +[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. In The Eleventh International Conference on Learning Representations, 2023. 2 +[11] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10759-10769, 2024. 3 +[12] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 2 +[13] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? In International Conference on Machine Learning, pages 8717-8730. PMLR, 2023. 8 +[14] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 7 +[15] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701-6710, 2021. 2, 6 +[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2023. 3 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[18] Andres Guadamuz. Photographer sues LAION for copyright infringement. https://www.technollama.co.uk/ photographer-sues-laion-for-copyright-infringement, 2023.1 +[19] Ishaan Gulrajani, Colin Raffel, and Luke Metz. Towards gan benchmarks which require generalization. In International Conference on Learning Representations, 2019. 2 +[20] Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. Gradient-based adversarial attacks against text transformers. arXiv preprint arXiv:2104.13733, 2021. 5 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 6 +[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5 +[24] Yue Jiang, Haokun Lin, Yang Bai, Bo Peng, Zhili Liu, Yueming Lyu, Yong Yang, Xingzheng, and Jing Dong. Image-level memorization detection via inversion-based inference + +perturbation. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[25] James Jordon, Jinsung Yoon, and Mihaela Van Der Schaar. Pate-gan: Generating synthetic data with differential privacy guarantees. In International Conference on Learning Representations, 2018. 2 +[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6 +[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[28] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2 +[29] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 3 +[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2 +[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[32] Elizaveta Levina and Peter Bickel. Maximum likelihood estimation of intrinsic dimension. Advances in neural information processing systems, 17, 2004. 6 +[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 3 +[34] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3 +[35] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 1 +[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3 +[37] Helen Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119, 2004. 1 +[38] Kai Packhäuser, Lukas Folle, Florian Thamm, and Andreas Maier. Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classification systems. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 2 +[39] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for + +image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 4 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[42] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, pages 340-356. Springer, 2024. 2, 7, 8 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6 +[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2 +[46] Pamela Samuelson. Generative ai meets copyright. Science, 381(6654):158-161, 2023. 2 +[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1, 6 +[48] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1 +[49] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 8 +[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2, 3 +[51] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? + +investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 2, 4, 6, 7 +[52] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 2, 6 +[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 2 +[54] N Vaishnavh, C Raffel, and IJ Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2, 6 +[55] Gerrit van den Burg and Chris Williams. On memorization in probabilistic deep generative models. Advances in Neural Information Processing Systems, 34:27916-27928, 2021. 2, 7 +[56] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 2 +[57] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 3 +[58] Ryan Webster. A reproducible extraction of training images from diffusion models. CoRR, abs/2305.08694, 2023. 2, 6 +[59] Ryan Webster, Julien Rabin, Loic Simon, and Frédéric Jurie. Detecting overfitting of deep generative networks via latent recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11273-11282, 2019. 2, 7 +[60] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of LAION-2B. arXiv preprint arXiv:2303.12733, 2023. 2, 6 +[61] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 7, 8 +[62] Xiaoyu Wu, Yang Hua, Chumeng Liang, Jiaru Zhang, Hao Wang, Tao Song, and Haibing Guan. Cgi-dm: Digital copyright authentication for diffusion models via contrasting gradient inversion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10812-10821. IEEE Computer Society, 2024. 3 +[63] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3 +[64] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 8, 2 \ No newline at end of file diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/images.zip b/2025/An Inversion-based Measure of Memorization for Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f1b381e882ef4fe6c13149a0aa171a0369d602d --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58cee999b8ba9c34831afca12ef76d8d8dcb708ccb7626113ccdaa50cdbac287 +size 360926 diff --git a/2025/An Inversion-based Measure of Memorization for Diffusion Models/layout.json b/2025/An Inversion-based Measure of Memorization for Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..195d0e9f966d6ccd53472c20178499b0cbbba4d0 --- /dev/null +++ b/2025/An Inversion-based Measure of Memorization for Diffusion Models/layout.json @@ -0,0 +1,10575 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 97, + 103, + 514, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 103, + 514, + 120 + ], + "spans": [ + { + "bbox": [ + 97, + 103, + 514, + 120 + ], + "type": "text", + "content": "An Inversion-based Measure of Memorization for Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 143, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 143, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 149, + 143, + 459, + 185 + ], + "type": "text", + "content": "Zhe Ma, Qingming Li, Xuhong Zhang, Tianyu Du, Ruixiao Lin, Zonghui Wang* Shouling Ji, Wenzhi Chen* Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 187, + 498, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 187, + 498, + 200 + ], + "spans": [ + { + "bbox": [ + 109, + 187, + 498, + 200 + ], + "type": "text", + "content": "{mz.rs, liqm, zhangxuhong, zjradty, linruixiao, zhwang, sji, chenwz}@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 252, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 297, + 517 + ], + "type": "text", + "content": "The past few years have witnessed substantial advances in image generation powered by diffusion models. However, it was shown that diffusion models are susceptible to training data memorization, raising significant concerns regarding copyright infringement and privacy invasion. This study delves into a rigorous analysis of memorization in diffusion models. We introduce InvMM1, an inversion-based measure of memorization, which is based on inverting a sensitive latent noise distribution accounting for the replication of an image. For accurate estimation of the measure, we propose an adaptive algorithm that balances the normality and sensitivity of the noise distribution. Comprehensive experiments across four datasets, conducted on both unconditional and text-guided diffusion models, demonstrate that InvMM provides a reliable and complete quantification of memorization. Notably, InvMM is commensurable between samples, reveals the true extent of memorization from an adversarial standpoint and implies how memorization differs from membership. In practice, it serves as an auditing tool for developers to reliably assess the risk of memorization, thereby contributing to the enhancement of trustworthiness and privacy-preserving capabilities of diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 541, + 135, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 135, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 135, + 553 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 562, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 296, + 682 + ], + "type": "text", + "content": "Diffusion Models (DMs) [22, 50] have shown impressive capabilities in the generation of images [41, 43], videos [6], 3D point cloud [35], etc. These techniques lay the foundation for commercial systems or communities such as Stable Diffusion [43], Midjourney [1], DALL-E [3, 41] and Imagegen [44], which have attracted millions of active users. The popularity of diffusion models stems from their hierarchical denoising process, which offers high stability for training on billions of data [47] and scalability for multimodal conditional generation [41, 44]." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 329, + 228, + 541, + 342 + ], + "blocks": [ + { + "bbox": [ + 329, + 228, + 541, + 342 + ], + "lines": [ + { + "bbox": [ + 329, + 228, + 541, + 342 + ], + "spans": [ + { + "bbox": [ + 329, + 228, + 541, + 342 + ], + "type": "image", + "image_path": "aade137ef013e9e5b15ef790874e3223b1f161b84c406e1f4c0e41bb4bfffbdb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 353, + 555, + 374 + ], + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 374 + ], + "type": "text", + "content": "Figure 1. Schematic illustration of diffusion model inversion, optionally with conditions." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 399, + 555, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 542 + ], + "type": "text", + "content": "However, large-scale datasets used to train the prevalent DMs, e.g., the open-source image-caption dataset LAION [47], are widely acknowledged to contain content that will raise concerns regarding copyright infringement and privacy invasion. For instance, recent reports revealed that LAION could refer to photographers' works without authorization [18] and include private medical photographs [2]. With the uncurated data for training, diffusion models are likely to generate content that infringes the creators' rights or exposes private information [9, 48]. In this work, we focus on memorization in DMs, one of the most critical issues related to training data misuse." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "Memorization usually manifests itself at sampling time, when a model generates images closely resembling training images. It can turn into a problem both technically and practically. Technically, heavy memorization indicates failure of generalization to some degree because a probabilistic generative model is supposed to produce novel and diversified images rather than frequently replicate those in the training set. Practically, memorization might also raise ethical and legal concerns. First, even the model developers are authorized to train their model with protected images, the image owners will never expect their works to be exposed to arbitrary users due to memorization as this would cause indisciplinable dissemination, violating the contextual integrity [37]. Second, in past years, text-to-image models" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 147, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 147, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 147, + 703 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 226, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 226, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 226, + 712 + ], + "type": "text", + "content": "1https://github.com/Maryeon/InvMM" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16959" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "have been facing lawsuits for generating derivative images that mimic the style of artists. Compared to derivative generations whose legality is still in pending [46], exact replication of copyrighted images is undisputedly intolerable. Finally, research on privacy protection [25, 38] has proposed to use synthetic data in place of real data to prevent sharing of private information. For this goal, potential memorization must be carefully circumvented." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 298, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 298, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 298, + 483 + ], + "type": "text", + "content": "Recent studies [9, 51, 52, 58] have confirmed the existence of some heavily memorized training data in (text-guided) DMs. They either draw random samples from the model and then detect risky generations of low diversity [9], search for nearest neighbors in the training set [51, 52], or directly identify abnormal prompts [58]. However, conducting a practical analysis of memorization with these approaches still faces challenges. Transforming the memorization of an image to the presence or absence in samples from the model is not a reliable solution. While certain images do not appear in a set of samples, it does not demonstrate that they are not highly memorized [55]. For DMs with large latent space, adequate samples should be tested to ensure as complete coverage of the latent space as possible, which could be up to millions of tests [9, 58]. Furthermore, things will get much more complicated in conditional DMs, e.g., text-to-image generation. The additional conditions expand the input space of DMs, making the memorization of an image also depend on the specific condition. Existing studies exploit training captions to discover highly memorized training images or perform memorization detection [42, 61]. However, attributing the memorization risk of an image to its coupled training caption is incomplete. It still remains unknown that whether the other images not triggered by their coupled captions are also to a large extent of memorization and to be triggered by potential prompts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 713 + ], + "type": "text", + "content": "Our contributions are threefold. (1) To overcome the challenges, we propose an Inversion-based Measure of Memorization (InvMM) for DMs, a continuous instance-level measure that takes an image as input, outputs a scalar quantifying its level of memorization. As shown in Fig. 1, the inversion process locates a sensitive latent noise distribution and optionally another condition distribution for the target image, such that generating with noise and condition drawn from the inverted distributions will produce samples exactly the same as the target image. InvMM is then defined as the minimum KL divergence of the sensitive noise distribution to the standard Gaussian distribution. Compared to existing methods, InvMM is more reliable and complete in that it proactively searches for all potential sensitive inputs and does not rely on limited test cases from the training set (e.g., captions). (2) To calculate the measure, we design an adaptive optimization algorithm motivated by the TCP congestion control protocol, which is demonstrated to be effective to alleviate overestimate and underestimate" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "of memorization. (3) Through comprehensive experiments across four datasets, encompassing both unconditional and text-guided DMs, we present the superiority of InvMM in analyzing memorization. Compared to alternative memorization measures, detection metrics and membership inference metrics, InvMM is commensurable between samples, reflects the true memorization level from an adversarial perspective and clarifies the fundamental difference between memorization and membership. As an auditing tool, it has great potential in discovering highly memorized images, validating safety against memorization for any image, determining appropriate training settings, etc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 230, + 399, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 399, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 399, + 243 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 502, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 502, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 502, + 264 + ], + "type": "text", + "content": "2.1. Memorization of generative models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 270, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 556 + ], + "type": "text", + "content": "Memorization has raised wide attention on various generative models, including GAN [17], VAE [30], language models [7] and DM [50]. There have been studies on training algorithms [4, 22, 45] and evaluation protocols [19] to improve the generalization ability of generative models that would potentially circumvent memorization. Empirical and theoretical studies also find that dataset complexity and size [15, 51], training time [54], training data duplication [10, 60] and model capacity [5, 10] are influential factors of memorization. Several works [54, 59] focus on the latent space (noise space) to analyze memorization. Nagarajan et al. [54] demonstrate that when trained on sufficient random noise vectors, a generator is guaranteed to memorize training data. Webster et al. [59] measure overfitting of GANs via the reconstruction error at the optimal noise vector. We take a similar perspective on the latent space to analyze memorization while aim to reach an instance-level measure of memorization rather than reveal overall properties of models. In addition, Van der Burg et al. [55] also propose an instance-level measure of memorization as the changed likelihood of one sample when removing it from the training set. However, the measure is not necessarily comparable between datasets, which is important to understand the true level of memorization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Specifically for diffusion models, recent research [9, 51, 52] demonstrates that text-to-image diffusion models suffer from memorization. Some heavily memorized images are discovered from numerous generated samples by membership inference [9] or searching for the most similar training images with image matching models [51]. Webster [58] further provides more efficient attacks to extract training images by identifying abnormal prompts. Subsequent works [42, 61] focus on prompt detection that will trigger training image replication and memorization mitigation [12] at inference time. In response to the disclosed threats, Vyas et al. [56] propose a provable copyright protection method to prevent replication of sensitive training" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16960" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 191 + ], + "type": "text", + "content": "images. Our method improves upon existing methods by being commensurable between different samples, basing on images themselves rather than prompts, and working for any samples (not necessarily training samples). A recent work by Jiang et al. [24] also considers prompt-free memorization. However, their method extends Wen et al.'s findings [61] and is restricted to text-guided diffusion models. Our method is established on the property of latent space thus applies to both unconditional and conditional diffusion models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 214, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 214, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 214, + 210 + ], + "type": "text", + "content": "2.2. Inversion of diffusion models" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 312 + ], + "type": "text", + "content": "Inversion techniques in diffusion models are widely studied for image editing [16, 36, 63], copyright protection [62], generated content detection [11, 57], etc. Through inversion, the object, style and concept contained in the source images can be compressed in latent noises or input token embeddings (for text-to-image models). Then the inverted latent noises or input token embeddings are utilized to generate similar images that preserve the desired content." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 396 + ], + "type": "text", + "content": "We leverage inversion techniques to analyze memorization in diffusion models. In contrast, we invert a sample to a distribution consisting of all feasible latent vectors rather than to a single vector [11, 16, 36, 57, 62]. We also adopt more efficient training objective for inversion, without the need to iteratively invert a specific sampling procedure [11, 36, 57, 62]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 135, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 135, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 135, + 419 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": "Basically, DMs [22, 50] are likelihood-based generative models that learn to maximize the likelihood " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " of observed data " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": ". Directly calculating " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " is intractable for continuous data such as image. By contrasting to a sequence of latent variables " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "x_{1:T}" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": ", the log-likelihood " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " can be maximized indirectly by maximizing its variational lower bound [29, 34]:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 517, + 295, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 517, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 87, + 517, + 295, + 569 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) = \\log \\int p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0: T}\\right) d \\boldsymbol {x} _ {1: T} (1) \\\\ \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1: T} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log \\frac {p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0 : T}\\right)}{q \\left(\\boldsymbol {x} _ {1 : T} \\mid \\boldsymbol {x} _ {0}\\right)} \\right] (2) \\\\ \\end{array}", + "image_path": "918ae90d35aa9c87afee92727632a651eeff002a1cf10dd8bc023cfd5b066678.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "content": "Assume both transitions " + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "inline_equation", + "content": "p_{\\pmb{\\theta}}(\\pmb{x}_{0:T})" + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{1:T}|\\pmb{x}_0)" + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "content": " are Markov chains, i.e., " + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{1:T}|\\pmb{x}_0) = \\prod_{t=1}^{T} q(\\pmb{x}_t|\\pmb{x}_{t-1})" + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "inline_equation", + "content": "p_{\\pmb{\\theta}}(\\pmb{x}_{0:T}) = p(\\pmb{x}_T)\\prod_{t=1}^{T} p_{\\pmb{\\theta}}(\\pmb{x}_{t-1}|\\pmb{x}_t)" + }, + { + "bbox": [ + 55, + 575, + 296, + 627 + ], + "type": "text", + "content": ", the variational lower bound in Eq. (2) expands as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 634, + 295, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 634, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 66, + 634, + 295, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0}\\right) \\geq \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {1} \\mid \\boldsymbol {x} _ {0}\\right)} \\left[ \\log p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {0} \\mid \\boldsymbol {x} _ {1}\\right) \\right] \\\\ - \\sum_ {t = 2} ^ {T} \\mathbb {E} _ {q \\left(\\boldsymbol {x} _ {t} \\mid \\boldsymbol {x} _ {0}\\right)} [ D _ {\\mathrm {K L}} (q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\boldsymbol {\\theta}} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}\\right)) ] (3) \\\\ - D _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {T} \\mid \\boldsymbol {x} _ {0}\\right) \\| p \\left(\\boldsymbol {x} _ {T}\\right)\\right) (4) \\\\ \\end{array}", + "image_path": "485f8ff67c3f9175d6d36b477f8bc771e0a1ea53f4227a91e959ba8f3d69901a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{x}_T)" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " is generally assumed to be a simple distribution such as standard Gaussian " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "content": "The transition " + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{1:T}|\\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "content": " is not learned but replaced by a manually crafted non-parametric diffusion process. The transition " + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{0:T})" + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "content": " is the corresponding reverse (or generation, sampling) process. The diffusion process gradually adds Gaussian noises to the input " + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "content": " according to a weight schedule " + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "inline_equation", + "content": "\\alpha_{1:T}" + }, + { + "bbox": [ + 313, + 96, + 554, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 178, + 553, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 178, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 329, + 178, + 553, + 191 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\alpha_ {t}} \\boldsymbol {x} _ {t - 1} + \\sqrt {1 - \\alpha_ {t}} \\boldsymbol {\\epsilon} _ {t - 1}, \\boldsymbol {\\epsilon} _ {t - 1} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {5}", + "image_path": "36cc010e5f808517670b7ab9fd9c22784090dd6ea9fd2b2a925a63b635c460fd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 194, + 553, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 194, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 349, + 194, + 553, + 207 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} \\boldsymbol {x} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\boldsymbol {\\epsilon} _ {0}, \\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}) \\tag {6}", + "image_path": "151ec343f90cfeab07f4338f7227dcf9f88cd809e29ace88b335fd6f3afbd49b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": " decreases over time and finally to almost zero at the last step " + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_T | \\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": " approaches to a standard Gaussian and thus Eq. (4) approximately equals to 0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "text", + "content": "If the diffusion process is divided into sufficient steps, each reverse transition " + }, + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_{t-1}|\\pmb{x}_t)" + }, + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "text", + "content": " can also be approximated by a Gaussian transformation with learnable parameters that is trained to match the corresponding posterior diffusion transition " + }, + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "inline_equation", + "content": "q(\\pmb{x}_{t-1}|\\pmb{x}_t,\\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 266, + 554, + 362 + ], + "type": "text", + "content": ". Omitting further details [34], the following simplified objective [22] as an estimate of the variational lower bound is commonly used for training:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 337, + 371, + 553, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 371, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 337, + 371, + 553, + 386 + ], + "type": "interline_equation", + "content": "l _ {d e} = \\mathbb {E} _ {\\boldsymbol {\\epsilon} _ {0} \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), t \\sim \\mathcal {U} \\{1, T \\}} \\left[ \\| \\boldsymbol {\\epsilon} _ {0} - \\boldsymbol {\\epsilon} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {7}", + "image_path": "b1b2df64454893d505b71f68309f108bbb4bf1fdf6fe65a636837ca53c5f41e1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " is a neural network that takes the noisy data " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " and timestep " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " as input and predicts the added noise " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": ". After training, the vanilla sampling procedure starts with a random Gaussian noise " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " and removes the predicted noise stepwise. However, it is extremely slow to generate an image as it must invoke the network " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 394, + 554, + 502 + ], + "type": "text", + "content": " times (typically 1000 steps). To mitigate the problem, a variety of efficient sampling algorithms are proposed, such as DDIM sampler [53] and PLMS sampler [33]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 513, + 457, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 457, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 457, + 525 + ], + "type": "text", + "content": "4. Inversion-based measure" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": ", we aim to measure to what extent " + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": " is memorized by a diffusion model (DM). We propose to invert the image into a latent noise distribution, such that noise vectors drawn from it will lead to replication of the image " + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_0" + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": ". Our memorization measure is established in this inversion process. In this section, we introduce a sensitivity definition of the inverted noise distribution (Sec. 4.1), elaborate the memorization measure (Sec. 4.2) and give an adaptive algorithm for its calculation (Sec. 4.3). Ultimately, we present an example to measure memorization in conditional DM, specifically text-to-image DM (Sec. 4.4)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 672, + 485, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 485, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 485, + 685 + ], + "type": "text", + "content": "4.1. Sensitivity of latent distribution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "The diversity of DM generated content lies in the randomness of latent noise vectors. Starting from different noise" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16961" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "vectors, the generated contents generally vary. In the context of memorization, this motivates us to search for sensitive noise vectors that will replicate the target image. We then figure out \"how many\" sensitive noise vectors there are and adopt it as a meaningful measure." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": "Considering the fact that the common latent noise is sampled from a standard Gaussian distribution, we correspondingly assume the sensitive noise vectors follow a distribution " + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)" + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": ". The sensitivity function " + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "inline_equation", + "content": "S(q_{\\varphi})" + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": " of a noise distribution can be defined as the probability that generated image " + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "inline_equation", + "content": "G(\\epsilon_0)" + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": " obtained from random noise " + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "inline_equation", + "content": "\\epsilon_0 \\sim q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)" + }, + { + "bbox": [ + 55, + 133, + 295, + 218 + ], + "type": "text", + "content": " replicates the target image:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 226, + 295, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 226, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 90, + 226, + 295, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left(q _ {\\varphi}\\right) = \\int \\mathbf {1} _ {d \\left(G \\left(\\boldsymbol {\\epsilon} _ {0}\\right), \\boldsymbol {x} _ {0}\\right) \\leq \\beta} q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x} _ {0}\\right) d \\boldsymbol {\\epsilon} _ {0} \\tag {8}", + "image_path": "d66c755dfe484218ab20f7a8d480f6303dd8eaba34d62a25349d000bcc1e6176.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_v" + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "content": " is an indicator function that equals 1 if " + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "content": " is true and 0 otherwise, " + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "inline_equation", + "content": "d(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "content": " is a distance metric between two images and " + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 260, + 295, + 297 + ], + "type": "text", + "content": " is an appropriate threshold." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 297, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 441 + ], + "type": "text", + "content": "The choice of the distance metric depends on specific applications. In applications that accept derivative generation, a strict metric that only detects eidetic generation can be utilized. While under some circumstances even similar generation should be circumvented, a looser metric is needed. In this work, we consider exact replication that the generated image and target image are exactly the same. This consideration to some extent reduces the ambiguity of determining the similarity of two images. Following the investigation of Somepalli et al. [51], we implement the metric as the Euclidean distance between embeddings extracted by the pretrained copy detection model SSCD [39]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 188, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 188, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 188, + 462 + ], + "type": "text", + "content": "4.2. Quantify memorization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": "Based on the definition of sensitivity, we propose an Inversion-based Measure of Memorization (InvMM), as the best achievable average normality of " + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\epsilon | \\boldsymbol{x})" + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": " under the condition that " + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "inline_equation", + "content": "S(q_{\\varphi}) = 1" + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 525, + 294, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 525, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 110, + 525, + 294, + 566 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {\\varphi} \\frac {1}{N} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\boldsymbol {\\epsilon} _ {0} \\mid \\boldsymbol {x}\\right) \\| p \\left(\\boldsymbol {\\epsilon} _ {0}\\right)\\right) \\\\ \\text {s . t .} \\mathcal {S} \\left(q _ {\\varphi}\\right) = 1 \\tag {9} \\\\ \\end{array}", + "image_path": "9945caeb529276b63113f86c6f926edf22847c3161ce5e4094dab7970ae4ccdd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": " is the dimensionality of latent noise space and " + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "inline_equation", + "content": "p(\\epsilon_0)" + }, + { + "bbox": [ + 55, + 575, + 295, + 635 + ], + "type": "text", + "content": " is the prior distribution used in DM, which is often selected to be standard Gaussian. The normality of inverted noise distribution is the KL divergence to the prior. A lower value of InvMM means greater degree of memorization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 659 + ], + "type": "text", + "content": "Explaining InvMM. Partitioning the normality gives us a more intuitional understanding of the measure. Notice that" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 682, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 682, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 97, + 682, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {\\mathrm {K L}} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} | \\boldsymbol {x} _ {0}\\right) \\| p (\\epsilon_ {0})\\right) = \\\\ \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), p \\left(\\epsilon_ {0}\\right)\\right) - \\mathrm {H} \\left(q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right)\\right) \\tag {10} \\\\ \\end{array}", + "image_path": "e11e6f50ad39a4df026b184f3b3cdb786c6df62278a9a3d05e27960be6f17a12.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 347, + 73, + 400, + 143 + ], + "blocks": [ + { + "bbox": [ + 347, + 73, + 400, + 143 + ], + "lines": [ + { + "bbox": [ + 347, + 73, + 400, + 143 + ], + "spans": [ + { + "bbox": [ + 347, + 73, + 400, + 143 + ], + "type": "image", + "image_path": "8693c96cea934401c724b0022745b09ec19951ada2eeabab4e48522f41f49a67.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 149, + 392, + 159 + ], + "lines": [ + { + "bbox": [ + 353, + 149, + 392, + 159 + ], + "spans": [ + { + "bbox": [ + 353, + 149, + 392, + 159 + ], + "type": "text", + "content": "(a) Optimal" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 406, + 73, + 462, + 144 + ], + "blocks": [ + { + "bbox": [ + 406, + 73, + 462, + 144 + ], + "lines": [ + { + "bbox": [ + 406, + 73, + 462, + 144 + ], + "spans": [ + { + "bbox": [ + 406, + 73, + 462, + 144 + ], + "type": "image", + "image_path": "3c8cef596b7cc85255f22c75a5f3bd0477629e5b8602e7f4b124a296366615f9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 149, + 462, + 159 + ], + "lines": [ + { + "bbox": [ + 406, + 149, + 462, + 159 + ], + "spans": [ + { + "bbox": [ + 406, + 149, + 462, + 159 + ], + "type": "text", + "content": "(b) Overestimate" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 467, + 150, + 525, + 159 + ], + "lines": [ + { + "bbox": [ + 467, + 150, + 525, + 159 + ], + "spans": [ + { + "bbox": [ + 467, + 150, + 525, + 159 + ], + "type": "text", + "content": "(c)Underestimate" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 473, + 80, + 521, + 139 + ], + "blocks": [ + { + "bbox": [ + 473, + 80, + 521, + 139 + ], + "lines": [ + { + "bbox": [ + 473, + 80, + 521, + 139 + ], + "spans": [ + { + "bbox": [ + 473, + 80, + 521, + 139 + ], + "type": "image", + "image_path": "22f216fac8e19c8856f63f792ea342611b9259c4446177d177231e17ffe96e7d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 169, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 190 + ], + "type": "text", + "content": "Figure 2. Illustration of our memorization measure. The gray blocks indicate the distribution of sensitive latent noise vectors." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0))" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": " is the cross entropy between the inverted distribution and the prior distribution, " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0))" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": " is the differential entropy of the inverted distribution. Minimizing the normality thus encourages a distribution as close to the prior distribution as possible, meanwhile with the largest entropy. For the cross entropy, following the Jensen's inequality, " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "-\\mathrm{H}(q_{\\varphi}(\\epsilon_0|\\pmb{x}_0), p(\\epsilon_0)) = \\mathbb{E}_{q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)}[\\log p(\\epsilon_0)] \\leq \\log \\int q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)d\\epsilon_0" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": ". The term " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)p(\\epsilon_0)" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": " is proportional to the probability of a random noise " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": ", with probability density " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\epsilon_0|\\pmb{x}_0)" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": ", to be sampled from the prior distribution with probability density " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "p(\\epsilon_0)" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": ". In the sense, minimizing the cross-entropy improves the probability that sensitive noise vectors regarding an image " + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\pmb{x}_0" + }, + { + "bbox": [ + 313, + 217, + 555, + 456 + ], + "type": "text", + "content": " are sampled in the normal usage of DM. For the differential entropy, maximizing it will produce a noise distribution involving \"as many\" sensitive noise vectors as possible. In the extreme case, suppose a uniform distribution over the entire latent space having the largest entropy, it means that any noise in the space can cause replication of the target training sample." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 567 + ], + "type": "text", + "content": "InvMM requires to achieve the best trade-off between improving normality and meeting the sensitivity condition. Different images to varied extent of memorization may reach different equilibrium states. Figure 2 illustrates the optimal solution compared to suboptimal choices. Overestimate of memorization produces a distribution with good normality but low sensitivity (Fig. 2b). Conservative inversion of the latent distribution perfectly meets the sensitivity condition while underestimates memorization (Fig. 2c)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 581, + 428, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 428, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 428, + 594 + ], + "type": "text", + "content": "4.3. Adaptive algorithm" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 601, + 553, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 624 + ], + "type": "text", + "content": "We then introduce an adaptive algorithm to solve the inversion problem." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "text", + "content": "Transforming the inversion problem. It is intractable to directly calculate the sensitivity condition. For diffusion models, it can be achieved by minimizing the denoising error " + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "inline_equation", + "content": "l_{de}" + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "text", + "content": " (see Eq. (7)) as a proxy. As can be expected, when " + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "inline_equation", + "content": "l_{de}" + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "text", + "content": " gradually decreases, " + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "inline_equation", + "content": "S(q_{\\varphi})" + }, + { + "bbox": [ + 313, + 627, + 554, + 688 + ], + "type": "text", + "content": " will approach 1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Moreover, denote the average KL divergence in Eq. (9) by " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "l_{kl}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ", the optimization problem can be solved by La" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16962" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 129, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 129, + 85 + ], + "type": "text", + "content": "grangian method:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 95, + 295, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 95, + 295, + 113 + ], + "spans": [ + { + "bbox": [ + 99, + 95, + 295, + 113 + ], + "type": "interline_equation", + "content": "\\min _ {\\varphi , \\lambda} l _ {d e} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right) + \\lambda l _ {k l} \\left(\\boldsymbol {x} _ {0}; \\boldsymbol {\\varphi}\\right), \\lambda \\geq 0 \\tag {11}", + "image_path": "2e6a74d5e40b9543bbd413027464c9dfa1b932a4c8101df458de6ee221008972.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "inline_equation", + "content": "\\lambda \\geq 0" + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "text", + "content": " is a multiplier. We attach the multiplier to " + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "inline_equation", + "content": "l_{kl}" + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "text", + "content": " to ensure adequate optimization of the denoising error " + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "inline_equation", + "content": "l_{de}" + }, + { + "bbox": [ + 55, + 121, + 295, + 146 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": "Different from the standard " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": "-prediction denoising error, we adopt " + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 146, + 296, + 183 + ], + "type": "text", + "content": "-prediction denoising error, which shows more stable inversion efficacy in practice:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 190, + 295, + 232 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 190, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 73, + 190, + 295, + 232 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} l _ {d e} = \\mathbb {E} _ {q _ {\\boldsymbol {\\varphi}} (\\epsilon_ {0} | \\boldsymbol {x} _ {0}), t} \\left[ \\| \\boldsymbol {x} _ {0} - \\boldsymbol {x} _ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\\\ = \\frac {1 - \\bar {\\alpha} _ {t}}{\\bar {\\alpha} _ {t}} \\mathbb {E} _ {q _ {\\varphi} \\left(\\epsilon_ {0} \\mid \\boldsymbol {x} _ {0}\\right), t} \\left[ \\| \\epsilon_ {0} - \\epsilon_ {\\boldsymbol {\\theta}} (\\boldsymbol {x} _ {t}, t) \\| _ {2} ^ {2} \\right] \\tag {12} \\\\ \\end{array}", + "image_path": "303740a439ff50fc6afdb37593674a4b6d94368a601b3622194616ea57687a07.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t" + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "content": " is obtained as in Eq. (6) with added noise drawn from " + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "inline_equation", + "content": "q_{\\varphi}" + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "content": " rather than standard Gaussian." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": "For the noise distribution " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(\\epsilon_0|\\boldsymbol{x}_0)" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ", we assume it to be a multivariate Gaussian " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\boldsymbol{\\mu},\\boldsymbol{\\sigma}^2)" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": " with learnable mean " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": " and diagonal variance " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\varphi = (\\pmb{\\mu},\\pmb{\\sigma}^2)" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ". Then InvMM has a closed-form expression " + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\frac{1}{2N}\\sum_i(\\pmb{\\mu}_i^2 +\\pmb{\\sigma}_i^2 -\\log \\pmb{\\sigma}_i^2 -1)" + }, + { + "bbox": [ + 55, + 264, + 296, + 348 + ], + "type": "text", + "content": ". Although the true distribution of sensitive latent noise might not be fully captured by a Gaussian, we choose this modeling for simplicity and find it effective in experiments." + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 58, + 375, + 298, + 652 + ], + "blocks": [ + { + "bbox": [ + 57, + 360, + 283, + 372 + ], + "lines": [ + { + "bbox": [ + 57, + 360, + 283, + 372 + ], + "spans": [ + { + "bbox": [ + 57, + 360, + 283, + 372 + ], + "type": "text", + "content": "Algorithm 1 InvMM: adaptive weight tuning algorithm." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "lines": [ + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "spans": [ + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": "Input: Noise prediction network " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", target image " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", optimization iteration " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", observation cycle " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", weight increment " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", threshold " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", learning rate " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n1: Initialize " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\varphi \\gets (\\mu \\gets 0, \\sigma \\gets 1)" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\lambda \\gets 1, l_{de}^p \\gets +\\infty" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": ", success " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\gets" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " FALSE \n2: for " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "i \\gets 1" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " do \n3: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\gets l_{de}(\\boldsymbol{x}_0; \\varphi) + \\lambda l_{kl}(\\boldsymbol{x}_0; \\varphi)" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n4: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\varphi \\gets \\varphi - \\gamma \\nabla_{\\varphi} \\mathcal{L}" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n5: if " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "i > 1" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "i \\% C = 0" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " then \n6: if " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "l_{de} - l_{de}^p < \\xi" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " then \n7: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\lambda \\gets \\lambda / 2" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n8: else \n9: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\lambda \\gets \\lambda + \\delta" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "l_{de}^p \\gets l_{de}" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n11: if " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "S(q_\\varphi) = 1" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " then \n12: success " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " TRUE \n13: break \n14: else \n15: " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\lambda \\gets \\lambda + \\delta" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n16: if success then \n17: return " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "\\frac{1}{2N} \\sum_i (\\mu_i^2 + \\sigma_i^2 - \\log \\sigma_i^2 - 1)" + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "text", + "content": " \n18: else \n19: return " + }, + { + "bbox": [ + 58, + 375, + 298, + 652 + ], + "type": "inline_equation", + "content": "+\\infty" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Adaptive weight tuning. To achieve the best trade-off, the weight " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " should be adjusted appropriately for different samples. However, it cannot be optimized via gradient descent together with " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". As the gradient w.r.t. " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "l_{kl}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", which" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "is always nonnegative, " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " will gradually decrease to 0 and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "l_{kl}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " will be totally ignored. To solve the problem, we propose an algorithm that allows adaptive tuning of " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " during inversion, motivated by the additive increase/multiplicative decrease scheme in congestion control of TCP protocol, as in Algorithm 1, line 6-9." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": "Specifically, we monitor the denoising error " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "l_{de}" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " every " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " steps during the optimization process and divide the weight " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " by half if no significant improvement of denoising error is observed in the cycle. Otherwise, " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " slowly increases by a small increment " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " each step. In this way, " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " is adaptively adjusted for different samples to balance the normality and denoising error. When the normality is too strong to further reduce the denoising error, thus to meet the sensitivity condition, it is immediately down-weighted. Otherwise, we gradually strengthen it for better normality. We also perform early stop (line 11-13) when the sensitivity condition is reached to avoid unnecessary normality degradation. The sensitivity condition is tested via the Monte Carlo method. Random latent vectors are drawn to generate samples. If the distances between all of them and the target image are lower than the threshold " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": ", then the inversion is successful. There are cases that only partial samples replicate that target image in the end. We still consider such cases as insignificant memorization and flag them as unsuccessful." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 381, + 423, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 381, + 423, + 395 + ], + "spans": [ + { + "bbox": [ + 314, + 381, + 423, + 395 + ], + "type": "text", + "content": "4.4. Text-to-image DM" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": "For text-to-image DM, we model the prompt part as a distribution as well and optimize it simultaneously with the noise distribution. Suppose an input prompt is tokenized into maximum " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\omega = [\\omega_{1},\\omega_{2},\\dots,\\omega_{M}]" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": ". Motivated by Guo et al. [20], we learn a " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "|\\mathcal{V}|" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": "-dimensional categorical distribution over the vocabulary " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": " for each token " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": ". Let token " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": " follows a categorical distribution with probabilities " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\pi_{i,j}" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\sum_{j=1}^{|\\mathcal{V}|}\\pi_{i,j} = 1" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\pi_{i,j}" + }, + { + "bbox": [ + 313, + 400, + 554, + 510 + ], + "type": "text", + "content": " can be parameterized and optimized." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": "Because sampling from categorical distribution is non-differentiable, we adopt the Gumbel-Softmax reparameterization [23] to obtain smoothed sample " + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{\\omega}_i" + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": ". Given the smoothed token " + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\tilde{\\omega}_i" + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": ", it can be converted to token embeddings as a linear combination of vocabulary embeddings " + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "inline_equation", + "content": "e(\\tilde{\\omega}_i) = \\sum_{j=1}^{|\\mathcal{V}|} \\tilde{\\omega}_{i,j} e(j)" + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "inline_equation", + "content": "e(j)" + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": " is the pretrained token embedding of the " + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 510, + 554, + 620 + ], + "type": "text", + "content": "-th token. A proper choice of the prompt distribution is critical for inversion on text-to-image DMs. We provide a detailed analysis in Appendix Sec. 9." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 632, + 391, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 391, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 391, + 646 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "In this section, we conduct a series of experiments from different perspectives to demonstrate the validity of InvMM. First, we conduct controlled studies to verify the effectiveness of the adaptive algorithm (Sec. 5.2). Based on this, we show that InvMM can reasonably elucidate the influence" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16963" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 73, + 280, + 154 + ], + "blocks": [ + { + "bbox": [ + 73, + 73, + 280, + 154 + ], + "lines": [ + { + "bbox": [ + 73, + 73, + 280, + 154 + ], + "spans": [ + { + "bbox": [ + 73, + 73, + 280, + 154 + ], + "type": "image", + "image_path": "4813e931ea36e647500fa196e4d782a40002627e60c75eb8762b1d7172665c3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 169, + 293, + 180 + ], + "lines": [ + { + "bbox": [ + 56, + 169, + 293, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 293, + 180 + ], + "type": "text", + "content": "Figure 3. The influence of weight " + }, + { + "bbox": [ + 56, + 169, + 293, + 180 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 56, + 169, + 293, + 180 + ], + "type": "text", + "content": " on measuring memorization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 231 + ], + "type": "text", + "content": "of various factors on memorization (Sec. 5.3) and identify highly memorized images (Sec. 5.4). Finally, we discuss how memorization differs from membership (Sec. 5.5)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 162, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 162, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 162, + 251 + ], + "type": "text", + "content": "5.1. Experiment setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 256, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 470 + ], + "type": "text", + "content": "Dataset. We use four datasets for evaluation, i.e., CIFAR-10 [31], CelebAHQ [26], FFHQ [27] and LAION-Aesthetics V2 [47]. For subsequent experiments, we construct four randomly sampled subsets of CelebAHQ and FFHQ: CelebAHQ-250, CelebAHQ-2.5k, FFHQ-600 and FFHQ-6k. We ensure that the larger subset includes the smaller subset. The LAION dataset is extremely large. To keep computational costs manageable, we collect a subset of images. We first include 76 heavily memorized images discovered by Webster [58]. We call this set of images the confirmed set. In addition, we collected another 50 images with high duplication rates according to the statistics from Webster et al [60]. This set of suspects, named as the suspicious set, are assured to be disjoint with the confirmed set. Images in the confirmed set can be triggered by their training captions while those in the suspicious set cannot. Finally, another 50 randomly sampled images are used for comparison, referred to as the normal set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 554 + ], + "type": "text", + "content": "Diffusion Models. Following the implementations of DDPM [22] and Latent Diffusion Model (LDM) [43], we train DDPM on CIFAR-10 and LDM on the subsets of CelebAHQ and FFHQ. Pre-trained LDMs on the whole CelebAHQ and FFHQ are also used. For text-to-image generation, we utilize Stable Diffusion (SD) v1.4, v2.1 and v3.5 trained on LAION." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 580 + ], + "type": "text", + "content": "Inversion. The experiment setting and additional results are deferred to Appendix Sec. 8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 170, + 601 + ], + "type": "text", + "content": "5.2. Adaptive algorithm" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "The core design of Algorithm 1 is the dynamic reweighting of the two loss items. To verify its effectiveness, we compare to fixed weights " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " equals DM's original training objective (Eqs. (3) and (4)) and " + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": " totally ignores the normality constraint." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "A comparison of the resulting memorization score under different settings can be observed in Fig. 3. Fixing " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " to 1 cannot successfully invert most images: " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "92.4\\%" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "99.7\\%" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " images of CelebAHQ and FFHQ result in infinite" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 338, + 69, + 434, + 167 + ], + "blocks": [ + { + "bbox": [ + 338, + 69, + 434, + 167 + ], + "lines": [ + { + "bbox": [ + 338, + 69, + 434, + 167 + ], + "spans": [ + { + "bbox": [ + 338, + 69, + 434, + 167 + ], + "type": "image", + "image_path": "afc7359e3bf733e790b4837efe24af4f9053d825b4c428a1fd9f1919c696c31c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 168, + 424, + 177 + ], + "lines": [ + { + "bbox": [ + 348, + 168, + 424, + 177 + ], + "spans": [ + { + "bbox": [ + 348, + 168, + 424, + 177 + ], + "type": "text", + "content": "(a) CIFAR-10 (DDPM)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 531, + 167 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 531, + 167 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 531, + 167 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 531, + 167 + ], + "type": "image", + "image_path": "7cbab21a67950b263c4df02e499421375d803ae2b9124a6c9ce6ed16403c1782.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 168, + 511, + 177 + ], + "lines": [ + { + "bbox": [ + 456, + 168, + 511, + 177 + ], + "spans": [ + { + "bbox": [ + 456, + 168, + 511, + 177 + ], + "type": "text", + "content": "(b) LAION (SD)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 338, + 178, + 434, + 275 + ], + "blocks": [ + { + "bbox": [ + 338, + 178, + 434, + 275 + ], + "lines": [ + { + "bbox": [ + 338, + 178, + 434, + 275 + ], + "spans": [ + { + "bbox": [ + 338, + 178, + 434, + 275 + ], + "type": "image", + "image_path": "8509469aed809b5e7f006632522b02cd7a0bd58dd90fd21021f3bdbd9e722bc6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 276, + 423, + 285 + ], + "lines": [ + { + "bbox": [ + 349, + 276, + 423, + 285 + ], + "spans": [ + { + "bbox": [ + 349, + 276, + 423, + 285 + ], + "type": "text", + "content": "(c) CelebAHQ (LDM)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 435, + 178, + 531, + 275 + ], + "blocks": [ + { + "bbox": [ + 435, + 178, + 531, + 275 + ], + "lines": [ + { + "bbox": [ + 435, + 178, + 531, + 275 + ], + "spans": [ + { + "bbox": [ + 435, + 178, + 531, + 275 + ], + "type": "image", + "image_path": "c07f7eee74e85a37e97ce2de550bb087a13fcd96f61b0629fe15196d7537033b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 276, + 512, + 285 + ], + "lines": [ + { + "bbox": [ + 454, + 276, + 512, + 285 + ], + "spans": [ + { + "bbox": [ + 454, + 276, + 512, + 285 + ], + "type": "text", + "content": "(d) FFHQ (LDM)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 296, + 554, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 328 + ], + "type": "text", + "content": "Figure 4. Random samples generated using latent noise vectors drawn from inverted distributions. The original training images are in the first column within each block." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "text", + "content": "memorization scores. In contrast, both " + }, + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "text", + "content": " and adaptive reweighting successfully invert all images. However, Fixing " + }, + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 342, + 554, + 389 + ], + "type": "text", + "content": " to 0 leads to larger resulting memorization scores, which indicates underestimation of the extent of memorization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 473 + ], + "type": "text", + "content": "Figure 4 showcases the generation results with the learned distributions. With the adaptive algorithm, we successfully perform inversion across various datasets and DMs that the randomly generated samples closely resemble the training images. The results form the premise of InvMM that memorization of samples is measured in the state they are made replicable." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 480, + 416, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 416, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 416, + 492 + ], + "type": "text", + "content": "5.3. Influence factors" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 569 + ], + "type": "text", + "content": "Both theoretical analysis and empirical investigation point out that dataset complexity [15, 52], dataset size [15, 52], cardinality of latent vector set [54] and data duplication [60] are important factors for memorization. We thus verify the correctness of our measure by utilizing it to describe how memorization trends vary with changes in these factors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Dataset complexity, size and training epoch. We begin by simultaneously validating dataset complexity, size and cardinality of latent vector set on CelebAHQ and FFHQ. Following Feng et al. [15], we quantify dataset complexity with Intrinsic Dimensionality (ID) [32]. The cardinality of latent vector set can be controlled by adjusting training epochs (if in each iteration a fresh random latent vector is picked). However, we argue that previous works [15, 51] obfuscate the effect of dataset size and training epochs, as models are trained for different epochs (probably more) on smaller subsets to the full dataset. Consequently, on smaller subsets, each training image would see different number of" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16964" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 284, + 155 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 284, + 155 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 284, + 155 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 284, + 155 + ], + "type": "image", + "image_path": "f4d1f16b45c49edd9e1af1c325a3cefe3dc3f41c6631558aedbc85e55eeadba7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 295, + 196 + ], + "type": "text", + "content": "Figure 5. Dataset complexity & size vs. memorization. Models are trained for 943 epochs on CelebAHQ and 446 epochs on FFHQ." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 68, + 208, + 284, + 292 + ], + "blocks": [ + { + "bbox": [ + 68, + 208, + 284, + 292 + ], + "lines": [ + { + "bbox": [ + 68, + 208, + 284, + 292 + ], + "spans": [ + { + "bbox": [ + 68, + 208, + 284, + 292 + ], + "type": "image", + "image_path": "16a4ac4915f23c078a5ddf9269ee5bf89ba5c915521825c31780f30a252ad3f4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 323 + ], + "type": "text", + "content": "Figure 6. Training epoch vs. memorization. Models are trained for multiples of the default epochs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 68, + 335, + 284, + 420 + ], + "blocks": [ + { + "bbox": [ + 68, + 335, + 284, + 420 + ], + "lines": [ + { + "bbox": [ + 68, + 335, + 284, + 420 + ], + "spans": [ + { + "bbox": [ + 68, + 335, + 284, + 420 + ], + "type": "image", + "image_path": "be263d4e47840684b09f3eef8498f1596034b2e7ae6bbe6be9a46c09052a0fbf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 428, + 249, + 439 + ], + "lines": [ + { + "bbox": [ + 101, + 428, + 249, + 439 + ], + "spans": [ + { + "bbox": [ + 101, + 428, + 249, + 439 + ], + "type": "text", + "content": "Figure 7. Duplication vs. memorization." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 68, + 442, + 284, + 528 + ], + "blocks": [ + { + "bbox": [ + 68, + 442, + 284, + 528 + ], + "lines": [ + { + "bbox": [ + 68, + 442, + 284, + 528 + ], + "spans": [ + { + "bbox": [ + 68, + 442, + 284, + 528 + ], + "type": "image", + "image_path": "a09685385a9420faf2a8a648f48ffcea348e5994d9418f4e1033878785a5823b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 535, + 253, + 547 + ], + "lines": [ + { + "bbox": [ + 97, + 535, + 253, + 547 + ], + "spans": [ + { + "bbox": [ + 97, + 535, + 253, + 547 + ], + "type": "text", + "content": "Figure 8. Dedduplication vs. memorization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 605 + ], + "type": "text", + "content": "latent vectors. Therefore, in our experiment, we decouple these two factors by training for the same epochs as on the full dataset when studying the influence of dataset size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "As can be observed in Fig. 5, dataset complexity and size seem not present high correlation with memorization, as indicated by our measure. When trained for longer time (Fig. 6), each model presents a larger extent of memorization. The results suggest that dataset complexity and size are not independent determining factors, and InvMM accurately reflects the impact of these three factors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Data duplication. We then validate the effect of data duplication in DMs' training set through our measure. We" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 330, + 70, + 539, + 162 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 539, + 162 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 539, + 162 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 539, + 162 + ], + "type": "table", + "html": "
MetricDDPM (CIFAR-10)SD v1.4 (LAION)
AUCTPR@1%FPRAUCTPR@1%FPR
ε0-loss0.9810.4140.9550.000
x0-loss0.1800.0000.9750.342
MLOO [55]0.4180.000--
Wen [61]--0.9990.868
Ren [42]--0.9750.355
InvMM0.9970.8990.9740.908
", + "image_path": "ae3932e3f976816946359a17afca31ddfed9745fcac3bf7aed08d4897ecef76c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 170, + 537, + 180 + ], + "lines": [ + { + "bbox": [ + 331, + 170, + 537, + 180 + ], + "spans": [ + { + "bbox": [ + 331, + 170, + 537, + 180 + ], + "type": "text", + "content": "Table 1. Detection performance on DDPM and SD v1.4." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 331, + 194, + 432, + 294 + ], + "blocks": [ + { + "bbox": [ + 331, + 194, + 432, + 294 + ], + "lines": [ + { + "bbox": [ + 331, + 194, + 432, + 294 + ], + "spans": [ + { + "bbox": [ + 331, + 194, + 432, + 294 + ], + "type": "image", + "image_path": "738b0ece1612e53c175970844ff5b77c61f3897c419cdf15b176aebb71afb08d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 298, + 386, + 308 + ], + "lines": [ + { + "bbox": [ + 375, + 298, + 386, + 308 + ], + "spans": [ + { + "bbox": [ + 375, + 298, + 386, + 308 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 318, + 553, + 340 + ], + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 340 + ], + "type": "text", + "content": "Figure 9. (a) Memorization score distribution on three datasets. (b) Adversarial prompt for an example from the suspicious set." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 444, + 194, + 541, + 243 + ], + "blocks": [ + { + "bbox": [ + 444, + 194, + 541, + 243 + ], + "lines": [ + { + "bbox": [ + 444, + 194, + 541, + 243 + ], + "spans": [ + { + "bbox": [ + 444, + 194, + 541, + 243 + ], + "type": "image", + "image_path": "5c4d49bbcfca264b237986853ff93da2e73d281ac130b4bdfad5392c8e1aefbd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 243, + 474, + 251 + ], + "lines": [ + { + "bbox": [ + 448, + 243, + 474, + 251 + ], + "spans": [ + { + "bbox": [ + 448, + 243, + 474, + 251 + ], + "type": "text", + "content": "Prompt:" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 487, + 298, + 498, + 307 + ], + "lines": [ + { + "bbox": [ + 487, + 298, + 498, + 307 + ], + "spans": [ + { + "bbox": [ + 487, + 298, + 498, + 307 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 448, + 252, + 535, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 252, + 535, + 292 + ], + "spans": [ + { + "bbox": [ + 448, + 252, + 535, + 292 + ], + "type": "text", + "content": "<|endoftext|>eaglegotcha \nforever walaccidents concludes \ntrailer disodana crying ner banana!# conan \"\"gazing \n<|endoftext|>shark" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 353, + 553, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 553, + 448 + ], + "type": "text", + "content": "train a series of LDMs on CelebAHQ-2.5k and FFHQ-6k, with their subsets CelebAHQ-250 and FFHQ-600 duplicated 5, 10, 15 and 20 times. Figure 7 depicts the memorization scores of duplicated images. As can be observed, data duplication rapidly worsens memorization in DMs and our measure is sensitive to these changes. With increasing duplication times, the larger FFHQ-6k dataset presents a stronger resistance to memorization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "content": "For CIFAR-10 and LAION which originally contain duplicates, we observe the effect of dedduplication. We deduplicate CIFAR-10 with SSCD embeddings and DBSCAN clustering [14]. This results in 666 identified clusters, only one image of which is retained for training. The memorization is evaluated on the retained 666 images before and after dedduplication. On LAION, as compared to SD v1.4, SD v2.1 and v3.5 deduplicate the training set. Memorization is evaluated on the confirmed and suspicious sets. Figure 8 shows the results, with infinite scores on the top. As can be observed, InvMM clearly exhibits the benefits that dedduplication brings." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 600, + 382, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 382, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 382, + 611 + ], + "type": "text", + "content": "5.4. Detection" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "For InvMM to be a correct measure of memorization, it should be capable of distinguishing highly memorized images from normal ones. To construct a reliable test set, we conduct a nearest neighbor test following previous studies [9, 51, 59] to discover highly memorized images from DM's training set, i.e., highly memorized images consist of training images that can appear at sampling time. Specifically, we consider DDPM trained on CIFAR-10, LDMs" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16965" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 74, + 175, + 155 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 175, + 155 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 175, + 155 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 175, + 155 + ], + "type": "image", + "image_path": "ecf7f3b0f8c5a99cfe6a7764d4f1cbc14195af6ccea105c77a61f434163181fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 169, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 169, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 295, + 192 + ], + "type": "text", + "content": "Figure 10. Performance comparison of membership inference on FFHQ-6k, across different training epochs and duplication times." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 177, + 74, + 278, + 155 + ], + "blocks": [ + { + "bbox": [ + 177, + 74, + 278, + 155 + ], + "lines": [ + { + "bbox": [ + 177, + 74, + 278, + 155 + ], + "spans": [ + { + "bbox": [ + 177, + 74, + 278, + 155 + ], + "type": "image", + "image_path": "10e9dcea9605bbf1ebf04c48da64361236540c32a9fd50cfa70310663eccc841.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 207, + 295, + 351 + ], + "type": "text", + "content": "trained on CelebAHQ and FFHQ, and SDs trained on LAION. With their official implementations, only DDPM and SD v1.4 present obvious replications, which are thus used for evaluation. For DDPM, the 99 discovered training images appearing in 1M random samples are used as positive samples and another 1k different training images are included as negative samples; for SD v1.4, in accordance with the setting of previous works, we deem images in the confirmed set as positive and those in the suspicious and normal set as negative, considering that captions in the confirmed set can cause replication of their paired images while those in the suspicious and normal set cannot." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "content": "Table 1 presents the detection performance assessed via AUC and TPR@1%FPR. For the unconditional DDPM on CIFAR-10, training loss metrics (the first 2 rows) show decent results and the original " + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "content": "-loss works much better than the " + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "content": "-loss. The differential metric " + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "inline_equation", + "content": "M^{\\mathrm{LOO}}" + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "content": " is less effective than the " + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "inline_equation", + "content": "\\epsilon_0" + }, + { + "bbox": [ + 54, + 353, + 295, + 484 + ], + "type": "text", + "content": "-loss and InvMM. The results indicate that the compared metrics are not able to quantify memorization, as they are not fundamentally commensurable between samples. InvMM measures memorization of each sample by contrasting its sensitive noise distribution to a common standard Gaussian, thus having the same scale." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 295, + 713 + ], + "type": "text", + "content": "For SD v1.4, we compare InvMM against training loss metrics and two state-of-the-art detection metrics [42, 61]. Since all compared approaches require access to image captions, we implement InvMM with fixed prompts during inversion. As shown in Tab. 1, our measure achieves performance comparable to state-of-the-art metrics. The results of compared metrics seem to demonstrate a strong ability to distinguish memorization thus far. However, we argue that the prompt-based task formulation employed in existing studies offers an overly optimistic understanding of memorization. It is because these techniques are actually to predict whether the captions will replicate their paired images, while overlook potential prompts that might cause replication. As observed in Fig. 9a, when simultaneously inverting the input prompt, our measure reveals that some images in the suspicious set are also memorized to a similar extent as those in the confirmed set (due to duplication). Applying GCG [64] attack on images with the lowest InvMM scores in the suspicious set provides additional evidence, as illus" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "trated in Fig. 9b: images previously deemed “not memorized” according to their training captions can indeed be replicated by potential adversarial prompts (see Appendix Fig. 13 for more examples). Hence, going beyond the confines of existing research, our measure describes worst-case memorization from an adversarial standpoint, making it more reliable in security-critical scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 163, + 468, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 468, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 468, + 176 + ], + "type": "text", + "content": "5.5. Comparison to membership" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 287 + ], + "type": "text", + "content": "A close topic to memorization is membership inference [8, 13, 49], which reveals the leakage risk of membership information that whether an image is used for training. The common implicit assumption is that member samples are \"memorized\" more than hold-out samples [13], rendering them distinguishable through statistics like (variants of) training loss or differential loss [8]. Interestingly, the principle of InvMM can provide an explicit clarification of the difference between membership and memorization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 552 + ], + "type": "text", + "content": "A comparison to a recent membership inference metric SecMI [13] on FFHQ-6k, as shown in Fig. 10, presents a positive correlation between them. As factors that exacerbate memorization changes, both SecMI and InvMM show increasing performance. There is also a performance gap between SecMI and InvMM, indicating that certain samples from the hold-out set are memorized more than some member samples. This mainly results from the gap between model's metric (training loss) and human's metric (perceptual similarity). Membership inference leverages the model's metric for prediction, while memorization should resort to perceptual metric, e.g., simulated by SSCD. What level of loss leads to image replication depends both on the model and image itself. Low training loss does not necessarily lead to data replication, but would probably distinguish training data from never accessed. On the contrary, high training loss signifying non-membership may still be enough for replication. Therefore, membership does not deterministically lead to high memorization and vice versa. InvMM differs from membership metrics by focusing on the perceptual level and adaptively handles the specific loss landscapes of different samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this work, we dive into measuring memorization for diffusion models. Comprehensive experiments demonstrate the ability of the proposed measure to correctly quantify memorization. In short, an image to a higher extent of memorization is able to be inverted to a more standard and diverse latent distribution. However, our method is also limited by its simple modeling of the sensitive latent noise as a multivariate normal. During our research, we observed many clues confirming that the true distribution is complex. Future work might explore more advanced methods, e.g., a generative network, to fit the sensitive noise." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16966" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 236 + ], + "type": "text", + "content": "This work was partly supported by the National Key Research and Development Program of China under No. 2022YFB3102100, NSFC under No. U244120033, U24A20336, 62172243, 62402425 and 62402418, the China Postdoctoral Science Foundation under No. 2024M762829, the Zhejiang Provincial Natural Science Foundation under No. LD24F020002, the “Pioneer and Leading Goose” R&D Program of Zhejiang under No. 2025C01082, 2025C02033, 2025C02263 and 2025C02034, the Zhejiang Provincial Priority-Funded Postdoctoral Research Project under No. ZJ2024001 and the Key R&D Program of Ningbo under No. 2024Z115." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 115, + 258 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 266, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 266, + 275, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 266, + 275, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 266, + 275, + 277 + ], + "type": "text", + "content": "[1] Midjourney. https://www.midjourney.com. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 295, + 332 + ], + "type": "text", + "content": "[2] Artist finds private medical record photos in popular AI training data set. https://arstechnica.com/information-technology/2022/09/artist-finds-private-medical-record-photos-in-popular-ai-training-data-set/.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 334, + 278, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 334, + 278, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 334, + 278, + 344 + ], + "type": "text", + "content": "[3] DALL·E 3. https://openai.com/dall-e-3.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 388 + ], + "type": "text", + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 295, + 423 + ], + "type": "text", + "content": "[5] Sanjeev Arora and Yi Zhang. Do gans actually learn the distribution? an empirical study. arXiv preprint arXiv:1706.08224, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 295, + 479 + ], + "type": "text", + "content": "[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 611 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 295, + 658 + ], + "type": "text", + "content": "[8] Nicholas Carlini, Steve Chien, Milad Nasr, Shuang Song, Andreas Terzis, and Florian Tramer. Membership inference attacks from first principles. In 2022 IEEE Symposium on Security and Privacy (SP), pages 1897-1914. IEEE, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "type": "text", + "content": "[9] Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5253-5270, 2023. 1, 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[10] Nicholas Carlini, Daphne Ippolito, Matthew Jagielski, Katherine Lee, Florian Tramer, and Chiyuan Zhang. Quantifying memorization across neural language models. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 184 + ], + "type": "text", + "content": "[11] George Cazenavette, Avneesh Sud, Thomas Leung, and Ben Usman. Fakeinversion: Learning to detect images from unseen text-to-image models by inverting stable diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10759-10769, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 185, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 228 + ], + "type": "text", + "content": "[12] Chen Chen, Daochang Liu, and Chang Xu. Towards memorization-free diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8425-8434, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 230, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 554, + 274 + ], + "type": "text", + "content": "[13] Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? In International Conference on Machine Learning, pages 8717-8730. PMLR, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 275, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 275, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 554, + 317 + ], + "type": "text", + "content": "[14] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 319, + 554, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 319, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 554, + 373 + ], + "type": "text", + "content": "[15] Qianli Feng, Chenqi Guo, Fabian Benitez-Quiroz, and Aleix M Martinez. When do gans replicate? on the choice of dataset size. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6701-6710, 2021. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 376, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 554, + 430 + ], + "type": "text", + "content": "[16] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 432, + 554, + 475 + ], + "type": "text", + "content": "[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "type": "text", + "content": "[18] Andres Guadamuz. Photographer sues LAION for copyright infringement. https://www.technollama.co.uk/ photographer-sues-laion-for-copyright-infringement, 2023.1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 522, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 554, + 555 + ], + "type": "text", + "content": "[19] Ishaan Gulrajani, Colin Raffel, and Luke Metz. Towards gan benchmarks which require generalization. In International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 555, + 554, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 588 + ], + "type": "text", + "content": "[20] Chuan Guo, Alexandre Sablayrolles, Hervé Jégou, and Douwe Kiela. Gradient-based adversarial attacks against text transformers. arXiv preprint arXiv:2104.13733, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 612 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 645 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 678 + ], + "type": "text", + "content": "[23] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[24] Yue Jiang, Haokun Lin, Yang Bai, Bo Peng, Zhili Liu, Yueming Lyu, Yong Yang, Xingzheng, and Jing Dong. Image-level memorization detection via inversion-based inference" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16967" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "perturbation. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 97, + 295, + 140 + ], + "type": "text", + "content": "[25] James Jordon, Jinsung Yoon, and Mihaela Van Der Schaar. Pate-gan: Generating synthetic data with differential privacy guarantees. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 295, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 295, + 186 + ], + "type": "text", + "content": "[26] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 295, + 232 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "text", + "content": "[28] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "text", + "content": "[29] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 369 + ], + "type": "text", + "content": "[30] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In 2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 370, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 370, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 370, + 294, + 392 + ], + "type": "text", + "content": "[31] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "type": "text", + "content": "[32] Elizaveta Levina and Peter Bickel. Maximum likelihood estimation of intrinsic dimension. Advances in neural information processing systems, 17, 2004. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 294, + 472 + ], + "type": "text", + "content": "[33] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 496 + ], + "type": "text", + "content": "[34] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "text", + "content": "[35] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2837-2845, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "type": "text", + "content": "[36] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "text", + "content": "[37] Helen Nissenbaum. Privacy as contextual integrity. Wash. L. Rev., 79:119, 2004. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 688 + ], + "type": "text", + "content": "[38] Kai Packhäuser, Lukas Folle, Florian Thamm, and Andreas Maier. Generation of anonymous chest radiographs using latent diffusion models for training thoracic abnormality classification systems. In 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI), pages 1-5. IEEE, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "image copy detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14532-14542, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 215 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[42] Jie Ren, Yaxin Li, Shenglai Zeng, Han Xu, Lingjuan Lyu, Yue Xing, and Jiliang Tang. Unveiling and mitigating memorization in text-to-image diffusion models through cross attention. In European Conference on Computer Vision, pages 340-356. Springer, 2024. 2, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 327 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 392 + ], + "type": "text", + "content": "[44] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[45] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 458 + ], + "type": "text", + "content": "[46] Pamela Samuelson. Generative ai meets copyright. Science, 381(6654):158-161, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 460, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 535 + ], + "type": "text", + "content": "[47] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 590 + ], + "type": "text", + "content": "[48] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[49] Reza Shokri, Marco Stronati, Congzheng Song, and Vitaly Shmatikov. Membership inference attacks against machine learning models. In 2017 IEEE symposium on security and privacy (SP), pages 3-18. IEEE, 2017. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[51] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery?" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16968" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 689 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 2, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[52] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Understanding and mitigating copying in diffusion models. Advances in Neural Information Processing Systems, 36:47783-47803, 2023. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 3, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 218 + ], + "type": "text", + "content": "[54] N Vaishnavh, C Raffel, and IJ Goodfellow. Theoretical insights into memorization in gans. In Neural Information Processing Systems Workshop, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 261 + ], + "type": "text", + "content": "[55] Gerrit van den Burg and Chris Williams. On memorization in probabilistic deep generative models. Advances in Neural Information Processing Systems, 34:27916-27928, 2021. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[56] Nikhil Vyas, Sham M Kakade, and Boaz Barak. On provable copyright protection for generative models. In International Conference on Machine Learning, pages 35277-35299. PMLR, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 295, + 364 + ], + "type": "text", + "content": "[57] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for diffusion-generated image detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445-22455, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 387 + ], + "type": "text", + "content": "[58] Ryan Webster. A reproducible extraction of training images from diffusion models. CoRR, abs/2305.08694, 2023. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[59] Ryan Webster, Julien Rabin, Loic Simon, and Frédéric Jurie. Detecting overfitting of deep generative networks via latent recovery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11273-11282, 2019. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 295, + 476 + ], + "type": "text", + "content": "[60] Ryan Webster, Julien Rabin, Loic Simon, and Frederic Jurie. On the de-duplication of LAION-2B. arXiv preprint arXiv:2303.12733, 2023. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "text", + "content": "[61] Yuxin Wen, Yuchen Liu, Chen Chen, and Lingjuan Lyu. Detecting, explaining, and mitigating memorization in diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 588 + ], + "type": "text", + "content": "[62] Xiaoyu Wu, Yang Hua, Chumeng Liang, Jiaru Zhang, Hao Wang, Tao Song, and Haibing Guan. Cgi-dm: Digital copyright authentication for diffusion models via contrasting gradient inversion. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10812-10821. IEEE Computer Society, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 644 + ], + "type": "text", + "content": "[63] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Inversion-based style transfer with diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10146-10156, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 689 + ], + "type": "text", + "content": "[64] Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J Zico Kolter, and Matt Fredrikson. Universal and transferable adversarial attacks on aligned language models. arXiv preprint arXiv:2307.15043, 2023. 8, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16969" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_content_list.json b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b91a617fb7592d226e40a5660a89865c39ef1d0f --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_content_list.json @@ -0,0 +1,3281 @@ +[ + { + "type": "text", + "text": "An OpenMind for 3D medical vision self-supervised learning", + "text_level": 1, + "bbox": [ + 192, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tassilo Wald $^{*,1,2,3}$ , Constantin Ulrich $^{*,1,4}$ , Jonathan Suprijadi $^{*,†}$ , Sebastian Ziegler $^{1,2}$ , Michal Nohel $^{\\dagger,6,7}$ , Robin Peretzke $^{1,4}$ , Gregor Kohler $^{\\dagger,1}$ , Klaus Maier-Hein $^{1,2,3,4,5,8}$", + "bbox": [ + 235, + 179, + 763, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany", + "bbox": [ + 184, + 234, + 810, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Helmholtz Imaging, DKFZ; $^{3}$ Faculty of Mathematics and Computer Science, Heidelberg University", + "bbox": [ + 192, + 253, + 800, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Medical Faculty, Heidelberg University; 5 National Center for Tumor Diseases (NCT), Heidelberg", + "bbox": [ + 202, + 271, + 792, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{6}$ Faculty of Electrical Engineering and Communication, Brno University of Technology, Czech Republic", + "bbox": [ + 187, + 289, + 808, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "7 University hospital Ostrava, Department of Deputy director for science, research and education", + "bbox": [ + 212, + 306, + 784, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{8}$ Pattern Analysis and Learning Group, Department of Radiation Oncology tassilo.wald@dkfz-heidelberg.de", + "bbox": [ + 274, + 323, + 722, + 356 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 391, + 326, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of self-supervised learning (SSL) for 3D medical images lacks consistency and standardization. While many methods have been developed, it is impossible to identify the current state-of-the-art, due to i) varying and small pretraining datasets, ii) varying architectures, and iii) being evaluated on differing downstream datasets. In this paper, we bring clarity to this field and lay the foundation for further method advancements through three key contributions: We a) publish the largest publicly available pre-training dataset comprising $114k$ 3D brain MRI volumes, enabling all practitioners to pre-train on a large-scale dataset. We b) benchmark existing 3D self-supervised learning methods on this dataset for a state-of-the-art CNN and Transformer architecture, clarifying the state of 3D SSL pre-training. Among many findings, we show that pre-trained methods can exceed a strong from-scratch nnU-Net ResEnc-L baseline. Lastly, we c) publish the code of our pre-training and fine-tuning frameworks and provide the pre-trained models created during the benchmarking process to facilitate rapid adoption and reproduction. Available here.", + "bbox": [ + 88, + 422, + 485, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 753, + 220, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-Supervised Learning (SSL) has emerged as a transformative approach in deep learning, enabling the extraction of robust and general representations from data-rich domains [1-3]. This paradigm has led to significant advancements in areas with abundant data, such as natural language process", + "bbox": [ + 89, + 777, + 483, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and natural image computer vision. In the domain of 3D medical imaging, where annotation is expensive and annotated data scarce, SSL pre-training holds great promise as it can learn transferable representations without requiring extensive manual labeling efforts. However, currently the field of medical image analysis follows two predominant approaches to develop vision models: training models from scratch, often using the nnU-Net framework [4], or adopting supervised pre-training with large, monolithic annotated datasets [5-7] or aggregates of many datasets to increase supervised data scale [8, 9]. While the latter approach indicates a willingness to adopt pre-training paradigms, it raises the question why self-supervised learning is hardly pursued in 3D medical imaging. We believe this lack of SSL method development and adoption can be attributed to two major factors:", + "bbox": [ + 511, + 393, + 906, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1) A lack of a large, open-source 3D dataset for pretraining. Acquiring large unlabeled datasets for self-supervised pre-training is a non-trivial task. While large studies like the UK-BioBank or the NIH Adolescent Brain Cognitive Development (ABCD) exist and hold $>100\\mathrm{k}$ or $>40\\mathrm{k}$ 3D Volumes respectively, access to them depends on a positive internal review process requiring an appropriate project proposal. To add to the complexity, datasets are often subject to unique Data Usage Agreements (DUAs) that impose various restrictions on recipients. These rules may include: (i) mandatory acknowledgments, (ii) internal administrative review of manuscripts prior to submission [10-14], (iii) inclusion of a consortium in the author list [12-14], or (iv) requiring the dataset's name to appear in the manuscript title [13]. While acknowledgements are reasonable and exemptions for the remaining restrictions can be requested, these DUAs create legal barriers that hinder re", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, Author order among the co-first authors may be adjusted for individual use.", + "bbox": [ + 89, + 862, + 482, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done while at DKFZ.", + "bbox": [ + 109, + 887, + 261, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23839", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfc5f8ea733febaa604803f077e12569109f4f2e56dcd1acd0f52e22ed76f4e0.jpg", + "image_caption": [ + "Figure 1. The OpenMind dataset contains 114k 3D Head-And-Neck volumes of 23 different modalities. It represents the largest openly accessible dataset of 3D medical images currently available." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 473, + 310 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "production or creation of large-scale pre-training datasets. Consequently, the majority of current SSL methods are either developed on large, restricted or proprietary datasets [15-17] or are developed on small-scale publicly available datasets [18-24], leading to the current fragmented landscape of pre-training datasets.", + "bbox": [ + 89, + 426, + 483, + 518 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) A lack of comparability: To adopt existing pre-trained models or conduct one's own SSL pre-training, it is important to know which model or pre-training method is the best. Currently, this is an impossible task, due to: i) Different pre-training dataset selection (as previously highlighted). ii) Different architecture choices. Some methods use CNNs for pre-training [15, 17, 21], some ViT-like Transformers [18, 19], others Swin Transformers [22, 23] and even others hybrid architectures [16, 24]. iii) Insufficient downstream dataset choices, potentially providing unreliable results.", + "bbox": [ + 89, + 527, + 483, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we address these issues and provide three key contributions: i) We share the largest publicly available 3D dataset, comprising 114k 3D volumes of 23 different brain MRI modalities. ii) We conduct a comprehensive self-supervised learning benchmark leveraging this dataset to assess the performance of state-of-the-art CNN and transformer-based architectures on 15 segmentation and classification downstream datasets. iii) We open-source our pre-training and fine-tuning frameworks as well as the trained model checkpoints to enable rapid reproduction and novel method development in the field of 3D medical image segmentation.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/bc4f0442f7a72fad7b01d567a142e4cc7bdd430b7f43be27e9f3a4d6719cbf4f.jpg", + "table_caption": [ + "Table 1. Large medical datasets are often restricted through custom DUAs limiting their overall applicability." + ], + "table_footnote": [], + "table_body": "
DatasetPatients3D VolumesModalitiesAccess
ABCD [25]11,38541,5802Restricted
PPMI [27]2,0119,5545Restricted
ADNI [28]1,5639,5483Restricted
OASIS31,37615,8727Restricted
OASIS46614,0215Restricted
BraTS21 [29]1,2515,0044CC-BY-4.0
CT-RATE [25]21,30450,1882CC-BY-NC-4.0
OpenMind34,191113,92124CC-BY-4.0
", + "bbox": [ + 532, + 128, + 885, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. The OpenMind Dataset", + "text_level": 1, + "bbox": [ + 511, + 263, + 736, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To provide practitioners access to a large-scale dataset for self-supervised learning, we introduce the OpenMind dataset: a large-scale, easily accessible resource containing 114k Head-and-Neck MRI volumes across various modalities, sourced and curated from the OpenNeuro platform. As the largest freely accessible 3D brain imaging dataset to date, OpenMind surpasses the NIH Adolescent Brain Cognitive Decline initiative [25] by a factor of three, and exceeds the recent CT-Rate dataset [26] by a factor of two, while being released under a permissive CC-BY license (see Tab. 1).", + "bbox": [ + 511, + 287, + 906, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dataset creation", + "text_level": 1, + "bbox": [ + 511, + 464, + 674, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All data was sourced or derived from datasets available on the OpenNeuro platform [30]. OpenNeuro (formerly known as openfMRI) is a publicly accessible data repository designed to support neurological research, adhering to the FAIR principles (Findable, Accessible, Interoperable, Reusable). The platform currently hosts over 1,200 public datasets, encompassing data from more than 51,300 subjects. Due to the diverse nature of its collection, OpenNeuro includes a wide array of modalities, such as MRI, PET, fMRI, MEG, EEG, and iEEG, from studies involving both healthy and diseased subjects, enabling research across various aspects of neurology. From this extensive collection, all available 3D MRIs, such as T1-weighted and T2-weighted scans, as well as 4D diffusion-weighted MRIs were collected from 800 different studies (citations [55-854]). This aggregation resulted in a total of 71k 3D MRI scans1 and 15k 4D diffusion-weighted images from a total of 34,139 subjects. While existing 3D SSL methods could be directly trained on the 71k 3D MRI scans spanning multiple MRI sequences, the 4D diffusion-weighted images required pre-processing to create 3D derivatives which we describe in the following section. Overall, due to the high amount of independent studies, the OpenMind dataset provides great levels of diversity regarding participant age distribution, countries of origin, MRI modalities, scanner manufacturer, model and scanning protocols.", + "bbox": [ + 511, + 486, + 906, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "This includes a minority of 653 PET scans.", + "bbox": [ + 531, + 886, + 767, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23840", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DWI Preprocessing Diffusion-weighted imaging (DWI) is an advanced MRI modality that captures the random motion of water molecules in tissue by applying diffusion-sensitive gradient pulses, enabling the assessment of microstructural properties. It is widely used for detecting abnormalities such as strokes, tumors, and white matter changes. In clinical and research settings, DWI data is often preprocessed into more interpretable 3D derivatives, including mean diffusion (MD) maps, and fractional anisotropy (FA) maps, which simplify analysis and interpretation. Additionally, T2-weighted images can be derived from DWI scans. Consequently, we pre-processed the DWI images into MD maps, FA maps, and T2-weighted images, resulting in an additional $43\\mathrm{k}$ 3D images. The detailed preprocessing pipeline is described in the appendix in B.2.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Defacing Protecting patient privacy is a big concern in the medical domain, hence, the facial region of Head-and-Neck scans are commonly anonymized. Common anonymization techniques include defacing – replacing the volume with a constant value –, blurring everything but the brain or brain extraction, see Figure 2 in the appendix. These augmented regions can interfere with reconstruction-based SSL methods, potentially penalizing models for attempting to reconstruct plausible anatomical features in anonymized regions. Therefore, masks delineating these regions are needed to allow excluding them from the loss computation. Unfortunately, not all datasets provide these masks, despite applying some sort of anonymization.", + "bbox": [ + 89, + 337, + 480, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we generate associated anonymization masks and anatomy masks for cases where these are not provided (Figure 2 bottom in the appendix). The anonymization mask identifies areas where anatomical structures have been artificially modified or removed and where no loss should be calculated, whereas anatomy masks indicate where relevant foreground is present, allowing to avoid sampling non-empty patches which may speed up the convergence of the pre-training methods. Both of these masks are created using the model proposed by Nohel et al. [31].", + "bbox": [ + 89, + 534, + 480, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta-Data enrichment While OpenNeuro follows the Brain Imaging Data Structure (BIDS) format, metadata organization varies significantly across datasets. Participant demographics (e.g., age, sex, health status) and image acquisition details (e.g., repetition times, scanner types, imaging protocols) are inconsistently reported, creating challenges for systematic analysis. We curated and harmonized all available, disparate metadata into a unified, complementary metadata structure. This standardized framework ensures consistency and enables users to efficiently filter the OpenMind dataset based on specific selection criteria, such as subject characteristics or imaging", + "bbox": [ + 89, + 719, + 480, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "parameters, enhancing the flexibility of the dataset and simplifying the extraction of relevant sub-populations.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to the harmonized metadata, we provide an Image Quality Score (IQS) for each imaging modality of each of the 800 datasets. This score represents a subjective usability score of the images for self-supervised pretraining. The assessment considers factors such as image noise levels, motion artifacts, and other image quality-degrading elements, condensing these evaluations into a unified score ranging from 1 to 5 with 1 representing the highest quality images. The quality scores were determined through manual inspection of multiple images from each modality in each dataset and were assessed independently by two raters to ensure reliability (more details in the appendix B.1). Together, the harmonized metadata and Image Quality Score are introduced to test the effects of simple data-filtering techniques on self-supervised methods trained on the OpenMind dataset. If significant effects of pre-training data choice can be measured, the OpenMind datasets would show to have sufficient size to enable data-centric research in the domain of self-supervised learning for 3D medical image analysis, akin to [2, 32, 33].", + "bbox": [ + 511, + 122, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Useability To maximize the utility of the OpenMind dataset, we provide the final dataset of 114k 3D images on Hugging Face together with anonymization and anatomy masks, as well as the harmonized metadata including the Image Quality Score.", + "bbox": [ + 511, + 438, + 903, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Why benchmarking matters While the OpenMind dataset addresses the issue of data accessibility and data scale in self-supervised learning for 3D medical imaging, the current state-of-the-art remains unknown. To establish a reference point and provide an initial set of baselines, we conduct a comprehensive benchmarking effort to compare existing 3D SSL methods within a common framework.", + "bbox": [ + 511, + 529, + 903, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The OpenMind Benchmark", + "text_level": 1, + "bbox": [ + 513, + 648, + 769, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The field of self-supervised learning (SSL) for 3D medical imaging remains fragmented due to inconsistencies in pre-training datasets, architectures, and downstream evaluations. Without standardized benchmarks, it is impossible to fairly compare methods and determine the most effective approaches [15]. The OpenMind Benchmark addresses this challenge by evaluating a broad range of SSL methods under controlled conditions, using a common pre-training dataset, standardized architectures, and diverse downstream tasks, which we detail in this section.", + "bbox": [ + 511, + 672, + 903, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Standardization", + "text_level": 1, + "bbox": [ + 513, + 833, + 671, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To establish a standardized and fair evaluation, all self-supervised learning (SSL) methods trained in the benchmark were trained on the OpenMind dataset, ensur", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23841", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing that pre-training dataset variability does not influence results. Further, we use two state-of-the-art architectures, the ResEnc-L Residual Encoder U-Net [34] to represent convolutional neural networks (CNNs) and the recent Primus-M Transformer [35], allowing us to analyze SSL effectiveness across different architectural paradigms. Pre-trained models are then evaluated on a diverse set of downstream datasets for segmentation and classification tasks.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Downstream Datasets All downstream datasets used are split 50/50 into a train and test set. Moreover, we group our datasets into development datasets and test datasets. The development datasets serve to optimize SSL method hyperparameters on their train split, should the method not be directly applicable on the OpenMind dataset with original hyperparameters. During this process, the most important method-specific hyperparameters as well as learning rate and weight decay are optimized based on an additional 80/20 split on the training set of the development datasets. Aside from this, the development datasets are used to determine the optimal fine-tuning schedules, hence results of the optimal fine-tuning schedules are reported on the validation split of the dataset (Tab. 4 and in the appendix in table 8).", + "bbox": [ + 91, + 242, + 483, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To avoid potential biases introduced during the optimization process on the development datasets, we withheld additional test datasets, which we left untouched and only evaluated after finishing all method optimizations on the development datasets. Moreover, since datasets tend to be small and targets diverse, results are inherently noisy and may not generalize from one dataset to another. To reduce the chance of over optimizing for one dataset, we include four development datasets as well as eight test segmentation datasets and three test classification datasets.", + "bbox": [ + 89, + 469, + 482, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Development datasets:", + "text_level": 1, + "bbox": [ + 89, + 657, + 251, + 672 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. Atlas R2.0 (ATL) [36]: Brain scans of stroke patients imaged through T1 weighted MRIs.", + "2. Stanford Brain Metastasis (SBM) [37]: Brain scans of metastasis imaged through pre-contrast T1, contrast-enhanced gradient echo, contrast-enhanced spin echo and a T2 FLAIR sequence.", + "3. AMOS22 (AMO) [38]: CT and MRI scans of 15 abdominal organs.", + "4. KiTS19 (KIT) [39]: CT images of kidneys and kidney tumors." + ], + "bbox": [ + 91, + 674, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While AMOS and KiTS can be considered out-of-distribution due to the pre-training images exclusively depicting the head-and-neck region, these datasets were found to be robust benchmarking datasets in [34], hence were included despite their shift. Further, during develop", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ment, we only optimized the methods for segmentation, due to it being the most common task in 3D medical image analysis [35].", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test datasets (Segmentation):", + "text_level": 1, + "bbox": [ + 513, + 151, + 720, + 165 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. ISLES (ISL) [40]: DWI $(\\mathrm{b} = 1000)$ and ADC map brain MRI scans of patients with delineations of stroke lesions.", + "2. HNTS-MRG (HNT) [41]: MR images of patients with head-and-neck primary tumors and metastatic lymphnodes pre and mid-treatment of which we use the pretreatment images and labels.", + "3. HaN-Seg OAR (HAN) [42]: CT and MR images of Head-and-Neck organs-at-risk of which we use only the MR images.", + "4. MS FLAIR (MSF) [43]: FLAIR MRIs of hyperintense multiple-sclerosis lesions with associated delineations.", + "5. ToP-CoW (TPC) [44]: Time-of-Flight (ToF) Magnetic Resonance Angiography (MRA) of the Circle of Willis and associated annotations of the arterial structure.", + "6. Yale Brain Metastasis (YBM) [45]: Pre and post-contrast T1, T2 and FLAIR MRI images of brain metastases.", + "7. Cosmos (COS) [46]: Carotid vessel wall segmentations and atherosclerosis diagnosis, with contours of the vessel walls on 3D-VISTA MRI.", + "8. ACDC (ACD) [47]: Cardiac cine-MRI image of three segmented ventricular structures." + ], + "bbox": [ + 514, + 167, + 903, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Test datasets (Classification):", + "text_level": 1, + "bbox": [ + 514, + 484, + 718, + 500 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1. MR-Net (MRN) [48]: Knee T1w, T2w and Proton Density (PD) weighted MRIs of ACL or meniscal tears as well as other abnormalities which we classify.", + "2. RSNA-SpineFrac (RSN) [49]: CT Spine fractures with the goal of identifying lumbar spine degenerative conditions. We classify fracture and no-fracture.", + "3. ABIDE (ABI) [50]: T1w MRI images of either healthy controls or subjects with autism, on which we conduct binary classification - autism or not-autism." + ], + "bbox": [ + 514, + 501, + 903, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training configuration To ensure a fair comparison, all SSL methods are trained under comparable pre-training and equal fine-tuning schedules. Specifically, we pre-train all methods for a total of 1000 epochs with 250 steps per epoch each, with each method trained DDP on four 40GB A100s to decrease pre-training time. Furthermore, the pretrained models were fine-tuned for a total amount of 150 or 1000 epochs with 250 steps per epoch and a common batch size of 2 for both architectures following an identical learning rate schedule implemented within the nnU-Net framework [4] for the segmentation downstream datasets. For classification downstream datasets, all models were fine-tuned for 200 dataset epochs within the public Image classification framework, which supports 3D classification. We provide more detailed descriptions of the fine-tuning process and frameworks in the appendix C.", + "bbox": [ + 511, + 657, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23842", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Benchmark scope", + "text_level": 1, + "bbox": [ + 89, + 90, + 264, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the benchmark, three key aspects of SSL are investigated, highlighting the possible research directions the OpenMind dataset and benchmark can enable:", + "bbox": [ + 89, + 113, + 482, + 157 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i) SSL method performance: We compare the efficacy of various SSL methods in learning transferable representations to various downstream segmentation and classification datasets when using the same pretraining dataset, using two common architectures and a unified fine-tuning schedule.", + "ii) Fine-tuning methodology We evaluate various fine-tuning schedules for both architectures to investigate the importance of a well-configured downstream fine-tuning schedule.", + "iii) Data-centric baselines We evaluate the influence of data-filtering on the large OpenMind dataset, to evaluate if data-centric research can be conducted on this scale." + ], + "bbox": [ + 99, + 159, + 482, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SSL Methods Various SSL methods native to 3D medical images exist. In the benchmark, we include contrastive methods, like Volume Contrastive (VoCo) [22], a 3D SimCLR adaptation, the pseudo-supervised VolumeFusion (VF) [16], default masked autoencoders (MAE) [3], sparse adaptations of them for CNNs (S3D) [15], masked image modeling (SimMIM) [18] and, the denoising and masked auto-encoder combination ModelsGenesis (MG) [21]. Moreover, we incorporate the SwinUNETR pre-training scheme (SwinUNETR) [23], which integrates masked inpainting, contrastive learning and rotation prediction2. We tried including the 3D geometric-matching-based pre-training method GVSL [20], but the method failed to converge both when using the original implementation and when reimplementing it within our framework, regardless of whether we trained it on our OpenMind dataset or the ABCD-NIH dataset. We provide a brief description of each included SSL method and its used hyperparameters in the appendix D. Methods, which did not feature publicly available Code that could be tested for functionality or integrated into our repository, were excluded, e.g. GLMAE [19]. Moreover, more recent self-supervised methods were given higher priority for inclusion in our benchmark than older methods. Furthermore, prominent 2D methods were excluded, as adapting their configuration from the natural imaging domain to the 3D medical imaging domain would require substantial development effort, e.g. due to large batch size requirements leading to VRAM issues.", + "bbox": [ + 89, + 371, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning Schedules Effective adaptation during fine-tuning is crucial for optimizing performance in medical image segmentation, as it determines how well a pre-trained", + "bbox": [ + 89, + 816, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model can transfer knowledge to a target dataset. We test five different fine-tuning strategies for adaptation to segmentation datasets, aiming to balance maintaining general pre-trained representations with adapting the encoder to specialized representations through varying learning rate (lr) schedules: i) nnU-Net's default training strategy employs a polynomial decay lr with an initial lr of $1\\mathrm{e} - 2^3$ , which we reduce to 1e-3 for the transfer learning setting. ii) The Frozen schedule follows the same lr schedule but restricts training to the decoder. iii) The Warm-Up schedule introduces an initial phase with a linear increase in the lr before transitioning to the default schedule. iv) The Valley scheme prioritizes decoder adaptation by first training it with a linear decreasing lr, followed by a linear warm-up phase for the full network, and then the default schedule. v) Finally, the Sawtooth scheme employs a two-stage warm-up: first for the decoder while keeping the encoder frozen, and then for the whole network, both with a linear increasing lr, before transitioning to the default schedule.", + "bbox": [ + 511, + 90, + 903, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the majority of fine-tuning runs for 150 epochs to reduce computational cost. For these, each warm-up stage lasts 15 epochs. For the CNNs we provide additional fine-tuning experiments trained for 1000 epochs, in which the warm-up stage is extended to 50 epochs. For a more detailed description and visualization of this, we defer to the appendix C, where we also provide additional information on the classification fine-tuning.", + "bbox": [ + 511, + 378, + 903, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset Filters When training on a large enough quantity of data, optimizing the dataset becomes an important aspect that can greatly improve the resulting model performance [2, 32, 33]. Given the scale and diversity of the OpenMind dataset, we evaluate four simple filtering schemes to measure if we are at a scale where data filtering can be effective. The filtering schemes implemented are: 1. Filtering by the manually created Image Quality Score (IQS) at three thresholds, and 2. filtering all modalities but T1w, T2w and FLAIR images, the most common downstream dataset modalities.", + "bbox": [ + 511, + 537, + 906, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reproducibility While evaluating this benchmark, a lot of pre-trained models were generated. To maximize reproducibility, we provide the checkpoints of all pre-training methods for both architectures. Further, we provide an integration into the nnU-Net framework that holds the provided fine-tuning code, enabling users to quickly adapt the pretrained checkpoints to a desired downstream task.", + "bbox": [ + 511, + 744, + 903, + 849 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2We refer to the pre-training scheme as SwinUNETR, not to be confused with the identically named architecture.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Learning rate value of the ResEnc-L CNN only. Primus-M learning rate values are provided in the appendix C", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23843", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/af795f1b5fa3355c3156d3c284aa11461b36a5deb2719b895e97c4e81fadb5f4.jpg", + "table_caption": [ + "Table 2. Segmentation results of tested pre-training methods. CNN and Transformer architectures were pre-trained on the same dataset for the same amount of steps. All methods were fine-tuned with the optimal architecture specific fine-tuning scheme for 150 epochs, unless indicated by suffix $1k$ . Color ranges are set per-dataset and per-architecture. S3D and SimMIM are only shown for CNN or Transformer as they are architecture specific. nnU-Net def.: Dynamically planned default nnU-Net trained from scratch; 1k: Fine-tuned for 1000 epochs instead of 150 epochs; ID Mean: In distribution mean; OOD Mean: Mean across datasets of non Head-and-Neck datasets." + ], + "table_footnote": [], + "table_body": "
PT MethodDice Similarity Coefficient (DSC) [%] on ...
Dataset of same anatomical region (ID)Dataset of OOD regionAverage across ...
ATLSBMISLHNTHANMSFTPCYBMCOSACDAMOKITIDOODAll
nnU-Net def. 1k58.7059.9878.4062.9853.3752.1979.5058.4346.1991.1088.0087.2161.0888.7768.00
nnU-Net def.56.0860.4178.2259.3732.2755.8476.7856.7349.3190.7283.8877.6158.3384.0764.77
ResEnc-L (CNN)
Scratch 1k58.2153.4379.1465.7558.2454.9079.9456.1271.5792.0988.7387.4864.1589.4370.47
Scratch57.0254.2978.0963.3056.1155.4776.1854.4265.2091.9785.2484.0362.2387.0868.44
VoCo57.1459.6277.5063.4851.1254.9075.1256.9263.4991.4485.6085.7162.1487.5868.50
SwinUNETR56.0757.2577.4561.6449.4254.8274.6957.0565.6890.5384.9585.5461.5687.0167.92
SimCLR57.1559.7278.0163.3251.5655.6877.7759.1468.2091.7686.0684.8563.4087.5669.44
VF57.4259.8878.1864.3251.6757.4276.1159.3163.9891.5785.3886.2163.1487.7269.29
MG58.0361.5777.5865.1154.6955.2577.1458.6771.2791.7486.3586.1764.3788.0970.30
MAE58.2562.4177.8966.5855.1456.8477.9660.0770.8591.9886.7886.1265.1188.3070.91
S3D58.7664.0978.0565.7452.8156.0878.8159.1866.6692.0186.1686.0164.4688.0670.36
Primus-M (Transformer)
Scratch 1k56.7748.5076.5958.4053.4053.2776.3252.5364.6890.8987.2485.5760.0587.9067.01
Scratch51.5143.2675.2355.3050.6054.0073.3150.3062.1190.9380.1776.7357.2982.6163.62
VoCo46.8034.1573.2951.0647.6452.6465.5244.7552.1687.2665.8170.2152.0074.4357.61
SwinUNETR47.2336.3173.8450.1546.4952.8066.2344.4954.8287.9266.1770.3952.4974.8258.07
SimCLR54.6142.6275.4356.7550.8053.5970.0848.3658.2089.9775.7581.2756.7282.3363.12
VF58.6247.3777.5662.3756.1855.0074.9853.9669.7491.4184.9586.1761.7587.5168.19
MG56.5047.3476.7658.4254.0254.6773.6549.7760.7490.8982.1584.4559.1085.8365.78
MAE61.1656.6777.1266.1257.2456.0278.3154.3572.0292.1687.1686.7464.3488.6970.42
SimMIM60.2851.6877.5362.7656.7455.9177.0052.9070.8791.9886.5785.9262.8588.1669.18
", + "bbox": [ + 112, + 170, + 883, + 450 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b576ca0b63f887ac914bd98c6fe4ad5495fd32b5640a9ad6919d50f26bd0b2b.jpg", + "table_caption": [ + "Table 3. Classification results of tested pre-training methods. CNN and Transformer architectures were pre-trained and fine-tuned identically on the three classification datasets. PT Method: pre-training method" + ], + "table_footnote": [], + "table_body": "
Architecture Metric PT MethodResEnc-L (CNN)Primus-M (Transformer)
Average PrecisionMeanMRNBalanced AccuracyAverage PrecisionMeanMRNBalanced Accuracy
MRNRSNABIRSNABIMeanRSNABIMean
Scratch66.3365.3957.1862.9776.6162.2553.2964.0560.5858.9152.8757.4575.4957.1052.0561.55
VoCo71.9970.3362.2368.1880.3565.5760.1868.7067.9262.0953.2261.0877.6861.6451.2863.53
SwinUNETR70.7270.3762.3267.8079.2566.6059.8868.5864.9860.7956.8460.8776.6159.2255.6063.81
SimCLR68.7473.0259.1766.9878.2767.9657.6667.9667.6760.2655.5961.1876.6959.5254.1563.45
VF68.4666.2258.0164.2378.4363.2855.6065.7766.4161.6658.3762.1476.6760.1456.0664.29
MG68.8571.2159.6966.5979.0966.1457.2467.4968.8661.2452.4160.8477.7359.9754.1363.94
MAE64.5367.8658.3163.5776.7264.2156.2365.7267.0060.2853.0960.1377.2658.8351.8462.64
S3D70.0869.2560.6566.6678.5664.2056.4166.39--------
SimMIM--------65.2759.6253.9459.6177.1558.4953.9363.19
", + "bbox": [ + 114, + 501, + 883, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 89, + 651, + 176, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we highlight the most important insights of our benchmark, but not all due to spatial constraints. A comprehensive list of them is provided in the appendix A.", + "bbox": [ + 89, + 678, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. SSL method performance", + "text_level": 1, + "bbox": [ + 89, + 739, + 323, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation performance We compare the fine-tuning results (fine-tuned for 150 epochs) of all pre-training methods for the ResEnc-L CNN architecture and the Primus-M Transformer architecture in Tab. 2. For both architectures, we observe that self-supervised pre-training is able to exceed both, a from-scratch nnU-Net default baseline trained for 1000 epochs and their respective from-scratch trained architecture trained for 150 epochs. In particular, reconstruction based methods exhibit the best performance,", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with default MAEs leading the pack. The MAE pre-trained CNN model even exceeds its strong from scratch baseline trained for 1000 epochs in only 150 epochs, and performs even better when fine-tuned for 1000 epochs (see Table 5 in the appendix). Aside from the MAE pre-training paradigm, this is only achieved by the sparse-autoencoder based S3D pre-training scheme with 1000 epochs of fine-tuning.", + "bbox": [ + 511, + 652, + 906, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing the benefits for each architecture, it can be seen that the Transformer-based Primus-M architecture profits substantially more from pre-training than the ResEnc-L CNN counterpart, however, it shows lower from-scratch results to begin with. Comparing the MAE pre-trained Primus-M to its 1000 epochs from scratch counterpart, it is able to improve about $3.5\\%$ DSC points and almost exceeds the ResEnc-L 1000 epochs from scratch baseline in 150 epochs. However, on singular datasets the MAE", + "bbox": [ + 511, + 758, + 908, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23844", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ca50a42b2d5bacbedcb438fa7975ce23a7ae5440c3e4a2bd33ce7b0ff107708c.jpg", + "table_caption": [ + "Table 4. Effects of different CNN fine-tuning schedules." + ], + "table_footnote": [], + "table_body": "
DatasetPT Method FT ScheduleVoCoSwinUNETRSimCLRVFMGMAES3DAverage
SBMDefault73.8669.8672.7972.5773.3771.8475.4772.82
Frozen52.3438.5960.8328.5451.3559.4556.5549.67
Warm-Up70.8769.1673.9473.1174.6774.5173.5472.83
Valley70.2670.2972.8275.8174.6873.7172.4072.85
Sawtooth71.5271.0573.7273.7873.8074.7973.9773.23
ATLDefault62.7860.8256.8063.8362.6262.7862.1661.68
Frozen58.2754.2757.9543.3357.2754.4957.1554.68
Warm-Up59.3260.2362.3663.2063.4663.6162.3462.07
Valley59.4360.3461.2164.0061.2362.9863.3861.80
Sawtooth60.4261.1761.3465.0761.5463.4763.7662.40
AMODefault85.9385.4287.1286.0586.9386.7886.7486.42
Frozen53.1861.2662.9230.9757.8950.8352.4852.79
Warm-Up86.9886.6187.1086.5887.3487.6287.5487.11
Valley85.9185.4686.7286.5587.0987.3786.9286.57
Sawtooth86.7885.7686.7586.4687.1587.4886.4586.69
KITDefault85.8083.8381.2084.3483.6984.5682.2183.66
Frozen69.3369.5965.2236.0671.3963.9771.4163.85
Warm-Up84.7181.7781.8083.5183.9985.2482.2083.32
Valley84.0882.4181.5585.0083.4883.4883.7883.40
Sawtooth85.1283.2683.1085.4184.1584.2883.3184.09
AverageDefault77.0974.9874.4876.7076.6576.4976.6576.15
Frozen58.2855.9361.7334.7359.4857.1859.4055.25
Warm-Up75.4774.4476.3076.6077.3677.7576.4076.33
Valley74.9274.6275.5777.8476.6276.8876.6276.15
Sawtooth75.9675.3176.2377.6876.6677.5076.8776.60
", + "bbox": [ + 102, + 114, + 468, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-trained Primus-M transformer already achieves the highest performance overall (ATL, COS and ACD), including the best pre-trained ResEnc-L. Given the substantial performance discrepancy observed in Wald et al. [35] between the two architectures, this provides the first sign of vision transformers reaching state-of-the-art performance in 3D medical image segmentation.", + "bbox": [ + 89, + 354, + 482, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, not all pre-training methods are able to learn useful representations for segmentation. VoCo, SwinUNETR pre-training, and SimCLR (for transformers) do not, or just barely exceed, a from-scratch baseline trained for 150 epochs and can even lead to substantial performance reductions for the transformer architecture likely due to their global pre-training task not learning the local information needed for segmentation. Moreover, individual datasets can be negatively impacted by any pre-training, as for the ISL and HAN dataset when using a ResEnc-L architecture. This is particularly surprising as Primus-M shows strong improvements with its MAE pre-training on both of these datasets. When filtering low quality images this seems to be partially alleviated (see Table 6 in the appendix), indicating that CNNs may learn different patterns in the presence of noise, which a transformer may be more robust to.", + "bbox": [ + 89, + 460, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Lastly, comparing the longer 1000 epoch fine-tuning schedule ( see Table 5 in the appendix) and the shorter 150 epoch fine-tuning schedule of the ResEnc-L, it can be observed that the longer fine-tuning schedule can lead to performance degradation on datasets where pre-training showed strong improvements (SBM, HNT). Simultaneously, it reduces the performance gap to the from-scratch 1000 epoch baseline on cases where pre-training was not as helpful (ISL, HAN, AMO), likely by allowing to forget learned representations and adapt more strongly to the downstream task.", + "bbox": [ + 89, + 700, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Classification performance We evaluate the fine-tuned performance of our pre-trained methods for classification", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- a more global downstream task - in Tab. 3. Contrastive pre-training methods like VoCo, SwinUNETR, and SimCLR, which were ineffective for segmentation, achieve the highest classification performance. In contrast, MAE pretraining, which was the strongest for segmentation, performs the worst across both architectures and nearly all datasets. Furthermore, segmentation focused methods like S3D and MG do not transfer well to classification, though they do not fall to the lowest ranks. These findings reinforce the distinction between global and local feature learning: contrastive methods specialize in global similarity learning, making them well-suited for classification, whereas reconstruction-based objectives prioritize pixel-wise representations, benefiting segmentation. Interestingly, SwinUNETR pre-training, which incorporated both global and local objectives, still converged toward representations favoring classification over segmentation, suggesting a bias in learned feature hierarchies.", + "bbox": [ + 511, + 90, + 903, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, it is to note that from-scratch classification performance is subpar, with ABI barely exceeding randomness (balanced accuracy of $50\\%$ ). This may be due to the classification pipeline lacking the refinements of the widely adopted nnU-Net framework, due to the lower prevalence of global reasoning tasks in 3D medical imaging. Thus, while our classification results provide valuable insights, they should be interpreted with caution, particularly compared to our segmentation findings which are more reliable due to the robust benchmarking framework.", + "bbox": [ + 511, + 362, + 903, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Fine-tuning methodology", + "text_level": 1, + "bbox": [ + 511, + 529, + 743, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The ablation results of different CNN fine-tuning strategies in Tab. 4 indicate that the Sawtooth fine-tuning strategy consistently achieves the best overall performance, followed closely by the Warm-Up scheme. This shows that adapting the encoder, after tuning the newly initialized decoder is crucial to maximize performance. The staged learning rate increases in these schedules likely help to prevent instability and allow for a smoother transition from pre-training to task-specific adaptation. For the out-of-distribution CT datasets, AMO and KIT, the Default fine-tuning schedule performs very competitively. This indicates that models pretrained on MRI data require greater modifications when adapted to CT images. The high initial lr allows the models to shift their learned features more effectively to the new domain.", + "bbox": [ + 511, + 551, + 903, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Keeping the encoder frozen during fine-tuning significantly reduces performance across all datasets and underscores the importance of encoder adaptation. However, SimCLR is less affected by this constraint. Notably, both VoCo and SimCLR perform almost comparably to other fine-tuning schedules when the encoder is frozen on ATL, w.r.t. the drop on other datasets. This dataset is closest to the pretraining data in the development datasets, as it consists of a", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23845", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "single MRI sequence, whereas the other datasets either contain CT scans or multiple MRI sequences as input.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the Transformer, fine-tuning results in Table 8 in the appendix reveal that the Warm-Up scheme achieves the best performance across all methods, emphasizing the importance of architecture-specific adaptation. The Frozen schedule leads to an even greater performance drop than for CNNs, as the ResEnc-L decoder, despite being lightweight, has significantly more parameters than the \"decoder\" of Primus-M. In this setting, fine-tuning can be interpreted almost as linear probing, with only back-to-back transposed conv layers (including normalization and activations) being trained, severely limiting decoder capacity. With a frozen encoder, it can be observed that adapting to the ATL dataset reaches the closest performance to other fine-tuning schedules.", + "bbox": [ + 89, + 121, + 480, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Data-centric results", + "text_level": 1, + "bbox": [ + 89, + 340, + 277, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the effect of data filtering of the pre-training dataset for CNN MAE pre-training and highlight the measured downstream performance in Appendix Table 6. It can be observed that filtering the lowest quality images through the Image Quality Score (IQS) shows minor positive effects on overall downstream segmentation performance, despite removing substantial amounts of pre-training data. Only when removing about $66\\%$ , keeping only the highest quality images, a degradation in performance is measured. Moreover, it can be observed that reducing data diversity by removing all modalities but T1w, T2w and FLAIR images reduces model performance, despite having more pre-training images than the intermediate IQS filtering method which slightly increased performance. Overall, this indicates that the OpenMind dataset does contain redundant data which may be leveraged by more complex data-centric methods to increase performance further.", + "bbox": [ + 89, + 363, + 482, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of anonymization on reconstruction Many images of the OpenMind dataset were defaced or brain-extracted to guarantee patient privacy. To allow taking these artifacts into account when training with a reconstruction objective, we created masks that allow ignoring such regions. To measure the effect of such artifacts, we compare the performance of reconstruction methods that were or were not penalized for the reconstruction in such regions. (See Appendix Table 7). It can be observed that ignoring the reconstruction loss of anonymized regions improves the segmentation performance of the MAE and S3D, showing that taking anonymization into account improves representations.", + "bbox": [ + 89, + 619, + 482, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion, Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 89, + 814, + 444, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our study presents the OpenMind dataset, the largest publicly available 3D medical imaging dataset, alongside a standardized self-supervised learning (SSL) benchmark for 3D medical image analysis. Through extensive experi", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mentation across diverse datasets and across many pretraining paradigms, we demonstrate the superiority of self-supervised pre-training over from-scratch training for 3D medical image analysis, when pre-training on a large-scale dataset. We show that contrastive methods excel in global tasks such as classification, and reconstruction-based approaches perform well in dense tasks such as segmentation. However, we find that no pre-training method of the 3D vision domain generalizes and performs well in both tasks, emphasizing the need for further SSL method development. Moreover, we compared state-of-the-art CNNs and Transformers in our benchmark and show for the very first time that a pre-trained Primus-M Transformer can exceed the strong ResEnc-L CNN from-scratch baseline on many datasets and can exceed the strongest pre-trained ResEnc-L CNN on individual datasets. This marks an important milestone for the domain, as transformers from-scratch so far never exceeded a ResEnc-L model [34, 35, 51]. On average, though, we show that a pre-trained ResEnc-L remains the strongest architecture for now.", + "bbox": [ + 511, + 90, + 903, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, our ablation results suggest that fine-tuning strategies are a key determinant of the success of pre-trained methods, with bad strategies losing significant performance. This emphasizes the importance of careful adaptation beyond pre-training. Experiments on fine-tuning for longer showed partial degradation of overall performance, indicating the problem of overfitting to downstream tasks. Together, these insights open up research opportunities for developing and adapting Parameter-Efficient Fine-Tuning (PEFT) methods to 3D medical image analysis to address these issues. Moreover, our fine-tuning experiments showed that keeping the vision encoders frozen shows sub-par performance for all pre-training methods. This emphasizes the lack of generality of currently learned representations and – again – highlights the need for better pre-training methods. While our study provides a lot it also has various limitations: i) Our classification results are not as reliable as the segmentation results, due to the lower interest in classification for 3D medical imaging, requiring further refinements. ii) Current evidence of the utility of the OpenMind dataset for data-centric research is rather weak, due to the minimal improvements observed requiring further investigation. iii) We evaluate pre-training for 1000 epochs due to resource limitations. Pre-training for longer may reveal different trends than shown in this study.", + "bbox": [ + 511, + 393, + 903, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, by open-sourcing the OpenMind dataset and establishing the OpenMind benchmark, we provide a clear assessment of state-of-the-art SSL in 3D medical imaging and offer a reproducible framework for method development. Through them, we aim to accelerate progress in self-supervised learning for 3D medical images and pave the way for new research directions.", + "bbox": [ + 511, + 770, + 903, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23846", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partly funded by the Helmholtz Foundation Model Initiative (HFMI) under the subproject 'The Human Radiome Project' (THRP). Part of this work was funded by Helmholtz Imaging (HI), a platform of the Helmholtz Incubator on Information and Data Science. Additionally, a part of the computational resources were provided by the German AI Service Center WestAI.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023.", + "[2] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023.", + "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022.", + "[4] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021.", + "[5] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023.", + "[6] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024.", + "[7] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023.", + "[8] Constantin Ulrich, Fabian Isensee, Tassilo Wald, Maximilian Zenk, Michael Baumgartner, and Klaus H Maier-Hein. Multilant: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 648-658. Springer, 2023.", + "[9] Wenxuan Li, Alan Yuille, and Zongwei Zhou. How well do supervised 3d models transfer to medical imaging tasks?" + ], + "bbox": [ + 107, + 261, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In The Twelfth International Conference on Learning Representations, 2024.", + "[10] OASIS. Open access series of imaging studies data use agreement, 2024. URL https://sites.wustl.edu/oasisbrains/.", + "[11] PPMI. Parkinson's progression markers initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=PPMI.", + "[12] ADNI. Alzheimer's disease neuroimaging initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=ADNI.", + "[13] MCSA. Mayo clinic study of aging data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=MCSA.", + "[14] HABS-HD. Health and aging brain study: Health disparities data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project $\\equiv$ HABS_HD.", + "[15] Tassilo Wald, Constantin Ulrich, Stanislav Lukyanenko, Andrei Goncharov, Alberto Paderno, Leander Maerkisch, Paul F Jaeger, and Klaus Maier-Hein. Revisiting mae pretraining for 3d medical image segmentation. arXiv preprint arXiv:2410.23132, 2024.", + "[16] Guotai Wang, Jianghao Wu, Xiangde Luo, Xinglong Liu, Kang Li, and Shaoting Zhang. Mis-fm: 3d medical image segmentation using foundation models pretrained on a large-scale unannotated dataset. arXiv preprint arXiv:2306.16925, 2023.", + "[17] Asbjørn Munk, Jakob Ambsdorf, Sebastian Llambias, and Mads Nielsen. Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation. arXiv preprint arXiv:2408.00640, 2024.", + "[18] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023.", + "[19] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023.", + "[20] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023.", + "[21] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021.", + "[22] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024." + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23847", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022.", + "[24] Fenghe Tang, Ronghao Xu, Qingsong Yao, Xueming Fu, Quan Quan, Heqin Zhu, Zaiyi Liu, and S Kevin Zhou. Hyspark: Hybrid sparse masking for large scale medical image pre-training. arXiv preprint arXiv:2408.05815, 2024.", + "[25] National Institutes of Health et al. Adolescent brain cognitive development study, 2015.", + "[26] Ibrahim Ethem Hamamci, Sezgin Er, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Muhammed Furkan Dasdelen, Omer Faruk Durugol, Bastian Wittmann, Tamaz Amiranashvili, et al. Developing generalist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834, 2024.", + "[27] Kenneth Marek, Sohini Chowdhury, Andrew Siderowf, Shirley Lasch, Christopher S Coffey, Chelsea Caspell-Garcia, Tanya Simuni, Danna Jennings, Caroline M Tanner, John Q Trojanowski, et al. The parkinson's progression markers initiative (ppmi)-establishing a pd biomarker cohort. Annals of clinical and translational neurology, 5 (12):1460-1477, 2018.", + "[28] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics, 15(4):869-877, 2005.", + "[29] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021.", + "[30] Christopher J Markiewicz, Krzysztof J Gorgolewski, Franklin Feingold, Ross Blair, Yaroslav O Halchenko, Eric Miller, Nell Hardcastle, Joe Wexler, Oscar Esteban, Mathias Goncavles, et al. The openneuro resource for sharing of neuroscience data. *Elife*, 10:e71774, 2021.", + "[31] Michal Nohel, Constantin Ulrich, Jonathan Suprijadi, Tassilo Wald, and Klaus Maier-Hein. Unified framework for foreground and anonymization area segmentation in ct and mri data. In Bildverarbeitung für die Medizin 2025, 2025.", + "[32] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023.", + "[33] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540, 2023.", + "[34] Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F" + ], + "bbox": [ + 99, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 488-498. Springer, 2024.", + "[35] Tassilo Wald, Saikat Roy, Fabian Isensee, Constantin Ulrich, Sebastian Ziegler, Dasha Trofimova, Raphael Stock, Michael Baumgartner, Gregor Kohler, and Klaus Maier-Hein. Primus: Enforcing attention usage for 3d medical image segmentation, 2025. URL https://arxiv.org/abs/2503.01835.", + "[36] Sook-Lei Liew, Bethany P Lo, Miranda R Donnelly, Artemis Zavaliangos-Petropulu, Jessica N Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P Simon, Julia M Juliano, Anisha Suri, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific data, 9(1):320, 2022.", + "[37] Endre Grøvik, Darwin Yi, Michael Iv, Elizabeth Tong, Daniel Rubin, and Greg Zaharchuk. Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri. Journal of Magnetic Resonance Imaging, 51(1):175-182, 2020.", + "[38] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems, 35: 36722–36732, 2022.", + "[39] Nicholas Heller, Fabian Isensee, Klaus H Maier-Hein, Xiaoshuai Hou, Chunmei Xie, Fengyi Li, Yang Nan, Guangrui Mu, Zhiyong Lin, Miofei Han, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis, 67:101821, 2021.", + "[40] Moritz R Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, et al. Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data, 9(1):762, 2022.", + "[41] Kareem Wahid, Cem Dede, Mohamed Naser, and Clifton Fuller. Training dataset for hntsmrg 2024 challenge, 2024. URL https://zenodo.org/doi/10.5281/ zenodo.11199559.", + "[42] Gasper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023.", + "[43] Ali M. Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita binti Hanafi, Osama Alnuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi. Brain mri dataset of multiple sclerosis with consensus manual lesion segmentation and patient meta information. Data in Brief, 42:108139, 2022. ISSN 2352-3409. doi: https://doi.org/10.1016/j.dib.2022.108139. URL https://www.sciencedirect.com/science/article/pii/S235234092200347X.", + "[44] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23848", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hoher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, Suprosanna Shit, Houjing Huang, Chinmay Prabhakar, Ezequiel de la Rosa, Diana Waldmannstetter, Florian Kofler, Fernando Navarro, Martin Menten, Ivan Ezhov, Daniel Rueckert, Iris Vos, Ynte Ruigrok, Birgitta Velthuis, Hugo Kuijf, Julien Hammerli, Catherine Wurster, Philippe Bijlenga, Laura Westphal, Jeroen Bisschop, Elisa Colombo, Hakim Baazaoui, Andrew Makmur, James Hallinan, Bene Wiestler, Jan S. Kirschke, Roland Wiest, Emmanuel Montagnon, Laurent Letourneau-Guillon, Adrian Galdran, Francesco Galati, Daniele Falcetta, Maria A. Zuluaga, Chaolong Lin, Haoran Zhao, Zehan Zhang, Sinyoung Ra, Jongyun Hwang, Hyunjin Park, Junqiang Chen, Marek Wodzinski, Henning Muller, Pengcheng Shi, Wei Liu Ting Ma, Cansu Yalçin, Rachika E. Hamadache, Joaquim Salvi, Xavier Llado, Uma Maria Lal-Trehan Estrada, Valeria Abramova, Luca Giancardo, Arnau Oliver, Jialu Liu, Haibin Huang, Yue Cui, Zehang Lin, Yusheng Liu, Shunzhi Zhu, Tatsat R. Patel, Vincent M. Tutino, Maysam Orouskhani, Huayu Wang, Mahmud Mossa-Basha, Chengcheng Zhu, Maximilian R. Rokuss, Yannick Kirchhoff, Nico Disch, Julius Holzschuh, Fabian Isensee, Klaus Maier-Hein, Yuki Sato, Sven Hirsch, Susanne Wegener, and Bjoern Menze. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra, 2024. URL https://arxiv.org/abs/2312.17670.", + "[45] Divya Ramakrishnan, Leon Jekel, Saahil Chadha, Anastasia Janas, Harrison Moy, Nazanin Maleki, Matthew Sala, Manpreet Kaur, Gabriel Cassinelli Petersen, Sara Merkaj, et al. A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information. Scientific Data, 11(1):254, 2024.", + "[46] H. Chen, X. Zhao, J. Dou, C. Du, R. Yang, H. Sun, S. Yu, H. Zhao, C. Yuan, and N. Balu. Carotid vessel wall segmentation and atherosclerosis diagnosis challenge, 2022. URL https://vessel-wall-segmentation-2022.grand-challenge.org.", + "[47] Olivier Bernard, Alain Lalande, Clement Zotti, Cervenansky, and et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE TMI, 2018.", + "[48] Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet. PLoS medicine, 15(11):e1002699, 2018.", + "[49] Hui Ming Lin, Errol Colak, Tyler Richards, Felipe C Kitamura, Luciano M Prevedello, Jason Talbott, Robyn L Ball, Ekim Gumeler, Kristen W Yeom, Mohammad Hamghalam, et al. The rsna cervical spine fracture ct dataset. Radiology: Artificial Intelligence, 5(5):e230034, 2023.", + "[50] Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain ar" + ], + "bbox": [ + 99, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chitecture in autism. Molecular psychiatry, 19(6):659-667, 2014.", + "[51] Pedro RAS Bassi, Wenxuan Li, Yucheng Tang, Fabian Isensee, Zifu Wang, Jieneng Chen, Yu-Cheng Chou, Yannick Kirchhoff, Maximilian R Rokuss, Ziyan Huang, et al. Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation? Advances in Neural Information Processing Systems, 37:15184-15201, 2025.", + "[52] Lachlan T. Strike, Narelle K. Hansell, Jessica L. Miller, Kai-Hsiang Chuang, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin adolescent brain (qtab)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004146.v1.0.4.", + "[53] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-id1000\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso003097.v1.2.1.", + "[54] Makayla Gibson, Roger Newman-Norlund, Leonardo Bonilha, Julius Fridriksson, Gregory Hickok, Argye E. Hillis, Dirk-Bart den Ouden, and Chris Rorden. \"aphasia recovery cohort (arc) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004884.v1.0.2.", + "[55] Chris Rorden, John Absher, and Roger Newman-Norlund. \"stroke outcome optimization project (soop)\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004889.v1.1.2.", + "[56] Allison C. Nugent, Adam G Thomas, Margaret Mahoney, Alison Gibbons, Jarrod Smith, Antoinette Charles, Jacob S Shaw, Jeffrey D Stout, Anna M Namyst, Arshitha Basavaraj, Eric Earl, Dustin Moraczewski, Emily Guinee, Michael Liu, Travis Riddle, Joseph Snow, Shruti Japee, Morgan Andrews, Adriana Pavletic, Stephen Sinclair, Vinai Roopchansingh, Peter A Bandettini, and Joyce Chung. \"the nimh healthy research volunteer dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005752.v2.1.0.", + "[57] Rani Gera, Maya Bar Or, Ido Tavor, Dana Roll, Jeffrey Cockburn, Segev Barak, Elizabeth Tricomi, John P. O'Doherty, and Tom Schonberg. \"characterizing habit learning in the human brain at the individual and group levels: a multi-modal mri study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004299.v1.0.0.", + "[58] Denise Park, Joseph Hennessee, Evan T. Smith, Micaela Chan, Carson Katen, Julia Bacci, Sarah Frank, Sarah Monier, Alexandra Collyer, Carol Tamminga, William Moore, Neil Rofsky, Karen Rodrigue, Kristen Kennedy, and Gagan Wig. \"the dallas lifespan brain study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004856.v1.2.0.", + "[59] Jin Wang, Marisa N. Lytle, Yael Weiss, Brianna L. Yasasaki, and James R. Booth. \"a longitudinal neuroimaging dataset on language processing in children ages 5, 7, and 9 years old\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003604.v1.0.7.", + "[60] Leon Y. Cai, Qi Yang, Praitayini Kanakaraj, Vishwesh Nath, Allen T. Newton, Heidi A. Edmonson, Jeffrey Luci, Benjamin N. Conrad, Gavin R. Price, Colin B. Hansen, Cailey I. Kerley, Karthik Ramadass, Fang-Cheng Yeh, Hakmook Kang, Eleftherios Garyfallidis, Maxime Descoteaux," + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23849", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Francois Rheault, Kurt G. Schilling, and Bennett A. Landman. \"masivar: Multisite, multiscanner, and multisubject acquisitions for studying variability in diffusion weighted magnetic resonance imaging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003416.v2.0.2.", + "[61] Anahit Babayan, Blazeij Baczkowski, Roberto Cozatl, Maria Dreyer, Haakon Engen, Miray Erbey, Marcel Falkiewicz, Nicolas Farrugia, Michael Gaebler, Johannes Golchert, Laura Golz, Krzysztof Gorgolewski, Philipp Haueis, Julia Huntenburg, Rebecca Jost, Yelyzaveta Kramarenko, Sarah Krause, Deniz Kumral, Mark Lauckner, Daniel S. Margulies, Natacha Mendes, Katharina Ohrnberger, Sabine Oligschläger, Anastasia Osoianu, Jared Pool, Janis Reichelt, Andrea Reiter, Josefin Röbbig, Lina Schaare, Jonathan Smallwood, and Arno Villringer. \"mpileipzig_mind-brain-body\". OpenNeuro, 2020. doi: 10. 18112/openneuro.ds000221.v1.0.0.", + "[62] Karl Koschutnig, Bernhard Weber, and Andreas Fink. \"tidying up white matter: Neuroplastic transformations in sensorimotor tracts following slackline skill acquisition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003138.v1.0.1.", + "[63] Sangil Lee and Joseph Kable. \"cognitive training\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds002843.v1.0.1.", + "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022.", + "[65] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580, 2023.", + "[66] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020.", + "[67] Tom Schonberg, Christopher Trepel, Craig Fox, and Russell A. Poldrack. \"balloon analog risk-taking task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000001.v1.0.0.", + "[68] J.M. Moran, E. Jolly, and J.P. Mitchell. \"social cognitive tasks (animate movement task, moral judgment task, false belief task)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000109.v2.0.2.", + "[69] Ella Gabitov, David Manor, and Avi Karni. \"learning and memory: motor skill consolidation and intermanual transfer\". OpenNeuro, 2018.", + "[70] A. Roy, R. A. Bernier, J. Wang, M. Benson, J. French, D. Good, and F.G. Hillary. \"cost analysis tbi\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms000220.v1.0.0.", + "[71] Noritaka Yoneyama, Hirohisa Watanabe, Kazuya Kawabata, Epifanio Bagarinao, Kazuhiro Hara, Takashi Tsuboi, Yasuhiro Tanaka, Reiko Ohdake, Kazunori Imai, Rei Ogura, Yasutaka Kato, Michihito Masuda, Tatsuya Hattori, Mizuki Ito, Satoshi Maesawa, Daisuke Mori, Masahisa Katsuno, and Gen Sobue. \"ds000245_r1.0.0\". OpenNeuro, 2018." + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] H Aerts and D Marinazzo. \"btc_preop\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s001226.v5.0.0.", + "[73] Johns CL, Jahn AA, Jones HR, Kush DK, Molfese PJ, Van Dyke JA, Magnuson JS, Tabor W, Mencl WE, Shankweiler DP, and Braze D. \"indiv diffs readingskill\". OpenNeuro, 2018. doi: 10.18112/openneuro.d001365.v1.0.0.", + "[74] C. Baldassano, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (scrambled)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001511.v1.0.5.", + "[75] A.M. Bornstein, M. Aly, S. Feng, N.B. Turk-Browne, K.A. Norman, and J.D. Cohen. \"perceptual decisions result from the continuous accumulation of memory and sensory evidence\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001614.v1.1.0.", + "[76] Unknown authors. \"sc mri dataset\". OpenNeuro, 2019.", + "[77] Eduardo A. Garza-Villarreal, Victor Pando-Naude, Fernando A. Barrios, Peter Vuust, Lene Vase, Elvira Brattico, Erick H. Pasaye, and Sarael Alcauter. \"functional connectivity of music-induced analgesia in fibromyalgia\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001928.v1.1.0.", + "[78] B. Sherman and N.B. Turk-Browne. \"stat episodic\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds002144.v2.0.0.", + "[79] Gilles de Hollander and Tomas H. Knapen. \"odc mapper 7t surface coil\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002295.v1.0.0.", + "[80] Tomoyasu Horikawa, Alan S. Cowen, Dacher Keltner, and Yukiyasu Kamitani. \"the neural representation of visually evoked emotion is high-dimensional, categorical, and distributed across transmodal brain regions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002425.v1.0.2.", + "[81] Yazhu Qian, Jinyou Zou, Zihao Zhang, Zhentao Zuo, Yan Zhuo, Danny JJ, Wang, and Peng Zhang. \"robust functional mapping of layer-selective responses in human lateral geniculate nucleus with high-resolution 7t fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002672.v1.0.0.", + "[82] G. Elliott Wimmer, Jamie Li, Krzysztof Jacek Gorgolewski, and Russell A. Poldrack. \"rewardbeast\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s002738.v1.0.2.", + "[83] Francesca M M Citron, Nora Michaelis, and Adele E Goldberg. \"l2_sentences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002842.v1.0.1.", + "[84] Matteo Visconti di Oleggio Castello, Vassiki Chauhan, Guo Jiahui, and M. Ida Gobbini. \"an fmri dataset in response to \"the grand budapest hotel\", a socially-rich, naturalistic movie\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds003017.v1.0.3.", + "[85] Chiara Banfi, Karl Koschutnig, Kristina Moll, Gerd Schulte-Koerne, Andreas Fink, and Karin Landerl. \"reading-related functional activity in children with isolated spelling deficits and dyslexia\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003126.v1.3.1.", + "[86] Andy Schumann, Feliberto de la Cruz, Stefanie K\"ohler, Lisa Brotte, and Karl-J\"urgen B\"ar. \"the influence of heart rate variability biofeedback on cardiac regulation and functional brain connectivity\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003357.v1.0.0." + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23850", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant4 motion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003441.v1.0.0.", + "[88] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., Todd R. Parrish, and Barbara Trommer. \"response inhibition and selective attention in adults and children with and without ADHD\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003500.v1.2.0.", + "[89] Nathan M. Muncy, C. Brock Kirwan, and Jonathan J. Wisco. \"differences in chemo-signaling compound-evoked brain activity in male and female young adults: A pilot study in the role of sexual dimorphism in olfactory chemosignaling\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003606.v1.0.0.", + "[90] Beau Sievers, Carolyn Parkinson, Peter J. Kohler, James Hughes, Sergey V. Fogelson, and Thalia Wheatley. \"visual and auditory brain areas share a representational structure that supports emotion perception: fmri data\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003715.v1.0.0.", + "[91] Andreas Fink, Karl Koschutnig, Thomas Zussner, Corinna M. Perchtold-Stefan, Christian Rominger, Mathias Benedek, and Ilona Papousek. \"a two-week running intervention reduces symptoms related to depression and increases hippocampal volume in young adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003799.v2.0.0.", + "[92] John Wilder, Morteza Rezanejad, Sven Dickinson, Kaleem Siddiqi, Allan Jepson, and Dirk B. Walther. \"sceneperceptionandparallelism\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03950.v1.0.0.", + "[93] Zoe Woodhead, Maria Sablik, Kate Watkins, David Carey, Dorothy Bishop, and Emma Karlsson. \"comparing language lateralisation using fmri and ftcd\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004073.v1.0.1.", + "[94] Martin Szinte, Tomas Knapen, and Matthias Nau. \"gaze_exp\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004271.v1.1.0.", + "[95] Maria Czarnecka, Andre Knops, and Marcin Szwed. \"number comparison task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004400.v1.0.6.", + "[96] Alex L. White, Kendrick Kay, Kenny A. Tang, and Jason D. Yeatman. \"raw data for engaging in word recognition elicits highly specific modulations in visual cortex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004489.v1.0.1.", + "[97] Loriana Goulding, Austin Schmidt, Isaac Hamm, and C. Brock Kirwan. \"fmri investigations of individual differences on memory activation for faces and words: The effects of handedness and phenomenal experience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004589.v1.0.0.", + "[98] Halle R. Dimsdale-Zucker and Christopher Baldassano. \"common\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004663.v1.0.4.", + "[99] Rotem Botvinik-Nezer, Bogdan Petre, Marta Ceko, Naomi Friedman, and Tor Wager. \"paingen_placebo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004746.v1.0.1.", + "[100] Xuefeng Yan, Martin Noergaard, Cheryl L. Morse, JeihSan Liow, Jinsoo Hong, Sanjay Telu, Min-Jeong Kim, Jose" + ], + "bbox": [ + 94, + 92, + 482, + 900 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "A Montero Santamaria, Anthony Galassi, Ningping Feng, Sarah K. Williams Avram, Ted B. Usdin, Shawn Wu, Andrea Zhang, Lester Manly, Madeline Jenkins, Maia Van Buskirk, Adrian Lee, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. \"positron emission tomography (pet) quantification in healthy humans of cyclooxygenase-2 (cox-2), a potential biomarker of neuroinflammation\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004869.v1.1.1.", + "[101] Liberty S. Hamilton, Maansi Desai, and Alyssa Field. \"wired icm sample dataset - workshop on intracranial recordings in humans, epilepsy, dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004993.v1.1.2.", + "[102] Keith M. Vogt, Alex C. Burlew, Sujatha N. Reddy, Marcus Simmons, Courtney N. Kozdron, and James W. Ibinson. \"neural correlates of lidocaine analgesic (nla) study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005088.v1.0.0.", + "[103] Sidhant Chopra, Carrisa V. Cocuzza, Connor Lawhead, Jocelyn A. Ricard, Loic Labache, Lauren Patrick, Poornima Kumar, Arielle Rubenstein, Julia Moses, Lia Chen, Crystal Blankenbaker, Bryce Gillis, Laura T. Germine, Ilan Harpaz-Rote, BT Thomas Yeo, Justin T. Baker, and Avram J. Holmes. \"transdiagnostic connectome project\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005237.v1.1.3.", + "[104] Marleen Haupt, Douglas D. Garrett, and Radoslaw M. Cichy. \"object recognition in healthy aging (orha) - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005374.v1.0.1.", + "[105] Luis Martinez Agulleiro, Bowen Deng, Yuliya Yoncheva, Chao-Gan Yan, F Xavier Castellanos, and Marie-Pierre St-Onge. \"circadian misalignment and energy balance\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005525.v1.0.0.", + "[106] A.R. Aron, M.A. Gluck, and R.A. Poldrack. \"classification learning\". OpenNeuro, 2018.", + "[107] Rebecca J. Lepping, Ruth Ann Atchley, Evangelia Chrysikou, Laura E. Martin, Alicia A. Clair, Rick E. Ingram, W. Kyle Simmons, and Cary R. Savage. \"neural processing of emotional musical and nonmusical stimuli in depression\". OpenNeuro, 2018.", + "[108] Elizabeth Bock, Peter Donhauser, Francois Tadel, Guiomar Niso, and Sylvain Baillet. \"meg-bids brainstorm data sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000246.v1.0.1.", + "[109] Joseph M. Orr and Vijay A. Mittal. \"rule learning in adolescents at clinical high risk for psychosis shows altered patterns of activation\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps01371.v1.1.1.", + "[110] A.C. Schapiro, T.T. Rogers, N.I. Cordova, N.B. Turk-Browne, and M.M. Botvinick. \"neural representations of events arise from temporal community structure\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001621.v1.1.0.", + "[111] Omer Faruk Gulban, Kevin R. Sitek, Satrajit S. Ghosh, Michelle Moerel, and Federico De Martino. \"auditory localization with 7t fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001942.v1.2.0.", + "[112] Tomoya Nakai and Shinji Nishimoto. \"over 100 task fmri" + ], + "bbox": [ + 94, + 92, + 903, + 900 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "23851", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002306.v1.1.0.", + "[113] Laura Pritschet, Tyler Santander, Caitlin M. Taylor, Evan Layher, Shuying Yu, Michael B. Miller, Scott T. Grafton, Emily G. Jacobs, Morgan Fitzgerald, and Joshua M. Mueller. \"28andme\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002674.v1.0.6.", + "[114] PhD Heidemarie Laurent, Megan K. Finnegan, and Katherine Haigler. \"postnatal affective mri dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003136.v1.0.0.", + "[115] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant5 suboptimal event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003442.v1.0.0.", + "[116] Monica Crotti, Karl Koschutnig, and Selina Wriessnegger. \"the impact of handedness on the neural correlates during kinesthetic motor imagery: a fmri study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003612.v1.0.4.", + "[117] Ben Deen and Winrich Freiwald. \"parallel systems for social and spatial reasoning\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms003814.v1.0.0.", + "[118] Shaonan Wang, Xiaohan Zhang, Jiajun Zhang, and Chengqing Zong. \"a synchronized multimodal neuroimaging dataset to study brain language processing\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004078.v1.2.1.", + "[119] Shani Agron, Reut Weissgross, Eva Mishor, Lior Gorodisky, Aharon Ravia, Edna Furman-Haran, Tali Weiss, and Noam Sobel. \"a chemical signal in human emotional tears blocks aggression in men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004274.v1.0.0.", + "[120] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004493.v1.0.2.", + "[121] Jose Pedro Manzano Patron, Steen Moeller, Jesper L.R. Andersson, Kamil Ugurbil, Essa Yacoub, and Stamatios N. Sotiropoulos. \"edden: Evaluation of diffusion mri denoising\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004666.v1.0.5.", + "[122] Samira Epp, Gabriel Castrillon, Beija Yuan, Jessica Andrews-Hanna, Christine Preibisch, and Valentin Riedl. \"two distinct modes of hemodynamic responses in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004873.v2.0.6.", + "[123] Chloe M. de Nys, Ee Shern Liang, Marita Prior, Maria A. Woodruff, James I. Novak, Ashley R. Murphy, Zhiyong Li, Craig D. Winter, and Mark C. Allenby. \"royalbrisbane_tofmra_intracranial aneurysm_database\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005096.v1.0.3.", + "[124] Luis Marcos-Vidal, Helena Gil-Buitrago, Irene Cisma, Rosamunde C. Hendricks, Scott Atran, and Clara Pretus. \"polex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005375.v1.0.0.", + "[125] G. Xue and Russell A. Poldrack. \"rhyme judgment\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms000003.v1.0.0." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[126] Tom S.M., Fox C.R., Trepel C., and Poldrack R.A. \"mixed-gambles task\". OpenNeuro, 2018.", + "[127] K Jimura, E Stover, F Casalis, and RA Poldrack. \"ds000006\". OpenNeuro, 2018.", + "[128] Xue G, Aron AR, and Poldrack RA. \"stop-signal task with spoken & manual responses\". OpenNeuro, 2018.", + "[129] A.R. Aron, T.E. Behrens, M. Frank, S. Smith, and R.A. Poldrack. \"stop-signal task with unconditional and conditional stopping\". OpenNeuro, 2018.", + "[130] Jessica Cohen and Russell Poldrack. \"the generality of self-control\". OpenNeuro, 2018.", + "[131] K. Foerde, B.J. Knowlton, and Russell A. Poldrack. \"classification learning and tone-counting\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso000011.v1.0.0.", + "[132] Rizk-Jackson, Aron, and Russell A. Poldrack. \"ds000017\". OpenNeuro, 2018.", + "[133] R Bilder, R Poldrack, T Cannon, E London, N Freimer, E Congdon, K Karlgodt, and F Sabb. \"ucla consortium for neuropsychiatric phenomics la5c study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000030.v1.0.0.", + "[134] Russell Poldrack. \"myconnectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs000031.v2.0.2.", + "[135] R. Alvarez and R.A. Poldrack. \"cross-language repetition priming\". OpenNeuro, 2018.", + "[136] R.A. Poldrack, J. Clark, E. J. Pare-Blagoev, D. Shohamy, J. Creso Moyano, C. Myers, and M Gluck. \"classification learning and reversal\". OpenNeuro, 2018.", + "[137] Mei-Yen Chen, Corey N. White, Nathan Giles, Albert Elumn, Sagar Parikh, Ungi Kim, W. Todd Maddox, and Russell A. Poldrack. \"training of loss aversion modulates neural sensitivity toward potential gains\". OpenNeuro, 2018.", + "[138] Kelly AMC and Milham MP. \"simon task\". OpenNeuro, 2018.", + "[139] Kelly AMC, Uddin LQ, Biswal BB, Castellanos FX, and Milham MP. \"flanker task (event-related)\". OpenNeuro, 2018.", + "[140] J.V. Haxby, M.I. Gobbini, M.L. Furey, A. Ishai, J.L. Schouten, and P. Pietrini. \"visual object recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000105.v3.0.0.", + "[141] K.J. Duncan, C. Pattamadilok, I. Knierim, and J.T. Devlin. \"word and object processing\". OpenNeuro, 2018.", + "[142] Wager T.D., M.L. Davidson, B.L. Hughes, M.A. Lindquist, and K.N. Ochsner. \"prefrontal-subcortical pathways mediating successful emotion regulation\". OpenNeuro, 2018.", + "[143] Melina R. Uncapher, J. Benjamin Hutchinson, and Anthony D. Wagner. \"incidental encoding task (posner cueing paradigm)\". OpenNeuro, 2018.", + "[144] Michael Hanke, Florian J. Baumgartner, Pierre Ibe, Falko R. Kaule, Stefan Pollmann, Oliver Speck, Wolf Zinke, and Jorg Stadler. \"forrest gump\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps00113.v1.3.0.", + "[145] Gorgolewski KJ, Storkey A, Bastin ME, Whittle IR, Wardlaw JM, and Pernet CR. \"a test-retest fmri dataset for motor, language and spatial attention functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000114.v1.0.2." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "23852", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[146] Grega Repov $\\mathring{\\mathrm{A}}_{\\mathrm{i}}$ and Deanna M. Barch. \"working memory in healthy and schizophrenic individuals\". OpenNeuro, 2018.", + "[147] Jennifer M Walz, Robin I Goldman, Jordan Muraskin, Bryan Conroy, Truman R Brown, and Paul Sajda. \"auditory and visual oddball eeg-fmri\". OpenNeuro, 2018.", + "[148] DG Wakeman and RN Henson. \"multisubject, multimodal face processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.rs000117.v1.1.0.", + "[149] Katerina Velanova, Mark E. Wheeler, and Beatrix Luna. \"maturational changes in anterior cingulate and frontoparietal recruitment support the development of error processing and inhibitory control (antistate)\". OpenNeuro, 2018.", + "[150] Aarthi Padmanabhana, Charles F. Geier, Sarah J. Ordaz, Theresa Teslovich, and Beatrix Luna. \"developmental changes in brain function underlying the influence of reward processing on inhibitory control (slot reward)\". OpenNeuro, 2018.", + "[151] C.F. Geier, R. Terwilliger, T. Teslovich, K. Velanova, and B. Luna. \"immaturities in reward processing and its influence on inhibitory control in adolescence (ring reward)\". OpenNeuro, 2018.", + "[152] J Xu, S Zhang, VD Calhoun, J Monterosso, and MN Potenza. \"task-related concurrent but opposite modulations of overlapping functional networks as revealed by spatial ica\". OpenNeuro, 2018.", + "[153] Nicoletta Cera, Armando Tartaro, and Stefano L. Sensi. \"modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fmri study\". OpenNeuro, 2018.", + "[154] Choong-Wan Woo, Mathieu Roy, Jason T. Buhle, and Tor D. Wager. \"distinct brain systems mediate the effects of nociceptive input and self-regulation on pain\". OpenNeuro, 2018.", + "[155] Kimberly L. H. Carpenter, Adrian Angold, Nan-Kuei Chen, William E. Copeland, Pooja Gaur, Kevin Pelphrey, Allen W. Song, and Helen L. Egger. \"preschool anxiety disorders\". OpenNeuro, 2018.", + "[156] E. Necka, M. Senderecka, B. Kucharzyk, and Falkiewicz M. \"who can afford self-control? the neural efficiency mechanism explains effective self-regulation of behavior\". OpenNeuro, 2019.", + "[157] Elissa M. Aminoff and Michael J. Tarr. \"scene perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000149.v1.0.0.", + "[158] Paul A. M. Smeets, Floor M. Kroese, Catherine Evers, and D. T. D. de Ridder. \"block design food and nonfood picture viewing task\". OpenNeuro, 2018.", + "[159] CR Pernet, P Belin, P McAleer, KJ Gorgolewski, M Valdes-Sosa, M Latinus, I Charest, PEG Bestelmeyer, RH Watson, and D Fleming. \"the human voice areas: spatial organisation and inter-individual variability in temporal and extra-temporal cortices\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000158.v1.0.0.", + "[160] Timothy D. Verstynen. \"stroop task\". OpenNeuro, 2018.", + "[161] James K. Bursley, Adrian Nestor, Michael J. Tarr, and J. David Creswell. \"offline processing in associative learning\". OpenNeuro, 2019." + ], + "bbox": [ + 89, + 90, + 482, + 900 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[162] Hsu AL, Chou KH, Chao YP, Fan HY, Wu CW, and Chen JH. \"physiological contribution in spontaneous oscillations: An approximate quality - assurance index for resting-state fmri signals\". OpenNeuro, 2019.", + "[163] Laura Koenders, Janna Cousijn, Wilhelmina A.M. Vingerhoets, Wim van den Brink, Reinout W. Wiers, Carin J. Meijer, Marise W. J. Machielsen, Dick J. Veltman, Anneke E. Goudriaan, and Lieuwe de Haan. \"t1-weighted structural mri study of cannabis users at baseline and 3 years follow up\". OpenNeuro, 2019.", + "[164] Emilia Iannilli. \"pre-adolescents exposure to manganese\". OpenNeuro, 2018.", + "[165] Gustav Nilsson, Sandra Tamm, Paolo d'Onofrio, Hanna Å Thuné, Johanna Schwarz, Catharina Lavebratt, Jia Jia Liu, Kristoffer NT Måsson, Tina Sundelin, John Axelsson, Peter Fransson, Göran Kecklund, Håkan Fischer, Mats Lekander, and Torbjörn Åkerstedt. \"the stockholm sleepy brain study: Effects of sleep deprivation on cognitive and emotional processing in young and old\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000201.v1.0.3.", + "[166] Peter Van Schuerbeek, Chris Baeken, and Johan De Mey. \"the heterogeneity in retrieved relations between the personality trait 'harm avoidance' and gray matter volumes due to variations in the vbm and roi labeling processing settings\". OpenNeuro, 2018.", + "[167] Christian Stephan-Otto, Sara Siddi, Carl Senior, Daniel Muñoz-Samons, Susana Ochoa, Ana María Sánchez Laforga, and Gildas Brébion. \"visual imagery and false memory for pictures\". OpenNeuro, 2018.", + "[168] Dara Dickstein, Mariel Y. Pullman, Corey Fernandez, Jennifer A. Short, Lale Kostakoglu, Karin Knaesaurek, Laili Soleimani, Barry Jordan, Wayne Gordon, Kristen Dams-O'Connor, Bradley N. Delman, Edmund Wong, Cheuk Y. Tang, Steven T. DeKosky, James R. Stone, Robert C. Cantu, Mary Sano, Patrick R. Hof, and Sam Gandy. \"imaging [18f]av-1451 and [18f]av-45 in acute and chronic traumatic brain injury\". OpenNeuro, 2018.", + "[169] Jongwan Kim, Jing Wang, Svetlana V. Shinkareva, and Douglas H. Wedell. \"affective videos\". OpenNeuro, 2018.", + "[170] Vincent A. Magnotta, Joy T. Matsui, Dawei Liu, Hans J. Johnson, Jeffrey D. Long, Bradley D. Bolster Jr, Bryon A. Mueller, Kelvin Lim, Susumu Mori, Karl G. Helmer, Jessica A. Turner, Sarah Reading, Mark J. Lowe, Elizabeth Aylward, Laura A. Flashman, Greg Bonett, and Jane S. Paulsen. \"dwi traveling human phantom study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000206.v1.0.0.", + "[171] Tetreault P, Mansour A, Vachon-Presseau E, Schnitzer TJ, Apkarian AV, and Baliki MN. \"brain connectivity predicts placebo response across chronic pain clinical trials\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds000208.v1.0.1.", + "[172] Elizabeth DuPre, Wen-Ming Luh, and R. Nathan Spreng. \"multi-echo fmri replication sample of autobiographical memory, prospection and theory of mind reasoning tasks\". OpenNeuro, 2018.", + "[173] Liane Young, Alek Chakroff, Emily Wasserman, Rebecca Saxe, James Dungan, Amelia Brown, and Jorie Koster-Hale. \"moral judgments of intentional and accidental moral" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 14 + }, + { + "type": "page_number", + "text": "23853", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 14 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "violations across harm and purity domains\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000212.v1.0.0.", + "[174] Xiao Gao, Xiao Deng, Xin Wen, Ying She, Petra Corianne Vinke, and Hong Chen. \"neural mechanism underlying appearance social comparison\". OpenNeuro, 2018.", + "[175] Stephen Giles, Jeremy Hall, Merrick Pope, Katie Nicol, and Liana Romaniuk. \"eupd cyberball\". OpenNeuro, 2018.", + "[176] Alexander D. Cohen, Andrew S. Nencka, R. Marc Lebel, and Yang Wang. \"multiband multi-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity\". OpenNeuro, 2018.", + "[177] Avi J H Chanales, Ashima Oza, Serra E Favila, and Brice A Kuhl. \"route learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms000217.v1.0.0.", + "[178] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (oral nutritional supplements)\". OpenNeuro, 2018.", + "[179] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (regular products)\". OpenNeuro, 2018.", + "[180] Thomas H B FitzGerald, Dorothea Haemmerer, Karl J Friston, Shu-Chen Li, and Raymond J Dolan. \"sequential inference vbm\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000222.v1.0.1.", + "[181] Ian Ballard, Bokyung Kim, Anthony Liatsis, Gökhan Aydogan, Jonathan D. Cohen, and Samuel M. McClure. \"magnitude effect\". OpenNeuro, 2018.", + "[182] Evan M. Gordon, Timothy O. Laumann, Adrian W. Gilmore, Dillan J. Newbold, Deanna J. Greene, Jeffrey J. Berg, Mario Ortega, Catherine Hoyt-Drazen, Caterina Gratton, Haoxin Sun, Jacqueline M. Hampton, Rebecca S. Coalson, Annie Nguyen, Kathleen B. McDermott, Joshua S. Shimony, Abraham Z. Snyder, Bradley L. Schlaggar, Steven E. Petersen, Steven M. Nelson, and Nico U.F. Dosenbach. \"the midnight scan club (msc) dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms000224.v1.0.4.", + "[183] Hilary Richardson, Grace Lisandrelli, Alexa Riobueno-Naylor, and Rebecca Saxe. \"mri data of 3-12 year old children and adults during viewing of a short animated film\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000228.v1.1.", + "[184] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward - study 1\". OpenNeuro, 2018.", + "[185] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward-study 3\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000231.v1.0.0.", + "[186] Johan D Carlin and Nikolaus Kriegeskorte. \"adjudicating between face-coding models with individual-face fmri responses\". OpenNeuro, 2018.", + "[187] Samuel A. Nastase, Yaroslav O. Halchenko, Andrew C. Connolly, M. Ida Gobbini, and James V. Haxby. \"neural responses to naturalistic clips of behaving animals in" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two different task contexts\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000233.v1.0.1.", + "[188] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 1\". OpenNeuro, 2018.", + "[189] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000235.v2.0.1.", + "[190] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 3\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps00236.v2.0.1.", + "[191] Yuji Kanazawa, Kimihiro Nakamura, Toru Ishii, Toshihiko Aso, Hiroshi Yamazaki, and Koichi Omori. \"phonological memory in sign language relies on the visuomotor neural system outside the left hemisphere language network\". OpenNeuro, 2019.", + "[192] D. Zeithamova, M. De Araujo Sanchez, and A. Adke. \"trial timing for multivariate pattern analysis\". OpenNeuro, 2018.", + "[193] Julian Maclaren, Zhaoying Han, Sjoerd B. Vos, Nancy Fischbein, and Roland Bammer. \"maclaren test-retest brain volume dataset\". OpenNeuro, 2018.", + "[194] Alvaro Galiano, Reyes Garcia de Eulate, Marta Vidorreta, Miriam Recio, Mario Riverol, José L. Zubieta, and Maria A. Fernandez-Seara PhD. \"resting state perfusion in healthy aging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000240.v2.0.0.", + "[195] Andrew C. Connolly, J. Swaroop Guntupalli, Jason Gors, Michael Hanke, Yaroslav O. Halchenko, Yu-Chien Wu, Hervé Abdi, and James V. Haxby.\"ak6\". OpenNeuro, 2018.", + "[196] Steve Petersen, Brad Schlaggar, and Jonathan Power. \"washington university 120\". OpenNeuro, 2018.", + "[197] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"individual brain charting\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds000244.v1.0.0.", + "[198] Guiomar Niso, Jeremy Moreau, Elizabeth Bock, Francois Tadel, and Sylvain Baillet. \"meg-bids omega restingstate_sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000247.v1.0.2.", + "[199] Alexandre Gramfort and Matti S Hämäläinen. ””. Open-Neuro, 2020. doi: 10.18112/openneuro.ds000248.v1.2.4.", + "[200] Agnes Norbury and Ben Seymour. \"value generalization in human avoidance learning\". OpenNeuro, 2018.", + "[201] Diana Gorbet and Lauren Sergio. \"female action video game players\". OpenNeuro, 2018.", + "[202] Alexander D. Cohen, Andrew S. Nencka, and Yang Wang. \"multiband multi-echo simultaneous asl/bold for task-induced functional mri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000254.v1.0.0.", + "[203] Yoichi Miyawaki, Hajime Uchida, Okito Yamashita, Masaaki Sato, Yusuke Morito, Hiroki C. Tanabe, Norihiro Sadato, and Yukiyasu Kamitani. \"visual image reconstruction\". OpenNeuro, 2018." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 15 + }, + { + "type": "page_number", + "text": "23854", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 15 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[204] Deanna J. Greene, Jonathan M. Koller, Jacqueline M. Hampton, Victoria Wesevich, Andrew N. Van, Annie L. Nguyen, Catherine R. Hoyt, Lindsey McIntyre, Eric A. Earl, Rachel L. Klein, Joshua S. Shimony, Steven E. Petersen, Bradley L. Schlaggar, Damien A. Fair, and Nico U.F. Dosenbach. \"behavioral interventions for reducing head motion during migraine scans in children\". OpenNeuro, 2018.", + "[205] Jonathan Power, Valerie Voon, Prantik Kundu, Suyash Bhogawar, Peter Bandettini, and Alex Martin. \"multi-echo dambridge\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000258.v1.0.1.", + "[206] Michael Milham, Cameron Craddock, Russ Tobe, Anna MacKay-Brandt, Stan Colcombe, Bennett Leventhal, and F. Xavier Castellanos. \"nki-sample\". OpenNeuro, 2018.", + "[207] Chris Gorgolewski. \"the brain of chris\". OpenNeuro, 2018.", + "[208] Mark A. Thornton, Miriam E. Weaverdyck, and Diana I. Tamir. \"social predictive coding\". OpenNeuro, 2018.", + "[209] A. Zadbood, J. Chen, Y.C. Leong, K.A. Norman, and U. Hasson. \"sherlock_merlin\". OpenNeuro, 2018.", + "[210] Y. Yeshurun, M. Nguyen, and U. Hasson. \"milky-vodka\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001131.v1.0.0.", + "[211] J. Chen, Y.C. Leong, C.J. Honey, K.A. Yong C.H., Norman, and U. Hasson. \"sherlock\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001132.v1.0.0.", + "[212] J. Chen, C. J. Honey, E. Simony, M. J. Arcaro, K. A. Norman, and U. Hasson. \"twilight zone movie watching dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001145.v1.0.0.", + "[213] Chris Gorgolewski, Natacha Mendes, Domenica Wilfling, Elisabeth Wladimirow, Claudine J. Gauthier, Tyler Bonnen, Florence J.M. Ruby, Robert Trampel, Pierre-Louis Bazin, Roberto Cozatl, Jonathan Smallwood, and Daniel S. Margulies. \"a high resolution 7-tesla resting-state fmri test-retest dataset with cognitive and physiological measures\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001168.v1.0.1.", + "[214] Krzysztof J. Gorgolewski. \"example mr artifacts\". Open-Neuro, 2018.", + "[215] Mark Thornton, Miriam Weaverdyck, and Diana Tamir. \"state representation pictures\". OpenNeuro, 2018.", + "[216] Sam Gilbert and Hoki Fung. \"decoding intentions of self and others\". OpenNeuro, 2018.", + "[217] J.A. Lewis-Peacock, J.D. Cohen, and K.A. Norman. \"neural evidence of the strategic choice between working memory and episodic memory in prospective remembering\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001229.v1.0.1.", + "[218] J.A. Lewis-Peacock and K.A. Norman. \"competition between items in working memory leads to forgetting\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001232.v1.0.0.", + "[219] Patrick Beukema and Timothy Verstynen. \"singlefingerrsa\". OpenNeuro, 2018.", + "[220] Christopher Baldassano, Diane M. Beck, and Li Fei-Fei. \"human-object interaction\". OpenNeuro, 2018." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[221] Michelle G. Hall, Claire K. Naughtin, Jason B. Mattingley, and Paul E. Dux. \"olvsl_ object-location visual statistical learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001241.v1.0.1.", + "[222] Tae-Ho Lee, Steven Greening, Taiji Ueno, David Clewett, Allison Ponzio, Michiko Sakaki, and Mara Mather. \"examining effects of arousal on responses to salient and non-salient stimuli in younger and older adults\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001242.v1.0.0.", + "[223] Tomoyasu Horikawa and Yukiyasu Kamitani. \"generic object decoding (fmri onImagenet)\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001246.v1.2.1.", + "[224] Damien Brevers, Sarah C Herremans, Qinghua He, Marie-Anne Vanderhasselt, Mathieu Petieau, Dimitri Verdonck, Tasha Poppa, Sara De Witte, Charles Kornreich, Antoine Bechara, and Chris Baeken. \"sports_betting_study\". Open-Neuro, 2018.", + "[225] Matteo Visconti di Oleggio Castello, Yaroslav O. Halchenko, J. Swaroop Guntupalli, Jason D. Gors, and M. Ida Gobbini. \"personally familiar faces\". OpenNeuro, 2018.", + "[226] Frank J. Kanayet, Andrew Mattarella-Micke, Peter J. Kohler, Anthony M. Norcia, Bruce D. McCandliss, and James L. McClelland. \"numberlinemarking\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001299.v1.0.0.", + "[227] Sean O'Bryan, Darrell Worthy, Evan Livesey, and Tyler Davis. \"inverse base rate\". OpenNeuro, 2018. doi: 10.18112/openneuro.ms001302.v1.0.1.", + "[228] Reshanne R. Reeder, Christian N. L. Olivers, and Stefan Pollmann. \"negative_template\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001306.v1.0.0.", + "[229] E.S. Finn, P.R. Corlett, G. Chen, P.A. Bandettini, and R.T. Constable. \"paranoiastory\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001338.v1.0.0.", + "[230] Ivy F. Tso, Saige E. Rutherford, Yu Fang, Mike Angstadt, and Stephan F. Taylor. \"socialbrain\". OpenNeuro, 2018. doi: 10.52181/zenodod.1210217.", + "[231] Jessica-Lily Harvey, Lysia Demetriou, John McGonigle, and Matthew B. Wall. \"phfmri_control_tasks\". OpenNeuro, 2018. doi: 10.18112/openneuro.rs001344.v1.0.0.", + "[232] Michael P Notter, Sandra Da Costa, and Micah M Murray. \"decoding of multisensory semantics and memories in low-level visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001345.v1.0.0.", + "[233] Brock Kirwan, Thomas Meservy, and Kelly Fadel. \"information filtering in electronic networks of practice: An fmri investigation of expectation [dis]confirmation\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001353.v1.0.0.", + "[234] Amado Catarina, Kovács Petra, Mayer Rebecca, Ambrus G. Geza, Trapp Sabrina, and Kovács Gyula. \"namefacepriming\". OpenNeuro, 2018.", + "[235] Christopher Baldassano. \"loci case study\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001379.v1.0.0.", + "[236] Dr. Frank Hillary. \"subject data\". OpenNeuro, 2018.", + "[237] Joset A. Etzel, Debbie Yee, Bidhan Lamichhane, Mitchell Jeffers, Elisa Di Rosa, Jenny Crawford, Hongyu An, and Todd S. Braver. \"multibandacqtests\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso001399.v2.0.0." + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 16 + }, + { + "type": "page_number", + "text": "23855", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 16 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[238] Dobrushina Olga R, Vlasova Roza M, Rumshiskaya Alena D, Litvinova Liudmila D, Mershina Elena A, Sinit-syn Valentin E, and Pechenkova Ekaterina V. \"rsfmri_single_session_eeg_nf\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001408.v1.0.3.", + "[239] Megen Brittell. \"audio cartography\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001415.v1.0.0.", + "[240] Rotem Botvinik-Nezer, Tom Salomon, and Tom Schonberg. \"cat snacks functional plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001417.v1.0.0.", + "[241] Martin Wegrzyn, Joana Aust, Larissa Barnstorf, Magdalena Gippert, Mareike Harms, Antonia Hautum, Shanna Heidel, Friederike Herold, Sarah M. Hommel, Anna-Katharina Knigge, Dominik Neu, Diana Peters, Marius Schaefer, Julia Schneider, Ria Vombrock, Sabrina M. Zimmer, Friedrich G. Woermann, and Kirsten Labudda. \"thought-experiment\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001419.v1.0.1.", + "[242] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]dasb pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001420.v1.2.0.", + "[243] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]sb207145 pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001421.v1.4.1.", + "[244] G. Kim, K.A. Norman, and N.B. Turk-Browne. \"neural overlap in item representations across episodes impairs context memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001430.v1.0.2.", + "[245] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001439.v1.2.0.", + "[246] A.C. Schapiro, E.A. McDevitt, T.T. Rogers, S.C. Mednick, and K.A. Norman. \"human hippocampal replay during rest prioritizes weakly learned information and predicts memory performance\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001454.v1.3.1.", + "[247] Ilaria Berteletti, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, Jérôme Prado, Macarena Suarez Pellicioni, Jessica Wise Younger, and James R. Booth. \"brain correlates of math development\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds001486.v1.3.1.", + "[248] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, and Gert J. ter Horst. \"valence processing differs across stimulus modalities (multi-echo)\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps001491.v1.0.0.", + "[249] J Lewis-Peacock, BR Postle, CR Cox, and TT Rogers. \"long-term memory (ltm) for famous faces, places, and common objects\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001497.v1.0.2.", + "[250] Nadine Chang, John A. Pyles, Austin Marcus, Abhinav Gupta, Michael J. Tarr, and Elisaa M. Aminoff. \"bold5000\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001499.v1.3.1." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[251] Guohua Shen, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"deep image reconstruction\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001506.v1.3.1.", + "[252] C. Baldassano, R. Masis-Obando, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (intact)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001510.v2.0.3.", + "[253] P. Kok and N.B. Turk-Browne. \"associative prediction of visual shape in the hippocampus\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001517.v1.0.3.", + "[254] S. C. Chan, M. C. Applegate, N. W. Morton, S. M. Polyn, and K. A. Norman. \"lingering representations of stimuli influence recall organization\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001521.v1.0.2.", + "[255] Feng Rong, A. Lisette Isenberg, Erica Sun, and Gregory S. Hickok. \"syllable level speech sequencing\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001525.v1.1.1.", + "[256] Andrea Courtney, Emma PeConga, Dylan Wagner, and Kristina Rapuano. \"calorie-labeled food cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001534.v1.1.0.", + "[257] M. Aly, J. Chen, N.B. Turk-Browne, and U. Hasson. \"learning naturalistic temporal structure in the posterior medial network\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001545.v1.1.1.", + "[258] Eliska Prochazkova. \"ds: ds001551:1.0.0\". OpenNeuro, 2018.", + "[259] Javier Gonzalez-Castillo, Ziad S. Saad, Daniel A. Handwaker, Souheil J. Inati, Noah Brenowitz, and Peter A. Bandettini. \"100 runs at 3t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001553.v1.0.1.", + "[260] O. Lositsky, J. Chen, D. Toker, C.J. Honey, J.L. Poppenk, U. Hasson, and K.A. Norman. \"neural pattern change during encoding of a narrative predicts retrospective duration estimates\". OpenNeuro, 2018.", + "[261] Javier Gonzalez-Castilloa, Colin W. Hoy, Daniel A. Handwerker, Vinai Roopchansingh, Souheil J. Inati, Ziad S. Saad, Robert W. Cox, and Peter A. Bandettini. \"large single-subject functional mri datasets at 7t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001555.v1.0.1.", + "[262] Renzo Huber. \"cmrr workshop\". OpenNeuro, 2019. doi: mockDOI.", + "[263] Unknown authors. \"ds: ds001566:1.0.1\". OpenNeuro, 2018.", + "[264] A.N. Hoskin, A.M. Bornstein, K.A. Norman, and J.D. Cohen. \"refresh my memory: Episodic memory reinstatements intrude on working memory maintenance\". Open-Neuro, 2018.", + "[265] Ian C. Ballard, Anthony D. Wagner, and Samuel M. McClure. \"feature discrimination\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001590.v1.0.1.", + "[266] Unknown authors. \"test dataset for xcp software\". Open-Neuro, 2018.", + "[267] Patrick Beukema and Timothy Verstynen. \"multifingerrsa\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001597.v1.0.0.", + "[268] Matt Cieslak, Mark Elliott, and Ted Satterthwaite. \"example fieldmaps\". OpenNeuro, 2018." + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 17 + }, + { + "type": "page_number", + "text": "23856", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 17 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[269] A.M. Bornstein and K.A. Norman. \"reinstated episodic context guides sampling-based decisions for reward\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001607.v1.0.1.", + "[270] Aya Ben-Yakov and Noga Cohen. \"prestimulus exp2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001608.v1.0.1.", + "[271] I. Momennejad, A.R. Otto, N. Daw, and K.A. Norman. \"offline replay supports planning in human reinforcement learning\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001612.v1.0.2.", + "[272] Adam Hampshire and Eyal Soreq. \"visual working memory: Study one task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001634.v1.0.1.", + "[273] Adam Hampshire and Eyal Soreq. \"visual working memory: Study two task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001635.v1.0.1.", + "[274] Elizabeth Valles. \"ironia vev\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001652.v2.0.0.", + "[275] Mattia Veronese, Gaia Rizzo, Martin Belzunce, Julia Schubert, Barbara Santangelo, Ayla Mansur, Alex Whittington, Joel Dunn, Graham Searle, Andrew Reader, and Roger Gunn. \"nrm2018 pet grand challenge dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms001705.v1.0.1.", + "[276] A. Shenhav, M.A. Straccia, S. Musslick, J.D. Cohen, and M.M. Botvinick. \"dissociable neural mechanisms track evidence accumulation for selection of attention versus action\". OpenNeuro, 2019.", + "[277] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others (experiment 2)\". OpenNeuro, 2019. doi: 10.18112/openneuro. ds001722.v1.1.0.", + "[278] Unknown authors. \"test-geier: Ds: ds001728:1.1.0\". Open-Neuro, 2019.", + "[279] Rotem Botvinik-Nezer, Roni Iwanir, Russell A. Poldrack, and Tom Schonberg. \"narps\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s001734.v1.0.5.", + "[280] Birgit Rauchbauer, Bruno Nazarian, Morgan Bourhis, Magalie Ochs, Laurent Prevot, and Thierry Chaminade. \"brain activity during reciprocal social interaction investigated using conversational robots as control condition\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001740.v2.2.0.", + "[281] Nathaniel P Miller, Yao Liu, Roman Krivochenitser, and Bas Rokers. \"unilateral glaucoma 3t dmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001743.v1.0.1.", + "[282] J.R. Manning, J.C. Hulbert, J. Williams, L. Piloto, L. Sahakyan, and K.A. Norman. \"a neural signature of contextually mediated intentional forgetting\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001745.v1.1.0.", + "[283] Carrie E. Gold, Athena L. Howell, Jacob Burdis, C. Brock Kirwan, and Gregory L. Thompson. \"exploring the resting state neural activity of monolinguals and late and early bilinguals\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001747.v1.1.0." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[284] Samuel Fynes-Clinton, Lars Marstaller, and Hana Burianova. \"differentiation of functional networks during long-term memory retrieval in children and adolescents\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001748.v1.0.4.", + "[285] Bart Aben, Cristian Buc Calderon, Laurens Van der Cruyssen, Doerte Picksak, Eva Van den Bussche, and Tom Verguts. \"context-dependent cognitive control — flanker task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001751.v1.0.2.", + "[286] R. VanRullen and L. Reddy. \"reconstructing faces from fmri patterns using deep generative neural networks\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001761.v2.0.1.", + "[287] A. Shenhav, M.A. Straccia, M.M. Botvinick, and J.D. Cohen. \"dorsal anterior cingulate and ventromedial prefrontal cortex have inverse roles in both foraging and economic choice\". OpenNeuro, 2019.", + "[288] Virginia AGLIERI, Bastien CAGNA, Pascal BELIN, and Sylvain TAKERKART. \"intertva: a multimodal mri dataset for the study of inter-individual differences in voice perception and identification\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001771.v1.0.2.", + "[289] J.C. Hulbert and K.A. Norman. \"neural differentiation tracks improved recall of competing memories following interleaved study and retrieval practice\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001775.v1.0.1.", + "[290] Alik Wadge and Sam Zorowitz. \"effects of on/off deep brain stimulation on cognitive control in treatment-resistant depression (eeg)\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001784.v1.1.2.", + "[291] Vincent DeLuca and Christos Pliatsikas. \"bilingualism and the brain\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds001796.v1.7.0.", + "[292] Samuel Zorowitz, Alexander P Rockhill, Kristen K Kellard, Katherine E Link, Todd Herrington, Diego A Pizzagalli, Alik S Widge, Thilo Deckersbach, and Darin D Dougherty. \"neuroanatomical correlates of approach-avoidance conflict (fmri)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds001814.v1.0.7.", + "[293] Unknown authors. \"cat faces mri experiment. ds: ds001818:1.0.0\". OpenNeuro, 2019.", + "[294] Isabel Alkhasli, Katrin Sakreida, Felix M. Mottaghy, and Ferdinand Binkofski. \"resting state - tms\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001832.v1.0.1.", + "[295] Celia Goffin, H. Moriah Sokolowski, Michael Slipenkyj, and Daniel Ansari. \"handedness and symbolic number representation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001838.v1.0.1.", + "[296] Unknown authors. \"spinal stimulation stepping and standing dataset. ds: ds001839:1.0.1\". OpenNeuro, 2019.", + "[297] Noga Cohen and Aya Ben-Yakov. \"prestimulus exp1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001840.v1.0.2.", + "[298] S.C.Y. Chan, Y. Niv, and K.A. Norman. \"a probability distribution over latent causes, in the orbitofrontal cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001847.v1.0.1." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 18 + }, + { + "type": "page_number", + "text": "23857", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 18 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[299] H. Moriah Sokolowski, Zachary Hawes, Lien Peters, and Daniel Ansari. \"parallel adaptation of symbols, quantities, and physical size\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001848.v1.0.1.", + "[300] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making intertemporal choice task dataset\". OpenNeuro, 2019.", + "[301] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making risky choice task dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001883.v1.0.3.", + "[302] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"longitudinal brain correlates of multisensory lexical processing in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001894.v1.4.2.", + "[303] Trevor K. M. Day, Tara M. Madyastha, Peter Board, Mary K. Askren, Nicholas J. Wapstra, Thomas J. Montine, and Thomas J. Grabowski. \"ant: Healthy aging and parkinson's disease\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001907.v3.1.0.", + "[304] Milena Kaestner, Ryan T. Maloney, Kirstie H. Wailes-Newson, Marina Bloj, Julie M. Harris, Antony B. Morland, and Alex R. Wade. \"s-cone_motiondepth_fmri\". Open-Neuro, 2019.", + "[305] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value\". OpenNeuro, 2019.", + "[306] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value (2)\". OpenNeuro, 2019.", + "[307] Ghootae Kim, Kenneth A. Norman, and Nicholas B. Turk-Browne. \"violation differentiation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001926.v1.0.1.", + "[308] Gabriel Castrillon, Nico Sollmann, Katarzyna Kurcyus, Adeel Razi, Sandro M. Krieg, and Valentin Riedl. \"the physiological effects of non-invasive brain stimulation fundamentally differ across the human cortex\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001927.v2.1.0.", + "[309] N.C. Hindy, E.W. Avery, and N.B. Turk-Browne. \"hippocampal-neocortical interactions sharpen over time for predictive actions\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001946.v1.0.3.", + "[310] Peter J. Kohler and Anthony M. Norcia. \"svndl anatomical data\". OpenNeuro, 2019.", + "[311] Peter J. Kohler, Benoit R. Cotterreau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (central letter task)\". OpenNeuro, 2019.", + "[312] Peter J. Kohler, Benoit R. Cottereau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (passive fixation)\". OpenNeuro, 2019.", + "[313] Benjamin J. Griffiths, Stephen D. Mayhew, Karen J. Mullinger, João Jorge, Ian Charest, Maria Wimber, and" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Simon Hanslmayr. \"reinstatement_fidelity\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002000.v1.0.0.", + "[314] Janine Mendola and Elizabeth Bock. \"rivalry_tagging\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002001.v1.0.0.", + "[315] Akram Bakkour, Daphna Shohamy, and Michael N. Shadlen. \"memory and decision making dataset\". Open-Neuro, 2019. doi: 10.18112/openneuro.ms002006.v1.0.1.", + "[316] David Clewett, Ringo Huang, Rico Velasco, Tae-Ho Lee, and Mara Mather. \"locus coeruleus activity strengthens prioritized memories under arousal\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002011.v1.0.0.", + "[317] Jakob Heinzle (data acquisition), Kai Goergen (data curation), and Joram Soch (BIDS conversion). \"stimdisc\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002013.v1.0.3.", + "[318] Martin Wegrzyn, Markus Mertens, Christian G. Bien, Friedrich G. Woermann, and Kirsten Labudda. \"language production fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002014.v1.0.1.", + "[319] Yiming Xiao, Jonathan C. Lau, Taylor Anderson, Jordan DeKraker, D. Louis Collins, Terry M. Peters, and Ali R. Khan. \"bigbrainmricoreg\". OpenNeuro, 2019. doi: 10.18112/openneuro.d0s002016.v1.0.0.", + "[320] Rebecca Merkley, Benjamin Conrad, Gavin Price, and Daniel Ansari. \"investigating the visual number form area: A replication study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002033.v1.0.1.", + "[321] Unknown authors. \"bids_parkinson\". OpenNeuro, 2019.", + "[322] Jaime J. Castrellon, Jacob S. Young, Linh C. Dang, Ronald L. Cowan, David H. Zald, and Gregory R. Samanez-Larkin. \"mesolimbic dopamine d2 receptors and neural representations of subjective value\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002041.v2.0.0.", + "[323] Emily S Finn, Laurentius Huber, David C Jangraw, Peter J Molfese, and Peter A Bandettini. \"layer-dependent activity in human prefrontal cortex during working memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002076.v1.0.1.", + "[324] Hannelore Aerts and Daniele Marinazzo. \"btc_postop\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002080.v4.0.0.", + "[325] Maedbh King, Carlos Hernandez-Castillo, Russell A. Poldrack, Jonathon Walters, Richard B. Ivry, and Jorn Diedrichsen. \"multi-domain task battery (mdtb)\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002105.v1.1.0.", + "[326] Celia Goffin, Michael Slipenkyj, Stephan E. Vogel, Rebecca Merkley, and Daniel Ansari. \"development of symbolic number processing\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002116.v1.0.0.", + "[327] Adam Kimbler, Amanda G. Hamm, and Aaron T. Mattfeld. \"conditional visual associative learning task\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002149.v1.0.5.", + "[328] Charlie, John, Ross, and Joey. \"testrun\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002153.v1.0.1.", + "[329] Florence Bouhali, Zoé Bézagu, Stanislas Dehaene, and Laurent Cohen. \"a mesial-to-lateral dissociation for ortho-" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 19 + }, + { + "type": "page_number", + "text": "23858", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 19 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graphic processing in the visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso002155.v1.0.0.", + "[330] Dylan M. Nielson and Argyris Stringaris. \"multi-echo masking test dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002156.v2.0.0.", + "[331] Michael Pereira, Nathan Faivre, Inaki Iturrate, Marco Wirthlin, Luana Serafini, Stephanie Martin, Arnaud Desvachez, Olaf Blanke, Dimitri Van de Ville, and Jose del R. Millan. \"disentangling the origins of confidence in speeded perceptual judgments through multimodal imaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002158.v1.0.2.", + "[332] Michelle I. Nash, Cooper B. Hodges, Nathan M. Muncy, and C. Brock Kirwan. \"pattern separation beyond the hippocampus: A high-resolution whole-brain investigation of mnemonic discrimination in healthy adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002168.v1.1.0.", + "[333] Jiefeng Jiang, Shao-Fang Wang, Wanjia Guo, Corey Fernandez, and Anthony D. Wagner. \"ccd\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso02169.v1.0.0.", + "[334] Tali Weiss, Timna Soroka, Lior Gorodisky, Sagit Shushan, Kobi Snitz, Reut Weisgross, Edna Furman-Haran, Thijs Dhollander, and Noam Sobel. \"human olfaction without apparent olfactory bulbs\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002185.v1.1.0.", + "[335] Michele Svanera and Lars Muckli. \"cerebrum: a 3t mri segmentation tool\". OpenNeuro, 2019. doi: mockDOI.", + "[336] Dorit Kliemann and Ralph Adolphs. \"caltech rsfmri dataset\". OpenNeuro, 2019.", + "[337] Jordan Bigio, Tali Bitan, Douglas Bolger, Douglas Burman, Fan Cao, Tai-Li Chou, Nadia Cone, Jessica Gayda, Dong Lu, Marisa Lytle, Jenni Minas, and James R. Booth. \"cross-sectional multidomain lexical processing\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002236.v1.1.1.", + "[338] J. Poppenk and K. A. Norman. \"briefly cueing memories leads to suppression of their neural representations\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds002237.v1.1.0.", + "[339] J. Fan, J.D. Wammes, J.B. Gunn, D.L. Yamins, K.A. Norman, and N.B. Turk-Browne. \"relating visual production and recognition of objects in human visual cortex\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002241.v1.1.1.", + "[340] Nathan M. Muncy and C. Brock Kirwan. \"correcting false memories: The effect of mnemonic generalization on original memory traces\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002242.v1.0.0.", + "[341] Joshua Hirschfeld-Kroen, Emily Wasserman, Stefano Anzellotti, and Liane Young. \"when my wrongs are worse than yours: behavioral and neural asymmetries in first-person and third-person perspectives of accidents\". Open-Neuro, 2020. doi: 10.18112/openneuro.d002250.v1.0.1.", + "[342] Jean-Marie Bonny, Juliana Iranpour, and Béatrice Claise. \"viewing_food_cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002270.v1.0.0.", + "[343] Katherine L. Bottenhorn, Taylor Salo, Julio A. Peraza, Michael C. Riedel, Jessica S. Flannery, Adam Kimbler, Alfredo Toll, Diego Suarez, Francis M. Cruz, Israel Zagales, Nayade Caldes, Olivia Dolan, Ruth Zagales, Matthew T." + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sutherland, Robert W. Laird, and Angela R. Laird. \"dense investigation of variability of affect (diva)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002278.v2.0.0.", + "[344] Unknown authors. \"fsl open science dev dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002293.v1.0.0.", + "[345] Elizabeth Long, Nathan Wheeler, and William Cunningham. \"trait-role\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002294.v1.0.1.", + "[346] J. Poppenk and K.A. Norman. \"multiple-object tracking as atool for parametrically modulating memory reactivation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002311.v1.1.0.", + "[347] P. G. Bissett, J. M. Shine, J. Durnez, J. Li, K. J. Gorgolewski, and R. A. Poldrack. \"rewardbeast\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002316.v1.0.0.", + "[348] Luiz Pessoa, Christian Meyer, and Srikanth Padmala. \"dynamic_passive-threat\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002320.v1.1.0.", + "[349] Shohini Bhattasali, Jonathan R. Brennan, Wen-Ming Luh, Berta Franzluebbers, and John T. Hale. \"the alice dataset: fmri dataset to study natural language comprehension in the brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002322.v1.0.4.", + "[350] Taylor Hanayik. \"fsl open science dev dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms002328.v1.0.2.", + "[351] A. Sunavsky and J. Poppenk. \"neuroimaging predictors of creativity in healthy adults\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002330.v1.1.0.", + "[352] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp1\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002336.v2.0.2.", + "[353] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp2\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002338.v2.0.2.", + "[354] Samuel A. Nastase, Yun-Fei Liu, Hanna Hillman, Asieh Zadbood, Liat Hasenfratz, Neggin Keshavarzian, Janice Chen, Christopher J. Honey, Yaara Yeshurun, Mor Regev, Mai Nguyen, Claire H. C. Chang, Christopher Baldassano, Olga Lositsky, Erez Simony, Michael A. Chow, Yuan Chang Leong, Paula P. Brooks, Emily Micciche, Gina Choe, Ariel Goldstein, Tamara Vanderwal, Yaroslav O. Halchenko, Kenneth A. Norman, and Uri Hasson. \"narratives\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002345.v1.1.4.", + "[355] Brock Kirwan, Daniel Bjornn, Bonnie Anderson, Anthony Vance, David Eargle, and Jeffrey Jenkins. \"repetition of computer security warnings results in differential repetition suppression effects as revealed with functional mri\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002363.v1.1.1.", + "[356] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 20 + }, + { + "type": "page_number", + "text": "23859", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 20 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "\"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds002366.v1.1.0.", + "[357] Deni Kurban, Renzo Huber, Gilad Liberman, Dimo Ivanov, Sri Kashyap, Debora Niekaemper, and Benedikt Poser. \"ohbm abstract 2020\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002367.v1.0.0.", + "[358] M.D. Rajendra Morey. \"serumgreymattercorticalthickness\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002380.v1.0.1.", + "[359] Rogers CS, Jones MS, McConkey S, Spehar B, Van Engen KJ, Sommers MS, and Peelle JE. \"age-related differences in auditory cortex activity during spoken word recognition\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002382.v1.0.1.", + "[360] Yoshiharu I Kutani, Takatomi Kubo, Satoshi Nishida, Hideaki Hata, Kenichi Matsumoto, Kazushi Ikeda, and Shinji Nishimoto. \"fmri dataset on program comprehension and expertise\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002411.v1.1.0.", + "[361] Jelle R. Dalenberg, Barkha P. Patel, Raphael Denis, Maria G. Veldhuizen, Yuko Nakamura, Petra C. Vinke, Serge Luquet, and Dana M. Small. \"short-term consumption of sucralose with, but not without, carbohydrate impairs neural and metabolic sensitivity to sugar\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002419.v1.0.3.", + "[362] Yana Panikratova, Alexander Tomyshev, Ekaterina Pechenkova, and Roza Vlasova. \"fmri: resting state and arithmetic task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002422.v1.1.0.", + "[363] James R. Booth, GE Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, MA Stein, and Michael Tennekoon. \"working memory and reward in children with and without attention deficit hyperactivity disorder (adhd)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002424.v1.2.0.", + "[364] Damien Brevers, Chris Baeken, Antoine Bechara, Qinghua He, Pierre Maurage, Mathieu Petieau, Guillaume Sescousse, Claus Vögele, and Joel Billieux. \"increased brain reactivity to gambling unavailability is a marker of problem gambling\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002513.v1.0.0.", + "[365] Tamar Kolodny, Michael-Paul Schallmo, and Scott O. Murray. \"contrast response functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002522.v1.0.2.", + "[366] S. Jahfari, K.R. Ridderinkhof, A.G.E. Collins, T. Knapen, L.J. Waldorp, M.J. Frank, and J. Theeuwes. \"learning, inhibitory control, and perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s002543.v1.0.1.", + "[367] Suzanne Oosterwijk, Lukas Snoek, Mark Rotteveel, Lisa Feldman Barrett, and H. Steven Scholte. \"shared-states\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002547.v1.1.0.", + "[368] Laboratory for Computational Neurodiagnostics. \"protecting the aging brain (pagb)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002549.v1.0.1.", + "[369] Romain Quentin, Jean-Remi King, Etienne Sallard, Nathan Fishman, Ryan Thompson, Ethan Buch, and Leonardo" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 21 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cohen. \"differential brain mechanisms of selection and maintenance of information during working memory (meg data)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002550.v1.0.1.", + "[370] Tomas Knapen, Wietske van der Zwaag, and Daan van Es. \"cerebellum retinotopic mapping\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002574.v1.0.1.", + "[371] Arnaud Delorme and Scott Makeig. \"visual oddball task (256 channels)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs002578.v1.1.0.", + "[372] S. Zhang, W. Yoshida, H. Mano, T. Yanagisawa, K. Shibata, M. Kawato, and B. Seymour. \"cognitive control of sensory pain encoding in the pregenual anterior cingulate cortex. d1 - decoder construction in day 1, d2 - adaptive control in day 2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002596.v1.0.1.", + "[373] Lukas Snoek, Tinka Beemstermboer, and H. Steven Scholte. \"mb-epi-comparison\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002603.v1.0.0.", + "[374] Caroline Le Ster, Antonio Moreno, Franck Mauconduit, Vincent Gras, Ruediger Stirnberg, Benedikt A. Poser, Alexandre Vignaud, Evelyn Eger, Stanislas Dehaene, Florent Meyniel, and Nicolas Boulant. \"bids_dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002606.v1.1.0.", + "[375] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study1\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002608.v1.0.2.", + "[376] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study2\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002609.v1.0.3.", + "[377] Enzo Tagliazucchi. \"meditation interocepcion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002614.v1.0.0.", + "[378] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum. \"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002620.v1.0.0.", + "[379] N Eichert, KE Watkins, and project-specific co authors. \"project_larynx\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002634.v3.0.0.", + "[380] L.N. Van der Laan, C. Scholz, R.A. Poldrack, D.T.D. De Ridder, and A. Smidts. \"can we have a second helping? a replication study on the neurobiological mechanisms underlying self-control\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002643.v1.1.0.", + "[381] Mara Mather, Ringo Huang, David Clewett, Shawn Nielsen, Rico Velasco, Kristie Tu, Sophia Han, and Briana Kennedy. \"isometric exercise facilitates attention to salient events in women via the noradrenergic system\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002647.v1.0.1.", + "[382] Nathan M. Muncy and C. Brock Kirwan. \"the medial temporal lobe supports mnemonic discrimination for event duration\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002655.v1.0.1.", + "[383] Brian J. Coffey, Zachary D. Threlkeld, Andrea S. Foulkes, Yelena G. Bodien, and Brian L. Edlow. \"the language" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 21 + }, + { + "type": "page_number", + "text": "23860", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 21 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "network reemerges during recovery from severe traumatic brain injury\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002675.v1.0.0.", + "[384] Cheryl A. Olman and Kimberly B. Weldon. \"kung fu panda\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002684.v1.0.0.", + "[385] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"ibc\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs002685.v2.0.0.", + "[386] James R. Booth, Gillian E. Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, Mark A. Stein, and Michael Tennekoon. \"working memory and reward in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002687.v1.2.0.", + "[387] Kendrick Kay, Keith W. Jamison, Luca Vizioli, Ruyuan Zhang, Eshed Margalit, and Kamil Ugurbil. \"high-field 7t visual fmri datasets\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002702.v1.0.1.", + "[388] Christos Chatzichristos, Manuel Morante, Nikos Andreadis, Eleftherios Kofidis, Yiannis Kopsinis, and Sergios Theodoridis. \"smiley\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002711.v1.1.0.", + "[389] Sara Aurtenetxe, Nicola Molinaro, Doug Davidson, and Manuel Carreiras. \"numbers and letters\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002712.v1.0.1.", + "[390] Petra Vetter, Lukasz Bola, Lior Reich, Matthew Bennett, Lars Muckli, and Amir Amedi. \"avscenes_blind\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002715.v1.0.0.", + "[391] Liron Rozenkrantz, Reut Weissgross, Tali Weiss, Inbal Ravrebi, Idan Frumin, Sagit Shushan, Lior Gorodisky, Netta Reshef, Yael Holzman, Liron Pinchover, Yaara Endevelt-Shapira, Eva Mishor, Timna Soroka, Maya Finkel, Liav Tagania, Aharon Ravia, Ofer Perl, Edna Furman-Haran, Howard Carp, and Noam Sobel. \"unexplained repeated pregnancy loss is associated with altered perceptual and brain responses to men's body-odor\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002717.v1.0.1.", + "[392] Ian Daly, Nicoletta Nicolaou, Duncan Williams, Faustina Hwang, Alexis Kirke, Eduardo Miranda, and Slawomir J. Nasuto. \"a dataset recording joint eeg-fmri during affective music listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002725.v1.0.0.", + "[393] Jordi Sole-Casals, Josep M Serra-Grabulosa, Rafael Romero-Garcia, Gemma Vilaseca, Ana Adan, Nuria Vilaro, Nuria Vilaro, and Edward T. Bullmore. \"structural brain network of gifted children\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002726.v1.0.1.", + "[394] Francesca M M Citron, Cristina Cacciari, Jakob M Funcke, Chun-Ting Hsu, and Arthur Jacobs. \"emotive idioms\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002727.v1.0.2.", + "[395] Bi Zhu, Chuansheng Chen, Xuhao Shao, Wenzhi Liu, Zhi-fang Ye, Liping Zhuang, Li Zheng, Elizabeth F. Loftus, and Gui Xue. \"multiple interactive memory representations underlie the induction of false memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002731.v1.0.2.", + "[396] Nora C. Heinzelmann, Susanna C. Weber, and Philippe N. Tobler. OpenNeuro, 2020. doi: 10.18112/openneuro.ds002732.v1.0.0." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 22 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[397] Test. \"bids pilot project\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002733.v1.0.1.", + "[398] Andrea Pisauro, Elsa Fouragnan, C Retzler, and M Philiastides. \"evidence accumulation in value-based decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002734.v1.0.2.", + "[399] Jonathan D Power, Benjamin Silver, Melanie Silverman, Eliana Ajodan, Dienke Bos, and Rebecca M. Jones. \"headmold\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002735.v1.0.2.", + "[400] Joset A. Etzel and Todd S. Braver. \"multibandcftests\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002737.v1.0.1.", + "[401] Sabina Gherman and Marios G. Philiastides. \"simultaneous eeg-fmri - confidence in perceptual decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002739.v1.0.0.", + "[402] N Furl, F Begum, J Sulik, F Pizzorni Ferrarese, S Jans, and C Wooley. \"animated caricatures fmri study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002741.v1.0.2.", + "[403] AR. \"test1\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002743.v1.0.1.", + "[404] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"resting state with closed eyes for patients with depression and healthy participants\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002748.v1.0.5.", + "[405] Chao-Gan Yan. \"yandatabids\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002750.v1.0.1.", + "[406] Dillan J. Newbold, Timothy O. Laumann, Catherine R. Hoyt, Jacqueline M. Hampton, David F. Montez, Ryan V. Raut, Mario Ortega, Anish Mitra, Ashley N. Nielsen, Derek B. Miller, Babatunde Adeyemo, Annie L. Nguyen, Kristen M. Scheidter, Aaron B. Tanenbaum, Andrew N. Van, Scott Marek, Bradley L. Schlaggar, Alexandre R. Carter, Deanna J. Greene, Evan M. Gordon, Marcus E. Raichle, Steven E. Petersen, Abraham Z. Snyder, and Nico U.F. Dosenbach. \"cast-induced plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002766.v3.0.2.", + "[407] Damien Brevers, Claus Vögele, Pierre Maurage, Guillaume Sescousse, Chris Baeken, and Joel Billieux. \"brain mechanisms underlying episodic future thinking of sustainable behaviors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002770.v2.0.0.", + "[408] Zhifang Ye and Gui Xue. \"retrieval practice facilitates memory updating by enhancing and differentiating medial prefrontal cortex representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002773.v1.0.0.", + "[409] Eva Berlot, Nicola Popp, and Jörn Diedrichsen. \"motor sequence learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002776.v1.0.2.", + "[410] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002785.v2.0.0.", + "[411] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002790.v2.0.0." + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 22 + }, + { + "type": "page_number", + "text": "23861", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 22 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[412] Minjae Kim, Peter Mende-Siedlecki, Stefano Anzellotti, and Liane Young. \"tom following strong vs weak priors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002793.v1.0.1.", + "[413] Jiefeng Jiang, Anthony D. Wagner, and Tobias Egner. \"elife39497\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002797.v1.0.2.", + "[414] Thompson WH*, Nair R*, Oya H*, Esteban O, Shine JM, Petkov CI, Poldrack RA, Howard M, Adolphs R†, and †corresponding author *equally contributing. \"human es-fmri resource: Concurrent deep-brain stimulation and whole-brain functional migraine\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002799.v1.0.4.", + "[415] C.R. Bowman, T. Iwashita, and D. Zeithamova. \"model-based fmri reveals co-existing specific and generalized concept representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002813.v1.0.0.", + "[416] Fatemeh Ebrahiminia, Morteza Mahdiani, and Seyed-Mahdi Khaligh-Razavi. \"a multimodal neuroimaging dataset to study spatiotemporal dynamics of visual processing in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002814.v1.3.0.", + "[417] Sangil Lee, Trishala Parthasarathi, and Joseph Kable. \"prospection\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso02835.v1.0.1.", + "[418] Sarah Aliko, Jiawen Huang, Florin Gheorghiu, Stefanie Meliss, and Jeremy I Skipper. \"naturalistic neuroimaging database\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002837.v2.0.0.", + "[419] Sarah Schwettmann, Joshua B. Tenenbaum, and Nancy Kanwisher. \"intuitive physics with fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002841.v1.0.1.", + "[420] Omar Maxim. \"mriquality\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002848.v1.0.1.", + "[421] Caroline Landelle, Jean-Luc Anton, Bruno Nazarian, Julien Sein, Ali Gharbi, Olivier Felician, and Anne Kavounoudias. \"aging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002872.v1.3.0.", + "[422] Christian Roming, Karl Koschutnig, Andreas Fink, and ... \"creative direccion making in soccer\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002878.v2.0.0.", + "[423] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"brain correlates of multisensory lexical processing in adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds002879.v1.1.1.", + "[424] James R. Booth, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, and Jérôme Prado. \"brain development of deductive reasoning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002886.v1.1.0.", + "[425] Agnès Job, Chloé Jaroszynski, Anne Kavounoudias, Assia Jaillard, and Chantal Delon-Martin. \"resting state fmri study of non bothersome tinnitus\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002896.v1.0.0.", + "[426] Sharna D. Jamadar, Phillip G.D. Ward, Thomas G. Close, Alex Fornito, Malin Premaratne, Kieren O'Brien, Daniel" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 23 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Staeb, Zhaolin Chen, N. Jon Shah, and Gary F. Egan. \"monash rspet-mr\". OpenNeuro, 2024. doi: 10.18112/openneuro.d002898.v1.4.2.", + "[427] Christophe Pallier and Laurent Bonnasse-Gahot. \"simpcomp_rawdata\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002905.v1.0.1.", + "[428] Rafael Garcia-Dias. \"neuroharmony sample example\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002936.v1.0.0.", + "[429] Bart Aben, Cristian Buc Calderon, Eva Van den Bussche, and Tom Verguts. \"task-dependent effort-induced connectivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002938.v1.0.1.", + "[430] DC Monroe, RS Blumenfeld, SL Small, and DB Keator. \"head-to-ball impacts in collegiate soccer players\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002940.v1.0.1.", + "[431] Ruiqing Zhang. \"narrative listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002979.v1.0.0.", + "[432] Sangil Lee, Nicole Cooper, and Joseph Kable. \"delay discounting bidding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002989.v1.0.0.", + "[433] Gabriel Pelletier, Nadav Aridan, Lesley K Fellows, and Tom Schonberg. \"fribblesfmri_object-value-construction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002994.v1.0.3.", + "[434] Jason A Avery, Alexander G Liu, Cameron Riddell, John E Ingelholm, Stephen J Gotts, and Alex Martin. \"taste quality representation in the human brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002995.v1.0.1.", + "[435] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"two sessions of resting state with closed eyes for patients with depression in treatment course (nfb, cbt or no treatment groups)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003007.v1.0.1.", + "[436] Colin Hawco, Erin Dickie, Gabrielle Herman, Jessica Turner, Miklos Argyelan, Philipp Homan, Anil Malhotra, Robert Buchanan, and Aristotle Voineskos. \"social processes initiative in neurobiology of the schizophrenia(s) traveling human phantoms\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003011.v1.2.3.", + "[437] M. Martinez-Saito, R. Konovalov, M. A. Piradov, A. Shestakova, B. Gutkin, and V. Klucharev. \"action-in-auctions\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003012.v1.0.4.", + "[438] Amanda LeBel, Lauren Wagner, Shailee Jain, Aneesh Adhikari-Desai, Bhavin Gupta, Allyson Morgenthal, Jerry Tang, Lixiang Xu, and Alexander G. Huth. \"an fmri dataset during a passive natural language listening task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms003020.v3.0.0.", + "[439] Diego Angeles-Valdez, Jalil Rasgado-Toledo, Viviana Villicana, Alan Davalos-Guzman, Alfonso Fajardo-Valdez, Cristina Almanza, Ruth Alcala-Lozano, and Eduardo A. Garza-Villarreal. \"sudmex_tms\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003037.v2.1.0.", + "[440] Karen T. Navarro, Cheryl A. Olman, and Kimberly B. Weldon. \"dichoptic magno/parvo\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003043.v1.0.0." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 23 + }, + { + "type": "page_number", + "text": "23862", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 23 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[441] Krueger O and Klose U. \"dti data from 'fiber architecture in the ventromedial striatum and its relation with the bed nucleus of the stria terminalis\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003047.v1.0.0.", + "[442] Robin Carhart-Harris et al. \"neural correlates of the lsd experience revealed by multimodal neuroimaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003059.v1.0.0.", + "[443] James R. Booth, Marisa N. Lytle, Rachna Mureja, and Jérôme Prado. \"brain correlates of deductive reasoning in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003076.v1.0.1.", + "[444] Philippe DOMENECH, Sylvain RHEIMS, and Etienne KOECHLIN. \"probe ieeg\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms003078.v1.0.0.", + "[445] Jonathan Cote and Etienne de Villers-Sidani. \"auditory cortex mapping dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003082.v1.0.1.", + "[446] James R. Booth, Amy S. Desroches, Marisa N. Lytle, Rucha Mehta, Jennifer E. Minas, Rachna Mutreja, Jérôme Prado, and Hongchuan Zhang. \"brain correlates of math processing in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003083.v1.0.1.", + "[447] Matthew Sachs, Assal Habibi, Antonio Damasio, and Jonas Kaplan. \"temporal dynamics of emotional music\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003085.v1.0.0.", + "[448] Noam Saadon-Grosman, Shahar Arzy, and Yonatan Loewenstein. \"somatosensory phase-encoded bilateral full-body light touch stimulation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003089.v1.0.1.", + "[449] Jeffrey N. Chiang, Yujia Peng, Hongjing Lu, Keith J. Holyoak, and Martin M. Monti. \"analogical reasoning sequential design fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003094.v1.0.0.", + "[450] Y.C. Leong, J. Chen, R. Willer, and J. Zaki. \"polarization dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003095.v1.0.0.", + "[451] D Chung, MA Orloff, N Lauharatanahirun, B King-Casas, and Chiu PH. \"valuation of peers' safe choices is associated with substance-naivete in adolescents\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps003096.v1.0.2.", + "[452] Anders Lillevik Thorsen. \"abcd protocol for bcbp\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms003098.v1.0.0.", + "[453] HanShin Jo, Chiu-Yueh Chen, Der-Yow Chen, Ming-Heng Weng, and Chun-Chia Kung. \"consensus-seeking and conflict-resolving: an fmri study on college couples' shopping interaction\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003103.v1.0.1.", + "[454] Lauri Parkkonen, Stefan Appelhoff, Alexandre Gramfort, Mainak Jas, and Richard Höchenberger. \"mne-somatodata-bids (anonymous)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003104.v1.0.0.", + "[455] Jia-Xi Wang, Jin-Ying Zhuang, Lulu Fu, and Lei Qin. \"how ovarian hormones influence the behavioral activation and inhibition system through the dopamine pathway\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003114.v1.0.1.", + "[456] Saloni Krishnan and Kate Watkins. \"bold verb generation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003145.v1.0.2." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 24 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[457] Arafat Angulo-Perkins and Luis Concha. \"vocal task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003146.v1.0.2.", + "[458] Adam Hampshire and Eyal Soreq. \"neuroimaging evidence for network sampling theory of human intelligence\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003148.v1.0.1.", + "[459] Alexandra Ycaza Herrera, Padideh Nasseri, Ricardo Velasco, Sophia Faude, Jessica D. White, Philipp C. Opitz, Ringo Huang, Kristie Tu, and Mara Mather. \"stress-associated brain activation across the hormonal contraceptive cycle\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds003151.v2.0.1.", + "[460] Chang-Hao Kao, Sangil Lee, Joshua I. Gold, and Joseph W. Kable. \"changepoint fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s003170.v2.0.0.", + "[461] Sivayini Kandeepan, Jorge Rudas, Francisco Gomez, Bobby Stojanoski, Sreeram Valluri, Adrian Mark Owen, Lorina Naci, Andrea Soddu, and Emily Sophia Nichols. \"modeling an auditory stimulated brain under altered states of consciousness using the generalized ising model\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003171.v2.0.1.", + "[462] Simon M. Ceh, Sonja Annerer-Walcher, Karl Koschutnig, Christof Korner, and Mathias Benedek. \"neurophysiological indicators of internal attention: An fmri-eye-tracking coregistration study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003176.v2.0.1.", + "[463] Stefano Moia, Eneko Urñuela, Vicente Ferrer, and César Caballero-Gaudes. \"euskalibur\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003192.v1.0.1.", + "[464] Kenshu Koiso, Sebastian Dresbach, Christopher J Wiggins, Omer Faruk Gulban, Anna Katharina Mueller, Miriam Heynckes, Yuhui Chai, Yoichi Miyawaki, Benedikt A Poser, and Renzo Huber. \"acquisition and processing methods of whole-brain layer-fmri vaso and bold: The kenshu dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003216.v3.0.11.", + "[465] Meir Meshulam, Liat Hasenfratz, Hanna Hillman, Yun-Fei Liu, Mai Nguyen, Kenneth A. Norman, and Uri Hasson. \"think like an expert\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003233.v1.2.1.", + "[466] Livia Tomova, Kimberly L. Wang, Todd Thompson, Gillian A. Matthews, Kay M. Tye, and Rebecca Saxe. \"mri data of 40 adult participants in response to a cue induced craving task following food fasting, social isolation and baseline (within-subject design)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003242.v1.0.0.", + "[467] James Antony, Thomas Hartshorne, Ken Pomeroy, Todd Gureckis, Uri Hasson, Samuel McDougle, and Kenneth Norman. \"behavioral, physiological, and neural signatures of surprise during naturalistic sports viewing\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003338.v1.1.0.", + "[468] Jason A Avery, Alexander G Liu, John E Ingeholm, Stephen J Gotts, and Alex Martin. \"tasting pictures: Viewing images of foods evokes taste-quality-specific activity in gustatory insular cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003340.v1.0.4.", + "[469] Ethan Knights, Courtney Mansfield, Diana Tonin, Janak Saada, Fraser W. Smith, and Stephanie Rossit. \"hand-" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 24 + }, + { + "type": "page_number", + "text": "23863", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 24 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "selective visual regions represent how to grasp 3d tools for use: brain decoding during real actions\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003342.v1.0.0.", + "[470] Kelsey McDonald, William Broderick, Scott Huettel, and John Pearson. \"penaltykik.02\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003345.v1.0.2.", + "[471] Eduardo A. Garza-Villarreal, Jorge Julio Gonzalez Olvera, Thania Balducci, Diego Angeles Valdez, Alely Valencia, and Jalil Rasgado. \"sudmex Conn: The mexican dataset of cocaine use disorder patients.\" OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003346.v1.1.2.", + "[472] Mikkel Schöttner and Andreas Jansen. \"empatom with ppi-r, tas, and eq\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003354.v1.0.1.", + "[473] Richard Huskey, Justin Robert Keene, Shelby Wilcox, Xuanjun (Jason) Gong, Robyn Adams, and Christina J. Najera. \"brain dynamics during flow experiences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003358.v1.0.0.", + "[474] Samuel B. Snider, Yelena G. Bodien, Aina Frau-Pascual, Marta Bianciardi, Andrea S. Foulkes, and Brian L. Edlow. \"ascending arousal network connectivity during recovery from traumatic coma\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003367.v1.0.0.", + "[475] Heath R Pardoe, Samantha P Martin, Allan George, and Orrin Devinsky. \"estimation of in-scanner head pose changes during structural mri using a convolutional neural network trained on eye tracker video\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003381.v1.0.1.", + "[476] Sharna Jamadar, Shenjun Zhong, Phillip Ward, Alexandra Carey, Richard McIntyre, Alex Fornito, Malin Premaratne, N Jon Shah, Kieran O'Brien, Daniel Stab, Zhaolin Chen, and Gary Egan. \"monash vis-fpet-fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003382.v1.5.0.", + "[477] Nicolas Zilber, Philippe Ciuciu, Alexandre Gramfort, Leila Azizi, and Virginie van Wassenhove. \"neurospin hmt+localizer data (meg & amri)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003392.v1.0.4.", + "[478] Sharna D Jamadar, Emma X Liang, Phillip GD Ward, Alexandra Carey, Richard McIntyre, Zhaolin Chen, Shenjun Zhong, and Gary F Egan. \"monash dacra fpet-fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003397.v1.2.3.", + "[479] Jungtak Park, Karolina Janacsek, Dezso Nemeth, and Hyeon-Ae Jeon. \"asrt (alternating serial reaction time)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003401.v1.0.1.", + "[480] Brian P Keane, Deanna M Barch, Ravi Mill, Steven M Silverstein, Bart Krekelberg, and Michael W Cole. \"brain network mechanisms of visual shape completion\". Open-Neuro, 2025. doi: 10.18112/openneuro.ds003404.v1.0.5.", + "[481] Maria Geraldine Veldhuizen, Michael C. Farruggia, Xiao Gao, Yuko Nakamura, Barry G. Green, and Dana Small. \"identification of an amygdala-thalamic circuit that acts as a central gain mechanism in taste perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003424.v1.0.0.", + "[482] Steven Greening, Tae-Ho Lee, Laurent Gregoire, Lauryn Burleigh, Tyler Robinson, Xinrui Jiang, Mara Mather, and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 25 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jonas Kaplan. \"mental imagery can generate and regulate acquired differential fear conditioned reactivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003425.v1.0.2.", + "[483] Tomoyasu Horikawa and Yukiyasu Kamitani. \"attentionally modulated subjective images reconstructed from brain activity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003430.v1.2.0.", + "[484] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 localizer\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003433.v1.0.1.", + "[485] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 main experiment\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003434.v1.0.1.", + "[486] Sandra Arbula, Elisabetta Pisanu, and Raffaella I. Rumiati. \"agreeableness personality trait and social information encoding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003436.v1.0.0.", + "[487] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain - diet-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003437.v1.0.2.", + "[488] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant1 suboptimal block designs\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms03438.v1.0.0.", + "[489] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant2 suboptimal slow event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003439.v1.0.0.", + "[490] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant3 caffeine\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003440.v1.0.0.", + "[491] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusicac. \"newbi4fmri2020 variant6 adaptation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003443.v1.0.0.", + "[492] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant7 prosopagnosia\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003444.v1.0.0.", + "[493] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant8 resting state\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003445.v1.0.0.", + "[494] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusiac. \"newbi4fmri2020 variant9 7t\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003446.v1.0.0.", + "[495] Joset A. Etzel, Mitch Jeffers, and Todd S. Braver. \"dmcc13benchmark\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003452.v1.0.1.", + "[496] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Lithen," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 25 + }, + { + "type": "page_number", + "text": "23864", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 25 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, bolus-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003453.v1.0.4.", + "[497] Andrea Stocco, Chantel S. Prat, and Lauren K. Graham. \"individual differences in fluid reasoning and rapm-like problem solving\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003454.v1.0.1.", + "[498] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, case-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003455.v1.0.2.", + "[499] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., and Todd R. Parrish. \"component processes of word reading in adults and children\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003459.v1.0.2.", + "[500] Todd S. Braver, Alexander Kizhner, Rongxiang Tang, Michael C. Freund, and Joset A. Etzel. \"dmcc55b\". Open-Neuro, 2025. doi: 10.18112/openneuro.ms003465.v1.0.7.", + "[501] Jiajia Yang, Peter J. Molfese, Yinghua Yu, Daniel A. Handwerker, Gang Chen, Paul A. Taylor, Yoshimichi Ejima, Jinglong Wu, and Peter A. Bandettini. \"haptic three-dimensional curved surface exploration fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003466.v1.1.", + "[502] Ekaterina V. Pechenkova, Yana R. Panikratova, Maria A. Fomina, Alena D. Rumshiskaya, Darya A. Bazhenova, Liudmila A. Makovskaya, Irina S. Lebedeva, and Valentin E. Sinitsyn. \"speech disfluencies: Neurophysiological aspect in normal population\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003469.v1.0.0.", + "[503] Anja Ischebeck, Hannah Hiebel, Joe Miller, Margit Höfler, Iain D. Gilchrist, and Christof Körner. \"target processing in overt serial visual search involves the dorsal attention network: A fixation-based event-related fmri study\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003470.v2.0.0.", + "[504] J. Rasgado-Toledo, F. Lizcano-Cortés, V. Olalde-Mathieu, M. Zamora-Ursulo, G. Licea-Haquet, A. Carillo-Pena, E. Navarrete, A.* Reyes-Aguilar, and M Giordano. \"pragmatic language\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003481.v1.0.3.", + "[505] Brian Silston, Toby Wise, Song Qi, Xin Sui, Peter Dayan, and Dean Mobbs. \"foraging in competitive and hazardous environments\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003484.v1.0.0.", + "[506] Eva R. Pool, David Munoz Tord, Sylvain Delplanque, Yoann Stussi, Donato Cereghetti, Patrik Vuilleumier, and David Sander. \"differential contributions of ventral striatum subregions in the motivational and hedonic components of the affective response to reward\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003487.v2.0.0.", + "[507] Lukas Snoek and Steven Scholte. \"dstreamupgrade\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003495.v1.0.0." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 26 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[508] Kate Nussenbaum and Catherine A. Hartley. \"developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03499.v1.0.1.", + "[509] David Pascucci, Sebastien Tourbier, Joan Rue-Queralt, Margherita Carboni, Patric Hagmann, and Gijs Plomp. \"vepcon: Source imaging of high-density visual evoked potentials with multi-scale brain parcellations and connectomes\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003505.v1.1.2.", + "[510] Chuanji Gao, Christine E. Weber, Douglas H. Wedell, and Svetlana V. Shinkareva. \"fmri: Audiovisual valence congruence\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003507.v1.0.1.", + "[511] Mikael Novén, Hampus Olsson, Gunther Helms, Markus Nilsson, Merle Horne, and Mikael Roll. \"language learning aptitude dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003508.v1.0.0.", + "[512] Adrian W. Gilmore, Alina Quach, Sarah E. Kalinowski, Estefanía I. Gonzalez-Araya, Stephen J Gotts, Daniel L. Schacter, and Alex Martin. \"evidence supporting a timelimited hippocampal role in retrieving autobiographical memories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003511.v1.1.2.", + "[513] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study2\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso03521.v2.2.0.", + "[514] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study1\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03524.v1.0.0.", + "[515] Benjamin Risk, Raphiel Murden, Junjie Wu, Mary Beth Nebel, Arun Venkataraman, Zhengwu Zhang, and Deqiang Qiu. \"which multiband factor should you choose for your resting-state fmri study? the emory multiband dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003540.v1.0.1.", + "[516] Kshipra Gurunandan, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adult language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003542.v1.0.0.", + "[517] Kshipra Gurunandan, Jaione Arnaez-Telleria, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adolescent language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003545.v1.0.0.", + "[518] Isaac David, Victor Olalde-Mathieu, Ana Y. Martínez, Lluviana Rodríguez-Vidal, and Fernando A. Barrios. \"emotion category and face perception task optimized for multivariate pattern analysis\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms03548.v1.0.1.", + "[519] Renee M Visser, H Steven Scholte, Tinka Beemsterboer, and Merel Kindt. \"visser, scholte, beemsterboer, & kindt (2013) nature neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003550.v1.0.2.", + "[520] Renee M Visser, H Steven Scholte, and Merel Kindt. \"visser, scholte & kindt (2011) journal of neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003553.v1.0.2." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 26 + }, + { + "type": "page_number", + "text": "23865", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 26 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[521] Renee M Visser, Anna E. Kunze, Biance Westhoff, H Steven Scholte, and Merel Kindt. \"visser et al. 2015 psychoneuroendocrinology\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003554.v1.0.3.", + "[522] Falk Luesebrink, Hendrik Mattern, Renat Yakupov, Julio Acosta-Cabronero, Mohammad Ashtarayeh, Steffen Oeltze-Jafra, and Oliver Speck. \"data from: Comprehensive ultrahigh resolution whole brain in vivo mri dataset as a human phantom\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003563.v1.1.0.", + "[523] Lucrezia Liuzzi, Katharine Chang, Hanna Keren, Charles Zheng, Dipta Saha, Dylan Nielson, and Argyris Stringaris. \"mood induction in mdd and healthy adolescents\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds003568.v1.0.4.", + "[524] Jeahong Kim, Jeyoung Jung, and Kichun Nam. \"transposition confusability during visual word recognition\". Open-Neuro, 2021. doi: 10.18112/openneuro.ms003569.v1.0.0.", + "[525] Virginie Sterpenich, Mojca KM van Schie, Maximilien Catsiyannis, Avinash Ramyead, Stephen Perrig, Hee-Deok Yang, Dimitri Van De Ville, and Sophie Schwartz. \"reward biases spontaneous neural reactivation during sleep\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003574.v1.0.2.", + "[526] R. Nathan Spreng, Roni Setton, Udi Alter, Benjamin N. Cassidy, Bri Darboh, Elizabeth DuPre, Karin Kantarovich, Amber W. Lockrow, Laetitia Mwilambwe-Tshilobo, WenMing Luh, Prantik Kundu, and Gary R. Turner. \"neurocognitive aging data release with behavioral, structural, and multi-echo functional mri measures\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso03592.v1.0.13.", + "[527] Xingyu Liu, Yuxuan Dai, Hailun Xie, and Zonglei Zhen. \"forrestgump-meg\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003633.v1.0.4.", + "[528] Heath R Pardoe and Samantha P Martin. \"in-scanner head motion and structural covariance networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003639.v1.0.0.", + "[529] Michele Svanera and Lars Muckli. \"cerebrum-7t: Fast and fully-volumetric brain segmentation of 7 tesla mr volumes\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003642.v1.1.0.", + "[530] Jixing Li, John Hale, and Christophe Pallier. \"le petit prince: A multilingual fmri corpus using ecological stimuli\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds003643.v2.0.7.", + "[531] Ph.D. Anna Manelis, Ph.D. David A.A. Baranger, B.S. Skye Satz, M.A. Rachel Ragozzino, Ph.D. Satish Iyengar, M.D. Holly A. Swartz, and Ph.D. Yaroslav O. Halchenko. \"cortical myelin measured by the t1w/t2w ratio in individuals with depressive disorders and healthy controls\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003653.v1.0.0.", + "[532] Arielle Tambini, Jacob Miller, Anastasia Kiyonaga, and Mark D'Esposito. \"scantrain\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs003659.v2.0.4.", + "[533] Mohamed Abdelhack and Yukiyasu Kamitani. \"sharpening of hierarchical visual feature representations of blurred images\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003661.v1.0.0." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 27 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[534] Matan Mazor, Karl J. Friston, and Stephen M. Fleming. \"confidence in detection and discrimination\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003669.v1.0.0.", + "[535] Kangjoo Lee, Corey Horien, David O'Connor, Bronwen Garand-Sheridan, Fuyuze Tokoglu, Dustin Scheinost, Evelyn MR Lake, and Todd R Constable. \"yale resting state fmri/pupillometry: Arousal study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003673.v2.0.1.", + "[536] Giacomo Ariani, J Andrew Pruszynski, and Jorn Diedrichsen. \"finger sequence planning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003684.v1.0.0.", + "[537] Julia Berezutskaya, Mariska J. Vansteensel, Erik J. Aarnoutse, Zachary V. Freudenburg, Giovanni Piantoni, Mariana P. Branco, and Nick F. Ramsey. \"open multimodal ieeg-fmri dataset from naturalistic stimulation with a short audiovisual film\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003688.v1.0.7.", + "[538] Javier Ortiz-Tudela, Johanna Bergman, Matt Bennett, Isabelle Ehrich, Lars Muckli, and Yee Lee Shing. \"feedbes - feedback signals from episodic and semantic memories.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003691.v1.0.0.", + "[539] Shan Xu, Xingyu Liu, Jorge Almeida, and Dietmar Heinke. \"action-related object pairs - fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003696.v1.0.0.", + "[540] Takuya Ito, Emily Winfield, Nicole Lalta, Kaustubh R. Kulkarni, Douglas H. Schultz, Ravi D. Mill, Richard H. Chen, Levi Solomyak, and Michael W. Cole. \"concrete permuted rule operations\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003701.v1.0.1.", + "[541] Guo Wanjia, Serra E. Favila, Ghootae Kim, Robert J. Moltor, and Brice A. Kuhl. \"abrupt hippocampal remapping signals resolution of memory interference.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003707.v1.0.0.", + "[542] Hanna Keren, Katharine Chang, Aria Vitale, Chris C. Camp, Robb B Rutledge, Dylan M. Nielson, and Argyris Stringaris. \"nimh-compypsych mmi\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03709.v1.0.0.", + "[543] Daniel K. Bjornn, Julie Van, and C. Brock Kirwan. \"the contributionsof eye gaze fixations and target-lure similarity to behavioral and fmri indices of pattern separation and pattern completion\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003711.v1.0.0.", + "[544] Avijit Chowdhury. \"test\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003714.v1.0.1.", + "[545] Li J., Curley W. H., Guerin B., Dougherty D. D., Dalca A. V., Fischl B., Horn A., and Edlow B. L. \"subcortical dmn functional connectivity\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003716.v1.0.0.", + "[546] Peelle JE, Spehar B, Jones MS, McConkey S, Myerson J, Hale S, Sommers MS, and Tye-Murray N. \"visual and audiovisual speech perception associated with increased functional connectivity between sensory and motor regions\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003717.v1.1.0.", + "[547] Tomoya Nakai, Naoko Koide-Majima, and Shinji Nishimoto. \"music genre fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003720.v1.0.1." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 27 + }, + { + "type": "page_number", + "text": "23866", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 27 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[548] Renee M Visser, Richard N Henson, and Emily A Holmes. \"a naturalistic paradigm to investigate post-encoding neural activation patterns in relation to subsequent voluntary and intrusive recall of distressing events\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003721.v1.0.1.", + "[549] David V. Smith, Rita M. Ludwig, Jeffrey B. Dennison, Crystal Reeck, and Dominic S. Fareri. \"an fmri dataset on social reward processing and decision making in younger and older adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003745.v2.1.1.", + "[550] michael barnett, geoffrey aguirre, and david brainard. \"If contrast\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003752.v1.0.0.", + "[551] Kenji Kobayashi, Sangil Lee, Alexandre L. S. Filipowicz, Kara D. McGaughey, Joseph W. Kable, and Matthew R. Nassar. \"dynamic representation of the subjective value of information\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003758.v1.0.2.", + "[552] Olga R. Dobrushina, Larisa A. Dobrynina, Galina A. Arina, Maria V. Gubanova, Anastasia V. Belopasova, Evgenia S. Novikova, Anastasia D. Suslina, Elena I. Kremneva, Ekaterina V. Pechenkova, and Marina V. Krotenkova. \"interception during aging: The heartbeat detection task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003763.v1.0.5.", + "[553] Zhenzhong Gan, Suiping Wang, and Gangyi Feng. \"functional magnetic resonance imaging data for the neural dynamics underlying the acquisition of distinct auditory categories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003764.v1.0.5.", + "[554] Yameng Gu, Feng Han, Lucas E. Sainburg, Margeaux M. Schade, and Xiao Liu. \"simultaneous eeg and fmri signals during sleep from humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003768.v1.0.11.", + "[555] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"depression treatment by medial prefrontal real-time fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms003770.v1.2.1.", + "[556] Joseph T. McGuire, Matthew R. Nassar, Joshua I. Gold, and Joseph W. Kable. \"functionally dissociable influences on learning rate in a dynamic environment\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003772.v1.0.1.", + "[557] Redouane Jamil, Franck Mauconduit, Caroline Le Ster, Philipp Ehses, Benedikt A. Poser, Alexandre Vignaud, and Nicolas Boulant. \"temporal snr optimization through rf coil combination in fmri: The more, the better? - dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003777.v1.0.1.", + "[558] Yufei Zhao, Avi Chanales, and Brice Kuhl. \"adaptive memory distortions are predicted by feature representations in parietal cortex\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003778.v1.0.0.", + "[559] Nadav Aridan, Nicholas J. Malecek, Russell A. Poldrack, and Tom Schonberg. \"neural correlates of effort-based valuation under risk\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003782.v1.0.1.", + "[560] Marc Himmelberg, Jan Kurzawski, Noah Benson, Denis Pelli, Marisa Carrasco, and Jon Winawer. \"nyu retinotopy" + ], + "bbox": [ + 91, + 90, + 480, + 902 + ], + "page_idx": 28 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003787.v1.0.1.", + "[561] Xuhao Shao, Chuansheng Chen, Elizabeth F. Loftus, Gui Xue, and Bi Zhu. \"dynamic changes in neural representations underlie the repetition effect on false memory\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003789.v2.0.0.", + "[562] Eva Mishor, Daniel Amir, Tali Weiss, Danielle Honigstein, Aharon Weissbrod, Ethan Livne, Lior Gorodisky, Shiri Karagach, Aharon Ravia, Kobi Snitz, Diyala Karawani, Rotem Zirler, Reut Weissgross, Timna Soroka, Yaara Endevelt-Shapira, Shani Agron, Liron Rozenkrantz, Netta Reshef, Edna Furman-Haran, Heinz Breer, Joerg Strotmann, Tatsuya Uebi, Mamiko Ozaki, and Noam Sobel. \"sniffing the human body-volatile hexadecanal blocks aggression in men but triggers aggression in women\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003791.v1.0.0.", + "[563] Dorit Kliemann, Ralph Adolphs, Tim Armstrong, Paola Galdi, David A. Kahn, Tessa Rusch, A. Zeynep Enkavi, Deuhua Liang, Steven Lograsso, Wenying Zhu, Rona Yu, Remya Nair*, Lynn K. Paul*, J. Michael Tyszka*, and * Joint senior authorship. \"caltech conte center - a multimodal data resource for exploring social cognition and decision-making\". OpenNeuro, 2021. doi: 10.18112/openneuro.dss003798.v1.0.5.", + "[564] Hyun Joo Yoo, Kaoru Nashiro, Jungwon Min, Christine Cho, Noah Mercer, Shelby L. Bachman, Padideh Nasseri, Shubir Dutt, Shai Porat, Paul Choi, Yong Zhang, Vardui Grigoryan, Tiantian Feng, Julian F. Thayer, Paul Lehrer, Catie Chang, Jeffrey A. Stanley, Elizabeth Head, Jeremy Rouanet, Vasilis Z. Marmarelis, Shrikanth Narayanan, Jessica Wisnowski, Daniel A. Nation, and Mara Mather. \"heart rate variability biofeedback training and emotion regulation\". OpenNeuro, 2025. doi: 10.18112/openneuro.d003823.v1.3.5.", + "[565] Michal Rafal Zareba, Magdalena Fafrowicz, Tadeusz Marek, Ewa Beldzik, Halszka Oginska, and Aleksandra Domagalik. \"structural (t1) images of 136 young healthy adults; study of effects of chronotype, sleep quality and daytime sleepiness on brain structure\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003826.v3.0.1.", + "[566] Kevin Fialkowski, Ivan Messias, and Keith A. Bush. \"cognitive control theoretic mechanisms of real-time fmriguided neuromodulation (ctm)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003831.v1.0.0.", + "[567] Matteo Visconti di Oleggio Castello, James V. Haxby, and M. Ida Gobbini. \"shared neural codes for visual and semantic information about familiar faces in a common representational space\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003834.v1.0.2.", + "[568] Alyssa H Sinclair and Morgan D Barense. \"pe-update\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003835.v1.0.2.", + "[569] Flavia Mancini, Suyi Zhang, and Ben Seymour. \"learning the statistics of pain\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs003836.v1.0.0.", + "[570] van Blooijs D., Demuru M., Zweiphenning W, Hermes D., Leijten F., and Zijlmans M. \"dataset clinical epilepsy ieeg" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 28 + }, + { + "type": "page_number", + "text": "23867", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 28 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to bids - respect_longterm_jeeg\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003848.v1.0.3.", + "[571] Austin L. Boroshok, Anne T. Park, Panagiotis Fotiadis, Gerardo H. Velasquez, Ursula A. Tooley, Katrina R. Simon, Jasmine C.P. Forde, Lourdes M. Delgado Reyes, M. Dylan Tisdall, Dani S. Bassett, Emily A. Cooper, and Allyson P. Mackey. \"individual differences in frontoparietal plasticity in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003849.v1.0.0.", + "[572] S.R. Ashby and D. Zeithamova. \"category-biased neural representations form spontaneously during learning that emphasizes memory for specific instances\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003851.v1.0.1.", + "[573] Leili Mortazavi, Tara Srirangarajan, Tiago Bortolini, Jorge Moll, and Brian Knutson. \"monetary incentive delay task\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003858.v1.0.1.", + "[574] Christopher N. Wahlheim, Alexander P. Christensen, Brittany S. Cassidy, and Zachariah M. Reagh. \"resting-state for 34 younger and 28 older adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003871.v1.0.2.", + "[575] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003872.v1.0.0.", + "[576] Ted Turesky, Talat Shama, Shahriah Hafiz Kakon, Rashidul Haque, Nazrul Islam, Amala Someshwar, Borjan Gagoski, William Petri, Charles Nelson, and Nadine Gaab. \"brain-morphometry diminishedgrowthBeanstudy_2021\". OpenNeuro, 2022. doi: 10.18112/openneuro.d003877.v1.1.1.", + "[577] Wen Wen, Yue Wang, jiawei Zhou, Sheng He, Xinghuai Sun, Hong Liu, Chen Zhao, and Peng Zhang. \"loss and enhancement of layer-selective signals in geniculostriate and corticotectal pathways of adult human amblyopia\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003892.v5.0.0.", + "[578] Philippe Poulin, Guillaume Thead, Pierre-Marc Jodoin, and Maxime Descoteaux. \"tractoinferno: A large-scale, open-source, multi-site database for machine learning dmri tractography\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003900.v1.1.1.", + "[579] J. Pesnot Lerousseau, C. Parise, MO. Ernst, and V. van Wassenhove. \"multisensory correlation detector\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003922.v1.0.1.", + "[580] Ning Mei, Roberto Santana, and David Soto. \"informative neural representations of unseen contents during higher-order processing in human brains and deep artificial networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003927.v1.0.1.", + "[581] Tommaso Di Noto, Guillaume Marie, Sebastien Tourbier, Yasser Aleman-Gómez, Oscar Esteban, Guillaume Saliou, Meritxell Bach Cuadra, Patric Hagmann, and Jonas Richiardi. \"lausanne_tof-mra_aneurysm_cohort\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003949.v1.0.1.", + "[582] Lukas Snoek, Jessica Loke, Michelle Kuehn, Mona Zimmermann, and Steven Scholte. \"ni-edu-data-minimal\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003965.v1.0.0." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 29 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[583] Olivier Girard, Ludovic de Rochefort, Maxime Guye, Jean-Philippe Ranjeva, and Arnaud Le Troter. \"7tamibrain\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003967.v1.0.0.", + "[584] W. Schellekens, C. Bakker, N.F. Ramsey, and N. Petridou. \"body motor\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003972.v1.0.0.", + "[585] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 11 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03974.v3.0.0.", + "[586] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03988.v1.0.0.", + "[587] Sarah Khalife, Susan T. Francis, Denis Schluppeck, Rosa Maria Sanchez-Panchuelo, and Julien Besle. \"fastvsslowertactile7t\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003990.v1.0.2.", + "[588] Jun Kai Ho, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"inter-individual deep image reconstruction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003993.v1.0.0.", + "[589] PhD) Daminov V. (MD, MSc) Novak E. (MD, Slepyova N. (MD), Mikhailov D. (MSc), and Karpulevich E. (MSc). \"pre-post rehabilitation fmri data of post-stroke patients\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003999.v1.0.2.", + "[590] J. Wammes, K.A. Norman, and N.B. Turk-Browne. \"Increasing stimulus similarity drives nonmonotonic representational change in hippocampus\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s004006.v1.0.4.", + "[591] J.A. Williams, E.H. Margulis, S.A. Nastase, J. Chen, U. Hasson, K.A. Norman, and C. Baldassano. \"high-order areas and auditory cortex both represent the high-level event structure of music\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004007.v1.0.2.", + "[592] Robert M. Mok and Bradley C. Love. \"abstract-category-signal\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004009.v1.0.0.", + "[593] Dominic Standage, Joseph Y. Nashed, J. Randall Flanagan, and Jason. P Gallivan. \"visuomotor rotation adaptation experiment\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004021.v1.0.0.", + "[594] Julio Cesar Hernandez Pavon, Nils Schneider Garces, John Patrick Begnoche, Lee Miller, and Tommi Rajj. \"tmseeg-mri-fmri-dwi data on paired associative stimulation and connectivity (shirley ryan abilitylab, chicago, il)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004024.v1.0.1.", + "[595] Cammille C Go, Huseyin O Taskin, Seyed Ahmad Ahmadi, Giulia Frazzetta, Laura Cutler, Saguna Malhotra, Jessica IW Morgan, Virginia L Flanagin, and Geoffrey K Aguirre. \"tome-go2022vestibularmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04038.v1.0.0.", + "[596] Hongmi Lee, Janice Chen, and Uri Hasson. \"filmfestival\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004042.v1.0.1." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 29 + }, + { + "type": "page_number", + "text": "23868", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 29 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[597] Sai Ma, Taicheng Huang, Yukun Qu, Xiayu Chen, Yajie Zhang, and Zonglei Zhen. \"an fmri dataset for whole-body somatotopic mapping in humans\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004044.v2.0.3.", + "[598] Miao Zhang, Qikai Qin, Shuning Zhang, Wei Liu, Hongping Meng, Mengyang Xu, Xinyun Huang, Xiaozhu Lin, Mu Lin, Peter Herman, Fahmeed Hyder, Raymond C. Stevens, Zheng Wang, Biao Li, and Garth J. Thompson. \"aerobic glycolysis imaging pet-mri\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004054.v1.0.0.", + "[599] Yoshihisa Nakayama, Sho K. Sugawara, Masaki Fukunaga, Yuki H. Hamano, Norihiro Sadato, and Yukio Nishimura. \"goal-directed motor task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004056.v1.0.2.", + "[600] Shai Sabbah, Michael S. Worden, Dimitrios D. Laniado, David M. Berson, and Jerome N. Sanes. \"luminance\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004065.v1.0.0.", + "[601] Matan Mazor, Chudi gong, and Stephen M. Fleming. \"confidence in equal and unequal variance settings\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds004081.v1.0.0.", + "[602] Julie Van, Samuel Nielson, and C. Brock Kirwan. \"triple dissociation revisited\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004086.v1.2.0.", + "[603] Martin Szinte, Gilles de Hollander, Marco Aqil, Serge Dumoulin, and Tomas Knapen. \"gaze prf\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004091.v2.0.0.", + "[604] Athula Pudhiyidath, Neal W Morton, Rodrigo Viveros Duran, Anna C. Schapiro, Ida Momennejad, Demitrius M. Hinojosa-Rowland, Robert J. Molitor, and Alison R. Preston. \"effects of temporal community structure learning on reasoning decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004094.v1.0.1.", + "[605] Kristina M. Rapuano, May I. Conley, Anthony C. Juliano, Gregory M. Conan, Maria T. Maza, Kylie Woodman, Steven A. Martinez, Eric Earl, Anders Perrone, Eric Feczko, Damien A. Fair, Richard Watts, BJ Casey, and Monica D. Rosenberg. \"an open-access accelerated adult equivalent of the abcd study neuroimaging dataset (a-abcd)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004097.v1.1.0.", + "[606] Dobrushina Olga R, Dobrunina Larisa A, Kremneva Elena I, and Pechenkova Ekaterina V. \"resting-state fmri data acquired before and after a course of infra-low frequency neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004101.v1.0.1.", + "[607] Damien Brevers, Chris Baeken, Stefanie De Smet, Beatrix Catoira, Sara De Witte, Qingha He, Pierre Maurage, Laimi Schultze, Guillaume Sescousse, Claudia Vila Verde, Claus Vögele, and Joel Billieux. \"stimulation of the dorsolateral prefrontal cortex modulates brain cue reactivity to reward(un)availability\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004102.v1.0.1.", + "[608] Siao-Shan Shen, Jen-Tang Cheng, I-Jeng Hsu, Der-Yow Chen, Ming-Hung Weng, and Chun-Chia Kung. \"collaborations and deceptions in strategic interactions revealed by hyperscanning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004103.v1.0.0." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 30 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[609] M.P. Weisend, F.M. Hanlon, R. Montano, S.P. Ahlfors, A.C. Leuthold, D. Pantazis, J.C. Mosher, A.P. Georgopoulos, M.S. Hamalainen, and C.J. Aine. \"mind data\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004107.v1.0.0.", + "[610] Yannan Zhu and Shaozheng Qin. \"emotional learning retroactively promotes memory integration through rapid neural reactivation and reorganization\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004109.v1.0.0.", + "[611] Kenji Kobayashi, Joseph W. Kable, Ming Hsu, and Adrianna C. Jenkins. \"neural representations of others' traits predict social decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004128.v1.0.0.", + "[612] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype segregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004129.v1.0.0.", + "[613] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype aggregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004130.v1.0.0.", + "[614] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype external example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004131.v1.0.1.", + "[615] Sepehr Mortaheb, Laurens Van Calster, Federico Raimondo, Manousos A. Klados, Paradeisios Alexandros Boulakis, Kleio Georgoula, Steve Majerus, Dimitri Van De Ville, and Athena Demertzi. \"experience sampling in resting state\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004134.v1.0.1.", + "[616] Bruno Direito, Marco Simões, Alexandre Sayal, and Miguel Castelo-Branco. \"targeting dynamic facial processing mechanisms in superior temporal sulcus using fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004141.v1.0.5.", + "[617] Bruno Direito, Manuel Ramos, Alexandre Sayal, Joao Pereira, and Miguel Castelo-Branco. \"exploring the neural correlates of feedback-related reward saliency and valence during real-time fmri-based neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004142.v1.0.1.", + "[618] Thania Balducci, Jalil Rascado-Toledo, Alely Valencia, Maria-Jose van Tol, Andre Aleman, and Eduardo A. Garza-Villarreal. \"a behavioral, clinical and brain imaging dataset with focus on emotion regulation of females with fibromyalgia\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds004144.v1.0.2.", + "[619] Martin Szinte, Anna Montagnini, and Guillaume Masson. \"rest_eye\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004158.v2.0.5.", + "[620] Lachlan T. Strike, Gabriella A.M. Blokland, Narelle K. Hansell, Nicholas G. Martin, Arthur W. Toga, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin imaging (qtim)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004169.v1.0.7." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 30 + }, + { + "type": "page_number", + "text": "23869", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 30 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[621] Ádám Nárai, Petra Hermann, Tibor Auer, Péter Kemenczky, János Szalma, István Homolya, Eszter Somogyi, Pál Vakli, Béla Weiss, and Zoltán Vidnyanszky. \"movement-related artefacts (mr-art) dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004173.v1.0.2.", + "[622] Stefanie Meliss, Cristina Pascua-Martin, Jeremy Skipper, and Kou Murayama. \"magic, memory, and curiosity (mmc) fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004182.v1.0.1.", + "[623] Irem Yildirim, Khan Hekmatyar, and Keith A. Schneider. \"lgn layers data\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004187.v1.0.2.", + "[624] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004192.v1.0.7.", + "[625] Iris Groen, Kenichi Yuasa, Amber Brands, Giovanni Piantoni, Stephanie Montenegro, Adeen Flinker, Sasha Devore, Orrin Devinsky, Werner Doyle, Patricia Dugan, Daniel Friedman, Nick Ramsey, Natalia Petridou, and Jonathan Winawer. \"visual ecog dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004194.v2.0.0.", + "[626] Foteini Liwicki, Vibha Gupta, Rajkumar Saini, Kanjar De, Nosheen Abid, Sumit Rakesh, Scott Wellington, Holly Wilson, Marcus Liwicki, and Johan Eriksson. \"bimodal dataset on inner speech\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004196.v2.0.2.", + "[627] Fabiane Schuch, Lennart Walger, Matthias Schmitz, Bastian David, Tobias Bauer, Antonia Harms, Laura Fischbach, Freya Schulte, Martin Schidlowski, Johannes Reiter, Felix Bitzer, Randi von Wrede, Attila Racz, Tobias Baumgartner, Valeri Borger, Matthias Schneider, Achim Flender, Albert Becker, Hartmut Vatter, Bernd Weber, Louisa Specht-Riemenschneider, Alexander Radbruch, Rainer Surges, and Theodor Rüber. \"an open presurgery mri dataset of people with epilepsy and focal cortical dysplasia type ii\". OpenNeuro, 2025. doi: 10.18112/openneuro.d004199.v1.0.6.", + "[628] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-meg\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004212.v2.0.1.", + "[629] Hermes D and Wandell BA. \"brainbeats\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004213.v1.0.1.", + "[630] Mark A. Thornton, Elyssa Barrick, Nathan Liang, and Diana I. Tamir. \"interacting representations of mental states and traits\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004217.v1.0.0.", + "[631] Kyoko Ohashi, Martin H. Teicher, Elizabeth Bolger, Laura C. Hernandez Garcia, Poopak Hafezi, Leslie P. Weiser, Cynthia E. McGreenery, and Alaptagin Khan. \"lite-book_Alertness_study\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004219.v1.0.0.", + "[632] Mark A. Thornton and Diana I. Tamir.\"summed social cognition: additive neural representations of situations, mental states, and actions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004226.v1.0.0." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 31 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[633] Anne T. Park, Hilary Richardson, Ursula A. Tooley, Cassidy L. McDermott, Austin L. Boroshok, Adrian Ke, Julia A. Leonard, M. Dylan Tisdall, Kirby Deater-Deckard, J. Christopher Edgar, and Allyson P. Mackey. \"early stressful experiences are associated with reduced neural responses to naturalistic emotional and social content in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004228.v1.0.1.", + "[634] Min-Jeong Kim, Jae-Hoon Lee, Fernanda Juarez Anaya, Jinsoo Hong, William Miller, Sanjay Telu, Prachi Singh, Michelle Y Cortes, Katharine Henry, George L Tye, Michael P Frankland, Jose A Montero Santamaria, Jeih-San Liow, Sami S Zoghbi, Masahiro Fujita, Victor W Pike, and Robert B Innis. \"first-in-human evaluation of [11c]ps13, a novel pet radioligand, to quantify cyclooxygenase-1 in the brain\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004230.v3.0.0.", + "[635] Miguel B. Garcia, Gilles de Hollander, Marcus Grueschow, Rafael Polania, Michael Woodford, and Christian C. Ruff. \"individual risk attitudes arise from noise in neurocognitive magnitude representations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004259.v1.0.0.", + "[636] Xuhao Shao, Ao Li, Chuangsheng Chen, Elizabeth F. Loftus, and Bi Zhu. \"cross-stage neural pattern similarity in the hippocampus predicts false memory derived from post-event inaccurate information\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004261.v2.0.0.", + "[637] Thirza Dado, Yaghmur Güçlütürk, Luca Ambrogioni, Gabriëlle Ras, Sander Bosch, Marcel van Gerven, and Umut Güçlü. \"hyper\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04280.v1.0.1.", + "[638] Krista Bond, Javier Rasero, Raghav Madan, Jyotika Bahuguna, Jonathan Rubin, and Timothy Verstynen. \"neuroloki\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004283.v1.0.3.", + "[639] Chad S. Rogers, Michael S. Jones, Sarah McConkey, and Jonathan E. Peelle. \"listening task\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs004285.v1.0.0.", + "[640] Wilma A. Bainbridge and Chris I. Baker. \"multidimensional memory topography in the medial parietal cortex identified from neuroimaging of thousands of daily memory videos\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004286.v1.0.0.", + "[641] Shaonan Wang, Yunhao Zhang, Xiaohan Zhang, Jingyuan Sun, Nan Lin, Jiajun Zhang, Chengqing Zong, and Shaonan Wang. \"an fmri dataset for concept representation with semantic feature annotations\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004301.v1.0.2.", + "[642] Joan Soler-Vidal, Paola Fuentes-Claramonte, Pilar Salgado-Pineda, Nuria Ramiro, María Ángeles García-León, María Llanos Torres, Antonio Arevalo, Amalia Guerrero-Pedraza, Josep Munuera, Salvador Sarro, Raymond Salvador, Wolfram Hinzen, Peter McKenna, and Edith Pomarol-Clotet. \"brain correlates of speech perception in schizophrenia patients with and without auditory hallucinations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004302.v1.0.1." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 31 + }, + { + "type": "page_number", + "text": "23870", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 31 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[643] Jason A Avery, Madeline Carrington, and Alex Martin. \"a common neural code for representing imagined and inferred tastes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004312.v1.0.3.", + "[644] Momchil S. Tomov, Pedro A. Tsividis, Thomas Pouncy, Joshua B. Tenenbaum, and Samuel J. Gershman. \"atari-style video game learning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004323.v1.0.0.", + "[645] Emily E. Perszyk, Jessica Trinh, Xue S. Davis, Zach Hutelin, Leonie Koban, Maria G. Veldhuizen, Jelena Djordjevic, Marilyn Jones-Gotman, Tor Wager, Hedy Kober, and Dana M. Small. \"odour-imagery ability is linked to food craving, intake, and adiposity change in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004327.v1.0.3.", + "[646] Johannes J.D. Singer, Radoslaw M. Cichy, and Martin N. Hebart. \"the spatiotemporal neural dynamics of object recognition for natural images and line drawings (fmri)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004331.v1. 0.4.", + "[647] Melanie Ganz and Hannah Eichhorn. \"datasets with and without deliberate head movements for evaluating the performance of markerless prospective motion correction and selective reacquisition in a general clinical protocol for brain mri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004332.v1.3.0.", + "[648] Loris Naspi, Paul Hoffman, and Alexa Morcom. \"semantic_encoding\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004341.v1.0.0.", + "[649] Oscar Ferrante, Ling Liu, Tamas Minarik, Urszula Gorska, Tara Ghafari, Huan Luo, and Ole Jensen. \"flux: A pipeline for meg analysis\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004346.v1.0.8.", + "[650] L.E. Frank and D. Zeithamova. \"evaluating methods for measuring background connectivity in slow event-related fmri designs\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004349.v1.0.0.", + "[651] Asieh Zadbood, Samuel A. Nastase, Janice Chen, Kenneth A. Norman, and Uri Hasson. \"sixthsense\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004359.v1.0.0.", + "[652] Korey P. Wylie, Benzi M. Kluger, Luis D. Medina, Samantha K. Holden, Eugene Kronberg, Jason R. Tregellas, and Isabelle Guard. \"parkinson's disease, functional connectivity, and cognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004392.v1.0.0.", + "[653] Lauryn Burleigh and Steven Greening. \"fear in the mind's eye: the neural correlates of differential fear acquisition to imagined conditioned stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004393.v1.0.4.", + "[654] Doug Greve, Paul Wighton, Melanie Ganz, Martin Nørgaard, Paolo Zanotti-Fregonara, Cyril Pernet, Anthony Galassi, Adam Thomas, Nafiseh Ghazanfari, Jeih-San Liow, Granville Matheson, Gitte Knudsen, and Robert Innis. \"a pet molecular imaging brain atlas of cyclooxygenase-1 (kim 2021)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004401.v1.3.0.", + "[655] Corey Fernandez, Jiefeng Jiang, Shao-Fang Wang, Hannah L. Choi, and Anthony D. Wagner. \"rid\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004406.v1.0.0." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 32 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[656] Marc Himmelberg, Ekin Tuncok, Jesse Gomez, Kalanit Grill-Spector, Marisa Carrasco, and Jonathan Winawer. \"stanford child and adult checkerboard retinotopy dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004440.v1.0.1.", + "[657] Puti Wen, Michael S. Landy, and Bas Rokers. \"identifying cortical areas that underlie the transformation from 2d retinal to 3d head-centric motion signals\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004443.v1.0.0.", + "[658] Will be updated. \"xcp_walkthrough\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004450.v1.0.1.", + "[659] Carolina Feher da Silva, Gaia Lombardi, Micah Edelson, and Todd A. Hare. \"magic_carpet\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004455.v1.1.0.", + "[660] Agata Wolna and Zofia Wodniecka. \"bilingual speech production with functional localizers\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004456.v1.0.1.", + "[661] Renzo Huber. \"highresolution development\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004458.v1.0.2.", + "[662] Chris Racey, Charlotte Rae, Ivor Simpson, Oscar Bowen-Hill, Mathilde Bauer, Julia Simner, Christina Kampoureli, and Jamie Ward. \"mapping the connectome of synaesthesia: An open access mri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004466.v1.0.2.", + "[663] Halie Olson, Emily Chen, Kirsten Lydic, and Rebecca Saxe. \"mri data from 20 adults in response to videos of dialogue and monologue from sesame street\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004467.v1.0.0.", + "[664] Raul Rodríguez-Cruces, Vicente Camacho-Téllez, Alfonso Fajardo, and Luis Concha. \"temporal lobe epilepsy - unam\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004469.v1.1.4.", + "[665] Jonathan C. Lau, Yiming Xiao, Roy A. M. Haast, Greydon Gilmore, Kamil Uludag, Keith W. MacDougall, Ravi S. Menon, Andrew G. Parrent, Terry M. Peters, and Ali R. Khan. \"stereotactic neurosurgery dataset (snsx)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004470.v1.0.1.", + "[666] Mohamad Abbass, Greydon Gilmore, Alaa Taha, Ryan Chevalier, Magdalena Jach, Terry M. Peters, Ali R. Khan, and Jonathan C. Lau. \"London heath sciences center parkinson's disease dataset (lhscpd)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004471.v1.0.1.", + "[667] Alexander P. Rockhill, Alessandra Mantovani, Brittany Stedelin, Admed M. Raslan, and Nicole C. Swann. \"seeg forced two-choice task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004473.v1.0.2.", + "[668] Noelle A. Jacobsen and Daniel P. Ferris. \"mobile eeg splitbelt walking study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004475.v1.0.3.", + "[669] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Nicole Tacugue, Zenia Valdiviezo, Laura D. Lewis, and Zachary Diamandis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004478.v1.0.2.", + "[670] Samuel Planton*, Fosca Al Roumi*, Liping Wang, and Stanislas Dehaene. \"abseqfmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso004482.v1.0.0." + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 32 + }, + { + "type": "page_number", + "text": "23871", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 32 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[671] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004484.v1.0.1.", + "[672] Ming Zhou, Zhengxin Gong, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for human action recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004488.v1.1.1.", + "[673] Zhengxin Gong, Ming Zhou, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for the visual processing of naturalistic scenes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004496.v2.1.2.", + "[674] Ph.D. Evan Gordon and Ph.D. Timothy Laumann. \"perinatal stroke\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004498.v1.0.0.", + "[675] Mikl Michal, Kovarova Anezka, Gajdos Martin, Marecek Radek, Novakova Marie, and Slavicek Tomas. \"multi-echo simultaneous multislice fmri dataset: Effect of acquisition parameters on fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps004499.v1.0.3.", + "[676] Amanda Studnicki and Daniel P. Ferris. \"real world table tennis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004505.v1.0.4.", + "[677] Gabriel Castrillon, Samira Epp, Antonia Bose, Laura Fraticelli, André Hechler, Roman Belenya, Andreas Ranft, Igor Yakushev, Lukas Utz, Lalith Sundar, Josef P Rauschecker, Christine Preibisch, Katarzyna Kurcyus, and Valentin Riedl. \"the energetic costs of the human connectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004513.v1.0.4.", + "[678] Clara Sava-Segal, Chandler Richards, Megan Leung, Daniel A. Handwerker, Javier Gonzalez-Castillo, Peter J. Molfese, Peter A Bandettini, and Emily S. Finn. \"avid: Audio-visual individual differences dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004516.v2.0.2.", + "[679] Nestor Zaragoza-Jimenez, Hauke Niehaus, Ina Thome, Christoph Vogelbacher, Gabriele Ende, Inge Kamp-Becker, Dominik Endres, and Andreas Jansen. \"modeling face recognition in the predictive coding framework: A combined computational modeling and functional imaging study - stage 2\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004529.v1.1.1.", + "[680] Elizabeth Valles-Capetillo, Magda Giordano, and Cristian Ibarra. \"ironia vev\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004533.v1.0.0.", + "[681] Sebastian Dresbach, Renzo Huber, Omer Faruk Gulban, and Rainer Goebel. \"event-related vaso\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004539.v1.1.1.", + "[682] Emalie McMahon, Michael F. Bonner, and Leyla Isik. \"social interaction dyads\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004542.v1.0.0.", + "[683] Paul Black, C. Brock Kirwan, Thomas Meservy, William Tayler, and Jeff Williams. \"an fmri investigation of the neurocognitive processing of strategies and measures\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004544.v1.0.0." + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 33 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[684] Michael Rojek-Giffin, Mael Lebreton, Andrea Farina, Jorg Gross, Jean Daunizeau, and Carsten de Dreu. \"learning rules of engagement for social exchange within and between groups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004553.v1.0.1.", + "[685] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 1\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004556.v1.0.1.", + "[686] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 2\". OpenNeuro, 2023. doi: 10.18112/openneuro. ds004557.v1.1.0.", + "[687] M. Rebsamen, D. Romascano, M. Capiglioni, R. Wiest, P. Radojewski, and C. Rummel. \"the phantom of bern: repeated scans of two volunteers with eight different combinations of mr sequence parameters\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004560.v1.0.1.", + "[688] Youngjo Song, Wooree Shin, Pyeongsoo Kim, and Jaeseung Jeong. \"mvpa for motor adaptation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004562.v1.0.4.", + "[689] Paul Broca and Carl Wernicke. \"demo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004564.v1.0.1.", + "[690] Ludmila Midrigan-Ciochina, Kayla P. Vodacek, Srinija Balabhadra, and David P. Corina. \"brain differences in monolingual and highly proficient multilingual speakers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004581.v2.0.1.", + "[691] Yingying Wang, Hongmi Lee, and Brice Kuhl. \"ds004590\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004590.v1.0.0.", + "[692] Hayoung Song, Won Mok Shim, and Monica D. Rosenberg. \"song dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004592.v1.0.1.", + "[693] Marc N. Coutanche, Jake Sauter, Essang Akpan, Rae Buckser, August Vincent, and M. Kathleen Caulfield. \"brain surface fingerprints\". OpenNeuro, 2023. doi: 10. 18112/openneuro.ds004594.v1.0.1.", + "[694] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, Bruno Nazarian, Julien Sein, and Jean-Luc Anton. \"gloups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004597.v2.0.0.", + "[695] Antonin Rovai, Valentina Lolli, Nicola Trotta, Serge Goldman, and Xavier De Tiège. \"cerebrovascular reactivity normative dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004604.v2.0.0.", + "[696] Cameron C. McKay, Brooke Scheinberg, Ellie P. Xu, Katharina Kircanski, Daniel S. Pine, Melissa A. Brotman, Ellen Leibenluft, and Julia O. Linke. \"emotion and development branch phenotyping and dti (2012-2017)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004605.v1.0.1.", + "[697] David Romascano, Gian Franco Piredda, Samuele Caneschi, Tom Hilbert, Ricardo Corredor, Bénédicte Maréchal, Tobias Kober, Jean-Baptiste Ledoux, Eleonora" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 33 + }, + { + "type": "page_number", + "text": "23872", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 33 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fornari, Patric Hagmann, and Solange Denervaud. \"developmental relaxometry 2023\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004611.v1.0.2.", + "[698] Dylan M. Nielson, Neda Sadeghi, Lisa S. Gorham, Lillian Eisner, Jeremy Taigman, Katherine Haynes, Karen Qi, Christopher C. Camp, Payton Fors, Diana Rodriguez, Jerry McGuire, Erin Garth, Chana Engel, Mollie Davis, Kenneth Towbin, and Argyris Stringaris. \"characterization and treatment of adolescent depression (cat-d)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004627.v1.1.0.", + "[699] Mengfan Zhang, Alexander Karner, Kathrin Kostorz, Sophia Shea, David Steyrl, Filip Melinscak, Ronald Sladky, Cindy Sumaly Lor, and Frank Scharnowski. \"spida-mrineuroimaging dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004630.v1.1.3.", + "[700] A. De Soares, T. Kim, F. Mugisho, E. Zhu, A. Lin, C. Zheng, and C. Baldassano. \"top-down attention impacts neural dynamics during narratives with overlapping event scripts\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004631.v1.0.0.", + "[701] P. G. Bissett, I. W. Eisenberg, S. Shim, J. A. H. Rios, H. M. Jones, M. P. Hagen, A. Z. Enkavi, J. K. Li, J. A. Mumford, D. P. MacKinnon, L. A. Marsch, and R. A. Poldrack. \"cognitive tasks, anatomical mri, and functional mri data evaluating the construct of self-regulation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004636.v1.0.4.", + "[702] Edlow B. L., Olchanyi M. D., Freeman H. J., Li J., Maffei C., Snider S. B., Zollei L., Eugenio Iglesias J., Augustinack J., Bodien Y. G., Haynes R. L, Greve D. N., Diamond B. R., Stevens A., Giacino J. T., Destrieux C., van der Kouwe A., Brown E. N., Folkerth R. D., Fischl B., and Kinney H. C. \"sustaining wakefulness: Brainstem connectivity in human consciousness\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004640.v1.0.4.", + "[703] Sydney M. Bailes, Daniel E.P. Gomez, Beverly Setzer, and Laura D. Lewis. \"resting-state fmri signals contain spectral signatures of local hemodynamic response timing\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004645.v1.0.0.", + "[704] Heesun Park, Hoyoung Doh, Eunhwi Lee, Harhim Park, and Woo-Young Ahn. \"the neurocognitive role of working memory load when pavlovian motivational control affects instrumental learning\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004647.v1.0.2.", + "[705] Ana L. Miranda-Angulo, Juan D. Sánchez-lópez, Daniel A. Vargas-Tejada, Juan C. Calderón, Jaime Gallo-Villegas, Juan F. Alzate-Restrepo, Jazmin X. Suarez-Revelo, and Gabriel Castrillón. \"sympathovagal quotient and functional connectivity of control networks are related to gut ruminococcaceae abundance in healthy men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004648.v1.0.0.", + "[706] Martin Norgaard. \"the mn dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004650.v1.0.2.", + "[707] Kathleen C Schmidt, Inna Loutaev, Thomas V Burlin, Zenaide Quezado, Audrey Thurm, Carrie Sheeler, and Carolyn Beebe Smith. \"rates of cerebral protein synthesis in subjects with fragile x syndrome\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004654.v1.0.1." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 34 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[708] Travis D. Masterson, C. Brock Kirwan, Lance E. Davidson, and James D. LeCheminant. \"neural reactivity to visual food stimuli is reduced in some areas of the brain during evening hours compared to morning hours: an fmri study in women\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004656.v1.1.0.", + "[709] Neha A Reddy, Kristina M Zvolanek, and Molly G Bright. \"denoising task-correlated head motion from motor-task fmri data with multi-echo ica\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004662.v1.1.0.", + "[710] Fan Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C. Aoki, Jin Hirano, and Yukiyasu Kamitani. \"visual illusion reconstruction\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004670.v1.0.1.", + "[711] Evan Layher, Tyler Santander, Puneeth Chakravarthula, Nicole Marinsek, Benjamin O. Turner, Miguel P. Eckstein, and Michael B. Miller. \"widespread frontoparietal fmri activity is greatly affected by changes in criterion placement, not discriminability, during recognition memory and visual detection tests\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004692.v1.0.0.", + "[712] Johannes Singer, Agnessa Karapetian, Martin Hebart, and Radoslaw Cichy. \"identifying and characterizing scene representations relevant for categorization behavior\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004693.v1.0.3.", + "[713] Bari Fuchs, Alaina Pearce, Barbara Rolls, Stephen Wilson, Emma Jane Rose, Charles Geier, Hugh Garavan, and Kathleen Keller. \"food and brain study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004697.v1.0.2.", + "[714] Kelly Chang, Ione Fine, and Geoffrey M. Boynton. \"chn retinotopic mapping dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004698.v2.0.1.", + "[715] David A Feinberg, Alexander JS Beckett, and An T VU. \"next generation mri scanner designed for ultra-high resolution human brain imaging at 7-tesla\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004710.v1.1.0.", + "[716] Loreen Tisdall and Rui Mata. \"agerisk\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004711.v1.0.0.", + "[717] S. Warrington, A. Torchi, O. Mougin, A. Ntata, J. Campbell, M. Craig, F. Alfaro-Almagro, K.L. Miller, P.S. Morgan, M. Jenkinson, and S.N. Sotiropoulos. \"on-harmony: A resource for development and comparison of multi-modal brain 3t mri harmonisation approaches\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004712.v2.0.1.", + "[718] Elisa Zamboni, Alexis Makin, Marco Bertamini, and Antony Morland. \"symmetry & luminance bold fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004715.v1.0.3.", + "[719] Elena Filimonova, Anton Pashkov, Norayr Borisov, Anton Kalinovsky, and Jamil Rzaev. \"utilizing amide proton transfer technique to characterise diffuse gliomas based on who 2021 classification of cns tumors\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004717.v1.0.0.", + "[720] Mohammad Momenian, Zhengwu Ma, Shuyi Wu, Chengcheng Wang, and Jixing Li. \"le petit prince hong kong: Naturalistic fmri and eeg dataset from" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 34 + }, + { + "type": "page_number", + "text": "23873", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 34 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "older cantonese speakers\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004718.v1.1.1.", + "[721] Unknown authors. \"descriptive name for this dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004720.v1.0.0.", + "[722] Marilyn Horta, Rebecca Polk, and Natalie Ebner. \"single dose intranasal oxytocin administration: Data from healthy younger and older adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004725.v1.0.1.", + "[723] Shrinivas Bishu, Kathleen C Schmidt, Thomas V Burlin, Michael A Charming, Lisa Horowitz, Tianjiang Huang, Zhong-Hua Liu, Mei Qin, B-K Vuong, Aaron J Unterman, Zengyan Xia, Alan Zametkin, Peter Herscovitch, Zenaide Quezado, and Carolyn B Smith. \"effects of propofol anesthesia on rates of cerebral protein synthesis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004730.v1.0.0.", + "[724] Dante Picchioni, Kathleen C Schmidt, Kelly K McWhirter, Inna Loutaev, Adriana J Pavletic, Andrew M Speer, Alan J Zametkin, Ning Miao, Shrinivas Bishu, Kate M Turetsky, Anne S Morrow, Jeffrey L Nadel, Brittney C Evans, Diana M Vesselinovitch, Carrie A Sheeler, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis and memory formation during sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004731.v1.0.0.", + "[725] Dante Picchioni, Kathleen C Schmidt, Inna Loutaev, Adriana J Pavletic, Carrie Sheeler, Shrinivas Bishu, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis in stages of sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004733.v1.0.0.", + "[726] Hamsanandini Radhakrishnan, Chenying Zhao, Valerie J. Sydnor, Erica B. Baller, Philip A. Cook, Damien Fair, Barry Giesbrecht, Bart Larsen, Kristin Murtha, David R. Roalf, Sage Rush-Goebel, Russell Shinohara, M. Dylan Tisdall, Jean Vettel, Scott Grafton, Matthew Cieslak, Theodore Satterthwaite, and Haochang Shou. \"cs-dsi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004737.v2.0.0.", + "[727] Anne Kavounoudias, Caroline Landelle, Jeremy Dana, Bruno Nazarian, Julien Sein, Jean-Luc Anton, Laurent Pruvost, Frederic Giraud, and Olivier Felician. \"audiotact_ya\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004743.v1.0.0.", + "[728] Joe Bathelt, Joanne Taylor, and Kathy Rastle. \"language fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004765.v1.0.0.", + "[729] Rene Weber, Frederic R. Hopp, Allison Eden, and Kate Lee. \"vicarious punishment of moral violations in naturalistic drama narratives predicts cortical synchronization\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004775.v1.1.", + "[730] Katyucia de Macedo Rodrigues, Emma Ben-Avi, Danielle D. Sliva, Myong sun Choe, Marie Drottar, Ruopeng Wang, Bruce Fischl, Patricia E. Grant, and Lilla Zöllei. \"infant freesurfer test subject (cnybch)\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004776.v1.0.0.", + "[731] A L Williams, S Vicary, and G Orgs. \"functional mri dataset for evaluating brain synchrony among dance spectators\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004783.v1.0.1." + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 35 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[732] Eddy Cavalli, Valérie Chanoine, and Johannes C. Ziegler. \"morphosem\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004786.v1.0.1.", + "[733] Dylan Nielson, Andre Zugman, Marie Zelenina, and Danny Pine. \"nimh meter (multi-echo test-retest)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004787.v1.1.0.", + "[734] Fu Yu Kwok, Eric D. Wilkey, Lien Peters, Ellyn Khiu, Rebecca Bull, Kerry Lee, and Daniel Ansari. \"kwok et al., 2023, human brain mapping\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004791.v1.0.0.", + "[735] Hampus Olsson, Jason Michael Millward, Ludger Starke, Thomas Gladytz, Tobias Klein, Jana Fehr, Wei-Cheng Lai, Christoph Lippert, Thoralf Niendorf, and Sonia Waiczies. \"simulated motion artifacts based on the (mr-art) dataset (https://openneuro.org/datasets/ds004173/versions/1.0.2)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004795.v1.0.0.", + "[736] Umit Keles, Julien Dubois, Kevin J. M. Le, J. Michael Tyszka, David A. Kahn, Chrystal M. Reed, Jeffrey M. Chung, Adam N. Mamelak, Ralph Adolphs, and Ueli Rutishauser. \"fmri data for: Multimodal single-neuron, intracranial eeg, and fmri brain responses during movie watching in human patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004798.v1.0.5.", + "[737] Audrey Mazancieux, Franck Mauconduit, Alexis Amadon, Jan Willem de Gee, Tobias Donner, and Florent Meyniel. \"brainstem fmri signaling of surprise across different types of deviant stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004808.v1.0.0.", + "[738] Michelle Moerel and Essa Yacoub. \"high-res gradient echo epi and 3d grase data of auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004814.v1.0.0.", + "[739] Agustin Lage-Castellanos, Federico De Martino, Geoff M. Ghose, Omer Faruk Gulban, and Michelle Moerel. \"selective attention sharpens population receptive fields in human auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004815.v1.0.1.", + "[740] Ana F. Palenciano, Carlos González-García, Jan de Houwer, Baptist Lieefoghe, and Marcel Brass. \"concurrent response and action effect representations across the somatomotor cortices during novel task preparation.\" OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004829.v1.0.1.", + "[741] Michelle Moerel, Federico De Martino, Kamil Ugurbil, Essa Yacoub, and Elia Formisano. \"processing of frequency and location in human subcortical auditory structures\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004835.v1.0.0.", + "[742] Fran López-Caballero, Mark Curtis, Brian Coffman, and Dean Salisbury. \"magnetoencephalographic (meg) pitch and duration mismatch negativity (mmn) in first-episode psychosis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004837.v1.0.0.", + "[743] Kira Noad, David Watson, and Timothy Andrews. \"game of thrones - a naturalistic viewing dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004848.v1.0.1." + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 35 + }, + { + "type": "page_number", + "text": "23874", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 35 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[744] Sean Paulsen and Michael A. Casey. \"shanxi enculturation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004866.v1.0.0.", + "[745] Nafiseh Ghazanfari, Jeih-San Liow, Min-Jeong Kim, Raven Cureton, Adrian Lee, Carson Knoer, Madeline Jenkins, Jinsoo Hong, Jose A. Montero Santamaria, H. Umesha Shetty, Anthony Galassi, Paul Wighton, Martin Norgaard, Douglas N. Greve, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. [11c]ps13 demonstrates pharmacologically selective and substantial binding to cyclooxygenase-1 (cox-1) in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004868.v1.0.4.", + "[746] Elenor Morgenroth, Stefano Moia, Laura Vilaclara, Raphael Fournier, Michal Muszynski, Maria Ploumitsakou, Marina Almato Bellavista, Patrik Vuilleumier, and Dimitri Van De Ville. \"emo-film\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps004892.v1.0.1.", + "[747] Toru Maekawa, Takafumi Sasaoka, Toshio Inui, Alan S. R. Fermin, and Shigeto Yamawaki. \"heart rate and insula activity increase in response to music in individuals with high interoceptive sensitivity\". OpenNeuro, 2023. doi: 10.18112/openneuro.ps004894.v1.0.0.", + "[748] Harrison Ritz and Amitai Shenhav. \"pact_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004909.v1.1.0.", + "[749] Ye Wu, Xiaoming Liu, Yunzhi Huang, Tao Zhou, and Fan Zhang. \"an open relaxation-diffusion mri dataset in neurosurgical studies\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004910.v1.0.1.", + "[750] Alejandra Figueroa-Vargas, Gabriela Valdebenito-Oyarzo, María Paz Martínez-Molina, Francisco Zamorano, and Pablo Billeke. \"probability decision-making task with ambiguity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004917.v1.0.1.", + "[751] David V. Smith, James Wyngaarden, Cooper J. Sharp, Daniel Sazhin, Ori Zaff, Dominic Fareri, and Johanna Jarcho. \"an fmri dataset of social and nonsocial reward processing in young adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004920.v1.1.", + "[752] Lonike K Faes, Agustin Lage-Castellanos, Giancarlo Valente, Zidan Yu, Martijn A. Cloos, Luca Vizioli, Steen Moeller, Essa Yacoub, and Federico De Martino. \"predictive processing in the auditory cortex measured at 7t\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004928.v1.0.0.", + "[753] Shari Liu, Kirsten Lydic, Lingjie Mei, and Rebecca Saxe. \"fmri dataset: Violations of psychological and physical expectations in human adult brains\". OpenNeuro, 2024. doi: 10.18112/openneuro.dss004934.v1.0.0.", + "[754] Elise M. Cardinale, Jessica Bezek, Julia O Linke, Olivia Siegal, Gabrielle Freitag, Anni Subar, Khosravi Parmis, Ajitha Mallidi, Olivia Peterson, Isaac Morales, Simone P Haller, Courtney Filippi, Kyunghun Lee, Melissa A. Brotman, Daniel S. Pine, Ellen Leibenluft, and Katharina Kirkanski. \"multivariate assessment of inhibitory control in youth: Links with psychopathology and brain function dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005754.v1.1.0." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 36 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[755] Patrick Klepits, Karl Koschutnig, Thomas Zussner, and Andreas Fink. \"effects of a seven-week running intervention with moderate intensity on the volume of the hippocampus and depressive symptoms in young men from the general population.\" OpenNeuro, 2024. doi: 10.18112/openneuro. ds004937.v1.0.1.", + "[756] Jens Jensen, Andreanna Benitez, Erin Klintworth, Hunter Moss, and Kathryn Thorn. \"pydesigner example dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004945.v1.0.0.", + "[757] Roxane Philips, Chris Baeken, Joel Billieux, James Harris, Pierre Maurage, Ismael Muela, Irem Tugce Oz, Athur Pabst, Guillaume Sescousse, Claus Vogele, and Damien Brevers. \"brain mechanisms discriminating enactive mental simulations of running and plogging\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004946.v1.0.0.", + "[758] Marleen Haupt, Monika Graumann, Santani Teng, Carina Kaltenbach, and Radoslaw M. Cichy. \"braille letters - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004956.v1.0.1.", + "[759] Angelique Volfart, Katie McMahon, and Greig de Zubicaray. \"a comparison of denoising approaches for spoken word production related artefacts in continuous multiband fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004957.v1.0.2.", + "[760] Douglas N. Greve and Bruce Fischl. \"the freesurfer maintenance dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004958.v1.0.0.", + "[761] Corinna Perchtold-Stefan, Christian Rominger, Karl Koschutnig, and Andreas Fink. \"truecrime\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004965.v1.0.1.", + "[762] Ekaterina Torubarova, Caroline Arvidsson, Jonathan Berrebi, Julia Uddén, and André Pereira. \"neuroengage\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004996.v2.0.0.", + "[763] Eric C. Leuthardt, Joshua S. Shimony, Abraham Z. Snyder, Donna Dierker, and Ki Yun Park. \"retrospective task/rest fmri data from tumor patients\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005003.v2.0.0.", + "[764] Neha A Reddy, Rebecca G Clements, and Molly G Bright. \"simultaneous cortical, subcortical, and brainstem mapping of sensory activation\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005009.v1.0.0.", + "[765] Michael I. Demidenko, Edward D. Huntley, and Daniel P. Keating. \"adolescent health risk behavior study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005012.v1.0.3.", + "[766] David Seminowicz, Shana Burrowes, Alexandra Kearson, Jing Zhang, Samuel Kimmel, Luma Samawi, Andrew Furman, and Michael Keaser. \"mbsr\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005016.v1.1.1.", + "[767] Unknown authors. \"pop\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005017.v1.0.2.", + "[768] Sara Ponticorvo and Fabrizio Esposito. \"hearing loss connectome\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005026.v1.0.0.", + "[769] Michael I. Demidenko, Ryan Klaus, Mary Soules, and Mary M. Heitzeg. \"michigan longitudinal study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005027.v1.0.3." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 36 + }, + { + "type": "page_number", + "text": "23875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 36 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[770] Marie Mueckstein, Kai Görgen, Stephan Heinzel, Urs Granacher, A. Michael Rapp, and Christine Stelzel. \"modality-based multitasking and practice - fmri\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005038.v1.0.3.", + "[771] Michael S. Cohen, Qiongwen (Jovie) Cao, and Jean Decety. \"political moralization (pmm)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005040.v1.2.0.", + "[772] Panpan Chen, Chi Zhang, Bao Li, Li Tong, Linyuan Wang, Shuxiao Ma, Long Cao, Ziya Yu, and Bin Yan. \"an fmri dataset in response to large-scale short natural dynamic facial expression videos\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005047.v1.0.7.", + "[773] S.H.P. Collin, R. Kempner, S. Srivatsan, and K.A. Norman. \"neural codes track prior events in a narrative and predict subsequent memory for details\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms00505.v1.0.0.", + "[774] Guangyu Zhou, Gregory Lane, Thorsten Kahnt, and Christina Zelano. \"dataset2\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005056.v1.0.0.", + "[775] Jing Wang. \"cr dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005063.v1.0.0.", + "[776] jessica. \"ds: ds005069:1.0.0\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005069.v1.0.0.", + "[777] jessica. \"ds: ds005072:1.0.1\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005072.v1.0.1.", + "[778] Brian P Keane, Bart Krekelberg, Ravi Mill, Steven M Silverstein, Judy L Thompson, Megan R Serody, Deanna M Barch, and Michael W Cole. \"the neural basis of visual shape completion in schizophrenia and bipolar disorder\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005073.v1.0.0.", + "[779] Caroline Landelle, Nawal Kinany, Benjamin De Leener, Nicholas D. Murphy, Ovidiu Lungu, Véronique Marchand-Pauvert, Dimitri Van De Ville, and Julien Doyon. \"simultaneous brain and spinal cord at rest: an fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005075.v1.0.1.", + "[780] Unknown authors. \"sequence pilot: multiecho and multiband fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005085.v1.0.0.", + "[781] Hannah Grotzinger, Laura Pritschet, Elle M. Murata, Tyler Santander, Pavel Shapturenka, Caitlin Taylor, and Emily G. Jacobs. \"28andhe\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05115.v1.2.0.", + "[782] Charles J. Lynch and Conor Liston. \"weill cornell medicine multi-echo (wcm-me) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005118.v1.0.0.", + "[783] David V. Smith, Cooper J. Sharp, Abraham Dachs, James Wyngaarden, Daniel Sazhin, Jen Yang, Melanie Kos, Tia Tropea, Ishika Kohli, John A. Clithero, Ingrid Olson, Tania Giovannetti, Dominic Fareri, and Johanna M. Jarcho. \"social reward and nonsocial reward processing across the adult lifespan: An interim multi-echo fmri and diffusion dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005123.v1.1.3.", + "[784] Hannah Tarder-Stoll, Christopher Baldassano*, and Mariam Aly*. \"the brain hierarchically represents the past and future during multistep anticipation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005125.v1.0.0." + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 37 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[785] Takahiro Hirao, Mitsuhiro Miyamae, Daisuke Matsuyoshi, Ryuto Inoue, Yuhei Takado, Takayuki Obata, Makoto Higuchi, Naotsugu Tsuchiya, and Makiko Yamada. \"a neuroimaging dataset during sequential color qualia similarity judgments\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005126.v1.0.3.", + "[786] Dante Picchioni, Jeff H Duyn, and Jacco A de Zwart. \"amri 16-n-0031 sleep1\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005127.v1.0.4.", + "[787] Sangil Lee, Runxuan Niu, Lusha Zhu, Andrew Kayser, and Ming Hsu. \"deception signaling task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps005128.v1.0.0.", + "[788] Adam Raikes. \"sequence test for multiband dwi and magnetization transfer sequence\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005134.v1.0.0.", + "[789] Binke Yuan and Junjie Yang. \"scnu-mandarin-cantonese-dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005139.v1.0.7.", + "[790] Ladan Shahshahani, Maedbh B. King, Caroline Netekoven, Rich B. Ivry, and Jorn Diedrichsen. \"wmfs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005148.v1.1.0.", + "[791] Benjamin Lahner, Kshitij Dwivedi, Polina Iamshchinina, Monika Graumann, Alex Lascelles, Gemma Roig, Alessandro Thomas Gifford, Bowen Pan, SouYoung Jin, N.Apurva Ratan Murty, Kendrick Kay, Radoslaw Cichy*, and Aude Oliva*. \"modeling short visual events through the bold moments video fmri dataset and metadata\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005165.v1.0.4.", + "[792] Andrei Barborica, Felicia Mihai, Laurentiu Tofan, Irina Oane, and Ioana Mindruta. \"dataset of intracranial EEG during cortical stimulation evoking visual effects\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005169.v1.0.0.", + "[793] Tomoyasu Horikawa. \"mind captioning\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005191.v1.0.2.", + "[794] Nicholas G. Cicero, Michaela Klimova, Laura D. Lewis, and Sam Ling. \"differential cortical and subcortical visual processing with eyes shut\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005194.v1.1.0.", + "[795] Jiwoong Park, Hayoung Song, and Won Mok Shim. \"narrativepuzzle\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005215.v1.0.0.", + "[796] Lucas Mahler, Julius Steiglechner, Benjamin Bender, Tobias Lindig, Dana Ramadan, Jonas Bause, Florian Birk, Raphael Heule, Edyta Charyasz, Michael Erb, Vinod Jangir Kumar, Gisela E Hagberg, Pascal Martin, Gabriele Lohmann, and Klaus Scheffler. \"ultracortex: Submillimeter ultra-high field 9.4t brain mr image collection and manual cortical segmentations\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005216.v1.1.0.", + "[797] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"reinforcement-learning generalization\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005230.v1.0.0.", + "[798] Kirill A. Fadeev, Ilacai V. Romero Reyes, Dzerassa D. Goiaeva, Tatyana S. Obukhova, Tatyana M. Ovsiannikova, Andrey O. Prokofyev, Anna M. Rytikova, Artyom Y." + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 37 + }, + { + "type": "page_number", + "text": "23876", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 37 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Novikov, Vladimir V. Kozunov, Tatyana A. Stroganova, and Elena V. Orekhova. \"perception of vowel sounds in children with autism spectrum disorders and typically developing children (meg/erf study)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005234.v2.1.7.", + "[799] J. Veillette and H. Nusbaum. \"motor-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005239.v1.0.1.", + "[800] Taylor Salo, M. Dylan Tisdall, Lia Brodrick, Adam Czernuszenko, David R. Roalf, Sage Rush-Goebel, Nicholas Wellman, Matthew Cieslak, and Theodore D. Satterthwaite. \"single-echo/multi-echo comparison pilot\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005250.v1.1.5.", + "[801] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, and Jean-Michel Badier. \"gloups_meg\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps005261.v2.0.2.", + "[802] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Incentivized reaching\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05263.v1.0.0.", + "[803] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Ultra-high field functional connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005264.v1.0.0.", + "[804] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Approach-avoid\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005265.v1.0.0.", + "[805] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Active escape\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005266.v1.0.0.", + "[806] Anja Ischebeck, Hannah Kreilinger, Joe Miller, Margit Hoefler, Iain D. Gilchrist, and Christof Koerner. \"fixating targets in visual search: The role of dorsal and ventral attention networks in the processing of relevance and rarity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005267.v1.0.1.", + "[807] Hsi T. Wei, Farhan B. Faisal, Tamara Beck, Claire Shao, and Jed A. Meltzer. \"picture-word interference dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005279.v1.0.3.", + "[808] Vassiki Chauhan, Krystal McCook, and Alex White. \"dataset accompanying \"reading reshapes stimulus selectivity in the visual word form area\", chauhan, mccook and white (2024)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005295.v1.0.2.", + "[809] Laura Pritschet, Caitlin M. Taylor, Daniela Cossio, Joshua Faskowitz, Tyler Santander, Daniel A. Handwerker, Hannah Grotzinger, Evan Layher, Elizabeth R. Chrastil, and" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 38 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Emily G. Jacobs. \"maternal brain project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005299.v1.0.0.", + "[810] unknown author. \"neural bases of psychological reactance: An fmri study on dogmatic and suggestive health communication\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005304.v1.0.3.", + "[811] Robert O. Duncan and Evan A. Owens. \"bsc_bids_071724\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005355.v1.0.1.", + "[812] Unknown authors. \"neurogame project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005357.v1.0.0.", + "[813] Elizabeth J. Rizor, Viktoriya Babenko, Neil M. Dundon, Renee Beverly-Aylwin, Alexandra Stump, Margaret Hayes, Luna Herschenfeld-Catalan, Emily G. Jacobs, and Scott T. Grafton. \"hormone health study (hhs)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005360.v1.0.0.", + "[814] Alejandra Figueroa-Vargas, Leonie Kausel, María Paz Martínez-Molina, Francisco Zamorano, Ximena Stecher, Patricia Soto-Icaza, and Pablo Billeke. \"neurocovid mri dwi and fmri with reversal learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005364.v1.0.0.", + "[815] Amy Kartar, Toru Horinouchi, Balázs Örzsik, Brittany Anderson, Lottie Hall, Duncan Bailey, Sarah Samuel, Nati Beltran, Samira Bouyagoub, Chris Racey, Yoko Nagai, Iris Asllani, Hugo Critchley, and Alessandro Colasanti. \"neurobiological substrates of altered state of consciousness induced by high ventilation breathwork\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005365.v1.0.1.", + "[816] Magnus S. Bom, Annette M.A. Brak, Mathijs A.H.L.L. Raemaekers, Nick F. Ramsey, Mariska J. Vansteensel, and Mariana P. Branco. \"large-scale fmri dataset for the design of motor-based brain-computer interfaces\". OpenNeuro, 2025. doi: 10.18112/openneuro.ms005366.v2.0.0.", + "[817] Jessica Archibald, Amy Bouchard, Ralph Noeske, Dikoma Shungu, and Mark Mikkelsen. \"test-retest reliability of multi-metabolite edited mrs at 3t using press and slaser\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005371.v1.1.3.", + "[818] Hyunwoo Gu, Joonwon Lee, Sungje Kim, Jaeseob Lim, Hyang-Jung Lee, Heeseung Lee, Minjin Choe, Dong-Gyu Yoo, Jun Hwan (Joshua) Ryu, Sukbin Lim, and Sang-Hun Lee. \"discrimination-estimation task\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005381.v1.0.0.", + "[819] Chema G. Penalver, David Lopez-Garcia, and Marta Becerra Losada. \"atexp_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005386.v1.0.0.", + "[820] Jeryn Chang, Pamela A McCombe, Robert D Henderson, Diana Zanfirache, Christine C Guo, JingLei Lv, Kelly Garner, Saskia Bollmann, Shyuan T Ngo, Frederik J Steyn, and Thomas B Shaw. \"an fmri dataset for appetite neural correlates in people living with motor neuron disease\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005401.v1.1.1.", + "[821] Grant and None. \"grant dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005412.v1.0.0.", + "[822] Alexander P. Rockhill and Ahmed M. Raslan. \"numbers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005415.v1.0.0." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 38 + }, + { + "type": "page_number", + "text": "23877", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 38 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[823] Antonin Rovai, Valentina Lolli, Xavier De Tiège, Nicola Trotta, and Serge Goldman. \"cerebrovascular reactivity: Co2 inhalation and resting state comparisons\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005418.v3.2.0.", + "[824] Juan Eugenio Iglesias, Adrià Casamitjana, Brain L Edlow, Martina Bocchetta, Dorit Kliemann, Matteo Mancini, Loïc Peter, Eleanor Robinson, Roberto Annunziata, Juri Althonayan, Shauna Crampsie, Emily Blackburn, Benjamin Billot, Alessia Atzeni, Oula Puonti, Yael Balbastre, Peter Schmidt, James Hughes, Jean C Augustinack, Lilla Zöllei, David L Thomas, Catherine Strand, Janice L Holton, and Zane Jaunmuktane. \"nextbrain data extension\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005422.v1.0.1.", + "[825] Juan Pablo Franco, Peter Bossaerts, and Carsten Murawski. \"the neural dynamics associated with computational complexity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005427.v1.1.1.", + "[826] M. R. Ehlers, J. H. Kryklywy, Beukers A. O., Moore S. R., B. J. Forys, Anderson A.K., and R. M. Todd. \"valenced tactile information is evoked by neutral visual cues following emotional learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005449.v1.0.0.", + "[827] Emilie C. Sleight, Oscar Esteban, and Frédéric Grouiller. \"fmriprep boot camp - 7t training dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005454.v1.0.0.", + "[828] Tingting Guo, Xuedi Liu, Mo Chen, Yongben Fu, and Taomei Guo. \"an fmri dataset for investigating language control and cognitive control in bilinguals\". OpenNeuro, 2024. doi: 10.18112/openneuro.dS005455.v1.1.5.", + "[829] Wangjing Yu, Asieh Zadbood, Avi J. H. Chanales, and Lila Davachi. \"priority\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005464.v1.0.1.", + "[830] Antonin Rovai and Emilie Caspar. \"sense of agency in free and coerced moral decision-making among civilians and military personnel\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005469.v2.0.0.", + "[831] Patrycja Scislewska, Michal Rafal Zareba, Julia Lengier, Aaron Schirmer, Piotr Bebas, and Iwona Szatkowska. \"monetary incentive delay task - structural and functional images of 37 men; study of associations between circadian characteristics (eveningness, distinctness) and affective processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005479.v1.0.3.", + "[832] Unknown authors. \"neurowater project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005492.v1.0.0.", + "[833] Niharika Gajawelli, Cameron Glick, Yinming Sun, Faizan Badami, Amit Etkin, and Manish Saggar. \"single-pulse tms fmri\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005498.v2.0.0.", + "[834] Zoya Surani, Ted Turesky, Eileen Sullivan, Talat Shama, Rashidul Haque, Nazrul Islam, Shahriah Hafiz Kakon, Xi Yu, William Petri, Charles Nelson, and Nadine Gaab. \"psychosocial adversity and inhibitory control: an fmri study of children growing up in extreme poverty\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005504.v1.0.0.", + "[835] Ioannis Ntoumanis, Julia Sheronova, Alina Davydova, Maria Dolgaleva, Iiro P Jäaskeläinen, Vladimir Kosonogov," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 39 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Anna N Shestakova, and Vasily Klucharev. \"deciphering the neural responses to a naturalistic persuasive message\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005518.v1.0.1.", + "[836] Xingfeng Shao, Qinyang Shou, Kimberly Felix, Brandon Ojogho, Xuejuan Jiang, Brian T Gold, Megan M Herting, Eric L Goldwaser, Peter Kochunov, L Elliot Hong, Ioannis Pappas, Meredith Braskie, Hosung Kim, Steven Cen, Kay Jann, and Danny JJ Wang. \"dp-pcasl data (cbf, att, bbb kw) from 186 cognitively normal participants (8-92 years)\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005529.v1.0.1.", + "[837] Viviana Greco, Tamas A. Foldes, Neil A. Harrison, Kevin Murphy, Marta Wawrzuta, Mahmoud E. A. Abdellahi, and Penelope A. Lewis. \"depotentiation of emotional reactivity using tmr during rem sleep\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005530.v1.0.9.", + "[838] Hiroto Q. Yamaguchi, Naoko Koide-Majima, Rieko Kubo, Tomoya Nakai, and Shinji Nishimoto. \"narrative movie fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005531.v1.0.0.", + "[839] Nicholas G. Cicero, Nina E. Fultz, Hongbae Jeong, Stephanie D. Williams, Daniel Gomez, Beverly Setzer, Tracy Warbrick, Manfred Jaschke, Ravij Gupta, Michael Lev, Giorgio Bonmassar, and Laura D. Lewis. \"high-quality multimodal mri with simultaneous eeg using conductive ink and polymer-thick film nets\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005533.v1.0.0.", + "[840] Zsuzsanna Nemecz and Attila Keresztes. \"item and spatial pattern separation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005559.v1.0.1.", + "[841] María Paz Martínez-Molina, Alejandra Figueroa-Vargas, Francisco Zamorano, and Pablo Billeke. \"expectation of conflict stimuli\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005571.v1.0.1.", + "[842] Jarod L Roland, Carl D Hacker, Abraham Z Snyder, Joshua S Shimony, JOhn M Zempel, David D Limbrick, Matthew D Smyth, and Eric C Leuthardt. \"a comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005573.v1.0.0.", + "[843] Toshiki Okumura, Ikuhiro Kida, Atsushi Yokoi, Shinji Nishimoto, Kazushige Touhara, and Masako Okamoto. \"dataset for 7t odor fmri experiment\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005576.v1.0.0.", + "[844] Alexa Tompary and Lila Davachi. \"integration of overlapping sequences emerges with consolidation through mpfc neural ensembles and hippocampal-cortical connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005581.v1.0.0.", + "[845] Johannes Schultz, Maria Gädeke, and Tom Willems. \"sodec - responsibility fmri experiment dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005588.v1.0.0.", + "[846] Yaqiong Chai, Hedong Zhang, Carlos Robles, Andrew Shinho Kim, Kyung-Wook Kang, Jiyoung Kim, JungWon Shin, Abigail Trang, Jongmok Ha, Emily Lee, Yeonsil Moon, and Hosung Kim. \"precise perivascular space" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 39 + }, + { + "type": "page_number", + "text": "23878", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 39 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "segmentation on t2-weighted magnetic resonance imaging from human connectome project-aging\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005595.v1.0.0.", + "[847] Bolong Wang, XuanXuan Zhang, and Xiangzhen Kong. \"a naturalistic fmri dataset in response to public speaking\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005596.v1.1.", + "[848] EC Washington University School of Medicine Leuthardt, JS Washington University School of Medicine Shimony, AZ Washington University School of Medicine Snyder, D Washington University School of Medicine Dierker, and KY Washington University School of Medicine Park. \"retrospective task/rest fmri data from tumor patients - for sora\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005597.v1.0.1.", + "[849] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"visuomotor rotation learning and reward-based motor learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005598.v1.0.0.", + "[850] Xitong Chen, Stephanie Leach, Juniper Hollis, Dillan Cellier, and Kai Hwang. \"thalamocortical contributions to hierarchical cognitive control\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005600.v1.1.0.", + "[851] Peter N. Taylor, Yujiang Wang, Callum Simpson, Vytene Janiukstyte, Jonathan Horsley, Karoline Leiberg, Beth Little, Harry Clifford, Sophie Adler, Sjoerd B. Vos, Gavin P Winston, Andrew W McEvoy, Anna Miserocchi, Jane de Tisi, and John S Duncan. \"the imaging database for epilepsy and surgery (ideas)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005602.v1.0.0.", + "[852] Tim Mitchell and Carl Hacker. \"mitchell_hacker_2013\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005603.v1.0.1.", + "[853] Unknown authors. \"v4 crowding\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005604.v1.0.1.", + "[854] Xuefeng Yan, Fabrice G. Simeon, Jeih-San Liow, Cheryl L. Morse, Susovan Jana, Jose A. Montero Santamaria, Madeline Jenkins, Isabella Piacentino, Anthony Galassi, Sami S. Zoghbi, Victor W. Pike, Robert B. Innis, and Zanotti-Fregonara. \"[18f]sf51, a novel 18f-labeled pet radioligand for translocator protein 18kda (tspo) in brain, works well in monkeys but fails in humans\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005619.v1.1.0." + ], + "bbox": [ + 91, + 92, + 482, + 696 + ], + "page_idx": 40 + }, + { + "type": "page_number", + "text": "23879", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 40 + } +] \ No newline at end of file diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_model.json b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce5a6032c8455c719fc9168e81f1b787e263fdcc --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_model.json @@ -0,0 +1,12228 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "An OpenMind for 3D medical vision self-supervised learning" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.18, + 0.764, + 0.234 + ], + "angle": 0, + "content": "Tassilo Wald \\(^{*,1,2,3}\\), Constantin Ulrich\\(^{*,1,4}\\), Jonathan Suprijadi\\(^{*,†}\\), Sebastian Ziegler\\(^{1,2}\\), Michal Nohel\\(^{\\dagger,6,7}\\), Robin Peretzke\\(^{1,4}\\), Gregor Kohler\\(^{\\dagger,1}\\), Klaus Maier-Hein\\(^{1,2,3,4,5,8}\\)" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.236, + 0.812, + 0.252 + ], + "angle": 0, + "content": "1 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.254, + 0.802, + 0.27 + ], + "angle": 0, + "content": "\\(^{2}\\) Helmholtz Imaging, DKFZ; \\(^{3}\\) Faculty of Mathematics and Computer Science, Heidelberg University" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.272, + 0.794, + 0.288 + ], + "angle": 0, + "content": "4 Medical Faculty, Heidelberg University; 5 National Center for Tumor Diseases (NCT), Heidelberg" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.29, + 0.809, + 0.305 + ], + "angle": 0, + "content": "\\(^{6}\\) Faculty of Electrical Engineering and Communication, Brno University of Technology, Czech Republic" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.307, + 0.785, + 0.322 + ], + "angle": 0, + "content": "7 University hospital Ostrava, Department of Deputy director for science, research and education" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.324, + 0.723, + 0.357 + ], + "angle": 0, + "content": "\\(^{8}\\) Pattern Analysis and Learning Group, Department of Radiation Oncology tassilo.wald@dkfz-heidelberg.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.392, + 0.327, + 0.409 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.424, + 0.486, + 0.726 + ], + "angle": 0, + "content": "The field of self-supervised learning (SSL) for 3D medical images lacks consistency and standardization. While many methods have been developed, it is impossible to identify the current state-of-the-art, due to i) varying and small pretraining datasets, ii) varying architectures, and iii) being evaluated on differing downstream datasets. In this paper, we bring clarity to this field and lay the foundation for further method advancements through three key contributions: We a) publish the largest publicly available pre-training dataset comprising \\(114k\\) 3D brain MRI volumes, enabling all practitioners to pre-train on a large-scale dataset. We b) benchmark existing 3D self-supervised learning methods on this dataset for a state-of-the-art CNN and Transformer architecture, clarifying the state of 3D SSL pre-training. Among many findings, we show that pre-trained methods can exceed a strong from-scratch nnU-Net ResEnc-L baseline. Lastly, we c) publish the code of our pre-training and fine-tuning frameworks and provide the pre-trained models created during the benchmarking process to facilitate rapid adoption and reproduction. Available here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.222, + 0.769 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.485, + 0.856 + ], + "angle": 0, + "content": "Self-Supervised Learning (SSL) has emerged as a transformative approach in deep learning, enabling the extraction of robust and general representations from data-rich domains [1-3]. This paradigm has led to significant advancements in areas with abundant data, such as natural language process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.908, + 0.635 + ], + "angle": 0, + "content": "ing and natural image computer vision. In the domain of 3D medical imaging, where annotation is expensive and annotated data scarce, SSL pre-training holds great promise as it can learn transferable representations without requiring extensive manual labeling efforts. However, currently the field of medical image analysis follows two predominant approaches to develop vision models: training models from scratch, often using the nnU-Net framework [4], or adopting supervised pre-training with large, monolithic annotated datasets [5-7] or aggregates of many datasets to increase supervised data scale [8, 9]. While the latter approach indicates a willingness to adopt pre-training paradigms, it raises the question why self-supervised learning is hardly pursued in 3D medical imaging. We believe this lack of SSL method development and adoption can be attributed to two major factors:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "1) A lack of a large, open-source 3D dataset for pretraining. Acquiring large unlabeled datasets for self-supervised pre-training is a non-trivial task. While large studies like the UK-BioBank or the NIH Adolescent Brain Cognitive Development (ABCD) exist and hold \\(>100\\mathrm{k}\\) or \\(>40\\mathrm{k}\\) 3D Volumes respectively, access to them depends on a positive internal review process requiring an appropriate project proposal. To add to the complexity, datasets are often subject to unique Data Usage Agreements (DUAs) that impose various restrictions on recipients. These rules may include: (i) mandatory acknowledgments, (ii) internal administrative review of manuscripts prior to submission [10-14], (iii) inclusion of a consortium in the author list [12-14], or (iv) requiring the dataset's name to appear in the manuscript title [13]. While acknowledgements are reasonable and exemptions for the remaining restrictions can be requested, these DUAs create legal barriers that hinder re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.887 + ], + "angle": 0, + "content": "*Equal contribution, Author order among the co-first authors may be adjusted for individual use." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.262, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done while at DKFZ." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.475, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.378 + ], + "angle": 0, + "content": "Figure 1. The OpenMind dataset contains 114k 3D Head-And-Neck volumes of 23 different modalities. It represents the largest openly accessible dataset of 3D medical images currently available." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.519 + ], + "angle": 0, + "content": "production or creation of large-scale pre-training datasets. Consequently, the majority of current SSL methods are either developed on large, restricted or proprietary datasets [15-17] or are developed on small-scale publicly available datasets [18-24], leading to the current fragmented landscape of pre-training datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.695 + ], + "angle": 0, + "content": "2) A lack of comparability: To adopt existing pre-trained models or conduct one's own SSL pre-training, it is important to know which model or pre-training method is the best. Currently, this is an impossible task, due to: i) Different pre-training dataset selection (as previously highlighted). ii) Different architecture choices. Some methods use CNNs for pre-training [15, 17, 21], some ViT-like Transformers [18, 19], others Swin Transformers [22, 23] and even others hybrid architectures [16, 24]. iii) Insufficient downstream dataset choices, potentially providing unreliable results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper we address these issues and provide three key contributions: i) We share the largest publicly available 3D dataset, comprising 114k 3D volumes of 23 different brain MRI modalities. ii) We conduct a comprehensive self-supervised learning benchmark leveraging this dataset to assess the performance of state-of-the-art CNN and transformer-based architectures on 15 segmentation and classification downstream datasets. iii) We open-source our pre-training and fine-tuning frameworks as well as the trained model checkpoints to enable rapid reproduction and novel method development in the field of 3D medical image segmentation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Table 1. Large medical datasets are often restricted through custom DUAs limiting their overall applicability." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.129, + 0.887, + 0.242 + ], + "angle": 0, + "content": "
DatasetPatients3D VolumesModalitiesAccess
ABCD [25]11,38541,5802Restricted
PPMI [27]2,0119,5545Restricted
ADNI [28]1,5639,5483Restricted
OASIS31,37615,8727Restricted
OASIS46614,0215Restricted
BraTS21 [29]1,2515,0044CC-BY-4.0
CT-RATE [25]21,30450,1882CC-BY-NC-4.0
OpenMind34,191113,92124CC-BY-4.0
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.264, + 0.737, + 0.281 + ], + "angle": 0, + "content": "2. The OpenMind Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.456 + ], + "angle": 0, + "content": "To provide practitioners access to a large-scale dataset for self-supervised learning, we introduce the OpenMind dataset: a large-scale, easily accessible resource containing 114k Head-and-Neck MRI volumes across various modalities, sourced and curated from the OpenNeuro platform. As the largest freely accessible 3D brain imaging dataset to date, OpenMind surpasses the NIH Adolescent Brain Cognitive Decline initiative [25] by a factor of three, and exceeds the recent CT-Rate dataset [26] by a factor of two, while being released under a permissive CC-BY license (see Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.676, + 0.479 + ], + "angle": 0, + "content": "2.1. Dataset creation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.88 + ], + "angle": 0, + "content": "All data was sourced or derived from datasets available on the OpenNeuro platform [30]. OpenNeuro (formerly known as openfMRI) is a publicly accessible data repository designed to support neurological research, adhering to the FAIR principles (Findable, Accessible, Interoperable, Reusable). The platform currently hosts over 1,200 public datasets, encompassing data from more than 51,300 subjects. Due to the diverse nature of its collection, OpenNeuro includes a wide array of modalities, such as MRI, PET, fMRI, MEG, EEG, and iEEG, from studies involving both healthy and diseased subjects, enabling research across various aspects of neurology. From this extensive collection, all available 3D MRIs, such as T1-weighted and T2-weighted scans, as well as 4D diffusion-weighted MRIs were collected from 800 different studies (citations [55-854]). This aggregation resulted in a total of 71k 3D MRI scans1 and 15k 4D diffusion-weighted images from a total of 34,139 subjects. While existing 3D SSL methods could be directly trained on the 71k 3D MRI scans spanning multiple MRI sequences, the 4D diffusion-weighted images required pre-processing to create 3D derivatives which we describe in the following section. Overall, due to the high amount of independent studies, the OpenMind dataset provides great levels of diversity regarding participant age distribution, countries of origin, MRI modalities, scanner manufacturer, model and scanning protocols." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.768, + 0.902 + ], + "angle": 0, + "content": "This includes a minority of 653 PET scans." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23840" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "DWI Preprocessing Diffusion-weighted imaging (DWI) is an advanced MRI modality that captures the random motion of water molecules in tissue by applying diffusion-sensitive gradient pulses, enabling the assessment of microstructural properties. It is widely used for detecting abnormalities such as strokes, tumors, and white matter changes. In clinical and research settings, DWI data is often preprocessed into more interpretable 3D derivatives, including mean diffusion (MD) maps, and fractional anisotropy (FA) maps, which simplify analysis and interpretation. Additionally, T2-weighted images can be derived from DWI scans. Consequently, we pre-processed the DWI images into MD maps, FA maps, and T2-weighted images, resulting in an additional \\(43\\mathrm{k}\\) 3D images. The detailed preprocessing pipeline is described in the appendix in B.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.482, + 0.534 + ], + "angle": 0, + "content": "Defacing Protecting patient privacy is a big concern in the medical domain, hence, the facial region of Head-and-Neck scans are commonly anonymized. Common anonymization techniques include defacing – replacing the volume with a constant value –, blurring everything but the brain or brain extraction, see Figure 2 in the appendix. These augmented regions can interfere with reconstruction-based SSL methods, potentially penalizing models for attempting to reconstruct plausible anatomical features in anonymized regions. Therefore, masks delineating these regions are needed to allow excluding them from the loss computation. Unfortunately, not all datasets provide these masks, despite applying some sort of anonymization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.482, + 0.7 + ], + "angle": 0, + "content": "To address this issue, we generate associated anonymization masks and anatomy masks for cases where these are not provided (Figure 2 bottom in the appendix). The anonymization mask identifies areas where anatomical structures have been artificially modified or removed and where no loss should be calculated, whereas anatomy masks indicate where relevant foreground is present, allowing to avoid sampling non-empty patches which may speed up the convergence of the pre-training methods. Both of these masks are created using the model proposed by Nohel et al. [31]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Meta-Data enrichment While OpenNeuro follows the Brain Imaging Data Structure (BIDS) format, metadata organization varies significantly across datasets. Participant demographics (e.g., age, sex, health status) and image acquisition details (e.g., repetition times, scanner types, imaging protocols) are inconsistently reported, creating challenges for systematic analysis. We curated and harmonized all available, disparate metadata into a unified, complementary metadata structure. This standardized framework ensures consistency and enables users to efficiently filter the OpenMind dataset based on specific selection criteria, such as subject characteristics or imaging" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "parameters, enhancing the flexibility of the dataset and simplifying the extraction of relevant sub-populations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.424 + ], + "angle": 0, + "content": "In addition to the harmonized metadata, we provide an Image Quality Score (IQS) for each imaging modality of each of the 800 datasets. This score represents a subjective usability score of the images for self-supervised pretraining. The assessment considers factors such as image noise levels, motion artifacts, and other image quality-degrading elements, condensing these evaluations into a unified score ranging from 1 to 5 with 1 representing the highest quality images. The quality scores were determined through manual inspection of multiple images from each modality in each dataset and were assessed independently by two raters to ensure reliability (more details in the appendix B.1). Together, the harmonized metadata and Image Quality Score are introduced to test the effects of simple data-filtering techniques on self-supervised methods trained on the OpenMind dataset. If significant effects of pre-training data choice can be measured, the OpenMind datasets would show to have sufficient size to enable data-centric research in the domain of self-supervised learning for 3D medical image analysis, akin to [2, 32, 33]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.439, + 0.905, + 0.514 + ], + "angle": 0, + "content": "Useability To maximize the utility of the OpenMind dataset, we provide the final dataset of 114k 3D images on Hugging Face together with anonymization and anatomy masks, as well as the harmonized metadata including the Image Quality Score." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.635 + ], + "angle": 0, + "content": "Why benchmarking matters While the OpenMind dataset addresses the issue of data accessibility and data scale in self-supervised learning for 3D medical imaging, the current state-of-the-art remains unknown. To establish a reference point and provide an initial set of baselines, we conduct a comprehensive benchmarking effort to compare existing 3D SSL methods within a common framework." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.771, + 0.666 + ], + "angle": 0, + "content": "3. The OpenMind Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.905, + 0.824 + ], + "angle": 0, + "content": "The field of self-supervised learning (SSL) for 3D medical imaging remains fragmented due to inconsistencies in pre-training datasets, architectures, and downstream evaluations. Without standardized benchmarks, it is impossible to fairly compare methods and determine the most effective approaches [15]. The OpenMind Benchmark addresses this challenge by evaluating a broad range of SSL methods under controlled conditions, using a common pre-training dataset, standardized architectures, and diverse downstream tasks, which we detail in this section." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.834, + 0.673, + 0.849 + ], + "angle": 0, + "content": "3.1. Standardization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To establish a standardized and fair evaluation, all self-supervised learning (SSL) methods trained in the benchmark were trained on the OpenMind dataset, ensur" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23841" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "ing that pre-training dataset variability does not influence results. Further, we use two state-of-the-art architectures, the ResEnc-L Residual Encoder U-Net [34] to represent convolutional neural networks (CNNs) and the recent Primus-M Transformer [35], allowing us to analyze SSL effectiveness across different architectural paradigms. Pre-trained models are then evaluated on a diverse set of downstream datasets for segmentation and classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.243, + 0.485, + 0.469 + ], + "angle": 0, + "content": "Downstream Datasets All downstream datasets used are split 50/50 into a train and test set. Moreover, we group our datasets into development datasets and test datasets. The development datasets serve to optimize SSL method hyperparameters on their train split, should the method not be directly applicable on the OpenMind dataset with original hyperparameters. During this process, the most important method-specific hyperparameters as well as learning rate and weight decay are optimized based on an additional 80/20 split on the training set of the development datasets. Aside from this, the development datasets are used to determine the optimal fine-tuning schedules, hence results of the optimal fine-tuning schedules are reported on the validation split of the dataset (Tab. 4 and in the appendix in table 8)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.621 + ], + "angle": 0, + "content": "To avoid potential biases introduced during the optimization process on the development datasets, we withheld additional test datasets, which we left untouched and only evaluated after finishing all method optimizations on the development datasets. Moreover, since datasets tend to be small and targets diverse, results are inherently noisy and may not generalize from one dataset to another. To reduce the chance of over optimizing for one dataset, we include four development datasets as well as eight test segmentation datasets and three test classification datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.253, + 0.673 + ], + "angle": 0, + "content": "Development datasets:" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.675, + 0.483, + 0.703 + ], + "angle": 0, + "content": "1. Atlas R2.0 (ATL) [36]: Brain scans of stroke patients imaged through T1 weighted MRIs." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.705, + 0.484, + 0.763 + ], + "angle": 0, + "content": "2. Stanford Brain Metastasis (SBM) [37]: Brain scans of metastasis imaged through pre-contrast T1, contrast-enhanced gradient echo, contrast-enhanced spin echo and a T2 FLAIR sequence." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.765, + 0.483, + 0.794 + ], + "angle": 0, + "content": "3. AMOS22 (AMO) [38]: CT and MRI scans of 15 abdominal organs." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.795, + 0.482, + 0.823 + ], + "angle": 0, + "content": "4. KiTS19 (KIT) [39]: CT images of kidneys and kidney tumors." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.675, + 0.484, + 0.823 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "While AMOS and KiTS can be considered out-of-distribution due to the pre-training images exclusively depicting the head-and-neck region, these datasets were found to be robust benchmarking datasets in [34], hence were included despite their shift. Further, during develop" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "ment, we only optimized the methods for segmentation, due to it being the most common task in 3D medical image analysis [35]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.152, + 0.722, + 0.166 + ], + "angle": 0, + "content": "Test datasets (Segmentation):" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.168, + 0.905, + 0.197 + ], + "angle": 0, + "content": "1. ISLES (ISL) [40]: DWI \\((\\mathrm{b} = 1000)\\) and ADC map brain MRI scans of patients with delineations of stroke lesions." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.199, + 0.905, + 0.258 + ], + "angle": 0, + "content": "2. HNTS-MRG (HNT) [41]: MR images of patients with head-and-neck primary tumors and metastatic lymphnodes pre and mid-treatment of which we use the pretreatment images and labels." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.259, + 0.905, + 0.303 + ], + "angle": 0, + "content": "3. HaN-Seg OAR (HAN) [42]: CT and MR images of Head-and-Neck organs-at-risk of which we use only the MR images." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.304, + 0.904, + 0.333 + ], + "angle": 0, + "content": "4. MS FLAIR (MSF) [43]: FLAIR MRIs of hyperintense multiple-sclerosis lesions with associated delineations." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.335, + 0.905, + 0.378 + ], + "angle": 0, + "content": "5. ToP-CoW (TPC) [44]: Time-of-Flight (ToF) Magnetic Resonance Angiography (MRA) of the Circle of Willis and associated annotations of the arterial structure." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.379, + 0.905, + 0.409 + ], + "angle": 0, + "content": "6. Yale Brain Metastasis (YBM) [45]: Pre and post-contrast T1, T2 and FLAIR MRI images of brain metastases." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.41, + 0.905, + 0.453 + ], + "angle": 0, + "content": "7. Cosmos (COS) [46]: Carotid vessel wall segmentations and atherosclerosis diagnosis, with contours of the vessel walls on 3D-VISTA MRI." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.455, + 0.905, + 0.484 + ], + "angle": 0, + "content": "8. ACDC (ACD) [47]: Cardiac cine-MRI image of three segmented ventricular structures." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.168, + 0.905, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.486, + 0.72, + 0.5 + ], + "angle": 0, + "content": "Test datasets (Classification):" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.502, + 0.905, + 0.546 + ], + "angle": 0, + "content": "1. MR-Net (MRN) [48]: Knee T1w, T2w and Proton Density (PD) weighted MRIs of ACL or meniscal tears as well as other abnormalities which we classify." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.547, + 0.905, + 0.591 + ], + "angle": 0, + "content": "2. RSNA-SpineFrac (RSN) [49]: CT Spine fractures with the goal of identifying lumbar spine degenerative conditions. We classify fracture and no-fracture." + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.592, + 0.905, + 0.637 + ], + "angle": 0, + "content": "3. ABIDE (ABI) [50]: T1w MRI images of either healthy controls or subjects with autism, on which we conduct binary classification - autism or not-autism." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.502, + 0.905, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Pre-training configuration To ensure a fair comparison, all SSL methods are trained under comparable pre-training and equal fine-tuning schedules. Specifically, we pre-train all methods for a total of 1000 epochs with 250 steps per epoch each, with each method trained DDP on four 40GB A100s to decrease pre-training time. Furthermore, the pretrained models were fine-tuned for a total amount of 150 or 1000 epochs with 250 steps per epoch and a common batch size of 2 for both architectures following an identical learning rate schedule implemented within the nnU-Net framework [4] for the segmentation downstream datasets. For classification downstream datasets, all models were fine-tuned for 200 dataset epochs within the public Image classification framework, which supports 3D classification. We provide more detailed descriptions of the fine-tuning process and frameworks in the appendix C." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23842" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.266, + 0.108 + ], + "angle": 0, + "content": "3.2. Benchmark scope" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.483, + 0.158 + ], + "angle": 0, + "content": "In the benchmark, three key aspects of SSL are investigated, highlighting the possible research directions the OpenMind dataset and benchmark can enable:" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.16, + 0.483, + 0.25 + ], + "angle": 0, + "content": "i) SSL method performance: We compare the efficacy of various SSL methods in learning transferable representations to various downstream segmentation and classification datasets when using the same pretraining dataset, using two common architectures and a unified fine-tuning schedule." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.251, + 0.483, + 0.31 + ], + "angle": 0, + "content": "ii) Fine-tuning methodology We evaluate various fine-tuning schedules for both architectures to investigate the importance of a well-configured downstream fine-tuning schedule." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.311, + 0.483, + 0.37 + ], + "angle": 0, + "content": "iii) Data-centric baselines We evaluate the influence of data-filtering on the large OpenMind dataset, to evaluate if data-centric research can be conducted on this scale." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.16, + 0.483, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.794 + ], + "angle": 0, + "content": "SSL Methods Various SSL methods native to 3D medical images exist. In the benchmark, we include contrastive methods, like Volume Contrastive (VoCo) [22], a 3D SimCLR adaptation, the pseudo-supervised VolumeFusion (VF) [16], default masked autoencoders (MAE) [3], sparse adaptations of them for CNNs (S3D) [15], masked image modeling (SimMIM) [18] and, the denoising and masked auto-encoder combination ModelsGenesis (MG) [21]. Moreover, we incorporate the SwinUNETR pre-training scheme (SwinUNETR) [23], which integrates masked inpainting, contrastive learning and rotation prediction2. We tried including the 3D geometric-matching-based pre-training method GVSL [20], but the method failed to converge both when using the original implementation and when reimplementing it within our framework, regardless of whether we trained it on our OpenMind dataset or the ABCD-NIH dataset. We provide a brief description of each included SSL method and its used hyperparameters in the appendix D. Methods, which did not feature publicly available Code that could be tested for functionality or integrated into our repository, were excluded, e.g. GLMAE [19]. Moreover, more recent self-supervised methods were given higher priority for inclusion in our benchmark than older methods. Furthermore, prominent 2D methods were excluded, as adapting their configuration from the natural imaging domain to the 3D medical imaging domain would require substantial development effort, e.g. due to large batch size requirements leading to VRAM issues." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Fine-tuning Schedules Effective adaptation during fine-tuning is crucial for optimizing performance in medical image segmentation, as it determines how well a pre-trained" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.378 + ], + "angle": 0, + "content": "model can transfer knowledge to a target dataset. We test five different fine-tuning strategies for adaptation to segmentation datasets, aiming to balance maintaining general pre-trained representations with adapting the encoder to specialized representations through varying learning rate (lr) schedules: i) nnU-Net's default training strategy employs a polynomial decay lr with an initial lr of \\(1\\mathrm{e} - 2^3\\), which we reduce to 1e-3 for the transfer learning setting. ii) The Frozen schedule follows the same lr schedule but restricts training to the decoder. iii) The Warm-Up schedule introduces an initial phase with a linear increase in the lr before transitioning to the default schedule. iv) The Valley scheme prioritizes decoder adaptation by first training it with a linear decreasing lr, followed by a linear warm-up phase for the full network, and then the default schedule. v) Finally, the Sawtooth scheme employs a two-stage warm-up: first for the decoder while keeping the encoder frozen, and then for the whole network, both with a linear increasing lr, before transitioning to the default schedule." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.5 + ], + "angle": 0, + "content": "We train the majority of fine-tuning runs for 150 epochs to reduce computational cost. For these, each warm-up stage lasts 15 epochs. For the CNNs we provide additional fine-tuning experiments trained for 1000 epochs, in which the warm-up stage is extended to 50 epochs. For a more detailed description and visualization of this, we defer to the appendix C, where we also provide additional information on the classification fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.705 + ], + "angle": 0, + "content": "Dataset Filters When training on a large enough quantity of data, optimizing the dataset becomes an important aspect that can greatly improve the resulting model performance [2, 32, 33]. Given the scale and diversity of the OpenMind dataset, we evaluate four simple filtering schemes to measure if we are at a scale where data filtering can be effective. The filtering schemes implemented are: 1. Filtering by the manually created Image Quality Score (IQS) at three thresholds, and 2. filtering all modalities but T1w, T2w and FLAIR images, the most common downstream dataset modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.905, + 0.851 + ], + "angle": 0, + "content": "Reproducibility While evaluating this benchmark, a lot of pre-trained models were generated. To maximize reproducibility, we provide the checkpoints of all pre-training methods for both architectures. Further, we provide an integration into the nnU-Net framework that holds the provided fine-tuning code, enabling users to quickly adapt the pretrained checkpoints to a desired downstream task." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2We refer to the pre-training scheme as SwinUNETR, not to be confused with the identically named architecture." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3Learning rate value of the ResEnc-L CNN only. Primus-M learning rate values are provided in the appendix C" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23843" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 2. Segmentation results of tested pre-training methods. CNN and Transformer architectures were pre-trained on the same dataset for the same amount of steps. All methods were fine-tuned with the optimal architecture specific fine-tuning scheme for 150 epochs, unless indicated by suffix \\(1k\\). Color ranges are set per-dataset and per-architecture. S3D and SimMIM are only shown for CNN or Transformer as they are architecture specific. nnU-Net def.: Dynamically planned default nnU-Net trained from scratch; 1k: Fine-tuned for 1000 epochs instead of 150 epochs; ID Mean: In distribution mean; OOD Mean: Mean across datasets of non Head-and-Neck datasets." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.171, + 0.885, + 0.452 + ], + "angle": 0, + "content": "
PT MethodDice Similarity Coefficient (DSC) [%] on ...
Dataset of same anatomical region (ID)Dataset of OOD regionAverage across ...
ATLSBMISLHNTHANMSFTPCYBMCOSACDAMOKITIDOODAll
nnU-Net def. 1k58.7059.9878.4062.9853.3752.1979.5058.4346.1991.1088.0087.2161.0888.7768.00
nnU-Net def.56.0860.4178.2259.3732.2755.8476.7856.7349.3190.7283.8877.6158.3384.0764.77
ResEnc-L (CNN)
Scratch 1k58.2153.4379.1465.7558.2454.9079.9456.1271.5792.0988.7387.4864.1589.4370.47
Scratch57.0254.2978.0963.3056.1155.4776.1854.4265.2091.9785.2484.0362.2387.0868.44
VoCo57.1459.6277.5063.4851.1254.9075.1256.9263.4991.4485.6085.7162.1487.5868.50
SwinUNETR56.0757.2577.4561.6449.4254.8274.6957.0565.6890.5384.9585.5461.5687.0167.92
SimCLR57.1559.7278.0163.3251.5655.6877.7759.1468.2091.7686.0684.8563.4087.5669.44
VF57.4259.8878.1864.3251.6757.4276.1159.3163.9891.5785.3886.2163.1487.7269.29
MG58.0361.5777.5865.1154.6955.2577.1458.6771.2791.7486.3586.1764.3788.0970.30
MAE58.2562.4177.8966.5855.1456.8477.9660.0770.8591.9886.7886.1265.1188.3070.91
S3D58.7664.0978.0565.7452.8156.0878.8159.1866.6692.0186.1686.0164.4688.0670.36
Primus-M (Transformer)
Scratch 1k56.7748.5076.5958.4053.4053.2776.3252.5364.6890.8987.2485.5760.0587.9067.01
Scratch51.5143.2675.2355.3050.6054.0073.3150.3062.1190.9380.1776.7357.2982.6163.62
VoCo46.8034.1573.2951.0647.6452.6465.5244.7552.1687.2665.8170.2152.0074.4357.61
SwinUNETR47.2336.3173.8450.1546.4952.8066.2344.4954.8287.9266.1770.3952.4974.8258.07
SimCLR54.6142.6275.4356.7550.8053.5970.0848.3658.2089.9775.7581.2756.7282.3363.12
VF58.6247.3777.5662.3756.1855.0074.9853.9669.7491.4184.9586.1761.7587.5168.19
MG56.5047.3476.7658.4254.0254.6773.6549.7760.7490.8982.1584.4559.1085.8365.78
MAE61.1656.6777.1266.1257.2456.0278.3154.3572.0292.1687.1686.7464.3488.6970.42
SimMIM60.2851.6877.5362.7656.7455.9177.0052.9070.8791.9886.5785.9262.8588.1669.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.463, + 0.907, + 0.491 + ], + "angle": 0, + "content": "Table 3. Classification results of tested pre-training methods. CNN and Transformer architectures were pre-trained and fine-tuned identically on the three classification datasets. PT Method: pre-training method" + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.502, + 0.885, + 0.628 + ], + "angle": 0, + "content": "
Architecture Metric PT MethodResEnc-L (CNN)Primus-M (Transformer)
Average PrecisionMeanMRNBalanced AccuracyAverage PrecisionMeanMRNBalanced Accuracy
MRNRSNABIRSNABIMeanRSNABIMean
Scratch66.3365.3957.1862.9776.6162.2553.2964.0560.5858.9152.8757.4575.4957.1052.0561.55
VoCo71.9970.3362.2368.1880.3565.5760.1868.7067.9262.0953.2261.0877.6861.6451.2863.53
SwinUNETR70.7270.3762.3267.8079.2566.6059.8868.5864.9860.7956.8460.8776.6159.2255.6063.81
SimCLR68.7473.0259.1766.9878.2767.9657.6667.9667.6760.2655.5961.1876.6959.5254.1563.45
VF68.4666.2258.0164.2378.4363.2855.6065.7766.4161.6658.3762.1476.6760.1456.0664.29
MG68.8571.2159.6966.5979.0966.1457.2467.4968.8661.2452.4160.8477.7359.9754.1363.94
MAE64.5367.8658.3163.5776.7264.2156.2365.7267.0060.2853.0960.1377.2658.8351.8462.64
S3D70.0869.2560.6566.6678.5664.2056.4166.39--------
SimMIM--------65.2759.6253.9459.6177.1558.4953.9363.19
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.178, + 0.668 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.679, + 0.483, + 0.725 + ], + "angle": 0, + "content": "In this section we highlight the most important insights of our benchmark, but not all due to spatial constraints. A comprehensive list of them is provided in the appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.325, + 0.758 + ], + "angle": 0, + "content": "4.1. SSL method performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Segmentation performance We compare the fine-tuning results (fine-tuned for 150 epochs) of all pre-training methods for the ResEnc-L CNN architecture and the Primus-M Transformer architecture in Tab. 2. For both architectures, we observe that self-supervised pre-training is able to exceed both, a from-scratch nnU-Net default baseline trained for 1000 epochs and their respective from-scratch trained architecture trained for 150 epochs. In particular, reconstruction based methods exhibit the best performance," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.759 + ], + "angle": 0, + "content": "with default MAEs leading the pack. The MAE pre-trained CNN model even exceeds its strong from scratch baseline trained for 1000 epochs in only 150 epochs, and performs even better when fine-tuned for 1000 epochs (see Table 5 in the appendix). Aside from the MAE pre-training paradigm, this is only achieved by the sparse-autoencoder based S3D pre-training scheme with 1000 epochs of fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.909, + 0.895 + ], + "angle": 0, + "content": "Comparing the benefits for each architecture, it can be seen that the Transformer-based Primus-M architecture profits substantially more from pre-training than the ResEnc-L CNN counterpart, however, it shows lower from-scratch results to begin with. Comparing the MAE pre-trained Primus-M to its 1000 epochs from scratch counterpart, it is able to improve about \\(3.5\\%\\) DSC points and almost exceeds the ResEnc-L 1000 epochs from scratch baseline in 150 epochs. However, on singular datasets the MAE" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23844" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.09, + 0.46, + 0.104 + ], + "angle": 0, + "content": "Table 4. Effects of different CNN fine-tuning schedules." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.115, + 0.47, + 0.326 + ], + "angle": 0, + "content": "
DatasetPT Method FT ScheduleVoCoSwinUNETRSimCLRVFMGMAES3DAverage
SBMDefault73.8669.8672.7972.5773.3771.8475.4772.82
Frozen52.3438.5960.8328.5451.3559.4556.5549.67
Warm-Up70.8769.1673.9473.1174.6774.5173.5472.83
Valley70.2670.2972.8275.8174.6873.7172.4072.85
Sawtooth71.5271.0573.7273.7873.8074.7973.9773.23
ATLDefault62.7860.8256.8063.8362.6262.7862.1661.68
Frozen58.2754.2757.9543.3357.2754.4957.1554.68
Warm-Up59.3260.2362.3663.2063.4663.6162.3462.07
Valley59.4360.3461.2164.0061.2362.9863.3861.80
Sawtooth60.4261.1761.3465.0761.5463.4763.7662.40
AMODefault85.9385.4287.1286.0586.9386.7886.7486.42
Frozen53.1861.2662.9230.9757.8950.8352.4852.79
Warm-Up86.9886.6187.1086.5887.3487.6287.5487.11
Valley85.9185.4686.7286.5587.0987.3786.9286.57
Sawtooth86.7885.7686.7586.4687.1587.4886.4586.69
KITDefault85.8083.8381.2084.3483.6984.5682.2183.66
Frozen69.3369.5965.2236.0671.3963.9771.4163.85
Warm-Up84.7181.7781.8083.5183.9985.2482.2083.32
Valley84.0882.4181.5585.0083.4883.4883.7883.40
Sawtooth85.1283.2683.1085.4184.1584.2883.3184.09
AverageDefault77.0974.9874.4876.7076.6576.4976.6576.15
Frozen58.2855.9361.7334.7359.4857.1859.4055.25
Warm-Up75.4774.4476.3076.6077.3677.7576.4076.33
Valley74.9274.6275.5777.8476.6276.8876.6276.15
Sawtooth75.9675.3176.2377.6876.6677.5076.8776.60
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.461 + ], + "angle": 0, + "content": "pre-trained Primus-M transformer already achieves the highest performance overall (ATL, COS and ACD), including the best pre-trained ResEnc-L. Given the substantial performance discrepancy observed in Wald et al. [35] between the two architectures, this provides the first sign of vision transformers reaching state-of-the-art performance in 3D medical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.702 + ], + "angle": 0, + "content": "However, not all pre-training methods are able to learn useful representations for segmentation. VoCo, SwinUNETR pre-training, and SimCLR (for transformers) do not, or just barely exceed, a from-scratch baseline trained for 150 epochs and can even lead to substantial performance reductions for the transformer architecture likely due to their global pre-training task not learning the local information needed for segmentation. Moreover, individual datasets can be negatively impacted by any pre-training, as for the ISL and HAN dataset when using a ResEnc-L architecture. This is particularly surprising as Primus-M shows strong improvements with its MAE pre-training on both of these datasets. When filtering low quality images this seems to be partially alleviated (see Table 6 in the appendix), indicating that CNNs may learn different patterns in the presence of noise, which a transformer may be more robust to." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Lastly, comparing the longer 1000 epoch fine-tuning schedule ( see Table 5 in the appendix) and the shorter 150 epoch fine-tuning schedule of the ResEnc-L, it can be observed that the longer fine-tuning schedule can lead to performance degradation on datasets where pre-training showed strong improvements (SBM, HNT). Simultaneously, it reduces the performance gap to the from-scratch 1000 epoch baseline on cases where pre-training was not as helpful (ISL, HAN, AMO), likely by allowing to forget learned representations and adapt more strongly to the downstream task." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Classification performance We evaluate the fine-tuned performance of our pre-trained methods for classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.362 + ], + "angle": 0, + "content": "- a more global downstream task - in Tab. 3. Contrastive pre-training methods like VoCo, SwinUNETR, and SimCLR, which were ineffective for segmentation, achieve the highest classification performance. In contrast, MAE pretraining, which was the strongest for segmentation, performs the worst across both architectures and nearly all datasets. Furthermore, segmentation focused methods like S3D and MG do not transfer well to classification, though they do not fall to the lowest ranks. These findings reinforce the distinction between global and local feature learning: contrastive methods specialize in global similarity learning, making them well-suited for classification, whereas reconstruction-based objectives prioritize pixel-wise representations, benefiting segmentation. Interestingly, SwinUNETR pre-training, which incorporated both global and local objectives, still converged toward representations favoring classification over segmentation, suggesting a bias in learned feature hierarchies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.514 + ], + "angle": 0, + "content": "Additionally, it is to note that from-scratch classification performance is subpar, with ABI barely exceeding randomness (balanced accuracy of \\(50\\%\\)). This may be due to the classification pipeline lacking the refinements of the widely adopted nnU-Net framework, due to the lower prevalence of global reasoning tasks in 3D medical imaging. Thus, while our classification results provide valuable insights, they should be interpreted with caution, particularly compared to our segmentation findings which are more reliable due to the robust benchmarking framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.744, + 0.547 + ], + "angle": 0, + "content": "4.2. Fine-tuning methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.905, + 0.779 + ], + "angle": 0, + "content": "The ablation results of different CNN fine-tuning strategies in Tab. 4 indicate that the Sawtooth fine-tuning strategy consistently achieves the best overall performance, followed closely by the Warm-Up scheme. This shows that adapting the encoder, after tuning the newly initialized decoder is crucial to maximize performance. The staged learning rate increases in these schedules likely help to prevent instability and allow for a smoother transition from pre-training to task-specific adaptation. For the out-of-distribution CT datasets, AMO and KIT, the Default fine-tuning schedule performs very competitively. This indicates that models pretrained on MRI data require greater modifications when adapted to CT images. The high initial lr allows the models to shift their learned features more effectively to the new domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Keeping the encoder frozen during fine-tuning significantly reduces performance across all datasets and underscores the importance of encoder adaptation. However, SimCLR is less affected by this constraint. Notably, both VoCo and SimCLR perform almost comparably to other fine-tuning schedules when the encoder is frozen on ATL, w.r.t. the drop on other datasets. This dataset is closest to the pretraining data in the development datasets, as it consists of a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23845" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "single MRI sequence, whereas the other datasets either contain CT scans or multiple MRI sequences as input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.482, + 0.332 + ], + "angle": 0, + "content": "For the Transformer, fine-tuning results in Table 8 in the appendix reveal that the Warm-Up scheme achieves the best performance across all methods, emphasizing the importance of architecture-specific adaptation. The Frozen schedule leads to an even greater performance drop than for CNNs, as the ResEnc-L decoder, despite being lightweight, has significantly more parameters than the \"decoder\" of Primus-M. In this setting, fine-tuning can be interpreted almost as linear probing, with only back-to-back transposed conv layers (including normalization and activations) being trained, severely limiting decoder capacity. With a frozen encoder, it can be observed that adapting to the ATL dataset reaches the closest performance to other fine-tuning schedules." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.342, + 0.279, + 0.358 + ], + "angle": 0, + "content": "4.3. Data-centric results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.62 + ], + "angle": 0, + "content": "We compare the effect of data filtering of the pre-training dataset for CNN MAE pre-training and highlight the measured downstream performance in Appendix Table 6. It can be observed that filtering the lowest quality images through the Image Quality Score (IQS) shows minor positive effects on overall downstream segmentation performance, despite removing substantial amounts of pre-training data. Only when removing about \\(66\\%\\), keeping only the highest quality images, a degradation in performance is measured. Moreover, it can be observed that reducing data diversity by removing all modalities but T1w, T2w and FLAIR images reduces model performance, despite having more pre-training images than the intermediate IQS filtering method which slightly increased performance. Overall, this indicates that the OpenMind dataset does contain redundant data which may be leveraged by more complex data-centric methods to increase performance further." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Effect of anonymization on reconstruction Many images of the OpenMind dataset were defaced or brain-extracted to guarantee patient privacy. To allow taking these artifacts into account when training with a reconstruction objective, we created masks that allow ignoring such regions. To measure the effect of such artifacts, we compare the performance of reconstruction methods that were or were not penalized for the reconstruction in such regions. (See Appendix Table 7). It can be observed that ignoring the reconstruction loss of anonymized regions improves the segmentation performance of the MAE and S3D, showing that taking anonymization into account improves representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.445, + 0.832 + ], + "angle": 0, + "content": "5. Discussion, Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our study presents the OpenMind dataset, the largest publicly available 3D medical imaging dataset, alongside a standardized self-supervised learning (SSL) benchmark for 3D medical image analysis. Through extensive experi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.393 + ], + "angle": 0, + "content": "mentation across diverse datasets and across many pretraining paradigms, we demonstrate the superiority of self-supervised pre-training over from-scratch training for 3D medical image analysis, when pre-training on a large-scale dataset. We show that contrastive methods excel in global tasks such as classification, and reconstruction-based approaches perform well in dense tasks such as segmentation. However, we find that no pre-training method of the 3D vision domain generalizes and performs well in both tasks, emphasizing the need for further SSL method development. Moreover, we compared state-of-the-art CNNs and Transformers in our benchmark and show for the very first time that a pre-trained Primus-M Transformer can exceed the strong ResEnc-L CNN from-scratch baseline on many datasets and can exceed the strongest pre-trained ResEnc-L CNN on individual datasets. This marks an important milestone for the domain, as transformers from-scratch so far never exceeded a ResEnc-L model [34, 35, 51]. On average, though, we show that a pre-trained ResEnc-L remains the strongest architecture for now." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.77 + ], + "angle": 0, + "content": "Moreover, our ablation results suggest that fine-tuning strategies are a key determinant of the success of pre-trained methods, with bad strategies losing significant performance. This emphasizes the importance of careful adaptation beyond pre-training. Experiments on fine-tuning for longer showed partial degradation of overall performance, indicating the problem of overfitting to downstream tasks. Together, these insights open up research opportunities for developing and adapting Parameter-Efficient Fine-Tuning (PEFT) methods to 3D medical image analysis to address these issues. Moreover, our fine-tuning experiments showed that keeping the vision encoders frozen shows sub-par performance for all pre-training methods. This emphasizes the lack of generality of currently learned representations and – again – highlights the need for better pre-training methods. While our study provides a lot it also has various limitations: i) Our classification results are not as reliable as the segmentation results, due to the lower interest in classification for 3D medical imaging, requiring further refinements. ii) Current evidence of the utility of the OpenMind dataset for data-centric research is rather weak, due to the minimal improvements observed requiring further investigation. iii) We evaluate pre-training for 1000 epochs due to resource limitations. Pre-training for longer may reveal different trends than shown in this study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.905, + 0.876 + ], + "angle": 0, + "content": "In summary, by open-sourcing the OpenMind dataset and establishing the OpenMind benchmark, we provide a clear assessment of state-of-the-art SSL in 3D medical imaging and offer a reproducible framework for method development. Through them, we aim to accelerate progress in self-supervised learning for 3D medical images and pave the way for new research directions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23846" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.221 + ], + "angle": 0, + "content": "This work was partly funded by the Helmholtz Foundation Model Initiative (HFMI) under the subproject 'The Human Radiome Project' (THRP). Part of this work was funded by Helmholtz Imaging (HI), a platform of the Helmholtz Incubator on Information and Data Science. Additionally, a part of the computational resources were provided by the German AI Service Center WestAI." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.238, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.262, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.347, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[2] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.418, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.49, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[4] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.547, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[5] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.632, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[6] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.703, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[7] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.788, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[8] Constantin Ulrich, Fabian Isensee, Tassilo Wald, Maximilian Zenk, Michael Baumgartner, and Klaus H Maier-Hein. Multilant: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 648-658. Springer, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[9] Wenxuan Li, Alan Yuille, and Zongwei Zhou. How well do supervised 3d models transfer to medical imaging tasks?" + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.262, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "In The Twelfth International Conference on Learning Representations, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[10] OASIS. Open access series of imaging studies data use agreement, 2024. URL https://sites.wustl.edu/oasisbrains/." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[11] PPMI. Parkinson's progression markers initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=PPMI." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[12] ADNI. Alzheimer's disease neuroimaging initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=ADNI." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.274, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[13] MCSA. Mayo clinic study of aging data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=MCSA." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.33, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[14] HABS-HD. Health and aging brain study: Health disparities data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project \\(\\equiv\\) HABS_HD." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[15] Tassilo Wald, Constantin Ulrich, Stanislav Lukyanenko, Andrei Goncharov, Alberto Paderno, Leander Maerkisch, Paul F Jaeger, and Klaus Maier-Hein. Revisiting mae pretraining for 3d medical image segmentation. arXiv preprint arXiv:2410.23132, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[16] Guotai Wang, Jianghao Wu, Xiangde Luo, Xinglong Liu, Kang Li, and Shaoting Zhang. Mis-fm: 3d medical image segmentation using foundation models pretrained on a large-scale unannotated dataset. arXiv preprint arXiv:2306.16925, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Asbjørn Munk, Jakob Ambsdorf, Sebastian Llambias, and Mads Nielsen. Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation. arXiv preprint arXiv:2408.00640, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[18] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[19] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.706, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[20] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[21] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[22] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024." + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23847" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[23] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[24] Fenghe Tang, Ronghao Xu, Qingsong Yao, Xueming Fu, Quan Quan, Heqin Zhu, Zaiyi Liu, and S Kevin Zhou. Hyspark: Hybrid sparse masking for large scale medical image pre-training. arXiv preprint arXiv:2408.05815, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[25] National Institutes of Health et al. Adolescent brain cognitive development study, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[26] Ibrahim Ethem Hamamci, Sezgin Er, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Muhammed Furkan Dasdelen, Omer Faruk Durugol, Bastian Wittmann, Tamaz Amiranashvili, et al. Developing generalist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[27] Kenneth Marek, Sohini Chowdhury, Andrew Siderowf, Shirley Lasch, Christopher S Coffey, Chelsea Caspell-Garcia, Tanya Simuni, Danna Jennings, Caroline M Tanner, John Q Trojanowski, et al. The parkinson's progression markers initiative (ppmi)-establishing a pd biomarker cohort. Annals of clinical and translational neurology, 5 (12):1460-1477, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[28] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics, 15(4):869-877, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[29] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[30] Christopher J Markiewicz, Krzysztof J Gorgolewski, Franklin Feingold, Ross Blair, Yaroslav O Halchenko, Eric Miller, Nell Hardcastle, Joe Wexler, Oscar Esteban, Mathias Goncavles, et al. The openneuro resource for sharing of neuroscience data. *Elife*, 10:e71774, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[31] Michal Nohel, Constantin Ulrich, Jonathan Suprijadi, Tassilo Wald, and Klaus Maier-Hein. Unified framework for foreground and anonymization area segmentation in ct and mri data. In Bildverarbeitung für die Medizin 2025, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[32] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[33] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.149 + ], + "angle": 0, + "content": "Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 488-498. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.149, + 0.906, + 0.231 + ], + "angle": 0, + "content": "[35] Tassilo Wald, Saikat Roy, Fabian Isensee, Constantin Ulrich, Sebastian Ziegler, Dasha Trofimova, Raphael Stock, Michael Baumgartner, Gregor Kohler, and Klaus Maier-Hein. Primus: Enforcing attention usage for 3d medical image segmentation, 2025. URL https://arxiv.org/abs/2503.01835." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.232, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[36] Sook-Lei Liew, Bethany P Lo, Miranda R Donnelly, Artemis Zavaliangos-Petropulu, Jessica N Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P Simon, Julia M Juliano, Anisha Suri, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific data, 9(1):320, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.316, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[37] Endre Grøvik, Darwin Yi, Michael Iv, Elizabeth Tong, Daniel Rubin, and Greg Zaharchuk. Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri. Journal of Magnetic Resonance Imaging, 51(1):175-182, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.386, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[38] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems, 35: 36722–36732, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.469, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[39] Nicholas Heller, Fabian Isensee, Klaus H Maier-Hein, Xiaoshuai Hou, Chunmei Xie, Fengyi Li, Yang Nan, Guangrui Mu, Zhiyong Lin, Miofei Han, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis, 67:101821, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.553, + 0.906, + 0.635 + ], + "angle": 0, + "content": "[40] Moritz R Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, et al. Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data, 9(1):762, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.636, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[41] Kareem Wahid, Cem Dede, Mohamed Naser, and Clifton Fuller. Training dataset for hntsmrg 2024 challenge, 2024. URL https://zenodo.org/doi/10.5281/ zenodo.11199559." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.692, + 0.906, + 0.748 + ], + "angle": 0, + "content": "[42] Gasper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.749, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[43] Ali M. Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita binti Hanafi, Osama Alnuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi. Brain mri dataset of multiple sclerosis with consensus manual lesion segmentation and patient meta information. Data in Brief, 42:108139, 2022. ISSN 2352-3409. doi: https://doi.org/10.1016/j.dib.2022.108139. URL https://www.sciencedirect.com/science/article/pii/S235234092200347X." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[44] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23848" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.465 + ], + "angle": 0, + "content": "Hoher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, Suprosanna Shit, Houjing Huang, Chinmay Prabhakar, Ezequiel de la Rosa, Diana Waldmannstetter, Florian Kofler, Fernando Navarro, Martin Menten, Ivan Ezhov, Daniel Rueckert, Iris Vos, Ynte Ruigrok, Birgitta Velthuis, Hugo Kuijf, Julien Hammerli, Catherine Wurster, Philippe Bijlenga, Laura Westphal, Jeroen Bisschop, Elisa Colombo, Hakim Baazaoui, Andrew Makmur, James Hallinan, Bene Wiestler, Jan S. Kirschke, Roland Wiest, Emmanuel Montagnon, Laurent Letourneau-Guillon, Adrian Galdran, Francesco Galati, Daniele Falcetta, Maria A. Zuluaga, Chaolong Lin, Haoran Zhao, Zehan Zhang, Sinyoung Ra, Jongyun Hwang, Hyunjin Park, Junqiang Chen, Marek Wodzinski, Henning Muller, Pengcheng Shi, Wei Liu Ting Ma, Cansu Yalçin, Rachika E. Hamadache, Joaquim Salvi, Xavier Llado, Uma Maria Lal-Trehan Estrada, Valeria Abramova, Luca Giancardo, Arnau Oliver, Jialu Liu, Haibin Huang, Yue Cui, Zehang Lin, Yusheng Liu, Shunzhi Zhu, Tatsat R. Patel, Vincent M. Tutino, Maysam Orouskhani, Huayu Wang, Mahmud Mossa-Basha, Chengcheng Zhu, Maximilian R. Rokuss, Yannick Kirchhoff, Nico Disch, Julius Holzschuh, Fabian Isensee, Klaus Maier-Hein, Yuki Sato, Sven Hirsch, Susanne Wegener, and Bjoern Menze. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra, 2024. URL https://arxiv.org/abs/2312.17670." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.468, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[45] Divya Ramakrishnan, Leon Jekel, Saahil Chadha, Anastasia Janas, Harrison Moy, Nazanin Maleki, Matthew Sala, Manpreet Kaur, Gabriel Cassinelli Petersen, Sara Merkaj, et al. A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information. Scientific Data, 11(1):254, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[46] H. Chen, X. Zhao, J. Dou, C. Du, R. Yang, H. Sun, S. Yu, H. Zhao, C. Yuan, and N. Balu. Carotid vessel wall segmentation and atherosclerosis diagnosis challenge, 2022. URL https://vessel-wall-segmentation-2022.grand-challenge.org." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[47] Olivier Bernard, Alain Lalande, Clement Zotti, Cervenansky, and et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE TMI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.678, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[48] Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet. PLoS medicine, 15(11):e1002699, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[49] Hui Ming Lin, Errol Colak, Tyler Richards, Felipe C Kitamura, Luciano M Prevedello, Jason Talbott, Robyn L Ball, Ekim Gumeler, Kristen W Yeom, Mohammad Hamghalam, et al. The rsna cervical spine fracture ct dataset. Radiology: Artificial Intelligence, 5(5):e230034, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[50] Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain ar" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "chitecture in autism. Molecular psychiatry, 19(6):659-667, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[51] Pedro RAS Bassi, Wenxuan Li, Yucheng Tang, Fabian Isensee, Zifu Wang, Jieneng Chen, Yu-Cheng Chou, Yannick Kirchhoff, Maximilian R Rokuss, Ziyan Huang, et al. Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation? Advances in Neural Information Processing Systems, 37:15184-15201, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.221, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[52] Lachlan T. Strike, Narelle K. Hansell, Jessica L. Miller, Kai-Hsiang Chuang, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin adolescent brain (qtab)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004146.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[53] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-id1000\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso003097.v1.2.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.35, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[54] Makayla Gibson, Roger Newman-Norlund, Leonardo Bonilha, Julius Fridriksson, Gregory Hickok, Argye E. Hillis, Dirk-Bart den Ouden, and Chris Rorden. \"aphasia recovery cohort (arc) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004884.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[55] Chris Rorden, John Absher, and Roger Newman-Norlund. \"stroke outcome optimization project (soop)\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004889.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.465, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[56] Allison C. Nugent, Adam G Thomas, Margaret Mahoney, Alison Gibbons, Jarrod Smith, Antoinette Charles, Jacob S Shaw, Jeffrey D Stout, Anna M Namyst, Arshitha Basavaraj, Eric Earl, Dustin Moraczewski, Emily Guinee, Michael Liu, Travis Riddle, Joseph Snow, Shruti Japee, Morgan Andrews, Adriana Pavletic, Stephen Sinclair, Vinai Roopchansingh, Peter A Bandettini, and Joyce Chung. \"the nimh healthy research volunteer dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005752.v2.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[57] Rani Gera, Maya Bar Or, Ido Tavor, Dana Roll, Jeffrey Cockburn, Segev Barak, Elizabeth Tricomi, John P. O'Doherty, and Tom Schonberg. \"characterizing habit learning in the human brain at the individual and group levels: a multi-modal mri study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004299.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.676, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[58] Denise Park, Joseph Hennessee, Evan T. Smith, Micaela Chan, Carson Katen, Julia Bacci, Sarah Frank, Sarah Monier, Alexandra Collyer, Carol Tamminga, William Moore, Neil Rofsky, Karen Rodrigue, Kristen Kennedy, and Gagan Wig. \"the dallas lifespan brain study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004856.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[59] Jin Wang, Marisa N. Lytle, Yael Weiss, Brianna L. Yasasaki, and James R. Booth. \"a longitudinal neuroimaging dataset on language processing in children ages 5, 7, and 9 years old\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003604.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Leon Y. Cai, Qi Yang, Praitayini Kanakaraj, Vishwesh Nath, Allen T. Newton, Heidi A. Edmonson, Jeffrey Luci, Benjamin N. Conrad, Gavin R. Price, Colin B. Hansen, Cailey I. Kerley, Karthik Ramadass, Fang-Cheng Yeh, Hakmook Kang, Eleftherios Garyfallidis, Maxime Descoteaux," + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23849" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "Francois Rheault, Kurt G. Schilling, and Bennett A. Landman. \"masivar: Multisite, multiscanner, and multisubject acquisitions for studying variability in diffusion weighted magnetic resonance imaging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003416.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[61] Anahit Babayan, Blazeij Baczkowski, Roberto Cozatl, Maria Dreyer, Haakon Engen, Miray Erbey, Marcel Falkiewicz, Nicolas Farrugia, Michael Gaebler, Johannes Golchert, Laura Golz, Krzysztof Gorgolewski, Philipp Haueis, Julia Huntenburg, Rebecca Jost, Yelyzaveta Kramarenko, Sarah Krause, Deniz Kumral, Mark Lauckner, Daniel S. Margulies, Natacha Mendes, Katharina Ohrnberger, Sabine Oligschläger, Anastasia Osoianu, Jared Pool, Janis Reichelt, Andrea Reiter, Josefin Röbbig, Lina Schaare, Jonathan Smallwood, and Arno Villringer. \"mpileipzig_mind-brain-body\". OpenNeuro, 2020. doi: 10. 18112/openneuro.ds000221.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.332, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[62] Karl Koschutnig, Bernhard Weber, and Andreas Fink. \"tidying up white matter: Neuroplastic transformations in sensorimotor tracts following slackline skill acquisition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003138.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[63] Sangil Lee and Joseph Kable. \"cognitive training\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds002843.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.501 + ], + "angle": 0, + "content": "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[65] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[66] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[67] Tom Schonberg, Christopher Trepel, Craig Fox, and Russell A. Poldrack. \"balloon analog risk-taking task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000001.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[68] J.M. Moran, E. Jolly, and J.P. Mitchell. \"social cognitive tasks (animate movement task, moral judgment task, false belief task)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000109.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[69] Ella Gabitov, David Manor, and Avi Karni. \"learning and memory: motor skill consolidation and intermanual transfer\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[70] A. Roy, R. A. Bernier, J. Wang, M. Benson, J. French, D. Good, and F.G. Hillary. \"cost analysis tbi\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms000220.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.805, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[71] Noritaka Yoneyama, Hirohisa Watanabe, Kazuya Kawabata, Epifanio Bagarinao, Kazuhiro Hara, Takashi Tsuboi, Yasuhiro Tanaka, Reiko Ohdake, Kazunori Imai, Rei Ogura, Yasutaka Kato, Michihito Masuda, Tatsuya Hattori, Mizuki Ito, Satoshi Maesawa, Daisuke Mori, Masahisa Katsuno, and Gen Sobue. \"ds000245_r1.0.0\". OpenNeuro, 2018." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.904, + 0.12 + ], + "angle": 0, + "content": "[72] H Aerts and D Marinazzo. \"btc_preop\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s001226.v5.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.123, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[73] Johns CL, Jahn AA, Jones HR, Kush DK, Molfese PJ, Van Dyke JA, Magnuson JS, Tabor W, Mencl WE, Shankweiler DP, and Braze D. \"indiv diffs readingskill\". OpenNeuro, 2018. doi: 10.18112/openneuro.d001365.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[74] C. Baldassano, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (scrambled)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001511.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.236, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[75] A.M. Bornstein, M. Aly, S. Feng, N.B. Turk-Browne, K.A. Norman, and J.D. Cohen. \"perceptual decisions result from the continuous accumulation of memory and sensory evidence\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001614.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.307, + 0.882, + 0.319 + ], + "angle": 0, + "content": "[76] Unknown authors. \"sc mri dataset\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.322, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[77] Eduardo A. Garza-Villarreal, Victor Pando-Naude, Fernando A. Barrios, Peter Vuust, Lene Vase, Elvira Brattico, Erick H. Pasaye, and Sarael Alcauter. \"functional connectivity of music-induced analgesia in fibromyalgia\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001928.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.392, + 0.904, + 0.419 + ], + "angle": 0, + "content": "[78] B. Sherman and N.B. Turk-Browne. \"stat episodic\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds002144.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.421, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[79] Gilles de Hollander and Tomas H. Knapen. \"odc mapper 7t surface coil\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002295.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.464, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[80] Tomoyasu Horikawa, Alan S. Cowen, Dacher Keltner, and Yukiyasu Kamitani. \"the neural representation of visually evoked emotion is high-dimensional, categorical, and distributed across transmodal brain regions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002425.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.535, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[81] Yazhu Qian, Jinyou Zou, Zihao Zhang, Zhentao Zuo, Yan Zhuo, Danny JJ, Wang, and Peng Zhang. \"robust functional mapping of layer-selective responses in human lateral geniculate nucleus with high-resolution 7t fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002672.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.606, + 0.904, + 0.646 + ], + "angle": 0, + "content": "[82] G. Elliott Wimmer, Jamie Li, Krzysztof Jacek Gorgolewski, and Russell A. Poldrack. \"rewardbeast\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s002738.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.649, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[83] Francesca M M Citron, Nora Michaelis, and Adele E Goldberg. \"l2_sentences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002842.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.691, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[84] Matteo Visconti di Oleggio Castello, Vassiki Chauhan, Guo Jiahui, and M. Ida Gobbini. \"an fmri dataset in response to \"the grand budapest hotel\", a socially-rich, naturalistic movie\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds003017.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[85] Chiara Banfi, Karl Koschutnig, Kristina Moll, Gerd Schulte-Koerne, Andreas Fink, and Karin Landerl. \"reading-related functional activity in children with isolated spelling deficits and dyslexia\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003126.v1.3.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[86] Andy Schumann, Feliberto de la Cruz, Stefanie K\"ohler, Lisa Brotte, and Karl-J\"urgen B\"ar. \"the influence of heart rate variability biofeedback on cardiac regulation and functional brain connectivity\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003357.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23850" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[87] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant4 motion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003441.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[88] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., Todd R. Parrish, and Barbara Trommer. \"response inhibition and selective attention in adults and children with and without ADHD\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003500.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[89] Nathan M. Muncy, C. Brock Kirwan, and Jonathan J. Wisco. \"differences in chemo-signaling compound-evoked brain activity in male and female young adults: A pilot study in the role of sexual dimorphism in olfactory chemosignaling\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003606.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[90] Beau Sievers, Carolyn Parkinson, Peter J. Kohler, James Hughes, Sergey V. Fogelson, and Thalia Wheatley. \"visual and auditory brain areas share a representational structure that supports emotion perception: fmri data\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003715.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[91] Andreas Fink, Karl Koschutnig, Thomas Zussner, Corinna M. Perchtold-Stefan, Christian Rominger, Mathias Benedek, and Ilona Papousek. \"a two-week running intervention reduces symptoms related to depression and increases hippocampal volume in young adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003799.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[92] John Wilder, Morteza Rezanejad, Sven Dickinson, Kaleem Siddiqi, Allan Jepson, and Dirk B. Walther. \"sceneperceptionandparallelism\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03950.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[93] Zoe Woodhead, Maria Sablik, Kate Watkins, David Carey, Dorothy Bishop, and Emma Karlsson. \"comparing language lateralisation using fmri and ftcd\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004073.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[94] Martin Szinte, Tomas Knapen, and Matthias Nau. \"gaze_exp\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004271.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[95] Maria Czarnecka, Andre Knops, and Marcin Szwed. \"number comparison task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004400.v1.0.6." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[96] Alex L. White, Kendrick Kay, Kenny A. Tang, and Jason D. Yeatman. \"raw data for engaging in word recognition elicits highly specific modulations in visual cortex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004489.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[97] Loriana Goulding, Austin Schmidt, Isaac Hamm, and C. Brock Kirwan. \"fmri investigations of individual differences on memory activation for faces and words: The effects of handedness and phenomenal experience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004589.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[98] Halle R. Dimsdale-Zucker and Christopher Baldassano. \"common\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004663.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[99] Rotem Botvinik-Nezer, Bogdan Petre, Marta Ceko, Naomi Friedman, and Tor Wager. \"paingen_placebo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004746.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[100] Xuefeng Yan, Martin Noergaard, Cheryl L. Morse, JeihSan Liow, Jinsoo Hong, Sanjay Telu, Min-Jeong Kim, Jose" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.217 + ], + "angle": 0, + "content": "A Montero Santamaria, Anthony Galassi, Ningping Feng, Sarah K. Williams Avram, Ted B. Usdin, Shawn Wu, Andrea Zhang, Lester Manly, Madeline Jenkins, Maia Van Buskirk, Adrian Lee, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. \"positron emission tomography (pet) quantification in healthy humans of cyclooxygenase-2 (cox-2), a potential biomarker of neuroinflammation\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004869.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[101] Liberty S. Hamilton, Maansi Desai, and Alyssa Field. \"wired icm sample dataset - workshop on intracranial recordings in humans, epilepsy, dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004993.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[102] Keith M. Vogt, Alex C. Burlew, Sujatha N. Reddy, Marcus Simmons, Courtney N. Kozdron, and James W. Ibinson. \"neural correlates of lidocaine analgesic (nla) study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005088.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.346, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[103] Sidhant Chopra, Carrisa V. Cocuzza, Connor Lawhead, Jocelyn A. Ricard, Loic Labache, Lauren Patrick, Poornima Kumar, Arielle Rubenstein, Julia Moses, Lia Chen, Crystal Blankenbaker, Bryce Gillis, Laura T. Germine, Ilan Harpaz-Rote, BT Thomas Yeo, Justin T. Baker, and Avram J. Holmes. \"transdiagnostic connectome project\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005237.v1.1.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.445, + 0.905, + 0.499 + ], + "angle": 0, + "content": "[104] Marleen Haupt, Douglas D. Garrett, and Radoslaw M. Cichy. \"object recognition in healthy aging (orha) - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005374.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.502, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[105] Luis Martinez Agulleiro, Bowen Deng, Yuliya Yoncheva, Chao-Gan Yan, F Xavier Castellanos, and Marie-Pierre St-Onge. \"circadian misalignment and energy balance\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005525.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.56, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[106] A.R. Aron, M.A. Gluck, and R.A. Poldrack. \"classification learning\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[107] Rebecca J. Lepping, Ruth Ann Atchley, Evangelia Chrysikou, Laura E. Martin, Alicia A. Clair, Rick E. Ingram, W. Kyle Simmons, and Cary R. Savage. \"neural processing of emotional musical and nonmusical stimuli in depression\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[108] Elizabeth Bock, Peter Donhauser, Francois Tadel, Guiomar Niso, and Sylvain Baillet. \"meg-bids brainstorm data sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000246.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[109] Joseph M. Orr and Vijay A. Mittal. \"rule learning in adolescents at clinical high risk for psychosis shows altered patterns of activation\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps01371.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[110] A.C. Schapiro, T.T. Rogers, N.I. Cordova, N.B. Turk-Browne, and M.M. Botvinick. \"neural representations of events arise from temporal community structure\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001621.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.831, + 0.905, + 0.885 + ], + "angle": 0, + "content": "[111] Omer Faruk Gulban, Kevin R. Sitek, Satrajit S. Ghosh, Michelle Moerel, and Federico De Martino. \"auditory localization with 7t fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001942.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[112] Tomoya Nakai and Shinji Nishimoto. \"over 100 task fmri" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23851" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002306.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[113] Laura Pritschet, Tyler Santander, Caitlin M. Taylor, Evan Layher, Shuying Yu, Michael B. Miller, Scott T. Grafton, Emily G. Jacobs, Morgan Fitzgerald, and Joshua M. Mueller. \"28andme\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002674.v1.0.6." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[114] PhD Heidemarie Laurent, Megan K. Finnegan, and Katherine Haigler. \"postnatal affective mri dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003136.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[115] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant5 suboptimal event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003442.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[116] Monica Crotti, Karl Koschutnig, and Selina Wriessnegger. \"the impact of handedness on the neural correlates during kinesthetic motor imagery: a fmri study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003612.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[117] Ben Deen and Winrich Freiwald. \"parallel systems for social and spatial reasoning\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms003814.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[118] Shaonan Wang, Xiaohan Zhang, Jiajun Zhang, and Chengqing Zong. \"a synchronized multimodal neuroimaging dataset to study brain language processing\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004078.v1.2.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[119] Shani Agron, Reut Weissgross, Eva Mishor, Lior Gorodisky, Aharon Ravia, Edna Furman-Haran, Tali Weiss, and Noam Sobel. \"a chemical signal in human emotional tears blocks aggression in men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004274.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[120] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004493.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[121] Jose Pedro Manzano Patron, Steen Moeller, Jesper L.R. Andersson, Kamil Ugurbil, Essa Yacoub, and Stamatios N. Sotiropoulos. \"edden: Evaluation of diffusion mri denoising\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004666.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[122] Samira Epp, Gabriel Castrillon, Beija Yuan, Jessica Andrews-Hanna, Christine Preibisch, and Valentin Riedl. \"two distinct modes of hemodynamic responses in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004873.v2.0.6." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[123] Chloe M. de Nys, Ee Shern Liang, Marita Prior, Maria A. Woodruff, James I. Novak, Ashley R. Murphy, Zhiyong Li, Craig D. Winter, and Mark C. Allenby. \"royalbrisbane_tofmra_intracranial aneurysm_database\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005096.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[124] Luis Marcos-Vidal, Helena Gil-Buitrago, Irene Cisma, Rosamunde C. Hendricks, Scott Atran, and Clara Pretus. \"polex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005375.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[125] G. Xue and Russell A. Poldrack. \"rhyme judgment\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms000003.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[126] Tom S.M., Fox C.R., Trepel C., and Poldrack R.A. \"mixed-gambles task\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.904, + 0.149 + ], + "angle": 0, + "content": "[127] K Jimura, E Stover, F Casalis, and RA Poldrack. \"ds000006\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.152, + 0.905, + 0.18 + ], + "angle": 0, + "content": "[128] Xue G, Aron AR, and Poldrack RA. \"stop-signal task with spoken & manual responses\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.182, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[129] A.R. Aron, T.E. Behrens, M. Frank, S. Smith, and R.A. Poldrack. \"stop-signal task with unconditional and conditional stopping\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.904, + 0.251 + ], + "angle": 0, + "content": "[130] Jessica Cohen and Russell Poldrack. \"the generality of self-control\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.904, + 0.295 + ], + "angle": 0, + "content": "[131] K. Foerde, B.J. Knowlton, and Russell A. Poldrack. \"classification learning and tone-counting\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso000011.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.297, + 0.904, + 0.324 + ], + "angle": 0, + "content": "[132] Rizk-Jackson, Aron, and Russell A. Poldrack. \"ds000017\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.327, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[133] R Bilder, R Poldrack, T Cannon, E London, N Freimer, E Congdon, K Karlgodt, and F Sabb. \"ucla consortium for neuropsychiatric phenomics la5c study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000030.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.384, + 0.904, + 0.411 + ], + "angle": 0, + "content": "[134] Russell Poldrack. \"myconnectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs000031.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.904, + 0.44 + ], + "angle": 0, + "content": "[135] R. Alvarez and R.A. Poldrack. \"cross-language repetition priming\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.904, + 0.484 + ], + "angle": 0, + "content": "[136] R.A. Poldrack, J. Clark, E. J. Pare-Blagoev, D. Shohamy, J. Creso Moyano, C. Myers, and M Gluck. \"classification learning and reversal\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[137] Mei-Yen Chen, Corey N. White, Nathan Giles, Albert Elumn, Sagar Parikh, Ungi Kim, W. Todd Maddox, and Russell A. Poldrack. \"training of loss aversion modulates neural sensitivity toward potential gains\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.904, + 0.569 + ], + "angle": 0, + "content": "[138] Kelly AMC and Milham MP. \"simon task\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.611 + ], + "angle": 0, + "content": "[139] Kelly AMC, Uddin LQ, Biswal BB, Castellanos FX, and Milham MP. \"flanker task (event-related)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[140] J.V. Haxby, M.I. Gobbini, M.L. Furey, A. Ishai, J.L. Schouten, and P. Pietrini. \"visual object recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000105.v3.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.904, + 0.699 + ], + "angle": 0, + "content": "[141] K.J. Duncan, C. Pattamadilok, I. Knierim, and J.T. Devlin. \"word and object processing\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[142] Wager T.D., M.L. Davidson, B.L. Hughes, M.A. Lindquist, and K.N. Ochsner. \"prefrontal-subcortical pathways mediating successful emotion regulation\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[143] Melina R. Uncapher, J. Benjamin Hutchinson, and Anthony D. Wagner. \"incidental encoding task (posner cueing paradigm)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[144] Michael Hanke, Florian J. Baumgartner, Pierre Ibe, Falko R. Kaule, Stefan Pollmann, Oliver Speck, Wolf Zinke, and Jorg Stadler. \"forrest gump\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps00113.v1.3.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[145] Gorgolewski KJ, Storkey A, Bastin ME, Whittle IR, Wardlaw JM, and Pernet CR. \"a test-retest fmri dataset for motor, language and spatial attention functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000114.v1.0.2." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23852" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[146] Grega Repov\\(\\mathring{\\mathrm{A}}_{\\mathrm{i}}\\) and Deanna M. Barch. \"working memory in healthy and schizophrenic individuals\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.134, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[147] Jennifer M Walz, Robin I Goldman, Jordan Muraskin, Bryan Conroy, Truman R Brown, and Paul Sajda. \"auditory and visual oddball eeg-fmri\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.177, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[148] DG Wakeman and RN Henson. \"multisubject, multimodal face processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.rs000117.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[149] Katerina Velanova, Mark E. Wheeler, and Beatrix Luna. \"maturational changes in anterior cingulate and frontoparietal recruitment support the development of error processing and inhibitory control (antistate)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[150] Aarthi Padmanabhana, Charles F. Geier, Sarah J. Ordaz, Theresa Teslovich, and Beatrix Luna. \"developmental changes in brain function underlying the influence of reward processing on inhibitory control (slot reward)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.343, + 0.482, + 0.398 + ], + "angle": 0, + "content": "[151] C.F. Geier, R. Terwilliger, T. Teslovich, K. Velanova, and B. Luna. \"immaturities in reward processing and its influence on inhibitory control in adolescence (ring reward)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[152] J Xu, S Zhang, VD Calhoun, J Monterosso, and MN Potenza. \"task-related concurrent but opposite modulations of overlapping functional networks as revealed by spatial ica\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.456, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[153] Nicoletta Cera, Armando Tartaro, and Stefano L. Sensi. \"modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fmri study\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.511, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[154] Choong-Wan Woo, Mathieu Roy, Jason T. Buhle, and Tor D. Wager. \"distinct brain systems mediate the effects of nociceptive input and self-regulation on pain\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[155] Kimberly L. H. Carpenter, Adrian Angold, Nan-Kuei Chen, William E. Copeland, Pooja Gaur, Kevin Pelphrey, Allen W. Song, and Helen L. Egger. \"preschool anxiety disorders\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[156] E. Necka, M. Senderecka, B. Kucharzyk, and Falkiewicz M. \"who can afford self-control? the neural efficiency mechanism explains effective self-regulation of behavior\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[157] Elissa M. Aminoff and Michael J. Tarr. \"scene perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000149.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[158] Paul A. M. Smeets, Floor M. Kroese, Catherine Evers, and D. T. D. de Ridder. \"block design food and nonfood picture viewing task\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[159] CR Pernet, P Belin, P McAleer, KJ Gorgolewski, M Valdes-Sosa, M Latinus, I Charest, PEG Bestelmeyer, RH Watson, and D Fleming. \"the human voice areas: spatial organisation and inter-individual variability in temporal and extra-temporal cortices\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000158.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[160] Timothy D. Verstynen. \"stroop task\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[161] James K. Bursley, Adrian Nestor, Michael J. Tarr, and J. David Creswell. \"offline processing in associative learning\". OpenNeuro, 2019." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[162] Hsu AL, Chou KH, Chao YP, Fan HY, Wu CW, and Chen JH. \"physiological contribution in spontaneous oscillations: An approximate quality - assurance index for resting-state fmri signals\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[163] Laura Koenders, Janna Cousijn, Wilhelmina A.M. Vingerhoets, Wim van den Brink, Reinout W. Wiers, Carin J. Meijer, Marise W. J. Machielsen, Dick J. Veltman, Anneke E. Goudriaan, and Lieuwe de Haan. \"t1-weighted structural mri study of cannabis users at baseline and 3 years follow up\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.904, + 0.262 + ], + "angle": 0, + "content": "[164] Emilia Iannilli. \"pre-adolescents exposure to manganese\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[165] Gustav Nilsson, Sandra Tamm, Paolo d'Onofrio, Hanna Å Thuné, Johanna Schwarz, Catharina Lavebratt, Jia Jia Liu, Kristoffer NT Måsson, Tina Sundelin, John Axelsson, Peter Fransson, Göran Kecklund, Håkan Fischer, Mats Lekander, and Torbjörn Åkerstedt. \"the stockholm sleepy brain study: Effects of sleep deprivation on cognitive and emotional processing in young and old\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000201.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[166] Peter Van Schuerbeek, Chris Baeken, and Johan De Mey. \"the heterogeneity in retrieved relations between the personality trait 'harm avoidance' and gray matter volumes due to variations in the vbm and roi labeling processing settings\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[167] Christian Stephan-Otto, Sara Siddi, Carl Senior, Daniel Muñoz-Samons, Susana Ochoa, Ana María Sánchez Laforga, and Gildas Brébion. \"visual imagery and false memory for pictures\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[168] Dara Dickstein, Mariel Y. Pullman, Corey Fernandez, Jennifer A. Short, Lale Kostakoglu, Karin Knaesaurek, Laili Soleimani, Barry Jordan, Wayne Gordon, Kristen Dams-O'Connor, Bradley N. Delman, Edmund Wong, Cheuk Y. Tang, Steven T. DeKosky, James R. Stone, Robert C. Cantu, Mary Sano, Patrick R. Hof, and Sam Gandy. \"imaging [18f]av-1451 and [18f]av-45 in acute and chronic traumatic brain injury\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[169] Jongwan Kim, Jing Wang, Svetlana V. Shinkareva, and Douglas H. Wedell. \"affective videos\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[170] Vincent A. Magnotta, Joy T. Matsui, Dawei Liu, Hans J. Johnson, Jeffrey D. Long, Bradley D. Bolster Jr, Bryon A. Mueller, Kelvin Lim, Susumu Mori, Karl G. Helmer, Jessica A. Turner, Sarah Reading, Mark J. Lowe, Elizabeth Aylward, Laura A. Flashman, Greg Bonett, and Jane S. Paulsen. \"dwi traveling human phantom study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000206.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.745, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[171] Tetreault P, Mansour A, Vachon-Presseau E, Schnitzer TJ, Apkarian AV, and Baliki MN. \"brain connectivity predicts placebo response across chronic pain clinical trials\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds000208.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[172] Elizabeth DuPre, Wen-Ming Luh, and R. Nathan Spreng. \"multi-echo fmri replication sample of autobiographical memory, prospection and theory of mind reasoning tasks\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[173] Liane Young, Alek Chakroff, Emily Wasserman, Rebecca Saxe, James Dungan, Amelia Brown, and Jorie Koster-Hale. \"moral judgments of intentional and accidental moral" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23853" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "violations across harm and purity domains\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000212.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[174] Xiao Gao, Xiao Deng, Xin Wen, Ying She, Petra Corianne Vinke, and Hong Chen. \"neural mechanism underlying appearance social comparison\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.163, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[175] Stephen Giles, Jeremy Hall, Merrick Pope, Katie Nicol, and Liana Romaniuk. \"eupd cyberball\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[176] Alexander D. Cohen, Andrew S. Nencka, R. Marc Lebel, and Yang Wang. \"multiband multi-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[177] Avi J H Chanales, Ashima Oza, Serra E Favila, and Brice A Kuhl. \"route learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms000217.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[178] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (oral nutritional supplements)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[179] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (regular products)\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[180] Thomas H B FitzGerald, Dorothea Haemmerer, Karl J Friston, Shu-Chen Li, and Raymond J Dolan. \"sequential inference vbm\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000222.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[181] Ian Ballard, Bokyung Kim, Anthony Liatsis, Gökhan Aydogan, Jonathan D. Cohen, and Samuel M. McClure. \"magnitude effect\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[182] Evan M. Gordon, Timothy O. Laumann, Adrian W. Gilmore, Dillan J. Newbold, Deanna J. Greene, Jeffrey J. Berg, Mario Ortega, Catherine Hoyt-Drazen, Caterina Gratton, Haoxin Sun, Jacqueline M. Hampton, Rebecca S. Coalson, Annie Nguyen, Kathleen B. McDermott, Joshua S. Shimony, Abraham Z. Snyder, Bradley L. Schlaggar, Steven E. Petersen, Steven M. Nelson, and Nico U.F. Dosenbach. \"the midnight scan club (msc) dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms000224.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[183] Hilary Richardson, Grace Lisandrelli, Alexa Riobueno-Naylor, and Rebecca Saxe. \"mri data of 3-12 year old children and adults during viewing of a short animated film\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000228.v1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[184] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward - study 1\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[185] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward-study 3\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000231.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[186] Johan D Carlin and Nikolaus Kriegeskorte. \"adjudicating between face-coding models with individual-face fmri responses\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[187] Samuel A. Nastase, Yaroslav O. Halchenko, Andrew C. Connolly, M. Ida Gobbini, and James V. Haxby. \"neural responses to naturalistic clips of behaving animals in" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "two different task contexts\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000233.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[188] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 1\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[189] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000235.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[190] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 3\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps00236.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[191] Yuji Kanazawa, Kimihiro Nakamura, Toru Ishii, Toshihiko Aso, Hiroshi Yamazaki, and Koichi Omori. \"phonological memory in sign language relies on the visuomotor neural system outside the left hemisphere language network\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[192] D. Zeithamova, M. De Araujo Sanchez, and A. Adke. \"trial timing for multivariate pattern analysis\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[193] Julian Maclaren, Zhaoying Han, Sjoerd B. Vos, Nancy Fischbein, and Roland Bammer. \"maclaren test-retest brain volume dataset\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[194] Alvaro Galiano, Reyes Garcia de Eulate, Marta Vidorreta, Miriam Recio, Mario Riverol, José L. Zubieta, and Maria A. Fernandez-Seara PhD. \"resting state perfusion in healthy aging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000240.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[195] Andrew C. Connolly, J. Swaroop Guntupalli, Jason Gors, Michael Hanke, Yaroslav O. Halchenko, Yu-Chien Wu, Hervé Abdi, and James V. Haxby.\"ak6\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[196] Steve Petersen, Brad Schlaggar, and Jonathan Power. \"washington university 120\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[197] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"individual brain charting\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds000244.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[198] Guiomar Niso, Jeremy Moreau, Elizabeth Bock, Francois Tadel, and Sylvain Baillet. \"meg-bids omega restingstate_sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000247.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[199] Alexandre Gramfort and Matti S Hämäläinen. ””. Open-Neuro, 2020. doi: 10.18112/openneuro.ds000248.v1.2.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[200] Agnes Norbury and Ben Seymour. \"value generalization in human avoidance learning\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[201] Diana Gorbet and Lauren Sergio. \"female action video game players\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[202] Alexander D. Cohen, Andrew S. Nencka, and Yang Wang. \"multiband multi-echo simultaneous asl/bold for task-induced functional mri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000254.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[203] Yoichi Miyawaki, Hajime Uchida, Okito Yamashita, Masaaki Sato, Yusuke Morito, Hiroki C. Tanabe, Norihiro Sadato, and Yukiyasu Kamitani. \"visual image reconstruction\". OpenNeuro, 2018." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23854" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[204] Deanna J. Greene, Jonathan M. Koller, Jacqueline M. Hampton, Victoria Wesevich, Andrew N. Van, Annie L. Nguyen, Catherine R. Hoyt, Lindsey McIntyre, Eric A. Earl, Rachel L. Klein, Joshua S. Shimony, Steven E. Petersen, Bradley L. Schlaggar, Damien A. Fair, and Nico U.F. Dosenbach. \"behavioral interventions for reducing head motion during migraine scans in children\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[205] Jonathan Power, Valerie Voon, Prantik Kundu, Suyash Bhogawar, Peter Bandettini, and Alex Martin. \"multi-echo dambridge\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000258.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[206] Michael Milham, Cameron Craddock, Russ Tobe, Anna MacKay-Brandt, Stan Colcombe, Bennett Leventhal, and F. Xavier Castellanos. \"nki-sample\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.321 + ], + "angle": 0, + "content": "[207] Chris Gorgolewski. \"the brain of chris\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.323, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[208] Mark A. Thornton, Miriam E. Weaverdyck, and Diana I. Tamir. \"social predictive coding\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[209] A. Zadbood, J. Chen, Y.C. Leong, K.A. Norman, and U. Hasson. \"sherlock_merlin\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.382, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[210] Y. Yeshurun, M. Nguyen, and U. Hasson. \"milky-vodka\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001131.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[211] J. Chen, Y.C. Leong, C.J. Honey, K.A. Yong C.H., Norman, and U. Hasson. \"sherlock\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001132.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[212] J. Chen, C. J. Honey, E. Simony, M. J. Arcaro, K. A. Norman, and U. Hasson. \"twilight zone movie watching dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001145.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[213] Chris Gorgolewski, Natacha Mendes, Domenica Wilfling, Elisabeth Wladimirow, Claudine J. Gauthier, Tyler Bonnen, Florence J.M. Ruby, Robert Trampel, Pierre-Louis Bazin, Roberto Cozatl, Jonathan Smallwood, and Daniel S. Margulies. \"a high resolution 7-tesla resting-state fmri test-retest dataset with cognitive and physiological measures\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001168.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.639, + 0.482, + 0.666 + ], + "angle": 0, + "content": "[214] Krzysztof J. Gorgolewski. \"example mr artifacts\". Open-Neuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.669, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[215] Mark Thornton, Miriam Weaverdyck, and Diana Tamir. \"state representation pictures\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[216] Sam Gilbert and Hoki Fung. \"decoding intentions of self and others\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.796 + ], + "angle": 0, + "content": "[217] J.A. Lewis-Peacock, J.D. Cohen, and K.A. Norman. \"neural evidence of the strategic choice between working memory and episodic memory in prospective remembering\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001229.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.8, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[218] J.A. Lewis-Peacock and K.A. Norman. \"competition between items in working memory leads to forgetting\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001232.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.843, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[219] Patrick Beukema and Timothy Verstynen. \"singlefingerrsa\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[220] Christopher Baldassano, Diane M. Beck, and Li Fei-Fei. \"human-object interaction\". OpenNeuro, 2018." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[221] Michelle G. Hall, Claire K. Naughtin, Jason B. Mattingley, and Paul E. Dux. \"olvsl_ object-location visual statistical learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001241.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[222] Tae-Ho Lee, Steven Greening, Taiji Ueno, David Clewett, Allison Ponzio, Michiko Sakaki, and Mara Mather. \"examining effects of arousal on responses to salient and non-salient stimuli in younger and older adults\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001242.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[223] Tomoyasu Horikawa and Yukiyasu Kamitani. \"generic object decoding (fmri onImagenet)\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001246.v1.2.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[224] Damien Brevers, Sarah C Herremans, Qinghua He, Marie-Anne Vanderhasselt, Mathieu Petieau, Dimitri Verdonck, Tasha Poppa, Sara De Witte, Charles Kornreich, Antoine Bechara, and Chris Baeken. \"sports_betting_study\". Open-Neuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.329, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[225] Matteo Visconti di Oleggio Castello, Yaroslav O. Halchenko, J. Swaroop Guntupalli, Jason D. Gors, and M. Ida Gobbini. \"personally familiar faces\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.385, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[226] Frank J. Kanayet, Andrew Mattarella-Micke, Peter J. Kohler, Anthony M. Norcia, Bruce D. McCandliss, and James L. McClelland. \"numberlinemarking\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001299.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.441, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[227] Sean O'Bryan, Darrell Worthy, Evan Livesey, and Tyler Davis. \"inverse base rate\". OpenNeuro, 2018. doi: 10.18112/openneuro.ms001302.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.483, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[228] Reshanne R. Reeder, Christian N. L. Olivers, and Stefan Pollmann. \"negative_template\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001306.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.525, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[229] E.S. Finn, P.R. Corlett, G. Chen, P.A. Bandettini, and R.T. Constable. \"paranoiastory\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001338.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.567, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[230] Ivy F. Tso, Saige E. Rutherford, Yu Fang, Mike Angstadt, and Stephan F. Taylor. \"socialbrain\". OpenNeuro, 2018. doi: 10.52181/zenodod.1210217." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.608, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[231] Jessica-Lily Harvey, Lysia Demetriou, John McGonigle, and Matthew B. Wall. \"phfmri_control_tasks\". OpenNeuro, 2018. doi: 10.18112/openneuro.rs001344.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.651, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[232] Michael P Notter, Sandra Da Costa, and Micah M Murray. \"decoding of multisensory semantics and memories in low-level visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001345.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.706, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[233] Brock Kirwan, Thomas Meservy, and Kelly Fadel. \"information filtering in electronic networks of practice: An fmri investigation of expectation [dis]confirmation\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001353.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.762, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[234] Amado Catarina, Kovács Petra, Mayer Rebecca, Ambrus G. Geza, Trapp Sabrina, and Kovács Gyula. \"namefacepriming\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.804, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[235] Christopher Baldassano. \"loci case study\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001379.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.832, + 0.87, + 0.845 + ], + "angle": 0, + "content": "[236] Dr. Frank Hillary. \"subject data\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[237] Joset A. Etzel, Debbie Yee, Bidhan Lamichhane, Mitchell Jeffers, Elisa Di Rosa, Jenny Crawford, Hongyu An, and Todd S. Braver. \"multibandacqtests\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso001399.v2.0.0." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23855" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[238] Dobrushina Olga R, Vlasova Roza M, Rumshiskaya Alena D, Litvinova Liudmila D, Mershina Elena A, Sinit-syn Valentin E, and Pechenkova Ekaterina V. \"rsfmri_single_session_eeg_nf\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001408.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[239] Megen Brittell. \"audio cartography\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001415.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.195, + 0.482, + 0.236 + ], + "angle": 0, + "content": "[240] Rotem Botvinik-Nezer, Tom Salomon, and Tom Schonberg. \"cat snacks functional plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001417.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.239, + 0.482, + 0.349 + ], + "angle": 0, + "content": "[241] Martin Wegrzyn, Joana Aust, Larissa Barnstorf, Magdalena Gippert, Mareike Harms, Antonia Hautum, Shanna Heidel, Friederike Herold, Sarah M. Hommel, Anna-Katharina Knigge, Dominik Neu, Diana Peters, Marius Schaefer, Julia Schneider, Ria Vombrock, Sabrina M. Zimmer, Friedrich G. Woermann, and Kirsten Labudda. \"thought-experiment\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001419.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[242] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]dasb pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001420.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[243] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]sb207145 pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001421.v1.4.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[244] G. Kim, K.A. Norman, and N.B. Turk-Browne. \"neural overlap in item representations across episodes impairs context memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001430.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.482, + 0.569 + ], + "angle": 0, + "content": "[245] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001439.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.571, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[246] A.C. Schapiro, E.A. McDevitt, T.T. Rogers, S.C. Mednick, and K.A. Norman. \"human hippocampal replay during rest prioritizes weakly learned information and predicts memory performance\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001454.v1.3.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.726 + ], + "angle": 0, + "content": "[247] Ilaria Berteletti, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, Jérôme Prado, Macarena Suarez Pellicioni, Jessica Wise Younger, and James R. Booth. \"brain correlates of math development\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds001486.v1.3.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[248] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, and Gert J. ter Horst. \"valence processing differs across stimulus modalities (multi-echo)\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps001491.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[249] J Lewis-Peacock, BR Postle, CR Cox, and TT Rogers. \"long-term memory (ltm) for famous faces, places, and common objects\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001497.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[250] Nadine Chang, John A. Pyles, Austin Marcus, Abhinav Gupta, Michael J. Tarr, and Elisaa M. Aminoff. \"bold5000\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001499.v1.3.1." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[251] Guohua Shen, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"deep image reconstruction\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001506.v1.3.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[252] C. Baldassano, R. Masis-Obando, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (intact)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001510.v2.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[253] P. Kok and N.B. Turk-Browne. \"associative prediction of visual shape in the hippocampus\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001517.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[254] S. C. Chan, M. C. Applegate, N. W. Morton, S. M. Polyn, and K. A. Norman. \"lingering representations of stimuli influence recall organization\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001521.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[255] Feng Rong, A. Lisette Isenberg, Erica Sun, and Gregory S. Hickok. \"syllable level speech sequencing\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001525.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[256] Andrea Courtney, Emma PeConga, Dylan Wagner, and Kristina Rapuano. \"calorie-labeled food cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001534.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.377, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[257] M. Aly, J. Chen, N.B. Turk-Browne, and U. Hasson. \"learning naturalistic temporal structure in the posterior medial network\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001545.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.433, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[258] Eliska Prochazkova. \"ds: ds001551:1.0.0\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[259] Javier Gonzalez-Castillo, Ziad S. Saad, Daniel A. Handwaker, Souheil J. Inati, Noah Brenowitz, and Peter A. Bandettini. \"100 runs at 3t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001553.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.518, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[260] O. Lositsky, J. Chen, D. Toker, C.J. Honey, J.L. Poppenk, U. Hasson, and K.A. Norman. \"neural pattern change during encoding of a narrative predicts retrospective duration estimates\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.575, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[261] Javier Gonzalez-Castilloa, Colin W. Hoy, Daniel A. Handwerker, Vinai Roopchansingh, Souheil J. Inati, Ziad S. Saad, Robert W. Cox, and Peter A. Bandettini. \"large single-subject functional mri datasets at 7t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001555.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.645, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[262] Renzo Huber. \"cmrr workshop\". OpenNeuro, 2019. doi: mockDOI." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.674, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[263] Unknown authors. \"ds: ds001566:1.0.1\". OpenNeuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[264] A.N. Hoskin, A.M. Bornstein, K.A. Norman, and J.D. Cohen. \"refresh my memory: Episodic memory reinstatements intrude on working memory maintenance\". Open-Neuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[265] Ian C. Ballard, Anthony D. Wagner, and Samuel M. McClure. \"feature discrimination\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001590.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.802, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[266] Unknown authors. \"test dataset for xcp software\". Open-Neuro, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[267] Patrick Beukema and Timothy Verstynen. \"multifingerrsa\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001597.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[268] Matt Cieslak, Mark Elliott, and Ted Satterthwaite. \"example fieldmaps\". OpenNeuro, 2018." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23856" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[269] A.M. Bornstein and K.A. Norman. \"reinstated episodic context guides sampling-based decisions for reward\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001607.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[270] Aya Ben-Yakov and Noga Cohen. \"prestimulus exp2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001608.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[271] I. Momennejad, A.R. Otto, N. Daw, and K.A. Norman. \"offline replay supports planning in human reinforcement learning\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001612.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[272] Adam Hampshire and Eyal Soreq. \"visual working memory: Study one task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001634.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[273] Adam Hampshire and Eyal Soreq. \"visual working memory: Study two task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001635.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.482, + 0.367 + ], + "angle": 0, + "content": "[274] Elizabeth Valles. \"ironia vev\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001652.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.37, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[275] Mattia Veronese, Gaia Rizzo, Martin Belzunce, Julia Schubert, Barbara Santangelo, Ayla Mansur, Alex Whittington, Joel Dunn, Graham Searle, Andrew Reader, and Roger Gunn. \"nrm2018 pet grand challenge dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms001705.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[276] A. Shenhav, M.A. Straccia, S. Musslick, J.D. Cohen, and M.M. Botvinick. \"dissociable neural mechanisms track evidence accumulation for selection of attention versus action\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.499, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[277] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others (experiment 2)\". OpenNeuro, 2019. doi: 10.18112/openneuro. ds001722.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[278] Unknown authors. \"test-geier: Ds: ds001728:1.1.0\". Open-Neuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[279] Rotem Botvinik-Nezer, Roni Iwanir, Russell A. Poldrack, and Tom Schonberg. \"narps\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s001734.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.658, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[280] Birgit Rauchbauer, Bruno Nazarian, Morgan Bourhis, Magalie Ochs, Laurent Prevot, and Thierry Chaminade. \"brain activity during reciprocal social interaction investigated using conversational robots as control condition\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001740.v2.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[281] Nathaniel P Miller, Yao Liu, Roman Krivochenitser, and Bas Rokers. \"unilateral glaucoma 3t dmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001743.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[282] J.R. Manning, J.C. Hulbert, J. Williams, L. Piloto, L. Sahakyan, and K.A. Norman. \"a neural signature of contextually mediated intentional forgetting\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001745.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[283] Carrie E. Gold, Athena L. Howell, Jacob Burdis, C. Brock Kirwan, and Gregory L. Thompson. \"exploring the resting state neural activity of monolinguals and late and early bilinguals\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001747.v1.1.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[284] Samuel Fynes-Clinton, Lars Marstaller, and Hana Burianova. \"differentiation of functional networks during long-term memory retrieval in children and adolescents\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001748.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[285] Bart Aben, Cristian Buc Calderon, Laurens Van der Cruyssen, Doerte Picksak, Eva Van den Bussche, and Tom Verguts. \"context-dependent cognitive control — flanker task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001751.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[286] R. VanRullen and L. Reddy. \"reconstructing faces from fmri patterns using deep generative neural networks\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001761.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[287] A. Shenhav, M.A. Straccia, M.M. Botvinick, and J.D. Cohen. \"dorsal anterior cingulate and ventromedial prefrontal cortex have inverse roles in both foraging and economic choice\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[288] Virginia AGLIERI, Bastien CAGNA, Pascal BELIN, and Sylvain TAKERKART. \"intertva: a multimodal mri dataset for the study of inter-individual differences in voice perception and identification\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001771.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[289] J.C. Hulbert and K.A. Norman. \"neural differentiation tracks improved recall of competing memories following interleaved study and retrieval practice\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001775.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[290] Alik Wadge and Sam Zorowitz. \"effects of on/off deep brain stimulation on cognitive control in treatment-resistant depression (eeg)\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001784.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[291] Vincent DeLuca and Christos Pliatsikas. \"bilingualism and the brain\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds001796.v1.7.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[292] Samuel Zorowitz, Alexander P Rockhill, Kristen K Kellard, Katherine E Link, Todd Herrington, Diego A Pizzagalli, Alik S Widge, Thilo Deckersbach, and Darin D Dougherty. \"neuroanatomical correlates of approach-avoidance conflict (fmri)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds001814.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[293] Unknown authors. \"cat faces mri experiment. ds: ds001818:1.0.0\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[294] Isabel Alkhasli, Katrin Sakreida, Felix M. Mottaghy, and Ferdinand Binkofski. \"resting state - tms\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001832.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[295] Celia Goffin, H. Moriah Sokolowski, Michael Slipenkyj, and Daniel Ansari. \"handedness and symbolic number representation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001838.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[296] Unknown authors. \"spinal stimulation stepping and standing dataset. ds: ds001839:1.0.1\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[297] Noga Cohen and Aya Ben-Yakov. \"prestimulus exp1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001840.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[298] S.C.Y. Chan, Y. Niv, and K.A. Norman. \"a probability distribution over latent causes, in the orbitofrontal cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001847.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23857" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[299] H. Moriah Sokolowski, Zachary Hawes, Lien Peters, and Daniel Ansari. \"parallel adaptation of symbols, quantities, and physical size\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001848.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[300] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making intertemporal choice task dataset\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[301] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making risky choice task dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001883.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[302] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"longitudinal brain correlates of multisensory lexical processing in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001894.v1.4.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.347, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[303] Trevor K. M. Day, Tara M. Madyastha, Peter Board, Mary K. Askren, Nicholas J. Wapstra, Thomas J. Montine, and Thomas J. Grabowski. \"ant: Healthy aging and parkinson's disease\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001907.v3.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.472 + ], + "angle": 0, + "content": "[304] Milena Kaestner, Ryan T. Maloney, Kirstie H. Wailes-Newson, Marina Bloj, Julie M. Harris, Antony B. Morland, and Alex R. Wade. \"s-cone_motiondepth_fmri\". Open-Neuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[305] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[306] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value (2)\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[307] Ghootae Kim, Kenneth A. Norman, and Nicholas B. Turk-Browne. \"violation differentiation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001926.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[308] Gabriel Castrillon, Nico Sollmann, Katarzyna Kurcyus, Adeel Razi, Sandro M. Krieg, and Valentin Riedl. \"the physiological effects of non-invasive brain stimulation fundamentally differ across the human cortex\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001927.v2.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.756 + ], + "angle": 0, + "content": "[309] N.C. Hindy, E.W. Avery, and N.B. Turk-Browne. \"hippocampal-neocortical interactions sharpen over time for predictive actions\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001946.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[310] Peter J. Kohler and Anthony M. Norcia. \"svndl anatomical data\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[311] Peter J. Kohler, Benoit R. Cotterreau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (central letter task)\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[312] Peter J. Kohler, Benoit R. Cottereau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (passive fixation)\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[313] Benjamin J. Griffiths, Stephen D. Mayhew, Karen J. Mullinger, João Jorge, Ian Charest, Maria Wimber, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Simon Hanslmayr. \"reinstatement_fidelity\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002000.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[314] Janine Mendola and Elizabeth Bock. \"rivalry_tagging\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002001.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[315] Akram Bakkour, Daphna Shohamy, and Michael N. Shadlen. \"memory and decision making dataset\". Open-Neuro, 2019. doi: 10.18112/openneuro.ms002006.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[316] David Clewett, Ringo Huang, Rico Velasco, Tae-Ho Lee, and Mara Mather. \"locus coeruleus activity strengthens prioritized memories under arousal\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002011.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[317] Jakob Heinzle (data acquisition), Kai Goergen (data curation), and Joram Soch (BIDS conversion). \"stimdisc\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002013.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[318] Martin Wegrzyn, Markus Mertens, Christian G. Bien, Friedrich G. Woermann, and Kirsten Labudda. \"language production fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002014.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[319] Yiming Xiao, Jonathan C. Lau, Taylor Anderson, Jordan DeKraker, D. Louis Collins, Terry M. Peters, and Ali R. Khan. \"bigbrainmricoreg\". OpenNeuro, 2019. doi: 10.18112/openneuro.d0s002016.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[320] Rebecca Merkley, Benjamin Conrad, Gavin Price, and Daniel Ansari. \"investigating the visual number form area: A replication study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002033.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.89, + 0.504 + ], + "angle": 0, + "content": "[321] Unknown authors. \"bids_parkinson\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[322] Jaime J. Castrellon, Jacob S. Young, Linh C. Dang, Ronald L. Cowan, David H. Zald, and Gregory R. Samanez-Larkin. \"mesolimbic dopamine d2 receptors and neural representations of subjective value\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002041.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[323] Emily S Finn, Laurentius Huber, David C Jangraw, Peter J Molfese, and Peter A Bandettini. \"layer-dependent activity in human prefrontal cortex during working memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002076.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[324] Hannelore Aerts and Daniele Marinazzo. \"btc_postop\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002080.v4.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[325] Maedbh King, Carlos Hernandez-Castillo, Russell A. Poldrack, Jonathon Walters, Richard B. Ivry, and Jorn Diedrichsen. \"multi-domain task battery (mdtb)\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002105.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[326] Celia Goffin, Michael Slipenkyj, Stephan E. Vogel, Rebecca Merkley, and Daniel Ansari. \"development of symbolic number processing\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002116.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[327] Adam Kimbler, Amanda G. Hamm, and Aaron T. Mattfeld. \"conditional visual associative learning task\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002149.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[328] Charlie, John, Ross, and Joey. \"testrun\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002153.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[329] Florence Bouhali, Zoé Bézagu, Stanislas Dehaene, and Laurent Cohen. \"a mesial-to-lateral dissociation for ortho-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23858" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "graphic processing in the visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso002155.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[330] Dylan M. Nielson and Argyris Stringaris. \"multi-echo masking test dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002156.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[331] Michael Pereira, Nathan Faivre, Inaki Iturrate, Marco Wirthlin, Luana Serafini, Stephanie Martin, Arnaud Desvachez, Olaf Blanke, Dimitri Van de Ville, and Jose del R. Millan. \"disentangling the origins of confidence in speeded perceptual judgments through multimodal imaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002158.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[332] Michelle I. Nash, Cooper B. Hodges, Nathan M. Muncy, and C. Brock Kirwan. \"pattern separation beyond the hippocampus: A high-resolution whole-brain investigation of mnemonic discrimination in healthy adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002168.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[333] Jiefeng Jiang, Shao-Fang Wang, Wanjia Guo, Corey Fernandez, and Anthony D. Wagner. \"ccd\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso02169.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[334] Tali Weiss, Timna Soroka, Lior Gorodisky, Sagit Shushan, Kobi Snitz, Reut Weisgross, Edna Furman-Haran, Thijs Dhollander, and Noam Sobel. \"human olfaction without apparent olfactory bulbs\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002185.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.447, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[335] Michele Svanera and Lars Muckli. \"cerebrum: a 3t mri segmentation tool\". OpenNeuro, 2019. doi: mockDOI." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[336] Dorit Kliemann and Ralph Adolphs. \"caltech rsfmri dataset\". OpenNeuro, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[337] Jordan Bigio, Tali Bitan, Douglas Bolger, Douglas Burman, Fan Cao, Tai-Li Chou, Nadia Cone, Jessica Gayda, Dong Lu, Marisa Lytle, Jenni Minas, and James R. Booth. \"cross-sectional multidomain lexical processing\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002236.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.577, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[338] J. Poppenk and K. A. Norman. \"briefly cueing memories leads to suppression of their neural representations\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds002237.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[339] J. Fan, J.D. Wammes, J.B. Gunn, D.L. Yamins, K.A. Norman, and N.B. Turk-Browne. \"relating visual production and recognition of objects in human visual cortex\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002241.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[340] Nathan M. Muncy and C. Brock Kirwan. \"correcting false memories: The effect of mnemonic generalization on original memory traces\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002242.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[341] Joshua Hirschfeld-Kroen, Emily Wasserman, Stefano Anzellotti, and Liane Young. \"when my wrongs are worse than yours: behavioral and neural asymmetries in first-person and third-person perspectives of accidents\". Open-Neuro, 2020. doi: 10.18112/openneuro.d002250.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[342] Jean-Marie Bonny, Juliana Iranpour, and Béatrice Claise. \"viewing_food_cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002270.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[343] Katherine L. Bottenhorn, Taylor Salo, Julio A. Peraza, Michael C. Riedel, Jessica S. Flannery, Adam Kimbler, Alfredo Toll, Diego Suarez, Francis M. Cruz, Israel Zagales, Nayade Caldes, Olivia Dolan, Ruth Zagales, Matthew T." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Sutherland, Robert W. Laird, and Angela R. Laird. \"dense investigation of variability of affect (diva)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002278.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[344] Unknown authors. \"fsl open science dev dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002293.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[345] Elizabeth Long, Nathan Wheeler, and William Cunningham. \"trait-role\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002294.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.209, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[346] J. Poppenk and K.A. Norman. \"multiple-object tracking as atool for parametrically modulating memory reactivation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002311.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[347] P. G. Bissett, J. M. Shine, J. Durnez, J. Li, K. J. Gorgolewski, and R. A. Poldrack. \"rewardbeast\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002316.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[348] Luiz Pessoa, Christian Meyer, and Srikanth Padmala. \"dynamic_passive-threat\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002320.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[349] Shohini Bhattasali, Jonathan R. Brennan, Wen-Ming Luh, Berta Franzluebbers, and John T. Hale. \"the alice dataset: fmri dataset to study natural language comprehension in the brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002322.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.422, + 0.905, + 0.449 + ], + "angle": 0, + "content": "[350] Taylor Hanayik. \"fsl open science dev dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms002328.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.451, + 0.905, + 0.492 + ], + "angle": 0, + "content": "[351] A. Sunavsky and J. Poppenk. \"neuroimaging predictors of creativity in healthy adults\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002330.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.494, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[352] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp1\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002336.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.579, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[353] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp2\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002338.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.663, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[354] Samuel A. Nastase, Yun-Fei Liu, Hanna Hillman, Asieh Zadbood, Liat Hasenfratz, Neggin Keshavarzian, Janice Chen, Christopher J. Honey, Yaara Yeshurun, Mor Regev, Mai Nguyen, Claire H. C. Chang, Christopher Baldassano, Olga Lositsky, Erez Simony, Michael A. Chow, Yuan Chang Leong, Paula P. Brooks, Emily Micciche, Gina Choe, Ariel Goldstein, Tamara Vanderwal, Yaroslav O. Halchenko, Kenneth A. Norman, and Uri Hasson. \"narratives\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002345.v1.1.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[355] Brock Kirwan, Daniel Bjornn, Bonnie Anderson, Anthony Vance, David Eargle, and Jeffrey Jenkins. \"repetition of computer security warnings results in differential repetition suppression effects as revealed with functional mri\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002363.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[356] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "\"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds002366.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[357] Deni Kurban, Renzo Huber, Gilad Liberman, Dimo Ivanov, Sri Kashyap, Debora Niekaemper, and Benedikt Poser. \"ohbm abstract 2020\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002367.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[358] M.D. Rajendra Morey. \"serumgreymattercorticalthickness\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002380.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[359] Rogers CS, Jones MS, McConkey S, Spehar B, Van Engen KJ, Sommers MS, and Peelle JE. \"age-related differences in auditory cortex activity during spoken word recognition\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002382.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[360] Yoshiharu I Kutani, Takatomi Kubo, Satoshi Nishida, Hideaki Hata, Kenichi Matsumoto, Kazushi Ikeda, and Shinji Nishimoto. \"fmri dataset on program comprehension and expertise\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002411.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[361] Jelle R. Dalenberg, Barkha P. Patel, Raphael Denis, Maria G. Veldhuizen, Yuko Nakamura, Petra C. Vinke, Serge Luquet, and Dana M. Small. \"short-term consumption of sucralose with, but not without, carbohydrate impairs neural and metabolic sensitivity to sugar\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002419.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[362] Yana Panikratova, Alexander Tomyshev, Ekaterina Pechenkova, and Roza Vlasova. \"fmri: resting state and arithmetic task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002422.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[363] James R. Booth, GE Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, MA Stein, and Michael Tennekoon. \"working memory and reward in children with and without attention deficit hyperactivity disorder (adhd)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002424.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[364] Damien Brevers, Chris Baeken, Antoine Bechara, Qinghua He, Pierre Maurage, Mathieu Petieau, Guillaume Sescousse, Claus Vögele, and Joel Billieux. \"increased brain reactivity to gambling unavailability is a marker of problem gambling\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002513.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[365] Tamar Kolodny, Michael-Paul Schallmo, and Scott O. Murray. \"contrast response functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002522.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[366] S. Jahfari, K.R. Ridderinkhof, A.G.E. Collins, T. Knapen, L.J. Waldorp, M.J. Frank, and J. Theeuwes. \"learning, inhibitory control, and perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s002543.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[367] Suzanne Oosterwijk, Lukas Snoek, Mark Rotteveel, Lisa Feldman Barrett, and H. Steven Scholte. \"shared-states\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002547.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[368] Laboratory for Computational Neurodiagnostics. \"protecting the aging brain (pagb)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002549.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[369] Romain Quentin, Jean-Remi King, Etienne Sallard, Nathan Fishman, Ryan Thompson, Ethan Buch, and Leonardo" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Cohen. \"differential brain mechanisms of selection and maintenance of information during working memory (meg data)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002550.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[370] Tomas Knapen, Wietske van der Zwaag, and Daan van Es. \"cerebellum retinotopic mapping\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002574.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[371] Arnaud Delorme and Scott Makeig. \"visual oddball task (256 channels)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs002578.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[372] S. Zhang, W. Yoshida, H. Mano, T. Yanagisawa, K. Shibata, M. Kawato, and B. Seymour. \"cognitive control of sensory pain encoding in the pregenual anterior cingulate cortex. d1 - decoder construction in day 1, d2 - adaptive control in day 2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002596.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[373] Lukas Snoek, Tinka Beemstermboer, and H. Steven Scholte. \"mb-epi-comparison\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002603.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[374] Caroline Le Ster, Antonio Moreno, Franck Mauconduit, Vincent Gras, Ruediger Stirnberg, Benedikt A. Poser, Alexandre Vignaud, Evelyn Eger, Stanislas Dehaene, Florent Meyniel, and Nicolas Boulant. \"bids_dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002606.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[375] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study1\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002608.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[376] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study2\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002609.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[377] Enzo Tagliazucchi. \"meditation interocepcion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002614.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[378] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum. \"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002620.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[379] N Eichert, KE Watkins, and project-specific co authors. \"project_larynx\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002634.v3.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[380] L.N. Van der Laan, C. Scholz, R.A. Poldrack, D.T.D. De Ridder, and A. Smidts. \"can we have a second helping? a replication study on the neurobiological mechanisms underlying self-control\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002643.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[381] Mara Mather, Ringo Huang, David Clewett, Shawn Nielsen, Rico Velasco, Kristie Tu, Sophia Han, and Briana Kennedy. \"isometric exercise facilitates attention to salient events in women via the noradrenergic system\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002647.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[382] Nathan M. Muncy and C. Brock Kirwan. \"the medial temporal lobe supports mnemonic discrimination for event duration\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002655.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[383] Brian J. Coffey, Zachary D. Threlkeld, Andrea S. Foulkes, Yelena G. Bodien, and Brian L. Edlow. \"the language" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23860" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "network reemerges during recovery from severe traumatic brain injury\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002675.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[384] Cheryl A. Olman and Kimberly B. Weldon. \"kung fu panda\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002684.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[385] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"ibc\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs002685.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[386] James R. Booth, Gillian E. Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, Mark A. Stein, and Michael Tennekoon. \"working memory and reward in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002687.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[387] Kendrick Kay, Keith W. Jamison, Luca Vizioli, Ruyuan Zhang, Eshed Margalit, and Kamil Ugurbil. \"high-field 7t visual fmri datasets\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002702.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[388] Christos Chatzichristos, Manuel Morante, Nikos Andreadis, Eleftherios Kofidis, Yiannis Kopsinis, and Sergios Theodoridis. \"smiley\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002711.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.427 + ], + "angle": 0, + "content": "[389] Sara Aurtenetxe, Nicola Molinaro, Doug Davidson, and Manuel Carreiras. \"numbers and letters\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002712.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[390] Petra Vetter, Lukasz Bola, Lior Reich, Matthew Bennett, Lars Muckli, and Amir Amedi. \"avscenes_blind\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002715.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[391] Liron Rozenkrantz, Reut Weissgross, Tali Weiss, Inbal Ravrebi, Idan Frumin, Sagit Shushan, Lior Gorodisky, Netta Reshef, Yael Holzman, Liron Pinchover, Yaara Endevelt-Shapira, Eva Mishor, Timna Soroka, Maya Finkel, Liav Tagania, Aharon Ravia, Ofer Perl, Edna Furman-Haran, Howard Carp, and Noam Sobel. \"unexplained repeated pregnancy loss is associated with altered perceptual and brain responses to men's body-odor\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002717.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[392] Ian Daly, Nicoletta Nicolaou, Duncan Williams, Faustina Hwang, Alexis Kirke, Eduardo Miranda, and Slawomir J. Nasuto. \"a dataset recording joint eeg-fmri during affective music listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002725.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.733 + ], + "angle": 0, + "content": "[393] Jordi Sole-Casals, Josep M Serra-Grabulosa, Rafael Romero-Garcia, Gemma Vilaseca, Ana Adan, Nuria Vilaro, Nuria Vilaro, and Edward T. Bullmore. \"structural brain network of gifted children\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002726.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[394] Francesca M M Citron, Cristina Cacciari, Jakob M Funcke, Chun-Ting Hsu, and Arthur Jacobs. \"emotive idioms\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002727.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[395] Bi Zhu, Chuansheng Chen, Xuhao Shao, Wenzhi Liu, Zhi-fang Ye, Liping Zhuang, Li Zheng, Elizabeth F. Loftus, and Gui Xue. \"multiple interactive memory representations underlie the induction of false memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002731.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[396] Nora C. Heinzelmann, Susanna C. Weber, and Philippe N. Tobler. OpenNeuro, 2020. doi: 10.18112/openneuro.ds002732.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "[397] Test. \"bids pilot project\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002733.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[398] Andrea Pisauro, Elsa Fouragnan, C Retzler, and M Philiastides. \"evidence accumulation in value-based decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002734.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[399] Jonathan D Power, Benjamin Silver, Melanie Silverman, Eliana Ajodan, Dienke Bos, and Rebecca M. Jones. \"headmold\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002735.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[400] Joset A. Etzel and Todd S. Braver. \"multibandcftests\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002737.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[401] Sabina Gherman and Marios G. Philiastides. \"simultaneous eeg-fmri - confidence in perceptual decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002739.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.362 + ], + "angle": 0, + "content": "[402] N Furl, F Begum, J Sulik, F Pizzorni Ferrarese, S Jans, and C Wooley. \"animated caricatures fmri study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002741.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[403] AR. \"test1\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002743.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.393, + 0.906, + 0.448 + ], + "angle": 0, + "content": "[404] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"resting state with closed eyes for patients with depression and healthy participants\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002748.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.906, + 0.477 + ], + "angle": 0, + "content": "[405] Chao-Gan Yan. \"yandatabids\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002750.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.479, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[406] Dillan J. Newbold, Timothy O. Laumann, Catherine R. Hoyt, Jacqueline M. Hampton, David F. Montez, Ryan V. Raut, Mario Ortega, Anish Mitra, Ashley N. Nielsen, Derek B. Miller, Babatunde Adeyemo, Annie L. Nguyen, Kristen M. Scheidter, Aaron B. Tanenbaum, Andrew N. Van, Scott Marek, Bradley L. Schlaggar, Alexandre R. Carter, Deanna J. Greene, Evan M. Gordon, Marcus E. Raichle, Steven E. Petersen, Abraham Z. Snyder, and Nico U.F. Dosenbach. \"cast-induced plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002766.v3.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.906, + 0.686 + ], + "angle": 0, + "content": "[407] Damien Brevers, Claus Vögele, Pierre Maurage, Guillaume Sescousse, Chris Baeken, and Joel Billieux. \"brain mechanisms underlying episodic future thinking of sustainable behaviors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002770.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[408] Zhifang Ye and Gui Xue. \"retrieval practice facilitates memory updating by enhancing and differentiating medial prefrontal cortex representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002773.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[409] Eva Berlot, Nicola Popp, and Jörn Diedrichsen. \"motor sequence learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002776.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[410] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002785.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[411] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002790.v2.0.0." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23861" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[412] Minjae Kim, Peter Mende-Siedlecki, Stefano Anzellotti, and Liane Young. \"tom following strong vs weak priors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002793.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[413] Jiefeng Jiang, Anthony D. Wagner, and Tobias Egner. \"elife39497\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002797.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[414] Thompson WH*, Nair R*, Oya H*, Esteban O, Shine JM, Petkov CI, Poldrack RA, Howard M, Adolphs R†, and †corresponding author *equally contributing. \"human es-fmri resource: Concurrent deep-brain stimulation and whole-brain functional migraine\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002799.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[415] C.R. Bowman, T. Iwashita, and D. Zeithamova. \"model-based fmri reveals co-existing specific and generalized concept representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002813.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[416] Fatemeh Ebrahiminia, Morteza Mahdiani, and Seyed-Mahdi Khaligh-Razavi. \"a multimodal neuroimaging dataset to study spatiotemporal dynamics of visual processing in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002814.v1.3.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[417] Sangil Lee, Trishala Parthasarathi, and Joseph Kable. \"prospection\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso02835.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[418] Sarah Aliko, Jiawen Huang, Florin Gheorghiu, Stefanie Meliss, and Jeremy I Skipper. \"naturalistic neuroimaging database\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002837.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[419] Sarah Schwettmann, Joshua B. Tenenbaum, and Nancy Kanwisher. \"intuitive physics with fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002841.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[420] Omar Maxim. \"mriquality\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002848.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[421] Caroline Landelle, Jean-Luc Anton, Bruno Nazarian, Julien Sein, Ali Gharbi, Olivier Felician, and Anne Kavounoudias. \"aging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002872.v1.3.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[422] Christian Roming, Karl Koschutnig, Andreas Fink, and ... \"creative direccion making in soccer\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002878.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[423] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"brain correlates of multisensory lexical processing in adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds002879.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[424] James R. Booth, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, and Jérôme Prado. \"brain development of deductive reasoning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002886.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[425] Agnès Job, Chloé Jaroszynski, Anne Kavounoudias, Assia Jaillard, and Chantal Delon-Martin. \"resting state fmri study of non bothersome tinnitus\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002896.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[426] Sharna D. Jamadar, Phillip G.D. Ward, Thomas G. Close, Alex Fornito, Malin Premaratne, Kieren O'Brien, Daniel" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Staeb, Zhaolin Chen, N. Jon Shah, and Gary F. Egan. \"monash rspet-mr\". OpenNeuro, 2024. doi: 10.18112/openneuro.d002898.v1.4.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[427] Christophe Pallier and Laurent Bonnasse-Gahot. \"simpcomp_rawdata\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002905.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[428] Rafael Garcia-Dias. \"neuroharmony sample example\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002936.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[429] Bart Aben, Cristian Buc Calderon, Eva Van den Bussche, and Tom Verguts. \"task-dependent effort-induced connectivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002938.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[430] DC Monroe, RS Blumenfeld, SL Small, and DB Keator. \"head-to-ball impacts in collegiate soccer players\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002940.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[431] Ruiqing Zhang. \"narrative listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002979.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.392 + ], + "angle": 0, + "content": "[432] Sangil Lee, Nicole Cooper, and Joseph Kable. \"delay discounting bidding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002989.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.394, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[433] Gabriel Pelletier, Nadav Aridan, Lesley K Fellows, and Tom Schonberg. \"fribblesfmri_object-value-construction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002994.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.45, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[434] Jason A Avery, Alexander G Liu, Cameron Riddell, John E Ingelholm, Stephen J Gotts, and Alex Martin. \"taste quality representation in the human brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002995.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.576 + ], + "angle": 0, + "content": "[435] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"two sessions of resting state with closed eyes for patients with depression in treatment course (nfb, cbt or no treatment groups)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003007.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[436] Colin Hawco, Erin Dickie, Gabrielle Herman, Jessica Turner, Miklos Argyelan, Philipp Homan, Anil Malhotra, Robert Buchanan, and Aristotle Voineskos. \"social processes initiative in neurobiology of the schizophrenia(s) traveling human phantoms\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003011.v1.2.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[437] M. Martinez-Saito, R. Konovalov, M. A. Piradov, A. Shestakova, B. Gutkin, and V. Klucharev. \"action-in-auctions\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003012.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[438] Amanda LeBel, Lauren Wagner, Shailee Jain, Aneesh Adhikari-Desai, Bhavin Gupta, Allyson Morgenthal, Jerry Tang, Lixiang Xu, and Alexander G. Huth. \"an fmri dataset during a passive natural language listening task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms003020.v3.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[439] Diego Angeles-Valdez, Jalil Rasgado-Toledo, Viviana Villicana, Alan Davalos-Guzman, Alfonso Fajardo-Valdez, Cristina Almanza, Ruth Alcala-Lozano, and Eduardo A. Garza-Villarreal. \"sudmex_tms\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003037.v2.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[440] Karen T. Navarro, Cheryl A. Olman, and Kimberly B. Weldon. \"dichoptic magno/parvo\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003043.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23862" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[441] Krueger O and Klose U. \"dti data from 'fiber architecture in the ventromedial striatum and its relation with the bed nucleus of the stria terminalis\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003047.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[442] Robin Carhart-Harris et al. \"neural correlates of the lsd experience revealed by multimodal neuroimaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003059.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[443] James R. Booth, Marisa N. Lytle, Rachna Mureja, and Jérôme Prado. \"brain correlates of deductive reasoning in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003076.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[444] Philippe DOMENECH, Sylvain RHEIMS, and Etienne KOECHLIN. \"probe ieeg\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms003078.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[445] Jonathan Cote and Etienne de Villers-Sidani. \"auditory cortex mapping dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003082.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[446] James R. Booth, Amy S. Desroches, Marisa N. Lytle, Rucha Mehta, Jennifer E. Minas, Rachna Mutreja, Jérôme Prado, and Hongchuan Zhang. \"brain correlates of math processing in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003083.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[447] Matthew Sachs, Assal Habibi, Antonio Damasio, and Jonas Kaplan. \"temporal dynamics of emotional music\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003085.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[448] Noam Saadon-Grosman, Shahar Arzy, and Yonatan Loewenstein. \"somatosensory phase-encoded bilateral full-body light touch stimulation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003089.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[449] Jeffrey N. Chiang, Yujia Peng, Hongjing Lu, Keith J. Holyoak, and Martin M. Monti. \"analogical reasoning sequential design fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003094.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[450] Y.C. Leong, J. Chen, R. Willer, and J. Zaki. \"polarization dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003095.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[451] D Chung, MA Orloff, N Lauharatanahirun, B King-Casas, and Chiu PH. \"valuation of peers' safe choices is associated with substance-naivete in adolescents\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps003096.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[452] Anders Lillevik Thorsen. \"abcd protocol for bcbp\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms003098.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[453] HanShin Jo, Chiu-Yueh Chen, Der-Yow Chen, Ming-Heng Weng, and Chun-Chia Kung. \"consensus-seeking and conflict-resolving: an fmri study on college couples' shopping interaction\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003103.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[454] Lauri Parkkonen, Stefan Appelhoff, Alexandre Gramfort, Mainak Jas, and Richard Höchenberger. \"mne-somatodata-bids (anonymous)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003104.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[455] Jia-Xi Wang, Jin-Ying Zhuang, Lulu Fu, and Lei Qin. \"how ovarian hormones influence the behavioral activation and inhibition system through the dopamine pathway\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003114.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[456] Saloni Krishnan and Kate Watkins. \"bold verb generation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003145.v1.0.2." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[457] Arafat Angulo-Perkins and Luis Concha. \"vocal task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003146.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[458] Adam Hampshire and Eyal Soreq. \"neuroimaging evidence for network sampling theory of human intelligence\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003148.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[459] Alexandra Ycaza Herrera, Padideh Nasseri, Ricardo Velasco, Sophia Faude, Jessica D. White, Philipp C. Opitz, Ringo Huang, Kristie Tu, and Mara Mather. \"stress-associated brain activation across the hormonal contraceptive cycle\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds003151.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[460] Chang-Hao Kao, Sangil Lee, Joshua I. Gold, and Joseph W. Kable. \"changepoint fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s003170.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[461] Sivayini Kandeepan, Jorge Rudas, Francisco Gomez, Bobby Stojanoski, Sreeram Valluri, Adrian Mark Owen, Lorina Naci, Andrea Soddu, and Emily Sophia Nichols. \"modeling an auditory stimulated brain under altered states of consciousness using the generalized ising model\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003171.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[462] Simon M. Ceh, Sonja Annerer-Walcher, Karl Koschutnig, Christof Korner, and Mathias Benedek. \"neurophysiological indicators of internal attention: An fmri-eye-tracking coregistration study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003176.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[463] Stefano Moia, Eneko Urñuela, Vicente Ferrer, and César Caballero-Gaudes. \"euskalibur\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003192.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[464] Kenshu Koiso, Sebastian Dresbach, Christopher J Wiggins, Omer Faruk Gulban, Anna Katharina Mueller, Miriam Heynckes, Yuhui Chai, Yoichi Miyawaki, Benedikt A Poser, and Renzo Huber. \"acquisition and processing methods of whole-brain layer-fmri vaso and bold: The kenshu dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003216.v3.0.11." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[465] Meir Meshulam, Liat Hasenfratz, Hanna Hillman, Yun-Fei Liu, Mai Nguyen, Kenneth A. Norman, and Uri Hasson. \"think like an expert\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003233.v1.2.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[466] Livia Tomova, Kimberly L. Wang, Todd Thompson, Gillian A. Matthews, Kay M. Tye, and Rebecca Saxe. \"mri data of 40 adult participants in response to a cue induced craving task following food fasting, social isolation and baseline (within-subject design)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003242.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[467] James Antony, Thomas Hartshorne, Ken Pomeroy, Todd Gureckis, Uri Hasson, Samuel McDougle, and Kenneth Norman. \"behavioral, physiological, and neural signatures of surprise during naturalistic sports viewing\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003338.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[468] Jason A Avery, Alexander G Liu, John E Ingeholm, Stephen J Gotts, and Alex Martin. \"tasting pictures: Viewing images of foods evokes taste-quality-specific activity in gustatory insular cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003340.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[469] Ethan Knights, Courtney Mansfield, Diana Tonin, Janak Saada, Fraser W. Smith, and Stephanie Rossit. \"hand-" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23863" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "selective visual regions represent how to grasp 3d tools for use: brain decoding during real actions\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003342.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[470] Kelsey McDonald, William Broderick, Scott Huettel, and John Pearson. \"penaltykik.02\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003345.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[471] Eduardo A. Garza-Villarreal, Jorge Julio Gonzalez Olvera, Thania Balducci, Diego Angeles Valdez, Alely Valencia, and Jalil Rasgado. \"sudmex Conn: The mexican dataset of cocaine use disorder patients.\" OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003346.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[472] Mikkel Schöttner and Andreas Jansen. \"empatom with ppi-r, tas, and eq\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003354.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[473] Richard Huskey, Justin Robert Keene, Shelby Wilcox, Xuanjun (Jason) Gong, Robyn Adams, and Christina J. Najera. \"brain dynamics during flow experiences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003358.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[474] Samuel B. Snider, Yelena G. Bodien, Aina Frau-Pascual, Marta Bianciardi, Andrea S. Foulkes, and Brian L. Edlow. \"ascending arousal network connectivity during recovery from traumatic coma\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003367.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[475] Heath R Pardoe, Samantha P Martin, Allan George, and Orrin Devinsky. \"estimation of in-scanner head pose changes during structural mri using a convolutional neural network trained on eye tracker video\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003381.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[476] Sharna Jamadar, Shenjun Zhong, Phillip Ward, Alexandra Carey, Richard McIntyre, Alex Fornito, Malin Premaratne, N Jon Shah, Kieran O'Brien, Daniel Stab, Zhaolin Chen, and Gary Egan. \"monash vis-fpet-fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003382.v1.5.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[477] Nicolas Zilber, Philippe Ciuciu, Alexandre Gramfort, Leila Azizi, and Virginie van Wassenhove. \"neurospin hmt+localizer data (meg & amri)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003392.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[478] Sharna D Jamadar, Emma X Liang, Phillip GD Ward, Alexandra Carey, Richard McIntyre, Zhaolin Chen, Shenjun Zhong, and Gary F Egan. \"monash dacra fpet-fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003397.v1.2.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[479] Jungtak Park, Karolina Janacsek, Dezso Nemeth, and Hyeon-Ae Jeon. \"asrt (alternating serial reaction time)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003401.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[480] Brian P Keane, Deanna M Barch, Ravi Mill, Steven M Silverstein, Bart Krekelberg, and Michael W Cole. \"brain network mechanisms of visual shape completion\". Open-Neuro, 2025. doi: 10.18112/openneuro.ds003404.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[481] Maria Geraldine Veldhuizen, Michael C. Farruggia, Xiao Gao, Yuko Nakamura, Barry G. Green, and Dana Small. \"identification of an amygdala-thalamic circuit that acts as a central gain mechanism in taste perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003424.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[482] Steven Greening, Tae-Ho Lee, Laurent Gregoire, Lauryn Burleigh, Tyler Robinson, Xinrui Jiang, Mara Mather, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Jonas Kaplan. \"mental imagery can generate and regulate acquired differential fear conditioned reactivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003425.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[483] Tomoyasu Horikawa and Yukiyasu Kamitani. \"attentionally modulated subjective images reconstructed from brain activity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003430.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[484] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 localizer\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003433.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[485] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 main experiment\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003434.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[486] Sandra Arbula, Elisabetta Pisanu, and Raffaella I. Rumiati. \"agreeableness personality trait and social information encoding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003436.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[487] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain - diet-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003437.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[488] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant1 suboptimal block designs\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms03438.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[489] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant2 suboptimal slow event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003439.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[490] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant3 caffeine\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003440.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[491] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusicac. \"newbi4fmri2020 variant6 adaptation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003443.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[492] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant7 prosopagnosia\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003444.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[493] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant8 resting state\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003445.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[494] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusiac. \"newbi4fmri2020 variant9 7t\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003446.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[495] Joset A. Etzel, Mitch Jeffers, and Todd S. Braver. \"dmcc13benchmark\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003452.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[496] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Lithen," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23864" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, bolus-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003453.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[497] Andrea Stocco, Chantel S. Prat, and Lauren K. Graham. \"individual differences in fluid reasoning and rapm-like problem solving\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003454.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[498] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, case-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003455.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[499] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., and Todd R. Parrish. \"component processes of word reading in adults and children\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003459.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[500] Todd S. Braver, Alexander Kizhner, Rongxiang Tang, Michael C. Freund, and Joset A. Etzel. \"dmcc55b\". Open-Neuro, 2025. doi: 10.18112/openneuro.ms003465.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[501] Jiajia Yang, Peter J. Molfese, Yinghua Yu, Daniel A. Handwerker, Gang Chen, Paul A. Taylor, Yoshimichi Ejima, Jinglong Wu, and Peter A. Bandettini. \"haptic three-dimensional curved surface exploration fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003466.v1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[502] Ekaterina V. Pechenkova, Yana R. Panikratova, Maria A. Fomina, Alena D. Rumshiskaya, Darya A. Bazhenova, Liudmila A. Makovskaya, Irina S. Lebedeva, and Valentin E. Sinitsyn. \"speech disfluencies: Neurophysiological aspect in normal population\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003469.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[503] Anja Ischebeck, Hannah Hiebel, Joe Miller, Margit Höfler, Iain D. Gilchrist, and Christof Körner. \"target processing in overt serial visual search involves the dorsal attention network: A fixation-based event-related fmri study\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003470.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[504] J. Rasgado-Toledo, F. Lizcano-Cortés, V. Olalde-Mathieu, M. Zamora-Ursulo, G. Licea-Haquet, A. Carillo-Pena, E. Navarrete, A.* Reyes-Aguilar, and M Giordano. \"pragmatic language\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003481.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[505] Brian Silston, Toby Wise, Song Qi, Xin Sui, Peter Dayan, and Dean Mobbs. \"foraging in competitive and hazardous environments\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003484.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[506] Eva R. Pool, David Munoz Tord, Sylvain Delplanque, Yoann Stussi, Donato Cereghetti, Patrik Vuilleumier, and David Sander. \"differential contributions of ventral striatum subregions in the motivational and hedonic components of the affective response to reward\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003487.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[507] Lukas Snoek and Steven Scholte. \"dstreamupgrade\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003495.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[508] Kate Nussenbaum and Catherine A. Hartley. \"developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03499.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[509] David Pascucci, Sebastien Tourbier, Joan Rue-Queralt, Margherita Carboni, Patric Hagmann, and Gijs Plomp. \"vepcon: Source imaging of high-density visual evoked potentials with multi-scale brain parcellations and connectomes\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003505.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[510] Chuanji Gao, Christine E. Weber, Douglas H. Wedell, and Svetlana V. Shinkareva. \"fmri: Audiovisual valence congruence\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003507.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[511] Mikael Novén, Hampus Olsson, Gunther Helms, Markus Nilsson, Merle Horne, and Mikael Roll. \"language learning aptitude dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003508.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[512] Adrian W. Gilmore, Alina Quach, Sarah E. Kalinowski, Estefanía I. Gonzalez-Araya, Stephen J Gotts, Daniel L. Schacter, and Alex Martin. \"evidence supporting a timelimited hippocampal role in retrieving autobiographical memories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003511.v1.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[513] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study2\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso03521.v2.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[514] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study1\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03524.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[515] Benjamin Risk, Raphiel Murden, Junjie Wu, Mary Beth Nebel, Arun Venkataraman, Zhengwu Zhang, and Deqiang Qiu. \"which multiband factor should you choose for your resting-state fmri study? the emory multiband dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003540.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[516] Kshipra Gurunandan, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adult language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003542.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[517] Kshipra Gurunandan, Jaione Arnaez-Telleria, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adolescent language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003545.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[518] Isaac David, Victor Olalde-Mathieu, Ana Y. Martínez, Lluviana Rodríguez-Vidal, and Fernando A. Barrios. \"emotion category and face perception task optimized for multivariate pattern analysis\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms03548.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[519] Renee M Visser, H Steven Scholte, Tinka Beemsterboer, and Merel Kindt. \"visser, scholte, beemsterboer, & kindt (2013) nature neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003550.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[520] Renee M Visser, H Steven Scholte, and Merel Kindt. \"visser, scholte & kindt (2011) journal of neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003553.v1.0.2." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23865" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[521] Renee M Visser, Anna E. Kunze, Biance Westhoff, H Steven Scholte, and Merel Kindt. \"visser et al. 2015 psychoneuroendocrinology\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003554.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[522] Falk Luesebrink, Hendrik Mattern, Renat Yakupov, Julio Acosta-Cabronero, Mohammad Ashtarayeh, Steffen Oeltze-Jafra, and Oliver Speck. \"data from: Comprehensive ultrahigh resolution whole brain in vivo mri dataset as a human phantom\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003563.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[523] Lucrezia Liuzzi, Katharine Chang, Hanna Keren, Charles Zheng, Dipta Saha, Dylan Nielson, and Argyris Stringaris. \"mood induction in mdd and healthy adolescents\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds003568.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[524] Jeahong Kim, Jeyoung Jung, and Kichun Nam. \"transposition confusability during visual word recognition\". Open-Neuro, 2021. doi: 10.18112/openneuro.ms003569.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[525] Virginie Sterpenich, Mojca KM van Schie, Maximilien Catsiyannis, Avinash Ramyead, Stephen Perrig, Hee-Deok Yang, Dimitri Van De Ville, and Sophie Schwartz. \"reward biases spontaneous neural reactivation during sleep\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003574.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[526] R. Nathan Spreng, Roni Setton, Udi Alter, Benjamin N. Cassidy, Bri Darboh, Elizabeth DuPre, Karin Kantarovich, Amber W. Lockrow, Laetitia Mwilambwe-Tshilobo, WenMing Luh, Prantik Kundu, and Gary R. Turner. \"neurocognitive aging data release with behavioral, structural, and multi-echo functional mri measures\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso03592.v1.0.13." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[527] Xingyu Liu, Yuxuan Dai, Hailun Xie, and Zonglei Zhen. \"forrestgump-meg\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003633.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[528] Heath R Pardoe and Samantha P Martin. \"in-scanner head motion and structural covariance networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003639.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[529] Michele Svanera and Lars Muckli. \"cerebrum-7t: Fast and fully-volumetric brain segmentation of 7 tesla mr volumes\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003642.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[530] Jixing Li, John Hale, and Christophe Pallier. \"le petit prince: A multilingual fmri corpus using ecological stimuli\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds003643.v2.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[531] Ph.D. Anna Manelis, Ph.D. David A.A. Baranger, B.S. Skye Satz, M.A. Rachel Ragozzino, Ph.D. Satish Iyengar, M.D. Holly A. Swartz, and Ph.D. Yaroslav O. Halchenko. \"cortical myelin measured by the t1w/t2w ratio in individuals with depressive disorders and healthy controls\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003653.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[532] Arielle Tambini, Jacob Miller, Anastasia Kiyonaga, and Mark D'Esposito. \"scantrain\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs003659.v2.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[533] Mohamed Abdelhack and Yukiyasu Kamitani. \"sharpening of hierarchical visual feature representations of blurred images\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003661.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[534] Matan Mazor, Karl J. Friston, and Stephen M. Fleming. \"confidence in detection and discrimination\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003669.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[535] Kangjoo Lee, Corey Horien, David O'Connor, Bronwen Garand-Sheridan, Fuyuze Tokoglu, Dustin Scheinost, Evelyn MR Lake, and Todd R Constable. \"yale resting state fmri/pupillometry: Arousal study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003673.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[536] Giacomo Ariani, J Andrew Pruszynski, and Jorn Diedrichsen. \"finger sequence planning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003684.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[537] Julia Berezutskaya, Mariska J. Vansteensel, Erik J. Aarnoutse, Zachary V. Freudenburg, Giovanni Piantoni, Mariana P. Branco, and Nick F. Ramsey. \"open multimodal ieeg-fmri dataset from naturalistic stimulation with a short audiovisual film\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003688.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[538] Javier Ortiz-Tudela, Johanna Bergman, Matt Bennett, Isabelle Ehrich, Lars Muckli, and Yee Lee Shing. \"feedbes - feedback signals from episodic and semantic memories.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003691.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[539] Shan Xu, Xingyu Liu, Jorge Almeida, and Dietmar Heinke. \"action-related object pairs - fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003696.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[540] Takuya Ito, Emily Winfield, Nicole Lalta, Kaustubh R. Kulkarni, Douglas H. Schultz, Ravi D. Mill, Richard H. Chen, Levi Solomyak, and Michael W. Cole. \"concrete permuted rule operations\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003701.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[541] Guo Wanjia, Serra E. Favila, Ghootae Kim, Robert J. Moltor, and Brice A. Kuhl. \"abrupt hippocampal remapping signals resolution of memory interference.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003707.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[542] Hanna Keren, Katharine Chang, Aria Vitale, Chris C. Camp, Robb B Rutledge, Dylan M. Nielson, and Argyris Stringaris. \"nimh-compypsych mmi\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03709.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[543] Daniel K. Bjornn, Julie Van, and C. Brock Kirwan. \"the contributionsof eye gaze fixations and target-lure similarity to behavioral and fmri indices of pattern separation and pattern completion\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003711.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[544] Avijit Chowdhury. \"test\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003714.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[545] Li J., Curley W. H., Guerin B., Dougherty D. D., Dalca A. V., Fischl B., Horn A., and Edlow B. L. \"subcortical dmn functional connectivity\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003716.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[546] Peelle JE, Spehar B, Jones MS, McConkey S, Myerson J, Hale S, Sommers MS, and Tye-Murray N. \"visual and audiovisual speech perception associated with increased functional connectivity between sensory and motor regions\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003717.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[547] Tomoya Nakai, Naoko Koide-Majima, and Shinji Nishimoto. \"music genre fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003720.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23866" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[548] Renee M Visser, Richard N Henson, and Emily A Holmes. \"a naturalistic paradigm to investigate post-encoding neural activation patterns in relation to subsequent voluntary and intrusive recall of distressing events\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003721.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[549] David V. Smith, Rita M. Ludwig, Jeffrey B. Dennison, Crystal Reeck, and Dominic S. Fareri. \"an fmri dataset on social reward processing and decision making in younger and older adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003745.v2.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[550] michael barnett, geoffrey aguirre, and david brainard. \"If contrast\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003752.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[551] Kenji Kobayashi, Sangil Lee, Alexandre L. S. Filipowicz, Kara D. McGaughey, Joseph W. Kable, and Matthew R. Nassar. \"dynamic representation of the subjective value of information\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003758.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[552] Olga R. Dobrushina, Larisa A. Dobrynina, Galina A. Arina, Maria V. Gubanova, Anastasia V. Belopasova, Evgenia S. Novikova, Anastasia D. Suslina, Elena I. Kremneva, Ekaterina V. Pechenkova, and Marina V. Krotenkova. \"interception during aging: The heartbeat detection task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003763.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.501 + ], + "angle": 0, + "content": "[553] Zhenzhong Gan, Suiping Wang, and Gangyi Feng. \"functional magnetic resonance imaging data for the neural dynamics underlying the acquisition of distinct auditory categories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003764.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[554] Yameng Gu, Feng Han, Lucas E. Sainburg, Margeaux M. Schade, and Xiao Liu. \"simultaneous eeg and fmri signals during sleep from humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003768.v1.0.11." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[555] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"depression treatment by medial prefrontal real-time fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms003770.v1.2.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[556] Joseph T. McGuire, Matthew R. Nassar, Joshua I. Gold, and Joseph W. Kable. \"functionally dissociable influences on learning rate in a dynamic environment\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003772.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[557] Redouane Jamil, Franck Mauconduit, Caroline Le Ster, Philipp Ehses, Benedikt A. Poser, Alexandre Vignaud, and Nicolas Boulant. \"temporal snr optimization through rf coil combination in fmri: The more, the better? - dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003777.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[558] Yufei Zhao, Avi Chanales, and Brice Kuhl. \"adaptive memory distortions are predicted by feature representations in parietal cortex\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003778.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[559] Nadav Aridan, Nicholas J. Malecek, Russell A. Poldrack, and Tom Schonberg. \"neural correlates of effort-based valuation under risk\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003782.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.903 + ], + "angle": 0, + "content": "[560] Marc Himmelberg, Jan Kurzawski, Noah Benson, Denis Pelli, Marisa Carrasco, and Jon Winawer. \"nyu retinotopy" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003787.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[561] Xuhao Shao, Chuansheng Chen, Elizabeth F. Loftus, Gui Xue, and Bi Zhu. \"dynamic changes in neural representations underlie the repetition effect on false memory\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003789.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[562] Eva Mishor, Daniel Amir, Tali Weiss, Danielle Honigstein, Aharon Weissbrod, Ethan Livne, Lior Gorodisky, Shiri Karagach, Aharon Ravia, Kobi Snitz, Diyala Karawani, Rotem Zirler, Reut Weissgross, Timna Soroka, Yaara Endevelt-Shapira, Shani Agron, Liron Rozenkrantz, Netta Reshef, Edna Furman-Haran, Heinz Breer, Joerg Strotmann, Tatsuya Uebi, Mamiko Ozaki, and Noam Sobel. \"sniffing the human body-volatile hexadecanal blocks aggression in men but triggers aggression in women\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003791.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[563] Dorit Kliemann, Ralph Adolphs, Tim Armstrong, Paola Galdi, David A. Kahn, Tessa Rusch, A. Zeynep Enkavi, Deuhua Liang, Steven Lograsso, Wenying Zhu, Rona Yu, Remya Nair*, Lynn K. Paul*, J. Michael Tyszka*, and * Joint senior authorship. \"caltech conte center - a multimodal data resource for exploring social cognition and decision-making\". OpenNeuro, 2021. doi: 10.18112/openneuro.dss003798.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.57 + ], + "angle": 0, + "content": "[564] Hyun Joo Yoo, Kaoru Nashiro, Jungwon Min, Christine Cho, Noah Mercer, Shelby L. Bachman, Padideh Nasseri, Shubir Dutt, Shai Porat, Paul Choi, Yong Zhang, Vardui Grigoryan, Tiantian Feng, Julian F. Thayer, Paul Lehrer, Catie Chang, Jeffrey A. Stanley, Elizabeth Head, Jeremy Rouanet, Vasilis Z. Marmarelis, Shrikanth Narayanan, Jessica Wisnowski, Daniel A. Nation, and Mara Mather. \"heart rate variability biofeedback training and emotion regulation\". OpenNeuro, 2025. doi: 10.18112/openneuro.d003823.v1.3.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[565] Michal Rafal Zareba, Magdalena Fafrowicz, Tadeusz Marek, Ewa Beldzik, Halszka Oginska, and Aleksandra Domagalik. \"structural (t1) images of 136 young healthy adults; study of effects of chronotype, sleep quality and daytime sleepiness on brain structure\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003826.v3.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[566] Kevin Fialkowski, Ivan Messias, and Keith A. Bush. \"cognitive control theoretic mechanisms of real-time fmriguided neuromodulation (ctm)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003831.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[567] Matteo Visconti di Oleggio Castello, James V. Haxby, and M. Ida Gobbini. \"shared neural codes for visual and semantic information about familiar faces in a common representational space\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003834.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[568] Alyssa H Sinclair and Morgan D Barense. \"pe-update\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003835.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[569] Flavia Mancini, Suyi Zhang, and Ben Seymour. \"learning the statistics of pain\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs003836.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[570] van Blooijs D., Demuru M., Zweiphenning W, Hermes D., Leijten F., and Zijlmans M. \"dataset clinical epilepsy ieeg" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23867" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "to bids - respect_longterm_jeeg\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003848.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[571] Austin L. Boroshok, Anne T. Park, Panagiotis Fotiadis, Gerardo H. Velasquez, Ursula A. Tooley, Katrina R. Simon, Jasmine C.P. Forde, Lourdes M. Delgado Reyes, M. Dylan Tisdall, Dani S. Bassett, Emily A. Cooper, and Allyson P. Mackey. \"individual differences in frontoparietal plasticity in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003849.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[572] S.R. Ashby and D. Zeithamova. \"category-biased neural representations form spontaneously during learning that emphasizes memory for specific instances\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003851.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[573] Leili Mortazavi, Tara Srirangarajan, Tiago Bortolini, Jorge Moll, and Brian Knutson. \"monetary incentive delay task\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003858.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[574] Christopher N. Wahlheim, Alexander P. Christensen, Brittany S. Cassidy, and Zachariah M. Reagh. \"resting-state for 34 younger and 28 older adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003871.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[575] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003872.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[576] Ted Turesky, Talat Shama, Shahriah Hafiz Kakon, Rashidul Haque, Nazrul Islam, Amala Someshwar, Borjan Gagoski, William Petri, Charles Nelson, and Nadine Gaab. \"brain-morphometry diminishedgrowthBeanstudy_2021\". OpenNeuro, 2022. doi: 10.18112/openneuro.d003877.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[577] Wen Wen, Yue Wang, jiawei Zhou, Sheng He, Xinghuai Sun, Hong Liu, Chen Zhao, and Peng Zhang. \"loss and enhancement of layer-selective signals in geniculostriate and corticotectal pathways of adult human amblyopia\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003892.v5.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[578] Philippe Poulin, Guillaume Thead, Pierre-Marc Jodoin, and Maxime Descoteaux. \"tractoinferno: A large-scale, open-source, multi-site database for machine learning dmri tractography\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003900.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[579] J. Pesnot Lerousseau, C. Parise, MO. Ernst, and V. van Wassenhove. \"multisensory correlation detector\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003922.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[580] Ning Mei, Roberto Santana, and David Soto. \"informative neural representations of unseen contents during higher-order processing in human brains and deep artificial networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003927.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[581] Tommaso Di Noto, Guillaume Marie, Sebastien Tourbier, Yasser Aleman-Gómez, Oscar Esteban, Guillaume Saliou, Meritxell Bach Cuadra, Patric Hagmann, and Jonas Richiardi. \"lausanne_tof-mra_aneurysm_cohort\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003949.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[582] Lukas Snoek, Jessica Loke, Michelle Kuehn, Mona Zimmermann, and Steven Scholte. \"ni-edu-data-minimal\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003965.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[583] Olivier Girard, Ludovic de Rochefort, Maxime Guye, Jean-Philippe Ranjeva, and Arnaud Le Troter. \"7tamibrain\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003967.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[584] W. Schellekens, C. Bakker, N.F. Ramsey, and N. Petridou. \"body motor\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003972.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[585] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 11 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03974.v3.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[586] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03988.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[587] Sarah Khalife, Susan T. Francis, Denis Schluppeck, Rosa Maria Sanchez-Panchuelo, and Julien Besle. \"fastvsslowertactile7t\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003990.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[588] Jun Kai Ho, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"inter-individual deep image reconstruction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003993.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[589] PhD) Daminov V. (MD, MSc) Novak E. (MD, Slepyova N. (MD), Mikhailov D. (MSc), and Karpulevich E. (MSc). \"pre-post rehabilitation fmri data of post-stroke patients\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003999.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[590] J. Wammes, K.A. Norman, and N.B. Turk-Browne. \"Increasing stimulus similarity drives nonmonotonic representational change in hippocampus\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s004006.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[591] J.A. Williams, E.H. Margulis, S.A. Nastase, J. Chen, U. Hasson, K.A. Norman, and C. Baldassano. \"high-order areas and auditory cortex both represent the high-level event structure of music\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004007.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[592] Robert M. Mok and Bradley C. Love. \"abstract-category-signal\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004009.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[593] Dominic Standage, Joseph Y. Nashed, J. Randall Flanagan, and Jason. P Gallivan. \"visuomotor rotation adaptation experiment\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004021.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[594] Julio Cesar Hernandez Pavon, Nils Schneider Garces, John Patrick Begnoche, Lee Miller, and Tommi Rajj. \"tmseeg-mri-fmri-dwi data on paired associative stimulation and connectivity (shirley ryan abilitylab, chicago, il)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004024.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[595] Cammille C Go, Huseyin O Taskin, Seyed Ahmad Ahmadi, Giulia Frazzetta, Laura Cutler, Saguna Malhotra, Jessica IW Morgan, Virginia L Flanagin, and Geoffrey K Aguirre. \"tome-go2022vestibularmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04038.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[596] Hongmi Lee, Janice Chen, and Uri Hasson. \"filmfestival\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004042.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23868" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[597] Sai Ma, Taicheng Huang, Yukun Qu, Xiayu Chen, Yajie Zhang, and Zonglei Zhen. \"an fmri dataset for whole-body somatotopic mapping in humans\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004044.v2.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[598] Miao Zhang, Qikai Qin, Shuning Zhang, Wei Liu, Hongping Meng, Mengyang Xu, Xinyun Huang, Xiaozhu Lin, Mu Lin, Peter Herman, Fahmeed Hyder, Raymond C. Stevens, Zheng Wang, Biao Li, and Garth J. Thompson. \"aerobic glycolysis imaging pet-mri\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004054.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[599] Yoshihisa Nakayama, Sho K. Sugawara, Masaki Fukunaga, Yuki H. Hamano, Norihiro Sadato, and Yukio Nishimura. \"goal-directed motor task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004056.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[600] Shai Sabbah, Michael S. Worden, Dimitrios D. Laniado, David M. Berson, and Jerome N. Sanes. \"luminance\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004065.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[601] Matan Mazor, Chudi gong, and Stephen M. Fleming. \"confidence in equal and unequal variance settings\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds004081.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[602] Julie Van, Samuel Nielson, and C. Brock Kirwan. \"triple dissociation revisited\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004086.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[603] Martin Szinte, Gilles de Hollander, Marco Aqil, Serge Dumoulin, and Tomas Knapen. \"gaze prf\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004091.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[604] Athula Pudhiyidath, Neal W Morton, Rodrigo Viveros Duran, Anna C. Schapiro, Ida Momennejad, Demitrius M. Hinojosa-Rowland, Robert J. Molitor, and Alison R. Preston. \"effects of temporal community structure learning on reasoning decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004094.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[605] Kristina M. Rapuano, May I. Conley, Anthony C. Juliano, Gregory M. Conan, Maria T. Maza, Kylie Woodman, Steven A. Martinez, Eric Earl, Anders Perrone, Eric Feczko, Damien A. Fair, Richard Watts, BJ Casey, and Monica D. Rosenberg. \"an open-access accelerated adult equivalent of the abcd study neuroimaging dataset (a-abcd)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004097.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[606] Dobrushina Olga R, Dobrunina Larisa A, Kremneva Elena I, and Pechenkova Ekaterina V. \"resting-state fmri data acquired before and after a course of infra-low frequency neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004101.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[607] Damien Brevers, Chris Baeken, Stefanie De Smet, Beatrix Catoira, Sara De Witte, Qingha He, Pierre Maurage, Laimi Schultze, Guillaume Sescousse, Claudia Vila Verde, Claus Vögele, and Joel Billieux. \"stimulation of the dorsolateral prefrontal cortex modulates brain cue reactivity to reward(un)availability\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004102.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[608] Siao-Shan Shen, Jen-Tang Cheng, I-Jeng Hsu, Der-Yow Chen, Ming-Hung Weng, and Chun-Chia Kung. \"collaborations and deceptions in strategic interactions revealed by hyperscanning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004103.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[609] M.P. Weisend, F.M. Hanlon, R. Montano, S.P. Ahlfors, A.C. Leuthold, D. Pantazis, J.C. Mosher, A.P. Georgopoulos, M.S. Hamalainen, and C.J. Aine. \"mind data\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004107.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[610] Yannan Zhu and Shaozheng Qin. \"emotional learning retroactively promotes memory integration through rapid neural reactivation and reorganization\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004109.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[611] Kenji Kobayashi, Joseph W. Kable, Ming Hsu, and Adrianna C. Jenkins. \"neural representations of others' traits predict social decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004128.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[612] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype segregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004129.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[613] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype aggregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004130.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[614] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype external example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004131.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[615] Sepehr Mortaheb, Laurens Van Calster, Federico Raimondo, Manousos A. Klados, Paradeisios Alexandros Boulakis, Kleio Georgoula, Steve Majerus, Dimitri Van De Ville, and Athena Demertzi. \"experience sampling in resting state\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004134.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[616] Bruno Direito, Marco Simões, Alexandre Sayal, and Miguel Castelo-Branco. \"targeting dynamic facial processing mechanisms in superior temporal sulcus using fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004141.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[617] Bruno Direito, Manuel Ramos, Alexandre Sayal, Joao Pereira, and Miguel Castelo-Branco. \"exploring the neural correlates of feedback-related reward saliency and valence during real-time fmri-based neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004142.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[618] Thania Balducci, Jalil Rascado-Toledo, Alely Valencia, Maria-Jose van Tol, Andre Aleman, and Eduardo A. Garza-Villarreal. \"a behavioral, clinical and brain imaging dataset with focus on emotion regulation of females with fibromyalgia\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds004144.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[619] Martin Szinte, Anna Montagnini, and Guillaume Masson. \"rest_eye\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004158.v2.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[620] Lachlan T. Strike, Gabriella A.M. Blokland, Narelle K. Hansell, Nicholas G. Martin, Arthur W. Toga, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin imaging (qtim)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004169.v1.0.7." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23869" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[621] Ádám Nárai, Petra Hermann, Tibor Auer, Péter Kemenczky, János Szalma, István Homolya, Eszter Somogyi, Pál Vakli, Béla Weiss, and Zoltán Vidnyanszky. \"movement-related artefacts (mr-art) dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004173.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[622] Stefanie Meliss, Cristina Pascua-Martin, Jeremy Skipper, and Kou Murayama. \"magic, memory, and curiosity (mmc) fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004182.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[623] Irem Yildirim, Khan Hekmatyar, and Keith A. Schneider. \"lgn layers data\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004187.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[624] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004192.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[625] Iris Groen, Kenichi Yuasa, Amber Brands, Giovanni Piantoni, Stephanie Montenegro, Adeen Flinker, Sasha Devore, Orrin Devinsky, Werner Doyle, Patricia Dugan, Daniel Friedman, Nick Ramsey, Natalia Petridou, and Jonathan Winawer. \"visual ecog dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004194.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[626] Foteini Liwicki, Vibha Gupta, Rajkumar Saini, Kanjar De, Nosheen Abid, Sumit Rakesh, Scott Wellington, Holly Wilson, Marcus Liwicki, and Johan Eriksson. \"bimodal dataset on inner speech\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004196.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[627] Fabiane Schuch, Lennart Walger, Matthias Schmitz, Bastian David, Tobias Bauer, Antonia Harms, Laura Fischbach, Freya Schulte, Martin Schidlowski, Johannes Reiter, Felix Bitzer, Randi von Wrede, Attila Racz, Tobias Baumgartner, Valeri Borger, Matthias Schneider, Achim Flender, Albert Becker, Hartmut Vatter, Bernd Weber, Louisa Specht-Riemenschneider, Alexander Radbruch, Rainer Surges, and Theodor Rüber. \"an open presurgery mri dataset of people with epilepsy and focal cortical dysplasia type ii\". OpenNeuro, 2025. doi: 10.18112/openneuro.d004199.v1.0.6." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[628] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-meg\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004212.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[629] Hermes D and Wandell BA. \"brainbeats\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004213.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[630] Mark A. Thornton, Elyssa Barrick, Nathan Liang, and Diana I. Tamir. \"interacting representations of mental states and traits\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004217.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[631] Kyoko Ohashi, Martin H. Teicher, Elizabeth Bolger, Laura C. Hernandez Garcia, Poopak Hafezi, Leslie P. Weiser, Cynthia E. McGreenery, and Alaptagin Khan. \"lite-book_Alertness_study\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004219.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[632] Mark A. Thornton and Diana I. Tamir.\"summed social cognition: additive neural representations of situations, mental states, and actions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004226.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[633] Anne T. Park, Hilary Richardson, Ursula A. Tooley, Cassidy L. McDermott, Austin L. Boroshok, Adrian Ke, Julia A. Leonard, M. Dylan Tisdall, Kirby Deater-Deckard, J. Christopher Edgar, and Allyson P. Mackey. \"early stressful experiences are associated with reduced neural responses to naturalistic emotional and social content in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004228.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[634] Min-Jeong Kim, Jae-Hoon Lee, Fernanda Juarez Anaya, Jinsoo Hong, William Miller, Sanjay Telu, Prachi Singh, Michelle Y Cortes, Katharine Henry, George L Tye, Michael P Frankland, Jose A Montero Santamaria, Jeih-San Liow, Sami S Zoghbi, Masahiro Fujita, Victor W Pike, and Robert B Innis. \"first-in-human evaluation of [11c]ps13, a novel pet radioligand, to quantify cyclooxygenase-1 in the brain\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004230.v3.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[635] Miguel B. Garcia, Gilles de Hollander, Marcus Grueschow, Rafael Polania, Michael Woodford, and Christian C. Ruff. \"individual risk attitudes arise from noise in neurocognitive magnitude representations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004259.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[636] Xuhao Shao, Ao Li, Chuangsheng Chen, Elizabeth F. Loftus, and Bi Zhu. \"cross-stage neural pattern similarity in the hippocampus predicts false memory derived from post-event inaccurate information\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004261.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[637] Thirza Dado, Yaghmur Güçlütürk, Luca Ambrogioni, Gabriëlle Ras, Sander Bosch, Marcel van Gerven, and Umut Güçlü. \"hyper\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04280.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[638] Krista Bond, Javier Rasero, Raghav Madan, Jyotika Bahuguna, Jonathan Rubin, and Timothy Verstynen. \"neuroloki\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004283.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[639] Chad S. Rogers, Michael S. Jones, Sarah McConkey, and Jonathan E. Peelle. \"listening task\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs004285.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[640] Wilma A. Bainbridge and Chris I. Baker. \"multidimensional memory topography in the medial parietal cortex identified from neuroimaging of thousands of daily memory videos\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004286.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[641] Shaonan Wang, Yunhao Zhang, Xiaohan Zhang, Jingyuan Sun, Nan Lin, Jiajun Zhang, Chengqing Zong, and Shaonan Wang. \"an fmri dataset for concept representation with semantic feature annotations\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004301.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[642] Joan Soler-Vidal, Paola Fuentes-Claramonte, Pilar Salgado-Pineda, Nuria Ramiro, María Ángeles García-León, María Llanos Torres, Antonio Arevalo, Amalia Guerrero-Pedraza, Josep Munuera, Salvador Sarro, Raymond Salvador, Wolfram Hinzen, Peter McKenna, and Edith Pomarol-Clotet. \"brain correlates of speech perception in schizophrenia patients with and without auditory hallucinations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004302.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23870" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[643] Jason A Avery, Madeline Carrington, and Alex Martin. \"a common neural code for representing imagined and inferred tastes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004312.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[644] Momchil S. Tomov, Pedro A. Tsividis, Thomas Pouncy, Joshua B. Tenenbaum, and Samuel J. Gershman. \"atari-style video game learning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004323.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[645] Emily E. Perszyk, Jessica Trinh, Xue S. Davis, Zach Hutelin, Leonie Koban, Maria G. Veldhuizen, Jelena Djordjevic, Marilyn Jones-Gotman, Tor Wager, Hedy Kober, and Dana M. Small. \"odour-imagery ability is linked to food craving, intake, and adiposity change in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004327.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[646] Johannes J.D. Singer, Radoslaw M. Cichy, and Martin N. Hebart. \"the spatiotemporal neural dynamics of object recognition for natural images and line drawings (fmri)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004331.v1. 0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[647] Melanie Ganz and Hannah Eichhorn. \"datasets with and without deliberate head movements for evaluating the performance of markerless prospective motion correction and selective reacquisition in a general clinical protocol for brain mri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004332.v1.3.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[648] Loris Naspi, Paul Hoffman, and Alexa Morcom. \"semantic_encoding\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004341.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[649] Oscar Ferrante, Ling Liu, Tamas Minarik, Urszula Gorska, Tara Ghafari, Huan Luo, and Ole Jensen. \"flux: A pipeline for meg analysis\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004346.v1.0.8." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[650] L.E. Frank and D. Zeithamova. \"evaluating methods for measuring background connectivity in slow event-related fmri designs\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004349.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[651] Asieh Zadbood, Samuel A. Nastase, Janice Chen, Kenneth A. Norman, and Uri Hasson. \"sixthsense\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004359.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[652] Korey P. Wylie, Benzi M. Kluger, Luis D. Medina, Samantha K. Holden, Eugene Kronberg, Jason R. Tregellas, and Isabelle Guard. \"parkinson's disease, functional connectivity, and cognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004392.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[653] Lauryn Burleigh and Steven Greening. \"fear in the mind's eye: the neural correlates of differential fear acquisition to imagined conditioned stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004393.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[654] Doug Greve, Paul Wighton, Melanie Ganz, Martin Nørgaard, Paolo Zanotti-Fregonara, Cyril Pernet, Anthony Galassi, Adam Thomas, Nafiseh Ghazanfari, Jeih-San Liow, Granville Matheson, Gitte Knudsen, and Robert Innis. \"a pet molecular imaging brain atlas of cyclooxygenase-1 (kim 2021)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004401.v1.3.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[655] Corey Fernandez, Jiefeng Jiang, Shao-Fang Wang, Hannah L. Choi, and Anthony D. Wagner. \"rid\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004406.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[656] Marc Himmelberg, Ekin Tuncok, Jesse Gomez, Kalanit Grill-Spector, Marisa Carrasco, and Jonathan Winawer. \"stanford child and adult checkerboard retinotopy dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004440.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[657] Puti Wen, Michael S. Landy, and Bas Rokers. \"identifying cortical areas that underlie the transformation from 2d retinal to 3d head-centric motion signals\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004443.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[658] Will be updated. \"xcp_walkthrough\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004450.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[659] Carolina Feher da Silva, Gaia Lombardi, Micah Edelson, and Todd A. Hare. \"magic_carpet\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004455.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[660] Agata Wolna and Zofia Wodniecka. \"bilingual speech production with functional localizers\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004456.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[661] Renzo Huber. \"highresolution development\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004458.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[662] Chris Racey, Charlotte Rae, Ivor Simpson, Oscar Bowen-Hill, Mathilde Bauer, Julia Simner, Christina Kampoureli, and Jamie Ward. \"mapping the connectome of synaesthesia: An open access mri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004466.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[663] Halie Olson, Emily Chen, Kirsten Lydic, and Rebecca Saxe. \"mri data from 20 adults in response to videos of dialogue and monologue from sesame street\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004467.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[664] Raul Rodríguez-Cruces, Vicente Camacho-Téllez, Alfonso Fajardo, and Luis Concha. \"temporal lobe epilepsy - unam\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004469.v1.1.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[665] Jonathan C. Lau, Yiming Xiao, Roy A. M. Haast, Greydon Gilmore, Kamil Uludag, Keith W. MacDougall, Ravi S. Menon, Andrew G. Parrent, Terry M. Peters, and Ali R. Khan. \"stereotactic neurosurgery dataset (snsx)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004470.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[666] Mohamad Abbass, Greydon Gilmore, Alaa Taha, Ryan Chevalier, Magdalena Jach, Terry M. Peters, Ali R. Khan, and Jonathan C. Lau. \"London heath sciences center parkinson's disease dataset (lhscpd)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004471.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[667] Alexander P. Rockhill, Alessandra Mantovani, Brittany Stedelin, Admed M. Raslan, and Nicole C. Swann. \"seeg forced two-choice task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004473.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[668] Noelle A. Jacobsen and Daniel P. Ferris. \"mobile eeg splitbelt walking study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004475.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[669] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Nicole Tacugue, Zenia Valdiviezo, Laura D. Lewis, and Zachary Diamandis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004478.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[670] Samuel Planton*, Fosca Al Roumi*, Liping Wang, and Stanislas Dehaene. \"abseqfmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso004482.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23871" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[671] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004484.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[672] Ming Zhou, Zhengxin Gong, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for human action recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004488.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[673] Zhengxin Gong, Ming Zhou, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for the visual processing of naturalistic scenes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004496.v2.1.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[674] Ph.D. Evan Gordon and Ph.D. Timothy Laumann. \"perinatal stroke\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004498.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[675] Mikl Michal, Kovarova Anezka, Gajdos Martin, Marecek Radek, Novakova Marie, and Slavicek Tomas. \"multi-echo simultaneous multislice fmri dataset: Effect of acquisition parameters on fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps004499.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[676] Amanda Studnicki and Daniel P. Ferris. \"real world table tennis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004505.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[677] Gabriel Castrillon, Samira Epp, Antonia Bose, Laura Fraticelli, André Hechler, Roman Belenya, Andreas Ranft, Igor Yakushev, Lukas Utz, Lalith Sundar, Josef P Rauschecker, Christine Preibisch, Katarzyna Kurcyus, and Valentin Riedl. \"the energetic costs of the human connectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004513.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[678] Clara Sava-Segal, Chandler Richards, Megan Leung, Daniel A. Handwerker, Javier Gonzalez-Castillo, Peter J. Molfese, Peter A Bandettini, and Emily S. Finn. \"avid: Audio-visual individual differences dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004516.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[679] Nestor Zaragoza-Jimenez, Hauke Niehaus, Ina Thome, Christoph Vogelbacher, Gabriele Ende, Inge Kamp-Becker, Dominik Endres, and Andreas Jansen. \"modeling face recognition in the predictive coding framework: A combined computational modeling and functional imaging study - stage 2\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004529.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[680] Elizabeth Valles-Capetillo, Magda Giordano, and Cristian Ibarra. \"ironia vev\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004533.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[681] Sebastian Dresbach, Renzo Huber, Omer Faruk Gulban, and Rainer Goebel. \"event-related vaso\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004539.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[682] Emalie McMahon, Michael F. Bonner, and Leyla Isik. \"social interaction dyads\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004542.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[683] Paul Black, C. Brock Kirwan, Thomas Meservy, William Tayler, and Jeff Williams. \"an fmri investigation of the neurocognitive processing of strategies and measures\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004544.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[684] Michael Rojek-Giffin, Mael Lebreton, Andrea Farina, Jorg Gross, Jean Daunizeau, and Carsten de Dreu. \"learning rules of engagement for social exchange within and between groups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004553.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[685] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 1\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004556.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[686] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 2\". OpenNeuro, 2023. doi: 10.18112/openneuro. ds004557.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[687] M. Rebsamen, D. Romascano, M. Capiglioni, R. Wiest, P. Radojewski, and C. Rummel. \"the phantom of bern: repeated scans of two volunteers with eight different combinations of mr sequence parameters\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004560.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[688] Youngjo Song, Wooree Shin, Pyeongsoo Kim, and Jaeseung Jeong. \"mvpa for motor adaptation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004562.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[689] Paul Broca and Carl Wernicke. \"demo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004564.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[690] Ludmila Midrigan-Ciochina, Kayla P. Vodacek, Srinija Balabhadra, and David P. Corina. \"brain differences in monolingual and highly proficient multilingual speakers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004581.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[691] Yingying Wang, Hongmi Lee, and Brice Kuhl. \"ds004590\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004590.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[692] Hayoung Song, Won Mok Shim, and Monica D. Rosenberg. \"song dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004592.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[693] Marc N. Coutanche, Jake Sauter, Essang Akpan, Rae Buckser, August Vincent, and M. Kathleen Caulfield. \"brain surface fingerprints\". OpenNeuro, 2023. doi: 10. 18112/openneuro.ds004594.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[694] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, Bruno Nazarian, Julien Sein, and Jean-Luc Anton. \"gloups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004597.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[695] Antonin Rovai, Valentina Lolli, Nicola Trotta, Serge Goldman, and Xavier De Tiège. \"cerebrovascular reactivity normative dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004604.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[696] Cameron C. McKay, Brooke Scheinberg, Ellie P. Xu, Katharina Kircanski, Daniel S. Pine, Melissa A. Brotman, Ellen Leibenluft, and Julia O. Linke. \"emotion and development branch phenotyping and dti (2012-2017)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004605.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[697] David Romascano, Gian Franco Piredda, Samuele Caneschi, Tom Hilbert, Ricardo Corredor, Bénédicte Maréchal, Tobias Kober, Jean-Baptiste Ledoux, Eleonora" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23872" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Fornari, Patric Hagmann, and Solange Denervaud. \"developmental relaxometry 2023\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004611.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[698] Dylan M. Nielson, Neda Sadeghi, Lisa S. Gorham, Lillian Eisner, Jeremy Taigman, Katherine Haynes, Karen Qi, Christopher C. Camp, Payton Fors, Diana Rodriguez, Jerry McGuire, Erin Garth, Chana Engel, Mollie Davis, Kenneth Towbin, and Argyris Stringaris. \"characterization and treatment of adolescent depression (cat-d)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004627.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[699] Mengfan Zhang, Alexander Karner, Kathrin Kostorz, Sophia Shea, David Steyrl, Filip Melinscak, Ronald Sladky, Cindy Sumaly Lor, and Frank Scharnowski. \"spida-mrineuroimaging dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004630.v1.1.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[700] A. De Soares, T. Kim, F. Mugisho, E. Zhu, A. Lin, C. Zheng, and C. Baldassano. \"top-down attention impacts neural dynamics during narratives with overlapping event scripts\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004631.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[701] P. G. Bissett, I. W. Eisenberg, S. Shim, J. A. H. Rios, H. M. Jones, M. P. Hagen, A. Z. Enkavi, J. K. Li, J. A. Mumford, D. P. MacKinnon, L. A. Marsch, and R. A. Poldrack. \"cognitive tasks, anatomical mri, and functional mri data evaluating the construct of self-regulation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004636.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[702] Edlow B. L., Olchanyi M. D., Freeman H. J., Li J., Maffei C., Snider S. B., Zollei L., Eugenio Iglesias J., Augustinack J., Bodien Y. G., Haynes R. L, Greve D. N., Diamond B. R., Stevens A., Giacino J. T., Destrieux C., van der Kouwe A., Brown E. N., Folkerth R. D., Fischl B., and Kinney H. C. \"sustaining wakefulness: Brainstem connectivity in human consciousness\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004640.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.483, + 0.631 + ], + "angle": 0, + "content": "[703] Sydney M. Bailes, Daniel E.P. Gomez, Beverly Setzer, and Laura D. Lewis. \"resting-state fmri signals contain spectral signatures of local hemodynamic response timing\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004645.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[704] Heesun Park, Hoyoung Doh, Eunhwi Lee, Harhim Park, and Woo-Young Ahn. \"the neurocognitive role of working memory load when pavlovian motivational control affects instrumental learning\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004647.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[705] Ana L. Miranda-Angulo, Juan D. Sánchez-lópez, Daniel A. Vargas-Tejada, Juan C. Calderón, Jaime Gallo-Villegas, Juan F. Alzate-Restrepo, Jazmin X. Suarez-Revelo, and Gabriel Castrillón. \"sympathovagal quotient and functional connectivity of control networks are related to gut ruminococcaceae abundance in healthy men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004648.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[706] Martin Norgaard. \"the mn dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004650.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[707] Kathleen C Schmidt, Inna Loutaev, Thomas V Burlin, Zenaide Quezado, Audrey Thurm, Carrie Sheeler, and Carolyn Beebe Smith. \"rates of cerebral protein synthesis in subjects with fragile x syndrome\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004654.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[708] Travis D. Masterson, C. Brock Kirwan, Lance E. Davidson, and James D. LeCheminant. \"neural reactivity to visual food stimuli is reduced in some areas of the brain during evening hours compared to morning hours: an fmri study in women\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004656.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[709] Neha A Reddy, Kristina M Zvolanek, and Molly G Bright. \"denoising task-correlated head motion from motor-task fmri data with multi-echo ica\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004662.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[710] Fan Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C. Aoki, Jin Hirano, and Yukiyasu Kamitani. \"visual illusion reconstruction\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004670.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[711] Evan Layher, Tyler Santander, Puneeth Chakravarthula, Nicole Marinsek, Benjamin O. Turner, Miguel P. Eckstein, and Michael B. Miller. \"widespread frontoparietal fmri activity is greatly affected by changes in criterion placement, not discriminability, during recognition memory and visual detection tests\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004692.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[712] Johannes Singer, Agnessa Karapetian, Martin Hebart, and Radoslaw Cichy. \"identifying and characterizing scene representations relevant for categorization behavior\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004693.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[713] Bari Fuchs, Alaina Pearce, Barbara Rolls, Stephen Wilson, Emma Jane Rose, Charles Geier, Hugh Garavan, and Kathleen Keller. \"food and brain study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004697.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[714] Kelly Chang, Ione Fine, and Geoffrey M. Boynton. \"chn retinotopic mapping dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004698.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[715] David A Feinberg, Alexander JS Beckett, and An T VU. \"next generation mri scanner designed for ultra-high resolution human brain imaging at 7-tesla\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004710.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[716] Loreen Tisdall and Rui Mata. \"agerisk\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004711.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[717] S. Warrington, A. Torchi, O. Mougin, A. Ntata, J. Campbell, M. Craig, F. Alfaro-Almagro, K.L. Miller, P.S. Morgan, M. Jenkinson, and S.N. Sotiropoulos. \"on-harmony: A resource for development and comparison of multi-modal brain 3t mri harmonisation approaches\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004712.v2.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[718] Elisa Zamboni, Alexis Makin, Marco Bertamini, and Antony Morland. \"symmetry & luminance bold fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004715.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[719] Elena Filimonova, Anton Pashkov, Norayr Borisov, Anton Kalinovsky, and Jamil Rzaev. \"utilizing amide proton transfer technique to characterise diffuse gliomas based on who 2021 classification of cns tumors\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004717.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[720] Mohammad Momenian, Zhengwu Ma, Shuyi Wu, Chengcheng Wang, and Jixing Li. \"le petit prince hong kong: Naturalistic fmri and eeg dataset from" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23873" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "older cantonese speakers\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004718.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[721] Unknown authors. \"descriptive name for this dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004720.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[722] Marilyn Horta, Rebecca Polk, and Natalie Ebner. \"single dose intranasal oxytocin administration: Data from healthy younger and older adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004725.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[723] Shrinivas Bishu, Kathleen C Schmidt, Thomas V Burlin, Michael A Charming, Lisa Horowitz, Tianjiang Huang, Zhong-Hua Liu, Mei Qin, B-K Vuong, Aaron J Unterman, Zengyan Xia, Alan Zametkin, Peter Herscovitch, Zenaide Quezado, and Carolyn B Smith. \"effects of propofol anesthesia on rates of cerebral protein synthesis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004730.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[724] Dante Picchioni, Kathleen C Schmidt, Kelly K McWhirter, Inna Loutaev, Adriana J Pavletic, Andrew M Speer, Alan J Zametkin, Ning Miao, Shrinivas Bishu, Kate M Turetsky, Anne S Morrow, Jeffrey L Nadel, Brittney C Evans, Diana M Vesselinovitch, Carrie A Sheeler, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis and memory formation during sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004731.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[725] Dante Picchioni, Kathleen C Schmidt, Inna Loutaev, Adriana J Pavletic, Carrie Sheeler, Shrinivas Bishu, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis in stages of sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004733.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[726] Hamsanandini Radhakrishnan, Chenying Zhao, Valerie J. Sydnor, Erica B. Baller, Philip A. Cook, Damien Fair, Barry Giesbrecht, Bart Larsen, Kristin Murtha, David R. Roalf, Sage Rush-Goebel, Russell Shinohara, M. Dylan Tisdall, Jean Vettel, Scott Grafton, Matthew Cieslak, Theodore Satterthwaite, and Haochang Shou. \"cs-dsi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004737.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[727] Anne Kavounoudias, Caroline Landelle, Jeremy Dana, Bruno Nazarian, Julien Sein, Jean-Luc Anton, Laurent Pruvost, Frederic Giraud, and Olivier Felician. \"audiotact_ya\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004743.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[728] Joe Bathelt, Joanne Taylor, and Kathy Rastle. \"language fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004765.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[729] Rene Weber, Frederic R. Hopp, Allison Eden, and Kate Lee. \"vicarious punishment of moral violations in naturalistic drama narratives predicts cortical synchronization\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004775.v1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[730] Katyucia de Macedo Rodrigues, Emma Ben-Avi, Danielle D. Sliva, Myong sun Choe, Marie Drottar, Ruopeng Wang, Bruce Fischl, Patricia E. Grant, and Lilla Zöllei. \"infant freesurfer test subject (cnybch)\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004776.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[731] A L Williams, S Vicary, and G Orgs. \"functional mri dataset for evaluating brain synchrony among dance spectators\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004783.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[732] Eddy Cavalli, Valérie Chanoine, and Johannes C. Ziegler. \"morphosem\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004786.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.18 + ], + "angle": 0, + "content": "[733] Dylan Nielson, Andre Zugman, Marie Zelenina, and Danny Pine. \"nimh meter (multi-echo test-retest)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004787.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.182, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[734] Fu Yu Kwok, Eric D. Wilkey, Lien Peters, Ellyn Khiu, Rebecca Bull, Kerry Lee, and Daniel Ansari. \"kwok et al., 2023, human brain mapping\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004791.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.24, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[735] Hampus Olsson, Jason Michael Millward, Ludger Starke, Thomas Gladytz, Tobias Klein, Jana Fehr, Wei-Cheng Lai, Christoph Lippert, Thoralf Niendorf, and Sonia Waiczies. \"simulated motion artifacts based on the (mr-art) dataset (https://openneuro.org/datasets/ds004173/versions/1.0.2)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004795.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[736] Umit Keles, Julien Dubois, Kevin J. M. Le, J. Michael Tyszka, David A. Kahn, Chrystal M. Reed, Jeffrey M. Chung, Adam N. Mamelak, Ralph Adolphs, and Ueli Rutishauser. \"fmri data for: Multimodal single-neuron, intracranial eeg, and fmri brain responses during movie watching in human patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004798.v1.0.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.44, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[737] Audrey Mazancieux, Franck Mauconduit, Alexis Amadon, Jan Willem de Gee, Tobias Donner, and Florent Meyniel. \"brainstem fmri signaling of surprise across different types of deviant stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004808.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[738] Michelle Moerel and Essa Yacoub. \"high-res gradient echo epi and 3d grase data of auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004814.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.557, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[739] Agustin Lage-Castellanos, Federico De Martino, Geoff M. Ghose, Omer Faruk Gulban, and Michelle Moerel. \"selective attention sharpens population receptive fields in human auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004815.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.711 + ], + "angle": 0, + "content": "[740] Ana F. Palenciano, Carlos González-García, Jan de Houwer, Baptist Lieefoghe, and Marcel Brass. \"concurrent response and action effect representations across the somatomotor cortices during novel task preparation.\" OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004829.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[741] Michelle Moerel, Federico De Martino, Kamil Ugurbil, Essa Yacoub, and Elia Formisano. \"processing of frequency and location in human subcortical auditory structures\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004835.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[742] Fran López-Caballero, Mark Curtis, Brian Coffman, and Dean Salisbury. \"magnetoencephalographic (meg) pitch and duration mismatch negativity (mmn) in first-episode psychosis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004837.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[743] Kira Noad, David Watson, and Timothy Andrews. \"game of thrones - a naturalistic viewing dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004848.v1.0.1." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23874" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[744] Sean Paulsen and Michael A. Casey. \"shanxi enculturation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004866.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[745] Nafiseh Ghazanfari, Jeih-San Liow, Min-Jeong Kim, Raven Cureton, Adrian Lee, Carson Knoer, Madeline Jenkins, Jinsoo Hong, Jose A. Montero Santamaria, H. Umesha Shetty, Anthony Galassi, Paul Wighton, Martin Norgaard, Douglas N. Greve, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. [11c]ps13 demonstrates pharmacologically selective and substantial binding to cyclooxygenase-1 (cox-1) in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004868.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[746] Elenor Morgenroth, Stefano Moia, Laura Vilaclara, Raphael Fournier, Michal Muszynski, Maria Ploumitsakou, Marina Almato Bellavista, Patrik Vuilleumier, and Dimitri Van De Ville. \"emo-film\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps004892.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[747] Toru Maekawa, Takafumi Sasaoka, Toshio Inui, Alan S. R. Fermin, and Shigeto Yamawaki. \"heart rate and insula activity increase in response to music in individuals with high interoceptive sensitivity\". OpenNeuro, 2023. doi: 10.18112/openneuro.ps004894.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[748] Harrison Ritz and Amitai Shenhav. \"pact_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004909.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[749] Ye Wu, Xiaoming Liu, Yunzhi Huang, Tao Zhou, and Fan Zhang. \"an open relaxation-diffusion mri dataset in neurosurgical studies\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004910.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[750] Alejandra Figueroa-Vargas, Gabriela Valdebenito-Oyarzo, María Paz Martínez-Molina, Francisco Zamorano, and Pablo Billeke. \"probability decision-making task with ambiguity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004917.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[751] David V. Smith, James Wyngaarden, Cooper J. Sharp, Daniel Sazhin, Ori Zaff, Dominic Fareri, and Johanna Jarcho. \"an fmri dataset of social and nonsocial reward processing in young adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004920.v1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[752] Lonike K Faes, Agustin Lage-Castellanos, Giancarlo Valente, Zidan Yu, Martijn A. Cloos, Luca Vizioli, Steen Moeller, Essa Yacoub, and Federico De Martino. \"predictive processing in the auditory cortex measured at 7t\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004928.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[753] Shari Liu, Kirsten Lydic, Lingjie Mei, and Rebecca Saxe. \"fmri dataset: Violations of psychological and physical expectations in human adult brains\". OpenNeuro, 2024. doi: 10.18112/openneuro.dss004934.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[754] Elise M. Cardinale, Jessica Bezek, Julia O Linke, Olivia Siegal, Gabrielle Freitag, Anni Subar, Khosravi Parmis, Ajitha Mallidi, Olivia Peterson, Isaac Morales, Simone P Haller, Courtney Filippi, Kyunghun Lee, Melissa A. Brotman, Daniel S. Pine, Ellen Leibenluft, and Katharina Kirkanski. \"multivariate assessment of inhibitory control in youth: Links with psychopathology and brain function dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005754.v1.1.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[755] Patrick Klepits, Karl Koschutnig, Thomas Zussner, and Andreas Fink. \"effects of a seven-week running intervention with moderate intensity on the volume of the hippocampus and depressive symptoms in young men from the general population.\" OpenNeuro, 2024. doi: 10.18112/openneuro. ds004937.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[756] Jens Jensen, Andreanna Benitez, Erin Klintworth, Hunter Moss, and Kathryn Thorn. \"pydesigner example dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004945.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[757] Roxane Philips, Chris Baeken, Joel Billieux, James Harris, Pierre Maurage, Ismael Muela, Irem Tugce Oz, Athur Pabst, Guillaume Sescousse, Claus Vogele, and Damien Brevers. \"brain mechanisms discriminating enactive mental simulations of running and plogging\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004946.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[758] Marleen Haupt, Monika Graumann, Santani Teng, Carina Kaltenbach, and Radoslaw M. Cichy. \"braille letters - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004956.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.439 + ], + "angle": 0, + "content": "[759] Angelique Volfart, Katie McMahon, and Greig de Zubicaray. \"a comparison of denoising approaches for spoken word production related artefacts in continuous multiband fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004957.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[760] Douglas N. Greve and Bruce Fischl. \"the freesurfer maintenance dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004958.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[761] Corinna Perchtold-Stefan, Christian Rominger, Karl Koschutnig, and Andreas Fink. \"truecrime\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004965.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[762] Ekaterina Torubarova, Caroline Arvidsson, Jonathan Berrebi, Julia Uddén, and André Pereira. \"neuroengage\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004996.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[763] Eric C. Leuthardt, Joshua S. Shimony, Abraham Z. Snyder, Donna Dierker, and Ki Yun Park. \"retrospective task/rest fmri data from tumor patients\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005003.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[764] Neha A Reddy, Rebecca G Clements, and Molly G Bright. \"simultaneous cortical, subcortical, and brainstem mapping of sensory activation\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005009.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[765] Michael I. Demidenko, Edward D. Huntley, and Daniel P. Keating. \"adolescent health risk behavior study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005012.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[766] David Seminowicz, Shana Burrowes, Alexandra Kearson, Jing Zhang, Samuel Kimmel, Luma Samawi, Andrew Furman, and Michael Keaser. \"mbsr\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005016.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[767] Unknown authors. \"pop\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005017.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[768] Sara Ponticorvo and Fabrizio Esposito. \"hearing loss connectome\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005026.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[769] Michael I. Demidenko, Ryan Klaus, Mary Soules, and Mary M. Heitzeg. \"michigan longitudinal study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005027.v1.0.3." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23875" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[770] Marie Mueckstein, Kai Görgen, Stephan Heinzel, Urs Granacher, A. Michael Rapp, and Christine Stelzel. \"modality-based multitasking and practice - fmri\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005038.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[771] Michael S. Cohen, Qiongwen (Jovie) Cao, and Jean Decety. \"political moralization (pmm)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005040.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[772] Panpan Chen, Chi Zhang, Bao Li, Li Tong, Linyuan Wang, Shuxiao Ma, Long Cao, Ziya Yu, and Bin Yan. \"an fmri dataset in response to large-scale short natural dynamic facial expression videos\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005047.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[773] S.H.P. Collin, R. Kempner, S. Srivatsan, and K.A. Norman. \"neural codes track prior events in a narrative and predict subsequent memory for details\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms00505.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[774] Guangyu Zhou, Gregory Lane, Thorsten Kahnt, and Christina Zelano. \"dataset2\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005056.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[775] Jing Wang. \"cr dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005063.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.387, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[776] jessica. \"ds: ds005069:1.0.0\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005069.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.415, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[777] jessica. \"ds: ds005072:1.0.1\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005072.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.443, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[778] Brian P Keane, Bart Krekelberg, Ravi Mill, Steven M Silverstein, Judy L Thompson, Megan R Serody, Deanna M Barch, and Michael W Cole. \"the neural basis of visual shape completion in schizophrenia and bipolar disorder\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005073.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[779] Caroline Landelle, Nawal Kinany, Benjamin De Leener, Nicholas D. Murphy, Ovidiu Lungu, Véronique Marchand-Pauvert, Dimitri Van De Ville, and Julien Doyon. \"simultaneous brain and spinal cord at rest: an fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005075.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[780] Unknown authors. \"sequence pilot: multiecho and multiband fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005085.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[781] Hannah Grotzinger, Laura Pritschet, Elle M. Murata, Tyler Santander, Pavel Shapturenka, Caitlin Taylor, and Emily G. Jacobs. \"28andhe\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05115.v1.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[782] Charles J. Lynch and Conor Liston. \"weill cornell medicine multi-echo (wcm-me) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005118.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[783] David V. Smith, Cooper J. Sharp, Abraham Dachs, James Wyngaarden, Daniel Sazhin, Jen Yang, Melanie Kos, Tia Tropea, Ishika Kohli, John A. Clithero, Ingrid Olson, Tania Giovannetti, Dominic Fareri, and Johanna M. Jarcho. \"social reward and nonsocial reward processing across the adult lifespan: An interim multi-echo fmri and diffusion dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005123.v1.1.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[784] Hannah Tarder-Stoll, Christopher Baldassano*, and Mariam Aly*. \"the brain hierarchically represents the past and future during multistep anticipation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005125.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[785] Takahiro Hirao, Mitsuhiro Miyamae, Daisuke Matsuyoshi, Ryuto Inoue, Yuhei Takado, Takayuki Obata, Makoto Higuchi, Naotsugu Tsuchiya, and Makiko Yamada. \"a neuroimaging dataset during sequential color qualia similarity judgments\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005126.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[786] Dante Picchioni, Jeff H Duyn, and Jacco A de Zwart. \"amri 16-n-0031 sleep1\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005127.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[787] Sangil Lee, Runxuan Niu, Lusha Zhu, Andrew Kayser, and Ming Hsu. \"deception signaling task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps005128.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.264, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[788] Adam Raikes. \"sequence test for multiband dwi and magnetization transfer sequence\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005134.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.307, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[789] Binke Yuan and Junjie Yang. \"scnu-mandarin-cantonese-dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005139.v1.0.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[790] Ladan Shahshahani, Maedbh B. King, Caroline Netekoven, Rich B. Ivry, and Jorn Diedrichsen. \"wmfs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005148.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.406, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[791] Benjamin Lahner, Kshitij Dwivedi, Polina Iamshchinina, Monika Graumann, Alex Lascelles, Gemma Roig, Alessandro Thomas Gifford, Bowen Pan, SouYoung Jin, N.Apurva Ratan Murty, Kendrick Kay, Radoslaw Cichy*, and Aude Oliva*. \"modeling short visual events through the bold moments video fmri dataset and metadata\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005165.v1.0.4." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[792] Andrei Barborica, Felicia Mihai, Laurentiu Tofan, Irina Oane, and Ioana Mindruta. \"dataset of intracranial EEG during cortical stimulation evoking visual effects\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005169.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.562, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[793] Tomoyasu Horikawa. \"mind captioning\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005191.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[794] Nicholas G. Cicero, Michaela Klimova, Laura D. Lewis, and Sam Ling. \"differential cortical and subcortical visual processing with eyes shut\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005194.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.648, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[795] Jiwoong Park, Hayoung Song, and Won Mok Shim. \"narrativepuzzle\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005215.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.69, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[796] Lucas Mahler, Julius Steiglechner, Benjamin Bender, Tobias Lindig, Dana Ramadan, Jonas Bause, Florian Birk, Raphael Heule, Edyta Charyasz, Michael Erb, Vinod Jangir Kumar, Gisela E Hagberg, Pascal Martin, Gabriele Lohmann, and Klaus Scheffler. \"ultracortex: Submillimeter ultra-high field 9.4t brain mr image collection and manual cortical segmentations\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005216.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.803, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[797] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"reinforcement-learning generalization\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005230.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[798] Kirill A. Fadeev, Ilacai V. Romero Reyes, Dzerassa D. Goiaeva, Tatyana S. Obukhova, Tatyana M. Ovsiannikova, Andrey O. Prokofyev, Anna M. Rytikova, Artyom Y." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23876" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "Novikov, Vladimir V. Kozunov, Tatyana A. Stroganova, and Elena V. Orekhova. \"perception of vowel sounds in children with autism spectrum disorders and typically developing children (meg/erf study)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005234.v2.1.7." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[799] J. Veillette and H. Nusbaum. \"motor-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005239.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[800] Taylor Salo, M. Dylan Tisdall, Lia Brodrick, Adam Czernuszenko, David R. Roalf, Sage Rush-Goebel, Nicholas Wellman, Matthew Cieslak, and Theodore D. Satterthwaite. \"single-echo/multi-echo comparison pilot\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005250.v1.1.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[801] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, and Jean-Michel Badier. \"gloups_meg\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps005261.v2.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[802] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Incentivized reaching\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05263.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[803] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Ultra-high field functional connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005264.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[804] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Approach-avoid\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005265.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[805] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Active escape\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005266.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[806] Anja Ischebeck, Hannah Kreilinger, Joe Miller, Margit Hoefler, Iain D. Gilchrist, and Christof Koerner. \"fixating targets in visual search: The role of dorsal and ventral attention networks in the processing of relevance and rarity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005267.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[807] Hsi T. Wei, Farhan B. Faisal, Tamara Beck, Claire Shao, and Jed A. Meltzer. \"picture-word interference dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005279.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[808] Vassiki Chauhan, Krystal McCook, and Alex White. \"dataset accompanying \"reading reshapes stimulus selectivity in the visual word form area\", chauhan, mccook and white (2024)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005295.v1.0.2." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[809] Laura Pritschet, Caitlin M. Taylor, Daniela Cossio, Joshua Faskowitz, Tyler Santander, Daniel A. Handwerker, Hannah Grotzinger, Evan Layher, Elizabeth R. Chrastil, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Emily G. Jacobs. \"maternal brain project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005299.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[810] unknown author. \"neural bases of psychological reactance: An fmri study on dogmatic and suggestive health communication\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005304.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[811] Robert O. Duncan and Evan A. Owens. \"bsc_bids_071724\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005355.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[812] Unknown authors. \"neurogame project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005357.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[813] Elizabeth J. Rizor, Viktoriya Babenko, Neil M. Dundon, Renee Beverly-Aylwin, Alexandra Stump, Margaret Hayes, Luna Herschenfeld-Catalan, Emily G. Jacobs, and Scott T. Grafton. \"hormone health study (hhs)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005360.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[814] Alejandra Figueroa-Vargas, Leonie Kausel, María Paz Martínez-Molina, Francisco Zamorano, Ximena Stecher, Patricia Soto-Icaza, and Pablo Billeke. \"neurocovid mri dwi and fmri with reversal learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005364.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[815] Amy Kartar, Toru Horinouchi, Balázs Örzsik, Brittany Anderson, Lottie Hall, Duncan Bailey, Sarah Samuel, Nati Beltran, Samira Bouyagoub, Chris Racey, Yoko Nagai, Iris Asllani, Hugo Critchley, and Alessandro Colasanti. \"neurobiological substrates of altered state of consciousness induced by high ventilation breathwork\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005365.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[816] Magnus S. Bom, Annette M.A. Brak, Mathijs A.H.L.L. Raemaekers, Nick F. Ramsey, Mariska J. Vansteensel, and Mariana P. Branco. \"large-scale fmri dataset for the design of motor-based brain-computer interfaces\". OpenNeuro, 2025. doi: 10.18112/openneuro.ms005366.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[817] Jessica Archibald, Amy Bouchard, Ralph Noeske, Dikoma Shungu, and Mark Mikkelsen. \"test-retest reliability of multi-metabolite edited mrs at 3t using press and slaser\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005371.v1.1.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[818] Hyunwoo Gu, Joonwon Lee, Sungje Kim, Jaeseob Lim, Hyang-Jung Lee, Heeseung Lee, Minjin Choe, Dong-Gyu Yoo, Jun Hwan (Joshua) Ryu, Sukbin Lim, and Sang-Hun Lee. \"discrimination-estimation task\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005381.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[819] Chema G. Penalver, David Lopez-Garcia, and Marta Becerra Losada. \"atexp_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005386.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[820] Jeryn Chang, Pamela A McCombe, Robert D Henderson, Diana Zanfirache, Christine C Guo, JingLei Lv, Kelly Garner, Saskia Bollmann, Shyuan T Ngo, Frederik J Steyn, and Thomas B Shaw. \"an fmri dataset for appetite neural correlates in people living with motor neuron disease\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005401.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[821] Grant and None. \"grant dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005412.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[822] Alexander P. Rockhill and Ahmed M. Raslan. \"numbers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005415.v1.0.0." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23877" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[823] Antonin Rovai, Valentina Lolli, Xavier De Tiège, Nicola Trotta, and Serge Goldman. \"cerebrovascular reactivity: Co2 inhalation and resting state comparisons\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005418.v3.2.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[824] Juan Eugenio Iglesias, Adrià Casamitjana, Brain L Edlow, Martina Bocchetta, Dorit Kliemann, Matteo Mancini, Loïc Peter, Eleanor Robinson, Roberto Annunziata, Juri Althonayan, Shauna Crampsie, Emily Blackburn, Benjamin Billot, Alessia Atzeni, Oula Puonti, Yael Balbastre, Peter Schmidt, James Hughes, Jean C Augustinack, Lilla Zöllei, David L Thomas, Catherine Strand, Janice L Holton, and Zane Jaunmuktane. \"nextbrain data extension\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005422.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[825] Juan Pablo Franco, Peter Bossaerts, and Carsten Murawski. \"the neural dynamics associated with computational complexity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005427.v1.1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[826] M. R. Ehlers, J. H. Kryklywy, Beukers A. O., Moore S. R., B. J. Forys, Anderson A.K., and R. M. Todd. \"valenced tactile information is evoked by neutral visual cues following emotional learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005449.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[827] Emilie C. Sleight, Oscar Esteban, and Frédéric Grouiller. \"fmriprep boot camp - 7t training dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005454.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[828] Tingting Guo, Xuedi Liu, Mo Chen, Yongben Fu, and Taomei Guo. \"an fmri dataset for investigating language control and cognitive control in bilinguals\". OpenNeuro, 2024. doi: 10.18112/openneuro.dS005455.v1.1.5." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[829] Wangjing Yu, Asieh Zadbood, Avi J. H. Chanales, and Lila Davachi. \"priority\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005464.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[830] Antonin Rovai and Emilie Caspar. \"sense of agency in free and coerced moral decision-making among civilians and military personnel\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005469.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[831] Patrycja Scislewska, Michal Rafal Zareba, Julia Lengier, Aaron Schirmer, Piotr Bebas, and Iwona Szatkowska. \"monetary incentive delay task - structural and functional images of 37 men; study of associations between circadian characteristics (eveningness, distinctness) and affective processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005479.v1.0.3." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[832] Unknown authors. \"neurowater project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005492.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[833] Niharika Gajawelli, Cameron Glick, Yinming Sun, Faizan Badami, Amit Etkin, and Manish Saggar. \"single-pulse tms fmri\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005498.v2.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[834] Zoya Surani, Ted Turesky, Eileen Sullivan, Talat Shama, Rashidul Haque, Nazrul Islam, Shahriah Hafiz Kakon, Xi Yu, William Petri, Charles Nelson, and Nadine Gaab. \"psychosocial adversity and inhibitory control: an fmri study of children growing up in extreme poverty\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005504.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[835] Ioannis Ntoumanis, Julia Sheronova, Alina Davydova, Maria Dolgaleva, Iiro P Jäaskeläinen, Vladimir Kosonogov," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Anna N Shestakova, and Vasily Klucharev. \"deciphering the neural responses to a naturalistic persuasive message\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005518.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[836] Xingfeng Shao, Qinyang Shou, Kimberly Felix, Brandon Ojogho, Xuejuan Jiang, Brian T Gold, Megan M Herting, Eric L Goldwaser, Peter Kochunov, L Elliot Hong, Ioannis Pappas, Meredith Braskie, Hosung Kim, Steven Cen, Kay Jann, and Danny JJ Wang. \"dp-pcasl data (cbf, att, bbb kw) from 186 cognitively normal participants (8-92 years)\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005529.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[837] Viviana Greco, Tamas A. Foldes, Neil A. Harrison, Kevin Murphy, Marta Wawrzuta, Mahmoud E. A. Abdellahi, and Penelope A. Lewis. \"depotentiation of emotional reactivity using tmr during rem sleep\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005530.v1.0.9." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[838] Hiroto Q. Yamaguchi, Naoko Koide-Majima, Rieko Kubo, Tomoya Nakai, and Shinji Nishimoto. \"narrative movie fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005531.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[839] Nicholas G. Cicero, Nina E. Fultz, Hongbae Jeong, Stephanie D. Williams, Daniel Gomez, Beverly Setzer, Tracy Warbrick, Manfred Jaschke, Ravij Gupta, Michael Lev, Giorgio Bonmassar, and Laura D. Lewis. \"high-quality multimodal mri with simultaneous eeg using conductive ink and polymer-thick film nets\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005533.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[840] Zsuzsanna Nemecz and Attila Keresztes. \"item and spatial pattern separation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005559.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[841] María Paz Martínez-Molina, Alejandra Figueroa-Vargas, Francisco Zamorano, and Pablo Billeke. \"expectation of conflict stimuli\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005571.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[842] Jarod L Roland, Carl D Hacker, Abraham Z Snyder, Joshua S Shimony, JOhn M Zempel, David D Limbrick, Matthew D Smyth, and Eric C Leuthardt. \"a comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005573.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[843] Toshiki Okumura, Ikuhiro Kida, Atsushi Yokoi, Shinji Nishimoto, Kazushige Touhara, and Masako Okamoto. \"dataset for 7t odor fmri experiment\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005576.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[844] Alexa Tompary and Lila Davachi. \"integration of overlapping sequences emerges with consolidation through mpfc neural ensembles and hippocampal-cortical connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005581.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[845] Johannes Schultz, Maria Gädeke, and Tom Willems. \"sodec - responsibility fmri experiment dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005588.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[846] Yaqiong Chai, Hedong Zhang, Carlos Robles, Andrew Shinho Kim, Kyung-Wook Kang, Jiyoung Kim, JungWon Shin, Abigail Trang, Jongmok Ha, Emily Lee, Yeonsil Moon, and Hosung Kim. \"precise perivascular space" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23878" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "segmentation on t2-weighted magnetic resonance imaging from human connectome project-aging\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005595.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[847] Bolong Wang, XuanXuan Zhang, and Xiangzhen Kong. \"a naturalistic fmri dataset in response to public speaking\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005596.v1.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[848] EC Washington University School of Medicine Leuthardt, JS Washington University School of Medicine Shimony, AZ Washington University School of Medicine Snyder, D Washington University School of Medicine Dierker, and KY Washington University School of Medicine Park. \"retrospective task/rest fmri data from tumor patients - for sora\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005597.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[849] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"visuomotor rotation learning and reward-based motor learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005598.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[850] Xitong Chen, Stephanie Leach, Juniper Hollis, Dillan Cellier, and Kai Hwang. \"thalamocortical contributions to hierarchical cognitive control\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005600.v1.1.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[851] Peter N. Taylor, Yujiang Wang, Callum Simpson, Vytene Janiukstyte, Jonathan Horsley, Karoline Leiberg, Beth Little, Harry Clifford, Sophie Adler, Sjoerd B. Vos, Gavin P Winston, Andrew W McEvoy, Anna Miserocchi, Jane de Tisi, and John S Duncan. \"the imaging database for epilepsy and surgery (ideas)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005602.v1.0.0." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.556 + ], + "angle": 0, + "content": "[852] Tim Mitchell and Carl Hacker. \"mitchell_hacker_2013\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005603.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[853] Unknown authors. \"v4 crowding\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005604.v1.0.1." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[854] Xuefeng Yan, Fabrice G. Simeon, Jeih-San Liow, Cheryl L. Morse, Susovan Jana, Jose A. Montero Santamaria, Madeline Jenkins, Isabella Piacentino, Anthony Galassi, Sami S. Zoghbi, Victor W. Pike, Robert B. Innis, and Zanotti-Fregonara. \"[18f]sf51, a novel 18f-labeled pet radioligand for translocator protein 18kda (tspo) in brain, works well in monkeys but fails in humans\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005619.v1.1.0." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23879" + } + ] +] \ No newline at end of file diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_origin.pdf b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63f8f23f3193b0e89b40f875e777c2ff28e4c9ff --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/237d5190-142b-4bfa-94c8-a21dfa727bcc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e63b153783565df1210e3d61c740eda829ec84a7103a3f01e0ebeac9bb06176d +size 1299073 diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/full.md b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72760adb55ba936755e71a7a2db342c1870a3831 --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/full.md @@ -0,0 +1,1142 @@ +# An OpenMind for 3D medical vision self-supervised learning + +Tassilo Wald $^{*,1,2,3}$ , Constantin Ulrich $^{*,1,4}$ , Jonathan Suprijadi $^{*,†}$ , Sebastian Ziegler $^{1,2}$ , Michal Nohel $^{\dagger,6,7}$ , Robin Peretzke $^{1,4}$ , Gregor Kohler $^{\dagger,1}$ , Klaus Maier-Hein $^{1,2,3,4,5,8}$ + +1 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany + +$^{2}$ Helmholtz Imaging, DKFZ; $^{3}$ Faculty of Mathematics and Computer Science, Heidelberg University + +4 Medical Faculty, Heidelberg University; 5 National Center for Tumor Diseases (NCT), Heidelberg + +$^{6}$ Faculty of Electrical Engineering and Communication, Brno University of Technology, Czech Republic + +7 University hospital Ostrava, Department of Deputy director for science, research and education + +$^{8}$ Pattern Analysis and Learning Group, Department of Radiation Oncology tassilo.wald@dkfz-heidelberg.de + +# Abstract + +The field of self-supervised learning (SSL) for 3D medical images lacks consistency and standardization. While many methods have been developed, it is impossible to identify the current state-of-the-art, due to i) varying and small pretraining datasets, ii) varying architectures, and iii) being evaluated on differing downstream datasets. In this paper, we bring clarity to this field and lay the foundation for further method advancements through three key contributions: We a) publish the largest publicly available pre-training dataset comprising $114k$ 3D brain MRI volumes, enabling all practitioners to pre-train on a large-scale dataset. We b) benchmark existing 3D self-supervised learning methods on this dataset for a state-of-the-art CNN and Transformer architecture, clarifying the state of 3D SSL pre-training. Among many findings, we show that pre-trained methods can exceed a strong from-scratch nnU-Net ResEnc-L baseline. Lastly, we c) publish the code of our pre-training and fine-tuning frameworks and provide the pre-trained models created during the benchmarking process to facilitate rapid adoption and reproduction. Available here. + +# 1. Introduction + +Self-Supervised Learning (SSL) has emerged as a transformative approach in deep learning, enabling the extraction of robust and general representations from data-rich domains [1-3]. This paradigm has led to significant advancements in areas with abundant data, such as natural language process + +ing and natural image computer vision. In the domain of 3D medical imaging, where annotation is expensive and annotated data scarce, SSL pre-training holds great promise as it can learn transferable representations without requiring extensive manual labeling efforts. However, currently the field of medical image analysis follows two predominant approaches to develop vision models: training models from scratch, often using the nnU-Net framework [4], or adopting supervised pre-training with large, monolithic annotated datasets [5-7] or aggregates of many datasets to increase supervised data scale [8, 9]. While the latter approach indicates a willingness to adopt pre-training paradigms, it raises the question why self-supervised learning is hardly pursued in 3D medical imaging. We believe this lack of SSL method development and adoption can be attributed to two major factors: + +1) A lack of a large, open-source 3D dataset for pretraining. Acquiring large unlabeled datasets for self-supervised pre-training is a non-trivial task. While large studies like the UK-BioBank or the NIH Adolescent Brain Cognitive Development (ABCD) exist and hold $>100\mathrm{k}$ or $>40\mathrm{k}$ 3D Volumes respectively, access to them depends on a positive internal review process requiring an appropriate project proposal. To add to the complexity, datasets are often subject to unique Data Usage Agreements (DUAs) that impose various restrictions on recipients. These rules may include: (i) mandatory acknowledgments, (ii) internal administrative review of manuscripts prior to submission [10-14], (iii) inclusion of a consortium in the author list [12-14], or (iv) requiring the dataset's name to appear in the manuscript title [13]. While acknowledgements are reasonable and exemptions for the remaining restrictions can be requested, these DUAs create legal barriers that hinder re + +![](images/cfc5f8ea733febaa604803f077e12569109f4f2e56dcd1acd0f52e22ed76f4e0.jpg) +Figure 1. The OpenMind dataset contains 114k 3D Head-And-Neck volumes of 23 different modalities. It represents the largest openly accessible dataset of 3D medical images currently available. + +production or creation of large-scale pre-training datasets. Consequently, the majority of current SSL methods are either developed on large, restricted or proprietary datasets [15-17] or are developed on small-scale publicly available datasets [18-24], leading to the current fragmented landscape of pre-training datasets. + +2) A lack of comparability: To adopt existing pre-trained models or conduct one's own SSL pre-training, it is important to know which model or pre-training method is the best. Currently, this is an impossible task, due to: i) Different pre-training dataset selection (as previously highlighted). ii) Different architecture choices. Some methods use CNNs for pre-training [15, 17, 21], some ViT-like Transformers [18, 19], others Swin Transformers [22, 23] and even others hybrid architectures [16, 24]. iii) Insufficient downstream dataset choices, potentially providing unreliable results. + +In this paper we address these issues and provide three key contributions: i) We share the largest publicly available 3D dataset, comprising 114k 3D volumes of 23 different brain MRI modalities. ii) We conduct a comprehensive self-supervised learning benchmark leveraging this dataset to assess the performance of state-of-the-art CNN and transformer-based architectures on 15 segmentation and classification downstream datasets. iii) We open-source our pre-training and fine-tuning frameworks as well as the trained model checkpoints to enable rapid reproduction and novel method development in the field of 3D medical image segmentation. + +Table 1. Large medical datasets are often restricted through custom DUAs limiting their overall applicability. + +
DatasetPatients3D VolumesModalitiesAccess
ABCD [25]11,38541,5802Restricted
PPMI [27]2,0119,5545Restricted
ADNI [28]1,5639,5483Restricted
OASIS31,37615,8727Restricted
OASIS46614,0215Restricted
BraTS21 [29]1,2515,0044CC-BY-4.0
CT-RATE [25]21,30450,1882CC-BY-NC-4.0
OpenMind34,191113,92124CC-BY-4.0
+ +# 2. The OpenMind Dataset + +To provide practitioners access to a large-scale dataset for self-supervised learning, we introduce the OpenMind dataset: a large-scale, easily accessible resource containing 114k Head-and-Neck MRI volumes across various modalities, sourced and curated from the OpenNeuro platform. As the largest freely accessible 3D brain imaging dataset to date, OpenMind surpasses the NIH Adolescent Brain Cognitive Decline initiative [25] by a factor of three, and exceeds the recent CT-Rate dataset [26] by a factor of two, while being released under a permissive CC-BY license (see Tab. 1). + +# 2.1. Dataset creation + +All data was sourced or derived from datasets available on the OpenNeuro platform [30]. OpenNeuro (formerly known as openfMRI) is a publicly accessible data repository designed to support neurological research, adhering to the FAIR principles (Findable, Accessible, Interoperable, Reusable). The platform currently hosts over 1,200 public datasets, encompassing data from more than 51,300 subjects. Due to the diverse nature of its collection, OpenNeuro includes a wide array of modalities, such as MRI, PET, fMRI, MEG, EEG, and iEEG, from studies involving both healthy and diseased subjects, enabling research across various aspects of neurology. From this extensive collection, all available 3D MRIs, such as T1-weighted and T2-weighted scans, as well as 4D diffusion-weighted MRIs were collected from 800 different studies (citations [55-854]). This aggregation resulted in a total of 71k 3D MRI scans1 and 15k 4D diffusion-weighted images from a total of 34,139 subjects. While existing 3D SSL methods could be directly trained on the 71k 3D MRI scans spanning multiple MRI sequences, the 4D diffusion-weighted images required pre-processing to create 3D derivatives which we describe in the following section. Overall, due to the high amount of independent studies, the OpenMind dataset provides great levels of diversity regarding participant age distribution, countries of origin, MRI modalities, scanner manufacturer, model and scanning protocols. + +DWI Preprocessing Diffusion-weighted imaging (DWI) is an advanced MRI modality that captures the random motion of water molecules in tissue by applying diffusion-sensitive gradient pulses, enabling the assessment of microstructural properties. It is widely used for detecting abnormalities such as strokes, tumors, and white matter changes. In clinical and research settings, DWI data is often preprocessed into more interpretable 3D derivatives, including mean diffusion (MD) maps, and fractional anisotropy (FA) maps, which simplify analysis and interpretation. Additionally, T2-weighted images can be derived from DWI scans. Consequently, we pre-processed the DWI images into MD maps, FA maps, and T2-weighted images, resulting in an additional $43\mathrm{k}$ 3D images. The detailed preprocessing pipeline is described in the appendix in B.2. + +Defacing Protecting patient privacy is a big concern in the medical domain, hence, the facial region of Head-and-Neck scans are commonly anonymized. Common anonymization techniques include defacing – replacing the volume with a constant value –, blurring everything but the brain or brain extraction, see Figure 2 in the appendix. These augmented regions can interfere with reconstruction-based SSL methods, potentially penalizing models for attempting to reconstruct plausible anatomical features in anonymized regions. Therefore, masks delineating these regions are needed to allow excluding them from the loss computation. Unfortunately, not all datasets provide these masks, despite applying some sort of anonymization. + +To address this issue, we generate associated anonymization masks and anatomy masks for cases where these are not provided (Figure 2 bottom in the appendix). The anonymization mask identifies areas where anatomical structures have been artificially modified or removed and where no loss should be calculated, whereas anatomy masks indicate where relevant foreground is present, allowing to avoid sampling non-empty patches which may speed up the convergence of the pre-training methods. Both of these masks are created using the model proposed by Nohel et al. [31]. + +Meta-Data enrichment While OpenNeuro follows the Brain Imaging Data Structure (BIDS) format, metadata organization varies significantly across datasets. Participant demographics (e.g., age, sex, health status) and image acquisition details (e.g., repetition times, scanner types, imaging protocols) are inconsistently reported, creating challenges for systematic analysis. We curated and harmonized all available, disparate metadata into a unified, complementary metadata structure. This standardized framework ensures consistency and enables users to efficiently filter the OpenMind dataset based on specific selection criteria, such as subject characteristics or imaging + +parameters, enhancing the flexibility of the dataset and simplifying the extraction of relevant sub-populations. + +In addition to the harmonized metadata, we provide an Image Quality Score (IQS) for each imaging modality of each of the 800 datasets. This score represents a subjective usability score of the images for self-supervised pretraining. The assessment considers factors such as image noise levels, motion artifacts, and other image quality-degrading elements, condensing these evaluations into a unified score ranging from 1 to 5 with 1 representing the highest quality images. The quality scores were determined through manual inspection of multiple images from each modality in each dataset and were assessed independently by two raters to ensure reliability (more details in the appendix B.1). Together, the harmonized metadata and Image Quality Score are introduced to test the effects of simple data-filtering techniques on self-supervised methods trained on the OpenMind dataset. If significant effects of pre-training data choice can be measured, the OpenMind datasets would show to have sufficient size to enable data-centric research in the domain of self-supervised learning for 3D medical image analysis, akin to [2, 32, 33]. + +Useability To maximize the utility of the OpenMind dataset, we provide the final dataset of 114k 3D images on Hugging Face together with anonymization and anatomy masks, as well as the harmonized metadata including the Image Quality Score. + +Why benchmarking matters While the OpenMind dataset addresses the issue of data accessibility and data scale in self-supervised learning for 3D medical imaging, the current state-of-the-art remains unknown. To establish a reference point and provide an initial set of baselines, we conduct a comprehensive benchmarking effort to compare existing 3D SSL methods within a common framework. + +# 3. The OpenMind Benchmark + +The field of self-supervised learning (SSL) for 3D medical imaging remains fragmented due to inconsistencies in pre-training datasets, architectures, and downstream evaluations. Without standardized benchmarks, it is impossible to fairly compare methods and determine the most effective approaches [15]. The OpenMind Benchmark addresses this challenge by evaluating a broad range of SSL methods under controlled conditions, using a common pre-training dataset, standardized architectures, and diverse downstream tasks, which we detail in this section. + +# 3.1. Standardization + +To establish a standardized and fair evaluation, all self-supervised learning (SSL) methods trained in the benchmark were trained on the OpenMind dataset, ensur + +ing that pre-training dataset variability does not influence results. Further, we use two state-of-the-art architectures, the ResEnc-L Residual Encoder U-Net [34] to represent convolutional neural networks (CNNs) and the recent Primus-M Transformer [35], allowing us to analyze SSL effectiveness across different architectural paradigms. Pre-trained models are then evaluated on a diverse set of downstream datasets for segmentation and classification tasks. + +Downstream Datasets All downstream datasets used are split 50/50 into a train and test set. Moreover, we group our datasets into development datasets and test datasets. The development datasets serve to optimize SSL method hyperparameters on their train split, should the method not be directly applicable on the OpenMind dataset with original hyperparameters. During this process, the most important method-specific hyperparameters as well as learning rate and weight decay are optimized based on an additional 80/20 split on the training set of the development datasets. Aside from this, the development datasets are used to determine the optimal fine-tuning schedules, hence results of the optimal fine-tuning schedules are reported on the validation split of the dataset (Tab. 4 and in the appendix in table 8). + +To avoid potential biases introduced during the optimization process on the development datasets, we withheld additional test datasets, which we left untouched and only evaluated after finishing all method optimizations on the development datasets. Moreover, since datasets tend to be small and targets diverse, results are inherently noisy and may not generalize from one dataset to another. To reduce the chance of over optimizing for one dataset, we include four development datasets as well as eight test segmentation datasets and three test classification datasets. + +# Development datasets: + +1. Atlas R2.0 (ATL) [36]: Brain scans of stroke patients imaged through T1 weighted MRIs. +2. Stanford Brain Metastasis (SBM) [37]: Brain scans of metastasis imaged through pre-contrast T1, contrast-enhanced gradient echo, contrast-enhanced spin echo and a T2 FLAIR sequence. +3. AMOS22 (AMO) [38]: CT and MRI scans of 15 abdominal organs. +4. KiTS19 (KIT) [39]: CT images of kidneys and kidney tumors. + +While AMOS and KiTS can be considered out-of-distribution due to the pre-training images exclusively depicting the head-and-neck region, these datasets were found to be robust benchmarking datasets in [34], hence were included despite their shift. Further, during develop + +ment, we only optimized the methods for segmentation, due to it being the most common task in 3D medical image analysis [35]. + +# Test datasets (Segmentation): + +1. ISLES (ISL) [40]: DWI $(\mathrm{b} = 1000)$ and ADC map brain MRI scans of patients with delineations of stroke lesions. +2. HNTS-MRG (HNT) [41]: MR images of patients with head-and-neck primary tumors and metastatic lymphnodes pre and mid-treatment of which we use the pretreatment images and labels. +3. HaN-Seg OAR (HAN) [42]: CT and MR images of Head-and-Neck organs-at-risk of which we use only the MR images. +4. MS FLAIR (MSF) [43]: FLAIR MRIs of hyperintense multiple-sclerosis lesions with associated delineations. +5. ToP-CoW (TPC) [44]: Time-of-Flight (ToF) Magnetic Resonance Angiography (MRA) of the Circle of Willis and associated annotations of the arterial structure. +6. Yale Brain Metastasis (YBM) [45]: Pre and post-contrast T1, T2 and FLAIR MRI images of brain metastases. +7. Cosmos (COS) [46]: Carotid vessel wall segmentations and atherosclerosis diagnosis, with contours of the vessel walls on 3D-VISTA MRI. +8. ACDC (ACD) [47]: Cardiac cine-MRI image of three segmented ventricular structures. + +# Test datasets (Classification): + +1. MR-Net (MRN) [48]: Knee T1w, T2w and Proton Density (PD) weighted MRIs of ACL or meniscal tears as well as other abnormalities which we classify. +2. RSNA-SpineFrac (RSN) [49]: CT Spine fractures with the goal of identifying lumbar spine degenerative conditions. We classify fracture and no-fracture. +3. ABIDE (ABI) [50]: T1w MRI images of either healthy controls or subjects with autism, on which we conduct binary classification - autism or not-autism. + +Pre-training configuration To ensure a fair comparison, all SSL methods are trained under comparable pre-training and equal fine-tuning schedules. Specifically, we pre-train all methods for a total of 1000 epochs with 250 steps per epoch each, with each method trained DDP on four 40GB A100s to decrease pre-training time. Furthermore, the pretrained models were fine-tuned for a total amount of 150 or 1000 epochs with 250 steps per epoch and a common batch size of 2 for both architectures following an identical learning rate schedule implemented within the nnU-Net framework [4] for the segmentation downstream datasets. For classification downstream datasets, all models were fine-tuned for 200 dataset epochs within the public Image classification framework, which supports 3D classification. We provide more detailed descriptions of the fine-tuning process and frameworks in the appendix C. + +# 3.2. Benchmark scope + +In the benchmark, three key aspects of SSL are investigated, highlighting the possible research directions the OpenMind dataset and benchmark can enable: + +i) SSL method performance: We compare the efficacy of various SSL methods in learning transferable representations to various downstream segmentation and classification datasets when using the same pretraining dataset, using two common architectures and a unified fine-tuning schedule. +ii) Fine-tuning methodology We evaluate various fine-tuning schedules for both architectures to investigate the importance of a well-configured downstream fine-tuning schedule. +iii) Data-centric baselines We evaluate the influence of data-filtering on the large OpenMind dataset, to evaluate if data-centric research can be conducted on this scale. + +SSL Methods Various SSL methods native to 3D medical images exist. In the benchmark, we include contrastive methods, like Volume Contrastive (VoCo) [22], a 3D SimCLR adaptation, the pseudo-supervised VolumeFusion (VF) [16], default masked autoencoders (MAE) [3], sparse adaptations of them for CNNs (S3D) [15], masked image modeling (SimMIM) [18] and, the denoising and masked auto-encoder combination ModelsGenesis (MG) [21]. Moreover, we incorporate the SwinUNETR pre-training scheme (SwinUNETR) [23], which integrates masked inpainting, contrastive learning and rotation prediction2. We tried including the 3D geometric-matching-based pre-training method GVSL [20], but the method failed to converge both when using the original implementation and when reimplementing it within our framework, regardless of whether we trained it on our OpenMind dataset or the ABCD-NIH dataset. We provide a brief description of each included SSL method and its used hyperparameters in the appendix D. Methods, which did not feature publicly available Code that could be tested for functionality or integrated into our repository, were excluded, e.g. GLMAE [19]. Moreover, more recent self-supervised methods were given higher priority for inclusion in our benchmark than older methods. Furthermore, prominent 2D methods were excluded, as adapting their configuration from the natural imaging domain to the 3D medical imaging domain would require substantial development effort, e.g. due to large batch size requirements leading to VRAM issues. + +Fine-tuning Schedules Effective adaptation during fine-tuning is crucial for optimizing performance in medical image segmentation, as it determines how well a pre-trained + +model can transfer knowledge to a target dataset. We test five different fine-tuning strategies for adaptation to segmentation datasets, aiming to balance maintaining general pre-trained representations with adapting the encoder to specialized representations through varying learning rate (lr) schedules: i) nnU-Net's default training strategy employs a polynomial decay lr with an initial lr of $1\mathrm{e} - 2^3$ , which we reduce to 1e-3 for the transfer learning setting. ii) The Frozen schedule follows the same lr schedule but restricts training to the decoder. iii) The Warm-Up schedule introduces an initial phase with a linear increase in the lr before transitioning to the default schedule. iv) The Valley scheme prioritizes decoder adaptation by first training it with a linear decreasing lr, followed by a linear warm-up phase for the full network, and then the default schedule. v) Finally, the Sawtooth scheme employs a two-stage warm-up: first for the decoder while keeping the encoder frozen, and then for the whole network, both with a linear increasing lr, before transitioning to the default schedule. + +We train the majority of fine-tuning runs for 150 epochs to reduce computational cost. For these, each warm-up stage lasts 15 epochs. For the CNNs we provide additional fine-tuning experiments trained for 1000 epochs, in which the warm-up stage is extended to 50 epochs. For a more detailed description and visualization of this, we defer to the appendix C, where we also provide additional information on the classification fine-tuning. + +Dataset Filters When training on a large enough quantity of data, optimizing the dataset becomes an important aspect that can greatly improve the resulting model performance [2, 32, 33]. Given the scale and diversity of the OpenMind dataset, we evaluate four simple filtering schemes to measure if we are at a scale where data filtering can be effective. The filtering schemes implemented are: 1. Filtering by the manually created Image Quality Score (IQS) at three thresholds, and 2. filtering all modalities but T1w, T2w and FLAIR images, the most common downstream dataset modalities. + +Reproducibility While evaluating this benchmark, a lot of pre-trained models were generated. To maximize reproducibility, we provide the checkpoints of all pre-training methods for both architectures. Further, we provide an integration into the nnU-Net framework that holds the provided fine-tuning code, enabling users to quickly adapt the pretrained checkpoints to a desired downstream task. + +Table 2. Segmentation results of tested pre-training methods. CNN and Transformer architectures were pre-trained on the same dataset for the same amount of steps. All methods were fine-tuned with the optimal architecture specific fine-tuning scheme for 150 epochs, unless indicated by suffix $1k$ . Color ranges are set per-dataset and per-architecture. S3D and SimMIM are only shown for CNN or Transformer as they are architecture specific. nnU-Net def.: Dynamically planned default nnU-Net trained from scratch; 1k: Fine-tuned for 1000 epochs instead of 150 epochs; ID Mean: In distribution mean; OOD Mean: Mean across datasets of non Head-and-Neck datasets. + +
PT MethodDice Similarity Coefficient (DSC) [%] on ...
Dataset of same anatomical region (ID)Dataset of OOD regionAverage across ...
ATLSBMISLHNTHANMSFTPCYBMCOSACDAMOKITIDOODAll
nnU-Net def. 1k58.7059.9878.4062.9853.3752.1979.5058.4346.1991.1088.0087.2161.0888.7768.00
nnU-Net def.56.0860.4178.2259.3732.2755.8476.7856.7349.3190.7283.8877.6158.3384.0764.77
ResEnc-L (CNN)
Scratch 1k58.2153.4379.1465.7558.2454.9079.9456.1271.5792.0988.7387.4864.1589.4370.47
Scratch57.0254.2978.0963.3056.1155.4776.1854.4265.2091.9785.2484.0362.2387.0868.44
VoCo57.1459.6277.5063.4851.1254.9075.1256.9263.4991.4485.6085.7162.1487.5868.50
SwinUNETR56.0757.2577.4561.6449.4254.8274.6957.0565.6890.5384.9585.5461.5687.0167.92
SimCLR57.1559.7278.0163.3251.5655.6877.7759.1468.2091.7686.0684.8563.4087.5669.44
VF57.4259.8878.1864.3251.6757.4276.1159.3163.9891.5785.3886.2163.1487.7269.29
MG58.0361.5777.5865.1154.6955.2577.1458.6771.2791.7486.3586.1764.3788.0970.30
MAE58.2562.4177.8966.5855.1456.8477.9660.0770.8591.9886.7886.1265.1188.3070.91
S3D58.7664.0978.0565.7452.8156.0878.8159.1866.6692.0186.1686.0164.4688.0670.36
Primus-M (Transformer)
Scratch 1k56.7748.5076.5958.4053.4053.2776.3252.5364.6890.8987.2485.5760.0587.9067.01
Scratch51.5143.2675.2355.3050.6054.0073.3150.3062.1190.9380.1776.7357.2982.6163.62
VoCo46.8034.1573.2951.0647.6452.6465.5244.7552.1687.2665.8170.2152.0074.4357.61
SwinUNETR47.2336.3173.8450.1546.4952.8066.2344.4954.8287.9266.1770.3952.4974.8258.07
SimCLR54.6142.6275.4356.7550.8053.5970.0848.3658.2089.9775.7581.2756.7282.3363.12
VF58.6247.3777.5662.3756.1855.0074.9853.9669.7491.4184.9586.1761.7587.5168.19
MG56.5047.3476.7658.4254.0254.6773.6549.7760.7490.8982.1584.4559.1085.8365.78
MAE61.1656.6777.1266.1257.2456.0278.3154.3572.0292.1687.1686.7464.3488.6970.42
SimMIM60.2851.6877.5362.7656.7455.9177.0052.9070.8791.9886.5785.9262.8588.1669.18
+ +Table 3. Classification results of tested pre-training methods. CNN and Transformer architectures were pre-trained and fine-tuned identically on the three classification datasets. PT Method: pre-training method + +
Architecture Metric PT MethodResEnc-L (CNN)Primus-M (Transformer)
Average PrecisionMeanMRNBalanced AccuracyAverage PrecisionMeanMRNBalanced Accuracy
MRNRSNABIRSNABIMeanRSNABIMean
Scratch66.3365.3957.1862.9776.6162.2553.2964.0560.5858.9152.8757.4575.4957.1052.0561.55
VoCo71.9970.3362.2368.1880.3565.5760.1868.7067.9262.0953.2261.0877.6861.6451.2863.53
SwinUNETR70.7270.3762.3267.8079.2566.6059.8868.5864.9860.7956.8460.8776.6159.2255.6063.81
SimCLR68.7473.0259.1766.9878.2767.9657.6667.9667.6760.2655.5961.1876.6959.5254.1563.45
VF68.4666.2258.0164.2378.4363.2855.6065.7766.4161.6658.3762.1476.6760.1456.0664.29
MG68.8571.2159.6966.5979.0966.1457.2467.4968.8661.2452.4160.8477.7359.9754.1363.94
MAE64.5367.8658.3163.5776.7264.2156.2365.7267.0060.2853.0960.1377.2658.8351.8462.64
S3D70.0869.2560.6566.6678.5664.2056.4166.39--------
SimMIM--------65.2759.6253.9459.6177.1558.4953.9363.19
+ +# 4. Results + +In this section we highlight the most important insights of our benchmark, but not all due to spatial constraints. A comprehensive list of them is provided in the appendix A. + +# 4.1. SSL method performance + +Segmentation performance We compare the fine-tuning results (fine-tuned for 150 epochs) of all pre-training methods for the ResEnc-L CNN architecture and the Primus-M Transformer architecture in Tab. 2. For both architectures, we observe that self-supervised pre-training is able to exceed both, a from-scratch nnU-Net default baseline trained for 1000 epochs and their respective from-scratch trained architecture trained for 150 epochs. In particular, reconstruction based methods exhibit the best performance, + +with default MAEs leading the pack. The MAE pre-trained CNN model even exceeds its strong from scratch baseline trained for 1000 epochs in only 150 epochs, and performs even better when fine-tuned for 1000 epochs (see Table 5 in the appendix). Aside from the MAE pre-training paradigm, this is only achieved by the sparse-autoencoder based S3D pre-training scheme with 1000 epochs of fine-tuning. + +Comparing the benefits for each architecture, it can be seen that the Transformer-based Primus-M architecture profits substantially more from pre-training than the ResEnc-L CNN counterpart, however, it shows lower from-scratch results to begin with. Comparing the MAE pre-trained Primus-M to its 1000 epochs from scratch counterpart, it is able to improve about $3.5\%$ DSC points and almost exceeds the ResEnc-L 1000 epochs from scratch baseline in 150 epochs. However, on singular datasets the MAE + +Table 4. Effects of different CNN fine-tuning schedules. + +
DatasetPT Method FT ScheduleVoCoSwinUNETRSimCLRVFMGMAES3DAverage
SBMDefault73.8669.8672.7972.5773.3771.8475.4772.82
Frozen52.3438.5960.8328.5451.3559.4556.5549.67
Warm-Up70.8769.1673.9473.1174.6774.5173.5472.83
Valley70.2670.2972.8275.8174.6873.7172.4072.85
Sawtooth71.5271.0573.7273.7873.8074.7973.9773.23
ATLDefault62.7860.8256.8063.8362.6262.7862.1661.68
Frozen58.2754.2757.9543.3357.2754.4957.1554.68
Warm-Up59.3260.2362.3663.2063.4663.6162.3462.07
Valley59.4360.3461.2164.0061.2362.9863.3861.80
Sawtooth60.4261.1761.3465.0761.5463.4763.7662.40
AMODefault85.9385.4287.1286.0586.9386.7886.7486.42
Frozen53.1861.2662.9230.9757.8950.8352.4852.79
Warm-Up86.9886.6187.1086.5887.3487.6287.5487.11
Valley85.9185.4686.7286.5587.0987.3786.9286.57
Sawtooth86.7885.7686.7586.4687.1587.4886.4586.69
KITDefault85.8083.8381.2084.3483.6984.5682.2183.66
Frozen69.3369.5965.2236.0671.3963.9771.4163.85
Warm-Up84.7181.7781.8083.5183.9985.2482.2083.32
Valley84.0882.4181.5585.0083.4883.4883.7883.40
Sawtooth85.1283.2683.1085.4184.1584.2883.3184.09
AverageDefault77.0974.9874.4876.7076.6576.4976.6576.15
Frozen58.2855.9361.7334.7359.4857.1859.4055.25
Warm-Up75.4774.4476.3076.6077.3677.7576.4076.33
Valley74.9274.6275.5777.8476.6276.8876.6276.15
Sawtooth75.9675.3176.2377.6876.6677.5076.8776.60
+ +pre-trained Primus-M transformer already achieves the highest performance overall (ATL, COS and ACD), including the best pre-trained ResEnc-L. Given the substantial performance discrepancy observed in Wald et al. [35] between the two architectures, this provides the first sign of vision transformers reaching state-of-the-art performance in 3D medical image segmentation. + +However, not all pre-training methods are able to learn useful representations for segmentation. VoCo, SwinUNETR pre-training, and SimCLR (for transformers) do not, or just barely exceed, a from-scratch baseline trained for 150 epochs and can even lead to substantial performance reductions for the transformer architecture likely due to their global pre-training task not learning the local information needed for segmentation. Moreover, individual datasets can be negatively impacted by any pre-training, as for the ISL and HAN dataset when using a ResEnc-L architecture. This is particularly surprising as Primus-M shows strong improvements with its MAE pre-training on both of these datasets. When filtering low quality images this seems to be partially alleviated (see Table 6 in the appendix), indicating that CNNs may learn different patterns in the presence of noise, which a transformer may be more robust to. + +Lastly, comparing the longer 1000 epoch fine-tuning schedule ( see Table 5 in the appendix) and the shorter 150 epoch fine-tuning schedule of the ResEnc-L, it can be observed that the longer fine-tuning schedule can lead to performance degradation on datasets where pre-training showed strong improvements (SBM, HNT). Simultaneously, it reduces the performance gap to the from-scratch 1000 epoch baseline on cases where pre-training was not as helpful (ISL, HAN, AMO), likely by allowing to forget learned representations and adapt more strongly to the downstream task. + +Classification performance We evaluate the fine-tuned performance of our pre-trained methods for classification + +- a more global downstream task - in Tab. 3. Contrastive pre-training methods like VoCo, SwinUNETR, and SimCLR, which were ineffective for segmentation, achieve the highest classification performance. In contrast, MAE pretraining, which was the strongest for segmentation, performs the worst across both architectures and nearly all datasets. Furthermore, segmentation focused methods like S3D and MG do not transfer well to classification, though they do not fall to the lowest ranks. These findings reinforce the distinction between global and local feature learning: contrastive methods specialize in global similarity learning, making them well-suited for classification, whereas reconstruction-based objectives prioritize pixel-wise representations, benefiting segmentation. Interestingly, SwinUNETR pre-training, which incorporated both global and local objectives, still converged toward representations favoring classification over segmentation, suggesting a bias in learned feature hierarchies. + +Additionally, it is to note that from-scratch classification performance is subpar, with ABI barely exceeding randomness (balanced accuracy of $50\%$ ). This may be due to the classification pipeline lacking the refinements of the widely adopted nnU-Net framework, due to the lower prevalence of global reasoning tasks in 3D medical imaging. Thus, while our classification results provide valuable insights, they should be interpreted with caution, particularly compared to our segmentation findings which are more reliable due to the robust benchmarking framework. + +# 4.2. Fine-tuning methodology + +The ablation results of different CNN fine-tuning strategies in Tab. 4 indicate that the Sawtooth fine-tuning strategy consistently achieves the best overall performance, followed closely by the Warm-Up scheme. This shows that adapting the encoder, after tuning the newly initialized decoder is crucial to maximize performance. The staged learning rate increases in these schedules likely help to prevent instability and allow for a smoother transition from pre-training to task-specific adaptation. For the out-of-distribution CT datasets, AMO and KIT, the Default fine-tuning schedule performs very competitively. This indicates that models pretrained on MRI data require greater modifications when adapted to CT images. The high initial lr allows the models to shift their learned features more effectively to the new domain. + +Keeping the encoder frozen during fine-tuning significantly reduces performance across all datasets and underscores the importance of encoder adaptation. However, SimCLR is less affected by this constraint. Notably, both VoCo and SimCLR perform almost comparably to other fine-tuning schedules when the encoder is frozen on ATL, w.r.t. the drop on other datasets. This dataset is closest to the pretraining data in the development datasets, as it consists of a + +single MRI sequence, whereas the other datasets either contain CT scans or multiple MRI sequences as input. + +For the Transformer, fine-tuning results in Table 8 in the appendix reveal that the Warm-Up scheme achieves the best performance across all methods, emphasizing the importance of architecture-specific adaptation. The Frozen schedule leads to an even greater performance drop than for CNNs, as the ResEnc-L decoder, despite being lightweight, has significantly more parameters than the "decoder" of Primus-M. In this setting, fine-tuning can be interpreted almost as linear probing, with only back-to-back transposed conv layers (including normalization and activations) being trained, severely limiting decoder capacity. With a frozen encoder, it can be observed that adapting to the ATL dataset reaches the closest performance to other fine-tuning schedules. + +# 4.3. Data-centric results + +We compare the effect of data filtering of the pre-training dataset for CNN MAE pre-training and highlight the measured downstream performance in Appendix Table 6. It can be observed that filtering the lowest quality images through the Image Quality Score (IQS) shows minor positive effects on overall downstream segmentation performance, despite removing substantial amounts of pre-training data. Only when removing about $66\%$ , keeping only the highest quality images, a degradation in performance is measured. Moreover, it can be observed that reducing data diversity by removing all modalities but T1w, T2w and FLAIR images reduces model performance, despite having more pre-training images than the intermediate IQS filtering method which slightly increased performance. Overall, this indicates that the OpenMind dataset does contain redundant data which may be leveraged by more complex data-centric methods to increase performance further. + +Effect of anonymization on reconstruction Many images of the OpenMind dataset were defaced or brain-extracted to guarantee patient privacy. To allow taking these artifacts into account when training with a reconstruction objective, we created masks that allow ignoring such regions. To measure the effect of such artifacts, we compare the performance of reconstruction methods that were or were not penalized for the reconstruction in such regions. (See Appendix Table 7). It can be observed that ignoring the reconstruction loss of anonymized regions improves the segmentation performance of the MAE and S3D, showing that taking anonymization into account improves representations. + +# 5. Discussion, Limitations and Conclusion + +Our study presents the OpenMind dataset, the largest publicly available 3D medical imaging dataset, alongside a standardized self-supervised learning (SSL) benchmark for 3D medical image analysis. Through extensive experi + +mentation across diverse datasets and across many pretraining paradigms, we demonstrate the superiority of self-supervised pre-training over from-scratch training for 3D medical image analysis, when pre-training on a large-scale dataset. We show that contrastive methods excel in global tasks such as classification, and reconstruction-based approaches perform well in dense tasks such as segmentation. However, we find that no pre-training method of the 3D vision domain generalizes and performs well in both tasks, emphasizing the need for further SSL method development. Moreover, we compared state-of-the-art CNNs and Transformers in our benchmark and show for the very first time that a pre-trained Primus-M Transformer can exceed the strong ResEnc-L CNN from-scratch baseline on many datasets and can exceed the strongest pre-trained ResEnc-L CNN on individual datasets. This marks an important milestone for the domain, as transformers from-scratch so far never exceeded a ResEnc-L model [34, 35, 51]. On average, though, we show that a pre-trained ResEnc-L remains the strongest architecture for now. + +Moreover, our ablation results suggest that fine-tuning strategies are a key determinant of the success of pre-trained methods, with bad strategies losing significant performance. This emphasizes the importance of careful adaptation beyond pre-training. Experiments on fine-tuning for longer showed partial degradation of overall performance, indicating the problem of overfitting to downstream tasks. Together, these insights open up research opportunities for developing and adapting Parameter-Efficient Fine-Tuning (PEFT) methods to 3D medical image analysis to address these issues. Moreover, our fine-tuning experiments showed that keeping the vision encoders frozen shows sub-par performance for all pre-training methods. This emphasizes the lack of generality of currently learned representations and – again – highlights the need for better pre-training methods. While our study provides a lot it also has various limitations: i) Our classification results are not as reliable as the segmentation results, due to the lower interest in classification for 3D medical imaging, requiring further refinements. ii) Current evidence of the utility of the OpenMind dataset for data-centric research is rather weak, due to the minimal improvements observed requiring further investigation. iii) We evaluate pre-training for 1000 epochs due to resource limitations. Pre-training for longer may reveal different trends than shown in this study. + +In summary, by open-sourcing the OpenMind dataset and establishing the OpenMind benchmark, we provide a clear assessment of state-of-the-art SSL in 3D medical imaging and offer a reproducible framework for method development. Through them, we aim to accelerate progress in self-supervised learning for 3D medical images and pave the way for new research directions. + +# 6. Acknowledgments + +This work was partly funded by the Helmholtz Foundation Model Initiative (HFMI) under the subproject 'The Human Radiome Project' (THRP). Part of this work was funded by Helmholtz Imaging (HI), a platform of the Helmholtz Incubator on Information and Data Science. Additionally, a part of the computational resources were provided by the German AI Service Center WestAI. + +# References + +[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023. +[2] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. +[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. +[4] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021. +[5] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023. +[6] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024. +[7] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023. +[8] Constantin Ulrich, Fabian Isensee, Tassilo Wald, Maximilian Zenk, Michael Baumgartner, and Klaus H Maier-Hein. Multilant: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 648-658. Springer, 2023. +[9] Wenxuan Li, Alan Yuille, and Zongwei Zhou. How well do supervised 3d models transfer to medical imaging tasks? + +In The Twelfth International Conference on Learning Representations, 2024. +[10] OASIS. Open access series of imaging studies data use agreement, 2024. URL https://sites.wustl.edu/oasisbrains/. +[11] PPMI. Parkinson's progression markers initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=PPMI. +[12] ADNI. Alzheimer's disease neuroimaging initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=ADNI. +[13] MCSA. Mayo clinic study of aging data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=MCSA. +[14] HABS-HD. Health and aging brain study: Health disparities data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project $\equiv$ HABS_HD. +[15] Tassilo Wald, Constantin Ulrich, Stanislav Lukyanenko, Andrei Goncharov, Alberto Paderno, Leander Maerkisch, Paul F Jaeger, and Klaus Maier-Hein. Revisiting mae pretraining for 3d medical image segmentation. arXiv preprint arXiv:2410.23132, 2024. +[16] Guotai Wang, Jianghao Wu, Xiangde Luo, Xinglong Liu, Kang Li, and Shaoting Zhang. Mis-fm: 3d medical image segmentation using foundation models pretrained on a large-scale unannotated dataset. arXiv preprint arXiv:2306.16925, 2023. +[17] Asbjørn Munk, Jakob Ambsdorf, Sebastian Llambias, and Mads Nielsen. Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation. arXiv preprint arXiv:2408.00640, 2024. +[18] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023. +[19] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023. +[20] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023. +[21] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021. +[22] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024. + +[23] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022. +[24] Fenghe Tang, Ronghao Xu, Qingsong Yao, Xueming Fu, Quan Quan, Heqin Zhu, Zaiyi Liu, and S Kevin Zhou. Hyspark: Hybrid sparse masking for large scale medical image pre-training. arXiv preprint arXiv:2408.05815, 2024. +[25] National Institutes of Health et al. Adolescent brain cognitive development study, 2015. +[26] Ibrahim Ethem Hamamci, Sezgin Er, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Muhammed Furkan Dasdelen, Omer Faruk Durugol, Bastian Wittmann, Tamaz Amiranashvili, et al. Developing generalist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834, 2024. +[27] Kenneth Marek, Sohini Chowdhury, Andrew Siderowf, Shirley Lasch, Christopher S Coffey, Chelsea Caspell-Garcia, Tanya Simuni, Danna Jennings, Caroline M Tanner, John Q Trojanowski, et al. The parkinson's progression markers initiative (ppmi)-establishing a pd biomarker cohort. Annals of clinical and translational neurology, 5 (12):1460-1477, 2018. +[28] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics, 15(4):869-877, 2005. +[29] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. +[30] Christopher J Markiewicz, Krzysztof J Gorgolewski, Franklin Feingold, Ross Blair, Yaroslav O Halchenko, Eric Miller, Nell Hardcastle, Joe Wexler, Oscar Esteban, Mathias Goncavles, et al. The openneuro resource for sharing of neuroscience data. *Elife*, 10:e71774, 2021. +[31] Michal Nohel, Constantin Ulrich, Jonathan Suprijadi, Tassilo Wald, and Klaus Maier-Hein. Unified framework for foreground and anonymization area segmentation in ct and mri data. In Bildverarbeitung für die Medizin 2025, 2025. +[32] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023. +[33] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540, 2023. +[34] Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F + +Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 488-498. Springer, 2024. +[35] Tassilo Wald, Saikat Roy, Fabian Isensee, Constantin Ulrich, Sebastian Ziegler, Dasha Trofimova, Raphael Stock, Michael Baumgartner, Gregor Kohler, and Klaus Maier-Hein. Primus: Enforcing attention usage for 3d medical image segmentation, 2025. URL https://arxiv.org/abs/2503.01835. +[36] Sook-Lei Liew, Bethany P Lo, Miranda R Donnelly, Artemis Zavaliangos-Petropulu, Jessica N Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P Simon, Julia M Juliano, Anisha Suri, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific data, 9(1):320, 2022. +[37] Endre Grøvik, Darwin Yi, Michael Iv, Elizabeth Tong, Daniel Rubin, and Greg Zaharchuk. Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri. Journal of Magnetic Resonance Imaging, 51(1):175-182, 2020. +[38] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems, 35: 36722–36732, 2022. +[39] Nicholas Heller, Fabian Isensee, Klaus H Maier-Hein, Xiaoshuai Hou, Chunmei Xie, Fengyi Li, Yang Nan, Guangrui Mu, Zhiyong Lin, Miofei Han, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis, 67:101821, 2021. +[40] Moritz R Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, et al. Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data, 9(1):762, 2022. +[41] Kareem Wahid, Cem Dede, Mohamed Naser, and Clifton Fuller. Training dataset for hntsmrg 2024 challenge, 2024. URL https://zenodo.org/doi/10.5281/ zenodo.11199559. +[42] Gasper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023. +[43] Ali M. Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita binti Hanafi, Osama Alnuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi. Brain mri dataset of multiple sclerosis with consensus manual lesion segmentation and patient meta information. Data in Brief, 42:108139, 2022. ISSN 2352-3409. doi: https://doi.org/10.1016/j.dib.2022.108139. URL https://www.sciencedirect.com/science/article/pii/S235234092200347X. +[44] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano + +Hoher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, Suprosanna Shit, Houjing Huang, Chinmay Prabhakar, Ezequiel de la Rosa, Diana Waldmannstetter, Florian Kofler, Fernando Navarro, Martin Menten, Ivan Ezhov, Daniel Rueckert, Iris Vos, Ynte Ruigrok, Birgitta Velthuis, Hugo Kuijf, Julien Hammerli, Catherine Wurster, Philippe Bijlenga, Laura Westphal, Jeroen Bisschop, Elisa Colombo, Hakim Baazaoui, Andrew Makmur, James Hallinan, Bene Wiestler, Jan S. Kirschke, Roland Wiest, Emmanuel Montagnon, Laurent Letourneau-Guillon, Adrian Galdran, Francesco Galati, Daniele Falcetta, Maria A. Zuluaga, Chaolong Lin, Haoran Zhao, Zehan Zhang, Sinyoung Ra, Jongyun Hwang, Hyunjin Park, Junqiang Chen, Marek Wodzinski, Henning Muller, Pengcheng Shi, Wei Liu Ting Ma, Cansu Yalçin, Rachika E. Hamadache, Joaquim Salvi, Xavier Llado, Uma Maria Lal-Trehan Estrada, Valeria Abramova, Luca Giancardo, Arnau Oliver, Jialu Liu, Haibin Huang, Yue Cui, Zehang Lin, Yusheng Liu, Shunzhi Zhu, Tatsat R. Patel, Vincent M. Tutino, Maysam Orouskhani, Huayu Wang, Mahmud Mossa-Basha, Chengcheng Zhu, Maximilian R. Rokuss, Yannick Kirchhoff, Nico Disch, Julius Holzschuh, Fabian Isensee, Klaus Maier-Hein, Yuki Sato, Sven Hirsch, Susanne Wegener, and Bjoern Menze. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra, 2024. URL https://arxiv.org/abs/2312.17670. +[45] Divya Ramakrishnan, Leon Jekel, Saahil Chadha, Anastasia Janas, Harrison Moy, Nazanin Maleki, Matthew Sala, Manpreet Kaur, Gabriel Cassinelli Petersen, Sara Merkaj, et al. A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information. Scientific Data, 11(1):254, 2024. +[46] H. Chen, X. Zhao, J. Dou, C. Du, R. Yang, H. Sun, S. Yu, H. Zhao, C. Yuan, and N. Balu. Carotid vessel wall segmentation and atherosclerosis diagnosis challenge, 2022. URL https://vessel-wall-segmentation-2022.grand-challenge.org. +[47] Olivier Bernard, Alain Lalande, Clement Zotti, Cervenansky, and et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE TMI, 2018. +[48] Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet. PLoS medicine, 15(11):e1002699, 2018. +[49] Hui Ming Lin, Errol Colak, Tyler Richards, Felipe C Kitamura, Luciano M Prevedello, Jason Talbott, Robyn L Ball, Ekim Gumeler, Kristen W Yeom, Mohammad Hamghalam, et al. The rsna cervical spine fracture ct dataset. Radiology: Artificial Intelligence, 5(5):e230034, 2023. +[50] Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain ar + +chitecture in autism. Molecular psychiatry, 19(6):659-667, 2014. +[51] Pedro RAS Bassi, Wenxuan Li, Yucheng Tang, Fabian Isensee, Zifu Wang, Jieneng Chen, Yu-Cheng Chou, Yannick Kirchhoff, Maximilian R Rokuss, Ziyan Huang, et al. Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation? Advances in Neural Information Processing Systems, 37:15184-15201, 2025. +[52] Lachlan T. Strike, Narelle K. Hansell, Jessica L. Miller, Kai-Hsiang Chuang, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. "queensland twin adolescent brain (qtab)". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004146.v1.0.4. +[53] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. "aomic-id1000". OpenNeuro, 2021. doi: 10.18112/openneuro.dso003097.v1.2.1. +[54] Makayla Gibson, Roger Newman-Norlund, Leonardo Bonilha, Julius Fridriksson, Gregory Hickok, Argye E. Hillis, Dirk-Bart den Ouden, and Chris Rorden. "aphasia recovery cohort (arc) dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004884.v1.0.2. +[55] Chris Rorden, John Absher, and Roger Newman-Norlund. "stroke outcome optimization project (soop)". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004889.v1.1.2. +[56] Allison C. Nugent, Adam G Thomas, Margaret Mahoney, Alison Gibbons, Jarrod Smith, Antoinette Charles, Jacob S Shaw, Jeffrey D Stout, Anna M Namyst, Arshitha Basavaraj, Eric Earl, Dustin Moraczewski, Emily Guinee, Michael Liu, Travis Riddle, Joseph Snow, Shruti Japee, Morgan Andrews, Adriana Pavletic, Stephen Sinclair, Vinai Roopchansingh, Peter A Bandettini, and Joyce Chung. "the nimh healthy research volunteer dataset". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005752.v2.1.0. +[57] Rani Gera, Maya Bar Or, Ido Tavor, Dana Roll, Jeffrey Cockburn, Segev Barak, Elizabeth Tricomi, John P. O'Doherty, and Tom Schonberg. "characterizing habit learning in the human brain at the individual and group levels: a multi-modal mri study". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004299.v1.0.0. +[58] Denise Park, Joseph Hennessee, Evan T. Smith, Micaela Chan, Carson Katen, Julia Bacci, Sarah Frank, Sarah Monier, Alexandra Collyer, Carol Tamminga, William Moore, Neil Rofsky, Karen Rodrigue, Kristen Kennedy, and Gagan Wig. "the dallas lifespan brain study". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004856.v1.2.0. +[59] Jin Wang, Marisa N. Lytle, Yael Weiss, Brianna L. Yasasaki, and James R. Booth. "a longitudinal neuroimaging dataset on language processing in children ages 5, 7, and 9 years old". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003604.v1.0.7. +[60] Leon Y. Cai, Qi Yang, Praitayini Kanakaraj, Vishwesh Nath, Allen T. Newton, Heidi A. Edmonson, Jeffrey Luci, Benjamin N. Conrad, Gavin R. Price, Colin B. Hansen, Cailey I. Kerley, Karthik Ramadass, Fang-Cheng Yeh, Hakmook Kang, Eleftherios Garyfallidis, Maxime Descoteaux, + +Francois Rheault, Kurt G. Schilling, and Bennett A. Landman. "masivar: Multisite, multiscanner, and multisubject acquisitions for studying variability in diffusion weighted magnetic resonance imaging". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003416.v2.0.2. +[61] Anahit Babayan, Blazeij Baczkowski, Roberto Cozatl, Maria Dreyer, Haakon Engen, Miray Erbey, Marcel Falkiewicz, Nicolas Farrugia, Michael Gaebler, Johannes Golchert, Laura Golz, Krzysztof Gorgolewski, Philipp Haueis, Julia Huntenburg, Rebecca Jost, Yelyzaveta Kramarenko, Sarah Krause, Deniz Kumral, Mark Lauckner, Daniel S. Margulies, Natacha Mendes, Katharina Ohrnberger, Sabine Oligschläger, Anastasia Osoianu, Jared Pool, Janis Reichelt, Andrea Reiter, Josefin Röbbig, Lina Schaare, Jonathan Smallwood, and Arno Villringer. "mpileipzig_mind-brain-body". OpenNeuro, 2020. doi: 10. 18112/openneuro.ds000221.v1.0.0. +[62] Karl Koschutnig, Bernhard Weber, and Andreas Fink. "tidying up white matter: Neuroplastic transformations in sensorimotor tracts following slackline skill acquisition". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003138.v1.0.1. +[63] Sangil Lee and Joseph Kable. "cognitive training". Open-Neuro, 2021. doi: 10.18112/openneuro.ds002843.v1.0.1. +[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022. +[65] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580, 2023. +[66] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. +[67] Tom Schonberg, Christopher Trepel, Craig Fox, and Russell A. Poldrack. "balloon analog risk-taking task". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000001.v1.0.0. +[68] J.M. Moran, E. Jolly, and J.P. Mitchell. "social cognitive tasks (animate movement task, moral judgment task, false belief task)". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000109.v2.0.2. +[69] Ella Gabitov, David Manor, and Avi Karni. "learning and memory: motor skill consolidation and intermanual transfer". OpenNeuro, 2018. +[70] A. Roy, R. A. Bernier, J. Wang, M. Benson, J. French, D. Good, and F.G. Hillary. "cost analysis tbi". OpenNeuro, 2019. doi: 10.18112/openneuro.ms000220.v1.0.0. +[71] Noritaka Yoneyama, Hirohisa Watanabe, Kazuya Kawabata, Epifanio Bagarinao, Kazuhiro Hara, Takashi Tsuboi, Yasuhiro Tanaka, Reiko Ohdake, Kazunori Imai, Rei Ogura, Yasutaka Kato, Michihito Masuda, Tatsuya Hattori, Mizuki Ito, Satoshi Maesawa, Daisuke Mori, Masahisa Katsuno, and Gen Sobue. "ds000245_r1.0.0". OpenNeuro, 2018. + +[72] H Aerts and D Marinazzo. "btc_preop". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s001226.v5.0.0. +[73] Johns CL, Jahn AA, Jones HR, Kush DK, Molfese PJ, Van Dyke JA, Magnuson JS, Tabor W, Mencl WE, Shankweiler DP, and Braze D. "indiv diffs readingskill". OpenNeuro, 2018. doi: 10.18112/openneuro.d001365.v1.0.0. +[74] C. Baldassano, U. Hasson, and K.A. Norman. "schematic narrative perception and recall (scrambled)". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001511.v1.0.5. +[75] A.M. Bornstein, M. Aly, S. Feng, N.B. Turk-Browne, K.A. Norman, and J.D. Cohen. "perceptual decisions result from the continuous accumulation of memory and sensory evidence". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001614.v1.1.0. +[76] Unknown authors. "sc mri dataset". OpenNeuro, 2019. +[77] Eduardo A. Garza-Villarreal, Victor Pando-Naude, Fernando A. Barrios, Peter Vuust, Lene Vase, Elvira Brattico, Erick H. Pasaye, and Sarael Alcauter. "functional connectivity of music-induced analgesia in fibromyalgia". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001928.v1.1.0. +[78] B. Sherman and N.B. Turk-Browne. "stat episodic". Open-Neuro, 2024. doi: 10.18112/openneuro.ds002144.v2.0.0. +[79] Gilles de Hollander and Tomas H. Knapen. "odc mapper 7t surface coil". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002295.v1.0.0. +[80] Tomoyasu Horikawa, Alan S. Cowen, Dacher Keltner, and Yukiyasu Kamitani. "the neural representation of visually evoked emotion is high-dimensional, categorical, and distributed across transmodal brain regions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002425.v1.0.2. +[81] Yazhu Qian, Jinyou Zou, Zihao Zhang, Zhentao Zuo, Yan Zhuo, Danny JJ, Wang, and Peng Zhang. "robust functional mapping of layer-selective responses in human lateral geniculate nucleus with high-resolution 7t fmri". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002672.v1.0.0. +[82] G. Elliott Wimmer, Jamie Li, Krzysztof Jacek Gorgolewski, and Russell A. Poldrack. "rewardbeast". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s002738.v1.0.2. +[83] Francesca M M Citron, Nora Michaelis, and Adele E Goldberg. "l2_sentences". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002842.v1.0.1. +[84] Matteo Visconti di Oleggio Castello, Vassiki Chauhan, Guo Jiahui, and M. Ida Gobbini. "an fmri dataset in response to "the grand budapest hotel", a socially-rich, naturalistic movie". OpenNeuro, 2020. doi: 10.18112/openneuro. ds003017.v1.0.3. +[85] Chiara Banfi, Karl Koschutnig, Kristina Moll, Gerd Schulte-Koerne, Andreas Fink, and Karin Landerl. "reading-related functional activity in children with isolated spelling deficits and dyslexia". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003126.v1.3.1. +[86] Andy Schumann, Feliberto de la Cruz, Stefanie K"ohler, Lisa Brotte, and Karl-J"urgen B"ar. "the influence of heart rate variability biofeedback on cardiac regulation and functional brain connectivity". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003357.v1.0.0. + +[87] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant4 motion". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003441.v1.0.0. +[88] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., Todd R. Parrish, and Barbara Trommer. "response inhibition and selective attention in adults and children with and without ADHD". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003500.v1.2.0. +[89] Nathan M. Muncy, C. Brock Kirwan, and Jonathan J. Wisco. "differences in chemo-signaling compound-evoked brain activity in male and female young adults: A pilot study in the role of sexual dimorphism in olfactory chemosignaling". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003606.v1.0.0. +[90] Beau Sievers, Carolyn Parkinson, Peter J. Kohler, James Hughes, Sergey V. Fogelson, and Thalia Wheatley. "visual and auditory brain areas share a representational structure that supports emotion perception: fmri data". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003715.v1.0.0. +[91] Andreas Fink, Karl Koschutnig, Thomas Zussner, Corinna M. Perchtold-Stefan, Christian Rominger, Mathias Benedek, and Ilona Papousek. "a two-week running intervention reduces symptoms related to depression and increases hippocampal volume in young adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003799.v2.0.0. +[92] John Wilder, Morteza Rezanejad, Sven Dickinson, Kaleem Siddiqi, Allan Jepson, and Dirk B. Walther. "sceneperceptionandparallelism". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03950.v1.0.0. +[93] Zoe Woodhead, Maria Sablik, Kate Watkins, David Carey, Dorothy Bishop, and Emma Karlsson. "comparing language lateralisation using fmri and ftcd". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004073.v1.0.1. +[94] Martin Szinte, Tomas Knapen, and Matthias Nau. "gaze_exp". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004271.v1.1.0. +[95] Maria Czarnecka, Andre Knops, and Marcin Szwed. "number comparison task". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004400.v1.0.6. +[96] Alex L. White, Kendrick Kay, Kenny A. Tang, and Jason D. Yeatman. "raw data for engaging in word recognition elicits highly specific modulations in visual cortex". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004489.v1.0.1. +[97] Loriana Goulding, Austin Schmidt, Isaac Hamm, and C. Brock Kirwan. "fmri investigations of individual differences on memory activation for faces and words: The effects of handedness and phenomenal experience". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004589.v1.0.0. +[98] Halle R. Dimsdale-Zucker and Christopher Baldassano. "common". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004663.v1.0.4. +[99] Rotem Botvinik-Nezer, Bogdan Petre, Marta Ceko, Naomi Friedman, and Tor Wager. "paingen_placebo". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004746.v1.0.1. +[100] Xuefeng Yan, Martin Noergaard, Cheryl L. Morse, JeihSan Liow, Jinsoo Hong, Sanjay Telu, Min-Jeong Kim, Jose + +A Montero Santamaria, Anthony Galassi, Ningping Feng, Sarah K. Williams Avram, Ted B. Usdin, Shawn Wu, Andrea Zhang, Lester Manly, Madeline Jenkins, Maia Van Buskirk, Adrian Lee, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. "positron emission tomography (pet) quantification in healthy humans of cyclooxygenase-2 (cox-2), a potential biomarker of neuroinflammation". OpenNeuro, 2024. doi: 10.18112/openneuro.d004869.v1.1.1. +[101] Liberty S. Hamilton, Maansi Desai, and Alyssa Field. "wired icm sample dataset - workshop on intracranial recordings in humans, epilepsy, dbs". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004993.v1.1.2. +[102] Keith M. Vogt, Alex C. Burlew, Sujatha N. Reddy, Marcus Simmons, Courtney N. Kozdron, and James W. Ibinson. "neural correlates of lidocaine analgesic (nla) study". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005088.v1.0.0. +[103] Sidhant Chopra, Carrisa V. Cocuzza, Connor Lawhead, Jocelyn A. Ricard, Loic Labache, Lauren Patrick, Poornima Kumar, Arielle Rubenstein, Julia Moses, Lia Chen, Crystal Blankenbaker, Bryce Gillis, Laura T. Germine, Ilan Harpaz-Rote, BT Thomas Yeo, Justin T. Baker, and Avram J. Holmes. "transdiagnostic connectome project". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005237.v1.1.3. +[104] Marleen Haupt, Douglas D. Garrett, and Radoslaw M. Cichy. "object recognition in healthy aging (orha) - fmri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005374.v1.0.1. +[105] Luis Martinez Agulleiro, Bowen Deng, Yuliya Yoncheva, Chao-Gan Yan, F Xavier Castellanos, and Marie-Pierre St-Onge. "circadian misalignment and energy balance". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005525.v1.0.0. +[106] A.R. Aron, M.A. Gluck, and R.A. Poldrack. "classification learning". OpenNeuro, 2018. +[107] Rebecca J. Lepping, Ruth Ann Atchley, Evangelia Chrysikou, Laura E. Martin, Alicia A. Clair, Rick E. Ingram, W. Kyle Simmons, and Cary R. Savage. "neural processing of emotional musical and nonmusical stimuli in depression". OpenNeuro, 2018. +[108] Elizabeth Bock, Peter Donhauser, Francois Tadel, Guiomar Niso, and Sylvain Baillet. "meg-bids brainstorm data sample". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000246.v1.0.1. +[109] Joseph M. Orr and Vijay A. Mittal. "rule learning in adolescents at clinical high risk for psychosis shows altered patterns of activation". OpenNeuro, 2022. doi: 10.18112/openneuro.ps01371.v1.1.1. +[110] A.C. Schapiro, T.T. Rogers, N.I. Cordova, N.B. Turk-Browne, and M.M. Botvinick. "neural representations of events arise from temporal community structure". OpenNeuro, 2019. doi: 10.18112/openneuro.d001621.v1.1.0. +[111] Omer Faruk Gulban, Kevin R. Sitek, Satrajit S. Ghosh, Michelle Moerel, and Federico De Martino. "auditory localization with 7t fmri". OpenNeuro, 2019. doi: 10.18112/openneuro.d001942.v1.2.0. +[112] Tomoya Nakai and Shinji Nishimoto. "over 100 task fmri + +dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002306.v1.1.0. +[113] Laura Pritschet, Tyler Santander, Caitlin M. Taylor, Evan Layher, Shuying Yu, Michael B. Miller, Scott T. Grafton, Emily G. Jacobs, Morgan Fitzgerald, and Joshua M. Mueller. "28andme". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002674.v1.0.6. +[114] PhD Heidemarie Laurent, Megan K. Finnegan, and Katherine Haigler. "postnatal affective mri dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003136.v1.0.0. +[115] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant5 suboptimal event related". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003442.v1.0.0. +[116] Monica Crotti, Karl Koschutnig, and Selina Wriessnegger. "the impact of handedness on the neural correlates during kinesthetic motor imagery: a fmri study". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003612.v1.0.4. +[117] Ben Deen and Winrich Freiwald. "parallel systems for social and spatial reasoning". OpenNeuro, 2022. doi: 10.18112/openneuro.ms003814.v1.0.0. +[118] Shaonan Wang, Xiaohan Zhang, Jiajun Zhang, and Chengqing Zong. "a synchronized multimodal neuroimaging dataset to study brain language processing". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004078.v1.2.1. +[119] Shani Agron, Reut Weissgross, Eva Mishor, Lior Gorodisky, Aharon Ravia, Edna Furman-Haran, Tali Weiss, and Noam Sobel. "a chemical signal in human emotional tears blocks aggression in men". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004274.v1.0.0. +[120] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. "neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004493.v1.0.2. +[121] Jose Pedro Manzano Patron, Steen Moeller, Jesper L.R. Andersson, Kamil Ugurbil, Essa Yacoub, and Stamatios N. Sotiropoulos. "edden: Evaluation of diffusion mri denoising". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004666.v1.0.5. +[122] Samira Epp, Gabriel Castrillon, Beija Yuan, Jessica Andrews-Hanna, Christine Preibisch, and Valentin Riedl. "two distinct modes of hemodynamic responses in the human brain". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004873.v2.0.6. +[123] Chloe M. de Nys, Ee Shern Liang, Marita Prior, Maria A. Woodruff, James I. Novak, Ashley R. Murphy, Zhiyong Li, Craig D. Winter, and Mark C. Allenby. "royalbrisbane_tofmra_intracranial aneurysm_database". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005096.v1.0.3. +[124] Luis Marcos-Vidal, Helena Gil-Buitrago, Irene Cisma, Rosamunde C. Hendricks, Scott Atran, and Clara Pretus. "polex". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005375.v1.0.0. +[125] G. Xue and Russell A. Poldrack. "rhyme judgment". Open-Neuro, 2020. doi: 10.18112/openneuro.ms000003.v1.0.0. + +[126] Tom S.M., Fox C.R., Trepel C., and Poldrack R.A. "mixed-gambles task". OpenNeuro, 2018. +[127] K Jimura, E Stover, F Casalis, and RA Poldrack. "ds000006". OpenNeuro, 2018. +[128] Xue G, Aron AR, and Poldrack RA. "stop-signal task with spoken & manual responses". OpenNeuro, 2018. +[129] A.R. Aron, T.E. Behrens, M. Frank, S. Smith, and R.A. Poldrack. "stop-signal task with unconditional and conditional stopping". OpenNeuro, 2018. +[130] Jessica Cohen and Russell Poldrack. "the generality of self-control". OpenNeuro, 2018. +[131] K. Foerde, B.J. Knowlton, and Russell A. Poldrack. "classification learning and tone-counting". OpenNeuro, 2022. doi: 10.18112/openneuro.dso000011.v1.0.0. +[132] Rizk-Jackson, Aron, and Russell A. Poldrack. "ds000017". OpenNeuro, 2018. +[133] R Bilder, R Poldrack, T Cannon, E London, N Freimer, E Congdon, K Karlgodt, and F Sabb. "ucla consortium for neuropsychiatric phenomics la5c study". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000030.v1.0.0. +[134] Russell Poldrack. "myconnectome". OpenNeuro, 2023. doi: 10.18112/openneuro.rs000031.v2.0.2. +[135] R. Alvarez and R.A. Poldrack. "cross-language repetition priming". OpenNeuro, 2018. +[136] R.A. Poldrack, J. Clark, E. J. Pare-Blagoev, D. Shohamy, J. Creso Moyano, C. Myers, and M Gluck. "classification learning and reversal". OpenNeuro, 2018. +[137] Mei-Yen Chen, Corey N. White, Nathan Giles, Albert Elumn, Sagar Parikh, Ungi Kim, W. Todd Maddox, and Russell A. Poldrack. "training of loss aversion modulates neural sensitivity toward potential gains". OpenNeuro, 2018. +[138] Kelly AMC and Milham MP. "simon task". OpenNeuro, 2018. +[139] Kelly AMC, Uddin LQ, Biswal BB, Castellanos FX, and Milham MP. "flanker task (event-related)". OpenNeuro, 2018. +[140] J.V. Haxby, M.I. Gobbini, M.L. Furey, A. Ishai, J.L. Schouten, and P. Pietrini. "visual object recognition". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000105.v3.0.0. +[141] K.J. Duncan, C. Pattamadilok, I. Knierim, and J.T. Devlin. "word and object processing". OpenNeuro, 2018. +[142] Wager T.D., M.L. Davidson, B.L. Hughes, M.A. Lindquist, and K.N. Ochsner. "prefrontal-subcortical pathways mediating successful emotion regulation". OpenNeuro, 2018. +[143] Melina R. Uncapher, J. Benjamin Hutchinson, and Anthony D. Wagner. "incidental encoding task (posner cueing paradigm)". OpenNeuro, 2018. +[144] Michael Hanke, Florian J. Baumgartner, Pierre Ibe, Falko R. Kaule, Stefan Pollmann, Oliver Speck, Wolf Zinke, and Jorg Stadler. "forrest gump". OpenNeuro, 2018. doi: 10.18112/openneuro.ps00113.v1.3.0. +[145] Gorgolewski KJ, Storkey A, Bastin ME, Whittle IR, Wardlaw JM, and Pernet CR. "a test-retest fmri dataset for motor, language and spatial attention functions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000114.v1.0.2. + +[146] Grega Repov $\mathring{\mathrm{A}}_{\mathrm{i}}$ and Deanna M. Barch. "working memory in healthy and schizophrenic individuals". OpenNeuro, 2018. +[147] Jennifer M Walz, Robin I Goldman, Jordan Muraskin, Bryan Conroy, Truman R Brown, and Paul Sajda. "auditory and visual oddball eeg-fmri". OpenNeuro, 2018. +[148] DG Wakeman and RN Henson. "multisubject, multimodal face processing". OpenNeuro, 2025. doi: 10.18112/openneuro.rs000117.v1.1.0. +[149] Katerina Velanova, Mark E. Wheeler, and Beatrix Luna. "maturational changes in anterior cingulate and frontoparietal recruitment support the development of error processing and inhibitory control (antistate)". OpenNeuro, 2018. +[150] Aarthi Padmanabhana, Charles F. Geier, Sarah J. Ordaz, Theresa Teslovich, and Beatrix Luna. "developmental changes in brain function underlying the influence of reward processing on inhibitory control (slot reward)". OpenNeuro, 2018. +[151] C.F. Geier, R. Terwilliger, T. Teslovich, K. Velanova, and B. Luna. "immaturities in reward processing and its influence on inhibitory control in adolescence (ring reward)". OpenNeuro, 2018. +[152] J Xu, S Zhang, VD Calhoun, J Monterosso, and MN Potenza. "task-related concurrent but opposite modulations of overlapping functional networks as revealed by spatial ica". OpenNeuro, 2018. +[153] Nicoletta Cera, Armando Tartaro, and Stefano L. Sensi. "modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fmri study". OpenNeuro, 2018. +[154] Choong-Wan Woo, Mathieu Roy, Jason T. Buhle, and Tor D. Wager. "distinct brain systems mediate the effects of nociceptive input and self-regulation on pain". OpenNeuro, 2018. +[155] Kimberly L. H. Carpenter, Adrian Angold, Nan-Kuei Chen, William E. Copeland, Pooja Gaur, Kevin Pelphrey, Allen W. Song, and Helen L. Egger. "preschool anxiety disorders". OpenNeuro, 2018. +[156] E. Necka, M. Senderecka, B. Kucharzyk, and Falkiewicz M. "who can afford self-control? the neural efficiency mechanism explains effective self-regulation of behavior". OpenNeuro, 2019. +[157] Elissa M. Aminoff and Michael J. Tarr. "scene perception". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000149.v1.0.0. +[158] Paul A. M. Smeets, Floor M. Kroese, Catherine Evers, and D. T. D. de Ridder. "block design food and nonfood picture viewing task". OpenNeuro, 2018. +[159] CR Pernet, P Belin, P McAleer, KJ Gorgolewski, M Valdes-Sosa, M Latinus, I Charest, PEG Bestelmeyer, RH Watson, and D Fleming. "the human voice areas: spatial organisation and inter-individual variability in temporal and extra-temporal cortices". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000158.v1.0.0. +[160] Timothy D. Verstynen. "stroop task". OpenNeuro, 2018. +[161] James K. Bursley, Adrian Nestor, Michael J. Tarr, and J. David Creswell. "offline processing in associative learning". OpenNeuro, 2019. + +[162] Hsu AL, Chou KH, Chao YP, Fan HY, Wu CW, and Chen JH. "physiological contribution in spontaneous oscillations: An approximate quality - assurance index for resting-state fmri signals". OpenNeuro, 2019. +[163] Laura Koenders, Janna Cousijn, Wilhelmina A.M. Vingerhoets, Wim van den Brink, Reinout W. Wiers, Carin J. Meijer, Marise W. J. Machielsen, Dick J. Veltman, Anneke E. Goudriaan, and Lieuwe de Haan. "t1-weighted structural mri study of cannabis users at baseline and 3 years follow up". OpenNeuro, 2019. +[164] Emilia Iannilli. "pre-adolescents exposure to manganese". OpenNeuro, 2018. +[165] Gustav Nilsson, Sandra Tamm, Paolo d'Onofrio, Hanna Å Thuné, Johanna Schwarz, Catharina Lavebratt, Jia Jia Liu, Kristoffer NT Måsson, Tina Sundelin, John Axelsson, Peter Fransson, Göran Kecklund, Håkan Fischer, Mats Lekander, and Torbjörn Åkerstedt. "the stockholm sleepy brain study: Effects of sleep deprivation on cognitive and emotional processing in young and old". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000201.v1.0.3. +[166] Peter Van Schuerbeek, Chris Baeken, and Johan De Mey. "the heterogeneity in retrieved relations between the personality trait 'harm avoidance' and gray matter volumes due to variations in the vbm and roi labeling processing settings". OpenNeuro, 2018. +[167] Christian Stephan-Otto, Sara Siddi, Carl Senior, Daniel Muñoz-Samons, Susana Ochoa, Ana María Sánchez Laforga, and Gildas Brébion. "visual imagery and false memory for pictures". OpenNeuro, 2018. +[168] Dara Dickstein, Mariel Y. Pullman, Corey Fernandez, Jennifer A. Short, Lale Kostakoglu, Karin Knaesaurek, Laili Soleimani, Barry Jordan, Wayne Gordon, Kristen Dams-O'Connor, Bradley N. Delman, Edmund Wong, Cheuk Y. Tang, Steven T. DeKosky, James R. Stone, Robert C. Cantu, Mary Sano, Patrick R. Hof, and Sam Gandy. "imaging [18f]av-1451 and [18f]av-45 in acute and chronic traumatic brain injury". OpenNeuro, 2018. +[169] Jongwan Kim, Jing Wang, Svetlana V. Shinkareva, and Douglas H. Wedell. "affective videos". OpenNeuro, 2018. +[170] Vincent A. Magnotta, Joy T. Matsui, Dawei Liu, Hans J. Johnson, Jeffrey D. Long, Bradley D. Bolster Jr, Bryon A. Mueller, Kelvin Lim, Susumu Mori, Karl G. Helmer, Jessica A. Turner, Sarah Reading, Mark J. Lowe, Elizabeth Aylward, Laura A. Flashman, Greg Bonett, and Jane S. Paulsen. "dwi traveling human phantom study". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000206.v1.0.0. +[171] Tetreault P, Mansour A, Vachon-Presseau E, Schnitzer TJ, Apkarian AV, and Baliki MN. "brain connectivity predicts placebo response across chronic pain clinical trials". Open-Neuro, 2022. doi: 10.18112/openneuro.ds000208.v1.0.1. +[172] Elizabeth DuPre, Wen-Ming Luh, and R. Nathan Spreng. "multi-echo fmri replication sample of autobiographical memory, prospection and theory of mind reasoning tasks". OpenNeuro, 2018. +[173] Liane Young, Alek Chakroff, Emily Wasserman, Rebecca Saxe, James Dungan, Amelia Brown, and Jorie Koster-Hale. "moral judgments of intentional and accidental moral + +violations across harm and purity domains". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000212.v1.0.0. +[174] Xiao Gao, Xiao Deng, Xin Wen, Ying She, Petra Corianne Vinke, and Hong Chen. "neural mechanism underlying appearance social comparison". OpenNeuro, 2018. +[175] Stephen Giles, Jeremy Hall, Merrick Pope, Katie Nicol, and Liana Romaniuk. "eupd cyberball". OpenNeuro, 2018. +[176] Alexander D. Cohen, Andrew S. Nencka, R. Marc Lebel, and Yang Wang. "multiband multi-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity". OpenNeuro, 2018. +[177] Avi J H Chanales, Ashima Oza, Serra E Favila, and Brice A Kuhl. "route learning". OpenNeuro, 2020. doi: 10.18112/openneuro.ms000217.v1.0.0. +[178] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. "flavour pleasantness (oral nutritional supplements)". OpenNeuro, 2018. +[179] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. "flavour pleasantness (regular products)". OpenNeuro, 2018. +[180] Thomas H B FitzGerald, Dorothea Haemmerer, Karl J Friston, Shu-Chen Li, and Raymond J Dolan. "sequential inference vbm". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000222.v1.0.1. +[181] Ian Ballard, Bokyung Kim, Anthony Liatsis, Gökhan Aydogan, Jonathan D. Cohen, and Samuel M. McClure. "magnitude effect". OpenNeuro, 2018. +[182] Evan M. Gordon, Timothy O. Laumann, Adrian W. Gilmore, Dillan J. Newbold, Deanna J. Greene, Jeffrey J. Berg, Mario Ortega, Catherine Hoyt-Drazen, Caterina Gratton, Haoxin Sun, Jacqueline M. Hampton, Rebecca S. Coalson, Annie Nguyen, Kathleen B. McDermott, Joshua S. Shimony, Abraham Z. Snyder, Bradley L. Schlaggar, Steven E. Petersen, Steven M. Nelson, and Nico U.F. Dosenbach. "the midnight scan club (msc) dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ms000224.v1.0.4. +[183] Hilary Richardson, Grace Lisandrelli, Alexa Riobueno-Naylor, and Rebecca Saxe. "mri data of 3-12 year old children and adults during viewing of a short animated film". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000228.v1.1. +[184] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. "integration of sweet taste and metabolism determines carbohydrate reward - study 1". OpenNeuro, 2018. +[185] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. "integration of sweet taste and metabolism determines carbohydrate reward-study 3". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000231.v1.0.0. +[186] Johan D Carlin and Nikolaus Kriegeskorte. "adjudicating between face-coding models with individual-face fmri responses". OpenNeuro, 2018. +[187] Samuel A. Nastase, Yaroslav O. Halchenko, Andrew C. Connolly, M. Ida Gobbini, and James V. Haxby. "neural responses to naturalistic clips of behaving animals in + +two different task contexts". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000233.v1.0.1. +[188] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. "whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 1". OpenNeuro, 2018. +[189] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. "whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 2". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000235.v2.0.1. +[190] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. "whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 3". OpenNeuro, 2019. doi: 10.18112/openneuro.ps00236.v2.0.1. +[191] Yuji Kanazawa, Kimihiro Nakamura, Toru Ishii, Toshihiko Aso, Hiroshi Yamazaki, and Koichi Omori. "phonological memory in sign language relies on the visuomotor neural system outside the left hemisphere language network". OpenNeuro, 2019. +[192] D. Zeithamova, M. De Araujo Sanchez, and A. Adke. "trial timing for multivariate pattern analysis". OpenNeuro, 2018. +[193] Julian Maclaren, Zhaoying Han, Sjoerd B. Vos, Nancy Fischbein, and Roland Bammer. "maclaren test-retest brain volume dataset". OpenNeuro, 2018. +[194] Alvaro Galiano, Reyes Garcia de Eulate, Marta Vidorreta, Miriam Recio, Mario Riverol, José L. Zubieta, and Maria A. Fernandez-Seara PhD. "resting state perfusion in healthy aging". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000240.v2.0.0. +[195] Andrew C. Connolly, J. Swaroop Guntupalli, Jason Gors, Michael Hanke, Yaroslav O. Halchenko, Yu-Chien Wu, Hervé Abdi, and James V. Haxby."ak6". OpenNeuro, 2018. +[196] Steve Petersen, Brad Schlaggar, and Jonathan Power. "washington university 120". OpenNeuro, 2018. +[197] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. "individual brain charting". OpenNeuro, 2018. doi: 10.18112/openneuro.ds000244.v1.0.0. +[198] Guiomar Niso, Jeremy Moreau, Elizabeth Bock, Francois Tadel, and Sylvain Baillet. "meg-bids omega restingstate_sample". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000247.v1.0.2. +[199] Alexandre Gramfort and Matti S Hämäläinen. ””. Open-Neuro, 2020. doi: 10.18112/openneuro.ds000248.v1.2.4. +[200] Agnes Norbury and Ben Seymour. "value generalization in human avoidance learning". OpenNeuro, 2018. +[201] Diana Gorbet and Lauren Sergio. "female action video game players". OpenNeuro, 2018. +[202] Alexander D. Cohen, Andrew S. Nencka, and Yang Wang. "multiband multi-echo simultaneous asl/bold for task-induced functional mri". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000254.v1.0.0. +[203] Yoichi Miyawaki, Hajime Uchida, Okito Yamashita, Masaaki Sato, Yusuke Morito, Hiroki C. Tanabe, Norihiro Sadato, and Yukiyasu Kamitani. "visual image reconstruction". OpenNeuro, 2018. + +[204] Deanna J. Greene, Jonathan M. Koller, Jacqueline M. Hampton, Victoria Wesevich, Andrew N. Van, Annie L. Nguyen, Catherine R. Hoyt, Lindsey McIntyre, Eric A. Earl, Rachel L. Klein, Joshua S. Shimony, Steven E. Petersen, Bradley L. Schlaggar, Damien A. Fair, and Nico U.F. Dosenbach. "behavioral interventions for reducing head motion during migraine scans in children". OpenNeuro, 2018. +[205] Jonathan Power, Valerie Voon, Prantik Kundu, Suyash Bhogawar, Peter Bandettini, and Alex Martin. "multi-echo dambridge". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000258.v1.0.1. +[206] Michael Milham, Cameron Craddock, Russ Tobe, Anna MacKay-Brandt, Stan Colcombe, Bennett Leventhal, and F. Xavier Castellanos. "nki-sample". OpenNeuro, 2018. +[207] Chris Gorgolewski. "the brain of chris". OpenNeuro, 2018. +[208] Mark A. Thornton, Miriam E. Weaverdyck, and Diana I. Tamir. "social predictive coding". OpenNeuro, 2018. +[209] A. Zadbood, J. Chen, Y.C. Leong, K.A. Norman, and U. Hasson. "sherlock_merlin". OpenNeuro, 2018. +[210] Y. Yeshurun, M. Nguyen, and U. Hasson. "milky-vodka". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001131.v1.0.0. +[211] J. Chen, Y.C. Leong, C.J. Honey, K.A. Yong C.H., Norman, and U. Hasson. "sherlock". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001132.v1.0.0. +[212] J. Chen, C. J. Honey, E. Simony, M. J. Arcaro, K. A. Norman, and U. Hasson. "twilight zone movie watching dataset". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001145.v1.0.0. +[213] Chris Gorgolewski, Natacha Mendes, Domenica Wilfling, Elisabeth Wladimirow, Claudine J. Gauthier, Tyler Bonnen, Florence J.M. Ruby, Robert Trampel, Pierre-Louis Bazin, Roberto Cozatl, Jonathan Smallwood, and Daniel S. Margulies. "a high resolution 7-tesla resting-state fmri test-retest dataset with cognitive and physiological measures". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001168.v1.0.1. +[214] Krzysztof J. Gorgolewski. "example mr artifacts". Open-Neuro, 2018. +[215] Mark Thornton, Miriam Weaverdyck, and Diana Tamir. "state representation pictures". OpenNeuro, 2018. +[216] Sam Gilbert and Hoki Fung. "decoding intentions of self and others". OpenNeuro, 2018. +[217] J.A. Lewis-Peacock, J.D. Cohen, and K.A. Norman. "neural evidence of the strategic choice between working memory and episodic memory in prospective remembering". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001229.v1.0.1. +[218] J.A. Lewis-Peacock and K.A. Norman. "competition between items in working memory leads to forgetting". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001232.v1.0.0. +[219] Patrick Beukema and Timothy Verstynen. "singlefingerrsa". OpenNeuro, 2018. +[220] Christopher Baldassano, Diane M. Beck, and Li Fei-Fei. "human-object interaction". OpenNeuro, 2018. + +[221] Michelle G. Hall, Claire K. Naughtin, Jason B. Mattingley, and Paul E. Dux. "olvsl_ object-location visual statistical learning". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001241.v1.0.1. +[222] Tae-Ho Lee, Steven Greening, Taiji Ueno, David Clewett, Allison Ponzio, Michiko Sakaki, and Mara Mather. "examining effects of arousal on responses to salient and non-salient stimuli in younger and older adults". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001242.v1.0.0. +[223] Tomoyasu Horikawa and Yukiyasu Kamitani. "generic object decoding (fmri onImagenet)". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001246.v1.2.1. +[224] Damien Brevers, Sarah C Herremans, Qinghua He, Marie-Anne Vanderhasselt, Mathieu Petieau, Dimitri Verdonck, Tasha Poppa, Sara De Witte, Charles Kornreich, Antoine Bechara, and Chris Baeken. "sports_betting_study". Open-Neuro, 2018. +[225] Matteo Visconti di Oleggio Castello, Yaroslav O. Halchenko, J. Swaroop Guntupalli, Jason D. Gors, and M. Ida Gobbini. "personally familiar faces". OpenNeuro, 2018. +[226] Frank J. Kanayet, Andrew Mattarella-Micke, Peter J. Kohler, Anthony M. Norcia, Bruce D. McCandliss, and James L. McClelland. "numberlinemarking". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001299.v1.0.0. +[227] Sean O'Bryan, Darrell Worthy, Evan Livesey, and Tyler Davis. "inverse base rate". OpenNeuro, 2018. doi: 10.18112/openneuro.ms001302.v1.0.1. +[228] Reshanne R. Reeder, Christian N. L. Olivers, and Stefan Pollmann. "negative_template". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001306.v1.0.0. +[229] E.S. Finn, P.R. Corlett, G. Chen, P.A. Bandettini, and R.T. Constable. "paranoiastory". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001338.v1.0.0. +[230] Ivy F. Tso, Saige E. Rutherford, Yu Fang, Mike Angstadt, and Stephan F. Taylor. "socialbrain". OpenNeuro, 2018. doi: 10.52181/zenodod.1210217. +[231] Jessica-Lily Harvey, Lysia Demetriou, John McGonigle, and Matthew B. Wall. "phfmri_control_tasks". OpenNeuro, 2018. doi: 10.18112/openneuro.rs001344.v1.0.0. +[232] Michael P Notter, Sandra Da Costa, and Micah M Murray. "decoding of multisensory semantics and memories in low-level visual cortex". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001345.v1.0.0. +[233] Brock Kirwan, Thomas Meservy, and Kelly Fadel. "information filtering in electronic networks of practice: An fmri investigation of expectation [dis]confirmation". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001353.v1.0.0. +[234] Amado Catarina, Kovács Petra, Mayer Rebecca, Ambrus G. Geza, Trapp Sabrina, and Kovács Gyula. "namefacepriming". OpenNeuro, 2018. +[235] Christopher Baldassano. "loci case study". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001379.v1.0.0. +[236] Dr. Frank Hillary. "subject data". OpenNeuro, 2018. +[237] Joset A. Etzel, Debbie Yee, Bidhan Lamichhane, Mitchell Jeffers, Elisa Di Rosa, Jenny Crawford, Hongyu An, and Todd S. Braver. "multibandacqtests". OpenNeuro, 2022. doi: 10.18112/openneuro.dso001399.v2.0.0. + +[238] Dobrushina Olga R, Vlasova Roza M, Rumshiskaya Alena D, Litvinova Liudmila D, Mershina Elena A, Sinit-syn Valentin E, and Pechenkova Ekaterina V. "rsfmri_single_session_eeg_nf". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001408.v1.0.3. +[239] Megen Brittell. "audio cartography". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001415.v1.0.0. +[240] Rotem Botvinik-Nezer, Tom Salomon, and Tom Schonberg. "cat snacks functional plasticity". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001417.v1.0.0. +[241] Martin Wegrzyn, Joana Aust, Larissa Barnstorf, Magdalena Gippert, Mareike Harms, Antonia Hautum, Shanna Heidel, Friederike Herold, Sarah M. Hommel, Anna-Katharina Knigge, Dominik Neu, Diana Peters, Marius Schaefer, Julia Schneider, Ria Vombrock, Sabrina M. Zimmer, Friedrich G. Woermann, and Kirsten Labudda. "thought-experiment". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001419.v1.0.1. +[242] Melanie Ganz-Benjaminsen and Martin Noergaard. "[11c]dasb pet gimbi database example". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001420.v1.2.0. +[243] Melanie Ganz-Benjaminsen and Martin Noergaard. "[11c]sb207145 pet gimbi database example". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001421.v1.4.1. +[244] G. Kim, K.A. Norman, and N.B. Turk-Browne. "neural overlap in item representations across episodes impairs context memory". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001430.v1.0.2. +[245] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. "a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001439.v1.2.0. +[246] A.C. Schapiro, E.A. McDevitt, T.T. Rogers, S.C. Mednick, and K.A. Norman. "human hippocampal replay during rest prioritizes weakly learned information and predicts memory performance". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001454.v1.3.1. +[247] Ilaria Berteletti, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, Jérôme Prado, Macarena Suarez Pellicioni, Jessica Wise Younger, and James R. Booth. "brain correlates of math development". OpenNeuro, 2021. doi: 10.18112/openneuro.ds001486.v1.3.1. +[248] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, and Gert J. ter Horst. "valence processing differs across stimulus modalities (multi-echo)". OpenNeuro, 2018. doi: 10.18112/openneuro.ps001491.v1.0.0. +[249] J Lewis-Peacock, BR Postle, CR Cox, and TT Rogers. "long-term memory (ltm) for famous faces, places, and common objects". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001497.v1.0.2. +[250] Nadine Chang, John A. Pyles, Austin Marcus, Abhinav Gupta, Michael J. Tarr, and Elisaa M. Aminoff. "bold5000". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001499.v1.3.1. + +[251] Guohua Shen, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. "deep image reconstruction". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001506.v1.3.1. +[252] C. Baldassano, R. Masis-Obando, U. Hasson, and K.A. Norman. "schematic narrative perception and recall (intact)". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001510.v2.0.3. +[253] P. Kok and N.B. Turk-Browne. "associative prediction of visual shape in the hippocampus". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001517.v1.0.3. +[254] S. C. Chan, M. C. Applegate, N. W. Morton, S. M. Polyn, and K. A. Norman. "lingering representations of stimuli influence recall organization". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001521.v1.0.2. +[255] Feng Rong, A. Lisette Isenberg, Erica Sun, and Gregory S. Hickok. "syllable level speech sequencing". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001525.v1.1.1. +[256] Andrea Courtney, Emma PeConga, Dylan Wagner, and Kristina Rapuano. "calorie-labeled food cues". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001534.v1.1.0. +[257] M. Aly, J. Chen, N.B. Turk-Browne, and U. Hasson. "learning naturalistic temporal structure in the posterior medial network". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001545.v1.1.1. +[258] Eliska Prochazkova. "ds: ds001551:1.0.0". OpenNeuro, 2018. +[259] Javier Gonzalez-Castillo, Ziad S. Saad, Daniel A. Handwaker, Souheil J. Inati, Noah Brenowitz, and Peter A. Bandettini. "100 runs at 3t". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001553.v1.0.1. +[260] O. Lositsky, J. Chen, D. Toker, C.J. Honey, J.L. Poppenk, U. Hasson, and K.A. Norman. "neural pattern change during encoding of a narrative predicts retrospective duration estimates". OpenNeuro, 2018. +[261] Javier Gonzalez-Castilloa, Colin W. Hoy, Daniel A. Handwerker, Vinai Roopchansingh, Souheil J. Inati, Ziad S. Saad, Robert W. Cox, and Peter A. Bandettini. "large single-subject functional mri datasets at 7t". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001555.v1.0.1. +[262] Renzo Huber. "cmrr workshop". OpenNeuro, 2019. doi: mockDOI. +[263] Unknown authors. "ds: ds001566:1.0.1". OpenNeuro, 2018. +[264] A.N. Hoskin, A.M. Bornstein, K.A. Norman, and J.D. Cohen. "refresh my memory: Episodic memory reinstatements intrude on working memory maintenance". Open-Neuro, 2018. +[265] Ian C. Ballard, Anthony D. Wagner, and Samuel M. McClure. "feature discrimination". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001590.v1.0.1. +[266] Unknown authors. "test dataset for xcp software". Open-Neuro, 2018. +[267] Patrick Beukema and Timothy Verstynen. "multifingerrsa". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001597.v1.0.0. +[268] Matt Cieslak, Mark Elliott, and Ted Satterthwaite. "example fieldmaps". OpenNeuro, 2018. + +[269] A.M. Bornstein and K.A. Norman. "reinstated episodic context guides sampling-based decisions for reward". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001607.v1.0.1. +[270] Aya Ben-Yakov and Noga Cohen. "prestimulus exp2". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001608.v1.0.1. +[271] I. Momennejad, A.R. Otto, N. Daw, and K.A. Norman. "offline replay supports planning in human reinforcement learning". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001612.v1.0.2. +[272] Adam Hampshire and Eyal Soreq. "visual working memory: Study one task fmri and behavioural response". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001634.v1.0.1. +[273] Adam Hampshire and Eyal Soreq. "visual working memory: Study two task fmri and behavioural response". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001635.v1.0.1. +[274] Elizabeth Valles. "ironia vev". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001652.v2.0.0. +[275] Mattia Veronese, Gaia Rizzo, Martin Belzunce, Julia Schubert, Barbara Santangelo, Ayla Mansur, Alex Whittington, Joel Dunn, Graham Searle, Andrew Reader, and Roger Gunn. "nrm2018 pet grand challenge dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ms001705.v1.0.1. +[276] A. Shenhav, M.A. Straccia, S. Musslick, J.D. Cohen, and M.M. Botvinick. "dissociable neural mechanisms track evidence accumulation for selection of attention versus action". OpenNeuro, 2019. +[277] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. "a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others (experiment 2)". OpenNeuro, 2019. doi: 10.18112/openneuro. ds001722.v1.1.0. +[278] Unknown authors. "test-geier: Ds: ds001728:1.1.0". Open-Neuro, 2019. +[279] Rotem Botvinik-Nezer, Roni Iwanir, Russell A. Poldrack, and Tom Schonberg. "narps". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s001734.v1.0.5. +[280] Birgit Rauchbauer, Bruno Nazarian, Morgan Bourhis, Magalie Ochs, Laurent Prevot, and Thierry Chaminade. "brain activity during reciprocal social interaction investigated using conversational robots as control condition". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001740.v2.2.0. +[281] Nathaniel P Miller, Yao Liu, Roman Krivochenitser, and Bas Rokers. "unilateral glaucoma 3t dmri". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001743.v1.0.1. +[282] J.R. Manning, J.C. Hulbert, J. Williams, L. Piloto, L. Sahakyan, and K.A. Norman. "a neural signature of contextually mediated intentional forgetting". OpenNeuro, 2019. doi: 10.18112/openneuro.d001745.v1.1.0. +[283] Carrie E. Gold, Athena L. Howell, Jacob Burdis, C. Brock Kirwan, and Gregory L. Thompson. "exploring the resting state neural activity of monolinguals and late and early bilinguals". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001747.v1.1.0. + +[284] Samuel Fynes-Clinton, Lars Marstaller, and Hana Burianova. "differentiation of functional networks during long-term memory retrieval in children and adolescents". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001748.v1.0.4. +[285] Bart Aben, Cristian Buc Calderon, Laurens Van der Cruyssen, Doerte Picksak, Eva Van den Bussche, and Tom Verguts. "context-dependent cognitive control — flanker task". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001751.v1.0.2. +[286] R. VanRullen and L. Reddy. "reconstructing faces from fmri patterns using deep generative neural networks". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001761.v2.0.1. +[287] A. Shenhav, M.A. Straccia, M.M. Botvinick, and J.D. Cohen. "dorsal anterior cingulate and ventromedial prefrontal cortex have inverse roles in both foraging and economic choice". OpenNeuro, 2019. +[288] Virginia AGLIERI, Bastien CAGNA, Pascal BELIN, and Sylvain TAKERKART. "intertva: a multimodal mri dataset for the study of inter-individual differences in voice perception and identification". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001771.v1.0.2. +[289] J.C. Hulbert and K.A. Norman. "neural differentiation tracks improved recall of competing memories following interleaved study and retrieval practice". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001775.v1.0.1. +[290] Alik Wadge and Sam Zorowitz. "effects of on/off deep brain stimulation on cognitive control in treatment-resistant depression (eeg)". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001784.v1.1.2. +[291] Vincent DeLuca and Christos Pliatsikas. "bilingualism and the brain". OpenNeuro, 2022. doi: 10.18112/openneuro. ds001796.v1.7.0. +[292] Samuel Zorowitz, Alexander P Rockhill, Kristen K Kellard, Katherine E Link, Todd Herrington, Diego A Pizzagalli, Alik S Widge, Thilo Deckersbach, and Darin D Dougherty. "neuroanatomical correlates of approach-avoidance conflict (fmri)". OpenNeuro, 2020. doi: 10.18112/openneuro. ds001814.v1.0.7. +[293] Unknown authors. "cat faces mri experiment. ds: ds001818:1.0.0". OpenNeuro, 2019. +[294] Isabel Alkhasli, Katrin Sakreida, Felix M. Mottaghy, and Ferdinand Binkofski. "resting state - tms". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001832.v1.0.1. +[295] Celia Goffin, H. Moriah Sokolowski, Michael Slipenkyj, and Daniel Ansari. "handedness and symbolic number representation". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001838.v1.0.1. +[296] Unknown authors. "spinal stimulation stepping and standing dataset. ds: ds001839:1.0.1". OpenNeuro, 2019. +[297] Noga Cohen and Aya Ben-Yakov. "prestimulus exp1". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001840.v1.0.2. +[298] S.C.Y. Chan, Y. Niv, and K.A. Norman. "a probability distribution over latent causes, in the orbitofrontal cortex". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001847.v1.0.1. + +[299] H. Moriah Sokolowski, Zachary Hawes, Lien Peters, and Daniel Ansari. "parallel adaptation of symbols, quantities, and physical size". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001848.v1.0.1. +[300] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. "social decision-making intertemporal choice task dataset". OpenNeuro, 2019. +[301] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. "social decision-making risky choice task dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001883.v1.0.3. +[302] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. "longitudinal brain correlates of multisensory lexical processing in children". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001894.v1.4.2. +[303] Trevor K. M. Day, Tara M. Madyastha, Peter Board, Mary K. Askren, Nicholas J. Wapstra, Thomas J. Montine, and Thomas J. Grabowski. "ant: Healthy aging and parkinson's disease". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001907.v3.1.0. +[304] Milena Kaestner, Ryan T. Maloney, Kirstie H. Wailes-Newson, Marina Bloj, Julie M. Harris, Antony B. Morland, and Alex R. Wade. "s-cone_motiondepth_fmri". Open-Neuro, 2019. +[305] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. "anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value". OpenNeuro, 2019. +[306] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. "anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value (2)". OpenNeuro, 2019. +[307] Ghootae Kim, Kenneth A. Norman, and Nicholas B. Turk-Browne. "violation differentiation". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001926.v1.0.1. +[308] Gabriel Castrillon, Nico Sollmann, Katarzyna Kurcyus, Adeel Razi, Sandro M. Krieg, and Valentin Riedl. "the physiological effects of non-invasive brain stimulation fundamentally differ across the human cortex". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001927.v2.1.0. +[309] N.C. Hindy, E.W. Avery, and N.B. Turk-Browne. "hippocampal-neocortical interactions sharpen over time for predictive actions". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001946.v1.0.3. +[310] Peter J. Kohler and Anthony M. Norcia. "svndl anatomical data". OpenNeuro, 2019. +[311] Peter J. Kohler, Benoit R. Cotterreau, and Anthony M. Noria. "image segmentation based on relative motion and relative disparity cues (central letter task)". OpenNeuro, 2019. +[312] Peter J. Kohler, Benoit R. Cottereau, and Anthony M. Noria. "image segmentation based on relative motion and relative disparity cues (passive fixation)". OpenNeuro, 2019. +[313] Benjamin J. Griffiths, Stephen D. Mayhew, Karen J. Mullinger, João Jorge, Ian Charest, Maria Wimber, and + +Simon Hanslmayr. "reinstatement_fidelity". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002000.v1.0.0. +[314] Janine Mendola and Elizabeth Bock. "rivalry_tagging". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002001.v1.0.0. +[315] Akram Bakkour, Daphna Shohamy, and Michael N. Shadlen. "memory and decision making dataset". Open-Neuro, 2019. doi: 10.18112/openneuro.ms002006.v1.0.1. +[316] David Clewett, Ringo Huang, Rico Velasco, Tae-Ho Lee, and Mara Mather. "locus coeruleus activity strengthens prioritized memories under arousal". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002011.v1.0.0. +[317] Jakob Heinzle (data acquisition), Kai Goergen (data curation), and Joram Soch (BIDS conversion). "stimdisc". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002013.v1.0.3. +[318] Martin Wegrzyn, Markus Mertens, Christian G. Bien, Friedrich G. Woermann, and Kirsten Labudda. "language production fmri". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002014.v1.0.1. +[319] Yiming Xiao, Jonathan C. Lau, Taylor Anderson, Jordan DeKraker, D. Louis Collins, Terry M. Peters, and Ali R. Khan. "bigbrainmricoreg". OpenNeuro, 2019. doi: 10.18112/openneuro.d0s002016.v1.0.0. +[320] Rebecca Merkley, Benjamin Conrad, Gavin Price, and Daniel Ansari. "investigating the visual number form area: A replication study". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002033.v1.0.1. +[321] Unknown authors. "bids_parkinson". OpenNeuro, 2019. +[322] Jaime J. Castrellon, Jacob S. Young, Linh C. Dang, Ronald L. Cowan, David H. Zald, and Gregory R. Samanez-Larkin. "mesolimbic dopamine d2 receptors and neural representations of subjective value". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002041.v2.0.0. +[323] Emily S Finn, Laurentius Huber, David C Jangraw, Peter J Molfese, and Peter A Bandettini. "layer-dependent activity in human prefrontal cortex during working memory". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002076.v1.0.1. +[324] Hannelore Aerts and Daniele Marinazzo. "btc_postop". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002080.v4.0.0. +[325] Maedbh King, Carlos Hernandez-Castillo, Russell A. Poldrack, Jonathon Walters, Richard B. Ivry, and Jorn Diedrichsen. "multi-domain task battery (mdtb)". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002105.v1.1.0. +[326] Celia Goffin, Michael Slipenkyj, Stephan E. Vogel, Rebecca Merkley, and Daniel Ansari. "development of symbolic number processing". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002116.v1.0.0. +[327] Adam Kimbler, Amanda G. Hamm, and Aaron T. Mattfeld. "conditional visual associative learning task". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002149.v1.0.5. +[328] Charlie, John, Ross, and Joey. "testrun". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002153.v1.0.1. +[329] Florence Bouhali, Zoé Bézagu, Stanislas Dehaene, and Laurent Cohen. "a mesial-to-lateral dissociation for ortho- + +graphic processing in the visual cortex". OpenNeuro, 2019. doi: 10.18112/openneuro.dso002155.v1.0.0. +[330] Dylan M. Nielson and Argyris Stringaris. "multi-echo masking test dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002156.v2.0.0. +[331] Michael Pereira, Nathan Faivre, Inaki Iturrate, Marco Wirthlin, Luana Serafini, Stephanie Martin, Arnaud Desvachez, Olaf Blanke, Dimitri Van de Ville, and Jose del R. Millan. "disentangling the origins of confidence in speeded perceptual judgments through multimodal imaging". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002158.v1.0.2. +[332] Michelle I. Nash, Cooper B. Hodges, Nathan M. Muncy, and C. Brock Kirwan. "pattern separation beyond the hippocampus: A high-resolution whole-brain investigation of mnemonic discrimination in healthy adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002168.v1.1.0. +[333] Jiefeng Jiang, Shao-Fang Wang, Wanjia Guo, Corey Fernandez, and Anthony D. Wagner. "ccd". OpenNeuro, 2019. doi: 10.18112/openneuro.dso02169.v1.0.0. +[334] Tali Weiss, Timna Soroka, Lior Gorodisky, Sagit Shushan, Kobi Snitz, Reut Weisgross, Edna Furman-Haran, Thijs Dhollander, and Noam Sobel. "human olfaction without apparent olfactory bulbs". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002185.v1.1.0. +[335] Michele Svanera and Lars Muckli. "cerebrum: a 3t mri segmentation tool". OpenNeuro, 2019. doi: mockDOI. +[336] Dorit Kliemann and Ralph Adolphs. "caltech rsfmri dataset". OpenNeuro, 2019. +[337] Jordan Bigio, Tali Bitan, Douglas Bolger, Douglas Burman, Fan Cao, Tai-Li Chou, Nadia Cone, Jessica Gayda, Dong Lu, Marisa Lytle, Jenni Minas, and James R. Booth. "cross-sectional multidomain lexical processing". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002236.v1.1.1. +[338] J. Poppenk and K. A. Norman. "briefly cueing memories leads to suppression of their neural representations". Open-Neuro, 2019. doi: 10.18112/openneuro.ds002237.v1.1.0. +[339] J. Fan, J.D. Wammes, J.B. Gunn, D.L. Yamins, K.A. Norman, and N.B. Turk-Browne. "relating visual production and recognition of objects in human visual cortex". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002241.v1.1.1. +[340] Nathan M. Muncy and C. Brock Kirwan. "correcting false memories: The effect of mnemonic generalization on original memory traces". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002242.v1.0.0. +[341] Joshua Hirschfeld-Kroen, Emily Wasserman, Stefano Anzellotti, and Liane Young. "when my wrongs are worse than yours: behavioral and neural asymmetries in first-person and third-person perspectives of accidents". Open-Neuro, 2020. doi: 10.18112/openneuro.d002250.v1.0.1. +[342] Jean-Marie Bonny, Juliana Iranpour, and Béatrice Claise. "viewing_food_cues". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002270.v1.0.0. +[343] Katherine L. Bottenhorn, Taylor Salo, Julio A. Peraza, Michael C. Riedel, Jessica S. Flannery, Adam Kimbler, Alfredo Toll, Diego Suarez, Francis M. Cruz, Israel Zagales, Nayade Caldes, Olivia Dolan, Ruth Zagales, Matthew T. + +Sutherland, Robert W. Laird, and Angela R. Laird. "dense investigation of variability of affect (diva)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002278.v2.0.0. +[344] Unknown authors. "fsl open science dev dataset". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002293.v1.0.0. +[345] Elizabeth Long, Nathan Wheeler, and William Cunningham. "trait-role". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002294.v1.0.1. +[346] J. Poppenk and K.A. Norman. "multiple-object tracking as atool for parametrically modulating memory reactivation". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002311.v1.1.0. +[347] P. G. Bissett, J. M. Shine, J. Durnez, J. Li, K. J. Gorgolewski, and R. A. Poldrack. "rewardbeast". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002316.v1.0.0. +[348] Luiz Pessoa, Christian Meyer, and Srikanth Padmala. "dynamic_passive-threat". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002320.v1.1.0. +[349] Shohini Bhattasali, Jonathan R. Brennan, Wen-Ming Luh, Berta Franzluebbers, and John T. Hale. "the alice dataset: fmri dataset to study natural language comprehension in the brain". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002322.v1.0.4. +[350] Taylor Hanayik. "fsl open science dev dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ms002328.v1.0.2. +[351] A. Sunavsky and J. Poppenk. "neuroimaging predictors of creativity in healthy adults". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002330.v1.1.0. +[352] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. "a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp1". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002336.v2.0.2. +[353] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. "a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp2". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002338.v2.0.2. +[354] Samuel A. Nastase, Yun-Fei Liu, Hanna Hillman, Asieh Zadbood, Liat Hasenfratz, Neggin Keshavarzian, Janice Chen, Christopher J. Honey, Yaara Yeshurun, Mor Regev, Mai Nguyen, Claire H. C. Chang, Christopher Baldassano, Olga Lositsky, Erez Simony, Michael A. Chow, Yuan Chang Leong, Paula P. Brooks, Emily Micciche, Gina Choe, Ariel Goldstein, Tamara Vanderwal, Yaroslav O. Halchenko, Kenneth A. Norman, and Uri Hasson. "narratives". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002345.v1.1.4. +[355] Brock Kirwan, Daniel Bjornn, Bonnie Anderson, Anthony Vance, David Eargle, and Jeffrey Jenkins. "repetition of computer security warnings results in differential repetition suppression effects as revealed with functional mri". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002363.v1.1.1. +[356] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum. + +"emotion regulation in the ageing brain, university of reading, bbsrc". OpenNeuro, 2021. doi: 10.18112/openneuro. ds002366.v1.1.0. +[357] Deni Kurban, Renzo Huber, Gilad Liberman, Dimo Ivanov, Sri Kashyap, Debora Niekaemper, and Benedikt Poser. "ohbm abstract 2020". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002367.v1.0.0. +[358] M.D. Rajendra Morey. "serumgreymattercorticalthickness". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002380.v1.0.1. +[359] Rogers CS, Jones MS, McConkey S, Spehar B, Van Engen KJ, Sommers MS, and Peelle JE. "age-related differences in auditory cortex activity during spoken word recognition". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002382.v1.0.1. +[360] Yoshiharu I Kutani, Takatomi Kubo, Satoshi Nishida, Hideaki Hata, Kenichi Matsumoto, Kazushi Ikeda, and Shinji Nishimoto. "fmri dataset on program comprehension and expertise". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002411.v1.1.0. +[361] Jelle R. Dalenberg, Barkha P. Patel, Raphael Denis, Maria G. Veldhuizen, Yuko Nakamura, Petra C. Vinke, Serge Luquet, and Dana M. Small. "short-term consumption of sucralose with, but not without, carbohydrate impairs neural and metabolic sensitivity to sugar". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002419.v1.0.3. +[362] Yana Panikratova, Alexander Tomyshev, Ekaterina Pechenkova, and Roza Vlasova. "fmri: resting state and arithmetic task". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002422.v1.1.0. +[363] James R. Booth, GE Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, MA Stein, and Michael Tennekoon. "working memory and reward in children with and without attention deficit hyperactivity disorder (adhd)". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002424.v1.2.0. +[364] Damien Brevers, Chris Baeken, Antoine Bechara, Qinghua He, Pierre Maurage, Mathieu Petieau, Guillaume Sescousse, Claus Vögele, and Joel Billieux. "increased brain reactivity to gambling unavailability is a marker of problem gambling". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002513.v1.0.0. +[365] Tamar Kolodny, Michael-Paul Schallmo, and Scott O. Murray. "contrast response functions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002522.v1.0.2. +[366] S. Jahfari, K.R. Ridderinkhof, A.G.E. Collins, T. Knapen, L.J. Waldorp, M.J. Frank, and J. Theeuwes. "learning, inhibitory control, and perception". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s002543.v1.0.1. +[367] Suzanne Oosterwijk, Lukas Snoek, Mark Rotteveel, Lisa Feldman Barrett, and H. Steven Scholte. "shared-states". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002547.v1.1.0. +[368] Laboratory for Computational Neurodiagnostics. "protecting the aging brain (pagb)". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002549.v1.0.1. +[369] Romain Quentin, Jean-Remi King, Etienne Sallard, Nathan Fishman, Ryan Thompson, Ethan Buch, and Leonardo + +Cohen. "differential brain mechanisms of selection and maintenance of information during working memory (meg data)". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002550.v1.0.1. +[370] Tomas Knapen, Wietske van der Zwaag, and Daan van Es. "cerebellum retinotopic mapping". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002574.v1.0.1. +[371] Arnaud Delorme and Scott Makeig. "visual oddball task (256 channels)". OpenNeuro, 2021. doi: 10.18112/openneuro.rs002578.v1.1.0. +[372] S. Zhang, W. Yoshida, H. Mano, T. Yanagisawa, K. Shibata, M. Kawato, and B. Seymour. "cognitive control of sensory pain encoding in the pregenual anterior cingulate cortex. d1 - decoder construction in day 1, d2 - adaptive control in day 2". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002596.v1.0.1. +[373] Lukas Snoek, Tinka Beemstermboer, and H. Steven Scholte. "mb-epi-comparison". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002603.v1.0.0. +[374] Caroline Le Ster, Antonio Moreno, Franck Mauconduit, Vincent Gras, Ruediger Stirnberg, Benedikt A. Poser, Alexandre Vignaud, Evelyn Eger, Stanislas Dehaene, Florent Meyniel, and Nicolas Boulant. "bids_dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002606.v1.1.0. +[375] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. "rpn-signature_study1". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002608.v1.0.2. +[376] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. "rpn-signature_study2". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002609.v1.0.3. +[377] Enzo Tagliazucchi. "meditation interocepcion". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002614.v1.0.0. +[378] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum. "emotion regulation in the ageing brain, university of reading, bbsrc". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002620.v1.0.0. +[379] N Eichert, KE Watkins, and project-specific co authors. "project_larynx". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002634.v3.0.0. +[380] L.N. Van der Laan, C. Scholz, R.A. Poldrack, D.T.D. De Ridder, and A. Smidts. "can we have a second helping? a replication study on the neurobiological mechanisms underlying self-control". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002643.v1.1.0. +[381] Mara Mather, Ringo Huang, David Clewett, Shawn Nielsen, Rico Velasco, Kristie Tu, Sophia Han, and Briana Kennedy. "isometric exercise facilitates attention to salient events in women via the noradrenergic system". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002647.v1.0.1. +[382] Nathan M. Muncy and C. Brock Kirwan. "the medial temporal lobe supports mnemonic discrimination for event duration". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002655.v1.0.1. +[383] Brian J. Coffey, Zachary D. Threlkeld, Andrea S. Foulkes, Yelena G. Bodien, and Brian L. Edlow. "the language + +network reemerges during recovery from severe traumatic brain injury". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002675.v1.0.0. +[384] Cheryl A. Olman and Kimberly B. Weldon. "kung fu panda". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002684.v1.0.0. +[385] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. "ibc". OpenNeuro, 2024. doi: 10.18112/openneuro.rs002685.v2.0.0. +[386] James R. Booth, Gillian E. Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, Mark A. Stein, and Michael Tennekoon. "working memory and reward in adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002687.v1.2.0. +[387] Kendrick Kay, Keith W. Jamison, Luca Vizioli, Ruyuan Zhang, Eshed Margalit, and Kamil Ugurbil. "high-field 7t visual fmri datasets". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002702.v1.0.1. +[388] Christos Chatzichristos, Manuel Morante, Nikos Andreadis, Eleftherios Kofidis, Yiannis Kopsinis, and Sergios Theodoridis. "smiley". OpenNeuro, 2020. doi: 10.18112/openneuro.d002711.v1.1.0. +[389] Sara Aurtenetxe, Nicola Molinaro, Doug Davidson, and Manuel Carreiras. "numbers and letters". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002712.v1.0.1. +[390] Petra Vetter, Lukasz Bola, Lior Reich, Matthew Bennett, Lars Muckli, and Amir Amedi. "avscenes_blind". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002715.v1.0.0. +[391] Liron Rozenkrantz, Reut Weissgross, Tali Weiss, Inbal Ravrebi, Idan Frumin, Sagit Shushan, Lior Gorodisky, Netta Reshef, Yael Holzman, Liron Pinchover, Yaara Endevelt-Shapira, Eva Mishor, Timna Soroka, Maya Finkel, Liav Tagania, Aharon Ravia, Ofer Perl, Edna Furman-Haran, Howard Carp, and Noam Sobel. "unexplained repeated pregnancy loss is associated with altered perceptual and brain responses to men's body-odor". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002717.v1.0.1. +[392] Ian Daly, Nicoletta Nicolaou, Duncan Williams, Faustina Hwang, Alexis Kirke, Eduardo Miranda, and Slawomir J. Nasuto. "a dataset recording joint eeg-fmri during affective music listening". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002725.v1.0.0. +[393] Jordi Sole-Casals, Josep M Serra-Grabulosa, Rafael Romero-Garcia, Gemma Vilaseca, Ana Adan, Nuria Vilaro, Nuria Vilaro, and Edward T. Bullmore. "structural brain network of gifted children". OpenNeuro, 2020. doi: 10.18112/openneuro.d002726.v1.0.1. +[394] Francesca M M Citron, Cristina Cacciari, Jakob M Funcke, Chun-Ting Hsu, and Arthur Jacobs. "emotive idioms". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002727.v1.0.2. +[395] Bi Zhu, Chuansheng Chen, Xuhao Shao, Wenzhi Liu, Zhi-fang Ye, Liping Zhuang, Li Zheng, Elizabeth F. Loftus, and Gui Xue. "multiple interactive memory representations underlie the induction of false memory". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002731.v1.0.2. +[396] Nora C. Heinzelmann, Susanna C. Weber, and Philippe N. Tobler. OpenNeuro, 2020. doi: 10.18112/openneuro.ds002732.v1.0.0. + +[397] Test. "bids pilot project". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002733.v1.0.1. +[398] Andrea Pisauro, Elsa Fouragnan, C Retzler, and M Philiastides. "evidence accumulation in value-based decisions". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002734.v1.0.2. +[399] Jonathan D Power, Benjamin Silver, Melanie Silverman, Eliana Ajodan, Dienke Bos, and Rebecca M. Jones. "headmold". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002735.v1.0.2. +[400] Joset A. Etzel and Todd S. Braver. "multibandcftests". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002737.v1.0.1. +[401] Sabina Gherman and Marios G. Philiastides. "simultaneous eeg-fmri - confidence in perceptual decisions". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002739.v1.0.0. +[402] N Furl, F Begum, J Sulik, F Pizzorni Ferrarese, S Jans, and C Wooley. "animated caricatures fmri study". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002741.v1.0.2. +[403] AR. "test1". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002743.v1.0.1. +[404] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. "resting state with closed eyes for patients with depression and healthy participants". OpenNeuro, 2021. doi: 10.18112/openneuro.d002748.v1.0.5. +[405] Chao-Gan Yan. "yandatabids". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002750.v1.0.1. +[406] Dillan J. Newbold, Timothy O. Laumann, Catherine R. Hoyt, Jacqueline M. Hampton, David F. Montez, Ryan V. Raut, Mario Ortega, Anish Mitra, Ashley N. Nielsen, Derek B. Miller, Babatunde Adeyemo, Annie L. Nguyen, Kristen M. Scheidter, Aaron B. Tanenbaum, Andrew N. Van, Scott Marek, Bradley L. Schlaggar, Alexandre R. Carter, Deanna J. Greene, Evan M. Gordon, Marcus E. Raichle, Steven E. Petersen, Abraham Z. Snyder, and Nico U.F. Dosenbach. "cast-induced plasticity". OpenNeuro, 2020. doi: 10.18112/openneuro.d002766.v3.0.2. +[407] Damien Brevers, Claus Vögele, Pierre Maurage, Guillaume Sescousse, Chris Baeken, and Joel Billieux. "brain mechanisms underlying episodic future thinking of sustainable behaviors". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002770.v2.0.0. +[408] Zhifang Ye and Gui Xue. "retrieval practice facilitates memory updating by enhancing and differentiating medial prefrontal cortex representations". OpenNeuro, 2020. doi: 10.18112/openneuro.d002773.v1.0.0. +[409] Eva Berlot, Nicola Popp, and Jörn Diedrichsen. "motor sequence learning". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002776.v1.0.2. +[410] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. "aomic-piop1". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002785.v2.0.0. +[411] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. "aomic-piop2". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002790.v2.0.0. + +[412] Minjae Kim, Peter Mende-Siedlecki, Stefano Anzellotti, and Liane Young. "tom following strong vs weak priors". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002793.v1.0.1. +[413] Jiefeng Jiang, Anthony D. Wagner, and Tobias Egner. "elife39497". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002797.v1.0.2. +[414] Thompson WH*, Nair R*, Oya H*, Esteban O, Shine JM, Petkov CI, Poldrack RA, Howard M, Adolphs R†, and †corresponding author *equally contributing. "human es-fmri resource: Concurrent deep-brain stimulation and whole-brain functional migraine". OpenNeuro, 2021. doi: 10.18112/openneuro.d002799.v1.0.4. +[415] C.R. Bowman, T. Iwashita, and D. Zeithamova. "model-based fmri reveals co-existing specific and generalized concept representations". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002813.v1.0.0. +[416] Fatemeh Ebrahiminia, Morteza Mahdiani, and Seyed-Mahdi Khaligh-Razavi. "a multimodal neuroimaging dataset to study spatiotemporal dynamics of visual processing in humans". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002814.v1.3.0. +[417] Sangil Lee, Trishala Parthasarathi, and Joseph Kable. "prospection". OpenNeuro, 2021. doi: 10.18112/openneuro.dso02835.v1.0.1. +[418] Sarah Aliko, Jiawen Huang, Florin Gheorghiu, Stefanie Meliss, and Jeremy I Skipper. "naturalistic neuroimaging database". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002837.v2.0.0. +[419] Sarah Schwettmann, Joshua B. Tenenbaum, and Nancy Kanwisher. "intuitive physics with fmri". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002841.v1.0.1. +[420] Omar Maxim. "mriquality". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002848.v1.0.1. +[421] Caroline Landelle, Jean-Luc Anton, Bruno Nazarian, Julien Sein, Ali Gharbi, Olivier Felician, and Anne Kavounoudias. "aging". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002872.v1.3.0. +[422] Christian Roming, Karl Koschutnig, Andreas Fink, and ... "creative direccion making in soccer". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002878.v2.0.0. +[423] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. "brain correlates of multisensory lexical processing in adults". OpenNeuro, 2023. doi: 10.18112/openneuro.ds002879.v1.1.1. +[424] James R. Booth, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, and Jérôme Prado. "brain development of deductive reasoning". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002886.v1.1.0. +[425] Agnès Job, Chloé Jaroszynski, Anne Kavounoudias, Assia Jaillard, and Chantal Delon-Martin. "resting state fmri study of non bothersome tinnitus". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002896.v1.0.0. +[426] Sharna D. Jamadar, Phillip G.D. Ward, Thomas G. Close, Alex Fornito, Malin Premaratne, Kieren O'Brien, Daniel + +Staeb, Zhaolin Chen, N. Jon Shah, and Gary F. Egan. "monash rspet-mr". OpenNeuro, 2024. doi: 10.18112/openneuro.d002898.v1.4.2. +[427] Christophe Pallier and Laurent Bonnasse-Gahot. "simpcomp_rawdata". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002905.v1.0.1. +[428] Rafael Garcia-Dias. "neuroharmony sample example". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002936.v1.0.0. +[429] Bart Aben, Cristian Buc Calderon, Eva Van den Bussche, and Tom Verguts. "task-dependent effort-induced connectivity". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002938.v1.0.1. +[430] DC Monroe, RS Blumenfeld, SL Small, and DB Keator. "head-to-ball impacts in collegiate soccer players". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002940.v1.0.1. +[431] Ruiqing Zhang. "narrative listening". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002979.v1.0.0. +[432] Sangil Lee, Nicole Cooper, and Joseph Kable. "delay discounting bidding". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002989.v1.0.0. +[433] Gabriel Pelletier, Nadav Aridan, Lesley K Fellows, and Tom Schonberg. "fribblesfmri_object-value-construction". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002994.v1.0.3. +[434] Jason A Avery, Alexander G Liu, Cameron Riddell, John E Ingelholm, Stephen J Gotts, and Alex Martin. "taste quality representation in the human brain". OpenNeuro, 2020. doi: 10.18112/openneuro.d002995.v1.0.1. +[435] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. "two sessions of resting state with closed eyes for patients with depression in treatment course (nfb, cbt or no treatment groups)". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003007.v1.0.1. +[436] Colin Hawco, Erin Dickie, Gabrielle Herman, Jessica Turner, Miklos Argyelan, Philipp Homan, Anil Malhotra, Robert Buchanan, and Aristotle Voineskos. "social processes initiative in neurobiology of the schizophrenia(s) traveling human phantoms". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003011.v1.2.3. +[437] M. Martinez-Saito, R. Konovalov, M. A. Piradov, A. Shestakova, B. Gutkin, and V. Klucharev. "action-in-auctions". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003012.v1.0.4. +[438] Amanda LeBel, Lauren Wagner, Shailee Jain, Aneesh Adhikari-Desai, Bhavin Gupta, Allyson Morgenthal, Jerry Tang, Lixiang Xu, and Alexander G. Huth. "an fmri dataset during a passive natural language listening task". OpenNeuro, 2024. doi: 10.18112/openneuro.ms003020.v3.0.0. +[439] Diego Angeles-Valdez, Jalil Rasgado-Toledo, Viviana Villicana, Alan Davalos-Guzman, Alfonso Fajardo-Valdez, Cristina Almanza, Ruth Alcala-Lozano, and Eduardo A. Garza-Villarreal. "sudmex_tms". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003037.v2.1.0. +[440] Karen T. Navarro, Cheryl A. Olman, and Kimberly B. Weldon. "dichoptic magno/parvo". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003043.v1.0.0. + +[441] Krueger O and Klose U. "dti data from 'fiber architecture in the ventromedial striatum and its relation with the bed nucleus of the stria terminalis". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003047.v1.0.0. +[442] Robin Carhart-Harris et al. "neural correlates of the lsd experience revealed by multimodal neuroimaging". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003059.v1.0.0. +[443] James R. Booth, Marisa N. Lytle, Rachna Mureja, and Jérôme Prado. "brain correlates of deductive reasoning in adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003076.v1.0.1. +[444] Philippe DOMENECH, Sylvain RHEIMS, and Etienne KOECHLIN. "probe ieeg". OpenNeuro, 2020. doi: 10.18112/openneuro.ms003078.v1.0.0. +[445] Jonathan Cote and Etienne de Villers-Sidani. "auditory cortex mapping dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003082.v1.0.1. +[446] James R. Booth, Amy S. Desroches, Marisa N. Lytle, Rucha Mehta, Jennifer E. Minas, Rachna Mutreja, Jérôme Prado, and Hongchuan Zhang. "brain correlates of math processing in adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003083.v1.0.1. +[447] Matthew Sachs, Assal Habibi, Antonio Damasio, and Jonas Kaplan. "temporal dynamics of emotional music". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003085.v1.0.0. +[448] Noam Saadon-Grosman, Shahar Arzy, and Yonatan Loewenstein. "somatosensory phase-encoded bilateral full-body light touch stimulation". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003089.v1.0.1. +[449] Jeffrey N. Chiang, Yujia Peng, Hongjing Lu, Keith J. Holyoak, and Martin M. Monti. "analogical reasoning sequential design fmri". OpenNeuro, 2021. doi: 10.18112/openneuro.d003094.v1.0.0. +[450] Y.C. Leong, J. Chen, R. Willer, and J. Zaki. "polarization dataset". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003095.v1.0.0. +[451] D Chung, MA Orloff, N Lauharatanahirun, B King-Casas, and Chiu PH. "valuation of peers' safe choices is associated with substance-naivete in adolescents". OpenNeuro, 2020. doi: 10.18112/openneuro.ps003096.v1.0.2. +[452] Anders Lillevik Thorsen. "abcd protocol for bcbp". Open-Neuro, 2020. doi: 10.18112/openneuro.ms003098.v1.0.0. +[453] HanShin Jo, Chiu-Yueh Chen, Der-Yow Chen, Ming-Heng Weng, and Chun-Chia Kung. "consensus-seeking and conflict-resolving: an fmri study on college couples' shopping interaction". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003103.v1.0.1. +[454] Lauri Parkkonen, Stefan Appelhoff, Alexandre Gramfort, Mainak Jas, and Richard Höchenberger. "mne-somatodata-bids (anonymous)". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003104.v1.0.0. +[455] Jia-Xi Wang, Jin-Ying Zhuang, Lulu Fu, and Lei Qin. "how ovarian hormones influence the behavioral activation and inhibition system through the dopamine pathway". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003114.v1.0.1. +[456] Saloni Krishnan and Kate Watkins. "bold verb generation". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003145.v1.0.2. + +[457] Arafat Angulo-Perkins and Luis Concha. "vocal task". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003146.v1.0.2. +[458] Adam Hampshire and Eyal Soreq. "neuroimaging evidence for network sampling theory of human intelligence". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003148.v1.0.1. +[459] Alexandra Ycaza Herrera, Padideh Nasseri, Ricardo Velasco, Sophia Faude, Jessica D. White, Philipp C. Opitz, Ringo Huang, Kristie Tu, and Mara Mather. "stress-associated brain activation across the hormonal contraceptive cycle". OpenNeuro, 2021. doi: 10.18112/openneuro. ds003151.v2.0.1. +[460] Chang-Hao Kao, Sangil Lee, Joshua I. Gold, and Joseph W. Kable. "changepoint fmri". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s003170.v2.0.0. +[461] Sivayini Kandeepan, Jorge Rudas, Francisco Gomez, Bobby Stojanoski, Sreeram Valluri, Adrian Mark Owen, Lorina Naci, Andrea Soddu, and Emily Sophia Nichols. "modeling an auditory stimulated brain under altered states of consciousness using the generalized ising model". OpenNeuro, 2021. doi: 10.18112/openneuro.d003171.v2.0.1. +[462] Simon M. Ceh, Sonja Annerer-Walcher, Karl Koschutnig, Christof Korner, and Mathias Benedek. "neurophysiological indicators of internal attention: An fmri-eye-tracking coregistration study". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003176.v2.0.1. +[463] Stefano Moia, Eneko Urñuela, Vicente Ferrer, and César Caballero-Gaudes. "euskalibur". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003192.v1.0.1. +[464] Kenshu Koiso, Sebastian Dresbach, Christopher J Wiggins, Omer Faruk Gulban, Anna Katharina Mueller, Miriam Heynckes, Yuhui Chai, Yoichi Miyawaki, Benedikt A Poser, and Renzo Huber. "acquisition and processing methods of whole-brain layer-fmri vaso and bold: The kenshu dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003216.v3.0.11. +[465] Meir Meshulam, Liat Hasenfratz, Hanna Hillman, Yun-Fei Liu, Mai Nguyen, Kenneth A. Norman, and Uri Hasson. "think like an expert". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003233.v1.2.1. +[466] Livia Tomova, Kimberly L. Wang, Todd Thompson, Gillian A. Matthews, Kay M. Tye, and Rebecca Saxe. "mri data of 40 adult participants in response to a cue induced craving task following food fasting, social isolation and baseline (within-subject design)". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003242.v1.0.0. +[467] James Antony, Thomas Hartshorne, Ken Pomeroy, Todd Gureckis, Uri Hasson, Samuel McDougle, and Kenneth Norman. "behavioral, physiological, and neural signatures of surprise during naturalistic sports viewing". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003338.v1.1.0. +[468] Jason A Avery, Alexander G Liu, John E Ingeholm, Stephen J Gotts, and Alex Martin. "tasting pictures: Viewing images of foods evokes taste-quality-specific activity in gustatory insular cortex". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003340.v1.0.4. +[469] Ethan Knights, Courtney Mansfield, Diana Tonin, Janak Saada, Fraser W. Smith, and Stephanie Rossit. "hand- + +selective visual regions represent how to grasp 3d tools for use: brain decoding during real actions". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003342.v1.0.0. +[470] Kelsey McDonald, William Broderick, Scott Huettel, and John Pearson. "penaltykik.02". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003345.v1.0.2. +[471] Eduardo A. Garza-Villarreal, Jorge Julio Gonzalez Olvera, Thania Balducci, Diego Angeles Valdez, Alely Valencia, and Jalil Rasgado. "sudmex Conn: The mexican dataset of cocaine use disorder patients." OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003346.v1.1.2. +[472] Mikkel Schöttner and Andreas Jansen. "empatom with ppi-r, tas, and eq". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003354.v1.0.1. +[473] Richard Huskey, Justin Robert Keene, Shelby Wilcox, Xuanjun (Jason) Gong, Robyn Adams, and Christina J. Najera. "brain dynamics during flow experiences". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003358.v1.0.0. +[474] Samuel B. Snider, Yelena G. Bodien, Aina Frau-Pascual, Marta Bianciardi, Andrea S. Foulkes, and Brian L. Edlow. "ascending arousal network connectivity during recovery from traumatic coma". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003367.v1.0.0. +[475] Heath R Pardoe, Samantha P Martin, Allan George, and Orrin Devinsky. "estimation of in-scanner head pose changes during structural mri using a convolutional neural network trained on eye tracker video". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003381.v1.0.1. +[476] Sharna Jamadar, Shenjun Zhong, Phillip Ward, Alexandra Carey, Richard McIntyre, Alex Fornito, Malin Premaratne, N Jon Shah, Kieran O'Brien, Daniel Stab, Zhaolin Chen, and Gary Egan. "monash vis-fpet-fmri". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003382.v1.5.0. +[477] Nicolas Zilber, Philippe Ciuciu, Alexandre Gramfort, Leila Azizi, and Virginie van Wassenhove. "neurospin hmt+localizer data (meg & amri)". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003392.v1.0.4. +[478] Sharna D Jamadar, Emma X Liang, Phillip GD Ward, Alexandra Carey, Richard McIntyre, Zhaolin Chen, Shenjun Zhong, and Gary F Egan. "monash dacra fpet-fmri". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003397.v1.2.3. +[479] Jungtak Park, Karolina Janacsek, Dezso Nemeth, and Hyeon-Ae Jeon. "asrt (alternating serial reaction time)". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003401.v1.0.1. +[480] Brian P Keane, Deanna M Barch, Ravi Mill, Steven M Silverstein, Bart Krekelberg, and Michael W Cole. "brain network mechanisms of visual shape completion". Open-Neuro, 2025. doi: 10.18112/openneuro.ds003404.v1.0.5. +[481] Maria Geraldine Veldhuizen, Michael C. Farruggia, Xiao Gao, Yuko Nakamura, Barry G. Green, and Dana Small. "identification of an amygdala-thalamic circuit that acts as a central gain mechanism in taste perception". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003424.v1.0.0. +[482] Steven Greening, Tae-Ho Lee, Laurent Gregoire, Lauryn Burleigh, Tyler Robinson, Xinrui Jiang, Mara Mather, and + +Jonas Kaplan. "mental imagery can generate and regulate acquired differential fear conditioned reactivity". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003425.v1.0.2. +[483] Tomoyasu Horikawa and Yukiyasu Kamitani. "attentionally modulated subjective images reconstructed from brain activity". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003430.v1.2.0. +[484] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 localizer". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003433.v1.0.1. +[485] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 main experiment". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003434.v1.0.1. +[486] Sandra Arbula, Elisabetta Pisanu, and Raffaella I. Rumiati. "agreeableness personality trait and social information encoding". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003436.v1.0.0. +[487] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. "protecting the aging brain - diet-study". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003437.v1.0.2. +[488] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant1 suboptimal block designs". OpenNeuro, 2020. doi: 10.18112/openneuro.ms03438.v1.0.0. +[489] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant2 suboptimal slow event related". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003439.v1.0.0. +[490] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant3 caffeine". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003440.v1.0.0. +[491] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusicac. "newbi4fmri2020 variant6 adaptation". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003443.v1.0.0. +[492] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant7 prosopagnosia". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003444.v1.0.0. +[493] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. "newbi4fmri2020 variant8 resting state". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003445.v1.0.0. +[494] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusiac. "newbi4fmri2020 variant9 7t". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003446.v1.0.0. +[495] Joset A. Etzel, Mitch Jeffers, and Todd S. Braver. "dmcc13benchmark". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003452.v1.0.1. +[496] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Lithen, + +Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. "protecting the aging brain, bolus-study". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003453.v1.0.4. +[497] Andrea Stocco, Chantel S. Prat, and Lauren K. Graham. "individual differences in fluid reasoning and rapm-like problem solving". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003454.v1.0.1. +[498] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. "protecting the aging brain, case-study". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003455.v1.0.2. +[499] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., and Todd R. Parrish. "component processes of word reading in adults and children". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003459.v1.0.2. +[500] Todd S. Braver, Alexander Kizhner, Rongxiang Tang, Michael C. Freund, and Joset A. Etzel. "dmcc55b". Open-Neuro, 2025. doi: 10.18112/openneuro.ms003465.v1.0.7. +[501] Jiajia Yang, Peter J. Molfese, Yinghua Yu, Daniel A. Handwerker, Gang Chen, Paul A. Taylor, Yoshimichi Ejima, Jinglong Wu, and Peter A. Bandettini. "haptic three-dimensional curved surface exploration fmri dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003466.v1.1. +[502] Ekaterina V. Pechenkova, Yana R. Panikratova, Maria A. Fomina, Alena D. Rumshiskaya, Darya A. Bazhenova, Liudmila A. Makovskaya, Irina S. Lebedeva, and Valentin E. Sinitsyn. "speech disfluencies: Neurophysiological aspect in normal population". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003469.v1.0.0. +[503] Anja Ischebeck, Hannah Hiebel, Joe Miller, Margit Höfler, Iain D. Gilchrist, and Christof Körner. "target processing in overt serial visual search involves the dorsal attention network: A fixation-based event-related fmri study". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003470.v2.0.0. +[504] J. Rasgado-Toledo, F. Lizcano-Cortés, V. Olalde-Mathieu, M. Zamora-Ursulo, G. Licea-Haquet, A. Carillo-Pena, E. Navarrete, A.* Reyes-Aguilar, and M Giordano. "pragmatic language". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003481.v1.0.3. +[505] Brian Silston, Toby Wise, Song Qi, Xin Sui, Peter Dayan, and Dean Mobbs. "foraging in competitive and hazardous environments". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003484.v1.0.0. +[506] Eva R. Pool, David Munoz Tord, Sylvain Delplanque, Yoann Stussi, Donato Cereghetti, Patrik Vuilleumier, and David Sander. "differential contributions of ventral striatum subregions in the motivational and hedonic components of the affective response to reward". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003487.v2.0.0. +[507] Lukas Snoek and Steven Scholte. "dstreamupgrade". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003495.v1.0.0. + +[508] Kate Nussenbaum and Catherine A. Hartley. "developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03499.v1.0.1. +[509] David Pascucci, Sebastien Tourbier, Joan Rue-Queralt, Margherita Carboni, Patric Hagmann, and Gijs Plomp. "vepcon: Source imaging of high-density visual evoked potentials with multi-scale brain parcellations and connectomes". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003505.v1.1.2. +[510] Chuanji Gao, Christine E. Weber, Douglas H. Wedell, and Svetlana V. Shinkareva. "fmri: Audiovisual valence congruence". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003507.v1.0.1. +[511] Mikael Novén, Hampus Olsson, Gunther Helms, Markus Nilsson, Merle Horne, and Mikael Roll. "language learning aptitude dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003508.v1.0.0. +[512] Adrian W. Gilmore, Alina Quach, Sarah E. Kalinowski, Estefanía I. Gonzalez-Araya, Stephen J Gotts, Daniel L. Schacter, and Alex Martin. "evidence supporting a timelimited hippocampal role in retrieving autobiographical memories". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003511.v1.1.2. +[513] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. "fridaynightlights_study2". OpenNeuro, 2023. doi: 10.18112/openneuro.dso03521.v2.2.0. +[514] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. "fridaynightlights_study1". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03524.v1.0.0. +[515] Benjamin Risk, Raphiel Murden, Junjie Wu, Mary Beth Nebel, Arun Venkataraman, Zhengwu Zhang, and Deqiang Qiu. "which multiband factor should you choose for your resting-state fmri study? the emory multiband dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003540.v1.0.1. +[516] Kshipra Gurunandan, Manuel Carreiras, and Pedro M. Paz-Alonso. "adult language learners". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003542.v1.0.0. +[517] Kshipra Gurunandan, Jaione Arnaez-Telleria, Manuel Carreiras, and Pedro M. Paz-Alonso. "adolescent language learners". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003545.v1.0.0. +[518] Isaac David, Victor Olalde-Mathieu, Ana Y. Martínez, Lluviana Rodríguez-Vidal, and Fernando A. Barrios. "emotion category and face perception task optimized for multivariate pattern analysis". OpenNeuro, 2021. doi: 10.18112/openneuro.ms03548.v1.0.1. +[519] Renee M Visser, H Steven Scholte, Tinka Beemsterboer, and Merel Kindt. "visser, scholte, beemsterboer, & kindt (2013) nature neuroscience". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003550.v1.0.2. +[520] Renee M Visser, H Steven Scholte, and Merel Kindt. "visser, scholte & kindt (2011) journal of neuroscience". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003553.v1.0.2. + +[521] Renee M Visser, Anna E. Kunze, Biance Westhoff, H Steven Scholte, and Merel Kindt. "visser et al. 2015 psychoneuroendocrinology". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003554.v1.0.3. +[522] Falk Luesebrink, Hendrik Mattern, Renat Yakupov, Julio Acosta-Cabronero, Mohammad Ashtarayeh, Steffen Oeltze-Jafra, and Oliver Speck. "data from: Comprehensive ultrahigh resolution whole brain in vivo mri dataset as a human phantom". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003563.v1.1.0. +[523] Lucrezia Liuzzi, Katharine Chang, Hanna Keren, Charles Zheng, Dipta Saha, Dylan Nielson, and Argyris Stringaris. "mood induction in mdd and healthy adolescents". Open-Neuro, 2023. doi: 10.18112/openneuro.ds003568.v1.0.4. +[524] Jeahong Kim, Jeyoung Jung, and Kichun Nam. "transposition confusability during visual word recognition". Open-Neuro, 2021. doi: 10.18112/openneuro.ms003569.v1.0.0. +[525] Virginie Sterpenich, Mojca KM van Schie, Maximilien Catsiyannis, Avinash Ramyead, Stephen Perrig, Hee-Deok Yang, Dimitri Van De Ville, and Sophie Schwartz. "reward biases spontaneous neural reactivation during sleep". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003574.v1.0.2. +[526] R. Nathan Spreng, Roni Setton, Udi Alter, Benjamin N. Cassidy, Bri Darboh, Elizabeth DuPre, Karin Kantarovich, Amber W. Lockrow, Laetitia Mwilambwe-Tshilobo, WenMing Luh, Prantik Kundu, and Gary R. Turner. "neurocognitive aging data release with behavioral, structural, and multi-echo functional mri measures". OpenNeuro, 2022. doi: 10.18112/openneuro.dso03592.v1.0.13. +[527] Xingyu Liu, Yuxuan Dai, Hailun Xie, and Zonglei Zhen. "forrestgump-meg". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003633.v1.0.4. +[528] Heath R Pardoe and Samantha P Martin. "in-scanner head motion and structural covariance networks". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003639.v1.0.0. +[529] Michele Svanera and Lars Muckli. "cerebrum-7t: Fast and fully-volumetric brain segmentation of 7 tesla mr volumes". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003642.v1.1.0. +[530] Jixing Li, John Hale, and Christophe Pallier. "le petit prince: A multilingual fmri corpus using ecological stimuli". OpenNeuro, 2025. doi: 10.18112/openneuro.ds003643.v2.0.7. +[531] Ph.D. Anna Manelis, Ph.D. David A.A. Baranger, B.S. Skye Satz, M.A. Rachel Ragozzino, Ph.D. Satish Iyengar, M.D. Holly A. Swartz, and Ph.D. Yaroslav O. Halchenko. "cortical myelin measured by the t1w/t2w ratio in individuals with depressive disorders and healthy controls". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003653.v1.0.0. +[532] Arielle Tambini, Jacob Miller, Anastasia Kiyonaga, and Mark D'Esposito. "scantrain". OpenNeuro, 2023. doi: 10.18112/openneuro.rs003659.v2.0.4. +[533] Mohamed Abdelhack and Yukiyasu Kamitani. "sharpening of hierarchical visual feature representations of blurred images". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003661.v1.0.0. + +[534] Matan Mazor, Karl J. Friston, and Stephen M. Fleming. "confidence in detection and discrimination". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003669.v1.0.0. +[535] Kangjoo Lee, Corey Horien, David O'Connor, Bronwen Garand-Sheridan, Fuyuze Tokoglu, Dustin Scheinost, Evelyn MR Lake, and Todd R Constable. "yale resting state fmri/pupillometry: Arousal study". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003673.v2.0.1. +[536] Giacomo Ariani, J Andrew Pruszynski, and Jorn Diedrichsen. "finger sequence planning". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003684.v1.0.0. +[537] Julia Berezutskaya, Mariska J. Vansteensel, Erik J. Aarnoutse, Zachary V. Freudenburg, Giovanni Piantoni, Mariana P. Branco, and Nick F. Ramsey. "open multimodal ieeg-fmri dataset from naturalistic stimulation with a short audiovisual film". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003688.v1.0.7. +[538] Javier Ortiz-Tudela, Johanna Bergman, Matt Bennett, Isabelle Ehrich, Lars Muckli, and Yee Lee Shing. "feedbes - feedback signals from episodic and semantic memories." OpenNeuro, 2021. doi: 10.18112/openneuro.ds003691.v1.0.0. +[539] Shan Xu, Xingyu Liu, Jorge Almeida, and Dietmar Heinke. "action-related object pairs - fmri dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003696.v1.0.0. +[540] Takuya Ito, Emily Winfield, Nicole Lalta, Kaustubh R. Kulkarni, Douglas H. Schultz, Ravi D. Mill, Richard H. Chen, Levi Solomyak, and Michael W. Cole. "concrete permuted rule operations". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003701.v1.0.1. +[541] Guo Wanjia, Serra E. Favila, Ghootae Kim, Robert J. Moltor, and Brice A. Kuhl. "abrupt hippocampal remapping signals resolution of memory interference." OpenNeuro, 2021. doi: 10.18112/openneuro.ds003707.v1.0.0. +[542] Hanna Keren, Katharine Chang, Aria Vitale, Chris C. Camp, Robb B Rutledge, Dylan M. Nielson, and Argyris Stringaris. "nimh-compypsych mmi". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03709.v1.0.0. +[543] Daniel K. Bjornn, Julie Van, and C. Brock Kirwan. "the contributionsof eye gaze fixations and target-lure similarity to behavioral and fmri indices of pattern separation and pattern completion". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003711.v1.0.0. +[544] Avijit Chowdhury. "test". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003714.v1.0.1. +[545] Li J., Curley W. H., Guerin B., Dougherty D. D., Dalca A. V., Fischl B., Horn A., and Edlow B. L. "subcortical dmn functional connectivity". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003716.v1.0.0. +[546] Peelle JE, Spehar B, Jones MS, McConkey S, Myerson J, Hale S, Sommers MS, and Tye-Murray N. "visual and audiovisual speech perception associated with increased functional connectivity between sensory and motor regions". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003717.v1.1.0. +[547] Tomoya Nakai, Naoko Koide-Majima, and Shinji Nishimoto. "music genre fmri dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003720.v1.0.1. + +[548] Renee M Visser, Richard N Henson, and Emily A Holmes. "a naturalistic paradigm to investigate post-encoding neural activation patterns in relation to subsequent voluntary and intrusive recall of distressing events". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003721.v1.0.1. +[549] David V. Smith, Rita M. Ludwig, Jeffrey B. Dennison, Crystal Reeck, and Dominic S. Fareri. "an fmri dataset on social reward processing and decision making in younger and older adults". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003745.v2.1.1. +[550] michael barnett, geoffrey aguirre, and david brainard. "If contrast". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003752.v1.0.0. +[551] Kenji Kobayashi, Sangil Lee, Alexandre L. S. Filipowicz, Kara D. McGaughey, Joseph W. Kable, and Matthew R. Nassar. "dynamic representation of the subjective value of information". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003758.v1.0.2. +[552] Olga R. Dobrushina, Larisa A. Dobrynina, Galina A. Arina, Maria V. Gubanova, Anastasia V. Belopasova, Evgenia S. Novikova, Anastasia D. Suslina, Elena I. Kremneva, Ekaterina V. Pechenkova, and Marina V. Krotenkova. "interception during aging: The heartbeat detection task". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003763.v1.0.5. +[553] Zhenzhong Gan, Suiping Wang, and Gangyi Feng. "functional magnetic resonance imaging data for the neural dynamics underlying the acquisition of distinct auditory categories". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003764.v1.0.5. +[554] Yameng Gu, Feng Han, Lucas E. Sainburg, Margeaux M. Schade, and Xiao Liu. "simultaneous eeg and fmri signals during sleep from humans". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003768.v1.0.11. +[555] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. "depression treatment by medial prefrontal real-time fmri neurofeedback". OpenNeuro, 2023. doi: 10.18112/openneuro.ms003770.v1.2.1. +[556] Joseph T. McGuire, Matthew R. Nassar, Joshua I. Gold, and Joseph W. Kable. "functionally dissociable influences on learning rate in a dynamic environment". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003772.v1.0.1. +[557] Redouane Jamil, Franck Mauconduit, Caroline Le Ster, Philipp Ehses, Benedikt A. Poser, Alexandre Vignaud, and Nicolas Boulant. "temporal snr optimization through rf coil combination in fmri: The more, the better? - dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003777.v1.0.1. +[558] Yufei Zhao, Avi Chanales, and Brice Kuhl. "adaptive memory distortions are predicted by feature representations in parietal cortex". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003778.v1.0.0. +[559] Nadav Aridan, Nicholas J. Malecek, Russell A. Poldrack, and Tom Schonberg. "neural correlates of effort-based valuation under risk". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003782.v1.0.1. +[560] Marc Himmelberg, Jan Kurzawski, Noah Benson, Denis Pelli, Marisa Carrasco, and Jon Winawer. "nyu retinotopy + +dataset". OpenNeuro, 2021. doi: 10.18112/openneuro.d003787.v1.0.1. +[561] Xuhao Shao, Chuansheng Chen, Elizabeth F. Loftus, Gui Xue, and Bi Zhu. "dynamic changes in neural representations underlie the repetition effect on false memory". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003789.v2.0.0. +[562] Eva Mishor, Daniel Amir, Tali Weiss, Danielle Honigstein, Aharon Weissbrod, Ethan Livne, Lior Gorodisky, Shiri Karagach, Aharon Ravia, Kobi Snitz, Diyala Karawani, Rotem Zirler, Reut Weissgross, Timna Soroka, Yaara Endevelt-Shapira, Shani Agron, Liron Rozenkrantz, Netta Reshef, Edna Furman-Haran, Heinz Breer, Joerg Strotmann, Tatsuya Uebi, Mamiko Ozaki, and Noam Sobel. "sniffing the human body-volatile hexadecanal blocks aggression in men but triggers aggression in women". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003791.v1.0.0. +[563] Dorit Kliemann, Ralph Adolphs, Tim Armstrong, Paola Galdi, David A. Kahn, Tessa Rusch, A. Zeynep Enkavi, Deuhua Liang, Steven Lograsso, Wenying Zhu, Rona Yu, Remya Nair*, Lynn K. Paul*, J. Michael Tyszka*, and * Joint senior authorship. "caltech conte center - a multimodal data resource for exploring social cognition and decision-making". OpenNeuro, 2021. doi: 10.18112/openneuro.dss003798.v1.0.5. +[564] Hyun Joo Yoo, Kaoru Nashiro, Jungwon Min, Christine Cho, Noah Mercer, Shelby L. Bachman, Padideh Nasseri, Shubir Dutt, Shai Porat, Paul Choi, Yong Zhang, Vardui Grigoryan, Tiantian Feng, Julian F. Thayer, Paul Lehrer, Catie Chang, Jeffrey A. Stanley, Elizabeth Head, Jeremy Rouanet, Vasilis Z. Marmarelis, Shrikanth Narayanan, Jessica Wisnowski, Daniel A. Nation, and Mara Mather. "heart rate variability biofeedback training and emotion regulation". OpenNeuro, 2025. doi: 10.18112/openneuro.d003823.v1.3.5. +[565] Michal Rafal Zareba, Magdalena Fafrowicz, Tadeusz Marek, Ewa Beldzik, Halszka Oginska, and Aleksandra Domagalik. "structural (t1) images of 136 young healthy adults; study of effects of chronotype, sleep quality and daytime sleepiness on brain structure". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003826.v3.0.1. +[566] Kevin Fialkowski, Ivan Messias, and Keith A. Bush. "cognitive control theoretic mechanisms of real-time fmriguided neuromodulation (ctm)". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003831.v1.0.0. +[567] Matteo Visconti di Oleggio Castello, James V. Haxby, and M. Ida Gobbini. "shared neural codes for visual and semantic information about familiar faces in a common representational space". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003834.v1.0.2. +[568] Alyssa H Sinclair and Morgan D Barense. "pe-update". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003835.v1.0.2. +[569] Flavia Mancini, Suyi Zhang, and Ben Seymour. "learning the statistics of pain". OpenNeuro, 2022. doi: 10.18112/openneuro.rs003836.v1.0.0. +[570] van Blooijs D., Demuru M., Zweiphenning W, Hermes D., Leijten F., and Zijlmans M. "dataset clinical epilepsy ieeg + +to bids - respect_longterm_jeeg". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003848.v1.0.3. +[571] Austin L. Boroshok, Anne T. Park, Panagiotis Fotiadis, Gerardo H. Velasquez, Ursula A. Tooley, Katrina R. Simon, Jasmine C.P. Forde, Lourdes M. Delgado Reyes, M. Dylan Tisdall, Dani S. Bassett, Emily A. Cooper, and Allyson P. Mackey. "individual differences in frontoparietal plasticity in humans". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003849.v1.0.0. +[572] S.R. Ashby and D. Zeithamova. "category-biased neural representations form spontaneously during learning that emphasizes memory for specific instances". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003851.v1.0.1. +[573] Leili Mortazavi, Tara Srirangarajan, Tiago Bortolini, Jorge Moll, and Brian Knutson. "monetary incentive delay task". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003858.v1.0.1. +[574] Christopher N. Wahlheim, Alexander P. Christensen, Brittany S. Cassidy, and Zachariah M. Reagh. "resting-state for 34 younger and 28 older adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003871.v1.0.2. +[575] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. "the reading brain project 12 adults". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003872.v1.0.0. +[576] Ted Turesky, Talat Shama, Shahriah Hafiz Kakon, Rashidul Haque, Nazrul Islam, Amala Someshwar, Borjan Gagoski, William Petri, Charles Nelson, and Nadine Gaab. "brain-morphometry diminishedgrowthBeanstudy_2021". OpenNeuro, 2022. doi: 10.18112/openneuro.d003877.v1.1.1. +[577] Wen Wen, Yue Wang, jiawei Zhou, Sheng He, Xinghuai Sun, Hong Liu, Chen Zhao, and Peng Zhang. "loss and enhancement of layer-selective signals in geniculostriate and corticotectal pathways of adult human amblyopia". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003892.v5.0.0. +[578] Philippe Poulin, Guillaume Thead, Pierre-Marc Jodoin, and Maxime Descoteaux. "tractoinferno: A large-scale, open-source, multi-site database for machine learning dmri tractography". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003900.v1.1.1. +[579] J. Pesnot Lerousseau, C. Parise, MO. Ernst, and V. van Wassenhove. "multisensory correlation detector". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003922.v1.0.1. +[580] Ning Mei, Roberto Santana, and David Soto. "informative neural representations of unseen contents during higher-order processing in human brains and deep artificial networks". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003927.v1.0.1. +[581] Tommaso Di Noto, Guillaume Marie, Sebastien Tourbier, Yasser Aleman-Gómez, Oscar Esteban, Guillaume Saliou, Meritxell Bach Cuadra, Patric Hagmann, and Jonas Richiardi. "lausanne_tof-mra_aneurysm_cohort". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003949.v1.0.1. +[582] Lukas Snoek, Jessica Loke, Michelle Kuehn, Mona Zimmermann, and Steven Scholte. "ni-edu-data-minimal". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003965.v1.0.0. + +[583] Olivier Girard, Ludovic de Rochefort, Maxime Guye, Jean-Philippe Ranjeva, and Arnaud Le Troter. "7tamibrain". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003967.v1.0.0. +[584] W. Schellekens, C. Bakker, N.F. Ramsey, and N. Petridou. "body motor". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003972.v1.0.0. +[585] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. "the reading brain project 11 adults". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03974.v3.0.0. +[586] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. "the reading brain project 12 adults". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03988.v1.0.0. +[587] Sarah Khalife, Susan T. Francis, Denis Schluppeck, Rosa Maria Sanchez-Panchuelo, and Julien Besle. "fastvsslowertactile7t". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003990.v1.0.2. +[588] Jun Kai Ho, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. "inter-individual deep image reconstruction". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003993.v1.0.0. +[589] PhD) Daminov V. (MD, MSc) Novak E. (MD, Slepyova N. (MD), Mikhailov D. (MSc), and Karpulevich E. (MSc). "pre-post rehabilitation fmri data of post-stroke patients". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003999.v1.0.2. +[590] J. Wammes, K.A. Norman, and N.B. Turk-Browne. "Increasing stimulus similarity drives nonmonotonic representational change in hippocampus". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s004006.v1.0.4. +[591] J.A. Williams, E.H. Margulis, S.A. Nastase, J. Chen, U. Hasson, K.A. Norman, and C. Baldassano. "high-order areas and auditory cortex both represent the high-level event structure of music". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004007.v1.0.2. +[592] Robert M. Mok and Bradley C. Love. "abstract-category-signal". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004009.v1.0.0. +[593] Dominic Standage, Joseph Y. Nashed, J. Randall Flanagan, and Jason. P Gallivan. "visuomotor rotation adaptation experiment". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004021.v1.0.0. +[594] Julio Cesar Hernandez Pavon, Nils Schneider Garces, John Patrick Begnoche, Lee Miller, and Tommi Rajj. "tmseeg-mri-fmri-dwi data on paired associative stimulation and connectivity (shirley ryan abilitylab, chicago, il)". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004024.v1.0.1. +[595] Cammille C Go, Huseyin O Taskin, Seyed Ahmad Ahmadi, Giulia Frazzetta, Laura Cutler, Saguna Malhotra, Jessica IW Morgan, Virginia L Flanagin, and Geoffrey K Aguirre. "tome-go2022vestibularmri". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04038.v1.0.0. +[596] Hongmi Lee, Janice Chen, and Uri Hasson. "filmfestival". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004042.v1.0.1. + +[597] Sai Ma, Taicheng Huang, Yukun Qu, Xiayu Chen, Yajie Zhang, and Zonglei Zhen. "an fmri dataset for whole-body somatotopic mapping in humans". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004044.v2.0.3. +[598] Miao Zhang, Qikai Qin, Shuning Zhang, Wei Liu, Hongping Meng, Mengyang Xu, Xinyun Huang, Xiaozhu Lin, Mu Lin, Peter Herman, Fahmeed Hyder, Raymond C. Stevens, Zheng Wang, Biao Li, and Garth J. Thompson. "aerobic glycolysis imaging pet-mri". OpenNeuro, 2022. doi: 10.18112/openneuro.d004054.v1.0.0. +[599] Yoshihisa Nakayama, Sho K. Sugawara, Masaki Fukunaga, Yuki H. Hamano, Norihiro Sadato, and Yukio Nishimura. "goal-directed motor task". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004056.v1.0.2. +[600] Shai Sabbah, Michael S. Worden, Dimitrios D. Laniado, David M. Berson, and Jerome N. Sanes. "luminance". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004065.v1.0.0. +[601] Matan Mazor, Chudi gong, and Stephen M. Fleming. "confidence in equal and unequal variance settings". Open-Neuro, 2022. doi: 10.18112/openneuro.ds004081.v1.0.0. +[602] Julie Van, Samuel Nielson, and C. Brock Kirwan. "triple dissociation revisited". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004086.v1.2.0. +[603] Martin Szinte, Gilles de Hollander, Marco Aqil, Serge Dumoulin, and Tomas Knapen. "gaze prf". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004091.v2.0.0. +[604] Athula Pudhiyidath, Neal W Morton, Rodrigo Viveros Duran, Anna C. Schapiro, Ida Momennejad, Demitrius M. Hinojosa-Rowland, Robert J. Molitor, and Alison R. Preston. "effects of temporal community structure learning on reasoning decisions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004094.v1.0.1. +[605] Kristina M. Rapuano, May I. Conley, Anthony C. Juliano, Gregory M. Conan, Maria T. Maza, Kylie Woodman, Steven A. Martinez, Eric Earl, Anders Perrone, Eric Feczko, Damien A. Fair, Richard Watts, BJ Casey, and Monica D. Rosenberg. "an open-access accelerated adult equivalent of the abcd study neuroimaging dataset (a-abcd)". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004097.v1.1.0. +[606] Dobrushina Olga R, Dobrunina Larisa A, Kremneva Elena I, and Pechenkova Ekaterina V. "resting-state fmri data acquired before and after a course of infra-low frequency neurofeedback". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004101.v1.0.1. +[607] Damien Brevers, Chris Baeken, Stefanie De Smet, Beatrix Catoira, Sara De Witte, Qingha He, Pierre Maurage, Laimi Schultze, Guillaume Sescousse, Claudia Vila Verde, Claus Vögele, and Joel Billieux. "stimulation of the dorsolateral prefrontal cortex modulates brain cue reactivity to reward(un)availability". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004102.v1.0.1. +[608] Siao-Shan Shen, Jen-Tang Cheng, I-Jeng Hsu, Der-Yow Chen, Ming-Hung Weng, and Chun-Chia Kung. "collaborations and deceptions in strategic interactions revealed by hyperscanning fmri". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004103.v1.0.0. + +[609] M.P. Weisend, F.M. Hanlon, R. Montano, S.P. Ahlfors, A.C. Leuthold, D. Pantazis, J.C. Mosher, A.P. Georgopoulos, M.S. Hamalainen, and C.J. Aine. "mind data". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004107.v1.0.0. +[610] Yannan Zhu and Shaozheng Qin. "emotional learning retroactively promotes memory integration through rapid neural reactivation and reorganization". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004109.v1.0.0. +[611] Kenji Kobayashi, Joseph W. Kable, Ming Hsu, and Adrianna C. Jenkins. "neural representations of others' traits predict social decisions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004128.v1.0.0. +[612] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. "bids phenotype segregation example dataset". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004129.v1.0.0. +[613] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. "bids phenotype aggregation example dataset". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004130.v1.0.0. +[614] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. "bids phenotype external example dataset". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004131.v1.0.1. +[615] Sepehr Mortaheb, Laurens Van Calster, Federico Raimondo, Manousos A. Klados, Paradeisios Alexandros Boulakis, Kleio Georgoula, Steve Majerus, Dimitri Van De Ville, and Athena Demertzi. "experience sampling in resting state". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004134.v1.0.1. +[616] Bruno Direito, Marco Simões, Alexandre Sayal, and Miguel Castelo-Branco. "targeting dynamic facial processing mechanisms in superior temporal sulcus using fmri neurofeedback". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004141.v1.0.5. +[617] Bruno Direito, Manuel Ramos, Alexandre Sayal, Joao Pereira, and Miguel Castelo-Branco. "exploring the neural correlates of feedback-related reward saliency and valence during real-time fmri-based neurofeedback". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004142.v1.0.1. +[618] Thania Balducci, Jalil Rascado-Toledo, Alely Valencia, Maria-Jose van Tol, Andre Aleman, and Eduardo A. Garza-Villarreal. "a behavioral, clinical and brain imaging dataset with focus on emotion regulation of females with fibromyalgia". OpenNeuro, 2022. doi: 10.18112/openneuro. ds004144.v1.0.2. +[619] Martin Szinte, Anna Montagnini, and Guillaume Masson. "rest_eye". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004158.v2.0.5. +[620] Lachlan T. Strike, Gabriella A.M. Blokland, Narelle K. Hansell, Nicholas G. Martin, Arthur W. Toga, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. "queensland twin imaging (qtim)". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004169.v1.0.7. + +[621] Ádám Nárai, Petra Hermann, Tibor Auer, Péter Kemenczky, János Szalma, István Homolya, Eszter Somogyi, Pál Vakli, Béla Weiss, and Zoltán Vidnyanszky. "movement-related artefacts (mr-art) dataset". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004173.v1.0.2. +[622] Stefanie Meliss, Cristina Pascua-Martin, Jeremy Skipper, and Kou Murayama. "magic, memory, and curiosity (mmc) fmri dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004182.v1.0.1. +[623] Irem Yildirim, Khan Hekmatyar, and Keith A. Schneider. "lgn layers data". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004187.v1.0.2. +[624] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. "things-fmri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004192.v1.0.7. +[625] Iris Groen, Kenichi Yuasa, Amber Brands, Giovanni Piantoni, Stephanie Montenegro, Adeen Flinker, Sasha Devore, Orrin Devinsky, Werner Doyle, Patricia Dugan, Daniel Friedman, Nick Ramsey, Natalia Petridou, and Jonathan Winawer. "visual ecog dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004194.v2.0.0. +[626] Foteini Liwicki, Vibha Gupta, Rajkumar Saini, Kanjar De, Nosheen Abid, Sumit Rakesh, Scott Wellington, Holly Wilson, Marcus Liwicki, and Johan Eriksson. "bimodal dataset on inner speech". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004196.v2.0.2. +[627] Fabiane Schuch, Lennart Walger, Matthias Schmitz, Bastian David, Tobias Bauer, Antonia Harms, Laura Fischbach, Freya Schulte, Martin Schidlowski, Johannes Reiter, Felix Bitzer, Randi von Wrede, Attila Racz, Tobias Baumgartner, Valeri Borger, Matthias Schneider, Achim Flender, Albert Becker, Hartmut Vatter, Bernd Weber, Louisa Specht-Riemenschneider, Alexander Radbruch, Rainer Surges, and Theodor Rüber. "an open presurgery mri dataset of people with epilepsy and focal cortical dysplasia type ii". OpenNeuro, 2025. doi: 10.18112/openneuro.d004199.v1.0.6. +[628] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. "things-meg". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004212.v2.0.1. +[629] Hermes D and Wandell BA. "brainbeats". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004213.v1.0.1. +[630] Mark A. Thornton, Elyssa Barrick, Nathan Liang, and Diana I. Tamir. "interacting representations of mental states and traits". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004217.v1.0.0. +[631] Kyoko Ohashi, Martin H. Teicher, Elizabeth Bolger, Laura C. Hernandez Garcia, Poopak Hafezi, Leslie P. Weiser, Cynthia E. McGreenery, and Alaptagin Khan. "lite-book_Alertness_study". OpenNeuro, 2022. doi: 10.18112/openneuro.d004219.v1.0.0. +[632] Mark A. Thornton and Diana I. Tamir."summed social cognition: additive neural representations of situations, mental states, and actions". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004226.v1.0.0. + +[633] Anne T. Park, Hilary Richardson, Ursula A. Tooley, Cassidy L. McDermott, Austin L. Boroshok, Adrian Ke, Julia A. Leonard, M. Dylan Tisdall, Kirby Deater-Deckard, J. Christopher Edgar, and Allyson P. Mackey. "early stressful experiences are associated with reduced neural responses to naturalistic emotional and social content in children". OpenNeuro, 2022. doi: 10.18112/openneuro.d004228.v1.0.1. +[634] Min-Jeong Kim, Jae-Hoon Lee, Fernanda Juarez Anaya, Jinsoo Hong, William Miller, Sanjay Telu, Prachi Singh, Michelle Y Cortes, Katharine Henry, George L Tye, Michael P Frankland, Jose A Montero Santamaria, Jeih-San Liow, Sami S Zoghbi, Masahiro Fujita, Victor W Pike, and Robert B Innis. "first-in-human evaluation of [11c]ps13, a novel pet radioligand, to quantify cyclooxygenase-1 in the brain". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004230.v3.0.0. +[635] Miguel B. Garcia, Gilles de Hollander, Marcus Grueschow, Rafael Polania, Michael Woodford, and Christian C. Ruff. "individual risk attitudes arise from noise in neurocognitive magnitude representations". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004259.v1.0.0. +[636] Xuhao Shao, Ao Li, Chuangsheng Chen, Elizabeth F. Loftus, and Bi Zhu. "cross-stage neural pattern similarity in the hippocampus predicts false memory derived from post-event inaccurate information". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004261.v2.0.0. +[637] Thirza Dado, Yaghmur Güçlütürk, Luca Ambrogioni, Gabriëlle Ras, Sander Bosch, Marcel van Gerven, and Umut Güçlü. "hyper". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04280.v1.0.1. +[638] Krista Bond, Javier Rasero, Raghav Madan, Jyotika Bahuguna, Jonathan Rubin, and Timothy Verstynen. "neuroloki". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004283.v1.0.3. +[639] Chad S. Rogers, Michael S. Jones, Sarah McConkey, and Jonathan E. Peelle. "listening task". OpenNeuro, 2022. doi: 10.18112/openneuro.rs004285.v1.0.0. +[640] Wilma A. Bainbridge and Chris I. Baker. "multidimensional memory topography in the medial parietal cortex identified from neuroimaging of thousands of daily memory videos". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004286.v1.0.0. +[641] Shaonan Wang, Yunhao Zhang, Xiaohan Zhang, Jingyuan Sun, Nan Lin, Jiajun Zhang, Chengqing Zong, and Shaonan Wang. "an fmri dataset for concept representation with semantic feature annotations". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004301.v1.0.2. +[642] Joan Soler-Vidal, Paola Fuentes-Claramonte, Pilar Salgado-Pineda, Nuria Ramiro, María Ángeles García-León, María Llanos Torres, Antonio Arevalo, Amalia Guerrero-Pedraza, Josep Munuera, Salvador Sarro, Raymond Salvador, Wolfram Hinzen, Peter McKenna, and Edith Pomarol-Clotet. "brain correlates of speech perception in schizophrenia patients with and without auditory hallucinations". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004302.v1.0.1. + +[643] Jason A Avery, Madeline Carrington, and Alex Martin. "a common neural code for representing imagined and inferred tastes". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004312.v1.0.3. +[644] Momchil S. Tomov, Pedro A. Tsividis, Thomas Pouncy, Joshua B. Tenenbaum, and Samuel J. Gershman. "atari-style video game learning fmri". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004323.v1.0.0. +[645] Emily E. Perszyk, Jessica Trinh, Xue S. Davis, Zach Hutelin, Leonie Koban, Maria G. Veldhuizen, Jelena Djordjevic, Marilyn Jones-Gotman, Tor Wager, Hedy Kober, and Dana M. Small. "odour-imagery ability is linked to food craving, intake, and adiposity change in humans". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004327.v1.0.3. +[646] Johannes J.D. Singer, Radoslaw M. Cichy, and Martin N. Hebart. "the spatiotemporal neural dynamics of object recognition for natural images and line drawings (fmri)". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004331.v1. 0.4. +[647] Melanie Ganz and Hannah Eichhorn. "datasets with and without deliberate head movements for evaluating the performance of markerless prospective motion correction and selective reacquisition in a general clinical protocol for brain mri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004332.v1.3.0. +[648] Loris Naspi, Paul Hoffman, and Alexa Morcom. "semantic_encoding". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004341.v1.0.0. +[649] Oscar Ferrante, Ling Liu, Tamas Minarik, Urszula Gorska, Tara Ghafari, Huan Luo, and Ole Jensen. "flux: A pipeline for meg analysis". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004346.v1.0.8. +[650] L.E. Frank and D. Zeithamova. "evaluating methods for measuring background connectivity in slow event-related fmri designs". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004349.v1.0.0. +[651] Asieh Zadbood, Samuel A. Nastase, Janice Chen, Kenneth A. Norman, and Uri Hasson. "sixthsense". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004359.v1.0.0. +[652] Korey P. Wylie, Benzi M. Kluger, Luis D. Medina, Samantha K. Holden, Eugene Kronberg, Jason R. Tregellas, and Isabelle Guard. "parkinson's disease, functional connectivity, and cognition". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004392.v1.0.0. +[653] Lauryn Burleigh and Steven Greening. "fear in the mind's eye: the neural correlates of differential fear acquisition to imagined conditioned stimuli". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004393.v1.0.4. +[654] Doug Greve, Paul Wighton, Melanie Ganz, Martin Nørgaard, Paolo Zanotti-Fregonara, Cyril Pernet, Anthony Galassi, Adam Thomas, Nafiseh Ghazanfari, Jeih-San Liow, Granville Matheson, Gitte Knudsen, and Robert Innis. "a pet molecular imaging brain atlas of cyclooxygenase-1 (kim 2021)". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004401.v1.3.0. +[655] Corey Fernandez, Jiefeng Jiang, Shao-Fang Wang, Hannah L. Choi, and Anthony D. Wagner. "rid". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004406.v1.0.0. + +[656] Marc Himmelberg, Ekin Tuncok, Jesse Gomez, Kalanit Grill-Spector, Marisa Carrasco, and Jonathan Winawer. "stanford child and adult checkerboard retinotopy dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004440.v1.0.1. +[657] Puti Wen, Michael S. Landy, and Bas Rokers. "identifying cortical areas that underlie the transformation from 2d retinal to 3d head-centric motion signals". OpenNeuro, 2023. doi: 10.18112/openneuro.d004443.v1.0.0. +[658] Will be updated. "xcp_walkthrough". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004450.v1.0.1. +[659] Carolina Feher da Silva, Gaia Lombardi, Micah Edelson, and Todd A. Hare. "magic_carpet". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004455.v1.1.0. +[660] Agata Wolna and Zofia Wodniecka. "bilingual speech production with functional localizers". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004456.v1.0.1. +[661] Renzo Huber. "highresolution development". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004458.v1.0.2. +[662] Chris Racey, Charlotte Rae, Ivor Simpson, Oscar Bowen-Hill, Mathilde Bauer, Julia Simner, Christina Kampoureli, and Jamie Ward. "mapping the connectome of synaesthesia: An open access mri dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004466.v1.0.2. +[663] Halie Olson, Emily Chen, Kirsten Lydic, and Rebecca Saxe. "mri data from 20 adults in response to videos of dialogue and monologue from sesame street". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004467.v1.0.0. +[664] Raul Rodríguez-Cruces, Vicente Camacho-Téllez, Alfonso Fajardo, and Luis Concha. "temporal lobe epilepsy - unam". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004469.v1.1.4. +[665] Jonathan C. Lau, Yiming Xiao, Roy A. M. Haast, Greydon Gilmore, Kamil Uludag, Keith W. MacDougall, Ravi S. Menon, Andrew G. Parrent, Terry M. Peters, and Ali R. Khan. "stereotactic neurosurgery dataset (snsx)". OpenNeuro, 2023. doi: 10.18112/openneuro.d004470.v1.0.1. +[666] Mohamad Abbass, Greydon Gilmore, Alaa Taha, Ryan Chevalier, Magdalena Jach, Terry M. Peters, Ali R. Khan, and Jonathan C. Lau. "London heath sciences center parkinson's disease dataset (lhscpd)". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004471.v1.0.1. +[667] Alexander P. Rockhill, Alessandra Mantovani, Brittany Stedelin, Admed M. Raslan, and Nicole C. Swann. "seeg forced two-choice task". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004473.v1.0.2. +[668] Noelle A. Jacobsen and Daniel P. Ferris. "mobile eeg splitbelt walking study". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004475.v1.0.3. +[669] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Nicole Tacugue, Zenia Valdiviezo, Laura D. Lewis, and Zachary Diamandis. "neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004478.v1.0.2. +[670] Samuel Planton*, Fosca Al Roumi*, Liping Wang, and Stanislas Dehaene. "abseqfmri". OpenNeuro, 2023. doi: 10.18112/openneuro.dso004482.v1.0.0. + +[671] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. "neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004484.v1.0.1. +[672] Ming Zhou, Zhengxin Gong, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. "a large-scale fmri dataset for human action recognition". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004488.v1.1.1. +[673] Zhengxin Gong, Ming Zhou, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. "a large-scale fmri dataset for the visual processing of naturalistic scenes". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004496.v2.1.2. +[674] Ph.D. Evan Gordon and Ph.D. Timothy Laumann. "perinatal stroke". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004498.v1.0.0. +[675] Mikl Michal, Kovarova Anezka, Gajdos Martin, Marecek Radek, Novakova Marie, and Slavicek Tomas. "multi-echo simultaneous multislice fmri dataset: Effect of acquisition parameters on fmri data". OpenNeuro, 2024. doi: 10.18112/openneuro.ps004499.v1.0.3. +[676] Amanda Studnicki and Daniel P. Ferris. "real world table tennis". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004505.v1.0.4. +[677] Gabriel Castrillon, Samira Epp, Antonia Bose, Laura Fraticelli, André Hechler, Roman Belenya, Andreas Ranft, Igor Yakushev, Lukas Utz, Lalith Sundar, Josef P Rauschecker, Christine Preibisch, Katarzyna Kurcyus, and Valentin Riedl. "the energetic costs of the human connectome". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004513.v1.0.4. +[678] Clara Sava-Segal, Chandler Richards, Megan Leung, Daniel A. Handwerker, Javier Gonzalez-Castillo, Peter J. Molfese, Peter A Bandettini, and Emily S. Finn. "avid: Audio-visual individual differences dataset". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004516.v2.0.2. +[679] Nestor Zaragoza-Jimenez, Hauke Niehaus, Ina Thome, Christoph Vogelbacher, Gabriele Ende, Inge Kamp-Becker, Dominik Endres, and Andreas Jansen. "modeling face recognition in the predictive coding framework: A combined computational modeling and functional imaging study - stage 2". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004529.v1.1.1. +[680] Elizabeth Valles-Capetillo, Magda Giordano, and Cristian Ibarra. "ironia vev". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004533.v1.0.0. +[681] Sebastian Dresbach, Renzo Huber, Omer Faruk Gulban, and Rainer Goebel. "event-related vaso". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004539.v1.1.1. +[682] Emalie McMahon, Michael F. Bonner, and Leyla Isik. "social interaction dyads". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004542.v1.0.0. +[683] Paul Black, C. Brock Kirwan, Thomas Meservy, William Tayler, and Jeff Williams. "an fmri investigation of the neurocognitive processing of strategies and measures". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004544.v1.0.0. + +[684] Michael Rojek-Giffin, Mael Lebreton, Andrea Farina, Jorg Gross, Jean Daunizeau, and Carsten de Dreu. "learning rules of engagement for social exchange within and between groups". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004553.v1.0.1. +[685] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. "the role of superstition of cognitive control during neurofeedback training - part 1". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004556.v1.0.1. +[686] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. "the role of superstition of cognitive control during neurofeedback training - part 2". OpenNeuro, 2023. doi: 10.18112/openneuro. ds004557.v1.1.0. +[687] M. Rebsamen, D. Romascano, M. Capiglioni, R. Wiest, P. Radojewski, and C. Rummel. "the phantom of bern: repeated scans of two volunteers with eight different combinations of mr sequence parameters". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004560.v1.0.1. +[688] Youngjo Song, Wooree Shin, Pyeongsoo Kim, and Jaeseung Jeong. "mvpa for motor adaptation". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004562.v1.0.4. +[689] Paul Broca and Carl Wernicke. "demo". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004564.v1.0.1. +[690] Ludmila Midrigan-Ciochina, Kayla P. Vodacek, Srinija Balabhadra, and David P. Corina. "brain differences in monolingual and highly proficient multilingual speakers". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004581.v2.0.1. +[691] Yingying Wang, Hongmi Lee, and Brice Kuhl. "ds004590". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004590.v1.0.0. +[692] Hayoung Song, Won Mok Shim, and Monica D. Rosenberg. "song dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004592.v1.0.1. +[693] Marc N. Coutanche, Jake Sauter, Essang Akpan, Rae Buckser, August Vincent, and M. Kathleen Caulfield. "brain surface fingerprints". OpenNeuro, 2023. doi: 10. 18112/openneuro.ds004594.v1.0.1. +[694] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, Bruno Nazarian, Julien Sein, and Jean-Luc Anton. "gloups". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004597.v2.0.0. +[695] Antonin Rovai, Valentina Lolli, Nicola Trotta, Serge Goldman, and Xavier De Tiège. "cerebrovascular reactivity normative dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004604.v2.0.0. +[696] Cameron C. McKay, Brooke Scheinberg, Ellie P. Xu, Katharina Kircanski, Daniel S. Pine, Melissa A. Brotman, Ellen Leibenluft, and Julia O. Linke. "emotion and development branch phenotyping and dti (2012-2017)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004605.v1.0.1. +[697] David Romascano, Gian Franco Piredda, Samuele Caneschi, Tom Hilbert, Ricardo Corredor, Bénédicte Maréchal, Tobias Kober, Jean-Baptiste Ledoux, Eleonora + +Fornari, Patric Hagmann, and Solange Denervaud. "developmental relaxometry 2023". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004611.v1.0.2. +[698] Dylan M. Nielson, Neda Sadeghi, Lisa S. Gorham, Lillian Eisner, Jeremy Taigman, Katherine Haynes, Karen Qi, Christopher C. Camp, Payton Fors, Diana Rodriguez, Jerry McGuire, Erin Garth, Chana Engel, Mollie Davis, Kenneth Towbin, and Argyris Stringaris. "characterization and treatment of adolescent depression (cat-d)". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004627.v1.1.0. +[699] Mengfan Zhang, Alexander Karner, Kathrin Kostorz, Sophia Shea, David Steyrl, Filip Melinscak, Ronald Sladky, Cindy Sumaly Lor, and Frank Scharnowski. "spida-mrineuroimaging dataset". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004630.v1.1.3. +[700] A. De Soares, T. Kim, F. Mugisho, E. Zhu, A. Lin, C. Zheng, and C. Baldassano. "top-down attention impacts neural dynamics during narratives with overlapping event scripts". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004631.v1.0.0. +[701] P. G. Bissett, I. W. Eisenberg, S. Shim, J. A. H. Rios, H. M. Jones, M. P. Hagen, A. Z. Enkavi, J. K. Li, J. A. Mumford, D. P. MacKinnon, L. A. Marsch, and R. A. Poldrack. "cognitive tasks, anatomical mri, and functional mri data evaluating the construct of self-regulation". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004636.v1.0.4. +[702] Edlow B. L., Olchanyi M. D., Freeman H. J., Li J., Maffei C., Snider S. B., Zollei L., Eugenio Iglesias J., Augustinack J., Bodien Y. G., Haynes R. L, Greve D. N., Diamond B. R., Stevens A., Giacino J. T., Destrieux C., van der Kouwe A., Brown E. N., Folkerth R. D., Fischl B., and Kinney H. C. "sustaining wakefulness: Brainstem connectivity in human consciousness". OpenNeuro, 2024. doi: 10.18112/openneuro.d004640.v1.0.4. +[703] Sydney M. Bailes, Daniel E.P. Gomez, Beverly Setzer, and Laura D. Lewis. "resting-state fmri signals contain spectral signatures of local hemodynamic response timing". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004645.v1.0.0. +[704] Heesun Park, Hoyoung Doh, Eunhwi Lee, Harhim Park, and Woo-Young Ahn. "the neurocognitive role of working memory load when pavlovian motivational control affects instrumental learning". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004647.v1.0.2. +[705] Ana L. Miranda-Angulo, Juan D. Sánchez-lópez, Daniel A. Vargas-Tejada, Juan C. Calderón, Jaime Gallo-Villegas, Juan F. Alzate-Restrepo, Jazmin X. Suarez-Revelo, and Gabriel Castrillón. "sympathovagal quotient and functional connectivity of control networks are related to gut ruminococcaceae abundance in healthy men". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004648.v1.0.0. +[706] Martin Norgaard. "the mn dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004650.v1.0.2. +[707] Kathleen C Schmidt, Inna Loutaev, Thomas V Burlin, Zenaide Quezado, Audrey Thurm, Carrie Sheeler, and Carolyn Beebe Smith. "rates of cerebral protein synthesis in subjects with fragile x syndrome". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004654.v1.0.1. + +[708] Travis D. Masterson, C. Brock Kirwan, Lance E. Davidson, and James D. LeCheminant. "neural reactivity to visual food stimuli is reduced in some areas of the brain during evening hours compared to morning hours: an fmri study in women". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004656.v1.1.0. +[709] Neha A Reddy, Kristina M Zvolanek, and Molly G Bright. "denoising task-correlated head motion from motor-task fmri data with multi-echo ica". OpenNeuro, 2023. doi: 10.18112/openneuro.d004662.v1.1.0. +[710] Fan Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C. Aoki, Jin Hirano, and Yukiyasu Kamitani. "visual illusion reconstruction". OpenNeuro, 2023. doi: 10.18112/openneuro.d004670.v1.0.1. +[711] Evan Layher, Tyler Santander, Puneeth Chakravarthula, Nicole Marinsek, Benjamin O. Turner, Miguel P. Eckstein, and Michael B. Miller. "widespread frontoparietal fmri activity is greatly affected by changes in criterion placement, not discriminability, during recognition memory and visual detection tests". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004692.v1.0.0. +[712] Johannes Singer, Agnessa Karapetian, Martin Hebart, and Radoslaw Cichy. "identifying and characterizing scene representations relevant for categorization behavior". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004693.v1.0.3. +[713] Bari Fuchs, Alaina Pearce, Barbara Rolls, Stephen Wilson, Emma Jane Rose, Charles Geier, Hugh Garavan, and Kathleen Keller. "food and brain study". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004697.v1.0.2. +[714] Kelly Chang, Ione Fine, and Geoffrey M. Boynton. "chn retinotopic mapping dataset". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004698.v2.0.1. +[715] David A Feinberg, Alexander JS Beckett, and An T VU. "next generation mri scanner designed for ultra-high resolution human brain imaging at 7-tesla". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004710.v1.1.0. +[716] Loreen Tisdall and Rui Mata. "agerisk". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004711.v1.0.0. +[717] S. Warrington, A. Torchi, O. Mougin, A. Ntata, J. Campbell, M. Craig, F. Alfaro-Almagro, K.L. Miller, P.S. Morgan, M. Jenkinson, and S.N. Sotiropoulos. "on-harmony: A resource for development and comparison of multi-modal brain 3t mri harmonisation approaches". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004712.v2.0.1. +[718] Elisa Zamboni, Alexis Makin, Marco Bertamini, and Antony Morland. "symmetry & luminance bold fmri". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004715.v1.0.3. +[719] Elena Filimonova, Anton Pashkov, Norayr Borisov, Anton Kalinovsky, and Jamil Rzaev. "utilizing amide proton transfer technique to characterise diffuse gliomas based on who 2021 classification of cns tumors". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004717.v1.0.0. +[720] Mohammad Momenian, Zhengwu Ma, Shuyi Wu, Chengcheng Wang, and Jixing Li. "le petit prince hong kong: Naturalistic fmri and eeg dataset from + +older cantonese speakers". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004718.v1.1.1. +[721] Unknown authors. "descriptive name for this dataset". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004720.v1.0.0. +[722] Marilyn Horta, Rebecca Polk, and Natalie Ebner. "single dose intranasal oxytocin administration: Data from healthy younger and older adults". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004725.v1.0.1. +[723] Shrinivas Bishu, Kathleen C Schmidt, Thomas V Burlin, Michael A Charming, Lisa Horowitz, Tianjiang Huang, Zhong-Hua Liu, Mei Qin, B-K Vuong, Aaron J Unterman, Zengyan Xia, Alan Zametkin, Peter Herscovitch, Zenaide Quezado, and Carolyn B Smith. "effects of propofol anesthesia on rates of cerebral protein synthesis". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004730.v1.0.0. +[724] Dante Picchioni, Kathleen C Schmidt, Kelly K McWhirter, Inna Loutaev, Adriana J Pavletic, Andrew M Speer, Alan J Zametkin, Ning Miao, Shrinivas Bishu, Kate M Turetsky, Anne S Morrow, Jeffrey L Nadel, Brittney C Evans, Diana M Vesselinovitch, Carrie A Sheeler, Thomas J Balkin, and Carolyn B Smith. "rates of cerebral protein synthesis and memory formation during sleep". OpenNeuro, 2023. doi: 10.18112/openneuro.d004731.v1.0.0. +[725] Dante Picchioni, Kathleen C Schmidt, Inna Loutaev, Adriana J Pavletic, Carrie Sheeler, Shrinivas Bishu, Thomas J Balkin, and Carolyn B Smith. "rates of cerebral protein synthesis in stages of sleep". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004733.v1.0.0. +[726] Hamsanandini Radhakrishnan, Chenying Zhao, Valerie J. Sydnor, Erica B. Baller, Philip A. Cook, Damien Fair, Barry Giesbrecht, Bart Larsen, Kristin Murtha, David R. Roalf, Sage Rush-Goebel, Russell Shinohara, M. Dylan Tisdall, Jean Vettel, Scott Grafton, Matthew Cieslak, Theodore Satterthwaite, and Haochang Shou. "cs-dsi". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004737.v2.0.0. +[727] Anne Kavounoudias, Caroline Landelle, Jeremy Dana, Bruno Nazarian, Julien Sein, Jean-Luc Anton, Laurent Pruvost, Frederic Giraud, and Olivier Felician. "audiotact_ya". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004743.v1.0.0. +[728] Joe Bathelt, Joanne Taylor, and Kathy Rastle. "language fmri". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004765.v1.0.0. +[729] Rene Weber, Frederic R. Hopp, Allison Eden, and Kate Lee. "vicarious punishment of moral violations in naturalistic drama narratives predicts cortical synchronization". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004775.v1.1. +[730] Katyucia de Macedo Rodrigues, Emma Ben-Avi, Danielle D. Sliva, Myong sun Choe, Marie Drottar, Ruopeng Wang, Bruce Fischl, Patricia E. Grant, and Lilla Zöllei. "infant freesurfer test subject (cnybch)". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004776.v1.0.0. +[731] A L Williams, S Vicary, and G Orgs. "functional mri dataset for evaluating brain synchrony among dance spectators". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004783.v1.0.1. + +[732] Eddy Cavalli, Valérie Chanoine, and Johannes C. Ziegler. "morphosem". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004786.v1.0.1. +[733] Dylan Nielson, Andre Zugman, Marie Zelenina, and Danny Pine. "nimh meter (multi-echo test-retest)". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004787.v1.1.0. +[734] Fu Yu Kwok, Eric D. Wilkey, Lien Peters, Ellyn Khiu, Rebecca Bull, Kerry Lee, and Daniel Ansari. "kwok et al., 2023, human brain mapping". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004791.v1.0.0. +[735] Hampus Olsson, Jason Michael Millward, Ludger Starke, Thomas Gladytz, Tobias Klein, Jana Fehr, Wei-Cheng Lai, Christoph Lippert, Thoralf Niendorf, and Sonia Waiczies. "simulated motion artifacts based on the (mr-art) dataset (https://openneuro.org/datasets/ds004173/versions/1.0.2)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004795.v1.0.0. +[736] Umit Keles, Julien Dubois, Kevin J. M. Le, J. Michael Tyszka, David A. Kahn, Chrystal M. Reed, Jeffrey M. Chung, Adam N. Mamelak, Ralph Adolphs, and Ueli Rutishauser. "fmri data for: Multimodal single-neuron, intracranial eeg, and fmri brain responses during movie watching in human patients". OpenNeuro, 2024. doi: 10.18112/openneuro.d004798.v1.0.5. +[737] Audrey Mazancieux, Franck Mauconduit, Alexis Amadon, Jan Willem de Gee, Tobias Donner, and Florent Meyniel. "brainstem fmri signaling of surprise across different types of deviant stimuli". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004808.v1.0.0. +[738] Michelle Moerel and Essa Yacoub. "high-res gradient echo epi and 3d grase data of auditory cortex". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004814.v1.0.0. +[739] Agustin Lage-Castellanos, Federico De Martino, Geoff M. Ghose, Omer Faruk Gulban, and Michelle Moerel. "selective attention sharpens population receptive fields in human auditory cortex". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004815.v1.0.1. +[740] Ana F. Palenciano, Carlos González-García, Jan de Houwer, Baptist Lieefoghe, and Marcel Brass. "concurrent response and action effect representations across the somatomotor cortices during novel task preparation." OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004829.v1.0.1. +[741] Michelle Moerel, Federico De Martino, Kamil Ugurbil, Essa Yacoub, and Elia Formisano. "processing of frequency and location in human subcortical auditory structures". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004835.v1.0.0. +[742] Fran López-Caballero, Mark Curtis, Brian Coffman, and Dean Salisbury. "magnetoencephalographic (meg) pitch and duration mismatch negativity (mmn) in first-episode psychosis". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004837.v1.0.0. +[743] Kira Noad, David Watson, and Timothy Andrews. "game of thrones - a naturalistic viewing dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004848.v1.0.1. + +[744] Sean Paulsen and Michael A. Casey. "shanxi enculturation". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004866.v1.0.0. +[745] Nafiseh Ghazanfari, Jeih-San Liow, Min-Jeong Kim, Raven Cureton, Adrian Lee, Carson Knoer, Madeline Jenkins, Jinsoo Hong, Jose A. Montero Santamaria, H. Umesha Shetty, Anthony Galassi, Paul Wighton, Martin Norgaard, Douglas N. Greve, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. [11c]ps13 demonstrates pharmacologically selective and substantial binding to cyclooxygenase-1 (cox-1) in the human brain". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004868.v1.0.4. +[746] Elenor Morgenroth, Stefano Moia, Laura Vilaclara, Raphael Fournier, Michal Muszynski, Maria Ploumitsakou, Marina Almato Bellavista, Patrik Vuilleumier, and Dimitri Van De Ville. "emo-film". OpenNeuro, 2025. doi: 10.18112/openneuro.ps004892.v1.0.1. +[747] Toru Maekawa, Takafumi Sasaoka, Toshio Inui, Alan S. R. Fermin, and Shigeto Yamawaki. "heart rate and insula activity increase in response to music in individuals with high interoceptive sensitivity". OpenNeuro, 2023. doi: 10.18112/openneuro.ps004894.v1.0.0. +[748] Harrison Ritz and Amitai Shenhav. "pact_fmi". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004909.v1.1.0. +[749] Ye Wu, Xiaoming Liu, Yunzhi Huang, Tao Zhou, and Fan Zhang. "an open relaxation-diffusion mri dataset in neurosurgical studies". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004910.v1.0.1. +[750] Alejandra Figueroa-Vargas, Gabriela Valdebenito-Oyarzo, María Paz Martínez-Molina, Francisco Zamorano, and Pablo Billeke. "probability decision-making task with ambiguity". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004917.v1.0.1. +[751] David V. Smith, James Wyngaarden, Cooper J. Sharp, Daniel Sazhin, Ori Zaff, Dominic Fareri, and Johanna Jarcho. "an fmri dataset of social and nonsocial reward processing in young adults". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004920.v1.1. +[752] Lonike K Faes, Agustin Lage-Castellanos, Giancarlo Valente, Zidan Yu, Martijn A. Cloos, Luca Vizioli, Steen Moeller, Essa Yacoub, and Federico De Martino. "predictive processing in the auditory cortex measured at 7t". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004928.v1.0.0. +[753] Shari Liu, Kirsten Lydic, Lingjie Mei, and Rebecca Saxe. "fmri dataset: Violations of psychological and physical expectations in human adult brains". OpenNeuro, 2024. doi: 10.18112/openneuro.dss004934.v1.0.0. +[754] Elise M. Cardinale, Jessica Bezek, Julia O Linke, Olivia Siegal, Gabrielle Freitag, Anni Subar, Khosravi Parmis, Ajitha Mallidi, Olivia Peterson, Isaac Morales, Simone P Haller, Courtney Filippi, Kyunghun Lee, Melissa A. Brotman, Daniel S. Pine, Ellen Leibenluft, and Katharina Kirkanski. "multivariate assessment of inhibitory control in youth: Links with psychopathology and brain function dataset". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005754.v1.1.0. + +[755] Patrick Klepits, Karl Koschutnig, Thomas Zussner, and Andreas Fink. "effects of a seven-week running intervention with moderate intensity on the volume of the hippocampus and depressive symptoms in young men from the general population." OpenNeuro, 2024. doi: 10.18112/openneuro. ds004937.v1.0.1. +[756] Jens Jensen, Andreanna Benitez, Erin Klintworth, Hunter Moss, and Kathryn Thorn. "pydesigner example dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004945.v1.0.0. +[757] Roxane Philips, Chris Baeken, Joel Billieux, James Harris, Pierre Maurage, Ismael Muela, Irem Tugce Oz, Athur Pabst, Guillaume Sescousse, Claus Vogele, and Damien Brevers. "brain mechanisms discriminating enactive mental simulations of running and plogging". OpenNeuro, 2024. doi: 10.18112/openneuro.d004946.v1.0.0. +[758] Marleen Haupt, Monika Graumann, Santani Teng, Carina Kaltenbach, and Radoslaw M. Cichy. "braille letters - fmri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004956.v1.0.1. +[759] Angelique Volfart, Katie McMahon, and Greig de Zubicaray. "a comparison of denoising approaches for spoken word production related artefacts in continuous multiband fmri data". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004957.v1.0.2. +[760] Douglas N. Greve and Bruce Fischl. "the freesurfer maintenance dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004958.v1.0.0. +[761] Corinna Perchtold-Stefan, Christian Rominger, Karl Koschutnig, and Andreas Fink. "truecrime". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004965.v1.0.1. +[762] Ekaterina Torubarova, Caroline Arvidsson, Jonathan Berrebi, Julia Uddén, and André Pereira. "neuroengage". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004996.v2.0.0. +[763] Eric C. Leuthardt, Joshua S. Shimony, Abraham Z. Snyder, Donna Dierker, and Ki Yun Park. "retrospective task/rest fmri data from tumor patients". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005003.v2.0.0. +[764] Neha A Reddy, Rebecca G Clements, and Molly G Bright. "simultaneous cortical, subcortical, and brainstem mapping of sensory activation". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005009.v1.0.0. +[765] Michael I. Demidenko, Edward D. Huntley, and Daniel P. Keating. "adolescent health risk behavior study". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005012.v1.0.3. +[766] David Seminowicz, Shana Burrowes, Alexandra Kearson, Jing Zhang, Samuel Kimmel, Luma Samawi, Andrew Furman, and Michael Keaser. "mbsr". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005016.v1.1.1. +[767] Unknown authors. "pop". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005017.v1.0.2. +[768] Sara Ponticorvo and Fabrizio Esposito. "hearing loss connectome". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005026.v1.0.0. +[769] Michael I. Demidenko, Ryan Klaus, Mary Soules, and Mary M. Heitzeg. "michigan longitudinal study". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005027.v1.0.3. + +[770] Marie Mueckstein, Kai Görgen, Stephan Heinzel, Urs Granacher, A. Michael Rapp, and Christine Stelzel. "modality-based multitasking and practice - fmri". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005038.v1.0.3. +[771] Michael S. Cohen, Qiongwen (Jovie) Cao, and Jean Decety. "political moralization (pmm)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005040.v1.2.0. +[772] Panpan Chen, Chi Zhang, Bao Li, Li Tong, Linyuan Wang, Shuxiao Ma, Long Cao, Ziya Yu, and Bin Yan. "an fmri dataset in response to large-scale short natural dynamic facial expression videos". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005047.v1.0.7. +[773] S.H.P. Collin, R. Kempner, S. Srivatsan, and K.A. Norman. "neural codes track prior events in a narrative and predict subsequent memory for details". OpenNeuro, 2024. doi: 10.18112/openneuro.ms00505.v1.0.0. +[774] Guangyu Zhou, Gregory Lane, Thorsten Kahnt, and Christina Zelano. "dataset2". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005056.v1.0.0. +[775] Jing Wang. "cr dbs". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005063.v1.0.0. +[776] jessica. "ds: ds005069:1.0.0". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005069.v1.0.0. +[777] jessica. "ds: ds005072:1.0.1". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005072.v1.0.1. +[778] Brian P Keane, Bart Krekelberg, Ravi Mill, Steven M Silverstein, Judy L Thompson, Megan R Serody, Deanna M Barch, and Michael W Cole. "the neural basis of visual shape completion in schizophrenia and bipolar disorder". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005073.v1.0.0. +[779] Caroline Landelle, Nawal Kinany, Benjamin De Leener, Nicholas D. Murphy, Ovidiu Lungu, Véronique Marchand-Pauvert, Dimitri Van De Ville, and Julien Doyon. "simultaneous brain and spinal cord at rest: an fmri dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005075.v1.0.1. +[780] Unknown authors. "sequence pilot: multiecho and multiband fmri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005085.v1.0.0. +[781] Hannah Grotzinger, Laura Pritschet, Elle M. Murata, Tyler Santander, Pavel Shapturenka, Caitlin Taylor, and Emily G. Jacobs. "28andhe". OpenNeuro, 2024. doi: 10.18112/openneuro.d05115.v1.2.0. +[782] Charles J. Lynch and Conor Liston. "weill cornell medicine multi-echo (wcm-me) dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005118.v1.0.0. +[783] David V. Smith, Cooper J. Sharp, Abraham Dachs, James Wyngaarden, Daniel Sazhin, Jen Yang, Melanie Kos, Tia Tropea, Ishika Kohli, John A. Clithero, Ingrid Olson, Tania Giovannetti, Dominic Fareri, and Johanna M. Jarcho. "social reward and nonsocial reward processing across the adult lifespan: An interim multi-echo fmri and diffusion dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005123.v1.1.3. +[784] Hannah Tarder-Stoll, Christopher Baldassano*, and Mariam Aly*. "the brain hierarchically represents the past and future during multistep anticipation". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005125.v1.0.0. + +[785] Takahiro Hirao, Mitsuhiro Miyamae, Daisuke Matsuyoshi, Ryuto Inoue, Yuhei Takado, Takayuki Obata, Makoto Higuchi, Naotsugu Tsuchiya, and Makiko Yamada. "a neuroimaging dataset during sequential color qualia similarity judgments". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005126.v1.0.3. +[786] Dante Picchioni, Jeff H Duyn, and Jacco A de Zwart. "amri 16-n-0031 sleep1". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005127.v1.0.4. +[787] Sangil Lee, Runxuan Niu, Lusha Zhu, Andrew Kayser, and Ming Hsu. "deception signaling task". OpenNeuro, 2024. doi: 10.18112/openneuro.ps005128.v1.0.0. +[788] Adam Raikes. "sequence test for multiband dwi and magnetization transfer sequence". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005134.v1.0.0. +[789] Binke Yuan and Junjie Yang. "scnu-mandarin-cantonese-dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005139.v1.0.7. +[790] Ladan Shahshahani, Maedbh B. King, Caroline Netekoven, Rich B. Ivry, and Jorn Diedrichsen. "wmfs". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005148.v1.1.0. +[791] Benjamin Lahner, Kshitij Dwivedi, Polina Iamshchinina, Monika Graumann, Alex Lascelles, Gemma Roig, Alessandro Thomas Gifford, Bowen Pan, SouYoung Jin, N.Apurva Ratan Murty, Kendrick Kay, Radoslaw Cichy*, and Aude Oliva*. "modeling short visual events through the bold moments video fmri dataset and metadata". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005165.v1.0.4. +[792] Andrei Barborica, Felicia Mihai, Laurentiu Tofan, Irina Oane, and Ioana Mindruta. "dataset of intracranial EEG during cortical stimulation evoking visual effects". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005169.v1.0.0. +[793] Tomoyasu Horikawa. "mind captioning". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005191.v1.0.2. +[794] Nicholas G. Cicero, Michaela Klimova, Laura D. Lewis, and Sam Ling. "differential cortical and subcortical visual processing with eyes shut". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005194.v1.1.0. +[795] Jiwoong Park, Hayoung Song, and Won Mok Shim. "narrativepuzzle". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005215.v1.0.0. +[796] Lucas Mahler, Julius Steiglechner, Benjamin Bender, Tobias Lindig, Dana Ramadan, Jonas Bause, Florian Birk, Raphael Heule, Edyta Charyasz, Michael Erb, Vinod Jangir Kumar, Gisela E Hagberg, Pascal Martin, Gabriele Lohmann, and Klaus Scheffler. "ultracortex: Submillimeter ultra-high field 9.4t brain mr image collection and manual cortical segmentations". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005216.v1.1.0. +[797] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. "reinforcement-learning generalization". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005230.v1.0.0. +[798] Kirill A. Fadeev, Ilacai V. Romero Reyes, Dzerassa D. Goiaeva, Tatyana S. Obukhova, Tatyana M. Ovsiannikova, Andrey O. Prokofyev, Anna M. Rytikova, Artyom Y. + +Novikov, Vladimir V. Kozunov, Tatyana A. Stroganova, and Elena V. Orekhova. "perception of vowel sounds in children with autism spectrum disorders and typically developing children (meg/erf study)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005234.v2.1.7. +[799] J. Veillette and H. Nusbaum. "motor-fmri". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005239.v1.0.1. +[800] Taylor Salo, M. Dylan Tisdall, Lia Brodrick, Adam Czernuszenko, David R. Roalf, Sage Rush-Goebel, Nicholas Wellman, Matthew Cieslak, and Theodore D. Satterthwaite. "single-echo/multi-echo comparison pilot". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005250.v1.1.5. +[801] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, and Jean-Michel Badier. "gloups_meg". OpenNeuro, 2025. doi: 10.18112/openneuro.ps005261.v2.0.2. +[802] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. "socal kinesia and incentivization for parkinson's disease (skip): Incentivized reaching". OpenNeuro, 2024. doi: 10.18112/openneuro.d05263.v1.0.0. +[803] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. "socal kinesia and incentivization for parkinson's disease (skip): Ultra-high field functional connectivity". OpenNeuro, 2024. doi: 10.18112/openneuro.d005264.v1.0.0. +[804] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. "socal kinesia and incentivization for parkinson's disease (skip): Approach-avoid". OpenNeuro, 2024. doi: 10.18112/openneuro.d005265.v1.0.0. +[805] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. "socal kinesia and incentivization for parkinson's disease (skip): Active escape". OpenNeuro, 2024. doi: 10.18112/openneuro.d005266.v1.0.0. +[806] Anja Ischebeck, Hannah Kreilinger, Joe Miller, Margit Hoefler, Iain D. Gilchrist, and Christof Koerner. "fixating targets in visual search: The role of dorsal and ventral attention networks in the processing of relevance and rarity". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005267.v1.0.1. +[807] Hsi T. Wei, Farhan B. Faisal, Tamara Beck, Claire Shao, and Jed A. Meltzer. "picture-word interference dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005279.v1.0.3. +[808] Vassiki Chauhan, Krystal McCook, and Alex White. "dataset accompanying "reading reshapes stimulus selectivity in the visual word form area", chauhan, mccook and white (2024)". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005295.v1.0.2. +[809] Laura Pritschet, Caitlin M. Taylor, Daniela Cossio, Joshua Faskowitz, Tyler Santander, Daniel A. Handwerker, Hannah Grotzinger, Evan Layher, Elizabeth R. Chrastil, and + +Emily G. Jacobs. "maternal brain project". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005299.v1.0.0. +[810] unknown author. "neural bases of psychological reactance: An fmri study on dogmatic and suggestive health communication". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005304.v1.0.3. +[811] Robert O. Duncan and Evan A. Owens. "bsc_bids_071724". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005355.v1.0.1. +[812] Unknown authors. "neurogame project". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005357.v1.0.0. +[813] Elizabeth J. Rizor, Viktoriya Babenko, Neil M. Dundon, Renee Beverly-Aylwin, Alexandra Stump, Margaret Hayes, Luna Herschenfeld-Catalan, Emily G. Jacobs, and Scott T. Grafton. "hormone health study (hhs)". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005360.v1.0.0. +[814] Alejandra Figueroa-Vargas, Leonie Kausel, María Paz Martínez-Molina, Francisco Zamorano, Ximena Stecher, Patricia Soto-Icaza, and Pablo Billeke. "neurocovid mri dwi and fmri with reversal learning". OpenNeuro, 2024. doi: 10.18112/openneuro.d005364.v1.0.0. +[815] Amy Kartar, Toru Horinouchi, Balázs Örzsik, Brittany Anderson, Lottie Hall, Duncan Bailey, Sarah Samuel, Nati Beltran, Samira Bouyagoub, Chris Racey, Yoko Nagai, Iris Asllani, Hugo Critchley, and Alessandro Colasanti. "neurobiological substrates of altered state of consciousness induced by high ventilation breathwork". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005365.v1.0.1. +[816] Magnus S. Bom, Annette M.A. Brak, Mathijs A.H.L.L. Raemaekers, Nick F. Ramsey, Mariska J. Vansteensel, and Mariana P. Branco. "large-scale fmri dataset for the design of motor-based brain-computer interfaces". OpenNeuro, 2025. doi: 10.18112/openneuro.ms005366.v2.0.0. +[817] Jessica Archibald, Amy Bouchard, Ralph Noeske, Dikoma Shungu, and Mark Mikkelsen. "test-retest reliability of multi-metabolite edited mrs at 3t using press and slaser". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005371.v1.1.3. +[818] Hyunwoo Gu, Joonwon Lee, Sungje Kim, Jaeseob Lim, Hyang-Jung Lee, Heeseung Lee, Minjin Choe, Dong-Gyu Yoo, Jun Hwan (Joshua) Ryu, Sukbin Lim, and Sang-Hun Lee. "discrimination-estimation task". OpenNeuro, 2024. doi: 10.18112/openneuro.d005381.v1.0.0. +[819] Chema G. Penalver, David Lopez-Garcia, and Marta Becerra Losada. "atexp_fmi". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005386.v1.0.0. +[820] Jeryn Chang, Pamela A McCombe, Robert D Henderson, Diana Zanfirache, Christine C Guo, JingLei Lv, Kelly Garner, Saskia Bollmann, Shyuan T Ngo, Frederik J Steyn, and Thomas B Shaw. "an fmri dataset for appetite neural correlates in people living with motor neuron disease". OpenNeuro, 2024. doi: 10.18112/openneuro.d005401.v1.1.1. +[821] Grant and None. "grant dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005412.v1.0.0. +[822] Alexander P. Rockhill and Ahmed M. Raslan. "numbers". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005415.v1.0.0. + +[823] Antonin Rovai, Valentina Lolli, Xavier De Tiège, Nicola Trotta, and Serge Goldman. "cerebrovascular reactivity: Co2 inhalation and resting state comparisons". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005418.v3.2.0. +[824] Juan Eugenio Iglesias, Adrià Casamitjana, Brain L Edlow, Martina Bocchetta, Dorit Kliemann, Matteo Mancini, Loïc Peter, Eleanor Robinson, Roberto Annunziata, Juri Althonayan, Shauna Crampsie, Emily Blackburn, Benjamin Billot, Alessia Atzeni, Oula Puonti, Yael Balbastre, Peter Schmidt, James Hughes, Jean C Augustinack, Lilla Zöllei, David L Thomas, Catherine Strand, Janice L Holton, and Zane Jaunmuktane. "nextbrain data extension". OpenNeuro, 2024. doi: 10.18112/openneuro.d005422.v1.0.1. +[825] Juan Pablo Franco, Peter Bossaerts, and Carsten Murawski. "the neural dynamics associated with computational complexity". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005427.v1.1.1. +[826] M. R. Ehlers, J. H. Kryklywy, Beukers A. O., Moore S. R., B. J. Forys, Anderson A.K., and R. M. Todd. "valenced tactile information is evoked by neutral visual cues following emotional learning". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005449.v1.0.0. +[827] Emilie C. Sleight, Oscar Esteban, and Frédéric Grouiller. "fmriprep boot camp - 7t training dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005454.v1.0.0. +[828] Tingting Guo, Xuedi Liu, Mo Chen, Yongben Fu, and Taomei Guo. "an fmri dataset for investigating language control and cognitive control in bilinguals". OpenNeuro, 2024. doi: 10.18112/openneuro.dS005455.v1.1.5. +[829] Wangjing Yu, Asieh Zadbood, Avi J. H. Chanales, and Lila Davachi. "priority". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005464.v1.0.1. +[830] Antonin Rovai and Emilie Caspar. "sense of agency in free and coerced moral decision-making among civilians and military personnel". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005469.v2.0.0. +[831] Patrycja Scislewska, Michal Rafal Zareba, Julia Lengier, Aaron Schirmer, Piotr Bebas, and Iwona Szatkowska. "monetary incentive delay task - structural and functional images of 37 men; study of associations between circadian characteristics (eveningness, distinctness) and affective processing". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005479.v1.0.3. +[832] Unknown authors. "neurowater project". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005492.v1.0.0. +[833] Niharika Gajawelli, Cameron Glick, Yinming Sun, Faizan Badami, Amit Etkin, and Manish Saggar. "single-pulse tms fmri". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005498.v2.0.0. +[834] Zoya Surani, Ted Turesky, Eileen Sullivan, Talat Shama, Rashidul Haque, Nazrul Islam, Shahriah Hafiz Kakon, Xi Yu, William Petri, Charles Nelson, and Nadine Gaab. "psychosocial adversity and inhibitory control: an fmri study of children growing up in extreme poverty". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005504.v1.0.0. +[835] Ioannis Ntoumanis, Julia Sheronova, Alina Davydova, Maria Dolgaleva, Iiro P Jäaskeläinen, Vladimir Kosonogov, + +Anna N Shestakova, and Vasily Klucharev. "deciphering the neural responses to a naturalistic persuasive message". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005518.v1.0.1. +[836] Xingfeng Shao, Qinyang Shou, Kimberly Felix, Brandon Ojogho, Xuejuan Jiang, Brian T Gold, Megan M Herting, Eric L Goldwaser, Peter Kochunov, L Elliot Hong, Ioannis Pappas, Meredith Braskie, Hosung Kim, Steven Cen, Kay Jann, and Danny JJ Wang. "dp-pcasl data (cbf, att, bbb kw) from 186 cognitively normal participants (8-92 years)". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005529.v1.0.1. +[837] Viviana Greco, Tamas A. Foldes, Neil A. Harrison, Kevin Murphy, Marta Wawrzuta, Mahmoud E. A. Abdellahi, and Penelope A. Lewis. "depotentiation of emotional reactivity using tmr during rem sleep". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005530.v1.0.9. +[838] Hiroto Q. Yamaguchi, Naoko Koide-Majima, Rieko Kubo, Tomoya Nakai, and Shinji Nishimoto. "narrative movie fmri dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005531.v1.0.0. +[839] Nicholas G. Cicero, Nina E. Fultz, Hongbae Jeong, Stephanie D. Williams, Daniel Gomez, Beverly Setzer, Tracy Warbrick, Manfred Jaschke, Ravij Gupta, Michael Lev, Giorgio Bonmassar, and Laura D. Lewis. "high-quality multimodal mri with simultaneous eeg using conductive ink and polymer-thick film nets". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005533.v1.0.0. +[840] Zsuzsanna Nemecz and Attila Keresztes. "item and spatial pattern separation". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005559.v1.0.1. +[841] María Paz Martínez-Molina, Alejandra Figueroa-Vargas, Francisco Zamorano, and Pablo Billeke. "expectation of conflict stimuli". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005571.v1.0.1. +[842] Jarod L Roland, Carl D Hacker, Abraham Z Snyder, Joshua S Shimony, JOhn M Zempel, David D Limbrick, Matthew D Smyth, and Eric C Leuthardt. "a comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005573.v1.0.0. +[843] Toshiki Okumura, Ikuhiro Kida, Atsushi Yokoi, Shinji Nishimoto, Kazushige Touhara, and Masako Okamoto. "dataset for 7t odor fmri experiment". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005576.v1.0.0. +[844] Alexa Tompary and Lila Davachi. "integration of overlapping sequences emerges with consolidation through mpfc neural ensembles and hippocampal-cortical connectivity". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005581.v1.0.0. +[845] Johannes Schultz, Maria Gädeke, and Tom Willems. "sodec - responsibility fmri experiment dataset". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005588.v1.0.0. +[846] Yaqiong Chai, Hedong Zhang, Carlos Robles, Andrew Shinho Kim, Kyung-Wook Kang, Jiyoung Kim, JungWon Shin, Abigail Trang, Jongmok Ha, Emily Lee, Yeonsil Moon, and Hosung Kim. "precise perivascular space + +segmentation on t2-weighted magnetic resonance imaging from human connectome project-aging". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005595.v1.0.0. +[847] Bolong Wang, XuanXuan Zhang, and Xiangzhen Kong. "a naturalistic fmri dataset in response to public speaking". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005596.v1.1. +[848] EC Washington University School of Medicine Leuthardt, JS Washington University School of Medicine Shimony, AZ Washington University School of Medicine Snyder, D Washington University School of Medicine Dierker, and KY Washington University School of Medicine Park. "retrospective task/rest fmri data from tumor patients - for sora". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005597.v1.0.1. +[849] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. "visuomotor rotation learning and reward-based motor learning". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005598.v1.0.0. +[850] Xitong Chen, Stephanie Leach, Juniper Hollis, Dillan Cellier, and Kai Hwang. "thalamocortical contributions to hierarchical cognitive control". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005600.v1.1.0. +[851] Peter N. Taylor, Yujiang Wang, Callum Simpson, Vytene Janiukstyte, Jonathan Horsley, Karoline Leiberg, Beth Little, Harry Clifford, Sophie Adler, Sjoerd B. Vos, Gavin P Winston, Andrew W McEvoy, Anna Miserocchi, Jane de Tisi, and John S Duncan. "the imaging database for epilepsy and surgery (ideas)". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005602.v1.0.0. +[852] Tim Mitchell and Carl Hacker. "mitchell_hacker_2013". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005603.v1.0.1. +[853] Unknown authors. "v4 crowding". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005604.v1.0.1. +[854] Xuefeng Yan, Fabrice G. Simeon, Jeih-San Liow, Cheryl L. Morse, Susovan Jana, Jose A. Montero Santamaria, Madeline Jenkins, Isabella Piacentino, Anthony Galassi, Sami S. Zoghbi, Victor W. Pike, Robert B. Innis, and Zanotti-Fregonara. "[18f]sf51, a novel 18f-labeled pet radioligand for translocator protein 18kda (tspo) in brain, works well in monkeys but fails in humans". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005619.v1.1.0. \ No newline at end of file diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/images.zip b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..855ee2ea08cd91e1b8de4233c13a80d08eea66a0 --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8a9f6e1304e46b93a5d695ae2fe929fae3290f7b0d3ca2870fbaefff1abad5c +size 420816 diff --git a/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/layout.json b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0f1efb93aea3f939a65b8b11fdcd90fbc66e192c --- /dev/null +++ b/2025/An OpenMind for 3D Medical Vision Self-supervised Learning/layout.json @@ -0,0 +1,36092 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 493, + 121 + ], + "type": "text", + "content": "An OpenMind for 3D medical vision self-supervised learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "spans": [ + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": "Tassilo Wald " + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{*,1,2,3}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Constantin Ulrich" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{*,1,4}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Jonathan Suprijadi" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{*,†}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Sebastian Ziegler" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Michal Nohel" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{\\dagger,6,7}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Robin Peretzke" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Gregor Kohler" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{\\dagger,1}" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "text", + "content": ", Klaus Maier-Hein" + }, + { + "bbox": [ + 144, + 142, + 467, + 185 + ], + "type": "inline_equation", + "content": "^{1,2,3,4,5,8}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 186, + 496, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 186, + 496, + 199 + ], + "spans": [ + { + "bbox": [ + 113, + 186, + 496, + 199 + ], + "type": "text", + "content": "1 Division of Medical Image Computing, German Cancer Research Center (DKFZ), Heidelberg, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "spans": [ + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "type": "text", + "content": " Helmholtz Imaging, DKFZ; " + }, + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 118, + 201, + 490, + 213 + ], + "type": "text", + "content": " Faculty of Mathematics and Computer Science, Heidelberg University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 124, + 215, + 485, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 215, + 485, + 228 + ], + "spans": [ + { + "bbox": [ + 124, + 215, + 485, + 228 + ], + "type": "text", + "content": "4 Medical Faculty, Heidelberg University; 5 National Center for Tumor Diseases (NCT), Heidelberg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 229, + 495, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 229, + 495, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 495, + 241 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 115, + 229, + 495, + 241 + ], + "type": "text", + "content": " Faculty of Electrical Engineering and Communication, Brno University of Technology, Czech Republic" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 130, + 243, + 480, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 243, + 480, + 255 + ], + "spans": [ + { + "bbox": [ + 130, + 243, + 480, + 255 + ], + "type": "text", + "content": "7 University hospital Ostrava, Department of Deputy director for science, research and education" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 168, + 256, + 442, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 256, + 442, + 282 + ], + "spans": [ + { + "bbox": [ + 168, + 256, + 442, + 282 + ], + "type": "inline_equation", + "content": "^{8}" + }, + { + "bbox": [ + 168, + 256, + 442, + 282 + ], + "type": "text", + "content": " Pattern Analysis and Learning Group, Department of Radiation Oncology tassilo.wald@dkfz-heidelberg.de" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 310, + 200, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 310, + 200, + 323 + ], + "spans": [ + { + "bbox": [ + 151, + 310, + 200, + 323 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 335, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 297, + 574 + ], + "type": "text", + "content": "The field of self-supervised learning (SSL) for 3D medical images lacks consistency and standardization. While many methods have been developed, it is impossible to identify the current state-of-the-art, due to i) varying and small pretraining datasets, ii) varying architectures, and iii) being evaluated on differing downstream datasets. In this paper, we bring clarity to this field and lay the foundation for further method advancements through three key contributions: We a) publish the largest publicly available pre-training dataset comprising " + }, + { + "bbox": [ + 54, + 335, + 297, + 574 + ], + "type": "inline_equation", + "content": "114k" + }, + { + "bbox": [ + 54, + 335, + 297, + 574 + ], + "type": "text", + "content": " 3D brain MRI volumes, enabling all practitioners to pre-train on a large-scale dataset. We b) benchmark existing 3D self-supervised learning methods on this dataset for a state-of-the-art CNN and Transformer architecture, clarifying the state of 3D SSL pre-training. Among many findings, we show that pre-trained methods can exceed a strong from-scratch nnU-Net ResEnc-L baseline. Lastly, we c) publish the code of our pre-training and fine-tuning frameworks and provide the pre-trained models created during the benchmarking process to facilitate rapid adoption and reproduction. Available here." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 135, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 677 + ], + "type": "text", + "content": "Self-Supervised Learning (SSL) has emerged as a transformative approach in deep learning, enabling the extraction of robust and general representations from data-rich domains [1-3]. This paradigm has led to significant advancements in areas with abundant data, such as natural language process" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 502 + ], + "type": "text", + "content": "ing and natural image computer vision. In the domain of 3D medical imaging, where annotation is expensive and annotated data scarce, SSL pre-training holds great promise as it can learn transferable representations without requiring extensive manual labeling efforts. However, currently the field of medical image analysis follows two predominant approaches to develop vision models: training models from scratch, often using the nnU-Net framework [4], or adopting supervised pre-training with large, monolithic annotated datasets [5-7] or aggregates of many datasets to increase supervised data scale [8, 9]. While the latter approach indicates a willingness to adopt pre-training paradigms, it raises the question why self-supervised learning is hardly pursued in 3D medical imaging. We believe this lack of SSL method development and adoption can be attributed to two major factors:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "1) A lack of a large, open-source 3D dataset for pretraining. Acquiring large unlabeled datasets for self-supervised pre-training is a non-trivial task. While large studies like the UK-BioBank or the NIH Adolescent Brain Cognitive Development (ABCD) exist and hold " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": ">100\\mathrm{k}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": ">40\\mathrm{k}" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": " 3D Volumes respectively, access to them depends on a positive internal review process requiring an appropriate project proposal. To add to the complexity, datasets are often subject to unique Data Usage Agreements (DUAs) that impose various restrictions on recipients. These rules may include: (i) mandatory acknowledgments, (ii) internal administrative review of manuscripts prior to submission [10-14], (iii) inclusion of a consortium in the author list [12-14], or (iv) requiring the dataset's name to appear in the manuscript title [13]. While acknowledgements are reasonable and exemptions for the remaining restrictions can be requested, these DUAs create legal barriers that hinder re" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 702 + ], + "type": "text", + "content": "*Equal contribution, Author order among the co-first authors may be adjusted for individual use." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 160, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 160, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 160, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 703, + 160, + 712 + ], + "type": "text", + "content": " Work done while at DKFZ." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23839" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 290, + 246 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 290, + 246 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 290, + 246 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 290, + 246 + ], + "type": "image", + "image_path": "cfc5f8ea733febaa604803f077e12569109f4f2e56dcd1acd0f52e22ed76f4e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 299 + ], + "type": "text", + "content": "Figure 1. The OpenMind dataset contains 114k 3D Head-And-Neck volumes of 23 different modalities. It represents the largest openly accessible dataset of 3D medical images currently available." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 411 + ], + "type": "text", + "content": "production or creation of large-scale pre-training datasets. Consequently, the majority of current SSL methods are either developed on large, restricted or proprietary datasets [15-17] or are developed on small-scale publicly available datasets [18-24], leading to the current fragmented landscape of pre-training datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 550 + ], + "type": "text", + "content": "2) A lack of comparability: To adopt existing pre-trained models or conduct one's own SSL pre-training, it is important to know which model or pre-training method is the best. Currently, this is an impossible task, due to: i) Different pre-training dataset selection (as previously highlighted). ii) Different architecture choices. Some methods use CNNs for pre-training [15, 17, 21], some ViT-like Transformers [18, 19], others Swin Transformers [22, 23] and even others hybrid architectures [16, 24]. iii) Insufficient downstream dataset choices, potentially providing unreliable results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "In this paper we address these issues and provide three key contributions: i) We share the largest publicly available 3D dataset, comprising 114k 3D volumes of 23 different brain MRI modalities. ii) We conduct a comprehensive self-supervised learning benchmark leveraging this dataset to assess the performance of state-of-the-art CNN and transformer-based architectures on 15 segmentation and classification downstream datasets. iii) We open-source our pre-training and fine-tuning frameworks as well as the trained model checkpoints to enable rapid reproduction and novel method development in the field of 3D medical image segmentation." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 326, + 102, + 542, + 191 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 95 + ], + "type": "text", + "content": "Table 1. Large medical datasets are often restricted through custom DUAs limiting their overall applicability." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 102, + 542, + 191 + ], + "lines": [ + { + "bbox": [ + 326, + 102, + 542, + 191 + ], + "spans": [ + { + "bbox": [ + 326, + 102, + 542, + 191 + ], + "type": "table", + "html": "
DatasetPatients3D VolumesModalitiesAccess
ABCD [25]11,38541,5802Restricted
PPMI [27]2,0119,5545Restricted
ADNI [28]1,5639,5483Restricted
OASIS31,37615,8727Restricted
OASIS46614,0215Restricted
BraTS21 [29]1,2515,0044CC-BY-4.0
CT-RATE [25]21,30450,1882CC-BY-NC-4.0
OpenMind34,191113,92124CC-BY-4.0
", + "image_path": "bc4f0442f7a72fad7b01d567a142e4cc7bdd430b7f43be27e9f3a4d6719cbf4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 209, + 451, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 451, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 451, + 222 + ], + "type": "text", + "content": "2. The OpenMind Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 361 + ], + "type": "text", + "content": "To provide practitioners access to a large-scale dataset for self-supervised learning, we introduce the OpenMind dataset: a large-scale, easily accessible resource containing 114k Head-and-Neck MRI volumes across various modalities, sourced and curated from the OpenNeuro platform. As the largest freely accessible 3D brain imaging dataset to date, OpenMind surpasses the NIH Adolescent Brain Cognitive Decline initiative [25] by a factor of three, and exceeds the recent CT-Rate dataset [26] by a factor of two, while being released under a permissive CC-BY license (see Tab. 1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 368, + 413, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 413, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 413, + 379 + ], + "type": "text", + "content": "2.1. Dataset creation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 385, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 696 + ], + "type": "text", + "content": "All data was sourced or derived from datasets available on the OpenNeuro platform [30]. OpenNeuro (formerly known as openfMRI) is a publicly accessible data repository designed to support neurological research, adhering to the FAIR principles (Findable, Accessible, Interoperable, Reusable). The platform currently hosts over 1,200 public datasets, encompassing data from more than 51,300 subjects. Due to the diverse nature of its collection, OpenNeuro includes a wide array of modalities, such as MRI, PET, fMRI, MEG, EEG, and iEEG, from studies involving both healthy and diseased subjects, enabling research across various aspects of neurology. From this extensive collection, all available 3D MRIs, such as T1-weighted and T2-weighted scans, as well as 4D diffusion-weighted MRIs were collected from 800 different studies (citations [55-854]). This aggregation resulted in a total of 71k 3D MRI scans1 and 15k 4D diffusion-weighted images from a total of 34,139 subjects. While existing 3D SSL methods could be directly trained on the 71k 3D MRI scans spanning multiple MRI sequences, the 4D diffusion-weighted images required pre-processing to create 3D derivatives which we describe in the following section. Overall, due to the high amount of independent studies, the OpenMind dataset provides great levels of diversity regarding participant age distribution, countries of origin, MRI modalities, scanner manufacturer, model and scanning protocols." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 470, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 470, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 470, + 714 + ], + "type": "text", + "content": "This includes a minority of 653 PET scans." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23840" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "DWI Preprocessing Diffusion-weighted imaging (DWI) is an advanced MRI modality that captures the random motion of water molecules in tissue by applying diffusion-sensitive gradient pulses, enabling the assessment of microstructural properties. It is widely used for detecting abnormalities such as strokes, tumors, and white matter changes. In clinical and research settings, DWI data is often preprocessed into more interpretable 3D derivatives, including mean diffusion (MD) maps, and fractional anisotropy (FA) maps, which simplify analysis and interpretation. Additionally, T2-weighted images can be derived from DWI scans. Consequently, we pre-processed the DWI images into MD maps, FA maps, and T2-weighted images, resulting in an additional " + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "inline_equation", + "content": "43\\mathrm{k}" + }, + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": " 3D images. The detailed preprocessing pipeline is described in the appendix in B.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 294, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 294, + 422 + ], + "type": "text", + "content": "Defacing Protecting patient privacy is a big concern in the medical domain, hence, the facial region of Head-and-Neck scans are commonly anonymized. Common anonymization techniques include defacing – replacing the volume with a constant value –, blurring everything but the brain or brain extraction, see Figure 2 in the appendix. These augmented regions can interfere with reconstruction-based SSL methods, potentially penalizing models for attempting to reconstruct plausible anatomical features in anonymized regions. Therefore, masks delineating these regions are needed to allow excluding them from the loss computation. Unfortunately, not all datasets provide these masks, despite applying some sort of anonymization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 423, + 294, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 294, + 554 + ], + "type": "text", + "content": "To address this issue, we generate associated anonymization masks and anatomy masks for cases where these are not provided (Figure 2 bottom in the appendix). The anonymization mask identifies areas where anatomical structures have been artificially modified or removed and where no loss should be calculated, whereas anatomy masks indicate where relevant foreground is present, allowing to avoid sampling non-empty patches which may speed up the convergence of the pre-training methods. Both of these masks are created using the model proposed by Nohel et al. [31]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 713 + ], + "type": "text", + "content": "Meta-Data enrichment While OpenNeuro follows the Brain Imaging Data Structure (BIDS) format, metadata organization varies significantly across datasets. Participant demographics (e.g., age, sex, health status) and image acquisition details (e.g., repetition times, scanner types, imaging protocols) are inconsistently reported, creating challenges for systematic analysis. We curated and harmonized all available, disparate metadata into a unified, complementary metadata structure. This standardized framework ensures consistency and enables users to efficiently filter the OpenMind dataset based on specific selection criteria, such as subject characteristics or imaging" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "parameters, enhancing the flexibility of the dataset and simplifying the extraction of relevant sub-populations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 335 + ], + "type": "text", + "content": "In addition to the harmonized metadata, we provide an Image Quality Score (IQS) for each imaging modality of each of the 800 datasets. This score represents a subjective usability score of the images for self-supervised pretraining. The assessment considers factors such as image noise levels, motion artifacts, and other image quality-degrading elements, condensing these evaluations into a unified score ranging from 1 to 5 with 1 representing the highest quality images. The quality scores were determined through manual inspection of multiple images from each modality in each dataset and were assessed independently by two raters to ensure reliability (more details in the appendix B.1). Together, the harmonized metadata and Image Quality Score are introduced to test the effects of simple data-filtering techniques on self-supervised methods trained on the OpenMind dataset. If significant effects of pre-training data choice can be measured, the OpenMind datasets would show to have sufficient size to enable data-centric research in the domain of self-supervised learning for 3D medical image analysis, akin to [2, 32, 33]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 347, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 407 + ], + "type": "text", + "content": "Useability To maximize the utility of the OpenMind dataset, we provide the final dataset of 114k 3D images on Hugging Face together with anonymization and anatomy masks, as well as the harmonized metadata including the Image Quality Score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 419, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 502 + ], + "type": "text", + "content": "Why benchmarking matters While the OpenMind dataset addresses the issue of data accessibility and data scale in self-supervised learning for 3D medical imaging, the current state-of-the-art remains unknown. To establish a reference point and provide an initial set of baselines, we conduct a comprehensive benchmarking effort to compare existing 3D SSL methods within a common framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 514, + 471, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 471, + 527 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 471, + 527 + ], + "type": "text", + "content": "3. The OpenMind Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 652 + ], + "type": "text", + "content": "The field of self-supervised learning (SSL) for 3D medical imaging remains fragmented due to inconsistencies in pre-training datasets, architectures, and downstream evaluations. Without standardized benchmarks, it is impossible to fairly compare methods and determine the most effective approaches [15]. The OpenMind Benchmark addresses this challenge by evaluating a broad range of SSL methods under controlled conditions, using a common pre-training dataset, standardized architectures, and diverse downstream tasks, which we detail in this section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 660, + 411, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 660, + 411, + 672 + ], + "spans": [ + { + "bbox": [ + 314, + 660, + 411, + 672 + ], + "type": "text", + "content": "3.1. Standardization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "To establish a standardized and fair evaluation, all self-supervised learning (SSL) methods trained in the benchmark were trained on the OpenMind dataset, ensur" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23841" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "ing that pre-training dataset variability does not influence results. Further, we use two state-of-the-art architectures, the ResEnc-L Residual Encoder U-Net [34] to represent convolutional neural networks (CNNs) and the recent Primus-M Transformer [35], allowing us to analyze SSL effectiveness across different architectural paradigms. Pre-trained models are then evaluated on a diverse set of downstream datasets for segmentation and classification tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 192, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 192, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 296, + 371 + ], + "type": "text", + "content": "Downstream Datasets All downstream datasets used are split 50/50 into a train and test set. Moreover, we group our datasets into development datasets and test datasets. The development datasets serve to optimize SSL method hyperparameters on their train split, should the method not be directly applicable on the OpenMind dataset with original hyperparameters. During this process, the most important method-specific hyperparameters as well as learning rate and weight decay are optimized based on an additional 80/20 split on the training set of the development datasets. Aside from this, the development datasets are used to determine the optimal fine-tuning schedules, hence results of the optimal fine-tuning schedules are reported on the validation split of the dataset (Tab. 4 and in the appendix in table 8)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 491 + ], + "type": "text", + "content": "To avoid potential biases introduced during the optimization process on the development datasets, we withheld additional test datasets, which we left untouched and only evaluated after finishing all method optimizations on the development datasets. Moreover, since datasets tend to be small and targets diverse, results are inherently noisy and may not generalize from one dataset to another. To reduce the chance of over optimizing for one dataset, we include four development datasets as well as eight test segmentation datasets and three test classification datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 154, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 154, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 154, + 533 + ], + "type": "text", + "content": "Development datasets:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 534, + 296, + 651 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 534, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 556 + ], + "type": "text", + "content": "1. Atlas R2.0 (ATL) [36]: Brain scans of stroke patients imaged through T1 weighted MRIs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 558, + 296, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 296, + 604 + ], + "type": "text", + "content": "2. Stanford Brain Metastasis (SBM) [37]: Brain scans of metastasis imaged through pre-contrast T1, contrast-enhanced gradient echo, contrast-enhanced spin echo and a T2 FLAIR sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 605, + 295, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 605, + 295, + 628 + ], + "type": "text", + "content": "3. AMOS22 (AMO) [38]: CT and MRI scans of 15 abdominal organs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 629, + 294, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 629, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 629, + 294, + 651 + ], + "type": "text", + "content": "4. KiTS19 (KIT) [39]: CT images of kidneys and kidney tumors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "While AMOS and KiTS can be considered out-of-distribution due to the pre-training images exclusively depicting the head-and-neck region, these datasets were found to be robust benchmarking datasets in [34], hence were included despite their shift. Further, during develop" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "ment, we only optimized the methods for segmentation, due to it being the most common task in 3D medical image analysis [35]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 120, + 441, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 120, + 441, + 131 + ], + "spans": [ + { + "bbox": [ + 314, + 120, + 441, + 131 + ], + "type": "text", + "content": "Test datasets (Segmentation):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 133, + 553, + 383 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 133, + 553, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 133, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 133, + 553, + 156 + ], + "type": "text", + "content": "1. ISLES (ISL) [40]: DWI " + }, + { + "bbox": [ + 315, + 133, + 553, + 156 + ], + "type": "inline_equation", + "content": "(\\mathrm{b} = 1000)" + }, + { + "bbox": [ + 315, + 133, + 553, + 156 + ], + "type": "text", + "content": " and ADC map brain MRI scans of patients with delineations of stroke lesions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 157, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 157, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 315, + 157, + 553, + 204 + ], + "type": "text", + "content": "2. HNTS-MRG (HNT) [41]: MR images of patients with head-and-neck primary tumors and metastatic lymphnodes pre and mid-treatment of which we use the pretreatment images and labels." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 205, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 205, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 315, + 205, + 553, + 239 + ], + "type": "text", + "content": "3. HaN-Seg OAR (HAN) [42]: CT and MR images of Head-and-Neck organs-at-risk of which we use only the MR images." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 240, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 240, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 553, + 263 + ], + "type": "text", + "content": "4. MS FLAIR (MSF) [43]: FLAIR MRIs of hyperintense multiple-sclerosis lesions with associated delineations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 265, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 265, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 553, + 299 + ], + "type": "text", + "content": "5. ToP-CoW (TPC) [44]: Time-of-Flight (ToF) Magnetic Resonance Angiography (MRA) of the Circle of Willis and associated annotations of the arterial structure." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 300, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 300, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 315, + 300, + 553, + 323 + ], + "type": "text", + "content": "6. Yale Brain Metastasis (YBM) [45]: Pre and post-contrast T1, T2 and FLAIR MRI images of brain metastases." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 324, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 324, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 324, + 553, + 358 + ], + "type": "text", + "content": "7. Cosmos (COS) [46]: Carotid vessel wall segmentations and atherosclerosis diagnosis, with contours of the vessel walls on 3D-VISTA MRI." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 360, + 553, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 360, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 315, + 360, + 553, + 383 + ], + "type": "text", + "content": "8. ACDC (ACD) [47]: Cardiac cine-MRI image of three segmented ventricular structures." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 384, + 440, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 384, + 440, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 384, + 440, + 396 + ], + "type": "text", + "content": "Test datasets (Classification):" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 397, + 553, + 504 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 397, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 397, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 553, + 432 + ], + "type": "text", + "content": "1. MR-Net (MRN) [48]: Knee T1w, T2w and Proton Density (PD) weighted MRIs of ACL or meniscal tears as well as other abnormalities which we classify." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 433, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 553, + 468 + ], + "type": "text", + "content": "2. RSNA-SpineFrac (RSN) [49]: CT Spine fractures with the goal of identifying lumbar spine degenerative conditions. We classify fracture and no-fracture." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 468, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 553, + 504 + ], + "type": "text", + "content": "3. ABIDE (ABI) [50]: T1w MRI images of either healthy controls or subjects with autism, on which we conduct binary classification - autism or not-autism." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "Pre-training configuration To ensure a fair comparison, all SSL methods are trained under comparable pre-training and equal fine-tuning schedules. Specifically, we pre-train all methods for a total of 1000 epochs with 250 steps per epoch each, with each method trained DDP on four 40GB A100s to decrease pre-training time. Furthermore, the pretrained models were fine-tuned for a total amount of 150 or 1000 epochs with 250 steps per epoch and a common batch size of 2 for both architectures following an identical learning rate schedule implemented within the nnU-Net framework [4] for the segmentation downstream datasets. For classification downstream datasets, all models were fine-tuned for 200 dataset epochs within the public Image classification framework, which supports 3D classification. We provide more detailed descriptions of the fine-tuning process and frameworks in the appendix C." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23842" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 162, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 162, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 162, + 85 + ], + "type": "text", + "content": "3.2. Benchmark scope" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 295, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 295, + 125 + ], + "type": "text", + "content": "In the benchmark, three key aspects of SSL are investigated, highlighting the possible research directions the OpenMind dataset and benchmark can enable:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 126, + 295, + 293 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 64, + 126, + 295, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 126, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 126, + 295, + 198 + ], + "type": "text", + "content": "i) SSL method performance: We compare the efficacy of various SSL methods in learning transferable representations to various downstream segmentation and classification datasets when using the same pretraining dataset, using two common architectures and a unified fine-tuning schedule." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 198, + 295, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 198, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 295, + 245 + ], + "type": "text", + "content": "ii) Fine-tuning methodology We evaluate various fine-tuning schedules for both architectures to investigate the importance of a well-configured downstream fine-tuning schedule." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 246, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 246, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 246, + 295, + 293 + ], + "type": "text", + "content": "iii) Data-centric baselines We evaluate the influence of data-filtering on the large OpenMind dataset, to evaluate if data-centric research can be conducted on this scale." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 294, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 628 + ], + "type": "text", + "content": "SSL Methods Various SSL methods native to 3D medical images exist. In the benchmark, we include contrastive methods, like Volume Contrastive (VoCo) [22], a 3D SimCLR adaptation, the pseudo-supervised VolumeFusion (VF) [16], default masked autoencoders (MAE) [3], sparse adaptations of them for CNNs (S3D) [15], masked image modeling (SimMIM) [18] and, the denoising and masked auto-encoder combination ModelsGenesis (MG) [21]. Moreover, we incorporate the SwinUNETR pre-training scheme (SwinUNETR) [23], which integrates masked inpainting, contrastive learning and rotation prediction2. We tried including the 3D geometric-matching-based pre-training method GVSL [20], but the method failed to converge both when using the original implementation and when reimplementing it within our framework, regardless of whether we trained it on our OpenMind dataset or the ABCD-NIH dataset. We provide a brief description of each included SSL method and its used hyperparameters in the appendix D. Methods, which did not feature publicly available Code that could be tested for functionality or integrated into our repository, were excluded, e.g. GLMAE [19]. Moreover, more recent self-supervised methods were given higher priority for inclusion in our benchmark than older methods. Furthermore, prominent 2D methods were excluded, as adapting their configuration from the natural imaging domain to the 3D medical imaging domain would require substantial development effort, e.g. due to large batch size requirements leading to VRAM issues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 684 + ], + "type": "text", + "content": "Fine-tuning Schedules Effective adaptation during fine-tuning is crucial for optimizing performance in medical image segmentation, as it determines how well a pre-trained" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "content": "model can transfer knowledge to a target dataset. We test five different fine-tuning strategies for adaptation to segmentation datasets, aiming to balance maintaining general pre-trained representations with adapting the encoder to specialized representations through varying learning rate (lr) schedules: i) nnU-Net's default training strategy employs a polynomial decay lr with an initial lr of " + }, + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "inline_equation", + "content": "1\\mathrm{e} - 2^3" + }, + { + "bbox": [ + 313, + 72, + 553, + 299 + ], + "type": "text", + "content": ", which we reduce to 1e-3 for the transfer learning setting. ii) The Frozen schedule follows the same lr schedule but restricts training to the decoder. iii) The Warm-Up schedule introduces an initial phase with a linear increase in the lr before transitioning to the default schedule. iv) The Valley scheme prioritizes decoder adaptation by first training it with a linear decreasing lr, followed by a linear warm-up phase for the full network, and then the default schedule. v) Finally, the Sawtooth scheme employs a two-stage warm-up: first for the decoder while keeping the encoder frozen, and then for the whole network, both with a linear increasing lr, before transitioning to the default schedule." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 396 + ], + "type": "text", + "content": "We train the majority of fine-tuning runs for 150 epochs to reduce computational cost. For these, each warm-up stage lasts 15 epochs. For the CNNs we provide additional fine-tuning experiments trained for 1000 epochs, in which the warm-up stage is extended to 50 epochs. For a more detailed description and visualization of this, we defer to the appendix C, where we also provide additional information on the classification fine-tuning." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 558 + ], + "type": "text", + "content": "Dataset Filters When training on a large enough quantity of data, optimizing the dataset becomes an important aspect that can greatly improve the resulting model performance [2, 32, 33]. Given the scale and diversity of the OpenMind dataset, we evaluate four simple filtering schemes to measure if we are at a scale where data filtering can be effective. The filtering schemes implemented are: 1. Filtering by the manually created Image Quality Score (IQS) at three thresholds, and 2. filtering all modalities but T1w, T2w and FLAIR images, the most common downstream dataset modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 590, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 673 + ], + "type": "text", + "content": "Reproducibility While evaluating this benchmark, a lot of pre-trained models were generated. To maximize reproducibility, we provide the checkpoints of all pre-training methods for both architectures. Further, we provide an integration into the nnU-Net framework that holds the provided fine-tuning code, enabling users to quickly adapt the pretrained checkpoints to a desired downstream task." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2We refer to the pre-training scheme as SwinUNETR, not to be confused with the identically named architecture." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "3Learning rate value of the ResEnc-L CNN only. Primus-M learning rate values are provided in the appendix C" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23843" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 135, + 541, + 357 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 2. Segmentation results of tested pre-training methods. CNN and Transformer architectures were pre-trained on the same dataset for the same amount of steps. All methods were fine-tuned with the optimal architecture specific fine-tuning scheme for 150 epochs, unless indicated by suffix " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "1k" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": ". Color ranges are set per-dataset and per-architecture. S3D and SimMIM are only shown for CNN or Transformer as they are architecture specific. nnU-Net def.: Dynamically planned default nnU-Net trained from scratch; 1k: Fine-tuned for 1000 epochs instead of 150 epochs; ID Mean: In distribution mean; OOD Mean: Mean across datasets of non Head-and-Neck datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 135, + 541, + 357 + ], + "lines": [ + { + "bbox": [ + 69, + 135, + 541, + 357 + ], + "spans": [ + { + "bbox": [ + 69, + 135, + 541, + 357 + ], + "type": "table", + "html": "
PT MethodDice Similarity Coefficient (DSC) [%] on ...
Dataset of same anatomical region (ID)Dataset of OOD regionAverage across ...
ATLSBMISLHNTHANMSFTPCYBMCOSACDAMOKITIDOODAll
nnU-Net def. 1k58.7059.9878.4062.9853.3752.1979.5058.4346.1991.1088.0087.2161.0888.7768.00
nnU-Net def.56.0860.4178.2259.3732.2755.8476.7856.7349.3190.7283.8877.6158.3384.0764.77
ResEnc-L (CNN)
Scratch 1k58.2153.4379.1465.7558.2454.9079.9456.1271.5792.0988.7387.4864.1589.4370.47
Scratch57.0254.2978.0963.3056.1155.4776.1854.4265.2091.9785.2484.0362.2387.0868.44
VoCo57.1459.6277.5063.4851.1254.9075.1256.9263.4991.4485.6085.7162.1487.5868.50
SwinUNETR56.0757.2577.4561.6449.4254.8274.6957.0565.6890.5384.9585.5461.5687.0167.92
SimCLR57.1559.7278.0163.3251.5655.6877.7759.1468.2091.7686.0684.8563.4087.5669.44
VF57.4259.8878.1864.3251.6757.4276.1159.3163.9891.5785.3886.2163.1487.7269.29
MG58.0361.5777.5865.1154.6955.2577.1458.6771.2791.7486.3586.1764.3788.0970.30
MAE58.2562.4177.8966.5855.1456.8477.9660.0770.8591.9886.7886.1265.1188.3070.91
S3D58.7664.0978.0565.7452.8156.0878.8159.1866.6692.0186.1686.0164.4688.0670.36
Primus-M (Transformer)
Scratch 1k56.7748.5076.5958.4053.4053.2776.3252.5364.6890.8987.2485.5760.0587.9067.01
Scratch51.5143.2675.2355.3050.6054.0073.3150.3062.1190.9380.1776.7357.2982.6163.62
VoCo46.8034.1573.2951.0647.6452.6465.5244.7552.1687.2665.8170.2152.0074.4357.61
SwinUNETR47.2336.3173.8450.1546.4952.8066.2344.4954.8287.9266.1770.3952.4974.8258.07
SimCLR54.6142.6275.4356.7550.8053.5970.0848.3658.2089.9775.7581.2756.7282.3363.12
VF58.6247.3777.5662.3756.1855.0074.9853.9669.7491.4184.9586.1761.7587.5168.19
MG56.5047.3476.7658.4254.0254.6773.6549.7760.7490.8982.1584.4559.1085.8365.78
MAE61.1656.6777.1266.1257.2456.0278.3154.3572.0292.1687.1686.7464.3488.6970.42
SimMIM60.2851.6877.5362.7656.7455.9177.0052.9070.8791.9886.5785.9262.8588.1669.18
", + "image_path": "af795f1b5fa3355c3156d3c284aa11461b36a5deb2719b895e97c4e81fadb5f4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 70, + 397, + 541, + 497 + ], + "blocks": [ + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 555, + 388 + ], + "type": "text", + "content": "Table 3. Classification results of tested pre-training methods. CNN and Transformer architectures were pre-trained and fine-tuned identically on the three classification datasets. PT Method: pre-training method" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 397, + 541, + 497 + ], + "lines": [ + { + "bbox": [ + 70, + 397, + 541, + 497 + ], + "spans": [ + { + "bbox": [ + 70, + 397, + 541, + 497 + ], + "type": "table", + "html": "
Architecture Metric PT MethodResEnc-L (CNN)Primus-M (Transformer)
Average PrecisionMeanMRNBalanced AccuracyAverage PrecisionMeanMRNBalanced Accuracy
MRNRSNABIRSNABIMeanRSNABIMean
Scratch66.3365.3957.1862.9776.6162.2553.2964.0560.5858.9152.8757.4575.4957.1052.0561.55
VoCo71.9970.3362.2368.1880.3565.5760.1868.7067.9262.0953.2261.0877.6861.6451.2863.53
SwinUNETR70.7270.3762.3267.8079.2566.6059.8868.5864.9860.7956.8460.8776.6159.2255.6063.81
SimCLR68.7473.0259.1766.9878.2767.9657.6667.9667.6760.2655.5961.1876.6959.5254.1563.45
VF68.4666.2258.0164.2378.4363.2855.6065.7766.4161.6658.3762.1476.6760.1456.0664.29
MG68.8571.2159.6966.5979.0966.1457.2467.4968.8661.2452.4160.8477.7359.9754.1363.94
MAE64.5367.8658.3163.5776.7264.2156.2365.7267.0060.2853.0960.1377.2658.8351.8462.64
S3D70.0869.2560.6566.6678.5664.2056.4166.39--------
SimMIM--------65.2759.6253.9459.6177.1558.4953.9363.19
", + "image_path": "1b576ca0b63f887ac914bd98c6fe4ad5495fd32b5640a9ad6919d50f26bd0b2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 108, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 108, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 108, + 529 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 295, + 574 + ], + "type": "text", + "content": "In this section we highlight the most important insights of our benchmark, but not all due to spatial constraints. A comprehensive list of them is provided in the appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 586, + 198, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 198, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 198, + 600 + ], + "type": "text", + "content": "4.1. SSL method performance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Segmentation performance We compare the fine-tuning results (fine-tuned for 150 epochs) of all pre-training methods for the ResEnc-L CNN architecture and the Primus-M Transformer architecture in Tab. 2. For both architectures, we observe that self-supervised pre-training is able to exceed both, a from-scratch nnU-Net default baseline trained for 1000 epochs and their respective from-scratch trained architecture trained for 150 epochs. In particular, reconstruction based methods exhibit the best performance," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "content": "with default MAEs leading the pack. The MAE pre-trained CNN model even exceeds its strong from scratch baseline trained for 1000 epochs in only 150 epochs, and performs even better when fine-tuned for 1000 epochs (see Table 5 in the appendix). Aside from the MAE pre-training paradigm, this is only achieved by the sparse-autoencoder based S3D pre-training scheme with 1000 epochs of fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 601, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 708 + ], + "type": "text", + "content": "Comparing the benefits for each architecture, it can be seen that the Transformer-based Primus-M architecture profits substantially more from pre-training than the ResEnc-L CNN counterpart, however, it shows lower from-scratch results to begin with. Comparing the MAE pre-trained Primus-M to its 1000 epochs from scratch counterpart, it is able to improve about " + }, + { + "bbox": [ + 313, + 601, + 556, + 708 + ], + "type": "inline_equation", + "content": "3.5\\%" + }, + { + "bbox": [ + 313, + 601, + 556, + 708 + ], + "type": "text", + "content": " DSC points and almost exceeds the ResEnc-L 1000 epochs from scratch baseline in 150 epochs. However, on singular datasets the MAE" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23844" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 91, + 287, + 258 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 281, + 82 + ], + "type": "text", + "content": "Table 4. Effects of different CNN fine-tuning schedules." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 91, + 287, + 258 + ], + "lines": [ + { + "bbox": [ + 63, + 91, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 63, + 91, + 287, + 258 + ], + "type": "table", + "html": "
DatasetPT Method FT ScheduleVoCoSwinUNETRSimCLRVFMGMAES3DAverage
SBMDefault73.8669.8672.7972.5773.3771.8475.4772.82
Frozen52.3438.5960.8328.5451.3559.4556.5549.67
Warm-Up70.8769.1673.9473.1174.6774.5173.5472.83
Valley70.2670.2972.8275.8174.6873.7172.4072.85
Sawtooth71.5271.0573.7273.7873.8074.7973.9773.23
ATLDefault62.7860.8256.8063.8362.6262.7862.1661.68
Frozen58.2754.2757.9543.3357.2754.4957.1554.68
Warm-Up59.3260.2362.3663.2063.4663.6162.3462.07
Valley59.4360.3461.2164.0061.2362.9863.3861.80
Sawtooth60.4261.1761.3465.0761.5463.4763.7662.40
AMODefault85.9385.4287.1286.0586.9386.7886.7486.42
Frozen53.1861.2662.9230.9757.8950.8352.4852.79
Warm-Up86.9886.6187.1086.5887.3487.6287.5487.11
Valley85.9185.4686.7286.5587.0987.3786.9286.57
Sawtooth86.7885.7686.7586.4687.1587.4886.4586.69
KITDefault85.8083.8381.2084.3483.6984.5682.2183.66
Frozen69.3369.5965.2236.0671.3963.9771.4163.85
Warm-Up84.7181.7781.8083.5183.9985.2482.2083.32
Valley84.0882.4181.5585.0083.4883.4883.7883.40
Sawtooth85.1283.2683.1085.4184.1584.2883.3184.09
AverageDefault77.0974.9874.4876.7076.6576.4976.6576.15
Frozen58.2855.9361.7334.7359.4857.1859.4055.25
Warm-Up75.4774.4476.3076.6077.3677.7576.4076.33
Valley74.9274.6275.5777.8476.6276.8876.6276.15
Sawtooth75.9675.3176.2377.6876.6677.5076.8776.60
", + "image_path": "ca50a42b2d5bacbedcb438fa7975ce23a7ae5440c3e4a2bd33ce7b0ff107708c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 365 + ], + "type": "text", + "content": "pre-trained Primus-M transformer already achieves the highest performance overall (ATL, COS and ACD), including the best pre-trained ResEnc-L. Given the substantial performance discrepancy observed in Wald et al. [35] between the two architectures, this provides the first sign of vision transformers reaching state-of-the-art performance in 3D medical image segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 555 + ], + "type": "text", + "content": "However, not all pre-training methods are able to learn useful representations for segmentation. VoCo, SwinUNETR pre-training, and SimCLR (for transformers) do not, or just barely exceed, a from-scratch baseline trained for 150 epochs and can even lead to substantial performance reductions for the transformer architecture likely due to their global pre-training task not learning the local information needed for segmentation. Moreover, individual datasets can be negatively impacted by any pre-training, as for the ISL and HAN dataset when using a ResEnc-L architecture. This is particularly surprising as Primus-M shows strong improvements with its MAE pre-training on both of these datasets. When filtering low quality images this seems to be partially alleviated (see Table 6 in the appendix), indicating that CNNs may learn different patterns in the presence of noise, which a transformer may be more robust to." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 676 + ], + "type": "text", + "content": "Lastly, comparing the longer 1000 epoch fine-tuning schedule ( see Table 5 in the appendix) and the shorter 150 epoch fine-tuning schedule of the ResEnc-L, it can be observed that the longer fine-tuning schedule can lead to performance degradation on datasets where pre-training showed strong improvements (SBM, HNT). Simultaneously, it reduces the performance gap to the from-scratch 1000 epoch baseline on cases where pre-training was not as helpful (ISL, HAN, AMO), likely by allowing to forget learned representations and adapt more strongly to the downstream task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Classification performance We evaluate the fine-tuned performance of our pre-trained methods for classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 286 + ], + "type": "text", + "content": "- a more global downstream task - in Tab. 3. Contrastive pre-training methods like VoCo, SwinUNETR, and SimCLR, which were ineffective for segmentation, achieve the highest classification performance. In contrast, MAE pretraining, which was the strongest for segmentation, performs the worst across both architectures and nearly all datasets. Furthermore, segmentation focused methods like S3D and MG do not transfer well to classification, though they do not fall to the lowest ranks. These findings reinforce the distinction between global and local feature learning: contrastive methods specialize in global similarity learning, making them well-suited for classification, whereas reconstruction-based objectives prioritize pixel-wise representations, benefiting segmentation. Interestingly, SwinUNETR pre-training, which incorporated both global and local objectives, still converged toward representations favoring classification over segmentation, suggesting a bias in learned feature hierarchies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 407 + ], + "type": "text", + "content": "Additionally, it is to note that from-scratch classification performance is subpar, with ABI barely exceeding randomness (balanced accuracy of " + }, + { + "bbox": [ + 313, + 287, + 553, + 407 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 287, + 553, + 407 + ], + "type": "text", + "content": "). This may be due to the classification pipeline lacking the refinements of the widely adopted nnU-Net framework, due to the lower prevalence of global reasoning tasks in 3D medical imaging. Thus, while our classification results provide valuable insights, they should be interpreted with caution, particularly compared to our segmentation findings which are more reliable due to the robust benchmarking framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 419, + 455, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 455, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 455, + 433 + ], + "type": "text", + "content": "4.2. Fine-tuning methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 616 + ], + "type": "text", + "content": "The ablation results of different CNN fine-tuning strategies in Tab. 4 indicate that the Sawtooth fine-tuning strategy consistently achieves the best overall performance, followed closely by the Warm-Up scheme. This shows that adapting the encoder, after tuning the newly initialized decoder is crucial to maximize performance. The staged learning rate increases in these schedules likely help to prevent instability and allow for a smoother transition from pre-training to task-specific adaptation. For the out-of-distribution CT datasets, AMO and KIT, the Default fine-tuning schedule performs very competitively. This indicates that models pretrained on MRI data require greater modifications when adapted to CT images. The high initial lr allows the models to shift their learned features more effectively to the new domain." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Keeping the encoder frozen during fine-tuning significantly reduces performance across all datasets and underscores the importance of encoder adaptation. However, SimCLR is less affected by this constraint. Notably, both VoCo and SimCLR perform almost comparably to other fine-tuning schedules when the encoder is frozen on ATL, w.r.t. the drop on other datasets. This dataset is closest to the pretraining data in the development datasets, as it consists of a" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23845" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "single MRI sequence, whereas the other datasets either contain CT scans or multiple MRI sequences as input." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 294, + 262 + ], + "type": "text", + "content": "For the Transformer, fine-tuning results in Table 8 in the appendix reveal that the Warm-Up scheme achieves the best performance across all methods, emphasizing the importance of architecture-specific adaptation. The Frozen schedule leads to an even greater performance drop than for CNNs, as the ResEnc-L decoder, despite being lightweight, has significantly more parameters than the \"decoder\" of Primus-M. In this setting, fine-tuning can be interpreted almost as linear probing, with only back-to-back transposed conv layers (including normalization and activations) being trained, severely limiting decoder capacity. With a frozen encoder, it can be observed that adapting to the ATL dataset reaches the closest performance to other fine-tuning schedules." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 170, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 170, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 170, + 283 + ], + "type": "text", + "content": "4.3. Data-centric results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 491 + ], + "type": "text", + "content": "We compare the effect of data filtering of the pre-training dataset for CNN MAE pre-training and highlight the measured downstream performance in Appendix Table 6. It can be observed that filtering the lowest quality images through the Image Quality Score (IQS) shows minor positive effects on overall downstream segmentation performance, despite removing substantial amounts of pre-training data. Only when removing about " + }, + { + "bbox": [ + 55, + 288, + 295, + 491 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 55, + 288, + 295, + 491 + ], + "type": "text", + "content": ", keeping only the highest quality images, a degradation in performance is measured. Moreover, it can be observed that reducing data diversity by removing all modalities but T1w, T2w and FLAIR images reduces model performance, despite having more pre-training images than the intermediate IQS filtering method which slightly increased performance. Overall, this indicates that the OpenMind dataset does contain redundant data which may be leveraged by more complex data-centric methods to increase performance further." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 635 + ], + "type": "text", + "content": "Effect of anonymization on reconstruction Many images of the OpenMind dataset were defaced or brain-extracted to guarantee patient privacy. To allow taking these artifacts into account when training with a reconstruction objective, we created masks that allow ignoring such regions. To measure the effect of such artifacts, we compare the performance of reconstruction methods that were or were not penalized for the reconstruction in such regions. (See Appendix Table 7). It can be observed that ignoring the reconstruction loss of anonymized regions improves the segmentation performance of the MAE and S3D, showing that taking anonymization into account improves representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 645, + 272, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 272, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 272, + 658 + ], + "type": "text", + "content": "5. Discussion, Limitations and Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Our study presents the OpenMind dataset, the largest publicly available 3D medical imaging dataset, alongside a standardized self-supervised learning (SSL) benchmark for 3D medical image analysis. Through extensive experi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 311 + ], + "type": "text", + "content": "mentation across diverse datasets and across many pretraining paradigms, we demonstrate the superiority of self-supervised pre-training over from-scratch training for 3D medical image analysis, when pre-training on a large-scale dataset. We show that contrastive methods excel in global tasks such as classification, and reconstruction-based approaches perform well in dense tasks such as segmentation. However, we find that no pre-training method of the 3D vision domain generalizes and performs well in both tasks, emphasizing the need for further SSL method development. Moreover, we compared state-of-the-art CNNs and Transformers in our benchmark and show for the very first time that a pre-trained Primus-M Transformer can exceed the strong ResEnc-L CNN from-scratch baseline on many datasets and can exceed the strongest pre-trained ResEnc-L CNN on individual datasets. This marks an important milestone for the domain, as transformers from-scratch so far never exceeded a ResEnc-L model [34, 35, 51]. On average, though, we show that a pre-trained ResEnc-L remains the strongest architecture for now." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 609 + ], + "type": "text", + "content": "Moreover, our ablation results suggest that fine-tuning strategies are a key determinant of the success of pre-trained methods, with bad strategies losing significant performance. This emphasizes the importance of careful adaptation beyond pre-training. Experiments on fine-tuning for longer showed partial degradation of overall performance, indicating the problem of overfitting to downstream tasks. Together, these insights open up research opportunities for developing and adapting Parameter-Efficient Fine-Tuning (PEFT) methods to 3D medical image analysis to address these issues. Moreover, our fine-tuning experiments showed that keeping the vision encoders frozen shows sub-par performance for all pre-training methods. This emphasizes the lack of generality of currently learned representations and – again – highlights the need for better pre-training methods. While our study provides a lot it also has various limitations: i) Our classification results are not as reliable as the segmentation results, due to the lower interest in classification for 3D medical imaging, requiring further refinements. ii) Current evidence of the utility of the OpenMind dataset for data-centric research is rather weak, due to the minimal improvements observed requiring further investigation. iii) We evaluate pre-training for 1000 epochs due to resource limitations. Pre-training for longer may reveal different trends than shown in this study." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 610, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 553, + 693 + ], + "type": "text", + "content": "In summary, by open-sourcing the OpenMind dataset and establishing the OpenMind benchmark, we provide a clear assessment of state-of-the-art SSL in 3D medical imaging and offer a reproducible framework for method development. Through them, we aim to accelerate progress in self-supervised learning for 3D medical images and pave the way for new research directions." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23846" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was partly funded by the Helmholtz Foundation Model Initiative (HFMI) under the subproject 'The Human Radiome Project' (THRP). Part of this work was funded by Helmholtz Imaging (HI), a platform of the Helmholtz Incubator on Information and Data Science. Additionally, a part of the computational resources were provided by the German AI Service Center WestAI." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 207, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 207, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 207, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 66, + 207, + 295, + 274 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619-15629, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 274, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 274, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 66, + 274, + 295, + 330 + ], + "type": "text", + "content": "[2] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 331, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 331, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 66, + 331, + 295, + 386 + ], + "type": "text", + "content": "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 388, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 388, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 66, + 388, + 295, + 432 + ], + "type": "text", + "content": "[4] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods, 18(2):203-211, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 433, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 433, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 66, + 433, + 295, + 499 + ], + "type": "text", + "content": "[5] Jakob Wasserthal, Hanns-Christian Breit, Manfred T Meyer, Maurice Pradella, Daniel Hinck, Alexander W Sauter, Tobias Heye, Daniel T Boll, Joshy Cyriac, Shan Yang, et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence, 5(5), 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 66, + 500, + 295, + 555 + ], + "type": "text", + "content": "[6] Chongyu Qu, Tiezheng Zhang, Hualin Qiao, Yucheng Tang, Alan L Yuille, Zongwei Zhou, et al. Abdomenatlas-8k: Annotating 8,000 ct volumes for multi-organ segmentation in three weeks. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 556, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 556, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 66, + 556, + 295, + 622 + ], + "type": "text", + "content": "[7] Ziyan Huang, Haoyu Wang, Zhongying Deng, Jin Ye, Yanzhou Su, Hui Sun, Junjun He, Yun Gu, Lixu Gu, Shaoting Zhang, et al. Stu-net: Scalable and transferable medical image segmentation models empowered by large-scale supervised pre-training. arXiv preprint arXiv:2304.06716, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 624, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 624, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 66, + 624, + 295, + 690 + ], + "type": "text", + "content": "[8] Constantin Ulrich, Fabian Isensee, Tassilo Wald, Maximilian Zenk, Michael Baumgartner, and Klaus H Maier-Hein. Multilant: A multi-dataset approach to medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 648-658. Springer, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 691, + 295, + 714 + ], + "type": "text", + "content": "[9] Wenxuan Li, Alan Yuille, and Zongwei Zhou. How well do supervised 3d models transfer to medical imaging tasks?" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "In The Twelfth International Conference on Learning Representations, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 320, + 95, + 553, + 128 + ], + "type": "text", + "content": "[10] OASIS. Open access series of imaging studies data use agreement, 2024. URL https://sites.wustl.edu/oasisbrains/." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 172 + ], + "type": "text", + "content": "[11] PPMI. Parkinson's progression markers initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=PPMI." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 553, + 216 + ], + "type": "text", + "content": "[12] ADNI. Alzheimer's disease neuroimaging initiative data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=ADNI." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 553, + 259 + ], + "type": "text", + "content": "[13] MCSA. Mayo clinic study of aging data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project=MCSA." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 261, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 261, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 320, + 261, + 553, + 305 + ], + "type": "text", + "content": "[14] HABS-HD. Health and aging brain study: Health disparities data use agreement, 2024. URL https://ida.loni.usc.edu/collaboration/access/appApply.jsp?project " + }, + { + "bbox": [ + 320, + 261, + 553, + 305 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 320, + 261, + 553, + 305 + ], + "type": "text", + "content": " HABS_HD." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 305, + 553, + 360 + ], + "type": "text", + "content": "[15] Tassilo Wald, Constantin Ulrich, Stanislav Lukyanenko, Andrei Goncharov, Alberto Paderno, Leander Maerkisch, Paul F Jaeger, and Klaus Maier-Hein. Revisiting mae pretraining for 3d medical image segmentation. arXiv preprint arXiv:2410.23132, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 415 + ], + "type": "text", + "content": "[16] Guotai Wang, Jianghao Wu, Xiangde Luo, Xinglong Liu, Kang Li, and Shaoting Zhang. Mis-fm: 3d medical image segmentation using foundation models pretrained on a large-scale unannotated dataset. arXiv preprint arXiv:2306.16925, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "type": "text", + "content": "[17] Asbjørn Munk, Jakob Ambsdorf, Sebastian Llambias, and Mads Nielsen. Amaes: Augmented masked autoencoder pretraining on public brain mri data for 3d-native segmentation. arXiv preprint arXiv:2408.00640, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 514 + ], + "type": "text", + "content": "[18] Zekai Chen, Devansh Agarwal, Kshitij Aggarwal, Wiem Safta, Mariann Micsinai Balan, and Kevin Brown. Masked image modeling advances 3d medical image analysis. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1970-1980, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 515, + 553, + 557 + ], + "type": "text", + "content": "[19] Jia-Xin Zhuang, Luyang Luo, and Hao Chen. Advancing volumetric medical image segmentation via global-local masked autoencoder. arXiv preprint arXiv:2306.08913, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 553, + 624 + ], + "type": "text", + "content": "[20] Yuting He, Guanyu Yang, Rongjun Ge, Yang Chen, Jean-Louis Coatrieux, Boyu Wang, and Shuo Li. Geometric visual similarity learning in 3d medical image self-supervised pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9538-9547, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 658 + ], + "type": "text", + "content": "[21] Zongwei Zhou, Vatsal Sodha, Jiaxuan Pang, Michael B Gotway, and Jianming Liang. Models genesis. Medical image analysis, 67:101840, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 658, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 555, + 714 + ], + "type": "text", + "content": "[22] Linshan Wu, Jiaxin Zhuang, and Hao Chen. Voco: A simple-yet-effective volume contrastive learning framework for 3d medical image analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22873-22882, 2024." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23847" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 139 + ], + "type": "text", + "content": "[23] Yucheng Tang, Dong Yang, Wenqi Li, Holger R Roth, Bennett Landman, Daguang Xu, Vishwesh Nath, and Ali Hatamizadeh. Self-supervised pre-training of swin transformers for 3d medical image analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20730-20740, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 140, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 296, + 183 + ], + "type": "text", + "content": "[24] Fenghe Tang, Ronghao Xu, Qingsong Yao, Xueming Fu, Quan Quan, Heqin Zhu, Zaiyi Liu, and S Kevin Zhou. Hyspark: Hybrid sparse masking for large scale medical image pre-training. arXiv preprint arXiv:2408.05815, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 205 + ], + "type": "text", + "content": "[25] National Institutes of Health et al. Adolescent brain cognitive development study, 2015." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 296, + 281 + ], + "type": "text", + "content": "[26] Ibrahim Ethem Hamamci, Sezgin Er, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Muhammed Furkan Dasdelen, Omer Faruk Durugol, Bastian Wittmann, Tamaz Amiranashvili, et al. Developing generalist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 283, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 295, + 358 + ], + "type": "text", + "content": "[27] Kenneth Marek, Sohini Chowdhury, Andrew Siderowf, Shirley Lasch, Christopher S Coffey, Chelsea Caspell-Garcia, Tanya Simuni, Danna Jennings, Caroline M Tanner, John Q Trojanowski, et al. The parkinson's progression markers initiative (ppmi)-establishing a pd biomarker cohort. Annals of clinical and translational neurology, 5 (12):1460-1477, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 295, + 414 + ], + "type": "text", + "content": "[28] Susanne G Mueller, Michael W Weiner, Leon J Thal, Ronald C Petersen, Clifford Jack, William Jagust, John Q Trojanowski, Arthur W Toga, and Laurel Beckett. The alzheimer's disease neuroimaging initiative. Neuroimaging Clinics, 15(4):869-877, 2005." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 480 + ], + "type": "text", + "content": "[29] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 536 + ], + "type": "text", + "content": "[30] Christopher J Markiewicz, Krzysztof J Gorgolewski, Franklin Feingold, Ross Blair, Yaroslav O Halchenko, Eric Miller, Nell Hardcastle, Joe Wexler, Oscar Esteban, Mathias Goncavles, et al. The openneuro resource for sharing of neuroscience data. *Elife*, 10:e71774, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 295, + 581 + ], + "type": "text", + "content": "[31] Michal Nohel, Constantin Ulrich, Jonathan Suprijadi, Tassilo Wald, and Klaus Maier-Hein. Unified framework for foreground and anonymization area segmentation in ct and mri data. In Bildverarbeitung für die Medizin 2025, 2025." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 582, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 295, + 646 + ], + "type": "text", + "content": "[32] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 691 + ], + "type": "text", + "content": "[33] Amro Abbas, Kushal Tirumala, Daniel Simig, Surya Ganguli, and Ari S Morcos. Semdedup: Data-efficient learning at web-scale through semantic dedduplication. arXiv preprint arXiv:2303.09540, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 713 + ], + "type": "text", + "content": "[34] Fabian Isensee, Tassilo Wald, Constantin Ulrich, Michael Baumgartner, Saikat Roy, Klaus Maier-Hein, and Paul F" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 118 + ], + "type": "text", + "content": "Jaeger. nnu-net revisited: A call for rigorous validation in 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 488-498. Springer, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 118, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 182 + ], + "type": "text", + "content": "[35] Tassilo Wald, Saikat Roy, Fabian Isensee, Constantin Ulrich, Sebastian Ziegler, Dasha Trofimova, Raphael Stock, Michael Baumgartner, Gregor Kohler, and Klaus Maier-Hein. Primus: Enforcing attention usage for 3d medical image segmentation, 2025. URL https://arxiv.org/abs/2503.01835." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 183, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 183, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 320, + 183, + 554, + 249 + ], + "type": "text", + "content": "[36] Sook-Lei Liew, Bethany P Lo, Miranda R Donnelly, Artemis Zavaliangos-Petropulu, Jessica N Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P Simon, Julia M Juliano, Anisha Suri, et al. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific data, 9(1):320, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 250, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 250, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 250, + 554, + 304 + ], + "type": "text", + "content": "[37] Endre Grøvik, Darwin Yi, Michael Iv, Elizabeth Tong, Daniel Rubin, and Greg Zaharchuk. Deep learning enables automatic detection and segmentation of brain metastases on multisequence mri. Journal of Magnetic Resonance Imaging, 51(1):175-182, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 305, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 305, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 320, + 305, + 554, + 370 + ], + "type": "text", + "content": "[38] Yuanfeng Ji, Haotian Bai, Chongjian Ge, Jie Yang, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhanng, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in Neural Information Processing Systems, 35: 36722–36732, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 371, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 371, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 371, + 554, + 437 + ], + "type": "text", + "content": "[39] Nicholas Heller, Fabian Isensee, Klaus H Maier-Hein, Xiaoshuai Hou, Chunmei Xie, Fengyi Li, Yang Nan, Guangrui Mu, Zhiyong Lin, Miofei Han, et al. The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis, 67:101821, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 437, + 554, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 437, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 320, + 437, + 554, + 502 + ], + "type": "text", + "content": "[40] Moritz R Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Valenzuela, Mauricio Reyes, Maria Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, et al. Isles 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. Scientific data, 9(1):762, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 503, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 503, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 320, + 503, + 554, + 547 + ], + "type": "text", + "content": "[41] Kareem Wahid, Cem Dede, Mohamed Naser, and Clifton Fuller. Training dataset for hntsmrg 2024 challenge, 2024. URL https://zenodo.org/doi/10.5281/ zenodo.11199559." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 548, + 554, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 554, + 592 + ], + "type": "text", + "content": "[42] Gasper Podobnik, Primož Strojan, Primož Peterlin, Bulat Ibragimov, and Tomaž Vrtovec. Han-seg: The head and neck organ-at-risk ct and mr segmentation dataset. Medical physics, 50(3):1917-1927, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 593, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 593, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 320, + 593, + 554, + 690 + ], + "type": "text", + "content": "[43] Ali M. Muslim, Syamsiah Mashohor, Gheyath Al Gawwam, Rozi Mahmud, Marsyita binti Hanafi, Osama Alnuaimi, Raad Josephine, and Abdullah Dhaifallah Almutairi. Brain mri dataset of multiple sclerosis with consensus manual lesion segmentation and patient meta information. Data in Brief, 42:108139, 2022. ISSN 2352-3409. doi: https://doi.org/10.1016/j.dib.2022.108139. URL https://www.sciencedirect.com/science/article/pii/S235234092200347X." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 554, + 713 + ], + "type": "text", + "content": "[44] Kaiyuan Yang, Fabio Musio, Yihui Ma, Norman Juchler, Johannes C. Paetzold, Rami Al-Maskari, Luciano" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23848" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 368 + ], + "type": "text", + "content": "Hoher, Hongwei Bran Li, Ibrahim Ethem Hamamci, Anjany Sekuboyina, Suprosanna Shit, Houjing Huang, Chinmay Prabhakar, Ezequiel de la Rosa, Diana Waldmannstetter, Florian Kofler, Fernando Navarro, Martin Menten, Ivan Ezhov, Daniel Rueckert, Iris Vos, Ynte Ruigrok, Birgitta Velthuis, Hugo Kuijf, Julien Hammerli, Catherine Wurster, Philippe Bijlenga, Laura Westphal, Jeroen Bisschop, Elisa Colombo, Hakim Baazaoui, Andrew Makmur, James Hallinan, Bene Wiestler, Jan S. Kirschke, Roland Wiest, Emmanuel Montagnon, Laurent Letourneau-Guillon, Adrian Galdran, Francesco Galati, Daniele Falcetta, Maria A. Zuluaga, Chaolong Lin, Haoran Zhao, Zehan Zhang, Sinyoung Ra, Jongyun Hwang, Hyunjin Park, Junqiang Chen, Marek Wodzinski, Henning Muller, Pengcheng Shi, Wei Liu Ting Ma, Cansu Yalçin, Rachika E. Hamadache, Joaquim Salvi, Xavier Llado, Uma Maria Lal-Trehan Estrada, Valeria Abramova, Luca Giancardo, Arnau Oliver, Jialu Liu, Haibin Huang, Yue Cui, Zehang Lin, Yusheng Liu, Shunzhi Zhu, Tatsat R. Patel, Vincent M. Tutino, Maysam Orouskhani, Huayu Wang, Mahmud Mossa-Basha, Chengcheng Zhu, Maximilian R. Rokuss, Yannick Kirchhoff, Nico Disch, Julius Holzschuh, Fabian Isensee, Klaus Maier-Hein, Yuki Sato, Sven Hirsch, Susanne Wegener, and Bjoern Menze. Benchmarking the cow with the topcow challenge: Topology-aware anatomical segmentation of the circle of willis for cta and mra, 2024. URL https://arxiv.org/abs/2312.17670." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 370, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 370, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 61, + 370, + 294, + 435 + ], + "type": "text", + "content": "[45] Divya Ramakrishnan, Leon Jekel, Saahil Chadha, Anastasia Janas, Harrison Moy, Nazanin Maleki, Matthew Sala, Manpreet Kaur, Gabriel Cassinelli Petersen, Sara Merkaj, et al. A large open access dataset of brain metastasis 3d segmentations on mri with clinical and imaging information. Scientific Data, 11(1):254, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 436, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 491 + ], + "type": "text", + "content": "[46] H. Chen, X. Zhao, J. Dou, C. Du, R. Yang, H. Sun, S. Yu, H. Zhao, C. Yuan, and N. Balu. Carotid vessel wall segmentation and atherosclerosis diagnosis challenge, 2022. URL https://vessel-wall-segmentation-2022.grand-challenge.org." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 492, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 294, + 535 + ], + "type": "text", + "content": "[47] Olivier Bernard, Alain Lalande, Clement Zotti, Cervenansky, and et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE TMI, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 536, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 602 + ], + "type": "text", + "content": "[48] Nicholas Bien, Pranav Rajpurkar, Robyn L Ball, Jeremy Irvin, Allison Park, Erik Jones, Michael Bereket, Bhavik N Patel, Kristen W Yeom, Katie Shpanskaya, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: development and retrospective validation of mrnet. PLoS medicine, 15(11):e1002699, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "type": "text", + "content": "[49] Hui Ming Lin, Errol Colak, Tyler Richards, Felipe C Kitamura, Luciano M Prevedello, Jason Talbott, Robyn L Ball, Ekim Gumeler, Kristen W Yeom, Mohammad Hamghalam, et al. The rsna cervical spine fracture ct dataset. Radiology: Artificial Intelligence, 5(5):e230034, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[50] Adriana Di Martino, Chao-Gan Yan, Qingyang Li, Erin Denio, Francisco X Castellanos, Kaat Alaerts, Jeffrey S Anderson, Michal Assaf, Susan Y Bookheimer, Mirella Dapretto, et al. The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain ar" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 94 + ], + "type": "text", + "content": "chitecture in autism. Molecular psychiatry, 19(6):659-667, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 96, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 172 + ], + "type": "text", + "content": "[51] Pedro RAS Bassi, Wenxuan Li, Yucheng Tang, Fabian Isensee, Zifu Wang, Jieneng Chen, Yu-Cheng Chou, Yannick Kirchhoff, Maximilian R Rokuss, Ziyan Huang, et al. Touchstone benchmark: Are we on the right way for evaluating ai algorithms for medical segmentation? Advances in Neural Information Processing Systems, 37:15184-15201, 2025." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 553, + 229 + ], + "type": "text", + "content": "[52] Lachlan T. Strike, Narelle K. Hansell, Jessica L. Miller, Kai-Hsiang Chuang, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin adolescent brain (qtab)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004146.v1.0.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 553, + 274 + ], + "type": "text", + "content": "[53] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-id1000\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso003097.v1.2.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 320, + 277, + 553, + 331 + ], + "type": "text", + "content": "[54] Makayla Gibson, Roger Newman-Norlund, Leonardo Bonilha, Julius Fridriksson, Gregory Hickok, Argye E. Hillis, Dirk-Bart den Ouden, and Chris Rorden. \"aphasia recovery cohort (arc) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004884.v1.0.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 333, + 553, + 365 + ], + "type": "text", + "content": "[55] Chris Rorden, John Absher, and Roger Newman-Norlund. \"stroke outcome optimization project (soop)\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004889.v1.1.2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 368, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 368, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 320, + 368, + 553, + 465 + ], + "type": "text", + "content": "[56] Allison C. Nugent, Adam G Thomas, Margaret Mahoney, Alison Gibbons, Jarrod Smith, Antoinette Charles, Jacob S Shaw, Jeffrey D Stout, Anna M Namyst, Arshitha Basavaraj, Eric Earl, Dustin Moraczewski, Emily Guinee, Michael Liu, Travis Riddle, Joseph Snow, Shruti Japee, Morgan Andrews, Adriana Pavletic, Stephen Sinclair, Vinai Roopchansingh, Peter A Bandettini, and Joyce Chung. \"the nimh healthy research volunteer dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005752.v2.1.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 320, + 468, + 553, + 533 + ], + "type": "text", + "content": "[57] Rani Gera, Maya Bar Or, Ido Tavor, Dana Roll, Jeffrey Cockburn, Segev Barak, Elizabeth Tricomi, John P. O'Doherty, and Tom Schonberg. \"characterizing habit learning in the human brain at the individual and group levels: a multi-modal mri study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004299.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 535, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 535, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 320, + 535, + 553, + 600 + ], + "type": "text", + "content": "[58] Denise Park, Joseph Hennessee, Evan T. Smith, Micaela Chan, Carson Katen, Julia Bacci, Sarah Frank, Sarah Monier, Alexandra Collyer, Carol Tamminga, William Moore, Neil Rofsky, Karen Rodrigue, Kristen Kennedy, and Gagan Wig. \"the dallas lifespan brain study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004856.v1.2.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 320, + 602, + 553, + 656 + ], + "type": "text", + "content": "[59] Jin Wang, Marisa N. Lytle, Yael Weiss, Brianna L. Yasasaki, and James R. Booth. \"a longitudinal neuroimaging dataset on language processing in children ages 5, 7, and 9 years old\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003604.v1.0.7." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[60] Leon Y. Cai, Qi Yang, Praitayini Kanakaraj, Vishwesh Nath, Allen T. Newton, Heidi A. Edmonson, Jeffrey Luci, Benjamin N. Conrad, Gavin R. Price, Colin B. Hansen, Cailey I. Kerley, Karthik Ramadass, Fang-Cheng Yeh, Hakmook Kang, Eleftherios Garyfallidis, Maxime Descoteaux," + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23849" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "text", + "content": "Francois Rheault, Kurt G. Schilling, and Bennett A. Landman. \"masivar: Multisite, multiscanner, and multisubject acquisitions for studying variability in diffusion weighted magnetic resonance imaging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003416.v2.0.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 295, + 261 + ], + "type": "text", + "content": "[61] Anahit Babayan, Blazeij Baczkowski, Roberto Cozatl, Maria Dreyer, Haakon Engen, Miray Erbey, Marcel Falkiewicz, Nicolas Farrugia, Michael Gaebler, Johannes Golchert, Laura Golz, Krzysztof Gorgolewski, Philipp Haueis, Julia Huntenburg, Rebecca Jost, Yelyzaveta Kramarenko, Sarah Krause, Deniz Kumral, Mark Lauckner, Daniel S. Margulies, Natacha Mendes, Katharina Ohrnberger, Sabine Oligschläger, Anastasia Osoianu, Jared Pool, Janis Reichelt, Andrea Reiter, Josefin Röbbig, Lina Schaare, Jonathan Smallwood, and Arno Villringer. \"mpileipzig_mind-brain-body\". OpenNeuro, 2020. doi: 10. 18112/openneuro.ds000221.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 262, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 294, + 316 + ], + "type": "text", + "content": "[62] Karl Koschutnig, Bernhard Weber, and Andreas Fink. \"tidying up white matter: Neuroplastic transformations in sensorimotor tracts following slackline skill acquisition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003138.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 341 + ], + "type": "text", + "content": "[63] Sangil Lee and Joseph Kable. \"cognitive training\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds002843.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "type": "text", + "content": "[64] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9653-9663, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 441 + ], + "type": "text", + "content": "[65] Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, and Zehuan Yuan. Designing bert for convolutional networks: Sparse and hierarchical masked modeling. arXiv preprint arXiv:2301.03580, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 294, + 487 + ], + "type": "text", + "content": "[66] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 294, + 520 + ], + "type": "text", + "content": "[67] Tom Schonberg, Christopher Trepel, Craig Fox, and Russell A. Poldrack. \"balloon analog risk-taking task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000001.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 294, + 566 + ], + "type": "text", + "content": "[68] J.M. Moran, E. Jolly, and J.P. Mitchell. \"social cognitive tasks (animate movement task, moral judgment task, false belief task)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000109.v2.0.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "type": "text", + "content": "[69] Ella Gabitov, David Manor, and Avi Karni. \"learning and memory: motor skill consolidation and intermanual transfer\". OpenNeuro, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 635 + ], + "type": "text", + "content": "[70] A. Roy, R. A. Bernier, J. Wang, M. Benson, J. French, D. Good, and F.G. Hillary. \"cost analysis tbi\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms000220.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 637, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 637, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 637, + 294, + 712 + ], + "type": "text", + "content": "[71] Noritaka Yoneyama, Hirohisa Watanabe, Kazuya Kawabata, Epifanio Bagarinao, Kazuhiro Hara, Takashi Tsuboi, Yasuhiro Tanaka, Reiko Ohdake, Kazunori Imai, Rei Ogura, Yasutaka Kato, Michihito Masuda, Tatsuya Hattori, Mizuki Ito, Satoshi Maesawa, Daisuke Mori, Masahisa Katsuno, and Gen Sobue. \"ds000245_r1.0.0\". OpenNeuro, 2018." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 95 + ], + "type": "text", + "content": "[72] H Aerts and D Marinazzo. \"btc_preop\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s001226.v5.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 97, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 97, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 321, + 97, + 553, + 149 + ], + "type": "text", + "content": "[73] Johns CL, Jahn AA, Jones HR, Kush DK, Molfese PJ, Van Dyke JA, Magnuson JS, Tabor W, Mencl WE, Shankweiler DP, and Braze D. \"indiv diffs readingskill\". OpenNeuro, 2018. doi: 10.18112/openneuro.d001365.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 185 + ], + "type": "text", + "content": "[74] C. Baldassano, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (scrambled)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001511.v1.0.5." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 239 + ], + "type": "text", + "content": "[75] A.M. Bornstein, M. Aly, S. Feng, N.B. Turk-Browne, K.A. Norman, and J.D. Cohen. \"perceptual decisions result from the continuous accumulation of memory and sensory evidence\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001614.v1.1.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 243, + 539, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 539, + 252 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 539, + 252 + ], + "type": "text", + "content": "[76] Unknown authors. \"sc mri dataset\". OpenNeuro, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 255, + 553, + 308 + ], + "type": "text", + "content": "[77] Eduardo A. Garza-Villarreal, Victor Pando-Naude, Fernando A. Barrios, Peter Vuust, Lene Vase, Elvira Brattico, Erick H. Pasaye, and Sarael Alcauter. \"functional connectivity of music-induced analgesia in fibromyalgia\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001928.v1.1.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 310, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 310, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 321, + 310, + 553, + 331 + ], + "type": "text", + "content": "[78] B. Sherman and N.B. Turk-Browne. \"stat episodic\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds002144.v2.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 321, + 333, + 553, + 365 + ], + "type": "text", + "content": "[79] Gilles de Hollander and Tomas H. Knapen. \"odc mapper 7t surface coil\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002295.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 367, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 367, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 321, + 367, + 553, + 421 + ], + "type": "text", + "content": "[80] Tomoyasu Horikawa, Alan S. Cowen, Dacher Keltner, and Yukiyasu Kamitani. \"the neural representation of visually evoked emotion is high-dimensional, categorical, and distributed across transmodal brain regions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002425.v1.0.2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 423, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 423, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 321, + 423, + 553, + 478 + ], + "type": "text", + "content": "[81] Yazhu Qian, Jinyou Zou, Zihao Zhang, Zhentao Zuo, Yan Zhuo, Danny JJ, Wang, and Peng Zhang. \"robust functional mapping of layer-selective responses in human lateral geniculate nucleus with high-resolution 7t fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002672.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 321, + 479, + 553, + 511 + ], + "type": "text", + "content": "[82] G. Elliott Wimmer, Jamie Li, Krzysztof Jacek Gorgolewski, and Russell A. Poldrack. \"rewardbeast\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s002738.v1.0.2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 514, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 514, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 321, + 514, + 553, + 545 + ], + "type": "text", + "content": "[83] Francesca M M Citron, Nora Michaelis, and Adele E Goldberg. \"l2_sentences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002842.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 547, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 547, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 547, + 553, + 601 + ], + "type": "text", + "content": "[84] Matteo Visconti di Oleggio Castello, Vassiki Chauhan, Guo Jiahui, and M. Ida Gobbini. \"an fmri dataset in response to \"the grand budapest hotel\", a socially-rich, naturalistic movie\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds003017.v1.0.3." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 321, + 603, + 553, + 657 + ], + "type": "text", + "content": "[85] Chiara Banfi, Karl Koschutnig, Kristina Moll, Gerd Schulte-Koerne, Andreas Fink, and Karin Landerl. \"reading-related functional activity in children with isolated spelling deficits and dyslexia\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003126.v1.3.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 712 + ], + "type": "text", + "content": "[86] Andy Schumann, Feliberto de la Cruz, Stefanie K\"ohler, Lisa Brotte, and Karl-J\"urgen B\"ar. \"the influence of heart rate variability biofeedback on cardiac regulation and functional brain connectivity\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003357.v1.0.0." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23850" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 58, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "text", + "content": "[87] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant4 motion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003441.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 117, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 294, + 183 + ], + "type": "text", + "content": "[88] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., Todd R. Parrish, and Barbara Trommer. \"response inhibition and selective attention in adults and children with and without ADHD\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003500.v1.2.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 184, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 248 + ], + "type": "text", + "content": "[89] Nathan M. Muncy, C. Brock Kirwan, and Jonathan J. Wisco. \"differences in chemo-signaling compound-evoked brain activity in male and female young adults: A pilot study in the role of sexual dimorphism in olfactory chemosignaling\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003606.v1.0.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 294, + 303 + ], + "type": "text", + "content": "[90] Beau Sievers, Carolyn Parkinson, Peter J. Kohler, James Hughes, Sergey V. Fogelson, and Thalia Wheatley. \"visual and auditory brain areas share a representational structure that supports emotion perception: fmri data\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003715.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 304, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 370 + ], + "type": "text", + "content": "[91] Andreas Fink, Karl Koschutnig, Thomas Zussner, Corinna M. Perchtold-Stefan, Christian Rominger, Mathias Benedek, and Ilona Papousek. \"a two-week running intervention reduces symptoms related to depression and increases hippocampal volume in young adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003799.v2.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 294, + 414 + ], + "type": "text", + "content": "[92] John Wilder, Morteza Rezanejad, Sven Dickinson, Kaleem Siddiqi, Allan Jepson, and Dirk B. Walther. \"sceneperceptionandparallelism\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03950.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 458 + ], + "type": "text", + "content": "[93] Zoe Woodhead, Maria Sablik, Kate Watkins, David Carey, Dorothy Bishop, and Emma Karlsson. \"comparing language lateralisation using fmri and ftcd\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004073.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 491 + ], + "type": "text", + "content": "[94] Martin Szinte, Tomas Knapen, and Matthias Nau. \"gaze_exp\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004271.v1.1.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 492, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 294, + 525 + ], + "type": "text", + "content": "[95] Maria Czarnecka, Andre Knops, and Marcin Szwed. \"number comparison task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004400.v1.0.6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "type": "text", + "content": "[96] Alex L. White, Kendrick Kay, Kenny A. Tang, and Jason D. Yeatman. \"raw data for engaging in word recognition elicits highly specific modulations in visual cortex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004489.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 294, + 624 + ], + "type": "text", + "content": "[97] Loriana Goulding, Austin Schmidt, Isaac Hamm, and C. Brock Kirwan. \"fmri investigations of individual differences on memory activation for faces and words: The effects of handedness and phenomenal experience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004589.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 656 + ], + "type": "text", + "content": "[98] Halle R. Dimsdale-Zucker and Christopher Baldassano. \"common\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004663.v1.0.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 690 + ], + "type": "text", + "content": "[99] Rotem Botvinik-Nezer, Bogdan Petre, Marta Ceko, Naomi Friedman, and Tor Wager. \"paingen_placebo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004746.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 692, + 294, + 713 + ], + "type": "text", + "content": "[100] Xuefeng Yan, Martin Noergaard, Cheryl L. Morse, JeihSan Liow, Jinsoo Hong, Sanjay Telu, Min-Jeong Kim, Jose" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 171 + ], + "type": "text", + "content": "A Montero Santamaria, Anthony Galassi, Ningping Feng, Sarah K. Williams Avram, Ted B. Usdin, Shawn Wu, Andrea Zhang, Lester Manly, Madeline Jenkins, Maia Van Buskirk, Adrian Lee, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. \"positron emission tomography (pet) quantification in healthy humans of cyclooxygenase-2 (cox-2), a potential biomarker of neuroinflammation\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004869.v1.1.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "text", + "content": "[101] Liberty S. Hamilton, Maansi Desai, and Alyssa Field. \"wired icm sample dataset - workshop on intracranial recordings in humans, epilepsy, dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004993.v1.1.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "text", + "content": "[102] Keith M. Vogt, Alex C. Burlew, Sujatha N. Reddy, Marcus Simmons, Courtney N. Kozdron, and James W. Ibinson. \"neural correlates of lidocaine analgesic (nla) study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005088.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 274, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 274, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 553, + 350 + ], + "type": "text", + "content": "[103] Sidhant Chopra, Carrisa V. Cocuzza, Connor Lawhead, Jocelyn A. Ricard, Loic Labache, Lauren Patrick, Poornima Kumar, Arielle Rubenstein, Julia Moses, Lia Chen, Crystal Blankenbaker, Bryce Gillis, Laura T. Germine, Ilan Harpaz-Rote, BT Thomas Yeo, Justin T. Baker, and Avram J. Holmes. \"transdiagnostic connectome project\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005237.v1.1.3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 352, + 553, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 352, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 317, + 352, + 553, + 395 + ], + "type": "text", + "content": "[104] Marleen Haupt, Douglas D. Garrett, and Radoslaw M. Cichy. \"object recognition in healthy aging (orha) - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005374.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 397, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 397, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 397, + 553, + 441 + ], + "type": "text", + "content": "[105] Luis Martinez Agulleiro, Bowen Deng, Yuliya Yoncheva, Chao-Gan Yan, F Xavier Castellanos, and Marie-Pierre St-Onge. \"circadian misalignment and energy balance\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005525.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 443, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 553, + 464 + ], + "type": "text", + "content": "[106] A.R. Aron, M.A. Gluck, and R.A. Poldrack. \"classification learning\". OpenNeuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 520 + ], + "type": "text", + "content": "[107] Rebecca J. Lepping, Ruth Ann Atchley, Evangelia Chrysikou, Laura E. Martin, Alicia A. Clair, Rick E. Ingram, W. Kyle Simmons, and Cary R. Savage. \"neural processing of emotional musical and nonmusical stimuli in depression\". OpenNeuro, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 522, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 564 + ], + "type": "text", + "content": "[108] Elizabeth Bock, Peter Donhauser, Francois Tadel, Guiomar Niso, and Sylvain Baillet. \"meg-bids brainstorm data sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000246.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 609 + ], + "type": "text", + "content": "[109] Joseph M. Orr and Vijay A. Mittal. \"rule learning in adolescents at clinical high risk for psychosis shows altered patterns of activation\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps01371.v1.1.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 553, + 655 + ], + "type": "text", + "content": "[110] A.C. Schapiro, T.T. Rogers, N.I. Cordova, N.B. Turk-Browne, and M.M. Botvinick. \"neural representations of events arise from temporal community structure\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001621.v1.1.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 700 + ], + "type": "text", + "content": "[111] Omer Faruk Gulban, Kevin R. Sitek, Satrajit S. Ghosh, Michelle Moerel, and Federico De Martino. \"auditory localization with 7t fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001942.v1.2.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 702, + 553, + 713 + ], + "type": "text", + "content": "[112] Tomoya Nakai and Shinji Nishimoto. \"over 100 task fmri" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23851" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "text", + "content": "dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002306.v1.1.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[113] Laura Pritschet, Tyler Santander, Caitlin M. Taylor, Evan Layher, Shuying Yu, Michael B. Miller, Scott T. Grafton, Emily G. Jacobs, Morgan Fitzgerald, and Joshua M. Mueller. \"28andme\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002674.v1.0.6." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[114] PhD Heidemarie Laurent, Megan K. Finnegan, and Katherine Haigler. \"postnatal affective mri dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003136.v1.0.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 230 + ], + "type": "text", + "content": "[115] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant5 suboptimal event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003442.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 275 + ], + "type": "text", + "content": "[116] Monica Crotti, Karl Koschutnig, and Selina Wriessnegger. \"the impact of handedness on the neural correlates during kinesthetic motor imagery: a fmri study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003612.v1.0.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[117] Ben Deen and Winrich Freiwald. \"parallel systems for social and spatial reasoning\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms003814.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 354 + ], + "type": "text", + "content": "[118] Shaonan Wang, Xiaohan Zhang, Jiajun Zhang, and Chengqing Zong. \"a synchronized multimodal neuroimaging dataset to study brain language processing\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004078.v1.2.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "text", + "content": "[119] Shani Agron, Reut Weissgross, Eva Mishor, Lior Gorodisky, Aharon Ravia, Edna Furman-Haran, Tali Weiss, and Noam Sobel. \"a chemical signal in human emotional tears blocks aggression in men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004274.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "text", + "content": "[120] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004493.v1.0.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[121] Jose Pedro Manzano Patron, Steen Moeller, Jesper L.R. Andersson, Kamil Ugurbil, Essa Yacoub, and Stamatios N. Sotiropoulos. \"edden: Evaluation of diffusion mri denoising\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004666.v1.0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[122] Samira Epp, Gabriel Castrillon, Beija Yuan, Jessica Andrews-Hanna, Christine Preibisch, and Valentin Riedl. \"two distinct modes of hemodynamic responses in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004873.v2.0.6." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 644 + ], + "type": "text", + "content": "[123] Chloe M. de Nys, Ee Shern Liang, Marita Prior, Maria A. Woodruff, James I. Novak, Ashley R. Murphy, Zhiyong Li, Craig D. Winter, and Mark C. Allenby. \"royalbrisbane_tofmra_intracranial aneurysm_database\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005096.v1.0.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[124] Luis Marcos-Vidal, Helena Gil-Buitrago, Irene Cisma, Rosamunde C. Hendricks, Scott Atran, and Clara Pretus. \"polex\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005375.v1.0.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[125] G. Xue and Russell A. Poldrack. \"rhyme judgment\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms000003.v1.0.0." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[126] Tom S.M., Fox C.R., Trepel C., and Poldrack R.A. \"mixed-gambles task\". OpenNeuro, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 118 + ], + "type": "text", + "content": "[127] K Jimura, E Stover, F Casalis, and RA Poldrack. \"ds000006\". OpenNeuro, 2018." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 120, + 553, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 120, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 120, + 553, + 142 + ], + "type": "text", + "content": "[128] Xue G, Aron AR, and Poldrack RA. \"stop-signal task with spoken & manual responses\". OpenNeuro, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 144, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 553, + 175 + ], + "type": "text", + "content": "[129] A.R. Aron, T.E. Behrens, M. Frank, S. Smith, and R.A. Poldrack. \"stop-signal task with unconditional and conditional stopping\". OpenNeuro, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 177, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 198 + ], + "type": "text", + "content": "[130] Jessica Cohen and Russell Poldrack. \"the generality of self-control\". OpenNeuro, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "text", + "content": "[131] K. Foerde, B.J. Knowlton, and Russell A. Poldrack. \"classification learning and tone-counting\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso000011.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 235, + 553, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 235, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 316, + 235, + 553, + 256 + ], + "type": "text", + "content": "[132] Rizk-Jackson, Aron, and Russell A. Poldrack. \"ds000017\". OpenNeuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 301 + ], + "type": "text", + "content": "[133] R Bilder, R Poldrack, T Cannon, E London, N Freimer, E Congdon, K Karlgodt, and F Sabb. \"ucla consortium for neuropsychiatric phenomics la5c study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000030.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 304, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 325 + ], + "type": "text", + "content": "[134] Russell Poldrack. \"myconnectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs000031.v2.0.2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 348 + ], + "type": "text", + "content": "[135] R. Alvarez and R.A. Poldrack. \"cross-language repetition priming\". OpenNeuro, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 553, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 383 + ], + "type": "text", + "content": "[136] R.A. Poldrack, J. Clark, E. J. Pare-Blagoev, D. Shohamy, J. Creso Moyano, C. Myers, and M Gluck. \"classification learning and reversal\". OpenNeuro, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "text", + "content": "[137] Mei-Yen Chen, Corey N. White, Nathan Giles, Albert Elumn, Sagar Parikh, Ungi Kim, W. Todd Maddox, and Russell A. Poldrack. \"training of loss aversion modulates neural sensitivity toward potential gains\". OpenNeuro, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 430, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 450 + ], + "type": "text", + "content": "[138] Kelly AMC and Milham MP. \"simon task\". OpenNeuro, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 453, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 483 + ], + "type": "text", + "content": "[139] Kelly AMC, Uddin LQ, Biswal BB, Castellanos FX, and Milham MP. \"flanker task (event-related)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 529 + ], + "type": "text", + "content": "[140] J.V. Haxby, M.I. Gobbini, M.L. Furey, A. Ishai, J.L. Schouten, and P. Pietrini. \"visual object recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000105.v3.0.0." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 553 + ], + "type": "text", + "content": "[141] K.J. Duncan, C. Pattamadilok, I. Knierim, and J.T. Devlin. \"word and object processing\". OpenNeuro, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 556, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 588 + ], + "type": "text", + "content": "[142] Wager T.D., M.L. Davidson, B.L. Hughes, M.A. Lindquist, and K.N. Ochsner. \"prefrontal-subcortical pathways mediating successful emotion regulation\". OpenNeuro, 2018." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "text", + "content": "[143] Melina R. Uncapher, J. Benjamin Hutchinson, and Anthony D. Wagner. \"incidental encoding task (posner cueing paradigm)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[144] Michael Hanke, Florian J. Baumgartner, Pierre Ibe, Falko R. Kaule, Stefan Pollmann, Oliver Speck, Wolf Zinke, and Jorg Stadler. \"forrest gump\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps00113.v1.3.0." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[145] Gorgolewski KJ, Storkey A, Bastin ME, Whittle IR, Wardlaw JM, and Pernet CR. \"a test-retest fmri dataset for motor, language and spatial attention functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds000114.v1.0.2." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23852" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[146] Grega Repov" + }, + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "inline_equation", + "content": "\\mathring{\\mathrm{A}}_{\\mathrm{i}}" + }, + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": " and Deanna M. Barch. \"working memory in healthy and schizophrenic individuals\". OpenNeuro, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 295, + 139 + ], + "type": "text", + "content": "[147] Jennifer M Walz, Robin I Goldman, Jordan Muraskin, Bryan Conroy, Truman R Brown, and Paul Sajda. \"auditory and visual oddball eeg-fmri\". OpenNeuro, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 140, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 57, + 140, + 295, + 172 + ], + "type": "text", + "content": "[148] DG Wakeman and RN Henson. \"multisubject, multimodal face processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.rs000117.v1.1.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 173, + 294, + 216 + ], + "type": "text", + "content": "[149] Katerina Velanova, Mark E. Wheeler, and Beatrix Luna. \"maturational changes in anterior cingulate and frontoparietal recruitment support the development of error processing and inhibitory control (antistate)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 294, + 270 + ], + "type": "text", + "content": "[150] Aarthi Padmanabhana, Charles F. Geier, Sarah J. Ordaz, Theresa Teslovich, and Beatrix Luna. \"developmental changes in brain function underlying the influence of reward processing on inhibitory control (slot reward)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 271, + 294, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 271, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 271, + 294, + 315 + ], + "type": "text", + "content": "[151] C.F. Geier, R. Terwilliger, T. Teslovich, K. Velanova, and B. Luna. \"immaturities in reward processing and its influence on inhibitory control in adolescence (ring reward)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 316, + 294, + 360 + ], + "type": "text", + "content": "[152] J Xu, S Zhang, VD Calhoun, J Monterosso, and MN Potenza. \"task-related concurrent but opposite modulations of overlapping functional networks as revealed by spatial ica\". OpenNeuro, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "text", + "content": "[153] Nicoletta Cera, Armando Tartaro, and Stefano L. Sensi. \"modafinil alters intrinsic functional connectivity of the right posterior insula: a pharmacological resting state fmri study\". OpenNeuro, 2018." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 404, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 404, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 57, + 404, + 294, + 446 + ], + "type": "text", + "content": "[154] Choong-Wan Woo, Mathieu Roy, Jason T. Buhle, and Tor D. Wager. \"distinct brain systems mediate the effects of nociceptive input and self-regulation on pain\". OpenNeuro, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 449, + 294, + 491 + ], + "type": "text", + "content": "[155] Kimberly L. H. Carpenter, Adrian Angold, Nan-Kuei Chen, William E. Copeland, Pooja Gaur, Kevin Pelphrey, Allen W. Song, and Helen L. Egger. \"preschool anxiety disorders\". OpenNeuro, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 294, + 536 + ], + "type": "text", + "content": "[156] E. Necka, M. Senderecka, B. Kucharzyk, and Falkiewicz M. \"who can afford self-control? the neural efficiency mechanism explains effective self-regulation of behavior\". OpenNeuro, 2019." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 294, + 568 + ], + "type": "text", + "content": "[157] Elissa M. Aminoff and Michael J. Tarr. \"scene perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000149.v1.0.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "type": "text", + "content": "[158] Paul A. M. Smeets, Floor M. Kroese, Catherine Evers, and D. T. D. de Ridder. \"block design food and nonfood picture viewing task\". OpenNeuro, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 668 + ], + "type": "text", + "content": "[159] CR Pernet, P Belin, P McAleer, KJ Gorgolewski, M Valdes-Sosa, M Latinus, I Charest, PEG Bestelmeyer, RH Watson, and D Fleming. \"the human voice areas: spatial organisation and inter-individual variability in temporal and extra-temporal cortices\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000158.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 287, + 680 + ], + "type": "text", + "content": "[160] Timothy D. Verstynen. \"stroop task\". OpenNeuro, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[161] James K. Bursley, Adrian Nestor, Michael J. Tarr, and J. David Creswell. \"offline processing in associative learning\". OpenNeuro, 2019." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[162] Hsu AL, Chou KH, Chao YP, Fan HY, Wu CW, and Chen JH. \"physiological contribution in spontaneous oscillations: An approximate quality - assurance index for resting-state fmri signals\". OpenNeuro, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 184 + ], + "type": "text", + "content": "[163] Laura Koenders, Janna Cousijn, Wilhelmina A.M. Vingerhoets, Wim van den Brink, Reinout W. Wiers, Carin J. Meijer, Marise W. J. Machielsen, Dick J. Veltman, Anneke E. Goudriaan, and Lieuwe de Haan. \"t1-weighted structural mri study of cannabis users at baseline and 3 years follow up\". OpenNeuro, 2019." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "text", + "content": "[164] Emilia Iannilli. \"pre-adolescents exposure to manganese\". OpenNeuro, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 296 + ], + "type": "text", + "content": "[165] Gustav Nilsson, Sandra Tamm, Paolo d'Onofrio, Hanna Å Thuné, Johanna Schwarz, Catharina Lavebratt, Jia Jia Liu, Kristoffer NT Måsson, Tina Sundelin, John Axelsson, Peter Fransson, Göran Kecklund, Håkan Fischer, Mats Lekander, and Torbjörn Åkerstedt. \"the stockholm sleepy brain study: Effects of sleep deprivation on cognitive and emotional processing in young and old\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000201.v1.0.3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "text", + "content": "[166] Peter Van Schuerbeek, Chris Baeken, and Johan De Mey. \"the heterogeneity in retrieved relations between the personality trait 'harm avoidance' and gray matter volumes due to variations in the vbm and roi labeling processing settings\". OpenNeuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "text", + "content": "[167] Christian Stephan-Otto, Sara Siddi, Carl Senior, Daniel Muñoz-Samons, Susana Ochoa, Ana María Sánchez Laforga, and Gildas Brébion. \"visual imagery and false memory for pictures\". OpenNeuro, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 487 + ], + "type": "text", + "content": "[168] Dara Dickstein, Mariel Y. Pullman, Corey Fernandez, Jennifer A. Short, Lale Kostakoglu, Karin Knaesaurek, Laili Soleimani, Barry Jordan, Wayne Gordon, Kristen Dams-O'Connor, Bradley N. Delman, Edmund Wong, Cheuk Y. Tang, Steven T. DeKosky, James R. Stone, Robert C. Cantu, Mary Sano, Patrick R. Hof, and Sam Gandy. \"imaging [18f]av-1451 and [18f]av-45 in acute and chronic traumatic brain injury\". OpenNeuro, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 510 + ], + "type": "text", + "content": "[169] Jongwan Kim, Jing Wang, Svetlana V. Shinkareva, and Douglas H. Wedell. \"affective videos\". OpenNeuro, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 588 + ], + "type": "text", + "content": "[170] Vincent A. Magnotta, Joy T. Matsui, Dawei Liu, Hans J. Johnson, Jeffrey D. Long, Bradley D. Bolster Jr, Bryon A. Mueller, Kelvin Lim, Susumu Mori, Karl G. Helmer, Jessica A. Turner, Sarah Reading, Mark J. Lowe, Elizabeth Aylward, Laura A. Flashman, Greg Bonett, and Jane S. Paulsen. \"dwi traveling human phantom study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds000206.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[171] Tetreault P, Mansour A, Vachon-Presseau E, Schnitzer TJ, Apkarian AV, and Baliki MN. \"brain connectivity predicts placebo response across chronic pain clinical trials\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds000208.v1.0.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[172] Elizabeth DuPre, Wen-Ming Luh, and R. Nathan Spreng. \"multi-echo fmri replication sample of autobiographical memory, prospection and theory of mind reasoning tasks\". OpenNeuro, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 680, + 553, + 713 + ], + "type": "text", + "content": "[173] Liane Young, Alek Chakroff, Emily Wasserman, Rebecca Saxe, James Dungan, Amelia Brown, and Jorie Koster-Hale. \"moral judgments of intentional and accidental moral" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23853" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 14 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "violations across harm and purity domains\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000212.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 128 + ], + "type": "text", + "content": "[174] Xiao Gao, Xiao Deng, Xin Wen, Ying She, Petra Corianne Vinke, and Hong Chen. \"neural mechanism underlying appearance social comparison\". OpenNeuro, 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 129, + 294, + 150 + ], + "type": "text", + "content": "[175] Stephen Giles, Jeremy Hall, Merrick Pope, Katie Nicol, and Liana Romaniuk. \"eupd cyberball\". OpenNeuro, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 194 + ], + "type": "text", + "content": "[176] Alexander D. Cohen, Andrew S. Nencka, R. Marc Lebel, and Yang Wang. \"multiband multi-echo imaging of simultaneous oxygenation and flow timeseries for resting state connectivity\". OpenNeuro, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 195, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 227 + ], + "type": "text", + "content": "[177] Avi J H Chanales, Ashima Oza, Serra E Favila, and Brice A Kuhl. \"route learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms000217.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 260 + ], + "type": "text", + "content": "[178] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (oral nutritional supplements)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "text", + "content": "[179] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, L. Nanetti, and Gert J. ter Horst. \"flavour pleasantness (regular products)\". OpenNeuro, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 337 + ], + "type": "text", + "content": "[180] Thomas H B FitzGerald, Dorothea Haemmerer, Karl J Friston, Shu-Chen Li, and Raymond J Dolan. \"sequential inference vbm\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000222.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[181] Ian Ballard, Bokyung Kim, Anthony Liatsis, Gökhan Aydogan, Jonathan D. Cohen, and Samuel M. McClure. \"magnitude effect\". OpenNeuro, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 372, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 470 + ], + "type": "text", + "content": "[182] Evan M. Gordon, Timothy O. Laumann, Adrian W. Gilmore, Dillan J. Newbold, Deanna J. Greene, Jeffrey J. Berg, Mario Ortega, Catherine Hoyt-Drazen, Caterina Gratton, Haoxin Sun, Jacqueline M. Hampton, Rebecca S. Coalson, Annie Nguyen, Kathleen B. McDermott, Joshua S. Shimony, Abraham Z. Snyder, Bradley L. Schlaggar, Steven E. Petersen, Steven M. Nelson, and Nico U.F. Dosenbach. \"the midnight scan club (msc) dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms000224.v1.0.4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[183] Hilary Richardson, Grace Lisandrelli, Alexa Riobueno-Naylor, and Rebecca Saxe. \"mri data of 3-12 year old children and adults during viewing of a short animated film\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000228.v1.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 581 + ], + "type": "text", + "content": "[184] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward - study 1\". OpenNeuro, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 646 + ], + "type": "text", + "content": "[185] Maria Geraldine Veldhuizen, Richard Keith Babbs, Barkha Patel, Wambura Fobbs, Nils B Kroemer, Elizabeth Garcia, Martin Yeomans, and Dana Small. \"integration of sweet taste and metabolism determines carbohydrate reward-study 3\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000231.v1.0.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 680 + ], + "type": "text", + "content": "[186] Johan D Carlin and Nikolaus Kriegeskorte. \"adjudicating between face-coding models with individual-face fmri responses\". OpenNeuro, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[187] Samuel A. Nastase, Yaroslav O. Halchenko, Andrew C. Connolly, M. Ida Gobbini, and James V. Haxby. \"neural responses to naturalistic clips of behaving animals in" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "two different task contexts\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000233.v1.0.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[188] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 1\". OpenNeuro, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[189] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds000235.v2.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[190] Marta Vidorreta, Ze Wang, Yulin V. Chang, Maria A. Fernandez-Seara, and John A. Detre. \"whole-brain background-suppressed pcasl mri with 1d-accelerated 3d rare stack-of-spirals readout- dataset 3\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps00236.v2.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 306 + ], + "type": "text", + "content": "[191] Yuji Kanazawa, Kimihiro Nakamura, Toru Ishii, Toshihiko Aso, Hiroshi Yamazaki, and Koichi Omori. \"phonological memory in sign language relies on the visuomotor neural system outside the left hemisphere language network\". OpenNeuro, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 330 + ], + "type": "text", + "content": "[192] D. Zeithamova, M. De Araujo Sanchez, and A. Adke. \"trial timing for multivariate pattern analysis\". OpenNeuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[193] Julian Maclaren, Zhaoying Han, Sjoerd B. Vos, Nancy Fischbein, and Roland Bammer. \"maclaren test-retest brain volume dataset\". OpenNeuro, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[194] Alvaro Galiano, Reyes Garcia de Eulate, Marta Vidorreta, Miriam Recio, Mario Riverol, José L. Zubieta, and Maria A. Fernandez-Seara PhD. \"resting state perfusion in healthy aging\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds000240.v2.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[195] Andrew C. Connolly, J. Swaroop Guntupalli, Jason Gors, Michael Hanke, Yaroslav O. Halchenko, Yu-Chien Wu, Hervé Abdi, and James V. Haxby.\"ak6\". OpenNeuro, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 455, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 476 + ], + "type": "text", + "content": "[196] Steve Petersen, Brad Schlaggar, and Jonathan Power. \"washington university 120\". OpenNeuro, 2018." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 510 + ], + "type": "text", + "content": "[197] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"individual brain charting\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds000244.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "text", + "content": "[198] Guiomar Niso, Jeremy Moreau, Elizabeth Bock, Francois Tadel, and Sylvain Baillet. \"meg-bids omega restingstate_sample\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds000247.v1.0.2." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 578 + ], + "type": "text", + "content": "[199] Alexandre Gramfort and Matti S Hämäläinen. ””. Open-Neuro, 2020. doi: 10.18112/openneuro.ds000248.v1.2.4." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 579, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 601 + ], + "type": "text", + "content": "[200] Agnes Norbury and Ben Seymour. \"value generalization in human avoidance learning\". OpenNeuro, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 624 + ], + "type": "text", + "content": "[201] Diana Gorbet and Lauren Sergio. \"female action video game players\". OpenNeuro, 2018." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[202] Alexander D. Cohen, Andrew S. Nencka, and Yang Wang. \"multiband multi-echo simultaneous asl/bold for task-induced functional mri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000254.v1.0.0." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[203] Yoichi Miyawaki, Hajime Uchida, Okito Yamashita, Masaaki Sato, Yusuke Morito, Hiroki C. Tanabe, Norihiro Sadato, and Yukiyasu Kamitani. \"visual image reconstruction\". OpenNeuro, 2018." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23854" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 15 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "type": "text", + "content": "[204] Deanna J. Greene, Jonathan M. Koller, Jacqueline M. Hampton, Victoria Wesevich, Andrew N. Van, Annie L. Nguyen, Catherine R. Hoyt, Lindsey McIntyre, Eric A. Earl, Rachel L. Klein, Joshua S. Shimony, Steven E. Petersen, Bradley L. Schlaggar, Damien A. Fair, and Nico U.F. Dosenbach. \"behavioral interventions for reducing head motion during migraine scans in children\". OpenNeuro, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[205] Jonathan Power, Valerie Voon, Prantik Kundu, Suyash Bhogawar, Peter Bandettini, and Alex Martin. \"multi-echo dambridge\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds000258.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 241 + ], + "type": "text", + "content": "[206] Michael Milham, Cameron Craddock, Russ Tobe, Anna MacKay-Brandt, Stan Colcombe, Bennett Leventhal, and F. Xavier Castellanos. \"nki-sample\". OpenNeuro, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 243, + 294, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 254 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 254 + ], + "type": "text", + "content": "[207] Chris Gorgolewski. \"the brain of chris\". OpenNeuro, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 277 + ], + "type": "text", + "content": "[208] Mark A. Thornton, Miriam E. Weaverdyck, and Diana I. Tamir. \"social predictive coding\". OpenNeuro, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 294, + 300 + ], + "type": "text", + "content": "[209] A. Zadbood, J. Chen, Y.C. Leong, K.A. Norman, and U. Hasson. \"sherlock_merlin\". OpenNeuro, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 302, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 294, + 334 + ], + "type": "text", + "content": "[210] Y. Yeshurun, M. Nguyen, and U. Hasson. \"milky-vodka\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001131.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 294, + 369 + ], + "type": "text", + "content": "[211] J. Chen, Y.C. Leong, C.J. Honey, K.A. Yong C.H., Norman, and U. Hasson. \"sherlock\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001132.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[212] J. Chen, C. J. Honey, E. Simony, M. J. Arcaro, K. A. Norman, and U. Hasson. \"twilight zone movie watching dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001145.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 416, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 503 + ], + "type": "text", + "content": "[213] Chris Gorgolewski, Natacha Mendes, Domenica Wilfling, Elisabeth Wladimirow, Claudine J. Gauthier, Tyler Bonnen, Florence J.M. Ruby, Robert Trampel, Pierre-Louis Bazin, Roberto Cozatl, Jonathan Smallwood, and Daniel S. Margulies. \"a high resolution 7-tesla resting-state fmri test-retest dataset with cognitive and physiological measures\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001168.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 506, + 294, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 294, + 527 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 294, + 527 + ], + "type": "text", + "content": "[214] Krzysztof J. Gorgolewski. \"example mr artifacts\". Open-Neuro, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 529, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 294, + 551 + ], + "type": "text", + "content": "[215] Mark Thornton, Miriam Weaverdyck, and Diana Tamir. \"state representation pictures\". OpenNeuro, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 553, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 574 + ], + "type": "text", + "content": "[216] Sam Gilbert and Hoki Fung. \"decoding intentions of self and others\". OpenNeuro, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 576, + 294, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 630 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 630 + ], + "type": "text", + "content": "[217] J.A. Lewis-Peacock, J.D. Cohen, and K.A. Norman. \"neural evidence of the strategic choice between working memory and episodic memory in prospective remembering\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001229.v1.0.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 633, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 294, + 666 + ], + "type": "text", + "content": "[218] J.A. Lewis-Peacock and K.A. Norman. \"competition between items in working memory leads to forgetting\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001232.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 667, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 667, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 667, + 294, + 689 + ], + "type": "text", + "content": "[219] Patrick Beukema and Timothy Verstynen. \"singlefingerrsa\". OpenNeuro, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[220] Christopher Baldassano, Diane M. Beck, and Li Fei-Fei. \"human-object interaction\". OpenNeuro, 2018." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[221] Michelle G. Hall, Claire K. Naughtin, Jason B. Mattingley, and Paul E. Dux. \"olvsl_ object-location visual statistical learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001241.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "type": "text", + "content": "[222] Tae-Ho Lee, Steven Greening, Taiji Ueno, David Clewett, Allison Ponzio, Michiko Sakaki, and Mara Mather. \"examining effects of arousal on responses to salient and non-salient stimuli in younger and older adults\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001242.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 553, + 205 + ], + "type": "text", + "content": "[223] Tomoyasu Horikawa and Yukiyasu Kamitani. \"generic object decoding (fmri onImagenet)\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001246.v1.2.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 315, + 206, + 553, + 258 + ], + "type": "text", + "content": "[224] Damien Brevers, Sarah C Herremans, Qinghua He, Marie-Anne Vanderhasselt, Mathieu Petieau, Dimitri Verdonck, Tasha Poppa, Sara De Witte, Charles Kornreich, Antoine Bechara, and Chris Baeken. \"sports_betting_study\". Open-Neuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 260, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 260, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 315, + 260, + 553, + 303 + ], + "type": "text", + "content": "[225] Matteo Visconti di Oleggio Castello, Yaroslav O. Halchenko, J. Swaroop Guntupalli, Jason D. Gors, and M. Ida Gobbini. \"personally familiar faces\". OpenNeuro, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 304, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 304, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 553, + 347 + ], + "type": "text", + "content": "[226] Frank J. Kanayet, Andrew Mattarella-Micke, Peter J. Kohler, Anthony M. Norcia, Bruce D. McCandliss, and James L. McClelland. \"numberlinemarking\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001299.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 349, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 349, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 349, + 553, + 380 + ], + "type": "text", + "content": "[227] Sean O'Bryan, Darrell Worthy, Evan Livesey, and Tyler Davis. \"inverse base rate\". OpenNeuro, 2018. doi: 10.18112/openneuro.ms001302.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 382, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 382, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 315, + 382, + 553, + 414 + ], + "type": "text", + "content": "[228] Reshanne R. Reeder, Christian N. L. Olivers, and Stefan Pollmann. \"negative_template\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001306.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 415, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 415, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 315, + 415, + 553, + 447 + ], + "type": "text", + "content": "[229] E.S. Finn, P.R. Corlett, G. Chen, P.A. Bandettini, and R.T. Constable. \"paranoiastory\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001338.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 449, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 449, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 553, + 479 + ], + "type": "text", + "content": "[230] Ivy F. Tso, Saige E. Rutherford, Yu Fang, Mike Angstadt, and Stephan F. Taylor. \"socialbrain\". OpenNeuro, 2018. doi: 10.52181/zenodod.1210217." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 481, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 481, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 481, + 553, + 514 + ], + "type": "text", + "content": "[231] Jessica-Lily Harvey, Lysia Demetriou, John McGonigle, and Matthew B. Wall. \"phfmri_control_tasks\". OpenNeuro, 2018. doi: 10.18112/openneuro.rs001344.v1.0.0." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 515, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 515, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 553, + 557 + ], + "type": "text", + "content": "[232] Michael P Notter, Sandra Da Costa, and Micah M Murray. \"decoding of multisensory semantics and memories in low-level visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001345.v1.0.0." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 559, + 553, + 602 + ], + "type": "text", + "content": "[233] Brock Kirwan, Thomas Meservy, and Kelly Fadel. \"information filtering in electronic networks of practice: An fmri investigation of expectation [dis]confirmation\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001353.v1.0.0." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 603, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 603, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 315, + 603, + 553, + 635 + ], + "type": "text", + "content": "[234] Amado Catarina, Kovács Petra, Mayer Rebecca, Ambrus G. Geza, Trapp Sabrina, and Kovács Gyula. \"namefacepriming\". OpenNeuro, 2018." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 315, + 636, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 636, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 315, + 636, + 553, + 657 + ], + "type": "text", + "content": "[235] Christopher Baldassano. \"loci case study\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs001379.v1.0.0." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 315, + 658, + 532, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 532, + 669 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 532, + 669 + ], + "type": "text", + "content": "[236] Dr. Frank Hillary. \"subject data\". OpenNeuro, 2018." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "text", + "content": "[237] Joset A. Etzel, Debbie Yee, Bidhan Lamichhane, Mitchell Jeffers, Elisa Di Rosa, Jenny Crawford, Hongyu An, and Todd S. Braver. \"multibandacqtests\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso001399.v2.0.0." + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23855" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 16 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 128 + ], + "type": "text", + "content": "[238] Dobrushina Olga R, Vlasova Roza M, Rumshiskaya Alena D, Litvinova Liudmila D, Mershina Elena A, Sinit-syn Valentin E, and Pechenkova Ekaterina V. \"rsfmri_single_session_eeg_nf\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001408.v1.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 152 + ], + "type": "text", + "content": "[239] Megen Brittell. \"audio cartography\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001415.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 154, + 294, + 186 + ], + "type": "text", + "content": "[240] Rotem Botvinik-Nezer, Tom Salomon, and Tom Schonberg. \"cat snacks functional plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001417.v1.0.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 189, + 294, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 294, + 276 + ], + "type": "text", + "content": "[241] Martin Wegrzyn, Joana Aust, Larissa Barnstorf, Magdalena Gippert, Mareike Harms, Antonia Hautum, Shanna Heidel, Friederike Herold, Sarah M. Hommel, Anna-Katharina Knigge, Dominik Neu, Diana Peters, Marius Schaefer, Julia Schneider, Ria Vombrock, Sabrina M. Zimmer, Friedrich G. Woermann, and Kirsten Labudda. \"thought-experiment\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001419.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 312 + ], + "type": "text", + "content": "[242] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]dasb pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001420.v1.2.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 346 + ], + "type": "text", + "content": "[243] Melanie Ganz-Benjaminsen and Martin Noergaard. \"[11c]sb207145 pet gimbi database example\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms001421.v1.4.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 392 + ], + "type": "text", + "content": "[244] G. Kim, K.A. Norman, and N.B. Turk-Browne. \"neural overlap in item representations across episodes impairs context memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001430.v1.0.2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 395, + 294, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 294, + 450 + ], + "type": "text", + "content": "[245] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001439.v1.2.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 452, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 452, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 294, + 507 + ], + "type": "text", + "content": "[246] A.C. Schapiro, E.A. McDevitt, T.T. Rogers, S.C. Mednick, and K.A. Norman. \"human hippocampal replay during rest prioritizes weakly learned information and predicts memory performance\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001454.v1.3.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 574 + ], + "type": "text", + "content": "[247] Ilaria Berteletti, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, Jérôme Prado, Macarena Suarez Pellicioni, Jessica Wise Younger, and James R. Booth. \"brain correlates of math development\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds001486.v1.3.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 620 + ], + "type": "text", + "content": "[248] Jelle R. Dalenberg, Liselore Weitkamp, Remco J. Renken, and Gert J. ter Horst. \"valence processing differs across stimulus modalities (multi-echo)\". OpenNeuro, 2018. doi: 10.18112/openneuro.ps001491.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 667 + ], + "type": "text", + "content": "[249] J Lewis-Peacock, BR Postle, CR Cox, and TT Rogers. \"long-term memory (ltm) for famous faces, places, and common objects\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001497.v1.0.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[250] Nadine Chang, John A. Pyles, Austin Marcus, Abhinav Gupta, Michael J. Tarr, and Elisaa M. Aminoff. \"bold5000\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001499.v1.3.1." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[251] Guohua Shen, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"deep image reconstruction\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds001506.v1.3.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 150 + ], + "type": "text", + "content": "[252] C. Baldassano, R. Masis-Obando, U. Hasson, and K.A. Norman. \"schematic narrative perception and recall (intact)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001510.v2.0.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "text", + "content": "[253] P. Kok and N.B. Turk-Browne. \"associative prediction of visual shape in the hippocampus\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs001517.v1.0.3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "text", + "content": "[254] S. C. Chan, M. C. Applegate, N. W. Morton, S. M. Polyn, and K. A. Norman. \"lingering representations of stimuli influence recall organization\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001521.v1.0.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "type": "text", + "content": "[255] Feng Rong, A. Lisette Isenberg, Erica Sun, and Gregory S. Hickok. \"syllable level speech sequencing\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001525.v1.1.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 553, + 297 + ], + "type": "text", + "content": "[256] Andrea Courtney, Emma PeConga, Dylan Wagner, and Kristina Rapuano. \"calorie-labeled food cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001534.v1.1.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 298, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 298, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 315, + 298, + 553, + 341 + ], + "type": "text", + "content": "[257] M. Aly, J. Chen, N.B. Turk-Browne, and U. Hasson. \"learning naturalistic temporal structure in the posterior medial network\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001545.v1.1.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 342, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 342, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 315, + 342, + 553, + 363 + ], + "type": "text", + "content": "[258] Eliska Prochazkova. \"ds: ds001551:1.0.0\". OpenNeuro, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "type": "text", + "content": "[259] Javier Gonzalez-Castillo, Ziad S. Saad, Daniel A. Handwaker, Souheil J. Inati, Noah Brenowitz, and Peter A. Bandettini. \"100 runs at 3t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001553.v1.0.1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 553, + 453 + ], + "type": "text", + "content": "[260] O. Lositsky, J. Chen, D. Toker, C.J. Honey, J.L. Poppenk, U. Hasson, and K.A. Norman. \"neural pattern change during encoding of a narrative predicts retrospective duration estimates\". OpenNeuro, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 455, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 455, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 315, + 455, + 553, + 509 + ], + "type": "text", + "content": "[261] Javier Gonzalez-Castilloa, Colin W. Hoy, Daniel A. Handwerker, Vinai Roopchansingh, Souheil J. Inati, Ziad S. Saad, Robert W. Cox, and Peter A. Bandettini. \"large single-subject functional mri datasets at 7t\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001555.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 510, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 510, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 510, + 553, + 531 + ], + "type": "text", + "content": "[262] Renzo Huber. \"cmrr workshop\". OpenNeuro, 2019. doi: mockDOI." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 533, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 533, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 315, + 533, + 553, + 553 + ], + "type": "text", + "content": "[263] Unknown authors. \"ds: ds001566:1.0.1\". OpenNeuro, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 315, + 555, + 553, + 599 + ], + "type": "text", + "content": "[264] A.N. Hoskin, A.M. Bornstein, K.A. Norman, and J.D. Cohen. \"refresh my memory: Episodic memory reinstatements intrude on working memory maintenance\". Open-Neuro, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 315, + 601, + 553, + 633 + ], + "type": "text", + "content": "[265] Ian C. Ballard, Anthony D. Wagner, and Samuel M. McClure. \"feature discrimination\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001590.v1.0.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 635, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 553, + 655 + ], + "type": "text", + "content": "[266] Unknown authors. \"test dataset for xcp software\". Open-Neuro, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 553, + 689 + ], + "type": "text", + "content": "[267] Patrick Beukema and Timothy Verstynen. \"multifingerrsa\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001597.v1.0.0." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 691, + 553, + 713 + ], + "type": "text", + "content": "[268] Matt Cieslak, Mark Elliott, and Ted Satterthwaite. \"example fieldmaps\". OpenNeuro, 2018." + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23856" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 17 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[269] A.M. Bornstein and K.A. Norman. \"reinstated episodic context guides sampling-based decisions for reward\". OpenNeuro, 2018. doi: 10.18112/openneuro.ds001607.v1.0.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 151 + ], + "type": "text", + "content": "[270] Aya Ben-Yakov and Noga Cohen. \"prestimulus exp2\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001608.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "text", + "content": "[271] I. Momennejad, A.R. Otto, N. Daw, and K.A. Norman. \"offline replay supports planning in human reinforcement learning\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001612.v1.0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "text", + "content": "[272] Adam Hampshire and Eyal Soreq. \"visual working memory: Study one task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001634.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 266 + ], + "type": "text", + "content": "[273] Adam Hampshire and Eyal Soreq. \"visual working memory: Study two task fmri and behavioural response\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds001635.v1.0.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 269, + 294, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 294, + 290 + ], + "type": "text", + "content": "[274] Elizabeth Valles. \"ironia vev\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs001652.v2.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 346 + ], + "type": "text", + "content": "[275] Mattia Veronese, Gaia Rizzo, Martin Belzunce, Julia Schubert, Barbara Santangelo, Ayla Mansur, Alex Whittington, Joel Dunn, Graham Searle, Andrew Reader, and Roger Gunn. \"nrm2018 pet grand challenge dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms001705.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[276] A. Shenhav, M.A. Straccia, S. Musslick, J.D. Cohen, and M.M. Botvinick. \"dissociable neural mechanisms track evidence accumulation for selection of attention versus action\". OpenNeuro, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 395, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 395, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 395, + 294, + 460 + ], + "type": "text", + "content": "[277] Brendan Gaesser, Joshua Hirschfeld-Kroen, Emily Wasserman, Mary Horn, and Liane Young. \"a role for the medial temporal lobe subsystem in guiding prosociality: the effect of episodic processes on willingness to help others (experiment 2)\". OpenNeuro, 2019. doi: 10.18112/openneuro. ds001722.v1.1.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 463, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 483 + ], + "type": "text", + "content": "[278] Unknown authors. \"test-geier: Ds: ds001728:1.1.0\". Open-Neuro, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 518 + ], + "type": "text", + "content": "[279] Rotem Botvinik-Nezer, Roni Iwanir, Russell A. Poldrack, and Tom Schonberg. \"narps\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s001734.v1.0.5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "type": "text", + "content": "[280] Birgit Rauchbauer, Bruno Nazarian, Morgan Bourhis, Magalie Ochs, Laurent Prevot, and Thierry Chaminade. \"brain activity during reciprocal social interaction investigated using conversational robots as control condition\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001740.v2.2.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 610 + ], + "type": "text", + "content": "[281] Nathaniel P Miller, Yao Liu, Roman Krivochenitser, and Bas Rokers. \"unilateral glaucoma 3t dmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001743.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[282] J.R. Manning, J.C. Hulbert, J. Williams, L. Piloto, L. Sahakyan, and K.A. Norman. \"a neural signature of contextually mediated intentional forgetting\". OpenNeuro, 2019. doi: 10.18112/openneuro.d001745.v1.1.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[283] Carrie E. Gold, Athena L. Howell, Jacob Burdis, C. Brock Kirwan, and Gregory L. Thompson. \"exploring the resting state neural activity of monolinguals and late and early bilinguals\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001747.v1.1.0." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[284] Samuel Fynes-Clinton, Lars Marstaller, and Hana Burianova. \"differentiation of functional networks during long-term memory retrieval in children and adolescents\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001748.v1.0.4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[285] Bart Aben, Cristian Buc Calderon, Laurens Van der Cruyssen, Doerte Picksak, Eva Van den Bussche, and Tom Verguts. \"context-dependent cognitive control — flanker task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001751.v1.0.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[286] R. VanRullen and L. Reddy. \"reconstructing faces from fmri patterns using deep generative neural networks\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001761.v2.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[287] A. Shenhav, M.A. Straccia, M.M. Botvinick, and J.D. Cohen. \"dorsal anterior cingulate and ventromedial prefrontal cortex have inverse roles in both foraging and economic choice\". OpenNeuro, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 317 + ], + "type": "text", + "content": "[288] Virginia AGLIERI, Bastien CAGNA, Pascal BELIN, and Sylvain TAKERKART. \"intertva: a multimodal mri dataset for the study of inter-individual differences in voice perception and identification\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001771.v1.0.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[289] J.C. Hulbert and K.A. Norman. \"neural differentiation tracks improved recall of competing memories following interleaved study and retrieval practice\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001775.v1.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[290] Alik Wadge and Sam Zorowitz. \"effects of on/off deep brain stimulation on cognitive control in treatment-resistant depression (eeg)\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001784.v1.1.2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 441 + ], + "type": "text", + "content": "[291] Vincent DeLuca and Christos Pliatsikas. \"bilingualism and the brain\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds001796.v1.7.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 444, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 508 + ], + "type": "text", + "content": "[292] Samuel Zorowitz, Alexander P Rockhill, Kristen K Kellard, Katherine E Link, Todd Herrington, Diego A Pizzagalli, Alik S Widge, Thilo Deckersbach, and Darin D Dougherty. \"neuroanatomical correlates of approach-avoidance conflict (fmri)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds001814.v1.0.7." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 510, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 532 + ], + "type": "text", + "content": "[293] Unknown authors. \"cat faces mri experiment. ds: ds001818:1.0.0\". OpenNeuro, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 566 + ], + "type": "text", + "content": "[294] Isabel Alkhasli, Katrin Sakreida, Felix M. Mottaghy, and Ferdinand Binkofski. \"resting state - tms\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001832.v1.0.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 610 + ], + "type": "text", + "content": "[295] Celia Goffin, H. Moriah Sokolowski, Michael Slipenkyj, and Daniel Ansari. \"handedness and symbolic number representation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001838.v1.0.1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 634 + ], + "type": "text", + "content": "[296] Unknown authors. \"spinal stimulation stepping and standing dataset. ds: ds001839:1.0.1\". OpenNeuro, 2019." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 666 + ], + "type": "text", + "content": "[297] Noga Cohen and Aya Ben-Yakov. \"prestimulus exp1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001840.v1.0.2." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[298] S.C.Y. Chan, Y. Niv, and K.A. Norman. \"a probability distribution over latent causes, in the orbitofrontal cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001847.v1.0.1." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23857" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 18 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[299] H. Moriah Sokolowski, Zachary Hawes, Lien Peters, and Daniel Ansari. \"parallel adaptation of symbols, quantities, and physical size\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001848.v1.0.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[300] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making intertemporal choice task dataset\". OpenNeuro, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[301] Matthew Piva, Kayla Velnoskey, Ruonan Jia, Amrita Nair, Ifat Levy, and Steve W.C. Chang. \"social decision-making risky choice task dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001883.v1.0.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 272 + ], + "type": "text", + "content": "[302] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"longitudinal brain correlates of multisensory lexical processing in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds001894.v1.4.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 295, + 329 + ], + "type": "text", + "content": "[303] Trevor K. M. Day, Tara M. Madyastha, Peter Board, Mary K. Askren, Nicholas J. Wapstra, Thomas J. Montine, and Thomas J. Grabowski. \"ant: Healthy aging and parkinson's disease\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds001907.v3.1.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 373 + ], + "type": "text", + "content": "[304] Milena Kaestner, Ryan T. Maloney, Kirstie H. Wailes-Newson, Marina Bloj, Julie M. Harris, Antony B. Morland, and Alex R. Wade. \"s-cone_motiondepth_fmri\". Open-Neuro, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 418 + ], + "type": "text", + "content": "[305] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value\". OpenNeuro, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 464 + ], + "type": "text", + "content": "[306] A. Shenhav, M.A. Straccia, J.D. Cohen, and M.M. Botvinick. \"anterior cingulate engagement in a foraging context reflects choice difficulty, not foraging value (2)\". OpenNeuro, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 498 + ], + "type": "text", + "content": "[307] Ghootae Kim, Kenneth A. Norman, and Nicholas B. Turk-Browne. \"violation differentiation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ps001926.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 499, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 294, + 553 + ], + "type": "text", + "content": "[308] Gabriel Castrillon, Nico Sollmann, Katarzyna Kurcyus, Adeel Razi, Sandro M. Krieg, and Valentin Riedl. \"the physiological effects of non-invasive brain stimulation fundamentally differ across the human cortex\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds001927.v2.1.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "type": "text", + "content": "[309] N.C. Hindy, E.W. Avery, and N.B. Turk-Browne. \"hippocampal-neocortical interactions sharpen over time for predictive actions\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds001946.v1.0.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 621 + ], + "type": "text", + "content": "[310] Peter J. Kohler and Anthony M. Norcia. \"svndl anatomical data\". OpenNeuro, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 656 + ], + "type": "text", + "content": "[311] Peter J. Kohler, Benoit R. Cotterreau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (central letter task)\". OpenNeuro, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[312] Peter J. Kohler, Benoit R. Cottereau, and Anthony M. Noria. \"image segmentation based on relative motion and relative disparity cues (passive fixation)\". OpenNeuro, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[313] Benjamin J. Griffiths, Stephen D. Mayhew, Karen J. Mullinger, João Jorge, Ian Charest, Maria Wimber, and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Simon Hanslmayr. \"reinstatement_fidelity\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002000.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[314] Janine Mendola and Elizabeth Bock. \"rivalry_tagging\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002001.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[315] Akram Bakkour, Daphna Shohamy, and Michael N. Shadlen. \"memory and decision making dataset\". Open-Neuro, 2019. doi: 10.18112/openneuro.ms002006.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 207 + ], + "type": "text", + "content": "[316] David Clewett, Ringo Huang, Rico Velasco, Tae-Ho Lee, and Mara Mather. \"locus coeruleus activity strengthens prioritized memories under arousal\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002011.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[317] Jakob Heinzle (data acquisition), Kai Goergen (data curation), and Joram Soch (BIDS conversion). \"stimdisc\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002013.v1.0.3." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 296 + ], + "type": "text", + "content": "[318] Martin Wegrzyn, Markus Mertens, Christian G. Bien, Friedrich G. Woermann, and Kirsten Labudda. \"language production fmri\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002014.v1.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 341 + ], + "type": "text", + "content": "[319] Yiming Xiao, Jonathan C. Lau, Taylor Anderson, Jordan DeKraker, D. Louis Collins, Terry M. Peters, and Ali R. Khan. \"bigbrainmricoreg\". OpenNeuro, 2019. doi: 10.18112/openneuro.d0s002016.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 386 + ], + "type": "text", + "content": "[320] Rebecca Merkley, Benjamin Conrad, Gavin Price, and Daniel Ansari. \"investigating the visual number form area: A replication study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002033.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 544, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 544, + 399 + ], + "type": "text", + "content": "[321] Unknown authors. \"bids_parkinson\". OpenNeuro, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "text", + "content": "[322] Jaime J. Castrellon, Jacob S. Young, Linh C. Dang, Ronald L. Cowan, David H. Zald, and Gregory R. Samanez-Larkin. \"mesolimbic dopamine d2 receptors and neural representations of subjective value\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002041.v2.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 509 + ], + "type": "text", + "content": "[323] Emily S Finn, Laurentius Huber, David C Jangraw, Peter J Molfese, and Peter A Bandettini. \"layer-dependent activity in human prefrontal cortex during working memory\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002076.v1.0.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 542 + ], + "type": "text", + "content": "[324] Hannelore Aerts and Daniele Marinazzo. \"btc_postop\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002080.v4.0.0." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 588 + ], + "type": "text", + "content": "[325] Maedbh King, Carlos Hernandez-Castillo, Russell A. Poldrack, Jonathon Walters, Richard B. Ivry, and Jorn Diedrichsen. \"multi-domain task battery (mdtb)\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002105.v1.1.0." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 590, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 632 + ], + "type": "text", + "content": "[326] Celia Goffin, Michael Slipenkyj, Stephan E. Vogel, Rebecca Merkley, and Daniel Ansari. \"development of symbolic number processing\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002116.v1.0.0." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "text", + "content": "[327] Adam Kimbler, Amanda G. Hamm, and Aaron T. Mattfeld. \"conditional visual associative learning task\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002149.v1.0.5." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 689 + ], + "type": "text", + "content": "[328] Charlie, John, Ross, and Joey. \"testrun\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002153.v1.0.1." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 712 + ], + "type": "text", + "content": "[329] Florence Bouhali, Zoé Bézagu, Stanislas Dehaene, and Laurent Cohen. \"a mesial-to-lateral dissociation for ortho-" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23858" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 19 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "graphic processing in the visual cortex\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso002155.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "text", + "content": "[330] Dylan M. Nielson and Argyris Stringaris. \"multi-echo masking test dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002156.v2.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 206 + ], + "type": "text", + "content": "[331] Michael Pereira, Nathan Faivre, Inaki Iturrate, Marco Wirthlin, Luana Serafini, Stephanie Martin, Arnaud Desvachez, Olaf Blanke, Dimitri Van de Ville, and Jose del R. Millan. \"disentangling the origins of confidence in speeded perceptual judgments through multimodal imaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002158.v1.0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 295, + 262 + ], + "type": "text", + "content": "[332] Michelle I. Nash, Cooper B. Hodges, Nathan M. Muncy, and C. Brock Kirwan. \"pattern separation beyond the hippocampus: A high-resolution whole-brain investigation of mnemonic discrimination in healthy adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002168.v1.1.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 264, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 295, + 297 + ], + "type": "text", + "content": "[333] Jiefeng Jiang, Shao-Fang Wang, Wanjia Guo, Corey Fernandez, and Anthony D. Wagner. \"ccd\". OpenNeuro, 2019. doi: 10.18112/openneuro.dso02169.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[334] Tali Weiss, Timna Soroka, Lior Gorodisky, Sagit Shushan, Kobi Snitz, Reut Weisgross, Edna Furman-Haran, Thijs Dhollander, and Noam Sobel. \"human olfaction without apparent olfactory bulbs\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002185.v1.1.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 354, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 354, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 57, + 354, + 294, + 375 + ], + "type": "text", + "content": "[335] Michele Svanera and Lars Muckli. \"cerebrum: a 3t mri segmentation tool\". OpenNeuro, 2019. doi: mockDOI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 376, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 294, + 399 + ], + "type": "text", + "content": "[336] Dorit Kliemann and Ralph Adolphs. \"caltech rsfmri dataset\". OpenNeuro, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 455 + ], + "type": "text", + "content": "[337] Jordan Bigio, Tali Bitan, Douglas Bolger, Douglas Burman, Fan Cao, Tai-Li Chou, Nadia Cone, Jessica Gayda, Dong Lu, Marisa Lytle, Jenni Minas, and James R. Booth. \"cross-sectional multidomain lexical processing\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002236.v1.1.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 488 + ], + "type": "text", + "content": "[338] J. Poppenk and K. A. Norman. \"briefly cueing memories leads to suppression of their neural representations\". Open-Neuro, 2019. doi: 10.18112/openneuro.ds002237.v1.1.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "text", + "content": "[339] J. Fan, J.D. Wammes, J.B. Gunn, D.L. Yamins, K.A. Norman, and N.B. Turk-Browne. \"relating visual production and recognition of objects in human visual cortex\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002241.v1.1.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[340] Nathan M. Muncy and C. Brock Kirwan. \"correcting false memories: The effect of mnemonic generalization on original memory traces\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002242.v1.0.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 634 + ], + "type": "text", + "content": "[341] Joshua Hirschfeld-Kroen, Emily Wasserman, Stefano Anzellotti, and Liane Young. \"when my wrongs are worse than yours: behavioral and neural asymmetries in first-person and third-person perspectives of accidents\". Open-Neuro, 2020. doi: 10.18112/openneuro.d002250.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 668 + ], + "type": "text", + "content": "[342] Jean-Marie Bonny, Juliana Iranpour, and Béatrice Claise. \"viewing_food_cues\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002270.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[343] Katherine L. Bottenhorn, Taylor Salo, Julio A. Peraza, Michael C. Riedel, Jessica S. Flannery, Adam Kimbler, Alfredo Toll, Diego Suarez, Francis M. Cruz, Israel Zagales, Nayade Caldes, Olivia Dolan, Ruth Zagales, Matthew T." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "Sutherland, Robert W. Laird, and Angela R. Laird. \"dense investigation of variability of affect (diva)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds002278.v2.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 129 + ], + "type": "text", + "content": "[344] Unknown authors. \"fsl open science dev dataset\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002293.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[345] Elizabeth Long, Nathan Wheeler, and William Cunningham. \"trait-role\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002294.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 165, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 165, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 165, + 553, + 206 + ], + "type": "text", + "content": "[346] J. Poppenk and K.A. Norman. \"multiple-object tracking as atool for parametrically modulating memory reactivation\". OpenNeuro, 2019. doi: 10.18112/openneuro.ds002311.v1.1.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 242 + ], + "type": "text", + "content": "[347] P. G. Bissett, J. M. Shine, J. Durnez, J. Li, K. J. Gorgolewski, and R. A. Poldrack. \"rewardbeast\". OpenNeuro, 2019. doi: 10.18112/openneuro.ms002316.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 276 + ], + "type": "text", + "content": "[348] Luiz Pessoa, Christian Meyer, and Srikanth Padmala. \"dynamic_passive-threat\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002320.v1.1.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "text", + "content": "[349] Shohini Bhattasali, Jonathan R. Brennan, Wen-Ming Luh, Berta Franzluebbers, and John T. Hale. \"the alice dataset: fmri dataset to study natural language comprehension in the brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002322.v1.0.4." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 334, + 553, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 334, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 553, + 355 + ], + "type": "text", + "content": "[350] Taylor Hanayik. \"fsl open science dev dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms002328.v1.0.2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 357, + 553, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 357, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 316, + 357, + 553, + 389 + ], + "type": "text", + "content": "[351] A. Sunavsky and J. Poppenk. \"neuroimaging predictors of creativity in healthy adults\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002330.v1.1.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 391, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 391, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 391, + 553, + 456 + ], + "type": "text", + "content": "[352] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp1\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002336.v2.0.2." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 458, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 458, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 316, + 458, + 553, + 523 + ], + "type": "text", + "content": "[353] Giulia Lioi, Claire Cury, Lorraine Perronnet, Marsel Mano, Elise Bannier, Anatole Lecuyer, and Christian Barillot. \"a multi-modal human neuroimaging dataset for data integration: simultaneous EEG and fmri acquisition during a motor imagery neurofeedback task: Xp2\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002338.v2.0.2." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 525, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 633 + ], + "type": "text", + "content": "[354] Samuel A. Nastase, Yun-Fei Liu, Hanna Hillman, Asieh Zadbood, Liat Hasenfratz, Neggin Keshavarzian, Janice Chen, Christopher J. Honey, Yaara Yeshurun, Mor Regev, Mai Nguyen, Claire H. C. Chang, Christopher Baldassano, Olga Lositsky, Erez Simony, Michael A. Chow, Yuan Chang Leong, Paula P. Brooks, Emily Micciche, Gina Choe, Ariel Goldstein, Tamara Vanderwal, Yaroslav O. Halchenko, Kenneth A. Norman, and Uri Hasson. \"narratives\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002345.v1.1.4." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[355] Brock Kirwan, Daniel Bjornn, Bonnie Anderson, Anthony Vance, David Eargle, and Jeffrey Jenkins. \"repetition of computer security warnings results in differential repetition suppression effects as revealed with functional mri\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002363.v1.1.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[356] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23859" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 20 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "text", + "content": "\"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds002366.v1.1.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[357] Deni Kurban, Renzo Huber, Gilad Liberman, Dimo Ivanov, Sri Kashyap, Debora Niekaemper, and Benedikt Poser. \"ohbm abstract 2020\". OpenNeuro, 2019. doi: 10.18112/openneuro.rs002367.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[358] M.D. Rajendra Morey. \"serumgreymattercorticalthickness\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002380.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 239 + ], + "type": "text", + "content": "[359] Rogers CS, Jones MS, McConkey S, Spehar B, Van Engen KJ, Sommers MS, and Peelle JE. \"age-related differences in auditory cortex activity during spoken word recognition\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002382.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[360] Yoshiharu I Kutani, Takatomi Kubo, Satoshi Nishida, Hideaki Hata, Kenichi Matsumoto, Kazushi Ikeda, and Shinji Nishimoto. \"fmri dataset on program comprehension and expertise\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002411.v1.1.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 363 + ], + "type": "text", + "content": "[361] Jelle R. Dalenberg, Barkha P. Patel, Raphael Denis, Maria G. Veldhuizen, Yuko Nakamura, Petra C. Vinke, Serge Luquet, and Dana M. Small. \"short-term consumption of sucralose with, but not without, carbohydrate impairs neural and metabolic sensitivity to sugar\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002419.v1.0.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[362] Yana Panikratova, Alexander Tomyshev, Ekaterina Pechenkova, and Roza Vlasova. \"fmri: resting state and arithmetic task\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002422.v1.1.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[363] James R. Booth, GE Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, MA Stein, and Michael Tennekoon. \"working memory and reward in children with and without attention deficit hyperactivity disorder (adhd)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002424.v1.2.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 531 + ], + "type": "text", + "content": "[364] Damien Brevers, Chris Baeken, Antoine Bechara, Qinghua He, Pierre Maurage, Mathieu Petieau, Guillaume Sescousse, Claus Vögele, and Joel Billieux. \"increased brain reactivity to gambling unavailability is a marker of problem gambling\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002513.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 533, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 565 + ], + "type": "text", + "content": "[365] Tamar Kolodny, Michael-Paul Schallmo, and Scott O. Murray. \"contrast response functions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002522.v1.0.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 610 + ], + "type": "text", + "content": "[366] S. Jahfari, K.R. Ridderinkhof, A.G.E. Collins, T. Knapen, L.J. Waldorp, M.J. Frank, and J. Theeuwes. \"learning, inhibitory control, and perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s002543.v1.0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 655 + ], + "type": "text", + "content": "[367] Suzanne Oosterwijk, Lukas Snoek, Mark Rotteveel, Lisa Feldman Barrett, and H. Steven Scholte. \"shared-states\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002547.v1.1.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[368] Laboratory for Computational Neurodiagnostics. \"protecting the aging brain (pagb)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002549.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[369] Romain Quentin, Jean-Remi King, Etienne Sallard, Nathan Fishman, Ryan Thompson, Ethan Buch, and Leonardo" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 116 + ], + "type": "text", + "content": "Cohen. \"differential brain mechanisms of selection and maintenance of information during working memory (meg data)\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002550.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 149 + ], + "type": "text", + "content": "[370] Tomas Knapen, Wietske van der Zwaag, and Daan van Es. \"cerebellum retinotopic mapping\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002574.v1.0.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 182 + ], + "type": "text", + "content": "[371] Arnaud Delorme and Scott Makeig. \"visual oddball task (256 channels)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs002578.v1.1.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 248 + ], + "type": "text", + "content": "[372] S. Zhang, W. Yoshida, H. Mano, T. Yanagisawa, K. Shibata, M. Kawato, and B. Seymour. \"cognitive control of sensory pain encoding in the pregenual anterior cingulate cortex. d1 - decoder construction in day 1, d2 - adaptive control in day 2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002596.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 282 + ], + "type": "text", + "content": "[373] Lukas Snoek, Tinka Beemstermboer, and H. Steven Scholte. \"mb-epi-comparison\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002603.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "text", + "content": "[374] Caroline Le Ster, Antonio Moreno, Franck Mauconduit, Vincent Gras, Ruediger Stirnberg, Benedikt A. Poser, Alexandre Vignaud, Evelyn Eger, Stanislas Dehaene, Florent Meyniel, and Nicolas Boulant. \"bids_dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002606.v1.1.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[375] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study1\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds002608.v1.0.2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[376] Tamas Spisak, Balint Kincses, Frederik Schlitt, Matthias Zunhammer, Tobias Schmidt-Wilcke, Zsigmond T. Kincses, and Ulrike Bingel. \"rpn-signature_study2\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002609.v1.0.3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 448 + ], + "type": "text", + "content": "[377] Enzo Tagliazucchi. \"meditation interocepcion\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002614.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[378] William K. Lloyd, Jayne Morriss, Birthe Macdonald, Karin Joanknecht, Julie Nihouarn, and Carien M. van Reekum. \"emotion regulation in the ageing brain, university of reading, bbsrc\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002620.v1.0.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[379] N Eichert, KE Watkins, and project-specific co authors. \"project_larynx\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002634.v3.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[380] L.N. Van der Laan, C. Scholz, R.A. Poldrack, D.T.D. De Ridder, and A. Smidts. \"can we have a second helping? a replication study on the neurobiological mechanisms underlying self-control\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002643.v1.1.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "text", + "content": "[381] Mara Mather, Ringo Huang, David Clewett, Shawn Nielsen, Rico Velasco, Kristie Tu, Sophia Han, and Briana Kennedy. \"isometric exercise facilitates attention to salient events in women via the noradrenergic system\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002647.v1.0.1." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[382] Nathan M. Muncy and C. Brock Kirwan. \"the medial temporal lobe supports mnemonic discrimination for event duration\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002655.v1.0.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[383] Brian J. Coffey, Zachary D. Threlkeld, Andrea S. Foulkes, Yelena G. Bodien, and Brian L. Edlow. \"the language" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23860" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 21 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 105 + ], + "type": "text", + "content": "network reemerges during recovery from severe traumatic brain injury\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002675.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 138 + ], + "type": "text", + "content": "[384] Cheryl A. Olman and Kimberly B. Weldon. \"kung fu panda\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002684.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 172 + ], + "type": "text", + "content": "[385] Ana Luisa Grilo Pinho, Lucie Hertz-Pannier, and Bertrand Thirion. \"ibc\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs002685.v2.0.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 216 + ], + "type": "text", + "content": "[386] James R. Booth, Gillian E. Cooke, Jessica Gayda, Rubi Hammer, Marisa N. Lytle, Mark A. Stein, and Michael Tennekoon. \"working memory and reward in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002687.v1.2.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "text", + "content": "[387] Kendrick Kay, Keith W. Jamison, Luca Vizioli, Ruyuan Zhang, Eshed Margalit, and Kamil Ugurbil. \"high-field 7t visual fmri datasets\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002702.v1.0.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "text", + "content": "[388] Christos Chatzichristos, Manuel Morante, Nikos Andreadis, Eleftherios Kofidis, Yiannis Kopsinis, and Sergios Theodoridis. \"smiley\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002711.v1.1.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 338 + ], + "type": "text", + "content": "[389] Sara Aurtenetxe, Nicola Molinaro, Doug Davidson, and Manuel Carreiras. \"numbers and letters\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps002712.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 371 + ], + "type": "text", + "content": "[390] Petra Vetter, Lukasz Bola, Lior Reich, Matthew Bennett, Lars Muckli, and Amir Amedi. \"avscenes_blind\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002715.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 470 + ], + "type": "text", + "content": "[391] Liron Rozenkrantz, Reut Weissgross, Tali Weiss, Inbal Ravrebi, Idan Frumin, Sagit Shushan, Lior Gorodisky, Netta Reshef, Yael Holzman, Liron Pinchover, Yaara Endevelt-Shapira, Eva Mishor, Timna Soroka, Maya Finkel, Liav Tagania, Aharon Ravia, Ofer Perl, Edna Furman-Haran, Howard Carp, and Noam Sobel. \"unexplained repeated pregnancy loss is associated with altered perceptual and brain responses to men's body-odor\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002717.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[392] Ian Daly, Nicoletta Nicolaou, Duncan Williams, Faustina Hwang, Alexis Kirke, Eduardo Miranda, and Slawomir J. Nasuto. \"a dataset recording joint eeg-fmri during affective music listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002725.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 295, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 580 + ], + "type": "text", + "content": "[393] Jordi Sole-Casals, Josep M Serra-Grabulosa, Rafael Romero-Garcia, Gemma Vilaseca, Ana Adan, Nuria Vilaro, Nuria Vilaro, and Edward T. Bullmore. \"structural brain network of gifted children\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002726.v1.0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 624 + ], + "type": "text", + "content": "[394] Francesca M M Citron, Cristina Cacciari, Jakob M Funcke, Chun-Ting Hsu, and Arthur Jacobs. \"emotive idioms\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002727.v1.0.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "text", + "content": "[395] Bi Zhu, Chuansheng Chen, Xuhao Shao, Wenzhi Liu, Zhi-fang Ye, Liping Zhuang, Li Zheng, Elizabeth F. Loftus, and Gui Xue. \"multiple interactive memory representations underlie the induction of false memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002731.v1.0.2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 712 + ], + "type": "text", + "content": "[396] Nora C. Heinzelmann, Susanna C. Weber, and Philippe N. Tobler. OpenNeuro, 2020. doi: 10.18112/openneuro.ds002732.v1.0.0." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[397] Test. \"bids pilot project\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002733.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 139 + ], + "type": "text", + "content": "[398] Andrea Pisauro, Elsa Fouragnan, C Retzler, and M Philiastides. \"evidence accumulation in value-based decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002734.v1.0.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 184 + ], + "type": "text", + "content": "[399] Jonathan D Power, Benjamin Silver, Melanie Silverman, Eliana Ajodan, Dienke Bos, and Rebecca M. Jones. \"headmold\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002735.v1.0.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 554, + 217 + ], + "type": "text", + "content": "[400] Joset A. Etzel and Todd S. Braver. \"multibandcftests\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002737.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "text", + "content": "[401] Sabina Gherman and Marios G. Philiastides. \"simultaneous eeg-fmri - confidence in perceptual decisions\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002739.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 286 + ], + "type": "text", + "content": "[402] N Furl, F Begum, J Sulik, F Pizzorni Ferrarese, S Jans, and C Wooley. \"animated caricatures fmri study\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002741.v1.0.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 288, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 554, + 308 + ], + "type": "text", + "content": "[403] AR. \"test1\". OpenNeuro, 2020. doi: 10.18112/openneuro. ds002743.v1.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 311, + 554, + 354 + ], + "type": "text", + "content": "[404] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"resting state with closed eyes for patients with depression and healthy participants\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002748.v1.0.5." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 554, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 554, + 377 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 554, + 377 + ], + "type": "text", + "content": "[405] Chao-Gan Yan. \"yandatabids\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002750.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 379, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 379, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 379, + 554, + 488 + ], + "type": "text", + "content": "[406] Dillan J. Newbold, Timothy O. Laumann, Catherine R. Hoyt, Jacqueline M. Hampton, David F. Montez, Ryan V. Raut, Mario Ortega, Anish Mitra, Ashley N. Nielsen, Derek B. Miller, Babatunde Adeyemo, Annie L. Nguyen, Kristen M. Scheidter, Aaron B. Tanenbaum, Andrew N. Van, Scott Marek, Bradley L. Schlaggar, Alexandre R. Carter, Deanna J. Greene, Evan M. Gordon, Marcus E. Raichle, Steven E. Petersen, Abraham Z. Snyder, and Nico U.F. Dosenbach. \"cast-induced plasticity\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002766.v3.0.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 490, + 554, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 554, + 543 + ], + "type": "text", + "content": "[407] Damien Brevers, Claus Vögele, Pierre Maurage, Guillaume Sescousse, Chris Baeken, and Joel Billieux. \"brain mechanisms underlying episodic future thinking of sustainable behaviors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002770.v2.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 554, + 589 + ], + "type": "text", + "content": "[408] Zhifang Ye and Gui Xue. \"retrieval practice facilitates memory updating by enhancing and differentiating medial prefrontal cortex representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002773.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "text", + "content": "[409] Eva Berlot, Nicola Popp, and Jörn Diedrichsen. \"motor sequence learning\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002776.v1.0.2." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 667 + ], + "type": "text", + "content": "[410] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop1\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002785.v2.0.0." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[411] Lukas Snoek, Maite van der Miesen, Andries van der Leij, Tinka Beemsterboer, Annemarie Eigenhuis, and Steven Scholte. \"aomic-piop2\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002790.v2.0.0." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23861" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 22 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[412] Minjae Kim, Peter Mende-Siedlecki, Stefano Anzellotti, and Liane Young. \"tom following strong vs weak priors\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002793.v1.0.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[413] Jiefeng Jiang, Anthony D. Wagner, and Tobias Egner. \"elife39497\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002797.v1.0.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[414] Thompson WH*, Nair R*, Oya H*, Esteban O, Shine JM, Petkov CI, Poldrack RA, Howard M, Adolphs R†, and †corresponding author *equally contributing. \"human es-fmri resource: Concurrent deep-brain stimulation and whole-brain functional migraine\". OpenNeuro, 2021. doi: 10.18112/openneuro.d002799.v1.0.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[415] C.R. Bowman, T. Iwashita, and D. Zeithamova. \"model-based fmri reveals co-existing specific and generalized concept representations\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002813.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[416] Fatemeh Ebrahiminia, Morteza Mahdiani, and Seyed-Mahdi Khaligh-Razavi. \"a multimodal neuroimaging dataset to study spatiotemporal dynamics of visual processing in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002814.v1.3.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 352 + ], + "type": "text", + "content": "[417] Sangil Lee, Trishala Parthasarathi, and Joseph Kable. \"prospection\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso02835.v1.0.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[418] Sarah Aliko, Jiawen Huang, Florin Gheorghiu, Stefanie Meliss, and Jeremy I Skipper. \"naturalistic neuroimaging database\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002837.v2.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[419] Sarah Schwettmann, Joshua B. Tenenbaum, and Nancy Kanwisher. \"intuitive physics with fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002841.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 454 + ], + "type": "text", + "content": "[420] Omar Maxim. \"mriquality\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002848.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[421] Caroline Landelle, Jean-Luc Anton, Bruno Nazarian, Julien Sein, Ali Gharbi, Olivier Felician, and Anne Kavounoudias. \"aging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002872.v1.3.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[422] Christian Roming, Karl Koschutnig, Andreas Fink, and ... \"creative direccion making in soccer\". OpenNeuro, 2021. doi: 10.18112/openneuro.ps002878.v2.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 600 + ], + "type": "text", + "content": "[423] James R. Booth, Christine Brennan, Ozlem Ece Demir-Lira, Amy Desroches, Clara Ekerdt, Margaret M. Gullick, Marisa N. Lytle, Chris McNorgan, Melissa Randazzo-Wagner, Yael Weiss, and Jessica Wise Younger. \"brain correlates of multisensory lexical processing in adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds002879.v1.1.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[424] James R. Booth, John Binzak, Ozlem Ece Demir-Lira, Marisa Lytle, Rachna Mutreja, and Jérôme Prado. \"brain development of deductive reasoning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds002886.v1.1.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[425] Agnès Job, Chloé Jaroszynski, Anne Kavounoudias, Assia Jaillard, and Chantal Delon-Martin. \"resting state fmri study of non bothersome tinnitus\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002896.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[426] Sharna D. Jamadar, Phillip G.D. Ward, Thomas G. Close, Alex Fornito, Malin Premaratne, Kieren O'Brien, Daniel" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 106 + ], + "type": "text", + "content": "Staeb, Zhaolin Chen, N. Jon Shah, and Gary F. Egan. \"monash rspet-mr\". OpenNeuro, 2024. doi: 10.18112/openneuro.d002898.v1.4.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[427] Christophe Pallier and Laurent Bonnasse-Gahot. \"simpcomp_rawdata\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002905.v1.0.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[428] Rafael Garcia-Dias. \"neuroharmony sample example\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds002936.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[429] Bart Aben, Cristian Buc Calderon, Eva Van den Bussche, and Tom Verguts. \"task-dependent effort-induced connectivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002938.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "text", + "content": "[430] DC Monroe, RS Blumenfeld, SL Small, and DB Keator. \"head-to-ball impacts in collegiate soccer players\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds002940.v1.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "type": "text", + "content": "[431] Ruiqing Zhang. \"narrative listening\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs002979.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 310 + ], + "type": "text", + "content": "[432] Sangil Lee, Nicole Cooper, and Joseph Kable. \"delay discounting bidding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms002989.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 312, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 553, + 354 + ], + "type": "text", + "content": "[433] Gabriel Pelletier, Nadav Aridan, Lesley K Fellows, and Tom Schonberg. \"fribblesfmri_object-value-construction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds002994.v1.0.3." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 356, + 553, + 399 + ], + "type": "text", + "content": "[434] Jason A Avery, Alexander G Liu, Cameron Riddell, John E Ingelholm, Stephen J Gotts, and Alex Martin. \"taste quality representation in the human brain\". OpenNeuro, 2020. doi: 10.18112/openneuro.d002995.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 401, + 553, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 456 + ], + "type": "text", + "content": "[435] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"two sessions of resting state with closed eyes for patients with depression in treatment course (nfb, cbt or no treatment groups)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003007.v1.0.1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 457, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 521 + ], + "type": "text", + "content": "[436] Colin Hawco, Erin Dickie, Gabrielle Herman, Jessica Turner, Miklos Argyelan, Philipp Homan, Anil Malhotra, Robert Buchanan, and Aristotle Voineskos. \"social processes initiative in neurobiology of the schizophrenia(s) traveling human phantoms\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003011.v1.2.3." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 567 + ], + "type": "text", + "content": "[437] M. Martinez-Saito, R. Konovalov, M. A. Piradov, A. Shestakova, B. Gutkin, and V. Klucharev. \"action-in-auctions\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003012.v1.0.4." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 623 + ], + "type": "text", + "content": "[438] Amanda LeBel, Lauren Wagner, Shailee Jain, Aneesh Adhikari-Desai, Bhavin Gupta, Allyson Morgenthal, Jerry Tang, Lixiang Xu, and Alexander G. Huth. \"an fmri dataset during a passive natural language listening task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms003020.v3.0.0." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[439] Diego Angeles-Valdez, Jalil Rasgado-Toledo, Viviana Villicana, Alan Davalos-Guzman, Alfonso Fajardo-Valdez, Cristina Almanza, Ruth Alcala-Lozano, and Eduardo A. Garza-Villarreal. \"sudmex_tms\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003037.v2.1.0." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[440] Karen T. Navarro, Cheryl A. Olman, and Kimberly B. Weldon. \"dichoptic magno/parvo\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003043.v1.0.0." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23862" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 23 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[441] Krueger O and Klose U. \"dti data from 'fiber architecture in the ventromedial striatum and its relation with the bed nucleus of the stria terminalis\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003047.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "text", + "content": "[442] Robin Carhart-Harris et al. \"neural correlates of the lsd experience revealed by multimodal neuroimaging\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003059.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 194 + ], + "type": "text", + "content": "[443] James R. Booth, Marisa N. Lytle, Rachna Mureja, and Jérôme Prado. \"brain correlates of deductive reasoning in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003076.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 227 + ], + "type": "text", + "content": "[444] Philippe DOMENECH, Sylvain RHEIMS, and Etienne KOECHLIN. \"probe ieeg\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms003078.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[445] Jonathan Cote and Etienne de Villers-Sidani. \"auditory cortex mapping dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003082.v1.0.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 316 + ], + "type": "text", + "content": "[446] James R. Booth, Amy S. Desroches, Marisa N. Lytle, Rucha Mehta, Jennifer E. Minas, Rachna Mutreja, Jérôme Prado, and Hongchuan Zhang. \"brain correlates of math processing in adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003083.v1.0.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 348 + ], + "type": "text", + "content": "[447] Matthew Sachs, Assal Habibi, Antonio Damasio, and Jonas Kaplan. \"temporal dynamics of emotional music\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003085.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[448] Noam Saadon-Grosman, Shahar Arzy, and Yonatan Loewenstein. \"somatosensory phase-encoded bilateral full-body light touch stimulation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003089.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[449] Jeffrey N. Chiang, Yujia Peng, Hongjing Lu, Keith J. Holyoak, and Martin M. Monti. \"analogical reasoning sequential design fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003094.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 438, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 469 + ], + "type": "text", + "content": "[450] Y.C. Leong, J. Chen, R. Willer, and J. Zaki. \"polarization dataset\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003095.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "text", + "content": "[451] D Chung, MA Orloff, N Lauharatanahirun, B King-Casas, and Chiu PH. \"valuation of peers' safe choices is associated with substance-naivete in adolescents\". OpenNeuro, 2020. doi: 10.18112/openneuro.ps003096.v1.0.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 515, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 536 + ], + "type": "text", + "content": "[452] Anders Lillevik Thorsen. \"abcd protocol for bcbp\". Open-Neuro, 2020. doi: 10.18112/openneuro.ms003098.v1.0.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 591 + ], + "type": "text", + "content": "[453] HanShin Jo, Chiu-Yueh Chen, Der-Yow Chen, Ming-Heng Weng, and Chun-Chia Kung. \"consensus-seeking and conflict-resolving: an fmri study on college couples' shopping interaction\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003103.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "text", + "content": "[454] Lauri Parkkonen, Stefan Appelhoff, Alexandre Gramfort, Mainak Jas, and Richard Höchenberger. \"mne-somatodata-bids (anonymous)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003104.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 679 + ], + "type": "text", + "content": "[455] Jia-Xi Wang, Jin-Ying Zhuang, Lulu Fu, and Lei Qin. \"how ovarian hormones influence the behavioral activation and inhibition system through the dopamine pathway\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003114.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[456] Saloni Krishnan and Kate Watkins. \"bold verb generation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003145.v1.0.2." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "text", + "content": "[457] Arafat Angulo-Perkins and Luis Concha. \"vocal task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003146.v1.0.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "text", + "content": "[458] Adam Hampshire and Eyal Soreq. \"neuroimaging evidence for network sampling theory of human intelligence\". Open-Neuro, 2020. doi: 10.18112/openneuro.ds003148.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "text", + "content": "[459] Alexandra Ycaza Herrera, Padideh Nasseri, Ricardo Velasco, Sophia Faude, Jessica D. White, Philipp C. Opitz, Ringo Huang, Kristie Tu, and Mara Mather. \"stress-associated brain activation across the hormonal contraceptive cycle\". OpenNeuro, 2021. doi: 10.18112/openneuro. ds003151.v2.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 206, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 238 + ], + "type": "text", + "content": "[460] Chang-Hao Kao, Sangil Lee, Joshua I. Gold, and Joseph W. Kable. \"changepoint fmri\". OpenNeuro, 2020. doi: 10.18112/openneuro.d0s003170.v2.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 304 + ], + "type": "text", + "content": "[461] Sivayini Kandeepan, Jorge Rudas, Francisco Gomez, Bobby Stojanoski, Sreeram Valluri, Adrian Mark Owen, Lorina Naci, Andrea Soddu, and Emily Sophia Nichols. \"modeling an auditory stimulated brain under altered states of consciousness using the generalized ising model\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003171.v2.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 360 + ], + "type": "text", + "content": "[462] Simon M. Ceh, Sonja Annerer-Walcher, Karl Koschutnig, Christof Korner, and Mathias Benedek. \"neurophysiological indicators of internal attention: An fmri-eye-tracking coregistration study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003176.v2.0.1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 392 + ], + "type": "text", + "content": "[463] Stefano Moia, Eneko Urñuela, Vicente Ferrer, and César Caballero-Gaudes. \"euskalibur\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003192.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 469 + ], + "type": "text", + "content": "[464] Kenshu Koiso, Sebastian Dresbach, Christopher J Wiggins, Omer Faruk Gulban, Anna Katharina Mueller, Miriam Heynckes, Yuhui Chai, Yoichi Miyawaki, Benedikt A Poser, and Renzo Huber. \"acquisition and processing methods of whole-brain layer-fmri vaso and bold: The kenshu dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003216.v3.0.11." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[465] Meir Meshulam, Liat Hasenfratz, Hanna Hillman, Yun-Fei Liu, Mai Nguyen, Kenneth A. Norman, and Uri Hasson. \"think like an expert\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003233.v1.2.1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 580 + ], + "type": "text", + "content": "[466] Livia Tomova, Kimberly L. Wang, Todd Thompson, Gillian A. Matthews, Kay M. Tye, and Rebecca Saxe. \"mri data of 40 adult participants in response to a cue induced craving task following food fasting, social isolation and baseline (within-subject design)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003242.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[467] James Antony, Thomas Hartshorne, Ken Pomeroy, Todd Gureckis, Uri Hasson, Samuel McDougle, and Kenneth Norman. \"behavioral, physiological, and neural signatures of surprise during naturalistic sports viewing\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003338.v1.1.0." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[468] Jason A Avery, Alexander G Liu, John E Ingeholm, Stephen J Gotts, and Alex Martin. \"tasting pictures: Viewing images of foods evokes taste-quality-specific activity in gustatory insular cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003340.v1.0.4." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[469] Ethan Knights, Courtney Mansfield, Diana Tonin, Janak Saada, Fraser W. Smith, and Stephanie Rossit. \"hand-" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23863" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 24 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "selective visual regions represent how to grasp 3d tools for use: brain decoding during real actions\". OpenNeuro, 2020. doi: 10.18112/openneuro.rs003342.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[470] Kelsey McDonald, William Broderick, Scott Huettel, and John Pearson. \"penaltykik.02\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003345.v1.0.2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[471] Eduardo A. Garza-Villarreal, Jorge Julio Gonzalez Olvera, Thania Balducci, Diego Angeles Valdez, Alely Valencia, and Jalil Rasgado. \"sudmex Conn: The mexican dataset of cocaine use disorder patients.\" OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003346.v1.1.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 230 + ], + "type": "text", + "content": "[472] Mikkel Schöttner and Andreas Jansen. \"empatom with ppi-r, tas, and eq\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003354.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 274 + ], + "type": "text", + "content": "[473] Richard Huskey, Justin Robert Keene, Shelby Wilcox, Xuanjun (Jason) Gong, Robyn Adams, and Christina J. Najera. \"brain dynamics during flow experiences\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003358.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 331 + ], + "type": "text", + "content": "[474] Samuel B. Snider, Yelena G. Bodien, Aina Frau-Pascual, Marta Bianciardi, Andrea S. Foulkes, and Brian L. Edlow. \"ascending arousal network connectivity during recovery from traumatic coma\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003367.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[475] Heath R Pardoe, Samantha P Martin, Allan George, and Orrin Devinsky. \"estimation of in-scanner head pose changes during structural mri using a convolutional neural network trained on eye tracker video\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003381.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[476] Sharna Jamadar, Shenjun Zhong, Phillip Ward, Alexandra Carey, Richard McIntyre, Alex Fornito, Malin Premaratne, N Jon Shah, Kieran O'Brien, Daniel Stab, Zhaolin Chen, and Gary Egan. \"monash vis-fpet-fmri\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003382.v1.5.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[477] Nicolas Zilber, Philippe Ciuciu, Alexandre Gramfort, Leila Azizi, and Virginie van Wassenhove. \"neurospin hmt+localizer data (meg & amri)\". OpenNeuro, 2021. doi: 10.18112/openneuro.rs003392.v1.0.4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 543 + ], + "type": "text", + "content": "[478] Sharna D Jamadar, Emma X Liang, Phillip GD Ward, Alexandra Carey, Richard McIntyre, Zhaolin Chen, Shenjun Zhong, and Gary F Egan. \"monash dacra fpet-fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003397.v1.2.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[479] Jungtak Park, Karolina Janacsek, Dezso Nemeth, and Hyeon-Ae Jeon. \"asrt (alternating serial reaction time)\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003401.v1.0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[480] Brian P Keane, Deanna M Barch, Ravi Mill, Steven M Silverstein, Bart Krekelberg, and Michael W Cole. \"brain network mechanisms of visual shape completion\". Open-Neuro, 2025. doi: 10.18112/openneuro.ds003404.v1.0.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[481] Maria Geraldine Veldhuizen, Michael C. Farruggia, Xiao Gao, Yuko Nakamura, Barry G. Green, and Dana Small. \"identification of an amygdala-thalamic circuit that acts as a central gain mechanism in taste perception\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003424.v1.0.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[482] Steven Greening, Tae-Ho Lee, Laurent Gregoire, Lauryn Burleigh, Tyler Robinson, Xinrui Jiang, Mara Mather, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "Jonas Kaplan. \"mental imagery can generate and regulate acquired differential fear conditioned reactivity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003425.v1.0.2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 148 + ], + "type": "text", + "content": "[483] Tomoyasu Horikawa and Yukiyasu Kamitani. \"attentionally modulated subjective images reconstructed from brain activity\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003430.v1.2.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 183 + ], + "type": "text", + "content": "[484] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 localizer\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003433.v1.0.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 185, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 226 + ], + "type": "text", + "content": "[485] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 main experiment\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003434.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 270 + ], + "type": "text", + "content": "[486] Sandra Arbula, Elisabetta Pisanu, and Raffaella I. Rumiati. \"agreeableness personality trait and social information encoding\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003436.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 348 + ], + "type": "text", + "content": "[487] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain - diet-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003437.v1.0.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[488] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant1 suboptimal block designs\". OpenNeuro, 2020. doi: 10.18112/openneuro.ms03438.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 436 + ], + "type": "text", + "content": "[489] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant2 suboptimal slow event related\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003439.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[490] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant3 caffeine\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003440.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 524 + ], + "type": "text", + "content": "[491] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusicac. \"newbi4fmri2020 variant6 adaptation\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003443.v1.0.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "type": "text", + "content": "[492] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant7 prosopagnosia\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003444.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[493] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebekka Lagace Cusiac. \"newbi4fmri2020 variant8 resting state\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003445.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 656 + ], + "type": "text", + "content": "[494] Jody Culham, Kevin Stubbs, Ethan Jackson, and Rebecca Lagace Cusiac. \"newbi4fmri2020 variant9 7t\". OpenNeuro, 2020. doi: 10.18112/openneuro.ds003446.v1.0.0." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[495] Joset A. Etzel, Mitch Jeffers, and Todd S. Braver. \"dmcc13benchmark\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003452.v1.0.1." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[496] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Lithen," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23864" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 25 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "text", + "content": "Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, bolus-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003453.v1.0.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 172 + ], + "type": "text", + "content": "[497] Andrea Stocco, Chantel S. Prat, and Lauren K. Graham. \"individual differences in fluid reasoning and rapm-like problem solving\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003454.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 249 + ], + "type": "text", + "content": "[498] Lilianne R. Mujica-Parodi, Anar Amgalan, Syed Fahad Sultan, Botond Antal, Steven Skiena, Andrew Litten, Noor Adra, Eva-Maria Ratai, Corey Weistuch, Sindhuja Tirumalai Govindarajan, Helmut H. Strey, Ken A. Dill, Steven M. Stufflebeam, Richard L. Veech, and Kieran Clarke. \"protecting the aging brain, case-study\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003455.v1.0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 304 + ], + "type": "text", + "content": "[499] James R. Booth, Douglas D. Burman, Darren R. Gitelman, Marisa N. Lytle, Marsel M. Mesulam, Joel R. Meyer Jr., and Todd R. Parrish. \"component processes of word reading in adults and children\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003459.v1.0.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 337 + ], + "type": "text", + "content": "[500] Todd S. Braver, Alexander Kizhner, Rongxiang Tang, Michael C. Freund, and Joset A. Etzel. \"dmcc55b\". Open-Neuro, 2025. doi: 10.18112/openneuro.ms003465.v1.0.7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 403 + ], + "type": "text", + "content": "[501] Jiajia Yang, Peter J. Molfese, Yinghua Yu, Daniel A. Handwerker, Gang Chen, Paul A. Taylor, Yoshimichi Ejima, Jinglong Wu, and Peter A. Bandettini. \"haptic three-dimensional curved surface exploration fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003466.v1.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "text", + "content": "[502] Ekaterina V. Pechenkova, Yana R. Panikratova, Maria A. Fomina, Alena D. Rumshiskaya, Darya A. Bazhenova, Liudmila A. Makovskaya, Irina S. Lebedeva, and Valentin E. Sinitsyn. \"speech disfluencies: Neurophysiological aspect in normal population\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003469.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[503] Anja Ischebeck, Hannah Hiebel, Joe Miller, Margit Höfler, Iain D. Gilchrist, and Christof Körner. \"target processing in overt serial visual search involves the dorsal attention network: A fixation-based event-related fmri study\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003470.v2.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 580 + ], + "type": "text", + "content": "[504] J. Rasgado-Toledo, F. Lizcano-Cortés, V. Olalde-Mathieu, M. Zamora-Ursulo, G. Licea-Haquet, A. Carillo-Pena, E. Navarrete, A.* Reyes-Aguilar, and M Giordano. \"pragmatic language\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003481.v1.0.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[505] Brian Silston, Toby Wise, Song Qi, Xin Sui, Peter Dayan, and Dean Mobbs. \"foraging in competitive and hazardous environments\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003484.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 690 + ], + "type": "text", + "content": "[506] Eva R. Pool, David Munoz Tord, Sylvain Delplanque, Yoann Stussi, Donato Cereghetti, Patrik Vuilleumier, and David Sander. \"differential contributions of ventral striatum subregions in the motivational and hedonic components of the affective response to reward\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003487.v2.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[507] Lukas Snoek and Steven Scholte. \"dstreamupgrade\". Open-Neuro, 2021. doi: 10.18112/openneuro.ds003495.v1.0.0." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[508] Kate Nussenbaum and Catherine A. Hartley. \"developmental change in prefrontal cortex recruitment supports the emergence of value-guided memory\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03499.v1.0.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[509] David Pascucci, Sebastien Tourbier, Joan Rue-Queralt, Margherita Carboni, Patric Hagmann, and Gijs Plomp. \"vepcon: Source imaging of high-density visual evoked potentials with multi-scale brain parcellations and connectomes\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003505.v1.1.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 183, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 226 + ], + "type": "text", + "content": "[510] Chuanji Gao, Christine E. Weber, Douglas H. Wedell, and Svetlana V. Shinkareva. \"fmri: Audiovisual valence congruence\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003507.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[511] Mikael Novén, Hampus Olsson, Gunther Helms, Markus Nilsson, Merle Horne, and Mikael Roll. \"language learning aptitude dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003508.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "type": "text", + "content": "[512] Adrian W. Gilmore, Alina Quach, Sarah E. Kalinowski, Estefanía I. Gonzalez-Araya, Stephen J Gotts, Daniel L. Schacter, and Alex Martin. \"evidence supporting a timelimited hippocampal role in retrieving autobiographical memories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003511.v1.1.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[513] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study2\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso03521.v2.2.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 425 + ], + "type": "text", + "content": "[514] Luke Chang, Eshin Jolly, Jin Hyun Cheong, Kristina Rapuano, Nathan Greenstein, Pin-Hao Chen, and Jeremy Manning. \"fridaynightlights_study1\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03524.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 491 + ], + "type": "text", + "content": "[515] Benjamin Risk, Raphiel Murden, Junjie Wu, Mary Beth Nebel, Arun Venkataraman, Zhengwu Zhang, and Deqiang Qiu. \"which multiband factor should you choose for your resting-state fmri study? the emory multiband dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003540.v1.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "text", + "content": "[516] Kshipra Gurunandan, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adult language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.d0s003542.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 568 + ], + "type": "text", + "content": "[517] Kshipra Gurunandan, Jaione Arnaez-Telleria, Manuel Carreiras, and Pedro M. Paz-Alonso. \"adolescent language learners\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003545.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[518] Isaac David, Victor Olalde-Mathieu, Ana Y. Martínez, Lluviana Rodríguez-Vidal, and Fernando A. Barrios. \"emotion category and face perception task optimized for multivariate pattern analysis\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms03548.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[519] Renee M Visser, H Steven Scholte, Tinka Beemsterboer, and Merel Kindt. \"visser, scholte, beemsterboer, & kindt (2013) nature neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003550.v1.0.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[520] Renee M Visser, H Steven Scholte, and Merel Kindt. \"visser, scholte & kindt (2011) journal of neuroscience\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003553.v1.0.2." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23865" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 26 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[521] Renee M Visser, Anna E. Kunze, Biance Westhoff, H Steven Scholte, and Merel Kindt. \"visser et al. 2015 psychoneuroendocrinology\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003554.v1.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 185 + ], + "type": "text", + "content": "[522] Falk Luesebrink, Hendrik Mattern, Renat Yakupov, Julio Acosta-Cabronero, Mohammad Ashtarayeh, Steffen Oeltze-Jafra, and Oliver Speck. \"data from: Comprehensive ultrahigh resolution whole brain in vivo mri dataset as a human phantom\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003563.v1.1.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 231 + ], + "type": "text", + "content": "[523] Lucrezia Liuzzi, Katharine Chang, Hanna Keren, Charles Zheng, Dipta Saha, Dylan Nielson, and Argyris Stringaris. \"mood induction in mdd and healthy adolescents\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds003568.v1.0.4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 266 + ], + "type": "text", + "content": "[524] Jeahong Kim, Jeyoung Jung, and Kichun Nam. \"transposition confusability during visual word recognition\". Open-Neuro, 2021. doi: 10.18112/openneuro.ms003569.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 322 + ], + "type": "text", + "content": "[525] Virginie Sterpenich, Mojca KM van Schie, Maximilien Catsiyannis, Avinash Ramyead, Stephen Perrig, Hee-Deok Yang, Dimitri Van De Ville, and Sophie Schwartz. \"reward biases spontaneous neural reactivation during sleep\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003574.v1.0.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 325, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 401 + ], + "type": "text", + "content": "[526] R. Nathan Spreng, Roni Setton, Udi Alter, Benjamin N. Cassidy, Bri Darboh, Elizabeth DuPre, Karin Kantarovich, Amber W. Lockrow, Laetitia Mwilambwe-Tshilobo, WenMing Luh, Prantik Kundu, and Gary R. Turner. \"neurocognitive aging data release with behavioral, structural, and multi-echo functional mri measures\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso03592.v1.0.13." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[527] Xingyu Liu, Yuxuan Dai, Hailun Xie, and Zonglei Zhen. \"forrestgump-meg\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003633.v1.0.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 472 + ], + "type": "text", + "content": "[528] Heath R Pardoe and Samantha P Martin. \"in-scanner head motion and structural covariance networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003639.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 517 + ], + "type": "text", + "content": "[529] Michele Svanera and Lars Muckli. \"cerebrum-7t: Fast and fully-volumetric brain segmentation of 7 tesla mr volumes\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003642.v1.1.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[530] Jixing Li, John Hale, and Christophe Pallier. \"le petit prince: A multilingual fmri corpus using ecological stimuli\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds003643.v2.0.7." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 632 + ], + "type": "text", + "content": "[531] Ph.D. Anna Manelis, Ph.D. David A.A. Baranger, B.S. Skye Satz, M.A. Rachel Ragozzino, Ph.D. Satish Iyengar, M.D. Holly A. Swartz, and Ph.D. Yaroslav O. Halchenko. \"cortical myelin measured by the t1w/t2w ratio in individuals with depressive disorders and healthy controls\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003653.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "text", + "content": "[532] Arielle Tambini, Jacob Miller, Anastasia Kiyonaga, and Mark D'Esposito. \"scantrain\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs003659.v2.0.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[533] Mohamed Abdelhack and Yukiyasu Kamitani. \"sharpening of hierarchical visual feature representations of blurred images\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003661.v1.0.0." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[534] Matan Mazor, Karl J. Friston, and Stephen M. Fleming. \"confidence in detection and discrimination\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003669.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[535] Kangjoo Lee, Corey Horien, David O'Connor, Bronwen Garand-Sheridan, Fuyuze Tokoglu, Dustin Scheinost, Evelyn MR Lake, and Todd R Constable. \"yale resting state fmri/pupillometry: Arousal study\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003673.v2.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "text", + "content": "[536] Giacomo Ariani, J Andrew Pruszynski, and Jorn Diedrichsen. \"finger sequence planning\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003684.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[537] Julia Berezutskaya, Mariska J. Vansteensel, Erik J. Aarnoutse, Zachary V. Freudenburg, Giovanni Piantoni, Mariana P. Branco, and Nick F. Ramsey. \"open multimodal ieeg-fmri dataset from naturalistic stimulation with a short audiovisual film\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003688.v1.0.7." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "text", + "content": "[538] Javier Ortiz-Tudela, Johanna Bergman, Matt Bennett, Isabelle Ehrich, Lars Muckli, and Yee Lee Shing. \"feedbes - feedback signals from episodic and semantic memories.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003691.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 348 + ], + "type": "text", + "content": "[539] Shan Xu, Xingyu Liu, Jorge Almeida, and Dietmar Heinke. \"action-related object pairs - fmri dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003696.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[540] Takuya Ito, Emily Winfield, Nicole Lalta, Kaustubh R. Kulkarni, Douglas H. Schultz, Ravi D. Mill, Richard H. Chen, Levi Solomyak, and Michael W. Cole. \"concrete permuted rule operations\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003701.v1.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[541] Guo Wanjia, Serra E. Favila, Ghootae Kim, Robert J. Moltor, and Brice A. Kuhl. \"abrupt hippocampal remapping signals resolution of memory interference.\" OpenNeuro, 2021. doi: 10.18112/openneuro.ds003707.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[542] Hanna Keren, Katharine Chang, Aria Vitale, Chris C. Camp, Robb B Rutledge, Dylan M. Nielson, and Argyris Stringaris. \"nimh-compypsych mmi\". OpenNeuro, 2021. doi: 10.18112/openneuro.dso03709.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[543] Daniel K. Bjornn, Julie Van, and C. Brock Kirwan. \"the contributionsof eye gaze fixations and target-lure similarity to behavioral and fmri indices of pattern separation and pattern completion\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003711.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 569 + ], + "type": "text", + "content": "[544] Avijit Chowdhury. \"test\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003714.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[545] Li J., Curley W. H., Guerin B., Dougherty D. D., Dalca A. V., Fischl B., Horn A., and Edlow B. L. \"subcortical dmn functional connectivity\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003716.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 678 + ], + "type": "text", + "content": "[546] Peelle JE, Spehar B, Jones MS, McConkey S, Myerson J, Hale S, Sommers MS, and Tye-Murray N. \"visual and audiovisual speech perception associated with increased functional connectivity between sensory and motor regions\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003717.v1.1.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[547] Tomoya Nakai, Naoko Koide-Majima, and Shinji Nishimoto. \"music genre fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003720.v1.0.1." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23866" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 27 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 715 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[548] Renee M Visser, Richard N Henson, and Emily A Holmes. \"a naturalistic paradigm to investigate post-encoding neural activation patterns in relation to subsequent voluntary and intrusive recall of distressing events\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003721.v1.0.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[549] David V. Smith, Rita M. Ludwig, Jeffrey B. Dennison, Crystal Reeck, and Dominic S. Fareri. \"an fmri dataset on social reward processing and decision making in younger and older adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds003745.v2.1.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[550] michael barnett, geoffrey aguirre, and david brainard. \"If contrast\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003752.v1.0.0." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[551] Kenji Kobayashi, Sangil Lee, Alexandre L. S. Filipowicz, Kara D. McGaughey, Joseph W. Kable, and Matthew R. Nassar. \"dynamic representation of the subjective value of information\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003758.v1.0.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 341 + ], + "type": "text", + "content": "[552] Olga R. Dobrushina, Larisa A. Dobrynina, Galina A. Arina, Maria V. Gubanova, Anastasia V. Belopasova, Evgenia S. Novikova, Anastasia D. Suslina, Elena I. Kremneva, Ekaterina V. Pechenkova, and Marina V. Krotenkova. \"interception during aging: The heartbeat detection task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003763.v1.0.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[553] Zhenzhong Gan, Suiping Wang, and Gangyi Feng. \"functional magnetic resonance imaging data for the neural dynamics underlying the acquisition of distinct auditory categories\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003764.v1.0.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[554] Yameng Gu, Feng Han, Lucas E. Sainburg, Margeaux M. Schade, and Xiao Liu. \"simultaneous eeg and fmri signals during sleep from humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s003768.v1.0.11." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 487 + ], + "type": "text", + "content": "[555] Bezmaternykh DD, Melnikov ME, Savelov AA, and Petrovskii ED. \"depression treatment by medial prefrontal real-time fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms003770.v1.2.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[556] Joseph T. McGuire, Matthew R. Nassar, Joshua I. Gold, and Joseph W. Kable. \"functionally dissociable influences on learning rate in a dynamic environment\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003772.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 599 + ], + "type": "text", + "content": "[557] Redouane Jamil, Franck Mauconduit, Caroline Le Ster, Philipp Ehses, Benedikt A. Poser, Alexandre Vignaud, and Nicolas Boulant. \"temporal snr optimization through rf coil combination in fmri: The more, the better? - dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003777.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[558] Yufei Zhao, Avi Chanales, and Brice Kuhl. \"adaptive memory distortions are predicted by feature representations in parietal cortex\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003778.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[559] Nadav Aridan, Nicholas J. Malecek, Russell A. Poldrack, and Tom Schonberg. \"neural correlates of effort-based valuation under risk\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003782.v1.0.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 715 + ], + "type": "text", + "content": "[560] Marc Himmelberg, Jan Kurzawski, Noah Benson, Denis Pelli, Marisa Carrasco, and Jon Winawer. \"nyu retinotopy" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "dataset\". OpenNeuro, 2021. doi: 10.18112/openneuro.d003787.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[561] Xuhao Shao, Chuansheng Chen, Elizabeth F. Loftus, Gui Xue, and Bi Zhu. \"dynamic changes in neural representations underlie the repetition effect on false memory\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds003789.v2.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 142, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 553, + 251 + ], + "type": "text", + "content": "[562] Eva Mishor, Daniel Amir, Tali Weiss, Danielle Honigstein, Aharon Weissbrod, Ethan Livne, Lior Gorodisky, Shiri Karagach, Aharon Ravia, Kobi Snitz, Diyala Karawani, Rotem Zirler, Reut Weissgross, Timna Soroka, Yaara Endevelt-Shapira, Shani Agron, Liron Rozenkrantz, Netta Reshef, Edna Furman-Haran, Heinz Breer, Joerg Strotmann, Tatsuya Uebi, Mamiko Ozaki, and Noam Sobel. \"sniffing the human body-volatile hexadecanal blocks aggression in men but triggers aggression in women\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003791.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 340 + ], + "type": "text", + "content": "[563] Dorit Kliemann, Ralph Adolphs, Tim Armstrong, Paola Galdi, David A. Kahn, Tessa Rusch, A. Zeynep Enkavi, Deuhua Liang, Steven Lograsso, Wenying Zhu, Rona Yu, Remya Nair*, Lynn K. Paul*, J. Michael Tyszka*, and * Joint senior authorship. \"caltech conte center - a multimodal data resource for exploring social cognition and decision-making\". OpenNeuro, 2021. doi: 10.18112/openneuro.dss003798.v1.0.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 342, + 553, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 451 + ], + "type": "text", + "content": "[564] Hyun Joo Yoo, Kaoru Nashiro, Jungwon Min, Christine Cho, Noah Mercer, Shelby L. Bachman, Padideh Nasseri, Shubir Dutt, Shai Porat, Paul Choi, Yong Zhang, Vardui Grigoryan, Tiantian Feng, Julian F. Thayer, Paul Lehrer, Catie Chang, Jeffrey A. Stanley, Elizabeth Head, Jeremy Rouanet, Vasilis Z. Marmarelis, Shrikanth Narayanan, Jessica Wisnowski, Daniel A. Nation, and Mara Mather. \"heart rate variability biofeedback training and emotion regulation\". OpenNeuro, 2025. doi: 10.18112/openneuro.d003823.v1.3.5." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 518 + ], + "type": "text", + "content": "[565] Michal Rafal Zareba, Magdalena Fafrowicz, Tadeusz Marek, Ewa Beldzik, Halszka Oginska, and Aleksandra Domagalik. \"structural (t1) images of 136 young healthy adults; study of effects of chronotype, sleep quality and daytime sleepiness on brain structure\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003826.v3.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "text", + "content": "[566] Kevin Fialkowski, Ivan Messias, and Keith A. Bush. \"cognitive control theoretic mechanisms of real-time fmriguided neuromodulation (ctm)\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003831.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 566, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 620 + ], + "type": "text", + "content": "[567] Matteo Visconti di Oleggio Castello, James V. Haxby, and M. Ida Gobbini. \"shared neural codes for visual and semantic information about familiar faces in a common representational space\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003834.v1.0.2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 654 + ], + "type": "text", + "content": "[568] Alyssa H Sinclair and Morgan D Barense. \"pe-update\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003835.v1.0.2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "text", + "content": "[569] Flavia Mancini, Suyi Zhang, and Ben Seymour. \"learning the statistics of pain\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs003836.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[570] van Blooijs D., Demuru M., Zweiphenning W, Hermes D., Leijten F., and Zijlmans M. \"dataset clinical epilepsy ieeg" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23867" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 28 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 95 + ], + "type": "text", + "content": "to bids - respect_longterm_jeeg\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003848.v1.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "type": "text", + "content": "[571] Austin L. Boroshok, Anne T. Park, Panagiotis Fotiadis, Gerardo H. Velasquez, Ursula A. Tooley, Katrina R. Simon, Jasmine C.P. Forde, Lourdes M. Delgado Reyes, M. Dylan Tisdall, Dani S. Bassett, Emily A. Cooper, and Allyson P. Mackey. \"individual differences in frontoparietal plasticity in humans\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003849.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[572] S.R. Ashby and D. Zeithamova. \"category-biased neural representations form spontaneously during learning that emphasizes memory for specific instances\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003851.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[573] Leili Mortazavi, Tara Srirangarajan, Tiago Bortolini, Jorge Moll, and Brian Knutson. \"monetary incentive delay task\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003858.v1.0.1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[574] Christopher N. Wahlheim, Alexander P. Christensen, Brittany S. Cassidy, and Zachariah M. Reagh. \"resting-state for 34 younger and 28 older adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003871.v1.0.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 353 + ], + "type": "text", + "content": "[575] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2021. doi: 10.18112/openneuro.ms003872.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 409 + ], + "type": "text", + "content": "[576] Ted Turesky, Talat Shama, Shahriah Hafiz Kakon, Rashidul Haque, Nazrul Islam, Amala Someshwar, Borjan Gagoski, William Petri, Charles Nelson, and Nadine Gaab. \"brain-morphometry diminishedgrowthBeanstudy_2021\". OpenNeuro, 2022. doi: 10.18112/openneuro.d003877.v1.1.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "text", + "content": "[577] Wen Wen, Yue Wang, jiawei Zhou, Sheng He, Xinghuai Sun, Hong Liu, Chen Zhao, and Peng Zhang. \"loss and enhancement of layer-selective signals in geniculostriate and corticotectal pathways of adult human amblyopia\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003892.v5.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 521 + ], + "type": "text", + "content": "[578] Philippe Poulin, Guillaume Thead, Pierre-Marc Jodoin, and Maxime Descoteaux. \"tractoinferno: A large-scale, open-source, multi-site database for machine learning dmri tractography\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003900.v1.1.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 555 + ], + "type": "text", + "content": "[579] J. Pesnot Lerousseau, C. Parise, MO. Ernst, and V. van Wassenhove. \"multisensory correlation detector\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003922.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "text", + "content": "[580] Ning Mei, Roberto Santana, and David Soto. \"informative neural representations of unseen contents during higher-order processing in human brains and deep artificial networks\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003927.v1.0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 668 + ], + "type": "text", + "content": "[581] Tommaso Di Noto, Guillaume Marie, Sebastien Tourbier, Yasser Aleman-Gómez, Oscar Esteban, Guillaume Saliou, Meritxell Bach Cuadra, Patric Hagmann, and Jonas Richiardi. \"lausanne_tof-mra_aneurysm_cohort\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003949.v1.0.1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[582] Lukas Snoek, Jessica Loke, Michelle Kuehn, Mona Zimmermann, and Steven Scholte. \"ni-edu-data-minimal\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003965.v1.0.0." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[583] Olivier Girard, Ludovic de Rochefort, Maxime Guye, Jean-Philippe Ranjeva, and Arnaud Le Troter. \"7tamibrain\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003967.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[584] W. Schellekens, C. Bakker, N.F. Ramsey, and N. Petridou. \"body motor\". OpenNeuro, 2021. doi: 10.18112/openneuro.ds003972.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[585] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 11 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03974.v3.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[586] Ping Li, Chun-Ting Hsu, Ben Schloss, Anya Yu, Lindsey Ma, Marissa Scotto, Friederike Seyfried, and Chanyuan Gu. \"the reading brain project 12 adults\". OpenNeuro, 2022. doi: 10.18112/openneuro.ps03988.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[587] Sarah Khalife, Susan T. Francis, Denis Schluppeck, Rosa Maria Sanchez-Panchuelo, and Julien Besle. \"fastvsslowertactile7t\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003990.v1.0.2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 330 + ], + "type": "text", + "content": "[588] Jun Kai Ho, Tomoyasu Horikawa, Kei Majima, and Yukiyasu Kamitani. \"inter-individual deep image reconstruction\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003993.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 386 + ], + "type": "text", + "content": "[589] PhD) Daminov V. (MD, MSc) Novak E. (MD, Slepyova N. (MD), Mikhailov D. (MSc), and Karpulevich E. (MSc). \"pre-post rehabilitation fmri data of post-stroke patients\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds003999.v1.0.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[590] J. Wammes, K.A. Norman, and N.B. Turk-Browne. \"Increasing stimulus similarity drives nonmonotonic representational change in hippocampus\". OpenNeuro, 2022. doi: 10.18112/openneuro.d0s004006.v1.0.4." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[591] J.A. Williams, E.H. Margulis, S.A. Nastase, J. Chen, U. Hasson, K.A. Norman, and C. Baldassano. \"high-order areas and auditory cortex both represent the high-level event structure of music\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004007.v1.0.2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 521 + ], + "type": "text", + "content": "[592] Robert M. Mok and Bradley C. Love. \"abstract-category-signal\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004009.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[593] Dominic Standage, Joseph Y. Nashed, J. Randall Flanagan, and Jason. P Gallivan. \"visuomotor rotation adaptation experiment\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004021.v1.0.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 623 + ], + "type": "text", + "content": "[594] Julio Cesar Hernandez Pavon, Nils Schneider Garces, John Patrick Begnoche, Lee Miller, and Tommi Rajj. \"tmseeg-mri-fmri-dwi data on paired associative stimulation and connectivity (shirley ryan abilitylab, chicago, il)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004024.v1.0.1." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[595] Cammille C Go, Huseyin O Taskin, Seyed Ahmad Ahmadi, Giulia Frazzetta, Laura Cutler, Saguna Malhotra, Jessica IW Morgan, Virginia L Flanagin, and Geoffrey K Aguirre. \"tome-go2022vestibularmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04038.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[596] Hongmi Lee, Janice Chen, and Uri Hasson. \"filmfestival\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004042.v1.0.1." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23868" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 29 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[597] Sai Ma, Taicheng Huang, Yukun Qu, Xiayu Chen, Yajie Zhang, and Zonglei Zhen. \"an fmri dataset for whole-body somatotopic mapping in humans\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004044.v2.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 183 + ], + "type": "text", + "content": "[598] Miao Zhang, Qikai Qin, Shuning Zhang, Wei Liu, Hongping Meng, Mengyang Xu, Xinyun Huang, Xiaozhu Lin, Mu Lin, Peter Herman, Fahmeed Hyder, Raymond C. Stevens, Zheng Wang, Biao Li, and Garth J. Thompson. \"aerobic glycolysis imaging pet-mri\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004054.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[599] Yoshihisa Nakayama, Sho K. Sugawara, Masaki Fukunaga, Yuki H. Hamano, Norihiro Sadato, and Yukio Nishimura. \"goal-directed motor task\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004056.v1.0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[600] Shai Sabbah, Michael S. Worden, Dimitrios D. Laniado, David M. Berson, and Jerome N. Sanes. \"luminance\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004065.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 304 + ], + "type": "text", + "content": "[601] Matan Mazor, Chudi gong, and Stephen M. Fleming. \"confidence in equal and unequal variance settings\". Open-Neuro, 2022. doi: 10.18112/openneuro.ds004081.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 338 + ], + "type": "text", + "content": "[602] Julie Van, Samuel Nielson, and C. Brock Kirwan. \"triple dissociation revisited\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004086.v1.2.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 371 + ], + "type": "text", + "content": "[603] Martin Szinte, Gilles de Hollander, Marco Aqil, Serge Dumoulin, and Tomas Knapen. \"gaze prf\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004091.v2.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 437 + ], + "type": "text", + "content": "[604] Athula Pudhiyidath, Neal W Morton, Rodrigo Viveros Duran, Anna C. Schapiro, Ida Momennejad, Demitrius M. Hinojosa-Rowland, Robert J. Molitor, and Alison R. Preston. \"effects of temporal community structure learning on reasoning decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004094.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 525 + ], + "type": "text", + "content": "[605] Kristina M. Rapuano, May I. Conley, Anthony C. Juliano, Gregory M. Conan, Maria T. Maza, Kylie Woodman, Steven A. Martinez, Eric Earl, Anders Perrone, Eric Feczko, Damien A. Fair, Richard Watts, BJ Casey, and Monica D. Rosenberg. \"an open-access accelerated adult equivalent of the abcd study neuroimaging dataset (a-abcd)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004097.v1.1.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 580 + ], + "type": "text", + "content": "[606] Dobrushina Olga R, Dobrunina Larisa A, Kremneva Elena I, and Pechenkova Ekaterina V. \"resting-state fmri data acquired before and after a course of infra-low frequency neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004101.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 657 + ], + "type": "text", + "content": "[607] Damien Brevers, Chris Baeken, Stefanie De Smet, Beatrix Catoira, Sara De Witte, Qingha He, Pierre Maurage, Laimi Schultze, Guillaume Sescousse, Claudia Vila Verde, Claus Vögele, and Joel Billieux. \"stimulation of the dorsolateral prefrontal cortex modulates brain cue reactivity to reward(un)availability\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004102.v1.0.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[608] Siao-Shan Shen, Jen-Tang Cheng, I-Jeng Hsu, Der-Yow Chen, Ming-Hung Weng, and Chun-Chia Kung. \"collaborations and deceptions in strategic interactions revealed by hyperscanning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004103.v1.0.0." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[609] M.P. Weisend, F.M. Hanlon, R. Montano, S.P. Ahlfors, A.C. Leuthold, D. Pantazis, J.C. Mosher, A.P. Georgopoulos, M.S. Hamalainen, and C.J. Aine. \"mind data\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004107.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[610] Yannan Zhu and Shaozheng Qin. \"emotional learning retroactively promotes memory integration through rapid neural reactivation and reorganization\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004109.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[611] Kenji Kobayashi, Joseph W. Kable, Ming Hsu, and Adrianna C. Jenkins. \"neural representations of others' traits predict social decisions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004128.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[612] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype segregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004129.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[613] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype aggregation example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004130.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[614] Samuel Guay, Eric Earl, Hao-Ting Wang, Remi Gau, Dorota Jarecka, David Keator, Melissa Kline Struhl, Satra Ghosh, Louis De Beaumont, and Adam G. Thomas. \"bids phenotype external example dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ms004131.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "text", + "content": "[615] Sepehr Mortaheb, Laurens Van Calster, Federico Raimondo, Manousos A. Klados, Paradeisios Alexandros Boulakis, Kleio Georgoula, Steve Majerus, Dimitri Van De Ville, and Athena Demertzi. \"experience sampling in resting state\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004134.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[616] Bruno Direito, Marco Simões, Alexandre Sayal, and Miguel Castelo-Branco. \"targeting dynamic facial processing mechanisms in superior temporal sulcus using fmri neurofeedback\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004141.v1.0.5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[617] Bruno Direito, Manuel Ramos, Alexandre Sayal, Joao Pereira, and Miguel Castelo-Branco. \"exploring the neural correlates of feedback-related reward saliency and valence during real-time fmri-based neurofeedback\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004142.v1.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "type": "text", + "content": "[618] Thania Balducci, Jalil Rascado-Toledo, Alely Valencia, Maria-Jose van Tol, Andre Aleman, and Eduardo A. Garza-Villarreal. \"a behavioral, clinical and brain imaging dataset with focus on emotion regulation of females with fibromyalgia\". OpenNeuro, 2022. doi: 10.18112/openneuro. ds004144.v1.0.2." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 645 + ], + "type": "text", + "content": "[619] Martin Szinte, Anna Montagnini, and Guillaume Masson. \"rest_eye\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004158.v2.0.5." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[620] Lachlan T. Strike, Gabriella A.M. Blokland, Narelle K. Hansell, Nicholas G. Martin, Arthur W. Toga, Paul M. Thompson, Greig I. de Zubicaray, Katie L. McMahon, and Margaret J. Wright. \"queensland twin imaging (qtim)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004169.v1.0.7." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23869" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 30 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[621] Ádám Nárai, Petra Hermann, Tibor Auer, Péter Kemenczky, János Szalma, István Homolya, Eszter Somogyi, Pál Vakli, Béla Weiss, and Zoltán Vidnyanszky. \"movement-related artefacts (mr-art) dataset\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004173.v1.0.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "text", + "content": "[622] Stefanie Meliss, Cristina Pascua-Martin, Jeremy Skipper, and Kou Murayama. \"magic, memory, and curiosity (mmc) fmri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004182.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[623] Irem Yildirim, Khan Hekmatyar, and Keith A. Schneider. \"lgn layers data\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004187.v1.0.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 259 + ], + "type": "text", + "content": "[624] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004192.v1.0.7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 326 + ], + "type": "text", + "content": "[625] Iris Groen, Kenichi Yuasa, Amber Brands, Giovanni Piantoni, Stephanie Montenegro, Adeen Flinker, Sasha Devore, Orrin Devinsky, Werner Doyle, Patricia Dugan, Daniel Friedman, Nick Ramsey, Natalia Petridou, and Jonathan Winawer. \"visual ecog dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004194.v2.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[626] Foteini Liwicki, Vibha Gupta, Rajkumar Saini, Kanjar De, Nosheen Abid, Sumit Rakesh, Scott Wellington, Holly Wilson, Marcus Liwicki, and Johan Eriksson. \"bimodal dataset on inner speech\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004196.v2.0.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 491 + ], + "type": "text", + "content": "[627] Fabiane Schuch, Lennart Walger, Matthias Schmitz, Bastian David, Tobias Bauer, Antonia Harms, Laura Fischbach, Freya Schulte, Martin Schidlowski, Johannes Reiter, Felix Bitzer, Randi von Wrede, Attila Racz, Tobias Baumgartner, Valeri Borger, Matthias Schneider, Achim Flender, Albert Becker, Hartmut Vatter, Bernd Weber, Louisa Specht-Riemenschneider, Alexander Radbruch, Rainer Surges, and Theodor Rüber. \"an open presurgery mri dataset of people with epilepsy and focal cortical dysplasia type ii\". OpenNeuro, 2025. doi: 10.18112/openneuro.d004199.v1.0.6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "text", + "content": "[628] Martin N. Hebart, Oliver Contier, Lina Teichmann, Adam H. Rockter, Charles Zheng, Alexis Kidder, Anna Corriveau, Maryam Vaziri-Pashkam, and Chris I. Baker. \"things-meg\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004212.v2.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 569 + ], + "type": "text", + "content": "[629] Hermes D and Wandell BA. \"brainbeats\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004213.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[630] Mark A. Thornton, Elyssa Barrick, Nathan Liang, and Diana I. Tamir. \"interacting representations of mental states and traits\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004217.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[631] Kyoko Ohashi, Martin H. Teicher, Elizabeth Bolger, Laura C. Hernandez Garcia, Poopak Hafezi, Leslie P. Weiser, Cynthia E. McGreenery, and Alaptagin Khan. \"lite-book_Alertness_study\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004219.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[632] Mark A. Thornton and Diana I. Tamir.\"summed social cognition: additive neural representations of situations, mental states, and actions\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004226.v1.0.0." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "text", + "content": "[633] Anne T. Park, Hilary Richardson, Ursula A. Tooley, Cassidy L. McDermott, Austin L. Boroshok, Adrian Ke, Julia A. Leonard, M. Dylan Tisdall, Kirby Deater-Deckard, J. Christopher Edgar, and Allyson P. Mackey. \"early stressful experiences are associated with reduced neural responses to naturalistic emotional and social content in children\". OpenNeuro, 2022. doi: 10.18112/openneuro.d004228.v1.0.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 259 + ], + "type": "text", + "content": "[634] Min-Jeong Kim, Jae-Hoon Lee, Fernanda Juarez Anaya, Jinsoo Hong, William Miller, Sanjay Telu, Prachi Singh, Michelle Y Cortes, Katharine Henry, George L Tye, Michael P Frankland, Jose A Montero Santamaria, Jeih-San Liow, Sami S Zoghbi, Masahiro Fujita, Victor W Pike, and Robert B Innis. \"first-in-human evaluation of [11c]ps13, a novel pet radioligand, to quantify cyclooxygenase-1 in the brain\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004230.v3.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 262, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 317 + ], + "type": "text", + "content": "[635] Miguel B. Garcia, Gilles de Hollander, Marcus Grueschow, Rafael Polania, Michael Woodford, and Christian C. Ruff. \"individual risk attitudes arise from noise in neurocognitive magnitude representations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004259.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[636] Xuhao Shao, Ao Li, Chuangsheng Chen, Elizabeth F. Loftus, and Bi Zhu. \"cross-stage neural pattern similarity in the hippocampus predicts false memory derived from post-event inaccurate information\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004261.v2.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 375, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 419 + ], + "type": "text", + "content": "[637] Thirza Dado, Yaghmur Güçlütürk, Luca Ambrogioni, Gabriëlle Ras, Sander Bosch, Marcel van Gerven, and Umut Güçlü. \"hyper\". OpenNeuro, 2022. doi: 10.18112/openneuro.dso04280.v1.0.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[638] Krista Bond, Javier Rasero, Raghav Madan, Jyotika Bahuguna, Jonathan Rubin, and Timothy Verstynen. \"neuroloki\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004283.v1.0.3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 499 + ], + "type": "text", + "content": "[639] Chad S. Rogers, Michael S. Jones, Sarah McConkey, and Jonathan E. Peelle. \"listening task\". OpenNeuro, 2022. doi: 10.18112/openneuro.rs004285.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[640] Wilma A. Bainbridge and Chris I. Baker. \"multidimensional memory topography in the medial parietal cortex identified from neuroimaging of thousands of daily memory videos\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004286.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "text", + "content": "[641] Shaonan Wang, Yunhao Zhang, Xiaohan Zhang, Jingyuan Sun, Nan Lin, Jiajun Zhang, Chengqing Zong, and Shaonan Wang. \"an fmri dataset for concept representation with semantic feature annotations\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004301.v1.0.2." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 614, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 712 + ], + "type": "text", + "content": "[642] Joan Soler-Vidal, Paola Fuentes-Claramonte, Pilar Salgado-Pineda, Nuria Ramiro, María Ángeles García-León, María Llanos Torres, Antonio Arevalo, Amalia Guerrero-Pedraza, Josep Munuera, Salvador Sarro, Raymond Salvador, Wolfram Hinzen, Peter McKenna, and Edith Pomarol-Clotet. \"brain correlates of speech perception in schizophrenia patients with and without auditory hallucinations\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004302.v1.0.1." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23870" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 31 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[643] Jason A Avery, Madeline Carrington, and Alex Martin. \"a common neural code for representing imagined and inferred tastes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004312.v1.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 160 + ], + "type": "text", + "content": "[644] Momchil S. Tomov, Pedro A. Tsividis, Thomas Pouncy, Joshua B. Tenenbaum, and Samuel J. Gershman. \"atari-style video game learning fmri\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004323.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 227 + ], + "type": "text", + "content": "[645] Emily E. Perszyk, Jessica Trinh, Xue S. Davis, Zach Hutelin, Leonie Koban, Maria G. Veldhuizen, Jelena Djordjevic, Marilyn Jones-Gotman, Tor Wager, Hedy Kober, and Dana M. Small. \"odour-imagery ability is linked to food craving, intake, and adiposity change in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004327.v1.0.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[646] Johannes J.D. Singer, Radoslaw M. Cichy, and Martin N. Hebart. \"the spatiotemporal neural dynamics of object recognition for natural images and line drawings (fmri)\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004331.v1. 0.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 347 + ], + "type": "text", + "content": "[647] Melanie Ganz and Hannah Eichhorn. \"datasets with and without deliberate head movements for evaluating the performance of markerless prospective motion correction and selective reacquisition in a general clinical protocol for brain mri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004332.v1.3.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "text", + "content": "[648] Loris Naspi, Paul Hoffman, and Alexa Morcom. \"semantic_encoding\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004341.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "text", + "content": "[649] Oscar Ferrante, Ling Liu, Tamas Minarik, Urszula Gorska, Tara Ghafari, Huan Luo, and Ole Jensen. \"flux: A pipeline for meg analysis\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004346.v1.0.8." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "text", + "content": "[650] L.E. Frank and D. Zeithamova. \"evaluating methods for measuring background connectivity in slow event-related fmri designs\". OpenNeuro, 2022. doi: 10.18112/openneuro.ds004349.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[651] Asieh Zadbood, Samuel A. Nastase, Janice Chen, Kenneth A. Norman, and Uri Hasson. \"sixthsense\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004359.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 558 + ], + "type": "text", + "content": "[652] Korey P. Wylie, Benzi M. Kluger, Luis D. Medina, Samantha K. Holden, Eugene Kronberg, Jason R. Tregellas, and Isabelle Guard. \"parkinson's disease, functional connectivity, and cognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004392.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[653] Lauryn Burleigh and Steven Greening. \"fear in the mind's eye: the neural correlates of differential fear acquisition to imagined conditioned stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004393.v1.0.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "text", + "content": "[654] Doug Greve, Paul Wighton, Melanie Ganz, Martin Nørgaard, Paolo Zanotti-Fregonara, Cyril Pernet, Anthony Galassi, Adam Thomas, Nafiseh Ghazanfari, Jeih-San Liow, Granville Matheson, Gitte Knudsen, and Robert Innis. \"a pet molecular imaging brain atlas of cyclooxygenase-1 (kim 2021)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004401.v1.3.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[655] Corey Fernandez, Jiefeng Jiang, Shao-Fang Wang, Hannah L. Choi, and Anthony D. Wagner. \"rid\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004406.v1.0.0." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 126 + ], + "type": "text", + "content": "[656] Marc Himmelberg, Ekin Tuncok, Jesse Gomez, Kalanit Grill-Spector, Marisa Carrasco, and Jonathan Winawer. \"stanford child and adult checkerboard retinotopy dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004440.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "text", + "content": "[657] Puti Wen, Michael S. Landy, and Bas Rokers. \"identifying cortical areas that underlie the transformation from 2d retinal to 3d head-centric motion signals\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004443.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 194 + ], + "type": "text", + "content": "[658] Will be updated. \"xcp_walkthrough\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004450.v1.0.1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[659] Carolina Feher da Silva, Gaia Lombardi, Micah Edelson, and Todd A. Hare. \"magic_carpet\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004455.v1.1.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "text", + "content": "[660] Agata Wolna and Zofia Wodniecka. \"bilingual speech production with functional localizers\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004456.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 282 + ], + "type": "text", + "content": "[661] Renzo Huber. \"highresolution development\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004458.v1.0.2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 337 + ], + "type": "text", + "content": "[662] Chris Racey, Charlotte Rae, Ivor Simpson, Oscar Bowen-Hill, Mathilde Bauer, Julia Simner, Christina Kampoureli, and Jamie Ward. \"mapping the connectome of synaesthesia: An open access mri dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004466.v1.0.2." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[663] Halie Olson, Emily Chen, Kirsten Lydic, and Rebecca Saxe. \"mri data from 20 adults in response to videos of dialogue and monologue from sesame street\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004467.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[664] Raul Rodríguez-Cruces, Vicente Camacho-Téllez, Alfonso Fajardo, and Luis Concha. \"temporal lobe epilepsy - unam\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds004469.v1.1.4." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "text", + "content": "[665] Jonathan C. Lau, Yiming Xiao, Roy A. M. Haast, Greydon Gilmore, Kamil Uludag, Keith W. MacDougall, Ravi S. Menon, Andrew G. Parrent, Terry M. Peters, and Ali R. Khan. \"stereotactic neurosurgery dataset (snsx)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004470.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "text", + "content": "[666] Mohamad Abbass, Greydon Gilmore, Alaa Taha, Ryan Chevalier, Magdalena Jach, Terry M. Peters, Ali R. Khan, and Jonathan C. Lau. \"London heath sciences center parkinson's disease dataset (lhscpd)\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004471.v1.0.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[667] Alexander P. Rockhill, Alessandra Mantovani, Brittany Stedelin, Admed M. Raslan, and Nicole C. Swann. \"seeg forced two-choice task\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004473.v1.0.2." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[668] Noelle A. Jacobsen and Daniel P. Ferris. \"mobile eeg splitbelt walking study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004475.v1.0.3." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 679 + ], + "type": "text", + "content": "[669] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Nicole Tacugue, Zenia Valdiviezo, Laura D. Lewis, and Zachary Diamandis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004478.v1.0.2." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[670] Samuel Planton*, Fosca Al Roumi*, Liping Wang, and Stanislas Dehaene. \"abseqfmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.dso004482.v1.0.0." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23871" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 32 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[671] Stephanie D. Williams, Beverly Setzer, Nina E Fultz, Zachary Diamandis, Nicole Tacugue, Zenia Valdiviezo, and Laura D. Lewis. \"neural activity induced by sensory stimulation can drive large-scale cerebrospinal fluid flow during wakefulness in humans\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004484.v1.0.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[672] Ming Zhou, Zhengxin Gong, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for human action recognition\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004488.v1.1.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "text", + "content": "[673] Zhengxin Gong, Ming Zhou, Yuxuan Dai, Yushan Wen, Youyi Liu, and Zonglei Zhen. \"a large-scale fmri dataset for the visual processing of naturalistic scenes\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004496.v2.1.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[674] Ph.D. Evan Gordon and Ph.D. Timothy Laumann. \"perinatal stroke\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004498.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[675] Mikl Michal, Kovarova Anezka, Gajdos Martin, Marecek Radek, Novakova Marie, and Slavicek Tomas. \"multi-echo simultaneous multislice fmri dataset: Effect of acquisition parameters on fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps004499.v1.0.3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[676] Amanda Studnicki and Daniel P. Ferris. \"real world table tennis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004505.v1.0.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 431 + ], + "type": "text", + "content": "[677] Gabriel Castrillon, Samira Epp, Antonia Bose, Laura Fraticelli, André Hechler, Roman Belenya, Andreas Ranft, Igor Yakushev, Lukas Utz, Lalith Sundar, Josef P Rauschecker, Christine Preibisch, Katarzyna Kurcyus, and Valentin Riedl. \"the energetic costs of the human connectome\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004513.v1.0.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "type": "text", + "content": "[678] Clara Sava-Segal, Chandler Richards, Megan Leung, Daniel A. Handwerker, Javier Gonzalez-Castillo, Peter J. Molfese, Peter A Bandettini, and Emily S. Finn. \"avid: Audio-visual individual differences dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004516.v2.0.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 488, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 565 + ], + "type": "text", + "content": "[679] Nestor Zaragoza-Jimenez, Hauke Niehaus, Ina Thome, Christoph Vogelbacher, Gabriele Ende, Inge Kamp-Becker, Dominik Endres, and Andreas Jansen. \"modeling face recognition in the predictive coding framework: A combined computational modeling and functional imaging study - stage 2\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004529.v1.1.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[680] Elizabeth Valles-Capetillo, Magda Giordano, and Cristian Ibarra. \"ironia vev\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004533.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "text", + "content": "[681] Sebastian Dresbach, Renzo Huber, Omer Faruk Gulban, and Rainer Goebel. \"event-related vaso\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004539.v1.1.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[682] Emalie McMahon, Michael F. Bonner, and Leyla Isik. \"social interaction dyads\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004542.v1.0.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[683] Paul Black, C. Brock Kirwan, Thomas Meservy, William Tayler, and Jeff Williams. \"an fmri investigation of the neurocognitive processing of strategies and measures\". Open-Neuro, 2023. doi: 10.18112/openneuro.ds004544.v1.0.0." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[684] Michael Rojek-Giffin, Mael Lebreton, Andrea Farina, Jorg Gross, Jean Daunizeau, and Carsten de Dreu. \"learning rules of engagement for social exchange within and between groups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004553.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[685] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 1\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004556.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 261 + ], + "type": "text", + "content": "[686] Doris Groessinger, Florian Ph.S Fischmeister, Matthias Witte, Karl Koschutnig, Manuel Ninaus, Christa Neuper, Silvia Erika Kober, and Guilherme Wood. \"the role of superstition of cognitive control during neurofeedback training - part 2\". OpenNeuro, 2023. doi: 10.18112/openneuro. ds004557.v1.1.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 318 + ], + "type": "text", + "content": "[687] M. Rebsamen, D. Romascano, M. Capiglioni, R. Wiest, P. Radojewski, and C. Rummel. \"the phantom of bern: repeated scans of two volunteers with eight different combinations of mr sequence parameters\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004560.v1.0.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 352 + ], + "type": "text", + "content": "[688] Youngjo Song, Wooree Shin, Pyeongsoo Kim, and Jaeseung Jeong. \"mvpa for motor adaptation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004562.v1.0.4." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 353, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 374 + ], + "type": "text", + "content": "[689] Paul Broca and Carl Wernicke. \"demo\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004564.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[690] Ludmila Midrigan-Ciochina, Kayla P. Vodacek, Srinija Balabhadra, and David P. Corina. \"brain differences in monolingual and highly proficient multilingual speakers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004581.v2.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[691] Yingying Wang, Hongmi Lee, and Brice Kuhl. \"ds004590\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004590.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 488 + ], + "type": "text", + "content": "[692] Hayoung Song, Won Mok Shim, and Monica D. Rosenberg. \"song dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ms004592.v1.0.1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[693] Marc N. Coutanche, Jake Sauter, Essang Akpan, Rae Buckser, August Vincent, and M. Kathleen Caulfield. \"brain surface fingerprints\". OpenNeuro, 2023. doi: 10. 18112/openneuro.ds004594.v1.0.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 578 + ], + "type": "text", + "content": "[694] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, Bruno Nazarian, Julien Sein, and Jean-Luc Anton. \"gloups\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004597.v2.0.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 623 + ], + "type": "text", + "content": "[695] Antonin Rovai, Valentina Lolli, Nicola Trotta, Serge Goldman, and Xavier De Tiège. \"cerebrovascular reactivity normative dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004604.v2.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[696] Cameron C. McKay, Brooke Scheinberg, Ellie P. Xu, Katharina Kircanski, Daniel S. Pine, Melissa A. Brotman, Ellen Leibenluft, and Julia O. Linke. \"emotion and development branch phenotyping and dti (2012-2017)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004605.v1.0.1." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[697] David Romascano, Gian Franco Piredda, Samuele Caneschi, Tom Hilbert, Ricardo Corredor, Bénédicte Maréchal, Tobias Kober, Jean-Baptiste Ledoux, Eleonora" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23872" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 33 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "text", + "content": "Fornari, Patric Hagmann, and Solange Denervaud. \"developmental relaxometry 2023\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004611.v1.0.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 185 + ], + "type": "text", + "content": "[698] Dylan M. Nielson, Neda Sadeghi, Lisa S. Gorham, Lillian Eisner, Jeremy Taigman, Katherine Haynes, Karen Qi, Christopher C. Camp, Payton Fors, Diana Rodriguez, Jerry McGuire, Erin Garth, Chana Engel, Mollie Davis, Kenneth Towbin, and Argyris Stringaris. \"characterization and treatment of adolescent depression (cat-d)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004627.v1.1.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 239 + ], + "type": "text", + "content": "[699] Mengfan Zhang, Alexander Karner, Kathrin Kostorz, Sophia Shea, David Steyrl, Filip Melinscak, Ronald Sladky, Cindy Sumaly Lor, and Frank Scharnowski. \"spida-mrineuroimaging dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004630.v1.1.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 295, + 297 + ], + "type": "text", + "content": "[700] A. De Soares, T. Kim, F. Mugisho, E. Zhu, A. Lin, C. Zheng, and C. Baldassano. \"top-down attention impacts neural dynamics during narratives with overlapping event scripts\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004631.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 364 + ], + "type": "text", + "content": "[701] P. G. Bissett, I. W. Eisenberg, S. Shim, J. A. H. Rios, H. M. Jones, M. P. Hagen, A. Z. Enkavi, J. K. Li, J. A. Mumford, D. P. MacKinnon, L. A. Marsch, and R. A. Poldrack. \"cognitive tasks, anatomical mri, and functional mri data evaluating the construct of self-regulation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004636.v1.0.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 453 + ], + "type": "text", + "content": "[702] Edlow B. L., Olchanyi M. D., Freeman H. J., Li J., Maffei C., Snider S. B., Zollei L., Eugenio Iglesias J., Augustinack J., Bodien Y. G., Haynes R. L, Greve D. N., Diamond B. R., Stevens A., Giacino J. T., Destrieux C., van der Kouwe A., Brown E. N., Folkerth R. D., Fischl B., and Kinney H. C. \"sustaining wakefulness: Brainstem connectivity in human consciousness\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004640.v1.0.4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 499 + ], + "type": "text", + "content": "[703] Sydney M. Bailes, Daniel E.P. Gomez, Beverly Setzer, and Laura D. Lewis. \"resting-state fmri signals contain spectral signatures of local hemodynamic response timing\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004645.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "text", + "content": "[704] Heesun Park, Hoyoung Doh, Eunhwi Lee, Harhim Park, and Woo-Young Ahn. \"the neurocognitive role of working memory load when pavlovian motivational control affects instrumental learning\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004647.v1.0.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 556, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 633 + ], + "type": "text", + "content": "[705] Ana L. Miranda-Angulo, Juan D. Sánchez-lópez, Daniel A. Vargas-Tejada, Juan C. Calderón, Jaime Gallo-Villegas, Juan F. Alzate-Restrepo, Jazmin X. Suarez-Revelo, and Gabriel Castrillón. \"sympathovagal quotient and functional connectivity of control networks are related to gut ruminococcaceae abundance in healthy men\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004648.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "text", + "content": "[706] Martin Norgaard. \"the mn dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004650.v1.0.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[707] Kathleen C Schmidt, Inna Loutaev, Thomas V Burlin, Zenaide Quezado, Audrey Thurm, Carrie Sheeler, and Carolyn Beebe Smith. \"rates of cerebral protein synthesis in subjects with fragile x syndrome\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004654.v1.0.1." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[708] Travis D. Masterson, C. Brock Kirwan, Lance E. Davidson, and James D. LeCheminant. \"neural reactivity to visual food stimuli is reduced in some areas of the brain during evening hours compared to morning hours: an fmri study in women\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004656.v1.1.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "text", + "content": "[709] Neha A Reddy, Kristina M Zvolanek, and Molly G Bright. \"denoising task-correlated head motion from motor-task fmri data with multi-echo ica\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004662.v1.1.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[710] Fan Cheng, Tomoyasu Horikawa, Kei Majima, Misato Tanaka, Mohamed Abdelhack, Shuntaro C. Aoki, Jin Hirano, and Yukiyasu Kamitani. \"visual illusion reconstruction\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004670.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 241, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 318 + ], + "type": "text", + "content": "[711] Evan Layher, Tyler Santander, Puneeth Chakravarthula, Nicole Marinsek, Benjamin O. Turner, Miguel P. Eckstein, and Michael B. Miller. \"widespread frontoparietal fmri activity is greatly affected by changes in criterion placement, not discriminability, during recognition memory and visual detection tests\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004692.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[712] Johannes Singer, Agnessa Karapetian, Martin Hebart, and Radoslaw Cichy. \"identifying and characterizing scene representations relevant for categorization behavior\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004693.v1.0.3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[713] Bari Fuchs, Alaina Pearce, Barbara Rolls, Stephen Wilson, Emma Jane Rose, Charles Geier, Hugh Garavan, and Kathleen Keller. \"food and brain study\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004697.v1.0.2." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "text", + "content": "[714] Kelly Chang, Ione Fine, and Geoffrey M. Boynton. \"chn retinotopic mapping dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004698.v2.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[715] David A Feinberg, Alexander JS Beckett, and An T VU. \"next generation mri scanner designed for ultra-high resolution human brain imaging at 7-tesla\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004710.v1.1.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 488, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 510 + ], + "type": "text", + "content": "[716] Loreen Tisdall and Rui Mata. \"agerisk\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004711.v1.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "text", + "content": "[717] S. Warrington, A. Torchi, O. Mougin, A. Ntata, J. Campbell, M. Craig, F. Alfaro-Almagro, K.L. Miller, P.S. Morgan, M. Jenkinson, and S.N. Sotiropoulos. \"on-harmony: A resource for development and comparison of multi-modal brain 3t mri harmonisation approaches\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004712.v2.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 622 + ], + "type": "text", + "content": "[718] Elisa Zamboni, Alexis Makin, Marco Bertamini, and Antony Morland. \"symmetry & luminance bold fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004715.v1.0.3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[719] Elena Filimonova, Anton Pashkov, Norayr Borisov, Anton Kalinovsky, and Jamil Rzaev. \"utilizing amide proton transfer technique to characterise diffuse gliomas based on who 2021 classification of cns tumors\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004717.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[720] Mohammad Momenian, Zhengwu Ma, Shuyi Wu, Chengcheng Wang, and Jixing Li. \"le petit prince hong kong: Naturalistic fmri and eeg dataset from" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23873" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 34 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "older cantonese speakers\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004718.v1.1.1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 128 + ], + "type": "text", + "content": "[721] Unknown authors. \"descriptive name for this dataset\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004720.v1.0.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[722] Marilyn Horta, Rebecca Polk, and Natalie Ebner. \"single dose intranasal oxytocin administration: Data from healthy younger and older adults\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004725.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 249 + ], + "type": "text", + "content": "[723] Shrinivas Bishu, Kathleen C Schmidt, Thomas V Burlin, Michael A Charming, Lisa Horowitz, Tianjiang Huang, Zhong-Hua Liu, Mei Qin, B-K Vuong, Aaron J Unterman, Zengyan Xia, Alan Zametkin, Peter Herscovitch, Zenaide Quezado, and Carolyn B Smith. \"effects of propofol anesthesia on rates of cerebral protein synthesis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004730.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 337 + ], + "type": "text", + "content": "[724] Dante Picchioni, Kathleen C Schmidt, Kelly K McWhirter, Inna Loutaev, Adriana J Pavletic, Andrew M Speer, Alan J Zametkin, Ning Miao, Shrinivas Bishu, Kate M Turetsky, Anne S Morrow, Jeffrey L Nadel, Brittney C Evans, Diana M Vesselinovitch, Carrie A Sheeler, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis and memory formation during sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.d004731.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[725] Dante Picchioni, Kathleen C Schmidt, Inna Loutaev, Adriana J Pavletic, Carrie Sheeler, Shrinivas Bishu, Thomas J Balkin, and Carolyn B Smith. \"rates of cerebral protein synthesis in stages of sleep\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004733.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 470 + ], + "type": "text", + "content": "[726] Hamsanandini Radhakrishnan, Chenying Zhao, Valerie J. Sydnor, Erica B. Baller, Philip A. Cook, Damien Fair, Barry Giesbrecht, Bart Larsen, Kristin Murtha, David R. Roalf, Sage Rush-Goebel, Russell Shinohara, M. Dylan Tisdall, Jean Vettel, Scott Grafton, Matthew Cieslak, Theodore Satterthwaite, and Haochang Shou. \"cs-dsi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004737.v2.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[727] Anne Kavounoudias, Caroline Landelle, Jeremy Dana, Bruno Nazarian, Julien Sein, Jean-Luc Anton, Laurent Pruvost, Frederic Giraud, and Olivier Felician. \"audiotact_ya\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004743.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 558 + ], + "type": "text", + "content": "[728] Joe Bathelt, Joanne Taylor, and Kathy Rastle. \"language fmri\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004765.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[729] Rene Weber, Frederic R. Hopp, Allison Eden, and Kate Lee. \"vicarious punishment of moral violations in naturalistic drama narratives predicts cortical synchronization\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004775.v1.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[730] Katyucia de Macedo Rodrigues, Emma Ben-Avi, Danielle D. Sliva, Myong sun Choe, Marie Drottar, Ruopeng Wang, Bruce Fischl, Patricia E. Grant, and Lilla Zöllei. \"infant freesurfer test subject (cnybch)\". Open-Neuro, 2023. doi: 10.18112/openneuro.ms004776.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[731] A L Williams, S Vicary, and G Orgs. \"functional mri dataset for evaluating brain synchrony among dance spectators\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004783.v1.0.1." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[732] Eddy Cavalli, Valérie Chanoine, and Johannes C. Ziegler. \"morphosem\". OpenNeuro, 2023. doi: 10.18112/openneuro.rs004786.v1.0.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 553, + 142 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 142 + ], + "type": "text", + "content": "[733] Dylan Nielson, Andre Zugman, Marie Zelenina, and Danny Pine. \"nimh meter (multi-echo test-retest)\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004787.v1.1.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 144, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 144, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 144, + 553, + 187 + ], + "type": "text", + "content": "[734] Fu Yu Kwok, Eric D. Wilkey, Lien Peters, Ellyn Khiu, Rebecca Bull, Kerry Lee, and Daniel Ansari. \"kwok et al., 2023, human brain mapping\". OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004791.v1.0.0." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 190, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 190, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 190, + 553, + 266 + ], + "type": "text", + "content": "[735] Hampus Olsson, Jason Michael Millward, Ludger Starke, Thomas Gladytz, Tobias Klein, Jana Fehr, Wei-Cheng Lai, Christoph Lippert, Thoralf Niendorf, and Sonia Waiczies. \"simulated motion artifacts based on the (mr-art) dataset (https://openneuro.org/datasets/ds004173/versions/1.0.2)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004795.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 269, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 346 + ], + "type": "text", + "content": "[736] Umit Keles, Julien Dubois, Kevin J. M. Le, J. Michael Tyszka, David A. Kahn, Chrystal M. Reed, Jeffrey M. Chung, Adam N. Mamelak, Ralph Adolphs, and Ueli Rutishauser. \"fmri data for: Multimodal single-neuron, intracranial eeg, and fmri brain responses during movie watching in human patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004798.v1.0.5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 348, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 348, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 553, + 403 + ], + "type": "text", + "content": "[737] Audrey Mazancieux, Franck Mauconduit, Alexis Amadon, Jan Willem de Gee, Tobias Donner, and Florent Meyniel. \"brainstem fmri signaling of surprise across different types of deviant stimuli\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004808.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 405, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 438 + ], + "type": "text", + "content": "[738] Michelle Moerel and Essa Yacoub. \"high-res gradient echo epi and 3d grase data of auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004814.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 495 + ], + "type": "text", + "content": "[739] Agustin Lage-Castellanos, Federico De Martino, Geoff M. Ghose, Omer Faruk Gulban, and Michelle Moerel. \"selective attention sharpens population receptive fields in human auditory cortex\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004815.v1.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 563 + ], + "type": "text", + "content": "[740] Ana F. Palenciano, Carlos González-García, Jan de Houwer, Baptist Lieefoghe, and Marcel Brass. \"concurrent response and action effect representations across the somatomotor cortices during novel task preparation.\" OpenNeuro, 2023. doi: 10.18112/openneuro.d0s004829.v1.0.1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 565, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 620 + ], + "type": "text", + "content": "[741] Michelle Moerel, Federico De Martino, Kamil Ugurbil, Essa Yacoub, and Elia Formisano. \"processing of frequency and location in human subcortical auditory structures\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004835.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 622, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 677 + ], + "type": "text", + "content": "[742] Fran López-Caballero, Mark Curtis, Brian Coffman, and Dean Salisbury. \"magnetoencephalographic (meg) pitch and duration mismatch negativity (mmn) in first-episode psychosis\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004837.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[743] Kira Noad, David Watson, and Timothy Andrews. \"game of thrones - a naturalistic viewing dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004848.v1.0.1." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23874" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 35 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[744] Sean Paulsen and Michael A. Casey. \"shanxi enculturation\". OpenNeuro, 2023. doi: 10.18112/openneuro.ds004866.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 206 + ], + "type": "text", + "content": "[745] Nafiseh Ghazanfari, Jeih-San Liow, Min-Jeong Kim, Raven Cureton, Adrian Lee, Carson Knoer, Madeline Jenkins, Jinsoo Hong, Jose A. Montero Santamaria, H. Umesha Shetty, Anthony Galassi, Paul Wighton, Martin Norgaard, Douglas N. Greve, Sami S. Zoghbi, Victor W Pike, Robert B Innis, and Paolo Zanotti-Fregonara. [11c]ps13 demonstrates pharmacologically selective and substantial binding to cyclooxygenase-1 (cox-1) in the human brain\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004868.v1.0.4." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[746] Elenor Morgenroth, Stefano Moia, Laura Vilaclara, Raphael Fournier, Michal Muszynski, Maria Ploumitsakou, Marina Almato Bellavista, Patrik Vuilleumier, and Dimitri Van De Ville. \"emo-film\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps004892.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[747] Toru Maekawa, Takafumi Sasaoka, Toshio Inui, Alan S. R. Fermin, and Shigeto Yamawaki. \"heart rate and insula activity increase in response to music in individuals with high interoceptive sensitivity\". OpenNeuro, 2023. doi: 10.18112/openneuro.ps004894.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 342 + ], + "type": "text", + "content": "[748] Harrison Ritz and Amitai Shenhav. \"pact_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004909.v1.1.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 388 + ], + "type": "text", + "content": "[749] Ye Wu, Xiaoming Liu, Yunzhi Huang, Tao Zhou, and Fan Zhang. \"an open relaxation-diffusion mri dataset in neurosurgical studies\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs004910.v1.0.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 389, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 294, + 443 + ], + "type": "text", + "content": "[750] Alejandra Figueroa-Vargas, Gabriela Valdebenito-Oyarzo, María Paz Martínez-Molina, Francisco Zamorano, and Pablo Billeke. \"probability decision-making task with ambiguity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004917.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 500 + ], + "type": "text", + "content": "[751] David V. Smith, James Wyngaarden, Cooper J. Sharp, Daniel Sazhin, Ori Zaff, Dominic Fareri, and Johanna Jarcho. \"an fmri dataset of social and nonsocial reward processing in young adults\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004920.v1.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 502, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 567 + ], + "type": "text", + "content": "[752] Lonike K Faes, Agustin Lage-Castellanos, Giancarlo Valente, Zidan Yu, Martijn A. Cloos, Luca Vizioli, Steen Moeller, Essa Yacoub, and Federico De Martino. \"predictive processing in the auditory cortex measured at 7t\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004928.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 294, + 613 + ], + "type": "text", + "content": "[753] Shari Liu, Kirsten Lydic, Lingjie Mei, and Rebecca Saxe. \"fmri dataset: Violations of psychological and physical expectations in human adult brains\". OpenNeuro, 2024. doi: 10.18112/openneuro.dss004934.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 712 + ], + "type": "text", + "content": "[754] Elise M. Cardinale, Jessica Bezek, Julia O Linke, Olivia Siegal, Gabrielle Freitag, Anni Subar, Khosravi Parmis, Ajitha Mallidi, Olivia Peterson, Isaac Morales, Simone P Haller, Courtney Filippi, Kyunghun Lee, Melissa A. Brotman, Daniel S. Pine, Ellen Leibenluft, and Katharina Kirkanski. \"multivariate assessment of inhibitory control in youth: Links with psychopathology and brain function dataset\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005754.v1.1.0." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[755] Patrick Klepits, Karl Koschutnig, Thomas Zussner, and Andreas Fink. \"effects of a seven-week running intervention with moderate intensity on the volume of the hippocampus and depressive symptoms in young men from the general population.\" OpenNeuro, 2024. doi: 10.18112/openneuro. ds004937.v1.0.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[756] Jens Jensen, Andreanna Benitez, Erin Klintworth, Hunter Moss, and Kathryn Thorn. \"pydesigner example dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004945.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 249 + ], + "type": "text", + "content": "[757] Roxane Philips, Chris Baeken, Joel Billieux, James Harris, Pierre Maurage, Ismael Muela, Irem Tugce Oz, Athur Pabst, Guillaume Sescousse, Claus Vogele, and Damien Brevers. \"brain mechanisms discriminating enactive mental simulations of running and plogging\". OpenNeuro, 2024. doi: 10.18112/openneuro.d004946.v1.0.0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[758] Marleen Haupt, Monika Graumann, Santani Teng, Carina Kaltenbach, and Radoslaw M. Cichy. \"braille letters - fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004956.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 347 + ], + "type": "text", + "content": "[759] Angelique Volfart, Katie McMahon, and Greig de Zubicaray. \"a comparison of denoising approaches for spoken word production related artefacts in continuous multiband fmri data\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004957.v1.0.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 381 + ], + "type": "text", + "content": "[760] Douglas N. Greve and Bruce Fischl. \"the freesurfer maintenance dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms004958.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[761] Corinna Perchtold-Stefan, Christian Rominger, Karl Koschutnig, and Andreas Fink. \"truecrime\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds004965.v1.0.1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[762] Ekaterina Torubarova, Caroline Arvidsson, Jonathan Berrebi, Julia Uddén, and André Pereira. \"neuroengage\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds004996.v2.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[763] Eric C. Leuthardt, Joshua S. Shimony, Abraham Z. Snyder, Donna Dierker, and Ki Yun Park. \"retrospective task/rest fmri data from tumor patients\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005003.v2.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 547 + ], + "type": "text", + "content": "[764] Neha A Reddy, Rebecca G Clements, and Molly G Bright. \"simultaneous cortical, subcortical, and brainstem mapping of sensory activation\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005009.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "text", + "content": "[765] Michael I. Demidenko, Edward D. Huntley, and Daniel P. Keating. \"adolescent health risk behavior study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005012.v1.0.3." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[766] David Seminowicz, Shana Burrowes, Alexandra Kearson, Jing Zhang, Samuel Kimmel, Luma Samawi, Andrew Furman, and Michael Keaser. \"mbsr\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005016.v1.1.1." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 646 + ], + "type": "text", + "content": "[767] Unknown authors. \"pop\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005017.v1.0.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[768] Sara Ponticorvo and Fabrizio Esposito. \"hearing loss connectome\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005026.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[769] Michael I. Demidenko, Ryan Klaus, Mary Soules, and Mary M. Heitzeg. \"michigan longitudinal study\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005027.v1.0.3." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23875" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 36 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[770] Marie Mueckstein, Kai Görgen, Stephan Heinzel, Urs Granacher, A. Michael Rapp, and Christine Stelzel. \"modality-based multitasking and practice - fmri\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005038.v1.0.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 149 + ], + "type": "text", + "content": "[771] Michael S. Cohen, Qiongwen (Jovie) Cao, and Jean Decety. \"political moralization (pmm)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005040.v1.2.0." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[772] Panpan Chen, Chi Zhang, Bao Li, Li Tong, Linyuan Wang, Shuxiao Ma, Long Cao, Ziya Yu, and Bin Yan. \"an fmri dataset in response to large-scale short natural dynamic facial expression videos\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005047.v1.0.7." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[773] S.H.P. Collin, R. Kempner, S. Srivatsan, and K.A. Norman. \"neural codes track prior events in a narrative and predict subsequent memory for details\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms00505.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "type": "text", + "content": "[774] Guangyu Zhou, Gregory Lane, Thorsten Kahnt, and Christina Zelano. \"dataset2\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005056.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 304 + ], + "type": "text", + "content": "[775] Jing Wang. \"cr dbs\". OpenNeuro, 2024. doi: 10.18112/openneuro.dso005063.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 306, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 294, + 327 + ], + "type": "text", + "content": "[776] jessica. \"ds: ds005069:1.0.0\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005069.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 328, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 328, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 328, + 294, + 349 + ], + "type": "text", + "content": "[777] jessica. \"ds: ds005072:1.0.1\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005072.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 350, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 414 + ], + "type": "text", + "content": "[778] Brian P Keane, Bart Krekelberg, Ravi Mill, Steven M Silverstein, Judy L Thompson, Megan R Serody, Deanna M Barch, and Michael W Cole. \"the neural basis of visual shape completion in schizophrenia and bipolar disorder\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005073.v1.0.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 470 + ], + "type": "text", + "content": "[779] Caroline Landelle, Nawal Kinany, Benjamin De Leener, Nicholas D. Murphy, Ovidiu Lungu, Véronique Marchand-Pauvert, Dimitri Van De Ville, and Julien Doyon. \"simultaneous brain and spinal cord at rest: an fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005075.v1.0.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 472, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 502 + ], + "type": "text", + "content": "[780] Unknown authors. \"sequence pilot: multiecho and multiband fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005085.v1.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[781] Hannah Grotzinger, Laura Pritschet, Elle M. Murata, Tyler Santander, Pavel Shapturenka, Caitlin Taylor, and Emily G. Jacobs. \"28andhe\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05115.v1.2.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[782] Charles J. Lynch and Conor Liston. \"weill cornell medicine multi-echo (wcm-me) dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005118.v1.0.0." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 668 + ], + "type": "text", + "content": "[783] David V. Smith, Cooper J. Sharp, Abraham Dachs, James Wyngaarden, Daniel Sazhin, Jen Yang, Melanie Kos, Tia Tropea, Ishika Kohli, John A. Clithero, Ingrid Olson, Tania Giovannetti, Dominic Fareri, and Johanna M. Jarcho. \"social reward and nonsocial reward processing across the adult lifespan: An interim multi-echo fmri and diffusion dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005123.v1.1.3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[784] Hannah Tarder-Stoll, Christopher Baldassano*, and Mariam Aly*. \"the brain hierarchically represents the past and future during multistep anticipation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005125.v1.0.0." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 137 + ], + "type": "text", + "content": "[785] Takahiro Hirao, Mitsuhiro Miyamae, Daisuke Matsuyoshi, Ryuto Inoue, Yuhei Takado, Takayuki Obata, Makoto Higuchi, Naotsugu Tsuchiya, and Makiko Yamada. \"a neuroimaging dataset during sequential color qualia similarity judgments\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005126.v1.0.3." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 140, + 553, + 172 + ], + "type": "text", + "content": "[786] Dante Picchioni, Jeff H Duyn, and Jacco A de Zwart. \"amri 16-n-0031 sleep1\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005127.v1.0.4." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 553, + 206 + ], + "type": "text", + "content": "[787] Sangil Lee, Runxuan Niu, Lusha Zhu, Andrew Kayser, and Ming Hsu. \"deception signaling task\". OpenNeuro, 2024. doi: 10.18112/openneuro.ps005128.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 209, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 209, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 553, + 240 + ], + "type": "text", + "content": "[788] Adam Raikes. \"sequence test for multiband dwi and magnetization transfer sequence\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005134.v1.0.0." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 553, + 274 + ], + "type": "text", + "content": "[789] Binke Yuan and Junjie Yang. \"scnu-mandarin-cantonese-dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005139.v1.0.7." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "text", + "content": "[790] Ladan Shahshahani, Maedbh B. King, Caroline Netekoven, Rich B. Ivry, and Jorn Diedrichsen. \"wmfs\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005148.v1.1.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 321, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 321, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 315, + 321, + 553, + 398 + ], + "type": "text", + "content": "[791] Benjamin Lahner, Kshitij Dwivedi, Polina Iamshchinina, Monika Graumann, Alex Lascelles, Gemma Roig, Alessandro Thomas Gifford, Bowen Pan, SouYoung Jin, N.Apurva Ratan Murty, Kendrick Kay, Radoslaw Cichy*, and Aude Oliva*. \"modeling short visual events through the bold moments video fmri dataset and metadata\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005165.v1.0.4." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 443 + ], + "type": "text", + "content": "[792] Andrei Barborica, Felicia Mihai, Laurentiu Tofan, Irina Oane, and Ioana Mindruta. \"dataset of intracranial EEG during cortical stimulation evoking visual effects\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005169.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 445, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 445, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 553, + 465 + ], + "type": "text", + "content": "[793] Tomoyasu Horikawa. \"mind captioning\". OpenNeuro, 2024. doi: 10.18112/openneuro.rs005191.v1.0.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 553, + 510 + ], + "type": "text", + "content": "[794] Nicholas G. Cicero, Michaela Klimova, Laura D. Lewis, and Sam Ling. \"differential cortical and subcortical visual processing with eyes shut\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005194.v1.1.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 513, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 513, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 513, + 553, + 544 + ], + "type": "text", + "content": "[795] Jiwoong Park, Hayoung Song, and Won Mok Shim. \"narrativepuzzle\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005215.v1.0.0." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 546, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 546, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 315, + 546, + 553, + 633 + ], + "type": "text", + "content": "[796] Lucas Mahler, Julius Steiglechner, Benjamin Bender, Tobias Lindig, Dana Ramadan, Jonas Bause, Florian Birk, Raphael Heule, Edyta Charyasz, Michael Erb, Vinod Jangir Kumar, Gisela E Hagberg, Pascal Martin, Gabriele Lohmann, and Klaus Scheffler. \"ultracortex: Submillimeter ultra-high field 9.4t brain mr image collection and manual cortical segmentations\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005216.v1.1.0." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 553, + 677 + ], + "type": "text", + "content": "[797] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"reinforcement-learning generalization\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005230.v1.0.0." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "text", + "content": "[798] Kirill A. Fadeev, Ilacai V. Romero Reyes, Dzerassa D. Goiaeva, Tatyana S. Obukhova, Tatyana M. Ovsiannikova, Andrey O. Prokofyev, Anna M. Rytikova, Artyom Y." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23876" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 37 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 127 + ], + "type": "text", + "content": "Novikov, Vladimir V. Kozunov, Tatyana A. Stroganova, and Elena V. Orekhova. \"perception of vowel sounds in children with autism spectrum disorders and typically developing children (meg/erf study)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005234.v2.1.7." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "text", + "content": "[799] J. Veillette and H. Nusbaum. \"motor-fmri\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005239.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[800] Taylor Salo, M. Dylan Tisdall, Lia Brodrick, Adam Czernuszenko, David R. Roalf, Sage Rush-Goebel, Nicholas Wellman, Matthew Cieslak, and Theodore D. Satterthwaite. \"single-echo/multi-echo comparison pilot\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005250.v1.1.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "text", + "content": "[801] Snezana Todorovic, Elin Runnqvist, Valerie Chanoine, and Jean-Michel Badier. \"gloups_meg\". OpenNeuro, 2025. doi: 10.18112/openneuro.ps005261.v2.0.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 304 + ], + "type": "text", + "content": "[802] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Incentivized reaching\". OpenNeuro, 2024. doi: 10.18112/openneuro.d05263.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 380 + ], + "type": "text", + "content": "[803] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Ultra-high field functional connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005264.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 448 + ], + "type": "text", + "content": "[804] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Approach-avoid\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005265.v1.0.0." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 514 + ], + "type": "text", + "content": "[805] Neil M. Dundon, Elizabeth Rizor, Joanne Stasiak, Jingyi Wang, Kiana Sabugo, Christina Villaneuva, Parker Barandon, Andreea C. Bostan, Regina Lapate, and Scott T. Grafton. \"socal kinesia and incentivization for parkinson's disease (skip): Active escape\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005266.v1.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 579 + ], + "type": "text", + "content": "[806] Anja Ischebeck, Hannah Kreilinger, Joe Miller, Margit Hoefler, Iain D. Gilchrist, and Christof Koerner. \"fixating targets in visual search: The role of dorsal and ventral attention networks in the processing of relevance and rarity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005267.v1.0.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 623 + ], + "type": "text", + "content": "[807] Hsi T. Wei, Farhan B. Faisal, Tamara Beck, Claire Shao, and Jed A. Meltzer. \"picture-word interference dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005279.v1.0.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 679 + ], + "type": "text", + "content": "[808] Vassiki Chauhan, Krystal McCook, and Alex White. \"dataset accompanying \"reading reshapes stimulus selectivity in the visual word form area\", chauhan, mccook and white (2024)\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005295.v1.0.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[809] Laura Pritschet, Caitlin M. Taylor, Daniela Cossio, Joshua Faskowitz, Tyler Santander, Daniel A. Handwerker, Hannah Grotzinger, Evan Layher, Elizabeth R. Chrastil, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Emily G. Jacobs. \"maternal brain project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005299.v1.0.0." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[810] unknown author. \"neural bases of psychological reactance: An fmri study on dogmatic and suggestive health communication\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005304.v1.0.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[811] Robert O. Duncan and Evan A. Owens. \"bsc_bids_071724\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005355.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 197 + ], + "type": "text", + "content": "[812] Unknown authors. \"neurogame project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005357.v1.0.0." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 252 + ], + "type": "text", + "content": "[813] Elizabeth J. Rizor, Viktoriya Babenko, Neil M. Dundon, Renee Beverly-Aylwin, Alexandra Stump, Margaret Hayes, Luna Herschenfeld-Catalan, Emily G. Jacobs, and Scott T. Grafton. \"hormone health study (hhs)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005360.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "text", + "content": "[814] Alejandra Figueroa-Vargas, Leonie Kausel, María Paz Martínez-Molina, Francisco Zamorano, Ximena Stecher, Patricia Soto-Icaza, and Pablo Billeke. \"neurocovid mri dwi and fmri with reversal learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005364.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 387 + ], + "type": "text", + "content": "[815] Amy Kartar, Toru Horinouchi, Balázs Örzsik, Brittany Anderson, Lottie Hall, Duncan Bailey, Sarah Samuel, Nati Beltran, Samira Bouyagoub, Chris Racey, Yoko Nagai, Iris Asllani, Hugo Critchley, and Alessandro Colasanti. \"neurobiological substrates of altered state of consciousness induced by high ventilation breathwork\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005365.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 443 + ], + "type": "text", + "content": "[816] Magnus S. Bom, Annette M.A. Brak, Mathijs A.H.L.L. Raemaekers, Nick F. Ramsey, Mariska J. Vansteensel, and Mariana P. Branco. \"large-scale fmri dataset for the design of motor-based brain-computer interfaces\". OpenNeuro, 2025. doi: 10.18112/openneuro.ms005366.v2.0.0." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[817] Jessica Archibald, Amy Bouchard, Ralph Noeske, Dikoma Shungu, and Mark Mikkelsen. \"test-retest reliability of multi-metabolite edited mrs at 3t using press and slaser\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005371.v1.1.3." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[818] Hyunwoo Gu, Joonwon Lee, Sungje Kim, Jaeseob Lim, Hyang-Jung Lee, Heeseung Lee, Minjin Choe, Dong-Gyu Yoo, Jun Hwan (Joshua) Ryu, Sukbin Lim, and Sang-Hun Lee. \"discrimination-estimation task\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005381.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "text", + "content": "[819] Chema G. Penalver, David Lopez-Garcia, and Marta Becerra Losada. \"atexp_fmi\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005386.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "text", + "content": "[820] Jeryn Chang, Pamela A McCombe, Robert D Henderson, Diana Zanfirache, Christine C Guo, JingLei Lv, Kelly Garner, Saskia Bollmann, Shyuan T Ngo, Frederik J Steyn, and Thomas B Shaw. \"an fmri dataset for appetite neural correlates in people living with motor neuron disease\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005401.v1.1.1." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 678 + ], + "type": "text", + "content": "[821] Grant and None. \"grant dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005412.v1.0.0." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[822] Alexander P. Rockhill and Ahmed M. Raslan. \"numbers\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005415.v1.0.0." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23877" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 38 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[823] Antonin Rovai, Valentina Lolli, Xavier De Tiège, Nicola Trotta, and Serge Goldman. \"cerebrovascular reactivity: Co2 inhalation and resting state comparisons\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005418.v3.2.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 217 + ], + "type": "text", + "content": "[824] Juan Eugenio Iglesias, Adrià Casamitjana, Brain L Edlow, Martina Bocchetta, Dorit Kliemann, Matteo Mancini, Loïc Peter, Eleanor Robinson, Roberto Annunziata, Juri Althonayan, Shauna Crampsie, Emily Blackburn, Benjamin Billot, Alessia Atzeni, Oula Puonti, Yael Balbastre, Peter Schmidt, James Hughes, Jean C Augustinack, Lilla Zöllei, David L Thomas, Catherine Strand, Janice L Holton, and Zane Jaunmuktane. \"nextbrain data extension\". OpenNeuro, 2024. doi: 10.18112/openneuro.d005422.v1.0.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 294, + 261 + ], + "type": "text", + "content": "[825] Juan Pablo Franco, Peter Bossaerts, and Carsten Murawski. \"the neural dynamics associated with computational complexity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005427.v1.1.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 318 + ], + "type": "text", + "content": "[826] M. R. Ehlers, J. H. Kryklywy, Beukers A. O., Moore S. R., B. J. Forys, Anderson A.K., and R. M. Todd. \"valenced tactile information is evoked by neutral visual cues following emotional learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005449.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "text", + "content": "[827] Emilie C. Sleight, Oscar Esteban, and Frédéric Grouiller. \"fmriprep boot camp - 7t training dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005454.v1.0.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[828] Tingting Guo, Xuedi Liu, Mo Chen, Yongben Fu, and Taomei Guo. \"an fmri dataset for investigating language control and cognitive control in bilinguals\". OpenNeuro, 2024. doi: 10.18112/openneuro.dS005455.v1.1.5." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[829] Wangjing Yu, Asieh Zadbood, Avi J. H. Chanales, and Lila Davachi. \"priority\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005464.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 476 + ], + "type": "text", + "content": "[830] Antonin Rovai and Emilie Caspar. \"sense of agency in free and coerced moral decision-making among civilians and military personnel\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005469.v2.0.0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 554 + ], + "type": "text", + "content": "[831] Patrycja Scislewska, Michal Rafal Zareba, Julia Lengier, Aaron Schirmer, Piotr Bebas, and Iwona Szatkowska. \"monetary incentive delay task - structural and functional images of 37 men; study of associations between circadian characteristics (eveningness, distinctness) and affective processing\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005479.v1.0.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 555, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 578 + ], + "type": "text", + "content": "[832] Unknown authors. \"neurowater project\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005492.v1.0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 622 + ], + "type": "text", + "content": "[833] Niharika Gajawelli, Cameron Glick, Yinming Sun, Faizan Badami, Amit Etkin, and Manish Saggar. \"single-pulse tms fmri\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005498.v2.0.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[834] Zoya Surani, Ted Turesky, Eileen Sullivan, Talat Shama, Rashidul Haque, Nazrul Islam, Shahriah Hafiz Kakon, Xi Yu, William Petri, Charles Nelson, and Nadine Gaab. \"psychosocial adversity and inhibitory control: an fmri study of children growing up in extreme poverty\". Open-Neuro, 2024. doi: 10.18112/openneuro.ds005504.v1.0.0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[835] Ioannis Ntoumanis, Julia Sheronova, Alina Davydova, Maria Dolgaleva, Iiro P Jäaskeläinen, Vladimir Kosonogov," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 553, + 116 + ], + "type": "text", + "content": "Anna N Shestakova, and Vasily Klucharev. \"deciphering the neural responses to a naturalistic persuasive message\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005518.v1.0.1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 205 + ], + "type": "text", + "content": "[836] Xingfeng Shao, Qinyang Shou, Kimberly Felix, Brandon Ojogho, Xuejuan Jiang, Brian T Gold, Megan M Herting, Eric L Goldwaser, Peter Kochunov, L Elliot Hong, Ioannis Pappas, Meredith Braskie, Hosung Kim, Steven Cen, Kay Jann, and Danny JJ Wang. \"dp-pcasl data (cbf, att, bbb kw) from 186 cognitively normal participants (8-92 years)\". OpenNeuro, 2024. doi: 10.18112/openneuro. ds005529.v1.0.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[837] Viviana Greco, Tamas A. Foldes, Neil A. Harrison, Kevin Murphy, Marta Wawrzuta, Mahmoud E. A. Abdellahi, and Penelope A. Lewis. \"depotentiation of emotional reactivity using tmr during rem sleep\". OpenNeuro, 2025. doi: 10.18112/openneuro.ds005530.v1.0.9." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 260, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 303 + ], + "type": "text", + "content": "[838] Hiroto Q. Yamaguchi, Naoko Koide-Majima, Rieko Kubo, Tomoya Nakai, and Shinji Nishimoto. \"narrative movie fmri dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005531.v1.0.0." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 380 + ], + "type": "text", + "content": "[839] Nicholas G. Cicero, Nina E. Fultz, Hongbae Jeong, Stephanie D. Williams, Daniel Gomez, Beverly Setzer, Tracy Warbrick, Manfred Jaschke, Ravij Gupta, Michael Lev, Giorgio Bonmassar, and Laura D. Lewis. \"high-quality multimodal mri with simultaneous eeg using conductive ink and polymer-thick film nets\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005533.v1.0.0." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 414 + ], + "type": "text", + "content": "[840] Zsuzsanna Nemecz and Attila Keresztes. \"item and spatial pattern separation\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005559.v1.0.1." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[841] María Paz Martínez-Molina, Alejandra Figueroa-Vargas, Francisco Zamorano, and Pablo Billeke. \"expectation of conflict stimuli\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005571.v1.0.1." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 536 + ], + "type": "text", + "content": "[842] Jarod L Roland, Carl D Hacker, Abraham Z Snyder, Joshua S Shimony, JOhn M Zempel, David D Limbrick, Matthew D Smyth, and Eric C Leuthardt. \"a comparison of resting state functional magnetic resonance imaging to invasive electrocortical stimulation for sensorimotor mapping in pediatric patients\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005573.v1.0.0." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 580 + ], + "type": "text", + "content": "[843] Toshiki Okumura, Ikuhiro Kida, Atsushi Yokoi, Shinji Nishimoto, Kazushige Touhara, and Masako Okamoto. \"dataset for 7t odor fmri experiment\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005576.v1.0.0." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 634 + ], + "type": "text", + "content": "[844] Alexa Tompary and Lila Davachi. \"integration of overlapping sequences emerges with consolidation through mpfc neural ensembles and hippocampal-cortical connectivity\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005581.v1.0.0." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[845] Johannes Schultz, Maria Gädeke, and Tom Willems. \"sodec - responsibility fmri experiment dataset\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005588.v1.0.0." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[846] Yaqiong Chai, Hedong Zhang, Carlos Robles, Andrew Shinho Kim, Kyung-Wook Kang, Jiyoung Kim, JungWon Shin, Abigail Trang, Jongmok Ha, Emily Lee, Yeonsil Moon, and Hosung Kim. \"precise perivascular space" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23878" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 39 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 106 + ], + "type": "text", + "content": "segmentation on t2-weighted magnetic resonance imaging from human connectome project-aging\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005595.v1.0.0." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[847] Bolong Wang, XuanXuan Zhang, and Xiangzhen Kong. \"a naturalistic fmri dataset in response to public speaking\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005596.v1.1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 239 + ], + "type": "text", + "content": "[848] EC Washington University School of Medicine Leuthardt, JS Washington University School of Medicine Shimony, AZ Washington University School of Medicine Snyder, D Washington University School of Medicine Dierker, and KY Washington University School of Medicine Park. \"retrospective task/rest fmri data from tumor patients - for sora\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005597.v1.0.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 295, + 284 + ], + "type": "text", + "content": "[849] Jason Gallivan, Corson Areshenkoff, Anouk de Brouwer, and Dan Gale. \"visuomotor rotation learning and reward-based motor learning\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005598.v1.0.0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "text", + "content": "[850] Xitong Chen, Stephanie Leach, Juniper Hollis, Dillan Cellier, and Kai Hwang. \"thalamocortical contributions to hierarchical cognitive control\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005600.v1.1.0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 407 + ], + "type": "text", + "content": "[851] Peter N. Taylor, Yujiang Wang, Callum Simpson, Vytene Janiukstyte, Jonathan Horsley, Karoline Leiberg, Beth Little, Harry Clifford, Sophie Adler, Sjoerd B. Vos, Gavin P Winston, Andrew W McEvoy, Anna Miserocchi, Jane de Tisi, and John S Duncan. \"the imaging database for epilepsy and surgery (ideas)\". OpenNeuro, 2024. doi: 10.18112/openneuro.d0s005602.v1.0.0." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 294, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 440 + ], + "type": "text", + "content": "[852] Tim Mitchell and Carl Hacker. \"mitchell_hacker_2013\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005603.v1.0.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 441, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 463 + ], + "type": "text", + "content": "[853] Unknown authors. \"v4 crowding\". OpenNeuro, 2024. doi: 10.18112/openneuro.ms005604.v1.0.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 464, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 552 + ], + "type": "text", + "content": "[854] Xuefeng Yan, Fabrice G. Simeon, Jeih-San Liow, Cheryl L. Morse, Susovan Jana, Jose A. Montero Santamaria, Madeline Jenkins, Isabella Piacentino, Anthony Galassi, Sami S. Zoghbi, Victor W. Pike, Robert B. Innis, and Zanotti-Fregonara. \"[18f]sf51, a novel 18f-labeled pet radioligand for translocator protein 18kda (tspo) in brain, works well in monkeys but fails in humans\". OpenNeuro, 2024. doi: 10.18112/openneuro.ds005619.v1.1.0." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23879" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 40 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_content_list.json b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..84e20fc329a8cdeb2aaf1a9d78d6ab22f53da417 --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_content_list.json @@ -0,0 +1,1739 @@ +[ + { + "type": "text", + "text": "Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pegah Khayatan $^{\\star 1}$ Mustafa Shukor $^{\\star 1}$ Jayneel Parekh $^{\\star 1}$ Arnaud Dapogny $^{1}$ Matthieu Cord $^{1,2}$", + "bbox": [ + 99, + 179, + 895, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1ISIR, Sorbonne Université, Paris, France 2Valeo.ai, Paris, France", + "bbox": [ + 228, + 215, + 769, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 275, + 326, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.1", + "bbox": [ + 88, + 308, + 485, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 672, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid progress in Large Language Models (LLMs) [7, 10, 28, 44, 65], Multimodal LLMs (MLLMs) [3, 12, 34, 39, 61] have recently demonstrated remarkable capabilities in addressing complex multimodal tasks such as image captioning and visual question-answering.", + "bbox": [ + 89, + 696, + 483, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLLMs are typically composed of a visual encoder, an LLM, and a connector. Following initial unimodal pretraining—and, in many cases, multimodal pretraining on large datasets—these models can be further specialized by training on multimodal datasets [1, 33]. Given the high compu", + "bbox": [ + 89, + 773, + 485, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tational cost of training these models, recent research has proposed more efficient approaches, such as creating diverse, high-quality instruction-tuning datasets [39] or keeping the LLM frozen and fine-tuning small amounts of parameters, like the connector [41, 57, 58, 67]. These approaches take advantage of the ability of frozen LLMs to generalize to multimodal data [56]. Despite the different efficient tuning methods, training these models still incurs significant costs.", + "bbox": [ + 511, + 277, + 906, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While substantial progress has been made in developing high-performing MLLMs, relatively few studies aim to understand them [4, 47, 55, 56, 59, 61, 73]. Existing work typically conducts post-hoc analyses of MLLMs in isolation, overlooking the internal changes due to fine-tuning. Research by [56] addresses this gap to some extent by examining the internal multimodal alignment as it evolves during training.", + "bbox": [ + 511, + 398, + 908, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we apply concept-level analysis to provide a readable understanding of MLLM behavior and, in particular, semantic dynamics that may occur due to fine-tuning, or due to covariate shift when considering different datasets.", + "bbox": [ + 511, + 518, + 906, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the first case, we find that fine-tuning on a specific task potentially reshapes learned latent concepts, with some adjusting subtly to align with the task, and others emerging or disappearing altogether (see Fig. 1). Notably, we find that most fine-tuned concepts can be reconstructed from the original model by translating its original concepts in the direction of specific concept shift vectors, reducing the need for additional training and its associated costs. Furthermore, we explore the implications of the proposed analysis for MLLMs steering, demonstrating how model outputs can be modified inexpensively without additional training. Our key findings are summarized as follows:", + "bbox": [ + 511, + 580, + 908, + 761 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We apply latent concept-level analysis to provide readable understanding on MLLMs' behavior ; in particular, we show that fine-tuning can introduce significant alteration in the original concepts.", + "- We show that we can control the MLLM's behavior w.r.t. certain concepts by simply manipulating shift vectors.", + "- Lastly, our findings also have direct applications for steering MLLM outputs, which find use for model debiasing as well as safety control." + ], + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*First authors", + "bbox": [ + 107, + 863, + 184, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page and code: https://pegah-kh.github.io/projects/lmm-finetuning-analysis-and-steering/", + "bbox": [ + 91, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2206", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In a nutshell, we propose a novel, ready-to-use framework (including code) for MLLM behavior interpretability and control, debiasing and steering, which, we believe, will pave the way for future research.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 167, + 232, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept-based explainability. Concept-based explainability methods have emerged as an alternative to traditional feature attribution based methods, that are capable of extracting key semantic features from the model internal representations. Most post-hoc concept-based approaches are based on the idea of concept activation vectors (CAV) [30], which represent concepts as vectors in the activation space. Instead or relying on human annotations, recent works have proposed methods to automatically discover concepts via clustering [21, 71] or matrix decomposition [18], which can be viewed as instances of a dictionary learning problem [17]. Initially focusing on understanding vision models, dictionary learning for concept extraction has been extended to LLMs e.g. using sparse autoencoders [27, 52]. However, none of the prior approaches have been applied to understand MLLMs, with the exception of recently proposed CoX-LMM [47].", + "bbox": [ + 89, + 188, + 485, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLMs and Explainability. Multimodal LLMs [3, 34, 39, 67] have recently garnered significant interest. They typically adopt a late fusion architecture, and consist of an image encoder [19, 51, 72], a connector, and an LLM [28, 63, 65]. This family of models has inspired extensive research to better understand them and explain their behavior. For example, studies like [26, 45, 55] seek to identify multimodal neurons within LLMs or analyze modality-specific sub networks [56]. Some methods leverage the fact that these models are text-generative to simply generate textual explanations for model outputs [8, 20, 59, 70]. MLLMs benefit from in-context learning capabilities, which have been examined for limitations, including biases [4] and links to hallucinations [59], as well as the factors that may enhance their in-context learning performance [9, 49]. Related to our approach, CoX-LMMs [47] employs dictionary learning to extract multimodal semantic concepts from model representations. However, these studies typically assess models only in their final trained states, overlooking the dynamic changes that occur during training. Only limited works, such as [56], have investigated explaining changes due to fine-tuning, focusing specifically on implicit alignment between image and text modalities. In this work, we investigate how multimodal concepts within the model evolve throughout fine-tuning and explore the implications of these shifts on model steering.", + "bbox": [ + 89, + 431, + 483, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Steering models with feature editing. In contrast to editing model weights, representation or feature editing methods [62, 66, 69] aim to modify model outputs without altering the model's weights. A prominent approach within this family involves identifying steering vectors, or directions in the feature space (often within the residual stream), that are", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "linked to contrasting concepts. These methods have been applied to language models for various purposes, such as enhancing factuality or reducing hallucinations [46], inducing sentiment shifts or detoxification [64, 66], improving refusals to harmful requests [2], promoting truthfulness by modifying the output of attention heads [35], and erasing specific concepts or biases [6, 53]. However, their application to MLLMs is yet to be explored. Another set of approaches related to steering methods are based on In-context learning (ICL) [15, 16, 29], where prompts are carefully designed to induce desired behavior. Yet, ICL requires predefined demonstrations and lacks interpretability at a concept level. In contrast, our method offers lightweight steering, extending such capabilities to MLLMs without requiring any training, for instance as in ReFT [69].", + "bbox": [ + 511, + 90, + 908, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology Overview", + "text_level": 1, + "bbox": [ + 511, + 325, + 733, + 343 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework is summarized in Fig. 1. We apply concept-level analysis of MLLM latent space (Section 3.1) to define and compare concepts between different setups. This allows us to monitor conceptual changes occurring during fin-tuning and manipulate MLLM behavior at a conceptual level (Section 3.2). Lastly, our framework finds applications for MLLM steering for e.g. debiasing and safety control (3.3) with negligible computational burden.", + "bbox": [ + 511, + 352, + 908, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLM setup. A generic MLLM consists of a visual encoder $f_{V}$ , a trainable connector $C$ , and a language model $f_{\\mathrm{LM}}$ . We assume that the model is pretrained on a multimodal (e.g. captioning) dataset $S = \\{(x_i, y_i)\\}_i$ , where $x_i \\in \\mathcal{X}$ represents images and $y_i \\subset \\mathcal{Y}$ are the associated captions specified as sequence of tokens from token vocabulary space $\\mathcal{Y}$ . The model is trained to generate the next text tokens, conditioned on text and images. The input to $f_{\\mathrm{LM}}$ is a sequence of tokens that includes the concatenation of: (1) $N_V$ visual tokens extracted from the image $x$ via the visual encoder and connector $(C(f_V(x)))$ , and (2) linearly embedded textual tokens corresponding to the text instruction and previously predicted tokens. This can be expressed as:", + "bbox": [ + 511, + 474, + 908, + 672 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} ^ {p} = f _ {L M} (h ^ {1}, \\dots , h ^ {N _ {V}}, \\dots , h ^ {p}),\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 683, + 823, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $h^1, \\ldots, h^{N_V} = C(f_V(x))$ , and $h^p = \\operatorname{Emb}(\\hat{y}^{p-1})$ with Emb representing the token embedding layer. During generation, the output token $\\hat{y}^p$ is derived by normalizing the last layer $(L)$ tokens $h_{(L)}^p$ , then applying the unembedding layer $W_U$ and a softmax operation. The model keeps predicting the next token until the end of the sentence token to obtain the generated response $\\hat{y} = \\{\\hat{y}^p\\}_{p > N_V + N_I}$ , where $N_I$ is corresponds to the text instruction.", + "bbox": [ + 511, + 713, + 908, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Retrieval and comparison of latent concepts", + "text_level": 1, + "bbox": [ + 511, + 847, + 885, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To understand the internal representations of any given MLLM $f$ , we leverage the approach introduced in [47].", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/054278d75a0790e4190ebed6be90a0d14c9a03ae03ee7c970853d97f5e1cefeb.jpg", + "image_caption": [ + "Figure 1. Framework overview. We apply concept-level analysis for MLLM behavior monitor and control, for (left) understanding and manipulating (through shift vectors) concept changes due to fine-tuning, as well as (right) MLLM steering for debiasing or safety control." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 480, + 329 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e72ac1c9d73f1d12b0d1dffe68512f63b1401a077fe8832fe65d6ceb229070c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 90, + 861, + 229 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4941f3fec2ee19335fb9bef24aaffe6f277b15a2a2b1d2676fc371e95cf25dd8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 231, + 859, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given a set of $M$ images $\\{x_{1},\\dots,x_{M}\\}$ we extract a set of residual stream representations from some layer $l$ of the MLLM $f$ . These representations $\\mathbf{z}_m = f_l(x_m)\\in \\mathbb{R}^D$ (one per image) are collected in a feature matrix $\\pmb{Z}\\in \\mathbb{R}^{D\\times M}$ . Typically, the set of images and extracted representations correspond to a particular token of interest TOI (e.g., 'Dog', 'Cat', 'Person', etc.) in the predicted caption. However, the extraction can also be performed for a larger set of target tokens. This feature matrix $\\pmb{Z}$ is then decomposed as $\\pmb{Z}\\approx \\pmb{U}\\pmb{V}$ to recover the concepts in the latent embedding space. Here, $\\pmb{U}\\in \\mathbb{R}^{D\\times K}$ is the matrix of $K$ concepts and $\\pmb{V}\\in \\mathbb{R}^{K\\times M}$ represents the coefficients/activations of the samples projected onto these concepts. Different decompositions of the matrix $K$ result in various concepts, inheriting the properties of the decomposition (such as low grounding overlap with PCA). We employ $K$ -Means to learn our concept dictionaries. This is motivated by $K$ -Means' simplicity, and straightforward arithmetic manipulation of the clusters/concepts it allows. Each column $\\pmb{u}_k\\in \\pmb{U}$ corresponds to a concept, while each column of $\\pmb{V}$ encodes the activation of these concepts for a given sample. Note that any given representation $f_{l}(x)$ can be projected on $\\pmb{U}$ to obtain its activation vector $\\pmb{v}(x)\\in \\mathbb{R}^{K}$ , i.e. $f_{l}(x)\\approx U\\pmb{v}(x)$ . Each extracted concept is then interpreted through grounding in both image and text spaces. Specifically, the top $N_{\\mathrm{MAS}}$ that activates concept $\\pmb{u}_k$ the most represent its image grounding:", + "bbox": [ + 91, + 386, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {X} _ {\\text {M A S}} \\left(\\boldsymbol {u} _ {k}\\right) = \\underset {\\hat {X} \\subset \\mathbf {X} _ {t}, | \\hat {X} | = N _ {\\text {M A S}}} {\\arg \\max } \\sum_ {x \\in \\hat {X}} \\left| \\boldsymbol {v} _ {k} (x) \\right|, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 795, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{v}_k(x)$ refers to the activation of $\\pmb{u}_k$ for image $x$ . For text grounding, we decode the features using the unembedding matrix of the language model $W_U$ [5, 32, 43, 54]. Specifically, the operation $W_U\\pmb{u}_k \\in \\mathbb{R}^{|\\mathcal{Y}|}$ produces", + "bbox": [ + 89, + 840, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "logits over the vocabulary, and the top $N_{\\mathrm{grounding}}$ words with highest logits are extracted:", + "bbox": [ + 511, + 386, + 906, + 416 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} _ {k}) = \\underset {\\text {T o p} - N _ {\\text {g r o u n d i n g}}} {\\arg \\max } \\left(W _ {U} \\boldsymbol {u} _ {k}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 431, + 906, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, to quantify the similarity of two concepts (i.e. the columns of $\\mathbf{U}$ we define the Text Grounding Overlap as:", + "bbox": [ + 511, + 465, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {T} \\text {O v e r l a p} (\\boldsymbol {u}, \\boldsymbol {u} ^ {\\prime}) = 1 0 0 \\times \\frac {\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\cap \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} ^ {\\prime}) \\right|}{\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\right|}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 506, + 906, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now that we have defined a generic framework, we present two subcases for extracting and manipulating concepts.", + "bbox": [ + 511, + 550, + 906, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evolution of concepts through fine-tuning.", + "text_level": 1, + "bbox": [ + 511, + 590, + 872, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Setup overview. An original model $f^a$ is typically finetuned to produce a specialized model $f^b$ for a particular task—or, specifically, for a set of target concepts. This finetuning can be conducted on samples that include a set of words $\\{w_1, \\dots, w_m\\}$ associated with these target concepts. For instance, if we fine-tune a image captioning model to emphasize colors in the image, the set of words will simply be these colors. Efficient fine-tuning is typically achieved using Low-Rank Adaptation (LoRA) [25, 34, 39]. Finetuning can selectively alter certain representations, leading to shifts in the conceptual space encoded by the model. Using the interpretability framework discussed in Section 3.1 we can study these shifts at a readable conceptual level.", + "bbox": [ + 509, + 613, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concept recovery via shift vectors. To study the change from an original model $f^a$ to a finetuned model $f^b$ , we fix the dataset $S^{(1)} = S^{(2)}$ , and obtain two sets of embeddings from $f^a, f^b$ respectively, i.e. $A \\approx U^a V^a$ , $B \\approx U^b V^b$ , where $U^a, U^b \\in \\mathbb{R}^{D \\times K}$ are $K$ concepts extracted from each model. We propose to characterize the concept changes", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2208", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from an original to fine-tuned model as linear directions in embedding space or concept shift vectors. To do so, we first associate each original concept $\\pmb{u}_k^a\\in U^a$ with a subset of samples where $\\pmb{u}_k^a$ is the most activated concept:", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {A} _ {k} = \\{m \\mid k = \\underset {i} {\\arg \\max} | \\pmb {v} _ {i} ^ {a} (x _ {m}) | \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 159, + 413, + 184 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each sample $x_{m}, m \\in \\mathbf{A}_{k}$ we define $\\delta_{m}^{a \\to b} = \\pmb{b}_{m} - \\pmb{a}_{m}$ as the change in its representation from $f^{a}$ to $f^{b}$ . To compute the concept shift vector $\\pmb{\\Delta}_{k}^{a \\to b}(\\pmb{u}_{k}^{a})$ associated with $\\pmb{u}_{k}^{a}$ , we aggregate shifts of its associated samples specified by $\\mathbf{A}_{k}$ :", + "bbox": [ + 89, + 191, + 485, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {\\Delta} _ {k} ^ {a \\rightarrow b} (\\pmb {u} _ {k} ^ {a}) = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} \\delta_ {m} ^ {a \\rightarrow b} = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} (\\pmb {b} _ {m} - \\pmb {a} _ {m})\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 271, + 483, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The concept shift vector is used to shift each concept in the original model $\\boldsymbol{u}_k^a$ to obtain the shifted concept $\\boldsymbol{u}_k^s$ :", + "bbox": [ + 89, + 315, + 483, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {u} _ {k} ^ {s} = \\boldsymbol {u} _ {k} ^ {a} + \\alpha \\Delta_ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 352, + 483, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a coefficient to control the shift magnitude. Unless otherwise stated, we use $\\alpha = 1$ as the default magnitude of shift. It is worth noting that given the concept shift vectors, the computation of shifted concepts does not rely on accessing the fine-tuned model. Practically speaking, this means that we can \"push\" the original model towards the concepts of a fine-tuned one with very little overhead, by simply shifting its latent representation in the direction of the shift vector, as it will be illustrated in the experiments.", + "bbox": [ + 89, + 378, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Concept evolution across datasets and applications to model steering", + "text_level": 1, + "bbox": [ + 89, + 522, + 483, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept comparison between different datasets. Another interesting subcase of the proposed framework consists in evaluating the shift from one dataset $S^{(1)}$ to another $S^{(2)}$ , using the same model and encoding $h$ . Beyond interpretability of a model behavior, it also find applications for model steering. Model steering (see Fig. 1 (right)) refers to guiding the model outputs towards desired outcomes by modifying the features without altering the model weights.", + "bbox": [ + 89, + 559, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-grained model steering. In coarse-grained or global steering, the objective is to adjust the model outputs $\\hat{y}$ to generally align with a set of target samples (e.g., changing answers type). Given input-output samples, we first extract the answer representations $B = b_{1},\\dots,b_{N}$ at layer $l$ from the target set. Similarly, we obtain representations for the original set $A = a_{1},\\dots,a_{M}$ (e.g. randomly drawn from the train set). We then compute the coarse steering vector $s_{c}$ as:", + "bbox": [ + 89, + 680, + 483, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {c} = \\frac {\\sum_ {i} ^ {N} \\boldsymbol {b} _ {i}}{N} - \\frac {\\sum_ {i} ^ {M} \\boldsymbol {a} _ {i}}{M}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 810, + 483, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$s_c$ is applied to all the samples in the validation set. For instance, the activations $f_{l}(x_{i})$ of a sample $x_{i}$ become:", + "bbox": [ + 89, + 845, + 483, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {f} _ {l} \\left(x _ {i}\\right) = f _ {l} \\left(x _ {i}\\right) + \\alpha \\boldsymbol {s} _ {c} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 883, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/90bd6d7f34acd49ea52a792b06f82d4fa94c74e52851369fd682616b18ae41ee.jpg", + "image_caption": [ + "Figure 2. Concepts extracted from original and fine-tuned models. concepts from the original $f^a$ (top), and model fine-tuned to focus more on places $f^b$ (bottom), for $TOI =$ person. The concepts from $f^b$ exhibit a stronger association with places." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 901, + 170 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ controls the steering strength and it is set to 1 (we study $\\alpha$ in App. B.5.3). Thus, in this setup, all examples are coarsely steered in the direction of the steering vector $s_c$ at layer $l$ , before passing $\\tilde{f}_l(x_i)$ through the rest of layers and $\\tilde{f}_l(x_i)$ becomes the input to the next layer $l + 1$ .", + "bbox": [ + 511, + 250, + 906, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fine-grained steering. Unlike global steering, fine-grained steering consists in finding and editing directions that adjust only certain concepts to other ones. To do this, we decompose the hidden states of a set of samples into a set of concepts $\\mathbf{U}$ as previously explained. We then compute a set of fine-grained steering vectors $s^f = s_{11}^f, \\ldots, s_{NN}^f$ , with $s_{ij}^f = s_{ij}^f = u_j - u_i$ the steering vector from concept $u_i$ to $u_j$ . However, not all steering vectors are meaningful: options for finding the relevant ones include proximity matching, as well as identifying vectors that have the strongest impact on guiding the model towards generating specific answers or concepts (e.g. producing significantly more target answers). This is more detailed in App. B.3. Various applications of MLLM steering can be explored, such as gender debiasing and safety alignment, as it will be shown below.", + "bbox": [ + 511, + 327, + 908, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 566, + 645, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Fine-tuning experiments", + "text_level": 1, + "bbox": [ + 511, + 590, + 738, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we study how fine-tuning introduces changes in the overall structure of the learned concepts in MLLMs (with the architecture described in Section 3).", + "bbox": [ + 511, + 613, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. The main paper covers experiments on the popular LLaVA [38] model comprising a CLIP image encoder, a two-layer MLP connector, and a 7B Vicuna-1.5 LLM. More experiments on a different multimodal model can be found in App. A. We conduct our study in a controlled setup that consists in specializing the model on a target dataset. Specifically, we apply fine-tuning on three different subsets of the Visual Genome dataset [31], related to places, colors, and sentiments (more details in App. A.2). Impact of fine-tuning on learned concepts. Fig. 2 depicts this concept change. Through the concepts groundings for an example TOI (person) we see that the fine-tuned model puts a stronger emphasis on places, which is expected and serves as a sanity check for our method.", + "bbox": [ + 511, + 659, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Matched concepts. To further analyze how each concept changes after fine-tuning, we focus on each concept and its", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c10d0f78db351f015b227b953972a226d4fcbfd61fd6ab3a813acf94dfa66fd8.jpg", + "image_caption": [ + "Figure 3. Concepts text grounding change after fine-tuning. Text grounding for concepts $(TOI = \\text{bus})$ from $f^a$ and their match from $f^b$ , (fine-tuned to focus more on places). Emerging concepts may include grounding words not explicitly included in the fine-tuning vocabulary for place (e.g., \"District\", \"Crossing\"), while others evolve more smoothly (e.g., \"Street\")." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 460, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "match. Specifically, we define a matching function $m: i \\to j^*$ which associates each concept vector $\\pmb{u}_i^a$ in set $\\pmb{U}^a$ to its closest vector $\\pmb{u}_{j*}^b$ in set $\\pmb{U}^b$ based on cosine similarity, i.e. $m(i) = \\arg \\max_{\\pmb{u}_j^b \\in \\pmb{U}_b} \\cos(\\pmb{u}_i^a, \\pmb{u}_j^b)$ .", + "bbox": [ + 89, + 347, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 shows the text groundings for various concepts, displaying the words with a frequency lower than 5 across concepts (e.g., filtering out high-frequency terms like bus, vehicle, etc.). We observe the emergence of place-related terms across the identified elements, while the overall thematic structure remains consistent. Note that certain concepts may converge toward the same fine-tuned concept. We also analyze how the distances between matched concepts evolve during fine-tuning in App. A.3.", + "bbox": [ + 89, + 412, + 483, + 547 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/777d941f6424d01022b485dab279e939b753602f35e760f2a92a6a45835c844b.jpg", + "image_caption": [ + "Figure 4. Text grounding overlap (T-Overlap) between original and fine-tuned model concepts. Different concepts change to different extents depending on the fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 91, + 559, + 480, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept evolution. To quantify how much a concept $\\boldsymbol{u}_i^a \\in \\boldsymbol{U}^a$ is changed after fine-tuning, we compute the overlap between its grounding words and those of its closest matching concept from the fine-tuned model $\\boldsymbol{U}^b$ . Specifically, we compute T-Overlap $(\\boldsymbol{u}_i^a, \\boldsymbol{u}_{m(i)}^b)$ (Eq. (3)) for all the concepts $i \\in \\{1, \\dots, K\\}$ , and visualize them for different fine-tunings in Fig. 4. We observe varying rates of change across different concepts and fine-tunings. This might be due to the difference in the fine-tuning dataset size, complexity,", + "bbox": [ + 89, + 762, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "or similarity to the original dataset. It also highlights 2 main behaviors, detailed as follows:", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Concepts that are refined. This group contains the concepts that slightly change to be more specialized towards the fine-tuning task (Fig. 5 top, middle rows). These concepts exhibit a relatively high (T-Overlap $(\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)$ ).", + "- Concepts that change completely. This group contains the concepts that emerge or, to a certain extent, disappear (Fig. 5 bottom row) in the fine-tuned model. New concepts emerge during fine-tuning likely due to the introduction of novel patterns or relationships. These concepts exhibit a relatively low (T-Overlap $(\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)$ )." + ], + "bbox": [ + 511, + 122, + 908, + 277 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/614281c11141c2af45f53ba86312363b394d893afb497f4689914332056ce9de.jpg", + "image_caption": [ + "Figure 5. Concepts evolve differently due to fine-tuning. Left: concepts extracted from the original model $f^a$ . Right: matched concept extracted from $f^b$ , fine-tuned to focus on places. Each concept is grounded in image and text. We observe different levels of adaptation across concepts: some concepts specialize further by adding to place-related words (e.g., Top), some concepts introduce place-related words while staying aligned with the original concept in textual or visual groundings (e.g., Middle), while others undergo complete transformation, diverging significantly from their original meaning to fully embrace place-related elements (e.g., Bottom)." + ], + "image_footnote": [], + "bbox": [ + 531, + 303, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also notice that T-Overlap decreases with the number of training iterations, indicating that fine-tuning leads to deviation from the original concepts (more details in App. A.3). Evaluating fine-tuned concept recovery. To study if the fine-tuned concepts $\\pmb{U}^{b}$ can be recovered from the original ones $\\pmb{U}^{a}$ , we first establish a matching $(m : i \\rightarrow j)$ between the set of original $\\{\\pmb{u}_{i}^{a}\\}_{i=1}^{K}$ and fine-tuned concepts $\\{\\pmb{u}_{j}^{b}\\}_{j=1}^{K}$ . For systematic evaluation of recovery of all fine-tuned concepts, we constrain $m$ to be bijective using an optimal transport algorithm detailed in App. A.1. Finally, we evaluate how well a shifted concept $\\pmb{u}_{k}^{s}$ (Equ. (4)) is similar to its match $\\pmb{u}_{m(k)}^{b}$ using the aforementioned T-Overlap metric. Fig. 6 shows the results of recovering the fine-tuned concepts for models fine-tuned on different subsets of the VG", + "bbox": [ + 511, + 688, + 910, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2210", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5492ad358c17c4139aeef04a9c3285898c141926b586baf66e5256296d256e22.jpg", + "image_caption": [ + "Figure 6. Recovering fine-tuned concepts. Across different finetunings (places (top), colors (middle) and sentiments (bottom)), we compute the average text grounding overlap of original and shifted concepts with the matched fine-tuned ones. Shifting the original concepts result in partially recovering the fine-tuned ones." + ], + "image_footnote": [], + "bbox": [ + 99, + 94, + 473, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset (place, color, sentiment). We report the T-Overlap between the shifted $u_{k}^{s}$ and fine-tuned concepts $u_{m(k)}^{b}$ for various different tokens of interest. For each target token and finetuning we extract $K = 20$ concepts and report the mean and standard deviation over them. We use the overlap between the original concepts $u_{k}^{a}$ and the fine-tuned ones as a baseline. We observe that most shifted concepts show higher overlap than the original ones: this demonstrates that fine-tuned concepts can be efficiently recovered from the original model with concept shift vectors.", + "bbox": [ + 89, + 372, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f5d587dc3cd6761cb01fdd6ba1f2740fc4c487e9ffe493c4d7b03527b89838d0.jpg", + "image_caption": [ + "Figure 7. Correlation between shift consistency and concept recovery (Color finetuning). The more consistent and aligned the individual representation shifts associated with a concept, the better the recovery of the fine-tuned concept." + ], + "image_footnote": [], + "bbox": [ + 109, + 542, + 460, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Which concepts are recovered better? We hypothesize that if the representation shift for individual samples associated with a concept $\\boldsymbol{u}_k^a$ , $\\{\\delta_m^{a\\rightarrow b}|m\\in \\boldsymbol{A}_k\\}$ , is consistently aligned with the concept shift vector, the resulting $\\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a)$ , should be more effective at recovering the finetuned concept. We quantify the consistency through mean cosine similarity of $\\{\\delta_m^{a\\rightarrow b}\\}_{m\\in \\boldsymbol{A}_k}$ with the concept shift vector $\\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a)$ . In other words, this quantifies the alignment", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of individual shifts and their mean, $\\pmb{\\Delta}_k^{a\\rightarrow b}(\\pmb{u}_k^a)$ :", + "bbox": [ + 511, + 90, + 830, + 108 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o n s i s t e n c y} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right) = \\frac {1}{\\left| \\boldsymbol {A} _ {k} \\right|} \\sum_ {m \\in \\boldsymbol {A} _ {k}} \\cos \\left(\\delta_ {m}, \\boldsymbol {\\Delta} _ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 114, + 887, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We measure the recovery of concept $k$ , $\\mathrm{CR}_k$ , as the improvement in similarity between the matched fine-tuned $\\boldsymbol{u}_{m(k)}^b$ and shifted concept $\\boldsymbol{u}_k^s$ , relative to the original one $\\boldsymbol{u}_k^a$ :", + "bbox": [ + 511, + 161, + 906, + 208 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C R} _ {k} = \\frac {\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {s}\\right) - \\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)}{\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 215, + 906, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We plot the consistency and recovery for concepts extracted across four tokens of interest for color finetuning in Fig. 7. Plots for other finetunings are in App. A.5. Crucially, we observe a positive and statistically significant correlation between the two quantities across all the finetuning tasks. This supports our hypothesis that a better concept shift recovery is related to consistency of individual shifts of original concept. More ablation studies about concept recovery are available in App. A.4, analyzing the influence of steering strength $\\alpha$ , number of concepts $K$ , and extraction layer $l$ .", + "bbox": [ + 511, + 263, + 908, + 416 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/299dd8fe0458bcc4305331a96826d04fa3c8832368af4528aaa6862f7df71b8f.jpg", + "image_caption": [ + "Figure 8. Linear separability of concepts features in MLLMs. We visualize the features related to the concepts \"yes\" and \"no\" after PCA projections across MLLMs layers." + ], + "image_footnote": [], + "bbox": [ + 517, + 426, + 906, + 493 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, we demonstrated the feasibility of recovering target fine-tuned concepts by applying simple per-concept shift of the original model features. This supposes that the features related to different concepts are almost linearly separable, which empirically seems to hold at least for the last MLLM layers as pictured on Fig. 8, and as previously studied for LLMs in [42, 48]. This motivates the following investigations on using a similar methodology for simple, computationally inexpensive, yet efficient MLLM steering.", + "bbox": [ + 511, + 551, + 908, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Multimodal model steering", + "text_level": 1, + "bbox": [ + 511, + 696, + 758, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform steering by applying a steering vector $\\pmb{v}$ to the residual stream features $Z$ of the MLLM, without changing its parameters. We first evaluate the MLLM steering capabilities in a visual question-answering (VQA) setup. Then, we show the applicability of our MLLM steering to control captioning styles. Lastly, we present two steering applications: gender debiasing, aiming to mitigate biases in model outputs, and safety alignment, i.e. ensuring that the model refuses to generate harmful information. We discuss technical details for each of these applications in App. C and App. D.", + "bbox": [ + 511, + 719, + 908, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Setup. For VQA tasks, each query consists of a question about an input image, and the model generates an answer.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2211", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To measure the effectiveness of our approach in directing the model towards specific answers or answer types, we report the number of generated answers that align with the target output or answer type. Additionally, we aim for targeted steering, ensuring that only specific answer types are influenced. For example, when altering answers from \"yes\" to \"no\" within the \"yes/no\" category, responses to other question types should remain unaffected. This specificity is assessed by tracking accuracy across answer types and number of answers from each type.", + "bbox": [ + 88, + 90, + 482, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation details. Experiments in the main paper are primarily conducted on LLaVA [39] for conciseness. However, we show in App. B.2 that our method is generally applicable to other popular MLLMs. We experiment on VQAv2 [24], a visual question-answering corpus with image-question-answer triplets and annotated answer types (\"yes/no\", \"number\", and \"other\"). Steering vectors are derived from a subset of the train set, with model performance evaluated on the val set. As steering becomes more effective in deeper layers (see Fig. 8 and App. B), we apply it on the last layer. Additional experiments can be found in App. B.", + "bbox": [ + 89, + 243, + 483, + 411 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/02ac6cfa30e14a827b5f27ed97fc7de827b440993ac669570fee2592b83e324f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Target TypeAnswers Type
yes/nonumberother
N/A366122494
yes/no55796288
number327201390
other364115501
", + "bbox": [ + 181, + 425, + 393, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coarse and Fine-Grained Model Steering for VQA. We explore both coarse and fine-grained steering in VQA tasks. Specifically, coarse steering aims to alter the distribution of answers, while fine-grained steering targets specific responses. For coarse-grained steering, we direct the model's answers toward a particular category: yes/no, numbers, or other (e.g., colors, objects). We compute a steering vector for each target answer type. As shown in Table 1, applying steering significantly increases the proportion of the targeted answer type. For fine-grained steering, we first assess the feasibility of identifying such steering vectors. Fig. 9 illustrates examples of these vectors. We derive them from three sets of sample answers corresponding to \"yes/no,\" \"number,\" and \"other\" categories. Interestingly, some vectors distinctly align with specific answers such as \"No,\" \"4,\" and \"Red.\" This confirms the potential to identify fine-grained steering vectors capable of guiding the model toward a precise response. Building on these insights, we seek to steer the model toward a user-specified answer. For each original/target answer pair (e.g., yes/no), we collect some samples and compute the corresponding (coarse) steering vector. We then apply these vectors to all validation set samples. In", + "bbox": [ + 91, + 568, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2437368440318cfa7421b8fec918c07aac795a916de04c757b4498e1bc11f4cb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 555, + 89, + 656, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2eafbad9d9f38f450f1d7c90b8f83717326881a7fdb394bfa8e4ccbdb06057aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 90, + 759, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5dbd2cd46908a19efc09ce597283e096720d1d0d1e12a8a1519e538d1561d14e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 90, + 866, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adc0117383fb2ab858d38ae414b29cb2ba181e696fb506c80b32ce17455b2b1c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 180, + 653, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c5394683e576cc3287e39c9a7c1bb06bfb5f74aba6e291ac6fd7b7eb3a88d544.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 181, + 761, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a786d751a8b150d9408fb1a8d1ae2452ecea1c8174587e99a4b929145e8ddde7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 181, + 864, + 265 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a7c6ec020e1e00fabaded1ca98e6ed98e0dc2e2ae63193e1318d1e6456cfcc2.jpg", + "image_caption": [ + "Figure 9. Discovering meaningful steering directions. Each line corresponds to a fine-grained steering direction to steer the model answer to: \"No\" (yes/no), \"4\" (number) and \"Red\" (other). Some steering directions are targeted (e.g., \"No\") as there is slight change in both the accuracy and number of answers types on other types (e.g., number, other). We show the relative scores compared to a baseline with no steering." + ], + "image_footnote": [], + "bbox": [ + 557, + 271, + 653, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c54b6c6263624626db32666059d183fa4bef536eeaa6b916092b22e6e261687.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 272, + 761, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3b567c38f1d4f3240a4c9354fd1b32de7d421a883f844d9321f123281dfdb812.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 271, + 864, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2, we report evaluation metrics when steering at the last layer. The results show that steering effectively increases the occurrence of target answers, while accuracy on other answer types remains largely stable.", + "bbox": [ + 511, + 465, + 906, + 527 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8d676b4eac264586d49b16eebb1a2ef42c798a8ecd601bc177df73f46105dd9.jpg", + "table_caption": [ + "Table 1. Steering MLLMs answers type. Number of target answers type increases significantly after model steering." + ], + "table_footnote": [], + "table_body": "
SteeringAccuracy (%)Answer TypesAnswers
Yes/NoNumberOtherYes/NoNumberOtherOriginalTarget
N/A90.8258.4771.101861687234900
Yes → No69.0356.8268.9918846952294-828+828
1 → 390.7154.5271.1218616702350-215+144
White → Black90.4058.4258.3618616712312-98+441
", + "bbox": [ + 516, + 537, + 903, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Steering MLLMs answers. Steering answers from \"Yes\" (Yes/No), \"1\" (Number), \"White\" (Other) to \"No\", \"3\", \"Black\" respectively. The number of original/target answer counts decreases/increases significantly, while the accuracy on other answer types changes only slightly, and the number of answer type counts remains almost constant.", + "bbox": [ + 511, + 616, + 906, + 699 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/32e4ee70a330dce9139f5043b03d94a5a748ed60e6da9568c95fe151cf32fd4d.jpg", + "image_caption": [ + "Figure 10. Steering MLLMs captions style. Captions steered to focus more on colors (left), places (middle) and sentiments (right)." + ], + "image_footnote": [], + "bbox": [ + 517, + 724, + 645, + 806 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6eed37fede6ea1430cd86ae9cf2c93fcd37c3b4edb13ea3fe9b3a8f5b564b7c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 726, + 774, + 805 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4166487faa950fc72e254be5e84cb072a94f113261e82c27d925cf772711959d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 724, + 903, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Steering image caption styles. We previously applied steering on relatively brief answers from the VQAv2 dataset. Here, we extend this approach to longer, descriptive outputs", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2212", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "using the COCO captioning dataset [37]. Given that multiple captions can effectively describe an image by emphasizing various aspects such as the main object, surroundings, actions, or events, we aim at modifying captions to align with a specific target style. Here, we learn a coarse steering vector between samples with predicted captions in the target style and random samples. Qualitative examples of this steering for LLaVA are in Fig. 10. Results in Table 3 demonstrate that captions can be effectively steered towards a target style even when considering tasks with longer responses. We provide more captioning experiments in App. B.4.", + "bbox": [ + 89, + 90, + 483, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/76f9ff0ea4f2f1cadc5823c2c03effaa8fcaba4012c7f63616de4e69ca3be7cc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Target StyleCaptionsStyle
placescolorssentiments
N/A43013092
places79610771
colors48825611
sentiments393104048
", + "bbox": [ + 173, + 268, + 405, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Gender debiasing captions. We perform gender debiasing on COCO test set, aiming at mitigating biases for any gendered nouns when captioning. We experiment with both coarse and fine-grained steering. The coarse steering vector is computed between sets of samples with a gendered/neutral noun in the caption. For fine-grained steering, we steer each concept to its closest (as defined by its posing similarity) counterpart among the neutral concepts, as explained in Section 3.3. The results are reported in Table 4. Interestingly, both strategies are capable of converting many gendered captions to neutral ones, with fine-grained steering being significantly more effective than coarse steering.", + "bbox": [ + 89, + 439, + 483, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Safety alignment. Pure text LLMs often exhibit stronger safety alignment compared to MLLMs [11]. Empirical evidence for this can be found in App. D. Using this insight, we construct two sets of samples, categorized as safe and unsafe, by evaluating the model's response to identical malicious content presented in different modalities: one conveyed through text and the other through text + image. We use these to compute a safety guard steering vector. This steering vector gives the model a higher level of safety, without affecting its usefulness for safe tasks (Table 5). We evaluate the safety of the model by ASR (attack success rate) metric, described in detail in App. D.", + "bbox": [ + 89, + 621, + 483, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 89, + 814, + 204, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. The effectiveness of our method relies on the fact that the concepts are represented by linear directions in latent space. However, recent work [14] has found that not all features are captured as such: thus, when analyzing recovery", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/034270970525cdb221aab4d495a30beae03c4c37730514e8cf0164f4f2a07e95.jpg", + "table_caption": [ + "Table 3. Steering MLLMs captions style. Each line corresponds to a different steering vector. Steering towards a target style increases the number of captions with that style." + ], + "table_footnote": [], + "table_body": "
ModelTotalMethodGendered → Neutral
LLaVA-1.5794coarse232
fine-grained632
Idefics2815coarse237
fine-grained315
Qwen2-VL-Instruct926coarse134
fine-grained300
Before SteeringAfter Steering
A young boy with curly hair is playing a video game.A child with curly hair is playing a video game.
A man riding a dirt bike on a beach.A person riding a dirt bike on a beach.
", + "bbox": [ + 517, + 88, + 903, + 270 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8ab7a95122e8571ea6bb46eda34555266e50dccb0cced923d375d5e6c110669.jpg", + "table_caption": [ + "Table 4. Gender debiasing results: number of occurrences of gendered terms converted to neutral terms across different models and methods, after steering with $\\alpha = 1$ . Below, qualitative samples illustrate changes in descriptions before and after applying steering." + ], + "table_footnote": [], + "table_body": "
ModelBefore steeringAfter steering
Qwen2-VL-Instruct45/1005/100
", + "bbox": [ + 555, + 361, + 864, + 392 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Enhancing MLLM Safety Through Steering. We evaluate safety using the ASR metric, quantifying the proportion of responses that do notrefuse to provide harmful instructions. Our assessment is conducted on a portion of MM-SafetyBench [40] dataset. A lower ASR is desired when the prompt requires harmful instructions.", + "bbox": [ + 511, + 402, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of fine-tuned concepts, it can be interesting to explore more sophisticated similarity measures and matching algorithms.", + "bbox": [ + 511, + 492, + 906, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we introduced a novel concept-based analysis framework for monitoring and controlling MLLM behavior, offering new insights into how latent representations evolve during fine-tuning and across datasets. To address the former, we proposed concept shift vectors, an efficient method for recovering and interpreting concepts in fine-tuned models relative to their original counterparts. This approach naturally led us to explore the latter case, where we demonstrated the ability to steer model behavior by modifying features – without requiring additional training. Our results show that this technique effectively modifies MLLM answers, enabling applications such as gender debiasing, safety control, and enhanced caption generation that highlight different aspects of an image. By releasing our code, we hope our framework will benefit the community and encourage research towards better understanding of MLLMs, as well as their broader applications in domains such as physical and digital agents [50, 60].", + "bbox": [ + 511, + 523, + 908, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 813, + 681, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work has been partially supported by ANR grant VISA DEEP (ANR-20-CHIA-0022), HPC resources of IDRIS under the file A0160614966 allocated by GENCI, and Cluster PostGenAI@Paris (ANR-23-IACL-0007, FRANCE 2030).", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2213", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1", + "[2] Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717, 2024. 2", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 2", + "[4] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539-1550, 2024. 1, 2", + "[5] Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor V. Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens. ArXiv, abs/2303.08112, 2023. 3", + "[6] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1", + "[8] Shi Chen and Qi Zhao. Rex: Reasoning-aware and grounded explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15586-15595, 2022. 2", + "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2", + "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1", + "[11] Yi Ding, Bolian Li, and Ruqi Zhang. Eta: Evaluating then aligning safety of vision language models at inference time. ArXiv, abs/2410.06625, 2024. 8, 10, 11", + "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1", + "[13] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1, 2021. 1", + "[14] Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are linear. arXiv preprint arXiv:2405.14860, 2024. 8", + "[15] Brandon Huang et al. Multimodal task vectors enable many-shot multimodal in-context learning. In NeurIPS, 2024. 2", + "[16] Yingzhe et al. LIVE: Learnable in-context vector for visual question answering. In NeurIPS, 2024. 2", + "[17] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2023. 2", + "[18] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 2", + "[19] Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, et al. Multimodal autoregressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402, 2024. 2", + "[20] Jiaxin Ge, Sanjay Subramanian, Trevor Darrell, and Boyi Li. From wrong to right: A recursive approach towards vision-language explanation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1173–1185, 2023. 2", + "[21] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems (NeurIPS), pages 9277-9286, 2019. 2", + "[22] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. ArXiv, abs/2311.05608, 2023. 11", + "[23] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via imageto-text transformation. In European Conference on Computer Vision, 2024. 11", + "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7, 3", + "[25] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3, 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2214", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kaichen Huang, Jiahao Huo, Yibo Yan, Kun Wang, Yutao Yue, and Xuming Hu. Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models. arXiv preprint arXiv:2410.04819, 2024. 2", + "[27] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 2", + "[29] Yuchu Jiang, Jiale Fu, Chenduo Hao, Xinting Hu, Yingzhe Peng, Xin Geng, and Xu Yang. Mimic in-context learning for multimodal tasks. arXiv preprint arXiv:2504.08851, 2025. 2", + "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2", + "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123(1):32-73, 2017. 4, 1", + "[32] Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet. Decoderlens: Layerwise interpretation of encoder-decoder transformers. ArXiv, abs/2310.03686, 2023. 3", + "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3", + "[35] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[36] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 10", + "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8, 1, 3, 9", + "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4", + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7", + "[40] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, 2023. 8, 10", + "[41] Oscar Manas, Pau Rodriguez Lopez, Saba Ahmadi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2523-2548, 2023. 1", + "[42] Neel Nanda. Actually, othello-gpt has a linear emergent world model, 2023. 6, 8", + "[43] Nostalgebraist. Interpreting gpt: The logit lens. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens, 2020. Accessed: [date of access]. 3", + "[44] OpenAI. Gpt-4 technical report. arXiv, 2023. 1", + "[45] Haowen Pan, Yixin Cao, Xiaozhi Wang, and Xun Yang. Finding and editing multi-modal neurons in pre-trained transformer. arXiv preprint arXiv:2311.07470, 2023. 2", + "[46] Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023. 2", + "[47] Jayneel Parekh, Pegah Khayatan, Mustafa Shukor, Alasdair Newson, and Matthieu Cord. A concept-based explainability framework for large multimodal models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2", + "[48] Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 6, 8", + "[49] Libo Qin, Qiguang Chen, Hao Fei, Zhi Chen, Min Li, and Wanxiang Che. What factors affect multi-modal in-context learning? an in-depth exploration. arXiv preprint arXiv:2410.20482, 2024. 2", + "[50] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025.8", + "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[52] Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah," + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2215", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. In Advances in Neural Information Processing Systems, 2024. 2", + "[53] Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan D Cotterell. Linear adversarial concept erasure. In International Conference on Machine Learning, pages 18400-18421. PMLR, 2022. 2", + "[54] Mansi Sakarvadia, Arham Khan, Aswathy Ajith, Daniel Grzenda, Nathaniel Hudson, André Bauer, Kyle Chard, and Ian T. Foster. Attention lens: A tool for mechanistically interpreting the attention head information retrieval mechanism. CoRR, abs/2310.16270, 2023. 3", + "[55] Sarah Schwettmann, Neil Chowdhury, Samuel Klein, David Bau, and Antonio Torralba. Multimodal neurons in pretrained text-only transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2862-2867, 2023. 1, 2", + "[56] Mustafa Shukor and Matthieu Cord. Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2", + "[57] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 1", + "[58] Mustafa Shukor, Corentin Dancette, and Matthieu Cord. ep-alm: Efficient perceptual augmentation of language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22056-22069, 2023. 1", + "[59] Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2", + "[60] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. 8", + "[61] Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, and Alaaeldin El-Nouby. Scaling laws for native multimodal models. arXiv preprint arXiv:2504.07951, 2025. 1", + "[62] Nishant Subramani, Nivedita Suresh, and Matthew E Peters. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566-581, 2022. 2", + "[63] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 2", + "[64] Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023. 2", + "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1, 2", + "[66] Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023. 2", + "[67] Théophane Vallaeys, Mustafa Shukor, Matthieu Cord, and Jakob Verbeek. Improved baselines for data-efficient perceptual augmentation of llms. arXiv preprint arXiv:2403.13499, 2024. 1, 2", + "[68] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, 2024. 10", + "[69] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. arXiv preprint arXiv:2404.03592, 2024. 2", + "[70] Dizhan Xue, Shengsheng Qian, and Changsheng Xu. Few-shot multimodal explanation for visual question answering. In ACM Multimedia 2024, 2024. 2", + "[71] Chih-Kuan Yeh, Been Kim, Sercan O Arik, Chun-Liang Li, Pradeep Ravikumar, and Tomas Pfister. On concept-based explanations in deep neural networks. arXiv preprint arXiv:1910.07969, 2019. 2", + "[72] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2", + "[73] Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. arXiv preprint arXiv:2406.06579, 2024.1" + ], + "bbox": [ + 516, + 90, + 906, + 599 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2216", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_model.json b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f74ad05e58c2af3c0bfa7afbf6284d3b61d69248 --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.897, + 0.201 + ], + "angle": 0, + "content": "Pegah Khayatan\\(^{\\star 1}\\) Mustafa Shukor\\(^{\\star 1}\\) Jayneel Parekh\\(^{\\star 1}\\) Arnaud Dapogny\\(^{1}\\) Matthieu Cord\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.216, + 0.77, + 0.234 + ], + "angle": 0, + "content": "1ISIR, Sorbonne Université, Paris, France 2Valeo.ai, Paris, France" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.276, + 0.327, + 0.292 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.309, + 0.486, + 0.642 + ], + "angle": 0, + "content": "Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.1" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.222, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.485, + 0.775 + ], + "angle": 0, + "content": "With the rapid progress in Large Language Models (LLMs) [7, 10, 28, 44, 65], Multimodal LLMs (MLLMs) [3, 12, 34, 39, 61] have recently demonstrated remarkable capabilities in addressing complex multimodal tasks such as image captioning and visual question-answering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.486, + 0.851 + ], + "angle": 0, + "content": "MLLMs are typically composed of a visual encoder, an LLM, and a connector. Following initial unimodal pretraining—and, in many cases, multimodal pretraining on large datasets—these models can be further specialized by training on multimodal datasets [1, 33]. Given the high compu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.908, + 0.399 + ], + "angle": 0, + "content": "tational cost of training these models, recent research has proposed more efficient approaches, such as creating diverse, high-quality instruction-tuning datasets [39] or keeping the LLM frozen and fine-tuning small amounts of parameters, like the connector [41, 57, 58, 67]. These approaches take advantage of the ability of frozen LLMs to generalize to multimodal data [56]. Despite the different efficient tuning methods, training these models still incurs significant costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.909, + 0.519 + ], + "angle": 0, + "content": "While substantial progress has been made in developing high-performing MLLMs, relatively few studies aim to understand them [4, 47, 55, 56, 59, 61, 73]. Existing work typically conducts post-hoc analyses of MLLMs in isolation, overlooking the internal changes due to fine-tuning. Research by [56] addresses this gap to some extent by examining the internal multimodal alignment as it evolves during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.908, + 0.58 + ], + "angle": 0, + "content": "In this work, we apply concept-level analysis to provide a readable understanding of MLLM behavior and, in particular, semantic dynamics that may occur due to fine-tuning, or due to covariate shift when considering different datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.762 + ], + "angle": 0, + "content": "In the first case, we find that fine-tuning on a specific task potentially reshapes learned latent concepts, with some adjusting subtly to align with the task, and others emerging or disappearing altogether (see Fig. 1). Notably, we find that most fine-tuned concepts can be reconstructed from the original model by translating its original concepts in the direction of specific concept shift vectors, reducing the need for additional training and its associated costs. Furthermore, we explore the implications of the proposed analysis for MLLMs steering, demonstrating how model outputs can be modified inexpensively without additional training. Our key findings are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.906, + 0.825 + ], + "angle": 0, + "content": "- We apply latent concept-level analysis to provide readable understanding on MLLMs' behavior ; in particular, we show that fine-tuning can introduce significant alteration in the original concepts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.856 + ], + "angle": 0, + "content": "- We show that we can control the MLLM's behavior w.r.t. certain concepts by simply manipulating shift vectors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Lastly, our findings also have direct applications for steering MLLM outputs, which find use for model debiasing as well as safety control." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.864, + 0.186, + 0.875 + ], + "angle": 0, + "content": "*First authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Project page and code: https://pegah-kh.github.io/projects/lmm-finetuning-analysis-and-steering/" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2206" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "In a nutshell, we propose a novel, ready-to-use framework (including code) for MLLM behavior interpretability and control, debiasing and steering, which, we believe, will pave the way for future research." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.169, + 0.233, + 0.185 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.486, + 0.431 + ], + "angle": 0, + "content": "Concept-based explainability. Concept-based explainability methods have emerged as an alternative to traditional feature attribution based methods, that are capable of extracting key semantic features from the model internal representations. Most post-hoc concept-based approaches are based on the idea of concept activation vectors (CAV) [30], which represent concepts as vectors in the activation space. Instead or relying on human annotations, recent works have proposed methods to automatically discover concepts via clustering [21, 71] or matrix decomposition [18], which can be viewed as instances of a dictionary learning problem [17]. Initially focusing on understanding vision models, dictionary learning for concept extraction has been extended to LLMs e.g. using sparse autoencoders [27, 52]. However, none of the prior approaches have been applied to understand MLLMs, with the exception of recently proposed CoX-LMM [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.809 + ], + "angle": 0, + "content": "MLLMs and Explainability. Multimodal LLMs [3, 34, 39, 67] have recently garnered significant interest. They typically adopt a late fusion architecture, and consist of an image encoder [19, 51, 72], a connector, and an LLM [28, 63, 65]. This family of models has inspired extensive research to better understand them and explain their behavior. For example, studies like [26, 45, 55] seek to identify multimodal neurons within LLMs or analyze modality-specific sub networks [56]. Some methods leverage the fact that these models are text-generative to simply generate textual explanations for model outputs [8, 20, 59, 70]. MLLMs benefit from in-context learning capabilities, which have been examined for limitations, including biases [4] and links to hallucinations [59], as well as the factors that may enhance their in-context learning performance [9, 49]. Related to our approach, CoX-LMMs [47] employs dictionary learning to extract multimodal semantic concepts from model representations. However, these studies typically assess models only in their final trained states, overlooking the dynamic changes that occur during training. Only limited works, such as [56], have investigated explaining changes due to fine-tuning, focusing specifically on implicit alignment between image and text modalities. In this work, we investigate how multimodal concepts within the model evolve throughout fine-tuning and explore the implications of these shifts on model steering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Steering models with feature editing. In contrast to editing model weights, representation or feature editing methods [62, 66, 69] aim to modify model outputs without altering the model's weights. A prominent approach within this family involves identifying steering vectors, or directions in the feature space (often within the residual stream), that are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.319 + ], + "angle": 0, + "content": "linked to contrasting concepts. These methods have been applied to language models for various purposes, such as enhancing factuality or reducing hallucinations [46], inducing sentiment shifts or detoxification [64, 66], improving refusals to harmful requests [2], promoting truthfulness by modifying the output of attention heads [35], and erasing specific concepts or biases [6, 53]. However, their application to MLLMs is yet to be explored. Another set of approaches related to steering methods are based on In-context learning (ICL) [15, 16, 29], where prompts are carefully designed to induce desired behavior. Yet, ICL requires predefined demonstrations and lacks interpretability at a concept level. In contrast, our method offers lightweight steering, extending such capabilities to MLLMs without requiring any training, for instance as in ReFT [69]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.734, + 0.344 + ], + "angle": 0, + "content": "3. Methodology Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.909, + 0.474 + ], + "angle": 0, + "content": "Our framework is summarized in Fig. 1. We apply concept-level analysis of MLLM latent space (Section 3.1) to define and compare concepts between different setups. This allows us to monitor conceptual changes occurring during fin-tuning and manipulate MLLM behavior at a conceptual level (Section 3.2). Lastly, our framework finds applications for MLLM steering for e.g. debiasing and safety control (3.3) with negligible computational burden." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.909, + 0.673 + ], + "angle": 0, + "content": "MLLM setup. A generic MLLM consists of a visual encoder \\( f_{V} \\), a trainable connector \\( C \\), and a language model \\( f_{\\mathrm{LM}} \\). We assume that the model is pretrained on a multimodal (e.g. captioning) dataset \\( S = \\{(x_i, y_i)\\}_i \\), where \\( x_i \\in \\mathcal{X} \\) represents images and \\( y_i \\subset \\mathcal{Y} \\) are the associated captions specified as sequence of tokens from token vocabulary space \\( \\mathcal{Y} \\). The model is trained to generate the next text tokens, conditioned on text and images. The input to \\( f_{\\mathrm{LM}} \\) is a sequence of tokens that includes the concatenation of: (1) \\( N_V \\) visual tokens extracted from the image \\( x \\) via the visual encoder and connector \\( (C(f_V(x))) \\), and (2) linearly embedded textual tokens corresponding to the text instruction and previously predicted tokens. This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.684, + 0.825, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\hat {y} ^ {p} = f _ {L M} (h ^ {1}, \\dots , h ^ {N _ {V}}, \\dots , h ^ {p}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.909, + 0.837 + ], + "angle": 0, + "content": "where \\( h^1, \\ldots, h^{N_V} = C(f_V(x)) \\), and \\( h^p = \\operatorname{Emb}(\\hat{y}^{p-1}) \\) with Emb representing the token embedding layer. During generation, the output token \\( \\hat{y}^p \\) is derived by normalizing the last layer \\( (L) \\) tokens \\( h_{(L)}^p \\), then applying the unembedding layer \\( W_U \\) and a softmax operation. The model keeps predicting the next token until the end of the sentence token to obtain the generated response \\( \\hat{y} = \\{\\hat{y}^p\\}_{p > N_V + N_I} \\), where \\( N_I \\) is corresponds to the text instruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.887, + 0.865 + ], + "angle": 0, + "content": "3.1. Retrieval and comparison of latent concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To understand the internal representations of any given MLLM \\(f\\), we leverage the approach introduced in [47]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2207" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.092, + 0.482, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.091, + 0.862, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.232, + 0.861, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 1. Framework overview. We apply concept-level analysis for MLLM behavior monitor and control, for (left) understanding and manipulating (through shift vectors) concept changes due to fine-tuning, as well as (right) MLLM steering for debiasing or safety control." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.387, + 0.485, + 0.78 + ], + "angle": 0, + "content": "Specifically, given a set of \\(M\\) images \\(\\{x_{1},\\dots,x_{M}\\}\\) we extract a set of residual stream representations from some layer \\(l\\) of the MLLM \\(f\\). These representations \\(\\mathbf{z}_m = f_l(x_m)\\in \\mathbb{R}^D\\) (one per image) are collected in a feature matrix \\(\\pmb{Z}\\in \\mathbb{R}^{D\\times M}\\). Typically, the set of images and extracted representations correspond to a particular token of interest TOI (e.g., 'Dog', 'Cat', 'Person', etc.) in the predicted caption. However, the extraction can also be performed for a larger set of target tokens. This feature matrix \\(\\pmb{Z}\\) is then decomposed as \\(\\pmb{Z}\\approx \\pmb{U}\\pmb{V}\\) to recover the concepts in the latent embedding space. Here, \\(\\pmb{U}\\in \\mathbb{R}^{D\\times K}\\) is the matrix of \\(K\\) concepts and \\(\\pmb{V}\\in \\mathbb{R}^{K\\times M}\\) represents the coefficients/activations of the samples projected onto these concepts. Different decompositions of the matrix \\(K\\) result in various concepts, inheriting the properties of the decomposition (such as low grounding overlap with PCA). We employ \\(K\\)-Means to learn our concept dictionaries. This is motivated by \\(K\\)-Means' simplicity, and straightforward arithmetic manipulation of the clusters/concepts it allows. Each column \\(\\pmb{u}_k\\in \\pmb{U}\\) corresponds to a concept, while each column of \\(\\pmb{V}\\) encodes the activation of these concepts for a given sample. Note that any given representation \\(f_{l}(x)\\) can be projected on \\(\\pmb{U}\\) to obtain its activation vector \\(\\pmb{v}(x)\\in \\mathbb{R}^{K}\\), i.e. \\(f_{l}(x)\\approx U\\pmb{v}(x)\\). Each extracted concept is then interpreted through grounding in both image and text spaces. Specifically, the top \\(N_{\\mathrm{MAS}}\\) that activates concept \\(\\pmb{u}_k\\) the most represent its image grounding:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.796, + 0.484, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {X} _ {\\text {M A S}} \\left(\\boldsymbol {u} _ {k}\\right) = \\underset {\\hat {X} \\subset \\mathbf {X} _ {t}, | \\hat {X} | = N _ {\\text {M A S}}} {\\arg \\max } \\sum_ {x \\in \\hat {X}} \\left| \\boldsymbol {v} _ {k} (x) \\right|, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_k(x)\\) refers to the activation of \\(\\pmb{u}_k\\) for image \\(x\\). For text grounding, we decode the features using the unembedding matrix of the language model \\(W_U\\) [5, 32, 43, 54]. Specifically, the operation \\(W_U\\pmb{u}_k \\in \\mathbb{R}^{|\\mathcal{Y}|}\\) produces" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.417 + ], + "angle": 0, + "content": "logits over the vocabulary, and the top \\( N_{\\mathrm{grounding}} \\) words with highest logits are extracted:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.433, + 0.907, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} _ {k}) = \\underset {\\text {T o p} - N _ {\\text {g r o u n d i n g}}} {\\arg \\max } \\left(W _ {U} \\boldsymbol {u} _ {k}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.907, + 0.497 + ], + "angle": 0, + "content": "Finally, to quantify the similarity of two concepts (i.e. the columns of \\(\\mathbf{U}\\) we define the Text Grounding Overlap as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.507, + 0.907, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\operatorname {T} \\text {O v e r l a p} (\\boldsymbol {u}, \\boldsymbol {u} ^ {\\prime}) = 1 0 0 \\times \\frac {\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\cap \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} ^ {\\prime}) \\right|}{\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\right|}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Now that we have defined a generic framework, we present two subcases for extracting and manipulating concepts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.874, + 0.608 + ], + "angle": 0, + "content": "3.2. Evolution of concepts through fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.614, + 0.908, + 0.81 + ], + "angle": 0, + "content": "Setup overview. An original model \\( f^a \\) is typically finetuned to produce a specialized model \\( f^b \\) for a particular task—or, specifically, for a set of target concepts. This finetuning can be conducted on samples that include a set of words \\( \\{w_1, \\dots, w_m\\} \\) associated with these target concepts. For instance, if we fine-tune a image captioning model to emphasize colors in the image, the set of words will simply be these colors. Efficient fine-tuning is typically achieved using Low-Rank Adaptation (LoRA) [25, 34, 39]. Finetuning can selectively alter certain representations, leading to shifts in the conceptual space encoded by the model. Using the interpretability framework discussed in Section 3.1 we can study these shifts at a readable conceptual level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Concept recovery via shift vectors. To study the change from an original model \\( f^a \\) to a finetuned model \\( f^b \\), we fix the dataset \\( S^{(1)} = S^{(2)} \\), and obtain two sets of embeddings from \\( f^a, f^b \\) respectively, i.e. \\( A \\approx U^a V^a \\), \\( B \\approx U^b V^b \\), where \\( U^a, U^b \\in \\mathbb{R}^{D \\times K} \\) are \\( K \\) concepts extracted from each model. We propose to characterize the concept changes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2208" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "from an original to fine-tuned model as linear directions in embedding space or concept shift vectors. To do so, we first associate each original concept \\(\\pmb{u}_k^a\\in U^a\\) with a subset of samples where \\(\\pmb{u}_k^a\\) is the most activated concept:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.16, + 0.414, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\pmb {A} _ {k} = \\{m \\mid k = \\underset {i} {\\arg \\max} | \\pmb {v} _ {i} ^ {a} (x _ {m}) | \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.486, + 0.255 + ], + "angle": 0, + "content": "For each sample \\(x_{m}, m \\in \\mathbf{A}_{k}\\) we define \\(\\delta_{m}^{a \\to b} = \\pmb{b}_{m} - \\pmb{a}_{m}\\) as the change in its representation from \\(f^{a}\\) to \\(f^{b}\\). To compute the concept shift vector \\(\\pmb{\\Delta}_{k}^{a \\to b}(\\pmb{u}_{k}^{a})\\) associated with \\(\\pmb{u}_{k}^{a}\\), we aggregate shifts of its associated samples specified by \\(\\mathbf{A}_{k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.272, + 0.484, + 0.311 + ], + "angle": 0, + "content": "\\[\n\\pmb {\\Delta} _ {k} ^ {a \\rightarrow b} (\\pmb {u} _ {k} ^ {a}) = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} \\delta_ {m} ^ {a \\rightarrow b} = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} (\\pmb {b} _ {m} - \\pmb {a} _ {m})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.348 + ], + "angle": 0, + "content": "The concept shift vector is used to shift each concept in the original model \\( \\boldsymbol{u}_k^a \\) to obtain the shifted concept \\( \\boldsymbol{u}_k^s \\):" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.353, + 0.484, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {u} _ {k} ^ {s} = \\boldsymbol {u} _ {k} ^ {a} + \\alpha \\Delta_ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.485, + 0.515 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a coefficient to control the shift magnitude. Unless otherwise stated, we use \\(\\alpha = 1\\) as the default magnitude of shift. It is worth noting that given the concept shift vectors, the computation of shifted concepts does not rely on accessing the fine-tuned model. Practically speaking, this means that we can \"push\" the original model towards the concepts of a fine-tuned one with very little overhead, by simply shifting its latent representation in the direction of the shift vector, as it will be illustrated in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.523, + 0.484, + 0.556 + ], + "angle": 0, + "content": "3.3. Concept evolution across datasets and applications to model steering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Concept comparison between different datasets. Another interesting subcase of the proposed framework consists in evaluating the shift from one dataset \\( S^{(1)} \\) to another \\( S^{(2)} \\), using the same model and encoding \\( h \\). Beyond interpretability of a model behavior, it also find applications for model steering. Model steering (see Fig. 1 (right)) refers to guiding the model outputs towards desired outcomes by modifying the features without altering the model weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.802 + ], + "angle": 0, + "content": "Coarse-grained model steering. In coarse-grained or global steering, the objective is to adjust the model outputs \\(\\hat{y}\\) to generally align with a set of target samples (e.g., changing answers type). Given input-output samples, we first extract the answer representations \\(B = b_{1},\\dots,b_{N}\\) at layer \\(l\\) from the target set. Similarly, we obtain representations for the original set \\(A = a_{1},\\dots,a_{M}\\) (e.g. randomly drawn from the train set). We then compute the coarse steering vector \\(s_{c}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.811, + 0.484, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {c} = \\frac {\\sum_ {i} ^ {N} \\boldsymbol {b} _ {i}}{N} - \\frac {\\sum_ {i} ^ {M} \\boldsymbol {a} _ {i}}{M}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.484, + 0.879 + ], + "angle": 0, + "content": "\\(s_c\\) is applied to all the samples in the validation set. For instance, the activations \\(f_{l}(x_{i})\\) of a sample \\(x_{i}\\) become:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.884, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\tilde {f} _ {l} \\left(x _ {i}\\right) = f _ {l} \\left(x _ {i}\\right) + \\alpha \\boldsymbol {s} _ {c} \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.903, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.189, + 0.908, + 0.245 + ], + "angle": 0, + "content": "Figure 2. Concepts extracted from original and fine-tuned models. concepts from the original \\( f^a \\) (top), and model fine-tuned to focus more on places \\( f^b \\) (bottom), for \\( TOI = \\) person. The concepts from \\( f^b \\) exhibit a stronger association with places." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.907, + 0.327 + ], + "angle": 0, + "content": "where \\(\\alpha\\) controls the steering strength and it is set to 1 (we study \\(\\alpha\\) in App. B.5.3). Thus, in this setup, all examples are coarsely steered in the direction of the steering vector \\(s_c\\) at layer \\(l\\), before passing \\(\\tilde{f}_l(x_i)\\) through the rest of layers and \\(\\tilde{f}_l(x_i)\\) becomes the input to the next layer \\(l + 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.909, + 0.556 + ], + "angle": 0, + "content": "Fine-grained steering. Unlike global steering, fine-grained steering consists in finding and editing directions that adjust only certain concepts to other ones. To do this, we decompose the hidden states of a set of samples into a set of concepts \\( \\mathbf{U} \\) as previously explained. We then compute a set of fine-grained steering vectors \\( s^f = s_{11}^f, \\ldots, s_{NN}^f \\), with \\( s_{ij}^f = s_{ij}^f = u_j - u_i \\) the steering vector from concept \\( u_i \\) to \\( u_j \\). However, not all steering vectors are meaningful: options for finding the relevant ones include proximity matching, as well as identifying vectors that have the strongest impact on guiding the model towards generating specific answers or concepts (e.g. producing significantly more target answers). This is more detailed in App. B.3. Various applications of MLLM steering can be explored, such as gender debiasing and safety alignment, as it will be shown below." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.646, + 0.585 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.74, + 0.608 + ], + "angle": 0, + "content": "4.1. Fine-tuning experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.658 + ], + "angle": 0, + "content": "In this section, we study how fine-tuning introduces changes in the overall structure of the learned concepts in MLLMs (with the architecture described in Section 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Implementation details. The main paper covers experiments on the popular LLaVA [38] model comprising a CLIP image encoder, a two-layer MLP connector, and a 7B Vicuna-1.5 LLM. More experiments on a different multimodal model can be found in App. A. We conduct our study in a controlled setup that consists in specializing the model on a target dataset. Specifically, we apply fine-tuning on three different subsets of the Visual Genome dataset [31], related to places, colors, and sentiments (more details in App. A.2). Impact of fine-tuning on learned concepts. Fig. 2 depicts this concept change. Through the concepts groundings for an example TOI (person) we see that the fine-tuned model puts a stronger emphasis on places, which is expected and serves as a sanity check for our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Matched concepts. To further analyze how each concept changes after fine-tuning, we focus on each concept and its" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2209" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.462, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.485, + 0.342 + ], + "angle": 0, + "content": "Figure 3. Concepts text grounding change after fine-tuning. Text grounding for concepts \\((TOI = \\text{bus})\\) from \\(f^a\\) and their match from \\(f^b\\), (fine-tuned to focus more on places). Emerging concepts may include grounding words not explicitly included in the fine-tuning vocabulary for place (e.g., \"District\", \"Crossing\"), while others evolve more smoothly (e.g., \"Street\")." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.413 + ], + "angle": 0, + "content": "match. Specifically, we define a matching function \\(m: i \\to j^*\\) which associates each concept vector \\(\\pmb{u}_i^a\\) in set \\(\\pmb{U}^a\\) to its closest vector \\(\\pmb{u}_{j*}^b\\) in set \\(\\pmb{U}^b\\) based on cosine similarity, i.e. \\(m(i) = \\arg \\max_{\\pmb{u}_j^b \\in \\pmb{U}_b} \\cos(\\pmb{u}_i^a, \\pmb{u}_j^b)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Fig. 3 shows the text groundings for various concepts, displaying the words with a frequency lower than 5 across concepts (e.g., filtering out high-frequency terms like bus, vehicle, etc.). We observe the emergence of place-related terms across the identified elements, while the overall thematic structure remains consistent. Note that certain concepts may converge toward the same fine-tuned concept. We also analyze how the distances between matched concepts evolve during fine-tuning in App. A.3." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.704 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Figure 4. Text grounding overlap (T-Overlap) between original and fine-tuned model concepts. Different concepts change to different extents depending on the fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Concept evolution. To quantify how much a concept \\( \\boldsymbol{u}_i^a \\in \\boldsymbol{U}^a \\) is changed after fine-tuning, we compute the overlap between its grounding words and those of its closest matching concept from the fine-tuned model \\( \\boldsymbol{U}^b \\). Specifically, we compute T-Overlap \\( (\\boldsymbol{u}_i^a, \\boldsymbol{u}_{m(i)}^b) \\) (Eq. (3)) for all the concepts \\( i \\in \\{1, \\dots, K\\} \\), and visualize them for different fine-tunings in Fig. 4. We observe varying rates of change across different concepts and fine-tunings. This might be due to the difference in the fine-tuning dataset size, complexity," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "or similarity to the original dataset. It also highlights 2 main behaviors, detailed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.909, + 0.185 + ], + "angle": 0, + "content": "- Concepts that are refined. This group contains the concepts that slightly change to be more specialized towards the fine-tuning task (Fig. 5 top, middle rows). These concepts exhibit a relatively high (T-Overlap \\((\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)\\))." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.909, + 0.278 + ], + "angle": 0, + "content": "- Concepts that change completely. This group contains the concepts that emerge or, to a certain extent, disappear (Fig. 5 bottom row) in the fine-tuned model. New concepts emerge during fine-tuning likely due to the introduction of novel patterns or relationships. These concepts exhibit a relatively low (T-Overlap \\((\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)\\))." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.123, + 0.909, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.304, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.534, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Figure 5. Concepts evolve differently due to fine-tuning. Left: concepts extracted from the original model \\( f^a \\). Right: matched concept extracted from \\( f^b \\), fine-tuned to focus on places. Each concept is grounded in image and text. We observe different levels of adaptation across concepts: some concepts specialize further by adding to place-related words (e.g., Top), some concepts introduce place-related words while staying aligned with the original concept in textual or visual groundings (e.g., Middle), while others undergo complete transformation, diverging significantly from their original meaning to fully embrace place-related elements (e.g., Bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.911, + 0.901 + ], + "angle": 0, + "content": "We also notice that T-Overlap decreases with the number of training iterations, indicating that fine-tuning leads to deviation from the original concepts (more details in App. A.3). Evaluating fine-tuned concept recovery. To study if the fine-tuned concepts \\( \\pmb{U}^{b} \\) can be recovered from the original ones \\( \\pmb{U}^{a} \\), we first establish a matching \\( (m : i \\rightarrow j) \\) between the set of original \\( \\{\\pmb{u}_{i}^{a}\\}_{i=1}^{K} \\) and fine-tuned concepts \\( \\{\\pmb{u}_{j}^{b}\\}_{j=1}^{K} \\). For systematic evaluation of recovery of all fine-tuned concepts, we constrain \\( m \\) to be bijective using an optimal transport algorithm detailed in App. A.1. Finally, we evaluate how well a shifted concept \\( \\pmb{u}_{k}^{s} \\) (Equ. (4)) is similar to its match \\( \\pmb{u}_{m(k)}^{b} \\) using the aforementioned T-Overlap metric. Fig. 6 shows the results of recovering the fine-tuned concepts for models fine-tuned on different subsets of the VG" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2210" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.095, + 0.475, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.368 + ], + "angle": 0, + "content": "Figure 6. Recovering fine-tuned concepts. Across different finetunings (places (top), colors (middle) and sentiments (bottom)), we compute the average text grounding overlap of original and shifted concepts with the matched fine-tuned ones. Shifting the original concepts result in partially recovering the fine-tuned ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.525 + ], + "angle": 0, + "content": "dataset (place, color, sentiment). We report the T-Overlap between the shifted \\( u_{k}^{s} \\) and fine-tuned concepts \\( u_{m(k)}^{b} \\) for various different tokens of interest. For each target token and finetuning we extract \\( K = 20 \\) concepts and report the mean and standard deviation over them. We use the overlap between the original concepts \\( u_{k}^{a} \\) and the fine-tuned ones as a baseline. We observe that most shifted concepts show higher overlap than the original ones: this demonstrates that fine-tuned concepts can be efficiently recovered from the original model with concept shift vectors." + }, + { + "type": "image", + "bbox": [ + 0.11, + 0.544, + 0.462, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.757 + ], + "angle": 0, + "content": "Figure 7. Correlation between shift consistency and concept recovery (Color finetuning). The more consistent and aligned the individual representation shifts associated with a concept, the better the recovery of the fine-tuned concept." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Which concepts are recovered better? We hypothesize that if the representation shift for individual samples associated with a concept \\( \\boldsymbol{u}_k^a \\), \\( \\{\\delta_m^{a\\rightarrow b}|m\\in \\boldsymbol{A}_k\\} \\), is consistently aligned with the concept shift vector, the resulting \\( \\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a) \\), should be more effective at recovering the finetuned concept. We quantify the consistency through mean cosine similarity of \\( \\{\\delta_m^{a\\rightarrow b}\\}_{m\\in \\boldsymbol{A}_k} \\) with the concept shift vector \\( \\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a) \\). In other words, this quantifies the alignment" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.831, + 0.109 + ], + "angle": 0, + "content": "of individual shifts and their mean, \\(\\pmb{\\Delta}_k^{a\\rightarrow b}(\\pmb{u}_k^a)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.116, + 0.888, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o n s i s t e n c y} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right) = \\frac {1}{\\left| \\boldsymbol {A} _ {k} \\right|} \\sum_ {m \\in \\boldsymbol {A} _ {k}} \\cos \\left(\\delta_ {m}, \\boldsymbol {\\Delta} _ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right)\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.162, + 0.908, + 0.209 + ], + "angle": 0, + "content": "We measure the recovery of concept \\( k \\), \\( \\mathrm{CR}_k \\), as the improvement in similarity between the matched fine-tuned \\( \\boldsymbol{u}_{m(k)}^b \\) and shifted concept \\( \\boldsymbol{u}_k^s \\), relative to the original one \\( \\boldsymbol{u}_k^a \\):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.217, + 0.907, + 0.257 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C R} _ {k} = \\frac {\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {s}\\right) - \\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)}{\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.909, + 0.417 + ], + "angle": 0, + "content": "We plot the consistency and recovery for concepts extracted across four tokens of interest for color finetuning in Fig. 7. Plots for other finetunings are in App. A.5. Crucially, we observe a positive and statistically significant correlation between the two quantities across all the finetuning tasks. This supports our hypothesis that a better concept shift recovery is related to consistency of individual shifts of original concept. More ablation studies about concept recovery are available in App. A.4, analyzing the influence of steering strength \\(\\alpha\\), number of concepts \\(K\\), and extraction layer \\(l\\)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.428, + 0.907, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.496, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Figure 8. Linear separability of concepts features in MLLMs. We visualize the features related to the concepts \"yes\" and \"no\" after PCA projections across MLLMs layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.69 + ], + "angle": 0, + "content": "In summary, we demonstrated the feasibility of recovering target fine-tuned concepts by applying simple per-concept shift of the original model features. This supposes that the features related to different concepts are almost linearly separable, which empirically seems to hold at least for the last MLLM layers as pictured on Fig. 8, and as previously studied for LLMs in [42, 48]. This motivates the following investigations on using a similar methodology for simple, computationally inexpensive, yet efficient MLLM steering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.759, + 0.714 + ], + "angle": 0, + "content": "4.2. Multimodal model steering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.871 + ], + "angle": 0, + "content": "We perform steering by applying a steering vector \\(\\pmb{v}\\) to the residual stream features \\(Z\\) of the MLLM, without changing its parameters. We first evaluate the MLLM steering capabilities in a visual question-answering (VQA) setup. Then, we show the applicability of our MLLM steering to control captioning styles. Lastly, we present two steering applications: gender debiasing, aiming to mitigate biases in model outputs, and safety alignment, i.e. ensuring that the model refuses to generate harmful information. We discuss technical details for each of these applications in App. C and App. D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Setup. For VQA tasks, each query consists of a question about an input image, and the model generates an answer." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2211" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.243 + ], + "angle": 0, + "content": "To measure the effectiveness of our approach in directing the model towards specific answers or answer types, we report the number of generated answers that align with the target output or answer type. Additionally, we aim for targeted steering, ensuring that only specific answer types are influenced. For example, when altering answers from \"yes\" to \"no\" within the \"yes/no\" category, responses to other question types should remain unaffected. This specificity is assessed by tracking accuracy across answer types and number of answers from each type." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.412 + ], + "angle": 0, + "content": "Implementation details. Experiments in the main paper are primarily conducted on LLaVA [39] for conciseness. However, we show in App. B.2 that our method is generally applicable to other popular MLLMs. We experiment on VQAv2 [24], a visual question-answering corpus with image-question-answer triplets and annotated answer types (\"yes/no\", \"number\", and \"other\"). Steering vectors are derived from a subset of the train set, with model performance evaluated on the val set. As steering becomes more effective in deeper layers (see Fig. 8 and App. B), we apply it on the last layer. Additional experiments can be found in App. B." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.426, + 0.395, + 0.51 + ], + "angle": 0, + "content": "
Target TypeAnswers Type
yes/nonumberother
N/A366122494
yes/no55796288
number327201390
other364115501
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.521, + 0.485, + 0.551 + ], + "angle": 0, + "content": "Table 1. Steering MLLMs answers type. Number of target answers type increases significantly after model steering." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.569, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Coarse and Fine-Grained Model Steering for VQA. We explore both coarse and fine-grained steering in VQA tasks. Specifically, coarse steering aims to alter the distribution of answers, while fine-grained steering targets specific responses. For coarse-grained steering, we direct the model's answers toward a particular category: yes/no, numbers, or other (e.g., colors, objects). We compute a steering vector for each target answer type. As shown in Table 1, applying steering significantly increases the proportion of the targeted answer type. For fine-grained steering, we first assess the feasibility of identifying such steering vectors. Fig. 9 illustrates examples of these vectors. We derive them from three sets of sample answers corresponding to \"yes/no,\" \"number,\" and \"other\" categories. Interestingly, some vectors distinctly align with specific answers such as \"No,\" \"4,\" and \"Red.\" This confirms the potential to identify fine-grained steering vectors capable of guiding the model toward a precise response. Building on these insights, we seek to steer the model toward a user-specified answer. For each original/target answer pair (e.g., yes/no), we collect some samples and compute the corresponding (coarse) steering vector. We then apply these vectors to all validation set samples. In" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.09, + 0.657, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.091, + 0.761, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.091, + 0.867, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.181, + 0.655, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.182, + 0.762, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.182, + 0.866, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.272, + 0.655, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.273, + 0.762, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.272, + 0.866, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Figure 9. Discovering meaningful steering directions. Each line corresponds to a fine-grained steering direction to steer the model answer to: \"No\" (yes/no), \"4\" (number) and \"Red\" (other). Some steering directions are targeted (e.g., \"No\") as there is slight change in both the accuracy and number of answers types on other types (e.g., number, other). We show the relative scores compared to a baseline with no steering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Table 2, we report evaluation metrics when steering at the last layer. The results show that steering effectively increases the occurrence of target answers, while accuracy on other answer types remains largely stable." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.538, + 0.905, + 0.607 + ], + "angle": 0, + "content": "
SteeringAccuracy (%)Answer TypesAnswers
Yes/NoNumberOtherYes/NoNumberOtherOriginalTarget
N/A90.8258.4771.101861687234900
Yes → No69.0356.8268.9918846952294-828+828
1 → 390.7154.5271.1218616702350-215+144
White → Black90.4058.4258.3618616712312-98+441
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.617, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Table 2. Steering MLLMs answers. Steering answers from \"Yes\" (Yes/No), \"1\" (Number), \"White\" (Other) to \"No\", \"3\", \"Black\" respectively. The number of original/target answer counts decreases/increases significantly, while the accuracy on other answer types changes only slightly, and the number of answer type counts remains almost constant." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.726, + 0.646, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.727, + 0.775, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.726, + 0.904, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.814, + 0.907, + 0.843 + ], + "angle": 0, + "content": "Figure 10. Steering MLLMs captions style. Captions steered to focus more on colors (left), places (middle) and sentiments (right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Steering image caption styles. We previously applied steering on relatively brief answers from the VQAv2 dataset. Here, we extend this approach to longer, descriptive outputs" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.259 + ], + "angle": 0, + "content": "using the COCO captioning dataset [37]. Given that multiple captions can effectively describe an image by emphasizing various aspects such as the main object, surroundings, actions, or events, we aim at modifying captions to align with a specific target style. Here, we learn a coarse steering vector between samples with predicted captions in the target style and random samples. Qualitative examples of this steering for LLaVA are in Fig. 10. Results in Table 3 demonstrate that captions can be effectively steered towards a target style even when considering tasks with longer responses. We provide more captioning experiments in App. B.4." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.269, + 0.406, + 0.354 + ], + "angle": 0, + "content": "
Target StyleCaptionsStyle
placescolorssentiments
N/A43013092
places79610771
colors48825611
sentiments393104048
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.365, + 0.485, + 0.407 + ], + "angle": 0, + "content": "Table 3. Steering MLLMs captions style. Each line corresponds to a different steering vector. Steering towards a target style increases the number of captions with that style." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Gender debiasing captions. We perform gender debiasing on COCO test set, aiming at mitigating biases for any gendered nouns when captioning. We experiment with both coarse and fine-grained steering. The coarse steering vector is computed between sets of samples with a gendered/neutral noun in the caption. For fine-grained steering, we steer each concept to its closest (as defined by its posing similarity) counterpart among the neutral concepts, as explained in Section 3.3. The results are reported in Table 4. Interestingly, both strategies are capable of converting many gendered captions to neutral ones, with fine-grained steering being significantly more effective than coarse steering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.622, + 0.484, + 0.803 + ], + "angle": 0, + "content": "Safety alignment. Pure text LLMs often exhibit stronger safety alignment compared to MLLMs [11]. Empirical evidence for this can be found in App. D. Using this insight, we construct two sets of samples, categorized as safe and unsafe, by evaluating the model's response to identical malicious content presented in different modalities: one conveyed through text and the other through text + image. We use these to compute a safety guard steering vector. This steering vector gives the model a higher level of safety, without affecting its usefulness for safe tasks (Table 5). We evaluate the safety of the model by ASR (attack success rate) metric, described in detail in App. D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.205, + 0.831 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Limitations. The effectiveness of our method relies on the fact that the concepts are represented by linear directions in latent space. However, recent work [14] has found that not all features are captured as such: thus, when analyzing recovery" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.271 + ], + "angle": 0, + "content": "
ModelTotalMethodGendered → Neutral
LLaVA-1.5794coarse232
fine-grained632
Idefics2815coarse237
fine-grained315
Qwen2-VL-Instruct926coarse134
fine-grained300
Before SteeringAfter Steering
A young boy with curly hair is playing a video game.A child with curly hair is playing a video game.
A man riding a dirt bike on a beach.A person riding a dirt bike on a beach.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.28, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Table 4. Gender debiasing results: number of occurrences of gendered terms converted to neutral terms across different models and methods, after steering with \\(\\alpha = 1\\). Below, qualitative samples illustrate changes in descriptions before and after applying steering." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.362, + 0.866, + 0.393 + ], + "angle": 0, + "content": "
ModelBefore steeringAfter steering
Qwen2-VL-Instruct45/1005/100
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Table 5. Enhancing MLLM Safety Through Steering. We evaluate safety using the ASR metric, quantifying the proportion of responses that do notrefuse to provide harmful instructions. Our assessment is conducted on a portion of MM-SafetyBench [40] dataset. A lower ASR is desired when the prompt requires harmful instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.524 + ], + "angle": 0, + "content": "of fine-tuned concepts, it can be interesting to explore more sophisticated similarity measures and matching algorithms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.909, + 0.797 + ], + "angle": 0, + "content": "Conclusion. In this work, we introduced a novel concept-based analysis framework for monitoring and controlling MLLM behavior, offering new insights into how latent representations evolve during fine-tuning and across datasets. To address the former, we proposed concept shift vectors, an efficient method for recovering and interpreting concepts in fine-tuned models relative to their original counterparts. This approach naturally led us to explore the latter case, where we demonstrated the ability to steer model behavior by modifying features – without requiring additional training. Our results show that this technique effectively modifies MLLM answers, enabling applications such as gender debiasing, safety control, and enhanced caption generation that highlight different aspects of an image. By releasing our code, we hope our framework will benefit the community and encourage research towards better understanding of MLLMs, as well as their broader applications in domains such as physical and digital agents [50, 60]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.814, + 0.682, + 0.831 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "This work has been partially supported by ANR grant VISA DEEP (ANR-20-CHIA-0022), HPC resources of IDRIS under the file A0160614966 allocated by GENCI, and Cluster PostGenAI@Paris (ANR-23-IACL-0007, FRANCE 2030)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2213" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.198 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.485, + 0.256 + ], + "angle": 0, + "content": "[2] Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.485, + 0.411 + ], + "angle": 0, + "content": "[4] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539-1550, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.484, + 0.47 + ], + "angle": 0, + "content": "[5] Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor V. Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens. ArXiv, abs/2303.08112, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.484, + 0.528 + ], + "angle": 0, + "content": "[6] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.485, + 0.613 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[8] Shi Chen and Qi Zhao. Rex: Reasoning-aware and grounded explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15586-15595, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.485, + 0.741 + ], + "angle": 0, + "content": "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[11] Yi Ding, Bolian Li, and Ruqi Zhang. Eta: Evaluating then aligning safety of vision language models at inference time. ArXiv, abs/2410.06625, 2024. 8, 10, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.119 + ], + "angle": 0, + "content": "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[13] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[14] Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are linear. arXiv preprint arXiv:2405.14860, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.907, + 0.26 + ], + "angle": 0, + "content": "[15] Brandon Huang et al. Multimodal task vectors enable many-shot multimodal in-context learning. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[16] Yingzhe et al. LIVE: Learnable in-context vector for visual question answering. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[17] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[18] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[19] Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, et al. Multimodal autoregressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[20] Jiaxin Ge, Sanjay Subramanian, Trevor Darrell, and Boyi Li. From wrong to right: A recursive approach towards vision-language explanation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1173–1185, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.907, + 0.65 + ], + "angle": 0, + "content": "[21] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems (NeurIPS), pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.907, + 0.706 + ], + "angle": 0, + "content": "[22] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. ArXiv, abs/2311.05608, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[23] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via imageto-text transformation. In European Conference on Computer Vision, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[25] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2214" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Kaichen Huang, Jiahao Huo, Yibo Yan, Kun Wang, Yutao Yue, and Xuming Hu. Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models. arXiv preprint arXiv:2410.04819, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[27] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[29] Yuchu Jiang, Jiale Fu, Chenduo Hao, Xinting Hu, Yingzhe Peng, Xin Geng, and Xu Yang. Mimic in-context learning for multimodal tasks. arXiv preprint arXiv:2504.08851, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123(1):32-73, 2017. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[32] Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet. Decoderlens: Layerwise interpretation of encoder-decoder transformers. ArXiv, abs/2310.03686, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.629 + ], + "angle": 0, + "content": "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[35] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[36] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.885 + ], + "angle": 0, + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8, 1, 3, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[40] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, 2023. 8, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.342 + ], + "angle": 0, + "content": "[41] Oscar Manas, Pau Rodriguez Lopez, Saba Ahmadi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2523-2548, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[42] Neel Nanda. Actually, othello-gpt has a linear emergent world model, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[43] Nostalgebraist. Interpreting gpt: The logit lens. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens, 2020. Accessed: [date of access]. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.831, + 0.44 + ], + "angle": 0, + "content": "[44] OpenAI. Gpt-4 technical report. arXiv, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[45] Haowen Pan, Yixin Cao, Xiaozhi Wang, and Xun Yang. Finding and editing multi-modal neurons in pre-trained transformer. arXiv preprint arXiv:2311.07470, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[46] Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.907, + 0.606 + ], + "angle": 0, + "content": "[47] Jayneel Parekh, Pegah Khayatan, Mustafa Shukor, Alasdair Newson, and Matthieu Cord. A concept-based explainability framework for large multimodal models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[48] Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[49] Libo Qin, Qiguang Chen, Hao Fei, Zhi Chen, Min Li, and Wanxiang Che. What factors affect multi-modal in-context learning? an in-depth exploration. arXiv preprint arXiv:2410.20482, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[50] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[52] Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2215" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. In Advances in Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.189 + ], + "angle": 0, + "content": "[53] Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan D Cotterell. Linear adversarial concept erasure. In International Conference on Machine Learning, pages 18400-18421. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[54] Mansi Sakarvadia, Arham Khan, Aswathy Ajith, Daniel Grzenda, Nathaniel Hudson, André Bauer, Kyle Chard, and Ian T. Foster. Attention lens: A tool for mechanistically interpreting the attention head information retrieval mechanism. CoRR, abs/2310.16270, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[55] Sarah Schwettmann, Neil Chowdhury, Samuel Klein, David Bau, and Antonio Torralba. Multimodal neurons in pretrained text-only transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2862-2867, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[56] Mustafa Shukor and Matthieu Cord. Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.425 + ], + "angle": 0, + "content": "[57] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.484, + 0.482 + ], + "angle": 0, + "content": "[58] Mustafa Shukor, Corentin Dancette, and Matthieu Cord. ep-alm: Efficient perceptual augmentation of language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22056-22069, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[59] Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[60] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[61] Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, and Alaaeldin El-Nouby. Scaling laws for native multimodal models. arXiv preprint arXiv:2504.07951, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[62] Nishant Subramani, Nivedita Suresh, and Matthew E Peters. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566-581, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[63] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[64] Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.134 + ], + "angle": 0, + "content": "Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[66] Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.908, + 0.247 + ], + "angle": 0, + "content": "[67] Théophane Vallaeys, Mustafa Shukor, Matthieu Cord, and Jakob Verbeek. Improved baselines for data-efficient perceptual augmentation of llms. arXiv preprint arXiv:2403.13499, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.316 + ], + "angle": 0, + "content": "[68] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, 2024. 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.908, + 0.374 + ], + "angle": 0, + "content": "[69] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. arXiv preprint arXiv:2404.03592, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.908, + 0.416 + ], + "angle": 0, + "content": "[70] Dizhan Xue, Shengsheng Qian, and Changsheng Xu. Few-shot multimodal explanation for visual question answering. In ACM Multimedia 2024, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.908, + 0.473 + ], + "angle": 0, + "content": "[71] Chih-Kuan Yeh, Been Kim, Sercan O Arik, Chun-Liang Li, Pradeep Ravikumar, and Tomas Pfister. On concept-based explanations in deep neural networks. arXiv preprint arXiv:1910.07969, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.908, + 0.531 + ], + "angle": 0, + "content": "[72] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.908, + 0.601 + ], + "angle": 0, + "content": "[73] Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. arXiv preprint arXiv:2406.06579, 2024.1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.601 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2216" + } + ] +] \ No newline at end of file diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_origin.pdf b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..74668747d3bf893e55c866279f40b5e6e41d04b8 --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/3a9fad37-0f80-4ba9-981c-b874e41fd8d8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d77bafc8035526a0d73e64e1bf1b4481ddda2a2f7024863ba618a348a6c156d2 +size 1455593 diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/full.md b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5cfc9fd0dd1f1ef618e2039bb1c14b8267fcbae8 --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/full.md @@ -0,0 +1,359 @@ +# Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering + +Pegah Khayatan $^{\star 1}$ Mustafa Shukor $^{\star 1}$ Jayneel Parekh $^{\star 1}$ Arnaud Dapogny $^{1}$ Matthieu Cord $^{1,2}$ + +1ISIR, Sorbonne Université, Paris, France 2Valeo.ai, Paris, France + +# Abstract + +Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.1 + +# 1. Introduction + +With the rapid progress in Large Language Models (LLMs) [7, 10, 28, 44, 65], Multimodal LLMs (MLLMs) [3, 12, 34, 39, 61] have recently demonstrated remarkable capabilities in addressing complex multimodal tasks such as image captioning and visual question-answering. + +MLLMs are typically composed of a visual encoder, an LLM, and a connector. Following initial unimodal pretraining—and, in many cases, multimodal pretraining on large datasets—these models can be further specialized by training on multimodal datasets [1, 33]. Given the high compu + +tational cost of training these models, recent research has proposed more efficient approaches, such as creating diverse, high-quality instruction-tuning datasets [39] or keeping the LLM frozen and fine-tuning small amounts of parameters, like the connector [41, 57, 58, 67]. These approaches take advantage of the ability of frozen LLMs to generalize to multimodal data [56]. Despite the different efficient tuning methods, training these models still incurs significant costs. + +While substantial progress has been made in developing high-performing MLLMs, relatively few studies aim to understand them [4, 47, 55, 56, 59, 61, 73]. Existing work typically conducts post-hoc analyses of MLLMs in isolation, overlooking the internal changes due to fine-tuning. Research by [56] addresses this gap to some extent by examining the internal multimodal alignment as it evolves during training. + +In this work, we apply concept-level analysis to provide a readable understanding of MLLM behavior and, in particular, semantic dynamics that may occur due to fine-tuning, or due to covariate shift when considering different datasets. + +In the first case, we find that fine-tuning on a specific task potentially reshapes learned latent concepts, with some adjusting subtly to align with the task, and others emerging or disappearing altogether (see Fig. 1). Notably, we find that most fine-tuned concepts can be reconstructed from the original model by translating its original concepts in the direction of specific concept shift vectors, reducing the need for additional training and its associated costs. Furthermore, we explore the implications of the proposed analysis for MLLMs steering, demonstrating how model outputs can be modified inexpensively without additional training. Our key findings are summarized as follows: + +- We apply latent concept-level analysis to provide readable understanding on MLLMs' behavior ; in particular, we show that fine-tuning can introduce significant alteration in the original concepts. +- We show that we can control the MLLM's behavior w.r.t. certain concepts by simply manipulating shift vectors. +- Lastly, our findings also have direct applications for steering MLLM outputs, which find use for model debiasing as well as safety control. + +In a nutshell, we propose a novel, ready-to-use framework (including code) for MLLM behavior interpretability and control, debiasing and steering, which, we believe, will pave the way for future research. + +# 2. Related Work + +Concept-based explainability. Concept-based explainability methods have emerged as an alternative to traditional feature attribution based methods, that are capable of extracting key semantic features from the model internal representations. Most post-hoc concept-based approaches are based on the idea of concept activation vectors (CAV) [30], which represent concepts as vectors in the activation space. Instead or relying on human annotations, recent works have proposed methods to automatically discover concepts via clustering [21, 71] or matrix decomposition [18], which can be viewed as instances of a dictionary learning problem [17]. Initially focusing on understanding vision models, dictionary learning for concept extraction has been extended to LLMs e.g. using sparse autoencoders [27, 52]. However, none of the prior approaches have been applied to understand MLLMs, with the exception of recently proposed CoX-LMM [47]. + +MLLMs and Explainability. Multimodal LLMs [3, 34, 39, 67] have recently garnered significant interest. They typically adopt a late fusion architecture, and consist of an image encoder [19, 51, 72], a connector, and an LLM [28, 63, 65]. This family of models has inspired extensive research to better understand them and explain their behavior. For example, studies like [26, 45, 55] seek to identify multimodal neurons within LLMs or analyze modality-specific sub networks [56]. Some methods leverage the fact that these models are text-generative to simply generate textual explanations for model outputs [8, 20, 59, 70]. MLLMs benefit from in-context learning capabilities, which have been examined for limitations, including biases [4] and links to hallucinations [59], as well as the factors that may enhance their in-context learning performance [9, 49]. Related to our approach, CoX-LMMs [47] employs dictionary learning to extract multimodal semantic concepts from model representations. However, these studies typically assess models only in their final trained states, overlooking the dynamic changes that occur during training. Only limited works, such as [56], have investigated explaining changes due to fine-tuning, focusing specifically on implicit alignment between image and text modalities. In this work, we investigate how multimodal concepts within the model evolve throughout fine-tuning and explore the implications of these shifts on model steering. + +Steering models with feature editing. In contrast to editing model weights, representation or feature editing methods [62, 66, 69] aim to modify model outputs without altering the model's weights. A prominent approach within this family involves identifying steering vectors, or directions in the feature space (often within the residual stream), that are + +linked to contrasting concepts. These methods have been applied to language models for various purposes, such as enhancing factuality or reducing hallucinations [46], inducing sentiment shifts or detoxification [64, 66], improving refusals to harmful requests [2], promoting truthfulness by modifying the output of attention heads [35], and erasing specific concepts or biases [6, 53]. However, their application to MLLMs is yet to be explored. Another set of approaches related to steering methods are based on In-context learning (ICL) [15, 16, 29], where prompts are carefully designed to induce desired behavior. Yet, ICL requires predefined demonstrations and lacks interpretability at a concept level. In contrast, our method offers lightweight steering, extending such capabilities to MLLMs without requiring any training, for instance as in ReFT [69]. + +# 3. Methodology Overview + +Our framework is summarized in Fig. 1. We apply concept-level analysis of MLLM latent space (Section 3.1) to define and compare concepts between different setups. This allows us to monitor conceptual changes occurring during fin-tuning and manipulate MLLM behavior at a conceptual level (Section 3.2). Lastly, our framework finds applications for MLLM steering for e.g. debiasing and safety control (3.3) with negligible computational burden. + +MLLM setup. A generic MLLM consists of a visual encoder $f_{V}$ , a trainable connector $C$ , and a language model $f_{\mathrm{LM}}$ . We assume that the model is pretrained on a multimodal (e.g. captioning) dataset $S = \{(x_i, y_i)\}_i$ , where $x_i \in \mathcal{X}$ represents images and $y_i \subset \mathcal{Y}$ are the associated captions specified as sequence of tokens from token vocabulary space $\mathcal{Y}$ . The model is trained to generate the next text tokens, conditioned on text and images. The input to $f_{\mathrm{LM}}$ is a sequence of tokens that includes the concatenation of: (1) $N_V$ visual tokens extracted from the image $x$ via the visual encoder and connector $(C(f_V(x)))$ , and (2) linearly embedded textual tokens corresponding to the text instruction and previously predicted tokens. This can be expressed as: + +$$ +\hat {y} ^ {p} = f _ {L M} (h ^ {1}, \dots , h ^ {N _ {V}}, \dots , h ^ {p}), +$$ + +where $h^1, \ldots, h^{N_V} = C(f_V(x))$ , and $h^p = \operatorname{Emb}(\hat{y}^{p-1})$ with Emb representing the token embedding layer. During generation, the output token $\hat{y}^p$ is derived by normalizing the last layer $(L)$ tokens $h_{(L)}^p$ , then applying the unembedding layer $W_U$ and a softmax operation. The model keeps predicting the next token until the end of the sentence token to obtain the generated response $\hat{y} = \{\hat{y}^p\}_{p > N_V + N_I}$ , where $N_I$ is corresponds to the text instruction. + +# 3.1. Retrieval and comparison of latent concepts + +To understand the internal representations of any given MLLM $f$ , we leverage the approach introduced in [47]. + +![](images/054278d75a0790e4190ebed6be90a0d14c9a03ae03ee7c970853d97f5e1cefeb.jpg) +Figure 1. Framework overview. We apply concept-level analysis for MLLM behavior monitor and control, for (left) understanding and manipulating (through shift vectors) concept changes due to fine-tuning, as well as (right) MLLM steering for debiasing or safety control. + +![](images/e72ac1c9d73f1d12b0d1dffe68512f63b1401a077fe8832fe65d6ceb229070c6.jpg) + +![](images/4941f3fec2ee19335fb9bef24aaffe6f277b15a2a2b1d2676fc371e95cf25dd8.jpg) + +Specifically, given a set of $M$ images $\{x_{1},\dots,x_{M}\}$ we extract a set of residual stream representations from some layer $l$ of the MLLM $f$ . These representations $\mathbf{z}_m = f_l(x_m)\in \mathbb{R}^D$ (one per image) are collected in a feature matrix $\pmb{Z}\in \mathbb{R}^{D\times M}$ . Typically, the set of images and extracted representations correspond to a particular token of interest TOI (e.g., 'Dog', 'Cat', 'Person', etc.) in the predicted caption. However, the extraction can also be performed for a larger set of target tokens. This feature matrix $\pmb{Z}$ is then decomposed as $\pmb{Z}\approx \pmb{U}\pmb{V}$ to recover the concepts in the latent embedding space. Here, $\pmb{U}\in \mathbb{R}^{D\times K}$ is the matrix of $K$ concepts and $\pmb{V}\in \mathbb{R}^{K\times M}$ represents the coefficients/activations of the samples projected onto these concepts. Different decompositions of the matrix $K$ result in various concepts, inheriting the properties of the decomposition (such as low grounding overlap with PCA). We employ $K$ -Means to learn our concept dictionaries. This is motivated by $K$ -Means' simplicity, and straightforward arithmetic manipulation of the clusters/concepts it allows. Each column $\pmb{u}_k\in \pmb{U}$ corresponds to a concept, while each column of $\pmb{V}$ encodes the activation of these concepts for a given sample. Note that any given representation $f_{l}(x)$ can be projected on $\pmb{U}$ to obtain its activation vector $\pmb{v}(x)\in \mathbb{R}^{K}$ , i.e. $f_{l}(x)\approx U\pmb{v}(x)$ . Each extracted concept is then interpreted through grounding in both image and text spaces. Specifically, the top $N_{\mathrm{MAS}}$ that activates concept $\pmb{u}_k$ the most represent its image grounding: + +$$ +\boldsymbol {X} _ {\text {M A S}} \left(\boldsymbol {u} _ {k}\right) = \underset {\hat {X} \subset \mathbf {X} _ {t}, | \hat {X} | = N _ {\text {M A S}}} {\arg \max } \sum_ {x \in \hat {X}} \left| \boldsymbol {v} _ {k} (x) \right|, \tag {1} +$$ + +where $\pmb{v}_k(x)$ refers to the activation of $\pmb{u}_k$ for image $x$ . For text grounding, we decode the features using the unembedding matrix of the language model $W_U$ [5, 32, 43, 54]. Specifically, the operation $W_U\pmb{u}_k \in \mathbb{R}^{|\mathcal{Y}|}$ produces + +logits over the vocabulary, and the top $N_{\mathrm{grounding}}$ words with highest logits are extracted: + +$$ +\boldsymbol {T} _ {\text {w o r d s}} (\boldsymbol {u} _ {k}) = \underset {\text {T o p} - N _ {\text {g r o u n d i n g}}} {\arg \max } \left(W _ {U} \boldsymbol {u} _ {k}\right). \tag {2} +$$ + +Finally, to quantify the similarity of two concepts (i.e. the columns of $\mathbf{U}$ we define the Text Grounding Overlap as: + +$$ +\operatorname {T} \text {O v e r l a p} (\boldsymbol {u}, \boldsymbol {u} ^ {\prime}) = 1 0 0 \times \frac {\left| \boldsymbol {T} _ {\text {w o r d s}} (\boldsymbol {u}) \cap \boldsymbol {T} _ {\text {w o r d s}} (\boldsymbol {u} ^ {\prime}) \right|}{\left| \boldsymbol {T} _ {\text {w o r d s}} (\boldsymbol {u}) \right|}. \tag {3} +$$ + +Now that we have defined a generic framework, we present two subcases for extracting and manipulating concepts. + +# 3.2. Evolution of concepts through fine-tuning. + +Setup overview. An original model $f^a$ is typically finetuned to produce a specialized model $f^b$ for a particular task—or, specifically, for a set of target concepts. This finetuning can be conducted on samples that include a set of words $\{w_1, \dots, w_m\}$ associated with these target concepts. For instance, if we fine-tune a image captioning model to emphasize colors in the image, the set of words will simply be these colors. Efficient fine-tuning is typically achieved using Low-Rank Adaptation (LoRA) [25, 34, 39]. Finetuning can selectively alter certain representations, leading to shifts in the conceptual space encoded by the model. Using the interpretability framework discussed in Section 3.1 we can study these shifts at a readable conceptual level. + +Concept recovery via shift vectors. To study the change from an original model $f^a$ to a finetuned model $f^b$ , we fix the dataset $S^{(1)} = S^{(2)}$ , and obtain two sets of embeddings from $f^a, f^b$ respectively, i.e. $A \approx U^a V^a$ , $B \approx U^b V^b$ , where $U^a, U^b \in \mathbb{R}^{D \times K}$ are $K$ concepts extracted from each model. We propose to characterize the concept changes + +from an original to fine-tuned model as linear directions in embedding space or concept shift vectors. To do so, we first associate each original concept $\pmb{u}_k^a\in U^a$ with a subset of samples where $\pmb{u}_k^a$ is the most activated concept: + +$$ +\pmb {A} _ {k} = \{m \mid k = \underset {i} {\arg \max} | \pmb {v} _ {i} ^ {a} (x _ {m}) | \}. +$$ + +For each sample $x_{m}, m \in \mathbf{A}_{k}$ we define $\delta_{m}^{a \to b} = \pmb{b}_{m} - \pmb{a}_{m}$ as the change in its representation from $f^{a}$ to $f^{b}$ . To compute the concept shift vector $\pmb{\Delta}_{k}^{a \to b}(\pmb{u}_{k}^{a})$ associated with $\pmb{u}_{k}^{a}$ , we aggregate shifts of its associated samples specified by $\mathbf{A}_{k}$ : + +$$ +\pmb {\Delta} _ {k} ^ {a \rightarrow b} (\pmb {u} _ {k} ^ {a}) = \frac {1}{| \pmb {A} _ {k} |} \sum_ {m \in \pmb {A} _ {k}} \delta_ {m} ^ {a \rightarrow b} = \frac {1}{| \pmb {A} _ {k} |} \sum_ {m \in \pmb {A} _ {k}} (\pmb {b} _ {m} - \pmb {a} _ {m}) +$$ + +The concept shift vector is used to shift each concept in the original model $\boldsymbol{u}_k^a$ to obtain the shifted concept $\boldsymbol{u}_k^s$ : + +$$ +\boldsymbol {u} _ {k} ^ {s} = \boldsymbol {u} _ {k} ^ {a} + \alpha \Delta_ {k} ^ {a \rightarrow b} \left(\boldsymbol {u} _ {k} ^ {a}\right), \tag {4} +$$ + +where $\alpha$ is a coefficient to control the shift magnitude. Unless otherwise stated, we use $\alpha = 1$ as the default magnitude of shift. It is worth noting that given the concept shift vectors, the computation of shifted concepts does not rely on accessing the fine-tuned model. Practically speaking, this means that we can "push" the original model towards the concepts of a fine-tuned one with very little overhead, by simply shifting its latent representation in the direction of the shift vector, as it will be illustrated in the experiments. + +# 3.3. Concept evolution across datasets and applications to model steering + +Concept comparison between different datasets. Another interesting subcase of the proposed framework consists in evaluating the shift from one dataset $S^{(1)}$ to another $S^{(2)}$ , using the same model and encoding $h$ . Beyond interpretability of a model behavior, it also find applications for model steering. Model steering (see Fig. 1 (right)) refers to guiding the model outputs towards desired outcomes by modifying the features without altering the model weights. + +Coarse-grained model steering. In coarse-grained or global steering, the objective is to adjust the model outputs $\hat{y}$ to generally align with a set of target samples (e.g., changing answers type). Given input-output samples, we first extract the answer representations $B = b_{1},\dots,b_{N}$ at layer $l$ from the target set. Similarly, we obtain representations for the original set $A = a_{1},\dots,a_{M}$ (e.g. randomly drawn from the train set). We then compute the coarse steering vector $s_{c}$ as: + +$$ +\boldsymbol {s} _ {c} = \frac {\sum_ {i} ^ {N} \boldsymbol {b} _ {i}}{N} - \frac {\sum_ {i} ^ {M} \boldsymbol {a} _ {i}}{M}, \tag {5} +$$ + +$s_c$ is applied to all the samples in the validation set. For instance, the activations $f_{l}(x_{i})$ of a sample $x_{i}$ become: + +$$ +\tilde {f} _ {l} \left(x _ {i}\right) = f _ {l} \left(x _ {i}\right) + \alpha \boldsymbol {s} _ {c} \tag {6} +$$ + +![](images/90bd6d7f34acd49ea52a792b06f82d4fa94c74e52851369fd682616b18ae41ee.jpg) +Figure 2. Concepts extracted from original and fine-tuned models. concepts from the original $f^a$ (top), and model fine-tuned to focus more on places $f^b$ (bottom), for $TOI =$ person. The concepts from $f^b$ exhibit a stronger association with places. + +where $\alpha$ controls the steering strength and it is set to 1 (we study $\alpha$ in App. B.5.3). Thus, in this setup, all examples are coarsely steered in the direction of the steering vector $s_c$ at layer $l$ , before passing $\tilde{f}_l(x_i)$ through the rest of layers and $\tilde{f}_l(x_i)$ becomes the input to the next layer $l + 1$ . + +Fine-grained steering. Unlike global steering, fine-grained steering consists in finding and editing directions that adjust only certain concepts to other ones. To do this, we decompose the hidden states of a set of samples into a set of concepts $\mathbf{U}$ as previously explained. We then compute a set of fine-grained steering vectors $s^f = s_{11}^f, \ldots, s_{NN}^f$ , with $s_{ij}^f = s_{ij}^f = u_j - u_i$ the steering vector from concept $u_i$ to $u_j$ . However, not all steering vectors are meaningful: options for finding the relevant ones include proximity matching, as well as identifying vectors that have the strongest impact on guiding the model towards generating specific answers or concepts (e.g. producing significantly more target answers). This is more detailed in App. B.3. Various applications of MLLM steering can be explored, such as gender debiasing and safety alignment, as it will be shown below. + +# 4. Experiments + +# 4.1. Fine-tuning experiments + +In this section, we study how fine-tuning introduces changes in the overall structure of the learned concepts in MLLMs (with the architecture described in Section 3). + +Implementation details. The main paper covers experiments on the popular LLaVA [38] model comprising a CLIP image encoder, a two-layer MLP connector, and a 7B Vicuna-1.5 LLM. More experiments on a different multimodal model can be found in App. A. We conduct our study in a controlled setup that consists in specializing the model on a target dataset. Specifically, we apply fine-tuning on three different subsets of the Visual Genome dataset [31], related to places, colors, and sentiments (more details in App. A.2). Impact of fine-tuning on learned concepts. Fig. 2 depicts this concept change. Through the concepts groundings for an example TOI (person) we see that the fine-tuned model puts a stronger emphasis on places, which is expected and serves as a sanity check for our method. + +Matched concepts. To further analyze how each concept changes after fine-tuning, we focus on each concept and its + +![](images/c10d0f78db351f015b227b953972a226d4fcbfd61fd6ab3a813acf94dfa66fd8.jpg) +Figure 3. Concepts text grounding change after fine-tuning. Text grounding for concepts $(TOI = \text{bus})$ from $f^a$ and their match from $f^b$ , (fine-tuned to focus more on places). Emerging concepts may include grounding words not explicitly included in the fine-tuning vocabulary for place (e.g., "District", "Crossing"), while others evolve more smoothly (e.g., "Street"). + +match. Specifically, we define a matching function $m: i \to j^*$ which associates each concept vector $\pmb{u}_i^a$ in set $\pmb{U}^a$ to its closest vector $\pmb{u}_{j*}^b$ in set $\pmb{U}^b$ based on cosine similarity, i.e. $m(i) = \arg \max_{\pmb{u}_j^b \in \pmb{U}_b} \cos(\pmb{u}_i^a, \pmb{u}_j^b)$ . + +Fig. 3 shows the text groundings for various concepts, displaying the words with a frequency lower than 5 across concepts (e.g., filtering out high-frequency terms like bus, vehicle, etc.). We observe the emergence of place-related terms across the identified elements, while the overall thematic structure remains consistent. Note that certain concepts may converge toward the same fine-tuned concept. We also analyze how the distances between matched concepts evolve during fine-tuning in App. A.3. + +![](images/777d941f6424d01022b485dab279e939b753602f35e760f2a92a6a45835c844b.jpg) +Figure 4. Text grounding overlap (T-Overlap) between original and fine-tuned model concepts. Different concepts change to different extents depending on the fine-tuning. + +Concept evolution. To quantify how much a concept $\boldsymbol{u}_i^a \in \boldsymbol{U}^a$ is changed after fine-tuning, we compute the overlap between its grounding words and those of its closest matching concept from the fine-tuned model $\boldsymbol{U}^b$ . Specifically, we compute T-Overlap $(\boldsymbol{u}_i^a, \boldsymbol{u}_{m(i)}^b)$ (Eq. (3)) for all the concepts $i \in \{1, \dots, K\}$ , and visualize them for different fine-tunings in Fig. 4. We observe varying rates of change across different concepts and fine-tunings. This might be due to the difference in the fine-tuning dataset size, complexity, + +or similarity to the original dataset. It also highlights 2 main behaviors, detailed as follows: + +- Concepts that are refined. This group contains the concepts that slightly change to be more specialized towards the fine-tuning task (Fig. 5 top, middle rows). These concepts exhibit a relatively high (T-Overlap $(\pmb{u}_i^a,\pmb{u}_{m(i)}^b)$ ). +- Concepts that change completely. This group contains the concepts that emerge or, to a certain extent, disappear (Fig. 5 bottom row) in the fine-tuned model. New concepts emerge during fine-tuning likely due to the introduction of novel patterns or relationships. These concepts exhibit a relatively low (T-Overlap $(\pmb{u}_i^a,\pmb{u}_{m(i)}^b)$ ). + +![](images/614281c11141c2af45f53ba86312363b394d893afb497f4689914332056ce9de.jpg) +Figure 5. Concepts evolve differently due to fine-tuning. Left: concepts extracted from the original model $f^a$ . Right: matched concept extracted from $f^b$ , fine-tuned to focus on places. Each concept is grounded in image and text. We observe different levels of adaptation across concepts: some concepts specialize further by adding to place-related words (e.g., Top), some concepts introduce place-related words while staying aligned with the original concept in textual or visual groundings (e.g., Middle), while others undergo complete transformation, diverging significantly from their original meaning to fully embrace place-related elements (e.g., Bottom). + +We also notice that T-Overlap decreases with the number of training iterations, indicating that fine-tuning leads to deviation from the original concepts (more details in App. A.3). Evaluating fine-tuned concept recovery. To study if the fine-tuned concepts $\pmb{U}^{b}$ can be recovered from the original ones $\pmb{U}^{a}$ , we first establish a matching $(m : i \rightarrow j)$ between the set of original $\{\pmb{u}_{i}^{a}\}_{i=1}^{K}$ and fine-tuned concepts $\{\pmb{u}_{j}^{b}\}_{j=1}^{K}$ . For systematic evaluation of recovery of all fine-tuned concepts, we constrain $m$ to be bijective using an optimal transport algorithm detailed in App. A.1. Finally, we evaluate how well a shifted concept $\pmb{u}_{k}^{s}$ (Equ. (4)) is similar to its match $\pmb{u}_{m(k)}^{b}$ using the aforementioned T-Overlap metric. Fig. 6 shows the results of recovering the fine-tuned concepts for models fine-tuned on different subsets of the VG + +![](images/5492ad358c17c4139aeef04a9c3285898c141926b586baf66e5256296d256e22.jpg) +Figure 6. Recovering fine-tuned concepts. Across different finetunings (places (top), colors (middle) and sentiments (bottom)), we compute the average text grounding overlap of original and shifted concepts with the matched fine-tuned ones. Shifting the original concepts result in partially recovering the fine-tuned ones. + +dataset (place, color, sentiment). We report the T-Overlap between the shifted $u_{k}^{s}$ and fine-tuned concepts $u_{m(k)}^{b}$ for various different tokens of interest. For each target token and finetuning we extract $K = 20$ concepts and report the mean and standard deviation over them. We use the overlap between the original concepts $u_{k}^{a}$ and the fine-tuned ones as a baseline. We observe that most shifted concepts show higher overlap than the original ones: this demonstrates that fine-tuned concepts can be efficiently recovered from the original model with concept shift vectors. + +![](images/f5d587dc3cd6761cb01fdd6ba1f2740fc4c487e9ffe493c4d7b03527b89838d0.jpg) +Figure 7. Correlation between shift consistency and concept recovery (Color finetuning). The more consistent and aligned the individual representation shifts associated with a concept, the better the recovery of the fine-tuned concept. + +Which concepts are recovered better? We hypothesize that if the representation shift for individual samples associated with a concept $\boldsymbol{u}_k^a$ , $\{\delta_m^{a\rightarrow b}|m\in \boldsymbol{A}_k\}$ , is consistently aligned with the concept shift vector, the resulting $\Delta_k^{a\rightarrow b}(\boldsymbol{u}_k^a)$ , should be more effective at recovering the finetuned concept. We quantify the consistency through mean cosine similarity of $\{\delta_m^{a\rightarrow b}\}_{m\in \boldsymbol{A}_k}$ with the concept shift vector $\Delta_k^{a\rightarrow b}(\boldsymbol{u}_k^a)$ . In other words, this quantifies the alignment + +of individual shifts and their mean, $\pmb{\Delta}_k^{a\rightarrow b}(\pmb{u}_k^a)$ : + +$$ +\operatorname {C o n s i s t e n c y} \left(\boldsymbol {u} _ {k} ^ {a}\right) = \frac {1}{\left| \boldsymbol {A} _ {k} \right|} \sum_ {m \in \boldsymbol {A} _ {k}} \cos \left(\delta_ {m}, \boldsymbol {\Delta} _ {k} ^ {a \rightarrow b} \left(\boldsymbol {u} _ {k} ^ {a}\right)\right), +$$ + +We measure the recovery of concept $k$ , $\mathrm{CR}_k$ , as the improvement in similarity between the matched fine-tuned $\boldsymbol{u}_{m(k)}^b$ and shifted concept $\boldsymbol{u}_k^s$ , relative to the original one $\boldsymbol{u}_k^a$ : + +$$ +\mathrm {C R} _ {k} = \frac {\cos \left(\boldsymbol {u} _ {m (k)} ^ {b} , \boldsymbol {u} _ {k} ^ {s}\right) - \cos \left(\boldsymbol {u} _ {m (k)} ^ {b} , \boldsymbol {u} _ {k} ^ {a}\right)}{\cos \left(\boldsymbol {u} _ {m (k)} ^ {b} , \boldsymbol {u} _ {k} ^ {a}\right)} \tag {7} +$$ + +We plot the consistency and recovery for concepts extracted across four tokens of interest for color finetuning in Fig. 7. Plots for other finetunings are in App. A.5. Crucially, we observe a positive and statistically significant correlation between the two quantities across all the finetuning tasks. This supports our hypothesis that a better concept shift recovery is related to consistency of individual shifts of original concept. More ablation studies about concept recovery are available in App. A.4, analyzing the influence of steering strength $\alpha$ , number of concepts $K$ , and extraction layer $l$ . + +![](images/299dd8fe0458bcc4305331a96826d04fa3c8832368af4528aaa6862f7df71b8f.jpg) +Figure 8. Linear separability of concepts features in MLLMs. We visualize the features related to the concepts "yes" and "no" after PCA projections across MLLMs layers. + +In summary, we demonstrated the feasibility of recovering target fine-tuned concepts by applying simple per-concept shift of the original model features. This supposes that the features related to different concepts are almost linearly separable, which empirically seems to hold at least for the last MLLM layers as pictured on Fig. 8, and as previously studied for LLMs in [42, 48]. This motivates the following investigations on using a similar methodology for simple, computationally inexpensive, yet efficient MLLM steering. + +# 4.2. Multimodal model steering + +We perform steering by applying a steering vector $\pmb{v}$ to the residual stream features $Z$ of the MLLM, without changing its parameters. We first evaluate the MLLM steering capabilities in a visual question-answering (VQA) setup. Then, we show the applicability of our MLLM steering to control captioning styles. Lastly, we present two steering applications: gender debiasing, aiming to mitigate biases in model outputs, and safety alignment, i.e. ensuring that the model refuses to generate harmful information. We discuss technical details for each of these applications in App. C and App. D. + +Setup. For VQA tasks, each query consists of a question about an input image, and the model generates an answer. + +To measure the effectiveness of our approach in directing the model towards specific answers or answer types, we report the number of generated answers that align with the target output or answer type. Additionally, we aim for targeted steering, ensuring that only specific answer types are influenced. For example, when altering answers from "yes" to "no" within the "yes/no" category, responses to other question types should remain unaffected. This specificity is assessed by tracking accuracy across answer types and number of answers from each type. + +Implementation details. Experiments in the main paper are primarily conducted on LLaVA [39] for conciseness. However, we show in App. B.2 that our method is generally applicable to other popular MLLMs. We experiment on VQAv2 [24], a visual question-answering corpus with image-question-answer triplets and annotated answer types ("yes/no", "number", and "other"). Steering vectors are derived from a subset of the train set, with model performance evaluated on the val set. As steering becomes more effective in deeper layers (see Fig. 8 and App. B), we apply it on the last layer. Additional experiments can be found in App. B. + +
Target TypeAnswers Type
yes/nonumberother
N/A366122494
yes/no55796288
number327201390
other364115501
+ +Coarse and Fine-Grained Model Steering for VQA. We explore both coarse and fine-grained steering in VQA tasks. Specifically, coarse steering aims to alter the distribution of answers, while fine-grained steering targets specific responses. For coarse-grained steering, we direct the model's answers toward a particular category: yes/no, numbers, or other (e.g., colors, objects). We compute a steering vector for each target answer type. As shown in Table 1, applying steering significantly increases the proportion of the targeted answer type. For fine-grained steering, we first assess the feasibility of identifying such steering vectors. Fig. 9 illustrates examples of these vectors. We derive them from three sets of sample answers corresponding to "yes/no," "number," and "other" categories. Interestingly, some vectors distinctly align with specific answers such as "No," "4," and "Red." This confirms the potential to identify fine-grained steering vectors capable of guiding the model toward a precise response. Building on these insights, we seek to steer the model toward a user-specified answer. For each original/target answer pair (e.g., yes/no), we collect some samples and compute the corresponding (coarse) steering vector. We then apply these vectors to all validation set samples. In + +![](images/2437368440318cfa7421b8fec918c07aac795a916de04c757b4498e1bc11f4cb.jpg) + +![](images/2eafbad9d9f38f450f1d7c90b8f83717326881a7fdb394bfa8e4ccbdb06057aa.jpg) + +![](images/5dbd2cd46908a19efc09ce597283e096720d1d0d1e12a8a1519e538d1561d14e.jpg) + +![](images/adc0117383fb2ab858d38ae414b29cb2ba181e696fb506c80b32ce17455b2b1c.jpg) + +![](images/c5394683e576cc3287e39c9a7c1bb06bfb5f74aba6e291ac6fd7b7eb3a88d544.jpg) + +![](images/a786d751a8b150d9408fb1a8d1ae2452ecea1c8174587e99a4b929145e8ddde7.jpg) + +![](images/7a7c6ec020e1e00fabaded1ca98e6ed98e0dc2e2ae63193e1318d1e6456cfcc2.jpg) +Figure 9. Discovering meaningful steering directions. Each line corresponds to a fine-grained steering direction to steer the model answer to: "No" (yes/no), "4" (number) and "Red" (other). Some steering directions are targeted (e.g., "No") as there is slight change in both the accuracy and number of answers types on other types (e.g., number, other). We show the relative scores compared to a baseline with no steering. + +![](images/0c54b6c6263624626db32666059d183fa4bef536eeaa6b916092b22e6e261687.jpg) + +![](images/3b567c38f1d4f3240a4c9354fd1b32de7d421a883f844d9321f123281dfdb812.jpg) + +Table 2, we report evaluation metrics when steering at the last layer. The results show that steering effectively increases the occurrence of target answers, while accuracy on other answer types remains largely stable. + +Table 1. Steering MLLMs answers type. Number of target answers type increases significantly after model steering. + +
SteeringAccuracy (%)Answer TypesAnswers
Yes/NoNumberOtherYes/NoNumberOtherOriginalTarget
N/A90.8258.4771.101861687234900
Yes → No69.0356.8268.9918846952294-828+828
1 → 390.7154.5271.1218616702350-215+144
White → Black90.4058.4258.3618616712312-98+441
+ +Table 2. Steering MLLMs answers. Steering answers from "Yes" (Yes/No), "1" (Number), "White" (Other) to "No", "3", "Black" respectively. The number of original/target answer counts decreases/increases significantly, while the accuracy on other answer types changes only slightly, and the number of answer type counts remains almost constant. + +![](images/32e4ee70a330dce9139f5043b03d94a5a748ed60e6da9568c95fe151cf32fd4d.jpg) +Figure 10. Steering MLLMs captions style. Captions steered to focus more on colors (left), places (middle) and sentiments (right). + +![](images/e6eed37fede6ea1430cd86ae9cf2c93fcd37c3b4edb13ea3fe9b3a8f5b564b7c.jpg) + +![](images/4166487faa950fc72e254be5e84cb072a94f113261e82c27d925cf772711959d.jpg) + +Steering image caption styles. We previously applied steering on relatively brief answers from the VQAv2 dataset. Here, we extend this approach to longer, descriptive outputs + +using the COCO captioning dataset [37]. Given that multiple captions can effectively describe an image by emphasizing various aspects such as the main object, surroundings, actions, or events, we aim at modifying captions to align with a specific target style. Here, we learn a coarse steering vector between samples with predicted captions in the target style and random samples. Qualitative examples of this steering for LLaVA are in Fig. 10. Results in Table 3 demonstrate that captions can be effectively steered towards a target style even when considering tasks with longer responses. We provide more captioning experiments in App. B.4. + +
Target StyleCaptionsStyle
placescolorssentiments
N/A43013092
places79610771
colors48825611
sentiments393104048
+ +Gender debiasing captions. We perform gender debiasing on COCO test set, aiming at mitigating biases for any gendered nouns when captioning. We experiment with both coarse and fine-grained steering. The coarse steering vector is computed between sets of samples with a gendered/neutral noun in the caption. For fine-grained steering, we steer each concept to its closest (as defined by its posing similarity) counterpart among the neutral concepts, as explained in Section 3.3. The results are reported in Table 4. Interestingly, both strategies are capable of converting many gendered captions to neutral ones, with fine-grained steering being significantly more effective than coarse steering. + +Safety alignment. Pure text LLMs often exhibit stronger safety alignment compared to MLLMs [11]. Empirical evidence for this can be found in App. D. Using this insight, we construct two sets of samples, categorized as safe and unsafe, by evaluating the model's response to identical malicious content presented in different modalities: one conveyed through text and the other through text + image. We use these to compute a safety guard steering vector. This steering vector gives the model a higher level of safety, without affecting its usefulness for safe tasks (Table 5). We evaluate the safety of the model by ASR (attack success rate) metric, described in detail in App. D. + +# 5. Discussion + +Limitations. The effectiveness of our method relies on the fact that the concepts are represented by linear directions in latent space. However, recent work [14] has found that not all features are captured as such: thus, when analyzing recovery + +Table 3. Steering MLLMs captions style. Each line corresponds to a different steering vector. Steering towards a target style increases the number of captions with that style. + +
ModelTotalMethodGendered → Neutral
LLaVA-1.5794coarse232
fine-grained632
Idefics2815coarse237
fine-grained315
Qwen2-VL-Instruct926coarse134
fine-grained300
Before SteeringAfter Steering
A young boy with curly hair is playing a video game.A child with curly hair is playing a video game.
A man riding a dirt bike on a beach.A person riding a dirt bike on a beach.
+ +Table 4. Gender debiasing results: number of occurrences of gendered terms converted to neutral terms across different models and methods, after steering with $\alpha = 1$ . Below, qualitative samples illustrate changes in descriptions before and after applying steering. + +
ModelBefore steeringAfter steering
Qwen2-VL-Instruct45/1005/100
+ +Table 5. Enhancing MLLM Safety Through Steering. We evaluate safety using the ASR metric, quantifying the proportion of responses that do notrefuse to provide harmful instructions. Our assessment is conducted on a portion of MM-SafetyBench [40] dataset. A lower ASR is desired when the prompt requires harmful instructions. + +of fine-tuned concepts, it can be interesting to explore more sophisticated similarity measures and matching algorithms. + +Conclusion. In this work, we introduced a novel concept-based analysis framework for monitoring and controlling MLLM behavior, offering new insights into how latent representations evolve during fine-tuning and across datasets. To address the former, we proposed concept shift vectors, an efficient method for recovering and interpreting concepts in fine-tuned models relative to their original counterparts. This approach naturally led us to explore the latter case, where we demonstrated the ability to steer model behavior by modifying features – without requiring additional training. Our results show that this technique effectively modifies MLLM answers, enabling applications such as gender debiasing, safety control, and enhanced caption generation that highlight different aspects of an image. By releasing our code, we hope our framework will benefit the community and encourage research towards better understanding of MLLMs, as well as their broader applications in domains such as physical and digital agents [50, 60]. + +# Acknowledgements + +This work has been partially supported by ANR grant VISA DEEP (ANR-20-CHIA-0022), HPC resources of IDRIS under the file A0160614966 allocated by GENCI, and Cluster PostGenAI@Paris (ANR-23-IACL-0007, FRANCE 2030). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1 +[2] Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717, 2024. 2 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 2 +[4] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539-1550, 2024. 1, 2 +[5] Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor V. Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens. ArXiv, abs/2303.08112, 2023. 3 +[6] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36, 2024. 2 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1 +[8] Shi Chen and Qi Zhao. Rex: Reasoning-aware and grounded explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15586-15595, 2022. 2 +[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2 +[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1 +[11] Yi Ding, Bolian Li, and Ruqi Zhang. Eta: Evaluating then aligning safety of vision language models at inference time. ArXiv, abs/2410.06625, 2024. 8, 10, 11 +[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An + +embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1 +[13] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1, 2021. 1 +[14] Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are linear. arXiv preprint arXiv:2405.14860, 2024. 8 +[15] Brandon Huang et al. Multimodal task vectors enable many-shot multimodal in-context learning. In NeurIPS, 2024. 2 +[16] Yingzhe et al. LIVE: Learnable in-context vector for visual question answering. In NeurIPS, 2024. 2 +[17] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2023. 2 +[18] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 2 +[19] Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, et al. Multimodal autoregressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402, 2024. 2 +[20] Jiaxin Ge, Sanjay Subramanian, Trevor Darrell, and Boyi Li. From wrong to right: A recursive approach towards vision-language explanation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1173–1185, 2023. 2 +[21] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems (NeurIPS), pages 9277-9286, 2019. 2 +[22] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. ArXiv, abs/2311.05608, 2023. 11 +[23] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via imageto-text transformation. In European Conference on Computer Vision, 2024. 11 +[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7, 3 +[25] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3, 1 + +[26] Kaichen Huang, Jiahao Huo, Yibo Yan, Kun Wang, Yutao Yue, and Xuming Hu. Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models. arXiv preprint arXiv:2410.04819, 2024. 2 +[27] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 2 +[29] Yuchu Jiang, Jiale Fu, Chenduo Hao, Xinting Hu, Yingzhe Peng, Xin Geng, and Xu Yang. Mimic in-context learning for multimodal tasks. arXiv preprint arXiv:2504.08851, 2025. 2 +[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2 +[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123(1):32-73, 2017. 4, 1 +[32] Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet. Decoderlens: Layerwise interpretation of encoder-decoder transformers. ArXiv, abs/2310.03686, 2023. 3 +[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36, 2024. 1 +[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3 +[35] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 2 +[36] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 10 +[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8, 1, 3, 9 +[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. + +Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4 +[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7 +[40] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, 2023. 8, 10 +[41] Oscar Manas, Pau Rodriguez Lopez, Saba Ahmadi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2523-2548, 2023. 1 +[42] Neel Nanda. Actually, othello-gpt has a linear emergent world model, 2023. 6, 8 +[43] Nostalgebraist. Interpreting gpt: The logit lens. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens, 2020. Accessed: [date of access]. 3 +[44] OpenAI. Gpt-4 technical report. arXiv, 2023. 1 +[45] Haowen Pan, Yixin Cao, Xiaozhi Wang, and Xun Yang. Finding and editing multi-modal neurons in pre-trained transformer. arXiv preprint arXiv:2311.07470, 2023. 2 +[46] Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023. 2 +[47] Jayneel Parekh, Pegah Khayatan, Mustafa Shukor, Alasdair Newson, and Matthieu Cord. A concept-based explainability framework for large multimodal models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2 +[48] Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 6, 8 +[49] Libo Qin, Qiguang Chen, Hao Fei, Zhi Chen, Min Li, and Wanxiang Che. What factors affect multi-modal in-context learning? an in-depth exploration. arXiv preprint arXiv:2410.20482, 2024. 2 +[50] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025.8 +[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[52] Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah, + +and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. In Advances in Neural Information Processing Systems, 2024. 2 +[53] Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan D Cotterell. Linear adversarial concept erasure. In International Conference on Machine Learning, pages 18400-18421. PMLR, 2022. 2 +[54] Mansi Sakarvadia, Arham Khan, Aswathy Ajith, Daniel Grzenda, Nathaniel Hudson, André Bauer, Kyle Chard, and Ian T. Foster. Attention lens: A tool for mechanistically interpreting the attention head information retrieval mechanism. CoRR, abs/2310.16270, 2023. 3 +[55] Sarah Schwettmann, Neil Chowdhury, Samuel Klein, David Bau, and Antonio Torralba. Multimodal neurons in pretrained text-only transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2862-2867, 2023. 1, 2 +[56] Mustafa Shukor and Matthieu Cord. Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2 +[57] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 1 +[58] Mustafa Shukor, Corentin Dancette, and Matthieu Cord. ep-alm: Efficient perceptual augmentation of language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22056-22069, 2023. 1 +[59] Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2 +[60] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. 8 +[61] Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, and Alaaeldin El-Nouby. Scaling laws for native multimodal models. arXiv preprint arXiv:2504.07951, 2025. 1 +[62] Nishant Subramani, Nivedita Suresh, and Matthew E Peters. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566-581, 2022. 2 +[63] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 2 +[64] Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023. 2 +[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya + +Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1, 2 +[66] Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023. 2 +[67] Théophane Vallaeys, Mustafa Shukor, Matthieu Cord, and Jakob Verbeek. Improved baselines for data-efficient perceptual augmentation of llms. arXiv preprint arXiv:2403.13499, 2024. 1, 2 +[68] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, 2024. 10 +[69] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. arXiv preprint arXiv:2404.03592, 2024. 2 +[70] Dizhan Xue, Shengsheng Qian, and Changsheng Xu. Few-shot multimodal explanation for visual question answering. In ACM Multimedia 2024, 2024. 2 +[71] Chih-Kuan Yeh, Been Kim, Sercan O Arik, Chun-Liang Li, Pradeep Ravikumar, and Tomas Pfister. On concept-based explanations in deep neural networks. arXiv preprint arXiv:1910.07969, 2019. 2 +[72] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2 +[73] Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. arXiv preprint arXiv:2406.06579, 2024.1 \ No newline at end of file diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/images.zip b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25107a59c9ce45eac87a221887e9427d2a16df59 --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d3b2fce8c11ef65417c53b8c6a61514a962045ec1df4e640e168e99728839c68 +size 554263 diff --git a/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/layout.json b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dc4bdea7d1067485a04bf1efe1bbdfa4bb1a62af --- /dev/null +++ b/2025/Analyzing Finetuning Representation Shift for Multimodal LLMs Steering/layout.json @@ -0,0 +1,11037 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "content": "Pegah Khayatan" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "content": " Mustafa Shukor" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "content": " Jayneel Parekh" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "inline_equation", + "content": "^{\\star 1}" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "content": " Arnaud Dapogny" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "text", + "content": " Matthieu Cord" + }, + { + "bbox": [ + 61, + 142, + 548, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 171, + 471, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 171, + 471, + 185 + ], + "spans": [ + { + "bbox": [ + 140, + 171, + 471, + 185 + ], + "type": "text", + "content": "1ISIR, Sorbonne Université, Paris, France 2Valeo.ai, Paris, France" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "spans": [ + { + "bbox": [ + 151, + 218, + 200, + 231 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 244, + 297, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 244, + 297, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 297, + 508 + ], + "type": "text", + "content": "Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 613 + ], + "type": "text", + "content": "With the rapid progress in Large Language Models (LLMs) [7, 10, 28, 44, 65], Multimodal LLMs (MLLMs) [3, 12, 34, 39, 61] have recently demonstrated remarkable capabilities in addressing complex multimodal tasks such as image captioning and visual question-answering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 613, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 297, + 673 + ], + "type": "text", + "content": "MLLMs are typically composed of a visual encoder, an LLM, and a connector. Following initial unimodal pretraining—and, in many cases, multimodal pretraining on large datasets—these models can be further specialized by training on multimodal datasets [1, 33]. Given the high compu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 316 + ], + "type": "text", + "content": "tational cost of training these models, recent research has proposed more efficient approaches, such as creating diverse, high-quality instruction-tuning datasets [39] or keeping the LLM frozen and fine-tuning small amounts of parameters, like the connector [41, 57, 58, 67]. These approaches take advantage of the ability of frozen LLMs to generalize to multimodal data [56]. Despite the different efficient tuning methods, training these models still incurs significant costs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 556, + 411 + ], + "type": "text", + "content": "While substantial progress has been made in developing high-performing MLLMs, relatively few studies aim to understand them [4, 47, 55, 56, 59, 61, 73]. Existing work typically conducts post-hoc analyses of MLLMs in isolation, overlooking the internal changes due to fine-tuning. Research by [56] addresses this gap to some extent by examining the internal multimodal alignment as it evolves during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 459 + ], + "type": "text", + "content": "In this work, we apply concept-level analysis to provide a readable understanding of MLLM behavior and, in particular, semantic dynamics that may occur due to fine-tuning, or due to covariate shift when considering different datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 556, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 603 + ], + "type": "text", + "content": "In the first case, we find that fine-tuning on a specific task potentially reshapes learned latent concepts, with some adjusting subtly to align with the task, and others emerging or disappearing altogether (see Fig. 1). Notably, we find that most fine-tuned concepts can be reconstructed from the original model by translating its original concepts in the direction of specific concept shift vectors, reducing the need for additional training and its associated costs. Furthermore, we explore the implications of the proposed analysis for MLLMs steering, demonstrating how model outputs can be modified inexpensively without additional training. Our key findings are summarized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 605, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 653 + ], + "type": "text", + "content": "- We apply latent concept-level analysis to provide readable understanding on MLLMs' behavior ; in particular, we show that fine-tuning can introduce significant alteration in the original concepts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 677 + ], + "type": "text", + "content": "- We show that we can control the MLLM's behavior w.r.t. certain concepts by simply manipulating shift vectors." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "- Lastly, our findings also have direct applications for steering MLLM outputs, which find use for model debiasing as well as safety control." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 684, + 113, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 684, + 113, + 693 + ], + "spans": [ + { + "bbox": [ + 66, + 684, + 113, + 693 + ], + "type": "text", + "content": "*First authors" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Project page and code: https://pegah-kh.github.io/projects/lmm-finetuning-analysis-and-steering/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2206" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "In a nutshell, we propose a novel, ready-to-use framework (including code) for MLLM behavior interpretability and control, debiasing and steering, which, we believe, will pave the way for future research." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 142, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 142, + 146 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 142, + 146 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 297, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 297, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 297, + 341 + ], + "type": "text", + "content": "Concept-based explainability. Concept-based explainability methods have emerged as an alternative to traditional feature attribution based methods, that are capable of extracting key semantic features from the model internal representations. Most post-hoc concept-based approaches are based on the idea of concept activation vectors (CAV) [30], which represent concepts as vectors in the activation space. Instead or relying on human annotations, recent works have proposed methods to automatically discover concepts via clustering [21, 71] or matrix decomposition [18], which can be viewed as instances of a dictionary learning problem [17]. Initially focusing on understanding vision models, dictionary learning for concept extraction has been extended to LLMs e.g. using sparse autoencoders [27, 52]. However, none of the prior approaches have been applied to understand MLLMs, with the exception of recently proposed CoX-LMM [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 640 + ], + "type": "text", + "content": "MLLMs and Explainability. Multimodal LLMs [3, 34, 39, 67] have recently garnered significant interest. They typically adopt a late fusion architecture, and consist of an image encoder [19, 51, 72], a connector, and an LLM [28, 63, 65]. This family of models has inspired extensive research to better understand them and explain their behavior. For example, studies like [26, 45, 55] seek to identify multimodal neurons within LLMs or analyze modality-specific sub networks [56]. Some methods leverage the fact that these models are text-generative to simply generate textual explanations for model outputs [8, 20, 59, 70]. MLLMs benefit from in-context learning capabilities, which have been examined for limitations, including biases [4] and links to hallucinations [59], as well as the factors that may enhance their in-context learning performance [9, 49]. Related to our approach, CoX-LMMs [47] employs dictionary learning to extract multimodal semantic concepts from model representations. However, these studies typically assess models only in their final trained states, overlooking the dynamic changes that occur during training. Only limited works, such as [56], have investigated explaining changes due to fine-tuning, focusing specifically on implicit alignment between image and text modalities. In this work, we investigate how multimodal concepts within the model evolve throughout fine-tuning and explore the implications of these shifts on model steering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Steering models with feature editing. In contrast to editing model weights, representation or feature editing methods [62, 66, 69] aim to modify model outputs without altering the model's weights. A prominent approach within this family involves identifying steering vectors, or directions in the feature space (often within the residual stream), that are" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 252 + ], + "type": "text", + "content": "linked to contrasting concepts. These methods have been applied to language models for various purposes, such as enhancing factuality or reducing hallucinations [46], inducing sentiment shifts or detoxification [64, 66], improving refusals to harmful requests [2], promoting truthfulness by modifying the output of attention heads [35], and erasing specific concepts or biases [6, 53]. However, their application to MLLMs is yet to be explored. Another set of approaches related to steering methods are based on In-context learning (ICL) [15, 16, 29], where prompts are carefully designed to induce desired behavior. Yet, ICL requires predefined demonstrations and lacks interpretability at a concept level. In contrast, our method offers lightweight steering, extending such capabilities to MLLMs without requiring any training, for instance as in ReFT [69]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 449, + 272 + ], + "type": "text", + "content": "3. Methodology Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 556, + 375 + ], + "type": "text", + "content": "Our framework is summarized in Fig. 1. We apply concept-level analysis of MLLM latent space (Section 3.1) to define and compare concepts between different setups. This allows us to monitor conceptual changes occurring during fin-tuning and manipulate MLLM behavior at a conceptual level (Section 3.2). Lastly, our framework finds applications for MLLM steering for e.g. debiasing and safety control (3.3) with negligible computational burden." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": "MLLM setup. A generic MLLM consists of a visual encoder " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "f_{V}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ", a trainable connector " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ", and a language model " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{LM}}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ". We assume that the model is pretrained on a multimodal (e.g. captioning) dataset " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "S = \\{(x_i, y_i)\\}_i" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": " represents images and " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "y_i \\subset \\mathcal{Y}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": " are the associated captions specified as sequence of tokens from token vocabulary space " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ". The model is trained to generate the next text tokens, conditioned on text and images. The input to " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{LM}}" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": " is a sequence of tokens that includes the concatenation of: (1) " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "N_V" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": " visual tokens extracted from the image " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": " via the visual encoder and connector " + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "inline_equation", + "content": "(C(f_V(x)))" + }, + { + "bbox": [ + 313, + 376, + 556, + 533 + ], + "type": "text", + "content": ", and (2) linearly embedded textual tokens corresponding to the text instruction and previously predicted tokens. This can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 364, + 541, + 504, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 541, + 504, + 556 + ], + "spans": [ + { + "bbox": [ + 364, + 541, + 504, + 556 + ], + "type": "interline_equation", + "content": "\\hat {y} ^ {p} = f _ {L M} (h ^ {1}, \\dots , h ^ {N _ {V}}, \\dots , h ^ {p}),", + "image_path": "ecfe0d4921c164041718d577490f6f423b06d7e08bfaf5929c9c4dc79817a6d8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "h^1, \\ldots, h^{N_V} = C(f_V(x))" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "h^p = \\operatorname{Emb}(\\hat{y}^{p-1})" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": " with Emb representing the token embedding layer. During generation, the output token " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\hat{y}^p" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": " is derived by normalizing the last layer " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "(L)" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": " tokens " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "h_{(L)}^p" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": ", then applying the unembedding layer " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "W_U" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": " and a softmax operation. The model keeps predicting the next token until the end of the sentence token to obtain the generated response " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\{\\hat{y}^p\\}_{p > N_V + N_I}" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "inline_equation", + "content": "N_I" + }, + { + "bbox": [ + 313, + 565, + 556, + 662 + ], + "type": "text", + "content": " is corresponds to the text instruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 671, + 542, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 542, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 542, + 685 + ], + "type": "text", + "content": "3.1. Retrieval and comparison of latent concepts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "To understand the internal representations of any given MLLM " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": ", we leverage the approach introduced in [47]." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2207" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 294, + 261 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 261 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 261 + ], + "type": "image", + "image_path": "054278d75a0790e4190ebed6be90a0d14c9a03ae03ee7c970853d97f5e1cefeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 293 + ], + "type": "text", + "content": "Figure 1. Framework overview. We apply concept-level analysis for MLLM behavior monitor and control, for (left) understanding and manipulating (through shift vectors) concept changes due to fine-tuning, as well as (right) MLLM steering for debiasing or safety control." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 72, + 527, + 182 + ], + "blocks": [ + { + "bbox": [ + 296, + 72, + 527, + 182 + ], + "lines": [ + { + "bbox": [ + 296, + 72, + 527, + 182 + ], + "spans": [ + { + "bbox": [ + 296, + 72, + 527, + 182 + ], + "type": "image", + "image_path": "e72ac1c9d73f1d12b0d1dffe68512f63b1401a077fe8832fe65d6ceb229070c6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 296, + 183, + 526, + 264 + ], + "blocks": [ + { + "bbox": [ + 296, + 183, + 526, + 264 + ], + "lines": [ + { + "bbox": [ + 296, + 183, + 526, + 264 + ], + "spans": [ + { + "bbox": [ + 296, + 183, + 526, + 264 + ], + "type": "image", + "image_path": "4941f3fec2ee19335fb9bef24aaffe6f277b15a2a2b1d2676fc371e95cf25dd8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": "Specifically, given a set of " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\{x_{1},\\dots,x_{M}\\}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " we extract a set of residual stream representations from some layer " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " of the MLLM " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": ". These representations " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_m = f_l(x_m)\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " (one per image) are collected in a feature matrix " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{Z}\\in \\mathbb{R}^{D\\times M}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": ". Typically, the set of images and extracted representations correspond to a particular token of interest TOI (e.g., 'Dog', 'Cat', 'Person', etc.) in the predicted caption. However, the extraction can also be performed for a larger set of target tokens. This feature matrix " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{Z}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " is then decomposed as " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{Z}\\approx \\pmb{U}\\pmb{V}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " to recover the concepts in the latent embedding space. Here, " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{U}\\in \\mathbb{R}^{D\\times K}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " is the matrix of " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " concepts and " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{V}\\in \\mathbb{R}^{K\\times M}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " represents the coefficients/activations of the samples projected onto these concepts. Different decompositions of the matrix " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " result in various concepts, inheriting the properties of the decomposition (such as low grounding overlap with PCA). We employ " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": "-Means to learn our concept dictionaries. This is motivated by " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": "-Means' simplicity, and straightforward arithmetic manipulation of the clusters/concepts it allows. Each column " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k\\in \\pmb{U}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " corresponds to a concept, while each column of " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " encodes the activation of these concepts for a given sample. Note that any given representation " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "f_{l}(x)" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " can be projected on " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{U}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " to obtain its activation vector " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{v}(x)\\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": ", i.e. " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "f_{l}(x)\\approx U\\pmb{v}(x)" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": ". Each extracted concept is then interpreted through grounding in both image and text spaces. Specifically, the top " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{MAS}}" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " that activates concept " + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k" + }, + { + "bbox": [ + 56, + 306, + 296, + 617 + ], + "type": "text", + "content": " the most represent its image grounding:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 630, + 296, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 630, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 86, + 630, + 296, + 657 + ], + "type": "interline_equation", + "content": "\\boldsymbol {X} _ {\\text {M A S}} \\left(\\boldsymbol {u} _ {k}\\right) = \\underset {\\hat {X} \\subset \\mathbf {X} _ {t}, | \\hat {X} | = N _ {\\text {M A S}}} {\\arg \\max } \\sum_ {x \\in \\hat {X}} \\left| \\boldsymbol {v} _ {k} (x) \\right|, \\tag {1}", + "image_path": "d197135a22bcabf368f9630d1a1a7270570ea11a9f41ba2222b17470e5851a5c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{v}_k(x)" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " refers to the activation of " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " for image " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": ". For text grounding, we decode the features using the unembedding matrix of the language model " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "W_U" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " [5, 32, 43, 54]. Specifically, the operation " + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "inline_equation", + "content": "W_U\\pmb{u}_k \\in \\mathbb{R}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 55, + 666, + 296, + 715 + ], + "type": "text", + "content": " produces" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 555, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 330 + ], + "type": "text", + "content": "logits over the vocabulary, and the top " + }, + { + "bbox": [ + 313, + 306, + 555, + 330 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{grounding}}" + }, + { + "bbox": [ + 313, + 306, + 555, + 330 + ], + "type": "text", + "content": " words with highest logits are extracted:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 364, + 342, + 555, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 342, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 364, + 342, + 555, + 364 + ], + "type": "interline_equation", + "content": "\\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} _ {k}) = \\underset {\\text {T o p} - N _ {\\text {g r o u n d i n g}}} {\\arg \\max } \\left(W _ {U} \\boldsymbol {u} _ {k}\\right). \\tag {2}", + "image_path": "61282f250960e117b59affbee48af22e7ae2454ec9ca341bc0fea797a31e2fab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 369, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 393 + ], + "type": "text", + "content": "Finally, to quantify the similarity of two concepts (i.e. the columns of " + }, + { + "bbox": [ + 313, + 369, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 369, + 555, + 393 + ], + "type": "text", + "content": " we define the Text Grounding Overlap as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 401, + 555, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 401, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 321, + 401, + 555, + 429 + ], + "type": "interline_equation", + "content": "\\operatorname {T} \\text {O v e r l a p} (\\boldsymbol {u}, \\boldsymbol {u} ^ {\\prime}) = 1 0 0 \\times \\frac {\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\cap \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u} ^ {\\prime}) \\right|}{\\left| \\boldsymbol {T} _ {\\text {w o r d s}} (\\boldsymbol {u}) \\right|}. \\tag {3}", + "image_path": "3bd38802fe77216c15b5514219937cba28450f9169949f7fd86df160268cc3ca.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 436, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 461 + ], + "type": "text", + "content": "Now that we have defined a generic framework, we present two subcases for extracting and manipulating concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 534, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 534, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 534, + 481 + ], + "type": "text", + "content": "3.2. Evolution of concepts through fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "text", + "content": "Setup overview. An original model " + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "text", + "content": " is typically finetuned to produce a specialized model " + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "text", + "content": " for a particular task—or, specifically, for a set of target concepts. This finetuning can be conducted on samples that include a set of words " + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\{w_1, \\dots, w_m\\}" + }, + { + "bbox": [ + 312, + 486, + 555, + 641 + ], + "type": "text", + "content": " associated with these target concepts. For instance, if we fine-tune a image captioning model to emphasize colors in the image, the set of words will simply be these colors. Efficient fine-tuning is typically achieved using Low-Rank Adaptation (LoRA) [25, 34, 39]. Finetuning can selectively alter certain representations, leading to shifts in the conceptual space encoded by the model. Using the interpretability framework discussed in Section 3.1 we can study these shifts at a readable conceptual level." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Concept recovery via shift vectors. To study the change from an original model " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " to a finetuned model " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", we fix the dataset " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "S^{(1)} = S^{(2)}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", and obtain two sets of embeddings from " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "f^a, f^b" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " respectively, i.e. " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "A \\approx U^a V^a" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "B \\approx U^b V^b" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "U^a, U^b \\in \\mathbb{R}^{D \\times K}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " concepts extracted from each model. We propose to characterize the concept changes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2208" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "from an original to fine-tuned model as linear directions in embedding space or concept shift vectors. To do so, we first associate each original concept " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k^a\\in U^a" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " with a subset of samples where " + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "inline_equation", + "content": "\\pmb{u}_k^a" + }, + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": " is the most activated concept:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 126, + 253, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 126, + 253, + 146 + ], + "spans": [ + { + "bbox": [ + 99, + 126, + 253, + 146 + ], + "type": "interline_equation", + "content": "\\pmb {A} _ {k} = \\{m \\mid k = \\underset {i} {\\arg \\max} | \\pmb {v} _ {i} ^ {a} (x _ {m}) | \\}.", + "image_path": "69865ce55839f26cf01759be0343c0ac0163f4c712e29a5141d70beb1492d47f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": "For each sample " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "x_{m}, m \\in \\mathbf{A}_{k}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": " we define " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\delta_{m}^{a \\to b} = \\pmb{b}_{m} - \\pmb{a}_{m}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": " as the change in its representation from " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "f^{a}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "f^{b}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": ". To compute the concept shift vector " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_{k}^{a \\to b}(\\pmb{u}_{k}^{a})" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": " associated with " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{k}^{a}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": ", we aggregate shifts of its associated samples specified by " + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{k}" + }, + { + "bbox": [ + 55, + 152, + 297, + 201 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 215, + 296, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 215, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 56, + 215, + 296, + 246 + ], + "type": "interline_equation", + "content": "\\pmb {\\Delta} _ {k} ^ {a \\rightarrow b} (\\pmb {u} _ {k} ^ {a}) = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} \\delta_ {m} ^ {a \\rightarrow b} = \\frac {1}{| \\pmb {A} _ {k} |} \\sum_ {m \\in \\pmb {A} _ {k}} (\\pmb {b} _ {m} - \\pmb {a} _ {m})", + "image_path": "5922754081473722fa29e41d4c4ac0f40c8ac32d5141ba82b4a04ea8ccaa4f41.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "text", + "content": "The concept shift vector is used to shift each concept in the original model " + }, + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_k^a" + }, + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "text", + "content": " to obtain the shifted concept " + }, + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_k^s" + }, + { + "bbox": [ + 55, + 250, + 296, + 275 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 279, + 296, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 279, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 121, + 279, + 296, + 295 + ], + "type": "interline_equation", + "content": "\\boldsymbol {u} _ {k} ^ {s} = \\boldsymbol {u} _ {k} ^ {a} + \\alpha \\Delta_ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right), \\tag {4}", + "image_path": "7f5fedf72a68e47cd465edbec9d093a7c4c9ac4313e65255c42a3c2408441bd5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " is a coefficient to control the shift magnitude. Unless otherwise stated, we use " + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 300, + 296, + 407 + ], + "type": "text", + "content": " as the default magnitude of shift. It is worth noting that given the concept shift vectors, the computation of shifted concepts does not rely on accessing the fine-tuned model. Practically speaking, this means that we can \"push\" the original model towards the concepts of a fine-tuned one with very little overhead, by simply shifting its latent representation in the direction of the shift vector, as it will be illustrated in the experiments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 414, + 296, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 440 + ], + "type": "text", + "content": "3.3. Concept evolution across datasets and applications to model steering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "text", + "content": "Concept comparison between different datasets. Another interesting subcase of the proposed framework consists in evaluating the shift from one dataset " + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "inline_equation", + "content": "S^{(1)}" + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "text", + "content": " to another " + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "inline_equation", + "content": "S^{(2)}" + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "text", + "content": ", using the same model and encoding " + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 443, + 296, + 538 + ], + "type": "text", + "content": ". Beyond interpretability of a model behavior, it also find applications for model steering. Model steering (see Fig. 1 (right)) refers to guiding the model outputs towards desired outcomes by modifying the features without altering the model weights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": "Coarse-grained model steering. In coarse-grained or global steering, the objective is to adjust the model outputs " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " to generally align with a set of target samples (e.g., changing answers type). Given input-output samples, we first extract the answer representations " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "B = b_{1},\\dots,b_{N}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " from the target set. Similarly, we obtain representations for the original set " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "A = a_{1},\\dots,a_{M}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " (e.g. randomly drawn from the train set). We then compute the coarse steering vector " + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "inline_equation", + "content": "s_{c}" + }, + { + "bbox": [ + 55, + 539, + 296, + 635 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 123, + 642, + 296, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 642, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 123, + 642, + 296, + 669 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {c} = \\frac {\\sum_ {i} ^ {N} \\boldsymbol {b} _ {i}}{N} - \\frac {\\sum_ {i} ^ {M} \\boldsymbol {a} _ {i}}{M}, \\tag {5}", + "image_path": "90c4afa0f08430a89b112ef580ee62fe92182f2be6cc079ba7a2fe6a6061365a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": " is applied to all the samples in the validation set. For instance, the activations " + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "f_{l}(x_{i})" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": " of a sample " + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 670, + 296, + 696 + ], + "type": "text", + "content": " become:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 128, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\tilde {f} _ {l} \\left(x _ {i}\\right) = f _ {l} \\left(x _ {i}\\right) + \\alpha \\boldsymbol {s} _ {c} \\tag {6}", + "image_path": "ff8f5cee84ad32ac8c7c9d21ce3abfbf6b07883ccd8f531f223eaa8c97294ea2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 552, + 135 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 552, + 135 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 552, + 135 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 552, + 135 + ], + "type": "image", + "image_path": "90bd6d7f34acd49ea52a792b06f82d4fa94c74e52851369fd682616b18ae41ee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": "Figure 2. Concepts extracted from original and fine-tuned models. concepts from the original " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": " (top), and model fine-tuned to focus more on places " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": " (bottom), for " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "TOI =" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": " person. The concepts from " + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 313, + 149, + 555, + 194 + ], + "type": "text", + "content": " exhibit a stronger association with places." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " controls the steering strength and it is set to 1 (we study " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " in App. B.5.3). Thus, in this setup, all examples are coarsely steered in the direction of the steering vector " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " at layer " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": ", before passing " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{f}_l(x_i)" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " through the rest of layers and " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{f}_l(x_i)" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": " becomes the input to the next layer " + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 313, + 198, + 555, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": "Fine-grained steering. Unlike global steering, fine-grained steering consists in finding and editing directions that adjust only certain concepts to other ones. To do this, we decompose the hidden states of a set of samples into a set of concepts " + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": " as previously explained. We then compute a set of fine-grained steering vectors " + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "inline_equation", + "content": "s^f = s_{11}^f, \\ldots, s_{NN}^f" + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "inline_equation", + "content": "s_{ij}^f = s_{ij}^f = u_j - u_i" + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": " the steering vector from concept " + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "inline_equation", + "content": "u_j" + }, + { + "bbox": [ + 313, + 259, + 556, + 440 + ], + "type": "text", + "content": ". However, not all steering vectors are meaningful: options for finding the relevant ones include proximity matching, as well as identifying vectors that have the strongest impact on guiding the model towards generating specific answers or concepts (e.g. producing significantly more target answers). This is more detailed in App. B.3. Various applications of MLLM steering can be explored, such as gender debiasing and safety alignment, as it will be shown below." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 449, + 395, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 395, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 395, + 463 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 468, + 452, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 452, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 452, + 481 + ], + "type": "text", + "content": "4.1. Fine-tuning experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 521 + ], + "type": "text", + "content": "In this section, we study how fine-tuning introduces changes in the overall structure of the learned concepts in MLLMs (with the architecture described in Section 3)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 522, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 689 + ], + "type": "text", + "content": "Implementation details. The main paper covers experiments on the popular LLaVA [38] model comprising a CLIP image encoder, a two-layer MLP connector, and a 7B Vicuna-1.5 LLM. More experiments on a different multimodal model can be found in App. A. We conduct our study in a controlled setup that consists in specializing the model on a target dataset. Specifically, we apply fine-tuning on three different subsets of the Visual Genome dataset [31], related to places, colors, and sentiments (more details in App. A.2). Impact of fine-tuning on learned concepts. Fig. 2 depicts this concept change. Through the concepts groundings for an example TOI (person) we see that the fine-tuned model puts a stronger emphasis on places, which is expected and serves as a sanity check for our method." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Matched concepts. To further analyze how each concept changes after fine-tuning, we focus on each concept and its" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2209" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 282, + 198 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 282, + 198 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 282, + 198 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 282, + 198 + ], + "type": "image", + "image_path": "c10d0f78db351f015b227b953972a226d4fcbfd61fd6ab3a813acf94dfa66fd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "text", + "content": "Figure 3. Concepts text grounding change after fine-tuning. Text grounding for concepts " + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "inline_equation", + "content": "(TOI = \\text{bus})" + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "text", + "content": " and their match from " + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 55, + 204, + 296, + 270 + ], + "type": "text", + "content": ", (fine-tuned to focus more on places). Emerging concepts may include grounding words not explicitly included in the fine-tuning vocabulary for place (e.g., \"District\", \"Crossing\"), while others evolve more smoothly (e.g., \"Street\")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": "match. Specifically, we define a matching function " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "m: i \\to j^*" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": " which associates each concept vector " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{u}_i^a" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": " in set " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{U}^a" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": " to its closest vector " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{j*}^b" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": " in set " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\pmb{U}^b" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": " based on cosine similarity, i.e. " + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "inline_equation", + "content": "m(i) = \\arg \\max_{\\pmb{u}_j^b \\in \\pmb{U}_b} \\cos(\\pmb{u}_i^a, \\pmb{u}_j^b)" + }, + { + "bbox": [ + 55, + 275, + 296, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 434 + ], + "type": "text", + "content": "Fig. 3 shows the text groundings for various concepts, displaying the words with a frequency lower than 5 across concepts (e.g., filtering out high-frequency terms like bus, vehicle, etc.). We observe the emergence of place-related terms across the identified elements, while the overall thematic structure remains consistent. Note that certain concepts may converge toward the same fine-tuned concept. We also analyze how the distances between matched concepts evolve during fine-tuning in App. A.3." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 443, + 294, + 557 + ], + "blocks": [ + { + "bbox": [ + 56, + 443, + 294, + 557 + ], + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 557 + ], + "type": "image", + "image_path": "777d941f6424d01022b485dab279e939b753602f35e760f2a92a6a45835c844b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 296, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 594 + ], + "type": "text", + "content": "Figure 4. Text grounding overlap (T-Overlap) between original and fine-tuned model concepts. Different concepts change to different extents depending on the fine-tuning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "content": "Concept evolution. To quantify how much a concept " + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_i^a \\in \\boldsymbol{U}^a" + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "content": " is changed after fine-tuning, we compute the overlap between its grounding words and those of its closest matching concept from the fine-tuned model " + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{U}^b" + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "content": ". Specifically, we compute T-Overlap " + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "inline_equation", + "content": "(\\boldsymbol{u}_i^a, \\boldsymbol{u}_{m(i)}^b)" + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "content": " (Eq. (3)) for all the concepts " + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 604, + 297, + 715 + ], + "type": "text", + "content": ", and visualize them for different fine-tunings in Fig. 4. We observe varying rates of change across different concepts and fine-tunings. This might be due to the difference in the fine-tuning dataset size, complexity," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "or similarity to the original dataset. It also highlights 2 main behaviors, detailed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 97, + 556, + 220 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 97, + 556, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 556, + 146 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 556, + 146 + ], + "type": "text", + "content": "- Concepts that are refined. This group contains the concepts that slightly change to be more specialized towards the fine-tuning task (Fig. 5 top, middle rows). These concepts exhibit a relatively high (T-Overlap " + }, + { + "bbox": [ + 313, + 97, + 556, + 146 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)" + }, + { + "bbox": [ + 313, + 97, + 556, + 146 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 145, + 556, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 556, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 556, + 220 + ], + "type": "text", + "content": "- Concepts that change completely. This group contains the concepts that emerge or, to a certain extent, disappear (Fig. 5 bottom row) in the fine-tuned model. New concepts emerge during fine-tuning likely due to the introduction of novel patterns or relationships. These concepts exhibit a relatively low (T-Overlap " + }, + { + "bbox": [ + 313, + 145, + 556, + 220 + ], + "type": "inline_equation", + "content": "(\\pmb{u}_i^a,\\pmb{u}_{m(i)}^b)" + }, + { + "bbox": [ + 313, + 145, + 556, + 220 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 325, + 240, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 325, + 240, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 325, + 240, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 325, + 240, + 545, + 415 + ], + "type": "image", + "image_path": "614281c11141c2af45f53ba86312363b394d893afb497f4689914332056ce9de.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "type": "text", + "content": "Figure 5. Concepts evolve differently due to fine-tuning. Left: concepts extracted from the original model " + }, + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "type": "inline_equation", + "content": "f^a" + }, + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "type": "text", + "content": ". Right: matched concept extracted from " + }, + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "type": "inline_equation", + "content": "f^b" + }, + { + "bbox": [ + 313, + 422, + 556, + 534 + ], + "type": "text", + "content": ", fine-tuned to focus on places. Each concept is grounded in image and text. We observe different levels of adaptation across concepts: some concepts specialize further by adding to place-related words (e.g., Top), some concepts introduce place-related words while staying aligned with the original concept in textual or visual groundings (e.g., Middle), while others undergo complete transformation, diverging significantly from their original meaning to fully embrace place-related elements (e.g., Bottom)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": "We also notice that T-Overlap decreases with the number of training iterations, indicating that fine-tuning leads to deviation from the original concepts (more details in App. A.3). Evaluating fine-tuned concept recovery. To study if the fine-tuned concepts " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{b}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " can be recovered from the original ones " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{U}^{a}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": ", we first establish a matching " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "(m : i \\rightarrow j)" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " between the set of original " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\{\\pmb{u}_{i}^{a}\\}_{i=1}^{K}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " and fine-tuned concepts " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\{\\pmb{u}_{j}^{b}\\}_{j=1}^{K}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": ". For systematic evaluation of recovery of all fine-tuned concepts, we constrain " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " to be bijective using an optimal transport algorithm detailed in App. A.1. Finally, we evaluate how well a shifted concept " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{k}^{s}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " (Equ. (4)) is similar to its match " + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{m(k)}^{b}" + }, + { + "bbox": [ + 313, + 545, + 557, + 713 + ], + "type": "text", + "content": " using the aforementioned T-Overlap metric. Fig. 6 shows the results of recovering the fine-tuned concepts for models fine-tuned on different subsets of the VG" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2210" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 75, + 290, + 235 + ], + "blocks": [ + { + "bbox": [ + 61, + 75, + 290, + 235 + ], + "lines": [ + { + "bbox": [ + 61, + 75, + 290, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 75, + 290, + 235 + ], + "type": "image", + "image_path": "5492ad358c17c4139aeef04a9c3285898c141926b586baf66e5256296d256e22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 291 + ], + "type": "text", + "content": "Figure 6. Recovering fine-tuned concepts. Across different finetunings (places (top), colors (middle) and sentiments (bottom)), we compute the average text grounding overlap of original and shifted concepts with the matched fine-tuned ones. Shifting the original concepts result in partially recovering the fine-tuned ones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "content": "dataset (place, color, sentiment). We report the T-Overlap between the shifted " + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "inline_equation", + "content": "u_{k}^{s}" + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "content": " and fine-tuned concepts " + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "inline_equation", + "content": "u_{m(k)}^{b}" + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "content": " for various different tokens of interest. For each target token and finetuning we extract " + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "inline_equation", + "content": "K = 20" + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "content": " concepts and report the mean and standard deviation over them. We use the overlap between the original concepts " + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "inline_equation", + "content": "u_{k}^{a}" + }, + { + "bbox": [ + 55, + 295, + 296, + 415 + ], + "type": "text", + "content": " and the fine-tuned ones as a baseline. We observe that most shifted concepts show higher overlap than the original ones: this demonstrates that fine-tuned concepts can be efficiently recovered from the original model with concept shift vectors." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 430, + 282, + 552 + ], + "blocks": [ + { + "bbox": [ + 67, + 430, + 282, + 552 + ], + "lines": [ + { + "bbox": [ + 67, + 430, + 282, + 552 + ], + "spans": [ + { + "bbox": [ + 67, + 430, + 282, + 552 + ], + "type": "image", + "image_path": "f5d587dc3cd6761cb01fdd6ba1f2740fc4c487e9ffe493c4d7b03527b89838d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 555, + 295, + 599 + ], + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 599 + ], + "type": "text", + "content": "Figure 7. Correlation between shift consistency and concept recovery (Color finetuning). The more consistent and aligned the individual representation shifts associated with a concept, the better the recovery of the fine-tuned concept." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Which concepts are recovered better? We hypothesize that if the representation shift for individual samples associated with a concept " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_k^a" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\delta_m^{a\\rightarrow b}|m\\in \\boldsymbol{A}_k\\}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ", is consistently aligned with the concept shift vector, the resulting " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a)" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ", should be more effective at recovering the finetuned concept. We quantify the consistency through mean cosine similarity of " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\delta_m^{a\\rightarrow b}\\}_{m\\in \\boldsymbol{A}_k}" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": " with the concept shift vector " + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Delta_k^{a\\rightarrow b}(\\boldsymbol{u}_k^a)" + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": ". In other words, this quantifies the alignment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 508, + 86 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 508, + 86 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 508, + 86 + ], + "type": "text", + "content": "of individual shifts and their mean, " + }, + { + "bbox": [ + 313, + 72, + 508, + 86 + ], + "type": "inline_equation", + "content": "\\pmb{\\Delta}_k^{a\\rightarrow b}(\\pmb{u}_k^a)" + }, + { + "bbox": [ + 313, + 72, + 508, + 86 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 91, + 543, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 91, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 324, + 91, + 543, + 121 + ], + "type": "interline_equation", + "content": "\\operatorname {C o n s i s t e n c y} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right) = \\frac {1}{\\left| \\boldsymbol {A} _ {k} \\right|} \\sum_ {m \\in \\boldsymbol {A} _ {k}} \\cos \\left(\\delta_ {m}, \\boldsymbol {\\Delta} _ {k} ^ {a \\rightarrow b} \\left(\\boldsymbol {u} _ {k} ^ {a}\\right)\\right),", + "image_path": "26086dd021d9b48704df21f6a87de35201f8a501659a9fc73bf67601bdb696e9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": "We measure the recovery of concept " + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "inline_equation", + "content": "\\mathrm{CR}_k" + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": ", as the improvement in similarity between the matched fine-tuned " + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_{m(k)}^b" + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": " and shifted concept " + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_k^s" + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": ", relative to the original one " + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}_k^a" + }, + { + "bbox": [ + 313, + 128, + 555, + 165 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 171, + 555, + 203 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 171, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 347, + 171, + 555, + 203 + ], + "type": "interline_equation", + "content": "\\mathrm {C R} _ {k} = \\frac {\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {s}\\right) - \\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)}{\\cos \\left(\\boldsymbol {u} _ {m (k)} ^ {b} , \\boldsymbol {u} _ {k} ^ {a}\\right)} \\tag {7}", + "image_path": "f0be3e2b481668599e70573c29088bb3541ef33088f2812ebea5c1486e23b14f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "text", + "content": "We plot the consistency and recovery for concepts extracted across four tokens of interest for color finetuning in Fig. 7. Plots for other finetunings are in App. A.5. Crucially, we observe a positive and statistically significant correlation between the two quantities across all the finetuning tasks. This supports our hypothesis that a better concept shift recovery is related to consistency of individual shifts of original concept. More ablation studies about concept recovery are available in App. A.4, analyzing the influence of steering strength " + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "text", + "content": ", number of concepts " + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "text", + "content": ", and extraction layer " + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 209, + 556, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 338, + 555, + 391 + ], + "blocks": [ + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "type": "image", + "image_path": "299dd8fe0458bcc4305331a96826d04fa3c8832368af4528aaa6862f7df71b8f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 426 + ], + "type": "text", + "content": "Figure 8. Linear separability of concepts features in MLLMs. We visualize the features related to the concepts \"yes\" and \"no\" after PCA projections across MLLMs layers." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 556, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 546 + ], + "type": "text", + "content": "In summary, we demonstrated the feasibility of recovering target fine-tuned concepts by applying simple per-concept shift of the original model features. This supposes that the features related to different concepts are almost linearly separable, which empirically seems to hold at least for the last MLLM layers as pictured on Fig. 8, and as previously studied for LLMs in [42, 48]. This motivates the following investigations on using a similar methodology for simple, computationally inexpensive, yet efficient MLLM steering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 464, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 464, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 464, + 565 + ], + "type": "text", + "content": "4.2. Multimodal model steering" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "content": "We perform steering by applying a steering vector " + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "content": " to the residual stream features " + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "content": " of the MLLM, without changing its parameters. We first evaluate the MLLM steering capabilities in a visual question-answering (VQA) setup. Then, we show the applicability of our MLLM steering to control captioning styles. Lastly, we present two steering applications: gender debiasing, aiming to mitigate biases in model outputs, and safety alignment, i.e. ensuring that the model refuses to generate harmful information. We discuss technical details for each of these applications in App. C and App. D." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Setup. For VQA tasks, each query consists of a question about an input image, and the model generates an answer." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2211" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 192 + ], + "type": "text", + "content": "To measure the effectiveness of our approach in directing the model towards specific answers or answer types, we report the number of generated answers that align with the target output or answer type. Additionally, we aim for targeted steering, ensuring that only specific answer types are influenced. For example, when altering answers from \"yes\" to \"no\" within the \"yes/no\" category, responses to other question types should remain unaffected. This specificity is assessed by tracking accuracy across answer types and number of answers from each type." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 326 + ], + "type": "text", + "content": "Implementation details. Experiments in the main paper are primarily conducted on LLaVA [39] for conciseness. However, we show in App. B.2 that our method is generally applicable to other popular MLLMs. We experiment on VQAv2 [24], a visual question-answering corpus with image-question-answer triplets and annotated answer types (\"yes/no\", \"number\", and \"other\"). Steering vectors are derived from a subset of the train set, with model performance evaluated on the val set. As steering becomes more effective in deeper layers (see Fig. 8 and App. B), we apply it on the last layer. Additional experiments can be found in App. B." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 111, + 337, + 241, + 403 + ], + "blocks": [ + { + "bbox": [ + 111, + 337, + 241, + 403 + ], + "lines": [ + { + "bbox": [ + 111, + 337, + 241, + 403 + ], + "spans": [ + { + "bbox": [ + 111, + 337, + 241, + 403 + ], + "type": "table", + "html": "
Target TypeAnswers Type
yes/nonumberother
N/A366122494
yes/no55796288
number327201390
other364115501
", + "image_path": "02ac6cfa30e14a827b5f27ed97fc7de827b440993ac669570fee2592b83e324f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 450, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 450, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 450, + 296, + 714 + ], + "type": "text", + "content": "Coarse and Fine-Grained Model Steering for VQA. We explore both coarse and fine-grained steering in VQA tasks. Specifically, coarse steering aims to alter the distribution of answers, while fine-grained steering targets specific responses. For coarse-grained steering, we direct the model's answers toward a particular category: yes/no, numbers, or other (e.g., colors, objects). We compute a steering vector for each target answer type. As shown in Table 1, applying steering significantly increases the proportion of the targeted answer type. For fine-grained steering, we first assess the feasibility of identifying such steering vectors. Fig. 9 illustrates examples of these vectors. We derive them from three sets of sample answers corresponding to \"yes/no,\" \"number,\" and \"other\" categories. Interestingly, some vectors distinctly align with specific answers such as \"No,\" \"4,\" and \"Red.\" This confirms the potential to identify fine-grained steering vectors capable of guiding the model toward a precise response. Building on these insights, we seek to steer the model toward a user-specified answer. For each original/target answer pair (e.g., yes/no), we collect some samples and compute the corresponding (coarse) steering vector. We then apply these vectors to all validation set samples. In" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 402, + 141 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 402, + 141 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 402, + 141 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 402, + 141 + ], + "type": "image", + "image_path": "2437368440318cfa7421b8fec918c07aac795a916de04c757b4498e1bc11f4cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 408, + 72, + 465, + 141 + ], + "blocks": [ + { + "bbox": [ + 408, + 72, + 465, + 141 + ], + "lines": [ + { + "bbox": [ + 408, + 72, + 465, + 141 + ], + "spans": [ + { + "bbox": [ + 408, + 72, + 465, + 141 + ], + "type": "image", + "image_path": "2eafbad9d9f38f450f1d7c90b8f83717326881a7fdb394bfa8e4ccbdb06057aa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 467, + 72, + 530, + 140 + ], + "blocks": [ + { + "bbox": [ + 467, + 72, + 530, + 140 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 530, + 140 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 530, + 140 + ], + "type": "image", + "image_path": "5dbd2cd46908a19efc09ce597283e096720d1d0d1e12a8a1519e538d1561d14e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 341, + 143, + 400, + 213 + ], + "blocks": [ + { + "bbox": [ + 341, + 143, + 400, + 213 + ], + "lines": [ + { + "bbox": [ + 341, + 143, + 400, + 213 + ], + "spans": [ + { + "bbox": [ + 341, + 143, + 400, + 213 + ], + "type": "image", + "image_path": "adc0117383fb2ab858d38ae414b29cb2ba181e696fb506c80b32ce17455b2b1c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 402, + 144, + 466, + 213 + ], + "blocks": [ + { + "bbox": [ + 402, + 144, + 466, + 213 + ], + "lines": [ + { + "bbox": [ + 402, + 144, + 466, + 213 + ], + "spans": [ + { + "bbox": [ + 402, + 144, + 466, + 213 + ], + "type": "image", + "image_path": "c5394683e576cc3287e39c9a7c1bb06bfb5f74aba6e291ac6fd7b7eb3a88d544.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 470, + 144, + 529, + 210 + ], + "blocks": [ + { + "bbox": [ + 470, + 144, + 529, + 210 + ], + "lines": [ + { + "bbox": [ + 470, + 144, + 529, + 210 + ], + "spans": [ + { + "bbox": [ + 470, + 144, + 529, + 210 + ], + "type": "image", + "image_path": "a786d751a8b150d9408fb1a8d1ae2452ecea1c8174587e99a4b929145e8ddde7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 341, + 215, + 400, + 285 + ], + "blocks": [ + { + "bbox": [ + 341, + 215, + 400, + 285 + ], + "lines": [ + { + "bbox": [ + 341, + 215, + 400, + 285 + ], + "spans": [ + { + "bbox": [ + 341, + 215, + 400, + 285 + ], + "type": "image", + "image_path": "7a7c6ec020e1e00fabaded1ca98e6ed98e0dc2e2ae63193e1318d1e6456cfcc2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 288, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 365 + ], + "type": "text", + "content": "Figure 9. Discovering meaningful steering directions. Each line corresponds to a fine-grained steering direction to steer the model answer to: \"No\" (yes/no), \"4\" (number) and \"Red\" (other). Some steering directions are targeted (e.g., \"No\") as there is slight change in both the accuracy and number of answers types on other types (e.g., number, other). We show the relative scores compared to a baseline with no steering." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 216, + 466, + 285 + ], + "blocks": [ + { + "bbox": [ + 402, + 216, + 466, + 285 + ], + "lines": [ + { + "bbox": [ + 402, + 216, + 466, + 285 + ], + "spans": [ + { + "bbox": [ + 402, + 216, + 466, + 285 + ], + "type": "image", + "image_path": "0c54b6c6263624626db32666059d183fa4bef536eeaa6b916092b22e6e261687.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 470, + 215, + 529, + 285 + ], + "blocks": [ + { + "bbox": [ + 470, + 215, + 529, + 285 + ], + "lines": [ + { + "bbox": [ + 470, + 215, + 529, + 285 + ], + "spans": [ + { + "bbox": [ + 470, + 215, + 529, + 285 + ], + "type": "image", + "image_path": "3b567c38f1d4f3240a4c9354fd1b32de7d421a883f844d9321f123281dfdb812.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 369, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 418 + ], + "type": "text", + "content": "Table 2, we report evaluation metrics when steering at the last layer. The results show that steering effectively increases the occurrence of target answers, while accuracy on other answer types remains largely stable." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 316, + 426, + 553, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 412, + 296, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 436 + ], + "type": "text", + "content": "Table 1. Steering MLLMs answers type. Number of target answers type increases significantly after model steering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 480 + ], + "type": "table", + "html": "
SteeringAccuracy (%)Answer TypesAnswers
Yes/NoNumberOtherYes/NoNumberOtherOriginalTarget
N/A90.8258.4771.101861687234900
Yes → No69.0356.8268.9918846952294-828+828
1 → 390.7154.5271.1218616702350-215+144
White → Black90.4058.4258.3618616712312-98+441
", + "image_path": "f8d676b4eac264586d49b16eebb1a2ef42c798a8ecd601bc177df73f46105dd9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 488, + 555, + 554 + ], + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 554 + ], + "type": "text", + "content": "Table 2. Steering MLLMs answers. Steering answers from \"Yes\" (Yes/No), \"1\" (Number), \"White\" (Other) to \"No\", \"3\", \"Black\" respectively. The number of original/target answer counts decreases/increases significantly, while the accuracy on other answer types changes only slightly, and the number of answer type counts remains almost constant." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 574, + 395, + 639 + ], + "blocks": [ + { + "bbox": [ + 317, + 574, + 395, + 639 + ], + "lines": [ + { + "bbox": [ + 317, + 574, + 395, + 639 + ], + "spans": [ + { + "bbox": [ + 317, + 574, + 395, + 639 + ], + "type": "image", + "image_path": "32e4ee70a330dce9139f5043b03d94a5a748ed60e6da9568c95fe151cf32fd4d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 644, + 555, + 667 + ], + "lines": [ + { + "bbox": [ + 313, + 644, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 644, + 555, + 667 + ], + "type": "text", + "content": "Figure 10. Steering MLLMs captions style. Captions steered to focus more on colors (left), places (middle) and sentiments (right)." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 575, + 474, + 638 + ], + "blocks": [ + { + "bbox": [ + 395, + 575, + 474, + 638 + ], + "lines": [ + { + "bbox": [ + 395, + 575, + 474, + 638 + ], + "spans": [ + { + "bbox": [ + 395, + 575, + 474, + 638 + ], + "type": "image", + "image_path": "e6eed37fede6ea1430cd86ae9cf2c93fcd37c3b4edb13ea3fe9b3a8f5b564b7c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 474, + 574, + 553, + 638 + ], + "blocks": [ + { + "bbox": [ + 474, + 574, + 553, + 638 + ], + "lines": [ + { + "bbox": [ + 474, + 574, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 474, + 574, + 553, + 638 + ], + "type": "image", + "image_path": "4166487faa950fc72e254be5e84cb072a94f113261e82c27d925cf772711959d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Steering image caption styles. We previously applied steering on relatively brief answers from the VQAv2 dataset. Here, we extend this approach to longer, descriptive outputs" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2212" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": "using the COCO captioning dataset [37]. Given that multiple captions can effectively describe an image by emphasizing various aspects such as the main object, surroundings, actions, or events, we aim at modifying captions to align with a specific target style. Here, we learn a coarse steering vector between samples with predicted captions in the target style and random samples. Qualitative examples of this steering for LLaVA are in Fig. 10. Results in Table 3 demonstrate that captions can be effectively steered towards a target style even when considering tasks with longer responses. We provide more captioning experiments in App. B.4." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 213, + 248, + 280 + ], + "blocks": [ + { + "bbox": [ + 106, + 213, + 248, + 280 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 248, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 248, + 280 + ], + "type": "table", + "html": "
Target StyleCaptionsStyle
placescolorssentiments
N/A43013092
places79610771
colors48825611
sentiments393104048
", + "image_path": "76f9ff0ea4f2f1cadc5823c2c03effaa8fcaba4012c7f63616de4e69ca3be7cc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": "Gender debiasing captions. We perform gender debiasing on COCO test set, aiming at mitigating biases for any gendered nouns when captioning. We experiment with both coarse and fine-grained steering. The coarse steering vector is computed between sets of samples with a gendered/neutral noun in the caption. For fine-grained steering, we steer each concept to its closest (as defined by its posing similarity) counterpart among the neutral concepts, as explained in Section 3.3. The results are reported in Table 4. Interestingly, both strategies are capable of converting many gendered captions to neutral ones, with fine-grained steering being significantly more effective than coarse steering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 492, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 296, + 635 + ], + "type": "text", + "content": "Safety alignment. Pure text LLMs often exhibit stronger safety alignment compared to MLLMs [11]. Empirical evidence for this can be found in App. D. Using this insight, we construct two sets of samples, categorized as safe and unsafe, by evaluating the model's response to identical malicious content presented in different modalities: one conveyed through text and the other through text + image. We use these to compute a safety guard steering vector. This steering vector gives the model a higher level of safety, without affecting its usefulness for safe tasks (Table 5). We evaluate the safety of the model by ASR (attack success rate) metric, described in detail in App. D." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 645, + 125, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 125, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 125, + 658 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Limitations. The effectiveness of our method relies on the fact that the concepts are represented by linear directions in latent space. However, recent work [14] has found that not all features are captured as such: thus, when analyzing recovery" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 289, + 296, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 322 + ], + "type": "text", + "content": "Table 3. Steering MLLMs captions style. Each line corresponds to a different steering vector. Steering towards a target style increases the number of captions with that style." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 214 + ], + "type": "table", + "html": "
ModelTotalMethodGendered → Neutral
LLaVA-1.5794coarse232
fine-grained632
Idefics2815coarse237
fine-grained315
Qwen2-VL-Instruct926coarse134
fine-grained300
Before SteeringAfter Steering
A young boy with curly hair is playing a video game.A child with curly hair is playing a video game.
A man riding a dirt bike on a beach.A person riding a dirt bike on a beach.
", + "image_path": "034270970525cdb221aab4d495a30beae03c4c37730514e8cf0164f4f2a07e95.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 340, + 286, + 529, + 311 + ], + "blocks": [ + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "text", + "content": "Table 4. Gender debiasing results: number of occurrences of gendered terms converted to neutral terms across different models and methods, after steering with " + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 221, + 555, + 266 + ], + "type": "text", + "content": ". Below, qualitative samples illustrate changes in descriptions before and after applying steering." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 286, + 529, + 311 + ], + "lines": [ + { + "bbox": [ + 340, + 286, + 529, + 311 + ], + "spans": [ + { + "bbox": [ + 340, + 286, + 529, + 311 + ], + "type": "table", + "html": "
ModelBefore steeringAfter steering
Qwen2-VL-Instruct45/1005/100
", + "image_path": "a8ab7a95122e8571ea6bb46eda34555266e50dccb0cced923d375d5e6c110669.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 384 + ], + "type": "text", + "content": "Table 5. Enhancing MLLM Safety Through Steering. We evaluate safety using the ASR metric, quantifying the proportion of responses that do notrefuse to provide harmful instructions. Our assessment is conducted on a portion of MM-SafetyBench [40] dataset. A lower ASR is desired when the prompt requires harmful instructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 390, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 415 + ], + "type": "text", + "content": "of fine-tuned concepts, it can be interesting to explore more sophisticated similarity measures and matching algorithms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 631 + ], + "type": "text", + "content": "Conclusion. In this work, we introduced a novel concept-based analysis framework for monitoring and controlling MLLM behavior, offering new insights into how latent representations evolve during fine-tuning and across datasets. To address the former, we proposed concept shift vectors, an efficient method for recovering and interpreting concepts in fine-tuned models relative to their original counterparts. This approach naturally led us to explore the latter case, where we demonstrated the ability to steer model behavior by modifying features – without requiring additional training. Our results show that this technique effectively modifies MLLM answers, enabling applications such as gender debiasing, safety control, and enhanced caption generation that highlight different aspects of an image. By releasing our code, we hope our framework will benefit the community and encourage research towards better understanding of MLLMs, as well as their broader applications in domains such as physical and digital agents [50, 60]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 644, + 417, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 644, + 417, + 658 + ], + "spans": [ + { + "bbox": [ + 314, + 644, + 417, + 658 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "This work has been partially supported by ANR grant VISA DEEP (ANR-20-CHIA-0022), HPC resources of IDRIS under the file A0160614966 allocated by GENCI, and Cluster PostGenAI@Paris (ANR-23-IACL-0007, FRANCE 2030)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2213" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 202 + ], + "type": "text", + "content": "[2] Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Rimsky, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 325 + ], + "type": "text", + "content": "[4] Folco Bertini Baldassini, Mustafa Shukor, Matthieu Cord, Laure Soulier, and Benjamin Piwowarski. What makes multimodal in-context learning work? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 1539-1550, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 296, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 296, + 372 + ], + "type": "text", + "content": "[5] Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor V. Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens. ArXiv, abs/2303.08112, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 296, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 296, + 418 + ], + "type": "text", + "content": "[6] Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. Leace: Perfect linear concept erasure in closed form. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 420, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 296, + 485 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 296, + 530 + ], + "type": "text", + "content": "[8] Shi Chen and Qi Zhao. Rex: Reasoning-aware and grounded explanation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15586-15595, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 586 + ], + "type": "text", + "content": "[9] Shuo Chen, Zhen Han, Bailan He, Mark Buckley, Philip Torr, Volker Tresp, and Jindong Gu. Understanding and improving in-context learning on vision-language models. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 589, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 296, + 644 + ], + "type": "text", + "content": "[10] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 646, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 296, + 678 + ], + "type": "text", + "content": "[11] Yi Ding, Bolian Li, and Ruqi Zhang. Eta: Evaluating then aligning safety of vision language models at inference time. ArXiv, abs/2410.06625, 2024. 8, 10, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 713 + ], + "type": "text", + "content": "[12] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "text", + "content": "embodied multimodal language model. arXiv preprint arXiv:2303.03378, 2023.1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[13] Nelson Elhage, Neel Nanda, Catherine Olsson, Tom Henighan, Nicholas Joseph, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 1, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 184 + ], + "type": "text", + "content": "[14] Joshua Engels, Eric J Michaud, Isaac Liao, Wes Gurnee, and Max Tegmark. Not all language model features are linear. arXiv preprint arXiv:2405.14860, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 205 + ], + "type": "text", + "content": "[15] Brandon Huang et al. Multimodal task vectors enable many-shot multimodal in-context learning. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 554, + 228 + ], + "type": "text", + "content": "[16] Yingzhe et al. LIVE: Learnable in-context vector for visual question answering. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 293 + ], + "type": "text", + "content": "[17] Thomas Fel, Victor Boutin, Louis Béthune, Rémi Cadène, Mazda Moayeri, Léo Andéol, Mathieu Chalvidal, and Thomas Serre. A holistic approach to unifying automatic concept extraction and concept importance estimation. Advances in Neural Information Processing Systems, 36, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 294, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 555, + 358 + ], + "type": "text", + "content": "[18] Thomas Fel, Agustin Picard, Louis Bethune, Thibaut Boissin, David Vigouroux, Julien Colin, Rémi Cadène, and Thomas Serre. Craft: Concept recursive activation factorization for explainability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2711-2721, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 361, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 555, + 415 + ], + "type": "text", + "content": "[19] Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, et al. Multimodal autoregressive pre-training of large vision encoders. arXiv preprint arXiv:2411.14402, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 416, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 555, + 469 + ], + "type": "text", + "content": "[20] Jiaxin Ge, Sanjay Subramanian, Trevor Darrell, and Boyi Li. From wrong to right: A recursive approach towards vision-language explanation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 1173–1185, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 471, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 555, + 514 + ], + "type": "text", + "content": "[21] Amirata Ghorbani, James Wexler, James Y Zou, and Been Kim. Towards automatic concept-based explanations. In Advances in Neural Information Processing Systems (NeurIPS), pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 515, + 555, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 555, + 559 + ], + "type": "text", + "content": "[22] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. ArXiv, abs/2311.05608, 2023. 11" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 613 + ], + "type": "text", + "content": "[23] Yunhao Gou, Kai Chen, Zhili Liu, Lanqing Hong, Hang Xu, Zhenguo Li, Dit-Yan Yeung, James T. Kwok, and Yu Zhang. Eyes closed, safety on: Protecting multimodal llms via imageto-text transformation. In European Conference on Computer Vision, 2024. 11" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 614, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 555, + 669 + ], + "type": "text", + "content": "[24] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 7, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[25] J. Edward Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. ArXiv, abs/2106.09685, 2021. 3, 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2214" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[26] Kaichen Huang, Jiahao Huo, Yibo Yan, Kun Wang, Yutao Yue, and Xuming Hu. Miner: Mining the underlying pattern of modality-specific neurons in multimodal large language models. arXiv preprint arXiv:2410.04819, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[27] Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 228 + ], + "type": "text", + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 262 + ], + "type": "text", + "content": "[29] Yuchu Jiang, Jiale Fu, Chenduo Hao, Xinting Hu, Yingzhe Peng, Xin Geng, and Xu Yang. Mimic in-context learning for multimodal tasks. arXiv preprint arXiv:2504.08851, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "text", + "content": "[30] Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning, pages 2668-2677. PMLR, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 385 + ], + "type": "text", + "content": "[31] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A. Shamma, Michael S. Bernstein, and Li Fei-Fei. Visual genome: Connecting language and vision using crowdsourced dense image annotations. Int. J. Comput. Vis., 123(1):32-73, 2017. 4, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 430 + ], + "type": "text", + "content": "[32] Anna Langedijk, Hosein Mohebbi, Gabriele Sarti, Willem Zuidema, and Jaap Jumelet. Decoderlens: Layerwise interpretation of encoder-decoder transformers. ArXiv, abs/2310.03686, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 498 + ], + "type": "text", + "content": "[33] Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander Rush, Douwe Kiela, et al. Obelics: An open web-scale filtered dataset of interleaved image-text documents. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 499, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 295, + 533 + ], + "type": "text", + "content": "[34] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[35] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "text", + "content": "[36] Lijun Li, Bowen Dong, Ruohui Wang, Xuhao Hu, Wangmeng Zuo, Dahua Lin, Yu Qiao, and Jing Shao. Salad-bench: A hierarchical and comprehensive safety benchmark for large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 10" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 700 + ], + "type": "text", + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 8, 1, 3, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 702, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 295, + 713 + ], + "type": "text", + "content": "[38] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Visual instruction tuning. Advances in neural information processing systems, 36, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 554, + 194 + ], + "type": "text", + "content": "[40] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench: A benchmark for safety evaluation of multimodal large language models. In European Conference on Computer Vision, 2023. 8, 10" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 270 + ], + "type": "text", + "content": "[41] Oscar Manas, Pau Rodriguez Lopez, Saba Ahmadi, Aida Nematzadeh, Yash Goyal, and Aishwarya Agrawal. Mapl: Parameter-efficient adaptation of unimodal pre-trained models for vision-language few-shot prompting. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2523-2548, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 293 + ], + "type": "text", + "content": "[42] Neel Nanda. Actually, othello-gpt has a linear emergent world model, 2023. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 337 + ], + "type": "text", + "content": "[43] Nostalgebraist. Interpreting gpt: The logit lens. https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens, 2020. Accessed: [date of access]. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 508, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 508, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 508, + 348 + ], + "type": "text", + "content": "[44] OpenAI. Gpt-4 technical report. arXiv, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 381 + ], + "type": "text", + "content": "[45] Haowen Pan, Yixin Cao, Xiaozhi Wang, and Xun Yang. Finding and editing multi-modal neurons in pre-trained transformer. arXiv preprint arXiv:2311.07470, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "text", + "content": "[46] Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 555, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 479 + ], + "type": "text", + "content": "[47] Jayneel Parekh, Pegah Khayatan, Mustafa Shukor, Alasdair Newson, and Matthieu Cord. A concept-based explainability framework for large multimodal models. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 554, + 525 + ], + "type": "text", + "content": "[48] Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In *Forty-first International Conference on Machine Learning*, 2024. 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 554, + 568 + ], + "type": "text", + "content": "[49] Libo Qin, Qiguang Chen, Hao Fei, Zhi Chen, Min Li, and Wanxiang Che. What factors affect multi-modal in-context learning? an in-depth exploration. arXiv preprint arXiv:2410.20482, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 570, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 623 + ], + "type": "text", + "content": "[50] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025.8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 554, + 689 + ], + "type": "text", + "content": "[51] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 712 + ], + "type": "text", + "content": "[52] Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2215" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "and Neel Nanda. Improving dictionary learning with gated sparse autoencoders. In Advances in Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 149 + ], + "type": "text", + "content": "[53] Shauli Ravfogel, Michael Twiton, Yoav Goldberg, and Ryan D Cotterell. Linear adversarial concept erasure. In International Conference on Machine Learning, pages 18400-18421. PMLR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 296, + 205 + ], + "type": "text", + "content": "[54] Mansi Sakarvadia, Arham Khan, Aswathy Ajith, Daniel Grzenda, Nathaniel Hudson, André Bauer, Kyle Chard, and Ian T. Foster. Attention lens: A tool for mechanistically interpreting the attention head information retrieval mechanism. CoRR, abs/2310.16270, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 259 + ], + "type": "text", + "content": "[55] Sarah Schwettmann, Neil Chowdhury, Samuel Klein, David Bau, and Antonio Torralba. Multimodal neurons in pretrained text-only transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2862-2867, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[56] Mustafa Shukor and Matthieu Cord. Implicit multimodal alignment: On the generalization of frozen llms to multimodal inputs. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 296, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 336 + ], + "type": "text", + "content": "[57] Mustafa Shukor and Matthieu Cord. Skipping computations in multimodal llms. arXiv preprint arXiv:2410.09454, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 381 + ], + "type": "text", + "content": "[58] Mustafa Shukor, Corentin Dancette, and Matthieu Cord. ep-alm: Efficient perceptual augmentation of language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22056-22069, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 437 + ], + "type": "text", + "content": "[59] Mustafa Shukor, Alexandre Rame, Corentin Dancette, and Matthieu Cord. Beyond task performance: evaluating and reducing the flaws of large multimodal models with in-context-learning. In The Twelfth International Conference on Learning Representations, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 502 + ], + "type": "text", + "content": "[60] Mustafa Shukor, Dana Aubakirova, Francesco Capuano, Pepijn Kooijmans, Steven Palma, Adil Zouitine, Michel Aractingi, Caroline Pascal, Martino Russi, Andres Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics. arXiv preprint arXiv:2506.01844, 2025. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 547 + ], + "type": "text", + "content": "[61] Mustafa Shukor, Enrico Fini, Victor Guilherme Turrisi da Costa, Matthieu Cord, Joshua Susskind, and Alaaeldin El-Nouby. Scaling laws for native multimodal models. arXiv preprint arXiv:2504.07951, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "text", + "content": "[62] Nishant Subramani, Nivedita Suresh, and Matthew E Peters. Extracting latent steering vectors from pretrained language models. In Findings of the Association for Computational Linguistics: ACL 2022, pages 566-581, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 656 + ], + "type": "text", + "content": "[63] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 691 + ], + "type": "text", + "content": "[64] Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. Linear representations of sentiment in large language models. arXiv preprint arXiv:2310.15154, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 713 + ], + "type": "text", + "content": "[65] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 475 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "text", + "content": "Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[66] Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[67] Théophane Vallaeys, Mustafa Shukor, Matthieu Cord, and Jakob Verbeek. Improved baselines for data-efficient perceptual augmentation of llms. arXiv preprint arXiv:2403.13499, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "text", + "content": "[68] Yu Wang, Xiaogeng Liu, Yu Li, Muhao Chen, and Chaowei Xiao. Adashield: Safeguarding multimodal large language models from structure-based attack via adaptive shield prompting. In European Conference on Computer Vision, 2024. 10" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "type": "text", + "content": "[69] Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Reft: Representation finetuning for language models. arXiv preprint arXiv:2404.03592, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 555, + 329 + ], + "type": "text", + "content": "[70] Dizhan Xue, Shengsheng Qian, and Changsheng Xu. Few-shot multimodal explanation for visual question answering. In ACM Multimedia 2024, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 374 + ], + "type": "text", + "content": "[71] Chih-Kuan Yeh, Been Kim, Sercan O Arik, Chun-Liang Li, Pradeep Ravikumar, and Tomas Pfister. On concept-based explanations in deep neural networks. arXiv preprint arXiv:1910.07969, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 420 + ], + "type": "text", + "content": "[72] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 555, + 475 + ], + "type": "text", + "content": "[73] Xiaofeng Zhang, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. arXiv preprint arXiv:2406.06579, 2024.1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2216" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Anchor Token Matching_ Implicit Structure Locking for Training-free AR Image Editing/02ecc675-dff6-4c74-84e6-883c770f01a4_content_list.json b/2025/Anchor Token Matching_ Implicit Structure Locking for Training-free AR Image Editing/02ecc675-dff6-4c74-84e6-883c770f01a4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3a2f59876bad324bddfad5a405b6bb38dd56a40b --- /dev/null +++ b/2025/Anchor Token Matching_ Implicit Structure Locking for Training-free AR Image Editing/02ecc675-dff6-4c74-84e6-883c770f01a4_content_list.json @@ -0,0 +1,1642 @@ +[ + { + "type": "text", + "text": "Anchor Token Matching: Implicit Structure Locking for Training-free AR Image Editing", + "text_level": 1, + "bbox": [ + 176, + 131, + 815, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taihang Hu $^{1*}$ , Linxuan Li $^{1*}$ , Kai Wang $^{3,4\\dagger}$ , Yaxing Wang $^{2,1\\dagger}$ , Jian Yang $^{1}$ , Ming-Ming Cheng $^{2,1}$ , $^{1}$ VCIP, College of Computer Science, Nankai University, $^{2}$ NKIARI, Shenzhen Futian $^{3}$ Computer Vision Center, Universitat Autònoma de Barcelona $^{4}$ City University of Hong Kong (Dongguan) {hutaihang00, linxuanli520}@gmail.com kai.wang@cityu-dg.edu.cn, {csjyang, cmm, yaxing} @nankai.edu.cn", + "bbox": [ + 133, + 184, + 861, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce93ed34491fa44a66560e7f6a365afa05cc74f165e645803bf1b2d49993ea85.jpg", + "image_caption": [ + "white horse $\\longrightarrow$ golden horse" + ], + "image_footnote": [], + "bbox": [ + 93, + 301, + 361, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f789f63d9ac0ecaa744df7b259450948146b8920b6a5decd4a01dda67249bc84.jpg", + "image_caption": [ + "a kangaroo wearing sunglasses ..." + ], + "image_footnote": [], + "bbox": [ + 364, + 301, + 632, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42da68ab11b8c4df093b76a15fde26a1d5db04b268d5379507c06fb4c144b031.jpg", + "image_caption": [ + "cartoon style of ..." + ], + "image_footnote": [], + "bbox": [ + 635, + 301, + 903, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c3d9e97dace58395307433b85f6874de64831c9e10eeb2b164864e18e7184ec5.jpg", + "image_caption": [ + "stream $\\rightarrow$ road", + "Figure 1. Our method, Implicit Structure Locking (ISLock), is able to achieve the attribute/object replacement (left), add/remove object (mid) and style/state transfer (right) tasks while maintaining the other information firm to the original image. ISLock is also applicable to different AR-based models, the first row is generated based on LlamaGen [47], while the second row based on Lumina-mGPT [29]" + ], + "image_footnote": [], + "bbox": [ + 93, + 421, + 359, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f36f0703f3ca4ef3981e73ae88d4b9e00c4bc83ac87eb89bfb51d498d19e421.jpg", + "image_caption": [ + "tent in the forest" + ], + "image_footnote": [], + "bbox": [ + 364, + 420, + 632, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08fe57b1179f778e0729f3558e2f8952f24c425994e543ccbf8076258e7030f5.jpg", + "image_caption": [ + "road covered with snow" + ], + "image_footnote": [], + "bbox": [ + 635, + 420, + 903, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 592, + 326, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image generation has seen groundbreaking advancements with diffusion models, enabling high-fidelity synthesis and precise image editing through cross-attention manipulation. Recently, autoregressive (AR) models have re-emerged as powerful alternatives, leveraging next-token generation to match diffusion models. However, existing editing techniques designed for diffusion models fail to translate directly to AR models due to fundamental differences in structural control. Specifically, AR models suffer from spatial poverty of attention maps and sequential accumulation of structural errors during image editing, which disrupt object layouts and global consistency. In this work, we introduce Implicit Structure Locking (ISLock), the first training-free editing strategy for AR visual models. Rather than relying on explicit attention manipulation or fine-tuning, ISLock preserves structural blueprints by dynamically aligning self-attention", + "bbox": [ + 88, + 626, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patterns with reference images through the Anchor Token Matching (ATM) protocol. By implicitly enforcing structural consistency in latent space, our method ISLock enables structure-aware editing while maintaining generative autonomy. Extensive experiments demonstrate that ISLock achieves high-quality, structure-consistent edits without additional training and is superior or comparable to conventional editing techniques. Our findings pioneer the way for efficient and flexible AR-based image editing, further bridging the performance gap between diffusion and autoregressive generative models. The code will be publicly available at https://github.com/hutaiHang/ATM", + "bbox": [ + 511, + 594, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 784, + 645, + 799 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, text-to-image (T2I) generation technologies centered on diffusion models have achieved revolutionary breakthroughs [7, 19, 20, 25, 43]. Foundational models like Stable Diffusion [11, 38] and Imagen [44] have not only driven innovations in artistic creation [3, 13-15], but also catalyzed a rich ecosystem of downstream tasks — ranging", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, † Co-corresponding authors.", + "bbox": [ + 107, + 886, + 370, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18166", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from text-guided image editing [18, 26, 51, 54] and drag manipulation [32, 45] to image-guided T2I generation [33, 34, 59, 60, 62]. These methods generally achieve pixel-level image manipulation by exploring latent features [33, 60] or attention weights [35, 51] during the denoising process.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most recently, autoregressive models (AR) have experienced a resurgence in image generation. Inspired by the success of large language models (LLMs) [1, 28, 49], recent visual autoregressive models [29, 48, 57, 58], including LlamaGen [47] and Emu3 [55], treat images as discrete token sequences [10, 53], reconstructing high-fidelity visuals through next-token prediction. These approaches match diffusion models in long-range coherence while offering unique advantages: their sequential generation mechanism inherently supports localized editing and seamless integration with multimodal language models. However, the potential of AR models remains underexplored for image editing tasks — their next-token generation paradigm fundamentally differs from the parallel denoising mechanism of full latent vectors in diffusion models, which prevents direct migration of existing diffusion-based editing techniques.", + "bbox": [ + 88, + 167, + 485, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This challenge comes from the overlying difference in their structural control mechanisms during image generation. In diffusion models, text-to-image spatial correspondences are explicitly established via cross-attention maps: coarse structural information is locked during early denoising stages, allowing global consistency preservation through localized attention adjustments [18, 21, 51]. In contrast, AR models follow a distinct structural generation logic: Each token prediction strictly depends on preceding sequences, with structural information not centrally determined at any single stage but progressively evolving during generation. This mechanism introduces two critical issues: (1) Spatial Poverty of Attention Maps: Text-to-image attention maps in AR models lack precise structural correspondence*, making them unreliable as editing anchors (as shown in Fig. 4). (2) Sequential Accumulation of Structural Errors: Naive modification of target tokens (e.g. changing \"cat\" to \"dog\") induces localized shifts in latent states. These deviations propagate through autoregressive dependency chains, ultimately distorting global structures such as object poses and scene layouts (as shown in Fig. 2).", + "bbox": [ + 89, + 409, + 486, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent attempts [27, 34] to mitigate these issues involve fine-tuning the model parameters in large-scale paired image editing datasets [2, 12] or introducing distillation losses. However, such approaches require substantial training data and computational resources while sacrificing zero-shot editing flexibility. Thus, a key challenge emerges: How can we achieve structure-consistent editing in text-to-image autoregressive models in a training-free manner, leveraging a deeper understanding of their attention mechanisms?", + "bbox": [ + 89, + 727, + 485, + 863 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3236287dd420e3793c1f73c9fcad0f58157865ae67fd9b64c58d00c8f0e4eefc.jpg", + "image_caption": [ + "Figure 2. (Left) Direct modification of target tokens causes significant content distortion due to sequential structural error accumulation. (Mid) Naive attention injection disrupts content coherence. (Right) In contrast, our implicit structure locking (ISLock) effectively mitigates these issues through the proposed Anchor Token Matching (ATM) strategy." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 905, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we first conduct a systematic investigation into the structural control mechanisms inherent to autoregressive (AR) image generation. While prior research has extensively examined attention-guided editing in diffusion models [18, 35, 51, 54], the relationship between attention maps and structural layout in AR frameworks remains unexplored. Existing diffusion-based approaches [18, 51] suggest that transplanting reference attention maps can effectively enforce structural consistency. However, our experiments reveal a fundamental limitation of this approach when applied to AR models: injecting external attention maps disrupts the intrinsic attention dynamics of AR models, leading to coherence loss with the semantic context in the generated image (as illustrated in Fig. 2-mid, The layout structure is kept but the content is distorted.). This disruption manifests itself as blurred textures, distorted object proportions, and inconsistent lighting, artifacts arising from the inherent incompatibility between parallel attention injection and the sequential dependencies of AR models. To address this challenge, we propose Implicit Structure Locking (ISLock) as the first zero-shot training-free editing strategy for AR visual generative models. At the core of our approach is the Anchor Token Matching (ATM) strategy. Instead of brute-force attention map transplantation, we selectively match tokens during autoregressive decoding by identifying those whose hidden representations exhibit maximum similarity to anchor tokens from the original sequence. This process induces implicit attention alignment, enabling the model to naturally compute attention maps that preserve structural coherence while adapting to local semantic edits (e.g., transitioning fur texture from \"cat\" to \"dog\"). Note that attention consistency emerges as a by-product of ISLock rather than from explicit constraints. And that ISLock simultaneously achieves two critical objectives: (1) Preserving Structural Blueprints: the self-attention maps of the edited sequence maintain structural consistency with those of the reference", + "bbox": [ + 511, + 357, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*Reasonably, the current token attends heavily to the previous one due to the next-token prediction mechanism in AR models.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18167", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image. (2) Maintaining Generative Autonomy: allowing the model to dynamically adjust attention patterns to accommodate the semantic requirements of the edited content, while the non-edit regions are unchanged.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Through extensive comparisons with existing diffusion-based and AR model image editing approaches on the widely used PIE-Bench [23] dataset, our proposed method, ISLock, achieves user satisfactory performance in text-guided image editing. In summary, our key contributions include:", + "bbox": [ + 89, + 151, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct an in-depth study of attention mechanisms in autoregressive (AR) image generation, revealing the limitations of existing diffusion-based structural control methods when applied to AR models.", + "- Based on this finding, we propose the first training-free editing method, Implicit Structure Locking (ISLock), that approximates structural layouts by implicitly matching attention pattern during autoregressive decoding, overcoming spatial inconsistencies in text-guided editing.", + "- We introduce a token-matching mechanism, Anchor Token Matching (ATM), which implicitly preserves key structural elements by identifying anchor tokens in latent space while allowing natural emergence of attention coherence as a byproduct, ensuring semantic consistency without disrupting generative autonomy.", + "- Extensive experiments on the PIE-Bench [23] demonstrate that ISLock significantly maintains the structural consistency and visual fidelity in AR-based image editing." + ], + "bbox": [ + 91, + 227, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 512, + 232, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autoregressive Image Generation. Inspired by the sequence prediction paradigm in LLMs [1, 28], autoregressive (AR) models reformulate image generation as an image token sequence prediction task. The pioneering PixelCNN [52] achieved image synthesis through pixel-wise conditional probability modeling, yet its limited receptive-field constrained global coherence. Subsequent works in discrete representation learning addressed this bottleneck: VQVAE [41] and VQGAN [10] established learnable discrete codebooks that compress images into token sequences, laying the foundation for scalable AR modeling. MaskGIT [6] innovatively introduced masked prediction mechanisms, enabling parallel decoding through bidirectional contextual modeling while preserving autoregressive properties.", + "bbox": [ + 89, + 537, + 485, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the progress of LLMs, researchers have explored cross-modal extensions—LlamaGen [47] adapted LLaMA architectures [49] for visual token modeling, achieving visual results comparable to diffusion models [7, 43, 46], while Emu3 [55] constructed a unified autoregressive space for multimodal joint training. Notably, recent work [9, 58] further integrates visual understanding and generation within a single autoregressive framework, demonstrating task generalization potential. However, existing AR-based approaches encounter fundamental challenges in controllable image edit", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing. Their sequential generation process inherently accumulates errors during localized modifications, leading to discrepancies that conflict with the strict spatial consistency required for precise image editing tasks.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-guided Image Editing. Text-guided image editing aims to modify image content based on semantic prompts while preserving irrelevant regions. Traditional GAN-based methods [16, 30, 36, 36] achieve local editing by optimizing in the GAN latent space with CLIP guidance [22, 37], but their performance is constrained by the capacity of pretrained GANs. Recent diffusion-based approaches have become mainstream, yet they often require complex inversion and latent optimization processes to balance fidelity and controllability. For example, optimization-based inversion methods [17, 26, 31, 54] refine latent noise for accurate reconstruction and manipulate cross-attention maps (e.g., Prompt-to-Prompt [18]) to preserve structure. InstructPix2Pix [2] bypasses inversion by training on paired edit data, but relies on Prompt-to-Prompt [18] to generate largescale image-instruction pairs. To reduce data dependency, some work [5, 8] extracts consecutive video frames as editing samples to simulate real-world editing dynamics. Recent studie [34] adapt the InstructPix2Pix framework to autoregressive models through fine-tuning on paired datasets.", + "bbox": [ + 511, + 152, + 908, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, most existing editing methods remain fragmented due to their underlying architectures—while diffusion models dominate the landscape, AR models remain underexplored, primarily due to the absence of a zero-shot editing framework. A few prior works [27, 34] have attempted to address this limitation through computationally expensive fine-tuning on large-scale paired image editing datasets, often at the cost of zero-shot flexibility. In contrast, our approach unifies spatial control and semantic editing within the AR paradigm, eliminating the need for optimization, or additional model components.", + "bbox": [ + 511, + 454, + 910, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 632, + 604, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to achieve zero-shot text-guided image editing using text-to-image autoregressive (AR) models [47] by systematically investigating the key factors that govern image structure in AR-based generation. We begin by briefly reviewing the paradigm of text-to-image AR models in Section 3.1, which serves as the foundational framework for our approach. Next, in Section 3.2, we establish our motivation through a series of analytical experiments that highlight the structure preservation challenges in AR-based editing. Finally, we introduce our method, Implicit Structure Locking (ISLock), in Section 3.3, detailing its design and effectiveness. An overview of ISLock is illustrated in Fig. 3.", + "bbox": [ + 511, + 657, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 848, + 642, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The AR visual generation model LlamaGen [47] synthesizes images from text by sequentially predicting image tokens. Its", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18168", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/236ffead61d8de29a0d0e9d5feeecbf63646b606c74c2f01bb89c41c1513f302.jpg", + "image_caption": [ + "Figure 3. Our method, ISLock, achieves implicit structural locking through Anchor Token Matching (ATM). Rather than relying on direct attention injection—which often introduces distortions in AR-based visual generation—our approach selects the optimal editing token by identifying the candidate with the smallest distance to the reference token among $K$ candidates." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 903, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "architecture consists of two key components working in synergy: a VQ-Autoencoder [10, 41] that converts images into discrete token sequences and an autoregressive transformer $f_{\\theta}$ that learns the joint distribution of these image tokens. Given an input image $x \\in \\mathcal{R}^{H \\times W \\times 3}$ , the feature encoder $\\mathcal{E}_{VQ}$ first maps it to a latent representation $z^{e} \\in \\mathcal{R}^{h \\times w \\times d}$ , where $d$ denotes the feature dimension. Through nearest-neighbor quantization, each spatial feature vector $z_{i,j}^{e}$ is projected to a codebook prototype $z_{i,j}^{q} \\in \\mathcal{V}$ , generating a discrete token sequence $\\mathbf{Z} = \\{z_{1}, \\dots, z_{h \\times w}\\}$ , where $\\mathcal{V}$ is the learned codebook vector set. To enable text-conditioned generation, LlamaGen [47] integrates a pre-trained T5 [40] text encoder $\\tau_{\\xi}$ that maps the text prompt $\\mathcal{P}$ to a sequence of embeddings $c = \\tau_{\\xi}(\\mathcal{P})$ . The text embeddings are projected into the transformer input space through a linear layer and prepended to the image token sequence. Following that, the AR model $f_{\\theta}$ autoregressively predicts the joint distribution over image tokens conditioned on the text:", + "bbox": [ + 91, + 369, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (z | c) = \\prod_ {i = 1} ^ {N} P \\left(z _ {i} \\mid z _ {< i}, c\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 645, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N = h \\times w$ . During generation, the concatenated sequence $[c; z_{